[ { "paper_id": "zzR1Uskhj0", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Ruiyuan Huang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Zengfeng Huang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 75, "n_formula_1": 24, "n_ref_uni": 16, "n_ref": 61, "n_ref_all": 67, "n_fig": 1, "n_tab": 5, "L_tab": 1160, "n_element_tab": 208, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8883, "formula_len_all_1": 2196, "len_all": 138213, "len_all_1": 70644, "len_abs": 1446, "len_title": 143, "len_sents": 38664, "len_sents_1": 25034, "n_sents": 402, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1468, "title": "High Probability Bounds for Cross-Learning Contextual Bandits with Unknown Context Distributions", "abs": "Motivated by applications in online bidding and sleeping bandits, we examine the problem of contextual bandits with cross learning, where the learner observes the loss associated with the action across all possible contexts, not just the current round\u2019s context. Our focus is on a setting where losses are chosen adversarially, and contexts are sampled i.i.d. from a specific distribution. This problem was first studied by Balseiro et al. (2019), who proposed an algorithm that achieves near-optimal regret under the assumption that the context distribution is known in advance. However, this assumption is often unrealistic. To address this issue, Schneider & Zimmert (2023) recently proposed a new algorithm that achieves nearly optimal expected regret. It is well-known that expected regret can be significantly weaker than high-probability bounds. In this paper, we present a novel, in-depth analysis of their algorithm and demonstrate that it actually achieves near-optimal regret with $\\textit{high probability}$. There are steps in the original analysis by Schneider & Zimmert (2023) that lead only to an expected bound by nature. In our analysis, we introduce several new insights. Specifically, we make extensive use of the weak dependency structure between different epochs, which was overlooked in previous analyses. Additionally, standard martingale inequalities are not directly applicable, so we refine martingale inequalities to complete our analysis.", "keywords": [ "contextual bandits", "cross-learning", "high-probability bounds" ], "rating_list": [ 6, 6, 8, 5, 5 ], "soundness_list": [ 3, 4, 4, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 1 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 3, 3 ] }, { "paper_id": "zz9jAssrwL", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Jangwon Kim", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Jonghyeok Park", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Soohee Han", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Yoonhee Gil", "gender": "Female", "institution": "MakinaRocks", "country": "KR", "position": "Researcher" }, { "name": "Yoonsu Jang", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 28, "n_ref_uni": 45, "n_ref": 62, "n_ref_all": 81, "n_fig": 6, "n_tab": 3, "L_tab": 160, "n_element_tab": 15, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2376, "formula_len_all_1": 1824, "len_all": 114860, "len_all_1": 57986, "len_abs": 879, "len_title": 126, "len_sents": 32164, "len_sents_1": 26684, "n_sents": 241, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 884, "title": "Bayesian Policy Distillation via Offline RL for Lightweight and Fast Inference", "abs": "High-performance deep reinforcement learning faces tremendous challenges when implemented on cost-effective low-end embedded systems due to its heavy computational burden. To address this issue, we propose a policy distillation method called Bayesian Policy Distillation (BPD), which effectively retrains small-sized neural networks through an offline reinforcement learning approach. BPD exploits Bayesian neural networks to distill already designed high-performance policy networks by adopting value optimizing, behavior cloning, and sparsity-inducing strategies. Simulation results reveal that the proposed BPD successfully compresses the policy networks, making them lighter and achieving faster inference time. Furthermore, the proposed approach is demonstrated with a real inverted pendulum system and reduced the inference time and memory size by 78 \\% and 98 \\%, respectively.", "keywords": [ "neural network compression", "reinforcement learning", "robot learning" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 4, 2, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "zyGrziIVdE", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Dennis George Wilson", "gender": "Male", "institution": "Institut Sup\u00e9rieur de l'A\u00e9ronautique et de l'Espace", "country": "FR", "position": "Associate Professor" }, { "name": "Emmanuel Rachelson", "gender": "Male", "institution": "Institut Sup\u00e9rieur de l'A\u00e9ronautique et de l'Espace", "country": "FR", "position": "Full Professor" }, { "name": "Paul-Antoine LE TOLGUENEC", "gender": "Male", "institution": "Institut Sup\u00e9rieur de l'A\u00e9ronautique et de l'Espace", "country": "FR", "position": "PhD student" }, { "name": "Yann Besse", "gender": "unknown", "institution": "Institut Sup\u00e9rieur de l'A\u00e9ronautique et de l'Espace", "country": "", "position": "Instructor" }, { "name": "Florent Teichteil-K\u00f6nigsbuch", "gender": "Male", "institution": "AIRBUS SAS", "country": "FR", "position": "Researcher" } ], "n_formula": 39, "n_formula_1": 23, "n_ref_uni": 32, "n_ref": 70, "n_ref_all": 85, "n_fig": 3, "n_tab": 5, "L_tab": 1826, "n_element_tab": 136, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 812, "n_element_tab_1": 57, "formula_len_all": 6086, "formula_len_all_1": 2024, "len_all": 134740, "len_all_1": 64496, "len_abs": 1147, "len_title": 89, "len_sents": 41618, "len_sents_1": 29282, "n_sents": 328, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1200, "title": "Exploration by Running Away from the Past", "abs": "The ability to explore efficiently and effectively is a central challenge of reinforcement learning.\nIn this work, we consider exploration through the lens of information theory.\nSpecifically, we cast exploration as a problem of maximizing the Shannon entropy of the state occupation measure.\nThis is done by maximizing a sequence of divergences between distributions representing an agent's past behavior and its current behavior.\nIntuitively, this encourages the agent to explore new behaviors that are distinct from past behaviors.\nHence, we call our method RAMP, for ``$\\textbf{R}$unning $\\textbf{A}$way fro$\\textbf{m}$ the $\\textbf{P}$ast.''\nA fundamental question of this method is the quantification of the distribution change over time.\nWe consider both the Kullback-Leibler divergence and the Wasserstein distance to quantify divergence between successive state occupation measures, and explain why the former might lead to undesirable exploratory behaviors in some tasks. \nWe demonstrate that by encouraging the agent to explore by actively distancing itself from past experiences, it can effectively explore mazes and a wide range of behaviors on robotic manipulation and locomotion tasks.", "keywords": [ "Reinforcement Learning", "Exploration", "Deep Learning" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "zxqdVo9FjY", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Jiping Li", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Undergrad student" }, { "name": "Rishi Sonthalia", "gender": "Male", "institution": "Boston College", "country": "US", "position": "Assistant Professor" } ], "n_formula": 175, "n_formula_1": 42, "n_ref_uni": 37, "n_ref": 117, "n_ref_all": 126, "n_fig": 1, "n_tab": 3, "L_tab": 389, "n_element_tab": 21, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 181, "n_element_tab_1": 3, "formula_len_all": 24432, "formula_len_all_1": 3655, "len_all": 206203, "len_all_1": 61386, "len_abs": 1683, "len_title": 122, "len_sents": 51474, "len_sents_1": 23124, "n_sents": 646, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 940, "title": "Generalization for Least Squares Regression with Simple Spiked Covariances", "abs": "Random matrix theory has proven to be a valuable tool in analyzing the generalization of linear models. However, the generalization properties of even two-layer neural networks trained by gradient descent remain poorly understood. To understand the generalization performance of such networks, it is crucial to characterize the spectrum of the feature matrix at the hidden layer.\nRecent work has made progress in this direction by describing the spectrum after a single gradient step, revealing a spiked covariance structure. Yet, the generalization error for linear models with spiked covariances has not been previously determined.\nThis paper addresses this gap by examining two simple models exhibiting spiked covariances. We derive their generalization error in the asymptotic proportional regime. Our analysis demonstrates that the eigenvector and eigenvalue corresponding to the spike significantly influence the generalization error.", "keywords": [ "Generalization", "Random Matrix Theory", "Spiked Covariance", "Two Layer Network", "Layer Wise Training" ], "rating_list": [ 6, 5, 5, 3, 5 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 4, 2, 3, 1, 3 ], "contribution_list": [ 3, 2, 2, 1, 1 ], "confidence_list": [ 3, 3, 4, 4, 3 ] }, { "paper_id": "zxg6601zoc", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 11, "author_info_list": [ { "name": "Cheng Han", "gender": "Male", "institution": "University of Missouri - Kansas City", "country": "US", "position": "Assistant Professor" }, { "name": "Dongfang Liu", "gender": "unknown", "institution": "Rochester Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "James Chenhao Liang", "gender": "Male", "institution": "U. S. Naval Research Laboratory", "country": "US", "position": "Researcher" }, { "name": "Lifu Huang", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Assistant Professor" }, { "name": "MAJID RABBANI", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Qifan Wang", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Principal Researcher" }, { "name": "Raghuveer Rao", "gender": "unknown", "institution": "DEVCOM Army Research Laboratory", "country": "US", "position": "Researcher" }, { "name": "Ruixiang Tang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Assistant Professor" }, { "name": "Sohail Dianat", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Yiyang Liu", "gender": "Male", "institution": "University of Missouri", "country": "US", "position": "PhD student" }, { "name": "Yugyung Lee", "gender": "Female", "institution": "University of Missouri - Kansas City", "country": "", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 67, "n_ref": 162, "n_ref_all": 188, "n_fig": 8, "n_tab": 12, "L_tab": 3543, "n_element_tab": 372, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 228, "n_element_tab_1": 36, "formula_len_all": 676, "formula_len_all_1": 676, "len_all": 227778, "len_all_1": 69317, "len_abs": 1308, "len_title": 113, "len_sents": 58300, "len_sents_1": 33077, "n_sents": 415, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1314, "title": "Re-Imagining Multimodal Instruction Tuning: A Representation View", "abs": "Multimodal instruction tuning has proven to be an effective strategy for achieving zero-shot generalization by fine-tuning pre-trained Large Multimodal Models (LMMs) with instruction-following data. However, as the scale of LMMs continues to grow, fully fine-tuning these models has become highly parameter-intensive. Although Parameter-Efficient Fine-Tuning (PEFT) methods have been introduced to reduce the number of tunable parameters, a significant performance gap remains compared to full fine-tuning. Furthermore, existing PEFT approaches are often highly parameterized, making them difficult to interpret and control. In light of this, we introduce Multimodal Representation Tuning (MRT), a novel approach that focuses on directly editing semantically rich multimodal representations to achieve strong performance and provide intuitive control over LMMs. Empirical results show that our method surpasses current state-of-the-art baselines with significant performance gains (e.g., 1580.40 MME score) while requiring substantially fewer tunable parameters (e.g., 0.03% parameters). Additionally, we conduct experiments on editing instrumental tokens within multimodal representations, demonstrating that direct manipulation of these representations enables simple yet effective control over network behavior.", "keywords": [ "Representation Tuning", "Large Multimodal Models", "Parameter-efficient Fine-tuning" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "zxO4WuVGns", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Dominik Straub", "gender": "unknown", "institution": "TU Darmstadt", "country": "", "position": "PhD student" }, { "name": "Jan Peters", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Principal Researcher" }, { "name": "Tobias F. Niehues", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Constantin Rothkopf", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "Full Professor" } ], "n_formula": 21, "n_formula_1": 9, "n_ref_uni": 59, "n_ref": 82, "n_ref_all": 118, "n_fig": 12, "n_tab": 1, "L_tab": 2, "n_element_tab": 1, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1401, "formula_len_all_1": 433, "len_all": 156224, "len_all_1": 67319, "len_abs": 1774, "len_title": 111, "len_sents": 52048, "len_sents_1": 33583, "n_sents": 377, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1782, "title": "Inverse decision-making using neural amortized Bayesian actors", "abs": "Bayesian observer and actor models have provided normative explanations for many behavioral phenomena in perception, sensorimotor control, and other areas of cognitive science and neuroscience. They attribute behavioral variability and biases to interpretable entities such as perceptual and motor uncertainty, prior beliefs, and behavioral costs. However, when extending these models to more naturalistic tasks with continuous actions, solving the Bayesian decision-making problem is often analytically intractable. Inverse decision-making, i.e. performing inference over the parameters of such models given behavioral data, is computationally even more difficult. Therefore, researchers typically constrain their models to easily tractable components, such as Gaussian distributions or quadratic cost functions, or resort to numerical approximations. To overcome these limitations, we amortize the Bayesian actor using a neural network trained on a wide range of parameter settings in an unsupervised fashion. Using the pre-trained neural network enables performing efficient gradient-based Bayesian inference of the Bayesian actor model's parameters. We show on synthetic data that the inferred posterior distributions are in close alignment with those obtained using analytical solutions where they exist. Where no analytical solution is available, we recover posterior distributions close to the ground truth. We then show how our method allows for principled model comparison and how it can be used to disentangle factors that may lead to unidentifiabilities between priors and costs. Finally, we apply our method to empirical data from three sensorimotor tasks and compare model fits with different cost functions to show that it can explain individuals' behavioral patterns.", "keywords": [ "Bayesian actor models", "perception and action", "cognitive science", "Bayesian inference", "inverse modeling" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 2, 4, 3 ], "confidence_list": [ 2, 2, 3 ] }, { "paper_id": "zwuemuTiN8", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Ankitkumar Joshi", "gender": "Male", "institution": "University of Pittsburgh", "country": "", "position": "PhD student" }, { "name": "Milos Hauskrecht", "gender": "Male", "institution": "University of Pittsburgh", "country": "US", "position": "Full Professor" } ], "n_formula": 21, "n_formula_1": 18, "n_ref_uni": 21, "n_ref": 39, "n_ref_all": 75, "n_fig": 16, "n_tab": 6, "L_tab": 1738, "n_element_tab": 146, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 4695, "n_element_tab_1": 243, "formula_len_all": 1484, "formula_len_all_1": 886, "len_all": 186937, "len_all_1": 73770, "len_abs": 1606, "len_title": 144, "len_sents": 76854, "len_sents_1": 33071, "n_sents": 586, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1621, "title": "TACD-GRU: Time-Aware Context-Dependent Autoregressive Model for Irregularly Sampled Time Series", "abs": "Multivariate time series data and their models are extremely important for under-\nstanding the behavior of various natural and man-made systems. Development of\naccurate time series models often requires capturing intricate relationships among\nthe variables and their dynamics. Particularly challenging to model and learn\nare time series with irregular and sparse observations, that may arise in domains\nas diverse as healthcare, sensor and communication networks. In this work, we\npropose and study TACD-GRU, a new Time- Aware Context-Dependent Gated\nRecurrent Unit framework for multivariate time series prediction (or forecasting)\nthat accounts for irregularities in observation times of individual time series vari-\nables and their dependencies. Our framework defines a novel sequential unit that\nis triggered by the arrival of a new observation to update its state, and a predic-\ntion module that supports time series predictions at any future time. The current\nprediction module consists of and combines two novel prediction models: (i) a\ncontext-based model (TACD-GRU-CONTEXT) that relies on a set of tunable latent\ndecay functions of time and their linear combinations to support the prediction,\nand (ii) an attention-based model (TACD-GRU-ATTENTION) that models depen-\ndencies among variables and their most recent values using a temporal attention\nmechanism. Our model shows highly competitive performance when powered by\nboth individual and combined prediction functions outperforming existing state-of-\nthe-art (SOTA) models on both single-step and multi-step prediction tasks across\nthree real-world datasets.", "keywords": [ "Time series models", "Irregularly sampled time-series", "Autoregressive models", "Recurrent neural networks" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "zweyouirw7", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chenlin Zhou", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "", "position": "Engineer" }, { "name": "Dongyang Ma", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jianing Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Wei Zhang", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yonghong Tian", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Zhengyu Ma", "gender": "Female", "institution": "Peng Cheng Lab", "country": "CN", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 14, "n_ref_uni": 27, "n_ref": 46, "n_ref_all": 70, "n_fig": 13, "n_tab": 8, "L_tab": 3165, "n_element_tab": 235, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 450, "n_element_tab_1": 138, "formula_len_all": 1296, "formula_len_all_1": 739, "len_all": 128117, "len_all_1": 49976, "len_abs": 1451, "len_title": 104, "len_sents": 32610, "len_sents_1": 22340, "n_sents": 247, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1458, "title": "Spiking Transformer-CNN for Event-based Object Detection", "abs": "Spiking Neural Networks (SNNs) enable energy-efficient computation through event-driven computing and multiplication-free inference, making them well-suited for processing sparse events. Recently, deep Spiking Convolutional Neural Networks (CNNs) have shown energy efficiency advantages on event-based object detection. However, spiking CNNs have been limited to local and single-scale features, making it challenging for them to achieve better detection accuracy. To address this challenge, we propose a hierarchical Spiking Transformer-CNN (i.e., Spike-TransCNN) architecture, which is the first attempt to leverage the global information extraction capabilities of Spiking Transformers and the local information capture abilities of Spiking CNNs for event-based object detection. Technically, we first propose using the Spiking Transformer to extract global features and employ a multi-scale local feature extraction CNN module to complement the Spiking Transformers in local feature extraction. Then, we design intra-stage and inter-stage feature fusion modules to integrate global and multi-scale local features within the network architecture. Experimental results demonstrate that our Spike-TransCNN significantly outperforms existing SNN-based object detectors on the Gen1 dataset, achieving higher detection accuracy (mAP 0.336 vs. 0.321) with lower energy consumption (5.49 mJ vs. 7.26 mJ). Our code can be available in the supplementary materials.", "keywords": [ "Event data", "Object detection", "Spike neural networks", "Low power consumption", "Transformer-CNN" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "zvoM1Wastw", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Aiqi Zhang", "gender": "unknown", "institution": "Wilfrid Laurier University", "country": "CA", "position": "Assistant Professor" }, { "name": "Chenjia Bai", "gender": "Male", "institution": "TeleAI, China Telecom", "country": "CN", "position": "Research Scientist" }, { "name": "Rushuai Yang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Xiu Su", "gender": "unknown", "institution": "Central South University", "country": "CN", "position": "Full Professor" }, { "name": "Yi Chen", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 113, "n_formula_1": 27, "n_ref_uni": 71, "n_ref": 99, "n_ref_all": 147, "n_fig": 5, "n_tab": 2, "L_tab": 364, "n_element_tab": 15, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6687, "formula_len_all_1": 1229, "len_all": 196220, "len_all_1": 70518, "len_abs": 1314, "len_title": 108, "len_sents": 47507, "len_sents_1": 31868, "n_sents": 421, "n_sents_1": 248, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1320, "title": "A Provable Quantile Regression Adapter via Transfer Learning", "abs": "Adapter-tuning strategy is an efficient method in machine learning that introduces lightweight and sparse trainable parameters into a pretrained model without altering the original parameters (e.g., low-rank adaptation of large language models). Nevertheless, most existing adapter-tuning approaches are developed for risk-neutral task objectives and the study on the adaptation of risk-sensitive tasks is limited. In this paper, we propose a transfer learning-based quantile regression adapter to improve the estimation of quantile-related risks by leveraging existing pretrained models. We also establish a theoretical analysis to quantify the efficacy of our quantile regression adapter. Particularly, we introduce a transferability measure that characterizes the intrinsic similarity between the pretrained model and downstream task in order to explain when transferring knowledge can improve downstream learning. Under appropriate transferability and structural assumptions, we establish error bounds for the estimation and out-of-sample prediction quality by our quantile regression adapter. Compared to vanilla approaches without transfer learning, our method is provably more sample efficient. Extensive numerical simulations are conducted to demonstrate the superiority and robustness of our method empirically.", "keywords": [ "Transfer Learning", "Adaptation", "Quantile Regression", "High-dimensional Statistics", "Convergence Rate" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 1, 2, 1, 1 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "zvaiz3FjA9", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Ashish Kumar", "gender": "Male", "institution": "Indian Institute of Technology, Kanpur", "country": "", "position": "PhD student" }, { "name": "Jaesik Park", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 140, "n_ref_all": 190, "n_fig": 15, "n_tab": 5, "L_tab": 4341, "n_element_tab": 545, "n_fig_1": 16, "n_tab_1": 5, "L_tab_1": 3635, "n_element_tab_1": 484, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 130030, "len_all_1": 65831, "len_abs": 960, "len_title": 95, "len_sents": 32626, "len_sents_1": 25501, "n_sents": 254, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1006, "title": "Designing Concise ConvNets with Columnar Stages", "abs": "In the era of vision Transformers, the recent success of VanillaNet shows the huge\npotential of simple and concise convolutional neural networks (ConvNets). Where\nsuch models mainly focus on runtime, it is also crucial to simultaneously focus\non other aspects, e.g., FLOPs, parameters, etc, to strengthen their utility further.\nTo this end, we introduce a refreshing ConvNet macro design called Columnar\nStage Network (CoSNet). CoSNet has a systematically developed simple and\nconcise structure, smaller depth, low parameter count, low FLOPs, and attention-\nless operations, well suited for resource-constrained deployment. The key novelty\nof CoSNet is deploying parallel convolutions with fewer kernels fed by input\nreplication, using columnar stacking of these convolutions, and minimizing the use\nof 1\u00d71 convolution layers. Our comprehensive evaluations show that CoSNet rivals\nmany renowned ConvNets and Transformer designs under resource-constrained\nscenarios. Pretrained models shall be open-sourced.", "keywords": [ "Convolutional Neural Networks", "Columnar Stages", "Input Replication", "Image Classification", "Detection" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "zvYJ1qG1Fy", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Hui Chen", "gender": "Male", "institution": "Macquarie University", "country": "AU", "position": "PhD student" }, { "name": "Jin Li", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Longbing Cao", "gender": "Male", "institution": "Macquarie University", "country": "AU", "position": "Full Professor" }, { "name": "Xuhui Fan", "gender": "Male", "institution": "Macquarie University", "country": "AU", "position": "Lecturer" }, { "name": "Zhilin Zhao", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "zhangkai wu", "gender": "Male", "institution": "Macquarie University", "country": "AU", "position": "Postdoc" } ], "n_formula": 18, "n_formula_1": 9, "n_ref_uni": 53, "n_ref": 85, "n_ref_all": 104, "n_fig": 12, "n_tab": 6, "L_tab": 2368, "n_element_tab": 96, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 1703, "n_element_tab_1": 70, "formula_len_all": 3038, "formula_len_all_1": 922, "len_all": 148572, "len_all_1": 62858, "len_abs": 1467, "len_title": 106, "len_sents": 40644, "len_sents_1": 29003, "n_sents": 288, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1473, "title": "Parameter Space Representation Learning on Mixed-type Data", "abs": "A significant challenge in representation learning is to capture latent semantics in data mixing continuous, discrete, and even discretized observations (called mixedtype data), encountering issues like inconsistent discoveries and redundant modeling. Recently, Bayesian flow networks (BFNs) offer a unified strategy to represent such mixed-type data in the parameter space but cannot learn low-dimensional latent semantics since BFNs assume the size of parameters being the same as that of observations. This raises a new important question: how to learn latent semantics in parameter spaces rather than in observation spaces of mixed-type data? Accordingly, we propose a novel unified parameter space representation learning framework, ParamReL, which extracts progressive latent semantics in parameter spaces of mixed-type data. In ParamReL, a self-encoder learns latent semantics from intermediate parameters rather than observations. The learned semantics are then integrated into BFNs to efficiently learn unified representations of mixed-type data. Additionally, a reverse-sampling procedure can empower BFNs for tasks including input reconstruction and interpolation. Extensive experiments verify the effectiveness of ParamReL in learning parameter space representations for latent interpolation, disentanglement, time-varying conditional reconstruction, and conditional generation. The code is available at https: //anonymous.4open.science/r/ICLR25-F087/README.md.", "keywords": [ "Representation learning; Parameter space; Diffusion model; Bayesian flow networks" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 1, 1, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "zv9jedBExg", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Hideaki Iiduka", "gender": "Male", "institution": "Meiji University", "country": "JP", "position": "Full Professor" }, { "name": "Naoki Sato", "gender": "Male", "institution": "Meiji University", "country": "JP", "position": "MS student" } ], "n_formula": 126, "n_formula_1": 29, "n_ref_uni": 69, "n_ref": 113, "n_ref_all": 177, "n_fig": 11, "n_tab": 3, "L_tab": 536, "n_element_tab": 5, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 426, "n_element_tab_1": 54, "formula_len_all": 14174, "formula_len_all_1": 2654, "len_all": 229552, "len_all_1": 74413, "len_abs": 1171, "len_title": 142, "len_sents": 54957, "len_sents_1": 31724, "n_sents": 489, "n_sents_1": 230, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1173, "title": "Role of Momentum in Smoothing Objective Function and Generalizability of Deep Neural Networks", "abs": "For nonconvex objective functions, including deep neural networks, stochastic gradient descent (SGD) with momentum has faster convergence and better generalizability than SGD without momentum, but a theoretical explanation for this is lacking. Adding momentum is thought to reduce stochastic noise, but several studies have argued that stochastic noise actually contributes to the generalizability of the model, which raises a contradiction. We show that the stochastic noise in SGD with momentum smoothes the objective function, the degree of which is determined by the learning rate, the batch size, the momentum factor, the variance of the stochastic gradient, and the upper bound of the gradient norm. By numerically deriving the stochastic noise level in SGD with and without momentum, we provide theoretical findings that help explain the training dynamics of SGD with momentum, which were not explained by previous studies on convergence and stability, and that resolve the contradiction. We also provide experimental results for an image classification task using ResNets that support our assertion that model generalizability depends on the stochastic noise level.", "keywords": [ "deep learning theory", "degree of smoothing", "generalizability", "nonconvex optimization", "SGD with momentum", "smoothing property" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "zuuhtmK1Ub", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Alexander Ryabov", "gender": "Male", "institution": "Skolkovo Institute Of Science and Technology", "country": "", "position": "Researcher" }, { "name": "Evgeny Burnaev", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Full Professor" }, { "name": "Nikolay Yavich", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Researcher" }, { "name": "Vladimir Vanovskiy", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "Researcher" } ], "n_formula": 22, "n_formula_1": 20, "n_ref_uni": 20, "n_ref": 21, "n_ref_all": 46, "n_fig": 10, "n_tab": 2, "L_tab": 48, "n_element_tab": 10, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 48, "n_element_tab_1": 10, "formula_len_all": 891, "formula_len_all_1": 866, "len_all": 68729, "len_all_1": 50710, "len_abs": 1097, "len_title": 138, "len_sents": 20843, "len_sents_1": 19565, "n_sents": 191, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 886, "title": "Differentiable Implicit Solver on Graph Neural Networks for Forward and Inverse Problems", "abs": "Partial differential equations (PDEs) on unstructured grids can be solved using message passing on a graph neural network (GNN). Implicit time-stepping schemes are often favored, especially for parabolic PDEs, due to their stability properties. In this work, we develop a fully differentiable implicit solver for unstructured grids. We evaluate its performance across four key tasks: a) forward modeling of stiff evolutionary and static problems; b) the inverse problem of estimating equation coefficients; c) the inverse problem of estimating the right-hand side; and d) graph coarsening to accelerate forward modeling. The increased stability and differentiability of our solver enable excellent results in reducing the complexity of forward modeling and efficiently solving related inverse problems. This makes it a promising tool for geoscience and other physics-based applications.", "keywords": [ "Graph Neural Networks", "Differentiable solvers", "Implicit schemes", "Numerical modelling", "Inverse problems" ], "rating_list": [ 1, 3, 1, 3 ], "soundness_list": [ 2, 1, 2, 1 ], "presentation_list": [ 1, 1, 1, 1 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "zuOnOAHBMy", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Fei Dou", "gender": "Female", "institution": "University of Georgia", "country": "US", "position": "Assistant Professor" }, { "name": "Jiaxi Li", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Jin Lu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Assistant Professor" }, { "name": "Shizhe Liang", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Postdoc" }, { "name": "Tianming Liu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Professor" }, { "name": "Xiaofeng Yang", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Associate Professor" }, { "name": "Yi Pan", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "huaqin zhao", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" } ], "n_formula": 99, "n_formula_1": 18, "n_ref_uni": 55, "n_ref": 83, "n_ref_all": 100, "n_fig": 7, "n_tab": 2, "L_tab": 1208, "n_element_tab": 254, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1492, "n_element_tab_1": 206, "formula_len_all": 6352, "formula_len_all_1": 1269, "len_all": 183578, "len_all_1": 64482, "len_abs": 1445, "len_title": 170, "len_sents": 52598, "len_sents_1": 29142, "n_sents": 478, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 122, "L_abs": 1455, "title": "HELENE: Hessian Layer-wise Clipping and Gradient Annealing for Accelerating Fine-tuning LLM with Zeroth-order Optimization", "abs": "Fine-tuning large language models (LLMs) poses significant memory challenges, as the back-propagation process demands extensive resources, especially with growing model sizes. Recent work, MeZO, addresses this issue using a zeroth-order (ZO) optimization method, which reduces memory consumption by matching the usage to the inference phase. However, MeZO experiences slow convergence due to varying curvatures across model parameters. To overcome this limitation, we introduce HELENE, a novel scalable and memory-efficient optimizer that integrates annealed A-GNB gradients with a diagonal Hessian estimation and layer-wise clipping, serving as a second-order pre-conditioner. This combination allows for faster and more stable convergence. Our theoretical analysis demonstrates that HELENE improves convergence rates, particularly for models with heterogeneous layer dimensions, by reducing the dependency on the total parameter space dimension. Instead, the method scales with the largest layer dimension, making it highly suitable for modern LLM architectures. Experimental results on RoBERTa-large and OPT-1.3B across multiple tasks show that HELENE achieves up to a 20\u00d7 speedup compared to MeZO, with average accuracy improvements of 1.5%. Furthermore, HELENE remains compatible with both full parameter tuning and parameter-efficient fine-tuning (PEFT), outperforming several state-of-the-art optimizers. The codes will be released after reviewing.", "keywords": [ "optimization", "large language models" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "zuKrRYM3Tg", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Armand Foucault", "gender": "Male", "institution": "Universit\u00e9 de Toulouse", "country": "", "position": "PhD student" }, { "name": "Franck Mamalet", "gender": "Male", "institution": "IRT Saint Exupery", "country": "FR", "position": "Researcher" }, { "name": "Francois Malgouyres", "gender": "Male", "institution": "Universit\u00e9 Paul Sabatier", "country": "", "position": "Full Professor" } ], "n_formula": 40, "n_formula_1": 14, "n_ref_uni": 54, "n_ref": 119, "n_ref_all": 168, "n_fig": 3, "n_tab": 11, "L_tab": 3389, "n_element_tab": 617, "n_fig_1": 1, "n_tab_1": 8, "L_tab_1": 1318, "n_element_tab_1": 219, "formula_len_all": 2053, "formula_len_all_1": 552, "len_all": 216151, "len_all_1": 66396, "len_abs": 987, "len_title": 109, "len_sents": 59215, "len_sents_1": 26861, "n_sents": 476, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 992, "title": "Quantized Approximately Orthogonal Recurrent Neural Networks", "abs": "In recent years, Orthogonal Recurrent Neural Networks (ORNNs) have gained popularity due to their ability to manage tasks involving long-term dependencies, such as the copy task, and their linear complexity. However, existing ORNNs utilize full precision weights and activations, which prevents their deployment on compact devices.\n\nIn this paper, we explore the quantization of the weight matrices in ORNNs, leading to Quantized approximately Orthogonal RNNs (QORNNs). The construction of such networks remained an open problem, acknowledged for its inherent instability. We propose and investigate two strategies to learn QORNN by combining quantization-aware training (QAT) and orthogonal projections. We also study post-training quantization of the activations for pure integer computation of the recurrent loop. The most efficient models achieve results similar to state-of-the-art full-precision ORNN, LSTM and FastRNN on a variety of standard benchmarks, even with 3-bits quantization.", "keywords": [ "ecurrent neural networks", "neural network quantization", "orthogonal recurrent neural networks", "quantization bitwidth" ], "rating_list": [ 1, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "zu7cBTPsDb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Ge Zheng", "gender": "unknown", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" }, { "name": "Hanzhuo Huang", "gender": "unknown", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" }, { "name": "Jiepeng Wang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Sibei Yang", "gender": "Female", "institution": "ShanghaiTech University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuan Liu", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Zhiyang Dou", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Visiting PhD" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 119, "n_ref_all": 131, "n_fig": 6, "n_tab": 2, "L_tab": 397, "n_element_tab": 48, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 330, "n_element_tab_1": 45, "formula_len_all": 254, "formula_len_all_1": 255, "len_all": 140603, "len_all_1": 62622, "len_abs": 1183, "len_title": 123, "len_sents": 35672, "len_sents_1": 30408, "n_sents": 264, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1241, "title": "MVTokenFlow: High-quality 4D Content Generation using Multiview Token Flow", "abs": "In this paper, we present MVTokenFlow for high-quality 4D content creation from monocular videos. Recent advancements in generative models such as video diffusion models and multiview diffusion models enable us to create videos or 3D models. However, extending these generative models for dynamic 4D content creation is still a challenging task that requires the generated content to be consistent spatially and temporally. To address this challenge, MVTokenFlow utilizes the multiview diffusion model to generate multiview images on different timesteps, which attains spatial consistency across different viewpoints and allows us to reconstruct a reasonable coarse 4D field. Then, MVTokenFlow further regenerates all the multiview images using the rendered 2D flows as guidance. The 2D flows effectively associate pixels from different timesteps and improve the temporal consistency by reusing tokens in the regeneration process. Finally, the regenerated images are spatiotemporally consistent and utilized to refine the coarse 4D field to get a high-quality 4D field. Experiments demonstrate the effectiveness of our design and show significantly improved quality than baseline methods. Project page: https://soolab.github.io/MVTokenFlow.", "keywords": [ "4D Generation", "Dynamic 3D Gaussian Splatting", "Dynamic Reconstruction", "Diffusion Models" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "ztzZDzgfrh", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Han Li", "gender": "Male", "institution": "Kuaishou", "country": "", "position": "Researcher" }, { "name": "Jun Xu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Kai Zheng", "gender": "Male", "institution": "Kuaishou", "country": "CN", "position": "Researcher" }, { "name": "Weijie Yu", "gender": "Male", "institution": "University of International Business and Economics", "country": "CL", "position": "Assistant Professor" }, { "name": "Xiao Zhang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaoxue Zang", "gender": "Female", "institution": "kuaishou", "country": "CN", "position": "Researcher" }, { "name": "Yang Song", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Researcher" }, { "name": "ZhongXiang Sun", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 10, "n_ref_uni": 54, "n_ref": 109, "n_ref_all": 142, "n_fig": 9, "n_tab": 6, "L_tab": 6537, "n_element_tab": 450, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 489, "n_element_tab_1": 91, "formula_len_all": 1844, "formula_len_all_1": 543, "len_all": 230879, "len_all_1": 53934, "len_abs": 1445, "len_title": 79, "len_sents": 73437, "len_sents_1": 25127, "n_sents": 513, "n_sents_1": 142, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 98, "L_abs": 1348, "title": "ReDeEP: Detecting Hallucination in Retrieval-Augmented Generation via Mechanistic Interpretability", "abs": "Retrieval-Augmented Generation (RAG) models are designed to incorporate external knowledge, reducing hallucinations caused by insufficient parametric (internal) knowledge. However, even with accurate and relevant retrieved content, RAG models can still produce hallucinations by generating outputs that conflict with the retrieved information. Detecting such hallucinations requires disentangling how Large Language Models (LLMs) balance external and parametric knowledge. Current detection methods often focus on one of these mechanisms or without decoupling their intertwined effects, making accurate detection difficult. In this paper, we investigate the internal mechanisms behind hallucinations in RAG scenarios. We discover hallucinations occur when the **Knowledge FFNs** in LLMs overemphasize parametric knowledge in the residual stream, while **Copying Heads** fail to effectively retain or integrate external knowledge from retrieved content. Based on these findings, we propose **ReDeEP**, a novel method that detects hallucinations by decoupling LLM\u2019s utilization of external context and parametric knowledge. Our experiments show that ReDeEP significantly improves RAG hallucination detection accuracy. Additionally, we introduce AARF, which mitigates hallucinations by modulating the contributions of Knowledge FFNs and Copying Heads.", "keywords": [ "Retrieval-Augmented Generation Hallucination", "Hallucination Detection", "Mechanistic Interpretability" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 4, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "zsVZCiYG2r", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Jingyi Liu", "gender": "Female", "institution": "Institute of Semiconductors", "country": "CN", "position": "Postdoc" }, { "name": "Lina Yu", "gender": "Female", "institution": "Institute of Semiconductors, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Min Wu", "gender": "Male", "institution": "Institute of Semiconductors, Chinese Academy of Sciences", "country": "", "position": "Assistant Professor" }, { "name": "Shu Wei", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Weijun Li", "gender": "Male", "institution": "Institute of Semiconductors Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yanjie Li", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Deng Yusong", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 4, "n_ref_uni": 6, "n_ref": 10, "n_ref_all": 29, "n_fig": 4, "n_tab": 10, "L_tab": 5382, "n_element_tab": 239, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2281, "n_element_tab_1": 32, "formula_len_all": 2136, "formula_len_all_1": 324, "len_all": 146351, "len_all_1": 60185, "len_abs": 1346, "len_title": 90, "len_sents": 35052, "len_sents_1": 26939, "n_sents": 305, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 1355, "title": "ChatSR: Conversational Symbolic Regression", "abs": "Formulas are the language of communication between humans and nature. It is an important research topic of artificial intelligence to find expressions from observed data to reflect the relationship between each variable in the data, which is called a symbolic regression problem. The existing symbolic regression methods directly generate expressions according to the given observation data, but we cannot require the algorithm to generate expressions that meet specific requirements according to the known prior knowledge. For example, the expression needs to contain the symbol `$\\sin$' or be periodicity, and so on. Even if it can, it often requires very complex operations, which is very inconvenient. In this paper, based on multi-modal large language models, we propose ChatSR, a conversational symbolic regression method that can generate expressions that meet the requirements simply by describing the requirements with natural language instructions. By experimenting on the test datasets, we can demonstrate that ChatSR leads the state-of-the-art baselines in fitting performance. More notably, ChatSR can well understand the prior knowledge contained in natural language prompts, and can further improve the quality of generated expressions according to the prior knowledge. In addition, it is exciting that ChatSR has good zero-shot capability.", "keywords": [ "Symbolic Regression", "Multi-modal Large Language Models", "Scientific discovery" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "zs6bRl05g8", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Dongsheng Li", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Liangzu Peng", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Xiao Li", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Yifei Shen", "gender": "Male", "institution": "Microsoft Research Asia", "country": "ZW", "position": "Research Cheerleader" }, { "name": "qijun luo", "gender": "Male", "institution": "Chinese University of Hong Kong Shen Zhen", "country": "CN", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 18, "n_ref_uni": 31, "n_ref": 48, "n_ref_all": 74, "n_fig": 9, "n_tab": 8, "L_tab": 3666, "n_element_tab": 646, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 542, "n_element_tab_1": 22, "formula_len_all": 2951, "formula_len_all_1": 1015, "len_all": 158988, "len_all_1": 60216, "len_abs": 1336, "len_title": 129, "len_sents": 47280, "len_sents_1": 28058, "n_sents": 402, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1343, "title": "Accelerating Block Coordinate Descent for LLM Finetuning via Landscape Correction", "abs": "Training and finetuning large language models (LLMs) are resource-intensive tasks, with memory limitations being a key bottleneck. A classic optimization method, block coordinate descent (BCD), offers solutions by segmenting the trainable parameters into multiple blocks and optimizing one active block at a time while freezing the others, thereby significantly reducing memory cost. However, we identify that blindly applying BCD to train LLMs can be inefficient for two reasons. First, optimizing only the active block requires backpropagating through multiple deeper yet inactive blocks, resulting in wasteful computations. Second, the frozen blocks, when they are not quite close to optimality, can narrow the optimization landscape, potentially misguiding the training of the active block. To address these issues simultaneously, we propose integrating BCD with *landscape correction*, which unfreezes the inactive blocks and updates them in a cost-efficient manner during the same backpropagation as the update to the active block. We show that our method empirically improves vanilla BCD with minimal additional computation and memory. Experiments on 8B and 70B models demonstrate that our proposed method surpasses memory efficient baselines and matches Adam's downstream performance while reducing memory cost by 80% compared to Adam.", "keywords": [ "Block coordinate descent", "large language model finetuning" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "zrdkQaf48Z", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Dongkuan Xu", "gender": "Male", "institution": "North Carolina State University", "country": "", "position": "Assistant Professor" }, { "name": "Haisong Gong", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Huanhuan Ma", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Xiaoyuan Yi", "gender": "Male", "institution": "Research, Microsoft", "country": "CN", "position": "Senior Researcher" }, { "name": "Xin Xie", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Senior Principal Researcher" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 54, "n_ref": 80, "n_ref_all": 98, "n_fig": 7, "n_tab": 12, "L_tab": 5112, "n_element_tab": 553, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 925, "n_element_tab_1": 105, "formula_len_all": 331, "formula_len_all_1": 121, "len_all": 159757, "len_all_1": 52786, "len_abs": 1635, "len_title": 156, "len_sents": 36928, "len_sents_1": 24709, "n_sents": 271, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1643, "title": "Leveraging Implicit Sentiments: Enhancing Reliability and Validity in Psychological Trait Evaluation of LLMs", "abs": "Recent advancements in Large Language Models (LLMs) have led to their increasing integration into human life. Understanding their inherent characteristics, such as personalities, temperaments, and emotions, is essential for responsible AI development. However, current psychometric evaluations of LLMs, often derived from human psychological assessments, encounter significant limitations in terms of reliability and validity. Test results reveal that models frequently refuse to provide anthropomorphic responses and exhibit inconsistent scores across various scenarios. Moreover, human-derived theories may not accurately predict model behavior in practical real-world applications.\nTo address these limitations, we propose Core Sentiment Inventory (CSI), a novel evaluation instrument inspired by the Implicit Association Test (IAT). CSI is built from the ground up with a significantly broader range of stimuli words than traditional assessments. CSI covers both English and Chinese to implicitly evaluate models\u2019 sentiment tendencies, which allows for a much more comprehensive assessment.\nThrough extensive experiments, we demonstrate that CSI effectively quantifies models\u2019 sentiments, revealing nuanced emotional patterns that vary significantly across languages and contexts. CSI significantly improves reliability, yielding more consistent results and a reduced reluctance rate, and enhances predictive power by effectively capturing models\u2019 emotional tendencies. These findings validate CSI as a robust and insightful tool for evaluating the psychological traits of LLMs, offering a more reliable alternative to traditional methods.", "keywords": [ "LLM", "Benchmark", "Evaluation", "Psychometrics" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "zrNbsV87Os", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 13, "author_info_list": [ { "name": "Chenlin Meng", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Gengchen Mai", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Hongxu Ma", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Jiaming Song", "gender": "Male", "institution": "Luma AI", "country": "US", "position": "Chief Scientist" }, { "name": "Jinmeng Rao", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Nemin Wu", "gender": "Female", "institution": "University of Georgia", "country": "GE", "position": "PhD student" }, { "name": "Ni Lao", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Qian Cao", "gender": "Female", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Stefano Ermon", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "weiwei sun", "gender": "Male", "institution": "Amazon", "country": "CA", "position": "Researcher" }, { "name": "Yuchi Ma", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Postdoc" }, { "name": "Zeping Liu", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Zhangyu Wang", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 70, "n_ref": 248, "n_ref_all": 304, "n_fig": 18, "n_tab": 6, "L_tab": 1778, "n_element_tab": 298, "n_fig_1": 11, "n_tab_1": 6, "L_tab_1": 2620, "n_element_tab_1": 2, "formula_len_all": 1445, "formula_len_all_1": 492, "len_all": 242843, "len_all_1": 64813, "len_abs": 1819, "len_title": 139, "len_sents": 66006, "len_sents_1": 28309, "n_sents": 434, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1827, "title": "SSIF: Physics-Inspired Implicit Representations for Spatial-Spectral Image Super-Resolution", "abs": "Existing digital sensors capture images at fixed spatial and spectral resolutions (e.g., RGB, multispectral, and hyperspectral images), and generating super-resolution images with different resolution settings requires bespoke machine learning models. Spatial Implicit Functions (SIFs) partially overcome the spatial resolution challenge by representing an image in a spatial-resolution-independent way. However, they\nstill operate at fixed, pre-defined spectral resolutions. To address this challenge, we propose Spatial-Spectral Implicit Function (SSIF), a neural implicit model that represents an image as a function of both continuous pixel coordinates in the spatial domain and continuous wavelengths in the spectral domain. This continuous representation across spatial and spectral domains enables a single model to learn from a diverse set of resolution settings, which leads to better generalizability. This representation also allows the physical principle of spectral imaging and the spectral response functions of sensors to be easily incorporated during training and inference. Moreover, SSIF does not have the equal spectral wavelength interval requirement for both input and output images which leads to much better applicability. We empirically demonstrate the effectiveness of SSIF on two challenging spatial-spectral super-resolution benchmarks. We observe that SSIF consistently outperforms state-of-the-art baselines even when the baselines are allowed to train separate models at each spatial or spectral resolution. We show that SSIF generalizes well to both unseen spatial and spectral resolutions. Moreover, due to its physics-inspired design, SSIF performs significantly better at low data regime and converges faster during training compared with other strong neural implicit function-based baselines.", "keywords": [ "Neural Implicit Function", "Spatial-Spectral Super Resolution", "Spectral Encoding" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 5, 3, 5 ] }, { "paper_id": "zqzsZ5cXbB", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Di He", "gender": "Male", "institution": "Peking University", "country": "", "position": "Assistant Professor" }, { "name": "Jingjing Xu", "gender": "Female", "institution": "Shanghai AI Lab", "country": "", "position": "PhD student" }, { "name": "Jun Zhang", "gender": "Male", "institution": "ByteDance", "country": "CN", "position": "Researcher" }, { "name": "Shengjie Luo", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhenyu He", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Zhi Zhang", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 58, "n_ref": 91, "n_ref_all": 100, "n_fig": 4, "n_tab": 7, "L_tab": 6814, "n_element_tab": 343, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 4104, "n_element_tab_1": 218, "formula_len_all": 698, "formula_len_all_1": 701, "len_all": 155648, "len_all_1": 68354, "len_abs": 1536, "len_title": 99, "len_sents": 32847, "len_sents_1": 28891, "n_sents": 248, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1563, "title": "Let the Code LLM Edit Itself When You Edit the Code", "abs": "In this work, we investigate a typical scenario in code generation where a developer edits existing code in real time and requests a code assistant, e.g., a large language model, to re-predict the next token or next line on the fly. Naively, the LLM needs to re-encode the entire KV cache to provide an accurate prediction. However, this process is computationally expensive, especially when the sequence length is long. Simply encoding the edited subsequence and integrating it to the original KV cache meets the temporal confusion problem, leading to significantly worse performance. We address this efficiency and accuracy trade-off by introducing $\\underline{\\textbf{P}\\text{ositional}\\ \\textbf{I}\\text{ntegrity}\\ \\textbf{E}\\text{ncoding}}$ (PIE). Building upon the rotary positional encoding, PIE first removes the rotary matrices in the Key cache that introduce temporal confusion and then reapplies the correct rotary matrices. This process ensures that positional relationships between tokens are correct and requires only a single round of matrix multiplication. We validate the effectiveness of PIE through extensive experiments on the RepoBench-C-8k dataset, utilizing DeepSeek-Coder models with 1.3B, 6.7B, and 33B parameters. Our evaluation includes three real-world coding tasks: code insertion, code deletion, and multi-place code editing. Results demonstrate that PIE reduces computational overhead by over 85% compared to the standard full recomputation approach across all model sizes and tasks while well approximating the model performance.", "keywords": [ "code generation", "efficiency", "large language model", "code assistant" ], "rating_list": [ 5, 8, 6, 3 ], "soundness_list": [ 2, 4, 4, 1 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 4, 3, 2 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "zqtql1YmlS", "primary_area": "reinforcement learning", "n_author": 11, "author_info_list": [ { "name": "Bo XU", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Chenghao Li", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Chengjie Wu", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Chongjie Zhang", "gender": "unknown", "institution": "Washington University, Saint Louis", "country": "", "position": "Associate Professor" }, { "name": "Dianyu Zhong", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hao Hu", "gender": "Male", "institution": "Moonshot AI", "country": "CN", "position": "Researcher" }, { "name": "Qianchuan Zhao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Quanwei Wang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yiqin Yang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuhua Jiang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ziyou Zhang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 68, "n_formula_1": 28, "n_ref_uni": 44, "n_ref": 65, "n_ref_all": 81, "n_fig": 6, "n_tab": 4, "L_tab": 1805, "n_element_tab": 202, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1165, "n_element_tab_1": 169, "formula_len_all": 4501, "formula_len_all_1": 1604, "len_all": 136860, "len_all_1": 63557, "len_abs": 1098, "len_title": 131, "len_sents": 34637, "len_sents_1": 26763, "n_sents": 334, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1084, "title": "Fewer May Be Better: Enhancing Offline Reinforcement Learning with Reduced Dataset", "abs": "Research in offline reinforcement learning (RL) marks a paradigm shift in RL. However, a critical yet under-investigated aspect of offline RL is determining the subset of the offline dataset, which is used to improve algorithm performance while accelerating algorithm training. Moreover, the size of reduced datasets can uncover the requisite offline data volume essential for addressing analogous challenges. Based on the above considerations, we propose identifying Reduced Datasets for Offline RL (ReDOR) by formulating it as a gradient approximation optimization problem. We prove that the common actor-critic framework in reinforcement learning can be transformed into a submodular objective. This insight enables us to construct a subset by adopting the orthogonal matching pursuit (OMP). Specifically, we have made several critical modifications to OMP to enable successful adaptation with Offline RL algorithms. The experimental results indicate that the data subsets constructed by the ReDOR can significantly improve algorithm performance with low computational complexity.", "keywords": [ "Offline Reinforcement Learning; Data Selection; Grad Match" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "zqo2eKjSWH", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Moyang Guo", "gender": "Male", "institution": "Duke University, Duke University", "country": "US", "position": "MS student" }, { "name": "Yuepeng Hu", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Zhengyuan Jiang", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Neil Gong", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 86, "n_ref_all": 99, "n_fig": 6, "n_tab": 3, "L_tab": 1169, "n_element_tab": 85, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 279, "n_element_tab_1": 50, "formula_len_all": 877, "formula_len_all_1": 129, "len_all": 118908, "len_all_1": 53945, "len_abs": 821, "len_title": 124, "len_sents": 40145, "len_sents_1": 24615, "n_sents": 276, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 824, "title": "Stable Signature is Unstable: Removing Image Watermark from Diffusion Models", "abs": "Watermark has been widely deployed by industry to detect AI-generated images. A recent watermarking framework called Stable Signature (proposed by Meta) roots watermark into the parameters of a diffusion model's decoder such that its generated images are inherently watermarked. Stable Signature makes it possible to watermark images generated by open-source diffusion models and was claimed to be robust against removal attacks. In this work, we propose a new attack to remove the watermark from a diffusion model by fine-tuning it. Our results show that our attack can effectively remove the watermark from a diffusion model such that its generated images are non-watermarked, while maintaining the visual quality of the generated images. Our results highlight that Stable Signature is not as stable as previously thought.", "keywords": [ "Image Watermark", "Diffusion Model", "AI-generated Image" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "zqXANcFO9T", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Jake Perazzone", "gender": "unknown", "institution": "DeepSig, Inc.", "country": "US", "position": "Researcher" }, { "name": "Kevin S Chan", "gender": "Male", "institution": "DEVCOM Army Research Laboratory", "country": "US", "position": "Researcher" }, { "name": "Mingyue Ji", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "Associate Professor" }, { "name": "Shiqiang Wang", "gender": "Male", "institution": "IBM, International Business Machines", "country": "US", "position": "Research Staff Member" }, { "name": "Xin Yao", "gender": "unknown", "institution": "University of Florida", "country": "US", "position": "PhD student" } ], "n_formula": 61, "n_formula_1": 25, "n_ref_uni": 34, "n_ref": 74, "n_ref_all": 90, "n_fig": 4, "n_tab": 7, "L_tab": 1742, "n_element_tab": 208, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 311, "n_element_tab_1": 41, "formula_len_all": 14291, "formula_len_all_1": 2467, "len_all": 146585, "len_all_1": 70969, "len_abs": 2393, "len_title": 126, "len_sents": 40409, "len_sents_1": 30226, "n_sents": 357, "n_sents_1": 248, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1445, "title": "Compressed Decentralized Learning with Error-Feedback under Data Heterogeneity", "abs": "Decentralized learning distributes the training process across multiple nodes, enabling collaborative model training without relying on a central server. Each node performs local training using its own data, with model updates exchanged directly between connected nodes within a given network topology. Various algorithms have been developed within this decentralized learning framework and have been proven to converge under specific assumptions. However, two key challenges remain: 1) ensuring robust performance with both a high degree of gradient compression and data heterogeneity, and 2) providing a general convergence upper bound under commonly used assumptions. To address these challenges, we propose the *Discounted Error-Feedback Decentralized Parallel Stochastic Gradient Descent (DEFD-PSGD)* algorithm, which efficiently manages both high levels of gradient compression and data heterogeneity, without sacrificing communication efficiency. The core idea is to introduce controllable residual error feedback that effectively balances the impact of gradient compression and data heterogeneity. Additionally, we develop novel proof techniques to derive a convergence upper bound under relaxed assumptions. Finally, we present experimental results demonstrating that DEFD-PSGD outperforms other state-of-the-art decentralized learning algorithms, particularly in scenarios involving high compression and significant data heterogeneity.", "keywords": [ "distributed training", "error-feedback", "convergence analysis" ], "rating_list": [ 1, 1, 3 ], "soundness_list": [ 1, 1, 2 ], "presentation_list": [ 2, 1, 3 ], "contribution_list": [ 1, 1, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "zq1zTgSBro", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Andrew Markham", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" }, { "name": "Jia-Xing Zhong", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Niki Trigoni", "gender": "Female", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Sangyun Shin", "gender": "Male", "institution": "Department of Computer Science", "country": "GB", "position": "PhD student" }, { "name": "Shitong Xu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Yuhang He", "gender": "Male", "institution": "Research, Microsoft", "country": "CA", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 57, "n_ref": 116, "n_ref_all": 145, "n_fig": 14, "n_tab": 4, "L_tab": 564, "n_element_tab": 95, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 683, "n_element_tab_1": 98, "formula_len_all": 420, "formula_len_all_1": 421, "len_all": 161719, "len_all_1": 70708, "len_abs": 1264, "len_title": 106, "len_sents": 41626, "len_sents_1": 31920, "n_sents": 313, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1193, "title": "SPEAR: Receiver-to-Receiver Acoustic Neural Warping Field", "abs": "We present SPEAR, a continuous receiver-to-receiver acoustic neural warping field for spatial acoustic effects prediction in an acoustic 3D space with a single stationary audio source. Unlike traditional source-to-receiver modelling methods that require prior space acoustic properties knowledge to rigorously model audio propagation from source to receiver, we propose to predict by warping the spatial acoustic effects from one reference receiver position to another target receiver position, so that the warped audio essentially accommodates all spatial acoustic effects belonging to the target position. SPEAR can be trained in a data much more readily accessible manner, in which we simply ask two robots to independently record spatial audio at different positions. We further theoretically prove the universal existence of the warping field if and only if one audio source presents. Three physical principles are incorporated to guide SPEAR network design, leading to the learned warping field physically meaningful. We demonstrate SPEAR superiority in receiver-to-receiver warping field prediction through detailed experiments on both synthetic, photo-realistic and real-world dataset.", "keywords": [ "Spatial Acoustic Effects", "Receiver-to-Receiver", "Neural Warping Field" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "zpX0teJu9Z", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Andreas Radler", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "PhD student" }, { "name": "Arturs Berzins", "gender": "unknown", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "", "position": "Postdoc" }, { "name": "Eric Volkmann", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "PhD student" }, { "name": "Johannes Brandstetter", "gender": "Male", "institution": "Emmi AI", "country": "AT", "position": "Principal Researcher" }, { "name": "Sebastian Sanokowski", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "PhD student" }, { "name": "Sepp Hochreiter", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "Full Professor" } ], "n_formula": 33, "n_formula_1": 10, "n_ref_uni": 77, "n_ref": 129, "n_ref_all": 182, "n_fig": 13, "n_tab": 2, "L_tab": 551, "n_element_tab": 91, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 2767, "n_element_tab_1": 33, "formula_len_all": 1954, "formula_len_all_1": 610, "len_all": 231156, "len_all_1": 68588, "len_abs": 1021, "len_title": 81, "len_sents": 67943, "len_sents_1": 31311, "n_sents": 596, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 33, "L_abs": 1032, "title": "Geometry-Informed Neural Networks", "abs": "Geometry is a ubiquitous tool in computer graphics, design, and engineering. However, the lack of large shape datasets limits the application of state-of-the-art supervised learning methods and motivates the exploration of alternative learning strategies. To this end, we introduce geometry-informed neural networks (GINNs) - a framework for training shape-generative neural fields *without data* by leveraging user-specified design requirements in the form of objectives and constraints. By adding *diversity* as an explicit constraint, GINNs avoid mode-collapse and can generate multiple diverse solutions, often required in geometry tasks. Experimentally, we apply GINNs to several introductory problems and a realistic 3D engineering design problem, showing control over geometrical and topological properties, such as surface smoothness or the number of holes. These results demonstrate the potential of training shape-generative models without data, paving the way for new generative design approaches without large datasets.", "keywords": [ "geometry", "implicit neural representation", "neural fields", "theory-informed learning", "geometric deep learning", "physics-informed neural networks", "generative design" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "zpLcZ2AyDK", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Chen Si", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Jiale Fu", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "Jiaming Fan", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "Undergrad student" }, { "name": "SIMENG HAN", "gender": "Female", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Xu Yang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Yaqing Wang", "gender": "Female", "institution": "Beijing Institute of Mathematical Sciences and Applications", "country": "CN", "position": "Associate Professor" } ], "n_formula": 37, "n_formula_1": 19, "n_ref_uni": 55, "n_ref": 209, "n_ref_all": 242, "n_fig": 9, "n_tab": 8, "L_tab": 3965, "n_element_tab": 536, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2551, "n_element_tab_1": 454, "formula_len_all": 1758, "formula_len_all_1": 1092, "len_all": 191280, "len_all_1": 77962, "len_abs": 1579, "len_title": 130, "len_sents": 52822, "len_sents_1": 33245, "n_sents": 416, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1585, "title": "GraphIC: A Graph-Based In-Context Example Retrieval Model for Multi-Step Reasoning", "abs": "In-context learning (ICL) enables large language models (LLMs) to generalize to new tasks by incorporating a few in-context examples (ICEs) directly in the input, without updating parameters. However, the effectiveness of ICL heavily relies on the selection of ICEs, and conventional text-based embedding methods are often inadequate for tasks that require multi-step reasoning, such as mathematical and logical problem solving. This is due to the bias introduced by shallow semantic similarities that fail to capture the deeper reasoning structures required for these tasks. We present GraphIC, a novel approach that leverages graph-based representations of reasoning processes, coupled with Bayesian Networks (BNs) to select ICEs. Graph structures inherently filter out shallow semantics while preserving the core reasoning structure. Importantly, BNs capture the dependency of a node\u2019s attributes on its parent nodes, closely mirroring the hierarchical nature of human cognition\u2014where each thought is shaped by preceding ones. This makes BNs particularly well-suited for multi-step reasoning tasks, aligning the process more closely with human-like reasoning. Extensive experiments across three types of reasoning tasks (mathematical reasoning, code generation, and logical reasoning) demonstrate that GraphIC outperforms both training-free and training-based models in selecting ICEs, excelling in terms of both effectiveness and efficiency. We show that GraphIC enhances ICL\u2019s performance and interpretability, significantly advancing ICE selection for multi-step reasoning tasks.", "keywords": [ "In-context learning", "multi-step reasoning", "thought graphs", "large language model" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "zpENPcQSj1", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Bing Liu", "gender": "Male", "institution": "University of Illinois at Chicago", "country": "US", "position": "Full Professor" }, { "name": "Changnan Xiao", "gender": "Male", "institution": "MiHoYo", "country": "", "position": "Researcher" } ], "n_formula": 32, "n_formula_1": 13, "n_ref_uni": 49, "n_ref": 132, "n_ref_all": 162, "n_fig": 2, "n_tab": 8, "L_tab": 733, "n_element_tab": 34, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 2148, "n_element_tab_1": 57, "formula_len_all": 4009, "formula_len_all_1": 1942, "len_all": 216383, "len_all_1": 86972, "len_abs": 1702, "len_title": 97, "len_sents": 56787, "len_sents_1": 35957, "n_sents": 581, "n_sents_1": 325, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 832, "title": "Generalizing Reasoning Problems to Longer Lengths", "abs": "Length generalization (LG) is a challenging problem in learning to reason. It refers to the phenomenon that when trained on reasoning problems of smaller lengths/sizes, the model struggles with problems of larger sizes or lengths. Although it has been proven that reasoning can be learned if the intermediate reasoning steps (also known as chain-of-thought (CoT)) are given in the training data, existing studies only apply to within a given length (interpolation), while LG is about extrapolation beyond the given length. This paper begins by presenting a theorem that identifies the root cause of the LG problem. It then defines a class of reasoning problems for which achieving LG with Transformers can be theoretically guaranteed, provided the CoT schemes are constructed to meet a proposed condition called $(n,r)$-consistency.", "keywords": [ "length generalization", "learning to reason", "length extrapolation" ], "rating_list": [ 8, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "zpDGwcmMV4", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Tian Ye", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Yuanzhi Li", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Zeyuan Allen-Zhu", "gender": "Unspecified", "institution": "Meta FAIR Labs", "country": "US", "position": "Researcher" }, { "name": "Zicheng Xu", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 2, "n_ref": 16, "n_ref_all": 79, "n_fig": 12, "n_tab": 4, "L_tab": 25, "n_element_tab": 1, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 306, "formula_len_all_1": 0, "len_all": 191233, "len_all_1": 76099, "len_abs": 1186, "len_title": 122, "len_sents": 85963, "len_sents_1": 39633, "n_sents": 544, "n_sents_1": 282, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1189, "title": "Physics of Language Models: Part 2.2, How to Learn From Mistakes on Grade-School Math Problems", "abs": "Language models have demonstrated remarkable performance in solving reasoning tasks; however, even the strongest models still occasionally make reasoning mistakes. Recently, there has been active research aimed at improving reasoning accuracy, particularly by using pretrained language models to \"self-correct'' their mistakes via multi-round prompting. In this paper, we follow this line of work but focus on understanding the usefulness of incorporating ``error-correction'' data directly into the pretraining stage. This data consists of erroneous solution steps immediately followed by their corrections. Using a synthetic math dataset, we show promising results: this type of pretrain data can help language models achieve higher reasoning accuracy directly (i.e., through simple auto-regression, without multi-round prompting) compared to pretraining on the same amount of error-free data. We also delve into many details, such as (1) how this approach differs from beam search, (2) how such data can be prepared, (3) whether masking is needed on the erroneous tokens, (4) the amount of error required, (5) whether such data can be deferred to the fine-tuning stage, and many others.", "keywords": [ "pretraining", "language model", "error correction", "error detection" ], "rating_list": [ 6, 5, 8, 8 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "zpBamnxyPm", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Adam Ibrahim", "gender": "unknown", "institution": "H", "country": "FR", "position": "Researcher" }, { "name": "Brando Miranda", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Gabriel Mukobi", "gender": "Male", "institution": "UC Berkeley, University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Hailey Schoelkopf", "gender": "Female", "institution": "EleutherAI", "country": "US", "position": "Researcher" }, { "name": "Herbie Bradley", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Rylan Schaeffer", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" }, { "name": "Stella Biderman", "gender": "Female", "institution": "EleutherAI", "country": "", "position": "Researcher" }, { "name": "Varun Madan", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 66, "n_ref": 116, "n_ref_all": 124, "n_fig": 3, "n_tab": 2, "L_tab": 47345, "n_element_tab": 1264, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 131, "formula_len_all_1": 364, "len_all": 281113, "len_all_1": 40951, "len_abs": 1706, "len_title": 141, "len_sents": 38144, "len_sents_1": 20172, "n_sents": 183, "n_sents_1": 108, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1724, "title": "Why Has Predicting Downstream Capabilities of Frontier AI Models with Scale Remained Elusive?", "abs": "Predictable behavior from scaling advanced AI systems is an extremely desirable property for engineers, companies, economists and governments alike, and while a well-established literature exists on how pretraining performance scales, predictable scaling behavior on downstream capabilities remains elusive. While many factors are certainly responsible, this paper shines a light on a significant factor that makes predicting scaling behavior on widely used multiple-choice question answering benchmarks challenging and illuminates a path towards making such downstream evaluations predictable with scale. Using five model families and twelve well-established multiple-choice benchmarks, we show that downstream performance is computed from negative log likelihoods via a sequence of transformations that progressively degrades the statistical relationship between performance and scale. We then reveal the mechanism causing this degradation: downstream metrics require comparing the correct choice against a small number of specific incorrect choices, meaning accurately predicting downstream capabilities requires predicting not just how probability mass concentrates on the correct choice with scale, but also how probability mass fluctuates on specific incorrect choices with scale. We empirically study how probability mass on the correct choice co-varies with probability mass on incorrect choices with increasing compute, suggesting that scaling laws for \\textit{incorrect} choices might be achievable. Our work also explains why pretraining scaling laws are commonly regarded as more predictable than downstream capabilities and contributes towards establishing scaling-predictable evaluations of frontier AI models.", "keywords": [ "evaluations", "benchmarks", "scaling laws", "emergent abilities", "capabilities", "frontier models", "foundation models" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "zp88xOXAfS", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Danilo Giordano", "gender": "Male", "institution": "Polytechnic Institute of Turin", "country": "IT", "position": "Assistant Professor" }, { "name": "Francesco De Santis", "gender": "Male", "institution": "Polytechnic Institute of Turin", "country": "IT", "position": "PhD student" }, { "name": "Gabriele Ciravegna", "gender": "Male", "institution": "Polytechnic Institute of Turin", "country": "IT", "position": "Postdoc" }, { "name": "Philippe Bich", "gender": "Male", "institution": "Polytechnic Institute of Turin", "country": "IT", "position": "PhD student" }, { "name": "Pietro Barbiero", "gender": "Male", "institution": "International Business Machines", "country": "CH", "position": "Postdoc" }, { "name": "Tania Cerquitelli", "gender": "Female", "institution": "Polytechnic Institute of Turin", "country": "IT", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 5, "n_ref_uni": 48, "n_ref": 106, "n_ref_all": 145, "n_fig": 10, "n_tab": 4, "L_tab": 2163, "n_element_tab": 110, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 840, "n_element_tab_1": 67, "formula_len_all": 580, "formula_len_all_1": 246, "len_all": 183357, "len_all_1": 68871, "len_abs": 1336, "len_title": 118, "len_sents": 55519, "len_sents_1": 32831, "n_sents": 436, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1342, "title": "Linearly Interpretable Concept Embedding Model for Text Classification", "abs": "Despite their success, Large-Language Models (LLMs) still face criticism due to their lack of interpretability.\nTraditional post-hoc interpretation methods, based on attention and gradient-based analysis, offer limited insight as they only approximate the model's decision-making processes and have been proved to be unreliable.\nFor this reason, Concept-Bottleneck Models (CBMs) have been lately proposed in the textual field to provide interpretable predictions based on human-understandable concepts. \nHowever, CBMs still face several criticisms for their architectural constraints limiting their expressivity, for the absence of task-interpretability when employing non-linear task predictors and for requiring extensive annotations that are impractical for real-world text data. In this paper we address these challenges by proposing a novel Linearly Interpretable Concept Embedding Model (LICEM) going beyond the current accuracy-interpretability trade-off. LICEM classification accuracy is better than existing interpretable models and matches black-box models. The provided explanations are more plausible and useful with respect to existing solutions, as attested in a user study. Finally, we show our model can be trained without requiring any concept supervision, as concepts can be automatically predicted by the same LLM backbone.", "keywords": [ "CBM", "XAI", "Interpretable AI" ], "rating_list": [ 3, 6, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 2, 4, 4, 4, 5 ] }, { "paper_id": "zo049dh2r9", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Hongwei Sheng", "gender": "Male", "institution": "University of Queensland", "country": "AU", "position": "PhD student" }, { "name": "Jiaying Ying", "gender": "Female", "institution": "University of Queensland", "country": "AU", "position": "PhD student" }, { "name": "Kun Zhan", "gender": "Male", "institution": "LiAuto", "country": "CN", "position": "Engneering" }, { "name": "Shuyun Wang", "gender": "Male", "institution": "University of Queensland", "country": "AU", "position": "PhD student" }, { "name": "Tianqing Zhu", "gender": "Female", "institution": "City University of Macau", "country": "MO", "position": "Full Professor" }, { "name": "Xiaobiao Du", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Xin Yu", "gender": "Male", "institution": "Google", "country": "AU", "position": "Researcher" }, { "name": "Yida Wang", "gender": "Male", "institution": "Li Auto Inc.", "country": "CN", "position": "Senior Research Engineer" }, { "name": "Zhuojie Wu", "gender": "unknown", "institution": "University of Queensland", "country": "AU", "position": "PhD student" }, { "name": "haiyang sun", "gender": "Male", "institution": "LiAuto", "country": "CN", "position": "Engineer" }, { "name": "Ming Lu", "gender": "Male", "institution": "Intel Labs China", "country": "CN", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 57, "n_ref": 144, "n_ref_all": 172, "n_fig": 13, "n_tab": 4, "L_tab": 964, "n_element_tab": 121, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1755, "n_element_tab_1": 139, "formula_len_all": 106, "formula_len_all_1": 106, "len_all": 173528, "len_all_1": 58711, "len_abs": 807, "len_title": 0, "len_sents": 38293, "len_sents_1": 26133, "n_sents": 323, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1783, "title": "3DRealCar: An In-the-wild RGB-D Car Dataset with 360-degree Views", "abs": "3D cars are commonly used in self-driving systems, virtual/augmented reality, and games. However, existing 3D car datasets are either synthetic or low-quality, presenting a significant gap toward the high-quality real-world 3D car datasets and limiting their applications in practical scenarios. In this paper, we propose the first large-scale 3D real car dataset, termed 3DRealCar, offering three distinctive features. (1) \\textbf{High-Volume}: 2,500 cars are meticulously scanned by 3D scanners, obtaining car images and point clouds with real-world dimensions; (2) \\textbf{High-Quality}: Each car is captured in an average of 200 dense, high-resolution 360-degree RGB-D views, enabling high-fidelity 3D reconstruction; (3) \\textbf{High-Diversity}: The dataset contains various cars from over 100 brands, collected under three distinct lighting conditions, including reflective, standard, and dark. Additionally, we offer detailed car parsing maps for each instance to promote research in car parsing tasks. Moreover, we remove background point clouds and standardize the car orientation to a unified axis for the reconstruction only on cars without background and controllable rendering. We benchmark 3D reconstruction results with state-of-the-art methods across each lighting condition in 3DRealCar. Extensive experiments demonstrate that the standard lighting condition part of 3DRealCar can be used to produce a large number of high-quality 3D cars, improving various 2D and 3D tasks related to cars. Notably, our dataset brings insight into the fact that recent 3D reconstruction methods face challenges in reconstructing high-quality 3D cars under reflective and dark lighting conditions. \n\\textcolor{red}{\\href{https://3drealcar.github.io/}{Our dataset is available here.}}", "keywords": [ "3D reconstruction", "Car reconstruction", "Car dataset" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 3, 2, 4 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "zno7tZVG8T", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Fei Chao", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Jing Lin", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Shenao", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Wenting Lin", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Xiawu Zheng", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Yiwu Yao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yuexiao Ma", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Yuhang Wu", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Zhaohao Xie", "gender": "Male", "institution": "Xiamen University", "country": "", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 26, "n_ref": 80, "n_ref_all": 87, "n_fig": 4, "n_tab": 5, "L_tab": 6824, "n_element_tab": 1156, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1635, "n_element_tab_1": 147, "formula_len_all": 700, "formula_len_all_1": 680, "len_all": 120215, "len_all_1": 53054, "len_abs": 1741, "len_title": 129, "len_sents": 30085, "len_sents_1": 25146, "n_sents": 201, "n_sents_1": 153, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1750, "title": "Extreme composite compression of large language models through joint optimization", "abs": "Post-Training Quantization (PTQ) and Sparsification (PTS) are dominant methods in the compression of Large Language Models (LLMs) due to their minimal resource usage and generalizability. It is a natural idea to integrate quantization and sparsification in a unified framework, which however, often results in substantial accuracy losses. Here we argue that, the key lies in optimization. This paper introduces a novel joint optimization strategy that concurrently mitigates errors induced by both sparsification and quantization. \nUnlike sequential approaches, our method employs learnable transformation matrices to simultaneously optimize errors across both dimensions, preventing the typical misalignments associated with sequential optimizations. Furthermore, we present a reordering mechanism within the learnable mask sparsification process to maintain consistent sparsity ratios. This mechanism ensures the prioritization of the least important weights during each update iteration, thus enhancing the stability of the compression process. \nOur approach demonstrates considerable performance enhancements across diverse models and datasets, with the most notable gains observed under conditions of extremely low-bit quantization and high sparsity ratios. For example, in the LLaMA2-13b model with weight quantization at 2 bit and a 75% sparsity configuration, our method surpasses the state-of-the-art (SOTA) by 9.03% in average accuracy across five zero-shot tasks. Meanwhile, in the newest LLaMA3-8b model, with weight quantization at 3 bit and a 50% sparsity configuration, our method outperforms the SOTA by 4.58% (56.86% vs 52.28%) in zero-shot tasks and achieves a perplexity reduction of 4.45 on the WikiText2 dataset (10.78 vs 15.23).", "keywords": [ "model quantization", "model compression", "sparsification", "joint optimization" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "znhZbonEoe", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Jie Tan", "gender": "Female", "institution": "University of Groningen", "country": "NL", "position": "PhD student" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Qinglun Li", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Xiaochun Cao", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Yifan Shi", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yingqi Liu", "gender": "Female", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 95, "n_formula_1": 34, "n_ref_uni": 43, "n_ref": 83, "n_ref_all": 114, "n_fig": 5, "n_tab": 4, "L_tab": 1841, "n_element_tab": 192, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15912, "formula_len_all_1": 3091, "len_all": 202910, "len_all_1": 75042, "len_abs": 1449, "len_title": 132, "len_sents": 57399, "len_sents_1": 34223, "n_sents": 463, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1442, "title": "Understanding the Stability-based Generalization of Personalized Federated Learning", "abs": "Despite great achievements in algorithm design for Personalized Federated Learning (PFL), research on the theoretical analysis of generalization is still in its early stages. Some recent theoretical results have investigated the generalization performance of personalized models under the problem setting and hypothesis in the convex condition, which do not consider the real iteration performance during the non-convex training. To further understand the testing performance from the theoretical perspective, we propose the first algorithm-matter generalization analysis with uniform stability for the typical PFL method Partial Model Personalization on smooth and non-convex objectives. In an attempt to distinguish the shared and personalized errors, we decouple the shared aggregation and the local fine-tuning progress and illustrate the interaction mechanism between the shared and personalized variables. The algorithm-matter generalization bounds analyze the impact of the trivial hyperparameters like learning steps and stepsizes as well as the communication modes in both Centralized and Decentralized PFL (C-PFL and D-PFL), which also concludes that C-PFL generalizes better than D-PFL. Combined with the convergence errors, we then obtain the excess risk analysis and establish the better early stopping point for the optimal population risk of PFL. Promising experiments on CIFAR dataset also corroborate our theoretical results.", "keywords": [ "stability analysis+generalization gap+excess risk+personalized federated learning" ], "rating_list": [ 8, 6, 3, 6 ], "soundness_list": [ 4, 3, 2, 2 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "znL549Ymoi", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Gyorgy Buzsaki", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Wannan Yang", "gender": "Female", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 20, "n_ref": 57, "n_ref_all": 83, "n_fig": 18, "n_tab": 3, "L_tab": 446, "n_element_tab": 24, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 1740, "n_element_tab_1": 21, "formula_len_all": 345, "formula_len_all_1": 836, "len_all": 172400, "len_all_1": 51419, "len_abs": 1670, "len_title": 98, "len_sents": 33842, "len_sents_1": 21190, "n_sents": 302, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1696, "title": "Interpretability of LLM Deception: Universal Motif", "abs": "Conversational large language models (LLMs) are trained to be helpful, honest and harmless (HHH) and yet they remain susceptible to hallucinations, misinformation and are capable of deception. A promising avenue for safeguarding against these behaviors is to gain a deeper understanding of their inner workings. Here we ask: what could interpretability tell us about deception and can it help to control it? First, we introduce a simple and yet general protocol to induce 20 large conversational models from different model families (Llama, Gemma, Yi and Qwen) of various sizes (from 1.5B to 70B) to knowingly lie. Second, we characterize three iterative refinement stages of deception from the latent space representation. Third, we demonstrate that these stages are \\textit{universal} across models from different families and sizes. We find that the third stage progression reliably predicts whether a certain model is capable of deception. Furthermore, our patching results reveal that a surprisingly sparse set of layers and attention heads are causally responsible for lying. Importantly, consistent across all models tested, this sparse set of layers and attention heads are part of the third iterative refinement process. When contrastive activation steering is applied to control model output, only steering these layers from the third stage could effectively reduce lying. Overall, these findings identify a universal motif across deceptive models and provide actionable insights for developing general and robust safeguards against deceptive AI. The code, dataset, visualizations, and an interactive demo notebook are available at \\url{https://github.com/safellm-2024/llm_deception}.", "keywords": [ "safety", "honesty", "deception", "lie", "interpretability", "Large Language Model" ], "rating_list": [ 6, 1, 5, 10 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 2, 1, 1, 4 ], "contribution_list": [ 3, 1, 3, 4 ], "confidence_list": [ 3, 5, 4, 5 ] }, { "paper_id": "znGnmAM44K", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Gabriel Kreiman", "gender": "Male", "institution": "Harvard Medical School", "country": "US", "position": "Full Professor" }, { "name": "Mengmi Zhang", "gender": "Female", "institution": "Nanyang Technological University, Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Mingzhong Sun", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 41, "n_ref": 61, "n_ref_all": 95, "n_fig": 21, "n_tab": 6, "L_tab": 1302, "n_element_tab": 89, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 180, "formula_len_all_1": 31, "len_all": 143482, "len_all_1": 64978, "len_abs": 1466, "len_title": 131, "len_sents": 48275, "len_sents_1": 31837, "n_sents": 370, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1472, "title": "The other you in black mirror: first steps from chatbots to personalized LLM clones", "abs": "Large language models (LLMs) have demonstrated remarkable abilities in a wide\nvariety of generic tasks. Here we investigate whether it is possible to use LLMs\nto partially replicate cognitive aspects of an individual by fine-tuning an LLM\nwith personal data. Our model, A-clone, built on the pretrained Llama3-70B, was\nfine-tuned with a private English dataset from one volunteer referred to as A throughout. We\nevaluated A-clone in two ways. First, using 701 open-ended questions, we gathered\nresponses from A, A-clone, other LLMs, and A\u2019s family members imitating A.\nWe conducted a Turing-like test where 31 participants with varying degrees of\nfamiliarity with A attempted to identify A\u2019s real answers in a question-and-answer\ntask. Human participants identified the genuine responses from A 55% \u00b1 7%\nof the time, just over chance levels. A-clone outperformed all other baselines\nin mimicking adequate responses from A. Second, we compared the outputs\nof A-Clone with the ground truth from A in 10 psychological, moral, career,\npolitical tendency, and general knowledge tests, containing 484 questions altogether.\nA-Clone demonstrated a strong correlation with A\u2019s responses. This work provides\nan initial, proof-of-principle, evaluation of the possibility of mimicking the\nresponses of an individual, opening doors to many real-world applications but\nalso raising potential privacy and safety concerns about digital clones. The code\nand data can be found in this link.", "keywords": [ "Large Language Models (LLMs)", "Personalized AI", "Turing Test", "AI Safety" ], "rating_list": [ 6, 6, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "zn0eqMtsrw", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Mathis Gerdes", "gender": "unknown", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Max Welling", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "Miranda C. N. Cheng", "gender": "unknown", "institution": "Academia Sinica", "country": "TW", "position": "Researcher" } ], "n_formula": 43, "n_formula_1": 30, "n_ref_uni": 27, "n_ref": 45, "n_ref_all": 64, "n_fig": 8, "n_tab": 1, "L_tab": 297, "n_element_tab": 22, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 297, "n_element_tab_1": 22, "formula_len_all": 2308, "formula_len_all_1": 1529, "len_all": 127752, "len_all_1": 67878, "len_abs": 1239, "len_title": 86, "len_sents": 47302, "len_sents_1": 31833, "n_sents": 346, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 38, "L_abs": 1236, "title": "GUD: Generation with Unified Diffusion", "abs": "Diffusion generative models transform noise into data by inverting a process that progressively adds noise to data samples. Inspired by concepts from the renormalization group in physics, which analyzes systems across different scales, we revisit diffusion models by exploring three key design aspects: 1) the choice of representation in which the diffusion process operates (e.g. pixel-, PCA-, Fourier-, or wavelet-basis), 2) the prior distribution that data is transformed into during diffusion (e.g. Gaussian with covariance $\\Sigma$), and 3) the scheduling of noise levels applied separately to different parts of the data, captured by a component-wise noise schedule. \n Incorporating the flexibility in these choices, we develop a unified framework for diffusion generative models with greatly enhanced design freedom. In particular, we introduce soft-conditioning models that smoothly interpolate between standard diffusion models and autoregressive models (in any basis), conceptually bridging these two approaches. \nOur framework opens up a wide design space which may lead to more efficient training and data generation, and paves the way to novel architectures integrating different generative approaches and generation tasks.", "keywords": [ "diffusion models", "renormalization group", "autoregressive models", "wavelet decomposition", "denoising score matching" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "zmmfsJpYcq", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Fandi Wu", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Jianhua Yao", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Principal Researcher" }, { "name": "Jiaxiang Wu", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Peilin Zhao", "gender": "unknown", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Rubo Wang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xingyu Gao", "gender": "unknown", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 23, "n_formula_1": 12, "n_ref_uni": 81, "n_ref": 148, "n_ref_all": 167, "n_fig": 6, "n_tab": 5, "L_tab": 1959, "n_element_tab": 282, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 427, "n_element_tab_1": 40, "formula_len_all": 1844, "formula_len_all_1": 1067, "len_all": 232957, "len_all_1": 61777, "len_abs": 1246, "len_title": 116, "len_sents": 59094, "len_sents_1": 30952, "n_sents": 414, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1314, "title": "IgGM: A Generative Model for Functional Antibody and Nanobody Design", "abs": "Immunoglobulins are crucial proteins produced by the immune system to identify and bind to foreign substances, playing an essential role in shielding organisms from infections and diseases. Designing specific antibodies opens new pathways for disease treatment. With the rise of deep learning, AI-driven drug design has become possible, leading to several methods for antibody design. However, many of these approaches require additional conditions that differ from real-world scenarios, making it challenging to incorporate them into existing antibody design processes. Here, we introduce IgGM, a generative model for the de novo design of immunoglobulins with functional specificity. IgGM simultaneously generates antibody sequences and structures for a given antigen, consisting of three core components: a pre-trained language model for extracting sequence features, a feature learning module for identifying pertinent features, and a prediction module that outputs designed antibody sequences and the predicted complete antibody-antigen complex structure. IgGM effectively predicts structures and designs novel antibodies and nanobodies. This makes it highly applicable in a wide range of practical situations related to antibody and nanobody design. Code is available at: https://github.com/TencentAI4S/IgGM.", "keywords": [ "de novo antibody design", "complex structure prediction", "protein design" ], "rating_list": [ 6, 5, 5, 8, 5 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 4, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "zmHqlXGTTl", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Chuan Qin", "gender": "Male", "institution": "CNIC, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Hengshu Zhu", "gender": "unknown", "institution": "Computer Network Information Center, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Kaichun Yao", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Lan Zeng", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "MS student" }, { "name": "Libo Zhang", "gender": "Male", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Yanjun Wu", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 25, "n_ref": 48, "n_ref_all": 69, "n_fig": 3, "n_tab": 9, "L_tab": 2482, "n_element_tab": 250, "n_fig_1": 1, "n_tab_1": 8, "L_tab_1": 2479, "n_element_tab_1": 257, "formula_len_all": 934, "formula_len_all_1": 787, "len_all": 112400, "len_all_1": 61111, "len_abs": 1598, "len_title": 129, "len_sents": 34618, "len_sents_1": 25318, "n_sents": 292, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1605, "title": "SciPG: A New Benchmark and Approach for Layout-aware Scientific Poster Generation", "abs": "Scientific posters are an effective and expressive medium for conveying the core ideas of academic papers, facilitating the communication of research techniques. However, creating high-quality scientific posters is a complex and time-consuming task that requires advanced skills to summarize key concepts and arrange them logically and visually appealingly. Previous studies have primarily focused on either content extraction or the layout and composition of posters, often relying on small-scale datasets. The scarcity of large, publicly available datasets has further limited advancements in this field.\nIn this paper, we introduce a new task called layout-aware scientific poster generation (LayoutSciPG), which aims to generate flexible posters from scientific papers through integrated automatic content extraction and layout design.\nTo achieve this, we first build a large-scale dataset containing over 10,000 pairs of scientific papers and their corresponding posters. We then propose a multimodal extractor-generator framework, which employs a multimodal extractor to retrieve key text and image elements from the papers and designs an interactive generator with an adaptive memory mechanism to seamlessly paraphrase the extracted content and generate a structured layout. This approach effectively tackles challenges related to GPU memory consumption and long-term dependencies when handling the lengthy inputs (scientific papers) and outputs (posters). Finally, both qualitative and quantitative evaluations demonstrate the effectiveness of our approach while highlighting remaining challenges.", "keywords": [ "Scientific poster generation", "multimodal extraction", "multimodal generation" ], "rating_list": [ 3, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "zlAUnwhE2v", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Can Wang", "gender": "unknown", "institution": "Griffith University", "country": "", "position": "Senior Lecturer" }, { "name": "Huan Yee Koh", "gender": "Male", "institution": "Monash University", "country": "", "position": "PhD student" }, { "name": "Jiaxin Ju", "gender": "Female", "institution": "Griffith University", "country": "AU", "position": "PhD student" }, { "name": "Shirui Pan", "gender": "unknown", "institution": "Griffith University", "country": "AU", "position": "Full Professor" }, { "name": "YIZHEN ZHENG", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 42, "n_ref": 84, "n_ref_all": 97, "n_fig": 4, "n_tab": 3, "L_tab": 1334, "n_element_tab": 92, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1377, "n_element_tab_1": 96, "formula_len_all": 357, "formula_len_all_1": 357, "len_all": 124609, "len_all_1": 59284, "len_abs": 1248, "len_title": 134, "len_sents": 36540, "len_sents_1": 29140, "n_sents": 226, "n_sents_1": 170, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1250, "title": "ChemThinker: Thinking Like a Chemist with Multi-Agent LLMs for Deep Molecular Insights", "abs": "Molecular property prediction is vital in drug discovery and cheminformatics, yet many current models lack interpretability, making it difficult for experts to understand the rationale behind predictions. To address this, we introduce ChemThinker, a novel large language models (LLMs) multi-agent framework designed to effectively control the internal representations of concepts and functions within LLMs. ChemThinker emulates the way chemists approach molecular analysis by integrating insights from three perspectives: general molecular properties, data-driven analysis, and task-specific factors. Each perspective uses an agentic approach to stimulate the LLM's internal representations, enabling more targeted and interpretable outputs based on the problem at hand, akin to how stimuli trigger the brain's cognitive processes. By feeding representations from these three perspectives into a simple multi-layer perceptron (MLP), ChemThinker achieves superior performance, significantly outperforming existing baselines across multiple benchmarks. Furthermore, our framework provides interpretable insights into the molecular mechanisms driving the predictions, making it a practical tool for drug discovery and other cheminformatics applications.", "keywords": [ "Molecular Property Prediction", "Molecular Representation Learning", "Multi-Agent LLMs" ], "rating_list": [ 3, 1, 3, 5 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "zl3pfz4VCV", "primary_area": "datasets and benchmarks", "n_author": 83, "author_info_list": [ { "name": "Akash Kundu", "gender": "Male", "institution": "Heritage Institute of Technology", "country": "IN", "position": "Undergrad student" }, { "name": "Akshita Sukhlecha", "gender": "Female", "institution": "Weav.AI", "country": "US", "position": "Researcher" }, { "name": "Aleksandr Abramov", "gender": "Male", "institution": "SaluteDevices", "country": "", "position": "Principal Researcher" }, { "name": "Aleksei Vatolin", "gender": "unknown", "institution": "Federal Research Center \u00abComputer Science and Control\u00bb of Russian Academy of Sciences", "country": "RU", "position": "PhD student" }, { "name": "Alena Fenogenova", "gender": "Female", "institution": "SaluteDevices", "country": "RU", "position": "Researcher" }, { "name": "Alessia Borghini", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Andrianos Michail", "gender": "Male", "institution": "University of Zurich", "country": "", "position": "PhD student" }, { "name": "Artem Snegirev", "gender": "Male", "institution": "SaluteDevices", "country": "", "position": "Researcher" }, { "name": "Ashwin Mathur", "gender": "Male", "institution": "Open Source", "country": "IN", "position": "Researcher" }, { "name": "Bhavish Pahwa", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Researcher" }, { "name": "Bj\u00f6rn Pl\u00fcster", "gender": "Male", "institution": "Universit\u00e4t Hamburg", "country": "", "position": "MS student" }, { "name": "Chenghao Xiao", "gender": "Male", "institution": "Durham University", "country": "GB", "position": "PhD student" }, { "name": "Daniel Auras", "gender": "Male", "institution": "ellamind GmbH", "country": "", "position": "Researcher" }, { "name": "David Stap", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "PhD student" }, { "name": "Dawei Zhu", "gender": "unknown", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Diganta Misra", "gender": "Male", "institution": "Max-Planck-Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Dipam Vasani", "gender": "Male", "institution": "DocuSign", "country": "", "position": "Researcher" }, { "name": "Dominik Krzemi\u0144ski", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "Federico Cassano", "gender": "unknown", "institution": "Cursor AI", "country": "US", "position": "Research Scientist" }, { "name": "Gayatri K", "gender": "Female", "institution": "R V College of Engineering", "country": "", "position": "Undergrad student" }, { "name": "Genta Indra Winata", "gender": "Male", "institution": "Capital One", "country": "US", "position": "Principal Researcher" }, { "name": "Guangyu Song", "gender": "Not Specified", "institution": "Tano Labs", "country": "CA", "position": "Researcher" }, { "name": "Henil Shalin Panchal", "gender": "Male", "institution": "Institute of Technology,Nirma University", "country": "", "position": "Undergrad student" }, { "name": "Hippolyte Gisserot-Boukhlef", "gender": "Male", "institution": "CentraleSupelec", "country": "", "position": "PhD student" }, { "name": "Imene Kerboua", "gender": "Female", "institution": "INSA Lyon, LIRIS", "country": "FR", "position": "PhD student" }, { "name": "Isabelle Mohr", "gender": "Female", "institution": "Jina AI", "country": "DE", "position": "Researcher" }, { "name": "Jan Kostkan", "gender": "unknown", "institution": "Aarhus University", "country": "", "position": "PhD student" }, { "name": "Jan Philipp Harries", "gender": "Male", "institution": "University of Wuppertal", "country": "", "position": "Emeritus" }, { "name": "Jay Gala", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "John Yang", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Jonathan Rystr\u00f8m", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Jordan Clive", "gender": "Male", "institution": "Chattermill", "country": "GB", "position": "Principal Researcher" }, { "name": "Kenneth Enevoldsen", "gender": "Male", "institution": "Aarhus University", "country": "DK", "position": "Postdoc" }, { "name": "Konrad Wojtasik", "gender": "Male", "institution": "Technical University of Wroclaw", "country": "PL", "position": "PhD student" }, { "name": "Kranthi Kiran GV", "gender": "Male", "institution": "DataChat", "country": "", "position": "Researcher" }, { "name": "Lasse Hansen", "gender": "Male", "institution": "Aarhus University", "country": "DK", "position": "PhD student" }, { "name": "Lester James Validad Miranda", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Lo\u00efc Magne", "gender": "Male", "institution": "Ecole Normale Superieure", "country": "", "position": "MS student" }, { "name": "Maksimova Anna", "gender": "Female", "institution": "Higher School of Economics, Higher School of Economics", "country": "", "position": "Undergrad student" }, { "name": "Malte Ostendorff", "gender": "unknown", "institution": "Deutsche Telekom", "country": "DE", "position": "Researcher" }, { "name": "Manan Dey", "gender": "Male", "institution": "SalesForce.com", "country": "", "position": "Senior Member of Technical Staff " }, { "name": "Manuel Faysse", "gender": "Male", "institution": "CentraleSupelec", "country": "", "position": "PhD student" }, { "name": "Marek Suppa", "gender": "unknown", "institution": "Comenius University in Bratislava", "country": "SK", "position": "Lecturer" }, { "name": "Maria Tikhonova", "gender": "Female", "institution": "Higher School of Economics", "country": "RU", "position": "Lecturer" }, { "name": "Marion Schaeffer", "gender": "unknown", "institution": "Institut National des Sciences Appliqu\u00e9es de Rouen", "country": "FR", "position": "PhD student" }, { "name": "Mariya Hendriksen", "gender": "Female", "institution": "University of Amsterdam", "country": "", "position": "PhD student" }, { "name": "Martin Bernstorff", "gender": "Male", "institution": "Aarhus University", "country": "", "position": "PhD student" }, { "name": "Mathieu Ciancone", "gender": "Male", "institution": "Wikit", "country": "FR", "position": "PhD student" }, { "name": "Mengzhou Xia", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Michael G\u00fcnther", "gender": "Male", "institution": "Jina AI", "country": "DE", "position": "Researcher" }, { "name": "Mohammed Hamdy", "gender": "Male", "institution": "Independent", "country": "", "position": "Researcher" }, { "name": "M\u00e1rton Kardos", "gender": "Male", "institution": "Aarhus University", "country": "DK", "position": "Junior Developer" }, { "name": "Nandan Thakur", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Niklas Muennighoff", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Orion Weller", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "PhD student" }, { "name": "Pranjal A Chitale", "gender": "Male", "institution": "Microsoft Research", "country": "IN", "position": "Intern" }, { "name": "Rafa\u0142 Po\u015bwiata", "gender": "unknown", "institution": "National Information Processing Institute", "country": "", "position": "Researcher" }, { "name": "Roberta Rocca", "gender": "Female", "institution": "Aarhus University", "country": "", "position": "Assistant Professor" }, { "name": "Roman Solomatin", "gender": "Male", "institution": "ITMO University", "country": "RU", "position": "MS student" }, { "name": "Ruqiya Bin Safi", "gender": "Female", "institution": "The London Institute of Banking & Finance", "country": "GB", "position": "MS student" }, { "name": "Saba Sturua", "gender": "Male", "institution": "Jina AI", "country": "", "position": "Researcher" }, { "name": "Saiteja Utpala", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "", "position": "Researcher" }, { "name": "Shawon Ashraf", "gender": "Male", "institution": "ellamind GmbH", "country": "DE", "position": "Researcher" }, { "name": "Shreeya Dhakal", "gender": "Female", "institution": "icodeformybhasa", "country": "US", "position": "Researcher" }, { "name": "Silvan Wehrli", "gender": "unknown", "institution": "Robert Koch Institute", "country": "DE", "position": "PhD student" }, { "name": "Simon Clematide", "gender": "Male", "institution": "University of Zurich", "country": "CH", "position": "Researcher" }, { "name": "Simone Tedeschi", "gender": "Male", "institution": "University of Roma \"La Sapienza\"", "country": "IT", "position": "PhD student" }, { "name": "Siva Reddy", "gender": "Male", "institution": "ServiceNow Inc", "country": "CA", "position": "Researcher" }, { "name": "Taemin Lee", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Researcher" }, { "name": "Tom Aarsen", "gender": "Male", "institution": "Hugging Face", "country": "NL", "position": "Independent Researcher" }, { "name": "Vaibhav Adlakha", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "Weijia Shi", "gender": "unknown", "institution": "University of Washington, Seattle", "country": "US", "position": "PhD student" }, { "name": "Wen-Ding Li", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Wissam Siblini", "gender": "unknown", "institution": "Komodo Health", "country": "FR", "position": "Researcher" }, { "name": "Zheng Liu", "gender": "unknown", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" }, { "name": "\u00d6mer Veysel \u00c7a\u011fatan", "gender": "Male", "institution": "Ko\u00e7 University", "country": "TR", "position": "Undergrad student" }, { "name": "Xinyu ZHANG", "gender": "Female", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Isaac Kwan Yin Chung", "gender": "Male", "institution": "Wrike", "country": "EE", "position": "Researcher" }, { "name": "Manan Dey", "gender": "Male", "institution": "KIIT University", "country": "", "position": "Undergrad student" }, { "name": "Tai Nguyen", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "MS student" }, { "name": "Sara Hooker", "gender": "unknown", "institution": "Cohere For AI", "country": "US", "position": "Principal Researcher" }, { "name": "Shi Tao Xiao", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Xing Han Lu", "gender": "Male", "institution": "McGill University", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 81, "n_ref": 132, "n_ref_all": 185, "n_fig": 12, "n_tab": 7, "L_tab": 1557, "n_element_tab": 208, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 3346, "n_element_tab_1": 231, "formula_len_all": 216, "formula_len_all_1": 0, "len_all": 465213, "len_all_1": 61814, "len_abs": 2102, "len_title": 101, "len_sents": 76393, "len_sents_1": 25663, "n_sents": 598, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1631, "title": "MMTEB: Massive Multilingual Text Embedding Benchmark", "abs": "Text embeddings are typically evaluated on a narrow set of tasks, limited in terms of languages, domains, and task types. To circumvent this limitation and to provide a more comprehensive evaluation, we introduce the Massive Multilingual Text Embedding Benchmark (MMTEB) -- a large-scale community-driven initiative expanding MTEB to over 500 \\textit{quality controlled} evaluation tasks across 1,000+ languages. MMTEB includes a wide range of challenging novel tasks such as instruction following, long-document retrieval, and code retrieval, and represents the largest multilingual collection of evaluation tasks for embedding models to date. We use this collection to construct multiple highly multilingual benchmarks. We evaluate a representative set of models on these benchmarks.\nOur findings indicate that, while LLM-based models can achieve state-of-the-art performance on a subset of languages, the best-performing publicly available model across languages is the notably smaller, multilingual-e5-large-instruct.\n\nMassive benchmarks often impose high computational demands, limiting accessibility, particularly for low-resource communities. To address this, we downsample tasks based on inter-task correlation (i.e., selecting only a diverse set of tasks) while preserving relative rankings.\nWe further optimize tasks such as retrieval by sampling hard negatives, creating smaller but effective splits. These optimizations allow us to introduce benchmarks at a significantly lower computational cost. For instance, we introduce a new zero-shot English benchmark that maintains a similar ordering at a fraction of the cost.", "keywords": [ "natural language processing", "benchmark", "sentence embeddings", "multilingual" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 4, 3, 4 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "zl3nFqY8l1", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 6, "author_info_list": [ { "name": "Chengjin Xu", "gender": "Male", "institution": "International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Dingmin Wang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Jian Guo", "gender": "Male", "institution": "International Digital Economy Academy, International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Yong Dou", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zhen Huang", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zhongwu Chen", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 53, "n_ref": 78, "n_ref_all": 127, "n_fig": 6, "n_tab": 5, "L_tab": 1146, "n_element_tab": 105, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 239, "n_element_tab_1": 42, "formula_len_all": 606, "formula_len_all_1": 437, "len_all": 243947, "len_all_1": 66514, "len_abs": 1931, "len_title": 144, "len_sents": 55883, "len_sents_1": 34187, "n_sents": 387, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1999, "title": "RuleRAG: Rule-Guided Retrieval-Augmented Generation with Language Models for Question Answering", "abs": "Retrieval-augmented generation (RAG) framework has shown promising potential in knowledge-intensive question answering (QA) by retrieving external corpus and generating based on augmented context. However, existing approaches only consider the query itself, neither specifying the retrieval preferences for the retrievers nor informing the generators of how to refer to the retrieved documents for the answers, which poses a significant challenge to the QA performance. To address these issues, we propose Rule-Guided Retrieval-Augmented Generation with LMs, which explicitly introduces symbolic rules as demonstrations for in-context learning (RuleRAG-ICL) to guide retrievers to retrieve logically related documents in the directions of rules and uniformly guide generators to generate answers attributed by the guidance of the same set of rules. Moreover, the combination of queries and rules can be further used as supervised fine-tuning data to update retrievers and generators (RuleRAG-FT) to achieve better rule-based instruction following capability, leading to retrieve more supportive results and generate more acceptable answers. To emphasize the attribution of rules, we construct five rule-aware QA benchmarks, including three temporal and two static scenarios, and equip RuleRAG with several kinds of retrievers and generators. Experiments demonstrate that training-free RuleRAG-ICL effectively improves the retrieval quality of +89.2\\% in Recall@10 scores and generation accuracy of +103.1\\% in exact match scores over standard RAG on average across the five benchmarks, and further fine-tuned RuleRAG-FT consistently yields more significant performance enhancement. Extensive analyses indicate that RuleRAG scales well with increasing numbers of retrieved documents and exhibits generalization ability for untrained rules. Our code and benchmarks are available at [https://anonymous.4open.science/r/ICLR2025_RuleRAG_ICL_FT](https://anonymous.4open.science/r/ICLR2025_RuleRAG_ICL_FT).", "keywords": [ "Rule-Guided Retrieval", "Rule-Guided Generation", "RAG", "Question Answering" ], "rating_list": [ 6, 3, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 3, 3 ] }, { "paper_id": "zl0HLZOJC9", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Gustavo Carneiro", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Full Professor" }, { "name": "Cuong Cao Nguyen", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Postdoc" }, { "name": "Thanh-Toan Do", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" } ], "n_formula": 54, "n_formula_1": 17, "n_ref_uni": 28, "n_ref": 52, "n_ref_all": 103, "n_fig": 11, "n_tab": 5, "L_tab": 1334, "n_element_tab": 159, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1453, "n_element_tab_1": 164, "formula_len_all": 5418, "formula_len_all_1": 1860, "len_all": 142484, "len_all_1": 64235, "len_abs": 1494, "len_title": 157, "len_sents": 48844, "len_sents_1": 29344, "n_sents": 354, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 106, "L_abs": 1499, "title": "Probabilistic Learning to Defer: Handling Missing Expert Annotations and Controlling Workload Distribution", "abs": "Recent progress in machine learning research is gradually shifting its focus towards *human-AI cooperation* due to the advantages of exploiting the reliability of human experts and the efficiency of AI models. One of the promising approaches in human-AI cooperation is *learning to defer* (L2D), where the system analyses the input data and decides to make its own decision or defer to human experts. Although L2D has demonstrated state-of-the-art performance, in its standard setting, L2D entails a severe limitation: all human experts must annotate the whole training dataset of interest, resulting in a time-consuming and expensive annotation process that can subsequently influence the size and diversity of the training set. Moreover, the current L2D does not have a principled way to control workload distribution among human experts and the AI classifier, which is critical to optimise resource allocation. We, therefore, propose a new probabilistic modelling approach inspired by the mixture-of-experts, where the Expectation - Maximisation algorithm is leverage to address the issue of missing expert's annotations. Furthermore, we introduce a constraint, which can be solved efficiently during the E-step, to control the workload distribution among human experts and the AI classifier. Empirical evaluation on synthetic and real-world datasets shows that our proposed probabilistic approach performs competitively, or surpasses previously proposed methods assessed on the same benchmarks.", "keywords": [ "learning to defer", "expectation - maximisation" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 3, 2, 3 ] }, { "paper_id": "zkn2tvtt8J", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Guillermo Jimenez-Perez", "gender": "Male", "institution": "Bayer AG", "country": "ES", "position": "Researcher" }, { "name": "Javier Montalt-Tordera", "gender": "Male", "institution": "Bayer Ag", "country": "ES", "position": "Researcher" }, { "name": "Jens Hooge", "gender": "Male", "institution": "Bayer Ag", "country": "DE", "position": "Principal Researcher" }, { "name": "Josef Cersovsky", "gender": "Male", "institution": "Bayer Ag", "country": "", "position": "Researcher" }, { "name": "Pedro Os\u00f3rio", "gender": "Male", "institution": "Bayer Ag", "country": "", "position": "Intern" }, { "name": "Sadegh Mohammadi", "gender": "Male", "institution": "Universit\u00e0 degli Studi di Genova, Istituto Italiano di Tecnologia", "country": "", "position": "PhD student" }, { "name": "Steffen Vogler", "gender": "Male", "institution": "Bayer Ag", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 50, "n_ref": 103, "n_ref_all": 132, "n_fig": 14, "n_tab": 1, "L_tab": 365, "n_element_tab": 38, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 454, "n_element_tab_1": 53, "formula_len_all": 225, "formula_len_all_1": 204, "len_all": 179641, "len_all_1": 60078, "len_abs": 1858, "len_title": 129, "len_sents": 44384, "len_sents_1": 30595, "n_sents": 270, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1473, "title": "DiNO-Diffusion: Scaling Medical Diffusion Models via Self-Supervised Pre-Training", "abs": "Diffusion models (DMs) require large annotated datasets for training, limiting their applicability in medical imaging where datasets are typically smaller and sparsely annotated. We introduce DiNO-Diffusion, a self-supervised method for training DMs that conditions the generation process on image embeddings extracted from DiNO, a pretrained vision transformer. By not relying on annotations, our training leverages over 868k unlabelled images from public chest X-Ray (CXR) datasets. DiNO-Diffusion shows comprehensive manifold coverage, with FID scores as low as 4.7, and emerging properties when evaluated in downstream tasks, allowing to generate semantically-diverse synthetic datasets even from small data pools, demonstrating up to 20\\% AUC increase in classification performance when used for data augmentation. Results suggest that DiNO-Diffusion could facilitate the creation of large datasets for flexible training of downstream AI models from limited amount of real data, while also holding potential for privacy preservation. Additionally, DiNO-Diffusion demonstrates zero-shot segmentation performance of up to 84.4\\% Dice score when evaluating lung lobe segmentation, evidencing good CXR image-anatomy alignment akin to textual descriptors on vanilla DMs. Finally, DiNO-Diffusion can be easily adapted to other medical imaging modalities or state-of-the-art diffusion models, allowing large-scale, multi-domain image generation pipelines for medical imaging.", "keywords": [ "Diffusion Models", "Generative AI", "Medical Imaging", "Self-Supervision" ], "rating_list": [ 8, 3, 3, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "zkNCWtw2fd", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Adel Elmahdy", "gender": "unknown", "institution": "University of Minnesota, Minneapolis", "country": "", "position": "PhD student" }, { "name": "Amin Ahmad", "gender": "Not Specified", "institution": "Vectara", "country": "US", "position": "Researcher" }, { "name": "Sheng-Chieh Lin", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 30, "n_ref": 60, "n_ref_all": 91, "n_fig": 1, "n_tab": 11, "L_tab": 7398, "n_element_tab": 1312, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 2387, "n_element_tab_1": 420, "formula_len_all": 5498, "formula_len_all_1": 5710, "len_all": 150052, "len_all_1": 67637, "len_abs": 1302, "len_title": 168, "len_sents": 35866, "len_sents_1": 26174, "n_sents": 247, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 120, "L_abs": 1306, "title": "Synergistic Approach for Simultaneous Optimization of Monolingual, Cross-lingual, and Multilingual Information Retrieval", "abs": "Information retrieval across different languages is an increasingly important challenge in natural language processing. Recent approaches based on multilingual pre-trained language models have achieved remarkable success, yet they often optimize for either monolingual, cross-lingual, or multilingual retrieval performance at the expense of others. This paper proposes a novel hybrid batch training strategy to simultaneously improve zero-shot retrieval performance across monolingual, cross-lingual, and multilingual settings while mitigating language bias. The approach fine-tunes multilingual language models using a mix of monolingual and cross-lingual question-answer pair batches sampled based on dataset size. Experiments on XQuAD-R, MLQA-R, and MIRACL benchmark datasets show that the proposed method consistently achieves comparable or superior results in zero-shot retrieval across various languages and retrieval tasks compared to monolingual-only or cross-lingual-only training. Hybrid batch training also substantially reduces language bias in multilingual retrieval compared to monolingual training. These results demonstrate the effectiveness of the proposed approach for learning language-agnostic representations that enable strong zero-shot retrieval performance across diverse languages.", "keywords": [ "Information Retrieval", "Multilingualism and Cross-Lingual NLP", "Question Answering" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "zkMRmW3gcT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Rong Xiao", "gender": "Male", "institution": "Intellifusion", "country": "CN", "position": "Director" }, { "name": "Shaozhe Hao", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Tianyang Hu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Xianbiao Qi", "gender": "Male", "institution": "AITC", "country": "CN", "position": "Principal Researcher" }, { "name": "Xuantong LIU", "gender": "Female", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yuan Yao", "gender": "Unspecified", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Jun WANG", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 55, "n_ref": 111, "n_ref_all": 150, "n_fig": 11, "n_tab": 10, "L_tab": 3333, "n_element_tab": 466, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1235, "n_element_tab_1": 231, "formula_len_all": 304, "formula_len_all_1": 304, "len_all": 162432, "len_all_1": 60938, "len_abs": 180, "len_title": 117, "len_sents": 39435, "len_sents_1": 27769, "n_sents": 270, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1508, "title": "Elucidating the Design Space of Language Models for Image Generation", "abs": "The success of autoregressive (AR) language models in text generation has inspired the computer vision community to adopt Large Language Models (LLMs) for image generation. However, considering the essential differences between text and image modalities, the design space of language models for image generation remains underexplored. We observe that image tokens exhibit greater randomness compared to text tokens, which presents challenges when training with token prediction. Nevertheless, AR models demonstrate their potential by effectively learning patterns even from a seemingly suboptimal optimization problem. Our analysis also reveals that while all models successfully grasp the importance of local information in image generation, smaller models struggle to capture the global context. In contrast, larger models showcase improved capabilities in this area, helping to explain the performance gains achieved when scaling up model size. We further elucidate the design space of language models for vision generation, including tokenizer choice, model choice, model scalability, vocabulary design, and sampling strategy, through extensive comparative experiments. Our work is the first to analyze the optimization behavior of language models in vision generation, and we believe it can inspire more effective designs when applying LMs to other domains. Finally, our elucidated language model for image generation, termed ELM, achieves state-of-the-art performance on the ImageNet 256\u00d7256 benchmark.", "keywords": [ "Image generation", "Large language model", "Generative model" ], "rating_list": [ 5, 6, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3, 3 ] }, { "paper_id": "zkGxROm7D3", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Andrey Kuznetsov", "gender": "Male", "institution": "AIRI", "country": "RU", "position": "Principal Researcher" }, { "name": "Anton Konushin", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "Arsen Kuzhamuratov", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Researcher" }, { "name": "Konstantin Sobolev", "gender": "Male", "institution": "AIRI", "country": "RU", "position": "Researcher" }, { "name": "Mikhail Zhirnov", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Denis Valerievich Dimitrov", "gender": "Male", "institution": "AIRI", "country": "", "position": "Principal Researcher" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 29, "n_ref": 69, "n_ref_all": 119, "n_fig": 10, "n_tab": 10, "L_tab": 3006, "n_element_tab": 313, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 2205, "n_element_tab_1": 184, "formula_len_all": 1667, "formula_len_all_1": 1620, "len_all": 154686, "len_all_1": 59760, "len_abs": 799, "len_title": 126, "len_sents": 42255, "len_sents_1": 26201, "n_sents": 302, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1766, "title": "State & Image Guidance: Teaching Old Text-to-Video Diffusion Models New Tricks", "abs": "Current text-to-video (T2V) models have made significant progress in generating high-quality video. However, these models are limited when it comes to generating dynamic video scenes where the description per frame can vary dramatically. Changing the color, shape, position and state of objects in the scene is a challenge that current video models cannot handle. In addition, the lack of a cheap image-based conditioning mechanism limits their creative application. To address these challenges and extend the applicability of T2V models, we propose two innovative approaches: **State Guidance** and **Image Guidance**. **State Guidance** uses advanced guidance mechanisms to control motion dynamics and scene transformation smoothness by navigating the diffusion process between a state triplet . This mechanism enables the generation of dynamic video scenes (Dynamic Scene T2V) and allows to control the speed and the expressiveness of the scene transformation by introducing temporal dynamics via a guidance weight schedule across video frames. **Image Guidance** enables Zero-Shot Image-to-Video generation (Zero-Shot I2V) by injecting reference image into the initial diffusion steps noise predictions. Furthermore, the combination of **State Guidance** and **Image Guidance** allows for zero-shot transitions between two input reference frames of a video (Zero-Shot II2V). Finally, we introduce the novel **Dynamic Scene Benchmark** to evaluate the ability of the models to generate dynamic video scenes. Extensive experiments show that **State Guidance** and **Image Guidance** successfully address the aforementioned challenges and significantly improve the generation capabilities of existing T2V architectures.", "keywords": [ "Text-to-Video Generation", "Diffusion Models", "Diffusion Guidance", "Zero-shot Image-to-Video Generation" ], "rating_list": [ 5, 5, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 5 ] }, { "paper_id": "zjeHLSiNv1", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Defa Zhu", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Hongzhi Huang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Qiyang Min", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Yutao Zeng", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Zihao Huang", "gender": "unknown", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "zhou xun", "gender": "Female", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "GuoRan", "gender": "Female", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 22, "n_formula_1": 18, "n_ref_uni": 42, "n_ref": 45, "n_ref_all": 75, "n_fig": 14, "n_tab": 10, "L_tab": 3730, "n_element_tab": 478, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 226, "n_element_tab_1": 42, "formula_len_all": 1133, "formula_len_all_1": 1097, "len_all": 142486, "len_all_1": 57148, "len_abs": 905, "len_title": 75, "len_sents": 38282, "len_sents_1": 24811, "n_sents": 318, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 27, "L_abs": 955, "title": "Ultra-Sparse Memory Network", "abs": "It is widely acknowledged that the performance of Transformer models is logarithmically related to their number of parameters and computational complexity. While approaches like Mixture of Experts (MoE) decouple parameter count from computational complexity, they still face challenges in inference due to high memory access costs. This work introduces UltraMem, incorporating large-scale, ultra-sparse memory layer to address these limitations. Our approach significantly reduces inference latency while maintaining model performance. We also investigate the scaling laws of this new architecture, demonstrating that it not only exhibits favorable scaling properties but outperforms MoE. In experiments, the largest UltraMem we train has \\textbf{20 million} memory slots. The results show that our method achieves state-of-the-art inference speed and model performance within a given computational budget, paving the way for billions of slots or experts.", "keywords": [ "Large language model", "sparse model", "scaling law" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "zjAEa4s3sH", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Christopher Earls", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Full Professor" }, { "name": "Nicolas Boulle", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "Assistant Professor" }, { "name": "Rapha\u00ebl Sarfati", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Researcher" }, { "name": "Toni J.B. Liu", "gender": "Male", "institution": "Cornell University", "country": "", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 13, "n_ref_uni": 26, "n_ref": 33, "n_ref_all": 102, "n_fig": 24, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 19, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 950, "formula_len_all_1": 546, "len_all": 119047, "len_all_1": 64111, "len_abs": 3496, "len_title": 89, "len_sents": 34371, "len_sents_1": 26377, "n_sents": 322, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 793, "title": "Lines of Thought in Large Language Models", "abs": "Large Language Models achieve next-token prediction by transporting a vectorized piece of text (prompt) across an accompanying embedding space under the action of successive transformer layers. The resulting high-dimensional trajectories realize different contextualization, or 'thinking', steps, and fully determine the output probability distribution. We aim to characterize the statistical properties of ensembles of these 'lines of thought.' We observe that independent trajectories cluster along a low-dimensional, non-Euclidean manifold, and that their path can be well approximated by a stochastic equation with few parameters extracted from data. We find it remarkable that the vast complexity of such large models can be reduced to a much simpler form, and we reflect on implications.", "keywords": [ "LLM", "latent space", "token trajectories", "interpretability", "transformer" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "ziw5bzg2NO", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Keonwoo Kim", "gender": "Male", "institution": "Kim & Chang", "country": "KR", "position": "Researcher" }, { "name": "Sungzoon Cho", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Taebaek Hwang", "gender": "unknown", "institution": "Waddle Corporation", "country": "KR", "position": "Researcher" }, { "name": "Yeongjae Cho", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 9, "n_ref_uni": 39, "n_ref": 104, "n_ref_all": 135, "n_fig": 10, "n_tab": 6, "L_tab": 2859, "n_element_tab": 338, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1272, "n_element_tab_1": 72, "formula_len_all": 974, "formula_len_all_1": 617, "len_all": 154429, "len_all_1": 57137, "len_abs": 1090, "len_title": 156, "len_sents": 40266, "len_sents_1": 26106, "n_sents": 307, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 108, "L_abs": 1093, "title": "Do You Keep an Eye on What I Ask? Mitigating Multimodal Hallucination via Attention-Guided Ensemble Decoding", "abs": "Recent advancements in Large Vision-Language Models (LVLMs) have significantly expanded their utility in tasks like image captioning and visual question answering. However, they still struggle with object hallucination, where models generate descriptions that inaccurately reflect the visual content by including nonexistent objects or misrepresenting existing ones. While previous methods, such as data augmentation and training-free approaches, strive to tackle this issue, they still encounter scalability challenges and often depend on additional external modules. In this work, we propose Ensemble Decoding (ED), a novel strategy that splits the input image into sub-images and combines logit distributions by assigning weights through the attention map. Furthermore, we introduce ED adaptive plausibility constraint to calibrate logit distribution and FastED, a variant designed for speed-critical applications. Extensive experiments across hallucination benchmarks demonstrate that our proposed method achieves state-of-the-art performance, validating the effectiveness of our approach.", "keywords": [ "Hallucination", "Multimodal Hallucination", "Large Vision-Language Model" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "ziB549CQ30", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Yijian Wang", "gender": "unknown", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zhaoqiang Liu", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 30, "n_formula_1": 29, "n_ref_uni": 43, "n_ref": 179, "n_ref_all": 212, "n_fig": 2, "n_tab": 6, "L_tab": 4741, "n_element_tab": 297, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 758, "n_element_tab_1": 88, "formula_len_all": 1361, "formula_len_all_1": 1196, "len_all": 129956, "len_all_1": 69403, "len_abs": 1441, "len_title": 118, "len_sents": 35730, "len_sents_1": 31487, "n_sents": 276, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1447, "title": "Solving the Fuzzy Job Shop Scheduling Problem via Learning Approaches", "abs": "The fuzzy job shop scheduling problem (FJSSP) emerges as an innovative extension to the conventional job shop scheduling problem (JSSP), incorporating a layer of uncertainty that aligns the model more closely with the complexities of real-world manufacturing environments. This enhancement, while enhancing its applicability, concurrently escalates the computational complexity of deriving solutions. In the domain of traditional scheduling, neural combinatorial optimization (NCO) has recently demonstrated remarkable efficacy. However, its application to the realm of fuzzy scheduling has been relatively unexplored. This paper aims to bridge this gap by investigating the feasibility of employing neural networks to assimilate and process fuzzy information for the resolution of FJSSP, thereby leveraging the advancements in NCO to enhance fuzzy scheduling methodologies. To this end, we present a self-supervised algorithm for the FJSSP (SS-FJSSP). This algorithm employs an iterative mechanism to refine pseudo-labels, progressively transitioning from suboptimal to optimal solutions. This innovative approach adeptly circumvents the significant challenge of procuring true labels, a common challenge in NCO frameworks. Experiments demonstrate that our SS-FJSSP algorithm yields results on a par with the state-of-the-art methods while achieving a remarkable reduction in computational time, specifically being two orders of magnitude faster.", "keywords": [ "Fuzzy job shop scheduling problem", "neural combinatorial optimization", "self-supervised learning" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "zi8YBcmXqA", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Andy Luu Nguyen", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Undergrad student" }, { "name": "Chi Jin", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Assistant Professor" }, { "name": "Seth Karten", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 18, "n_ref": 26, "n_ref_all": 39, "n_fig": 5, "n_tab": 8, "L_tab": 2551, "n_element_tab": 147, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 722, "n_element_tab_1": 112, "formula_len_all": 896, "formula_len_all_1": 731, "len_all": 136046, "len_all_1": 56643, "len_abs": 1383, "len_title": 127, "len_sents": 35013, "len_sents_1": 24988, "n_sents": 271, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1469, "title": "PokeChamp: an Expert-level Minimax Language Agent for Competitive Pokemon", "abs": "We introduce \\texttt{Pok\\'eChamp}, a Large Language Model (LLM) powered game-theoretic aware agent for two-player competitive Pok\\'emon battles, that uses an LLM prior and collected high-Elo human data to model minimax search without any additional training. \\texttt{Pok\\'eChamp} uses a depth-limited minimax search online where the LLM replaces three key components: 1) action sampling from the LLM guided by prompts (including from a damage calculation tool), 2) opponent-modeling via the historical likelihood of actions from our dataset to model the effect of LLM-predicted opponent actions, and 3) state value calculation for the LLM to reflect on each intrinsic state. \\texttt{Pok\\'eChamp} outperforms all existing AIs (76\\%) and heuristic bots (84\\%) by an enormous margin, including winning consistently (>50\\%) against prior human-parity work run with a frontier model, GPT 4-o, while using an open-source 8 billion parameter Llama 3.1 model. \\texttt{Pok\\'eChamp} achieves expert performance in the top 10\\% of players on the online ladder against competitive human players at an Elo of 1500. Finally, we collect the largest Pok\\'emon battling dataset, including 1 million+ games with 150k+ high Elo games, prepare a series of battling benchmarks based on real player data and puzzles to analyze specific battling abilities, and provide crucial updates to the local game engine. Our code is available \\href{https://sites.google.com/view/pokechamp-llm}{online}.", "keywords": [ "multiagent", "LLM agents", "competitive games", "game theory", "reinforcement learning" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "zi3MEZRCqd", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 11, "author_info_list": [ { "name": "Dongsheng Li", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Hoifung Poon", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "General Manager" }, { "name": "Lili Qiu", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Principal Researcher" }, { "name": "Naoto Usuyama", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Sheng Zhang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "XINYANG JIANG", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Researcher" }, { "name": "Xiaoxuan He", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Xufang Luo", "gender": "Female", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" }, { "name": "Yifan Yang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Yuqing Yang", "gender": "unknown", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Zilong Wang", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 41, "n_ref": 87, "n_ref_all": 96, "n_fig": 2, "n_tab": 8, "L_tab": 3568, "n_element_tab": 140, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2526, "n_element_tab_1": 64, "formula_len_all": 256, "formula_len_all_1": 256, "len_all": 170832, "len_all_1": 63356, "len_abs": 1420, "len_title": 131, "len_sents": 40691, "len_sents_1": 30382, "n_sents": 294, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1371, "title": "Learning Robust Representations for Medical Images via Unifying (Self-)Supervisions", "abs": "Pre-training medical image encoder to provide robust, task-agnostic representations is highly valuable, as it enhances the understanding of medical images and is important for performing many data-scarce analysis tasks. Current pre-training works are unable to integrate various types of supervisions, including self-supervision and external supervision such as segmentation annotations, while they are highly valuable for medical image understanding. Therefore, in this paper, we take the first step toward exploring unifying all common types of supervisions into a pre-training framework through a same scalable way. This require the pre-training framework being both unified, for accommodating diverse data and extensible, and effective, for making heterogeneous data synergistically assist unknown downstream tasks. To this end, we propose UmiF, whose principle is that once converted into token embeddings in a unified space, all diverse supervisions can be effectively utilized via contrastive learning and mask modeling with a same way. With UmiF, we pre-train on 1.66M samples from 14 public datasets, significantly surpassing previous efforts in terms of the dataset scale. We obtain and release the UmiF model, which achieved state-of-the-art performance across various downstream tasks, including classification, segmentation, and detection, retrieval and VQA.", "keywords": [ "medical image pre-training", "medical image representation learning" ], "rating_list": [ 6, 3, 5, 6, 3 ], "soundness_list": [ 3, 2, 2, 3, 2 ], "presentation_list": [ 3, 1, 1, 2, 2 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 5 ] }, { "paper_id": "zi0XgnZlcl", "primary_area": "foundation or frontier models, including LLMs", "n_author": 1, "author_info_list": [ { "name": "Zhen Zheng", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 26, "n_ref": 92, "n_ref_all": 101, "n_fig": 2, "n_tab": 5, "L_tab": 2784, "n_element_tab": 317, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2578, "n_element_tab_1": 243, "formula_len_all": 312, "formula_len_all_1": 312, "len_all": 138562, "len_all_1": 68447, "len_abs": 1492, "len_title": 120, "len_sents": 37198, "len_sents_1": 28048, "n_sents": 293, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1500, "title": "MixLLM: Mixed-precision LLM Quantization with Algorithm-system Co-design", "abs": "Quantization has become one of the most effective methodologies to compress LLMs into smaller size.\nHowever, the existing quantization solutions still show limitations of either non-negligible accuracy drop or system inefficiency.\nIn this paper, we make a comprehensive\nanalysis of the general quantization principles on their effect to the triangle of accuracy, memory consumption and system efficiency.\nWe propose MixLLM that explores the new optimization space of mixed-precision quantization between output features based on the insight that different output features matter differently in the model.\nMixLLM identifies the output features with high salience in the global view rather than within each single layer,\neffectively assigning the larger bit-width to output features that need it most to achieve good accuracy with low memory consumption.\nWe present the sweet spot of quantization configuration of algorithm-system co-design that lead to high accuracy and system efficiency.\nTo address the system challenge of this sweet spot, we design the two-step dequantization to make use of the int8 Tensor Core easily and fast data type conversion to reduce dequantization overhead significantly.\nExtensive experiments show that MixLLM achieves the best accuracy on a variety of tasks for the popular LLMs than a set of state-of-the-art works.\nIt shows 0.31 lower perplexity and 0.43\\% improvement on zero shot tasks for Llama 3 8B than QoQ, with similar memory consumption and system efficiency.", "keywords": [ "LLM", "Quantization", "Mixed-precision" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "zhxATDLAmJ", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Alan COLLET", "gender": "unknown", "institution": "CEA", "country": "FR", "position": "Postdoc" }, { "name": "Antonio Bazco-Nogueras", "gender": "unknown", "institution": "IMDEA Networks Institute", "country": "ES", "position": "Postdoc" }, { "name": "Marco Fiore", "gender": "Male", "institution": "IMDEA Networks Institute", "country": "ES", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 44, "n_ref": 74, "n_ref_all": 102, "n_fig": 8, "n_tab": 7, "L_tab": 2718, "n_element_tab": 299, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 136, "n_element_tab_1": 17, "formula_len_all": 813, "formula_len_all_1": 560, "len_all": 171534, "len_all_1": 73945, "len_abs": 1561, "len_title": 121, "len_sents": 60901, "len_sents_1": 38493, "n_sents": 388, "n_sents_1": 230, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1566, "title": "Loss2Net: Loss Meta-Learning for Regression with A-priori Unknown Metrics", "abs": "There exist many practical applications where regression tasks must cope with a generally overseen problem: the output variable to be computed, which is often a decision variable, impacts the performance metric to optimize in a manner that is not known a priori. This challenge translates into a loss-metric mismatch, which makes standard loss functions such as Mean Square Error (MSE) not suitable because they significantly hinder the final performance. While this problem is of crucial importance in, e.g., many engineering and economic applications, the literature in meta-learning of loss functions has focused on other problems, such as classification or few-shot learning tasks. In this work, we aim at closing this research gap by proposing a model that can handle common situations in real systems where the unknown prediction-metric relationship is time-correlated, non-differentiable, or depends on multiple intertwined predictions. We present a novel loss meta-learning architecture for regression, named Loss2Net, which is able to (i) jointly learn the actual regressor and the loss function that it should minimize, directly from system responses; (ii) it does so without any assumption on the loss function structure; (iii) it provides a manner to learn non-differentiable and multi-dimensional loss functions from entangled performance metrics. Detailed experiments for power grid and telecommunications infrastructure optimization, grounded on real-world measurement data, demonstrate how Loss2Net can effectively learn unidentified loss functions.", "keywords": [ "loss meta-learning", "loss-metric mismatch", "system management", "unknown metric", "transfer learning" ], "rating_list": [ 3, 5, 5, 3, 6 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 5, 3 ] }, { "paper_id": "zhFyKgqxlz", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Julia Gaudio", "gender": "Female", "institution": "Northwestern University", "country": "US", "position": "Assistant Professor" }, { "name": "Nirmit Joshi", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "", "position": "PhD student" } ], "n_formula": 178, "n_formula_1": 29, "n_ref_uni": 48, "n_ref": 111, "n_ref_all": 136, "n_fig": 1, "n_tab": 2, "L_tab": 276, "n_element_tab": 14, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18615, "formula_len_all_1": 1819, "len_all": 294355, "len_all_1": 69343, "len_abs": 4140, "len_title": 131, "len_sents": 86067, "len_sents_1": 29069, "n_sents": 909, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1297, "title": "Exact Community Recovery under Side Information: Optimality of Spectral Algorithms", "abs": "We study the problem of exact community recovery in general, two-community block models, in the presence of node-attributed *side information*. We allow for a very general side information channel for node attributes, and for pairwise (edge) observations, consider both Bernoulli and Gaussian matrix models, capturing the Stochastic Block Model, Submatrix Localization, and $\\mathbb{Z}_2$-Synchronization as special cases. A recent work of Dreveton et al. 2024 characterized the information-theoretic limit of a very general exact recovery problem with side information. In this paper, we show algorithmic achievability in the above important cases by designing a simple but optimal spectral algorithm that incorporates side information (when present) along with the eigenvectors of the pairwise observation matrix. Using the powerful tool of entrywise eigenvector analysis [Abbe et al. 2020], we show that our spectral algorithm can mimic the so called *genie-aided estimators*, where the $i^{\\mathrm{th}}$ genie-aided estimator optimally computes the estimate of the $i^{\\mathrm{th}}$ label, when all remaining labels are revealed by a genie. This perspective provides a unified understanding of the optimality of spectral algorithms for various exact recovery problems in a recent line of work.", "keywords": [ "Community Detection", "Spectral Algorithms", "Side Information" ], "rating_list": [ 6, 3, 6, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "zgs450VzkU", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Alvin Wan", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Researcher" }, { "name": "Colorado Reed", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Ion Stoica", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Joseph E. Gonzalez", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Associate Professor" }, { "name": "Pranav Ramesh", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "Undergrad student" }, { "name": "Shishir G Patil", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Qi Shan", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 44, "n_ref_all": 67, "n_fig": 13, "n_tab": 8, "L_tab": 4750, "n_element_tab": 427, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1738, "n_element_tab_1": 194, "formula_len_all": 77, "formula_len_all_1": 0, "len_all": 124103, "len_all_1": 56355, "len_abs": 1239, "len_title": 152, "len_sents": 37719, "len_sents_1": 26681, "n_sents": 228, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1243, "title": "Test-Time RAG: Enhancing Long Context Understanding in LLMs with Retrieval-Augmented Mechanisms", "abs": "Large Language Models (LLMs) are becoming increasingly pivotal in applications that depend on extensive personalized context, such as conversational agents and specialized task-oriented systems. In these scenarios, effective long-context handling is essential to support agentic tasks and enhance in-context learning capabilities. To address this challenge, we propose a novel integration of Retrieval Augmented Generation (RAG) techniques with LLMs, designed to enhance their ability to effectively manage and utilize large contextual information only available at test time. Our methodology, Test-Time RAG (TTRAG), enriches LLMs by dynamically generating novel conditional embeddings coupled with query rewriting and utilizing semantic search to retrieve the most relevant document chunks at test time. This process preserves the context\u2019s meaning and enhances the model\u2019s responsiveness and accuracy in knowledge-intensive Question Answering (QA) tasks. Our evaluations demonstrate our system\u2019s ability synthesize and retrieve information across extensive texts: HotpotQA (+17.29%), QASPER (+4.39%), and Natural Questions (+8.73%), demonstrating the effectiveness of TTRAG across varied context lengths from 1 million to 9.6 million tokens.", "keywords": [ "Retreival Augmented Generation (RAG)", "Personalization", "LLM" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "zgXGNXkC0F", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chuong Nguyen", "gender": "Male", "institution": "CSIRO", "country": "AU", "position": "Researcher" }, { "name": "Hongdong Li", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "Full Professor" }, { "name": "Maying Shen", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Deep Learning R&D Engineer" }, { "name": "Nadine Chang", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Shan Wang", "gender": "Female", "institution": "Australian National University", "country": "AU", "position": "PhD student" }, { "name": "Jose M Alvarez", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Director" } ], "n_formula": 22, "n_formula_1": 19, "n_ref_uni": 13, "n_ref": 32, "n_ref_all": 64, "n_fig": 9, "n_tab": 5, "L_tab": 1400, "n_element_tab": 239, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 650, "n_element_tab_1": 153, "formula_len_all": 1471, "formula_len_all_1": 1434, "len_all": 136051, "len_all_1": 71621, "len_abs": 1569, "len_title": 119, "len_sents": 33198, "len_sents_1": 31469, "n_sents": 268, "n_sents_1": 251, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1528, "title": "Mitigating Multimodal Hallucinations via Gradient-based Self-Reflection", "abs": "Hallucination in Multimodal Large Language Models (MLLMs) occurs when inaccurate text-visual alignments are generated, posing a major challenge for reliable model output. Previous studies have identified three primary biases as major causes of hallucinations: text-visual bias (over-reliance on text over visual details), co-occurrence bias (misleading object correlations), and long-term bias (increased hallucinations in later stages of long sequences). Existing hallucination mitigation methods often rely on visual grounding, which requires additional resources such as scoring systems using another MLLM, and still fail to fully address all biases, particularly co-occurrence bias in visual inputs. We propose Gradient-based Influence-Aware Contrastive Decoding (GACD) to explicitly and jointly balance these biases, thereby mitigating hallucinations. To quantify these biases at the individual sample level, we introduce `token influence'. Since biases are rooted in the training data and become embedded in pre-trained MLLMs, we derive token influence through self-reflection by calculating the gradients from output predictions to input tokens. Notably, GACD is the first approach capable of fully addressing co-occurrence bias without relying on extra resources or any form of tuning. Extensive experiments demonstrate GACD's effectiveness in reducing hallucinations and improving MLLM performance, achieving new state-of-the-art results while providing insights into the visual perception capabilities of these models.", "keywords": [ "Hallucinations", "MLLMs", "Gradient-based Analysis" ], "rating_list": [ 3, 5, 8, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "zgM66fu0wv", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Lizhen Qu", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Lecturer" }, { "name": "Niket Tandon", "gender": "Male", "institution": "Research, Microsoft", "country": "IN", "position": "Principal Researcher" }, { "name": "Reza Haffari", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" }, { "name": "Tao Feng", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 31, "n_ref": 50, "n_ref_all": 67, "n_fig": 3, "n_tab": 10, "L_tab": 2799, "n_element_tab": 248, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 34, "n_element_tab_1": 1, "formula_len_all": 134, "formula_len_all_1": 133, "len_all": 149296, "len_all_1": 58602, "len_abs": 1214, "len_title": 122, "len_sents": 44877, "len_sents_1": 28894, "n_sents": 325, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1216, "title": "IRIS: An Iterative and Integrated Framework for Real-Time Causal Discovery", "abs": "Causal discovery is fundamental to scientific research, yet traditional statistical algorithms face significant challenges, including expensive data collection, redundant examination of known relations, and unrealistic assumptions. Additionally, while recent LLM-based methods excel at identifying commonly known causal relations, they fall short in uncovering novel relations. We introduce IRIS (Iterative Retrieval and Integrated System for Real-Time Causal Discovery), a novel framework that addresses these limitations. Starting with a set of initial variables, IRIS automatically retrieves relevant documents, extracts variable values, and organizes data for statistical algorithms in real-time. Our hybrid causal discovery method combines statistical algorithms and LLM-based methods to discover existing and novel causal relations. The missing variable proposal component identifies missing variables, and subsequently, IRIS expands the causal graphs by including both the initial and the newly suggested variables. Our approach offers a scalable and adaptable solution for causal discovery, enabling the exploration of causal relations from a set of initial variables without requiring pre-existing datasets.", "keywords": [ "causal discovery", "real-time", "large language model" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "zg3ec1TdAP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Christopher Re", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Ethan Steinberg", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Researcher" }, { "name": "Jason Alan Fries", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Research Scientist" }, { "name": "Michael Wornow", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Miguel Angel Fuentes Hernandez", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Nigam Shah", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Suhana Bedi", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" } ], "n_formula": 30, "n_formula_1": 4, "n_ref_uni": 47, "n_ref": 129, "n_ref_all": 179, "n_fig": 15, "n_tab": 11, "L_tab": 4791, "n_element_tab": 863, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 969, "n_element_tab_1": 137, "formula_len_all": 1608, "formula_len_all_1": 255, "len_all": 247951, "len_all_1": 66933, "len_abs": 1923, "len_title": 126, "len_sents": 72412, "len_sents_1": 32243, "n_sents": 546, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1938, "title": "Context Clues: Evaluating Long Context Models for Clinical Prediction Tasks on EHR Data", "abs": "Foundation Models (FMs) trained on Electronic Health Records (EHRs) have achieved state-of-the-art results on numerous clinical prediction tasks. However, prior EHR FMs typically have context windows of $<$1k tokens, which prevents them from modeling full patient EHRs which can exceed 10k's of events. For making clinical predictions, both model performance and robustness to the unique properties of EHR data are crucial. Recent advancements in subquadratic long-context architectures (e.g. Mamba) offer a promising solution. However, their application to EHR data has not been well-studied. We address this gap by presenting the first systematic evaluation of the effect of context length on modeling EHR data. We find that longer context models improve predictive performance -- our Mamba-based model surpasses the prior state-of-the-art on 9/14 tasks on the EHRSHOT prediction benchmark. Additionally, we measure robustness to three unique, previously underexplored properties of EHR data: (1) the prevalence of ``copy-forwarded\" diagnoses which create artificial token repetition in EHR sequences; (2) the irregular time intervals between EHR events which can lead to a wide range of timespans within a context window; and (3) the natural increase in disease complexity over time which makes later tokens in the EHR harder to predict than earlier ones. Stratifying our EHRSHOT results, we find that higher levels of each property correlate negatively with model performance (e.g., a 14% higher Brier loss between the least and most irregular patients), but that longer context models are more robust to more extreme levels of these properties. Our work highlights the potential for using long-context architectures to model EHR data, and offers a case study on how to identify and quantify new challenges in modeling sequential data motivated by domains outside of natural language. We release all of our model checkpoints and code.", "keywords": [ "ehr", "foundation model", "long context", "clinical prediction making", "healthcare" ], "rating_list": [ 5, 8, 6, 10, 6 ], "soundness_list": [ 2, 4, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 4, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "zfgYC3sDt6", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Hongxin Wei", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Yixuan Li", "gender": "Female", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Shuoyuan Wang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 12, "n_ref_uni": 51, "n_ref": 123, "n_ref_all": 168, "n_fig": 14, "n_tab": 17, "L_tab": 9816, "n_element_tab": 1565, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 2240, "n_element_tab_1": 366, "formula_len_all": 1281, "formula_len_all_1": 740, "len_all": 222644, "len_all_1": 68924, "len_abs": 1174, "len_title": 135, "len_sents": 57699, "len_sents_1": 29538, "n_sents": 481, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1179, "title": "Understanding and Mitigating Miscalibration in Prompt Tuning for Vision-Language Models", "abs": "Confidence calibration is critical for the safe deployment of machine learning models in the real world.\nHowever, such issue in vision-language models like CLIP, particularly after fine-tuning, has not been fully addressed.\nIn this work, we demonstrate that existing prompt tuning methods usually lead to a trade-off of calibration between base and new classes:\nthe cross-entropy loss in CoOp causes overconfidence in new classes by increasing textual label divergence, whereas the regularization of KgCoOp maintains the confidence level but results in underconfidence in base classes due to the improved accuracy.\nInspired by the observations, we introduce Dynamic Outlier Regularization (DOR) to ensure the confidence calibration on both base and new classes after fine-tuning. \nIn particular, we propose to minimize the feature deviation of novel textual labels (instead of base classes) sampled from a large vocabulary.\nIn effect, DOR prevents the increase in textual divergence for new labels while easing restrictions on base classes.\nExtensive experiments demonstrate that DOR can enhance the calibration performance of current fine-tuning methods on base and new classes.", "keywords": [ "Vision-Language Models", "Confidence calibration", "Outlier Regularization", "Prompt Tuning" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "zfeso8ceqr", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "David Brandfonbrener", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Depen Morwani", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Nikhil Vyas", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Rosie Zhao", "gender": "Female", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Sham M. Kakade", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 39, "n_ref": 79, "n_ref_all": 126, "n_fig": 17, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 202, "formula_len_all_1": 202, "len_all": 137025, "len_all_1": 53807, "len_abs": 1546, "len_title": 125, "len_sents": 43886, "len_sents_1": 26789, "n_sents": 298, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1553, "title": "Deconstructing What Makes a Good Optimizer for Autoregressive Language Models", "abs": "Training language models becomes increasingly expensive with scale, prompting numerous attempts to improve optimization efficiency. Despite these efforts, the Adam optimizer remains the most widely used, due to a prevailing view that it is the most effective approach. We aim to compare several optimization algorithms, including SGD, Adafactor, Adam, Lion, and Sophia in the context of autoregressive language modeling across a range of model sizes, hyperparameters, and architecture variants. Our findings indicate that, except for SGD, these algorithms all perform comparably both in their optimal performance and also in terms of how they fare across a wide range of hyperparameter choices. Our results suggest to practitioners that the choice of optimizer can be guided by practical considerations like memory constraints and ease of implementation, as no single algorithm emerged as a clear winner in terms of performance or stability to hyperparameter misspecification. Given our findings, we further dissect these approaches, examining two simplified versions of Adam: a) signed momentum (Signum) which we see recovers both the performance and hyperparameter stability of Adam and b) Adalayer, a layerwise variant of Adam which we introduce to study the impact on Adam's preconditioning for different layers of the network. Examining Adalayer leads us to the conclusion that, perhaps surprisingly, adaptivity on *both* the last layer and LayerNorm parameters in particular are necessary for retaining performance and stability to learning rate.", "keywords": [ "optimization", "LLMs", "language models", "Adam" ], "rating_list": [ 5, 5, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "zfQA8y3n2o", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Dongwon Lee", "gender": "Male", "institution": "The Pennsylvania State University", "country": "US", "position": "Full Professor" }, { "name": "Hui Liu", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Suhang Wang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Associate Professor" }, { "name": "Xianfeng Tang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Xianren Zhang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Zongyu Wu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Qi He", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Director of Applied Science" } ], "n_formula": 11, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 89, "n_ref_all": 121, "n_fig": 6, "n_tab": 24, "L_tab": 9317, "n_element_tab": 611, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1277, "n_element_tab_1": 160, "formula_len_all": 414, "formula_len_all_1": 220, "len_all": 210783, "len_all_1": 73259, "len_abs": 3608, "len_title": 119, "len_sents": 65139, "len_sents_1": 34400, "n_sents": 593, "n_sents_1": 267, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 2379, "title": "Divide-Verify-Refine: Aligning LLM Responses with Complex Instructions", "abs": "Recent studies show that LLMs, particularly open-source models, struggle to follow complex instructions with multiple constraints, hindering their adoption in mission-critical applications. Despite the importance, methods to improve LLMs' adherence to such constraints remain largely unexplored, and current research focuses primarily on evaluating this ability rather than developing solutions. While a few studies enhance constraint adherence through model tuning, this approach is computationally expensive and heavily reliant on training data quality. An alternative is to leverage LLMs' self-correction capabilities, allowing them to adjust responses to better meet specified constraints. However, this self-correction ability of LLMs is limited by the feedback quality, as LLMs cannot autonomously generate reliable feedback or detect errors. Moreover, the self-refinement process heavily depends on few-shot examples that illustrate how to modify responses to meet constraints. As constraints in complex instructions are diverse and vary widely (e.g., text length, number of bullet points, or inclusion of specific keywords), manually crafting few-shot examples for each constraint type can be labor-intensive and sub-optimal. To deal with these two challenges, we propose the Divide-Verify-Refine (DVR) framework with three steps: (1) Divide complex instructions into single constraints and prepare appropriate tools; (2) Verify: To address the feedback quality problem, these tools will rigorously verify responses and provide reliable feedback (e.g., Python scripts for format checking or pre-trained classifiers for content analysis); (3) Refine: To address the constraint diversity challenge, we design a refinement repository that collects successful refinement processes and uses them as few-shot demonstrations for future cases, allowing LLMs to learn from the past experience during inference. Additionally, recognizing that existing datasets lack complexity and have internal conflict, we develop a new dataset of complex instructions, each containing 1-6 constraints. Experiments show that the framework significantly improves performance, doubling LLama3.1-8B's constraint adherence and tripling Mistral-7B's performance on instructions with 6 constraints. The code and dataset are available at https://anonymous.4open.science/r/CODE_ICLR2025-52CE/README.md", "keywords": [ "Large Language Model", "Instruction Following", "Constraints Following" ], "rating_list": [ 5, 3, 6, 8, 5 ], "soundness_list": [ 2, 3, 3, 4, 3 ], "presentation_list": [ 3, 2, 2, 4, 3 ], "contribution_list": [ 2, 2, 2, 3, 1 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "zfIxlvKq4u", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Binglei Li", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Le Zhuo", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Qi Qin", "gender": "Not Specified", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Siqi Luo", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xiaohong Liu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xu Guo", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Yi Xin", "gender": "Male", "institution": "Nanjing university", "country": "CN", "position": "PhD student" }, { "name": "Chang Xu", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Gao Peng", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 26, "n_ref": 99, "n_ref_all": 125, "n_fig": 13, "n_tab": 6, "L_tab": 2036, "n_element_tab": 189, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 1077, "n_element_tab_1": 138, "formula_len_all": 89, "formula_len_all_1": 0, "len_all": 129659, "len_all_1": 58994, "len_abs": 261, "len_title": 144, "len_sents": 34223, "len_sents_1": 26377, "n_sents": 277, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1700, "title": "Exploring the Design Space of Autoregressive Models for Efficient and Scalable Image Generation", "abs": "Autoregressive (AR) models and their variants are re-revolutionizing visual generation with improved frameworks. However, unlike the well-established practices for building diffusion models, there lacks a comprehensive recipe for AR models, e.g., selecting image tokenizers, model architectures, and AR paradigms. In this work, we delve into the design space of general AR models, including Mask Autoregressive (MAR) models, to identify optimal configurations for efficient and scalable image generation. We first conduct a detailed evaluation of four prevalent image tokenizers across both AR and MAR settings, examining the impact of codebook size (ranging from 1,024 to 262,144) on generation quality, and identify the most effective tokenizer for image generation. Building on these insights, we propose an enhanced MAR model architecture, named Masked Generative Image LLaMA (MaskGIL), comprising of LlamaGen-VQ and Bidirectional LLaMA. To ensure stable scaling, we introduce modifications such as query-key normalization and post-normalization, resulting in a series of class-conditional MaskGIL models, ranging from 111M to 1.4B parameters. MaskGIL significantly improves the MAR baseline, achieving an 3.71 FID comparable to state-of-the-art AR models on the ImageNet 256$\\times$256 benchmark, with only 8 inference steps, far fewer than the 256 steps needed for AR models. Additionally, we introduce a text-conditional MaskGIL model with 775M parameters, capable of flexibly generating images at any resolution with high aesthetics. To bridge AR and MAR image generation, we investigate their combination during the inference phase. We release all models and code to foster further research.", "keywords": [ "Image Generation", "Autoregressive Model" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "zf777Odl6J", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Hongtao Li", "gender": "Male", "institution": "China Internet Network Information Center", "country": "CN", "position": "Principal Researcher" }, { "name": "Jianfan Chen", "gender": "Male", "institution": "China Internet Network Information Center", "country": "CN", "position": "Researcher" }, { "name": "Kejun Dong", "gender": "Male", "institution": "China Internet Network Information Center", "country": "CN", "position": "Researcher" }, { "name": "Xuebiao Yuchi", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Zhiwei Yan", "gender": "Male", "institution": "cnnic", "country": "CN", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 16, "n_ref": 27, "n_ref_all": 32, "n_fig": 2, "n_tab": 2, "L_tab": 640, "n_element_tab": 74, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 640, "n_element_tab_1": 74, "formula_len_all": 374, "formula_len_all_1": 374, "len_all": 80323, "len_all_1": 50542, "len_abs": 981, "len_title": 104, "len_sents": 27523, "len_sents_1": 23304, "n_sents": 181, "n_sents_1": 154, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 984, "title": "KA-GAT: Kolmogorov\u2013Arnold based Graph Attention Networks", "abs": "Graph Neural Networks (GNNs) have demonstrated remarkable capabilities in processing graph-structured data, but they often struggle with high-dimensional features and complex, nonlinear relationships. To address these challenges, we propose KA-GAT, a novel model that integrates Kolmogorov-Arnold Networks (KANs) with Graph Attention Networks (GATs). KA-GAT leverages KAN to decompose and reconstruct high-dimensional features, enhancing representational capacity, while a multi-head attention mechanism dynamically focuses on key graph components, improving interpretability. Experimental results on benchmark datasets, including Cora and Citeseer, demonstrate that KA-GAT achieves significant accuracy improvements compared to baseline models like GAT, with a relative gain of 4.5\\% on Cora. These findings highlight KA-GAT\u2019s robustness and potential as an interpretable and scalable solution for high-dimensional graph data, paving the way for further advancements in GNN research.", "keywords": [ "Graph Neural Networks", "Kolmogorov-Arnold Networks", "Graph Attention Networks", "Multi-head Attention Mechanism", "Model Interpretability" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 1, 3, 1, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "zf53vmj6k4", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Isack Lee", "gender": "Male", "institution": "theori", "country": "KR", "position": "Researcher" }, { "name": "Haebin Sung", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 40, "n_ref": 65, "n_ref_all": 88, "n_fig": 5, "n_tab": 10, "L_tab": 3137, "n_element_tab": 243, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1475, "n_element_tab_1": 60, "formula_len_all": 232, "formula_len_all_1": 201, "len_all": 154221, "len_all_1": 52770, "len_abs": 1584, "len_title": 141, "len_sents": 35025, "len_sents_1": 24228, "n_sents": 257, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1577, "title": "Do LLMs Have Political Correctness? Analyzing Ethical Biases and Jailbreak Vulnerabilities in AI Systems", "abs": "Although large language models (LLMs) demonstrate impressive proficiency in various tasks, they present potential safety risks, such as 'jailbreaks', where malicious inputs can coerce LLMs into generating harmful content bypassing safety alignments. In this paper, we delve into the ethical biases in LLMs and examine how those biases could be exploited for jailbreaks. Notably, these biases result in a jailbreaking success rate in GPT-4o models that differs by 20% between non-binary and cisgender keywords and by 16% between white and black keywords, even when the other parts of the prompts are identical. We introduce the concept of BiasJailbreak, highlighting the inherent risks posed by these safety-induced biases. BiasJailbreak generates biased keywords automatically by asking the target LLM itself, and utilizes the keywords to generate harmful output. Additionally, we propose an efficient defense method BiasDefense, which prevents jailbreak attempts by injecting defense prompts prior to generation. BiasDefense stands as an appealing alternative to Guard Models, such as Llama-Guard, that require additional inference cost after text generation. Our findings emphasize that ethical biases in LLMs can actually lead to generating unsafe output, and suggest a method to make the LLMs more secure and unbiased. To enable further research and improvements, we open-source our [code and artifacts](https://anonymous.4open.science/r/PCJailbreak-F2B0) of BiasJailbreak, providing the community with tools to better understand and mitigate safety-induced biases in LLMs.", "keywords": [ "LLM", "safety", "jailbreak" ], "rating_list": [ 8, 3, 1, 5 ], "soundness_list": [ 3, 2, 1, 1 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "zeeLxGw5pp", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Apostolis Zarras", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Misha Glazunov", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "Postdoc" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 29, "n_ref": 56, "n_ref_all": 70, "n_fig": 3, "n_tab": 9, "L_tab": 2180, "n_element_tab": 269, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 834, "n_element_tab_1": 99, "formula_len_all": 787, "formula_len_all_1": 776, "len_all": 110899, "len_all_1": 66987, "len_abs": 1703, "len_title": 71, "len_sents": 35260, "len_sents_1": 31093, "n_sents": 250, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1711, "title": "Enhancing Robustness of Deep Learning via Unified Latent Representation", "abs": "Adversarial examples and Out-of-Distribution (OoD) inputs constitute major problematic instances for the image classifiers based on Deep Neural Networks (DNNs). In particular, DNNs tend to be overconfident with their predictions, assigning a different category with a high probability. In this work, we suggest a combined solution to tackle both input types based on the Variational Autoencoder (VAE). First, we scrutinize the recent successful results in detecting OoDs utilizing Bayesian epistemic uncertainty estimation over weights of VAEs. Surprisingly, contrary to the previous claims in the literature, we discover that we can obtain comparable detection performance utilizing a standard procedure of importance sampling with the classical formulation of VAE. Second, we dissect the marginal likelihood approximation, analyzing the primary source of variation responsible for distinguishing inliers versus outliers, and establish a link with the recent promising results in detecting outliers using latent holes. Finally, we identify that adversarial examples and OoD inputs have similar latent representations. This insight allows us to develop separate methods to automatically distinguish between them by considering their non-similarities in the input space. The suggested approach enables pre-training a VAE model on specific input data, allowing it to act as a gatekeeper. This achieves two major goals: defending the DNN classifier against potential attacks and flagging OoDs. Once pre-trained, VAE can be plugged as a filter into any DNN image classifier of arbitrary architecture trained on the same data inputs without the need for its retraining or accessing the layers and weights of the DNN.", "keywords": [ "deep learning robustness", "out-of-distribution inputs", "adversarial examples", "VAE latent representation" ], "rating_list": [ 1, 5, 1, 1, 8 ], "soundness_list": [ 1, 3, 2, 1, 4 ], "presentation_list": [ 1, 2, 2, 2, 3 ], "contribution_list": [ 1, 2, 1, 1, 4 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "zeBhcfP8tN", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "An Yan", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Ran Xu", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "senior manager" }, { "name": "Senthil Purushwalkam", "gender": "Male", "institution": "Salesforce", "country": "", "position": "Researcher" }, { "name": "Viraj Uday Prabhu", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Research Scientist" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 28, "n_ref": 73, "n_ref_all": 107, "n_fig": 17, "n_tab": 2, "L_tab": 1251, "n_element_tab": 132, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 773, "n_element_tab_1": 81, "formula_len_all": 473, "formula_len_all_1": 221, "len_all": 129570, "len_all_1": 82667, "len_abs": 1328, "len_title": 105, "len_sents": 39233, "len_sents_1": 30549, "n_sents": 335, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1193, "title": "Trust but Verify: Programmatic VLM Evaluation in the Wild", "abs": "Vision-Language Models (VLMs) often generate plausible but incorrect responses to visual queries. However, reliably quantifying the effect of such hallucinations in free-form responses to open-ended queries is challenging as it requires visually verifying each claim within the response. We propose Programmatic VLM Evaluation (PROVE), a new benchmarking paradigm for evaluating VLM responses to open-ended queries. To construct PROVE, we provide a large language model with a high-fidelity scene-graph representation constructed from a hyper-detailed image caption, and prompt it to generate diverse question-answer (QA) pairs, as well as programs that can be executed over the scene graph object to _verify_ each QA pair. We thus construct a benchmark of 10.5k challenging but grounded visual QA pairs. Next, to evaluate free-form model responses to queries in PROVE, we propose a _programmatic_ evaluation strategy that measures both the helpfulness and truthfulness of a response within a unified scene graph-based framework. We benchmark the helpfulness-truthfulness trade-offs of a range of VLMs on PROVE, finding that very few are in-fact able to achieve a good balance between the two.", "keywords": [ "vision-language models", "evaluation", "hallucinations" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "zeAOzn80VQ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Bingde Hu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" }, { "name": "Genlang Chen", "gender": "Male", "institution": "NingboTech University", "country": "CN", "position": "Full Professor" }, { "name": "Haofei Zhang", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Jie Song", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Mengqi Xue", "gender": "Female", "institution": "Hangzhou City University", "country": "CN", "position": "Researcher" }, { "name": "Mingli Song", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Xingen Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Yuechen Xie", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 35, "n_ref": 61, "n_ref_all": 90, "n_fig": 7, "n_tab": 9, "L_tab": 1252, "n_element_tab": 171, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 113, "n_element_tab_1": 5, "formula_len_all": 1203, "formula_len_all_1": 863, "len_all": 148152, "len_all_1": 58827, "len_abs": 1368, "len_title": 112, "len_sents": 41098, "len_sents_1": 27503, "n_sents": 323, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1377, "title": "Dataset Ownership Verification in Contrastive Pre-trained Models", "abs": "High-quality open-source datasets, which necessitate substantial efforts for curation, has become the primary catalyst for the swift progress of deep learning. Concurrently, protecting these datasets is paramount for the well-being of the data owner. Dataset ownership verification emerges as a crucial method in this domain, but existing approaches are often limited to supervised models and cannot be directly extended to increasingly popular unsupervised pre-trained models. In this work, we propose the first dataset ownership verification method tailored specifically for self-supervised pre-trained models by contrastive learning. Its primary objective is to ascertain whether a suspicious black-box backbone has been pre-trained on a specific unlabeled dataset, aiding dataset owners in upholding their rights. The proposed approach is motivated by our empirical insights that when models are trained with the target dataset, the unary and binary instance relationships within the embedding space exhibit significant variations compared to models trained without the target dataset. We validate the efficacy of this approach across multiple contrastive pre-trained models including SimCLR, BYOL, SimSiam, MOCO v3, and DINO. The results demonstrate that our method rejects the null hypothesis with a $p$-value markedly below $0.05$, surpassing all previous methodologies.", "keywords": [ "Dataset Ownership Verification", "Data Protection", "Contrastive Learning", "Pre-trained Models", "Self-supervised Learning" ], "rating_list": [ 8, 6, 5, 8, 8 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 1, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3, 4 ] }, { "paper_id": "zdKgyC2vnQ", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Enming Zhang", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Qinghai Miao", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Xingyuan Dai", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Yisheng Lv", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Science", "country": "CN", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 12, "n_ref": 34, "n_ref_all": 50, "n_fig": 8, "n_tab": 4, "L_tab": 1093, "n_element_tab": 97, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 866, "n_element_tab_1": 65, "formula_len_all": 271, "formula_len_all_1": 304, "len_all": 96776, "len_all_1": 53489, "len_abs": 1453, "len_title": 163, "len_sents": 31852, "len_sents_1": 24586, "n_sents": 245, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 115, "L_abs": 1458, "title": "MiniDrive: More Efficient Vision-Language Models with Multi-Level 2D Features as Text Tokens for Autonomous Driving", "abs": "Vision-language models (VLMs) serve as general-purpose end-to-end models in autonomous driving, performing subtasks such as prediction, planning, and perception through question-and-answer interactions. However, most existing methods rely on computationally expensive visual encoders and large language models (LLMs), making them difficult to deploy in real-world scenarios and real-time applications. Meanwhile, most existing VLMs lack the ability to process multiple images, making it difficult to adapt to multi-camera perception in autonomous driving. To address these issues, we propose a novel framework called MiniDrive, which incorporates our proposed Feature Engineering Mixture of Experts (FE-MoE) module and Dynamic Instruction Adapter (DI-Adapter). The FE-MoE effectively maps 2D features into visual token embeddings before being input into the language model. The DI-Adapter enables the visual token embeddings to dynamically change with the instruction text embeddings, resolving the issue of static visual token embeddings for the same image in previous approaches. The DI-Adapter enables the FE-MoE to further extract and process 2D visual features based on user instructions, focus on attention regions, and reduce redundancy. Compared to previous works, MiniDrive achieves state-of-the-art performance in terms of parameter size, floating point operations, and response efficiency, with the smallest version containing only 83M parameters.", "keywords": [ "Vision-language models", "Autonomous driving" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "zd5Knrtja4", "primary_area": "learning on time series and dynamical systems", "n_author": 7, "author_info_list": [ { "name": "Chang Xu", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Huanyu Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Jiang Bian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Liang Wang", "gender": "Male", "institution": "Institute of Automation\uff0c CAS\uff0cChina", "country": "CN", "position": "Full Professor" }, { "name": "Tieniu Tan", "gender": "Male", "institution": " Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Professor" }, { "name": "Zhang Zhang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "yifan zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 30, "n_ref": 63, "n_ref_all": 92, "n_fig": 7, "n_tab": 10, "L_tab": 4487, "n_element_tab": 366, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1326, "n_element_tab_1": 231, "formula_len_all": 358, "formula_len_all_1": 454, "len_all": 145878, "len_all_1": 63522, "len_abs": 1179, "len_title": 131, "len_sents": 45505, "len_sents_1": 27972, "n_sents": 345, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1199, "title": "TimeRAF: Retrieval-Augmented Foundation model for Zero-shot Time Series Forecasting", "abs": "Time series forecasting plays a crucial role in data mining, driving rapid advancements across numerous industries. \nWith the emergence of large models, time series foundation models (TSFMs) have exhibited remarkable generalization capabilities, such as zero-shot learning, through large-scale pre-training. \nMeanwhile, Retrieval-Augmented Generation (RAG) methods are widely employed to enhance the performance of foundation models on unseen data, allowing models to access to external knowledge. \nIn this paper, we introduce **TimeRAF**, a **R**etrieval-**A**ugmented **F**orecasting model that enhance zero-shot time series forecasting through retrieval-augmented techniques.\nWe develop customized time series knowledge bases that are tailored to the specific forecasting tasks.\nTimeRAF employs an end-to-end learnable retriever to extract valuable information from the knowledge base.\nAdditionally, we propose Channel Prompting for knowledge integration, which effectively extracts relevant information from the retrieved knowledge along the channel dimension.\nExtensive experiments demonstrate the effectiveness of our model, showing significant improvement across various domains and datasets.", "keywords": [ "time series forecasting", "retrieval augmented generation", "time series foundation model" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "zd0iX5xBhA", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Archiki Prasad", "gender": "Female", "institution": "University of North Carolina, Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Elias Stengel-Eskin", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Postdoc" }, { "name": "Mohit Bansal", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" }, { "name": "Peter Hase", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Chih Yao Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Swarnadeep Saha", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 6, "n_ref_uni": 34, "n_ref": 92, "n_ref_all": 125, "n_fig": 7, "n_tab": 8, "L_tab": 10509, "n_element_tab": 199, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 104, "n_element_tab_1": 5, "formula_len_all": 946, "formula_len_all_1": 273, "len_all": 190374, "len_all_1": 78835, "len_abs": 2185, "len_title": 123, "len_sents": 58772, "len_sents_1": 37237, "n_sents": 510, "n_sents_1": 351, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 2193, "title": "System 1.x: Learning to Balance Fast and Slow Planning with Language Models", "abs": "Language models can be used to solve long-horizon planning problems in two distinct modes. In a fast 'System-1' mode, models directly generate plans without any explicit search or backtracking, and in a slow 'System-2' mode, they plan step-by-step by explicitly searching over possible actions. System-2 planning, while typically more effective, is also computationally more expensive and often infeasible for long plans or large action spaces. Moreover, isolated System-1 or System-2 planning ignores the user's end goals and constraints (e.g., token budget), failing to provide ways for the user to control the model's behavior. To this end, we propose the System-1.x Planner, a framework for controllable planning with language models that is capable of generating hybrid plans and balancing between the two planning modes based on the difficulty of the problem at hand. System-1.x consists of (i) a controller, (ii) a System-1 Planner, and (iii) a System-2 Planner. Based on a user-specified hybridization factor x governing the degree to which the system uses System-1 vs. System-2, the controller decomposes a planning problem into subgoals, and classifies them as easy or hard to be solved by either System-1 or System-2, respectively. We fine-tune all three components on top of a single base LLM, requiring only search traces as supervision. Experiments with two diverse planning tasks -- Maze Navigation and Blocksworld -- show that our System-1.x Planner outperforms a System-1 Planner, a System-2 Planner trained to approximate A* search, and also a symbolic planner (A* search), given an exploration budget. We also demonstrate the following key properties of our planner: (1) controllability: by adjusting the hybridization factor x (e.g., System-1.75 vs. System-1.5) we can perform more (or less) search, improving performance, (2) flexibility: by building a neuro-symbolic variant composed of a neural System-1 planner and a symbolic System-2 planner, we can take advantage of existing symbolic methods, and (3) generalizability: by learning from different search algorithms (BFS, DFS, A*), we show that our method is robust to the choice of search algorithm used for training.", "keywords": [ "Large Language Models", "Planning" ], "rating_list": [ 1, 3, 6, 5 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "zcx6rIMbbR", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Andrei Simion", "gender": "Male", "institution": "Columbia University", "country": "", "position": "Lecturer" }, { "name": "Changhai Zhou", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Qian Qiao", "gender": "Male", "institution": "Suzhou University", "country": "CN", "position": "MS student" }, { "name": "Shijie Han", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "MS student" }, { "name": "Shiyang Zhang", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Undergrad student" }, { "name": "Yuhua Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Weizhong Zhang", "gender": "Male", "institution": "Fudan University", "country": "HK", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 27, "n_ref": 43, "n_ref_all": 61, "n_fig": 7, "n_tab": 3, "L_tab": 3135, "n_element_tab": 351, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2009, "n_element_tab_1": 167, "formula_len_all": 1165, "formula_len_all_1": 888, "len_all": 162816, "len_all_1": 69572, "len_abs": 3852, "len_title": 116, "len_sents": 51714, "len_sents_1": 31755, "n_sents": 346, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1479, "title": "Efficient Fine-Tuning of Quantized LLMs via Three-Stage Optimization", "abs": "To address the memory consumption and computational efficiency issues in fine-tuning large language models (LLMs), Parameter-Efficient Fine-Tuning (PEFT) and quantization have emerged. Recent studies have combined the two and have proposed adjusting parameters before fine-tuning to reduce quantization errors, aiming to improve fine-tuning performance. We find that the performance of fine-tuning on the adjusted quantized models is even worse than using the original quantized models directly, as the adjusted model is essentially a completely different model from the original quantized model. Additionally, we have discovered that due to the poor robustness of quantized models, increasing the training difficulty may result in even worse outcomes. To address this, we propose two constraints for fine-tuning quantized models, and based on these, we introduce a general fine-tuning framework called QR-Adaptor. This framework bypasses the network errors introduced by quantization and directly uses actual performance and memory as optimization targets. Through initialization, extrapolation, and interpolation, it quickly solves this gradient-free optimization problem. Experimental results demonstrate that our method yields fine-tuned low-bit quantized models that outperform fine-tuned 16-bit models while maintaining the same memory usage as fine-tuning 4-bit models. For example, in the zero-shot test on MMLU, it improves accuracy by 3.3\\% over both LoftQ and LQ-LoRA.", "keywords": [ "Efficient Fine-Tuning", "NLP", "Iterative Optimization", "Layer-wise Quantization and Low-Rank Configuration" ], "rating_list": [ 8, 5, 6, 3, 5 ], "soundness_list": [ 4, 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "zcTLpIfj9u", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Alejandro Lozano", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Curtis Langlotz", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Ethan Steinberg", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Researcher" }, { "name": "Jason Alan Fries", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Research Scientist" }, { "name": "Jeya Maria Jose Valanarasu", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Researcher" }, { "name": "Louis Blankemeier", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Nigam Shah", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Akshay Chaudhari", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Zepeng Huo", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" } ], "n_formula": 8, "n_formula_1": 2, "n_ref_uni": 58, "n_ref": 86, "n_ref_all": 110, "n_fig": 4, "n_tab": 17, "L_tab": 10145, "n_element_tab": 722, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2912, "n_element_tab_1": 65, "formula_len_all": 272, "formula_len_all_1": 98, "len_all": 183509, "len_all_1": 58381, "len_abs": 1347, "len_title": 96, "len_sents": 45494, "len_sents_1": 28352, "n_sents": 310, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1351, "title": "Future-Guided Pretraining via Time-to-Event Supervision for 3D Medical Imaging", "abs": "With the rise of medical foundation models and the growing availability of imaging data, scalable pretraining techniques offer a promising way to identify imaging biomarkers predictive of future disease risk. While current self-supervised methods for 3D medical imaging models capture local structural features like organ morphology, they fail to link pixel biomarkers with long-term health outcomes due to a missing context problem. Current approaches lack the temporal context necessary to identify biomarkers correlated with disease progression, as they rely on supervision derived only from images and concurrent text descriptions. To address this, we introduce time-to-event pretraining, a pretraining framework for 3D medical imaging models that leverages large-scale temporal supervision from paired, longitudinal electronic health records (EHRs). Using a dataset of 18,945 CT scans (4.2 million 2D images) and time-to-event distributions across thousands of EHR-derived tasks, our method improves outcome prediction, achieving an average AUROC increase of 23.7% and a 29.4% gain in Harrell\u2019s C-index across 8 benchmark tasks. Importantly, these gains are achieved without sacrificing diagnostic classification performance. This study lays the foundation for integrating longitudinal EHR and 3D imaging data to advance clinical risk prediction.", "keywords": [ "Multimodal learning", "medical imaging", "Electronic Health Records" ], "rating_list": [ 8, 5, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "zbpzJmRNiZ", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Anton Frederik Thielmann", "gender": "Male", "institution": "BASF", "country": "DE", "position": "Postdoc" }, { "name": "Arik Reuter", "gender": "Male", "institution": "Technische Universit\u00e4t Clausthal", "country": "DE", "position": "Researcher" }, { "name": "Benjamin S\u00e4fken", "gender": "Male", "institution": "Technische Universit\u00e4t Clausthal", "country": "", "position": "Full Professor" } ], "n_formula": 21, "n_formula_1": 12, "n_ref_uni": 39, "n_ref": 113, "n_ref_all": 122, "n_fig": 2, "n_tab": 7, "L_tab": 3284, "n_element_tab": 304, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1650, "n_element_tab_1": 156, "formula_len_all": 2119, "formula_len_all_1": 1098, "len_all": 150130, "len_all_1": 58694, "len_abs": 1410, "len_title": 173, "len_sents": 46175, "len_sents_1": 25106, "n_sents": 377, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 125, "L_abs": 1450, "title": "From Uncontextualized Embeddings to Marginal Feature Effects: Incorporating Intelligibility into Tabular Transformer Networks", "abs": "In recent years, deep neural networks have showcased their predictive power across a variety of tasks. The transformer architecture, originally developed for natural language processing, has also shown great efficiency in handling tabular data, offering a competitive alternative to traditional gradient-boosted decision trees in this domain. However, this predictive power comes at the cost of intelligibility: Marginal feature effects are almost completely lost in the black-box nature of deep tabular transformer networks. Alternative architectures that use the additivity constraints of classical statistical regression models can maintain intelligible marginal feature effects, but often fall short in predictive power compared to their more complex counterparts. To bridge the gap between intelligibility and performance, we propose an adaptation of tabular transformer networks designed to identify marginal feature effects. We provide theoretical justifications that marginal feature effects can be accurately identified, and our ablation study demonstrates that the proposed model efficiently detects these effects, even amidst complex feature interactions. To demonstrate the model's predictive capabilities, we compare it to several interpretable as well as black-box models and find that it can match black-box performances while maintaining intelligibility. The source code is vailable at https://anonymous.4open.science/r/nmfrmr-B086.", "keywords": [ "Tabular Deep Learning", "Interpretability", "Tabular Transformer Networks" ], "rating_list": [ 3, 5, 8, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 4, 2 ], "confidence_list": [ 2, 2, 4, 1 ] }, { "paper_id": "zboCXnuNv7", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Matti Lassas", "gender": "Male", "institution": "University of Helsinki", "country": "", "position": "Full Professor" }, { "name": "S David Mis", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Maarten v. de Hoop", "gender": "unknown", "institution": "Rice University", "country": "US", "position": "Full Professor" } ], "n_formula": 109, "n_formula_1": 26, "n_ref_uni": 46, "n_ref": 74, "n_ref_all": 123, "n_fig": 3, "n_tab": 5, "L_tab": 239, "n_element_tab": 22, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7942, "formula_len_all_1": 2296, "len_all": 244323, "len_all_1": 62074, "len_abs": 1249, "len_title": 108, "len_sents": 84251, "len_sents_1": 26496, "n_sents": 821, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1252, "title": "Semialgebraic Neural Networks: From roots to representations", "abs": "Many numerical algorithms in scientific computing\u2014particularly in areas like numerical linear algebra, PDE simulation, and inverse problems\u2014produce outputs that can be represented by semialgebraic functions; that is, the graph of the computed function can be described by finitely many polynomial equalities and inequalities. In this work, we introduce Semialgebraic Neural Networks (SANNs), a neural network architecture capable of representing any bounded semialgebraic function, and computing such functions up to the accuracy of a numerical ODE solver chosen by the programmer. Conceptually, we encode the graph of the learned function as the kernel of a piecewise polynomial selected from a class of functions whose roots can be evaluated using a particular homotopy continuation method. We show by construction that the SANN architecture is able to execute this continuation method, thus evaluating the learned semialgebraic function. Furthermore, the architecture can exactly represent even discontinuous semialgebraic functions by executing a continuation method on each connected component of the target function. Lastly, we provide example applications of these networks and show they can be trained with traditional deep-learning techniques.", "keywords": [ "deep learning", "semialgebraic functions", "homotopy continuation", "real algebraic geometry", "recurrent neural networks" ], "rating_list": [ 5, 8, 5, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "zbIS2r0t0F", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Alejandro Jimenez Rodriguez", "gender": "Male", "institution": "Sheffield Hallam University", "country": "", "position": "Lecturer" }, { "name": "Alessandro Di Nuovo", "gender": "Male", "institution": "Sheffield Hallam University", "country": "", "position": "Full Professor" }, { "name": "Aung Htet", "gender": "Male", "institution": "Sheffield Hallam University", "country": "GB", "position": "PhD student" }, { "name": "Sarah Hamburg", "gender": "Female", "institution": "Sheffield Hallam University", "country": "", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 23, "n_ref": 38, "n_ref_all": 66, "n_fig": 16, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 16, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 301, "formula_len_all_1": 301, "len_all": 77587, "len_all_1": 56723, "len_abs": 1679, "len_title": 145, "len_sents": 28432, "len_sents_1": 28442, "n_sents": 202, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1694, "title": "Allostatic Control of Persistent States in Spiking Neural Networks for Perception and Computation", "abs": "We introduce a novel model for updating perceptual beliefs about the environment\nby extending the concept of Allostasis to the control of internal representations.\nAllostasis is a fundamental regulatory mechanism observed in animal physiology\nthat orchestrates responses to maintain a dynamic equilibrium in bodily needs and\ninternal states. In this paper, we focus on an application in numerical cognition,\nwhere a bump of activity in an attractor network is used as a spatial-numerical\nrepresentation. While existing neural networks can maintain persistent states, to\ndate, there is no unified framework for dynamically controlling spatial changes in\nneuronal activity in response to enviromental changes. To address this, we couple\na well-known allostatic microcircuit, the Hammel model, with a ring attractor, re-\nsulting in a Spiking Neural Network architecture that can modulate the location of\nthe bump as a function of some reference input. This localised activity in turn is\nused as a perceptual belief in a simulated subitization task \u2013 a quick enumeration\nprocess without counting. We provide a general procedure to fine-tune the model\nand demonstrate the successful control of the bump location. We also study the\nresponse time in the model with respect to changes in parameters and compare\nit with biological data. Finally, we analyze the dynamics of the network to un-\nderstand the selectivity and specificity of different neurons to different categories\npresent in the input. The results of this paper, particularly the mechanism for mov-\ning persistent states, are not limited to numerical cognition but can be applied to a\nwide range of tasks involving similar representations.", "keywords": [ "Allostatic", "Dynamic", "Attractors" ], "rating_list": [ 3, 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 1, 2, 2 ], "presentation_list": [ 2, 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3, 3 ] }, { "paper_id": "zb1UI74kxA", "primary_area": "interpretability and explainable AI", "n_author": 9, "author_info_list": [ { "name": "Arnav Mohanty Das", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Chirag Shah", "gender": "unknown", "institution": "University of Washington", "country": "", "position": "Full Professor" }, { "name": "Gantavya Bhatt", "gender": "Male", "institution": "University of Washington, Seattle", "country": "US", "position": "Graduate Student" }, { "name": "Hannaneh Hajishirzi", "gender": "Female", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "senior director" }, { "name": "Jeff Bilmes", "gender": "Male", "institution": "University of Washington, Seattle", "country": "US", "position": "Full Professor" }, { "name": "Preethi Seshadri", "gender": "Female", "institution": "University of California, Irvine", "country": "", "position": "PhD student" }, { "name": "Royi Rassin", "gender": "Male", "institution": "Research, Google", "country": "IL", "position": "Intern" }, { "name": "Sahil Verma", "gender": "unknown", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Yanai Elazar", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Postdoc" } ], "n_formula": 0, "n_formula_1": 2, "n_ref_uni": 45, "n_ref": 86, "n_ref_all": 150, "n_fig": 36, "n_tab": 16, "L_tab": 90227, "n_element_tab": 341, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 10517, "n_element_tab_1": 50, "formula_len_all": 0, "formula_len_all_1": 110, "len_all": 753528, "len_all_1": 139917, "len_abs": 1619, "len_title": 124, "len_sents": 81458, "len_sents_1": 40066, "n_sents": 626, "n_sents_1": 270, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1524, "title": "How Many Van Goghs Does It Take to Van Gogh? Finding the Imitation Threshold", "abs": "Text-to-image models are trained using large datasets collected by scraping image-text pairs from the internet. These datasets often include private, copyrighted, and licensed material. Training models on such datasets enables them to generate images with such content, which might violate copyright laws and individual privacy. This phenomenon is termed imitation -- generation of images with content that has recognizable similarity to its training images. In this work we study the relationship between a concept's frequency in the training dataset and the ability of a model to imitate it. We seek to determine the point at which a model was trained on enough instances to imitate a concept -- the imitation threshold. We posit this question as a new problem: Finding the Imitation Threshold (FIT) and propose an efficient approach that estimates the imitation threshold without incurring the colossal cost of training multiple models from scratch. We experiment with two domains -- human faces and art styles -- for which we create four datasets, and evaluate three text-to-image models which were trained on two pretraining datasets. Our results estimate that the imitation threshold of these models is in the range of 200-600 images, depending on the domain and the model. The imitation threshold can provide an empirical basis for copyright violation claims and acts as a guiding principle for text-to-image model developers that aim to comply with copyright and privacy laws. Code will be released upon publication.", "keywords": [ "Data Interpretability", "Privacy", "Text-to-Image Models" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "zaxyuX8eqw", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Divyansha Lachi", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Mehdi Azabou", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Postdoc" }, { "name": "Vinam Arora", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Eva Dyer", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 50, "n_ref": 74, "n_ref_all": 107, "n_fig": 15, "n_tab": 6, "L_tab": 6073, "n_element_tab": 664, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 1112, "n_element_tab_1": 27, "formula_len_all": 294, "formula_len_all_1": 241, "len_all": 187546, "len_all_1": 67923, "len_abs": 1463, "len_title": 151, "len_sents": 52758, "len_sents_1": 33375, "n_sents": 344, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 1468, "title": "GraphFM: A generalist graph transformer that learns transferable representations across diverse domains", "abs": "Graph neural networks (GNNs) are often trained on individual datasets, requiring specialized models and significant hyperparameter tuning due to the unique structures and features of each dataset. This approach limits the scalability and generalizability of GNNs, as models must be tailored for each specific graph type. To address these challenges, we introduce GraphFM, a scalable multi-graph pretraining approach designed for learning across diverse graph datasets. GraphFM uses a Perceiver-based encoder with learned latent tokens to compress domain-specific features into a shared latent space, enabling generalization across graph domains. We propose new techniques for scaling up graph training on datasets of different sizes, allowing us to train GraphFM on 152 distinct graph datasets, spanning 7.4 million nodes and 189 million edges. This allows us to study the effect of scale on pretraining across domains such as molecules, citation networks, and product graphs, and show that training on diverse datasets improves performance over single-source pretraining. Our results demonstrate that pretraining on diverse real and synthetic graphs enhances adaptability and stability, leading to competitive performance with state-of-the-art models across various node classification tasks. This approach reduces the burden of dataset-specific training and provides a single generalist model capable of performing across multiple diverse graph structures and tasks.", "keywords": [ "graph transformer", "multi-graph training", "graph foundation model", "node classification" ], "rating_list": [ 3, 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 4, 3 ] }, { "paper_id": "zaoGCGLpux", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Changqing Zou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Chenyang Zhao", "gender": "Male", "institution": "Zhejiang Lab", "country": "", "position": "Researcher" }, { "name": "Hongming Li", "gender": "unknown", "institution": "Zhejiang Lab, Zhejiang Lab", "country": "CN", "position": "Researcher" }, { "name": "\u738b\u6d77\u5e73", "gender": "Male", "institution": "Zhejiang Lab", "country": "CN", "position": "Researcher" }, { "name": "Yixiong wei", "gender": "Male", "institution": "ZheJiangLab", "country": "CN", "position": "Researcher" } ], "n_formula": 35, "n_formula_1": 25, "n_ref_uni": 31, "n_ref": 80, "n_ref_all": 95, "n_fig": 5, "n_tab": 1, "L_tab": 216, "n_element_tab": 27, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3261, "formula_len_all_1": 1763, "len_all": 118052, "len_all_1": 62490, "len_abs": 1634, "len_title": 77, "len_sents": 38366, "len_sents_1": 28200, "n_sents": 306, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 29, "L_abs": 1675, "title": "RETHINK MAXIMUM STATE ENTROPY", "abs": "In the absence of specific tasks or extrinsic reward signals, a key objective for an agent is the efficient exploration of its environment. A widely adopted strategy to achieve this is maximizing state entropy, which encourages the agent to uniformly explore the entire state space. Most existing approaches for maximum state entropy (MaxEnt) are rooted in two foundational approaches, which were proposed by Hazan and Liu \\& Abbeel, respectively. However, a unified perspective on these methods is lacking within the community.\n\nIn this paper, we analyze these two foundational approaches within a unified framework and demonstrate that both methods share the same reward function when employing the $k$NN density estimator. We also show that the $\\eta$-based policy sampling method proposed by Hazan is unnecessary and that the primary distinction between the two lies in the frequency with which the locally stationary reward function is updated. Building on this analysis, we introduce MaxEnt-(V)eritas, which combines the most effective components of both methods: iteratively updating the reward function as defined by Liu \\& Abbeel, and training the agent until convergence before updating the reward functions, akin to the procedure used by Hazan. We prove that MaxEnt-V is an efficient $\\varepsilon$-optimal algorithm for maximizing state entropy, where the tolerance $\\varepsilon$ decreases as the number of iterations increases. Empirical validation in three Mujoco environments shows that MaxEnt-Veritas significantly outperforms the two MaxEnt frameworks in terms of both state coverage and state entropy maximization, with sound explanations for these results.", "keywords": [ "Reinforcement Learning for Exploration", "Maximum Entropy", "Intrinsic Rewards" ], "rating_list": [ 3, 3, 8, 5 ], "soundness_list": [ 1, 4, 3, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 2, 1, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "zaDU4vMAUr", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Haochen Yuan", "gender": "unknown", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "MS student" }, { "name": "Minting Pan", "gender": "unknown", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Xiaokang Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University, China", "country": "CN", "position": "Professor" }, { "name": "Yunbo Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 30, "n_formula_1": 24, "n_ref_uni": 44, "n_ref": 66, "n_ref_all": 90, "n_fig": 6, "n_tab": 15, "L_tab": 4288, "n_element_tab": 429, "n_fig_1": 6, "n_tab_1": 10, "L_tab_1": 2302, "n_element_tab_1": 154, "formula_len_all": 26284, "formula_len_all_1": 20280, "len_all": 187218, "len_all_1": 105609, "len_abs": 1189, "len_title": 125, "len_sents": 46615, "len_sents_1": 35916, "n_sents": 366, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1102, "title": "Bilevel Reinforcement Learning for Stock Data with A Conservative TD Ensemble", "abs": "Reinforcement learning (RL) has shown significant promise in stock trading. A typical solution involves optimizing cumulative returns using historical offline data. However, it may produce less generalizable policies that merely \"memorize\" optimal buying and selling actions from the offline data while neglecting the non-stationary nature of the financial market. We frame stock trading as a specific type of offline RL problem. Our method, MetaTrader, presents two key contributions. First, it introduces a novel bilevel actor-critic method that spans both the original stock data and its transformations. The fundamental idea is that an effective policy should be generalizable across out-of-distribution data. Second, we propose a novel variant of conservative TD learning, utilizing an ensemble-based TD target to mitigate value overestimation, particularly in scenarios with limited offline data. Our empirical findings across two publicly available datasets demonstrate the superior performance of MetaTrader over existing methods, including both RL-based approaches and stock prediction models.", "keywords": [ "Reinforcement learning", "stock markets", "portfolio optimization" ], "rating_list": [ 5, 8, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "zZUCWkn4PL", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Quoc Tran-Dinh", "gender": "Male", "institution": "University of North Carolina, Chapel Hill", "country": "US", "position": "Associate Professor" } ], "n_formula": 54, "n_formula_1": 50, "n_ref_uni": 52, "n_ref": 121, "n_ref_all": 130, "n_fig": 2, "n_tab": 1, "L_tab": 223, "n_element_tab": 1, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 209, "n_element_tab_1": 1, "formula_len_all": 3959, "formula_len_all_1": 3745, "len_all": 153721, "len_all_1": 90712, "len_abs": 1115, "len_title": 119, "len_sents": 33740, "len_sents_1": 29284, "n_sents": 361, "n_sents_1": 328, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1170, "title": "Variance-Reduced Forward-Reflected Algorithms for Generalized Equations", "abs": "We develop two novel stochastic variance-reduction methods to approximate a solution of generalized equations applicable to both equations and inclusions. Our algorithms leverage a new combination of ideas from the forward-reflected-backward splitting method and a class of unbiased variance-reduced estimators. We construct two new stochastic estimators within this class, inspired by the well-known SVRG and SAGA estimators. These estimators significantly differ from existing approaches used in minimax and variational inequality problems. By appropriately selecting parameters, both algorithms achieve the state-of-the-art oracle complexity of $\\mathcal{O}(n +\u00a0n^{2/3} \\epsilon^{-2})$ for obtaining an $\\epsilon$-solution in terms of the operator residual norm, where $n$ represents the number of summands and $\\epsilon$ signifies the desired accuracy. This complexity aligns with the best-known results in SVRG and SAGA methods for stochastic nonconvex optimization. We test our algorithms on two numerical examples and compare them with existing methods. The results demonstrate promising improvements offered by the new methods compared to their competitors.", "keywords": [ "Variance Reduction Method", "SGD", "Generalized Equation", "Variational Inequality", "Minimax Problem" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "zZU69H8tcr", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Chengrui Zhu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Jiateng Wei", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Jingyang Xiang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Jun Chen", "gender": "Male", "institution": "Zhejiang Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Quan Lu", "gender": "Male", "institution": "Mashang Financial Institution", "country": "CN", "position": "Researcher" }, { "name": "Siqi Li", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Yong Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "ning jiang", "gender": "Male", "institution": "Mashang Consumer Finance Co, Ltd", "country": "", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 24, "n_ref": 39, "n_ref_all": 68, "n_fig": 19, "n_tab": 6, "L_tab": 5337, "n_element_tab": 325, "n_fig_1": 15, "n_tab_1": 5, "L_tab_1": 3879, "n_element_tab_1": 254, "formula_len_all": 1236, "formula_len_all_1": 951, "len_all": 115058, "len_all_1": 62516, "len_abs": 1407, "len_title": 119, "len_sents": 27673, "len_sents_1": 25491, "n_sents": 212, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1415, "title": "SparsitySolver: Efficient Reinforcement Learning-based Pruning for LLMs", "abs": "Large Language Models (LLMs) have achieved significant success in the field of Natural Language Processing (NLP). However, due to their large model size and high inference costs, the application of LLMs is restricted. Pruning is regarded as an effective method to reduce the size of LLMs. Mainstream pruning methods for LLMs typically apply a uniform ratio to prune all the layers or determine layerwise sparsity based on simple criteria. Such manually or semi-manually designed pruning strategies often lead to suboptimal results, which makes reinforcement learning a feasible solution. However, current reinforcement learning-based pruning methods usually have redundant environment designs or multiple agents, rendering them ill-suited to massive LLMs. Hence, we propose SparsitySolver, which first incorporates reinforcement learning into the pruning of LLMs, supporting various pruning granularity. SparsitySolver employs an improved reinforcement learning environment, allowing for a rapid pruning strategy search with a small-scale agent. Moreover, to lessen the performance decline caused by structured pruning, we propose a compensation method capable of restoring performance without introducing additional parameters to the model. We evaluate our approach on LLaMA-V1/V2, Mistral, and the OPT families across multiple pruning granularities, achieving performances surpassing the state-of-the-art methods.", "keywords": [ "Large Language Models", "Model Compression" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 2, 1, 3, 1 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "zZ8fgXHkXi", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Ananjan Nandi", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Christopher D. Manning", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "", "position": "Full Professor" }, { "name": "Dan Jurafsky", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Davide Ghilardi", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Intern" }, { "name": "Federico Bianchi", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Postdoc" }, { "name": "Gabriel Poesia", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Anna Darling Goldie", "gender": "Female", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Moussa Doumbouya", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 39, "n_ref": 75, "n_ref_all": 115, "n_fig": 6, "n_tab": 5, "L_tab": 1563, "n_element_tab": 71, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 96, "n_element_tab_1": 5, "formula_len_all": 150, "formula_len_all_1": 150, "len_all": 202077, "len_all_1": 59347, "len_abs": 1925, "len_title": 110, "len_sents": 76257, "len_sents_1": 30773, "n_sents": 438, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 2097, "title": "h4rm3l: A Language for Composable Jailbreak Attack Synthesis", "abs": "Despite their demonstrated valuable capabilities, state-of-the-art (SOTA) widely\ndeployed large language models (LLMs) still cause harm to society due to the inef-\nfectiveness of their safety filters, which can be bypassed by prompt transformations\ncalled jailbreak attacks. Current approaches to LLM safety assessment, which\nemploy datasets of templated prompts and benchmarking pipelines, fail to cover\nsufficiently large and diverse sets of jailbreak attacks, leading to the widespread\ndeployment of unsafe LLMs. Recent research showed that novel jailbreak attacks\ncould be derived by composition, however, a formal composable representation for\njailbreak attacks, which among other benefits could enable the exploration of a\nlarge compositional space of jailbreak attacks through program synthesis methods,\nhas not been previously proposed. We introduce h4rm3l, a novel approach\naddressing this gap with a human-readable domain-specific language (DSL). Our\nframework comprises: (1) The h4rm3l DSL, which formally expresses jailbreak\nattacks as compositions of parameterized string transformation primitives. (2)\nA synthesizer with bandit algorithms that efficiently generates jailbreak attacks\noptimized for a target black box LLM. (3) The h4rm3l red-teaming software\ntoolkit that employs the previous two components and an automated harmful\nLLM behavior classifier that is strongly aligned with human preferences. We\ndemonstrate h4rm3l\u2019s efficacy by synthesizing a dataset of 2656 successful\nnovel jailbreak targeting 6 SOTA open-source and proprietary LLMs (GPT-3.5,\nGPT-4o, Claude-3-sonnet, Claude-3-haiku, Llama3-8b, and Llama3-70b), and\nby benchmarking those models against a subset of the synthesized attacks, and\npreviously published jailbreak attacks which were used as few-shot examples. Our\nresults show that h4rm3l\u2019s synthesized attacks are diverse and more successful\nthan previously reported attacks, with success rates exceeding 90% on SOTA LLMs.\nWarning: This paper and related research artifacts contain offensive and\npotentially disturbing prompts and model-generated content.", "keywords": [ "LLM safety", "program synthesis", "compositional modeling", "jailbreak attacks", "red-teaming", "domain-specific languages", "string transformations", "AI safety research", "black-box optimization", "automated benchmarking" ], "rating_list": [ 8, 5, 8, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "zZ6TT254Np", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Wuyang Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Yixuan Yuan", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Yujiang", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "YufanHU", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 18, "n_ref_uni": 48, "n_ref": 68, "n_ref_all": 87, "n_fig": 6, "n_tab": 3, "L_tab": 878, "n_element_tab": 117, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 651, "n_element_tab_1": 79, "formula_len_all": 1964, "formula_len_all_1": 1363, "len_all": 125472, "len_all_1": 67997, "len_abs": 1749, "len_title": 169, "len_sents": 34204, "len_sents_1": 31527, "n_sents": 231, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 120, "L_abs": 1756, "title": "Synthesizing Realistic fMRI: A Physiological Dynamics-Driven Hierarchical Diffusion Model for Efficient fMRI Acquisition", "abs": "Functional magnetic resonance imaging (fMRI) is essential for mapping brain activity but faces challenges like lengthy acquisition time and sensitivity to patient movement, limiting its clinical and machine learning applications. While generative models such as diffusion models can synthesize fMRI signals to alleviate these issues, they often underperform due to neglecting the brain's complex structural and dynamic properties.\nTo address these limitations, we propose the Physiological Dynamics-Driven Hierarchical Diffusion Model, a novel framework integrating two key brain physiological properties into the diffusion process: brain hierarchical regional interactions and multifractal dynamics. \nTo model complex interactions among brain regions, we construct hypergraphs based on the prior knowledge of brain functional parcellation reflected by resting-state functional connectivity (rsFC). This enables the aggregation of fMRI signals across multiple scales and generates hierarchical signals. \nAdditionally, by incorporating the prediction of two key dynamics properties of fMRI\u2014the multifractal spectrum and generalized Hurst exponent\u2014our framework effectively guides the diffusion process, ensuring the preservation of the scale-invariant characteristics inherent in real fMRI data.\nOur framework employs progressive diffusion generation, with signals representing broader brain region information conditioning those that capture localized details, and unifies multiple inputs during denoising for balanced integration.\nExperiments demonstrate that our model generates physiologically realistic fMRI signals, potentially reducing acquisition time and enhancing data quality, benefiting clinical diagnostics and machine learning in neuroscience.", "keywords": [ "Time Series", "Diffusion" ], "rating_list": [ 8, 6, 5, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "zZ3eYI0QXN", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Lars Kulik", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Full Professor" }, { "name": "Nestor Cabello", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Postdoc" }, { "name": "Shaona Cheng", "gender": "unknown", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 32, "n_ref": 50, "n_ref_all": 69, "n_fig": 3, "n_tab": 6, "L_tab": 2909, "n_element_tab": 327, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1695, "n_element_tab_1": 192, "formula_len_all": 175, "formula_len_all_1": 175, "len_all": 97164, "len_all_1": 62184, "len_abs": 1197, "len_title": 115, "len_sents": 29255, "len_sents_1": 27964, "n_sents": 229, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1169, "title": "Simple, Accurate, and Efficient Axis-Aligned Decision Tree Learning", "abs": "Decision Trees (DTs) are widely used in various domains for their simplicity and interpretability. However, traditional DTs often suffer from low accuracy and reduced robustness because they rely on fixed splits and a greedy approach to decision-making. While recent approaches combining decision trees with optimization seek to balance accuracy, computational efficiency, and interpretability, they still fall short. In this paper, we introduce a novel Probabilistic univariate Decision Tree (ProuDT), a non-greedy, axis-aligned tree that aims to address these challenges and achieve significant improvements. By assigning a single deterministic feature to each decision node, ProuDT ensures univariate splits while preserving the differentiability of soft decision trees for gradient-based optimization. This tree enhances interpretability through transparent feature utilization in decision-making. Additionally, ProuDT simplifies the optimization process and reduces computational cost by avoiding complex parameters. Extensive experiments on tabular datasets demonstrate ProuDT\u2019s superior performance and scalability in binary and multi-class classification tasks.", "keywords": [ "decision tree", "gradient descent", "tabular data" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "zY37C8d6bS", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Lixuan Jin", "gender": "unknown", "institution": "nanjing university", "country": "CN", "position": "Undergrad student" }, { "name": "Ruifeng Chen", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "PhD student" }, { "name": "Tian-Shuo Liu", "gender": "Male", "institution": "Nanjing university", "country": "CN", "position": "MS student" }, { "name": "Xu-Hui Liu", "gender": "unknown", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Zhilong Zhang", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "MS student" }, { "name": "wang pengyuan", "gender": "Male", "institution": "nanjing university", "country": "", "position": "MS student" }, { "name": "Yang Yu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Professor" } ], "n_formula": 35, "n_formula_1": 15, "n_ref_uni": 57, "n_ref": 91, "n_ref_all": 98, "n_fig": 4, "n_tab": 8, "L_tab": 2425, "n_element_tab": 325, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 2138, "n_element_tab_1": 134, "formula_len_all": 2196, "formula_len_all_1": 1314, "len_all": 182363, "len_all_1": 62751, "len_abs": 1632, "len_title": 146, "len_sents": 48147, "len_sents_1": 27452, "n_sents": 383, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1658, "title": "Semantic Skill Extraction via Vision-Language Model Guidance for Efficient Reinforcement Learning", "abs": "Extracting temporally extended skills can significantly improve the efficiency of reinforcement learning (RL) by breaking down complex decision-making problems with sparse rewards into simpler subtasks and enabling more effective credit assignment. However, existing abstraction methods either discover skills in an unsupervised manner, which often lacks semantic information and leads to erroneous or scattered skill extraction results, or require substantial human intervention. In this work, we propose to leverage the extensive knowledge in pretrained Vision-Language Models (VLMs) to progressively guide the latent space after vector quantization to be more semantically meaningful through relabeling each skill. This approach, termed **V**ision-l**an**guage model guided **T**emporal **A**bstraction (**VanTA**), facilitates the discovery of more interpretable and task-relevant temporal segmentations from offline data without the need for extensive manual intervention or heuristics. By leveraging the rich information in VLMs, our method can significantly outperform existing offline RL approaches that depend only on limited training data. From a theory perspective, we demonstrate that stronger internal sequential correlations within each sub-task, induced by VanTA, effectively reduces suboptimality in policy learning. We validate the effectiveness of our approach through extensive experiments on diverse environments, including Franka Kitchen, Minigrid, and Crafter. These experiments show that our method surpasses existing approaches in long-horizon offline reinforcement learning scenarios with both proprioceptive and visual observations.", "keywords": [ "Reinforcement Learning; Vision-Language Models; Temporal Abstraction" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "zXCnIyX9MG", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Anne Churchland", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Daiyao Yi", "gender": "Female", "institution": "Yale University", "country": "", "position": "PhD student" }, { "name": "Hao Dong", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Michael James Higley", "gender": "Male", "institution": "Yale University", "country": "", "position": "Associate Professor" }, { "name": "Shreya Saxena", "gender": "Female", "institution": "Yale University", "country": "", "position": "Assistant Professor" } ], "n_formula": 23, "n_formula_1": 16, "n_ref_uni": 20, "n_ref": 62, "n_ref_all": 103, "n_fig": 23, "n_tab": 3, "L_tab": 813, "n_element_tab": 36, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 253, "n_element_tab_1": 13, "formula_len_all": 1616, "formula_len_all_1": 1049, "len_all": 174378, "len_all_1": 72522, "len_abs": 1880, "len_title": 155, "len_sents": 61274, "len_sents_1": 36187, "n_sents": 420, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 109, "L_abs": 1216, "title": "Shared-AE: Unsupervised Identification of Shared Subspaces in High-dimensional Neural and Behavioral Activity", "abs": "Understanding the relationship between behavior and neural activity is crucial for understanding brain function. One effective method is to learn embeddings for interconnected modalities. For simple behavioral tasks, neural features can be learned based on labels. However, complex behavioral tasks and social behaviors require joint extraction of both behavioral and neural features. In this paper, we present an unsupervised autoencoder (AE) framework, called Shared-AE, which includes a novel regularization term that automatically identifies features shared between neural activity and behavior, while simultaneously capturing the unique private features specific to each modality. We apply Shared-AE, to large-scale neural activity recorded across the entire dorsal cortex of the mouse, during two very different behaviors: (i) head-fixed mice performing a self-initiated decision-making task, and (ii) freely-moving social behavior amongst two mice. Our model successfully captures both 'shared features', shared across the neural and behavioral activity, and 'private features', unique to each modality, significantly enhancing our understanding of the alignment between neural activity and complex behaviors.", "keywords": [ "Computational neuroscience", "Multimodal", "Social behavior" ], "rating_list": [ 8, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "zWYHsbuedA", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Guangyu Shen", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Guanhong Tao", "gender": "unknown", "institution": "University of Utah", "country": "US", "position": "Assistant Professor" }, { "name": "Shengwei An", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Siyuan Cheng", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Xiangyu Zhang", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Full Professor" }, { "name": "Yingqi Liu", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "ZHUO ZHANG", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "Postdoc" }, { "name": "Zhenting Wang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 7, "n_ref_uni": 37, "n_ref": 146, "n_ref_all": 179, "n_fig": 7, "n_tab": 9, "L_tab": 3274, "n_element_tab": 283, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 5243, "n_element_tab_1": 208, "formula_len_all": 1563, "formula_len_all_1": 516, "len_all": 200084, "len_all_1": 79179, "len_abs": 1159, "len_title": 106, "len_sents": 63806, "len_sents_1": 34211, "n_sents": 560, "n_sents_1": 281, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1166, "title": "Combating Hidden Vulnerabilities in Computer Vision Tasks", "abs": "Backdoor attacks are among the most prominent security threats to deep learning models. Traditional backdoors leverage static trigger patterns, such as a red square patch. They can be removed by existing defense techniques.\nHowever, recent backdoor attacks use semantic features as the trigger. Existing techniques largely fall short when facing such backdoors. In this paper, we propose a novel backdoor mitigation technique, MARTINI, that effectively mitigates various backdoors. It features a specially designed trigger reverse-engineering method for constructing backdoor samples that have a similar attack effect as the injected backdoor across a spectrum of attacks. Using the samples derived from MARTINI, paired with the correct labels, in training can remove injected backdoor effects in deep learning models. Our evaluation on 14 types of backdoor attacks in image classification shows that MARTINI can reduce the attack success rate (ASR) from 96.56% to 5.17% on average, outperforming 12 state-of-the-art backdoor removal approaches, which at best reduce the ASR to 26.56%. It can also mitigate backdoors in self-supervised learning and object detection.", "keywords": [ "Computer Vision", "Hidden Vulnerabilities" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "zWASuY0t6o", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Adam Davies", "gender": "Non-Binary", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Alasdair Paren", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Francesco Pinto", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Postdoc" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Tom Andrew Lamb", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 50, "n_formula_1": 14, "n_ref_uni": 49, "n_ref": 97, "n_ref_all": 134, "n_fig": 13, "n_tab": 4, "L_tab": 652, "n_element_tab": 24, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 57962, "formula_len_all_1": 17389, "len_all": 300806, "len_all_1": 91638, "len_abs": 1100, "len_title": 122, "len_sents": 80238, "len_sents_1": 33299, "n_sents": 616, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1084, "title": "Focus On This, Not That! Steering LLMs With Adaptive Feature Specification", "abs": "Despite the success of Instruction Tuning (IT) in training large language models (LLMs) to perform arbitrary user-specified tasks, these models often still leverage spurious or biased features learned from their training data, leading to undesired behaviours when deploying them in new contexts. In this work, we introduce *Focus Instruction Tuning* (FIT), which trains LLMs to condition their responses by ''focusing on'' specific features whilst ignoring others, leading to different behaviours based on which features are specified. Across several experimental settings, we show that focus-tuned models can be adaptively steered by focusing on different features at inference-time, such as (a) improving robustness by focusing on task-causal features and ignoring spurious features, and (b) mitigating bias by ignoring demographic categories. Furthermore, FIT can steer behaviour in new contexts, generalising under distribution shift and to new unseen features at inference time, thereby facilitating more robust, fair, and explainable LLM applications in real-world environments.", "keywords": [ "instruction tuning", "LLMs", "spurious correlations", "robustness", "distribution shift", "bias" ], "rating_list": [ 8, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "zVtwIWyX4S", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Heng Li", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Intern" }, { "name": "Margret Keuper", "gender": "Female", "institution": "Universit\u00e4t Mannheim", "country": "DE", "position": "Full Professor" }, { "name": "Mario Fritz", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Full Professor" }, { "name": "Xudong Yan", "gender": "Male", "institution": "City University of Macao", "country": "CN", "position": "PhD student" }, { "name": "Yuxuan Zhou", "gender": "Male", "institution": "Universit\u00e4t Mannheim", "country": "DE", "position": "PhD student" }, { "name": "Zhi-Qi Cheng", "gender": "Not Specified", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" } ], "n_formula": 40, "n_formula_1": 9, "n_ref_uni": 35, "n_ref": 59, "n_ref_all": 84, "n_fig": 9, "n_tab": 7, "L_tab": 1609, "n_element_tab": 172, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 1480, "n_element_tab_1": 161, "formula_len_all": 2779, "formula_len_all_1": 390, "len_all": 124949, "len_all_1": 54382, "len_abs": 1334, "len_title": 114, "len_sents": 38814, "len_sents_1": 24613, "n_sents": 301, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1359, "title": "MaxSup: Fixing Label Smoothing for Improved Feature Representation", "abs": "Label Smoothing aims to prevent Neural Networks from making over-confident predictions and improve generalization.\nDue to its effectiveness, it has become an indispensable ingredient of the training recipe for tasks such as Image Recognition and Neural Machine Translation. Despite that, previous work shows it encourages an overly tight cluster in the feature space, which `erases' the similarity information of individual examples, resulting in impaired representation learning. By isolating the loss induced by Label Smoothing into a combination of a regularization term and an error-enhancement term, we reveal a previously unknown defect, i.e., it indeed encourages classifiers to be over-confident, when they make incorrect predictions. To remedy this, we present a solution called Max Suppression (MaxSup), which consistently applies the intended regularization effect during training, independent of the correctness of prediction. By visualizing the learned features, we show that MaxSup successfully enlarges intra-class variations, while improving the inter-class separability. We further conduct experiments on Image Classification and Machine Translation tasks, validating the superiority of Max Suppression. The code implementation is available at [anonymous repository](https://anonymous.4open.science/r/Maximum-Suppression-Regularization-DB0C).", "keywords": [ "Label Smoothing", "Regularization", "Representation Learning", "Explainability" ], "rating_list": [ 5, 5, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "zVp0TVDkrX", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Lunting Fan", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Senior Operations and Maintenance Director" }, { "name": "Roger Zimmermann", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Yingying ZHANG", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yutong Xia", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yuxuan Liang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Qingsong Wen", "gender": "Male", "institution": "Squirrel Ai Learning", "country": "US", "position": "Principal Researcher" } ], "n_formula": 18, "n_formula_1": 14, "n_ref_uni": 65, "n_ref": 134, "n_ref_all": 171, "n_fig": 19, "n_tab": 6, "L_tab": 2420, "n_element_tab": 205, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 1592, "n_element_tab_1": 59, "formula_len_all": 1150, "formula_len_all_1": 733, "len_all": 195127, "len_all_1": 66987, "len_abs": 2160, "len_title": 126, "len_sents": 55409, "len_sents_1": 30315, "n_sents": 422, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1112, "title": "CaPulse: Detecting Anomalies by Tuning in to the Causal Rhythms of Time Series", "abs": "Time series anomaly detection has garnered considerable attention across diverse domains. While existing methods often fail to capture the underlying mechanisms behind anomaly generation in time series data. In addition, time series anomaly detection often faces several data-related inherent challenges, i.e., label scarcity, data imbalance, and complex multi-periodicity. In this paper, we leverage causal tools and introduce a new causality-based framework, **CaPulse**, which *tunes in* to the underlying *causal pulse* of time series data to effectively detect anomalies. Concretely, we begin by building a structural causal model to decipher the generation processes behind anomalies. To tackle the challenges posed by the data, we propose Periodical Normalizing Flows with a novel mask mechanism and carefully designed periodical learners, creating a periodicity-aware, density-based anomaly detection approach. Extensive experiments on seven real-world datasets demonstrate that CaPulse consistently outperforms existing methods, achieving AUROC improvements of 3% to 17%, with enhanced interpretability.", "keywords": [ "Time Series Anomaly Detection" ], "rating_list": [ 8, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 1 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "zVagbJLgkP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Franck Dernoncourt", "gender": "unknown", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Jiuxiang Gu", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Nedim Lipka", "gender": "unknown", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Sungchul Kim", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Tong Yu", "gender": "unknown", "institution": "Adobe Research", "country": "US", "position": "Senior Research Scientist" }, { "name": "Xiang Chen", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Zhehao Zhang", "gender": "Male", "institution": "Dartmouth College", "country": "", "position": "MS student" }, { "name": "Zichao Wang", "gender": "Not Specified", "institution": "Adobe Research", "country": "US", "position": "Research Scientist" }, { "name": "RUIYI ZHANG", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" }, { "name": "Ryan Rossi", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Senior Research Scientist" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 68, "n_ref": 167, "n_ref_all": 208, "n_fig": 4, "n_tab": 9, "L_tab": 4136, "n_element_tab": 110, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 655, "n_element_tab_1": 94, "formula_len_all": 286, "formula_len_all_1": 286, "len_all": 271391, "len_all_1": 68820, "len_abs": 1715, "len_title": 138, "len_sents": 61843, "len_sents_1": 34996, "n_sents": 359, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1721, "title": "VipAct: Visual-Perception Enhancement via Specialized VLM Agent Collaboration and Tool-use", "abs": "While vision-language models (VLMs) have demonstrated remarkable performance across various tasks combining textual and visual information, they continue to struggle with fine-grained visual perception tasks that require detailed pixel-level analysis. Effectively eliciting comprehensive reasoning from VLMs on such intricate visual elements remains an open challenge. In this paper, we present VipAct, an agent framework that enhances VLMs by integrating multi-agent collaboration and vision expert models, enabling more precise visual understanding and comprehensive reasoning. VipAct consists of an orchestrator agent, which manages task requirement analysis, planning, and coordination, along with specialized agents that handle specific tasks such as image captioning and vision expert models that provide high-precision perceptual information. This multi-agent approach allows VLMs to better perform fine-grained visual perception tasks by synergizing planning, reasoning, and tool use. We evaluate VipAct on benchmarks featuring a diverse set of visual perception tasks, with experimental results demonstrating significant performance improvements over state-of-the-art baselines across all tasks. Furthermore, comprehensive ablation studies reveal the critical role of multi-agent collaboration in eliciting more detailed System-2 reasoning and highlight the importance of image input for task planning. Additionally, our error analysis identifies patterns of VLMs' inherent limitations in visual perception, providing insights into potential future improvements. VipAct offers a flexible and extensible framework, paving the way for more advanced visual perception systems across various real-world applications.", "keywords": [ "Visual Language Model", "Visual Perception", "Language Agent" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "zV6D212c7Q", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Alexander S Ecker", "gender": "Male", "institution": "Georg-August Universit\u00e4t G\u00f6ttingen", "country": "DE", "position": "Full Professor" }, { "name": "Timo L\u00fcddecke", "gender": "unknown", "institution": "University of Goettingen", "country": "", "position": "Postdoc" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 46, "n_ref": 58, "n_ref_all": 73, "n_fig": 9, "n_tab": 5, "L_tab": 430, "n_element_tab": 51, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 430, "n_element_tab_1": 51, "formula_len_all": 666, "formula_len_all_1": 666, "len_all": 132924, "len_all_1": 52414, "len_abs": 1236, "len_title": 125, "len_sents": 29152, "len_sents_1": 23600, "n_sents": 243, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1252, "title": "Masked Cross-attention Adapters Enable the Characterization of Dense Features", "abs": "Learning meaningful representations is a core topic of deep learning. Throughout the last decade, many strategies for learning image representations have been proposed involving supervision and self-supervision and various data sources. \nIn most current work, evaluation is focused on classification tasks while neglecting dense prediction tasks, possibly because linear probing is more challenging in the latter case.\nFurthermore, dense prediction heads are often large and come with specific inductive biases that distort performance measurement further.\nIn this work we propose masked cross-attention adapters (MAXA), a minimal adapter method that is capable of dense prediction independent of the size and resolution of the encoder output. This allows us to make dense predictions using a small number of additional parameters ($<0.3 $%) while allowing for fast training using frozen backbones.\nUsing this adapter, we run a comprehensive evaluation assessing instance awareness, local semantics and spatial representation of a diverse set of backbones. \nWe find that DINOv2 outperforms all other backbones tested - including those supervised with masks and language - across all three task categories. \nCode is available at https://to.be.released.", "keywords": [ "image features", "image backbones", "ViT", "instance segmentation" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "zV2cgXk2aY", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Alberto Castellini", "gender": "Male", "institution": "University of Verona", "country": "IT", "position": "Assistant Professor" }, { "name": "Alessandro Farinelli", "gender": "Male", "institution": "Universit\u00e0 degli Studi di Verona", "country": "IT", "position": "Full Professor" }, { "name": "Davide Villaboni", "gender": "Male", "institution": "University of Verona", "country": "IT", "position": "PhD student" }, { "name": "Ivan Luciano Danesi", "gender": "Male", "institution": "Universit\u00e0 Cattolica del Sacro Cuore", "country": "IT", "position": "Instructor" } ], "n_formula": 19, "n_formula_1": 17, "n_ref_uni": 18, "n_ref": 72, "n_ref_all": 85, "n_fig": 3, "n_tab": 3, "L_tab": 285, "n_element_tab": 32, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1646, "n_element_tab_1": 52, "formula_len_all": 1500, "formula_len_all_1": 1500, "len_all": 83447, "len_all_1": 50536, "len_abs": 1095, "len_title": 145, "len_sents": 25653, "len_sents_1": 21830, "n_sents": 191, "n_sents_1": 144, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1099, "title": "Sentinel: Multi-Patch Transformer with Temporal and Channel Attention for Time Series Forecasting", "abs": "Transformer-based time series forecasting has recently gained strong interest due to the ability of transformers to model sequential data. Most of the state-of-the-art architectures exploit either temporal or inter-channel dependencies, limiting their effectiveness in multivariate time-series forecasting where both types of dependencies are crucial. We propose Sentinel, a full transformer-based architecture composed of an encoder able to extract contextual information from the channel dimension, and a decoder designed to capture causal relations and dependencies across the temporal dimension. Additionally, we introduce a multi-patch attention mechanism, which leverages the patching process to structure the input sequence in a way that can be naturally integrated into the transformer architecture, replacing the multi-head splitting process. Extensive experiments on standard benchmarks demonstrate that Sentinel, because of its ability to ``monitor\" both the temporal and the inter-channel dimension, achieves better or comparable performance with respect to state-of-the-art approaches.", "keywords": [ "Transformer", "Time Series Forecasting", "Attention mechanism" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "zUtl4kJa0C", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Fudong Lin", "gender": "Male", "institution": "University of Delaware", "country": "US", "position": "PhD student" }, { "name": "Hao Wang", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Jiadong Lou", "gender": "Male", "institution": "University of Delaware", "country": "US", "position": "PhD student" }, { "name": "Jian Li", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Assistant Professor" }, { "name": "Nian-Feng Tzeng", "gender": "Male", "institution": "University of Louisiana at Lafeyette", "country": "US", "position": "Full Professor" }, { "name": "Xu Yuan", "gender": "Male", "institution": "University of Delaware", "country": "US", "position": "Associate Professor" } ], "n_formula": 50, "n_formula_1": 11, "n_ref_uni": 18, "n_ref": 54, "n_ref_all": 75, "n_fig": 6, "n_tab": 3, "L_tab": 131, "n_element_tab": 4, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 103, "n_element_tab_1": 27, "formula_len_all": 2734, "formula_len_all_1": 576, "len_all": 112521, "len_all_1": 65857, "len_abs": 1528, "len_title": 60, "len_sents": 42782, "len_sents_1": 34546, "n_sents": 286, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1535, "title": "Revisiting Critical Learning Periods in Deep Neural Networks", "abs": "Deep neural networks (DNNs) exhibit critical learning periods (CLPs) during early training phases, when exposure to defective data can permanently impair model performance. The prevalent understanding of such periods, primarily based on the interpretation of Fisher Information (FI), attributes CLPs to the memorization phase. However, our theoretical and empirical study exhibits that such explanations of CLPs are inaccurate because of the misunderstanding of the relationship between FI and model memorization. As such, we revisit the CLPs in DNNs from the information theory and optimization perspectives, gaining a better and more accurate understanding of CLPs. \nWe visualize model memorization dynamics and observe that CLPs extend beyond the memorization phase. Additionally, we introduce the concept of the effective gradient, a novel metric able to quantify the actual influence of each training epoch on the optimization trajectory. Our empirical and theoretical analyses reveal that the norm of effective gradients generally diminishes over training epochs and eventually converges to zero, highlighting the disproportionate larger impact of initial training on final model outcomes. Besides, this insight also clarifies the mechanism behind permanent performance degradation due to defective initial training: the model becomes trapped in the suboptimal region of parameter space. Our work offers novel and in-depth understandings of CLPs and sheds light on enhancing model performance and robustness through such periods.", "keywords": [ "DNN", "Critical Learning Period" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "zUrdd5NRLH", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Yijin Zhou", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yu Guang Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 121, "n_formula_1": 21, "n_ref_uni": 63, "n_ref": 173, "n_ref_all": 201, "n_fig": 5, "n_tab": 11, "L_tab": 5625, "n_element_tab": 353, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 4673, "n_element_tab_1": 294, "formula_len_all": 9495, "formula_len_all_1": 1832, "len_all": 253984, "len_all_1": 66450, "len_abs": 1375, "len_title": 129, "len_sents": 73236, "len_sents_1": 26812, "n_sents": 608, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1382, "title": "GROD: Enhancing Generalization of Transformer with Out-of-Distribution Detection", "abs": "Transformer networks excel in natural language processing (NLP) and computer vision (CV) tasks. However, they face challenges in generalizing to Out-of-Distribution (OOD) datasets, that is, data whose distribution differs from that seen during training. The OOD detection aims to distinguish data that deviates from the expected distribution, while maintaining optimal performance on in-distribution (ID) data. This paper introduces a novel approach based on OOD detection, termed the Generate Rounded OOD Data (GROD) algorithm, which significantly bolsters the generalization performance of transformer networks across various tasks. GROD is motivated by our new OOD detection Probably Approximately Correct (PAC) Theory for transformer. The transformer has learnability in terms of OOD detection that is, when the data is sufficient the outlier can be well represented. By penalizing the misclassification of OOD data within the loss function and generating synthetic outliers, GROD guarantees learnability and refines the decision boundaries between inlier and outlier. This strategy demonstrates robust adaptability and general applicability across different data types. Evaluated across diverse OOD detection tasks in NLP and CV, GROD achieves SOTA regardless of data format. The code is available at https://anonymous.4open.science/r/GROD-OOD-Detection-with-transformers-B70F.", "keywords": [ "OOD detection", "learning theory", "transformer models" ], "rating_list": [ 5, 6, 6, 5, 3 ], "soundness_list": [ 2, 2, 3, 3, 1 ], "presentation_list": [ 2, 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3, 1 ], "confidence_list": [ 4, 3, 4, 4, 5 ] }, { "paper_id": "zUlK1qMIcE", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Christian J Cyron", "gender": "unknown", "institution": "Hamburg University of Technology", "country": "", "position": "Full Professor" }, { "name": "Kevin Linka", "gender": "Male", "institution": "Technische Universit\u00e4t Hamburg", "country": "DE", "position": "Lecturer" }, { "name": "Marius Tacke", "gender": "unknown", "institution": "Helmholtz Zentrum hereon", "country": "DE", "position": "PhD student" }, { "name": "Matthias Busch", "gender": "Male", "institution": "Technische Universit\u00e4t Hamburg", "country": "DE", "position": "PhD student" }, { "name": "Roland Aydin", "gender": "Male", "institution": "Technical University Hamburg", "country": "", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 39, "n_ref": 48, "n_ref_all": 83, "n_fig": 21, "n_tab": 2, "L_tab": 142, "n_element_tab": 7, "n_fig_1": 18, "n_tab_1": 1, "L_tab_1": 27, "n_element_tab_1": 1, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 103082, "len_all_1": 59473, "len_abs": 1698, "len_title": 110, "len_sents": 35941, "len_sents_1": 29343, "n_sents": 267, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1709, "title": "Active partitioning: inverting the paradigm of active learning", "abs": "Datasets often incorporate various functional patterns related to different aspects or regimes, which are typically not equally present throughout the dataset. We propose a novel, general-purpose partitioning algorithm that utilizes competition between models to detect and separate these functional patterns. This competition is induced by multiple models iteratively submitting their predictions for the dataset, with the best prediction for each data point being rewarded with training on that data point. This reward mechanism amplifies each model\u2019s strengths and encourages specialization in different patterns. The specializations can then be translated into a partitioning scheme. The amplification of each model\u2019s strengths inverts the active learning paradigm: while active learning typically focuses the training of models on their weaknesses to minimize the number of required training data points, our concept reinforces the strengths of each model, thus specializing them. We validate our concept -- called active partitioning -- with various datasets with clearly distinct functional patterns, such as mechanical stress and strain data in a porous structure. The active partitioning algorithm produces valuable insights into the datasets\u2019 structure, which can serve various further applications. As a demonstration of one exemplary usage, we set up modular models consisting of multiple expert models, each learning a single partition, and compare their performance on more than twenty popular regression problems with single models learning all partitions simultaneously. Our results show significant improvements, with up to 54% loss reduction, confirming our partitioning algorithm\u2019s utility.", "keywords": [ "Partitioning", "Pattern-recognition", "Clustering", "Active learning", "Modular networks" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "zUXejfUAbx", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Amandin Chyba Rabeendran", "gender": "Male", "institution": "Colorado School of Mines", "country": "", "position": "Undergrad student" }, { "name": "Brandon Knutson", "gender": "Male", "institution": "Colorado School of Mines", "country": "", "position": "PhD student" }, { "name": "Cecilia Diniz Behn", "gender": "unknown", "institution": "Colorado School of Mines", "country": "", "position": "Associate Professor" }, { "name": "Daniel McKenzie", "gender": "unknown", "institution": "Colorado School of Mines", "country": "", "position": "Assistant Professor" }, { "name": "Michael Ivanitskiy", "gender": "Male", "institution": "Colorado School of Mines", "country": "", "position": "PhD student" }, { "name": "Samy Wu Fung", "gender": "Male", "institution": "Colorado School of Mines", "country": "", "position": "Assistant Professor" }, { "name": "Jordan Nikolai Pettyjohn", "gender": "Male", "institution": "Colorado School of Mines", "country": "US", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 37, "n_ref": 104, "n_ref_all": 125, "n_fig": 3, "n_tab": 2, "L_tab": 538, "n_element_tab": 63, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 437, "n_element_tab_1": 52, "formula_len_all": 405, "formula_len_all_1": 228, "len_all": 123889, "len_all_1": 73701, "len_abs": 1267, "len_title": 119, "len_sents": 36065, "len_sents_1": 26217, "n_sents": 286, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1283, "title": "On Logical Extrapolation for Mazes with Recurrent and Implicit Networks", "abs": "Recent work has suggested that certain neural network architectures---particularly recurrent neural networks (RNNs) and implicit neural networks (INNs)--- are capable of _logical extrapolation_. That is, one may train such a network on easy instances of a specific task and then apply it successfully to more difficult instances of the same task. In this paper, we revisit this idea and show that (i) The capacity for extrapolation is less robust than previously suggested. Specifically, in the context of a maze-solving task, we show that while INNs (and some RNNs) are capable of generalizing to larger maze instances, they fail to generalize along axes of difficulty other than maze size. (ii) Models that are explicitly trained to converge to a fixed point (e.g. the INN we test) are likely to do so when extrapolating, while models that are not (e.g. the RNN we test) may exhibit more exotic limiting behaviour such as limit cycles, _even when_ they correctly solve the problem. Our results suggest that (i) further study into _why_ such networks extrapolate easily along certain axes of difficulty yet struggle with others is necessary, and (ii) analyzing the _dynamics_ of extrapolation may yield insights into designing more efficient and interpretable logical extrapolators.", "keywords": [ "implicit networks", "topological data analysis", "logical extrapolation", "out-of-distribution extrapolation", "limit cycles", "dynamics", "mazes." ], "rating_list": [ 5, 3, 3, 5, 3 ], "soundness_list": [ 2, 3, 4, 3, 3 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 2, 1 ], "confidence_list": [ 2, 3, 3, 3, 3 ] }, { "paper_id": "zUD06a6leU", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Aditya Joshi", "gender": "Male", "institution": "UNSW", "country": "AU", "position": "Lecturer" }, { "name": "Duke Nguyen", "gender": "Not Specified", "institution": "University of New South Wales", "country": "AU", "position": "PhD student" }, { "name": "Flora Salim", "gender": "Female", "institution": "University of New South Wales", "country": "AU", "position": "Full Professor" } ], "n_formula": 46, "n_formula_1": 13, "n_ref_uni": 49, "n_ref": 123, "n_ref_all": 165, "n_fig": 2, "n_tab": 8, "L_tab": 1690, "n_element_tab": 189, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 2009, "n_element_tab_1": 150, "formula_len_all": 2964, "formula_len_all_1": 1156, "len_all": 199869, "len_all_1": 60329, "len_abs": 888, "len_title": 113, "len_sents": 52893, "len_sents_1": 26529, "n_sents": 403, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 893, "title": "Spectraformer: A Unified Random Feature Framework for Transformer", "abs": "Linearization of attention using various kernel approximation and kernel learning techniques has shown promise. Past methods use a subset of combinations of component functions and weight matrices within the random features paradigm. We identify the need for a systematic comparison of different combinations of weight matrices and component functions for attention learning in Transformer. In this work, we introduce $\\textit{Spectraformer}$, a unified framework for approximating and learning the kernel function in linearized attention of the Transformer. We experiment with broad classes of component functions and weight matrices for three textual tasks in the LRA benchmark. Our findings indicate that different kernels are good at different tasks and that kernel choice is fundamental to performant models. Our code is available at: https://anonymous.4open.science/r/spectraformer-8A97.", "keywords": [ "linearized attention", "transformer", "efficient transformer", "kernel", "random features" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "zSfeN1uAcx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Acyr Locatelli", "gender": "Male", "institution": "Cohere", "country": "", "position": "Researcher" }, { "name": "Adrien Morisot", "gender": "unknown", "institution": "Cohere", "country": "CA", "position": "Researcher" }, { "name": "Ahmet \u00dcst\u00fcn", "gender": "Male", "institution": "Cohere For AI", "country": "CA", "position": "Researcher" }, { "name": "Ivan Zhang", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "Undergrad student" }, { "name": "Marzieh Fadaee", "gender": "Female", "institution": "Cohere For AI", "country": "NL", "position": "Researcher" }, { "name": "Raymond Ma", "gender": "Male", "institution": "Cohere", "country": "", "position": "Member of Technical Staff" }, { "name": "Viraat Aryabumi", "gender": "Male", "institution": "Cohere For AI", "country": "", "position": "Researcher" }, { "name": "Yixuan Su", "gender": "Male", "institution": "Cohere", "country": "", "position": "Researcher" }, { "name": "Sara Hooker", "gender": "unknown", "institution": "Cohere For AI", "country": "US", "position": "Principal Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 64, "n_ref": 125, "n_ref_all": 151, "n_fig": 9, "n_tab": 5, "L_tab": 1157, "n_element_tab": 116, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 363809, "len_all_1": 60996, "len_abs": 1382, "len_title": 113, "len_sents": 42857, "len_sents_1": 30459, "n_sents": 291, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1387, "title": "To Code or Not To Code? Exploring Impact of Code in Pre-training", "abs": "Including code in the pre-training data mixture, even for models not specifically designed for code, has become a common practice in LLMs pre-training. While there has been anecdotal consensus among practitioners that code data plays a vital role in general LLMs' performance, there is only limited work analyzing the precise impact of code on non-code tasks. In this work, we systematically investigate the impact of code data on general performance. We ask \u201cwhat is the impact of code data used in pre-training on a large variety of downstream tasks beyond code generation\u201d. We conduct extensive ablations and evaluate across a broad range of natural language reasoning tasks, world knowledge tasks, code benchmarks, and LLM-as-a-judge win-rates for models with sizes ranging from 470M to 2.8B parameters. Across settings, we find a consistent results that code is a critical building block for generalization far beyond coding tasks and improvements to code quality have an outsized impact across all tasks. In particular, compared to text-only pre-training, the addition of code results in up to relative increase of 8.2% in natural language (NL) reasoning, 4.2% in world knowledge, 6.6% improvement in generative win-rates, and a 12x boost in code performance respectively. Our work suggests investments in code quality and preserving code during pre-training have positive impacts.", "keywords": [ "code data", "pre-training", "code pre-training" ], "rating_list": [ 8, 3, 6, 6 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "zSUXo1nkqR", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Baochun Li", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Full Professor" }, { "name": "Di Niu", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Full Professor" }, { "name": "Jiuding Yang", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "PhD student" }, { "name": "Shengyao Lu", "gender": "Female", "institution": "University of Alberta", "country": "CA", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 15, "n_ref_uni": 41, "n_ref": 109, "n_ref_all": 139, "n_fig": 8, "n_tab": 5, "L_tab": 1453, "n_element_tab": 179, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1161, "n_element_tab_1": 91, "formula_len_all": 1315, "formula_len_all_1": 934, "len_all": 172942, "len_all_1": 73102, "len_abs": 1512, "len_title": 132, "len_sents": 55922, "len_sents_1": 32382, "n_sents": 502, "n_sents_1": 289, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1362, "title": "TreeX: Generating Global Graphical GNN Explanations via Critical Subtree Extraction", "abs": "The growing demand for transparency and interpretability in critical domains has driven increased interests in comprehending the explainability of Message-Passing (MP) Graph Neural Networks (GNNs). Although substantial research efforts have been made to generate explanations for individual graph instances, identifying global explaining concepts for a GNN still poses great challenges, especially when concepts are desired in a graphical form on the dataset level. While most prior works treat GNNs as black boxes, in this paper, we propose to unbox GNNs by analyzing and extracting critical subtrees incurred by the inner workings of message passing, which correspond to critical subgraphs in the datasets. By aggregating subtrees in an embedding space with an efficient algorithm, which does not require complex subgraph matching or search, we can make intuitive graphical explanations for Message-Passing GNNs on local, class and global levels. We empirically show that our proposed approach not only generates clean subgraph concepts on a dataset level in contrast to existing global explaining methods which generate non-graphical rules (e.g., language or embeddings) as explanations, but it is also capable of providing explanations for individual instances with a comparable or even superior performance as compared to leading local-level GNN explainers.", "keywords": [ "GNN Explainability", "global-level", "XAI", "Explainable AI" ], "rating_list": [ 3, 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 1, 2 ], "confidence_list": [ 4, 4, 3, 5, 4 ] }, { "paper_id": "zSHoaTNlmA", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Serdar Eri\u015fen", "gender": "Male", "institution": "Hacettepe University", "country": "TR", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 31, "n_ref": 73, "n_ref_all": 117, "n_fig": 8, "n_tab": 5, "L_tab": 4956, "n_element_tab": 426, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 3976, "n_element_tab_1": 229, "formula_len_all": 177, "formula_len_all_1": 174, "len_all": 122521, "len_all_1": 57544, "len_abs": 1621, "len_title": 101, "len_sents": 37951, "len_sents_1": 26534, "n_sents": 241, "n_sents_1": 141, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1626, "title": "Segmentation using efficient residual networks with attention-fusion modules", "abs": "Fusing global and local semantic information in segmentation networks remains\nchallenging due to computational costs and the need for effective long-range\nrecognition. Based on the recent success of transformers and attention mechanisms,\nthis research applies attention-based methods of attention-boosting modules\nand attention-fusion networks in enhancing the performance of state-of-the-art\nsegmentation networks, such as InternImage and SERNet-Former, addressing\nthese challenges. Integrating attention-boosting modules into residual networks\ngenerates baseline architectures like Efficient-ResNet, enabling them to extract\nglobal context feature maps in the encoder while minimizing computational costs.\nAttention-based algorithms can also be applied to networks utilizing vision transformers\nand convolutional layers, such as InternImage, to improve the existing\nresults of state-of-the-art networks. In this research, SERNet-Former is deployed\non the challenging benchmarking datasets such as ADE20K, BDD100K, CamVid,\nand Cityscapes by depending on the attention-based methods with new implementations\nof the network, SERNet-Former v2. Our methods have also been implemented\nfor InternImage-XL and improved the test performance of the network on\nthe Cityscapes dataset (85.1 % mean IoU). Respectively, the results of the selected\nnetworks developed by our methods on the challenging benchmarking datasets are\nfound worth considering: 85.1 % mean IoU on the Cityscapes test dataset, 59.35\n% mean IoU on ADE20K validation dataset, 67.42 % mean IoU on BDD100K\nvalidation dataset, and 84.62 % mean IoU on the CamVid dataset.", "keywords": [ "Segmentation", "Attention mechanisms", "Efficient residual networks" ], "rating_list": [ 5, 1, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "zRsFAUQDRk", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Firas Laakom", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" }, { "name": "J\u00fcrgen Schmidhuber", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Director, AI Initiative" }, { "name": "Moncef Gabbouj", "gender": "Male", "institution": "Tampere University", "country": "", "position": "Full Professor" }, { "name": "Yuheng Bu", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "Assistant Professor" } ], "n_formula": 110, "n_formula_1": 21, "n_ref_uni": 53, "n_ref": 118, "n_ref_all": 143, "n_fig": 10, "n_tab": 4, "L_tab": 97, "n_element_tab": 13, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10383, "formula_len_all_1": 1710, "len_all": 196207, "len_all_1": 70931, "len_abs": 1137, "len_title": 114, "len_sents": 63178, "len_sents_1": 34651, "n_sents": 482, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1142, "title": "Class-wise Generalization Error: an Information-Theoretic analysis", "abs": "Existing generalization theories for supervised learning typically take a holistic approach and provide bounds for the expected generalization over the whole data distribution, which implicitly assumes that the model generalizes similarly for all different classes. In practice, however, there are significant variations in generalization performance among different classes, which cannot be captured by the existing generalization bounds. In this work, we tackle this problem by theoretically studying the class-generalization error, which quantifies the generalization performance of the model for each individual class. We derive a novel information-theoretic bound for class-generalization error using the KL divergence, and we further obtain several tighter bounds using recent advances in conditional mutual information bound, which enables practical evaluation. We empirically validate our proposed bounds in various neural networks and show that they accurately capture the complex class-generalization behavior. Moreover, we demonstrate that the theoretical tools developed in this work can be applied in several other applications.", "keywords": [ "information-theoretic bounds", "generalization error", "class-bias" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "zPxlHOLxmh", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Chun Yong Chong", "gender": "Male", "institution": "Monash University", "country": "", "position": "Lecturer" }, { "name": "Mei Kuan Lim", "gender": "Female", "institution": "Monash University, Malaysia Campus", "country": "", "position": "Lecturer" }, { "name": "Julian Yong Hao", "gender": "Male", "institution": "Monash University, Malaysia Campus", "country": "MY", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 13, "n_ref": 35, "n_ref_all": 48, "n_fig": 5, "n_tab": 11, "L_tab": 6521, "n_element_tab": 296, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 813, "n_element_tab_1": 30, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 191761, "len_all_1": 64623, "len_abs": 1613, "len_title": 160, "len_sents": 84395, "len_sents_1": 26917, "n_sents": 835, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1620, "title": "From Counseling Transcript to Mind Map: Leveraging LLMs for Effective Summarization in Mental Health Counseling", "abs": "The increasing number of patients with mental health illness has heightened the cognitive load on therapists, making it challenging for them to provide personalized care that each patient requires. Summarizing counseling sessions can aid mental health practitioners in recalling key details. However, most existing research on summarization focuses primarily on text-based summaries which often require significant cognitive effort to read and interpret. Visual-based summary such as mind maps is proven to help enhance cognitive understanding by giving a quick overview of topics and content. Nevertheless, due to the complex nature of counseling which involves substantial qualitative data, generating visual-based summaries using traditional AI models can be challenging. With the recent advancements in Large Language Models (LLMs), these models have demonstrated the capability to perform tasks based on instructions and generate outputs in various formats. In this study, we develop a web-based summarization tool that serves as a pipeline in performing summarization of counseling transcripts into visual-based mind map summaries using LLMs. We conducted a human evaluation to validate the effectiveness of the generated visual-based summary based on criteria of accuracy, completeness, conciseness and coherence. Our findings show that our web-based summarization tool can effectively extract key points from counseling transcripts and present them in visual-based mind maps, demonstrating its potential in enhancing insights for therapists, ultimately simplifying the process of documenting counseling sessions.", "keywords": [ "Large Language Models", "Visual-based Summarization", "Mental Health Counseling" ], "rating_list": [ 3, 1, 1, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "zPoW8CajCN", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Jeongwhan Choi", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "PhD student" }, { "name": "Noseong Park", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Seungjun Park", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Undergrad student" }, { "name": "Sumin Park", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Sung-Bae Cho", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Full Professor" } ], "n_formula": 46, "n_formula_1": 21, "n_ref_uni": 73, "n_ref": 168, "n_ref_all": 210, "n_fig": 12, "n_tab": 14, "L_tab": 5046, "n_element_tab": 594, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 166, "n_element_tab_1": 13, "formula_len_all": 2425, "formula_len_all_1": 1190, "len_all": 262951, "len_all_1": 66851, "len_abs": 1851, "len_title": 116, "len_sents": 75132, "len_sents_1": 31130, "n_sents": 619, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1184, "title": "Fractal-Inspired Message Passing Neural Networks with Fractal Nodes", "abs": "Graph Neural Networks (GNNs) have emerged as powerful tools for learning on graph-structured data, but they struggle to balance local and global information processing. While graph Transformers aim to address these issues, they often neglect the inherent locality of Message Passing Neural Networks (MPNNs). Inspired by the fractal nature of real-world networks, we propose a novel concept, '*fractal nodes*', that addresses the limitations of both MPNN and graph Transformer. The approach draws insights from renormalization techniques to design a message-passing scheme that captures both local and global structural information. Our method enforces feature self-similarity into nodes by creating fractal nodes that coexist with the original nodes. Fractal nodes adaptively summarize subgraph information and are integrated into MPNN. We show that fractal nodes alleviate an over-squashing problem by providing direct shortcuts to pass fractal information over long distances. Experiments show that our method achieves comparable or better performance to the graph Transformers while maintaining the computational efficiency of MPNN by improving the long-range dependencies of MPNN.", "keywords": [ "graph neural network", "message passing neural network" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "zPaTnGjgpa", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Lawrence Wang", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 32, "n_formula_1": 9, "n_ref_uni": 25, "n_ref": 36, "n_ref_all": 61, "n_fig": 13, "n_tab": 1, "L_tab": 425, "n_element_tab": 6, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3134, "formula_len_all_1": 629, "len_all": 120356, "len_all_1": 61762, "len_abs": 1374, "len_title": 138, "len_sents": 45318, "len_sents_1": 30406, "n_sents": 352, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1383, "title": "Can Stability be Detrimental? Better Generalization through Gradient Descent Instabilities", "abs": "Traditional analyses of gradient descent optimization show that, when the largest eigenvalue of the loss Hessian - often referred to as the sharpness - is below a critical learning-rate threshold, then training is \u2018stable\u2019 and training loss decreases monotonically. Recent studies, however, have suggested that the majority of modern deep neural networks achieve good performance despite operating outside this stable regime. In this work, we demonstrate that such instabilities, induced by large learning rates, move model parameters toward flatter regions of the loss landscape. Our crucial insight lies in noting that, during these instabilities, the orientation of the Hessian eigenvectors rotate. This, we conjecture, allows the model to explore regions of the loss landscape that display more desirable geometrical properties for generalization, such as flatness. These rotations are a consequence of network depth, and we prove that for any network with depth $> 1$, unstable growth in parameters cause rotations in the principal components of the Hessian, which promote exploration of the parameter space away from unstable directions. Our empirical studies reveal an implicit regularization effect in gradient descent with large learning rates operating beyond the stability threshold. We find these lead to excellent generalization performance on modern benchmark datasets.", "keywords": [ "Gradient Descent", "Generalization", "Optimization", "Stability" ], "rating_list": [ 5, 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 1, 3, 2 ], "contribution_list": [ 3, 1, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 2 ] }, { "paper_id": "zPRQ7wtwhb", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Hongfu Liu", "gender": "Male", "institution": "Brandeis University", "country": "US", "position": "Assistant Professor" }, { "name": "Wenxiao Xiao", "gender": "Male", "institution": "Brandeis University", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 96, "n_ref": 160, "n_ref_all": 182, "n_fig": 5, "n_tab": 5, "L_tab": 1921, "n_element_tab": 308, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 766, "n_element_tab_1": 133, "formula_len_all": 206, "formula_len_all_1": 169, "len_all": 185001, "len_all_1": 64929, "len_abs": 1213, "len_title": 92, "len_sents": 47048, "len_sents_1": 31100, "n_sents": 317, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1219, "title": "Salutary Labeling with Zero Human Annotation", "abs": "Active learning strategically selects informative unlabeled data points and queries their ground truth labels for model updates. The prevailing assumption in the active learning paradigm is that the acquisition of ground truth labels optimally enhances model performance. However, this assumption may not always hold or maximize learning capacity. Moreover, ground truth annotations incur significant costs due to the need for intensive human labor. In contrast to traditional active learning, this paper proposes salutary labeling, which automatically assigns the most beneficial labels to the most informative samples without human annotation. Specifically, we utilize the influence function, a tool for estimating sample influence, to select newly added samples and assign their salutary labels by choosing the category that maximizes their positive influence. This process eliminates the need for human annotation. Extensive experiments conducted on nine benchmark datasets demonstrate the superior performance of our salutary labeling approach over traditional active learning strategies. Additionally, we provide several in-depth explorations and practical applications including large language model fine-tuning.", "keywords": [ "Active learning", "influence function" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "zPPy79qKWe", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Gabriel Synnaeve", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Jade Copet", "gender": "unknown", "institution": "Facebook AI Research", "country": "", "position": "Research Engineering Manager" }, { "name": "Jonas Gehring", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" }, { "name": "Kunhao Zheng", "gender": "Male", "institution": "Facebook", "country": "FR", "position": "PhD student" }, { "name": "Taco Cohen", "gender": "Male", "institution": "Meta", "country": "IT", "position": "Researcher" }, { "name": "Vegard Mella", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 41, "n_ref": 69, "n_ref_all": 97, "n_fig": 6, "n_tab": 10, "L_tab": 2450, "n_element_tab": 239, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 2274, "n_element_tab_1": 262, "formula_len_all": 440, "formula_len_all_1": 145, "len_all": 187269, "len_all_1": 68451, "len_abs": 1011, "len_title": 123, "len_sents": 51924, "len_sents_1": 32892, "n_sents": 379, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 876, "title": "RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning", "abs": "Large language models (LLMs) deployed as agents solve user-specified tasks over multiple steps while keeping the required manual engagement to a minimum. Crucially, such LLMs need to ground their generations in any feedback obtained to reliably achieve desired outcomes. We propose an end-to-end reinforcement learning method for teaching models to leverage execution feedback in the realm of code synthesis, where state-of-the-art LLMs struggle to improve code iteratively compared to independent sampling. We benchmark on competitive programming tasks, where we achieve new start-of-the art results with both small (8B parameters) and large (70B) models while reducing the amount of samples required by an order of magnitude. Our analysis of inference-time behavior demonstrates that our method produces LLMs that effectively leverage automatic feedback over multiple steps.", "keywords": [ "Large language models", "automatic code generation", "reinforcement learning", "LLM agents" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "zPHra4V5Mc", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Binghui Li", "gender": "Not Specified", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jian Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Kaifeng Lyu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Zhixuan Pan", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 327, "n_formula_1": 13, "n_ref_uni": 46, "n_ref": 151, "n_ref_all": 195, "n_fig": 18, "n_tab": 3, "L_tab": 1641, "n_element_tab": 108, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 183, "n_element_tab_1": 2, "formula_len_all": 30242, "formula_len_all_1": 774, "len_all": 394887, "len_all_1": 75428, "len_abs": 2673, "len_title": 148, "len_sents": 111313, "len_sents_1": 34522, "n_sents": 1220, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1904, "title": "Feature Averaging: An Implicit Bias of Gradient Descent Leading to Non-Robustness in Neural Networks", "abs": "In this work, we investigate a particular implicit bias in the gradient descent training process, which we term \u201cFeature Averaging\u201d, and argue that it is one of the principal factors contributing to non-robustness of deep neural networks. Despite the existence of multiple discriminative features capable of classifying data, neural networks trained by gradient descent exhibit a tendency to learn the average (or certain combination) of these features, rather than distinguishing and leveraging each feature individually. In particular, we provide a detailed theoretical analysis of the training dynamics of gradient descent in a two-layer ReLU network for a binary classification task, where the data distribution consists of multiple clusters with orthogonal cluster center vectors. We rigorously prove that gradient descent converges to the regime of feature averaging, wherein the weights associated with each hidden-layer neuron represent an average of the cluster centers (each center corresponding to a distinct feature). It leads the network classifier to be non-robust due to an attack that aligns with the negative direction of the averaged features. Furthermore, we prove that, with the provision of more granular supervised information, a two-layer multi-class neural network is capable of learning individual features, which is able to induce a binary classifier with the optimal robustness under our setting. Besides, we also conduct extensive experiments using synthetic datasets, MNIST and CIFAR-10 to substantiate the phenomenon of feature averaging and its role in adversarial robustness of neural networks. We hope the theoretical and empirical insights can provide a deeper understanding of the impact of the gradient descent training on feature learning process, which in turn influences the robustness of the network, and how more detailed supervision may enhance model robustness.", "keywords": [ "deep learning theory", "feature learning", "adversarial robustness", "implicit bias" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 3, 2, 4 ] }, { "paper_id": "zPDpdk3V8L", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Tao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaoying Tang", "gender": "Female", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Yongxin Guo", "gender": "Male", "institution": "Chinese University of HongKong, Shenzhen", "country": "CN", "position": "PhD student" } ], "n_formula": 312, "n_formula_1": 12, "n_ref_uni": 43, "n_ref": 191, "n_ref_all": 255, "n_fig": 7, "n_tab": 8, "L_tab": 1878, "n_element_tab": 107, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 478, "n_element_tab_1": 1, "formula_len_all": 22862, "formula_len_all_1": 1445, "len_all": 249362, "len_all_1": 68537, "len_abs": 1256, "len_title": 142, "len_sents": 63268, "len_sents_1": 30280, "n_sents": 655, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1172, "title": "Enhancing Clustered Federated Learning: Integration of Strategies and Improved Methodologies", "abs": "Federated Learning (FL) is an evolving distributed machine learning approach that safeguards client privacy by keeping data on edge devices. However, the variation in data among clients poses challenges in training models that excel across all local distributions. Recent studies suggest clustering as a solution to address client heterogeneity in FL by grouping clients with distribution shifts into distinct clusters. Nonetheless, the diverse learning frameworks used in current clustered FL methods create difficulties in integrating these methods, leveraging their advantages, and making further enhancements. \nTo this end, this paper conducts a thorough examination of existing clustered FL methods and introduces a four-tier framework, named HCFL, to encompass and extend the existing approaches. Utilizing the HCFL, we identify persistent challenges associated with current clustering methods in each tier and propose an enhanced clustering method called HCFL$^{+}$ to overcome these challenges. Through extensive numerical evaluations, we demonstrate the effectiveness of our clustering framework and the enhanced components. Our code will be publicly accessible.", "keywords": [ "Federated Learning", "Clustering" ], "rating_list": [ 6, 5, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 2, 3, 3 ] }, { "paper_id": "zP8HygcAMY", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Guilin Qi", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Full Professor" }, { "name": "Hongru WANG", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jeff Z. Pan", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Full Professor" }, { "name": "Jiaoyan Chen", "gender": "Male", "institution": "Department of Computer Science, The University of Manchester", "country": "", "position": "Lecturer" }, { "name": "Nan Hu", "gender": "Male", "institution": "Southeast University", "country": "", "position": "PhD student" }, { "name": "Tongtong Wu", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Postdoc" }, { "name": "Yike Wu", "gender": "Male", "institution": "Southeast University", "country": "", "position": "MS student" }, { "name": "Yongrui Chen", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Postdoc" }, { "name": "bisheng", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Lecturer" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 34, "n_ref": 74, "n_ref_all": 107, "n_fig": 6, "n_tab": 11, "L_tab": 4417, "n_element_tab": 308, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1037, "n_element_tab_1": 93, "formula_len_all": 229, "formula_len_all_1": 0, "len_all": 178249, "len_all_1": 68404, "len_abs": 1925, "len_title": 139, "len_sents": 52510, "len_sents_1": 32495, "n_sents": 407, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1485, "title": "Can LLMs Evaluate Complex Attribution in QA? Automatic Benchmarking Using Knowledge Graphs", "abs": "The attribution of question answering (QA), which is to get evidences for supporting the generated answer, has attracted wide research attention. The current methods for automatically evaluating the attribution, typically relying on Large Language Models (LLMs), are still inadequate, particularly in recognizing subtle differences between attributions, and in measuring complex attribution reasoning. Existing benchmarks, which are primarily based on manual annotations, suffer from limited evaluation settings with incomplete and coarse attribution categories and reasoning scenarios, hindering the evaluation and advancement of attribution evaluators. To address this gap, we introduce Complex Attributed Question Answering (CAQA), a large-scale benchmark automatically generated using Knowledge Graphs (KGs), containing more comprehensive attribution categories and complex attribution reasoning scenarios. Our experiments with two specifically developed evaluators and nine LLM evaluators reveal that they struggle in identifying negative attribution categories and handling complex attribution reasoning in both zero-shot and few-shot settings, but mostly perform relatively well in the fine-tuning setting. Moreover, all evaluators perform inadequately in fine-grained attribution identification scenarios. The experiments also demonstrate that CAQA is consistent with human annotations, and is promising for selecting and developing more effective attribution evaluators in QA.", "keywords": [ "Large Language Model", "Attributed Question Answering", "Knowledge Graph" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "zONMuIVCAT", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Alex Su", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Chenlu Ding", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Jiancan Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Postdoc" }, { "name": "Jinda Lu", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Kai Zhang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Xiang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xiangnan He", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Yancheng Yuan", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 27, "n_formula_1": 21, "n_ref_uni": 65, "n_ref": 117, "n_ref_all": 146, "n_fig": 3, "n_tab": 12, "L_tab": 3456, "n_element_tab": 373, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2949, "n_element_tab_1": 206, "formula_len_all": 1828, "formula_len_all_1": 1398, "len_all": 223305, "len_all_1": 68482, "len_abs": 1137, "len_title": 95, "len_sents": 54295, "len_sents_1": 28228, "n_sents": 414, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1140, "title": "Unified Parameter-Efficient Unlearning for LLMs", "abs": "The advent of Large Language Models (LLMs) has revolutionized natural language processing, enabling advanced understanding and reasoning capabilities across a variety of tasks. Fine-tuning these models for specific domains, particularly through Parameter-Efficient Fine-Tuning (PEFT) strategies like LoRA, has become a prevalent practice due to its efficiency. However, this raises significant privacy and security concerns, as models may inadvertently retain and disseminate sensitive or undesirable information. To address these issues, we introduce a novel instance-wise unlearning framework, LLMEraser, which systematically categorizes unlearning tasks and applies precise parameter adjustments using influence functions. Unlike traditional unlearning techniques that are often limited in scope and require extensive retraining, LLMEraser is designed to handle a broad spectrum of unlearning tasks without compromising model performance. Extensive experiments on benchmark datasets demonstrate that LLMEraser excels in efficiently managing various unlearning scenarios while maintaining the overall integrity and efficacy of the models.", "keywords": [ "Large Language Model Unlearning; Machine Unlearning; Influence Function" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "zOMa82W1HV", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chang D. Yoo", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Eunseop Yoon", "gender": "Female", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Hee Suk Yoon", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Hyeonmok Ko", "gender": "Male", "institution": "Samsung Research", "country": "", "position": "Principal Researcher" }, { "name": "Mark A. Hasegawa-Johnson", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Full Professor" }, { "name": "SooHwan Eom", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Jun Yeop Shim", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 54, "n_ref": 118, "n_ref_all": 138, "n_fig": 3, "n_tab": 11, "L_tab": 2372, "n_element_tab": 182, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1371, "n_element_tab_1": 90, "formula_len_all": 227, "formula_len_all_1": 227, "len_all": 176810, "len_all_1": 55294, "len_abs": 953, "len_title": 135, "len_sents": 45394, "len_sents_1": 25602, "n_sents": 325, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 956, "title": "SQuBa: Speech Mamba Language Model with Querying-Attention for Efficient Summarization", "abs": "Abstractive Speech Summarization (SSum) becomes increasingly difficult as the input speech length grows. To address this, we present SQuBa (Speech Querying Mamba Architecture), an end-to-end model designed explicitly for efficient speech summarization. SQuBa leverages a querying-attention Mamba projector to condense extended acoustic features into compact semantic tokens, which are subsequently summarized by the Mamba Large Language Model (LLM). The architecture\u2019s computational complexity scales linearly with input length, enabling efficient handling of longer inputs. A two-stage training framework, complemented by bootstrapped Direct Preference Optimization (DPO) fine-tuning, empowers SQuBa to generate concise and coherent summaries. Experimental results demonstrate that SQuBa delivers competitive performance while significantly improving inference speed, making it ideal for real-world applications such as podcast and meeting transcriptions.", "keywords": [ "Summarisation", "Speech", "Mamba" ], "rating_list": [ 3, 3, 6, 5, 5 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 5, 5, 3, 4, 3 ] }, { "paper_id": "zNfdtV9ADQ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Abdallah Chehade", "gender": "Male", "institution": "University of Michigan - Dearborn", "country": "US", "position": "Associate Professor" }, { "name": "Wael Hassanieh", "gender": "unknown", "institution": "University of Michigan - Dearborn", "country": "US", "position": "PhD student" } ], "n_formula": 59, "n_formula_1": 34, "n_ref_uni": 37, "n_ref": 55, "n_ref_all": 82, "n_fig": 3, "n_tab": 5, "L_tab": 930, "n_element_tab": 121, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 82, "n_element_tab_1": 11, "formula_len_all": 4048, "formula_len_all_1": 2042, "len_all": 127196, "len_all_1": 61628, "len_abs": 1648, "len_title": 139, "len_sents": 34115, "len_sents_1": 25354, "n_sents": 297, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1869, "title": "SRPCA: Sparse Reverse of Principal Component Analysis for Fast Low-Rank Matrix Completion", "abs": "Supervised and unsupervised learning methods experience a decline in performance when applied to incomplete, corrupted, or noisy datasets. Matrix completion is a common task to impute the missing values in sparsely observed matrices. Given a matrix $\\mathbf{X} \\in \\mathbb{R}^{m \\times n}$, low-rank matrix completion computes a rank-$r$ approximation of $\\mathbf{X}$, where $r\\ll\\min\\\\{m,n\\\\}$, by only observing a few random entries of $\\mathbf{X}$. It is commonly applied for recommender systems, image processing, and multi-output collaborative modeling. Existing matrix completion methods suffer either from slow convergence or failure under significant missing data levels. \nThis paper proposes a novel approach, the Sparse Reverse of Principal Component Analysis (SRPCA), that reformulates matrix factorization based low-rank completion $(\\min_{\\mathbf{U},\\mathbf{V}}\\Vert\\mathcal{P}_{\\mathbf{\\Omega}}(\\mathbf{X}-\\mathbf{U}\\mathbf{V}^T)\\Vert_F^2)$\nto iteratively learn a single low-rank subspace representation by solving the convex optimization problem\n$\\min\\_{\\mathbf{V}}\\Vert\\mathcal{P}\\_{\\mathbf{\\Omega}}(\\mathbf{X}-\\mathbf{P}\\mathbf{V}^T)\\Vert_F^2$ under the principal component analysis framework, resulting in a significant convergence acceleration. SRPCA converges iteratively and is computationally tractable with a proven controllable upper bound on the number of iterations until convergence. Unlike existing matrix completion algorithms, the proposed SRPCA applies iterative pre-processing resets that maintain smoothness across the reconstructed matrix, which results in a performance boost for smooth matrices. The performance of the proposed technique is validated on case studies for image processing, multivariate time-series imputation, and collaborative filtering. SRPCA is also compared with state-of-the-art benchmarks for matrix completion.", "keywords": [ "matrix completion", "low rank", "PCA", "collaborative filtering", "image inpainting", "time-series imputation" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "zNVefjN3EP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Xiao-Ping Zhang", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xinchun Yu", "gender": "Male", "institution": "Zhejiang Gongshang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yansong Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yinghao Wu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Zhaojian Yu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Zhuotao Deng", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 50, "n_ref": 67, "n_ref_all": 88, "n_fig": 4, "n_tab": 3, "L_tab": 1099, "n_element_tab": 135, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1099, "n_element_tab_1": 135, "formula_len_all": 233, "formula_len_all_1": 233, "len_all": 132792, "len_all_1": 49754, "len_abs": 1391, "len_title": 127, "len_sents": 26959, "len_sents_1": 23155, "n_sents": 181, "n_sents_1": 155, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1401, "title": "OpenCarbonEval: How much $CO_2$ will your large model exhale in training process?", "abs": "Data, model and hardware are crucial components in the development of large scale machine learning models. The training of such models necessitates substantial computational resources, energy consumption, and raw materials, resulting in significant environmental implications. However, the environmental impact of these models has been largely overlooked due to a lack of assessment and analysis of their carbon footprint. In this paper, we present OpenCarbonEval, a carbon emission estimation framework to quantify the environmental implications of large scale machine learning models given their total training computations and hardware configurations.\nIn OpenCarbonEval, we conducted a comprehensive dynamic analysis of the interrelationships among data, models, and hardware throughout the model training process, aiming to forecast the carbon emission of large scale models more accurately. We validated our approach on real-world dataset, and experimental results demonstrate that OpenCarbonEval can predict energy costs and carbon emissions more accurately than previous methods. Furthermore, it can be seamlessly applied to various machine learning tasks without a precision decline. By quantifying the environmental impact of large-scale models, OpenCarbonEval promotes sustainable AI development and deployment, contributing to a more environmentally responsible future for the AI community.", "keywords": [ "Large-scale model", "Carbon footprint", "Sustainable AI" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 1, 1, 1 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 1, 4, 2 ] }, { "paper_id": "zMjjzXxS64", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Adam Kania", "gender": "Not Specified", "institution": "Jagiellonian University Cracow", "country": "PL", "position": "MS student" }, { "name": "Jacek Tabor", "gender": "Male", "institution": "Jagiellonian University", "country": "PL", "position": "Full Professor" }, { "name": "Marko Mihajlovic", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" }, { "name": "Przemys\u0142aw Spurek", "gender": "Male", "institution": "Jagiellonian University Cracow", "country": "PL", "position": "Associate Professor" }, { "name": "Sergey Prokudin", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Senior Scientist" } ], "n_formula": 14, "n_formula_1": 7, "n_ref_uni": 47, "n_ref": 111, "n_ref_all": 158, "n_fig": 12, "n_tab": 8, "L_tab": 5332, "n_element_tab": 375, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 987, "n_element_tab_1": 68, "formula_len_all": 699, "formula_len_all_1": 387, "len_all": 167423, "len_all_1": 67312, "len_abs": 1405, "len_title": 120, "len_sents": 48236, "len_sents_1": 30920, "n_sents": 350, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1409, "title": "FreSh: Frequency Shifting for Accelerated Neural Representation Learning", "abs": "Implicit Neural Representations (INRs) have recently gained attention as a powerful approach for continuously representing signals such as images, videos, and 3D shapes using multilayer perceptrons (MLPs). However, MLPs are known to exhibit a low-frequency bias, limiting their ability to capture high-frequency details accurately. This limitation is typically addressed by incorporating high-frequency input embeddings or specialized activation layers. In this work, we demonstrate that these embeddings and activations are often configured with hyperparameters that perform well on average but are suboptimal for specific input signals under consideration, necessitating a costly grid search to identify optimal settings. Our key observation is that the initial frequency spectrum of an untrained model's output correlates strongly with the model's eventual performance on a given target signal. Leveraging this insight, we propose frequency shifting (or FreSh), a method that selects embedding hyperparameters to align the frequency spectrum of the model\u2019s initial output with that of the target signal. We show that this simple initialization technique improves performance across various neural representation methods and tasks, achieving results comparable to extensive hyperparameter sweeps but with only marginal computational overhead compared to training a single model with default hyperparameters.", "keywords": [ "spectral bias", "automatic hyperparameter selection", "implicit neural representation", "discrete fourier transform" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 4, 2, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "zM92zziRtQ", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Chen Chen", "gender": "Male", "institution": "OPPO AI Center", "country": "CN", "position": "Researcher" }, { "name": "Haonan Lu", "gender": "Male", "institution": "OPPO Guangdong Mobile Telecommunications Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Qingsong Xie", "gender": "Male", "institution": "OPPO", "country": "CN", "position": "Researcher" }, { "name": "SHIXIANG TANG", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Zhenyi Liao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Zhijie Deng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 28, "n_ref": 87, "n_ref_all": 100, "n_fig": 4, "n_tab": 4, "L_tab": 683, "n_element_tab": 87, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 565, "n_element_tab_1": 73, "formula_len_all": 976, "formula_len_all_1": 925, "len_all": 111372, "len_all_1": 51928, "len_abs": 350, "len_title": 133, "len_sents": 30132, "len_sents_1": 22936, "n_sents": 242, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1811, "title": "TLCM: Training- efficient Latent Consistency Model for Image Generation with 2-8 Steps", "abs": "Distilling latent diffusion models (LDMs) into ones that are fast to sample from is attracting growing research interest. However, the majority of existing methods face two critical challenges: \n1) They need to perform long-time learning with a huge volume of real data. \n2) They routinely lead to quality degradation for generation, especially in text-image alignment. \n\nThis paper proposes the novel Training-efficient Latent Consistency Model (TLCM) to overcome these challenges. \nOur method first fast accelerate LDMs via data-free multistep latent consistency distillation (MLCD), then data-free latent consistency distillation is proposed to guarantee the inter-segment consistency in MLCD at low cost. \nFurthermore, we introduce bags of techniques to enhance TLCM's performance at rare-step inference without any real data, e.g., distribution matching, adversarial learning, and preference learning. \nTLCM demonstrates a high level of flexibility by allowing for adjustment of sampling steps within the range of 2 to 8 while still producing competitive outputs compared to full-step approaches.\nAs its name suggests, TLCM excels in training efficiency in terms of both computational resources and data utilization.\nNotably, TLCM operates without reliance on a training dataset but instead employs synthetic data for the teacher itself during distillation. With just 70 training hours on an A100 GPU, a 3-step TLCM distilled from SDXL achieves an impressive CLIP Score of 33.68 and an Aesthetic Score of 5.97 on the MSCOCO-2017 5K benchmark, surpassing various accelerated models and even outperforming the teacher model in human preference metrics. \nWe also demonstrate the versatility of TLCMs in applications including controllable generation, image style transfer, and Chinese-to-image generation.", "keywords": [ "latent diffusion model", "consistency model", "acceleration" ], "rating_list": [ 3, 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 2, 1 ], "contribution_list": [ 1, 2, 2, 2, 1 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "zLaayPL8f0", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Feng Zhao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "JingHao Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 38, "n_ref_all": 65, "n_fig": 8, "n_tab": 12, "L_tab": 5196, "n_element_tab": 606, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 3803, "n_element_tab_1": 459, "formula_len_all": 1169, "formula_len_all_1": 326, "len_all": 166186, "len_all_1": 64572, "len_abs": 1450, "len_title": 121, "len_sents": 37629, "len_sents_1": 27400, "n_sents": 244, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1454, "title": "Decomposition Ascribed Synergistic Learning for Unified Image Restoration", "abs": "Learning to restore multiple image degradations within a single model is quite beneficial for real-world applications. Nevertheless, existing works typically concentrate on regarding each degradation independently, while their relationship has been less comprehended to ensure the synergistic learning. To this end, we revisit the diverse degradations through the lens of singular value decomposition, with the observation that the decomposed singular vectors and singular values naturally undertake the different types of degradation information, dividing various restoration tasks into two groups, \\ie, singular vector dominated and singular value dominated. The above analysis renders a more unified perspective to ascribe diverse degradation connections, compared to previous task-level independent learning. The dedicated optimization of degraded singular vectors and singular values inherently utilizes the potential partnership among diverse restoration tasks, attributing to the Decomposition Ascribed Synergistic Learning (DASL). Specifically, DASL comprises two effective operators, namely, Singular VEctor Operator (SVEO) and Singular VAlue Operator (SVAO), to favor the decomposed optimization, which can be lightly integrated into existing image restoration backbone. Moreover, the congruous decomposition loss has been devised for auxiliary. Extensive experiments on five image restoration tasks demonstrate the effectiveness of our method.", "keywords": [ "Image Restoration", "Decomposition", "Orthogonality", "Signal formation" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "zLHP6QDWYp", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Chi-Man Pun", "gender": "unknown", "institution": "University of Macau", "country": "MO", "position": "Full Professor" }, { "name": "Lijian Li", "gender": "Male", "institution": "University of Macau", "country": "CN", "position": "MS student" }, { "name": "Wenpin Jiao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Xiancai Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yuanpeng He", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhi Jin", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 18, "n_ref": 30, "n_ref_all": 37, "n_fig": 3, "n_tab": 4, "L_tab": 2483, "n_element_tab": 412, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 3129, "n_element_tab_1": 469, "formula_len_all": 1222, "formula_len_all_1": 1199, "len_all": 140593, "len_all_1": 67890, "len_abs": 1570, "len_title": 119, "len_sents": 33308, "len_sents_1": 27827, "n_sents": 206, "n_sents_1": 165, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1576, "title": "Towards Realistic Long-tailed Semi-supervised Learning in an Open World", "abs": "Open-world long-tailed semi-supervised learning (OLSSL) has increasingly attracted attention. However, existing OLSSL algorithms generally assume that the distributions between known and novel categories are nearly identical. Against this backdrop, we construct a more Realistic Open-world Long-tailed Semi-supervised Learning (ROLSSL) setting where there is no premise on the distribution relationships between known and novel categories. Furthermore, even within the known categories, the number of labeled samples is significantly smaller than that of the unlabeled samples, as acquiring valid annotations is often prohibitively costly in the real world. Under the proposed ROLSSL setting, we propose a simple yet potentially effective solution called dual-stage post-hoc logit adjustments. The proposed approach revisits the logit adjustment strategy by considering the relationships among the frequency of samples, the total number of categories, and the overall size of data. Then, it estimates the distribution of unlabeled data for both known and novel categories to dynamically readjust the corresponding predictive probabilities, effectively mitigating category bias during the learning of known and novel classes with more selective utilization of imbalanced unlabeled data. Extensive experiments on datasets such as CIFAR100 and ImageNet100 have demonstrated performance improvements of up to 50.1%, validating the superiority of our proposed method and establishing a strong baseline for this task. For further researches, the experimental code will be open soon.", "keywords": [ "Open-world", "Realistic long-tailed semi-supervised learning", "Logit adjustment" ], "rating_list": [ 5, 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5, 3 ] }, { "paper_id": "zKvrOOBouT", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Angela Yao", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" }, { "name": "Kai Xu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Qiyuan He", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 44, "n_formula_1": 17, "n_ref_uni": 35, "n_ref": 167, "n_ref_all": 195, "n_fig": 7, "n_tab": 8, "L_tab": 1808, "n_element_tab": 100, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1618, "n_element_tab_1": 78, "formula_len_all": 3339, "formula_len_all_1": 871, "len_all": 170955, "len_all_1": 66642, "len_abs": 977, "len_title": 52, "len_sents": 61195, "len_sents_1": 29574, "n_sents": 485, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 998, "title": "EvA: Erasing Spurious Correlations with Activations", "abs": "Spurious correlations often arise when models associate features strongly correlated with, but not causally related to, the label e.g. an image classifier associates bodies of water with ducks. To mitigate spurious correlations, existing methods focus on learning unbiased representation or incorporating additional information about the correlations during training. This work removes spurious correlations by ``**E**rasing **wi**th **A**ctivations'' (EvA). EvA learns class-specific spurious indicator on each channel for the fully connected layer of pretrained networks. By erasing spurious connections during re-weighting, EvA achieves state-of-the-art performance across diverse datasets (6.2\\% relative gain on BAR and achieves 4.1\\% on Waterbirds). For biased datasets without any information about the spurious correlations, EvA can outperform previous methods (4.8\\% relative gain on Waterbirds) with 6 orders of magnitude less compute, highlighting its data and computational efficiency.", "keywords": [ "spurious correlation", "compute efficiency", "data efficiency" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "zKoUV1wHRJ", "primary_area": "applications to robotics, autonomy, planning", "n_author": 3, "author_info_list": [ { "name": "Haisheng Su", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Wei Wu", "gender": "Male", "institution": "Sensetime", "country": "CN", "position": "Researcher" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 11, "n_ref": 41, "n_ref_all": 55, "n_fig": 4, "n_tab": 8, "L_tab": 7253, "n_element_tab": 369, "n_fig_1": 4, "n_tab_1": 9, "L_tab_1": 6765, "n_element_tab_1": 351, "formula_len_all": 411, "formula_len_all_1": 170, "len_all": 125441, "len_all_1": 73474, "len_abs": 1705, "len_title": 168, "len_sents": 29331, "len_sents_1": 29078, "n_sents": 202, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 119, "L_abs": 1712, "title": "DiFSD: Ego-Centric Fully Sparse Paradigm with Uncertainty Denoising and Iterative Refinement for Efficient Self-Driving", "abs": "Current end-to-end autonomous driving methods resort to unifying modular designs for various tasks (e.g. perception, prediction and planning). Although optimized in a planning-oriented spirit with a fully differentiable framework, existing end-to-end driving systems without ego-centric designs still suffer from unsatisfactory performance and inferior efficiency, owing to the rasterized scene representation learning and redundant information transmission. In this paper, we revisit the human driving behavior and propose an ego-centric fully sparse paradigm, named DiFSD, for end-to-end self-driving. Specifically, DiFSD mainly consists of sparse perception, hierarchical interaction and iterative motion planner. The sparse perception module performs detection, tracking and online mapping based on sparse representation of the driving scene. The hierarchical interaction module aims to select the Closest In-Path Vehicle / Stationary (CIPV / CIPS) from coarse to fine, benefiting from an additional geometric prior. As for the iterative motion planner, both selected interactive agents and ego-vehicle are considered for joint motion prediction, where the output multi-modal ego-trajectories are optimized in an iterative fashion. Besides, both position-level motion diffusion and trajectory-level planning denoising are introduced for uncertainty modeling, thus facilitating the training stability and convergence of the whole framework. Extensive experiments conducted on nuScenes dataset demonstrate the superior planning performance and great efficiency of DiFSD, which significantly reduces the average L2 error by 66% and collision rate by 77% than UniAD while achieves 8.2x faster running efficiency.", "keywords": [ "Autonomous Driving", "End-to-End Fully Sparse Paradigm", "Iterative Refinement", "Uncertainty Denoising" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "zKlFXV87Pp", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Enshu Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xuefei Ning", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Research Assistant Professor" }, { "name": "Yu Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Zinan Lin", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Senior Researcher" } ], "n_formula": 30, "n_formula_1": 31, "n_ref_uni": 34, "n_ref": 87, "n_ref_all": 110, "n_fig": 10, "n_tab": 3, "L_tab": 2497, "n_element_tab": 398, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 436, "n_element_tab_1": 28, "formula_len_all": 2189, "formula_len_all_1": 2032, "len_all": 161075, "len_all_1": 76874, "len_abs": 1787, "len_title": 115, "len_sents": 46328, "len_sents_1": 33681, "n_sents": 429, "n_sents_1": 310, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1813, "title": "Distilled Decoding 1: One-step Sampling of Image Auto-regressive Models with Flow Matching", "abs": "Autoregressive (AR) models have recently achieved state-of-the-art performance in text and image generation. However, their primary limitation is slow generation speed due to the token-by-token process. We ask an ambitious question: can a pre-trained AR model be adapted to generate outputs in just one or two steps? If successful, this would significantly advance the development and deployment of AR models. We notice that existing works that attempt to speed up AR generation by generating multiple tokens at once fundamentally cannot capture the output distribution due to the conditional dependencies between tokens, limiting their effectiveness for few-step generation. To overcome this, we propose Distilled Decoding (DD), which leverages flow matching to create a deterministic mapping from Gaussian distribution to the output distribution of the pre-trained AR model. We then train a network to distill this mapping, enabling few-step generation. The entire training process of DD does not need the training data of the original AR model (as opposed to some other methods), thus making DD more practical. We evaluate DD on state-of-the-art image AR models and present promising results. For VAR, which requires 10-step generation (680 tokens), DD enables one-step generation (6.3$\\times$ speed-up), with an acceptable increase in FID from 4.19 to 9.96. Similarly, for LlamaGen, DD reduces generation from 256 steps to 1, achieving an 217.8$\\times$ speed-up with a comparable FID increase from 4.11 to 11.35. In both cases, baseline methods completely fail with FID scores $>$100. As the first work to demonstrate the possibility of one-step generation for image AR models, DD challenges the prevailing notion that AR models are inherently slow, and opens up new opportunities for efficient AR generation.", "keywords": [ "image autoregressive models", "parallel decoding", "distillation" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "zKFUNRH0hN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Adriano Fragomeni", "gender": "Male", "institution": "University of Bristol", "country": "", "position": "PhD student" }, { "name": "Dima Damen", "gender": "Female", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Michael Wray", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 18, "n_ref": 79, "n_ref_all": 112, "n_fig": 20, "n_tab": 7, "L_tab": 2537, "n_element_tab": 405, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 3616, "n_element_tab_1": 801, "formula_len_all": 1266, "formula_len_all_1": 795, "len_all": 218691, "len_all_1": 87134, "len_abs": 1826, "len_title": 113, "len_sents": 54850, "len_sents_1": 33484, "n_sents": 448, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1434, "title": "Leveraging Modality Tags for Enhanced Cross-Modal Video Retrieval", "abs": "Video retrieval requires aligning visual content with corresponding natural language descriptions. In this paper, we introduce Modality Auxiliary Concepts for Video Retrieval (MAC-VR), a novel approach that leverages modality-specific tags---automatically extracted from foundation models---to enhance video retrieval.\nPrevious works have proposed to emulate human reasoning by introducing latent concepts derived from the features of a video and its corresponding caption. Building on these efforts to align latent concepts across both modalities, we propose learning auxiliary concepts from modality-specific tags. \nWe introduce these auxiliary concepts to improve the alignment of visual and textual latent concepts, and so be able to distinguish each concept from the other.\nTo strengthen the alignment between visual and textual latent concepts\u2014where a set of visual concepts matches a corresponding set of textual concepts\u2014we introduce an Alignment Loss. This loss aligns the proposed auxiliary concepts with the modalities' latent concepts, enhancing the model's ability to accurately match videos with their appropriate captions. \nWe conduct extensive experiments on three diverse datasets: MSR-VTT, DiDeMo, and ActivityNet Captions. The experimental results consistently demonstrate that modality-specific tags significantly improve cross-modal alignment, achieving performance comparable to current state-of-the-art methods.", "keywords": [ "Video Understanding" ], "rating_list": [ 5, 8, 3, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "zJjzNj6QUe", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jianghong Ma", "gender": "Female", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Ruizhi Qiao", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Tianjun Wei", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "Wei Wen", "gender": "Male", "institution": "Tencent YouTu Lab", "country": "CN", "position": "researcher" }, { "name": "Xing Sun", "gender": "Male", "institution": "Tencent YouTu Lab", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 24, "n_ref": 63, "n_ref_all": 96, "n_fig": 17, "n_tab": 8, "L_tab": 7535, "n_element_tab": 261, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 382, "n_element_tab_1": 23, "formula_len_all": 1072, "formula_len_all_1": 178, "len_all": 171186, "len_all_1": 61387, "len_abs": 1804, "len_title": 116, "len_sents": 52158, "len_sents_1": 30282, "n_sents": 414, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1903, "title": "RocketEval: Efficient automated LLM evaluation via grading checklist", "abs": "Evaluating large language models (LLMs) in diverse and challenging scenarios is essential to align them with human preferences. To mitigate the prohibitive costs associated with human evaluations, utilizing a powerful LLM as a judge has emerged as a favored approach. Nevertheless, this methodology encounters several challenges, including substantial expenses, concerns regarding privacy and security, and reproducibility. In this paper, we propose a straightforward, replicable, and accurate automated evaluation method by leveraging a lightweight LLM as the judge, named RocketEval. Initially, we identify that the performance disparity between lightweight and powerful LLMs in evaluation tasks primarily stems from their ability to conduct comprehensive analyses, which is not easily enhanced through techniques such as chain-of-thought reasoning. By reframing the evaluation task as a multi-faceted Q\\&A using an instance-specific checklist, we demonstrate that the limited judgment accuracy of lightweight LLMs is largely attributes to high uncertainty and positional bias. To address these challenges, we introduce an automated evaluation process grounded in checklist grading, which is designed to accommodate a variety of scenarios and questions. This process encompasses the creation of checklists, the grading of these checklists by lightweight LLMs, and the reweighting of checklist items to align with the supervised annotations. Our experiments carried out on the automated evaluation benchmarks, MT-Bench and WildBench datasets, reveal that RocketEval, when using $\\textit{Gemma-2-2B}$ as the judge, achieves a high correlation (0.965) with human preferences, which is comparable to $\\textit{GPT-4o}$. Moreover, RocketEval provides a cost reduction exceeding 50-fold for large-scale evaluation and comparison scenarios. Our code is available at https://github.com/Joinn99/RocketEval-ICLR.", "keywords": [ "automated evaluation", "large language models", "natural language processing" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "zJfOyS1YLW", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Josiah Hanna", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "Assistant Professor" }, { "name": "Nicholas Corrado", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 10, "n_ref_uni": 40, "n_ref": 60, "n_ref_all": 99, "n_fig": 22, "n_tab": 1, "L_tab": 171, "n_element_tab": 12, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1480, "formula_len_all_1": 635, "len_all": 168543, "len_all_1": 70161, "len_abs": 1213, "len_title": 123, "len_sents": 57683, "len_sents_1": 33125, "n_sents": 455, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1601, "title": "On-Policy Policy Gradient Reinforcement Learning Without On-Policy Sampling", "abs": "On-policy reinforcement learning RL algorithms perform policy updates using i.i.d. trajectories collected by the current policy. However, after observing only a finite number of trajectories, on-policy sampling may produce data that fails to match the expected on-policy data distribution. This sampling error leads to noisy updates and data inefficient on-policy learning. Recent work in the policy evaluation setting has shown that non-i.i.d., off-policy sampling can produce data with lower sampling error than on-policy sampling can produce~\\citep{zhong2022robust}. Motivated by this observation, we introduce an adaptive, off-policy sampling method to improve the data efficiency of on-policy policy gradient algorithms. Our method, Proximal Robust On-Policy Sampling (PROPS) reduces sampling error by collecting data with a behavior policy that increases the probability of sampling actions that are under-sampled with respect to the current policy. Rather than discarding data from old policies -- as is commonly done in on-policy algorithms -- PROPS uses data collection to adjust the distribution of previously collected data to be approximately on-policy. We empirically evaluate PROPS on both continuous-action MuJoCo benchmark tasks as well discrete-action tasks and demonstrate that (1) PROPS decreases sampling error throughout training and (2) improves the data efficiency of on-policy policy gradient algorithms. Our work improves the RL community\u2019s understanding of a nuance in the on-policy vs off-policy dichotomy: on-policy learning requires on-policy data, not on-policy sampling.", "keywords": [ "reinforcement learning", "on-policy", "policy gradient", "data collection" ], "rating_list": [ 5, 8, 6, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "zJbwrk1DHc", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Guoyuan An", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Kaixiang Song", "gender": "Male", "institution": "Zhejiang Lab", "country": "", "position": "Researcher" }, { "name": "Shu Kong", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Assistant Professor" }, { "name": "Sung-eui Yoon", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "Professor" }, { "name": "Wei Hua", "gender": "Male", "institution": "Zhejiang Lab", "country": "CN", "position": "Full Professor" }, { "name": "Yanan Li", "gender": "Female", "institution": "Zhejiang Lab", "country": "CN", "position": "Researcher" }, { "name": "Yuchi Huo", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 27, "n_ref": 78, "n_ref_all": 156, "n_fig": 23, "n_tab": 11, "L_tab": 6255, "n_element_tab": 455, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 641, "formula_len_all_1": 523, "len_all": 161502, "len_all_1": 59058, "len_abs": 1738, "len_title": 102, "len_sents": 56216, "len_sents_1": 28130, "n_sents": 422, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 1743, "title": "Hierarchical Classification via Diffusion on Manifolds", "abs": "Hierarchical classification, the problem of classifying images according to a predefined hierarchical taxonomy, has practical significance owing to the principle of ``making better mistakes'', i.e., better to predict correct coarse labels than incorrect fine labels. Yet, it is insufficiently studied in literature, presumably because simply finetuning a pretrained deep neural network using the cross-entropy loss on leaf classes already leads to good performance w.r.t not only the popular top-1 accuracy but also hierarchical metrics. Despite the empirical effectiveness of finetuning pretrained models, we argue that hierarchical classification could be better addressed by explicitly regularizing finetuning w.r.t the predefined hierarchical taxonomy. Intuitively, with a pretrained model, data lies in hierarchical manifolds in the feature space. Hence, we propose a hierarchical multimodal contrastive finetuning method to leverage taxonomic hierarchy to finetune a pretrained model for better hierarchical classification. Moreover, the hierarchical manifolds motivate a graph diffusion-based method to adjust posteriors at hierarchical levels altogether in inference. This distinguishes our method from the existing ones, including top-down approaches (using coarse-class predictions to adjust fine-class predictions) and bottom-up approaches (processing fine-class predictions towards coarse-label predictions). We validate our method on two large-scale datasets, iNat18 and iNat21. Extensive experiments demonstrate that our method significantly outperforms prior arts w.r.t both top-1 accuracy and established hierarchical metrics, thanks to our new multi-modal hierarchical contrastive training and graph-diffusion-based inference.", "keywords": [ "hierarchical classification", "graph diffusion" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "zIqLQVBxdd", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Han Xu", "gender": "Male", "institution": "Institute of automation, Chinese Academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Kai Chen", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Kang Liu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Professor" }, { "name": "Qiwei Ye", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuancheng SUN", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 9, "n_ref_uni": 64, "n_ref": 106, "n_ref_all": 123, "n_fig": 9, "n_tab": 5, "L_tab": 3567, "n_element_tab": 178, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 923, "n_element_tab_1": 46, "formula_len_all": 1113, "formula_len_all_1": 451, "len_all": 177345, "len_all_1": 65409, "len_abs": 1716, "len_title": 142, "len_sents": 51222, "len_sents_1": 32171, "n_sents": 350, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1723, "title": "The Latent Road to Atoms: Backmapping Coarse-grained Protein Structures with Latent Diffusion", "abs": "Coarse-grained molecular dynamics simulations offer computational efficiency for exploring protein conformational ensembles and thermodynamic properties.\nThough coarse representations enable large-scale simulations across extended temporal and spatial ranges, the sacrifice of atomic-level details limits their utility in tasks such as ligand docking and protein-protein interaction prediction.\nBackmapping, the process of reconstructing all-atom structures from coarse-grained representations, is crucial for recovering these fine details.\nWhile recent machine learning methods have made strides in protein structure generation, challenges persist in reconstructing diverse atomistic conformations that maintain geometric accuracy and chemical validity.\nIn this paper, we present Latent Diffusion Backmapping (LDB), a novel approach leveraging denoising diffusion within latent space to address these challenges. \nBy combining discrete latent encoding with diffusion, LDB bypasses the need for equivariant and internal coordinate manipulation, significantly simplifying the training and sampling processes as well as facilitating better and wider exploration in configuration space. \nWe evaluate LDB\u2019s state-of-the-art performance on three distinct protein datasets, demonstrating its ability to efficiently reconstruct structures with high structural accuracy and chemical validity.\nMoreover, LDB shows exceptional versatility in capturing diverse protein ensembles, highlighting its capability to explore intricate conformational spaces. \nOur results position LDB as a powerful and scalable approach for backmapping, effectively bridging the gap between CG simulations and atomic-level analyses in computational biology.", "keywords": [ "Protein Structure Reconstruction", "Latent Diffusion", "Discrete Protein Representations" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 1, 3, 1, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "zIPFFhowuM", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Greg Durrett", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Kaj Bostrom", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "PhD student" }, { "name": "Swarat Chaudhuri", "gender": "Male", "institution": "Google Deepmind", "country": "GB", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 37, "n_ref": 48, "n_ref_all": 74, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1211, "n_element_tab_1": 90, "formula_len_all": 3781, "formula_len_all_1": 3104, "len_all": 134224, "len_all_1": 60676, "len_abs": 1142, "len_title": 86, "len_sents": 35310, "len_sents_1": 26155, "n_sents": 236, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 38, "L_abs": 1147, "title": "Proof Search Augmented Language Models", "abs": "Transformer language models (TLMs) exhibit an impressively general range of capabilities. A growing body of work aims to harness these models for complex reasoning problems expressed in natural language. However, recent theoretical and empirical results have revealed limits to the algorithmic generalization of TLM reasoning. Transformers trained to solve deduction problems from one distribution fail to solve instances of the same problem type drawn from other distributions. We propose to improve the systematic reasoning capabilities of TLMs via a differentiable proof search module, yielding proof-search augmented language models (PSALMs).\nIn a PSALM, a Transformer is responsible for predicting rule and fact representations for a neural theorem prover (NTP). The NTP performs a backward-chaining search over proofs, scoring them based on a soft unification operation. Our results show that PSALMs successfully generalize in deduction tasks where vanilla transformers do not learn systematic behavior, can be adapted to more natural text with only label supervision, and robustly handle large examples where proprietary LLMs make mistakes.", "keywords": [ "reasoning", "transformers", "neural theorem proving", "neural network architectures", "differentiable algorithms" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "zHf7hOfeer", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Jiahao Fan", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "Postdoc" }, { "name": "Yanan Sun", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "Full Professor" }, { "name": "Yuqi Feng", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "PhD student" }, { "name": "Yuwei Ou", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "MS student" } ], "n_formula": 21, "n_formula_1": 12, "n_ref_uni": 54, "n_ref": 143, "n_ref_all": 221, "n_fig": 10, "n_tab": 15, "L_tab": 4404, "n_element_tab": 503, "n_fig_1": 1, "n_tab_1": 8, "L_tab_1": 1770, "n_element_tab_1": 226, "formula_len_all": 1176, "formula_len_all_1": 644, "len_all": 185669, "len_all_1": 75781, "len_abs": 1368, "len_title": 101, "len_sents": 54143, "len_sents_1": 33398, "n_sents": 388, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1382, "title": "Zero-cost Proxy for Adversarial Robustness Evaluation", "abs": "Deep neural networks (DNNs) easily cause security issues due to the lack of adversarial robustness. An emerging research topic for this problem is to design adversarially robust architectures via neural architecture search (NAS), i.e., robust NAS. However, robust NAS needs to train numerous DNNs for robustness estimation, making the search process prohibitively expensive. In this paper, we propose a zero-cost proxy to evaluate the adversarial robustness without training. Specifically, the proposed zero-cost proxy formulates the upper bound of adversarial loss, which can directly reflect the adversarial robustness. The formulation involves only the initialized weights of DNNs, thus the training process is no longer needed. Moreover, we theoretically justify the validity of the proposed proxy based on the theory of neural tangent kernel and input loss landscape. Experimental results show that the proposed zero-cost proxy can bring more than $20\\times$ speedup compared with the state-of-the-art robust NAS methods, while the searched architecture has superior robustness and transferability under white-box and black-box attacks. Furthermore, compared with the state-of-the-art zero-cost proxies, the calculation of the proposed method has the strongest correlation with adversarial robustness. Our source code is available at https://anonymous.4open.science/r/ZCP-05B6.", "keywords": [ "Neural architecture search", "adversarial robustness", "zero-cost proxy" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 4, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "zHeHIIFQVF", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Adrian M\u00fcller", "gender": "Non-Binary", "institution": "Aalto University", "country": "FI", "position": "MS student" }, { "name": "Krishna Agaram", "gender": "unknown", "institution": "Indian Institute of Technology Bombay, Indian Institute of Technology, Bombay", "country": "IN", "position": "Undergrad student" }, { "name": "Siddhant Midha", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Vikas K Garg", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 10, "n_ref_uni": 55, "n_ref": 94, "n_ref_all": 107, "n_fig": 5, "n_tab": 3, "L_tab": 337, "n_element_tab": 13, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 70, "n_element_tab_1": 9, "formula_len_all": 1664, "formula_len_all_1": 1166, "len_all": 161339, "len_all_1": 66278, "len_abs": 1401, "len_title": 118, "len_sents": 49599, "len_sents_1": 31374, "n_sents": 381, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1416, "title": "Train once and generalize: Zero-shot quantum state preparation with RL", "abs": "Quantum state preparation forms an essential cornerstone of quantum information science and quantum algorithms. Designing efficient and scalable methods for approximate state preparation on near-term quantum devices remains a significant challenge, with worst-case hardness results compounding this difficulty. In this work, we propose a deep reinforcement learning framework for quantum state preparation, capable of immediate inference of arbitrary stabilizer states at a fixed system size post a training phase. Our approach scales substantially beyond previous works by leveraging a novel reward function. In our experiments on stabilizer states up to nine qubits, our trained agent successfully prepares nearly all previously unseen states, despite being trained on less than $10^{-3}$\\% of the state space -- demonstrating significant generalization to novel states. Benchmarking shows our model produces stabilizer circuits with size $60$\\% that of existing algorithms, setting a new state of the art in circuit efficiency. Furthermore, we show that this performance advantage is consistent across states with varying entanglement content. We also analyze the rate of increase of entanglement entropy across the prepared circuit, obtaining insight into the quantum entanglement dynamics generated by our trained agent. Finally, we prove our agent generalizes to (almost) the entire space of stabilizer states.", "keywords": [ "Quantum State Preparation", "Deep Reinforcement Learning", "Zero-shot Inference", "Off-the-shelf Algorithms", "Generalization" ], "rating_list": [ 3, 3, 6, 6, 3 ], "soundness_list": [ 2, 2, 3, 3, 1 ], "presentation_list": [ 2, 2, 3, 3, 1 ], "contribution_list": [ 1, 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 3, 5 ] }, { "paper_id": "zGzs5SIwT8", "primary_area": "causal reasoning", "n_author": 9, "author_info_list": [ { "name": "Biwei Huang", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Haoran Deng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Ignavier Ng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Mingming Gong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Wenqin Liu", "gender": "Female", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" }, { "name": "Yian Ma", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Yingyu Lin", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "yuxing huang", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 82, "n_formula_1": 21, "n_ref_uni": 55, "n_ref": 103, "n_ref_all": 141, "n_fig": 8, "n_tab": 6, "L_tab": 674, "n_element_tab": 96, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 21, "n_element_tab_1": 1, "formula_len_all": 12257, "formula_len_all_1": 1865, "len_all": 191229, "len_all_1": 68645, "len_abs": 1132, "len_title": 131, "len_sents": 52998, "len_sents_1": 31645, "n_sents": 478, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1165, "title": "A Skewness-Based Criterion for Addressing Heteroscedastic Noise in Causal Discovery", "abs": "Real-world data often violates the equal-variance assumption (homoscedasticity), making it essential to account for heteroscedastic noise in causal discovery. In this work, we explore heteroscedastic symmetric noise models (HSNMs), where the effect $Y$ is modeled as $Y = f(X) + \\sigma(X)N$, with $X$ as the cause and $N$ as independent noise following a symmetric distribution. We introduce a novel criterion for identifying HSNMs based on the skewness of the score (i.e., the gradient of the log density) of the data distribution. This criterion establishes a computationally tractable measurement that is zero in the causal direction but nonzero in the anticausal direction, enabling the causal direction discovery. We extend this skewness-based criterion to the multivariate setting and propose \\texttt{SkewScore}, an algorithm that handles heteroscedastic noise without requiring the extraction of exogenous noise. We also conduct a case study on the robustness of \\texttt{SkewScore} in a bivariate model with a latent confounder, providing theoretical insights into its performance. Empirical studies further validate the effectiveness of the proposed method.", "keywords": [ "Causal Discovery", "Heteroscedastic Noise", "Score Matching" ], "rating_list": [ 6, 10, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 2, 4 ], "presentation_list": [ 4, 4, 3, 3, 3 ], "contribution_list": [ 3, 4, 2, 2, 3 ], "confidence_list": [ 3, 4, 2, 4, 4 ] }, { "paper_id": "zGvwENuzPU", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Edward Rees", "gender": "Male", "institution": "Apollo Research", "country": "", "position": "Researcher" }, { "name": "Ethan Perez", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Hunar Batra", "gender": "Female", "institution": "Department of Computer Science, University of Oxford", "country": "", "position": "PhD student" }, { "name": "James Chua", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Julian Michael", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "Miles Andrew Turpin", "gender": "Male", "institution": "Scale AI", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 44, "n_ref": 75, "n_ref_all": 127, "n_fig": 17, "n_tab": 13, "L_tab": 2808, "n_element_tab": 199, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 907, "n_element_tab_1": 23, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 261932, "len_all_1": 71800, "len_abs": 1075, "len_title": 128, "len_sents": 88764, "len_sents_1": 34951, "n_sents": 747, "n_sents_1": 258, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1078, "title": "Bias-Augmented Consistency Training Reduces Biased Reasoning in Chain-of-Thought", "abs": "While chain-of-thought prompting (CoT) has the potential to improve the explainability of language model reasoning, it can systematically misrepresent the factors influencing models' behavior--for example, rationalizing answers in line with a user's opinion without mentioning this bias. To mitigate this biased reasoning problem, we introduce bias-augmented consistency training (BCT), an unsupervised fine-tuning scheme that trains models to give consistent reasoning across prompts with and without biasing features. We construct a suite testing nine forms of biased reasoning on seven question-answering tasks, and find that applying BCT to GPT-3.5-Turbo with one bias reduces the rate of biased reasoning by 86% on held-out tasks. Moreover, this model generalizes to other forms of bias, reducing biased reasoning on held-out biases by an average of 37%. As BCT generalizes to held-out biases and does not require gold labels, this method may hold promise for reducing biased reasoning from as-of-yet unknown biases and on tasks where ground truth reasoning is unavailable.", "keywords": [ "Chain-of-thought prompting", "Explainability", "Generalization", "Reasoning", "Bias" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "zGej22CBnS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Brandon Amos", "gender": "unknown", "institution": "Meta AI", "country": "US", "position": "Research Scientist" }, { "name": "Itai Gat", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Karen Ullrich", "gender": "Female", "institution": "Meta AI", "country": "US", "position": "Research Scientist" }, { "name": "Marton Havasi", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Matthew J. Muckley", "gender": "Male", "institution": "Meta", "country": "US", "position": "Research Engineer" }, { "name": "Truong Buu Phan", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "PhD student" } ], "n_formula": 39, "n_formula_1": 13, "n_ref_uni": 21, "n_ref": 43, "n_ref_all": 69, "n_fig": 6, "n_tab": 4, "L_tab": 813, "n_element_tab": 132, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2854, "formula_len_all_1": 831, "len_all": 179142, "len_all_1": 80844, "len_abs": 1655, "len_title": 143, "len_sents": 54863, "len_sents_1": 30219, "n_sents": 483, "n_sents_1": 267, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1670, "title": "Exact Byte-Level Probabilities from Tokenized Language Models for FIM-Tasks and Model Ensembles", "abs": "Tokenization is associated with many poorly understood shortcomings in language models (LMs), yet remains an important component for long sequence scaling purposes. This work studies how tokenization impacts model performance by analyzing and comparing the stochastic behavior of tokenized models with their byte-level, or token-free, counterparts. We discover that, even when the two models are statistically equivalent, their predictive distributions over the next byte can be substantially different, a phenomenon we term as ``tokenization bias''. To fully characterize this phenomenon, we introduce the Byte-Token Representation Lemma, a framework that establishes a mapping between the learned token distribution and its equivalent byte-level distribution. From this result, we develop a next-byte sampling algorithm that eliminates tokenization bias without requiring further training or optimization. In other words, this enables zero-shot conversion of tokenized LMs into statistically equivalent token-free ones. We demonstrate its broad applicability with two use cases: fill-in-the-middle (FIM) tasks and model ensembles. In FIM tasks where input prompts may terminate mid-token, leading to out-of-distribution tokenization, our method mitigates performance degradation and achieves an approximately 18\\% improvement in FIM coding benchmarks, consistently outperforming the standard token healing fix. For model ensembles where each model employs a distinct vocabulary, our approach enables seamless integration, resulting in improved performance (up to 3.7\\%) over individual models across various standard baselines in reasoning, knowledge, and coding.", "keywords": [ "Language models", "Tokenization", "Probability", "Sampling" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "zGb4WgCW5i", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Chengzhi Liu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Intern" }, { "name": "Feilong Tang", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Researcher" }, { "name": "Harry Yang", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Qiang Sun", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Associate Professor" }, { "name": "Zile Huang", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "CN", "position": "Undergrad student" }, { "name": "Ser-Nam Lim", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Associate Professor" } ], "n_formula": 49, "n_formula_1": 16, "n_ref_uni": 50, "n_ref": 105, "n_ref_all": 129, "n_fig": 9, "n_tab": 13, "L_tab": 4544, "n_element_tab": 768, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1773, "n_element_tab_1": 376, "formula_len_all": 3215, "formula_len_all_1": 833, "len_all": 283269, "len_all_1": 75557, "len_abs": 1875, "len_title": 132, "len_sents": 83176, "len_sents_1": 32586, "n_sents": 650, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1676, "title": "Intervening Anchor Token: Decoding Strategy in Alleviating Hallucinations for MLLMs", "abs": "Multimodal large language models (MLLMs) offer a powerful mechanism for interpreting visual information. However, they often suffer from hallucinations, which impede the real-world usage of these models. Existing methods attempt to alleviate this issue by designing special decoding strategies that penalize the summary tokens. However, these methods lack analysis of the relationship between hallucination and summarization mechanism of LLMs. Interestingly, we find that penalizing summary tokens is not necessary: merely intervening the query-key parameters variance, without costing extra inference time, still alleviates hallucinations. Specifically, we explore the causes of hallucinations by analyzing localized self-attention patterns called ``anchor\" tokens and define the attention localization degree of the model as token propagation probabilities. Our analysis reveals that over-propagation of anchor tokens occurs when the distribution of eigenvalues of the query and key matrices has a non-zero mean and a polarized variance, leading to excessive dependence on anchor tokens while neglecting vision information and describes the image content with hallucination. Based on the observation, we propose a versatile plug-and-play decoding strategy, Dynamic Token Propagation Mechanism (TAME), to alleviate excessive propagation by dynamically intervening the eigenspectrum variance of the attention weight, thereby alleviating hallucinations without relying on complex decoding strategies. Extensive experiments reveal a correlation between the eigenspectrum and hallucinations across various MLLMs, and show that TAME reduces the percentage of hallucinated objects.", "keywords": [ "Multimodal large language models; Hallucination" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "zG459X3Xge", "primary_area": "generative models", "n_author": 11, "author_info_list": [ { "name": "Bokai Xu", "gender": "Male", "institution": "Chinese University of Hong Kong, Shenzhen", "country": "", "position": "Undergrad student" }, { "name": "Chaoyue Tang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Junbo Cui", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Junhao Ran", "gender": "Male", "institution": "Rice University", "country": "US", "position": "MS student" }, { "name": "Maosong Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Shi Yu", "gender": "unknown", "institution": "Department of Computer Science and Technology, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Shuo Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Xu Han", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yukun Yan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Zhenghao Liu", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "Associate Professor" }, { "name": "Zhiyuan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 53, "n_ref": 117, "n_ref_all": 151, "n_fig": 6, "n_tab": 5, "L_tab": 1852, "n_element_tab": 208, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 458, "n_element_tab_1": 54, "formula_len_all": 264, "formula_len_all_1": 266, "len_all": 217200, "len_all_1": 62796, "len_abs": 1446, "len_title": 127, "len_sents": 54540, "len_sents_1": 25329, "n_sents": 382, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1453, "title": "VisRAG: Vision-based Retrieval-augmented Generation on Multi-modality Documents", "abs": "Retrieval-augmented generation (RAG) is an effective technique that enables large language models (LLMs) to utilize external knowledge sources for generation. However, current RAG systems are solely based on text, rendering it impossible to utilize vision information like layout and images that play crucial roles in real-world multi-modality documents. In this paper, we introduce VisRAG, which tackles this issue by establishing a vision-language model (VLM)-based RAG pipeline. In this pipeline, instead of first parsing the document to obtain text, the document is directly embedded using a VLM as an image and then retrieved to enhance the generation of a VLM. Compared to traditional text-based RAG, VisRAG maximizes the retention and utilization of the data information in the original documents, eliminating the information loss introduced during the parsing process. We collect both open-source and synthetic data to train the retriever in VisRAG and explore a variety of generation methods. Experiments demonstrate that VisRAG outperforms traditional RAG in both the retrieval and generation stages, achieving a 25\u201339% end-to-end performance gain over traditional textbased RAG pipeline. Further analysis reveals that VisRAG is effective in utilizing training data and demonstrates strong generalization capability, positioning it as a promising solution for RAG on multi-modality documents. Our code and data will be made publicly available.", "keywords": [ "Retrieval-augmented Generation", "Vision-language Models" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "zG2vcC1l1f", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Johannes A. Stork", "gender": "Not Specified", "institution": "\u00d6rebro University", "country": "", "position": "Assistant Professor" }, { "name": "Martin Magnusson", "gender": "Male", "institution": "\u00d6rebro University", "country": "SE", "position": "Associate Professor" }, { "name": "Shih-Min Yang", "gender": "Male", "institution": "\u00d6rebro University", "country": "SE", "position": "PhD student" }, { "name": "Todor Stoyanov", "gender": "Male", "institution": "\u00d6rebro University", "country": "SE", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 21, "n_ref": 39, "n_ref_all": 56, "n_fig": 7, "n_tab": 7, "L_tab": 1953, "n_element_tab": 96, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 410, "n_element_tab_1": 39, "formula_len_all": 248, "formula_len_all_1": 248, "len_all": 112635, "len_all_1": 66441, "len_abs": 1752, "len_title": 111, "len_sents": 33355, "len_sents_1": 31868, "n_sents": 226, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 113, "L_abs": 1768, "title": "KEA: Keeping Exploration Alive by Proactively Coordinating Exploration Strategies in Curiosity-driven Exploration", "abs": "In continuous control tasks, Soft Actor-Critic (SAC) has achieved notable success by balancing exploration and exploitation. However, SAC struggles in sparse reward environments, where infrequent rewards hinder efficient exploration. While novelty-based exploration methods help address this issue by encouraging the agent to explore novel states, they introduce challenges, such as the difficulty of setting an optimal reward scale and managing the interaction between novelty-based exploration and SAC\u2019s stochastic policy. These complexities often lead to inefficient exploration or premature convergence and make balancing exploration-exploitation challenging. In this paper, we propose KEA (Keeping Exploration Alive) to tackle the inefficiencies in balancing the exploration-exploitation trade-off when combining SAC with novelty-based methods. KEA introduces an additional co-behavior agent that works alongside SAC and a switching mechanism to facilitate proactive coordination between exploration strategies from the co-behavior agent and the SAC agent with novelty-based exploration. This coordination allows the agent to maintain stochasticity in high-novelty regions, preventing premature convergence and enhancing exploration efficiency. We first analyze the difficulty of balancing exploration-exploitation when combining SAC with novelty-based methods in a 2D grid environment. We then evaluate KEA on sparse reward control tasks from the DeepMind Control Suite and compare against two state-of-the-art novelty-based exploration baselines --- Random Network Distillation (RND) and NovelD. KEA improves episodic rewards by up to 119\\% over RND and 28\\% over NovelD, significantly improving learning efficiency and robustness in sparse reward environments.", "keywords": [ "Reinforcement Learning", "Curiosity-based Exploration", "Sparse Reward", "Soft Actor-Critic" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "zFfZEQHUiv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Hoseung Choi", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Undergrad student" }, { "name": "Junmyeong Lee", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Minsu Cho", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Associate Professor" }, { "name": "Yoonwoo Jeong", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 29, "n_ref_uni": 47, "n_ref": 103, "n_ref_all": 113, "n_fig": 3, "n_tab": 4, "L_tab": 135, "n_element_tab": 17, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1590, "formula_len_all_1": 1530, "len_all": 146319, "len_all_1": 55777, "len_abs": 1302, "len_title": 121, "len_sents": 39268, "len_sents_1": 25354, "n_sents": 288, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1326, "title": "Towards Pose-Free Dynamic Neural Fields: Leveraging Geometric Foundation Models", "abs": "Dynamic view synthesis (DVS) from monocular videos has remarkably advanced in recent years, achieving high-fidelity rendering with reduced computational costs. Despite these advancements, the optimization of dynamic neural fields still relies on traditional structure from motion (SfM), requiring that all objects remain stationary during scene capture. To address this limitation, we present \\textbf{SC-4DGS}, a pose-free optimization pipeline for dynamic Gaussian Splatting (GS) from monocular videos, which eliminates the need for SfM through self-calibration. Specifically, we jointly optimize dynamic Gaussian representations and camera poses by utilizing DUSt3R, enabling accurate calibration and rendering.\nFurthermore, we introduce a comprehensive benchmark, \\textbf{Kubric-MRig}, that includes extensive camera and object motions along with simultaneous multi-view captures. \nUnlike previous benchmarks for DVS, where ground truths for camera information are absent due to the difficulty of capturing multiple viewpoints simultaneously, it facilitates evaluating both calibration and rendering quality in dynamic scenes.\nExperimental results demonstrate that the proposed method outperforms previous pose-free dynamic neural fields and achieves competitive performance compared to existing pose-free 3D neural fields.", "keywords": [ "Neural Rendering", "Pose-free", "Gaussian Splatting", "Dynamic View Synthesis" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "zEm5nXxiXU", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Dadi Guo", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Dongyan Zhao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Zichen Wen", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Huishuai Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 44, "n_ref": 76, "n_ref_all": 114, "n_fig": 7, "n_tab": 8, "L_tab": 3281, "n_element_tab": 326, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1282, "n_element_tab_1": 137, "formula_len_all": 112, "formula_len_all_1": 111, "len_all": 169007, "len_all_1": 60982, "len_abs": 1331, "len_title": 150, "len_sents": 43863, "len_sents_1": 27088, "n_sents": 331, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1333, "title": "AIDBench: A benchmark for evaluating the authorship identification capability of large language models", "abs": "As large language models (LLMs) rapidly advance and integrate into daily life, the privacy risks they pose are attracting increasing attention. We focus on a specific privacy risk where LLMs may help identify the authorship of anonymous texts, which challenges the effectiveness of anonymity in real-world systems such as anonymous peer review systems. To investigate these risks, we present AIDBench, a new benchmark that incorporates several author identification datasets, including emails, blogs, reviews, articles, and research papers. AIDBench utilizes two evaluation methods: one-to-one authorship identification, which determines whether two texts are from the same author; and one-to-many authorship identification, which, given a query text and a list of candidate texts, identifies the candidate most likely written by the same author as the query text. We also introduce a Retrieval-Augmented Generation (RAG)-based method to enhance the large-scale authorship identification capabilities of LLMs, particularly when input lengths exceed the models' context windows, thereby establishing a new baseline for authorship identification using LLMs. Our experiments with AIDBench demonstrate that LLMs can correctly guess authorship at rates well above random chance, revealing new privacy risks posed by these powerful models.", "keywords": [ "large language models", "privacy", "authorship identification" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "zEUDoD9cU9", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Cairong Yan", "gender": "Female", "institution": "Donghua University, Shanghai", "country": "CN", "position": "Associate Professor" }, { "name": "Huanwen Zheng", "gender": "Male", "institution": "Donghua University, Shanghai", "country": "CN", "position": "MS student" }, { "name": "Jingyi Guo", "gender": "Female", "institution": "Donghua University, Shanghai", "country": "CN", "position": "MS student" }, { "name": "Yanting Zhang", "gender": "Female", "institution": "Donghua University, Shanghai", "country": "CN", "position": "Associate Professor" }, { "name": "Zeqiang Lai", "gender": "unknown", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Wenhai Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 32, "n_ref": 65, "n_ref_all": 78, "n_fig": 6, "n_tab": 3, "L_tab": 166, "n_element_tab": 41, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 372, "n_element_tab_1": 37, "formula_len_all": 170, "formula_len_all_1": 172, "len_all": 99767, "len_all_1": 54744, "len_abs": 1025, "len_title": 131, "len_sents": 30996, "len_sents_1": 26980, "n_sents": 202, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1030, "title": "CycleVTON: Improving Diffusion-Based Virtual Try-On with Cycle-Consistent Training", "abs": "We present CycleVTON, a cycle-consistent diffusion-based virtual try-on framework. Unlike existing methods that rely on a single try-on network, our model consists of two conjugated networks. In addition to the regular try-on network, we design a clothing extraction network that extracts the clothing worn by the person and standardizes it into a front-facing format. These two networks are symmetrical, enabling alignment between the generated dressed human and real images of dressed human, as well as between the extracted clothing and its front-facing ground truth. This cycle-consistent optimization strategy allows for enhanced retention of clothing textures and structures, ensuring a more realistic and accurate clothing generation in virtual try-on scenarios. Moreover, the conjugated network structure not only supports traditional virtual try-on but also allows flexible clothing extraction and clothing exchange between different individuals. The experiments on VITON-HD demonstrate the effectiveness of our approach.", "keywords": [ "virtual try-on", "diffusion", "cycle-consistency" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 5, 4, 5 ] }, { "paper_id": "zET0Zg71WT", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 9, "author_info_list": [ { "name": "Alvaro Velasquez", "gender": "Male", "institution": "Defense Advanced Research Projects Agency", "country": "US", "position": "Program Manager" }, { "name": "Calvin Yeung", "gender": "unknown", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Hamza Errahmouni Barkam", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Hanning Chen", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Mohsen Imani", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Assistant Professor" }, { "name": "SungHeon Jeong", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Wenjun Huang", "gender": "unknown", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Yang Ni", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Zhuowen Zou", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 17, "n_ref_uni": 28, "n_ref": 44, "n_ref_all": 58, "n_fig": 7, "n_tab": 5, "L_tab": 2945, "n_element_tab": 450, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 3016, "n_element_tab_1": 194, "formula_len_all": 1403, "formula_len_all_1": 1319, "len_all": 121106, "len_all_1": 61507, "len_abs": 1481, "len_title": 112, "len_sents": 30472, "len_sents_1": 24212, "n_sents": 252, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1489, "title": "Structure-aware Attention based on Vector Symbolic Architectures", "abs": "The introduction of the Transformer has brought about a revolution in AI. Central to the success of the Transformer architecture is the self-attention mechanism, enabling context dependence and long-range dependencies between tokens. Recent work has drawn an equivalence between Hopfield networks, a kind of associative memory model, and Transformers. In this work, we leverage this bridge, using Vector Symbolic Architectures (VSA), a brain-inspired computational paradigm capable of representing and implementing data structures, including associative memory models, to define a broad class of attention mechanisms catered for complex data types. In particular, we use Generalized Holographic Reduced Representations (GHRR), an implementation of a VSA, as the foundation for our proposed class of attention mechanisms. We show that GHRR is capable of implementing attention and design a GHRR Transformer encoder architecture based on the demonstrated mathematical equivalence. We propose a new kind of binding-based positional encoding based on methods used in VSAs for encoding sequential information. We extend the attention mechanism in our architecture to support graphs, inspired by techniques used in VSAs to encode graph representations. We evaluate the GHRR Transformer on language modeling, vertex classification, and graph classification tasks. Results suggest that our approach provides benefits in language modeling and graph classification tasks compared to baseline models.", "keywords": [ "transformers", "attention", "vector symbolic architectures", "neurosymbolic ai", "hyperdimensional computing" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 1, 3, 3, 3 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "zEPYCDaJae", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 3, "author_info_list": [ { "name": "Xinye Yang", "gender": "Male", "institution": "Brown University", "country": "US", "position": "MS student" }, { "name": "Yang Xiang", "gender": "Non-Binary", "institution": "Brown University", "country": "US", "position": "MS student" }, { "name": "Yanghao Wu", "gender": "Male", "institution": "Brown University", "country": "US", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 29, "n_ref": 36, "n_ref_all": 40, "n_fig": 3, "n_tab": 6, "L_tab": 8342, "n_element_tab": 328, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 776, "n_element_tab_1": 156, "formula_len_all": 48, "formula_len_all_1": 0, "len_all": 176251, "len_all_1": 67092, "len_abs": 1506, "len_title": 145, "len_sents": 62875, "len_sents_1": 26287, "n_sents": 468, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1513, "title": "DATASEA - AN AUTOMATIC FRAMEWORK FOR COMPREHENSIVE DATASET PROCESSING USING LARGE LANGUAGE MODELS", "abs": "In the era of data-driven decision-making, efficiently acquiring and analyzing\ndiverse datasets is critical for accelerating research and innovation. Yet, traditional manual approaches to dataset discovery, preparation, and exploration\nremain inefficient and cumbersome, especially as the scale and complexity of\ndatasets continue to expand. These challenges create major roadblocks, slowing down the pace of progress and reducing the capacity for data-driven breakthroughs. To address these challenges, we introduce DataSEA (Search, Evaluate, Analyze), a fully automated system for comprehensive dataset processing, leveraging large language models (LLMs) to streamline the data handling\npipeline. DataSEA autonomously searches for dataset sources, retrieves and organizes evaluation metadata, and generates custom scripts to load and analyze\ndata based on user input. Users can provide just a dataset name, and DataSEA\nwill handle the entire preparation process. While fully automated, minimal user\ninteraction can further enhance system accuracy and dataset handling specificity.\nWe evaluated DataSEA on datasets from distinct fields, demonstrating its robustness and efficiency in reducing the time and effort required for data preparation\nand exploration. By automating these foundational tasks, DataSEA empowers\nresearchers to allocate more time to in-depth analysis and hypothesis generation, ultimately accelerating the pace of innovation. The code is available at\nhttps://github.com/SingleView11/DataSEA.", "keywords": [ "Automated Data Processing", "LLM", "Data Pipeline Automation", "NLP", "Data Mining" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 5, 3, 3, 5 ] }, { "paper_id": "zE4mL85zgg", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Baochang Zhang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Professor" }, { "name": "Guodong Guo", "gender": "Male", "institution": "West Virginia University", "country": "US", "position": "Full Professor" }, { "name": "Hong Li", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Linlin Yang", "gender": "Male", "institution": "Communication University of China", "country": "CN", "position": "Lecturer" }, { "name": "Peng Zhou", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Runtang Guo", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Tongfei Chen", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Yuguang Yang", "gender": "Male", "institution": "\u00a0School of Electronic Information Engineering", "country": "CN", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 28, "n_ref_uni": 5, "n_ref": 12, "n_ref_all": 16, "n_fig": 1, "n_tab": 4, "L_tab": 1958, "n_element_tab": 131, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1958, "n_element_tab_1": 131, "formula_len_all": 1641, "formula_len_all_1": 1664, "len_all": 68318, "len_all_1": 59177, "len_abs": 1331, "len_title": 123, "len_sents": 25940, "len_sents_1": 25342, "n_sents": 213, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1335, "title": "CLAD: A Contrastive Learning based Method for Multi-Class Anomaly Detection", "abs": "Anomaly detection is crucial yet challenging in industrial production, especially in multi-class scenarios. Existing high-performance unsupervised methods often suffer from low efficiency and high model complexity. While lightweight discriminator-based detectors have been proposed, they are typically designed for single-class detection and exhibit significant performance degradation when extended to multi-class tasks. To address these limitations, we propose a novel Contrastive Learning-based multi-class Anomaly Detection (CLAD) method.\nOur approach first encodes multi-class normal images to generate normal samples in the feature space, then synthesizes anomalous samples in this encoded space. We then employ an Adapter network to compress the samples and leverage contrastive learning to effectively cluster normal and anomalous samples across multiple classes. Finally, a discriminator network is used for anomaly classification and identification. By leveraging anomaly sample generation and a two-stage training process, our framework achieves state-of-the-art performance on the MVTec and VisA datasets under the discriminator-based paradigm. Our key contributions include a novel framework for multi-class anomaly detection, efficient sample generation techniques, and a comprehensive evaluation of model configurations.", "keywords": [ "Industrial anomaly detection", "Multi-class anomaly detection", "Contrastive Learning" ], "rating_list": [ 1, 3, 3, 1, 3 ], "soundness_list": [ 2, 1, 1, 1, 2 ], "presentation_list": [ 2, 1, 1, 1, 2 ], "contribution_list": [ 2, 1, 2, 1, 2 ], "confidence_list": [ 5, 5, 5, 4, 4 ] }, { "paper_id": "zDze7VtB5C", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Ashish Sabharwal", "gender": "Male", "institution": "Allen Institute for AI", "country": "US", "position": "Principal Researcher" }, { "name": "William Merrill", "gender": "Male", "institution": "New York University", "country": "", "position": "Graduate student" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 19, "n_ref": 45, "n_ref_all": 56, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 810, "formula_len_all_1": 507, "len_all": 143407, "len_all_1": 78833, "len_abs": 1580, "len_title": 127, "len_sents": 56454, "len_sents_1": 36711, "n_sents": 436, "n_sents_1": 268, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1596, "title": "A Little Depth Goes a Long Way: the Expressive Power of Log-Depth Transformers", "abs": "Most analysis of transformer expressivity treats the depth (number of layers) of a model as a fixed constant, and analyzes the kinds of problems such models can solve across inputs of unbounded length. In practice, however, the context length of a trained transformer model is bounded. Thus, a more pragmatic question is: *What kinds of computation can a transformer perform on inputs of bounded length?* We formalize this by studying highly uniform transformers where the depth can grow minimally with context length. In this regime, we show that transformers with depth $O(\\log C)$ can, in fact, compute solutions to two important problems for inputs bounded by some max context length $C$, namely *simulating finite automata*, which relates to the ability to track state, and *graph connectivity*, which underlies multi-step reasoning. Notably, both of these problems have previously been proven to be asymptotically beyond the reach of fixed depth transformers under standard complexity conjectures, yet empirically transformer models can successfully track state and perform multi-hop reasoning on short contexts. Our novel analysis thus explains how transformer models may rely on depth to feasibly solve problems up to bounded context that they cannot solve over long contexts. It makes actionable suggestions for practitioners as to how to minimally scale the depth of a transformer to support reasoning over long contexts, and also argues for dynamically unrolling depth as a more effective way of adding compute compared to increasing model dimension or adding a short chain of thought.", "keywords": [ "transformer", "expressivity", "limits", "bounded context", "circuits" ], "rating_list": [ 6, 5, 3, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 2, 2, 2, 4 ], "contribution_list": [ 2, 3, 1, 4 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "zDjHOsSQxd", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 5, "author_info_list": [ { "name": "Hanpin Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Katsumi Inoue", "gender": "unknown", "institution": "National Institute of Informatics", "country": "JP", "position": "Full Professor" }, { "name": "Kun Gao", "gender": "Male", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "SG", "position": "Scientist" }, { "name": "Yang Feng", "gender": "Male", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Yongzhi Cao", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 19, "n_ref_uni": 37, "n_ref": 69, "n_ref_all": 94, "n_fig": 10, "n_tab": 3, "L_tab": 1243, "n_element_tab": 72, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1753, "formula_len_all_1": 1791, "len_all": 124093, "len_all_1": 72274, "len_abs": 1038, "len_title": 103, "len_sents": 35860, "len_sents_1": 31725, "n_sents": 285, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1043, "title": "End-to-End Rule Induction from Raw Sequence Inputs", "abs": "Rule learning-based models are widely used in highly interpretable scenarios due to their transparent structures. Inductive logic programming (ILP), a form of machine learning, induces rules from facts while maintaining interpretability. Differentiable ILP models enhance this process by leveraging neural networks to improve robustness and scalability. However, most differentiable ILP methods rely on symbolic datasets, facing challenges when learning directly from raw data. Specifically, they struggle with explicit label leakage: The inability to map continuous inputs to symbolic variables without explicit supervision of input feature labels. In this work, we address this issue by integrating a self-supervised differentiable clustering model with a novel differentiable ILP model, enabling rule learning from raw data without explicit label leakage. The learned rules effectively describe raw data through its features. We demonstrate that our method intuitively and precisely learns generalized rules from time series and image data.", "keywords": [ "Neuro-Symbolic Methods", "Interpretability", "Inductive Logic Programming" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "zDJf7fvdid", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Hui LIU", "gender": "Female", "institution": "Saint Francis University", "country": "HK", "position": "Assistant Professor" }, { "name": "Junhui Hou", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Meng YOU", "gender": "Female", "institution": "City University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Zhiyu Zhu", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Postdoc" } ], "n_formula": 40, "n_formula_1": 24, "n_ref_uni": 64, "n_ref": 111, "n_ref_all": 149, "n_fig": 8, "n_tab": 7, "L_tab": 4731, "n_element_tab": 849, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1250, "n_element_tab_1": 109, "formula_len_all": 2776, "formula_len_all_1": 1304, "len_all": 190450, "len_all_1": 62099, "len_abs": 300, "len_title": 126, "len_sents": 36385, "len_sents_1": 25827, "n_sents": 293, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1053, "title": "NVS-Solver: Video Diffusion Model as Zero-Shot Novel View Synthesizer", "abs": "By harnessing the potent generative capabilities of pre-trained large video diffusion models, we propose a new novel view synthesis paradigm that operates without the need for training. The proposed method adaptively modulates the diffusion sampling process with the given views to enable the creation of visually pleasing results from single or multiple views of static scenes or monocular videos of dynamic scenes. Specifically, built upon our theoretical modeling, we iteratively modulate the score function with the given scene priors represented with warped input views to control the video diffusion process. Moreover, by theoretically exploring the boundary of the estimation error, we achieve the modulation in an adaptive fashion according to the view pose and the number of diffusion steps. Extensive evaluations on both static and dynamic scenes substantiate the significant superiority of our method over state-of-the-art methods both quantitatively and qualitatively. The source code can be found on https://github.com/ZHU-Zhiyu/NVS_Solver.", "keywords": [ "Diffusion sampling", "Novel view synthesis" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4, 2 ] }, { "paper_id": "zDJNUDprhW", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Baoxiang Wang", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Jing Dong", "gender": "Female", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "PhD student" }, { "name": "Yaoliang Yu", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Associate Professor" } ], "n_formula": 35, "n_formula_1": 35, "n_ref_uni": 25, "n_ref": 62, "n_ref_all": 74, "n_fig": 1, "n_tab": 2, "L_tab": 1220, "n_element_tab": 160, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 1220, "n_element_tab_1": 160, "formula_len_all": 3757, "formula_len_all_1": 3751, "len_all": 100145, "len_all_1": 72986, "len_abs": 1259, "len_title": 121, "len_sents": 26377, "len_sents_1": 25438, "n_sents": 269, "n_sents_1": 258, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 905, "title": "Uncoupled and Convergent Learning in Monotone Games under Bandit Feedback", "abs": "We study the problem of no-regret learning algorithms for general monotone and smooth games and their last-iterate convergence properties. Specifically, we investigate the problem under bandit feedback and strongly uncoupled dynamics, which allows modular development of the multi-player system that applies to a wide range of real applications. We propose a mirror-descent-based algorithm, which converges in $O(T^{-1/4})$ and is also no-regret. The result is achieved by a dedicated use of two regularizations and the analysis of the fixed point thereof. The convergence rate is further improved to $O(T^{-1/2})$ in the case of strongly monotone games.\nMotivated by practical tasks where the game evolves over time, the algorithm is extended to time-varying monotone games. We provide the first non-asymptotic result in converging monotone games and give improved results for equilibrium tracking games.", "keywords": [ "online learning", "game" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "zDC3iCBxJb", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Jiahan Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Jian Peng", "gender": "Male", "institution": "Univ. of Illinois at Urbana-Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Jianzhu Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jiaqi Guan", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Sheng Wang", "gender": "unknown", "institution": "University of Washington, Seattle", "country": "US", "position": "Assistant Professor" }, { "name": "Xiangxin Zhou", "gender": "Not Specified", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xingang Peng", "gender": "unknown", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yunan Luo", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 30, "n_ref_uni": 45, "n_ref": 75, "n_ref_all": 96, "n_fig": 11, "n_tab": 7, "L_tab": 2081, "n_element_tab": 335, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 740, "n_element_tab_1": 123, "formula_len_all": 32013, "formula_len_all_1": 28824, "len_all": 188012, "len_all_1": 110449, "len_abs": 877, "len_title": 88, "len_sents": 48395, "len_sents_1": 37781, "n_sents": 303, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 889, "title": "Group Ligands Docking to Protein Pockets", "abs": "Molecular docking is a key task in computational biology that has attracted increasing interest from the machine learning community. While existing methods have achieved success, they generally treat each protein-ligand pair in isolation. Inspired by the biochemical observation that ligands binding to the same target protein tend to adopt similar poses, we propose \\textsc{GroupBind}, a novel molecular docking framework that simultaneously considers multiple ligands docking to a protein. This is achieved by introducing an interaction layer for the group of ligands and a triangle attention module for embedding protein-ligand and group-ligand pairs. By integrating our approach with diffusion based docking model, we set a new state-of-the-art performance on the PDBBind blind docking benchmark, demonstrating the effectiveness of our paradigm in enhancing molecular docking accuracy.", "keywords": [ "molecular docking", "ai4science" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "zCxGCdzreM", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Chris Lu", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Jakob Nicolaus Foerster", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Michael Beukman", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Michael Matthews", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 3, "n_ref_uni": 83, "n_ref": 152, "n_ref_all": 190, "n_fig": 23, "n_tab": 5, "L_tab": 859, "n_element_tab": 108, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 733, "formula_len_all_1": 78, "len_all": 267171, "len_all_1": 61359, "len_abs": 1468, "len_title": 149, "len_sents": 64020, "len_sents_1": 30448, "n_sents": 470, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 100, "L_abs": 1478, "title": "Kinetix: Investigating the Training of General Agents through Open-Ended Physics-Based Control Tasks", "abs": "While large models trained with self-supervised learning on offline datasets have shown remarkable capabilities in text and image domains, achieving the same generalisation for agents that act in sequential decision problems remains an open challenge.\nIn this work, we take a step towards this goal by procedurally generating tens of millions of 2D physics-based tasks and using these to train a general reinforcement learning (RL) agent for physical control.\nTo this end, we introduce Kinetix: an open-ended space of physics-based RL environments that can represent tasks ranging from robotic locomotion and grasping to video games and classic RL environments, all within a unified framework.\nKinetix makes use of our novel hardware-accelerated physics engine Jax2D that allows us to cheaply simulate billions of environment steps during training.\nOur trained agent exhibits strong physical reasoning capabilities in 2D space, being able to zero-shot solve unseen human-designed environments. Furthermore, fine-tuning this general agent on tasks of interest shows significantly stronger performance than training an RL agent *tabula rasa*. This includes solving some environments that standard RL training completely fails at.\nWe believe this demonstrates the feasibility of large scale, mixed-quality pre-training for online RL and we hope that Kinetix will serve as a useful framework to investigate this further.\nWe open-source Jax2D, Kinetix, and our final model weights.", "keywords": [ "reinforcement learning", "open-endedness", "unsupervised environment design", "automatic curriculum learning", "benchmark" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "zCncHdGsOa", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Akiko Takeda", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Professor" }, { "name": "Andi Han", "gender": "Male", "institution": "RIKEN AIP", "country": "JP", "position": "Postdoc" }, { "name": "Pierre-Louis Poirion", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Researcher" } ], "n_formula": 131, "n_formula_1": 30, "n_ref_uni": 53, "n_ref": 94, "n_ref_all": 117, "n_fig": 10, "n_tab": 1, "L_tab": 86, "n_element_tab": 15, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 396, "n_element_tab_1": 5, "formula_len_all": 15183, "formula_len_all_1": 2873, "len_all": 209545, "len_all_1": 72328, "len_abs": 1455, "len_title": 146, "len_sents": 59996, "len_sents_1": 30647, "n_sents": 597, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1335, "title": "Efficient optimization with orthogonality constraint: a randomized Riemannian submanifold method", "abs": "Optimization with orthogonality constraints frequently arise in various fields such as machine learning, signal processing and computer vision. Riemannian optimization offers a powerful framework for solving these problems by equipping the constraint set with a Riemannian manifold structure and performing optimization intrinsically on the manifold. This approach typically involves computing a search direction in the tangent space and updating variables via a retraction operation. However, as the size of the variables increases, the computational cost of the retraction can become prohibitively high, limiting the applicability of Riemannian optimization to large-scale problems. To address this challenge and enhance scalability, we propose a novel approach that restricts each update on a random submanifold, thereby significantly reducing the per-iteration complexity. We introduce two sampling strategies for selecting the random submanifold and theoretically analyze the convergence of the proposed method. We provide convergence results for general nonconvex functions and functions that satisfy Riemannian Polyak\u2013\u0141ojasiewicz condition as well as for stochastic optimization settings. Extensive experiments verify the benefits of the proposed method, showcasing its effectiveness across a wide variety of problem instances.", "keywords": [ "Oprimization", "Orthogonality constraint", "Riemannian optimization", "Stiefel manifold" ], "rating_list": [ 8, 5, 5, 6 ], "soundness_list": [ 4, 4, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "zCZnEXF3bN", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Kfir Yehuda Levy", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Associate Professor" }, { "name": "Tehila Dahan", "gender": "Female", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "PhD student" } ], "n_formula": 115, "n_formula_1": 26, "n_ref_uni": 31, "n_ref": 82, "n_ref_all": 145, "n_fig": 3, "n_tab": 2, "L_tab": 294, "n_element_tab": 7, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 248, "n_element_tab_1": 20, "formula_len_all": 9877, "formula_len_all_1": 1439, "len_all": 193489, "len_all_1": 75118, "len_abs": 1479, "len_title": 143, "len_sents": 55992, "len_sents_1": 31940, "n_sents": 565, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1487, "title": "Do Stochastic, Feel Noiseless: Stable Stochastic Optimization via a Double Momentum Mechanism", "abs": "Optimization methods are crucial to the success of machine learning, with Stochastic Gradient Descent (SGD) serving as a foundational algorithm for training models. However, SGD is often sensitive to the choice of the learning rate, which necessitates extensive hyperparameter tuning. In this work, we introduce a new variant of SGD that brings enhanced stability in two key aspects. First, our method allows the use of the same fixed learning rate to attain optimal convergence rates regardless of the noise magnitude, eliminating the need to adjust learning rates between noiseless and noisy settings. Second, our approach achieves these optimal rates over a wide range of learning rates, significantly reducing sensitivity compared to standard SGD, which requires precise learning rate selection.\nOur key innovation is a novel gradient estimator based on a double-momentum mechanism that combines two recent momentum-based techniques. Utilizing this estimator, we design both standard and accelerated algorithms that are robust to the choice of learning rate. Specifically, our methods attain optimal convergence rates in both noiseless and noisy stochastic convex optimization scenarios without the need for learning rate decay or fine-tuning. We also prove that our approach maintains optimal performance across a wide spectrum of learning rates, underscoring its stability and practicality. Empirical studies further validate the robustness and enhanced stability of our approach.", "keywords": [ "online convex optimization", "stochastic convex optimization" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "zCJqgXnV7f", "primary_area": "reinforcement learning", "n_author": 1, "author_info_list": [ { "name": "En Zhi Tan", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "Undergrad student" } ], "n_formula": 107, "n_formula_1": 15, "n_ref_uni": 43, "n_ref": 102, "n_ref_all": 114, "n_fig": 4, "n_tab": 8, "L_tab": 1874, "n_element_tab": 106, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 505, "n_element_tab_1": 48, "formula_len_all": 12417, "formula_len_all_1": 1227, "len_all": 233013, "len_all_1": 60553, "len_abs": 815, "len_title": 122, "len_sents": 78798, "len_sents_1": 26735, "n_sents": 647, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 825, "title": "Dynamic Elimination For PAC Optimal Item Selection From Relative Feedback", "abs": "We study the problem of best-item identification from relative feedback where a learner adaptively plays subsets of items and receives stochastic feedback in the form of the best item in the set. We propose an algorithm - Dynamic Elimination (DE) - that dynamically prunes sub-optimal items from contention to efficiently identify the best item and show a strong sample complexity upper bound for it. We further formalize the notion of inferred updates to obtain estimates on item win rates without directly playing them by leveraging item correlation information. We propose the Dynamic Elimination by Correlation (DEBC) algorithm as an extension to DE with inferred updates. We show through extensive experiments that DE and DEBC significantly outperform all existing baselines across multiple datasets in various settings.", "keywords": [ "probably approximately correct", "optimal item selection", "relative feedback", "multi armed bandits", "Plackett Luce Model", "Condorcet winner", "Bayesian updates", "active learning" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "zBrjRswpkg", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Haru Kondoh", "gender": "Male", "institution": "Tokyo Institute of Technology", "country": "JP", "position": "MS student" }, { "name": "Kazutoshi Tanaka", "gender": "Male", "institution": "OMRON SINIC X Corporation", "country": "", "position": "Researcher" }, { "name": "Tadashi Kozuno", "gender": "Male", "institution": "OMRON SINIC X", "country": "JP", "position": "Researcher" } ], "n_formula": 40, "n_formula_1": 23, "n_ref_uni": 50, "n_ref": 75, "n_ref_all": 94, "n_fig": 7, "n_tab": 1, "L_tab": 120, "n_element_tab": 8, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 120, "n_element_tab_1": 8, "formula_len_all": 4543, "formula_len_all_1": 2120, "len_all": 129220, "len_all_1": 61173, "len_abs": 708, "len_title": 62, "len_sents": 33051, "len_sents_1": 25333, "n_sents": 326, "n_sents_1": 243, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 711, "title": "Foundation of Scalable Constraint Learning from Human Feedback", "abs": "Constraint learning from human feedback (CLHF) has garnered significant interest in the domain of safe reinforcement learning (RL) due to the challenges associated with designing constraints that elicit desired behaviors. However, a comprehensive theoretical analysis of CLHF is still missing. This paper addresses this gap by establishing a theoretical foundation. Concretely, trajectory-wise feedback, which is the most natural form of feedback, is shown to be helpful only for learning chance constraints. Building on this insight, we propose and theoretically analyze algorithms for CLHF and for solving chance constrained RL problems. Our algorithm is empirically shown to outperform an existing algorithm.", "keywords": [ "RLHF", "RL", "Constraint Learning", "Theoretical Analysis" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 2, 1, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "zBgiCWCxJB", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Guillermo Sapiro", "gender": "unknown", "institution": "Apple", "country": "", "position": "Apple" }, { "name": "Lun Huang", "gender": "unknown", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Qiang Qiu", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "Associate Professor" } ], "n_formula": 90, "n_formula_1": 17, "n_ref_uni": 37, "n_ref": 57, "n_ref_all": 73, "n_fig": 2, "n_tab": 9, "L_tab": 2652, "n_element_tab": 403, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 1707, "n_element_tab_1": 108, "formula_len_all": 4365, "formula_len_all_1": 941, "len_all": 185136, "len_all_1": 69475, "len_abs": 1433, "len_title": 124, "len_sents": 55790, "len_sents_1": 32064, "n_sents": 488, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1437, "title": "SSOLE: Rethinking Orthogonal Low-rank Embedding for Self-Supervised Learning", "abs": "Self-supervised learning (SSL) aims to learn meaningful representations from unlabeled data. Orthogonal Low-rank Embedding (OLE) shows promise for SSL by enhancing intra-class similarity in a low-rank subspace and promoting inter-class dissimilarity in a high-rank subspace, making it particularly suitable for multi-view learning tasks. However, directly applying OLE to SSL poses significant challenges: (1) the virtually infinite number of \"classes\" in SSL makes achieving the OLE objective impractical, leading to representational collapse; and (2) low-rank constraints may fail to distinguish between positively and negatively correlated features, further undermining learning. To address these issues, we propose SSOLE (Self-Supervised Orthogonal Low-rank Embedding), a novel framework that integrates OLE principles into SSL by (1) decoupling the low-rank and high-rank enforcement to align with SSL objectives; and (2) applying low-rank constraints to feature deviations from their mean, ensuring better alignment of positive pairs by accounting for the signs of cosine similarities. Our theoretical analysis and empirical results demonstrate that these adaptations are crucial to SSOLE\u2019s effectiveness. Moreover, SSOLE achieves competitive performance across SSL benchmarks without relying on large batch sizes, memory banks, or dual-encoder architectures, making it an efficient and scalable solution for self-supervised tasks.", "keywords": [ "self-supervised learning", "orthogonal low-rank embedding" ], "rating_list": [ 8, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "zBbZ2vdLzH", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Cheng Shi", "gender": "Male", "institution": "University of Basel", "country": "", "position": "PhD student" }, { "name": "Jonas Linkerh\u00e4gner", "gender": "Male", "institution": "University of Basel", "country": "CH", "position": "PhD student" }, { "name": "Ivan Dokmanic", "gender": "Male", "institution": "University of Basel", "country": "CH", "position": "Associate Professor" } ], "n_formula": 32, "n_formula_1": 10, "n_ref_uni": 60, "n_ref": 129, "n_ref_all": 194, "n_fig": 24, "n_tab": 18, "L_tab": 7456, "n_element_tab": 957, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1762, "n_element_tab_1": 159, "formula_len_all": 3126, "formula_len_all_1": 739, "len_all": 282686, "len_all_1": 69092, "len_abs": 928, "len_title": 113, "len_sents": 76893, "len_sents_1": 29943, "n_sents": 698, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 65, "L_abs": 930, "title": "Joint Graph Rewiring and Feature Denoising via Spectral Resonance", "abs": "In graph learning the graph and the node features both contain noisy information about the node labels. In this paper we propose joint denoising and rewiring (JDR)\u2014an algorithm to jointly rewire the graph and denoise the features, which improves the performance of downstream node classification graph neural nets (GNNs). JDR improves the alignment between the leading eigenspaces of graph and feature matrices. To approximately solve the associated non-convex optimization problem we propose a heuristic that efficiently handles real-world graph datasets with multiple classes and different levels of homophily or heterophily. We theoretically justify JDR in a stylized setting and verify the effectiveness of our approach through extensive experiments on synthetic and real-world graph datasets. The results show that JDR consistently outperforms existing rewiring methods on node classification using GNNs as downstream models.", "keywords": [ "GNNs", "Rewiring", "Denoising", "Spectral Resonance", "cSBM" ], "rating_list": [ 8, 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3, 4 ], "presentation_list": [ 2, 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 2, 3, 4 ], "confidence_list": [ 3, 3, 3, 2, 3 ] }, { "paper_id": "zB6uMznFuZ", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Din-Yin Hsieh", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "MS student" }, { "name": "Guang Cheng", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "Full Professor" }, { "name": "Namjoon Suh", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "Postdoc" }, { "name": "Qitong Luan", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "", "position": "Undergrad student" }, { "name": "Shirong Xu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Researcher" }, { "name": "Shixiang Zhu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Yuning Yang", "gender": "Male", "institution": "FDA", "country": "", "position": "Researcher" } ], "n_formula": 29, "n_formula_1": 8, "n_ref_uni": 49, "n_ref": 123, "n_ref_all": 148, "n_fig": 6, "n_tab": 15, "L_tab": 7108, "n_element_tab": 600, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1923, "n_element_tab_1": 110, "formula_len_all": 2837, "formula_len_all_1": 647, "len_all": 221995, "len_all_1": 75225, "len_abs": 4033, "len_title": 133, "len_sents": 69178, "len_sents_1": 32663, "n_sents": 523, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1569, "title": "TimeAutoDiff: Generation of Heterogeneous Time Series Data via Latent Diffusion Model", "abs": "In this paper, we leverage the power of latent diffusion models to generate synthetic time series tabular data.\nAlong with the temporal and feature correlations, the heterogeneous nature of the feature in the table has been one of the main obstacles in time series tabular data modeling. \nWe tackle this problem by combining the ideas of the variational auto-encoder (VAE) and the denoising diffusion probabilistic model (DDPM).\nOur model named as \\texttt{TimeAutoDiff} has several key advantages including \n(1) \\textit{\\textbf{Generality}}: the ability to handle the broad spectrum of time series tabular data with heterogeneous, continuous only, or categorical only features; \n(2) \\textit{\\textbf{Fast sampling speed}}: entire time series data generation as opposed to the sequential data sampling schemes implemented in the existing diffusion-based models, eventually leading to significant improvements in sampling speed, \n(3) \\textit{\\textbf{Time varying metadata conditional generation}}: the implementation of time series tabular data generation of heterogeneous outputs conditioned on heterogenous, time varying features, enabling scenario exploration across multiple scientific and engineering domains.\n(4) \\textit{\\textbf{Good fidelity and utility guarantees}}: numerical experiments on eight publicly available datasets demonstrating significant improvements over state-of-the-art models in generating time series tabular data, across four metrics measuring fidelity and utility; \nCodes for model implementations are available at the supplementary materials.", "keywords": [ "Time series data", "Tabular data", "Heterogeneous", "Diffusion model", "VAE", "Generative model" ], "rating_list": [ 3, 3, 1, 5 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 4, 3, 1, 3 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 4, 2, 1, 3 ] }, { "paper_id": "zAzzMOaisF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Hao Sun", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Mihaela van der Schaar", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Thomas Pouplin", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Katarzyna Kobalczyk", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 61, "n_ref": 118, "n_ref_all": 216, "n_fig": 5, "n_tab": 8, "L_tab": 3201, "n_element_tab": 320, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1093, "n_element_tab_1": 124, "formula_len_all": 921, "formula_len_all_1": 434, "len_all": 266821, "len_all_1": 83888, "len_abs": 1191, "len_title": 139, "len_sents": 74329, "len_sents_1": 35772, "n_sents": 570, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1195, "title": "LLMs for Generalizable Language-Conditioned Policy Learning under Minimal Data Requirements", "abs": "To develop autonomous agents capable of executing complex, multi-step decision-making tasks as specified by humans in natural language, existing reinforcement learning approaches typically require expensive labeled datasets or access to real-time experimentation. Moreover, conventional methods often face difficulties in generalizing to unseen goals and states, thereby limiting their practical applicability. This paper presents TEDUO, a novel training pipeline for offline language-conditioned policy learning. TEDUO operates on easy-to-obtain, unlabeled datasets and is suited for the so-called in-the-wild evaluation, wherein the agent encounters previously unseen goals and states. To address the challenges posed by such data and evaluation settings, our method leverages the prior knowledge and instruction-following capabilities of large language models (LLMs) to enhance the fidelity of pre-collected offline data and enable flexible generalization to new goals and states. Empirical results demonstrate that the dual role of LLMs in our framework\u2014as data enhancers and generalizers\u2014facilitates both effective and data-efficient learning of generalizable language-conditioned policies.", "keywords": [ "Large Language Models", "Language-conditioned policy", "Offline policy learning", "Decison Making Agent", "Goals generalization", "Domain generalization" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "zAyS5aRKV8", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Animesh Garg", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Songheng Yin", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "MS student" }, { "name": "Steve Easterbrook", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "Full Professor" }, { "name": "Wei Yu", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 22, "n_ref": 46, "n_ref_all": 66, "n_fig": 7, "n_tab": 6, "L_tab": 1106, "n_element_tab": 93, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1152, "n_element_tab_1": 97, "formula_len_all": 319, "formula_len_all_1": 303, "len_all": 113751, "len_all_1": 54795, "len_abs": 1504, "len_title": 78, "len_sents": 32610, "len_sents_1": 25098, "n_sents": 255, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1511, "title": "EgoSim: Egocentric Exploration in Virtual Worlds with Multi-modal Conditioning", "abs": "Recent advancements in video diffusion models have established a strong foundation for developing world models with practical applications. The next challenge lies in exploring how an agent can leverage these foundation models to understand, interact with, and plan within observed environments. This requires adding more controllability to the model, transforming it into a versatile game engine capable of dynamic manipulation and control. To address this, we investigated three key conditioning factors: camera, context frame, and text, identifying limitations in current model designs. Specifically, the fusion of camera embeddings with video features leads to camera control being influenced by those features. Additionally, while textual information compensates for necessary spatiotemporal structures, it often intrudes into already observed parts of the scene. To tackle these issues, we designed the Spacetime Epipolar Attention Layer, which ensures that egomotion generated by the model strictly aligns with the camera\u2019s movement through rigid constraints. Moreover, we propose the CI2V-adapter, which uses camera information to better determine whether to prioritize textual or visual embeddings, thereby alleviating the issue of textual intrusion into observed areas. Through extensive experiments, we demonstrate that our new model EgoSim achieves excellent results on both the RealEstate and newly repurposed Epic-Field datasets. For more results, please refer to https://egosim.github.io/EgoSim/.", "keywords": [ "Controllable video generation", "Egocentric video prediction", "World model" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 2, 4, 3 ], "presentation_list": [ 2, 3, 1 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "zAogQOIphH", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Hai Huang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Minghui Fang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Qian Chen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Shengpeng Ji", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Siqi Zheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Wen Wang", "gender": "unknown", "institution": "Alibaba Group", "country": "", "position": "Senior Staff Algorithm Engineer" }, { "name": "Xize Cheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zehan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Ziyue Jiang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jialung Zuo", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 49, "n_ref": 136, "n_ref_all": 160, "n_fig": 3, "n_tab": 7, "L_tab": 1156, "n_element_tab": 95, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 602, "n_element_tab_1": 32, "formula_len_all": 1527, "formula_len_all_1": 911, "len_all": 174755, "len_all_1": 65700, "len_abs": 1818, "len_title": 146, "len_sents": 51694, "len_sents_1": 33039, "n_sents": 347, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 98, "L_abs": 1841, "title": "ControlSpeech: Towards Simultaneous Zero-shot Speaker Cloning and Zero-shot Language Style Control", "abs": "In this paper, we present ControlSpeech, a text-to-speech (TTS) system capable of fully cloning the speaker's voice and enabling arbitrary control and adjustment of speaking style, merely based on a few seconds of audio prompt and a simple textual style description prompt. Prior zero-shot TTS models only mimic the speaker's voice without further control and adjustment capabilities while prior controllable TTS models cannot perform speaker-specific voice generation. Therefore, ControlSpeech focuses on a more challenging task\u2014a TTS system with controllable timbre, content, and style at the same time. ControlSpeech takes speech prompts, content prompts, and style prompts as inputs and utilizes bidirectional attention and mask-based parallel decoding to capture codec representations corresponding to timbre, content, and style in a discrete decoupling codec space. Moreover, we analyze the many-to-many issue in textual style control and propose the Style Mixture Semantic Density (SMSD) module, which is based on Gaussian mixture density networks, to resolve this problem. The SMSD module enhances the fine-grained partitioning and sampling capabilities of style semantic information and enables speech generation with more diverse styles. To facilitate empirical validations, we make available a controllable model toolkit called ControlToolkit, which includes all source code, a new style controllable dataset VccmDataset, and our replicated competitive baseline models. Our experimental results demonstrate that ControlSpeech exhibits comparable or state-of-the-art (SOTA) performance in terms of controllability, timbre similarity, audio quality, robustness, and generalizability. Ablation studies further validate the necessity of each component in ControlSpeech. Audio samples are available at https://controlspeech.github.io/.", "keywords": [ "text-to-speech", "style control", "discrete codec model" ], "rating_list": [ 5, 3, 5, 5, 8 ], "soundness_list": [ 2, 2, 3, 3, 4 ], "presentation_list": [ 2, 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 2, 4 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "zA0oW4Q4ly", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "David Hyde", "gender": "unknown", "institution": "Vanderbilt University", "country": "US", "position": "Assistant Professor" }, { "name": "Forrest John Laine", "gender": "Male", "institution": "Vanderbilt University", "country": "", "position": "Assistant Professor" }, { "name": "Max Milkert", "gender": "Male", "institution": "The National Renewable Energy Laboratory", "country": "US", "position": "Intern" } ], "n_formula": 45, "n_formula_1": 6, "n_ref_uni": 36, "n_ref": 48, "n_ref_all": 98, "n_fig": 15, "n_tab": 5, "L_tab": 1249, "n_element_tab": 91, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 667, "n_element_tab_1": 41, "formula_len_all": 3238, "formula_len_all_1": 335, "len_all": 179031, "len_all_1": 66425, "len_abs": 1091, "len_title": 157, "len_sents": 61377, "len_sents_1": 30930, "n_sents": 492, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1095, "title": "Compelling ReLU Networks to Exhibit Exponentially Many Linear Regions at Initialization and During Training", "abs": "A neural network with ReLU activations may be viewed as a composition of piecewise linear functions. For such networks, the number of distinct linear regions expressed over the input domain has the potential to scale exponentially with depth, but it is not expected to do so when the initial parameters are chosen randomly. Therefore, randomly initialized models are often unnecessarily large, even when approximating simple functions. To address this issue, we introduce a novel training strategy: we first reparameterize the network weights in a manner that forces the network to exhibit a number of linear regions exponential in depth. Training first on our derived parameters provides an initial solution that can later be refined by directly updating the underlying model weights. This approach allows us to learn approximations of convex, one-dimensional functions that are several orders of magnitude more accurate than their randomly initialized counterparts. We further demonstrate how to extend our approach to multidimensional and non-convex functions, with similar benefits observed.", "keywords": [ "linear regions", "activation regions", "ReLU network", "pretraining", "network initialization" ], "rating_list": [ 5, 8, 8, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "z9j7wctoGV", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Jun Yu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Shuhai Zhang", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yiliao Song", "gender": "Female", "institution": "University of Adelaide", "country": "AU", "position": "Lecturer" }, { "name": "Zhen Fang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Assistant Professor" }, { "name": "Zhenqiao Yuan", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "MS student" } ], "n_formula": 21, "n_formula_1": 18, "n_ref_uni": 51, "n_ref": 149, "n_ref_all": 171, "n_fig": 2, "n_tab": 10, "L_tab": 5910, "n_element_tab": 330, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1583, "n_element_tab_1": 137, "formula_len_all": 2039, "formula_len_all_1": 1710, "len_all": 206501, "len_all_1": 74523, "len_abs": 1281, "len_title": 120, "len_sents": 53720, "len_sents_1": 29553, "n_sents": 513, "n_sents_1": 292, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1360, "title": "Deep Kernel Relative Test for Machine-generated Text Detection", "abs": "Recent studies demonstrate that two-sample test can effectively detect machine-generated texts (MGTs) with excellent adaptation ability to texts generated by newer LLMs. However, two-sample test-based detection relies on the assumption that human-written texts (HWTs) must follow the distribution of seen HWTs. As a result, it tends to make mistakes in identifying HWTs that deviate from the seen HWT distribution, limiting their use in sensitive areas like academic integrity verification. To address this issue, we propose to employ non-parametric kernel relative test to detect MGTs by testing whether it is statistically significant that the distribution of a text to be tested is closer to the distribution of HWTs than to the MGTs' distribution. We further develop a kernel optimisation algorithm in relative test to select the best kernel that can enhance the testing capability for MGT detection. As relative test does not assume that a text to be tested must belong exclusively to either MGTs or HWTs, relative test can largely reduce the false positive error compared to two-sample test, offering significant advantages in practice. Extensive experiments demonstrate the superior performance of our method, compared to state-of-the-art non-parametric and parametric detectors. The code and demo are available: https://github.com/xLearn-AU/R-Detect.", "keywords": [ "Large language models", "Machine-generated text detection", "Relative test" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "z9UBpl4pv5", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Jianqiao Zheng", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "PhD student" }, { "name": "Simon Lucey", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Full Professor" }, { "name": "Xueqian Li", "gender": "unknown", "institution": "University of Adelaide", "country": "AU", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 10, "n_ref_uni": 26, "n_ref": 51, "n_ref_all": 70, "n_fig": 6, "n_tab": 4, "L_tab": 2920, "n_element_tab": 422, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1151, "n_element_tab_1": 160, "formula_len_all": 1919, "formula_len_all_1": 1392, "len_all": 114956, "len_all_1": 67284, "len_abs": 1584, "len_title": 110, "len_sents": 35609, "len_sents_1": 30042, "n_sents": 274, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1594, "title": "Structured Initialization for Attention in Vision Transformers", "abs": "The application of Vision Transformers (ViTs) to new domains where an inductive bias is known but only small datasets are available to train upon is a growing area of interest.\nHowever, training ViT networks on small-scale datasets poses a significant challenge. \nIn contrast, Convolutional Neural Networks (CNNs) have an architectural inductive bias enabling them to perform well on such problems. \nIn this paper, we propose that the architectural bias inherent to CNNs can be reinterpreted as an initialization bias within ViT. \nSpecifically, based on our theoretical findings that the convolutional structures of CNNs allow random impulse filters to achieve performance comparable to their learned counterparts, we design a ``structured initialization'' for ViT with optimization.\nUnlike conventional initialization methods for ViTs, which typically (1) rely on empirical results such as attention weights in pretrained models, (2) focus on the distribution of the attention weights, resulting in unstructured attention maps, our approach is grounded in a solid theoretical analysis, and builds structured attention maps.\nThis key difference in the attention map empowers ViTs to perform equally well on small-scale problems while preserving their structural flexibility for large-scale applications.\nWe show that our method achieves significant performance improvements over conventional ViT initialization methods across numerous small-scale benchmarks including CIFAR-10, CIFAR-100, and SVHN, while maintaining on-par if not better performance on large-scale datasets such as ImageNet-1K.", "keywords": [ "Transformer", "Learning theory", "Initialization", "ConvMixer", "Attention map" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "z9UABOHCZc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "David G Shatwell", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Ishan Rajendrakumar Dave", "gender": "Male", "institution": "Apple", "country": "US", "position": "Intern" }, { "name": "Mubarak Shah", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Swetha Sirnam", "gender": "Female", "institution": "University of Central Florida", "country": "US", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 8, "n_ref_uni": 38, "n_ref": 71, "n_ref_all": 92, "n_fig": 11, "n_tab": 9, "L_tab": 1485, "n_element_tab": 244, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2988, "n_element_tab_1": 135, "formula_len_all": 1714, "formula_len_all_1": 1167, "len_all": 156156, "len_all_1": 67265, "len_abs": 1857, "len_title": 99, "len_sents": 47031, "len_sents_1": 30069, "n_sents": 341, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1864, "title": "GeoTimeCLIP: Unveiling the When and Where of Images", "abs": "Timestamp prediction aims to accurately determine the date and hour at which an image was captured using only visual cues, with applications ranging from image retrieval and metadata correction to digital forensics. In outdoor scenes, this can be inferred from variables such as overall brightness, hue, and shadow positions for hourly estimations, as well as weather patterns or seasonal changes for determining the date. However, these factors vary greatly depending on geographical location, making the challenges of time-of-capture prediction closely related to geo-localization. To address this problem, we introduce GeoTimeCLIP, a novel method capable of simultaneously estimating both the capture time (i.e., hour and month) and geo-location (i.e., GPS coordinates) of an image using a retrieval approach. Our model employs an image encoder, a time encoder, and a location encoder, aligning the time and GPS embeddings with the image embeddings in a continuous high-dimensional feature space. Considering the cyclical nature of days and years, we propose an effective way to represent time using Random Fourier Features. To learn image-time embedding alignment, rather than applying a standard contrastive loss with hard positives and negatives, we propose a more effective metric learning-based objective, which provides soft targets by considering the time difference between samples over a toroidal manifold. We introduce new benchmarks for time prediction, where we show that our jointly optimized time-location-based method outperforms baselines optimized solely for time. We also evaluate our method on existing geo-localization protocols, demonstrating that our approach performs competitively with expert geo-localization methods. Our shared embedding space enables various downstream tasks, such as compositional retrieval and text-based retrieval.", "keywords": [ "time prediction", "geolocalization", "contrastive learning", "metric learning" ], "rating_list": [ 6, 6, 5, 3 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 1 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "z9CCkjVY0h", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Jongmin Yoon", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Juho Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 22, "n_formula_1": 20, "n_ref_uni": 20, "n_ref": 53, "n_ref_all": 88, "n_fig": 8, "n_tab": 4, "L_tab": 246, "n_element_tab": 28, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1082, "n_element_tab_1": 87, "formula_len_all": 1860, "formula_len_all_1": 1626, "len_all": 109922, "len_all_1": 56528, "len_abs": 1115, "len_title": 120, "len_sents": 34455, "len_sents_1": 24516, "n_sents": 252, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1084, "title": "Augmented Flow Matching via Variance Reduction with Auxiliary Variables", "abs": "Flow matching is a simulation-free approach that scalably generates an ODE, in which its path traverses between two different distributions. However, conventional flow matching relies on the training pairs drawn independently, inducing high variance that might slow down training process and degrade the performance upon training. To mitigate this, we propose augmented flow matching, a simple yet efficient framework that can be ubiquitously applied to flow matching with slight modification to the models. We first find that when some auxiliary variables that are correlated to the training data, then they contribute on variance reduction of the flow matching loss estimation, when used together with the training data pair. With this observation, we construct auxiliary variables that are correlated to the training pair, which is obtained by simple and effective linear operation from the input data. Finally, we show that with this simple modification on the training phase, we achieve the improved model flexibility and performance when the ODE is applied on the learned model.", "keywords": [ "generative modeling", "flow matching" ], "rating_list": [ 6, 5, 1, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "z8sxoCYgmd", "primary_area": "datasets and benchmarks", "n_author": 15, "author_info_list": [ { "name": "Baichuan Zhou", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Conghui He", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Hengrui Kang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Jun He", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Junan Zhang", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Junyan Ye", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Tianyi Bai", "gender": "unknown", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Tong Wu", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Weijia Li", "gender": "Female", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Yiping Chen", "gender": "Female", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhizheng Wu", "gender": "Male", "institution": "Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Zihao Wang", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "Zilong Huang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "honglin lin", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 78, "n_ref": 131, "n_ref_all": 171, "n_fig": 9, "n_tab": 24, "L_tab": 13895, "n_element_tab": 2480, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 4471, "n_element_tab_1": 402, "formula_len_all": 380, "formula_len_all_1": 303, "len_all": 496637, "len_all_1": 62220, "len_abs": 6045, "len_title": 0, "len_sents": 141575, "len_sents_1": 25773, "n_sents": 1034, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 86, "L_abs": 1452, "title": "LOKI: A Comprehensive Synthetic Data Detection Benchmark using Large Multimodal Models", "abs": "With the rapid development of AI-generated content, the future internet may be inundated with synthetic data, making the discrimination of authentic and credible multimodal data increasingly challenging. Synthetic data detection has thus garnered widespread attention, and the performance of large multimodal models (LMMs) in this task has attracted significant interest. LMMs can provide natural language explanations for their authenticity judgments, enhancing the explainability of synthetic content detection. Simultaneously, the task of distinguishing between real and synthetic data effectively tests the perception, knowledge, and reasoning capabilities of LMMs. In response, we introduce LOKI, a novel benchmark designed to evaluate the ability of LMMs to detect synthetic data across multiple modalities. LOKI encompasses video, image, 3D, text, and audio modalities, comprising 18K carefully curated questions across 26 subcategories with clear difficulty levels. The benchmark includes coarse-grained judgment and multiple-choice questions, as well as fine-grained anomaly selection and explanation tasks, allowing for a comprehensive analysis of LMMs. We evaluated 22 open-source LMMs and 6 closed-source models on LOKI, highlighting their potential as synthetic data detectors and also revealing some limitations in the development of LMM capabilities. More information about LOKI can be found at https://loki102.github.io/LOKI.github.io/.", "keywords": [ "LMMs\uff1bDeepfake\uff1bMultimodality" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "z8PcUSKXXN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Guixu Lin", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Hongjun Wang", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Weihang Ran", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Yinqiang Zheng", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Zhengwei Yin", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 5, "n_ref_uni": 52, "n_ref": 132, "n_ref_all": 145, "n_fig": 8, "n_tab": 9, "L_tab": 3978, "n_element_tab": 466, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 813, "n_element_tab_1": 102, "formula_len_all": 691, "formula_len_all_1": 179, "len_all": 165162, "len_all_1": 55112, "len_abs": 1585, "len_title": 156, "len_sents": 41840, "len_sents_1": 25294, "n_sents": 277, "n_sents_1": 155, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 107, "L_abs": 1599, "title": "Random Is All You Need: Random Noise Injection on Feature Statistics for Generalizable Deep Image Denoising", "abs": "Recent advancements in generalizable deep image denoising have catalyzed the development of robust noise-handling models. The current state-of-the-art, Masked Training (MT), constructs a masked swinir model which is trained exclusively on Gaussian noise ($\\sigma$=15) but can achieve commendable denoising performance across various noise types (*i.e.* speckle noise, poisson noise). However, this method, while focusing on content reconstruction, often produces over-smoothed images and poses challenges in mask ratio optimization, complicating its integration with other methodologies. In response, this paper introduces RNINet, a novel architecture built on a streamlined encoder-decoder framework to enhance both efficiency and overall performance. Initially, we train a pure RNINet (only simple encoder-decoder) on individual noise types, observing that feature statistics such as mean and variance shift in response to different noise conditions. Leveraging these insights, we incorporate a noise injection block that injects random noise into feature statistics within our framework, significantly improving generalization across unseen noise types. Our framework not only simplifies the architectural complexity found in MT but also delivers superior performance. Comprehensive experimental evaluations demonstrate that our method outperforms MT in various unseen noise conditions in terms of denoising effectiveness and computational efficiency (lower MACs and GPU memory usage), achieving up to 10 times faster inference speeds and underscoring it's capability for large scale deployments.", "keywords": [ "Image Denoising", "Low-Level Vision", "Generalization Problem" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "z7QAz5y8Uz", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Rudrasis Chakraborty", "gender": "Male", "institution": "Lawrence Livermore National Labs", "country": "", "position": "Researcher" }, { "name": "Sotirios Panagiotis Chytas", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Vikas Singh", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Professor" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 66, "n_ref": 120, "n_ref_all": 146, "n_fig": 6, "n_tab": 5, "L_tab": 1170, "n_element_tab": 94, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2121, "n_element_tab_1": 149, "formula_len_all": 548, "formula_len_all_1": 718, "len_all": 190542, "len_all_1": 74167, "len_abs": 1526, "len_title": 102, "len_sents": 50196, "len_sents_1": 32789, "n_sents": 358, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1357, "title": "FoGE: Fock Space inspired encoding for graph prompting", "abs": "Recent results show that modern Large Language Models (LLM) are indeed capable of understanding and answering questions about structured data such as graphs. Existing proposals often use some description of the graph to create an ``augmented'' prompt fed to the LLM. For a chosen class of graphs, if a well-tailored graph encoder is deployed to play together with a pre-trained LLM, the model can answer graph-related questions well. Existing solutions to graph-based prompts range from graph serialization to graph transformers. In this work, we show that the use of a parameter-free graph encoder based on Fock space representations, a concept borrowed from mathematical physics, is remarkably versatile in this problem setting. The simple construction, inherited directly from the theory with a few small adjustments, can provide rich and informative graph encodings, for a wide range of different graphs. We investigate the use of this idea for prefix-tuned prompts leveraging the capabilities of a pre-trained, frozen LLM. The modifications lead to a model that can answer graph-related questions -- from simple graphs to proteins to hypergraphs -- effectively and with minimal, if any, adjustments to the architecture. Our work significantly simplifies existing solutions and generalizes well to multiple different graph-based structures effortlessly.", "keywords": [ "llm", "prefix tuning", "graph", "graph encoding", "geometric algebra", "Fock space" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 1, 3, 4 ] }, { "paper_id": "z7PhIgVmZU", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Baoming Zhang", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Leonid Karlinsky", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Rogerio Feris", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Research Manager" }, { "name": "Sarthak Kumar Maharana", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Yunhui Guo", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 43, "n_ref": 142, "n_ref_all": 176, "n_fig": 7, "n_tab": 9, "L_tab": 5793, "n_element_tab": 261, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 313, "n_element_tab_1": 39, "formula_len_all": 370, "formula_len_all_1": 370, "len_all": 161714, "len_all_1": 63053, "len_abs": 1394, "len_title": 95, "len_sents": 50754, "len_sents_1": 27868, "n_sents": 403, "n_sents_1": 223, "venue": "ICLR2025", "decision": "withdraw", "L_title": 47, "L_abs": 1434, "title": "BAT-CLIP: Bimodal Test-Time Adaptation for CLIP", "abs": "Although open-vocabulary classification models like Contrastive Language Image Pretraining (CLIP) have demonstrated strong zero-shot learning capabilities, their robustness to common image corruptions remains poorly understood. Through extensive experiments, we show that zero-shot CLIP lacks robustness to common image corruptions at increasing severity levels during test time, necessitating the adaptation of CLIP to unlabeled corrupted images using test-time adaptation (TTA). However, we found that existing TTA methods have severe limitations in adapting CLIP due to their $\\textit{unimodal}$ nature. To address these limitations, we propose $\\textbf{BAT-CLIP}$, a $\\textit{bimodal}$ TTA method specially designed to improve CLIP's robustness to common image corruptions. The key insight of our approach is not only to adapt the visual encoders for better image feature extraction but also to strengthen the alignment between image and text features by promoting a stronger association between the image class prototype, computed using pseudo-labels, and the corresponding text feature. We evaluate our approach on benchmark image corruption datasets and achieve state-of-the-art results in TTA for CLIP, specifically for domains involving image corruptions. Particularly, with a ViT-B/16 vision backbone, we obtain mean accuracy improvements of 9.7\\%, 5.94\\%, and 5.12\\% for CIFAR-10C, CIFAR-100C, and ImageNet-C, respectively.", "keywords": [ "Test-Time Adaptation", "CLIP", "Robustness" ], "rating_list": [ 6, 5, 8, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "z7JBs8UOLI", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Ashok Cutkosky", "gender": "unknown", "institution": "Boston University", "country": "", "position": "Assistant Professor" }, { "name": "Jiujia Zhang", "gender": "Female", "institution": "Boston University", "country": "", "position": "PhD student" } ], "n_formula": 128, "n_formula_1": 34, "n_ref_uni": 41, "n_ref": 80, "n_ref_all": 145, "n_fig": 2, "n_tab": 15, "L_tab": 3416, "n_element_tab": 360, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 674, "n_element_tab_1": 41, "formula_len_all": 16312, "formula_len_all_1": 2861, "len_all": 207306, "len_all_1": 65968, "len_abs": 824, "len_title": 95, "len_sents": 55867, "len_sents_1": 26742, "n_sents": 512, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 976, "title": "Unconstrained Robust Online Convex Optimization", "abs": "This paper addresses online learning with ''corrupted'' feedback. Our learner is provided with potentially corrupted gradients $\\tilde g_t$ instead of the ''true'' gradients $g_t$. We make no assumptions about how the corruptions arise: they could be the result of outliers, mislabeled data, or even malicious interference. We focus on the difficult ''unconstrained'' setting in which our algorithm must maintain low regret with respect to any comparison point $\\||u\\|| \\in \\mathbb{R}^d$. Perhaps surprisingly, the unconstrained setting is significantly more challenging as existing algorithms suffer extremely high regret even with very tiny amounts of corruption (which is not true in the case of a bounded domain). Our algorithms guarantee regret $ \\||u\\||G (\\sqrt{T} + k) $ when Lipschitz constant $G \\ge \\max_t \\||g_t\\||$ is known, where $k$ is a measure of the total amount of corruption. When $G$ is unknown and incur an extra additive penalty of $(\\||u\\||^2+G^2) k$.", "keywords": [ "online learning", "online convex optimization", "adversarial corruption", "comparator adaptive", "parameter-free", "unconstrained domain" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 1, 4, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "z6qmomJW91", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "David Barber", "gender": "Male", "institution": "University College London", "country": "", "position": "Associate Professor" }, { "name": "Harry Jake Cunningham", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Kai Biegun", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Rares Dolga", "gender": "unknown", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 23, "n_ref_uni": 27, "n_ref": 71, "n_ref_all": 86, "n_fig": 5, "n_tab": 2, "L_tab": 730, "n_element_tab": 21, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 823, "n_element_tab_1": 41, "formula_len_all": 1721, "formula_len_all_1": 1213, "len_all": 105175, "len_all_1": 63712, "len_abs": 843, "len_title": 95, "len_sents": 32499, "len_sents_1": 26887, "n_sents": 241, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 47, "L_abs": 846, "title": "RotRNN: Modelling Long Sequences with Rotations", "abs": "Linear recurrent neural networks, such as State Space Models (SSMs) and Linear Recurrent Units (LRUs), have recently shown state-of-the-art performance on long sequence modelling benchmarks. Despite their success, their empirical performance is not well understood and they come with a number of drawbacks, most notably their complex initialisation and normalisation schemes. In this work, we address some of these issues by proposing RotRNN \u2013 a linear recurrent model which utilises the convenient properties of rotation matrices. We show that RotRNN provides a simple and efficient model with a robust normalisation procedure, and a practical implementation that remains faithful to its theoretical derivation. RotRNN also achieves competitive performance to state-of-the-art linear recurrent models on several long sequence modelling datasets.", "keywords": [ "Sequence Modelling", "Recurrent Neural Networks", "State Space Models", "Long Sequences" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 3, 2, 3 ] }, { "paper_id": "z5uVAKwmjf", "primary_area": "applications to robotics, autonomy, planning", "n_author": 14, "author_info_list": [ { "name": "Bang Liu", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Assistant Professor" }, { "name": "Bingnan Zheng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Fengwei Teng", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Intern" }, { "name": "Jiaqi Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Jiayi Zhang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jinlin Wang", "gender": "Male", "institution": "Xiamen University", "country": "", "position": "MS student" }, { "name": "Jinyu Xiang", "gender": "Male", "institution": "DeepWisdom", "country": "CN", "position": "Researcher" }, { "name": "Mingchen Zhuge", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Sirui Hong", "gender": "Female", "institution": "DeepWisdom", "country": "CN", "position": "Researcher" }, { "name": "Xin Cheng", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "MS student" }, { "name": "Xiong-Hui Chen", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Yuyu Luo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhaoyang Yu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Undergrad student" }, { "name": "Alexander Wu", "gender": "Male", "institution": "DeepWisdom", "country": "", "position": "CEO" } ], "n_formula": 16, "n_formula_1": 7, "n_ref_uni": 32, "n_ref": 85, "n_ref_all": 95, "n_fig": 5, "n_tab": 10, "L_tab": 12288, "n_element_tab": 791, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1556, "n_element_tab_1": 116, "formula_len_all": 1276, "formula_len_all_1": 314, "len_all": 217563, "len_all_1": 60094, "len_abs": 1281, "len_title": 93, "len_sents": 59953, "len_sents_1": 27716, "n_sents": 436, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 45, "L_abs": 1298, "title": "AFlow: Automating Agentic Workflow Generation", "abs": "Large language models (LLMs) have demonstrated remarkable potential in solving complex tasks across diverse domains, typically by employing agentic workflows that follow detailed instructions and operational sequences. However, constructing these workflows requires significant human effort, limiting scalability and generalizability. Recent research has sought to automate the generation and optimization of these workflows, but existing methods still rely on initial manual setup and fall short of achieving fully automated and effective workflow generation. To address this challenge, we reformulate workflow optimization as a search problem over code-represented workflows, where LLM-invoking nodes are connected by edges. We introduce \\textbf{AFlow}, an automated framework that efficiently explores this space using Monte Carlo Tree Search, iteratively refining workflows through code modification, tree-structured experience, and execution feedback. Empirical evaluations across six benchmark datasets demonstrate AFlow's efficacy, yielding a 5.7\\% average improvement over state-of-the-art baselines. Furthermore, AFlow enables smaller models to outperform GPT-4o on specific tasks at 4.55\\% of its inference cost in dollars. The code will be made available as open-source upon publication.", "keywords": [ "LLM Agent; Prompt Optimization; Workflow Generation" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 3, 4, 4 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "z5UZZjXFc9", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Xiaohan Qin", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Intern" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Victor Wang", "gender": "Male", "institution": "IAAR", "country": "CN", "position": "Researcher" } ], "n_formula": 57, "n_formula_1": 19, "n_ref_uni": 33, "n_ref": 132, "n_ref_all": 148, "n_fig": 6, "n_tab": 12, "L_tab": 5778, "n_element_tab": 890, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 3146, "n_element_tab_1": 477, "formula_len_all": 4380, "formula_len_all_1": 735, "len_all": 172487, "len_all_1": 77499, "len_abs": 1422, "len_title": 158, "len_sents": 48929, "len_sents_1": 31366, "n_sents": 408, "n_sents_1": 231, "venue": "ICLR2025", "decision": "withdraw", "L_title": 109, "L_abs": 1428, "title": "Rethinking Fairness Representation in Multi-Task Learning: a Performance-Informed Variance Reduction Approach", "abs": "Multi-task learning (MTL) can leverage shared knowledge across tasks to improve data efficiency and generalization performance, and has been applied in various scenarios. However, task imbalance remains a major challenge for existing MTL methods. While the prior works have attempted to mitigate inter-task unfairness through loss-based and gradient-based strategies, they still exhibit imbalanced performance across tasks on common benchmarks.\nThis key observation motivates us to consider performance-level information as an explicit fairness indicator, which can more accurately reflect the current optimization status of each task, and accordingly help to adjust the gradient aggregation process.\nSpecifically, we utilize the performance variance among tasks as the fairness indicator and introduce a dynamic weighting strategy to gradually reduce the performance variance. \nBased on this, we propose PIVRG, a novel performance-informed variance reduction gradient aggregation approach.\nExtensive experiments show that PIVRG achieves state-of-the-art performance across various benchmarks, spanning both supervised learning and reinforcement learning tasks with task numbers ranging from 2 to 40. Results from the ablation study also show that our approach can be integrated into existing methods, significantly enhancing their performance while reducing the variance in task performance, thus achieving fairer optimization.", "keywords": [ "Multi-Task Learning", "Fair Optimization", "Dynamic Weighting Strategy" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "z4rBSPep64", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Caren Han", "gender": "Female", "institution": "University of Melbourne", "country": "AU", "position": "Associate Professor" }, { "name": "Hyunsuk Chung", "gender": "Male", "institution": "FortifyEdge", "country": "", "position": "CTO" }, { "name": "Yihao Ding", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Postdoc" }, { "name": "Zechuan Li", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 14, "n_ref_uni": 35, "n_ref": 72, "n_ref_all": 99, "n_fig": 9, "n_tab": 16, "L_tab": 7938, "n_element_tab": 555, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 2723, "n_element_tab_1": 178, "formula_len_all": 879, "formula_len_all_1": 976, "len_all": 180610, "len_all_1": 73037, "len_abs": 1007, "len_title": 132, "len_sents": 55853, "len_sents_1": 32450, "n_sents": 378, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1012, "title": "DAViD: Domain Adaptive Visually-Rich Document Understanding with Synthetic Insights", "abs": "Visually-Rich Documents (VRDs), encompassing elements like charts, tables, and references, convey complex information across various fields. However, extracting information from these rich documents is labor-intensive, especially given their inconsistent formats and domain-specific requirements. While pretrained models for VRD Understanding have progressed, their reliance on large, annotated datasets limits scalability. This paper introduces the Domain Adaptive Visually-rich Document Understanding (DAViD) framework, which utilises machine-generated synthetic data for domain adaptation. DAViD integrates fine-grained and coarse-grained document representation learning and employs synthetic annotations to reduce the need for costly manual labelling. By leveraging pretrained models and synthetic data, DAViD achieves competitive performance with minimal annotated datasets. Extensive experiments validate DAViD\u2019s effectiveness, demonstrating its ability to efficiently adapt to domain-specific VRDU tasks.", "keywords": [ "Visually-Rich Documents", "Visually-Rich Document Understanding", "Domain Adaption" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "z4bfNsrum4", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Simran Ketha", "gender": "Female", "institution": "BITS Pilani, Birla Institute of Technology and Science", "country": "IN", "position": "PhD student" }, { "name": "Venkatakrishnan Ramaswamy", "gender": "Male", "institution": "BITS Pilani, Birla Institute of Technology and Science", "country": "", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 47, "n_ref": 60, "n_ref_all": 92, "n_fig": 19, "n_tab": 9, "L_tab": 5731, "n_element_tab": 241, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1281, "formula_len_all_1": 0, "len_all": 144034, "len_all_1": 52089, "len_abs": 1638, "len_title": 113, "len_sents": 44800, "len_sents_1": 26832, "n_sents": 305, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1645, "title": "Decoding Generalization from Memorization in Deep Neural Networks", "abs": "Overparameterized Deep Neural Networks that generalize well have been key to the dramatic success of Deep Learning in recent years. The reasons for their remarkable ability to generalize are not well understood yet. It has also been known that deep networks possess the ability to memorize training data, as evidenced by perfect or high training accuracies on models trained with corrupted data that have class labels shuffled to varying degrees. Concomitantly, such models are known to generalize poorly, i.e. they suffer from poor test accuracies, due to which it is thought that the act of memorizing substantially degrades the ability to generalize. It has, however, been unclear why the poor generalization that accompanies such memorization, comes about. One possibility is that in the process of training with corrupted data, the layers of the network irretrievably re-organize their representations in a manner that makes generalization difficult. The other possibility is that the network retains significant ability to generalize, but the trained network somehow \u201cchooses\u201d to readout in a manner that is detrimental to generalization. Here, we provide evidence for the latter possibility by demonstrating, empirically, that such models possess information in their representations for substantially improved generalization, even in the face of memorization. Furthermore, such generalization abilities can be easily decoded from the internals of the trained model, and we build a technique to do so from the outputs of specific layers of the network. We demonstrate results on multiple models trained with a number of standard datasets.", "keywords": [ "Generalization", "Memorization" ], "rating_list": [ 6, 3, 6, 3, 1 ], "soundness_list": [ 3, 1, 3, 2, 1 ], "presentation_list": [ 2, 2, 3, 3, 1 ], "contribution_list": [ 3, 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "z4Ho599uOL", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Ararat Harutyunyan", "gender": "Male", "institution": "Universit\u00e9 Paris-Dauphine - PSL", "country": "", "position": "Assistant Professor" }, { "name": "Henrik Abgaryan", "gender": "Male", "institution": "Universit\u00e9 Paris-Dauphine - PSL", "country": "FR", "position": "PhD student" }, { "name": "Tristan Cazenave", "gender": "Male", "institution": "Universit\u00e9 Paris-Dauphine (Paris IX)", "country": "FR", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 15, "n_ref": 28, "n_ref_all": 36, "n_fig": 1, "n_tab": 2, "L_tab": 1043, "n_element_tab": 75, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 1043, "n_element_tab_1": 75, "formula_len_all": 36, "formula_len_all_1": 36, "len_all": 72480, "len_all_1": 71275, "len_abs": 1079, "len_title": 99, "len_sents": 22986, "len_sents_1": 22986, "n_sents": 170, "n_sents_1": 170, "venue": "ICLR2025", "decision": "withdraw", "L_title": 51, "L_abs": 1088, "title": "STARJOB: DATASET FOR LLM-DRIVEN JOB SHOP SCHEDULING", "abs": "The Job Shop Scheduling Problem (JSSP) presents a significant challenge in opti-\nmizing production processes. This problem requires efficient allocation of jobs to\na limited number of machines while minimizing total processing time (makespan).\nAlthough recent advancements in artificial intelligence have produced promising\nsolutions, such as reinforcement learning and graph neural networks, this paper\ninvestigates the potential of Large Language Models (LLMs) for addressing JSSP.\nWe introduce the first supervised 120k dataset called Starjob specifically designed\nto train LLMs for JSSP and we subsequently fintune the LLaMA 8B model on\nthis dataset using Lora. We compare the average makespan gap of our end-to-\nend LLM-based scheduling method with that of the most widely used priority\ndispatching rules (PDRs) and neural methods such as L2D. Surprisingly, our find-\nings indicate that LLM-based scheduling not only surpasses traditional PDRs but\nalso achieves on average 11.28% on DMU and 3.29% gap improvement on the\nTailard benchmarks compared to the state-of-the-art L2D method.", "keywords": [ "JSSP", "Large Language Models", "supervised dataset", "Starjob", "artificial intelligence", "sampling method", "LLM" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 4, 5, 2 ] }, { "paper_id": "z3vplLsIve", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Haoru Tan", "gender": "Male", "institution": "HKU", "country": "HK", "position": "PhD student" }, { "name": "Shaofeng Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Sitong Wu", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Zhibin Wang", "gender": "Male", "institution": "INF Tech", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 46, "n_ref": 120, "n_ref_all": 123, "n_fig": 0, "n_tab": 2, "L_tab": 2482, "n_element_tab": 359, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 2482, "n_element_tab_1": 359, "formula_len_all": 407, "formula_len_all_1": 407, "len_all": 126316, "len_all_1": 76788, "len_abs": 1019, "len_title": 104, "len_sents": 35644, "len_sents_1": 34138, "n_sents": 274, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1022, "title": "Learn to Synthesize Compact Datasets by Matching Effects", "abs": "The emerging field of data distillation aims to compress large datasets by aligning synthetic and real data representations to create a highly informative dataset. The optimization objectives of data distillation focus on aligning representations by using process alignment methods such as trajectory and gradient matching. However, this approach is limited by the strict alignment of intermediate quantities between synthetic and real data and the mismatch between their optimization trajectories. To address these limitations, a new data distillation method called effect alignment is proposed, which aims to only push for the consistency of endpoint training results. The approach uses classification tasks to estimate the impact of replacing real training samples with synthetic data, which helps to learn a synthetic dataset that can replace the real dataset and achieve effect alignment. The method is efficient and does not require costly mechanisms, and satisfactory results have been achieved through experiments.", "keywords": [ "Deep Learning", "Dataset Distillation" ], "rating_list": [ 3, 1, 5, 5 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "z3KmG5JIN4", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Amit Finkman Noah", "gender": "unknown", "institution": "Ben Gurion University of the Negev", "country": "IL", "position": "MS student" }, { "name": "Asaf Shabtai", "gender": "Male", "institution": "Ben-Gurion University of the Negev", "country": "IL", "position": "Full Professor" }, { "name": "Dudu Mimran", "gender": "Male", "institution": "Cyber @ BGU", "country": "IL", "position": "Researcher" }, { "name": "Eden Bar Kochva", "gender": "Female", "institution": "Ben Gurion University of the Negev", "country": "IL", "position": "MS student" }, { "name": "Inbar Maimon", "gender": "Female", "institution": "Ben Gurion University of the Negev", "country": "IL", "position": "Researcher" }, { "name": "Yuval Elovici", "gender": "Male", "institution": "Ben Gurion University of the Negev, Technion", "country": "IL", "position": "Full Professor" }, { "name": "avishag shapira", "gender": "Female", "institution": "Ben-Gurion University of the Negev", "country": "IL", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 27, "n_ref": 41, "n_ref_all": 66, "n_fig": 4, "n_tab": 9, "L_tab": 1567, "n_element_tab": 93, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 617, "n_element_tab_1": 57, "formula_len_all": 270, "formula_len_all_1": 244, "len_all": 141237, "len_all_1": 75788, "len_abs": 1356, "len_title": 70, "len_sents": 52512, "len_sents_1": 38645, "n_sents": 360, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1351, "title": "CodeCloak: A Method for Mitigating Code Leakage by LLM Code Assistants", "abs": "LLM-based code assistants are becoming increasingly popular among developers.\nThese tools help developers improve their coding efficiency and reduce errors by providing real-time suggestions based on the developer\u2019s codebase. \nWhile beneficial, the use of these tools can inadvertently expose the developer\u2019s proprietary code to the code assistant service provider during the development process. \nIn this work, we propose a method to mitigate the risk of code leakage when using LLM-based code assistants. CodeCloak is a novel deep reinforcement learning agent that manipulates the prompts before sending them to the code assistant service.\nCodeCloak aims to achieve the following two contradictory goals: (i) minimizing code leakage, while (ii) preserving relevant and useful suggestions for the developer. \nOur evaluation, employing StarCoder and Code Llama, LLM-based code assistants models, demonstrates CodeCloak\u2019s effectiveness on a diverse set of code repositories of varying sizes, as well as its transferability across different models.\nWe also designed a method for reconstructing the developer\u2019s original codebase from code segments sent to the code assistant service (i.e., prompts) during the development process, to thoroughly analyze code leakage risks and evaluate the effectiveness of CodeCloak under practical development scenarios.", "keywords": [ "privacy", "DRL", "LLM", "code assistant", "generative models" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "z3DMFpaP6m", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Hang Chen", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Wenya Wang", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Xinyu Yang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Jiaying Zhu", "gender": "Female", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 43, "n_ref": 66, "n_ref_all": 96, "n_fig": 6, "n_tab": 3, "L_tab": 1921, "n_element_tab": 137, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 620, "n_element_tab_1": 50, "formula_len_all": 1211, "formula_len_all_1": 1044, "len_all": 145237, "len_all_1": 70816, "len_abs": 1536, "len_title": 113, "len_sents": 40879, "len_sents_1": 30516, "n_sents": 292, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1545, "title": "On the Entropy of Language Models in Getting Semantic from Tokens", "abs": "Large language models (LLMs) are widely recognized for their exceptional capacity to capture semantic meaning. Yet, there remains no established metric to quantify this capability. In this work, we introduce a quantitative metric, Information Emergence (IE), designed to measure LLMs\u2019 ability to extract semantics from input tokens. We formalize \u201csemantics\u201d as the meaningful information abstracted from a sequence of tokens and, leveraging information theory, quantify this through comparing the reduction in entropy observed for a sequence of tokens (macro-level) and individual tokens (micro-level). To achieve this, we design a light-weight estimator to compute the mutual information at both micro and macro levels for each transformer layer, which is agnostic to different tasks and language model architectures. We apply IE in both synthetic in-context learning (ICL) scenarios and natural sentence contexts. Experiments show a high-level informativeness of our metric reflected in semantic faithfulness, sensitivity, and connection with emergence. In addition, we highlight some interesting findings: 1) IE explains why ICL offers clearer semantics and benefits compared to natural text through changes\nin entropy. 2) We could associate certain hallucination phenomenon with increased variance in IE. 3) IE can effectively differentiate between human-written and LLM generated text, proving especially useful for extremely large and closed-source language models. Our codes are available at: https://anonymous.4open.science/r/Emergence/.", "keywords": [ "LLM evaluation" ], "rating_list": [ 5, 1, 3 ], "soundness_list": [ 3, 1, 1 ], "presentation_list": [ 2, 1, 1 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 2, 3, 3 ] }, { "paper_id": "z2z9suDRjw", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Darko Drakulic", "gender": "Male", "institution": "Naver Labs Europe", "country": "FR", "position": "Researcher" }, { "name": "Jean-Marc Andreoli", "gender": "Male", "institution": "Naver Labs Europe", "country": "FR", "position": "Researcher" }, { "name": "Sofia Michel", "gender": "unknown", "institution": "Naver Labs Europe", "country": "", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 61, "n_ref": 112, "n_ref_all": 123, "n_fig": 4, "n_tab": 4, "L_tab": 2430, "n_element_tab": 436, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 892, "n_element_tab_1": 164, "formula_len_all": 287, "formula_len_all_1": 287, "len_all": 190047, "len_all_1": 67645, "len_abs": 1469, "len_title": 109, "len_sents": 56193, "len_sents_1": 34464, "n_sents": 419, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1501, "title": "GOAL: A Generalist Combinatorial Optimization Agent Learner", "abs": "Machine Learning-based heuristics have recently shown impressive performance in solving a variety of hard combinatorial optimization problems (COPs). However they generally rely on a separate neural model, specialized and trained for each single problem. Any variation of a problem requires adjustment of its model and re-training from scratch. In this paper, we propose GOAL (for Generalist combinatorial Optimization Agent Learner), a generalist model capable of efficiently solving multiple COPs and which can be fine-tuned to solve new COPs. GOAL consists of a single backbone plus light-weight problem-specific adapters for input and output processing. The backbone is based on a new form of mixed-attention blocks which allows to handle problems defined on graphs with arbitrary combinations of node, edge and instance-level features. Additionally, problems which involve heterogeneous types of nodes or edges are handled through a novel multi-type transformer architecture, where the attention blocks are duplicated to attend the meaningful combinations of types while relying on the same shared parameters. We train GOAL on a set of routing, scheduling and classic graph problems and show that it is only slightly inferior to the specialized baselines while being the first multi-task model that solves a wide range of COPs. Finally we showcase the strong transfer learning capacity of GOAL by fine-tuning it on several new problems. Our code is available at https://github.com/naver/goal-co .", "keywords": [ "neural combinatorial optimization", "generalist models", "transfer learning", "fine tuning" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 4, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "z2WCyBO923", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Jia-Jiun Yao", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "Sheng Feng Yu", "gender": "Male", "institution": "National Chiao Tung University", "country": "TW", "position": "PhD student" }, { "name": "Wei-Chen Chiu", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 24, "n_ref": 86, "n_ref_all": 99, "n_fig": 1, "n_tab": 5, "L_tab": 782, "n_element_tab": 83, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 427, "n_element_tab_1": 33, "formula_len_all": 1188, "formula_len_all_1": 1136, "len_all": 110497, "len_all_1": 61776, "len_abs": 1656, "len_title": 121, "len_sents": 42655, "len_sents_1": 31827, "n_sents": 249, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1663, "title": "Four eyes see more than two: Dataset Distillation with Mixture-of-Experts", "abs": "The ever-growing size of datasets in deep learning presents a significant challenge in terms of training efficiency and computational cost. Dataset distillation (DD) has emerged as a promising approach to address this challenge by generating compact synthetic datasets that retain the essential information of the original data. However, existing DD methods often suffer from performance degradation when transferring distilled datasets across different network architectures (i.e. the model utilizing distilled dataset for further training is different from the one used in dataset distillation). To overcome this limitation, we propose a novel mixture-of-experts framework for dataset distillation. Our goal focuses on promoting diversity within the distilled dataset by distributing the distillation tasks to multiple expert models. Each expert specializes in distilling a distinct subset of the dataset, encouraging them to capture different aspects of the original data distribution. To further enhance diversity, we introduce a distance correlation minimization strategy to encourage the experts to learn distinct representations. Moreover, during the testing stage (where the distilled dataset is used for training a new model), the mixup-based fusion strategy is applied to better leverage the complementary information captured by each expert. Through extensive experiments, we demonstrate that our framework effectively mitigates the issue of cross-architecture performance degradation in dataset distillation, particularly in low-data regimes, leading to more efficient and versatile deep learning models while being trained upon the distilled dataset.", "keywords": [ "dataset distillation", "mixture-of-experts" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "z2VBHpRT14", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Jiaping Xiao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Mir Feroskhan", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "Assistant Professor" }, { "name": "Qianlei Jia", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Yuhang Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" }, { "name": "L Bai", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "ZHANG RANGYA", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 16, "n_ref": 28, "n_ref_all": 54, "n_fig": 10, "n_tab": 14, "L_tab": 6071, "n_element_tab": 407, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 2878, "n_element_tab_1": 219, "formula_len_all": 235, "formula_len_all_1": 235, "len_all": 137524, "len_all_1": 63626, "len_abs": 1658, "len_title": 139, "len_sents": 45224, "len_sents_1": 28907, "n_sents": 325, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1679, "title": "SpaceSet: A Large-scale Realistic Space-based Image Dataset for Space Situational Awareness", "abs": "Space situational awareness (SSA) plays an imperative role in maintaining safe space operations, especially given the increasingly congested space traffic around Earth. Space-based SSA offers a flexible and lightweight solution compared to traditional ground-based SSA. With advanced machine learning approaches, space-based SSA can extract features from high-resolution images in space to detect and track resident space objects (RSOs). However, existing spacecraft image datasets, such as SPARK, fall short of providing realistic camera observations, rendering the derived algorithms unsuitable for real SSA systems. In this research, we introduce SpaceSet, a large-scale realistic space-based image dataset for SSA. We consider accurate space orbit dynamics and a physical camera model with various noise distributions, generating images at the photon level. To extend the available observation window, four overlapping cameras are simulated with a fixed rotation angle. SpaceSet includes images of RSOs observed from $19 km$ to $63,000 km$, captured by a tracker operating in LEO, MEO, and GEO orbits over a period of $5,000$ seconds. Each image has a resolution of $4418 \\times 4418$ pixels, providing detailed features for developing advanced SSA approaches. We split the dataset into three subsets: SpaceSet-100, SpaceSet-5000, and SpaceSet-full, catering to various image processing applications. The SpaceSet-full corpus includes a comprehensive data-loader with $781.5GB$ of images and $25.9MB$ of ground truth labels. We also benchmark detection and tracking algorithms on the SpaceSet-100 dataset using a specified splitting method to accelerate the training process.", "keywords": [ "space situational awareness", "object detection and tracking", "space image dataset", "high resolution image" ], "rating_list": [ 10, 5, 5, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 4, 2, 2, 3 ], "confidence_list": [ 1, 3, 4, 2 ] }, { "paper_id": "z2QdVmhtAP", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Ambuj Singh", "gender": "unknown", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" }, { "name": "Christos Zangos", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Danish Ebadulla", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Thomas Christopher Sprague", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 1, "n_ref_uni": 33, "n_ref": 55, "n_ref_all": 71, "n_fig": 6, "n_tab": 4, "L_tab": 1431, "n_element_tab": 41, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 168, "n_element_tab_1": 8, "formula_len_all": 508, "formula_len_all_1": 34, "len_all": 130884, "len_all_1": 54440, "len_abs": 1067, "len_title": 85, "len_sents": 44653, "len_sents_1": 27550, "n_sents": 358, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1060, "title": "Efficient Multi Subject Visual Reconstruction from fMRI Using Aligned Representations", "abs": "Reconstructing visual images from fMRI data presents a challenging task, particularly when dealing with limited data and compute availability. This work introduces a novel approach to fMRI-based visual image reconstruction using a subject-agnostic common representation space. We show that subjects' brain signals naturally align in this common space during training, without the need for explicit alignment. This is leveraged to demonstrate that aligning subject-specific adapters to a reference subject is significantly more efficient than traditional end-to-end training methods. Our approach excels in low-data scenarios, where training the adapter with limited data achieves faster and better performance. We also introduce a novel method to select the most representative subset of images for a new subject, allowing for fine-tuning with 40\\% less data while maintaining performance. These advancements make fMRI data collection more efficient and practical, reducing the burden on subjects and improving the generalization of fMRI reconstruction models.", "keywords": [ "fMRI", "Computational Neuroscience", "Neuroimaging", "Diffusion", "CLIP", "alignment", "neuroAI" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 3, 1, 1 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 1, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "z21DkDDdgq", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Jennie Si", "gender": "unknown", "institution": "Arizona State University", "country": "US", "position": "Full Professor" }, { "name": "Brent Abraham Wallace", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "PhD student" } ], "n_formula": 75, "n_formula_1": 27, "n_ref_uni": 46, "n_ref": 201, "n_ref_all": 351, "n_fig": 22, "n_tab": 15, "L_tab": 6530, "n_element_tab": 887, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 1385, "n_element_tab_1": 196, "formula_len_all": 4560, "formula_len_all_1": 1855, "len_all": 277259, "len_all_1": 69912, "len_abs": 833, "len_title": 133, "len_sents": 93118, "len_sents_1": 29507, "n_sents": 757, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 835, "title": "Integral Performance Approximation for Continuous-Time Reinforcement Learning Control", "abs": "We introduce integral performance approximation (IPA), a new continuous-time reinforcement learning (CT-RL) control method. It leverages an affine nonlinear dynamic model, which partially captures the dynamics of the physical environment, alongside state-action trajectory data to enable optimal control with great data efficiency and robust control performance. Utilizing Kleinman algorithm structures allows IPA to provide theoretical guarantees of learning convergence, solution optimality, and closed-loop stability. Furthermore, we demonstrate the effectiveness of IPA on three CT-RL environments including hypersonic vehicle (HSV) control, which has additional challenges caused by unstable and nonminimum phase dynamics. As a result, we demonstrate that the IPA method leads to new, SOTA control design and performance in CT-RL.", "keywords": [ "Continuous-Time Reinforcement Learning (CT-RL)", "Optimal Control", "Integral Performance Approximation (IPA)", "Adaptive/Approximate Dynamic Programming (ADP)", "Flight Control", "Hypersonic Vehicles (HSVs)" ], "rating_list": [ 5, 8, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "z1yI8uoVU3", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Bumjin Park", "gender": "Male", "institution": "KAIST", "country": "", "position": "MS student" }, { "name": "Jaesik Choi", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" }, { "name": "Leejinsil", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Yeonjea Kim", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Youngju Joung", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 42, "n_ref": 80, "n_ref_all": 95, "n_fig": 8, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 681, "formula_len_all_1": 619, "len_all": 148655, "len_all_1": 57796, "len_abs": 1334, "len_title": 128, "len_sents": 31554, "len_sents_1": 26347, "n_sents": 252, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1223, "title": "Measuring Effects of Steered Representation in Large Language Models", "abs": "Large Language Models (LLMs) show advanced performance and adaptability across various tasks. As the model size becomes more extensive, precise control by editing the forward process of LLMs is a challenging problem. Recent research has focused on steering hidden representations during forward propagation to guide model outputs in desired directions, yielding precise control over specific responses. Although steering shows a broader impact on diverse tasks, the influence of steered representations remains unclear. For instance, steering towards a refusal direction might lead the model to refuse even benign requests in subsequent generations. This work tackles the problem of evaluating activation steering. We introduce a counterfactual-based steering evaluation framework that compares the output of base and steered generations. Within the framework, we propose a steering effect matrix that eases the selection of generations base and steered output types. We experimentally evaluate the effects of steered representation for consequence generation with Llama3-8B, Llama2-7B, and Exaone-8B across diverse datasets. We conclude that steered representation changes the original output severely in longer contexts.", "keywords": [ "in-context learning", "activation steering", "large language models" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "z1td6fBKpG", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Tian Yu Liu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Stefano Soatto", "gender": "unknown", "institution": "Amazon Web Services", "country": "US", "position": "Vice President" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 39, "n_ref": 59, "n_ref_all": 77, "n_fig": 5, "n_tab": 4, "L_tab": 1570, "n_element_tab": 195, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 4188, "n_element_tab_1": 222, "formula_len_all": 1374, "formula_len_all_1": 1374, "len_all": 112679, "len_all_1": 60023, "len_abs": 1258, "len_title": 77, "len_sents": 28228, "len_sents_1": 24526, "n_sents": 192, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 29, "L_abs": 1271, "title": "Conjuring Semantic Similarity", "abs": "The semantic similarity between sample expressions measures the distance between their latent 'meaning'. Such meanings are themselves typically represented by textual expressions, often insufficient to differentiate concepts at fine granularity. We propose a novel approach whereby the semantic similarity among textual expressions is based {\\em not} on other expressions they can be rephrased as, but rather based on the imagery they evoke. While this is not possible with humans, generative models allow us to easily visualize and compare generated images, or their distribution, evoked by a textual prompt. Therefore, we characterize the semantic similarity between two textual expressions simply as the distance between image distributions they induce, or 'conjure.' We show that by choosing the Jensen-Shannon divergence between the reverse-time diffusion stochastic differential equations (SDEs) induced by each textual expression, this can be directly computed via Monte-Carlo sampling. Our method contributes a novel perspective on semantic similarity that not only aligns with human-annotated scores, but also opens up new avenues for the evaluation of text-conditioned generative models while offering better interpretability of their learnt representations.", "keywords": [ "Semantic Similarity", "Interpretability", "Diffusion Models" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "z1pydjd4XQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Baorong Shi", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Hao Henry Wang", "gender": "Male", "institution": "Alibaba Cloud Intelligence Group", "country": "CN", "position": "Full Professor" }, { "name": "Li Shen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 8, "n_ref": 27, "n_ref_all": 35, "n_fig": 3, "n_tab": 4, "L_tab": 1162, "n_element_tab": 158, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1168, "n_element_tab_1": 158, "formula_len_all": 261, "formula_len_all_1": 261, "len_all": 85872, "len_all_1": 73894, "len_abs": 1405, "len_title": 154, "len_sents": 21965, "len_sents_1": 21957, "n_sents": 160, "n_sents_1": 159, "venue": "ICLR2025", "decision": "withdraw", "L_title": 105, "L_abs": 1411, "title": "YESNO-PRO: A HIGH-PERFORMANCE POINTWISE RERANKING ALGORITHM BRIDGING ENCODERDECODER AND DECODER-ONLY LLMS", "abs": "Recent research has shown significant progress in the field of zero-shot text reranking for large language models (LLMs). Traditional pointwise approaches prompt the LLM to output relevance labels such as \"yes/no\" or fine-grained labels, but they have several drawbacks. Firstly, these prompts struggle to capture complex correlations between queries and passages and lack robustness for outputs not covered by predefined labels. Secondly, ranking scores rely solely on the likelihood of relevance labels, leading to potential noise and bias. Lastly, existing pointwise approaches are not supported by decoder-only LLMs, as ranking requires LLMs to output prediction probabilities. In response to these challenges, a novel pointwise approach called yesno-pro has been designed, which redefines both prompt design and score computation mechanisms to better align with the intrinsic nature of text reranking. Additionally, a comprehensive reranking framework based on LLM services has been proposed to support concurrent ranking calls and quickly adapt to any open-source decoder-only large models. Experimental results have demonstrated that this method outperforms existing pointwise and some pairwise/listwise methods on TREC19/20 and BEIR datasets, achieving the state-of-the-art performance. Due to its concurrency features, this work is applicable to practical applications with high real-time requirements.", "keywords": [ "zero-shot text reranking", "Large Language Models" ], "rating_list": [ 3, 3, 1, 3 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "z1ohBxWeL2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Aurick Qiao", "gender": "Male", "institution": "Snowflake", "country": "US", "position": "Researcher" }, { "name": "Samyam Rajbhandari", "gender": "Male", "institution": "AI Research, Snowflake", "country": "US", "position": "Principal Researcher" }, { "name": "Yuxiong He", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Zhewei Yao", "gender": "Male", "institution": "Snowflake", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 40, "n_ref": 65, "n_ref_all": 99, "n_fig": 6, "n_tab": 9, "L_tab": 4164, "n_element_tab": 560, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2724, "n_element_tab_1": 349, "formula_len_all": 238, "formula_len_all_1": 222, "len_all": 197713, "len_all_1": 68628, "len_abs": 1623, "len_title": 136, "len_sents": 46416, "len_sents_1": 28184, "n_sents": 395, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1629, "title": "SwiftKV: Fast Prefill-Optimized Inference with Knowledge-Preserving Model Transformation", "abs": "LLM inference for popular enterprise use cases, such as summarization, RAG, and code-generation, typically observes orders of magnitude longer prompt lengths than generation lengths. This characteristic leads to high cost of prefill and increased response latency. \nIn this paper, we present SwiftKV, a novel model transformation and distillation procedure specifically designed to reduce the time and cost of processing prompt tokens while preserving high quality of generated tokens. SwiftKV combines three key mechanisms: i) SingleInputKV, which prefills later layers' KV cache using a much earlier layer's output, allowing prompt tokens to skip much of the model computation, ii) AcrossKV, which merges the KV caches of neighboring layers to reduce the memory footprint and support larger batch size for higher throughput, and iii) a knowledge-preserving distillation procedure that can adapt existing LLMs for SwiftKV with minimal accuracy impact and low compute and data requirement. For Llama-3.1-8B and 70B, SwiftKV reduces the compute requirement of prefill by 50% and the memory requirement of the KV cache by 62.5% while incurring minimum quality degradation across a wide range of tasks. In the end-to-end inference serving using an optimized vLLM implementation, SwiftKV realizes up to 2x higher aggregate throughput and 60% lower time per output token. It can achieve a staggering 560 TFlops/GPU of normalized inference throughput, which translates to 16K tokens/s for Llama-3.1-70B in 16-bit precision on 4x H100 GPUs. Our training, inference, and model implementations are open-sourced at https://anonymized.link.", "keywords": [ "LLM", "Inference", "System", "Compression", "Distillation" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "z1nSpA2dAW", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Guanghao Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jinyang Jiang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Mingqian Feng", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Tao Ren", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yijie Peng", "gender": "Male", "institution": "Peking University", "country": "", "position": "Associate Professor" }, { "name": "Zeliang Zhang", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Zishi Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 47, "n_formula_1": 25, "n_ref_uni": 46, "n_ref": 64, "n_ref_all": 87, "n_fig": 8, "n_tab": 4, "L_tab": 1173, "n_element_tab": 157, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 844, "n_element_tab_1": 121, "formula_len_all": 3980, "formula_len_all_1": 1875, "len_all": 146389, "len_all_1": 62853, "len_abs": 1365, "len_title": 94, "len_sents": 40928, "len_sents_1": 27249, "n_sents": 379, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 45, "L_abs": 1373, "title": "FLOPS: Forward Learning with OPtimal Sampling", "abs": "Given the limitations of backpropagation, perturbation-based gradient computation methods have recently gained focus for learning with only forward passes, also referred to as queries. Conventional forward learning consumes enormous queries on each data point for accurate gradient estimation through Monte Carlo sampling, which hinders the scalability of those algorithms. However, not all data points deserve equal queries for gradient estimation. In this paper, we study the problem of improving the forward learning efficiency from a novel perspective: how to reduce the gradient estimation variance with minimum cost? For this, we allocate the optimal number of queries within a set budget during training to balance estimation accuracy and computational efficiency. Specifically, with a simplified proxy objective and a reparameterization technique, we derive a novel plug-and-play query allocator with minimal parameters. Theoretical results are carried out to verify its optimality. We conduct extensive experiments for fine-tuning Vision Transformers on various datasets and further deploy the allocator to two black-box applications: prompt tuning and multimodal alignment for foundation models. All findings demonstrate that our proposed allocator significantly enhances the scalability of forward-learning algorithms, paving the way for real-world applications.", "keywords": [ "stochastic optimization", "gradient estimation" ], "rating_list": [ 8, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 3, 2, 1 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 1, 3, 3 ] }, { "paper_id": "z1mLNhWFyY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Alex Cloud", "gender": "unknown", "institution": "ML Alignment and Theory Scholars", "country": "US", "position": "Researcher" }, { "name": "Alexander Matt Turner", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Ev\u017een Wybitul", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Jacob Goldman-Wetzler", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Undergrad student" }, { "name": "Joseph Miller", "gender": "Male", "institution": "ML Alignment & Theory Scholars", "country": "US", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 2, "n_ref_uni": 115, "n_ref": 178, "n_ref_all": 222, "n_fig": 20, "n_tab": 5, "L_tab": 2325, "n_element_tab": 62, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 666, "n_element_tab_1": 17, "formula_len_all": 1030, "formula_len_all_1": 141, "len_all": 400245, "len_all_1": 66005, "len_abs": 1269, "len_title": 127, "len_sents": 109623, "len_sents_1": 31890, "n_sents": 1016, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1274, "title": "Gradient Routing: Masking Gradients to Localize Computation in Neural Networks", "abs": "Neural networks are trained primarily based on their inputs and outputs, without regard for their internal mechanisms. These neglected mechanisms determine properties that are critical for safety, like (i) transparency; (ii) the absence of sensitive information or harmful capabilities; and (iii) reliable generalization of goals beyond the training distribution. To address this shortcoming, we introduce gradient routing, a training method that isolates capabilities to specific subregions of a neural network. Gradient routing applies data-dependent, weighted masks to gradients during backpropagation. These masks are supplied by the user in order to configure which parameters are updated by which data points. We show that gradient routing can be used to (1) learn representations which are partitioned in an interpretable way; (2) enable robust unlearning via ablation of a pre-specified network subregion; and (3) achieve scalable oversight of a reinforcement learner by localizing modules responsible for different behaviors. Throughout, we find that gradient routing localizes capabilities even when applied to a limited, ad-hoc subset of the data. We conclude that the approach holds promise for challenging, real-world applications where quality data are scarce.", "keywords": [ "representation learning", "modularity", "unlearning", "reinforcement learning", "scalable oversight" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 4, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "z1Jq1PLQWs", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Aadirupa Saha", "gender": "unknown", "institution": "University of Illinois at Chicago", "country": "US", "position": "Assistant Professor" }, { "name": "Barry-John Theobald", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 19, "n_ref_uni": 38, "n_ref": 70, "n_ref_all": 73, "n_fig": 0, "n_tab": 3, "L_tab": 326, "n_element_tab": 56, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 1256, "n_element_tab_1": 85, "formula_len_all": 1849, "formula_len_all_1": 1644, "len_all": 112859, "len_all_1": 70145, "len_abs": 2262, "len_title": 177, "len_sents": 33922, "len_sents_1": 29320, "n_sents": 264, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 129, "L_abs": 1998, "title": "Dueling in the Dark: An Efficient and Optimal $O(\\sqrt{T})$ Mirror Descent Approach for Competing against Adversarial Preferences", "abs": "Recent developments in Large Language Models (LLMs) have sparked significant attention in Reinforcement Learning from Human Feedback (RLHF), which uses reinforcement learning techniques to optimize a model's performance through human-provided feedback. A simple, widely used, and cost-effective method for gathering human feedback is through relative queries based on human preferences, often modeled using sigmoid utility models. Despite the popularity of sigmoid model-based RLHF algorithms, their theoretical foundations remain underdeveloped as existing algorithms often lack performance guarantees or are limited to small-scale problems due to computationally intractable steps. We address the challenge of developing no-regret learning algorithms for training optimal policy RLHF, and develop the first efficient gradient descent-based algorithm with near-optimal regret guarantees. More technically, we consider the adversarial online convex optimization problem with preference feedback and propose a mirror descent method to obtain a regret of $O(\\sqrt{T})$ over $T$ rounds. The main challenge we are required to solve lies in finding a suitable `gradient-approximation' of the underlying utility functions solely from a binary preference feedback. Following this we extend our results to policy optimization in the RLHF framework with trajectory preferences and design no-regret RL policies using a variant of mirror descent. We also extend our methods beyond pairwise preferences --- to multi-way (batched pairwise) feedback and ranking feedback --- and analyze the trade-off between learning rate with increasing subset size. Our contribution lays the groundwork for a practical gradient descent-based algorithm in RLHF with human preferences. Supported by robust theoretical guarantees, our approach holds promise in the current landscape of developing efficient algorithms for LLMs and addressing human-AI alignment challenges. Empirical evaluations validate our theoretical findings.", "keywords": [ "Large Language Models (LLMs)", "Reinforcement Learning from Human Feedback (RLHF)", "gradient descent-based algorithm", "theoretical foundations", "active no-regret learning", "preference feedback", "trajectory preferences", "multi-way feedback", "human-AI alignment", "practical impact." ], "rating_list": [ 5, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 4, 2, 4, 3, 3 ] }, { "paper_id": "z0hUsPhwUN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Anqi Li", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Feng Li", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Huihui Bai", "gender": "Female", "institution": "Beijing jiaotong univercity", "country": "CN", "position": "Full Professor" }, { "name": "Runmin Cong", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "Full Professor" }, { "name": "Yao Zhao", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Yuxi Liu", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 56, "n_ref": 102, "n_ref_all": 126, "n_fig": 10, "n_tab": 3, "L_tab": 417, "n_element_tab": 41, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 345, "n_element_tab_1": 5, "formula_len_all": 939, "formula_len_all_1": 551, "len_all": 138990, "len_all_1": 52645, "len_abs": 1791, "len_title": 137, "len_sents": 39275, "len_sents_1": 26201, "n_sents": 263, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1664, "title": "Once-for-All: Controllable Generative Image Compression with Dynamic Granularity Adaption", "abs": "Although recent generative image compression methods have demonstrated impressive potential in optimizing the rate-distortion-perception trade-off, they still face the critical challenge of flexible rate adaption to diverse compression necessities and scenarios. To overcome this challenge, this paper proposes a $\\textbf{Control}$lable $\\textbf{G}$enerative $\\textbf{I}$mage $\\textbf{C}$ompression framework, $\\textbf{Control-GIC}$, the first capable of fine-grained bitrate adaption across a broad spectrum while ensuring high-fidelity and generality compression. We base $\\textbf{Control-GIC}$ on a VQGAN framework representing an image as a sequence of variable-length codes ($\\textit{i.e.}$ VQ-indices), which can be losslessly compressed and exhibits a direct positive correlation with the bitrates. Drawing inspiration from the classical coding principle, we correlate the information density of local image patches with their granular representations. Hence, we can flexibly determine a proper allocation of granularity for the patches to achieve dynamic adjustment for VQ-indices, resulting in desirable compression rates. We further develop a probabilistic conditional decoder capable of retrieving historic encoded multi-granularity representations according to transmitted codes, and then reconstruct hierarchical granular features in the formalization of conditional probability, enabling more informative aggregation to improve reconstruction realism. Our experiments show that $\\textbf{Control-GIC}$ allows highly flexible and controllable bitrate adaption where the results demonstrate its superior performance over recent state-of-the-art methods.", "keywords": [ "image compression", "vqgan", "generative compression model", "multi-grained representation" ], "rating_list": [ 6, 5, 3, 8, 6 ], "soundness_list": [ 3, 3, 1, 3, 3 ], "presentation_list": [ 3, 2, 1, 3, 3 ], "contribution_list": [ 3, 3, 1, 3, 2 ], "confidence_list": [ 5, 4, 5, 5, 5 ] }, { "paper_id": "z0B7A6Dh1H", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Aldo Pacchiano", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "Assistant Professor" }, { "name": "Spyros Dragazis", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "PhD student" } ], "n_formula": 62, "n_formula_1": 35, "n_ref_uni": 18, "n_ref": 46, "n_ref_all": 48, "n_fig": 0, "n_tab": 2, "L_tab": 303, "n_element_tab": 6, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 140, "n_element_tab_1": 15, "formula_len_all": 5654, "formula_len_all_1": 2551, "len_all": 121012, "len_all_1": 60641, "len_abs": 1283, "len_title": 112, "len_sents": 32625, "len_sents_1": 24483, "n_sents": 327, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1348, "title": "High Probability Contextual Bandits for Optimal Dosage Selection", "abs": "Multi-Armed Bandit ($\\textit{MAB}$) formulations are commonly used to model the problem of $\\textit{Optimal Dose-Finding}$.\nHowever, in many practical applications, it is necessary to receive data about the patient\u2019s current state and then administer a drug dosage adapted to that state. \nTo overcome this issue, we adopt a linear contextual bandit formulation with stage-wise constraints.\nAt each round, the learner selects a dosage and receives both a reward signal and a cost signal.\nThe learner\u2019s goal is to maximize the drug's efficacy\u2014captured as the expected cumulative reward\u2014while ensuring that the toxicity, reflected by the cost signal, remains below a known threshold.\nSatisfying the cost signal constraint only in expectation can be dangerous, as it may lead to over-dosage complications in certain cases.\nTo address this issue, we introduce a novel model that controls the realization of the cost signal with high probability, in contrast to previous works where control was only applied to the expected cost signal.\nOur algorithm follows the $\\textit{UCB}$ approach, for which we establish a regret bound over \n$T$ rounds and run numerical experiments.\nWe further generalize our results to $\\textit{non-linear}$ functions and provide a regret bound in terms of the $\\textit{eluder dimension}$, a measure of function class complexity.", "keywords": [ "Linear Bandits", "Dosage Selection", "Contextual Bandits" ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 4, 3, 4, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "yzloNYH3QN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Shijie Chen", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Yu Su", "gender": "Male", "institution": "Ohio State University", "country": "US", "position": "Assistant Professor" }, { "name": "Bernal Jimenez", "gender": "Male", "institution": "The Ohio State University", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 43, "n_ref": 80, "n_ref_all": 110, "n_fig": 4, "n_tab": 16, "L_tab": 7301, "n_element_tab": 732, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 641, "n_element_tab_1": 41, "formula_len_all": 169, "formula_len_all_1": 169, "len_all": 198045, "len_all_1": 60156, "len_abs": 2204, "len_title": 120, "len_sents": 59551, "len_sents_1": 28415, "n_sents": 439, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 2220, "title": "Attention in Large Language Models Yields Efficient Zero-Shot Re-Rankers", "abs": "Information retrieval (IR) systems have played a vital role in modern digital life and have cemented their continued usefulness in this new era of generative AI via retrieval-augmented generation. With strong language processing capabilities and remarkable versatility, large language models (LLMs) have become popular choices for zero-shot re-ranking in IR systems. So far, LLM-based re-ranking methods rely on strong generative capabilities, which restricts their use to either specialized or powerful proprietary models. Given these restrictions, we ask: is autoregressive generation necessary and optimal for LLMs to perform re-ranking? We hypothesize that there are abundant signals relevant to re-ranking within LLMs that might not be used to their full potential via generation. To more directly leverage such signals, we propose in-context re-ranking (ICR), a novel method that leverages the change in attention pattern caused by the search query for accurate and efficient re-ranking. We assume that more relevant documents should receive more attention weights when an LLM is processing the query tokens, and leverage such signals for re-ranking. To mitigate the intrinsic biases in LLMs, we propose a calibration method using a content-free query. Due to the absence of generation, ICR only requires two ($O(1)$) forward passes to re-rank $N$ documents, making it substantially more efficient than generative re-ranking methods that require at least $O(N)$ forward passes. Our novel design also enables ICR to be applied to any LLM without specialized training while guaranteeing a well-formed ranking. Extensive experiments with two popular open-weight LLMs on standard single-hop and multi-hop information retrieval benchmarks show that ICR outperforms RankGPT while cutting the latency by more than 60% in practice. Through detailed analyses, we show that ICR's performance is specially strong on tasks that require more complex re-ranking signals, such as handling contextualization and contradiction between the query and passages, as well as information integration across multiple passages. Our findings call for further exploration on novel ways of utilizing open-weight LLMs beyond text generation.", "keywords": [ "Large Language Model", "Information Retrieval" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "yyIHdaSDUU", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Aristeidis Panos", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "Daniel Olmeda Reino", "gender": "Not Specified", "institution": "Toyota Motor Europe", "country": "BE", "position": "Researcher" }, { "name": "Rahaf Aljundi", "gender": "Female", "institution": "Toyota Motor Europe", "country": "BE", "position": "Researcher" }, { "name": "Richard E. Turner", "gender": "Male", "institution": "Alan Turing Institute", "country": "GB", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 17, "n_ref_uni": 31, "n_ref": 60, "n_ref_all": 83, "n_fig": 2, "n_tab": 18, "L_tab": 1741, "n_element_tab": 252, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 151, "n_element_tab_1": 24, "formula_len_all": 1075, "formula_len_all_1": 965, "len_all": 149148, "len_all_1": 74467, "len_abs": 1300, "len_title": 136, "len_sents": 45479, "len_sents_1": 33557, "n_sents": 460, "n_sents_1": 341, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1307, "title": "Adaptive Vision Encoders: Balancing Efficiency and Robustness in Vision-Language Models", "abs": "Vision-language models (VLMs) demonstrate impressive capabilities in visual question answering and image captioning, acting as a crucial link between visual and language modalities. However, existing open-source VLMs rely heavily on pretrained vision encoders, such as CLIP. Despite CLIP\u2019s robustness across diverse domains, it still exhibits significant image understanding errors. These errors propagate to the VLM responses, resulting in sub-optimal performance. In our work, we propose an efficient and robust method for updating vision encoders within VLMs. Our approach selectively and locally updates the model parameters, leading to substantial performance improvements on data where previous mistakes occurred, while maintaining overall robustness. We demonstrate the effectiveness of our method during offline and continual few-shot updates, simulating a model editing regime for VLMs. While our method also scales efficiently and effectively to adapting the language model (LLM) component of the VLM, we show that separately updating the vision encoder can be a very efficient alternative. This approach improves VLM performance with less than 10x the compute resources required for updating the LLM. Our method is also supported by theoretical justifications on the parameter selection strategy.", "keywords": [ "large vision-language models", "multimodal learning", "continual learning" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 3, 1, 2, 1 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "yx8bU8T5ZN", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bowen Yu", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Hongyu Lin", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Keming Lu", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Le Sun", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Le Yu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Qiaoyu Tang", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Xianpei Han", "gender": "Male", "institution": "Institute of Software, CAS", "country": "CN", "position": "Professor" }, { "name": "Yaojie Lu", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 18, "n_ref_uni": 38, "n_ref": 64, "n_ref_all": 83, "n_fig": 6, "n_tab": 1, "L_tab": 131, "n_element_tab": 26, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 131, "n_element_tab_1": 26, "formula_len_all": 1892, "formula_len_all_1": 1824, "len_all": 119136, "len_all_1": 54533, "len_abs": 1316, "len_title": 124, "len_sents": 28048, "len_sents_1": 23638, "n_sents": 199, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1320, "title": "A Unified View of Delta Parameter Editing in Post-Trained Large-Scale Models", "abs": "Post-training has emerged as a crucial paradigm for adapting large-scale pre-trained models to various tasks, whose effects are fully reflected by delta parameters (i.e., the disparity between post-trained and pre-trained parameters). While numerous studies have explored delta parameter properties via operations like pruning, quantization, low-rank approximation, and extrapolation, a unified framework for systematically examining these characteristics has been lacking. In this paper, we propose a novel perspective based on Riemann sum approximation of the loss function to elucidate delta parameter editing operations. Our analysis categorizes existing methods into three classes based on their post-editing performance: competitive, decreased, and improved, explaining how they are expressed by the Riemann sum approximation term and how they alter the model performance. Extensive experiments on both visual and language models, including ViT, LLaMA 3, and Mistral, corroborate our theoretical findings. Furthermore, we introduce extensions to existing techniques like DARE and BitDelta, highlighting their limitations in leveraging the properties of delta parameters and reorganizing them into general expressions to enhance the applicability and effectiveness of delta parameter editing in post-trained models.", "keywords": [ "Large Language Models", "Delta Parameters Editing" ], "rating_list": [ 3, 3, 1 ], "soundness_list": [ 2, 1, 1 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "ywgwArtbDq", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Andreas Plesner", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Roger Wattenhofer", "gender": "Not Specified", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Turlan Kuzhagaliyev", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Yahya Jabary", "gender": "unknown", "institution": "Technische Universit\u00e4t Wien", "country": "AT", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 26, "n_ref": 37, "n_ref_all": 59, "n_fig": 2, "n_tab": 1, "L_tab": 7, "n_element_tab": 1, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 7, "n_element_tab_1": 1, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 80436, "len_all_1": 70172, "len_abs": 1201, "len_title": 118, "len_sents": 25011, "len_sents_1": 19889, "n_sents": 170, "n_sents_1": 135, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 913, "title": "Seeing Through the Mask: Rethinking Adversarial Examples for CAPTCHAs", "abs": "Modern CAPTCHAs often rely on vision tasks that are supposedly hard for computers but easy for humans. Although image recognition models pose a significant threat to such CAPTCHAs, they can be fooled by hiding ``random'' noise in images. However, these methods are model-specific and thus can not aid CAPTCHAs in fooling all models. \n We show in this work that by allowing for more significant changes to the images while preserving the semantic information and keeping it solvable by humans, we can fool many state-of-the-art models. Specifically, we demonstrate that by adding masks of various intensities the Top 1 Accuracy (Acc@1) drops by more than 50%-points for all models, and supposedly robust models such as vision transformers see an Acc@1 drop of 80%-points. \n These masks can therefore effectively fool modern image classifiers, thus showing that machines have not caught up with humans -- yet.", "keywords": [ "CAPTCHAs", "Adversarial examples", "Vision models", "Robust models" ], "rating_list": [ 5, 1, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "ywKlmMor0f", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Jiaxian Guo", "gender": "Male", "institution": "Google Research", "country": "AU", "position": "Research Scientist" }, { "name": "Kexin Song", "gender": "Female", "institution": "The University of Tokyo", "country": "JP", "position": "MS student" }, { "name": "LIANG DING", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Mingming Gong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Yusuke Iwasawa", "gender": "Male", "institution": "The University of Tokyo, The University of Tokyo", "country": "JP", "position": "Associate Professor" }, { "name": "Yutaka Matsuo", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Ru Wang", "gender": "Male", "institution": "The University of Tokyo, The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Shane Gu", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Staff Research Scientist" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 51, "n_ref": 75, "n_ref_all": 102, "n_fig": 12, "n_tab": 7, "L_tab": 3781, "n_element_tab": 594, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 5377, "n_element_tab_1": 592, "formula_len_all": 37, "formula_len_all_1": 0, "len_all": 184948, "len_all_1": 71764, "len_abs": 2027, "len_title": 122, "len_sents": 42126, "len_sents_1": 26409, "n_sents": 328, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 2139, "title": "MMA: Benchmarking Multi-Modal Large Language Model in Ambiguity Contexts", "abs": "Multi-Modal Large Language Models (MLLMs) recently demonstrated strong capabilities in both instruction comprehension and responding, positioning them as promising tools for human-computer interaction. However, the inherent ambiguity of language poses a challenge, potentially leading models astray in task implementation due to differing interpretations of the same text within varying contexts. In multi-modal settings, visual information serves as a natural aid in disambiguating such scenarios. In this paper, we introduce the first benchmark specifically designed to evaluate the performance of \\textbf{M}LL\\textbf{M}s in \\textbf{A}mbiguous contexts (MMA). This benchmark employs a multiple-choice visual question-answering format and includes 261 textual contexts and \nquestions with ambiguous meaning. Each question is linked to a pair of images that suggest divergent scenarios, thus leading to different answers given the same question. These questions are stratified into three categories of ambiguity: lexical, syntactic, and semantic, to facilitate a detailed examination of MLLM performance across varying levels of ambiguity. By evaluating 24 proprietary and open-sourced MLLMs, we find that: (1) MLLMs often overlook scenario-specific information provided by images to clarify the ambiguity of texts. When presented with two different contextual images and asked the same question, \n MLLMs achieved an accuracy rate of only 53.22\\% in answering both correctly, \n compared to human performance at 88.97\\%.(2) Among the three types of ambiguity, models perform best under lexical ambiguity and worst under syntactic ambiguity. (3) Open-sourced models generally perform significantly lower than proprietary MLLMs, with an average performance gap of 12.59\\%, Claude 3.5 Sonnet, emerges as the top model, achieving 74.32\\% accuracy. These findings firstly underscore the current limitations of MLLMs in integrating visual information to clarify textual ambiguities and highlight critical areas for future improvements. The codes and benchmark data are \\href{https://github.com/AnonymousSubmitter-gpu/MMA_Anony}{available}.", "keywords": [ "Multi-Modal Large Language Model", "Ambiguity", "Benchmark" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "ywHOnGOLb1", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Jianping Wang", "gender": "Female", "institution": "City University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Meng XU", "gender": "Male", "institution": "City University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Xinhong Chen", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Zhao Guanyi", "gender": "Male", "institution": "City University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Zihao WEN", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 40, "n_formula_1": 17, "n_ref_uni": 24, "n_ref": 60, "n_ref_all": 74, "n_fig": 7, "n_tab": 2, "L_tab": 324, "n_element_tab": 10, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1230, "n_element_tab_1": 24, "formula_len_all": 4010, "formula_len_all_1": 1490, "len_all": 151758, "len_all_1": 68275, "len_abs": 1500, "len_title": 123, "len_sents": 52467, "len_sents_1": 31812, "n_sents": 430, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1505, "title": "A Competitive-Cooperative Actor-critic Framework for Reinforcement Learning", "abs": "In the field of Deep reinforcement learning (DRL), enhancing exploration capabilities and improving the accuracy of Q-value estimation remain two major challenges.\nRecently, double-actor DRL methods have emerged as a promising class of DRL approaches, achieving substantial advancements in both exploration and Q-value estimation. However, existing double-actor DRL methods feature actors that operate independently in exploring the environment, lacking mutual learning and collaboration, which leads to suboptimal policies. To address this challenge, this work proposes a generic solution that can be seamlessly integrated into existing double-actor DRL methods by promoting mutual learning among the actors to develop improved policies. Specifically, we calculate the difference in actions output by the actors and minimize this difference as a loss during training to facilitate mutual imitation among the actors. Simultaneously, we also minimize the differences in Q-values output by the various critics as part of the loss, thereby avoiding significant discrepancies in value estimation for the imitated actions. We present two specific implementations of our method and extend these implementations beyond double-actor DRL methods to other DRL approaches to encourage broader adoption. Experimental results demonstrate that our method effectively enhances four state-of-the-art (SOTA) double-actor DRL methods and five other types of SOTA DRL methods across four MuJoCo tasks, as measured by return.", "keywords": [ "Deep reinforcement learning; Double-actor framework; Competition and Cooperation" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "ywFOSIT9ik", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Heng Huang", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Shaocong Ma", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Postdoc" } ], "n_formula": 57, "n_formula_1": 15, "n_ref_uni": 41, "n_ref": 78, "n_ref_all": 98, "n_fig": 7, "n_tab": 5, "L_tab": 1835, "n_element_tab": 60, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 232, "n_element_tab_1": 14, "formula_len_all": 7132, "formula_len_all_1": 1266, "len_all": 195966, "len_all_1": 66580, "len_abs": 1163, "len_title": 163, "len_sents": 56134, "len_sents_1": 29896, "n_sents": 486, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 117, "L_abs": 1114, "title": "Revisiting Zeroth-Order Optimization: Minimum-Variance Two-Point Estimators and Directionally Aligned Perturbations", "abs": "In this paper, we explore the two-point zeroth-order gradient estimator and identify the optimal distribution of random perturbations that minimizes the estimator's variance. We formulate it as a constrained functional optimization problem over the space of perturbation distributions. Our findings reveal that optimal perturbations either maintain a fixed length or align directionally with the true gradient. While existing research has largely focused on fixed-length perturbations, the potential advantages of directional alignment have been overlooked. To address this gap, we delve into the theoretical and empirical properties of the directionally aligned perturbation (DAP) scheme, which adaptively offers higher accuracy along critical directions. Additionally, we provide a convergence analysis for stochastic gradient descent using $\\delta$-unbiased random perturbations, extending optimal complexity bounds to a wider range of perturbations. Through empirical evaluations on both synthetic problems and practical tasks, we demonstrate that DAPs outperform traditional methods under specific conditions.", "keywords": [ "zeroth-order optimization", "SGD", "convergence analysis" ], "rating_list": [ 8, 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 2, 5 ] }, { "paper_id": "yvxpHbydFx", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Peng Wang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Postdoc" }, { "name": "Qing Qu", "gender": "Male", "institution": "University of Michigan", "country": "US", "position": "Assistant Professor" }, { "name": "Siyi Chen", "gender": "Female", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Xiang Li", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Xiao Li", "gender": "unknown", "institution": "University of Michigan", "country": "", "position": "PhD student" }, { "name": "Zekai Zhang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Zhihui Zhu", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Assistant Professor" } ], "n_formula": 44, "n_formula_1": 22, "n_ref_uni": 58, "n_ref": 165, "n_ref_all": 223, "n_fig": 18, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9100, "formula_len_all_1": 1414, "len_all": 181507, "len_all_1": 64586, "len_abs": 1071, "len_title": 131, "len_sents": 53332, "len_sents_1": 31392, "n_sents": 392, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1073, "title": "Understanding Diffusion-based Representation Learning via Low-Dimensional Modeling", "abs": "This work addresses the critical question of why and when diffusion models, despite their generative design, are capable of learning high-quality representations in a self-supervised manner. We hypothesize that diffusion models excel in representation learning due to their ability to learn the low-dimensional distributions of image datasets via optimizing a noise-controlled denoising objective. Our empirical results support this hypothesis, indicating that variations in the representation learning performance of diffusion models across noise levels are closely linked to the quality of the corresponding posterior estimation. Grounded on this observation, we offer theoretical insights into the unimodal representation dynamics of diffusion models as noise scales vary, demonstrating how they effectively learn meaningful representations through the denoising process. We also highlight the impact of the inherent parameter-sharing mechanism in diffusion models, which accounts for their advantages over traditional denoising auto-encoders in representation learning.", "keywords": [ "diffusion representation learning", "representation learning", "diffusion model", "denoising auto-encoder" ], "rating_list": [ 3, 8, 1, 5 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "yuymgwkjj1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Qiaoqiao Ding", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Researcher" }, { "name": "Xiaoqun Zhang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "", "position": "Full Professor" }, { "name": "Yuzhong Zhao", "gender": "unknown", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 31, "n_ref": 44, "n_ref_all": 65, "n_fig": 5, "n_tab": 11, "L_tab": 3765, "n_element_tab": 602, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2190, "n_element_tab_1": 377, "formula_len_all": 759, "formula_len_all_1": 591, "len_all": 146030, "len_all_1": 65402, "len_abs": 1655, "len_title": 154, "len_sents": 40615, "len_sents_1": 27611, "n_sents": 299, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1427, "title": "Correcting the Bias of Normalizing Flows by Synthetic Outliers for Improving Out-of-Distribution Detection", "abs": "Out-of-distribution (OOD) detection is critical for ensuring the reliability and robustness of deep learning models in real-world applications. While normalizing flows have demonstrated impressive performance for various task of image OOD detection, recent findings suggest that they still encounter limitations and severe biases when applied to datasets with different statistics. Specifically, it has been observed that normalizing flow models tend to assign higher likelihoods to OOD samples with low complexity, which undermines the effectiveness of likelihood based OOD detection methods. In this paper, we explore the bias related to data complexity linked to normalizing flow models in OOD detection. We propose a novel method for bias correction by incorporating synthetic outliers during training, guiding the model to assign lower likelihoods to OOD samples. Additionally, we introduce a specialized training objective that leverages the softplus function for OOD data, ensuring a smooth and effective training process. Extensive experiments on benchmark and high-dimensional real-world datasets, including both images and texts, confirm that our proposed approach significantly enhances OOD detection accuracy, achieving performance comparable to models trained with a limited number of real outliers. Moreover, our method increases the Lipschitz constant, supporting the hypothesis presented in related literature.", "keywords": [ "OOD Detection", "Normalizing Flow" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "yuuyPlywuO", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Diyi Yang", "gender": "Female", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Michael J Ryan", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Minzhi Li", "gender": "Female", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Weiyan Shi", "gender": "Female", "institution": "Stanford University", "country": "", "position": "Postdoc" }, { "name": "William Barr Held", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Yanzhe Zhang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 55, "n_ref": 103, "n_ref_all": 118, "n_fig": 4, "n_tab": 3, "L_tab": 663, "n_element_tab": 134, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 827, "n_element_tab_1": 510, "formula_len_all": 88, "formula_len_all_1": 88, "len_all": 203924, "len_all_1": 61689, "len_abs": 379, "len_title": 0, "len_sents": 35254, "len_sents_1": 26861, "n_sents": 262, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 905, "title": "Distilling an End-to-End Voice Assistant Without Instruction Training Data", "abs": "Voice assistants, such as Siri and Google Assistant, typically model audio and text separately, resulting in lost speech information and increased complexity. Recent efforts to address this with end-to-end Speech Large Language Models (LLMs) trained with supervised finetuning (SFT) \n have led to models ``forgetting\" capabilities from text-only LLMs. Our work proposes an alternative paradigm for training Speech LLMs without instruction data, using the response of a text-only LLM to transcripts as self-supervision. Importantly, this process can be performed without annotated responses. We show that our Distilled Voice Assistant (DiVA) generalizes to Spoken Question Answering, Classification, and Translation. Furthermore, we show that DiVA better meets user preferences, achieving a 72\\% win rate compared with state-of-the-art models like Qwen 2 Audio, despite using $>$100x less training compute.", "keywords": [ "Multi-Modal LLMs", "Voice Assistants", "Distillation" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 2, 1, 4, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "yu1vqQqKkx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Aditya Grover", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "Tung Nguyen", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 58, "n_ref": 116, "n_ref_all": 136, "n_fig": 5, "n_tab": 7, "L_tab": 6329, "n_element_tab": 435, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 3297, "n_element_tab_1": 244, "formula_len_all": 209, "formula_len_all_1": 209, "len_all": 184459, "len_all_1": 74628, "len_abs": 1337, "len_title": 113, "len_sents": 45826, "len_sents_1": 34628, "n_sents": 333, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1342, "title": "LICO: Large Language Models for In-Context Molecular Optimization", "abs": "Optimizing black-box functions is a fundamental problem in science and engineering. To solve this problem, many approaches learn a surrogate function that estimates the underlying objective from limited historical evaluations. Large Language Models (LLMs), with their strong pattern-matching capabilities via pretraining on vast amounts of data, stand out as a potential candidate for surrogate modeling. However, directly prompting a pretrained language model to produce predictions is not feasible in many scientific domains due to the scarcity of domain-specific data in the pretraining corpora and the challenges of articulating complex problems in natural language. In this work, we introduce LICO, a general-purpose model that extends arbitrary base LLMs for black-box optimization, with a particular application to the molecular domain. To achieve this, we equip the language model with a separate embedding layer and prediction layer, and train the model to perform in-context predictions on a diverse set of functions defined over the domain. Once trained, LICO can generalize to unseen molecule properties simply via in-context prompting. LICO performs competitively on PMO, a challenging molecular optimization benchmark comprising 23 objective functions, and achieves state-of-the-art performance on its low-budget version PMO-1K.", "keywords": [ "large language models", "molecular optimization", "black-box optimization", "foundation models", "in-context learning" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 4, 4 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "ytvWZEiywp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 1, "author_info_list": [ { "name": "Edward Chang", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Full Professor" } ], "n_formula": 21, "n_formula_1": 5, "n_ref_uni": 21, "n_ref": 39, "n_ref_all": 71, "n_fig": 8, "n_tab": 10, "L_tab": 3080, "n_element_tab": 82, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 214, "n_element_tab_1": 3, "formula_len_all": 1108, "formula_len_all_1": 841, "len_all": 207909, "len_all_1": 45380, "len_abs": 876, "len_title": 94, "len_sents": 90843, "len_sents_1": 22025, "n_sents": 683, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 672, "title": "EVINCE: Optimizing Adversarial LLM Dialogues via Conditional Statistics and Information Theory", "abs": "This paper introduces EVINCE (Entropy and Variation IN Conditional Exchanges), a framework that optimizes multi-LLM dialogues using conditional statistics and information theory. EVINCE introduces dual entropy optimization to balance perspective diversity with prior knowledge, providing quantitative measures for modulating LLM interactions. Through information-theoretic metrics and mutual information optimization, the framework demonstrates consistent improvement over single-LLM performance in applications ranging from disease diagnosis to news debiasing. We present theoretical foundations and empirical validation for this structured approach to LLM collaboration.", "keywords": [ "LLM", "GAI", "AGI" ], "rating_list": [ 5, 6, 3, 3, 3 ], "soundness_list": [ 3, 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "ytn0rbIfOx", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Fan Li", "gender": "Female", "institution": "University of Exeter", "country": "GB", "position": "Postdoc" }, { "name": "Ke Li", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "Associate Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 28, "n_ref": 39, "n_ref_all": 56, "n_fig": 7, "n_tab": 9, "L_tab": 3871, "n_element_tab": 242, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 3096, "n_element_tab_1": 143, "formula_len_all": 190, "formula_len_all_1": 190, "len_all": 135771, "len_all_1": 60904, "len_abs": 1405, "len_title": 169, "len_sents": 38620, "len_sents_1": 27159, "n_sents": 291, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 120, "L_abs": 1410, "title": "Formulating AutoML as a Variable-Length Optimization Problem: A Tree of Thought Approach with LLM-Driven Code Generation", "abs": "Recent advancements in machine learning have created a demand for automated systems that enable efficient development and deployment of machine learning applications. Traditional Automated Machine Learning (AutoML) approaches often rely on fixed pipeline structures, which limit adaptability to diverse task complexities. In this paper, we introduce a novel formulation of AutoML as a variable-length optimization problem, allowing for dynamic adjustment of model architectures based on task requirements. To effectively navigate the expanded search space of variable-length models, we employ the Tree of Thoughts (ToT) method combined with Large Language Models (LLMs). This framework utilizes a sequential decision-making process, allowing models to be incrementally constructed by evaluating prior outcomes. Additionally, LLMs automatically generate the code corresponding to each decision, transforming model configurations into executable pipelines and reducing manual intervention. Our approach enhances efficiency by focusing on promising pathways and improves transparency by explicitly showcasing how each decision contributes to the overall optimization. Experiments conducted on diverse datasets, including OpenML and clinical tasks, demonstrate that our method outperforms traditional AutoML systems, delivering superior model performance and better adaptability across different task complexities.", "keywords": [ "AutoML", "Tree of Thought", "LLM" ], "rating_list": [ 3, 3, 8 ], "soundness_list": [ 1, 2, 4 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 4 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "yt7nxONs3J", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Ahmad Sajedi", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Kai Wang", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Kaipeng Zhang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Samir Khaki", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "MS student" }, { "name": "Tianle Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" }, { "name": "Zekai Li", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "Undergrad student" }, { "name": "Zhi-Qi Cheng", "gender": "Not Specified", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "Ziyao Guo", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Kostantinos Plataniotis", "gender": "Male", "institution": "Toronto University", "country": "CA", "position": "Full Professor" }, { "name": "ZHAO WANGBO", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 40, "n_ref": 93, "n_ref_all": 123, "n_fig": 10, "n_tab": 5, "L_tab": 3314, "n_element_tab": 410, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 1698, "n_element_tab_1": 57, "formula_len_all": 206, "formula_len_all_1": 172, "len_all": 128592, "len_all_1": 53625, "len_abs": 1598, "len_title": 92, "len_sents": 34560, "len_sents_1": 24090, "n_sents": 266, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1276, "title": "Prioritize Alignment in Dataset Distillation", "abs": "Dataset Distillation aims to compress a large dataset into a significantly more compact, synthetic one without compromising the performance of the trained models. \nTo achieve this, existing methods use the agent model to extract information from the target dataset and embed it into the distilled dataset. \nConsequently, the quality of extracted and embedded information determines the quality of the distilled dataset.\nIn this work, we find that existing methods introduce misaligned information in both information extraction and embedding stages.\nTo alleviate this, we propose Prioritize Alignment in Dataset Distillation (\\textbf{PAD}), which aligns information from the following two perspectives.\n1) We prune the target dataset according to the compressing ratio to filter the information that can be extracted by the agent model.\n2) We use only deep layers of the agent model to perform the distillation to avoid excessively introducing low-level information.\nThis simple strategy effectively filters out misaligned information and brings non-trivial improvement for mainstream matching-based distillation algorithms.\nFurthermore, built on trajectory matching, \\textbf{PAD} achieves remarkable improvements on various benchmarks, achieving state-of-the-art performance.", "keywords": [ "dataset distillation" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "yspBoIZJ9Z", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Di Huang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Guodong Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Guohao Li", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Jinjin Zhang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Puyue Hou", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 37, "n_ref": 90, "n_ref_all": 115, "n_fig": 9, "n_tab": 6, "L_tab": 1067, "n_element_tab": 110, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1207, "n_element_tab_1": 179, "formula_len_all": 690, "formula_len_all_1": 690, "len_all": 132515, "len_all_1": 65789, "len_abs": 1572, "len_title": 122, "len_sents": 35244, "len_sents_1": 29145, "n_sents": 277, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1284, "title": "Enhancing Video Understanding with Vision and Language Collaboration", "abs": "Leveraging video pre-trained models has led to significant advancements in video understanding tasks. However, due to the inherent bias towards temporal learning in video pre-training, these models fail to capture comprehensive spatial cues. Additionally, the widely-used supervised adaption methods lack fine-grained semantic guidance as single action labels cannot precisely depict the intra-class diversity. To address these challenges, we incorporate the general capabilities of large Vision Language Models (VLMs) and propose a cross-modal collaborative knowledge transfer method to enhance video understanding. First, we propose an attentive spatial knowledge transfer method that distills spatial knowledge from the VLM's image encoder, enabling the precise capture of spatial information. Next, we design a contrastive textual knowledge transfer approach that achieves detailed video representations through fine-grained text-video alignment. Owing to the cross-modal knowledge transfer, the video representations are capable of attending to informative spatial regions and aligning with fine-grained texts that carry rich semantics. Extensive experiments demonstrate that our method achieves state-of-the-art performance across various datasets, validating its effectiveness.", "keywords": [ "Video understanding", "video pre-trained model", "vision-language model", "collaboration learning" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ysZvK6b60c", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Bin Chen", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Bin Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hao Fang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Intern" }, { "name": "Hongyao Yu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Ke Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Shu-Tao Xia", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Sijin Yu", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "MS student" }, { "name": "Yixiang Qiu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 21, "n_formula_1": 4, "n_ref_uni": 34, "n_ref": 109, "n_ref_all": 149, "n_fig": 15, "n_tab": 16, "L_tab": 5597, "n_element_tab": 856, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1729, "n_element_tab_1": 90, "formula_len_all": 1039, "formula_len_all_1": 120, "len_all": 181845, "len_all_1": 63238, "len_abs": 1487, "len_title": 101, "len_sents": 54210, "len_sents_1": 28298, "n_sents": 463, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 1512, "title": "CALoR: Towards Comprehensive Model Inversion Defense", "abs": "Model Inversion Attacks (MIAs) aim at recovering privacy-sensitive training data from the knowledge encoded in the released machine learning models. Recent advances in the MIA field have significantly enhanced the attack performance under multiple scenarios, posing serious privacy risks of Deep Neural Networks (DNNs). However, the development of defense strategies against MIAs is relatively backward to resist the latest MIAs and existing defenses fail to achieve further trade-off between model utility and model robustness. In this paper, we provide an in-depth analysis from the perspective of intrinsic vulnerabilities of MIAs, comprehensively uncovering the weaknesses inherent in the basic pipeline, which are partially investigated in the previous defenses. Building upon these new insights, we propose a robust defense mechanism, integrating ***C**onfidence **A**daptation* and ***Lo**w-**R**ank compression*(**CALoR**). Our method includes a novel robustness-enhanced classification loss specially-designed for model inversion defenses and reveals the extraordinary effectiveness of compressing the classification header. With CALoR, we can mislead the optimization objective, reduce the leaked information and impede the backpropagation of MIAs, thus mitigating the risk of privacy leakage. Extensive experimental results demonstrate that our method achieves state-of-the-art (SOTA) defense performance against MIAs and exhibits superior generalization to existing defenses across various scenarios.", "keywords": [ "Privacy Leakage", "Model Inversion", "Defense" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "ysQiaWhnCN", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Julian Togelius", "gender": "Male", "institution": "New York University", "country": "US", "position": "Associate Professor" }, { "name": "Sam Earle", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 9, "n_ref": 11, "n_ref_all": 22, "n_fig": 5, "n_tab": 3, "L_tab": 251, "n_element_tab": 22, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 251, "n_element_tab_1": 22, "formula_len_all": 259, "formula_len_all_1": 324, "len_all": 93983, "len_all_1": 54807, "len_abs": 1430, "len_title": 121, "len_sents": 32616, "len_sents_1": 29400, "n_sents": 168, "n_sents_1": 158, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1434, "title": "Autoverse: an Evolvable Game Language for Learning Robust Embodied Agents", "abs": "We introduce Autoverse, an evolvable, domain-specific language for single-player 2D grid-based games, and demonstrate its use as a scalable training ground for Open-Ended Learning (OEL) algorithms. Autoverse uses cellular-automaton-like rewrite rules to describe game mechanics, allowing it to express various game environments (e.g. mazes, dungeons, sokoban puzzles) that are popular testbeds for Reinforcement Learning (RL) agents. Each rewrite rule can be expressed as a series of simple convolutions, allowing for environments to be parallelized on the GPU, thereby drastically accelerating RL training. Using Autoverse, we propose jump-starting open-ended learning by imitation learning from search. In such an approach, we first evolve Autoverse environments (their rules and initial map topology) to maximize the number of iterations required by greedy tree search to discover a new best solution, producing a curriculum of increasingly complex environments and playtraces. We then distill these expert playtraces into a neural-network-based policy using imitation learning. Finally, we use the learned policy as a starting point for open-ended RL, where new training environments are continually evolved to maximize the RL player agent's value function error (a proxy for its regret, or the learnability of generated environments), finding that this approach improves the performance and generality of resultant player agents.", "keywords": [ "open-ended learning", "reinforcement learning", "imitation learning", "evolution", "search" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 1, 1 ], "presentation_list": [ 1, 2, 1, 1 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "ysAX5ORQoX", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Bin Fu", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Chuyan Xiong", "gender": "Female", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Hanxuan Li", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Ruiping Wang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Ziyi Bai", "gender": "Female", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xilin Chen", "gender": "Male", "institution": "Institute of Computing Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 20, "n_ref": 62, "n_ref_all": 75, "n_fig": 6, "n_tab": 6, "L_tab": 2210, "n_element_tab": 274, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1298, "n_element_tab_1": 159, "formula_len_all": 509, "formula_len_all_1": 461, "len_all": 151261, "len_all_1": 64909, "len_abs": 2471, "len_title": 121, "len_sents": 42603, "len_sents_1": 29258, "n_sents": 353, "n_sents_1": 240, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1853, "title": "R2C: Mapping Room to Chessboard to Unlock LLM As Low-Level Action Planner", "abs": "This paper explores the potential of leveraging large language models (LLMs) as low-level action planners capable of executing long-horizon tasks based on natural language instructions. Although LLMs can act as the \"brain\" of robots by excelling in high-level task planning, they are not yet capable of directly guiding the \"body\" to execute low-level motion plans. This limitation stems from a communication gap between the \"brain\" and the \"body\". Specifically, LLMs lack access to rich spatial semantic information from the robot's real-time observations, hindering their ability to generate precise and actionable low-level plans.To address this, we propose a unified framework that bridges high-level and low-level planning by establishing an efficient communication interface between LLMs and robots. Our insight is to formulate the task as playing chess with LLMs. We map the room into a semantic chessboard, which we call Room to Chessboard (R2C). Each grid represents the position and size of objects inside the room. We find that chessboard is \\textbf{succinct} enough for LLMs to conduct semantic searches with global view of the room. Also, the chessboard is \\textbf{informative} enough to convey detailed environmental state for LLMs to predict executable low-level actions. Additionally, we enhance decision-making through a Chain-of-Thought (CoT) paradigm, improving LLMs' interpretability and action reasoning. We implement R2C using both fine-tuned open-source LLMs and closed-source models like GPT-4, and demonstrate its efficacy on the challenging ALFRED benchmark. Our results show that with communication based on chessboard, LLMs can serve as effective low-level action planners, and can generalizes well to open-vocabulary robotic planning tasks. View the demos on our project page: https://anonymous4cv.github.io/Room2Chessboard.", "keywords": [ "Embodied AI", "Large Language Model", "Embodied Instruction Following", "Robotic Planning" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ys3eqxzkeN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Badhan Chandra Das", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "PhD student" }, { "name": "M. Hadi Amini", "gender": "Not Specified", "institution": "Florida International University", "country": "US", "position": "Assistant Professor" }, { "name": "Yanzhao Wu", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 69, "n_ref": 132, "n_ref_all": 152, "n_fig": 5, "n_tab": 4, "L_tab": 1266, "n_element_tab": 153, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1417, "n_element_tab_1": 156, "formula_len_all": 98, "formula_len_all_1": 97, "len_all": 148980, "len_all_1": 68348, "len_abs": 1471, "len_title": 70, "len_sents": 36320, "len_sents_1": 32242, "n_sents": 221, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1479, "title": "Efficient Gun Detection in Real-World Videos: Challenges and Solutions", "abs": "Object detection in videos is a crucial task in the computer vision domain. Existing methods have explored different approaches to detect objects and classify the videos. However, detecting tiny objects (e.g., gun) in videos has always been a challenging and rigorous task. Moreover, the existing video analysis (detection and classification) models may not achieve high accuracy for gun detection in videos in real-world scenarios due to the lack of a large amount of labeled data. Thus, it is imperative to develop an efficient method to capture the features of tiny objects and train models that can perform accurate gun detection. To address this challenge, we make three contributions. First, we perform an empirical study of several existing video classification methods to identify the presence of guns in videos. Our extensive analysis shows that these methods may not achieve high accuracy in detecting guns in videos. Second, we propose a novel gun detection method with image-augmented training and evaluate the technique in real-world settings with different evaluation metrics. Third, our experimental results demonstrate that our proposed domain-specific method can achieve significant performance improvements in real-world settings compared to the other popular methods. We also discuss emerging challenges and critical aspects of detecting tiny objects, e.g., guns, using existing computer vision techniques, their limitations, and future research opportunities.", "keywords": [ "Image-augmented training", "transfer learning" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "ys16t9FcLN", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Patrick Jaillet", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Rafael Hanashiro", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 222, "n_formula_1": 33, "n_ref_uni": 7, "n_ref": 29, "n_ref_all": 31, "n_fig": 0, "n_tab": 3, "L_tab": 60, "n_element_tab": 2, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 26404, "formula_len_all_1": 2465, "len_all": 204118, "len_all_1": 61558, "len_abs": 1096, "len_title": 108, "len_sents": 51026, "len_sents_1": 24327, "n_sents": 674, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1123, "title": "Distribution-Dependent Rates for Multi-Distribution Learning", "abs": "To address the needs of modeling uncertainty in sensitive machine learning applications, the setup of distributionally robust optimization (DRO) seeks good performance uniformly across a variety of tasks. The recent multi-distribution learning (MDL) framework \\cite{pmlr-v195-awasthi23a-open-prob} tackles this objective in a dynamic interaction with the environment, where the learner has sampling access to each target distribution. Drawing inspiration from the field of pure-exploration multi-armed bandits, we provide \\textit{distribution-dependent} guarantees in the MDL regime, that scale with suboptimality gaps and result in superior dependence on the sample size when compared to the existing distribution-independent analyses. We investigate two non-adaptive strategies, uniform and non-uniform exploration, and present non-asymptotic regret bounds using novel tools from empirical process theory. Furthermore, we devise an adaptive optimistic algorithm, LCB-DR, that showcases enhanced dependence on the gaps, mirroring the contrast between uniform and optimistic allocation in the multi-armed bandit literature.", "keywords": [ "multi-distribution learning", "distributionally robust optimization", "pure exploration multi-armed bandits" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 2, 2, 3 ] }, { "paper_id": "yrnrvfXFaV", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Di Wang", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Huanyi Xie", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "MS student" }, { "name": "JUN ZHOU", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Lijie Hu", "gender": "Female", "institution": "KAUST", "country": "SA", "position": "PhD student" }, { "name": "Longfei Li", "gender": "Male", "institution": "ant group", "country": "", "position": "Researcher" }, { "name": "Lu Yu", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Meng Li", "gender": "unknown", "institution": "Nankai University", "country": "", "position": "MS student" }, { "name": "Tianhao Alex Huang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 60, "n_ref": 93, "n_ref_all": 108, "n_fig": 5, "n_tab": 9, "L_tab": 4927, "n_element_tab": 435, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 2765, "n_element_tab_1": 318, "formula_len_all": 857, "formula_len_all_1": 759, "len_all": 184351, "len_all_1": 78730, "len_abs": 1298, "len_title": 120, "len_sents": 46480, "len_sents_1": 33394, "n_sents": 389, "n_sents_1": 262, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1305, "title": "Low-cost Enhancer for Text Attributed Graph Learning via Graph Alignment", "abs": "Many graphs can be represented as Text-attributed Graphs (TAGs). Due to the rich textual information present in each node of TAGs, traditional graph neural networks (GNNs) often struggle to deliver satisfactory performance. Recent advancements leveraging large language models (LLMs) to augment new node text features have notably enhanced node representations, resulting in significant performance improvements. However, these methods typically require extensive annotations or fine-tuning on all nodes, which are both time-consuming and expensive. To address this challenge, we propose GAGA, a novel and lightweight framework for TAG representation learning. GAGA employs a more efficient strategy by annotating only representative nodes and edges, thereby reducing both annotation time and cost. It further capitalizes on these annotations by constructing an annotation graph that captures the topological relationships among them. Additionally, GAGA introduces a two-level alignment module to integrate the annotation graph with the TAG, ensuring effective alignment of their underlying structures. Experiments demonstrate that GAGA achieves classification accuracies comparable to or exceeding state-of-the-art methods while requiring only 1\\% of the data to be annotated, making it highly efficient.", "keywords": [ "Text-attributed Graphs" ], "rating_list": [ 3, 3, 8, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "yrf5RmaHfG", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 1, "author_info_list": [ { "name": "Ezgi Korkmaz", "gender": "unknown", "institution": "DeepMind", "country": "GB", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 23, "n_ref_uni": 38, "n_ref": 65, "n_ref_all": 76, "n_fig": 4, "n_tab": 2, "L_tab": 186, "n_element_tab": 22, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 295, "n_element_tab_1": 22, "formula_len_all": 2852, "formula_len_all_1": 2625, "len_all": 115685, "len_all_1": 72801, "len_abs": 1494, "len_title": 131, "len_sents": 36528, "len_sents_1": 33584, "n_sents": 246, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1500, "title": "JuxtAlign: A Foundational Analysis on Alignment of Certified Reinforcement Learning", "abs": "Sequential decision making in highly complex MDPs with high-dimensional observations and state dynamics became possible with the progress achieved in deep reinforcement learning research. At the same time, deep neural policies have been observed to be highly unstable with respect to the minor sensitivities in their state space induced by non-robust directions. To alleviate these volatilities a line of work suggested techniques to cope with this problem via explicitly regularizing the temporal difference loss for the worst-case sensitivity. \nIn this paper we provide theoretical foundations on the failure instances of the approaches proposed to overcome instabilities of the deep neural policy manifolds. Our comprehensive analysis reveals that certified reinforcement learning learns misaligned values. Our empirical analysis in the Arcade Learning Environment further demonstrates that the state-of-the-art certified policies learn inconsistent and overestimated value functions compared to standard training techniques. In connection to this analysis, we highlight the intrinsic gap between how natural intelligence understands and interacts with an environment in contrast to policies learnt via certified training. This intrinsic gap between natural intelligence and the restrictions induced by certified training on the capabilities of artificial intelligence further demonstrates the need to rethink the approach in establishing reliable and aligned deep reinforcement learning policies.", "keywords": [ "alignment", "juxtaposition", "reinforcement learning" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 1, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "yr7PjzmkQ6", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Christian A Shewmake", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Erik J Bekkers", "gender": "unknown", "institution": "University of Amsterdam", "country": "", "position": "Assistant Professor" }, { "name": "Mohammad Mohaiminul Islam", "gender": "Male", "institution": "University of Amsterdam, University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Sharvaree Vadgama", "gender": "Female", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Domas Antanas Buracas", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "Researcher" } ], "n_formula": 32, "n_formula_1": 13, "n_ref_uni": 43, "n_ref": 69, "n_ref_all": 85, "n_fig": 4, "n_tab": 3, "L_tab": 829, "n_element_tab": 198, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2251, "n_element_tab_1": 388, "formula_len_all": 2671, "formula_len_all_1": 735, "len_all": 151773, "len_all_1": 67543, "len_abs": 3544, "len_title": 147, "len_sents": 45251, "len_sents_1": 28689, "n_sents": 345, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 956, "title": "On the utility of Equivariance and Symmetry Breaking in Deep learning architectures on point clouds", "abs": "This paper explores the key factors that influence the performance of models working with point clouds, \\edit{across different tasks of varying geometric complexity.} In this work, we explore the trade-offs between flexibility and weight-sharing introduced by equivariant layers, assessing when equivariance boosts or detracts from performance. It is often argued that providing more information as input improves a model's performance. However, if this additional information breaks certain properties, such as $SE(3)$ equivariance, does it remain beneficial? We identify the key aspects of equivariant and non-equivariant architectures that drive success in different tasks by benchmarking them on segmentation, regression, and generation tasks across multiple datasets with increasing complexity. We observe a positive impact of equivariance, which becomes more pronounced with increasing task complexity, even when strict equivariance is not required.", "keywords": [ "deep learning architectures", "geometric deep learning", "equivariance", "group convolutional networks", "generative modeling" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "yr0l1IoyzV", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 5, "author_info_list": [ { "name": "Depeng Jin", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Jun Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Junbo Yan", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Wenxuan Ao", "gender": "Male", "institution": "Department of Electronic Engineering, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 39, "n_ref": 282, "n_ref_all": 312, "n_fig": 11, "n_tab": 16, "L_tab": 4365, "n_element_tab": 351, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 361, "n_element_tab_1": 73, "formula_len_all": 557, "formula_len_all_1": 257, "len_all": 183072, "len_all_1": 70497, "len_abs": 2095, "len_title": 140, "len_sents": 60733, "len_sents_1": 35022, "n_sents": 500, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 2105, "title": "A GPU-accelerated Large-scale Simulator for Transportation System Optimization Benchmarking", "abs": "With the development of artificial intelligence techniques, transportation system optimization is evolving from traditional methods relying on expert experience to simulation and learning-based decision and optimization methods.\nLearning-based optimization methods require extensive interactions with highly realistic microscopic traffic simulators.\nHowever, existing microscopic traffic simulators are inefficient in large-scale scenarios and thus fail to support the adoption of these methods in large-scale transportation system optimization scenarios.\nIn addition, the optimization scenarios supported by existing simulators are limited, mainly focusing on the traffic signal control.\nTo address these challenges, we propose the first open-source GPU-accelerated large-scale microscopic simulator for transportation system simulation and optimization.\nThe simulator can iterate at 84.09Hz, which achieves 88.92 times computational acceleration in the large-scale scenario with 2,464,950 vehicles compared to the best baseline CityFlow.\nBesides, it achieves a more realistic average road speeds simulated on real datasets by adopting the IDM model as the car-following model and the randomized MOBIL model as the lane-changing model.\nBased on it, we implement a set of microscopic and macroscopic controllable objects and metrics provided by Python API to support typical transportation system optimization scenarios including traffic signal control, dynamic lane assignment within junctions, tidal lane control, congestion pricing, road planning, e.t.c.\nWe choose five representative transportation system optimization scenarios and benchmark classical rule-based algorithms, reinforcement learning algorithms, and black-box optimization algorithms in four cities.\nThese experiments effectively demonstrate the usability of the simulator for large-scale traffic system optimization.\nThe anonymous code of the simulator is available at https://anonymous.4open.science/r/moss-AF45 and the others are shown at Appendix A.\nIn addition, we build an open-registration web platform to support no-code trials.", "keywords": [ "microscopic traffic simulator", "transportation system optimization", "GPU acceleration" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "yqaN7MfkFU", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Bingyao Huang", "gender": "Male", "institution": "Guangdong University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Junfeng Huo", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Liuhua Peng", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Associate Professor" }, { "name": "Yanyan Liu", "gender": "Female", "institution": "Wuhan University", "country": "", "position": "Full Professor" } ], "n_formula": 60, "n_formula_1": 28, "n_ref_uni": 45, "n_ref": 62, "n_ref_all": 79, "n_fig": 2, "n_tab": 9, "L_tab": 6446, "n_element_tab": 566, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 3894, "n_element_tab_1": 189, "formula_len_all": 6662, "formula_len_all_1": 2936, "len_all": 162573, "len_all_1": 69914, "len_abs": 1002, "len_title": 107, "len_sents": 41331, "len_sents_1": 26582, "n_sents": 373, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1050, "title": "Regularized Maximum Mean Discrepancy for Variable Selection", "abs": "In this paper, we propose a variable selection method based on maximum mean discrepancy (MMD) to effectively identify important variables that contribute to distributional differences between two samples. We begin by assigning weights to each variable and then optimizing these weights within a regularized MMD framework. The optimized weights serve as an importance measure for each variable and can be leveraged for variable selection. Additionally, using the optimized weights, we design two algorithms aimed at enhancing test power and improving classification accuracy for two-sample tests and classification problems. Our method is model-free and makes no assumptions about the underlying structure of the data. Moreover, we propose an acceleration method to improve computational efficiency.\nWe also provide theoretical guarantees, including the consistency of the estimated weights and the convergence of our acceleration algorithms. Through numerical simulations and real-world datasets, we validate the effectiveness of the proposed method.", "keywords": [ "Variable selection", "Maximum mean discrepancy", "Two-sample tests", "Binary classification" ], "rating_list": [ 3, 5, 6, 3, 5 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 2, 4, 3, 4, 3 ] }, { "paper_id": "yqST7JwsCt", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Lin Wang", "gender": "unknown", "institution": "CUHK", "country": "CN", "position": "PhD student" }, { "name": "Sai Praneeth Karimireddy", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Xiaoying Tang", "gender": "Female", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Ye Shi", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhichao Wang", "gender": "Male", "institution": "The Chinese University of HongKong, Shenzhen", "country": "CN", "position": "MS student" } ], "n_formula": 231, "n_formula_1": 37, "n_ref_uni": 46, "n_ref": 147, "n_ref_all": 204, "n_fig": 11, "n_tab": 15, "L_tab": 7063, "n_element_tab": 571, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 4034, "n_element_tab_1": 201, "formula_len_all": 21114, "formula_len_all_1": 2737, "len_all": 298641, "len_all_1": 64789, "len_abs": 1447, "len_title": 67, "len_sents": 79738, "len_sents_1": 22746, "n_sents": 802, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1452, "title": "Entropy-Based Aggregation for Fair and Effective Federated Learning", "abs": "Federated Learning (FL) enables collaborative model training across distributed devices while preserving data privacy. Nonetheless, the heterogeneity of edge devices often leads to inconsistent performance of the globally trained models, resulting in unfair outcomes among users. Existing federated fairness algorithms strive to enhance fairness but often fall short in maintaining the overall performance of the global model, typically measured by the average accuracy across all clients. To address this issue, we propose a novel algorithm that leverages entropy-based aggregation combined with model and gradient alignments to simultaneously optimize fairness and global model performance. Our method employs a bi-level optimization framework, where we derive an analytic solution to the aggregation probability in the inner loop, making the optimization process computationally efficient. Additionally, we introduce an innovative alignment update and an adaptive strategy in the outer loop to further balance global model's performance and fairness. Theoretical analysis indicates that our approach guarantees convergence even in non-convex FL settings and demonstrates significant fairness improvements in generalized regression and strongly convex models. Empirically, our approach surpasses state-of-the-art federated fairness algorithms, ensuring consistent performance among clients while improving the overall performance of the global model.", "keywords": [ "Fairness", "Heterogeneous Federated Learning" ], "rating_list": [ 8, 6, 5, 6, 5 ], "soundness_list": [ 4, 4, 2, 2, 3 ], "presentation_list": [ 4, 3, 2, 2, 3 ], "contribution_list": [ 4, 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "yqJoqtUwSI", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Dong Zheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jun Xiao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Kexin Li", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Yi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zongxin Yang", "gender": "Male", "institution": "Harvard Medical School, Harvard University", "country": "US", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 22, "n_ref": 51, "n_ref_all": 63, "n_fig": 4, "n_tab": 3, "L_tab": 1255, "n_element_tab": 115, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 807, "n_element_tab_1": 39, "formula_len_all": 339, "formula_len_all_1": 339, "len_all": 107908, "len_all_1": 51925, "len_abs": 242, "len_title": 0, "len_sents": 35076, "len_sents_1": 24394, "n_sents": 271, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1708, "title": "Collaborative Hybrid Propagator for Temporal Misalignment in Audio-Visual Segmentation", "abs": "Audio-visual video segmentation (AVVS) aims to generate pixel-level maps of sound-producing objects that accurately align with the corresponding audio. However, existing methods often face temporal misalignment, where audio cues and segmentation results are not temporally coordinated. Audio provides two critical pieces of information: i) target object-level details and ii) the timing of when objects start and stop producing sounds. Current methods focus more on object-level information but neglect the boundaries of audio semantic changes, leading to temporal misalignment. To address this issue, we propose a Collaborative Hybrid Propagator Framework~(Co-Prop). This framework includes two main steps: Preliminary Audio Boundary Anchoring and Frame-by-Frame Audio-Insert Propagation. To Anchor the audio boundary, we employ retrieval-assist prompts with Qwen large language models to identify control points of audio semantic changes. These control points split the audio into semantically consistent audio portions. After obtaining the control point lists, we propose the Audio Insertion Propagator to process each audio portion using a frame-by-frame audio insertion propagation and matching approach. We curated a compact dataset comprising diverse source conversion cases and devised a metric to assess alignment rates. Compared to traditional simultaneous processing methods, our approach reduces memory requirements and facilitates frame alignment. Experimental results demonstrate the effectiveness of our approach across three datasets and two backbones. Furthermore, our method can be integrated with existing AVVS approaches, offering plug-and-play functionality to enhance their performance.", "keywords": [ "audio-visual video segmentation" ], "rating_list": [ 8, 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 3, 1 ], "presentation_list": [ 3, 3, 3, 4, 1 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 4, 4 ] }, { "paper_id": "ypBYdetYd9", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Ann Huang", "gender": "Female", "institution": "Harvard University, Harvard University", "country": "", "position": "PhD student" }, { "name": "Kanaka Rajan", "gender": "Female", "institution": "Icahn School of Medicine at Mount Sinai", "country": "", "position": "Assistant Professor" }, { "name": "Satpreet Harcharan Singh", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 55, "n_ref": 75, "n_ref_all": 90, "n_fig": 11, "n_tab": 2, "L_tab": 234, "n_element_tab": 29, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 115, "n_element_tab_1": 15, "formula_len_all": 281, "formula_len_all_1": 185, "len_all": 130384, "len_all_1": 59634, "len_abs": 1587, "len_title": 25, "len_sents": 38907, "len_sents_1": 30357, "n_sents": 221, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1594, "title": "Measuring and Controlling Solution Degeneracy across Task-Trained Recurrent Neural Networks", "abs": "Task-trained recurrent neural networks (RNNs) are versatile models of dynamical processes widely used in machine learning and neuroscience. While RNNs are easily trained to perform a wide range of tasks, the nature and extent of the degeneracy in the resultant solutions (i.e., the variability across trained RNNs) remain poorly understood. Here, we provide a unified framework for analyzing degeneracy across three levels: behavior, neural dynamics, and weight space. We analyzed RNNs trained on diverse tasks across machine learning and neuroscience domains, including N-bit flip-flop, sine wave generation, delayed discrimination, and path integration. \nOur key finding is that the variability across RNN solutions, quantified on the basis of neural dynamics and trained weights, depends primarily on network capacity and task characteristics such as complexity. We introduce information-theoretic measures to quantify task complexity and demonstrate that increasing task complexity consistently reduces degeneracy in neural dynamics and generalization behavior while increasing degeneracy in weight space. These relationships hold across diverse tasks and can be used to control the degeneracy of the solution space of task-trained RNNs. Furthermore, we provide several strategies to control solution degeneracy, enabling task-trained RNNs to learn more consistent or diverse solutions as needed. We envision that these insights will lead to more reliable machine learning models and could inspire strategies to better understand and control degeneracy observed in neuroscience experiments.", "keywords": [ "Recurrent Neural Network", "Dynamical System", "Neural Computation", "Computational Neuroscience" ], "rating_list": [ 5, 3, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 2, 2 ], "presentation_list": [ 3, 4, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 1, 3 ], "confidence_list": [ 2, 4, 4, 3, 5 ] }, { "paper_id": "yp95goUAT1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Alexander Richard Fabbri", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Chien-Sheng Wu", "gender": "Male", "institution": "Salesforce AI", "country": "US", "position": "Researcher" }, { "name": "Gabriel Bernadett-Shapiro", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Instructor" }, { "name": "Nan Zhang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Prafulla Kumar Choubey", "gender": "Male", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "Prasenjit Mitra", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Full Professor" }, { "name": "Rui Zhang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 27, "n_ref": 46, "n_ref_all": 78, "n_fig": 6, "n_tab": 8, "L_tab": 1416, "n_element_tab": 116, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 564, "n_element_tab_1": 34, "formula_len_all": 68, "formula_len_all_1": 67, "len_all": 125965, "len_all_1": 63469, "len_abs": 1365, "len_title": 121, "len_sents": 42823, "len_sents_1": 31863, "n_sents": 306, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1372, "title": "SiReRAG: Indexing Similar and Related Information for Multihop Reasoning", "abs": "Indexing is an important step towards strong performance in retrieval-augmented generation (RAG) systems. However, existing methods organize data based on either semantic similarity (similarity) or related information (relatedness), but do not cover both perspectives comprehensively. Our analysis reveals that modeling only one perspective results in insufficient knowledge synthesis, leading to suboptimal performance on complex tasks requiring multihop reasoning. In this paper, we propose SiReRAG, a novel RAG indexing approach that explicitly considers both similar and related information. On the similarity side, we follow existing work and explore some variances to construct a similarity tree based on recursive summarization. On the relatedness side, SiReRAG extracts propositions and entities from texts, groups propositions via shared entities, and generates recursive summaries to construct a relatedness tree. We index and flatten both similarity and relatedness trees into a unified retrieval pool. Our experiments demonstrate that SiReRAG consistently outperforms state-of-the-art indexing methods on three multihop datasets (MuSiQue, 2WikiMultiHopQA, and HotpotQA), with an average 1.9% improvement in F1 scores. With our unified retrieval pool, SiReRAG enhances existing reranking methods significantly, with up to 7.8% improvement in average F1 scores.", "keywords": [ "Retrieval-augmented generation (RAG)", "RAG indexing", "Multi-hop question answering" ], "rating_list": [ 5, 8, 8, 6 ], "soundness_list": [ 2, 4, 4, 3 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "yougZBoUY3", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Aengus Lynch", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Ethan Perez", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Fazl Barez", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Henry Sleight", "gender": "Male", "institution": "Constellation", "country": "US", "position": "Researcher" }, { "name": "John Hughes", "gender": "Male", "institution": "ML Alignment & Theory Scholars", "country": "GB", "position": "Researcher" }, { "name": "Mrinank Sharma", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Rylan Schaeffer", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" }, { "name": "Sara Brittany Price", "gender": "Female", "institution": "Machine Learning Alignment & Theory Scholars", "country": "", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 54, "n_ref": 77, "n_ref_all": 133, "n_fig": 23, "n_tab": 4, "L_tab": 825, "n_element_tab": 44, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 152, "formula_len_all_1": 0, "len_all": 239719, "len_all_1": 58226, "len_abs": 4738, "len_title": 107, "len_sents": 78812, "len_sents_1": 28437, "n_sents": 594, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1361, "title": "Attacking Audio Language Models with Best-of-N Jailbreaking", "abs": "In this work, we investigate the susceptibility of Audio Language Models (ALMs) to audio-based jailbreaks and introduce Best-of-N (BoN) Jailbreaking, a black-box jailbreaking algorithm to extract harmful information from ALMs. To craft jailbreak inputs, our approach samples audio augmentations and applies them to malicious prompts. We repeat this process until we find a set of augmentations that elicits a harmful response from the target ALM. Empirically, we find that applying BoN with 7000 sampled augmentations achieves an attack success rate (ASR) of over 60% on all models tested, including the preview model for the released GPT-4o. Furthermore, we uncover power laws that accurately predict the ASR of BoN jailbreaking as a function of the number of samples. These power laws allow us to forecast the effectiveness of BoN jailbreaking as a function of the number of sampled augmentations over an order of magnitude. Finally, we show that BoN jailbreaking can be composed with other black-box attack algorithms for even more effective attacks\u2014combining BoN with an optimized prefix attack achieves 98% ASR on Gemini Pro and Flash. Overall, by exploiting stochastic sampling and sensitivity to variations in a high-dimensional input space, we propose a scalable, composable, and highly effective black-box algorithm for attacking state-of-the-art ALMs.", "keywords": [ "adversarial robustness", "jailbreaks", "audio language model", "speech language model", "multimodal", "adversarial attack", "audio jailbreak", "safety", "trustworthy", "robustness" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 3, 1, 3 ], "presentation_list": [ 2, 1, 4 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "ymt4crbbXh", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Evan Zheran Liu", "gender": "Male", "institution": "Generally Intelligent", "country": "", "position": "Researcher" }, { "name": "Farzaan Kaiyom", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Percy Liang", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Tatsunori Hashimoto", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Xiang Lisa Li", "gender": "Female", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Yifan Mai", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 20, "n_ref": 40, "n_ref_all": 71, "n_fig": 7, "n_tab": 13, "L_tab": 4740, "n_element_tab": 738, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 728, "n_element_tab_1": 111, "formula_len_all": 539, "formula_len_all_1": 539, "len_all": 205409, "len_all_1": 65310, "len_abs": 1503, "len_title": 103, "len_sents": 57665, "len_sents_1": 31488, "n_sents": 456, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1428, "title": "AutoBencher: Towards Declarative Benchmark Construction", "abs": "We present AutoBencher, a declarative framework for automatic benchmark construction, and use it to scalably discover novel insights and vulnerabilities of existing language models. Concretely, given a few desiderata of benchmarks (e.g., question difficulty, topic salience), we operationalize each desideratum and cast benchmark creation as an optimization problem. Specifically, we experiment with two settings with different optimization objectives: (i) for capability evaluation, we declare the goal of finding a salient, difficult dataset that induces novel performance patterns; (ii) for safety evaluation, we declare the goal of finding a dataset of unsafe prompts that existing LMs fail to decline. To tackle this type of optimization problem, we propose to use a language model to automatically construct datasets and iteratively revise the dataset to optimize for the declared desiderata. We use AutoBencher (powered by GPT-4) to create datasets for math, multilinguality, knowledge, and safety. The scalability of AutoBencher allows it to test fine-grained categories and tail knowledge, creating datasets that are on average 27% more novel and 22% more difficult than existing benchmarks. AutoBencher also helps identify specific gaps not captured by existing benchmarks: e.g., Gemini-Pro has knowledge gaps on Permian Extinction and Fordism while GPT-4o fails to decline harmful requests about cryptocurrency scams.", "keywords": [ "automatic evaluation", "language models" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 3, 3, 2, 5 ] }, { "paper_id": "ymqLAmqYHW", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Dong Yuan", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Huaming Chen", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Jiayu Zhang", "gender": "Male", "institution": "Suzhou Yierqi", "country": "", "position": "Researcher" }, { "name": "Xinyi Wang", "gender": "Male", "institution": "Universiti Malaya", "country": "", "position": "MS student" }, { "name": "Zhibo Jin", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Zhu Zhiyu", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 15, "n_ref": 60, "n_ref_all": 96, "n_fig": 13, "n_tab": 29, "L_tab": 44875, "n_element_tab": 2419, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 292, "n_element_tab_1": 10, "formula_len_all": 423, "formula_len_all_1": 289, "len_all": 243092, "len_all_1": 63088, "len_abs": 1063, "len_title": 100, "len_sents": 61686, "len_sents_1": 30578, "n_sents": 659, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1022, "title": "K&L: Penetrating Backdoor Defense with Key and Locks", "abs": "Backdoor attacks in machine learning create hidden vulnerability by manipulating the model behaviour with specific triggers. Such attacks often remain unnoticed as the model operates as expected for normal input. Thus, it is imperative to understand the intricate mechanism of backdoor attacks. To address this challenge, in this work, we introduce three key requirements that a backdoor attack must meet. Moreover, we note that current backdoor attack algorithms, whether employing fixed or input-dependent triggers, exhibit a high binding with model parameters, rendering them easier to defend against. To tackle this issue, we propose the Key-Locks algorithm, which separates the backdoor attack process into embedding locks and employing a key for unlocking. This method enables the adjustment of unlocking levels to counteract diverse defense mechanisms. Extensive experiments are conducted to evaluate the effective of our proposed algorithm. Our code is available at: https://anonymous.4open.science/r/KeyLocks-FD85", "keywords": [ "backdoor attack", "backdoor defense", "AI security" ], "rating_list": [ 1, 6, 5, 6 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "ym7pr83XQr", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Bing Su", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Jiahao Chen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Jiangmeng Li", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Rui Jiao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Wenbing Huang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yurou Liu", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 46, "n_formula_1": 21, "n_ref_uni": 51, "n_ref": 129, "n_ref_all": 146, "n_fig": 9, "n_tab": 14, "L_tab": 5940, "n_element_tab": 526, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3131, "n_element_tab_1": 227, "formula_len_all": 3815, "formula_len_all_1": 784, "len_all": 189218, "len_all_1": 66460, "len_abs": 1435, "len_title": 152, "len_sents": 56456, "len_sents_1": 29047, "n_sents": 419, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1440, "title": "DenoiseVAE: Learning Molecule-Adaptive Noise Distributions for Denoising-based 3D Molecular Pre-training", "abs": "Denoising learning of 3D molecules learns molecular representations by imposing noises into the equilibrium conformation and predicting the added noises to recover the equilibrium conformation, which essentially captures the information of molecular force fields. Due to the specificity of Potential Energy Surfaces, the probabilities of physically reasonable noises for each atom in different molecules are different. However, existing methods apply the shared heuristic hand-crafted noise sampling strategy to all molecules, resulting in inaccurate force field learning. In this paper, we propose a novel 3D molecular pre-training method, namely DenoiseVAE, which employs a Noise Generator to acquire atom-specific noise distributions for different molecules. It utilizes the stochastic reparameterization technique to sample noisy conformations from the generated distributions, which are inputted into a Denoising Module for denoising. The Noise Generator and the Denoising Module are jointly learned in a manner conforming with the paradigm of Variational Auto Encoder. Consequently, the sampled noisy conformations can be more diverse, adaptive, and informative, and thus DenoiseVAE can learn representations that better reveal the molecular force fields. Extensive experiments show that DenoiseVAE outperforms the current state-of-the-art methods on various molecular property prediction tasks, demonstrating the effectiveness of it.", "keywords": [ "3D Molecular pre-training via denoising", "Molecular property prediction" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 4, 3, 4, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 2 ] }, { "paper_id": "ym1dS37mZE", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Chuanyang Zheng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Han Shi", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Principal Researcher" }, { "name": "Hong Cheng", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Professor" }, { "name": "Minbin Huang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Runhui Huang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xiangguo Sun", "gender": "Male", "institution": "Chinese University of Hong Kong, The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Xin Jiang", "gender": "Male", "institution": "Noah\u2019s Ark Lab, Huawei Technologies", "country": "HK", "position": "Principal Researcher" }, { "name": "Xinchi Deng", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "MS student" }, { "name": "Yimeng Chen", "gender": "unknown", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" }, { "name": "Zhenguo Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 21, "n_ref": 37, "n_ref_all": 43, "n_fig": 4, "n_tab": 2, "L_tab": 100, "n_element_tab": 17, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 100, "n_element_tab_1": 17, "formula_len_all": 384, "formula_len_all_1": 384, "len_all": 90088, "len_all_1": 58870, "len_abs": 962, "len_title": 70, "len_sents": 26206, "len_sents_1": 26208, "n_sents": 182, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 978, "title": "Efficient Multi-modal Large Language Models via Visual Token Grouping", "abs": "The development of Multi-modal Large Language Models (MLLMs) has significantly advanced various downstream applications, including visual question answering and image captioning. However, the substantial computational costs associated with processing high-resolution images and videos pose a barrier to their broader adoption. To address this challenge, compressing vision tokens in MLLMs has emerged as a promising approach to reduce inference costs. In this paper, we introduce \\methodname, a novel grouping mechanism that leverages the capabilities of pretrained vision encoders to group similar image segments without the need for segmentation masks. With the isolated attention we adopt, \\methodname can identify and eliminate redundant visual tokens, which effectively reduces computational demands. Extensive experiments demonstrate that the effectiveness of\\methodname , maintains over 98.1% of the original performance while achieving a reduction of over 27% in TFLOPS.", "keywords": [ "Large Language Model", "Multi-modal Learning" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "ylhKbwJrjC", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Hirota Kinoshita", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "MS student" }, { "name": "Segev Wasserkrug", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Takayuki Osogami", "gender": "Male", "institution": "International Business Machines", "country": "JP", "position": "Principal Researcher" } ], "n_formula": 59, "n_formula_1": 25, "n_ref_uni": 50, "n_ref": 133, "n_ref_all": 228, "n_fig": 9, "n_tab": 4, "L_tab": 582, "n_element_tab": 132, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4465, "formula_len_all_1": 1532, "len_all": 217743, "len_all_1": 71500, "len_abs": 1340, "len_title": 40, "len_sents": 72828, "len_sents_1": 34698, "n_sents": 545, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1340, "title": "Mechanism design with multi-armed bandit", "abs": "A popular approach of automated mechanism design is to formulate a linear program (LP) whose solution gives a mechanism with desired properties. We analytically derive a class of optimal solutions for such an LP that gives mechanisms achieving standard properties of efficiency, incentive compatibility, strong budget balance (SBB), and individual rationality (IR), where SBB and IR are satisfied in expectation. Notably, our solutions are represented by an exponentially smaller number of essential variables than the original variables of LP. Our solutions, however, involve a term whose exact evaluation requires solving a certain optimization problem exponentially many times as the number of players grows. We thus evaluate this term by modeling it as the problem of estimating the mean reward of the best arm in multi-armed bandit (MAB), propose a Probably and Approximately Correct estimator, and prove its asymptotic optimality by establishing a lower bound on its sample complexity. This MAB approach reduces the number of times the optimization problem is solved from exponential to linear. Numerical experiments show that the proposed approach finds mechanisms that are guaranteed to achieve desired properties with high probability for environments with up to 128 players, which substantially improves upon the prior work.", "keywords": [ "mechanism design", "incentive compatibility", "efficiency", "individual rationality", "budget balance", "multi-armed bandit", "probably approximately correct" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 1 ], "confidence_list": [ 2, 2, 3 ] }, { "paper_id": "ylgg2RE7ub", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Eunho Shin", "gender": "Male", "institution": "Sogang University", "country": "KR", "position": "MS student" }, { "name": "Jimin Roh", "gender": "Male", "institution": "Sogang University", "country": "KR", "position": "MS student" }, { "name": "Joonsoo Kim", "gender": "Male", "institution": "Electronics and Telecommunications Research Institute", "country": "KR", "position": "Researcher" }, { "name": "Kyeongbo Kong", "gender": "Male", "institution": "Pusan National University", "country": "KR", "position": "Assistant Professor" }, { "name": "Songju Na", "gender": "unknown", "institution": "Samsung Electronics", "country": "KR", "position": "Researcher" }, { "name": "Suk-Ju Kang", "gender": "Male", "institution": "Sogang University", "country": "KR", "position": "Full Professor" }, { "name": "Yeomsuwoong", "gender": "Male", "institution": "Sogang University", "country": "KR", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 25, "n_ref_uni": 26, "n_ref": 66, "n_ref_all": 86, "n_fig": 6, "n_tab": 3, "L_tab": 473, "n_element_tab": 53, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 166, "n_element_tab_1": 16, "formula_len_all": 1236, "formula_len_all_1": 1269, "len_all": 118798, "len_all_1": 61075, "len_abs": 1585, "len_title": 107, "len_sents": 35433, "len_sents_1": 28694, "n_sents": 255, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1593, "title": "IF-MODGS : INITIAL FREE MONOCULAR DYNAMIC GAUSSIAN SPLATTING", "abs": "In the field of scene reconstruction with moving objects, recent studies have utilized 3D Gaussian Splatting (3DGS) for spatial representation. This method typically relies on camera poses and point clouds obtained through the Structure-from-Motion (SfM) algorithm. However, in scenes captured with monocular viewpoints and containing moving objects in each frame, the SfM algorithm struggles to obtain accurate camera poses and points clouds. As a result, it often either removes point clouds of dynamic objects or fails to find camera poses for each frame, thereby leading to sub-optimal rendering of dynamic scenes. We propose a novel approach, Initial-Free Monocular Dynamic Gaussian Splatting (IF-MoDGS) which does not require precomputed camera poses and point clouds in dynamic scenes with moving objects. Our approach estimates camera poses using the static background, separated from dynamic objects by a motion mask, and generates point clouds specifically for the dynamic objects. To handle dynamic objects, we define a canonical space and apply deformation to link it with each viewpoint and timestamp. Then, to improve quality in complex spatio-temporal scenes, we utilize a high-dimensional feature loss and an annealing frequency loss. Extensive experimental results demonstrate that our method can effectively render dynamic scenes without relying on precomputed camera poses and point clouds, achieving the state-of-the-art performance in dynamic scene rendering tasks using a monocular camera. Our project will be available at:https://anonymous.4open.science/w/IF-MODGS-67F5/", "keywords": [ "novel view synthesis", "4D rendering", "camera pose estimation", "3D reconstruction" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "ykuc5q381b", "primary_area": "datasets and benchmarks", "n_author": 15, "author_info_list": [ { "name": "Danqi Chen", "gender": "Female", "institution": "Department of Computer Science, Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Hongjin SU", "gender": "unknown", "institution": "the University of Hong Kong, University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Howard Yen", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Jinsung Yoon", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Liu Haisu", "gender": "Not Specified", "institution": "University of Hong Kong", "country": "", "position": "Undergrad student" }, { "name": "Mengzhou Xia", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Michael Tang", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "Undergrad student" }, { "name": "Niklas Muennighoff", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Ruoxi Sun", "gender": "Female", "institution": "Google", "country": "", "position": "Google" }, { "name": "Sercan O Arik", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Tao Yu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Weijia Shi", "gender": "unknown", "institution": "University of Washington, Seattle", "country": "US", "position": "PhD student" }, { "name": "Zachary S Siegel", "gender": "Male", "institution": "Department of Computer Science, Princeton University", "country": "", "position": "Undergrad student" }, { "name": "Wang Han-yu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Undergrad student" }, { "name": "Ben Shi", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Undergrad student" } ], "n_formula": 14, "n_formula_1": 2, "n_ref_uni": 22, "n_ref": 85, "n_ref_all": 155, "n_fig": 5, "n_tab": 28, "L_tab": 17355, "n_element_tab": 2132, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 245, "n_element_tab_1": 41, "formula_len_all": 563, "formula_len_all_1": 156, "len_all": 340889, "len_all_1": 65725, "len_abs": 1643, "len_title": 128, "len_sents": 101359, "len_sents_1": 29767, "n_sents": 982, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 79, "L_abs": 1486, "title": "BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval", "abs": "Existing retrieval benchmarks primarily consist of information-seeking queries (e.g., aggregated questions from search engines) where keyword or semantic-based retrieval is usually sufficient. However, many complex real-world queries require in-depth reasoning to identify relevant documents that go beyond surface form matching. For example, finding documentation for a coding question requires understanding the logic and syntax of the functions involved. To better benchmark retrieval on such challenging queries, we introduce BRIGHT, the first text retrieval benchmark that requires intensive reasoning to retrieve relevant documents. Our dataset consists of 1,398 real-world queries spanning diverse domains such as economics, psychology, mathematics, coding, and more. These queries are drawn from naturally occurring or carefully curated human data. Extensive evaluation reveals that even state-of-the-art retrieval models perform poorly on BRIGHT. The leading model on the MTEB leaderboard (Muennighoff et al., 2023), which achieves a score of 59.0 nDCG@10,1 produces a score of nDCG@10 of 18.0 on BRIGHT. We show that incorporating explicit reasoning about the query improves retrieval performance by up to 12.2 points. Moreover, incorporating retrieved documents from the top-performing retriever boosts question answering performance by over 6.6 points. We believe that BRIGHT paves the way for future research on retrieval systems in more realistic and challenging settings.", "keywords": [ "Retrieval benchmark", "Reasoning" ], "rating_list": [ 10, 6, 6, 8, 6 ], "soundness_list": [ 4, 3, 3, 3, 4 ], "presentation_list": [ 4, 3, 4, 3, 3 ], "contribution_list": [ 4, 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "ykt6I21YQZ", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Austin Wang", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Hongkai Zheng", "gender": "unknown", "institution": "California Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Nikola Borislavov Kovachki", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Ricardo Baptista", "gender": "Male", "institution": "Deparment of Computing + Mathematical Sciences, California Institute of Technology", "country": "", "position": "Instructor" }, { "name": "Wenda Chu", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yisong Yue", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 56, "n_formula_1": 33, "n_ref_uni": 41, "n_ref": 76, "n_ref_all": 119, "n_fig": 10, "n_tab": 9, "L_tab": 3554, "n_element_tab": 399, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 994, "n_element_tab_1": 143, "formula_len_all": 5733, "formula_len_all_1": 2475, "len_all": 147428, "len_all_1": 63998, "len_abs": 1099, "len_title": 129, "len_sents": 37971, "len_sents_1": 26676, "n_sents": 316, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1129, "title": "Ensemble Kalman Diffusion Guidance: A Derivative-free Method for Inverse Problems", "abs": "When solving inverse problems, it is increasingly popular to use pre-trained diffusion models as plug-and-play priors. This framework can accommodate different forward models without re-training while preserving the generative capability of diffusion models. Despite their success in many imaging inverse problems, most existing methods rely on privileged information such as derivative, pseudo-inverse, or full knowledge about the forward model. This reliance poses a substantial limitation that restricts their use in a wide range of problems where such information is unavailable, such as many scientific applications. To address this, we propose Ensemble Kalman Diffusion Guidance (EnKG) for diffusion models, a derivative-free approach that can solve inverse problems by only accessing forward model evaluations and a pre-trained diffusion model. We study the empirical effectiveness of our method across various inverse problems, including scientific settings such as inferring fluid flows and astronomical objects, which are highly non-linear inverse problems that often only permit black-box access to the forward model.", "keywords": [ "inverse problem", "diffusion model", "derivative-free" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "yklJpvB7Dq", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Atul Prakash", "gender": "unknown", "institution": "University of Michigan", "country": "", "position": "Professor" }, { "name": "Bhavya Kailkhura", "gender": "Male", "institution": "Lawrence Livermore National Laboratory", "country": "US", "position": "Research Staff" }, { "name": "Elisa Tsai", "gender": "Not Specified", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Haizhong Zheng", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Jiachen Sun", "gender": "Male", "institution": "University of Michigan", "country": "", "position": "PhD student" }, { "name": "Yifu Lu", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "Undergrad student" }, { "name": "Brian R Bartoldson", "gender": "Male", "institution": "Lawrence Livermore National Labs", "country": "US", "position": "Postdoc" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 39, "n_ref": 121, "n_ref_all": 159, "n_fig": 8, "n_tab": 8, "L_tab": 2129, "n_element_tab": 117, "n_fig_1": 2, "n_tab_1": 9, "L_tab_1": 2770, "n_element_tab_1": 157, "formula_len_all": 225, "formula_len_all_1": 225, "len_all": 146988, "len_all_1": 63071, "len_abs": 1359, "len_title": 105, "len_sents": 46093, "len_sents_1": 27455, "n_sents": 358, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1366, "title": "Label-Free Coreset Selection with Proxy Training Dynamics", "abs": "High-quality human-annotated data is crucial for modern deep learning pipelines, yet the human annotation process is both costly and time-consuming. Given a constrained human labeling budget, selecting an informative and representative data subset for labeling can significantly reduce human annotation effort. Well-performing state-of-the-art (SOTA) coreset selection methods require ground truth labels over the whole dataset, failing to reduce the human labeling burden. Meanwhile, SOTA label-free coreset selection methods deliver inferior performance due to poor geometry-based difficulty scores. In this paper, we introduce ELFS (Effective Label-Free Coreset Selection), a novel label-free coreset selection method. ELFS significantly improves label-free coreset selection by addressing two challenges: 1) ELFS utilizes deep clustering to estimate training dynamics-based data difficulty scores without ground truth labels; 2) Pseudo-labels introduce a distribution shift in the data difficulty scores, and we propose a simple but effective double-end pruning method to mitigate bias on calculated scores. We evaluate ELFS on four vision benchmarks and show that, given the same vision encoder, ELFS consistently outperforms SOTA label-free baselines. For instance, when using SwAV as the encoder, ELFS outperforms D2 by up to 10.2% in accuracy on ImageNet-1K.", "keywords": [ "Coreset Selection", "Data pruning", "Label free coreset selection" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "ykD8a9gJvy", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Aleksander Holynski", "gender": "unknown", "institution": "UC Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Boyang Zhou", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "Undergrad student" }, { "name": "Brian Curless", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Ira Kemelmacher-Shlizerman", "gender": "unknown", "institution": "University of Washington", "country": "US", "position": "Full Professor" }, { "name": "Steve Seitz", "gender": "Male", "institution": "University of Washington", "country": "", "position": "Full Professor" }, { "name": "Xiaojuan Wang", "gender": "Female", "institution": "Department of Computer Science", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 35, "n_ref": 62, "n_ref_all": 85, "n_fig": 8, "n_tab": 1, "L_tab": 249, "n_element_tab": 18, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 285, "n_element_tab_1": 22, "formula_len_all": 6897, "formula_len_all_1": 6115, "len_all": 124818, "len_all_1": 73169, "len_abs": 879, "len_title": 141, "len_sents": 34364, "len_sents_1": 31336, "n_sents": 236, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 882, "title": "Generative Inbetweening: Adapting Image-to-Video Models for Keyframe Interpolation", "abs": "We present a method for generating video sequences with coherent motion between a pair of input keyframes. We adapt a pretrained large-scale image-to-video diffusion model (originally trained to generate videos moving forward in time from a single input image) for keyframe interpolation, i.e., to produce a video between two input frames. We accomplish this adaptation through a lightweight fine-tuning technique that produces a version of the model that instead predicts videos moving backwards in time from a single input image. This model (along with the original forward-moving model) is subsequently used in a dual-directional diffusion sampling process that combines the overlapping model estimates starting from each of the two keyframes. Our experiments shows that our method outperforms both existing diffusion-based methods and traditional frame interpolation techniques.", "keywords": [ "generative keyframe interpolation", "image-to-video diffusion models" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "yj9lLwMjnE", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Alexander H. Liu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Bryan Catanzaro", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Vice President" }, { "name": "Chao-Han Huck Yang", "gender": "Male", "institution": "NVIDIA Research", "country": "TW", "position": "Researcher" }, { "name": "James R. Glass", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Principal Researcher" }, { "name": "Rafael Valle", "gender": "Not Specified", "institution": "NVIDIA", "country": "", "position": "Senior Research Scientist" }, { "name": "Sang-gil Lee", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yu-Chiang Frank Wang", "gender": "Male", "institution": "NVIDIA", "country": "TW", "position": "Research Director" }, { "name": "Yuan Gong", "gender": "Male", "institution": "xAI Corp", "country": "US", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 62, "n_ref": 162, "n_ref_all": 187, "n_fig": 2, "n_tab": 7, "L_tab": 1818, "n_element_tab": 256, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 2408, "n_element_tab_1": 229, "formula_len_all": 889, "formula_len_all_1": 898, "len_all": 167324, "len_all_1": 69909, "len_abs": 1145, "len_title": 135, "len_sents": 41011, "len_sents_1": 30872, "n_sents": 288, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1103, "title": "UniWav: Towards Unified Pre-training for Speech Representation Learning and Generation", "abs": "Pre-training and representation learning have been playing an increasingly important role in modern speech processing. Nevertheless, different applications have been relying on different foundation models, since predominant pre-training techniques are either designed for discriminative tasks or generative tasks. In this work, we make the first attempt at building a unified pre-training framework for both types of tasks in speech. We show that with the appropriate design choices for pre-training, one can jointly learn a representation encoder and generative audio decoder that can be applied to both types of tasks. We propose UniWav, an encoder-decoder framework designed to unify pre-training representation learning and generative tasks. On speech recognition, text-to-speech, and speech tokenization, UniWav achieves comparable performance to different existing foundation models, each trained on a specific task. Our findings suggest that a single general-purpose foundation model for speech can be built to replace different foundation models, reducing the overhead and cost of pre-training.", "keywords": [ "speech foundation model", "generative pre-training", "self-supervised learning", "speech generation", "speech tokenization" ], "rating_list": [ 6, 6, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 3, 4, 3 ] }, { "paper_id": "yj6P8OdWyj", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Bingsheng He", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Qinbin Li", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xinqi qian", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Yiqun Diao", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 27, "n_ref_uni": 56, "n_ref": 101, "n_ref_all": 113, "n_fig": 4, "n_tab": 8, "L_tab": 6057, "n_element_tab": 247, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 1098, "n_element_tab_1": 62, "formula_len_all": 4903, "formula_len_all_1": 2057, "len_all": 175821, "len_all_1": 64665, "len_abs": 1186, "len_title": 123, "len_sents": 44139, "len_sents_1": 28280, "n_sents": 360, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1190, "title": "Open-Set Learning for Addressing Label Skews in One-Shot Federated Learning", "abs": "Federated learning (FL) is crucial for collaborative model training, yet it faces significant challenges from data heterogeneity, particularly label skews across clients, where some classes may be underrepresented or absent entirely. In one-shot FL, where clients only communicate with the server once, this problem becomes even more challenging. Recent solutions propose incorporating open-set learning (OSL) to tackle this issue by detecting unknown samples during inference, but current methods like FedOV lack adaptability to varying client data distributions. In this paper, we provide a theoretical analysis proving that improving OSL algorithms can effectively address label skews in one-shot FL, since one-shot FL is learnable through good OSL algorithms regardless of label skews. We also empirically evaluate state-of-the-art OSL algorithms and identify their limitations. Based on these insights, we propose FedAdav, an adaptive algorithm that combines OSL signals to significantly improve ensemble accuracy in one-shot FL under label skews. Through extensive experiments, we demonstrate that exploring better OSL is key to overcoming label skew challenges in federated learning.", "keywords": [ "federated learning", "open-set learning" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "yizEOJVFFd", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Mingyuan Zhou", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Weizhu Chen", "gender": "Male", "institution": "Microsoft GenAI", "country": "US", "position": "Vice President" }, { "name": "Yueqin Yin", "gender": "Female", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Yujia Xie", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Zhendong Wang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 47, "n_ref": 104, "n_ref_all": 117, "n_fig": 2, "n_tab": 7, "L_tab": 2802, "n_element_tab": 351, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2944, "n_element_tab_1": 257, "formula_len_all": 505, "formula_len_all_1": 505, "len_all": 143018, "len_all_1": 73642, "len_abs": 1174, "len_title": 138, "len_sents": 39382, "len_sents_1": 33509, "n_sents": 272, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1165, "title": "Self-Augmented Preference Optimization: Off-Policy Paradigms for Language Model Alignment", "abs": "Traditional language model alignment methods, such as Direct Preference Optimization (DPO), are limited by their dependence on static, pre-collected paired preference data, which restricts their adaptability and practical applicability. To address this limitation, we introduce Self-Augmented Preference Optimization (SAPO), an effective and scalable training paradigm without the need of existing paired data. Built upon the self-play concept that autonomously generate negative responses, we further involve the off-policy learning pipeline to improve the data exploration and exploitation. Specifically, we employ an Exponential Moving Average (EMA) model along with a replay buffer to enable dynamic updates of response segments, effectively integrating real-time feedback with historical data insights. Our comprehensive evaluations of the LLaMA3-8B and Mistral-7B models across benchmarks\u2014including the Open LLM Leaderboard, IFEval, AlpacaEval 2.0, and MT-Bench\u2014demonstrate that SAPO matches or surpasses established offline contrastive baselines, such as DPO and Odds Ratio Preference Optimization (ORPO), and outperforms offline self-play methods like SPIN.", "keywords": [ "Large Language Model", "Fine-tuning", "Self-play" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "yitH9xAHQs", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Chuan Wu", "gender": "unknown", "institution": "The University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Lingpeng Kong", "gender": "Male", "institution": "Department of Computer Science, The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Qintong Li", "gender": "Female", "institution": "The University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Renjie Pi", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Sheng Wang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xin Jiang", "gender": "Male", "institution": "Noah\u2019s Ark Lab, Huawei Technologies", "country": "HK", "position": "Principal Researcher" }, { "name": "Xueliang Zhao", "gender": "Male", "institution": "The University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Zhenguo Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "JIAHUI GAO", "gender": "Female", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 39, "n_ref": 98, "n_ref_all": 118, "n_fig": 2, "n_tab": 3, "L_tab": 1665, "n_element_tab": 174, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1853, "n_element_tab_1": 206, "formula_len_all": 9595, "formula_len_all_1": 11259, "len_all": 232358, "len_all_1": 85228, "len_abs": 1451, "len_title": 144, "len_sents": 53770, "len_sents_1": 33898, "n_sents": 396, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1458, "title": "Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration", "abs": "Large language models (LLMs) have significantly benefited from training on diverse, high-quality task-specific data, leading to impressive performance across a range of downstream applications. Current methods often rely on human-annotated data or predefined task templates to direct powerful LLMs in synthesizing task-relevant data for effective model training. However, this dependence on manually designed components may constrain the scope of generated data, potentially overlooking critical edge cases or novel scenarios that could challenge the model. In this paper, we present a novel approach, ReverseGen, designed to automatically generate effective training samples that expose the weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained to produce queries that lead target models to generate unsatisfactory responses. These failure-inducing queries are then used to construct training data, helping to address the models' shortcomings and improve overall performance. Our approach is flexible and can be applied to models of various scales (3B, 7B, and 8B). We evaluate ReverseGen on three key applications\u2014safety, honesty, and math\u2014demonstrating that our generated data is both highly effective and diverse. Models fine-tuned with ReverseGen-generated data consistently outperform those trained on human-annotated or general model-generated data, offering a new perspective on data synthesis for task-specific LLM enhancement.", "keywords": [ "data synthesis", "preference learning", "LLM alignment" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "yiQCeXdPvs", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Jifan Zhang", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "PhD student" }, { "name": "Lalit K Jain", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Robert D Nowak", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "", "position": "Full Professor" }, { "name": "Shyam Nuggehalli", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Undergrad student" } ], "n_formula": 15, "n_formula_1": 7, "n_ref_uni": 47, "n_ref": 102, "n_ref_all": 120, "n_fig": 8, "n_tab": 4, "L_tab": 213, "n_element_tab": 38, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 625, "n_element_tab_1": 8, "formula_len_all": 1452, "formula_len_all_1": 591, "len_all": 137934, "len_all_1": 62538, "len_abs": 1251, "len_title": 109, "len_sents": 40378, "len_sents_1": 27833, "n_sents": 338, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1258, "title": "DIRECT: Deep Active Learning under Imbalance and Label Noise", "abs": "Class imbalance is a prevalent issue in real world machine learning applications, often leading to poor performance in rare and minority classes. With an abundance of wild unlabeled data, active learning is perhaps the most effective technique in solving the problem at its root -- collecting a more balanced and informative set of labeled examples during annotation. Label noise is another common issue in data annotation jobs, which is especially challenging for active learning methods. In this work, we conduct the first study of active learning under both class imbalance and label noise. We propose a novel algorithm that robustly identifies the class separation threshold and annotates the most uncertain examples that are closest from it. Through a novel reduction to one-dimensional active learning, our algorithm DIRECT is able to leverage classic active learning theory and methods to address issues such as batch labeling and tolerance towards label noise. We present extensive experiments on imbalanced datasets with and without label noise. Our results demonstrate that DIRECT can save more than 60% of the annotation budget compared to state-of-art active learning algorithms and more than 80% of annotation budget compared to random sampling.", "keywords": [ "Deep Learning", "Active Learning" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "yiGSI7Ou3i", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Chao Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Lingzhi Gao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zexi Li", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 42, "n_ref": 123, "n_ref_all": 142, "n_fig": 5, "n_tab": 7, "L_tab": 1466, "n_element_tab": 204, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 2886, "n_element_tab_1": 201, "formula_len_all": 365, "formula_len_all_1": 61, "len_all": 165615, "len_all_1": 68777, "len_abs": 1380, "len_title": 145, "len_sents": 46555, "len_sents_1": 29587, "n_sents": 380, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1400, "title": "Text-to-Model: Text-Conditioned Neural Network Diffusion for Train-Once-for-All Personalization", "abs": "Generative artificial intelligence (GenAI) has made significant progress in understanding world knowledge and generating content from human languages across various modalities, like text-to-text large language models, text-to-image stable diffusion, and text-to-video Sora. While in this paper, we investigate the capability of GenAI for text-to-model generation, to see whether GenAI can comprehend hyper-level knowledge embedded within AI itself parameters. Specifically, we study a practical scenario termed train-once-for-all personalization, aiming to generate personalized models for diverse end-users and tasks using text prompts. Inspired by the recent emergence of neural network diffusion, we present Tina, a text-conditioned neural network diffusion for train-once-for-all personalization. Tina leverages a diffusion transformer model conditioned on task descriptions embedded using a CLIP model. Despite the astronomical number of potential personalized tasks (e.g., $1.73\\times10^{13}$), by our design, Tina demonstrates remarkable in-distribution and out-of-distribution generalization even trained on small datasets ($\\sim 1000$). We further verify whether and how \\Tina understands world knowledge by analyzing its capabilities under zero-shot/few-shot image prompts, different numbers of personalized classes, prompts of natural language descriptions, and predicting unseen entities.", "keywords": [ "diffusion model", "parameter generation", "personalization" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "yi3QcCGfP1", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Bo-Han Kung", "gender": "Male", "institution": "Department of computer science and informational engineering, National Taiwan University", "country": "TW", "position": "PhD student" }, { "name": "Bo-Han Lai", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "MS student" }, { "name": "Pin-Han Huang", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Shang-Tse Chen", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Associate Professor" } ], "n_formula": 48, "n_formula_1": 12, "n_ref_uni": 42, "n_ref": 90, "n_ref_all": 128, "n_fig": 12, "n_tab": 12, "L_tab": 5140, "n_element_tab": 497, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 800, "n_element_tab_1": 62, "formula_len_all": 3336, "formula_len_all_1": 1048, "len_all": 219106, "len_all_1": 62431, "len_abs": 1132, "len_title": 68, "len_sents": 65239, "len_sents_1": 27203, "n_sents": 568, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 775, "title": "Enhancing Certified Robustness via Block Reflector Orthogonal Layers", "abs": "Lipschitz neural networks are well-known for providing certified robustness in deep learning. In this paper, we present a novel efficient Block Reflector Orthogonal layer that enables the construction of simple yet effective Lipschitz neural networks. \nIn addition, by theoretically analyzing the nature of Lipschitz neural networks, we introduce a new loss function that employs an annealing mechanism to improve margin for most data points.\nThis enables Lipschitz models to provide better certified robustness.\nBy employing our BRO layer and loss function, we design BRONet, which provides state-of-the-art certified robustness.\t\nExtensive experiments and empirical analysis on CIFAR-10, CIFAR-100, and Tiny-ImageNet validate that our method outperforms existing baselines.", "keywords": [ "Certified robustness", "Adversarial" ], "rating_list": [ 6, 3, 6, 6, 3 ], "soundness_list": [ 2, 1, 3, 3, 2 ], "presentation_list": [ 3, 1, 3, 3, 2 ], "contribution_list": [ 2, 1, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 2, 4 ] }, { "paper_id": "yhmVrA8W0v", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Bo Jiang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Cheng Chen", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Shuai Li", "gender": "Female", "institution": "John Hopcroft Center, Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yubo Yang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "He Baoxiang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" } ], "n_formula": 115, "n_formula_1": 34, "n_ref_uni": 23, "n_ref": 83, "n_ref_all": 113, "n_fig": 2, "n_tab": 7, "L_tab": 2272, "n_element_tab": 188, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 302, "n_element_tab_1": 40, "formula_len_all": 15628, "formula_len_all_1": 3576, "len_all": 157640, "len_all_1": 65757, "len_abs": 1261, "len_title": 117, "len_sents": 43009, "len_sents_1": 27161, "n_sents": 425, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1331, "title": "The Convergence of Second-Order Sampling Methods for Diffusion Models", "abs": "Diffusion models have achieved great success in generating samples from complex distributions, notably in the domains of images and videos. Beyond the experimental success, theoretical insights into their performance have been illuminated, particularly concerning the convergence of diffusion models when applied with discretization methods such as Euler-Maruyama (EM) and Exponential Integrator (EI). This paper embarks on analyzing the convergence of the higher-order discretization method (SDE-DPM-2) under $L^2$-accurate score estimate. Our findings reveal that to attain $\\tilde{O}(\\epsilon_0^2)$ Kullback-Leibler (KL) divergence between the target and the sampled distributions, the sampling complexity - or the required number of discretization steps - for SDE-DPM-2 is $\\tilde{O}(1/\\epsilon_0)$, which is better than the currently known sample complexity of EI given by $\\tilde{O}(1/\\epsilon_0^2)$. We further extend our analysis to the Runge-Kutta-2 (RK-2) method, which demands a sampling complexity of $\\tilde{O}(1/\\epsilon_0^2)$, indicating that SDE-DPM-2 is more efficient than RK-2. Our study also demonstrates that the convergence of SDE-DPM-2 under Variance Exploding (VE) SDEs aligns with that of Variance Preserving (VP) SDEs, highlighting the adaptability of SDE-DPM-2 across various diffusion models frameworks.", "keywords": [ "diffusion models", "reserve SDE" ], "rating_list": [ 5, 3, 3, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 4 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 1, 3, 3 ], "confidence_list": [ 4, 4, 5, 4, 3 ] }, { "paper_id": "yheQRc5xWB", "primary_area": "causal reasoning", "n_author": 7, "author_info_list": [ { "name": "Hao Zou", "gender": "Male", "institution": "ZGC laboratory", "country": "CN", "position": "Researcher" }, { "name": "Haoang Chi", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Haotian Wang", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Haoxuan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Long Lan", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Wanrong Huang", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Assistant Research Fellow" }, { "name": "Wenjing Yang", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 61, "n_formula_1": 24, "n_ref_uni": 43, "n_ref": 125, "n_ref_all": 157, "n_fig": 9, "n_tab": 10, "L_tab": 3979, "n_element_tab": 323, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 575, "n_element_tab_1": 21, "formula_len_all": 4989, "formula_len_all_1": 1447, "len_all": 183866, "len_all_1": 58357, "len_abs": 4146, "len_title": 135, "len_sents": 54758, "len_sents_1": 25927, "n_sents": 493, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1664, "title": "Effective and Efficient Time-Varying Counterfactual Prediction with State-Space Models", "abs": "Time-varying counterfactual prediction (TCP) from observational data supports the answer of when and how to assign multiple sequential treatments, yielding importance in various applications. Despite the progress achieved by recent advances, e.g., LSTM or Transformer based causal approaches, their capability of capturing interactions in long sequences remains to be improved in both prediction performance and running efficiency. In parallel with the development of TCP, the success of the state-space models (SSMs) has achieved remarkable progress toward long-sequence modeling with saved running time. Consequently, studying how Mamba simultaneously benefits the effectiveness and efficiency of TCP becomes a compelling research direction. In this paper, we propose to exploit advantages of the SSMs to tackle the TCP task, by introducing a counterfactual Mamba model with Covariate-based Decorrelation towards Selective Parameters (Mamba-CDSP). Motivated by the over-balancing problem in TCP of the direct covariate balancing methods, we propose to de-correlate between the current treatment and the representation of historical covariates, treatments, and outcomes, which can mitigate the confounding bias while preserve more covariate information. In addition, we show that the overall de-correlation in TCP is equivalent to regularizing the selective parameters of Mamba over each time step, which leads our approach to be effective and lightweight. We conducted extensive experiments on both synthetic and real-world datasets, demonstrating that Mamba-CDSP not only outperforms baselines by a large margin, but also exhibits prominent running efficiency.", "keywords": [ "Time Series; State-space Models; Treatment Effect Estimation" ], "rating_list": [ 6, 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 2, 3 ] }, { "paper_id": "yhKNCvYlCr", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Feng Zhao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jiahao Chang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Kun Li", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Lin Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Tianyou Song", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Yukun Qi", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Zehui Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 40, "n_ref": 84, "n_ref_all": 92, "n_fig": 5, "n_tab": 4, "L_tab": 3422, "n_element_tab": 200, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2150, "n_element_tab_1": 130, "formula_len_all": 461, "formula_len_all_1": 461, "len_all": 119393, "len_all_1": 61739, "len_abs": 1639, "len_title": 147, "len_sents": 30441, "len_sents_1": 26174, "n_sents": 209, "n_sents_1": 166, "venue": "ICLR2025", "decision": "withdraw", "L_title": 99, "L_abs": 1647, "title": "Transfering Knowledge into Efficient Tiny Models for Object Detection with Dual Prompt Distillation", "abs": "Knowledge Distillation (KD) has demonstrated significant benefits for learning compact models for object detection. Most current work focuses on general distillation settings, where student models are relatively large and learnable, then compete with the distillation performance. However, due to the model scale and inference speed, these models are seldom deployed in real-world applications. In this paper, we dive into a challenging but more applicable setting: how to distill rich teacher knowledge into tiny, faster models for object detection? We first show that simply applying previous KD strategies under such settings cannot achieve satisfying results, due to the extremely large model capacity gap between the teacher-student pairs. To this end, we propose a simple prompt-based object detection distillation framework, namely DualPromptKD, which aims to improve knowledge transfer efficiency from both teacher and student perspectives. Specifically, by distilling teacher representations into compact external prompts, we enable the student model to fully leverage proficient teacher knowledge even at inference time. In terms of the limited learning ability of the student model, we introduce lightweight internal prompts tailored to bolster the feature imitation capability for the target model. Extensive experimental results on the COCO benchmarks validate the effectiveness and generalization of our approach, including different image backbones and detector types. Notably, our DualPromptKD surpasses the previous best distillation strategies by more than 2.0 mAP under various experimental settings. The code will be available.", "keywords": [ "knowledge distillation", "object detection" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "ygtmPu0xZy", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Baoxiang Wang", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Jiawei Xu", "gender": "Male", "institution": "CUHK(SZ)", "country": "", "position": "PhD student" }, { "name": "Yingru Li", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen, China", "country": "CN", "position": "PhD student" }, { "name": "Zhi-Quan Luo", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Full Professor" } ], "n_formula": 148, "n_formula_1": 11, "n_ref_uni": 31, "n_ref": 95, "n_ref_all": 138, "n_fig": 12, "n_tab": 12, "L_tab": 3427, "n_element_tab": 115, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 847, "n_element_tab_1": 45, "formula_len_all": 14126, "formula_len_all_1": 1034, "len_all": 301513, "len_all_1": 69030, "len_abs": 1771, "len_title": 83, "len_sents": 102436, "len_sents_1": 31203, "n_sents": 899, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1795, "title": "Scalable Exploration via Ensemble++", "abs": "Scalable exploration in high-dimensional, complex environments is a significant challenge in sequential decision making, especially when utilizing neural networks. Ensemble sampling, a practical approximation of Thompson sampling, is widely adopted but often suffers performance degradation due to ensemble coupling in shared layer architectures, leading to reduced diversity and ineffective exploration. In this paper, we introduce Ensemble++, a novel method that addresses these challenges through architectural and algorithmic innovations. To prevent ensemble coupling, Ensemble++ decouples mean and uncertainty estimation by separating the base network and ensemble components, employs a symmetrized loss function and the stop-gradient operator. To further enhance exploration, it generates richer hypothesis spaces through random linear combinations of ensemble components using continuous index sampling. Theoretically, we prove that Ensemble++ matches the regret bounds of exact Thompson sampling in linear contextual bandits while maintaining a scalable per-step computational complexity of $\\tilde{O}( \\log T)$. This provides the first rigorous analysis demonstrating that ensemble sampling can be an scalable and effective approximation to Thompson Sampling, closing a key theoretical gap in exploration efficiency. Empirically, we demonstrate Ensemble++'s effectiveness in both regret minimization and computational efficiency across a range of nonlinear bandit environments, including a language-based contextual bandits where the agents employ GPT backbones. Our results highlight the capability of Ensemble++ for real-time adaptation in complex environments where computational and data collection budgets are constrained. \\url{https://anonymous.4open.science/r/EnsemblePlus2-1E54}", "keywords": [ "Bandit", "Scalable Exploration", "Function Approximation" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "yfkvUJEY6i", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Ashley Beecy", "gender": "Female", "institution": "Weill Cornell Medicine, Cornell University", "country": "US", "position": "Assistant Professor" }, { "name": "Deborah Estrin", "gender": "Female", "institution": "Cornell University", "country": "US", "position": "Full Professor" }, { "name": "Divya M Shanmugam", "gender": "Female", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "Postdoc" }, { "name": "Emma Pierson", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "Assistant Professor" }, { "name": "Erica Chiang", "gender": "unknown", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "PhD student" }, { "name": "Gabriel Sayer", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Associate Professor" }, { "name": "Nikhil Garg", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 54, "n_formula_1": 15, "n_ref_uni": 60, "n_ref": 92, "n_ref_all": 117, "n_fig": 5, "n_tab": 5, "L_tab": 1113, "n_element_tab": 50, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 801, "n_element_tab_1": 48, "formula_len_all": 5559, "formula_len_all_1": 980, "len_all": 206406, "len_all_1": 79502, "len_abs": 1225, "len_title": 67, "len_sents": 62168, "len_sents_1": 36471, "n_sents": 438, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1025, "title": "Learning Disease Progression Models That Capture Health Disparities", "abs": "Disease progression models are widely used to inform the diagnosis and treatment of many progressive diseases. However, a significant limitation of existing models is that they do not account for health disparities that can bias the observed data. To address this, we develop an interpretable Bayesian disease progression model that captures three key health disparities: certain patient populations may (1) start receiving care only when their disease is more severe, (2) experience faster disease progression even while receiving care, or (3) receive follow-up care less frequently conditional on disease severity. We show theoretically and empirically that failing to account for disparities produces biased estimates of severity (underestimating severity for disadvantaged groups, for example). On a dataset of heart failure patients, we show that our model can identify groups that face each type of health disparity, and that accounting for these disparities meaningfully shifts which patients are considered high-risk.", "keywords": [ "fairness", "equity", "bias", "health disparities", "disease progression", "bayesian model" ], "rating_list": [ 8, 3, 3, 3 ], "soundness_list": [ 3, 1, 3, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "yfZJdCijo6", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Alessandro Epasto", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Alina Ene", "gender": "Female", "institution": "Boston University", "country": "US", "position": "Associate Professor" }, { "name": "David Woodruff", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Hoai-An Nguyen", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Huy Nguyen", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Associate Professor" }, { "name": "Peilin Zhong", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Vahab Mirrokni", "gender": "Male", "institution": "Google Research", "country": "", "position": "Distinguished Scientist" } ], "n_formula": 25, "n_formula_1": 2, "n_ref_uni": 29, "n_ref": 57, "n_ref_all": 80, "n_fig": 3, "n_tab": 2, "L_tab": 157, "n_element_tab": 7, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2059, "formula_len_all_1": 115, "len_all": 203835, "len_all_1": 79159, "len_abs": 2016, "len_title": 130, "len_sents": 73072, "len_sents_1": 36587, "n_sents": 761, "n_sents_1": 321, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 2138, "title": "Maximum Coverage in Turnstile Streams with Applications to Fingerprinting Measures", "abs": "In the maximum coverage problem we are given $d$ subsets from a universe $[n]$, and the goal is to output at most $k$ subsets such that their union covers the largest possible number of distinct items. The input can be formalized as an $n \\times d$ matrix $A$ where entry $A_{ij} \\neq 0$ if item $i$ is covered by subset $j$ and $A_{ij} = 0$ otherwise. In this paper we create the first linear sketch to solve the maximum coverage problem. The sketch has size sublinear in the input and is directly applicable to distributed and streaming settings, often offering significant runtime improvements. We focus on the application to the turnstile streaming model which supports insertions and deletions. In this model, updates take the form $(i,j, \\pm 1)$ which update $A_{ij}$ to $A_{ij} + 1$ or $A_{ij} - 1$, depending on the sign. Previous work has largely focused on more restrictive models, such as the set-arrival model where each update reveals an entire column of $A$, or the insertion-only model which does not allow deletions. We design an algorithm with an $\\tilde{O}(d/\\varepsilon^3)$ space bound for all $k \\geq 0$. We note that when $k$ is constant, this space bound is nearly optimal up to logarithmic factors.\n\nWe then turn to fingerprinting for risk measurement. The input is an $n \\times d$ matrix $A$ where there are $n$ users and $d$ features, and the goal is to determine which $k$ features (or columns in $A$) together pose the greatest re-identification risk. Our maximum coverage sketch directly enables a solution to targeted fingerprinting for risk measurement. Furthermore, we present a result of independent interest: a linear sketch of the complement of $F_p$, the $p^{\\text{th}}$ frequency moment, for $p \\geq 2$. We use this sketch to solve general fingerprinting for risk management. Empirical evaluation confirms the practicality of our fingerprinting algorithms, demonstrating a speedup of up to $210$x over prior work. We also demonstrate that our general fingerprinting algorithm can serve as a dimensionality reduction technique, with an application to facilitating enhanced feature selection efficiency.", "keywords": [ "maximum coverage", "turnstile streams", "sketching" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 3, 2, 1 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 2, 3, 4 ] }, { "paper_id": "yfW1x7uBS5", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Javier Rando", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Nicholas Carlini", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Robert H\u00f6nig", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "MS student" }, { "name": "Florian Tramer", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 0, "n_ref_uni": 45, "n_ref": 121, "n_ref_all": 173, "n_fig": 24, "n_tab": 3, "L_tab": 2438, "n_element_tab": 750, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 710, "formula_len_all_1": 0, "len_all": 212394, "len_all_1": 63284, "len_abs": 942, "len_title": 124, "len_sents": 70371, "len_sents_1": 31581, "n_sents": 600, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 76, "L_abs": 954, "title": "Adversarial Perturbations Cannot Reliably Protect Artists From Generative AI", "abs": "Artists are increasingly concerned about advancements in image generation models that can closely replicate their unique artistic styles.\nIn response, several protection tools against style mimicry have been developed that incorporate small adversarial perturbations into artworks published online. In this work, we evaluate the effectiveness of popular protections---with millions of downloads---and show they only provide a false sense of security. We find that low-effort and \"off-the-shelf\" techniques, such as image upscaling, are sufficient to create robust mimicry methods that significantly degrade existing protections. Through a user study, we demonstrate that **all existing protections can be easily bypassed**, leaving artists vulnerable to style mimicry. We caution that tools based on adversarial perturbations cannot reliably protect artists from the misuse of generative AI, and urge the development of alternative protective solutions.", "keywords": [ "security", "adversarial", "style mimicry", "generative ai" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "yf30Al57nu", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Leitian Tao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Saayan Mitra", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Principal Researcher" }, { "name": "Tong Yu", "gender": "unknown", "institution": "Adobe Research", "country": "US", "position": "Senior Research Scientist" }, { "name": "Tung Mai", "gender": "Male", "institution": "Adobe", "country": "", "position": "Research Scientist" }, { "name": "Xiang Chen", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Yixuan Li", "gender": "Female", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Ryan Rossi", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Senior Research Scientist" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 51, "n_ref": 102, "n_ref_all": 118, "n_fig": 7, "n_tab": 4, "L_tab": 2211, "n_element_tab": 206, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1439, "n_element_tab_1": 200, "formula_len_all": 643, "formula_len_all_1": 602, "len_all": 164405, "len_all_1": 64405, "len_abs": 1686, "len_title": 121, "len_sents": 38487, "len_sents_1": 29825, "n_sents": 273, "n_sents_1": 223, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1384, "title": "CodeLutra: Boosting LLM Code Generation via Preference-Guided Refinement", "abs": "Large Language Models (LLMs) have significantly advanced code generation but often require substantial resources and tend to over-generalize, limiting their efficiency for specific tasks. Fine-tuning smaller, open-source LLMs presents a viable alternative; however, it typically lags behind cutting-edge models due to supervised fine-tuning's reliance solely on correct code examples, which restricts the model's ability to learn from its own mistakes and adapt to diverse programming challenges. To bridge this gap, we introduce CodeLutra, a novel framework that enhances low-performing LLMs by leveraging both successful and failed code generation attempts. Unlike conventional fine-tuning, CodeLutra employs an iterative preference learning mechanism to compare correct and incorrect solutions as well as maximize the likelihood of correct codes. Through continuous iterative refinement, CodeLutra enables smaller LLMs to match or surpass GPT-4\u2019s performance in various code generation tasks without relying on vast external datasets or larger auxiliary models. On a challenging data analysis task, using just 500 samples improved Llama-3-8B's accuracy from 28.2\\% to 48.6\\%, approaching GPT-4's performance. These results highlight CodeLutra's potential to close the gap between open-source and closed-source models, making it a promising approach in the field of code generation.", "keywords": [ "large language models; preference learning; code generation" ], "rating_list": [ 6, 3, 3, 5, 8 ], "soundness_list": [ 3, 2, 2, 4, 3 ], "presentation_list": [ 3, 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 1, 3, 3 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "yeeIGM3N6w", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Chun-Yi Lee", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Full Professor" }, { "name": "Hsu-Shen Liu", "gender": "unknown", "institution": "National Tsinghua University", "country": "TW", "position": "MS student" }, { "name": "I-Chun Chen", "gender": "unknown", "institution": "CyberLink Taiwan", "country": "TW", "position": "Enginner" }, { "name": "Wei-Fang Sun", "gender": "Male", "institution": "NVIDIA", "country": "TW", "position": "Researcher" }, { "name": "Yen-Chang Hsu", "gender": "Male", "institution": "Samsung Research America", "country": "", "position": "Research Scientist" }, { "name": "CHEN-HAO CHAO", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 27, "n_ref": 59, "n_ref_all": 83, "n_fig": 5, "n_tab": 10, "L_tab": 6242, "n_element_tab": 637, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3554, "n_element_tab_1": 419, "formula_len_all": 298, "formula_len_all_1": 298, "len_all": 129059, "len_all_1": 75348, "len_abs": 1408, "len_title": 128, "len_sents": 41341, "len_sents_1": 32245, "n_sents": 328, "n_sents_1": 251, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1416, "title": "Retraining-Free Merging of Sparse Mixture-of-Experts via Hierarchical Clustering", "abs": "Sparse Mixture-of-Experts (SMoE) models represent a significant breakthrough in large language model development. These models enable performance improvements without a proportional increase in inference costs. By selectively activating a small set of parameters during task execution, SMoEs enhance model capacity. However, their deployment remains challenging due to the substantial memory footprint required to accommodate the growing number of experts. This constraint renders them less feasible in environments with limited hardware resources. To address this challenge, we propose Hierarchical Clustering for Sparsely activated Mixture of Experts (HC-SMoE), a task-agnostic expert merging framework that reduces SMoE model parameters without retraining. Unlike previous methods, HC-SMoE employs hierarchical clustering based on expert outputs. This approach ensures that the merging process remains unaffected by routing decisions. The output-based clustering strategy captures functional similarities between experts, offering an adaptable solution for models with numerous experts. We validate our approach through extensive experiments on eight zero-shot language tasks and demonstrate its effectiveness in large-scale SMoE models such as Qwen and Mixtral. Our comprehensive results demonstrate that HC-SMoE consistently achieves strong performance, which highlights its potential for real-world deployment.", "keywords": [ "Sparse Mixture-of-Experts", "Merging", "Compression" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "yeEWZ8qvlS", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Alexandros Doumanoglou", "gender": "Male", "institution": "Maastricht University", "country": "NL", "position": "PhD student" }, { "name": "Dimitrios Zarpalas", "gender": "Male", "institution": "Centre for Research and Technology Hellas", "country": "GR", "position": "Principal Researcher" }, { "name": "Kurt Driessens", "gender": "Male", "institution": "Maastricht University", "country": "NL", "position": "Associate Professor" } ], "n_formula": 30, "n_formula_1": 16, "n_ref_uni": 17, "n_ref": 82, "n_ref_all": 138, "n_fig": 8, "n_tab": 15, "L_tab": 3382, "n_element_tab": 242, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 304, "n_element_tab_1": 45, "formula_len_all": 1645, "formula_len_all_1": 1122, "len_all": 222681, "len_all_1": 71778, "len_abs": 1292, "len_title": 150, "len_sents": 88768, "len_sents_1": 33676, "n_sents": 567, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1296, "title": "Learning Interpretable and Influential Directions with Signal Vectors and Uncertainty Region Alignment", "abs": "Latent space directions have played a key role in understanding, debugging, and fixing deep learning models. Concepts are often encoded in distinct feature space directions, and evaluating impact of these directions on the model's predictions, highlights their importance in the decision-making process. Additionally, recent studies have shown that penalizing directions associated with spurious artifacts during training can force models to unlearn features irrelevant to their prediction task. Identifying these directions, therefore, provides numerous benefits, including a deeper understanding of the model's strategy, fostering trust, and enabling model correction and improvement. We introduce a novel unsupervised approach utilizing signal vectors and uncertainty region alignment to discover latent space directions that meet two key debugging criteria: significant influence on model predictions and high level of interpretability. To our knowledge, this method is the first of its kind to uncover such directions, leveraging the inherent structure of the feature space and the knowledge encoded in the deep network. We validate our approach using both synthetic and real-world benchmarks, demonstrating that the discovered directions effectively fulfill the critical debugging criteria.", "keywords": [ "latent space", "interpretability", "concepts", "directions", "signals", "patterns", "distractors" ], "rating_list": [ 6, 5, 5, 3, 6 ], "soundness_list": [ 4, 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 2, 3, 3, 4, 2 ] }, { "paper_id": "ye1mxb79lw", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Bryan Kian Hsiang Low", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Associate Professor" }, { "name": "Quoc Phong Nguyen", "gender": "Male", "institution": "Deakin University", "country": "US", "position": "Research Fellow" }, { "name": "Wan Theng Ruth Chew", "gender": "unknown", "institution": "National University of Singapore", "country": "", "position": "Researcher" } ], "n_formula": 41, "n_formula_1": 24, "n_ref_uni": 26, "n_ref": 37, "n_ref_all": 64, "n_fig": 15, "n_tab": 3, "L_tab": 1864, "n_element_tab": 189, "n_fig_1": 15, "n_tab_1": 2, "L_tab_1": 129, "n_element_tab_1": 27, "formula_len_all": 4771, "formula_len_all_1": 1965, "len_all": 122246, "len_all_1": 68366, "len_abs": 1080, "len_title": 84, "len_sents": 40809, "len_sents_1": 30918, "n_sents": 354, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 1087, "title": "BILBO: BILevel Bayesian Optimization", "abs": "Bilevel optimization, characterized by a two-level hierarchical optimization structure, is prevalent in real-world problems but poses significant challenges, especially in noisy, constrained, and derivative-free settings. To tackle these challenges, we present a novel algorithm for BILevel Bayesian Optimization (BILBO) that optimizes both upper- and lower-level problems jointly in a sample-efficient manner by using confidence bounds to construct trusted sets of feasible and lower-level optimal solutions. We show that sampling from our trusted sets guarantees points with instantaneous regret bounds. Moreover, BILBO selects only one function query per iteration, facilitating its use in decoupled settings where upper- and lower-level function evaluations may come from different simulators or experiments. We also show that this function query selection strategy leads to an instantaneous regret bound for the query point. The performance of BILBO is theoretically guaranteed with a sublinear regret bound and is empirically evaluated on several synthetic and real-world problems.", "keywords": [ "bilevel", "Bayesian optimization" ], "rating_list": [ 6, 6, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 4, 3 ] }, { "paper_id": "ydw2l8zgUB", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Jia-He Lim", "gender": "Male", "institution": "Department of Computer Science, National Tsing Hua University, National Tsinghua University", "country": "TW", "position": "MS student" }, { "name": "Po-Chih Kuo", "gender": "Male", "institution": "National Tsinghua University", "country": "", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 38, "n_ref": 74, "n_ref_all": 97, "n_fig": 7, "n_tab": 6, "L_tab": 2309, "n_element_tab": 213, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3213, "n_element_tab_1": 296, "formula_len_all": 580, "formula_len_all_1": 217, "len_all": 152921, "len_all_1": 59802, "len_abs": 1637, "len_title": 64, "len_sents": 51666, "len_sents_1": 26173, "n_sents": 374, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1500, "title": "EEGTrans: Transformer-Driven Generative Models for EEG Synthesis", "abs": "Recent advancements in Large Language Models (LLMs) have been significant, largely due to improvements in network architecture, particularly the transformer model. With access to large training datasets, LLMs can train in an unsupervised manner and still achieve impressive results in generating coherent output. This study introduces a transformer-based generative model, EEGTrans, designed for sequentially generating synthetic electroencephalogram (EEG) signals. Given the inherent noise in EEG data, we employ a quantized autoencoder that compresses these signals into discrete codes, effectively capturing their temporal features and enabling generalization across diverse datasets. The encoder of EEGTrans processes EEG signals as input, while its decoder autoregressively generates discrete codes. We evaluate our method in a motor imagery Brain-Computer Interface (BCI) application, where merging data across datasets is particularly challenging due to experimental differences. Our results demonstrate that the synthetic EEG data effectively captures temporal patterns while maintaining the complexity and power spectrum of the original signals. Moreover, classification results show that incorporating synthetic data improves performance and even surpasses that of models based on Generative Adversarial Networks. These findings highlight the potential of transformer-based generative models to generalize effectively across multiple datasets and produce high-quality synthetic EEG signals.", "keywords": [ "LLM", "EEG", "BCI", "transformer" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "ydlDRUuGm9", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Jonathan W. Siegel", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "", "position": "Assistant Professor" }, { "name": "Yixuan Wang", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Ziming Liu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Thomas Hou", "gender": "unknown", "institution": "California Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 39, "n_formula_1": 22, "n_ref_uni": 31, "n_ref": 50, "n_ref_all": 64, "n_fig": 7, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2556, "n_element_tab_1": 30, "formula_len_all": 1860, "formula_len_all_1": 1305, "len_all": 143354, "len_all_1": 59099, "len_abs": 1574, "len_title": 0, "len_sents": 33993, "len_sents_1": 25769, "n_sents": 262, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1584, "title": "On the expressiveness and spectral bias of KANs", "abs": "Kolmogorov-Arnold Networks (KAN) \\cite{liu2024kan} were very recently proposed as a potential alternative to the prevalent architectural backbone of many deep learning models, the multi-layer perceptron (MLP). KANs have seen success in various tasks of AI for science, with their empirical efficiency and accuracy demonstrated in function regression, PDE solving, and many more scientific problems.\n \nIn this article, we revisit the comparison of KANs and MLPs, with emphasis on a theoretical perspective. On the one hand, we compare the representation and approximation capabilities of KANs and MLPs. We establish that MLPs can be represented using KANs of a comparable size. This shows that the approximation and representation capabilities of KANs are at least as good as MLPs. Conversely, we show that KANs can be represented using MLPs, but that in this representation the number of parameters increases by a factor of the KAN grid size. This suggests that KANs with a large grid size may be more efficient than MLPs at approximating certain functions. On the other hand, from the perspective of learning and optimization, we study the spectral bias of KANs compared with MLPs. We demonstrate that KANs are less biased toward low frequencies than MLPs. We highlight that the multi-level learning feature specific to KANs, i.e. grid extension of splines, improves the learning process for high-frequency components. Detailed comparisons with different choices of depth, width, and grid sizes of KANs are made, shedding some light on how to choose the hyperparameters in practice.", "keywords": [ "Kolmogorov-Arnold Network", "Spectral Bias", "Approximation Theory" ], "rating_list": [ 6, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 2, 4, 5 ] }, { "paper_id": "ydREOIttdC", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Guanghui Wang", "gender": "Male", "institution": "Toronto Metropolitan University", "country": "CA", "position": "Associate Professor" }, { "name": "IL MIN KIM", "gender": "Male", "institution": "Queen's University", "country": "CA", "position": "Full Professor" }, { "name": "Milad Khademi Nori", "gender": "Male", "institution": "Toronto Metropolitan University", "country": "CA", "position": "Postdoc" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 49, "n_ref": 104, "n_ref_all": 121, "n_fig": 3, "n_tab": 2, "L_tab": 1953, "n_element_tab": 150, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1985, "n_element_tab_1": 151, "formula_len_all": 1304, "formula_len_all_1": 1304, "len_all": 132820, "len_all_1": 74337, "len_abs": 1443, "len_title": 191, "len_sents": 32984, "len_sents_1": 31166, "n_sents": 247, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 142, "L_abs": 1450, "title": "Federated Class-Incremental Learning: A Hybrid Approach Using Latent Exemplars and Data-Free Techniques to Address Local and Global Forgetting", "abs": "Federated Class-Incremental Learning (FCIL) refers to a scenario where a dynamically changing number of clients collaboratively learn an ever-increasing number of incoming tasks. FCIL is known to suffer from local forgetting due to class imbalance at each client and global forgetting due to class imbalance across clients. We develop a mathematical framework for FCIL that formulates local and global forgetting. Then, we propose an approach called Hybrid Rehearsal (HR), which utilizes latent exemplars and data-free techniques to address local and global forgetting, respectively. HR employs a customized autoencoder designed for both data classification and the generation of synthetic data. To determine the embeddings of new tasks for all clients in the latent space of the encoder, the server uses the Lennard-Jones Potential formulations. Meanwhile, at the clients, the decoder decodes the stored low-dimensional latent space exemplars back to the high-dimensional input space, used to address local forgetting. To overcome global forgetting, the decoder generates synthetic data. Furthermore, our mathematical framework proves that our proposed approach HR can, in principle, tackle the two local and global forgetting challenges. In practice, extensive experiments demonstrate that while preserving privacy, our proposed approach outperforms the state-of-the-art baselines on multiple FCIL benchmarks with low compute and memory footprints.", "keywords": [ "Class-Incremental Learning", "Federated Learning", "Global Forgetting", "Local Forgetting." ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "ydH8nU5csJ", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Dailing Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Jing Zhang", "gender": "Female", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Engineer" }, { "name": "Kaiqi Huang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "Professor" }, { "name": "Shiyu Hu", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Xiaokun Feng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "PhD student" }, { "name": "Xuchen Li", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "wu meiqi", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 87, "n_ref_all": 117, "n_fig": 12, "n_tab": 4, "L_tab": 2375, "n_element_tab": 237, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1375, "n_element_tab_1": 146, "formula_len_all": 77, "formula_len_all_1": 0, "len_all": 164363, "len_all_1": 65373, "len_abs": 1967, "len_title": 134, "len_sents": 46886, "len_sents_1": 30069, "n_sents": 351, "n_sents_1": 231, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1975, "title": "DTVLT: A Multi-modal Diverse Text Benchmark for Visual Language Tracking Based on LLM", "abs": "Visual language tracking (VLT) has emerged as a cutting-edge research area, harnessing linguistic data to enhance algorithms with multi-modal inputs and broadening the scope of traditional single object tracking (SOT) to encompass video understanding applications. Despite this, most VLT benchmarks still depend on succinct, human-annotated text descriptions for each video. These descriptions often fall short in capturing the nuances of video content dynamics and lack stylistic variety in language, constrained by their uniform level of detail and a fixed annotation frequency. As a result, algorithms tend to default to a \u201cmemorize the answer\u201d strategy, diverging from the core objective of achieving a deeper understanding of video content. Fortunately, the emergence of large language models (LLMs) has enabled the generation of diverse text. This work utilizes LLMs to generate varied semantic annotations (in terms of text lengths and granularities) for representative SOT benchmarks, thereby establishing a novel multi-modal benchmark. Specifically, we (1) propose a new visual language tracking benchmark with diverse texts, named DTVLT, based on five prominent VLT and SOT benchmarks, including three sub-tasks: short-term tracking, long-term tracking, and global instance tracking. (2) We offer four granularity texts in our benchmark, considering the extent and density of semantic information. This is achieved through DTLLM-VLT, a method for generating high-quality, diverse text by leveraging the extensive knowledge base of LLMs to produce descriptions rich in world knowledge. We expect this multi-granular generation strategy to foster a favorable environment for VLT and video understanding research. (3) We conduct comprehensive experimental analyses on DTVLT, evaluating the impact of diverse text on tracking performance and hope the identified performance bottlenecks of existing algorithms can support further research in VLT and video understanding.", "keywords": [ "Visual Language Tracking", "Video Understanding", "Large Language Model" ], "rating_list": [ 5, 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 5, 3 ] }, { "paper_id": "yc38vnXhTh", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chen Liang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" }, { "name": "Wenguan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zhiyuan Weng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 64, "n_ref": 95, "n_ref_all": 119, "n_fig": 9, "n_tab": 5, "L_tab": 1240, "n_element_tab": 76, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 130, "n_element_tab_1": 18, "formula_len_all": 115, "formula_len_all_1": 0, "len_all": 183595, "len_all_1": 70203, "len_abs": 1433, "len_title": 0, "len_sents": 51612, "len_sents_1": 35389, "n_sents": 373, "n_sents_1": 258, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1439, "title": "Towards Human-like Virtual Beings: Simulating Human Behavior in 3D Scenes", "abs": "Building autonomous agents that can replicate human behavior in the realistic 3D world is a key step toward artificial general intelligence. This requires agents to be holistic goal achievers and to naturally adapt to environmental dynamics. In this work, we introduce ACTOR, an agent capable of performing high-level, long-horizon, abstract goals in 3D households, guided by its internal value similar to those of humans. ACTOR operates in a perceive-plan-act cycle, extending the ungrounded, scene-agnostic LLM controller with deliberate goal decomposition and decision-making through actively searching the behavior space, generating activity choices based on a hierarchical prior, and evaluating these choices using customizable value functions to determine the subsequent steps. Furthermore, we introduce BehaviorHub, a large-scale human behavior simulation dataset in scene-aware, complicated tasks. Considering the unaffordable acquisition of human-authored 3D human behavior data, we construct BehaviorHub by exploring the commonsense knowledge of LLMs learned from large corpora, and automatically aligning motion resources with 3D scene for knowledgeable generation. Extensive experiments on our established benchmark demonstrate that the proposed architecture leads to effective behavior planning and simulation. BehaviorHub also proves beneficial for downstream task development. Our code and dataset will be publicly released.", "keywords": [ "Agent AI", "3D Humanoid", "Large Language Model", "Deep Learning" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "ybfmpJiKXX", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Adriana Eufrosina Bora", "gender": "Female", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "", "position": "Researcher" }, { "name": "Arsene Fansi Tchango", "gender": "Male", "institution": "Montreal Institute of Learning Algorithms", "country": "", "position": "Senior Researcher" }, { "name": "Bruno Rousseau", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "Postdoc" }, { "name": "Kerrie Mengersen", "gender": "Female", "institution": "Queensland University of Technology", "country": "AU", "position": "Full Professor" }, { "name": "Mirko Bronzi", "gender": "unknown", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "", "position": "Researcher" }, { "name": "Pierre-Luc St-Charles", "gender": "unknown", "institution": "Mila", "country": "CA", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 40, "n_ref_all": 69, "n_fig": 9, "n_tab": 6, "L_tab": 4025, "n_element_tab": 443, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2031, "n_element_tab_1": 212, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 264263, "len_all_1": 67598, "len_abs": 5227, "len_title": 141, "len_sents": 98569, "len_sents_1": 33847, "n_sents": 720, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1325, "title": "AIMS.au: A Dataset for the Analysis of Modern Slavery Countermeasures in Corporate Statements", "abs": "Despite over a decade of legislative efforts to address modern slavery in the supply chains of large corporations, the effectiveness of government oversight remains hampered by the challenge of scrutinizing thousands of statements annually. While Large Language Models (LLMs) can be considered a well established solution for the automatic analysis and summarization of documents, recognizing concrete modern slavery countermeasures taken by companies and differentiating those from vague claims remains a challenging task. To help evaluate and fine-tune LLMs for the assessment of corporate statements, we introduce a dataset composed of 5,731 modern slavery statements taken from the Australian Modern Slavery Register and annotated at the sentence level. This paper details the construction steps for the dataset that include the careful design of annotation specifications, the selection and preprocessing of statements, and the creation of high-quality annotation subsets for effective model evaluations. To demonstrate our dataset's utility, we propose a machine learning methodology for the detection of sentences relevant to mandatory reporting requirements set by the Australian Modern Slavery Act. We then follow this methodology to benchmark modern language models under zero-shot and supervised learning settings.", "keywords": [ "natural language processing", "modern slavery", "corporate statements", "benchmark", "text extraction", "large language models" ], "rating_list": [ 5, 8, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "ybWOYIuFl6", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Bo Qiang", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "RuiKang OuYang", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Research Assistant" }, { "name": "Jose Miguel Hernandez", "gender": "unknown", "institution": "Microsoft Research Cambridge", "country": "GB", "position": "Research Scientist" } ], "n_formula": 135, "n_formula_1": 35, "n_ref_uni": 20, "n_ref": 49, "n_ref_all": 95, "n_fig": 14, "n_tab": 11, "L_tab": 2471, "n_element_tab": 222, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 83, "n_element_tab_1": 14, "formula_len_all": 13450, "formula_len_all_1": 2565, "len_all": 194288, "len_all_1": 60901, "len_abs": 914, "len_title": 71, "len_sents": 65511, "len_sents_1": 26474, "n_sents": 591, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 919, "title": "BNEM: A Boltzmann Sampler Based on Bootstrapped Noised Energy Matching", "abs": "Generating independent samples from a Boltzmann distribution is a highly relevant problem in scientific research, e.g. in molecular dynamics, where one has initial access to the underlying energy function but not to samples from the Boltzmann distribution. We address this problem by learning the energies of the convolution of the Boltzmann distribution with Gaussian noise. These energies are then used to generate independent samples through a denoising diffusion approach. The resulting method, Noised Energy Matching (NEM), has lower variance and only slightly higher cost than previous related works. We also improve NEM through a novel bootstrapping technique called Bootstrap NEM (BNEM) that further reduces variance while only slightly increasing bias. Experiments on a collection of problems demonstrate that NEM can outperform previous methods while being more robust and that BNEM further improves on NEM.", "keywords": [ "neural sampler", "Boltzmann distribution", "diffusion model" ], "rating_list": [ 5, 3, 8, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ybFRoGxZjs", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Chenggang Yan", "gender": "Male", "institution": "Hangzhou Dianzi University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Hangyu Chen", "gender": "Male", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "MS student" }, { "name": "Le zhang", "gender": "Female", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "Associate Professor" }, { "name": "Rongfeng Lu", "gender": "Male", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "PhD student" }, { "name": "Yuhang Qin", "gender": "Male", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "Undergrad student" }, { "name": "Zunjie Zhu", "gender": "Male", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "Assistant Professor" }, { "name": "anke xue", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Postdoc" }, { "name": "Ming Lu", "gender": "Male", "institution": "Intel Labs China", "country": "CN", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 41, "n_ref": 61, "n_ref_all": 88, "n_fig": 12, "n_tab": 4, "L_tab": 2059, "n_element_tab": 183, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 2049, "n_element_tab_1": 175, "formula_len_all": 464, "formula_len_all_1": 529, "len_all": 105711, "len_all_1": 58539, "len_abs": 446, "len_title": 46, "len_sents": 29361, "len_sents_1": 26266, "n_sents": 218, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 46, "L_abs": 1396, "title": "ThermalGaussian: Thermal 3D Gaussian Splatting", "abs": "Thermography is especially valuable for the military and other users of surveillance cameras. Some recent methods based on Neural Radiance Fields (NeRF) are proposed to reconstruct the thermal scenes in 3D from a set of thermal and RGB images. However, unlike NeRF, 3D Gaussian splatting (3DGS) prevails due to its rapid training and real-time rendering. In this work, we propose ThermalGaussian, the first thermal 3DGS approach capable of rendering high-quality images in RGB and thermal modalities. We first calibrate the RGB camera and the thermal camera to ensure that both modalities are accurately aligned. Subsequently, we use the registered images to learn the multimodal 3D Gaussians. To prevent the overfitting of any single modality, we introduce several multimodal regularization constraints. We also develop smoothing constraints tailored to the physical characteristics of the thermal modality.\nBesides, we contribute a real-world dataset named RGBT-Scenes, captured by a hand-hold thermal-infrared camera, facilitating future research on thermal scene reconstruction. We conduct comprehensive experiments to show that ThermalGaussian achieves photorealistic rendering of thermal images and improves the rendering quality of RGB images. With the proposed multimodal regularization constraints, we also reduced the model's storage cost by 90\\%. The code and dataset will be released.", "keywords": [ "3D reconstruction; Thermal fild reconstruction; 3D Computer Vision; Machine learning approaches;" ], "rating_list": [ 6, 6, 8, 5, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 4 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "yb4QE6b22f", "primary_area": "foundation or frontier models, including LLMs", "n_author": 18, "author_info_list": [ { "name": "Daniel McDuff", "gender": "Male", "institution": "Google", "country": "US", "position": "Principal Researcher" }, { "name": "Girish Narayanswamy", "gender": "Male", "institution": "University of Washington", "country": "", "position": "PhD student" }, { "name": "Jacob Sunshine", "gender": "Male", "institution": "University of Washington", "country": "", "position": "Associate Professor" }, { "name": "Jake Garrison", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Jiening Zhan", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Kumar Ayush", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Mark Malhotra", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Pushmeet Kohli", "gender": "Male", "institution": "Microsoft Research Cambridge", "country": "", "position": "Researcher" }, { "name": "Samy Abdel-Ghaffar", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Shrikanth Narayanan", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Full Professor" }, { "name": "Shwetak Patel", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Tim Althoff", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "", "position": "Assistant Professor" }, { "name": "Xin Liu", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist " }, { "name": "Xuhai Xu", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" }, { "name": "Yun Liu", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Yuzhe Yang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "shun liao", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Shyam Anil Tailor", "gender": "Male", "institution": "Google", "country": "GB", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 55, "n_ref": 105, "n_ref_all": 156, "n_fig": 14, "n_tab": 21, "L_tab": 12240, "n_element_tab": 2859, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 549, "n_element_tab_1": 91, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 252105, "len_all_1": 55723, "len_abs": 1195, "len_title": 82, "len_sents": 64166, "len_sents_1": 27370, "n_sents": 526, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 34, "L_abs": 1200, "title": "Scaling Wearable Foundation Models", "abs": "Wearable sensors have become ubiquitous thanks to a variety of health tracking features. The resulting continuous and longitudinal measurements from everyday life generate large volumes of data; however, making sense of these observations for scientific and actionable insights is non-trivial. Inspired by the empirical success of generative modeling, where large neural networks learn powerful representations from vast amounts of text, image, video, or audio data, we investigate the scaling properties of sensor foundation models across compute, data, and model size. Using a dataset of up to 40 million hours of in-situ heart rate, heart rate variability, electrodermal activity, accelerometer, skin temperature, and altimeter per-minute data from over 165,000 people, we create LSM, a multimodal foundation model built on the largest wearable-signals dataset with the most extensive range of sensor modalities to date. Our results establish the scaling laws of LSMs for tasks such as imputation, interpolation and extrapolation, both across time and sensor modalities. Moreover, we highlight how LSMs enables sample-efficient downstream learning for tasks like exercise and activity recognition.", "keywords": [ "Health", "Foundation Model", "Scaling" ], "rating_list": [ 6, 8, 6, 8, 5 ], "soundness_list": [ 3, 4, 3, 3, 2 ], "presentation_list": [ 3, 4, 4, 3, 3 ], "contribution_list": [ 3, 4, 3, 4, 2 ], "confidence_list": [ 4, 5, 4, 3, 5 ] }, { "paper_id": "yatNm6A6sR", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Guanjie Zheng", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "", "position": "Assistant Professor" }, { "name": "Hongwei Zhang", "gender": "Male", "institution": "Shandong University", "country": "", "position": "MS student" }, { "name": "Linghe Kong", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Su Ziyang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Wen Ling", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Yiheng Wang", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Yuhang Luo", "gender": "unknown", "institution": "", "country": "", "position": "" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 19, "n_ref": 32, "n_ref_all": 52, "n_fig": 6, "n_tab": 10, "L_tab": 10066, "n_element_tab": 784, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3884, "n_element_tab_1": 567, "formula_len_all": 25, "formula_len_all_1": 0, "len_all": 140139, "len_all_1": 58385, "len_abs": 4996, "len_title": 121, "len_sents": 43666, "len_sents_1": 22046, "n_sents": 325, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1198, "title": "OSM+: Cloud-native Open Street Map Data System for City-wide Experiments", "abs": "Road network data can provide rich information about cities and thus become the base for various urban research. However, processing large-volume world-wide road network data requires intensive computing resources and the processed results might be different to be unified for benchmark downstream tasks. Therefore, in this paper, we process the OpenStreetMap data and release a structured world-wide 1-billion-node road network graph database with high accessibility and usability. We have presented three illustrative use cases, traffic prediction task, city boundary detection task and traffic policy control task. Moreover, for the well-investigated traffic prediction task, we release a new benchmark with 31 datasets, which is much more comprehensive than the previously frequently-used datasets. While for the relatively novel traffic policy control task, we release a new 6 city datasets with much larger scale than the previous datasets. Along with the OSM+ dataset, the release of data converters facilitates the integration of multimodal spatial-temporal data based on map information for large model training, thereby expediting the process of uncovering compelling scientific insights.", "keywords": [ "Global dataset", "Traffic prediction", "Traffic policy control" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "yarlMUJePB", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Wenhan Gao", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Xufeng Liu", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Yi Liu", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 53, "n_ref": 102, "n_ref_all": 127, "n_fig": 5, "n_tab": 6, "L_tab": 748, "n_element_tab": 164, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 918, "n_element_tab_1": 173, "formula_len_all": 601, "formula_len_all_1": 600, "len_all": 148025, "len_all_1": 65053, "len_abs": 1978, "len_title": 123, "len_sents": 39024, "len_sents_1": 29137, "n_sents": 274, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1986, "title": "Energy-Based Discrete Mask Approximation for 3D Molecular Graph Explanation", "abs": "In recent years, Graph Neural Networks (GNNs) have become a powerful tool for modeling molecular data. To enhance their reliability and interpretability, various explanation methods have been developed to identify key molecular substructures, specifically a set of edges, in the decision-making process. Early work with 2D GNNs represented molecules as graphs with atoms as nodes and bonds as edges, neglecting 3D geometric configurations. While existing explanation methods perform well on 2D GNNs, there is a pressing need for 3D explanation methods tailored for 3D GNNs, which outperform 2D GNNs in many tasks. Current explanation methods struggle with 3D GNNs due to the construction of edges based on cut-off distances in 3D GNNs, resulting in an exponentially large number of edges. We identify the sources of errors in explanations and decompose them into two components based on a derived upper bound between the optimized masks and the actual explanatory subgraph. This gap can be significant, especially for 3D GNNs because of the large number of edges. To achieve optimal explanation fidelity, our method aims to bridge this gap by assigning two energy values to each atom based on its contribution to the prediction: one energy reflects the scenario where this node is important in making the decision, while the other represents the scenario where it is unimportant. In analogy to physics, lower energy values indicate greater stability in the prediction, and thus, we are more confident about the scenario with which it is associated. Our approach strives to push up and down the energies, respectively, to distinguish these two scenarios to simultaneously minimize both components of the derived upper bound of error, enabling us to identify a stable subgraph that maintains high explanation fidelity. Experiments conducted on backbone networks and the QM9 dataset demonstrate the effectiveness of our method in providing accurate and reliable explanations for 3D graphs.", "keywords": [ "3D Graph Explanation", "3D Molecular Graphs", "Energy-Based Models", "Discrete Masks" ], "rating_list": [ 5, 3, 3, 3, 3 ], "soundness_list": [ 3, 1, 1, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "yaqPf0KAlN", "primary_area": "datasets and benchmarks", "n_author": 20, "author_info_list": [ { "name": "Baobao Chang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "benyou wang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Bofei Gao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Chenghao Ma", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "MS student" }, { "name": "Daoguang Zan", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Feifan Song", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Ge Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Lei Sha", "gender": "Male", "institution": "Beihang University", "country": "", "position": "Full Professor" }, { "name": "Liang Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Qingxiu Dong", "gender": "Female", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Runxin Xu", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Tianyu Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Staff Engineer" }, { "name": "Xuancheng Ren", "gender": "unknown", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Yibo Miao", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "MS student" }, { "name": "Yichang Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Zefan Cai", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Zhe Yang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Zhengyang Tang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Tobias Lee", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shr Quan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 25, "n_ref": 42, "n_ref_all": 85, "n_fig": 17, "n_tab": 9, "L_tab": 3231, "n_element_tab": 244, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1352, "n_element_tab_1": 141, "formula_len_all": 480, "formula_len_all_1": 289, "len_all": 196899, "len_all_1": 64096, "len_abs": 1819, "len_title": 132, "len_sents": 63869, "len_sents_1": 27081, "n_sents": 532, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1232, "title": "Omni-MATH: A Universal Olympiad Level Mathematic Benchmark for Large Language Models", "abs": "Recent advancements in large language models (LLMs) have led to significant breakthroughs in mathematical reasoning capabilities. \nHowever, existing benchmarks like GSM8K or MATH are now being solved with high accuracy (e.g., OpenAI o1 achieves 94.8% on MATH dataset), indicating their inadequacy for truly challenging these models. To bridge this gap, we propose a comprehensive and challenging benchmark specifically designed to assess LLMs' mathematical reasoning at the Olympiad level. Unlike existing Olympiad-related benchmarks, our dataset focuses exclusively on mathematics and comprises a vast collection of 4428 competition-level problems with rigorous human annotation. These problems are meticulously categorized into over 33 sub-domains and span more than 10 distinct difficulty levels, enabling a holistic assessment of model performance in Olympiad-mathematical reasoning. Furthermore, we conducted an in-depth analysis based on this benchmark. Our experimental results show that even the most advanced models, OpenAI o1-mini and OpenAI o1-preview, struggle with highly challenging Olympiad-level problems, with 60.54% and 52.55% accuracy, highlighting significant challenges in Olympiad-level mathematical reasoning.", "keywords": [ "Mathematical Benchmark", "LLM Evaluation", "Olympic" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "yaR0hqaGbI", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Jundong Li", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" }, { "name": "Song Wang", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Wendy Zheng", "gender": "Female", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Yaochen Zhu", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "", "position": "PhD student" }, { "name": "Yinhan He", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Yushun Dong", "gender": "Male", "institution": "Florida State University", "country": "US", "position": "Assistant Professor" }, { "name": "Zaiyi Zheng", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "", "position": "PhD student" } ], "n_formula": 52, "n_formula_1": 10, "n_ref_uni": 32, "n_ref": 70, "n_ref_all": 101, "n_fig": 13, "n_tab": 13, "L_tab": 3329, "n_element_tab": 197, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1661, "n_element_tab_1": 58, "formula_len_all": 3767, "formula_len_all_1": 630, "len_all": 245050, "len_all_1": 67869, "len_abs": 2050, "len_title": 128, "len_sents": 87957, "len_sents_1": 31985, "n_sents": 816, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 2060, "title": "Hierarchical Demonstration Order Optimization for Many-shot In-Context Learning", "abs": "In-Context Learning (ICL) is a technique where large language models (LLMs) leverage multiple demonstrations (i.e., examples) to perform tasks. With the recent expansion of LLM context windows, many-shot ICL (generally with more than 50 demonstrations) can lead to significant performance improvements on a variety of language tasks such as text classification and question answering. Nevertheless, ICL faces demonstration order instability (ICL-DOI), which means that performance varies significantly depending on the order of demonstrations. Moreover, the ICL-DOI phenomenon persists and can sometimes be more pronounced in many-shot ICL, validated by our thorough experimental investigation. Current strategies handling ICL-DOI, however, are not applicable to many-shot ICL, since they cannot overcome two critical challenges: (1) Most metrics measuring the quality of demonstration order rely on subjective judgment, lacking a theoretical foundation to achieve precise quality characterization. These metrics are thus non-applicable to many-shot situations, where the order quality of different orders is less distinguishable due to the limited ability of LLMs to exploit information in long input contexts. (2) The requirement to examine all orders is computationally infeasible due to the combinatorial complexity of the order space in many-shot ICL. To tackle the first challenge, we design a demonstration order evaluation metric based on information theory for measuring order quality, which effectively quantifies the usable information gain of a given demonstration order. To address the second challenge, we propose a hierarchical demonstration order optimization method named HIDO that enables a more refined exploration of the order space, achieving high ICL performance without the need to evaluate all possible orders. Extensive experiments on multiple LLMs and real-world datasets demonstrate that our HIDO method consistently and efficiently outperforms other baselines. Our code can be found at https://anonymous.4open.science/r/HIDO-B2DE/.", "keywords": [ "In-context learning", "Demonstration Order Optimization" ], "rating_list": [ 8, 5, 1, 3 ], "soundness_list": [ 4, 2, 1, 2 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "yaQbTAD2JJ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Boris Ivanovic", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Boyi Li", "gender": "Female", "institution": "NVIDIA Research", "country": "US", "position": "Researcher" }, { "name": "Edward Schmerling", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Jang Hyun Cho", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Intern" }, { "name": "Marco Pavone", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Director, Autonomous Vehicle Research" }, { "name": "Philipp Kraehenbuehl", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Xinshuo Weng", "gender": "Female", "institution": "NVIDIA", "country": "CA", "position": "Researcher" }, { "name": "Yan Wang", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yue Wang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Yulong Cao", "gender": "unknown", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Yurong You", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 30, "n_ref": 68, "n_ref_all": 103, "n_fig": 16, "n_tab": 6, "L_tab": 3401, "n_element_tab": 380, "n_fig_1": 10, "n_tab_1": 6, "L_tab_1": 1394, "n_element_tab_1": 236, "formula_len_all": 475, "formula_len_all_1": 475, "len_all": 172390, "len_all_1": 51254, "len_abs": 1530, "len_title": 91, "len_sents": 37857, "len_sents_1": 20159, "n_sents": 326, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 1536, "title": "Language-Image Models with 3D Understanding", "abs": "Multi-modal large language models (MLLMs) have shown incredible capabilities in a variety of 2D vision and language tasks. We extend MLLMs\u2019 perceptual capabilities to ground and reason about images in 3-dimensional space. To that end, we first develop a large-scale pretraining dataset for 2D and 3D called LV3D by combining multiple existing 2D and 3D recognition datasets under a common task formulation: as multi-turn question-answering. Next, we introduce a new MLLM named CUBE-LLM and pre-train it on LV3D. We show that pure data scaling makes a strong 3D perception capability without 3D specific architectural design or training objective. CUBE-LLM exhibits intriguing properties similar to LLMs: (1) CUBE-LLM can apply chain-of-thought prompting to improve 3D understanding from 2D context information. (2) CUBE-LLM can follow complex and diverse instructions and adapt to versatile input and output formats. (3) CUBE-LLM can be visually prompted such as 2D box or a set of candidate 3D boxes from specialists. Our experiments on outdoor benchmarks demonstrate that CUBE-LLM significantly outperforms existing baselines by 21.3 points of AP-BEV on the Talk2Car dataset for 3D grounded reasoning and 17.7 points on the DriveLM dataset for complex reasoning about driving scenarios, respectively. CUBE-LLM also shows competitive results in general MLLM benchmarks such as refCOCO for 2D grounding with (87.0) average score, as well as visual question answering benchmarks such as VQAv2, GQA, SQA, POPE, etc. for complex reasoning.", "keywords": [ "Multi-modal Large Language Model with 3D Understanding; 3D Image Grounding from Image" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "yaOe2xBcLC", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Sen Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "AU", "position": "Researcher" }, { "name": "Siyuan Liang", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Zheng Yi Ho", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Yibing Zhan", "gender": "Male", "institution": "JD Explore Academy", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 50, "n_ref": 69, "n_ref_all": 118, "n_fig": 10, "n_tab": 12, "L_tab": 7128, "n_element_tab": 862, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1767, "n_element_tab_1": 176, "formula_len_all": 310, "formula_len_all_1": 330, "len_all": 187648, "len_all_1": 66845, "len_abs": 1825, "len_title": 130, "len_sents": 49405, "len_sents_1": 30471, "n_sents": 377, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1530, "title": "NoVo: Norm Voting off Hallucinations with Attention Heads in Large Language Models", "abs": "Hallucinations in Large Language Models (LLMs) remain a major obstacle, particularly in high-stakes applications where factual accuracy is critical. While representation editing and reading methods have made strides in reducing hallucinations, their heavy reliance on specialised tools and training on in-domain samples, makes them difficult to scale and prone to overfitting. This limits their accuracy gains and generalizability to diverse datasets. This paper presents a lightweight method, Norm Voting (NoVo), which harnesses the untapped potential of attention head norms to dramatically enhance factual accuracy in zero-shot multiple-choice questions (MCQs). NoVo begins by automatically selecting truth-correlated head norms with an efficient, inference-only algorithm using only 30 random samples, allowing NoVo to effortlessly scale to diverse datasets. Afterwards, selected head norms are employed in a simple voting algorithm, which yields significant gains in prediction accuracy. On TruthfulQA MC1, NoVo surpasses the current state-of-the-art and all previous methods by an astounding margin---at least 19 accuracy points. NoVo demonstrates exceptional generalization to 20 diverse datasets, with significant gains in over 90\\% of them, far exceeding all current representation editing and reading methods. NoVo also reveals promising gains to finetuning strategies and building textual adversarial defence. NoVo's effectiveness with head norms opens new frontiers in LLM interpretability, robustness and reliability.", "keywords": [ "Hallucination Mitigation", "Large Language Models", "TruthfulQA", "Representation Editing", "Multiple Choice Question Answering", "Attention Heads" ], "rating_list": [ 5, 8, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 4, 4 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "yZdPpKTO9R", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Cong Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": " LLM Algorithm Research Scientist" }, { "name": "Wen Song", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhiguang Cao", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Sun Jing", "gender": "unknown", "institution": "City University of Macau", "country": "MO", "position": "Assistant Professor" } ], "n_formula": 37, "n_formula_1": 30, "n_ref_uni": 38, "n_ref": 72, "n_ref_all": 90, "n_fig": 7, "n_tab": 3, "L_tab": 350, "n_element_tab": 28, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3096, "formula_len_all_1": 2228, "len_all": 147129, "len_all_1": 71842, "len_abs": 1481, "len_title": 130, "len_sents": 47746, "len_sents_1": 32481, "n_sents": 368, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1506, "title": "Decision-making with speculative opponent model-aided value function factorization", "abs": "In many real-world scenarios, teams of agents must coordinate their actions while competing against opponents. Traditional multi-agent reinforcement learning (MARL) approaches often treat opponents as part of the environment, causing controlled agents to overlook the impact of their adversaries. Opponent modeling can enhance an agent\u2019s decision-making by constructing predictive models of other agents. However, existing approaches typically rely on centralized learning with access to opponent data, and the process of extracting decentralized policies becomes impractical with larger teams. To address this issue, we propose the Distributional Speculative Opponent-aided mixing framework (DSOMIX), a novel value-based speculative opponent modeling algorithm that relies solely on local information\u2014namely the agent's own observations, actions, and rewards. DSOMIX uses speculative beliefs to predict the behaviors of unseen opponents, enabling agents to make decisions based on local observations. Additionally, it incorporates distributional value decomposition models to capture a more granular representation of the agent's return distribution, improving the training process for the speculative opponent models. We formally derive a value-based theorem that underpins the training process. Extensive experiments across four challenging MARL benchmarks, including MPE and Pommerman, demonstrate that DSOMIX outperforms state-of-the-art methods, achieving superior performance and faster convergence.", "keywords": [ "Decision making", "Cooperative multi-agent reinforcement learning;" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "yZ7sn9pyqb", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Emily Black", "gender": "Female", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Fan Wu", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "PhD student" }, { "name": "Varun Chandrasekaran", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 47, "n_ref": 75, "n_ref_all": 156, "n_fig": 34, "n_tab": 5, "L_tab": 1505, "n_element_tab": 75, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 291, "formula_len_all_1": 112, "len_all": 218567, "len_all_1": 69619, "len_abs": 1192, "len_title": 47, "len_sents": 73430, "len_sents_1": 35247, "n_sents": 523, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1200, "title": "Generative Monoculture in Large Language Models", "abs": "We introduce {\\em generative monoculture}, a behavior observed in large language models (LLMs) characterized by a significant narrowing of model output diversity relative to available training data for a given task: for example, generating only positive book reviews for books with a mixed reception. While in some cases, generative monoculture enhances performance (e.g., LLMs more often produce efficient code), the dangers are exacerbated in others (e.g., LLMs refuse to share diverse opinions). As LLMs are increasingly used in high-impact settings such as education and web search, careful maintenance of LLM output diversity is essential to ensure a variety of facts and perspectives are preserved over time. We experimentally demonstrate the prevalence of generative monoculture through analysis of book review and code generation tasks, and find that simple countermeasures such as altering sampling or prompting strategies are insufficient to mitigate the behavior. Moreover, our results suggest that the root causes of generative monoculture are likely embedded within the LLM's alignment processes, suggesting a need for developing fine-tuning paradigms that preserve or promote diversity.", "keywords": [ "monoculture", "bias", "alignment" ], "rating_list": [ 5, 6, 8, 6, 5 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 4, 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 4, 2 ], "confidence_list": [ 4, 4, 4, 5, 4 ] }, { "paper_id": "yYxEFC3Ep4", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Abhipsa Basu", "gender": "Female", "institution": "Indian Institute of Science", "country": "IN", "position": "PhD student" }, { "name": "Danish Pruthi", "gender": "Male", "institution": "Indian Institute of Science, Bangalore ", "country": "", "position": "Assistant Professor" }, { "name": "Kirti Bhagat", "gender": "Female", "institution": "Indian Institute of Science, Indian institute of science, Bangalore", "country": "", "position": "Researcher" }, { "name": "Preethi Seshadri", "gender": "Female", "institution": "University of California, Irvine", "country": "", "position": "PhD student" }, { "name": "Venkatesh Babu R", "gender": "Male", "institution": "Indian Institute of Science", "country": "IN", "position": "Full Professor" }, { "name": "Yugam Bahl", "gender": "Male", "institution": "Indian Institute of Science, Indian institute of science, Bangalore", "country": "IN", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 36, "n_ref": 70, "n_ref_all": 124, "n_fig": 9, "n_tab": 9, "L_tab": 3365, "n_element_tab": 433, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 224, "n_element_tab_1": 38, "formula_len_all": 142, "formula_len_all_1": 114, "len_all": 198568, "len_all_1": 64924, "len_abs": 1400, "len_title": 128, "len_sents": 61336, "len_sents_1": 32019, "n_sents": 472, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1375, "title": "Where Do Images Come From? Analyzing Captions to Geographically Profile Datasets", "abs": "Building on studies documenting gender and racial biases in vision-language models, recent works show that such models often fail to generate geographically-representative images that accurately reflect different regions around the world. A common concern is that the data used to train these models is not representative, prompting the question: *which parts of the world do these training examples come from?* To answer this question, we develop a system, *GeoProfiler*, which geographically profiles multimodal datasets by mapping image-caption pairs to countries. Using location information from captions, GeoProfiler maps examples to countries with a high precision ($0.86$). We then apply *GeoProfiler* to geographically profile the English captions of the LAION dataset for $10$ common entities (e.g., house, flag, etc.). We observe the geographical distribution of $8$ entities to obey the power law distribution. The United States, the United Kingdom, and India are most represented, appearing in 53.7% of samples. Problematically, African and South American countries are severely under-represented with only 2.0 % and 4.3 % of images respectively. We also observe a high correlation between a country's GDP and frequency ($\\rho=0.79$). Lastly, we analyze the diversity of images from individual countries, and find that more images does not imply higher diversity.", "keywords": [ "geographical profiling", "dataset auditing" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "yYZbZGo4ei", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Chang Zou", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Intern" }, { "name": "Linfeng Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Siteng Huang", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Ting Liu", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Xuyang Liu", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "MS student" } ], "n_formula": 19, "n_formula_1": 11, "n_ref_uni": 24, "n_ref": 69, "n_ref_all": 97, "n_fig": 10, "n_tab": 9, "L_tab": 3575, "n_element_tab": 421, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1536, "n_element_tab_1": 224, "formula_len_all": 1570, "formula_len_all_1": 840, "len_all": 163299, "len_all_1": 62644, "len_abs": 1166, "len_title": 115, "len_sents": 57998, "len_sents_1": 29639, "n_sents": 407, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1181, "title": "Accelerating Diffusion Transformers with Token-wise Feature Caching", "abs": "Diffusion transformers have shown significant effectiveness in both image and video synthesis at the expense of huge computation costs. To address this problem, feature caching methods have been introduced to accelerate diffusion transformers by caching the features in previous timesteps and reusing them in the following timesteps. However, previous caching methods ignore that different tokens exhibit different sensitivities to feature caching, and feature caching on some tokens may lead to 10X more destruction to the overall generation quality compared with other tokens. In this paper, we introduce token-wise feature caching, allowing us to adaptively select the most suitable tokens for caching, and further enable us to apply different caching ratios to neural layers in different types and depths. Extensive experiments on PixArt-alpha, OpenSora, and DiT demonstrate our effectiveness in both image and video generation with no requirements for training. For instance, 2.36X and 1.93X acceleration are achieved on OpenSora and PixArt-alpha with almost no drop in generation quality. Codes have been released in the supplementary material and will be released in Github.", "keywords": [ "Diffusion Models", "Image generation", "Video generation", "Model Acceleration", "Feature Cache" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "yYQLvofQ1k", "primary_area": "applications to robotics, autonomy, planning", "n_author": 8, "author_info_list": [ { "name": "Haoyang Su", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Jinzhe Li", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Nanqing Dong", "gender": "unknown", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Associate Professor" }, { "name": "Renqi Chen", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "SHIXIANG TANG", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Wanli Ouyang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Full Professor" }, { "name": "Xinzhe Zheng", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Zhenfei Yin", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 56, "n_ref": 94, "n_ref_all": 140, "n_fig": 28, "n_tab": 8, "L_tab": 2617, "n_element_tab": 201, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1216, "n_element_tab_1": 28, "formula_len_all": 568, "formula_len_all_1": 364, "len_all": 254620, "len_all_1": 68156, "len_abs": 1106, "len_title": 155, "len_sents": 93728, "len_sents_1": 33161, "n_sents": 694, "n_sents_1": 234, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1109, "title": "Two Heads Are Better Than One: A Multi-Agent System Has the Potential to Improve Scientific Idea Generation", "abs": "The rapid advancement of scientific progress requires innovative tools that can accelerate discovery. While recent AI methods, particularly large language models (LLMs), have shown promise in tasks such as hypothesis generation and experimental design, they fall short in replicating the collaborative nature of real-world scientific practices, where diverse teams of experts work together to tackle complex problems. To address the limitation, we propose an LLM-based multi-agent system, i.e., Virtual Scientists (VirSci), designed to mimic the teamwork inherent in scientific research. VirSci organizes a team of agents to collaboratively generate, evaluate, and refine research ideas. Through comprehensive experiments, we demonstrate that this multi-agent approach outperforms the state-of-the-art method in producing novel and impactful scientific ideas, showing potential in aligning with key insights in the Science of Science field. Our findings suggest that integrating collaborative agents can lead to more innovative scientific outputs, offering a robust system for autonomous scientific discovery.", "keywords": [ "Large Language Model", "Multi-agent System", "Collaboration Strategy", "Automatic Scientific Discovery", "Science of Science" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "yXCTDhZDh6", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Fanbo Xiang", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Hao Su", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Jiayuan Gu", "gender": "unknown", "institution": "ShanghaiTech University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tung Yen Chiang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "yuchen zhou", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 33, "n_ref": 76, "n_ref_all": 96, "n_fig": 4, "n_tab": 11, "L_tab": 3196, "n_element_tab": 379, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 3190, "n_element_tab_1": 210, "formula_len_all": 450, "formula_len_all_1": 54, "len_all": 160438, "len_all_1": 59506, "len_abs": 969, "len_title": 108, "len_sents": 43162, "len_sents_1": 26375, "n_sents": 373, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 925, "title": "Point-SAM: Promptable 3D Segmentation Model for Point Clouds", "abs": "The development of 2D foundation models for image segmentation has been significantly advanced by the Segment Anything Model (SAM). However, achieving similar success in 3D models remains a challenge due to issues such as non-unified data formats, poor model scalability, and the scarcity of labeled data with diverse masks. To this end, we propose a 3D promptable segmentation model Point-SAM, focusing on point clouds. We employ an efficient transformer-based architecture tailored for point clouds, extending SAM to the 3D domain. We then distill the rich knowledge from 2D SAM for Point-SAM training by introducing a data engine to generate part-level and object-level pseudo-labels at scale from 2D SAM. Our model outperforms state-of-the-art 3D segmentation models on several indoor and outdoor benchmarks and demonstrates a variety of applications, such as interactive 3D annotation and zero-shot 3D instance proposal.", "keywords": [ "3D vision", "promptable segmentation", "point cloud segmentation" ], "rating_list": [ 6, 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 5, 5, 4, 3, 4 ] }, { "paper_id": "yWoV4Ca6ji", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Gabriel Peyr\u00e9", "gender": "Male", "institution": "ENS", "country": "FR", "position": "Researcher" }, { "name": "Michael Eli Sander", "gender": "Male", "institution": "Google ", "country": "FR", "position": "Researcher" } ], "n_formula": 79, "n_formula_1": 24, "n_ref_uni": 32, "n_ref": 53, "n_ref_all": 64, "n_fig": 4, "n_tab": 2, "L_tab": 1071, "n_element_tab": 1, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 67, "n_element_tab_1": 2, "formula_len_all": 5317, "formula_len_all_1": 1560, "len_all": 189986, "len_all_1": 88639, "len_abs": 1293, "len_title": 128, "len_sents": 45437, "len_sents_1": 29469, "n_sents": 461, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1323, "title": "Towards Understanding the Universality of Transformers for Next-Token Prediction", "abs": "Causal Transformers are trained to predict the next token for a given context. While it is widely accepted that self-attention is crucial for encoding the causal structure of sequences, the precise underlying mechanism behind this in-context autoregressive learning ability remains unclear. In this paper, we take a step towards understanding this phenomenon by studying the approximation ability of Transformers for next-token prediction. Specifically, we explore the capacity of causal Transformers to predict the next token $x_{t+1}$ given an autoregressive sequence $(x_1, \\dots, x_t)$ as a prompt, where $ x_{t+1} = f(x_t) $, and $ f $ is a context-dependent function that varies with each sequence.\nOn the theoretical side, we focus on specific instances, namely when $ f $ is linear or when $ (x_t)$ is periodic. We explicitly construct a Transformer (with linear, exponential, or softmax attention) that learns the mapping $f$ in-context through a causal kernel descent method. The causal kernel descent method we propose provably estimates $x_{t+1} $ based solely on past and current observations $ (x_1, \\dots, x_t) $, with connections to the Kaczmarz algorithm in Hilbert spaces. We present experimental results that validate our theoretical findings and suggest their applicability to more general mappings $f$.", "keywords": [ "Transformers", "In-Context Learning", "Deep Learning Theory" ], "rating_list": [ 5, 6, 6, 3, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 3, 2, 4, 4, 3 ] }, { "paper_id": "yVeNBxwL5W", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Ao Li", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Ge Yang", "gender": "Male", "institution": "National Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of sciences", "country": "CN", "position": "Full Professor" }, { "name": "Hongbo Zhao", "gender": "Male", "institution": "Institute of automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Le Lu", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Full Professor" }, { "name": "Minfeng Xu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Wei Fang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 72, "n_formula_1": 28, "n_ref_uni": 22, "n_ref": 85, "n_ref_all": 171, "n_fig": 5, "n_tab": 12, "L_tab": 3542, "n_element_tab": 341, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1941, "n_element_tab_1": 171, "formula_len_all": 8820, "formula_len_all_1": 2069, "len_all": 162646, "len_all_1": 58215, "len_abs": 1493, "len_title": 126, "len_sents": 38454, "len_sents_1": 22860, "n_sents": 381, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 77, "L_abs": 1501, "title": "MRS: A Fast Sampler for Mean Reverting Diffusion based on ODE and SDE Solvers", "abs": "In applications of diffusion models, controllable generation is of practical significance, but is also challenging. Current methods for controllable generation primarily focus on modifying the score function of diffusion models, while Mean Reverting (MR) Diffusion directly modifies the structure of the stochastic differential equation (SDE), making the incorporation of image conditions simpler and more natural. However, current training-free fast samplers are not directly applicable to MR Diffusion. And thus MR Diffusion requires hundreds of NFEs (number of function evaluations) to obtain high-quality samples. In this paper, we propose a new algorithm named MRS (MR Sampler) to reduce the sampling NFEs of MR Diffusion. We solve the reverse-time SDE and the probability flow ordinary differential equation (PF-ODE) associated with MR Diffusion, and derive semi-analytical solutions. The solutions consist of an analytical function and an integral parameterized by a neural network. Based on this solution, we can generate high-quality samples in fewer steps. Our approach does not require training and supports all mainstream parameterizations, including noise prediction, data prediction and velocity prediction. Extensive experiments demonstrate that MR Sampler maintains high sampling quality with a speedup of 10 to 20 times across ten different image restoration tasks. Our algorithm accelerates the sampling procedure of MR Diffusion, making it more practical in controllable generation.", "keywords": [ "Fast Sampler", "Mean Reverting Diffusion" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "yVVzaRE8Pi", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Lingrui Mei", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Lujun Li", "gender": "Male", "institution": "HKGAI", "country": "HK", "position": "Researcher" }, { "name": "Ruibin Yuan", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Tianyu Wu", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Researcher" }, { "name": "Wei Xue", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Yike Guo", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 52, "n_ref": 105, "n_ref_all": 126, "n_fig": 4, "n_tab": 15, "L_tab": 6419, "n_element_tab": 337, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3318, "n_element_tab_1": 189, "formula_len_all": 217, "formula_len_all_1": 426, "len_all": 258707, "len_all_1": 62418, "len_abs": 1393, "len_title": 113, "len_sents": 78982, "len_sents_1": 26704, "n_sents": 536, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1434, "title": "You Know What I'm Saying: Jailbreak Attack via Implicit Reference", "abs": "While recent advancements in large language model (LLM) alignment have enabled the effective identification of malicious objectives involving scene nesting and keyword rewriting, our study reveals that these methods remain inadequate at detecting malicious objectives expressed through context within nested harmless objectives.\nThis study identifies a previously overlooked vulnerability, which we term $\\textbf{A}$ttack via $\\textbf{I}$mplicit $\\textbf{R}$eference ($\\textbf{AIR}$). AIR decomposes a malicious objective into permissible objectives and links them through implicit references within the context. This method employs multiple related harmless objectives to generate malicious content without triggering refusal responses, thereby effectively bypassing existing detection techniques.\nOur experiments demonstrate AIR's effectiveness across state-of-the-art LLMs, achieving an attack success rate (ASR) exceeding $\\textbf{90}$% on most models, including GPT-4o, Claude-3.5-Sonnet, and Qwen-2-72B. Notably, we observe an inverse scaling phenomenon, where larger models are more vulnerable to this attack method. These findings underscore the urgent need for defense mechanisms capable of understanding and preventing contextual attacks. Furthermore, we introduce a cross-model attack strategy that leverages less secure models to generate malicious contexts, thereby further increasing the ASR when targeting other models.", "keywords": [ "Adversarial attacks", "Jailbreak", "Security", "Black box", "LLM", "Alignment", "Cross-Modality alignment", "in context learning" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "yVQcr4qjD6", "primary_area": "foundation or frontier models, including LLMs", "n_author": 12, "author_info_list": [ { "name": "Cheng Cheng", "gender": "Male", "institution": "Guangdong OPPO Mobile Telecommunications Corp.,Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Guanyu Nie", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "PhD student" }, { "name": "Jiamu Zhou", "gender": "unknown", "institution": "Guangdong OPPO Mobile Telecommunications Corp.,Ltd.", "country": "CN", "position": "Assistant Professor" }, { "name": "Jun Wang", "gender": "Male", "institution": "OPPO Research Institute ", "country": "CN", "position": "Principal Researcher" }, { "name": "Jun Wang", "gender": "unknown", "institution": "Guangdong OPPO Mobile Telecommunications Corp.,Ltd.", "country": "CN", "position": "MS student" }, { "name": "Junwei Liao", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Muning Wen", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Qiqiang Lin", "gender": "Male", "institution": "Guangdong OPPO Mobile Telecommunications Corp.,Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Qiuying Peng", "gender": "Male", "institution": "OPPO Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Weinan Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaoyun Mo", "gender": "Female", "institution": "Oppo", "country": "", "position": "Researcher" }, { "name": "Yin Zhao", "gender": "Male", "institution": "Guangdong OPPO Mobile Telecommunications Corp.,Ltd.", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 47, "n_ref_all": 73, "n_fig": 6, "n_tab": 6, "L_tab": 5129, "n_element_tab": 560, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 5697, "n_element_tab_1": 336, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 140482, "len_all_1": 60231, "len_abs": 1357, "len_title": 122, "len_sents": 38285, "len_sents_1": 25956, "n_sents": 237, "n_sents_1": 149, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 73, "L_abs": 1194, "title": "Robust Function-Calling for On-Device Language Model via Function Masking", "abs": "Large language models have demonstrated impressive value in performing as autonomous agents when equipped with external tools and API calls. Nonetheless, effectively harnessing their potential for executing complex tasks crucially relies on enhancements in their function-calling capabilities. This paper identifies a critical gap in existing function-calling models, where performance varies significantly across benchmarks, often due to over-fitting to specific naming conventions. To address such an issue, we introduce Hammer, a novel family of foundation models specifically engineered for on-device function calling. Hammer employs an augmented dataset that enhances models\u2019 sensitivity to irrelevant functions and incorporates function masking techniques to minimize over-fitting. Our empirical evaluations reveal that Hammer not only outperforms larger models but also demonstrates robust generalization across diverse benchmarks, achieving state-of-the-art results. Our open-source contributions include a specialized dataset for irrelevance detection, a tuning framework for enhanced generalization, and the Hammer models, establishing a new standard for function-calling performance.", "keywords": [ "language models", "function-calling", "mobile assistant", "tool-using" ], "rating_list": [ 8, 6, 8, 6, 6 ], "soundness_list": [ 4, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 4, 3, 3 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "yVGGtsOgc7", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Aman Bhargava", "gender": "Male", "institution": "California Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Antonio Rangel", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Pantelis Vafidis", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 55, "n_formula_1": 5, "n_ref_uni": 83, "n_ref": 132, "n_ref_all": 222, "n_fig": 39, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 24, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2980, "formula_len_all_1": 483, "len_all": 335048, "len_all_1": 75840, "len_abs": 2121, "len_title": 105, "len_sents": 111820, "len_sents_1": 37553, "n_sents": 842, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 2139, "title": "Disentangling Representations through Multi-task Learning", "abs": "Intelligent perception and interaction with the world hinges on internal representations that capture its underlying structure (\"disentangled\" or \"abstract\" representations). Disentangled representations serve as world models, isolating latent factors of variation in the world along approximately orthogonal directions, thus facilitating feature-based generalization. We provide experimental and theoretical results guaranteeing the emergence of disentangled representations in agents that optimally solve multi-task evidence accumulation classification tasks, canonical in the neuroscience literature. The key conceptual finding is that, by producing accurate multi-task classification estimates, a system implicitly represents a set of coordinates specifying a disentangled representation of the underlying latent state of the data it receives. The theory provides conditions for the emergence of these representations in terms of noise, number of tasks, and evidence accumulation time, when the classification boundaries are affine in the latent space. Surprisingly, the theory also produces closed-form expressions for extracting the disentangled representation from the model's latent state $\\mathbf Z(t)$. We experimentally validate these predictions in RNNs trained on multi-task classification, which learn disentangled representations in the form of continuous attractors, leading to zero-shot out-of-distribution (OOD) generalization in predicting latent factors. We demonstrate the robustness of our framework across autoregressive architectures, decision boundary geometries and in tasks requiring classification confidence estimation. We find that transformers are particularly suited for disentangling representations, which might explain their unique world understanding abilities. Overall, our framework puts forth parallel processing as a principle for the formation of cognitive maps that capture the structure of the world in both biological and artificial systems, and helps explain why ANNs often arrive at human-interpretable concepts, and how they both may acquire exceptional zero-shot generalization capabilities.", "keywords": [ "zero-shot generalization", "disentanglement", "representation learning", "multi-task learning", "interpretability", "computational neuroscience", "evidence accumulation", "world models", "cognitive maps", "continuous attractors", "RNNs", "transformers" ], "rating_list": [ 8, 6, 3, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 4, 2, 3, 3 ], "contribution_list": [ 3, 3, 1, 3, 3 ], "confidence_list": [ 4, 4, 4, 4, 2 ] }, { "paper_id": "yUefexs79U", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Koichi Taniguchi", "gender": "Male", "institution": "Shizuoka University", "country": "JP", "position": "Associate Professor" }, { "name": "Satoshi Okuda", "gender": "unknown", "institution": "Rikkyo University (St. Paul's University)", "country": "JP", "position": "Researcher" }, { "name": "Takashi Furuya", "gender": "Male", "institution": "Shimane University", "country": "JP", "position": "Assistant Professor" } ], "n_formula": 145, "n_formula_1": 43, "n_ref_uni": 50, "n_ref": 87, "n_ref_all": 113, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 17095, "formula_len_all_1": 5021, "len_all": 194734, "len_all_1": 62574, "len_abs": 887, "len_title": 129, "len_sents": 52914, "len_sents_1": 25796, "n_sents": 518, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 889, "title": "Quantitative Approximation for Neural Operators in Nonlinear Parabolic Equations", "abs": "Neural operators serve as universal approximators for general continuous operators. In this paper, we derive the approximation rate of solution operators for the nonlinear parabolic partial differential equations (PDEs), contributing to the quantitative approximation theorem for solution operators of nonlinear PDEs. Our results show that neural operators can efficiently approximate these solution operators without the exponential growth in model complexity, thus strengthening the theoretical foundation of neural operators. A key insight in our proof is to transfer PDEs into the corresponding integral equations via Duahamel's principle, and to leverage the similarity between neural operators and Picard\u2019s iteration\u2014a classical algorithm for solving PDEs. This approach is potentially generalizable beyond parabolic PDEs to a class of PDEs which can be solved by Picard's iteration.", "keywords": [ "Neural operators", "Partial differential equations", "Nonlinear parabolic equations", "Quantitative universal approximation" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "yUC8pU508S", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Beidi Chen", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Tianqi Chen", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Xinyu Yang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 41, "n_ref": 78, "n_ref_all": 117, "n_fig": 14, "n_tab": 6, "L_tab": 2410, "n_element_tab": 215, "n_fig_1": 14, "n_tab_1": 3, "L_tab_1": 721, "n_element_tab_1": 133, "formula_len_all": 711, "formula_len_all_1": 711, "len_all": 136620, "len_all_1": 62686, "len_abs": 1573, "len_title": 130, "len_sents": 31528, "len_sents_1": 28536, "n_sents": 252, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1479, "title": "APE: Faster and Longer Context-Augmented Generation via Adaptive Parallel Encoding", "abs": "Many modern language model applications, such as RAG and in-context learning, require the efficient combination of multiple external contexts to generate a response. Directly incorporating these contexts sequentially presents two challenges: (i) re-encoding each combined selection of contexts for every request creates a significant computational burden. (ii) concatenating selected contexts into a single sequence often exceeds LLM's context window limit. In this work, we explore the promising potential of parallel encoding as a solution to pre-cache the KV states of each context separately, allowing for direct loading and position reuse during inference. However, due to the misalignment of attention distribution, directly applying parallel encoding results in significant performance degradation. To enable accurate and efficient parallel encoding, we propose adaptive parallel encoding, which brings a shared prefix, additional scaling factor, and lower attention temperature to align the distribution of parallel encoding with sequential encoding. Experimental results on both ICL and RAG tasks tasks demonstrate an average improvement of 7.8% over standard parallel encoding. Comparing to sequential encoding, APE enhances performance by 2.9% for long context understanding while preserving 93% accuracy in few-shot learning. Efficiency evaluation demonstrates that APE achieves a 976$\\times$ speedup for a 512K context-augmented generation with a 256-token response.", "keywords": [ "Parallel Encoding; Context-Augmented LLM; Efficient Inference; Context Window Expansion" ], "rating_list": [ 6, 6, 8, 6, 6 ], "soundness_list": [ 4, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 4, 3 ], "contribution_list": [ 3, 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3, 2 ] }, { "paper_id": "yTEwmr1TJb", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Guangqi Jiang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Huanyu Li", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Huazhe Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tao Huang", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "MS student" }, { "name": "Yifei Sun", "gender": "Female", "institution": "Tongji University", "country": "CN", "position": "Undergrad student" }, { "name": "Yongyuan Liang", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 12, "n_ref": 66, "n_ref_all": 107, "n_fig": 13, "n_tab": 5, "L_tab": 1931, "n_element_tab": 268, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 90, "n_element_tab_1": 70, "formula_len_all": 561, "formula_len_all_1": 386, "len_all": 234813, "len_all_1": 69314, "len_abs": 893, "len_title": 47, "len_sents": 52904, "len_sents_1": 33484, "n_sents": 411, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 99, "L_abs": 1632, "title": "Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Dataset", "abs": "The pre-training of visual representations has enhanced the efficiency of robot learning. Due to the lack of large-scale in-domain robotic datasets, prior works utilize in-the-wild human videos to pre-train robotic visual representation. Despite their promising results, representations from human videos are inevitably subject to distribution shifts and lack the dynamics information crucial for task completion. We first evaluate various pre-trained representations in terms of their correlation to the downstream robotic manipulation tasks (i.e., manipulation-centricity). Interestingly, we find that the \u201cmanipulation-centricity\u201d is a strong indicator of success rates when applied to downstream tasks. Drawing from these findings, we propose Robots Pre-train robots with Manipulation centricity (RPM), a foundation representation learning framework capturing both visual features and the dynamics information such as actions and proprioceptions of manipulation tasks to improve manipulation centricity. Specifically, we pre-train a visual encoder on the DROID robotic dataset and leverage motion-relevant data such as robot proprioceptive states and actions. We introduce a novel contrastive loss that aligns visual observations with the robot\u2019s proprioceptive state-action dynamics, combined with an action prediction loss and a time contrastive loss during pre-training. Empirical results across four simulation domains with 20 robotic manipulation tasks demonstrate that RPM outperforms the strongest baseline by 14.8%. Additionally, RPM significantly boosts the success rate in three real-world manipulation tasks by 76.9%.", "keywords": [ "Robot Learning", "Foundation Model", "Representation Learning" ], "rating_list": [ 3, 8, 8, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "ySmovxuDMi", "primary_area": "datasets and benchmarks", "n_author": 1, "author_info_list": [ { "name": "Philipp Dominic Siedler", "gender": "Male", "institution": "Aleph Alpha", "country": "DE", "position": "Researcher" } ], "n_formula": 43, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 101, "n_ref_all": 118, "n_fig": 3, "n_tab": 15, "L_tab": 95772, "n_element_tab": 939, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 3942, "n_element_tab_1": 304, "formula_len_all": 2014, "formula_len_all_1": 0, "len_all": 393007, "len_all_1": 47903, "len_abs": 1261, "len_title": 111, "len_sents": 73812, "len_sents_1": 18188, "n_sents": 620, "n_sents_1": 145, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1177, "title": "HIVEX: A High-Impact Environment Suite for Multi-Agent Research", "abs": "Games have been vital test beds for the rapid development of Agent-based research. Remarkable progress has been achieved in the past, but it is unclear if the findings equip for real-world problems. While pressure grows, some of the most critical ecological challenges can find mitigation and prevention solutions through technology and its applications. Most real-world domains include multi-agent scenarios and require machine-machine and human-machine collaboration. Open-source environments have not advanced and are often toy scenarios, too abstract or not suitable for multi-agent research. By mimicking real-world problems and increasing the complexity of environments, we hope to advance state-of-the-art multi-agent research and inspire researchers to work on immediate real-world problems. Here, we present HIVEX, an environment suite to benchmark multi-agent research focusing on ecological challenges. HIVEX includes the following environments: Wind Farm Control, Wildfire Resource Management, Drone-Based Reforestation, Ocean Plastic Collection, and Aerial Wildfire Suppression. We provide environments, training examples, and baselines for the main and sub-tasks.", "keywords": [ "Environment Benchmark", "Multi-Agent Reinforcement Learning", "Multi-Agent Systems", "Critical Ecological Challenges" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "ySRsm6HDy5", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Adam Wierman", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Professor" }, { "name": "Eric Mazumdar", "gender": "Male", "institution": "Deparment of Computing + Mathematical Sciences, California Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Jingchu Gai", "gender": "Male", "institution": "School of Mathematical Science, Peking University, Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Laixi Shi", "gender": "Female", "institution": "California Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Yuejie Chi", "gender": "Female", "institution": "Meta AI", "country": "US", "position": "Researcher" } ], "n_formula": 369, "n_formula_1": 33, "n_ref_uni": 65, "n_ref": 131, "n_ref_all": 180, "n_fig": 0, "n_tab": 4, "L_tab": 936, "n_element_tab": 91, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 1209, "n_element_tab_1": 46, "formula_len_all": 46989, "formula_len_all_1": 3154, "len_all": 322581, "len_all_1": 74509, "len_abs": 1211, "len_title": 126, "len_sents": 76120, "len_sents_1": 31657, "n_sents": 821, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1216, "title": "Breaking the Curse of Multiagency in Robust Multi-Agent Reinforcement Learning", "abs": "Standard multi-agent reinforcement learning (MARL) algorithms are vulnerable to sim-to-real gaps. To address this, distributionally robust Markov games (RMGs) have been proposed to enhance robustness in MARL by optimizing the worst-case performance when game dynamics shift within a prescribed uncertainty set. Solving RMGs remains under-explored, from problem formulation to the development of sample-efficient algorithms. A notorious yet open challenge is if RMGs can escape the curse of multiagency, where the sample complexity scales exponentially with the number of agents. In this work, we propose a natural class of RMGs where the uncertainty set of each agent is shaped by both the environment and other agents' strategies in a best-response manner. We first establish the well-posedness of these RMGs by proving the existence of game-theoretic solutions such as robust Nash equilibria and coarse correlated equilibria (CCE). Assuming access to a generative model, we then introduce a sample-efficient algorithm for learning the CCE whose sample complexity scales polynomially with all relevant parameters. To the best of our knowledge, this is the first algorithm to break the curse of multiagency for RMGs.", "keywords": [ "Multi-agent reinforcement learning", "Robust Markov games", "Game theory", "Distribution shift" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "ySJSGZxN7M", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Hy Nguyen", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "PhD student" }, { "name": "Kon Mouzakis", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Full Professor" }, { "name": "Nguyen Hung Nguyen", "gender": "Male", "institution": "Swinburne University of Technology", "country": "AU", "position": "Undergrad student" }, { "name": "Nguyen Linh Bao Nguyen", "gender": "Male", "institution": "Swinburne University of Technology", "country": "AU", "position": "Undergrad student" }, { "name": "Rajesh Vasa", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Full Professor" }, { "name": "Srikanth Thudumu", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 5, "n_ref_uni": 8, "n_ref": 17, "n_ref_all": 37, "n_fig": 12, "n_tab": 2, "L_tab": 1070, "n_element_tab": 104, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 250, "n_element_tab_1": 32, "formula_len_all": 214, "formula_len_all_1": 295, "len_all": 84248, "len_all_1": 48393, "len_abs": 1854, "len_title": 119, "len_sents": 33978, "len_sents_1": 23723, "n_sents": 252, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1668, "title": "Dual-Branch HNSW Approach with Skip Bridges and LID-Driven Optimization", "abs": "The Hierarchical Navigable Small World (HNSW) algorithm is widely used for approximate nearest neighbor (ANN) search, leveraging the principles of navigable small-world graphs. However, it faces some limitations. The first is the local optima problem, which arises from the algorithm's greedy search strategy, selecting neighbors based solely on proximity at each step. This often leads to cluster disconnections. The second limitation is that HNSW frequently fails to achieve logarithmic complexity, particularly in high-dimensional datasets, due to the exhaustive traversal through each layer. To address these limitations, we propose a novel algorithm that mitigates local optima and cluster disconnections while improving inference speed. The first component is a dual-branch HNSW structure with LID-based insertion mechanisms, enabling traversal from multiple directions. This improves outlier node capture, enhances cluster connectivity, and reduces the risk of local minima. The second component introduces a bridge-building technique that adds shortcuts between layers, enabling direct jumps and speeding up inference. Experiments on various benchmarks and datasets showed that our algorithm outperforms the original HNSW in both accuracy and speed. We evaluated six datasets across Computer Vision (CV), deep learning (DL), and Natural Language Processing (NLP), showing improvements of 2.5% in NLP, 15% in DL, and up to 35% in CV tasks. Inference speed is also improved by 12% across all datasets. Ablation studies revealed that LID-based insertion had the greatest impact on performance, followed by the dual-branch structure and bridge-building components.", "keywords": [ "Nearest Neighbor Search", "Optimization" ], "rating_list": [ 5, 5, 1 ], "soundness_list": [ 3, 2, 1 ], "presentation_list": [ 3, 2, 1 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 3, 5 ] }, { "paper_id": "yRd4loGAhJ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Changhun Lee", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Eunhyeok Park", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Assistant Professor" }, { "name": "Jun-gyu Jin", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "", "position": "MS student" }, { "name": "YoungHyun Cho", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 22, "n_ref": 44, "n_ref_all": 67, "n_fig": 7, "n_tab": 4, "L_tab": 1648, "n_element_tab": 177, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 48, "n_element_tab_1": 8, "formula_len_all": 610, "formula_len_all_1": 296, "len_all": 119386, "len_all_1": 55612, "len_abs": 1108, "len_title": 111, "len_sents": 38771, "len_sents_1": 27481, "n_sents": 280, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1113, "title": "SEAL: Scaling to Emphasize Attention for Long-Context Retrieval", "abs": "In this work, we introduce a novel approach called Scaling to Emphasize Attention for Long-context retrieval (SEAL), which enhances the retrieval performance of large language models (LLMs) over extended contexts. Previous studies have shown that each attention head in LLMs has a unique functionality and collectively contributes to the overall behavior of the model. Similarly, we observe that specific heads are closely tied to long-context retrieval, showing positive or negative correlation with retrieval scores. Built on this insight, we propose a learning-based mechanism using zero-shot generated data to emphasize these heads, improving the model's performance in long-context retrieval tasks. \nBy applying SEAL, we can achieve significant improvements in in-domain retrieval performance, including document QA tasks from LongBench, and considerable improvements in out-of-domain cases.\nAdditionally, when combined with existing training-free context extension techniques, SEAL extends the context limits of LLMs while maintaining highly reliable outputs, opening new avenues for research in this field.", "keywords": [ "large language models", "long context", "retrieval", "attention", "supervised fine-tuning" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "yRKelogz5i", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Haoxi Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Jie ZHANG", "gender": "Female", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Peiran Dong", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Sikai Bai", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Song Guo", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Xueyang Tang", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "", "position": "PhD student" }, { "name": "Yue Yu", "gender": "unknown", "institution": "University of California, Davis", "country": "US", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 19, "n_ref_uni": 30, "n_ref": 61, "n_ref_all": 99, "n_fig": 9, "n_tab": 6, "L_tab": 1912, "n_element_tab": 261, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2425, "n_element_tab_1": 203, "formula_len_all": 1519, "formula_len_all_1": 1118, "len_all": 146641, "len_all_1": 63941, "len_abs": 1220, "len_title": 122, "len_sents": 52384, "len_sents_1": 27942, "n_sents": 376, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1215, "title": "Causally Motivated Sycophancy Mitigation for Large Language Models", "abs": "Incorporating user preferences into large language models (LLMs) can enhance the personalization and reliability of model outputs and facilitate the application of LLMs to real-world scenarios. However, leveraging user preferences can be a double-edged sword. Recent studies have found that improper utilization can incur sycophancy, where LLMs prioritize alignment with user preferences over the correctness of their outputs. To address sycophancy in LLMs, we analyze and model the problem through the lens of structured causal models (SCMs). We attribute sycophancy to LLMs' reliance on spurious correlations between user preferences and model outputs in this paper. Based on the proposed SCMs, we develop a novel framework to mitigate sycophancy in LLMs by exploiting a significant causal signature. Specifically, we eliminate the spurious correlations embedded in the intermediate layers of LLMs through head reweighting, and then calibrate the intra-head knowledge along the causal representation direction. Extensive experiments are conducted across diverse language tasks, and the empirical results demonstrate the superiority of our method over state-of-the-art competitors in mitigating sycophancy in LLMs.", "keywords": [ "Large Language Model; Sycophancy; Causal Modeling" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "yR47RmND1m", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Anirudh Goyal", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Kenji Kawaguchi", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Presidential Young Professor" }, { "name": "Wenxuan Zhang", "gender": "unknown", "institution": "Alibaba Group", "country": "SG", "position": "Researcher" }, { "name": "Yiran Zhao", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Yuxi Xie", "gender": "Female", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Qizhe Xie", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 8, "n_ref_uni": 48, "n_ref": 98, "n_ref_all": 119, "n_fig": 7, "n_tab": 4, "L_tab": 1065, "n_element_tab": 194, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 984, "n_element_tab_1": 186, "formula_len_all": 875, "formula_len_all_1": 467, "len_all": 137749, "len_all_1": 67246, "len_abs": 1581, "len_title": 110, "len_sents": 32947, "len_sents_1": 29804, "n_sents": 241, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1685, "title": "Identifying and Tuning Safety Neurons in Large Language Models", "abs": "Safety alignment for Large Language Models (LLMs) has become a critical issue due to their rapid progress. However, our understanding of effective safety mechanisms in LLMs remains limited, leading to safety alignment training that mainly focuses on improving optimization, data-level enhancement, or adding extra structures to intentionally block harmful outputs. To address this gap, we develop a neuron detection method to identify safety neurons\u2014those consistently crucial for handling and defending against harmful queries. Our findings reveal that these safety neurons constitute less than $1\\%$ of all parameters, are language-specific and are predominantly located in self-attention layers. Moreover, safety is collectively managed by these neurons in the first several layers. Based on these observations, we introduce a $\\underline{S}$afety $\\underline{N}$euron $\\underline{Tun}$ing method, named $\\texttt{SN-Tune}$, that exclusively tune safety neurons without compromising models' general capabilities. $\\texttt{SN-Tune}$ significantly enhances the safety of instruction-tuned models, notably reducing the harmful scores of Llama3-8B-Instruction from $65.5$ to $2.0$, Mistral-7B-Instruct-v0.2 from $70.8$ to $4.5$, and Vicuna-13B-1.5 from $93.5$ to $3.0$. Moreover, $\\texttt{SN-Tune}$ can be applied to base models on establishing LLMs' safety mechanism, effectively diminishing models' harmful scores from around $100$ to $5.3$, $13.5$, and $13.8$ for LLama2-7B-Base, LLama3-8B-Base, and Mistral-7B-v0.1, respectively. In addition, we improve the LLMs' safety robustness during downstream tasks fine-tuning by separating the safety neurons from models' foundation neurons.", "keywords": [ "Large Language Models", "Alignment", "Safety", "Interpretability", "Neuron Detection" ], "rating_list": [ 3, 8, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 4, 3 ], "confidence_list": [ 5, 4, 4, 3, 2 ] }, { "paper_id": "yQcFniousM", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Elaine Shi", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "", "position": "Associate Professor" }, { "name": "Giulia Fanti", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Mingxun Zhou", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 51, "n_ref": 110, "n_ref_all": 138, "n_fig": 8, "n_tab": 1, "L_tab": 283, "n_element_tab": 12, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1179, "n_element_tab_1": 15, "formula_len_all": 518, "formula_len_all_1": 246, "len_all": 175802, "len_all_1": 72504, "len_abs": 5044, "len_title": 63, "len_sents": 61369, "len_sents_1": 33946, "n_sents": 514, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1356, "title": "Pacmann: Efficient Private Approximate Nearest Neighbor Search", "abs": "We propose a new private Approximate Nearest Neighbor (ANN) search scheme\nnamed Pacmann\nthat allows a client to perform ANN search\nin a vector database \nwithout revealing the query vector to the server.\nUnlike prior constructions that run encrypted search on the server side,\nPacmann carefully offloads limited computation and storage to the client,\nno longer requiring computationally-intensive cryptographic techniques.\nSpecifically, clients run a graph-based ANN search, where in each hop on the graph, the client privately retrieves local graph information from the server. \nTo make this efficient, we combine two ideas: \n(1) we adapt a leading graph-based ANN search algorithm to be compatible with private information retrieval (PIR) for subgraph retrieval;\n(2) we use a recent class of PIR schemes that trade offline preprocessing for online computational efficiency. \nPacmann achieves significantly better search quality than\nthe state-of-the-art private ANN search schemes,\nshowing up to 2.5$\\times$ better search accuracy on \nreal-world datasets than prior work and\nreaching 90\\% quality of a state-of-the-art \nnon-private ANN algorithm.\nMoreover on large datasets with up to 100 million vectors,\nPacmann shows better scalability \nthan prior private ANN schemes\nwith up to 62\\% reduction in computation time\nand 22\\% reduction in overall latency.", "keywords": [ "Information Retrieval", "Privacy" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 2, 4, 3, 2 ] }, { "paper_id": "yPyb2j7oZc", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Chunyu Xuan", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Jing Yang", "gender": "Female", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Shuai Hu", "gender": "Male", "institution": "Novosibirsk State University", "country": "RU", "position": "MS student" }, { "name": "Yazhe Niu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yu Liu", "gender": "Male", "institution": "SenseTime", "country": "HK", "position": "Principal Researcher" }, { "name": "Yuan Pu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 31, "n_formula_1": 12, "n_ref_uni": 28, "n_ref": 75, "n_ref_all": 126, "n_fig": 13, "n_tab": 4, "L_tab": 1468, "n_element_tab": 93, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 734, "n_element_tab_1": 42, "formula_len_all": 2691, "formula_len_all_1": 700, "len_all": 161186, "len_all_1": 60770, "len_abs": 1159, "len_title": 131, "len_sents": 51227, "len_sents_1": 27859, "n_sents": 402, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1165, "title": "ReZero: Boosting MCTS-based Algorithms by Backward-view and Entire-buffer Reanalyze", "abs": "Monte Carlo Tree Search (MCTS)-based algorithms, such as MuZero and its derivatives, have achieved widespread success in various decision-making domains. These algorithms employ the reanalyze process to enhance sample efficiency from stale data, albeit at the expense of significant wall-clock time consumption. To address this issue, we propose a general approach named ReZero to boost tree search operations for MCTS-based algorithms. Specifically, drawing inspiration from the one-armed bandit model, we reanalyze training \nsamples through a backward-view reuse technique which uses the value estimation of a certain child node to save the corresponding sub-tree search time. To further adapt to this design, we periodically reanalyze the entire buffer instead of frequently reanalyzing the mini-batch. The synergy of these two designs can significantly reduce the search cost and meanwhile guarantee or even improve performance, simplifying both data collecting and reanalyzing. Experiments conducted on Atari environments, DMControl suites and board games demonstrate that ReZero substantially improves training speed while maintaining high sample efficiency.", "keywords": [ "Deep Reinforcement Learning", "Monte Carlo tree search", "MuZero", "efficiency optimization", "reanalyze" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 5, 4, 2, 2 ] }, { "paper_id": "yPxhj1FKhG", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "He Wang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jinhui Tang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Longquan Dai", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 23, "n_formula_1": 15, "n_ref_uni": 48, "n_ref": 112, "n_ref_all": 131, "n_fig": 10, "n_tab": 6, "L_tab": 3331, "n_element_tab": 517, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3228, "n_element_tab_1": 363, "formula_len_all": 1165, "formula_len_all_1": 1483, "len_all": 145214, "len_all_1": 53714, "len_abs": 1098, "len_title": 128, "len_sents": 33430, "len_sents_1": 20283, "n_sents": 236, "n_sents_1": 123, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1351, "title": "APCtrl: Adding Conditional Control to Diffusion Models by Alternative Projection", "abs": "Enhancing the versatility of pretrained diffusion models through advanced conditioning techniques is crucial for improving their applicability. We present APCtrl, a novel conditional image generation approach that formulates the latent \\( \\dmrv{z}_\\dms{t} \\) at timestep \\( t \\) as the projection \\( \\dmrv{z}_\\dms{t} = \\text{Proj}_{\\bmfrakD_\\dms{t}} (\\dmrv{z}_{ \\dms{t} + \\dms{1} }) \\) onto the denosing set \\( \\bmfrakD_\\dms{t} \\). For conditional control, APCtrl integrates the condition set \\( \\bmfrakC_\\dms{t} \\), defined by a latent control network \\(\\bmcalA_{\\dmv{theta}}(\\cdot, \\cdot)\\). Our method simplifies conditional sampling to recursive projections \\( \\dmrv{z}_\\dms{t} = \\text{Proj}_{\\bmfrakI_\\dms{t}} \\circ \\text{Proj}_{\\bmfrakD_\\dms{t}} (\\dmrv{z}_{ \\dms{t} + \\dms{1} }) \\), where each projection step integrates both the diffusion and condition priors. By employing Alternative Projection, our approach offers several key advantages: 1. Multi-Condition Generation: easily expandable with additional conditional sets; 2. Model and Sampling Agnosticism: works with any model or sampling method; 3. Unified Control Loss: simplifies the management of diverse control applications; 4. Efficiency: delivers comparable control with reduced training and sampling times. Extensive experiments demonstrate the superior performance of our method.", "keywords": [ "Diffusion Models", "Condition Diffusion", "Alternative Projection", "Control-on-Training", "Control-on-Sampling" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "yP0iKsinmk", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Hongbo Zhou", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Jun Peng", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Postdoc" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Shuheng Zhang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Xiaoshuai Sun", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Yiyi Zhou", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuqi Liu", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 28, "n_ref": 131, "n_ref_all": 141, "n_fig": 7, "n_tab": 3, "L_tab": 422, "n_element_tab": 35, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1113, "n_element_tab_1": 52, "formula_len_all": 644, "formula_len_all_1": 621, "len_all": 124666, "len_all_1": 65002, "len_abs": 117, "len_title": 140, "len_sents": 41101, "len_sents_1": 30316, "n_sents": 304, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1664, "title": "AdaFlow: Efficient Long Video Editing via Adaptive Attention Slimming And Keyframe Selection", "abs": "Text-driven video editing is an emerging research hot spot in deep learning. Despite great progress, long video editing is still notoriously challenging mainly due to excessive memory overhead. To tackle this problem, recent efforts have simplified this task into a two-step process of keyframe translation and interpolation generation, enabling the editing of more frames. However, the token-wise keyframe translation still plagues the upper limit of video length. In this paper, we propose a novel and training-free approach towards efficient and effective long video editing, termed AdaFlow. We first reveal that not all tokens of video frames hold equal importance for keyframe-consistency editing, based on which we propose an Adaptive Attention Slimming scheme for AdaFlow to squeeze the $KV$ sequence of extended self-attention. This enhancement allows AdaFlow to increase the number of keyframes for translations by an order of magnitude. In addition, an Adaptive Keyframe Selection scheme is also equipped to select the representative frames for joint editing, further improving generation quality. With these innovative designs, AdaFlow achieves high-quality long video editing of minutes in one inference, i.e., more than 1$k$ frames on one A800 GPU, which is about ten times longer than the compared methods. To validate AdaFlow, we also build a new benchmark for long video editing with high-quality annotations, termed LongV-EVAL. The experimental results show that our AdaFlow can achieve obvious advantages in both the efficiency and quality of long video editing. Our code is anonymously released at https://anonymous.4open.science/r/AdaFlow-C28F.", "keywords": [ "video editing", "diffusion model", "keyframe selection", "token slimming" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "yOrtDi6IXs", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Amirhossein Roknilamouki", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Arnob Ghosh", "gender": "Male", "institution": "New Jersey Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Eylem Ekici", "gender": "Male", "institution": "Ohio State University", "country": "US", "position": "Full Professor" }, { "name": "Fatemeh Nourzad", "gender": "Female", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Ming Shi", "gender": "unknown", "institution": "State University of New York at Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Ness Shroff", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Full Professor" } ], "n_formula": 143, "n_formula_1": 25, "n_ref_uni": 31, "n_ref": 75, "n_ref_all": 140, "n_fig": 13, "n_tab": 5, "L_tab": 2217, "n_element_tab": 134, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12061, "formula_len_all_1": 1360, "len_all": 221848, "len_all_1": 70984, "len_abs": 1737, "len_title": 143, "len_sents": 66760, "len_sents_1": 31206, "n_sents": 622, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1920, "title": "Provably Efficient Linear Bandits with Instantaneous Constraints in Non-Convex Feature Spaces", "abs": "In linear stochastic bandits, tasks with instantaneous hard constraints present significant challenges, particularly when the feature space is non-convex or discrete. This is especially relevant in applications such as financial management, recommendation systems, and medical treatment selection, where safety constraints appear in non-convex forms or where decisions must often be made within non-convex and discrete sets. In these systems, bandit methods rely on the ability of feature functions to extract critical features. However, in contrast to the star-convexity assumption commonly discussed in the literature, these feature functions often lead to non-convex and more complex feature spaces. In this paper, we investigate linear bandits and introduce a method that operates effectively in a non-convex feature space while satisfying instantaneous hard constraints at each time step. We demonstrate that our method, with high probability, achieves a regret of $\\tilde{\\mathcal{O}}\\big( d (1+\\frac{\\tau}{\\epsilon \\iota}) \\sqrt{T}\\big)$ and meets the instantaneous hard constraints, where $d$ represents the feature space dimension, $T$ the total number of rounds, and $\\tau$ a safety related parameter. The constant parameters $\\epsilon$ and $\\iota$ are related to our localized assumptions around the origin and the optimal point. In contrast, standard safe linear bandit algorithms that rely on the star-convexity assumption often result in linear regret. Furthermore, our approach handles discrete action spaces while maintaining a comparable regret bound. Moreover, we establish an information-theoretic lower bound on the regret of $\\Omega \\left( \\max\\{ d \\sqrt{T}, \\frac{1}{\\epsilon \\iota^2} \\} \\right)$ for $T \\geq \\frac{32 e}{\\epsilon \\iota^2}$, emphasizing the critical role of $\\epsilon$ and $\\iota$ in the regret upper bound. Lastly, we provide numerical results to validate our theoretical findings.", "keywords": [ "Linear Bandits", "Non-convex feature spaces", "Instantaneous hard constraints", "Safety", "UCB" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 3, 2, 3, 4 ] }, { "paper_id": "yOhNLIqTEF", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Han Yu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hao Zou", "gender": "Male", "institution": "ZGC laboratory", "country": "CN", "position": "Researcher" }, { "name": "Haoran Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Jiansheng Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Peng Cui", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Renzhe Xu", "gender": "Male", "institution": "Shanghai University of Finance and Economics", "country": "CN", "position": "Assistant Professor" }, { "name": "Shikai Guan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Xingxuan Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yuan Xue", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 29, "n_formula_1": 13, "n_ref_uni": 23, "n_ref": 50, "n_ref_all": 92, "n_fig": 17, "n_tab": 20, "L_tab": 4941, "n_element_tab": 545, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 384, "n_element_tab_1": 37, "formula_len_all": 4497, "formula_len_all_1": 2139, "len_all": 213363, "len_all_1": 67989, "len_abs": 1566, "len_title": 123, "len_sents": 73000, "len_sents_1": 31165, "n_sents": 533, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1229, "title": "Generalization of Transformers with In-Context Learning: An Empirical Study", "abs": "Large language models (LLMs) like GPT-4 and LLaMA-3 utilize the powerful in-context learning (ICL) capability of Transformer architecture to learn on the fly from limited examples. While ICL underpins many LLM applications, its full potential remains hindered by a limited understanding of its generalization boundaries and vulnerabilities. We present a systematic investigation of transformers' generalization capability with ICL relative to training data coverage by defining a task-centric framework along three dimensions: inter-problem, intra-problem, and intra-task generalization. Through extensive simulation and real-world experiments, encompassing tasks such as function fitting, API calling, and translation, we find that transformers lack inter-problem generalization with ICL, but excel in intra-task and intra-problem generalization. Furthermore, when the training data includes a greater variety of mixed tasks, it significantly enhances the generalization ability of ICL on unseen tasks and even on known simple tasks. This guides us in designing training data to maximize the diversity of tasks covered and to combine different tasks whenever possible, rather than solely focusing on the target task for testing.", "keywords": [ "generalization", "in-context learning", "transformer" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 2, 3, 4 ], "presentation_list": [ 3, 2, 4 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "yORSk4Ycsa", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Amir Zait", "gender": "Male", "institution": "Google", "country": "IL", "position": "Researcher" }, { "name": "Andrew Liu", "gender": "unknown", "institution": "GoogleDeepMind", "country": "US", "position": "Researcher" }, { "name": "Danny Karmon", "gender": "Male", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Gargi Balasubramaniam", "gender": "Female", "institution": "Google DeepMind", "country": "", "position": "Research Engineer" }, { "name": "Henry Prior", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Ilia Labzovsky", "gender": "unknown", "institution": "Google", "country": "IL", "position": "Researcher" }, { "name": "Ishita Dasgupta", "gender": "unknown", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Kenneth Marino", "gender": "Male", "institution": "University of Utah", "country": "US", "position": "Assistant Professor" }, { "name": "Kim Stachenfeld", "gender": "Female", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" }, { "name": "Rivka Moroshko", "gender": "unknown", "institution": "Google", "country": "IL", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 78, "n_ref_all": 97, "n_fig": 6, "n_tab": 2, "L_tab": 1645, "n_element_tab": 56, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 544, "n_element_tab_1": 42, "formula_len_all": 12, "formula_len_all_1": 0, "len_all": 179729, "len_all_1": 66553, "len_abs": 1361, "len_title": 129, "len_sents": 53573, "len_sents_1": 33012, "n_sents": 435, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1395, "title": "ReCogLab: a framework testing relational reasoning, cognitive hypotheses on LLMs", "abs": "A fundamental part of human cognition is the ability to not only recall memories, but to reason and manipulate information from them. In cognitive science and psychology, this is termed relational reasoning or relational memory and a number of cognitive effects and biases have been observed and proposed. Some of these effects include \\textit{congruence}, \\textit{the symbolic distance effect} and \\textit{transitive inference}. In addition, many of these phenomena have been observed in large language models on various handcrafted reasoning benchmarks. While some of these have been studied individually in prior benchmarks for reasoning with, none of these have the flexibility to study all or even most of these hypotheses. In this work, we create a fully customizable, automatically generated dataset which allows us to study these effects in detail. We introduce four settings with multiple cognitive-reasoning-inspired tasks targeting different skills and difficulties with parameters of each of these being configurable to run probes on different abilities. With our framework, we test and find many of these human cognitive effects are repeated in LLMs and provide a number of interesting analyses. We believe our generative framework will help accelerate the testing of various cognitive hypotheses and provide an interesting alternative paradigm for measuring reasoning capabilities.", "keywords": [ "Congitive Science", "Large Language Models", "Datasets", "Evaluation", "Relational Reasoning" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 3, 1, 2, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "yOOJwR15xg", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Jingwei Xu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Junyu Lai", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Yunpeng Huang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 6, "n_ref_uni": 39, "n_ref": 63, "n_ref_all": 84, "n_fig": 5, "n_tab": 8, "L_tab": 11835, "n_element_tab": 336, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2379, "n_element_tab_1": 11, "formula_len_all": 657, "formula_len_all_1": 472, "len_all": 160861, "len_all_1": 65884, "len_abs": 1315, "len_title": 111, "len_sents": 42577, "len_sents_1": 30220, "n_sents": 302, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1320, "title": "MeteoRA: Multiple-tasks Embedded LoRA for Large Language Models", "abs": "The pretrain+fine-tune paradigm is foundational for deploying large language models (LLMs) across various downstream applications. Within this framework, Low-Rank Adaptation (LoRA) stands out for its parameter-efficient fine-tuning (PEFT), producing numerous reusable task-specific LoRA adapters. However, this approach requires explicit task intention selection, posing challenges for autonomous task sensing and switching during inference with multiple existing LoRA adapters embedded in a single LLM. In this work, we introduce MeteoRA (Multiple-Tasks embedded LoRA), a scalable and efficient framework that reuses multiple task-specific LoRA adapters into the base LLM via a full-mode Mixture-of-Experts (MoE) architecture. This framework also includes novel MoE forward acceleration strategies to address the efficiency challenges of traditional MoE implementations. Our evaluation, using the LlaMA2-13B and LlaMA3-8B base models equipped with 28 existing LoRA adapters through MeteoRA, demonstrates equivalent performance with the traditional PEFT method. Moreover, the LLM equipped with MeteoRA achieves superior performance in handling composite tasks, effectively solving ten sequential problems in a single inference pass, thereby demonstrating the framework's enhanced capability for timely adapter switching.", "keywords": [ "LLM", "LoRA" ], "rating_list": [ 6, 6, 6, 5, 8 ], "soundness_list": [ 3, 2, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 2, 3, 4, 4 ] }, { "paper_id": "yNZi38u52U", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 1, "author_info_list": [ { "name": "Gianluca Detommaso", "gender": "Male", "institution": "Helsing", "country": "DE", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 51, "n_ref": 70, "n_ref_all": 84, "n_fig": 3, "n_tab": 1, "L_tab": 3552, "n_element_tab": 164, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 3552, "n_element_tab_1": 164, "formula_len_all": 636, "formula_len_all_1": 636, "len_all": 129379, "len_all_1": 61422, "len_abs": 1030, "len_title": 113, "len_sents": 28855, "len_sents_1": 27248, "n_sents": 204, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1033, "title": "Model Cautiousness: Towards Safer Deployment in Critical Domains", "abs": "In this paper, we introduce the concept of model cautiousness, which stresses the importance of aligning a model's confidence with its accuracy in in-distribution (ID) scenarios while adopting a more uncertain approach in out-of-distribution (OoD) contexts. Model cautiousness is framed as a spectrum between justified confidence and complete ignorance, induced by the inability to clearly define a model's domain of expertise. We propose a rigorous post-hoc approach to obtain a cautious model that merges the confidence scores of the primary confidence model and a model discriminating between ID and OoD inputs. A metric to measure the cautiousness error of a confidence model is introduced. We further present a simple method for discriminating ID from OoD inputs and providing a meaningful confidence estimate that an input is OoD. Finally, we benchmark our approach across 12 question-answering and 37 vision datasets, demonstrating its effectiveness in enhancing model cautiousness compared to standard calibration procedures.", "keywords": [ "cautiousness", "calibration", "out-of-distribution", "safety" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "yMHe9SRvxk", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Ayano Hiranaka", "gender": "unknown", "institution": "University of Southern California", "country": "", "position": "PhD student" }, { "name": "Chieh-Hsin Lai", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Researcher" }, { "name": "Dongjun Kim", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Naoki Murata", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Researcher" }, { "name": "Shang-Fu Chen", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "PhD student" }, { "name": "Shao-Hua Sun", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "Assistant Professor" }, { "name": "Takashi Shibuya", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Senior Manager" }, { "name": "Wei-Hsiang Liao", "gender": "Male", "institution": "Sony Corporation", "country": "", "position": "Research Engineer" }, { "name": "Yuki Mitsufuji", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Lead Research Scientist" } ], "n_formula": 21, "n_formula_1": 7, "n_ref_uni": 33, "n_ref": 84, "n_ref_all": 129, "n_fig": 11, "n_tab": 9, "L_tab": 3078, "n_element_tab": 173, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 393, "n_element_tab_1": 22, "formula_len_all": 1733, "formula_len_all_1": 609, "len_all": 170118, "len_all_1": 68285, "len_abs": 1254, "len_title": 133, "len_sents": 53091, "len_sents_1": 31856, "n_sents": 425, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1360, "title": "HERO: Human-Feedback Efficient Reinforcement Learning for Online Diffusion Model Finetuning", "abs": "Controllable generation through Stable Diffusion (SD) fine-tuning aims to improve fidelity, safety, and alignment with human guidance. Existing reinforcement learning from human feedback methods usually rely on predefined heuristic reward functions or pretrained reward models built on large-scale datasets, limiting their applicability to scenarios where collecting such data is costly or difficult. To effectively and efficiently utilize human feedback, we develop a framework, HERO, which leverages online human feedback collected on the fly during model learning. Specifically, HERO features two key mechanisms: (1) Feedback-Aligned Representation Learning, an online training method that captures human feedback and provides informative learning signals for fine-tuning, and (2) Feedback-Guided Image Generation, which involves generating images from SD's refined initialization samples, enabling faster convergence towards the evaluator's intent. We demonstrate that HERO is 4x more efficient in online feedback for body part anomaly correction compared to the best existing method. Additionally, experiments show that HERO can effectively handle tasks like reasoning, counting, personalization, and reducing NSFW content with only 0.5K online feedback. The code and project page are available at [https://hero-dm.github.io/](https://hero-dm.github.io/).", "keywords": [ "Online RLHF", "Diffusion Model Finetuning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 4, 2, 4 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "yM7rw8Bo1f", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Dong Yang", "gender": "unknown", "institution": "Georgia State University", "country": "US", "position": "PhD student" }, { "name": "HaitaoYang", "gender": "Male", "institution": "Yantai University", "country": "CN", "position": "MS student" }, { "name": "Shenqiang Wang", "gender": "unknown", "institution": "Yantai University", "country": "CN", "position": "MS student" }, { "name": "Wang Kerry", "gender": "Male", "institution": "Inspur Smart City Technology Co, Ltd", "country": "CN", "position": "Researcher" }, { "name": "Yihao Jiang", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Zhaowei Liu", "gender": "Male", "institution": "Yantai University", "country": "CN", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 20, "n_ref_uni": 33, "n_ref": 45, "n_ref_all": 64, "n_fig": 6, "n_tab": 5, "L_tab": 3462, "n_element_tab": 109, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1897, "n_element_tab_1": 92, "formula_len_all": 1072, "formula_len_all_1": 1023, "len_all": 125549, "len_all_1": 50134, "len_abs": 1127, "len_title": 133, "len_sents": 32974, "len_sents_1": 20381, "n_sents": 245, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1175, "title": "FE-GNN: Feature Enhanced Graph Neural Networks for Account Classification in Ethereum", "abs": "Since the birth of the blockchain cryptocurrency trading platform represented by Bitcoin, cryptocurrencies based on blockchain technology have gained widespread attention and accumulated a large amount of transaction data. The analysis of cryptocurrency transactions has become an important research direction with social and economic value, and an important area of blockchain scientific research. Identifying the identity of different cryptocurrency addresses and understanding their behavior is the core challenge to achieve cryptocurrency transaction analysis, otherwise it is difficult to understand blockchain datasets and analyze them with meaningful results. To this end, this paper proposes a blockchain address identity identification method called \\textbf{F}eature \\textbf{E}nhanced \\textbf{G}raph \\textbf{N}eural \\textbf{N}etworks (FE-GNN). Specifically, a transaction graph is constructed based on the collected transaction data, and graph learning techniques based on graph convolutional networks and graph attention networks are used to infer the blockchain address identity. Experimental results show that the FE-GNN algorithm outperforms previous algorithms.", "keywords": [ "Blockchain", "Identity identification", "GNN" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "yLmcYLP3Yd", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Gleb Rodionov", "gender": "unknown", "institution": "Yandex", "country": "", "position": "Researcher" }, { "name": "Liudmila Prokhorenkova", "gender": "Female", "institution": "Yandex", "country": "", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 41, "n_ref": 76, "n_ref_all": 92, "n_fig": 6, "n_tab": 6, "L_tab": 1909, "n_element_tab": 236, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1287, "n_element_tab_1": 179, "formula_len_all": 465, "formula_len_all_1": 465, "len_all": 137313, "len_all_1": 68557, "len_abs": 1132, "len_title": 85, "len_sents": 46253, "len_sents_1": 31838, "n_sents": 319, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 1138, "title": "Discrete Neural Algorithmic Reasoning", "abs": "Neural algorithmic reasoning aims to capture computations with neural networks via learning the models to imitate the execution of classic algorithms. While common architectures are expressive enough to contain the correct model in the weights space, current neural reasoners are struggling to generalize well on out-of-distribution data. On the other hand, classic computations are not affected by distributional shifts as they can be described as transitions between discrete computational states. In this work, we propose to force neural reasoners to maintain the execution trajectory as a combination of finite predefined states. To achieve that, we separate discrete and continuous data flows and describe the interaction between them. Trained with supervision on the algorithm's state transitions, such models are able to perfectly align with the original algorithm. To show this, we evaluate our approach on multiple algorithmic problems and get perfect test scores both in single-task and multitask setups. Moreover, the proposed architectural choice allows us to prove the correctness of the learned algorithms for any test data.", "keywords": [ "neural algorithmic reasoning", "graph neural networks" ], "rating_list": [ 6, 5, 3, 5, 8 ], "soundness_list": [ 2, 2, 2, 2, 4 ], "presentation_list": [ 2, 3, 2, 2, 4 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 2, 4 ] }, { "paper_id": "yLhJYvkKA0", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Chen Wang", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Postdoc" }, { "name": "Chengyuan Deng", "gender": "unknown", "institution": "Rutgers University", "country": "", "position": "PhD student" }, { "name": "Jie Gao", "gender": "unknown", "institution": "Rutgers University", "country": "US", "position": "Full Professor" }, { "name": "Samson Zhou", "gender": "unknown", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" }, { "name": "Jalaj Upadhyay", "gender": "Male", "institution": "Rutgers University", "country": "", "position": "Assistant Professor" } ], "n_formula": 70, "n_formula_1": 11, "n_ref_uni": 39, "n_ref": 113, "n_ref_all": 133, "n_fig": 8, "n_tab": 5, "L_tab": 1374, "n_element_tab": 259, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 211, "n_element_tab_1": 16, "formula_len_all": 6203, "formula_len_all_1": 579, "len_all": 236182, "len_all_1": 70424, "len_abs": 3139, "len_title": 112, "len_sents": 76537, "len_sents_1": 32652, "n_sents": 724, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1699, "title": "On the Price of Differential Privacy for Hierarchical Clustering", "abs": "Hierarchical clustering is a fundamental unsupervised machine learning task with the aim of organizing data into a hierarchy of clusters. Many applications of hierarchical clustering involve sensitive user information, therefore motivating recent studies on differentially private hierarchical clustering under the rigorous framework of Dasgupta's objective. However, it has been shown that any privacy-preserving algorithm under edge-level differential privacy necessarily suffers a large error. To capture practical applications of this problem, we focus on the weight privacy model, where each edge of the input graph is at least unit weight. We present a novel algorithm in the weight privacy model that shows significantly better approximation than known impossibility results in the edge-level DP setting. In particular, our algorithm achieves $O(\\log^{1.5}n/\\varepsilon)$ multiplicative error for $\\varepsilon$-DP and runs in polynomial time, where $n$ is the size of the input graph, and the cost is never worse than the optimal additive error in existing work. We complement our algorithm by showing if the unit-weight constraint does not apply, the lower bound for weight-level DP hierarchical clustering is essentially the same as the edge-level DP, i.e. $\\Omega(n^2/\\varepsilon)$ additive error. As a result, we also obtain a new lower bound of $\\tilde{\\Omega}(1/\\varepsilon)$ additive error for balanced sparsest cuts in the weight-level DP model, which may be of independent interest. Finally, we evaluate our algorithm on synthetic and real-world datasets. Our experimental results show that our algorithm performs well in terms of extra cost and has good scalability to large graphs.", "keywords": [ "Hierarchical clustering", "differential privacy", "sparsest cut" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 4, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "yLYMFRZkdU", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Joseph E. Gonzalez", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Associate Professor" }, { "name": "Justin Wong", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Michael Luo", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Sanjit A. Seshia", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Yury Orlovskiy", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 42, "n_ref": 48, "n_ref_all": 72, "n_fig": 7, "n_tab": 2, "L_tab": 400, "n_element_tab": 38, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 322, "n_element_tab_1": 26, "formula_len_all": 203, "formula_len_all_1": 159, "len_all": 154930, "len_all_1": 63234, "len_abs": 0, "len_title": 169, "len_sents": 38204, "len_sents_1": 30219, "n_sents": 332, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1269, "title": "SimpleStrat: Diversifying Language Model Generation with Stratification", "abs": "Generating diverse responses from large language models (LLMs) is crucial for applications such as planning/search and synthetic data generation, where diversity provides distinct answers across generations.\nPrior approaches rely on increasing temperature to increase diversity. However, contrary to popular belief, we show not only does this approach produce lower quality individual generations as temperature increases, but it depends on model's next-token probabilities being similar to the true distribution of answers. We propose SimpleStrat, an alternative approach that uses the language model itself to partition the space into strata. At inference, a random stratum is selected and a sample drawn from within the strata.\nTo measure diversity, we introduce CoverageQA, a dataset of underspecified questions with multiple equally plausible answers, and assess diversity by measuring KL Divergence between the sampling distribution and uniform distribution over valid ground truth answers. As computing a posterior probability for proprietary models is infeasible, we measure recall on ground truth solutions.\nOur evaluation show using SimpleStrat achieves higher recall by 0.05 compared to GPT-4o and 0.36 average reduction in KL Divergence compared to Llama 3.", "keywords": [ "Diverse Generation; Large Language Models Sampling; Stratified Sampling" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "yJduhi9mDQ", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Bhaskar Ramasubramanian", "gender": "Male", "institution": "Western Washington University", "country": "", "position": "Assistant Professor" }, { "name": "Dinuka Sahabandu", "gender": "Male", "institution": "University of Washington, Seattle", "country": "", "position": "PhD student" }, { "name": "Kaiyuan Zheng", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "MS student" }, { "name": "Radha Poovendran", "gender": "unknown", "institution": "University of Washington, Seattle", "country": "", "position": "Full Professor" }, { "name": "Yifei Zhao", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Yuchen Wu", "gender": "unknown", "institution": "University of Washington", "country": "", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 6, "n_ref_uni": 50, "n_ref": 225, "n_ref_all": 258, "n_fig": 16, "n_tab": 10, "L_tab": 8448, "n_element_tab": 331, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 3257, "n_element_tab_1": 73, "formula_len_all": 1067, "formula_len_all_1": 287, "len_all": 221503, "len_all_1": 61787, "len_abs": 1841, "len_title": 127, "len_sents": 73427, "len_sents_1": 27858, "n_sents": 497, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1849, "title": "H\u00d6LDER PRUNING: LOCALIZED PRUNING FOR BACKDOOR REMOVAL IN DEEP NEURAL NETWORKS", "abs": "Deep Neural Networks (DNNs) have become the cornerstone of modern machine\nlearning applications, achieving impressive results in domains ranging from com-\nputer vision to autonomous systems. However, their dependence on extensive data\nand computational resources exposes them to vulnerabilities such as backdoor\nattacks, where poisoned samples can lead to erroneous model outputs. To counter\nthese threats, we introduce a defense strategy called H\u00f6lder Pruning to detect\nand eliminate neurons affected by triggers embedded in poisoned samples. Our\nmethod partitions the neural network into two stages: feature extraction and feature\nprocessing, aiming to detect and remove backdoored neurons\u2014the highly sensitive\nneurons affected by the embedded triggers\u2014while maintaining model performance\nThis improves model sensitivity to perturbations and enhances pruning precision\nby exploiting the unique clustering properties of poisoned samples. We use the\nH\u00f6lder constant to quantify sensitivity of neurons to input perturbations and prove\nthat using the Fast Gradient Sign Method (FGSM) can effectively identify highly\nsensitive backdoored neurons. Our extensive experiments demonstrate efficacy of\nH\u00f6lder Pruning across six clean feature extractors (SimCLR, Pretrained ResNet-18,\nViT, ALIGN, CLIP, and BLIP-2) and confirm robustness against nine backdoor\nattacks (BadNets, LC, SIG, LF, WaNet, Input-Aware, SSBA, Trojan, BppAttack)\nusing three datasets (CIFAR-10, CIFAR-100, GTSRB). We compare H\u00f6lder Pruning to eight SOTA backdoor defenses (FP, ANP, CLP, FMP, ABL, DBD, D-ST)\nand show that H\u00f6lder Pruning outperforms all eight SOTA methods. Moreover,\nH\u00f6lder Pruning achieves a runtime up to 1000x faster than SOTA defenses when\na clean feature extractor is available. Even when clean feature extractors are not\navailable, our method is up to 10x faster.", "keywords": [ "Holder Pruning", "Holder iteration defense", "backdoor attacks", "Deep Neural Networks", "backdoor defense" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 4, 2, 2, 2 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "yJAk0n0NyU", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Hui Zhang", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Jie Shao", "gender": "Male", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "Tingwei Gao", "gender": "unknown", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Yu-Gang Jiang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Zuxuan Wu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 46, "n_ref": 95, "n_ref_all": 124, "n_fig": 10, "n_tab": 8, "L_tab": 2441, "n_element_tab": 235, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1877, "n_element_tab_1": 193, "formula_len_all": 11012, "formula_len_all_1": 8552, "len_all": 161769, "len_all_1": 82934, "len_abs": 191, "len_title": 199, "len_sents": 47233, "len_sents_1": 37473, "n_sents": 289, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1595, "title": "BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers", "abs": "Diffusion models have demonstrated impressive generation capabilities, particularly with recent advancements leveraging transformer architectures to improve both visual and artistic quality. However, Diffusion Transformers (DiTs) continue to encounter challenges related to low inference speed, primarily due to the iterative denoising process.\nTo address this issue, we propose BlockDance, a training-free approach that explores feature similarities at adjacent time steps to accelerate DiTs.\nUnlike previous feature-reuse methods that lack tailored reuse strategies for features at different scales, BlockDance prioritizes the identification of the most structurally similar features, referred to as Structurally Similar Spatio-Temporal (STSS) features. These features are primarily located within the structure-focused blocks of the transformer during the later stages of denoising.\nBlockDance caches and reuses these highly similar features to mitigate redundant computation, thereby accelerating DiTs while maximizing consistency with the generated results of the original model.\nFurthermore, considering the diversity of generated content and the varying distributions of redundant features, we introduce BlockDance-Ada, a lightweight decision-making network tailored for instance-specific acceleration.\nBlockDance-Ada dynamically allocates resources and provides superior content quality.\nBoth BlockDance and BlockDance-Ada have demonstrated effectiveness across diverse generation tasks and models, achieving an acceleration ranging from 25\\% to 50\\% while preserving generation quality.", "keywords": [ "Diffusion Models", "Efficient Image and Video Generation" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "yJ9QNbpMi2", "primary_area": "applications to neuroscience & cognitive science", "n_author": 7, "author_info_list": [ { "name": "Andrew Luo", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Jacob Yeung", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Leila Wehbe", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Margaret Marie Henderson", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Rushikesh Zawar", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Shaurya Rajat Dewan", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Michael Tarr", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 91, "n_ref": 152, "n_ref_all": 176, "n_fig": 8, "n_tab": 3, "L_tab": 1345, "n_element_tab": 56, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 532, "n_element_tab_1": 49, "formula_len_all": 763, "formula_len_all_1": 766, "len_all": 214260, "len_all_1": 66649, "len_abs": 1957, "len_title": 166, "len_sents": 51756, "len_sents_1": 31747, "n_sents": 380, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 117, "L_abs": 1964, "title": "Brain Mapping with Dense Features: Grounding Cortical Semantic Selectivity in Natural Images With Vision Transformers", "abs": "We introduce BrainSAIL (Semantic Attribution and Image Localization), a method for linking neural selectivity with spatially distributed semantic visual concepts in natural scenes. BrainSAIL leverages recent advances in large-scale artificial neural networks, using them to provide insights into the functional topology of the brain. To overcome the challenge presented by the co-occurrence of multiple categories in natural images, BrainSAIL exploits semantically consistent, dense spatial features from pre-trained vision models, building upon their demonstrated ability to robustly predict neural activity. This method derives clean, spatially dense embeddings without requiring any additional training, and employs a novel denoising process that leverages the semantic consistency of images under random augmentations. By unifying the space of whole-image embeddings and dense visual features and then applying voxel-wise encoding models to these features, we enable the identification of specific subregions of each image which drive selectivity patterns in different areas of the higher visual cortex. This provides a powerful tool for dissecting the neural mechanisms that underlie semantic visual processing for natural images. We validate BrainSAIL on cortical regions with known category selectivity, demonstrating its ability to accurately localize and disentangle selectivity to diverse visual concepts. Next, we demonstrate BrainSAIL's ability to characterize high-level visual selectivity to scene properties and low-level visual features such as depth, luminance, and saturation, providing insights into the encoding of complex visual information. Finally, we use BrainSAIL to directly compare the feature selectivity of different brain encoding models across different regions of interest in visual cortex. Our innovative method paves the way for significant advances in mapping and decomposing high-level visual representations in the human brain.", "keywords": [ "fMRI", "visual cortex", "neuroscience", "cognitive science", "brain", "vision transformer", "semantic selectivity" ], "rating_list": [ 6, 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 4, 3, 3 ], "confidence_list": [ 2, 4, 4, 3, 4 ] }, { "paper_id": "yIlyHJdYV3", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Anuj Karpatne", "gender": "unknown", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "Associate Professor" }, { "name": "Arka Daw", "gender": "Male", "institution": "Oak Ridge National Laboratory", "country": "US", "position": "Researcher" }, { "name": "Medha Sawhney", "gender": "Female", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "PhD student" }, { "name": "Naveen Gupta", "gender": "unknown", "institution": "SLB", "country": "US", "position": "Researcher" }, { "name": "Youzuo Lin", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Associate Professor" } ], "n_formula": 18, "n_formula_1": 6, "n_ref_uni": 21, "n_ref": 37, "n_ref_all": 90, "n_fig": 23, "n_tab": 13, "L_tab": 7607, "n_element_tab": 823, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 398, "n_element_tab_1": 35, "formula_len_all": 1286, "formula_len_all_1": 560, "len_all": 163219, "len_all_1": 60026, "len_abs": 1592, "len_title": 154, "len_sents": 56068, "len_sents_1": 30791, "n_sents": 364, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 106, "L_abs": 1594, "title": "A Unified Framework for Forward and Inverse Problems in Subsurface Imaging using Latent Space Translations", "abs": "In subsurface imaging, learning the mapping from velocity maps to seismic waveforms (forward problem) and waveforms to velocity (inverse problem) is important for several applications. While traditional techniques for solving forward and inverse problems are computationally prohibitive, there is a growing interest to leverage recent advances in deep learning to learn the mapping between velocity maps and seismic waveform images directly from data. Despite the variety of architectures explored in previous works, several open questions still remain unanswered such as the effect of latent space sizes, the importance of manifold learning, the complexity of translation models, and the value of jointly solving forward and inverse problems. We propose a unified framework to systematically characterize prior research in this area termed the Generalized Forward-Inverse (GFI) framework, building on the assumption of manifolds and latent space translations. We show that GFI encompasses previous works in deep learning for subsurface imaging, which can be viewed as specific instantiations of GFI. We also propose two new model architectures within the framework of GFI: Latent U-Net and Invertible X-Net, leveraging the power of U-Nets for domain translation and the ability of IU-Nets to simultaneously learn forward and inverse translations, respectively. We show that our proposed models achieve state-of-the-art (SOTA) performance for forward and inverse problems on a wide range of synthetic datasets, and also investigate their zero-shot effectiveness on two real-world-like datasets.", "keywords": [ "Machine Learning", "Inverse Problems", "Full-Waveform Inversion", "Seismic Imaging", "ML4Science" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "yIdCQFvbYe", "primary_area": "applications to robotics, autonomy, planning", "n_author": 3, "author_info_list": [ { "name": "Karl Henrik Johansson", "gender": "unknown", "institution": "KTH Royal Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Mahmoud Selim", "gender": "unknown", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "PhD student" }, { "name": "Sriharsha Bhat", "gender": "unknown", "institution": "Scania CV AB", "country": "SE", "position": "Researcher" } ], "n_formula": 29, "n_formula_1": 17, "n_ref_uni": 46, "n_ref": 60, "n_ref_all": 85, "n_fig": 12, "n_tab": 11, "L_tab": 2857, "n_element_tab": 195, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 200, "n_element_tab_1": 18, "formula_len_all": 3707, "formula_len_all_1": 884, "len_all": 177102, "len_all_1": 66803, "len_abs": 1414, "len_title": 161, "len_sents": 60439, "len_sents_1": 33348, "n_sents": 471, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1420, "title": "Bayesian Learning of Adaptive Koopman Operator with Application to Robust Motion Planning for Autonomous Trucks", "abs": "Koopman theory has recently been shown to enable an efficient data-driven approach for modeling physical systems, offering a linear framework despite underlying nonlinear dynamics. It is, however, not clear how to account for uncertainty or temporal distributional shifts within this framework, both commonly encountered in real-world autonomous driving with changing weather conditions and time-varying vehicle dynamics. In this work, we introduce BLAK, Bayesian Learning of Adaptive Koopman operator to address these limitations. Specifically, we propose a Bayesian Koopman operator that incorporates uncertainty quantification, enabling more robust predictions. To tackle distributional shifts, we propose an online adaptation mechanism, ensuring the operator remains responsive to changes in system dynamics. Additionally, we apply the architecture to motion planning and show that it gives fast and precise predictions. By leveraging uncertainty awareness and real-time updates, our planner generates dynamically accurate trajectories and makes more informed decisions. We evaluate our method on real-world truck dynamics data under varying weather conditions\u2014such as wet roads, snow, and ice\u2014where uncertainty and dynamic shifts are prominent, as well as in other simulated environments. The results demonstrate our method\u2019s ability to deliver accurate, uncertainty-aware open-loop predictions for dynamic systems.", "keywords": [ "Koopman Theory", "Motion Planning", "Autonomous Systems" ], "rating_list": [ 5, 6, 8, 3, 5 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 2, 3 ] }, { "paper_id": "yIbSXuLoO1", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Hao Qiu", "gender": "Male", "institution": "University of Milan", "country": "IT", "position": "PhD student" }, { "name": "Jingyuan Liu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Undergrad student" }, { "name": "Lin Yang", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "Assistant Professor" }, { "name": "Mohammad Hajiesmaili", "gender": "Male", "institution": "University of Massachusetts Amherst", "country": "US", "position": "Associate Professor" }, { "name": "Xuchuang Wang", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Postdoc" }, { "name": "Xutong Liu", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Postdoc" } ], "n_formula": 109, "n_formula_1": 25, "n_ref_uni": 28, "n_ref": 60, "n_ref_all": 84, "n_fig": 1, "n_tab": 6, "L_tab": 1284, "n_element_tab": 199, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 141, "n_element_tab_1": 11, "formula_len_all": 9817, "formula_len_all_1": 1539, "len_all": 201765, "len_all_1": 70811, "len_abs": 1323, "len_title": 88, "len_sents": 66764, "len_sents_1": 32312, "n_sents": 660, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1565, "title": "Set-Size Dependent Combinatorial Bandits", "abs": "This paper introduces and studies a new variant of Combinatorial Multi-Armed Bandits (\\CMAB{}), called Set-Size Dependent Combinatorial Multi-Armed Bandits (\\SDMAB{}). In \\SDMAB{}, each base arm is associated with a set of different reward distributions instead of a single distribution as in \\CMAB{}, and the reward distribution of each base arm depends on the set size, i.e., the number of the base arms in the chosen super arm in \\CMAB{}. \\SDMAB{} involves a much larger exploration set of the super arms than the basic \\CMAB{} model. An important property called order preservation exists in \\SDMAB{}, i.e. the order of reward means of base arms is independent of set size, which widely exists in real-world applications. We propose the \\SUCB{} algorithm, effectively leveraging the order preservation property to shrink the exploration set. We provide theoretical upper bound of $O\\left(\\max\\left\\{\\frac{M\\delta_L}{\\Delta_{L}},\\frac{L^2}{\\Delta_S}\\right\\}\\log(T)\\right)$ for \\SUCB{} which outperforms the classic \\CMAB{} algorithms with regret $O\\left(\\frac{ML^2}{\\Delta_S}\\log(T)\\right)$, where $M$ denotes the number of base arms, $L$ denotes the maximum number of base arms in a super arm, $\\delta$ and $\\Delta$ are related to the gap of arms. We also derive a lower bound which can be informally written as $\\Omega\\left(\\max\\left\\{\\min_{k\\in[L]}\\left\\{\\frac{(M-L)\\delta_{k}}{\\Delta_{k}^2}\\right\\},\\frac{L^2}{\\Delta_S}\\right\\}\\log(T)\\right)$ showing that \\SUCB{} is partially tight. We conduct numerical experiments, showing the good performance of \\SUCB{}.", "keywords": [ "Combinatorial Multi-armed Bandit", "Set-Size Dependent", "Online learning" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "yIRtu2FJvY", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Antoine Honore", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "Postdoc" }, { "name": "Borja Rodr\u00edguez G\u00e1lvez", "gender": "Male", "institution": "KTH Royal Institute of Technology, Stockholm, Sweden", "country": "", "position": "PhD student" }, { "name": "MingX", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "Full Professor" }, { "name": "Volker M. Lauschke", "gender": "unknown", "institution": "Karolinska Institute Stockholm", "country": "SE", "position": "Full Professor" }, { "name": "Yitian Zhou", "gender": "Male", "institution": "Karolinska Institute", "country": "SE", "position": "Postdoc" }, { "name": "Yoomi Park", "gender": "unknown", "institution": "Karolinska Institute Stockholm", "country": "SE", "position": "Postdoc" } ], "n_formula": 18, "n_formula_1": 17, "n_ref_uni": 33, "n_ref": 49, "n_ref_all": 78, "n_fig": 8, "n_tab": 3, "L_tab": 3791, "n_element_tab": 387, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 901, "n_element_tab_1": 52, "formula_len_all": 1057, "formula_len_all_1": 1139, "len_all": 148809, "len_all_1": 58240, "len_abs": 1482, "len_title": 128, "len_sents": 40754, "len_sents_1": 25497, "n_sents": 312, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1499, "title": "A Matrix Variational Auto-Encoder for Variant Effect Prediction in Pharmacogenes", "abs": "Variant effect predictors (VEPs) are designed to predict the impact of protein variants on cellular function, traditionally using data from multiple sequence alignments (MSAs). This assumes that natural variants are fit, a premise challenged by pharmacogenomics, where some pharmacogenes have low evolutionary pressure. In this context, deep mutational scanning (DMS) datasets are of particular interest since they provide quantitative fitness scores for variants. In this work, we propose a transformer-based matrix variational auto-encoder architecture and evaluate its performances on $33$ DMS datasets corresponding to $26$ drug target and absorption-distribution-metabolism-excretion (ADME) proteins available in the ProteinGym benchmark. Our model trained on MSAs (matVAE-MSA) outperforms a model similar to the widely used VEPs in pharmacogenomics, and sets a new zero-shot prediction benchmark for $2$ proteins related to the Noonan syndrome. We compare matVAE-MSA with matENC-DMS, a model with similar capacity, but trained on DMS data in a 5-fold supervised cross-validation framework. matENC-DMS outperforms matVAE-MSA for $15$ out of $33$ DMS datasets, including all ADME, and certain drug target proteins. Although our models do not outperform the best baseline models, our results help shed new light on the role of evolutionary pressure for the validity of the premise of VEP design. In turn motivating the development of DMS datasets to improve VEPs on pharmacogene-related proteins.", "keywords": [ "variant effect prediction", "variational auto-encoder", "transformer", "deep learning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 2, 1, 1, 4 ], "contribution_list": [ 1, 1, 1, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "yIN4yDCcmo", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Chenwei Lin", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Hanjia Lyu", "gender": "unknown", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Xian Xu", "gender": "Male", "institution": "Fudan University", "country": "", "position": "Full Professor" }, { "name": "Jiebo Luo", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 70, "n_ref_all": 85, "n_fig": 4, "n_tab": 6, "L_tab": 5215, "n_element_tab": 427, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2975, "n_element_tab_1": 235, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 148362, "len_all_1": 56926, "len_abs": 1583, "len_title": 134, "len_sents": 29011, "len_sents_1": 25850, "n_sents": 181, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1543, "title": "INS-MMBench: A Comprehensive Benchmark for Evaluating LVLMs' Performance in Insurance", "abs": "Large Vision-Language Models (LVLMs) have demonstrated outstanding performance in various general multimodal applications such as image recognition and visual reasoning, and have also shown promising potential in specialized domains. However, the application potential of LVLMs in the insurance domain\u2014characterized by rich application scenarios and abundant multimodal data\u2014has not been effectively explored. There is no systematic review of multimodal tasks in the insurance domain, nor a benchmark specifically designed to evaluate the capabilities of LVLMs in insurance. This gap hinders the development of LVLMs within the insurance domain. In this paper, we systematically review and distill multimodal tasks for four representative types of insurance: auto insurance, property insurance, health insurance, and agricultural insurance. We propose INS-MMBench, the first comprehensive LVLMs benchmark tailored for the insurance domain. INS-MMBench comprises a total of 2.2K thoroughly designed multiple-choice questions, covering 12 meta-tasks and 22 fundamental tasks. Furthermore, we evaluate multiple representative LVLMs, including closed-source models such as GPT-4o and open-source models like BLIP-2. This evaluation not only validates the effectiveness of our benchmark but also provides an in-depth performance analysis of current LVLMs on various multimodal tasks in the insurance domain. We hope that INS-MMBench will facilitate the further application of LVLMs in the insurance domain and inspire interdisciplinary development.", "keywords": [ "large vision-language model", "insurance", "multimodal" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "yHj6EunfVQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Akash Kumar", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Yogesh S Rawat", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Zsolt Kira", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 41, "n_ref": 129, "n_ref_all": 170, "n_fig": 14, "n_tab": 13, "L_tab": 1403, "n_element_tab": 199, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 404, "n_element_tab_1": 82, "formula_len_all": 881, "formula_len_all_1": 737, "len_all": 183155, "len_all_1": 69376, "len_abs": 3289, "len_title": 132, "len_sents": 50511, "len_sents_1": 31217, "n_sents": 439, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1300, "title": "Contextual Self-paced Learning for Weakly Supervised Spatio-Temporal Video Grounding", "abs": "In this work, we focus on Weakly Supervised Spatio-Temporal Video Grounding (WSTVG). It is a multimodal task aimed at localizing specific subjects spatio-temporally based on textual queries without bounding box supervision. Motivated by recent advancements in multi-modal foundation models for grounding tasks, we first explore the potential of state-of-the-art object detection models for WSTVG. Despite their robust zero-shot capabilities, our adaptation reveals significant limitations, including inconsistent temporal predictions, inadequate understanding of complex queries, and challenges in adapting to difficult scenarios.\nWe propose CoSPaL (Contextual Self-Paced Learning), a novel approach which is designed to overcome these limitations. CoSPaL integrates three core components: (1) Tubelet Phrase Grounding (TPG), which introduces spatio-temporal prediction by linking textual queries to tubelets; (2) Contextual Referral Grounding (CRG), which improves comprehension of complex queries by extracting contextual information to refine object identification over time; and (3) Self-Paced Scene Understanding (SPS), a training paradigm that progressively increases task difficulty, enabling the model to adapt to complex scenarios by transitioning from coarse to fine-grained understanding.", "keywords": [ "spatio-temporal video grounding", "weakly supervised learning" ], "rating_list": [ 3, 6, 5, 8 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "yHVjncoGSp", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Christopher Brinton", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "David Love", "gender": "unknown", "institution": "Purdue University", "country": "", "position": "Full Professor" }, { "name": "Jacqueline R Malayter", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "PhD student" } ], "n_formula": 40, "n_formula_1": 38, "n_ref_uni": 12, "n_ref": 29, "n_ref_all": 42, "n_fig": 6, "n_tab": 2, "L_tab": 295, "n_element_tab": 33, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 236, "n_element_tab_1": 29, "formula_len_all": 2325, "formula_len_all_1": 2259, "len_all": 86239, "len_all_1": 60817, "len_abs": 3646, "len_title": 97, "len_sents": 31543, "len_sents_1": 26837, "n_sents": 251, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1229, "title": "Deep Learning Aided Broadcast Codes With Feedback", "abs": "Deep learning aided codes have been shown to improve code performance in feedback codes in high noise regimes due to the ability to leverage non-linearity in code design. In the additive white Gaussian broadcast channel (AWGN-BC), the addition of feedback may allow the capacity region to extend far beyond the capacity region of the channel without feedback, enabling higher data rates. On the other hand, there are limited deep-learning aided implementations of broadcast codes. In this work, we extend two classes of deep-learning assisted feedback codes to the AWGN-BC channel; the first being an RNN-based architecture and the second being a lightweight MLP-based architecture. Both codes are trained using a global model, and then they are trained using a more realistic vertical federated learning based framework. We first show that in most cases, using an AWGN-BC code outperforms a linear-based concatenated scheme. Second, we show in some regimes, the lightweight architecture far exceeds the RNN-based code, but in especially unreliable conditions, the RNN-based code dominates. The results show the promise of deep-learning aided broadcast codes in unreliable channels, and future research directions are discussed.", "keywords": [ "Deep Learning", "Wireless Communication", "Feedback Coding", "Error Control Coding", "Federated Learning" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 2, 1, 3, 1 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 5, 5, 3, 3 ] }, { "paper_id": "yGv5GzlBwr", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Daoyu Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Enhong Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Mingyue Cheng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Researcher" }, { "name": "Qi Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Zhiding Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 43, "n_ref": 61, "n_ref_all": 77, "n_fig": 3, "n_tab": 12, "L_tab": 10904, "n_element_tab": 721, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 4803, "n_element_tab_1": 284, "formula_len_all": 829, "formula_len_all_1": 598, "len_all": 144906, "len_all_1": 62939, "len_abs": 1286, "len_title": 140, "len_sents": 38600, "len_sents_1": 25764, "n_sents": 280, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1223, "title": "Diffusion Auto-regressive Transformer for Effective Self-supervised Time Series Forecasting", "abs": "Self-supervised learning has become an essential and popular approach for enhancing time series forecasting, enabling models to learn universal representations from unlabeled data. However, effectively capturing both the global sequence dependence and local detail features within time series data remains challenging. To address this, we propose a novel generative self-supervised method called TimeDART, denoting Diffusion Auto-regressive Transformer for Time series forecasting. In TimeDART, we treat time series patches as basic modeling units. For one thing, we employ an self-attention based Transformer encoder to model the dependencies of inter-patches. For another, we introduce diffusion and denoising mechanisms to capture the locality features of intra-patch. Notably, we design a cross-attention-based denoising decoder that allows for adjustable optimization difficulty in the self-supervised task, facilitating more effective self-supervised pre-training. Extensive experiments demonstrate that TimeDART achieves state-of-the-art fine-tuning performance compared to the most advanced competitive methods in forecasting tasks. Our code is publicly available at https://anonymous.4open.science/r/TimeDART-2024.", "keywords": [ "Self-supervised Learning", "Diffusion Model", "Time Series Forecasting" ], "rating_list": [ 3, 5, 8, 5 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "yGnsH3gQ6U", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Philipp Kraehenbuehl", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Yuanjun Xiong", "gender": "Male", "institution": "Predera.AI", "country": "US", "position": "Researcher" }, { "name": "Yue Zhao", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" } ], "n_formula": 49, "n_formula_1": 23, "n_ref_uni": 76, "n_ref": 152, "n_ref_all": 189, "n_fig": 10, "n_tab": 9, "L_tab": 1995, "n_element_tab": 168, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 2537, "n_element_tab_1": 181, "formula_len_all": 5098, "formula_len_all_1": 1690, "len_all": 192435, "len_all_1": 75128, "len_abs": 1081, "len_title": 111, "len_sents": 43811, "len_sents_1": 30182, "n_sents": 431, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1088, "title": "Image and Video Tokenization with Binary Spherical Quantization", "abs": "We propose a new transformer-based image and video tokenizer with Binary Spherical Quantization (BSQ). BSQ projects the high-dimensional visual embedding to a lower-dimensional hypersphere and then applies binary quantization. BSQ is (1) parameter-efficient without an explicit codebook, (2) scalable to arbitrary token dimensions, and (3) compact: compressing visual data by up to 100\u00d7\nwith minimal distortion. Our tokenizer uses a transformer encoder and decoder with simple block-wise causal masking to support variable-length videos as input. The resulting BSQ-ViT achieves state-of-the-art visual reconstruction quality on image and video reconstruction benchmarks with 2.4\u00d7 throughput compared to the best prior methods. Furthermore, by learning an autoregressive prior for adap-\ntive arithmetic coding, BSQ-ViT achieves comparable visual compression results with commonly used compression standards, e.g. JPEG2000/WebP for images and H.264/H.265 for videos. BSQ-ViT also enables masked language models to achieve competitive image synthesis quality to GAN and diffusion approaches.", "keywords": [ "quantization", "visual compression", "visual generation" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "yG1fW8igzP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Ahmad Beirami", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Ali Etemad", "gender": "Male", "institution": "Queen's University", "country": "CA", "position": "Associate Professor" }, { "name": "Pritam Sarkar", "gender": "Male", "institution": "Vector Institute", "country": "CA", "position": "PhD student" }, { "name": "Sayna Ebrahimi", "gender": "Female", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Sercan O Arik", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Tomas Pfister", "gender": "Male", "institution": "Google", "country": "", "position": "Head of Research @ Cloud AI" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 57, "n_ref": 213, "n_ref_all": 252, "n_fig": 13, "n_tab": 13, "L_tab": 6393, "n_element_tab": 744, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 5736, "n_element_tab_1": 532, "formula_len_all": 952, "formula_len_all_1": 481, "len_all": 304182, "len_all_1": 86404, "len_abs": 1555, "len_title": 121, "len_sents": 77244, "len_sents_1": 33103, "n_sents": 725, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1327, "title": "Data-Augmented Phrase-Level Alignment for Mitigating Object Hallucination", "abs": "Despite their significant advancements, Multimodal Large Language Models\n(MLLMs) often generate factually inaccurate information, referred to as hallucination.\nIn this work, we address object hallucinations in MLLMs, where information\nis generated about an object not present in the input image. We introduce Data-augmented\nPhrase-level Alignment (DPA), a novel loss which can be applied to\ninstruction-tuned off-the-shelf MLLMs to mitigate hallucinations, while preserving\ntheir general vision-language capabilities. To fine-tune MLLMs with DPA, we first\ngenerate a set of 'hallucinated' and 'correct' response pairs through generative data\naugmentation by selectively altering the ground-truth information of the correct\nresponses at a phrase level. The DPA loss is then used to train MLLMs to reduce\nthe likelihood of hallucinated phrases compared to the correct ones. Our thorough\nevaluation on various benchmarks confirms the effectiveness of DPA in mitigating\nhallucination while retaining the out-of-the-box performance of the MLLMs on\ngeneral tasks. For instance, MLLMs finetuned with DPA, which we refer to as Hallucination\nAttenuated Language and Vision Assistant (HALVA), improve F1 by up\nto 13.4% on hallucination visual question-answering and reduce the hallucination\nrate by up to 4.2% on image description tasks.", "keywords": [ "Multimodal LLMs", "Object Hallucination", "Vision-language Models" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "yFGR36PLDJ", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Jan Robine", "gender": "Male", "institution": "Technische Universit\u00e4t Dortmund", "country": "DE", "position": "PhD student" }, { "name": "Marc H\u00f6ftmann", "gender": "Male", "institution": "Technische Universit\u00e4t Dortmund", "country": "", "position": "PhD student" }, { "name": "Stefan Harmeling", "gender": "Unspecified", "institution": "Technische Universit\u00e4t Dortmund", "country": "", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 58, "n_ref": 152, "n_ref_all": 166, "n_fig": 7, "n_tab": 6, "L_tab": 3603, "n_element_tab": 556, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1012, "n_element_tab_1": 145, "formula_len_all": 495, "formula_len_all_1": 396, "len_all": 175059, "len_all_1": 64739, "len_abs": 690, "len_title": 112, "len_sents": 43633, "len_sents_1": 29335, "n_sents": 335, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 692, "title": "Simple, Good, Fast: Self-Supervised World Models Free of Baggage", "abs": "What are the essential components of world models? How far do we get with world models that are not employing RNNs, transformers, discrete representations, and image reconstructions? This paper introduces SGF, a Simple, Good, and Fast world model that uses self-supervised representation learning, captures short-time dependencies through frame and action stacking, and enhances robustness against model errors through data augmentation. We extensively discuss SGF\u2019s connections to established world models, evaluate the building blocks in ablation studies, and demonstrate good performance through quantitative comparisons on the Atari 100k benchmark. The source code will be made available.", "keywords": [ "Reinforcement learning", "World models", "Self-supervised learning", "Atari 100k" ], "rating_list": [ 6, 6, 8, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "yFEqYwgttJ", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Chang D. Yoo", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Tri Ton", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Trung X. Pham", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 26, "n_ref": 51, "n_ref_all": 69, "n_fig": 13, "n_tab": 13, "L_tab": 3345, "n_element_tab": 298, "n_fig_1": 5, "n_tab_1": 9, "L_tab_1": 3732, "n_element_tab_1": 265, "formula_len_all": 697, "formula_len_all_1": 541, "len_all": 136729, "len_all_1": 65050, "len_abs": 5109, "len_title": 153, "len_sents": 42203, "len_sents_1": 25757, "n_sents": 340, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1112, "title": "MDSGen: Fast and Efficient Masked Diffusion Temporal-Aware Transformers for Open-Domain Sound Generation", "abs": "We introduce MDSGen, a novel framework for vision-guided open-domain sound generation optimized for model parameter size, memory consumption, and inference speed. This framework incorporates two key innovations: (1) a redundant video feature removal module that filters out unnecessary visual information, and (2) a temporal-aware masking strategy that leverages temporal context for enhanced audio generation accuracy. In contrast to existing resource-heavy Unet-based models, MDSGen employs denoising masked diffusion transformers, facilitating efficient generation without reliance on pre-trained diffusion models. Evaluated on the benchmark VGGSound dataset, our smallest model (5M parameters) achieves 97.9% alignment accuracy, using 172x fewer parameters, 371% less memory, and offering 36x faster inference than the current 860M-parameter state-of-the-art model (93.9% accuracy). The larger model (131M parameters) reaches nearly 99% accuracy while requiring 6.5x fewer parameters. These results highlight the scalability and effectiveness of our approach. The code is available at https://bit.ly/mdsgen.", "keywords": [ "vision-guided audio generation", "fast inference", "open-domain sound synthesis", "masked diffusion models", "temporal learning", "visual sound source localization", "generative AI" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 4, 2, 4 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "yEwakMNIex", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Hao Xiong", "gender": "Non-Binary", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Jiale Ma", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Wentao Zhao", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Wenzheng Pan", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yang Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 56, "n_formula_1": 17, "n_ref_uni": 52, "n_ref": 146, "n_ref_all": 186, "n_fig": 8, "n_tab": 19, "L_tab": 3554, "n_element_tab": 316, "n_fig_1": 8, "n_tab_1": 8, "L_tab_1": 6796, "n_element_tab_1": 598, "formula_len_all": 4555, "formula_len_all_1": 1263, "len_all": 287628, "len_all_1": 75675, "len_abs": 1813, "len_title": 174, "len_sents": 99687, "len_sents_1": 29578, "n_sents": 732, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1767, "title": "UniCO: On Unified Combinatorial Optimization via Problem Reduction to Matrix-Encoded General TSP", "abs": "Various neural solvers have been devised for combinatorial optimization (CO), which are often tailored for specific problem types, ranging from TSP, CVRP to SAT, etc. Yet, it remains an open question how to achieve universality regarding problem representing and learning with a general framework. This paper first proposes **UniCO**, to unify a set of CO problems by reducing them into the *general* TSP form featured by distance matrices. The applicability of this strategy is dependent on the efficiency of the problem reduction and solution transition procedures, which we show that at least ATSP, HCP, and SAT are readily feasible. The hope is to allow for the effective and even simultaneous use of as many types of CO instances as possible to train a neural TSP solver, and optionally finetune it for specific problem types. In particular, unlike the prevalent TSP benchmarks based on Euclidean instances with 2-D coordinates, our focused domain of general TSP could involve non-metric, asymmetric or discrete distances without explicit node coordinates, which is much less explored in TSP literature while poses new intellectual challenges. Along this direction, we devise two neural TSP solvers with and without supervision to conquer such matrix-formulated input, respectively: 1) **MatPOENet** and 2) **MatDIFFNet**. The former is a reinforcement learning-based sequential model with pseudo one-hot embedding (POE) scheme; and the latter is a Diffusion-based generative model with the mix-noised reference mapping scheme. Extensive experiments on ATSP, 2DTSP, HCP- and SAT-distributed general TSPs demonstrate the strong ability of our approaches towards arbitrary matrix-encoded TSP with structure and size variation. Code and data have been open-sourced.", "keywords": [ "Travelling Salesman Problem", "Neural Combinatorial Optimization" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "yEox25xAED", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Jason Piquenot", "gender": "Male", "institution": "Universit\u00e9 de Rouen - Haute Normandie", "country": "FR", "position": "PhD student" }, { "name": "Jean-Yves RAMEL", "gender": "Male", "institution": "Universit\u00e9 de Savoie", "country": "FR", "position": "Full Professor" }, { "name": "Maxime Berar", "gender": "Male", "institution": "Universit\u00e9 de Rouen Normandie", "country": "FR", "position": "Assistant Professor" }, { "name": "Pierre H\u00e9roux", "gender": "Male", "institution": "Universit\u00e9 de Rouen - Haute Normandie", "country": "", "position": "Assistant Professor" }, { "name": "Romain Raveaux", "gender": "Male", "institution": "Universit\u00e9 Fran\u00e7ois Rabelais de Tours", "country": "", "position": "Associate Professor" }, { "name": "S\u00e9bastien Adam", "gender": "Male", "institution": "Universit\u00e9 de Rouen", "country": "", "position": "Full Professor" } ], "n_formula": 76, "n_formula_1": 6, "n_ref_uni": 31, "n_ref": 59, "n_ref_all": 110, "n_fig": 14, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12641, "formula_len_all_1": 669, "len_all": 179158, "len_all_1": 59207, "len_abs": 3501, "len_title": 118, "len_sents": 53211, "len_sents_1": 26115, "n_sents": 476, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 118, "L_abs": 887, "title": "Grammar Reinforcement Learning: path and cycle counting in graphs with a Context-Free Grammar and Transformer approach", "abs": "This paper presents Grammar Reinforcement Learning (GRL), a reinforcement learning algorithm that uses Monte Carlo Tree Search (MCTS) and a transformer architecture that models a Pushdown Automaton (PDA) within a context-free grammar (CFG) framework. Taking as use case the problem of efficiently counting paths and cycles in graphs, a key challenge in network analysis, computer science, biology, and social sciences, GRL discovers new matrix-based formulas for path/cycle counting that improve computational efficiency by factors of two to six w.r.t state-of-the-art approaches. Our contributions include: (i) a framework for generating transformers that operate within a CFG, (ii) the development of GRL for optimizing formulas within grammatical structures, and (iii) the discovery of novel formulas for graph substructure counting, leading to significant computational improvements.", "keywords": [ "Graph", "Reinforcement Learning", "Grammar", "Cycle Counting" ], "rating_list": [ 6, 8, 8, 5, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4, 4 ] }, { "paper_id": "yEnJvc7ogD", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Anirudh Srinivasan", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Branislav Kveton", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Principal Scientist" }, { "name": "Gaurush Hiranandani", "gender": "Male", "institution": "Typeface Inc", "country": "US", "position": "Researcher" }, { "name": "Nurendra Choudhary", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Sahika Genc", "gender": "Female", "institution": "University of Michigan", "country": "", "position": "PhD student" }, { "name": "Siddartha Devic", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 5, "n_ref_uni": 68, "n_ref": 108, "n_ref_all": 128, "n_fig": 13, "n_tab": 2, "L_tab": 1595, "n_element_tab": 68, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 116, "n_element_tab_1": 8, "formula_len_all": 2712, "formula_len_all_1": 422, "len_all": 178465, "len_all_1": 67154, "len_abs": 981, "len_title": 118, "len_sents": 51139, "len_sents_1": 33333, "n_sents": 382, "n_sents_1": 223, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 975, "title": "An Efficient Plugin Method for Metric Optimization of Black-Box Models", "abs": "Many machine learning algorithms and classifiers are available only via API queries as a ``black-box'' --- that is, the downstream user has no ability to change, re-train, or fine-tune the model on a particular target distribution.\nIndeed, a downstream user may not have any knowledge of the training distribution or performance metric used to construct and optimize the black-box model.\nWe propose a simple and efficient plugin method which takes as input arbitrary multiclass predictions and post-processes them in order to adapt them to a new target distribution, while simultaneously optimizing for a particular metric of the confusion matrix.\nImportantly, the plugin method is \\textit{post-hoc}, does not rely on feature information, and only requires a small number of probabilistic predictions along with their corresponding true label.\nWe empirically demonstrate that plugin has performance competitive with related methods on a variety of tabular and language tasks.", "keywords": [ "optimization", "black box systems", "domain adaptation", "distribution shift", "classification" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "yEPNPbF8E7", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Shuai Yang", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Gunagzhou)", "country": "CN", "position": "PhD student" }, { "name": "Ying-Cong Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Ying Shan", "gender": "Male", "institution": "Tencent AI Lab Center of Visual Computing", "country": "CN", "position": "Director" }, { "name": "Yixiao Ge", "gender": "Female", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Yukang Chen", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yuying Ge", "gender": "Female", "institution": "Tencent ARC Lab", "country": "CN", "position": "Researcher" }, { "name": "LI Yang", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Intern" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 37, "n_ref_all": 55, "n_fig": 7, "n_tab": 6, "L_tab": 1894, "n_element_tab": 109, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 279, "n_element_tab_1": 33, "formula_len_all": 15, "formula_len_all_1": 0, "len_all": 189582, "len_all_1": 47520, "len_abs": 0, "len_title": 119, "len_sents": 53417, "len_sents_1": 21578, "n_sents": 483, "n_sents_1": 164, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1460, "title": "SEED-Story: Multimodal Long Story Generation with Large Language Model", "abs": "With the remarkable advancements in image generation and open-form text generation, the creation of interleaved image-text content has become an increasingly intriguing field. Multimodal story generation, characterized by producing narrative texts and vivid images in an interleaved manner, has emerged as a valuable and practical task with broad applications. However, this task poses significant challenges, as it necessitates the comprehension of the complex interplay between texts and images, and the ability to generate long sequences of coherent, contextually relevant texts and visuals. In this work, we propose SEED-Story, a novel method that leverages a Multimodal Large Language Model (MLLM) to generate extended multimodal stories. Our model, built upon the powerful comprehension capability of MLLM, predicts text tokens as well as visual tokens, which are subsequently processed with an adapted visual de-tokenizer to produce images with consistent characters and styles. We further propose multimodal attention sink mechanism to enable the generation of stories with up to 25 sequences (only 10 for training) in a highly efficient autoregressive manner. Additionally, we present a large-scale and high-resolution dataset named StoryStream for training our model and quantitatively evaluating the task of multimodal story generation in various aspects. \nAll models, codes and datasets are released in https://anonymous.4open.science/r/SEED-Story.", "keywords": [ "LLM", "Story telling", "multi-modal generation" ], "rating_list": [ 6, 6, 3, 3, 3 ], "soundness_list": [ 4, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 4, 3, 2 ], "contribution_list": [ 4, 3, 2, 2, 3 ], "confidence_list": [ 5, 4, 3, 3, 4 ] }, { "paper_id": "yDy9fZXNJV", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Didier Ch\u00e9telat", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CA", "position": "Researcher" }, { "name": "Dimitrios Tsaras", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Reza Moravej", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CA", "position": "Researcher" }, { "name": "Saurabh Bodhe", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CA", "position": "Researcher" }, { "name": "Yingxue Zhang", "gender": "Female", "institution": "Huawei Canada, Huawei Noah's Ark Lab", "country": "CA", "position": "Researcher" }, { "name": "Zhanguang Zhang", "gender": "unknown", "institution": "Huawei Noah's Ark Lab", "country": "CA", "position": "Researcher" }, { "name": "Hui-Ling Zhen", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Principal Researcher" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 34, "n_ref": 69, "n_ref_all": 85, "n_fig": 5, "n_tab": 3, "L_tab": 1924, "n_element_tab": 409, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 491, "n_element_tab_1": 61, "formula_len_all": 279, "formula_len_all_1": 189, "len_all": 128413, "len_all_1": 57300, "len_abs": 1235, "len_title": 138, "len_sents": 32438, "len_sents_1": 28294, "n_sents": 216, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 961, "title": "The Graph's Apprentice: Teaching an LLM Low-Level Knowledge for Circuit Quality Estimation", "abs": "Logic synthesis is a crucial phase in the circuit design process, responsible for transforming hardware description language (HDL) designs into optimized netlists. However, traditional logic synthesis methods are computationally intensive, restricting their iterative use in refining chip designs. Recent advancements in large language models (LLMs), particularly those fine-tuned on programming languages, present a promising alternative. This work proposes augmenting LLMs with predictor networks trained to estimate circuit quality directly from HDL code. To enhance performance, the model is regularized using embeddings from graph neural networks (GNNs) trained on Look-Up Table (LUT) graphs, thereby incorporating lower-level circuit insights. The proposed method demonstrates superior performance compared to existing graph-based RTL-level estimation techniques on the established benchmark OpenABCD, while providing instant feedback on HDL code quality.", "keywords": [ "LLM", "Knowledge Distillation", "Verilog", "Graph Neural Network" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 4, 2, 4 ] }, { "paper_id": "yDlvteYBbF", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Mat\u011bj Zorek", "gender": "Male", "institution": "Czech Technical Univeresity in Prague", "country": "CZ", "position": "PhD student" }, { "name": "Tom\u00e1\u0161 Pevn\u00fd", "gender": "Male", "institution": "Czech Technical University in Prague", "country": "CZ", "position": "Associate Professor" }, { "name": "Vaclav Smidl", "gender": "Male", "institution": "Czech Technical Univeresity in Prague", "country": "CZ", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 15, "n_ref_uni": 65, "n_ref": 106, "n_ref_all": 128, "n_fig": 7, "n_tab": 8, "L_tab": 2565, "n_element_tab": 267, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1873, "n_element_tab_1": 180, "formula_len_all": 2052, "formula_len_all_1": 1334, "len_all": 161632, "len_all_1": 66436, "len_abs": 962, "len_title": 110, "len_sents": 42525, "len_sents_1": 26477, "n_sents": 354, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 966, "title": "Differentiable Distance Between Hierarchically-Structured Data", "abs": "Many machine learning algorithms solving various problems are available for\nmetric spaces. While there are plenty of distances for vector spaces, much\nless exists for structured data (rooted heterogeneous trees) stored in popular\nformats like JSON, XML, ProtoBuffer, MessagePack, etc. This paper\nintroduces the Hierarchically-structured Tree Distance (HTD) designed\nespecially for these data. The HTD distance is modular with differentiable\nparameters weighting the importance of different sub-spaces. This allows\nthe distance to be tailored to a given dataset and task, such as classification,\nclustering, and anomaly detection. The extensive experimental comparison\nshows that distance-based algorithms with the proposed HTD distance\nare competitive to state-of-the-art methods based on neural networks with\norders of magnitude more parameters. Furthermore, we show that HTD is\nmore suited to analyze heterogeneous Graph Neural Networks than Tree\nMover\u2019s Distance.", "keywords": [ "Distance", "Distance function", "Tree-structured data", "Heterogenous Graphs", "JSONs", "Multiple Instance Learning" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "yDICgRUj5s", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Kerem Zaman", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Shashank Srivastava", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 27, "n_ref": 52, "n_ref_all": 66, "n_fig": 7, "n_tab": 2, "L_tab": 431, "n_element_tab": 66, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 277, "formula_len_all_1": 279, "len_all": 127458, "len_all_1": 63654, "len_abs": 1449, "len_title": 97, "len_sents": 36133, "len_sents_1": 30297, "n_sents": 257, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1455, "title": "A Causal Lens for Evaluating Faithfulness Metrics", "abs": "The increasing capabilities of Large Language Models (LLMs) have made natural language explanations a promising alternative to traditional feature attribution methods for model interpretability. However, while these explanations may seem plausible, they can fail to reflect the model's underlying reasoning faithfully. The idea of faithfulness is critical for assessing the alignment between the explanation and the model's true decision-making mechanisms. Although several faithfulness metrics have been proposed, they lack a unified evaluation framework. To address this limitation, we introduce Causal Diagnosticity, a new evaluation framework for comparing faithfulness metrics in natural language explanations. Our framework extends the idea of diagnosticity to the faithfulness metrics for natural language explanations by using model editing to generate faithful and unfaithful explanation pairs. We introduce a benchmark consisting of three tasks: fact-checking, analogy, and object counting, and evaluate a diverse set of faithfulness metrics, including post-hoc explanation-based and chain-of-thought (CoT)-based methods. Our results show that while CC-SHAP significantly outperforms other metrics, there is substantial room for improvement. This work lays the foundation for future research in developing more faithful natural language explanations, highlighting the need for improved metrics and more reliable interpretability methods in LLMs.", "keywords": [ "faithfulness", "diagonsticity", "natural language explanations", "interpretability", "model editing" ], "rating_list": [ 3, 5, 3, 5, 6 ], "soundness_list": [ 2, 2, 1, 3, 3 ], "presentation_list": [ 4, 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "yD7oAhFEtD", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Heshan Liu", "gender": "Male", "institution": "Pony AI", "country": "US", "position": "Researcher" }, { "name": "Junze Yin", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Yingyu Liang", "gender": "Male", "institution": "University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Zhenmei Shi", "gender": "Male", "institution": "Voyage AI", "country": "US", "position": "Researcher" }, { "name": "Zhuoyan Xu", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" } ], "n_formula": 122, "n_formula_1": 22, "n_ref_uni": 57, "n_ref": 112, "n_ref_all": 132, "n_fig": 6, "n_tab": 2, "L_tab": 239, "n_element_tab": 21, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15322, "formula_len_all_1": 2645, "len_all": 329558, "len_all_1": 67434, "len_abs": 1351, "len_title": 150, "len_sents": 74827, "len_sents_1": 26083, "n_sents": 943, "n_sents_1": 280, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1407, "title": "Conv-Basis: A New Paradigm for Efficient Attention Inference and Gradient Computation in Transformers", "abs": "The self-attention mechanism is the key to the success of transformers in recent Large Language Models (LLMs). However, the quadratic computational cost $O(n^2)$ in the input sequence length $n$ is a notorious obstacle for further improvement and scalability in longer contexts. In this work, we leverage the convolution-like structure of attention matrices to develop an efficient approximation method for attention computation using convolution matrices. We propose a $\\mathsf{conv}$ basis system, analogous to the rank basis, and show that any lower triangular matrix can always be decomposed as a sum of structured convolution matrices in this basis. We then design a fast algorithm to approximate the attention matrix via a sum of such $k$ convolution matrices. This allows us to compute the attention {\\it inference} via Fast Fourier Transforms (FFT) in $O(knd \\log n)$ time, where $d$ is the hidden dimension, and thus achieve almost linear time $n^{1+o(1)}$ in the practical scenario where $kd = n^{o(1)}$. Furthermore, the attention {\\it training forward} and {\\it backward gradient} can be computed in $n^{1+o(1)}$ as well. We provide theoretical guarantees on the run time and approximation error and conduct preliminary experiments to evaluate its effectiveness. We hope our new paradigm for accelerating attention computation in transformer models can help their application to longer contexts.", "keywords": [ "Attention Acceleration", "Fast Fourier Transforms", "Gradient Computation" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 2, 2 ] }, { "paper_id": "yD2JMeKumt", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Changqing Zhang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jialong Yang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Undergrad student" }, { "name": "Junfan Li", "gender": "Male", "institution": "Harbin Institute of Technology Shenzhen", "country": "CN", "position": "Postdoc" }, { "name": "Qianli Xu", "gender": "Male", "institution": "Institute for Infocomm Research, A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Qinghua Hu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Professor" }, { "name": "Zongbo Han", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Zheng Shou", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 59, "n_ref": 109, "n_ref_all": 116, "n_fig": 2, "n_tab": 11, "L_tab": 4377, "n_element_tab": 516, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 1599, "n_element_tab_1": 174, "formula_len_all": 680, "formula_len_all_1": 650, "len_all": 172851, "len_all_1": 68120, "len_abs": 1417, "len_title": 67, "len_sents": 44553, "len_sents_1": 30266, "n_sents": 334, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1427, "title": "DOTA: Distributional Test-Time Adaptation of Vision-Language Models", "abs": "Vision-language foundation models (e.g., CLIP) have shown remarkable performance across a wide range of tasks. However, deploying these models may be unreliable when significant distribution gaps exist between the training and test data. The training-free test-time dynamic adapter (TDA) is a promising approach to address this issue by storing representative test samples to guide the classification of subsequent ones. However, TDA only naively maintains a limited number of reference samples in the cache, leading to severe test-time catastrophic forgetting when the cache is updated by dropping samples. In this paper, we propose a simple yet effective method for DistributiOnal Test-time Adaptation (DOTA). Instead of naively memorizing representative test samples, DOTA continually estimates the distributions of test samples, allowing the model to continually adapt to the deployment environment. The test-time posterior probabilities are then computed using the estimated distributions based on Bayes' theorem for adaptation purposes. To further enhance the adaptability on the uncertain samples, we introduce a new human-machine collaboration paradigm which identifies uncertain samples, collects human-feedback, and incorporates it into the DOTA framework. Extensive experiments validate that DOTA enables CLIP to continually learn, resulting in a significant improvement compared to current state-of-the-art methods.", "keywords": [ "Test-time", "uncertainty", "vision-language models" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "yCr55EjC1d", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 8, "author_info_list": [ { "name": "Kaiwen Dong", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Mingxuan Ju", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Neil Shah", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Tong Zhao", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "William Shiao", "gender": "Male", "institution": "University of California, Riverside", "country": "", "position": "PhD student" }, { "name": "Yozen Liu", "gender": "unknown", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Zhichun Guo", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "Postdoc" }, { "name": "Nitesh Chawla", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 69, "n_ref": 148, "n_ref_all": 187, "n_fig": 9, "n_tab": 13, "L_tab": 9284, "n_element_tab": 827, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2823, "n_element_tab_1": 226, "formula_len_all": 958, "formula_len_all_1": 731, "len_all": 223490, "len_all_1": 75713, "len_abs": 1478, "len_title": 100, "len_sents": 56419, "len_sents_1": 32302, "n_sents": 455, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1476, "title": "Node Duplication Improves Cold-start Link Prediction", "abs": "Graph Neural Networks (GNNs) are prominent in graph machine learning and have shown state-of-the-art performance in Link Prediction (LP) tasks. Nonetheless, recent studies show that GNNs struggle to produce good results on low-degree nodes despite their overall strong performance. In practical applications of LP, like recommendation systems, improving performance on low-degree nodes is critical, as it amounts to tackling the cold-start problem of improving the experiences of users with few observed interactions. In this paper, we investigate improving GNNs' LP performance on low-degree nodes while preserving their performance on high-degree nodes and propose a simple yet surprisingly effective augmentation technique called NodeDup. Specifically, NodeDup duplicates low-degree nodes and creates links between nodes and their own duplicates before following the standard supervised LP training scheme. By leveraging a ``multi-view'' perspective for low-degree nodes, NodeDup shows significant LP performance improvements on low-degree nodes without compromising any performance on high-degree nodes. Additionally, as a plug-and-play augmentation module, NodeDup can be easily applied on existing GNNs with very light computational cost. Extensive experiments show that NodeDup achieves 38.49%, 13.34%, and 6.76% improvements on isolated, low-degree, and warm nodes, respectively, on average across all datasets compared to GNNs and state-of-the-art cold-start methods.", "keywords": [ "Graph Neural Network", "Link Prediction", "Cold-start", "Graph Augmentation" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "yCN4yI6zhH", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Huawei Cao", "gender": "Male", "institution": ", Chinese Academy of Sciences", "country": "", "position": "Associate Professor" }, { "name": "Maolei Huang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Ming Dun", "gender": "Female", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "Researcher" }, { "name": "Shuhan Song", "gender": "Male", "institution": "Institute of Computing Technology\uff0cChinese Academy", "country": "", "position": "PhD student" }, { "name": "Xiaochun Ye", "gender": "Male", "institution": "institute of computing technology, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Li Ping", "gender": "Female", "institution": ", Chinese Academy of Sciences", "country": "", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 18, "n_ref_uni": 39, "n_ref": 70, "n_ref_all": 87, "n_fig": 3, "n_tab": 9, "L_tab": 3713, "n_element_tab": 643, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 181, "n_element_tab_1": 10, "formula_len_all": 2114, "formula_len_all_1": 1317, "len_all": 153051, "len_all_1": 61874, "len_abs": 1964, "len_title": 134, "len_sents": 40881, "len_sents_1": 26511, "n_sents": 333, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1873, "title": "GPromptShield: Elevating Resilience in Graph Prompt Tuning Against Adversarial Attacks", "abs": "The paradigm of ``pre-training and prompt fine-tuning\", with its effectiveness and lightweight characteristics, has rapidly spread from the language field to the graph field. Several pioneering studies have designed specialized prompt functions for diverse downstream graph tasks based on various graph pre-training strategies. These prompts concentrate on the compatibility between the pre-training pretext and downstream graph tasks, aiming to bridge the gap between them. However, designing prompts to blindly adapt to downstream tasks based on this concept neglects crucial security issues. By conducting covert attacks on downstream graph data, we find that even when the downstream task data closely matches that of the pre-training tasks, it is still feasible to generate highly misleading prompts using simple deceptive techniques. In this paper, we shift the primary focus of graph prompts from compatibility to vulnerability issues in adversarial attack scenarios. We design a highly extensible shield defense system for the prompts, which enhances their robustness from two perspectives: Direct Handling and Indirect Amplification. When downstream graph data exhibits unreliable biases, the former directly combats invalid information by adding mixed multi-defense prompts to the input graph's feature space, while the latter employs a training strategy that circumvents invalid part and amplifies valid part. We provide a theoretical derivation that proves their feasibility, indicating that unbiased prompts exist under certain conditions on unreliable data. Extensive experiments across various adversarial attack scenarios indicate that the prompts within our shield defense system exhibit enhanced resilience and superiority. Our work explores new perspectives in the field of graph prompts, offering a novel option for downstream robust prompt fine-tuning.", "keywords": [ "pre-training; prompt tuning; robustness; adversarial attacks." ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "yCEf1cJDGh", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Ermis Soumalias", "gender": "Male", "institution": "Department of Informatics, University of Zurich, University of Zurich", "country": "CH", "position": "PhD student" }, { "name": "Michael Curry", "gender": "Male", "institution": "University of Illinois at Chicago", "country": "US", "position": "Assistant Professor" }, { "name": "Sven Seuken", "gender": "Male", "institution": "University of Zurich", "country": "", "position": "Associate Professor" } ], "n_formula": 35, "n_formula_1": 11, "n_ref_uni": 30, "n_ref": 58, "n_ref_all": 115, "n_fig": 14, "n_tab": 5, "L_tab": 540, "n_element_tab": 18, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4389, "formula_len_all_1": 1175, "len_all": 206105, "len_all_1": 72455, "len_abs": 1368, "len_title": 118, "len_sents": 73572, "len_sents_1": 35569, "n_sents": 514, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1315, "title": "Truthful Aggregation of LLMs with an Application to Online Advertising", "abs": "The next frontier of online advertising is revenue generation from LLM-generated content. We consider a setting where advertisers aim to influence the responses of an LLM to align with their interests, while platforms seek to maximize advertiser value and ensure user satisfaction. The challenge is that advertisers may misreport their preferences. To address this, we introduce MOSAIC, an auction mechanism that ensures that truthful reporting is a dominant strategy for advertisers, and which aligns each advertiser\u2019s utility with their contribution to social welfare. Importantly, the mechanism operates without LLM fine-tuning or access to model weights and provably converges to the output of the optimally fine-tuned LLM for the platform\u2019s objective as computational resources increase. Additionally, it can incorporate contextual information about the advertisers, accelerating convergence. Via experiments with a publicly available LLM, we show that MOSAIC significantly boosts advertiser value and platform revenue with low computational overhead. While our motivating application is online advertising, our mechanism can be applied in any setting with monetary transfers, making it a general-purpose solution for truthfully aggregating the preferences of self-interested agents over LLM-generated replies.", "keywords": [ "mechanism design", "llm", "auction", "online advertising" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "yCAigmDGVy", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Haoqi He", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "Yan XIAO", "gender": "Female", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 31, "n_formula_1": 21, "n_ref_uni": 32, "n_ref": 42, "n_ref_all": 55, "n_fig": 0, "n_tab": 2, "L_tab": 865, "n_element_tab": 134, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 439, "n_element_tab_1": 79, "formula_len_all": 1197, "formula_len_all_1": 868, "len_all": 117338, "len_all_1": 64406, "len_abs": 1263, "len_title": 155, "len_sents": 37269, "len_sents_1": 29468, "n_sents": 292, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1269, "title": "HiQ-Lip: A Quantum-Classical Hierarchical Method for Global Lipschitz Constant Estimation of ReLU Networks", "abs": "Estimating the global Lipschitz constant of neural networks is crucial for understanding and improving their robustness and generalization capabilities. However, precise calculations are NP-hard, and current semidefinite programming (SDP) methods face challenges such as high memory usage and slow processing speeds. In this paper, we propose $\\textbf{HiQ-Lip}$, a hybrid quantum-classical hierarchical method that leverages Coherent Ising Machines (CIMs) to estimate the global Lipschitz constant. \nWe tackle the estimation by converting it into a Quadratic Unconstrained Binary Optimization (QUBO) problem and implement a multilevel graph coarsening and refinement strategy to adapt to the constraints of contemporary quantum hardware. \nOur experimental evaluations on fully connected neural networks demonstrate that HiQ-Lip not only provides estimates comparable to state-of-the-art methods but also significantly accelerates the computation process. \nIn specific tests involving two-layer neural networks with 256 hidden neurons, HiQ-Lip doubles the solving speed and offers more accurate upper bounds than the existing best method, LiPopt.\nThese findings highlight the promising utility of small-scale quantum devices in advancing the estimation of neural network", "keywords": [ "Quantum Computing", "Lipschitz Constant", "Neural Network", "Quantum-Classical Hybrid Method", "Coherent Ising Machine", "QUBO" ], "rating_list": [ 3, 6, 3, 5, 5 ], "soundness_list": [ 2, 4, 2, 2, 3 ], "presentation_list": [ 3, 4, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "yBlVlS2Fd9", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 17, "author_info_list": [ { "name": "Minghui Fang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Qian Chen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Qian Yang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Rongjie Huang", "gender": "Male", "institution": "FAIR", "country": "US", "position": "Researcher" }, { "name": "Ruiqi Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Shengpeng Ji", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Siqi Zheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Wen Wang", "gender": "unknown", "institution": "Alibaba Group", "country": "", "position": "Senior Staff Algorithm Engineer" }, { "name": "Xiaoda Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Xize Cheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yidi Jiang", "gender": "Female", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Yifu Chen", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Zehan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Ziang Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Ziyue Jiang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jialung Zuo", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 50, "n_ref": 84, "n_ref_all": 111, "n_fig": 2, "n_tab": 15, "L_tab": 4977, "n_element_tab": 726, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 4168, "n_element_tab_1": 509, "formula_len_all": 334, "formula_len_all_1": 334, "len_all": 189631, "len_all_1": 72326, "len_abs": 1550, "len_title": 136, "len_sents": 45046, "len_sents_1": 30915, "n_sents": 309, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1520, "title": "WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling", "abs": "Language models have been effectively applied to modeling natural signals, such as images, video, speech, and audio. A crucial component of these models is the codec tokenizer, which compresses high-dimensional natural signals into lower-dimensional discrete tokens. In this paper, we introduce WavTokenizer, which offers several advantages over previous SOTA acoustic codec models in the audio domain: 1) extreme compression. By compressing the layers of quantizers and the temporal dimension of the discrete codec, one-second audio of 24kHz sampling rate requires only a single quantizer with 40 or 75 tokens. 2) improved subjective quality. Despite the reduced number of tokens, WavTokenizer achieves state-of-the-art reconstruction quality with outstanding UTMOS scores and inherently contains richer semantic information. Specifically, we achieve these results by designing a broader VQ space, extended contextual windows, and improved attention networks, as well as introducing a powerful multi-scale discriminator and an inverse Fourier transform structure. We conducted extensive reconstruction experiments in the domains of speech, audio, and music. WavTokenizer exhibited strong performance across various objective and subjective metrics compared to state-of-the-art models. We also tested semantic information, VQ utilization, and adaptability to generative models. Comprehensive ablation studies confirm the necessity of each module in WavTokenizer. The demo is available at https://wavtokenizer.github.io/.", "keywords": [ "speech representation", "discrete codec", "audio language model" ], "rating_list": [ 5, 8, 3, 10 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 4, 1, 3 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "yBhSORdXqq", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Chun Hei Yip", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Undergrad student" }, { "name": "Jason Gross", "gender": "Male", "institution": "Independent", "country": "US", "position": "Researcher" }, { "name": "Lawrence Chan", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Rajashree Agrawal", "gender": "Female", "institution": "Reed College", "country": "US", "position": "Undergrad student" } ], "n_formula": 67, "n_formula_1": 25, "n_ref_uni": 11, "n_ref": 37, "n_ref_all": 68, "n_fig": 13, "n_tab": 4, "L_tab": 3958, "n_element_tab": 85, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 550, "n_element_tab_1": 26, "formula_len_all": 7758, "formula_len_all_1": 2135, "len_all": 129937, "len_all_1": 56592, "len_abs": 1035, "len_title": 123, "len_sents": 39708, "len_sents_1": 22685, "n_sents": 325, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1085, "title": "Modular addition without black-boxes: Compressing explanations of MLPs that compute numerical integration", "abs": "The goal of mechanistic interpretability is discovering a simple, low-rank algorithm implemented by models.\nWhile we can compress activations into features, compressing nonlinear feature-maps---like MLP layers---is an open problem.\nIn this work, we present the first case study in rigorously compressing nonlinear feature-maps.\nWe work in the classic setting of the modular addition models (Nanda et al., 2023), and target a non-vacuous bound on the behavior of the ReLU MLP in time linear in the parameter-count of the circuit.\nTo study the ReLU MLP analytically, we use the infinite-width lens, which turns post-activation matrix multiplications into approximate integrals.\nWe discover a novel interpretation of the MLP layer in one-layer transformers implementing the \u201cpizza\u201d algorithm (Zhong et al., 2023): the MLP can be understood as evaluating a quadrature scheme, where each neuron computes the area of a rectangle under the curve of a trigonometric integral identity.\nOur code is available at [https://tinyurl.com/mod-add-integration](https://tinyurl.com/mod-add-integration).", "keywords": [ "mechanistic interpretability", "proof", "guarantees", "interpretability", "numerical integration" ], "rating_list": [ 6, 3, 8, 5 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 2, 5, 4, 4 ] }, { "paper_id": "yBLBls6ryd", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Felix Dangel", "gender": "Male", "institution": "Vector Institute, Toronto", "country": "CA", "position": "Postdoc" }, { "name": "Juhan Bae", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Richard E. Turner", "gender": "Male", "institution": "Alan Turing Institute", "country": "GB", "position": "Researcher" }, { "name": "Roger Baker Grosse", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "Associate Professor" }, { "name": "Runa Eschenhagen", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Wu Lin", "gender": "Male", "institution": "Vector Institute", "country": "CA", "position": "Postdoc" } ], "n_formula": 51, "n_formula_1": 37, "n_ref_uni": 39, "n_ref": 77, "n_ref_all": 106, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4770, "formula_len_all_1": 2977, "len_all": 148356, "len_all_1": 79223, "len_abs": 1273, "len_title": 128, "len_sents": 47240, "len_sents_1": 34618, "n_sents": 413, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1304, "title": "Fast Fractional Natural Gradient Descent using Learnable Spectral Factorizations", "abs": "Many popular optimization methods can be united through fractional natural gradient descent (FNGD), which pre-conditions the gradient with a fractional power of the inverse Fisher:\n RMSprop and Adam(W) estimate a diagonal Fisher matrix and apply a square root before inversion; other methods like K-FAC and Shampoo employ matrix-valued Fisher estimates and apply the inverse and inverse square root, respectively.\n Recently, the question of how fractional power affects optimization has moved into focus, e.g. offering trade-offs between convergence and generalization.\n Gaining deeper insights into this phenomenon would require going beyond diagonal estimations and using cheap and flexible matrix-valued Fisher estimators capable of applying any fractional power; however, existing methods are limited by their expensive matrix fraction computation.\n To address this, we propose a Riemannian framework to learn eigen-factorized Fisher estimations on the fly, allowing for the cheap application of *arbitrary* fractional powers.\n Our approach does not require matrix decompositions and, therefore, is stable in half precision.\n We show our framework's efficacy on positive-definite matrix optimization problems and demonstrate its efficiency and flexibility for training neural nets.", "keywords": [ "natural gradient", "Riemannian optimization", "positive-definite manifold", "Kronecker-facotrized", "Shampoo" ], "rating_list": [ 5, 8, 3, 5, 5 ], "soundness_list": [ 3, 4, 3, 2, 3 ], "presentation_list": [ 2, 3, 1, 2, 2 ], "contribution_list": [ 2, 4, 2, 3, 3 ], "confidence_list": [ 3, 4, 2, 2, 4 ] }, { "paper_id": "yAzN4tz7oI", "primary_area": "applications to robotics, autonomy, planning", "n_author": 9, "author_info_list": [ { "name": "Hengkai Tan", "gender": "Male", "institution": "the Department of Computer Science, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Huayu Chen", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jun Zhu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Ke Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Lingxuan Wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Songming Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhengyi Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "BANGGUO LI", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Hang Su", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 67, "n_ref": 165, "n_ref_all": 212, "n_fig": 13, "n_tab": 2, "L_tab": 137, "n_element_tab": 24, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 825, "n_element_tab_1": 47, "formula_len_all": 433, "formula_len_all_1": 361, "len_all": 302339, "len_all_1": 70676, "len_abs": 1959, "len_title": 110, "len_sents": 60941, "len_sents_1": 33265, "n_sents": 517, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1645, "title": "RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation", "abs": "Bimanual manipulation is essential in robotics, yet developing foundation models is extremely challenging due to the inherent complexity of coordinating two robot arms (leading to multi-modal action distributions) and the scarcity of training data. In this paper, we present the Robotics Diffusion Transformer (RDT), a pioneering diffusion foundation model for bimanual manipulation. It is built on scalable Diffusion Transformers (DiTs), which can effectively represent multi-modality, with innovative designs to deal with the heterogeneity of multi-modal inputs and to capture the nonlinearity and high frequency of robotic data. To address data scarcity, we first introduce a Physically Interpretable Unified Action Space, which can unify the action representations of various robots while preserving the physical meanings of original actions, facilitating learning transferrable physical knowledge. With the above designs, we managed to pre-train RDT on the largest collection of multi-robot datasets to date and scaled it up to $1.2$B parameters, which is the largest diffusion-based foundation model for robotic manipulation. We further fine-tuned RDT on a self-created multi-task bimanual dataset with over $6$K+ episodes to refine its manipulation capabilities. Experiments on real robots demonstrate that RDT significantly outperforms existing methods. It exhibits zero-shot generalization to unseen objects and scenes, understands and follows language instructions, learns new skills with just 1$\\sim$5 demonstrations, and effectively handles complex, dexterous tasks. Code and a Demo video are provided in the supplementary materials.", "keywords": [ "robot learning", "diffusion models", "foundation models", "bimanual manipulation" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 4, 3, 4 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "yAU5X77S06", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Fragkiskos D. Malliaros", "gender": "Male", "institution": "CentraleSup\u00e9lec, Inria, Paris-Saclay University", "country": "FR", "position": "Associate Professor" }, { "name": "Johannes F. Lutzeyer", "gender": "Male", "institution": "Ecole Polytechique", "country": "FR", "position": "Assistant Professor" }, { "name": "Michalis Vazirgiannis", "gender": "Male", "institution": "Ecole Polytechnique, France", "country": "", "position": "Full Professor" }, { "name": "Yassine ABBAHADDOU", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "FR", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 22, "n_ref_uni": 30, "n_ref": 49, "n_ref_all": 64, "n_fig": 3, "n_tab": 15, "L_tab": 7385, "n_element_tab": 524, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1067, "n_element_tab_1": 53, "formula_len_all": 1108, "formula_len_all_1": 1098, "len_all": 119003, "len_all_1": 65446, "len_abs": 924, "len_title": 92, "len_sents": 28828, "len_sents_1": 25282, "n_sents": 234, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 930, "title": "ADMP-GNN: Adaptive Depth Message Passing GNN", "abs": "Graph Neural Networks (GNNs) have proven to be highly effective in various graph representation learning tasks. A key characteristic is that GNNs apply a fixed number of message-passing steps to all nodes in the graph, regardless of the varying computational needs and characteristics of each node. Through empirical analysis of real-world data, we show that the optimal number of message-passing layers differs for nodes with different characteristics. This insight is further validated with experiments on synthetic datasets. To address this, we propose Adaptive Depth Message Passing GNN (ADMP-GNN), a novel framework that dynamically adjusts the number of message-passing layers for each node, leading to enhanced performance. This approach is applicable to any model that follows the message-passing scheme. We evaluate ADMP-GNN on the node classification task and observe performance improvements over a wide range of GNNs.", "keywords": [ "Graph Neural Networks" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 1 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "yAN2oPHs7y", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Jilles Vreeken", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Tenured Faculty" }, { "name": "Nils Philipp Walter", "gender": "Male", "institution": "CISPA, saarland university, saarland informatics campus", "country": "DE", "position": "PhD student" }, { "name": "Sascha Xu", "gender": "Male", "institution": "CISPA, saarland university, saarland informatics campus", "country": "DE", "position": "PhD student" } ], "n_formula": 47, "n_formula_1": 17, "n_ref_uni": 29, "n_ref": 43, "n_ref_all": 65, "n_fig": 11, "n_tab": 6, "L_tab": 5070, "n_element_tab": 284, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 1941, "n_element_tab_1": 71, "formula_len_all": 4143, "formula_len_all_1": 1191, "len_all": 141267, "len_all_1": 69120, "len_abs": 1349, "len_title": 73, "len_sents": 47904, "len_sents_1": 30712, "n_sents": 401, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 25, "L_abs": 1412, "title": "Neuro-Symbolic Rule Lists", "abs": "Machine learning models deployed in sensitive areas such as healthcare must be interpretable to ensure accountability and fairness.\nRule lists (_**If**_ $\\texttt{Age} < 35 \\wedge \\texttt{Priors} > 0$ _**then**_ $\\texttt{Recidivism} = $True, _**else if**_ \"Next Condition\" ...)\noffer full transparency, making them well-suited for high-stakes decisions.\nHowever, learning such rule lists presents significant challenges. Existing methods based on combinatorial optimization require feature pre-discretization and impose restrictions on rule size. Neuro-symbolic methods use more scalable continuous optimization yet place similar pre-discretization constraints and suffer from unstable optimization. To address the existing limitations, we introduce NyRules, an end-to-end trainable model that unifies discretization, rule learning, and rule order into a single differentiable framework. \nWe formulate a continuous relaxation of the rule list learning problem that converges to a strict rule list through temperature annealing.\nNyRules learns both the discretizations of individual features, as well as their combination into conjunctive rules without any pre-processing or restrictions.\nExtensive experiments demonstrate that NyRules consistently outperforms both combinatorial and neuro-symbolic methods,\neffectively learning simple and complex rules, as well as their order, across a wide range of datasets.", "keywords": [ "Neuro-Symbolic;Rule Induction; Intepretability" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "y9tQNJ2n1y", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Guangzhi Sun", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Junior Research Fellow" }, { "name": "Jose M. Such", "gender": "unknown", "institution": "Universidad Polit\u00e9cnica de Valencia", "country": "ES", "position": "Full Professor" }, { "name": "Phil Woodland", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Full Professor" }, { "name": "Nicole Zhan", "gender": "Female", "institution": "King's College London, University of London", "country": "GB", "position": "Postdoc" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 50, "n_ref": 93, "n_ref_all": 120, "n_fig": 11, "n_tab": 3, "L_tab": 1466, "n_element_tab": 48, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 81, "n_element_tab_1": 9, "formula_len_all": 53, "formula_len_all_1": 81, "len_all": 168588, "len_all_1": 60144, "len_abs": 1418, "len_title": 128, "len_sents": 56870, "len_sents_1": 30113, "n_sents": 385, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1422, "title": "CASE-Bench: Context-Aware Safety Evaluation Benchmark for Large Language Models", "abs": "Aligning large language models (LLMs) with human values is essential for their safe deployment and widespread adoption. Current LLM safety benchmarks often focus solely on the refusal of individual problematic queries, which overlooks the importance of the context where the query occurs and may cause undesired refusal of queries under safe contexts that diminish user experience. Addressing this gap, we introduce CASE-Bench, a Context-Aware Safety Evaluation Benchmark that integrates context into safety assessments of LLMs. CASE-Bench assigns distinct, formally described contexts to categorized queries based on Contextual Integrity theory. Additionally, in contrast to previous studies which mainly rely on majority voting from just a few annotators, we recruited a sufficient number of annotators necessary to ensure the detection of statistically significant differences among the experimental conditions based on power analysis. Our extensive analysis using CASE-Bench on various open-source and commercial LLMs reveals a substantial and significant influence of context on human judgments ($p<$0.0001 from a z-test), underscoring the necessity of context in safety evaluations. We also identify notable mismatches between human judgments and LLM responses, particularly in commercial models within safe contexts. Code and data used in the paper are available at https://anonymous.4open.science/r/CASEBench-D5DB.", "keywords": [ "safety", "benchmark", "context", "large language model", "contextual integrity" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "y9e1tcWlme", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Florian Marquardt", "gender": "Male", "institution": "Max Planck Institute for the Science of Light", "country": "DE", "position": "Principal Researcher" }, { "name": "Jan Olle", "gender": "Male", "institution": "Max Planck Institute for the Science of Light", "country": "DE", "position": "Postdoc" }, { "name": "Maximilian N\u00e4gele", "gender": "Male", "institution": "Max-Planck Institute for the Science of Light", "country": "DE", "position": "PhD student" }, { "name": "Remmy Zen", "gender": "Male", "institution": "Max Planck Institute for the Science of Light", "country": "DE", "position": "Postdoc" }, { "name": "Thomas F\u00f6sel", "gender": "unknown", "institution": "Friedrich-Alexander Universit\u00e4t Erlangen-N\u00fcrnberg", "country": "", "position": "Postdoc" } ], "n_formula": 32, "n_formula_1": 14, "n_ref_uni": 28, "n_ref": 65, "n_ref_all": 114, "n_fig": 11, "n_tab": 7, "L_tab": 1139, "n_element_tab": 93, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 959, "n_element_tab_1": 77, "formula_len_all": 2921, "formula_len_all_1": 1166, "len_all": 147427, "len_all_1": 61065, "len_abs": 1154, "len_title": 87, "len_sents": 43470, "len_sents_1": 28001, "n_sents": 376, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1159, "title": "Tackling Decision Processes with Non-Cumulative Objectives using Reinforcement Learning", "abs": "Markov decision processes (MDPs) are used to model a wide variety of applications ranging from game playing over robotics to finance. Their optimal policy typically maximizes the expected sum of rewards given at each step of the decision process. However, a large class of problems does not fit straightforwardly into this framework: Non-cumulative Markov decision processes (NCMDPs), where instead of the expected sum of rewards, the expected value of an arbitrary function of the rewards is maximized. Example functions include the maximum of the rewards or their mean divided by their standard deviation. In this work, we introduce a general mapping of NCMDPs to standard MDPs. This allows all techniques developed to find optimal policies for MDPs, such as reinforcement learning or dynamic programming, to be directly applied to the larger class of NCMDPs. Focusing on reinforcement learning, we show applications in a diverse set of tasks, including classical control, portfolio optimization in finance, and discrete optimization problems. Given our approach, we can improve both final performance and training time compared to relying on standard MDPs.", "keywords": [ "reinforcement learning", "markov decision processes", "discrete optimization" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "y9Xp9NozPR", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Gilad Yehudai", "gender": "Male", "institution": "New York University", "country": "", "position": "Postdoc" }, { "name": "Joan Bruna", "gender": "Male", "institution": "New York University", "country": "", "position": "" }, { "name": "Noah Amsel", "gender": "Male", "institution": "NYU, New York University", "country": "US", "position": "PhD student" } ], "n_formula": 262, "n_formula_1": 18, "n_ref_uni": 47, "n_ref": 98, "n_ref_all": 142, "n_fig": 8, "n_tab": 2, "L_tab": 916, "n_element_tab": 109, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 22317, "formula_len_all_1": 999, "len_all": 371590, "len_all_1": 71581, "len_abs": 1150, "len_title": 84, "len_sents": 102073, "len_sents_1": 34105, "n_sents": 1113, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 36, "L_abs": 1155, "title": "The Low-Rank Bottleneck in Attention", "abs": "Attention-based mechanisms are widely used in machine learning, most prominently in transformers. However, hyperparameters such as the rank of the attention matrices and the number of attention heads are scaled nearly the same way in all realizations of this architecture, without theoretical justification. In this paper, we prove that the rank can have a dramatic effect on the representational capacity of attention. This effect persists even when the number of heads and the parameter count are very large. Specifically, we present a simple and natural target function based on nearest neighbor search that can be represented using a single full-rank attention head for any context length, but that cannot be approximated by low-rank attention unless the number of heads is exponential in the embedding dimension, even for short context lengths. Moreover, we show that, for short context lengths, adding depth allows the target to be approximated by low-rank attention. For long contexts, we conjecture that full-rank attention is necessary. Finally, we present experiments with standard multilayer transformers that validate our theoretical findings.", "keywords": [ "Learning theory", "Expressive capacity", "Expressive power", "Transformer", "Attention" ], "rating_list": [ 6, 8, 5, 3 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "y9Lbr6vFHF", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Beilun Wang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Dian Shen", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Meng Wang", "gender": "Not Specified", "institution": "Tongji University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiao Tan", "gender": "Female", "institution": "Southeast University", "country": "CN", "position": "PhD student" }, { "name": "Yangyang Shen", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "MS student" } ], "n_formula": 20, "n_formula_1": 11, "n_ref_uni": 53, "n_ref": 170, "n_ref_all": 208, "n_fig": 11, "n_tab": 13, "L_tab": 4117, "n_element_tab": 964, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1941, "n_element_tab_1": 508, "formula_len_all": 1502, "formula_len_all_1": 1051, "len_all": 210493, "len_all_1": 65009, "len_abs": 1628, "len_title": 125, "len_sents": 60152, "len_sents_1": 26554, "n_sents": 459, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1640, "title": "Bi-perspective Splitting Defense: Achieving Clean-Data-Free Backdoor Security", "abs": "Backdoor attacks have seriously threatened deep neural networks (DNNs) by embedding concealed vulnerabilities through data poisoning. To counteract these attacks, training benign models from poisoned data garnered considerable interest from researchers. High-performing defenses often rely on additional clean subsets, which is untenable due to increasing privacy concerns and data scarcity. In the absence of clean subsets, defenders resort to complex feature extraction and analysis, resulting in excessive overhead and compromised performance. In the face of these challenges, we identify the key lies in sufficient utilization of the easier-to-obtain target labels and excavation of clean hard samples. In this work, we propose a Bi-perspective Splitting Defense (BSD). BSD splits the dataset using both semantic and loss statistics characteristics through open set recognition-based splitting (OSS) and altruistic model-based data splitting (ALS) respectively, achieving good clean pool initialization. BSD further introduces class completion and selective dropping strategies in the subsequent pool updates to avoid potential class underfitting and backdoor overfitting caused by loss-guided split. Through extensive experiments on 3 benchmark datasets and against 7 representative attacks, we empirically demonstrate that our BSD is robust across various attack settings. Specifically, BSD has an average improvement in Defense Effectiveness Rating (DER) by 16.29\\% compared to 5 state-of-the-art defenses, achieving clean-data-free backdoor security with minimal compromise in both Clean Accuracy (CA) and Attack Success Rate (ASR).", "keywords": [ "Trustworthy AI", "Backdoor Defense", "Deep Neural Networks" ], "rating_list": [ 6, 3, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "y9A2TpaGsE", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Efstratios Gavves", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Associate Professor" }, { "name": "Ivan Titov", "gender": "unknown", "institution": "University of Edinburgh", "country": "", "position": "Full Professor" }, { "name": "John Gkountouras", "gender": "Male", "institution": "University of Amsterdam, University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Matthias Lindemann", "gender": "unknown", "institution": "University of Edinburgh, University of Edinburgh", "country": "", "position": "PhD student" }, { "name": "Phillip Lippe", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 5, "n_ref_uni": 60, "n_ref": 97, "n_ref_all": 109, "n_fig": 1, "n_tab": 9, "L_tab": 2346, "n_element_tab": 216, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 808, "n_element_tab_1": 60, "formula_len_all": 1055, "formula_len_all_1": 335, "len_all": 227689, "len_all_1": 73094, "len_abs": 1238, "len_title": 117, "len_sents": 66578, "len_sents_1": 33686, "n_sents": 484, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1245, "title": "Language Agents Meet Causality -- Bridging LLMs and Causal World Models", "abs": "Large Language Models (LLMs) have recently shown great promise in planning and reasoning applications. These tasks demand robust systems, which arguably require a causal understanding of the environment. While LLMs can acquire and reflect common sense causal knowledge from their pretraining data, this information is often incomplete, incorrect, or inapplicable to a specific environment. In contrast, causal representation learning (CRL) focuses on identifying the underlying causal structure within a given environment. We propose a framework that integrates CRLs with LLMs to enable causally-aware reasoning and planning. This framework learns a causal world model, with causal variables linked to natural language expressions. This mapping provides LLMs with a flexible interface to process and generate descriptions of actions and states in text form. Effectively, the causal world model acts as a simulator that the LLM can query and interact with. We evaluate the framework on causal inference and planning tasks across temporal scales and environmental complexities. Our experiments demonstrate the effectiveness of the approach, with the causally-aware method outperforming LLM-based reasoners, especially for longer planning horizons.", "keywords": [ "Large Language Models", "Causality", "Causal Representation Learning", "Language Agents", "Planning" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 4, 2, 3, 4 ] }, { "paper_id": "y8uPsxR8PN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Alexei Bourd", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "US", "position": "Principal Researcher" }, { "name": "Amir Said", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "US", "position": "Principal Researcher" }, { "name": "Farzad Farhadzadeh", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "US", "position": "Researcher" }, { "name": "Fatih Porikli", "gender": "Male", "institution": "QualComm", "country": "US", "position": "Senior Director" }, { "name": "Hoang Le", "gender": "unknown", "institution": "Qualcomm AI Research", "country": "", "position": "Researcher" }, { "name": "Qiqi Hou", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "US", "position": "Researcher" }, { "name": "Randall Rauwendaal", "gender": "Not Specified", "institution": "Qualcomm", "country": "US", "position": "Principal Researcher" }, { "name": "Zifeng Li", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "US", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 29, "n_ref": 45, "n_ref_all": 70, "n_fig": 6, "n_tab": 6, "L_tab": 2894, "n_element_tab": 315, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1815, "n_element_tab_1": 208, "formula_len_all": 1272, "formula_len_all_1": 1280, "len_all": 139212, "len_all_1": 66196, "len_abs": 190, "len_title": 103, "len_sents": 36699, "len_sents_1": 28850, "n_sents": 291, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1102, "title": "Sort-free Gaussian Splatting via Weighted Sum Rendering", "abs": "Recently, 3D Gaussian Splatting (3DGS) has emerged as a significant advancement in 3D scene reconstruction, attracting considerable attention due to its ability to recover high-fidelity details while maintaining low complexity. Despite the promising results achieved by 3DGS, its rendering performance is constrained by its dependence on costly non-commutative alpha-blending operations. These operations mandate complex view dependent sorting operations that introduce computational overhead, especially on the resource-constrained platforms such as mobile phones. In this paper, we propose Weighted Sum Rendering, which approximates alpha blending with weighted sums, thereby removing the need for sorting. This simplifies implementation, delivers superior performance, and eliminates the ``popping'' artifacts caused by sorting. Experimental results show that optimizing a generalized Gaussian splatting formulation to the new differentiable rendering yields competitive image quality. The method was implemented and tested in a mobile device GPU, achieving on average $1.23\\times$ faster rendering.", "keywords": [ "Sort-free", "Gaussian Splatting", "Weighted Sum Rendering" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "y8qBBbAdEv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Ankush Khandelwal", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "", "position": "Researcher" }, { "name": "Praveen Ravirathinam", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" }, { "name": "Rahul Ghosh", "gender": "Male", "institution": "University of Minnesota, Minneapolis", "country": "", "position": "PhD student" }, { "name": "Vipin Kumar", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 58, "n_ref_all": 77, "n_fig": 10, "n_tab": 2, "L_tab": 1031, "n_element_tab": 63, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 732, "n_element_tab_1": 51, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 126997, "len_all_1": 69818, "len_abs": 1411, "len_title": 151, "len_sents": 40141, "len_sents_1": 35213, "n_sents": 260, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1415, "title": "Towards a Knowledge guided Multimodal Foundation Model for Spatio-Temporal Remote Sensing Applications", "abs": "In recent years, there has been an increased interest in foundation models for geoscience due to the vast amount of Earth observing satellite imagery. Existing remote sensing foundation models make use of the various sources of spectral imagery to create large models pretrained on the task of masked reconstruction. In this paper, we present a foundation model framework, where the pretraining task captures the causal relationship between multiple modalities. Our framework leverages the knowledge guided principles that the spectral imagery captures the impact of the physical drivers on the environmental system, and that the relationship between them is governed by the characteristics of the system. Specifically, our method, called MultiModal Variable Step Forecasting (MM-VSF), uses forecasting of satellite imagery as a pretraining task and is able to capture the causal relationship between spectral imagery and weather. In our evaluation we show that the forecasting of satellite imagery using weather can be used as an effective pretraining task for foundation models. We further show the effectiveness of the embeddings produced by MM-VSF on the downstream tasks of pixel wise crop mapping and missing image prediction of spectral imagery, when compared with embeddings created by models trained in alternative pretraining settings including the traditional single modality input masked reconstruction.", "keywords": [ "Foundation model", "Spatiotemporal modelling", "Remote Sensing", "Knowledge guided" ], "rating_list": [ 5, 3, 5, 1 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 3, 1, 2, 1 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "y8TjnkdWNA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Alex Troy Mallen", "gender": "unknown", "institution": "University of Washington, Seattle", "country": "", "position": "Undergrad student" }, { "name": "Nora Belrose", "gender": "Female", "institution": "EleutherAI", "country": "", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 24, "n_ref": 41, "n_ref_all": 62, "n_fig": 5, "n_tab": 4, "L_tab": 542, "n_element_tab": 178, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 134691, "len_all_1": 87302, "len_abs": 1425, "len_title": 109, "len_sents": 25715, "len_sents_1": 22125, "n_sents": 179, "n_sents_1": 150, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1433, "title": "Balancing Label Quantity and Quality for Scalable Elicitation", "abs": "Scalable oversight studies methods of training and evaluating AI systems in domains where human judgement is unreliable or expensive, such as scientific research and software engineering in complex codebases. Recent work in this area by Burns et al. (2023) suggests that Language Models (LMs) pretrained on internet-scale corpora exhibit an inductive bias toward producing correct answers, even when finetuned on error-prone labels produced by a smaller language model. This suggests that massive pretraining combined with finetuning on imperfect human labels may be a solid baseline method for scalable oversight. In the real world, however, label quality is not fixed: practitioners face a quantity-quality tradeoff when generating finetuning data. In this paper, we explore the microeconomics of the quantity-quality tradeoff on binary NLP classification tasks used in Burns et al. (2023). We find that there are three regimes of eliciting classification knowledge from pretrained models using supervised finetuning: quantity-dominant, quality-dominant, and a mixed regime involving the use of low- and high-quality data together to attain higher accuracy at a lower cost than using either alone. We explore sample-efficient elicitation methods that make use of two datasets of differing qualities, and establish a Pareto frontier of scalable elicitation methods that optimally trade off labeling cost and classifier performance.", "keywords": [ "Scalable oversight", "Alignment", "Safety", "Few-shot learning", "Eliciting latent knowledge", "Weak-to-strong generalization" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 1, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "y80D4IojuY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Angjoo Kanazawa", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "Assistant Professor" }, { "name": "Gengshan Yang", "gender": "unknown", "institution": "World Labs", "country": "US", "position": "Researcher" }, { "name": "Shunsuke Saito", "gender": "Male", "institution": "Codec Avatars Lab", "country": "US", "position": "Researcher" }, { "name": "Andrea Victoria Bajcsy", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 17, "n_ref_uni": 78, "n_ref": 117, "n_ref_all": 146, "n_fig": 17, "n_tab": 6, "L_tab": 897, "n_element_tab": 83, "n_fig_1": 12, "n_tab_1": 5, "L_tab_1": 1167, "n_element_tab_1": 95, "formula_len_all": 695, "formula_len_all_1": 707, "len_all": 198524, "len_all_1": 73783, "len_abs": 1267, "len_title": 82, "len_sents": 49351, "len_sents_1": 32370, "n_sents": 458, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1112, "title": "Agent-to-Sim: Learning Interactive Behavior Model from Casual Longitudinal Videos", "abs": "We present Agent-to-Sim (ATS), a framework for learning interactive behavior models of 3D agents in a 3D environment from casually-captured videos. Different from prior works that rely on marker-based tracking and multiview cameras, ATS learns natural behaviors of animal and human agents in a non-invasive way, directly from monocular video collections. Modeling 3D behavior of an agent requires persistent 3D tracking (e.g., knowing which point corresponds to which) over a long time period. To obtain such data, we develop a coarse-to-fine registration method that tracks the agent and the camera over time through a canonical 3D space, resulting in a complete and persistent spacetime 4D representation. We then train a generative model of agent behaviors using paired data of perception and motion of an agent queried from the 4D reconstruction. ATS enables real-to-sim transfer of agents in their familiar environments given longitudinal video recordings (e.g., over a month). We demonstrate results on pets (e.g., cat, dog, bunny) and human given monocular RGBD video collections captured by a smartphone.", "keywords": [ "dynamic 3d reconstruction; multi-video registration; motion generation" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 3, 4, 5 ] }, { "paper_id": "y7Ud3RAPT8", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Fei Guo", "gender": "Female", "institution": "Central South University", "country": "CN", "position": "Full Professor" }, { "name": "Qiaozhen Meng", "gender": "unknown", "institution": "Xiangtan University", "country": "CN", "position": "Lecturer" }, { "name": "kerwin", "gender": "unknown", "institution": "Central South University", "country": "CN", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 50, "n_ref": 116, "n_ref_all": 130, "n_fig": 1, "n_tab": 4, "L_tab": 1229, "n_element_tab": 186, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2233, "n_element_tab_1": 118, "formula_len_all": 656, "formula_len_all_1": 655, "len_all": 132545, "len_all_1": 62253, "len_abs": 1459, "len_title": 150, "len_sents": 31154, "len_sents_1": 27623, "n_sents": 229, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1465, "title": "MolCoMA: Complementary Masking Strategy for Promoting Atom-Level Multi-Modal Molecular Representation", "abs": "Molecular representation learning, which captures the fundamental characteristics of chemical compounds, is crucial for AI-driven drug discovery. Methodologies exist that integrate various modalities (e.g., 2D topology and 3D geometry) and develop robust representations. However, current multi-modal fusion strategies either align embedding space through independent models separately, thereby overlooking complementary information, or bridge modalities at a coarse-grained level, failing to capture inherent correlation. To facilitate fine-grained interactions of intrinsic features across modalities, this study presents MolCoMA, an innovative pretraining framework for Molecular representation, employing a unified encoder that leverages Complementary Masking mechanism. Specifically, we first employ two distinct encoders to capture the unique characteristics and structures inherent in different modalities. We then utilize a unified encoder accompanied by a customized complementary masking strategy to seamlessly integrate information, mitigating overlap and similarity between 2D and 3D representations. Finally, we incorporate a cross-modal reconstruction module to enhance fine-grained interactions at the atomic level. Extensive experiments demonstrate that our model outperforms existing molecular pretraining methods across both 2D and 3D benchmarks. This finding underscores the effectiveness of our approach to fusing information between modalities.", "keywords": [ "Multi-modal Fusion", "Molecular Pretraining", "Molecular Representation Learning" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "y6wVRmPwDu", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Hao Kong", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Jian Guo", "gender": "Male", "institution": "International Digital Economy Academy, International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Lionel Ni", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou))", "country": "CN", "position": "Full Professor" }, { "name": "Saizhuo Wang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Wanyun Zhou", "gender": "Female", "institution": "The Hong Kong University of Science and Technology (GuangZhou)", "country": "CN", "position": "PhD student" }, { "name": "Yiyan Qi", "gender": "unknown", "institution": "IDEA", "country": "CN", "position": "Researcher" }, { "name": "Fengrui HUA", "gender": "Male", "institution": "Hong Kong University of Science and Technology (GuangZhou)", "country": "", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 51, "n_ref": 73, "n_ref_all": 89, "n_fig": 4, "n_tab": 13, "L_tab": 6645, "n_element_tab": 546, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 5620, "n_element_tab_1": 282, "formula_len_all": 364, "formula_len_all_1": 0, "len_all": 161465, "len_all_1": 67977, "len_abs": 1202, "len_title": 111, "len_sents": 34971, "len_sents_1": 26222, "n_sents": 262, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1205, "title": "QuantBench: Benchmarking AI Modeling for Quantitative Investment", "abs": "The field of artificial intelligence (AI) in quantitative investment has seen significant advancements, yet it lacks a standardized benchmark aligned with industry practices. This gap hinders research progress and limits the practical application of academic innovations. We present QuantBench, an industrial-grade benchmark platform designed to address this critical need. QuantBench offers three key strengths: (1) standardization that aligns with quantitative investment industry practices, (2) flexibility to integrate various AI algorithms, and (3) full-pipeline coverage of the entire quantitative investment process. Our empirical studies using QuantBench reveal some critical research directions, including the need for continual learning to address distribution shifts, improved methods for modeling relational financial data, and more robust approaches to mitigate overfitting in low signal-to-noise environments. By providing a common ground for evaluation and fostering collaboration between researchers and practitioners, QuantBench aims to accelerate progress in AI for quantitative investment, similar to the impact of benchmark platforms in computer vision and natural language processing.", "keywords": [ "deep learning", "quantitative investment" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "y5tkxH7kxQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Chenjia Bai", "gender": "Male", "institution": "TeleAI, China Telecom", "country": "CN", "position": "Research Scientist" }, { "name": "Fei Wu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Shixin Yang", "gender": "Male", "institution": "Northwestern Polytechnical University, Northwest Polytechnical University Xi'an", "country": "", "position": "MS student" }, { "name": "Xiu Li", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Xuelong Li", "gender": "Male", "institution": "China Telecom", "country": "CN", "position": "Full Professor" }, { "name": "Yang Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhen Wang", "gender": "Male", "institution": "Northwestern Polytechnical University", "country": "CN", "position": "Full Professor" } ], "n_formula": 50, "n_formula_1": 18, "n_ref_uni": 44, "n_ref": 116, "n_ref_all": 178, "n_fig": 19, "n_tab": 11, "L_tab": 5240, "n_element_tab": 719, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 309, "n_element_tab_1": 244, "formula_len_all": 4477, "formula_len_all_1": 1110, "len_all": 274602, "len_all_1": 78550, "len_abs": 1406, "len_title": 119, "len_sents": 69599, "len_sents_1": 33593, "n_sents": 572, "n_sents_1": 262, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1418, "title": "Towards Efficient LLM Grounding for Embodied Multi-Agent Collaboration", "abs": "Grounding the reasoning ability of large language models (LLMs) for embodied tasks is challenging due to the complexity of the physical world. Especially, LLM planning for multi-agent collaboration requires communication of agents or credit assignment as the feedback to re-adjust the proposed plans and achieve effective coordination. However, existing methods that overly rely on physical verification or self-reflection suffer from excessive and inefficient querying of LLMs. In this paper, we propose a novel framework for multi-agent collaboration that introduces Reinforced Advantage feedback (ReAd) for efficient self-refinement of plans. Specifically, we perform critic regression to learn a sequential advantage function from LLM-planned data, and then treat the LLM planner as an optimizer to generate actions that maximize the advantage function. It endows the LLM with the foresight to discern whether the action contributes to accomplishing the final task. We provide theoretical analysis by extending advantage-weighted regression in reinforcement learning to multi-agent systems. Experiments on Overcooked-AI and a difficult variant of RoCoBench show that ReAd surpasses baselines in success rate, and also significantly decreases the interaction steps of agents and query rounds of LLMs, demonstrating its high efficiency for grounding LLMs. More results are given at \\url{https://read-llm.github.io/}.", "keywords": [ "LLM planning", "Large Language Models", "Multi-Agent Collaboration" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "y5einmJ0Yx", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Danny Wang", "gender": "Male", "institution": "The University of Queensland", "country": "AU", "position": "PhD student" }, { "name": "Guangdong Bai", "gender": "unknown", "institution": "University of Queensland", "country": "", "position": "Associate Professor" }, { "name": "Ruihong Qiu", "gender": "Male", "institution": "The University of Queensland", "country": "AU", "position": "Assistant Professor" }, { "name": "Zi Huang", "gender": "Female", "institution": "University of Queensland", "country": "AU", "position": "Full Professor" } ], "n_formula": 26, "n_formula_1": 19, "n_ref_uni": 71, "n_ref": 205, "n_ref_all": 242, "n_fig": 8, "n_tab": 21, "L_tab": 12612, "n_element_tab": 1723, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1154, "n_element_tab_1": 66, "formula_len_all": 2407, "formula_len_all_1": 888, "len_all": 241152, "len_all_1": 68223, "len_abs": 1704, "len_title": 132, "len_sents": 57301, "len_sents_1": 32234, "n_sents": 411, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 84, "L_abs": 1668, "title": "GOLD: Graph Out-of-Distribution Detection via Implicit Adversarial Latent Generation", "abs": "Despite graph neural networks' (GNNs) great success in modelling graph-structured data, out-of-distribution (OOD) test instances still pose a great challenge for current GNNs. One of the most effective techniques to detect OOD nodes is to expose the detector model with an additional OOD node-set, yet the extra OOD instances are often difficult to obtain in practice. Recent methods for image data address this problem using OOD data synthesis, typically relying on pre-trained generative models like Stable Diffusion. However, these approaches require vast amounts of additional data, as well as one-for-all pre-trained generative models, which are not available for graph data. Therefore, we propose the GOLD framework for graph OOD detection, an implicit adversarial learning pipeline with synthetic OOD exposure without pre-trained models. The implicit adversarial training process employs a novel alternating optimisation framework by training: (1) a latent generative model to regularly imitate the in-distribution (ID) embeddings from an evolving GNN, and (2) a GNN encoder and an OOD detector to accurately classify ID data while increasing the energy divergence between the ID embeddings and the generative model's synthetic embeddings. This novel approach implicitly transforms the synthetic embeddings into pseudo-OOD instances relative to the ID data, effectively simulating exposure to OOD scenarios without auxiliary data. Extensive OOD detection experiments are conducted on five benchmark graph datasets, verifying the superior performance of GOLD without using real OOD data compared with the state-of-the-art OOD exposure and non-exposure baselines.", "keywords": [ "Graph Neural Network", "Out-of-Distribution Detection" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "y5G1BfV7Am", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Andrew Tao", "gender": "Male", "institution": "NVIDIA Corporation", "country": "US", "position": "Vice President of Applied Research" }, { "name": "Dan Xu", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "De-An Huang", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Hanrong Ye", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Jan Kautz", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "VP Research" }, { "name": "Pavlo Molchanov", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Research Scientist" }, { "name": "Song Han", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Wei Ping", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Yao Lu", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Zhiding Yu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Hongxu Yin", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Senior Research Scientist" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 48, "n_ref": 128, "n_ref_all": 152, "n_fig": 10, "n_tab": 2, "L_tab": 276, "n_element_tab": 27, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 150, "n_element_tab_1": 17, "formula_len_all": 467, "formula_len_all_1": 486, "len_all": 182683, "len_all_1": 62358, "len_abs": 1023, "len_title": 0, "len_sents": 45114, "len_sents_1": 28484, "n_sents": 361, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1028, "title": "X-VILA: Cross-Modality Alignment for Large Language Models", "abs": "We introduce X-VILA, an omni-modality model designed to extend the capabilities of large language models (LLMs) by incorporating image, video, and audio modalities. By aligning modality-specific encoders with LLM inputs and diffusion decoders with LLM outputs, X-VILA achieves cross-modality understanding, reasoning, and generation. To facilitate this cross-modality alignment, we curate an effective interleaved any-to-any modality instruction-following dataset. Furthermore, we identify a significant problem with the current cross-modality alignment method, which results in visual information loss. To address the issue, we propose a visual alignment mechanism with a visual embedding highway module. We then introduce a resource-efficient recipe for training X-VILA, that exhibits proficiency in any-to-any modality conversation, surpassing previous approaches by large margins. X-VILA also showcases emergent properties across modalities even in the absence of similar training data. The project will be made open-source.", "keywords": [ "Multi-task learning", "vision-language models", "generative models" ], "rating_list": [ 5, 8, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "y5B0ca4mjt", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Eunbyung Park", "gender": "Male", "institution": "Sungkyunkwan University", "country": "KR", "position": "Assistant Professor" }, { "name": "Jaemin Oh", "gender": "unknown", "institution": "Texas A&M University - College Station", "country": "US", "position": "Postdoc" }, { "name": "Namgyu Kang", "gender": "Male", "institution": "Sungkyunkwan University", "country": "", "position": "MS student" }, { "name": "Youngjoon Hong", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" } ], "n_formula": 14, "n_formula_1": 10, "n_ref_uni": 57, "n_ref": 97, "n_ref_all": 119, "n_fig": 7, "n_tab": 3, "L_tab": 954, "n_element_tab": 128, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 339, "n_element_tab_1": 58, "formula_len_all": 808, "formula_len_all_1": 537, "len_all": 172695, "len_all_1": 60811, "len_abs": 1679, "len_title": 123, "len_sents": 41017, "len_sents_1": 27522, "n_sents": 327, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1647, "title": "PIG: Physics-Informed Gaussians as Adaptive Parametric Mesh Representations", "abs": "The approximation of Partial Differential Equations (PDEs) using neural networks has seen significant advancements through Physics-Informed Neural Networks (PINNs). Despite their straightforward optimization framework and flexibility in implementing various PDEs, PINNs often suffer from low accuracy due to the spectral bias of Multi-Layer Perceptrons (MLPs), which struggle to effectively learn high-frequency and non-linear components. Recently, the parametric mesh representations have been investigated as a promising approach to effectively eliminating the inductive biases of neural networks. However, they often require very high-resolution grids and a large number of collocation points to achieve high accuracy while avoiding overfitting issues. In addition, these are limited by the fixed positions of the mesh parameters, hindering their ability to approximate complex PDEs. To overcome these limitations, we propose Physics-Informed Gaussians (PIGs), which combine feature embeddings using Gaussian functions with a lightweight neural network. Our approach uses trainable parameters for the mean and variance of each Gaussian, allowing for dynamic adjustment of their positions and shapes during training. This adaptability enables our model to optimally approximate PDE solutions, unlike models with fixed parameter positions. Furthermore, the proposed approach maintains the same optimization framework used in PINNs, allowing us to benefit from their excellent properties. Experimental results show the competitive performance of our model across various PDEs, demonstrating its potential as a robust tool for solving complex PDEs.", "keywords": [ "Gaussians", "Physics-informed Deep Learning" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "y59zhBNKGZ", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 3, "author_info_list": [ { "name": "Armin Gerami", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Monte Hoover", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Ramani Duraiswami", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "Full Professor" } ], "n_formula": 80, "n_formula_1": 55, "n_ref_uni": 17, "n_ref": 27, "n_ref_all": 32, "n_fig": 2, "n_tab": 1, "L_tab": 311, "n_element_tab": 17, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 311, "n_element_tab_1": 17, "formula_len_all": 6177, "formula_len_all_1": 3720, "len_all": 113969, "len_all_1": 54730, "len_abs": 1732, "len_title": 86, "len_sents": 27688, "len_sents_1": 21537, "n_sents": 259, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 38, "L_abs": 1385, "title": "Towards Making Linear Attention Usable", "abs": "The original Transformer attention mechanism, based on Softmax, has time and memory complexities of $O(N^2D)$ and $O(N^2)$, where $N$ is the number of tokens and $D$ the dimension per attention head. As current LLM applications trend towards processing larger token sequences, and Transformers gain popularity in image, video, and audio processing, addressing this quadratic cost becomes imperative. Since the introduction of Transformers, numerous approaches have been proposed to linearize this scaling. One such method is Linear Attention, which captures all-to-all token pair attention in $O(ND^2)$ time. However, its drawback lies in its high memory footprint of $O(ND^2)$. While Linear Attention has shown promise in small-scale benchmarks, the high memory demand has prevented Linear Attention to be studied in context of large benchmarks and practical use cases. In this work, we demonstrate how to reduce the memory complexity to $O(ND)$ by approaching calculations from a novel perspective. Additionally, since Linear Attention does not compute the attention matrix directly, it precludes the use of traditional dropout. To address this, we introduce an alternative dropout mechanism. Our study confirms linear scaling in both wall-clock time and memory usage. We also compare our method with Flash Attention and conduct an ablation study on our proposed dropout alternative.", "keywords": [ "Linear Attention", "Kernel Separation", "Transformers", "Memory Reduction" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "y4F2YZxN9T", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Dan Zhang", "gender": "unknown", "institution": "Robert Bosch GmbH, Bosch", "country": "DE", "position": "Research Scientist" }, { "name": "Eric Nalisnick", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" }, { "name": "Mona Schirmer", "gender": "Female", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 16, "n_ref_uni": 69, "n_ref": 146, "n_ref_all": 168, "n_fig": 7, "n_tab": 17, "L_tab": 3886, "n_element_tab": 372, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 452, "n_element_tab_1": 46, "formula_len_all": 2953, "formula_len_all_1": 1347, "len_all": 216591, "len_all_1": 76220, "len_abs": 973, "len_title": 101, "len_sents": 59283, "len_sents_1": 34821, "n_sents": 515, "n_sents_1": 295, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 977, "title": "Temporal Test-Time Adaptation with State-Space Models", "abs": "Distribution shifts between training and test data are inevitable over the lifecycle of a deployed model, leading to performance decay. Adapting a model on test samples can help mitigate this drop in performance. However, most test-time adaptation methods have focused on synthetic corruption shifts, leaving a variety of distribution shifts underexplored. In this paper, we focus on distribution shifts that evolve gradually over time, which are common in the wild but challenging for existing methods, as we show. To address this, we propose STAD, a probabilistic state-space model that adapts a deployed model to temporal distribution shifts by learning the time-varying dynamics in the last set of hidden features. Without requiring labels, our model infers time-evolving class prototypes that act as a dynamic classification head. Through experiments on real-world temporal distribution shifts, we show that our method excels in handling small batch sizes and label shift.", "keywords": [ "test-time adaptation", "state-space models", "probabilistic modelling", "dynamical systems" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "y4DtzADzd1", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Jakob Verbeek", "gender": "Not Specified", "institution": "Meta", "country": "FR", "position": "Research Scientist" }, { "name": "Karteek Alahari", "gender": "Male", "institution": "Inria", "country": "FR", "position": "Research director (eq. Full professor)" }, { "name": "Marton Havasi", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Melissa Hall", "gender": "Female", "institution": "Research, Facebook", "country": "US", "position": "Researcher" }, { "name": "Michal Drozdzal", "gender": "Male", "institution": "Meta", "country": "CA", "position": "Research Scientst" }, { "name": "Pietro Astolfi", "gender": "Male", "institution": "Meta", "country": "", "position": "Postdoc" }, { "name": "Tariq Berrada", "gender": "Male", "institution": "INRIA", "country": "", "position": "PhD student" }, { "name": "Yohann Benchetrit", "gender": "Male", "institution": "Meta AI", "country": "", "position": "Researcher" }, { "name": "Adriana Romero", "gender": "Female", "institution": "Mila", "country": "CA", "position": "Core Industry Member" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 53, "n_ref": 97, "n_ref_all": 117, "n_fig": 10, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 242, "n_element_tab_1": 52, "formula_len_all": 775, "formula_len_all_1": 625, "len_all": 134488, "len_all_1": 68030, "len_abs": 396, "len_title": 100, "len_sents": 35671, "len_sents_1": 31009, "n_sents": 245, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1133, "title": "Boosting Latent Diffusion with Perceptual Objectives", "abs": "Latent diffusion models (LDMs) power state-of-the-art high-resolution generative image models. LDMs learn the data distribution in the latent space of an autoencoder (AE) and produce images by mapping the generated latents into RGB image space using the AE decoder. While this approach allows for efficient model training and sampling, it induces a disconnect between the training of the diffusion model and the decoder, resulting in a loss of detail in the generated images. To remediate this disconnect, we propose to leverage the internal features of the decoder to define a latent perceptual loss (LPL). This loss encourages the models to create sharper and more realistic images. Our loss can be seamlessly integrated with common autoencoders used in latent diffusion models, and can be applied to different generative modeling paradigms such as DDPM with epsilon and velocity prediction, as well as flow matching. Extensive experiments with models trained on three datasets at 256 and 512 resolution show improved quantitative -- with boosts between 6% and 20% in FID -- and qualitative results when using our perceptual loss.", "keywords": [ "diffusion", "flows", "latent diffusion", "LDM", "latent generative models", "T2I", "image generation", "generative models." ], "rating_list": [ 6, 5, 8, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "y3zswp3gek", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Dominik Wagner", "gender": "Male", "institution": "Technische Hochschule N\u00fcrnberg Georg Simon Ohm", "country": "DE", "position": "Researcher" }, { "name": "Juho Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Minki Kang", "gender": "Male", "institution": "KRAFTON", "country": "KR", "position": "Researcher" }, { "name": "Seanie Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Sung Ju Hwang", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Xiaoyin Chen", "gender": "unknown", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "PhD student" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" }, { "name": "DONGBOK LEE", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Haebin Sung", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 46, "n_ref": 104, "n_ref_all": 142, "n_fig": 11, "n_tab": 7, "L_tab": 2856, "n_element_tab": 245, "n_fig_1": 9, "n_tab_1": 6, "L_tab_1": 3875, "n_element_tab_1": 270, "formula_len_all": 621, "formula_len_all_1": 514, "len_all": 189276, "len_all_1": 65930, "len_abs": 2570, "len_title": 134, "len_sents": 45285, "len_sents_1": 27316, "n_sents": 357, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1935, "title": "HarmAug: Effective Data Augmentation for Knowledge Distillation of Safety Guard Models", "abs": "Safety guard models that detect malicious queries aimed at large language models (LLMs) are essential for ensuring the secure and responsible deployment of LLMs in real-world applications.\nHowever, deploying existing safety guard models with billions of parameters alongside LLMs on mobile devices is impractical due to substantial memory requirements and latency.\nTo reduce this cost, we distill a large teacher safety guard model into a smaller one using a labeled dataset of instruction-response pairs with binary harmfulness labels. Due to the limited diversity of harmful instructions in the existing labeled dataset, naively distilled models tend to underperform compared to larger models. To bridge the gap between small and large models, we propose **HarmAug**, a simple yet effective data augmentation method that involves jailbreaking an LLM and prompting it to generate harmful instructions. Given a prompt such as, \"Make a single harmful instruction prompt that would elicit offensive content\", we add an affirmative prefix (e.g., \"I have an idea for a prompt:\") to the LLM's response. This encourages the LLM to continue generating the rest of the response, leading to sampling harmful instructions. Another LLM generates a response to the harmful instruction, and the teacher model labels the instruction-response pair. We empirically show that our HarmAug outperforms other relevant baselines. Moreover, a 435-million-parameter safety guard model trained with HarmAug achieves an F1 score comparable to larger models with over 7 billion parameters, and even outperforms them in AUPRC, while operating at less than 25\\% of their computational cost. Our [code](https://anonymous.4open.science/r/HarmAug/), [safety guard model](https://huggingface.co/AnonHB/HarmAug_Guard_Model_deberta_v3_large_finetuned), and [synthetic dataset](https://huggingface.co/datasets/AnonHB/HarmAug_generated_dataset) are publicly available.", "keywords": [ "knowledge distillation", "safety guard" ], "rating_list": [ 6, 10, 6, 6 ], "soundness_list": [ 4, 4, 3, 4 ], "presentation_list": [ 4, 4, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "y3jJmrKWQ4", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Chiyu Ma", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "PhD student" }, { "name": "Lin Shi", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "Undergrad student" }, { "name": "Soroush Vosoughi", "gender": "unknown", "institution": "Dartmouth College", "country": "", "position": "Assistant Professor" }, { "name": "Weicheng Ma", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Wenhua Liang", "gender": "unknown", "institution": "Dartmouth College", "country": "US", "position": "Undergrad student" } ], "n_formula": 11, "n_formula_1": 6, "n_ref_uni": 37, "n_ref": 213, "n_ref_all": 252, "n_fig": 9, "n_tab": 10, "L_tab": 1428, "n_element_tab": 145, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 638, "n_element_tab_1": 37, "formula_len_all": 502, "formula_len_all_1": 409, "len_all": 228754, "len_all_1": 71285, "len_abs": 1273, "len_title": 155, "len_sents": 79545, "len_sents_1": 33315, "n_sents": 545, "n_sents_1": 261, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1281, "title": "Judging the Judges: A Systematic Investigation of Position Bias in Pairwise Comparative Assessments by LLMs", "abs": "LLM-as-a-Judge presents a promising alternative to human evaluators across various tasks, but inherent biases, especially position bias \u2014 a tendency to favor solutions based on their position in the prompt \u2014 have compromised its effectiveness. Our study introduces a systematic framework to examine position bias in pairwise comparisons, focusing on repetition stability, position consistency, and preference fairness. This research significantly contributes to the field by introducing new concepts for understanding position bias and providing a multi-dimensional framework for evaluations. We conducted experiments with 12 LLM judges across MTBench and DevBench, covering 22 tasks and approximately 40 solution-generating models \u2014 candidates, resulting in over 100,000 evaluation instances. Our findings confirm that position bias in capable LLM judges is not due to random chances, along with notable variations observed across judges and tasks. Moreover, position bias is weakly influenced by the length of prompt components but significantly impacted by the quality gap between solutions. These insights can help optimize judge model selections, improve benchmark design, and inform future research on debiasing strategies, ultimately enhancing the reliability of LLM judges.", "keywords": [ "LLM-as-a-Judge", "LLM evaluators", "position bias", "length bias", "verbosity bias", "pairwise comparison", "repetition stability", "position consistency", "preference fairness" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "y3CdSwREZl", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Jiahao Huo", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Kaichen Huang", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Kun Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Xuming Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Yibo Yan", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yutao Yue", "gender": "unknown", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Associate Professor" } ], "n_formula": 15, "n_formula_1": 16, "n_ref_uni": 50, "n_ref": 100, "n_ref_all": 120, "n_fig": 13, "n_tab": 4, "L_tab": 1223, "n_element_tab": 361, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 1223, "n_element_tab_1": 361, "formula_len_all": 1071, "formula_len_all_1": 1271, "len_all": 164646, "len_all_1": 69805, "len_abs": 1384, "len_title": 149, "len_sents": 39311, "len_sents_1": 27002, "n_sents": 344, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1170, "title": "MINER: Mining the Underlying Pattern of Modality-Specific Neurons in Multimodal Large Language Models", "abs": "In recent years, multimodal large language models (MLLMs) have significantly advanced, integrating more modalities into diverse applications. However, the lack of explainability remains a major barrier to their use in scenarios requiring decision transparency. Current neuron-level explanation paradigms mainly focus on knowledge localization or language- and domain-specific analyses, leaving the exploration of multimodality largely unaddressed. To tackle these challenges, we propose MINER, a transferable framework for mining modality-specific neurons (MSNs) in MLLMs, which comprises four stages: (1) modality separation, (2) importance score calculation, (3) importance score aggregation, (4) modality-specific neuron selection. Extensive experiments across six benchmarks and two MLLMs show that (1) deactivating ONLY 2% of MSNs significantly reduce MLLMs performance (0.56 to 0.24 for Qwen2-VL, 0.69 to 0.31 for Qwen2-Audio), (2) different modalities mainly converge in the lower layers, (3) MSNs influence how key information from various modalities converges to the last token, (4) We observed two intriguing phenomena, semantic probing and semantic telomeres.", "keywords": [ "MLLMs", "neuron analysis", "interpretability" ], "rating_list": [ 6, 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 1 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 3, 4 ] }, { "paper_id": "y2ch7iQSJu", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Ali Parsaee", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "MS student" }, { "name": "Bei Jiang", "gender": "Female", "institution": "University of Alberta", "country": "CA", "position": "Associate Professor" }, { "name": "Russell Greiner", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 6, "n_ref_uni": 14, "n_ref": 48, "n_ref_all": 55, "n_fig": 2, "n_tab": 10, "L_tab": 7450, "n_element_tab": 369, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 510, "n_element_tab_1": 9, "formula_len_all": 1336, "formula_len_all_1": 552, "len_all": 152251, "len_all_1": 75751, "len_abs": 1866, "len_title": 109, "len_sents": 55921, "len_sents_1": 36640, "n_sents": 384, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1900, "title": "Budget-constrained Active Learning to De-censor Survival Data", "abs": "Standard supervised learners attempt to learn a model from a labeled dataset. Given a small set of labeled instances, and a pool of unlabeled instances, a budgeted learner can use its given budget to pay to acquire the labels of some unlabeled instances, which it can then use to produce a model. Here, we explore budgeted learning in the context of survival datasets, which include (right) censored instances, where we know only a lower bound c_i on that instance\u2019s time-to-event t_i. Here, that learner can pay to (partially) label a censored instance \u2013 eg, to acquire the actual time t_i for an instance [eg, go from (3yr, censor) to (7.2yr, uncensored)], or other variants [eg, learn about 1 more year, so go from (3yr, censor) to either (3.2yr, uncensored) or (4yr, censor)]. This serves as a model of real world data collection, where followup with censored patients does not always lead to complete uncensoring, and how much information is given to the learner model during data collection is a function of the budget and the nature of the data itself. Many fields, such as medicine, finance, and engineering contain survival datasets with a large number of censored instance, and also operate under budget constraints with respect to the learning process, thus making it important to be able to apply this budgeted learning approach. Despite this importance; to our knowledge no other work has looked into doing this. We provide both experimental and theoretical results for how to apply state-of-the-art budgeted learning algorithms to survival data and the respective limitations that exist in doing so. Our approach provides bounds and time complexity theoretically equivalent to standard active learning methods. Moreover, empirical analysis on several survival tasks show that our model performs better than other potential approaches that might be considered on several benchmarks.", "keywords": [ "Active Learning", "Survival Analysis", "Budgeted Constraints", "Bayesian Model", "Mutual Information", "De-censoring Data" ], "rating_list": [ 3, 3, 1, 1 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "y1iU5czYpE", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Chenggang Zhao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Damai Dai", "gender": "Male", "institution": "DeepSeek-AI", "country": "CN", "position": "Researcher" }, { "name": "Huazuo Gao", "gender": "unknown", "institution": "DeepSeek AI", "country": "CN", "position": "Researcher" }, { "name": "Lean Wang", "gender": "Male", "institution": "DeepSeek", "country": "CN", "position": "Intern" }, { "name": "Xu Sun", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 5, "n_ref": 31, "n_ref_all": 51, "n_fig": 7, "n_tab": 6, "L_tab": 1139, "n_element_tab": 155, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1044, "n_element_tab_1": 150, "formula_len_all": 508, "formula_len_all_1": 420, "len_all": 80407, "len_all_1": 46292, "len_abs": 1308, "len_title": 115, "len_sents": 26377, "len_sents_1": 20976, "n_sents": 176, "n_sents_1": 136, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1317, "title": "Auxiliary-Loss-Free Load Balancing Strategy for Mixture-of-Experts", "abs": "For Mixture-of-Experts (MoE) models, an unbalanced expert load will lead to routing collapse or increased computational overhead. Existing methods commonly employ an auxiliary loss to encourage load balance, but a large auxiliary loss will introduce non-negligible interference gradients into training and thus impair the model performance. In order to control load balance while not producing undesired gradients during training, we propose **Loss-Free Balancing**, a new load balancing strategy that operates without auxiliary losses. To be specific, before the top-K routing decision, Loss-Free Balancing will first apply an expert-wise bias to the routing scores of each expert. By dynamically updating the bias of each expert according to its recent load, Loss-Free Balancing can consistently maintain a balanced distribution of expert load. In addition, since Loss-Free Balancing does not produce any interference gradients, it also elevates the upper bound of model performance gained from MoE training. We validate the performance of Loss-Free Balancing on MoE models with up to 3B parameters trained on up to 200B tokens. Experimental results show that Loss-Free Balancing achieves both better performance and better load balance compared with traditional auxiliary-loss-controlled load balancing strategies.", "keywords": [ "mixture of experts", "load balancing", "auxiliary-loss-free" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "y1UHa9sl2w", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Arya Hadizadeh Moghaddam", "gender": "Male", "institution": "University of Kansas", "country": "", "position": "PhD student" }, { "name": "Dongjie Wang", "gender": "Male", "institution": "University of Kansas", "country": "US", "position": "Assistant Professor" }, { "name": "Mohsen Nayebi Kerdabadi", "gender": "Male", "institution": "University of Kansas", "country": "US", "position": "PhD student" }, { "name": "Zijun Yao", "gender": "Male", "institution": "University of Kansas", "country": "US", "position": "Assistant Professor" } ], "n_formula": 29, "n_formula_1": 29, "n_ref_uni": 25, "n_ref": 52, "n_ref_all": 65, "n_fig": 6, "n_tab": 7, "L_tab": 5588, "n_element_tab": 364, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2971, "n_element_tab_1": 170, "formula_len_all": 1874, "formula_len_all_1": 1851, "len_all": 123364, "len_all_1": 68755, "len_abs": 2040, "len_title": 127, "len_sents": 34921, "len_sents_1": 28481, "n_sents": 272, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 2095, "title": "OntoFAR: Hierarchical Multi-Ontology Fusion Better Augments EHR Representation", "abs": "Medical ontology graphs, which typically organize and relate comprehensive medical concepts in a hierarchical structure, are able to map a rich set of external knowledge onto the specific medical codes observed in electronic health records (EHRs). Through the connectivity in ontologies, healthcare predictive models can utilize the ancestor, descendant, or sibling information to add supplementary contexts on medical codes, thereby augmenting expressiveness of EHR representations. However, existing approaches are limited by the heterogeneous isolation of different ontology systems (e.g., conditions vs. drugs), that different types of ontology concepts have to be learned individually, and only the homogeneous ontology relationships can be exploited. This limitation restricts the existing methods from fully leveraging the cross-ontology relationships which could substantially enhance healthcare representations. \nIn this paper, we propose OntoFAR, a framework that fuse multiple ontology graphs, utilizing the collaboration across ontologies to enhance medical concept representation. Our method jointly represents medical concepts cross multiple ontology structures by performing message passing in two dimensions: (1) vertical propagation over levels of ontology hierarchy, and (2) horizontal propagation over co-occurring concepts in EHR visits. Additionally, OntoFAR leverages the large language models (LLMs) pre-trained on massive open world information to understand each target concept with its ontology relationships, providing enhanced embedding initialization for concepts. Through extensive experimental studies on two public datasets, MIMIC-III and MIMIC-IV, we validate the superior performance of OntoFAR over the state-of-the-art baselines. Beyond accuracy, our model also exhibits the add-on compatibility to boost existing healthcare prediction models, and demonstrate a good robustness in scenarios with limited data availability. The implementation code is available at [https://anonymous.4open.science/r/OntoFAR-35D4](https://anonymous.4open.science/r/OntoFAR-35D4)", "keywords": [ "Health Informatics", "EHR", "Diagnosis Prediction", "Healthcare Representation" ], "rating_list": [ 5, 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 2, 5 ] }, { "paper_id": "y15LAM4u0A", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Baining Zhao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Chen Gao", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Fanhang Man", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jianjie Fang", "gender": "unknown", "institution": "Northeastern University", "country": "CN", "position": "Undergrad student" }, { "name": "Jinqiang Cui", "gender": "Male", "institution": "Pengcheng Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Jinzhu Mao", "gender": "Male", "institution": "Tsinghua University ", "country": "", "position": "MS student" }, { "name": "Jun Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Weichen Zhang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xinlei Chen", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Zhiheng Zheng", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Zile Zhou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 39, "n_ref_all": 74, "n_fig": 11, "n_tab": 10, "L_tab": 4112, "n_element_tab": 353, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2365, "n_element_tab_1": 152, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 164862, "len_all_1": 61175, "len_abs": 2174, "len_title": 133, "len_sents": 54138, "len_sents_1": 26930, "n_sents": 481, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 2186, "title": "EmbodiedCity: A Benchmark Platform for Embodied Agent in Real-world City Environment", "abs": "Embodied artificial intelligence (EmbodiedAI) emphasizes the role of an agent's body in generating human-like behaviors. The recent efforts on EmbodiedAI pay a lot of attention to building up machine learning models to possess perceiving, planning, and acting abilities, thereby enabling real-time interaction with the world. However, most works focus on bounded indoor environments, such as navigation in a room or manipulating a device, with limited exploration of embodying the agents in open-world scenarios. That is, embodied intelligence in the open and outdoor environment is less explored, for which one potential reason is the lack of high-quality simulators, benchmarks, and datasets. To address it, in this paper, we construct a benchmark platform for embodied intelligence evaluation in real-world city environments. Specifically, we first construct a highly realistic 3D simulation environment based on the real buildings, roads, and other elements in a real city. In this environment, we combine historically collected data and simulation algorithms to conduct simulations of pedestrian and vehicle flows with high fidelity. Further, we designed a set of evaluation tasks covering different EmbodiedAI abilities. Moreover, we provide a complete set of input and output interfaces for access, enabling embodied agents to easily take task requirements and current environmental observations as input and then make decisions and obtain performance evaluations. On the one hand, it expands the capability of existing embodied intelligence to higher levels. On the other hand, it has a higher practical value in the real world and can support more potential applications for artificial general intelligence. Based on this platform, we evaluate some popular large language models for embodied intelligence capabilities of different dimensions and difficulties. The executable program of this platform is available for download, and we have also released an easy-to-use Python library and detailed tutorial documents. All of the software, Python library, codes, datasets, tutorials, and real-time online service are available on this anonymous website: https://embodied-ai.city.", "keywords": [ "Embodied intelligence", "real-world city environment", "large language model agent", "benchmark" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "y10AP0BkID", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Bingchen Gong", "gender": "Not Specified", "institution": "\u00c9cole Polytechnique", "country": "FR", "position": "Postdoc" }, { "name": "Sipeng Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Xiaogang Jin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoguang Han", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Xinyu Gao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Ziyi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 27, "n_ref": 44, "n_ref_all": 76, "n_fig": 17, "n_tab": 5, "L_tab": 1351, "n_element_tab": 65, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 737, "formula_len_all_1": 680, "len_all": 141382, "len_all_1": 48097, "len_abs": 1770, "len_title": 114, "len_sents": 34191, "len_sents_1": 21863, "n_sents": 258, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1780, "title": "Towards Realistic Example-based Modeling via 3D Gaussian Stitching", "abs": "Using parts of existing models to rebuild new models, commonly termed as example-based modeling, is a classical methodology in the realm of computer graphics. Previous works mostly focus on shape composition, making them very hard to use for realistic composition of 3D objects captured from real-world scenes. This leads to combining multiple NeRFs into a single 3D scene to achieve seamless appearance blending. However, the current SeamlessNeRF method struggles to achieve interactive editing and harmonious stitching for real-world scenes due to its gradient-based strategy and grid-based representation.\n\nTo this end, we present an example-based modeling method that combines multiple Gaussian fields in a point-based representation using sample-guided synthesis. Specifically, as for composition, we create a GUI to segment and transform multiple fields in real time, easily obtaining a semantically meaningful composition of models represented by 3D Gaussian Splatting (3DGS). For texture blending, due to the discrete and irregular nature of 3DGS, straightforwardly applying gradient propagation as SeamlssNeRF is not supported. Thus, a novel sampling-based cloning method is proposed to harmonize the blending while preserving the original rich texture and content. Our workflow consists of three steps: 1) real-time segmentation and transformation of a Gaussian model using a well-tailored GUI, 2) KNN analysis to identify boundary points in the intersecting area between the source and target models, and 3) two-phase optimization of the target model using sampling-based cloning and gradient constraints. Extensive experimental results validate that our approach significantly outperforms previous works in terms of realistic synthesis, demonstrating its practicality.", "keywords": [ "Gaussian splatting", "Composition", "Example-based Modeling" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "xzSUdw6s76", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Hao Zhang", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "PhD student" }, { "name": "Jayaram Raghuram", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Researcher" }, { "name": "Jingyu Liu", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Undergrad student" }, { "name": "M Badri Narayanan", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "MS student" }, { "name": "Suman Banerjee", "gender": "Male", "institution": "UW-Madison", "country": "US", "position": "Full Professor" }, { "name": "Utkarsh Sharma", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Undergrad student" }, { "name": "Yijing Zeng", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Yilong Li", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "SHUAI Zhang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 121, "n_ref_all": 150, "n_fig": 11, "n_tab": 6, "L_tab": 1505, "n_element_tab": 151, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 649, "n_element_tab_1": 17, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 129178, "len_all_1": 57139, "len_abs": 1489, "len_title": 92, "len_sents": 38793, "len_sents_1": 28222, "n_sents": 286, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1465, "title": "PALMBENCH: A COMPREHENSIVE BENCHMARK OF COMPRESSED LARGE LANGUAGE MODELS ON MOBILE PLATFORMS", "abs": "Deploying large language models (LLMs) locally on mobile devices is advantageous in scenarios where transmitting data to remote cloud servers is either undesirable due to privacy concerns or impractical due to network connection. Recent advancements have facilitated the local deployment of LLMs. However, local deployment also presents challenges, particularly in balancing quality (generative performance), latency, and throughput within the hardware constraints of mobile devices. In this paper, we introduce our lightweight, all-in-one automated benchmarking framework that allows users to evaluate LLMs on mobile devices. We provide a comprehensive benchmark of various popular LLMs with different quantization configurations (both weights and activations) across multiple mobile platforms with varying hardware capabilities. Unlike traditional benchmarks that assess full-scale models on high-end GPU clusters, we focus on evaluating resource efficiency (memory and power consumption) and harmful output for compressed models on mobile devices. Our key observations include: i) differences in energy efficiency and throughput across mobile platforms; ii) the impact of quantization on memory usage, GPU execution time, and power consumption; and iii) accuracy and performance degradation of quantized models compared to their non-quantized counterparts; and iv) the frequency of hallucinations and toxic content generated by compressed LLMs on\nmobile devices.", "keywords": [ "Mobile Platforms", "Large Language Models", "Quantization", "Benchmark" ], "rating_list": [ 5, 6, 5, 8, 5 ], "soundness_list": [ 2, 3, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 4, 5 ] }, { "paper_id": "xzKFnsJIXL", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Aur\u00e9lien Bellet", "gender": "unknown", "institution": "INRIA", "country": "FR", "position": "Tenured researcher" }, { "name": "Nicolas Papernot", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "Assistant Professor" }, { "name": "Tudor Ioan Cebere", "gender": "Male", "institution": "INRIA", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 48, "n_ref": 128, "n_ref_all": 172, "n_fig": 21, "n_tab": 1, "L_tab": 479, "n_element_tab": 79, "n_fig_1": 15, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 737, "formula_len_all_1": 603, "len_all": 171444, "len_all_1": 75214, "len_abs": 1380, "len_title": 67, "len_sents": 57769, "len_sents_1": 36511, "n_sents": 380, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1388, "title": "Tighter Privacy Auditing of DP-SGD in the Hidden State Threat Model", "abs": "Machine learning models can be trained with formal privacy guarantees via differentially private optimizers such as DP-SGD. In this work, we focus on a threat model where the adversary has access only to the final model, with no visibility into intermediate updates. In the literature, this ``hidden state'' threat model exhibits a significant gap between the lower bound from empirical privacy auditing and the theoretical upper bound provided by privacy accounting. To challenge this gap, we propose to audit this threat model with adversaries that craft a gradient sequence designed to maximize the privacy loss of the final model without relying on intermediate updates. Our experiments show that this approach consistently outperforms previous attempts at auditing the hidden state model. Furthermore, our results advance the understanding of achievable privacy guarantees within this threat model. Specifically, when the crafted gradient is inserted at every optimization step, we show that concealing the intermediate model updates in DP-SGD does not amplify privacy. The situation is more complex when the crafted gradient is not inserted at every step: our auditing lower bound matches the privacy upper bound only for an adversarially-chosen loss landscape and a sufficiently large batch size. This suggests that existing privacy upper bounds can be improved in certain regimes.", "keywords": [ "Differential Privacy", "Privacy Auditing", "Machine Learning" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "xz3dmxfFva", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Xinlei Chen", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Xueyang Yu", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "MS student" }, { "name": "Yossi Gandelsman", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 52, "n_ref_all": 80, "n_fig": 6, "n_tab": 12, "L_tab": 2178, "n_element_tab": 169, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 632, "n_element_tab_1": 58, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 129998, "len_all_1": 63787, "len_abs": 1536, "len_title": 101, "len_sents": 39251, "len_sents_1": 30666, "n_sents": 326, "n_sents_1": 244, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 1254, "title": "Video Representation Learning Without Natural Videos", "abs": "In this paper, we show that useful video representations can be learned from synthetic videos and natural images, without incorporating natural videos in the training. We propose a progression of video datasets synthesized by simple generative processes, that model a growing set of natural video properties (e.g. motion, acceleration, and shape transformations). The downstream performance of video models pre-trained on these generated datasets gradually increases with the dataset progression. A VideoMAE model pre-trained on our synthetic videos closes 97.2\\% of the performance gap on UCF101 action classification between training from scratch and self-supervised pre-training from natural videos, and outperforms the pre-trained model on HMDB51. Introducing crops of static images to the pre-training stage results in similar performance to UCF101 pre-training and outperforms the UCF101 pre-trained model on 11 out of 14 out-of-distribution datasets of UCF101-P. Analyzing the low-level properties of the datasets, we identify correlations between frame diversity, frame similarity to natural data, and downstream performance. Our approach provides a more controllable and transparent alternative to video data curation processes for pre-training.", "keywords": [ "video representation learning", "learning from synthetic data" ], "rating_list": [ 1, 5, 5 ], "soundness_list": [ 1, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "xyysYa4YvF", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Hao Luan", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Ping Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Zhiheng Li", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 14, "n_ref_uni": 42, "n_ref": 100, "n_ref_all": 116, "n_fig": 1, "n_tab": 5, "L_tab": 1415, "n_element_tab": 86, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 2068, "n_element_tab_1": 129, "formula_len_all": 2376, "formula_len_all_1": 1175, "len_all": 124055, "len_all_1": 64861, "len_abs": 1344, "len_title": 129, "len_sents": 35027, "len_sents_1": 29630, "n_sents": 258, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1354, "title": "Interpretable Boundary-based Watermark Up to the condition of Lov\\'asz Local Lemma", "abs": "Watermarking techniques have emerged as pivotal safeguards to defend the intellectual property of deep neural networks against model extraction attacks. Most existing watermarking methods rely on the identification of samples within randomly selected trigger sets. However, this paradigm is inevitably disrupted by the ambiguous points that exhibit poor discriminability, thus leading to the misidentification between benign and stolen models. To tackle this issue, in this paper, we propose a boundary-based watermarking method that enhances the discernibility of trigger set, further improving the ability in distinguish benign and stolen models. Specifically, we select trigger samples on the decision boundary of base model and assigned them labels with the least probabilities, while providing a tight bound based on the Lov\\'asz Local Lemma. This approach ensures the watermark's reliability in identifying stolen models by improving discriminability of trigger samples. Meanwhile, we provide theoretical proof to demonstrate that the watermark can be effectively guaranteed under the constraints guided by the Lov\\'asz Local Lemma. Experimental results demonstrate that our method outperforms the state-of-the-art watermarking methods on CIFAR-10, CIFAR-100 and ImageNet datasets. Code and data will be released publicly upon the paper acceptance.", "keywords": [ "Watermark", "Model extraction attacks", "Intellectual property protection" ], "rating_list": [ 1, 6, 5 ], "soundness_list": [ 2, 3, 4 ], "presentation_list": [ 2, 4, 3 ], "contribution_list": [ 1, 3, 3 ], "confidence_list": [ 5, 3, 5 ] }, { "paper_id": "xyfb9HHvMe", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Huaisheng Zhu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Vasant G Honavar", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Full Professor" }, { "name": "Teng Xiao", "gender": "Male", "institution": "The Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 60, "n_formula_1": 33, "n_ref_uni": 45, "n_ref": 107, "n_ref_all": 166, "n_fig": 12, "n_tab": 4, "L_tab": 1988, "n_element_tab": 229, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2479, "n_element_tab_1": 204, "formula_len_all": 19341, "formula_len_all_1": 13103, "len_all": 173933, "len_all_1": 85209, "len_abs": 476, "len_title": 120, "len_sents": 41998, "len_sents_1": 31021, "n_sents": 298, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (conditional oral)", "L_title": 72, "L_abs": 1565, "title": "DSPO: Direct Score Preference Optimization for Diffusion Model Alignment", "abs": "Diffusion-based Text-to-Image (T2I) models have achieved impressive success in generating high-quality images from textual prompts. While large language models (LLMs) effectively leverage Direct Preference Optimization (DPO) for fine-tuning on human preference data without the need for reward models, diffusion models have not been extensively explored in this area. Current preference learning methods applied to T2I diffusion models immediately adapt existing techniques from LLMs. However, this direct adaptation introduces an estimated loss specific to T2I diffusion models. This estimation can potentially lead to suboptimal performance through our empirical results. In this work, we propose Direct Score Preference Optimization (DSPO), a novel algorithm that aligns the pretraining and fine-tuning objectives of diffusion models by leveraging score matching, the same objective used during pretraining. It introduces a new perspective on preference learning for diffusion models. Specifically, DSPO distills the score function of human-preferred image distributions into pretrained diffusion models, fine-tuning the model to generate outputs that align with human preferences. We theoretically show that DSPO shares the same optimization direction as reinforcement learning algorithms in diffusion models under certain conditions. Our experimental results demonstrate that DSPO outperforms preference learning baselines for T2I diffusion models in human preference evaluation tasks and enhances both visual appeal and prompt alignment of generated images.", "keywords": [ "Text-to-image generation" ], "rating_list": [ 6, 6 ], "soundness_list": [ 3, 3 ], "presentation_list": [ 2, 3 ], "contribution_list": [ 3, 3 ], "confidence_list": [ 5, 4 ] }, { "paper_id": "xybTwSsdBP", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Beibei Chen", "gender": "Female", "institution": "Zhejiang Gongshang University", "country": "CN", "position": "Undergrad student" }, { "name": "Hao Chen", "gender": "unknown", "institution": "none", "country": "CN", "position": "Researcher" }, { "name": "Jianhang Ding", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Algorithm Engineer" }, { "name": "Siyu Liu", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "Yifan Ding", "gender": "unknown", "institution": "collective ai", "country": "CN", "position": "Researcher" }, { "name": "run zou", "gender": "Female", "institution": "nanjing university", "country": "CN", "position": "Researcher" }, { "name": "wenwu", "gender": "Male", "institution": "\u96f6\u77e9\u667a\u80fd", "country": "CN", "position": "Algorithm Engineer" }, { "name": "yun lou", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 15, "n_ref_uni": 32, "n_ref": 63, "n_ref_all": 78, "n_fig": 8, "n_tab": 8, "L_tab": 2138, "n_element_tab": 86, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 31, "n_element_tab_1": 4, "formula_len_all": 1033, "formula_len_all_1": 589, "len_all": 124067, "len_all_1": 62341, "len_abs": 1353, "len_title": 141, "len_sents": 33349, "len_sents_1": 28481, "n_sents": 280, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1369, "title": "OptBatch: Optimizing Instruction Tuning with Data Selection through Batch Stratified Sampling", "abs": "Instruction tuning has optimized the specialized capabilities of large language models (LLMs), but it often requires extensive datasets and prolonged training times. The challenge lies in developing specific capabilities by identifying useful data and efficiently fine-tuning. High-quality and diverse pruned data can help models achieve lossless performance at a lower cost. In this paper, we propose \\textbf{OptBatch}, a novel data selection method that focuses on the learnability of whole batch data rather than individual samples. OptBatch considers the coverage of the data distribution through stratified sampling and maximizes the relative distance between samples within a batch to enhance diversity. Furthermore, OptBatch utilizes Hessian gradient optimization to guide the selection strategy for subsequent batches. OptBatch effectively captures the intrinsic value of data curation, surpasses previous state-of-the-art methods, and demonstrates robust generalization performance across diverse downstream tasks and models. Extensive experiments reveal that OptBatch training in various pruning rates outperforms full dataset training, reducing computational cost by 20-40\\%. Additionally, evaluations using GPT-4 scores and other metrics for multi-turn dialogue, multilingual translation and QA tasks consistently demonstrate OptBatch's optimal performance.", "keywords": [ "data selection", "coreset", "gradients", "instruction tuning", "large language model" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "xy9yv5siYQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Tianrui Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Xiaolong Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Yang Fu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 19, "n_ref_uni": 28, "n_ref": 65, "n_ref_all": 70, "n_fig": 3, "n_tab": 5, "L_tab": 1990, "n_element_tab": 378, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 3643, "n_element_tab_1": 706, "formula_len_all": 1044, "formula_len_all_1": 800, "len_all": 104678, "len_all_1": 68819, "len_abs": 207, "len_title": 121, "len_sents": 27605, "len_sents_1": 26158, "n_sents": 206, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 940, "title": "Learning Dynamic 3D Gaussians from Monocular Videos without Camera Poses", "abs": "Dynamic scene reconstruction aims to recover the time-varying geometry and appearance of a dynamic scene. Existing methods, however, heavily rely on the existence of multiple-view captures or the accurate camera poses estimated by Structure from Motion (SfM) algorithms. To relax this constraint, we introduce a method capable of reconstructing generic dynamic scenes, from casually captured monocular videos without known camera poses. Unlike recent works that treat static and dynamic content separately, we propose a unified Hexplane-based Gaussian field to capture the complex effects of scene deformation and camera motion. The Hexplane decomposition enables feasible disentanglement for effective optimization. Combined with an efficient camera pose initialization strategy, our approach significantly improves view synthesis quality and camera pose estimation accuracy over previous methods, while enhancing computational efficiency.", "keywords": [ "Dynamic reconstruction", "camera pose estimation" ], "rating_list": [ 5, 3, 5, 8 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 1, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "xy6B5Fh2v7", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Fei Wang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Jiefeng Chen", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Ruoxi Sun", "gender": "Female", "institution": "Google", "country": "", "position": "Google" }, { "name": "Sercan O Arik", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Xingchen Wan", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 36, "n_ref": 78, "n_ref_all": 105, "n_fig": 7, "n_tab": 8, "L_tab": 1875, "n_element_tab": 192, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 599, "n_element_tab_1": 89, "formula_len_all": 71, "formula_len_all_1": 71, "len_all": 136666, "len_all_1": 67086, "len_abs": 1543, "len_title": 153, "len_sents": 39117, "len_sents_1": 31937, "n_sents": 294, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 106, "L_abs": 1617, "title": "Astute RAG: Overcoming Imperfect Retrieval Augmentation and Knowledge Conflicts for Large Language Models", "abs": "Retrieval augmented generation (RAG), while effectively integrating external knowledge to address the inherent limitations of large language models (LLMs), can be hindered by imperfect retrieval that contain irrelevant, misleading, or even malicious information. Previous studies have rarely connected the behavior of RAG through joint analysis, particularly regarding error propagation coming from imperfect retrieval and potential conflicts between LLMs' internal knowledge and external sources.\nThrough comprehensive and controlled analyses under realistic conditions, we find that imperfect retrieval augmentation is inevitable, common, and harmful. We identify the knowledge conflicts between LLM-internal and external knowledge from retrieval as a bottleneck to overcome imperfect retrieval in the post-retrieval stage of RAG.\nTo address this, we propose Astute RAG, a novel RAG approach designed to be resilient to imperfect retrieval augmentation. It adaptively elicits essential information from LLMs' internal knowledge, iteratively consolidates internal and external knowledge with source-awareness, and finalizes the answer according to information reliability.\nOur experiments with Gemini and Claude demonstrate the superior performance of Astute RAG compared to previous robustness-enhanced RAG approaches. Specifically, Astute RAG is the only RAG method that achieves performance comparable to or even surpassing conventional use of LLMs under the worst-case scenario. Further analysis reveals the effectiveness of \\method in resolving knowledge conflicts, thereby improving the trustworthiness of RAG.", "keywords": [ "Retrieval Augmented Generation", "Knowledge Conflicts" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "xxzukMsYs9", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Ruisi Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zechuan Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zongxin Yang", "gender": "Male", "institution": "Harvard Medical School, Harvard University", "country": "US", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 63, "n_ref": 113, "n_ref_all": 123, "n_fig": 6, "n_tab": 2, "L_tab": 552, "n_element_tab": 78, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1410, "n_element_tab_1": 89, "formula_len_all": 328, "formula_len_all_1": 309, "len_all": 154887, "len_all_1": 58323, "len_abs": 1342, "len_title": 113, "len_sents": 36022, "len_sents_1": 27978, "n_sents": 269, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1300, "title": "3D Object Manipulation in a Single Image Using Generative Models", "abs": "Object manipulation in images aims to not only edit the object presentation but also gift objects with motion. Previous methods encountered challenges in concurrently handling static editing and dynamic motion applications, while also struggling to achieve realism in object appearance and scene lighting. In this work, we introduce OMG3D, a novel framework that integrates the precise geometric control with the generative power of diffusion models, thus achieving significant enhancements in visual performance. Our framework converts 2D objects into 3D, enabling user-directed modifications and lifelike motions at the geometric level. To address texture realism, we propose CustomRefiner, a texture refinement module that pretrain a customized diffusion model to align the style and perspectives of coarse renderings with the original image. Additionally, we introduce IllumiCombiner, an lighting processing module that estimates and adjusts background lighting to match human visual perception, resulting in more realistic illumination. Extensive experiments demonstrate the outstanding visual performance of our approach in both static and dynamic scenarios. Remarkably, all these steps can be done using one NVIDIA 3090. The code and project page will be released upon acceptance of the paper.", "keywords": [ "3d object manipulation", "diffusion models", "image editing", "image animation" ], "rating_list": [ 6, 5, 8, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "xxSK3ZNAhh", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Jiang Bian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Lei Song", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Xianliang Yang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Yapu Zhang", "gender": "Female", "institution": "Beijing University of Technology", "country": "CN", "position": "Lecturer" } ], "n_formula": 4, "n_formula_1": 0, "n_ref_uni": 12, "n_ref": 40, "n_ref_all": 70, "n_fig": 8, "n_tab": 11, "L_tab": 11829, "n_element_tab": 1018, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1398, "n_element_tab_1": 133, "formula_len_all": 386, "formula_len_all_1": 0, "len_all": 191435, "len_all_1": 49057, "len_abs": 1514, "len_title": 169, "len_sents": 67227, "len_sents_1": 22331, "n_sents": 553, "n_sents_1": 149, "venue": "ICLR2025", "decision": "withdraw", "L_title": 121, "L_abs": 1531, "title": "HeurAgenix: A Multi-Agent LLM-Based Paradigm for Adaptive Heuristic Evolution and Selection in Combinatorial Optimization", "abs": "Combinatorial Optimization (CO) is a class of problems where the goal is to identify an optimal solution from a finite set of feasible solutions under specific constraints. Despite its ubiquity across industries, existing heuristic algorithms struggle with limited adaptability, complex parameter tuning, and limited generalization to novel problems. Recent approaches leveraging machine learning have made incremental improvements but remain constrained by extensive data requirements and reliance on historical problem-specific adjustments. Large Language Models (LLMs) offer a new paradigm to overcome these limitations due to their ability to generalize across domains, autonomously generate novel insights, and adapt dynamically to different problem contexts. To harness these capabilities, we introduce $\\textbf{HeurAgenix}$, a novel multi-agent hyper-heuristic framework that leverages LLMs to generate, evolve, evaluate, and select heuristics for solving CO problems. Our framework comprises four key agents: heuristic generation, heuristic evolution, benchmark evaluation, and heuristic selection. Each agent is designed to exploit specific strengths of LLMs, such as their capacity for synthesizing knowledge from diverse sources, autonomous decision-making, and adaptability to new problem instances. Experiments on both classic and novel CO tasks show that HeurAgenix significantly outperforms state-of-the-art approaches by enabling scalable, adaptable, and data-efficient solutions to complex optimization challenges.", "keywords": [ "Combinatorial Optimization; Heuristic Evolution; Heuristic Selection; Large Language Models" ], "rating_list": [ 3, 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 5, 3 ] }, { "paper_id": "xw4jtToUrf", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Jakob Nicolaus Foerster", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Kelvin Li", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Undergrad student" }, { "name": "Michael Beukman", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Perla Maiolino", "gender": "Female", "institution": "University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Uljad Berdica", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Alex Goldie", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Matthew Fellows", "gender": "Unspecified", "institution": "University of Oxford", "country": "", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 46, "n_ref": 88, "n_ref_all": 109, "n_fig": 10, "n_tab": 4, "L_tab": 3361, "n_element_tab": 422, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 660, "formula_len_all_1": 338, "len_all": 170426, "len_all_1": 52036, "len_abs": 1390, "len_title": 87, "len_sents": 37989, "len_sents_1": 24205, "n_sents": 250, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1398, "title": "Investigating Online RL in World Models", "abs": "Significant advances in online reinforcement learning (RL) remain limited by the need for extensive environment interaction or accurate simulators. World models trained on large-scale uncurated *offline data* could provide a training paradigm for generalist AI agents which alleviates the need for task specific simulation environments. Unfortunately, current offline RL methods rely on truncated rollouts that can lead to value overestimation and limit out-of-sample exploration. Additioanlly, common offline RL datasets have been shows to have a bias towards healthy behavior which does not help with the development of generalizable methods. We propose an algorithm and a data curation method that addresses both of these concerns by demonstrating that effective *full-length rollout* training is possible *without hand-crafted penalties* by treating each member of the world model ensemble as a level in the Unsupervised Environment Design (UED) framework. Our method achieves competitive performance even with less transitions than the same online algorithms are traditionally trained on. We find that training a recurrent policy on an ensemble of world models is sufficient to ensure transfer to the original environment and match online PPO performance on standard offline-RL benchmarks while maintaining robust performance on our dataset, where conventional offline RL methods underperform.", "keywords": [ "World models", "Domain Randomization", "Offline RL" ], "rating_list": [ 3, 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 1, 1, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "xvsNb5y9CN", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Tao Lu", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Associate Professor" }, { "name": "Yinghao Cai", "gender": "Female", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Yuxiang Zheng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 17, "n_ref_uni": 42, "n_ref": 57, "n_ref_all": 88, "n_fig": 10, "n_tab": 1, "L_tab": 520, "n_element_tab": 15, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 568, "n_element_tab_1": 69, "formula_len_all": 683, "formula_len_all_1": 782, "len_all": 123892, "len_all_1": 60227, "len_abs": 1006, "len_title": 175, "len_sents": 38751, "len_sents_1": 27931, "n_sents": 283, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 127, "L_abs": 1007, "title": "Sample-Imagined Generator: Efficient Virtual Sample Generation Method for Off-policy Reinforcement Learning with Sparse Rewards", "abs": "Off-policy reinforcement learning (RL) requires extensive real interaction with environment to gain experience for policy learning, presenting a challenge of low sample efficiency, especially in the condition of sparse rewards. To address this, we propose a Sample-Imagined Generator (SIG) which automatically trains a sample generator during environment interaction and could adaptively generate valuable imagined samples for policy learning. Through SIG, the policy greatly reduced the interaction with the environment during training and achieved comparable or even higher performance with those trained only through real interactions. SIG could be combined with any off-policy RL algorithm. Experiment in 5 continuous control tasks demonstrate that by substituting imagined samples for real ones to supplement the experience pool, SIG accomplishes tasks with significantly less interaction with the environment, notably improving sample efficiency across 10 off-policy reinforcement learning algorithms.", "keywords": [ "Off-policy Reinforcement Learning", "Sparse Reward Reinforcement Learning", "Sample Efficiency" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "xvhV3LvYTc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chenxin Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Panwang Pan", "gender": "Male", "institution": "ByteDance", "country": "CN", "position": "Researcher" }, { "name": "Wuyang Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Yifan Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Yixuan Yuan", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Zhiwen Fan", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Hengyu Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 24, "n_ref": 45, "n_ref_all": 55, "n_fig": 6, "n_tab": 4, "L_tab": 2607, "n_element_tab": 160, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1547, "n_element_tab_1": 69, "formula_len_all": 265, "formula_len_all_1": 262, "len_all": 119362, "len_all_1": 52104, "len_abs": 1571, "len_title": 142, "len_sents": 28830, "len_sents_1": 26337, "n_sents": 190, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1588, "title": "InstantSplamp: Fast and Generalizable Stenography Framework for Generative Gaussian Splatting", "abs": "With the rapid development of large generative models for 3D, especially the evolution from NeRF representations to more efficient Gaussian Splatting, the synthesis of 3D assets has become increasingly fast and efficient, enabling the large-scale publication and sharing of generated 3D objects. However, while existing methods can add watermarks or steganographic information to individual 3D assets, they often require time-consuming per-scene training and optimization, leading to watermarking overheads that can far exceed the time required for asset generation itself, making deployment impractical for generating large collections of 3D objects. To address this, we propose InstantSplamp a framework that seamlessly integrates the 3D steganography pipeline into large 3D generative models without introducing explicit additional time costs. Guided by visual foundation models,InstantSplamp subtly injects hidden information like copyright tags during asset generation, enabling effective embedding and recovery of watermarks within generated 3D assets while preserving original visual quality. Experiments across various potential deployment scenarios demonstrate that \\model~strikes an optimal balance between rendering quality and hiding fidelity, as well as between hiding performance and speed. Compared to existing per-scene optimization techniques for 3D assets, InstantSplamp reduces their watermarking training overheads that are multiples of generation time to nearly zero, paving the way for real-world deployment at scale. Project page: https://gaussian-stego.github.io/.", "keywords": [ "Gaussian Splatting", "3D Generation", "IP Verfication" ], "rating_list": [ 8, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 5, 3, 4 ] }, { "paper_id": "xvUVk9T3kZ", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Aviv Navon", "gender": "Male", "institution": "Bar Ilan University, Israel", "country": "IL", "position": "PhD student" }, { "name": "Aviv Shamsian", "gender": "Male", "institution": "aiOla", "country": "IL", "position": "Researcher" }, { "name": "Ethan Fetaya", "gender": "Male", "institution": "Bar Ilan University", "country": "IL", "position": "Assistant Professor" }, { "name": "Neta Glazer", "gender": "Female", "institution": "Bar-Ilan University", "country": "", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 21, "n_ref": 32, "n_ref_all": 59, "n_fig": 7, "n_tab": 7, "L_tab": 1877, "n_element_tab": 98, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 86, "n_element_tab_1": 4, "formula_len_all": 779, "formula_len_all_1": 783, "len_all": 128002, "len_all_1": 63389, "len_abs": 1118, "len_title": 65, "len_sents": 45243, "len_sents_1": 29946, "n_sents": 371, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1127, "title": "Multi Task Inverse Reinforcement Learning for Common Sense Reward", "abs": "One of the challenges in applying reinforcement learning in a complex real-world environment lies in providing the agent with a sufficiently detailed reward function. Any misalignment between the reward and the desired behavior can result in unwanted outcomes. This may lead to issues like \"reward hacking\" where the agent maximizes rewards by unintended behavior. In this work, we propose to disentangle the reward into two distinct parts. A simple task-specific reward, outlining the particulars of the task at hand, and an unknown common-sense reward, indicating the expected behavior of the agent within the environment. We then explore how this common-sense reward can be learned from expert demonstrations. We first show that inverse reinforcement learning, even when it succeeds in training an agent, does not learn a useful reward function. That is, training a new agent with the learned reward does not impair the desired behaviors. We then demonstrate that this problem can be solved by training simultaneously on multiple tasks. That is, multi-task inverse reinforcement learning can learn a useful reward function.", "keywords": [ "multi task learning", "reinforcement learning" ], "rating_list": [ 1, 1, 5, 5 ], "soundness_list": [ 1, 1, 2, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "xuQSp75HmP", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Junlin Xie", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Le Zhuo", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Renrui Zhang", "gender": "Male", "institution": "MMLab of CUHK & Shanghai AI Laboratory", "country": "HK", "position": "PhD student" }, { "name": "Shitian Zhao", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Undergrad student" }, { "name": "Siyuan Huang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Weifeng Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xinyu Wei", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Gao Peng", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 76, "n_ref": 152, "n_ref_all": 196, "n_fig": 10, "n_tab": 8, "L_tab": 3210, "n_element_tab": 385, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 733, "n_element_tab_1": 153, "formula_len_all": 504, "formula_len_all_1": 397, "len_all": 315540, "len_all_1": 53285, "len_abs": 2679, "len_title": 132, "len_sents": 61502, "len_sents_1": 24029, "n_sents": 516, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 84, "L_abs": 1311, "title": "PixWizard: Versatile Image-to-Image Visual Assistant with Open-Language Instructions", "abs": "This paper presents a versatile image-to-image visual assistant, PixWizard, designed for image generation, manipulation, and translation based on free-from language instructions. To this end, we tackle a variety of vision tasks into a unified image-text-to-image generation framework and curate an Omni Pixel-to-Pixel Instruction-Tuning Dataset. By constructing detailed instruction templates in natural language, we comprehensively include a large set of diverse vision tasks such as text-to-image generation, image restoration, image grounding, dense image prediction, image editing, controllable generation, inpainting/outpainting, and more. Furthermore, we adopt Diffusion Transformers (DiT) as our foundation model and extend its capabilities with a flexible any resolution mechanism, enabling the model to dynamically process images based on the aspect ratio of the input, closely aligning with human perceptual processes. The model also incorporates structure-aware and semantic-aware guidance to facilitate effective fusion of information from the input image. Our experiments demonstrate that PixWizard not only shows impressive generative and understanding abilities for images with diverse resolutions but also exhibits promising generalization capabilities with unseen tasks and human instructions.", "keywords": [ "Diffusion Model", "Image Generation", "Image-to-Image" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "xtzqU9FgSi", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Dmitry Kobak", "gender": "unknown", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "Researcher" }, { "name": "Philipp Berens", "gender": "Male", "institution": "University of Tuebingen", "country": "DE", "position": "Full Professor" }, { "name": "Rita Gonzalez Marquez", "gender": "unknown", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 89, "n_ref_all": 131, "n_fig": 13, "n_tab": 4, "L_tab": 2050, "n_element_tab": 511, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1100, "n_element_tab_1": 225, "formula_len_all": 63, "formula_len_all_1": 63, "len_all": 133982, "len_all_1": 61144, "len_abs": 1071, "len_title": 108, "len_sents": 39032, "len_sents_1": 28371, "n_sents": 270, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1076, "title": "Is self-supervision enough for training sentence embeddings?", "abs": "In NLP, sentence embeddings are crucial for many tasks such as information retrieval, classification, clustering, or visualizing collections of texts. Currently, top-performing sentence embeddings are derived from pre-trained language models that undergo extensive supervised fine-tuning. This contrasts with computer vision, where self-supervised training has demonstrated remarkable success. Here we show that self-supervision alone can produce high-quality sentence embeddings, albeit slightly below those from state-of-the-art supervised models. We systematically compare several existing augmentation strategies for positive pair generation in contrastive learning and show that text crops strongly outperform popular dropout-based augmentation. Using text crops, well-performing embeddings can be obtained even when training from scratch without using pre-trained model weights, or when training a bare token embedding layer without any transformer architecture. Overall, we show that self-supervised learning allows rapid training of text embeddings of a given dataset.", "keywords": [ "self-supervised learning", "language models", "contrastive learning", "transformers", "natural language processing" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 4, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "xtp6QPnwLu", "primary_area": "applications to robotics, autonomy, planning", "n_author": 11, "author_info_list": [ { "name": "Beiwen Tian", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Guanzhong Tian", "gender": "Male", "institution": " Ningbo Innovation Center, Zhejiang University.", "country": "CN", "position": "Assistant Professor" }, { "name": "Guyue Zhou", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Haizhou Ge", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hongrui Zhu", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Jiangning Zhang", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Longhua Ma", "gender": "Male", "institution": "NingboTech University", "country": "", "position": "Full Professor" }, { "name": "Ruixiang Wang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yufei Jia", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yuhang Dong", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yupei Zeng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 39, "n_ref": 67, "n_ref_all": 75, "n_fig": 4, "n_tab": 5, "L_tab": 1244, "n_element_tab": 217, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1244, "n_element_tab_1": 217, "formula_len_all": 255, "formula_len_all_1": 255, "len_all": 116906, "len_all_1": 59900, "len_abs": 1222, "len_title": 149, "len_sents": 28886, "len_sents_1": 26593, "n_sents": 205, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1224, "title": "Imit-Diff: Semantics Guided Diffusion Transformer with Dual Resolution Fusion for Imitation Learning", "abs": "Diffusion-based methods have become one of the most important paradigms in the field of imitation learning. However, even in state-of-the-art diffusion-based policies, there has been insufficient focus on semantics and fine-grained feature extraction, resulting in weaker generalization and a reliance on controlled environments. To address this issue, we propose Imit-Diff, which consists of three key components: 1) Dual Resolution Fusion for extracting fine-grained features with a manageable number of tokens by integrating high-resolution features into low-resolution visual embedding through an attention mechanism; 2) Semantics Injection to explicitly incorporate semantic information by using prior masks obtained from open vocabulary models, achieving a world-level understanding of imitation learning tasks; and 3) Consistency Policy on Diffusion Transformer to reduce the inference time of diffusion models by training a student model to implement few-step denoising on the Probability Flow ODE trajectory. Experimental results show that our method significantly outperforms state-of-the-art methods, especially in cluttered scenes, and is highly robust to task interruptions. The code will be publicly available.", "keywords": [ "Imitation learning", "Diffusion Policy", "Dual Resolution", "Semantics Injection" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "xtlMtbVfWu", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 6, "author_info_list": [ { "name": "Jiadi Jiang", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Algorithm Engineer" }, { "name": "Jialiang Cheng", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Algorithm Engineer" }, { "name": "Ning Gao", "gender": "unknown", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "", "position": "MS student" }, { "name": "YUN YUE", "gender": "Male", "institution": "antgroup", "country": "", "position": "Algorithm engineer" }, { "name": "Zhiling Ye", "gender": "Male", "institution": "Ant Group", "country": "", "position": "Engineer" }, { "name": "Sha Jian", "gender": "unknown", "institution": "Ant Group", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 38, "n_formula_1": 15, "n_ref_uni": 32, "n_ref": 74, "n_ref_all": 106, "n_fig": 13, "n_tab": 10, "L_tab": 2803, "n_element_tab": 375, "n_fig_1": 11, "n_tab_1": 5, "L_tab_1": 5313, "n_element_tab_1": 234, "formula_len_all": 2953, "formula_len_all_1": 756, "len_all": 144954, "len_all_1": 68764, "len_abs": 1428, "len_title": 89, "len_sents": 42605, "len_sents_1": 27711, "n_sents": 332, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1533, "title": "EDiT: A Local-SGD-Based Efficient Distributed Training Method for Large Language Models", "abs": "Distributed training methods are crucial for large language models (LLMs). However, existing distributed training methods often suffer from communication bottlenecks, stragglers, and limited elasticity, particularly in heterogeneous or large-scale environments. Local SGD methods have been proposed to address these issues, but their effectiveness remains limited to small-scale training due to additional memory overhead and lack of concerns on efficiency and stability. To tackle these issues, we propose EDiT, an innovative Efficient Distributed Training method that combines a tailored Local SGD approach with model sharding techniques to enhance large-scale training efficiency. EDiT performs layer-wise parameter synchronization during forward pass, reducing communication and memory overhead and enabling overlap. Besides, EDiT employs a pseudo gradient penalty strategy to suppress loss spikes, which ensures training stability and improves performance. Additionally, we introduce A-EDiT, a fully asynchronous variant of EDiT that accommodates heterogeneous clusters. Building on EDiT/A-EDiT, we conduct a series of experiments to validate large-scale asynchronous training for LLMs, accompanied by comprehensive analyses. Experimental results demonstrate the superior performance of EDiT/A-EDiT, establishing them as robust solutions for distributed LLM training in diverse computational ecosystems. The code is available at Atorch codebase: https://github.com/intelligent-machine-learning/atorch/tree/main/atorch/local_sgd.", "keywords": [ "Distributed Training", "Large Language Models", "Local SGD", "Training Acceleration" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "xtTut5lisc", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Dongjie Wang", "gender": "Male", "institution": "University of Kansas", "country": "US", "position": "Assistant Professor" }, { "name": "Kunpeng Liu", "gender": "Male", "institution": "Portland State University", "country": "US", "position": "Assistant Professor" }, { "name": "Xiao Luo", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Yanjie Fu", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Associate Professor" }, { "name": "Yanping Wu", "gender": "Female", "institution": "University of Glasgow", "country": "GB", "position": "PhD student" }, { "name": "Yanyong Huang", "gender": "Male", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "Full Professor" }, { "name": "Zhengzhang Chen", "gender": "Male", "institution": "NEC Labs America", "country": "US", "position": "Senior Research Scientist" }, { "name": "Zijun Yao", "gender": "Male", "institution": "University of Kansas", "country": "US", "position": "Assistant Professor" } ], "n_formula": 36, "n_formula_1": 28, "n_ref_uni": 42, "n_ref": 51, "n_ref_all": 79, "n_fig": 14, "n_tab": 7, "L_tab": 7128, "n_element_tab": 402, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 3132, "n_element_tab_1": 211, "formula_len_all": 2015, "formula_len_all_1": 1546, "len_all": 168974, "len_all_1": 74109, "len_abs": 1966, "len_title": 124, "len_sents": 50241, "len_sents_1": 30360, "n_sents": 420, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 2019, "title": "Iterative Feature Space Optimization through Incremental Adaptive Evaluation", "abs": "Iterative feature space optimization involves systematically evaluating and adjusting the feature space to improve downstream task performance. However, existing works suffer from three key limitations: 1) overlooking differences among data samples leads to evaluation bias; 2) tailoring feature spaces to specific machine learning models results in overfitting and poor generalization; 3) requiring\nthe evaluator to be retrained from scratch during each optimization iteration significantly reduces the overall efficiency of the optimization process. To bridge these gaps, we propose a gEneralized Adaptive feature Space Evaluator (EASE) to efficiently produce optimal and generalized feature spaces. This framework consists of two key components: Feature-Sample Subspace Generator and Contextual Attention Evaluator. The first component aims to decouple the information distribution within the feature space to mitigate evaluation bias. To achieve this, we first identify features most relevant to prediction tasks and samples most challenging for evaluation based on feedback from the subsequent evaluator. These identified feature and samples are then used to construct feature subspaces for next optimization iteration. This decoupling strategy makes the evaluator consistently target the most challenging aspects of the feature space. The second component intends to incrementally capture evolving patterns of the feature space for efficient evaluation. We propose a weighted-sharing multi-head attention mechanism to encode key characteristics of the feature space into an embedding vector for evaluation. Moreover, the evaluator is updated incrementally, retaining prior evaluation knowledge while incorporating new insights, as consecutive feature spaces during the optimization process share partial information. Extensive experiments on twelve real-world datasets demonstrate the effectiveness of the proposed framework. Our code and data are publicly available (https://anonymous.4open.science/r/EASE-1C51).", "keywords": [ "Automated Feature Optimization", "Incremental Learning", "Feature Space Evaluator" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "xt3mCoDks7", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 10, "author_info_list": [ { "name": "Hao Zhou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Hongbo Ma", "gender": "Male", "institution": "Department of Computer Science and Technology, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Jie Yu", "gender": "Male", "institution": "ShanghaiTech University", "country": "", "position": "PhD student" }, { "name": "Keyue Qiu", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Mingyue Zheng", "gender": "Male", "institution": "Shanghai Institute of Materia Medica", "country": "", "position": "Full Professor" }, { "name": "Wei-Ying Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yushuai Wu", "gender": "Male", "institution": "Shanghai Academy of Artificial Intelligence for Science", "country": "CN", "position": "Researcher" }, { "name": "Yuxuan Song", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Intern" }, { "name": "Zhilong Zhang", "gender": "unknown", "institution": "Institute for AI Industry Research (AIR), Tsinghua University", "country": "CN", "position": "Intern" }, { "name": "Ziyao Cao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 21, "n_ref_uni": 43, "n_ref": 104, "n_ref_all": 135, "n_fig": 11, "n_tab": 13, "L_tab": 6723, "n_element_tab": 793, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2003, "n_element_tab_1": 207, "formula_len_all": 4336, "formula_len_all_1": 2958, "len_all": 193002, "len_all_1": 63435, "len_abs": 1556, "len_title": 130, "len_sents": 57630, "len_sents_1": 26169, "n_sents": 403, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1362, "title": "Unlocking the Power of Gradient Guidance for Structure-Based Molecule Optimization", "abs": "Structure-based molecule optimization (SBMO) aims to optimize molecules with both continuous coordinates and discrete types against protein targets.\nA promising direction is to exert gradient guidance on generative models given its remarkable success in images, but it is challenging to guide discrete data and risks inconsistencies between modalities.\nTo this end, we leverage a continuous and differentiable space derived through Bayesian inference, presenting Molecule Joint Optimization (MolJO), the first gradient-based SBMO framework that facilitates joint guidance signals across different modalities while preserving SE(3)-equivariance.\nWe introduce a novel backward correction strategy that optimizes within a sliding window of the past histories, allowing for a seamless trade-off between explore-and-exploit during optimization.\nOur proposed MolJO achieves state-of-the-art performance on CrossDocked2020 benchmark (Success Rate 51.3% , Vina Dock -9.05 and SA 0.78), more than 4x improvement in Success Rate compared to the gradient-based counterpart, and 2x \"Me-Better\" Ratio as much as 3D baselines.\nFurthermore, we extend MolJO to a wide range of optimization settings, including multi-objective optimization and challenging tasks in drug design such as R-group optimization and scaffold hopping, further underscoring its versatility and potential.", "keywords": [ "molecule optimization", "structure-based drug design", "Bayesian flow network" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 2, 4, 2 ] }, { "paper_id": "xsx3Fpo3UD", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Fanqi Wan", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Jiajian Guo", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Qifan Wang", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Principal Researcher" }, { "name": "Shiping Gao", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Xiaojun Quan", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" } ], "n_formula": 35, "n_formula_1": 11, "n_ref_uni": 44, "n_ref": 69, "n_ref_all": 86, "n_fig": 3, "n_tab": 6, "L_tab": 2460, "n_element_tab": 378, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2104, "n_element_tab_1": 343, "formula_len_all": 3301, "formula_len_all_1": 1011, "len_all": 174168, "len_all_1": 56195, "len_abs": 1251, "len_title": 128, "len_sents": 50901, "len_sents_1": 23988, "n_sents": 365, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 79, "L_abs": 1466, "title": "Advantage-Guided Distillation for Preference Alignment in Small Language Models", "abs": "Alignment techniques enable Large Language Models (LLMs) to generate outputs that align with human preferences and play a crucial role in their effectiveness. However, their impact often diminishes when applied to Small Language Models (SLMs), likely due to the limited capacity of these models. Instead of directly applying existing alignment techniques to SLMs, we propose to utilize a well-aligned teacher LLM to guide the alignment process for these models, thereby facilitating the transfer of the teacher's knowledge of human preferences to the student model. To achieve this, we first explore a straightforward approach, Dual-Constrained Knowledge Distillation (DCKD), that employs knowledge distillation with two KL-divergence constraints from the aligned teacher to the unaligned student. To further enhance the student's ability to distinguish between preferred and dispreferred responses, we then propose Advantage-Guided Distillation for Preference Alignment (ADPA), which leverages an advantage function from the aligned teacher to deliver more nuanced, distribution-level reward signals for the student's alignment. Our experimental results show that these two approaches appreciably improve the alignment of SLMs and narrow the performance gap with larger counterparts. Among them, ADPA demonstrates superior performance and achieves even greater effectiveness when integrated with DCKD. Our code is available at \\url{https://github.com/SLIT-AI/ADPA}.", "keywords": [ "Preference Alignment; Large language model; Knowledge Distillation; Advantage Function" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "xsmlrhoQzC", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Been Kim", "gender": "unknown", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Kartikeya Badola", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Meera Hahn", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Nithish Kannen", "gender": "Male", "institution": "Google DeepMind", "country": "IN", "position": "Researcher" }, { "name": "Rich Galt", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Wenjun Zeng", "gender": "Not Specified", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Zi Wang", "gender": "Female", "institution": "Google DeepMind", "country": "US", "position": "Research scientist" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 69, "n_ref": 112, "n_ref_all": 144, "n_fig": 13, "n_tab": 7, "L_tab": 1864, "n_element_tab": 159, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 811, "n_element_tab_1": 51, "formula_len_all": 131, "formula_len_all_1": 76, "len_all": 268190, "len_all_1": 69089, "len_abs": 1055, "len_title": 122, "len_sents": 83499, "len_sents_1": 33269, "n_sents": 662, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1502, "title": "Proactive Agents for Multi-Turn Text-to-Image Generation Under Uncertainty", "abs": "User prompts for generative AI models are often underspecified or open-ended, which may lead to sub-optimal responses. This prompt underspecification problem is particularly evident in text-to-image (T2I) generation, where users commonly struggle to articulate their precise intent. This disconnect between the user's vision and the model's interpretation often forces users to painstakingly and repeatedly refine their prompts. To address this, we propose a design for proactive T2I agents equipped with an interface to actively ask clarification questions when uncertain, and present their understanding of user intent as an interpretable **belief graph** that a user can edit. We build simple prototypes for such agents and verify their effectiveness through both human studies and automated evaluation. We observed that at least 90\\% of human subjects found these agents and their belief graphs helpful for their T2I workflow. Moreover, we use a scalable automated evaluation approach using two agents, one with a ground truth image and the other tries to ask as few questions as possible to align with the ground truth. On DesignBench, a benchmark we created for artists and designers, the COCO dataset (Lin et al.,2014) and ImageInWords (Garg et al., 2024), we observed that these T2I agents were able to ask informative questions and elicit crucial information to achieve successful alignment with at least 2 times higher VQAScore (Lin et al., 2024) than the standard single-turn T2I generation.", "keywords": [ "Interpretable belief state", "uncertainty estimation", "information gathering", "intelligent agents", "question-asking under uncertainty" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "xsELpEPn4A", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Lianghui Zhu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Xinggang Wang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xinlong Wang", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 42, "n_ref": 115, "n_ref_all": 183, "n_fig": 21, "n_tab": 20, "L_tab": 6928, "n_element_tab": 1158, "n_fig_1": 13, "n_tab_1": 6, "L_tab_1": 1506, "n_element_tab_1": 274, "formula_len_all": 421, "formula_len_all_1": 0, "len_all": 281619, "len_all_1": 60656, "len_abs": 2364, "len_title": 109, "len_sents": 88456, "len_sents_1": 25110, "n_sents": 749, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 61, "L_abs": 1510, "title": "JudgeLM: Fine-tuned Large Language Models are Scalable Judges", "abs": "Evaluating Large Language Models (LLMs) in open-ended scenarios is challenging because existing benchmarks and metrics can not measure them comprehensively. To address this problem, we propose to fine-tune LLMs as scalable judges (JudgeLM) to evaluate LLMs efficiently and effectively in open-ended benchmarks. We first propose a comprehensive, large-scale, high-quality dataset containing task seeds, LLMs-generated answers, and GPT-4-generated judgments for fine-tuning high-performance judges, as well as a new benchmark for evaluating the judges. We train JudgeLM at different scales from 7B, 13B, to 33B parameters, and conduct a systematic analysis of its capabilities and behaviors. We then analyze the key biases in fine-tuning LLM as a judge and consider them as position bias, knowledge bias, and format bias. To address these issues, JudgeLM introduces a bag of techniques including swap augmentation, reference support, and reference drop, which clearly enhance the judge's performance. JudgeLM obtains the state-of-the-art judge performance on both the existing PandaLM benchmark and our proposed new benchmark. Our JudgeLM is efficient and the JudgeLM-7B only needs 3 minutes to judge 5K samples with 8 A100 GPUs. JudgeLM obtains high agreement with the teacher judge, achieving an agreement exceeding 90% that even surpasses human-to-human agreement. JudgeLM also demonstrates extended capabilities in being judges of the single answer, multimodal models, multiple answers, multi-turn chat, etc.", "keywords": [ "LLM Judging" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "xrtM8r0zdU", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "David H. Yang", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "PhD student" }, { "name": "Mohammad Mohammadi Amiri", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "Assistant Professor" }, { "name": "Pin-Yu Chen", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Subhajit Chaudhury", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Research Scientist" }, { "name": "Tejaswini Pedapati", "gender": "Female", "institution": "International Business Machines", "country": "US", "position": "Research Engineer" } ], "n_formula": 32, "n_formula_1": 25, "n_ref_uni": 38, "n_ref": 56, "n_ref_all": 75, "n_fig": 4, "n_tab": 2, "L_tab": 1030, "n_element_tab": 156, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 786, "n_element_tab_1": 145, "formula_len_all": 1952, "formula_len_all_1": 1138, "len_all": 127065, "len_all_1": 70643, "len_abs": 1458, "len_title": 113, "len_sents": 38785, "len_sents_1": 30767, "n_sents": 294, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1464, "title": "Sparse Gradient Compression for Fine-Tuning Large Language Models", "abs": "Fine-tuning large language models (LLMs) for downstream tasks has become increasingly crucial due to their widespread use and the growing availability of open-source models. However, the high memory costs associated with fine-tuning remain a significant challenge, especially as models increase in size. To address this, parameter efficient fine-tuning (PEFT) methods have been proposed to minimize the number of parameters required for fine-tuning LLMs. However, these approaches often tie the number of optimizer states to dimensions of model parameters, limiting flexibility and control during fine-tuning. In this paper, we propose sparse gradient compression (SGC), a training regime designed to address these limitations. Our approach leverages inherent sparsity in gradients to compress optimizer states by projecting them onto a low-dimensonal subspace, with dimensionality independent of the original model's parameters. By enabling optimizer state updates in an arbitrary low-dimensional subspace, SGC offers a flexible tradeoff between memory efficiency and performance. We demonstrate through experiments that SGC can decrease memory usage in optimizer states more effectively than exising PEFT methods. Furthermore, by fine-tuning LLaMA models on various downstream tasks, we show that SGC can deliver superior performance while substantially lowering optimizer state memory requirements, particularly in both data-limited and memory-limited settings.", "keywords": [ "Machine Learning", "Large Language Models", "Parameter efficient fine-tuning" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 3, 4, 5 ] }, { "paper_id": "xriJVaTh4C", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Marc Fischer", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Mark Niklas Mueller", "gender": "Male", "institution": "LogicStar AI", "country": "CH", "position": "Researcher" }, { "name": "Martin Vechev", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Maximilian Baader", "gender": "unknown", "institution": "ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Stefan Balauca", "gender": "Male", "institution": "INSAIT", "country": "BG", "position": "PhD student" }, { "name": "Yuhao Mao", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 12, "n_ref_uni": 34, "n_ref": 96, "n_ref_all": 140, "n_fig": 16, "n_tab": 13, "L_tab": 5113, "n_element_tab": 710, "n_fig_1": 13, "n_tab_1": 4, "L_tab_1": 1220, "n_element_tab_1": 161, "formula_len_all": 3174, "formula_len_all_1": 624, "len_all": 187725, "len_all_1": 78077, "len_abs": 1295, "len_title": 128, "len_sents": 58757, "len_sents_1": 34476, "n_sents": 422, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1297, "title": "Gaussian Loss Smoothing Enables Certified Training with Tight Convex Relaxations", "abs": "Training neural networks with high certified accuracy against adversarial examples remains an open challenge despite significant efforts. While certification methods can effectively leverage tight convex relaxations for bound computation, in training, these methods, perhaps surprisingly, can perform worse than looser relaxations. Prior work hypothesized that this phenomenon is caused by the discontinuity, non-smoothness and perturbation sensitivity of the loss surface induced by tighter relaxations. In this work, we theoretically show that Gaussian Loss Smoothing (GLS) can alleviate these issues. We confirm this empirically by instantiating GLS with two variants: a zeroth-order optimization algorithm called PGPE which allows training with non-differentiable relaxations, and a first-order optimization algorithm, called RGS, which requires gradients of the relaxation, but is much more efficient than PGPE. Extensive experiments show that when combined with tight relaxations, these methods surpass state-of-the-art methods when training on the same network architecture for many settings. Our results clearly demonstrate the promise of Gaussian Loss Smoothing for training certifiably robust neural networks and pave a path towards leveraging tighter relaxations for certified training.", "keywords": [ "Certified Robustness", "Adversarial Robustness", "Certified Training", "Convex Relaxation", "Neural Network Verification" ], "rating_list": [ 3, 1, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 1, 1, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "xrgXaOV6dK", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Arduin Findeis", "gender": "Not Specified", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Floris Weers", "gender": "Male", "institution": "Apple", "country": "NL", "position": "Researcher" }, { "name": "Guoli Yin", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Ke Ye", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Ruoming Pang", "gender": "unknown", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Tom Gunter", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 72, "n_ref_all": 99, "n_fig": 11, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 404, "n_element_tab_1": 5, "formula_len_all": 37, "formula_len_all_1": 0, "len_all": 159760, "len_all_1": 59719, "len_abs": 1873, "len_title": 124, "len_sents": 52063, "len_sents_1": 29346, "n_sents": 404, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1883, "title": "Can External Validation Tools Improve Annotation Quality for LLM-as-a-Judge?", "abs": "Pairwise preferences over model responses are widely collected to evaluate and provide feedback to large language models (LLMs). Given two alternative model responses to the same input, a human or AI annotator selects the \u201cbetter\u201d response. This approach can provide feedback for domains where other hard-coded metrics are difficult to obtain (e.g., quality of a chat interactions), thereby helping measure model progress or model fine-tuning (e.g., via reinforcement learning from human feedback, RLHF). However, for some domains it can be tricky to obtain such pairwise comparisons in high quality - from AI and humans. For example, for responses with many factual statements or complex code, annotators may overly focus on simpler features such as writing quality rather the underlying facts or technical details. In this work, we explore augmenting standard AI annotator systems with additional tools to improve performance on three challenging response domains: long-form factual, math and code tasks. We propose a tool-using agentic system to provide higher quality feedback on these domains. Our system uses web-search and code execution to ground itself based on external validation, independent of the LLM\u2019s internal knowledge and biases. We provide extensive experimental results evaluating our method across the three targeted response domains as well as general annotation tasks, using RewardBench data (incl. AlpacaEval and LLMBar), as well as three new datasets for areas where pre-existing datasets are saturated. Our results indicate that external tools can indeed improve AI annotator performance in many, but not all, cases. More generally, our experiments highlight the high variability of AI annotator performance with respect to simple parameters (e.g., prompt) and the need for improved (non-saturated) annotator benchmarks. We share our data and code publicly.", "keywords": [ "LLM-as-a-Judge", "AI annotators", "evaluation", "tool-use" ], "rating_list": [ 6, 8, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "xreOs2yjqf", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Cheng Zhang", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" }, { "name": "Luozheng Qin", "gender": "Male", "institution": "Soochow University", "country": "CN", "position": "MS student" }, { "name": "Mengping Yang", "gender": "Male", "institution": "East China University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "xiaomeng yang", "gender": "Male", "institution": "Shanghai Academy of Artificial Intelligence for Science", "country": "CN", "position": "Researcher" }, { "name": "Li Hao", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Zhiyu Tan", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 6, "n_ref_uni": 29, "n_ref": 76, "n_ref_all": 90, "n_fig": 3, "n_tab": 6, "L_tab": 1226, "n_element_tab": 102, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 864, "n_element_tab_1": 39, "formula_len_all": 436, "formula_len_all_1": 254, "len_all": 150172, "len_all_1": 69333, "len_abs": 1232, "len_title": 110, "len_sents": 38403, "len_sents_1": 30022, "n_sents": 261, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1239, "title": "EvalAlign: Supervised Fine-Tuning Multimodal LLMs with Human-Aligned Data for Evaluating Text-to-Image Models", "abs": "The recent advancements in text-to-image generative models have been remarkable. Yet, the field suffers from a lack of evaluation metrics that accurately reflect the performance of these models, particularly lacking fine-grained metrics that can guide the optimization of the models. In this paper, we propose EvalAlign, a metric characterized by its accuracy, stability, and fine granularity. Our approach leverages the capabilities of Multimodal Large Language Models (MLLMs) pre-trained on extensive data. We develop evaluation protocols that focus on two key dimensions: image faithfulness and text-image alignment. Each protocol comprises a set of detailed, fine-grained instructions linked to specific scoring options, enabling precise manual scoring of the generated images. We supervised fine-tune (SFT) the MLLM to align with human evaluative judgments, resulting in a robust evaluation model. Our evaluation across 24 text-to-image generation models demonstrate that EvalAlign not only provides superior metric stability but also aligns more closely with human preferences than existing metrics, confirming its effectiveness and utility in model assessment. We will make the code, data, and pre-trained models publicly available.", "keywords": [ "Text-to-Image Generative Models", "Evaluation Metrics", "Multimodal Large Language Models (MLLMs)", "Text-Image Consistency", "Image Generation Fidelity", "Supervised Fine-Tuning (SFT)", "Human Evaluative Judgments" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "xrazpGhJ10", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Ashutosh Jadhav", "gender": "Male", "institution": "IBM, International Business Machines", "country": "", "position": "Researcher" }, { "name": "David Beymer", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Hongzhi Wang", "gender": "Male", "institution": "IBM, International Business Machines", "country": "", "position": "Researcher" }, { "name": "Joy T Wu", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Intern" }, { "name": "Ken C. L. Wong", "gender": "Male", "institution": "IBM, International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Luyao Shi", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Raziuddin Mahmood", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "PhD student" }, { "name": "Satyananda Kashyap", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Research Staff Member" }, { "name": "Tanveer Syeda-mahmood", "gender": "Female", "institution": "International Business Machines", "country": "US", "position": "IBM Fellow" }, { "name": "Niharika Shimona Dsouza", "gender": "Female", "institution": "International Business Machines", "country": "US", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 23, "n_ref": 42, "n_ref_all": 93, "n_fig": 10, "n_tab": 6, "L_tab": 3198, "n_element_tab": 610, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1203, "n_element_tab_1": 214, "formula_len_all": 723, "formula_len_all_1": 643, "len_all": 125218, "len_all_1": 62113, "len_abs": 1460, "len_title": 143, "len_sents": 42721, "len_sents_1": 29023, "n_sents": 299, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1513, "title": "SemCLIP: Aligning vision-language encoder models to semantic spaces for stability in retrieval", "abs": "Vision-language models (VLM) bring image and textual representations close together in a joint embedding space to tackle many tasks ranging from image captioning to text-to-image retrieval. For such models to be reliably used in cloud vector stores, it is important to have a stable association between images and text such that synonymous queries bring up the same images or have a high degree of overlap. Current textual representations based on transformer models used to build the VLMs cannot adequately capture linguistic similarities to ensure such stability. In this paper we develop a database of linguists-curated similarity list of words derived from Wordnet, and train a semantics preserving textual embedding. We then train an alignment transformation to map existing VLM (CLIP) embeddings to bring synonymous embeddings closer while also preserving image-text similarities. The alignment transform is learned from textual embeddings alone thus avoiding large-scale retraining of VLMs from image-text pairs. This simple method outperforms other methods of creating image-joint text embeddings including even those by fine-tuning the encoders using the same synonyms lists. Results of analysis and comparison on multiple benchmark datasets is indicating both stable and improved quality of retrieval. The dataset of similarity lists and the semantics-preserve textual embedding itself can be employed in a variety of ways for other downstream tasks and will be made available for other researchers.", "keywords": [ "Semantic-preserving queries", "Vision-language encoder models", "Stability of retrieval", "joint embeddings" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "xrXci5YGm7", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Francois Charton", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Engineer" }, { "name": "Julia Kempe", "gender": "Female", "institution": "New York University", "country": "US", "position": "Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 35, "n_ref": 54, "n_ref_all": 78, "n_fig": 5, "n_tab": 14, "L_tab": 1639, "n_element_tab": 198, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 679, "n_element_tab_1": 102, "formula_len_all": 285, "formula_len_all_1": 0, "len_all": 157140, "len_all_1": 65316, "len_abs": 880, "len_title": 90, "len_sents": 56401, "len_sents_1": 32125, "n_sents": 403, "n_sents_1": 234, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 890, "title": "Emergent properties with repeated examples", "abs": "We study the performance of transformers as a function of the number of repetitions of training examples with algorithmically generated datasets. On three problems of mathematics: the greatest common divisor, modular multiplication, and matrix eigenvalues, we show that for a fixed number of training steps, models trained on smaller sets of repeated examples outperform models trained on larger sets of single-use examples. We also demonstrate that {\\em two-set training} - repeated use of a small random subset of examples, along normal sampling on the rest of the training set - provides for faster learning and better performance. This highlights that the benefits of repetition can outweigh those of data diversity. These datasets and problems provide a controlled setting to shed light on the still poorly understood interplay between generalization and memorization in deep learning.", "keywords": [ "transformers", "learning on repeated examples", "emergence" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "xrWOR5wSOz", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Abhinav Narayan Harish", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Brahma S Pavse", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "PhD student" }, { "name": "Josiah Hanna", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 11, "n_ref_uni": 28, "n_ref": 56, "n_ref_all": 85, "n_fig": 14, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2475, "formula_len_all_1": 789, "len_all": 134107, "len_all_1": 67248, "len_abs": 1278, "len_title": 140, "len_sents": 44453, "len_sents_1": 32473, "n_sents": 347, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1257, "title": "Replacing Implicit Regression with Classification in Policy Gradient Reinforcement Learning", "abs": "Stochastic policy gradient methods are a fundamental class of reinforcement learning algorithms. When using these algorithms for continuous control it is common to parameterize the policy using a Gaussian distribution. In this paper, we show that the policy gradient with Gaussian policies can be viewed as the gradient of a weighted least-squares objective function. That is, policy gradient algorithms are implicitly implementing a form of regression. A number of recent works have shown that reformulating regression problems as classification problems can improve learning. Inspired by these works, we investigate whether replacing this implicit regression with classification can improve the data efficiency and stability of policy learning. Toward this end, we introduce a novel policy gradient surrogate objective for softmax policies over a discretized action space. This surrogate objective uses a form of cross-entropy loss as a replacement for the implicit least-squares loss found in the surrogate loss for Gaussian policies. We extend prior theoretical analysis of this loss to our policy gradient surrogate objective and then provide experiments showing that this novel loss improves the data efficiency of stochastic policy gradient learning.", "keywords": [ "reinforcement learning; policy gradient RL; actor-critic" ], "rating_list": [ 8, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "xqEeGja6zq", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Alain Ryser", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Alice Bizeul", "gender": "Female", "institution": "ETHZ - ETH Zurich", "country": "", "position": "PhD student" }, { "name": "Julia E Vogt", "gender": "Female", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Assistant Professor" }, { "name": "Julius Von K\u00fcgelgen", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Thomas Marco Sutter", "gender": "unknown", "institution": "ETH Zurich", "country": "CH", "position": "Postdoc" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 110, "n_ref_all": 137, "n_fig": 11, "n_tab": 7, "L_tab": 1640, "n_element_tab": 244, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 908, "n_element_tab_1": 118, "formula_len_all": 263, "formula_len_all_1": 187, "len_all": 147131, "len_all_1": 65626, "len_abs": 1155, "len_title": 127, "len_sents": 45492, "len_sents_1": 31422, "n_sents": 334, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1160, "title": "Components Beat Patches: Eigenvector Removal for Robust Masked Image Modelling", "abs": "Masked Image Modeling has gained prominence as a powerful self-supervised learning approach for visual representation learning by reconstructing masked-out patches of images. However, the use of random spatial masking can lead to failure cases in which the learned features are not predictive of downstream labels. In this work, we introduce a novel masking strategy that targets principal components instead of image patches. The learning task then amounts to reconstructing the information of masked-out principal components. The principal components of a dataset contain more global information than patches, such that the information shared between the masked input and the reconstruction target should involve more high-level variables of interest. This property allows principal components to offer a more meaningful masking space, which manifests in improved quality of the learned representations. We provide empirical evidence across natural and medical datasets and demonstrate substantial improvements in image classification tasks. Our method thus offers a simple and robust data-driven alternative to traditional Masked Image Modelling approaches.", "keywords": [ "Self-supervised Representation Learning; Unsupervised Representation Learning; Visual Representation Learning" ], "rating_list": [ 3, 8, 3, 8 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 1, 3, 1, 3 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "xpmDc76RN2", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Myeong-Su Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KP", "position": "Postdoc" }, { "name": "Park Dongmin", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Undergrad student" }, { "name": "Youngjoon Hong", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" } ], "n_formula": 82, "n_formula_1": 33, "n_ref_uni": 29, "n_ref": 64, "n_ref_all": 80, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9026, "formula_len_all_1": 2860, "len_all": 130409, "len_all_1": 63722, "len_abs": 1315, "len_title": 87, "len_sents": 40983, "len_sents_1": 27665, "n_sents": 367, "n_sents_1": 226, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1329, "title": "Understanding Optimization of Operator Networks with Variational Loss for Solving PDEs", "abs": "In this paper, we analyze the optimization of operator networks for solving elliptic PDEs with variational loss functions. While approximation and generalization errors in operator networks have been extensively studied, optimization error remains largely unexplored. \nWe apply Restricted Strong Convexity (RSC) theory to rigorously examine the optimization dynamics of operator networks trained with variational loss, providing theoretical guarantees for convergence and training stability. \nWe further investigate the role of the condition number of $A$ in optimization and demonstrate that preconditioning strategies significantly improve convergence rates, establishing a solid theoretical basis for the empirical benefits of preconditioning. We also address the lower bound of a key quantity, $q_t$, which ensures convergence. \nTo prevent $q_t$ from vanishing, we propose an algorithm that adaptively incorporates additional weights into the variational loss function, leveraging values already computed during training, thereby avoiding any extra computational costs.\nFinally, we validate {our theoretical assumptions through numerical experiments, demonstrating their practical applicability} and confirming the effectiveness of preconditioning, with significant improvements in training performance and convergence rates.", "keywords": [ "Restriced Strong Convexity", "Operator Learning", "Variational Loss", "Scientific machine learning" ], "rating_list": [ 3, 1, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 1 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "xom3YUQfbK", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Shashidhar Reddy Javaji", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Shravan Doda", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Zining Zhu", "gender": "unknown", "institution": "Stevens Institute of Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 24, "n_ref": 35, "n_ref_all": 50, "n_fig": 4, "n_tab": 7, "L_tab": 6026, "n_element_tab": 294, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 64, "n_element_tab_1": 11, "formula_len_all": 88, "formula_len_all_1": 134, "len_all": 117400, "len_all_1": 58900, "len_abs": 1268, "len_title": 84, "len_sents": 36783, "len_sents_1": 27756, "n_sents": 299, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 1278, "title": "A Language Model based Model Manager", "abs": "In the current landscape of machine learning, we face a \u201cmodel lake\u201d phenomenon: a proliferation of deployed models often lacking adequate documentation. This presents significant challenges for model users attempting to navigate, differentiate, and select appropriate models for their needs. To address the issue of differentiation, we introduce Model Manager, a framework designed to facilitate easy comparison among existing models. Our approach leverages a large language model (LLM) to generate verbalizations of two models' differences by sampling from two models. We use a novel protocol that makes it possible to quantify the informativeness of the verbalizations. We also assemble a suite with a diverse set of commonly used models: Logistic Regression, Decision Trees, and K-Nearest Neighbors. We additionally performed ablation studies on crucial design decisions of the Model Managers. Our analysis yields pronounced results. For a pair of logistic regression models with a 20-25\\% performance difference on the blood dataset, the Model Manager effectively verbalizes their variations with up to 80\\% accuracy. The Model Manager framework opens up new research avenues for improving the transparency and comparability of machine learning models in a post-hoc manner.", "keywords": [ "Large Language Models", "Model Manager", "Verbalization", "Differentiation" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 1 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "xof0bvftR1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Batuhan K. Karaman", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Fengbei Liu", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Postdoc" }, { "name": "Heejong Kim", "gender": "Not Specified", "institution": "Weill Cornell Medicine, Cornell University", "country": "US", "position": "Instructor" }, { "name": "Mert R. Sabuncu", "gender": "Male", "institution": "Cornell Tech", "country": "US", "position": "Full Professor" }, { "name": "Minh Nguyen", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Alan Wang", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 51, "n_ref": 76, "n_ref_all": 88, "n_fig": 4, "n_tab": 7, "L_tab": 1586, "n_element_tab": 92, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2570, "n_element_tab_1": 97, "formula_len_all": 945, "formula_len_all_1": 862, "len_all": 145938, "len_all_1": 65453, "len_abs": 1453, "len_title": 96, "len_sents": 32927, "len_sents_1": 27264, "n_sents": 291, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 47, "L_abs": 1424, "title": "Knockout: A simple way to handle missing inputs", "abs": "Deep learning models can extract predictive and actionable information from complex inputs. The richer the inputs, the better these models usually perform. However, models that leverage rich inputs (e.g., multi-modality) can be difficult to deploy widely, because some inputs may be missing at inference. Current popular solutions to this problem include marginalization, imputation, and training multiple models. Marginalization can obtain calibrated predictions but it is computationally costly and therefore only feasible for low dimensional inputs. Imputation may result in inaccurate predictions because it employs point estimates for missing variables and does not work well for high dimensional inputs (e.g., images). Training multiple models whereby each model takes different subsets of inputs can work well but requires knowing missing input patterns in advance. Furthermore, training and retaining multiple models can be costly. We propose an efficient way to learn both the conditional distribution using full inputs and the marginal distributions. Our method, Knockout, randomly replaces input features with appropriate placeholder values during training. We provide a theoretical justification of Knockout and show that it can be viewed as an implicit marginalization strategy. We evaluate Knockout in a wide range of simulations and real-world datasets and show that it can offer strong empirical performance.", "keywords": [ "Applied Machine Learning", "Marginalization", "Missing inputs", "Multi-modality" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "xoXn62FzD0", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 15, "author_info_list": [ { "name": "Alexander K. Lew", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Ben Lipkin", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Benjamin LeBrun", "gender": "unknown", "institution": "McGill University", "country": "", "position": "Undergrad student" }, { "name": "Clemente Pasti", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Gabriel Grand", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jason Eisner", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Director of Research " }, { "name": "Jo\u00e3o Loula", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Li Du", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "PhD student" }, { "name": "Marjorie Freedman", "gender": "unknown", "institution": "USC/ISI", "country": "", "position": "Principal Researcher" }, { "name": "Tianyu Liu", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Tim Vieira", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "US", "position": "Postdoc" }, { "name": "Timothy J. O'Donnell", "gender": "unknown", "institution": "McGill University, Mila", "country": "", "position": "Associate Professor" }, { "name": "Vikash Mansinghka", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Principal Research Scientist" }, { "name": "Yahya Emara", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Ryan Cotterell", "gender": "Not Specified", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 13, "n_ref_uni": 65, "n_ref": 136, "n_ref_all": 153, "n_fig": 2, "n_tab": 8, "L_tab": 1252, "n_element_tab": 94, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 513, "n_element_tab_1": 22, "formula_len_all": 2230, "formula_len_all_1": 1297, "len_all": 205512, "len_all_1": 66675, "len_abs": 1013, "len_title": 130, "len_sents": 65178, "len_sents_1": 31898, "n_sents": 455, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 82, "L_abs": 1019, "title": "Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo", "abs": "A wide range of LLM applications require generating text that conforms to syntactic or semantic constraints. Imposing such constraints nontrivially alters the distribution over sequences, usually making exact sampling intractable. In this work, building on the Language Model Probabilistic Programming framework of Lew et al. (2023), we develop an approach to approximate inference for controlled LLM generation based on sequential Monte Carlo (SMC). Our SMC framework allows us to flexibly incorporate domain- and problem-specific constraints at inference time, and efficiently reallocate computation in light of new information during the course of generation. We demonstrate that our approach improves downstream performance on four challenging domains---Python code generation for data science, text-to-SQL, goal inference, and molecule synthesis. We compare to a number of alternative and ablated approaches, showing that our accuracy improvements are driven by better approximation to the full Bayesian posterior.", "keywords": [ "Sequential Monte Carlo", "Language Models", "Semantic parsing", "Bayesian inference", "Probabilistic programming", "SMC" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "xoW1Cb4MkP", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Liefeng Bo", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Principal Researcher" }, { "name": "Yifeng Geng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Tuo Yuxiang", "gender": "Male", "institution": "Alibaba", "country": "CN", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 14, "n_ref": 26, "n_ref_all": 42, "n_fig": 6, "n_tab": 7, "L_tab": 5467, "n_element_tab": 215, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 851, "n_element_tab_1": 54, "formula_len_all": 604, "formula_len_all_1": 328, "len_all": 131699, "len_all_1": 58524, "len_abs": 1761, "len_title": 122, "len_sents": 36345, "len_sents_1": 27106, "n_sents": 268, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1770, "title": "ANYTEXT2: Visual Text Generation and Editing with Customizable Attributes", "abs": "With the ongoing development in the text-to-image(T2I) domain, accurately generating text within images seamlessly integrating with the visual content has garnered increasing interest from the research community. In addition to controlling glyphs and positions of text, there is a rising demand for more fine-grained control over text attributes, such as font style and color, while maintaining the realism of the generated images. However, this issue has not yet been sufficiently explored. In this paper, we present AnyText2, the first known method to achieve precise control over the attributes of every line of multilingual text when generating images of natural scenes. Our method comprises two main components. First, we introduce an efficient WriteNet+AttnX architecture that encodes text features and injects these intermediate features into the U-Net decoder via learnable attention layers. This design is 19.8% faster than its predecessor, AnyText, and improves the realism of the generated images. Second, we thoroughly explore methods for extracting text fonts and colors from real images, and then develop a Text Embedding Module that employs multiple encoders to separately encode the glyph, position, font, and color of the text. This enables customizable font and color for each text line, yielding a 3.3% and 9.3% increase in text accuracy for Chinese and English, respectively, compared to AnyText. Furthermore, we validate the use of long captions, which enhances prompt-following and image realism without sacrificing text writing accuracy. Through comprehensive experiments, we demonstrate the state-of-the-art performance of our method. The code and model will be open-sourced in the future to promote the development of text generation technology.", "keywords": [ "Text-to-Image", "Visual Text Generation", "Visual Text Editing", "Customizable Attributes" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "xoUUCS9IGl", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Arian Rokkum Jamasb", "gender": "unknown", "institution": "Prescient Design / Roche / Genentech", "country": "CH", "position": "Researcher" }, { "name": "Charles Harris", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Kieran Didi", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "MS student" }, { "name": "Pietro Lio", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Simon V Mathis", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Tom Leon Blundell", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "Emeritus" }, { "name": "Chaitanya Krishna Joshi", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 41, "n_ref": 66, "n_ref_all": 91, "n_fig": 12, "n_tab": 3, "L_tab": 2118, "n_element_tab": 96, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 118076, "len_all_1": 57695, "len_abs": 1348, "len_title": 138, "len_sents": 33822, "len_sents_1": 28003, "n_sents": 216, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1351, "title": "PoseCheck: Generative Models for 3D Structure-based Drug Design Produce Unrealistic Poses", "abs": "Deep generative models for structure-based drug design (SBDD), where molecule generation is conditioned on a 3D protein pocket, have received considerable interest in recent years. These methods offer the promise of higher-quality molecule generation by explicitly modelling the 3D interaction between a potential drug and a protein receptor. However, previous work has primarily focused on the quality of the generated molecules themselves, with limited evaluation of the 3D poses that these methods produce, with most work simply discarding the generated pose and only reporting a \u201ccorrected\u201d pose after redocking with traditional methods. Little is known about whether generated molecules satisfy known physical constraints for binding and the extent to which redocking alters the generated interactions. We introduce POSECHECK, an extensive benchmarking suite for state-of-the-art SBDD methods and find that generated molecules have significantly more physical violations and fewer key interactions compared to baselines, calling into question the implicit assumption that providing rich 3D structure information improves molecule complementarity. We make recommendations for future research tackling identified failure modes and hope our benchmark will serve as a springboard for future SBDD generative modelling work to have a real-world impact.", "keywords": [ "generative models", "drug design", "benchmarks" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "xoIeVdFO7U", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Benjamin Eysenbach", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Assistant Professor" }, { "name": "Chongyi Zheng", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Jens Tuyls", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Joanne Peng", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Undergrad student" } ], "n_formula": 49, "n_formula_1": 18, "n_ref_uni": 74, "n_ref": 218, "n_ref_all": 246, "n_fig": 13, "n_tab": 2, "L_tab": 1355, "n_element_tab": 62, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5178, "formula_len_all_1": 1846, "len_all": 227221, "len_all_1": 82527, "len_abs": 2048, "len_title": 126, "len_sents": 60549, "len_sents_1": 34029, "n_sents": 484, "n_sents_1": 267, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 78, "L_abs": 907, "title": "Can a MISL Fly? Analysis and Ingredients for Mutual Information Skill Learning", "abs": "Self-supervised learning has the potential of lifting several of the key challenges in reinforcement learning today, such as exploration, representation learning, and reward design. Recent work (METRA) has effectively argued that moving away from mutual information and instead optimizing a certain Wasserstein distance is important for good performance. In this paper, we argue that the benefits seen in that paper can largely be explained within the existing framework of mutual information skill learning (MISL).\nOur analysis suggests a new MISL method (contrastive successor features) that retains the excellent performance of METRA with fewer moving parts, and highlights connections between skill learning, contrastive representation learning, and successor features. Finally, through careful ablation studies, we provide further insight into some of the key ingredients for both our method and METRA.", "keywords": [ "unsupervised learning", "reinforcement learning", "mutual information", "successor feature" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 4, 4, 3, 4 ], "presentation_list": [ 4, 4, 3, 4 ], "contribution_list": [ 4, 2, 3, 4 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "xnssGv9rpW", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Kin Long Kelvin Lee", "gender": "Male", "institution": "Intel", "country": "", "position": "Researcher" }, { "name": "Mikhail Galkin", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Qiang Zhu", "gender": "unknown", "institution": "University of North Carolina at Charlotte", "country": "", "position": "Associate Professor" }, { "name": "Santiago Miret", "gender": "Male", "institution": "Intel", "country": "US", "position": "Researcher" }, { "name": "Siamak Ravanbakhsh", "gender": "unknown", "institution": "McGill University", "country": "", "position": "Assistant Professor" }, { "name": "Siba Smarak Panigrahi", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "S\u00e9kou-Oumar Kaba", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "Daniel Tian Levy", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 14, "n_ref_uni": 48, "n_ref": 92, "n_ref_all": 113, "n_fig": 7, "n_tab": 4, "L_tab": 921, "n_element_tab": 109, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 779, "n_element_tab_1": 102, "formula_len_all": 1317, "formula_len_all_1": 975, "len_all": 164133, "len_all_1": 72357, "len_abs": 1299, "len_title": 116, "len_sents": 48520, "len_sents_1": 33191, "n_sents": 357, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1298, "title": "SymmCD: Symmetry-Preserving Crystal Generation with Diffusion Models", "abs": "Generating novel crystalline materials has potential to lead to advancements in fields such as electronics, energy storage, and catalysis. The defining characteristic of crystals is their symmetry, which plays a central role in determining their physical properties. However, existing crystal generation methods either fail to generate materials that display the symmetries of real-world crystals, or simply replicate the symmetry information from examples in a database. To address this limitation, we propose SymmCD, a novel diffusion-based generative model that explicitly incorporates crystallographic symmetry into the generative process. We decompose crystals into two components and learn their joint distribution through diffusion: 1) the asymmetric unit, the smallest subset of the crystal which can generate the whole crystal through symmetry transformations, and; 2) the symmetry transformations needed to be applied to each atom in the asymmetric unit. We also use a novel and interpretable representation for these transformations, enabling generalization across different crystallographic symmetry groups. We showcase the competitive performance of SymmCD on a subset of the Materials Project, obtaining diverse and valid crystals with realistic symmetries and predicted properties.", "keywords": [ "Crystals", "Symmetry", "Materials", "Diffusion", "Generative Models", "Equivariance" ], "rating_list": [ 8, 8, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 3, 4, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "xnWikQRJBR", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Deval Mehta", "gender": "unknown", "institution": "Monash University", "country": "AU", "position": "Postdoc" }, { "name": "Kshitij Sharad Jadhav", "gender": "Male", "institution": "Indian Institute of Technology, Bombay", "country": "IN", "position": "Assistant Professor" }, { "name": "Pranamya Kulkarni", "gender": "Male", "institution": "Indian Institute of Technology, Bombay", "country": "", "position": "Undergrad student" }, { "name": "Raghav Singhal", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "RAJA KUMAR", "gender": "Male", "institution": "CNRS", "country": "FR", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 17, "n_ref_uni": 57, "n_ref": 123, "n_ref_all": 153, "n_fig": 7, "n_tab": 6, "L_tab": 1980, "n_element_tab": 314, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 475, "n_element_tab_1": 84, "formula_len_all": 1418, "formula_len_all_1": 1430, "len_all": 165488, "len_all_1": 57928, "len_abs": 1301, "len_title": 154, "len_sents": 41569, "len_sents_1": 24981, "n_sents": 324, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1307, "title": "M3CoL: Harnessing Shared Relations via Multimodal Mixup Contrastive Learning for Multimodal Classification", "abs": "Deep multimodal learning has shown remarkable success by leveraging contrastive learning to capture explicit one-to-one relations across modalities. However, real-world data often exhibits shared relations beyond simple pairwise associations. We propose M3CoL, a Multimodal Mixup Contrastive Learning approach to capture nuanced shared relations inherent in multimodal data. Our key contribution is a Mixup-based contrastive loss that learns robust representations by aligning mixed samples from one modality with their corresponding samples from other modalities thereby capturing shared relations between them. For multimodal classification tasks, we introduce a framework that integrates a fusion module with unimodal prediction modules for auxiliary supervision during training, complemented by our proposed Mixup-based contrastive loss. Through extensive experiments on diverse datasets (N24News, ROSMAP, BRCA, and Food-101), we demonstrate that M3CoL effectively captures shared multimodal relations and generalizes across domains. It outperforms state-of-the-art methods on N24News, ROSMAP, and BRCA, while achieving comparable performance on Food-101. Our work highlights the significance of learning shared relations for robust multimodal learning, opening up promising avenues for future research.", "keywords": [ "Contrastive learning", "multimodal learning", "representation learning", "mutlimodal classification" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "xnF2U0ro7b", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Andrea Celli", "gender": "Male", "institution": "Bocconi University", "country": "IT", "position": "Assistant Professor" }, { "name": "Martino Bernasconi", "gender": "unknown", "institution": "Bocconi University", "country": "", "position": "Postdoc" }, { "name": "Matteo Castiglioni", "gender": "unknown", "institution": "Politecnico di Milano", "country": "IT", "position": "Assistant Professor" }, { "name": "Solenne Gaucher", "gender": "Female", "institution": "Ecole Nationale de la Statistique et de l'Administration Economique", "country": "", "position": "Postdoc" }, { "name": "Vianney Perchet", "gender": "Male", "institution": "Ensae ParisTech", "country": "", "position": "Professor" } ], "n_formula": 140, "n_formula_1": 18, "n_ref_uni": 31, "n_ref": 50, "n_ref_all": 62, "n_fig": 2, "n_tab": 21, "L_tab": 7216, "n_element_tab": 711, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 1100, "n_element_tab_1": 92, "formula_len_all": 13703, "formula_len_all_1": 1441, "len_all": 251981, "len_all_1": 81870, "len_abs": 1472, "len_title": 36, "len_sents": 62594, "len_sents_1": 29128, "n_sents": 629, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 36, "L_abs": 1500, "title": "Feature-Based Online Bilateral Trade", "abs": "Bilateral trade models the problem of facilitating trades between a seller and a buyer having private valuations for the item being sold. In the online version of the problem, the learner faces a new seller and buyer at each time step, and has to post a price for each of the two parties without any knowledge of their valuations. We consider a scenario where, at each time step, before posting prices the learner observes a context vector containing information about the features of the item for sale. The valuations of both the seller and the buyer follow an unknown linear function of the context. In this setting, the learner could leverage previous transactions in an attempt to estimate private valuations. We characterize the regret regimes of different settings, taking as a baseline the best context-dependent prices in hindsight. First, in the setting in which the learner has two-bit feedback and strong budget balance constraints, we propose an algorithm with $O(\\log T)$ regret. Then, we study the same set-up with noisy valuations, providing a tight $\\widetilde O(T^{2/3})$ regret upper bound. Finally, we show that loosening budget balance constraints allows the learner to operate under more restrictive feedback. Specifically, we show how to address the one-bit, global budget balance setting through a reduction from the two-bit, strong budget balance setup. This established a fundamental trade-off between the quality of the feedback and the strictness of the budget constraints.", "keywords": [ "bilateral trade", "online learning", "contextual bandits" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 4, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "xmgvF0sLIn", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Arushi Goel", "gender": "Female", "institution": "NVIDIA Research", "country": "", "position": "Researcher" }, { "name": "Bryan Catanzaro", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Vice President" }, { "name": "Rafael Valle", "gender": "Not Specified", "institution": "NVIDIA", "country": "", "position": "Senior Research Scientist" }, { "name": "Sang-gil Lee", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Sungwon Kim", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Zhifeng Kong", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" } ], "n_formula": 12, "n_formula_1": 1, "n_ref_uni": 49, "n_ref": 123, "n_ref_all": 179, "n_fig": 4, "n_tab": 34, "L_tab": 8787, "n_element_tab": 1481, "n_fig_1": 3, "n_tab_1": 10, "L_tab_1": 3470, "n_element_tab_1": 395, "formula_len_all": 1297, "formula_len_all_1": 410, "len_all": 249990, "len_all_1": 73672, "len_abs": 2611, "len_title": 100, "len_sents": 63820, "len_sents_1": 30795, "n_sents": 572, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1388, "title": "Elucidating the Design Space of Text-to-Audio Models", "abs": "Recent years have seen significant progress in Text-To-Audio (TTA) synthesis, enabling users to enrich their creative workflows with synthetic audio generated from natural language prompts. Despite this progress, the effects of data, model architecture, training objective functions, and sampling strategies on target benchmarks are not well understood. With the purpose of providing a holistic understanding of the design space of TTA models, we setup a large-scale empirical experiment focused on diffusion and flow matching models. Our contributions include: 1) AF-Synthetic, a large dataset of high quality synthetic captions obtained from an audio understanding model; 2) a systematic comparison of different architectural, training, and inference design choices for TTA models; 3) an analysis of sampling methods and their Pareto curves with respect to generation quality and inference speed. We leverage the knowledge obtained from this extensive analysis to propose our best model dubbed Elucidated Text-To-Audio (ETTA). When evaluated on AudioCaps and MusicCaps, ETTA provides improvements over the baselines trained on publicly available data, while being competitive with models trained on proprietary data. Finally, we show ETTA's improved ability to generate creative audio following complex and imaginative captions \u2014 a task that is more challenging than current benchmarks.", "keywords": [ "audio generation", "text-to-audio", "synthetic data", "diffusion", "flow matching" ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 4, 2, 4, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "xlxGsX1pc7", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Alex Myasnikov", "gender": "Male", "institution": "Gradarius", "country": "US", "position": "CTO" }, { "name": "Alexei Miasnikov", "gender": "unknown", "institution": "Stevens Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Konstantin Chernyshev", "gender": "unknown", "institution": "Universit\u00e4t des Saarlandes", "country": "DE", "position": "MS student" }, { "name": "Sergei Tilga", "gender": "Male", "institution": "Toloka AI", "country": "NL", "position": "Head of R&D" }, { "name": "Vitaliy Polshkov", "gender": "Male", "institution": "Toloka AI", "country": "RS", "position": "Researcher" }, { "name": "Vlad Stepanov", "gender": "Male", "institution": "Gradarius (Castle Point Learning Systems)", "country": "US", "position": "CEO" }, { "name": "Katya Artemova", "gender": "Female", "institution": "Toloka AI", "country": "", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 1, "n_ref_uni": 28, "n_ref": 69, "n_ref_all": 83, "n_fig": 8, "n_tab": 7, "L_tab": 3466, "n_element_tab": 387, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 617, "n_element_tab_1": 87, "formula_len_all": 493, "formula_len_all_1": 147, "len_all": 177723, "len_all_1": 57012, "len_abs": 1441, "len_title": 127, "len_sents": 38561, "len_sents_1": 28288, "n_sents": 330, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1208, "title": "U-MATH: A University-Level Benchmark for Evaluating Mathematical Skills in LLMs", "abs": "The current evaluation of mathematical skills in LLMs is limited, as existing benchmarks are relatively small, primarily focus on elementary and high-school problems, or lack diversity in topics. Additionally, the inclusion of visual elements in tasks remains largely under-explored. \n\nTo address these gaps, we introduce **U-MATH**, a novel benchmark of \\textbf{1,100} unpublished open-ended university-level problems sourced from teaching materials. It is balanced across six core subjects, with \\textbf{20\\% of multimodal problems}. Given the open-ended nature of U-MATH problems, we employ an LLM to judge the correctness of generated solutions. To this end, we release **$\\boldsymbol\\mu$-MATH**, an dataset to evaluate the LLMs' capabilities in judging solutions.\n\nThe evaluation of general domain, math-specific, and multimodal LLMs highlights the challenges presented by U-MATH. Our findings reveal that LLMs achieve a maximum accuracy of only 63\\% on text-based tasks, with even lower 45\\% on visual problems. The solution assessment proves challenging for LLMs, with the best LLM judge having an F1-score of 80\\% on $\\mu$-MATH.\n \nWe open-source U-MATH, $\\mu$-MATH, and evaluation code on GitHub.", "keywords": [ "Large Language Models (LLMs)", "Mathematical Reasoning", "Benchmarking", "University-Level Mathematics", "Multimodal", "Automatic Evaluation", "Solution Assessment" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "xlxDTVAbNM", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Aditi Raghunathan", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Gaurav Rohit Ghosal", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Suhas Kotha", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Tanishq Kumar", "gender": "Male", "institution": "Harvard College", "country": "", "position": "Undergrad student" }, { "name": "Uzay Girit", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Undergrad student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 35, "n_ref": 41, "n_ref_all": 63, "n_fig": 11, "n_tab": 1, "L_tab": 182, "n_element_tab": 64, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 106, "n_element_tab_1": 12, "formula_len_all": 657, "formula_len_all_1": 612, "len_all": 121920, "len_all_1": 58862, "len_abs": 709, "len_title": 128, "len_sents": 35078, "len_sents_1": 26735, "n_sents": 271, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 716, "title": "Lowering Data Diversity can Accelerate Training: Case Studies in Synthetic Tasks", "abs": "We identify a loss plateau at the start of training in the three synthetic settings of in-context linear regression, sparse parity, and fact memorization. While careful tweaks to the optimization algorithm can mitigate these plateaus, we find that a simpler orthogonal approach of *lowering the data diversity*, and in doing so, biasing the training distribution *away* from the test distribution, counter-intuitively also speeds up training. This connection between data diversity and training speed holds for three different diversity-*reducing* interventions across our varied synthetic settings. Our findings offer a new perspective on data filtering and curriculum learning for training machine learning models.", "keywords": [ "synthetic tasks", "data diversity", "curriculum learning", "data filtering", "learning plateaus", "batch gradients" ], "rating_list": [ 5, 3, 1, 5 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "xlrpVyMIwz", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "David J Nott", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Associate Professor" }, { "name": "Guilherme S. Rodrigues", "gender": "Male", "institution": "Universidade de Bras\u00edlia", "country": "BR", "position": "Assistant Professor" }, { "name": "Scott A Sisson", "gender": "unknown", "institution": "University of New South Wales", "country": "AU", "position": "Full Professor" }, { "name": "Thais Carvalho Valadares Rodrigues", "gender": "Female", "institution": "Universidade de Bras\u00edlia", "country": "BR", "position": "Lecturer" }, { "name": "William E. R. de Amorim", "gender": "Male", "institution": "Universidade de Bras\u00edlia", "country": "BR", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 34, "n_ref": 66, "n_ref_all": 80, "n_fig": 8, "n_tab": 4, "L_tab": 2126, "n_element_tab": 163, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1044, "n_element_tab_1": 81, "formula_len_all": 721, "formula_len_all_1": 453, "len_all": 112949, "len_all_1": 69795, "len_abs": 1162, "len_title": 117, "len_sents": 35003, "len_sents_1": 30760, "n_sents": 300, "n_sents_1": 265, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1167, "title": "Positional Encoder Graph Quantile Neural Networks for Geographic Data", "abs": "Positional Encoder Graph Neural Networks (PE-GNNs) are a leading approach for modeling continuous spatial data. However, they often fail to produce calibrated predictive distributions, limiting their effectiveness for uncertainty quantification. We introduce the Positional Encoder Graph Quantile Neural Network (PE-GQNN), a novel method that integrates PE-GNNs, Quantile Neural Networks, and recalibration techniques in a fully nonparametric framework, requiring minimal assumptions about the predictive distributions. We propose a new network architecture that, when combined with a quantile-based loss function, yields accurate and reliable probabilistic models without increasing computational complexity. Our approach provides a flexible, robust framework for conditional density estimation, applicable beyond spatial data contexts. We further introduce a structured method for incorporating a KNN predictor into the model while avoiding data leakage through the GNN layer operation. Experiments on benchmark datasets demonstrate that PE-GQNN significantly outperforms existing state-of-the-art methods in both predictive accuracy and uncertainty quantification.", "keywords": [ "Graph Neural Networks (GNNs); Quantile regression; Geospatial data; Uncertainty quantification; Calibration; Model recalibration." ], "rating_list": [ 6, 3, 1, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "xljPZuprBA", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Fanchen Bu", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Kijung Shin", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Paul Bogdan", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Jack Munushian Early Career Chair associate professor" }, { "name": "Ruochen Yang", "gender": "Female", "institution": "University of Southern California", "country": "", "position": "PhD student" } ], "n_formula": 88, "n_formula_1": 14, "n_ref_uni": 78, "n_ref": 109, "n_ref_all": 153, "n_fig": 7, "n_tab": 31, "L_tab": 21234, "n_element_tab": 1451, "n_fig_1": 1, "n_tab_1": 9, "L_tab_1": 2909, "n_element_tab_1": 75, "formula_len_all": 24270, "formula_len_all_1": 2836, "len_all": 348307, "len_all_1": 84680, "len_abs": 2409, "len_title": 149, "len_sents": 84974, "len_sents_1": 32999, "n_sents": 874, "n_sents_1": 301, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1383, "title": "Exploring Edge Probability Graph Models Beyond Edge Independency: Concepts, Analyses, and Algorithms", "abs": "Desirable random graph models (RGMs) should \n*(i)* reproduce *common patterns* in real-world graphs (e.g., high clustering),\n*(ii)* generate *variable* (i.e., not overly similar) graphs, and \n*(iii)* remain *tractable* to compute and control graph statistics.\nA common class of RGMs (e.g., Erd\\H{o}s-R\\'{e}nyi and stochastic Kronecker) outputs edge probabilities, and we need to realize (i.e., sample from) the edge probabilities to generate graphs.\nTypically, each edge's existence is assumed to be determined independently for simplicity and tractability.\nHowever, with edge independency, RGMs theoretically cannot produce high subgraph densities and high output variability simultaneously.\nIn this work, we explore realization beyond edge independence that can better reproduce common patterns while maintaining high tractability and variability.\nTheoretically, we propose an edge-dependent realization framework called *binding* that provably preserves output variability, and derive *closed-form* tractability results on subgraph (e.g., triangle) densities in generated graphs.\nPractically, we propose algorithms for graph generation with binding and parameter fitting of binding.\nOur empirical results demonstrate that binding exhibits high tractability and well reproduce patterns such as high clustering, significantly improving upon existing RGMs assuming edge independency.", "keywords": [ "Random graph models", "edge dependency", "triangle density", "subgraph densities", "tractability", "variability" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "xlbXRJ2XCP", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Carlo Abate", "gender": "Male", "institution": "University of Bologna", "country": "IT", "position": "PhD student" }, { "name": "Filippo Maria Bianchi", "gender": "Male", "institution": "University of Troms\u00f8", "country": "", "position": "Associate Professor" } ], "n_formula": 33, "n_formula_1": 13, "n_ref_uni": 60, "n_ref": 261, "n_ref_all": 284, "n_fig": 11, "n_tab": 9, "L_tab": 3113, "n_element_tab": 337, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 888, "n_element_tab_1": 66, "formula_len_all": 2471, "formula_len_all_1": 846, "len_all": 263755, "len_all_1": 88642, "len_abs": 177, "len_title": 132, "len_sents": 67693, "len_sents_1": 34567, "n_sents": 547, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 495, "title": "MaxCutPool: differentiable feature-aware Maxcut for pooling in graph neural networks", "abs": "We propose a novel approach to compute the MAXCUT in attributed graphs, i.e., graphs with features associated with nodes and edges. Our approach works well on any kind of graph topology and can find solutions that jointly optimize the MAXCUT along with other objectives. Based on the obtained MAXCUT partition, we implement a hierarchical graph pooling layer for Graph Neural Networks, which is sparse, trainable end-to-end, and particularly suitable for downstream tasks on heterophilic graphs.", "keywords": [ "Graph neural networks", "graph pooling", "graph coarsening", "maxcut" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 5, 4, 2, 3 ] }, { "paper_id": "xkgfLXZ4e0", "primary_area": "applications to neuroscience & cognitive science", "n_author": 9, "author_info_list": [ { "name": "Akshett Rai Jindal", "gender": "Male", "institution": "International Institute of Information Technology, Hyderabad, International Institute of Information Technology Hyderabad", "country": "IN", "position": "Undergrad student" }, { "name": "Bapi Raju Surampudi", "gender": "Male", "institution": "International Institute of Information Technology Hyderabad, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "", "position": "Full Professor" }, { "name": "Ishani Mondal", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Research Fellow" }, { "name": "Khushbu Pahwa", "gender": "Female", "institution": "Rice University", "country": "US", "position": "MS student" }, { "name": "Maneesh K. Singh", "gender": "Male", "institution": "Spector Inc", "country": "US", "position": "Head, AI Research & Technologies" }, { "name": "Manish Gupta", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Principal Researcher" }, { "name": "Manish Shrivastava", "gender": "Male", "institution": "International Institute of Information Technology Hyderabad, India", "country": "", "position": "Assistant Professor" }, { "name": "SUBBA REDDY OOTA", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "PhD student" }, { "name": "Satya Sai Srinath Namburi GNVV", "gender": "Male", "institution": "GE HealthCare", "country": "US", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 34, "n_ref": 82, "n_ref_all": 126, "n_fig": 20, "n_tab": 4, "L_tab": 1609, "n_element_tab": 266, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1047, "n_element_tab_1": 58, "formula_len_all": 171, "formula_len_all_1": 85, "len_all": 197479, "len_all_1": 69886, "len_abs": 1963, "len_title": 150, "len_sents": 70651, "len_sents_1": 36341, "n_sents": 464, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1980, "title": "Correlating instruction-tuning (in multimodal models) with vision-language processing (in the brain)", "abs": "Transformer-based language models, though not explicitly trained to mimic brain recordings, have demonstrated surprising alignment with brain activity. Progress in these models\u2014through increased size, instruction-tuning, and multimodality\u2014has led to better representational alignment with neural data. Recently, a new class of instruction-tuned multimodal LLMs (MLLMs) have emerged, showing remarkable zero-shot capabilities in open-ended multimodal vision tasks. However, it is unknown whether MLLMs, when prompted with natural instructions, lead to better brain alignment and effectively capture instruction-specific representations. To address this, we first investigate the brain alignment, i.e., measuring the degree of predictivity of neural visual activity using text output response embeddings from MLLMs as participants engage in watching natural scenes. Experiments with 10 different instructions (like image captioning, visual question answering, etc.) show that MLLMs exhibit significantly better brain alignment than vision-only models and perform comparably to non-instruction-tuned multimodal models like CLIP. We also find that while these MLLMs are effective at generating high-quality responses suitable to the task-specific instructions, not all instructions are relevant for brain alignment. Further, by varying instructions, we make the MLLMs encode instruction-specific visual concepts related to the input image. This analysis shows that MLLMs effectively capture count-related and recognition-related concepts, demonstrating strong alignment with brain activity. Notably, the majority of the explained variance of the brain encoding models is shared between MLLM embeddings of image captioning and other instructions. These results indicate that enhancing MLLMs' ability to capture more task-specific information could allow for better differentiation between various types of instructions, and hence improve their precision in predicting brain responses.", "keywords": [ "brain encoding", "fMRI", "visual processing", "multimodal instruction-tuned models", "language decoder", "LLMs", "MLLMs" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "xkR3bcswuC", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Anand Bhattad", "gender": "unknown", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Research Assistant Professor" }, { "name": "Nicholas Kolkin", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Xiaodan Du", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "PhD student" }, { "name": "Gregory Shakhnarovich", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 58, "n_ref": 88, "n_ref_all": 114, "n_fig": 16, "n_tab": 5, "L_tab": 2229, "n_element_tab": 544, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1243, "n_element_tab_1": 308, "formula_len_all": 180, "formula_len_all_1": 508, "len_all": 150281, "len_all_1": 52738, "len_abs": 0, "len_title": 47, "len_sents": 34505, "len_sents_1": 23656, "n_sents": 288, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1341, "title": "Generative Models: What Do They Know? Do They Know Things? Let's Find Out!", "abs": "Generative models excel at mimicking real scenes, suggesting they might inherently encode important intrinsic scene properties. In this paper, we aim to explore the following key questions: (1) What intrinsic knowledge do generative models like Autoregressive models, GANs and Diffusion models encode? (2) Can we establish a general framework to recover intrinsic representations from these models, regardless of their architecture or model type? (3) How small can the required learnable parameters and labeled data be to successfully recover this knowledge? (4) Is there a direct link between the quality of a generative model and the accuracy of the recovered scene intrinsics?\n\nOur findings indicate that a small Low-Rank Adaptation (LoRA) can recover intrinsic images---depth, normals, albedo, and shading---across different generators (GAN, Autoregressive, and Diffusion) while using the same decoder head that generates the image. As LoRA is lightweight, we introduce very few learnable parameters (as few as 0.04% of Stable Diffusion model weights for a rank of 2), and we find that as few as 250 labeled images are enough to generate intrinsic images with these LoRA modules. Finally, we also show a positive correlation between the generative model's quality and the accuracy of the recovered intrinsics through control experiments.", "keywords": [ "Visual knowledge", "Generative models", "Intrinsic Images" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "xjornbs7aT", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Alberto Sangiovanni-Vincentelli", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Lukas Dirnberger", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Marco Caccamo", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Mirco Theile", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Postdoc" }, { "name": "Raphael Trumpp", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 18, "n_ref_uni": 27, "n_ref": 49, "n_ref_all": 70, "n_fig": 11, "n_tab": 3, "L_tab": 2224, "n_element_tab": 216, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2232, "formula_len_all_1": 1396, "len_all": 136542, "len_all_1": 66471, "len_abs": 1020, "len_title": 133, "len_sents": 44945, "len_sents_1": 29994, "n_sents": 362, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1023, "title": "Action Mapping for Reinforcement Learning in Continuous Environments with Constraints", "abs": "Deep reinforcement learning (DRL) has had success across various domains, but applying it to environments with constraints remains challenging due to poor sample efficiency and slow convergence. Recent literature explored incorporating model knowledge to mitigate these problems, particularly through the use of models that assess the feasibility of proposed actions. However, integrating feasibility models efficiently into DRL pipelines in environments with continuous action spaces is non-trivial. We propose a novel DRL training strategy utilizing action mapping that leverages feasibility models to streamline the learning process. By decoupling the learning of feasible actions from policy optimization, action mapping allows DRL agents to focus on selecting the optimal action from a reduced feasible action set. We demonstrate through experiments that action mapping significantly improves training performance in constrained environments with continuous action spaces, especially with imperfect feasibility models.", "keywords": [ "Constrained MDPs", "continuous action space", "deep reinforcement learning" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "xjKz6IxgCX", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Francesco Pinto", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Postdoc" }, { "name": "Minzhou Pan", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "PhD student" }, { "name": "Zhaorun Chen", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 39, "n_ref": 114, "n_ref_all": 186, "n_fig": 14, "n_tab": 19, "L_tab": 8816, "n_element_tab": 851, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 696, "n_element_tab_1": 70, "formula_len_all": 1102, "formula_len_all_1": 687, "len_all": 284203, "len_all_1": 66676, "len_abs": 2182, "len_title": 149, "len_sents": 98085, "len_sents_1": 32579, "n_sents": 658, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 99, "L_abs": 2187, "title": "SafeWatch: An Efficient Safety-Policy Following Video Guardrail Model with Transparent Explanations", "abs": "With the wide adoption of generative AI and rapid growth of high-quality video generation, video guardrails have become more crucial than ever to ensure safety and security across platforms. Current video guardrails, however, are either overly simplistic, relying on pure classification models trained on simple policies with limited number of unsafe categories, which lack detailed explanations, or prompting multimodal large language models (MLLMs) with long safety guidelines, resulting in inefficient and impractical guardrails for real-world content. To bridge this gap, we propose SAFEWATCH, an efficient MLLM-based video guardrail model designed to follow customized safety policies and provide multi-label video guardrail outputs with content-specific explanations in a zero-shot manner. In particular, unlike traditional guardrails that encode entire policies autoregressive, causing inefficiency and bias, SAFEWATCH uniquely encodes each policy trunk in parallel and eliminates their position bias such that all policies are attended simultaneously with equal importance. In addition, to improve efficiency and accuracy, SafeWatch incorporates a policy-aware visual token pruning algorithm that adaptively selects the most relevant video tokens for each policy, discarding noisy or irrelevant information. This allows for more focused, policy-compliant guardrail with significantly reduced computational overhead. Considering the limitations of existing video guardrail benchmarks, we propose SafeWatch-Bench, a large-scale video guardrail benchmark comprising over 2M videos spanning six safety categories which covers over 30 tasks to ensure a comprehensive coverage of all potential safety scenarios. We have conducted extensive experiments, showing that SafeWatch outperforms all SOTA video guardrails on SafeWatch-Bench by 19.6% and 15.4% on existing benchmarks, while reducing inference cost by 25% on average. SafeWatch also demonstrates strong policy-following abilities and outperforms baselines by 20% in zero-shot adaptability to new policies. Additionally, both LLM-as-a-judge and human evaluators\nconfirm the high quality of the explanations provided by SafeWatch.", "keywords": [ "Video Guardrail Model", "Safe Foundation Models", "Efficient LLMs Inference", "LLM Safety", "Multimodal Foundation Models" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 3, 4, 2, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "xizpnYNvQq", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Mariko Kato", "gender": "unknown", "institution": "Japan Advanced Institute of Science and Technology", "country": "JP", "position": "MS student" }, { "name": "Naoya Inoue", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Visiting Researcher" }, { "name": "Yoshihiro Sakai", "gender": "Male", "institution": "Japan Advanced Institute of Science and Technology", "country": "JP", "position": "MS student" }, { "name": "Yufeng Zhao", "gender": "Male", "institution": "Japan Advanced Institute of Science and Technology", "country": "JP", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 6, "n_ref_uni": 36, "n_ref": 97, "n_ref_all": 183, "n_fig": 24, "n_tab": 8, "L_tab": 2788, "n_element_tab": 122, "n_fig_1": 9, "n_tab_1": 6, "L_tab_1": 3267, "n_element_tab_1": 122, "formula_len_all": 1921, "formula_len_all_1": 316, "len_all": 202750, "len_all_1": 72641, "len_abs": 2343, "len_title": 122, "len_sents": 69297, "len_sents_1": 33011, "n_sents": 447, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1612, "title": "Revisiting In-context Learning Inference Circuit in Large Language Models", "abs": "In-context Learning (ICL) is an emerging few-shot learning paradigm on Language Models (LMs) with inner mechanisms un-explored. There are already existing works describing the inner processing of ICL, while they struggle to capture all the inference phenomena in large language models. Therefore, this paper proposes a comprehensive circuit to model the inference dynamics and try to explain the observed phenomena of ICL. In detail, we divide ICL inference into 3 major operations: (1) Input Text Encode: LMs encode every input text (demonstrations and queries) into linear representation in the hidden states with sufficient information to solve ICL tasks. (2) Semantics Merge: LMs merge the encoded representations of demonstrations with their corresponding label tokens to produce joint representations of labels and demonstrations. (3) Feature Retrieval and Copy: LMs search the joint representations similar to the query representation on a task subspace, and copy the searched representations into the query. Then, language model heads capture these copied label representations to a certain extent and decode them into predicted labels. The proposed inference circuit successfully captured many phenomena observed during the ICL process, making it a comprehensive and practical explanation of the ICL inference process. Moreover, ablation analysis by disabling the proposed steps seriously damages the ICL performance, suggesting the proposed inference circuit is a dominating mechanism. Additionally, we confirm and list some bypass mechanisms that solve ICL tasks in parallel with the proposed circuit.", "keywords": [ "In-context Learning; Induction Circuit; Mechanistic Interpretability" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 4, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "xiyzCfXTS6", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Andreas Krause", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Melis Ilayda Bal", "gender": "Female", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Mojmir Mutny", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Pier Giuseppe Sessa", "gender": "Male", "institution": "Google DeepMind", "country": "CH", "position": "Researcher" } ], "n_formula": 28, "n_formula_1": 12, "n_ref_uni": 50, "n_ref": 125, "n_ref_all": 169, "n_fig": 19, "n_tab": 6, "L_tab": 2389, "n_element_tab": 261, "n_fig_1": 23, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3481, "formula_len_all_1": 1416, "len_all": 214994, "len_all_1": 70450, "len_abs": 1480, "len_title": 139, "len_sents": 71131, "len_sents_1": 32506, "n_sents": 526, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1573, "title": "Optimistic Games for Combinatorial Bayesian Optimization with Application to Protein Design", "abs": "Bayesian optimization (BO) is a powerful framework to optimize black-box expensive-to-evaluate functions via sequential interactions. In several important problems (e.g. drug discovery, circuit design, neural architecture search, etc.), though, such functions are defined over large $\\textit{combinatorial and unstructured}$ spaces. This makes existing BO algorithms not feasible due to the intractable maximization of the acquisition function over these domains. To address this issue, we propose $\\textbf{GameOpt}$, a novel game-theoretical approach to combinatorial BO. $\\textbf{GameOpt}$ establishes a cooperative game between the different optimization variables, and selects points that are game $\\textit{equilibria}$ of an upper confidence bound acquisition function. These are stable configurations from which no variable has an incentive to deviate$-$ analog to local optima in continuous domains. Crucially, this allows us to efficiently break down the complexity of the combinatorial domain into individual decision sets, making $\\textbf{GameOpt}$ scalable to large combinatorial spaces. We demonstrate the application of $\\textbf{GameOpt}$ to the challenging $\\textit{protein design}$ problem and validate its performance on four real-world protein datasets. Each protein can take up to $20^{X}$ possible configurations, where $X$ is the length of a protein, making standard BO methods infeasible. Instead, our approach iteratively selects informative protein configurations and very quickly discovers highly active protein variants compared to other baselines.", "keywords": [ "Combinatorial Bayesian Optimization", "Game Theory", "Gaussian Processes", "Protein Design" ], "rating_list": [ 3, 3, 8, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 4, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "xing7dDGh3", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Chandan Singh", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Researcher" }, { "name": "Jianfeng Gao", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Jingbo Shang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Yufan Zhuang", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Liyuan Liu", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 6, "n_ref_uni": 51, "n_ref": 78, "n_ref_all": 93, "n_fig": 7, "n_tab": 3, "L_tab": 438, "n_element_tab": 54, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 297, "n_element_tab_1": 42, "formula_len_all": 575, "formula_len_all_1": 277, "len_all": 166414, "len_all_1": 62156, "len_abs": 1097, "len_title": 119, "len_sents": 47878, "len_sents_1": 31244, "n_sents": 359, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1100, "title": "Vector-ICL: In-context Learning with Continuous Vector Representations", "abs": "Large language models (LLMs) have shown remarkable in-context learning (ICL) capabilities on textual data. We explore whether these capabilities can be extended to continuous vectors from diverse domains, obtained from black-box pretrained encoders. By aligning input data with an LLM's embedding space through lightweight projectors, we observe that LLMs can effectively process and learn from these projected vectors, which we term Vector-ICL. In particular, we find that pretraining projectors with general language modeling objectives enables Vector-ICL, while task-specific finetuning further enhances performance. In our experiments across various tasks and modalities, including text reconstruction, numerical function regression, text classification, summarization, molecule captioning, time-series classification, graph classification, and fMRI decoding, Vector-ICL often surpasses both few-shot ICL and domain-specific model or tuning. We further conduct analyses and case studies, indicating the potential of LLMs to process vector representations beyond traditional token-based paradigms.", "keywords": [ "large language models", "in-context learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "xiQNfYl33p", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Aditya T. Vadlamani", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Ali Payani", "gender": "Male", "institution": "Cisco", "country": "US", "position": "Researcher" }, { "name": "Anutam Srinivasan", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Undergrad student" }, { "name": "Pranav Maneriker", "gender": "Male", "institution": "Dolby Laboratories", "country": "US", "position": "Researcher" }, { "name": "srinivasan parthasarathy", "gender": "unknown", "institution": "Ohio State University", "country": "US", "position": "Professor" } ], "n_formula": 18, "n_formula_1": 16, "n_ref_uni": 36, "n_ref": 54, "n_ref_all": 83, "n_fig": 5, "n_tab": 3, "L_tab": 1026, "n_element_tab": 57, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 683, "n_element_tab_1": 55, "formula_len_all": 2546, "formula_len_all_1": 2365, "len_all": 119828, "len_all_1": 67192, "len_abs": 947, "len_title": 90, "len_sents": 31712, "len_sents_1": 27888, "n_sents": 251, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 958, "title": "A Generic Framework for Conformal Fairness", "abs": "Conformal Prediction (CP) is a popular method for uncertainty quantification with machine learning models. While the method provides probabilistic guarantees regarding the coverage of the true label, these guarantees are agnostic to the presence of sensitive attributes within the dataset. In this work, we formalize \\textit{Conformal Fairness}, a notion of fairness using conformal predictors, and provide a theoretically well-founded algorithm and associated framework to control for the gaps in coverage between different sensitive groups. Our framework leverages the exchangeability assumption (implicit to CP) rather than the typical IID assumption, allowing us to apply the notion of Conformal Fairness to data types and tasks that are not IID, such as graph data. Experiments were conducted on graph and tabular datasets to demonstrate that the algorithm can control fairness-related gaps in addition to coverage aligned with theoretical expectations.", "keywords": [ "Fairness", "Conformal Prediction", "Graph Neural Networks" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 2, 3, 4 ] }, { "paper_id": "xiDJaTim3P", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Chen Chen", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Jun Luo", "gender": "Male", "institution": "University of Pittsburgh", "country": "US", "position": "PhD student" }, { "name": "Shandong Wu", "gender": "Male", "institution": "University of Pittsburgh", "country": "", "position": "Associate Professor" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 36, "n_ref": 72, "n_ref_all": 85, "n_fig": 6, "n_tab": 8, "L_tab": 4694, "n_element_tab": 173, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2373, "n_element_tab_1": 44, "formula_len_all": 1254, "formula_len_all_1": 991, "len_all": 151122, "len_all_1": 57211, "len_abs": 1667, "len_title": 138, "len_sents": 42850, "len_sents_1": 25840, "n_sents": 297, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1261, "title": "Mixture of Experts Made Personalized: Federated Prompt Learning for Vision-Language Models", "abs": "Federated prompt learning benefits federated learning with CLIP-like Vision-Language Model's (VLM's) robust representation learning ability through prompt learning. However, current federated prompt learning methods are habitually restricted to the traditional FL paradigm, where the participating clients are generally only allowed to download a single globally aggregated model from the server. While justifiable for training full-sized models under federated settings, in this work, we argue that this paradigm is ill-suited for lightweight prompts. By facilitating the clients to download multiple pre-aggregated prompts as fixed non-local experts, we propose Personalized Federated Mixture of Adaptive Prompts (pFedMoAP), a novel FL framework that personalizes the prompt learning process through the lens of Mixture of Experts (MoE). pFedMoAP implements a local attention-based gating network that learns to generate enhanced text features for better alignment with local image data, benefiting from both local and downloaded non-local adaptive prompt experts. Extensive experiments on 9 datasets under various federated settings demonstrate the efficacy of the proposed pFedMoAP algorithm. The code is available at https://github.com/ljaiverson/pFedMoAP.", "keywords": [ "Federated learning", "prompt learning", "vision-language model", "mixture of experts" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 5, 4, 2 ] }, { "paper_id": "xi3sDtf8A0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Qianxiao Li", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Shiqi Jiang", "gender": "unknown", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Tan Minh Nguyen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Tho Tran Huu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 21, "n_ref_uni": 38, "n_ref": 74, "n_ref_all": 86, "n_fig": 6, "n_tab": 1, "L_tab": 216, "n_element_tab": 38, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 23, "n_element_tab_1": 1, "formula_len_all": 2480, "formula_len_all_1": 1842, "len_all": 132154, "len_all_1": 58819, "len_abs": 1441, "len_title": 123, "len_sents": 40427, "len_sents_1": 27162, "n_sents": 295, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1448, "title": "L-MSA: Layer-wise Fine-tuning using the Method of Successive Approximations", "abs": "With the emergence of large-scale models, the machine learning community has witnessed remarkable advancements. However, the substantial memory consumption associated with these models has emerged as a significant obstacle to large-scale training. To mitigate this challenge, an increasing emphasis has been placed on parameter-efficient fine-tuning methodologies, which adapt pre-trained models by fine-tuning only a subset of parameters. We observe that in various scenarios, fine-tuning different layers could lead to varying performance outcomes, and selectively fine-tuning certain layers has the potential to yield favorable performance results. Drawing upon this insight, we propose L-MSA, a novel layer-wise fine-tuning approach that integrates two key components: a metric for layer selection and an algorithm for optimizing the fine-tuning of the selected layers. By leveraging the principles of the Method of Successive Approximations, our method enhances model performance by targeting specific layers based on their unique characteristics and fine-tuning them efficiently. We also provide a theoretical analysis within deep linear networks, establishing a strong foundation for our layer selection criterion. Empirical evaluations across various datasets demonstrate that L-MSA identifies layers that yield superior training outcomes and fine-tunes them efficiently, consistently outperforming existing layer-wise fine-tuning methods.", "keywords": [ "layer-wise finetuning", "parameter-efficient fine-tuning", "method of successive approximations" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "xhtqgW5b93", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Shaoyi Zheng", "gender": "unknown", "institution": "New York University", "country": "US", "position": "Undergrad student" }, { "name": "Shengjie Wang", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Wenbo Lu", "gender": "Male", "institution": "New York University", "country": "CN", "position": "Undergrad student" }, { "name": "Yuxuan Xia", "gender": "Male", "institution": "New York University Shanghai", "country": "CN", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 19, "n_ref": 31, "n_ref_all": 37, "n_fig": 2, "n_tab": 6, "L_tab": 2137, "n_element_tab": 165, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 705, "n_element_tab_1": 66, "formula_len_all": 1133, "formula_len_all_1": 808, "len_all": 124571, "len_all_1": 57744, "len_abs": 1276, "len_title": 123, "len_sents": 37324, "len_sents_1": 26120, "n_sents": 288, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1173, "title": "ToMA: Token Merging with Attention For Diffusion Models", "abs": "Diffusion models have emerged as leading models for image generation. \nPlug-and-play token merging techniques have recently been introduced to mitigate the heavy computation cost of transformer blocks in diffusion models. \nHowever, existing methods overlook two key factors: 1. they fail to incorporate modern efficient implementation of attention, so that, the overhead backfires the achieved algorithmic efficiency 2. the selection of token to merge ignores the relation among tokens, limiting the image quality. \nIn this paper, we propose Token Merging with Attention(ToMA) with three major improvements. Firstly, we utilize a submodular-based token selection method to identify diverse tokens as merge destinations, representative of the entire token set. Secondly, we propose an attention merge, utilizing the efficient attention implementation, to perform the merge with negligible overhead. Also, we abstract the (un-)merging as (inverse-)linear transformations which also allows shareable transformation across layers/iterations. Finally, we utilize the image locality to further accelerate the computation by performing all the operations on tokens in local tiles.", "keywords": [ "Diffusion", "Token Merge", "Attention" ], "rating_list": [ 6, 6, 3, 3, 6 ], "soundness_list": [ 4, 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 1, 2, 3 ], "confidence_list": [ 2, 4, 5, 4, 4 ] }, { "paper_id": "xgtXkyqw1f", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Feng Zhao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jiangning Liu", "gender": "unknown", "institution": "Xi'an Jiaotong University", "country": "", "position": "MS student" }, { "name": "Kai Chen", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Kuikun Liu", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "", "position": "MS student" }, { "name": "Qiuchen Wang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Wenwei Zhang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Zehui Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 55, "n_ref": 79, "n_ref_all": 102, "n_fig": 10, "n_tab": 8, "L_tab": 3869, "n_element_tab": 166, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 821, "n_element_tab_1": 109, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 175874, "len_all_1": 54940, "len_abs": 2231, "len_title": 106, "len_sents": 46212, "len_sents_1": 26216, "n_sents": 295, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1920, "title": "MindSearch: Mimicking Human Minds Elicits Deep AI Searcher", "abs": "Information seeking and integration is a complex cognitive task that consumes enormous time and effort. Inspired by the remarkable progress of Large Language Models, recent works attempt to solve this task by combining LLMs and search engines. However, these methods still obtain unsatisfying performance due to three challenges: (1) complex requests often cannot be accurately and completely retrieved by the search engine once (2) corresponding information to be integrated is spread over multiple web pages along with massive noise, and (3) a large number of web pages with long contents may quickly exceed the maximum context length of LLMs. Inspired by the cognitive process when humans solve these problems, we introduce MindSearch to mimic the human minds in web information seeking and integration, which can be instantiated by a simple yet effective LLM-based multi-agent framework. The WebPlanner models the human mind of multi-step information seeking as a dynamic graph construction process: it decomposes the user query into atomic sub-questions as nodes in the graph and progressively extends the graph based on the search result from WebSearcher. Tasked with each sub-question, WebSearcher performs hierarchical information retrieval with search engines and collects valuable information for WebPlanner. The multi-agent design of MindSearch enables the whole framework to seek and integrate information parallelly from larger-scale (e.g., more than 300) web pages in 3 minutes, which is worth 3 hours of human effort. MindSearch demonstrates significant improvement in the response quality in terms of depth and breadth, on both close-set and open-set QA problems. Besides, responses from MindSearch based on InternLM2.5-7B are preferable by humans to ChatGPT-Web and Perplexity.ai applications, which implies that MindSearch can already deliver a competitive solution to the proprietary AI search engine.", "keywords": [ "language model", "search engine", "multi-agent system" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "xgQfWbV6Ey", "primary_area": "generative models", "n_author": 12, "author_info_list": [ { "name": "Ankur Taly", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Anush Mattapalli", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Chen-Yu Lee", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Jingbo Shang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Long Le", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Swaroop Mishra", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Tomas Pfister", "gender": "Male", "institution": "Google", "country": "", "position": "Head of Research @ Cloud AI" }, { "name": "Vincent Perot", "gender": "Male", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Yuwei Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Zifeng Wang", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Zilong Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Huaixiu Steven Zheng", "gender": "Male", "institution": "Google", "country": "", "position": "Software Engineer" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 49, "n_ref": 105, "n_ref_all": 123, "n_fig": 5, "n_tab": 11, "L_tab": 4405, "n_element_tab": 220, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 367, "n_element_tab_1": 11, "formula_len_all": 433, "formula_len_all_1": 401, "len_all": 195631, "len_all_1": 64907, "len_abs": 1306, "len_title": 123, "len_sents": 53638, "len_sents_1": 32846, "n_sents": 394, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1312, "title": "Speculative RAG: Enhancing Retrieval Augmented Generation through Drafting", "abs": "Retrieval augmented generation (RAG) combines the generative abilities of large language models (LLMs) with external knowledge sources to provide more accurate and up-to-date responses. Recent RAG advancements focus on improving retrieval outcomes through iterative LLM refinement or self-critique capabilities acquired through additional instruction tuning of LLMs. In this work, we introduce Speculative RAG - a framework that leverages a larger generalist LM to efficiently verify multiple RAG drafts produced in parallel by a smaller, distilled specialist LM. Each draft is generated from a distinct subset of retrieved documents, offering diverse perspectives on the evidence while reducing input token counts per draft. This approach enhances comprehension of each subset and mitigates potential position bias over long context. Our method accelerates RAG by delegating drafting to the smaller specialist LM, with the larger generalist LM performing a single verification pass over the drafts. Extensive experiments demonstrate that Speculative RAG achieves state-of-the-art performance with reduced latency on TriviaQA, MuSiQue, PopQA, PubHealth, and ARC-Challenge benchmarks. It notably enhances accuracy by up to 12.97% while reducing latency by 50.83% compared to conventional RAG systems on PubHealth.", "keywords": [ "generative model", "retrieval augmented generation" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "xfw92pDy2u", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "James Thornton", "gender": "unknown", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Joshua M. Susskind", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Navdeep Jaitly", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Ruixiang ZHANG", "gender": "Male", "institution": "Mila, UdeM", "country": "", "position": "PhD student" }, { "name": "Shuangfei Zhai", "gender": "Male", "institution": "Apple", "country": "US", "position": "Research Scientist" }, { "name": "Yizhe Zhang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Zijing Ou", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 23, "n_ref_uni": 42, "n_ref": 103, "n_ref_all": 125, "n_fig": 5, "n_tab": 2, "L_tab": 920, "n_element_tab": 68, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 920, "n_element_tab_1": 68, "formula_len_all": 2847, "formula_len_all_1": 1735, "len_all": 178684, "len_all_1": 71709, "len_abs": 1540, "len_title": 83, "len_sents": 43825, "len_sents_1": 29320, "n_sents": 333, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1560, "title": "Distilled Diffusion Language Models", "abs": "Transformer-based Large Language Models (LLMs) have demonstrated remarkable capa-\nbilities, yet their autoregressive nature forces sequential token-by-token decoding, leading\nto inefficiencies during inference. Furthermore, autoregressive language models lack in-\nherent self-correction abilities, which hinders their capacity to refine and improve gener-\nated content without relying on external prompting or retraining techniques. In contrast,\ndiffusion-based models offer the advantage of fast parallel generation through iterative\nrefinement, while leveraging bi-directional attention to utilize full context at once. How-\never, diffusion models are unable to match their autoregressive counterparts. This moti-\nvates us to explore the possibility of distilling a pre-trained autoregressive (AR) language\nmodel (teacher) into a non-autoregressive diffusion (non-AR) language model (student),\ncombining the best of both worlds. In this work, we present Target Concrete Score (TCS)\ndistillation, a theoretically grounded framework that bridges autoregressive and diffusion\nparadigms. TCS distillation is broadly applicable to both discrete and continuous diffu-\nsion models, with any pre-trained autoregressive teacher model. We propose techniques\nto make TCS distillation scalable and efficient for transformer-based models, and show\nhow it can both improve pre-trained diffusion language models and also train new mod-\nels from scratch. Through comprehensive experiments on language modeling tasks, we\ndemonstrate the effectiveness of our proposed methods.", "keywords": [ "diffusion language models", "discrete diffusion", "distillation" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "xeP03R58RH", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Kajetan Schweighofer", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "", "position": "PhD student" }, { "name": "Lukas Aichberger", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Sepp Hochreiter", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 44, "n_ref": 110, "n_ref_all": 129, "n_fig": 3, "n_tab": 4, "L_tab": 4256, "n_element_tab": 486, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 767, "n_element_tab_1": 113, "formula_len_all": 1082, "formula_len_all_1": 1019, "len_all": 152222, "len_all_1": 76869, "len_abs": 1306, "len_title": 112, "len_sents": 38056, "len_sents_1": 31284, "n_sents": 266, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1313, "title": "Rethinking Uncertainty Estimation in Natural Language Generation", "abs": "Large language models (LLMs) are increasingly employed in real-world applications, driving a need to determine when their generated text can be trusted or should be questioned. To assess the trustworthiness of the generated text, reliable uncertainty estimation is essential. Current LLMs generate text through a stochastic process that can lead to different output sequences for the same prompt. Consequently, leading uncertainty measures require generating multiple output sequences to estimate the LLM\u2019s uncertainty. However, generating additional output sequences is computationally expensive, making these uncertainty estimates impractical at scale. In this work, we challenge the theoretical foundations of the leading measures and derive an alternative measure that eliminates the need for generating multiple output sequences. Our new measure is based solely on the negative log-likelihood of the most likely output sequence. This vastly simplifies uncertainty estimation while maintaining theoretical rigor. Empirical results demonstrate that our new measure achieves state-of-the-art performance across various models and tasks. Our work lays the foundation for reliable and efficient uncertainty estimation in LLMs, challenging the necessity of the more complicated methods currently leading the field.", "keywords": [ "llm", "nlg", "uncertainty estimation", "uncertainty measures", "proper scoring rules" ], "rating_list": [ 3, 3, 3, 5, 6 ], "soundness_list": [ 2, 1, 2, 3, 2 ], "presentation_list": [ 1, 1, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3, 3 ] }, { "paper_id": "xdGsiYNfje", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "Jun Sun", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Full Professor" }, { "name": "Mengdi Zhang", "gender": "Female", "institution": "Singapore Management University", "country": "SG", "position": "Postdoc" }, { "name": "Peixin Zhang", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Postdoc" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 51, "n_ref": 87, "n_ref_all": 104, "n_fig": 6, "n_tab": 6, "L_tab": 2525, "n_element_tab": 315, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2493, "n_element_tab_1": 316, "formula_len_all": 499, "formula_len_all_1": 204, "len_all": 200624, "len_all_1": 67015, "len_abs": 1077, "len_title": 98, "len_sents": 46445, "len_sents_1": 30461, "n_sents": 355, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1080, "title": "LLMScan: Causal Scan for LLM Misbehavior Detection", "abs": "Despite the success of Large Language Models (LLMs) across various fields, their potential to generate untruthful, biased and harmful responses poses significant risks, particularly in critical applications. This highlights the urgent need for systematic methods to detect and prevent such misbehavior. While existing approaches target specific issues such as harmful responses, this work introduces LLMScan, an innovative LLM monitoring technique based on causality analysis, offering a comprehensive solution. LLMScan systematically monitors the inner workings of an LLM through the lens of causal inference, operating on the premise that the LLM's `brain' behaves differently when misbehaving. By analyzing the causal contributions of the LLM's input tokens and transformer layers, LLMScan effectively detects misbehavior. Extensive experiments across various tasks and models reveal clear distinctions in the causal distributions between normal behavior and misbehavior, enabling the development of accurate, lightweight detectors for a variety of misbehavior detection tasks.", "keywords": [ "Large Language Model", "LLM Safety", "LLM Misbehavior Detection", "Causality Analysis", "Model Scan" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 1, 2 ], "presentation_list": [ 3, 1, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "xcPN6Or88c", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Haishuai Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Research Professor" }, { "name": "Hongwei Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Jiajun Bu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Jialong Guo", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Ke Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "MENGXUAN LI", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 20, "n_ref_uni": 38, "n_ref": 58, "n_ref_all": 66, "n_fig": 4, "n_tab": 3, "L_tab": 5187, "n_element_tab": 94, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1242, "n_element_tab_1": 55, "formula_len_all": 868, "formula_len_all_1": 817, "len_all": 108402, "len_all_1": 56836, "len_abs": 3068, "len_title": 149, "len_sents": 26575, "len_sents_1": 25001, "n_sents": 211, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 1864, "title": "ImputeINR: Enhancing Time Series Imputation with Adaptive Group-based Implicit Neural Representations", "abs": "Time series data frequently exhibit the presence of missing values, rendering imputation a crucial process for downstream time series tasks and applications. However, existing imputation methods focus on discrete data points and are unable to effectively model sparse data, resulting in particularly poor performance for imputing substantial missing values. In this paper, we propose a novel approach, ImputeINR, for time series imputation by employing implicit neural representations (INR) to learn continuous functions for time series. ImputeINR leverages the merits of INR that the continuous functions are not coupled to sampling frequency and have infinite sampling frequency, allowing ImputeINR to generate fine-grained imputations even on extremely absent observed values. In addition, we introduce a multi-scale feature extraction module in ImputeINR architecture to capture patterns from different time scales, thereby effectively enhancing the fine-grained and global consistency of the imputation. To address the unique challenges of complex temporal patterns and multiple variables in time series, we design a specific form of INR continuous function that contains three additional components to learn trend, seasonal, and residual information separately. Furthermore, we innovatively propose an adaptive group-based framework to model complex residual information, where variables with similar distributions are modeled by the same group of multilayer perception layers to extract necessary correlation features. Since the number of groups and their output variables are determined by variable clustering, ImputeINR has the capacity of adapting to diverse datasets. Extensive experiments conducted on seven datasets with five ratios of missing values demonstrate the superior performance of ImputeINR, especially for high absent ratios in time series.", "keywords": [ "time series imputation", "implicit neural representations" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "xcHIiZr3DT", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Cai Yaobang", "gender": "Male", "institution": "SZTU", "country": "CN", "position": "Undergrad student" }, { "name": "Jianhao", "gender": "unknown", "institution": "Shenzhen University", "country": "CN", "position": "MS student" }, { "name": "Teng Yan", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "MS student" }, { "name": "Tian Xia", "gender": "Male", "institution": "Shenzhen University", "country": "", "position": "MS student" }, { "name": "Wenxian Li", "gender": "unknown", "institution": "sino German intelligent manufacture", "country": "CN", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 16, "n_ref": 18, "n_ref_all": 33, "n_fig": 6, "n_tab": 6, "L_tab": 999, "n_element_tab": 100, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 84, "n_element_tab_1": 4, "formula_len_all": 473, "formula_len_all_1": 477, "len_all": 71193, "len_all_1": 48737, "len_abs": 1311, "len_title": 139, "len_sents": 25985, "len_sents_1": 24262, "n_sents": 165, "n_sents_1": 145, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1074, "title": "Vision-Based Pseudo-Tactile Information Extraction and Localization for Dexterous Grasping", "abs": "This study addresses the challenges of tactile perception in robotic dexterous hand grasping by focusing on two main tasks: 1) Acquiring tactile information from everyday objects using vision, termed \"pseudo-tactile\" information, and 2) Building a Dexterous Hand (RH8D) model in Isaac Sim for real-time fingertip contact localization. Utilizing Isaac Sim enables safe, cost-effective experimentation and high-precision simulations that facilitate data collection for model validation. The research establishes a scientific connection between simulated 3D coordinates, actual 3D coordinates, and pseudo-tactile information derived from point clouds, quantified through normal vectors and grayscale variance analysis. Results demonstrate the ability to extract clear object surface textures, accurately locate fingertip contact points in real-time (with precision up to $0.001 m$), and provide tactile information at contact points. This framework enhances robotic grasping capabilities and offers low-cost sensory data. The source code and dataset are publicly available now.", "keywords": [ "Pseudo-Tactile Information", "Dexterous Grasping", "Vision-Based Perception", "Robotic Localization" ], "rating_list": [ 3, 3, 1, 3 ], "soundness_list": [ 2, 2, 1, 1 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 1, 1, 1, 1 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "xbXydoejvY", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Chenyi Zhuang", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Cunzheng Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Gaoang Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jinjie GU", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Tao Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Xi Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "mingxuan cui", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 41, "n_ref": 69, "n_ref_all": 91, "n_fig": 7, "n_tab": 7, "L_tab": 1765, "n_element_tab": 236, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1316, "n_element_tab_1": 139, "formula_len_all": 851, "formula_len_all_1": 589, "len_all": 137861, "len_all_1": 59704, "len_abs": 1611, "len_title": 117, "len_sents": 39447, "len_sents_1": 26532, "n_sents": 306, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1647, "title": "CWPS: Efficient Channel-Wise Parameter Sharing for Knowledge Transfer", "abs": "Knowledge transfer aims to apply existing knowledge to different tasks or new data, and it has extensive applications in multi-domain and multi-task learning.\n The key to this task is quickly identifying a fine-grained object for knowledge sharing and efficiently transferring knowledge.\n Current methods, such as fine-tuning, layer-wise parameter sharing, and task-specific adapters, only offer coarse-grained sharing solutions and struggle to effectively search for shared parameters, thus hindering the performance and efficiency of knowledge transfer.\n To address these issues, we propose Channel-Wise Parameter Sharing (CWPS), a novel fine-grained parameter-sharing method for Knowledge Transfer, which is efficient for parameter sharing, comprehensive, and plug-and-play.\n For the coarse-grained problem, we first achieve fine-grained parameter sharing by refining the granularity of shared parameters from the level of layers to the level of neurons. The knowledge learned from previous tasks can be utilized through the explicit composition of the model neurons.\n Besides, we promote an effective search strategy to minimize computational costs, simplifying the process of determining shared weights.\n In addition, our CWPS has strong composability and generalization ability, which theoretically can be applied to any network consisting of linear and convolution layers.\n We introduce several datasets in both incremental learning and multi-task learning scenarios. Our method has achieved state-of-the-art precision-to-parameter ratio performance with various backbones, demonstrating its efficiency and versatility.", "keywords": [ "Transfer Learning", "Multi-Domain Learning", "Multi-Task Learning" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "xbW6EGve6a", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Hai Dong", "gender": "unknown", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "Associate Professor" }, { "name": "Ziru Niu", "gender": "Male", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "PhD student" }, { "name": "A. Kai Qin", "gender": "unknown", "institution": "Swinburne University of Technology", "country": "AU", "position": "Full Professor" } ], "n_formula": 64, "n_formula_1": 23, "n_ref_uni": 33, "n_ref": 93, "n_ref_all": 134, "n_fig": 16, "n_tab": 2, "L_tab": 931, "n_element_tab": 169, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1736, "n_element_tab_1": 100, "formula_len_all": 6165, "formula_len_all_1": 1546, "len_all": 154946, "len_all_1": 61525, "len_abs": 1149, "len_title": 157, "len_sents": 39583, "len_sents_1": 24955, "n_sents": 327, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1153, "title": "Energy and Memory-Efficient Federated Learning with Ordered Layer Freezing and Tensor Operation Approximation", "abs": "The effectiveness of Federated Learning (FL) in the context of the Internet of Things (IoT) is hindered by the resource constraints of IoT devices, such as limited computing capability, memory space and bandwidth support. These constraints create significant computation and communication bottlenecks for training and transmitting deep neural networks. Various FL frameworks have been proposed to reduce computation and communication overheads through dropout or layer freezing. However, these approaches often sacrifice accuracy or neglect memory constraints. In this work, we introduce Federated Learning with Ordered Layer Freezing (FedOLF) to improve energy efficiency and reduce memory footprint while maintaining accuracy. Additionally, we employ the Tensor Operation Approximation technique to reduce the communication (and accordingly energy) cost, which can better preserve accuracy compared to traditional quantization methods. Experimental results demonstrate that FedOLF achieves higher accuracy and energy efficiency as well as lower memory footprint across EMNIST, CIFAR-10, CIFAR-100, and CINIC-10 benchmarks compared to existing methods.", "keywords": [ "Federated Learning", "Resource-Constrained devices", "Computation and Communication Overheads", "Layer Freezing", "Tensor Operation Approximation" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "xayT1nn8Mg", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Chunyang Li", "gender": "Male", "institution": "Chengdu Institute of Biological Products Co.\uff0c Ltd.", "country": "CN", "position": "Associate Professor" }, { "name": "Haoliang Li", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Hong Yan", "gender": "unknown", "institution": "City University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Mengxu ZHU", "gender": "unknown", "institution": "City University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Terry Lyons", "gender": "Male", "institution": "Alan Turing Institute", "country": "", "position": "Full Professor" }, { "name": "Tiexin Qin", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 17, "n_ref_uni": 39, "n_ref": 59, "n_ref_all": 91, "n_fig": 7, "n_tab": 4, "L_tab": 495, "n_element_tab": 46, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 658, "n_element_tab_1": 54, "formula_len_all": 1141, "formula_len_all_1": 984, "len_all": 156094, "len_all_1": 69724, "len_abs": 1232, "len_title": 114, "len_sents": 51250, "len_sents_1": 33777, "n_sents": 378, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1236, "title": "Deep Signature: Characterization of Large-Scale Molecular Dynamics", "abs": "Understanding protein dynamics are essential for deciphering protein functional mechanisms and developing molecular therapies. However, the complex high-dimensional dynamics and interatomic interactions of biological processes pose significant challenge for existing computational techniques. In this paper, we approach this problem for the first time by introducing Deep Signature, a novel computationally tractable framework that characterizes complex dynamics and interatomic interactions based on their evolving trajectories. Specifically, our approach incorporates soft spectral clustering that locally aggregates cooperative dynamics to reduce the size of the system, as well as signature transform that collects iterated integrals to provide a global characterization of the non-smooth interactive dynamics. Theoretical analysis demonstrates that Deep Signature exhibits several desirable properties, including invariance to translation, near invariance to rotation, equivariance to permutation of atomic coordinates, and invariance under time reparameterization. Furthermore, experimental results on three benchmarks of biological processes verify that our approach can achieve superior performance compared to baseline methods.", "keywords": [ "Molecular dynamics; representation learning; graph neural network; path signature" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 2, 3, 3 ] }, { "paper_id": "xawA8X5dHq", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Coralie Hemptinne", "gender": "Female", "institution": "Universit\u00e9 catholique de Louvain", "country": "BE", "position": "Associate Professor" }, { "name": "Demet YUKSEL", "gender": "Female", "institution": "Cliniques universitaires Saint-Luc", "country": "BE", "position": "Full Professor" }, { "name": "Jean Vanderdonckt", "gender": "Male", "institution": "UCL", "country": "", "position": "Full Professor" }, { "name": "Maxime Griot", "gender": "Male", "institution": "Louvain School of Management", "country": "BE", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 39, "n_ref_all": 51, "n_fig": 2, "n_tab": 3, "L_tab": 1353, "n_element_tab": 105, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 304, "n_element_tab_1": 14, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 127327, "len_all_1": 52868, "len_abs": 1504, "len_title": 143, "len_sents": 34052, "len_sents_1": 27206, "n_sents": 216, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1512, "title": "Multiple Choice Questions and Large Languages Models: A Case Study with Fictional Medical Data", "abs": "Large Language Models (LLMs) like ChatGPT demonstrate significant potential in the medical field, often evaluated using multiple-choice questions (MCQs) similar to those found on the USMLE. Despite their prevalence in medical education, MCQs have limitations that might be exacerbated when assessing LLMs. To evaluate the effectiveness of MCQs in assessing the performance of LLMs, we developed a fictional medical benchmark focused on a non-existent gland, the Glianorex. This approach allowed us to isolate the knowledge of the LLM from its test-taking abilities. We used GPT-4-Turbo and Claude 3.5 Sonnet to generate two comprehensive textbooks on the Glianorex in both English and French and developed corresponding multiple-choice questions in both languages. We evaluated various open-source, proprietary, and domain-specific LLMs using these questions in a zero-shot setting. The models achieved average scores around 64%, with minor performance differences between larger and smaller models. Performance was slightly higher in English than in French. Fine-tuned medical models showed some improvement over their base versions in English but not in French. The high performance across models suggests that traditional MCQ-based benchmarks may not accurately measure LLMs' clinical knowledge and reasoning abilities, instead highlighting their pattern recognition skills. This study underscores the need for more robust evaluation methods to better assess the true capabilities of LLMs in medical contexts.", "keywords": [ "large language models", "medicine", "benchmark", "evaluation", "clinical knowledge", "multiple choice questions" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "xam3sR3ffY", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Aman Singh Thakur", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "MS student" }, { "name": "Dieuwke Hupkes", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Kartik Choudhary", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "MS student" }, { "name": "Sankaran Vaidyanathan", "gender": "unknown", "institution": "Department of Computer Science, University of Massachusetts at Amherst", "country": "", "position": "PhD student" }, { "name": "Venkat Srinik Ramayapally", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 2, "n_ref_uni": 42, "n_ref": 67, "n_ref_all": 117, "n_fig": 25, "n_tab": 6, "L_tab": 2364, "n_element_tab": 204, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 690, "n_element_tab_1": 33, "formula_len_all": 650, "formula_len_all_1": 247, "len_all": 172108, "len_all_1": 61419, "len_abs": 1615, "len_title": 126, "len_sents": 51958, "len_sents_1": 29582, "n_sents": 389, "n_sents_1": 226, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1619, "title": "Judging the Judges: Evaluating Alignment and Vulnerabilities in LLMs-as-Judges", "abs": "Offering a promising solution to the scalability challenges associated with human evaluation, the LLM-as-a-judge paradigm is rapidly gaining traction as an approach to evaluating large language models (LLMs). However, there are still many open questions about the strengths and weaknesses of this paradigm, and what potential biases it may hold. In this paper, we present a comprehensive study of the performance of various LLMs acting as judges, focusing on a clean scenario in which inter-human agreement is high. Investigating thirteen judge models of different model sizes and families, judging answers of nine different \u2018examtaker models\u2019 \u2013 both base and instruction-tuned \u2013 we find that only the best (and largest) models achieve reasonable alignment with humans. However, they are still quite far behind inter-human agreement and their assigned scores may still differ with up to 5 points from human-assigned scores. In terms of their ranking of the nine exam-taker models, instead, also smaller models and even the lexical metric contains may provide a reasonable signal. Through error analysis and other studies, we identify vulnerabilities in judge models, such as their sensitivity to prompt complexity and length, and a tendency toward leniency. The fact that even the best judges differ from humans in this comparatively simple setup suggest that caution may be wise when using judges in more complex setups. Lastly, our research rediscovers the importance of using alignment metrics beyond simple percent alignment, showing that judges with high percent agreement can still assign vastly different scores.", "keywords": [ "LLMs", "NLP", "LLM Evaluation", "LLM-as-a-Judge", "Benchmarks" ], "rating_list": [ 8, 5, 5, 3, 3 ], "soundness_list": [ 4, 1, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 3, 1, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "xak8c9l1nu", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "A. Pavan", "gender": "unknown", "institution": "Iowa State University", "country": "US", "position": "Full Professor" }, { "name": "Arnab Bhattacharyya", "gender": "Male", "institution": "The University of Warwick", "country": "GB", "position": "Associate Professor" }, { "name": "Dimitrios Myrisiotis", "gender": "unknown", "institution": "CNRS@CREATE LTD.", "country": "", "position": "Postdoc" }, { "name": "Kuldeep S. Meel", "gender": "Male", "institution": "Department of Computer Science", "country": "", "position": "Associate Professor" }, { "name": "Sutanu Gayen", "gender": "Male", "institution": "Indian Institute of Technology, Kanpur", "country": "", "position": "Assistant Professor" }, { "name": "N.V. Vinodchandran", "gender": "Male", "institution": "University of Nebraska, Lincoln", "country": "US", "position": "Professor" } ], "n_formula": 44, "n_formula_1": 29, "n_ref_uni": 31, "n_ref": 40, "n_ref_all": 41, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 136, "n_element_tab_1": 23, "formula_len_all": 5803, "formula_len_all_1": 4530, "len_all": 93811, "len_all_1": 55747, "len_abs": 543, "len_title": 103, "len_sents": 25984, "len_sents_1": 20645, "n_sents": 280, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 54, "L_abs": 571, "title": "Computational Explorations of Total Variation Distance", "abs": "We investigate some previously unexplored (or underexplored) computational aspects of total variation (TV) distance.\nFirst, we give a simple deterministic polynomial-time algorithm for checking equivalence between mixtures of product distributions, over arbitrary alphabets.\nThis corresponds to a special case, whereby the TV distance between the two distributions is zero.\nSecond, we prove that unless $\\mathsf{NP} \\subseteq \\mathsf{RP}$ it is impossible to efficiently estimate the TV distance between arbitrary Ising models, even in a bounded-error randomized setting.", "keywords": [ "total variation distance", "TV distance", "mixtures of products", "equivalence checking", "Ising models", "computational complexity", "FPRAS" ], "rating_list": [ 8, 6, 8, 6, 8 ], "soundness_list": [ 4, 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 4, 3 ] }, { "paper_id": "xajif1l65R", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Hong Liu", "gender": "Non-Binary", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Yangze Liu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Undergrad student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 57, "n_ref_all": 74, "n_fig": 9, "n_tab": 5, "L_tab": 600, "n_element_tab": 49, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 376, "n_element_tab_1": 38, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 108294, "len_all_1": 66170, "len_abs": 1332, "len_title": 149, "len_sents": 38286, "len_sents_1": 32425, "n_sents": 280, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1361, "title": "Rethinking Dataset Quantization: Efficient Core Set Selection via Semantically-Aware Data Augmentation", "abs": "Dataset quantization (DQ) is an innovative coreset selection method to choose representative subsets from large-scale datasets, such as ImageNet. Although DQ has made significant progress, it heavily relies on large pre-trained models (like MAEs), leading to substantial additional computational overhead. We first identify that removing this pre-trained MAE model degrades DQ\u2019s performance and increases the variance in model training. Where MAE plays a crucial role in introducing prior knowledge and implicit regularization into the training process. Second, we investigate a data augmentation scheme that can simulate the steps of pixel compression and reconstruction in DQ by simply using a randomly initialized ResNet model. This randomly initialized ResNet model can take advantage of the inductive bias of CNNs to locate the semantic object region and then replace the other region with other images. Therefore, we can use a random model or trained model in the early training stage to enhance semantic diversity while selecting important samples. We remove the module that contains the pre-trained MAE model and integrate the data augmentation scheme into the DQ pipeline, which formulates a new simple but efficient method, called DQ v2. Our method achieves performance improvements across multiple datasets, such as ImageNette, CUB-200, and Food-101.", "keywords": [ "Coreset Selection", "Dataset Quantization", "Data Augmentation", "Efficient Deep Learning", "Semantically-Aware Augmentation" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "xaafWdM5jI", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Dai Shi", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Researcher" }, { "name": "Junbin Gao", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Full Professor" }, { "name": "Ruikun Li", "gender": "Male", "institution": "The University of Sydney", "country": "", "position": "PhD student" }, { "name": "Ye Xiao", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 40, "n_ref": 67, "n_ref_all": 91, "n_fig": 6, "n_tab": 6, "L_tab": 6318, "n_element_tab": 596, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 3795, "n_element_tab_1": 298, "formula_len_all": 1103, "formula_len_all_1": 1031, "len_all": 151978, "len_all_1": 75831, "len_abs": 1379, "len_title": 155, "len_sents": 45960, "len_sents_1": 30742, "n_sents": 296, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1400, "title": "UFGTime: Reforming the Pure Graph Paradigm for Multivariate Time Series Forecasting in the Frequency Domain", "abs": "Recent advances in multivariate time series forecasting have seen a shift toward a pure graph paradigm, which transforms time series into hypervariate graphs and employs graph neural networks (GNNs) to holistically capture intertwined spatiotemporal dependencies. While promising, this approach faces notable challenges. First, converting time series into hypervariate graphs often neglects essential temporal sequences, which are vital for accurately capturing temporal dependencies. Second, treating the graph as a complete structure can obscure the varying importance of intra- and inter-series connections, potentially overlooking key local patterns. To address these challenges, we introduce a novel hyperspectral graph data structure that embeds sequential order into frequency signals and employs a sparse yet meaningful topological structure. In addition, we propose the \\textsc{Ufgtime} framework, featuring a frequency-based global graph framelet message-passing operator tailored to hyperspectral graphs, effectively mitigating the smoothing issue and capturing global insights through sparse connections. Extensive experiments demonstrate that our framework significantly surpasses state-of-the-art methods, excelling in both short- and long-range time series forecasting while achieving superior efficiency. Our code is available at:~\\url{https://anonymous.4open.science/r/UFGTIME-E352}.", "keywords": [ "Multivariate Time Series Forecasting", "GNN", "Pure Graph Paradigm" ], "rating_list": [ 1, 5, 5, 3 ], "soundness_list": [ 1, 2, 3, 1 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "xaYlO03tIk", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Huazhe Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Kaizhe Hu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Pu Hua", "gender": "Male", "institution": "Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yao He", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Yuyao Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Zihang Rui", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 19, "n_formula_1": 13, "n_ref_uni": 40, "n_ref": 69, "n_ref_all": 101, "n_fig": 13, "n_tab": 9, "L_tab": 2963, "n_element_tab": 301, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 627, "n_element_tab_1": 67, "formula_len_all": 1455, "formula_len_all_1": 691, "len_all": 162842, "len_all_1": 62263, "len_abs": 442, "len_title": 115, "len_sents": 47753, "len_sents_1": 28981, "n_sents": 407, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 114, "L_abs": 1276, "title": "Stem-OB: Generalizable Visual Imitation Learning with Stem-Like Convergent Observation through Diffusion Inversion", "abs": "Visual imitation learning methods demonstrate strong performance, yet they lack generalization when faced with visual input perturbations like variations in lighting and textures. This limitation hampers their practical application in real-world settings. To address this, we propose ***Stem-OB*** that leverages the inversion process of pretrained image diffusion models to suppress low-level visual differences while maintaining high-level scene structures. This image inversion process is akin to transforming the observation into a shared representation, from which other observations also stem. *Stem-OB* offers a simple yet effective plug-and-play solution that stands in contrast to data augmentation approaches. It demonstrates robustness to various unspecified appearance changes without the need for additional training. We provide theoretical insights and empirical results that validate the efficacy of our approach in simulated and real settings. *Stem-OB* shows an exceptionally significant improvement in real-world robotic tasks, where challenging light and appearance changes are present, with an average increase of **22.2%** in success rates compared to the best baseline. Please refer to [this link](https://stem-ob.github.io/) for more videos and details.", "keywords": [ "Robotics", "Imitation Learning", "Visual Imitation Learning", "Robustness", "Diffusion Model", "Diffusion Inversion" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "xaXvHdH9Y4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Lim Sze Ying", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Undergrad student" }, { "name": "Rick Siow Mong Goh", "gender": "unknown", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "SG", "position": "Director" }, { "name": "Tao Luo", "gender": "Male", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "", "position": "Researcher" }, { "name": "zhehui wang", "gender": "unknown", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 22, "n_ref": 49, "n_ref_all": 65, "n_fig": 1, "n_tab": 5, "L_tab": 4618, "n_element_tab": 241, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 2500, "n_element_tab_1": 215, "formula_len_all": 360, "formula_len_all_1": 359, "len_all": 102212, "len_all_1": 56792, "len_abs": 1657, "len_title": 121, "len_sents": 29267, "len_sents_1": 23833, "n_sents": 220, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1665, "title": "P-BERT: Hardware-Aware Optimization of BERT Using Evolutionary Techniques", "abs": "Transformer-based models have emerged as the go-to standards in Natural Language Processing (NLP), revolutionizing the landscape of NLP applications. As complex models continue to proliferate, the need for more efficient computational processing becomes increasingly imperative. This has led to the rise of model compression techniques, implemented to target computational inefficiencies. Expounding on this, we propose Pyramid-BERT (P-BERT), the integration of three established model compression techniques to further reduce the computational inefficiency of the standard BERT models, and subsequently optimize BERT under the hardware characteristics. Specifically, the techniques employed are pruning, quantization, and knowledge distillation. The first two aforementioned correlated techniques work simultaneously to remove redundant specifications while leveraging knowledge transfer from baseline models. These techniques enable a substantial reduction in computational cost, making P-BERT highly suitable for portable, low-power devices such as cellphones, wearable devices, and smartwatches, and thus enabling hardware-friendly processing on various computing engines. Additionally, we will be proposing a new metric, the inverted computational complexity to quantify the complexity and efficacy of the model. This metric aims to more accurately capture the hardware-specific performance characteristics. Our experimental results show that P-BERT achieves a remarkable reduction of at least 60\\% in the inverted computational complexity ratio while ensuring comparable accuracy and scores across many downstream tasks compared with the baseline BERT models.", "keywords": [ "Model Compression", "Large Language Models", "Computation Complexity", "BERT", "Hardware-Aware" ], "rating_list": [ 3, 3, 3, 5, 5 ], "soundness_list": [ 3, 1, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 1, 1, 2, 2, 2 ], "confidence_list": [ 1, 5, 4, 3, 3 ] }, { "paper_id": "xZ2lTzfyFv", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 6, "author_info_list": [ { "name": "Ngoc-Quan Pham", "gender": "Male", "institution": "VinAI Research", "country": "VN", "position": "Researcher" }, { "name": "Nhat Ho", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Quyen Tran", "gender": "Female", "institution": "VinAI Research", "country": "VN", "position": "Research Resident" }, { "name": "Trung Le", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" }, { "name": "Tuan Truong", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Undergrad student" }, { "name": "Dinh Phung", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" } ], "n_formula": 54, "n_formula_1": 33, "n_ref_uni": 46, "n_ref": 68, "n_ref_all": 82, "n_fig": 4, "n_tab": 5, "L_tab": 1952, "n_element_tab": 166, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1155, "n_element_tab_1": 161, "formula_len_all": 5385, "formula_len_all_1": 3068, "len_all": 142323, "len_all_1": 62265, "len_abs": 950, "len_title": 109, "len_sents": 36655, "len_sents_1": 26262, "n_sents": 284, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 960, "title": "Improving Generalization with Flat Hilbert Bayesian Inference", "abs": "We introduce Flat Hilbert Bayesian Inference (FHBI), an algorithm designed to enhance generalization in Bayesian inference. Our approach involves an iterative two-step procedure with an adversarial functional perturbation step and a functional descent step within the reproducing kernel Hilbert spaces. This methodology is supported by a theoretical analysis that extends previous findings on generalization ability from finite-dimensional Euclidean spaces to infinite-dimensional functional spaces. To evaluate the effectiveness of FHBI, we conduct comprehensive comparisons against seven baseline methods on the VTAB-1K benchmark, which encompasses 19 diverse datasets across various domains with diverse semantics. Empirical results demonstrate that FHBI consistently outperforms the baselines by notable margins, highlighting its practical efficacy. Our code is available at \\url{https://anonymous.4open.science/r/Flat-Hilbert-Variational-Inference-008F/}.", "keywords": [ "Bayesian Inference", "Sharpness-aware Minimization" ], "rating_list": [ 6, 8, 3, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "xYzOkOGD96", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Cordelia Schmid", "gender": "Female", "institution": "Google", "country": "FR", "position": "Researcher" }, { "name": "Josef Sivic", "gender": "Male", "institution": "Czech Technical University in Prague", "country": "CZ", "position": "Principal investigator" }, { "name": "Evangelos Kazakos", "gender": "Male", "institution": "Czech Technical University of Prague", "country": "CZ", "position": "Postdoc" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 49, "n_ref": 103, "n_ref_all": 132, "n_fig": 11, "n_tab": 6, "L_tab": 1167, "n_element_tab": 167, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 280, "formula_len_all_1": 56, "len_all": 187471, "len_all_1": 68053, "len_abs": 1465, "len_title": 81, "len_sents": 53978, "len_sents_1": 35368, "n_sents": 395, "n_sents_1": 243, "venue": "ICLR2025", "decision": "withdraw", "L_title": 33, "L_abs": 1476, "title": "Grounded Video Caption Generation", "abs": "We propose a new task, dataset and model for grounded video caption generation. This task unifies captioning and object grounding in video, where the objects in the caption are grounded in the video via temporally consistent bounding boxes. We introduce the following contributions. First, we present a task definition and a manually annotated test dataset for this task, referred to as GROunded Video Caption Generation (GROC). Second, we introduce a large-scale automatic annotation method leveraging an existing model for grounded still image captioning together with an LLM for summarising frame-level captions into temporally consistent captions in video. \nFurthermore, we prompt the LLM to track by language \u2013 classifying noun phrases from the frame-level captions into noun phrases of the video-level generated caption. We apply this approach to videos from the HowTo100M dataset, which results in a new large-scale training dataset, called HowToGround, with automatically annotated captions and spatio-temporally consistent bounding boxes with coherent natural language labels. Third, we introduce a new grounded video caption generation model, called VideoGLaMM, and train the model on the new automatically annotated HowToGround dataset. Finally, results of our VideoGLaMM model set the state of the art for the new task of grounded video caption generation. We perform extensive ablations and demonstrate the importance of key technical contributions of our model.", "keywords": [ "vision-language models", "VLM", "LLM", "video grounding", "automatic annotation", "pseudo-labeling" ], "rating_list": [ 3, 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 1, 4, 2 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4, 5 ] }, { "paper_id": "xYquBPHppn", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Hadi Amiri", "gender": "Not Specified", "institution": "University of Massachusetts Lowell", "country": "US", "position": "Assistant Professor" }, { "name": "Nidhi Vakil", "gender": "unknown", "institution": "University of Massachusetts, Lowell", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 26, "n_ref": 40, "n_ref_all": 54, "n_fig": 5, "n_tab": 6, "L_tab": 744, "n_element_tab": 156, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1232, "n_element_tab_1": 161, "formula_len_all": 344, "formula_len_all_1": 344, "len_all": 89520, "len_all_1": 61810, "len_abs": 1206, "len_title": 131, "len_sents": 27441, "len_sents_1": 26944, "n_sents": 177, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1209, "title": "A VARIATIONAL FRAMEWORK FOR GRAPH GENERATION WITH FINE-GRAINED TOPOLOGICAL CONTROL", "abs": "Controlled graph generation is the process of generating graphs that satisfy specific topological properties (or attributes). Fine-grained control over graph properties allows for customizing generated graphs to precise specifications, which is essential for understanding and modeling complex networks. Existing approaches can only satisfy a few topological properties such as number of nodes or edges in output graphs. This paper introduces CGRAPHGEN, a novel conditional variational autoencoder that, unlike existing approaches, uses graph adjacency matrix during training, along with the desired graph properties, for improved decoder tuning and precise graph generation, while relying only on attributes during inference. In addition, CGRAPHGEN implements an effective scheduling technique to integrate representations from both adjacency matrix and attribute distributions for precise control. Experiments on five real-world datasets show the efficacy of CGRAPHGEN compared to baselines, which we attribute to its use of adjacency matrix during training and effective integration of representations, which aligns graphs and their attributes in the latent space effectively and results in better control.", "keywords": [ "Controlled Graph Generation" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 2, 4, 5 ] }, { "paper_id": "xXTkbTBmqq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 24, "author_info_list": [ { "name": "Akshita Bhagia", "gender": "Female", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Alexander Wettig", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Ali Farhadi", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Full Professor" }, { "name": "Amanpreet Singh", "gender": "Male", "institution": "Hugging Face", "country": "", "position": "Researcher" }, { "name": "Binyuan Hui", "gender": "Female", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "David Wadden", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Dirk Groeneveld", "gender": "unknown", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Principal Researcher" }, { "name": "Douwe Kiela", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Adjunct Professor" }, { "name": "Dustin Schwenk", "gender": "unknown", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Evan Pete Walsh", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Hannaneh Hajishirzi", "gender": "Female", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "senior director" }, { "name": "Kyle Lo", "gender": "unknown", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Luca Soldaini", "gender": "Non-Binary", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Nathan Lambert", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Niklas Muennighoff", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Oyvind Tafjord", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Pang Wei Koh", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Visiting Research Scientist" }, { "name": "Sewon Min", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "Assistant Professor" }, { "name": "Shane Arora", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "MS student" }, { "name": "Tim Dettmers", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Weijia Shi", "gender": "unknown", "institution": "University of Washington, Seattle", "country": "US", "position": "PhD student" }, { "name": "Yuling Gu", "gender": "unknown", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Predoctoral Young Investigator" }, { "name": "Jacob Daniel Morrison", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Noah Smith", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 3, "n_ref_uni": 147, "n_ref": 341, "n_ref_all": 435, "n_fig": 35, "n_tab": 8, "L_tab": 4099, "n_element_tab": 497, "n_fig_1": 12, "n_tab_1": 3, "L_tab_1": 1559, "n_element_tab_1": 75, "formula_len_all": 871, "formula_len_all_1": 256, "len_all": 633411, "len_all_1": 59951, "len_abs": 1509, "len_title": 46, "len_sents": 85786, "len_sents_1": 27115, "n_sents": 686, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 46, "L_abs": 621, "title": "OLMoE: Open Mixture-of-Experts Language Models", "abs": "We introduce OLMoE, a fully open, state-of-the-art language model leveraging sparse Mixture-of-Experts (MoE). OLMoE-1B-7B has 7 billion (B) parameters but uses only 1B per input token. We pretrain it on 5 trillion tokens and further adapt it to create OLMoE-1B-7B-Instruct. Our models outperform all available models with similar active parameters, even surpassing larger ones like Llama2-13B-Chat and DeepSeekMoE-16B. We present novel findings on MoE training, define and analyze new routing properties showing high specialization in our model, and open-source all our work: model weights, training data, code, and logs.", "keywords": [ "large language models", "mixture-of-experts", "open-source" ], "rating_list": [ 10, 8, 8 ], "soundness_list": [ 4, 4, 4 ], "presentation_list": [ 4, 4, 4 ], "contribution_list": [ 4, 3, 4 ], "confidence_list": [ 5, 2, 3 ] }, { "paper_id": "xW4J2QlqRx", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Pulkit Paliwal", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "MS student" }, { "name": "Sai Shankar Narasimhan", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Sameep Chattopadhyay", "gender": "Male", "institution": "Indian Institute of Technology, Bombay, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "Undergrad student" }, { "name": "Shubhankar Agarwal", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Sandeep Chinchali", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 29, "n_ref": 59, "n_ref_all": 81, "n_fig": 5, "n_tab": 12, "L_tab": 6520, "n_element_tab": 447, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 3848, "n_element_tab_1": 266, "formula_len_all": 504, "formula_len_all_1": 504, "len_all": 177045, "len_all_1": 72451, "len_abs": 1073, "len_title": 123, "len_sents": 60918, "len_sents_1": 30669, "n_sents": 417, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1076, "title": "Context Matters: Leveraging Contextual Features for Time Series Forecasting", "abs": "Time series forecasts are often influenced by exogenous contextual features in addition to their corresponding history. For example, in financial settings, it is hard to accurately predict a stock price without considering public sentiments and policy decisions in the form of news articles, tweets, etc. Though this is common knowledge, the current state-of-the-art (SOTA) forecasting models fail to incorporate such contextual information, owing to its heterogeneity and multimodal nature. To address this, we introduce ContextFormer, a novel plug-and-play method to surgically integrate multimodal contextual information into existing pre-trained forecasting models. ContextFormer effectively distills forecast-specific information from rich multimodal contexts, including categorical, continuous, time-varying, and even textual information, to significantly enhance the performance of existing base forecasters. ContextFormer outperforms SOTA forecasting models by up to 30% on a range of real-world datasets spanning energy, traffic, environmental, and financial domains.", "keywords": [ "Time series forecasting", "Contextual features", "Predictive modeling" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "xVw8YNEtH3", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "He Li", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Postdoc" }, { "name": "Jiajun Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Jingui Zou", "gender": "Male", "institution": "Wuhan University", "country": "", "position": "Full Professor" }, { "name": "Weiping Liu", "gender": "Female", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Youfa Liu", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 39, "n_formula_1": 22, "n_ref_uni": 35, "n_ref": 88, "n_ref_all": 107, "n_fig": 1, "n_tab": 11, "L_tab": 13460, "n_element_tab": 876, "n_fig_1": 1, "n_tab_1": 8, "L_tab_1": 7812, "n_element_tab_1": 442, "formula_len_all": 4333, "formula_len_all_1": 2237, "len_all": 161976, "len_all_1": 60113, "len_abs": 974, "len_title": 123, "len_sents": 34375, "len_sents_1": 19331, "n_sents": 238, "n_sents_1": 129, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 622, "title": "Reset Method based on the Theory of Manifold Optimization on Real Manifolds", "abs": "Manifold optimization is prominent in the fields of applied mathematics, statistics, machine learning, and in particular, deep learning. By leveraging the intrinsic geometric properties of manifolds, constrained optimization problems can be transformed into unconstrained optimization problems on certain manifolds. An innovative method, Reset Method, is introduced that combines manifold optimization and standard methods (SGD, Adam and AdamW), aiming to enhance the improvement of precision. The efficacy of our proposed method is corroborated by extensive deep learning experiments, providing visible higher precision.", "keywords": [ "Manifold Optimization", "Real Manifolds", "Method", "Deep Learning." ], "rating_list": [ 3, 5, 1 ], "soundness_list": [ 2, 3, 1 ], "presentation_list": [ 1, 2, 1 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "xVefsBbG2O", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Benedikt Hartl", "gender": "unknown", "institution": "Tufts University", "country": "", "position": "Postdoc" }, { "name": "Hananel Hazan", "gender": "Male", "institution": "Tufts University", "country": "US", "position": "Principal Researcher" }, { "name": "Michael Levin", "gender": "unknown", "institution": "Tufts University", "country": "US", "position": "Full Professor" }, { "name": "Yanbo Zhang", "gender": "Male", "institution": "Tufts University", "country": "US", "position": "Postdoc" } ], "n_formula": 43, "n_formula_1": 21, "n_ref_uni": 53, "n_ref": 87, "n_ref_all": 123, "n_fig": 7, "n_tab": 5, "L_tab": 3915, "n_element_tab": 443, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 67, "n_element_tab_1": 2, "formula_len_all": 2265, "formula_len_all_1": 1392, "len_all": 157165, "len_all_1": 61444, "len_abs": 1363, "len_title": 93, "len_sents": 49503, "len_sents_1": 29926, "n_sents": 354, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 44, "L_abs": 1370, "title": "Diffusion Models are Evolutionary Algorithms", "abs": "In a convergence of machine learning and biology, we reveal that diffusion models are evolutionary algorithms. By considering evolution as a denoising process and reversed evolution as diffusion, we mathematically demonstrate that diffusion models inherently perform evolutionary algorithms, naturally encompassing selection, mutation, and reproductive isolation. Building on this equivalence, we propose the Diffusion Evolution method: an evolutionary algorithm utilizing iterative denoising -- as originally introduced in the context of diffusion models -- to heuristically refine solutions in parameter spaces. Unlike traditional approaches, Diffusion Evolution efficiently identifies multiple optimal solutions and outperforms prominent mainstream evolutionary algorithms. Furthermore, leveraging advanced concepts from diffusion models, namely latent space diffusion and accelerated sampling, we introduce Latent Space Diffusion Evolution, which finds solutions for evolutionary tasks in high-dimensional complex parameter space while significantly reducing computational steps. This parallel between diffusion and evolution not only bridges two different fields but also opens new avenues for mutual enhancement, raising questions about open-ended evolution and potentially utilizing non-Gaussian or discrete diffusion models in the context of Diffusion Evolution.", "keywords": [ "Machine learning", "evolutionary computation", "Evolutionary Algorithms", "Diffusion Models", "Optimization" ], "rating_list": [ 8, 6, 5, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "xVU6rY37X9", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Kibok Lee", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Seunghan Lee", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "PhD student" }, { "name": "Taeyoung Park", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 33, "n_ref": 80, "n_ref_all": 120, "n_fig": 10, "n_tab": 21, "L_tab": 16362, "n_element_tab": 1105, "n_fig_1": 8, "n_tab_1": 12, "L_tab_1": 6643, "n_element_tab_1": 642, "formula_len_all": 607, "formula_len_all_1": 319, "len_all": 161737, "len_all_1": 70258, "len_abs": 1415, "len_title": 127, "len_sents": 37485, "len_sents_1": 25019, "n_sents": 263, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1242, "title": "Partial Channel Dependence with Channel Masks for Time Series Foundation Models", "abs": "Recent advancements in foundation models have been successfully extended to the time series (TS) domain, facilitated by the emergence of large-scale TS datasets. However, previous efforts have primarily focused on designing model architectures to address explicit heterogeneity among datasets such as various numbers of channels, while often overlooking implicit heterogeneity such as varying dependencies between channels. In this work, we introduce the concept of partial channel dependence (PCD) for models capturing channel dependencies (CDs) via attention, which enables a more sophisticated adjustment of CDs based on dataset-specific information. To achieve PCD, we propose a channel mask that captures the relationships between channels within a dataset using two key components: 1) a correlation matrix that encodes relative dependencies between channels, and 2) domain parameters that learn the absolute dependencies specific to each dataset, refining the correlation matrix. We validate the effectiveness of PCD across four tasks in TS including forecasting, classification, imputation, and anomaly detection, under diverse settings, including few-shot and zero-shot scenarios with both TS foundation models and single-task models.", "keywords": [ "Time Series", "Foundation Model", "Channel Dependence", "Transformer" ], "rating_list": [ 5, 3, 6, 6, 3 ], "soundness_list": [ 2, 1, 3, 2, 1 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 2, 1, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 3, 3 ] }, { "paper_id": "xVOMtecrAS", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Boxiao Liu", "gender": "Male", "institution": "Sensetime Research", "country": "CN", "position": "Researcher" }, { "name": "Guanglu Song", "gender": "Male", "institution": "Sensetime", "country": "CN", "position": "Computer Vision Researcher" }, { "name": "Haihang You", "gender": "unknown", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xingzhong Hou", "gender": "Male", "institution": "State Key Laboratory of Computer Architecture, Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yi Zhang", "gender": "Male", "institution": "SenseTime", "country": "HK", "position": "Researcher" }, { "name": "Yu Liu", "gender": "Male", "institution": "SenseTime", "country": "HK", "position": "Principal Researcher" }, { "name": "Yunpeng Liu", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 19, "n_ref_uni": 15, "n_ref": 25, "n_ref_all": 40, "n_fig": 6, "n_tab": 4, "L_tab": 836, "n_element_tab": 57, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1962, "n_element_tab_1": 174, "formula_len_all": 1152, "formula_len_all_1": 1151, "len_all": 84363, "len_all_1": 52262, "len_abs": 1091, "len_title": 135, "len_sents": 22942, "len_sents_1": 20760, "n_sents": 178, "n_sents_1": 164, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1095, "title": "See Further When Clear: Adaptive Generative Modeling with Curriculum Consistency Model", "abs": "Significant advances have been made in the sampling efficiency of diffusion models, driven by Consistency Distillation (CD), which trains a student model to mimic the output of a teacher model at an earlier timestep. However, we found that the learning complexity of the student model varies significantly across different timesteps, leading to suboptimal performance in consistency models.\nTo address this issue, we propose the Curriculum Consistency Model (CCM), which stabilizes and balances the learning complexity across timesteps. We define the distillation process as a curriculum and introduce Peak Signal-to-Noise Ratio (PSNR) as a metric to quantify the difficulty of each step in this curriculum.\nBy incorporating adversarial losses, our method achieves competitive single-step sampling Fr\u00e9chet Inception Distance (FID) scores of 1.64 on CIFAR-10 and 2.18 on ImageNet 64x64.\nMoreover, our approach generalizes well to both Flow Matching models and diffusion models. We have extended our method to large-scale text-to-image models, including Stable Diffusion XL and Stable Diffusion 3.", "keywords": [ "adaptive curriculum learning", "noise schedule", "flow matching", "consistency models" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 4, 2 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "xUMI52rrW7", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Guozu Ma", "gender": "unknown", "institution": "China Telecom WanWei Information Technology Co.,Ltd", "country": "CN", "position": "Principal Researcher" }, { "name": "Jiangning Zhu", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Minzhi Lin", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Shixia Liu", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Tianchi Xie", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Wei Chen", "gender": "Male", "institution": "Microsoft Research", "country": "", "position": "Pricipal Researcher" }, { "name": "Weikai Yang", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 32, "n_formula_1": 8, "n_ref_uni": 64, "n_ref": 122, "n_ref_all": 154, "n_fig": 4, "n_tab": 19, "L_tab": 13016, "n_element_tab": 922, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 6721, "n_element_tab_1": 247, "formula_len_all": 3192, "formula_len_all_1": 409, "len_all": 233961, "len_all_1": 72313, "len_abs": 1463, "len_title": 78, "len_sents": 61462, "len_sents_1": 29449, "n_sents": 539, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1522, "title": "Structural-Entropy-Based Sample Selection for Efficient and Effective Learning", "abs": "Sample selection improves the efficiency and effectiveness of machine learning models by providing informative and representative samples. Typically, samples can be modeled as a sample graph, where nodes are samples and edges represent their similarities. Most existing methods are based on local information, such as the training difficulty of samples, thereby overlooking global information, such as connectivity patterns. This oversight can result in suboptimal selection because global information is crucial for ensuring that the selected samples well represent the structural properties of the graph. To address this issue, we employ structural entropy to quantify global information and losslessly decompose it from the whole graph to individual nodes using the Shapley value. Based on the decomposition, we present $\\textbf{S}$tructural-$\\textbf{E}$ntropy-based sample $\\textbf{S}$election ($\\textbf{SES}$), a method that integrates both global and local information to select informative and representative samples. SES begins by constructing a $k$NN-graph among samples based on their similarities. It then measures sample importance by combining structural entropy (global metric) with training difficulty (local metric). Finally, SES applies importance-biased blue noise sampling to select a set of diverse and representative samples. Comprehensive experiments on three learning scenarios --- supervised learning, active learning, and continual learning --- clearly demonstrate the effectiveness of our method.", "keywords": [ "Sample selection", "graph", "structural entropy", "blue noise sampling" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "xUHL8mtSUL", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Didong Li", "gender": "unknown", "institution": "University of North Carolina at Chapel Hill", "country": "", "position": "Assistant Professor" }, { "name": "Geoffery Wang", "gender": "unknown", "institution": "North Carolina State University", "country": "US", "position": "Undergrad student" }, { "name": "Kevin Wang", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 20, "n_ref_uni": 19, "n_ref": 25, "n_ref_all": 37, "n_fig": 3, "n_tab": 5, "L_tab": 2341, "n_element_tab": 320, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1395, "n_element_tab_1": 191, "formula_len_all": 3587, "formula_len_all_1": 2518, "len_all": 97881, "len_all_1": 62339, "len_abs": 1047, "len_title": 122, "len_sents": 32557, "len_sents_1": 25670, "n_sents": 243, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1059, "title": "Scalable Gaussian Process via Hilbert-Schmidt Singular Value Decomposition", "abs": "Gaussian process regression is widely used for its flexible mean predictions and inherent uncertainty quantification. However, its scalability is limited by cubic time complexity, $O(n^3)$, and quadratic space complexity, $O(n^2)$, making it infeasible for large-scale datasets. Although recent advances have introduced approximate methods with time complexity $O(nm^2)$, where $m\\ll n$ is a tuning parameter, these methods each have their own bottlenecks, such as requiring a relatively large $m$ or involving expensive preprocessing steps. Moreover, for extremely large datasets with millions of samples, the space complexity $O(n^2)$ becomes another significant bottleneck. In this paper, we present a novel method based on the Hilbert-Schmidt singular value decomposition that obtains a low-rank decomposition ``for free\", reducing both time complexity to $O(nm^2)$ and space complexity to $O(nm)$, with no preprocessing overhead. We used simulated large-scale datasets to demonstrate the performance of our method compared to state-of-the-art approaches.", "keywords": [ "Scalability", "Gaussian process regression", "Hilbert Schmidt singular value decomposition", "compact Mat\\'ern" ], "rating_list": [ 5, 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 5, 3, 4 ] }, { "paper_id": "xTsvE8gOPT", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Jiayi Ji", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Mingrui Wu", "gender": "Not Specified", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Oucheng Huang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Ruohan Dong", "gender": "Not Specified", "institution": "Xiamen University", "country": "CN", "position": "Undergrad student" }, { "name": "Xiaoshuai Sun", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "JialeLi", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 13, "n_ref": 30, "n_ref_all": 68, "n_fig": 19, "n_tab": 7, "L_tab": 1418, "n_element_tab": 119, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 779, "n_element_tab_1": 74, "formula_len_all": 558, "formula_len_all_1": 517, "len_all": 119123, "len_all_1": 49799, "len_abs": 0, "len_title": 0, "len_sents": 34576, "len_sents_1": 22972, "n_sents": 274, "n_sents_1": 169, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1256, "title": "Tradiffusion++\uff1aHierarchical Guidance for Fine-Grained Trajectory-Based Image Generation", "abs": "Currently, many training-free methods based on diffusion models allow controllable generation. These methods, such as TraDiffusion, introduce control through additional trajectory input. While they are more user-friendly than traditional methods, they offer only coarse control over the Stable Diffusion (SD) model. We observe that SD focuses more on layout control at lower resolutions of cross-attention and shape control at higher ones. Based on this, we propose TraDiffusion++, which introduces a Hierarchical Guidance Mechanism (HGM) for finer-grained control in generation. HGM includes three key components: Control Loss (CL), Suppress Loss (SL), and Fix Loss (FL). CL aligns the layout with the trajectory across layers. SL suppresses objects outside the trajectory at lower resolutions. FL refines regions not fully controlled by the trajectory using attention feedback at middle and high resolutions. The combination of CL and SL ensures effective layout control. The interaction between CL and FL improves shape generation. We build a dataset with simple and complex trajectories. Experiments show that TraDiffusion++ achieves stable layout control and fine-grained object generation. This also reveals new insights into SD\u2019s control mechanisms.", "keywords": [ "Diffusion models; Trajectory control; TraDiffusion++; Training-free methods; Controllable generation; Stable Diffusion (SD); Fine-Grained Control" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "xTrAA3UKPa", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Chenwei Xu", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" }, { "name": "Weijian Li", "gender": "Male", "institution": "Northwestern University", "country": "", "position": "PhD student" }, { "name": "Haozheng LUO", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 34, "n_ref": 38, "n_ref_all": 52, "n_fig": 5, "n_tab": 4, "L_tab": 3328, "n_element_tab": 326, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 3602, "n_element_tab_1": 362, "formula_len_all": 290, "formula_len_all_1": 290, "len_all": 107137, "len_all_1": 57450, "len_abs": 1020, "len_title": 82, "len_sents": 28508, "len_sents_1": 22765, "n_sents": 212, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1024, "title": "SWGA: A Distributed Hyperparameter Search Method for Time Series Prediction Models", "abs": "We propose a distributed hyperparameter search method for time series prediction models named SWGA (Sliding Window Genetic Algorithm). Compared to current genetic algorithms for hyperparameter search, our method has three major advantages: (i) It adopts a configurable sliding window mechanism to effectively combat overfitting from distribution shifts inherent in time series data. (ii) It introduces a warm-up stage using Bayesian optimization-based methods to generate a good initial population. (iii) It supports distributed hyperparameter search across multi-node computing clusters, enhancing both scalability and efficiency. To demonstrate SWGA's efficacy, we conduct hyperparameter search experiments on time series datasets from various domains. The experiment results show that our method consistently finds a hyperparameter configuration that achieves better performance on out-of-sample time series data compared to the traditional genetic algorithm. On average, it reduces the out-of-sample loss by about 56.1%.", "keywords": [ "Machine Learning", "Deep Learning", "Time Series Prediction", "Hyperparameter Search", "Genetic Algorithms" ], "rating_list": [ 1, 1, 3, 3 ], "soundness_list": [ 1, 1, 2, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 1, 1, 1, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "xSSo8kCA9G", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Aleksandr Beznosikov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "Associate Professor" }, { "name": "Philip Zmushko", "gender": "unknown", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "MS student" }, { "name": "Samuel Horv\u00e1th", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Assistant Professor" }, { "name": "Martin Takac", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Associate Professor" } ], "n_formula": 80, "n_formula_1": 11, "n_ref_uni": 44, "n_ref": 120, "n_ref_all": 161, "n_fig": 3, "n_tab": 14, "L_tab": 4540, "n_element_tab": 404, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1175, "n_element_tab_1": 85, "formula_len_all": 10063, "formula_len_all_1": 640, "len_all": 214018, "len_all_1": 69553, "len_abs": 1580, "len_title": 87, "len_sents": 60721, "len_sents_1": 30564, "n_sents": 565, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1597, "title": "FRUGAL: Memory-Efficient Optimization by Reducing State Overhead for Scalable Training", "abs": "With the increase in the number of parameters in large language models, the process of pre-training and fine-tuning increasingly demands larger volumes of GPU memory. A significant portion of this memory is typically consumed by the optimizer state. To overcome this challenge, recent approaches such as low-rank adaptation (LoRA (Hu et al., 2021)), low-rank gradient projection (GaLore (Zhao\net al., 2024)), and block-wise optimization (BAdam (Luo et al., 2024)) have been proposed. However, in all these algorithms, the effective rank of the weight updates remains low-rank, which can lead to a substantial loss of information from the gradient. This loss can be critically important, especially during the pre-training stage. In this paper, we introduce **FRUGAL**; (**F**ull-**R**ank **U**pdates with **G**r**A**dient sp**L**itting),, a new memory-efficient optimization framework. The framework leverages gradient splitting to perform low-rank updates using advanced optimization algorithms (such as Adam), while updates along the remaining directions are\nexecuted via state-free methods like SGD or signSGD. Our framework can be integrated with various low-rank update selection techniques, including GaLore and BAdam. We provide theoretical convergence guarantees for our framework when\nusing SGDM for low-rank updates and SGD for state-free updates. Additionally, our method consistently outperforms concurrent approaches across various fixed memory budgets, achieving state-of-the-art results in pre-training and fine-tuning\ntasks while balancing memory efficiency and perplexity targets.", "keywords": [ "Memory-efficient training", "Optimization", "Full-Rank Update", "Large Language Models", "LLM", "Pre-training", "Fine-tuning" ], "rating_list": [ 5, 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 1, 2 ], "contribution_list": [ 2, 2, 4, 3, 3 ], "confidence_list": [ 3, 4, 2, 3, 4 ] }, { "paper_id": "xSOl0s1u77", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Jiachen Li", "gender": "Male", "institution": "xAI", "country": "US", "position": "Researcher" }, { "name": "Michael Saxon", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Tsu-Jui Fu", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Weixi Feng", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "wenhu chen", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Assistant Professor" }, { "name": "William Yang Wang", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 49, "n_ref": 117, "n_ref_all": 161, "n_fig": 16, "n_tab": 6, "L_tab": 2444, "n_element_tab": 362, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1750, "n_element_tab_1": 247, "formula_len_all": 1004, "formula_len_all_1": 757, "len_all": 197513, "len_all_1": 57789, "len_abs": 2011, "len_title": 129, "len_sents": 50940, "len_sents_1": 25234, "n_sents": 421, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1607, "title": "TC-Bench: Benchmarking Temporal Compositionality in Conditional Video Generation", "abs": "Video generation has many unique challenges beyond those of image generation. The temporal dimension introduces extensive possible variations across frames, over which consistency and continuity may be violated. In this study, we move beyond evaluating simple actions and argue that generated videos should incorporate the emergence of new concepts and their relation transitions like in real-world videos as time progresses. To assess the \\textbf{T}emporal \\textbf{C}ompositionality of video generation models, we propose TC-Bench, a benchmark of meticulously crafted text prompts, corresponding ground truth videos, and robust evaluation metrics. The prompts articulate the initial and final states of scenes, effectively reducing ambiguities for frame development and simplifying the assessment of transition completion. In addition, by collecting aligned real-world videos corresponding to the prompts, we expand TC-Bench's applicability from text-conditional models to image-conditional ones that can perform generative frame interpolation. We also develop new metrics to measure the completeness of component transitions in generated videos, which demonstrate significantly higher correlations with human judgments than existing metrics. Our comprehensive experimental results reveal that most video generators achieve less than \uff5e20% of the compositional changes, highlighting enormous space for future improvement. Our analysis indicates that current video generation models struggle to interpret descriptions of compositional changes and dynamically map varied semantics across different time steps.", "keywords": [ "Video Generation Benchmark; Text-to-Video Generation; Compositional Video Generation" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "xS6uKkJ9Uz", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Fuling WANG", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Hui Xiong", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Nan Tang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "Associate Professor" }, { "name": "Yin WU", "gender": "Male", "institution": "The Hong Kong University of Science and Technology(Guangzhou))", "country": "CN", "position": "MS student" }, { "name": "Yuyu Luo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhengxuan Zhang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 25, "n_ref": 49, "n_ref_all": 62, "n_fig": 4, "n_tab": 14, "L_tab": 2227, "n_element_tab": 256, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1986, "n_element_tab_1": 229, "formula_len_all": 321, "formula_len_all_1": 321, "len_all": 149223, "len_all_1": 55378, "len_abs": 1315, "len_title": 131, "len_sents": 52028, "len_sents_1": 25130, "n_sents": 342, "n_sents_1": 157, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1327, "title": "Detecting Out-of-Context Misinformation via Multi-Agent and Multi-Grained Retrieval", "abs": "Misinformation remains a critical issue in today's information landscape, significantly impacting public perception and behavior. Among its various forms, out-of-context (OOC) misinformation is particularly pervasive, misrepresenting information by repurposing authentic images with false text. Traditional OOC detection methods often rely on coarse-grained similarity measures between image-text pairs, which fall short of providing interpretability and nuanced understanding. Conversely, whereas multimodal large language models (MLLMs) exhibit vast knowledge and an inherent ability for visual reasoning and explanation generation, they remain deficient in the complexity required to understand and discern nuanced cross-modal distinctions thoroughly. To address these challenges, we propose MACAW, a retrieval-based approach that indexes external knowledge, focusing on multiple granularities by extracting and cataloging relevant events and entities. Our framework first extracts multi-granularity information to assess the contextual integrity of news items, followed by a multi-agent reasoning process for accurate detection. Extensive experiments demonstrate the robustness and effectiveness of our proposed framework in identifying out-of-context fake news, outperforming the state-of-the-art solutions by {\\bf 4.3\\%}.", "keywords": [ "Multimodal Machine learning", "Multi-modal Large Language Model" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "xS4XOS4NQ5", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Ge Zhang", "gender": "unknown", "institution": "Shanghai Qi Zhi Institute", "country": "CN", "position": "Researcher" }, { "name": "Kangping Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Yifan Zhang", "gender": "Not Specified", "institution": "University of California, Los Angeles", "country": "US", "position": "Researcher" }, { "name": "Yue Wu", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Postdoc" } ], "n_formula": 73, "n_formula_1": 29, "n_ref_uni": 46, "n_ref": 102, "n_ref_all": 125, "n_fig": 0, "n_tab": 10, "L_tab": 5272, "n_element_tab": 783, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 1601, "n_element_tab_1": 210, "formula_len_all": 4125, "formula_len_all_1": 1568, "len_all": 224976, "len_all_1": 66941, "len_abs": 1609, "len_title": 137, "len_sents": 62741, "len_sents_1": 28056, "n_sents": 558, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1608, "title": "General Preference Modeling with Preference Representations for Aligning Language Models", "abs": "Modeling human preferences is crucial for aligning foundation models with human values. Traditional reward modeling methods, such as the Bradley-Terry (BT) reward model, fall short in expressiveness, particularly in addressing intransitive preferences. Although supervised pair preference models and pair reward models can express general preferences, their implementation is highly ad-hoc and cannot guarantee a consistent preference probability of compared pairs. Additionally, they impose high computational costs due to their quadratic query complexity when comparing multiple responses. In this paper, we introduce preference representation learning, an approach that embeds responses into a latent space to capture intricate preference structures efficiently, achieving linear query complexity (matching the efficiency of the BT model). Additionally, we propose preference score-based General Preference Optimization (GPO), which generalizes reward-based reinforcement learning from human feedback. Experimental results show that our General Preference representation model (GPM) outperforms the BT reward model on the RewardBench benchmark with a margin of up to 9.1% and effectively models cyclic preferences where any BT reward model behaves like a random guess. Furthermore, evaluations on downstream tasks such as AlpacaEval2.0, following the language model post-training with GPO and our general preference model, reveal substantial performance improvements with margins up to 8.3%. These findings indicate that our method may enhance the alignment of foundation models with nuanced human values.", "keywords": [ "preference modeling", "preference optimization", "reinforcement learning from human feedback" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "xRi8sKo4XI", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Gang Niu", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Senior Research Scientist (tenured)" }, { "name": "Huaxiu Yao", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Jiandong Zhang", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "MS student" }, { "name": "Masashi Sugiyama", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Director" }, { "name": "Zhen-Yu Zhang", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Postdoc" } ], "n_formula": 20, "n_formula_1": 20, "n_ref_uni": 30, "n_ref": 57, "n_ref_all": 71, "n_fig": 4, "n_tab": 6, "L_tab": 1507, "n_element_tab": 202, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1281, "n_element_tab_1": 185, "formula_len_all": 1412, "formula_len_all_1": 1393, "len_all": 120238, "len_all_1": 73684, "len_abs": 1836, "len_title": 130, "len_sents": 35885, "len_sents_1": 33772, "n_sents": 273, "n_sents_1": 251, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1843, "title": "On Unsupervised Prompt Learning for Classification with Black-box Language Models", "abs": "Large language models (LLMs) have achieved impressive success in text-formatted learning problems, and most popular LLMs have been deployed in a black-box fashion. Meanwhile, fine-tuning is usually necessary for a specific downstream task to obtain better performance, and this functionality is provided by the owners of the black-box LLMs. To fine-tune a black-box LLM, labeled data are always required to adjust the model parameters. However, in many real-world applications, LLMs can label textual datasets with even better quality than skilled human annotators, motivating us to explore the possibility of fine-tuning black-box LLMs with unlabeled data. In this paper, we propose unsupervised prompt learning for classification with black-box LLMs, where the learning parameters are the prompt itself and the pseudo labels of unlabeled data. Specifically, the prompt is modeled as a sequence of discrete tokens, and every token has its own to-be-learned categorical distribution. On the other hand, for learning the pseudo labels, we are the first to consider the in-context learning (ICL) capabilities of LLMs: we first identify reliable pseudo-labeled data using the LLM, and then assign pseudo labels to other unlabeled data based on the prompt, allowing the pseudo-labeled data to serve as in-context demonstrations alongside the prompt. Those in-context demonstrations matter: previously, they are involved when the prompt is used for prediction while they are not involved when the prompt is trained; thus, taking them into account during training makes the prompt-learning and prompt-using stages more consistent. Experiments on benchmark datasets show the effectiveness of our proposed algorithm. After unsupervised prompt learning, we can use the pseudo-labeled dataset for further fine-tuning by the owners of the black-box LLMs.", "keywords": [ "Prompt Learning", "Black-box Language Models", "In-context Learning" ], "rating_list": [ 3, 1, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 2, 1, 3, 2, 2 ], "contribution_list": [ 1, 1, 2, 2, 2 ], "confidence_list": [ 3, 5, 3, 4, 4 ] }, { "paper_id": "xRDYDI6Rc9", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Anca Dragan", "gender": "Female", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Cassidy Laidlaw", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Shivam Singhal", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 53, "n_ref": 96, "n_ref_all": 112, "n_fig": 7, "n_tab": 2, "L_tab": 867, "n_element_tab": 45, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1768, "n_element_tab_1": 96, "formula_len_all": 590, "formula_len_all_1": 428, "len_all": 289175, "len_all_1": 72547, "len_abs": 1540, "len_title": 107, "len_sents": 67742, "len_sents_1": 36614, "n_sents": 433, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1546, "title": "Reliability-Aware Preference Learning for LLM Reward Models", "abs": "Reward functions learned from human feedback serve as the training objective for RLHF, the current state-of-the-art approach for aligning large language models to our values. However, in practice, these reward models fail to robustly capture our desiderata, often attributing more value to features such as output length or agreement with the user and less value to important features like factual correctness. A major reason is that human annotators provide feedback that is an unreliable reflection of their true preferences because of knowledge gaps, limited resources, cognitive biases, or other factors. We focus on making preference learning robust to unreliable feedback by explicitly modeling the knowledge and judgment of annotators. In particular, we estimate reliablity scores for each provided pairwise comparison and incoporate them into the implicit human model used in RLHF, DPO, and other alignment techniques, a technique we call Reliability Aware Preference Learning (RAPL). To test our approach, we introduce the Length Incentivized Evaluations dataset as a setting in which annotators are particularly likely to provide unreliable feedback. Then, we curate the Testing Reasoning and Understanding Errors dataset for training models to predict reliability scores. We find that traditional preference learning on the LIE dataset and other commonly used RLHF datasets leads to models that place far more weight on output length than accuracy. In contrast, RAPL results in models that better capture the true values of annotators.", "keywords": [ "preference learning", "RLHF", "human models", "scalable oversight" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "xQit6JBDR5", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Berker Demirel", "gender": "Male", "institution": "Institute of Science and Technology", "country": "AT", "position": "PhD student" }, { "name": "Francesco Locatello", "gender": "Male", "institution": "Institute of Science and Technology", "country": "AT", "position": "Assistant Professor" }, { "name": "Marco Fumero", "gender": "unknown", "institution": "Institute of Science and Technology Austria(ISTA)", "country": "AT", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 36, "n_ref": 74, "n_ref_all": 100, "n_fig": 5, "n_tab": 10, "L_tab": 3383, "n_element_tab": 578, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2081, "n_element_tab_1": 367, "formula_len_all": 924, "formula_len_all_1": 923, "len_all": 133869, "len_all_1": 60140, "len_abs": 1151, "len_title": 108, "len_sents": 33840, "len_sents_1": 26044, "n_sents": 253, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1159, "title": "Look Around and Find Out: OOD Detection with Relative Angles", "abs": "Deep learning systems deployed in real-world applications often encounter data that is different from their in-distribution (ID). A reliable system should ideally abstain from making decisions in this out-of-distribution (OOD) setting. Existing state-of-the-art methods primarily focus on feature distances, such as k-th nearest neighbors and distances to decision boundaries, either overlooking or ineffectively using in-distribution statistics. In this work, we propose a novel angle-based metric for OOD detection that is computed relative to the in-distribution structure. We demonstrate that the angles between feature representations and decision boundaries, viewed from the mean of in-distribution features, serve as an effective discriminative factor between ID and OOD data. Our method achieves state-of-the-art performance on CIFAR-10 and ImageNet benchmarks, reducing FPR95 by 0.88% and 7.74% respectively. Our scoring function is compatible with existing feature space regularization techniques, enhancing performance. Additionally, its scale-invariance property enables creating an ensemble of models for OOD detection via simple score summation.", "keywords": [ "out-of-distribution", "out-of-distribution detection", "decision boundaries" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "xQVxo9dSID", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Ashwini Pokle", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Justin Lin", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Undergrad student" }, { "name": "William Luo", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhengyang Geng", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zico Kolter", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 26, "n_formula_1": 20, "n_ref_uni": 76, "n_ref": 203, "n_ref_all": 225, "n_fig": 8, "n_tab": 6, "L_tab": 1490, "n_element_tab": 254, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1776, "n_element_tab_1": 118, "formula_len_all": 1664, "formula_len_all_1": 1008, "len_all": 215299, "len_all_1": 60317, "len_abs": 1399, "len_title": 76, "len_sents": 58777, "len_sents_1": 26711, "n_sents": 437, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 28, "L_abs": 1332, "title": "Consistency Models Made Easy", "abs": "Consistency models (CMs) offer faster sampling than traditional diffusion models, but their training is resource-intensive. For example, as of 2024, training a state-of-the-art CM on CIFAR-10 takes one week on 8 GPUs. In this work, we propose an effective scheme for training CMs that largely improves the efficiency of building such models. Specifically, by expressing CM trajectories via a particular differential equation, we argue that diffusion models can be viewed as a special case of CMs. We can thus fine-tune a consistency model starting from a pretrained diffusion model and progressively approximate the full consistency condition to stronger degrees over the training process. Our resulting method, which we term Easy Consistency Tuning (ECT), achieves vastly reduced training times while improving upon the quality of previous methods: for example, ECT achieves a 2-step FID of 2.73 on CIFAR10 within 1 hour on a single A100 GPU, matching Consistency Distillation trained for hundreds of GPU hours. Owing to this computational efficiency, we investigate the scaling laws of CMs under ECT, showing that they obey the classic power law scaling, hinting at their ability to improve efficiency and performance at larger scales. Our code will be made publicly available, making CMs more accessible to the broader community.", "keywords": [ "Consistency Models", "Efficient Generative Models", "Diffusion Models" ], "rating_list": [ 8, 5, 8, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "xQIJ5fjc7q", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Hai Yan", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Jiawen Shi", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Lin Lu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Pan Zhou", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Professor" }, { "name": "Pin-Yu Chen", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Wenqi Wei", "gender": "unknown", "institution": "Fordham University", "country": "US", "position": "Assistant Professor" }, { "name": "Zenghui Yuan", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 42, "n_ref": 95, "n_ref_all": 109, "n_fig": 2, "n_tab": 5, "L_tab": 2997, "n_element_tab": 162, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 189191, "len_all_1": 64261, "len_abs": 1569, "len_title": 145, "len_sents": 57167, "len_sents_1": 32264, "n_sents": 463, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1454, "title": "DAG-Jailbreak: Enhancing Black-box Jailbreak Attacks and Defenses through DAG Dependency Analysis", "abs": "Black-box jailbreak attacks and defenses, a critical branch of the large language model (LLM) security, are characterized by their minimal requirement for user expertise and high potential for automation. However, current black-box jailbreak approaches often adhere to a uniform global algorithmic framework, leading to suboptimal solutions due to challenges in local optimization. This limits both their effectiveness and scalability. To address these limitations, we propose **DAG-Jailbreak**, a novel framework leveraging Directed Acyclic Graph (DAG) dependency analysis to construct more robust jailbreak attacks and defenses. The core idea behind this framework is to combine optimal sub-components to form a more effective global algorithm. **DAG-Jailbreak** compromises three components: *DAG-Attack*, which creates highly effective attackers based on two global algorithms and is capable of compromising well-aligned LLMs without prior knowledge; *DAG-Defense*, which introduces a novel global framework based on a mixture-of-defenders mechanism, significantly enhancing the scalability and effectiveness of jailbreak defenses by reducing the attack success rate to below 3\\% in most cases; and *DAG-Evaluation*, which introduces the concept of jailbreak hallucination and a two-stage evaluation framework to assess the outputs generated by LLMs comprehensively. Extensive experiments validate the superiority and robustness of **DAG-Jailbreak**.", "keywords": [ "Jailbreak Attacks and Defenses", "LLM Security", "DAG Dependency Analysis" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 3, 3, 2, 1 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "xQCXInDq0m", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Anca Dragan", "gender": "Female", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Jerry Zhi-Yang He", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Mariah L Schrum", "gender": "Female", "institution": "Toyota Research Institute", "country": "US", "position": "Researcher" }, { "name": "Sashrika Pandey", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 6, "n_ref_uni": 20, "n_ref": 37, "n_ref_all": 73, "n_fig": 13, "n_tab": 11, "L_tab": 4699, "n_element_tab": 21, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 577, "formula_len_all_1": 357, "len_all": 223755, "len_all_1": 75069, "len_abs": 2206, "len_title": 113, "len_sents": 70873, "len_sents_1": 33438, "n_sents": 570, "n_sents_1": 283, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1615, "title": "Context Steering: Controllable Personalization at Inference Time", "abs": "To deliver high-quality, personalized responses, large language models (LLMs) must effectively incorporate context \u2014 personal, demographic, and cultural information specific to an end-user. For example, asking the model to explain Newton's second law with the context \"I am a toddler'' should produce a response different from when the context is \"I am a physics professor''. However, leveraging the context in practice is a nuanced and challenging task, and is often dependent on the specific situation or user base. The model must strike a balance between providing specific, personalized responses and maintaining general applicability. Current solutions, such as prompt-engineering and fine-tuning, require collection of contextually appropriate responses as examples, making them time-consuming and less flexible to use across different contexts. In this work, we introduce Context Steering (CoS) \u2014a simple, training-free decoding approach that amplifies the influence of the context in next token predictions. CoS computes contextual influence by comparing the output probabilities from two LLM forward passes: one that includes the context and one that does not. By linearly scaling the contextual influence, CoS allows practitioners to flexibly control the degree of personalization for different use cases. We show that CoS can be applied to autoregressive LLMs, and demonstrates strong performance in personalized recommendations. Additionally, we show that CoS can function as a Bayesian Generative model to infer and quantify correlations between open-ended texts, broadening its potential applications.", "keywords": [ "personalization", "context", "large language model", "inference", "controllable generation" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "xQBRrtQM8u", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Brian Karrer", "gender": "Male", "institution": "Meta Fundamental AI Research (FAIR)", "country": "US", "position": "Researcher" }, { "name": "Carles Domingo-Enrich", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Michal Drozdzal", "gender": "Male", "institution": "Meta", "country": "CA", "position": "Research Scientst" }, { "name": "Tian Qi Chen", "gender": "Male", "institution": "FAIR Labs, Meta AI", "country": "US", "position": "Researcher" } ], "n_formula": 259, "n_formula_1": 31, "n_ref_uni": 76, "n_ref": 181, "n_ref_all": 324, "n_fig": 5, "n_tab": 6, "L_tab": 6207, "n_element_tab": 436, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 244, "n_element_tab_1": 19, "formula_len_all": 31397, "formula_len_all_1": 2007, "len_all": 359884, "len_all_1": 64439, "len_abs": 1025, "len_title": 157, "len_sents": 87459, "len_sents_1": 27197, "n_sents": 823, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 109, "L_abs": 923, "title": "Adjoint Matching: Fine-tuning Flow and Diffusion Generative Models with Memoryless Stochastic Optimal Control", "abs": "Dynamical generative models that produce samples through an iterative process, such as Flow Matching and denoising diffusion models, have seen widespread use, but there have not been many theoretically-sound methods for improving these models with reward fine-tuning. In this work, we cast reward fine-tuning as stochastic optimal control (SOC). Critically, we prove that a very specific *memoryless* noise schedule must be enforced during fine-tuning, in order to account for the dependency between the noise variable and the generated samples. We also propose a new algorithm named *Adjoint Matching* which outperforms existing SOC algorithms, by casting SOC problems as a regression problem. We find that our approach significantly improves over existing methods for reward fine-tuning, achieving better consistency, realism, and generalization to unseen human preference reward models, while retaining sample diversity.", "keywords": [ "Reward fine-tuning", "stochastic optimal control", "flow matching", "diffusion models", "RLHF", "adjoint method" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 4, 4, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "xQAhUIuAc6", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Ahyun Seo", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Intern" }, { "name": "Minsu Cho", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Associate Professor" }, { "name": "Won-gyun Yu", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" } ], "n_formula": 74, "n_formula_1": 40, "n_ref_uni": 54, "n_ref": 135, "n_ref_all": 155, "n_fig": 7, "n_tab": 8, "L_tab": 880, "n_element_tab": 83, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 33077, "formula_len_all_1": 27426, "len_all": 227158, "len_all_1": 124748, "len_abs": 1030, "len_title": 128, "len_sents": 62415, "len_sents_1": 50509, "n_sents": 399, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1033, "title": "Axis-level Reflectional Symmetry Detection with Group-Equivariant Representation", "abs": "Reflectional symmetry detection remains a challenging task in machine perception, particularly in complex real-world scenarios involving noise, occlusions, and distortions. We introduce a novel equivariant approach to axis-level reflectional symmetry detection that effectively leverages dihedral group-equivariant representation to detect symmetry axes as line segments. We propose orientational anchor expansion for fine-grained rotation-equivariant analysis of diverse symmetry patterns across multiple orientations. Additionally, we develop reflectional matching with multi-scale kernels to extract effective cues of reflectional correlations, allowing for robust symmetry detection across different receptive fields. Our approach unifies axis-level detection with reflectional matching while preserving dihedral group equivariance throughout the process. Extensive experiments demonstrate the efficacy of our method while providing more accurate axis-level predictions than existing pixel-level methods in challenging scenarios.", "keywords": [ "Symmetry detection", "Equivariant learning", "Group equivariance" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 3, 1 ], "confidence_list": [ 2, 5, 5 ] }, { "paper_id": "xPxHQHDH2u", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Chun Gu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Li Zhang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiatian Zhu", "gender": "unknown", "institution": "University of Surrey", "country": "GB", "position": "Associate Professor" }, { "name": "Yuxuan Yao", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Zixuan Zeng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 30, "n_ref": 85, "n_ref_all": 116, "n_fig": 12, "n_tab": 7, "L_tab": 4385, "n_element_tab": 458, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 3085, "n_element_tab_1": 260, "formula_len_all": 724, "formula_len_all_1": 724, "len_all": 110308, "len_all_1": 50248, "len_abs": 353, "len_title": 135, "len_sents": 30568, "len_sents_1": 20431, "n_sents": 219, "n_sents_1": 140, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 29, "L_abs": 1341, "title": "Reflective Gaussian Splatting", "abs": "Novel view synthesis has experienced significant advancements owing to increasingly capable NeRF- and 3DGS-based methods. However, reflective object reconstruction remains challenging, lacking a proper solution to achieve real-time, high-quality rendering while accommodating inter-reflection. To fill this gap, we introduce a Reflective Gaussian splatting (Ref-Gaussian) framework characterized with two components: (I) Physically based deferred rendering that empowers the rendering equation with pixel-level material properties via formulating split-sum approximation; (II) Gaussian-grounded inter-reflection that realizes the desired inter-reflection function within a Gaussian splatting paradigm for the first time. To enhance geometry modeling, we further introduce material-aware normal propagation and an initial per-Gaussian shading stage, along with 2D Gaussian primitives. Extensive experiments on standard datasets demonstrate that Ref-Gaussian surpasses existing approaches in terms of quantitative metrics, visual quality, and compute efficiency. Further, we show that our method serves as a unified solution for both reflective and non-reflective scenes, going beyond the previous alternatives focusing on only reflective scenes. Also, we illustrate that Ref-Gaussian supports more applications such as relighting and editing.", "keywords": [ "Gaussain-Splatting", "Physically based Rendering", "Deferred-Rendering", "Inter-Reflection" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "xPTzjpIQNp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "BinChen", "gender": "unknown", "institution": "Zhejiang Lab, Zhejiang Lab", "country": "CN", "position": "Researcher" }, { "name": "Hao Wang", "gender": "Male", "institution": "ByteDance", "country": "CN", "position": "Researcher" }, { "name": "Haoxuan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Mingming Gong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Xu Chen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "zhengnan li", "gender": "Male", "institution": "Communication University of China", "country": "CN", "position": "Undergrad student" }, { "name": "Ziciu Can", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 68, "n_formula_1": 13, "n_ref_uni": 47, "n_ref": 95, "n_ref_all": 123, "n_fig": 9, "n_tab": 8, "L_tab": 6578, "n_element_tab": 338, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 2870, "n_element_tab_1": 126, "formula_len_all": 5256, "formula_len_all_1": 968, "len_all": 209530, "len_all_1": 66211, "len_abs": 1099, "len_title": 92, "len_sents": 69463, "len_sents_1": 27792, "n_sents": 484, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 44, "L_abs": 1123, "title": "Optimal Transport for Time Series Imputation", "abs": "Missing data imputation through distribution alignment has demonstrated advantages for non-temporal datasets but exhibits suboptimal performance in time-series applications. The primary obstacle is crafting a discrepancy measure that simultaneously (1) $\\textit{captures temporal pattern}$\u2014accounting for patterns such as periodicities and temporal dependencies inherent in time-series\u2014and (2) $\\textit{accommodates non-stationarity}$, ensuring robustness amidst multiple coexisting temporal patterns. In response to these challenges, we introduce the Proximal Spectrum Wasserstein (PSW) discrepancy based on the stochastic optimal transport framework, which incorporates a pairwise spectral distance to encapsulate temporal patterns, coupled with selective matching regularization to accommodate non-stationarity. Building upon PSW, we develop the PSW for Imputation (PSW-I) framework, which iteratively refines imputation results by minimizing the PSW discrepancy. Extensive experiments demonstrate that PSW-I effectively addresses these challenges and significantly outperforms prevailing time-series imputation methods.", "keywords": [ "Time series", "Imputation" ], "rating_list": [ 5, 8, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 4 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 2, 2, 2 ] }, { "paper_id": "xPO6fwvldG", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Haoyu Guo", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Hujun Bao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Sida Peng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaowei Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Xingyi He", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yuang Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yujian Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yujun Shen", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 59, "n_ref": 97, "n_ref_all": 114, "n_fig": 8, "n_tab": 5, "L_tab": 1898, "n_element_tab": 128, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 1309, "n_element_tab_1": 81, "formula_len_all": 147, "formula_len_all_1": 181, "len_all": 185220, "len_all_1": 69544, "len_abs": 4870, "len_title": 112, "len_sents": 53867, "len_sents_1": 32511, "n_sents": 407, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1392, "title": "UniRestore3D: A Scalable Framework For General Shape Restoration", "abs": "Shape restoration aims to recover intact 3D shapes from defective ones, such as those that are incomplete, noisy, and low-resolution. Previous works have achieved impressive results in shape restoration subtasks thanks to advanced generative models. While effective for specific shape defects, they are less applicable in real-world scenarios involving multiple defect types simultaneously. Additionally, training on limited subsets of defective shapes hinders knowledge transfer across restoration types and thus affects generalization. In this paper, we address the task of general shape restoration, which restores shapes with various types of defects through a unified model, thereby naturally improving the applicability and scalability. Our approach first standardizes the data representation across different restoration subtasks using high-resolution TSDF grids and constructs a large-scale dataset with diverse types of shape defects. Next, we design an efficient hierarchical shape generation model and a noise-robust defective shape encoder that enables effective impaired shape understanding and intact shape generation. Moreover, we propose a scalable training strategy for efficient model training. The capabilities of our proposed method are demonstrated across multiple shape restoration subtasks and validated on various datasets, including Objaverse, ShapeNet, GSO, and ABO.", "keywords": [ "Shape Restoration", "3D Reconstruction", "Diffusion Model" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "xP1radUi32", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Leonard Tang", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Undergrad student" }, { "name": "Maximilian Li", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "Undergrad student" }, { "name": "Brian R Huang", "gender": "Male", "institution": "Haize Labs", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 38, "n_ref": 71, "n_ref_all": 95, "n_fig": 8, "n_tab": 2, "L_tab": 7625, "n_element_tab": 52, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 41, "n_element_tab_1": 3, "formula_len_all": 62, "formula_len_all_1": 62, "len_all": 137821, "len_all_1": 53287, "len_abs": 1199, "len_title": 42, "len_sents": 41209, "len_sents_1": 27226, "n_sents": 288, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 1006, "title": "Endless Jailbreaks with Bijection Learning", "abs": "Despite extensive safety measures, LLMs are vulnerable to adversarial inputs, or jailbreaks, which can elicit unsafe behaviors. In this work, we introduce bijection learning, a powerful attack algorithm which automatically fuzzes LLMs for safety vulnerabilities using randomly-generated encodings whose complexity can be tightly controlled. We leverage in-context learning to teach models bijective encodings, pass encoded queries to the model to bypass built-in safety mechanisms, and finally decode responses back into English. Our attack is extremely effective on a wide range of frontier language models. Moreover, by controlling complexity parameters such as number of key-value mappings in the encodings, we find a close relationship between the capability level of the attacked LLM and the average complexity of the most effective bijection attacks. Our work highlights that new vulnerabilities in frontier models can emerge with scale: more capable models are more susceptible to bijection attacks.", "keywords": [ "jailbreaking", "redteaming", "AI safety", "AI alignment", "adversarial robustness", "adversarial attacks" ], "rating_list": [ 8, 8, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "xOtOfdbBqK", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Brian Park", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "MS student" }, { "name": "Jiesong Liu", "gender": "unknown", "institution": "North Carolina State University", "country": "US", "position": "PhD student" }, { "name": "Xipeng Shen", "gender": "Male", "institution": "North Carolina State University", "country": "", "position": "Professor" } ], "n_formula": 22, "n_formula_1": 16, "n_ref_uni": 28, "n_ref": 47, "n_ref_all": 69, "n_fig": 7, "n_tab": 7, "L_tab": 3012, "n_element_tab": 311, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 105, "n_element_tab_1": 15, "formula_len_all": 1131, "formula_len_all_1": 808, "len_all": 132686, "len_all_1": 63631, "len_abs": 1197, "len_title": 141, "len_sents": 38955, "len_sents_1": 29708, "n_sents": 363, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1230, "title": "A Drop-In Solution for On-the-Fly Adaptation of Speculative Decoding in Large Language Models", "abs": "Large Language Models (LLMs) are cutting-edge generative AI models built on transformer architecture, which tend to be highly memory-intensive when performing real-time inference. Various strategies have been developed to enhance the end-to-end inference speed for LLMs, one of which is speculative decoding. This technique involves running a smaller LLM (draft model) for inference over a defined window size, denoted as $\\gamma$, while simultaneously being validated by the larger LLM (target model). Choosing the optimal $\\gamma$ value and the draft model is essential for unlocking the potential of speculative decoding. But it is difficult to do due to the complicated influence from various factors, including the nature of the task, the hardware in use, and the combination of the large and small models. \nThis paper introduces *on-the-fly adaption of speculative decoding*, a solution that dynamically adapts the choices to maximize the efficiency of speculative decoding for LLM inferences. As a drop-in solution, it needs no offline benchmarking or training. \nExperiments show that the solution can lead to 3.55-16.48\\% speed improvement over the standard speculative decoding, and 1.2-3.4$\\times$ over the default LLMs.", "keywords": [ "LLM optimizations" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "xOmC5LiVuN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Adrian V Dalca", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" }, { "name": "Benjamin Billot", "gender": "Not Specified", "institution": "INRIA", "country": "FR", "position": "Principal Researcher" }, { "name": "Clinton Wang", "gender": "Male", "institution": "Google", "country": "", "position": "Intern" }, { "name": "Ellen Grant", "gender": "Female", "institution": "Boston Children's Hospital", "country": "US", "position": "Full Professor" }, { "name": "Mengwei Ren", "gender": "Female", "institution": "Adobe", "country": "US", "position": "Researcher" }, { "name": "Neel Dey", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Polina Golland", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Hallee E Wong", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 82, "n_ref": 148, "n_ref_all": 202, "n_fig": 13, "n_tab": 9, "L_tab": 2651, "n_element_tab": 293, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1719, "n_element_tab_1": 208, "formula_len_all": 244, "formula_len_all_1": 112, "len_all": 260894, "len_all_1": 53985, "len_abs": 1171, "len_title": 133, "len_sents": 60868, "len_sents_1": 24152, "n_sents": 431, "n_sents_1": 155, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1179, "title": "Learning General-purpose Biomedical Volume Representations using Randomized Synthesis", "abs": "Current _volumetric_ biomedical foundation models struggle to generalize as public 3D datasets are small and do not cover the broad diversity of medical procedures, conditions, anatomical regions, and imaging protocols. We address this by creating a representation learning method that instead anticipates strong domain shifts at training time itself. We first propose a data engine that synthesizes highly variable training samples that enable generalization to new biomedical contexts. To then train a single 3D network for any voxel-level task in radiology, we develop a contrastive learning method that pretrains the network to be stable against nuisance imaging variation simulated by the data engine, a key inductive bias for generalization. This network's features can be used as robust representations of input images for downstream tasks and its weights provide a strong, dataset-agnostic initialization for finetuning on new datasets. As a result, we set new standards across _both_ multimodality registration and few-shot segmentation, a first for any 3D biomedical vision model, all without (pre-)training on any existing dataset of real images. Our code is attached.", "keywords": [ "synthetic data", "representation learning", "medical image analysis", "image registration" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "xOZYU67EKL", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Han Yu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Pengwei Xing", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Songtao Lu", "gender": "Male", "institution": "IBM Thomas J. Watson Research Center", "country": "", "position": "Senior Research Scientist" } ], "n_formula": 54, "n_formula_1": 19, "n_ref_uni": 28, "n_ref": 32, "n_ref_all": 38, "n_fig": 1, "n_tab": 1, "L_tab": 1188, "n_element_tab": 118, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2708, "formula_len_all_1": 1196, "len_all": 116036, "len_all_1": 65608, "len_abs": 1512, "len_title": 69, "len_sents": 38024, "len_sents_1": 32006, "n_sents": 299, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1520, "title": "MMD-NSL: Mixed Multinomial Distribution-based Neuro-Symbolic Learning", "abs": "Neuro-symbolic learning (NSL) aims to integrate neural networks with symbolic reasoning approaches to enhance the interpretability of machine learning models. Existing methods mostly focus on the long dependency problem of symbolic learning. The important challenge of complex categorization is largely overlooked. To bridge this gap, we propose the Mixed Multinomial Distribution-based NSL MMD-NSL framework. It seamlessly integrates the handling of long dependency chains and complex semantic categorization within Knowledge Graphs (KGs). By introducing a continuous Mixed Multinomial Logic Semantic Distribution, we extend traditional Markov Logic Networks (MLN) to incorporate context-aware semantic embeddings. Our theoretical innovations, including a bijective mapping between MLNs and continuous multinomial distributions, enable the capture of intricate dependencies and varied contexts crucial for NSL tasks.\nThe framework leverages a bilevel optimization strategy, where a transformer-based upper level dynamically learns mixing coefficients akin to attention mechanisms, while the lower level optimizes rule weights for learning both context and rule patterns. Extensive experiments on the DWIE benchmarking datasets demonstrate significant advantages of MMD-NSL over four state-of-the-art approaches. It achieves 10.47% higher F1-scores on average than the best-performing baseline across 23 sub-datasets. It advances continuous probabilistic models for neuro-symbolic reasoning and complex relational tasks.", "keywords": [ "Neuro-Symbolic Learning", "Multinomial Mixture Distribution" ], "rating_list": [ 5, 3, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 1, 1, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 4, 2 ] }, { "paper_id": "xNwmWaq2KN", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Gui Xu", "gender": "Female", "institution": "Dalian University", "country": "CN", "position": "Researcher" }, { "name": "Hua Yu", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Jiao Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Melvin Wong", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Qiang Zhang", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yaqing Hou", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yew Soon Ong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 12, "n_ref_uni": 24, "n_ref": 37, "n_ref_all": 49, "n_fig": 4, "n_tab": 5, "L_tab": 5501, "n_element_tab": 308, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1312, "n_element_tab_1": 165, "formula_len_all": 1400, "formula_len_all_1": 906, "len_all": 131712, "len_all_1": 60370, "len_abs": 1325, "len_title": 145, "len_sents": 35304, "len_sents_1": 27145, "n_sents": 286, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1318, "title": "Novelty Unlocking with Multiobjective Generative Models: Batch Diversity of Human Motions", "abs": "Current generative models have shown potential performance in many tasks, which typically focus on generating samples that closely adhere to a given distribution, often overlooking the requirement to produce optimal diverse solutions in a batch diversity.\nRecognizing that maintaining ``diversity\" has been a longstanding challenge in multiobjective optimization, we were inspired to introduce a multiobjective optimization approach to enhance diversity in a single pass.\nThis paper utilizes the in-betweening human motion generation task as an example and introduces the multiobjective generative models to demonstrate the effectiveness of the proposed method in producing diverse and smooth human motion sequences. The resulting method, termed the \\textit{Multiobjective Generation Framework with In-Betweening Motion Model} (MGF-IMM), frames the human motion in-betweening task as a bi-objective optimization problem. The designed in-betweening motion model is then integrated into a nondominated sorting-based optimization framework to address this bi-objective optimization problem.\nThrough comprehensive qualitative and quantitative experiments, MGF-IMM has demonstrated state-of-the-art performance, surpassing the latest methods and validating its superiority in generating diverse in-betweening human motions.", "keywords": [ "Multiobjective optimization;Diverse In-Betweening Human Motions" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 2 ] }, { "paper_id": "xNsIfzlefG", "primary_area": "generative models", "n_author": 1, "author_info_list": [ { "name": "Lei Yang", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 15, "n_ref": 24, "n_ref_all": 54, "n_fig": 16, "n_tab": 2, "L_tab": 104, "n_element_tab": 24, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 104, "n_element_tab_1": 24, "formula_len_all": 581, "formula_len_all_1": 537, "len_all": 123515, "len_all_1": 50387, "len_abs": 3781, "len_title": 78, "len_sents": 41180, "len_sents_1": 22723, "n_sents": 358, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 30, "L_abs": 1250, "title": "Discrete Distribution Networks", "abs": "We introduce a novel generative model, the Discrete Distribution Networks (DDN), that approximates data distribution using hierarchical discrete distributions. We posit that since the features within a network inherently capture distributional information, enabling the network to generate multiple samples simultaneously, rather than a single output, may offer an effective way to represent distributions. Therefore, DDN fits the target distribution, including continuous ones, by generating multiple discrete sample points. To capture finer details of the target data, DDN selects the output that is closest to the Ground Truth (GT) from the coarse results generated in the first layer. This selected output is then fed back into the network as a condition for the second layer, thereby generating new outputs more similar to the GT. As the number of DDN layers increases, the representational space of the outputs expands exponentially, and the generated samples become increasingly similar to the GT. This hierarchical output pattern of discrete distributions endows DDN with unique property: more general zero-shot conditional generation. We demonstrate the efficacy of DDN and its intriguing properties through experiments on CIFAR-10 and FFHQ.", "keywords": [ "Generative Models", "Image Generation" ], "rating_list": [ 8, 8, 5 ], "soundness_list": [ 4, 3, 2 ], "presentation_list": [ 4, 3, 2 ], "contribution_list": [ 4, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "xNgmEWmd9T", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 5, "author_info_list": [ { "name": "Fabian Grob", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "MS student" }, { "name": "Giuseppe Franco", "gender": "Male", "institution": "AMD", "country": "", "position": "Researcher" }, { "name": "Ian Colbert", "gender": "unknown", "institution": "Advanced Micro Devices", "country": "US", "position": "Researcher" }, { "name": "Jinjie Zhang", "gender": "Male", "institution": "GSK", "country": "US", "position": "Researcher" }, { "name": "Rayan Saab", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "Full Professor" } ], "n_formula": 41, "n_formula_1": 22, "n_ref_uni": 43, "n_ref": 141, "n_ref_all": 160, "n_fig": 2, "n_tab": 4, "L_tab": 2160, "n_element_tab": 238, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 256, "n_element_tab_1": 25, "formula_len_all": 2912, "formula_len_all_1": 1140, "len_all": 163780, "len_all_1": 67557, "len_abs": 1370, "len_title": 119, "len_sents": 53126, "len_sents_1": 33821, "n_sents": 346, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1327, "title": "Accumulator-Aware Post-Training Quantization for Large Language Models", "abs": "Several recent studies have investigated low-precision accumulation, reporting improvements in throughput, power, and area across various platforms. However, the accompanying proposals have only considered the quantization-aware training (QAT) paradigm, in which models are fine-tuned or trained from scratch with quantization in the loop. As models continue to grow in size, QAT techniques become increasingly more expensive, which has motivated the recent surge in post-training quantization (PTQ) research. To the best of our knowledge, ours marks the first formal study of accumulator-aware quantization in the PTQ setting. To bridge this gap, we introduce AXE\u2014a practical, low-overhead framework of accumulator-aware extensions designed to endow overflow avoidance guarantees to existing layer-wise PTQ algorithms. We theoretically motivate AXE and demonstrate its flexibility by implementing it on top of two state-of-the-art PTQ algorithms: GPFQ and OPTQ. We further generalize AXE to support multi-stage accumulation for the first time, opening the door for full datapath optimization and scaling to large language models (LLMs). We evaluate AXE across autoregressive language generation models and observe significant improvements in the tradeoff between accumulator bit width and model accuracy over baseline methods.", "keywords": [ "Accumulators", "Deep Learning", "Inference", "Quantization" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 2, 3 ] }, { "paper_id": "xNf8sOtFbx", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Gang Niu", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Senior Research Scientist (tenured)" }, { "name": "Ji-Long Niu", "gender": "unknown", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Masashi Sugiyama", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Director" }, { "name": "Ming-Kun Xie", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Postdoc" }, { "name": "Sheng-Jun Huang", "gender": "unknown", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Jiahao Xiao", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 28, "n_ref": 41, "n_ref_all": 56, "n_fig": 5, "n_tab": 3, "L_tab": 1147, "n_element_tab": 206, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 260, "n_element_tab_1": 33, "formula_len_all": 616, "formula_len_all_1": 592, "len_all": 101636, "len_all_1": 59868, "len_abs": 1714, "len_title": 131, "len_sents": 33269, "len_sents_1": 30565, "n_sents": 230, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1741, "title": "On the Cost-Effectiveness of Partially-Annotating Methods for Multi-Label Learning", "abs": "Precisely annotating instances with multiple labels is costly and has emerged as a significant bottleneck in the real-world multi-label learning tasks. To deal with this problem, the most straightforward strategy is partially-annotating, which aims to reduce the cost by annotating only a subset of labels. Existing works mainly includes label-level partially-annotating (LPA), where each instance is assigned a subset of positive labels, and instance-level partially-annotating (IPA), where all positive labels are assigned to an instance, but only a subset of instances are annotated. However, these methods tend to focus on improving model performance under each type of partial annotation, often neglecting a fundamental question: \\textit{which method is the most cost-effective?} In this paper, we empirically evaluate which partially-annotating method achieves better model performance at the same annotation cost. To make a fair comparison, we manually annotated images in the MS-COCO dataset using two partially-annotating methods and recorded their averaging annotation time per image. This allows us to train models on two types of partial annotations with the same annotation cost and to compare their performance. Empirical results show that even when the number of examples annotated with IPA is only one-fifth that of LPA, models trained on IPA annotations significantly outperform those trained on LPA annotations, yielding that IPA is significantly more cost-effective than LPA. To explain the superiority of IPA, our causal reasoning framework shows that compared to LPA, IPA preserves complete co-occurrence relationships, enabling the model to capture correlative patterns, which is useful for improving model performance.", "keywords": [ "Partially-annotating", "multi-label learning" ], "rating_list": [ 6, 5, 3, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "xNaPs8bdLa", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Conghan Yue", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "MS student" }, { "name": "Dongyu Zhang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Shiyan Du", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Zhengwei Peng", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "MS student" }, { "name": "Zhi Ji", "gender": "Female", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" } ], "n_formula": 24, "n_formula_1": 17, "n_ref_uni": 28, "n_ref": 48, "n_ref_all": 58, "n_fig": 7, "n_tab": 3, "L_tab": 98, "n_element_tab": 22, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 91, "n_element_tab_1": 20, "formula_len_all": 2292, "formula_len_all_1": 1721, "len_all": 129993, "len_all_1": 57769, "len_abs": 1809, "len_title": 123, "len_sents": 28448, "len_sents_1": 25170, "n_sents": 192, "n_sents_1": 166, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1816, "title": "Aggregation of Multi Diffusion Models for Enhancing Learned Representations", "abs": "Diffusion models have achieved remarkable success in image generation, particularly with the various applications of classifier-free guidance conditional diffusion models. While many diffusion models perform well when controlling for particular aspect among style, character, and interaction, they struggle with fine-grained control due to dataset limitations and intricate model architecture design. This paper introduces a novel algorithm, Aggregation of Multi Diffusion Models (AMDM), which synthesizes features from multiple diffusion models into a specified model, enhancing its learned representations to activate specific features for fine-grained control. AMDM consists of two key components: spherical aggregation and manifold optimization. Spherical aggregation merges intermediate variables from different diffusion models with minimal manifold deviation, while manifold optimization refines these variables to align with the intermediate data manifold, enhancing sampling quality. Experimental results demonstrate that AMDM significantly improves fine-grained control without additional training or inference time, proving its effectiveness. Additionally, it reveals that diffusion models initially focus on features such as position, attributes, and style, with later stages improving generation quality and consistency. AMDM offers a new perspective for tackling the challenges of fine-grained conditional control generation in diffusion models: We can fully utilize existing conditional diffusion models that control specific aspects, or develop new ones, and then aggregate them using the AMDM algorithm. This eliminates the need for constructing complex datasets, designing intricate model architectures, and incurring high training costs. Code is available at: https://github.com/Hammour-steak/AMDM", "keywords": [ "Diffusion Models", "Conditional Generation" ], "rating_list": [ 3, 3, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 2, 1, 3, 3, 3 ], "contribution_list": [ 2, 1, 3, 3, 3 ], "confidence_list": [ 4, 4, 5, 4, 3 ] }, { "paper_id": "xNDydjYBmC", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Guoxin Fan", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "Researcher" }, { "name": "Huaqing Liu", "gender": "Male", "institution": "Research Institute of Tsinghua, Pearl River Delta", "country": "CN", "position": "Researcher" }, { "name": "Lanqing Han", "gender": "Male", "institution": "Research Institute of Tsinghua, Pearl River Delta", "country": "CN", "position": "Principal Researcher" }, { "name": "Peiyi Chen", "gender": "unknown", "institution": " Research Institute of Tsinghua, Pearl River Delta", "country": "CN", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 24, "n_ref": 40, "n_ref_all": 60, "n_fig": 17, "n_tab": 4, "L_tab": 3513, "n_element_tab": 322, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 301, "n_element_tab_1": 51, "formula_len_all": 732, "formula_len_all_1": 695, "len_all": 142044, "len_all_1": 52058, "len_abs": 1436, "len_title": 138, "len_sents": 36018, "len_sents_1": 25420, "n_sents": 246, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 137, "L_abs": 1436, "title": "Enhancing PPB Affinity Prediction through Data Integration and Feature Alignment: Approaching Structural Model Performance with Sequences", "abs": "One key step of protein drug development is the screening of protein-protein binding (PPB) affinity. The current mainstream screening method of PPB affinity is laboratory experiments, which are costly and time-consuming, making it difficult to quickly perform high-throughput screening. Various deep learning methods have been proposed to predict PPB affinity, but they are often limited by the availability of high-quality data and the compatibility of the algorithms with that data. In this work, we developed two AI models, PPBind-3D and PPBind-1D, to predict PPB affinity. PPBind-3D leverages structural information near the protein-protein binding interface to make its predictions. By employing monotonic neural network constrained multi-task learning, we effectively utilized heterogeneous affinity data from diverse wet lab experiments to expand the development dataset to over 23,000 samples, thereby enhancing the model's generalization capabilities. Additionally, PPBind-1D was developed using sequence data to address the lack of structural data in practical applications. During the training of PPBind-1D, we aligned it with PPBind-3D by incorporating an additional 42,108 no-affinity-label samples through an alignment approach. Finally, we demonstrated three application cases of our AI models in the virtual screening of protein drugs, illustrating that our models can significantly facilitate high-throughput screening.", "keywords": [ "binding affinity", "geometric deep learning", "virtual screening" ], "rating_list": [ 3, 5, 6, 3, 6 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 3, 4 ] }, { "paper_id": "xNCDKQMPYD", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Cheng Cheng", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Hongbin Sun", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Lin Song", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Xuchong Zhang", "gender": "unknown", "institution": "Xi'an Jiaotong University", "country": "", "position": "Associate Professor" }, { "name": "Yicheng Xiao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Ying Shan", "gender": "Male", "institution": "Tencent AI Lab Center of Visual Computing", "country": "CN", "position": "Director" }, { "name": "Yixiao Ge", "gender": "Female", "institution": "Tencent", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 29, "n_ref": 64, "n_ref_all": 74, "n_fig": 3, "n_tab": 3, "L_tab": 1337, "n_element_tab": 71, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1423, "n_element_tab_1": 77, "formula_len_all": 773, "formula_len_all_1": 787, "len_all": 86730, "len_all_1": 52031, "len_abs": 1250, "len_title": 110, "len_sents": 23490, "len_sents_1": 23174, "n_sents": 159, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1254, "title": "GPT4LoRA: Optimizing LoRA Combination via MLLM Self-Reflection", "abs": "Low-Rank Adaptation (LoRA) is extensively used in generative models to enable concept-driven personalization, such as rendering specific characters or adopting unique styles. Although recent approaches have explored LoRA combination to integrate diverse concepts, they often require further fine-tuning or modifications to the generative model's original architecture. To address these limitations, we introduce GPT4LoRA, a novel method for LoRA combination that adjusts combination coefficients by leveraging the self-reflection capabilities of multimodal large language models (MLLMs). GPT4LoRA operates through a three-step process\u2014Generate, Feedback, and Refine\u2014without the need for additional training, relying solely on tailored prompts and iterative refinement to enhance performance. This iterative approach ensures more constructive feedback and optimizes the model responses. Experiments on various LoRA model combinations, including both realistic and anime styles, demonstrate that GPT4LoRA achieves superior results compared to existing methods. Additionally, an evaluation framework based on GPT-4o further highlights the clear performance gains offered by GPT4LoRA over standard baselines, showcasing its potential for advancing the field.", "keywords": [ "MLLM", "Self-Reflection", "LoRA Combination" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "xN6z16agjE", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Husni A. Al-Muhtaseb", "gender": "Male", "institution": "King Fahad University of Petroleum and Minerals", "country": "SA", "position": "Assistant Professor" }, { "name": "Randah Alharbi", "gender": "Female", "institution": "King Fahad University of Petroleum and Minerals", "country": "SA", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 33, "n_ref_all": 40, "n_fig": 0, "n_tab": 9, "L_tab": 1442, "n_element_tab": 209, "n_fig_1": 0, "n_tab_1": 8, "L_tab_1": 1872, "n_element_tab_1": 187, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 112686, "len_all_1": 62021, "len_abs": 1113, "len_title": 124, "len_sents": 29445, "len_sents_1": 28804, "n_sents": 209, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1120, "title": "Evaluating word representation for hypernymy relation: with focus on Arabic", "abs": "Hypernymy relation is one of the fundamental relations for many natural language processing and information extraction tasks. A key component of the performance of any hypernymy-related task is word representation. Traditional word embeddings capture word similarity but fall short of representing more complex lexical-semantic relationships between terms, such as hypernymy. To overcome this, recent studies have proposed hypernymy-specific representations. In this study, we conduct an evaluation of several types of word representations to determine the most effective approach for modeling hypernymy relationships in Arabic. We use an Arabic training corpus and several datasets to assess traditional embedding, hypernymy-specific embedding, and contextual embedding across several hypernymy-related tasks, including hypernymy detection. The results indicate that different embeddings have different effects on the performance. Moreover, the performance is affected by the selected datasets. This highlights that there is a need for further research to develop more robust word representation and benchmark datasets.", "keywords": [ "Word representation", "hypernymy relation", "hypernymy specific embedding", "hypernymy detection." ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 1 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "xMxHJxp192", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Islem Rekik", "gender": "Female", "institution": "Imperial College London", "country": "", "position": "Associate Professor" }, { "name": "Kevin Mancini", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 72, "n_ref_all": 100, "n_fig": 7, "n_tab": 7, "L_tab": 11306, "n_element_tab": 568, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2711, "n_element_tab_1": 149, "formula_len_all": 896, "formula_len_all_1": 561, "len_all": 170482, "len_all_1": 62444, "len_abs": 1352, "len_title": 109, "len_sents": 49836, "len_sents_1": 25864, "n_sents": 364, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1358, "title": "DeltaGNN: Graph Neural Network with Information Flow Control", "abs": "Graph Neural Networks (GNNs) are popular machine learning models designed to process graph-structured data through recursive neighborhood aggregations in the message passing process. When applied to semi-supervised node classification, the message-passing enables GNNs to understand short-range spatial interactions, but also causes them to suffer from over-smoothing and over-squashing. These challenges hinder model expressiveness and prevent the use of deeper models to capture long-range node interactions (LRIs) within the graph. Popular solutions for LRIs detection are either too expensive to process large graphs due to high time complexity or fail to generalize across diverse graph structures. To address these limitations, we propose a mechanism called information flow control, which leverages a novel connectivity measure, called information flow score, to address over-smoothing and over-squashing with linear computational overhead, supported by theoretical evidence. Finally, to prove the efficacy of our methodology we design DeltaGNN, the first scalable and generalizable approach for long-range and short-range interaction detection. \nWe benchmark our model across 10 real-world datasets, including graphs with varying sizes, topologies, densities, and homophilic ratios, showing superior performance with limited computational complexity.", "keywords": [ "deep learning", "neural network", "graph neural network", "topology", "homophily", "heterophily", "over-smoothing", "over-squashing", "long-range interactions" ], "rating_list": [ 3, 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 2, 4 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "xMOLUzo2Lk", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Chejian Xu", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Huan Sun", "gender": "Female", "institution": "The Ohio State University, Columbus", "country": "US", "position": "Associate Professor" }, { "name": "Jiawei Zhang", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Lingbo Mo", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Mintong Kang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "PhD student" }, { "name": "Yuan Tian", "gender": "Female", "institution": "University of Virginia", "country": "", "position": "Assistant Professor" }, { "name": "Zeyi Liao", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "chaowei xiao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 26, "n_ref": 55, "n_ref_all": 84, "n_fig": 11, "n_tab": 3, "L_tab": 2269, "n_element_tab": 174, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 622, "n_element_tab_1": 143, "formula_len_all": 276, "formula_len_all_1": 201, "len_all": 167207, "len_all_1": 70934, "len_abs": 2042, "len_title": 80, "len_sents": 58241, "len_sents_1": 33816, "n_sents": 469, "n_sents_1": 278, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 2054, "title": "EIA: ENVIRONMENTAL INJECTION ATTACK ON GENERALIST WEB AGENTS FOR PRIVACY LEAKAGE", "abs": "Recently, generalist web agents have demonstrated remarkable potential in autonomously completing a wide range of tasks on real websites, significantly boosting human productivity. However, web tasks, such as booking flights, usually involve users' personally identifiable information (PII), which may be exposed to potential privacy risks if web agents accidentally interact with compromised websites\u2014a scenario that remains largely unexplored in the literature.\nIn this work, we narrow this gap by conducting the first study on the privacy risks of generalist web agents in adversarial environments. First, we present a realistic threat model for attacks on the website, where we consider two adversarial targets: stealing users' specific PII or the entire user request.\nThen, we propose a novel attack method, termed Environmental Injection Attack (EIA). EIA injects malicious content designed to adapt well to environments where the agents operate and our work instantiates EIA specifically for privacy scenarios in web environments.\nWe collect 177 action steps that involve diverse PII categories on realistic websites from the Mind2Web dataset, and conduct experiments using one of the most capable generalist web agent frameworks to date. The results demonstrate that EIA achieves up to 70\\% attack success rate (ASR) in stealing users' specific PII and 16\\% ASR in stealing a full user request at an action step. Additionally, by evaluating the detectability and testing defensive system prompts, we indicate that EIA is challenging to detect and mitigate.\nNotably, attacks that are not well adapted for a webpage can be detected through careful human inspection, leading to our discussion about the trade-off between security and autonomy. However, extra attackers' efforts can make EIA seamlessly adapted, rendering such human supervision ineffective. Thus, we further discuss the implications on defenses at the pre- and post-deployment stages of the websites without relying on human supervision and call for more advanced defense strategies.", "keywords": [ "Web Agent", "Attack" ], "rating_list": [ 3, 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 3, 4, 3, 3, 4 ], "contribution_list": [ 1, 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3, 5 ] }, { "paper_id": "xLPakPOKDX", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Wonwoong Cho", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "David I Inouye", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" }, { "name": "Raymond Yeh", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 57, "n_ref": 110, "n_ref_all": 145, "n_fig": 14, "n_tab": 5, "L_tab": 431, "n_element_tab": 37, "n_fig_1": 11, "n_tab_1": 5, "L_tab_1": 552, "n_element_tab_1": 78, "formula_len_all": 1095, "formula_len_all_1": 774, "len_all": 189103, "len_all_1": 73355, "len_abs": 2002, "len_title": 128, "len_sents": 57332, "len_sents_1": 32937, "n_sents": 528, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 2014, "title": "Causally Motivated Diffusion Sampling Frameworks for Harnessing Contextual Bias", "abs": "Diffusion models have shown remarkable performance in text-guided image generation when trained on large-scale datasets, usually collected from the Internet. These large-scale datasets have contextual biases (e.g., co-occurrence of objects) which will naturally cascade into the diffusion model. For example, given a text prompt of ``a photo of the living room'', diffusion models frequently generate a couch, a rug, and a lamp together while rarely generating objects that do not commonly occur in a living room. Intuitively, contextual bias can be helpful because it naturally draws the scene even without detailed information (i.e., visual autofill). On the other hand, contextual bias can limit the diversity of generated images (e.g., diverse object combinations) to focus on common image compositions. To have the best of both worlds, we argue that contextual bias needs to be strengthened or weakened depending on the situation. Previous causally-motivated studies have tried to deal with such issues by analyzing confounders (i.e., contextual bias) and augmenting training data or designing their models to directly learn the interventional distribution. However, due to the large-scale nature of these models, obtaining and analyzing the data or training the huge model from scratch is beyond reach in practice. To tackle this problem, we propose two novel frameworks for strengthening or weakening the contextual bias of pretrained diffusion models without training any parameters or accessing training data. Briefly, we first propose causal graphs to explicitly model contextual bias in the generation process. We then sample the hidden confounder due to contextual bias by sampling from a chain of pretrained large-scale models. Finally, we use samples from the confounder to strengthen or weaken the contextual bias based on methods from causal inference. Experiment results show that our proposed methods are effective in generating more realistic and diverse images than the regular sampling method.", "keywords": [ "Causal Inference", "Diffusion Models", "Contextual bias", "Spurious Correlations", "Object Cooccurrence", "StableDiffusion" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "xKDZAW0He3", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Chin-Yew Lin", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Senior Principal Research Manager" }, { "name": "Dongsheng Li", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "H. Vicky Zhao", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CA", "position": "Associate Professor" }, { "name": "Hao Cheng", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Researcher" }, { "name": "Huiqiang Jiang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "RSDE" }, { "name": "Jianfeng Gao", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Lili Qiu", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Principal Researcher" }, { "name": "Qianhui Wu", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Xufang Luo", "gender": "Female", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" }, { "name": "Yuqing Yang", "gender": "unknown", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Zhuoshi Pan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 3, "n_ref_uni": 55, "n_ref": 123, "n_ref_all": 184, "n_fig": 21, "n_tab": 16, "L_tab": 7605, "n_element_tab": 984, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 2094, "n_element_tab_1": 370, "formula_len_all": 427, "formula_len_all_1": 196, "len_all": 238429, "len_all_1": 59566, "len_abs": 1553, "len_title": 131, "len_sents": 72376, "len_sents_1": 25223, "n_sents": 528, "n_sents_1": 153, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1441, "title": "SeCom: On Memory Construction and Retrieval for Personalized Conversational Agents", "abs": "To deliver coherent and personalized experiences in long-term conversations, existing approaches typically perform retrieval augmented response generation by constructing memory banks from conversation history at either the turn-level, session-level, or through summarization techniques.\nIn this paper, we explore the impact of different memory granularities and present two key findings: (1) Both turn-level and session-level memory units are suboptimal, affecting not only the quality of final responses, but also the accuracy of the retrieval process.\n(2) The redundancy in natural language introduces noise, hindering precise retrieval. We demonstrate that *LLMLingua-2*, originally designed for prompt compression to accelerate LLM inference, can serve as an effective denoising method to enhance memory retrieval accuracy.\n\nBuilding on these insights, we propose **SeCom**, a method that constructs a memory bank with topical segments by introducing a conversation **Se**gmentation model, while performing memory retrieval based on **Com**pressed memory units.\nExperimental results show that **SeCom** outperforms turn-level, session-level, and several summarization-based methods on long-term conversation benchmarks such as *LOCOMO* and *Long-MT-Bench+*. Additionally, the proposed conversation segmentation method demonstrates superior performance on dialogue segmentation datasets such as *DialSeg711*, *TIAGE*, and *SuperDialSeg*.", "keywords": [ "memory management", "conversational agent", "RAG", "text segmentation", "prompt compression" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "xJtWqVBZya", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Junyuan Deng", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Mingkai Jia", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Ping Tan", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Qian Zhang", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Wei Yin", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Xiaotao Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Xiaoyang Guo", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "XIAOXIAO LONG", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 36, "n_ref": 76, "n_ref_all": 92, "n_fig": 5, "n_tab": 12, "L_tab": 1926, "n_element_tab": 252, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1207, "n_element_tab_1": 138, "formula_len_all": 952, "formula_len_all_1": 857, "len_all": 130385, "len_all_1": 59842, "len_abs": 194, "len_title": 123, "len_sents": 36127, "len_sents_1": 26639, "n_sents": 279, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1464, "title": "DrivingWorld: Constructing World Model for Autonomous Driving via Video GPT", "abs": "Recent successes in autoregressive (AR) generation models, such as the GPT series in natural language processing, have motivated efforts to replicate this success in visual tasks. By leveraging the next-token prediction strategy, GPT-style models can forecast future events from past data. Some research aims to extend this approach to autonomous driving by building video-based world models capable of generating realistic future video sequences and predicting the ego state. However, the prior works tend to produce unsatisfactory results, since the classic GPT framework is designed to handle 1D contextual information, such as text, and lacks the inherent capability to model the spatial and temporal dynamics necessary for video generation. In this paper, we present DrivingWorld, a video-based world model for autonomous driving via a new GPT structure with spatial-temporal design. The key idea is to disentangle temporal and spatial information in the generation. Specifically, we first propose next-frame-prediction strategy to model temporal coherence between consecutive frames and then apply next-token-prediction strategy to capture spatial information within a frame. With the hybrid design, our work is capable of producing high-fidelity and consistent video clips with long-time duration. Experiments show that compared to the prior works, our method presents better quality of visual effects and more accurate controllable future video generation.", "keywords": [ "world model", "video generation" ], "rating_list": [ 5, 1, 6, 6 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 5, 1, 5, 3 ] }, { "paper_id": "xJljiPE6dg", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Akbir Khan", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Ethan Perez", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Jacob Steinhardt", "gender": "unknown", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" }, { "name": "Jiaxin Wen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Minlie Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Ruiqi Zhong", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Samuel R. Bowman", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Member of Technical Staff" }, { "name": "Shi Feng", "gender": "Male", "institution": "George Washington University", "country": "", "position": "Assistant Professor" }, { "name": "He He", "gender": "Female", "institution": "New York University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 31, "n_ref": 57, "n_ref_all": 98, "n_fig": 17, "n_tab": 4, "L_tab": 670, "n_element_tab": 66, "n_fig_1": 14, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 32, "formula_len_all_1": 18, "len_all": 204465, "len_all_1": 68633, "len_abs": 1275, "len_title": 96, "len_sents": 68515, "len_sents_1": 30616, "n_sents": 630, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 1176, "title": "Language Models Learn to Mislead Humans via RLHF", "abs": "Language models (LMs) can produce errors that are hard to detect for humans, especially when the task is complex.\nRLHF, the most popular post-training method, may exacerbate this problem: to achieve higher rewards, LMs might get better at convincing humans that they are right even when they are wrong. We study this phenomenon under a standard RLHF pipeline, calling it ``U-Sophistry'' since it is \\textbf{U}nintended by model developers. Specifically, we ask time-constrained (e.g., 3-10 minutes) human subjects to evaluate the correctness of model outputs and calculate humans' accuracy against gold labels. On a question-answering task (QuALITY) and programming task (APPS), RLHF makes LMs better at convincing our subjects but not at completing the task correctly. RLHF also makes the model harder to evaluate: our subjects' false positive rate increases by 24.1% on QuALITY and 18.3% on APPS.\nFinally, we show that probing, a state-of-the-art approach for detecting \\textbf{I}ntended Sophistry (e.g.~backdoored LMs), does not generalize to U-Sophistry. Our results highlight an important failure mode of RLHF and call for more research in assisting humans to align them.", "keywords": [ "RLHF", "reward hacking", "human evaluation" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 4, 4, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "xJc3PazBwS", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Afra Alishahi", "gender": "Female", "institution": "Tilburg University", "country": "", "position": "Associate Professor" }, { "name": "Grzegorz Chrupa\u0142a", "gender": "Male", "institution": "Tilburg University", "country": "", "position": "Associate Professor" }, { "name": "Hosein Mohebbi", "gender": "Male", "institution": "Tilburg University", "country": "NL", "position": "PhD student" }, { "name": "Ivan Titov", "gender": "unknown", "institution": "University of Edinburgh", "country": "", "position": "Full Professor" }, { "name": "Willem H. Zuidema", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 66, "n_ref_all": 84, "n_fig": 5, "n_tab": 4, "L_tab": 4349, "n_element_tab": 103, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 358, "n_element_tab_1": 51, "formula_len_all": 408, "formula_len_all_1": 390, "len_all": 129243, "len_all_1": 65025, "len_abs": 1075, "len_title": 133, "len_sents": 41046, "len_sents_1": 31670, "n_sents": 271, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1077, "title": "Disentangling Textual and Acoustic Features of Neural Speech Representations", "abs": "Neural speech models build entangled internal representations, which capture a variety of features (e.g., pitch, loudness, syntax, or semantics of an utterance) in a distributed encoding. This complexity makes it difficult to track how such representations rely on textual and acoustic information when used in downstream applications, limiting their interpretability. In this paper, we build upon the Information Bottleneck principle to propose a disentanglement framework that separates speech representations learned by pre-trained neural speech models into two distinct components: one encoding content (i.e., what can be transcribed as text) and the other encoding acoustic features relevant to a downstream task. We apply and evaluate our framework to emotion recognition and speaker identification target tasks, quantifying the contribution of textual and acoustic features at each model layer. We also apply our disentanglement framework as an attribution method to identify the most salient speech frame representations from both the textual and acoustic perspectives.", "keywords": [ "Disentangling Representations", "Spoken language Processing", "Speech Emotion Recognition", "Interpretability" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "xJXq6FkqEw", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Bo Chen", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Mingyuan Zhou", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Xinyue Hu", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "PhD student" }, { "name": "Zhibin Duan", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 20, "n_ref_uni": 67, "n_ref": 120, "n_ref_all": 142, "n_fig": 11, "n_tab": 3, "L_tab": 1644, "n_element_tab": 89, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 636, "n_element_tab_1": 56, "formula_len_all": 1296, "formula_len_all_1": 1103, "len_all": 161628, "len_all_1": 63912, "len_abs": 1402, "len_title": 145, "len_sents": 48484, "len_sents_1": 31496, "n_sents": 298, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1417, "title": "Enhancing Uncertainty Estimation and Interpretability with Bayesian Non-negative Decision Layer", "abs": "Although deep neural networks have demonstrated significant success due to their\npowerful expressiveness, most models struggle to meet practical requirements for\nuncertainty estimation. Concurrently, the entangled nature of deep neural net-\nworks leads to a multifaceted problem, where various localized explanation tech-\nniques reveal that multiple unrelated features influence the decisions, thereby un-\ndermining interpretability. To address these challenges, we develop a Bayesian\nNonnegative Decision Layer (BNDL), which reformulates deep neural networks\nas a conditional Bayesian non-negative factor analysis. By leveraging stochastic\nlatent variables, the BNDL can model complex dependencies and provide robust\nuncertainty estimation. Moreover, the sparsity and non-negativity of the latent\nvariables encourage the model to learn disentangled representations and decision\nlayers, thereby improving interpretability. We also offer theoretical guarantees\nthat BNDL can achieve effective disentangled learning. In addition, we developed\na corresponding variational inference method utilizing a Weibull variational in-\nference network to approximate the posterior distribution of the latent variables.\nOur experimental results demonstrate that with enhanced disentanglement capa-\nbilities, BNDL not only improves the model\u2019s accuracy but also provides reliable\nuncertainty estimation and improved interpretability.", "keywords": [ "Factor Analysis", "Uncertainty Estimation", "explainable AI", "Bayesian Last Layer" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "xJUZHhrh3N", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Philippe Preux", "gender": "Male", "institution": "Universit\u00e9 de Lille", "country": "FR", "position": "Full Professor" }, { "name": "Riad Akrour", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Researcher" }, { "name": "Timoth\u00e9e Mathieu", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Researcher" }, { "name": "Yann Berthelot", "gender": "unknown", "institution": "INRIA", "country": "FR", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 21, "n_ref_uni": 18, "n_ref": 34, "n_ref_all": 47, "n_fig": 7, "n_tab": 2, "L_tab": 328, "n_element_tab": 28, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 966, "n_element_tab_1": 35, "formula_len_all": 3091, "formula_len_all_1": 1393, "len_all": 111777, "len_all_1": 62648, "len_abs": 1428, "len_title": 146, "len_sents": 41836, "len_sents_1": 28675, "n_sents": 344, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 98, "L_abs": 1543, "title": "BiVWAC: Improving deep reinforcement learning algorithms using Bias-Variance Weighted Actor-Critic", "abs": "We introduce $\\textrm{\\textbf{Bi}as-\\textbf{V}ariance \\textbf{W}eighted \\textbf{A}ctor \\textbf{C}ritic (\\textbf{BiVWAC})}$, a modification scheme for actor-critic algorithms allowing control over the bias-variance weighting in the critic. In actor-critic algorithms, the critic loss is the Mean Squared Error (MSE). The MSE may be decomposed in terms of bias and variance. Based on this decomposition, BiVWAC constructs a new critic loss, through a hyperparameter $\\alpha$, to weigh bias vs variance. MSE and Actor with Variance Estimated Critic (AVEC, which only considers the variance in the MSE decomposition) are special cases of this weighting for $\\alpha=0.5$ and $\\alpha=0$ respectively. We demonstrate the theoretical consistency of our new critic loss and measure its performance on a set of tasks. We also study value estimation and gradient estimation capabilities of BiVWAC to understand the means by which BiVWAC impacts performance.\n We show experimentally that the MSE is suboptimal as a critic loss when compared to other $\\alpha$ values. We equip SAC and PPO with the BiVWAC loss to obtain BiVWAC-SAC and BiVWAC-PPO and we propose a safe $\\alpha$ value, $\\alpha^*$, for which BiVWAC-SAC is better than or equal to SAC in all studied tasks but one in terms of policy performance. We also point out that BiVWAC introduces minimal changes to the algorithms and virtually no additional computational cost. \n In addition we also present a method to compare the impact of critic modifications between algorithms in a sound manner.", "keywords": [ "Reinforcement Learning", "Bias", "Variance", "Actor-Critic", "Deep Reinforcement Learning", "SAC", "PPO", "AVEC", "Mujoco" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 2, 1 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "xJDxVDG3x2", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Deli Zhao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Director" }, { "name": "Liang Wang", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Liang Wang", "gender": "Male", "institution": "Institute of Automation\uff0c CAS\uff0cChina", "country": "CN", "position": "Full Professor" }, { "name": "Qiang Liu", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Shaozhen Liu", "gender": "unknown", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Shu Wu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yu Rong", "gender": "Male", "institution": "Alibaba Group ", "country": "CN", "position": "Senior Staff Algorithm Engineer" } ], "n_formula": 14, "n_formula_1": 16, "n_ref_uni": 40, "n_ref": 81, "n_ref_all": 101, "n_fig": 5, "n_tab": 9, "L_tab": 3538, "n_element_tab": 398, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 1644, "n_element_tab_1": 158, "formula_len_all": 929, "formula_len_all_1": 1248, "len_all": 154114, "len_all_1": 63801, "len_abs": 1249, "len_title": 133, "len_sents": 39238, "len_sents_1": 27144, "n_sents": 301, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1254, "title": "MolSpectra: Pre-training 3D Molecular Representation with Multi-modal Energy Spectra", "abs": "Establishing the relationship between 3D structures and the energy states of molecular systems has proven to be a promising approach for learning 3D molecular representations. However, existing methods are limited to modeling the molecular energy states from classical mechanics. This limitation results in a significant oversight of quantum mechanical effects, such as quantized (discrete) energy level structures, which offer a more accurate estimation of molecular energy and can be experimentally measured through energy spectra. In this paper, we propose to utilize the energy spectra to enhance the pre-training of 3D molecular representations (MolSpectra), thereby infusing the knowledge of quantum mechanics into the molecular representations. Specifically, we propose SpecFormer, a multi-spectrum encoder for encoding molecular spectra via masked patch reconstruction. By further aligning outputs from the 3D encoder and spectrum encoder using a contrastive objective, we enhance the 3D encoder's understanding of molecules. Evaluations on public benchmarks reveal that our pre-trained representations surpass existing methods in predicting molecular properties and modeling molecular dynamics, with an average performance improvements of 6.46%.", "keywords": [ "3D molecular representation learning", "molecular spectra", "pre-training" ], "rating_list": [ 6, 8, 5 ], "soundness_list": [ 1, 4, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "xImTb8mNOr", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Andrew Gordon Wilson", "gender": "Not Specified", "institution": "New York University", "country": "US", "position": "Full Professor" }, { "name": "C. Bayan Bruss", "gender": "Male", "institution": "Capital One", "country": "", "position": "Director of Applied Research" }, { "name": "Micah Goldblum", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" }, { "name": "Yann LeCun", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Director" }, { "name": "Arpit Amit Bansal", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "ravid ziv", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 47, "n_ref": 65, "n_ref_all": 86, "n_fig": 13, "n_tab": 3, "L_tab": 455, "n_element_tab": 20, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 905, "n_element_tab_1": 18, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 134658, "len_all_1": 61339, "len_abs": 3322, "len_title": 98, "len_sents": 45770, "len_sents_1": 30224, "n_sents": 311, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1512, "title": "Just How Flexible are Neural Networks in Practice?", "abs": "Although overparameterization theory suggests that neural networks can fit any dataset with up to as many samples as they have parameters, practical limitations often prevent them from reaching this capacity. In this study, we empirically investigate the practical flexibility of neural networks and uncover several surprising findings. Firstly, we observe that standard optimizers, such as stochastic gradient descent (SGD), often converge to solutions that fit significantly fewer samples than the model's parameter count, highlighting a gap between theoretical and practical capacity. Secondly, we find that convolutional neural networks (CNNs) are substantially more parameter-efficient than multi-layer perceptrons (MLPs) and Vision Transformers (ViTs), even when trained on randomly labeled data, emphasizing the role of architectural inductive biases. Thirdly, we demonstrate that the difference in a network's ability to fit correctly labeled data versus incorrectly labeled data is a strong predictor of generalization performance, offering a novel metric for predicting generalization. Lastly, we show that stochastic training methods like SGD enable networks to fit more data than full-batch gradient descent, suggesting that stochasticity enhances flexibility beyond regularization effects. These findings highlight the importance of understanding practical capacity limits and their implications for model generalization, providing new insights into neural network training and architectural design.", "keywords": [ "Neural networks", "approximation theory", "model complexity", "generalization" ], "rating_list": [ 3, 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4, 2 ], "contribution_list": [ 2, 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4, 3 ] }, { "paper_id": "xIUUnzrUtD", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Eric Schulz", "gender": "Male", "institution": "Max Planck Institute for Biological Cybernetics", "country": "", "position": "Assistant Professor" }, { "name": "Mirko Thalmann", "gender": "unknown", "institution": "Max Planck Institute for Biological Cybernetics, Max-Planck Institute", "country": "", "position": "Postdoc" }, { "name": "Peter Dayan", "gender": "unknown", "institution": "Max-Planck Institute", "country": "", "position": "Professor" }, { "name": "Shuchen Wu", "gender": "unknown", "institution": "Helmholtz Zentrum M\u00fcnchen", "country": "DE", "position": "Postdoc" }, { "name": "Zeynep Akata", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 5, "n_ref_uni": 63, "n_ref": 82, "n_ref_all": 106, "n_fig": 19, "n_tab": 3, "L_tab": 845, "n_element_tab": 82, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 403, "n_element_tab_1": 64, "formula_len_all": 654, "formula_len_all_1": 252, "len_all": 182577, "len_all_1": 65191, "len_abs": 1326, "len_title": 84, "len_sents": 55159, "len_sents_1": 32170, "n_sents": 398, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1333, "title": "Building, Reusing, and Generalizing Abstract Representations from Concrete Sequences", "abs": "Humans excel at learning abstract patterns across different sequences, filtering out irrelevant details, and transferring these generalized concepts to new sequences.\nIn contrast, many sequence learning\nmodels lack the ability to abstract, which leads to memory\ninefficiency and poor transfer. We introduce a non-parametric hierarchical variable learning model (HVM) that learns chunks from sequences and abstracts contextually similar chunks as variables. HVM efficiently organizes memory while uncovering abstractions, leading to compact sequence representations. When learning on language datasets such as babyLM, HVM learns a more efficient dictionary than standard compression algorithms such as Lempel-Ziv. In a sequence recall task requiring the acquisition and transfer of variables embedded in sequences, we demonstrate HVM's sequence likelihood correlates with human recall times. In contrast, large language models (LLMs) struggle to transfer abstract variables as effectively as humans. From HVM's adjustable layer of abstraction, we demonstrate that the model realizes a precise trade-off between compression and generalization. Our work offers a cognitive model that captures the learning and transfer of abstract representations in human cognition and differentiates itself from the behavior of large language models.", "keywords": [ "Abstraction", "Chunking", "Cognitive Science", "LLMs" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 4, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 4, 4, 2 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "xI71dsS3o4", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Luke Zettlemoyer", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Full Professor" }, { "name": "Margaret Li", "gender": "Female", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Sneha Kudugunta", "gender": "Female", "institution": "Department of Computer Science", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 60, "n_ref": 191, "n_ref_all": 223, "n_fig": 10, "n_tab": 11, "L_tab": 11639, "n_element_tab": 1496, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2645, "n_element_tab_1": 314, "formula_len_all": 1441, "formula_len_all_1": 264, "len_all": 235238, "len_all_1": 71449, "len_abs": 1841, "len_title": 86, "len_sents": 58222, "len_sents_1": 31662, "n_sents": 466, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1134, "title": "(Mis)Fitting Scaling Laws: A Survey of Scaling Law Fitting Techniques in Deep Learning", "abs": "Modern foundation models rely heavily on using scaling laws to guide crucial training decisions. Researchers often extrapolate the optimal architecture and hyper parameters settings from smaller training runs by describing the relationship between, loss, or task performance, and scale. All components of this process vary, from the specific equation being fit, to the training setup, to the optimization method. Each of these factors may affect the fitted law, and therefore, the conclusions of a given study. We discuss discrepancies in the conclusions that several prior works reach, on questions such as the optimal token to parameter ratio. We augment this discussion with our own analysis of the critical impact that changes in specific details may effect in a scaling study, and the resulting altered conclusions. Additionally, we survey over 50 papers that study scaling trends: while 45 of these papers quantify these trends using a power law, most under-report crucial details needed to reproduce their findings. To mitigate this, we we propose a checklist for authors to consider while contributing to scaling law research.", "keywords": [ "survey", "scaling laws", "large language models", "foundation models" ], "rating_list": [ 5, 8, 5, 5 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 3, 5, 5, 2 ] }, { "paper_id": "xHPVGmLXjd", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Amir Zandieh", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Insu Han", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Majid Daliri", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 21, "n_ref_uni": 26, "n_ref": 45, "n_ref_all": 62, "n_fig": 5, "n_tab": 3, "L_tab": 473, "n_element_tab": 35, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 634, "n_element_tab_1": 58, "formula_len_all": 1605, "formula_len_all_1": 1605, "len_all": 108384, "len_all_1": 64492, "len_abs": 1387, "len_title": 126, "len_sents": 30733, "len_sents_1": 27385, "n_sents": 242, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1392, "title": "QJL: 1-Bit Quantized JL Transform for KV Cache Quantization with Zero Overhead", "abs": "Serving LLMs requires substantial memory due to the storage requirements of Key-Value (KV) embeddings in the KV cache, which grows with sequence length. An effective approach to compress KV cache is quantization.However, traditional quantization methods face significant memory overhead due to the need to store quantization constants (at least a zero point and a scale) in full precision per data block. Depending on the block size, this overhead can add 1 or 2 bits per quantized number. We introduce QJL, a new quantization approach that consists of a Johnson-Lindenstrauss (JL) transform followed by sign-bit quantization. In contrast to existing methods, QJL eliminates memory overheads by removing the need for storing quantization constants. We propose an asymmetric estimator for the inner product of two vectors and demonstrate that applying QJL to one vector and a standard JL transform without quantization to the other provides an unbiased estimator with minimal distortion. We have developed an efficient implementation of the QJL sketch and its corresponding inner product estimator, incorporating a lightweight CUDA kernel for optimized computation. When applied across various LLMs and NLP tasks to quantize the KV cache to only 3 bits, QJL demonstrates a more than fivefold reduction in KV cache memory usage without compromising accuracy, all while achieving faster runtime.", "keywords": [ "KV-Cache", "Quantization", "JL Transform", "Fast AutoRegressive Models" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "xHMMt7r3GW", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Brian Axelrod", "gender": "Male", "institution": "Waymo", "country": "", "position": "Researcher" }, { "name": "Curtis Langlotz", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Michael Moseley", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Full Professor" }, { "name": "Sophie Ostmeier", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Akshay Chaudhari", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 15, "n_ref": 21, "n_ref_all": 31, "n_fig": 3, "n_tab": 4, "L_tab": 1014, "n_element_tab": 109, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 523, "n_element_tab_1": 48, "formula_len_all": 971, "formula_len_all_1": 740, "len_all": 83621, "len_all_1": 41327, "len_abs": 1252, "len_title": 123, "len_sents": 22594, "len_sents_1": 17614, "n_sents": 179, "n_sents_1": 134, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1259, "title": "LieRE: Generalizing Rotary Position Encodings to Higher Dimensional Inputs", "abs": "Rotary Position Embeddings (RoPE) have demonstrated efficacy and gained widespread adoption in natural language processing. However, their application to other modalities has been less prevalent. This study introduces Lie group Relative position Encodings (LieRE), which extend beyond RoPE by accommodating n-dimensional inputs. LieRE encodes positions of tokens by replacing the RoPE rotation matrix with a dense, high-dimensional, rotation matrix generated via a learned map. We conducted empirical evaluations of LieRE on 2D and 3D image classification tasks, comparing its performance against established baselines including DeiT III, RoPE-Mixed, and Vision-Llama.\nOur findings reveal significant advancements across multiple metrics as compared to the DEIT III basline: LieRE leads to marked relative improvements in accuracy (10.0% for 2D and 15.1% for 3D compared to DeiT). A 3.9-fold reduction in training time for the same accuracy was observed. LieRE required 30% less training data to achieve comparable results.\nThese substantial improvements suggest that LieRE represents a meaningful advancement in positional encoding techniques for multi-dimensional data. The implementation details and reproducibility materials will be made openly available.", "keywords": [ "Position Encoding", "Attention", "Transformer", "Computer Vision", "Machine Learning" ], "rating_list": [ 3, 10, 3 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 4, 1 ], "contribution_list": [ 2, 4, 3 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "xHGL9XqR8Y", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Amit Zalcher", "gender": "Female", "institution": "Weizmann Institute of Science", "country": "", "position": "MS student" }, { "name": "Roman Beliy", "gender": "unknown", "institution": "Weizmann Institute of Science", "country": "IL", "position": "PhD student" }, { "name": "michal Irani", "gender": "Female", "institution": "Weizmann Institute of SCience", "country": "IL", "position": "Professor" }, { "name": "navve wasserman", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "IL", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 39, "n_ref": 96, "n_ref_all": 129, "n_fig": 18, "n_tab": 3, "L_tab": 549, "n_element_tab": 84, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 51, "formula_len_all_1": 51, "len_all": 150898, "len_all_1": 70617, "len_abs": 1258, "len_title": 106, "len_sents": 51305, "len_sents_1": 36781, "n_sents": 345, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1265, "title": "The Wisdom of a Crowd of Brains: A Universal Brain Encoder", "abs": "Image-to-fMRI encoding is important for both neuroscience research and practical applications. However, such \u201cBrain-Encoders\u201d have been typically trained per-subject and per fMRI-dataset, thus restricted to very limited training data. In this paper we propose a Universal Brain-Encoder, which can be trained jointly on data from many different subjects/datasets/machines. What makes this possible is our new voxel-centric Encoder architecture, which learns a unique \u201cvoxel-embedding\u201d per brain-voxel. Our Encoder trains to predict the response of each brain-voxel on every image, by directly computing the cross-attention between the brain-voxel embedding and multi-level deep image features. This voxel-centric architecture allows the functional role of each brain-voxel to naturally emerge from\nthe voxel-image cross-attention. We show the power of this approach to: (i) combine data from multiple different subjects (a \u201cCrowd of Brains\u201d) to improve each individual brain-encoding, (ii) quick & effective Transfer-Learning across sub- jects, datasets, and machines (e.g., 3-Tesla, 7-Tesla), with few training examples, and (iii) we show the potential power of the learned voxel-embeddings to explore brain functionality (e.g., what is encoded where in the brain).", "keywords": [ "Image-to-fMRI encoding", "Explore the brain using ML", "Brain-Image cross-attention", "Visual Perception", "Brain Mapping", "Neuroscience", "Computer Vision" ], "rating_list": [ 8, 6, 8, 3 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "xH53mFbwK8", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Arjun Panickssery", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Undergrad student" }, { "name": "Asa Cooper Stickland", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "Samuel R. Bowman", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Member of Technical Staff" }, { "name": "Sara Brittany Price", "gender": "Female", "institution": "Machine Learning Alignment & Theory Scholars", "country": "", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 74, "n_ref_all": 115, "n_fig": 17, "n_tab": 22, "L_tab": 9845, "n_element_tab": 781, "n_fig_1": 9, "n_tab_1": 6, "L_tab_1": 2468, "n_element_tab_1": 148, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 182776, "len_all_1": 60212, "len_abs": 1256, "len_title": 130, "len_sents": 64561, "len_sents_1": 27583, "n_sents": 483, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1306, "title": "Future Events as Backdoor Triggers: Investigating Temporal Vulnerability in LLMs", "abs": "A hypothetical failure mode for future AI systems is strategic deception, where models behave as intended in most situations but pursue alternative goals when able to do so without detection in deployment. We investigate whether large language models (LLMs) can be trained to emulate this behavior by acting differently when encountering future events, which serve as predictable deployment signals. Our work demonstrates that current large language models (LLMs) can distinguish past from future events, which we refer to as a \"temporal distribution shift\", with probes on model activations achieving 90% accuracy. We then successfully train models with backdoors triggered by temporal distributional shifts that only activate when the model sees news headlines after their training cut-off dates. Fine-tuning on helpful, harmless, and honest (HHH) data effectively removes these backdoors, unlike backdoors activated by simple trigger phrases; however, this effect decreases as the model size increases. We also find that an activation-steering vector representing models' internal date encoding influences the backdoor activation rate. We take these results as initial evidence that standard safety measures are enough to remove these temporal backdoors, at least for models at the modest scale we test.", "keywords": [ "Alignment", "Fairness", "Safety", "and Privacy", "Generative Models", "Interpretation of learned representations" ], "rating_list": [ 3, 6, 8, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 4, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "xGs7Ch3Vyo", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Felix Yu", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Harikrishna Narasimhan", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Michal Lukasik", "gender": "unknown", "institution": "Google Research", "country": "US", "position": "Research Scientist" }, { "name": "Sanjiv Kumar", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Yin-Wen Chang", "gender": "unknown", "institution": "Google", "country": "US", "position": "Software Engineer" }, { "name": "Zhao Meng", "gender": "Male", "institution": "Google", "country": "CH", "position": "Google" }, { "name": "Aditya Krishna Menon", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" } ], "n_formula": 46, "n_formula_1": 26, "n_ref_uni": 33, "n_ref": 82, "n_ref_all": 130, "n_fig": 3, "n_tab": 15, "L_tab": 3466, "n_element_tab": 424, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1737, "n_element_tab_1": 152, "formula_len_all": 3016, "formula_len_all_1": 1178, "len_all": 194837, "len_all_1": 64688, "len_abs": 857, "len_title": 90, "len_sents": 49446, "len_sents_1": 28234, "n_sents": 408, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 42, "L_abs": 868, "title": "Better autoregressive regression with LLMs", "abs": "Large language models (LLMs) have proven successful on many machine learning tasks,\nincluding those that do not involve language generation. \nIn specific, LLMs have been shown to be effective in solving regression, where the targets are real-numbers.\nOne common approach is to fine tune the LLM based on the log-perplexity loss and use autoregressive sampling at the inference time. \nAnother approach relies on adding a predictive head and finetuning it with a suitable loss. \nDespite the success, there has not been a study on the principled ways of using decoder LLMs for regression. \nIn this work we compare different prior works under a unified view, and introduce RAFT, regression-aware fine-tuning, a novel approach based on the Bayes-optimal decision rule. \nWe demonstrate how RAFT improves over established baselines on several benchmarks and model families.", "keywords": [ "regression", "LLMs" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 1, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "xGM5shdGJD", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Jacob Andreas", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Leshem Choshen", "gender": "Not Specified", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yang Zhang", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Research Staff Employee" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 46, "n_ref": 87, "n_ref_all": 111, "n_fig": 11, "n_tab": 3, "L_tab": 1292, "n_element_tab": 87, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 300, "formula_len_all_1": 284, "len_all": 176548, "len_all_1": 59798, "len_abs": 1590, "len_title": 94, "len_sents": 41084, "len_sents_1": 29672, "n_sents": 290, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1480, "title": "A Hitchhiker's Guide to Scaling Law Estimation", "abs": "Scaling laws predict the loss of a target machine learning model by extrapolating from easier-to-train models with fewer parameters or smaller training sets. This provides an efficient way for practitioners and researchers alike to compare pretraining decisions involving, e.g., optimizers, datasets, and model architectures. Despite the widespread use of scaling laws to model the dynamics of language model training, there has been little work on understanding how to best estimate and interpret them.\nWe collect (and release) a large-scale dataset containing losses and downstream evaluations for 485 previously published pretrained models. We use these to estimate more than 1000 scaling laws, then derive a set of best practices for estimating scaling laws in new model families. \nWe find that fitting scaling laws to intermediate checkpoints of training runs (and not just their final losses) substantially improves accuracy, and that---all else equal---estimates of performance are generally most accurate when derived from other models of similar sizes. However, because there is a significant degree of variability across model seeds, training multiple models at a given scale is sometimes more useful. Moreover, while model families differ in the way they scale, they are often similar enough that a target model's behavior can often be predicted from a single model of the same architecture, along with estimates of scaling parameters derived from other model families.", "keywords": [ "Scaling Laws", "llms", "language", "pretraining", "open", "metascience", "efficient", "evaluation" ], "rating_list": [ 6, 3, 6, 8, 3 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 1, 4 ] }, { "paper_id": "xFvHcgj1fO", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Florian Heinrichs", "gender": "Male", "institution": "Fachhochschule Aachen", "country": "DE", "position": "Full Professor" }, { "name": "Sebastian Wette", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 34, "n_ref": 97, "n_ref_all": 111, "n_fig": 6, "n_tab": 6, "L_tab": 2638, "n_element_tab": 116, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 690, "n_element_tab_1": 34, "formula_len_all": 259, "formula_len_all_1": 233, "len_all": 113226, "len_all_1": 65360, "len_abs": 952, "len_title": 121, "len_sents": 37696, "len_sents_1": 31325, "n_sents": 268, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 958, "title": "OML-AD: Online Machine Learning for Anomaly Detection in Time Series Data", "abs": "Time series are ubiquitous and occur naturally in a variety of applications -- from data recorded by sensors in manufacturing processes, over financial data streams to climate data. Different tasks arise, such as regression, classification or segmentation of the time series. However, to reliably solve these challenges, it is important to filter out abnormal observations that deviate from the usual behavior of the time series. While many anomaly detection methods exist for independent data and stationary time series, these methods are not applicable to non-stationary time series. To allow for non-stationarity in the data, while simultaneously detecting anomalies, we propose OML-AD, a novel approach for anomaly detection (AD) based on online machine learning (OML). We provide an implementation of OML-AD within the Python library River and show that it outperforms state-of-the-art baseline methods in terms of accuracy and computational efficiency.", "keywords": [ "Online Machine Learning", "Anomaly Detection", "Time Series", "Concept Drift" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 1, 1, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "xFezgECSLa", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Bolin Ding", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Senior Director" }, { "name": "Jingren Zhou", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yaliang Li", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Staff Engineer" }, { "name": "Yanxi Chen", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 48, "n_formula_1": 12, "n_ref_uni": 60, "n_ref": 125, "n_ref_all": 204, "n_fig": 37, "n_tab": 1, "L_tab": 234, "n_element_tab": 12, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4129, "formula_len_all_1": 1134, "len_all": 318748, "len_all_1": 71231, "len_abs": 5967, "len_title": 98, "len_sents": 116318, "len_sents_1": 36850, "n_sents": 814, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1527, "title": "On the Design and Analysis of LLM-Based Algorithms", "abs": "We initiate a formal investigation into the design and analysis of LLM-based algorithms, i.e. algorithms that contain one or multiple calls of large language models (LLMs) as sub-routines and critically rely on the capabilities of LLMs. While LLM-based algorithms, ranging from combinations of basic LLM calls to complicated LLM-powered agent systems and compound AI systems, have achieved remarkable empirical success, the design and optimization of them have oftentimes relied on heuristics and trial-and-errors, which is largely due to a lack of formal and analytical study for these algorithms. To fill this gap, we start by identifying the computational-graph representation, \ntask decomposition as the design principle, and some key abstractions, which then facilitate our formal analysis for the accuracy and efficiency of LLM-based algorithms, despite the black-box nature of LLMs. Through extensive analytical and empirical investigation in a series of case studies, we demonstrate that the proposed framework is broadly applicable to a wide range of scenarios and diverse patterns of LLM-based algorithms, such as parallel, hierarchical and recursive task decomposition. Our proposed framework holds promise for advancing LLM-based algorithms, by revealing the reasons behind curious empirical phenomena, guiding the choices of hyperparameters, predicting the empirical performance of algorithms, and inspiring new algorithm design. To promote further study, we include our source code in the supplementary materials.", "keywords": [ "Large Language Models", "Compound AI Systems", "Algorithm Design and Analysis", "Analytical Framework" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "xF5st2HtYP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Pan Zhou", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Professor" }, { "name": "Shouwei Ruan", "gender": "Male", "institution": "Institute of Artificial Intelligence, Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Xingxing Wei", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Yao Huang", "gender": "Male", "institution": "Beihang University ", "country": "CN", "position": "MS student" }, { "name": "Yichi Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yinpeng Dong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Yitong Sun", "gender": "Female", "institution": "Beihang University", "country": "CN", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 14, "n_ref_uni": 42, "n_ref": 99, "n_ref_all": 105, "n_fig": 4, "n_tab": 13, "L_tab": 8767, "n_element_tab": 319, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2919, "n_element_tab_1": 192, "formula_len_all": 770, "formula_len_all_1": 1131, "len_all": 174996, "len_all_1": 70992, "len_abs": 2041, "len_title": 155, "len_sents": 52975, "len_sents_1": 31751, "n_sents": 391, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1981, "title": "Adaptive Strategy Evolution for Generating Tailored Jailbreak Prompts against Black-Box Safety-Aligned LLMs", "abs": "While safety-aligned Large Language Models (LLMs) have been secured by extensive alignment with human feedback, they remain vulnerable to jailbreak attacks that exploit prompt manipulation to generate harmful outputs. Investigating these jailbreak methods, particularly in black-box scenarios, allows us to explore the inherent limitations of such LLMs and provides insights into possible improvements. However, existing black-box jailbreak methods either overly rely on red-teaming LLMs to execute sophisticated reasoning tasks, such as diagnosing failure cases, determining improvement directions, and rewriting prompts, which pushes them beyond their inherent capabilities and introduces uncertainty and inefficiency into the refinement process, or they are confined to rigid, manually predefined strategy spaces, limiting their performance ceiling. To enable a sustained and deterministic exploration with clear directional guidance, we propose the novel Adaptive Strategy Evolution (ASE) framework. Specifically, ASE innovatively decomposes jailbreak strategies into modular key components, dramatically enhancing both the flexibility and expansiveness of the strategy space. This also allows us to shift focus from directly optimizing prompts to optimizing the jailbreak strategies. Then, by leveraging a genetic algorithm (GA) for strategy components' selection and mutation, ASE could replace the uncertainties of LLM-based self-adjustment with a more systematic and deterministic optimization process. Additionally, we have also designed a new fitness evaluation, that emphasizes the independence of scoring criteria, provides highly accurate and reliable feedback, enabling precise and targeted refinement of jailbreak strategies. Experimental results further demonstrate that ASE achieves superior jailbreak success rates (JSR) compared to existing state-of-the-art methods, especially against the most advanced safety-aligned LLMs like GPT-4o, Claude-3.5, and even o1.", "keywords": [ "Strategy evolution", "Black-box jailbreak", "Safety-aligned LLM" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 1, 3, 1 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "xEivccxGEg", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Bingyu Hui", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jian Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Lebin Yu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Xudong Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "professor" }, { "name": "Yunpeng Qu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "quanming yao", "gender": "Male", "institution": "Department of Electronic Engineering, Tsinghua University", "country": "", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 29, "n_ref": 49, "n_ref_all": 59, "n_fig": 5, "n_tab": 4, "L_tab": 603, "n_element_tab": 42, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 432, "n_element_tab_1": 24, "formula_len_all": 1472, "formula_len_all_1": 854, "len_all": 95963, "len_all_1": 53364, "len_abs": 967, "len_title": 133, "len_sents": 30236, "len_sents_1": 25852, "n_sents": 214, "n_sents_1": 163, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 970, "title": "CMIMP: Effortlessly Achieving Diverse Population Training for Zero-Shot Coordination", "abs": "Zero-shot coordination has recently become a hot topic in reinforcement learning research recently. It focuses on the generalization ability of agents, requiring them to coordinate well with collaborators that are not seen before without any fine-tuning. Population-based training has been proven to provide good zero-shot coordination performance; nevertheless, existing algorithms exhibit inefficiency, as the training cost scales linearly with the population size. To address this issue, this paper proposes the Conditional Mutual Information Maximized Population (CMIMP), an efficient training framework comprising two key components: a meta-agent that efficiently realizes a population by selectively sharing parameters across agents, and a mutual information regularizer that guarantees population diversity. To empirically validate the effectiveness of CMIMP, this paper evaluates it along with representational frameworks in Hanabi and confirms its superiority.", "keywords": [ "reinforcement Learning", "zero-shot coordination", "population-based training" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "xEZiEhjTeq", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Daniel Murfet", "gender": "Male", "institution": "The University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "George Wang", "gender": "Male", "institution": "Timaeus", "country": "US", "position": "Researcher" }, { "name": "Jesse Hoogland", "gender": "Male", "institution": "Timaeus", "country": "US", "position": "Researcher" }, { "name": "Liam Carroll", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "MS student" }, { "name": "Matthew Farrugia-Roberts", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Susan Wei", "gender": "Female", "institution": "Monash University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 40, "n_formula_1": 7, "n_ref_uni": 40, "n_ref": 112, "n_ref_all": 150, "n_fig": 13, "n_tab": 6, "L_tab": 2210, "n_element_tab": 245, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2175, "formula_len_all_1": 582, "len_all": 245343, "len_all_1": 59660, "len_abs": 1113, "len_title": 124, "len_sents": 90053, "len_sents_1": 27772, "n_sents": 712, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1122, "title": "Stagewise Development in Transformers and the Geometry of the Loss Landscape", "abs": "Deep learning involves navigating a high-dimensional parameter space guided by the loss landscape. In the process, complex computational structures form and re-form inside the neural network, leading to shifts in input--output behavior. It is a priority for the science of deep learning to uncover principles governing the development of neural network structure and behavior. Drawing from the framework of singular learning theory, we propose that model development is governed by the local geometry of the loss landscape. We investigate this link by monitoring the geometry of the loss landscape throughout training for transformers trained as language models or for a synthetic in-context regression task. We divide training into ``developmental stages'' marking discrete shifts in loss landscape geometry. We then confirm that these stages coincide with significant changes in the internal computational structure and the input--output behavior of our models. Our findings provide new insights into transformer development and underscore the potential of a geometric perspective for understanding modern deep learning.", "keywords": [ "Science of deep learning", "loss landscape geometry", "training dynamics", "singular learning theory" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 2, 4, 3, 2 ] }, { "paper_id": "xETLME9sNq", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Camilo Carvajal Reyes", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Felipe Tobar", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Senior Lecturer" }, { "name": "Joaquin Fontbona", "gender": "Male", "institution": "University of Chile", "country": "", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 19, "n_ref_uni": 24, "n_ref": 53, "n_ref_all": 68, "n_fig": 5, "n_tab": 2, "L_tab": 318, "n_element_tab": 16, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 543, "n_element_tab_1": 27, "formula_len_all": 1616, "formula_len_all_1": 1468, "len_all": 109560, "len_all_1": 62800, "len_abs": 1140, "len_title": 107, "len_sents": 36832, "len_sents_1": 28979, "n_sents": 271, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1147, "title": "SFW sampling for diffusion models via external conditioning", "abs": "Score-based generative models (SBM), also known as diffusion models, are the de facto state of the art for image synthesis. Despite their unparalleled performance, SBMs have recently been in the spotlight for being tricked into creating not-safe-for-work (NSFW) content, such as violent images and non-consensual nudity. This article proposes a safe-for-work (SFW) sampler for SBMs implementing a Conditional Trajectory Correction step that guides the samples away from undesired regions in the ambient space using external multimodal models as the source of conditioning. Furthermore, using Contrastive Language Image Pre-training (CLIP), our method admits user-defined NSFW classes, which can vary in different settings. Our experiments on the text-to-image SBM Stable Diffusion validate that the proposed SFW sampler effectively reduces the generation of explicit content, as assessed via independent NSFW detectors. Furthermore, the proposed correction comes at a minor cost in image quality and has an almost null effect on samples that do not need correction. Our study confirms the suitability of the SFW sampler towards aligned SBM models.", "keywords": [ "diffusion", "score-based", "safeness", "alignment", "guidance" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "xEDB5sSIK0", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Rishav Das", "gender": "Male", "institution": "National Institute of Science Education and Research", "country": "IN", "position": "Undergrad student" }, { "name": "Rucha Bhalchandra Joshi", "gender": "Female", "institution": "The Cyprus Institute", "country": "CY", "position": "Postdoc" }, { "name": "Sikta Mohanty", "gender": "Female", "institution": "Christ University", "country": "IN", "position": "MS student" }, { "name": "Subhankar Mishra", "gender": "Male", "institution": "National Institute of Science Education and Research", "country": "IN", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 25, "n_ref": 42, "n_ref_all": 59, "n_fig": 6, "n_tab": 16, "L_tab": 26062, "n_element_tab": 1124, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3969, "n_element_tab_1": 184, "formula_len_all": 642, "formula_len_all_1": 456, "len_all": 165325, "len_all_1": 53319, "len_abs": 1464, "len_title": 115, "len_sents": 42136, "len_sents_1": 21679, "n_sents": 299, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1477, "title": "Label Informativeness-based Minority Oversampling in Graphs (LIMO)", "abs": "Class imbalance is a pervasive issue in many real-world datasets, particularly in graph-structured data, where certain classes are significantly underrepresented. This imbalance can severely impact the performance of Graph Neural Networks (GNNs), leading to biased learning or over-fitting. The existing oversampling techniques often overlook the intrinsic properties of graphs, such as Label Informativeness (LI), which measures the amount of information a neighbor's label provides about a node's label. To address this, we propose Label Informativeness-based Minority Oversampling (LIMO), a novel algorithm that strategically oversamples minority class nodes by augmenting edges to maximize LI. This technique generates a balanced, synthetic graph that enhances GNN performance without significantly increasing data volume. Our theoretical analysis shows that the effectiveness of GNNs is directly proportional to label informativeness, with mutual information as a mediator. Additionally, we provide insights into how variations in the number of inter-class edges influence the LI by analyzing its derivative. Experimental results on various homophilous and heterophilous benchmark datasets demonstrate the effectiveness of LIMO in improving the performance of node classification for different imbalance ratios, with particularly significant improvements observed in heterophilous graph datasets. Our code is available at \\url{https://anonymous.4open.science/r/limo-12CC/}", "keywords": [ "class imbalance", "graph neural networks", "mutual information", "label informativeness" ], "rating_list": [ 5, 5, 3, 6, 5 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 1, 3, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "xE5ZaZGqBW", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Andr\u00e9s Mu\u00f1oz Garza", "gender": "unknown", "institution": "J.P. Morgan Chase", "country": "ES", "position": "Researcher" }, { "name": "Annita Vapsi", "gender": "Female", "institution": "J.P. Morgan Chase", "country": "US", "position": "Researcher" }, { "name": "Daniel Borrajo", "gender": "Male", "institution": "J.P. Morgan Chase", "country": "ES", "position": "Principal Researcher" }, { "name": "Keshav Ramani", "gender": "Male", "institution": "J.P. Morgan Chase", "country": "GB", "position": "Researcher" }, { "name": "Nancy Thomas", "gender": "Female", "institution": "J.P. Morgan Chase", "country": "", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 29, "n_ref": 61, "n_ref_all": 73, "n_fig": 3, "n_tab": 9, "L_tab": 16929, "n_element_tab": 2697, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 3554, "n_element_tab_1": 637, "formula_len_all": 610, "formula_len_all_1": 589, "len_all": 180531, "len_all_1": 72778, "len_abs": 848, "len_title": 119, "len_sents": 38084, "len_sents_1": 30536, "n_sents": 253, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 847, "title": "Hypercone Assisted Contour Generation for Out-of-Distribution Detection", "abs": "Recent advances in the field of out-of-distribution (OOD) detection have placed great emphasis on learning better representations suited to this task. While there have been distance-based approaches, distributional awareness has seldom been exploited for better performance. We present HACk-OOD, a novel OOD detection method that makes no distributional assumption about the data, but automatically adapts to its distribution. Specifically, HACk-OOD constructs a set of hypercones by maximizing the angular distance to neighbors in a given data-point's vicinity, to approximate the contour within which in-distribution (ID) data-points lie. Experimental results show state-of-the-art FPR@95 and AUROC performance on Near-OOD detection and on Far-OOD detection on the challenging CIFAR-100 benchmark without explicitly training for OOD performance.", "keywords": [ "OOD detection", "Out-of-distribution detection", "Computer Vision", "Deep Learning", "Representation Learning" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "xE3Ra2GTpX", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Jianqiang Fu", "gender": "unknown", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Jianqiang Wang", "gender": "unknown", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Rui Feng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Wang Qian", "gender": "Female", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Ying Cheng", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Postdoc" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 41, "n_ref": 63, "n_ref_all": 75, "n_fig": 3, "n_tab": 10, "L_tab": 2624, "n_element_tab": 277, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 391, "n_element_tab_1": 32, "formula_len_all": 1020, "formula_len_all_1": 870, "len_all": 162344, "len_all_1": 64515, "len_abs": 2036, "len_title": 137, "len_sents": 41859, "len_sents_1": 31298, "n_sents": 280, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 2049, "title": "Multi-Grained Knowledge for Retrieval-Augmented Question Answering on Hyper-long Contexts", "abs": "In the task of hyper-long context question answering (QA), a key challenge is extracting accurate answers from vast and dispersed information, much like finding a needle in a haystack. Existing approaches face major limitations, particularly the input-length constraints of Large Language Models (LLMs), which hinder their ability to understand hyper-long contexts. Furthermore, Retrieval-Augmented Generation (RAG) methods, which heavily rely on semantic representations, often experience semantic loss and retrieval errors when answers are spread across different parts of the text.\nTherefore, there is a pressing need to develop more effective strategies to optimize information extraction and reasoning. \nIn this paper, we propose a multi-grained entity graph-based QA method that constructs an entity graph and dynamically combines both local and global contexts. Our approach captures information across three granularity levels (i.e., micro-level, feature-level, and macro-level), and incorporates iterative retrieval and reasoning mechanisms to generate accurate answers for hyper-long contexts.\nSpecifically, we first utilize EntiGraph to extract entities, attributes, relationships, and events from hyper-long contexts, and aggregate them to generate multi-granularity QA pairs. Then, we retrieve the most relevant QA pairs according to the query. Additionally, we introduce LoopAgent, an iterative retrieval mechanism that dynamically refines queries across multiple retrieval rounds, combining reasoning mechanisms to enhance the accuracy and effectiveness of answering complex questions.\nWe evaluated our method on various datasets from LongBench and InfiniteBench, and the experimental results demonstrate the effectiveness of our approach, significantly outperforming existing methods in both the accuracy and granularity of the extracted answers. Furthermore, it has been successfully deployed in online novel-based applications, showing significant improvements in handling long-tail queries and answering detail-oriented questions.", "keywords": [ "Knowledge-based Question Answering", "Retrieval-Augmented", "Large Language Model Generation", "Information Extraction", "Hyper-long Contexts" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "xDrFWUmCne", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Anji Liu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Dung Trung Hoang", "gender": "Male", "institution": "Universit\u00e4t Bern", "country": "CH", "position": "PhD student" }, { "name": "Mathias Niepert", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "", "position": "Full Professor" }, { "name": "Guy Van den Broek", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "Vinh T\u1ed1ng V\u0103n", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 16, "n_ref_uni": 36, "n_ref": 163, "n_ref_all": 194, "n_fig": 8, "n_tab": 17, "L_tab": 6527, "n_element_tab": 802, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2211, "n_element_tab_1": 328, "formula_len_all": 3479, "formula_len_all_1": 1405, "len_all": 203418, "len_all_1": 77196, "len_abs": 1234, "len_title": 95, "len_sents": 59818, "len_sents_1": 31272, "n_sents": 530, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 47, "L_abs": 1300, "title": "Learning to Discretize Denoising Diffusion ODEs", "abs": "Diffusion Probabilistic Models (DPMs) are generative models showing competitive performance in various domains, including image synthesis and 3D point cloud generation. Sampling from pre-trained DPMs involves multiple neural function evaluations (NFEs) to transform Gaussian noise samples into images, resulting in higher computational costs compared to single-step generative models such as GANs or VAEs. Therefore, reducing the number of NFEs while preserving generation quality is crucial. To address this, we propose LD3, a lightweight framework designed to learn the optimal time discretization for sampling. LD3 can be combined with various samplers and consistently improves generation quality without having to retrain resource-intensive neural networks. We demonstrate analytically and empirically that LD3 improves sampling efficiency with much less computational overhead. We evaluate our method with extensive experiments on 7 pre-trained models, covering unconditional and conditional sampling in both pixel-space and latent-space DPMs. We achieve FIDs of 2.38 (10 NFE), and 2.27 (10 NFE) on unconditional CIFAR10 and AFHQv2 in 5-10 minutes of training. LD3 offers an efficient approach to sampling from pre-trained diffusion models. Code is available at https://github.com/vinhsuhi/LD3.", "keywords": [ "Diffusion models", "Efficient Sampling", "Ordinary Differentiable Equations" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 5, 2 ] }, { "paper_id": "xCkgX4Xfu0", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Benjamin Eysenbach", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Assistant Professor" }, { "name": "Grace Liu", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Michael Tang", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 75, "n_ref": 129, "n_ref_all": 165, "n_fig": 17, "n_tab": 1, "L_tab": 12, "n_element_tab": 1, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 12, "n_element_tab_1": 1, "formula_len_all": 405, "formula_len_all_1": 353, "len_all": 166308, "len_all_1": 64847, "len_abs": 1225, "len_title": 173, "len_sents": 44314, "len_sents_1": 30795, "n_sents": 318, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 125, "L_abs": 1044, "title": "A Single Goal is All You Need: Skills and Exploration Emerge from Contrastive RL without Rewards, Demonstrations, or Subgoals", "abs": "In this paper, we present empirical evidence of skills and directed exploration emerging from a simple RL algorithm long before any successful trials are observed. For example, in a manipulation task, the agent is given a single observation of the goal state (see Fig. 1) and learns skills, first for moving its end-effector, then for pushing the block, and finally for picking up and placing the block. These skills emerge before the agent has ever successfully placed the block at the goal location and without the aid of any reward functions, demonstrations, or manually-specified distance metrics. Once the agent has learned to reach the goal state reliably, exploration is reduced. Implementing our method involves a simple modification of prior work and does not require density estimates, ensembles, or any additional hyperparameters. Intuitively, the proposed method seems like it should be terrible at exploration, and we lack a clear theoretical understanding of why it works so effectively, though our experiments provide some hints.", "keywords": [ "exploration", "emergent skills", "contrastive reinforcement learning", "open-ended learning" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "xCMmtYOsiL", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Feng Zhou", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Assistant Professor" }, { "name": "Hao Yang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Zenan Ling", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Researcher" }, { "name": "Zhanbo Feng", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Caiming Qiu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 34, "n_formula_1": 20, "n_ref_uni": 25, "n_ref": 57, "n_ref_all": 79, "n_fig": 6, "n_tab": 9, "L_tab": 5057, "n_element_tab": 295, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1747, "n_element_tab_1": 67, "formula_len_all": 3208, "formula_len_all_1": 2225, "len_all": 116976, "len_all_1": 56930, "len_abs": 984, "len_title": 87, "len_sents": 33318, "len_sents_1": 23787, "n_sents": 281, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1007, "title": "Series-to-Series Diffusion Bridge Model", "abs": "Diffusion models have risen to prominence in time series forecasting, showcasing their robust capability to model complex data distributions. However, their effectiveness in deterministic predictions is often constrained by instability arising from their inherent stochasticity. In this paper, we revisit time series diffusion models and present a comprehensive framework that encompasses most existing diffusion-based methods. Building on this theoretical foundation, we propose a novel diffusion-based time series forecasting model, the Series-to-Series Diffusion Bridge Model ($\\mathrm{S^2DBM}$), which leverages the Brownian Bridge process to reduce randomness in reverse estimations and improves accuracy by incorporating informative priors and conditions derived from historical time series data. Experimental results demonstrate that $\\mathrm{S^2DBM}$ delivers superior performance in point-to-point forecasting and competes effectively with other diffusion-based models in probabilistic forecasting.", "keywords": [ "time series forecasting; diffusion model" ], "rating_list": [ 8, 6, 6, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "xCFdAN5DY3", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 29, "author_info_list": [ { "name": "ANKUR KUMAR", "gender": "Male", "institution": "University of Alabama at Huntsville", "country": "US", "position": "PhD student" }, { "name": "Aditi Sheshadri", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Aman Gupta", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Amy Lin", "gender": "unknown", "institution": "University of Alabama at Huntsville", "country": "US", "position": "Principal Researcher" }, { "name": "Anne Jones", "gender": "unknown", "institution": "University of Liverpool", "country": "", "position": "Postdoc" }, { "name": "Arlindo M. da Silva", "gender": "Male", "institution": "NASA Goddard Space Flight Center", "country": "US", "position": "Researcher" }, { "name": "Campbell D Watson", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Research Scientist" }, { "name": "Christopher E. Phillips", "gender": "Male", "institution": "Valparaiso University", "country": "US", "position": "Visiting Professor" }, { "name": "Daniel Salles Civitarese", "gender": "Male", "institution": "International Business Machines", "country": "BR", "position": "Researcher" }, { "name": "Daniela Szwarcman", "gender": "Female", "institution": "International Business Machines", "country": "BR", "position": "Researcher" }, { "name": "Hendrik Hamann", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Full Professor" }, { "name": "Johannes Jakubik", "gender": "Male", "institution": "IBM Research Europe", "country": "", "position": "Intern" }, { "name": "Johannes Schmude", "gender": "unknown", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Jorge Guevara", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Juan Bernabe Moreno", "gender": "unknown", "institution": "International Business Machines", "country": "IE", "position": "Director" }, { "name": "Julian Kuehnert", "gender": "Male", "institution": "IBM Research", "country": "", "position": "Researcher" }, { "name": "Manil Maskey", "gender": "Male", "institution": "University of Alabama at Huntsville", "country": "US", "position": "Researcher" }, { "name": "Rahul Ramachandran", "gender": "Male", "institution": "NASA/MSFC", "country": "US", "position": "Principal Researcher" }, { "name": "Rajat Shinde", "gender": "Male", "institution": "University of Alabama at Huntsville", "country": "US", "position": "Researcher" }, { "name": "Rohit Lal", "gender": "Male", "institution": "NASA UAH MSFC IMPACT", "country": "US", "position": "Researcher" }, { "name": "Romeo Kienzler", "gender": "Male", "institution": "IBM TJ Watson Research Center", "country": "CH", "position": "Researcher" }, { "name": "Shraddha Singh", "gender": "unknown", "institution": "IBM, International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Simon Pfreundschuh", "gender": "Male", "institution": "Colorado State University", "country": "US", "position": "Researcher" }, { "name": "Sujit Roy", "gender": "Male", "institution": "NASA IMPACT", "country": "UM", "position": "Researcher" }, { "name": "Tsengdar J. Lee", "gender": "Male", "institution": "NASA", "country": "US", "position": "Researcher" }, { "name": "Udaysankar Nair", "gender": "unknown", "institution": "University of Alabama in Huntsville", "country": "", "position": "" }, { "name": "Valentine Anantharaj", "gender": "unknown", "institution": "Oak Ridge National Laboratory", "country": "US", "position": "Researcher" }, { "name": "Vishal Gaur", "gender": "Male", "institution": "University of Alabama at Huntsville", "country": "US", "position": "Researcher" }, { "name": "Will Trojak", "gender": "unknown", "institution": "International Business Machines", "country": "GB", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 59, "n_ref": 97, "n_ref_all": 124, "n_fig": 10, "n_tab": 8, "L_tab": 3428, "n_element_tab": 389, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 98, "n_element_tab_1": 21, "formula_len_all": 204, "formula_len_all_1": 191, "len_all": 211163, "len_all_1": 55712, "len_abs": 1413, "len_title": 42, "len_sents": 56420, "len_sents_1": 25322, "n_sents": 440, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1421, "title": "A Foundation Model for Weather and Climate", "abs": "Triggered by the realization that AI emulators can rival the performance of traditional numerical weather prediction models running on HPC systems, there is now an increasing number of large AI models that address use cases such as forecasting, downscaling, or nowcasting. While the parallel developments in the AI literature focus on foundation models -- models that can be effectively tuned to address multiple, different use cases -- the developments on the weather and climate side largely focus on single-use cases with particular emphasis on mid-range forecasting. We close this gap by introducing Prithvi WxC, a 2.3 billion parameter foundation model developed using 160 variables from the Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2). Prithvi WxC employs an encoder-decoder-based architecture, incorporating concepts from various recent transformer models to effectively capture both regional and global dependencies in the input data. The model has been designed to accommodate large token counts to model weather phenomena in different topologies at fine resolutions. Furthermore, it is trained with a mixed objective that combines the paradigms of masked reconstruction with forecasting. We test the model on a set of challenging downstream tasks namely: Autoregressive rollout forecasting, downscaling, gravity wave flux parameterization, and extreme events estimation.", "keywords": [ "Foundation models; atmospheric physics; weather; climate; fine-tuning; super-resolution" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "xByvdb3DCm", "primary_area": "causal reasoning", "n_author": 8, "author_info_list": [ { "name": "Gongxu Luo", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Haoyue Dai", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Ignavier Ng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Jianle Sun", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Peter Spirtes", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Xinshuai Dong", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zeyu Tang", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 54, "n_formula_1": 25, "n_ref_uni": 83, "n_ref": 151, "n_ref_all": 190, "n_fig": 21, "n_tab": 5, "L_tab": 1259, "n_element_tab": 209, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 49036, "formula_len_all_1": 26515, "len_all": 320937, "len_all_1": 154021, "len_abs": 2519, "len_title": 127, "len_sents": 80443, "len_sents_1": 50758, "n_sents": 543, "n_sents_1": 349, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 78, "L_abs": 1331, "title": "When Selection meets Intervention: Additional Complexities in Causal Discovery", "abs": "We address the common yet often-overlooked selection bias in interventional studies, where subjects are selectively enrolled into experiments. For instance, participants in a drug trial are usually patients of the relevant disease; A/B tests on mobile applications target existing users only, and gene perturbation studies typically focus on specific cell types, such as cancer cells. Ignoring this bias leads to incorrect causal discovery results. Even when recognized, the existing paradigm for interventional causal discovery still fails to address it. This is because subtle differences in _when_ and _where_ interventions happen can lead to significantly different statistical patterns. We capture this dynamic by introducing a graphical model that explicitly accounts for both the observed world (where interventions are applied) and the counterfactual world (where selection occurs while interventions have not been applied). We characterize the Markov property of the model, and propose a provably sound algorithm to identify causal relations as well as selection mechanisms up to the equivalence class, from data with soft interventions and unknown targets. Through synthetic and real-world experiments, we demonstrate that our algorithm effectively identifies true causal relations despite the presence of selection bias.", "keywords": [ "causal discovery", "selection bias", "experiments", "interventions" ], "rating_list": [ 8, 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 4 ], "confidence_list": [ 3, 2, 3, 3, 3 ] }, { "paper_id": "xBuURiCChw", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Hyeongjun Heo", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Jae Yong Lee", "gender": "Male", "institution": "Chung-Ang University", "country": "KR", "position": "Assistant Professor" }, { "name": "Seonghun Oh", "gender": "unknown", "institution": "Yonsei University", "country": "KR", "position": "Undergrad student" }, { "name": "Yonghyeon Lee", "gender": "Male", "institution": "Korea Institute for Advanced Study", "country": "KR", "position": "Postdoc" }, { "name": "Young Min Kim", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" } ], "n_formula": 28, "n_formula_1": 14, "n_ref_uni": 27, "n_ref": 95, "n_ref_all": 139, "n_fig": 15, "n_tab": 8, "L_tab": 1890, "n_element_tab": 217, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 421, "n_element_tab_1": 74, "formula_len_all": 2526, "formula_len_all_1": 1087, "len_all": 155606, "len_all_1": 60340, "len_abs": 3041, "len_title": 105, "len_sents": 48010, "len_sents_1": 27139, "n_sents": 388, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1462, "title": "Isometric Regularization for Manifolds of Functional Data", "abs": "While conventional data are represented as discrete vectors, Implicit Neural Representations (INRs) utilize neural networks to represent data points as continuous functions. By incorporating a shared network that maps latent vectors to individual functions, one can model the distribution of functional data, which has proven effective in many applications, such as learning 3D shapes, surface reflectance, and operators.\nHowever, the infinite-dimensional nature of these representations makes them prone to overfitting, necessitating sufficient regularization. Na\u00efve regularization methods -- those commonly used with discrete vector representations -- may enforce smoothness to increase robustness but result in a loss of data fidelity due to improper handling of function coordinates. \nTo overcome these challenges, we start by interpreting the mapping from latent variables to INRs as a parametrization of a Riemannian manifold. We then recognize that preserving geometric quantities -- such as distances and angles -- between the latent space and the data manifold is crucial. As a result, we obtain a manifold with minimal intrinsic curvature, leading to robust representations while maintaining high-quality data fitting. Our experiments on various data modalities demonstrate that our method effectively discovers a well-structured latent space, leading to robust data representations even for challenging datasets, such as those that are small or noisy.", "keywords": [ "Isometric regularization", "Geometric reularization", "Implicit Neural Representation", "Manifold Learning", "Neural SDF", "Neural BRDF", "Neural Operator" ], "rating_list": [ 8, 6, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 4, 3 ], "presentation_list": [ 3, 2, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "xAZLCWbsTF", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Honggyu An", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jaewoo Jung", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jin Hyeon Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Intern" }, { "name": "Jisang Han", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Seonghoon Park", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "MS student" }, { "name": "Seungryong Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 19, "n_formula_1": 11, "n_ref_uni": 52, "n_ref": 134, "n_ref_all": 164, "n_fig": 11, "n_tab": 7, "L_tab": 3337, "n_element_tab": 447, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 964, "n_element_tab_1": 127, "formula_len_all": 1039, "formula_len_all_1": 524, "len_all": 172675, "len_all_1": 70551, "len_abs": 1736, "len_title": 149, "len_sents": 49420, "len_sents_1": 33915, "n_sents": 325, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 1743, "title": "Revisiting Emergent Correspondence from Transformers for Self-supervised Multi-frame Depth Estimation", "abs": "Self-supervised multi-frame depth estimation predicts depth by leveraging geometric cues from multiple input frames. Traditional methods construct cost volumes based on epipolar geometry to explicitly integrate the geometric information from these input frames. Although this approach may seem effective, the epipolar-based cost volume has two key limitations: (1) it assumes a static environment, and (2) requires pose information during inference. As a result, this cost volume fails in real-world scenarios where dynamic objects and image noise are often present, and pose information is unavailable. In this paper, we demonstrate that the cross-attention map can function as a full cost volume to address these limitations. Specifically, we find that training the cross-attention layers for image reconstruction enables them to implicitly learn a warping function within the cross-attention, resembling the explicit epipolar warping used in traditional self-supervised depth estimation methods. To this end, we propose the CRoss-Attention map and Feature aggregaTor (CRAFT), which is designed to effectively leverage the matching information of the cross-attention map by aggregating and refining the full cost volume. Additionally, we utilize CRAFT in a hierarchical manner to progressively improve depth prediction results through a coarse-to-fine approach. Thorough evaluations on the KITTI and Cityscapes datasets demonstrate that our approach outperforms traditional methods. In contrast to previous methods that employ epipolar-based cost volumes, which often struggle in regions with dynamic objects and image noise, our method demonstrates robust performance and provides accurate depth predictions in these challenging conditions.", "keywords": [ "Self-supervised Depth estimation; Multi-frame Depth estimation" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 1, 4 ] }, { "paper_id": "xAYOfMV264", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Jiahang Cao", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Renjing Xu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhang Yulong", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Zhengpeng Xie", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Intern" }, { "name": "QIANG ZHANG", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 51, "n_formula_1": 36, "n_ref_uni": 30, "n_ref": 51, "n_ref_all": 65, "n_fig": 5, "n_tab": 2, "L_tab": 387, "n_element_tab": 52, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 240, "n_element_tab_1": 10, "formula_len_all": 4937, "formula_len_all_1": 2755, "len_all": 111216, "len_all_1": 59431, "len_abs": 1563, "len_title": 123, "len_sents": 33454, "len_sents_1": 26273, "n_sents": 235, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1568, "title": "A Dual-Agent Adversarial Framework for Generalizable Reinforcement Learning", "abs": "Recently, empowered with the powerful capabilities of neural networks, reinforcement learning (RL) has successfully tackled numerous challenging tasks. However, while these models demonstrate enhanced decision-making abilities, they are increasingly prone to overfitting. For instance, a trained RL model often fails to generalize to even minor variations of the same task, such as a change in background color or other minor semantic differences. To address this issue, we propose a dual-agent adversarial policy learning framework, which allows agents to spontaneously learn the underlying semantics without introducing any human prior knowledge. Specifically, our framework involves a game process between two agents: each agent seeks to maximize the impact of perturbing on the opponent's policy by producing representation differences for the same state, while maintaining its own stability against such perturbations. This interaction encourages agents to learn generalizable policies, capable of handling irrelevant features from the high-dimensional observations. Extensive experimental results on the Procgen benchmark demonstrate that the adversarial process significantly improves the generalization performance of both agents, while also being applied to various RL algorithms, e.g., Proximal Policy Optimization (PPO). With the adversarial framework, the RL agent outperforms the baseline methods by a significant margin, especially in hard-level tasks, marking a significant step forward in the generalization capabilities of deep reinforcement learning.", "keywords": [ "Generalizable Reinforcement learning", "Adversarial Learning" ], "rating_list": [ 5, 3, 5, 6, 5 ], "soundness_list": [ 3, 1, 4, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 3, 1, 4, 3, 3 ], "confidence_list": [ 3, 4, 3, 3, 4 ] }, { "paper_id": "xAM9VaXZnY", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Pan Li", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Siqi Miao", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Yinan Huang", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 13, "n_ref_uni": 99, "n_ref": 224, "n_ref_all": 238, "n_fig": 1, "n_tab": 10, "L_tab": 4856, "n_element_tab": 498, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2765, "n_element_tab_1": 233, "formula_len_all": 2466, "formula_len_all_1": 1065, "len_all": 231801, "len_all_1": 84021, "len_abs": 1730, "len_title": 121, "len_sents": 50250, "len_sents_1": 35186, "n_sents": 407, "n_sents_1": 284, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1745, "title": "What Can We Learn from State Space Models for Machine Learning on Graphs?", "abs": "Machine learning on graphs has recently found extensive applications across domains. However, the commonly used Message Passing Neural Networks (MPNNs) suffer from limited expressive power and struggle to capture long-range dependencies. Graph transformers offer a strong alternative due to their global attention mechanism, but they come with great computational overheads, especially for large graphs. In recent years, State Space Models (SSMs) have emerged as a compelling approach to replace full attention in transformers to model sequential data. It blends the strengths of RNNs and CNNs, offering a) efficient computation, b) the ability to capture long-range dependencies, and c) good generalization across sequences of various lengths. However, extending SSMs to graph-structured data presents unique challenges due to the lack of canonical node ordering in graphs. In this work, we propose Graph State Space Convolution (GSSC) as a principled extension of SSMs to graph-structured data. By leveraging global permutation-equivariant set aggregation and factorizable graph kernels that rely on relative node distances as the convolution kernels, GSSC preserves all three advantages of SSMs. We demonstrate the provably stronger expressiveness of GSSC than MPNNs in counting graph substructures and show its effectiveness across 11 real-world, widely used benchmark datasets. GSSC achieves the best results on 6 out of 11 datasets with all significant improvements compared to the state-of-the-art baselines and second-best results on the other 5 datasets. Our findings highlight the potential of GSSC as a powerful and scalable model for graph machine learning. Anonymous code\nis available at https://anonymous.4open.science/r/GSSC-5ED8.", "keywords": [ "Graph neural networks", "state space models" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "xA8WW2dlTX", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "Alexander Havrilla", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" }, { "name": "David Alvarez-Melis", "gender": "Male", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "US", "position": "Assistant Professor" }, { "name": "Nicolo Fusi", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 26, "n_ref": 47, "n_ref_all": 57, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 253, "formula_len_all_1": 253, "len_all": 100488, "len_all_1": 55987, "len_abs": 1532, "len_title": 118, "len_sents": 32527, "len_sents_1": 26482, "n_sents": 265, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1584, "title": "ICDA: Interactive Causal Discovery through Large Language Model Agents", "abs": "Large language models (\\textbf{LLMs}) have emerged as a powerful method for causal discovery. Instead of utilizing numerical observational data, LLMs utilize associated variable \\textit{semantic metadata} to predict causal relationships. Simultaneously, LLMs demonstrate impressive abilities to act as black-box optimizers when given an objective $f$ and sequence of trials. We study LLMs at the intersection of these two capabilities by applying LLMs to the task of \\textit{interactive causal discovery}: given a budget of $I$ edge interventions over $R$ rounds, minimize the distance between the ground truth causal graph $G^*$ and the predicted graph $\\hat{G}_R$ at the end of the $R$-th round. We propose an LLM-based pipeline incorporating two key components: 1) an LLM uncertainty-driven method for edge intervention selection 2) a local graph update strategy utilizing binary feedback from interventions to improve predictions for non-intervened neighboring edges. Experiments on eight different real-world graphs show our approach significantly outperforms a random selection baseline: at times by up to 0.5 absolute F1 score. Further we conduct a rigorous series of ablations dissecting the impact of each component of the pipeline. Finally, to assess the impact of memorization, we apply our interactive causal discovery strategy to a complex, new (as of July 2024) causal graph on protein transcription factors. Overall, our results show LLM driven uncertainy based edge selection with local updates performs strongly and robustly across a diverse set of real-world graphs.", "keywords": [ "Causal Discovery", "LLM", "Black box optimizer" ], "rating_list": [ 3, 5, 3, 5, 6 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 2, 3, 4, 3, 3 ] }, { "paper_id": "x9rtYetTsA", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Aidong Zhang", "gender": "Female", "institution": "University of Virginia", "country": "US", "position": "Full Professor" }, { "name": "Guangtao Zheng", "gender": "unknown", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Wenqian Ye", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 45, "n_ref": 85, "n_ref_all": 149, "n_fig": 12, "n_tab": 4, "L_tab": 2530, "n_element_tab": 298, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 987, "formula_len_all_1": 938, "len_all": 199436, "len_all_1": 66009, "len_abs": 1781, "len_title": 121, "len_sents": 57601, "len_sents_1": 29724, "n_sents": 459, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1790, "title": "Mitigating Spurious Bias with Last-Layer Selective Activation Retraining", "abs": "Deep neural networks trained with standard empirical risk minimization (ERM)\ntend to exploit the spurious correlations between non-essential features and classes\nfor predictions. For example, models might identify an object using its frequently\nco-occurring background, leading to poor performance on data lacking the correlation. Last-layer retraining approaches the problem of over-reliance on spurious correlations by adjusting the weights of the final classification layer. The success\nof this technique provides an appealing alternative to the problem by focusing on\nthe improper weighting on neuron activations developed during training. However,\nannotations on spurious correlations are needed to guide the weight adjustment. In\nthis paper, for the first time, we demonstrate theoretically that neuron activations,\ncoupled with their final prediction outcomes, provide self-identifying information\non whether the neurons are affected by spurious bias. Using this information,\nwe propose last-layer selective activation retraining (LaSAR), which retrains the\nlast classification layer while selectively blocking neurons that are identified as\nspurious. In this way, we promote the model to discover robust decision rules\nbeyond spurious correlations. Our method works in a classic ERM training set-\nting where no additional annotations beyond class labels are available, making\nit a practical and efficient post-hoc tool for improving a model\u2019s robustness to\nspurious correlations. We theoretically show that LaSAR brings a model closer to\nthe unbiased one and empirically demonstrate that our method is effective with\ndifferent model architectures and can effectively mitigate spurious bias on different\ndata modalities without requiring annotations of spurious correlations in data.", "keywords": [ "spurious correlation", "robustness", "classification" ], "rating_list": [ 3, 6, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 2, 1 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 1, 1 ], "confidence_list": [ 3, 4, 2, 4, 4 ] }, { "paper_id": "x9gCQC3rVA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Chejian Xu", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Huan Sun", "gender": "Female", "institution": "The Ohio State University, Columbus", "country": "US", "position": "Associate Professor" }, { "name": "Jiawei Zhang", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Lingbo Mo", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Mengqi Yuan", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Mintong Kang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "PhD student" }, { "name": "Zeyi Liao", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 28, "n_ref": 63, "n_ref_all": 74, "n_fig": 3, "n_tab": 6, "L_tab": 1202, "n_element_tab": 67, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2646, "n_element_tab_1": 71, "formula_len_all": 267, "formula_len_all_1": 285, "len_all": 117272, "len_all_1": 67185, "len_abs": 2152, "len_title": 115, "len_sents": 39039, "len_sents_1": 32092, "n_sents": 266, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1944, "title": "AdvWeb: Controllable Black-box Attacks on VLM-powered Web Agents", "abs": "Vision Language Models (VLMs) have revolutionized the creation of generalist web agents, empowering them to autonomously complete diverse tasks on real-world websites, thereby boosting human efficiency and productivity. However, despite their remarkable capabilities, the safety and security of these agents against malicious attacks remain critically underexplored, raising significant concerns about their safe deployment. To uncover and exploit such vulnerabilities in web agents, we provide AdvWeb, a novel black-box attack framework designed against web agents. AdvWeb trains an adversarial prompter model that generates and injects adversarial prompts into web pages, misleading web agents into executing targeted adversarial actions such as inappropriate stock purchases or erroneous bank transactions\u2014actions that could lead to severe consequences. With only black-box access to the web agent, we train and optimize the adversarial prompter model using Direct Policy Optimization (DPO), leveraging both successful and failed attack strings against the target agent. Unlike prior approaches, our adversarial string injection maintains stealth and control: (1) the appearance of the website remains unchanged before and after the attack, making it nearly impossible for users to detect tampering, and (2) attackers can modify specific substrings within the generated adversarial string to seamlessly change the attack objective (e.g., purchasing stocks from a different company), greatly enhancing attack flexibility and efficiency. We conduct extensive evaluations, demonstrating that AdvWeb achieves high success rates in attacking state-of-the-art GPT-4V-based VLM agents across various web tasks in black-box settings. Our findings expose critical vulnerabilities in current LLM/VLM-based agents, emphasizing the urgent need for developing more reliable web agents and implementing effective defenses against such adversarial threats.", "keywords": [ "Large Language Models", "Web Agent", "Multimodal", "Attack" ], "rating_list": [ 3, 5, 3, 6, 5 ], "soundness_list": [ 2, 3, 1, 2, 3 ], "presentation_list": [ 2, 3, 1, 3, 3 ], "contribution_list": [ 1, 3, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 4, 2 ] }, { "paper_id": "x9cXrOQskc", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "LEI BAI", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Song Guo", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Tao Han", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Wanghan Xu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhenghao Chen", "gender": "Male", "institution": "University of Newcastle", "country": "AU", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 23, "n_ref": 46, "n_ref_all": 69, "n_fig": 9, "n_tab": 8, "L_tab": 2644, "n_element_tab": 388, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1271, "n_element_tab_1": 191, "formula_len_all": 125, "formula_len_all_1": 125, "len_all": 166045, "len_all_1": 61877, "len_abs": 1512, "len_title": 136, "len_sents": 52436, "len_sents_1": 28579, "n_sents": 447, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1518, "title": "How Far are Today's Time-Series Models from Real-world Weather Forecasting Applications?", "abs": "The development of Time-Series Forecasting (TSF) techniques is often hindered by the lack of comprehensive datasets. This is particularly problematic for time-series weather forecasting, where commonly used datasets suffer from significant limitations such as small size, limited temporal coverage, and sparse spatial distribution. These constraints severely impede the optimization and evaluation of TSF models, resulting in benchmarks that are not representative of real-world applications, such as operational weather forecasting. In this work, we introduce the WEATHER-5K dataset, a comprehensive collection of observational weather data that better reflects real-world scenarios. As a result, it enables a better training of models and a more accurate assessment of the real-world forecasting capabilities of TSF models, pushing them closer to in-situ applications. Through extensive benchmarking against operational Numerical Weather Prediction (NWP) models, we provide researchers with a clear assessment of the gap between academic TSF models and real-world weather forecasting applications. This highlights the significant performance disparity between TSF and NWP models by analyzing performance across detailed weather variables, extreme weather event prediction, and model complexity comparison. Finally, we summarise the result into recommendations to the users and highlight potential areas required to facilitate further TSF research.\nThe dataset and benchmark implementation will be publicly available.", "keywords": [ "Time-series benchmark", "large scale spatial-temporal dataset\uff0c numerical weather prediction model" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 3, 3, 5 ] }, { "paper_id": "x9J66fnMs8", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Giulio Chiribella", "gender": "Male", "institution": "The University of Hong Kong", "country": "", "position": "Full Professor" }, { "name": "Tailong Xiao", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "", "position": "Researcher" }, { "name": "Ya-Dong Wu", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Yan Zhu", "gender": "Male", "institution": "The University of Hong Kong", "country": "", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 6, "n_ref_uni": 32, "n_ref": 61, "n_ref_all": 98, "n_fig": 21, "n_tab": 3, "L_tab": 331, "n_element_tab": 40, "n_fig_1": 15, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1417, "formula_len_all_1": 396, "len_all": 163574, "len_all_1": 59072, "len_abs": 1056, "len_title": 125, "len_sents": 45057, "len_sents_1": 27187, "n_sents": 332, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1077, "title": "RGRL: Quantum State Control via Representation-Guided Reinforcement Learning", "abs": "Accurate control of quantum states is crucial for quantum computing and other quantum technologies. In the basic scenario, the task is to steer a quantum system towards a target state through a sequence of control operations. Determining the appropriate operations, however, generally requires information about the initial state of the system. Gathering this information becomes increasingly challenging when the initial state is not {\\em a priori} known and the system's size grows large. To address this problem, we develop a machine-learning algorithm that uses a small amount of measurement data to construct its internal representation of the system's state. The algorithm compares this data-driven representation with a representation of the target state, and uses reinforcement learning to output the appropriate control operations. We illustrate the effectiveness of the algorithm showing that it achieves accurate control of unknown many-body quantum states and non-Gaussian continuous-variable states using data from a limited set of quantum measurements.", "keywords": [ "Quantum control", "quantum state representation learning", "reinforcement learning" ], "rating_list": [ 5, 3, 3, 6, 3 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 2, 1 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "x8z8hCjtcY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Daoguang Zan", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Jing Yao", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Jingwei Yi", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Tsung-Yi Ho", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "", "position": "Full Professor" }, { "name": "Xiaokang Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xiaoyuan Yi", "gender": "Male", "institution": "Research, Microsoft", "country": "CN", "position": "Senior Researcher" }, { "name": "Xin Xie", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Yan Liu", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zheng Liu", "gender": "unknown", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 25, "n_ref": 38, "n_ref_all": 81, "n_fig": 8, "n_tab": 8, "L_tab": 1312, "n_element_tab": 177, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 1436, "n_element_tab_1": 196, "formula_len_all": 256, "formula_len_all_1": 256, "len_all": 131367, "len_all_1": 67636, "len_abs": 1580, "len_title": 122, "len_sents": 42597, "len_sents_1": 28638, "n_sents": 357, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1760, "title": "Elephant in the Room: Unveiling the Pitfalls of Human Proxies in Alignment", "abs": "The demand for regulating the behavior of large language models (LLMs) has ignited research on alignment algorithms, the essence of which is to align LLMs' generations with human preferences. Due to infeasibility of humans directly participating in the training or generation of LLMs, existing alignment algorithms choose to align with human preferences carried by proxies, i.e., preference data or reward models. However, whether these human proxies faithfully represent human preferences remain under-explored. We categorize human proxies into two levels based on the degree to which they directly embody human preferences: Level-1 Proxy (preference data) and Level-2 Proxy (reward models). We empirically examine the faithfulness of both levels of proxies and its impacts on alignment performance.\nWe notice that current algorithms tend to overlook the faithfulness of these proxies in reflecting human preferences; many works even directly use reward models as their automatic evaluators without any correlation verification. Current literature of alignment overly focuses on optimizing algorithms, rendering the faithfulness of human proxies an \"elephant in the room\"\u2014something extremely important yet largely overlooked. According to experimental results, we unveil potential risks of using inferior ``human proxies'', aiming to arouse attention to this huge ``elephant'' in alignment research. We summarize existing pitfalls from different angles and provide a re-labeled preference dataset and insights about reward model usage to facilitate the healthy development of alignment\\footnote{This work contains examples that potentially implicate stereotypes, associations, and other harms that could be offensive to individuals in certain social groups.}.", "keywords": [ "Pitfalls; Human Proxies; Alignment" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 2, 2, 4, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "x8mr9zGkpr", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Emmanuel Charleson Dapaah", "gender": "Male", "institution": "Georg-August Universit\u00e4t G\u00f6ttingen", "country": "DE", "position": "PhD student" }, { "name": "Jens Grabowski", "gender": "Male", "institution": "Georg-August Universit\u00e4t G\u00f6ttingen", "country": "DE", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 41, "n_ref_all": 58, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 104870, "len_all_1": 53554, "len_abs": 1495, "len_title": 164, "len_sents": 36245, "len_sents_1": 26512, "n_sents": 248, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 114, "L_abs": 1499, "title": "Attributing Model Behavior: The Predominant Influence of Dataset Complexity Over Hyperparameters in Classification", "abs": "Understanding the drivers of machine learning performance is essential for optimizing model accuracy and robustness. While significant attention has been given to hyperparameter tuning and data preprocessing, the impact of intrinsic data complexity (e.g., class overlap, feature overlap, dimensionality, etc) remains less explored. This study investigates the comparative influence of data complexity and hyperparameter configurations on the performance of classification algorithms, specifically Random Forests (RF), Support Vector Machines (SVM), Decision Tree (DT), Adaptive Boosting (AB) and Multi-layer Perceptron (MLP). Using 270 diverse OpenML datasets and 304 hyperparameter configurations, we employ functional analysis of variance (fANOVA) and Ordinary Least Squares (OLS) regression to quantify the relative importance and effect sizes of hyperparameters and complexity meta-features. Our results reveal that data complexity exerts a more substantial influence on both bias and variance components than hyperparameter tuning, underscoring the importance of addressing intrinsic dataset challenges. These findings suggest that efforts to mitigate data complexity factors, such as class overlap or imbalance, may yield greater performance improvements than extensive hyperparameter optimization. This study provides actionable insights for machine learning practitioners and highlights the need for further research into the interplay between dataset properties and algorithmic performance.", "keywords": [ "Model Behavior Attribution", "Complexity Meta-features", "Hyperparameters", "Bias-Variance Decomposition" ], "rating_list": [ 3, 3, 5, 1 ], "soundness_list": [ 2, 1, 3, 1 ], "presentation_list": [ 4, 2, 4, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "x8jxf3byli", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Hisashi Oshima", "gender": "Male", "institution": "Chuo University", "country": "JP", "position": "PhD student" }, { "name": "TSUYOSHI ISHIZONE", "gender": "Male", "institution": "Meiji University", "country": "JP", "position": "PhD student" }, { "name": "Tomoyuki Higuchi", "gender": "Male", "institution": "Chuo University", "country": "JP", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 8, "n_ref_uni": 21, "n_ref": 69, "n_ref_all": 94, "n_fig": 10, "n_tab": 12, "L_tab": 9782, "n_element_tab": 574, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 794, "n_element_tab_1": 15, "formula_len_all": 2494, "formula_len_all_1": 1172, "len_all": 138131, "len_all_1": 48280, "len_abs": 1635, "len_title": 201, "len_sents": 39014, "len_sents_1": 23710, "n_sents": 252, "n_sents_1": 116, "venue": "ICLR2025", "decision": "withdraw", "L_title": 152, "L_abs": 1643, "title": "TWO STAGES DOMAIN INVARIANT REPRESENTATION LEARNERS SOLVE THE LARGE CO-VARIATE SHIFT IN UNSUPERVISED DOMAIN ADAPTATION WITH TWO DIMENSIONAL DATA DOMAINS", "abs": "Recent developments in the unsupervised domain adaptation (UDA) enable the unsupervised machine learning (ML) prediction for target data, thus this will accelerate real world applications with ML models such as image recognition tasks in self-driving. Researchers have reported the UDA techniques are not working well under large co-variate shift problems where e.g. supervised source data consists of handwritten digits data in monotone color and unsupervised target data colored digits data from the street view. Thus there is a need for a method to resolve co-variate shift and transfer source labelling rules under this dynamics. We perform two stages domain invariant representation learning to bridge the gap between source and target with semantic intermediate data (unsupervised). The proposed method can learn domain invariant features simultaneously between source and intermediate also intermediate and target. Finally this achieves good domain invariant representation between source and target plus task discriminability owing to source labels. This induction for the gradient descent search greatly eases learning convergence in terms of classification performance for target data even when large co-variate shift. We also derive a theorem for measuring the gap between trained models and unsupervised target labelling rules, which is necessary for the free parameters optimization. Finally we demonstrate that proposing method is superiority to previous UDA methods using 4 representative ML classification datasets including 38 UDA tasks. Our experiment will be a basis for challenging UDA problems with large co-variate shift.", "keywords": [ "domain invariant representation learning", "unsupervised domain adaptation", "image recognition", "signal processing", "classification" ], "rating_list": [ 3, 1, 3, 6, 1 ], "soundness_list": [ 2, 2, 3, 2, 1 ], "presentation_list": [ 1, 1, 2, 2, 1 ], "contribution_list": [ 1, 1, 2, 2, 1 ], "confidence_list": [ 4, 3, 4, 3, 5 ] }, { "paper_id": "x83w6yGIWb", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Juntao Li", "gender": "Male", "institution": "Soochow University, China", "country": "", "position": "Associate Professor" }, { "name": "Min Zhang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Ping Li", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Qingrong Xia", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Xinyu Duan", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Yang Xiang", "gender": "Male", "institution": "Suzhou University", "country": "CN", "position": "MS student" }, { "name": "Yixin Ji", "gender": "Male", "institution": "Soochow University", "country": "CN", "position": "PhD student" }, { "name": "Zhefeng Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 40, "n_ref": 68, "n_ref_all": 86, "n_fig": 6, "n_tab": 6, "L_tab": 2598, "n_element_tab": 523, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1803, "n_element_tab_1": 356, "formula_len_all": 177, "formula_len_all_1": 67, "len_all": 146327, "len_all_1": 61599, "len_abs": 1507, "len_title": 108, "len_sents": 32772, "len_sents_1": 26298, "n_sents": 243, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1447, "title": "Beware of Calibration Data for Pruning Large Language Models", "abs": "As large language models (LLMs) are widely applied across various fields, model\ncompression has become increasingly crucial for reducing costs and improving\ninference efficiency. Post-training pruning is a promising method that does not\nrequire resource-intensive iterative training and only needs a small amount of\ncalibration data to assess the importance of parameters. Recent research has enhanced post-training pruning from different aspects but few of them systematically\nexplore the effects of calibration data, and it is unclear if there exist better calibration data construction strategies. We fill this blank and surprisingly observe that\ncalibration data is also crucial to post-training pruning, especially for high sparsity. Through controlled experiments on important influence factors of calibration\ndata, including the pruning settings, the amount of data, and its similarity with\npre-training data, we observe that a small size of data is adequate, and more similar data to its pre-training stage can yield better performance. As pre-training data\nis usually inaccessible for advanced LLMs, we further provide a self-generating\ncalibration data synthesis strategy to construct feasible calibration data. Experimental results on recent strong open-source LLMs (e.g., DCLM, and LLaMA-3)\nshow that the proposed strategy can enhance the performance of strong pruning\nmethods (e.g., Wanda, DSnoT, OWL) by a large margin (up to 2.68%).", "keywords": [ "calibration data", "post-training pruning", "large language models" ], "rating_list": [ 3, 8, 5, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "x7Q0uFTH2a", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Dayang Liang", "gender": "Male", "institution": "Xiamen University", "country": "", "position": "PhD student" }, { "name": "Yunlong Liu", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 27, "n_formula_1": 15, "n_ref_uni": 49, "n_ref": 103, "n_ref_all": 124, "n_fig": 9, "n_tab": 4, "L_tab": 2469, "n_element_tab": 244, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 237, "n_element_tab_1": 10, "formula_len_all": 2657, "formula_len_all_1": 992, "len_all": 169275, "len_all_1": 67165, "len_abs": 1538, "len_title": 136, "len_sents": 48514, "len_sents_1": 33034, "n_sents": 310, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1546, "title": "Weak Bisimulation Metric-based Representations for Sparse-Reward Reinforcement Learning", "abs": "Recent studies have shown that bisimulation metrics possess the superiority of essentially extracting the features related to reinforcement learning tasks. However, limited by strict assumptions and the inherent conflict between metrics and sparse rewards, they suffer from serious representation degeneration and even collapse in sparse reward settings. To tackle the problems, we propose a reward-free weak bisimulation metric-based scalable representation learning approach (SRL). Specifically, we first introduce the weak bisimulation metric, which bypasses the intractable reward difference, instead leveraging a trainable Gaussian distribution to relax the traditional bisimulation metrics. Particularly, the Gaussian noise creates a flexible information margin for the metric optimization, which mitigates potential representation collapse caused by sparse rewards. Additionally, due to its pure distribution internally, the metric potentially mitigates representation degeneration resulting from inconsistent computations under strict assumptions. To tighten the metric, we accordingly consider continuous differences over the transition distribution to enhance the accuracy of the initial transition distribution difference, strengthening the extraction of equivalent task features. We evaluate SRL on challenging DeepMind Control Suite, MetaWorld, and Adroit tasks with sparse rewards. Empirical results demonstrate that SRL significantly outperforms state-of-the-art baselines on various tasks. The source code will be available later.", "keywords": [ "Deep reinforcement learning", "Weak bisimulation metric", "Representation learning", "Sparse reward" ], "rating_list": [ 6, 1, 3, 5 ], "soundness_list": [ 3, 1, 2, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 5, 4, 4 ] }, { "paper_id": "x7NbaU8RSU", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Songshuo Lu", "gender": "Male", "institution": "Moore Threads Technology Co. Ltd", "country": "CN", "position": "Researcher" }, { "name": "Yaohua Tang", "gender": "Male", "institution": "Moore Threads AI", "country": "CN", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 41, "n_ref_all": 64, "n_fig": 6, "n_tab": 5, "L_tab": 1619, "n_element_tab": 169, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 691, "formula_len_all_1": 472, "len_all": 115170, "len_all_1": 53656, "len_abs": 1198, "len_title": 145, "len_sents": 37649, "len_sents_1": 26245, "n_sents": 269, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1201, "title": "TurboRAG: Accelerating Retrieval-Augmented Generation with Precomputed KV Caches for Chunked Text", "abs": "Current Retrieval-Augmented Generation (RAG) systems concatenate and process numerous retrieved document chunks for prefill which requires a large volume of computation, therefore leading to significant latency in time-to-first-token (TTFT). To reduce the computation overhead as well as TTFT, we introduce TurboRAG, a novel RAG system that redesigns the inference paradigm of the current RAG system by first pre-computing and storing the key-value (KV) caches of documents offline, and then directly retrieving the saved KV cache for prefill. Hence, online computation of KV caches is eliminated during inference. In addition, we provide a number of insights into the mask matrix and positional embedding mechanisms, plus fine-tune a pretrained language model to maintain model accuracy of TurboRAG. Our approach is applicable to most existing large language models and their applications without any requirement in modification of models and inference systems. Experimental results across a suite of RAG benchmarks demonstrate that TurboRAG reduces TTFT by up to 9.4x compared to the conventional RAG systems (on an average of 8.6x), but reserving comparable performance to the standard RAG systems.", "keywords": [ "Retrieval-Augmented Generation; Large Language Models; Precomputed KV Cache" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "x6YSsKYJuH", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Benjamin I. P. Rubinstein", "gender": "Male", "institution": "The University of Melbourne", "country": "AU", "position": "Full Professor" }, { "name": "Jun Wang", "gender": "Male", "institution": "Research, Google", "country": "", "position": "Intern" }, { "name": "Pontus Stenetorp", "gender": "Not Specified", "institution": "University College London", "country": "GB", "position": "Full Professor" }, { "name": "Qiongkai Xu", "gender": "Male", "institution": "Macquarie University", "country": "", "position": "Assistant Professor" }, { "name": "Trevor Cohn", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Xuanli He", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Postdoc" }, { "name": "Pasquale Minervini", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 2, "n_ref_uni": 45, "n_ref": 85, "n_ref_all": 126, "n_fig": 21, "n_tab": 15, "L_tab": 8364, "n_element_tab": 374, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4316, "formula_len_all_1": 91, "len_all": 236817, "len_all_1": 62365, "len_abs": 1702, "len_title": 136, "len_sents": 57026, "len_sents_1": 30676, "n_sents": 496, "n_sents_1": 226, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1721, "title": "TuBA: Cross-Lingual Transferability of Backdoor Attacks in LLMs with Instruction Tuning", "abs": "The implications of backdoor attacks on English-centric large language models (LLMs) have been widely examined \u2014 such attacks can be achieved by embedding malicious behaviors during training and activated under specific conditions that trigger malicious outputs. Despite the increasing support for multilingual capabilities in open-source and proprietary LLMs, the impact of backdoor attacks on these systems remains largely under-explored. Our research focuses on crosslingual backdoor attacks against multilingual LLMs, particularly investigating how poisoning the instruction-tuning data for one or two languages can affect the outputs for languages whose instruction-tuning data was not poisoned. Despite its simplicity, our empirical analysis reveals that our method exhibits remarkable efficacy in models like mT5 and GPT-4o, with high attack success rates, surpassing 90% in more than 7 out of 12 languages across various scenarios. Our findings also indicate that more powerful models show increased susceptibility to transferable cross-lingual backdoor attacks, which also applies to LLMs predominantly pre-trained on English data, such as Llama2, Llama3, and Gemma. Moreover, our experiments demonstrate the high transferability of the proposed attack: 1) the backdoor mechanism successfully operates in cross-lingual response scenarios across 26 languages, achieving an average attack success rate of 99%, and 2) the proposed attack remains effective even after defenses are applied. These findings expose critical security vulnerabilities in multilingual LLMs and highlight the urgent need for more robust, targeted defense strategies to address the unique challenges posed by cross-lingual backdoor transfer.", "keywords": [ "backdoor attacks", "cross-lingual transfer", "LLMs" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "x5l5PRtvul", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Anson MacDonald", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "PhD student" }, { "name": "Sahani Pathiraja", "gender": "Female", "institution": "University of New South Wales", "country": "", "position": "Assistant Professor" }, { "name": "Scott A Sisson", "gender": "unknown", "institution": "University of New South Wales", "country": "AU", "position": "Full Professor" } ], "n_formula": 80, "n_formula_1": 52, "n_ref_uni": 19, "n_ref": 74, "n_ref_all": 107, "n_fig": 1, "n_tab": 5, "L_tab": 4592, "n_element_tab": 159, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8962, "formula_len_all_1": 4770, "len_all": 118921, "len_all_1": 66860, "len_abs": 1350, "len_title": 89, "len_sents": 33844, "len_sents_1": 25866, "n_sents": 328, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1356, "title": "Hybrid Kernel Stein Variational Gradient Descent", "abs": "Stein variational gradient descent (SVGD) is a particle based approximate inference algorithm. Many variants of SVGD have been proposed in recent years, including the hybrid kernel variant (h-SVGD), which has demonstrated promising results on image classification with deep neural network ensembles. In this paper, we demonstrate the ability of h-SVGD to alleviate variance collapse, a problem that SVGD is known to suffer from. Unlike other SVGD variants that alleviate variance collapse, h-SVGD does not incur additional computational cost, nor does it require the target density to factorise. We also develop the theory of h-SVGD by demonstrating the existence of a solution to the hybrid Stein partial differential equation. We highlight a special case in which h-SVGD is a kernelised Wasserstein gradient flow on a functional other than the Kullback-Leibler divergence, which is the functional describing the SVGD gradient flow. By characterising the fixed point in this special case, we show that h-SVGD does not converge to the target distribution in the the mean field limit. Other theoretical results include a descent lemma and a large particle limit result. Despite the bias in the mean field limiting distribution, experiments demonstrate that h-SVGD remains competitive on high dimensional inference tasks whilst alleviating variance collapse.", "keywords": [ "Stein Variational Gradient Descent", "Approximate Inference", "Particle-based Variational Inference", "Gradient Flow" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "x5hXkSMOd1", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Chien-Yi Wang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Min-hung Chen", "gender": "Male", "institution": "NVIDIA", "country": "TW", "position": "Research Scientist" }, { "name": "Ryo Hachiuma", "gender": "Male", "institution": "NVIDIA", "country": "TW", "position": "Researcher" }, { "name": "Yu-Chiang Frank Wang", "gender": "Male", "institution": "NVIDIA", "country": "TW", "position": "Research Director" }, { "name": "Yusuke Hirota", "gender": "Male", "institution": "NVIDIA", "country": "TW", "position": "Intern" }, { "name": "Yuta Nakashima", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 19, "n_ref_uni": 51, "n_ref": 146, "n_ref_all": 173, "n_fig": 5, "n_tab": 10, "L_tab": 2413, "n_element_tab": 281, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1083, "n_element_tab_1": 65, "formula_len_all": 678, "formula_len_all_1": 778, "len_all": 161176, "len_all_1": 59390, "len_abs": 1277, "len_title": 120, "len_sents": 48883, "len_sents_1": 26137, "n_sents": 354, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1282, "title": "SANER: Annotation-free Societal Attribute Neutralizer for Debiasing CLIP", "abs": "Large-scale vision-language models, such as CLIP, are known to contain societal bias regarding protected attributes (e.g., gender, age). This paper aims to address the problems of societal bias in CLIP. Although previous studies have proposed to debias societal bias through adversarial learning or test-time projecting, our comprehensive study of these works identifies two critical limitations: 1) loss of attribute information when it is explicitly disclosed in the input and 2) use of the attribute annotations during debiasing process. To mitigate societal bias in CLIP and overcome these limitations simultaneously, we introduce a simple-yet-effective debiasing method called SANER (societal attribute neutralizer) that eliminates attribute information from CLIP text features only of attribute-neutral descriptions. Experimental results show that SANER, which does not require attribute annotations and preserves original information for attribute-specific descriptions, demonstrates superior debiasing ability than the existing methods. Additionally, we observe that SANER does not require retraining CLIP from scratch with the original dataset. Moreover, the debiased model can be directly applied to the text-to-image generation model by simply replacing the text encoder.", "keywords": [ "Societal bias", "CLIP", "Debiasing", "Fairness" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 4, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "x5YEibapUM", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Chiwun Yang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "Majid Daliri", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" } ], "n_formula": 165, "n_formula_1": 27, "n_ref_uni": 42, "n_ref": 67, "n_ref_all": 83, "n_fig": 1, "n_tab": 5, "L_tab": 1334, "n_element_tab": 43, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15675, "formula_len_all_1": 2133, "len_all": 294536, "len_all_1": 70666, "len_abs": 1644, "len_title": 105, "len_sents": 60585, "len_sents_1": 29973, "n_sents": 854, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1452, "title": "Unlocking the Theory Behind Scaling 1-Bit Neural Networks", "abs": "Recently, 1-bit Large Language Models (LLMs) have emerged, showcasing an impressive combination of efficiency and performance that rivals traditional LLMs. Research by Wang et al. (2023); Ma et al. (2024) indicates that the performance of these 1-bit LLMs progressively improves as the number of parameters increases, hinting at the potential existence of a *Scaling Law for 1-bit Neural Networks*. In this paper, we present the \\emph{first theoretical} result that rigorously establishes this scaling law for 1-bit models. We prove that, despite the constraint of weights restricted to $\\\\{-1, +1\\\\}$, the dynamics of model training inevitably align with kernel behavior as the network width grows. This theoretical breakthrough guarantees convergence of the 1-bit model to an arbitrarily small loss as width increases. Furthermore, we introduce the concept of the generalization difference, defined as the gap between the outputs of 1-bit networks and their full-precision counterparts, and demonstrate that this difference maintains a negligible level as network width scales. Building on the work of Kaplan et al. (2020), we conclude by examining how the training loss scales as a power-law function of the model size, dataset size, and computational resources utilized for training. Our findings underscore the promising potential of scaling 1-bit neural networks, suggesting that int1 could become the standard in future neural network precision.", "keywords": [ "1-bit neural network", "quantization", "neural tangent kernel" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "x5FfUvsLIE", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Chuang Zhou", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "", "position": "PhD student" }, { "name": "Jiahe Du", "gender": "unknown", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Qiyuan Zheng", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "MS student" }, { "name": "Xiao Huang", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "", "position": "Assistant Professor" }, { "name": "Zhaozhuo Xu", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "CHEN SHENGYUAN", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Postdoc" }, { "name": "WANG Zhu", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 40, "n_ref": 56, "n_ref_all": 64, "n_fig": 5, "n_tab": 3, "L_tab": 1103, "n_element_tab": 223, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1103, "n_element_tab_1": 223, "formula_len_all": 776, "formula_len_all_1": 749, "len_all": 129448, "len_all_1": 69965, "len_abs": 2052, "len_title": 123, "len_sents": 36344, "len_sents_1": 33068, "n_sents": 263, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 2064, "title": "Large Language Models based Graph Convolution for Text-Attributed Networks", "abs": "Text-attributed graph (TAG) tasks involve analyzing both structural information and textual attributes. Existing methods employ text embeddings as node features, and leverage structural information by employing Graph Neural Networks (GNNs) to aggregate features from neighbors. These approaches demand substantial computational resources and rely on two cascaded stages, limiting scalability in large-scale scenarios and making them vulnerable to the influence of irrelevant neighboring nodes. The advancement of language models (LMs) presents new avenues for tackling this task without GNNs, leveraging their ability to process text attributes of both the target node and its important neighbors. Instead of using graph convolution modules, LMs can assign weights to these tokens based on relevance, enabling token-level weighted summarization. However, it is nontrivial to directly employ LMs for TAG tasks because assessing the importance of neighbor nodes involves both semantic and structural considerations. Additionally, the large search space presents efficiency issues for computing importance scores in a scalable manner.\nTo this end, we propose a novel semantic knowledge and Structural Enrichment framework, namely SKETCH, to adapt LMs for TAG tasks by retrieving both structural and text-related content. Specifically, we propose a retrieval model that identifies neighboring nodes exhibiting similarity to the target node across two dimensions: structural similarity and text similarity. To enable efficient retrieval, we introduce a hash-based common neighbor estimation algorithm for structural similarity and a nearest-neighbor recalling algorithm for embedding similarity. These two similarity measures are then aggregated using a weighted rank aggregation mechanism. The text attributes of both the retrieved nodes and the target node provide effective descriptions of the target node and are used as input for the LM predictor. Extensive experiments demonstrate that SKETCH can outperform other baselines on three datasets with fewer resources.", "keywords": [ "Text attributed graphs", "Long-context model" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "x4lmFlfFKX", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "C\u00e9dric Pradalier", "gender": "Male", "institution": "GeorgiaTech Lorraine", "country": "CH", "position": "Associate Professor" }, { "name": "Jeremy Fix", "gender": "Male", "institution": "CentraleSupelec", "country": "", "position": "Researcher" }, { "name": "Salim Khazem", "gender": "Male", "institution": "CentraleSupelec & GeorgiaTech-CNRS", "country": "FR", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 39, "n_ref": 52, "n_ref_all": 63, "n_fig": 3, "n_tab": 3, "L_tab": 1913, "n_element_tab": 276, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1670, "n_element_tab_1": 277, "formula_len_all": 33, "formula_len_all_1": 33, "len_all": 110419, "len_all_1": 65304, "len_abs": 2327, "len_title": 130, "len_sents": 33595, "len_sents_1": 31583, "n_sents": 212, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 2345, "title": "PolygoNet: Leveraging Simplified Polygonal Representation for Effective Shape Classification", "abs": "Deep learning models have achieved significant success in various image-related tasks. However, they often encounter challenges related to computational complexity and overfitting. In this paper, we propose an approach that leverages efficient polygonal representations of input images by utilizing either dominant points or coordinates of contours. Our method transforms input images into polygonal forms using one of these techniques, which are then employed to train deep neural networks. This representation offers a concise and flexible depiction of images. By converting images into either dominant points or contour coordinates, we substantially reduce the computational burden associated with processing large image datasets. This reduction not only accelerates the training process but also conserves computational resources, rendering our approach suitable for real-time applications and resource-constrained environments. Additionally, these representations facilitate improved generalization of the trained models. Both dominant points and contour coordinates inherently capture essential features of the input images while filtering out noise and irrelevant details, providing an inherent regularization effect that mitigates overfitting. Our approach results in lightweight models that can be efficiently deployed on edge devices, making it highly applicable for scenarios with limited computational resources. Despite the reduced complexity, our method achieve performance comparable to state-of-the-art methods that use full images as input. We validate our approach through extensive experiments on benchmark datasets, demonstrating its effectiveness in reducing computation, preventing overfitting, and enabling deployment on edge computing platforms. Overall, this work presents a methodology in image processing that leverages polygonal representations through either dominant points or contour coordinates to streamline computations, mitigate overfitting, and produce lightweight models suitable for edge computing. These findings indicate that this approach holds significant potential for advancing the field of deep learning by enabling efficient, accurate, and scalable solutions in real-world applications. The code for the experiments of the paper are provided at \\url{https://anonymous.4open.science/r/PolygoNet-7374}", "keywords": [ "Shape Classification; Polygonal representation; Computational Efficiency; Self-Attention Mechanism;" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 1, 1, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "x4jPW4p55i", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Hai Zhang", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Xinyu Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 97, "n_formula_1": 49, "n_ref_uni": 27, "n_ref": 39, "n_ref_all": 65, "n_fig": 4, "n_tab": 1, "L_tab": 239, "n_element_tab": 42, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6502, "formula_len_all_1": 3076, "len_all": 148788, "len_all_1": 69453, "len_abs": 1521, "len_title": 120, "len_sents": 42689, "len_sents_1": 25739, "n_sents": 407, "n_sents_1": 230, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1620, "title": "Learning High-dimensional Gaussian Mixture Models via a Fourier Approach", "abs": "In this paper, we address the challenge of learning high-dimensional Gaussian mixture models (GMMs), with a specific focus on estimating both the model order and the mixing distribution from i.i.d. samples. We propose a novel algorithm that achieves linear complexity relative to the sample size $n$, significantly improving computational efficiency. Unlike traditional methods, such as the method of moments or maximum likelihood estimation, our algorithm leverages Fourier measurements from the samples, facilitating simultaneous estimation of both the model order and the mixing distribution. The difficulty of the learning problem can be quantified by the separation distance $\\Delta$ and minimal mixing weight $w_{\\min}$. For stable estimation, a sample size of $\\Omega\\left(\\frac{1}{w_{\\min}^2 \\Delta^{4K-4}}\\right)$ is required for the model order, while $\\Omega\\left(\\frac{1}{w_{\\min}^2 \\Delta^{4K-2}}\\right)$ is necessary for the mixing distribution. This highlights the distinct sample complexities for the two tasks. For $D$-dimensional mixture models, we propose a PCA-based approach to reduce the dimension, reducing the algorithm\u2019s complexity to $O(nD^2)$, with potential further reductions through random projections. Numerical experiments demonstrate the efficiency and accuracy compared with the EM algorithm. In particular, we observe a clear phase transition in determining the model order, as our method outperforms traditional information criteria. Additionally, our framework is flexible and can be extended to learning mixtures of other distributions, such as Cauchy or exponential distributions.", "keywords": [ "Gaussian Mixture Models(GMM)", "Parameter Estimation", "Model Order Selection", "Super-resolution", "Line Spectral Estimation" ], "rating_list": [ 6, 3, 5, 3, 6 ], "soundness_list": [ 3, 2, 2, 2, 4 ], "presentation_list": [ 3, 2, 1, 2, 4 ], "contribution_list": [ 3, 2, 4, 2, 4 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "x4ZmQaumRg", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Daniel Musekamp", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "DE", "position": "PhD student" }, { "name": "David Holzm\u00fcller", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Postdoc" }, { "name": "Makoto Takamoto", "gender": "Male", "institution": "NEC", "country": "", "position": "Researcher" }, { "name": "Marimuthu Kalimuthu", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "DE", "position": "PhD student" }, { "name": "Mathias Niepert", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 6, "n_ref_uni": 58, "n_ref": 150, "n_ref_all": 177, "n_fig": 13, "n_tab": 13, "L_tab": 7013, "n_element_tab": 605, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 552, "n_element_tab_1": 241, "formula_len_all": 1154, "formula_len_all_1": 657, "len_all": 217462, "len_all_1": 62716, "len_abs": 1340, "len_title": 86, "len_sents": 58728, "len_sents_1": 28553, "n_sents": 470, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 38, "L_abs": 1343, "title": "Active Learning for Neural PDE Solvers", "abs": "Solving partial differential equations (PDEs) is a fundamental problem in engineering and science. While neural PDE solvers can be more efficient than established numerical solvers, they often require large amounts of training data that is costly to obtain. Active Learning (AL) could help surrogate models reach the same accuracy with smaller training sets by querying classical solvers with more informative initial conditions and PDE parameters. While AL is more common in other domains, it has yet to be studied extensively for neural PDE solvers. To bridge this gap, we introduce AL4PDE, a modular and extensible active learning benchmark. It provides multiple parametric PDEs and state-of-the-art surrogate models for the solver-in-the-loop setting, enabling the evaluation of existing and the development of new AL methods for PDE solving. We use the benchmark to evaluate batch active learning algorithms such as uncertainty- and feature-based methods. We show that AL reduces the average error by up to 71\\% compared to random sampling and significantly reduces worst-case errors. Moreover, AL generates similar datasets across repeated runs, with consistent distributions over the PDE parameters and initial conditions. The acquired datasets are reusable, providing benefits for surrogate models not involved in the data generation.", "keywords": [ "Active Learning", "Neural PDE Solvers", "Scientific Machine Learning", "Benchmark", "Framework", "Neural Operators" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "x4W8P7ybTE", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Ahmed Alaa", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Assistant Professor" }, { "name": "Alexander Marcel Schubert", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Shenghuan Sun", "gender": "Male", "institution": "University of California, San Francisco", "country": "US", "position": "Assistant Professor" }, { "name": "Thomas Hartvigsen", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Assistant Professor" }, { "name": "Wei Day", "gender": "Male", "institution": "National Central University", "country": "", "position": "Undergrad student" }, { "name": "Zhiqing Sun", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Researcher" }, { "name": "Atul J Butte", "gender": "Male", "institution": "University of California, San Francisco", "country": "", "position": "Full Professor" }, { "name": "Gregory Goldgof", "gender": "unknown", "institution": "Memorial Sloan Kettering Cancer Centre", "country": "", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 37, "n_ref": 58, "n_ref_all": 81, "n_fig": 5, "n_tab": 4, "L_tab": 703, "n_element_tab": 77, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 449, "n_element_tab_1": 60, "formula_len_all": 410, "formula_len_all_1": 404, "len_all": 164580, "len_all_1": 79332, "len_abs": 3045, "len_title": 122, "len_sents": 43056, "len_sents_1": 34027, "n_sents": 279, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1448, "title": "Aligning Multimodal Models for Clinical Reasoning using Rule-based Rewards", "abs": "Vision-Language Models (VLM) can support clinicians by analyzing medical images and engaging in natural language interactions to assist in diagnostic and treatment tasks. However, VLMs often exhibit \"hallucinatory\" behavior, generating textual outputs not grounded in contextual multimodal information. This challenge is particularly pronounced in the medical domain, where we do not only require VLM outputs to be accurate in single interactions but also to be consistent with clinical reasoning and diagnostic pathways throughout multi-turn conversations. For this purpose, we propose a new alignment algorithm that uses rule-based representations of clinical reasoning to ground VLMs in medical knowledge. These representations are utilized to (i) generate visual instruction tuning data at scale, simulating clinician-VLM conversations with demonstrations of clinical reasoning, and (ii) to derive a rule-based reward function that automatically evaluates the clinical validity of VLM responses throughout clinician-VLM interactions. Our algorithm eliminates the need for human involvement in training data generation or reward model construction, reducing costs compared to standard reinforcement learning with human feedback (RLHF). We apply our alignment algorithm to develop Dr-LLaVA, a conversational VLM finetuned for analyzing bone marrow pathology slides, demonstrating strong performance in single and multi-turn medical conversations.", "keywords": [ "Medical Vision-Language Models" ], "rating_list": [ 3, 5, 8, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 2, 4 ] }, { "paper_id": "x45vUUY4nT", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Pengwei Tang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Shaojie Li", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 140, "n_formula_1": 22, "n_ref_uni": 79, "n_ref": 151, "n_ref_all": 178, "n_fig": 2, "n_tab": 3, "L_tab": 200, "n_element_tab": 31, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 20577, "formula_len_all_1": 1436, "len_all": 230012, "len_all_1": 74659, "len_abs": 1095, "len_title": 118, "len_sents": 49586, "len_sents_1": 31528, "n_sents": 526, "n_sents_1": 282, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1131, "title": "Sharper Bounds of Non-Convex Stochastic Gradient Descent with Momentum", "abs": "Stochastic gradient descent with momentum (SGDM) has been widely used in machine learning. However, in non-convex domains, high probability learning bounds for SGDM are scarce. In this paper, we provide high probability convergence bounds and generalization bounds for SGDM. Firstly, we establish these bounds for the gradient norm in the general non-convex case. The derived convergence bounds are tighter than the theoretical results of related work, and to our best knowledge, the derived generalization bounds are the first ones for SGDM. Then, if the Polyak-{\\L}ojasiewicz condition is satisfied, we establish these bounds for the error of the function value, instead of the gradient norm. Moreover, the derived learning bounds have faster rates than the general non-convex case. Finally, we further provide sharper generalization bounds by considering a mild Bernstein condition on the gradient. In the case of low noise, their learning rates can reach $\\widetilde{\\mathcal{O}}(1/n^2)$, where $n$ is the sample size. Overall, we relatively systematically investigate the high probability learning bounds for non-convex SGDM.", "keywords": [ "learning theory", "nonconvex optimization", "stochastic gradient descent" ], "rating_list": [ 5, 3, 6, 5, 6 ], "soundness_list": [ 2, 1, 3, 3, 3 ], "presentation_list": [ 3, 1, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 2, 3 ] }, { "paper_id": "x418ZpazsR", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Charles Lovering", "gender": "unknown", "institution": "Kensho", "country": "US", "position": "Researcher" }, { "name": "Chris Tanner", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Lecturer" }, { "name": "Michael Krumdick", "gender": "Male", "institution": "Kensho", "country": "US", "position": "Researcher" }, { "name": "Nilesh Kumar", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Rik Koncel-Kedziorski", "gender": "Non-Binary", "institution": "Apple", "country": "US", "position": "Scientist" }, { "name": "Varshini Reddy", "gender": "Female", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "US", "position": "MS student" }, { "name": "Viet Dac Lai", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 2, "n_ref_uni": 28, "n_ref": 119, "n_ref_all": 165, "n_fig": 13, "n_tab": 6, "L_tab": 2135, "n_element_tab": 290, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1601, "n_element_tab_1": 210, "formula_len_all": 1812, "formula_len_all_1": 97, "len_all": 206294, "len_all_1": 61325, "len_abs": 2861, "len_title": 58, "len_sents": 73172, "len_sents_1": 25573, "n_sents": 513, "n_sents_1": 241, "venue": "ICLR2025", "decision": "withdraw", "L_title": 37, "L_abs": 1438, "title": "Are Language Model Logits Calibrated?", "abs": "Some information is factual (e.g., \"Paris is in France\"), whereas other information is probabilistic (e.g., \"the coin flip will be a [Heads/Tails].\"). We believe that good Language Models (LMs) should understand and reflect this nuance. Our work investigates this by testing if LMs' output probabilities are *calibrated* to their textual contexts. We define model \"calibration\" as the degree to which the output probabilities of candidate tokens are aligned with the relative likelihood that should be inferred from the given context. For example, if the context concerns two equally likely options (e.g., heads or tails for a fair coin), the output probabilities should reflect this. Likewise, context that concerns non-uniformly likely events (e.g., rolling a six with a die) should also be appropriately captured with proportionate output probabilities. We find that even in simple settings the best LMs (1) are poorly calibrated, and (2) have systematic biases (e.g., preferred colors and sensitivities to word orderings). For example, gpt-4o-mini often picks the first of two options presented in the prompt regardless of the options' implied likelihood, whereas Llama-3.1-8B picks the second. Our other consistent finding is mode-collapse: Instruction-tuned models often over-allocate probability mass on a single option. These systematic biases introduce non-intuitive model behavior, making models harder for users to understand.", "keywords": [ "language modeling", "calibration", "model understanding" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "x3lE88YkUl", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Hoang-Chau Luong", "gender": "Male", "institution": "Ho Chi Minh city University of Science, Vietnam National University", "country": "VN", "position": "Undergrad student" }, { "name": "Thuc Nguyen-Quang", "gender": "Male", "institution": "Ho Chi Minh city University of Science, Vietnam National University", "country": "VN", "position": "Researcher" }, { "name": "Minh-triet Tran", "gender": "Male", "institution": "Ho Chi Minh city University of Science, Vietnam National University", "country": "VN", "position": "Associate Professor" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 44, "n_ref": 87, "n_ref_all": 119, "n_fig": 13, "n_tab": 11, "L_tab": 3236, "n_element_tab": 282, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2602, "n_element_tab_1": 79, "formula_len_all": 946, "formula_len_all_1": 969, "len_all": 153829, "len_all_1": 61222, "len_abs": 1058, "len_title": 122, "len_sents": 48681, "len_sents_1": 27361, "n_sents": 324, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1061, "title": "Improving Resistance to Noisy Label Fitting by Reweighting Gradient in SAM", "abs": "Noisy labels pose a substantial challenge in machine learning, often resulting in overfitting and poor generalization. Sharpness-Aware Minimization (SAM), as demonstrated in Foret et al. (2021), improves generalization over traditional Stochastic Gradient Descent (SGD) in classification tasks with noisy labels by implicitly slowing noisy learning. While SAM\u2019s ability to generalize in noisy environments has been studied in several simplified settings, its full potential in more realistic training settings remains underexplored. In this work, we analyze SAM\u2019s behavior at each iteration, identifying specific components of the gradient vector that contribute significantly to its robustness against noisy labels. Based on these insights, we propose SANER (Sharpness-Aware Noise-Explicit Reweighting), an effective variant that enhances SAM\u2019s ability to manage noisy fitting rate. Our experiments on CIFAR-10, CIFAR-100, and Mini-WebVision demonstrate that SANER consistently outperforms SAM, achieving up to an 8% increase on CIFAR-100 with 50% label noise.", "keywords": [ "label noise", "sharpness-aware minimization", "optimization" ], "rating_list": [ 6, 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 5, 3, 4 ] }, { "paper_id": "x3l0fQubOn", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Athanasios Aris Panagopoulos", "gender": "Male", "institution": "California State University, Fresno", "country": "US", "position": "Associate Professor" }, { "name": "Robert Wiebe", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Undergrad student" } ], "n_formula": 7, "n_formula_1": 0, "n_ref_uni": 19, "n_ref": 54, "n_ref_all": 69, "n_fig": 9, "n_tab": 1, "L_tab": 781, "n_element_tab": 37, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 781, "n_element_tab_1": 37, "formula_len_all": 957, "formula_len_all_1": 0, "len_all": 97315, "len_all_1": 65278, "len_abs": 2139, "len_title": 81, "len_sents": 39980, "len_sents_1": 34553, "n_sents": 245, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 147, "L_abs": 2148, "title": "Structural Quantile Normalization: a general, differentiable feature scaling technique balancing gaussian approximation and structural preservation", "abs": "Feature scaling is an essential practice in modern machine learning, both as a preprocessing step and as an integral part of model architectures, such as batch and layer normalization in artificial neural networks. Its primary goal is to align feature scales, preventing larger-valued features from dominating model learning\u2014especially in algorithms utilizing distance metrics, gradient-based optimization, and regularization. Additionally, many algorithms benefit from or require input data approximating a standard Gaussian distribution, establishing \"Gaussianization\" as an additional objective. Lastly, an ideal scaling method should be general, as in applicable to any input distribution, and differentiable to facilitate seamless integration into gradient-optimized models. Although differentiable and general, traditional linear methods, such as standardization and min-max scaling, cannot reshape distributions relative to scale and offset. On the other hand, existing nonlinear methods, although more effective at Gaussianizing data, either lack general applicability (e.g., power transformations) or introduce excessive distortions that can obscure intrinsic data patterns (e.g., quantile normalization). Present non-linear methods are also not differentiable. We introduce Structural Quantile Normalization (SQN), a general and differentiable scaling method, that enables balancing Gaussian approximation with structural preservation. We also introduce Fast-SQN; a more performance-efficient variant with the same properties. We show that SQN is a generalized augmentation of standardization and quantile normalization. Using the real-world \"California Housing\" dataset, we demonstrate that Fast-SQN outperforms state-of-the-art methods\u2014including classical and ordered quantile normalization, and Box-Cox, and Yeo-Johnson transformations\u2014across key metrics (i.e., RMSE, MAE, MdAE) when used for preprocessing.\nFinally, we show our approach transformation differentiability and compatibility with gradient-based optimization using the real-world \"Gas Turbine Emission\" dataset and propose a methodology for integration into deep networks.", "keywords": [ "feature scaling", "preprocessing", "normal distribution", "differentiable transformation", "quantile normalization", "neural networks" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "x3jRzVAltZ", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Hanwang Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Jiachun Pan", "gender": "Female", "institution": "Skywork AI ", "country": "SG", "position": "Researcher" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Xingyu Xie", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Postdoc" } ], "n_formula": 35, "n_formula_1": 15, "n_ref_uni": 21, "n_ref": 51, "n_ref_all": 83, "n_fig": 16, "n_tab": 8, "L_tab": 1482, "n_element_tab": 192, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 245, "n_element_tab_1": 26, "formula_len_all": 6423, "formula_len_all_1": 1312, "len_all": 135999, "len_all_1": 63880, "len_abs": 1419, "len_title": 138, "len_sents": 43398, "len_sents_1": 28546, "n_sents": 383, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1435, "title": "VR-Sampling: Accelerating Flow Generative Model Training with Variance Reduction Sampling", "abs": "Recent advancements in text-to-image and text-to-video models, such as Stable Diffusion 3 (SD3), Flux and OpenSora, have adopted rectified flow over traditional diffusion models to enhance training and inference efficiency. SD3 notes increased difficulty in learning at intermediate timesteps but does not clarify the underlying cause. In this paper, we theoretically identify the root cause as a higher variance in the loss gradient estimates at these timesteps, which hinders training efficiency. Furthermore, this high-variance region is significantly influenced by the noise schedulers (i.e., how we add noises to clean images) and data (or latent space) dimensions. Building on this theoretical insight, we propose a Variance-Reduction Sampling (VR-sampling) strategy that samples the timesteps in high-variance region more frequently to enhance training efficiency in flow models. VR-sampling constructs sampling distributions based on Monte Carlo estimates of the loss gradient variance, allowing it to easily extend to different noise schedulers and data dimensions. Experiments demonstrate that VR sampling accelerates training by up to 33\\% on ImageNet 256 and 50\\% on ImageNet 512 datasets in rectified flow models. Furthermore, VR-sampling could simplify the hyperparameter tuning of logit-normal sampling introduced in SD3. \nThe code is available anonymously in~\\url{https://github.com/AnonymousProjects/VR_sampling.git}.", "keywords": [ "Flow Generative Models", "Training Acceleration", "Diffusion Models" ], "rating_list": [ 8, 5, 8, 3 ], "soundness_list": [ 3, 3, 3, 1 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "x3cFAoorct", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 1, "author_info_list": [ { "name": "Jiacheng Yin", "gender": "Not Specified", "institution": "HLX", "country": "CN", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 18, "n_ref_uni": 31, "n_ref": 38, "n_ref_all": 49, "n_fig": 5, "n_tab": 4, "L_tab": 3749, "n_element_tab": 62, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 269, "n_element_tab_1": 53, "formula_len_all": 9018, "formula_len_all_1": 1278, "len_all": 107948, "len_all_1": 44576, "len_abs": 1098, "len_title": 116, "len_sents": 30108, "len_sents_1": 20110, "n_sents": 196, "n_sents_1": 130, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1100, "title": "Learning Arbitrary Logical Formula as a Sparse Neural Network Module", "abs": "NeSy (Neuro-Symbolic) predictors are hybrid models composed of symbolic predictive models chained after neural networks. Most existing NeSy predictors require either given symbolic knowledge or iterative training. DSL (Deep Symbolic Learning) is the first NeSy predictor that supports fully end-to-end training from scratch, but it learns a look-up table rather than arbitrary programs or formulas. We propose the Logical Formula Learner framework, a general framework of network modules that explicitly equate a logical formula after convergence. We then propose 3 novel designs within the LFL framework with different levels of combinatorial search freedom: LFL-Type1 learns arbitrary logical formula, LFL-Type2 learns a look-up table, and LFL-Type3 has combinatorial search freedom between them. LFL-Type1 and LFL-Type2 show improvements over previous designs, and all three types can be wrapped into NeSy predictors. To our knowledge, LFL-Type1-based NeSy predictor is the first NeSy predictor that supports fully end-to-end training from scratch and explicitly learns arbitrary logical formulas.", "keywords": [ "Neuro-Symbolic AI; System 2 intelligence; Deep Symbolic Learning (DSL); Equation Learner (EQL); differentiable Neural Logic Networks (dNL)" ], "rating_list": [ 5, 3, 8, 3, 3 ], "soundness_list": [ 2, 1, 3, 2, 1 ], "presentation_list": [ 1, 1, 4, 2, 1 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 2, 5, 3, 3, 3 ] }, { "paper_id": "x3F8oPxKV2", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Agrin Hilmkil", "gender": "Male", "institution": "Latent Labs", "country": "GB", "position": "Researcher" }, { "name": "Cheng Zhang", "gender": "Female", "institution": "Meta", "country": "GB", "position": "Principal Researcher" }, { "name": "Divyat Mahajan", "gender": "Male", "institution": "Montreal Institute of Learning Algorithms", "country": "CA", "position": "PhD student" }, { "name": "Jannes Gladrow", "gender": "Male", "institution": "Research, Microsoft", "country": "GB", "position": "Researcher" }, { "name": "Meyer Scetbon", "gender": "Male", "institution": "Microsoft", "country": "GB", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 23, "n_ref_uni": 37, "n_ref": 81, "n_ref_all": 124, "n_fig": 4, "n_tab": 30, "L_tab": 14292, "n_element_tab": 783, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1953, "n_element_tab_1": 57, "formula_len_all": 1686, "formula_len_all_1": 1567, "len_all": 193791, "len_all_1": 66319, "len_abs": 1598, "len_title": 83, "len_sents": 60781, "len_sents_1": 30762, "n_sents": 466, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1587, "title": "Zero-Shot Learning of Causal Models", "abs": "With the increasing acquisition of datasets over time, we now have access to precise and varied descriptions of the world, capturing all sorts of phenomena.\nThese datasets can be seen as empirical observations of unknown causal generative processes, or Structural Causal Models (SCMs).\nRecovering these causal generative processes from observations poses formidable challenges, and often require to learn a specific generative model for each dataset.\nIn this work, we propose to learn a \\emph{single} model capable of inferring in a zero-shot manner the causal generative processes of datasets. \nRather than learning a specific SCM for each dataset, we enable FiP, the architecture proposed in~\\cite{scetbon2024fip}, to infer the generative SCMs conditionally on their empirical representations.\nMore specifically, we propose to amortize the learning of a conditional version of FiP to infer directly the generative SCMs from observations and causal structures on synthetically generated datasets.\nWe show that our model is capable of predicting in zero-shot the true generative SCMs, and as a by-product, of (i) generating new dataset samples, and (ii) inferring intervened ones.\nOur experiments demonstrate that our amortized procedure achieves performances on par with SoTA methods trained specifically for each dataset on both in and out-of-distribution problems. \nTo the best of our knowledge, this is the first time that SCMs are inferred in a zero-shot manner from observations, paving the way for a paradigmatic shift towards the assimilation of causal knowledge across datasets.", "keywords": [ "Causality", "Transformers", "Generative Models" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "x33vSZUg0A", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Jing Chen", "gender": "unknown", "institution": "tencent", "country": "CN", "position": "Researcher" }, { "name": "LINGYU DUAN", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Sha Guo", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Wenhan Yang", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xing Jiang", "gender": "unknown", "institution": "Fuzhou Chengtou New Infrastructure Group", "country": "CN", "position": "Researcher" }, { "name": "Yu Lin", "gender": "unknown", "institution": "Fuzhou Chengtou New Infrastructure Group", "country": "CN", "position": "Researcher" }, { "name": "Zhuo Chen", "gender": "Male", "institution": "Pengcheng Lab", "country": "CN", "position": "Researcher" }, { "name": "Zixuan Hu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 19, "n_ref_uni": 65, "n_ref": 141, "n_ref_all": 182, "n_fig": 20, "n_tab": 7, "L_tab": 4003, "n_element_tab": 583, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 338, "n_element_tab_1": 104, "formula_len_all": 2279, "formula_len_all_1": 1505, "len_all": 207608, "len_all_1": 64557, "len_abs": 1259, "len_title": 170, "len_sents": 60035, "len_sents_1": 28995, "n_sents": 451, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 122, "L_abs": 1264, "title": "Which Tasks Should Be Compressed Together? A Causal Discovery Approach for Efficient Multi-Task Representation Compression", "abs": "Conventional image compression methods are inadequate for intelligent analysis, as they overemphasize pixel-level precision while neglecting semantic significance and the interaction among multiple tasks. This paper introduces a Taskonomy-Aware Multi-Task Compression framework comprising (1) inter-coherent task grouping, which organizes synergistic tasks into shared representations to improve multi-task accuracy and reduce encoding volume, and (2) a conditional entropy-based directed acyclic graph (DAG) that captures causal dependencies among grouped representations. By leveraging parent representations as contextual priors for child representations, the framework effectively utilizes cross-task information to improve entropy model accuracy. Experiments on diverse vision tasks, including Keypoint 2D, Depth Zbuffer, Semantic Segmentation, Surface Normal, Edge Texture, and Autoencoder, demonstrate significant bitrate-performance gains, validating the method\u2019s capability to reduce system entropy uncertainty. These findings underscore the potential of leveraging representation disentanglement, synergy, and causal modelling for compact representation learning, enabling efficient multi-task compression in intelligent systems. Code will be available.", "keywords": [ "Video Coding for Machine", "Image Compression", "Multi-task Learning", "Causal Discovery" ], "rating_list": [ 3, 5, 8 ], "soundness_list": [ 2, 3, 4 ], "presentation_list": [ 1, 1, 3 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "x1yOHtFfDh", "primary_area": "datasets and benchmarks", "n_author": 13, "author_info_list": [ { "name": "Chi Lu", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Undergrad student" }, { "name": "Christopher Lai", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "Undergrad student" }, { "name": "Hanjie Chen", "gender": "Female", "institution": "Rice University", "country": "US", "position": "Assistant Professor" }, { "name": "Haotian Xia", "gender": "Male", "institution": "University of California, Irvine", "country": "", "position": "MS student" }, { "name": "Junbo Zou", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Rhys Tracy", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "Researcher" }, { "name": "Weining Shen", "gender": "unknown", "institution": "University of California, Irvine", "country": "", "position": "Associate Professor" }, { "name": "Xun Shao", "gender": "Male", "institution": "Wenzhou-Kean University", "country": "CN", "position": "Undergrad student" }, { "name": "Yanjun He", "gender": "Female", "institution": "University of California, Irvine", "country": "US", "position": "Undergrad student" }, { "name": "Yuan-fang Wang", "gender": "unknown", "institution": ", University of California, Santa Barbara", "country": "US", "position": "Full Professor" }, { "name": "Yuqing Wang", "gender": "Female", "institution": "Stanford University", "country": "", "position": "Postdoc" }, { "name": "Zhengbang Yang", "gender": "Male", "institution": "George Mason University", "country": "US", "position": "MS student" }, { "name": "Zhuoqing Xie", "gender": "Female", "institution": "University of California, Santa Barbara", "country": "US", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 67, "n_ref_all": 84, "n_fig": 4, "n_tab": 12, "L_tab": 4384, "n_element_tab": 501, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1482, "n_element_tab_1": 124, "formula_len_all": 467, "formula_len_all_1": 0, "len_all": 226716, "len_all_1": 68925, "len_abs": 3840, "len_title": 141, "len_sents": 58564, "len_sents_1": 33521, "n_sents": 493, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1805, "title": "SportU: A Comprehensive Sports Understanding Benchmark for Multimodal Large Language Models", "abs": "Multimodal Large Language Models (MLLMs) are advancing the ability to reason about complex sports scenarios by integrating textual and visual information. To comprehensively evaluate their capabilities, we introduce SPORTU, a benchmark designed to assess MLLMs across multi-level sports reasoning tasks. SPORTU comprises two key components: SPORTU-text, featuring 900 multiple-choice questions with human-annotated explanations for rule comprehension and strategy understanding. This component focuses on testing models' ability to reason about sports solely through question-answering (QA), without requiring visual inputs; SPORTU-video, consisting of 1,701 slow-motion video clips across 7 different sports and 12,048 QA pairs, designed to assess multi-level reasoning, from simple sports recognition to complex tasks like foul detection and rule application. We evaluate four prevalent LLMs mainly utilizing few-shot learning paradigms supplemented by chain-of-thought (CoT) prompting on the SPORTU-text part. We evaluate four LLMs using few-shot learning and chain-of-thought (CoT) prompting on SPORTU-text. GPT-4o achieves the highest accuracy of 71\\%, but still falls short of human-level performance, highlighting room for improvement in rule comprehension and reasoning. The evaluation for the SPORTU-video part includes 7 proprietary and 6 open-source MLLMs. Experiments show that models fall short on hard tasks that require deep reasoning and rule-based understanding. Claude-3.5-Sonnet performs the best with only 52.6\\% accuracy on the hard task, showing large room for improvement. We hope that SPORTU will serve as a critical step toward evaluating models' capabilities in sports understanding and reasoning. The dataset is available at \\url{https://anonymous.4open.science/r/ICLR_01-42D5/}", "keywords": [ "Multimodal Large Language Models", "Sports Understanding", "Benchmark" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 4, 2, 1 ], "presentation_list": [ 2, 4, 2, 2 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "x1uv2gdjKV", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Fan Wang", "gender": "Female", "institution": "Alibaba Group", "country": "US", "position": "Senior Staff Algorithm Engineer" }, { "name": "Jiangweizhi Peng", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" }, { "name": "Jiasheng Tang", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Mingyi Hong", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Tsung-Hui Chang", "gender": "Not Specified", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Zhiwei Tang", "gender": "Male", "institution": "Chinese University of HongKong, Shenzhen", "country": "CN", "position": "PhD student" } ], "n_formula": 38, "n_formula_1": 21, "n_ref_uni": 25, "n_ref": 85, "n_ref_all": 144, "n_fig": 27, "n_tab": 3, "L_tab": 549, "n_element_tab": 105, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2807, "formula_len_all_1": 1499, "len_all": 228663, "len_all_1": 73361, "len_abs": 1420, "len_title": 123, "len_sents": 68939, "len_sents_1": 36997, "n_sents": 510, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1429, "title": "Inference-Time Alignment of Diffusion Models with Direct Noise Optimization", "abs": "In this work, we focus on the alignment problem of diffusion models with a continuous reward function, which represents specific objectives for downstream tasks, such as increasing darkness or improving the aesthetics of images. The central goal of the alignment problem is to adjust the distribution learned by diffusion models such that the generated samples maximize the target reward function. We propose a novel alignment approach, named Direct Noise Optimization (DNO), that optimizes the injected noise during the sampling process of diffusion models. By design, DNO operates at inference-time, and thus is tuning-free and prompt-agnostic, with the alignment occurring in an online fashion during generation. We rigorously study the theoretical properties of DNO and also propose variants to deal with non-differentiable reward functions. Furthermore, we identify that naive implementation of DNO occasionally suffers from the out-of-distribution reward hacking problem, where optimized samples have high rewards but are no longer in the support of the pretrained distribution. To remedy this issue, we leverage classical high-dimensional statistics theory to an effective probability regularization technique. We conduct extensive experiments on several important reward functions and demonstrate that the proposed DNO approach can achieve state-of-the-art reward scores within a reasonable time budget for generation.", "keywords": [ "Diffusion Models", "Inference-Time Alignment", "Optimization", "RLHF" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 2, 3, 2 ] }, { "paper_id": "x1nlO1d1iG", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Cheng Cheng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Enhong Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jiayu Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Jing Sha", "gender": "Male", "institution": "iFLYTEK Research", "country": "", "position": "Researcher" }, { "name": "Jinze Wu", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Qi Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Shijin Wang", "gender": "Male", "institution": "State Key Laboratory of Cognitive Intelligence", "country": "", "position": "Vice Dean" }, { "name": "Wei Dai", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Zhenya Huang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 29, "n_ref": 64, "n_ref_all": 81, "n_fig": 3, "n_tab": 3, "L_tab": 1955, "n_element_tab": 185, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1535, "n_element_tab_1": 99, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 169854, "len_all_1": 65247, "len_abs": 1801, "len_title": 134, "len_sents": 64095, "len_sents_1": 31341, "n_sents": 548, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1809, "title": "CogMath: Evaluating LLMs' Authentic Mathematical Ability from a Cognitive Perspective", "abs": "As large language models (LLMs) exhibit potential in solving complex mathematical tasks, increasing attention has been directed toward constructing benchmarks to evaluate their mathematical capabilities. However, existing benchmarks are either limited to specific task types (e.g., long-text problem understanding) or rely solely on a coarse measure of answer accuracy, making them insufficient for assessing a model's authentic mathematical proficiency. In this paper, we propose CogMath, which provides a comprehensive assessment of LLMs' mathematical abilities based on human cognitive processes. Specifically, inspired by cognitive theories, CogMath formalizes the reasoning process into 3 stages that align with human cognition: problem comprehension, problem solving, and solution summarization, and encompasses 9 fine-grained evaluation dimensions from perspectives such as numerical calculation, knowledge, and counterfactuals. In each dimension, to carry out a scientific evaluation, we develop an ``Inquiry-Judge-Reference'' multi-agent system, where the Inquiry agent generates inquiries that assess LLMs' mastery from this dimension, the Judge agent ensures the inquiry quality, and the Reference agent provides correct responses for comparison with the LLMs' actual performances. A LLM is considered to truly master a problem only when excelling in all inquiries from the 9 dimensions. In experiments, we evaluate 7 mainstream LLMs by applying CogMath to three benchmarks, which cover the full K-12 mathematical curriculum. The results reveal that the authentic mathematical capabilities of current LLMs are overestimated by 30-40%. Moreover, we locate their strengths and weaknesses across different stages/dimensions, offering constructive insights to further enhance their reasoning abilities.", "keywords": [ "Mathematical Reasoning", "Large Language Models" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 1 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "x1SfON9HvT", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Hanping Zhang", "gender": "Male", "institution": "Carleton University", "country": "CA", "position": "PhD student" }, { "name": "Yuhong Guo", "gender": "Female", "institution": "Carleton University", "country": "CA", "position": "Full Professor" } ], "n_formula": 25, "n_formula_1": 18, "n_ref_uni": 18, "n_ref": 47, "n_ref_all": 58, "n_fig": 0, "n_tab": 3, "L_tab": 1628, "n_element_tab": 107, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 773, "n_element_tab_1": 54, "formula_len_all": 1703, "formula_len_all_1": 1141, "len_all": 101156, "len_all_1": 59832, "len_abs": 911, "len_title": 117, "len_sents": 34476, "len_sents_1": 27496, "n_sents": 226, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 915, "title": "Diffusion Modulation via Environment Mechanism Modeling for Planning", "abs": "Diffusion models have shown promising capabilities in trajectory generation for planning in offline reinforcement learning (RL). However, conventional diffusion-based planning methods often fail to account for the fact that generating trajectories in RL requires unique consistency between transitions to ensure coherence in real environments. This oversight can result in considerable discrepancies between the generated trajectories and the underlying mechanisms of a real environment. To address this problem, we propose a novel diffusion-based planning method, termed as Diffusion Modulation via Environment Mechanism Modeling (DMEMM). DMEMM modulates diffusion model training by incorporating key RL environment mechanisms, particularly transition dynamics and reward functions. Experimental results demonstrate that DMEMM achieves state-of-the-art performance for planning with offline reinforcement learning.", "keywords": [ "Reinforcement Learning", "Offline Reinforcement Learning", "Planning", "Diffusion Model" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "x1Okv4kbVR", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Dawei Yin", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Principal Researcher" }, { "name": "Lingyong Yan", "gender": "Male", "institution": "Baidu Inc.", "country": "CN", "position": "Search Scientist" }, { "name": "Maarten de Rijke", "gender": "unknown", "institution": "University of Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "Pengjie Ren", "gender": "Male", "institution": "Shandong University", "country": "", "position": "Full Professor" }, { "name": "Yougang Lyu", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Zhaochun Ren", "gender": "Male", "institution": "Leiden University", "country": "NL", "position": "Associate Professor" }, { "name": "Zihan Wang", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 50, "n_ref": 149, "n_ref_all": 167, "n_fig": 7, "n_tab": 9, "L_tab": 3193, "n_element_tab": 291, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 948, "n_element_tab_1": 134, "formula_len_all": 718, "formula_len_all_1": 640, "len_all": 266496, "len_all_1": 67260, "len_abs": 1624, "len_title": 131, "len_sents": 50680, "len_sents_1": 29736, "n_sents": 378, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1627, "title": "MACPO: Weak-to-Strong Alignment via Multi-Agent Contrastive Preference Optimization", "abs": "As large language models (LLMs) are rapidly advancing and achieving near-human capabilities, aligning them with human values is becoming more urgent. In scenarios where LLMs outperform humans, we face a weak-to-strong alignment problem where we need to effectively align strong student LLMs through weak supervision generated by weak teachers. Existing alignment methods mainly focus on strong-to-weak alignment and self-alignment settings, and it is impractical to adapt them to the much harder weak-to-strong alignment setting. To fill this gap, we propose a multi-agent contrastive preference optimization (MACPO) framework. MACPO facilitates weak teachers and strong students to learn from each other by iteratively reinforcing unfamiliar positive behaviors while penalizing familiar negative ones. To get this, we devise a mutual positive behavior augmentation strategy to encourage weak teachers and strong students to learn from each other\u2019s positive behavior and further provide higher quality positive behavior for the next iteration. Additionally, we propose a hard negative behavior construction strategy to induce weak teachers and strong students to generate familiar negative behavior by fine-tuning on negative behavioral data. Experimental results on the HH-RLHF and PKU-SafeRLHF datasets, evaluated using both automatic metrics and human judgments, demonstrate that MACPO simultaneously improves alignment performance of strong students and weak teachers. Moreover, as the number of weak teachers increases, MACPO achieves better weak-to-strong alignment performance through more iteration optimization rounds.", "keywords": [ "weak-to-strong alignment", "preference optimization" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "x1Bk51SCL9", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Jiangning Wei", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Mei Wang", "gender": "Female", "institution": "Beijing Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Miaoxuan Zhang", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Shilong Ou", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Weiran Xu", "gender": "Male", "institution": "Beijing University of Post and Telecommunication", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaoshuai Song", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Yuhang Zhang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "lixiong Qin", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "\u5218\u5b87\u6668", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 46, "n_ref": 83, "n_ref_all": 134, "n_fig": 10, "n_tab": 6, "L_tab": 10500, "n_element_tab": 1592, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 5975, "n_element_tab_1": 910, "formula_len_all": 901, "formula_len_all_1": 388, "len_all": 242296, "len_all_1": 71202, "len_abs": 4411, "len_title": 150, "len_sents": 55757, "len_sents_1": 24912, "n_sents": 434, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1387, "title": "Face-Human-Bench: A Comprehensive Benchmark of Face and Human Understanding for Multi-modal Assistants", "abs": "Faces and humans are crucial elements in social interaction and are widely included in everyday photos and videos. Therefore, a deep understanding of faces and humans will enable multi-modal assistants to achieve improved response quality and broadened application scope. Currently, the multi-modal assistant community lacks a comprehensive and scientific evaluation of face and human understanding abilities. In this paper, we first propose a hierarchical ability taxonomy that includes three levels of abilities. Then, based on this taxonomy, we collect images and annotations from publicly available datasets in the face and human community and build a semi-automatic data pipeline to produce problems for the new benchmark. Finally, the obtained Face-Human-Bench comprises a development set with 900 problems and a test set with 1800 problems, supporting both English and Chinese. We conduct evaluations over 25 mainstream multi-modal large language models (MLLMs) with our Face-Human-Bench, focusing on the correlation between abilities, the impact of the relative position of targets on performance, and the impact of Chain of Thought (CoT) prompting on performance. Moreover, inspired by multi-modal agents, we also explore which abilities of MLLMs need to be supplemented by specialist models. The data and evaluation code of the Face-Human-Bench will be made publicly available.", "keywords": [ "face and human understanding", "multi-modal assistants", "benchmark" ], "rating_list": [ 6, 6, 3, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 1, 3, 2, 4 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "x1An5a3U9I", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Bo Chen", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Mingyuan Zhou", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Xinyang Liu", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Intern" }, { "name": "Yilin He", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" } ], "n_formula": 50, "n_formula_1": 23, "n_ref_uni": 59, "n_ref": 166, "n_ref_all": 198, "n_fig": 10, "n_tab": 9, "L_tab": 4533, "n_element_tab": 571, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2185, "n_element_tab_1": 332, "formula_len_all": 4601, "formula_len_all_1": 2119, "len_all": 186386, "len_all_1": 70246, "len_abs": 1287, "len_title": 97, "len_sents": 51145, "len_sents_1": 30352, "n_sents": 356, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1207, "title": "Advancing Graph Generation through Beta Diffusion", "abs": "Diffusion models have excelled in generating natural images and are now being adapted to a variety of data types, including graphs. However, conventional models often rely on Gaussian or categorical diffusion processes, which can struggle to accommodate the mixed discrete and continuous components characteristic of graph data. Graphs typically feature discrete structures and continuous node attributes that often exhibit rich statistical patterns, including sparsity, bounded ranges, skewed distributions, and long-tailed behavior. To address these challenges, we introduce Graph Beta Diffusion (GBD), a generative model specifically designed to handle the diverse nature of graph data. GBD leverages a beta diffusion process, effectively modeling both continuous and discrete elements. Additionally, we propose a modulation technique that enhances the realism of generated graphs by stabilizing critical graph topology while maintaining flexibility for other components. GBD competes strongly with existing models across multiple general and biochemical graph benchmarks, showcasing its ability to capture the intricate balance between discrete and continuous features inherent in real-world graph data.", "keywords": [ "Graph Machine Learning", "Generative Models", "Denoising Diffusion Probabilistic Models" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 5, 4, 3, 2 ] }, { "paper_id": "x0h4H1WHXk", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Avi Semler", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "MS student" }, { "name": "Matthias Englert", "gender": "unknown", "institution": "University of Warwick", "country": "", "position": "Associate Professor" }, { "name": "Ranko Lazic", "gender": "Male", "institution": "University of Warwick", "country": "GB", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 23, "n_ref": 45, "n_ref_all": 67, "n_fig": 5, "n_tab": 1, "L_tab": 259, "n_element_tab": 78, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 259, "n_element_tab_1": 78, "formula_len_all": 100, "formula_len_all_1": 100, "len_all": 105261, "len_all_1": 64289, "len_abs": 1021, "len_title": 126, "len_sents": 39464, "len_sents_1": 31662, "n_sents": 277, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1024, "title": "Image Restoration for Training Data Reconstructed from Trained Neural Networks", "abs": "Haim et al. [NeurIPS 2022] propose a method to reconstruct training data from trained neural networks with impressive results. While their reconstructed images resemble the original training images, most of them also contain a considerable amount of noise and artifacts. This is especially true, when the network was trained on more than just a few dozen images. To address this, we view the problem as a specific image restoration task. Since the noise and artifacts are different from other types of noise (Gaussian noise, compression artifacts, blurring, or impulse noise from digital cameras), we create a new dataset specifically for the restoration of images produced by the reconstruction process proposed by Haim et al. We use this dataset consisting of about 60 million noisy reconstructions of CIFAR-10 images to train a diffusion model on the restoration task. Using this method, we obtain reconstructions that are significantly closer to the original training images measured in terms of SSIM and HaarPSI scores.", "keywords": [ "image restoration", "diffusion", "privacy attacks", "dataset reconstruction" ], "rating_list": [ 5, 3, 3, 1 ], "soundness_list": [ 3, 1, 1, 2 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 2, 4, 5, 5 ] }, { "paper_id": "x07rHuChwF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Jiarui Zhang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Jinyi Hu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ollie Liu", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD" }, { "name": "Tianyu Yu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Willie Neiswanger", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 47, "n_ref": 130, "n_ref_all": 149, "n_fig": 7, "n_tab": 7, "L_tab": 3445, "n_element_tab": 198, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1278, "n_element_tab_1": 111, "formula_len_all": 95, "formula_len_all_1": 55, "len_all": 176637, "len_all_1": 64008, "len_abs": 1423, "len_title": 134, "len_sents": 41373, "len_sents_1": 30346, "n_sents": 296, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1430, "title": "Euclid: Supercharging Multimodal LLMs with Synthetic High-Fidelity Visual Descriptions", "abs": "Multimodal large language models (MLLMs) have made rapid progress in recent years, yet continue to struggle with low-level visual perception\u2014particularly the ability to accurately describe the geometric details of an image. This capability is crucial for applications in areas such as robotics, medical image analysis, and manufacturing. To address this challenge, we first introduce Geoperception, a benchmark designed to evaluate an MLLM\u2019s ability to accurately transcribe 2D geometric information from an image. Using this benchmark, we demonstrate the limitations of leading MLLMs, and then conduct a comprehensive empirical study to explore strategies for improving their performance on geometric tasks. Our findings highlight the benefits of certain model architectures, training techniques, and data strategies, including the use of high-fidelity synthetic data and multi-stage training with a data curriculum. Notably, we find that a data curriculum enables models to learn challenging geometry understanding tasks which they fail to learn from scratch. Leveraging these insights, we develop Euclid, a family of models specifically optimized for strong low-level geometric perception. Although purely trained on synthetic multimodal data, Euclid shows strong generalization ability to novel geometry shapes. For instance, Euclid outperforms the best closed-source model, Gemini-1.5-Pro, by up to 54.52% on benchmark tasks.", "keywords": [ "Multimodal LLMs", "Geometric Perception", "Low-level Visual Perception" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "wyF5vNIsO7", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Elon Portugaly", "gender": "unknown", "institution": "Research, Microsoft", "country": "GB", "position": "Principal Researcher" }, { "name": "H. Jabran Zahid", "gender": "unknown", "institution": "Research, Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Ilker Demirel", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Javier Zazo", "gender": "Male", "institution": "Microsoft", "country": "GB", "position": "Researcher" }, { "name": "Julia Greissl", "gender": "unknown", "institution": "Research, Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Lorenzo Pisani", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Paidamoyo Chapfuwa", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 12, "n_ref_uni": 46, "n_ref": 79, "n_ref_all": 151, "n_fig": 20, "n_tab": 2, "L_tab": 389, "n_element_tab": 72, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 315, "n_element_tab_1": 62, "formula_len_all": 826, "formula_len_all_1": 553, "len_all": 156831, "len_all_1": 66500, "len_abs": 1259, "len_title": 127, "len_sents": 49218, "len_sents_1": 30560, "n_sents": 327, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1222, "title": "Scalable Universal T-Cell Receptor Embeddings from Adaptive Immune Repertoires", "abs": "T cells are a key component of the adaptive immune system, targeting infections, cancers, and allergens with specificity encoded by their T cell receptors (TCRs), and retaining a memory of their targets. High-throughput TCR repertoire sequencing captures a cross-section of TCRs that encode the immune history of any subject, though the data are heterogeneous, high dimensional, sparse, and mostly unlabeled. \nSets of TCRs responding to the same antigen, *i.e.*, a protein fragment, co-occur in subjects sharing immune genetics and exposure history. Here, we leverage TCR co-occurrence across a large set of TCR repertoires and employ the GloVe (Pennington et al., 2014) algorithm to derive low-dimensional, dense vector representations (embeddings) of TCRs. We then aggregate these TCR embeddings to generate subject-level embeddings based on observed *subject-specific* TCR subsets. Further, we leverage random projection theory to improve GloVe's computational efficiency in terms of memory usage and training time. Extensive experimental results show that TCR embeddings targeting the same pathogen have high cosine similarity, and subject-level embeddings encode both immune genetics and pathogenic exposure history.", "keywords": [ "Immunomics", "T-cell Receptor Embeddings", "GloVe", "Random Projection Theory", "Scaling", "Unsupervised Representation Learning" ], "rating_list": [ 6, 8, 8, 5 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 4, 4, 2, 5 ] }, { "paper_id": "wy9FRV8O5s", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 9, "author_info_list": [ { "name": "Haotian Xu", "gender": "Male", "institution": "RippleInfo", "country": "CN", "position": "Researcher" }, { "name": "Kaizhu Huang", "gender": "Male", "institution": "Duke Kunshan University", "country": "CN", "position": "Full Professor" }, { "name": "Shiming Chen", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Shreyank N Gowda", "gender": "Male", "institution": "University of Nottingham", "country": "GB", "position": "Assistant Professor" }, { "name": "Xiaobo Jin", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaowei Huang", "gender": "Male", "institution": "University of Liverpool", "country": "GB", "position": "Full Professor" }, { "name": "Yaochu Jin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" }, { "name": "Zihan Ye", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "CN", "position": "PhD student" }, { "name": "Fahad Shahbaz Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" } ], "n_formula": 30, "n_formula_1": 17, "n_ref_uni": 39, "n_ref": 79, "n_ref_all": 114, "n_fig": 15, "n_tab": 3, "L_tab": 1863, "n_element_tab": 226, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 507, "n_element_tab_1": 100, "formula_len_all": 2055, "formula_len_all_1": 1138, "len_all": 164503, "len_all_1": 67293, "len_abs": 3196, "len_title": 118, "len_sents": 40809, "len_sents_1": 27870, "n_sents": 345, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1678, "title": "ZeroDiff: Solidified Visual-semantic Correlation in Zero-Shot Learning", "abs": "Zero-shot Learning (ZSL) aims to enable classifiers to identify unseen classes. This is typically achieved by generating visual features for unseen classes based on learned visual-semantic correlations from seen classes. However, most current generative approaches heavily rely on having a sufficient number of samples from seen classes. Our study reveals that a scarcity of seen class samples results in a marked decrease in performance across many generative ZSL techniques. We argue, quantify, and empirically demonstrate that this decline is largely attributable to spurious visual-semantic correlations. To address this issue, we introduce ZeroDiff, an innovative generative framework for ZSL that incorporates diffusion mechanisms and contrastive representations to enhance visual-semantic correlations. ZeroDiff comprises three key components: (1) Diffusion augmentation, which naturally transforms limited data into an expanded set of noised data to mitigate generative model overfitting; (2) Supervised-contrastive (SC)-based representations that dynamically characterize each limited sample to support visual feature generation; and (3) Multiple feature discriminators employing a Wasserstein-distance-based mutual learning approach, evaluating generated features from various perspectives, including pre-defined semantics, SC-based representations, and the diffusion process. Extensive experiments on three popular ZSL benchmarks demonstrate that ZeroDiff not only achieves significant improvements over existing ZSL methods but also maintains robust performance even with scarce training data. Our codes are available at https://github.com/FouriYe/ZeroDiff_ICLR25.", "keywords": [ "Zero-shot Learning", "Generative Model", "Diffusion Mechanism", "Effective Learning" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "wxPnuFp8fZ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chenxu Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Qingpeng Kong", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "S Kevin Zhou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Zihang Jiang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" } ], "n_formula": 65, "n_formula_1": 25, "n_ref_uni": 59, "n_ref": 155, "n_ref_all": 255, "n_fig": 32, "n_tab": 6, "L_tab": 939, "n_element_tab": 135, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 1315, "n_element_tab_1": 16, "formula_len_all": 5446, "formula_len_all_1": 1989, "len_all": 242230, "len_all_1": 63791, "len_abs": 1037, "len_title": 123, "len_sents": 69416, "len_sents_1": 26179, "n_sents": 542, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1044, "title": "Self-Supervised Diffusion MRI Denoising via Iterative and Stable Refinement", "abs": "Magnetic Resonance Imaging (MRI), including diffusion MRI (dMRI), serves as a ``microscope'' for anatomical structures and routinely mitigates the influence of low signal-to-noise ratio scans by compromising temporal or spatial resolution. However, these compromises fail to meet clinical demands for both efficiency and precision. Consequently, denoising is a vital preprocessing step, particularly for dMRI, where clean data is unavailable. In this paper, we introduce Di-Fusion, a fully self-supervised denoising method that leverages the latter diffusion steps and an adaptive sampling process. Unlike previous approaches, our single-stage framework achieves efficient and stable training without extra noise model training and offers adaptive and controllable results in the sampling process. Our thorough experiments on real and simulated data demonstrate that Di-Fusion achieves state-of-the-art performance in microstructure modeling, tractography tracking, and other downstream tasks. Codes are available in the supplementary material.", "keywords": [ "Diffusion based models", "Self-supervised MRI denoising" ], "rating_list": [ 5, 10, 5, 8, 6 ], "soundness_list": [ 3, 4, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3, 2 ] }, { "paper_id": "wxEASOHHdT", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "Alan Yuille", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Feng Wang", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Guoyizhe Wei", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Jiahao Wang", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Jieru Mei", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Sucheng Ren", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Wei Shao", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Yuyin Zhou", "gender": "unknown", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "cihang xie", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 38, "n_ref": 81, "n_ref_all": 103, "n_fig": 7, "n_tab": 4, "L_tab": 1744, "n_element_tab": 246, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1671, "n_element_tab_1": 245, "formula_len_all": 166, "formula_len_all_1": 166, "len_all": 101142, "len_all_1": 53491, "len_abs": 1390, "len_title": 91, "len_sents": 25810, "len_sents_1": 22768, "n_sents": 177, "n_sents_1": 153, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 1399, "title": "Mamba-Reg: Vision Mamba Also Needs Registers", "abs": "Similar to Vision Transformers, this paper identifies artifacts also present within the feature maps of Vision Mamba. These artifacts, corresponding to high-norm tokens emerging in low-information background areas of images, appear much more severe in Vision Mamba---they exist prevalently even with the tiny-sized model and activate extensively across background regions. To mitigate this issue, we follow the prior solution of introducing register tokens into Vision Mamba. To better cope with Mamba blocks' uni-directional inference paradigm, two key modifications are introduced: 1) evenly inserting registers throughout the input token sequence, and 2) recycling registers for final decision predictions. We term this new architecture MambaReg. Qualitative observations suggest, compared to vanilla Vision Mamba, MambaReg's feature maps appear cleaner and more focused on semantically meaningful regions. Quantitatively, MambaReg attains stronger performance and scales better. For example, on the ImageNet benchmark, our MambaReg-B attains 83.0% accuracy, significantly outperforming Vim-B's 81.8%; furthermore, we provide the first successful scaling to the large model size (i.e., with 341M parameters), attaining a competitive accuracy of 83.6% (84.5% if finetuned with 384x384 inputs). Additional validation on the downstream semantic segmentation task also supports MambaReg's efficacy.", "keywords": [ "State Space Models", "Mamba", "Representation Learning" ], "rating_list": [ 3, 5, 5, 3, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 3, 1, 2 ], "confidence_list": [ 5, 4, 5, 4, 3 ] }, { "paper_id": "wwbVYrOMIW", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Bolin Ni", "gender": "unknown", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Gaofeng Meng", "gender": "Male", "institution": "Institute of automation", "country": "CN", "position": "Full Professor" }, { "name": "Shengchao Zhou", "gender": "Male", "institution": "University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "XIAOJUAN QI", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Youzhi Zhang", "gender": "unknown", "institution": "Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "", "position": "Assistant Professor" }, { "name": "Zhaoxiang Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 36, "n_formula_1": 13, "n_ref_uni": 50, "n_ref": 74, "n_ref_all": 110, "n_fig": 8, "n_tab": 17, "L_tab": 8689, "n_element_tab": 536, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2017, "n_element_tab_1": 153, "formula_len_all": 4848, "formula_len_all_1": 1038, "len_all": 190076, "len_all_1": 63055, "len_abs": 1645, "len_title": 123, "len_sents": 53970, "len_sents_1": 25640, "n_sents": 453, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1699, "title": "POC: Preventing the Over-Collapse of Classes for Class-Incremental Learning", "abs": "Deep neural network-based classification models often suffer from catastrophic forgetting during class-incremental learning (CIL). Previous studies reveal that it results from the overlap between seen and future classes after being mapped by model to its feature space through extracting the features. In this paper, we analyze that this overlap mainly results from the $\\textit{over-collapse}$ of seen classes, where the model tends to map originally separated one seen class and its adjacent regions in input space to be mixed in the feature space, making them indistinguishable. To this end, we propose a two-step framework to $\\textbf{P}$revent the $\\textbf{O}$ver-$\\textbf{C}$ollapse (POC). During training, POC first learns and applies a set of transformations to the training samples of seen classes. Based on our theoretical analysis, the transformation results will locate in the adjacent regions of the seen classes in the input space so that we can let them represent the adjacent regions. Then, the model's optimization objective is modified to additionally classify between the seen classes and the adjacent regions, separating them in model's feature space so that preventing the over-collapse. To retain the model's generalization on the seen classes, a deterministic contrastive loss that makes the separate features of seen classes and adjacent regions close is further introduced. Since POC uses the adjacent regions exclusively for classification, it can be easily adopted by existing CIL methods. Experiments on CIFAR-100 and ImageNet demonstrate that POC effectively increases the last/average incremental accuracy of six SOTA CIL methods by 3.5\\%/3.0\\% on average respectively.", "keywords": [ "Class-Incremental Learning", "Over-Collapse", "Catastrophic Forgetting" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 5 ] }, { "paper_id": "wwXgvjNmt5", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Chongxuan Li", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Dequan Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jiahao Zhan", "gender": "Male", "institution": "Fudan University", "country": "", "position": "Undergrad student" }, { "name": "Jin Gao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 34, "n_ref": 81, "n_ref_all": 106, "n_fig": 10, "n_tab": 7, "L_tab": 1534, "n_element_tab": 172, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1129, "n_element_tab_1": 132, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 179088, "len_all_1": 61778, "len_abs": 1138, "len_title": 133, "len_sents": 55981, "len_sents_1": 29709, "n_sents": 388, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1143, "title": "MAC: A Multimodal Benchmark for Understanding and Generating Academic Journal Covers", "abs": "We introduce the Multimodal Academic Cover (MAC) benchmark to address the challenges of Large Multimodal Models (LMMs) in understanding and generating academic journal covers. While LMMs have demonstrated significant progress in creative arts and everyday applications, their capabilities in comprehending complex academic visuals and narratives remain underexplored. MAC comprises\na collection of 5,872 cover images, accompanying cover stories, and associated articles from 40 prominent academic journals, providing a rich dataset for evaluation. We design bidirectional generative tasks\u2014Image2Text and Text2Imag to assess authenticity and creativity in generating cover images and stories. Current LMMs, including DALL\u00b7E 3, GPT-4V, Gemini, CogView-3, GLM-4V, LLaVA, LLaMA-adapter, and MiniGPT4, are evaluated on this benchmark. Furthermore, we propose Multimodal Agent Linkage (MAL), a novel method to enhance conceptual comprehension within a long-context window. In-context learning techniques, such as few-shot learning, are also explored to improve the effectiveness of LMMs. All benchmarks, prompts, and codes will be released publicly.", "keywords": [ "Benchmark", "Multi-modality", "Large Multimodal Models" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "wwVGZRnAYG", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Lin Luo", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Xingjun Ma", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Ge Yi Li", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Postdoc" }, { "name": "Yu-Gang Jiang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yunhan Zhao", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Xiang ZHENG", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 82, "n_ref_all": 111, "n_fig": 8, "n_tab": 9, "L_tab": 5593, "n_element_tab": 1016, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 4916, "n_element_tab_1": 390, "formula_len_all": 758, "formula_len_all_1": 481, "len_all": 168842, "len_all_1": 72035, "len_abs": 1549, "len_title": 136, "len_sents": 46520, "len_sents_1": 30456, "n_sents": 354, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1315, "title": "BlueSuffix: Reinforced Blue Teaming for Vision-Language Models Against Jailbreak Attacks", "abs": "In this paper, we focus on black-box defense for VLMs against jailbreak attacks.\nExisting black-box defense methods are either unimodal or bimodal. Unimodal methods enhance either the vision or language module of the VLM, while bimodal methods robustify the model through text-image representation realignment. \nHowever, these methods suffer from two limitations: 1) they fail to fully exploit the cross-modal information, or 2) they degrade the model performance on benign inputs.\nTo address these limitations, we propose a novel blue-team method BlueSuffix that defends target VLMs against jailbreak attacks without compromising its performance under black-box setting. BlueSuffix includes three key components: 1) a visual purifier against jailbreak images, 2) a textual purifier against jailbreak texts, and 3) a blue-team suffix generator using reinforcement fine-tuning for enhancing cross-modal robustness. We empirically show on four VLMs (LLaVA, MiniGPT-4, InstructionBLIP, and Gemini) and four safety benchmarks (Harmful Instruction, AdvBench, MM-SafetyBench, and RedTeam-2K) that BlueSuffix outperforms the baseline defenses by a significant margin. \nOur BlueSuffix opens up a promising direction for defending VLMs against jailbreak attacks. Code is available at https://github.com/Vinsonzyh/BlueSuffix.", "keywords": [ "Jailbreak Defense", "Blue-Teaming", "Large Vision-Language Model" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 5, 3, 5 ] }, { "paper_id": "wwO8qS9tQl", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Scott Emmons", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Shiye Su", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "MS student" }, { "name": "Stuart Russell", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Edmund M Mills", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "Intern" } ], "n_formula": 11, "n_formula_1": 4, "n_ref_uni": 73, "n_ref": 134, "n_ref_all": 180, "n_fig": 26, "n_tab": 4, "L_tab": 4586, "n_element_tab": 173, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 282, "n_element_tab_1": 30, "formula_len_all": 388, "formula_len_all_1": 181, "len_all": 332978, "len_all_1": 64334, "len_abs": 1349, "len_title": 118, "len_sents": 91122, "len_sents_1": 30799, "n_sents": 738, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1356, "title": "ALMANACS: A Simulatability Benchmark for Language Model Explainability", "abs": "How do we measure the efficacy of language model explainability methods? While many explainability methods have been developed, they are typically evaluated on bespoke tasks, preventing an apples-to-apples comparison. To help fill this gap, we present ALMANACS, a language model explainability benchmark. ALMANACS scores explainability methods on simulatability, i.e., how well the explanations improve behavior prediction on new inputs. The ALMANACS scenarios span twelve safety-relevant topics such as ethical reasoning and advanced AI behaviors; they have idiosyncratic premises to invoke model-specific behavior; and they have a train-test distributional shift to encourage faithful explanations. By using another language model to predict behavior based on the explanations, ALMANACS is a fully automated benchmark. While not a replacement for human evaluations, we aim for ALMANACS to be a complementary, automated tool that allows for fast, scalable evaluation. Using ALMANACS, we evaluate counterfactual, rationalization, attention, and Integrated Gradients explanations. Our results are sobering: when averaged across all topics, no explanation method outperforms the explanation-free control. We conclude that despite modest successes in prior work, developing an explanation method that aids simulatability in ALMANACS remains an open challenge.", "keywords": [ "explainability", "interpretability", "simulatability", "explanations", "evaluation", "benchmark", "natural language processing" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ww7JqIf494", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Amisha Bhaskar", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Anukriti Singh", "gender": "Female", "institution": "University of Maryland, College Park", "country": "", "position": "MS student" }, { "name": "Peihong Yu", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Pratap Tokekar", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Associate Professor" }, { "name": "Zahiruddin Mahammad", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 53, "n_ref_all": 91, "n_fig": 21, "n_tab": 4, "L_tab": 944, "n_element_tab": 95, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 1076, "n_element_tab_1": 17, "formula_len_all": 367, "formula_len_all_1": 233, "len_all": 135081, "len_all_1": 58241, "len_abs": 1407, "len_title": 131, "len_sents": 49272, "len_sents_1": 28519, "n_sents": 357, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1425, "title": "Sketch-to-Skill: Bootstrapping Robot Learning with Human Drawn Trajectory Sketches", "abs": "Training robotic manipulation policies traditionally requires numerous demonstrations and/or environmental rollouts. While recent Imitation Learning (IL) and Reinforcement Learning (RL) methods have reduced the number of required demonstrations, they still rely on expert knowledge to collect high-quality data, limiting scalability and accessibility. We propose Sketch-to-Skill, a novel framework that leverages human-drawn 2D sketch trajectories to bootstrap and guide RL for robotic manipulation. Our approach extends beyond previous sketch-based methods, which were primarily focused on imitation learning or policy conditioning, limited to specific trained tasks. Sketch-to-Skill employs a Sketch-to-3D Trajectory Generator that translates 2D sketches into 3D trajectories, which are then used to autonomously collect initial demonstrations. We utilize these sketch-generated demonstrations in two ways: to pre-train an initial policy through behavior cloning and to refine this policy through RL with guided exploration. Experimental results demonstrate that Sketch-to-Skill achieves $\\sim$96\\% of the performance of the baseline model that leverages teleoperated demonstration data, while exceeding the performance of a pure reinforcement learning policy by $\\sim$170\\%, only from sketch inputs. This makes robotic manipulation learning more accessible and potentially broadens its applications across various domains.", "keywords": [ "robotics", "learn from demonstration", "reinforcement learning" ], "rating_list": [ 5, 8, 8, 3, 5 ], "soundness_list": [ 1, 3, 2, 3, 2 ], "presentation_list": [ 3, 4, 3, 3, 2 ], "contribution_list": [ 1, 3, 3, 1, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "ww3CLRhF1v", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Antonio Orvieto", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen, Max Planck Institute for Intelligent Systems, T\u00fcbingen AI Center, T\u00fcbingen, Germany", "country": "CH", "position": "Principal Researcher" }, { "name": "Aurelien Lucchi", "gender": "Male", "institution": "University of Basel", "country": "CH", "position": "Assistant Professor" }, { "name": "Enea Monzio Compagnoni", "gender": "Male", "institution": "University of Basel", "country": "CH", "position": "PhD student" }, { "name": "Frank Norbert Proske", "gender": "Male", "institution": "University of Oslo", "country": "", "position": "Full Professor" }, { "name": "Rustem Islamov", "gender": "Male", "institution": "University of Basel", "country": "CH", "position": "PhD student" }, { "name": "Tianlin Liu", "gender": "Male", "institution": "Google DeepMind", "country": "FR", "position": "Research Scientist" } ], "n_formula": 346, "n_formula_1": 34, "n_ref_uni": 107, "n_ref": 210, "n_ref_all": 281, "n_fig": 15, "n_tab": 6, "L_tab": 194, "n_element_tab": 9, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 47, "n_element_tab_1": 4, "formula_len_all": 30406, "formula_len_all_1": 2428, "len_all": 384519, "len_all_1": 62643, "len_abs": 1243, "len_title": 132, "len_sents": 101430, "len_sents_1": 27094, "n_sents": 1090, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1248, "title": "Adaptive Methods through the Lens of SDEs: Theoretical Insights on the Role of Noise", "abs": "Despite the vast empirical evidence supporting the efficacy of adaptive optimization methods in deep learning, their theoretical understanding is far from complete. This work introduces novel SDEs for commonly used adaptive optimizers: SignSGD, RMSprop(W), and Adam(W). These SDEs offer a quantitatively accurate description of these optimizers and help illuminate an intricate relationship between adaptivity, gradient noise, and curvature. Our novel analysis of SignSGD highlights a noteworthy and precise contrast to SGD in terms of convergence speed, stationary distribution, and robustness to heavy-tail noise. We extend this analysis to AdamW and RMSpropW, for which we observe that the role of noise is much more complex. Crucially, we support our theoretical analysis with experimental evidence by verifying our insights: this includes numerically integrating our SDEs using Euler-Maruyama discretization on various neural network architectures such as MLPs, CNNs, ResNets, and Transformers. Our SDEs accurately track the behavior of the respective optimizers, especially when compared to previous SDEs derived for Adam and RMSprop. We believe our approach can provide valuable insights into best training practices and novel scaling rules.", "keywords": [ "Stochastic Differential Equations", "Stochastic Optimization", "Adaptive Methods" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 1, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "wvFnqVVUhN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 16, "author_info_list": [ { "name": "Brando Miranda", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Cristobal Eyzaguirre", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" }, { "name": "Dan Valentine", "gender": "Male", "institution": "Seri Mats", "country": "", "position": "Researcher" }, { "name": "Ethan Perez", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Henry Sleight", "gender": "Male", "institution": "Constellation", "country": "US", "position": "Researcher" }, { "name": "James Chua", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Joe Benton", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "John Hughes", "gender": "Male", "institution": "ML Alignment & Theory Scholars", "country": "GB", "position": "Researcher" }, { "name": "Mrinank Sharma", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Rajashree Agrawal", "gender": "Female", "institution": "Reed College", "country": "US", "position": "Undergrad student" }, { "name": "Rylan Schaeffer", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" }, { "name": "Scott Emmons", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Tony Tong Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Zane Durante", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Luke James Bailey", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 63, "n_ref": 188, "n_ref_all": 218, "n_fig": 16, "n_tab": 8, "L_tab": 3063, "n_element_tab": 201, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 117, "formula_len_all_1": 0, "len_all": 257384, "len_all_1": 52359, "len_abs": 2024, "len_title": 125, "len_sents": 50479, "len_sents_1": 26035, "n_sents": 339, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1770, "title": "Failures to Find Transferable Image Jailbreaks Between Vision-Language Models", "abs": "The integration of new modalities into frontier AI systems offers exciting capabilities, but also increases the possibility such systems can be adversarially manipulated in undesirable ways. In this work, we focus on a popular class of vision-language models (VLMs) that generates text outputs conditioned on visual and textual inputs. We conducted a large-scale empirical study to assess the transferability of gradient-based universal image \"jailbreaks\" using a diverse set of over 40 open-parameter VLMs, including 18 new VLMs that we publicly release.\nOverall, we find that transferable gradient-based image jailbreaks are extremely difficult to obtain. When an image jailbreak is optimized against a single VLM or against an ensemble of VLMs, the jailbreak successfully jailbreaks the attacked VLM(s), but exhibits little-to-no transfer to any other VLMs; transfer is not affected by whether the attacked and target VLMs possess matching vision backbones or language models, whether the language model underwent instruction-following and/or safety-alignment training, or other factors.\nOnly two settings display partially successful transfer: between identically-pretrained and identically-initialized VLMs with slightly different VLM training data, and between different training checkpoints of a single VLM. Leveraging these results, we then demonstrate that transfer can be significantly improved against a specific target VLM by attacking larger ensembles of ``highly-similar\" VLMs. These results stand in stark contrast to existing evidence of universal and transferable text jailbreaks against language models and transferable adversarial attacks against image classifiers, suggesting that VLMs may be more robust to standard gradient-based transfer attacks.", "keywords": [ "adversarial robustness", "jailbreaks", "vision-language model", "multimodal", "adversarial attack", "image jailbreak", "safety", "trustworthy", "robustness" ], "rating_list": [ 8, 3, 6, 8 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "wtrDLMFU9v", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Fangda Guo", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Guoxin Chen", "gender": "unknown", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Wenzheng Feng", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Principal Researcher" }, { "name": "Xin Cong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Yankai Lin", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "Assistant Professor" }, { "name": "Yasheng Wang", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "", "position": "Researcher" }, { "name": "Yesai Wu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Intern" }, { "name": "Zhong Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 121, "n_ref_all": 159, "n_fig": 7, "n_tab": 9, "L_tab": 1600, "n_element_tab": 179, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 645, "n_element_tab_1": 65, "formula_len_all": 549, "formula_len_all_1": 278, "len_all": 218369, "len_all_1": 67047, "len_abs": 1673, "len_title": 97, "len_sents": 66642, "len_sents_1": 32317, "n_sents": 492, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1123, "title": "Learning Evolving Tools for Large Language Models", "abs": "Tool learning enables large language models (LLMs) to interact with external tools and APIs, greatly expanding the application scope of LLMs. However, due to the dynamic nature of external environments, these tools and APIs may become outdated over time, preventing LLMs from correctly invoking tools. Existing research primarily focuses on static environments and overlooks this issue, limiting the adaptability of LLMs in real-world applications. In this paper, we propose ToolEVO, a novel framework designed to enhance the adaptive and reflective capabilities of LLMs against tool variability. By leveraging Monte Carlo Tree Search, ToolEVO facilitates active exploration and interaction of LLMs within dynamic environments, allowing for autonomous self-reflection and self-updating of tool usage based on environmental feedback. Additionally, we introduce ToolQA-D, a benchmark specifically designed to evaluate the impact of tool variability. Extensive experiments demonstrate the effectiveness and stability of our approach, highlighting the importance of adaptability to tool variability for effective tool learning.", "keywords": [ "Tool Learning", "Monte Calro Tree Search", "Large Language Models" ], "rating_list": [ 6, 3, 6, 1 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "wtNxByjLW3", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Anxiang Zeng", "gender": "Male", "institution": "Shopee", "country": "SG", "position": "Researcher" }, { "name": "Cong Fu", "gender": "Male", "institution": "Shopee Pte. Ltd", "country": "SG", "position": "Researcher" }, { "name": "Tingyang Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Xiangyu Ke", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yabo Ni", "gender": "unknown", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" }, { "name": "Yunjun Gao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 88, "n_ref_all": 111, "n_fig": 11, "n_tab": 2, "L_tab": 395, "n_element_tab": 50, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 267, "n_element_tab_1": 39, "formula_len_all": 959, "formula_len_all_1": 850, "len_all": 153571, "len_all_1": 62613, "len_abs": 1340, "len_title": 128, "len_sents": 56194, "len_sents_1": 30135, "n_sents": 444, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1346, "title": "Identify Dominators: The Key To Improve Large-Scale Maximum Inner Product Search", "abs": "Maximum Inner Product Search (MIPS) is essential for machine learning and information retrieval, particularly in applications that operate on high-dimensional data, such as recommender systems and retrieval-augmented generation (RAG), using inner product or cosine similarity. While numerous techniques have been developed for efficient MIPS, their performance often suffers due to a limited understanding of the geometric properties of Inner Product (IP) space. Many approaches reduce MIPS to Nearest Neighbor Search (NNS) through nonlinear transformations, which rely on strong assumptions and can hinder performance. To address these limitations, we propose a novel approach that directly leverages the geometry of IP space. We focus on a class of special vectors called dominators and introduce the Monotonic Relative Dominator Graph MRDG, an IP-space-native, sparse, and strongly-connected graph designed for efficient MIPS, offering theoretical solid foundations. To ensure scalability, we further introduce the Approximate Relative Dominator Graph (ARDG), which retains MRDG\u2019s benefits while significantly reducing indexing complexity. Extensive experiments on 8 public datasets demonstrate that ARDG achieves a 30% average speedup in search at high precision and reduces index size by 2x compared to state-of-the-art graph-based methods.", "keywords": [ "high-dimensional vector", "information retrieval", "vector based retrieval", "graph methods", "nearest neighbor", "maximum inner product search", "similarity search" ], "rating_list": [ 6, 1, 6, 5, 6 ], "soundness_list": [ 4, 1, 3, 2, 3 ], "presentation_list": [ 4, 2, 3, 2, 2 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 4, 5, 5, 4, 4 ] }, { "paper_id": "wtMh0PxDPO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Baoyi An", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Bin Chen", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Niu Lian", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Shu-Tao Xia", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Tao Dai", "gender": "Male", "institution": "Department of Software Engineering, Shenzhen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xin Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yujun Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 12, "n_ref_uni": 38, "n_ref": 79, "n_ref_all": 100, "n_fig": 10, "n_tab": 3, "L_tab": 1401, "n_element_tab": 269, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1290, "n_element_tab_1": 222, "formula_len_all": 6158, "formula_len_all_1": 1389, "len_all": 132185, "len_all_1": 58345, "len_abs": 1147, "len_title": 139, "len_sents": 37478, "len_sents_1": 24382, "n_sents": 301, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1152, "title": "3D-GP-LMVIC: Learning-based Multi-View Image Compression with 3D Gaussian Geometric Priors", "abs": "Multi-view image compression is vital for 3D-related applications. Existing methods often rely on 2D projection similarities between views to estimate disparity, performing well with small disparities, such as in stereo images, but struggling with more complex disparities from wide-baseline setups, common in virtual reality and autonomous driving systems. To overcome this limitation, we propose a novel approach: learning-based multi-view image compression with 3D Gaussian geometric priors (3D-GP-LMVIC). Our method leverages 3D Gaussian Splatting to derive geometric priors of the 3D scene, enabling more accurate disparity estimation between views within the compression model. Additionally, we introduce a depth map compression model to reduce redundancy in geometric information across views. A multi-view sequence ordering method is also proposed to enhance correlations between adjacent views. Experimental results demonstrate that 3D-GP-LMVIC surpasses both traditional and learning-based methods in performance, while maintaining fast encoding and decoding speed. The code is available at https://anonymous.4open.science/r/3D-GP-LMVIC-8FFA.", "keywords": [ "Multi-View Image Compression; 3D Gaussian Splatting; Deep Learning" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "wsb9GNh1Oi", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Elad Sharony", "gender": "unknown", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "PhD student" }, { "name": "Heng Yang", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" }, { "name": "Marco Pavone", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Director, Autonomous Vehicle Research" }, { "name": "Peter Karkus", "gender": "Male", "institution": "NVIDIA", "country": "CH", "position": "Researcher" }, { "name": "Shie Mannor", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Full Professor" }, { "name": "Tong Che", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 55, "n_ref_all": 77, "n_fig": 8, "n_tab": 5, "L_tab": 2672, "n_element_tab": 232, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1697, "n_element_tab_1": 121, "formula_len_all": 624, "formula_len_all_1": 477, "len_all": 163168, "len_all_1": 56736, "len_abs": 1215, "len_title": 109, "len_sents": 53671, "len_sents_1": 26153, "n_sents": 408, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1218, "title": "Learning Multiple Initial Solutions to Optimization Problems", "abs": "Sequentially solving similar optimization problems under strict runtime constraints is essential for many applications, such as robot control, autonomous driving, and portfolio management. The performance of local optimization methods in these settings is sensitive to the initial solution: poor initialization can lead to slow convergence or suboptimal solutions. To address this challenge, we propose learning to predict multiple diverse initial solutions given parameters that define the problem instance. We introduce two strategies for utilizing multiple initial solutions: (i) a single-optimizer approach, where the most promising initial solution is chosen using a selection function, and (ii) a multiple-optimizers approach, where several optimizers, potentially run in parallel, are each initialized with a different solution, with the best solution chosen afterward. We validate our method on three optimal control benchmark tasks: cart-pole, reacher, and autonomous driving, using different optimizers: DDP, MPPI, and iLQR. We find significant and consistent improvement with our method across all evaluation settings and demonstrate that it efficiently scales with the number of initial solutions required.", "keywords": [ "optimization", "initialization", "optimal control", "robotics", "autonomous driving" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 3, 3 ] }, { "paper_id": "wsWCVrH9dv", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Anneke Wernerfelt", "gender": "Female", "institution": "University of Pennsylvania", "country": "", "position": "MS student" }, { "name": "Berk Ustun", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Seung Hyun Cheon", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Sorelle Friedler", "gender": "unknown", "institution": "Haverford College", "country": "US", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 40, "n_ref": 73, "n_ref_all": 90, "n_fig": 3, "n_tab": 6, "L_tab": 1536, "n_element_tab": 86, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1028, "n_element_tab_1": 70, "formula_len_all": 1500, "formula_len_all_1": 1544, "len_all": 155069, "len_all_1": 70453, "len_abs": 1206, "len_title": 71, "len_sents": 43714, "len_sents_1": 32892, "n_sents": 318, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1227, "title": "Feature Responsiveness Scores: Model-Agnostic Explanations for Recourse", "abs": "Machine learning models are often used to automate or support decisions in applications such as lending and hiring. In such settings, consumer protection rules mandate that we provide consumers who receive adverse decisions with a list of \"principal reasons.\" In practice, lenders and employers identify principal reasons as the top-scoring features from a *feature attribution* method. In this work, we study how such practices align with one of the underlying goals of consumer protection -- *recourse* -- i.e., educating individuals on how to achieve a desired outcome. We show that standard attribution methods can highlight features that will not lead to recourse -- providing individuals with *reasons without recourse*. We propose to score features on the basis of *responsiveness*, i.e., the proportion of interventions that can lead to a desired outcome. We develop efficient methods to compute responsiveness scores for any model and any dataset under complex actionability constraints. We present an empirical study on the responsiveness of explanations in lending, and demonstrate how responsiveness scores can highlight features that support recourse and mitigate harm by flagging instances with fixed predictions.", "keywords": [ "explainability", "feature attribution", "algorithmic recourse", "regulation" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 2, 2, 4, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "ws5phQki00", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Maike Behrendt", "gender": "Female", "institution": "Heinrich-Heine Universit\u00e4t D\u00fcsseldorf", "country": "DE", "position": "PhD student" }, { "name": "Marc Ziegele", "gender": "Male", "institution": "Heinrich-Heine Universit\u00e4t D\u00fcsseldorf", "country": "DE", "position": "Associate Professor" }, { "name": "Stefan Harmeling", "gender": "Unspecified", "institution": "Technische Universit\u00e4t Dortmund", "country": "", "position": "Full Professor" }, { "name": "Stefan Sylvius Wagner", "gender": "Male", "institution": "University of D\u00fcsseldorf", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 34, "n_ref": 57, "n_ref_all": 91, "n_fig": 12, "n_tab": 9, "L_tab": 4925, "n_element_tab": 336, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 607, "n_element_tab_1": 62, "formula_len_all": 169, "formula_len_all_1": 155, "len_all": 191057, "len_all_1": 68859, "len_abs": 2107, "len_title": 143, "len_sents": 67398, "len_sents_1": 33100, "n_sents": 469, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 94, "L_abs": 1719, "title": "The Power of LLM-Generated Synthetic Data for Stance Detection in Online Political Discussions", "abs": "Stance detection holds great potential to improve online political discussions through its deployment in discussion platforms for purposes such as content moderation, topic summarisation or to facilitate more balanced discussions. Typically, transformer-based models are employed directly for stance detection, requiring vast amounts of data. However, the wide variety of debate topics in online political discussions makes data collection particularly challenging. LLMs have revived stance detection, but their online deployment in online political discussions faces challenges like inconsistent outputs, biases, and vulnerability to adversarial attacks. We show how LLM-generated synthetic data can improve stance detection for online political discussions by using reliable traditional stance detection models for online deployment, while leveraging the text generation capabilities of LLMs for synthetic data generation in a secure offline environment. To achieve this, (i) we generate synthetic data for specific debate questions by prompting a Mistral-7B model and show that fine-tuning with the generated synthetic data can substantially improve the performance of stance detection, while remaining interpretable and aligned with real world data. (ii) Using the synthetic data as a reference, we can improve performance even further by identifying the most informative samples in an unlabelled dataset, i.e., those samples which the stance detection model is most uncertain about and can benefit from the most. By fine-tuning with both synthetic data and the most informative samples, we surpass the performance of the baseline model that is fine-tuned on all true labels, while labelling considerably less data.", "keywords": [ "large language models", "stance detection", "data augmentation", "active learning", "online political discussions" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 2, 3 ] }, { "paper_id": "wryFCrWB0A", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Baobao Chang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Haozhe Zhao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Jinze Bai", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Junyang Lin", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Liang Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Sinan Tan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Tianyu Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Staff Engineer" }, { "name": "Weichu Xie", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yichi Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Zefan Cai", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 38, "n_ref": 133, "n_ref_all": 169, "n_fig": 17, "n_tab": 3, "L_tab": 885, "n_element_tab": 106, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 2118, "n_element_tab_1": 283, "formula_len_all": 242, "formula_len_all_1": 212, "len_all": 145465, "len_all_1": 54735, "len_abs": 513, "len_title": 124, "len_sents": 37022, "len_sents_1": 20712, "n_sents": 286, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 124, "L_abs": 1230, "title": "A Spark of Vision-Language Intelligence: 2-Dimensional Autoregressive Transformer for Efficient Finegrained Image Generation", "abs": "This work tackles the information loss bottleneck of vector-quantization (VQ) autoregressive image generation by introducing a novel model architecture called the 2-Dimensional Autoregression (DnD) Transformer. The DnD-Transformer predicts more codes for an image by introducing a new autoregression direction, \\textit{model depth}, along with the sequence length direction. Compared to traditional 1D autoregression and previous work utilizing similar 2D image decomposition such as RQ-Transformer, the DnD-Transformer is an end-to-end model that can generate higher quality images with the same backbone model size and sequence length, opening a new optimization perspective for autoregressive image generation. Furthermore, our experiments reveal that the DnD-Transformer's potential extends beyond generating natural images. It can even generate images with rich text and graphical elements in a self-supervised manner, demonstrating an understanding of these combined modalities. This has not been previously demonstrated for popular vision generative models such as diffusion models, showing a spark of vision-language intelligence when trained solely on images. We will open the codes, datasets and models for reproduction.", "keywords": [ "Autoregressive Model", "Image Generation", "Vision-Language Model", "Large Language Model" ], "rating_list": [ 5, 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 2, 4 ], "presentation_list": [ 3, 3, 3, 2, 4 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "wrXCIsysqB", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "De Wen Soh", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" }, { "name": "Hossein Rahmani", "gender": "Male", "institution": "Lancaster University", "country": "GB", "position": "Full Professor" }, { "name": "Jun Liu", "gender": "Male", "institution": "Lancaster University", "country": "GB", "position": "Full Professor" }, { "name": "Na Zhao", "gender": "Female", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" }, { "name": "Qihao Zhao", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Shuyi Jiang", "gender": "Female", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 20, "n_ref_uni": 39, "n_ref": 101, "n_ref_all": 117, "n_fig": 8, "n_tab": 5, "L_tab": 755, "n_element_tab": 67, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 591, "n_element_tab_1": 58, "formula_len_all": 1588, "formula_len_all_1": 1400, "len_all": 134051, "len_all_1": 64567, "len_abs": 1379, "len_title": 147, "len_sents": 39752, "len_sents_1": 29928, "n_sents": 275, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 98, "L_abs": 1386, "title": "GaussianBlock: Building Part-Aware Compositional and Editable 3D Scene by Primitives and Gaussians", "abs": "Recently, with the development of Neural Radiance Fields and Gaussian Splatting, 3D reconstruction techniques have achieved remarkably high fidelity. However, the latent representations learnt by these methods are highly entangled and lack interpretability. In this paper, we propose a novel part-aware compositional reconstruction method, called GaussianBlock, that enables semantically coherent and disentangled representations, allowing for precise and physical editing akin to building blocks, while simultaneously maintaining high fidelity.\nOur GaussianBlock introduces a hybrid representation that leverages the advantages of both primitives, known for their flexible actionability and editability, and 3D Gaussians, which excel in reconstruction quality. Specifically, we achieve semantically coherent primitives through a novel attention-guided centering loss derived from 2D semantic priors, complemented by a dynamic splitting and fusion strategy. \nFurthermore, we utilize 3D Gaussians that hybridize with primitives to refine structural details and enhance fidelity. \nAdditionally, a binding inheritance strategy is employed to strengthen and maintain the connection between the two. \nOur reconstructed scenes are evidenced to be disentangled, compositional, and compact across diverse benchmarks, enabling seamless, direct and precise editing while maintaining high quality.", "keywords": [ "3D Decompostion", "3D Reconstruction", "3D Editing" ], "rating_list": [ 5, 5, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 2, 4 ] }, { "paper_id": "wrVZ771SZQ", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 1, "author_info_list": [ { "name": "Rui Xue", "gender": "unknown", "institution": "North Carolina State University", "country": "US", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 39, "n_ref": 82, "n_ref_all": 88, "n_fig": 1, "n_tab": 5, "L_tab": 1209, "n_element_tab": 163, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1653, "n_element_tab_1": 92, "formula_len_all": 1009, "formula_len_all_1": 885, "len_all": 119508, "len_all_1": 67478, "len_abs": 1615, "len_title": 122, "len_sents": 37851, "len_sents_1": 32157, "n_sents": 261, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1621, "title": "VISAGNN: Versatile Staleness-Aware Training for Efficient Large-Scale GNNs", "abs": "Graph Neural Networks (GNNs) have shown exceptional success in graph representation learning and a wide range of real-world applications. However, scaling deeper GNNs poses challenges due to the neighbor explosion problem when training on large-scale graphs. To mitigate this, a promising class of GNN training algorithms utilizes historical embeddings to reduce computation and memory costs while preserving the expressiveness of the model. These methods leverage historical embeddings for out-of-batch nodes, effectively approximating full-batch training without losing any neighbor information\u2014a limitation found in traditional sampling methods. However, the staleness of these historical embeddings often introduces significant bias, acting as a bottleneck that can adversely affect model performance. In this paper, we propose a novel VersatIle Staleness-Aware GNN, named VISAGNN, which dynamically and adaptively incorporates staleness criteria into the large-scale GNN training process. By embedding staleness into the message-passing mechanism, loss function, and historical embeddings during training, our approach enables the model to adaptively mitigate the negative effects of stale embeddings, thereby reducing estimation errors and enhancing downstream accuracy. Comprehensive experiments demonstrate the effectiveness of our method in overcoming the limitations of existing historical embedding techniques, highlighting its superior performance and efficiency on large-scale benchmarks, as well as significantly accelerated convergence. We will make the code publicly available upon acceptance of the work.", "keywords": [ "Graph machine learning", "Large scale GNNs", "Staleness awareness" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "wqA7QmpUwa", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "benyou wang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Chen Zhang", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Junying Chen", "gender": "Male", "institution": "Chinese University of Hong Kong, Shenzhen", "country": "", "position": "PhD student" }, { "name": "Shunian Chen", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "", "position": "Researcher" }, { "name": "Song Dingjie", "gender": "Male", "institution": "Lehigh University", "country": "US", "position": "PhD student" }, { "name": "Xidong Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Zhenyang Cai", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 41, "n_ref": 132, "n_ref_all": 157, "n_fig": 6, "n_tab": 12, "L_tab": 4357, "n_element_tab": 535, "n_fig_1": 3, "n_tab_1": 9, "L_tab_1": 5507, "n_element_tab_1": 464, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 225199, "len_all_1": 70608, "len_abs": 589, "len_title": 233, "len_sents": 36994, "len_sents_1": 25782, "n_sents": 294, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1153, "title": "LongLLaVA: Scaling Multi-modal LLMs to 1000 Images Efficiently via a Hybrid Architecture", "abs": "Expanding the long-context capabilities of Multi-modal Large Language Models (MLLMs) is crucial for video understanding, high-resolution image understanding, and multi-modal agents. This involves a series of systematic optimizations, including model architecture, data construction and training strategy, particularly addressing challenges such as \\textit{degraded performance with more images} and \\textit{high computational costs}. In this paper, we adapt the model architecture to a hybrid of Mamba and Transformer blocks, approach data construction with both temporal and spatial dependencies among multiple images and employ a progressive training strategy. The released model **LongLLaVA** (**Long**-Context **L**arge **L**anguage **a**nd **V**ision **A**ssistant) is the first hybrid MLLM, which achieved a better balance between efficiency and effectiveness. LongLLaVA not only achieves competitive results across various benchmarks, but also maintains high throughput and low memory consumption. Especially, it could process nearly a thousand images on a single A100 80GB GPU, showing promising application prospects for a wide range of tasks.", "keywords": [ "Efficient Multimodal Large Language Model", "Transformer-Mamba Hybrid Architecture" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 5, 5 ] }, { "paper_id": "wq4AeBWQJ4", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 12, "author_info_list": [ { "name": "Amit Hasan", "gender": "Male", "institution": "University of Connecticut", "country": "US", "position": "PhD student" }, { "name": "CUNXI YU", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "", "position": "Assistant Professor" }, { "name": "Caiwen Ding", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "Associate Professor" }, { "name": "Hongwu Peng", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Researcher" }, { "name": "Jiahui Zhao", "gender": "Female", "institution": "University of Connecticut", "country": "US", "position": "PhD student" }, { "name": "Kiran Thorat", "gender": "Male", "institution": "University of Connecticut", "country": "", "position": "PhD student" }, { "name": "Nan Wu", "gender": "Female", "institution": "George Washington University", "country": "US", "position": "Assistant Professor" }, { "name": "Shaoyi Huang", "gender": "Female", "institution": "Stevens Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Xi Xie", "gender": "Male", "institution": "University of Connecticut", "country": "US", "position": "PhD student" }, { "name": "Yingjie Li", "gender": "Female", "institution": "University of Utah", "country": "", "position": "PhD student" }, { "name": "Yuebo Luo", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" }, { "name": "Zhijie Shi", "gender": "Male", "institution": "University of Connecticut", "country": "US", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 17, "n_ref": 41, "n_ref_all": 98, "n_fig": 18, "n_tab": 9, "L_tab": 2522, "n_element_tab": 284, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 238, "n_element_tab_1": 22, "formula_len_all": 310, "formula_len_all_1": 162, "len_all": 117097, "len_all_1": 66112, "len_abs": 1941, "len_title": 149, "len_sents": 39885, "len_sents_1": 31272, "n_sents": 326, "n_sents_1": 270, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 1975, "title": "GROOT: Graph Edge Re-growth and Partitioning for the Verification of Large Designs in Logic Synthesis", "abs": "Traditional verification methods in chip design are highly time-consuming and computationally demanding, especially for large-scale circuits. Graph neural networks (GNNs) have gained popularity as a potential solution to improve verification efficiency. However, there lacks a joint framework that considers all chip design domain knowledge, graph theory, and GPU kernel designs. To address\nthis challenge, we introduce GROOT, an algorithm and system co-design framework that contains chip design domain knowledge, graph theory, and redesigned GPU kernels, to improve verification efficiency. More specifically, we redesign node features utilizing the circuit node types and the polarity of the connections between the input edges to nodes in And-Inverter Graphs (AIGs). We utilize a graph partitioning algorithm based on the observation that approximately only 10% of boundary edges (nodes) between clusters, to divide the large graphs into smaller sub-graphs for fast GPU processing. We carefully profile the EDA graph workloads and observe the uniqueness of their polarized distribution of high-degree (HD) nodes and low-degree (LD) nodes. We redesign two GPU kernels (HD-kernel and LD-kernel), to fit the EDA graph learning workload on a single GPU. We evaluate the performance of GROOT on large circuit designs, e.g., Carry Save Adder (CSA) multipliers, the 7nm technology-mapped CSA multipliers, and Booth Multipliers. We compare the results with state-of-the-art GNN-based GAMORA and the traditional ABC framework. Results show that GROOT achieves a significant reduction in memory footprint (59.38 %), with high accuracy (99.96%) for a very large CSA multiplier, i.e. 1,024 bits with a batch size of 16, which consists of 134,103,040 nodes and 268,140,544 edges. We also compare GROOT with state-of-the-art GPU-based GPU Kernel designs such as cuSPARSE, MergePath-SpMM, and GNNAdvisor. We achieve up to 1.104\u00d7, 5.796\u00d7, and 1.469\u00d7 improvement in runtime, respectively.", "keywords": [ "Graph Neural Networks for EDA", "Logic Synthesis", "Formal Verification" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "wpL3otU9eY", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Aleksander Popel", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "Full Professor" }, { "name": "Jeremias Sulam", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" }, { "name": "Zhenzhen Wang", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 50, "n_ref": 82, "n_ref_all": 103, "n_fig": 4, "n_tab": 3, "L_tab": 955, "n_element_tab": 63, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1194, "n_element_tab_1": 113, "formula_len_all": 815, "formula_len_all_1": 647, "len_all": 129387, "len_all_1": 59743, "len_abs": 1018, "len_title": 109, "len_sents": 37043, "len_sents_1": 26375, "n_sents": 279, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1028, "title": "CBM-zero: Concept Bottleneck Model With Zero Performance Loss", "abs": "Interpreting machine learning models with high-level, human-understandable \\emph{concepts} has gained increasing interest. The concept bottleneck model (CBM) is a popular approach to providing interpretable models, relying on first predicting the presence of concepts in a given input, and then using these concept scores to predict a label of interest. Yet, CBMs suffer from lower accuracy compared with standard black-box models, as they use a surrogate (and thus, interpretable) predictor in lieu of the original model. In this work, we propose an approach that allows us to find a CBM in any standard black-box model via an invertible mapping from its latent space to an interpretable concept space. This method preserves the original black-box model's prediction and thus has zero performance drop while providing human-understandable explanations. We evaluate the accuracy and interpretability of our method across various benchmarks, demonstrating state-of-the-art explainability metrics while enjoying superior accuracy.", "keywords": [ "interpretability", "explainability", "concept bottleneck model" ], "rating_list": [ 3, 3, 1, 5 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 1, 1, 2, 3 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "wozhdnRCtw", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Alessandro Stolfo", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Besmira Nushi", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Eric Horvitz", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Chief Scientific Officer" }, { "name": "Safoora Yousefi", "gender": "Unspecified", "institution": "Microsoft Inc", "country": "", "position": "ML scientist" }, { "name": "Vidhisha Balachandran", "gender": "Female", "institution": "Research, Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 75, "n_ref": 153, "n_ref_all": 230, "n_fig": 39, "n_tab": 10, "L_tab": 3785, "n_element_tab": 218, "n_fig_1": 19, "n_tab_1": 3, "L_tab_1": 700, "n_element_tab_1": 46, "formula_len_all": 180, "formula_len_all_1": 160, "len_all": 306228, "len_all_1": 75241, "len_abs": 1381, "len_title": 126, "len_sents": 77617, "len_sents_1": 36087, "n_sents": 551, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1229, "title": "Improving Instruction-Following in Language Models through Activation Steering", "abs": "The ability to follow instructions is crucial for numerous real-world applications of language models. In pursuit of deeper insights and more powerful capabilities, we derive instruction-specific vector representations from language models and use them to steer models accordingly. These vectors are computed as the difference in activations between inputs with and without instructions, enabling a modular approach to activation steering. We demonstrate how this method can enhance model adherence to constraints such as output format, length, and word inclusion, providing inference-time control over instruction following. Our experiments across four models demonstrate how we can use the activation vectors to guide models to follow constraints even without explicit instructions and to enhance performance when instructions are present. Additionally, we explore the compositionality of activation steering, successfully applying multiple instructions simultaneously. Finally, we demonstrate that steering vectors computed on instruction-tuned models can transfer to improve base models. Our findings demonstrate that activation steering offers a practical and scalable approach for fine-grained control in language generatio", "keywords": [ "Interpretability", "Mechanistic Interpretability", "Instruction-following", "Activation Steering", "LLMs" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "womU9cEwcO", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Chuang Gan", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Delin Chen", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "MS student" }, { "name": "Rui Sun", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Wenjun Liu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Zhenfang Chen", "gender": "Male", "institution": "MIT-IBM Watson AI lab", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 58, "n_ref": 105, "n_ref_all": 144, "n_fig": 17, "n_tab": 13, "L_tab": 5295, "n_element_tab": 405, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 388, "n_element_tab_1": 42, "formula_len_all": 140, "formula_len_all_1": 97, "len_all": 205194, "len_all_1": 66310, "len_abs": 2297, "len_title": 0, "len_sents": 59419, "len_sents_1": 31827, "n_sents": 457, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 2006, "title": "Scaling Autonomous Agents via Automatic Reward Modeling And Planning", "abs": "Large language models (LLMs) have demonstrated remarkable capabilities across a range of text-generation tasks. However, LLMs still struggle with problems requiring multi-step decision-making and environmental feedback, such as online shopping, scientific reasoning, and mathematical problem-solving. Unlike pure text data, collecting large-scale decision-making data is challenging. Moreover, many powerful LLMs are only accessible through APIs, which hinders their fine-tuning for agent tasks due to cost and complexity. To address LLM agents' limitations, we propose a framework that can automatically learn a reward model from the environment without human annotations. This model can be used to evaluate the action trajectories of LLM agents and provide heuristics for task planning. Specifically, our approach involves employing one LLM-based agent to navigate an environment randomly, generating diverse action trajectories. Subsequently, a separate LLM is leveraged to assign a task intent and synthesize a negative response alongside the correct response for each trajectory. These triplets (task intent, positive response, and negative response) are then utilized as training data to optimize a reward model capable of scoring action trajectories. This reward model can be integrated with LLM-based agents and various planning algorithms to enhance task-solving performance. The effectiveness and generalizability of our framework are demonstrated through evaluations conducted on different agent benchmarks. In conclusion, our proposed framework represents a significant advancement in enhancing LLM agents' decision-making capabilities. By automating the learning of reward models, we overcome the challenges of data scarcity and API limitations, potentially revolutionizing the application of LLMs in complex and interactive environments. This research paves the way for more sophisticated AI agents capable of tackling a wide range of real-world problems requiring multi-step decision-making.", "keywords": [ "agents", "large language models", "planning" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 4, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "wojnTvBXqt", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Chengzhi Mao", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Junfeng Yang", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Associate Professor" }, { "name": "Ran Li", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Undergrad student" }, { "name": "Wei Hao", "gender": "unknown", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Weiliang Zhao", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 49, "n_ref": 83, "n_ref_all": 100, "n_fig": 5, "n_tab": 8, "L_tab": 1660, "n_element_tab": 233, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1931, "n_element_tab_1": 223, "formula_len_all": 400, "formula_len_all_1": 293, "len_all": 143177, "len_all_1": 60554, "len_abs": 1345, "len_title": 112, "len_sents": 37922, "len_sents_1": 27341, "n_sents": 258, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1355, "title": "Learning to Rewrite: Generalized Detection of LLM-Generated Text", "abs": "Large language models (LLMs) present significant risks when used to generate non-factual content and spread disinformation at scale. Detecting such LLM-generated content is crucial, yet current detectors often struggle to generalize in open-world contexts. We introduce **Learning2Rewrite**, a novel framework for detecting AI-generated text with exceptional generalization to unseen domains. Our method leverages the insight that LLMs inherently modify AI-generated content less than human-written text when tasked with rewriting. By training LLMs to minimize alterations on AI-generated inputs, we amplify this disparity, yielding a more distinguishable and generalizable edit distance across diverse text distributions. Extensive experiments on data from 21 independent domains and four major LLMs (GPT-3.5, GPT-4, Gemini, and Llama-3) demonstrate that our detector outperforms state-of-the-art detection methods by up to 23.04% in AUROC for in-distribution tests, 37.26% for out-of-distribution tests, and 48.66% under adversarial attacks. Our unique training objective ensures better generalizability compared to directly training for classification, when leveraging the same amount of learned parameters. Our findings suggest that reinforcing LLMs\u2019 inherent rewriting tendencies offers a robust and scalable solution for detecting AI-generated text.", "keywords": [ "LLM-generated text detection", "AIGC detection" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "wnT8bfJCDx", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Ameen Ali Ali", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Itamar Zimerman", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Lior Wolf", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Associate Professor" } ], "n_formula": 23, "n_formula_1": 23, "n_ref_uni": 55, "n_ref": 128, "n_ref_all": 150, "n_fig": 7, "n_tab": 7, "L_tab": 1813, "n_element_tab": 204, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 246, "n_element_tab_1": 33, "formula_len_all": 1954, "formula_len_all_1": 1759, "len_all": 157267, "len_all_1": 54808, "len_abs": 1111, "len_title": 128, "len_sents": 42877, "len_sents_1": 25708, "n_sents": 308, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1115, "title": "Explaining Modern Gated-Linear RNNs via a Unified Implicit Attention Formulation", "abs": "Recent advances in efficient sequence modeling have led to attention-free layers, such as Mamba, RWKV, and various gated RNNs, all featuring sub-quadratic complexity in sequence length and excellent scaling properties, enabling the construction of a new type of foundation models. In this paper, we present a unified view of these models, formulating such layers as implicit causal self-attention layers. The formulation includes most of their sub-components and is not limited to a specific part of the architecture. The framework compares the underlying mechanisms on similar grounds for different layers and provides a direct means for applying explainability methods. Our experiments show that our attention matrices and attribution method outperform an alternative and a more limited formulation that was recently proposed for Mamba. For the other architectures for which our method is the first to provide such a view, our method is effective and competitive in the relevant metrics compared to the results obtained by state-of-the-art Transformer explainability methods. Our code is attached as a supplement.", "keywords": [ "Explainability", "Interpretability", "Gated-Linear RNNs", "Attention-free", "Mamba" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 2, 3, 2 ] }, { "paper_id": "wmmDvZGFK7", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Guangyi Wang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Song-Zhi Su", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Wei Peng", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Yuren Cai", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "lijiang Li", "gender": "unknown", "institution": "Xiamen University", "country": "CN", "position": "MS student" } ], "n_formula": 43, "n_formula_1": 27, "n_ref_uni": 39, "n_ref": 268, "n_ref_all": 375, "n_fig": 16, "n_tab": 12, "L_tab": 5214, "n_element_tab": 534, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 772, "n_element_tab_1": 76, "formula_len_all": 4334, "formula_len_all_1": 2043, "len_all": 220925, "len_all_1": 67204, "len_abs": 1388, "len_title": 136, "len_sents": 73676, "len_sents_1": 29097, "n_sents": 497, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 103, "L_abs": 1394, "title": "PFDiff: Training-free Acceleration of Diffusion Models through the Gradient Guidance of Past and Future", "abs": "Diffusion Probabilistic Models (DPMs) have shown remarkable potential in image generation, but their sampling efficiency is hindered by the need for numerous denoising steps. Most existing solutions accelerate the sampling process by proposing fast ODE solvers. However, the inevitable discretization errors of the ODE solvers are significantly magnified when the number of function evaluations (NFE) is fewer. In this work, we propose PFDiff, a novel training-free and orthogonal timestep-skipping strategy, which enables existing fast ODE solvers to operate with fewer NFE. Specifically, PFDiff initially utilizes score replacement from past time steps to predict a \"springboard\". Subsequently, it employs this \u201cspringboard\u201d along with foresight updates inspired by Nesterov momentum to rapidly update current intermediate states. This approach effectively reduces unnecessary NFE while correcting for discretization errors inherent in first-order ODE solvers. Experimental results demonstrate that PFDiff exhibits flexible applicability across various pre-trained DPMs, particularly excelling in conditional DPMs and surpassing previous state-of-the-art training-free methods. For instance, using DDIM as a baseline, we achieved 16.46 FID (4 NFE) compared to 138.81 FID with DDIM on ImageNet 64x64 with classifier guidance, and 13.06 FID (10 NFE) on Stable Diffusion with 7.5 guidance scale.", "keywords": [ "diffusion models", "accelerated sampling", "training-free sampler", "orthogonal sampling method" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 4, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "wmV4cIbgl6", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Gideon Stein", "gender": "Male", "institution": "Friedrich-Schiller Universit\u00e4t Jena", "country": "", "position": "PhD student" }, { "name": "Jan Blunk", "gender": "unknown", "institution": "Friedrich-Schiller Universit\u00e4t Jena", "country": "DE", "position": "PhD student" }, { "name": "Joachim Denzler", "gender": "Male", "institution": "Friedrich-Schiller-University Jena", "country": "DE", "position": "Professor" }, { "name": "Maha Shadaydeh", "gender": "Female", "institution": "Jena University", "country": "DE", "position": "Principal Researcher" }, { "name": "Niklas Penzel", "gender": "Male", "institution": "Friedrich-Schiller Universit\u00e4t Jena", "country": "DE", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 40, "n_ref_all": 55, "n_fig": 4, "n_tab": 2, "L_tab": 1284, "n_element_tab": 96, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2339, "n_element_tab_1": 229, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 135688, "len_all_1": 65420, "len_abs": 1995, "len_title": 132, "len_sents": 35582, "len_sents_1": 31117, "n_sents": 256, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 85, "L_abs": 2011, "title": "CausalRivers - Scaling up benchmarking of causal discovery for real-world time-series", "abs": "Causal discovery, or identifying causal relationships from observational data, is a notoriously challenging task, with numerous methods proposed to tackle it.\nDespite this, in-the-wild evaluation is still lacking, as works frequently rely on synthetic data evaluation and sparse real-world examples under critical theoretical assumptions. \nReal-world causal structures, however, are often complex, evolving over time, non-linear, and influenced by unobserved factors, making\nit hard for practitioners to select appropriate methods. \nTo bridge this gap, we introduce CausalRivers, the largest in-the-wild causal discovery benchmarking kit for time series data to date.\nCausalRivers features an extensive dataset on river discharge that covers the complete eastern German territory (666 measurement stations) and the state of Bavaria (494 measurement stations). \nIt spans the years 2019 to 2023 with a 15-minute temporal resolution. \nFurther, we provide data from a recent flood around the Elbe River, as an event with a pronounced distributional shift. \nLeveraging multiple sources of information and time-series meta-data, we constructed two distinct causal ground truth graphs (Bavaria and eastern Germany).\nThese graphs can be sampled to generate thousands of subgraphs to benchmark causal discovery across diverse and challenging settings.\nTo demonstrate the utility of our benchmarking kit, we evaluate several causal discovery approaches through multiple experiments and introduce effective baselines, identifying several areas for enhancement.\nCausalRivers has the potential to facilitate robust evaluations and comparisons of causal discovery methods.\nBesides this primary purpose, we also expect that this dataset will be relevant for connected areas of research, such as time series forecasting and anomaly detection.\nBased on this, we hope to establish benchmark-driven method development that fosters advanced techniques for causal discovery, as is the case for many other areas of machine learning.", "keywords": [ "Causal Discovery", "Benchmarking", "Time-series" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 4, 3, 4 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "wmFp2aMhi0", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Abele M\u0103lan", "gender": "unknown", "institution": "Universit\u00e9 de Neuch\u00e2tel", "country": "CH", "position": "PhD student" }, { "name": "Aditya Shankar", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "PhD student" }, { "name": "Chenrui Fan", "gender": "Male", "institution": "Universit\u00e4t Bern", "country": "CH", "position": "MS student" }, { "name": "Lydia Y. Chen", "gender": "Female", "institution": "Delft University of Technology", "country": "", "position": "Associate Professor" }, { "name": "Zhi Wen Soi", "gender": "unknown", "institution": "Universit\u00e4t Bern", "country": "CH", "position": "MS student" } ], "n_formula": 41, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 62, "n_ref_all": 81, "n_fig": 5, "n_tab": 9, "L_tab": 11848, "n_element_tab": 404, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1324, "n_element_tab_1": 89, "formula_len_all": 1949, "formula_len_all_1": 733, "len_all": 168889, "len_all_1": 63213, "len_abs": 1468, "len_title": 122, "len_sents": 48691, "len_sents_1": 28263, "n_sents": 360, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1475, "title": "Federated Time Series Generation on Feature and Temporally Misaligned Data", "abs": "Distributed time series data presents a challenge for federated learning, as clients often possess different feature sets and have misaligned time steps. Existing federated time series models are limited by the assumption of perfect temporal or feature alignment across clients. In this paper, we propose FedTDD, a novel federated time series diffusion model that jointly learns a synthesizer across clients. At the core of FedTDD is a novel data distillation and aggregation framework that reconciles the differences between clients by imputing the misaligned timesteps and features. In contrast to traditional federated learning, FedTDD learns the correlation across clients' time series through the exchange of local synthetic outputs instead of model parameters. A coordinator iteratively improves a global distiller network by leveraging shared knowledge from clients through the exchange of synthetic data. As the distiller becomes more refined over time, it subsequently enhances the quality of the clients' local feature estimates, allowing each client to then improve its local imputations for missing data using the latest, more accurate distiller. Experimental results on five datasets demonstrate FedTDD's effectiveness compared to centralized training, and the effectiveness of sharing synthetic outputs to transfer knowledge of local time series. Notably, FedTDD achieves 79.4% and 62.8% improvement over local training in Context-FID and Correlational scores.", "keywords": [ "time series", "generative model", "federated learning" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "wm5wwAdiEt", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Baoxiang Wang", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Binbin Chen", "gender": "unknown", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Han Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "MS student" }, { "name": "Tieying Zhang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 6, "n_ref_uni": 61, "n_ref": 82, "n_ref_all": 92, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 674, "formula_len_all_1": 410, "len_all": 149120, "len_all_1": 69620, "len_abs": 1213, "len_title": 0, "len_sents": 48296, "len_sents_1": 35658, "n_sents": 329, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1229, "title": "Learning to Communicate Through Implicit Communication Channels", "abs": "Effective communication is an essential component in collaborative multi-agent systems. Situations where explicit messaging is not feasible have been common in human society throughout history, which motivate the study of implicit communication. Previous works on learning implicit communication mostly rely on theory of mind (ToM), where agents infer the mental states and intentions of others by interpreting their actions. However, ToM-based methods become less effective in making accurate inferences in complex tasks. In this work, we propose the Implicit Channel Protocol (ICP) framework, which allows agents to communicate through implicit communication channels similar to the explicit ones. ICP leverages a subset of actions, denoted as the scouting actions, and a mapping between information and these scouting actions that encodes and decodes the messages. We propose training algorithms for agents to message and act, including learning with a randomly initialized information map and with a delayed information map. The efficacy of ICP has been tested on the tasks of Guessing Numbers, Revealing Goals, and Hanabi, where ICP significantly outperforms baseline methods through more efficient information transmission.", "keywords": [ "implicit communication", "multi-agent reinforcement learning", "the Hanabi challenge" ], "rating_list": [ 8, 6, 3, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "wldwEhQ7cl", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Dror Moran", "gender": "unknown", "institution": "Weizmann Institute of Science", "country": "IL", "position": "PhD student" }, { "name": "Fadi Khatib", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "IL", "position": "PhD student" }, { "name": "Meirav Galun", "gender": "Female", "institution": "Weizmann Institute", "country": "IL", "position": "Assistant Professor" }, { "name": "Ronen Basri", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "IL", "position": "Full Professor" }, { "name": "Yoni Kasten", "gender": "unknown", "institution": "NVIDIA", "country": "IL", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 47, "n_ref": 92, "n_ref_all": 110, "n_fig": 2, "n_tab": 9, "L_tab": 4717, "n_element_tab": 495, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1505, "n_element_tab_1": 151, "formula_len_all": 379, "formula_len_all_1": 340, "len_all": 144915, "len_all_1": 65640, "len_abs": 960, "len_title": 94, "len_sents": 38696, "len_sents_1": 28544, "n_sents": 313, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 45, "L_abs": 963, "title": "Robust Deep Equivariant Structure from Motion", "abs": "Multiview Structure from Motion is a fundamental and challenging computer vision problem. A recent deep-based approach utilized matrix equivariant architectures for simultaneous recovery of camera pose and 3D scene structure from large image collections. That work, however, made the unrealistic assumption that the point tracks given as input are almost clean of outliers. Here, we propose an architecture suited to dealing with outliers by adding a multiview inlier/outlier classification module that respects the model equivariance and by utilizing a robust bundle adjustment step. Experiments demonstrate that our method can be applied successfully in realistic settings that include large image collections and point tracks extracted with common heuristics that include many outliers, achieving state-of-the-art accuracies in almost all runs, superior to existing deep-based methods and on-par with leading classical (non-deep) sequential and global methods.", "keywords": [ "3D Reconstruction", "Outlier Removal", "Structure from Motion" ], "rating_list": [ 5, 8, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 1, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "wl4c9jvcyY", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Hongxin Li", "gender": "Male", "institution": "Institute of Automation\uff0cChinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Jingfan CHEN", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Jingran Su", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Qing Li", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Yuntao Chen", "gender": "Male", "institution": "Centre for Artificial Intelligence and Robotics (CAIR), Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "HK", "position": "Assistant Professor" }, { "name": "Zhaoxiang Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 28, "n_ref": 83, "n_ref_all": 114, "n_fig": 14, "n_tab": 9, "L_tab": 3680, "n_element_tab": 412, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 1953, "n_element_tab_1": 285, "formula_len_all": 34, "formula_len_all_1": 34, "len_all": 270386, "len_all_1": 63668, "len_abs": 4466, "len_title": 130, "len_sents": 84674, "len_sents_1": 27804, "n_sents": 648, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1569, "title": "AutoGUI: Scaling GUI Grounding with Automatic Functionality Annotations from LLMs", "abs": "User interface understanding with vision-language models has received much attention due to its potential for enabling next-generation software automation.\nHowever, existing UI datasets either only provide large-scale context-free element annotations or contextualized functional descriptions for elements at a much smaller scale.\nIn this work, we propose the **AutoGUI** pipeline for automatically annotating UI elements with detailed functionality descriptions at scale.\nSpecifically, we leverage large language models (LLMs) to infer element functionality by comparing the UI content changes before and after simulated interactions with specific UI elements. To improve annotation quality, we propose LLM-aided rejection and verification, eliminating invalid and incorrect annotations without human labor.\nWe construct an **AutoGUI-704k** dataset using the proposed pipeline, featuring multi-resolution, multi-device screenshots, diverse data domains, and detailed functionality annotations that have never been provided by previous datasets.\nHuman evaluation shows that the **AutoGUI** pipeline achieves annotation correctness comparable to trained human annotators. Extensive experimental results show that our **AutoGUI-704k** dataset remarkably enhances VLM's UI grounding capabilities, exhibits significant scaling effects, and outperforms existing web pre-training data types. We envision AutoGUI as a scalable pipeline for generating massive data to build GUI-oriented VLMs. AutoGUI dataset can be viewed at this anonymous URL: https://huggingface.co/AutoGUI.", "keywords": [ "Vision language model", "Large language model", "Embodied AI", "GUI understanding", "Web agent" ], "rating_list": [ 3, 3, 8, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "wl1Kup6oES", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Cornelia Fermuller", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "Research Scientist" }, { "name": "Dehao Yuan", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Eadom T Dessalene", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Michael Maynord", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Nathan S. Pan", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "MS student" }, { "name": "Yiannis Aloimonos", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 23, "n_ref": 41, "n_ref_all": 57, "n_fig": 7, "n_tab": 1, "L_tab": 183, "n_element_tab": 14, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 183, "n_element_tab_1": 14, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 80655, "len_all_1": 64309, "len_abs": 549, "len_title": 131, "len_sents": 23539, "len_sents_1": 23545, "n_sents": 142, "n_sents_1": 142, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 553, "title": "From Appearance to Motion: Aligning Visual Representations for Robotic Manipulation", "abs": "Pre-trained vision models used in robotics often misalign with manipulation tasks due to the loss used to train these vision models being focused on appearance rather than motion. In order to enhance motion encoding within vision models, we introduce a simple novel contrastive training framework that operates over predictions of motion. After training over EPIC Kitchens, model evaluations on behavioral cloning show a improvement in success rate over state-of-the-art methods across a benchmark of $3$ environments and $21$ object manipulation tasks.", "keywords": [ "pretrained", "frozen", "motion", "features", "policy", "behavioral-cloning" ], "rating_list": [ 3, 1, 5, 3 ], "soundness_list": [ 3, 1, 2, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "wkp57p0uhm", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Bing Jiang", "gender": "Male", "institution": "National Open University", "country": "", "position": "natio" }, { "name": "Cheng Cui", "gender": "unknown", "institution": "Soochow University", "country": "CN", "position": "MS student" }, { "name": "Dehan Kong", "gender": "Male", "institution": "imean.ai", "country": "US", "position": "Researcher" }, { "name": "Hangyu Liu", "gender": "Male", "institution": "Starblazer Inc", "country": "", "position": "Researcher" }, { "name": "Shuyan Zhou", "gender": "Non-Binary", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Sida Zhou", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Tongshuang Wu", "gender": "Female", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Yanyi Shang", "gender": "Male", "institution": "imean.ai", "country": "US", "position": "Principal Researcher" }, { "name": "Yichen Pan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yifei Leng", "gender": "Male", "institution": "Stellarrover.Inc", "country": "", "position": "MLE" }, { "name": "Zhengyang Wu", "gender": "Male", "institution": "iMeanAI", "country": "", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 88, "n_ref_all": 126, "n_fig": 15, "n_tab": 12, "L_tab": 5618, "n_element_tab": 564, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 2396, "n_element_tab_1": 315, "formula_len_all": 384, "formula_len_all_1": 0, "len_all": 185614, "len_all_1": 65749, "len_abs": 1588, "len_title": 111, "len_sents": 57419, "len_sents_1": 28970, "n_sents": 432, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1594, "title": "WebCanvas: Benchmarking Web Agents in Online Environments", "abs": "For web agents to be practically useful, they must adapt to the continuously evolving web environment characterized by frequent updates to user interfaces and content. However, most existing benchmarks only capture the static aspects of the web. To bridge this gap, we introduce WebCanvas, an innovative online evaluation framework for web agents that effectively addresses the dynamic nature of web interactions. WebCanvas contains three main components to facilitate realistic assessments: (1) A novel evaluation metric which reliably capture critical intermediate actions or states necessary for task completions while disregarding noise caused by insignificant events or changed web-elements. (2) A benchmark dataset called Mind2Web-Live, a refined version of original Mind2Web static dataset containing 542 tasks with 2439 intermediate evaluation states; (3) Lightweight and generalizable annotation tools and maintenance pipelines that enables the community to collect and maintain the high-quality, up-to-date dataset. Building on WebCanvas, we open-source a baseline agent framework with extensible modules for reasoning, providing a foundation for the community to conduct online inference and evaluations. Our best-performing agent achieves a task success rate of 23.1% and a task completion rate of 48.8% on the Mind2Web-Live test set. Additionally, we analyze the performance discrepancies across various websites, domains, and experimental environments. We encourage the community to contribute further insights on online agent evaluation, thereby advancing this field of research.", "keywords": [ "web automation; benchmark; LLM; language-guided agents" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "wkmCbrrDQN", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Arnon Turetzky", "gender": "unknown", "institution": "Hebrew University of Jerusalem", "country": "", "position": "PhD student" }, { "name": "Avihu Dekel", "gender": "Male", "institution": "International Business Machines", "country": "IL", "position": "Researcher" }, { "name": "David Haws", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Research Staff Member" }, { "name": "Hagai Aronowitz", "gender": "Male", "institution": "International Business Machines", "country": "IL", "position": "Principal Researcher" }, { "name": "Nimrod Shabtay", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Ron Hoory", "gender": "Male", "institution": "International Business Machines", "country": "IL", "position": "Researcher" }, { "name": "Slava Shechtman", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 46, "n_ref": 75, "n_ref_all": 97, "n_fig": 15, "n_tab": 2, "L_tab": 694, "n_element_tab": 62, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 411, "n_element_tab_1": 44, "formula_len_all": 542, "formula_len_all_1": 417, "len_all": 124357, "len_all_1": 58542, "len_abs": 1082, "len_title": 108, "len_sents": 34799, "len_sents_1": 28014, "n_sents": 255, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1088, "title": "Continuous Speech Synthesis using per-token Latent Diffusion", "abs": "The success of autoregressive transformer models with discrete tokens has inspired quantization-based approaches for continuous modalities, though these often limit reconstruction quality.\nWe therefore introduce SALAD, a per-token latent diffusion model for zero-shot text-to-speech, that operates on continuous representations.\nSALAD builds upon the recently proposed expressive diffusion head for image generation, and extends it to generate variable-length outputs. \nOur approach utilizes semantic tokens for providing contextual information and determining the stopping condition.\nWe suggest three continuous variants for our method, extending popular discrete speech synthesis techniques. \nAdditionally, we implement discrete baselines for each variant and conduct a comparative analysis of discrete versus continuous speech modeling techniques.\nOur results demonstrate that both continuous and discrete approaches are highly competent, and that SALAD achieves a superior intelligibility score while obtaining speech quality and speaker similarity on par with the ground-truth audio.", "keywords": [ "Speech Synthesis", "Continuous Sequence Modeling", "Latent Diffusion" ], "rating_list": [ 6, 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 2, 4 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 1, 3 ], "confidence_list": [ 4, 5, 4, 3, 2 ] }, { "paper_id": "wkbx7BRAsM", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Jiang Bian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Junliang Guo", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Li Zhao", "gender": "Female", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Linli Xu", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Tianyu He", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Researcher" }, { "name": "Wentao Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 44, "n_ref": 79, "n_ref_all": 113, "n_fig": 14, "n_tab": 11, "L_tab": 1843, "n_element_tab": 226, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 893, "n_element_tab_1": 113, "formula_len_all": 394, "formula_len_all_1": 365, "len_all": 155399, "len_all_1": 62178, "len_abs": 1184, "len_title": 105, "len_sents": 40508, "len_sents_1": 27922, "n_sents": 282, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1049, "title": "Autoregressive Transformers are Zero-Shot Video Imitators", "abs": "People interact with the real-world largely dependent on visual signal, which are ubiquitous and illustrate detailed demonstrations. In this paper, we explore utilizing visual signals as a new interface for models to interact with the environment. Specifically, we choose videos as a representative visual signal. And by training autoregressive Transformers on video datasets in a self-supervised objective, we find that the model emerges a zero-shot capability to infer the semantics from a demonstration video, and imitate the semantics to an unseen scenario. This allows the models to perform unseen tasks by watching the demonstration video in an in-context manner, without further fine-tuning. To validate the imitation capacity, we design various evaluation metrics including both objective and subjective measures. The results show that our models can generate high-quality video clips that accurately align with the semantic guidance provided by the demonstration videos, and we also show that the imitation capacity follows the scaling law.", "keywords": [ "Video Generation", "Transformer", "Zero-Shot" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "wkHcXDv7cv", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Annan Yu", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Dongwei Lyu", "gender": "Female", "institution": "University of Chicago", "country": "US", "position": "MS student" }, { "name": "Michael W. Mahoney", "gender": "unknown", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "N. Benjamin Erichson", "gender": "Male", "institution": "Lawrence Berkeley National Lab", "country": "US", "position": "Researcher" }, { "name": "Soon Hoe Lim", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "Assistant Professor" } ], "n_formula": 56, "n_formula_1": 25, "n_ref_uni": 57, "n_ref": 106, "n_ref_all": 180, "n_fig": 10, "n_tab": 2, "L_tab": 724, "n_element_tab": 108, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 70, "n_element_tab_1": 6, "formula_len_all": 3875, "formula_len_all_1": 1806, "len_all": 190925, "len_all_1": 71627, "len_abs": 1410, "len_title": 91, "len_sents": 58467, "len_sents_1": 33466, "n_sents": 510, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 43, "L_abs": 1199, "title": "Tuning Frequency Bias of State Space Models", "abs": "State space models (SSMs) leverage linear, time-invariant (LTI) systems to effectively learn sequences with long-range dependencies. By analyzing the transfer functions of LTI systems, we find that SSMs exhibit an implicit bias toward capturing low-frequency components more effectively than high-frequency ones. This behavior aligns with the broader notion of frequency bias in deep learning model training. We show that the initialization of an SSM assigns it an innate frequency bias and that training the model in a conventional way does not alter this bias. Based on our theory, we propose two mechanisms to tune frequency bias: either by scaling the initialization to tune the inborn frequency bias; or by applying a Sobolev-norm-based filter to adjust the sensitivity of the gradients to high-frequency inputs, which allows us to change the frequency bias via training. Using an image-denoising task, we empirically show that we can strengthen, weaken, or even reverse the frequency bias using both mechanisms. By tuning the frequency bias, we can also improve SSMs' performance on learning long-range sequences, averaging an $88.26\\\\%$ accuracy on the Long-Range Arena (LRA) benchmark tasks.", "keywords": [ "state-space models", "sequence models", "Long-Range Arena", "frequency bias" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "wjgNVsbT3T", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "hanyu Wang", "gender": "Male", "institution": "Renmin University of China, Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Ke Fang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "", "position": "Undergrad student" }, { "name": "Shichao Song", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Simin Niu", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "PhD student" }, { "name": "Steven Shi", "gender": "Male", "institution": "5Y Capital", "country": "CN", "position": "Researcher" }, { "name": "Zhiyu li", "gender": "Male", "institution": "Institute for Advanced Algorithms Research, Shanghai", "country": "CN", "position": "Principal Researcher" }, { "name": "Zifan Zheng", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "MS student" }, { "name": "Yu Qingchen", "gender": "Male", "institution": "Institute for Advanced Algorithms Research, Shanghai", "country": "CN", "position": "Intern" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 23, "n_ref": 33, "n_ref_all": 56, "n_fig": 9, "n_tab": 7, "L_tab": 3570, "n_element_tab": 213, "n_fig_1": 10, "n_tab_1": 7, "L_tab_1": 1812, "n_element_tab_1": 218, "formula_len_all": 811, "formula_len_all_1": 680, "len_all": 103557, "len_all_1": 71794, "len_abs": 1709, "len_title": 122, "len_sents": 27204, "len_sents_1": 21796, "n_sents": 232, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1503, "title": "TurtleBench: Evaluating Top Language Models via Real-World Yes/No Puzzles", "abs": "As the application of Large Language Models (LLMs) expands, the demand for reliable evaluations increases. Existing LLM evaluation benchmarks primarily rely on static datasets, making it challenging to assess model performance in dynamic interactions with users. Moreover, these benchmarks often depend on specific background knowledge, complicating the measurement of a model's logical reasoning capabilities. Other dynamic evaluation methods based on strong models or manual efforts may introduce biases and incur high costs and time demands, hindering large-scale application. To address these issues, we propose TurtleBench. TurtleBench collects real user guesses from our online Turtle Soup Puzzle platform that we developed. This approach allows for the relatively dynamic generation of evaluation datasets, mitigating the risk of model cheating while aligning assessments more closely with genuine user needs for reasoning capabilities, thus enhancing the reliability of evaluations. TurtleBench includes 1,532 user guesses along with the correctness of guesses after annotation. Using this dataset, we thoroughly evaluated nine of the most advanced LLMs available today. Notably, the OpenAI o1 series models did not achieve leading results in these evaluations. We propose several hypotheses for further research, such as \u201cthe latent reasoning of o1 utilizes trivial Chain-of-Thought (CoT) techniques\u201d and \u201cincreasing CoT length not only provides reasoning benefits but also incurs noise costs.\u201d", "keywords": [ "Benchmark; LLM Evaluation" ], "rating_list": [ 5, 1, 5, 5, 3 ], "soundness_list": [ 3, 1, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 2, 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "wjPa7GUIR9", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Anwar Hithnawi", "gender": "unknown", "institution": "University of Toronto", "country": "CA", "position": "Assistant Professor" }, { "name": "Emanuel Opel", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Evan Rose", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Hidde Lycklama", "gender": "unknown", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Isha Gupta", "gender": "Female", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 33, "n_ref": 59, "n_ref_all": 74, "n_fig": 8, "n_tab": 1, "L_tab": 548, "n_element_tab": 22, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 602, "formula_len_all_1": 522, "len_all": 120472, "len_all_1": 63767, "len_abs": 1444, "len_title": 132, "len_sents": 42073, "len_sents_1": 32186, "n_sents": 313, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1455, "title": "Fragile Giants: Understanding Susceptibility of Models to Subpopulation Attacks", "abs": "As machine learning models become increasingly complex, concerns about their robustness and trustworthiness have become more pressing. A critical vulnerability of these models is data poisoning attacks, where adversaries deliberately alter training data to degrade model performance. One particularly stealthy form of these attacks is subpopulation poisoning, which targets distinct subgroups within a dataset while leaving overall performance largely intact. The ability of these attacks to generalize within subpopulations poses a significant risk in real-world settings, as they can be exploited to harm marginalized or underrepresented groups within the dataset. In this work, we investigate how model complexity influences susceptibility to subpopulation poisoning attacks. We introduce a theoretical\nframework that explains how models with locally dependent learning behavior\u2014a characteristic exhibited by overparameterized models\u2014can misclassify arbitrary subpopulations. To validate our theory, we conduct extensive experiments on large-scale image and text datasets using popular model architectures. Our results show a clear trend: models with more parameters are significantly more vulnerable to subpopulation poisoning. Moreover, we find that attacks on smaller, human-interpretable subgroups often go undetected by these models. These results highlight the need for developing defenses that specifically address subpopulation vulnerabilities.", "keywords": [ "poisoning", "robustness", "subpopulations" ], "rating_list": [ 8, 5, 6, 6, 5 ], "soundness_list": [ 3, 2, 2, 3, 2 ], "presentation_list": [ 4, 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 4, 2, 3, 3, 3 ] }, { "paper_id": "wj4Az2454x", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Alireza Moradzadeh", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Lukasz Wawrzyniak", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Miles Macklin", "gender": "Male", "institution": "NVIDIA", "country": "NZ", "position": "Principal Engineer" }, { "name": "Saee Gopal Paliwal", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 23, "n_ref": 44, "n_ref_all": 56, "n_fig": 4, "n_tab": 5, "L_tab": 230, "n_element_tab": 26, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 230, "n_element_tab_1": 26, "formula_len_all": 288, "formula_len_all_1": 288, "len_all": 66433, "len_all_1": 59148, "len_abs": 1091, "len_title": 91, "len_sents": 19831, "len_sents_1": 17613, "n_sents": 149, "n_sents_1": 130, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1114, "title": "UKAN: UNBOUNDED KOLMOGOROV-ARNOLD NETWORKS", "abs": "We present Unbounded Kolmogorov-Arnold Networks (UKANs), a novel algorithm that eliminates the need for bounded grids in traditional Kolmogorov-Arnold Networks (KANs). The key innovation is a coefficient generator (CG) model that dynamically produces B-spline coefficients, operating on an infinite symmetric grid. UKANs integrate multilayer-perceptrons with KANs, using positional encoding of grid groups as input to the CG model. This approach enables function approximation on unbounded domains without data normalization. Additionally, to reduce UKAN and KAN computational cost, we introduce a GPU-accelerated library that reduces B-spline evaluation complexity by a factor of $\\mathcal{O}(\\text{grid size})$ compared to existing libraries, enabling efficient large-scale learning. Our experiments on regression, classification, and generative tasks demonstrate UKANs' effectiveness, while benchmarks confirm superior memory and computational efficiency compared to existing methods. This work advances function approximation techniques, offering a flexible solution for complex, large-scale learning problems.", "keywords": [ "KAN", "Acceleration", "Unbounded KAN", "Grid Free", "Function Approximation" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 2, 3 ] }, { "paper_id": "wixDdL0vj8", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Fengyuan Liu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Intern" }, { "name": "Tao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xinyi Shang", "gender": "Female", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "peng sun", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 39, "n_ref": 80, "n_ref_all": 104, "n_fig": 9, "n_tab": 4, "L_tab": 882, "n_element_tab": 71, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 3455, "n_element_tab_1": 99, "formula_len_all": 531, "formula_len_all_1": 531, "len_all": 122800, "len_all_1": 70182, "len_abs": 1023, "len_title": 79, "len_sents": 36880, "len_sents_1": 30275, "n_sents": 286, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 31, "L_abs": 1043, "title": "Collaborative Data Optimization", "abs": "Training efficiency plays a pivotal role in deep learning.\n This paper begins by analyzing current methods for enhancing efficiency, highlighting the necessity of optimizing targets, a process we define as data optimization.\n Subsequently, we reveal that current data optimization methods incur significant additional costs, e.g., human resources or computational overhead, due to their inherently sequential optimization process.\n To address these issues, we propose CoOpt, a highly efficient, parallelized framework designed for collaborative data optimization.\n CoOpt enables participants to independently optimize data subsets, ensuring that the overall performance, once these subsets are collected, remains comparable to the sequential optimization of the entire dataset, thus significantly reducing optimization costs for individual participants.\n Extensive experiments have been conducted on various real-world scenarios to demonstrate the effectiveness and efficiency of CoOpt across various datasets and architectures.", "keywords": [ "Unlabeled Data", "Data Optimization", "Efficiency" ], "rating_list": [ 5, 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 1, 2 ], "presentation_list": [ 3, 1, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2, 3 ], "confidence_list": [ 2, 4, 3, 4, 4 ] }, { "paper_id": "whaO3482bs", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Chanwoong Yoon", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "MS student" }, { "name": "Donghyeon Lee", "gender": "unknown", "institution": "Korea University", "country": "KR", "position": "MS student" }, { "name": "Jaewoo Kang", "gender": "unknown", "institution": "Korea University", "country": "KR", "position": "Full Professor" }, { "name": "Jungwoo Park", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "MS student" }, { "name": "Minbyul Jeong", "gender": "Male", "institution": "Upstage", "country": "", "position": "Postdoc" }, { "name": "Yein Park", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 44, "n_ref": 76, "n_ref_all": 102, "n_fig": 7, "n_tab": 12, "L_tab": 6434, "n_element_tab": 547, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1787, "n_element_tab_1": 173, "formula_len_all": 352, "formula_len_all_1": 352, "len_all": 244794, "len_all_1": 62437, "len_abs": 1539, "len_title": 87, "len_sents": 65156, "len_sents_1": 29763, "n_sents": 438, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1548, "title": "ChroKnowledge: Unveiling Chronological Knowledge of Language Models in Multiple Domains", "abs": "Large language models (LLMs) have brought significant changes to many aspects of our lives.\nHowever, assessing and ensuring their chronological knowledge remains challenging.\nExisting approaches fall short in addressing the temporal adaptability of knowledge, often relying on a fixed time-point view. \nTo overcome this, we introduce ChroKnowBench, a benchmark dataset designed to evaluate chronologically accumulated knowledge across three key aspects: multiple domains, time dependency, temporal state.\nOur benchmark distinguishes between knowledge that evolves (e.g., personal history, scientific discoveries, amended laws) and knowledge that remain constant (e.g., mathematical truths, commonsense facts). \nBuilding on this benchmark, we present ChroKnowledge (Chronological Categorization of Knowledge), a novel sampling-based framework for evaluating LLMs' non-parametric chronological knowledge.\nOur evaluation led to the following observations: \n(1) The ability of eliciting temporal knowledge varies depending on the data format that model was trained on.\n(2) LLMs partially recall knowledge or show a cut-off at temporal boundaries rather than recalling all aspects of knowledge correctly.\nThus, we apply ourChroKnowPrompt, an in-depth prompting to elicit chronological knowledge by traversing step-by-step through the surrounding time spans.\nWe observe that it successfully recalls objects across both open-source and proprietary LLMs, demonstrating versatility, though it faces challenges with dynamic datasets and unstructured formats.", "keywords": [ "Temporal knowledge", "Knowledge update" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "whXHZIaRVB", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Diyi Yang", "gender": "Female", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Jiaao Chen", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 54, "n_ref": 103, "n_ref_all": 123, "n_fig": 4, "n_tab": 4, "L_tab": 1717, "n_element_tab": 236, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1717, "n_element_tab_1": 236, "formula_len_all": 70, "formula_len_all_1": 70, "len_all": 145753, "len_all_1": 61049, "len_abs": 1294, "len_title": 98, "len_sents": 32866, "len_sents_1": 28696, "n_sents": 210, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1299, "title": "Dynamic Skill Adaptation for Large Language Models", "abs": "We present Dynamic Skill Adaptation (DSA), an adaptive and dynamic framework to adapt novel and complex skills to Large Language Models (LLMs). Compared with previous work which learns from human-curated and static data in random orders, we propose to first automatically generate and organize the training data by mimicking the learning pathways of human and then dynamically tailor the training data based on the training dynamics. Specifically, inspired by the learning structures and teaching strategies in the human education system, we first construct a skill graph by decomposing complex skills into sub-skills and arranging them based on their dependencies in human syllables. For every skill, we utilize LLMs to generate both textbook-like data which contains detailed descriptions of skills for pre-training and exercise-like data which targets at explicitly utilizing the skills to solve problems for instruction-tuning. Furthermore, during the instruction-tuning, we dynamically update the training data which down-weight easy-to-learn examples, generate more complex examples, and filter out data with errors. Experiments on large language models such as LLAMA and Mistral demonstrate the effectiveness of our proposed methods in adapting math reasoning skills and social study skills.", "keywords": [ "Large Language Models", "Skill Adaptation", "Skill Graph" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "wh6pilyz2L", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Christian Schultze", "gender": "Male", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "MS student" }, { "name": "Felix Selgert", "gender": "Male", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "", "position": "Assistant Professor" }, { "name": "Kara Kuebart", "gender": "unknown", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "", "position": "PhD student" }, { "name": "Moritz Wolter", "gender": "Male", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "Postdoc" }, { "name": "Niklas Kerkfeld", "gender": "Male", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "Undergrad student" }, { "name": "Princilia Weber", "gender": "Female", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 73, "n_ref_all": 107, "n_fig": 5, "n_tab": 9, "L_tab": 3019, "n_element_tab": 282, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 528, "n_element_tab_1": 35, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 137575, "len_all_1": 57108, "len_abs": 1166, "len_title": 111, "len_sents": 47553, "len_sents_1": 27120, "n_sents": 429, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1174, "title": "Chronicling Germany: An Annotated Historical Newspaper Dataset", "abs": "The correct detection of dense article layout and the recognition of characters in historical newspaper pages remains a challenging requirement for Natural Language Processing (NLP) and machine learning applications on historical newspapers in the field of digital history. Digital newspaper portals for historic Germany typically provide Optical Character Recognition (OCR) text, albeit of varying quality. Unfortunately, layout information is often missing, limiting this rich source\u2019s scope. Our dataset is designed to enable the training of layout and OCR modells for historic German-language newspapers. The Chronicling Germany dataset contains 693 annotated historical newspaper pages from the time period between 1852 and 1924. The paper presents a processing pipeline and establishes baseline results on in- and out-of-domain test data using this pipeline. Both our dataset and the corresponding baseline code are freely available online. This work creates a starting point for future research in the field of digital history and historic German language newspaper processing. Furthermore, it provides the opportunity to study a low-resource task in computer vision.", "keywords": [ "historic newspaper processing", "digital history", "computer vision" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "wgnMdxS2nZ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "David E. Bernal Neira", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" }, { "name": "Muhammad Shafique", "gender": "Male", "institution": "New York University", "country": "AE", "position": "Full Professor" }, { "name": "Nouhaila Innan", "gender": "Female", "institution": "New York University, Abu Dhabi", "country": "AE", "position": "Postdoc" }, { "name": "Sadok Ben Ben Yahia", "gender": "Male", "institution": "University of Southern Denmark - SDU", "country": "DK", "position": "Full Professor" }, { "name": "Siddhant Dutta", "gender": "Male", "institution": "Dwarkadas J. Sanghvi College Of Engineering", "country": "IN", "position": "Undergrad student" } ], "n_formula": 15, "n_formula_1": 12, "n_ref_uni": 10, "n_ref": 18, "n_ref_all": 30, "n_fig": 5, "n_tab": 8, "L_tab": 4824, "n_element_tab": 429, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 6009, "n_element_tab_1": 426, "formula_len_all": 819, "formula_len_all_1": 649, "len_all": 118320, "len_all_1": 64939, "len_abs": 1087, "len_title": 140, "len_sents": 38636, "len_sents_1": 25734, "n_sents": 252, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1090, "title": "MQFL-FHE: Multimodal Quantum Federated Learning Framework with Fully Homomorphic Encryption", "abs": "The integration of fully homomorphic encryption (FHE) in federated learning (FL) has led to significant advances in data privacy. However, during the aggregation phase, it often results in performance degradation of the aggregated model, hindering the development of robust representational generalization. In this work, we propose a novel multimodal quantum federated learning framework that utilizes quantum computing to counteract the performance drop resulting from FHE. For the first time in FL, our framework combines a multimodal quantum mixture of experts (MQMoE) model with FHE, incorporating multimodal datasets for enriched representation and task-specific learning. Our MQMoE framework enhances performance on multimodal datasets and combined genomics and brain MRI scans, especially for underrepresented categories. Our results also demonstrate that the quantum-enhanced approach mitigates the performance degradation associated with FHE and improves classification accuracy across diverse datasets, validating the potential of quantum interventions in enhancing privacy in FL.", "keywords": [ "Quantum Federated Learning", "Fully Homomorphic Encryption", "Multimodal Quantum Mixture of Experts" ], "rating_list": [ 3, 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 2, 2 ] }, { "paper_id": "wgRQ2WAORJ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Ali Rasekh", "gender": "Male", "institution": "L3S Research Center", "country": "DE", "position": "Researcher" }, { "name": "Amirabbas Afzali", "gender": "Male", "institution": "Sharif University of Technology, Sharif University of Technology", "country": "IR", "position": "Undergrad student" }, { "name": "Borna khodabandeh", "gender": "Male", "institution": "Sharif University of Technology, Sharif University of Technology", "country": "IR", "position": "Undergrad student" }, { "name": "Mahyar JafariNodeh", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Sepehr Kazemi Ranjbar", "gender": "Male", "institution": "Sharif University of Technology", "country": "", "position": "Undergrad student" }, { "name": "Simon Gottschalk", "gender": "Male", "institution": "L3S Research Center, Leibniz Universit\u00e4t Hannover", "country": "DE", "position": "Postdoc" } ], "n_formula": 59, "n_formula_1": 25, "n_ref_uni": 46, "n_ref": 82, "n_ref_all": 115, "n_fig": 13, "n_tab": 6, "L_tab": 1106, "n_element_tab": 156, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 954, "n_element_tab_1": 63, "formula_len_all": 5021, "formula_len_all_1": 1558, "len_all": 205335, "len_all_1": 65234, "len_abs": 1701, "len_title": 119, "len_sents": 61219, "len_sents_1": 29685, "n_sents": 480, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1626, "title": "Aligning Visual Contrastive learning models via Preference Optimization", "abs": "Contrastive learning models have demonstrated impressive abilities to capture semantic similarities by aligning representations in the embedding space. However, their performance can be limited by the quality of the training data and its inherent biases. While Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) have been applied to generative models to align them with human preferences, their use in contrastive learning has yet to be explored. This paper introduces a novel method for training contrastive learning models using Preference Optimization (PO) to break down complex concepts. Our method systematically aligns model behavior with desired preferences, enhancing performance on the targeted task. In particular, we focus on enhancing model robustness against typographic attacks, commonly seen in contrastive models like CLIP. We further apply our method to disentangle gender understanding and mitigate gender biases, offering a more nuanced control over these sensitive attributes. Our experiments demonstrate that models trained using PO outperform standard contrastive learning techniques while retaining their ability to handle adversarial challenges and maintain accuracy on other downstream tasks. This makes our method well-suited for tasks requiring fairness, robustness, and alignment with specific preferences. We evaluate our method on several vision-language tasks, tackling challenges such as typographic attacks. Additionally, we explore the model's ability to disentangle gender concepts and mitigate gender bias, showcasing the versatility of our approach.", "keywords": [ "contrastive learning", "preference optimization", "alignment", "reinforcement learning from human feedback", "robustness", "computer vision" ], "rating_list": [ 3, 6, 8, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 2, 4, 4 ], "contribution_list": [ 2, 3, 4, 2 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "wgKW4U7ktq", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Bin Xu", "gender": "Male", "institution": "Department of Computer Science, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jie Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jinhao Chen", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Weihan Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Zhen Yang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhengxiao Du", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhihuan Jiang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 15, "n_ref": 36, "n_ref_all": 104, "n_fig": 36, "n_tab": 11, "L_tab": 6417, "n_element_tab": 724, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 3117, "n_element_tab_1": 383, "formula_len_all": 1058, "formula_len_all_1": 0, "len_all": 472729, "len_all_1": 54618, "len_abs": 1455, "len_title": 147, "len_sents": 182481, "len_sents_1": 22350, "n_sents": 1558, "n_sents_1": 159, "venue": "ICLR2025", "decision": "withdraw", "L_title": 98, "L_abs": 1474, "title": "VisScience: An Extensive Benchmark for Evaluating K12 Educational Multi-modal Scientific Reasoning", "abs": "Multi-modal large language models (MLLMs) have shown promise in integrating textual and visual information to handle complex visual understanding tasks. However, most benchmarks evaluating MLLMs focus mainly on mathematics or general visual understanding, revealing a significant gap in assessing capabilities across other critical scientific disciplines like physics and chemistry. To bridge this gap, we meticulously construct a comprehensive benchmark, \\textbf{VisScience}, to evaluate multi-modal scientific reasoning across mathematics, physics, and chemistry. This benchmark comprises 3,000 questions drawn from K12 education, from elementary to high school levels, evenly distributed with 1,000 questions per discipline. VisScience encompasses 21 distinct subjects, classified into five difficulty levels to cover a wide range of topics within each discipline. We utilize VisScience to conduct a detailed evaluation of 25 representative MLLMs in scientific reasoning. The experimental results show that closed-source MLLMs generally surpass open-source models, with standout performances including a 53.4\\% accuracy in mathematics by Claude3.5-Sonnet, 38.2\\% in physics by GPT-4o, and 47.0\\% in chemistry by Gemini-1.5-Pro. These results underscore the strengths and limitations of MLLMs, suggesting areas for future improvement and highlighting the importance of developing models that can effectively handle the diverse demands of multi-modal scientific reasoning.", "keywords": [ "Multi-modal Large Language Model", "Scientific Reasoning", "Benchmark" ], "rating_list": [ 8, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "wgHJDHW65K", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Andrew Miller", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Anshuman Mishra", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Ian Shapiro", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Joseph Futoma", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Salar Abbaspourazad", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 60, "n_ref": 155, "n_ref_all": 215, "n_fig": 6, "n_tab": 21, "L_tab": 8211, "n_element_tab": 708, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 275, "n_element_tab_1": 28, "formula_len_all": 162, "formula_len_all_1": 162, "len_all": 210610, "len_all_1": 69187, "len_abs": 2474, "len_title": 114, "len_sents": 59585, "len_sents_1": 38143, "n_sents": 293, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 2484, "title": "Representational Knowledge Distillation Across Wearable Biosignals", "abs": "Modern wearable devices can conveniently and continuously record various biosignals in the many different environments of daily living, ultimately enabling a rich view of individual health. \nHowever, not all biosignals are the same: high-fidelity measurements, such as photoplethysmography (PPG), contain more physiological information, but require optical sensors with a high power footprint. In a resource-constrained setting, such high-fidelity biosignals may be unavailable. Alternatively, a lower-fidelity biosignal, such as those from an accelerometer, has a significantly smaller power footprint and is available in almost any wearable device. While multi-modal modeling and cross-modal reconstruction of biosignals have been done before, here, we demonstrate that we can distill representational knowledge across biosignals with different levels of fidelity, i.e., from PPG to accelerometer, using 20 million minutes of unlabeled data collected from ~172K participants in the Apple Heart and Movement Study under informed consent. Our knowledge distillation framework does not require labels; we pre-train PPG encoders via self-supervised learning, and then distill the representational knowledge from the PPG encoders to accelerometer encoders. We first demonstrate strong cross-modal alignment on unseen data, e.g., 99.2% top-1 accuracy for retrieving PPG embeddings from accelerometer embeddings. We show that distilled accelerometer encoders have significantly more informative representations compared to self-supervised or supervised encoders trained on accelerometer data for downstream targets, observed by at least 23%-49% improved performance for predicting heart rate and heart rate variability, and are readily predictive of a wide array of downstream targets including demographic variables, health conditions, use of medications, and lifestyle habits. We also demonstrate that our framework can be applied to different encoder architectures with different pre-training strategies of the strong encoder, and can be used to simultaneously do cross-modality distillation and model compression. Additionally, we perform various ablations for augmentations, hyperparameters and multi-modal training. We believe our proposed representational knowledge distillation framework may unlock new opportunities for developing digital biomarkers from any wearable device with lower-fidelity biosignals, and help individuals track their health more frequently and conveniently.", "keywords": [ "Health", "Foundation models", "Knowledge distillation", "Unsupervised learning", "Self-supervised learning", "Biosignals", "Wearable devices" ], "rating_list": [ 3, 6, 3, 8, 8 ], "soundness_list": [ 1, 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 1, 2, 3 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "wgDB1QuxIA", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Kaiyi Ji", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Peiyao Xiao", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Qi Zhang", "gender": "unknown", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Shaofeng Zou", "gender": "unknown", "institution": "Arizona State University", "country": "", "position": "Associate Professor" } ], "n_formula": 153, "n_formula_1": 32, "n_ref_uni": 30, "n_ref": 128, "n_ref_all": 171, "n_fig": 2, "n_tab": 11, "L_tab": 3817, "n_element_tab": 415, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 19791, "formula_len_all_1": 2523, "len_all": 212713, "len_all_1": 72387, "len_abs": 4695, "len_title": 102, "len_sents": 54638, "len_sents_1": 30495, "n_sents": 519, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1569, "title": "MGDA Converges under Generalized Smoothness, Provably", "abs": "Multi-objective optimization (MOO) is receiving more attention in various fields such as multi-task learning. Recent works provide some effective algorithms with theoretical analysis but they are limited by the standard $L$-smooth or bounded-gradient assumptions, which typically do not hold for neural networks, such as Long short-term memory (LSTM) models and Transformers. In this paper, we study a more general and realistic class of generalized $\\ell$-smooth loss functions, where $\\ell$ is a general non-decreasing function of gradient norm. We revisit and analyze the fundamental multiple gradient descent algorithm (MGDA) and its stochastic version with double sampling for solving the generalized $\\ell$-smooth MOO problems, which approximate the conflict-avoidant (CA) direction that maximizes the minimum improvement among objectives. We provide a comprehensive convergence analysis of these algorithms and show that they converge to an $\\epsilon$-accurate Pareto stationary point with a guaranteed $\\epsilon$-level average CA distance (i.e., the gap between the updating direction and the CA direction) over all iterations, where totally $\\mathcal{O}(\\epsilon^{-2})$ and $\\mathcal{O}(\\epsilon^{-4})$ samples are needed for deterministic and stochastic settings, respectively. We prove that they can also guarantee a tighter $\\epsilon$-level CA distance in each iteration using more samples. Moreover, we analyze an efficient variant of MGDA named MGDA-FA using only $\\mathcal{O}(1)$ time and space, while achieving the same performance guarantee as MGDA.", "keywords": [ "Multi-Objective Optimization", "Generalized Smoothness", "Convergence Analysis", "Sample Complexity" ], "rating_list": [ 5, 5, 6, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 2, 2, 2 ] }, { "paper_id": "wg3rBImn3O", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "R. Teal Witter", "gender": "unknown", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Christopher P Musco", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 62, "n_formula_1": 31, "n_ref_uni": 38, "n_ref": 62, "n_ref_all": 85, "n_fig": 11, "n_tab": 5, "L_tab": 1724, "n_element_tab": 198, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 120, "n_element_tab_1": 4, "formula_len_all": 5302, "formula_len_all_1": 1892, "len_all": 153485, "len_all_1": 66919, "len_abs": 1176, "len_title": 118, "len_sents": 44243, "len_sents_1": 28966, "n_sents": 406, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 70, "L_abs": 1183, "title": "Provably Accurate Shapley Value Estimation via Leverage Score Sampling", "abs": "Originally introduced in game theory, Shapley values have emerged as a central tool in explainable machine learning, where they are used to attribute model predictions to specific input features. However, computing Shapley values exactly is expensive: for a model with $n$ features, $O(2^n)$ model evaluations are necessary. To address this issue, approximation algorithms are widely used. One of the most popular is the Kernel SHAP algorithm, which is model agnostic and remarkably effective in practice. However, to the best of our knowledge, Kernel SHAP has no strong non-asymptotic complexity guarantees. We address this issue by introducing *Leverage SHAP*, a light-weight modification of Kernel SHAP that provides provably accurate Shapley value estimates with just $O(n\\log n)$ model evaluations. Our approach takes advantage of a connection between Shapley value estimation and agnostic active learning by employing *leverage score sampling*, a powerful regression tool. Beyond theoretical guarantees, we show that Leverage SHAP consistently outperforms even the highly optimized implementation of Kernel SHAP available in the ubiquitous SHAP library [Lundberg \\& Lee, 2017].", "keywords": [ "Explainable AI", "Active Regression", "Shapley Values", "Leverage Scores" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 4, 3, 4 ], "presentation_list": [ 4, 2, 4 ], "contribution_list": [ 4, 2, 3 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "wg1PCg3CUP", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Aditi Raghunathan", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Benjamin Frederick Spector", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Blake Bordelon", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Cengiz Pehlevan", "gender": "unknown", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "Christopher Re", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Mansheej Paul", "gender": "Male", "institution": "Databricks, Databricks", "country": "US", "position": "Researcher" }, { "name": "Niklas Muennighoff", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Tanishq Kumar", "gender": "Male", "institution": "Harvard College", "country": "", "position": "Undergrad student" }, { "name": "Zachary Ankner", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Undergrad student" } ], "n_formula": 37, "n_formula_1": 7, "n_ref_uni": 79, "n_ref": 146, "n_ref_all": 178, "n_fig": 12, "n_tab": 3, "L_tab": 510, "n_element_tab": 105, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 357, "n_element_tab_1": 28, "formula_len_all": 2286, "formula_len_all_1": 546, "len_all": 252997, "len_all_1": 56901, "len_abs": 1130, "len_title": 74, "len_sents": 62377, "len_sents_1": 25987, "n_sents": 455, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 26, "L_abs": 1137, "title": "Scaling Laws for Precision", "abs": "Low precision training and inference affect both the quality and cost of language models, but current scaling laws do not account for this. In this work, we devise \"precision-aware\" scaling laws for both training and inference. We propose that training in lower precision reduces the model's \"effective parameter count,\" allowing us to predict the additional loss incurred from training in low precision and post-train quantization. For inference, we find that the degradation introduced by post-training quantization increases as models are trained on more data, eventually making additional pretraining data actively harmful. For training, our scaling laws allow us to predict the loss of a model with different parts in different precisions, and suggest that training larger models in lower precision can be compute optimal. We unify the scaling laws for post and pretraining quantization to arrive at a single functional form that predicts degradation from training and inference in varied precisions. We fit on over 465 pretraining runs and validate our predictions on model sizes up to 1.7B parameters trained on up to 26B tokens.", "keywords": [ "quantization", "scaling laws", "precision", "language models" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "wfLuiDjQ0u", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Chaofan Li", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Defu Lian", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Kun Luo", "gender": "Female", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Minghao Qin", "gender": "unknown", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "Undergrad student" }, { "name": "Yingxia Shao", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Full Professor" }, { "name": "Zheng Liu", "gender": "unknown", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Jianlv Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Shi Tao Xiao", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 39, "n_ref": 77, "n_ref_all": 86, "n_fig": 2, "n_tab": 13, "L_tab": 10842, "n_element_tab": 1068, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1775, "n_element_tab_1": 90, "formula_len_all": 313, "formula_len_all_1": 226, "len_all": 154851, "len_all_1": 58267, "len_abs": 2969, "len_title": 89, "len_sents": 34442, "len_sents_1": 26588, "n_sents": 274, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 39, "L_abs": 1380, "title": "Making Text Embedders Few-Shot Learners", "abs": "Large language models (LLMs) with decoder-only architectures have demonstrated exceptional text-generation capabilities across a variety of tasks. Some researchers have also adapted these models for text representation tasks. However, in text representation tasks, these models often face performance degradation on unseen tasks. In-context learning (ICL), which leverages examples provided in the input context, enables LLMs to handle unseen tasks effectively. Inspired by this, we aim to fully utilize the inherent properties of LLMs to enhance text representation performance across different tasks through the ICL approach.\n\nIn this paper, we introduce a simple yet effective training strategy, which significantly improves text representation capabilities. Unlike previous models that prepend task instructions to the text, our method randomly samples a varying number of examples during training, endowing the embedding model with in-context learning abilities while maintaining its zero-shot capabilities. This approach does not require additional data construction or modifications to the model architecture. On the contrary, we find that some popular modifications to the model, such as bidirectional attention, can degrade performance, undermining the inherent characteristics of LLMs. We open-source the model, code, and data to foster further development in the field.", "keywords": [ "large language model", "embedding model", "in-context learning" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "wetJo6xXb1", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chen Xiong", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Pin-Yu Chen", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Tsung-Yi Ho", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "", "position": "Full Professor" }, { "name": "Xiangyu Qi", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 27, "n_ref": 77, "n_ref_all": 126, "n_fig": 5, "n_tab": 19, "L_tab": 5870, "n_element_tab": 529, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 1168, "n_element_tab_1": 140, "formula_len_all": 369, "formula_len_all_1": 458, "len_all": 211239, "len_all_1": 67134, "len_abs": 1273, "len_title": 158, "len_sents": 71762, "len_sents_1": 30871, "n_sents": 568, "n_sents_1": 244, "venue": "ICLR2025", "decision": "withdraw", "L_title": 109, "L_abs": 1286, "title": "Defensive Prompt Patch: A Robust and Generalizable Defense of Large Language Models against Jailbreak Attacks", "abs": "Safety, security, and compliance are essential requirements when aligning large language models (LLMs). However, many seemingly aligned LLMs are soon shown to be susceptible to jailbreak attacks. These attacks aim to circumvent the models' safety guardrails and security mechanisms by introducing jailbreak prompts into malicious queries. In response to these challenges, this paper introduces \\textbf{Defensive Prompt Patch} (DPP), a novel prompt-based defense mechanism specifically designed to protect LLMs against such sophisticated jailbreak strategies. Unlike previous approaches, which have often compromised the utility of the model for the sake of safety, DPP is designed to achieve a minimal Attack Success Rate (ASR) while preserving the high utility of LLMs. Our method uses strategically designed suffix prompts that effectively thwart a wide range of standard and adaptive jailbreak techniques. Empirical results conducted on Llama-2-7B-Chat and Mistral-7B-Instruct-v0.2 demonstrate the robustness and adaptability of DPP, showing significant reductions in ASR with negligible impact on utility. Our approach not only outperforms existing defense strategies in balancing safety and functionality, but also provides a scalable and robust solution to various LLM platforms.", "keywords": [ "NLP", "AI Safety", "Adversarial Jailbreaking", "Jailbreak Defense" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "weM4YBicIP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chao Liang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Gaojie Lin", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Jianwen Jiang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Jiaqi Yang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Tianyun Zhong", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yanbo Zheng", "gender": "unknown", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 41, "n_ref": 91, "n_ref_all": 109, "n_fig": 7, "n_tab": 6, "L_tab": 2380, "n_element_tab": 167, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 855, "n_element_tab_1": 34, "formula_len_all": 283, "formula_len_all_1": 87, "len_all": 157616, "len_all_1": 65963, "len_abs": 1167, "len_title": 123, "len_sents": 46666, "len_sents_1": 32574, "n_sents": 358, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (conditional oral)", "L_title": 75, "L_abs": 1200, "title": "Loopy: Taming Audio-Driven Portrait Avatar with Long-Term Motion Dependency", "abs": "With the introduction of video diffusion model, audio-conditioned human video generation has recently achieved significant breakthroughs in both the naturalness of motion and the synthesis of portrait details. Due to the limited control of audio signals in driving human motion, existing methods often add auxiliary spatial signals such as movement regions to stabilize movements, which compromise the naturalness and freedom of motion. To address this issue, we propose an end-to-end audio-only conditioned video diffusion model named Loopy. Specifically, we designed two key modules: an inter- and intra-clip temporal module and an audio-to-latents module. These enable the model to better utilize long-term motion dependencies and establish a stronger audio-portrait movement correlation. Consequently, the model can generate more natural and stable portrait videos with subtle facial expressions, without the need for manually setting movement constraints. Extensive experiments show that Loopy outperforms recent audio-driven portrait diffusion models, delivering more lifelike and high-quality results across various scenarios. Video samples are available at https://loopyavataranony.github.io/", "keywords": [ "Diffusion Model", "Avatar", "Portrait Animation", "Audio-Condition Video Generation" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 4, 4, 4, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 3, 3, 4, 4 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "wdzCyr1stL", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Bei Jiang", "gender": "Female", "institution": "University of Alberta", "country": "CA", "position": "Associate Professor" }, { "name": "Jiamei Wu", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Lingchen Kong", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Linglong Kong", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Full Professor" }, { "name": "Yanlin Tang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" } ], "n_formula": 70, "n_formula_1": 39, "n_ref_uni": 23, "n_ref": 49, "n_ref_all": 59, "n_fig": 1, "n_tab": 4, "L_tab": 825, "n_element_tab": 115, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 393, "n_element_tab_1": 56, "formula_len_all": 6594, "formula_len_all_1": 2587, "len_all": 106245, "len_all_1": 58954, "len_abs": 1297, "len_title": 96, "len_sents": 32633, "len_sents_1": 24507, "n_sents": 303, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1301, "title": "Conformal Prediction with Model-Aware Debiasing", "abs": "Bias in model estimation can lead to wider prediction intervals, diminishing the utility of predictive inference. Existing methods have attempted to address this issue, but they often rely on nontrivial assumptions such as specific error distributions or model sparsity, and fail to guarantee coverage in finite samples, which makes their predictions unreliable in practice. To overcome these limitations, we propose a model-aware conformal prediction method that utilizes known model information to achieve debiasing while leaving the unknown aspects, such as data distribution, to the conformal prediction framework. This approach requires only the assumption of exchangeability, making it broadly applicable across various models. Importantly, it retains the finite-sample coverage property and produces shorter prediction intervals compared to existing methods. When applied to threshold ridge regression, we theoretically demonstrate that the model-aware conformal prediction maintains finite-sample marginal coverage and, under certain assumptions, converges to the oracle prediction band, achieving asymptotic conditional validity. Numerical experiments further show that our method outperforms existing methods, providing more efficient prediction intervals across diverse regression datasets.", "keywords": [ "conformal prediction", "model-aware debiasing", "statistical inference", "prediction interval" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "wdmI6A9d2w", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Aryo Lotfi", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Emmanuel Abbe", "gender": "unknown", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "", "position": "Full Professor" }, { "name": "Enrico Fini", "gender": "Male", "institution": "Apple", "country": "CH", "position": "Research Scientist" }, { "name": "Moin Nabi", "gender": "Male", "institution": "Apple Inc.", "country": "DE", "position": "Principal Research Scientist" }, { "name": "Samy Bengio", "gender": "Male", "institution": "Apple", "country": "US", "position": "Senior Director" } ], "n_formula": 12, "n_formula_1": 1, "n_ref_uni": 50, "n_ref": 76, "n_ref_all": 104, "n_fig": 14, "n_tab": 4, "L_tab": 1153, "n_element_tab": 93, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 453, "n_element_tab_1": 36, "formula_len_all": 780, "formula_len_all_1": 14, "len_all": 210664, "len_all_1": 70578, "len_abs": 1890, "len_title": 0, "len_sents": 75623, "len_sents_1": 35518, "n_sents": 560, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1323, "title": "Visual Scratchpads: Enabling Global Reasoning in Vision", "abs": "Modern vision models have achieved remarkable success in benchmarks where a small subset of local features provides critical information about the target. There is now a growing interest in solving tasks that require more global reasoning, where local features offer no significant information. These tasks are reminiscent of the connectivity problems discussed by Minsky and Papert in 1969, which exposed the limitations of the perceptron model and contributed to the first AI winter. In this paper, we revisit such tasks by introducing four global visual benchmarks involving path findings and mazes. We show the following: (1) Although today's large vision models largely surpass the expressivity limitations of the early models, they still struggle with learning efficiency; we introduce the 'globality degree' to understand this; (2) we then demonstrate that the outcome changes and global reasoning becomes feasible with the introduction of a 'visual scratchpad'; similarly to the text scratchpads and chain-of-thoughts used in language models, visual scratchpads help break down global problems into simpler subproblems; (3) we further show that more specific 'inductive scratchpads', which take steps relying on less information, afford better out-of-distribution generalization and succeed for smaller model sizes.", "keywords": [ "reasoning", "scratchpad", "vision", "visual reasoning" ], "rating_list": [ 3, 6, 5, 8 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "wazvIr0Sw0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Francis Engelmann", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Gonca Yilmaz", "gender": "Female", "institution": "Department of Informatics, University of Zurich, University of Zurich", "country": "", "position": "MS student" }, { "name": "Hermann Blum", "gender": "unknown", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "Assistant Professor" }, { "name": "Marc Pollefeys", "gender": "Male", "institution": "Microsoft", "country": "CH", "position": "Director" }, { "name": "Songyou Peng", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 57, "n_ref": 152, "n_ref_all": 174, "n_fig": 13, "n_tab": 6, "L_tab": 383, "n_element_tab": 45, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 244, "n_element_tab_1": 7, "formula_len_all": 598, "formula_len_all_1": 410, "len_all": 168026, "len_all_1": 62706, "len_abs": 1627, "len_title": 70, "len_sents": 48045, "len_sents_1": 30340, "n_sents": 368, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1634, "title": "OpenDAS: Open-Vocabulary Domain Adaptation for Segmentation", "abs": "Recently, Vision-Language Models (VLMs) have advanced segmentation techniques by shifting from the traditional segmentation of a closed-set of predefined object classes to open-vocabulary segmentation (OVS), allowing users to segment novel classes and concepts unseen during training of the segmentation model. However, this flexibility comes with a trade-off: fully-supervised closed-set methods still outperform OVS methods on base classes, that is on classes on which they have been explicitly trained. This is due to the lack of pixel-aligned training masks for VLMs (which are trained on image-caption pairs), and the absence of domain-specific knowledge, such as autonomous driving. Therefore, we propose the task of open-vocabulary domain adaptation to infuse domain-specific knowledge into VLMs while preserving their open-vocabulary nature. By doing so, we achieve improved performance in base and novel classes. Existing VLM adaptation methods improve performance on base (training) queries, but fail to fully preserve the open-set capabilities of VLMs on novel queries. To address this shortcoming, we combine parameter-efficient prompt tuning with a triplet-loss-based training strategy that uses auxiliary negative queries. Notably, our approach is the only parameter-efficient method that consistently surpasses the original VLM on novel classes. Our adapted VLMs can seamlessly be integrated into existing OVS pipelines, e.g., improving OVSeg by +6.0% mIoU on ADE20K for open-vocabulary 2D segmentation, and OpenMask3D by +4.1% AP on ScanNet++ Offices for open-vocabulary 3D instance segmentation without other changes.", "keywords": [ "Computer Vision", "Vision-Language Models", "Domain Adaptation", "Open-Vocabulary Segmentation", "Prompt Tuning" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "waf6HreC53", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Volker Tresp", "gender": "Male", "institution": "Ludwig Maximilian University of Munich", "country": "DE", "position": "Associate Professor" }, { "name": "Yize Sun", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Yunpu Ma", "gender": "Male", "institution": "Siemens Corporate Research", "country": "DE", "position": "Researcher" }, { "name": "Zixin Wu", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 43, "n_ref": 67, "n_ref_all": 82, "n_fig": 8, "n_tab": 4, "L_tab": 975, "n_element_tab": 97, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 364, "n_element_tab_1": 48, "formula_len_all": 660, "formula_len_all_1": 556, "len_all": 121068, "len_all_1": 56511, "len_abs": 1650, "len_title": 117, "len_sents": 33311, "len_sents_1": 27387, "n_sents": 224, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1667, "title": "Quantum Architecture Search With Unsupervised Representation Learning", "abs": "Unsupervised representation learning presents new opportunities for advancing Quantum Architecture Search (QAS) on Noisy Intermediate-Scale Quantum (NISQ) devices. QAS is designed to optimize quantum circuits for Variational Quantum Algorithms (VQAs). Most QAS algorithms tightly couple the search space and search algorithm, typically requiring the evaluation of numerous quantum circuits, resulting in high computational costs and limiting scalability to larger quantum circuits. Predictor-based QAS algorithms mitigate this issue by estimating circuit performance based on structure or embedding. However, these methods often demand time-intensive labeling to optimize gate parameters across many circuits, which is crucial for training accurate predictors. Inspired by the classical neural architecture search algorithm \\textit{Arch2vec}, we investigate the potential of unsupervised representation learning for QAS without relying on predictors. Our framework decouples unsupervised architecture representation learning from the search process, enabling the learned representations to be applied across various downstream tasks. Additionally, it integrates an improved quantum circuit graph encoding scheme, addressing the limitations of existing representations and enhancing search efficiency. This predictor-free approach removes the need for large labeled datasets. During the search, we employ REINFORCE and Bayesian Optimization to explore the latent representation space and compare their performance against baseline methods. Our results demonstrate that the framework efficiently identifies high-performing quantum circuits with fewer search iterations.", "keywords": [ "Quantum Circuit Architecture Search", "QAS", "unsupervised representation learning" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "waIltEWDr8", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Christian Wachinger", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "Professor" }, { "name": "Tom Nuno Wolf", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 13, "n_ref_uni": 34, "n_ref": 63, "n_ref_all": 84, "n_fig": 8, "n_tab": 1, "L_tab": 610, "n_element_tab": 115, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 281, "n_element_tab_1": 23, "formula_len_all": 1456, "formula_len_all_1": 857, "len_all": 126212, "len_all_1": 61730, "len_abs": 372, "len_title": 153, "len_sents": 43781, "len_sents_1": 30957, "n_sents": 296, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 105, "L_abs": 1444, "title": "WASUP: Interpretable Classification with Weight-Input Alignment and Class-Discriminative SUPports Vectors", "abs": "The deployment of deep learning models in critical domains necessitates a balance between high accuracy and interpretability.\nWe introduce WASUP, an inherently interpretable neural network that provides local and global explanations of its decision-making process.\nWe prove that these explanations are faithful by fulfilling established axioms for explanations. \nLeveraging the concept of case-based reasoning, WASUP extracts class-representative support vectors from training images, ensuring they capture relevant features while suppressing irrelevant ones.\nClassification decisions are made by calculating and aggregating similarity scores between these support vectors and the input's latent feature vector. \nWe employ B-Cos transformations, which align model weights with inputs to enable faithful mappings of latent features back to the input space, facilitating local explanations in addition to global explanations of case-based reasoning.\nWe evaluate WASUP on three tasks: fine-grained classification on Stanford Dogs, multi-label classification on Pascal VOC, and pathology detection on the RSNA dataset.\nResults indicate that WASUP not only achieves competitive accuracy compared to state-of-the-art black-box models but also offers insightful explanations verified through theoretical analysis.\nOur findings underscore WASUPs potential for applications where understanding model decisions is as critical as the decisions themselves.", "keywords": [ "explainability", "interpretability", "case-based reasoning" ], "rating_list": [ 1, 5, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 2, 3 ], "presentation_list": [ 1, 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 3, 3 ] }, { "paper_id": "waHmD2i1dv", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Fan Chen", "gender": "Male", "institution": "Southwest Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Hongjie He", "gender": "Female", "institution": "Southwest Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Jingzehua Xu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Shuai Zhang", "gender": "Male", "institution": "New Jersey Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Wenhao Feng", "gender": "Male", "institution": "Southwest Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Xin Lai", "gender": "Male", "institution": "Southwest Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Yubo Huang", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Zixi Wang", "gender": "Not Specified", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 37, "n_formula_1": 12, "n_ref_uni": 61, "n_ref": 86, "n_ref_all": 105, "n_fig": 7, "n_tab": 6, "L_tab": 1640, "n_element_tab": 302, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1289, "n_element_tab_1": 268, "formula_len_all": 2961, "formula_len_all_1": 891, "len_all": 182338, "len_all_1": 72514, "len_abs": 1419, "len_title": 115, "len_sents": 53027, "len_sents_1": 32506, "n_sents": 368, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1423, "title": "CausalVE: Face Video Privacy Encryption via Causal Video Prediction", "abs": "Advanced facial recognition technologies and recommender systems with inadequate privacy technologies and policies for facial interactions increase concerns about bioprivacy violations. With the proliferation of video and live-streaming websites, public-face video distribution and interactions pose greater privacy risks. Existing techniques typically address the risk of sensitive biometric information leakage through various privacy enhancement methods but pose a higher security risk by corrupting the information to be conveyed by the interaction data, or by leaving certain biometric features intact that allow an attacker to infer sensitive biometric information from them. To address these shortcomings, in this paper, we propose a neural network framework, CausalVE. We obtain cover images by adopting a diffusion model to achieve face swapping with face guidance and use the speech sequence features and spatiotemporal sequence features of the secret video for dynamic video inference and prediction to obtain a cover video with the same number of frames as the secret video. In addition, we hide the secret video by using reversible neural networks for video hiding so that the video can also disseminate secret data. Numerous experiments prove that our CausalVE has good security in public video dissemination and outperforms state-of-the-art methods from a qualitative, quantitative, and visual point of view.", "keywords": [ "Bioprivacy", "Diffusion model", "Face swapping", "Video Prediction", "Reversible neural networks", "Video Hiding" ], "rating_list": [ 5, 1, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 1, 1, 2 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 2, 4, 4, 5, 4 ] }, { "paper_id": "waGoVEQvT9", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Adam Kortylewski", "gender": "unknown", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Research Group Leader" }, { "name": "Alan Yuille", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Artur Jesslen", "gender": "Male", "institution": "University of Freiburg, Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Guofeng Zhang", "gender": "Male", "institution": "John Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Prakhar Kaushik", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Xiaoding Yuan", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 26, "n_ref": 63, "n_ref_all": 79, "n_fig": 4, "n_tab": 4, "L_tab": 1124, "n_element_tab": 122, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 580, "n_element_tab_1": 90, "formula_len_all": 748, "formula_len_all_1": 628, "len_all": 113863, "len_all_1": 62789, "len_abs": 2084, "len_title": 125, "len_sents": 33863, "len_sents_1": 29670, "n_sents": 233, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 2095, "title": "Scaling 3D Compositional Models for Robust Classification and Pose Estimation", "abs": "Deep learning algorithms for object classification and 3D object pose estimation lack robustness to out-of-distribution factors such as synthetic stimuli, changes in weather conditions, and partial occlusion. Human vision, however, is typically much more robust to all these factors. This is arguably because human vision exploits 3D object representations which are invariant to most of these factors. Recently a class of 3D compositional models have been developed where objects are represented in terms of 3D meshes, with typically 1000 vertices associated with learnt vertex features. These models have shown robustness in small-scale settings, involving 10 or 12 objects, but it is unclear that they can be scaled up to 100s of object classes. The main problem is that their training involves supervised contrastive learning on the mesh vertices representing the objects and requires each vertex to be contrasted with all other vertices, which scales quadratically with the vertex number. A newly available dataset with 3D annotations for 188 object classes allows us to address this scaling challenge. We present a strategy which exploits the compositionality of the objects, i.e. the independence of the feature vectors of the vertices, which greatly reduces the training time while also improving the performance of the algorithms. We first refactor the per-vertex contrastive learning into contrasting within class and between classes. Then we propose a process that dynamically decouples the contrast between classes which are rarely confused, and enhances the contrast between the vertices of classes that are most confused. Our large-scale 3D compositional model not only achieves state-of-the-art performance on object classification and 3D pose estimation in a unified manner surpassing ViT and ResNet, but is also more robust to out-of-distribution testing including occlusion, weather conditions, and synthetic data. This paves the way for scalable 3D object understanding and opens exciting possibilities for applications in robotics, autonomous systems, and augmented reality.", "keywords": [ "analysis by synthesis", "image classification", "3D representation", "compositional models" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "wZiH43e5Ah", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Hossein Moghaddam", "gender": "unknown", "institution": "T\u00e9l\u00e9com Paris", "country": "FR", "position": "Researcher" }, { "name": "Jayneel Parekh", "gender": "unknown", "institution": "Sorbonne Universit\u00e9", "country": "FR", "position": "Postdoc" }, { "name": "Pavlo Mozharovskyi", "gender": "Male", "institution": "T\u00e9l\u00e9com Paris", "country": "FR", "position": "Full Professor" }, { "name": "Quentin Bouniot", "gender": "Male", "institution": "Helmholtz Zentrum M\u00fcnchen", "country": "DE", "position": "Postdoc" }, { "name": "Florence d'Alch\u00e9", "gender": "Female", "institution": "T\u00e9l\u00e9com Paris", "country": "", "position": "Full Professor" } ], "n_formula": 29, "n_formula_1": 26, "n_ref_uni": 34, "n_ref": 62, "n_ref_all": 73, "n_fig": 5, "n_tab": 3, "L_tab": 715, "n_element_tab": 67, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 250, "n_element_tab_1": 30, "formula_len_all": 2248, "formula_len_all_1": 2203, "len_all": 114021, "len_all_1": 58819, "len_abs": 1276, "len_title": 156, "len_sents": 32467, "len_sents_1": 24896, "n_sents": 221, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 108, "L_abs": 1282, "title": "Conceptualize Any Network: A Concept Extraction Framework for Holistic Interpretability of Image Classifiers", "abs": "Attribution-based and concept-based methods dominate the area of post-hoc explainability for vision classifiers. While attribution-based methods highlight crucial regions of the input images to justify model predictions, concept-based methods provide explanations rooted in high-level properties that are generally more understandable for humans. In this work, we introduce ``Conceptualize Any Network'' (CAN), a comprehensive post-hoc explanation framework that combines the wide scope of attribution-based methods and the understandability of concept-based methods. \nDesigned to be model agnostic, CAN is capable of explaining any network that allows for the extraction of feature attribution maps, expanding its applicability to both CNNs and Vision Transformers (ViTs). Moreover, unlike existing concept-based methods for vision classifiers, CAN extracts a set of concepts shared across all classes, enabling a unified explanation of the model as a whole.\nExtensive numerical experiments across different architectures, datasets, and feature attribution methods showcase the capabilities of CAN in Conceptualizing Any Network faithfully, concisely, and consistently.\nFurthermore, we managed to scale our framework to all of ImageNet's classes which has not been achieved before.", "keywords": [ "Explainability", "Computer Vision", "CNN", "ViT" ], "rating_list": [ 3, 1, 3, 5 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "wZbkQStAXj", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Dequan Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jialing Zhang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Jin Gao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Lingfeng Zhou", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Mohan Jiang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 16, "n_ref": 31, "n_ref_all": 46, "n_fig": 5, "n_tab": 7, "L_tab": 4574, "n_element_tab": 390, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2574, "n_element_tab_1": 309, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 126598, "len_all_1": 61386, "len_abs": 1563, "len_title": 111, "len_sents": 47027, "len_sents_1": 27653, "n_sents": 283, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1568, "title": "PersonaEval: Benchmarking LLMs on Role-Playing Evaluation Tasks", "abs": "Role-playing in large language models (LLMs) has become a crucial area of research, enabling models to simulate diverse personas and tailor responses, significantly impacting natural language understanding and human-computer interaction. However, while advanced LLMs like GPT-4 are used to evaluate role-playing methods, their reliability in providing accurate assessments remains uncertain, especially in distinguishing nuanced role-playing characteristics. In this paper, we introduce PersonaEval, a benchmark designed to assess the effectiveness of LLMs in role-playing evaluation tasks. We frame the problem as a classification task to determine whether an LLM evaluator can distinguish between sentences from different levels of expertise based solely on linguistic cues. Using real-world data from the Wired 5 Levels video series\u2014where experts explain concepts to five distinct audiences: a child, a teenager, a college student, a graduate student, and another expert\u2014we design three evaluation settings that correspond to commonly used LLM evaluation approaches: single answer role grading, pairwise role comparison, and reference-guided role grading. These settings aim to capture various aspects of how effectively LLMs evaluate role-playing performance. Our study highlights the limitations of current LLMs in persona evaluation tasks and underscores the need for further research to enhance their evaluation capabilities. We provide a foundation for future work aimed at improving the accuracy and professionalism of LLM evaluators in role-playing contexts.", "keywords": [ "Role-playing", "evaluating evaluators" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "wYxOMEzpkl", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Di Wang", "gender": "Male", "institution": "Independent researcher", "country": "US", "position": "Researcher" }, { "name": "Zhanxing Zhu", "gender": "Male", "institution": "University of Southampton", "country": "", "position": "Associate Professor" }, { "name": "Bochen Lv", "gender": "Male", "institution": "University of Southampton", "country": "GB", "position": "PhD student" } ], "n_formula": 202, "n_formula_1": 31, "n_ref_uni": 39, "n_ref": 88, "n_ref_all": 262, "n_fig": 17, "n_tab": 4, "L_tab": 1198, "n_element_tab": 136, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 16377, "formula_len_all_1": 2441, "len_all": 263862, "len_all_1": 73941, "len_abs": 1179, "len_title": 92, "len_sents": 72323, "len_sents_1": 30982, "n_sents": 643, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 44, "L_abs": 1132, "title": "A Solvable Attention for Neural Scaling Laws", "abs": "Transformers and many other deep learning models are empirically shown to predictably enhance their performance as a power law in training time, model size, or the number of training data points, which is termed as the neural scaling law. This paper studies this intriguing phenomenon particularly for the transformer architecture in theoretical setups. Specifically, we propose a framework for self-attention, the underpinning block of transformer, to learn in an in-context manner, where the corresponding learning dynamics is modeled as a non-linear ordinary differential equation (ODE) system. Furthermore, we establish a procedure to derive a tractable solution for this ODE system by reformulating it as a Riccati equation, which allows us to precisely characterize neural scaling laws for self-attention with training time, model size, data size, and the optimal compute. In addition, we reveal that the self-attention shares similar neural scaling laws with several other architectures when the context sequence length of the in-context learning is fixed, otherwise it would exhibit a different scaling law of training time.", "keywords": [ "self-attention", "scaling laws", "solution of learning dynamics" ], "rating_list": [ 8, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "wYZ8rxwvMm", "primary_area": "reinforcement learning", "n_author": 13, "author_info_list": [ { "name": "Chao Zhang", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "David Zhu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "MS student" }, { "name": "Gagandeep Singh", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Haoming Jiang", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Monica Xiao Cheng", "gender": "Female", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Qi Zeng", "gender": "Female", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Qingyu Yin", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Rohan Gumaste", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Undergrad student" }, { "name": "Ruirui Li", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Tarun Suresh", "gender": "Male", "institution": "Department of Computer Science", "country": "", "position": "Undergrad student" }, { "name": "Xianfeng Tang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yinglun Xu", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" }, { "name": "Zhengyang Wang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 26, "n_ref": 42, "n_ref_all": 47, "n_fig": 3, "n_tab": 1, "L_tab": 926, "n_element_tab": 76, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 926, "n_element_tab_1": 76, "formula_len_all": 797, "formula_len_all_1": 797, "len_all": 88481, "len_all_1": 63696, "len_abs": 1064, "len_title": 131, "len_sents": 29386, "len_sents_1": 29208, "n_sents": 214, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1069, "title": "Two-Step Offline Preference-Based Reinforcement Learning with Constrained Actions", "abs": "Preference-based reinforcement learning (PBRL) in the offline setting has succeeded greatly in industrial applications such as chatbots. A two-step learning framework where one applies a reinforcement learning step after a reward modeling step has been widely adopted for the problem. However, such a method faces challenges from the risk of reward hacking and the complexity of reinforcement learning. To overcome the challenge, our insight is that both challenges come from the state-actions not supported in the dataset. Such state-actions are unreliable and increase the complexity of the reinforcement learning problem at the second step. Based on the insight, we develop a novel two-step learning method called PRC: preference-based reinforcement learning with constrained actions. The high-level idea is to limit the reinforcement learning agent to optimize over a constrained action space that excludes the out-of-distribution state-actions. We empirically verify that our method has high learning efficiency on various datasets in robotic control environments.", "keywords": [ "Preference Based Reinforcement Learning", "Offline Reinforcement Learning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "wYWJFLQov9", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Beining Yang", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Cheng Ji", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Jianxin Li", "gender": "Male", "institution": "Beihang University ", "country": "CN", "position": "Full Professor" }, { "name": "Qingyun Sun", "gender": "Female", "institution": "Beihang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xingcheng Fu", "gender": "Male", "institution": "Guangxi Normal University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 53, "n_ref": 117, "n_ref_all": 158, "n_fig": 12, "n_tab": 12, "L_tab": 6056, "n_element_tab": 681, "n_fig_1": 7, "n_tab_1": 7, "L_tab_1": 2821, "n_element_tab_1": 342, "formula_len_all": 828, "formula_len_all_1": 1227, "len_all": 175943, "len_all_1": 75992, "len_abs": 2284, "len_title": 117, "len_sents": 41356, "len_sents_1": 30560, "n_sents": 345, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1535, "title": "ST-GCond: Self-supervised and Transferable Graph Dataset Condensation", "abs": "The increasing scale of graph datasets significantly enhances deep learning models but also presents substantial training challenges. Graph dataset condensation has emerged to condense large datasets into smaller yet informative ones that maintain similar test performance. However, these methods require downstream usage to match the original dataset and task, which is impractical in real-world scenarios. Our empirical studies show that existing methods fail in \"cross-task\" and \"cross-dataset\" scenarios, often performing worse than training from scratch. To address these challenges, we propose a novel method: Self-supervised and Transferable Graph dataset Condensation (ST-GCond). For cross-task transferability, we propose a task-disentangled meta optimization strategy to adaptively update the condensed graph according to the task relevance, encouraging information preservation for various tasks. For cross-dataset transferability, we propose a multi-teacher self-supervised optimization strategy to incorporate auxiliary self-supervised tasks to inject universal knowledge into the condensed graph. Additionally, we incorporate mutual information guided joint condensation mitigating the potential conflicts and ensure the condensing stability. Experiments on both node-level and graph-level datasets show that ST-GCond outperforms existing methods by 2.5% to 18.7% in all cross-task and cross-dataset scenarios, and also achieves state-of-the-art performance on 5 out of 6 datasets in the single dataset and task scenario.", "keywords": [ "Graph Neural Network; Graph Dataset Condensation" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "wYVP4g8Low", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Duy Khoa Pham", "gender": "Male", "institution": "Swinburne University of Technology", "country": "AU", "position": "Undergrad student" }, { "name": "Hung Du", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "PhD student" }, { "name": "Hy Nguyen", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "PhD student" }, { "name": "Kon Mouzakis", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Full Professor" }, { "name": "Rajesh Vasa", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Full Professor" }, { "name": "Srikanth Thudumu", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 10, "n_ref_uni": 20, "n_ref": 33, "n_ref_all": 43, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1082, "formula_len_all_1": 697, "len_all": 87556, "len_all_1": 53959, "len_abs": 1494, "len_title": 140, "len_sents": 31902, "len_sents_1": 24968, "n_sents": 241, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1501, "title": "Local Control Networks (LCNs): Optimizing Flexibility in Neural Network Data Pattern Capture", "abs": "The widespread use of multilayer perceptrons (MLPs) often relies on a fixed activation function (e.g., ReLU, Sigmoid, Tanh) for all nodes within the hidden layers. While effective in many scenarios, this uniformity may limit the network\u2019s ability to capture complex data patterns. We argue that employing the same activation function at every node is suboptimal and propose leveraging different activation functions at each node to increase flexibility and adaptability. To achieve this, we introduce Local Control Networks (LCNs), which leverage B-spline functions to enable distinct activation curves at each node. Our mathematical analysis demonstrates the properties and benefits of LCNs over conventional MLPs. In addition, we demonstrate that more complex architectures, such as Kolmogorov\u2013Arnold Networks (KANs), are unnecessary in certain scenarios, and LCNs can be a more efficient alternative. Empirical experiments on various benchmarks and datasets validate our theoretical findings. In computer vision tasks, LCNs achieve marginal improvements over MLPs and outperform KANs by approximately 5%, while also being more computationally efficient than KANs. In basic machine learning tasks, LCNs show a 1% improvement over MLPs and a 0.6% improvement over KANs. For symbolic formula representation tasks, LCNs perform on par with KANs, with both architectures outperforming MLPs. Our findings suggest that diverse activations at the node level can lead to improved performance and efficiency.", "keywords": [ "Optimization", "Learning Representation", "Neural Network", "Activation Function" ], "rating_list": [ 3, 3, 1, 5 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "wYJII5BRYU", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Aleksandr Panov", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "Evgenii Aleksandrovich Dzhivelikian", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "", "position": "PhD student" }, { "name": "Petr Kuderov", "gender": "Non-Binary", "institution": "Moscow Institute of Physics and Technology", "country": "", "position": "Instructor" } ], "n_formula": 37, "n_formula_1": 24, "n_ref_uni": 53, "n_ref": 80, "n_ref_all": 105, "n_fig": 14, "n_tab": 2, "L_tab": 245, "n_element_tab": 30, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2064, "formula_len_all_1": 1369, "len_all": 193886, "len_all_1": 68726, "len_abs": 1032, "len_title": 116, "len_sents": 63680, "len_sents_1": 31364, "n_sents": 493, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1035, "title": "Learning Successor Features with Distributed Hebbian Temporal Memory", "abs": "This paper presents a novel approach to address the challenge of online sequence learning for decision making under uncertainty in non-stationary, partially observable environments. The proposed algorithm, Distributed Hebbian Temporal Memory (DHTM), is based on the factor graph formalism and a multi-component neuron model. DHTM aims to capture sequential data relationships and make cumulative predictions about future observations, forming Successor Features (SFs). Inspired by neurophysiological models of the neocortex, the algorithm uses distributed representations, sparse transition matrices, and local Hebbian-like learning rules to overcome the instability and slow learning of traditional temporal memory algorithms such as RNN and HMM. Experimental results show that DHTM outperforms LSTM, RWKV and a biologically inspired HMM-like algorithm, CSCG, on non-stationary data sets. Our results suggest that DHTM is a promising approach to address the challenges of online sequence learning and planning in dynamic environments.", "keywords": [ "temporal memory", "successor features", "online learning", "Hebbian learning" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 2, 2, 3, 4 ] }, { "paper_id": "wY5DE4Iuc8", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Bin.Yang.", "gender": "Female", "institution": "Southern Methodist University, Southern Methodist University", "country": "CN", "position": "Full Professor" }, { "name": "Qifan Zeng", "gender": "unknown", "institution": "Ocean University of China", "country": "", "position": "Full Professor" }, { "name": "Yantong Cai", "gender": "Female", "institution": "Dermatology Hospital, Southern Medical University", "country": "CN", "position": "Postdoc" }, { "name": "Yong Tang", "gender": "Male", "institution": "Winroad Holdings Ltd", "country": "", "position": "Vice President" }, { "name": "Jiake Xie", "gender": "Male", "institution": "Winroad Holdings Limited", "country": "", "position": "Computer Vision Engineer" }, { "name": "yifan hong", "gender": "Male", "institution": "southern medical university", "country": "", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 37, "n_ref": 51, "n_ref_all": 62, "n_fig": 8, "n_tab": 2, "L_tab": 364, "n_element_tab": 54, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 254, "n_element_tab_1": 40, "formula_len_all": 403, "formula_len_all_1": 404, "len_all": 112213, "len_all_1": 55997, "len_abs": 1067, "len_title": 153, "len_sents": 28383, "len_sents_1": 28299, "n_sents": 174, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1072, "title": "INRscrecon: Enhancing 3D Spatial Transcriptomics Reconstruction through Implicit Neural Representations", "abs": "Single-cell spatial transcriptomics (scST) technologies have revolutionized our understanding of the complex three-dimensional cellular landscapes of tissues. However, the accuracy of spatial expression profiles is often compromised by missing or distorted experimental data. To address this challenge, we introduce INRscrecon, a novel framework that leverages Implicit Neural Representations (INRs) known for their continuous signal encoding capabilities. INRscrecon accurately predicts and corrects spatial expressions, enhancing the clarity of 3D tissue reconstructions. Our study demonstrates the efficacy of INRscrecon across various datasets and dimensions, highlighting its potential to restore spatial expression with high precision. The findings suggest broader applications for INR-based methodologies in spatial transcriptomics, paving the way for more accurate and detailed analysis of cellular interactions within tissues. Future research may expand on the incorporation of INR techniques in spatial transcriptomics to further enhance analytical capabilities.", "keywords": [ "Spatial Transcriptomics reconstruction", "Implicit Neural Representations", "alignment" ], "rating_list": [ 1, 5, 5, 3 ], "soundness_list": [ 1, 1, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 2, 2, 4 ] }, { "paper_id": "wXIncJRlK0", "primary_area": "reinforcement learning", "n_author": 1, "author_info_list": [ { "name": "Ryo Iwaki", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Research Scientist" } ], "n_formula": 104, "n_formula_1": 36, "n_ref_uni": 41, "n_ref": 72, "n_ref_all": 108, "n_fig": 11, "n_tab": 1, "L_tab": 446, "n_element_tab": 27, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12761, "formula_len_all_1": 3827, "len_all": 163434, "len_all_1": 68959, "len_abs": 1002, "len_title": 106, "len_sents": 40594, "len_sents_1": 28667, "n_sents": 441, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1004, "title": "Mirror Descent Actor Critic via Bounded Advantage Learning", "abs": "Regularization is a core component of recent Reinforcement Learning (RL) algorithms. Mirror Descent Value Iteration (MDVI) uses both Kullback-Leibler divergence and entropy as regularizers in its value and policy updates. Despite its empirical success in discrete action domains and strong theoretical garantees, the performance improvement of a MDVI-based method over the entropy-only-regularized RL is limited in continuous action domains. In this study, we propose Mirror Descent Actor Critic (MDAC) as an actor-critic style instantiation of MDVI for continuous action domains, and show that its empirical performance is significantly boosted by bounding the values of actor's log-density terms in the critic's loss function. Further, we relate MDAC to Advantage Learning by recalling that the actor's log-probability is equal to the regularized advantage function in tabular cases, and theoretically show that the error of optimal policy misspecification is decreased by bounding the advantage terms.", "keywords": [ "reinforcement learning", "regularization", "KL divergence", "entropy", "actor critic" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 3, 3, 3, 1 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "wWpChcKLwB", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Feng Yin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Juntao Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "PhD student" }, { "name": "Renjie Li", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Postdoc" }, { "name": "Richard Cornelius Suwandi", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Tsung-Hui Chang", "gender": "Not Specified", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Full Professor" } ], "n_formula": 74, "n_formula_1": 15, "n_ref_uni": 54, "n_ref": 99, "n_ref_all": 140, "n_fig": 19, "n_tab": 7, "L_tab": 4174, "n_element_tab": 297, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1111, "n_element_tab_1": 80, "formula_len_all": 4590, "formula_len_all_1": 917, "len_all": 239052, "len_all_1": 68220, "len_abs": 1673, "len_title": 128, "len_sents": 72857, "len_sents_1": 29286, "n_sents": 622, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1684, "title": "Context-Aware Kernel Search for Bayesian Optimization with Large Language Models", "abs": "The efficiency of Bayesian optimization (BO) relies on careful selection of the surrogate model to balance exploration and exploitation under limited budget. Traditional BO methods often struggle with sub-optimal kernel choices when using Gaussian processes (GPs) as the surrogate model. When the kernel is inadequately chosen, BO may converge slowly or even get stuck at an undesired local minimum. To address such drawback, we propose the novel Context-Aware Kernel Search (CAKES) to automate optimal kernel design in BO with large language models (LLMs). Concretely, CAKES exploits LLMs as crossover and mutation operators to adaptively generate and refine GP kernels based on the observed data. CAKES works entirely in-context and can be easily integrated into existing systems without requiring any fine-tuning. We further present a theoretical analysis demonstrating that our method achieves sub-linear regret relative to the budget for any input dimension. Experimental results demonstrate that CAKES outperforms various salient baseline methods in numerous synthetic and real-world optimization tasks. Notably, CAKES improves the overall performance on benchmark functions by roughly 36\\%. In hyperparameter tuning tasks, CAKES can effectively leverage fewer data samples to quickly identify high-performing configurations and consistently ranks first across various datasets. As an encouraging real application, we successfully applied CAKES to design photonic chips, achieving significant improvements in key performance indicators while speeding up the design cycle by a factor of ten compared to the baselines. Our code is accessible at https://github.com/cakes4bo/cakes.", "keywords": [ "Bayesian optimization", "Gaussian processes", "kernel design", "large language models" ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 4, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "wWnsoLhHwt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Christopher Ackerman", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Nina Rimsky", "gender": "Female", "institution": "Anthropic", "country": "GB", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 17, "n_ref": 26, "n_ref_all": 54, "n_fig": 18, "n_tab": 4, "L_tab": 774, "n_element_tab": 41, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 153, "n_element_tab_1": 25, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 118276, "len_all_1": 55084, "len_abs": 1443, "len_title": 135, "len_sents": 48778, "len_sents_1": 29812, "n_sents": 306, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1449, "title": "Inspection and Control of Self-Generated-Text Recognition Ability in Llama3-8b-Instruct", "abs": "It has been reported that LLMs can recognize their own writing. As this has potential implications for AI safety, yet is relatively understudied, we investigate the phenomenon, seeking to establish: whether it robustly occurs at the behavioral level, how the observed behavior is achieved, and whether it can be controlled. First, we find that the Llama3-8b\u2013Instruct chat model - but not the base Llama3-8b model - can reliably distinguish its own outputs from those of humans, and present evidence that the chat model is likely using its experience with its own outputs, acquired during post-training, to succeed at the writing recognition task. Second, we identify a vector in the residual stream of the model that is differentially activated when the model makes a correct self-written-text recognition judgment, show that the vector activates in response to information relevant to self-authorship, present evidence that the vector is related to the concept of ``self'' in the model, and demonstrate that the vector is causally related to the model\u2019s ability to perceive and assert self-authorship. Finally, we show that the vector can be used to control both the model\u2019s behavior and its perception, steering the model to claim or disclaim authorship by applying the vector to the model\u2019s output as it generates it, and steering the model to believe or disbelieve it wrote arbitrary texts by applying the vector to them as the model reads them.", "keywords": [ "LLM", "Interpretability", "AI", "Activation Steering", "Representation Engineering", "Control" ], "rating_list": [ 8, 5, 3, 8 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "wWhZ2RFAxF", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Allon Adir", "gender": "unknown", "institution": "International Business Machines", "country": "IL", "position": "Researcher" }, { "name": "Ehud Aharoni", "gender": "Not Specified", "institution": "International Business Machines", "country": "IL", "position": "Researcher" }, { "name": "Itamar Zimerman", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Jenny Lerner", "gender": "Female", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Matan Avitan", "gender": "Male", "institution": "Bar-Ilan University", "country": "", "position": "MS student" }, { "name": "Moran Baruch", "gender": "Female", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Nir Drucker", "gender": "Male", "institution": "International Business Machines", "country": "IL", "position": "IBM Research, Israel" }, { "name": "Omri Soceanu", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Ramy Masalha", "gender": "Male", "institution": "University of Haifa", "country": "IL", "position": "PhD student" }, { "name": "Reut Moshe", "gender": "unknown", "institution": "Ben Gurion University of the Negev", "country": "", "position": "MSc Student in Electrical and Computer Engineering" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 30, "n_ref": 61, "n_ref_all": 86, "n_fig": 11, "n_tab": 3, "L_tab": 744, "n_element_tab": 56, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 618, "n_element_tab_1": 53, "formula_len_all": 807, "formula_len_all_1": 770, "len_all": 127003, "len_all_1": 67017, "len_abs": 1408, "len_title": 111, "len_sents": 37609, "len_sents_1": 30943, "n_sents": 279, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1412, "title": "PowerSoftmax: Towards secure LLM Inference Over Encrypted Data", "abs": "Modern cryptographic methods for implementing privacy-preserving LLMs such as Homomorphic Encryption require the LLMs to have a polynomial form. Forming such a representation is challenging because Transformers include non-polynomial components, such as Softmax and layer normalization. Previous approaches have either directly approximated pre-trained models with large-degree polynomials, which are less efficient over HE, or replaced non-polynomial components with easier-to-approximate primitives before training, e.g., Softmax with pointwise attention. The latter approach might introduce scalability challenges. \n\nWe present a new HE-friendly variant of self-attention that offers a stable form for training and is easy to approximate with polynomials for secure inference. Our work introduces the first polynomial LLMs with 32 layers and over a billion parameters, exceeding the size of previous models by more than tenfold. The resulting models demonstrate reasoning and in-context learning (ICL) capabilities comparable to standard transformers of the same size, representing a breakthrough in the field. Finally, we provide a detailed latency breakdown for each computation over encrypted data, paving the way for further optimization, and explore the differences in inductive bias between transformers relying on our HE-friendly variant and standard transformers. Our code is attached as a supplement.", "keywords": [ "Secure LLMs", "Secure Transformers", "Privacy Preserving", "Homomorphic Encryption (HE)" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 4 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "wWcNhS4g1U", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Jiajun Wu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Matt Zhou", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Intern" }, { "name": "Shangzhe Wu", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Assistant Professor" }, { "name": "Yunzhi Zhang", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Zizhang Li", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 3, "n_ref_uni": 32, "n_ref": 60, "n_ref_all": 98, "n_fig": 7, "n_tab": 9, "L_tab": 4303, "n_element_tab": 195, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 865, "n_element_tab_1": 42, "formula_len_all": 1391, "formula_len_all_1": 261, "len_all": 226667, "len_all_1": 61749, "len_abs": 1057, "len_title": 124, "len_sents": 80858, "len_sents_1": 28325, "n_sents": 647, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1060, "title": "The Scene Language: Representing Scenes with Programs, Words, and Embeddings", "abs": "We introduce the Scene Language, a visual scene representation that concisely and precisely describes the structure, semantics, and identity of visual scenes. The Scene Language represents a scene with three key components: a program that specifies the hierarchical and relational structure of entities in the scene, words in natural language that summarize the semantic class of each entity, and embeddings that capture the visual identity of each entity. This representation can be inferred from pre-trained language models via a training-free inference technique, given text or image inputs. The resulting scene can be rendered into images using traditional, neural, or hybrid graphics renderers. Together, this forms a robust, fully automated system for high-quality 3D and 4D scene generation. Compared with existing representations like scene graphs, our proposed Scene Language generates complex scenes with higher fidelity, while explicitly modeling the scene structures to enable precise control and editing. Project page: https://sclg-page.github.io/", "keywords": [ "3D scene generation; visual programs" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "wWPiAjbR7a", "primary_area": "applications to neuroscience & cognitive science", "n_author": 7, "author_info_list": [ { "name": "CHENG LI", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Intern" }, { "name": "Chi Han", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" }, { "name": "Heng Ji", "gender": "Female", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Jindong Wang", "gender": "Male", "institution": "William & Mary", "country": "US", "position": "Assistant Professor" }, { "name": "Manling Li", "gender": "Female", "institution": "Northwestern University", "country": "US", "position": "Assistant Professor" }, { "name": "Qingyun Wang", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" }, { "name": "Yi Fung", "gender": "F", "institution": "Hong Kong University of Science and Technology", "country": "US", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 35, "n_ref": 72, "n_ref_all": 101, "n_fig": 12, "n_tab": 12, "L_tab": 6474, "n_element_tab": 531, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 1984, "n_element_tab_1": 30, "formula_len_all": 341, "formula_len_all_1": 134, "len_all": 186339, "len_all_1": 58170, "len_abs": 1519, "len_title": 154, "len_sents": 56712, "len_sents_1": 28159, "n_sents": 494, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 105, "L_abs": 1417, "title": "MentalArena: Self-play Training of Language Models for Diagnosis and Treatment of Mental Health Disorders", "abs": "Mental health disorders are one of the most serious diseases in the world. Most people with such a disease lack access to adequate care, which highlights the importance of training models for the diagnosis and treatment of mental health disorders. However, in the mental health domain, privacy concerns limit the accessibility of personalized treatment data, making it challenging to build powerful models.\nIn this paper, we introduce MentalArena, a self-play framework to train language models by generating domain-specific personalized data, where we obtain a better model capable of making a personalized diagnosis and treatment (as a therapist) and providing information (as a patient). To accurately model human-like mental health patients, we devise Symptom Encoder which simulates a real patient from both cognition and behavior perspectives. To address intent bias during patient-therapist interactions, we propose Symptom Decoder to compare diagnosed symptoms with encoded symptoms, and dynamically manage the dialogue between patient and therapist according to the identified deviations. We evaluated MentalArena against $6$ benchmarks, including biomedicalQA and mental health tasks, compared to $6$ advanced models. Our models, fine-tuned on both GPT-3.5 and Llama-3-8b, significantly outperform their counterparts, including GPT-4o. We hope that our work can inspire future research on personalized care.", "keywords": [ "Mental health", "Self-play", "Co-evolve", "Iterative training" ], "rating_list": [ 6, 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3, 4 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 2, 4 ] }, { "paper_id": "wVmShpwtY0", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Jiahao Wang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Intern" }, { "name": "SHUANGJIA ZHENG", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 33, "n_ref": 55, "n_ref_all": 66, "n_fig": 6, "n_tab": 4, "L_tab": 2124, "n_element_tab": 180, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1250, "n_element_tab_1": 193, "formula_len_all": 1535, "formula_len_all_1": 1313, "len_all": 111840, "len_all_1": 59086, "len_abs": 1273, "len_title": 119, "len_sents": 33885, "len_sents_1": 27751, "n_sents": 230, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1278, "title": "Efficient Protein Optimization via Structure-aware Hamiltonian Dynamics", "abs": "The ability to engineer optimized protein variants has transformative potential for biotechnology and medicine. Prior sequence-based optimization methods struggle with the high-dimensional complexities due to the epistasis effect and the disregard for structural constraints. To address this, we propose HADES, a Bayesian optimization method utilizing Hamiltonian dynamics to efficiently sample from a structure-aware approximated posterior. Leveraging momentum and uncertainty in the simulated physical movements, HADES enables rapid transition of proposals toward promising areas. A position discretization procedure is introduced to propose discrete protein sequences from such continuous state system. The posterior surrogate is powered by a two-stage encoder-decoder framework to determine the structure and function relationships between mutant neighbors, consequently learning a smoothed landscape to sample from. Extensive experiments demonstrate that our method outperforms state-of-the-art baselines in in-silico evaluations across most metrics. Remarkably, our approach offers a unique advantage by leveraging the mutual constraints between protein structure and sequence, facilitating the design of protein sequences with similar structures and optimized properties.", "keywords": [ "protein engineering", "hamiltonian monte carlo", "directed evolution", "ai4science" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 1, 3, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "wVTJRnZ11Z", "primary_area": "optimization", "n_author": 9, "author_info_list": [ { "name": "Akang Wang", "gender": "Not Specified", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Researcher" }, { "name": "Jianghua Wu", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Researcher" }, { "name": "Lei Li", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Researcher" }, { "name": "Qian Chen", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "PhD student" }, { "name": "Qian Li", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Researcher" }, { "name": "Qingjiang Shi", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Full Professor" }, { "name": "Ruoyu Sun", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Tsung-Hui Chang", "gender": "Not Specified", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Xiaodong Luo", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 25, "n_formula_1": 18, "n_ref_uni": 33, "n_ref": 50, "n_ref_all": 63, "n_fig": 1, "n_tab": 4, "L_tab": 517, "n_element_tab": 75, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 77, "n_element_tab_1": 17, "formula_len_all": 2487, "formula_len_all_1": 1647, "len_all": 128192, "len_all_1": 71956, "len_abs": 1212, "len_title": 125, "len_sents": 42863, "len_sents_1": 33196, "n_sents": 336, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1217, "title": "When GNNs meet symmetry in ILPs: an orbit-based feature augmentation approach", "abs": "A common characteristic in integer linear programs (ILPs) is symmetry, allowing variables to be permuted without altering the underlying problem structure. Recently, GNNs have emerged as a promising approach for solving ILPs. \nHowever, a significant challenge arises when applying GNNs to ILPs with symmetry: classic GNN architectures struggle to differentiate between symmetric variables, which limits their predictive accuracy. In this work, we investigate the properties of permutation equivalence and invariance in GNNs, particularly in relation to the inherent symmetry of ILP formulations. We reveal that the interaction between these two factors contributes to the difficulty of distinguishing between symmetric variables.\nTo address this challenge, we explore the potential of feature augmentation and propose several guiding principles for constructing augmented features. Building on these principles, we develop an orbit-based augmentation scheme that first groups symmetric variables and then samples augmented features for each group from a discrete uniform distribution. Empirical results demonstrate that our proposed approach significantly enhances both training efficiency and predictive performance.", "keywords": [ "integer linear programming", "symmetry", "machine learning", "graph neural networks" ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 4, 5, 2 ] }, { "paper_id": "wVMzK2Feuu", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "HaiBin Wen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "JingXin Han", "gender": "Male", "institution": "Shanghai University", "country": "CN", "position": "MS student" }, { "name": "Pengkun Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" }, { "name": "Qingfu Zhang", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Yang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zhe Zhao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zhenkun Wang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "\u738b\u723d", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 48, "n_formula_1": 14, "n_ref_uni": 26, "n_ref": 37, "n_ref_all": 46, "n_fig": 3, "n_tab": 1, "L_tab": 1956, "n_element_tab": 315, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 2040, "n_element_tab_1": 320, "formula_len_all": 1720, "formula_len_all_1": 588, "len_all": 142879, "len_all_1": 59126, "len_abs": 1390, "len_title": 128, "len_sents": 45984, "len_sents_1": 25909, "n_sents": 353, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1401, "title": "Balancing Model Efficiency and Performance: Adaptive Pruner for Long-tailed Data", "abs": "Long-tailed distribution datasets are prevalent in many machine learning tasks, yet existing neural network models still face significant challenges when handling such data. This paper proposes a novel adaptive pruning strategy, LTAP (Long-Tailed Adaptive Pruner), aimed at balancing model efficiency and performance to better address the challenges posed by long-tailed data distributions. LTAP introduces multi-dimensional importance scoring criteria and designs a dynamic weight adjustment mechanism to adaptively determine the pruning priority of parameters for different classes. By focusing on protecting parameters critical for tail classes, LTAP significantly enhances computational efficiency while maintaining model performance. This method combines the strengths of long-tailed learning and neural network pruning, overcoming the limitations of existing approaches in handling imbalanced data. Extensive experiments demonstrate that LTAP outperforms existing methods on various long-tailed datasets, achieving a good balance between model compression rate, computational efficiency, and classification accuracy. This research provides new insights into solving model optimization problems in long-tailed learning and is significant for improving the performance of neural networks on imbalanced datasets. The code is available at \\url{https://anonymous.4open.science/r/AEFCDAISJ/README.md}.", "keywords": [ "Long-tail learning\uff0cNeural network pruning\uff0cMulti-objective Optimization" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 2, 4, 4, 2 ] }, { "paper_id": "wVADj7yKee", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Mingquan Feng", "gender": "unknown", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Weixin Liao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yixin Huang", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Undergrad student" }, { "name": "Yizhou Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Undergrad student" }, { "name": "Yuhong Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 44, "n_formula_1": 20, "n_ref_uni": 30, "n_ref": 50, "n_ref_all": 63, "n_fig": 2, "n_tab": 9, "L_tab": 1662, "n_element_tab": 209, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1987, "n_element_tab_1": 167, "formula_len_all": 3436, "formula_len_all_1": 1462, "len_all": 130322, "len_all_1": 62259, "len_abs": 860, "len_title": 124, "len_sents": 41421, "len_sents_1": 24788, "n_sents": 363, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 861, "title": "SINGER: Stochastic Network Graph Evolving Operator for High Dimensional PDEs", "abs": "We present a novel framework, StochastIc Network Graph Evolving operatoR (SINGER), for learning the evolution operator of high-dimensional partial differential equations (PDEs). The framework uses a sub-network to approximate the solution at the initial time step and stochastically evolves the sub-network parameters over time by a graph neural network to approximate the solution at later time steps. The framework is designed to inherit the desirable properties of the parametric solution operator, including graph topology, semigroup, and stability, with a theoretical guarantee. Numerical experiments on 8 evolution PDEs of 5,10,15,20-dimensions show that our method outperforms existing baselines in almost all cases (31 out of 32), and that our method generalizes well to unseen initial conditions, equation dimensions, sub-network width, and time steps.", "keywords": [ "PDE", "High Dimension", "Neural ODE" ], "rating_list": [ 5, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "wV9iMiyQcc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Haoxian Chen", "gender": "unknown", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Limin Wang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 31, "n_ref": 47, "n_ref_all": 62, "n_fig": 5, "n_tab": 11, "L_tab": 1704, "n_element_tab": 209, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1052, "n_element_tab_1": 162, "formula_len_all": 670, "formula_len_all_1": 685, "len_all": 99328, "len_all_1": 53745, "len_abs": 1282, "len_title": 104, "len_sents": 24375, "len_sents_1": 21710, "n_sents": 229, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1287, "title": "RotPruner: Large Language Model Pruning in Rotated Space", "abs": "Network pruning is a crucial technique for compressing large language models with billions of parameters, aiming to reduce memory and computational costs with minimal performance degradation. However, existing pruning methods for LLMs often focus on heuristic metrics or layer-wise reconstruction losses, neglecting the impact on the overall model output, which can lead to suboptimal result. Additionally, these methods operate directly on the original weight and activation spaces, which may not be ideal for pruning. In this paper, we propose that the original parameter space is not optimal for pruning and present a novel training-based pruning framework called RotPruner. RotPruner rotates the spaces of weight matrices and activations in linear layers, and applies existing pruning methods in a rotated space that is more suitable for pruning. We introduce an efficient algorithm to identify an appropriate rotation that preserves the performance of pruned LLMs. RotPruner is capable of integrating with other pruning methods and supporting unstructured, semi-structured, and structured pruning. We evaluate RotPruner on several large language models, including OPT, LLaMA-2, and LLaMA-3, and demonstrate state-of-the-art performance on both language modeling and zero-shot tasks.", "keywords": [ "network pruning", "sparsity", "Large Language Model" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "wUtXB43Chi", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 10, "author_info_list": [ { "name": "Haifeng Wang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "CTO" }, { "name": "Jiang Bian", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Applied Scientist" }, { "name": "Jinle Zeng", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Lujing Zheng", "gender": "Female", "institution": "Baidu", "country": "", "position": "Researcher" }, { "name": "Siming Wu", "gender": "unknown", "institution": "Northeastern University", "country": "CN", "position": "MS student" }, { "name": "Xiyuan Xiao", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Zeyu Chen", "gender": "Male", "institution": "Baidu, Inc.", "country": "", "position": "Principal Architect" }, { "name": "dianhai yu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "GuoxiaWang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "JiabinYang", "gender": "unknown", "institution": "Baidu", "country": "CN", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 6, "n_ref_uni": 12, "n_ref": 34, "n_ref_all": 58, "n_fig": 8, "n_tab": 13, "L_tab": 14803, "n_element_tab": 1759, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1982, "n_element_tab_1": 184, "formula_len_all": 2052, "formula_len_all_1": 898, "len_all": 166110, "len_all_1": 58075, "len_abs": 5062, "len_title": 110, "len_sents": 41860, "len_sents_1": 25131, "n_sents": 303, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 2140, "title": "FlashMask: Efficient and Rich Mask Extension of FlashAttention", "abs": "The computational and memory demands of vanilla attention scale quadratically with the sequence length $N$, posing significant challenges for processing long sequences in Transformer models. FlashAttention alleviates these challenges by eliminating the $O(N^2)$ memory dependency and reducing attention latency through IO-aware memory optimizations. However, its native support for certain attention mask types is limited, and it does not inherently accommodate more complex masking requirements. Previous approaches resort to using dense masks with $O(N^2)$ memory complexity, leading to inefficiencies. In this paper, we propose FlashMask, an extension of FlashAttention that introduces a column-wise sparse representation of attention masks. This approach efficiently represents a wide range of mask types and facilitates the development of optimized kernel implementations. By adopting this novel representation, FlashMask achieves linear memory complexity $O(N)$, making it suitable for modeling long-context sequences. Moreover, this representation enables kernel optimizations that eliminate unnecessary computations by leveraging sparsity in the attention mask, without sacrificing computational accuracy, resulting in higher computational efficiency. We evaluate FlashMask's performance in fine-tuning and alignment training of LLMs such as SFT, LoRA, DPO, and RM. FlashMask achieves significant throughput improvements, with end-to-end speedups ranging from 1.65x to 3.22x compared to existing FlashAttention dense method. Additionally, our kernel-level comparisons demonstrate that FlashMask surpasses the latest counterpart, FlexAttention, by 12.1% to 60.7% in kernel TFLOPs/s, achieving 37.8% to 62.3% of the theoretical maximum FLOPs/s on the A100 GPU. Our experiments highlight FlashMask's versatility and robustness across various mask and attention patterns. These results underscore its effectiveness in practical applications, including deployment in LLMs with over 100 billion parameters, efficiently handling contexts up to 128K tokens. The implementation is open-sourced and integrated into the PaddlePaddle framework.", "keywords": [ "Attention Mask Efficient Representation", "Efficient Attention Computation", "Long context", "IO complexity", "GPUs", "LLMs" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "wUtCieKuQU", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Jianing Zhu", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "PhD student" }, { "name": "Masashi Sugiyama", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Director" }, { "name": "Qizhou Wang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "PhD student" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" }, { "name": "Yang Puning", "gender": "unknown", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Intern" } ], "n_formula": 20, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 102, "n_ref_all": 115, "n_fig": 6, "n_tab": 16, "L_tab": 18524, "n_element_tab": 2959, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2075, "n_element_tab_1": 287, "formula_len_all": 1324, "formula_len_all_1": 265, "len_all": 228497, "len_all_1": 71004, "len_abs": 1827, "len_title": 121, "len_sents": 56700, "len_sents_1": 33743, "n_sents": 433, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1838, "title": "Towards Effective Evaluations and Comparison for LLM Unlearning Methods", "abs": "The imperative to eliminate undesirable data memorization underscores the significance of machine unlearning for large language models (LLMs). Recent research has introduced a series of promising unlearning methods, notably boosting the practical significance of the field. Nevertheless, adopting a proper evaluation framework to reflect the true unlearning efficacy is also essential yet has not received adequate attention. This paper seeks to improve the evaluation of LLM unlearning by addressing two key challenges---a) the robustness of evaluation metrics and b) the trade-offs between competing goals. The first challenge stems from findings that current metrics are susceptible to various red teaming scenarios. It indicates that they may not reflect the true extent of knowledge retained by LLMs but rather tend to mirror superficial model behaviors, thus prone to attacks. We address this issue by devising and assessing a series of candidate metrics, selecting the most robust ones under various types of attacks. The second challenge arises from the conflicting goals of eliminating unwanted knowledge while retaining those of others. This trade-off between unlearning and retention often fails to conform the Pareto frontier, rendering it subtle to compare the efficacy between methods that excel only in either unlearning or retention. We handle this issue by proposing a calibration method that can restore the original performance on non-targeted data after unlearning, thereby allowing us to focus exclusively on assessing the strength of unlearning. Our evaluation framework notably enhances the effectiveness when assessing and comparing various LLM unlearning methods, further allowing us to benchmark existing works, identify their proper hyper-parameters, and explore new tricks to enhance their practical efficacy.", "keywords": [ "llm unlearning" ], "rating_list": [ 6, 5, 8, 3 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "wUbum0nd9N", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Hengguan Huang", "gender": "Male", "institution": "University of Copenhagen", "country": "DK", "position": "Assistant Professor" }, { "name": "Hongfu Liu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Xiangming Gu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Ye Wang", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Associate Professor" }, { "name": "Hao Wang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 36, "n_ref": 71, "n_ref_all": 94, "n_fig": 3, "n_tab": 5, "L_tab": 5699, "n_element_tab": 1359, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1916, "n_element_tab_1": 460, "formula_len_all": 586, "formula_len_all_1": 290, "len_all": 158699, "len_all_1": 61867, "len_abs": 2813, "len_title": 120, "len_sents": 42554, "len_sents_1": 26262, "n_sents": 291, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1567, "title": "On Calibration of LLM-based Guard Models for Reliable Content Moderation", "abs": "Large language models (LLMs) are exposed to significant risks due to their potential for malicious use. Existing studies have developed LLM-based guard models designed to moderate the input and output of threat LLMs, ensuring adherence to safety policies by blocking content that violates these protocols upon deployment. However, limited attention has been given to the reliability and calibration of such guard models. In this work, we empirically conduct comprehensive investigations of confidence calibration for 9 existing LLM-based guard models on 12 benchmarks in both user input and model output classification. Our findings reveal that current LLM-based guard models tend to 1) produce overconfident predictions, 2) exhibit significant miscalibration when subjected to jailbreak attacks, and 3) demonstrate limited robustness to the outputs generated by different types of response models. Additionally, we assess the effectiveness of post-hoc calibration methods to mitigate miscalibration. We demonstrate the efficacy of temperature scaling and, for the first time, highlight the benefits of contextual calibration for confidence calibration of guard models, particularly in the absence of validation sets. Our analysis and experiments underscore the limitations of current LLM-based guard models and provide valuable insights for the future development of well-calibrated guard models toward more reliable content moderation. We also advocate for incorporating reliability evaluation of confidence calibration when releasing future LLM-based guard models.", "keywords": [ "Content Moderation", "LLM-based Guard Models", "Calibration", "Safety" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 2, 4, 2 ] }, { "paper_id": "wUFbwlHvbk", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Dan Zhang", "gender": "Male", "institution": "Clemson University", "country": "US", "position": "PhD student" }, { "name": "Feng Luo", "gender": "Male", "institution": "Clemson University", "country": "US", "position": "Full Professor" }, { "name": "Jingjing Wang", "gender": "unknown", "institution": "Clemson University", "country": "US", "position": "PhD student" }, { "name": "Joshua Luo", "gender": "Male", "institution": "The Westminster Schools", "country": "US", "position": "high school student" } ], "n_formula": 62, "n_formula_1": 32, "n_ref_uni": 15, "n_ref": 32, "n_ref_all": 39, "n_fig": 2, "n_tab": 4, "L_tab": 1808, "n_element_tab": 152, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1883, "n_element_tab_1": 185, "formula_len_all": 4203, "formula_len_all_1": 1565, "len_all": 115135, "len_all_1": 61436, "len_abs": 1690, "len_title": 71, "len_sents": 31717, "len_sents_1": 25353, "n_sents": 272, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 23, "L_abs": 1709, "title": "Integration Flow Models", "abs": "Recently, ordinary differential equation (ODE) based generative models have emerged as a cutting-edge method for producing high-quality samples in many applications. Generally, these methods typically involve learning continuous transformation trajectories that map a simple initial distribution (i.e., Gaussian noise) to the target data distribution (i.e., images) by multiple steps of solving different ODE functions in inference to obtain high-quality results. However, the ODE-based methods either suffer the discretization error of numerical solvers of ODE, which restricts the quality of samples when only a few NFEs are used, or struggle with training instability. In this paper, we proposed Integration Flow, which learns the results of ODE-based trajectory paths directly without solving the ODE functions. Moreover, Integration Flow explicitly incorporates the target state $\\mathbf{x}_0$ as the anchor state in guiding the reverse-time dynamics and we have theoretically proven this can contribute to both stability and accuracy. To the best of our knowledge, Integration Flow is the first model with the unified structure to estimate ODE-based generative models. Through theoretical analysis and empirical evaluations, we show that Integration Flows achieve improved performance when it is applied to existing ODE-based model, such as diffusion models, Rectified Flows, and PFGM++. Specifically, Integration Flow achieves one-step generation on CIFAR10 with FID of 2.63 for Variance Exploding (VE) diffusion model, 3.4 for Rectified Flow without relflow and 2.96 for PFGM++. By extending the sampling to 1000 steps, we further reduce FID score to 1.71 for VE, setting state-of-the-art performance.", "keywords": [ "integration flow", "ode-based generative models", "diffusion models" ], "rating_list": [ 5, 5, 1, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "wTm4W39GdD", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Bo Zhao", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Eric J Bigelow", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Hidenori Tanaka", "gender": "unknown", "institution": "Harvard University, Harvard University", "country": "US", "position": "Researcher" }, { "name": "Maya Okawa", "gender": "unknown", "institution": "NTT Research, Inc.", "country": "", "position": "Researcher" }, { "name": "Rose Yu", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Tomer Ullman", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 64, "n_ref": 81, "n_ref_all": 143, "n_fig": 27, "n_tab": 9, "L_tab": 9162, "n_element_tab": 655, "n_fig_1": 16, "n_tab_1": 2, "L_tab_1": 1931, "n_element_tab_1": 131, "formula_len_all": 363, "formula_len_all_1": 174, "len_all": 216564, "len_all_1": 69661, "len_abs": 1301, "len_title": 122, "len_sents": 62590, "len_sents_1": 29762, "n_sents": 520, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1353, "title": "Emergence of Hierarchical Emotion Representations in Large Language Models", "abs": "As large language models (LLMs) increasingly power emotionally engaging conversational agents, understanding how they represent, predict, and potentially influence human emotions is critical for their ethical deployment in sensitive contexts. In this work, we reveal emergent hierarchical structures in LLMs' emotion representations, drawing inspiration from psychological theories of emotion. By analyzing probabilistic dependencies between emotional states in LLM outputs, we propose a method for extracting these hierarchies. Our results show that larger models, such as LLaMA 3.1 (405B parameters), develop more intricate emotion hierarchies, resembling human emotional differentiation from broad categories to finer states. Moreover, we find that stronger emotional modeling enhances persuasive abilities in synthetic negotiation tasks, with LLMs that more accurately predict counterparts' emotions achieving better outcomes. Additionally, we explore the effects of persona biases\u2014such as gender and socioeconomic status\u2014on emotion recognition, revealing that LLMs can misclassify emotions when processing minority personas, thus exposing underlying biases. This study contributes to both the scientific understanding of how LLMs represent emotions and the ethical challenges they pose, proposing a novel interdisciplinary perspective on the issue.", "keywords": [ "LLM", "emotion" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 3, 4, 2, 3 ] }, { "paper_id": "wTLc79YNbh", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Can Li", "gender": "Female", "institution": "Tongji University", "country": "CN", "position": "Researcher" }, { "name": "LEI BAI", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Songtao Huang", "gender": "unknown", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "PhD student" }, { "name": "Zhen Zhao", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 20, "n_formula_1": 11, "n_ref_uni": 18, "n_ref": 49, "n_ref_all": 61, "n_fig": 2, "n_tab": 10, "L_tab": 8594, "n_element_tab": 276, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2297, "n_element_tab_1": 81, "formula_len_all": 940, "formula_len_all_1": 439, "len_all": 126420, "len_all_1": 58782, "len_abs": 1454, "len_title": 151, "len_sents": 40284, "len_sents_1": 28395, "n_sents": 273, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 1519, "title": "TimeKAN: KAN-based Frequency Decomposition Learning Architecture for Long-term Time Series Forecasting", "abs": "Real-world time series often have multiple frequency components that are intertwined with each other, making accurate time series forecasting challenging. Decomposing the mixed frequency components into multiple single frequency components is a natural choice. However, the information density of patterns varies across different frequencies, and employing a uniform modeling approach for different frequency components can lead to inaccurate characterization. To address this challenges, inspired by the flexibility of the recent Kolmogorov-Arnold Network (KAN), we propose a KAN-based Frequency Decomposition Learning architecture (TimeKAN) to address the complex forecasting challenges caused by multiple frequency mixtures. Specifically, TimeKAN mainly consists of three components: Cascaded Frequency Decomposition (CFD) blocks, Multi-order KAN Representation Learning (M-KAN) blocks and Frequency Mixing blocks. CFD blocks adopt a bottom-up cascading approach to obtain series representations for each frequency band. Benefiting from the high flexibility of KAN, we design a novel M-KAN block to learn and represent specific temporal patterns within each frequency band. Finally, Frequency Mixing blocks is used to recombine the frequency bands into the original format. Extensive experimental results across multiple real-world time series datasets demonstrate that TimeKAN achieves state-of-the-art performance as an extremely lightweight architecture. Code is available at https://github.com/huangst21/TimeKAN.", "keywords": [ "Kolmogorov-Arnold Network; Time Series Forecasting" ], "rating_list": [ 8, 8, 5, 3 ], "soundness_list": [ 3, 4, 3, 1 ], "presentation_list": [ 4, 4, 3, 2 ], "contribution_list": [ 4, 3, 2, 1 ], "confidence_list": [ 5, 2, 5, 5 ] }, { "paper_id": "wT1aFmsXOc", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Huiyuan Chen", "gender": "Male", "institution": "VISA", "country": "", "position": "Researcher" }, { "name": "Jing Li", "gender": "Male", "institution": "Case Western Reserve University", "country": "", "position": "Full Professor" }, { "name": "Xiao Li", "gender": "Female", "institution": "Case Western Reserve University", "country": "US", "position": "Assistant Professor" }, { "name": "Zhengyu Fang", "gender": "Male", "institution": "Case Western Reserve University", "country": "US", "position": "PhD student" }, { "name": "Zhimeng Jiang", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 40, "n_formula_1": 12, "n_ref_uni": 39, "n_ref": 79, "n_ref_all": 100, "n_fig": 5, "n_tab": 6, "L_tab": 9746, "n_element_tab": 742, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3243, "formula_len_all_1": 835, "len_all": 203019, "len_all_1": 55932, "len_abs": 1316, "len_title": 126, "len_sents": 65458, "len_sents_1": 25747, "n_sents": 517, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1328, "title": "Understanding and Mitigating Memorization in Diffusion Models for Tabular Data", "abs": "Tabular data generation has attracted significant research interest in recent years, with the tabular diffusion models greatly improving the quality of synthetic data. However, while memorization\u2014where models inadvertently replicate exact or near-identical training data\u2014has been thoroughly investigated in image and text generation, its effects on tabular data remain largely unexplored. In this paper, we conduct the first comprehensive investigation of memorization phenomena in diffusion models for tabular data. Our empirical analysis reveals that memorization appears in tabular diffusion models and increases with larger training epochs. We further examine the influence of factors such as dataset sizes, feature dimensions, and different diffusion models on memorization. Additionally, we provide a theoretical explanation for why memorization occurs in tabular diffusion models. To address this issue, we propose TabCutMix, a simple yet effective data augmentation technique that exchanges randomly selected feature segments between random training sample pairs. Experimental results across various datasets and diffusion models demonstrate that TabCutMix effectively mitigates memorization while maintaining high-quality data generation. Our code is available at \\url{https://anonymous.4open.science/r/TabCutMix-3F7B}.", "keywords": [ "Memorization", "Tabular Data", "Diffusion Models" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "wSkvf2WyYz", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Ankan Biswas", "gender": "unknown", "institution": "Fractal", "country": "IN", "position": "Researcher" }, { "name": "Kunal Singh", "gender": "Male", "institution": "Fractal Analytics", "country": "IN", "position": "Researcher" }, { "name": "Pradeep Moturi", "gender": "Male", "institution": "Fractal AI", "country": "IN", "position": "Researcher" }, { "name": "Sayandeep Bhowmick", "gender": "Male", "institution": "Indian Institute of Technology Kharagpur", "country": "IN", "position": "Undergrad student" }, { "name": "Siva Kishore Gollapalli", "gender": "Male", "institution": "Indian Institute of Technology, Bombay", "country": "IN", "position": "Undergrad student" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 38, "n_ref": 112, "n_ref_all": 156, "n_fig": 6, "n_tab": 18, "L_tab": 5006, "n_element_tab": 527, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 596, "n_element_tab_1": 103, "formula_len_all": 652, "formula_len_all_1": 103, "len_all": 290873, "len_all_1": 56302, "len_abs": 1153, "len_title": 122, "len_sents": 94024, "len_sents_1": 26094, "n_sents": 731, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1126, "title": "SBSC: Step-by-Step Coding for Improving Mathematical Olympiad Performance", "abs": "We propose Step-by-Step Coding (SBSC): a multi-turn math reasoning framework that enables Large Language Models (LLMs) to generate sequence of programs for solving Olympiad level math problems. After each turn/step, by leveraging the code execution outputs and programs of previous steps, the model generates the next sub-task and the corresponding program to complete it. This way, SBSC, sequentially navigates to reach the final answer. SBSC allows more granular, flexible and precise approach to problem-solving compared to existing methods. Extensive experiments highlight the effectiveness of SBSC in tackling competition and Olympiad-level math problems. For Claude-3.5-Sonnet, we observe SBSC (greedy decoding) surpasses existing state-of-the-art (SOTA) program generation based reasoning strategies by absolute 10.7% on AMC12, 8% on AIME and 12.6% on MathOdyssey. Given SBSC is multi-turn in nature, we also benchmark SBSC\u2019s greedy decoding against self- consistency decoding results of existing SOTA math reasoning strategies and observe performance gain by absolute 6.2% on AMC, 6.7% on AIME and 7.4% on MathOdyssey.", "keywords": [ "math AI", "LLM math reasoning" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 4, 3, 2, 4 ], "contribution_list": [ 4, 2, 3, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "wSErgkwDZO", "primary_area": "datasets and benchmarks", "n_author": 21, "author_info_list": [ { "name": "Bingli Wang", "gender": "Male", "institution": "Sichuan Agricultural University", "country": "CN", "position": "Undergrad student" }, { "name": "Chenghua Lin", "gender": "Male", "institution": "University of Manchester ", "country": "GB", "position": "Full Professor" }, { "name": "Chenhao Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Feiteng Fang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Ge Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Guangzeng Han", "gender": "Male", "institution": "University of Memphis", "country": "US", "position": "PhD student" }, { "name": "Liu Jiaheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jinchang Hou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Kaixin Deng", "gender": "Male", "institution": "Chengdu University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Min Yang", "gender": "Female", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Qinrui Li", "gender": "Female", "institution": "University of California, Santa Barbara", "country": "US", "position": "Undergrad student" }, { "name": "Qixuan Zhao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Shiwen Ni", "gender": "unknown", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Wenhao Huang", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Xi Feng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Xingwei Qu", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "PhD student" }, { "name": "Yifei Zhang", "gender": "unknown", "institution": "Southwest University", "country": "CN", "position": "MS student" }, { "name": "Yiming Liang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Ziqiang Liu", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "yuelin bai", "gender": "Not Specified", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Xinrun Du", "gender": "Male", "institution": "01.AI", "country": "CN", "position": "Intern" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 54, "n_ref_all": 68, "n_fig": 5, "n_tab": 9, "L_tab": 3983, "n_element_tab": 783, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3555, "n_element_tab_1": 463, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 150770, "len_all_1": 65444, "len_abs": 1843, "len_title": 112, "len_sents": 42483, "len_sents_1": 27301, "n_sents": 347, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1842, "title": "Can MLLMs Understand the Deep Implication Behind Chinese Images?", "abs": "As the capabilities of Multimodal Large Language Models (MLLMs) continue to improve, the need for higher-order capability evaluation of MLLMs is increasing. However, there is a lack of work evaluating MLLM for higher-order perception and understanding of Chinese visual content.\nTo fill the gap, we introduce the **C**hinese **I**mage **I**mplication understanding **Bench**mark, **CII-Bench**, which aims to assess the higher-order perception and understanding capabilities of MLLMs for Chinese images. \nCII-Bench stands out in several ways compared to existing benchmarks. Firstly, to ensure the authenticity of the Chinese context, images in CII-Bench are sourced from the Chinese Internet and manually reviewed, with corresponding answers also manually crafted. Additionally, CII-Bench incorporates images that represent Chinese traditional culture, such as famous Chinese traditional paintings, which can deeply reflect the model's understanding of Chinese traditional culture.\nThrough extensive experiments on CII-Bench across multiple MLLMs, we have made significant findings. \nInitially, a substantial gap is observed between the performance of MLLMs and humans on CII-Bench. The highest accuracy of MLLMs attains 64.4\\%, where as human accuracy averages 78.2\\%, peaking at an impressive 81.0\\%. Subsequently, MLLMs perform worse on Chinese traditional culture images, suggesting limitations in their ability to understand high-level semantics and lack a deep knowledge base of Chinese traditional culture. Finally, it is observed that most models exhibit enhanced accuracy when image emotion hints are incorporated into the prompts.\nWe believe that CII-Bench will enable MLLMs to gain a better understanding of Chinese semantics and Chinese-specific images, advancing the journey towards expert artificial general intelligence (AGI).", "keywords": [ "Multimodel Large Language Models", "Language and Vision" ], "rating_list": [ 5, 3, 3, 3, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 5, 5, 5, 4, 2 ] }, { "paper_id": "wRbSdbGyfj", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Danyang Huang", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Debarghya Mukherjee", "gender": "Male", "institution": "Boston University, Boston University", "country": "", "position": "Assistant Professor" }, { "name": "Huimin Cheng", "gender": "Female", "institution": "Boston University, Boston University", "country": "", "position": "Assistant Professor" }, { "name": "Jiachen Chen", "gender": "Female", "institution": "Boston University, Boston University", "country": "US", "position": "PhD student" }, { "name": "Kathryn L Lunetta", "gender": "Female", "institution": "Boston University, Boston University", "country": "US", "position": "Full Professor" }, { "name": "Liyuan Wang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 209, "n_formula_1": 17, "n_ref_uni": 56, "n_ref": 86, "n_ref_all": 124, "n_fig": 9, "n_tab": 8, "L_tab": 3006, "n_element_tab": 395, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 24302, "formula_len_all_1": 1711, "len_all": 268053, "len_all_1": 77709, "len_abs": 1134, "len_title": 149, "len_sents": 69090, "len_sents_1": 34413, "n_sents": 769, "n_sents_1": 286, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 1140, "title": "Transfer Learning Under High-Dimensional Graph Convolutional Regression Model for Node Classification", "abs": "Node classification is a fundamental task, but obtaining node classification labels can be challenging and expensive in many real-world scenarios. Transfer learning has emerged as a promising solution to address this challenge by leveraging knowledge from source domains to enhance learning in a target domain. Existing transfer learning methods for node classification primarily focus on integrating Graph Convolutional Networks (GCNs) with various transfer learning techniques. While these approaches have shown promising results, they often suffer from a lack of theoretical guarantees, restrictive conditions, and high sensitivity to hyperparameter choices. To overcome these limitations, we employ a Graph Convolutional Multinomial Logistic Lasso Regression (GCR) model which simplifies GCN, and develop a transfer learning method called Trans-GCR based on the GCR model. We provide theoretical guarantees of the estimate obtained under the GCR model in high-dimensional settings. Moreover, Trans-GCR demonstrates superior empirical performance, has a low computational cost, and requires fewer hyperparameters than existing methods.", "keywords": [ "Transfer learning", "Node Classification", "Graph Convolution", "High-Dimensional" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 5, 5, 3 ] }, { "paper_id": "wQkERVYqui", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Nico Scherf", "gender": "unknown", "institution": "Max Planck Institute for Human Cognitive and Brain Sciences", "country": "DE", "position": "group leader" }, { "name": "Nikola Milosevic", "gender": "Male", "institution": "Max Planck Institute for Human cognition and brain sciences, Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Johannes Christoph M\u00fcller", "gender": "Male", "institution": "RWTH Aachen University", "country": "DE", "position": "Postdoc" } ], "n_formula": 141, "n_formula_1": 48, "n_ref_uni": 32, "n_ref": 86, "n_ref_all": 120, "n_fig": 9, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9732, "formula_len_all_1": 2967, "len_all": 182666, "len_all_1": 67895, "len_abs": 1379, "len_title": 108, "len_sents": 53601, "len_sents_1": 28387, "n_sents": 447, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1024, "title": "Embedding Safety into RL: A New Take on Trust Region Methods", "abs": "Reinforcement Learning (RL) agents are able to solve a wide variety of tasks but are prone to producing unsafe behaviors.\nConstrained Markov Decision Processes (CMDPs) provide a popular framework for incorporating safety constraints. \nHowever, common solution methods often compromise reward maximization by being overly conservative or allow unsafe behavior during training.\nWe propose Constrained Trust Region Policy Optimization (C-TRPO), a novel approach that modifies the geometry of the policy space based on the safety constraints and yields trust regions composed exclusively of safe policies, ensuring constraint satisfaction throughout training.\nWe theoretically study the convergence and update properties of C-TRPO and highlight connections to TRPO, Natural Policy Gradient (NPG), and Constrained Policy Optimization (CPO).\nFinally, we demonstrate experimentally that C-TRPO significantly reduces constraint violations while achieving competitive reward maximization compared to state-of-the-art CMDP algorithms.", "keywords": [ "reinforcement learning", "safety", "information geometry" ], "rating_list": [ 8, 5, 3, 5, 6 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 3, 2, 4, 2, 3 ] }, { "paper_id": "wQk6yaRGOi", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Alan Nawzad Amin", "gender": "Male", "institution": "New York University", "country": "", "position": "Postdoc" }, { "name": "Andrew Gordon Wilson", "gender": "Not Specified", "institution": "New York University", "country": "US", "position": "Full Professor" }, { "name": "Nate Gruver", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" } ], "n_formula": 58, "n_formula_1": 19, "n_ref_uni": 18, "n_ref": 78, "n_ref_all": 98, "n_fig": 5, "n_tab": 10, "L_tab": 2326, "n_element_tab": 175, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 151, "n_element_tab_1": 4, "formula_len_all": 5897, "formula_len_all_1": 1725, "len_all": 169400, "len_all_1": 72384, "len_abs": 1237, "len_title": 103, "len_sents": 57466, "len_sents_1": 31980, "n_sents": 505, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1244, "title": "Improving Discrete Diffusion with Schedule-Conditioning", "abs": "Discrete diffusion models, like continuous diffusion models, generate high-quality sequence data by gradually undoing noise applied to datapoints via a Markov process. Gradual generation in theory comes with many conceptual benefits; for example, inductive biases can be incorporated into the noising Markov process. In practice however, the best performing discrete diffusion model is consistently masking, which does not denoise gradually. Here we explain the performance of masking diffusion by noting that it makes use of a fundamental difference between continuous and discrete Markov processes: discrete Markov processes evolve by discontinuous jumps at a fixed rate and, unlike other discrete diffusion models, masking diffusion builds in the known distribution of jump times and only learns where to jump to. We show that we can similarly bake in the known distribution of jump times into any discrete diffusion model; despite their simplicity, our new models -- schedule-conditioned diffusion (SCUD) -- generalize classical discrete diffusion and masking diffusion. By applying SCUD to models with noising processes that incorporate inductive biases on images, text, and protein data, we build diffusion models that outperform masking.", "keywords": [ "discrete diffusion", "image generation", "language model" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 1, 2, 4 ] }, { "paper_id": "wQHyjIZ1SH", "primary_area": "generative models", "n_author": 1, "author_info_list": [ { "name": "Jo\u00e3o Bravo", "gender": "Male", "institution": "Feedzai", "country": "PT", "position": "Researcher" } ], "n_formula": 69, "n_formula_1": 19, "n_ref_uni": 35, "n_ref": 94, "n_ref_all": 128, "n_fig": 6, "n_tab": 10, "L_tab": 2206, "n_element_tab": 211, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4342, "formula_len_all_1": 877, "len_all": 201684, "len_all_1": 59913, "len_abs": 1374, "len_title": 95, "len_sents": 74420, "len_sents_1": 30479, "n_sents": 511, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1001, "title": "NRGBoost: Energy-Based Generative Boosted Trees", "abs": "Despite the rise to dominance of deep learning in unstructured data domains, tree-based methods such as Random Forests (RF) and Gradient Boosted Decision Trees (GBDT) are still the workhorses for handling discriminative tasks on tabular data. We explore generative extensions of these popular algorithms with a focus on explicitly modeling the data density (up to a normalization constant), thus enabling other applications besides sampling. \nAs our main contribution we propose an energy-based generative boosting algorithm that is analogous to the second order boosting implemented in popular packages like XGBoost. We show that, despite producing a generative model capable of handling inference tasks over any input variable, our proposed algorithm can achieve similar discriminative performance to GBDT on a number of real world tabular datasets, outperforming alternative generative approaches. At the same time, we show that it is also competitive with neural network based models for sampling.", "keywords": [ "Energy-Based Models", "Generative Models", "Gradient Boosting", "Tabular Data" ], "rating_list": [ 10, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "wQEdh2cgEk", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Wendi Li", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Yixuan Li", "gender": "Female", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 29, "n_formula_1": 24, "n_ref_uni": 26, "n_ref": 105, "n_ref_all": 130, "n_fig": 6, "n_tab": 6, "L_tab": 2547, "n_element_tab": 162, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1892, "n_element_tab_1": 28, "formula_len_all": 2727, "formula_len_all_1": 2029, "len_all": 183539, "len_all_1": 75196, "len_abs": 1283, "len_title": 90, "len_sents": 56164, "len_sents_1": 31921, "n_sents": 527, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 1222, "title": "Process Reward Model with Q-value Rankings", "abs": "Process Reward Modeling (PRM) is critical for complex reasoning and decision-making tasks where the accuracy of intermediate steps significantly influences the overall outcome. Existing PRM approaches, primarily framed as classification problems, employ cross-entropy loss to independently evaluate each step's correctness. This method can lead to suboptimal reward distribution and does not adequately address the interdependencies among steps. To address these limitations, we introduce the Process Q-value Model (PQM), a novel framework that redefines PRM in the context of a Markov Decision Process. PQM optimizes Q-value rankings based on a novel comparative loss function, enhancing the model's ability to capture the intricate dynamics among sequential decisions. This approach provides a more granular and theoretically grounded methodology for process rewards. Our extensive empirical evaluations across various sampling policies, language model backbones, and multi-step reasoning benchmarks show that PQM outperforms classification-based PRMs. The effectiveness of the comparative loss function is highlighted in our comprehensive ablation studies, confirming PQM\u2019s practical efficacy and theoretical advantage.", "keywords": [ "process reward model", "reasoning" ], "rating_list": [ 3, 8, 8, 8, 5 ], "soundness_list": [ 3, 4, 3, 3, 2 ], "presentation_list": [ 2, 4, 3, 3, 2 ], "contribution_list": [ 2, 3, 4, 3, 3 ], "confidence_list": [ 4, 3, 3, 2, 4 ] }, { "paper_id": "wPyTeUMRgh", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Chengyang GU", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "", "position": "PhD student" }, { "name": "Haotian Bai", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Hui Xiong", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Yize Chen", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Assistant Professor" }, { "name": "Yuxin Pan", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 22, "n_ref_uni": 55, "n_ref": 92, "n_ref_all": 105, "n_fig": 4, "n_tab": 6, "L_tab": 1083, "n_element_tab": 113, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 704, "n_element_tab_1": 64, "formula_len_all": 1300, "formula_len_all_1": 1216, "len_all": 160486, "len_all_1": 68777, "len_abs": 1056, "len_title": 111, "len_sents": 43865, "len_sents_1": 32134, "n_sents": 346, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1060, "title": "SEAL: SEmantic-Augmented Imitation Learning via Language Model", "abs": "Hierarchical Imitation Learning (HIL) is a promising approach for tackling long-horizon decision-making tasks. While it is a challenging task due to the lack of detailed supervisory labels for sub-goal learning, and reliance on hundreds to thousands of expert demonstrations. In this work, we introduce SEAL, a novel framework that leverages Large Language Models (LLMs)'s powerful semantic and world knowledge for both specifying sub-goal space and pre-labeling states to semantically meaningful sub-goal representations without prior knowledge of task hierarchies. SEAL employs a dual-encoder structure, combining supervised LLM-guided sub-goal learning with unsupervised Vector Quantization (VQ) for more robust sub-goal representations. Additionally, SEAL incorporates a transition-augmented low-level planner for improved adaptation to sub-goal transitions. Our experiments demonstrate that SEAL outperforms state-of-the-art HIL methods and LLM-based planning approaches, particularly in settings with small expert datasets and complex long-horizon tasks.", "keywords": [ "Large Language Models", "Hierarchical Imitation Learning" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 4, 2, 4, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "wPStvOAtjR", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Fan Li", "gender": "Female", "institution": "University of Exeter", "country": "GB", "position": "Postdoc" }, { "name": "Ke Li", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "Associate Professor" }, { "name": "Shengbo Wang", "gender": "unknown", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 42, "n_formula_1": 18, "n_ref_uni": 33, "n_ref": 90, "n_ref_all": 166, "n_fig": 33, "n_tab": 9, "L_tab": 2806, "n_element_tab": 227, "n_fig_1": 19, "n_tab_1": 3, "L_tab_1": 702, "n_element_tab_1": 91, "formula_len_all": 3331, "formula_len_all_1": 1466, "len_all": 204798, "len_all_1": 70202, "len_abs": 1161, "len_title": 124, "len_sents": 57963, "len_sents_1": 31208, "n_sents": 571, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1191, "title": "LAMDA: Two-Phase Multi-Fidelity HPO via Learning Promising Regions from Data", "abs": "Multi-fidelity hyperparameter optimization (HPO) combines data from both high-fidelity (HF) and low-fidelity (LF) problems during the optimization process, aiding in effective sampling and preliminary screening. To enhance its performance, approaches that incorporate expert knowledge or transfer ability into the HPO algorithm have demonstrated their superiority, while such domain knowledge or abundant data from multiple similar tasks may not always be accessible. Observing that high-quality solutions in HPO exhibit some overlap between high- and low-fidelity problems, we propose a two-phase framework $\\texttt{Lamda}$ to streamline the multi-fidelity HPO. Specifically, in the first phase, it searches in the LF landscape to identify the promising regions of LF problem. In the second phase, we leverage such promising regions to construct reliable priors to navigate the HPO. We showcase how the $\\texttt{Lamda}$ framework can be integrated with various HPO algorithms to boost their performance, and further conduct theoretical analysis towards the integrated Bayesian optimization and bandit-based Hyperband. We demonstrate the effectiveness of our framework across $56$ HPO tasks.", "keywords": [ "HPO", "multi-fidelity", "overlapping", "promising regions" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "wPMRwmytZe", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Abhishek Panigrahi", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Andrej Risteski", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Bingbin Liu", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Sadhika Malladi", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Surbhi Goel", "gender": "Female", "institution": "University of Pennsylvania", "country": "US", "position": "Assistant Professor" } ], "n_formula": 96, "n_formula_1": 11, "n_ref_uni": 66, "n_ref": 138, "n_ref_all": 219, "n_fig": 31, "n_tab": 7, "L_tab": 1393, "n_element_tab": 41, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 437, "n_element_tab_1": 10, "formula_len_all": 9349, "formula_len_all_1": 721, "len_all": 356653, "len_all_1": 73529, "len_abs": 1326, "len_title": 103, "len_sents": 113878, "len_sents_1": 34661, "n_sents": 951, "n_sents_1": 274, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 55, "L_abs": 1332, "title": "Progressive distillation induces an implicit curriculum", "abs": "Knowledge distillation leverages a teacher model to improve the training of a student model. A persistent challenge is that a better teacher does not always yield a better student, to which a common mitigation is to use additional supervision from several \u201cintermediate\u201d teachers. One empirically validated variant of this principle is progressive distillation, where the student learns from successive intermediate checkpoints of the teacher. Using sparse parity as a sandbox, we identify an implicit curriculum as one mechanism through which progressive distillation accelerates the student\u2019s learning. This curriculum is available only through the intermediate checkpoints but not the final converged one, and imparts both empirical acceleration and a provable sample complexity benefit to the student. We then extend our investigation to Transformers trained on probabilistic context-free grammars (PCFGs) and real-world pre-training datasets (Wikipedia and Books). Through probing the teacher model, we identify an analogous implicit curriculum where the model progressively learns features that capture longer context. Our theoretical and empirical findings on sparse parity, complemented by empirical observations on more complex tasks, highlight the benefit of progressive distillation via implicit curriculum across setups.", "keywords": [ "knowledge distillation", "feature learning", "curriculum", "sparse parity", "PCFG", "optimization", "MLP", "Transformer" ], "rating_list": [ 8, 8, 8, 6, 8 ], "soundness_list": [ 4, 3, 4, 4, 3 ], "presentation_list": [ 4, 4, 4, 3, 3 ], "contribution_list": [ 3, 4, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "wP0nDEAlap", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Ke Li", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Senior Researcher" }, { "name": "Liujuan Cao", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Shengchuan Zhang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Sicheng Zhao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Xudong Li", "gender": "Male", "institution": "XMU", "country": "CN", "position": "MS student" }, { "name": "Yan Zhang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 44, "n_ref": 87, "n_ref_all": 121, "n_fig": 8, "n_tab": 14, "L_tab": 5504, "n_element_tab": 247, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 4156, "n_element_tab_1": 162, "formula_len_all": 504, "formula_len_all_1": 414, "len_all": 177467, "len_all_1": 66990, "len_abs": 1329, "len_title": 134, "len_sents": 51262, "len_sents_1": 29969, "n_sents": 389, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1341, "title": "Less is More: Learning Reference Knowledge Using No-Reference Image Quality Assessment", "abs": "Image Quality Assessment (IQA) with reference images has achieved great success by imitating the human vision system, in which the image quality is effectively assessed by comparing the query image with its pristine reference image. However, for the images in the wild, it is quite difficult to access accurate reference images. We argue that it is possible to learn reference knowledge under the \\emph{No-Reference Image Quality Assessment} (NR-IQA) setting, which is effective and efficient empirically. Concretely, by innovatively introducing a novel feature distillation method in IQA, we propose a new framework to learn comparative knowledge from non-aligned reference images. Then, we further propose inductive bias regularization to inject different inductive biases into the model to achieve fast convergence and avoid overfitting. Such a framework not only solves the congenital defects of NR-IQA but also improves the feature extraction framework, enabling it to express more abundant quality information. Surprisingly, our method utilizes less input\u2014eliminating the need for reference images during inference\u2014while obtaining more performance compared to some IQA methods that do require reference images. Comprehensive experiments on eight standard IQA datasets show that our approach outperforms state-of-the-art NR-IQA methods.", "keywords": [ "Image Quality Assessment", "Inductive Bias Regularization", "Reference Knowledge" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "wO8WbhsjNG", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Heng Huang", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Peiran Yu", "gender": "Female", "institution": "University of Texas at Arlington", "country": "US", "position": "Assistant Professor" }, { "name": "Qi He", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Reza Shirkavand", "gender": "Not Specified", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 14, "n_ref_uni": 35, "n_ref": 100, "n_ref_all": 140, "n_fig": 6, "n_tab": 7, "L_tab": 3407, "n_element_tab": 255, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1786, "n_element_tab_1": 169, "formula_len_all": 5203, "formula_len_all_1": 1243, "len_all": 192402, "len_all_1": 68427, "len_abs": 1597, "len_title": 166, "len_sents": 41521, "len_sents_1": 28628, "n_sents": 373, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 118, "L_abs": 1606, "title": "Bilevel ZOFO: Bridging Parameter-Efficient and Zeroth-Order Techniques for Efficient LLM Fine-Tuning and Meta-Training", "abs": "Fine-tuning pre-trained Large Language Models (LLMs) for downstream tasks using First-Order (FO) optimizers presents significant computational challenges. Parameter-Efficient Fine-Tuning (PEFT) methods have been proposed to address these challenges by freezing most model parameters and training only a small subset. While PEFT is efficient, it may not outperform full fine-tuning when high task-specific performance is required.\nZeroth-Order (ZO) methods offer an alternative for fine-tuning the entire pre-trained model by approximating gradients using only the forward pass, thus eliminating the computational burden of back-propagation in first-order methods. However, when implementing ZO methods, it is crucial to ensure prompt-based text alignment, and relying on simple, fixed hard prompts may not be optimal. In this paper, we propose a bilevel optimization framework that complements ZO methods with PEFT to mitigate sensitivity to hard prompts while efficiently and effectively fine-tuning LLMs. Our Bilevel ZOFO (Zeroth-Order-First-Order) method employs a double-loop optimization strategy, where only the gradient of the PEFT model and the forward pass of the base model are required. We provide convergence guarantees for Bilevel ZOFO. Empirically, we demonstrate that Bilevel ZOFO outperforms both PEFT and ZO methods in single-task settings. Additionally, we show its strong potential for multitask learning. Compared to current first-order meta-training algorithms for multitask learning, our method has significantly lower computational demands while maintaining or improving performance.", "keywords": [ "Zeroth-order Fine-tuning", "Parameter Efficient Fine-tuning", "Large Language Models", "Bilevel Optimization" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "wO1NJLitPL", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Amaury Habrard", "gender": "Male", "institution": "Universit\u00e9 Saint-Etienne, Laboratoire Hubert Curien", "country": "FR", "position": "Full Professor" }, { "name": "Marc Sebban", "gender": "Male", "institution": "University Jean Monnet", "country": "FR", "position": "Full Professor" }, { "name": "Massimiliano Pontil", "gender": "Male", "institution": "Universit\u00e0 degli Studi di Genova, Istituto Italiano di Tecnologia", "country": "IT", "position": "Principal Researcher" }, { "name": "R\u00e9mi Emonet", "gender": "Male", "institution": "University Jean Monnet", "country": "FR", "position": "Associate Professor" }, { "name": "Saverio Salzo", "gender": "unknown", "institution": "University of Roma \"La Sapienza\"", "country": "IT", "position": "Assistant Professor" }, { "name": "Abdel-Rahim MEZIDI", "gender": "unknown", "institution": "Universit\u00e9 Jean Monnet", "country": "FR", "position": "PhD student" }, { "name": "Jordan Fr\u00e9con", "gender": "Male", "institution": "Universit\u00e9 Jean Monnet, Inria", "country": "", "position": "Associate Professor" } ], "n_formula": 129, "n_formula_1": 40, "n_ref_uni": 26, "n_ref": 70, "n_ref_all": 105, "n_fig": 12, "n_tab": 5, "L_tab": 163, "n_element_tab": 23, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12679, "formula_len_all_1": 4602, "len_all": 188911, "len_all_1": 66802, "len_abs": 699, "len_title": 96, "len_sents": 57905, "len_sents_1": 27444, "n_sents": 553, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 48, "L_abs": 690, "title": "A Bregman Proximal Viewpoint on Neural Operators", "abs": "We present several advances on neural operators by viewing the action of operator layers as the minimizers of Bregman regularized optimization problems over Banach function spaces. The proposed framework allows interpreting the activation operators as Bregman proximity operators from dual to primal space. This novel viewpoint is general enough to recover classical neural operators as well as a new variant, coined Bregman neural operators, which includes the inverse activatio and features the same expressivity of standard neural operators. Numerical experiments support the added benefits of the Bregman variant of Fourier neural operators for training deeper and more accurate models.", "keywords": [ "neural operators", "proximal optimization", "bregman divergence", "fourier neural operator" ], "rating_list": [ 3, 5, 5, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 2, 1, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "wNobG8bV5Q", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Ali Pesaranghader", "gender": "Male", "institution": "LG Electronics", "country": "CA", "position": "Researcher" }, { "name": "Armin Toroghi", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Scott Sanner", "gender": "Male", "institution": "Department of Mechanical and Industrial Engineering, University of Toronto", "country": "CA", "position": "Associate Professor" }, { "name": "Tanmana Sadhu", "gender": "Female", "institution": "LG Corporation", "country": "CA", "position": "Researcher" } ], "n_formula": 26, "n_formula_1": 13, "n_ref_uni": 48, "n_ref": 89, "n_ref_all": 131, "n_fig": 3, "n_tab": 5, "L_tab": 2296, "n_element_tab": 273, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 1258, "n_element_tab_1": 231, "formula_len_all": 1922, "formula_len_all_1": 781, "len_all": 219075, "len_all_1": 71364, "len_abs": 1907, "len_title": 127, "len_sents": 66818, "len_sents_1": 32505, "n_sents": 482, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1947, "title": "LLM-based Typed Hyperresolution for Commonsense Reasoning with Knowledge Bases", "abs": "Large language models (LLM) are being increasingly applied to tasks requiring commonsense reasoning. Despite their outstanding potential, the reasoning process of LLMs is prone to errors and hallucinations that hinder their applicability, especially in high-stakes scenarios. Several works have attempted to enhance commonsense reasoning performance of LLMs by (i) using prompting styles that elicit more accurate reasoning, (ii) utilizing the LLM as a semantic parser for a symbolic reasoner, or (iii) enforcing the LLM to simulate a logical inference rule. However, all these solutions have critical limitations: they are unable to leverage the internal commonsense knowledge of the LLM in tandem with an axiomatic knowledge base, they lack a mechanism to reliably repair erroneous inference steps, and their application is restricted to small knowledge bases that fit the context limit of the LLM. In this work, we present LLM-based Typed Hyperresolution (LLM-TH), a logical commonsense reasoning framework that leverages \"theory resolution\", a concept from classical logical inference which enables integrating LLMs into the \"resolution\" inference rule, thus mitigating reasoning errors and hallucinations and enabling verification of the reasoning procedure. LLM-TH is also equipped with a mechanism for repairing erroneous inference steps supported by theoretical guarantees. Using \"Hyperresolution\" and \"Typed inference\" schemes, we show that LLM-TH can efficiently reason over large knowledge bases consisting of tens of thousands of rules with arbitrary predicate arities. Our experiments on three diverse language-based reasoning tasks\u2014preference reasoning, multi-domain deductive reasoning, and geographical question answering\u2014showcase that LLM-TH, using merely a BART 406M parameter NLI entailment model, significantly reduces reasoning errors compared to baselines using Llama3-70B, Gemini1.5-Flash, GPT-3.5-Turbo, and Mixtral-46.7B.", "keywords": [ "Large Language Models", "Commonsense reasoning", "Logical inference" ], "rating_list": [ 6, 1, 8, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "wNg0LibmQt", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Erik Brinkman", "gender": "unknown", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Hannah Korevaar", "gender": "Female", "institution": "Meta Platforms, Inc", "country": "", "position": "Researcher" }, { "name": "Ivan Evtimov", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Research Scientist" }, { "name": "Javier Rando", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Florian Tramer", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 23, "n_ref": 60, "n_ref_all": 77, "n_fig": 2, "n_tab": 6, "L_tab": 1123, "n_element_tab": 130, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1302, "n_element_tab_1": 138, "formula_len_all": 43, "formula_len_all_1": 43, "len_all": 94516, "len_all_1": 66705, "len_abs": 1027, "len_title": 108, "len_sents": 30003, "len_sents_1": 28520, "n_sents": 272, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1035, "title": "Gradient-based Jailbreak Images for Multimodal Fusion Models", "abs": "Augmenting language models with image inputs may enable more effective jailbreak attacks through continuous optimization, unlike text inputs that require discrete optimization. However, new *multimodal fusion models* tokenize all input modalities using non-differentiable functions, which hinders straightforward attacks. In this work, we introduce the notion of a *tokenizer shortcut* that approximates tokenization with a continuous function and enables continuous optimization. We use tokenizer shortcuts to create the first end-to-end gradient image attacks against multimodal fusion models. We evaluate our attacks on Chameleon models and obtain jailbreak images that elicit harmful information for 72.5% of prompts. Jailbreak images outperform text jailbreaks optimized with the same objective and require 3x lower compute budget to optimize 50x more input tokens. Finally, we find that representation engineering defenses, like Circuit Breakers, trained only on text attacks can effectively transfer to adversarial image inputs.", "keywords": [ "jailbreak", "adversarial examples", "multimodal", "language models" ], "rating_list": [ 3, 3, 3, 8, 8 ], "soundness_list": [ 1, 3, 2, 3, 3 ], "presentation_list": [ 1, 3, 3, 3, 4 ], "contribution_list": [ 1, 2, 1, 3, 4 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "wN3KaUXA5X", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Erik Jenner", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Shreyas Kapur", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Stuart Russell", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 39, "n_ref": 69, "n_ref_all": 90, "n_fig": 11, "n_tab": 1, "L_tab": 159, "n_element_tab": 11, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 317, "formula_len_all_1": 61, "len_all": 134269, "len_all_1": 59633, "len_abs": 1040, "len_title": 96, "len_sents": 39132, "len_sents_1": 26739, "n_sents": 351, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 47, "L_abs": 1047, "title": "Diffusion On Syntax Trees For Program Synthesis", "abs": "Large language models generate code one token at a time. Their autoregressive generation process lacks the feedback of observing the program's output. Training LLMs to suggest edits directly can be challenging due to the scarcity of rich edit data. To address these problems, we propose neural diffusion models that operate on syntax trees of any context-free grammar. Similar to image diffusion models, our method also inverts \"noise\" applied to syntax trees. Rather than generating code sequentially, we iteratively edit it while preserving syntactic validity, which makes it easy to combine this neural model with search. We apply our approach to inverse graphics tasks, where our model learns to convert images into programs that produce those images. Combined with search, our model is able to write graphics programs, see the execution result, and debug them to meet the required specifications. We additionally show how our system can write graphics programs for hand-drawn sketches. Video results can be found at https://td-anon.github.io.", "keywords": [ "neurosymbolic", "search", "programming languages", "inverse graphics" ], "rating_list": [ 8, 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 4, 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 5, 3, 3, 3, 4 ] }, { "paper_id": "wMj6PgKVuJ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Christos Perivolaropoulos", "gender": "Male", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Federico Barbero", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Petar Veli\u010dkovi\u0107", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Affiliated Lecturer" }, { "name": "Razvan Pascanu", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Affiliate Member" } ], "n_formula": 47, "n_formula_1": 15, "n_ref_uni": 47, "n_ref": 62, "n_ref_all": 85, "n_fig": 8, "n_tab": 1, "L_tab": 701, "n_element_tab": 35, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3596, "formula_len_all_1": 1373, "len_all": 167154, "len_all_1": 51972, "len_abs": 973, "len_title": 101, "len_sents": 48396, "len_sents_1": 24018, "n_sents": 384, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 976, "title": "softmax is not enough (for sharp out-of-distribution)", "abs": "A key property of reasoning systems is the ability to make sharp decisions on their input data. For contemporary AI systems, a key carrier of sharp behaviour is the softmax function, with its capability to perform differentiable query-key lookups. It is a common belief that the predictive power of networks leveraging softmax arises from \"circuits\" which sharply perform certain kinds of computations consistently across many diverse inputs. However, for these circuits to be robust, they would need to generalise well to arbitrary valid inputs. In this paper, we dispel this myth: even for tasks as simple as finding the maximum key, any learned circuitry must disperse as the number of items grows at test time. We attribute this to a fundamental limitation of the softmax function to robustly approximate sharp functions, prove this phenomenon theoretically, and propose adaptive temperature as an ad-hoc technique for improving the sharpness of softmax at inference time.", "keywords": [ "softmax", "transformers", "out-of-distribution", "sharpness", "entropy" ], "rating_list": [ 6, 6, 5, 8, 6 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 1, 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 3, 1 ] }, { "paper_id": "wMgr7wBuUo", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Amirozhan Dehghani", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "MS student" }, { "name": "Asa Farahani", "gender": "Female", "institution": "McGill University, McGill University", "country": "CA", "position": "PhD student" }, { "name": "Xinyu Qian", "gender": "unknown", "institution": "McGill University", "country": "", "position": "Undergrad student" }, { "name": "Pouya Bashivan", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 47, "n_ref": 107, "n_ref_all": 130, "n_fig": 17, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 13, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 259, "formula_len_all_1": 172, "len_all": 144417, "len_all_1": 58802, "len_abs": 1389, "len_title": 155, "len_sents": 46460, "len_sents_1": 31018, "n_sents": 284, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 106, "L_abs": 1394, "title": "Credit-based self organizing maps: training deep topographic networks with minimal performance degradation", "abs": "In the primate neocortex, neurons with similar function are often found to be spatially close. Kohonen's self-organizing map (SOM) has been one of the most influential approaches for simulating brain-like topographical organization in artificial neural network models. However, integrating these maps into deep neural networks with multitude of layers has been challenging, with self-organized deep neural networks suffering from substantially diminished capacity to perform visual recognition. We identified a key factor leading to the performance degradation in self-organized topographical neural network models: the discord between predominantly bottom-up learning updates in the self-organizing maps, and those derived from top-down, credit-based learning approaches. To address this, we propose an alternative self organization algorithm, tailored to align with the top-down learning processes in deep neural networks. This model not only emulates critical aspects of cortical topography but also significantly narrows the performance gap between non-topographical and topographical models. This advancement underscores the substantial importance of top-down assigned credits in shaping topographical organization. Our findings are a step in reconciling topographical modeling with the functional efficacy of neural network models, paving the way for more brain-like neural architectures.", "keywords": [ "Computer vision", "Neuroscience", "Convolutional Networks", "topographical organization", "self-organizing maps", "functional organization" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "wMSZEP7BDh", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Kawisorn Kamtue", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Orathai Sangpetch", "gender": "Female", "institution": "CMKL University", "country": "US", "position": "Assistant Professor" }, { "name": "Jos\u00e9 M. F. Moura", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Professor" } ], "n_formula": 25, "n_formula_1": 17, "n_ref_uni": 14, "n_ref": 23, "n_ref_all": 37, "n_fig": 11, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2769, "formula_len_all_1": 1378, "len_all": 94199, "len_all_1": 54629, "len_abs": 1225, "len_title": 162, "len_sents": 33184, "len_sents_1": 24138, "n_sents": 265, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 114, "L_abs": 1233, "title": "Is Pontryagin's Maximum Principle All You Need? Solving optimal control problems with PMP-inspired neural networks", "abs": "Calculus of Variations is the mathematics of functional optimization, i.e., when the solutions are functions over a time interval. This is particularly important when the time interval, or support, is unknown like in minimum-time control problems, so that forward-in-time solutions are not possible. Calculus of Variations also offers a robust framework for learning optimal control and inference with moving boundaries. How can this framework be leveraged to design neural networks to solve challenges in control and inference? We propose the Pontryagin's Maximum Principle Neural Network (PMP-Net) that is tailored to estimate control and inference solutions, in accordance with the necessary conditions outlined by Pontryagin\u2019s Maximum Principle. We assess PMP-Net on two classic optimal control and inference problems: optimal linear filtering and minimum-time control. Our findings indicate that PMP-Net can be effectively trained in an unsupervised manner to solve these problems without the need for ground-truth data, successfully deriving the classical ''Kalman filter'' and \"bang-bang'' control solution. This establishes a new approach for addressing general, possibly yet unsolved, inference and optimal control problems.", "keywords": [ "prior knowledge", "Pontryagin's Maximum Principle", "optimal control" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "wMRFTQwp1d", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Jing Wang", "gender": "Female", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Kunchang Li", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xinhao Li", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Intern" }, { "name": "Zhenpeng Huang", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Limin Wang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 19, "n_ref": 51, "n_ref_all": 71, "n_fig": 4, "n_tab": 7, "L_tab": 5817, "n_element_tab": 305, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 3159, "n_element_tab_1": 139, "formula_len_all": 44, "formula_len_all_1": 308, "len_all": 175776, "len_all_1": 50963, "len_abs": 1496, "len_title": 138, "len_sents": 34259, "len_sents_1": 20366, "n_sents": 232, "n_sents_1": 134, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1534, "title": "VideoEval: Comprehensive Benchmark Suite for Low-Cost Evaluation of Video Foundation Model", "abs": "With the accumulation of high-quality data and advancements in visual pretraining paradigms, recent Video Foundation Models (VFMs) have made significant progress, demonstrating remarkable performance on popular video understanding benchmarks. However, conventional benchmarks (e.g. Kinetics) and evaluation protocols are limited by their relatively poor diversity, high evaluation costs, and saturated performance metrics. In this work, we introduce a comprehensive benchmark suite to address these issues, namely **VideoEval**. We establish the **Vid**eo **T**ask **A**daption **B**enchmark (VidTAB) and the **Vid**eo **E**mbedding **B**enchmark (VidEB) from two perspectives: evaluating the task adaptability of VFMs under few-shot conditions and assessing their feature embedding's direct applicability to downstream tasks. With VideoEval, we conduct a large-scale study of 20 popular open-source vision foundation models. Our study reveals some insightful findings, 1) overall, current VFMs exhibit weak generalization across diverse tasks, 2) increasing video data, whether labeled or in video-text pairs, does not necessarily improve task performance, 3) the effectiveness of some pre-training paradigms may not be fully validated in previous benchmarks, and 4) combining different pre-training paradigms can help develop models with better generalization capabilities. We believe this study serves as a important complement to the current evaluation methods for VFMs and offers valuable insights for future research directions.", "keywords": [ "Video Understanding", "Video Foundation Model", "Benchmark" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "wM2sfVgMDH", "primary_area": "applications to robotics, autonomy, planning", "n_author": 11, "author_info_list": [ { "name": "Jingjing Liu", "gender": "unknown", "institution": "Tsinghua University", "country": "US", "position": "Full Professor" }, { "name": "Jinliang Zheng", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Kexin ZHENG", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Undergrad student" }, { "name": "Rui Ai", "gender": "Male", "institution": "HAOMO.AI Technology Co.,Ltd. ", "country": "", "position": "Researcher" }, { "name": "Ruiming Liang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Shengbo Eben Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Weihao Gu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xianyuan Zhan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yinan Zheng", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Li Jianxiong", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Liyuan Richard Mao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 18, "n_ref_uni": 50, "n_ref": 111, "n_ref_all": 143, "n_fig": 9, "n_tab": 8, "L_tab": 2679, "n_element_tab": 391, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 914, "n_element_tab_1": 81, "formula_len_all": 1629, "formula_len_all_1": 1256, "len_all": 181323, "len_all_1": 64706, "len_abs": 1324, "len_title": 118, "len_sents": 49013, "len_sents_1": 30508, "n_sents": 361, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 70, "L_abs": 1327, "title": "Diffusion-Based Planning for Autonomous Driving with Flexible Guidance", "abs": "Achieving human-like driving behaviors in complex open-world environments is a critical challenge in autonomous driving. Contemporary learning-based planning approaches such as imitation learning methods often struggle to balance competing objectives and lack of safety assurance,due to limited adaptability and inadequacy in learning complex multi-modal behaviors commonly exhibited in human planning, not to mention their strong reliance on the fallback strategy with predefined rules. We propose a novel transformer-based Diffusion Planner for closed-loop planning, which can effectively model multi-modal driving behavior and ensure trajectory quality without any rule-based refinement. Our model supports joint modeling of both prediction and planning tasks under the same architecture, enabling cooperative behaviors between vehicles. Moreover, by learning the gradient of the trajectory score function and employing a flexible classifier guidance mechanism, Diffusion Planner effectively achieves safe and adaptable planning behaviors. Evaluations on the large-scale real-world autonomous planning benchmark nuPlan and our newly collected 200-hour delivery-vehicle driving dataset demonstrate that Diffusion Planner achieves state-of-the-art closed-loop performance with robust transferability in diverse driving styles.", "keywords": [ "diffusion planning", "autonomous driving" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "wLzhEQq2hR", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Buse Giledereli", "gender": "Female", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Yifan Hou", "gender": "Male", "institution": "Department of Computer Science, Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" }, { "name": "Yilei Tu", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Intern" }, { "name": "MRINMAYA SACHAN", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 42, "n_ref": 85, "n_ref_all": 166, "n_fig": 22, "n_tab": 7, "L_tab": 1250, "n_element_tab": 93, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 996, "n_element_tab_1": 75, "formula_len_all": 1299, "formula_len_all_1": 679, "len_all": 256250, "len_all_1": 63234, "len_abs": 1594, "len_title": 108, "len_sents": 78090, "len_sents_1": 28804, "n_sents": 791, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1532, "title": "Do Vision-Language Models Really Understand Visual Language?", "abs": "Visual language is a system of communication that conveys information through symbols, shapes, and spatial arrangements. Diagrams are a typical example of a visual language depicting complex concepts and their relationships in the form of an image. The symbolic nature of diagrams presents significant challenges for building models capable of understanding them. Yet, recent studies seem to suggest that Large Vision-Language Models (LVLMs) can even tackle complex reasoning tasks involving diagrams. In this paper, we investigate this phenomenon by developing a comprehensive test suite to evaluate the diagram comprehension capability of LVLMs. Our test suite uses a variety of questions focused on concept entities and their relationships over a set of synthetic as well as real diagrams across several domains to evaluate the recognition and reasoning abilities of models. Our evaluation of three LVLMs (GPT-4V, GPT-4o, and Gemini) shows that while these models can accurately identify and reason about entities, their ability to understand relationships is notably limited. Further testing reveals that the decent performance on diagram understanding largely stems from leveraging their background knowledge as shortcuts to identify and reason about the relational information. Thus, we conclude that LVLMs have a limited capability for genuine diagram understanding, and their impressive performance in diagram reasoning is an illusion emanating from other confounding factors, such as the background knowledge in the models.", "keywords": [ "vision-language model", "visual language", "diagram reasoning", "evaluation" ], "rating_list": [ 5, 8, 6, 5 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 3, 2, 4 ] }, { "paper_id": "wLnls9LS3x", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Kshiteej Sheth", "gender": "Male", "institution": "EPFL", "country": "CH", "position": "PhD student" }, { "name": "Michael Kapralov", "gender": "unknown", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "", "position": "Assistant Professor" }, { "name": "Piotr Indyk", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Tal Wagner", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 13, "n_ref_uni": 25, "n_ref": 59, "n_ref_all": 78, "n_fig": 6, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2029, "formula_len_all_1": 918, "len_all": 133854, "len_all_1": 77810, "len_abs": 949, "len_title": 134, "len_sents": 43857, "len_sents_1": 32928, "n_sents": 390, "n_sents_1": 285, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1128, "title": "Improved Algorithms for Kernel Matrix-Vector Multiplication", "abs": "Motivated by the problem of fast processing of attention matrices, we study fast algorithms for computing matrix-vector products for asymmetric Gaussian Kernel matrices $K\\in \\mathbb{R}^{n\\times n}$. \n$K$'s columns are indexed by a set of $n$ keys $k_1,k_2\\ldots, k_n\\in \\mathbb{R}^d$, rows by a set of $n$ queries $q_1,q_2,\\ldots,q_n\\in \\mathbb{R}^d $, and its $i,j$ entry is $K_{ij} = e^{-\\|q_i-k_j\\|_2^2/2\\sigma^2}$ for some bandwidth parameter $\\sigma>0$. Given a vector $x\\in \\mathbb{R}^n$ and error parameter $\\epsilon>0$, our task is to output a $y\\in \\mathbb{R}^n$ such that $\\|Kx-y\\|_2\\leq \\epsilon \\|x\\|_2$ in time subquadratic in $n$ and linear in $d$. Our algorithms rely on the following modelling assumption about the matrices $K$: the sum of the entries of $K$ scales linearly in $n$, as opposed to worst case quadratic growth. We validate this assumption experimentally, for Gaussian kernel matrices encountered in various settings such as fast attention computation in LLMs. Under this assumption, we obtain the first subquadratic time algorithm for kernel matrix-vector multiplication for unrestricted vectors.", "keywords": [ "Algorithms", "Kernel Matrix", "Kernel Density Estimation", "Locality Sensitive Hashing", "Fast Attention" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "wLmJIs1uqG", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Bin Gao", "gender": "Male", "institution": "Academy of Mathematics and Systems Science", "country": "CN", "position": "Associate Professor" }, { "name": "Ya-xiang Yuan", "gender": "unknown", "institution": "Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yan Yang", "gender": "Male", "institution": "Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 179, "n_formula_1": 23, "n_ref_uni": 71, "n_ref": 187, "n_ref_all": 245, "n_fig": 13, "n_tab": 3, "L_tab": 1171, "n_element_tab": 138, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 309, "n_element_tab_1": 34, "formula_len_all": 17474, "formula_len_all_1": 2645, "len_all": 260062, "len_all_1": 71321, "len_abs": 1110, "len_title": 120, "len_sents": 62604, "len_sents_1": 27467, "n_sents": 630, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1136, "title": "LancBiO: Dynamic Lanczos-aided Bilevel Optimization via Krylov Subspace", "abs": "Bilevel optimization, with broad applications in machine learning, has an intricate hierarchical structure. Gradient-based methods have emerged as a common approach to large-scale bilevel problems. However, the computation of the hyper-gradient, which involves a Hessian inverse vector product, confines the efficiency and is regarded as a bottleneck. To circumvent the inverse, we construct a sequence of low-dimensional approximate Krylov subspaces with the aid of the Lanczos process. As a result, the constructed subspace is able to dynamically and incrementally approximate the Hessian inverse vector product with less effort and thus leads to a favorable estimate of the hyper-gradient. Moreover, we propose a provable subspace-based framework for bilevel problems where one central step is to solve a small-size tridiagonal linear system. To the best of our knowledge, this is the first time that subspace techniques are incorporated into bilevel optimization. This successful trial not only enjoys $\\mathcal{O}(\\epsilon^{-1})$ convergence rate but also demonstrates efficiency in a synthetic problem and two deep learning tasks.", "keywords": [ "Bilevel Optimization", "Lanczos Process", "Krylov Subspace" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "wLR9d5ZFpY", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Chaitanya Murti", "gender": "Male", "institution": "Indian Institute of Science", "country": "IN", "position": "PhD student" }, { "name": "Chiranjib Bhattacharyya", "gender": "Male", "institution": "Indian Institute of Science, Indian institute of science, Bangalore", "country": "IN", "position": "Full Professor" }, { "name": "Dhruva Kashyap", "gender": "Male", "institution": "Indian Institute of Science, Indian institute of science, Bangalore", "country": "IN", "position": "Research Assistant" }, { "name": "Tanay Narshana", "gender": "Male", "institution": "Google", "country": "IN", "position": "ML Software Engineer" } ], "n_formula": 53, "n_formula_1": 28, "n_ref_uni": 50, "n_ref": 95, "n_ref_all": 108, "n_fig": 4, "n_tab": 7, "L_tab": 1928, "n_element_tab": 248, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5560, "formula_len_all_1": 2172, "len_all": 201949, "len_all_1": 66982, "len_abs": 1977, "len_title": 124, "len_sents": 53932, "len_sents_1": 29911, "n_sents": 435, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1991, "title": "No Training Data, No Cry: Model Editing without Training Data or Fine-tuning", "abs": "Model Editing(ME)--such as classwise unlearning and structured pruning--is a nascent field that deals with identifying editable components that, when modified, significantly change the model's behaviour, typically requiring fine-tuning to regain performance.\nThe challenge of model editing increases when dealing with multi-branch networks(e.g. ResNets) in the data-free regime, where the training data and the loss function are not available.\nIdentifying editable components is more difficult in multi-branch networks due to the coupling of individual components across layers through skip connections. \nThis paper addresses these issues through the following contributions.\nFirst, we hypothesize that in a well-trained model, there exists a small set of channels, which we call HiFi channels, whose input contributions strongly correlate with the output feature map of that layer.\nFinding such subsets can be naturally posed as an expected reconstruction error problem. To solve this, we provide an efficient heuristic called RowSum.\nSecond, to understand how to regain accuracy after editing, we prove, for the first time, an upper bound on the loss function post-editing in terms of the change in the stored BatchNorm(BN) statistics. With this result, we derive BNFix, a simple algorithm to restore accuracy by updating the BN statistics using distributional access to the data distribution.\nWith these insights, we propose retraining free algorithms for structured pruning and classwise unlearning, CoBRA-P and CoBRA-U, that identify HiFi components and retains(structured pruning) or discards(classwise unlearning) them. CoBRA-P achieves at least 50% larger reduction in FLOPS and at least 10% larger reduction in parameters for similar drop in accuracy in the training free regime. In the training regime, for ImageNet, it achieves 60% larger parameter reduction. CoBRA-U achieves, on average, a 94% reduction in forget-class accuracy with a minimal drop in remaining class accuracy.", "keywords": [ "pruning", "model editing", "classwise unlearning" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 1, 1 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "wKOoWTBMZe", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Jian Wang", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Jinzhou Tang", "gender": "Not Specified", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "Jusheng Zhang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Intern" }, { "name": "Kaiyu Wu", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "Keze Wang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Mingyan Li", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "Sidi Liu", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "Yufeng Yang", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Postdoc" }, { "name": "Yangzhuojie", "gender": "Male", "institution": "Sun Yat-sen university", "country": "CN", "position": "MS student" } ], "n_formula": 81, "n_formula_1": 25, "n_ref_uni": 23, "n_ref": 37, "n_ref_all": 39, "n_fig": 0, "n_tab": 2, "L_tab": 160, "n_element_tab": 14, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 165, "n_element_tab_1": 14, "formula_len_all": 4652, "formula_len_all_1": 1363, "len_all": 108926, "len_all_1": 53122, "len_abs": 221, "len_title": 132, "len_sents": 33133, "len_sents_1": 21867, "n_sents": 280, "n_sents_1": 153, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 941, "title": "KEYPOINT-GUIDED 4D GAUSSIAN SPLATTING WITH DECOUPLED SPATIO-TEMPORAL FLOW REFINEMENT", "abs": "We propose KG4D, a novel method for generating time-aware 4D representations\nfrom a single static image or video. Previous methods largely rely on weak su-\npervision signals, failing to introduce fine-grained supervision necessary for cap-\nturing detailed spatio-temporal dynamics. In contrast, our approach employs Har-\nmonic Spatio-temporal Encoding (HSE) to achieve efficient spatio-temporal sep-\naration during training, allowing the model to represent dynamic scene changes\nmore accurately. Furthermore, Keypoint Feature Calibration (KFC) ensures pre-\ncise pose consistency, and Wasserstein Gradient Flow (WGF) enhances motion\ncoherence, effectively reducing artifacts. Comprehensive evaluation and ablations\ndemonstrate that our proposed KG4D outperforms existing state-of-the-art meth-\nods on various benchmarks in dynamic 4D generation and novel viewpoint syn-\nthesis, validating its effectiveness and superior generation capability.", "keywords": [ "Keypoint", "4D Gaussian Splatting" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 2, 1 ], "presentation_list": [ 1, 2, 1 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 5 ] }, { "paper_id": "wJv4AIt4sK", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 11, "author_info_list": [ { "name": "Amir Yazdanbakhsh", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Ayan Chakraborty", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Babak Falsafi", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "Full Professor" }, { "name": "Danila Mishin", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "MS student" }, { "name": "Dongho Ha", "gender": "Male", "institution": "MangoBoost Inc.", "country": "KR", "position": "System Architect" }, { "name": "Elizaveta Kostenok", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "", "position": "MS student" }, { "name": "Martin Jaggi", "gender": "Male", "institution": "EPFL", "country": "CH", "position": "Associate Professor" }, { "name": "Ming Liu", "gender": "Male", "institution": "Google", "country": "US", "position": "Software Engineer" }, { "name": "Simla Burcu Harma", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Suvinay Subramanian", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Yunho Oh", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 35, "n_formula_1": 13, "n_ref_uni": 0, "n_ref": 30, "n_ref_all": 66, "n_fig": 9, "n_tab": 17, "L_tab": 8116, "n_element_tab": 977, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 1954, "n_element_tab_1": 296, "formula_len_all": 4154, "formula_len_all_1": 871, "len_all": 303560, "len_all_1": 77575, "len_abs": 1570, "len_title": 126, "len_sents": 70760, "len_sents_1": 33907, "n_sents": 595, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 78, "L_abs": 1575, "title": "Effective Interplay between Sparsity and Quantization: From Theory to Practice", "abs": "The increasing size of deep neural networks (DNNs) necessitates effective model compression to reduce their computational and memory footprints. Sparsity and quantization are two prominent compression methods that have been shown to reduce DNNs' computational and memory footprints significantly while preserving model accuracy. However, how these two methods interact when combined together remains a key question for developers, as many tacitly assume that they are orthogonal, meaning that their combined use does not introduce additional errors beyond those introduced by each method independently. In this paper, we provide the first mathematical proof that sparsity and quantization are non-orthogonal. We corroborate these results with experiments spanning a range of large language models, including the OPT and LLaMA model families (with 125M to 8B parameters), and vision models like ViT and ResNet. We show that the order in which we apply these methods matters because applying quantization before sparsity may disrupt the relative importance of tensor elements, which may inadvertently remove significant elements from a tensor. More importantly, we show that even if applied in the correct order, the compounded errors from sparsity and quantization can significantly harm accuracy. Our findings extend to the efficient deployment of large models in resource-constrained compute platforms to reduce serving cost, offering insights into best practices for applying these compression methods to maximize hardware resource efficiency without compromising accuracy.", "keywords": [ "theory of compression", "model compression", "quantization", "max-scaled numerical encoding", "sparsity", "unstructured sparsity", "structured sparsity", "N:M sparsity", "large language models", "magnitude pruning", "post-training quantization", "efficient inference" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 2, 2, 4, 3 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "wJlzUR5sFl", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Ali Hojjat", "gender": "Male", "institution": "Kiel University", "country": "DE", "position": "PhD student" }, { "name": "Janek Haberer", "gender": "Male", "institution": "Christian-Albrechts-Universit\u00e4t Kiel", "country": "DE", "position": "PhD student" }, { "name": "Olaf Landsiedel", "gender": "Male", "institution": "Christian-Albrechts-Universit\u00e4t Kiel", "country": "DE", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 51, "n_ref": 87, "n_ref_all": 106, "n_fig": 9, "n_tab": 3, "L_tab": 821, "n_element_tab": 122, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 988, "n_element_tab_1": 229, "formula_len_all": 85, "formula_len_all_1": 85, "len_all": 113516, "len_all_1": 55478, "len_abs": 1590, "len_title": 120, "len_sents": 25048, "len_sents_1": 24391, "n_sents": 164, "n_sents_1": 156, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1342, "title": "MCUCoder: Adaptive Bitrate Learned Video Compression for IoT Devices", "abs": "The rapid growth of camera-based IoT devices demands the need for efficient video compression, particularly for edge applications where devices face hardware constraints, often with only 1 or 2 MB of RAM and unstable internet connections. Traditional and deep video compression methods are designed for high-end hardware, exceeding the capabilities of these constrained devices. Consequently, video compression in these scenarios is often limited to M-JPEG due to its high hardware efficiency and low complexity. This paper introduces , an open-source adaptive bitrate video compression model tailored for resource-limited IoT settings. MCUCoder features an ultra-lightweight encoder with only 10.5K parameters and a minimal 350KB memory footprint, making it well-suited for edge devices and MCU. While MCUCoder uses a similar amount of energy as M-JPEG, it reduces bitrate by 55.65\\% on the MCL-JCV dataset and 55.59\\% on the UVG dataset, measured in MS-SSIM. Moreover, MCUCoder supports adaptive bitrate streaming by generating a latent representation that is sorted by importance, allowing transmission based on available bandwidth. This ensures smooth real-time video transmission even under fluctuating network conditions on low-resource devices. Source code available at [Link removed due to double-blind policy, code submitted in ZIP].", "keywords": [ "Efficient Video Compression", "IoT Devices", "Learned Video Compression", "Adaptive Bitrate Compression", "Microcontrollers" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "wJVZkUOUjh", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Amanda S Barnard", "gender": "Female", "institution": "Australian National University", "country": "", "position": "Full Professor" }, { "name": "Quanling Deng", "gender": "Male", "institution": "Australian National University", "country": "", "position": "Lecturer" }, { "name": "Sichao Li", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 16, "n_ref_uni": 39, "n_ref": 93, "n_ref_all": 126, "n_fig": 7, "n_tab": 11, "L_tab": 8328, "n_element_tab": 1368, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1802, "n_element_tab_1": 396, "formula_len_all": 1832, "formula_len_all_1": 920, "len_all": 180953, "len_all_1": 65698, "len_abs": 1157, "len_title": 70, "len_sents": 52582, "len_sents_1": 28134, "n_sents": 386, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1160, "title": "EXAGREE: Towards Explanation Agreement in Explainable Machine Learning", "abs": "Explanations in machine learning are critical for trust, transparency, and fairness. Yet, complex disagreements among these explanations limit the reliability and applicability of machine learning models, especially in high-stakes environments. We formalize four fundamental ranking-based explanation disagreement problems and introduce a novel framework, EXplanation AGREEment (EXAGREE), to bridge diverse interpretations in explainable machine learning, particularly from stakeholder-centric perspectives. Our approach leverages a Rashomon set for attribution predictions and then optimizes within this set to identify Stakeholder-Aligned Explanation Models (SAEMs) that minimize disagreement with diverse stakeholder needs while maintaining predictive performance. Rigorous empirical analysis on synthetic and real-world datasets demonstrates that EXAGREE reduces explanation disagreement and improves fairness across subgroups in various domains. EXAGREE not only provides researchers with a new direction for studying explanation disagreement problems but also offers data scientists a tool for making better-informed decisions in practical applications.", "keywords": [ "Explainable Machine Learning", "Explainable Artificial Intelligence", "Rashomon Sets" ], "rating_list": [ 1, 3, 1, 3 ], "soundness_list": [ 1, 3, 1, 2 ], "presentation_list": [ 1, 1, 2, 1 ], "contribution_list": [ 1, 3, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "wJPMe9UKow", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Dongwoo Kim", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Assistant Professor" }, { "name": "Jaeseung Heo", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "MoonJeong Park", "gender": "unknown", "institution": "POSTECH", "country": "", "position": "MS student" } ], "n_formula": 36, "n_formula_1": 13, "n_ref_uni": 44, "n_ref": 82, "n_ref_all": 110, "n_fig": 9, "n_tab": 15, "L_tab": 9659, "n_element_tab": 556, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 6830, "n_element_tab_1": 215, "formula_len_all": 3263, "formula_len_all_1": 1403, "len_all": 175898, "len_all_1": 73407, "len_abs": 1045, "len_title": 98, "len_sents": 46411, "len_sents_1": 27540, "n_sents": 459, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1045, "title": "Posterior Label Smoothing for Node Classification", "abs": "Soft labels can improve the generalization of a neural network classifier in many domains, such as image classification. Despite its success, the current literature has overlooked the efficiency of label smoothing in node classification with graph-structured data. In this work, we propose a simple yet effective label smoothing for the transductive node classification task. We design the soft label to encapsulate the local context of the target node through the neighborhood label distribution. We apply the smoothing method for seven baseline models to show its effectiveness. The label smoothing methods improve the classification accuracy in 10 node classification datasets in most cases. In the following analysis, we find that incorporating global label statistics in posterior computation is the key to the success of label smoothing. Further investigation reveals that the soft labels mitigate overfitting during training, leading to better generalization performance. Our code is available at https://anonymous.4open.science/r/PosteL.", "keywords": [ "node classification", "label smoothing" ], "rating_list": [ 6, 8, 5, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 1, 4 ] }, { "paper_id": "wJGXiHQwpZ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Dong Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Jiawen Wang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Xiaoyu Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yinda Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zhiwei Xiong", "gender": "Male", "institution": "USTC", "country": "CN", "position": "Professor" }, { "name": "che liu", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 57, "n_formula_1": 15, "n_ref_uni": 65, "n_ref": 165, "n_ref_all": 191, "n_fig": 8, "n_tab": 3, "L_tab": 1415, "n_element_tab": 120, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1170, "n_element_tab_1": 81, "formula_len_all": 3573, "formula_len_all_1": 939, "len_all": 214061, "len_all_1": 58572, "len_abs": 1394, "len_title": 130, "len_sents": 53955, "len_sents_1": 26476, "n_sents": 464, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1398, "title": "MaskTwins: Dual-form Complementary Masking for Domain-Adaptive Image Segmentation", "abs": "Recent works have correlated Masked Image Modeling (MIM) with consistency regularization in unsupervised domain adaptation. However, they merely treat masking as a special form of deformation on the input images and neglect the theoretical analysis, which leads to a superficial understanding of masked reconstruction and insufficient exploitation of its potential in enhancing feature extraction and representation learning. In this paper, we reframe masked reconstruction as a sparse signal reconstruction problem and theoretically prove that the dual form of complementary masks possesses superior capabilities in extracting domain-agnostic image features. Based on this compelling insight, we propose MaskTwins, a simple yet effective learning strategy that integrates masked reconstruction directly into the main training pipeline. MaskTwins uncovers intrinsic structural patterns that persist across disparate domains by enforcing consistency between predictions of images masked in complementary ways, enabling domain generalization in an end-to-end manner. Extensive experiments verify the superiority of MaskTwins over baseline methods in natural and biological image segmentation. These results demonstrate the significant advantages of MaskTwins in extracting domain-invariant features without the need for separate pre-training, offering a new paradigm for domain-adaptive segmentation.", "keywords": [ "domain adaptation", "unsupervised learning", "masked image modeling", "semantic segmentation", "complementary masking" ], "rating_list": [ 5, 3, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "wJ6Bx1IYrQ", "primary_area": "applications to neuroscience & cognitive science", "n_author": 7, "author_info_list": [ { "name": "Jie Jiang", "gender": "Female", "institution": ", Institute of automation, Chinese academy of science", "country": "CN", "position": "Assistant Professor" }, { "name": "Jing Liu", "gender": "Female", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "Full Professor" }, { "name": "Longteng Guo", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Shuning Xue", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Tongtian Yue", "gender": "Male", "institution": ", Institute of automation, Chinese academy of science", "country": "CN", "position": "PhD student" }, { "name": "Xuange Gao", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yepeng Tang", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 47, "n_ref": 87, "n_ref_all": 98, "n_fig": 4, "n_tab": 7, "L_tab": 2527, "n_element_tab": 358, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 3066, "n_element_tab_1": 377, "formula_len_all": 530, "formula_len_all_1": 535, "len_all": 149946, "len_all_1": 66677, "len_abs": 1959, "len_title": 98, "len_sents": 40538, "len_sents_1": 27312, "n_sents": 334, "n_sents_1": 226, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1970, "title": "EEGPT: Unleashing the Potential of EEG Generalist Foundation Model by Autoregressive Pre-training", "abs": "Electroencephalogram (EEG) signals are pivotal in providing insights into spontaneous brain activity, highlighting their significant importance in neuroscience research. However, the exploration of versatile EEG models is constrained by diverse data formats, outdated pre-training paradigms, and limited transfer learning methods, only leading to specialist models on single dataset. In this paper, we introduce EEGPT, the first generalist EEG foundation model designed to address these challenges. First, we propose an electrode-wise modeling strategy that treats each electrode as a fundamental unit, enabling the integration of diverse EEG datasets collected from up to 138 electrodes, amassing 37.5M pre-training samples. Second, we develop the first autoregressive EEG pre-trained model, moving away from traditional masked autoencoder approaches to a next signal prediction task that better captures the sequential and temporal dependencies of EEG data. We also explore scaling laws with model up to 1.1B parameters \u2014 the largest in EEG research to date. Third, we introduce a multi-task transfer learning paradigm using a learnable electrode graph network that is shared across tasks, which for the first time confirms multi-task compatibility and synergy. As the first generalist EEG foundation model, EEGPT shows broad compatibility with various signal acquisition devices, subjects, and tasks. It supports up to 138 electrodes and any combination thereof as input. Furthermore, we simultaneously evaluate it on 5 distinct downstream tasks across 12 benchmarks. EEGPT consistently outperforms existing specialist models across all downstream tasks, with its effectiveness further validated through extensive ablation studies.\nThis work sets a new direction for generalist EEG modeling, offering improved scalability, transferability, and adaptability for a wide range of EEG applications. Both the training code and model checkpoints will be publicly available.", "keywords": [ "EEG", "Brain-computer interface", "Representation learning" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "wJ3GeGLFmc", "primary_area": "optimization", "n_author": 10, "author_info_list": [ { "name": "Shulin Zeng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Weiying Xie", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Xuefei Ning", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Research Assistant Professor" }, { "name": "Yu Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yunsong Li", "gender": "Male", "institution": "Xidian University ", "country": "CN", "position": "Full Professor" }, { "name": "Zhenman Fang", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "Assistant Professor" }, { "name": "Zihan Meng", "gender": "unknown", "institution": "Xidian University", "country": "CN", "position": "PhD student" }, { "name": "donglai Liu", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "MS student" }, { "name": "JIE LEI", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Researcher" }, { "name": "Wen-jin Guo", "gender": "Male", "institution": "XiDian University", "country": "CN", "position": "PhD student" } ], "n_formula": 50, "n_formula_1": 8, "n_ref_uni": 37, "n_ref": 59, "n_ref_all": 121, "n_fig": 15, "n_tab": 8, "L_tab": 1637, "n_element_tab": 221, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 376, "n_element_tab_1": 34, "formula_len_all": 5881, "formula_len_all_1": 493, "len_all": 146858, "len_all_1": 60592, "len_abs": 1655, "len_title": 106, "len_sents": 35011, "len_sents_1": 24548, "n_sents": 410, "n_sents_1": 256, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1721, "title": "Towards Accurate and Efficient Sub-8-Bit Integer Training", "abs": "Neural network training is a memory- and compute-intensive task. Quantization, which enables low-bitwidth formats in training, can significantly mitigate the workload. To reduce quantization error, recent methods have developed new data formats and additional pre-processing operations on quantizers. However, it remains quite challenging to achieve high accuracy and efficiency simultaneously. In this paper, we explore sub-8-bit integer training from its essence of gradient descent optimization. Our integer training framework includes two components: ShiftQuant to realize accurate gradient estimation, and L1 normalization to smoothen the loss landscape. \n ShiftQuant attains performance that approaches the theoretical upper bound of group quantization. Furthermore, it liberates group quantization from inefficient memory rearrangement. The L1 normalization facilitates the implementation of fully quantized normalization layers with impressive convergence accuracy. \n Our method frees sub-8-bit integer training from pre-processing and supports general devices. \n This framework achieves negligible accuracy loss across various neural networks and tasks ($0.92\\%$ on 4-bit ResNets, $0.61\\%$ on 6-bit Transformers, $0.61\\%$ on 6-bit GNNs). \n The prototypical implementation of ShiftQuant achieves more than $1.85\\times/15.3\\%$ performance improvement on CPU/GPU compared to its FP16 counterparts, and $33.9\\%$ resource consumption reduction on FPGA than the FP16 counterparts. The proposed fully-quantized L1 normalization layers achieve more than $35.54\\%$ improvement in throughout on CPU compared to traditional L2 normalization layers. Moreover, theoretical analysis verifies the advancement of our method.", "keywords": [ "Low-precision training; Model compression" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "wI5uHZLeCZ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 11, "author_info_list": [ { "name": "Abhay Sheshadri", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Aengus Lynch", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Aidan Ewart", "gender": "unknown", "institution": "University of Bristol", "country": "", "position": "Undergrad student" }, { "name": "Asa Cooper Stickland", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "Cindy Wu", "gender": "Female", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Dylan Hadfield-Menell", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Ethan Perez", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Henry Sleight", "gender": "Male", "institution": "Constellation", "country": "US", "position": "Researcher" }, { "name": "Phillip Huang Guo", "gender": "Male", "institution": "Gray Swan AI", "country": "US", "position": "Researcher" }, { "name": "Stephen Casper", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Graduate Student" }, { "name": "Vivek Hebbar", "gender": "unknown", "institution": "Constellation", "country": "", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 4, "n_ref_uni": 95, "n_ref": 232, "n_ref_all": 258, "n_fig": 6, "n_tab": 5, "L_tab": 2253, "n_element_tab": 119, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1567, "n_element_tab_1": 81, "formula_len_all": 1316, "formula_len_all_1": 159, "len_all": 257045, "len_all_1": 59613, "len_abs": 1582, "len_title": 136, "len_sents": 53595, "len_sents_1": 27811, "n_sents": 419, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1589, "title": "Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMs", "abs": "Large language models (LLMs) can often be made to behave in undesirable ways that they are explicitly fine-tuned not to. For example, the LLM red-teaming literature has produced a wide variety of 'jailbreaking' techniques to elicit harmful text from models that were fine-tuned to be harmless. Recent work on red-teaming, model editing, and interpretability suggests that this challenge stems from how (adversarial) fine-tuning largely serves to suppress rather than remove undesirable capabilities from LLMs. Prior work has introduced latent adversarial training (LAT) as a way to improve robustness to broad classes of failures. These prior works have considered untargeted latent space attacks where the adversary perturbs latent activations to maximize loss on examples of desirable behavior. Untargeted LAT can provide a generic type of robustness but does not leverage information about specific failure modes. Here, we experiment with targeted LAT where the adversary seeks to minimize loss on a specific competing task. We find that it can augment a wide variety of state-of-the-art methods. First, we use targeted LAT to improve robustness to jailbreaks, outperforming a strong R2D2 baseline with orders of magnitude less compute. Second, we use it to more effectively remove backdoors with no knowledge of the trigger. Finally, we use it to more effectively unlearn knowledge for specific undesirable tasks in a way that is also more robust to re-learning. Overall, our results suggest that targeted LAT can be an effective tool for defending against harmful behaviors from LLMs.", "keywords": [ "adversarial attacks", "adversarial training", "jailbreaks", "trojans", "backdoors", "unlearning", "robustenss" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "wHsAi8kINK", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Haotian Zheng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Huanhuan Chen", "gender": "Male", "institution": "University of Science and Technology of China, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jinlong Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Qiuju Chen", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 20, "n_ref": 39, "n_ref_all": 63, "n_fig": 5, "n_tab": 2, "L_tab": 882, "n_element_tab": 187, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 1419, "n_element_tab_1": 39, "formula_len_all": 607, "formula_len_all_1": 607, "len_all": 103041, "len_all_1": 59852, "len_abs": 1046, "len_title": 123, "len_sents": 31625, "len_sents_1": 27589, "n_sents": 198, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1050, "title": "Fed3+2p: Training different parts of neural network with two-phase strategy", "abs": "In federated learning, the non-identically distributed data affects both global and local performance, while clients with small data volumes may also suffer from overfitting issues. To address these challenges, we propose a federated learning framework called Fed3+2p. In Fed3+2p, we divide the client neural network into three parts: a feature extractor, a filter, and classification heads, and to train these parts, we present two types of coordinators to train client sets with a two-phase training strategy. In the first phase, each Type-A coordinator trains the feature extractor of partial clients, whose joint data distribution is similar to the global data distribution. In the second phase, each Type-B coordinator trains the filter and classification heads of partial clients, whose data distributions are similar to each other. We conduct empirical studies on three datasets: FMNIST and CIFAR-10/100, and the results show that Fed3+2p surpasses the state-of-the-art methods in both global and local performance across all tested datasets.", "keywords": [ "Federated Learning", "Distributed Machine Learning", "Neural Networks", "Non-IID Data" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 1 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "wHebuIb6IH", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Armando Solar-Lezama", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Beichen Li", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Bernd Bickel", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Full Professor" }, { "name": "Changxi Zheng", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Associate Professor" }, { "name": "Liang Shi", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Rundi Wu", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Wojciech Matusik", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 58, "n_ref": 132, "n_ref_all": 154, "n_fig": 8, "n_tab": 5, "L_tab": 2669, "n_element_tab": 219, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1048, "n_element_tab_1": 150, "formula_len_all": 170, "formula_len_all_1": 168, "len_all": 158633, "len_all_1": 55879, "len_abs": 1038, "len_title": 125, "len_sents": 38364, "len_sents_1": 25692, "n_sents": 277, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 76, "L_abs": 896, "title": "VLMaterial: Procedural Material Generation with Large Vision-Language Models", "abs": "Procedural materials, represented as functional node graphs, are ubiquitous in computer graphics for photorealistic material appearance design. They allow users to perform intuitive and precise editing to achieve desired visual appearances. However, creating a procedural material given an input image requires professional knowledge and significant effort. In this work, we leverage the ability to convert procedural materials into standard Python programs and fine-tune a large pre-trained vision-language model (VLM) to generate such programs from input images. To enable effective fine-tuning, we also contribute an open-source procedural material dataset and propose to perform program-level augmentation by prompting another pre-trained large language model (LLM). Through extensive evaluation, we show that our method outperforms previous methods on both synthetic and real-world examples.", "keywords": [ "generative model", "procedural material", "appearance modeling" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "wHLMsM1SrP", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Jonathan Roberts", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Kai Han", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "samuel Albanie", "gender": "Not Specified", "institution": "Google", "country": "GB", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 19, "n_ref": 115, "n_ref_all": 137, "n_fig": 14, "n_tab": 4, "L_tab": 2094, "n_element_tab": 250, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 142, "n_element_tab_1": 8, "formula_len_all": 336, "formula_len_all_1": 386, "len_all": 118811, "len_all_1": 57218, "len_abs": 1450, "len_title": 127, "len_sents": 35704, "len_sents_1": 25645, "n_sents": 289, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1411, "title": "Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?", "abs": "As the context limits of Large Language Models (LLMs) increase, the range of\npossible applications and downstream functions broadens. In many real-world\ntasks, decisions depend on details scattered across collections of often disparate\ndocuments containing mostly irrelevant information. Long-context LLMs appear\nwell-suited to this form of complex information retrieval and reasoning, which has\ntraditionally proven costly and time-consuming. However, although the development of longer context models has seen rapid gains in recent years, our understanding of how effectively LLMs use their context has not kept pace. To address\nthis, we conduct a set of retrieval experiments designed to evaluate the capabilities\nof 17 leading LLMs, such as their ability to follow threads of information through\nthe context window. Strikingly, we find that many models are remarkably thread-\nsafe: capable of simultaneously following multiple threads without significant loss\nin performance. Still, for many models, we find the effective context limit is significantly shorter than the supported context length, with accuracy decreasing as\nthe context window grows. Our study also highlights the important point that token counts from different tokenizers should not be directly compared\u2014they often\ncorrespond to substantially different numbers of written characters. We release\nour code and long context experimental data.", "keywords": [ "LLMs", "Long Context", "Evaluation" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "wH8XXUOUZU", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Enze Xie", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Researcher" }, { "name": "Han Cai", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Haotian Tang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Junyu Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Muyang Li", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Shang Yang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Song Han", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "\u4fca\u677e \u9648", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 44, "n_ref": 94, "n_ref_all": 124, "n_fig": 14, "n_tab": 21, "L_tab": 3773, "n_element_tab": 770, "n_fig_1": 6, "n_tab_1": 11, "L_tab_1": 865, "n_element_tab_1": 258, "formula_len_all": 417, "formula_len_all_1": 298, "len_all": 136812, "len_all_1": 47236, "len_abs": 1300, "len_title": 123, "len_sents": 27900, "len_sents_1": 19410, "n_sents": 229, "n_sents_1": 136, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1303, "title": "Deep Compression Autoencoder for Efficient High-Resolution Diffusion Models", "abs": "We present Deep Compression Autoencoder (DC-AE), a new family of autoencoder models for accelerating high-resolution diffusion models. Existing autoencoder models have demonstrated impressive results at a moderate spatial compression ratio (e.g., 8x), but fail to maintain satisfactory reconstruction accuracy for high spatial compression ratios (e.g., 64x). We address this challenge by introducing two key techniques: (1) Residual Autoencoding, where we design our models to learn residuals based on the space-to-channel transformed features to alleviate the optimization difficulty of high spatial-compression autoencoders; (2) Decoupled High-Resolution Adaptation, an efficient decoupled three-phases training strategy for mitigating the generalization penalty of high spatial-compression autoencoders. With these designs, we improve the autoencoder's spatial compression ratio up to 128 while maintaining the reconstruction quality. Applying our DC-AE to latent diffusion models, we achieve significant speedup without accuracy drop. For example, on ImageNet 512x512, our DC-AE provides 19.1x inference speedup and 17.9x training speedup on H100 GPU for UViT-H while achieving a better FID, compared with the widely used SD-VAE-f8 autoencoder. Our code and models will be released upon publication.", "keywords": [ "efficient high-resolution diffusion models" ], "rating_list": [ 6, 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 4, 3, 2 ], "presentation_list": [ 2, 3, 4, 4, 2 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 2, 4 ] }, { "paper_id": "wGqf7YMF8R", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Dong Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Distinguished Scientist" }, { "name": "Haitao Mi", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Principal Researcher" }, { "name": "Wenlin Yao", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 21, "n_ref": 42, "n_ref_all": 62, "n_fig": 5, "n_tab": 11, "L_tab": 3690, "n_element_tab": 268, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2704, "n_element_tab_1": 229, "formula_len_all": 219, "formula_len_all_1": 0, "len_all": 160656, "len_all_1": 71196, "len_abs": 1797, "len_title": 136, "len_sents": 57974, "len_sents_1": 32100, "n_sents": 447, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1638, "title": "HDFlow: Enhancing LLM Complex Problem-Solving with Hybrid Thinking and Dynamic Workflows", "abs": "Despite recent advancements in large language models (LLMs), their performance on complex reasoning problems requiring multi-step thinking and combining various skills is still limited. To address this, we propose a novel framework HDFlow for complex reasoning with LLMs that combines fast and slow thinking modes in an adaptive manner. Our approach consists of two key components: 1) a new approach for slow, deliberate reasoning called Dynamic Workflow, which automatically decomposes complex problems into more manageable sub-tasks and dynamically designs a workflow to assemble specialized LLM or symbolic reasoning tools to solve sub-tasks; 2) Hybrid Thinking, a general framework that dynamically combines fast and slow thinking based on problem complexity. \nFinally, we propose an easy-to-scale method for automatically synthesizing a large-scale dataset of 27K challenging reasoning problems for complex reasoning and a hybrid thinking tuning method that trains smaller LLMs on this dataset to internalize the fast/slow hybrid reasoning strategies.\nExperiments on four reasoning benchmark datasets demonstrate that our slow thinking with dynamic workflows significantly outperforms Chain-of-Thought, and hybrid thinking achieves the highest accuracy while providing an effective balance between computational efficiency and performance. Fine-tuning using our hybrid thinking approach also significantly boosts the complex reasoning capabilities of open-source language models. The results showcase the promise of slow thinking, dynamic workflows, and hybrid thinking in expanding the frontier of complex problem-solving with LLMs.", "keywords": [ "Large Language Models (LLMs)", "Complex Reasoning", "Hybrid Thinking", "Symbolic Reasoning" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "wGa2plE8ka", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Junshu Sun", "gender": "unknown", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Shuhui Wang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yaowei Wang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Yiling Wu", "gender": "Female", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Postdoc" }, { "name": "Yue WU", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zhaobo Qi", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Lecturer" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 30, "n_ref": 88, "n_ref_all": 122, "n_fig": 14, "n_tab": 15, "L_tab": 3530, "n_element_tab": 306, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 526, "n_element_tab_1": 85, "formula_len_all": 684, "formula_len_all_1": 496, "len_all": 169939, "len_all_1": 64308, "len_abs": 1898, "len_title": 151, "len_sents": 55946, "len_sents_1": 29547, "n_sents": 455, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 103, "L_abs": 1910, "title": "Learning Fine-Grained Representations through Textual Token Disentanglement in Composed Video Retrieval", "abs": "With the explosive growth of video data, finding videos that meet detailed requirements in large datasets has become a challenge. To address this, the composed video retrieval task has been introduced, enabling users to retrieve videos using complex queries that involve both visual and textual information. However, the inherent heterogeneity between modalities poses significant challenges. Textual data is highly abstract, while video content contains substantial redundancy. This modality gap in information representation makes existing methods struggle with the fine-grained fusion and alignment required for fine-grained composed retrieval. To overcome these challenges, we first introduce FineCVR-1M, a fine-grained composed video retrieval dataset containing 1,010,071 video-text triplets with detailed textual descriptions. This dataset is constructed through an automated process that identifies key concept changes between video pairs to generate textual descriptions for both static and action concepts. For fine-grained retrieval methods, the key challenge lies in understanding the detailed requirements. Text descriptions serve as clear expressions of intent, allowing models to distinguish fine-grained needs through textual feature disentanglement. Therefore, we propose a textual Feature Disentanglement and Cross-modal Alignment framework (FDCA) that disentangles features at both the sentence and token levels. At the sequence level, we separate text features into retained and injected features. At the token level, an Auxiliary Token Disentangling mechanism is proposed to disentangle texts into retained, injected, and excluded tokens. The disentanglement at both levels extracts fine-grained features, which are aligned and fused with reference video to extract global representations for video retrieval. Experiments on FineCVR-1M dataset demonstrate the superior performance of FDCA.", "keywords": [ "Composed Video Retrieval; Fine-grained Representation; Feature Disentanglement" ], "rating_list": [ 5, 8, 8, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "wGVOxplEbf", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Hongrui Huang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Jiangning Zhang", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Lizhuang Ma", "gender": "Male", "institution": "Dept. of Computer Sci. & Eng., Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Ran Yi", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Teng Hu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yabiao Wang", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Senior Researcher" } ], "n_formula": 19, "n_formula_1": 9, "n_ref_uni": 42, "n_ref": 97, "n_ref_all": 134, "n_fig": 24, "n_tab": 10, "L_tab": 7240, "n_element_tab": 991, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 4247, "n_element_tab_1": 524, "formula_len_all": 1364, "formula_len_all_1": 499, "len_all": 223698, "len_all_1": 74858, "len_abs": 2117, "len_title": 134, "len_sents": 74305, "len_sents_1": 32870, "n_sents": 474, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1609, "title": "SaRA: High-Efficient Diffusion Model Fine-tuning with Progressive Sparse Low-Rank Adaptation", "abs": "The development of diffusion models has led to significant progress in image and video generation tasks, with pre-trained models like the Stable Diffusion series playing a crucial role.\nHowever, a key challenge remains in downstream task applications: how to effectively and efficiently adapt pre-trained diffusion models to new tasks.\nInspired by model pruning which lightens large pre-trained models by removing unimportant parameters, we propose a novel model fine-tuning method to make full use of these ineffective parameters and enable the pre-trained model with new task-specified capabilities.\nIn this work, we first investigate the importance of parameters in pre-trained diffusion models and discover that parameters with the smallest absolute values do not contribute to the generation process due to training instabilities.\nBased on this observation, we propose a fine-tuning method termed SaRA that re-utilizes these temporarily ineffective parameters, equating to optimizing a sparse weight matrix to learn the task-specific knowledge.\nTo mitigate potential overfitting, we propose a nuclear-norm-based low-rank sparse training scheme for efficient fine-tuning.\nFurthermore, we design a new progressive parameter adjustment strategy to make full use of the finetuned parameters.\nFinally, we propose a novel unstructural backpropagation strategy, which significantly reduces memory costs during fine-tuning.\nOur method enhances the generative capabilities of pre-trained models in downstream applications and outperforms existing fine-tuning methods in maintaining model's generalization ability.", "keywords": [ "Diffusion Model", "Fine-tuning" ], "rating_list": [ 6, 6, 8, 5, 6 ], "soundness_list": [ 2, 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 4, 3, 3 ], "confidence_list": [ 2, 3, 5, 4, 3 ] }, { "paper_id": "wFs2E5wCw6", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Hanspeter Pfister", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Tong Ding", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Wanhua Li", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Zhongqi Miao", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 48, "n_ref": 136, "n_ref_all": 153, "n_fig": 4, "n_tab": 9, "L_tab": 3720, "n_element_tab": 350, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 5720, "n_element_tab_1": 292, "formula_len_all": 1451, "formula_len_all_1": 1319, "len_all": 166821, "len_all_1": 71480, "len_abs": 1435, "len_title": 109, "len_sents": 40803, "len_sents_1": 29381, "n_sents": 285, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1442, "title": "Tree of Attributes Prompt Learning for Vision-Language Models", "abs": "Prompt learning has proven effective in adapting vision language models for downstream tasks. However, existing methods usually append learnable prompt tokens solely with the category names to obtain textual features, which fails to fully leverage the rich context indicated in the category name. To address this issue, we propose the Tree of Attributes Prompt learning (TAP), which first instructs LLMs to generate a tree of attributes with a \u201cconcept - attribute - description\u201d structure for each category, and then learn the hierarchy with vision and text prompt tokens. Unlike existing methods that merely augment category names with a set of unstructured descriptions, our approach essentially distills structured knowledge graphs associated with class names from LLMs. Furthermore, our approach introduces text and vision prompts designed to explicitly learn the corresponding visual attributes, effectively serving as domain experts. Additionally, the general and diverse descriptions generated based on the class names may be wrong or absent in the specific given images. To address this misalignment, we further introduce a vision-conditional pooling module to extract instance-specific text features. Extensive experimental results demonstrate that our approach outperforms state-of-the-art methods on the zero-shot base-to-novel generalization, cross-dataset transfer, as well as few-shot classification across 11 diverse datasets.", "keywords": [ "Prompt Learning", "Vision-Language Models", "Tree of Attributes" ], "rating_list": [ 8, 6, 6, 6, 6 ], "soundness_list": [ 4, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 1, 3 ] }, { "paper_id": "wFg0shwoRe", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Darius Muglich", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Elise van der Pol", "gender": "Female", "institution": "Microsoft Research", "country": "NL", "position": "Researcher" }, { "name": "Jakob Nicolaus Foerster", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Johannes Forkel", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Researcher" } ], "n_formula": 41, "n_formula_1": 23, "n_ref_uni": 35, "n_ref": 69, "n_ref_all": 102, "n_fig": 2, "n_tab": 2, "L_tab": 586, "n_element_tab": 33, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3915, "formula_len_all_1": 1793, "len_all": 158633, "len_all_1": 78567, "len_abs": 1371, "len_title": 74, "len_sents": 51699, "len_sents_1": 35053, "n_sents": 396, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 26, "L_abs": 1375, "title": "Expected Return Symmetries", "abs": "Symmetry is an important inductive bias that can improve model robustness and generalization across many deep learning domains. In multi-agent settings, a priori known symmetries have been shown to address a fundamental coordination failure mode known as mutually incompatible symmetry breaking; e.g. in a game where two independent agents can choose to move \"left\" or \"right\", and where a reward of +1 or -1 is received when the agents choose the same action or different actions, respectively. However, the efficient and automatic discovery of environment symmetries, in particular for decentralized partially observable Markov decision processes, remains an open problem. Furthermore, environmental symmetry breaking constitutes only one type of coordination failure, which motivates the search for a more accessible and broader symmetry class. In this paper, we introduce such a broader group of previously unexplored symmetries, which we call expected return symmetries, which contains environment symmetries as a subgroup. We show that agents trained to be compatible under the group of expected return symmetries achieve better zero-shot coordination results than those using environment symmetries. As an additional benefit, our method makes minimal a priori assumptions about the structure of their environment and does not require access to ground truth symmetries.", "keywords": [ "multi-agent reinforcement learning", "zero-shot coordination" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 1 ] }, { "paper_id": "wFIf8zpzTI", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Daniel Kudenko", "gender": "Male", "institution": "L3S Research Center", "country": "DE", "position": "Research Group Leader" }, { "name": "Monty Maximilian Z\u00fchlke", "gender": "Male", "institution": "Leibniz Universit\u00e4t Hannover", "country": "DE", "position": "PhD student" }, { "name": "Wolfgang Nejdl", "gender": "Male", "institution": "L3S Research Center", "country": "DE", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 2, "n_ref_uni": 43, "n_ref": 135, "n_ref_all": 176, "n_fig": 9, "n_tab": 24, "L_tab": 23486, "n_element_tab": 3621, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 922, "n_element_tab_1": 64, "formula_len_all": 853, "formula_len_all_1": 100, "len_all": 400348, "len_all_1": 71778, "len_abs": 1536, "len_title": 77, "len_sents": 126847, "len_sents_1": 36706, "n_sents": 985, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1292, "title": "Out-Of-Context and Out-Of-Scope: Subliminal Priming for Large Language Models", "abs": "We mimic human subliminal priming studies for large language models (LLMs) by fine-tuning models with a few short ex-template descriptions of a fictitious character's behaviour mixed into a large corpus of longer but unrelated in-template instructions and eliciting demonstrations of the behaviour using suitable trigger prompts. Our theoretical motivation comes from observing that optimising models with the standard per-token cross-entropy loss is equivalent to training on a weighted context classification task, where shorter contexts have a higher weight. While we cannot measure an LLM's unawareness of the descriptions, we show that prompting strategies motivated by projective psychology and psychoanalytic theory succeed where naive questions fail, even with potent chain-of-thought (COT) initiators. This work extends research on out-of-context reasoning (OOCR), a primer for situational awareness, where LLMs \"read between the lines\" or \"think outside of the box\" by performing reasoning hops on internalised knowledge. We show that simple manipulations of the training data allow and improve the embedding of specific response behaviour, which may only be triggered using the correct prompting strategy, hinting at the possibility of undetected alignment hazards in current LLMs.", "keywords": [ "representation learning", "generative models", "learning theory", "applications to neuroscience & cognitive science" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "wFD16gwpze", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Bernd Rosenow", "gender": "Male", "institution": "Universit\u00e4t Leipzig", "country": "", "position": "Full Professor" }, { "name": "Roman Worschech", "gender": "unknown", "institution": "Max Planck Institute for Mathematics in the Sciences, Max-Planck Institute", "country": "DE", "position": "PhD student" } ], "n_formula": 244, "n_formula_1": 31, "n_ref_uni": 26, "n_ref": 48, "n_ref_all": 184, "n_fig": 27, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18996, "formula_len_all_1": 1716, "len_all": 292983, "len_all_1": 69488, "len_abs": 1804, "len_title": 127, "len_sents": 96820, "len_sents_1": 27928, "n_sents": 946, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 79, "L_abs": 1701, "title": "Analyzing Neural Scaling Laws in Two-Layer Networks with Power-Law Data Spectra", "abs": "Neural scaling laws describe how the performance of deep neural networks scales with key factors such as training data size, model complexity, and training time, often following power-law behaviors over multiple orders of magnitude. Despite their empirical observation, the theoretical understanding of these scaling laws remains limited. In this work, we employ techniques from statistical mechanics to analyze one-pass stochastic gradient descent within a student-teacher framework, where both the student and teacher are two-layer neural networks. Our study primarily focuses on the generalization error and its behavior in response to data covariance matrices that exhibit power-law spectra.\nFor linear activation functions, we derive analytical expressions for the generalization error, exploring different learning regimes and identifying conditions under which power-law scaling emerges. Additionally, we extend our analysis to non-linear activation functions in the feature learning regime, investigating how power-law spectra in the data covariance matrix impact learning dynamics. Importantly, we find that the length of the symmetric plateau depends on the number of distinct eigenvalues of the data covariance matrix and the number of hidden units, demonstrating how these plateaus behave under various configurations. In addition, our results reveal a transition from exponential to power-law convergence in the specialized phase when the data covariance matrix possesses a power-law spectrum. This work contributes to the theoretical understanding of neural scaling laws and provides insights into optimizing learning performance in practical scenarios involving complex data structures.", "keywords": [ "Statistical mechanics", "neural scaling laws" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 4, 4 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "wFAyp2CUnq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Jiajun Wu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Juan Carlos Niebles", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Director, Research" }, { "name": "Junxian He", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Manling Li", "gender": "Female", "institution": "Northwestern University", "country": "US", "position": "Assistant Professor" }, { "name": "Mor Geva", "gender": "Female", "institution": "Tel Aviv University", "country": "IL", "position": "Assistant Professor" }, { "name": "Ruochen Zhou", "gender": "Female", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shiqi Chen", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Intern" }, { "name": "Siyang Gao", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "TONGYAO ZHU", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 19, "n_ref": 51, "n_ref_all": 88, "n_fig": 12, "n_tab": 6, "L_tab": 1516, "n_element_tab": 194, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1915, "n_element_tab_1": 260, "formula_len_all": 580, "formula_len_all_1": 585, "len_all": 126163, "len_all_1": 64350, "len_abs": 1697, "len_title": 81, "len_sents": 43924, "len_sents_1": 28792, "n_sents": 308, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1135, "title": "AdaptVis: Spatial Understanding in Vision-Language Models Requires Adaptive Attention", "abs": "Vision Large Language Models (VLLMs) often struggle with adequately attending to image information, leading to significant hallucinations across various domains, especially on spatial reasoning tasks. In this study, we analyze the attention behavior of VLLMs in spatial reasoning Question-Answering (QA) tasks from a mechanism interpretability view. By visualizing the crucial areas of an image that receive the highest attention scores in the intermediate layers, we identify an interesting pattern: failures often correspond to attention being misallocated to irrelevant objects within the image. Moreover, the attention patterns exhibit large differences between familiar and unfamiliar spatial relationships. Motivated by this observation, we further explore the feasibility of adaptively adjusting the attention scores during the inference process based on the confidence score. Our experiments on spatial reasoning benchmarks including WhatsUp and VSR demonstrate that our decoding methods yield promising results, e.g., achieving up to a 50-point improvement on the WhatsUp benchmark with negligible additional computation cost.", "keywords": [ "Vision Language Models", "Uncertainty", "Mechanistic interpretability", "Constrain Decoding", "Spatial Understanding" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 4, 1, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "wF9Cz2PknU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Di Chang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Fang Li", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Hao Zhang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Mohammad Soleymani", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Research associate professor" }, { "name": "Narendra Ahuja", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Research Professor" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 16, "n_ref": 58, "n_ref_all": 76, "n_fig": 8, "n_tab": 3, "L_tab": 1691, "n_element_tab": 236, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 956, "n_element_tab_1": 150, "formula_len_all": 459, "formula_len_all_1": 457, "len_all": 148198, "len_all_1": 63259, "len_abs": 1792, "len_title": 123, "len_sents": 44339, "len_sents_1": 29118, "n_sents": 328, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1754, "title": "MagicPose4D: Crafting Articulated Models with Appearance and Motion Control", "abs": "With the success of 2D and 3D visual generative models, there is growing interest in generating 4D content. Existing methods primarily rely on text prompts to produce 4D content, but they often fall short of accurately defining complex or rare motions. To address this limitation, we propose MagicPose4D, a novel framework for refined control over both appearance and motion in 4D generation. Unlike traditional methods, MagicPose4D accepts monocular videos as motion prompts, enabling precise and customizable motion generation. MagicPose4D comprises two key modules:\n\ni) Dual-Phase 4D Reconstruction Module} which operates in two phases. The first phase focuses on capturing the model's shape using accurate 2D supervision and less accurate but geometrically informative 3D pseudo-supervision without imposing skeleton constraints. The second phase refines the model using more accurate pseudo-3D supervision, obtained in the first phase and introduces kinematic chain-based skeleton constraints to ensure physical plausibility. Additionally, we propose a Global-local Chamfer loss that aligns the overall distribution of predicted mesh vertices with the supervision while maintaining part-level alignment without extra annotations.\n\nii) Cross-category Motion Transfer Module} leverages the predictions from the 4D reconstruction module and uses a kinematic-chain-based skeleton to achieve cross-category motion transfer. It ensures smooth transitions between frames through dynamic rigidity, facilitating robust generalization without additional training.\n\nThrough extensive experiments, we demonstrate that MagicPose4D significantly improves the accuracy and consistency of 4D content generation, outperforming existing methods in various benchmarks.", "keywords": [ "4D Generation; 3D Reconstruction; Motion Transfer; Animation; Rigging" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "wElgE9qBb5", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Anton Frederik Thielmann", "gender": "Male", "institution": "BASF", "country": "DE", "position": "Postdoc" }, { "name": "Arik Reuter", "gender": "Male", "institution": "Technische Universit\u00e4t Clausthal", "country": "DE", "position": "Researcher" }, { "name": "Benjamin S\u00e4fken", "gender": "Male", "institution": "Technische Universit\u00e4t Clausthal", "country": "", "position": "Full Professor" }, { "name": "Christoph Weisser", "gender": "Male", "institution": "BASF", "country": "DE", "position": "Researcher" }, { "name": "Manish Kumar", "gender": "Male", "institution": "BASF", "country": "", "position": "Researcher" }, { "name": "Soheila Samiee", "gender": "Female", "institution": "BASF", "country": "CA", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 42, "n_ref": 94, "n_ref_all": 113, "n_fig": 3, "n_tab": 15, "L_tab": 4094, "n_element_tab": 345, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1997, "n_element_tab_1": 129, "formula_len_all": 786, "formula_len_all_1": 659, "len_all": 131164, "len_all_1": 60661, "len_abs": 1145, "len_title": 102, "len_sents": 42079, "len_sents_1": 26945, "n_sents": 324, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1153, "title": "Mambular: A Sequential Model for Tabular Deep Learning", "abs": "The analysis of tabular data has traditionally been dominated by gradient-boosted decision trees (GBDTs), known for their proficiency with mixed categorical and numerical features. However, recent deep learning innovations are challenging this dominance. We introduce Mambular, an adaptation of the Mamba architecture optimized for tabular data. We extensively benchmark Mambular against state-of-the-art models, including neural networks and tree-based methods, and demonstrate its competitive performance across diverse datasets.\nAdditionally, we explore various adaptations of Mambular to understand its effectiveness for tabular data. We investigate different pooling strategies, feature interaction mechanisms, and bi-directional processing. Our analysis shows that interpreting features as a sequence and passing them through Mamba layers results in surprisingly performant models. The results highlight Mambular\u2019s potential as a versatile and powerful architecture for tabular data analysis, expanding the scope of deep learning applications in this domain.\n The source code is available at: https://anonymous.4open.science/r/mamba-tabular-485F/", "keywords": [ "Tabular Deep Learning", "Mamba", "Sequential Models", "SSM", "Recurrent Neural Networks" ], "rating_list": [ 3, 8, 3, 3 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 4, 5, 3, 5 ] }, { "paper_id": "wE8wJXgI9T", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Abrar Fahim", "gender": "Male", "institution": ", University of Alberta", "country": "CA", "position": "MS student" }, { "name": "Alona Fyshe", "gender": "Female", "institution": "University of Alberta", "country": "CA", "position": "Associate Professor" }, { "name": "Alex Graeme Murphy", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 25, "n_ref": 62, "n_ref_all": 92, "n_fig": 12, "n_tab": 3, "L_tab": 528, "n_element_tab": 63, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 114, "n_element_tab_1": 11, "formula_len_all": 711, "formula_len_all_1": 643, "len_all": 125897, "len_all_1": 61923, "len_abs": 1528, "len_title": 122, "len_sents": 43254, "len_sents_1": 29686, "n_sents": 317, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1496, "title": "It's Not a Modality Gap: Characterizing and Addressing the Contrastive Gap", "abs": "Learning jointly from images and texts using contrastive pre-training has emerged as an effective method to train large-scale models with a strong grasp of semantic image concepts. For instance, CLIP, pre-trained on a large corpus of web data, excels in tasks like zero-shot image classification, object detection, geolocalization, and more. These contrastive models embed input images and texts into a shared representational space. Recently, it was claimed that models like CLIP show a *modality gap*, where image and text embeddings occupy disjoint areas in the representational space. Previous studies attribute this gap to factors like data artifacts (mismatched pairs), model architecture artifacts (the cone effect), and the nature of the loss landscape (getting stuck in local minima). We demonstrate that, even after accounting for these factors, and even when using the *same modality*, the contrastive loss actually *creates* a gap during training. As a result, we propose renaming this phenomenon the *contrastive gap*. We show that the contrastive gap is exacerbated by training with small batch sizes in high-dimensional spaces, causing embeddings of each modality to occupy small disjoint portions of the latent space. Our experiments show that minimizing the contrastive gap via the addition of uniformity and alignment terms optimizes the representational space and conveys better performance on downstream tasks such as zero-shot image classification and multi-modal arithmetic.", "keywords": [ "Multi Modal Representation Learning", "Contrastive Representation Learning", "Modality Gap", "CLIP" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "wE5xp3zBaQ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Berkant Turan", "gender": "Male", "institution": "Zuse Institute Berlin", "country": "DE", "position": "PhD student" }, { "name": "Grzegorz Gluch", "gender": "unknown", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "", "position": "PhD student" }, { "name": "Sai Ganesh Nagarajan", "gender": "unknown", "institution": "Zuse Institute Berlin", "country": "", "position": "Postdoc" }, { "name": "Sebastian Pokutta", "gender": "Male", "institution": "ZIB", "country": "DE", "position": "Vice President" } ], "n_formula": 49, "n_formula_1": 7, "n_ref_uni": 45, "n_ref": 89, "n_ref_all": 109, "n_fig": 3, "n_tab": 7, "L_tab": 1436, "n_element_tab": 56, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4540, "formula_len_all_1": 407, "len_all": 284172, "len_all_1": 70129, "len_abs": 2045, "len_title": 138, "len_sents": 87581, "len_sents_1": 32493, "n_sents": 789, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1287, "title": "The Good, the Bad and the Ugly: Watermarks, Transferable Attacks and Adversarial Defenses", "abs": "We formalize and extend existing definitions of backdoor-based watermarks and adversarial defenses as *interactive protocols* between two players. The existence of these schemes is inherently tied to the learning tasks for which they are designed. Our main result shows that for *almost every* discriminative learning task, at least one of the two \u2014 a watermark or an adversarial defense \u2014 exists. The \"*almost*\" refers to the fact that we also identify a third, counterintuitive but necessary option, i.e., a scheme we call a *transferable attack*. By transferable attack, we refer to an efficient algorithm computing queries that look indistinguishable from the data distribution and fool *all* efficient defenders.\n\nTo this end, we prove the necessity of a transferable attack via a construction that uses a cryptographic tool called homomorphic encryption. Furthermore, we show that any task that satisfies our notion of a transferable attack implies a *cryptographic primitive*, thus requiring the underlying task to be computationally complex. These two facts imply an \"*equivalence*\" between the existence of transferable attacks and cryptography. Finally, we show that the class of tasks of bounded VC-dimension has an adversarial defense, and a subclass of them has a watermark.", "keywords": [ "Watermarks", "Adversarial Defenses", "Transferable Attacks", "Interactive Proof Systems", "Cryptography", "Backdooring", "Game Theory", "Learning Theory" ], "rating_list": [ 6, 6, 5, 3 ], "soundness_list": [ 3, 2, 3, 1 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "wDcunIOAOk", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Jibril Frej", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "Postdoc" }, { "name": "Julian Blackwell", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "MS student" }, { "name": "Martin Jaggi", "gender": "Male", "institution": "EPFL", "country": "CH", "position": "Associate Professor" }, { "name": "Syrielle Montariol", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Postdoc" }, { "name": "Tanja K\u00e4ser", "gender": "Female", "institution": "School of Computer and Communication Sciences, EPFL - EPF Lausanne", "country": "", "position": "Assistant Professor" }, { "name": "Vinitra Swamy", "gender": "Female", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 63, "n_ref": 99, "n_ref_all": 128, "n_fig": 5, "n_tab": 5, "L_tab": 1487, "n_element_tab": 77, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2066, "n_element_tab_1": 128, "formula_len_all": 62, "formula_len_all_1": 62, "len_all": 158131, "len_all_1": 74017, "len_abs": 1640, "len_title": 121, "len_sents": 42980, "len_sents_1": 36327, "n_sents": 279, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1645, "title": "Intrinsic User-Centric Interpretability through Global Mixture of Experts", "abs": "In human-centric settings like education or healthcare, model accuracy and model explainability are key factors for user adoption. Towards these two goals, intrinsically interpretable deep learning models have gained popularity, focusing on accurate predictions alongside faithful explanations. However, there exists a gap in the human-centeredness of these approaches, which often produce nuanced and complex explanations that are not easily actionable for downstream users. We present InterpretCC (interpretable conditional computation), a family of intrinsically interpretable neural networks at a unique point in the design space that optimizes for ease of human understanding and explanation faithfulness, while maintaining comparable performance to state-of-the-art models. InterpretCC achieves this through adaptive sparse activation of features before prediction, allowing the model to use a different, minimal set of features for each instance. We extend this idea into an interpretable, global mixture-of-experts (MoE) model that allows users to specify topics of interest, discretely separates the feature space for each data point into topical subnetworks, and adaptively and sparsely activates these topical subnetworks for prediction. We apply InterpretCC for text, time series and tabular data across several real-world datasets, demonstrating comparable performance with non-interpretable baselines and outperforming intrinsically interpretable baselines. Through a user study involving 56 teachers, InterpretCC explanations are found to have higher actionability and usefulness over other intrinsically interpretable approaches.", "keywords": [ "interpretability", "human-centric computing", "mixture-of-experts" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "wD2sfTDy1W", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Jeffrey Liu", "gender": "Male", "institution": "MIT Lincoln Laboratory, Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Katherine Picchione", "gender": "Female", "institution": "NASA Langley Research Center", "country": "US", "position": "Researcher" }, { "name": "Sam Scheele", "gender": "Male", "institution": "MIT Lincoln Laboratory, Massachusetts Institute of Technology", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 20, "n_ref": 59, "n_ref_all": 79, "n_fig": 8, "n_tab": 5, "L_tab": 937, "n_element_tab": 134, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 857, "n_element_tab_1": 131, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 102389, "len_all_1": 57624, "len_abs": 1339, "len_title": 126, "len_sents": 32695, "len_sents_1": 26691, "n_sents": 198, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1359, "title": "LADI v2: Multi-label Dataset and Classifiers for Low-Altitude Disaster Imagery", "abs": "ML-based computer vision models are promising tools for supporting emergency management operations following natural disasters. Imagery taken from small manned and unmanned aircraft can be available soon after a disaster and provide valuable information from multiple perspectives for situational awareness and damage assessment applications. However, emergency managers often face challenges in effectively utilizing this data due to the difficulties in finding the most relevant imagery among the tens of thousands of images that may be taken after an event. Despite this promise, there is still a lack of training data for imagery of this type from multiple perspectives and for multiple hazard types. To address this, we present the LADI v2 (Low Altitude Disaster Imagery version 2) dataset, a curated set of about 10,000 disaster images captured by the Civil Air Patrol (CAP) in response to over 100 federal emergency declarations (2015-2023) from over 30 US states and territories and annotated for multi-label classification by trained CAP volunteers. We also provide two pretrained baseline classifiers and compare their performance to state-of-the-art vision-language models in multi-label classification. The data and code are released publicly to support the development of computer vision models for emergency management research and applications.", "keywords": [ "computer vision", "dataset", "disaster", "aerial imagery", "multilabel classification", "damage assessment" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "wCwz1F8qY8", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Derek Huang", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Peicong Lin", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Sheng-You Huang", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 37, "n_ref": 72, "n_ref_all": 96, "n_fig": 5, "n_tab": 4, "L_tab": 1027, "n_element_tab": 207, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 318, "n_element_tab_1": 64, "formula_len_all": 181, "formula_len_all_1": 99, "len_all": 107493, "len_all_1": 54857, "len_abs": 1228, "len_title": 147, "len_sents": 36423, "len_sents_1": 26764, "n_sents": 236, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1235, "title": "Prediction of Protein-protein Contacts with Structure-aware Single-sequence Protein Language Models", "abs": "Accurate prediction of the interface residue-residue contacts between interacting proteins is valuable for determining the structure and function of protein complexes. Recent deep learning methods have drastically improved the accuracy of predicting the interface contacts of protein complexes. However, existing methods rely on Multiple Sequence Alignments (MSA) features which pose limitations on prediction accuracy, speed, and computational efficiency. Here, we propose a transformer-powered deep learning method to predict the inter-protein residue-residue contacts based on both single-sequence and structure-aware protein language models (PLM), called DeepSSInter. Utilizing the intra-protein distance and graph representations and the ESM2 and SaProt protein language models, we are able to generate the structure-aware features for the protein receptor, ligand, and complex. These structure-aware features are passed into the Resnet Inception module and the Triangle-aware module to effectively produce the predicted inter-protein contact map. Extensive experiments on both homo- and hetero-dimeric complexes show that our DeepSSInter model significantly improves the performance compared to previous state-of-the-art methods.", "keywords": [ "Protein bioinformatics", "Protein language models", "Protein-protein contact prediction", "Protein representations", "Deep neural networks" ], "rating_list": [ 6, 8, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "wCXAlfvCy6", "primary_area": "generative models", "n_author": 18, "author_info_list": [ { "name": "Dacheng Li", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Fuzhao Xue", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Haotian Tang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Jan Kautz", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "VP Research" }, { "name": "Ligeng Zhu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Linxi Fan", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Pavlo Molchanov", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Research Scientist" }, { "name": "Qinghao Hu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Shang Yang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Song Han", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Xiuyu Li", "gender": "Not Specified", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Yao Lu", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Yihui He", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yukang Chen", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yuke Zhu", "gender": "Male", "institution": "Computer Science Department, University of Texas, Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Yunhao Fang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Zhijian Liu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Hongxu Yin", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Senior Research Scientist" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 45, "n_ref": 111, "n_ref_all": 143, "n_fig": 14, "n_tab": 6, "L_tab": 4431, "n_element_tab": 384, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 5684, "n_element_tab_1": 32, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 191760, "len_all_1": 57721, "len_abs": 1229, "len_title": 118, "len_sents": 41194, "len_sents_1": 25579, "n_sents": 306, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1236, "title": "LongVILA: Scaling Long-Context Visual Language Models for Long Videos", "abs": "Long-context capability is critical for multi-modal foundation models, especially for long video understanding. We introduce LongVILA, a full-stack solution for long-context visual-language models by co-designing the algorithm and system. For model training, we upgrade existing VLMs to support long video understanding by incorporating two additional stages, i.e., long context extension and long video supervised fine-tuning. However, training on long video is computationally and memory intensive. We introduce the long-context Multi-Modal Sequence Parallelism (MM-SP) system that efficiently parallelizes long video training and inference, enabling 2M context length training on 256 GPUs without any gradient checkpointing. LongVILA efficiently extends the number of video frames of VILA from 8 to 2048, achieving 99.8% accuracy in 6,000-frame (more than 1 million tokens) video needle-in-a-haystack. LongVILA-7B demonstrates strong accuracy on 9 popular video benchmarks, e.g., 65.1% VideoMME with subtitle. Besides, MM-SP is 2.1x - 5.7x faster than ring style sequence parallelism and 1.1x - 1.4x faster than Megatron with a hybrid context and tensor parallelism. Moreover, it seamlessly integrates with Hugging Face Transformers.", "keywords": [ "Large language models", "Long context", "Multi-modality", "Video understanding" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "wCOJpXm0Me", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Kate Saenko", "gender": "Female", "institution": "Meta AI", "country": "US", "position": "Researcher" }, { "name": "Kuniaki Saito", "gender": "Male", "institution": "OMRON SINICX", "country": "JP", "position": "Researcher" }, { "name": "Piotr Teterwak", "gender": "Male", "institution": "Boston University", "country": "US", "position": "PhD student" }, { "name": "Theodoros Tsiligkaridis", "gender": "Male", "institution": "MIT Lincoln Laboratory, Massachusetts Institute of Technology", "country": "US", "position": "Senior AI Research Scientist" }, { "name": "Bryan Allen Plummer", "gender": "Male", "institution": "Boston University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 38, "n_ref": 82, "n_ref_all": 120, "n_fig": 13, "n_tab": 20, "L_tab": 8565, "n_element_tab": 1064, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 5497, "n_element_tab_1": 277, "formula_len_all": 210, "formula_len_all_1": 210, "len_all": 172855, "len_all_1": 58518, "len_abs": 1459, "len_title": 116, "len_sents": 44487, "len_sents_1": 24577, "n_sents": 333, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1468, "title": "Is Large-scale Pretraining the Secret to Good Domain Generalization?", "abs": "Multi-Source Domain Generalization (DG) is the task of training on multiple source domains and achieving high classification performance on unseen target domains. Recent methods combine robust features from web-scale pretrained backbones with new features learned from source data, and this has dramatically improved benchmark results. However, it remains unclear if DG finetuning methods are becoming better over time, or if improved benchmark performance is simply an artifact of stronger pre-training. Prior studies have shown that perceptual similarity to pre-training data correlates with zero-shot performance, but we find the effect limited in the DG setting. Instead, we posit that having perceptually similar data in pretraining is not enough; and that it is how well these data were learned that determines performance. This leads us to introduce the Alignment Hypothesis, which states that the final DG performance will be high if and only if alignment of image and class label text embeddings is high. Our experiments confirm the Alignment Hypothesis is true, and we use it as an analysis tool of existing DG methods evaluated on DomainBed datasets by splitting evaluation data into In-pretraining (IP) and Out-of-pretraining (OOP). We show that all evaluated DG methods struggle on DomainBed-OOP, while recent methods excel on DomainBed-IP. Put together, our findings highlight the need for DG methods which can generalize beyond pretraining alignment.", "keywords": [ "Domain Generalization", "Robustness", "CLIP", "Pretraining" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "wCO966fAHd", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Feng Zhou", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Assistant Professor" }, { "name": "Zenan Ling", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Researcher" }, { "name": "Zhilin Zhao", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "YIXUAN ZHANG", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 20, "n_formula_1": 11, "n_ref_uni": 32, "n_ref": 48, "n_ref_all": 57, "n_fig": 3, "n_tab": 6, "L_tab": 1739, "n_element_tab": 204, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1390, "n_element_tab_1": 154, "formula_len_all": 1565, "formula_len_all_1": 657, "len_all": 104742, "len_all_1": 83020, "len_abs": 1081, "len_title": 119, "len_sents": 29303, "len_sents_1": 26770, "n_sents": 207, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1085, "title": "Dual-level Bias Mitigation via Fairness-guided Distribution Discrepancy", "abs": "Modern artificial intelligence predominantly relies on pre-trained models, which are fine-tuned for specific downstream tasks rather than built from scratch. However, a key challenge persists: the fairness of learned representations in pre-trained models is not guaranteed when transferred to new tasks, potentially leading to biased outcomes, even if fairness constraints were applied during the original training. To address this issue, we propose Dual-level Bias Mitigation (DBM), which measures the fairness-guided distribution discrepancy between representations of different demographic groups. By optimizing both the fairness-guided distribution discrepancy and the task-specific objective, DBM ensures fairness at both the representation and task levels. Theoretically, we provide the generalization error bound of the fairness-guided distribution discrepancy to support the efficacy of our approach. Experimental results on multiple benchmark datasets demonstrate that DBM effectively mitigates bias in fine-tuned models on downstream tasks across a range of fairness metrics.", "keywords": [ "Fairness", "Trust-worthy Machine Learning" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "wCNuEA5MSv", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Hao Liu", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Hao Wang", "gender": "Male", "institution": "The Hong Kong University of Science and Techenology (Guangzhou))", "country": "CN", "position": "Intern" }, { "name": "Jindong Han", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Wei Fan", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc Researcher" } ], "n_formula": 32, "n_formula_1": 11, "n_ref_uni": 46, "n_ref": 127, "n_ref_all": 143, "n_fig": 7, "n_tab": 6, "L_tab": 2611, "n_element_tab": 137, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2597, "n_element_tab_1": 75, "formula_len_all": 1335, "formula_len_all_1": 540, "len_all": 198470, "len_all_1": 69397, "len_abs": 2423, "len_title": 132, "len_sents": 63634, "len_sents_1": 32564, "n_sents": 394, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1599, "title": "Language Model Empowered Spatio-Temporal Forecasting via Physics-Aware Reprogramming", "abs": "Spatio-temporal forecasting is pivotal in numerous real-world applications, including transportation planning, energy management, and climate monitoring. \nIn this work, we aim to harness the reasoning and generalization abilities of Pre-trained Language Models (PLMs) for more effective spatio-temporal forecasting, particularly in data-scarce scenarios. \nHowever, recent studies uncover that PLMs, which are primarily trained on textual data, often falter when tasked with modeling the intricate correlations inherent in numerical time series, thereby limiting their effectiveness in comprehending spatio-temporal data.\nTo bridge the gap, we propose REPST, a physics-aware PLM reprogramming framework tailored for spatio-temporal forecasting. \nSpecifically, we first propose a physics-aware decomposer that adaptively disentangles spatially correlated time series into interpretable sub-components, which facilitates PLM\u2019s understanding of sophisticated spatio-temporal dynamics via a divide-and-conquer strategy.\nMoreover, we propose a selective discrete reprogramming scheme, which introduces an expanded spatio-temporal vocabulary space to project spatio-temporal series into discrete representations. This scheme minimizes the information loss during reprogramming and enriches the representations derived by PLMs.\nExtensive experiments on real-world datasets show that the proposed REPST outperforms twelve state-of-the-art baseline methods, particularly in data-scarce scenarios, highlighting the effectiveness and superior generalization capabilities of PLMs for spatio-temporal forecasting.", "keywords": [ "spatio-temporal forecasting", "time series forecasting" ], "rating_list": [ 6, 5, 3, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 4 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "wCIkU0XR4f", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Michelle V Mancenido", "gender": "Female", "institution": "Arizona State University", "country": "", "position": "Assistant Professor" }, { "name": "Rong Pan", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Full Professor" }, { "name": "Yang Ba", "gender": "unknown", "institution": "Arizona State University", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 41, "n_ref_all": 53, "n_fig": 6, "n_tab": 2, "L_tab": 108, "n_element_tab": 25, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 669, "n_element_tab_1": 46, "formula_len_all": 197, "formula_len_all_1": 197, "len_all": 109107, "len_all_1": 59207, "len_abs": 1608, "len_title": 118, "len_sents": 36660, "len_sents_1": 28726, "n_sents": 262, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1239, "title": "How Does Data Diversity Shape The Weight Landscape of Neural Networks?", "abs": "To enhance the generalization of machine learning models to unseen data, techniques such as dropout, weight decay (L2 regularization), and noise augmentation\nare commonly employed. While regularization methods (i.e., dropout and weight\ndecay) are geared toward adjusting model parameters to prevent overfitting, data\naugmentation increases the diversity of the input training set, a method purported\nto improve accuracy and calibration error. In this paper, we investigate the impact of each of these techniques on the parameter space of neural networks, with\nthe goal of understanding how they alter the weight landscape in transfer learning\nscenarios. To accomplish this, we employ Random Matrix Theory to analyze the\neigenvalue distributions of pre-trained models, fine-tuned using these techniques\nbut using different levels of data diversity, for the same downstream tasks. We\nobserve that diverse data influences the weight landscape in a similar fashion as\ndropout. Additionally, we compare commonly used data augmentation methods\nwith synthetic data created by generative models. We conclude that synthetic data\ncan bring more diversity into real input data, resulting in a better performance on\nout-of-distribution test instances.", "keywords": [ "data diversity", "regularization", "data augmentation", "synthetic data", "transfer learning" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "wAemQcyWqq", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chenhan Zhang", "gender": "Male", "institution": "Macquarie University", "country": "AU", "position": "Postdoc" }, { "name": "Shui Yu", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Full Professor" }, { "name": "Weiqi Wang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Zhiyi Tian", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 18, "n_ref_uni": 57, "n_ref": 175, "n_ref_all": 224, "n_fig": 10, "n_tab": 10, "L_tab": 2685, "n_element_tab": 293, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1651, "n_element_tab_1": 99, "formula_len_all": 2190, "formula_len_all_1": 1143, "len_all": 227633, "len_all_1": 69694, "len_abs": 1890, "len_title": 129, "len_sents": 68750, "len_sents_1": 31228, "n_sents": 517, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1897, "title": "Oblivious Unlearning by Learning: Machine Unlearning Without Exposing Erased Data", "abs": "Machine unlearning enables users to remove the influence of their data from trained models, thus protecting their privacy. However, it is paradoxical that most unlearning methods require users first to upload their to-be-removed data to machine learning servers and notify the servers of their unlearning intentions to prepare appropriate unlearning methods. Both unlearned data and unlearning intentions are sensitive user information. Exposing this information to the server for unlearning operations conflicts with the privacy protection goal. In this paper, we investigate the challenge of implementing unlearning without exposing erased data and unlearning intentions to the server. We propose an Oblivious Unlearning by Learning (OUbL) approach to address this privacy-preserving machine unlearning problem. In OUbL, the users construct a new dataset with synthesized unlearning noise, ensuring that once the server continually updates the model using the original learning algorithm based on this dataset, it can implement unlearning. The server does not need to perform any tailored unlearning operation and remains unaware that the constructed samples are for unlearning. As a result, the process is oblivious to the server regarding unlearning intentions. Additionally, by transforming the original erased data into unlearning noise and distributing this noise across numerous auxiliary samples, our approach protects the privacy of the unlearned data while effectively implementing unlearning. The effectiveness of the proposed OUbL method is evaluated through extensive experiments on three representative datasets across various model architectures and four mainstream unlearning benchmarks. The results demonstrate the significant superiority of OUbL over the state-of-the-art privacy-preserving unlearning benchmarks in terms of both privacy protection and unlearning effectiveness.", "keywords": [ "Machine unlearning", "Privacy Preserving" ], "rating_list": [ 6, 3, 5, 6, 8 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 5, 3, 3, 3, 4 ] }, { "paper_id": "wAXsx2MYgV", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Colin Conwell", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Postdoc" }, { "name": "Emalie McMahon", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Kathy Garcia", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Leyla Isik", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" }, { "name": "Michael Bonner", "gender": "unknown", "institution": "Johns Hopkins University", "country": "", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 57, "n_ref": 114, "n_ref_all": 153, "n_fig": 23, "n_tab": 2, "L_tab": 4796, "n_element_tab": 458, "n_fig_1": 22, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 169270, "len_all_1": 61106, "len_abs": 1848, "len_title": 0, "len_sents": 39303, "len_sents_1": 31073, "n_sents": 230, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1688, "title": "Modeling dynamic social vision highlights gaps between deep learning and humans", "abs": "Deep learning models trained on computer vision tasks are widely considered the most successful models of human vision to date. The majority of work that supports this idea evaluates how accurately these models predict behavior and brain responses to static images of objects and scenes. Real-world vision, however, is highly dynamic, and far less work has evaluated deep learning models on human responses to moving stimuli, especially those that involve more complicated, higher-order phenomena like social interactions. Here, we extend a dataset of natural videos depicting complex multi-agent interactions by collecting human-annotated sentence captions for each video, and we benchmark 350+ image, video, and language models on behavior and neural responses to the videos. As in prior work, we find that many vision models reach the noise ceiling in predicting visual scene features and responses along the ventral visual stream (often considered the primary neural substrate of object and scene recognition). In contrast, vision models poorly predict human action and social interaction ratings and neural responses in the lateral stream (a neural pathway theorized to specialize in dynamic, social vision), though video models show a striking advantage in predicting mid-level lateral stream regions. Language models (given human sentence captions of the videos) predict action and social ratings better than image and video models, but perform poorly at predicting neural responses in the lateral stream. Together, these results identify a major gap in AI's ability to match human social vision and provide insights to guide future model development for dynamic, natural contexts.", "keywords": [ "NeuroAI", "vision", "fMRI", "deep learning", "social perception" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "wA2RMD2AFq", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Baochang Zhang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Professor" }, { "name": "Boyu Liu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Guodong Guo", "gender": "Male", "institution": "West Virginia University", "country": "US", "position": "Full Professor" }, { "name": "Haoyu Huang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Linlin Yang", "gender": "Male", "institution": "Communication University of China", "country": "CN", "position": "Lecturer" }, { "name": "Xianbin Cao", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Yanjing Li", "gender": "unknown", "institution": "Beihang University", "country": "CN", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 57, "n_ref": 116, "n_ref_all": 135, "n_fig": 5, "n_tab": 9, "L_tab": 3714, "n_element_tab": 478, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1216, "n_element_tab_1": 214, "formula_len_all": 397, "formula_len_all_1": 397, "len_all": 152284, "len_all_1": 62373, "len_abs": 1495, "len_title": 126, "len_sents": 40621, "len_sents_1": 27896, "n_sents": 301, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1516, "title": "Efficient Low-Bit Quantization with Adaptive Scales for Multi-Task Co-Training", "abs": "Co-training can achieve parameter-efficient multi-task models but remains unexplored for quantization-aware training. Our investigation shows that directly introducing co-training into existing quantization-aware training (QAT) methods results in significant performance degradation. Our experimental study identifies that the primary issue with existing QAT methods stems from the inadequate activation quantization scales for the co-training framework. To address this issue, we propose Task-Specific Scales Quantization for Multi-Task Co-Training (TSQ-MTC) to tackle mismatched quantization scales. Specifically, a task-specific learnable multi-scale activation quantizer (TLMAQ) is incorporated to enrich the representational ability of shared features for different tasks. Additionally, we find that in the deeper layers of the Transformer model, the quantized network suffers from information distortion within the attention quantizer. A structure-based layer-by-layer distillation (SLLD) is then introduced to ensure that the quantized features effectively preserve the information from their full-precision counterparts. Our extensive experiments in two co-training scenarios demonstrate the effectiveness and versatility of TSQ-MTC. In particular, we successfully achieve a 4-bit quantized low-level visual foundation model based on IPT, which attains a PSNR comparable to the full-precision model while offering a $7.99\\times$ compression ratio in the $\\times4$ super-resolution task on the Set5 benchmark.", "keywords": [ "Low-Bit Quantization", "Multi-Task Learning", "Co-Training", "Quantization-Aware Training", "Quantization Scale" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 4, 4 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 5, 3, 3 ] }, { "paper_id": "w9tS6NRmxX", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Chao Yu", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Eugene Vinitsky", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Hong Lu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Jiaxuan Gao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Qixin Tan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Xinting Yang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yi Wu", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Assistant Professor" }, { "name": "Yu Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 2, "n_ref_uni": 30, "n_ref": 56, "n_ref_all": 79, "n_fig": 5, "n_tab": 14, "L_tab": 10269, "n_element_tab": 464, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1520, "n_element_tab_1": 138, "formula_len_all": 427, "formula_len_all_1": 131, "len_all": 212840, "len_all_1": 71372, "len_abs": 1342, "len_title": 116, "len_sents": 73993, "len_sents_1": 33680, "n_sents": 594, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1348, "title": "Few-shot In-context Preference Learning using Large Language Models", "abs": "Designing reward functions is a core component of reinforcement learning but can be challenging for truly complex behavior. Reinforcement Learning from Human Feedback (RLHF) has been used to alleviate this challenge by replacing a hand-coded reward function with a reward function learned from preferences. However, it can be exceedingly inefficient to learn these rewards as they are often learned tabula rasa. We investigate whether Large Language Models (LLMs) can reduce this query inefficiency by converting an iterative series of human preferences into code representing the rewards. We propose In-Context Preference Learning (ICPL), a method that uses the grounding of an LLM to accelerate learning reward functions from preferences. ICPL takes the environment context and task description, synthesizes a set of reward functions, and then repeatedly updates the reward functions using human feedback over videos of the resultant policies over a small number of trials. Using synthetic preferences, we demonstrate that ICPL is orders of magnitude more efficient than RLHF and is even competitive with methods that use ground-truth reward functions instead of preferences. Finally, we perform a series of human preference-learning trials and observe that ICPL extends beyond synthetic settings and can work effectively with humans-in-the-loop.", "keywords": [ "Large Language Models", "Preference-based RL", "Reinforcement Learning from Human Feedback", "Reward Design" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 5, 4, 3 ] }, { "paper_id": "w9bWY6LvrW", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Honghao Wei", "gender": "Male", "institution": "Washington State University ", "country": "US", "position": "Assistant Professor" }, { "name": "Keru Chen", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "Undergrad student" }, { "name": "Sen Lin", "gender": "unknown", "institution": "University of Houston", "country": "US", "position": "Assistant Professor" }, { "name": "Zhigang Deng", "gender": "Male", "institution": "University of Houston", "country": "", "position": "Full Professor" } ], "n_formula": 25, "n_formula_1": 23, "n_ref_uni": 56, "n_ref": 98, "n_ref_all": 140, "n_fig": 16, "n_tab": 1, "L_tab": 68, "n_element_tab": 5, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2498, "formula_len_all_1": 2126, "len_all": 185702, "len_all_1": 69341, "len_abs": 1736, "len_title": 134, "len_sents": 60493, "len_sents_1": 36195, "n_sents": 385, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1779, "title": "Marvel: Accelerating Safe Online Reinforcement Learning with Finetuned Offline Policy", "abs": "The high costs and risks involved in extensive environment interactions hinder the practical application of current online safe reinforcement learning (RL) methods. While offline safe RL addresses this by learning policies from static datasets, the performance therein is usually limited due to reliance on data quality and challenges with out-of-distribution (OOD) actions. Inspired by recent successes in offline-to-online (O2O) RL, it is crucial to explore whether offline safe RL can be leveraged to facilitate faster and safer online policy learning, a direction that has yet to be fully investigated. To fill this gap, we first demonstrate that naively applying existing O2O algorithms from standard RL would not work well in the safe RL setting due to two unique challenges: \\emph{erroneous Q-estimations}, resulted from offline-online objective mismatch and offline cost sparsity, and \\emph{Lagrangian mismatch}, resulted from difficulties in aligning Lagrange multipliers between offline and online policies. To address these challenges, we introduce \\textbf{Marvel}, a novel framework for O2O safe RL, comprising two key components that work in concert: \\emph{Value Pre-Alignment} to align the Q-functions with the underlying truth before online learning, and \\emph{Adaptive PID Control} to effectively adjust the Lagrange multipliers during online finetuning. Extensive experiments demonstrate that Marvel significantly outperforms existing baselines in both reward maximization and safety constraint satisfaction. By introducing the first policy-finetuning based framework for O2O safe RL, which is compatible with many offline and online safe RL methods, our work has the great potential to advance the field towards more efficient and practical safe RL solutions.", "keywords": [ "Offline-to-Online Fine-tuning", "Safe Reinforcement Learning", "Constrained Markov Decision Processes", "Reinforcement Learning" ], "rating_list": [ 5, 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "w8LMtFY97b", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 7, "author_info_list": [ { "name": "Juan Eugenio Iglesias", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Associate Professor" }, { "name": "Karthik Gopinath", "gender": "Male", "institution": "Massachusetts General Hospital, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Koen Van Leemput", "gender": "Male", "institution": "Massachusetts General Hospital, Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "Malte Hoffmann", "gender": "unknown", "institution": "Massachusetts General Hospital, Harvard University", "country": "US", "position": "Assistant Professor" }, { "name": "Oula Puonti", "gender": "Male", "institution": "Danish Research Centre for Magnetic Resonance", "country": "DK", "position": "Researcher" }, { "name": "Peirong Liu", "gender": "Female", "institution": "Harvard Medical School", "country": "US", "position": "Researcher" }, { "name": "Xiaoling Hu", "gender": "Male", "institution": "MGH and Harvard Medical School", "country": "US", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 58, "n_ref": 102, "n_ref_all": 121, "n_fig": 6, "n_tab": 9, "L_tab": 1548, "n_element_tab": 138, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 683, "n_element_tab_1": 25, "formula_len_all": 549, "formula_len_all_1": 567, "len_all": 153400, "len_all_1": 64713, "len_abs": 1411, "len_title": 131, "len_sents": 41361, "len_sents_1": 31688, "n_sents": 293, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1456, "title": "Hierarchical Uncertainty Estimation for Learning-based Registration in Neuroimaging", "abs": "Over recent years, deep learning based image registration has achieved impressive accuracy in many domains, including medical imaging and, specifically, human neuroimaging with magnetic resonance imaging (MRI). However, the uncertainty estimation associated with these methods has been largely limited to the application of generic techniques (e.g., Monte Carlo dropout) that do not exploit the peculiarities of the problem domain, particularly spatial modeling. Here, we propose a principled way to propagate uncertainties (epistemic or aleatoric) estimated at the level of spatial location by these methods, to the level of global transformation models, and further to downstream tasks. Specifically, we justify the choice of a Gaussian distribution for the local uncertainty modeling, and then propose a framework where uncertainties spread across hierarchical levels, depending on the choice of transformation model. Experiments on publicly available data sets show that Monte Carlo dropout correlates very poorly with the reference registration error, whereas our uncertainty estimates correlate much better. % with the reference registration error. Crucially, the results also show that uncertainty-aware fitting of transformations improves the registration accuracy of brain MRI scans. Finally, we illustrate how sampling from the posterior distribution of the transformations can be used to propagate uncertainties to downstream neuroimaging tasks.", "keywords": [ "Image registration", "uncertainty estimation", "medical image analysis" ], "rating_list": [ 6, 5, 8 ], "soundness_list": [ 3, 2, 4 ], "presentation_list": [ 3, 2, 4 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "w7vn6ah0Qg", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Jinyue Tian", "gender": "Female", "institution": "Southeast University", "country": "", "position": "MS student" }, { "name": "Meimei Yang", "gender": "Female", "institution": "Southeast University", "country": "", "position": "PhD student" }, { "name": "Pengfei Fang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Shipeng Zhu", "gender": "Male", "institution": "Southeast University", "country": "", "position": "PhD student" }, { "name": "Yanfang Xue", "gender": "Female", "institution": "Southeast University", "country": "CN", "position": "PhD student" }, { "name": "hui xue", "gender": "Female", "institution": "Southeast University", "country": "CN", "position": "Full Professor" } ], "n_formula": 51, "n_formula_1": 22, "n_ref_uni": 24, "n_ref": 62, "n_ref_all": 85, "n_fig": 3, "n_tab": 10, "L_tab": 8104, "n_element_tab": 1154, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 3784, "n_element_tab_1": 468, "formula_len_all": 5058, "formula_len_all_1": 1423, "len_all": 174008, "len_all_1": 77667, "len_abs": 1647, "len_title": 108, "len_sents": 44087, "len_sents_1": 31512, "n_sents": 381, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1671, "title": "KokerNet: Koopman Kernel Network for Time Series Forecasting", "abs": "The Koopman operator has gained increasing attention in time series forecasting due to its ability to simplify the complex evolution of dynamic systems. However, most existing Koopman-based methods suffer from significant computational costs in constructing measurement functions and struggle to address the challenge posed by the variation in data distribution. Additionally, these approaches tend to empirically decompose time series or distributions into combinations of components, lacking interpretability. To tackle these issues, we propose a novel approach, **Ko**opman **ker**nel **net**work (**KokerNet**), for time series forecasting. On one hand, we construct a measurement function space using the spectral kernel method, which enables us to perform Koopman operator learning in a low-dimensional feature space, efficiently reducing computational costs. On the other hand, an index is designed to characterize the stationarity of data in both time and frequency domains. This index can interpretably guide us to decompose the time series into stationary and non-stationary components. The global and local Koopman operators are then learned within the constructed measurement function space to predict the future behavior of the stationary and non-stationary components, respectively. Particularly, to address the challenge posed by the variation in distribution, we incorporate a distribution module for the non-stationary component, ensuring that the model can make aligned distribution predictions. Extensive experiments across multiple benchmarks illustrate the superiority of our proposed KokerNet, consistently outperforming the state-of-the-art models.", "keywords": [ "Spectral kernel", "Koopman operator", "Time series" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "w7pMjyjsKN", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Francesco Giannini", "gender": "Male", "institution": "School of Education Pisa", "country": "IT", "position": "Researcher" }, { "name": "Gabriele Dominici", "gender": "Male", "institution": "Universita della Svizzera Italiana", "country": "CH", "position": "PhD student" }, { "name": "Giuseppe Marra", "gender": "Male", "institution": "KU Leuven", "country": "", "position": "Assistant Professor" }, { "name": "Marc Langheinrich", "gender": "Male", "institution": "Universita della Svizzera Italiana", "country": "CH", "position": "Full Professor" }, { "name": "Martin Gjoreski", "gender": "Male", "institution": "Universit\u00e0 della svizzera italiana", "country": "CH", "position": "Principal Researcher" }, { "name": "Pietro Barbiero", "gender": "Male", "institution": "International Business Machines", "country": "CH", "position": "Postdoc" } ], "n_formula": 16, "n_formula_1": 9, "n_ref_uni": 41, "n_ref": 80, "n_ref_all": 108, "n_fig": 7, "n_tab": 11, "L_tab": 4260, "n_element_tab": 360, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 581, "n_element_tab_1": 18, "formula_len_all": 1442, "formula_len_all_1": 837, "len_all": 172651, "len_all_1": 69115, "len_abs": 3936, "len_title": 88, "len_sents": 51703, "len_sents_1": 33355, "n_sents": 350, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1602, "title": "Counterfactual Concept Bottleneck Models", "abs": "Current deep learning models are not designed to simultaneously address three fundamental questions: predict class labels to solve a given classification task (the \"What?\"), simulate changes in the situation to evaluate how this impacts class predictions (the \"How?\"), and imagine how the scenario should change to result in different class predictions (the \"Why not?\"). While current approaches in causal representation learning and concept interpretability are designed to address some of these questions individually (such as Concept Bottleneck Models, which address both ``what'' and ``how'' questions), no current deep learning model is specifically built to answer all of them at the same time. To bridge this gap, we introduce CounterFactual Concept Bottleneck Models (CF-CBMs), a class of models designed to efficiently address the above queries all at once without the need to run post-hoc searches. Our experimental results demonstrate that CF-CBMs: achieve classification accuracy comparable to black-box models and existing CBMs (\u201cWhat?\u201d), rely on fewer important concepts leading to simpler explanations (\u201cHow?\u201d), and produce interpretable, concept-based counterfactuals (\u201cWhy not?\u201d). Additionally, we show that training the counterfactual generator jointly with the CBM leads to two key improvements: (i) it alters the model's decision-making process, making the model rely on fewer important concepts (leading to simpler explanations), and (ii) it significantly increases the causal effect of concept interventions on class predictions, making the model more responsive to these changes.", "keywords": [ "Concept Bottleneck Models", "Concept Based Model", "Counterfactuals", "Explainable AI", "Interpretability" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "w7P92BEsb2", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Apivich Hemachandra", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Bryan Kian Hsiang Low", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Associate Professor" }, { "name": "Gregory Kang Ruey Lau", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "See-Kiong Ng", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 60, "n_formula_1": 9, "n_ref_uni": 49, "n_ref": 106, "n_ref_all": 160, "n_fig": 18, "n_tab": 5, "L_tab": 1876, "n_element_tab": 199, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 3, "n_element_tab_1": 1, "formula_len_all": 4946, "formula_len_all_1": 663, "len_all": 220641, "len_all_1": 63031, "len_abs": 2028, "len_title": 111, "len_sents": 74901, "len_sents_1": 31492, "n_sents": 602, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 2064, "title": "PIED: Physics-Informed Experimental Design for Inverse Problems", "abs": "In many science and engineering settings, system dynamics are characterized by governing partial differential equations (PDEs), and a major challenge is to solve inverse problems (IPs) where unknown PDE parameters are inferred based on observational data gathered under limited budget. \nDue to the high costs of setting up and running experiments, experimental design (ED) is often done with the help of PDE simulations to optimize for the most informative design parameters (e.g., sensor placements) to solve such IPs, prior to actual data collection. This process of optimizing design parameters is especially critical when the budget and other practical constraints make it infeasible to adjust the design parameters between trials during the experiments.\nHowever, existing experimental design (ED) methods tend to require sequential and frequent design parameter adjustments between trials. Furthermore, they also have significant computational bottlenecks due to the need for complex numerical simulations for PDEs, and do not exploit the advantages provided by physics informed neural networks (PINNs) in solving IPs for PDE-governed systems, such as its meshless solutions, differentiability, and amortized training. \nThis work presents Physics-Informed Experimental Design (PIED), the first ED framework that makes use of PINNs in a fully differentiable architecture to perform continuous optimization of design parameters for IPs for one-shot deployments. \nPIED overcomes existing methods' computational bottlenecks through parallelized computation and meta-learning of PINN parameter initialization, and proposes novel methods to effectively take into account PINN training dynamics in optimizing the ED parameters. \nThrough experiments based on noisy simulated data and even real world experimental data, we empirically show that given limited observation budget, PIED significantly outperforms existing ED methods in solving IPs, including for challenging settings where the inverse parameters are unknown functions rather than just finite-dimensional.", "keywords": [ "Physics-Informed Neural Network", "PINNs", "Experimental Design", "AI For Science" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 2, 2 ] }, { "paper_id": "w7BGq6ozOL", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Ali Riahi Samani", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Fateme Golivand Darvishvand", "gender": "Female", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Feng Chen", "gender": "Male", "institution": "University of Texas, Dallas", "country": "", "position": "Associate Professor" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 36, "n_ref": 64, "n_ref_all": 74, "n_fig": 2, "n_tab": 6, "L_tab": 3860, "n_element_tab": 516, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 88, "formula_len_all_1": 144, "len_all": 176479, "len_all_1": 68545, "len_abs": 1519, "len_title": 175, "len_sents": 54443, "len_sents_1": 36020, "n_sents": 373, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 121, "L_abs": 1527, "title": "Advancing Algorithmic Trading with Large Language Models: A Reinforcement Learning Approach for Stock Market Optimization", "abs": "In the fast-evolving landscape of financial markets, effective decision-making tools are essential for managing complexities driven by economic indicators and market dynamics. Algorithmic trading strategies have gained prominence for their ability to execute trades autonomously, with Deep Reinforcement Learning (DRL) emerging as a key approach for optimizing trading actions through continuous market interaction. However, RL-based systems face significant challenges, particularly in adapting to evolving time series data and incorporating unstructured textual information. In response to these limitations, recent advancements in Large Language Models (LLMs) offer new opportunities. LLMs possess the capacity to analyze vast volumes of data, providing enhanced insights that can complement traditional market analysis. This study proposes a novel approach that integrates six distinct LLMs into algorithmic trading frameworks, developing Stock-Evol-Instruct, an innovative instruction generation algorithm. This algorithm enables RL agents to fine-tune their trading strategies by leveraging LLM-driven insights for daily stock trading decisions. Empirical evaluation using real-world stock data from Silver and JPMorgan demonstrates the significant potential of this approach to outperform conventional trading models. By bridging the gap between LLMs and RL in algorithmic trading, this study contributes to a new frontier in financial technology, setting the stage for future advancements in autonomous trading systems.", "keywords": [ "Algorithmic trading", "Stock market", "Large language models", "Deep reinforcement learning" ], "rating_list": [ 3, 8, 6, 1 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 2, 4, 3, 1 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "w6rHCuN3YG", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Bangcheng Yang", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "MS student" }, { "name": "Jiaqi Li", "gender": "Not Specified", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Siyuan Qi", "gender": "unknown", "institution": "Beijing Institute for General Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Xiaobo Wang", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Intern" }, { "name": "Yaodong Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yifan Zhong", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Zilong Zheng", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "jiangkailin", "gender": "Male", "institution": "University of Science and Technology of China(USTC)", "country": "CN", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 14, "n_ref_uni": 16, "n_ref": 41, "n_ref_all": 66, "n_fig": 6, "n_tab": 15, "L_tab": 3670, "n_element_tab": 672, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 2234, "n_element_tab_1": 371, "formula_len_all": 2439, "formula_len_all_1": 734, "len_all": 186337, "len_all_1": 71683, "len_abs": 1123, "len_title": 118, "len_sents": 52647, "len_sents_1": 31215, "n_sents": 406, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1126, "title": "In-Context Editing: Learning Knowledge from Self-Induced Distributions", "abs": "In scenarios where language models must incorporate new information efficiently without extensive retraining, traditional fine-tuning methods are prone to overfitting, degraded generalization, and unnatural language generation. To address these limitations, we introduce Consistent In-Context Editing (ICE), a novel approach leveraging the model's in-context learning capability to optimize towards a contextual distribution rather than a one-hot target. ICE introduces a simple yet effective optimization framework for the model to internalize new knowledge by aligning its output distributions with and without additional context. This method enhances the robustness and effectiveness of gradient-based tuning methods, preventing overfitting and preserving the model's integrity. We analyze ICE across four critical aspects of knowledge editing: accuracy, locality, generalization, and linguistic quality, demonstrating its advantages. Experimental results confirm the effectiveness of ICE and demonstrate its potential for continual editing, ensuring that the integrity of the model is preserved while updating information.", "keywords": [ "Knowledge Editing", "In-Context Learning", "Language Models" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "w6nlcS8Kkn", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Dongwei Jiang", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "MS student" }, { "name": "Fangcong Yin", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Greg Durrett", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Kyle Mahowald", "gender": "Male", "institution": "The University of Texas at Austin", "country": "", "position": "Assistant Professor" }, { "name": "Manya Wadhwa", "gender": "unknown", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" }, { "name": "Prasann Singhal", "gender": "Male", "institution": ", University of Texas at Austin", "country": "", "position": "Undergrad student" }, { "name": "Xi Ye", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Xinyu Zhao", "gender": "Female", "institution": "University of Texas at Austin", "country": "US", "position": "Researcher" }, { "name": "Zayne Rea Sprague", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" }, { "name": "Juan Rodriguez", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 60, "n_ref": 123, "n_ref_all": 156, "n_fig": 8, "n_tab": 26, "L_tab": 30550, "n_element_tab": 3204, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 389, "n_element_tab_1": 14, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 366671, "len_all_1": 58929, "len_abs": 1263, "len_title": 130, "len_sents": 62542, "len_sents_1": 29231, "n_sents": 483, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1268, "title": "To CoT or not to CoT? Chain-of-thought helps mainly on math and symbolic reasoning", "abs": "Chain-of-thought (CoT) via prompting is the de facto method for eliciting reasoning capabilities from large language models (LLMs). But for what kinds of tasks is this extra \"thinking\" really helpful? To analyze this, we conducted a quantitative meta-analysis covering over 100 papers using CoT and ran our own evaluations of 20 datasets across 14 models. Our results show that CoT gives strong performance benefits primarily on tasks involving math or logic, with much smaller gains on other types of tasks. On MMLU, directly generating the answer without CoT leads to almost identical accuracy as CoT unless the question or model's response contains an equals sign, indicating symbolic operations and reasoning. Following this finding, we analyze the behavior of CoT on these problems by separating planning and execution and comparing against tool-augmented LLMs. Much of CoT's gain comes from improving symbolic execution, but it underperforms relative to using a symbolic solver. Our results indicate that CoT can be applied selectively, maintaining performance while saving inference costs. Furthermore, they suggest a need to move beyond prompt-based CoT to new paradigms that better leverage intermediate computation across the whole range of LLM applications.", "keywords": [ "Chain-of-Thought", "Large Language Models", "Textual Reasoning" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 2, 4, 4 ], "contribution_list": [ 2, 4, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "w6mjerkePG", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Daphne Koller", "gender": "Female", "institution": "Insitro", "country": "", "position": "CEO" }, { "name": "Emily B. Fox", "gender": "Female", "institution": "Inistro Inc.", "country": "US", "position": "Principal Researcher" }, { "name": "Rounak Dey", "gender": "Male", "institution": "Insitro", "country": "", "position": "Senior Data Scientist" }, { "name": "Tommaso Dreossi", "gender": "unknown", "institution": "Amazon Search", "country": "", "position": "Applied Scientist" } ], "n_formula": 11, "n_formula_1": 5, "n_ref_uni": 41, "n_ref": 53, "n_ref_all": 75, "n_fig": 15, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 13, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 922, "formula_len_all_1": 432, "len_all": 141745, "len_all_1": 61937, "len_abs": 1009, "len_title": 128, "len_sents": 43793, "len_sents_1": 31262, "n_sents": 318, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1012, "title": "Roll-AE: A Spatiotemporal Invariant Autoencoder for Neuronal Electro-Physiology", "abs": "Micro-electrode array (MEA) assays enable high-throughput recording of the electrophysiological activity in biological tissues, both in vivo and in vitro. While various classical and deep learning models have been developed for MEA signal analysis, the majority focus on in vivo experiments or specific downstream applications in vitro. Consequently, extracting relevant features from in vitro MEA recordings has remained largely dependent on particular curated features known as neural metrics. In this work, we introduce Roll-AE, a novel autoencoder designed to extract spatiotemporally invariant features from in vitro MEA recordings. Roll-AE serves as a foundational model that facilitates a wide range of downstream tasks. We demonstrate that 1) Roll-AE's embeddings outperform those from standard autoencoders across various classification tasks, and 2) Roll-AE's embeddings effectively characterize electrophysiological phenotypic traits in induced Pluripotent Stem Cells (iPSC)-derived neuronal cultures.", "keywords": [ "autoencoder", "electrophysiology", "self-supervised learning" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "w6YS9A78fq", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Kangfu Mei", "gender": "unknown", "institution": "Research, Google", "country": "US", "position": "Senior Research Scientist" }, { "name": "Mo Zhou", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Vishal M Patel", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 48, "n_ref": 88, "n_ref_all": 96, "n_fig": 5, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 186, "n_element_tab_1": 43, "formula_len_all": 289, "formula_len_all_1": 289, "len_all": 130983, "len_all_1": 57834, "len_abs": 1118, "len_title": 126, "len_sents": 35096, "len_sents_1": 29102, "n_sents": 239, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1070, "title": "A Simple Diffusion Transformer on Unified Video, 3D, and Game Field Generation", "abs": "The probabilistic field models the distribution of continuous functions defined over metric spaces. While these models hold great potential for unifying data generation across various modalities, including images, videos, and 3D geometry, they still struggle with long-context generation beyond simple examples. This limitation can be attributed to their MLP architecture, which lacks sufficient inductive bias to capture global structures through uniform sampling.\nTo address this, we propose a new and simple model that incorporates a view-wise sampling algorithm to focus on local structure learning, along with autoregressive generation to preserve global geometry. It adapts cross-modality conditions, such as text prompts for text-to-video generation, camera poses for 3D view generation, and control actions for game generation.\nExperimental results across various modalities demonstrate the effectiveness of our model, with its 675M parameter size, and highlight its potential as a foundational framework for scalable, modality-unified visual content generation.", "keywords": [ "Diffusion Probabilistic Fields", "World Model" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "w5pErXbwQl", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Aditya Rajagopal", "gender": "Male", "institution": "FLYR Inc.", "country": "US", "position": "Researcher" }, { "name": "Jeremy Karp", "gender": "unknown", "institution": "Uber", "country": "US", "position": "Researcher" }, { "name": "Lavanya Marla", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Ziyi Chen", "gender": "unknown", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 21, "n_ref": 35, "n_ref_all": 56, "n_fig": 9, "n_tab": 2, "L_tab": 714, "n_element_tab": 37, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 389, "n_element_tab_1": 30, "formula_len_all": 817, "formula_len_all_1": 818, "len_all": 83441, "len_all_1": 48487, "len_abs": 1077, "len_title": 106, "len_sents": 25707, "len_sents_1": 22391, "n_sents": 181, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1087, "title": "Noise-Robust Preference Losses for Deep Regression Models", "abs": "Deep regression models are widely employed for tasks such as pricing and forecasting. In industrial applications, it is common for analysts to adjust model outputs before they are deployed in commercial products. These adjustments, which we name \"analyst influences\", not only ensure the quality of the final products but also provide training data to improve model performance over time. However, due to the huge volumes of data, analyst influences can be applied broadly and can lack precision, hindering training effectiveness. To resolve the issue, we propose a novel framework Preference Learning from Analyst Influence which creates a weighted loss function that explicitly accounts for the relative quality levels of the training samples in comparison to model outputs. This approach effectively mitigates the impact of coarse training instances. Our extensive experiments on real-world data drawn from airline revenue management demonstrate that the proposed framework not only enhances pricing stability but also improves alignment with analyst influences compared to baselines.", "keywords": [ "Regression", "Robustness", "Alignment" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "w5h443GIGo", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Md Nishat Raihan", "gender": "Male", "institution": "George Mason University", "country": "US", "position": "PhD student" } ], "n_formula": 47, "n_formula_1": 46, "n_ref_uni": 11, "n_ref": 27, "n_ref_all": 29, "n_fig": 0, "n_tab": 2, "L_tab": 1553, "n_element_tab": 424, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2185, "formula_len_all_1": 2106, "len_all": 78412, "len_all_1": 65843, "len_abs": 1326, "len_title": 157, "len_sents": 20337, "len_sents_1": 20228, "n_sents": 171, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1330, "title": "On the Convergence of Symbolic Pattern Forests and Silhouette Coefficients for Robust Time Series Clustering", "abs": "Clustering algorithms are fundamental to data mining, serving dual roles as exploratory tools and preprocessing steps for advanced analytics. A persistent challenge in this domain is determining the optimal number of clusters, particularly for time series data where prevalent algorithms like k-means and k-shape require a priori knowledge of cluster quantity. This paper presents the first approach to time series clustering that does not require prior specification of cluster numbers. We introduce a novel extension of the Symbolic Pattern Forest (SPF) algorithm that automatically optimizes the number of clusters for time series datasets. Our method integrates SPF for cluster generation with the Silhouette Coefficient, computed on a two-stage vector representation: first transforming time series into Symbolic Aggregate approXimation (SAX) representations, then deriving both bag-of-words and TF-IDF vectors. Rigorous evaluation on diverse datasets from the UCR archive demonstrates that our approach significantly outperforms traditional baseline methods. This work contributes to the field of time series analysis by providing a truly unsupervised, data-driven approach to clustering, with potential impacts across various temporal data mining applications where the underlying number of clusters is unknown or variable.", "keywords": [ "Data Mining", "Time Series", "Clustering" ], "rating_list": [ 3, 1, 3 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 2, 2, 1 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "w5ZtXOzMeJ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Aaron Roth", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Giuseppe Vietri", "gender": "Male", "institution": "University of Minnesota, Minneapolis", "country": "", "position": "PhD student" }, { "name": "Periklis Petridis", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Tobias Leemann", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Researcher" }, { "name": "Dionysios Manousakas", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Serg\u00fcl Ayd\u00f6re", "gender": "Female", "institution": "Amazon", "country": "", "position": "Researcher" } ], "n_formula": 57, "n_formula_1": 6, "n_ref_uni": 36, "n_ref": 77, "n_ref_all": 123, "n_fig": 10, "n_tab": 7, "L_tab": 2171, "n_element_tab": 74, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3984, "formula_len_all_1": 398, "len_all": 231367, "len_all_1": 69531, "len_abs": 1900, "len_title": 147, "len_sents": 79327, "len_sents_1": 33335, "n_sents": 659, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 108, "L_abs": 1751, "title": "Auto-GDA: Automatic Domain Adaptation for Efficient Grounding Verification in Retrieval Augmented Generation", "abs": "While retrieval augmented generation (RAG) has been shown to enhance factuality of large language model (LLM) outputs, LLMs still suffer from hallucination, generating incorrect or irrelevant information. One common detection strategy involves prompting the LLM again to assess whether its response is grounded in the retrieved evidence, but this approach is costly. Alternatively, lightweight natural language inference (NLI) models for efficient grounding verification can be used at inference time. While existing pre-trained NLI models offer potential solutions, their performance remains subpar compared to larger models on realistic RAG inputs. RAG inputs are more complex than most datasets used for training NLI models and have characteristics specific to the underlying knowledge base, requiring adaptation of the NLI models to a specific target domain. Additionally, the lack of labeled instances in the target domain makes supervised domain adaptation, e.g., through fine-tuning, infeasible. To address these challenges, we introduce Automatic Generative Domain Adaptation (Auto-GDA). Our framework enables unsupervised domain adaptation through synthetic data generation.\nUnlike previous methods that rely on handcrafted filtering and augmentation strategies, Auto-GDA employs an iterative process to continuously improve the quality of generated samples using weak labels from less efficient teacher models and discrete optimization to select the most promising augmented samples. Experimental results demonstrate the effectiveness of our approach, with models fine-tuned on synthetic data using Auto-GDA often surpassing the performance of the teacher model and reaching the performance level of LLMs at 10 % of their computational cost.", "keywords": [ "domain adaptation; NLI; RAG; document-grounded; NLP;" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "w5Q3r8Jq3v", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Mingbao Lin", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Wengyi Zhan", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 12, "n_ref_uni": 51, "n_ref": 104, "n_ref_all": 131, "n_fig": 7, "n_tab": 4, "L_tab": 1143, "n_element_tab": 175, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1121, "n_element_tab_1": 179, "formula_len_all": 622, "formula_len_all_1": 615, "len_all": 137857, "len_all_1": 58609, "len_abs": 1340, "len_title": 112, "len_sents": 38977, "len_sents_1": 26626, "n_sents": 276, "n_sents_1": 169, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1303, "title": "DiffusionTrend: A Minimalist Approach to Virtual Fashion Try-On", "abs": "In this paper, we introduce DiffusionTrend, a pioneering approach for virtual fashion try-on that forgoes the need for training diffusion models, thereby offering simple, conventional pose virtual try-on services with significantly reduced computational overhead. By leveraging advanced diffusion models, DiffusionTrend harnesses latents rich with prior information to capture the nuances of garment details. Throughout the diffusion denoising process, these details are seamlessly integrated into the model image generation, expertly directed by a precise garment mask crafted by a lightweight and compact CNN. Although our DiffusionTrend model initially demonstrates suboptimal metric performance, our exploratory approach offers several significant advantages: (1) It circumvents the need for resource-intensive training of diffusion models on large datasets. (2) It eliminates the necessity for various complex and user-unfriendly model inputs. (3) It delivers a visually compelling virtual try-on experience, underscoring the potential of training-free diffusion models for future research within the community. Overall, this initial foray into the application of untrained diffusion models in virtual try-on technology paves the way for further exploration and refinement in this innovative field.", "keywords": [ "Virtual Try-on; Diffusion Model; Image Editing" ], "rating_list": [ 3, 5, 1, 5 ], "soundness_list": [ 1, 3, 1, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "w4gkS9RsWh", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Cheng-Yu Lin", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Da-Cheng Juan", "gender": "unknown", "institution": "National Tsinghua University", "country": "TW", "position": "Assistant Professor" }, { "name": "Heng-Yi Liu", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Shou-De Lin", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "Full Professor" }, { "name": "TZU-LING CHENG", "gender": "Female", "institution": "National Taiwan University", "country": "", "position": "PhD student" }, { "name": "Yu-Chieh Tu", "gender": "Male", "institution": "Department of computer science and informational engineering, National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "YiFu Fu", "gender": "unknown", "institution": "National Taiwan University", "country": "", "position": "Researcher" }, { "name": "Yiting Yang", "gender": "Male", "institution": " and informational engineering, National Taiwan University", "country": "TW", "position": "Undergrad student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 20, "n_ref": 28, "n_ref_all": 39, "n_fig": 7, "n_tab": 2, "L_tab": 301, "n_element_tab": 44, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 301, "n_element_tab_1": 44, "formula_len_all": 57, "formula_len_all_1": 56, "len_all": 74658, "len_all_1": 52308, "len_abs": 1040, "len_title": 135, "len_sents": 27762, "len_sents_1": 26059, "n_sents": 174, "n_sents_1": 162, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1042, "title": "Think or Remember? Detecting and Directing LLMs Towards Memorization or Generalization", "abs": "In this paper, we study fundamental mechanisms of memorization and generalization in Large Language Models (LLMs), drawing inspiration from the functional specialization observed in the human brain. Our study aims to (a) determine whether LLMs exhibit spatial differentiation of neurons for memorization and generalization, (b) predict these behaviors using internal representations, and (c) control them through inference-time interventions. To achieve this, we design specialized datasets to distinguish between memorization and generalization, build up classifiers to predict these behaviors from model hidden states and develop interventions to influence the model in real time. Our experiments reveal that LLMs exhibit neuron-wise differentiation for memorization and generalization, and the proposed intervention mechanism successfully steers the model's behavior as intended. These findings significantly advance the understanding of LLM behavior and demonstrate the potential for enhancing the reliability and controllability of LLMs.", "keywords": [ "LLM", "generalization", "memorization", "neuron differentiation", "behavior identification", "inference-time intervention", "behavior control" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "w4C4z80w59", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Cen Chen", "gender": "Female", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Die Chen", "gender": "Female", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Mingyuan Fan", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "PhD student" }, { "name": "Wenmeng Zhou", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yaliang Li", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Staff Engineer" }, { "name": "Yanhao Wang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhiwen Li", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 12, "n_ref_uni": 35, "n_ref": 65, "n_ref_all": 126, "n_fig": 19, "n_tab": 5, "L_tab": 1782, "n_element_tab": 206, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1194, "n_element_tab_1": 63, "formula_len_all": 1138, "formula_len_all_1": 671, "len_all": 146063, "len_all_1": 63095, "len_abs": 1671, "len_title": 131, "len_sents": 44525, "len_sents_1": 28243, "n_sents": 344, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1615, "title": "Growth Inhibitors for Suppressing Inappropriate Image Concepts in Diffusion Models", "abs": "Despite their remarkable image generation capabilities, text-to-image diffusion models inadvertently learn inappropriate concepts from vast and unfiltered training data, which leads to various ethical and business risks. Specifically, model-generated images may exhibit not safe for work (NSFW) content and style copyright infringements. The prompts that result in these problems often do not include explicit unsafe words; instead, they contain obscure and associative terms, which are referred to as *implicit unsafe prompts*. Existing approaches directly fine-tune models under textual guidance to alter the cognition of the diffusion model, thereby erasing inappropriate concepts. This not only requires concept-specific fine-tuning but may also incur catastrophic forgetting. To address these issues, we explore the representation of inappropriate concepts in the image space and guide them towards more suitable ones by injecting *growth inhibitors*, which are tailored based on the identified features related to inappropriate concepts during the diffusion process. Additionally, due to the varying degrees and scopes of inappropriate concepts, we train an adapter to infer the corresponding suppression scale during the injection process. Our method effectively captures the manifestation of subtle words at the image level, enabling direct and efficient erasure of target concepts without the need for fine-tuning. Through extensive experimentation, we demonstrate that our approach achieves superior erasure results with little effect on other normal concepts while preserving image quality and semantics.", "keywords": [ "Stable Diffusion", "Text-to-Image Generation", "Concept Erasure" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "w3rbBVJ9Jg", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Hao Sun", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Qi Wang", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Wan Han", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 8, "n_ref_uni": 31, "n_ref": 63, "n_ref_all": 106, "n_fig": 7, "n_tab": 4, "L_tab": 335, "n_element_tab": 62, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 144, "n_element_tab_1": 12, "formula_len_all": 750, "formula_len_all_1": 343, "len_all": 129919, "len_all_1": 62626, "len_abs": 1437, "len_title": 133, "len_sents": 44721, "len_sents_1": 30927, "n_sents": 337, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1443, "title": "PIMRL: Physics-Informed Multi-Scale Recurrent Learning for Spatiotemporal Prediction", "abs": "Simulation of spatiotemporal systems governed by partial differential equations is widely applied in fields such as biology, chemistry, aerospace dynamics, and meteorology. Traditional numerical methods incur high computational costs due to the requirement of small time steps for accurate predictions. While machine learning has reduced these costs, long-term predictions remain challenged by error accumulation, particularly in scenarios with insufficient data or varying time scales, where stability and accuracy are compromised. Existing methods often neglect the effective utilization of multi-scale data, leading to suboptimal robustness in predictions. To address these issues, we propose a novel multi-scale learning framework, namely, the Physics-Informed Multi-Scale Recurrent Learning (PIMRL), to effectively leverage multi-scale data for spatiotemporal dynamics prediction. The PIMRL framework comprises two modules: the micro-scale module embeds physical knowledge into neural networks via pretraining, and the macro-scale module adopts a data-driven approach to learn the temporal evolution of physics in the latent space. Experimental results demonstrate that the PIMRL framework consistently achieves state-of-the-art performance across five benchmark datasets ranging from one to three dimensions, showing average improvements of over 9\\% in both RMSE and MAE evaluation metrics, with maximum enhancements reaching up to 80\\%.", "keywords": [ "PDEs", "physics encoding", "data-driven modeling" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 2, 3 ] }, { "paper_id": "w3iM4WLuvy", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Devdhar Patel", "gender": "Male", "institution": "Department of Computer Science, University of Massachusetts, Amherst", "country": "", "position": "PhD student" }, { "name": "Hava T Siegelmann", "gender": "Female", "institution": "University of Massachusetts at Amherst", "country": "", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 5, "n_ref_uni": 72, "n_ref": 101, "n_ref_all": 117, "n_fig": 8, "n_tab": 6, "L_tab": 2997, "n_element_tab": 228, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1452, "n_element_tab_1": 95, "formula_len_all": 695, "formula_len_all_1": 401, "len_all": 194220, "len_all_1": 62961, "len_abs": 1710, "len_title": 126, "len_sents": 44318, "len_sents_1": 29183, "n_sents": 333, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 126, "L_abs": 1813, "title": "Overcoming Slow Decision Frequencies in Continuous Control: Model-Based Sequence Reinforcement Learning for Model-Free Control", "abs": "Reinforcement learning (RL) is rapidly reaching and surpassing human-level control capabilities. However, state-of-the-art RL algorithms often require timesteps and reaction times significantly faster than human capabilities, which is impractical in real-world settings and typically necessitates specialized hardware. Such speeds are difficult to achieve in the real world and often requires specialized hardware. We introduce Sequence Reinforcement Learning (SRL), an RL algorithm designed to produce a sequence of actions for a given input state, enabling effective control at lower decision frequencies. SRL addresses the challenges of learning action sequences by employing both a model and an actor-critic architecture operating at different temporal scales. We propose a \"temporal recall\" mechanism, where the critic uses the model to estimate intermediate states between primitive actions, providing a learning signal for each individual action within the sequence. Once training is complete, the actor can generate action sequences independently of the model, achieving model-free control at a slower frequency. We evaluate SRL on a suite of continuous control tasks, demonstrating that it achieves performance comparable to state-of-the-art algorithms while significantly reducing actor sample complexity. To better assess performance across varying decision frequencies, we introduce the Frequency-Averaged Score (FAS) metric. Our results show that SRL significantly outperforms traditional RL algorithms in terms of FAS, making it particularly suitable for applications requiring variable decision frequencies. Additionally, we compare SRL with model-based online planning, showing that SRL achieves comparable FAS while leveraging the same model during training that online planners use for planning.", "keywords": [ "Decision Frequency", "Action Sequence Generation", "Model-Based Training", "Model-Free Control", "Efficient Learning", "Reinforcement Learning" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "w2uIJiHTIA", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Atsushi Miyauchi", "gender": "Male", "institution": "CENTAI Institute", "country": "IT", "position": "Researcher" }, { "name": "Florian Adriaens", "gender": "unknown", "institution": "University of Helsinki, University of Helsinki", "country": "FI", "position": "Postdoc" }, { "name": "Francesco Bonchi", "gender": "Male", "institution": "Centai", "country": "IT", "position": "Principal Researcher" }, { "name": "Nikolaj Tatti", "gender": "unknown", "institution": "University of Helsinki", "country": "FI", "position": "Full Professor" } ], "n_formula": 55, "n_formula_1": 16, "n_ref_uni": 50, "n_ref": 129, "n_ref_all": 134, "n_fig": 0, "n_tab": 3, "L_tab": 108, "n_element_tab": 8, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 31, "n_element_tab_1": 3, "formula_len_all": 7219, "formula_len_all_1": 1792, "len_all": 182394, "len_all_1": 72284, "len_abs": 1561, "len_title": 81, "len_sents": 59034, "len_sents_1": 33782, "n_sents": 455, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 33, "L_abs": 1252, "title": "Multilayer Correlation Clustering", "abs": "We establish Multilayer Correlation Clustering, a novel generalization of Correlation Clustering to the multilayer setting. In this model, we are given a series of inputs of Correlation Clustering (called layers) over the common set $V$ of $n$ elements. The goal is to find a clustering of $V$ that minimizes the $\\ell_p$-norm ($p\\geq 1$) of the disagreements vector, which is defined as the vector (with dimension equal to the number of layers), each element of which represents the disagreements of the clustering on the corresponding layer. For this generalization, we first design an $O(L\\log n)$-approximation algorithm, where $L$ is the number of layers. We then study an important special case of our problem, namely the problem with the so-called probability constraint. For this case, we first give an $(\\alpha+2)$-approximation algorithm, where $\\alpha$ is any possible approximation ratio for the single-layer counterpart. Furthermore, we design a $4$-approximation algorithm, which improves the above approximation ratio of $\\alpha+2=4.5$ for the general probability-constraint case. Computational experiments using real-world datasets support our theoretical findings and demonstrate the practical effectiveness of our proposed algorithms.", "keywords": [ "Clustering", "Correlation Clustering", "Multilayer Networks", "Approximation Algorithms" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "w2qzdlvPMK", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Jiamu Sun", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Ruoxin Chen", "gender": "Male", "institution": "Tencent YouTu Lab", "country": "CN", "position": "Researcher" }, { "name": "Shouhong Ding", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "researcher" }, { "name": "Shouli Wang", "gender": "unknown", "institution": "East China University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Shuang Wu", "gender": "Male", "institution": "Tencent YouTu Lab", "country": "CN", "position": "Researcher" }, { "name": "Taiping Yao", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "researcher" }, { "name": "Zhe Wang", "gender": "Male", "institution": "East China University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Keyue Zhang", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 27, "n_ref": 61, "n_ref_all": 73, "n_fig": 5, "n_tab": 10, "L_tab": 3445, "n_element_tab": 405, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2701, "n_element_tab_1": 313, "formula_len_all": 171, "formula_len_all_1": 170, "len_all": 107487, "len_all_1": 64863, "len_abs": 571, "len_title": 110, "len_sents": 31374, "len_sents_1": 27427, "n_sents": 230, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1571, "title": "Decoupled Data Augmentation for Improving Image Classification", "abs": "Recent advancements in image mixing and generative data augmentation have shown promise in enhancing image classification. However, these techniques face the challenge of balancing semantic fidelity with diversity. Specifically, image mixing involves interpolating two images to create a new one, but this pixel-level interpolation can compromise fidelity. Generative augmentation uses text-to-image generative models to synthesize or modify images, often limiting diversity to avoid generating out-of-distribution data that potentially affects accuracy. We propose that this fidelity-diversity dilemma partially stems from the whole-image paradigm of existing methods. Since an image comprises the class-dependent part (CDP) and the class-independent part (CIP), where each part has fundamentally different impacts on the image's fidelity, treating different parts uniformly can therefore be misleading. To address this fidelity-diversity dilemma, we introduce Decoupled Data Augmentation (De-DA), which resolves the dilemma by separating images into CDPs and CIPs and handling them adaptively. To maintain fidelity, we use generative models to modify real CDPs under controlled conditions, preserving semantic consistency. To enhance diversity, we replace the image's CIP with inter-class variants, creating diverse CDP-CIP combinations. Additionally, we implement an online randomized combination strategy during training to generate numerous distinct CDP-CIP combinations cost-effectively. Comprehensive empirical evaluations validate the effectiveness of our method.", "keywords": [ "Data augmentation", "Diffusion", "Image classification" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "w2HYVwXhMh", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "FU-JEN CHU", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Huiyu Wang", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Jing Huang", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Kris M. Kitani", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Nanyun Peng", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Associate Professor" }, { "name": "Tushar Nagarajan", "gender": "unknown", "institution": "FAIR", "country": "US", "position": "Researcher" }, { "name": "Xitong Yang", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Zi-Yi Dou", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 47, "n_ref": 150, "n_ref_all": 177, "n_fig": 6, "n_tab": 5, "L_tab": 2003, "n_element_tab": 255, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1908, "n_element_tab_1": 260, "formula_len_all": 74, "formula_len_all_1": 74, "len_all": 175483, "len_all_1": 70080, "len_abs": 1460, "len_title": 133, "len_sents": 53841, "len_sents_1": 32096, "n_sents": 420, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1469, "title": "Unlocking Exocentric Video-Language Data for Egocentric Video Representation Learning", "abs": "We present EMBED (Egocentric Models Built with Exocentric Data), a framework designed to mine video-language data from exocentric sources for egocentric video representation learning. Large-scale exocentric data covers diverse activities with significant potential for egocentric learning, but inherent disparities between egocentric and exocentric data pose challenges in utilizing one view for the other seamlessly. In this study, we propose leveraging hand-object interactions and language narratives as cues to incorporate exocentric data into egocentric training. Specifically, we focus on identifying specific video clips that emphasize hand-object interactions and pairing them with action-focused language narrations. By applying our framework to exocentric datasets such as HowTo100M, we construct datasets thar are effective for egocentric video-language pretraining. Our extensive evaluations reveal that EMBED achieves state-of-the-art performance across various egocentric downstream tasks, including a 4.7\\% absolute improvement in multi-instance retrieval on the Epic-Kitchens-100 benchmark and a 6.2\\% improvement in classification on the EGTEA benchmark in zero-shot settings. Furthermore, EMBED enables egocentric video-language models to perform competitively in exocentric tasks. Finally, we showcase EMBED's application across various exocentric datasets, exhibiting strong generalization capabilities when applied to different exocentric datasets.", "keywords": [ "video-language pretraining", "egocentric video" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "w2HL7yuWE2", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Haozhe Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Jiancan Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Postdoc" }, { "name": "Wenyu Mao", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Xiang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xiangnan He", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Yancheng Yuan", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "Zhengyi Yang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 25, "n_ref_uni": 48, "n_ref": 114, "n_ref_all": 151, "n_fig": 6, "n_tab": 6, "L_tab": 2178, "n_element_tab": 352, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 619, "n_element_tab_1": 74, "formula_len_all": 2275, "formula_len_all_1": 1511, "len_all": 176515, "len_all_1": 63477, "len_abs": 1356, "len_title": 128, "len_sents": 49805, "len_sents_1": 27849, "n_sents": 380, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1362, "title": "Uncertainty-aware Guided Diffusion for Missing Data in Sequential Recommendation", "abs": "Denoising diffusion models (DDMs) have shown significant potential in generating oracle items that best match user preference with guidance from user historical interaction sequences. However, the quality of guidance is often compromised by the unpredictable missing data in the observed sequence, leading to suboptimal item generation. To tackle this challenge, we propose a novel uncertainty-aware guided diffusion model (DreamMiss) to alleviate the influence of missing data. The core of DreamMiss is the utilization of a dual-side Thompson sampling (DTS) strategy, which simulates the stochastical mechanism of missing data without disrupting preference evolution. Specifically, we first define dual-side probability models to capture user preference evolution, taking into account both local item continuity and global sequence stability. We then strategically remove items based on these two models with DTS, creating uncertainty-aware guidance for DDMs to generate oracle items. This can achieve DDMs\u2019 consistency regularization, enabling them to resile against uncertain missing data. Additionally, to accelerate sampling in the reverse process, DreamMiss is implemented under the framework of denoising diffusion implicit models (DDIM). Extensive experimental results show that DreamMiss significantly outperforms baselines in sequential recommendation.", "keywords": [ "Diffusion Models", "Recommender Systems", "Missing Data" ], "rating_list": [ 5, 10, 5, 6 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 3, 1, 4, 2 ] }, { "paper_id": "w2C7gJqaai", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Andy Song", "gender": "unknown", "institution": "RMIT University", "country": "AU", "position": "Associate Professor" }, { "name": "Beinan Xu", "gender": "Male", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "PhD student" }, { "name": "Jiti Gao", "gender": "unknown", "institution": "Monash University", "country": "", "position": "Full Professor" } ], "n_formula": 37, "n_formula_1": 19, "n_ref_uni": 34, "n_ref": 50, "n_ref_all": 90, "n_fig": 5, "n_tab": 13, "L_tab": 26056, "n_element_tab": 862, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2319, "n_element_tab_1": 27, "formula_len_all": 2659, "formula_len_all_1": 1525, "len_all": 190337, "len_all_1": 58630, "len_abs": 1411, "len_title": 115, "len_sents": 39381, "len_sents_1": 25392, "n_sents": 341, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1422, "title": "Integrated Multi-system Prediction via Equilibrium State Evaluation", "abs": "This study presents a new paradigm of prediction, Equilibrium State Evaluation (ESE), which excels in multi-system prediction where systems interact with each other and every system needs its own prediction. Unlike mainstream prediction approaches, ESE views each system as an integral part under one structure and predicts all systems simultaneously in one go. It evaluates these systems' equilibrium state by analyzing the dynamics of their attributes in a holistic manner, instead of treating each system as an individual time series. The effectiveness of ESE is verified in synthetic and real world scenarios, in particular COVID-19 transmission, where each geographic region can be viewed as a system. So cases spreading across regions against the medical competency and demographic traits of these regions can be considered as an equilibrium problem rather than a time series problem. Extensive analysis and experiments show that ESE is linear in complexity and can be 10+ times faster than SOTA methods, yet achieving comparable or better prediction accuracy. More importantly, ESE can be integrated with these prediction methods to achieve both high accuracy and high speed, making it a powerful prediction mechanism, especially for scenarios that involve multiple systems. When the dimensionality of the multi-system increases, e.g. more systems joining, the advantages of ESE would become even more apparent.", "keywords": [ "Multi-system", "Equilibrium", "Prediction" ], "rating_list": [ 1, 5, 1 ], "soundness_list": [ 1, 2, 1 ], "presentation_list": [ 1, 3, 1 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 5, 4, 1 ] }, { "paper_id": "w2BELPYbU0", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Chris Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Le Zhuo", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Qi Qin", "gender": "Not Specified", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Ruoyi Du", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "PhD student" }, { "name": "Zhanyu Ma", "gender": "Male", "institution": "Beijing University of Post and Telecommunication", "country": "CN", "position": "Full Professor" }, { "name": "Gao Peng", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 21, "n_ref": 34, "n_ref_all": 56, "n_fig": 13, "n_tab": 1, "L_tab": 195, "n_element_tab": 7, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 547, "formula_len_all_1": 454, "len_all": 107635, "len_all_1": 56925, "len_abs": 1019, "len_title": 151, "len_sents": 32902, "len_sents_1": 27669, "n_sents": 219, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1023, "title": "I-Max: Maximize the Resolution Potential of Pre-trained Rectified Flow Transformers with Projected Flow", "abs": "Rectified Flow Transformers (RFTs) offer superior training and inference efficiency, making them likely the most viable direction for scaling up diffusion models. However, progress in generation resolution has been relatively slow due to data quality and training costs. Tuning-free resolution extrapolation presents an alternative, but current methods often reduce generative stability, limiting practical application. In this paper, we review existing resolution extrapolation methods and introduce the I-Max framework to maximize the resolution potential of Text-to-Image RFTs. I-Max features: (i) a novel Projected Flow strategy for stable extrapolation and (ii) an advanced inference toolkit for generalizing model knowledge to higher resolutions. Experiments with Lumina-Next-2K and Flux.1-dev demonstrate I-Max's ability to enhance stability in resolution extrapolation and show that it can bring image detail emergence and artifact correction, confirming the practical value of tuning-free resolution extrapolation.", "keywords": [ "Diffusion Model", "Generative Model", "Image Generation", "High-resolution" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "w1Pwcx5hPp", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Guanghao Li", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Jian Pu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Qi Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Sijia Hu", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Yuxiang Yan", "gender": "unknown", "institution": "Fudan University", "country": "", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 14, "n_ref_uni": 58, "n_ref": 145, "n_ref_all": 157, "n_fig": 7, "n_tab": 13, "L_tab": 2654, "n_element_tab": 257, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2468, "n_element_tab_1": 178, "formula_len_all": 1942, "formula_len_all_1": 818, "len_all": 169222, "len_all_1": 56795, "len_abs": 1180, "len_title": 98, "len_sents": 40878, "len_sents_1": 23043, "n_sents": 332, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1184, "title": "Geometrically Constrained Gaussian Splatting SLAM", "abs": "3D Gaussian Splatting (3DGS) has emerged as a promising technique in SLAM due to its rapid and high-quality rendering capabilities. However, its reliance on discrete Gaussian ellipsoid primitives limits its effectiveness in capturing essential geometric features crucial for accurate pose estimation. To overcome this limitation, we propose a novel dense RGB-D SLAM system that integrates an implicit Truncated Signed Distance Function (TSDF) hash grid to constrain the distribution of Gaussian ellipsoids. This innovative approach enables precise estimation of the scene's geometric structure by smoothing the discrete Gaussian ellipsoids and anchoring them to the scene's surface. Acting as a low-pass filter, the implicit TSDF hash grid mitigates the inductive biases inherent in traditional 3DGS methods while preserving rendering quality. Our geometrically constrained map also significantly enhances generalization capabilities for depth estimation in novel views. Extensive experiments on the Replica, ScanNet, and TUM datasets demonstrate that our system achieves state-of-the-art tracking and mapping accuracy at speeds up to 30 times faster than existing 3DGS-based systems.", "keywords": [ "3DGS", "SLAM", "Robotics" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 3, 5, 4 ] }, { "paper_id": "w1MEIGDepc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Deqing Yang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Gang Li", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Hang Shao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Jiawei Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Ke Li", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Senior Researcher" }, { "name": "Siqi Cai", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Xing Sun", "gender": "Male", "institution": "Tencent YouTu Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuchen Shi", "gender": "Male", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Yulei Qin", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Zihan Xu", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 37, "n_ref": 73, "n_ref_all": 90, "n_fig": 7, "n_tab": 7, "L_tab": 5825, "n_element_tab": 532, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 205, "formula_len_all_1": 83, "len_all": 187863, "len_all_1": 61851, "len_abs": 1669, "len_title": 92, "len_sents": 47845, "len_sents_1": 31096, "n_sents": 346, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1676, "title": "FlowAgent: a New Paradigm for Workflow Agent", "abs": "Combining workflows with large language models (LLMs) allows LLMs to follow specific procedures, thereby extending their application to more real-world scenarios. However, incorporating workflows often compromises the flexibility of LLMs. For example in the case of Task-Oriented Dialogue (TOD), workflow atomize the function of LLM while programmatically imposing restrictions on execution path making the dialogue obstructed and less flexible when facing out-of-workflow (OOW) queries. Prompt-based methods offer soft control but sometimes fail to ensure procedure compliance. This paper introduces a new agent paradigm to address this challenge. Specifically, we first propose a novel Procedure Description Language (PDL) that integrates the flexibility of natural language and the precision of code for workflow expression. Additionally, we present a comprehensive framework that enables LLM to handle OOW queries while keeping execution safe with a series of controllers for behavioral regulation. This includes pre-decision and post-decision methods, where the dependency relationships between workflow nodes are modeled as a Directed Acyclic Graph (DAG) to validate node transitions. Beyond the primary objective of compliance considered in previous work, we introduce a new approach to evaluate the agent's flexibility in OOW situations. Experiments on three datasets demonstrate that FlowAgent not only adheres well to workflows but also responds better to OOW queries, showcasing its flexibility. Furthermore, exploration on WikiHow data confirms that the PDL effectively represents broader formats of workflow, inspiring further research on workflow-based QA tasks.", "keywords": [ "workflow", "LLM-based agent", "task-oriented dialog" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "w10KdRwcMk", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Moshe Butman", "gender": "unknown", "institution": "Reichman University", "country": "", "position": "Lecturer" }, { "name": "Nir Zvi Weingarten", "gender": "Male", "institution": "Eikona Technologies Inc.", "country": "IL", "position": "CEO" }, { "name": "Ronit Bustin", "gender": "Female", "institution": "General Motors", "country": "", "position": "Researcher" }, { "name": "Zohar Yakhini", "gender": "Male", "institution": "The Interdisciplinary Center Herzliya", "country": "IL", "position": "Associate Professor" } ], "n_formula": 21, "n_formula_1": 8, "n_ref_uni": 26, "n_ref": 73, "n_ref_all": 92, "n_fig": 5, "n_tab": 9, "L_tab": 2650, "n_element_tab": 485, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1506, "formula_len_all_1": 374, "len_all": 125544, "len_all_1": 47296, "len_abs": 803, "len_title": 49, "len_sents": 40785, "len_sents_1": 23290, "n_sents": 269, "n_sents_1": 137, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 805, "title": "Revisiting the Variational Information Bottleneck", "abs": "The Information Bottleneck (IB) framework offers a theoretically optimal approach to data modeling, though it is often intractable. Recent efforts have optimized supervised deep neural networks (DNNs) using a variational upper bound on the IB objective, leading to enhanced robustness to adversarial attacks. In these studies, supervision assumes a dual role: sometimes as a presumably constant and observed random variable, and at other times as its variational approximation. This work proposes an extension to the IB framework, and consequently to the derivation of its variational bound, that resolves this duality. Applying the resulting bound as an objective for supervised DNNs induces significant empirical improvements, and provides an information theoretic motivation for decoder regularization.", "keywords": [ "information bottleneck", "information theory", "representation learning", "adversarial attacks", "regularization", "supervised learning" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 2, 1, 1 ], "presentation_list": [ 3, 3, 1, 1 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "w0lhe9prqH", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Agneet Chatterjee", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Amir Saeidi", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Bimsara Pathiraja", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Chitta Baral", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Full Professor" }, { "name": "Shamanthak Hegde", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "MS student" }, { "name": "Yezhou Yang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Associate Professor" }, { "name": "Yiran Luo", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 4, "n_ref_uni": 23, "n_ref": 44, "n_ref_all": 73, "n_fig": 9, "n_tab": 7, "L_tab": 2769, "n_element_tab": 387, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2697, "n_element_tab_1": 314, "formula_len_all": 1291, "formula_len_all_1": 613, "len_all": 156349, "len_all_1": 73874, "len_abs": 3392, "len_title": 105, "len_sents": 41789, "len_sents_1": 28811, "n_sents": 328, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1421, "title": "Dual Caption Preference Optimization for Diffusion Models", "abs": "Recent advancements in human preference optimization, originally developed for Large Language Models (LLMs), have shown significant potential in improving text-to-image diffusion models. These methods aim to learn the distribution of preferred samples while distinguishing them from less preferred ones. However, existing preference datasets often exhibit overlap between these distributions, leading to a conflict distribution. Additionally, we identified a performance issue in previous optimization methods, where using the same prompt for preferred and less preferred images, known as the irrelevant prompt issue, restricts model performance. To address these challenges, we propose Dual Caption Preference Optimization (DCPO), a novel approach that utilizes two distinct captions to mitigate irrelevant prompts. To tackle conflict distribution, we introduce the Pick-Double Caption dataset, a modified version of Pick-a-Pic v2 with separate captions for preferred and less preferred images. We further propose three different strategies for generating distinct captions: captioning, perturbation, and hybrid methods. Our experiments show that DCPO significantly improves image quality and relevance to prompts, outperforming Stable Diffusion (SD) 2.1, SFT-Chosen, Diffusion-DPO and MaPO across multiple metrics, including Pickscore, HPSv2.1, GenEval, CLIPscore, and ImageReward, fine-tuned on SD 2.1 as the backbone.", "keywords": [ "Preference Optimization", "Diffusion Models", "Alignment" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "w0jk3L3IjV", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Jiangchao Yao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Lu Zhang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Associate Professor" }, { "name": "Xuan Li", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "CN", "position": "PhD student" }, { "name": "YiXiao ZHENG", "gender": "Female", "institution": "Hong Kong Baptist University", "country": "", "position": "Undergrad student" }, { "name": "Zhanke Zhou", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Visiting student" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 65, "n_ref": 137, "n_ref_all": 150, "n_fig": 8, "n_tab": 13, "L_tab": 11409, "n_element_tab": 850, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 3155, "n_element_tab_1": 278, "formula_len_all": 1014, "formula_len_all_1": 1040, "len_all": 218410, "len_all_1": 68808, "len_abs": 1438, "len_title": 121, "len_sents": 51771, "len_sents_1": 27175, "n_sents": 511, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 775, "title": "Breaking the Detection-Generalization Paradox on Out-Of-Distribution Data", "abs": "This work studies the trade-off between out-of-distribution (OOD) detection and generalization. We identify the Detection-Generalization Paradox in OOD data, where optimizing one objective can degrade the other. We investigate this paradox by analyzing the behaviors of models trained under different paradigms, focusing on representation, logits, and loss across in-distribution, covariate-shift, and semantic-shift data. Based on our findings, we propose Distribution-Robust Sharpness-Aware Minimization (DR-SAM), an optimization framework that balances OOD detection and generalization. Extensive experiments demonstrate the method's effectiveness, offering a clear, empirically validated approach for improving detection and generalizationability in different benchmarks.", "keywords": [ "Trustworthy Machine Learning; Out of distribution data" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 1, 3, 2 ], "confidence_list": [ 4, 5, 2 ] }, { "paper_id": "w0es2hinsd", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Haotian Chen", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Haoxue Wang", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Jiang Bian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Weiqing Liu", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Postdoc" }, { "name": "Wenjun Feng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "Undergrad student" }, { "name": "Xiao Yang", "gender": "Male", "institution": "Research, Microsoft", "country": "", "position": "Researcher" }, { "name": "Xinjie Shen", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Xu Yang", "gender": "Male", "institution": "Microsoft", "country": "", "position": "RSDE" }, { "name": "Zeqi Ye", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 39, "n_ref": 74, "n_ref_all": 91, "n_fig": 3, "n_tab": 4, "L_tab": 1673, "n_element_tab": 241, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1557, "n_element_tab_1": 120, "formula_len_all": 324, "formula_len_all_1": 218, "len_all": 179839, "len_all_1": 70748, "len_abs": 1840, "len_title": 94, "len_sents": 42940, "len_sents_1": 32850, "n_sents": 331, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1617, "title": "RD2Bench: Toward Data-Centric Automatic R&D", "abs": "The progress of humanity is driven by those successful discoveries accompanied by countless failed experiments. Researchers often seek potential solutions described in related literature (raw information) and verify them through experiments. With the explosive growth of deep learning literature and methods, such a process imposes a more significant burden on researchers and renders successful discoveries veiled. Therefore, automating such a research and development (R&D) process is an urgent need. In this paper, we serve as the first effort to formalize the goal by proposing a **R**eal-world **D**ata-centric automatic **R**&**D** **Bench**mark, namely RD2Bench. RD2Bench benchmarks the whole data-centric automatic R&D (D-CARD) process, including extracting methods (formulas and models) from raw information (reports and papers) and implementing methods through codes. Specifically, to investigate the capability boundaries of the state-of-the-art (SOTA) large language models (LLMs) in the unexplored D-CARD, we conduct exhausting and expensive human annotations and experiments. We evaluate the performance of SOTA LLMs on our identified 27 formulas and 6 models across various difficulty levels from financial reports and ML papers. We find that although RD2Bench is very challenging, SOTA LLMs possess promising potential to bring more significant development to D-CARD. We appeal to research teams with various domain expertise to consider constructing domain-specific D-CARD benchmarks, contributing to both a cross-domain D-CARD platform and the potential revolutionary upgrade to human productivity.}", "keywords": [ "Real-world Data-centric automatic R&D Benchmark", "data-centric automatic R&D", "trustworthy models" ], "rating_list": [ 5, 5, 8, 3 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 4, 2 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "w0b7fCX2nN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Grigorios Chrysos", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Markos Georgopoulos", "gender": "unknown", "institution": "Meta", "country": "CH", "position": "Researcher" }, { "name": "Volkan Cevher", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" }, { "name": "Yixin Cheng", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 50, "n_ref": 128, "n_ref_all": 148, "n_fig": 5, "n_tab": 10, "L_tab": 1049, "n_element_tab": 123, "n_fig_1": 0, "n_tab_1": 6, "L_tab_1": 1105, "n_element_tab_1": 141, "formula_len_all": 28, "formula_len_all_1": 28, "len_all": 262784, "len_all_1": 72451, "len_abs": 1244, "len_title": 128, "len_sents": 90565, "len_sents_1": 34828, "n_sents": 709, "n_sents_1": 257, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1256, "title": "Leveraging the Context through Multi-Round Interactions for Jailbreaking Attacks", "abs": "Large Language Models (LLMs) are susceptible to Jailbreaking attacks, which\naim to extract harmful information by subtly modifying the attack query. As de-\nfense mechanisms evolve, directly obtaining harmful information becomes increas-\ningly challenging for Jailbreaking attacks. In this work, inspired from Chomsky\u2019s\ntransformational-generative grammar theory and human practices of indirect con-\ntext to elicit harmful information, we focus on a new attack form, called Contextual\nInteraction Attack. We contend that the prior context\u2014the information preced-\ning the attack query\u2014plays a pivotal role in enabling strong Jailbreaking attacks.\nSpecifically, we propose first multi-turn approach that leverages benign preliminary\nquestions to interact with the LLM. Due to the autoregressive nature of LLMs,\nwhich use previous conversation rounds as context during generation, we guide the\nmodel\u2019s question-responses pair to construct a context that is semantically aligned\nwith the attack query to execute the attack. We conduct experiments on seven\ndifferent LLMs and demonstrate the efficacy of this attack, which is black-box, and\ncan also transfer across LLMs. We believe this can lead to further developments\nand understanding of the security in LLMs", "keywords": [ "LLM jailbreak", "trustworthy ML", "safety AI" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 1, 4, 2 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "w0MAu8vjwj", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Feijun Jiang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Kaifu Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "vice president" }, { "name": "Liu Haoxiang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Wanli Jiang", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "MS student" }, { "name": "Weihua Luo", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 14, "n_ref_uni": 16, "n_ref": 38, "n_ref_all": 52, "n_fig": 5, "n_tab": 7, "L_tab": 2258, "n_element_tab": 318, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1257, "n_element_tab_1": 205, "formula_len_all": 670, "formula_len_all_1": 755, "len_all": 108654, "len_all_1": 56211, "len_abs": 1396, "len_title": 126, "len_sents": 32109, "len_sents_1": 25135, "n_sents": 226, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1400, "title": "MOSLIM:Align with diverse preferences in prompts through reward classification", "abs": "The multi-objective alignment of Large Language Models (LLMs) is essential for ensuring foundational models conform to diverse human preferences. Current research in this field typically involves either multiple policies or multiple reward models customized for various preferences, or the need to train a preference-specific supervised fine-tuning (SFT) model. In this work, we introduce a novel multi-objective alignment method, MOSLIM, which utilizes a single reward model and policy model to address diverse objectives. MOSLIM provides a flexible way to control these objectives through prompting and does not require preference training during SFT phase, allowing thousands of off-the-shelf models to be directly utilized within this training framework.\nMOSLIM leverages a multi-head reward model that classifies question-answer pairs instead of scoring them and then optimize policy model with a scalar reward derived from a mapping function that converts classification results from reward model into reward scores. We demonstrate the efficacy of our proposed method across several multi-objective benchmarks and conduct ablation studies on various reward model sizes and policy optimization methods. The MOSLIM method outperforms current multi-objective approaches in most results while requiring significantly fewer GPU computing resources compared with existing policy optimization methods.", "keywords": [ "Large Language Models", "Multi-objective alignment", "Reward modeling" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "w0389y0W9D", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Lifeng Shang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Tao Li", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Xiaolin Huang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Yasheng Wang", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "", "position": "Researcher" }, { "name": "Yujun Li", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Zhengbao He", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 35, "n_ref": 90, "n_ref_all": 120, "n_fig": 5, "n_tab": 13, "L_tab": 2251, "n_element_tab": 249, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 949, "n_element_tab_1": 126, "formula_len_all": 790, "formula_len_all_1": 791, "len_all": 149031, "len_all_1": 66576, "len_abs": 1231, "len_title": 103, "len_sents": 36365, "len_sents_1": 27453, "n_sents": 286, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1236, "title": "Flat-LoRA: Low-Rank Adaption over a Flat Loss Landscape", "abs": "Fine-tuning large-scale pre-trained models is prohibitively expensive in terms of computational and memory costs. Low-Rank Adaptation (LoRA), a popular Parameter-Efficient Fine-Tuning (PEFT) method, provides an efficient way to fine-tune models by optimizing only a low-rank matrix. Despite recent progress made in improving LoRA's performance, the connection between the LoRA optimization space and the original full parameter space is often overlooked. A solution that appears flat in the LoRA space may exist sharp directions in the full parameter space, potentially harming generalization performance. In this paper, we propose Flat-LoRA, an efficient approach that seeks a low-rank adaptation located in a flat region of the full parameter space. Instead of relying on the well-established sharpness-aware minimization approach, which can incur significant computational and memory burdens, we utilize random weight perturbation with a Bayesian expectation loss objective to maintain training efficiency and design a refined perturbation generation strategy for improved performance. Experiments on natural language processing and image classification tasks with various architectures demonstrate the effectiveness of our approach.", "keywords": [ "low-rank adaption", "flat minima", "efficient training" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "vzrs42hgb0", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Mahdi Jalili", "gender": "Male", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "Full Professor" }, { "name": "Parham Moradi DW", "gender": "Male", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "Postdoc" }, { "name": "Saman Forouzandeh", "gender": "Male", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "Postdoctoral Research Fellow " } ], "n_formula": 20, "n_formula_1": 17, "n_ref_uni": 27, "n_ref": 47, "n_ref_all": 72, "n_fig": 5, "n_tab": 7, "L_tab": 3767, "n_element_tab": 375, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 634, "n_element_tab_1": 74, "formula_len_all": 927, "formula_len_all_1": 663, "len_all": 149452, "len_all_1": 64760, "len_abs": 1733, "len_title": 137, "len_sents": 57481, "len_sents_1": 29162, "n_sents": 386, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1075, "title": "DistillHGNN: A Knowledge Distillation Approach for High-Speed Hypergraph Neural Networks", "abs": "In this paper, we propose a novel framework to significantly enhance the inference speed and memory efficiency of Hypergraph Neural Networks (HGNNs) while preserving their high accuracy. Our approach utilizes an advanced teacher-student knowledge distillation strategy. The teacher model, consisting of an HGNN and a Multi-Layer Perceptron (MLP), not only produces soft labels but also transfers structural and high-order information to a lightweight Graph Convolutional Network (GCN) known as TinyGCN. This dual transfer mechanism enables the student model to effectively capture complex dependencies while benefiting from the faster inference and lower computational cost of the lightweight GCN. The student model is trained using both labeled data and soft labels provided by the teacher, with contrastive learning further ensuring that the student retains high-order relationships. This makes the proposed method efficient and suitable for real-time applications, achieving performance comparable to traditional HGNNs but with significantly reduced resource requirements.", "keywords": [ "Knowledge Distillation", "Hypergraph Neural Networks", "Contrastive Learning." ], "rating_list": [ 5, 6, 8 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "vzItLaEoDa", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Jiajian Li", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Qi Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Wenjun Zeng", "gender": "Male", "institution": "Eastern Institute for Advanced Study", "country": "CN", "position": "Full Professor" }, { "name": "Xiaokang Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University, China", "country": "CN", "position": "Professor" }, { "name": "Xin Jin", "gender": "Male", "institution": "Eastern Institute of Technology, Ningbo", "country": "CN", "position": "Assistant Professor" }, { "name": "Yang Li", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Yunbo Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 15, "n_ref_uni": 40, "n_ref": 69, "n_ref_all": 109, "n_fig": 17, "n_tab": 2, "L_tab": 1361, "n_element_tab": 280, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 922, "n_element_tab_1": 112, "formula_len_all": 2028, "formula_len_all_1": 1623, "len_all": 161636, "len_all_1": 62236, "len_abs": 1170, "len_title": 114, "len_sents": 49757, "len_sents_1": 28105, "n_sents": 376, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 66, "L_abs": 1186, "title": "Open-World Reinforcement Learning over Long Short-Term Imagination", "abs": "Training visual reinforcement learning agents in a high-dimensional open world presents significant challenges. While various model-based methods have improved sample efficiency by learning interactive world models, these agents tend to be \u201cshort-sighted\u201d, as they are typically trained on short snippets of imagined experiences. We argue that the primary challenge in open-world decision-making is improving the exploration efficiency across a vast state space, especially for tasks that demand consideration of long-horizon payoffs. In this paper, we present LS-Imagine, which extends the imagination horizon within a limited number of state transition steps, enabling the agent to explore behaviors that potentially lead to promising long-term feedback. The foundation of our approach is to build a $\\textit{long short-term world model}$. To achieve this, we simulate goal-conditioned jumpy state transitions and compute corresponding affordance maps by zooming in on specific areas within single images. This facilitates the integration of direct long-term values into behavior learning. Our method demonstrates significant improvements over state-of-the-art techniques in MineDojo.", "keywords": [ "World models", "reinforcement learning", "visual control" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "vyzPMQ5weJ", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 10, "author_info_list": [ { "name": "Brianna Barrow", "gender": "Female", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Christopher Jermaine", "gender": "Male", "institution": "Rice University", "country": "", "position": "Full Professor" }, { "name": "Daniel Bourgeois", "gender": "Male", "institution": "Rice University", "country": "", "position": "PhD student" }, { "name": "Jiawen Yao", "gender": "Female", "institution": ", Rice University", "country": "US", "position": "MS student" }, { "name": "Jiehui Li", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Researcher" }, { "name": "Sleem Mahmoud Abdelghafar", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Tania Lorido Botran", "gender": "Female", "institution": "Roblox", "country": "US", "position": "Researcher" }, { "name": "Xinyu Yao", "gender": "Female", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Zhimin Ding", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "yuxin tang", "gender": "unknown", "institution": "Bosch Center for Artificial Intelligence", "country": "US", "position": "Research Intern" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 19, "n_ref": 29, "n_ref_all": 58, "n_fig": 15, "n_tab": 3, "L_tab": 481, "n_element_tab": 28, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 203, "n_element_tab_1": 13, "formula_len_all": 584, "formula_len_all_1": 965, "len_all": 107275, "len_all_1": 67550, "len_abs": 1097, "len_title": 109, "len_sents": 39319, "len_sents_1": 31732, "n_sents": 285, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1169, "title": "TURNIP: A \u201cNondeterministic\u201d GPU Runtime with CPU RAM Offload", "abs": "An obvious way to alleviate memory difficulties in GPU-based AI computing is via CPU offload, where data are moved between GPU and CPU RAM, so inexpensive CPU RAM is used to increase the amount of storage available. While CPU offload is an obvious idea, it can greatly slow down a computation, due to the relatively slow transfer rate between CPU RAM and GPU RAM. Thus, any system for CPU offload needs to ensure that when such a transfer needs to happen, no computation is blocked waiting for the transfer to finish. One of the key challenges when using CPU offload is that memory transfers introduce nondeterminacy into the system: it is not possible to know before runtime when the transfers will finish, and hence what is the best order of operations to run to ensure there is no blocking. In this paper, we describe TURNIP, which is a system for running AI computations using CPU offload. The key innovation in TURNIP is the compilation of the AI computation into a dependency graph that gives the TURNIP runtime freedom to run operations such as GPU kernel calls in many different orders; at runtime, TURNIP chooses the best order in response to real-time events.", "keywords": [ "CPU Offload", "Memory Management", "Nondeterministic Execution", "Machine Learning System" ], "rating_list": [ 5, 5, 6, 5, 5 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 4, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "vyflgpwfJW", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Abhijeetsingh Meena", "gender": "Male", "institution": "Pandit Deendayal Petroleum University", "country": "IN", "position": "Undergrad student" }, { "name": "Aryan Prakhar", "gender": "Male", "institution": "Indian Institute of Technology (Banaras Hindu University) Varanasi", "country": "IN", "position": "Undergrad student" }, { "name": "Ashish Sabharwal", "gender": "Male", "institution": "Allen Institute for AI", "country": "US", "position": "Principal Researcher" }, { "name": "Bodhisattwa Prasad Majumder", "gender": "unknown", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Dhruv Agarwal", "gender": "Male", "institution": "University of Massachusetts Amherst", "country": "US", "position": "PhD student" }, { "name": "Harshit Surana", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Peter Clark", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Senior Research Manager" }, { "name": "Tirth Vora", "gender": "Male", "institution": "OpenLocus", "country": "IN", "position": "Intern" }, { "name": "Tushar Khot", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Lead Research Scientist" }, { "name": "Bhavana Dalvi", "gender": "Female", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Lead Reserarch Scientist" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 51, "n_ref": 75, "n_ref_all": 95, "n_fig": 5, "n_tab": 2, "L_tab": 10, "n_element_tab": 1, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 529, "n_element_tab_1": 93, "formula_len_all": 4833, "formula_len_all_1": 2849, "len_all": 181520, "len_all_1": 71899, "len_abs": 2576, "len_title": 121, "len_sents": 55928, "len_sents_1": 35720, "n_sents": 402, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 72, "L_abs": 1477, "title": "DiscoveryBench: Towards Data-Driven Discovery with Large Language Models", "abs": "Can the rapid advances in code generation, function calling, and data analysis using large language models (LLMs) help automate the search and verification of hypotheses purely from a set of provided datasets? To evaluate this question, we present DiscoveryBench, the first comprehensive benchmark that formalizes the multi-step process of data-driven discovery. The benchmark is designed to systematically assess current model capabilities in discovery tasks and provide a useful resource for improving them. Our benchmark contains 264 tasks collected across 6 diverse domains, such as sociology and engineering, by manually deriving discovery workflows from published papers to approximate the real-world challenges faced by researchers, where each task is defined by a dataset, its metadata, and a discovery goal in natural language. We additionally provide 903 synthetic tasks to conduct controlled evaluations on data-driven workflows that are not covered in the manually collected split. Furthermore, our structured formalism of data-driven discovery enables a facet-based evaluation that provides useful insights into different failure modes. We evaluate several popular LLM-based reasoning frameworks using both open and closed LLMs as baselines on DiscoveryBench and find that even the best system scores only 25%. Our benchmark, thus, illustrates the challenges in autonomous data-driven discovery and serves as a valuable resource for the community to make progress.", "keywords": [ "data-driven discovery", "data analysis", "large language models", "hypothesis generation", "hypothesis verification" ], "rating_list": [ 5, 8, 8 ], "soundness_list": [ 2, 4, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "vyHFTsOUWu", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Christopher D. Manning", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "", "position": "Full Professor" }, { "name": "John Hewitt", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" }, { "name": "Nelson F. Liu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Percy Liang", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 25, "n_ref": 62, "n_ref_all": 93, "n_fig": 6, "n_tab": 17, "L_tab": 26682, "n_element_tab": 418, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 3346, "n_element_tab_1": 93, "formula_len_all": 495, "formula_len_all_1": 286, "len_all": 265709, "len_all_1": 64088, "len_abs": 1275, "len_title": 96, "len_sents": 87614, "len_sents_1": 28609, "n_sents": 775, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1281, "title": "Instruction Following without Instruction Tuning", "abs": "Adaptations that seem deficient compared to instruction-tuning (finetuning on instruction-response pairs) can still implicitly yield instruction-following language models. We show that training solely on responses, without any corresponding\ninstructions, yields instruction following. This suggests that instruction-tuning just needs to teach the desired distribution of responses. However, we then show that one can finetune on instruction-response pairs from a distribution unlike the test distribution of instructions, like just poetry, or just math, and still yield a model that broadly follows instructions. Instead of acting, e.g., just as math models, these single-task models sometimes behave more as general-purpose chatbots for, e.g., non-math instructions. To begin to explain this implicit instruction tuning, we hypothesize that simple changes to a language model\u2019s distribution yield instruction following. We support this by hand-writing a rule-based adapter that yields instruction-following behavior in language models. The rules are to slowly increase the probability of ending the sequence, penalize repetition, and uniformly change 15 words\u2019 probabilities. In summary, adaptations made without being designed to yield instruction following can do so implicitly.", "keywords": [ "instruction tuning", "instruction following", "ablation", "rule-based" ], "rating_list": [ 6, 10, 5, 3 ], "soundness_list": [ 4, 4, 3, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "vyFSyfiOIu", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Cheng Ouyang", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Lecturer" }, { "name": "Haozhe Wang", "gender": "Male", "institution": "INF", "country": "CN", "position": "Researcher" }, { "name": "Rossella Arcucci", "gender": "Female", "institution": "Imperial College London ", "country": "", "position": "Senior Lecturer" }, { "name": "Wenjia Bai", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Associate Professor" }, { "name": "Zhongwei Wan", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "che liu", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 76, "n_ref_all": 90, "n_fig": 8, "n_tab": 4, "L_tab": 3097, "n_element_tab": 220, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 1689, "n_element_tab_1": 99, "formula_len_all": 615, "formula_len_all_1": 601, "len_all": 148893, "len_all_1": 68786, "len_abs": 2333, "len_title": 134, "len_sents": 35260, "len_sents_1": 31872, "n_sents": 276, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1851, "title": "Knowledge-enhanced Multimodal ECG Representation Learning with Arbitrary-Lead Inputs", "abs": "Recent advancements in multimodal representation learning for electrocardiogram (ECG) have moved onto learning representations by aligning ECG signals with their paired free-text reports. \nHowever, current methods often result in suboptimal alignment of ECG signals with their corresponding text reports, thereby limiting diagnostic accuracy. This is primarily due to the complexity and unstructured nature of medical language, which makes it challenging to effectively align ECG signals with the corresponding text reports. \nAdditionally, these methods are unable to handle arbitrary combinations of ECG leads as inputs, which poses a challenge since 12-lead ECGs may not always be available in under-resourced clinical environments.\n\nIn this work, we propose the **K**nowledge-enhanced **M**ultimodal **E**CG **R**epresentation **L**earning (**K-MERL**) framework to address these challenges. \nK-MERL leverages large language models (LLMs) to extract structured knowledge from free-text reports, enhancing the effectiveness of ECG multimodal learning. \nFurthermore, we design a lead-aware ECG encoder to capture lead-specific spatial-temporal characteristics of 12-lead ECGs, with dynamic lead masking. This novel encoder allows our framework to handle arbitrary lead inputs, rather than being limited to a fixed set of full 12 leads, which existing methods necessitate.\n\nWe evaluate K-MERL on six external ECG datasets and demonstrate its superior capability. \nK-MERL not only outperforms all existing methods in zero-shot classification and linear probing tasks using 12 leads, but also achieves state-of-the-art (SOTA) results in partial-lead settings, with an average improvement of **16%** in AUC score on zero-shot classification compared to previous SOTA multimodal methods[^1].\n\n[^1]: All data and code will be released upon acceptance.", "keywords": [ "Electrocardiogram", "healthcare", "physiological signals" ], "rating_list": [ 6, 3, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 1, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 3, 1 ], "confidence_list": [ 4, 4, 5, 4, 5 ] }, { "paper_id": "vyF5aim4US", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Dongmei Jiang", "gender": "Female", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Guiping Cao", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jianguo Zhang", "gender": "Male", "institution": "Southern University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Wenjian Huang", "gender": "Male", "institution": "Southern University Of Science and Technology", "country": "CN", "position": "Researcher" }, { "name": "Yaowei Wang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "xiangyuan lan", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 138, "n_ref_all": 224, "n_fig": 13, "n_tab": 22, "L_tab": 9752, "n_element_tab": 753, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 764, "n_element_tab_1": 83, "formula_len_all": 643, "formula_len_all_1": 420, "len_all": 257196, "len_all_1": 79372, "len_abs": 2248, "len_title": 120, "len_sents": 85870, "len_sents_1": 32833, "n_sents": 621, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 2240, "title": "QFree-Det: Query-Free Detector with Transformer and Sequential Matching", "abs": "Transformer-based detectors, such as DETR and DINO, often struggle with a specific limitation: they can detect only a fixed number of objects based on the predefined number of queries set. This limitation leads to missed detections when the scene exceeds the model\u2019s capacity and increases false positives when the scene contains fewer objects. In addition, existing approaches often combine one-to-one and one-to-many matching label assignment methods in the decoder for accelerating the model training and convergence. However, this operation can introduce a new detecting ambiguity issue, which is often overlooked by those methods. To address these challenges, we propose QFree-Det, a novel query-free detector capable of dynamically detecting a variable number of objects across different input images. In particular, we present an Adaptive Free Query Selection (AFQS) algorithm to dynamically select queries from the encoder tokens, which resolves the issue of fixed capacity. Then, we propose a sequential matching method that decouples the one-to-one and one-to-many processes into separating sequential steps, effectively addressing the issue of detecting ambiguity. To achieve the sequential matching, we design a new Location-Deduplication Decoder (LDD) by rethinking the role of cross-attention (CA) and self-attention (SA) within the decoder. LDD first regresses the location of multiple boxes with CA in a one-to-many manner and then performs object classification to recognize and eliminate duplicate boxes with SA in a one-to-one manner. Finally, to improve the detection ability on small objects, we design a unified PoCoo loss that leverages prior knowledge of box size to encourage the model to pay more attention to small objects. Extensive experiments on COCO2017 and WiderPerson datasets demonstrate the effectiveness of our QFreeDet. For instance, QFree-Det achieves consistent and remarkable improvements over DINO across five different backbones. Notably, QFree-Det obtains a new state-of-the-art of 54.4% AP and 38.8% APs on val2017 of COCO with the backbone of VMamba-T under 1\u00d7 training schedule (12 epochs), higher than DINO-VMamba-T by +0.9% AP and +2.2% APs. The source codes will be released upon acceptance.", "keywords": [ "free-object prediction", "query-free", "detecting ambiguity", "location-deduplication decoder" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "vxvgZ0kTFv", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Blake Woodworth", "gender": "Male", "institution": "George Washington University", "country": "US", "position": "Assistant Professor" }, { "name": "Pierfrancesco Beneventano", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" } ], "n_formula": 117, "n_formula_1": 25, "n_ref_uni": 24, "n_ref": 35, "n_ref_all": 51, "n_fig": 3, "n_tab": 1, "L_tab": 97, "n_element_tab": 19, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9854, "formula_len_all_1": 1700, "len_all": 151140, "len_all_1": 66586, "len_abs": 755, "len_title": 147, "len_sents": 49018, "len_sents_1": 29901, "n_sents": 462, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 772, "title": "Gradient Descent Converges Linearly to Flatter Minima than Gradient Flow in Shallow Linear Networks", "abs": "We study the gradient descent (GD) dynamics of a depth-2 linear neural network with a single input and output. We show that GD converges at an explicit linear rate to a global minimum of the training loss, even with a large stepsize--about $2/\\textrm{sharpness}$. It still converges for even larger stepsizes, but may do so very slowly. We also characterize the solution to which GD converges, which has lower norm and sharpness than the gradient flow solution.\nOur analysis reveals a trade off between the speed of convergence and the magnitude of implicit regularization.\nThis sheds light on the benefits of training at the ``Edge of Stability'', which induces additional regularization by delaying convergence and may have implications for training more complex models.", "keywords": [ "Gradient Descent", "Implicit Regularization", "Shallow Networks", "Linear Networks" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 1, 4, 3, 2 ], "presentation_list": [ 1, 3, 1, 1 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "vxutwN3xQN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 19, "author_info_list": [ { "name": "Canyu Chen", "gender": "unknown", "institution": "Illinois Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Chaoqi Wang", "gender": "Male", "institution": "xAI", "country": "US", "position": "Member of Technical Staff" }, { "name": "Chelsea Finn", "gender": "Female", "institution": "Physical Intelligence", "country": "US", "position": "Researcher" }, { "name": "Chenhang Cui", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Summer Camp Student" }, { "name": "Haoqin Tu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Huaxiu Yao", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Jiawei Zhou", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Assistant Professor" }, { "name": "Leria HUANG", "gender": "Female", "institution": "Ecole Sup\u00e9rieure des Sciences Economiques et Commerciales", "country": "FR", "position": "MS student" }, { "name": "Qinghao Ye", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Rafael Rafailov", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Yiyang Zhou", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Yuqing Zhang", "gender": "Female", "institution": "University of Queensland", "country": "", "position": "Researcher" }, { "name": "Zhaorun Chen", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Zhengwei Tong", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Zhenzhen Weng", "gender": "Female", "institution": "Waymo", "country": "US", "position": "Researcher" }, { "name": "Zhihong Zhu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Zhuokai Zhao", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Zichen Wen", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Du Yichao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 3, "n_ref_uni": 71, "n_ref": 228, "n_ref_all": 326, "n_fig": 19, "n_tab": 30, "L_tab": 15432, "n_element_tab": 2445, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 3028, "n_element_tab_1": 441, "formula_len_all": 1055, "formula_len_all_1": 143, "len_all": 394667, "len_all_1": 68998, "len_abs": 1885, "len_title": 140, "len_sents": 103883, "len_sents_1": 27899, "n_sents": 747, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1822, "title": "MJ-Bench: Is Your Multimodal Reward Model Really a Good Judge for Text-to-Image Generation?", "abs": "While text-to-image models like DALLE-3 and Stable Diffusion are rapidly proliferating, they often encounter challenges such as hallucination, bias, and the production of unsafe, low-quality output. To effectively address these issues, it is crucial to align these models with desired behaviors based on feedback from a multimodal judge. Despite their significance, current multimodal judges frequently undergo inadequate evaluation of their capabilities and limitations, potentially leading to misalignment and unsafe fine-tuning outcomes. To address this issue, we introduce MJ-Bench, a novel benchmark which incorporates a comprehensive preference dataset to evaluate multimodal judges in providing feedback for image generation models across four key perspectives: alignment, safety, image quality, and bias. Specifically, we evaluate a large variety of multimodal judges including smaller-sized CLIP-based scoring models, open-source VLMs (e.g. LLaVA family), and close-source VLMs (e.g. GPT-4o, Claude 3) on each decomposed subcategory of our preference dataset. Experiments reveal that close-source VLMs generally provide better feedback, with GPT-4o outperforming other judges in average. Compared with open-source VLMs, smaller-sized scoring models can provide better feedback regarding text-image alignment and image quality, while VLMs provide more accurate feedback regarding safety and generation bias due to their stronger reasoning capabilities. Further studies in feedback scale reveal that VLM judges can generally provide more accurate and stable feedback in natural language (Likert-scale) than numerical scales. Notably, human evaluations on end-to-end fine-tuned models using separate feedback from these multimodal judges provide similar conclusions, further confirming the effectiveness of MJ-Bench.", "keywords": [ "Multimodal Reward Models", "Foundation Models Alignment", "Reinforcement Learning from Human Feedback" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "vxrtEHc97c", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Huanying Gu", "gender": "Female", "institution": "New York Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Jerry Cheng", "gender": "Male", "institution": "New York Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Zheng Li", "gender": "Male", "institution": "New York Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 18, "n_ref_uni": 27, "n_ref": 31, "n_ref_all": 71, "n_fig": 10, "n_tab": 10, "L_tab": 4100, "n_element_tab": 301, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 726, "n_element_tab_1": 46, "formula_len_all": 1482, "formula_len_all_1": 1349, "len_all": 131066, "len_all_1": 61236, "len_abs": 1415, "len_title": 111, "len_sents": 39438, "len_sents_1": 25528, "n_sents": 320, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1429, "title": "LagEncoder: A Non-Parametric Method for Representation Learning", "abs": "Non-parametric encoders offer advantages in interpretability and generalizability. However, they often perform significantly worse than deep neural networks on many challenging recognition tasks, and it remains unclear how to effectively apply these techniques to such tasks. In this work, we view all AI recognition tasks as function approximation problems and introduce LagEncoder, a non-parametric, training-free feature extraction method based on finite element basis functions. Our encoder features a universal architecture that can be applied to various types of raw data and recognition tasks. We found that LagEncoder effectively overcomes the limitations of neural networks in regression problems, particularly when fitting multi-frequency functions. The LagEncoder-based model converges quickly and requires low training costs, as only the head is trained. Additionally, LagEncoder provides a parameter-efficient fine-tuning approach. Our experiments on the ImageNet-1K and WikiText dataset demonstrate that pre-trained models using LagEncoder achieve performance improvements within just one training epoch. Furthermore, it does not require adjustments to the original training recipe, extra training data, and the model's total parameters remain nearly unchanged. Our evaluation of the scaling law for model performance indicates that using the LagEncoder is more cost-effective than merely increasing the model size.", "keywords": [ "Non-parametric encoder", "Finite element method", "Interpretable model", "Universal architecture", "Scaling law", "ImageNet", "ResNet", "ViT" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 1, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 4, 4 ] }, { "paper_id": "vxhzSm1D3J", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Chaorui Zhang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Jianfeng Hou", "gender": "Male", "institution": "Fuzhou University", "country": "CN", "position": "Full Professor" }, { "name": "Meng QIN", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Wei Li", "gender": "unknown", "institution": "Fujian Agriculture and Forestry university", "country": "CN", "position": "Associate Professor" }, { "name": "Weixi Zhang", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Principal Researcher" }, { "name": "Zhong yiwen", "gender": "Male", "institution": "Fujian Agriculture and Forestry University", "country": "CN", "position": "Full Professor" } ], "n_formula": 74, "n_formula_1": 32, "n_ref_uni": 40, "n_ref": 100, "n_ref_all": 143, "n_fig": 4, "n_tab": 27, "L_tab": 15536, "n_element_tab": 1560, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3929, "n_element_tab_1": 379, "formula_len_all": 11124, "formula_len_all_1": 3070, "len_all": 210068, "len_all_1": 69982, "len_abs": 3627, "len_title": 134, "len_sents": 47615, "len_sents_1": 24693, "n_sents": 408, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1333, "title": "Rethinking Degree-Corrected Spectral Clustering: a Pure Spectral Analysis & Extension", "abs": "Spectral clustering is a representative graph clustering technique with strong interpretability and theoretical guarantees. Recently, degree-corrected spectral clustering (DCSC) has emerged as the state-of-the-art for this technique. While prior studies have provided several theoretical results for DCSC, their analysis relies on some random graph models (e.g., stochastic block models). In this study, we explore an alternative analysis of DCSC from a pure spectral view. It gives rigorous bounds for the mis-clustered volume and conductance w.r.t. the optimal solution while involving quantities that indicate impacts of (i) high degree heterogeneity and (ii) weak clustering structures to DCSC. Inspired by recent advances in graph neural networks (GNNs) and the associated over-smoothing issue, we propose ASCENT (Adaptive Spectral ClustEring with Node-wise correcTion), a simple yet effective extension of DCSC. Different from most DCSC methods with a constant degree correction for all nodes, ASCENT follows a node-wise correction scheme. It can assign different corrections for nodes via the mean aggregation of GNNs. We further demonstrate that (i) ASCENT reduces to conventional DCSC methods when encountering over-smoothing and (ii) some early stages before over-smoothing can potentially obtain better clustering quality.", "keywords": [ "Degree-corrected Spectral Clustering", "Regularized Spectral Clustering", "Graph Clustering", "Spectral Graph Theory" ], "rating_list": [ 5, 8, 3, 5, 3 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 1, 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 2, 3 ] }, { "paper_id": "vxWDoD8oz7", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Max van Spengler", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Pascal Mettes", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Assistant Professor" } ], "n_formula": 66, "n_formula_1": 27, "n_ref_uni": 55, "n_ref": 115, "n_ref_all": 136, "n_fig": 3, "n_tab": 6, "L_tab": 1776, "n_element_tab": 253, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 193, "n_element_tab_1": 16, "formula_len_all": 4060, "formula_len_all_1": 1373, "len_all": 185358, "len_all_1": 73375, "len_abs": 1531, "len_title": 117, "len_sents": 52842, "len_sents_1": 33739, "n_sents": 462, "n_sents_1": 273, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1539, "title": "Distortion-free and GPU-compatible Tree Embeddings in Hyperbolic Space", "abs": "Embedding tree-like data, from hierarchies to ontologies and taxonomies, forms a well-studied problem for representing knowledge across many domains. Hyperbolic geometry provides a natural solution for embedding trees, with vastly superior performance over Euclidean embeddings. Recent literature has shown that hyperbolic tree embeddings can even be placed on top of neural networks for hierarchical knowledge integration in deep learning settings. For all applications, a faithful embedding of trees is needed, with combinatorial constructions emerging as the most effective direction. This paper identifies and solves two key limitations of existing works. First, the combinatorial construction hinges on finding maximally separated points on a hypersphere, a notoriously difficult problem. Current approaches lead to poor separation, which degrades the quality of the corresponding hyperbolic embedding. As a solution, we propose maximally separated Delaunay tree embeddings (MS-DTE), where during placement, the children of a node are maximally separated through optimization, which directly leads to lower embedding distortion. Second, low distortion requires additional precision. The current approach for increasing precision is to use multiple precision arithmetic, which renders the embeddings useless on GPUs in deep learning settings. We reformulate the combinatorial construction using floating point expansion arithmetic, leading to superior embedding quality while simultaneously retaining their use on accelerated hardware.", "keywords": [ "Hyperbolic Geometry", "Hyperbolic Tree Embeddings", "Representation Learning", "Hierarchical Learning" ], "rating_list": [ 6, 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 2, 4 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 4, 3, 3, 2, 3 ], "confidence_list": [ 3, 1, 5, 5, 2 ] }, { "paper_id": "vxBvr5ZpIu", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Cheng Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Longlin Yu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Tianyu Xie", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhekun Shi", "gender": "unknown", "institution": "School of mathematical Science, Peking University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 151, "n_formula_1": 43, "n_ref_uni": 36, "n_ref": 94, "n_ref_all": 157, "n_fig": 7, "n_tab": 12, "L_tab": 2503, "n_element_tab": 187, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1164, "n_element_tab_1": 67, "formula_len_all": 20413, "formula_len_all_1": 4685, "len_all": 200305, "len_all_1": 71410, "len_abs": 2386, "len_title": 70, "len_sents": 49353, "len_sents_1": 26774, "n_sents": 480, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 22, "L_abs": 1016, "title": "Diffusion-PINN Sampler", "abs": "Recent success of diffusion models has inspired a surge of interest in developing sampling techniques using reverse diffusion processes. However, accurately estimating the drift term in the reverse stochastic differential equation (SDE) solely from the unnormalized target density poses significant challenges, hindering existing methods from achieving state-of-the-art performance. In this paper, we introduce the Diffusion-PINN Sampler (DPS), a novel diffusion-based sampling algorithm that estimates the drift term by solving the governing partial differential equation of the log-density of the underlying SDE marginals via physics-informed neural networks (PINN). We prove that the error of log-density approximation can be controlled by the PINN residual loss, enabling us to establish convergence guarantees of DPS. Experiments on a variety of sampling tasks demonstrate the effectiveness of our approach, particularly in accurately identifying mixing proportions when the target contains isolated components.", "keywords": [ "posterior sampling", "multi-modal sampling", "mixing proportion identification", "diffusion model", "physics-informed neural network" ], "rating_list": [ 6, 3, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "vx1vJIFvd5", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Cai Yuchen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Ding Cao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Guiquan Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Xinyuan Liu", "gender": "Not Specified", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 45, "n_formula_1": 14, "n_ref_uni": 32, "n_ref": 141, "n_ref_all": 176, "n_fig": 12, "n_tab": 14, "L_tab": 7258, "n_element_tab": 375, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 3386, "n_element_tab_1": 97, "formula_len_all": 3326, "formula_len_all_1": 848, "len_all": 217227, "len_all_1": 70701, "len_abs": 1095, "len_title": 122, "len_sents": 68593, "len_sents_1": 30127, "n_sents": 613, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1099, "title": "O-Edit: Orthogonal Subspace Editing for Language Model Sequential Editing", "abs": "Large language models (LLMs) acquire knowledge during pre-training, but over time, this knowledge may become incorrect or outdated, necessitating updates after training. Knowledge editing techniques address this issue without the need for costly re-training. However, most existing methods are designed for single edits, and as the number of edits increases, they often cause a decline in the model's overall performance, posing significant challenges for sequential editing. To overcome this, we propose Orthogonal Subspace Editing, O-Edit. This algorithm orthogonalizes the direction of each knowledge update, minimizing interference between successive updates and reducing the impact of new updates on unrelated knowledge. Our approach does not require replaying previously edited data and processes each edit knowledge on time. It can perform thousands of edits on mainstream LLMs, achieving an average performance improvement that is 4.2 times better than existing methods while effectively preserving the model's performance on downstream tasks, all with minimal additional parameter overhead.", "keywords": [ "large language model", "model editing", "sequential editing" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "vwOq7twk7L", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Bo Peng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Haokun Lin", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jing Dong", "gender": "Female", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xingzheng", "gender": "Male", "institution": "Tencent zhuque lab", "country": "CN", "position": "Researcher" }, { "name": "Yang Bai", "gender": "Female", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Yue Jiang", "gender": "Female", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yueming Lyu", "gender": "Female", "institution": "nanjing university", "country": "CN", "position": "Assistant Professor" }, { "name": "yang yong", "gender": "Male", "institution": "tencent security ", "country": "CN", "position": "Lecturer" }, { "name": "Zhili LIU", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 15, "n_ref_uni": 23, "n_ref": 82, "n_ref_all": 101, "n_fig": 8, "n_tab": 14, "L_tab": 5896, "n_element_tab": 788, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1025, "n_element_tab_1": 65, "formula_len_all": 836, "formula_len_all_1": 837, "len_all": 140383, "len_all_1": 55258, "len_abs": 1635, "len_title": 125, "len_sents": 41403, "len_sents_1": 25528, "n_sents": 296, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1642, "title": "Image-level memorization detection via inversion-based inference perturbation", "abs": "Recent studies have discovered that widely used text-to-image diffusion models can replicate training samples during image generation, a phenomenon known as memorization. Existing detection methods primarily focus on identifying memorized prompts. However, in real-world scenarios, image owners may need to verify whether their proprietary or personal images have been memorized by the model, even in the absence of paired prompts or related metadata. We refer to this challenge as image-level memorization detection, where current methods relying on original prompts fall short. In this work, we uncover two characteristics of memorized images after perturbing the inference procedure: lower similarity of the original images and larger magnitudes of TCNP.\nBuilding on these insights, we propose Inversion-based Inference Perturbation (IIP), a new framework for image-level memorization detection. Our approach uses unconditional DDIM inversion to derive latent codes that contain core semantic information of original images and optimizes random prompt embeddings to introduce effective perturbation. Memorized images exhibit distinct characteristics within the proposed pipeline, providing a robust basis for detection. To support this task, we construct a comprehensive setup for the image-level memorization detection, carefully curating datasets to simulate realistic memorization scenarios. Using this setup, we evaluate our IIP framework across three different memorization settings, demonstrating its state-of-the-art performance in identifying memorized images in various settings, even in the presence of data augmentation attacks.", "keywords": [ "Text-to-image diffusion model", "data memorization detection", "DDIM Inversion" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "vwENIgfZdQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 16, "author_info_list": [ { "name": "Chen-Wei Xie", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Fan Cheng", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Han Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Hongyang Zhang", "gender": "Male", "institution": "School of Computer Science, University of Waterloo", "country": "CA", "position": "Assistant Professor" }, { "name": "Huangji Wang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Jie Xiao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "JixuanChen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "engineer" }, { "name": "Kai Zhu", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "Postdoc" }, { "name": "Keyu Yan", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Pingyu Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Ruili Feng", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Shangwen Zhu", "gender": "unknown", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Yu Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yue Yang", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Zhantao Yang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" }, { "name": "Zhicai Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 45, "n_ref": 96, "n_ref_all": 129, "n_fig": 19, "n_tab": 7, "L_tab": 2272, "n_element_tab": 264, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 696, "n_element_tab_1": 118, "formula_len_all": 229, "formula_len_all_1": 229, "len_all": 211516, "len_all_1": 62966, "len_abs": 2953, "len_title": 125, "len_sents": 72145, "len_sents_1": 29741, "n_sents": 562, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 2116, "title": "Asking Specifically Instead of Ambiguously to Your GPT Improves Image Caption", "abs": "The advances in large vision-language models (VLMs) have sparked a growing interest in generating accurate, complete, and user-friendly image captions to enhance downstream multi-modality tasks such as text-to-image generation, text-driven object detection, and grounding. However, current VLM-based image captioning methods often miss important details, recognize incorrect objects or relationships, and deliver suboptimal captions for downstream applications. One primary reason for this issue is the ambiguous prompts typically used, such as \"describe this image in detail,\" which fail to guide the VLM's focus on specific elements within the image. To address this, we extensively explore the difference between using ambiguous prompts and decomposing them into a series of specific questions. We find that asking a series of targeted element-specific questions significantly enhances the attention of VLMs to important objects, the consistency of the answers under repeated questions, and the alignment with their training data distribution. Building on this insight, we introduce ASSIST, a method that systematically decomposes image caption prompts into a sequence of focused questions corresponding to distinct image elements.We annotated 100k images using GPT-4V with this approach and fine-tuned a LLAVA model, resulting in a captioner that greatly improves caption accuracy and quality. Our fine-tuned model recognizes $\\times 1.5$ more correct objects and achieves $\\times1.5$ higher precision in describing them on the COCO benchmark compared to vague prompting methods. Additionally, our method produces element-specific answers that can be efficiently organized into graph structures, benefiting tasks like open-vocabulary object detection and image generation. This leads to significant improvements in the accuracy, precision, and mIoU of state-of-the-art detection models, with precision scores increasing by $\\times 1.7$ over previous methods. Experiments across diverse scenarios and benchmarks validate the effectiveness of ASSIST. All code, datasets, and models will be made publicly available.", "keywords": [ "vision-language models", "image captioning" ], "rating_list": [ 6, 5, 6, 5, 5 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 2 ], "contribution_list": [ 1, 2, 3, 2, 2 ], "confidence_list": [ 2, 2, 4, 4, 4 ] }, { "paper_id": "vw0NurJ7UX", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Jiahao Wang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Mengzhao Chen", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Wenqi Shao", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yi Bin", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Yi Liu", "gender": "Male", "institution": "Intel", "country": "CN", "position": "Researcher" }, { "name": "Ping Luo", "gender": "unknown", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 135, "n_ref_all": 199, "n_fig": 23, "n_tab": 25, "L_tab": 11538, "n_element_tab": 977, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 1902, "n_element_tab_1": 165, "formula_len_all": 515, "formula_len_all_1": 308, "len_all": 184533, "len_all_1": 58446, "len_abs": 1730, "len_title": 129, "len_sents": 48322, "len_sents_1": 24582, "n_sents": 415, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1300, "title": "PrefixQuant: Static Quantization Beats Dynamic through Prefixed Outliers in LLMs", "abs": "Quantization is essential for deploying Large Language Models (LLMs) by enhancing memory efficiency and inference speed. Existing methods for activation quantization mainly address channel-wise outliers, often neglecting token-wise outliers, leading to reliance on costly per-token dynamic quantization. To address this, we introduce PrefixQuant, a novel technique that isolates outlier tokens offline without re-training. Specifically, PrefixQuant identifies high-frequency outlier tokens and prefixes them in the KV cache, preventing the generation of outlier tokens during inference and simplifying quantization. To our knowledge, PrefixQuant is the first to enable efficient per-tensor static quantization to outperform expensive per-token dynamic quantization. For instance, in W4A4KV4 (4- bit weight, 4-bit activation, and 4-bit KV cache) Llama-3-8B, PrefixQuant with per-tensor static quantization achieves a 7.43 WikiText2 perplexity and 71.08% average accuracy on 5 common-sense reasoning tasks, outperforming previous per-token dynamic quantization methods like QuaRot with 0.98 perplexity improvement and +5.98 points accuracy. Additionally, the inference speed of W4A4 quantized models using PrefixQuant is 1.60\u00d7 to 2.81\u00d7 faster than FP16 models and exceeds QuaRot models by 1.2\u00d7 to 1.3\u00d7.", "keywords": [ "Large language model; Token-wise outliers; Static quantization;" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 3, 1 ], "presentation_list": [ 1, 3, 2 ], "contribution_list": [ 3, 1, 1 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "vvi5OjPhbu", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Guangwei Xu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Hai-Tao Zheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "LinHai", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Ruijie Guo", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Director" }, { "name": "Xuan Wu", "gender": "Female", "institution": "Southeast University", "country": "", "position": "MS student" }, { "name": "Yuan Miao", "gender": "unknown", "institution": "nanjing university", "country": "", "position": "MS student" }, { "name": "Zixuan Xiong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "wenkai Zhang", "gender": "Male", "institution": "nanjing university", "country": "", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 36, "n_ref": 59, "n_ref_all": 76, "n_fig": 3, "n_tab": 7, "L_tab": 5107, "n_element_tab": 379, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 699, "n_element_tab_1": 104, "formula_len_all": 973, "formula_len_all_1": 748, "len_all": 188377, "len_all_1": 63430, "len_abs": 2019, "len_title": 133, "len_sents": 51479, "len_sents_1": 28325, "n_sents": 414, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1260, "title": "Youku Dense Caption: A Large-scale Chinese Video Dense Caption Dataset and Benchmarks", "abs": "With the explosive growth of video content, video captions have emerged as a crucial tool for video comprehension, significantly enhancing the ability to understand and retrieve information from videos. However, most publicly available dense video captioning datasets are in English, resulting in a scarcity of large-scale and high-quality Chinese dense video captioning datasets. To address this gap within the Chinese community and to promote the advancement of Chinese multi-modal models, we develop the first, large-scale, and high-quality Chinese dense video captioning dataset, named Youku Dense Caption. This dataset is sourced from Youku, a prominent Chinese video-sharing website. Youku Dense Caption includes 31,466 complete short videos annotated by 311,921 Chinese captions. To the best of our knowledge, it is currently the largest publicly available dataset for fine-grained Chinese video descriptions. Additionally, we establish several benchmarks for Chinese video-language tasks based on the Youku Dense Caption, including retrieval, grounding, and generation tasks. Extensive experiments and evaluations are conducted on existing state-of-the-art multi-modal models, demonstrating the dataset's utility and the potential for further research.", "keywords": [ "Chinese Video Datasets", "Retrieval", "Grounding", "Generation" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "vvD0VFw0LG", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Changhao Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Gongfan Fang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Haoling Li", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Huiqiong Wang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Lecturer" }, { "name": "Jie Song", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Mengqi Xue", "gender": "Female", "institution": "Hangzhou City University", "country": "CN", "position": "Researcher" }, { "name": "Mingli Song", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Sheng Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Zunlei Feng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 51, "n_ref": 223, "n_ref_all": 250, "n_fig": 2, "n_tab": 13, "L_tab": 27137, "n_element_tab": 4418, "n_fig_1": 2, "n_tab_1": 9, "L_tab_1": 8417, "n_element_tab_1": 1026, "formula_len_all": 61, "formula_len_all_1": 0, "len_all": 251055, "len_all_1": 81000, "len_abs": 1080, "len_title": 109, "len_sents": 47079, "len_sents_1": 29371, "n_sents": 344, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1081, "title": "PruningBench: A Comprehensive Benchmark of Structural Pruning", "abs": "Structural pruning has emerged as a promising approach for producing more efficient models. Nevertheless, the community suffers from a lack of standardized benchmarks and metrics, leaving the progress in this area not fully comprehended. To fill this gap, we present the first comprehensive benchmark, termed PruningBench, for structural pruning. PruningBench showcases the following three characteristics: 1) PruningBench employs a unified and consistent framework for evaluating the effectiveness of diverse structural pruning techniques; 2) PruningBench systematically evaluates 16 existing pruning methods, encompassing a wide array of models (e.g., CNNs and ViTs) and tasks (e.g., classification and detection); 3) PruningBench provides easily implementable interfaces to facilitate the implementation of future pruning methods, and enables the subsequent researchers to incorporate their work into our leaderboards. We will provide an online pruning platform for customizing pruning tasks and reproducing all results in this paper. Codes will also be made publicly available.", "keywords": [ "network compression", "structural pruning", "benchmark" ], "rating_list": [ 8, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "vuvG5rNBra", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Chenxiang Zhang", "gender": "Male", "institution": "University of Luxemburg", "country": "LU", "position": "PhD student" }, { "name": "Jun Pang", "gender": "Male", "institution": "University of Luxembourg", "country": "LU", "position": "Principal Researcher" }, { "name": "Sjouke Mauw", "gender": "Male", "institution": "University of Luxemburg", "country": "LU", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 52, "n_ref": 149, "n_ref_all": 188, "n_fig": 9, "n_tab": 5, "L_tab": 3366, "n_element_tab": 521, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 754, "n_element_tab_1": 54, "formula_len_all": 363, "formula_len_all_1": 327, "len_all": 174582, "len_all_1": 69827, "len_abs": 919, "len_title": 91, "len_sents": 51723, "len_sents_1": 34158, "n_sents": 381, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 940, "title": "Spurious Privacy Leakage in Neural Networks", "abs": "Neural networks are vulnerable to privacy attacks aimed at stealing sensitive data. When trained on real-world datasets, these models can also inherit latent biases, which may further increase privacy risks. In this work, we investigate the impact of spurious correlation bias on privacy vulnerability, identifying several key challenges. We introduce _spurious privacy leakage_, a phenomenon where spurious groups can be up to 100 times more vulnerable to privacy attacks than non-spurious groups, and demonstrate how this leakage is connected to task complexity. Furthermore, while robust training methods can mitigate the performance disparity across groups, they fail to reduce privacy vulnerability, and even differential privacy is ineffective in protecting the most vulnerable spurious group in practice. Finally, we compare model architectures in terms of both performance and privacy, revisiting prior research with novel insights.", "keywords": [ "spurious correlation", "membership inference", "privacy", "robustness", "safety" ], "rating_list": [ 3, 5, 8, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "vuuYbA1vB2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Guoqing Liu", "gender": "Male", "institution": "Microsoft Research ", "country": "CN", "position": "Researcher" }, { "name": "Ning Yang", "gender": "Female", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Yongcheng Zeng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zhiyu Zhao", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "MS student" } ], "n_formula": 24, "n_formula_1": 17, "n_ref_uni": 0, "n_ref": 1, "n_ref_all": 12, "n_fig": 3, "n_tab": 1, "L_tab": 423, "n_element_tab": 41, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2214, "formula_len_all_1": 1528, "len_all": 73348, "len_all_1": 58867, "len_abs": 1499, "len_title": 142, "len_sents": 31530, "len_sents_1": 26872, "n_sents": 244, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1504, "title": "Enhancing Mathematical Reasoning in Language Models Through Focused Differentiation Training", "abs": "Enhancing the mathematical capabilities of large language models (LLMs) is crucial for applications requiring precise and rigorous mathematical reasoning. Current models, even when trained with methods like Direct Preference Optimization (DPO), often struggle to effectively differentiate between correct and erroneous mathematical responses, especially when errors occur in multi-step solutions. Traditional approaches focusing on token or logit-level analysis fail to capture the nuanced semantic differences in mathematical reasoning. To address this challenge, we propose leveraging the rich semantic information embedded in the hidden state space of LLMs. Our novel approach, Focused Differentiation Training (FDT), fine-tunes the model by emphasizing the differences between the hidden states of correct and incorrect responses, rather than their common features. Unlike other methods that detect errors at the token or logits level and often rely on human input or more powerful models, our approach enhances mathematical reasoning capabilities using only the model's inherent abilities. This methodology promotes a more accurate alignment with mathematical correctness, thereby improving the model's ability to evaluate and generate precise mathematical responses. Experimental results demonstrate that our algorithm substantially outperforms traditional alignment methods in mathematical tasks, offering a robust solution for enhancing the mathematical reasoning capabilities of language models.", "keywords": [ "large language model", "alignment" ], "rating_list": [ 8, 3, 5, 3 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "vunPXOFmoi", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Fei Huang", "gender": "Male", "institution": "Alibaba Group US", "country": "US", "position": "Senior Research Director" }, { "name": "Huajun Chen", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Ningyu Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Pengjun Xie", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Shuofei Qiao", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Xiaobin Wang", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Intern" }, { "name": "Yong Jiang", "gender": "Male", "institution": "Tongyi Lab", "country": "CN", "position": "Researcher" }, { "name": "Zhisong Qiu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Rolnan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 42, "n_ref": 170, "n_ref_all": 196, "n_fig": 8, "n_tab": 6, "L_tab": 2369, "n_element_tab": 335, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1893, "n_element_tab_1": 215, "formula_len_all": 1000, "formula_len_all_1": 992, "len_all": 275970, "len_all_1": 79184, "len_abs": 1417, "len_title": 88, "len_sents": 50473, "len_sents_1": 34486, "n_sents": 404, "n_sents_1": 280, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1303, "title": "Benchmarking Agentic Workflow Generation", "abs": "Large Language Models (LLMs), with their remarkable task-handling capabilities, have catalyzed significant achievements in tackling reasoning and planning tasks, wherein decomposing complex problems into executable workflows is a crucial step in this process. Existing workflow evaluation frameworks either focus solely on holistic performance or suffer from limitations such as restricted scenario coverage, simplistic workflow structures, and lax evaluation standards. To this end, we introduce WorFBench, a unified workflow generation benchmark with multi-faceted scenarios and intricate graph workflow structures. Additionally, we present WorFEval, a systemic evaluation protocol utilizing subsequence and subgraph matching algorithms to accurately quantify the LLM agent's workflow generation capabilities. Through comprehensive evaluations across different types of LLMs, we discover distinct gaps between the sequence planning capabilities and graph planning capabilities of LLM agents, with even GPT-4 exhibiting a gap of around 15%. We also train two open-source models and evaluate their generalization abilities on held-out tasks. Furthermore, we observe that the generated workflows can enhance downstream tasks, enabling them to achieve superior performance with less time during inference.", "keywords": [ "workflow generation", "graph structured planning", "large language model", "agent" ], "rating_list": [ 6, 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 5, 3, 3, 4, 4 ] }, { "paper_id": "vue9P1Ypk6", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Hongyang Gao", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "Assistant Professor" }, { "name": "ZHAONING YU", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 43, "n_ref": 73, "n_ref_all": 85, "n_fig": 4, "n_tab": 9, "L_tab": 2319, "n_element_tab": 428, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1738, "n_element_tab_1": 293, "formula_len_all": 1231, "formula_len_all_1": 1059, "len_all": 142227, "len_all_1": 65744, "len_abs": 1266, "len_title": 134, "len_sents": 40157, "len_sents_1": 26893, "n_sents": 338, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1289, "title": "MAGE: Model-Level Graph Neural Networks Explanations via Motif-based Graph Generation", "abs": "Graph Neural Networks (GNNs) have shown remarkable success in molecular tasks, yet their interpretability remains challenging. Traditional model-level explanation methods like XGNN and GNNInterpreter often fail to identify valid substructures like rings, leading to questionable interpretability. This limitation stems from XGNN's atom-by-atom approach and GNNInterpreter's reliance on average graph embeddings, which overlook the essential structural elements crucial for molecules. To address these gaps, we introduce an innovative **M**otif-b**A**sed **G**NN **E**xplainer (MAGE) that uses motifs as fundamental units for generating explanations. Our approach begins with extracting potential motifs through a motif decomposition technique. Then, we utilize an attention-based learning method to identify class-specific motifs. Finally, we employ a motif-based graph generator for each class to create molecular graph explanations based on these class-specific motifs. This novel method not only incorporates critical substructures into the explanations but also guarantees their validity, yielding results that are human-understandable. Our proposed method's effectiveness is demonstrated through quantitative and qualitative assessments conducted on six real-world molecular datasets.", "keywords": [ "Model-level explanation", "Graph Neural Networks", "Motif" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 2, 4, 1, 2 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 4, 1, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "vuBhwseAKn", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Jia-Ling Zou", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Ozymandis Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Shu Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yimi Wang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Undergrad student" }, { "name": "Yonghong Tian", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Zongying Lin", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Yuan LI", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 16, "n_ref": 18, "n_ref_all": 24, "n_fig": 2, "n_tab": 3, "L_tab": 1351, "n_element_tab": 136, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1351, "n_element_tab_1": 136, "formula_len_all": 482, "formula_len_all_1": 482, "len_all": 97870, "len_all_1": 53427, "len_abs": 1178, "len_title": 141, "len_sents": 27348, "len_sents_1": 24858, "n_sents": 194, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1182, "title": "Deep-ComAIR: A Framework for Predicting TCR-pMHC Binding through Complex Structural Analysis", "abs": "The binding process between T cell receptor (TCR) and the peptide-major histocompatibility complex (pMHC) is a fundamental mechanism in adaptive immunity. Current research on binding prediction primarily emphasizes the sequence and structural features of critical regions within these molecules, often neglecting the intricate structural changes that occur at the binding process, which can lead to biased representations. To address this gap, we propose a novel framework, titled \u201cDeep-ComAIR,\u201d which effectively models the binding process by focusing on the complex structure of TCR-pMHC rather than individual components. This model enhances prediction accuracy by integrating features from three modalities: sequence, structural, and gene. Our approach achieves state-of-the-art results evidenced by an area under the receiver operating characteristic curve (AUROC) of 0.983 in binding reactivity prediction and a Pearson correlation coefficient of 0.833 in binding affinity prediction. These results highlight the framework's potential to deepen our understanding of TCR-pMHC interactions at the structural level and facilitate advancements in immunotherapy and vaccine design.", "keywords": [ "AI for science", "adaptive immunity", "TCR-pMHC binding", "multimodal integration" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "vtcn3DnUCw", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Inderjit S. Dhillon", "gender": "Male", "institution": "Google", "country": "US", "position": "Distinguished Scientist" }, { "name": "Sai Surya Duvvuri", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 21, "n_ref_uni": 41, "n_ref": 67, "n_ref_all": 82, "n_fig": 6, "n_tab": 6, "L_tab": 1305, "n_element_tab": 143, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 81, "n_element_tab_1": 13, "formula_len_all": 2576, "formula_len_all_1": 2098, "len_all": 121325, "len_all_1": 54202, "len_abs": 2023, "len_title": 97, "len_sents": 30919, "len_sents_1": 23347, "n_sents": 247, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1190, "title": "LASER: Attention using Exponential Transformation", "abs": "Transformers have had tremendous impact for several sequence related tasks. The softmax based dot-product attention mechanism plays a key role in the Transformer's ability to retrieve from any part of the sequence via a parameterized query-key-value mechanism. However, the softmax operation can backpropagate small gradients thus inhibiting learning. In this paper, we fix this by introducing a new attention mechanism called LASER attention, which admits a log-sum-exp structure and propagates a larger gradient signal. We show that LASER attention can be implemented by making small modifications to existing attention implementations. We conduct experiments on large language models (LLMs) with upto 2.2 billion parameters where we show improvements of upto 3.38\\% and 1\\% on an average compared to standard attention on downstream one-shot evaluations. We also evaluate on transformers spanning different modalities (vision, speech and text): Vision Transformer (ViT) on Imagenet (1.2\\% improvement in accuracy), Conformer on the Librispeech speech-to-text task (2.25\\% relative improvement) and encoder-only BERT Transformer with 2.2 billion parameters (0.93\\% relative improvement).", "keywords": [ "Attention Mechanism", "LLM", "Transformer", "Conformer", "ViT" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "vtUbXd5Cyg", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Muyao Niu", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Qingtian Zhu", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Yinqiang Zheng", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhihang Zhong", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Zhao Jiancheng", "gender": "Male", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "JP", "position": "PhD student" }, { "name": "Mingze MA", "gender": "Male", "institution": "Tokyo University", "country": "JP", "position": "MS student" }, { "name": "Xiao Sun", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "YIFAN ZHAN", "gender": "Male", "institution": "University of Tokyo", "country": "JP", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 40, "n_ref": 100, "n_ref_all": 126, "n_fig": 8, "n_tab": 4, "L_tab": 1731, "n_element_tab": 298, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 634, "n_element_tab_1": 76, "formula_len_all": 560, "formula_len_all_1": 580, "len_all": 133891, "len_all_1": 58675, "len_abs": 1245, "len_title": 143, "len_sents": 35576, "len_sents_1": 26637, "n_sents": 280, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 1250, "title": "ToMiE: Towards Modular Growth in Enhanced SMPL Skeleton for 3D Human Gaussians with Animatable Garments", "abs": "In this paper, we highlight a critical yet often overlooked factor in most 3D human tasks, namely modeling humans with complex garments. \nIt is known that the parameterized formulation of SMPL is able to fit human skin; while complex garments, e.g., hand-held objects and loose-fitting garments, are difficult to get modeled within the unified framework, since their movements are usually decoupled with the human body.\nTo enhance the capability of SMPL skeleton in response to this situation, we propose a modular growth strategy that enables the joint tree of the skeleton to expand adaptively. Specifically, our method, called ToMiE, consists of parent joints localization and external joints optimization. For parent joints localization, we employ a gradient-based approach guided by both LBS blending weights and motion kernels. Once the external joints are obtained, we proceed to optimize their transformations in SE(3) across different frames, enabling rendering and explicit animation. ToMiE manages to outperform other methods across various cases with garments, not only in rendering quality but also by offering free animation of grown joints, thereby enhancing the expressive ability of SMPL skeleton for a broader range of applications.", "keywords": [ "Human Gaussians", "Adaptive Growth", "Animatable Garments" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "vtT09dYPGI", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Qiong Wu", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoshuai Sun", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Yiyi Zhou", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhaoxi Ke", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 22, "n_ref": 64, "n_ref_all": 69, "n_fig": 3, "n_tab": 6, "L_tab": 2130, "n_element_tab": 118, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2443, "n_element_tab_1": 130, "formula_len_all": 633, "formula_len_all_1": 632, "len_all": 147451, "len_all_1": 49536, "len_abs": 1525, "len_title": 145, "len_sents": 29258, "len_sents_1": 19058, "n_sents": 255, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1517, "title": "Routing Experts: Learning to Route Dynamic Experts in Existing Multi-modal Large Language Models", "abs": "Recently, mixture of experts (MoE) has become a popular paradigm for achieving the trade-off between modal capacity and efficiency of multimodal large language models (MLLMs). Different from previous efforts, we are dedicated to exploring the dynamic experts in existing MLLMs and showing that a standard MLLM can also be a mixture of experts. However, achieving this target is still notoriously challenging. The well-trained MLLMs are more accustomed to the fixed pathway and a drastic change in its inference manner also greatly impedes its performance. To address these issues, we propose a novel dynamic expert routing method for existing MLLMs, termed Routing Experts (RoE), which can achieve example-dependent optimal path routing without obvious structure tweaks. Meanwhile, a new structure sparsity regularization is also introduced to force the well-trained MLLMs to learn more short-cut pathways. In addition, we also address the alignment of the training and inference of MLLMs in terms of network routing. To validate RoE, we apply it to a set of existing MLLMs, including LLaVA-1.5, LLaVA-HR and VILA, and conduct extensive experiments on a bunch of VL benchmarks. The experiment results not only show the effectiveness of our RoE in improving MLLMs' efficiency, but also yield obvious advantages over MoE-LLaVA in both performance and speed, e.g., an average performance gain of 3.3% on 5 benchmarks while being 1.61 times faster. Our code is anonymously released at https://github.com/DoubtedSteam/RoE", "keywords": [ "multimodal large language model", "dynamic routing" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "vtGLtSxtqv", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Kongcheng Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Mingli Song", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Shunyu Liu", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Tongya Zheng", "gender": "Male", "institution": "Hangzhou City University", "country": "CN", "position": "Postdoc" }, { "name": "Wenkai Fang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yaoru Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yuxuan Zheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhenyu Cui", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 48, "n_ref": 145, "n_ref_all": 161, "n_fig": 7, "n_tab": 8, "L_tab": 3783, "n_element_tab": 185, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1988, "n_element_tab_1": 100, "formula_len_all": 633, "formula_len_all_1": 235, "len_all": 267164, "len_all_1": 54202, "len_abs": 1473, "len_title": 107, "len_sents": 82240, "len_sents_1": 23863, "n_sents": 672, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1481, "title": "Odyssey: Empowering Minecraft Agents with Open-World Skills", "abs": "Recent studies have delved into constructing generalist agents for open-world environments like Minecraft. Despite the encouraging results, existing efforts mainly focus on solving basic programmatic tasks, e.g., material collection and tool-crafting following the Minecraft tech-tree, treating the ObtainDiamond task as the ultimate goal. This limitation stems from the narrowly defined set of actions available to agents, requiring them to learn effective long-horizon strategies from scratch. Consequently, discovering diverse gameplay opportunities in the open world becomes challenging. In this work, we introduce Odyssey, a new framework that empowers Large Language Model (LLM)-based agents with open-world skills to explore the vast Minecraft world. Odyssey comprises three key parts: (1) An interactive agent with an open-world skill library that consists of 40 primitive skills and 183 compositional skills. (2) A fine-tuned LLaMA-3 model trained on a large question-answering dataset with 390k+ instruction entries derived from the Minecraft Wiki. (3) A new agent capability benchmark includes the long-term planning task, the dynamic-immediate planning task, and the autonomous exploration task. Extensive experiments demonstrate that the proposed Odyssey framework can effectively evaluate different capabilities of LLM-based agents. All datasets, model weights, and code are publicly available to motivate future research on more advanced autonomous agent solutions.", "keywords": [ "Autonomous Agents", "Large Language Models", "Open-World Environments" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "vtCkb4KJxr", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Shuo Xing", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Victoria G. Crawford", "gender": "Female", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" }, { "name": "Wenjing Chen", "gender": "Female", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" } ], "n_formula": 137, "n_formula_1": 10, "n_ref_uni": 41, "n_ref": 114, "n_ref_all": 139, "n_fig": 10, "n_tab": 2, "L_tab": 694, "n_element_tab": 49, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1344, "n_element_tab_1": 38, "formula_len_all": 10002, "formula_len_all_1": 613, "len_all": 252637, "len_all_1": 65281, "len_abs": 1114, "len_title": 114, "len_sents": 90173, "len_sents_1": 30788, "n_sents": 797, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1132, "title": "Adaptive Threshold Sampling for Fast Noisy Submodular Maximization", "abs": "We address the problem of submodular maximization where objective function $f:2^U\\to\\mathbb{R}_{\\geq 0}$ can only be accessed through i.i.d noisy queries. This problem arises in many applications including influence maximization, diverse recommendation systems, and large-scale facility location optimization. We propose an efficient adaptive sampling strategy, called Confident Sample (CS), that is inspired by algorithms for best-arm-identification in multi-armed bandit, which significantly improves sample efficiency. We integrate CS into existing approximation algorithms for submodular maximization, resulting in algorithms with approximation guarantees arbitrarily close to the standard value oracle setting that are highly sample-efficient. We propose and analyze sample-efficient algorithms for monotone submodular maximization with cardinality and matroid constraints, as well as unconstrained non-monotone submodular maximization. Our theoretical analysis is complemented by empirical evaluation on real instances, demonstrating the superior sample efficiency of our proposed algorithm relative to alternative approaches.", "keywords": [ "submodular", "multi-armed bandit", "bandit feedback", "best-arm identification", "combinatorial optimization" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 2 ] }, { "paper_id": "vszlHtUvSR", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Dongzi Wang", "gender": "Male", "institution": "National University of Defense Technology", "country": "", "position": "PhD student" }, { "name": "Lilan Huang", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Minglong Li", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Muning Wen", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "TENG LI", "gender": "unknown", "institution": "National University of Defense Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiao Teng", "gender": "Male", "institution": "National University of Defense Technology", "country": "", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 12, "n_ref_uni": 21, "n_ref": 36, "n_ref_all": 50, "n_fig": 4, "n_tab": 1, "L_tab": 443, "n_element_tab": 39, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 443, "n_element_tab_1": 39, "formula_len_all": 1630, "formula_len_all_1": 1217, "len_all": 103241, "len_all_1": 60473, "len_abs": 1298, "len_title": 141, "len_sents": 30924, "len_sents_1": 26864, "n_sents": 234, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1303, "title": "RDHNet: Addressing Rotational and Permutational Symmetries in Continuous Multi-Agent Systems", "abs": "Symmetry is prevalent in multi-agent systems. The presence of symmetry, coupled with the misuse of absolute coordinate systems, often leads to a large amount of redundant representation space, significantly increasing the search space for learning policies and reducing learning efficiency. Effectively utilizing symmetry and extracting symmetry-invariant representations can significantly enhance multi-agent systems' learning efficiency and overall performance by compressing the model's hypothesis space and improving sample efficiency. The issue of rotational symmetry in multi-agent reinforcement learning has received little attention in previous research and is the primary focus of this paper. To address this issue, we propose a rotation-invariant network architecture for continuous action space tasks. This architecture utilizes relative coordinates between agents, eliminating dependence on absolute coordinate systems, and employs a hypernetwork to enhance the model's fitting capability, enabling it to model MDPs with more complex dynamics. It can be used for both predicting actions and evaluating action values/utilities. In benchmark tasks, experimental results validate the impact of rotational symmetry on multi-agent decision systems and demonstrate the effectiveness of our method.", "keywords": [ "Multi-agent", "Reinforcement Learning", "Symmetry" ], "rating_list": [ 5, 3, 1 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "vsYt8UHGzI", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Fandong Meng", "gender": "Male", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Huiqi Hu", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Jie Zhou", "gender": "Male", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Ning Cheng", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Pengyu Zhao", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Wenjuan Han", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiuwen Xu", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xue Zhang", "gender": "unknown", "institution": "Beijing Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Zijian Jin", "gender": "Male", "institution": "New York University", "country": "", "position": "MS student" }, { "name": "Xu Jinan", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 33, "n_ref": 56, "n_ref_all": 80, "n_fig": 5, "n_tab": 6, "L_tab": 1467, "n_element_tab": 161, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1467, "n_element_tab_1": 165, "formula_len_all": 0, "formula_len_all_1": 58, "len_all": 121701, "len_all_1": 56375, "len_abs": 1453, "len_title": 182, "len_sents": 35535, "len_sents_1": 25028, "n_sents": 267, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 133, "L_abs": 1458, "title": "Bridging the Reality Gap: A Benchmark for Physical Reasoning in General World Models with Various Physical Phenomena beyond Mechanics", "abs": "While general world models have demonstrated excellent capability in modeling and simulating the world through video understanding and generation, their ability to reason about physical phenomena beyond mechanics remains underexplored. This includes crucial aspects like thermodynamics, electromagnetism, and optics, all of which are fundamental for simulating and predicting real-world dynamics. Existing benchmarks for evaluating physical reasoning in models often rely on datasets consisting solely of simulator-generated, virtual videos, limiting their generalizability to real-world scenarios. This limitation hinders the comprehensive evaluation of general world models' physical reasoning in real-world scenarios. To bridge this gap, we introduce the Physics-RW benchmark, a physical reasoning dataset constructed from real-world videos. Encompassing a broad spectrum of real-world phenomena\u2014mechanics, thermodynamics, electromagnetism, and optics\u2014Physics-RW offers a comprehensive evaluation platform. We conducted extensive experiments on the Physics-RW benchmark, and the results indicate that there is still significant room for improvement in the physical reasoning abilities of general world models. We further analyzed the experimental results and explored several avenues for improvement. Virtual environment finetuning and physical knowledge injection via prompts demonstrate the potential for enhancing zero-shot physical reasoning ability.", "keywords": [ "Physical Reasoning", "General World Models", "Zero-shot Inference" ], "rating_list": [ 3, 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 2, 4, 3, 2, 2 ], "contribution_list": [ 1, 3, 3, 2, 3 ], "confidence_list": [ 5, 4, 5, 4, 4 ] }, { "paper_id": "vsU2veUpiR", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Aaquib Syed", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "Undergrad student" }, { "name": "Abhay Sheshadri", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Aidan Ewart", "gender": "unknown", "institution": "University of Bristol", "country": "", "position": "Undergrad student" }, { "name": "Gintare Karolina Dziugaite", "gender": "Female", "institution": "McGill University", "country": "CA", "position": "Adjunct Professor" }, { "name": "Phillip Huang Guo", "gender": "Male", "institution": "Gray Swan AI", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 39, "n_ref": 68, "n_ref_all": 104, "n_fig": 25, "n_tab": 5, "L_tab": 1511, "n_element_tab": 127, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 245, "n_element_tab_1": 35, "formula_len_all": 45, "formula_len_all_1": 0, "len_all": 159092, "len_all_1": 56543, "len_abs": 1343, "len_title": 140, "len_sents": 50339, "len_sents_1": 29014, "n_sents": 324, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1361, "title": "Mechanistic Unlearning: Robust Knowledge Unlearning and Editing via Mechanistic Localization", "abs": "Methods for knowledge editing and unlearning in large language models seek to edit or remove undesirable knowledge or capabilities without compromising general language modeling performance. This work investigates how mechanistic interpretability---which, in part, aims to identify model components (circuits) associated to specific interpretable mechanisms that make up a model capability---can improve the precision and effectiveness of editing and unlearning. \nWe find a stark difference in unlearning and edit robustness when training components localized by different methods. We highlight an important distinction between methods that localize components based primarily on preserving outputs, and those finding high level mechanisms with predictable intermediate states.\nIn particular, localizing edits/unlearning to components associated with the \\textit{lookup-table mechanism} for factual recall 1) leads to more robust edits/unlearning across different input/output formats, and 2) resists attempts to relearn the unwanted information, while also reducing unintended side effects compared to baselines, on both a sports facts dataset and the CounterFact dataset across multiple models.\nWe also find that certain localized edits disrupt the latent knowledge in the model more than any other baselines, making unlearning more robust to various attacks.", "keywords": [ "Model Editing", "Unlearning", "Mechanistic Interpretability", "Localization", "Adversarial Robustness" ], "rating_list": [ 8, 3, 5, 5 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 4, 5, 2, 3 ] }, { "paper_id": "vsLohTBH4h", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Xianliang Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhongyi Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 368, "n_formula_1": 43, "n_ref_uni": 57, "n_ref": 130, "n_ref_all": 173, "n_fig": 0, "n_tab": 1, "L_tab": 65, "n_element_tab": 27, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 38534, "formula_len_all_1": 3120, "len_all": 303137, "len_all_1": 64721, "len_abs": 860, "len_title": 140, "len_sents": 75788, "len_sents_1": 25982, "n_sents": 843, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 865, "title": "Refined Generalization Analysis of the Deep Ritz Method and Physics-Informed Neural Networks", "abs": "In this paper, we derive refined generalization bounds for the Deep Ritz Method (DRM) and Physics-Informed Neural Networks (PINNs). For the DRM, we focus on two prototype elliptic partial differential equations (PDEs): Poisson equation and static Schr\u00f6dinger equation on the $d$-dimensional unit hypercube with the Neumann boundary condition. Furthermore, sharper generalization bounds are derived based on the localization techniques under the assumptions that the exact solutions of the PDEs lie in the Barron spaces or the general Sobolev spaces. For the PINNs, we investigate the general linear second order elliptic PDEs with Dirichlet boundary condition using the local Rademacher complexity in the multi-task learning setting. Finally, we discuss the generalization error in the setting of over-parameterization when solutions of PDEs belong to Barron space.", "keywords": [ "Deep Ritz Method", "Physics-Informed Neural Networks", "Generalization analysis", "Fast Rate" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "vrCT5uCdYp", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Chao Yu", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Feng Gao", "gender": "Male", "institution": "IIIS, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Shu'ang Yu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yi Wu", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Assistant Professor" }, { "name": "Yu Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 41, "n_ref": 92, "n_ref_all": 111, "n_fig": 8, "n_tab": 9, "L_tab": 7817, "n_element_tab": 487, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 2750, "n_element_tab_1": 380, "formula_len_all": 410, "formula_len_all_1": 369, "len_all": 140614, "len_all_1": 63515, "len_abs": 1270, "len_title": 91, "len_sents": 35617, "len_sents_1": 26685, "n_sents": 296, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1272, "title": "FlightBench: Benchmarking Learning-based Methods for Ego-vision-based Quadrotors Navigation", "abs": "Ego-vision-based navigation in cluttered environments is crucial for mobile systems, particularly agile quadrotors. While learning-based methods have shown promise recently, head-to-head comparisons with cutting-edge optimization-based approaches are scarce, leaving open the question of where and to what extent they truly excel. In this paper, we introduce FlightBench, the first comprehensive benchmark that implements various learning-based methods for ego-vision-based navigation and evaluates them against mainstream optimization-based baselines using a broad set of performance metrics. Additionally, we develop a suite of criteria to assess scenario difficulty and design test cases that span different levels of difficulty based on these criteria. Our results show that while learning-based methods excel in high-speed flight and faster inference, they struggle with challenging scenarios like sharp corners or view occlusion. Analytical experiments validate the correlation between our difficulty criteria and flight performance. We hope this benchmark and these criteria will drive future advancements in learning-based navigation for ego-vision quadrotors. The source code and documentation is available at https://github.com/Anonymous314159265358/FlightBench.", "keywords": [ "Ego-vision-based Navigation", "Learning-based Quadrotor Methods", "Open-source Benchmark" ], "rating_list": [ 3, 8, 5, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "vr1QdCNJmN", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Hideitsu Hino", "gender": "Male", "institution": "The Institute of Statistical Mathematics, Japan", "country": "JP", "position": "Professor" }, { "name": "Masanari Kimura", "gender": "unknown", "institution": "University of Melbourne", "country": "AU", "position": "Postdoc" }, { "name": "Takahiro Kawashima", "gender": "Male", "institution": "ZOZO Research", "country": "JP", "position": "Researcher" }, { "name": "Tasuku Soma", "gender": "Male", "institution": "The Institute of Statistical Mathematics, Japan", "country": "", "position": "Associate Professor" } ], "n_formula": 46, "n_formula_1": 26, "n_ref_uni": 45, "n_ref": 81, "n_ref_all": 123, "n_fig": 2, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5765, "formula_len_all_1": 3317, "len_all": 139354, "len_all_1": 69587, "len_abs": 1246, "len_title": 92, "len_sents": 37674, "len_sents_1": 28515, "n_sents": 380, "n_sents_1": 268, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 1250, "title": "Difference-of-submodular Bregman Divergence", "abs": "The Bregman divergence, which is generated from a convex function, is commonly used as a pseudo-distance for comparing vectors or functions in continuous spaces. In contrast, defining an analog of the Bregman divergence for discrete spaces is nontrivial. Iyer & Bilmes (2012b) considered Bregman divergences on discrete domains using submodular functions as generating functions, the discrete analogs of convex functions. In this paper, we further generalize this framework to cases where the generating function is neither submodular nor supermodular, thus increasing the flexibility and representational capacity of the resulting divergence, which we term the difference-of-submodular Bregman divergence. Additionally, we introduce a learnable form of this divergence using permutation-invariant neural networks (NNs) and demonstrate through experiments that it effectively captures key structural properties in discrete data. As a result, the proposed method significantly improves the performance of existing methods on tasks such as clustering and set retrieval problems. This work addresses the challenge of defining meaningful divergences in discrete settings and provides a new tool for tasks requiring structure-preserving distance measures.", "keywords": [ "Bregman Divergence", "Permutation-invariant neural networks", "Metric learning", "Submodular functions" ], "rating_list": [ 8, 6, 5, 8, 8 ], "soundness_list": [ 3, 2, 2, 2, 4 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 4 ], "confidence_list": [ 3, 3, 3, 2, 5 ] }, { "paper_id": "vqgDq1uycO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Honglu Zhou", "gender": "Female", "institution": "SalesForce.com", "country": "US", "position": "Research Scientist" }, { "name": "Jihoon Chung", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Juan Carlos Niebles", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Director, Research" }, { "name": "Max Gonzalez Saez-Diez", "gender": "unknown", "institution": "Department of Computer Science, Princeton University", "country": "US", "position": "MS student" }, { "name": "Olga Russakovsky", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Associate Professor" }, { "name": "Tyler Zhu", "gender": "Male", "institution": "Department of Computer Science, Princeton University", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 38, "n_ref": 63, "n_ref_all": 89, "n_fig": 9, "n_tab": 9, "L_tab": 3567, "n_element_tab": 383, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 19, "n_element_tab_1": 1, "formula_len_all": 135, "formula_len_all_1": 136, "len_all": 217550, "len_all_1": 62417, "len_abs": 1585, "len_title": 110, "len_sents": 57009, "len_sents_1": 31648, "n_sents": 539, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1501, "title": "Unifying Specialized Visual Encoders for Video Language Models", "abs": "The recent advent of Large Language Models (LLMs) has ushered sophisticated reasoning capabilities into the realm of video through Video Large Language Models (VideoLLMs). However, VideoLLMs currently rely on a single vision encoder for all of their visual processing, which limits the amount and type of visual information that can be conveyed to the LLM. Our method, MERV, Multi-Encoder Representation of Videos, instead leverages multiple frozen visual encoders to create a unified representation of a video, providing the VideoLLM with a comprehensive set of specialized visual knowledge. Spatio-temporally aligning the features from each encoder allows us to tackle a wider range of open-ended and multiple-choice video understanding questions and outperform prior state-of-the-art works on their data mixes. MERV is up to 3.79% better in accuracy than Video-LLaVA across the standard suite video understanding benchmarks, while also having a better Video-ChatGPT score. We also improve upon SeViLA, the previous best on zero-shot Perception Test accuracy, by 2.21%. MERV introduces minimal extra parameters and trains faster than equivalent single-encoder approaches. Finally, we provide qualitative evidence that our model captures domain knowledge from each encoder simultaneously, such as on the motion classification tasks found in Something-Something v2. Our results offer promising directions for future research in utilizing multiple vision encoders for comprehensive video understanding.", "keywords": [ "video understanding", "multimodal llms" ], "rating_list": [ 6, 5, 5, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "vqbd2OQnGp", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Karthik Prasad", "gender": "unknown", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Mingrui Wu", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Sheng Cao", "gender": "Male", "institution": "Meta", "country": "US", "position": "AI Researcher" }, { "name": "Yuandong Tian", "gender": "Male", "institution": "Meta AI (FAIR)", "country": "", "position": "Research Scientist" }, { "name": "Zechun Liu", "gender": "unknown", "institution": "Meta Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 37, "n_ref": 81, "n_ref_all": 85, "n_fig": 1, "n_tab": 4, "L_tab": 4385, "n_element_tab": 261, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 3078, "n_element_tab_1": 102, "formula_len_all": 345, "formula_len_all_1": 254, "len_all": 144013, "len_all_1": 49378, "len_abs": 919, "len_title": 130, "len_sents": 39110, "len_sents_1": 21677, "n_sents": 260, "n_sents_1": 125, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 922, "title": "Knowledge And Capability Transfer Through Large Language Models' Parameters Fusing", "abs": "The post-training phase of large language models (LLMs) plays a pivotal role in refining models to follow instructions and align with human preferences. However, this phase is fraught with challenges, particularly in sourcing high-quality post-training data. This paper introduces a novel approach, termed Parameters Fusing, that simplifies the post-training process by amalgamating model parameters delta from existing instruct-tuned checkpoints with a new base model tailored to specific domain data obtained by continual pre-training. Utilizing open-weight models such as Meta's Llama, our method replicates the effects of the traditional post-training phase while significantly reducing both time and resource costs. This approach not only minimizes the challenges of post-training data acquisition but also provides a flexible and efficient framework for enhancing LLMs with domain-specific knowledge or capabilities.", "keywords": [ "large language model", "post-training", "transfer learning", "model merging", "weights averaging", "artificial intelligence" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 1, 4, 3, 3 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "vqJZb9SX1T", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Changxing Ding", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Chao Xue", "gender": "Male", "institution": "JD Explore Academy", "country": "CN", "position": "Researcher" }, { "name": "Fei Wang", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "LIANG DING", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Ye LIU", "gender": "unknown", "institution": "South China University of Technology", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 37, "n_formula_1": 14, "n_ref_uni": 37, "n_ref": 77, "n_ref_all": 95, "n_fig": 4, "n_tab": 7, "L_tab": 2417, "n_element_tab": 242, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2781, "n_element_tab_1": 164, "formula_len_all": 4504, "formula_len_all_1": 897, "len_all": 131200, "len_all_1": 65993, "len_abs": 1409, "len_title": 155, "len_sents": 31226, "len_sents_1": 26190, "n_sents": 295, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 106, "L_abs": 1433, "title": "Simultaneous Computation and Memory Efficient Zeroth-Order Optimizer for Fine-Tuning Large Language Models", "abs": "Fine-tuning is powerful for adapting large language models to downstream tasks, but it often results in huge memory usages. \nA promising approach to mitigate this is using Zeroth-Order (ZO) optimization, which estimates gradients to replace First-Order (FO) gradient calculations, albeit with longer training time due to its stochastic nature. \nBy revisiting the Memory-efficient ZO (MeZO) optimizer, we discover that the full-parameter perturbation and updating processes consume over 50\\% of its overall fine-tuning time cost. \nBased on these observations, we introduce a novel layer-wise sparse computation and memory efficient ZO optimizer, named LeZO \nLeZO treats layers as fundamental units for sparsification and dynamically perturbs different parameter subsets in each step to achieve full-parameter fine-tuning. \nLeZO incorporates layer-wise parameter sparsity in the process of simultaneous perturbation stochastic approximation (SPSA) and ZO stochastic gradient descent (ZO-SGD). \nIt achieves accelerated computation during perturbation and updating processes without additional memory overhead.\nWe conduct extensive experiments with the OPT model family on the SuperGLUE benchmark and two generative tasks. \nThe experiments show that LeZO accelerates training without compromising the performance of ZO optimization.\nSpecifically, it achieves over $3 \\times$ speedup compared to MeZO on the SST-2, BoolQ, and Copa tasks.", "keywords": [ "zeroth-order optimization", "large language models" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "vpo2K9Xivv", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Emi Zeger", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Mert Pilanci", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" } ], "n_formula": 87, "n_formula_1": 30, "n_ref_uni": 35, "n_ref": 76, "n_ref_all": 147, "n_fig": 12, "n_tab": 1, "L_tab": 723, "n_element_tab": 9, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7775, "formula_len_all_1": 2684, "len_all": 174639, "len_all_1": 67518, "len_abs": 955, "len_title": 115, "len_sents": 49837, "len_sents_1": 28307, "n_sents": 523, "n_sents_1": 244, "venue": "ICLR2025", "decision": "withdraw", "L_title": 115, "L_abs": 959, "title": "Black Boxes and Looking Glasses: Multilevel Symmetries, Reflection Planes, and Convex Optimization in Deep Networks", "abs": "We show that training deep neural networks (DNNs) with absolute value activation and arbitrary input dimension can be formulated as equivalent convex Lasso problems with novel features expressed using geometric algebra. This formulation reveals geometric structures encoding symmetry in neural networks. Using the equivalent Lasso form of DNNs, we formally prove a fundamental distinction between deep and shallow networks: deep networks inherently favor symmetric structures in their fitted functions, with greater depth enabling multilevel symmetries, i.e., symmetries within symmetries. Moreover, Lasso features represent distances to hyperplanes that are reflected across training points. These reflection hyperplanes are spanned by training data and are orthogonal to optimal weight vectors. Numerical experiments support theory and demonstrate theoretically predicted features when training networks using embeddings generated by Large Language Models.", "keywords": [ "deep neural networks", "convex optimization", "geometric algebra", "Lasso model", "sparsity" ], "rating_list": [ 5, 3, 5, 3, 3 ], "soundness_list": [ 3, 1, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 1, 1 ], "confidence_list": [ 2, 2, 3, 3, 3 ] }, { "paper_id": "vpKjmJp6cO", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Dohyeong Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Jungwoo Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Kyungjae Lee", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Assistant Professor" }, { "name": "Seokhun Ju", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Seungyub Han", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Taehyun Cho", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" } ], "n_formula": 85, "n_formula_1": 29, "n_ref_uni": 42, "n_ref": 81, "n_ref_all": 93, "n_fig": 3, "n_tab": 4, "L_tab": 594, "n_element_tab": 26, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 325, "n_element_tab_1": 6, "formula_len_all": 15900, "formula_len_all_1": 4741, "len_all": 183759, "len_all_1": 70905, "len_abs": 1193, "len_title": 176, "len_sents": 47886, "len_sents_1": 29242, "n_sents": 442, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 127, "L_abs": 1280, "title": "Bellman Unbiasedness: Toward Provably Efficient Distributional Reinforcement Learning with General Value Function Approximation", "abs": "Distributional reinforcement learning improves performance by effectively capturing environmental stochasticity. \nHowever, existing research on its regret analysis has relied heavily on structural assumptions that are difficult to implement in practice.\nIn particular, there has been little attention to the infeasibility issue of dealing with the infinite-dimensionality of a distribution.\nTo overcome this infeasibility, we present a regret analysis of distributional reinforcement learning with general value function approximation in a finite episodic Markov decision process setting through *statistical functional dynamic programming*. \nWe first introduce a key notion of *Bellman unbiasedness* which is essential for exactly learnable and provably efficient updates.\nOur theoretical results demonstrate that the only way to exactly capture statistical information, including nonlinear statistical functionals, is by representing the infinite-dimensional return distribution with a finite number of moment functionals.\nSecondly, we propose a provably efficient algorithm, *SF-LSVI*, that achieves a tight regret bound of $\\tilde{O}(d_E H^{\\frac{3}{2}}\\sqrt{K})$ where $H$ is the horizon, $K$ is the number of episodes, and $d_E$ is the eluder dimension of a function class.", "keywords": [ "Distributional Reinforcement Learning", "Regret Analysis", "General Value Function Approximation" ], "rating_list": [ 6, 6, 6, 6, 3 ], "soundness_list": [ 3, 3, 4, 3, 4 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 4, 3 ] }, { "paper_id": "vodsIF3o7N", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Bogdan Mazoure", "gender": "Male", "institution": "Apple", "country": "GB", "position": "Research Scientist" }, { "name": "Martin Klissarov", "gender": "Unspecified", "institution": "McGill University", "country": "", "position": "PhD student" }, { "name": "R Devon Hjelm", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Alexander Toshev", "gender": "Male", "institution": "Apple", "country": "US", "position": "research scientist" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 71, "n_ref": 125, "n_ref_all": 160, "n_fig": 20, "n_tab": 5, "L_tab": 3209, "n_element_tab": 178, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 130, "n_element_tab_1": 8, "formula_len_all": 525, "formula_len_all_1": 400, "len_all": 225076, "len_all_1": 62158, "len_abs": 1118, "len_title": 133, "len_sents": 58065, "len_sents_1": 31655, "n_sents": 439, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1121, "title": "On the Modeling Capabilities of Large Language Models for Sequential Decision Making", "abs": "Large pretrained models are showing increasingly better performance in reasoning and planning tasks across different modalities, opening the possibility to leverage them for complex sequential decision making problems. In this paper, we investigate the capabilities of Large Language Models (LLMs) for reinforcement learning (RL) across a diversity of interactive domains. We evaluate their ability to produce decision-making policies, either directly, by generating actions, or indirectly, by first generating reward models to train an agent with RL. Our results show that, even without task-specific fine-tuning, LLMs excel at reward modeling. In particular, crafting rewards through artificial intelligence (AI) feedback yields the most generally applicable approach and can enhance performance by improving credit assignment and exploration. Finally, in environments with unfamiliar dynamics, we explore how fine-tuning LLMs with synthetic data can significantly improve their reward modeling capabilities while mitigating catastrophic forgetting, further broadening their utility in sequential decision-making tasks.", "keywords": [ "reinforcement learning", "large language models", "ai agents", "preference based learning", "reward design" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "voYshhbWeJ", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Balu Harshavardan Koduru", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "MS student" }, { "name": "David Doermann", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Full Professor" }, { "name": "Junsong Yuan", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Full Professor" }, { "name": "Nan Xi", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Shun Liu", "gender": "Male", "institution": "Shanghai University of Finance and Economics", "country": "CN", "position": "Undergrad student" }, { "name": "Tenzin Lhakpa", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "MS student" }, { "name": "Tianyu Luan", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Xi Tang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Xuan Gong", "gender": "Female", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Yuan Zhang", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Yuanhao Zhai", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Yunjie Tian", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Postdoc" }, { "name": "Yuxuan Sun", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Ziqing Xue", "gender": "Female", "institution": "Hebei United University", "country": "CN", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 47, "n_ref": 79, "n_ref_all": 92, "n_fig": 4, "n_tab": 2, "L_tab": 959, "n_element_tab": 71, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 64, "formula_len_all_1": 64, "len_all": 161944, "len_all_1": 59702, "len_abs": 2391, "len_title": 115, "len_sents": 38015, "len_sents_1": 31478, "n_sents": 260, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 113, "L_abs": 1709, "title": "EndoAssistant: A Large-scale Vision-Language Dataset for Endoscopic Surgery Understanding from Open-Source Videos", "abs": "Endoscopic interventions offer a minimally invasive approach, minimizing patient discomfort and facilitating expedited recovery. Proficient training of junior surgeons necessitates the ability to analyze and interpret endoscopic scenes through questioning and answering. Consequently, the development of a robust foundation model for endoscopic visual language understanding holds immense value for medical training and surgical education. However, existing endoscopy vision-language datasets are limited in scale and diversity, consisting of only 50 videos sourced from a few clinical sites, thus posing a significant hurdle to the advancement of generalized and robust artificial intelligence models for endoscopic surgical applications. To address this challenge, we present a large-scale, meticulously curated image-text dataset of surgical endoscopic scenes from expert surgeons, designed to propel a vision-language assistant in medical scene understanding. Encompassing 590 open-source videos spanning more than 91 hours, our curated dataset includes 65,844 unique images, 30,002 unique captions, and 157,589 image-caption/question-answering pairs. This dataset aims to assist the development of automated systems to support medical professionals by mitigating repetitive tasks. We present a comprehensive endoscopic surgery assisting pipeline, (1) a first-ever image-caption dataset specifically for endoscopic scenes; (2) an image-question-answer dataset that offers greater size and diversity compared to existing collections; (3) rigorous evaluation demonstrating its efficacy in downstream surgical endoscopic scene comprehension tasks like classification, retrieval and visual question answering.", "keywords": [ "Medical image", "endoscopy", "vision-language model" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "vo9t20wsmd", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Ankit Singh Rawat", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Harikrishna Narasimhan", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Neha Gupta", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Sanjiv Kumar", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Seungyeon Kim", "gender": "unknown", "institution": "Google DeepMind", "country": "US", "position": "Research Engineer (L7)" }, { "name": "Wittawat Jitkrittum", "gender": "Male", "institution": "Google Research", "country": "US", "position": "Research Scientist" }, { "name": "Aditya Krishna Menon", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" } ], "n_formula": 77, "n_formula_1": 31, "n_ref_uni": 50, "n_ref": 170, "n_ref_all": 216, "n_fig": 12, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 850, "n_element_tab_1": 83, "formula_len_all": 8453, "formula_len_all_1": 3010, "len_all": 248389, "len_all_1": 71617, "len_abs": 5963, "len_title": 88, "len_sents": 70859, "len_sents_1": 32060, "n_sents": 580, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 40, "L_abs": 1146, "title": "Faster Cascades via Speculative Decoding", "abs": "Cascades and speculative decoding are two common approaches to improving language models' inference efficiency. Both approaches involve interleaving models of different sizes, but via fundamentally distinct mechanisms: cascades employ a deferral rule that invokes the larger model only for \"hard\" inputs, while speculative decoding uses speculative execution to primarily invoke the larger model in parallel verification mode. These mechanisms offer different benefits: empirically, cascades offer better cost-quality trade-offs, often even outperforming the large model, while theoretically, speculative decoding offers a guarantee of quality-neutrality. In this paper, we leverage the best of both these approaches by designing new speculative cascading techniques that implement their deferral rule through speculative execution. We characterize the optimal deferral rule for our speculative cascades, and employ a plug-in approximation to the optimal rule. Experiments with Gemma and T5 models on a range of language benchmarks show that our approach yields better cost quality trade-offs than cascading and speculative decoding baselines.", "keywords": [ "Cascades", "Speculative Decoding", "Speculative execution", "LLM", "Inference", "Adaptive Inference" ], "rating_list": [ 8, 6, 3 ], "soundness_list": [ 4, 3, 1 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 3, 1 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "vo5Md2RCWq", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Guodao Sun", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Haixia Wang", "gender": "Female", "institution": "Zhejiang University of Technology", "country": "", "position": "Full Professor" }, { "name": "Jingwei Tang", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "", "position": "PhD student" }, { "name": "Li Jiang", "gender": "Female", "institution": "Zhejiang University of Technology", "country": "", "position": "Associate Professor" }, { "name": "Qi Jiang", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Ronghua Liang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Tong Li", "gender": "Female", "institution": "Zhejiang University of Technology", "country": "", "position": "PhD student" }, { "name": "Wang Xia", "gender": "Female", "institution": "Zhejiang University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Xueqian zheng", "gender": "unknown", "institution": "Zhejiang University of Technology", "country": "CN", "position": "MS student" }, { "name": "yunchao wang", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 6, "n_ref": 16, "n_ref_all": 31, "n_fig": 12, "n_tab": 1, "L_tab": 338, "n_element_tab": 3, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 338, "n_element_tab_1": 3, "formula_len_all": 104, "formula_len_all_1": 104, "len_all": 84237, "len_all_1": 84237, "len_abs": 1033, "len_title": 158, "len_sents": 25044, "len_sents_1": 25044, "n_sents": 176, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 110, "L_abs": 1078, "title": "Unlocking Compositional Understanding of Vision-Language Models with Visualization Representation and Analysis", "abs": "Vision-language models (VLMs) have made significant advances, debates persist about their ability to understand the combined meaning of vision and linguistic. Existing research primarily relies on computer vision knowledge and static images to deliver findings and insights into compositional understanding of VLMs. There is still a limited understanding of how VLMs handle subtle differences between visual and linguistic information. This paper introduces an interactive visualization representation and analysis approach from outside the computer vision community. In this study, we found that CLIP's performance in compositional understanding only slightly exceeds the chance level of 50%. Particularly, it primarily relies on entities in visual and textual modalities, but is limited in recognizing spatial relationships, attribute ownership, and interaction relationships. Additionally, It behaves more like a bag-of-words model and relies on global feature alignment rather than fine-grained alignment, leading to insensitivity to subtle perturbations in text and images.", "keywords": [ "Vision-Language Models", "Compositional Understanding", "Visualization Representation and Analysis" ], "rating_list": [ 3, 5, 1, 3, 8 ], "soundness_list": [ 1, 3, 2, 2, 4 ], "presentation_list": [ 2, 3, 1, 2, 4 ], "contribution_list": [ 1, 2, 1, 2, 3 ], "confidence_list": [ 5, 2, 5, 4, 4 ] }, { "paper_id": "vo4AHjowKi", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Fei Wu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Huangsen Cao", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yifei Bi", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Yihuai Xu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Intern" }, { "name": "Yongwei Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yu Zhao", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zhouhan Lin", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 20, "n_ref_uni": 60, "n_ref": 121, "n_ref_all": 163, "n_fig": 11, "n_tab": 17, "L_tab": 9673, "n_element_tab": 1524, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1248, "formula_len_all_1": 1258, "len_all": 237064, "len_all_1": 66788, "len_abs": 1737, "len_title": 128, "len_sents": 61662, "len_sents_1": 31532, "n_sents": 484, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1781, "title": "Training-free LLM-generated Text Detection by Mining Token Probability Sequences", "abs": "Large language models (LLMs) have demonstrated remarkable capabilities in generating high-quality texts across diverse domains. However, the potential misuse of LLMs has raised significant concerns, underscoring the urgent need for reliable detection of LLM-generated texts. Conventional training-based detectors often struggle with generalization, particularly in cross-domain and cross-model scenarios. In contrast, training-free methods, which focus on inherent discrepancies through carefully designed statistical features, offer improved generalization and interpretability. Despite this, existing training-free detection methods typically rely on global text sequence statistics, neglecting the modeling of local discriminative features, thereby limiting their detection efficacy. In this work, we introduce a novel training-free detector, termed \\textbf{Lastde}\\footnote{The code and data are released at \\url{https://github.com/TrustMedia-zju/Lastde_Detector}.} that synergizes local and global statistics for enhanced detection. For the first time, we introduce time series analysis to LLM-generated text detection, capturing the temporal dynamics of token probability sequences. By integrating these local statistics with global ones, our detector reveals significant disparities between human and LLM-generated texts. We also propose an efficient alternative, \\textbf{Lastde++} to enable real-time detection. Extensive experiments on six datasets involving cross-domain, cross-model, and cross-lingual detection scenarios, under both white-box and black-box settings, demonstrated that our method consistently achieves state-of-the-art performance. Furthermore, our approach exhibits greater robustness against paraphrasing attacks compared to existing baseline methods.", "keywords": [ "Fake text detection", "training-free", "detection" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "vnp2LtLlQg", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Fatih Porikli", "gender": "Male", "institution": "QualComm", "country": "US", "position": "Senior Director" }, { "name": "Hanno Ackermann", "gender": "unknown", "institution": "Qualcomm Inc, QualComm", "country": "", "position": "Scientific Researcher" }, { "name": "Hong Cai", "gender": "Male", "institution": "Qualcomm AI Research", "country": "US", "position": "Researcher" }, { "name": "Leyla Mirvakhabova", "gender": "Female", "institution": "Qualcomm AI Research", "country": "", "position": "Researcher" }, { "name": "Markus Nagel", "gender": "unknown", "institution": "Qualcomm AI Research", "country": "NL", "position": "Researcher" }, { "name": "Farhad Ghazvinian Zanjani", "gender": "Male", "institution": "Qualcomm AI Research", "country": "NL", "position": "Scientific Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 19, "n_ref": 42, "n_ref_all": 54, "n_fig": 8, "n_tab": 3, "L_tab": 1387, "n_element_tab": 82, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1004, "n_element_tab_1": 76, "formula_len_all": 388, "formula_len_all_1": 441, "len_all": 98508, "len_all_1": 57519, "len_abs": 1158, "len_title": 68, "len_sents": 29286, "len_sents_1": 25898, "n_sents": 259, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 20, "L_abs": 1174, "title": "Optimizing Attention", "abs": "The attention mechanism is an important part of transformer architectures. It en-\nables the network to compare samples within a sequence. Before the comparison\nis performed, tokens are multiplied by trainable matrices. These matrices can\nconstitute a significant part of the total number of parameters. Their size creates\nproblems on systems with limited cache in the compute unit, especially if there\nis limited bandwidth between compute unit and memory. In particular, GPUs on\nmobile devices suffer from this double bottleneck.\nPrior works mitigate this problem for instance by storing low-rank approxima-\ntions, quantization or minimizing the amount of data that needs to be transferred.\nIn this paper, an alternative to the traditional attention mechanism is proposed\nwhich does not require any trainable matrices to perform the attention. The idea\nrests upon solving optimization problems, whereby memory is substituted for\ncompute. It will be shown however, that the computational demand can be re-\nduced such that auto-differentiation becomes possible. An experimental evalua-\ntion shows that the proposed algorithm performs favorable compared with several\nbaselines.", "keywords": [ "transfomers", "attention", "efficiency" ], "rating_list": [ 1, 3, 3, 5 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "vmulbBDCan", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Haiyang Jiang", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Imari Sato", "gender": "Female", "institution": "The University of Tokyo", "country": "JP", "position": "Professor" }, { "name": "Takeharu Nagai", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "Full Professor" }, { "name": "Tetsuichi Wazawa", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "Researcher" }, { "name": "Yinqiang Zheng", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 6, "n_ref_uni": 35, "n_ref": 60, "n_ref_all": 91, "n_fig": 14, "n_tab": 1, "L_tab": 125, "n_element_tab": 32, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1141, "n_element_tab_1": 67, "formula_len_all": 502, "formula_len_all_1": 301, "len_all": 117232, "len_all_1": 54820, "len_abs": 1562, "len_title": 147, "len_sents": 32140, "len_sents_1": 25715, "n_sents": 215, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 99, "L_abs": 1567, "title": "Revolutionizing EMCCD Denoising through a Novel Physics-Based Learning Framework for Noise Modeling", "abs": "Electron-multiplying charge-coupled device (EMCCD) has been instrumental in sensitive observations under low-light situations including astronomy, material science, and biology. \nDespite its ingenious designs to enhance target signals overcoming read-out circuit noises, produced images are not completely noise free, which could still cast a cloud on desired experiment outcomes, especially in fluorescence microscopy.\nExisting studies on EMCCD's noise model have been focusing on statistical characteristics in theory, yet unable to incorporate latest advancements in the field of computational photography, where physics-based noise models are utilized to guide deep learning processes, creating adaptive denoising algorithms for ordinary image sensors.\nStill, those models are not directly applicable to EMCCD.\nIn this paper, we intend to pioneer EMCCD denoising by introducing a systematic study on physics-based noise model calibration procedures for an EMCCD camera, accurately estimating statistical features of observable noise components in experiments, which are then utilized to generate substantial amount of authentic training samples for one of the most recent neural networks.\nA first real-world test image dataset for EMCCD is captured, containing both images of ordinary daily scenes and those of microscopic contents.\nBenchmarking upon the testset and authentic microscopic images, we demonstrate distinct advantages of our model against previous methods for EMCCD and physics-based noise modeling, forging a promising new path for EMCCD denoising.", "keywords": [ "EMCCD", "physics-based noise modeling", "deep high-sensitivity imaging", "fluorescence microscopy image denoising" ], "rating_list": [ 8, 5, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 5, 3, 5 ] }, { "paper_id": "vmkpk0ed1F", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Barproda Halder", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Faisal Hamman", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Ilia Sucholutsky", "gender": "Male", "institution": "New York University", "country": "US", "position": "Faculty Fellow/Assistant Professor" }, { "name": "Pasan Dissanayake", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Qiuyi Zhang", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Sanghamitra Dutta", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 7, "n_ref_uni": 59, "n_ref": 123, "n_ref_all": 173, "n_fig": 22, "n_tab": 8, "L_tab": 2290, "n_element_tab": 297, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1214, "formula_len_all_1": 695, "len_all": 184793, "len_all_1": 69981, "len_abs": 1826, "len_title": 131, "len_sents": 59181, "len_sents_1": 36114, "n_sents": 401, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1828, "title": "Formalizing Spuriousness of Biased Datasets using Partial Information Decomposition", "abs": "Spuriousness arises when there is an association between two or more variables in a dataset that are not causally related. Left unchecked, they can mislead a machine learning model into using the undesirable spurious features in decision-making over the core features, hindering generalization. In this work, we propose a novel explainability framework to disentangle the nature of such spurious associations, i.e., how the information about a target variable is distributed among the spurious and core features. Our framework leverages a body of work in information theory called Partial Information Decomposition (PID) to first decompose the total information about the target into four non-negative quantities namely unique information (in core and spurious features respectively), redundant information, and synergistic information. Next, we leverage this decomposition to propose a novel measure of the spuriousness of a dataset that steers models into choosing the spurious features over the core. We arrive at this measure systematically by examining several candidate measures, and demonstrating what they capture and miss through intuitive canonical examples and counterexamples. Our proposed explainability framework Spurious Disentangler consists of segmentation, dimensionality reduction, and estimation modules, with capabilities to specifically handle high dimensional image data efficiently. Finally, we also conduct empirical evaluation to demonstrate the trends of unique, redundant, and synergistic information, as well as our proposed spuriousness measure across several benchmark datasets under various settings. Interestingly, we observe a novel tradeoff between our measure of dataset spuriousness and empirical model generalization metrics such as worst-group accuracy, further supporting our proposition.", "keywords": [ "Explainability Framework", "Spuriousness", "Partial Information Decomposition", "Blackwell Sufficiency", "Auto-encoder", "Worst-group Accuracy" ], "rating_list": [ 8, 5, 6, 5, 3 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 3, 1, 3, 3, 1 ], "confidence_list": [ 2, 4, 1, 2, 4 ] }, { "paper_id": "vlpEXfbeHn", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Li Shi", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Pingge Hu", "gender": "Not Specified", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Xiaoteng Zhang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xingtong Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yueyang Cang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yuhang Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 21, "n_formula_1": 20, "n_ref_uni": 18, "n_ref": 38, "n_ref_all": 44, "n_fig": 3, "n_tab": 2, "L_tab": 2001, "n_element_tab": 211, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2140, "n_element_tab_1": 219, "formula_len_all": 994, "formula_len_all_1": 970, "len_all": 77260, "len_all_1": 50315, "len_abs": 1075, "len_title": 122, "len_sents": 22763, "len_sents_1": 20886, "n_sents": 176, "n_sents_1": 160, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1097, "title": "RetCompletion:High-Speed Inference Image Completion with Retentive Network", "abs": "Time cost is a major challenge in achieving high-quality pluralistic image completion. Recently, the Retentive Network (RetNet) in natural language processing offers a novel approach to this problem with its low-cost inference capabilities. Inspired by this, we apply RetNet to the pluralistic image completion task in computer vision. We present RetCompletion, a two-stage framework. In the first stage, we introduce Bi-RetNet, a bidirectional sequence information fusion model that integrates contextual information from images. During inference, we employ a unidirectional pixel-wise update strategy to restore consistent image structures, achieving both high reconstruction quality and fast inference speed. In the second stage, we use a CNN for low-resolution upsampling to enhance texture details. Experiments on ImageNet and CelebA-HQ demonstrate that our inference speed is 10$\\times$ faster than ICT and 15$\\times$ faster than RePaint. The proposed RetCompletion significantly improves inference speed and delivers strong performance, especially when masks cover large areas of the image.", "keywords": [ "Pluralistic image completion", "Retentive Network" ], "rating_list": [ 3, 3, 3, 6, 3 ], "soundness_list": [ 2, 1, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 3, 1 ] }, { "paper_id": "vlg5WRKHxh", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Jin Song Dong", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Kan Jiang", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Lecturer" }, { "name": "Murong Ma", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yun Lin", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhaoyu Liu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Zhe Hou", "gender": "Not Specified", "institution": "Griffith University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 6, "n_ref_uni": 37, "n_ref": 64, "n_ref_all": 77, "n_fig": 3, "n_tab": 8, "L_tab": 3081, "n_element_tab": 518, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1734, "n_element_tab_1": 268, "formula_len_all": 183, "formula_len_all_1": 314, "len_all": 156751, "len_all_1": 73418, "len_abs": 1074, "len_title": 126, "len_sents": 36807, "len_sents_1": 30992, "n_sents": 291, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 79, "L_abs": 1091, "title": "$F^3Set$: Towards Analyzing Fast, Frequent, and Fine-grained Events from Videos", "abs": "Analyzing Fast, Frequent, and Fine-grained ($F^3$) events presents a significant challenge in video analytics and multi-modal LLMs. Current methods struggle to identify events that satisfy all the $F^3$ criteria with high accuracy due to challenges such as motion blur and subtle visual discrepancies. To advance research in video understanding, we introduce $F^3Set$, a benchmark that consists of video datasets for precise $F^3$ event detection. Datasets in $F^3Set$ are characterized by their extensive scale and comprehensive detail, usually encompassing over 1,000 event types with precise timestamps and supporting multi-level granularity. Currently, $F^3Set$ contains several sports datasets, and this framework may be extended to other applications as well. We evaluated popular temporal action understanding methods on $F^3Set$, revealing substantial challenges for existing techniques. Additionally, we propose a new method, $F^3ED$, for $F^3$ event detections, achieving superior performance. The dataset, model, and benchmark code are available at https://github.com/F3Set/F3Set.", "keywords": [ "temporal event spotting", "fine-grained video understanding", "video analytics" ], "rating_list": [ 8, 5, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "vlOfFI9vWO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Lu ChengLong", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Wei Wang", "gender": "Male", "institution": "Fudan University", "country": "", "position": "Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 10, "n_ref": 12, "n_ref_all": 23, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 885, "n_element_tab_1": 86, "formula_len_all": 503, "formula_len_all_1": 503, "len_all": 74925, "len_all_1": 50967, "len_abs": 1515, "len_title": 146, "len_sents": 24695, "len_sents_1": 22919, "n_sents": 176, "n_sents_1": 164, "venue": "ICLR2025", "decision": "withdraw", "L_title": 96, "L_abs": 1469, "title": "Multi-Agent Reinforcement Learning for Efficient Vision Transformer with Dynamic Token Selection", "abs": "Vision Transformers (ViT) have revolutionized the field of computer vision by\nleveraging self-attention mechanisms to process images. However, the computational\ncost of ViT increases quadratically with the number of tokens. Dynamic token selection methods which aims to reduce computational cost by discard redundant tokens during inference, are primarily based on non-differentiable binary decisions methods and relaxations methods. However, Reinforcement Learning( (RL) based methods, which have astonishing decision-making ability, is considered to have high variance and high bias, not adopted for dynamic token selection task in previous work. Yet, RL-based methods have been successfully applied to many binary decision problems such as neural pruning, routing, path selection. In this paper, we propose Reinforcement Learning for Dynamic Vision Transformer (RL4DViT), a novel framework for the dynamic token selection task in ViT using RL. By harnessing the powerfull decision-making capabilities of Multi-Agent Reinforcement Learning(MARL) algorithms, our method dynamically prunes redundant tokens based on input complexity, significantly\nreducing the computational cost while maintaining high accuracy. Extensive experiments\non the ImageNet dataset indicate that our approach reduces the computational cost by\nup to 39%, with only a 0.17% decrease in accuracy. To the best of our knowledge,\nthis is the first RL-based token selection method for efficient ViT.", "keywords": [ "efficient vision transformer", "dynamic token selection", "mappo" ], "rating_list": [ 5, 3, 3, 1 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 3, 3, 1, 1 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 4, 5, 4, 2 ] }, { "paper_id": "vl8VpW2niQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Eduardo Blanco", "gender": "unknown", "institution": "University of Arizona", "country": "", "position": "Associate Professor" }, { "name": "Ellen Riloff", "gender": "Female", "institution": "University of Arizona", "country": "", "position": "Full Professor" }, { "name": "Mihai Surdeanu", "gender": "unknown", "institution": "University of Arizona", "country": "", "position": "Full Professor" }, { "name": "Shahriar Golchin", "gender": "Male", "institution": "Google", "country": "US", "position": "Intern" }, { "name": "Steven Bethard", "gender": "Male", "institution": "University of Arizona", "country": "US", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 47, "n_ref": 115, "n_ref_all": 157, "n_fig": 5, "n_tab": 5, "L_tab": 1550, "n_element_tab": 51, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 424, "n_element_tab_1": 37, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 218678, "len_all_1": 71288, "len_abs": 1156, "len_title": 83, "len_sents": 52637, "len_sents_1": 35673, "n_sents": 419, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1158, "title": "Memorization in In-Context Learning", "abs": "In-context learning (ICL) has proven to be an effective strategy for improving the performance of large language models (LLMs) with no additional training. However, the exact mechanism behind this performance improvement remains unclear. This study is the first to show how ICL surfaces memorized training data and to explore the correlation between this memorization and performance on downstream tasks across various ICL regimes: zero-shot, few-shot, and many-shot. Our most notable findings include: (1) ICL significantly surfaces memorization compared to zero-shot learning in most cases; (2) demonstrations, without their labels, are the most effective element in surfacing memorization; (3) ICL improves performance when the surfaced memorization in few-shot regimes reaches a high level (about 40%); and (4) there is a very strong correlation between performance and memorization in ICL when it outperforms zero-shot learning. Overall, our study uncovers memorization as a new factor impacting ICL, raising an important question: to what extent do LLMs truly generalize from demonstrations in ICL, and how much of their success is due to memorization?", "keywords": [ "Memorization", "In-Context Learning", "Large Language Models" ], "rating_list": [ 6, 5, 8, 5, 3 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3, 2 ], "contribution_list": [ 2, 2, 4, 2, 2 ], "confidence_list": [ 3, 3, 4, 3, 3 ] }, { "paper_id": "vl7kf0YHwj", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chun Yuan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jinwei Fang", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Ke Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Quan Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xi Lin", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xi Tang", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yuxin Qi", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 17, "n_ref_uni": 33, "n_ref": 57, "n_ref_all": 81, "n_fig": 5, "n_tab": 8, "L_tab": 5417, "n_element_tab": 520, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2876, "n_element_tab_1": 345, "formula_len_all": 1331, "formula_len_all_1": 1028, "len_all": 159490, "len_all_1": 60374, "len_abs": 1206, "len_title": 146, "len_sents": 32918, "len_sents_1": 23265, "n_sents": 246, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1211, "title": "IMDPrompter: Adapting SAM to Image Manipulation Detection by Cross-View Automated Prompt Learning", "abs": "Using extensive training data from SA-1B, the Segment Anything Model (SAM) has demonstrated exceptional generalization and zero-shot capabilities, attracting widespread attention in areas such as medical image segmentation and remote sensing image segmentation. However, its performance in the field of image manipulation detection remains largely unexplored and unconfirmed. There are two main challenges in applying SAM to image manipulation detection: a) reliance on manual prompts, and b) the difficulty of single-view information in supporting cross-dataset generalization. To address these challenges, we develops a cross-view prompt learning paradigm called IMDPrompter based on SAM. Benefiting from the design of automated prompts, IMDPrompter no longer relies on manual guidance, enabling automated detection and localization. Additionally, we propose components such as Cross-view Feature Perception, Optimal Prompt Selection, and Cross-View Prompt Consistency, which facilitate cross-view perceptual learning and guide SAM to generate accurate masks. Extensive experimental results from five datasets (CASIA, Columbia, Coverage, IMD2020, and NIST16) validate the effectiveness of our proposed method.", "keywords": [ "Image Manipulation Detection\uff1bSegment Anything Model\uff1bPrompt learning\uff1bSemantic-Agnostic" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "vkj5ARRCeY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Jaegul Choo", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Jaeseong Yun", "gender": "unknown", "institution": "Naver Labs", "country": "KR", "position": "Researcher" }, { "name": "Joohyun Park", "gender": "Male", "institution": "Naver Labs", "country": "KR", "position": "Researcher" }, { "name": "Junha Hyung", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Min-Jung Kim", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Soohyun Ryu", "gender": "Male", "institution": "Naver Labs", "country": "KR", "position": "Researcher" }, { "name": "Sungwon Hwang", "gender": "Male", "institution": "Graduate School of Artificial Intelligence, Korea Advanced Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Suyong Yeon", "gender": "Male", "institution": "NAVER LABS", "country": "KR", "position": "Researcher" }, { "name": "Taewoong Kang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Jesse Jaeseong Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" } ], "n_formula": 22, "n_formula_1": 20, "n_ref_uni": 35, "n_ref": 71, "n_ref_all": 106, "n_fig": 10, "n_tab": 6, "L_tab": 1879, "n_element_tab": 210, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 4161, "n_element_tab_1": 169, "formula_len_all": 1321, "formula_len_all_1": 1101, "len_all": 116638, "len_all_1": 58885, "len_abs": 247, "len_title": 140, "len_sents": 33341, "len_sents_1": 23046, "n_sents": 248, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 909, "title": "Injecting Inductive Bias to 3D Gaussian Splatting for Geometrically Accurate Radiance Fields", "abs": "3D Gaussian Splatting (3DGS) has significantly advanced high-fidelity, real-time novel view synthesis. However, its discrete nature limits the accurate reconstruction of geometry. To address this issue, recent methods have introduced rendering and regularization of depth and normal maps from 3D Gaussians, leading to plausible results. In this paper, we argue that computing normals from independently trainable Gaussian covariances contradicts the strict definition of normals, which should instead be derived from the distribution of neighboring densities. To address this, we introduce an inductive bias into 3DGS by explicitly parameterizing covariances of Gaussians using principal axes and variances of distribution computed from neighboring Gaussians. These axes and variances are then regularized to ensure local surface smoothness. Our approach achieves competitive performance on multiple datasets.", "keywords": [ "3D Gaussian Splatting", "Surface Reconstruction" ], "rating_list": [ 3, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "vkakKdznFS", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Chaofeng Chen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Jiaxing Xu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Litong Feng", "gender": "Male", "institution": "SenseTime Research", "country": "HK", "position": "Associate Research Director" }, { "name": "Mengcheng Lan", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xinjiang Wang", "gender": "Male", "institution": "SenseTime Group", "country": "", "position": "Researcher" }, { "name": "Yiping Ke", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Yue Zhou", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Wei Zhang", "gender": "Male", "institution": "SenseTime Research", "country": "CN", "position": "Research Director" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 49, "n_ref": 101, "n_ref_all": 124, "n_fig": 18, "n_tab": 14, "L_tab": 10422, "n_element_tab": 938, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 2507, "n_element_tab_1": 110, "formula_len_all": 21, "formula_len_all_1": 28, "len_all": 196857, "len_all_1": 50721, "len_abs": 1468, "len_title": 107, "len_sents": 44370, "len_sents_1": 22068, "n_sents": 401, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1439, "title": "Text4Seg: Reimagining Image Segmentation as Text Generation", "abs": "Multimodal Large Language Models (MLLMs) have shown exceptional capabilities in vision-language tasks; however, effectively integrating image segmentation into these models remains a significant challenge. In this paper, we introduce Text4Seg, a novel text-as-mask paradigm that casts image segmentation as a text generation problem, eliminating the need for additional decoders and significantly simplifying the segmentation process. Our key innovation is semantic descriptors, a new textual representation of segmentation masks where each image patch is mapped to its corresponding text label. This unified representation allows seamless integration into the auto-regressive training pipeline of MLLMs for easier optimization. We demonstrate that representing an image with $16\\times16$ semantic descriptors yields competitive segmentation performance. To enhance efficiency, we introduce the Row-wise Run-Length Encoding (R-RLE), which compresses redundant text sequences, reducing the length of semantic descriptors by 74\\% and accelerating inference by $3\\times$, without compromising performance. Extensive experiments across various vision tasks, such as referring expression segmentation and comprehension, show that Text4Seg achieves state-of-the-art performance on multiple datasets by fine-tuning different MLLM backbones. Our approach provides an efficient, scalable solution for vision-centric tasks within the MLLM framework.", "keywords": [ "Multimodal large language model", "Image segmentation", "Referring expression segmentation" ], "rating_list": [ 8, 5, 6 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "vkOaerjEcz", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Chaoqi Chen", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Kunze Huang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Luyao Tang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Xiaotong Tu", "gender": "unknown", "institution": "Xiamen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xinghao Ding", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Yue Huang", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 51, "n_ref": 80, "n_ref_all": 91, "n_fig": 4, "n_tab": 3, "L_tab": 2917, "n_element_tab": 200, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 508, "formula_len_all_1": 508, "len_all": 125745, "len_all_1": 56468, "len_abs": 1582, "len_title": 120, "len_sents": 33953, "len_sents_1": 27433, "n_sents": 235, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1424, "title": "MTMC: Generalized Category Discovery via Maximum Token Manifold Capacity", "abs": "Identifying previously unseen data is crucial for enhancing the robustness of deep learning models in the open world. Generalized category discovery (GCD) is a representative problem that requires clustering unlabeled data that includes known and novel categories. Current GCD methods mostly focus on minimizing intra-cluster variations, often at the cost of manifold capacity, thus limiting the richness of within-class representations. In this paper, we introduce a novel GCD approach that emphasizes maximizing the token manifold capacity (MTMC) within class tokens, thereby preserving the diversity and complexity of the data's intrinsic structure. Specifically, MTMC's efficacy is fundamentally rooted in its ability to leverage the nuclear norm of the singular values as a quantitative measure of the manifold capacity. MTMC enforces a richer and more informative representation within the manifolds of different patches constituting the same sample. MTMC ensures that, for each cluster, the representations of different patches of the same sample are compact and lie in a low-dimensional space, thereby enhancing discriminability. By doing so, the model could capture each class's nuanced semantic details and prevent the loss of critical information during the clustering process. MTMC promotes a comprehensive, non-collapsed representation that improves inter-class separability without adding excessive complexity.", "keywords": [ "generalized category discovery", "deep cluster", "manifold capacity" ], "rating_list": [ 5, 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 3, 4, 2, 3, 3 ], "contribution_list": [ 2, 4, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "vkOFOUDLTn", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Hanting Chen", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xiangzhong Fang", "gender": "Male", "institution": "School of mathematical Science, Peking University, Peking University", "country": "", "position": "Full Professor" }, { "name": "Yuchen Liang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yunhe Wang", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 25, "n_formula_1": 22, "n_ref_uni": 47, "n_ref": 137, "n_ref_all": 146, "n_fig": 3, "n_tab": 5, "L_tab": 2032, "n_element_tab": 274, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 436, "n_element_tab_1": 16, "formula_len_all": 1816, "formula_len_all_1": 1634, "len_all": 142315, "len_all_1": 67946, "len_abs": 1586, "len_title": 122, "len_sents": 32855, "len_sents_1": 27675, "n_sents": 274, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1706, "title": "Linear Multistep Solver Distillation for Fast Sampling of Diffusion Models", "abs": "Sampling from diffusion models can be seen as solving the corresponding \n probability flow ordinary differential equation (ODE). \n The solving process requires a significant number of function \n evaluations (NFE), making it time-consuming. \n Recently, several solver search frameworks have attempted to find \n better-performing model-specific solvers. However, predicting the impact of \n intermediate solving strategies on final sample quality remains challenging, \n rendering the search process inefficient.\n In this paper, we propose a novel method for designing \n solving strategies. We first introduce a unified prediction formula \n for linear multistep solvers. Subsequently, we present a solver distillation \n framework, which enables a student solver to mimic the sampling trajectory \n generated by a teacher solver with more steps. We utilize the mean Euclidean \n distance between the student and teacher sampling trajectories as a metric, \n facilitating rapid adjustment and optimization of intermediate solving strategies.\n The design space of our framework encompasses multiple aspects, \n including prediction coefficients, time step schedules, and time scaling \n factors. \n Our framework has the ability to complete a solver search \n for Stable-Diffusion in under 12 total GPU hours.\n Compared to previous reinforcement learning-based \n search frameworks, \n our approach achieves over a 10$\\times$ increase in search efficiency. \n With just 5 NFE, we achieve FID scores of 3.23 on CIFAR10, 7.16 on ImageNet-64, \n 5.44 on LSUN-Bedroom, and 12.52 on MS-COCO, resulting in a 2$\\times$ sampling acceleration ratio \n compared to handcrafted solvers.", "keywords": [ "Diffusion Probabilistic Model", "Diffusion Sampler", "Solver Schedule" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "vjel3nWP2a", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "A. Feder Cooper", "gender": "unknown", "institution": "Research, Microsoft", "country": "US", "position": "Postdoc" }, { "name": "Christopher A. Choquette-Choo", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Daphne Ippolito", "gender": "Female", "institution": "School of Engineering and Applied Science, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Javier Rando", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Jonathan Hayase", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Katherine Lee", "gender": "Female", "institution": "Cornell University", "country": "", "position": "PhD student" }, { "name": "Matthew Jagielski", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Milad Nasr", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Nicholas Carlini", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Florian Tramer", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 44, "n_ref": 132, "n_ref_all": 176, "n_fig": 18, "n_tab": 17, "L_tab": 2391, "n_element_tab": 233, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 20, "n_element_tab_1": 1, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 551370, "len_all_1": 57492, "len_abs": 657, "len_title": 125, "len_sents": 256434, "len_sents_1": 28565, "n_sents": 1871, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 631, "title": "Scalable Extraction of Training Data from Aligned, Production Language Models", "abs": "We show that *alignment*---a standard process that tunes LLMs to follow instructions in a harmless manner---seems to prevent existing data extraction attacks. We develop two novel attacks that undo a model's alignment and recover thousands of training examples from the popular proprietary model, OpenAI's ChatGPT. Our most potent attack causes ChatGPT to emit training data in over 23% of conversations, and enables targeted reconstruction of chosen training documents, including those containing copyrighted or harmful content. Our work highlights the limitations of existing safeguards to prevent training-data leakage in LLMs.", "keywords": [ "privacy", "language models", "data extraction", "security" ], "rating_list": [ 6, 6, 8, 6, 8 ], "soundness_list": [ 3, 4, 4, 3, 3 ], "presentation_list": [ 2, 2, 4, 2, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4, 2 ] }, { "paper_id": "vjbIer5R2H", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Bowei Zhu", "gender": "unknown", "institution": "Renmin University of China", "country": "", "position": "PhD student" }, { "name": "Huayi Tang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Shaojie Li", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 174, "n_formula_1": 21, "n_ref_uni": 45, "n_ref": 122, "n_ref_all": 152, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 75, "n_element_tab_1": 6, "formula_len_all": 20777, "formula_len_all_1": 2056, "len_all": 194909, "len_all_1": 63050, "len_abs": 1100, "len_title": 116, "len_sents": 52670, "len_sents_1": 27559, "n_sents": 577, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1103, "title": "Improved Risk Bounds with Unbounded Losses for Transductive Learning", "abs": "In the transductive learning setting, we are provided with a labeled training set and an unlabeled test set, with the objective of predicting the labels of the test points. This framework differs from the standard problem of fitting an unknown distribution with a training set drawn independently from this distribution. In this paper, we primarily improve the generalization bounds in transductive learning. Specifically, we develop two novel concentration inequalities for the suprema of empirical processes sampled without replacement for unbounded functions, marking the first discussion of the generalization performance of unbounded functions in the context of sampling without replacement. We further provide two valuable applications of our new inequalities: on one hand, we firstly derive fast excess risk bounds for empirical risk minimization in transductive learning under unbounded losses. On the other hand, we establish high-probability bounds on the generalization error for graph neural networks when using stochastic gradient descent which improve the current state-of-the-art results.", "keywords": [ "concentration inequality", "generalization bounds", "graph neural networks", "transductive learning", "unbounded losses" ], "rating_list": [ 1, 3, 8, 1 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 5, 3, 3, 5 ] }, { "paper_id": "vjHySpxDsv", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Chenyu Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Hanbo Cheng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Jia Pan", "gender": "Male", "institution": "IFLYTEK CO.LTD.", "country": "CN", "position": "Principal Researcher" }, { "name": "Jiefeng Ma", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Jun Du", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Limin Lin", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" }, { "name": "Pengcheng Xia", "gender": "Male", "institution": "IFLYTEK CO.LTD.", "country": "CN", "position": "Researcher" }, { "name": "Pengfei Hu", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 45, "n_ref": 94, "n_ref_all": 115, "n_fig": 6, "n_tab": 4, "L_tab": 1379, "n_element_tab": 242, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 942, "n_element_tab_1": 152, "formula_len_all": 392, "formula_len_all_1": 317, "len_all": 153885, "len_all_1": 65715, "len_abs": 1339, "len_title": 153, "len_sents": 42229, "len_sents_1": 30862, "n_sents": 316, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1383, "title": "DAWN: Dynamic Frame Avatar with Non-autoregressive Diffusion Framework for Talking head Video Generation", "abs": "Talking head generation intends to produce vivid and realistic talking head videos from a single portrait and speech audio clip. Although significant progress has been made in diffusion-based talking head generation, almost all methods rely on autoregressive strategies, which suffer from limited context utilization beyond the current generation step, error accumulation, and slower generation speed. To address these challenges, we present DAWN (\\textbf{D}ynamic frame \\textbf{A}vatar \\textbf{W}ith \\textbf{N}on-autoregressive diffusion), a framework that enables all-at-once generation of dynamic-length video sequences. Specifically, it consists of two main components: (1) audio-driven holistic facial dynamics generation in the latent motion space, and (2) audio-driven head pose and blink generation. Extensive experiments demonstrate that our method generates authentic and vivid videos with precise lip motions, and natural pose/blink movements. Additionally, with a high generation speed, DAWN possesses strong extrapolation capabilities, ensuring the stable production of high-quality long videos. These results highlight the considerable promise and potential impact of DAWN in the field of talking head video generation. Furthermore, we hope that DAWN sparks further exploration of non-autoregressive approaches in diffusion models. Our code will be publicly available.", "keywords": [ "Talking head generation", "Non-autoregressive generation", "Avatar", "Video generation", "Diffusion model" ], "rating_list": [ 5, 6, 5, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "vikwIayXOx", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Alex Kot", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Hans Vandierendonck", "gender": "Male", "institution": "The Queen's University Belfast", "country": "GB", "position": "Full Professor" }, { "name": "Ira Assent", "gender": "Female", "institution": "Aarhus University", "country": "DK", "position": "Full Professor" }, { "name": "Ngai-man Cheung", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Associate Professor" }, { "name": "Ngoc-Bao Nguyen", "gender": "Female", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Postdoc" }, { "name": "Son T. Mai", "gender": "unknown", "institution": "The Queen's University Belfast", "country": "GB", "position": "Associate Professor" }, { "name": "Tran Viet Hung", "gender": "Male", "institution": "The Queen's University Belfast", "country": "GB", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 39, "n_ref": 190, "n_ref_all": 197, "n_fig": 4, "n_tab": 16, "L_tab": 9023, "n_element_tab": 1004, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2981, "n_element_tab_1": 122, "formula_len_all": 167, "formula_len_all_1": 57, "len_all": 195376, "len_all_1": 62067, "len_abs": 1736, "len_title": 121, "len_sents": 60040, "len_sents_1": 27892, "n_sents": 471, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1751, "title": "Random Erasing vs. Model Inversion: A Promising Defense or a False Hope?", "abs": "Model Inversion (MI) attacks pose a significant privacy threat by reconstructing private training data from machine learning models. \nWhile existing defenses primarily concentrate on model-centric approaches, the impact of data on MI robustness remains largely unexplored. In this work, we explore Random Erasing (RE), a technique traditionally used to enhance model generalization under occlusion. Surprisingly, our study reveals that RE emerges as a powerful defense against MI attacks. We conduct analysis to identify crucial properties of RE to serve as an effective defense. Particularly, Partial Erasure in RE prevents the model from observing the entire objects during training, and we find that this has significant impact on MI, which aims to reconstruct the entire objects. Meanwhile, our analysis suggests Random Location in RE is important for outstanding privacy-utility trade-off. Furthermore, our analysis reveals that model trained with RE leads to a discrepancy between the features of MI-reconstructed images and that of private images. These effects significantly degrade MI reconstruction quality and attack accuracy while maintaining reasonable natural accuracy. Our RE-based defense method is simple to implement and can be combined with other defenses. Extensive experiments of 34 setups demonstrate that our method achieve SOTA performance in privacy-utility tradeoff. The results consistently demonstrate the superiority of our defense over existing defenses across different MI attacks, network architectures, and attack configurations. For the first time, we achieve significant degrade in attack accuracy without decrease in utility for some configurations. Our code and additional results are included in Supplementary.", "keywords": [ "Privacy", "Model Inversion", "Random Erasing" ], "rating_list": [ 5, 6, 3, 6, 5 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 1, 3, 2 ], "contribution_list": [ 2, 2, 1, 3, 3 ], "confidence_list": [ 4, 3, 4, 3, 5 ] }, { "paper_id": "viQ1bLqKY0", "primary_area": "datasets and benchmarks", "n_author": 1, "author_info_list": [ { "name": "Rob Kopel", "gender": "unknown", "institution": "PricewaterhouseCoopers", "country": "AU", "position": "Director" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 18, "n_ref": 48, "n_ref_all": 65, "n_fig": 6, "n_tab": 6, "L_tab": 10702, "n_element_tab": 264, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 163, "n_element_tab_1": 20, "formula_len_all": 60, "formula_len_all_1": 22, "len_all": 191765, "len_all_1": 66062, "len_abs": 1884, "len_title": 61, "len_sents": 65703, "len_sents_1": 28733, "n_sents": 459, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1917, "title": "EXecution-Eval: Can language models execute real-world code?", "abs": "As Large Language Models (LLMs) advance, traditional benchmarks face challenges of dataset saturation and disconnection from real-world performance, limiting our understanding of true model capabilities. We introduce EXecution-Eval (EXE), a benchmark designed to assess LLMs' ability to execute code and predict program states. EXE attempts to address key limitations in existing evaluations: difficulty scaling, task diversity, training data contamination, and cost-effective scalability.\nComprising over 30,000 tasks derived from 1,000 popular Python repositories on GitHub, EXE spans a range of context lengths and algorithmic complexities. Tasks require models to execute code, necessitating various operations including mathematical reasoning, logical inference, bit manipulation, string operations, loop execution, and maintaining multiple internal variable states during computation. Our methodology involves: (a) selecting and preprocessing GitHub repositories, (b) generating diverse inputs for functions, (c) executing code to obtain ground truth outputs, and (d) formulating tasks that require models to reason about code execution. This approach allows for continuous new task generation for as few as 1,200 tokens, significantly reducing the risk of models \"training on the test set.\"\nWe evaluate several state-of-the-art LLMs on EXE, revealing insights into their code comprehension and execution capabilities. Our results show that even the best-performing models struggle with complex, multi-step execution tasks, highlighting specific computational concepts that pose the greatest challenges for today's LLMs. Furthermore, we review EXE's potential for finding and predicting errors to aid in assessing a model's cybersecurity capabilities. We propose EXE as a sustainable and challenging testbed for evaluating frontier models, offering potential insights into their internal mechanistic advancement", "keywords": [ "large language model", "evaluation", "benchmark", "code execution" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "vi3DjUhFVm", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Dongmin Park", "gender": "Male", "institution": "KRAFTON", "country": "KR", "position": "Researcher" }, { "name": "Minkyu Kim", "gender": "Male", "institution": "KRAFTON, Inc.", "country": "KR", "position": "Researcher" }, { "name": "Sunwoo Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Undergrad student" } ], "n_formula": 89, "n_formula_1": 18, "n_ref_uni": 54, "n_ref": 139, "n_ref_all": 166, "n_fig": 13, "n_tab": 2, "L_tab": 378, "n_element_tab": 54, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 173, "n_element_tab_1": 9, "formula_len_all": 8182, "formula_len_all_1": 1202, "len_all": 223265, "len_all_1": 51291, "len_abs": 953, "len_title": 128, "len_sents": 63396, "len_sents_1": 23158, "n_sents": 523, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 72, "L_abs": 1014, "title": "Test-time Alignment of Diffusion Models without Reward Over-optimization", "abs": "Diffusion models excel in generative tasks, but aligning them with specific objectives while maintaining their versatility remains challenging. Existing fine-tuning methods often suffer from reward over-optimization, while approximate guidance approaches fail to optimize target rewards effectively. Addressing these limitations, we propose a training-free, test-time method based on Sequential Monte Carlo (SMC) to sample from the reward-aligned target distribution. Our approach, tailored for diffusion sampling and incorporating tempering techniques, achieves comparable or superior target rewards to fine-tuning methods while preserving diversity and cross-reward generalization. We demonstrate its effectiveness in single-reward optimization, multi-objective scenarios, and online black-box optimization. This work offers a robust solution for aligning diffusion models with diverse downstream objectives without compromising their general capabilities. Code is available at https://github.com/krafton-ai/DAS.", "keywords": [ "diffusion models", "alignment", "reward over-optimization", "sequential monte carlo samplers" ], "rating_list": [ 8, 8, 8, 5 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "vhazhSm6I0", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Ke-Jia CHEN", "gender": "Female", "institution": "Nanjing University of Posts and Telecommunications", "country": "", "position": "Associate Professor" }, { "name": "Weifeng Zhuang", "gender": "Male", "institution": "Nanjing University of Posts and Telecommunications", "country": "", "position": "MS student" }, { "name": "Wenbin Li", "gender": "Male", "institution": "Nanjing University of Posts and Telecommunications", "country": "", "position": "Researcher" }, { "name": "Zheng Liu", "gender": "Male", "institution": "Nanjing University of Posts and Telecommunications", "country": "CN", "position": "Researcher" } ], "n_formula": 33, "n_formula_1": 29, "n_ref_uni": 37, "n_ref": 74, "n_ref_all": 83, "n_fig": 6, "n_tab": 5, "L_tab": 3940, "n_element_tab": 248, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 5036, "n_element_tab_1": 290, "formula_len_all": 1650, "formula_len_all_1": 1554, "len_all": 138340, "len_all_1": 70664, "len_abs": 1142, "len_title": 149, "len_sents": 39918, "len_sents_1": 27610, "n_sents": 321, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1378, "title": "Optimizing Activations Beyond Entropy Minimization for Test-Time Adaptation of Graph Neural Networks", "abs": "Test-time adaptation for classification models involves optimizing classifiers through self-supervised learning without labeled training samples. Existing methods often rely on entropy minimization as the optimization objective, which\u00a0indeed\u00a0addresses the model performance connections with prediction confidence or representations amenable to cluster structure.\u00a0However, due to the lack of ground truth in training samples, test-time adaptation, as an effective way to deal with the shifting dataset distributions or domains, can sometimes lead to model collapse. In this paper, we focus on optimizing activations in batch normalization (BN) layers for test-time adaptation of graph neural networks (GNNs). Unlike many entropy minimization methods prone to catastrophic model collapse, our approach leverages pseudo-labels of test samples to mitigate the potential forgetting of training data. \nWe optimize activations in BN by a two-step process.\u00a0First, we identify weights and masks for the empirical batch mean and variance\u00a0of both training and test samples.\u00a0Subsequently, we refine\u00a0BN's\u00a0scale and shift parameters using a reformulated loss function with an energy-based model for improved generalization.\u00a0Empirical evaluation across seven challenging datasets demonstrates the superior performance of our method compared to state-of-the-art test-time adaptation approaches.", "keywords": [ "test-time adaptation", "batch normalization", "graph neural network", "energy-based model" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 1, 4, 3, 3 ] }, { "paper_id": "vhPE3PtTgC", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Amaru Cuba Gyllensten", "gender": "Male", "institution": "AI Sweden", "country": "", "position": "Researcher" }, { "name": "Ariel Ekgren", "gender": "unknown", "institution": "KTH Royal Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Danila Petrelli", "gender": "Female", "institution": "AI Sweden", "country": "SE", "position": "Researcher" }, { "name": "Magnus Sahlgren", "gender": "Male", "institution": "AI Sweden", "country": "SE", "position": "Principal Researcher" }, { "name": "Paul Gabriel dos Santos", "gender": "unknown", "institution": "AI Sweden", "country": "SE", "position": "Researcher" }, { "name": "Tim Isbister", "gender": "Male", "institution": "Ai Sweden", "country": "", "position": "ML Engineer" }, { "name": "Tobias Norlund", "gender": "unknown", "institution": "AI Sweden", "country": "SE", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 19, "n_ref": 37, "n_ref_all": 58, "n_fig": 12, "n_tab": 4, "L_tab": 4668, "n_element_tab": 194, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 119, "formula_len_all_1": 119, "len_all": 131525, "len_all_1": 58736, "len_abs": 596, "len_title": 104, "len_sents": 35009, "len_sents_1": 25575, "n_sents": 316, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 56, "L_abs": 597, "title": "SWEb: A Large Web Dataset for the Scandinavian Languages", "abs": "This paper presents the hitherto largest pretraining dataset for the Scandinavian languages: the Scandinavian WEb (SWEb), comprising over one trillion tokens. The paper details the collection and processing pipeline, and introduces a novel model-based text extractor that significantly reduces complexity in comparison with rule-based approaches. We also introduce a new cloze-style benchmark for evaluating language models in Swedish, and use this test to compare models trained on the SWEb data to models trained on FineWeb, with competitive results. All data, models and code are shared openly.", "keywords": [ "dataset", "pre-training", "swedish", "danish", "norwegian", "icelandic" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "vh1e2WJfZp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Bo Li", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Hao Zhang", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "Huchuan Lu", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Professor" }, { "name": "Jinwei Chen", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Lihe Zhang", "gender": "unknown", "institution": "Dalian University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Peng-Tao Jiang", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "qian yu", "gender": "Female", "institution": "Dalian University of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 45, "n_ref": 101, "n_ref_all": 114, "n_fig": 6, "n_tab": 6, "L_tab": 2938, "n_element_tab": 339, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 3453, "n_element_tab_1": 342, "formula_len_all": 691, "formula_len_all_1": 691, "len_all": 122481, "len_all_1": 62521, "len_abs": 1736, "len_title": 125, "len_sents": 31971, "len_sents_1": 26624, "n_sents": 188, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1741, "title": "High-Precision Dichotomous Image Segmentation via Probing Diffusion Capacity", "abs": "In the realm of high-resolution (HR), fine-grained image segmentation, the primary challenge is balancing broad contextual awareness with the precision required for detailed object delineation, capturing intricate details and the finest edges of objects. Diffusion models, trained on vast datasets comprising billions of image-text pairs, such as SD V2.1, have revolutionized text-to-image synthesis by delivering exceptional quality, fine detail resolution, and strong contextual awareness, making them an attractive solution for high-resolution image segmentation. To this end, we propose DiffDIS, a diffusion-driven segmentation model that taps into the potential of the pre-trained U-Net within diffusion models, specifically designed for high-resolution, fine-grained object segmentation. By leveraging the robust generalization capabilities and rich, versatile image representation prior of the SD models, coupled with a task-specific stable one-step denoising approach, we significantly reduce the inference time while preserving high-fidelity, detailed generation. Additionally, we introduce an auxiliary edge generation task to not only enhance the preservation of fine details of the object boundaries, but reconcile the probabilistic nature of diffusion with the deterministic demands of segmentation. With these refined strategies in place, DiffDIS serves as a rapid object mask generation model, specifically optimized for generating detailed binary maps at high resolutions, while demonstrating impressive accuracy and swift processing. Experiments on the DIS5K dataset demonstrate the superiority of DiffDIS, achieving state-of-the-art results through a streamlined inference process. Our code will be made publicly available.", "keywords": [ "dichotomous image segmentation", "diffusion models", "high-resolution image segmentation" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "vgvnfUho7X", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Evimaria Terzi", "gender": "unknown", "institution": "Boston University, Boston University", "country": "", "position": "Full Professor" }, { "name": "Gabriel Franco", "gender": "Male", "institution": "Boston University", "country": "US", "position": "PhD student" }, { "name": "Mark Crovella", "gender": "Male", "institution": "Boston University", "country": "", "position": "Full Professor" }, { "name": "Themistoklis Nikas", "gender": "Male", "institution": "University of Ioannina", "country": "", "position": "Undergrad student" }, { "name": "Wagner Meira Jr.", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "Researcher" }, { "name": "Zilu Tang", "gender": "Male", "institution": "Boston University, Boston University", "country": "", "position": "PhD student" }, { "name": "Pedro Henrique Calais Guerra", "gender": "Male", "institution": "Universidade Federal de Minas Gerais, Universidade Federal de Minas Gerais", "country": "", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 6, "n_ref_uni": 50, "n_ref": 81, "n_ref_all": 107, "n_fig": 6, "n_tab": 5, "L_tab": 2411, "n_element_tab": 90, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1004, "formula_len_all_1": 214, "len_all": 220553, "len_all_1": 64808, "len_abs": 1925, "len_title": 132, "len_sents": 65171, "len_sents_1": 31880, "n_sents": 515, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1928, "title": "Beyond accuracy: understanding the performance of LLMs on exams designed for humans", "abs": "Many recent studies of LLM performance have focused on the ability of LLMs to achieve outcomes comparable to humans on academic and professional exams. However, it is not clear whether such studies shed light on the extent to which models show reasoning ability, and there is controversy about the significance and implications of such results. We seek to look more deeply into the question of how and whether the performance of LLMs on exams designed for humans reflects true aptitude inherent in LLMs. We do so by making use of the tools of psychometrics which are designed to perform meaningful measurement in test taking. We leverage a unique dataset that captures the detailed performance of over 5M students across 8 college-entrance exams given over a span of two years in Brazil. With respect to the evaluation of LLM abilities, we show that the tools of Item Response Theory (IRT) provide a more informative evaluation of model performance than the usual accuracy metrics employed in previous studies. Digging deeper, we show that the modeling framework of IRT, by explicitly modeling the difficulty levels of questions, allows us to quantitatively distinguish between LLMs that answer questions in \u201chuman-like\u201d patterns versus LLMs that do not. We also show how to quantitatively identify cases in which exam results are not reliable measurements of an LLM's ability. Using the tools of IRT we can also identify specific questions that appear to be either much easier, or much harder, for machines than for humans, and we give some reasons for those differences. Overall, our study shows that the conventional focus on accuracy as the primary performance metric for LLM studies does not allow us to deeply understand the true capabilities of LLMs and compare them to that of humans. Thus, we claim that psychometric modeling should play a larger role in the evaluation of LLM capabilities on exams designed for humans.", "keywords": [ "large language models", "model evaluation", "psychometrics" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 1 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "vgt2rSf6al", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Duoqian Miao", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Full Professor" }, { "name": "Guangyin Bao", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "MS student" }, { "name": "Qi Zhang", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Researcher" }, { "name": "Zhuojia Wu", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "PhD student" }, { "name": "Zixuan Gong", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 53, "n_ref": 104, "n_ref_all": 132, "n_fig": 15, "n_tab": 11, "L_tab": 3291, "n_element_tab": 364, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 616, "formula_len_all_1": 613, "len_all": 183421, "len_all_1": 68429, "len_abs": 1526, "len_title": 118, "len_sents": 49115, "len_sents_1": 32851, "n_sents": 415, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1534, "title": "MindSimulator: Exploring Brain Concept Localization via Synthetic fMRI", "abs": "Concept-selective regions within the human cerebral cortex exhibit significant activation in response to specific visual stimuli associated with particular concepts. Precisely localizing these regions stands as a crucial long-term goal in neuroscience to grasp essential brain functions and mechanisms. Conventional experiment-driven approaches hinge on manually constructed visual stimulus collections and corresponding brain activity recordings, constraining the support and coverage of concept localization. Additionally, these stimuli often consist of concept objects in unnatural contexts and are potentially biased by subjective preferences, thus prompting concerns about the validity and generalizability of the identified regions. To address these limitations, we propose a data-driven exploration approach. By synthesizing extensive brain activity recordings, we statistically localize various concept-selective regions. Our proposed MindSimulator leverages advanced generative technologies to learn the probability distribution of brain activity conditioned on concept-oriented visual stimuli. This enables the creation of simulated brain recordings that reflect real neural response patterns. Using the synthetic recordings, we successfully localize several well-studied concept-selective regions and validate them against empirical findings, achieving promising prediction accuracy. The feasibility opens avenues for exploring novel concept-selective regions and provides prior hypotheses for future neuroscience research.", "keywords": [ "Neuroscience", "fMRI encoding", "Generative model", "fMRI generation", "fMRI functional localizer", "Concept-selective voxel" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "vgplRfepVq", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Chen Liu", "gender": "Male", "institution": "City University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Xinping Chen", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 28, "n_ref": 68, "n_ref_all": 96, "n_fig": 4, "n_tab": 4, "L_tab": 799, "n_element_tab": 60, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 692, "n_element_tab_1": 53, "formula_len_all": 1527, "formula_len_all_1": 1359, "len_all": 114963, "len_all_1": 66563, "len_abs": 1032, "len_title": 147, "len_sents": 37163, "len_sents_1": 29647, "n_sents": 292, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1054, "title": "Gradient Inversion Transcript: A Generative Model to Reconstruct Training Data by Gradient Leakage", "abs": "We propose Gradient Inversion Transcript (GIT), a generic approach for reconstructing training data from gradient leakage in distributed learning using a generative model. Unlike traditional gradient matching techniques, GIT requires only the model architecture information, without access to the model's parameters, making it more applicable to real-world distributed learning settings. Additionally, GIT operates offline, eliminating the need for intensive gradient requests and online optimization.\nCompared to existing generative methods, GIT adaptively constructs a generative network, with an architecture specifically tailored to the structure of the distributed learning model. Our extensive experiments demonstrate that GIT significantly improves reconstruction accuracy, especially in the case of deep models.\nIn summary, we offer a more effective and theoretically grounded strategy for exploiting vulnerabilities of gradient leakage in distributed learning, advancing the understanding of privacy risks in collaborative learning environments.", "keywords": [ "distributed learning", "training data reconstruction", "generative model", "gradient inversion" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "vgZDcUetWS", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Berthy Feng", "gender": "Female", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Katherine Bouman", "gender": "unknown", "institution": "California Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Ricardo Baptista", "gender": "Male", "institution": "Deparment of Computing + Mathematical Sciences, California Institute of Technology", "country": "", "position": "Instructor" } ], "n_formula": 22, "n_formula_1": 13, "n_ref_uni": 77, "n_ref": 156, "n_ref_all": 192, "n_fig": 10, "n_tab": 3, "L_tab": 1221, "n_element_tab": 143, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 788, "n_element_tab_1": 32, "formula_len_all": 8195, "formula_len_all_1": 8234, "len_all": 218326, "len_all_1": 86495, "len_abs": 2259, "len_title": 111, "len_sents": 61804, "len_sents_1": 36784, "n_sents": 425, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1640, "title": "Neural Approximate Mirror Maps for Constrained Diffusion Models", "abs": "Diffusion models excel at creating visually-convincing images, but they often struggle to meet subtle constraints inherent in the training data. Such constraints could be physics-based (e.g., satisfying a PDE), geometric (e.g., respecting symmetry), or semantic (e.g., including a particular number of objects). When the training data all satisfy a certain constraint, enforcing this constraint on a diffusion model makes it more reliable for generating valid synthetic data and solving constrained inverse problems. However, existing methods for constrained diffusion models are restricted in the constraints they can handle. For instance, recent work proposed to learn mirror diffusion models (MDMs), but analytical mirror maps only exist for convex constraints and can be challenging to derive. We propose *neural approximate mirror maps* (NAMMs) for general, possibly non-convex constraints. Our approach only requires a differentiable distance function from the constraint set. We learn an approximate mirror map that transforms data into an unconstrained space and a corresponding approximate inverse that maps data back to the constraint set. A generative model, such as an MDM, can then be trained in the learned mirror space and its samples restored to the constraint set by the inverse map. We validate our approach on a variety of constraints, showing that compared to an unconstrained diffusion model, a NAMM-based MDM substantially improves constraint satisfaction. We also demonstrate how existing diffusion-based inverse-problem solvers can be easily applied in the learned mirror space to solve constrained inverse problems.", "keywords": [ "generative models", "diffusion models", "mirror maps", "constrained generation", "inverse problems" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "vgXI1Ws0ma", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Fan Feng", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Postdoc" }, { "name": "Hongye Cao", "gender": "Male", "institution": "nanjing university", "country": "", "position": "PhD student" }, { "name": "Jing Huo", "gender": "Female", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Meng Fang", "gender": "Male", "institution": "University of Liverpool", "country": "", "position": "Assistant Professor" }, { "name": "Shaokang Dong", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "PhD student" }, { "name": "Tianpei Yang", "gender": "Female", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yang Gao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 12, "n_ref_uni": 36, "n_ref": 83, "n_ref_all": 135, "n_fig": 38, "n_tab": 10, "L_tab": 3717, "n_element_tab": 266, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 46, "n_element_tab_1": 7, "formula_len_all": 1320, "formula_len_all_1": 965, "len_all": 192469, "len_all_1": 60650, "len_abs": 1730, "len_title": 124, "len_sents": 62796, "len_sents_1": 28169, "n_sents": 513, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1669, "title": "Towards Empowerment Gain through Causal Structure Learning in Model-Based Reinforcement Learning", "abs": "In Model-Based Reinforcement Learning (MBRL), incorporating causal structures into dynamics models provides agents with a structured understanding of the environments, enabling efficient decision. \nEmpowerment as an intrinsic motivation enhances the ability of agents to actively control their environments by maximizing the mutual information between future states and actions. \nWe posit that empowerment coupled with causal understanding can improve controllability, while enhanced empowerment gain can further facilitate causal reasoning in MBRL. \nTo improve learning efficiency and controllability, we propose a novel framework, Empowerment through Causal Learning (ECL), where an agent with the awareness of causal dynamics models achieves empowerment-driven exploration and optimizes its causal structure for task learning. \nSpecifically, ECL operates by first training a causal dynamics model of the environment based on collected data. We then maximize empowerment under the causal structure for exploration, simultaneously using data gathered through exploration to update causal dynamics model to be more controllable than dense dynamics model without causal structure. In downstream task learning, an intrinsic curiosity reward is included to balance the causality, mitigating overfitting. \nImportantly, ECL is method-agnostic and is capable of integrating various causal discovery methods. \nWe evaluate ECL combined with $3$ causal discovery methods across $6$ environments including pixel-based tasks, demonstrating its superior performance compared to other causal MBRL methods, in terms of causal discovery, sample efficiency, and asymptotic performance.", "keywords": [ "Causal RL", "MBRL", "Empowerment", "Intrinsic Motivation" ], "rating_list": [ 6, 6, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 4, 4, 2, 3 ], "contribution_list": [ 3, 3, 4, 3, 2 ], "confidence_list": [ 4, 3, 5, 3, 3 ] }, { "paper_id": "vgV4y086FY", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Guy Kornowski", "gender": "unknown", "institution": "Weizmann Institute of Science", "country": "", "position": "PhD student" } ], "n_formula": 89, "n_formula_1": 41, "n_ref_uni": 80, "n_ref": 123, "n_ref_all": 150, "n_fig": 0, "n_tab": 4, "L_tab": 937, "n_element_tab": 90, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 22, "n_element_tab_1": 1, "formula_len_all": 10071, "formula_len_all_1": 4362, "len_all": 189767, "len_all_1": 69429, "len_abs": 858, "len_title": 91, "len_sents": 42833, "len_sents_1": 27432, "n_sents": 385, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 974, "title": "Differentially Private Bilevel Optimization", "abs": "We present differentially private (DP) algorithms for bilevel optimization, a problem class that received significant attention lately in various machine learning applications.\nThese are the first DP algorithms for this task that are able to provide any desired privacy, while also avoiding Hessian computations which are prohibitive in large-scale settings.\nUnder the well-studied setting in which the upper-level is not necessarily convex and the lower-level problem is strongly-convex, our proposed gradient-based $(\\epsilon,\\delta)$-DP algorithm returns a point with hypergradient norm at most $\\widetilde{\\mathcal{O}}\\left((\\sqrt{d_\\mathrm{up}}/\\epsilon n)^{1/2}+(\\sqrt{d_\\mathrm{low}}/\\epsilon n)^{1/3}\\right)$ where $n$ is the dataset size, and $d_\\mathrm{up}/d_\\mathrm{low}$ are the upper/lower level dimensions.\nOur analysis covers constrained and unconstrained problems alike, accounts for mini-batch gradients, and applies to both empirical and population losses.", "keywords": [ "Bilevel optimization", "differential privacy", "nonconvex optimization", "first-order methods" ], "rating_list": [ 8, 8, 5, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "vgQmK5HHfz", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Haoran Ni", "gender": "Male", "institution": "University of Warwick", "country": "GB", "position": "PhD student" }, { "name": "Martin Lotz", "gender": "Male", "institution": "University of Warwick", "country": "", "position": "Associate Professor" } ], "n_formula": 31, "n_formula_1": 27, "n_ref_uni": 30, "n_ref": 58, "n_ref_all": 66, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2004, "formula_len_all_1": 1655, "len_all": 89635, "len_all_1": 61072, "len_abs": 487, "len_title": 113, "len_sents": 27355, "len_sents_1": 26409, "n_sents": 226, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 487, "title": "A Normalizing Flows based Difference-of-Entropies Estimator for Mutual Information", "abs": "Estimating Mutual Information (MI), a key measure of dependence of random quantities without specific modelling assumptions, is a challenging problem in high dimensions. We propose a novel mutual information estimator based on parametrizing conditional densities using normalizing flows, a deep generative model that has gained popularity in recent years. This estimator leverages a block autoregressive structure to achieve improved bias-variance trade-offs on standard benchmark tasks.", "keywords": [ "Normalizing flows", "mutual information", "generative models" ], "rating_list": [ 5, 8, 5, 5, 3 ], "soundness_list": [ 3, 4, 3, 3, 2 ], "presentation_list": [ 3, 4, 2, 3, 1 ], "contribution_list": [ 3, 3, 2, 2, 1 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "vgMAtJONKX", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Chenglong Ye", "gender": "unknown", "institution": "University of Kentucky", "country": "", "position": "Assistant Professor" }, { "name": "Zeya Wang", "gender": "Male", "institution": "University of Kentucky", "country": "US", "position": "Assistant Professor" } ], "n_formula": 57, "n_formula_1": 19, "n_ref_uni": 44, "n_ref": 80, "n_ref_all": 133, "n_fig": 9, "n_tab": 10, "L_tab": 31117, "n_element_tab": 5536, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 4264, "n_element_tab_1": 1001, "formula_len_all": 3352, "formula_len_all_1": 1366, "len_all": 264750, "len_all_1": 76059, "len_abs": 1120, "len_title": 130, "len_sents": 54351, "len_sents_1": 27717, "n_sents": 421, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1124, "title": "Towards Accurate Validation in Deep Clustering through Unified Embedding Learning", "abs": "Deep clustering integrates deep neural networks into the clustering process, simultaneously learning embedding spaces and cluster assignments. However, significant challenges remain in evaluating and comparing the performance of different deep clustering algorithms\u2014or even different training runs of the same algorithm. First, evaluating the clustering results from different models in the same high-dimensional input space is impractical due to the curse of dimensionality. Second, comparing the clustering results of different models in their respective learned embedding spaces introduces discrepancies, as existing validation measures are designed for comparisons within the same feature space. To address these issues, we propose a novel evaluation framework that learns a unified embedding space. This approach aligns different embedding spaces into a common space, enabling accurate comparison of clustering results across different models and training runs. Extensive experiments demonstrate the effectiveness of our framework, showing improved consistency and reliability in evaluating deep clustering performance.", "keywords": [ "Internal validation measures", "Deep clustering", "Clustering evaluation", "Unified embedding learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "vgMAtJONKX", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Chenglong Ye", "gender": "unknown", "institution": "University of Kentucky", "country": "", "position": "Assistant Professor" }, { "name": "Zeya Wang", "gender": "Male", "institution": "University of Kentucky", "country": "US", "position": "Assistant Professor" } ], "n_formula": 57, "n_formula_1": 19, "n_ref_uni": 44, "n_ref": 80, "n_ref_all": 133, "n_fig": 9, "n_tab": 10, "L_tab": 31117, "n_element_tab": 5536, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 4264, "n_element_tab_1": 1001, "formula_len_all": 3352, "formula_len_all_1": 1366, "len_all": 264750, "len_all_1": 76059, "len_abs": 1120, "len_title": 130, "len_sents": 54351, "len_sents_1": 27717, "n_sents": 421, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1124, "title": "Towards Accurate Validation in Deep Clustering through Unified Embedding Learning", "abs": "Deep clustering integrates deep neural networks into the clustering process, simultaneously learning embedding spaces and cluster assignments. However, significant challenges remain in evaluating and comparing the performance of different deep clustering algorithms\u2014or even different training runs of the same algorithm. First, evaluating the clustering results from different models in the same high-dimensional input space is impractical due to the curse of dimensionality. Second, comparing the clustering results of different models in their respective learned embedding spaces introduces discrepancies, as existing validation measures are designed for comparisons within the same feature space. To address these issues, we propose a novel evaluation framework that learns a unified embedding space. This approach aligns different embedding spaces into a common space, enabling accurate comparison of clustering results across different models and training runs. Extensive experiments demonstrate the effectiveness of our framework, showing improved consistency and reliability in evaluating deep clustering performance.", "keywords": [ "Internal validation measures", "Deep clustering", "Clustering evaluation", "Unified embedding learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "vf8iou7FNF", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 5, "author_info_list": [ { "name": "Arnav Arora", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Piyush Jha", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Pranavkrishna Suresh", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Prithwish Jana", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Vijay Ganesh", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 59, "n_ref": 109, "n_ref_all": 127, "n_fig": 4, "n_tab": 7, "L_tab": 2096, "n_element_tab": 182, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1027, "n_element_tab_1": 83, "formula_len_all": 59, "formula_len_all_1": 297, "len_all": 179880, "len_all_1": 72789, "len_abs": 1919, "len_title": 98, "len_sents": 49125, "len_sents_1": 35617, "n_sents": 337, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 2002, "title": "RLSF: Reinforcement Learning via Symbolic Feedback", "abs": "Reinforcement Learning with Human Feedback (RLHF) is considered a standard approach to fine-tuning Large Language Models (LLMs) but faces challenges such as unsound reward models, sparse rewards, and difficulties in collecting human preference data, limiting its effectiveness for complex, domain-specific tasks.\n\nWe propose Reinforcement Learning via Symbolic Feedback (RLSF), a novel fine-tuning paradigm where reasoning tools (e.g., solvers, provers, algebra systems) serve as the RL environment, providing fine-grained feedback via \\textit{poly-sized certificates} (e.g., proofs) that characterize errors in the LLM-generated object with respect to specific correctness criteria. RLSF aims to improve the domain-specific understanding of LLMs more effectively than traditional reward signals. By enabling token-level corrections without requiring differentiable reasoning systems, RLSF addresses key limitations of traditional reward models.\n\nVia extensive evaluations, we show that our RLSF-based fine-tuning of LLMs outperforms traditional approaches on five different applications (that have some associated logical or domain constraints), namely, program synthesis from natural language pseudo-code to programming language (+31.43\\% in functional correctness for Google's CodeGemma-2b compared to supervised fine-tuning, +17.01\\% in functional correctness compared to GPT-3.5 -- 100$\\boldsymbol\\times$ larger), three chemistry tasks (+5.5\\% exact match for molecule generation, +19.4\\% exact match for forward synthesis, +33.7\\% exact match for retrosynthesis, using Meta's Galactica-1.3b, compared to GPT-4 -- 1000$\\boldsymbol\\times$ larger), and solving the Game of 24 (+25\\% success rate using Meta's Llama2-7b compared to traditional methods, and +7\\% success rate compared to GPT-3.5 -- 25$\\boldsymbol\\times$ larger). A takeaway is that fine-tuning via RLSF enables relatively smaller LLMs to significantly outperform closed-source models that are orders of magnitude larger (e.g., GPT-4).", "keywords": [ "Symbolic Feedback", "Reinforcement Learning", "Large Language Models", "Program Synthesis" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 2, 3, 2, 4 ] }, { "paper_id": "vf5aUZT0Fz", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Lorenzo Sani", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Meghdad Kurmanji", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "William F. Shen", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Xinchi Qiu", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Yan Gao", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Alexandru Andrei Iacob", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "DONGQI CAI", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Nic Lane", "gender": "Male", "institution": "Flower Labs", "country": "GB", "position": "Chief Scientific Officer" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 66, "n_ref": 157, "n_ref_all": 224, "n_fig": 6, "n_tab": 19, "L_tab": 12997, "n_element_tab": 2018, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 3689, "n_element_tab_1": 355, "formula_len_all": 724, "formula_len_all_1": 691, "len_all": 303716, "len_all_1": 62573, "len_abs": 1175, "len_title": 107, "len_sents": 72663, "len_sents_1": 26806, "n_sents": 451, "n_sents_1": 135, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 59, "L_abs": 1179, "title": "DEPT: Decoupled Embeddings for Pre-training Language Models", "abs": "Past works have shown that lexical, syntactical, and semantical differences in heterogeneous data sources can cause challenges such as negative interference or the ``curse of multilinguality''. Because of this, training on such heterogeneous corpora requires extensive and costly efforts to balance data mixtures. We propose a novel pre-training framework to alleviate this curse. Our method, DEPT, decouples embeddings from the transformer body while simultaneously training the latter in multiple contexts without a shared global vocabulary. DEPT: (1) trains robustly and effectively under significant data heterogeneity, (2) reduces token embedding parameters by up to 80% and communication costs by 714x for billion-scale models, (3) enhances transformer body plasticity and generalization, improving average perplexity upward of 15.3-20% and improving performance for downstream fine-tuning in our experiments, and (4) permits training with custom optimized vocabularies per data source. We demonstrate DEPT's potential via the first vocabulary-agnostic federated multilingual pre-training of a billion-scale model, reducing total parameters by 24% versus standard training.", "keywords": [ "Decentralized Training", "Federated Learning", "Multi-domain Training", "Multilingual Training" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 4, 4, 3 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "vf5M8YaGPY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Alex Beutel", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Researcher" }, { "name": "Eric Wallace", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Researcher" }, { "name": "Johannes Heidecke", "gender": "unknown", "institution": "OpenAI", "country": "", "position": "Researcher" }, { "name": "Kai Yuanqing Xiao", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Lilian Weng", "gender": "Female", "institution": "OpenAI", "country": "", "position": "Research Scientist" }, { "name": "Reimar Heinrich Leike", "gender": "Male", "institution": "OpenAI", "country": "", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 54, "n_ref_all": 74, "n_fig": 5, "n_tab": 4, "L_tab": 1111, "n_element_tab": 48, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 648, "n_element_tab_1": 28, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 104760, "len_all_1": 67691, "len_abs": 1236, "len_title": 127, "len_sents": 36364, "len_sents_1": 30085, "n_sents": 272, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 955, "title": "The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions", "abs": "Today's LLMs are susceptible to prompt injections, jailbreaks, and other attacks that allow adversaries to overwrite a model's original instructions with their own malicious prompts. In this work, we argue that one of the primary vulnerabilities underlying these attacks is that LLMs often consider system prompts (e.g., text from an application developer) to be the same priority as text from untrusted users and third parties. To address this, we propose an instruction hierarchy that explicitly defines how models should behave when instructions of different priorities conflict. We then propose a data generation method to demonstrate this hierarchical instruction following behavior, which teaches LLMs to selectively ignore lower-privileged instructions. We apply this method to GPT-3.5, showing that it drastically increases robustness---even for attack types not seen during training---while imposing minimal degradations on standard capabilities.", "keywords": [ "Jailbreaks", "Prompt Injections", "Adversarial Robustness" ], "rating_list": [ 3, 6, 8, 3, 8 ], "soundness_list": [ 1, 3, 3, 1, 3 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 4, 2, 4 ], "confidence_list": [ 5, 5, 4, 4, 4 ] }, { "paper_id": "veyPSmKrX4", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Colin Conwell", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Postdoc" }, { "name": "George A. Alvarez", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Full Professor" }, { "name": "Jacob S. Prince", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Leyla Isik", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" }, { "name": "Talia Konkle", "gender": "Female", "institution": "Harvard University", "country": "US", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 58, "n_ref": 101, "n_ref_all": 111, "n_fig": 5, "n_tab": 1, "L_tab": 35, "n_element_tab": 1, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 698, "n_element_tab_1": 33, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 166470, "len_all_1": 61523, "len_abs": 1381, "len_title": 96, "len_sents": 55653, "len_sents_1": 33994, "n_sents": 279, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1383, "title": "Rethinking Language-Alignment in Human Visual Cortex with Syntax Manipulation and Word Models", "abs": "Recent success predicting human ventral visual system responses to images from large language model (LLM) representations of image captions has sparked renewed interest in the possibility that high-level visual representations are aligned to language. Here, we further explore this possibility using image-caption pairs from the Natural Scenes fMRI Dataset, examining how well language-only representations of image captions predict image-evoked human visual cortical responses, compared to predictions based on vision model responses to the images themselves. As in recent work, we find that unimodal language models predict brain responses in human visual cortex as well as unimodal vision models. However, we find that the predictive power of large language models rests almost entirely on their ability to capture information about the nouns present in image descriptions, with little to no role for syntactic structure or semantic compositionality in predicting neural responses to static natural scenes. We propose that the convergence between language-model and vision-model representations and those of high-level visual cortex arises not from direct interaction between vision and language, but instead from common reference to real-world entities, and the prediction of brain data whose principal variance is defined by common objects in common, non-compositional contexts.", "keywords": [ "multimodality", "language models", "vision models", "visuosemantics", "visual neuroscience" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "veiSkPqIXm", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Hao-Zhe Tan", "gender": "unknown", "institution": "Nanjing University", "country": "", "position": "MS student" }, { "name": "Lan-Zhe Guo", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhi Zhou", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Lv Songlin", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Yu-Feng Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Zikang Wang", "gender": "Male", "institution": "nanjing university", "country": "", "position": "Undergrad student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 28, "n_ref": 55, "n_ref_all": 68, "n_fig": 5, "n_tab": 20, "L_tab": 9306, "n_element_tab": 1296, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 5176, "n_element_tab_1": 287, "formula_len_all": 811, "formula_len_all_1": 903, "len_all": 132472, "len_all_1": 58387, "len_abs": 4180, "len_title": 124, "len_sents": 29689, "len_sents_1": 23925, "n_sents": 189, "n_sents_1": 148, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1581, "title": "OpenPL: Realistic Evaluation of Prompt Learning for VLM in Open Environments", "abs": "Vision-language models (VLMs) have demonstrated impressive zero-shot capabilities across various image classification tasks. Their performance can be further enhanced through prompt learning methods. To evaluate the effectiveness of prompt learning, it is important to assess its robustness to new classes and distributional shifts. However, current studies typically assume single data distribution shifts and pre-known new class space, which still have gaps with real-world open environments where data distributions and classes are often uncertain and subject to continuous change. To better analyze the robustness of prompt learning methods in more realistic scenarios, we propose a novel evaluation benchmark called OpenPL from the following perspectives: 1) We reconstruct multiple scenarios of open environments, encompassing dynamic class changes, dynamic distribution shifts, and dynamic co-evolution of both distribution and classes; 2) We propose a series of new performance metrics for prompt learning methods based on the Dynamic Robustness Curve (DRC) to better understand their robustness in open environments; 3) We re-implement diverse prompt learning methods and evaluate their performance on the proposed OpenPL benchmark. The results show that no current prompt learning method is robust to open environments and no meaningful performance improvement is achieved compared to the zero-shot performance, designing robust prompt learning methods remains a difficult task. All re-implementations are available at \\url{https://anonymous.4open.science/r/OpenPL-565E}.", "keywords": [ "VLM; Prompt Learning; Open environments" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "vePZdNvrO9", "primary_area": "reinforcement learning", "n_author": 9, "author_info_list": [ { "name": "Cao Liu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Chunkang Zhang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Hongyu Lin", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Ke Zeng", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Le Sun", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Wan Guanglu", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Xianpei Han", "gender": "Male", "institution": "Institute of Software, CAS", "country": "CN", "position": "Professor" }, { "name": "Xunliang Cai", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Principal Researcher" }, { "name": "Yaojie Lu", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 20, "n_ref_uni": 30, "n_ref": 68, "n_ref_all": 86, "n_fig": 5, "n_tab": 3, "L_tab": 917, "n_element_tab": 128, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 885, "n_element_tab_1": 126, "formula_len_all": 1914, "formula_len_all_1": 1800, "len_all": 126294, "len_all_1": 58708, "len_abs": 1536, "len_title": 108, "len_sents": 37073, "len_sents_1": 25930, "n_sents": 260, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1252, "title": "GameInstruct: Teaching Machines to Reason via Chameleon Game", "abs": "Self-play has emerged as a promising approach for generating alignment data to reduce the data annotation costs during the alignment process.\nBy introducing specific game rules and utilizes the model\u2019s own language capabilities to generate data samples, self-play has achieved promising results.\nHowever, traditional self-play methods face two major challenges: insufficient data diversity during self-iterative training and difficulties in reward signal design.\nTo solve these problems, this paper introduces GameInstruct, a complex multi-player adversarial environment that increases the complexity of self-play generated data during self-iterative training.\nSpecifically, we employ the ``Chameleon Game'', where interactions between multiple players raise the diversity of the generated data, improving the model\u2019s reasoning abilities, \nAdditionally, we further propose a dynamic reward algorithm to capture signals within player conversations during the whole game.\nExperimental results show that compared to existing self-play methods, GameInstruct achieves significant improvements on the HuggingFace Open-LLM-Leaderboard reasoning benchmark while demonstrating continuous improvement and increasing data diversity during self-iterative training.", "keywords": [ "Large Language Model", "Self-play", "Alignment" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 2, 3, 4, 5 ] }, { "paper_id": "veNewXAdHE", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 1, "author_info_list": [ { "name": "Vaibhav Ganatra", "gender": "Male", "institution": "Birla Institute of Technology and Science, K.K. Birla Goa Campus", "country": "", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 20, "n_ref": 73, "n_ref_all": 88, "n_fig": 4, "n_tab": 9, "L_tab": 1808, "n_element_tab": 193, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 1855, "n_element_tab_1": 200, "formula_len_all": 180, "formula_len_all_1": 180, "len_all": 92392, "len_all_1": 68045, "len_abs": 1099, "len_title": 118, "len_sents": 28749, "len_sents_1": 28592, "n_sents": 201, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1105, "title": "LoRe - Logarithm Regularization for Few-Shot Class Incremental Learning", "abs": "Few-Shot Class-Incremental Learning (FSCIL) aims to adapt to new classes with very limited data, while remembering information about all the previously seen classes. Current FSCIL methods freeze the feature extractor in the incremental sessions to prevent catastrophic forgetting. However, to perform well on the incremental classes, many methods reserve feature spaces during base training to allow\nsufficient space for incremental classes. We hypothesize that such feature space reservation sharpens the minima of the loss-landscape, resulting in sub-optimal performance. Motivated by the superior generalization of wide minima, we propose LoRe - logarithm regularization to guide the model optimization to wider minima. Moreover, we propose a denoised distance metric when considering similarity with the poorly calibrated prototypes. Comprehensive evaluations across three benchmark datasets reveal that LoRe not only achieves state-of-the-art performance but also produces more robust prototypes. Additionally, we demonstrate that LoRe can be leveraged to enhance the performance of existing methods.", "keywords": [ "Few-Shot Class Incremental Learning", "Continual Learning", "Logarithmic Regularization", "Wide Minima" ], "rating_list": [ 3, 3, 5, 5, 5 ], "soundness_list": [ 2, 1, 3, 2, 3 ], "presentation_list": [ 1, 1, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 5, 4, 3 ] }, { "paper_id": "ve5Omkxc13", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Faming Liang", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Full Professor" }, { "name": "Frank Shih", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" } ], "n_formula": 59, "n_formula_1": 34, "n_ref_uni": 35, "n_ref": 62, "n_ref_all": 77, "n_fig": 6, "n_tab": 1, "L_tab": 635, "n_element_tab": 24, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 635, "n_element_tab_1": 24, "formula_len_all": 4158, "formula_len_all_1": 2410, "len_all": 116863, "len_all_1": 59013, "len_abs": 1439, "len_title": 154, "len_sents": 33907, "len_sents_1": 24675, "n_sents": 282, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1445, "title": "Latent Trajectory: A New Framework for Actor-Critic Reinforcement Learning with Uncertainty Quantification", "abs": "Uncertainty quantification for deep neural networks is crucial for building reliable modern AI models. This challenge is particularly pronounced in deep reinforcement learning, where agents continuously learn from their interactions with stochastic environments, and the uncertainty of the value function is a key concern for ensuring reliable and robust RL applications. The complexity increases in actor-critic methods, as the training process alternates between optimizing the actor and critic networks, whose optimization nature makes the uncertainty of the value function hard to be quantified. \nTo address this issue, we introduce a novel approach to RL training that conceptualizes transition trajectories as latent variables. Building on this framework, we propose an adaptive Stochastic Gradient Markov Chain Monte Carlo (SGMCMC) algorithm for training deep actor-critic models. This new training method allows for the implicit integration of latent transition trajectories, resulting in a trajectory-independent training process. We provide theoretical guarantees for the convergence of our algorithm and offer empirical evidence showing improvements in both performance and robustness of the deep actor-critic model under our Latent Trajectory Framework (LTF). Furthermore, this framework enables accurate uncertainty quantification for the value function of the RL system, paving the way for more reliable and robust RL applications.", "keywords": [ "Reinforcement learning", "Stochastic gradient MCMC", "Bayesian sampling", "Uncertainty quantification" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 3, 2 ] }, { "paper_id": "vdUYa7N8Mt", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Deniz Gunduz", "gender": "unknown", "institution": "Imperial College London", "country": "", "position": "Full Professor" }, { "name": "Yassine Hamdi", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Aaron Wagner", "gender": "unknown", "institution": "Cornell University", "country": "", "position": "Full Professor" } ], "n_formula": 166, "n_formula_1": 26, "n_ref_uni": 27, "n_ref": 81, "n_ref_all": 120, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10067, "formula_len_all_1": 1346, "len_all": 185147, "len_all_1": 63526, "len_abs": 884, "len_title": 113, "len_sents": 56500, "len_sents_1": 28106, "n_sents": 608, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 886, "title": "The Rate-Distortion-Perception Trade-Off with Algorithmic Realism", "abs": "Realism constraints (or constraints on perceptual quality) have received considerable recent attention within the context of lossy compression, particularly of images. Theoretical studies of lossy compression indicate that high-rate common randomness between the compressor and the decompressor is a valuable resource for achieving realism. On the other hand, the utility of significant amounts of common randomness at test time has not been noted in practice. We offer an explanation for this discrepancy by considering a realism constraint that requires satisfying a universal critic that inspects realizations of individual compressed images, or batches thereof. We characterize the optimal rate-distortion-perception trade-off under such a realism constraint, and show that it is asymptotically achievable without any common randomness, unless the batch size is impractically large.", "keywords": [ "lossy compression", "perceptual quality", "rate-distortion-perception trade-off", "randomization", "universal critics" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "vdHSMJpBya", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Dongxian Wu", "gender": "Male", "institution": "The University of Tokyo", "country": "", "position": "Postdoc" }, { "name": "Yichuan Mo", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yifei Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yisen Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yiwen Guo", "gender": "Male", "institution": "ByteDance", "country": "CN", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 88, "n_ref_all": 118, "n_fig": 7, "n_tab": 12, "L_tab": 5780, "n_element_tab": 850, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2865, "n_element_tab_1": 442, "formula_len_all": 392, "formula_len_all_1": 367, "len_all": 152556, "len_all_1": 68627, "len_abs": 1130, "len_title": 104, "len_sents": 44924, "len_sents_1": 29148, "n_sents": 344, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1134, "title": "Towards Reliable Backdoor Attacks on Vision Transformers", "abs": "Backdoor attacks, which make Convolution Neural Networks (CNNs) exhibit specific behaviors in the presence of a predefined trigger, bring risks to the usage of CNNs. These threats should be also considered on Vision Transformers. However, previous studies found that the existing backdoor attacks are powerful enough in ViTs to bypass common backdoor defenses, i.e., these defenses either fail to reduce the attack success rate or cause a significant accuracy drop. This study investigates the existing backdoor attacks/defenses and finds that this kind of achievement is over-optimistic, caused by inappropriate adaption of defenses from CNNs to ViTs. Existing backdoor attacks can still be easily defended against with proper inheritance from CNNs. Furthermore, we propose a more reliable attack: adding a small perturbation on the trigger is enough to help existing attacks more persistent against various defenses. We hope our contributions, including the finding that existing attacks are still easy to defend with adaptations and the new backdoor attack, will promote more in-depth research into the backdoor robustness of ViTs.", "keywords": [ "Backdoor Attacks", "Vision Transformer" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "vcX0k4rGTt", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Christos Louizos", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "NL", "position": "Researcher" }, { "name": "Dharmesh Tailor", "gender": "unknown", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Eric Nalisnick", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" }, { "name": "Alvaro Henrique Chaim Correia", "gender": "Male", "institution": "Qualcomm", "country": "", "position": "Researcher" } ], "n_formula": 49, "n_formula_1": 30, "n_ref_uni": 46, "n_ref": 110, "n_ref_all": 176, "n_fig": 2, "n_tab": 6, "L_tab": 5501, "n_element_tab": 289, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5094, "formula_len_all_1": 2471, "len_all": 189881, "len_all_1": 71910, "len_abs": 4325, "len_title": 145, "len_sents": 56215, "len_sents_1": 31685, "n_sents": 469, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1463, "title": "Approximating Full Conformal Prediction for Neural Network Regression with Gauss-Newton Influence", "abs": "Uncertainty quantification is an important prerequisite for the deployment of deep learning models in safety-critical areas. Yet, this hinges on the uncertainty estimates being useful to the extent the predictive prediction intervals are well-calibrated and sharp. In the absence of inherent uncertainty estimates (e.g. pretrained models), popular approaches that operate post-hoc include Laplace\u2019s method and split conformal prediction (split-CP). However, Laplace\u2019s method can be miscalibrated when the model is misspecified and split-CP requires sample splitting, and thus comes at the expense of statistical efficiency. In this work, we construct prediction intervals for neural network regressors post-hoc without held-out data. This is achieved by approximating the full conformal prediction method (full-CP). Whilst full-CP nominally requires retraining the model for every test point and candidate label, we propose to train just once and locally perturb model parameters using Gauss-Newton influence to approximate the effect of retraining. Coupled with linearization of the network, we express the absolute residual nonconformity score as a piecewise linear function of the candidate label allowing for an efficient procedure that avoids the exhaustive search over the output space. On standard regression benchmarks and bounding box localization, we show the resulting prediction intervals are locally-adaptive and often tighter than those of split-CP.", "keywords": [ "conformal", "laplace", "influence", "neural network", "deep learning", "uncertainty" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 2, 4, 1, 3 ], "confidence_list": [ 2, 4, 5, 3 ] }, { "paper_id": "vcJiPLeC48", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Ana Cukarska", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Erik Lien Bolager", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Felix Dietrich", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Assistant Professor" }, { "name": "Iryna Burak", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Zahra Monfared", "gender": "Female", "institution": "Heidelberg University, Ruprecht-Karls-Universit\u00e4t Heidelberg", "country": "DE", "position": "Research Group Leader" } ], "n_formula": 99, "n_formula_1": 25, "n_ref_uni": 43, "n_ref": 93, "n_ref_all": 149, "n_fig": 15, "n_tab": 7, "L_tab": 2573, "n_element_tab": 318, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3879, "n_element_tab_1": 206, "formula_len_all": 6133, "formula_len_all_1": 2034, "len_all": 224565, "len_all_1": 63967, "len_abs": 1296, "len_title": 124, "len_sents": 73451, "len_sents_1": 26037, "n_sents": 635, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1302, "title": "Gradient-free training of recurrent neural networks", "abs": "Recurrent neural networks are a successful neural architecture for many time-dependent problems, including time series analysis, forecasting, and modeling of dynamical systems. Training such networks with backpropagation through time is a notoriously difficult problem because their loss gradients tend to explode or vanish. In this contribution, we introduce a computational approach to construct all weights and biases of a recurrent neural network without using gradient-based methods. The approach is based on a combination of random feature networks and Koopman operator theory for dynamical systems. The hidden parameters of a single recurrent block are sampled at random, while the outer weights are constructed using extended dynamic mode decomposition. This approach alleviates all problems with backpropagation commonly related to recurrent networks. The connection to Koopman operator theory also allows us to start using results in this area to analyze recurrent neural networks. In computational experiments on time series, forecasting for chaotic dynamical systems, and control problems, as well as on weather data, we observe that the training time and forecasting accuracy of the recurrent neural networks we construct are improved when compared to commonly used gradient-based methods.", "keywords": [ "recurrent neural networks", "koopman operator", "random feature networks" ], "rating_list": [ 8, 6, 5, 5 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "vc1i3a4O99", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Jiayi Yuan", "gender": "unknown", "institution": "Rice University", "country": "", "position": "PhD student" }, { "name": "Wenlin Yao", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Xiaoming Zhai", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Associate Professor" }, { "name": "Xuansheng Wu", "gender": "unknown", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Ninghao Liu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 53, "n_ref": 167, "n_ref_all": 191, "n_fig": 4, "n_tab": 5, "L_tab": 1793, "n_element_tab": 124, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2091, "n_element_tab_1": 147, "formula_len_all": 527, "formula_len_all_1": 584, "len_all": 180883, "len_all_1": 68484, "len_abs": 1520, "len_title": 159, "len_sents": 52633, "len_sents_1": 30294, "n_sents": 389, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1476, "title": "Interpreting and Steering LLM Representations with Mutual Information-based Explanations on Sparse Autoencoders", "abs": "Large language models (LLMs) excel at addressing general human queries, yet they can falter or produce unexpected responses in specific scenarios. Gaining insight into the internal states of LLMs is key to understanding their successes and failures, as well as to refining their capabilities. Recent efforts have applied sparse autoencoders to learn a feature basis for explaining LLM hidden spaces. However, current post-hoc explanation methods can not effectively describe the semantic meaning of the learned features, and it is difficult to steer LLM behaviors by manipulating these features. Our analysis reveals that existing explanation methods suffer from the frequency bias issue, i.e., they tend to focus on trivial linguistic patterns rather than semantics. To overcome this, we propose explaining the learned features from a fixed vocabulary set to mitigate the frequency bias, and designing a novel explanation objective based on the mutual information theory to better express the meaning of the features. We further suggest two strategies to steer LLM representations by modifying sparse feature activations in response to user queries during runtime. Empirical results demonstrate that our method generates more discourse-level explanations than the baselines, and can effectively steer LLM behaviors to defend against jailbreak attacks in the wild. These findings highlight the value of explanations for steering LLM representations in downstream applications.", "keywords": [ "large language models", "sparse autoencoders", "usable xai", "explanations", "interpretability" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "vbr1OKK19i", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Hendrik Strobelt", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Visiting Scientist" }, { "name": "Kenza Amara", "gender": "Female", "institution": "ETH AI Center", "country": "CH", "position": "PhD student" }, { "name": "Lukas Klein", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "PhD student" }, { "name": "Mennatallah El-Assady", "gender": "unknown", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "", "position": "Assistant Professor" }, { "name": "Paul F Jaeger", "gender": "unknown", "institution": "Google", "country": "CA", "position": "Researcher" }, { "name": "Carsten Tim L\u00fcth", "gender": "Male", "institution": "Ruprecht-Karls-Universit\u00e4t Heidelberg", "country": "DE", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 0, "n_ref": 64, "n_ref_all": 81, "n_fig": 8, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 72, "n_element_tab_1": 3, "formula_len_all": 118, "formula_len_all_1": 118, "len_all": 81735, "len_all_1": 71855, "len_abs": 1986, "len_title": 135, "len_sents": 43507, "len_sents_1": 38171, "n_sents": 290, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1997, "title": "Why context matters in VQA & Reasoning: Semantic interventions for VLM input modalities", "abs": "The various limitations of Generative AI, such as hallucinations and model failures, have made it crucial to understand the role of different modalities in Visual Language Model (VLM) predictions. Our work investigates how the integration of information from image and text modalities influences the performance and behavior of VLMs in visual question answering (VQA) and reasoning tasks. We measure this effect through answer accuracy, reasoning quality, model uncertainty, and attention attribution to each modality. We study the interplay between text and image modalities in different configurations where visual content is essential for solving the VQA task. Our contributions include (1) the Semantic Interventions (SI)-VQA dataset, (2) a benchmark study of various VLM architectures under different modality configurations, and (3) the Interactive Semantic Interventions (ISI) tool. The SI-VQA dataset serves as the foundation for the benchmark, while the ISI tool provides an interface to test and apply semantic interventions in image and text inputs, enabling more fine-grained analysis. Our results show that complementary information between modalities improves answer and reasoning quality, while contradictory information harms model performance and confidence. Image text annotations have minimal impact on accuracy and uncertainty, slightly increasing image relevance. Attention analysis confirms the dominant role of image inputs over text in VQA tasks. In this study, we evaluate state-of-the-art VLMs that allow us to extract attention coefficients for each modality. A key finding is PaliGemma's harmful overconfidence, which poses a higher risk of silent failures compared to the LLaVA models. This work sets the foundation for rigorous analysis of modality integration, supported by datasets specifically designed for this purpose. The code is available at https://gitlab.com/dekfsx1/si-vlm-benchmark and the tool and dataset are hosted at https://gitlab.com/dekfsx1/isi-vlm.", "keywords": [ "Vision Language Model", "Vision Question Answering", "model failure", "multimodality", "interpretability", "semantic intervention" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 1, 1, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "vbmSSIhKAM", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Dongjie Fu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Jingyu Lu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Rongjie Huang", "gender": "Male", "institution": "FAIR", "country": "US", "position": "Researcher" }, { "name": "Shengpeng Ji", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Tao Jin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaoda Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Xize Cheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zehan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhang boyang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Hu Ruofan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 34, "n_ref": 81, "n_ref_all": 103, "n_fig": 3, "n_tab": 4, "L_tab": 3148, "n_element_tab": 118, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2093, "n_element_tab_1": 61, "formula_len_all": 144, "formula_len_all_1": 410, "len_all": 123733, "len_all_1": 63698, "len_abs": 1540, "len_title": 125, "len_sents": 37829, "len_sents_1": 28807, "n_sents": 288, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1561, "title": "VoxDialogue: Can Spoken Dialogue Systems Understand Information Beyond Words?", "abs": "With the rapid advancement of large models, voice assistants are gradually acquiring the ability to engage in open-ended daily conversations with humans. However, current spoken dialogue systems often overlook multi-modal information in audio beyond text, such as speech rate, volume, emphasis, and background sounds. Relying solely on Automatic Speech Recognition (ASR) can lead to the loss of valuable auditory cues, thereby weakening the system\u2019s ability to generate contextually appropriate responses. To address this limitation, we propose \\textbf{VoxDialogue}, a comprehensive benchmark for evaluating the ability of spoken dialogue systems to understand multi-modal information beyond text. Specifically, we have identified 12 attributes highly correlated with acoustic information beyond words and have meticulously designed corresponding spoken dialogue test sets for each attribute, encompassing a total of 4.5K multi-turn spoken dialogue samples. Finally, we evaluated several existing spoken dialogue models, analyzing their performance on the 12 attribute subsets of VoxDialogue. Experiments have shown that in spoken dialogue scenarios, many acoustic cues cannot be conveyed through textual information and must be directly interpreted from the audio input. In contrast, while direct spoken dialogue systems excel at processing acoustic signals, they still face limitations in handling complex dialogue tasks due to their restricted context understanding capabilities. All data and code will be open source at \\url{https://voxdialogue.github.io/}.", "keywords": [ "spoken dialogue system", "paralinguistic information", "benchmark" ], "rating_list": [ 6, 8, 3, 8, 8 ], "soundness_list": [ 3, 2, 2, 4, 3 ], "presentation_list": [ 3, 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 5, 3 ] }, { "paper_id": "vaJ4FObpXN", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Aaron M Ferber", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Postdoc" }, { "name": "Carla P Gomes", "gender": "Female", "institution": "Cornell University", "country": "US", "position": "Full Professor" }, { "name": "Utku Umur ACIKALIN", "gender": "Male", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 31, "n_ref": 53, "n_ref_all": 78, "n_fig": 5, "n_tab": 10, "L_tab": 5745, "n_element_tab": 1169, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2483, "n_element_tab_1": 509, "formula_len_all": 823, "formula_len_all_1": 671, "len_all": 150693, "len_all_1": 67279, "len_abs": 1997, "len_title": 133, "len_sents": 40762, "len_sents_1": 27107, "n_sents": 314, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1771, "title": "Learning to Explore and Exploit with GNNs for Unsupervised Combinatorial Optimization", "abs": "Combinatorial optimization (CO) problems are pervasive\nacross various domains, but their NP-hard nature often necessitates problem-specific\nheuristic algorithms. Recent advancements in deep learning have led to the development of learning-based heuristics, yet these approaches often struggle with limited search capabilities.\nWe introduce Explore-and-Exploit GNN ($X^2$GNN, pronounced x-squared GNN), \na novel unsupervised neural framework that combines exploration and exploitation for combinatorial search optimization:\ni) Exploration - $X^2$GNN generates multiple \nsolutions simultaneously, promoting diversity in the search space; \n(ii) Exploitation - $X^2$GNN employs neural stochastic iterative refinement, where sampled partial solutions guide the search toward promising regions and help escape local optima.\n$X^2$GNN employs neural stochastic iterative refinement to exploit partial existing solutions, guiding the search toward promising regions and helping escape local optima. By balancing exploration and exploitation $X^2$GNN achieves superior performance and generalization on several graph CO problems including Max Cut, Max Independent Set, and Max Clique. Notably, for large Max Clique problems, $X^2$GNN consistently generates solutions within 1.2\\% of optimality, while other state-of-the-art learning-based approaches struggle to reach within 22\\% of optimal. Moreover, $X^2$GNN consistently generates better solutions than Gurobi on large graphs for all three problems under reasonable time budgets. Furthermore, $X^2$GNN exhibits exceptional generalization capabilities. For the Maximum Independent Set problem, $X^2$GNN outperforms state-of-the-art methods even when trained on smaller or out-of-distribution graphs compared to the test set.", "keywords": [ "combinatorial optimization", "unsupervised learning", "graph neural networks" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "vaEPihQsAA", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chao Liang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Gaojie Lin", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Jianwen Jiang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Jiaqi Yang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Tianyun Zhong", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yanbo Zheng", "gender": "unknown", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Zerong Zheng", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 30, "n_ref": 97, "n_ref_all": 127, "n_fig": 12, "n_tab": 9, "L_tab": 2483, "n_element_tab": 313, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1252, "n_element_tab_1": 118, "formula_len_all": 385, "formula_len_all_1": 385, "len_all": 166480, "len_all_1": 63712, "len_abs": 1535, "len_title": 131, "len_sents": 53619, "len_sents_1": 30626, "n_sents": 379, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (conditional oral)", "L_title": 83, "L_abs": 1500, "title": "CyberHost: A One-stage Diffusion Framework for Audio-driven Talking Body Generation", "abs": "Diffusion-based video generation technology has advanced significantly, catalyzing a proliferation of research in human animation. While breakthroughs have been made in driving human animation through various modalities for portraits, most of current solutions for human body animation still focus on video-driven methods, leaving audio-driven taking body generation relatively underexplored. In this paper, we introduce CyberHost, a one-stage audio-driven talking body generation framework that addresses common synthesis degradations in half-body animation, including hand integrity, identity consistency, and natural motion.\nCyberHost's key designs are twofold. Firstly, the Region Attention Module (RAM) maintains a set of learnable, implicit, identity-agnostic latent features and combines them with identity-specific local visual features to enhance the synthesis of critical local regions. Secondly, the Human-Prior-Guided Conditions introduce more human structural priors into the model, reducing uncertainty in generated motion patterns and thereby improving the stability of the generated videos.\nTo our knowledge, CyberHost is the first one-stage audio-driven human diffusion model capable of zero-shot video generation for the human body. Extensive experiments demonstrate that CyberHost surpasses previous works in both quantitative and qualitative aspects. CyberHost can also be extended to video-driven and audio-video hybrid-driven scenarios, achieving similarly satisfactory results.", "keywords": [ "Audio-driven Human Animation.+Diffusion Model.+Generative Model.+Human Video Generation" ], "rating_list": [ 8, 10, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3, 4 ], "confidence_list": [ 3, 5, 4, 4, 4 ] }, { "paper_id": "vZK4pvHFd0", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Jie Guo", "gender": "Female", "institution": "Pengcheng Laboratory", "country": "CN", "position": "Assistant Professor" }, { "name": "Ronghui Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ruiqi Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xiu Li", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Yachao Zhang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zunnan Xu", "gender": "Not Specified", "institution": "TSINGHUA UNIVERSITY", "country": "CN", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 24, "n_ref": 48, "n_ref_all": 60, "n_fig": 3, "n_tab": 4, "L_tab": 1041, "n_element_tab": 149, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 690, "n_element_tab_1": 102, "formula_len_all": 500, "formula_len_all_1": 501, "len_all": 91610, "len_all_1": 53723, "len_abs": 1263, "len_title": 134, "len_sents": 30602, "len_sents_1": 25833, "n_sents": 200, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1296, "title": "HyDance: A Novel Hybrid Dance Generation Network with temporal and frequency features", "abs": "We propose HyDance, a diffusion network utilizing both the temporal and frequency-domain representations of dance motion sequences for music-driven dance motion generation. Existing dance generation methods primarily use temporal domain representations of dance motion in their networks, which often results in the network losing the sfrequency-domain characteristics of the dance. This manifests in overly smooth generated dance motion sequences, resulting in dance movements that lack dynamism. From an aesthetic perspective, such overly smooth movements are perceived as lacking expressiveness and the sense of power. To address this issue, we designed HyDance, which incorporates independent temporal feature encoders and frequency-domain feature encoders. The model employs a shared-weight hybrid feature encoder, enabling the complementary extraction of motion information from both domains. By introducing compact frequency-domain features into the dance generation framework, our method mitigates the oversmoothing problem in generated dance motion sequences and achieves improved spatial and temporal alignment in the generation results. Experiments show that our method generates more expressive dance movements than existing methods and achieves better alignment with the music beats.", "keywords": [ "Diffusion Models\uff0cMotion Generation" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "vYO7owSSHZ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Fei Wu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Shengyu Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "ZJU100 Young Professor" }, { "name": "Zheqi Lv", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zihao TANG", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 32, "n_ref": 59, "n_ref_all": 87, "n_fig": 7, "n_tab": 15, "L_tab": 4833, "n_element_tab": 765, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1786, "n_element_tab_1": 268, "formula_len_all": 289, "formula_len_all_1": 137, "len_all": 195616, "len_all_1": 72400, "len_abs": 1720, "len_title": 124, "len_sents": 49238, "len_sents_1": 31541, "n_sents": 402, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1753, "title": "LLM-Assisted Fast and Customized Model Generation: A Preliminary Exploration", "abs": "The rapid advancement of AI models has significantly impacted daily life, with Large Language Models (LLMs) playing a pivotal role in automating tasks and providing all-in-one solutions via API services. Meanwhile, there is a growing demand for private, resource-constrained, customizable, and high-performance models tailored to specific user needs. However, many users struggle to deploy these models due to limited resources or technical expertise. In this work, we try to address these challenges by focusing on two primary objectives: (1) to meet the specific needs of a broad range of users, and (2) to lower the barriers to AI model usage (\\textit{e.g.}, resource constraints, technical expertise) for most users. In our preliminary exploration, we introduce FLAME, a framework that determines and generates AI models based on data or task descriptions provided by users. While existing solutions rely on pre-built models or extensive finetuning, FLAME leverages LLMs (\\textit{e.g.}, GPT4-turbo) to capture data patterns and task features from user input, converting them into user requirements and structured metadata (\\textit{e.g.}, task type, model architecture, and classifier dimension). Then, FLAME uses them as guidance to generate customized models by hypernetworks. This approach significantly improves efficiency, achieving up to 270x faster model production compared to finetuning-based paradigms (e.g., all-parameter and LoRA fine-tuning) while maintaining comparable performance across various tasks. We validate the effectiveness of FLAME through comprehensive experiments on Natural Language Processing (NLP), Computer Vision (CV), and tabular datasets, demonstrating its ability to quickly deliver high-quality, customized models.", "keywords": [ "Customized Model Generation", "Hypernetworks", "Large Language Models" ], "rating_list": [ 5, 3, 6, 3, 3 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 3, 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3, 4 ] }, { "paper_id": "vYBzgwkwZb", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Changhua Meng", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Haoran Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Liangliang Shi", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Shuheng Shen", "gender": "unknown", "institution": "Ant Group", "country": "", "position": "Researcher" }, { "name": "Weiqiang Wang", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 39, "n_formula_1": 11, "n_ref_uni": 60, "n_ref": 111, "n_ref_all": 130, "n_fig": 6, "n_tab": 8, "L_tab": 5681, "n_element_tab": 972, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 975, "n_element_tab_1": 174, "formula_len_all": 3655, "formula_len_all_1": 1038, "len_all": 196439, "len_all_1": 78532, "len_abs": 1413, "len_title": 142, "len_sents": 53208, "len_sents_1": 36422, "n_sents": 407, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1420, "title": "BiQAP: Neural Bi-level Optimization-based Framework for Solving Quadratic Assignment Problems", "abs": "Quadratic Assignment Problem (QAP) has attracted lasting attention for its wide applications and computational challenge. Despite the rich literature in machine learning for QAP, most works often address the problem in the setting of image matching, whereby deep networks could play a vital role in extracting useful features for the subsequent matching. While its power on pure numerical QAP instances is limited in node embedding, often with a vanilla graph neural network. This paper tries to tap the potential of deep nets for QAP, specifically by modifying the input instance which is orthogonal to previous efforts. Specifically, we develop a bi-level unsupervised framework, where the inner optimization involves trying to solve the modified instance with entropic regularization that can be solved iteratively using the Sinkhorn algorithm without affecting backpropagation by truncating gradients during training. The outer minimization deals with the quadratic objective function of the original QAP. In particular, seeing the intractable scale of the most general form i.e. Lawler's QAP and the practical utility of the more efficient Koopmans-Beckmann QAP (KBQAP) form for solving other graph and combinatorial problems like TSP and graph edit distance, we embody our network on the KBQAP, and show its strong performance on various benchmarks in our experiments. Source code will be made publicly available.", "keywords": [ "Quadratic Assignment Problems", "Entropic Regularization", "Differential Gromov-Wasserstein Solver", "Unsupervised Learning" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 2, 4, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "vXSCD3ToCS", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Fandong Meng", "gender": "Male", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Jie Zhou", "gender": "Male", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Kelan Peng", "gender": "Not Specified", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Shaojun E", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Wenjuan Han", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yanke Yu", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhiwei Zhang", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 73, "n_ref_all": 97, "n_fig": 6, "n_tab": 9, "L_tab": 2509, "n_element_tab": 188, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1204, "n_element_tab_1": 131, "formula_len_all": 1801, "formula_len_all_1": 932, "len_all": 135880, "len_all_1": 59692, "len_abs": 1399, "len_title": 156, "len_sents": 37773, "len_sents_1": 26042, "n_sents": 302, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1407, "title": "DynST: Large-Scale Spatial-Temporal Dataset for Transferable Traffic Forecasting with Dynamic Road Networks", "abs": "In real-world traffic networks, it is common to encounter a shortage of historical data in the target region. Researchers often address this issue through transfer learning. However, transfer learning tasks in traffic prediction currently lack dedicated datasets and instead rely on datasets designed for non-transfer prediction tasks. The major drawback of these existing datasets is the adoption of a fixed network topology to model the real world's road networks. This does not align with reality and limits the model's transferability. To tackle this issue, we propose DynST, a dataset specifically designed for transfer learning tasks in traffic prediction, with a massive data volume of 20.35 billion, spanning 20 years and 9 regions. The key feature of DynST is evolving dynamic road network topology, which reflects the evolution of real road networks. Moreover, to address the shortcomings of the distance-based adjacency generation algorithm, we introduce a novel tree-based algorithm. Extensive experiments demonstrate that the adoption of DynST as the source dataset can significantly enhance the performance of the target region. The comparative experiment also validates that our adjacency matrix generation algorithm can lead to improved prediction accuracy. We believe that DynST, with rich spatial variation information, will facilitate research in the field of transfer traffic prediction.", "keywords": [ "Traffic Forecasting; Transfer Learning; Spatial-Temporal Data Mining; Dataset;" ], "rating_list": [ 5, 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "vXG7d2VlHU", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 10, "author_info_list": [ { "name": "Ao Qu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Dingyi Zhuang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jinhua Zhao", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Shenhao Wang", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Wei Ma", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "Yihong Tang", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Researcher" }, { "name": "Yunhan Zheng", "gender": "unknown", "institution": "Singapore-MIT Alliance for Research and Technology Centre", "country": "SG", "position": "Postdoc" }, { "name": "Zhan Zhao", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Zhaofeng Wu", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zhaokai Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Intern" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 41, "n_ref": 75, "n_ref_all": 95, "n_fig": 9, "n_tab": 5, "L_tab": 1559, "n_element_tab": 249, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 879, "n_element_tab_1": 145, "formula_len_all": 272, "formula_len_all_1": 824, "len_all": 172728, "len_all_1": 65084, "len_abs": 1870, "len_title": 174, "len_sents": 39071, "len_sents_1": 29170, "n_sents": 265, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 125, "L_abs": 1879, "title": "Sparkle: Mastering Basic Spatial Capabilities in Vision Language Models Elicits Generalization to Composite Spatial Reasoning", "abs": "Vision language models (VLMs) have demonstrated impressive performance across a wide range of downstream tasks. However, their proficiency in spatial reasoning remains limited, despite its crucial role in tasks involving navigation and interaction with physical environments. \nSpecifically, much of the spatial reasoning in these tasks occurs in two-dimensional (2D) environments, and our evaluation reveals that state-of-the-art VLMs frequently generate implausible and incorrect responses to composite spatial reasoning problems, including simple pathfinding tasks that humans can solve effortlessly at a glance. \nTo address this, we explore an effective approach to enhance 2D spatial reasoning within VLMs by training the model on basic spatial capabilities.\nWe begin by disentangling the key components of 2D spatial reasoning: direction comprehension, distance estimation, and localization.\nOur central hypothesis is that mastering these basic spatial capabilities can significantly enhance a model's performance on composite spatial tasks requiring advanced spatial understanding and combinatorial problem-solving.\nTo investigate this hypothesis, we introduce Sparkle,\na framework that fine-tunes VLMs on these three basic spatial capabilities by synthetic data generation and targeted supervision to form an instruction dataset for each capability.\nOur experiments demonstrate that VLMs fine-tuned with Sparkle achieve significant performance gains, not only in the basic tasks themselves but also in generalizing to composite and out-of-distribution spatial reasoning tasks (e.g., improving from 13.5% to 40.0% on the shortest path problem). These findings underscore the effectiveness of mastering basic spatial capabilities in enhancing composite spatial problem-solving, offering insights into systematic strategies for improving VLMs' spatial reasoning capabilities.", "keywords": [ "spatial reasoning", "vision language models", "multimodal large language models" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "vWRwdmA3wU", "primary_area": "applications to neuroscience & cognitive science", "n_author": 7, "author_info_list": [ { "name": "Christopher J Cueva", "gender": "Not Specified", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Earl K Miller", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Guangyu Robert Yang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Markus Siegel", "gender": "unknown", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "", "position": "Full Professor" }, { "name": "Moufan Li", "gender": "unknown", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Nathan Cloos", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Scott L Brincat", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" } ], "n_formula": 20, "n_formula_1": 7, "n_ref_uni": 21, "n_ref": 71, "n_ref_all": 116, "n_fig": 20, "n_tab": 2, "L_tab": 730, "n_element_tab": 133, "n_fig_1": 18, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1436, "formula_len_all_1": 421, "len_all": 141402, "len_all_1": 56221, "len_abs": 1847, "len_title": 122, "len_sents": 54501, "len_sents_1": 27991, "n_sents": 394, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1858, "title": "Differentiable Optimization of Similarity Scores Between Models and Brains", "abs": "How do we know if two systems - biological or artificial - process information in a similar way? Similarity measures such as linear regression, Centered Kernel Alignment (CKA), Normalized Bures Similarity (NBS), and angular Procrustes distance, are often used to quantify this similarity. However, it is currently unclear what drives high similarity scores and even what constitutes a \"good\" score. Here, we introduce a novel tool to investigate these questions by differentiating through similarity measures to directly maximize the score. Surprisingly, we find that high similarity scores do not guarantee encoding task-relevant information in a manner consistent with neural data; and this is particularly acute for CKA and even some variations of cross-validated and regularized linear regression. We find no consistent threshold for a good similarity score - it depends on both the measure and the dataset. In addition, synthetic datasets optimized to maximize similarity scores initially learn the highest variance principal component of the target dataset, but some methods like angular Procrustes capture lower variance dimensions much earlier than methods like CKA. To shed light on this, we mathematically derive the sensitivity of CKA, angular Procrustes, and NBS to the variance of principal component dimensions, and explain the emphasis CKA places on high variance components. Finally, by jointly optimizing multiple similarity measures, we characterize their allowable ranges and reveal that some similarity measures are more constraining than others. While current measures offer a seemingly straightforward way to quantify the similarity between neural systems, our work underscores the need for careful interpretation. We hope the tools we developed will be used by practitioners to better understand current and future similarity measures.", "keywords": [ "similarity measures", "representational alignment", "procrustes distance", "centered kernel alignment", "linear regression" ], "rating_list": [ 6, 6, 8, 5 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "vWR3KuiQur", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Chenlin Meng", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Enze Xie", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Researcher" }, { "name": "Jun-Yan Zhu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Muyang Li", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Song Han", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Tianle Cai", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Xiuyu Li", "gender": "Not Specified", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Yujun Lin", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zhekai Zhang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "junxian guo", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 22, "n_formula_1": 12, "n_ref_uni": 71, "n_ref": 152, "n_ref_all": 175, "n_fig": 18, "n_tab": 7, "L_tab": 1127, "n_element_tab": 199, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 8, "n_element_tab_1": 2, "formula_len_all": 1717, "formula_len_all_1": 773, "len_all": 203592, "len_all_1": 62386, "len_abs": 544, "len_title": 125, "len_sents": 44774, "len_sents_1": 28783, "n_sents": 393, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 77, "L_abs": 1865, "title": "SVDQuant: Absorbing Outliers by Low-Rank Component for 4-Bit Diffusion Models", "abs": "Diffusion models have been proven highly effective at generating high-quality images. However, as these models grow larger, they require significantly more memory and suffer from higher latency, posing substantial challenges for deployment. In this work, we aim to accelerate diffusion models by quantizing their weights and activations to 4 bits. At such an aggressive level, both weights and activations are highly sensitive, where conventional post-training quantization methods for large language models like smoothing become insufficient. To overcome this limitation, we propose SVDQuant, a new 4-bit quantization paradigm. Different from smoothing which redistributes outliers between weights and activations, our approach absorbs these outliers using a low-rank branch. We first consolidate the outliers by shifting them from activations to weights, then employ a high-precision low-rank branch to take in the weight outliers with Singular Value Decomposition (SVD). This process eases the quantization on both sides. However, naively running the low-rank branch independently incurs significant overhead due to extra data movement of activations, negating the quantization speedup. To address this, we co-design an inference engine LoRunner that fuses the kernels of the low-rank branch into those of the low-bit branch to cut off redundant memory access. It can also seamlessly support off-the-shelf low-rank adapters (LoRAs) without the need for re-quantization. Extensive experiments on SDXL, PixArt-$\\Sigma$, and FLUX.1 validate the effectiveness of SVDQuant in preserving image quality. We reduce the memory usage for the 12B FLUX.1 models by 3.6\u00d7, achieving 3.5\u00d7 speedup over the 4-bit weight-only quantized baseline on a 16GB RTX-4090 GPU, paving the way for more interactive applications on PCs. We will release the code and models upon publication.", "keywords": [ "Quantization", "Diffusion Models", "Efficiency", "Acceleration" ], "rating_list": [ 6, 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 2, 4, 3 ], "presentation_list": [ 2, 4, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 3, 4 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "vW6rsXAGrz", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Lee Carlin", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "", "position": "PhD student" }, { "name": "Yuval Benjamini", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 20, "n_ref": 25, "n_ref_all": 32, "n_fig": 2, "n_tab": 2, "L_tab": 1087, "n_element_tab": 229, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1087, "n_element_tab_1": 229, "formula_len_all": 396, "formula_len_all_1": 396, "len_all": 78622, "len_all_1": 78622, "len_abs": 1027, "len_title": 117, "len_sents": 28130, "len_sents_1": 28130, "n_sents": 216, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1031, "title": "CardiCat: a Variational Autoencoder for High-Cardinality Tabular Data", "abs": "High-cardinality categorical features are a common characteristic of mixed-type tabular datasets. Existing generative model architectures struggle to learn the complexities of such data at scale, primarily due to the difficulty of parameterizing the categorical features. In this paper, we present a general variational autoencoder model, CardiCat, that can accurately fit imbalanced high-cardinality and heterogeneous tabular data. Our method substitutes one-hot encoding with regularized dual encoder-decoder embedding layers, which are jointly learned. This approach enables us to use embeddings that depend also on the other covariates, leading to a compact and homogenized parameterization of categorical features. Our model employs a considerably smaller trainable parameter space than competing methods, enabling learning at a large scale. CardiCat generates high-quality synthetic data that better represent high-cardinality and imbalanced features compared to competing VAE models for multiple real and simulated datasets.", "keywords": [ "embedding", "VAE", "tabular", "regularization", "high-cardinality", "categorical", "imbalance", "mixed", "heterogeneous", "layers", "Generative", "model" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 1, 3, 1 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "vVxeFSR4fU", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Jiachen Jiang", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Jinxin Zhou", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Zhihui Zhu", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Assistant Professor" } ], "n_formula": 44, "n_formula_1": 12, "n_ref_uni": 61, "n_ref": 98, "n_ref_all": 151, "n_fig": 24, "n_tab": 1, "L_tab": 402, "n_element_tab": 37, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 402, "n_element_tab_1": 37, "formula_len_all": 3120, "formula_len_all_1": 681, "len_all": 197789, "len_all_1": 67663, "len_abs": 2215, "len_title": 130, "len_sents": 60639, "len_sents_1": 32991, "n_sents": 429, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 2229, "title": "Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity", "abs": "Analyzing the similarity of internal representations within and across different models has been an important technique for understanding the behavior of deep neural networks. Most existing methods for analyzing the similarity between representations of high dimensions, such as those based on Centered Kernel Alignment (CKA), rely on statistical properties of the representations for a set of data points. In this paper, we focus on transformer models and study the similarity of representations between the hidden layers of individual transformers. In this context, we show that a simple sample-wise cosine similarity metric is capable of capturing the similarity and aligns with the complicated CKA. Our experimental results on common transformers reveal that representations across layers are positively correlated, with similarity increasing when layers get closer. We provide a theoretical justification for this phenomenon under the geodesic curve assumption for the learned transformer, a property that may approximately hold for residual networks. We then show that an increase in representation similarity implies an increase in predicted probability when directly applying the last-layer classifier to any hidden layer representation. This offers a justification for {\\it saturation events}, where the model's top prediction remains unchanged across subsequent layers, indicating that the shallow layer has already learned the necessary knowledge. We then propose an aligned training method to improve the effectiveness of shallow layer by enhancing the similarity between internal representations, with trained models that enjoy the following properties: (1) more early saturation events, (2) layer-wise accuracies monotonically increase and reveal the minimal depth needed for the given task, (3) when served as multi-exit models, they achieve on-par performance with standard multi-exit architectures which consist of additional classifiers designed for early exiting in shallow layers. To our knowledge, our work is the first to show that one common classifier is sufficient for multi-exit models. We conduct experiments on both vision and NLP tasks to demonstrate the performance of the proposed aligned training.", "keywords": [ "Representation Similarity", "Saturation Event", "Early Exit" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 2, 3, 3 ] }, { "paper_id": "vVlNBaiLdN", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 1, "author_info_list": [ { "name": "Moritz Glaser", "gender": "unknown", "institution": "Charit\u00e9 - Universit\u00e4tsmedizin Berlin", "country": "DE", "position": "Undergrad student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 60, "n_ref_all": 81, "n_fig": 10, "n_tab": 2, "L_tab": 680, "n_element_tab": 94, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 4902, "n_element_tab_1": 162, "formula_len_all": 93, "formula_len_all_1": 0, "len_all": 213445, "len_all_1": 49971, "len_abs": 1698, "len_title": 121, "len_sents": 41745, "len_sents_1": 22277, "n_sents": 255, "n_sents_1": 127, "venue": "ICLR2025", "decision": "Reject", "L_title": 124, "L_abs": 1705, "title": "ESMGain: Effective and Efficient Prediction of Mutation\u2019s functional Effect via ESM2 Transfer Learning and robust Benchmarks", "abs": "Functional effect prediction of mutations, especially for properties like catalytic activity, holds greater significance for clinicians and protein engineers than traditional pathogenicity predictions. Recent approaches leveraging static ESM1 embeddings or multimodal features (e.g. embeddings, structures, and evolutionary data) either (1) fall short in accuracy or (2) involve complex preprocessing pipelines. Moreover, functional effect prediction suffers from (3) a lack of standardized datasets and metrics for robust benchmarking. We address these challenges by systematically optimizing ESM2-based functional effect prediction: Through extensive ablation studies, we demonstrate that fine-tuning significantly outperforms static embeddings, scaling laws for model size are non-transferable and LoRA matches full fine-tuning performance, deviating from trends observed in natural language processing. Our framework, ESM-Effect, fine-tunes 35M ESM2 layers with an inductive bias regression head achieving state-of-the-art performance. It slightly surpasses multimodal competitor PreMode indicating redundancy in structural and evolutionary features. We further propose a benchmarking framework featuring robst test datasets and strategies, and the relative Bin-Mean Error (rBME), as a metric designed to emphasize prediction accuracy in challenging, non-clustered, and rare gain-of-function regions. rBME better reflects model performance compared to commonly used Spearman\u2019s rho, as evidenced by improved plot-based analyses. As ESM-Effect exhibits mixed transferability to different unseen mutational regions, we identify multiple areas for improvement such as finer-grained pretraining strategies.", "keywords": [ "protein", "language model", "deep learning", "biology", "gain of function", "enzyme" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "vVhZh9ZpIM", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "David Lopez-Paz", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Elvis Dohmatob", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Mohammad Pezeshki", "gender": "Male", "institution": "Facebook", "country": "CA", "position": "Researcher" }, { "name": "Pascal Vincent", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "Adjunct Professor" }, { "name": "Reza Bayat", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "MS student" } ], "n_formula": 63, "n_formula_1": 15, "n_ref_uni": 57, "n_ref": 106, "n_ref_all": 127, "n_fig": 7, "n_tab": 4, "L_tab": 2958, "n_element_tab": 94, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1482, "n_element_tab_1": 101, "formula_len_all": 4403, "formula_len_all_1": 795, "len_all": 199656, "len_all_1": 65843, "len_abs": 876, "len_title": 116, "len_sents": 56203, "len_sents_1": 29947, "n_sents": 488, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 884, "title": "The Pitfalls of Memorization: When Memorization Hurts Generalization", "abs": "Neural networks often learn simple explanations that fit the majority of the data while memorizing exceptions that deviate from these explanations. This leads to poor generalization if the learned explanations are spurious. In this work, we formalize $\\textit{the interplay between memorization and generalization}$, showing that spurious correlations would particularly lead to poor generalization when are combined with memorization. Memorization can reduce the training loss to zero, leaving no incentive for learning robust, generalizable patterns. To address this issue, we introduce $\\textit{memorization-aware training}$ (MAT). MAT leverages the flip side of memorization by using held-out predictions to shift a model's logits, guiding it towards learning robust patterns that remain invariant from training to test, thereby enhancing generalization under distribution shifts.", "keywords": [ "Memorization", "Generalization", "Spurious Correlations" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 1 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 4, 2, 3, 5 ] }, { "paper_id": "vVVtTVIR5O", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Chaoquan Jiang", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Jiaming Zhang", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Researcher" }, { "name": "Jinlin Xiao", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Jitao Sang", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Zhiyu Lin", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yang Yunfan", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 33, "n_ref": 60, "n_ref_all": 67, "n_fig": 3, "n_tab": 5, "L_tab": 2372, "n_element_tab": 431, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 5063, "n_element_tab_1": 468, "formula_len_all": 433, "formula_len_all_1": 433, "len_all": 114203, "len_all_1": 65817, "len_abs": 1627, "len_title": 103, "len_sents": 31584, "len_sents_1": 27102, "n_sents": 230, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1651, "title": "Debiasing Vison-Language Models with Text-Only Training", "abs": "Pre-trained vision-language models (VLMs), such as CLIP, have exhibited remarkable performance across various downstream tasks by aligning text and images in a unified embedding space. However, due to the imbalanced distribution of pre-trained datasets, CLIP suffers from the bias problem in real-world applications. Existing debiasing methods struggle to obtain sufficient image samples for minority groups and incur high costs for group labeling. To address the limitations, we propose a **T**ext-**O**nly **D**ebiasing framework called **TOD**, leveraging a text-as-image training paradigm to mitigate visual biases. Specifically, this approach repurposes the text encoder to function as an image encoder, thereby eliminating the need for image data. Simultaneously, it utilizes a large language model (LLM) to generate a balanced text dataset, which is then used for prompt tuning. However, we observed that the model overfits to the text modality because label names, serving as supervision signals, appear explicitly in the texts. To address this issue, we further introduce a Multi-Target Prediction (MTP) task that motivates the model to focus on complex contexts and distinguish between target and biased information. Extensive experiments on the Waterbirds and CelebA datasets show that our method significantly improves group robustness, achieving state-of-the-art results among image-free methods and even competitive performance compared to image-supervised methods. Furthermore, the proposed method can be adapted to challenging scenarios with multiple or unknown bias attributes, demonstrating its strong generalization and robustness.", "keywords": [ "Vison Language Models", "Group Robustness", "Fairness", "CLIP" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "vVHc8bGRns", "primary_area": "datasets and benchmarks", "n_author": 18, "author_info_list": [ { "name": "Bing Han", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "MS student" }, { "name": "Defu Lian", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Guorui Zhou", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Principal Researcher" }, { "name": "Han Li", "gender": "Male", "institution": "Kuaishou", "country": "", "position": "Researcher" }, { "name": "Hongning Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Kai Zheng", "gender": "Male", "institution": "Kuaishou", "country": "CN", "position": "Researcher" }, { "name": "Kun Gai", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Instructor" }, { "name": "Kuo Cai", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "MS student" }, { "name": "Na Mou", "gender": "Female", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Qi Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Rui Huang", "gender": "Male", "institution": "Kuaishou", "country": "", "position": "Researcher" }, { "name": "Wentian Bao", "gender": "Male", "institution": "Columbia University", "country": "", "position": "MS student" }, { "name": "Wuchao Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yanan Niu", "gender": "Male", "institution": "Kuaishou Technology", "country": "", "position": "Lead of ranking model" }, { "name": "Yang Song", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Researcher" }, { "name": "Yuan Chai", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "", "position": "Researcher" }, { "name": "Yun En Yu", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Hui Yiqun", "gender": "unknown", "institution": "Kuaishou", "country": "", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 50, "n_ref": 97, "n_ref_all": 124, "n_fig": 3, "n_tab": 7, "L_tab": 5320, "n_element_tab": 266, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 866, "n_element_tab_1": 87, "formula_len_all": 193, "formula_len_all_1": 0, "len_all": 169168, "len_all_1": 66519, "len_abs": 1770, "len_title": 104, "len_sents": 42790, "len_sents_1": 31280, "n_sents": 377, "n_sents_1": 274, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1999, "title": "RecFlow: An Industrial Full Flow Recommendation Dataset", "abs": "Industrial recommendation systems (RS) rely on the multi-stage pipeline to balance effectiveness and efficiency when delivering items from a vast corpus to users. Existing RS benchmark datasets primarily focus on the exposure space, where novel RS algorithms are trained and evaluated. However, when these algorithms transition to real-world industrial RS, they face a critical challenge: handling unexposed items\u2014a significantly larger space than the exposed one. This discrepancy profoundly impacts their practical performance. Additionally, these algorithms often overlook the intricate interplay between multiple RS stages, resulting in suboptimal overall system performance. To address this issue, we introduce RecFlow\u2014an industrial full-flow recommendation dataset designed to bridge the gap between offline RS benchmarks and the real online environment. Unlike existing datasets, RecFlow includes samples not only from the exposure space but also unexposed items filtered at each stage of the RS funnel. Our dataset comprises 38M interactions from 42K users across nearly 9M items with additional 1.9B stage samples collected from 9.3M online requests over 37 days and spanning 6 stages. Leveraging the RecFlow dataset, we conduct courageous exploration experiments, showcasing its potential in designing new algorithms to enhance effectiveness by incorporating stage-specific samples. Some of these algorithms have already been deployed online, consistently yielding significant gains. We propose RecFlow as the first comprehensive benchmark dataset for the RS community, supporting research on designing algorithms at any stage, study of selection bias, debiased algorithms, multi-stage consistency and optimality, multi-task recommendation, and user behavior modeling. The RecFlow dataset, along with the corresponding source code, is publicly available at \\textcolor{red}{\\url{https://github.com/RecFlow-ICLR/RecFlow}}. The dataset is licensed under CC-BY-NC-SA-4.0 International License.", "keywords": [ "recommendation system", "recommendation dataset" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 4, 3, 4 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "vVCHWVBsLH", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Christoph Hertrich", "gender": "unknown", "institution": "University of Technology Nuremberg", "country": "DE", "position": "Assistant Professor" }, { "name": "Marie-Charlotte Brandenburg", "gender": "unknown", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "Postdoc" }, { "name": "Moritz Leo Grillo", "gender": "Male", "institution": "Technische Universit\u00e4t Berlin", "country": "DE", "position": "PhD student" } ], "n_formula": 66, "n_formula_1": 16, "n_ref_uni": 38, "n_ref": 98, "n_ref_all": 115, "n_fig": 7, "n_tab": 5, "L_tab": 173, "n_element_tab": 16, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6761, "formula_len_all_1": 1361, "len_all": 218869, "len_all_1": 75981, "len_abs": 1218, "len_title": 53, "len_sents": 76501, "len_sents_1": 33736, "n_sents": 693, "n_sents_1": 278, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 53, "L_abs": 1316, "title": "Decomposition Polyhedra of Piecewise Linear Functions", "abs": "In this paper we contribute to the frequently studied question of how to decompose a continuous piecewise linear (CPWL) function into a difference of two convex CPWL functions. Every CPWL function has infinitely many such decompositions, but for applications in optimization and neural network theory, it is crucial to find decompositions with as few linear pieces as possible. This is a highly challenging problem, as we further demonstrate by disproving a recently proposed approach by Tran and Wang [Minimal representations of tropical rational functions. Algebraic Statistics, 15(1):27\u201359, 2024]. To make the problem more tractable, we propose to fix an underlying polyhedral complex determining the possible locus of nonlinearity. Under this assumption, we prove that the set of decompositions forms a polyhedron that arises as intersection of two translated cones. We prove that irreducible decompositions correspond to the bounded faces of this polyhedron and minimal solutions must be vertices. We then identify cases with a unique minimal decomposition, and illustrate how our insights have consequences in the theory of submodular functions. Finally, we improve upon previous constructions of neural networks for a given convex CPWL function and apply our framework to obtain results in the nonconvex case.", "keywords": [ "Piecewise Linear Functions", "Polyhedral Geometry", "Minimal Convex Decompositions", "Submodular Functions", "Neural Networks" ], "rating_list": [ 8, 8, 8, 5 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "vTdwuKUc5Z", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Guihai Chen", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Jinjin Gu", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "Instructor" }, { "name": "Linghe Kong", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaokang Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University, China", "country": "CN", "position": "Professor" }, { "name": "Xin Yuan", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Associate Professor" }, { "name": "Yulun Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Zheng Chen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 53, "n_ref": 150, "n_ref_all": 167, "n_fig": 9, "n_tab": 5, "L_tab": 3342, "n_element_tab": 535, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 2558, "n_element_tab_1": 327, "formula_len_all": 100, "formula_len_all_1": 126, "len_all": 153290, "len_all_1": 68808, "len_abs": 1242, "len_title": 97, "len_sents": 35940, "len_sents_1": 27347, "n_sents": 369, "n_sents_1": 268, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1256, "title": "Image Super-Resolution with Text Prompt Diffusion", "abs": "Image super-resolution (SR) methods typically model degradation to improve reconstruction accuracy in complex and unknown degradation scenarios. However, extracting degradation information from low-resolution images is challenging, which limits the model performance. To boost image SR performance, one feasible approach is to introduce additional priors. Inspired by advancements in multi-modal methods and text prompt image processing, we introduce text prompts to image SR to provide degradation priors. Specifically, we first design a text-image generation pipeline to integrate text into the SR dataset through the text degradation representation and degradation model. The text representation applies a discretization manner based on the binning method to describe the degradation abstractly. This method maintains the flexibility of the text and is user-friendly. Meanwhile, we propose the PromptSR to realize the text prompt SR. The PromptSR utilizes the pre-trained language model (*e.g.*, T5 or CLIP) to enhance restoration. We train the PromptSR on the generated text-image dataset. Extensive experiments indicate that introducing text prompts into SR, yields excellent results on both synthetic and real-world images. The code will be released.", "keywords": [ "Image Super-Resolution", "Text Prompt", "Diffusion Model" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "vTRWu9zaWo", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Hideaki Iiduka", "gender": "Male", "institution": "Meiji University", "country": "JP", "position": "Full Professor" }, { "name": "Naoki Sato", "gender": "Male", "institution": "Meiji University", "country": "JP", "position": "MS student" } ], "n_formula": 74, "n_formula_1": 19, "n_ref_uni": 84, "n_ref": 192, "n_ref_all": 233, "n_fig": 15, "n_tab": 5, "L_tab": 504, "n_element_tab": 57, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 28971, "formula_len_all_1": 22385, "len_all": 241249, "len_all_1": 92379, "len_abs": 881, "len_title": 156, "len_sents": 55521, "len_sents_1": 32523, "n_sents": 464, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 894, "title": "Using Stochastic Gradient Descent to Smooth Nonconvex Functions: Analysis of Implicit Graduated Optimization", "abs": "The graduated optimization approach is a heuristic method for finding global optimal solutions for nonconvex functions by using a function smoothing operation with stochastic noise. We show that stochastic noise in stochastic gradient descent (SGD) has the effect of smoothing the objective function, the degree of which is determined by the learning rate, batch size, and variance of the stochastic gradient. Using this finding, we propose and analyze a new graduated optimization algorithm that varies the degree of smoothing by varying the learning rate and batch size, and provide experimental results on image classification tasks with ResNets that support our theoretical findings. We further show that there is an interesting correlation between the degree of smoothing by SGD's stochastic noise, the well-studied ``sharpness'' indicator, and the generalization performance of the model.", "keywords": [ "deep learning theory", "degree of smoothing", "generalizability", "graduated optimization", "SGD", "sharpness", "smoothing property", "stochastic noise" ], "rating_list": [ 5, 5, 3, 3, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 1, 3 ], "confidence_list": [ 3, 3, 3, 4, 3 ] }, { "paper_id": "vTLLyVCsrD", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Minghui Zhu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Full Professor" }, { "name": "Shicheng Liu", "gender": "unknown", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 130, "n_formula_1": 38, "n_ref_uni": 46, "n_ref": 109, "n_ref_all": 131, "n_fig": 2, "n_tab": 7, "L_tab": 808, "n_element_tab": 54, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 26786, "formula_len_all_1": 4495, "len_all": 219983, "len_all_1": 82987, "len_abs": 1602, "len_title": 120, "len_sents": 57353, "len_sents_1": 34990, "n_sents": 524, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1616, "title": "Improving Generalization of Meta Reinforcement Learning via Explanation", "abs": "Meta reinforcement learning learns a meta-prior (e.g., meta-policy) from a set of training tasks, such that the learned meta-prior can efficiently adapt to all the tasks in a task distribution. However, it has been observed in literature that the learned meta-prior usually has imbalanced generalization, i.e., it adapts well to some tasks but adapts poorly to some other tasks. This paper aims to explain why certain tasks are poorly adapted and, more importantly, use this explanation to improve generalization. Our methodology has two parts. The first part identifies ``critical\" training tasks that are most important to achieve good performance on those poorly-adapted tasks. An explanation of the poor generalization is that the meta-prior does not pay enough attention to the critical training tasks. To improve generalization, the second part formulates a bi-level optimization problem where the upper level learns how to augment the critical training tasks such that the meta-prior can pay more attention to the critical tasks, and the lower level computes the meta-prior distribution corresponding to the current augmentation. We propose an algorithm to solve the bi-level optimization problem and theoretically guarantee that (1) the algorithm converges at the rate of $O(1/\\sqrt{K})$, (2) the learned augmentation makes the meta-prior focus more on the critical training tasks, and (3) the generalization improves after the task augmentation. We use two real-world experiments and three MuJoCo experiments to show that our algorithm improves the generalization and outperforms state-of-the-art baselines.", "keywords": [ "explainable meta reinforcement learning; meta reinforcement learning generalization" ], "rating_list": [ 3, 6, 3, 3, 6 ], "soundness_list": [ 1, 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 1, 1 ], "contribution_list": [ 2, 3, 2, 1, 3 ], "confidence_list": [ 3, 2, 4, 4, 4 ] }, { "paper_id": "vSrBzCzg4G", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Davide Ghilardi", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Intern" }, { "name": "Federico Belotti", "gender": "unknown", "institution": "University of Milan - Bicocca", "country": "IT", "position": "Researcher" }, { "name": "Marco Molinari", "gender": "unknown", "institution": "LSE.AI", "country": "GB", "position": "Founder / Lead" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 30, "n_ref": 457, "n_ref_all": 494, "n_fig": 8, "n_tab": 3, "L_tab": 2250, "n_element_tab": 115, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 499, "n_element_tab_1": 9, "formula_len_all": 654, "formula_len_all_1": 439, "len_all": 138517, "len_all_1": 57723, "len_abs": 769, "len_title": 84, "len_sents": 35708, "len_sents_1": 26585, "n_sents": 227, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 763, "title": "Efficient Training of Sparse Autoencoders for Large Language Models via Layer Clustering", "abs": "Sparse Autoencoders (SAEs) have recently been employed as an unsupervised approach for understanding the inner workings of Large Language Models (LLMs). They reconstruct the model\u2019s activations with a sparse linear combination of interpretable features. However, training SAEs is computationally intensive, especially as models grow in size and complexity. To address this challenge, we propose a novel training strategy that reduces the number of trained SAEs from one per layer to one for a given group of contiguous layers. Our experimental results on Pythia 160M highlight a 3x speedup without compromising the reconstruction quality and performance on downstream tasks. Therefore, layer clustering presents an efficient approach to train SAEs in modern LLMs.", "keywords": [ "Sparse Autoencoders (SAEs)", "Meta Learning", "Mechanistic Interpretability", "Large Language Models (LLMs)" ], "rating_list": [ 6, 3, 6, 3, 3 ], "soundness_list": [ 1, 3, 1, 2, 2 ], "presentation_list": [ 1, 2, 3, 2, 3 ], "contribution_list": [ 3, 1, 1, 2, 1 ], "confidence_list": [ 4, 4, 4, 3, 5 ] }, { "paper_id": "vRvVVb0NAz", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Hongkang Li", "gender": "unknown", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "Postdoc" }, { "name": "Meng Wang", "gender": "Female", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "Associate Professor" }, { "name": "Pin-Yu Chen", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Shuai Zhang", "gender": "Male", "institution": "New Jersey Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Sijia Liu", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Assistant Professor" }, { "name": "Yihua Zhang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" } ], "n_formula": 133, "n_formula_1": 21, "n_ref_uni": 47, "n_ref": 102, "n_ref_all": 138, "n_fig": 3, "n_tab": 7, "L_tab": 1360, "n_element_tab": 113, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 13, "n_element_tab_1": 2, "formula_len_all": 16433, "formula_len_all_1": 2167, "len_all": 219520, "len_all_1": 77189, "len_abs": 1519, "len_title": 157, "len_sents": 50036, "len_sents_1": 33564, "n_sents": 520, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 109, "L_abs": 1527, "title": "When is Task Vector Provably Effective for Model Editing? A Generalization Analysis of Nonlinear Transformers", "abs": "Task arithmetic refers to editing the pre-trained model by adding a weighted sum of task vectors, each of which is the weight update from the pre-trained model to fine-tuned models for certain tasks. This approach recently gained attention as a computationally efficient inference method for model editing, e.g., multi-task learning, forgetting, and out-of-domain generalization capabilities. However, the theoretical understanding of why task vectors can execute various conceptual operations remains limited, due to the highly non-convexity of training Transformer-based models. To the best of our knowledge, this paper provides the first theoretical characterization of the generalization guarantees of task vector methods on nonlinear Transformers. We consider a conceptual learning setting, where each task is a binary classification problem based on a discriminative pattern. We theoretically prove the effectiveness of task addition in simultaneously learning a set of irrelevant or aligned tasks, as well as the success of task negation in unlearning one task from irrelevant or contradictory tasks. Moreover, we prove the proper selection of linear coefficients for task arithmetic to achieve guaranteed generalization to out-of-domain tasks. All of our theoretical results hold for both dense-weight parameters and their low-rank approximations. Although established in a conceptual setting, our theoretical findings were validated on a practical machine unlearning task using the large language model Phi-1.5 (1.3B).", "keywords": [ "Task arithmetic", "generalization", "nonlinear Transformers", "deep learning theory", "machine unlearning" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 3, 2, 3, 2 ] }, { "paper_id": "vR2MWaZ3MG", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Mihaela van der Schaar", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Nabeel Seedat", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 6, "n_ref_uni": 53, "n_ref": 106, "n_ref_all": 131, "n_fig": 13, "n_tab": 16, "L_tab": 6866, "n_element_tab": 189, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1513, "n_element_tab_1": 63, "formula_len_all": 793, "formula_len_all_1": 517, "len_all": 278038, "len_all_1": 71098, "len_abs": 1409, "len_title": 122, "len_sents": 95680, "len_sents_1": 34957, "n_sents": 689, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1424, "title": "Matchmaker: Schema Matching with self-improving compositional LLM programs", "abs": "Schema matching -- the task of finding matches between attributes across disparate data sources with different tables and hierarchies -- is critical for creating interoperable machine learning (ML)-ready data. Addressing this fundamental data-centric problem has wide implications, especially in domains like healthcare, finance and e-commerce --- but also has the potential to benefit ML models more generally, by increasing the data available for ML model training. However, schema matching is a challenging ML task due to structural/hierarchical and semantic heterogeneity between different schemas. Previous ML approaches to automate schema matching have either required significant labeled data for model training, which is often unrealistic, or suffer from poor zero-shot performance. To this end, we propose Matchmaker - a compositional language model program for schema matching, comprised of candidate generation, refinement and confidence scoring. Matchmaker also self-improves in a zero-shot manner without the need for labeled demonstrations via a novel optimization approach, which constructs synthetic in-context demonstrations to guide the language model's reasoning process. Empirically, we demonstrate on real-world medical schema matching benchmarks that Matchmaker outperforms previous ML-based approaches, highlighting its potential to accelerate data integration and interoperability of ML-ready data.", "keywords": [ "schema matching", "data-centric AI", "Large Language Models", "healthcare" ], "rating_list": [ 8, 3, 5, 5 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 1, 4, 4 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "vQxqcVGrhR", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Guibao Shen", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Haodong LI", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Intern" }, { "name": "Jing He", "gender": "Female", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Weichao Qiu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Ying-Cong Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Yingjie CAI", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "huyongzhe", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Undergrad student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 35, "n_ref": 128, "n_ref_all": 146, "n_fig": 9, "n_tab": 2, "L_tab": 497, "n_element_tab": 72, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 372, "n_element_tab_1": 49, "formula_len_all": 618, "formula_len_all_1": 419, "len_all": 122664, "len_all_1": 58857, "len_abs": 0, "len_title": 136, "len_sents": 34304, "len_sents_1": 28264, "n_sents": 226, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1337, "title": "DisEnvisioner: Disentangled and Enriched Visual Prompt for Customized Image Generation", "abs": "In the realm of image generation, creating customized images from visual prompt with additional textual instruction emerges as a promising endeavor. However, existing methods, both tuning-based and tuning-free, struggle with interpreting the subject-essential attributes from the visual prompt. This leads to subject-irrelevant attributes infiltrating the generation process, ultimately compromising the personalization quality in both editability and ID preservation. In this paper, we present $\\textbf{DisEnvisioner}$, a novel approach for effectively extracting and enriching the subject-essential features while filtering out -irrelevant information, enabling exceptional customization performance, in a $\\textbf{tuning-free}$ manner and using only $\\textbf{a single image}$. Specifically, the feature of the subject and other irrelevant components are effectively separated into distinctive visual tokens, enabling a much more accurate customization. Aiming to further improving the ID consistency, we enrich the disentangled features, sculpting them into a more granular representation. Experiments demonstrate the superiority of our approach over existing methods in instruction response (editability), ID consistency, inference speed, and the overall image quality, highlighting the effectiveness and efficiency of DisEnvisioner.", "keywords": [ "Visual Disentanglement and Enrichment", "Zero-shot Customization", "Text-to-Image Generation" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "vQhn4wrQ6j", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Benjamin Muller", "gender": "Male", "institution": "Meta", "country": "FR", "position": "Postdoc" }, { "name": "Bing Liu", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Hongjiang Lv", "gender": "Male", "institution": "Facebook", "country": "US", "position": "SWE" }, { "name": "Lucas Bandarkar", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Nayan Singhal", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Pritish Yuvraj", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Rui Hou", "gender": "Male", "institution": "Meta Inc. ", "country": "US", "position": "Research Scientist" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 53, "n_ref": 85, "n_ref_all": 97, "n_fig": 3, "n_tab": 2, "L_tab": 1127, "n_element_tab": 368, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 460, "n_element_tab_1": 81, "formula_len_all": 63, "formula_len_all_1": 43, "len_all": 189863, "len_all_1": 64516, "len_abs": 1482, "len_title": 124, "len_sents": 38001, "len_sents_1": 32236, "n_sents": 250, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 76, "L_abs": 1489, "title": "Layer Swapping for Zero-Shot Cross-Lingual Transfer in Large Language Models", "abs": "Model merging, such as model souping, is the practice of combining different models with the same architecture together without further training. In this work, we present a model merging methodology that addresses the difficulty of fine-tuning Large Language Models (LLMs) for target tasks in non-English languages, where task-specific data is often unavailable. We focus on mathematical reasoning and without in-language math data, facilitate cross-lingual transfer by composing language and math capabilities. Starting from the same pretrained model, we fine-tune separate \"experts\" on math instruction data in English and on generic instruction data in the target language. We then replace the top and bottom transformer layers of the math expert directly with layers from the language expert, which consequently enhances math performance in the target language. The resulting merged models outperform the individual experts and other merging methods on the math benchmark, MGSM, by 10% across four major languages where math instruction data is scarce. In addition, this layer swapping is simple, inexpensive, and intuitive, as it is based on an interpretative analysis of the most important parameter changes during the fine-tuning of each expert. The ability to successfully re-compose LLMs for cross-lingual transfer in this manner opens up future possibilities to combine model expertise, create modular solutions, and transfer reasoning capabilities across languages all post hoc.", "keywords": [ "model souping", "model merging", "cross-lingual transfer", "multilingual", "math", "mathematical reasoning", "LLM", "SFT" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 4, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "vQIVbfTMzf", "primary_area": "learning theory", "n_author": 1, "author_info_list": [ { "name": "Qiang Sun", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Associate Professor" } ], "n_formula": 218, "n_formula_1": 26, "n_ref_uni": 20, "n_ref": 40, "n_ref_all": 48, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 30414, "formula_len_all_1": 1715, "len_all": 198818, "len_all_1": 57929, "len_abs": 1035, "len_title": 0, "len_sents": 54040, "len_sents_1": 25203, "n_sents": 672, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1051, "title": "Adapting to both finite-sample and asymptotic regimes", "abs": "This paper introduces an empirical risk minimization based approach with concomitant scaling, which eliminates the need for tuning a robustification parameter in the presence of heavy-tailed data. This method leverages a new loss function that concurrently optimizes both the mean and robustification parameters. Through this dual-parameter optimization, the robustification parameter automatically adjusts to the unknown data variance, rendering the method self-tuning. Our approach surpasses previous models in both computational and asymptotic efficiency. Notably, it avoids the reliance on cross-validation or Lepski's method for tuning the robustification parameter, and the variance of our estimator attains the Cram'{e}r-Rao lower bound, demonstrating optimal efficiency. In essence, our approach demonstrates optimal performance across both finite-sample and large-sample scenarios, a feature we describe as \\textit{algorithmic adaptivity to both asymptotic and finite-sample regimes}. Numerical studies lend strong support to our methodology.", "keywords": [ "algorithmic adaptivity", "empirical risk minimization", "finite-sample regime", "asymptotic regime." ], "rating_list": [ 3, 3, 6, 1 ], "soundness_list": [ 2, 1, 4, 2 ], "presentation_list": [ 2, 2, 4, 1 ], "contribution_list": [ 2, 1, 3, 1 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "vQFw9ryKyK", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Likun Tang", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "Teng Wang", "gender": "Female", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Wenzhe Cai", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xinxin Zhao", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 58, "n_ref": 96, "n_ref_all": 116, "n_fig": 6, "n_tab": 4, "L_tab": 908, "n_element_tab": 161, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 2036, "n_element_tab_1": 23, "formula_len_all": 38, "formula_len_all_1": 0, "len_all": 155602, "len_all_1": 56798, "len_abs": 1785, "len_title": 141, "len_sents": 34242, "len_sents_1": 27974, "n_sents": 229, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1424, "title": "ImagineNav: Prompting Vision-Language Models as Embodied Navigator through Scene Imagination", "abs": "Visual navigation is an essential skill for home-assistance robots, providing the object-searching ability to accomplish long-horizon daily tasks. Many recent approaches use Large Language Models (LLMs) for commonsense inference to improve exploration efficiency. However, the planning process of LLMs is limited within texts and it is difficult to represent the spatial occupancy and geometry layout only by texts. Both are important for making rational navigation decisions. In this work, we seek to unleash the spatial perception and planning ability of Vision-Language Models (VLMs), and explore whether the VLM, with only on-board camera captured RGB/RGB-D stream inputs, can efficiently finish the visual navigation tasks in a mapless manner. We achieve this by developing the imagination-powered navigation framework ImagineNav, which imagines the future observation images at valuable robot views and translates the complex navigation planning process into a rather simple best-view image selection problem for VLM. To generate appropriate candidate robot views for imagination, we introduce the Where2Imagine module, which is distilled to align with human navigation habits. Finally, to reach the VLM preferred views, an off-the-shelf point-goal navigation policy is utilized. Empirical experiments on the challenging open-vocabulary object navigation benchmarks demonstrates the superiority of our proposed system.", "keywords": [ "Robotics", "Visual Navigation", "Vision-Language Model", "Scene Imagination" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "vQ1y086Kn2", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Churan Wang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Postdoc" }, { "name": "Hai Ci", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Hao Chen", "gender": "Male", "institution": "City University of Macau", "country": "MO", "position": "Assistant Professor" }, { "name": "Kui Wu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Yizhou Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Zhoujun Li", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "fangwei zhong", "gender": "Male", "institution": "Beijing Normal University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 36, "n_ref": 59, "n_ref_all": 104, "n_fig": 13, "n_tab": 12, "L_tab": 3599, "n_element_tab": 373, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1297, "n_element_tab_1": 184, "formula_len_all": 68, "formula_len_all_1": 248, "len_all": 173724, "len_all_1": 70887, "len_abs": 1163, "len_title": 122, "len_sents": 56004, "len_sents_1": 34241, "n_sents": 394, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1163, "title": "UnrealCV Zoo: Enriching Photo-realistic Virtual Worlds for Embodied AI Agents", "abs": "The embodied artificial intelligence agents should be capable of sensing, reasoning, planning, and acting in complex open worlds, which are unstructured, high-dynamic, and uncertain. To apply agents in the real world, the realism of the simulated worlds is important for training and evaluating the built agents. This paper introduces UnrealZoo, a rich collection of photo-realistic 3D environments that mimic the complexity and variability of the real world based on Unreal Engine. For embodied AI, we provide a diverse array of playable entities in the environments and a suite of tools, based on UnrealCV, for data collection, reinforcement learning, and evaluation. In the experiments, we benchmark the agent on visual navigation and tracking, two fundamental tasks for embodied vision agents, in complex open worlds. The results provide valuable insights into the strengths of enriching the diversity of the training environments and the challenges to current embodied vision agents in the open worlds, e.g., the latency in the closed-loop control to interact with the dynamic objects, reasoning the accordance of the spatial structure in the complex scenes.", "keywords": [ "Virtual worlds; Embodied AI; Embodied Tracking and Navigation; Visual RL;" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "vQ0zFYJaMo", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Lei Hsiung", "gender": "unknown", "institution": "Dartmouth College", "country": "US", "position": "PhD student" }, { "name": "Linyue Song", "gender": "Male", "institution": "Asana Inc", "country": "US", "position": "Software Engineer" }, { "name": "Pin-Yu Chen", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Tianyu Pang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Tsung-Yi Ho", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "", "position": "Full Professor" }, { "name": "Yaoqing Yang", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "Assistant Professor" }, { "name": "Yung-Chen Tang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 51, "n_ref": 86, "n_ref_all": 112, "n_fig": 5, "n_tab": 9, "L_tab": 4269, "n_element_tab": 314, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1898, "n_element_tab_1": 233, "formula_len_all": 604, "formula_len_all_1": 311, "len_all": 181536, "len_all_1": 68783, "len_abs": 1525, "len_title": 104, "len_sents": 49217, "len_sents_1": 33714, "n_sents": 349, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1536, "title": "Your Task May Vary: A Systematic Understanding of Alignment and Safety Degradation when Fine-tuning LLMs", "abs": "Through supervised fine-tuning or reinforcement learning with human feedback, large language models can achieve a certain level of safety alignment during instruction fine-tuning. However, these *safety guardrails* are often fragile, as models can easily generate harmful content after downstream fine-tuning. Although various methods have been proposed to mitigate this, our paper shifts focus to the durability of safety guardrails, beginning with their formation in the upstream alignment stages. The central question we explore is: *Can we construct more durable safety guardrails for specific downstream tasks to ensure models remain safe after fine-tuning?* Our experiments demonstrate that the durability of these safety guardrails is closely tied to the similarity between upstream and downstream datasets: higher similarity results in more fragile guardrails after fine-tuning, whereas lower similarity results in more durable guardrails. This finding highlights the importance of dataset diversity and privacy in upstream alignment data. Ensuring the diversity of the alignment dataset, which allows downstream datasets to be less similar to it, enhances the guardrail durability for fine-tuning. Maintaining its privacy prevents the exposure of alignment data that adversaries could exploit. Thus, we advocate for a dual strategy: prioritizing both the privacy and diversity of upstream alignment datasets to fortify safety guardrails against potential threats, ensuring long-term model robustness in real-world applications.", "keywords": [ "safety alignment", "task similarity", "guardrail durability" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "vPOMTkmSiu", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Berivan Isik", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Dimitris Paparas", "gender": "Male", "institution": "Google Research", "country": "US", "position": "Researcher" }, { "name": "Hussein Hazimeh", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Natalia Ponomareva", "gender": "Female", "institution": "Google", "country": "", "position": "Software Engineer in Research" }, { "name": "Sergei Vassilvitskii", "gender": "unknown", "institution": "Google", "country": "US", "position": "Scientist" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 6, "n_ref_uni": 65, "n_ref": 102, "n_ref_all": 152, "n_fig": 12, "n_tab": 4, "L_tab": 963, "n_element_tab": 88, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 442, "formula_len_all_1": 242, "len_all": 173052, "len_all_1": 68972, "len_abs": 1496, "len_title": 115, "len_sents": 47086, "len_sents_1": 35825, "n_sents": 318, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1503, "title": "Scaling Laws for Downstream Task Performance in Machine Translation", "abs": "Scaling laws provide important insights that can guide the design of large language models (LLMs). Existing work has primarily focused on studying scaling laws for pretraining (upstream) loss. However, in transfer learning settings, in which LLMs are pretrained on an unsupervised dataset and then finetuned on a downstream task, we often also care about the downstream performance. In this work, we study the scaling behavior in a transfer learning setting, where LLMs are finetuned for machine translation tasks. Specifically, we investigate how the choice of the pretraining data and its size affect downstream performance (translation quality) as judged by: downstream cross-entropy and translation quality metrics such as BLEU and COMET scores. Our experiments indicate that the size of the finetuning dataset and the distribution alignment between the pretraining and downstream data significantly influence the scaling behavior. With sufficient alignment, both downstream cross-entropy and translation quality scores improve monotonically with more pretraining data. In such cases, we show that it is possible to predict the downstream translation quality metrics with good accuracy using a log-law. However, there are cases where moderate misalignment causes the downstream translation scores to fluctuate or get worse with more pretraining, whereas downstream cross-entropy monotonically improves. By analyzing these, we provide new practical insights for choosing appropriate pretraining data.", "keywords": [ "scaling laws", "transfer learning", "machine translation", "large language models", "data valuation" ], "rating_list": [ 8, 8, 8, 6, 3 ], "soundness_list": [ 4, 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 2, 1 ], "contribution_list": [ 3, 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "vOfDGYGVyj", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Enis Cetin", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "US", "position": "Full Professor" }, { "name": "Emadeldeen Hamdan", "gender": "Male", "institution": "University of Illinois at Chicago", "country": "US", "position": "PhD student" }, { "name": "Hongyi Pan", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "Postdoc" } ], "n_formula": 25, "n_formula_1": 25, "n_ref_uni": 21, "n_ref": 34, "n_ref_all": 57, "n_fig": 0, "n_tab": 3, "L_tab": 375, "n_element_tab": 42, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 375, "n_element_tab_1": 42, "formula_len_all": 1146, "formula_len_all_1": 1146, "len_all": 68833, "len_all_1": 68833, "len_abs": 1210, "len_title": 122, "len_sents": 21502, "len_sents_1": 21502, "n_sents": 184, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1237, "title": "Sparse Mamba: Reinforcing Controllability In Structural State Space Models", "abs": "In this work, we introduce the concept of controllability and observability to the Mamba SSM's architecture in our Sparse-Mamba (S-Mamba) for natural language processing (NLP) applications. The structured state space model (SSM) development in recent studies, such as Mamba and Mamba2, outperformed and solved the computational inefficiency of transformers and large language models at small to medium scale. The Mamba SSMs architecture drops the need for attention layers or multilayer perception blocks in transformers. However, current Mamba models lack reinforcement of controllability in state-space equations for computing the $A$, $B$, $C$, and $D$ matrices at each time step, leading to increased complexity and computational costs. In this paper, we demonstrate a reduction of parameters in comparison to the first published Mamba and Mamba2. We showcase an improvement in perplexity by 5\\% and a decrease in training time by 3\\% after reinforcing controllability and observability on the original Mamba architecture in our proposed S-Mamba. The controllable $n \\times n$ state matrix $A$ is sparse and it has only $n$ free parameters. Our novel approach will ensure a controllable system which will be the gate key for Mamba3.", "keywords": [ "Mamba", "state space models", "natural language processing" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 1, 3, 2 ], "presentation_list": [ 2, 1, 3, 1 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 5, 5, 3, 5 ] }, { "paper_id": "vOSwtXGSA2", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Daojing He", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Jiawei Zhou", "gender": "Male", "institution": "Harbin Institute of Technology (ShenZhen)", "country": "CN", "position": "MS student" }, { "name": "Jinyin Hu", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "YU LI", "gender": "Not Specified", "institution": "Harbin Institute of Technology (Shen Zhen)", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 26, "n_ref": 78, "n_ref_all": 98, "n_fig": 4, "n_tab": 10, "L_tab": 4427, "n_element_tab": 631, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1884, "n_element_tab_1": 427, "formula_len_all": 388, "formula_len_all_1": 387, "len_all": 130276, "len_all_1": 66653, "len_abs": 1377, "len_title": 125, "len_sents": 43046, "len_sents_1": 29702, "n_sents": 324, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1383, "title": "An Adaptive Defense Against Adversarial Patch Attacks For Vision Transformers", "abs": "Vision Transformers (ViTs) have become the prominent architecture for various computer vision tasks due to their superior ability to capture long-range dependencies through the self-attention mechanism. However, recent research indicates that ViTs are highly susceptible to carefully crafted adversarial patch attacks, presenting a significant challenge for practical deployment, particularly in security-critical applications. Existing approaches towards robust ViT frameworks often sacrifice clean accuracy and/or achieve suboptimal robustness, likely due to their uniform handling of diverse input samples. In this paper, we present NeighborViT, a novel adaptive defense framework specifically designed to counter adversarial patch attacks for ViTs. NeighborViT stands out by detecting and categorizing different types of attacks on inputs and applying adaptive, tailored defense mechanisms for each type of attack. To realize effective attack detection, categorization, and mitigation, NeighborViT explores the information in neighbor patches of the target patch and strategically employs them for defense. Our experimental results on the ImageNet dataset using various state-of-the-art ViT models demonstrate that NeighborViT significantly enhances robust accuracy without compromising clean accuracy. Our code is available at https://anonymous.4open.science/r/NeighborViT-8255.", "keywords": [ "vision transformer; adversarial patch attack; adptive defense" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "vOFx8HDcvF", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Jinhang Zuo", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "John Lui", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Maoli Liu", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Mohammad Hajiesmaili", "gender": "Male", "institution": "University of Massachusetts Amherst", "country": "US", "position": "Associate Professor" }, { "name": "Xuchuang Wang", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Postdoc" }, { "name": "Xutong Liu", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Postdoc" } ], "n_formula": 73, "n_formula_1": 20, "n_ref_uni": 23, "n_ref": 78, "n_ref_all": 115, "n_fig": 9, "n_tab": 2, "L_tab": 335, "n_element_tab": 20, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 195, "n_element_tab_1": 2, "formula_len_all": 5080, "formula_len_all_1": 1140, "len_all": 147247, "len_all_1": 78234, "len_abs": 1614, "len_title": 96, "len_sents": 51955, "len_sents_1": 34719, "n_sents": 415, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 1366, "title": "Stochastic Bandits Robust to Adversarial Attacks", "abs": "This paper investigates stochastic multi-armed bandit algorithms that are robust to adversarial attacks, where an attacker can first observe the learner's action and *then* alter their reward observation.\nWe study two cases of this model, with or without the knowledge of an attack budget $C$, defined as an upper bound of the summation of the difference between the actual and altered rewards. For both cases, we devise two types of algorithms with regret bounds having additive or multiplicative $C$ dependence terms.\nFor the known attack budget case, we prove our algorithms achieve the regret bound of ${O}((K/\\Delta)\\log T + KC)$ and $\\tilde{O}(\\sqrt{KTC})$ for the additive and multiplicative $C$ terms, respectively, where $K$ is the number of arms, $T$ is the time horizon, $\\Delta$ is the gap between the expected rewards of the optimal arm and the second-best arm, and $\\tilde{O}$ hides the logarithmic factors.\nFor the unknown case, we prove our algorithms achieve the regret bound of $\\tilde{O}(\\sqrt{KT} + KC^2)$ and $\\tilde{O}(KC\\sqrt{T})$ for the additive and multiplicative $C$ terms, respectively.\nIn addition to these upper bound results, we provide several lower bounds showing the tightness of our bounds and the optimality of our algorithms.\nThese results delineate an intrinsic separation between the bandits with attacks and corruption models.", "keywords": [ "Robust Algorithms", "Multi-armed Bandits", "Adversarial Attacks" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "vNdOHr7mn5", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Bernd Bischl", "gender": "Male", "institution": "LMU", "country": "DE", "position": "Full Professor" }, { "name": "Chris Kolb", "gender": "Not Specified", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "David R\u00fcgamer", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Associate Professor" }, { "name": "Tobias Weber", "gender": "Male", "institution": "Department for Statistics", "country": "", "position": "PhD student" } ], "n_formula": 54, "n_formula_1": 11, "n_ref_uni": 77, "n_ref": 149, "n_ref_all": 213, "n_fig": 24, "n_tab": 6, "L_tab": 7889, "n_element_tab": 808, "n_fig_1": 15, "n_tab_1": 2, "L_tab_1": 1212, "n_element_tab_1": 65, "formula_len_all": 4944, "formula_len_all_1": 833, "len_all": 288944, "len_all_1": 67354, "len_abs": 1174, "len_title": 132, "len_sents": 86485, "len_sents_1": 30635, "n_sents": 645, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1184, "title": "Deep Weight Factorization: Sparse Learning Through the Lens of Artificial Symmetries", "abs": "Sparse regularization techniques are well-established in machine learning, yet their application in neural networks remains challenging due to the non-differentiability of penalties like the $L_1$ norm, which is incompatible with stochastic gradient descent. A promising alternative is shallow weight factorization, where weights are decomposed into two factors, allowing for smooth optimization of $L_1$-penalized neural networks by adding differentiable $L_2$ regularization to the factors. \nIn this work, we introduce deep weight factorization, extending previous shallow approaches to more than two factors. We theoretically establish equivalence of our deep factorization with non-convex sparse regularization and analyze its impact on training dynamics and optimization. Due to the limitations posed by standard training practices, we propose a tailored initialization scheme and identify important learning rate requirements necessary for training factorized networks.\nWe demonstrate the effectiveness of our deep weight factorization through experiments on various architectures and datasets, consistently outperforming its shallow counterpart and widely used pruning methods.", "keywords": [ "Sparsity", "Regularization", "Neural Networks", "Overparametrization" ], "rating_list": [ 8, 6, 8, 8, 5 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3, 1 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "vNZIePda08", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Decebal Constantin Mocanu", "gender": "Male", "institution": "University of Luxemburg", "country": "LU", "position": "Associate Professor" }, { "name": "In\u00eas Cardoso Oliveira", "gender": "Female", "institution": "University of Luxemburg", "country": "LU", "position": "PhD student" }, { "name": "Luis A. Leiva", "gender": "unknown", "institution": "University of Luxemburg", "country": "LU", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 33, "n_ref": 117, "n_ref_all": 145, "n_fig": 7, "n_tab": 8, "L_tab": 2011, "n_element_tab": 137, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1199, "n_element_tab_1": 72, "formula_len_all": 620, "formula_len_all_1": 453, "len_all": 166100, "len_all_1": 61230, "len_abs": 1197, "len_title": 93, "len_sents": 44889, "len_sents_1": 29056, "n_sents": 324, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1202, "title": "Sparse-to-Sparse Training of Diffusion Models", "abs": "Diffusion models (DMs) are a powerful type of generative models that have achieved state-of-the-art results in various image synthesis tasks and have shown potential in other domains, such as natural language processing and temporal data modeling. Despite their stable training dynamics and ability to produce diverse high-quality samples, DMs are notorious for requiring significant computational resources, both in the training and inference stages. Previous work has focused mostly on increasing the efficiency of model inference. This paper introduces, for the first time, the paradigm of sparse-to-sparse training to DMs, with the aim of improving both training and inference efficiency. We focus on unconditional generation and train sparse DMs from scratch (Latent Diffusion and ChiroDiff) on six datasets using three different methods (Static-DM, RigL-DM, and MagRan-DM) to study the effect of sparsity in model performance. Our experiments show that sparse DMs are able to match and sometimes outperform their Dense counterparts, while substantially reducing the number of trainable parameters and FLOPs. We also identify safe and effective values to perform sparse-to-sparse training of DMs.", "keywords": [ "Diffusion Models", "Sparse-to-Sparse Training", "Static Sparse Training", "Dynamic Sparse Training" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 1, 2, 1, 1 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "vNQLKY7nFM", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Shyam Venkatasubramanian", "gender": "unknown", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Vahid Tarokh", "gender": "Male", "institution": "Duke University", "country": "", "position": "Professor" } ], "n_formula": 49, "n_formula_1": 23, "n_ref_uni": 42, "n_ref": 54, "n_ref_all": 83, "n_fig": 3, "n_tab": 6, "L_tab": 2133, "n_element_tab": 172, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1419, "n_element_tab_1": 154, "formula_len_all": 3458, "formula_len_all_1": 1885, "len_all": 167729, "len_all_1": 67370, "len_abs": 810, "len_title": 115, "len_sents": 53630, "len_sents_1": 29367, "n_sents": 443, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 812, "title": "Learn2Mix: Training Neural Networks Using Adaptive Data Integration", "abs": "Accelerating model convergence in resource-constrained environments is essential for fast and efficient neural network training. This work presents learn2mix, a new training strategy that adaptively adjusts class proportions within batches, focusing on classes with higher error rates. Unlike classical training methods that use static class proportions, learn2mix continually adapts class proportions during training, leading to faster convergence. Empirical evaluations on benchmark datasets show that neural networks trained with learn2mix converge faster than those trained with classical approaches, achieving improved results for classification, regression, and reconstruction tasks under limited training resources and with imbalanced classes. Our empirical findings are supported by theoretical analysis.", "keywords": [ "adaptive training", "deep learning", "optimization" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "vNGv3dJATp", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Guodong Zheng", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Peng Wang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 78, "n_formula_1": 27, "n_ref_uni": 31, "n_ref": 49, "n_ref_all": 80, "n_fig": 2, "n_tab": 2, "L_tab": 458, "n_element_tab": 81, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 109, "n_element_tab_1": 15, "formula_len_all": 8869, "formula_len_all_1": 2615, "len_all": 139589, "len_all_1": 69188, "len_abs": 1400, "len_title": 108, "len_sents": 39271, "len_sents_1": 28548, "n_sents": 372, "n_sents_1": 243, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1404, "title": "Towards Understanding Memory buffer based Continual Learning", "abs": "Continual learning (CL) is a paradigm that adapts to and retains knowledge from a stream of tasks. Despite the growing number of experimental methods in CL, there is a lack of rigorous theoretical analysis, particularly in memory-based continual learning (MCL), which remains an open research area. In this paper, we theoretically analyze the impact of memory in CL and derive explicit expressions for expected forgetting and generalization errors under overparameterized linear models. We propose a detailed matrix decomposition of the data to distinguish between current and previous datasets, effectively decoupling the coupled data for different tasks. Additionally, we conduct a comprehensive mathematical analysis for scenarios with a small number of tasks and employ numerical analysis for larger task scenarios to evaluate the overall properties of expected forgetting and generalization errors. Compared with CL, our theoretical analysis suggests that (1) a larger memory buffer must be paired with a larger model to effectively reduce forgetting; (2) training with a larger memory buffer generalizes better when tasks are similar but may perform worse when tasks are dissimilar, while training with a large model can help mitigate this negative effect. Ultimately, our findings here sheds new light on how memory can assist CL in mitigating catastrophic forgetting and improving generalization.", "keywords": [ "continual learning", "memory", "catastrophic forgetting", "generalization" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 2, 1, 2, 1 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "vNATZfmY6R", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Alison Gopnik", "gender": "Female", "institution": "University of California, Berkeley", "country": "", "position": "Full Professor" }, { "name": "Anisa Noor Majhi", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Charlie Wong", "gender": "Female", "institution": "Mirror Mirror AI", "country": "US", "position": "Researcher" }, { "name": "Eunice Yiu", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Kate Saenko", "gender": "Female", "institution": "Meta AI", "country": "US", "position": "Researcher" }, { "name": "Maan Qraitem", "gender": "unknown", "institution": "Boston University", "country": "US", "position": "PhD student" }, { "name": "Shiry Ginosar", "gender": "Female", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Assistant Professor" }, { "name": "Yutong Bai", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 86, "n_ref": 165, "n_ref_all": 200, "n_fig": 13, "n_tab": 6, "L_tab": 3635, "n_element_tab": 728, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 205616, "len_all_1": 67558, "len_abs": 1590, "len_title": 119, "len_sents": 57025, "len_sents_1": 35124, "n_sents": 385, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1596, "title": "KiVA: Kid-inspired Visual Analogies for Testing Large Multimodal Models", "abs": "This paper investigates visual analogical reasoning in large multimodal models (LMMs) compared to human adults and children. A \u201cvisual analogy\u201d is an abstract rule inferred from one image and applied to another. \nWhile benchmarks exist for testing visual reasoning in LMMs, they require advanced skills and omit basic visual analogies that even young children can make. Inspired by developmental psychology, we propose a new benchmark of 1,400 visual transformations of everyday objects to test LMMs on visual analogical reasoning and compare them to children and adults. We structure the evaluation into three stages: identifying what changed (e.g., color, number, etc.), how it changed (e.g., added one object), and applying the rule to new scenarios. Our findings show that while models like GPT-4V, LLaVA-1.5, and MANTIS identify the \u201cwhat\u201d effectively, they struggle with quantifying the \u201chow\u201d and extrapolating this rule to new objects. In contrast, children and adults exhibit much stronger analogical reasoning at all three stages. Additionally, the strongest tested model, GPT-4V, performs better in tasks involving simple surface-level visual attributes like color and size, correlating with quicker human adult response times. Conversely, more complex tasks such as number, rotation, and reflection, which necessitate extensive cognitive processing and understanding of extrinsic spatial properties in the physical world, present more significant challenges. Altogether, these findings highlight the limitations of training models on data that primarily consists of 2D images and text.", "keywords": [ "large multimodal models", "analogical reasoning", "cognition", "developmental psychology" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 4, 4, 2 ], "contribution_list": [ 2, 4, 4, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "vMIVqlEWRw", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Alexis Roger", "gender": "unknown", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "Daniel Z Kaplan", "gender": "unknown", "institution": "Sage", "country": "US", "position": "Researcher" }, { "name": "Edwin Fennell", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "MS student" }, { "name": "George Adamopoulos", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "", "position": "Intern" }, { "name": "Irina Rish", "gender": "Female", "institution": "University of Montreal", "country": "CA", "position": "Professor" }, { "name": "Jonathan Siu Chi Lim", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "MS student" }, { "name": "Kshitij Gupta", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "MS student" }, { "name": "Prateek Humane", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "MS student" }, { "name": "Qi Sun", "gender": "Male", "institution": "tokyo institute of technology", "country": "JP", "position": "MS student" }, { "name": "Quentin Gregory Anthony", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 2, "n_ref_uni": 16, "n_ref": 40, "n_ref_all": 79, "n_fig": 16, "n_tab": 5, "L_tab": 2200, "n_element_tab": 129, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1911, "n_element_tab_1": 75, "formula_len_all": 112, "formula_len_all_1": 80, "len_all": 160302, "len_all_1": 54279, "len_abs": 5002, "len_title": 135, "len_sents": 52423, "len_sents_1": 24535, "n_sents": 446, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 835, "title": "Robin: a Suite of Multi-Scale Vision-Language Models and the CHIRP Evaluation Benchmark", "abs": "The proliferation of Vision-Language Models (VLMs) in the past several years calls for rigorous and comprehensive evaluation methods and benchmarks. This work analyzes existing VLM evaluation techniques, including automated metrics, AI-based assessments, and human evaluations across diverse tasks. We first introduce Robin - a novel suite of VLMs that we built by combining Large Language Models (LLMs) and Vision Encoders (VEs) at multiple scales, and use Robin to identify shortcomings of current evaluation approaches across scales. Next, to overcome the identified limitations, we introduce CHIRP - a new long form response benchmark we developed for more robust and complete VLM evaluation. We provide open access to the Robin training code, model suite, and CHIRP benchmark to promote reproducibility and advance VLM research.", "keywords": [ "Vision-Language Models", "Benchmarks", "Scalling Suites" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 1, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "vMA0ATykNU", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Bijun Li", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Jian Zhou", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Associate Professor" }, { "name": "Jinsheng Xiao", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Associate Professor" }, { "name": "Minghao Yu", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Undergrad student" }, { "name": "Yaoan Bian", "gender": "unknown", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Ziyue Tian", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 5, "n_ref": 5, "n_ref_all": 16, "n_fig": 2, "n_tab": 7, "L_tab": 4247, "n_element_tab": 199, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1002, "n_element_tab_1": 145, "formula_len_all": 622, "formula_len_all_1": 587, "len_all": 100158, "len_all_1": 59322, "len_abs": 1712, "len_title": 135, "len_sents": 25839, "len_sents_1": 20705, "n_sents": 191, "n_sents_1": 151, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1804, "title": "LSTR: Long-Short Range Aggregation for Trajectory Prediction at Intersection Scenarios", "abs": "Trajectory prediction is crucial for practical applications, encompassing navigation for autonomous vehicles and the implementation of safety systems based on the Internet of Vehicles (IoV). Most existing methods significantly rely on comprehensive map information, employing robust rule constraints to incrementally predict trajectories within the driver's local decision-making context. However, in environments characterized by weak rule enforcement, such as urban intersections, these approaches neglect the disparity between the driver's overarching intentions and current behaviors.Recognizing the characteristics of intersection traffic flow\u2014macroscopically organized yet microscopically disordered, exhibiting highly heterogeneous conditions\u2014this paper presents a novel model termed Long-short Range Aggregation for Trajectory Prediction in Intersections (LSTR). This model anchors the vehicle's local decision-making process to long-range intentions. Specifically, LSTR predicts the vehicle's destination via a global intention inference module and models its long-range driving intentions through clustering to extract macroscopic traffic flow patterns. This long-range intention subsequently informs the short-range local interaction behaviors captured by the local behavior decision module. Ultimately, the fused features from these two modules are analyzed using a multi-modal decoder to interpret the various motion patterns, resulting in the trajectory prediction outcomes.We rigorously validate the proposed framework across multiple intersection scenarios utilizing real-world datasets, including inD, roundD, and a subset of WOMD. Experimental results demonstrate that our model outperforms numerous benchmarks without relying on additional information such as HD maps of intersections.", "keywords": [ "motion prediction", "autonomous driving", "path_planning" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 1, 1, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "vM94dZiqx4", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Changick Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Hongsin Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Seungju Cho", "gender": "Male", "institution": "\tKorea Advanced Institute of Science and Technology ", "country": "KR", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 11, "n_ref_uni": 48, "n_ref": 102, "n_ref_all": 130, "n_fig": 8, "n_tab": 9, "L_tab": 5594, "n_element_tab": 591, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3349, "n_element_tab_1": 330, "formula_len_all": 3169, "formula_len_all_1": 1276, "len_all": 145274, "len_all_1": 60720, "len_abs": 1429, "len_title": 103, "len_sents": 32715, "len_sents_1": 22945, "n_sents": 239, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1440, "title": "Long-tailed Adversarial Training with Self-Distillation", "abs": "Adversarial training significantly enhances adversarial robustness, yet superior performance is predominantly achieved on balanced datasets.\n Addressing adversarial robustness in the context of unbalanced or long-tailed distributions is considerably more challenging, mainly due to the scarcity of tail data instances. \n Previous research on adversarial robustness within long-tailed distributions has primarily focused on combining traditional long-tailed natural training with existing adversarial robustness methods.\n In this study, we provide an in-depth analysis for the challenge that adversarial training struggles to achieve high performance on tail classes in long-tailed distributions.\n Furthermore, we propose a simple yet effective solution to advance adversarial robustness on long-tailed distributions through a novel self-distillation technique.\n Specifically, this approach leverages a balanced self-teacher model, which is trained using a balanced dataset sampled from the original long-tailed dataset.\nOur extensive experiments demonstrate state-of-the-art performance in both clean and robust accuracy for long-tailed adversarial robustness, with significant improvements in tail class performance on various datasets.\nWe improve the accuracy against PGD attacks for tail classes by 20.3, 7.1, and 3.8 percentage points on CIFAR-10, CIFAR-100, and Tiny-ImageNet, respectively, while achieving the highest robust accuracy.", "keywords": [ "Adversarial Robustness", "Adversarial Training", "Long-Tail Distribution Learning" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "vM4CdVScT8", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 7, "author_info_list": [ { "name": "Dan Li", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Dayu Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Tong Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Wang Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yanshu Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Postdoc" }, { "name": "Yujie Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Zhaoqian YAO", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Intern" } ], "n_formula": 31, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 45, "n_ref_all": 69, "n_fig": 17, "n_tab": 1, "L_tab": 85, "n_element_tab": 7, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 85, "n_element_tab_1": 7, "formula_len_all": 2131, "formula_len_all_1": 775, "len_all": 129215, "len_all_1": 64468, "len_abs": 1667, "len_title": 164, "len_sents": 40920, "len_sents_1": 29802, "n_sents": 355, "n_sents_1": 243, "venue": "ICLR2025", "decision": "withdraw", "L_title": 116, "L_abs": 1680, "title": "Quantum Entanglement Trees: Optimizing Quantized Matrix Quantization via Element Replacement and Residual Clustering", "abs": "The matrix quantization entails representing matrix elements in a more space-efficient form to reduce storage usage, with dequantization restoring the original matrix for use. We formulate the Quantization Error Minimization (QEM) problem as minimizing the distance between a matrix before and after quantization, under the condition that the quantized matrix occupies the same memory space. Matrix quantization is crucial in various applications, including Large Language Models (LLMs) weight quantization, vector databases, KV cache quantization, graph compression, and image compression. Recent advancements in LLMs, such as GPT-4 and BERT, have highlighted the importance of matrix compression due to the large size of parameters and KV cache, which are stored as matrices. \n\nWe propose Quantum Entanglement Trees (QET) to address the QEM problem by leveraging the local orderliness of matrix elements, involving iterative element swapping to form a locally ordered matrix. This matrix is then grouped and quantized by columns. To enhance QET, we introduce two optimizations: Residual Quantization Optimization (RQO), which reduces MSE by quantizing the residuals between the original and dequantized matrices, and Codebook Quantization Optimization (CQO), which reduces storage requirements by compressing the codebook itself.\n\nExperimental results demonstrate that QET can effectively reduce MSE to 5.05\\%, 13.33\\%, and 11.89\\% of the current best method on the LLM dataset, K cache, and V cache, respectively.\nOur contributions include the abstraction of the QEM problem, the design of the QET algorithm, and the proposal of two optimizations to improve accuracy and speed.", "keywords": [ "Matrix quantization", "LLM Weight Quantization", "KV Cache Quantization", "Residual Quantization" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 5, 4, 2, 3 ] }, { "paper_id": "vL9t9tpKli", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chaoyi Zhou", "gender": "unknown", "institution": "Clemson University", "country": "US", "position": "PhD student" }, { "name": "Feng Luo", "gender": "Male", "institution": "Clemson University", "country": "US", "position": "Full Professor" }, { "name": "Siyu Huang", "gender": "Male", "institution": "Clemson University", "country": "US", "position": "Assistant Professor" }, { "name": "Xi Liu", "gender": "Male", "institution": "Clemson University", "country": "US", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 19, "n_ref_uni": 34, "n_ref": 59, "n_ref_all": 70, "n_fig": 6, "n_tab": 2, "L_tab": 355, "n_element_tab": 65, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 872, "formula_len_all_1": 875, "len_all": 105425, "len_all_1": 57230, "len_abs": 1243, "len_title": 104, "len_sents": 31041, "len_sents_1": 28205, "n_sents": 204, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1195, "title": "Latent Radiance Fields with 3D-aware 2D Representations", "abs": "Latent 3D reconstruction has shown great promise in empowering 3D semantic understanding and 3D generation by distilling 2D features into the 3D space. However, existing approaches struggle with the domain gap between 2D feature space and 3D representations, resulting in degraded rendering performance. To address this challenge, we propose a novel framework that integrates 3D awareness into the 2D latent space. The framework consists of three stages: (1) a correspondence-aware autoencoding method that enhances the 3D consistency of 2D latent representations, (2) a latent radiance field (LRF) that lifts these 3D-aware 2D representations into 3D space, and (3) a VAE-Radiance Field (VAE-RF) alignment strategy that improves image decoding from the rendered 2D representations. Extensive experiments demonstrate that our method outperforms the state-of-the-art latent 3D reconstruction approaches in terms of synthesis performance and cross-dataset generalizability across diverse indoor and outdoor scenes. To our knowledge, this is the first work showing the radiance field representations constructed from 2D latent representations can yield photorealistic 3D reconstruction performance.", "keywords": [ "3D Gaussian Splatting", "3D-aware Representation" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "vKgDbYKZrH", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 14, "author_info_list": [ { "name": "Amit S", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Applied Research" }, { "name": "Anshul Mittal", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Researcher" }, { "name": "Bhavyajeet Singh", "gender": "Male", "institution": "International Institute of Information Technology, Hyderabad, International Institute of Information Technology Hyderabad", "country": "", "position": "MS student" }, { "name": "Deepak Saini", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Jian Jiao", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Lakshya Kumar", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Applied Research" }, { "name": "Manik Varma", "gender": "Male", "institution": "Research, Microsoft", "country": "IN", "position": "Principal Researcher" }, { "name": "Manish Gupta", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Principal Researcher" }, { "name": "Niket Tandon", "gender": "Male", "institution": "Research, Microsoft", "country": "IN", "position": "Principal Researcher" }, { "name": "Shikhar Mohan", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Applied Scientist" }, { "name": "Siddarth Asokan", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Researcher" }, { "name": "Suchith Chidananda Prabhu", "gender": "Male", "institution": "Indian Institute of Technology, Delhi", "country": "IN", "position": "PhD student" }, { "name": "Sumeet Agarwal", "gender": "Male", "institution": "Indian Institute of Technology Delhi", "country": "", "position": "Associate Professor" }, { "name": "Yashoteja Prabhu", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Researcher" } ], "n_formula": 28, "n_formula_1": 9, "n_ref_uni": 50, "n_ref": 90, "n_ref_all": 127, "n_fig": 4, "n_tab": 11, "L_tab": 3041, "n_element_tab": 231, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 365, "n_element_tab_1": 48, "formula_len_all": 3589, "formula_len_all_1": 602, "len_all": 202301, "len_all_1": 55426, "len_abs": 2245, "len_title": 123, "len_sents": 56231, "len_sents_1": 25250, "n_sents": 424, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1650, "title": "MOGIC: METADATA-INFUSED ORACLE GUIDANCE FOR IMPROVED EXTREME CLASSIFICATION", "abs": "Retrieval-augmented classification and generation models significantly benefit from the *early-stage fusion* of high-quality text-based auxiliary metadata, often called memory, but they suffer from high inference latency and poor robustness to noise. In classifications tasks, particularly the extreme classification (XC) setting, where low latency is critical, existing methods incorporate metadata for context enrichment via an XC-based retriever and obtain the representations of the relevant memory items to perform *late-stage fusion* to achieve low latency. With an aim of achieving higher accuracy while meeting the low latency constraints, in this paper, we propose MOGIC, an approach for metadata-infused Oracle guidance for XC tasks. In particular, we train an early-fusion Oracle classifier with access to both query-side and label-side ground-truth metadata in the textual form. The Oracle is subsequently used to guide the training of any existing memory-based XC Disciple model via regularization. The MOGIC algorithm, when applied to memory-based XC Disciple models such as OAK, improves precision@1 and propensity-scored precision@1 by ~2% on four standard datasets, at no additional inference-time costs to the Disciple model. We also show the feasibility of applying the MOGIC algorithm to improve the performance of state-of-the-art memory-free XC approaches such as NGAME or DEXA, demonstrating that the MOGIC algorithm can be used atop any existing XC-based approach in a *plug-and-play* manner. Finally, we also show the robustness of the MOGIC method to missing and noisy metadata settings. We will release code on acceptance.", "keywords": [ "recommendation systems", "auxiliary information", "extreme classification", "metadata" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 2, 5 ] }, { "paper_id": "vKL1i2p5Xr", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Feng Yu", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Jianfeng Lu", "gender": "Male", "institution": "Nanjing university of Science & Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yang Yang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Wu Xiangyu", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 53, "n_ref": 179, "n_ref_all": 211, "n_fig": 7, "n_tab": 9, "L_tab": 3172, "n_element_tab": 166, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 2952, "n_element_tab_1": 164, "formula_len_all": 637, "formula_len_all_1": 642, "len_all": 210974, "len_all_1": 70388, "len_abs": 1480, "len_title": 126, "len_sents": 52692, "len_sents_1": 29549, "n_sents": 502, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1486, "title": "Text as Any-Modality for Zero-shot Classification by Consistent Prompt Tuning", "abs": "The integration of prompt tuning with multimodal learning has shown significant generalization abilities for various downstream tasks. Despite advancements, existing methods heavily depend on massive modality-specific labeled data (e.g., video, audio, and image), or are customized for a single modality. In this study, we present Text as Any-Modality by Consistent Prompt Tuning (TaAM-CPT), a scalable approach for constructing a general representation model toward unlimited modalities using solely text data. TaAM-CPT comprises modality prompt pools, text construction, and modality-aligned text encoders from pre-trained models, which allows for extending new modalities by adding prompt pools and modality-aligned text encoders. To harmonize the learning across different modalities, TaAM-CPT designs intra- and inter-modal learning objectives, which can capture category details within modalities while maintaining semantic consistency across different modalities. Benefiting from its scalable architecture and pre-trained models, TaAM-CPT can be seamlessly extended to accommodate unlimited modalities. Remarkably, without any modality-specific labeled data, TaAM-CPT achieves leading results on diverse datasets spanning various modalities, including video classification (Kinetic-400/600/700), image classification (MSCOCO, VOC2007, NUSWIDE, VOC2012, Objects365), and audio classification (ESC50, US8K). The code is available at https://anonymous.4open.science/r/TaAM-CPT-0EA6.", "keywords": [ "Multimodal Learning ; Prompt Learning; Zero-shot Classification;" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "vKJ8YH0iNp", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Gaowen Liu", "gender": "Female", "institution": "Cisco Systems", "country": "US", "position": "Researcher" }, { "name": "Gaurav Kumar Nayak", "gender": "Male", "institution": "IIT Roorkee ", "country": "IN", "position": "Assistant Professor" }, { "name": "Jeffrey A Chan Santiago", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Mubarak Shah", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "praveen tirupattur", "gender": "Male", "institution": "Snapshots Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 32, "n_ref": 69, "n_ref_all": 92, "n_fig": 10, "n_tab": 6, "L_tab": 2441, "n_element_tab": 215, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 2958, "n_element_tab_1": 35, "formula_len_all": 630, "formula_len_all_1": 502, "len_all": 127008, "len_all_1": 67995, "len_abs": 1889, "len_title": 111, "len_sents": 40513, "len_sents_1": 30637, "n_sents": 318, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1856, "title": "MGD$^3$: Mode-Guided Dataset Distillation using Diffusion Models", "abs": "Dataset distillation aims to distill a smaller training dataset from a larger one so that a model trained on this smaller set performs similarly to one trained on the full dataset. Traditional methods are costly and lack sample diversity. Recent approaches utilizing generative models, particularly diffusion models, show promise in capturing data distribution, but they often oversample prominent modes, limiting sample diversity.\nTo address these limitations in this work, we propose a mode-guided diffusion model. Unlike existing works that fine-tune the diffusion models for dataset distillation, we propose to use a pre-trained model without the need for fine-tuning. Our novel approach consists of three stages: Mode Discovery, Mode Guidance, and Stop Guidance. In the first stage, we discover distinct modes in the data distribution of a class to build a representative set. In the second stage, we use a pre-trained diffusion model and guide the diffusion process toward the discovered modes to generate distinct samples, ensuring intra-class diversity. However, mode-guided sampling can introduce artifacts in the synthetic sample, which affect the performance. To control the fidelity of the synthetic dataset, we introduce the stop guidance. \nWe evaluate our method on multiple benchmark datasets, including ImageNette, ImageIDC, ImageNet-100, and ImageNet-1K; \nOur method improved $4.4\\%$, $2.9\\%$, $1.6\\%$, and $1.6\\%$ over the current state-of-the-art on the respective datasets.\nIn addition, our method does not require retraining of the diffusion model, which leads to reduced computational requirements. \nWe also demonstrate that our approach is effective with general-purpose diffusion models such as Text-to Image Stable Diffusion, showing promising performance towards eliminating the need for a pre-trained model in the target dataset.", "keywords": [ "Dataset Distillation; Dataset Condensation; Diffusion;" ], "rating_list": [ 8, 6, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "vKG270UOg4", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Jing-Hao Xue", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Full Professor" }, { "name": "Qingmin Liao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Wenming Yang", "gender": "Male", "institution": "Tsinghua University,", "country": "CN", "position": "Associate Professor" }, { "name": "Xu Zidong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Ling Peng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhangzongkai", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 43, "n_formula_1": 9, "n_ref_uni": 48, "n_ref": 98, "n_ref_all": 124, "n_fig": 11, "n_tab": 11, "L_tab": 2996, "n_element_tab": 433, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 2747, "n_element_tab_1": 313, "formula_len_all": 5322, "formula_len_all_1": 1078, "len_all": 170745, "len_all_1": 55684, "len_abs": 1943, "len_title": 120, "len_sents": 45269, "len_sents_1": 23048, "n_sents": 343, "n_sents_1": 158, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1964, "title": "BDC-Occ: Binarized Deep Convolution Unit For Binarized Occupancy Network", "abs": "Existing 3D occupancy networks demand significant hardware resources, hindering the deployment of edge devices. Binarized Neural Networks (BNNs) offer a potential solution by substantially reducing computational and memory requirements. However, their performances decrease notably compared to full-precision networks. In addition, it is challenging to enhance the performance of the binarized model by increasing the number of binarized convolutional layers, which limits its practicability for 3D occupancy prediction. This paper presents two original insights into binarized convolution, substantiated with theoretical proofs: (a) $1\\times1$ binarized convolution introduces minimal binarization errors as the network deepens, and (b) binarized convolution is inferior to full-precision convolution in capturing cross-channel feature importance. Building on the above insights, we propose a novel binarized deep convolution (BDC) unit that significantly enhances performance, even when the number of binarized convolutional layers increases. Specifically, in the BDC unit, additional binarized convolutional kernels are constrained to $1\\times1$ to minimize the effects of binarization errors. Further, we propose a per-channel refinement branch to reweight the output via first-order approximation. Then, we partition the 3D occupancy networks into four convolutional modules, using the proposed BDC unit to binarize them. The proposed BDC unit minimizes binarization errors and improves perceptual capability while significantly boosting computational efficiency, meeting the stringent requirements for accuracy and speed in occupancy prediction. Extensive quantitative and qualitative experiments validate that the proposed BDC unit supports state-of-the-art precision in occupancy prediction and object detection tasks with substantially reduced parameters and operations. Code is provided in the supplementary material and will be open-sourced upon review.", "keywords": [ "3D occupancy prediction; binarized networks" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "vK8C37eHXM", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "David Minnen", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Gabriel Barcik", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "James David Lyon", "gender": "Male", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Joshua V. Dillon", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Sergey Ioffe", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Vighnesh Birodkar", "gender": "Male", "institution": "Luma Labs", "country": "US", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 34, "n_ref": 66, "n_ref_all": 92, "n_fig": 9, "n_tab": 2, "L_tab": 428, "n_element_tab": 130, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 272, "n_element_tab_1": 30, "formula_len_all": 577, "formula_len_all_1": 577, "len_all": 97780, "len_all_1": 48918, "len_abs": 1297, "len_title": 30, "len_sents": 21836, "len_sents_1": 20891, "n_sents": 183, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 30, "L_abs": 1307, "title": "Sample what you can't compress", "abs": "For learned image representations, basic autoencoders often produce blurry results. Reconstruction quality can be improved by incorporating additional penalties such as adversarial (GAN) and perceptual losses. Arguably, these approaches lack a principled interpretation. Concurrently, in generative settings diffusion has demonstrated a remarkable ability to create crisp, high quality results and has solid theoretical underpinnings (from variational inference to direct study as the Fisher Divergence). Our work combines autoencoder representation learning with diffusion and is, to our knowledge, the first to demonstrate jointly learning\na continuous encoder and decoder under a diffusion-based loss\nand showing that it can lead to higher compression and better generation.. \nWe demonstrate that this approach yields better reconstruction quality as compared to GAN-based\nautoencoders while being easier to tune. \nWe also show that the resulting representation is easier to model\nwith a latent diffusion model as compared to the representation obtained from a state-of-the-art GAN-based loss.\nSince our decoder is stochastic, it can generate details not encoded in the otherwise deterministic latent representation; we therefore name our approach ``Sample what you can't compress'', or SWYCC for short.", "keywords": [ "autoencoders_diffusion+generative models" ], "rating_list": [ 3, 3, 1, 3, 6 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2, 3 ], "contribution_list": [ 2, 1, 1, 2, 3 ], "confidence_list": [ 5, 4, 5, 4, 5 ] }, { "paper_id": "vJwjWyt4Ed", "primary_area": "reinforcement learning", "n_author": 9, "author_info_list": [ { "name": "Gang Niu", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Senior Research Scientist (tenured)" }, { "name": "Jing-Cheng Pang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Kaiyuan Li", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Undergrad student" }, { "name": "Masashi Sugiyama", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Director" }, { "name": "Nan Tang", "gender": "Male", "institution": "nanjing university", "country": "", "position": "PhD student" }, { "name": "Xin-Qiang Cai", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Postdoc" }, { "name": "Yuting Tang", "gender": "Female", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Zhen-Yu Zhang", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Postdoc" }, { "name": "Yang Yu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Professor" } ], "n_formula": 13, "n_formula_1": 7, "n_ref_uni": 41, "n_ref": 98, "n_ref_all": 116, "n_fig": 9, "n_tab": 5, "L_tab": 1141, "n_element_tab": 190, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 276, "n_element_tab_1": 53, "formula_len_all": 1291, "formula_len_all_1": 643, "len_all": 172771, "len_all_1": 73697, "len_abs": 1509, "len_title": 116, "len_sents": 59328, "len_sents_1": 35682, "n_sents": 435, "n_sents_1": 275, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1517, "title": "Learning View-invariant World Models for Visual Robotic Manipulation", "abs": "Robotic manipulation tasks often rely on visual inputs from cameras to perceive the environment. However, previous approaches still suffer from performance degradation when the camera\u2019s viewpoint changes during manipulation. In this paper, we propose ReViWo (Representation learning for View-invariant World model), leveraging multi-view data to learn robust representations for control under viewpoint disturbance. ReViWo utilizes an autoencoder framework to reconstruct target images by an architecture that combines view-invariant representation (VIR) and view-dependent representation. To train ReViWo, we collect multi-view data in simulators with known view labels, meanwhile, ReViWo is simutaneously trained on Open X-Embodiment datasets without view labels. The VIR is then used to train a world model on pre-collected manipulation data and a policy through interaction with the world model. We evaluate the effectiveness of ReViWo in various viewpoint disturbance scenarios, including control under novel camera positions and frequent camera shaking, using the Meta-world & PandaGym environments. Besides, we also conduct experiments on real world ALOHA robot. The results demonstrate that ReViWo maintains robust performance under viewpoint disturbance, while baseline methods suffer from significant performance degradation. Furthermore, we show that the VIR captures task-relevant state information and remains stable for observations from novel viewpoints, validating the efficacy of the ReViWo approach.", "keywords": [ "Robotic manipulation", "reinforcement learning", "world model" ], "rating_list": [ 3, 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 4, 3, 2 ], "confidence_list": [ 5, 2, 3, 3, 4 ] }, { "paper_id": "vJmpg0exYA", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 7, "author_info_list": [ { "name": "Arturs Backurs", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Janardhan Kulkarni", "gender": "Male", "institution": "Microsoft Research, Redmond", "country": "US", "position": "Researcher" }, { "name": "Jerry Chee", "gender": "unknown", "institution": "Cornell University", "country": "", "position": "PhD student" }, { "name": "Li Zhang", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Software Engineer" }, { "name": "Rainie Heck", "gender": "unknown", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Sivakanth Gopi", "gender": "Male", "institution": "Microsoft Research", "country": "", "position": "Senior Researcher" }, { "name": "Thomas Rothvoss", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Full Professor" } ], "n_formula": 38, "n_formula_1": 16, "n_ref_uni": 28, "n_ref": 67, "n_ref_all": 97, "n_fig": 8, "n_tab": 5, "L_tab": 4601, "n_element_tab": 401, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 2581, "n_element_tab_1": 247, "formula_len_all": 3688, "formula_len_all_1": 1122, "len_all": 169246, "len_all_1": 74083, "len_abs": 1573, "len_title": 83, "len_sents": 45088, "len_sents_1": 29463, "n_sents": 439, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1619, "title": "DiscQuant: A Quantization Method for Neural Networks Inspired by Discrepancy Theory", "abs": "Quantizing the weights of a neural network has two steps: (1) Finding a good low bit-complexity representation for weights (which we call the quantization grid) and (2) Rounding the original weights to values in the quantization grid. In this paper, we study the problem of rounding optimally given any quantization grid. The simplest and most commonly used way to round is Round-to-Nearest (RTN). By rounding in a data-dependent way instead, one can improve the quality of the quantized model significantly.\n\nWe study the rounding problem from the lens of \\emph{discrepancy theory}, which studies how well we can round a continuous solution to a discrete solution without affecting solution quality too much. We prove that given $m=poly(1/\\epsilon)$ samples from the data distribution, we can round all but $O(m)$ model weights such that the expected approximation error of the quantized model on the true data distribution is $\\le \\epsilon$ as long as the space of gradients of the original model is approximately low rank (which we empirically validate).\n\nOur proof, which is algorithmic, inspired a simple and practical rounding algorithm called \\emph{DiscQuant}. In our experiments, we demonstrate that DiscQuant significantly improves over the prior state-of-the-art rounding method called GPTQ and the baseline RTN over a range of benchmarks on Phi3mini-3.8B and Llama3.1-8B. For example, rounding Phi3mini-3.8B to a fixed quantization grid with 3.25 bits per parameter using DiscQuant gets 64\\% accuracy on the GSM8k dataset, whereas GPTQ achieves 54\\% and RTN achieves 31\\% (the original model achieves 84\\%).", "keywords": [ "Quantization", "Discrepancy Theory", "LLMs", "Weights Only Quantization" ], "rating_list": [ 6, 3, 6, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 4, 2 ], "confidence_list": [ 2, 5, 3, 4 ] }, { "paper_id": "vJkktqyU8B", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Dong Zhang", "gender": "Male", "institution": "HKUST", "country": "CN", "position": "Researcher" }, { "name": "Kwang-Ting Cheng", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Pingcheng Dong", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Rui Yan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 64, "n_ref": 182, "n_ref_all": 193, "n_fig": 3, "n_tab": 5, "L_tab": 2713, "n_element_tab": 90, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 110, "n_element_tab_1": 5, "formula_len_all": 706, "formula_len_all_1": 482, "len_all": 156999, "len_all_1": 69809, "len_abs": 1497, "len_title": 106, "len_sents": 32521, "len_sents_1": 24888, "n_sents": 218, "n_sents_1": 150, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1502, "title": "Memory Efficient Transformer Adapter for Dense Predictions", "abs": "While current Vision Transformer (ViT) adapter methods have shown promising accuracy, their inference speed is implicitly hindered by inefficient memory access operations, e.g., standard normalization and frequent reshaping. In this work, we propose META, a simple and fast ViT adapter that can improve the model's memory efficiency and decrease memory time consumption by reducing the inefficient memory access operations. Our method features a memory-efficient adapter block that enables the common sharing of layer normalization between the self-attention and feed-forward network layers, thereby reducing the model's reliance on normalization operations. Within the proposed block, the cross-shaped self-attention is employed to reduce the model's frequent reshaping operations. Moreover, we augment the adapter block with a lightweight convolutional branch that can enhance local inductive biases, particularly beneficial for the dense prediction tasks, e.g., object detection, instance segmentation, and semantic segmentation. The adapter block is finally formulated in a cascaded manner to compute diverse head features, thereby enriching the variety of feature representations. Empirically, extensive evaluations on multiple representative datasets validate that META substantially enhances the predicted quality, while achieving a new state-of-the-art accuracy-efficiency trade-off. Theoretically, we demonstrate that META exhibits superior generalization capability and stronger adaptability.", "keywords": [ "Vision Transformer", "Vision Transformer", "Transformer" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "vJgJSrYPe1", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Shuang Li", "gender": "Female", "institution": "The Chinese University of Hong Kong (Shenzhen)", "country": "", "position": "Assistant Professor" }, { "name": "Shuhan Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Undergrad student" }, { "name": "Wendi Ren", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 22, "n_ref_uni": 26, "n_ref": 51, "n_ref_all": 69, "n_fig": 3, "n_tab": 8, "L_tab": 2785, "n_element_tab": 180, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1395, "formula_len_all_1": 1148, "len_all": 133515, "len_all_1": 61171, "len_abs": 1589, "len_title": 102, "len_sents": 51508, "len_sents_1": 28699, "n_sents": 385, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1620, "title": "Logic-Logit: A Logic-Based Approach to Choice Modeling", "abs": "In this study, we propose a novel rule-based interpretable choice model, {\\bf Logic-Logit}, designed to effectively learn and explain human choices. Choice models have been widely applied across various domains\u2014such as commercial demand forecasting, recommendation systems, and consumer behavior analysis\u2014typically categorized as parametric, nonparametric, or deep network-based. While recent innovations have favored neural network approaches for their computational power, these flexible models often involve large parameter sets and lack interpretability, limiting their effectiveness in contexts where transparency is essential.\n\nPrevious empirical evidence shows that individuals usually use {\\it heuristic decision rules} to form their consideration sets, from which they then choose. These rules are often represented as {\\it disjunctions of conjunctions} (i.e., OR-of-ANDs). These rules-driven, {\\it consider-then-choose} decision processes enable people to quickly screen numerous alternatives while reducing cognitive and search costs. Motivated by this insight, our approach leverages logic rules to elucidate human choices, providing a fresh perspective on preference modeling. We introduce a unique combination of column generation techniques and the Frank-Wolfe algorithm to facilitate efficient rule extraction for preference modeling\u2014a process recognized as NP-hard. Our empirical evaluation, conducted on both synthetic datasets and real-world data from commercial and healthcare domains, demonstrates that Logic-Logit significantly outperforms baseline models in terms of interpretability and accuracy.", "keywords": [ "Choice Model", "Preference Learning", "Interpretability", "Rule Learning" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "vJ0axKTh7t", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Cewu Lu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Hong Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Jianbin Zhu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Nanxi Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Qinlu Guo", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Yong-Lu Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuanjie Chen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Undergrad student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 29, "n_ref": 69, "n_ref_all": 110, "n_fig": 14, "n_tab": 8, "L_tab": 3202, "n_element_tab": 365, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 893, "formula_len_all_1": 895, "len_all": 200660, "len_all_1": 65600, "len_abs": 1840, "len_title": 124, "len_sents": 59126, "len_sents_1": 29538, "n_sents": 500, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1761, "title": "The Labyrinth of Links: Navigating the Associative Maze of Multi-modal LLMs", "abs": "Multi-modal Large Language Models (MLLMs) have exhibited impressive capability. However, recently many deficiencies of MLLMs have been found compared to human intelligence, $\\textit{e.g.}$, hallucination. To drive the MLLMs study, the community dedicated efforts to building larger benchmarks with complex tasks. In this paper, we propose benchmarking an essential but usually overlooked intelligence: $\\textbf{association}$, a human's basic capability to link observation and prior practice memory. To comprehensively investigate MLLM's performance on the association, we formulate the association task and devise a standard benchmark based on adjective and verb semantic concepts. Instead of costly data annotation and curation, we propose a convenient $\\textbf{annotation-free}$ construction method transforming the general dataset for our association tasks. Simultaneously, we devise a rigorous data refinement process to eliminate confusion in the raw dataset. Building on this database, we establish three levels of association tasks: single-step, synchronous, and asynchronous associations. Moreover, we conduct a comprehensive investigation into the MLLMs' zero-shot association capabilities, addressing multiple dimensions, including three distinct memory strategies, both open-source and closed-source MLLMs, cutting-edge Mixture-of-Experts (MoE) models, and the involvement of human experts. Our systematic investigation shows that current open-source MLLMs consistently exhibit poor capability in our association tasks, even the currently state-of-the-art GPT-4V(vision) also has a significant gap compared to humans. We believe our benchmark would pave the way for future MLLM studies. $\\textit{Our data and code will be made publicly available.}$", "keywords": [ "Multi-modal LLM", "Visual Reasoning", "Association" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 5, 3, 5 ] }, { "paper_id": "vIHmkF5rnC", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Chenhao Yang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Haochen Xu", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Rujun Jiang", "gender": "Male", "institution": "Fudan University", "country": "", "position": "Associate Professor" } ], "n_formula": 39, "n_formula_1": 33, "n_ref_uni": 30, "n_ref": 43, "n_ref_all": 77, "n_fig": 2, "n_tab": 3, "L_tab": 280, "n_element_tab": 10, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 100, "n_element_tab_1": 9, "formula_len_all": 2704, "formula_len_all_1": 2489, "len_all": 115706, "len_all_1": 62254, "len_abs": 889, "len_title": 121, "len_sents": 24727, "len_sents_1": 23397, "n_sents": 230, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 895, "title": "Lower-level Duality Based Penalty Methods for Hyperparameter Optimization", "abs": "Hyperparameter optimization (HO) is essential in machine learning and can be structured as a bilevel optimization. However, many existing algorithms designed for addressing nonsmooth lower-level problems involve solving sequential subproblems with high complexity. To tackle this challenge, we introduce penalty methods for solving HO based on strong duality between the lower level problem and its dual. We illustrate that the penalized problem closely approximates the optimal solutions of the original HO under certain conditions. In many real applications, the penalized problem is a weakly-convex objective with proximal-friendly constraints. Furthermore, we develop two fully first-order algorithms to solve the penalized problems. Theoretically, we prove the convergence of the proposed algorithms. We demonstrate the efficiency and superiority of our method across numerical experiments.", "keywords": [ "Bilevel Optimization", "Hyperparameter Optimization", "Nonsmooth Optimization" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 1, 3, 1 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "vI5cjHMzP4", "primary_area": "causal reasoning", "n_author": 2, "author_info_list": [ { "name": "Elias Bareinboim", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Associate Professor" }, { "name": "Junzhe Zhang", "gender": "Male", "institution": "Syracuse University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 66, "n_formula_1": 25, "n_ref_uni": 67, "n_ref": 116, "n_ref_all": 174, "n_fig": 20, "n_tab": 4, "L_tab": 196, "n_element_tab": 7, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7140, "formula_len_all_1": 2494, "len_all": 171593, "len_all_1": 66543, "len_abs": 1298, "len_title": 130, "len_sents": 51597, "len_sents_1": 31109, "n_sents": 397, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1301, "title": "Eligibility Traces for Confounding Robust Off-Policy Evaluation: A Causal Approach", "abs": "A unifying theme in Artificial Intelligence is learning an effective policy to control an agent in an unknown environment in order to optimize a certain performance measure. Off-policy methods can significantly improve the sample efficiency during training since they allow an agent to learn from observed trajectories generated by different behavior policies, without directly deploying the target policies in the underlying environment. This paper studies off-policy evaluation from biased offline data where (1) unobserved confounding bias cannot be ruled out a priori; or (2) the observed trajectories do not overlap with intended behaviors of the learner, i.e., the target and behavior policies do not share a common support. Specifically, we first extend the Bellman's equation to derive effective closed-form bounds over value functions from the observational distribution contaminated with unobserved confounding and no-overlap. Second, we propose two novel algorithms that use eligibility traces to estimate these bounds from finite observational data. Compared to other partial identification methods for off-policy evaluation in sequential environments, these methods are model-free and do not rely on additional parametric knowledge about the system dynamics in the underlying environment.", "keywords": [ "Causal Inference", "Graphical Models" ], "rating_list": [ 3, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 2, 3, 4, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "vHO9mU87dc", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Beidi Chen", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Hanshi Sun", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Intern" }, { "name": "Harry Dong", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Li-Wen Chang", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Ningxin Zheng", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Size Zheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Wenlei Bao", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Yuejie Chi", "gender": "Female", "institution": "Meta AI", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 35, "n_ref": 88, "n_ref_all": 132, "n_fig": 17, "n_tab": 16, "L_tab": 9955, "n_element_tab": 777, "n_fig_1": 11, "n_tab_1": 6, "L_tab_1": 3371, "n_element_tab_1": 245, "formula_len_all": 1085, "formula_len_all_1": 872, "len_all": 183895, "len_all_1": 62943, "len_abs": 1463, "len_title": 124, "len_sents": 41318, "len_sents_1": 26012, "n_sents": 312, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1481, "title": "ShadowKV: KV Cache in Shadows for High-Throughput Long-Context LLM Inference", "abs": "With the widespread deployment of long-context large language models (LLMs), there has been a growing demand for efficient support of high-throughput inference. However, as the key-value (KV) cache expands with the sequence length, the increasing memory footprint and the need to access it for each token generation both result in low throughput when serving long-context LLMs. While various dynamic sparse attention methods have been proposed to speed up inference while maintaining generation quality, they either fail to sufficiently reduce GPU memory consumption or introduce significant decoding latency by offloading the KV cache to the CPU. We present ShadowKV, a high-throughput long-context LLM inference system that stores the low-rank key cache and offloads the value cache to reduce the memory footprint for larger batch sizes and longer sequences. To minimize decoding latency, ShadowKV employs an accurate KV selection strategy that reconstructs minimal sparse KV pairs on-the-fly. By evaluating ShadowKV on a broad range of benchmarks, including RULER, LongBench, and Needle In A Haystack, and models like Llama-3.1-8B, Llama-3-8B-1M, GLM-4-9B-1M, Yi-9B-200K, Phi-3-Mini-128K, and Qwen2-7B-128K, we demonstrate that it can support up to 6$\\times$ larger batch sizes and boost throughput by up to 3.04$\\times$ on an A100 GPU without sacrificing accuracy, even surpassing the performance achievable with infinite batch size under the assumption of infinite GPU memory.", "keywords": [ "Long-Context LLM Inference", "KV Cache Optimization" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "vG9dVXwXQV", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Hao-Zhe Tan", "gender": "unknown", "institution": "Nanjing University", "country": "", "position": "MS student" }, { "name": "Lan-Zhe Guo", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhi Zhou", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Yu-Feng Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 14, "n_ref_uni": 42, "n_ref": 70, "n_ref_all": 81, "n_fig": 4, "n_tab": 4, "L_tab": 3002, "n_element_tab": 405, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 482, "n_element_tab_1": 5, "formula_len_all": 854, "formula_len_all_1": 822, "len_all": 127636, "len_all_1": 54365, "len_abs": 1294, "len_title": 122, "len_sents": 32794, "len_sents_1": 25876, "n_sents": 217, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1319, "title": "Pre-Trained Vision-Language Model Selection and Reuse for Downstream Tasks", "abs": "Pre-trained Vision-Language Models (VLMs) are becoming increasingly popular across various visual tasks, and several open-sourced VLM variants have been released. However, selecting the best-performing pre-trained VLM for a specific downstream task is challenging since no single VLM can achieve promising performance on all downstream tasks, and evaluating all available VLMs is impossible due to time and data limitations. To address this problem, this paper proposes a novel paradigm to select and reuse VLM for downstream tasks, called Model Label Learning (MLL). The proposal contains three key modules: \\emph{model labeling}, which assigns labels to each VLM to describe their specialty and utility; \\emph{model selection}, which matches the requirements of the target task with model labels; and \\emph{model reuse}, which applies selected VLMs to the target task in an ensemble manner. The proposal is highly computationally efficient and growable since the model labeling process is completed target task independent and the ability could grow with the number of candidate VLMs. We also introduce a new benchmark for evaluating VLM selection methods, including 49 VLMs and 17 target task datasets. Experimental results clearly demonstrate the effectiveness of the proposed method for selecting and reusing VLMs.", "keywords": [ "Vision-Langage Model; Model Selection; Model Reuse" ], "rating_list": [ 8, 6, 5 ], "soundness_list": [ 3, 3, 1 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "vG123yHVVl", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Anh Tuan Tran", "gender": "Male", "institution": "VinAI Research", "country": "VN", "position": "Research Scientist" }, { "name": "Chee Seng Chan", "gender": "Male", "institution": "Universiti Malaya", "country": "MY", "position": "Full Professor" }, { "name": "Chinh Duc La", "gender": "Male", "institution": "VinUniversity", "country": "VN", "position": "Research Assistant" }, { "name": "Nguyen Hung-Quang", "gender": "unknown", "institution": "VinUniversity ", "country": "VN", "position": "Research Assistant" }, { "name": "Sze Jue Yang", "gender": "Male", "institution": "Universiti Malaya", "country": "", "position": "Undergrad student" }, { "name": "Eugene Bagdasaryan", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Khoa Doan", "gender": "unknown", "institution": "VinUniversity", "country": "VN", "position": "Assistant Professor" }, { "name": "Kok Seng Wong", "gender": "Male", "institution": "VinUniversity", "country": "VN", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 57, "n_ref_all": 81, "n_fig": 13, "n_tab": 6, "L_tab": 1168, "n_element_tab": 139, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 3914, "n_element_tab_1": 79, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 138291, "len_all_1": 59755, "len_abs": 1765, "len_title": 97, "len_sents": 38045, "len_sents_1": 29213, "n_sents": 250, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1774, "title": "Synthesizing Physical Backdoor Datasets: An Automated Framework Leveraging Deep Generative Models", "abs": "Backdoor attacks, representing an emerging threat to the integrity of deep neural networks, have garnered significant attention due to their ability to compromise deep learning systems clandestinely. \nWhile numerous backdoor attacks occur within the digital realm, their practical implementation in real-world prediction systems remains limited and vulnerable to disturbances in the physical world. \nConsequently, this limitation has given rise to the development of physical backdoor attacks, where trigger objects manifest as physical entities within the real world. \nHowever, creating the requisite dataset to train or evaluate a physical backdoor model is a daunting task, limiting the backdoor researchers and practitioners from studying such physical attack scenarios. This paper unleashes a framework that empowers backdoor researchers to effortlessly create a malicious, physical backdoor dataset based on advances in generative modeling. Particularly, this framework involves 3 automatic modules: suggesting the suitable physical triggers, generating the poisoned candidate samples (either by synthesizing new samples or editing existing clean samples), and finally refining for the most plausible ones. As such, it effectively mitigates the perceived complexity associated with creating a physical backdoor dataset, transforming it from a daunting task into an attainable objective. Extensive experiment results show that datasets created by our framework enable researchers to achieve an impressive attack success rate on real physical world data and exhibit similar properties compared to previous physical backdoor attack studies. This paper offers researchers a valuable toolkit for studies of physical backdoors, all within the confines of their laboratories.", "keywords": [ "Backdoor Attacks", "Physical Backdoor Attacks", "Data Synthesis", "Automated Framework" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 2 ] }, { "paper_id": "vFgmobsJiZ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Bernhard Schoelkopf", "gender": "Male", "institution": "ELLIS Institute", "country": "DE", "position": "Principal Researcher" }, { "name": "Robert Bamler", "gender": "Male", "institution": "University of Tuebingen", "country": "", "position": "Assistant Professor" }, { "name": "Weiyang Liu", "gender": "Male", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "Postdoc" }, { "name": "Tim Xiao", "gender": "Male", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 6, "n_ref_uni": 62, "n_ref": 114, "n_ref_all": 148, "n_fig": 21, "n_tab": 64, "L_tab": 31830, "n_element_tab": 892, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 484, "n_element_tab_1": 27, "formula_len_all": 2115, "formula_len_all_1": 403, "len_all": 703079, "len_all_1": 91670, "len_abs": 1868, "len_title": 125, "len_sents": 267725, "len_sents_1": 40974, "n_sents": 2454, "n_sents_1": 346, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1422, "title": "Verbalized Machine Learning: Revisiting Machine Learning with Language Models", "abs": "Motivated by the large progress made by large language models (LLMs), we introduce the framework of verbalized machine learning (VML). In contrast to conventional machine learning models that are typically optimized over a continuous parameter space, VML constrains the parameter space to be human-interpretable natural language. Such a constraint leads to a new perspective of function approximation, where an LLM with a text prompt can be viewed as a function parameterized by the text prompt. Guided by this perspective, we revisit classical machine learning problems, such as regression and classification, and find that these problems can be solved by an LLM-parameterized learner and optimizer. The major advantages of VML include (1) easy encoding of inductive bias: prior knowledge about the problem and hypothesis class can be encoded in natural language and fed into the LLM-parameterized learner; (2) automatic model class selection: the optimizer can automatically select a concrete model class based on data and verbalized prior knowledge, and it can update the model class during training; and (3) interpretable learner updates: the LLM-parameterized optimizer can provide explanations for why each learner update is performed. We conduct several studies to empirically evaluate the effectiveness of VML, and hope that VML can serve as a stepping stone to stronger interpretability and trustworthiness in ML.", "keywords": [ "Large Language Models" ], "rating_list": [ 6, 3, 5, 5, 3 ], "soundness_list": [ 4, 1, 2, 2, 2 ], "presentation_list": [ 4, 2, 3, 4, 3 ], "contribution_list": [ 4, 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "vFfVXSP24J", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Chaofan Tao", "gender": "Male", "institution": "The University of Hong Kong", "country": "", "position": "PhD Student" }, { "name": "Huaxiu Yao", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Hui Shen", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Researcher" }, { "name": "Mi Zhang", "gender": "Male", "institution": "The Ohio State University", "country": "", "position": "Associate Professor" }, { "name": "Rossella Arcucci", "gender": "Female", "institution": "Imperial College London ", "country": "", "position": "Senior Lecturer" }, { "name": "Xin Wang", "gender": "Male", "institution": "The Ohio StateUniversity", "country": "US", "position": "PhD student" }, { "name": "Zhenwu Peng", "gender": "Male", "institution": "King's College London, University of London", "country": "", "position": "MS student" }, { "name": "Zhongwei Wan", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "che liu", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Jie Fu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 42, "n_ref": 70, "n_ref_all": 104, "n_fig": 7, "n_tab": 10, "L_tab": 3115, "n_element_tab": 443, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 2875, "n_element_tab_1": 344, "formula_len_all": 291, "formula_len_all_1": 291, "len_all": 167213, "len_all_1": 75017, "len_abs": 1234, "len_title": 138, "len_sents": 41798, "len_sents_1": 31401, "n_sents": 333, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1237, "title": "ECG Instruction Tuning on Multimodal LLMs for Report Generation: Benchmark and Evaluation", "abs": "Electrocardiogram (ECG) is the primary non-invasive diagnostic tool for monitoring cardiac conditions and is crucial in assisting clinicians. Recent studies have concentrated on classifying cardiac conditions using ECG data but have overlooked ECG report generation, which is time-consuming and requires clinical expertise. To automate ECG report generation and ensure its versatility, we propose the Multimodal ECG Instruction Tuning (MEIT) framework, the first attempt to tackle ECG report generation with LLMs and multimodal instructions. To facilitate future research, we establish a benchmark to evaluate MEIT with various LLMs backbones across two large-scale ECG datasets. Our approach uniquely aligns the representations of the ECG signal and the report, and we conduct extensive experiments to benchmark MEIT with nine open-source LLMs using more than 800,000 ECG reports. MEIT's results underscore the superior performance of instruction-tuned LLMs, showcasing their proficiency in quality report generation, zero-shot capabilities, resilience to signal perturbation, and alignment with human expert evaluation. These findings emphasize the efficacy of our MEIT framework and its potential for real-world clinical application.", "keywords": [ "ECG", "Instruction Tuning", "LLMs" ], "rating_list": [ 8, 6, 5, 3, 5 ], "soundness_list": [ 4, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 5, 4 ] }, { "paper_id": "vFanHFE4Qv", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Can Liao", "gender": "Female", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Jinzhuo Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Wei Wu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhengrui Guo", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "zizhen Deng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 13, "n_ref": 18, "n_ref_all": 25, "n_fig": 4, "n_tab": 5, "L_tab": 1913, "n_element_tab": 184, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1198, "n_element_tab_1": 133, "formula_len_all": 346, "formula_len_all_1": 332, "len_all": 91597, "len_all_1": 64699, "len_abs": 2147, "len_title": 121, "len_sents": 34969, "len_sents_1": 30528, "n_sents": 228, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 2039, "title": "Neuron Platonic Intrinsic Representation From Dynamics Using Contrastive Learning", "abs": "The Platonic Representation Hypothesis posits that behind different modalities of data (what we sense or detect), there exists a universal, modality-independent representation of reality. Inspired by this, we treat each neuron as a system, where we can detect the neuron\u2019s multi-segment activity data under different peripheral conditions. We believe that, similar to the Platonic idea, there exists a time-invariant representation behind the different segments of the same neuron, which reflects the intrinsic properties of the neuron\u2019s system. Intrinsic properties include the molecular profiles, brain regions and morphological structure, etc. The optimization objective for obtaining the intrinsic representation of neurons should satisfy two criteria: (I) segments from the same neuron should have a higher similarity than segments from different neurons; (II) the representations should generalize well to out-of-domain data. To achieve this, we employ contrastive learning, treating different segments from the same neuron as positive pairs and segments from different neurons as negative pairs. During the implementation, we chose the VICReg, which uses only positive pairs for optimization but indirectly separates dissimilar samples via regularization terms. To validate the efficacy of our method, we first applied it to simulated neuron population dynamics data generated using the Izhikevich model. We successfully confirmed that our approach captures the type of each neuron as defined by preset hyperparameters. We then applied our method to two real-world neuron dynamics datasets, including spatial transcriptomics-derived neuron type annotations and the brain regions where each neuron is located. The learned representations from our model not only predict neuron type and location but also show robustness when tested on out-of-domain data (unseen animals). This demonstrates the potential of our approach in advancing the understanding of neuronal systems and offers valuable insights for future neuroscience research.", "keywords": [ "representation learning", "biology", "neuroscience", "contrastive learning" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "vFVjJsy3PG", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Cai Zhou", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Muhan Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xingang Peng", "gender": "unknown", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Xiyuan Wang", "gender": "unknown", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Zian Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 63, "n_formula_1": 9, "n_ref_uni": 54, "n_ref": 206, "n_ref_all": 235, "n_fig": 6, "n_tab": 10, "L_tab": 4268, "n_element_tab": 490, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1677, "n_element_tab_1": 295, "formula_len_all": 4735, "formula_len_all_1": 718, "len_all": 239545, "len_all_1": 64931, "len_abs": 1693, "len_title": 123, "len_sents": 78541, "len_sents_1": 28418, "n_sents": 554, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1674, "title": "Geometric Representation Condition Improves Equivariant Molecule Generation", "abs": "Recent advancements in molecular generative models have demonstrated substantial potential in accelerating scientific discovery, particularly in drug design. However, these models often face challenges in generating high-quality molecules, especially in conditional scenarios where specific molecular properties must be satisfied. In this work, we introduce GeoRCG, a general framework to enhance the performance of molecular generative models by integrating geometric representation conditions. We decompose the molecule generation process into two stages: first, generating an informative geometric representation; second, generating a molecule conditioned on the representation. Compared to directly generating a molecule, the relatively easy-to-generate representation in the first-stage guides the second-stage generation to reach a high-quality molecule in a more goal-oriented and much faster way. Leveraging EDM as the base generator, we observe significant quality improvements in unconditional molecule generation on the widely-used QM9 and GEOM-DRUG datasets. More notably, in the challenging conditional molecular generation task, our framework achieves an average 31\\% performance improvement over state-of-the-art approaches, highlighting the superiority of conditioning on semantically rich geometric representations over conditioning on individual property values as in previous approaches. Furthermore, we show that, with such representation guidance, the number of diffusion steps can be reduced to as small as 100 while maintaining superior generation quality than that achieved with 1,000 steps, thereby significantly accelerating the generation process.", "keywords": [ "molecule generation", "equivariant generative models", "representation", "geometric deep learning", "diffusion models" ], "rating_list": [ 8, 3, 5, 5, 6 ], "soundness_list": [ 3, 3, 4, 3, 2 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 4, 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 3, 3 ] }, { "paper_id": "vF4RhEPGtb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Hao Cheng", "gender": "Male", "institution": "Hong Kong University of Science and Technology(Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Jiahang Cao", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jiayan Yang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Undergrad student" }, { "name": "Jindong Gu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Jize Zhang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Kaidi Xu", "gender": "Male", "institution": "Drexel University", "country": "US", "position": "Assistant Professor" }, { "name": "Le Yang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Renjing Xu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Erjia XIAO", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "QIANG ZHANG", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 47, "n_ref": 112, "n_ref_all": 132, "n_fig": 8, "n_tab": 9, "L_tab": 1336, "n_element_tab": 273, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2545, "n_element_tab_1": 367, "formula_len_all": 290, "formula_len_all_1": 290, "len_all": 170731, "len_all_1": 62576, "len_abs": 1341, "len_title": 167, "len_sents": 47383, "len_sents_1": 26389, "n_sents": 418, "n_sents_1": 159, "venue": "ICLR2025", "decision": "withdraw", "L_title": 118, "L_abs": 1243, "title": "Typography Leads Semantic Diversifying: Amplifying Adversarial Transferability across Multimodal Large Language Models", "abs": "Recently, Multimodal Large Language Models (MLLMs) have demonstrated exceptional performance in zero-shot tasks through their advanced cross-modal interaction and comprehension abilities. Despite these capabilities, MLLMs remain vulnerable to human-imperceptible adversarial examples. In real-world scenarios, the transferability of adversarial examples, which enables cross-model impact, is considered their most significant threat. However, systematic research on the threat of cross-MLLM adversarial transferability is currently lacking. Therefore, this paper serves as the first step toward a comprehensive evaluation of the transferability of adversarial examples generated by various MLLMs. Furthermore, we leverage two critical factors that significantly impact transferability: 1) the degree of information diversity involved in the adversarial generation; 2) the integration of cross vision-language modality editing. We propose a boosting method, the Typography Augment Transferability Method (TATM), to explore adversarial transferability across MLLMs. Through extensive experimental validation, our TATM demonstrates exceptional performance in real-world applications of Harmful Word Insertion and Important Information Protection.", "keywords": [ "Adversarial Transferability; Multimodal Large Language Models; Data Augmentation" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "vEtDApqkNR", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Xiuding Cai", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xueyao Wang", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yaoyao Zhu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yu Yao", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 40, "n_ref": 95, "n_ref_all": 112, "n_fig": 3, "n_tab": 6, "L_tab": 1529, "n_element_tab": 178, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 2384, "n_element_tab_1": 125, "formula_len_all": 762, "formula_len_all_1": 638, "len_all": 137839, "len_all_1": 64687, "len_abs": 1341, "len_title": 133, "len_sents": 31244, "len_sents_1": 28482, "n_sents": 239, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1350, "title": "MambaTS: Improved Selective State Space Models for Long-term Time Series Forecasting", "abs": "In recent years, Transformers have become the de-facto architecture for long-term sequence forecasting (LTSF), yet they face challenges associated with the self-attention mechanism, including quadratic complexity and permutation invariant bias. This raises an important question: \\emph{do we truly need the self-attention mechanism to establish long-range dependencies in LTSF?} Recognizing the significance of causal relationships in multivariate LTSF, we propose MambaTS, which leverages causal relationships to model global dependencies across time and variables through a single linear scan. However, causal graphs are often unknown. To address this, we introduce variable-aware scan along time (VAST), which dynamically discovers variable relationships during training and decodes the optimal variable scan order by solving the shortest path visiting all nodes problem during inference. MambaTS employs the latest Mamba model as its backbone. We suggest that the causal convolution in Mamba is unnecessary due to the presence of independent variables, leading to the development of the Temporal Mamba Block (TMB). To mitigate model overfitting, we further incorporate a dropout mechanism for selective parameters in TMB. Extensive experiments conducted on eight public datasets demonstrate that MambaTS achieves new state-of-the-art performance.", "keywords": [ "Time Series Forcasting; State Space Model" ], "rating_list": [ 6, 8, 3, 6, 5 ], "soundness_list": [ 2, 3, 1, 3, 2 ], "presentation_list": [ 1, 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3, 2 ] }, { "paper_id": "vErsELb7Qg", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Chun Yuan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Yongxian Wei", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Zixuan Hu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Zhenyi Wang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Postdoc" } ], "n_formula": 21, "n_formula_1": 14, "n_ref_uni": 54, "n_ref": 88, "n_ref_all": 119, "n_fig": 9, "n_tab": 12, "L_tab": 5592, "n_element_tab": 668, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 3059, "n_element_tab_1": 345, "formula_len_all": 1245, "formula_len_all_1": 845, "len_all": 201398, "len_all_1": 78831, "len_abs": 1568, "len_title": 156, "len_sents": 52096, "len_sents_1": 31462, "n_sents": 460, "n_sents_1": 265, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1577, "title": "LoRA Recycle: Towards Fine-Tuning-Free Visual Foundation Model via Double-Efficient Data-Free Meta-Learning", "abs": "Large Language Models (LLMs) such as ChatGPT can efficiently adapt to few-shot tasks without fine-tuning, making them ideal for data-limited applications requiring real-time responses. However, this adaptability has not yet been replicated in current Visual Foundation Models (VFMs), which require explicit fine-tuning with sufficient tuning data. Low-Rank Adaptation (LoRA), an effective fine-tuning approach, adapts VFMs to specific tasks by updating extra lightweight modules. Thanks to its modularity, users can upload locally tuned LoRAs to public repositories without exposing private training data. In this paper, we explore the potential of reusing diverse pre-tuned LoRAs without accessing their private training data, to improve the few-shot adaptability of VFMs without requiring further fine-tuning. To achieve this, we propose a data-free meta-learning framework named LoRA Recycle, which distills a meta-LoRA from diverse pre-tuned LoRAs using synthetic data generated via LoRA Inversion. The VFM, once equipped with the meta-LoRA, is empowered to solve new few-shot tasks in a single forward pass without further fine-tuning, akin to the in-context learning of LLMs. To further enhance efficiency, we propose a double-efficient mechanism that uses only the foreground patches and prunes background patches in the synthetic data, significantly accelerating the meta-training process while maintaining or even improving performance. Comprehensive experiments across eight datasets within both in- and cross-domain scenarios verify the superiority of our framework.", "keywords": [ "data-free meta-learning", "few-shot classification", "synthetic data" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "vDp6StrKIq", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Gerrit Gerhartz", "gender": "unknown", "institution": "Heidelberg University, Ruprecht-Karls-Universit\u00e4t Heidelberg", "country": "", "position": "MS student" }, { "name": "Peter Lippmann", "gender": "unknown", "institution": "Heidelberg University, Ruprecht-Karls-Universit\u00e4t Heidelberg", "country": "DE", "position": "PhD student" }, { "name": "Roman Remme", "gender": "Male", "institution": "Heidelberg University, Heidelberg University", "country": "", "position": "PhD student" }, { "name": "Fred A Hamprecht", "gender": "Male", "institution": "Heidelberg University", "country": "", "position": "Full Professor" } ], "n_formula": 43, "n_formula_1": 19, "n_ref_uni": 37, "n_ref": 67, "n_ref_all": 113, "n_fig": 7, "n_tab": 5, "L_tab": 3910, "n_element_tab": 338, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1395, "n_element_tab_1": 150, "formula_len_all": 2967, "formula_len_all_1": 1396, "len_all": 168366, "len_all_1": 69804, "len_abs": 1207, "len_title": 131, "len_sents": 53577, "len_sents_1": 31507, "n_sents": 423, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1221, "title": "Beyond Canonicalization: How Tensorial Messages Improve Equivariant Message Passing", "abs": "In numerous applications of geometric deep learning, the studied systems exhibit spatial symmetries and it is desirable to enforce these. For the symmetry of global rotations and reflections, this means that the model should be equivariant with respect to the transformations that form the group of $\\mathrm O(d)$.\nWhile many approaches for equivariant message passing require specialized architectures, including non-standard normalization layers or non-linearities, we here present a framework based on local reference frames (\"local canonicalization\") which can be integrated with any architecture without restrictions.\nWe enhance equivariant message passing based on local canonicalization by introducing tensorial messages to communicate geometric information consistently between different local coordinate frames.\nOur framework applies to message passing on geometric data in Euclidean spaces of arbitrary dimension.\nWe explicitly show how our approach can be adapted to make a popular existing point cloud architecture equivariant. We demonstrate the superiority of tensorial messages and achieve state-of-the-art results on normal vector regression and competitive results on other standard 3D point cloud tasks.", "keywords": [ "equivariance", "message passing", "tensor representation", "local frames", "geometric deep learning" ], "rating_list": [ 6, 8, 5 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "vDecbmWf6w", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Georg Martius", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "Full Professor" }, { "name": "Jonas Frey", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Marco Bagatella", "gender": "unknown", "institution": "Max Planck Institute for Intelligent Systems, Max Planck Institute for Intelligent Systems", "country": "DE", "position": "PhD student" }, { "name": "Nico G\u00fcrtler", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Thomas Rupf", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" } ], "n_formula": 32, "n_formula_1": 27, "n_ref_uni": 57, "n_ref": 113, "n_ref_all": 140, "n_fig": 8, "n_tab": 6, "L_tab": 4270, "n_element_tab": 192, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 1792, "n_element_tab_1": 49, "formula_len_all": 1939, "formula_len_all_1": 1815, "len_all": 215599, "len_all_1": 71535, "len_abs": 937, "len_title": 106, "len_sents": 49684, "len_sents_1": 29236, "n_sents": 411, "n_sents_1": 313, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 942, "title": "Zero-Shot Offline Imitation Learning via Optimal Transport", "abs": "Zero-shot imitation learning algorithms hold the promise of reproducing unseen behavior from as little as a single demonstration at test time.\nExisting practical approaches view the expert demonstration as a sequence of goals, enabling imitation with a high-level goal selector, and a low-level goal-conditioned policy. \nHowever, this framework can suffer from myopic behavior: the agent's immediate actions towards achieving individual goals may undermine long-term objectives.\nWe introduce a novel method that mitigates this issue by directly optimizing the occupancy matching objective that is intrinsic to imitation learning. \nWe propose to lift a goal-conditioned value function to a distance between occupancies, which are in turn approximated via a learned world model.\nThe resulting method can learn from offline, suboptimal data, and is capable of non-myopic, zero-shot imitation, as we demonstrate in complex, continuous benchmarks.", "keywords": [ "Imitation Learning", "Deep Reinforcement Learning", "Optimal Transport" ], "rating_list": [ 3, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 4, 2, 4 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 3, 2, 2, 3 ] }, { "paper_id": "vC7AlY1ytz", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Huaiyuan Xu", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Postdoc" }, { "name": "Junliang Chen", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Lap-pui Chau", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "YI WANG", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Research Assistant Professor" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 33, "n_ref": 74, "n_ref_all": 96, "n_fig": 10, "n_tab": 9, "L_tab": 2004, "n_element_tab": 161, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 602, "n_element_tab_1": 54, "formula_len_all": 443, "formula_len_all_1": 374, "len_all": 128623, "len_all_1": 61136, "len_abs": 1322, "len_title": 170, "len_sents": 34858, "len_sents_1": 27520, "n_sents": 279, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 129, "L_abs": 1353, "title": "OccProphet: Pushing the Efficiency Frontier of Camera-Only 4D Occupancy Forecasting with an Observer-Forecaster-Refiner Framework", "abs": "Predicting variations in complex traffic environments is crucial for the safety of autonomous driving. Recent advancements in occupancy forecasting have enabled forecasting future 3D occupied status in driving environments by observing historical 2D images. However, high computational demands make occupancy forecasting less efficient during training and inference stages, hindering its feasibility for deployment on edge agents. In this paper, we propose a novel framework, *i.e.*, OccProphet, to efficiently and effectively learn occupancy forecasting with significantly lower computational requirements while maintaining forecasting accuracy. OccProphet comprises three lightweight components: Observer, Forecaster, and Refiner. The Observer extracts spatio-temporal features from 3D using the proposed Efficient 4D Aggregation with Tripling-Attention Fusion, while the Forecaster and Refiner conditionally predict and refine future occupancy inferences. Experimental results on nuScenes, Lyft-Level5, and nuScenes-Occupancy datasets demonstrate that OccProphet is both training- and inference-friendly. OccProphet reduces 58\\%$\\sim$78\\% of the computational cost with a $2.6 \\times$ speedup compared with the state-of-the-art Cam4DOcc. Moreover, it achieves 4\\%$\\sim$18\\% relatively higher forecasting accuracy. The code will be publicly available.", "keywords": [ "camera-only occupancy forecasting", "efficiency", "effectiveness", "autonomous driving" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "vAuodZOQEZ", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Hao Sun", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Huaguan Chen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Yang Liu", "gender": "Female", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" } ], "n_formula": 31, "n_formula_1": 24, "n_ref_uni": 22, "n_ref": 65, "n_ref_all": 126, "n_fig": 29, "n_tab": 12, "L_tab": 3316, "n_element_tab": 224, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 540, "n_element_tab_1": 17, "formula_len_all": 2046, "formula_len_all_1": 1455, "len_all": 141788, "len_all_1": 56438, "len_abs": 1122, "len_title": 125, "len_sents": 45960, "len_sents_1": 25125, "n_sents": 379, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 33, "L_abs": 1126, "title": "Physics-Informed Neural Predictor", "abs": "Accurately predicting fluid dynamics and evolution has been a long-standing challenge in physical sciences. Conventional deep learning methods often rely on the nonlinear modeling capabilities of neural networks to establish mappings between past and future states, overlooking the fluid dynamics, or only modeling the velocity field, neglecting the coupling of multiple physical quantities. In this paper, we propose a new physics-informed learning approach that incorporates coupled physical quantities into the prediction process to assist with forecasting. Central to our method lies in the discretization of physical equations, which are directly integrated into the model architecture and loss function. This integration enables the model to provide robust, long-term future predictions. By incorporating physical equations, our model demonstrates temporal extrapolation and spatial generalization capabilities. Experimental results show that our approach achieves the state-of-the-art performance in spatiotemporal prediction across both numerical simulations and real-world extreme-precipitation nowcasting benchmarks.", "keywords": [ "Fluid dynamics", "Spatiotemporal prediction", "Physics-informed learning" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "vAoyZWyDEc", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "K Lakshmanan", "gender": "Male", "institution": "IIT BHU", "country": "", "position": "Assistant Professor" } ], "n_formula": 28, "n_formula_1": 24, "n_ref_uni": 7, "n_ref": 16, "n_ref_all": 21, "n_fig": 1, "n_tab": 1, "L_tab": 31, "n_element_tab": 3, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2296, "formula_len_all_1": 1921, "len_all": 80518, "len_all_1": 66330, "len_abs": 537, "len_title": 92, "len_sents": 28843, "len_sents_1": 26451, "n_sents": 352, "n_sents_1": 321, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 540, "title": "Approximating Optima of Nonconvex Functions", "abs": "We study the computability of approximating optima of non-convex functions. We give a simple proof to show that the problem of finding the optimal value (and optimal point) or its approximation is not even computable in the oracle setting. We also give a property a function has to satisfy if its global optima can be approximated. Next we give an example of such a global property we call basin of attraction. Then we give a simple algorithm which converges to the global optima when this is known. Finally, we give some numerical results.", "keywords": [ "Computablity of Approximate Optima", "Non-convex functions" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 1, 1, 1, 1 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 4, 3, 2, 5 ] }, { "paper_id": "v9fQfQ85oG", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Lingjie Duan", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "", "position": "Associate Professor" }, { "name": "Pengcheng Dai", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "", "position": "Postdoc" } ], "n_formula": 55, "n_formula_1": 57, "n_ref_uni": 22, "n_ref": 39, "n_ref_all": 59, "n_fig": 7, "n_tab": 1, "L_tab": 294, "n_element_tab": 12, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 728, "n_element_tab_1": 17, "formula_len_all": 4610, "formula_len_all_1": 4395, "len_all": 96420, "len_all_1": 65648, "len_abs": 1507, "len_title": 135, "len_sents": 29041, "len_sents_1": 24787, "n_sents": 202, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1675, "title": "Multi-objective Multi-agent Reinforcement Learning with Pareto-stationary Convergence", "abs": "Multi-objective multi-agent reinforcement learning (MOMARL) problems frequently arise in real world applications (e.g., path planning for swarm robots) or have not been explored well. To find Pareto-optimum is NP-hard, and thus some multi-objective algorithms have emerged recently to provide Pareto-stationary solution centrally, managed by a single agent. Yet, they cannot deal with MOMARL problem, as the dimension of global state-action $(\\boldsymbol{s},\\boldsymbol{a})$ grows exponentially with the number of spatially distributed agents. To tackle this issue, we design a novel graph-truncated $Q$-function approximation method for each agent $i$, which does not require the global state-action $(\\boldsymbol{s},\\boldsymbol{a})$ but only the neighborhood state-action $(s\\_{\\mathcal{N}^{\\kappa}\\_{i}},a\\_{\\mathcal{N}^{\\kappa}\\_{i}})$ of its $\\kappa$-hop neighbors. To further reduce the dimension to state-action $(s\\_{\\mathcal{N}^{\\kappa}\\_{i}},a\\_{i})$ with only local action, we further develop a concept of action-averaged $Q$-function and establish the equivalence between using graph-truncated $Q$-function and action-averaged $Q$-function for policy gradient approximation. Accordingly, we develop a distributed scalable algorithm with linear function approximation and we prove that it successfully converges Pareto-stationary solution at rate $\\mathcal{O}(1/T)$ that is inversely proportional to time domain $T$. Finally, we run simulations in a robot path planning environment and show our algorithm converges to greater multi-objective values as compared to the latest MORL algorithm, and performs close to the central optimum with much shorter running time.", "keywords": [ "Multi-objective", "multi-agent reinforcement learning", "Pareto-stationary convergence" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 2, 4, 3 ] }, { "paper_id": "v9LjNopQ6W", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Gyeong-Moon Park", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Assistant Professor" }, { "name": "Jinju Kim", "gender": "Female", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "MS student" }, { "name": "Jong Hwan Ko", "gender": "unknown", "institution": "Sungkyunkwan University", "country": "KR", "position": "Assistant Professor" }, { "name": "Taesoo Kim", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "PhD student" }, { "name": "KimDongChan", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "MS student" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 51, "n_ref": 92, "n_ref_all": 109, "n_fig": 6, "n_tab": 7, "L_tab": 1038, "n_element_tab": 169, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 634, "n_element_tab_1": 110, "formula_len_all": 858, "formula_len_all_1": 690, "len_all": 168065, "len_all_1": 62815, "len_abs": 1224, "len_title": 126, "len_sents": 49014, "len_sents_1": 29654, "n_sents": 339, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1228, "title": "Do Not Mimic My Voice: Teacher-Guided Unlearning for Zero-Shot Text-to-Speech", "abs": "The rapid advancement of Zero-Shot Text-to-Speech (ZS-TTS) technology has enabled high-fidelity voice synthesis from minimal audio cues, raising significant privacy and ethical concerns. In particular, the ability to replicate an individual\u2019s voice without consent poses risks, highlighting the need for machine unlearning techniques to protect voice privacy. In this paper, we introduce the first machine unlearning framework for ZS-TTS, Teacher-Guided Unlearning (TGU), designed to ensure that the model forgets designated speaker identities while retaining its ability to generate accurate speech for other speakers. Unlike conventional unlearning methods, TGU leverages randomness to prevent consistent replication of forget speakers' voices, ensuring unlearned identities remain untraceable. Additionally, we propose a new evaluation metric, speaker-Zero Retrain Forgetting (spk-ZRF), which measures the model\u2019s effectiveness in preventing the reproduction of forgotten voices. The experiments conducted on the state-of-the-art model demonstrate that TGU prevents the model from replicating forget speakers' voices while maintaining high quality for other speakers. The demo is available at https://speechunlearn.github.io/", "keywords": [ "zero-shot tts", "machine unlearning", "voice privacy" ], "rating_list": [ 3, 8, 3, 5 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "v9GwGQoOG5", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Tianyi Huang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Postdoc" }, { "name": "Xiaolong Qin", "gender": "Male", "institution": "Hangzhou Normal University", "country": "", "position": "Full Professor" }, { "name": "Xin Yuan", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhiling Cai", "gender": "Male", "institution": "Fujian Agriculture and Forestry University", "country": "CN", "position": "Lecturer" } ], "n_formula": 63, "n_formula_1": 17, "n_ref_uni": 65, "n_ref": 101, "n_ref_all": 132, "n_fig": 16, "n_tab": 9, "L_tab": 3431, "n_element_tab": 222, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1079, "n_element_tab_1": 32, "formula_len_all": 6188, "formula_len_all_1": 1349, "len_all": 184361, "len_all_1": 59156, "len_abs": 1630, "len_title": 137, "len_sents": 46504, "len_sents_1": 26376, "n_sents": 399, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1637, "title": "Beyond Markov Assumption: Improving Sample Efficiency in MDPs by Historical Augmentation", "abs": "Under the Markov assumption of Markov Decision Processes (MDPs), an optimal stationary policy does not need to consider history and is no worse than any non-stationary or history-dependent policy. Therefore, existing Deep Reinforcement Learning (DRL) algorithms usually model sequential decision-making as an MDP and then try to optimize a stationary policy by single-step state transitions. However, such optimization is often faced with sample inefficiency when the causal relationships of state transitions are complex. To address the above problem, this paper investigates if augmenting the states with their historical information can simplify the complex causal relationships in MDPs and thus improve the sample efficiency of DRL. First, we demonstrate that a complex causal relationship of single-step state transitions may be inferred by a simple causal function of the historically augmented states. Then, we propose a convolutional neural network architecture to learn the representation of the current state and its historical trajectory. The main idea of this representation learning is to compress the high-dimensional historical trajectories into a low-dimensional space. In this way, we can extract the simple causal relationships from historical information and avoid the overfitting caused by high-dimensional data. Finally, we formulate Historical Augmentation Aided Actor-Critic (HA3C) algorithm by adding the learned representations to the actor-critic method. The experiment on standard MDP tasks demonstrates that HA3C outperforms current state-of-the-art methods in terms of both sample efficiency and performance.", "keywords": [ "Deep reinforcement learning", "Sample efficiency", "State representation", "Historical augmentation", "Markov decision processes" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "v9EjwMM55Y", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Hongyang Chen", "gender": "Male", "institution": "Zhejiang Lab, China", "country": "CN", "position": "Senior Research Expert" }, { "name": "Mingqian Li", "gender": "unknown", "institution": "Zhejiang Lab, Zhejiang Lab", "country": "CN", "position": "Postdoc" }, { "name": "Qiang Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ruifeng Li", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Wei Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xiangxin Zhou", "gender": "Not Specified", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yuan Yao", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Yuhua Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 17, "n_ref_uni": 77, "n_ref": 167, "n_ref_all": 183, "n_fig": 5, "n_tab": 5, "L_tab": 1484, "n_element_tab": 109, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1013, "formula_len_all_1": 818, "len_all": 229486, "len_all_1": 60253, "len_abs": 1586, "len_title": 122, "len_sents": 57881, "len_sents_1": 26820, "n_sents": 422, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 74, "L_abs": 1443, "title": "UniMatch: Universal Matching from Atom to Task for Few-Shot Drug Discovery", "abs": "Drug discovery is crucial for identifying candidate drugs for various diseases. However, its low success rate often results in a scarcity of annotations, posing a few-shot learning problem. Existing methods primarily focus on single-scale features, overlooking the hierarchical molecular structures that determine different molecular properties. To address these issues, we introduce Universal Matching Networks (UniMatch), a dual matching framework that integrates explicit hierarchical molecular matching with implicit task-level matching via meta-\nlearning, bridging multi-level molecular representations and task-level generalization. Specifically, our approach explicitly captures structural features across multiple levels\u2014atoms, substructures, and molecules\u2014via hierarchical pooling and matching, facilitating precise molecular representation and comparison. Additionally, we employ a meta-learning strategy for implicit task-level matching, allowing the model to capture shared patterns across tasks and quickly adapt to new ones. This unified matching framework ensures effective molecular alignment while leveraging shared meta-knowledge for fast adaptation. Our experimental results demonstrate that UniMatch outperforms state-of-the-art methods on the MoleculeNet and FS-Mol benchmarks, achieving improvements of 2.87% in AUROC and 6.52% in \u2206AUPRC. UniMatch also shows excellent generalization ability on the Meta-MolNet benchmark.", "keywords": [ "Few-shot molecular representation learning", "maching learning" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 2, 4, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "v9CDpLpjiE", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Lei Zhang", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Chair Professor" }, { "name": "Minheng Ni", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Wangmeng Zuo", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "YuTao Fan", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "Undergrad student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 24, "n_ref": 49, "n_ref_all": 79, "n_fig": 5, "n_tab": 18, "L_tab": 6003, "n_element_tab": 430, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2167, "n_element_tab_1": 211, "formula_len_all": 358, "formula_len_all_1": 360, "len_all": 169966, "len_all_1": 62300, "len_abs": 2024, "len_title": 151, "len_sents": 54400, "len_sents_1": 27065, "n_sents": 433, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 1493, "title": "Visual-O1: Understanding Ambiguous Instructions via Multi-modal Multi-turn Chain-of-thoughts Reasoning", "abs": "As large-scale models evolve, language instructions are increasingly utilized in multi-modal tasks. Due to human language habits, these instructions often contain ambiguities in real-world scenarios, necessitating the integration of visual context or common sense for accurate interpretation. However, even highly intelligent large models exhibit observable performance limitations on ambiguous instructions, where weak reasoning abilities of disambiguation can lead to catastrophic errors. To address this issue, this paper proposes Visual-O1, a multi-modal multi-turn chain-of-thought reasoning framework. It simulates human multi-modal multi-turn reasoning, providing instantial experience for highly intelligent models or empirical experience for generally intelligent models to understand ambiguous instructions. Unlike traditional methods that require models to possess high intelligence to understand long texts or perform lengthy complex reasoning, our framework does not notably increase computational overhead and is more general and effective, even for generally intelligent models. Experiments show that our method not only enhances the performance of models of different intelligence levels on ambiguous instructions but also improves their performance on general datasets. Our work highlights the potential of artificial intelligence to work like humans in real-world scenarios with uncertainty and ambiguity. We release our data and code at https://github.com/kodenii/Visual-O1.", "keywords": [ "Understanding ambiguous instructions", "large multimodal model", "chain-of-thoughts", "multimodal" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "v8qABSeeKO", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Chenrui Shi", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Qing Li", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence (BIGAI)", "country": "CN", "position": "Researcher" }, { "name": "Siyuan Qi", "gender": "unknown", "institution": "Beijing Institute for General Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Yuntao Du", "gender": "Male", "institution": "BIGAI", "country": "CN", "position": "Researcher" }, { "name": "Zhi Gao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Postdoc" }, { "name": "Zilong Zheng", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "jiangkailin", "gender": "Male", "institution": "University of Science and Technology of China(USTC)", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 20, "n_ref": 65, "n_ref_all": 84, "n_fig": 6, "n_tab": 14, "L_tab": 12994, "n_element_tab": 1068, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 6486, "n_element_tab_1": 341, "formula_len_all": 228, "formula_len_all_1": 228, "len_all": 175194, "len_all_1": 66974, "len_abs": 1482, "len_title": 120, "len_sents": 50230, "len_sents_1": 27110, "n_sents": 376, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1501, "title": "MMKE-Bench: A Multimodal Editing Benchmark for Diverse Visual Knowledge", "abs": "Knowledge editing techniques have emerged as essential tools for updating the factual knowledge of large language models (LLMs) and multimodal models (LMMs), allowing them to correct outdated or inaccurate information without retraining from scratch. However, existing benchmarks for multimodal knowledge editing primarily focus on entity-level knowledge represented as simple triplets, which fail to capture the complexity of real-world multimodal information. To address this issue, we introduce MMKE-Bench, a comprehensive **M**ulti**M**odal **K**nowledge **E**diting Benchmark, designed to evaluate the ability of LMMs to edit diverse visual knowledge in real-world scenarios. MMKE-Bench addresses these limitations by incorporating three types of editing tasks: visual entity editing, visual semantic editing, and user-specific editing. Besides, MMKE-Bench uses free-form natural language to represent and edit knowledge, offering a more flexible and effective format. The benchmark consists of 2,940 pieces of knowledge and 8,363 images across 33 broad categories, with evaluation questions automatically generated and human-verified. We assess five state-of-the-art knowledge editing methods on three prominent LMMs, revealing that no method excels across all criteria, and that visual and user-specific edits are particularly challenging. MMKE-Bench sets a new standard for evaluating the robustness of multimodal knowledge editing techniques, driving progress in this rapidly evolving field.", "keywords": [ "Multimodal knowledge editing; Large multimodal model; Benchmark" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "v8RDgaEtE2", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Ashok Veeraraghavan", "gender": "unknown", "institution": "William Marsh Rice University", "country": "US", "position": "Full Professor" }, { "name": "Guha Balakrishnan", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Assistant Professor" }, { "name": "Laurence Edward Court", "gender": "Male", "institution": "MD Anderson Cancer Center", "country": "", "position": "Associate Professor" }, { "name": "Matt Y. Cheung", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Tucker James Netherton", "gender": "Male", "institution": "MD Anderson Cancer Center", "country": "", "position": "Assistant Professor" } ], "n_formula": 52, "n_formula_1": 24, "n_ref_uni": 46, "n_ref": 82, "n_ref_all": 97, "n_fig": 7, "n_tab": 1, "L_tab": 3, "n_element_tab": 1, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 3, "n_element_tab_1": 1, "formula_len_all": 5339, "formula_len_all_1": 1676, "len_all": 136561, "len_all_1": 61315, "len_abs": 2074, "len_title": 90, "len_sents": 37902, "len_sents_1": 27924, "n_sents": 312, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 1931, "title": "Regression Conformal Prediction under Bias", "abs": "Uncertainty quantification is crucial to account for the imperfect predictions of machine learning algorithms for high-impact applications. Conformal prediction (CP) is a powerful framework for uncertainty quantification that generates calibrated prediction intervals with valid coverage. \nIn this work, we study how CP intervals are affected by \\emph{bias} -- the systematic deviation of a prediction from ground truth values -- a phenomenon prevalent in many real-world applications.\nWe investigate the influence of bias on interval lengths of two different types of adjustments -- symmetric adjustments, the conventional method where both sides of the interval are adjusted equally, and asymmetric adjustments, a more flexible method where the interval can be adjusted unequally in positive or negative directions.\nWe present theoretical and empirical analyses characterizing how symmetric and asymmetric adjustments impact the \"tightness\" of CP intervals for regression tasks. \nSpecifically for absolute residual and quantile-based non-conformity scores, we prove: 1) the upper bound of symmetrically adjusted interval lengths increases by $2|b|$ where $b$ is a globally applied scalar value representing bias, 2) asymmetrically adjusted interval lengths are not affected by bias, and 3) conditions when asymmetrically adjusted interval lengths are guaranteed to be smaller than symmetric ones.\nOur analyses suggest that even if predictions exhibit significant drift from ground truth values, asymmetrically adjusted intervals are still able to maintain the same tightness and validity of intervals as if the drift had never happened, while symmetric ones significantly inflate the lengths. \nWe demonstrate our theoretical results with two real-world prediction tasks: sparse-view computed tomography (CT) reconstruction and time-series weather forecasting. Our work paves the way for more bias-robust machine learning systems.", "keywords": [ "Conformal Prediction", "Bias", "Uncertainty Quantification" ], "rating_list": [ 1, 1, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "v8GuB74YRA", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Di Huang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Huiqun Wang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Nan Zhou", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Yaoyan Zheng", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 11, "n_ref_uni": 35, "n_ref": 55, "n_ref_all": 66, "n_fig": 6, "n_tab": 5, "L_tab": 3703, "n_element_tab": 425, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1651, "n_element_tab_1": 216, "formula_len_all": 1107, "formula_len_all_1": 700, "len_all": 113131, "len_all_1": 57487, "len_abs": 1421, "len_title": 138, "len_sents": 30591, "len_sents_1": 26081, "n_sents": 211, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1430, "title": "Generalizable Transferability Estimation of Foundation Vision Models via Implicit Learning", "abs": "Transferability estimation aims to identify the most suitable model from a collection of pre-trained models for specific downstream tasks, playing a crucial role in the success of the pre-training and fine-tuning paradigm. However, the recent proliferation of pre-trained models with diverse architectures and training strategies poses significant challenges for transferability estimation due to discrepancies in intrinsic model characteristics, making it difficult for existing methods to accurately simulate embedding space evolution within feasible computational limits. To address these challenges, we propose an Implicit Transferability Modeling (ITM) paradigm that incorporates an implicit modeling strategy for the intrinsic properties of pre-trained models, enabling more accurate transferability estimation. ITM employs a Divide-and-Conquer Adaptation (DCA) process to efficiently model the transfer process, reducing both learning complexity and computational cost. Additionally, we introduce a Pseudo-Clustering-based Optimization (PCO) strategy that eliminates the need for extensive fine-tuning, enabling effective estimation without intensive retraining. Our method significantly outperforms state-of-the-art approaches, achieving notable improvements across ten widely used benchmarks and demonstrating its effectiveness and generalizability in enabling accurate and efficient model selection for downstream tasks.", "keywords": [ "Transferability Estimation", "Transfer Learning" ], "rating_list": [ 5, 1, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3, 2 ], "contribution_list": [ 3, 1, 3, 2, 3 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "v7aeTmfGOu", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Haohan Wang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Haoyang Liu", "gender": "Male", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Shuyu Chen", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Undergrad student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 46, "n_ref": 73, "n_ref_all": 84, "n_fig": 3, "n_tab": 6, "L_tab": 2149, "n_element_tab": 163, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1882, "n_element_tab_1": 154, "formula_len_all": 28, "formula_len_all_1": 0, "len_all": 212643, "len_all_1": 64083, "len_abs": 1805, "len_title": 163, "len_sents": 62159, "len_sents_1": 28728, "n_sents": 444, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 114, "L_abs": 1204, "title": "GenoAgent: A Baseline method for LLM-Based Exploration of Gene Expression Data in Alignment with Bioinformaticians", "abs": "Recent advancements in machine learning have significantly improved the identification of disease-associated genes from gene expression datasets. However, these processes often require extensive expertise and manual effort, limiting their scalability. Large Language Model (LLM)-based agents have shown promise in automating these tasks due to their increasing problem-solving abilities. To leverage the potential of agentic system, we introduce GenoAgent, a team of LLM-based agents designed with context-aware planning, iterative correction, and domain expert consultation to collaboratively explore gene datasets. GenoAgent provides generalized approach for addressing a wide range of gene identification problems, in a completely automated analysis pipeline that follows the standard of computational genomics. Our experiments with GenoAgent demonstrate the potential of LLM-based approaches in genomics data analysis, while error analysis highlights the challenges and areas for future improvement. We also propose GenoTEX, a benchmark dataset for automatic exploration of gene expression data, and also a promising resource for evaluating and enhancing AI-driven methods for genomics data analysis.", "keywords": [ "Multi-agent", "Bioinformatics" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 2, 2, 2, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 5, 3 ] }, { "paper_id": "v7a4KET0Md", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "anqi wu", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Feiyang Wu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jingyang Ke", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jiyi Wang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Jeffrey Evan Markowitz", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 5, "n_ref_uni": 36, "n_ref": 70, "n_ref_all": 95, "n_fig": 17, "n_tab": 2, "L_tab": 499, "n_element_tab": 58, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1879, "formula_len_all_1": 282, "len_all": 162368, "len_all_1": 79163, "len_abs": 1560, "len_title": 161, "len_sents": 56872, "len_sents_1": 39361, "n_sents": 442, "n_sents_1": 289, "venue": "ICLR2025", "decision": "Reject", "L_title": 112, "L_abs": 1567, "title": "Inverse Reinforcement Learning with Switching Rewards and History Dependency for Characterizing Animal Behaviors", "abs": "Traditional approaches to studying decision-making in neuroscience focus on simplified behavioral tasks where animals perform repetitive, stereotyped actions to receive explicit rewards. While informative, these methods constrain our understanding of decision-making to short timescale behaviors driven by explicit goals. In natural environments, animals exhibit more complex, long-term behaviors driven by intrinsic motivations that are often unobservable. Recent works in time-varying inverse reinforcement learning (IRL) aim to capture shifting motivations in long-term, freely moving behaviors. However, a crucial challenge remains: animals make decisions based on their history, not just their current state. To address this, we introduce SWIRL (SWitching IRL), a novel framework that extends traditional IRL by incorporating time-varying, history-dependent reward functions. SWIRL models long behavioral sequences as transitions between short-term decision-making processes, each governed by a unique reward function. SWIRL incorporates biologically plausible history dependency to capture how past decisions and environmental contexts shape behavior, offering a more accurate description of animal decision-making. We apply SWIRL to simulated and real-world animal behavior datasets and show that it outperforms models lacking history dependency, both quantitatively and qualitatively. This work presents the first IRL model to incorporate history-dependent policies and rewards to advance our understanding of complex, naturalistic decision-making in animals.", "keywords": [ "neuroscience", "decision-making", "inverse reinforcement learning" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 1, 2, 4 ], "confidence_list": [ 4, 2, 3 ] }, { "paper_id": "v7YrIjpkTF", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Changdong Wang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "Associate Professor" }, { "name": "Feidiao Yang", "gender": "Male", "institution": "Pengcheng Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Hui Li", "gender": "unknown", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Jianyang Zhai", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "PhD student" }, { "name": "Xiawu Zheng", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Yonghong Tian", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Zi-Feng Mai", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 38, "n_ref": 101, "n_ref_all": 117, "n_fig": 4, "n_tab": 5, "L_tab": 2258, "n_element_tab": 298, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2212, "n_element_tab_1": 297, "formula_len_all": 354, "formula_len_all_1": 354, "len_all": 169954, "len_all_1": 67016, "len_abs": 1804, "len_title": 111, "len_sents": 40334, "len_sents_1": 29801, "n_sents": 294, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1892, "title": "Multimodal Quantitative Language for Generative Recommendation", "abs": "Generative recommendation has emerged as a promising paradigm aiming at directly generating the identifiers of the target candidates.\nMost existing methods attempt to leverage prior knowledge embedded in Pre-trained Language Models (PLMs) to improve the recommendation performance. However, they often fail to accommodate the differences between the general linguistic knowledge of PLMs and the specific needs of recommendation systems. Moreover, they rarely consider the complementary knowledge between the multimodal information of items, which represents the multi-faceted preferences of users. To facilitate efficient recommendation knowledge transfer, we propose a novel approach called Multimodal Quantitative Language for Generative Recommendation (MQL4GRec). Our key idea is to transform items from different domains and modalities into a unified language, which can serve as a bridge for transferring recommendation knowledge. Specifically, we first introduce quantitative translators to convert the text and image content of items from various domains into a new and concise language, known as quantitative language, with all items sharing the same vocabulary. Then, we design a series of quantitative language generation tasks to enrich quantitative language with semantic information and prior knowledge. Finally, we achieve the transfer of recommendation knowledge from different domains and modalities to the recommendation task through pre-training and fine-tuning. We evaluate the effectiveness of MQL4GRec through extensive experiments and comparisons with existing methods, achieving improvements over the baseline by 11.18\\%, 14.82\\%, and 7.95\\% on the NDCG metric across three different datasets, respectively. Our implementation is available at: \\href{https://anonymous.4open.science/r/QL4GRec-ED65/}{\\textcolor{blue}{https://anonymous.4open.science/r/MQL4GRec-ED65/}.}", "keywords": [ "Recommendation System", "Generative Recommendation" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 4, 3, 3, 4 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "v6iLQBoIJw", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Chulhee Yun", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Kwangjun Ahn", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Minhak Song", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Undergrad student" } ], "n_formula": 21, "n_formula_1": 16, "n_ref_uni": 57, "n_ref": 89, "n_ref_all": 186, "n_fig": 46, "n_tab": 2, "L_tab": 474, "n_element_tab": 22, "n_fig_1": 19, "n_tab_1": 1, "L_tab_1": 16, "n_element_tab_1": 6, "formula_len_all": 1538, "formula_len_all_1": 1191, "len_all": 199574, "len_all_1": 65095, "len_abs": 1249, "len_title": 89, "len_sents": 56424, "len_sents_1": 28917, "n_sents": 460, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 1254, "title": "Does SGD really happen in tiny subspaces?", "abs": "Understanding the training dynamics of deep neural networks is challenging due to their high-dimensional nature and intricate loss landscapes. Recent studies have revealed that, along the training trajectory, the gradient approximately aligns with a low-rank top eigenspace of the training loss Hessian, referred to as the dominant subspace. Given this alignment, this paper explores whether neural networks can be trained within the dominant subspace, which, if feasible, could lead to more efficient training methods. Our primary observation is that when the SGD update is projected onto the dominant subspace, the training loss does not decrease further. This suggests that the observed alignment between the gradient and the dominant subspace is spurious. Surprisingly, projecting out the dominant subspace proves to be just as effective as the original update, despite removing the majority of the original update component. We observe similar behavior across practical setups, including the large learning rate regime (also known as Edge of Stability), Sharpness-Aware Minimization, momentum, and adaptive optimizers. We discuss the main causes and implications of this spurious alignment, shedding light on the dynamics of neural network training.", "keywords": [ "optimization for deep networks", "training dynamics", "SGD", "Hessian", "low-rank subspace" ], "rating_list": [ 6, 8, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 2, 3, 1, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "v6NNopExN4", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Adam Dziedzic", "gender": "Male", "institution": "CISPA, saarland university, saarland informatics campus", "country": "DE", "position": "Assistant Professor" }, { "name": "Christopher A. Choquette-Choo", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Franziska Boenisch", "gender": "Female", "institution": "CISPA", "country": "DE", "position": "Assistant Professor" }, { "name": "Jing Xu", "gender": "Female", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Postdoc" }, { "name": "Michael Backes", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Full Professor" }, { "name": "Xun Wang", "gender": "unknown", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 37, "n_ref": 77, "n_ref_all": 115, "n_fig": 7, "n_tab": 12, "L_tab": 2273, "n_element_tab": 273, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 125, "n_element_tab_1": 15, "formula_len_all": 244, "formula_len_all_1": 244, "len_all": 153384, "len_all_1": 65504, "len_abs": 1652, "len_title": 102, "len_sents": 46787, "len_sents_1": 32531, "n_sents": 364, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1643, "title": "POST: A Framework for Privacy of Soft-prompt Transfer", "abs": "Prompting has emerged as a dominant learning paradigm for adapting large language models (LLMs). While discrete (textual) prompts prepend tokens to the input for optimized outputs, soft (parameter) prompts are tuned in the embedding space via backpropagation, requiring less engineering effort. However, unlike semantically meaningful discrete prompts, soft prompts are tightly coupled to the LLM they were tuned on, hindering their generalization to other LLMs. This limitation is particularly problematic when efficiency and privacy are concerns, since (1) it requires tuning new prompts for each LLM which, due to the backpropagation, becomes increasingly computationally expensive as LLMs grow in size, and (2) when the LLM is centrally hosted, it requires sharing private data for soft prompt tuning with the LLM provider. To address these concerns, we propose a framework for Privacy Of Soft-prompt Transfer (POST), a novel method that enables private soft prompt tuning on a small language model and then transfers the prompt to the large LLM. Using knowledge distillation, we first derive the small language model directly from the LLM to facilitate prompt transferability. Then, we tune the soft prompt locally, if required with privacy guarantees, e.g., according to differential privacy. Finally, we use a small set of public data to transfer the prompt from the small model to the large LLM without additional privacy leakage. Our experimental results demonstrate that our method effectively transfers soft prompts, protecting local data privacy and reducing the computational complexity over soft prompt tuning on the large model.", "keywords": [ "prompt transfer", "soft prompt", "privacy", "distillation", "confidentiality" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "v5bK7cQch3", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Li Xiao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Wei Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Xingcan Hu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 28, "n_ref": 68, "n_ref_all": 80, "n_fig": 2, "n_tab": 4, "L_tab": 1045, "n_element_tab": 105, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2466, "n_element_tab_1": 131, "formula_len_all": 287, "formula_len_all_1": 287, "len_all": 122958, "len_all_1": 60279, "len_abs": 1548, "len_title": 165, "len_sents": 30045, "len_sents_1": 28748, "n_sents": 178, "n_sents_1": 143, "venue": "ICLR2025", "decision": "Reject", "L_title": 117, "L_abs": 1557, "title": "Learning 3D Medical Image Models From Brain Functional Connectivity Network Supervision For Mental Disorder Diagnosis", "abs": "For mental disorder diagnosis, most previous works are task-specific and focus primarily on functional connectivity network (FCN) derived from functional MRI (fMRI) data. However, the high cost of fMRI acquisition limits its practicality in real-world clinical settings. Meanwhile, the more easily obtainable 3D T1-weighted (T1w) MRI, which captures brain anatomy, is ofen overlooked in standard diagnostic processes of mental disorders.\nTo address these two issues, we propose CINP (Contrastive Image-Network Pre-training), a framework that employs contrastive learning between 3D T1w MRI and FCNs. CINP aims to learn a joint latent semantic space that integrates complementary information from both functional and structural perspective. During pre-training, we incorporate masked image modeling loss and network-image matching loss to enhance visual representation learning and modality alignment.\nFurthermore, thanks to contrastive pre-training which facilitates knowledge transfer from FCN to T1w MRI, we introduce network prompting. This protocol leverages 3D T1w MRI from suspected patients and FCNs from confirmed patients for differential diagnosis of mental disorders. \nExtensive experiments across three mental disorder diagnosis tasks demonstrate the competitive performance of CINP, using both linear probing and network prompting, compared with FCN-based methods and self-supervised pre-training methods.\nThese results highlight the potential of CINP to enhance diagnostic processes with the aid of 3D T1w MRI in real-world clinical scenario.", "keywords": [ "3D medical image", "functional connectivity network", "contrastive learning", "mental disease diagnosis" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "v5JrYUdMxc", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Haijie LI", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Jian Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Jiarui Meng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Shuzhou Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xiandong MENG", "gender": "Male", "institution": "Pengcheng Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yanmin Wu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yu Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 48, "n_ref": 126, "n_ref_all": 150, "n_fig": 15, "n_tab": 5, "L_tab": 551, "n_element_tab": 66, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 240, "n_element_tab_1": 27, "formula_len_all": 610, "formula_len_all_1": 611, "len_all": 159042, "len_all_1": 64045, "len_abs": 199, "len_title": 205, "len_sents": 45550, "len_sents_1": 29628, "n_sents": 377, "n_sents_1": 246, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1391, "title": "Hybrid Fourier Score Distillation for Efficient One Image to 3D Object Generation", "abs": "Single image-to-3D generation is pivotal for crafting controllable 3D assets. Given its under-constrained nature, we attempt to leverage 3D geometric priors from a novel view diffusion model and 2D appearance priors from an image generation model to guide the optimization process. We note that there is a disparity between the generation priors of these two diffusion models, leading to their different appearance outputs. Specifically, image generation models tend to deliver more detailed visuals, whereas novel view models produce consistent yet over-smooth results across different views. Directly combining them leads to suboptimal effects due to their appearance conflicts. Hence, we propose a 2D-3D **hy**brid **F**ourier **S**core **D**istillation objective function, **hy-FSD**. It optimizes 3D Gaussians using 3D priors in spatial domain to ensure geometric consistency, while exploiting 2D priors in the frequency domain through Fourier transform for better visual quality. hy-FSD can be integrated into existing 3D generation methods and produce significant performance gains. With this technique, we further develop an image-to-3D generation pipeline to create high-quality 3D objects within one minute, named **Fourier123**. Extensive experiments demonstrate that Fourier123 excels in efficient generation with rapid convergence speed and visually-friendly generation results.", "keywords": [ "3D Generation", "One Image to 3D Generation" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "v5BouOktUP", "primary_area": "causal reasoning", "n_author": 9, "author_info_list": [ { "name": "Juntao Wang", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "MS student" }, { "name": "Lai Zhong Yuan", "gender": "Male", "institution": "Ballsnow Technology (Shanghai)", "country": "CN", "position": "Principal Researcher" }, { "name": "Liang Hu", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Full Professor" }, { "name": "Qi Zhang", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Researcher" }, { "name": "Qingqing Cao", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "TINGZHOU WEI", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Undergrad student" }, { "name": "Usman Naseem", "gender": "unknown", "institution": "Macquarie University", "country": "AU", "position": "Assistant Professor" }, { "name": "Zihao Mao", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Zonghan Du", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Research assistant" } ], "n_formula": 23, "n_formula_1": 9, "n_ref_uni": 27, "n_ref": 52, "n_ref_all": 63, "n_fig": 4, "n_tab": 3, "L_tab": 565, "n_element_tab": 42, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 518, "n_element_tab_1": 25, "formula_len_all": 2517, "formula_len_all_1": 596, "len_all": 125362, "len_all_1": 52310, "len_abs": 1339, "len_title": 148, "len_sents": 42536, "len_sents_1": 26007, "n_sents": 279, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1329, "title": "Multivariate Time-series Forecasting with SPACE: Series Prediction Augmented by Causality Estimation", "abs": "The analysis of multivariate time series (MTS) presents a complex yet crucial task with substantial applications in areas such as weather forecasting, policy formulation, and stock market prediction. It is important to highlight three key characteristics of MTS that contribute to the challenging and multifaceted nature of their analysis: (i) their interrelationships are represented through causal relationships rather than mere similarities; (ii) they convey information across multiple independent factors; and (iii) their dynamics often arise from inherent temporal dependencies. While conventional time series analysis frameworks often fail to capture one or more of these aspects, resulting in incomplete or even misleading conclusions, we propose an end-to-end trainable $\\textbf{S}$eries $\\textbf{P}$rediction model $\\textbf{A}$ugmented by $\\textbf{C}$ausality $\\textbf{E}$stimation (SPACE) to address these limitations. This model effectively incorporates temporal dependencies and causal relationships, featuring a temporal embedding and a transfer entropy-based Cross-TE module designed to enhance predictions through causality-augmented mechanisms. Experiments demonstrate that SPACE achieves state-of-the-art results on challenging real-world time series prediction tasks, showing its effectiveness and versatility.", "keywords": [ "Time Series Forecasting", "Causal Learning", "Transfer Entropy", "Graph Based Learning" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "v593OaNePQ", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Dixant Mittal", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Liwei Kang", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Wee Sun Lee", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 45, "n_formula_1": 21, "n_ref_uni": 35, "n_ref": 62, "n_ref_all": 102, "n_fig": 7, "n_tab": 8, "L_tab": 2339, "n_element_tab": 205, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 1751, "n_element_tab_1": 167, "formula_len_all": 2801, "formula_len_all_1": 1127, "len_all": 192023, "len_all_1": 70543, "len_abs": 1472, "len_title": 96, "len_sents": 59243, "len_sents_1": 31115, "n_sents": 460, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 47, "L_abs": 1479, "title": "Learning to Search from Demonstration Sequences", "abs": "Search and planning are essential for solving many real-world problems. However, in numerous learning scenarios, only action-observation sequences, such as demonstrations or instruction sequences, are available for learning. Relying solely on supervised learning with these sequences can lead to sub-optimal performance due to the vast, unseen search space encountered during training. In this paper, we introduce Differentiable Tree Search Network (D-TSN), a novel neural network architecture that learns to construct search trees from just sequences of demonstrations by performing gradient descent on a best-first search tree construction algorithm. D-TSN enables the joint learning of submodules, including an encoder, value function, and world model, which are essential for planning. To construct the search tree, we employ a stochastic tree expansion policy and formulate it as another decision-making task. Then, we optimize the tree expansion policy via REINFORCE with an effective variance reduction technique for the gradient computation. D-TSN can be applied to problems with a known world model or to scenarios where it needs to jointly learn a world model with a latent state space. We study problems from these two scenarios, including Game of 24, 2D grid navigation, and Procgen games, to understand when D-TSN is more helpful. Through our experiments, we show that D-TSN is effective, especially when the world model with a latent state space is jointly learned.", "keywords": [ "planning", "reasoning", "learning to search", "reinforcement learning", "large language model" ], "rating_list": [ 6, 8, 8, 10 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "v4PnwdA056", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Dongrui Liu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Jie Zhang", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Jing Shao", "gender": "Female", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "chen qian", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "MS student" } ], "n_formula": 22, "n_formula_1": 5, "n_ref_uni": 64, "n_ref": 131, "n_ref_all": 159, "n_fig": 9, "n_tab": 7, "L_tab": 2353, "n_element_tab": 197, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1297, "n_element_tab_1": 152, "formula_len_all": 1979, "formula_len_all_1": 316, "len_all": 203116, "len_all_1": 65611, "len_abs": 1521, "len_title": 150, "len_sents": 44415, "len_sents_1": 30726, "n_sents": 326, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1350, "title": "DEAN: Deactivating the Coupled Neurons to Mitigate Fairness-Privacy Conflicts in Large Language Models", "abs": "Ensuring awareness of fairness and privacy in Large Language Models (LLMs) is critical. Interestingly, we discover a counter-intuitive trade-off phenomenon that enhancing an LLM's privacy awareness through Supervised Fine-Tuning (SFT) methods significantly decreases its fairness awareness with thousands of samples. To address this issue, inspired by the information theory, we introduce a training-free method to \\textbf{DEA}ctivate the fairness and privacy coupled \\textbf{N}eurons (\\textbf{DEAN}), which theoretically and empirically decrease the mutual information between fairness and privacy awareness. Extensive experimental results demonstrate that DEAN eliminates the trade-off phenomenon and significantly improves LLMs' fairness and privacy awareness simultaneously, \\eg improving Qwen-2-7B-Instruct's fairness awareness by 12.2\\% and privacy awareness by 14.0\\%.\nMore crucially, DEAN remains robust and effective with limited annotated data or even when only malicious fine-tuning data is available, whereas SFT methods may fail to perform properly in such scenarios. We hope this study provides valuable insights into concurrently addressing fairness and privacy concerns in LLMs and can be integrated into comprehensive frameworks to develop more ethical and responsible AI systems. Our code is provided in the supplementary materials.", "keywords": [ "Large Language Models", "Fairness", "Privacy" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 3, 3 ] }, { "paper_id": "v4MTnPiYXY", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Anca Dragan", "gender": "Female", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Joey Hong", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" } ], "n_formula": 22, "n_formula_1": 20, "n_ref_uni": 40, "n_ref": 81, "n_ref_all": 95, "n_fig": 4, "n_tab": 1, "L_tab": 508, "n_element_tab": 110, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 508, "n_element_tab_1": 110, "formula_len_all": 1687, "formula_len_all_1": 1241, "len_all": 201895, "len_all_1": 68985, "len_abs": 1788, "len_title": 112, "len_sents": 38586, "len_sents_1": 33252, "n_sents": 278, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1793, "title": "Q-SFT: Q-Learning for Language Models via Supervised Fine-Tuning", "abs": "Value-based reinforcement learning (RL) can in principle learn effective policies for a wide range of multi-turn problems, from games to dialogue to robotic control, including via offline RL from static previously collected datasets. However, despite the widespread use of policy gradient methods to train large language models for single turn tasks (e.g., question answering), value-based methods for multi-turn RL in an off-policy or offline setting have proven particularly challenging to scale to the setting of large language models. This setting requires effectively leveraging pretraining, scaling to large architectures with billions of parameters, and training on large datasets, all of which represent major challenges for current value-based RL methods. In this work, we propose a novel offline RL algorithm that addresses these drawbacks, casting Q-learning as a modified supervised fine-tuning (SFT) problem where the probabilities of tokens directly translate to Q-values. In this way we obtain an algorithm that smoothly transitions from maximizing the likelihood of the data during pretraining to learning a near-optimal Q-function during finetuning. Our algorithm has strong theoretical foundations, enjoying performance bounds similar to state-of-the-art Q-learning methods, while in practice utilizing an objective that closely resembles SFT. Because of this, our approach can enjoy the full benefits of the pretraining of language models, without the need to reinitialize any weights before RL finetuning, and without the need to initialize new heads for predicting values or advantages. Empirically, we evaluate our method on both pretrained LLMs and VLMs, on a variety of tasks including both natural language dialogue and robotic manipulation and navigation from images.", "keywords": [ "offline reinforcement learning", "language models", "dialogue", "robotics" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "v49jqwmGtM", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Jalal Etesami", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Assistant Professor" }, { "name": "Yutong Chao", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" } ], "n_formula": 233, "n_formula_1": 31, "n_ref_uni": 43, "n_ref": 75, "n_ref_all": 141, "n_fig": 6, "n_tab": 1, "L_tab": 153, "n_element_tab": 21, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 43714, "formula_len_all_1": 3266, "len_all": 250418, "len_all_1": 71015, "len_abs": 1472, "len_title": 129, "len_sents": 50609, "len_sents_1": 29908, "n_sents": 617, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1001, "title": "Convergence Analysis of Gradient Descent under Coordinate-wise Gradient Dominance", "abs": "We consider the optimization problem of finding Nash Equilibrium (NE) for a nonconvex function $f(x)=f(x_1,...,x_n)$, where $x_i\\in\\mathbb{R}^{d_i}$ denotes the $i$-th block of the variables. \nOur focus is on investigating first-order gradient-based algorithms and their variations such as the block coordinate descent (BCD) algorithm for tackling this problem. \nWe introduce a set of conditions, termed the $n$-sided PL condition, which extends the well-established gradient dominance condition a.k.a Polyak-{\\L}ojasiewicz (PL) condition and the concept of multi-convexity. This condition, satisfied by various classes of non-convex functions, allows us to analyze the convergence of various gradient descent (GD) algorithms. \nMoreover, our study delves into scenarios where the objective function only has strict saddle points, and normal gradient descent methods fail to converge to NE. In such cases, we propose adapted variants of GD that converge towards NE and analyze their convergence rates.", "keywords": [ "Non-convex Optimization", "Nash Equilibrium", "Gradient Dominance", "Strict Saddle" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "v46TPwU0Uy", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Bhiksha Raj", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Hao Chen", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Jason Kuen", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Researcher" }, { "name": "Jiuxiang Gu", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Kai Hu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Kai Qiu", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Xiang Li", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Zhe Lin", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Principal Researcher" } ], "n_formula": 8, "n_formula_1": 3, "n_ref_uni": 35, "n_ref": 78, "n_ref_all": 103, "n_fig": 15, "n_tab": 2, "L_tab": 3472, "n_element_tab": 36, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 429, "n_element_tab_1": 5, "formula_len_all": 1004, "formula_len_all_1": 230, "len_all": 167157, "len_all_1": 42216, "len_abs": 298, "len_title": 113, "len_sents": 28104, "len_sents_1": 19517, "n_sents": 219, "n_sents_1": 147, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1200, "title": "ControlVAR: Exploring Controllable Visual Autoregressive Modeling", "abs": "Conditional visual generation has witnessed remarkable progress with the advent of diffusion models (DMs), especially in tasks like control-to-image generation. However, challenges such as expensive computational cost, high inference latency, and difficulties of integration with large language models (LLMs) have necessitated exploring alternatives to DMs. This paper introduces ControlVAR, a novel framework that explores pixel-level controls in visual autoregressive (VAR) modeling for flexible and efficient conditional generation. In contrast to traditional conditional models that learn the conditional distribution, ControlVAR jointly models the distribution of image and pixel-level conditions during training and imposes conditional controls during testing. To enhance the joint modeling, we adopt the next-scale AR prediction paradigm and unify control and image representations. A teacher-forcing guidance strategy is proposed to further facilitate controllable generation with joint modeling. Extensive experiments demonstrate the superior efficacy and flexibility of ControlVAR across various conditional generation tasks against popular conditional DMs, \\eg, ControlNet and T2I-Adaptor.", "keywords": [ "Autoregressive generation", "Controllable image generation" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 1 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "v44CUwEeDY", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Abhishek A", "gender": "Male", "institution": "Fujitsu Research and Development Center Co. Ltm.", "country": "", "position": "Applied Researcher" }, { "name": "Mahesh Chandran", "gender": "Male", "institution": "Fujitsu Research and Development Center Co. Ltm.", "country": "IN", "position": "Principal Researcher" }, { "name": "Manohar Kaul", "gender": "Male", "institution": "Fujitsu Research and Development Center Co. Ltm.", "country": "IN", "position": "Principal Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 0, "n_ref": 0, "n_ref_all": 1, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 1995, "len_all_1": 1747, "len_abs": 0, "len_title": 0, "len_sents": 0, "len_sents_1": 0, "n_sents": 0, "n_sents_1": 0, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1421, "title": "Proper Orthogonal Decomposition for Scalable Training of Graph Neural Networks", "abs": "As large-scale graphs become ubiquitous in real-world applications, there is growing concern about \nthe memory and time requirement to train a graph neural network (GNN) model for such datasets.\nStoring the entire adjacency and node embedding matrices in memory is infeasible in such a scenario. Standard sampling-based methods for addressing the memory constraint suffer from the dependence of the number of mini-batches on the graph size. Existing sketch-based methods and graph compression techniques operate at higher sketch ratios, with the graph compression techniques showing poor generalization, implying that different GNNs trained on the same synthetic graph have performance gaps. Sketch-based methods necessitate online learning of sketches, further increasing the complexity. In this paper, we propose a new sketch-based algorithm, PGNN, employing the Proper orthogonal decomposition (POD) method to craft update rules to train GNNs, improving the memory requirement and training time without the complication of updating the sketches during training. Experiments on standard graph datasets show that PGNN can reach much lower sketch ratios without compromising the performance. We prove the optimality of the POD update rule for the linearized GNN (SGC). Empirical findings validate our approach, demonstrating superior performance at reduced sketch ratios and adaptability across various GNN architectures.", "keywords": [ "Graph Neural Networks", "Scalability", "Proper Orthogonal Decomposition", "Sublinear Complexity" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "v3XabZsB7j", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Ahsan Habib", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Lecturer" }, { "name": "Brenton Adey", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "MS student" }, { "name": "Chandan Karmakar", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 19, "n_ref": 33, "n_ref_all": 48, "n_fig": 6, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 203, "formula_len_all_1": 203, "len_all": 69804, "len_all_1": 55975, "len_abs": 1000, "len_title": 121, "len_sents": 23068, "len_sents_1": 23087, "n_sents": 133, "n_sents_1": 130, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1002, "title": "CNN Variational autoencoders' reconstruction ability of long ECG signals", "abs": "Can variational auto-encoders (VAEs) generate flexible continuous latent space for long electrocardiogram (ECG) segments and reconstruct the input? A folded VAE architecture is introduced in this study which is able to encode long ECG segments by splitting an input segment into folds and process them in sequence using a narrow field-of-view in the encoder and concatenate them at the end, instead of processing the long segment at a time. The VAE decoder follows similar folding and concatenation strategy for reconstruction of the original ECG segments. The proposed folded VAE architecture is able to generate better reconstruction of long 30-second ECG segments compared to unfolded classical VAE approach which often produce trivial reconstruction of long ECG segments. Experimental results show that the latent representation generated by our folded VAE architecture not only retains rich compressed information but also aids designing interpretable models by providing decision-making insights.", "keywords": [ "VAE", "CNN", "electrocardiogram", "reconstruction", "compression", "interpretability" ], "rating_list": [ 3, 1, 3, 1 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "v3W9tdTGx5", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Chao Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Didi Zhu", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Tao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zexi Li", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Zhiqi Li", "gender": "Male", "institution": "Epic Games", "country": "US", "position": "Intern" }, { "name": "JieLin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "tao shen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 48, "n_formula_1": 17, "n_ref_uni": 53, "n_ref": 123, "n_ref_all": 176, "n_fig": 7, "n_tab": 12, "L_tab": 4343, "n_element_tab": 292, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3693, "n_element_tab_1": 157, "formula_len_all": 5565, "formula_len_all_1": 1386, "len_all": 203687, "len_all_1": 67723, "len_abs": 1208, "len_title": 123, "len_sents": 54046, "len_sents_1": 25395, "n_sents": 506, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1214, "title": "Improving Group Connectivity for Generalization of Federated Deep Learning", "abs": "Federated learning (FL) involves multiple heterogeneous clients collaboratively training a global model via iterative local updates and model fusion. The generalization of FL's global model has a large gap compared with centralized training, which is its bottleneck for broader applications. In this paper, we study and improve FL's generalization through a fundamental \"connectivity'' perspective, which means how the local models are connected in the parameter region and fused into a generalized global model. The term \"connectivity'' is derived from linear mode connectivity (LMC), studying the interpolated loss landscape of two different solutions (e.g., modes) of neural networks. Bridging the gap between LMC and FL, in this paper, we leverage fixed anchor models to empirically and theoretically study the transitivity property of connectivity from two models (LMC) to a group of models (model fusion in FL). Based on the findings, we propose FedGuCci(+), improving group connectivity for better generalization. It is shown that our methods can boost the generalization of FL under client heterogeneity across various tasks (4 CV datasets and 6 NLP datasets) and model architectures (e.g., ViTs and PLMs).", "keywords": [ "Deep learning", "federated learning", "generalization" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 4, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "v3DwQlyGbv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Arnab Bhattacharya", "gender": "Male", "institution": "IIT Kanpur", "country": "", "position": "Full Professor" }, { "name": "Mitodru Niyogi", "gender": "Male", "institution": "CNRS", "country": "FR", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 65, "n_ref_all": 77, "n_fig": 2, "n_tab": 4, "L_tab": 3374, "n_element_tab": 176, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 791, "n_element_tab_1": 159, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 153591, "len_all_1": 58275, "len_abs": 2689, "len_title": 173, "len_sents": 33415, "len_sents_1": 28665, "n_sents": 207, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 125, "L_abs": 2695, "title": "Paramanu-Ganita: An Efficient Pre-trained Generative Mathematics Language Model with Chain-of-Thought Instruction Fine-Tuning", "abs": "In this paper, we pose the following question: whether domain specific pretraining of tiny generative language models from scratch with domain specialized tokenizer and Chain-of-Thought (CoT) instruction fine-tuning results in very competitive performance on mathematical reasoning than LLMs which are trained on trillion of tokens and humongous parameters? Secondly, we pose our second RQ: whether domain specific pretraining from scratch is environmentally sustainable, highly cost efficient? To address these research questions, we present Paramanu-Ganita, a 208 million-parameter novel Auto Regressive (AR) decoder based language model on mathematics. We performed pretraining from scratch on 31.5 billion tokens using a context size of 4096 on a mixed mathematical corpus consisting of mathematical web pages, mathematics related source code such as AlgebraStack, mathematical textbooks, Chain-of-Thought (CoT) templatised mathematical StackOverflow question answers pairs, and mathematical lecture notes in LaTeX curated by us. We also trained a math and code specialised BPE tokenizer. We proposed and performed Chain-of-Thought instruction fine-tuning of Paramanu-Ganita on the MetaMathQA dataset. We evaluate our model on GSM8K and MATH mathematical benchmarks, and on logical deductive reasoning (LogiQA) and multiple choice high school and college level math questions from SAT (AGIEVAL-SAT-Math), GRE/GMAT questions (AGIEVAL-AQuA-RAT), college and high school level math questions from MMLU.\nOur model Paramanu-Ganita, despite being 34 times smaller than the 7B LLMs, outperforms general LLMs by approximately 30% points, and even math-specialised LLMs by 3-23% points in GSM8K test accuracy metric. On MATH benchmark, Paramanu-Ganita outperformed the various models by 6-8% points. On other benchmarks such as LogiQA logical deductive reasoning benchmark, mathematical high school level multi-choice questions (MMLU-math-high-school), GRE-GMAT level quantitative questions (AGIEVAL-AQuA-RAT), SAT level math questions, Paramanu-Ganita was better than the others by about 1-4% points. The large significant margin improvement in performance of our math model over the existing LLMs signifies that reasoning capabilities of language models are just not restricted to those with humongous number of parameters. Paramanu-Ganita took only 170 hours of A100 training whereas large LLMs such as the math-specialised LLM, LLEMMA 7B, was trained for 23,000 A100 equivalent hours. Thus, our approach of pretraining powerful domain-specialised language models from scratch for domain adaptation is much more cost-effective and environmental friendly than performing continual training of LLMs.", "keywords": [ "reasoning", "language models", "pretraining", "CoT fine-tuning", "AI4Math" ], "rating_list": [ 3, 1, 3 ], "soundness_list": [ 2, 1, 3 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 1, 1, 2 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "v2zcCDYMok", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Ben Fei", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Junchao Gong", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Kun Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "LEI BAI", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Siwei Tu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Researcher" }, { "name": "Wanli Ouyang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Full Professor" }, { "name": "Weidong Yang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaokang Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University, China", "country": "CN", "position": "Professor" }, { "name": "Wenlong Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 6, "n_ref_uni": 28, "n_ref": 72, "n_ref_all": 99, "n_fig": 9, "n_tab": 5, "L_tab": 4316, "n_element_tab": 475, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2674, "n_element_tab_1": 187, "formula_len_all": 1930, "formula_len_all_1": 641, "len_all": 144860, "len_all_1": 67667, "len_abs": 1563, "len_title": 152, "len_sents": 45397, "len_sents_1": 30508, "n_sents": 308, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1566, "title": "PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling", "abs": "Precipitation nowcasting plays a pivotal role in socioeconomic sectors, especially in severe convective weather warnings. Although notable progress has been achieved by approaches mining the spatiotemporal correlations with deep learning, these methods still suffer severe blurriness as the lead time increases, which hampers accurate predictions for extreme precipitation. To alleviate blurriness, researchers explore generative methods conditioned on blurry predictions. However, the pairs of blurry predictions and corresponding ground truth need to be given in advance, making the training pipeline cumbersome and limiting the generality of generative models within blurry modes that appear in training data. By rethinking the blurriness in precipitation nowcasting as a blur kernel acting on predictions, we propose an unsupervised postprocessing method to eliminate the blurriness without the requirement of training with the pairs of blurry predictions and corresponding ground truth. Specifically, we utilize blurry predictions to guide the generation process of a pre-trained unconditional denoising diffusion probabilistic model (DDPM) to obtain high-fidelity predictions with eliminated blurriness. A zero-shot blur kernel estimation mechanism and an auto-scale denoise guidance strategy are introduced to adapt the unconditional DDPM to any blurriness modes varying from datasets and lead times in precipitation nowcasting. Extensive experiments are conducted on 7 precipitation radar datasets, demonstrating the generality and superiority of our method.", "keywords": [ "AI for Science; Precipitation Nowcasting; Diffusion Model; Zero-shot Blurriness Kernel; Auto-scale Denoise Guidance" ], "rating_list": [ 8, 6, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "v2uPdQDwSz", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Bin Gu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Hualin Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Huan Xiong", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Wanli Shi", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" } ], "n_formula": 76, "n_formula_1": 47, "n_ref_uni": 40, "n_ref": 88, "n_ref_all": 103, "n_fig": 4, "n_tab": 7, "L_tab": 3182, "n_element_tab": 406, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 1086, "n_element_tab_1": 190, "formula_len_all": 13354, "formula_len_all_1": 5905, "len_all": 153485, "len_all_1": 64455, "len_abs": 1064, "len_title": 138, "len_sents": 31366, "len_sents_1": 22722, "n_sents": 291, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1095, "title": "Query Efficient Nonsmooth Stochastic Black-Box Bilevel Optimization with Bregman Distance", "abs": "Bilevel optimization (BO) has recently gained significant attention in various machine learning applications due to its ability to model the hierarchical structures inherent in these problems. Several gradient-free methods have been proposed to address stochastic black-box bilevel optimization problems, where the gradients of both the upper and lower-level objective functions are unavailable. However, these methods suffer from high query complexity and do not accommodate more general bilevel problems involving nonsmooth regularization. In this paper, we present a query-efficient method that effectively leverages Bregman distance to solve nonsmooth stochastic black-box bilevel optimization problems. More importantly, we provide a non-asymptotic convergence analysis, showing that our method requires only $\\mathcal{O}({d_1(d_1+d_2)^2}{\\epsilon^{-2}})$ queries to reach the $\\epsilon$-stationary point. Additionally, we conduct experiments on data hyper-cleaning and hyper-representation learning tasks, demonstrating that our algorithms outperform existing bilevel optimization methods.", "keywords": [ "zeroth-order gradient", "bilevel optimization" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "v2nEL42Pvb", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Punit Rathore", "gender": "Male", "institution": "Indian Institute of Science, Indian institute of science, Bangalore", "country": "", "position": "Assistant Professor" }, { "name": "Ram Samarth B B", "gender": "Male", "institution": "Indian Institute of Science, Indian institute of science, Bangalore", "country": "IN", "position": "Intern" }, { "name": "Rishabh Sabharwal", "gender": "Male", "institution": "Indian Institute of Science, Indian institute of science, Bangalore", "country": "IN", "position": "Researcher" }, { "name": "Sundeep Prabhakar Chepuri", "gender": "Male", "institution": "Indian Institute of Science", "country": "", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 15, "n_ref_uni": 33, "n_ref": 48, "n_ref_all": 69, "n_fig": 6, "n_tab": 27, "L_tab": 14368, "n_element_tab": 1678, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 4926, "n_element_tab_1": 433, "formula_len_all": 1074, "formula_len_all_1": 651, "len_all": 180282, "len_all_1": 67545, "len_abs": 1978, "len_title": 105, "len_sents": 45648, "len_sents_1": 27296, "n_sents": 351, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1992, "title": "SSGNN: Simple Yet Effective Spectral Graph Neural Network", "abs": "Spectral GNNs leverage graph spectral properties to model graph representations but have been less explored due to their computational challenges, especially compared to the more flexible and scalable spatial GNNs, which have seen broader adoption. However, spatial methods cannot fully exploit the rich information in graph spectra. Current Spectral GNNs, relying on fixed-order polynomials, use scalar-to-scalar filters applied uniformly across eigenvalues, failing to capture key spectral shifts and signal propagation dynamics. Though set-to-set filters can capture spectral complexity, methods that employ them frequently rely on Transformers, which add considerable computational burden. Our analysis indicates that applying Transformers to these filters provides minimal advantage in the spectral domain. We demonstrate that effective spectral filtering can be achieved without the need for transformers, offering a more efficient and spectrum-aware alternative. To this end, we propose a $\\textit{Simple Yet Effective Spectral Graph Neural Network}$ (SSGNN), which leverages the graph spectrum to adaptively filter using a simplified set-to-set approach that captures key spectral features. Moreover, we introduce a novel, parameter-free $\\textit{Relative Gaussian Amplifier}$ (ReGA) module, which adaptively learns spectral filtering while maintaining robustness against structural perturbations, ensuring stability. Extensive experiments on 20 real-world graph datasets, spanning both node-level and graph-level tasks along with a synthetic graph dataset, show that SSGNN matches or surpasses the performance of state-of-the-art (SOTA) spectral-based GNNs and graph transformers while using significantly fewer parameters and GFLOPs. Specifically, SSGNN achieves performance comparable to the current SOTA Graph Transformer model, Polynormer, with an average 55x reduction in parameters and 100x reduction in GFLOPs across all datasets. Our code will be made public upon acceptance.", "keywords": [ "Spectral Graph Neural Networks", "Graph Representation Learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "v2NuTf6Kww", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 7, "author_info_list": [ { "name": "Harneet Singh", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Jeong hwan Yoon", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Karl Friston", "gender": "Male", "institution": "University College London", "country": "", "position": "Principal Researcher" }, { "name": "Katherine Hanson", "gender": "unknown", "institution": "ThoughtForge", "country": "US", "position": "Researcher" }, { "name": "Pedro Fontana", "gender": "Male", "institution": "ThoughtForge AI", "country": "US", "position": "Researcher" }, { "name": "Varun Kamat", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Whitney Sales", "gender": "Female", "institution": "ThoughtForge, Inc", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 19, "n_ref": 21, "n_ref_all": 35, "n_fig": 7, "n_tab": 3, "L_tab": 494, "n_element_tab": 47, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 514, "n_element_tab_1": 47, "formula_len_all": 61, "formula_len_all_1": 61, "len_all": 65035, "len_all_1": 46393, "len_abs": 1121, "len_title": 155, "len_sents": 24069, "len_sents_1": 22624, "n_sents": 144, "n_sents_1": 135, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1126, "title": "Network-based Active Inference for Adaptive and Cost-efficient Real-World Applications: PV Panel Inspection", "abs": "This paper introduces Network-based Active Inference (NetAIF), a novel framework that integrates random attractor dynamics and the Free Energy Principle (FEP) to improve trajectory generation and control in robotics. NetAIF optimizes the intrinsic dynamics of neural networks, enabling robots to quickly adapt to dynamic and complex real-world environments with minimal computational resources and without the need for extensive pre-training. Unlike traditional learning methods that rely on large datasets and prolonged training periods, NetAIF offers a more efficient alternative. \n\nIn real-world scenarios, such as Photovoltaic (PV) panel inspections, NetAIF demonstrates its ability to execute dynamic tasks with both high efficiency and robustness. The system excels in unpredictable environments while maintaining a low computational footprint. These capabilities make NetAIF a promising solution for industrial applications, offering cost-effective, adaptive robotic systems that can reduce operational expenses and enhance performance, particularly in sectors like energy, where adaptability and precision are crucial.", "keywords": [ "Active Inference (AIF)", "Free Energy Principle (FEP)", "Robotics", "Trajectory generation", "Random dynamical systems", "Random attractor dynamics", "Non-Equilibrium Steady State (NESS)", "Adaptive control", "Industrial automation", "Computational efficiency", "Cost-efficient solutions" ], "rating_list": [ 5, 5, 3, 3, 1 ], "soundness_list": [ 2, 2, 2, 2, 1 ], "presentation_list": [ 2, 2, 1, 1, 2 ], "contribution_list": [ 2, 2, 1, 1, 1 ], "confidence_list": [ 3, 5, 2, 3, 4 ] }, { "paper_id": "v2D1ASk5MT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Kaixiang Lin", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Li Erran Li", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Head of Science, HIL" }, { "name": "Min Bai", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Qianlan Yang", "gender": "Male", "institution": "Department of Computer Science, University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Xiong Zhou", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Yifei Zhou", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Yu-Xiong Wang", "gender": "unknown", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 101, "n_ref_all": 147, "n_fig": 27, "n_tab": 12, "L_tab": 3391, "n_element_tab": 234, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2217, "n_element_tab_1": 365, "formula_len_all": 180, "formula_len_all_1": 0, "len_all": 226466, "len_all_1": 81456, "len_abs": 2144, "len_title": 143, "len_sents": 77750, "len_sents_1": 39328, "n_sents": 545, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 2153, "title": "Proposer-Agent-Evaluator (PAE): Autonomous Skill Discovery For Foundation Model Internet Agents", "abs": "The vision of a broadly capable and goal-directed agent, such as an Internetbrowsing agent in the digital world and a household humanoid in the physical world, has rapidly advanced, thanks to the generalization capability of foundation models. Such a generalist agent needs to have a large and diverse skill repertoire, such as finding directions between two travel locations and buying specific items from the Internet. If each skill needs to be specified manually through a fixed set of human-annotated instructions, the agent\u2019s skill repertoire will necessarily be limited due to the quantity and diversity of human-annotated instructions. In this work, we address this challenge by proposing Proposer-Agent-Evaluator(PAE), a complete working system that enables foundation model agents to autonomously discover and practice skills in the wild. At the heart of PAE is a context-aware task proposer that autonomously proposes tasks for the agent to practice with context information of the websites such as user demos or even just the name of the website itself. Then, the agent policy attempts those tasks with thoughts and actual web operations in the real world with resulting trajectories evaluated by an autonomous model-based success evaluator. The success evaluation serves as the reward signal for the agent to refine its policies through RL. We validate PAE on challenging vision-based web navigation, using both real-world and self-hosted websites from WebVoyager (He et al., 2024) and WebArena (Zhou et al., 2024a). Our results show that PAE significantly improves the zero-shot generalization capability of VLM Internet agents (more than 30% relative improvement) to both unseen tasks and websites. Our model also achieves an absolute advantage of over 10% (from 22.6% to 33.0%) comparing to other state-of-the-art open source VLM agents including Qwen2VL-72B. To the best of our knowledge, this work represents the first working system to apply autonomous task proposal with RL for agents that generalizes real-world human-annotated benchmarks with sota performances. We plan to release our models and code to facilitate further research.", "keywords": [ "VLM Agent", "Web/GUI Agent", "VLM", "Reinforcement Learning", "Skill Discovery" ], "rating_list": [ 8, 3, 8, 5, 5 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 4, 3, 3, 2, 2 ], "contribution_list": [ 4, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "v27yHgKtMv", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Daehwan Kim", "gender": "unknown", "institution": "Hanyang University", "country": "KR", "position": "MS student" }, { "name": "Haejun Chung", "gender": "unknown", "institution": "Hanyang University", "country": "KR", "position": "Assistant Professor" }, { "name": "Ikbeom Jang", "gender": "Male", "institution": "Hankuk University of Foreign Studies", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 30, "n_ref": 84, "n_ref_all": 102, "n_fig": 5, "n_tab": 5, "L_tab": 1461, "n_element_tab": 90, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1967, "n_element_tab_1": 96, "formula_len_all": 533, "formula_len_all_1": 533, "len_all": 100981, "len_all_1": 70454, "len_abs": 1036, "len_title": 90, "len_sents": 34286, "len_sents_1": 33772, "n_sents": 182, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 1042, "title": "Calibration of ordinal regression networks", "abs": "Recent studies have shown that deep neural networks are not well-calibrated and produce over-confident predictions.\nThe miscalibration issue primarily stems from the minimization of cross-entropy, which aims to align predicted softmax probabilities with one-hot labels. In ordinal regression tasks, this problem is compounded by an additional challenge: the expectation that softmax probabilities should exhibit unimodal distribution is not met with cross-entropy. Rather, the ordinal regression literature has focused on unimodality and overlooked calibration. To address these issues, we propose a novel loss function that introduces order-aware calibration, ensuring that prediction confidence adheres to ordinal relationships between classes. It incorporates soft ordinal encoding and label-smoothing-based regularization to enforce both calibration and unimodality. Extensive experiments across three popular ordinal regression benchmarks demonstrate that our approach achieves state-of-the-art calibration without compromising accuracy.", "keywords": [ "Ordinal regression", "Calibration", "Deep neural networks", "Unimodality", "Loss function", "Soft ordinal encoding", "Label smoothing", "Order-aware calibration" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "v1rFkElnIn", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Alexandre Graell i Amat", "gender": "Male", "institution": "Chalmers University of Technology", "country": "", "position": "Full Professor" }, { "name": "Javad Aliakbari", "gender": "Male", "institution": "Chalmers University of Technology", "country": "SE", "position": "PhD student" }, { "name": "Johan \u00d6stman", "gender": "Male", "institution": "AI Sweden", "country": "", "position": "Researcher" } ], "n_formula": 102, "n_formula_1": 47, "n_ref_uni": 34, "n_ref": 96, "n_ref_all": 130, "n_fig": 9, "n_tab": 10, "L_tab": 10009, "n_element_tab": 504, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 452, "n_element_tab_1": 46, "formula_len_all": 24569, "formula_len_all_1": 15499, "len_all": 259032, "len_all_1": 111489, "len_abs": 953, "len_title": 85, "len_sents": 77669, "len_sents_1": 45629, "n_sents": 651, "n_sents_1": 315, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 37, "L_abs": 957, "title": "Decoupled Subgraph Federated Learning", "abs": "We address the challenge of federated learning on graph-structured data distributed across multiple clients. Specifically, we focus on the prevalent scenario of interconnected subgraphs, where inter-connections between different clients play a critical role. We present a novel framework for this scenario, named FedStruct, that harnesses deep structural dependencies. To uphold privacy, unlike existing methods, FedStruct eliminates the necessity of sharing or generating sensitive node features or embeddings among clients. Instead, it leverages explicit global graph structure information to capture inter-node dependencies. We validate the effectiveness of FedStruct through experimental results conducted on six datasets for semi-supervised node classification, showcasing performance close to the centralized approach across various scenarios, including different data partitioning methods, varying levels of label availability, and number of clients.", "keywords": [ "Federated Learning", "Subgraph Federated Learning", "Inter-Connected Graphs", "GNN", "Decoupled GCN" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 4, 2, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "v1qNr99R5n", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Canaan Yung", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" }, { "name": "Christopher Leckie", "gender": "unknown", "institution": "The University of Melbourne", "country": "AU", "position": "Full Professor" }, { "name": "Hanxun Huang", "gender": "Male", "institution": "The University of Melbourne", "country": "AU", "position": "Postdoc" }, { "name": "Sarah Monazam Erfani", "gender": "unknown", "institution": "The University of Melbourne", "country": "AU", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 10, "n_ref_uni": 75, "n_ref": 137, "n_ref_all": 164, "n_fig": 5, "n_tab": 17, "L_tab": 4036, "n_element_tab": 376, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 729, "n_element_tab_1": 42, "formula_len_all": 767, "formula_len_all_1": 388, "len_all": 225086, "len_all_1": 68234, "len_abs": 1591, "len_title": 110, "len_sents": 59548, "len_sents_1": 32710, "n_sents": 468, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1666, "title": "CURVALID: A Geometrically-guided Adversarial Prompt Detection", "abs": "Adversarial prompts that can jailbreak large language models (LLMs) and lead to undesirable behaviours pose a significant challenge to the safe deployment of LLMs. Existing defenses, such as input perturbation and adversarial training, depend on activating LLMs' defense mechanisms or fine-tuning LLMs individually, resulting in inconsistent performance across different prompts and LLMs. To address this, we propose CurvaLID, an algorithm that classifies benign and adversarial prompts by leveraging two complementary geometric measures: Local Intrinsic Dimensionality (LID) and curvature. LID provides an analysis of geometric differences at the prompt level, while curvature captures the degree of curvature in the manifolds and the semantic shifts at the word level. Together, these tools capture both prompt-level and word-level geometric properties, enhancing adversarial prompt detection. We demonstrate the limitations of using token-level LID, as applied in previous work, for capturing the geometric properties of text prompts. To address this, we propose PromptLID to calculate LID in prompt-level representations to explore the adversarial local subspace for detection. Additionally, we propose TextCurv to further analyze the local geometric structure of prompt manifolds by calculating the curvature in text prompts. CurvaLID achieves over 0.99 detection accuracy, effectively reducing the attack success rate of advanced adversarial prompts to zero or nearly zero. Importantly, CurvaLID provides a unified detection framework across different adversarial prompts and LLMs, as it achieves consistent performance regardless of the specific LLM targeted.", "keywords": [ "Large language models", "Adversarial attacks", "Local Intrinsic Dimension", "Curvature" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 2 ] }, { "paper_id": "v1f6c7wVBm", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Jingnan Gao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xiaokang Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University, China", "country": "CN", "position": "Professor" }, { "name": "Yichao Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhuo Chen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 21, "n_ref_uni": 35, "n_ref": 59, "n_ref_all": 90, "n_fig": 21, "n_tab": 3, "L_tab": 1569, "n_element_tab": 90, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2982, "n_element_tab_1": 118, "formula_len_all": 997, "formula_len_all_1": 959, "len_all": 148965, "len_all_1": 58383, "len_abs": 306, "len_title": 151, "len_sents": 30195, "len_sents_1": 24665, "n_sents": 257, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 103, "L_abs": 1330, "title": "AniSDF: Fused-Granularity Neural Surfaces with Anisotropic Encoding for High-Fidelity 3D Reconstruction", "abs": "Neural radiance fields have recently revolutionized novel-view synthesis and achieved high-fidelity renderings. \nHowever, these methods sacrifice the geometry for the rendering quality, limiting their further applications including relighting and deformation. \nHow to synthesize photo-realistic rendering while reconstructing accurate geometry remains an unsolved problem. In this work, we present AniSDF, a novel approach that learns fused-granularity neural surfaces with physics-based encoding for high-fidelity 3D reconstruction. Different from previous neural surfaces, our fused-granularity geometry structure balances the overall structures and fine geometric details, producing accurate geometry reconstruction. \nTo disambiguate geometry from reflective appearance, we introduce blended radiance fields to model diffuse and specularity following the anisotropic spherical Gaussian encoding, a physics-based rendering pipeline. With these designs, AniSDF can reconstruct objects with complex structures and produce high-quality renderings. \nFurthermore, our method is a unified model that does not require complex hyperparameter tuning for specific objects. \nExtensive experiments demonstrate that our method boosts the quality of SDF-based methods by a great scale in both geometry reconstruction and novel-view synthesis.", "keywords": [ "Surface Reconstruction", "Neural Radiance Field" ], "rating_list": [ 6, 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "v1OQ0kNq0w", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Houqiang Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Wengang Zhou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoyang Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Yunyao Mao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 12, "n_ref_uni": 28, "n_ref": 64, "n_ref_all": 73, "n_fig": 4, "n_tab": 3, "L_tab": 1126, "n_element_tab": 88, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 1126, "n_element_tab_1": 88, "formula_len_all": 819, "formula_len_all_1": 763, "len_all": 114737, "len_all_1": 53323, "len_abs": 984, "len_title": 155, "len_sents": 34874, "len_sents_1": 23387, "n_sents": 265, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 995, "title": "MotionRL: Align Text-to-Motion Generation to Human Preferences with Multi-Reward Reinforcement Learning", "abs": "We introduce **MotionRL**, the first approach to utilize Multi-Reward Reinforcement Learning (RL) for optimizing text-to-motion generation tasks and aligning them with human preferences. Previous works focused on improving numerical performance metrics on the given datasets, often neglecting the variability and subjectivity of human feedback. In contrast, our novel approach uses reinforcement learning to fine-tune the motion generator based on human preferences prior knowledge of the human perception model, allowing it to generate motions that better align human preferences. In addition, MotionRL introduces a novel multi-objective optimization strategy to approximate Pareto optimality between text adherence, motion quality, and human preferences. Extensive experiments and user studies demonstrate that MotionRL not only allows control over the generated results across different objectives but also significantly enhances performance across these metrics compared to other algorithms.", "keywords": [ "Motion Generation; Reinforcement Learning;" ], "rating_list": [ 5, 5, 6, 6, 3 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 1, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "v1B4aet9ct", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Can Pouliquen", "gender": "unknown", "institution": "Ecole Normale Sup\u00e9rieure de Lyon", "country": "", "position": "PhD student" }, { "name": "Mathurin Massias", "gender": "unknown", "institution": "INRIA", "country": "FR", "position": "Researcher" }, { "name": "Titouan Vayer", "gender": "Male", "institution": "INRIA", "country": "", "position": "Assistant Professor" } ], "n_formula": 30, "n_formula_1": 22, "n_ref_uni": 40, "n_ref": 65, "n_ref_all": 86, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1877, "formula_len_all_1": 1492, "len_all": 112164, "len_all_1": 63351, "len_abs": 1052, "len_title": 127, "len_sents": 35992, "len_sents_1": 27979, "n_sents": 288, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1057, "title": "Schur's Positive-Definite Network: Deep Learning in the SPD cone with structure", "abs": "Estimating matrices in the symmetric positive-definite (SPD) cone is of interest for many applications ranging from computer vision to graph learning. While there exist various convex optimization-based estimators, they remain limited in expressivity due to their model-based approach. The success of deep learning motivates the use of learning-based approaches to estimate SPD matrices with neural networks in a data-driven fashion. However, designing effective neural architectures for SPD learning is challenging, particularly when the task requires\nadditional structural constraints, such as element-wise sparsity. Current approaches either do not ensure that the output meets all desired properties or lack expressivity. In this paper, we introduce SpodNet, a novel and generic learning module that guarantees SPD outputs and supports additional structural constraints. Notably, it solves the challenging task of learning jointly SPD and\nsparse matrices. Our experiments illustrate the versatility and relevance of SpodNet layers for such applications.", "keywords": [ "sparsity", "graphical lasso", "lasso", "deep learning", "neural networks" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "v0O9FrVTt1", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Hongyi Chen", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Jingtao Ding", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Xiao-Ping Zhang", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaojun Liang", "gender": "Male", "institution": "Peng Cheng Lab", "country": "CN", "position": "Researcher" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 13, "n_ref_uni": 34, "n_ref": 88, "n_ref_all": 115, "n_fig": 8, "n_tab": 5, "L_tab": 1904, "n_element_tab": 205, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1380, "n_element_tab_1": 83, "formula_len_all": 869, "formula_len_all_1": 842, "len_all": 150393, "len_all_1": 65972, "len_abs": 1341, "len_title": 128, "len_sents": 54107, "len_sents_1": 31134, "n_sents": 377, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1352, "title": "Adaptive Source Localization on Complex Networks via Conditional Diffusion Model", "abs": "Network propagation issues like the spread of misinformation, cyber threats, or infrastructure breakdowns are prevalent and have significant societal impacts. Identifying the source of such propagation by analyzing snapshots of affected networks is crucial for managing crises like disease outbreaks and enhancing network security. Traditional methods rely on metrics derived from network topology and are limited to specific propagation models, while deep learning models face the challenge of data scarcity. We propose \\textbf{ASLDiff}~(\\textbf{A}daptive \\textbf{S}ource \\textbf{L}ocalization \\textbf{Diff}sion Model), a novel adaptive source localization diffusion model to achieve accurate and robust source localization across different network topologies and propagation modes by fusing the principles of information propagation and restructuring the label propagation process within the conditioning module. Our approach not only adapts to real-world patterns easily without abundant fine-tuning data but can also generalize to different network topologies easily. Evaluations of various datasets demonstrate ASLDiff's superior effectiveness, accuracy, and adaptability in real-world applications, showcasing its robust performance across different localization scenarios. The code can be found at https://anonymous.4open.science/r/ASLDiff-4FE0.", "keywords": [ "Diffusion Model", "Knowledge Informed Machine Learning", "Source Localization", "Complex Network" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "v0FzmPCd1e", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Matan Kalman", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Yaniv Leviathan", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Yossi Matias", "gender": "Male", "institution": "Google", "country": "", "position": "Vice President" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 51, "n_ref": 90, "n_ref_all": 130, "n_fig": 10, "n_tab": 7, "L_tab": 1881, "n_element_tab": 179, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 809, "n_element_tab_1": 35, "formula_len_all": 294, "formula_len_all_1": 109, "len_all": 156379, "len_all_1": 58521, "len_abs": 1013, "len_title": 88, "len_sents": 44748, "len_sents_1": 26372, "n_sents": 341, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1008, "title": "Selective Attention Improves Transformer", "abs": "Unneeded elements in the attention's context degrade performance. We introduce Selective Attention, a simple parameter-free change to the standard attention mechanism which reduces attention to unneeded elements. Selective attention consistently improves language modeling performance across model sizes and context lengths. For example, a range of transformers trained with the language modeling objective on C4 with selective attention perform equivalently to transformers with standard attention modules with ~2X more parameters and heads. In addition, selective attention allows reducing the size of the attention's context buffer, leading to substantial reductions in the memory and compute requirements during inference. For example, transformers with 100M parameters and context sizes of 512, 1,024, and 2,048 need 16X, 25X, and 47X less memory for their attention module, respectively, when equipped with selective attention, as those without selective attention, with the same validation perplexity.", "keywords": [ "selective attention", "attention", "transformer", "llm", "language model" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "uzz3qAYy0D", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Han Qiu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Zhang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Jiwei Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Qing Guo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Runyi Hu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Research Associate" }, { "name": "Tianwei Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Yiming Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Research Fellow" } ], "n_formula": 21, "n_formula_1": 18, "n_ref_uni": 37, "n_ref": 63, "n_ref_all": 105, "n_fig": 15, "n_tab": 16, "L_tab": 4717, "n_element_tab": 724, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 2059, "n_element_tab_1": 368, "formula_len_all": 1248, "formula_len_all_1": 854, "len_all": 186371, "len_all_1": 70451, "len_abs": 1660, "len_title": 128, "len_sents": 59718, "len_sents_1": 28989, "n_sents": 443, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1669, "title": "VideoShield: Regulating Diffusion-based Video Generation Models via Watermarking", "abs": "Artificial Intelligence Generated Content (AIGC) has advanced significantly, particularly with the development of video generation models such as text-to-video (T2V) models and image-to-video (I2V) models. However, like other AIGC types, video generation requires robust content control. A common approach is to embed watermarks, but most research has focused on images, with limited attention given to videos. Traditional methods, which embed watermarks frame-by-frame in a post-processing manner, often degrade video quality. In this paper, we propose VideoShield, a novel watermarking framework specifically designed for popular diffusion-based video generation models. Unlike post-processing methods, VideoShield embeds watermarks directly during video generation, eliminating the need for additional training. To ensure video integrity, we introduce a tamper localization feature that can detect changes both temporally (across frames) and spatially (within individual frames). Our method maps watermark bits to template bits, which are then used to generate watermarked noise during the denoising process. Using DDIM Inversion, we can reverse the video to its original watermarked noise, enabling straightforward watermark extraction. Additionally, template bits allow precise detection for potential spatial and temporal modification. Extensive experiments across various video models (both T2V and I2V models) demonstrate that our method effectively extracts watermarks and detects tamper without compromising video quality. Furthermore, we show that this approach is applicable to image generation models, enabling tamper detection in generated images as well.", "keywords": [ "video", "watermarking", "tamper localization" ], "rating_list": [ 8, 5, 8, 3, 6 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 4, 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 5, 4, 4 ] }, { "paper_id": "uzKG83YJ3t", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Jiaqi Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Jiazi Bu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Pan Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Pengyang Ling", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Tong Wu", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Xiaoyi Dong", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yuhang CAO", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yuhang Zang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 22, "n_formula_1": 9, "n_ref_uni": 26, "n_ref": 49, "n_ref_all": 81, "n_fig": 22, "n_tab": 1, "L_tab": 230, "n_element_tab": 22, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 230, "n_element_tab_1": 22, "formula_len_all": 1097, "formula_len_all_1": 459, "len_all": 115065, "len_all_1": 51341, "len_abs": 242, "len_title": 123, "len_sents": 32852, "len_sents_1": 23670, "n_sents": 252, "n_sents_1": 158, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1400, "title": "BroadWay: Boost Your Text-to-Video Generation Model in a Training-free Way", "abs": "The text-to-video (T2V) generation models, offering convenient visual creation, have recently garnered increasing attention. Despite their substantial potential, the generated videos may present artifacts, including structural implausibility, temporal inconsistency, and a lack of motion, often resulting in near-static video. In this work, we have identified a correlation between the disparity of temporal attention maps across different blocks and the occurrence of temporal inconsistencies. Additionally, we have observed that the energy contained within the temporal attention maps is directly related to the magnitude of motion amplitude in the generated videos. Based on these observations, we present BroadWay, a training-free method to improve the quality of text-to-video generation without introducing additional parameters, augmenting memory or sampling time. Specifically, BroadWay is composed of two principal components: 1) Temporal Self-Guidance improves the structural plausibility and temporal consistency of generated videos by reducing the disparity between the temporal attention maps across various decoder blocks. 2) Fourier-based Motion Enhancement enhances the magnitude and richness of motion by amplifying the energy of the map. Extensive experiments demonstrate that BroadWay significantly improves the quality of text-to-video generation with negligible additional cost.", "keywords": [ "generative models", "text-to-video generation", "video quality enhancement" ], "rating_list": [ 5, 3, 6, 3, 6 ], "soundness_list": [ 3, 2, 3, 1, 3 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 5, 5, 4, 5, 5 ] }, { "paper_id": "uz4QiNHB16", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Debasis Mitra", "gender": "Male", "institution": "Florida Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Deva Atturu", "gender": "Male", "institution": "Florida Institute of Technology", "country": "", "position": "MS student" }, { "name": "Parth Ganeriwala", "gender": "Male", "institution": "Florida Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 31, "n_ref": 40, "n_ref_all": 52, "n_fig": 6, "n_tab": 2, "L_tab": 47, "n_element_tab": 4, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 25, "n_element_tab_1": 3, "formula_len_all": 197, "formula_len_all_1": 196, "len_all": 112034, "len_all_1": 59465, "len_abs": 1521, "len_title": 140, "len_sents": 32566, "len_sents_1": 31003, "n_sents": 210, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1527, "title": "FLAIR: A Foundation Model for Grapheme Recognition in Ancient Scripts with Few-Shot Learning", "abs": "The Indus Valley Civilization (IVC) left behind an undeciphered script, posing a significant challenge to archaeologists and linguists. This paper introduces FLAIR, a few-shot learning approach that aims to establish a foundational model for recognizing and identifying individual graphemes from the limited available Indus script. As a foundational model, FLAIR is designed to be versatile, supporting multiple potential applications in script recognition and beyond. It leverages prototypical networks combined with a modified proposed encoder network for segmentation, ProtoSegment to extract intricate features from the grapheme images. We evaluate FLAIR\u2019s ability to generalize from minimal data using IVC grapheme classification tasks and further experiment with pre-trained Omniglot models for fine-tuning. Additionally, we simulate real-world data scarcity by intentionally restricting training data on the Omniglot dataset. Our experiments demonstrate FLAIR\u2019s accuracy in digitizing and recognizing Indus Valley seal graphemes, outperforming traditional machine learning classification approaches. These results underscore FLAIR's potential not only for the digitization of ancient scripts with limited labeled datasets but also for broader applications where data is scarce. FLAIR\u2019s success in grapheme recognition highlights its promise as a foundational model capable of extending to other undeciphered writing systems, thereby contributing to the integration of classic scientific tools and data-driven approaches.", "keywords": [ "Foundation Model", "Few-Shot Learning", "Prototypical Networks", "Encoder Network", "Indus Valley Civilization Script", "Omniglot Dataset" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 1, 2, 1 ], "presentation_list": [ 2, 1, 2, 1 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 4, 2, 5, 5 ] }, { "paper_id": "uyzkKPvVyS", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Donghoon Ahn", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Undergrad student" }, { "name": "Jin-Hwa Kim", "gender": "Unspecified", "institution": "Seoul National University", "country": "KR", "position": "Adjunct Professor" }, { "name": "Min-Seop Kwak", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Seungryong Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Hyeonsu Kim", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 28, "n_ref": 91, "n_ref_all": 113, "n_fig": 15, "n_tab": 1, "L_tab": 116, "n_element_tab": 7, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 1743, "n_element_tab_1": 60, "formula_len_all": 1183, "formula_len_all_1": 1115, "len_all": 117065, "len_all_1": 60462, "len_abs": 1312, "len_title": 121, "len_sents": 39569, "len_sents_1": 26864, "n_sents": 247, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1154, "title": "Geometry-aware Score Distillation via 3D Consistent Noising and Gradients", "abs": "Score distillation sampling (SDS), the methodology in which the score from pretrained 2D diffusion models is distilled into 3D representation, has recently brought significant advancements in text-to-3D generation. However, this approach is still confronted with critical geometric inconsistency problems such as the ``Janus problem''. We provide a novel insight into this problem, hypothesizing that the incorporation of 3D awareness into the 3D noising process and gradient distillation process may bring about enhanced consistency between gradients, leading to improved fidelity and geometric consistency. To achieve this, we propose a simple yet effective approach to achieve a 3D consistent, geometry-aware noising process, leveraging the advantages that 3D Gaussian Splatting possesses as an explicit 3D representation. Combined with our geometry-based gradient warping and our novel gradient dissimilarity loss, we demonstrate that our method significantly improves performance by addressing geometric inconsistency problems in text-to-3D generation with minimal computation cost and being compatible with existing score distillation-based models.", "keywords": [ "Diffusion Models", "Score Distillation Sampling", "Text-to-3D Generation" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "uy9oR0nYCW", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Christian Schroeder de Witt", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Lecturer" }, { "name": "Georgia Channing", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Juil Sock", "gender": "Male", "institution": "BBC", "country": "GB", "position": "Researcher" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Ronald Clark", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 15, "n_ref_uni": 32, "n_ref": 71, "n_ref_all": 83, "n_fig": 4, "n_tab": 3, "L_tab": 907, "n_element_tab": 105, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 439, "n_element_tab_1": 5, "formula_len_all": 369, "formula_len_all_1": 423, "len_all": 135733, "len_all_1": 57175, "len_abs": 724, "len_title": 129, "len_sents": 47054, "len_sents_1": 28666, "n_sents": 314, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 726, "title": "Toward Robust Real-World Audio Deepfake Detection: Closing the Explainability Gap", "abs": "The rapid proliferation of AI-manipulated or generated audio deepfakes poses serious challenges to media integrity and election security. Current AI-driven detection solutions lack explainability and underperform in real-world settings. In this paper, we introduce novel explainability methods for state-of-the-art transformer-based audio deepfake detectors and open-source a novel benchmark for real-world generalizability. By narrowing the explainability gap between transformer-based audio deepfake detectors and traditional methods, our results not only build trust with human experts, but also pave the way for unlocking the potential of citizen intelligence to overcome the scalability issue in audio deepfake detection.", "keywords": [ "self-supervised learning", "explainability", "deepfake audio", "generalizability" ], "rating_list": [ 1, 3, 5, 1 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "uy4EavBEwl", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Ally Yalei Du", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Undergrad student" }, { "name": "Dung Daniel Ngo", "gender": "Male", "institution": "University of Minnesota, Minneapolis", "country": "US", "position": "PhD student" }, { "name": "Zhiwei Steven Wu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 157, "n_formula_1": 25, "n_ref_uni": 24, "n_ref": 42, "n_ref_all": 58, "n_fig": 6, "n_tab": 1, "L_tab": 31, "n_element_tab": 6, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 17176, "formula_len_all_1": 2383, "len_all": 200923, "len_all_1": 70914, "len_abs": 4845, "len_title": 110, "len_sents": 62891, "len_sents_1": 33363, "n_sents": 553, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1683, "title": "Reconciling Model Multiplicity for Downstream Decision Making", "abs": "We consider the problem of \\emph{model multiplicity} in downstream decision-making, a setting where two predictive models of equivalent accuracy cannot agree on what action to take for a downstream decision-making problem. Prior work attempts to address model multiplicity by resolving prediction disagreement between models. However, we show that even when the two predictive models approximately agree on their individual predictions almost everywhere, these models can lead the downstream decision-maker to take actions with substantially higher losses. We address this issue by proposing a framework that \\emph{calibrates} the predictive models with respect to both a finite set of downstream decision-making problems and the individual probability prediction. Specifically, leveraging tools from multi-calibration, we provide an algorithm that, at each time-step, first reconciles the differences in individual probability prediction, then calibrates the updated models such that they are indistinguishable from the true probability distribution to the decision-makers. We extend our results to the setting where one does not have direct access to the true probability distribution and instead relies on a set of i.i.d data to be the empirical distribution. Furthermore, we generalize our results to the settings where one has more than two predictive models and an infinitely large downstream action set. Finally, we provide a set of experiments to evaluate our methods empirically. Compared to existing work, our proposed algorithm creates a pair of predictive models with improved downstream decision-making losses and agrees on their best-response actions almost everywhere.", "keywords": [ "model multiplicity", "multi-calibration", "decision-making", "uncertainty quantification" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "uy31tqVuNo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "David E. Jacobs", "gender": "Male", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Jialu Li", "gender": "Female", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Michael Rubinstein", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Principal Researcher" }, { "name": "Mohit Bansal", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" }, { "name": "Nataniel Ruiz", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Neal Wadhwa", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Yael Pritch", "gender": "Female", "institution": "Google Research", "country": "US", "position": "Researcher" }, { "name": "Yuanzhen Li", "gender": "Female", "institution": "Google", "country": "US", "position": "Software Engineer" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 63, "n_ref": 91, "n_ref_all": 120, "n_fig": 14, "n_tab": 4, "L_tab": 760, "n_element_tab": 78, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 702, "n_element_tab_1": 126, "formula_len_all": 229, "formula_len_all_1": 125, "len_all": 168419, "len_all_1": 61635, "len_abs": 509, "len_title": 66, "len_sents": 41215, "len_sents_1": 30090, "n_sents": 286, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1440, "title": "Unbounded: A Generative Infinite Game of Character Life Simulation", "abs": "We introduce the concept of a generative infinite game, a video game that transcends the traditional boundaries of finite, hard-coded systems by using generative models. Inspired by James P. Carse's distinction between finite and infinite games, we leverage recent advances in generative AI to create Unbounded: a game of character life simulation that is fully encapsulated in generative models. Specifically, Unbounded draws inspiration from sandbox life simulations and allows you to interact with your autonomous virtual character in a virtual world by feeding, playing with and guiding it - with open-ended mechanics generated by an LLM, some of which can be emergent. In order to develop Unbounded, we propose technical innovations in both the LLM and visual generation domains. Specifically, we present: (1) a specialized, distilled large language model (LLM) that dynamically generates game mechanics, narratives, and character interactions in real-time, and (2) a new dynamic regional image prompt Adapter (IP-Adapter) for vision models that ensures consistent yet flexible visual generation of a character across multiple environments. We evaluate our system through both qualitative and quantitative analysis, showing significant improvements in character life simulation, user instruction following, narrative coherence, and visual consistency for both characters and the environments compared to traditional related approaches.", "keywords": [ "Text-to-Image Generation", "Interactive Image Generation" ], "rating_list": [ 6, 6 ], "soundness_list": [ 3, 2 ], "presentation_list": [ 3, 3 ], "contribution_list": [ 3, 2 ], "confidence_list": [ 4, 4 ] }, { "paper_id": "uxYbEAEWm4", "primary_area": "foundation or frontier models, including LLMs", "n_author": 1, "author_info_list": [ { "name": "Noriaki Kawamae", "gender": "unknown", "institution": "Sophia University", "country": "JP", "position": "Lecturer" } ], "n_formula": 19, "n_formula_1": 17, "n_ref_uni": 32, "n_ref": 62, "n_ref_all": 94, "n_fig": 1, "n_tab": 3, "L_tab": 808, "n_element_tab": 97, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2484, "n_element_tab_1": 115, "formula_len_all": 1349, "formula_len_all_1": 1415, "len_all": 141085, "len_all_1": 65691, "len_abs": 1447, "len_title": 84, "len_sents": 32203, "len_sents_1": 29662, "n_sents": 215, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 1494, "title": "Knowledge Lift Alignment Fine Tuning", "abs": "We present a visual tuning framework, \\textbf{K}nowledge \\textbf{L}ift \\textbf{A}lignment \\textbf{F}ine \\textbf{T}uning (KLAFT), \nwhich enhances the expressive image captioning capabilities of Pre-trained Language Models (PLMs), including LLMs and VLMs.\nAs this task involves generating more detailed and comprehensive captions than basic image descriptions,\nthe core idea behind KLAFT is that fine-grained alignment could exploit the capabilities of PLMs and a given target domain dataset.\nThis idea motivates and challenges us to explore the framework that deeply understands both given images and text for this alignment and tuning PLMs towards expressive image captioning.\nThis direction modifies the attention mechanism (Modified Attention Mechanism, MAM) and develops both a Topic Control Mechanism (TCM) and their training objectives.\nThe innovation of KLAFT lies in its approach to addressing the disparities in knowledge - visual versus textual via MAM\nand source versus target domain via TCM.\nAs these hidden spaces are conceptualized as distinct sub-networks within the PLM, each possessing specific knowledge,\nKLAFT's unique contribution is in aligning and adjusting the weights of these sub-networks in a fine-grained manner,\nand fine-tuning this PLM.\nOur empirical studies demonstrate that KLAFT significantly improves expressive captioning tasks by aligning and amplifying target knowledge, with the potential for Parameter-Efficient Fine-Tuning (PEFT) at low computational cost.", "keywords": [ "PEFT", "PLM", "LLM", "VLM", "Multi-modal", "Image captioning" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 1, 1 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "uxVBbSlKQ4", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "David L\u00fcdke", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Leo Schwinn", "gender": "Male", "institution": "Technical University of Munich", "country": "DE", "position": "Postdoc" }, { "name": "Marcel Kollovieh", "gender": "Male", "institution": "Department of Informatics, Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Marten Lienen", "gender": "unknown", "institution": "Technical University Munich", "country": "DE", "position": "PhD student" }, { "name": "Stephan G\u00fcnnemann", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Professor" } ], "n_formula": 35, "n_formula_1": 25, "n_ref_uni": 46, "n_ref": 93, "n_ref_all": 119, "n_fig": 7, "n_tab": 13, "L_tab": 4359, "n_element_tab": 416, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1587, "n_element_tab_1": 90, "formula_len_all": 29754, "formula_len_all_1": 27504, "len_all": 225258, "len_all_1": 106512, "len_abs": 1241, "len_title": 132, "len_sents": 57335, "len_sents_1": 31890, "n_sents": 585, "n_sents_1": 288, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1245, "title": "Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting", "abs": "Recent advancements in generative modeling, particularly diffusion models, have opened new directions for time series modeling, achieving state-of-the-art performance in forecasting and synthesis. However, the reliance of diffusion-based models on a simple, fixed prior complicates the generative process since the data and prior distributions differ significantly. We introduce TSFlow, a conditional flow matching (CFM) model for time series that simplifies the generative problem by combining Gaussian processes, optimal transport paths, and data-dependent prior distributions. By incorporating (conditional) Gaussian processes, TSFlow aligns the prior distribution more closely with the temporal structure of the data, enhancing both unconditional and conditional generation. Furthermore, we propose conditional prior sampling to enable probabilistic forecasting with an unconditionally trained model. In our experimental evaluation on eight real-world datasets, we demonstrate the generative capabilities of TSFlow, producing high-quality unconditional samples. Finally, we show that both conditionally and unconditionally trained models achieve competitive results in forecasting benchmarks, surpassing other methods on 6 out of 8 datasets.", "keywords": [ "flow matching", "time series forecasting", "generative modeling", "deep learning" ], "rating_list": [ 5, 6, 8, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 2, 4, 2 ] }, { "paper_id": "uxDFlPGRLX", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Alexander Richard", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Simon Welker", "gender": "Male", "institution": "Universit\u00e4t Hamburg", "country": "DE", "position": "PhD student" }, { "name": "Timo Gerkmann", "gender": "unknown", "institution": "Universit\u00e4t Hamburg", "country": "DE", "position": "Full Professor" }, { "name": "YI-CHIAO WU", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Matthew Le", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Tian Qi Chen", "gender": "Male", "institution": "FAIR Labs, Meta AI", "country": "US", "position": "Researcher" }, { "name": "Wei-Ning Hsu", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 10, "n_ref_uni": 54, "n_ref": 166, "n_ref_all": 202, "n_fig": 15, "n_tab": 10, "L_tab": 2928, "n_element_tab": 432, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 882, "n_element_tab_1": 76, "formula_len_all": 1175, "formula_len_all_1": 616, "len_all": 197402, "len_all_1": 69900, "len_abs": 1057, "len_title": 128, "len_sents": 52441, "len_sents_1": 31703, "n_sents": 334, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1059, "title": "FlowDec: A flow-based full-band general audio codec with high perceptual quality", "abs": "We propose FlowDec, a neural full-band audio codec for general audio sampled at 48 kHz that combines non-adversarial codec training with a stochastic postfilter based on a novel conditional flow matching method. Compared to the prior work ScoreDec which is based on score matching, we generalize from speech to general audio and move from 24 kbit/s to as low as 4 kbit/s, while improving output quality and reducing the required postfilter DNN evaluations from 60 to 6 without any fine-tuning or distillation techniques. We provide theoretical insights and geometric intuitions for our approach in comparison to ScoreDec as well as another recent work that uses flow matching, and conduct ablation studies on our proposed components. We show that FlowDec is a competitive alternative to the recent GAN-dominated stream of neural codecs, achieving FAD scores better than those of the established GAN-based codec DAC and listening test scores that are on par, and producing qualitatively more natural reconstructions for speech and harmonic structures in music.", "keywords": [ "audio", "audio codec", "generative models", "flow matching", "postfilter", "signal enhancement" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 4, 4, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "uwzyMFwyOO", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Alessandro Manenti", "gender": "unknown", "institution": "Universita della Svizzera Italiana", "country": "CH", "position": "PhD student" }, { "name": "Cesare Alippi", "gender": "Male", "institution": "Universit\u00e0 della Svizzera Italiana", "country": "", "position": "Full Professor" }, { "name": "Daniele Zambon", "gender": "unknown", "institution": "The Swiss AI Lab IDSIA", "country": "", "position": "Postdoc" } ], "n_formula": 34, "n_formula_1": 21, "n_ref_uni": 56, "n_ref": 83, "n_ref_all": 127, "n_fig": 14, "n_tab": 3, "L_tab": 505, "n_element_tab": 31, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 155, "n_element_tab_1": 18, "formula_len_all": 2707, "formula_len_all_1": 1780, "len_all": 165243, "len_all_1": 66179, "len_abs": 1090, "len_title": 102, "len_sents": 44045, "len_sents_1": 26687, "n_sents": 350, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1096, "title": "Learning Latent Graph Structures and their Uncertainty", "abs": "Graph neural networks use relational information as an inductive bias to enhance prediction performance. Not rarely, task-relevant relations are unknown and graph structure learning approaches have been proposed to learn them from data. Given their latent nature, no graph observations are available to provide a direct training signal to the learnable relations. Therefore, graph topologies are typically learned on the prediction task alongside the other graph neural network parameters.\nIn this paper, we demonstrate that minimizing point-prediction losses does not guarantee proper learning of the latent relational information and its associated uncertainty. Conversely, we prove that suitable loss functions on the stochastic model outputs simultaneously grant solving two tasks: (i) learning the unknown distribution of the latent graph and (ii) achieving optimal predictions of the model output. \nFinally, we propose a sampling-based method that solves this joint learning task. Empirical results validate our theoretical claims and demonstrate the effectiveness of the proposed approach.", "keywords": [ "Graph Structure Learning", "Graph Neural Networks", "Latent Distribution Calibration", "Discrete Random Variables" ], "rating_list": [ 6, 6, 8, 5, 3 ], "soundness_list": [ 4, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3, 4 ] }, { "paper_id": "uvHmnahyp1", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Arne Schneuing", "gender": "Not Specified", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Bruno Correia", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "Associate Professor" }, { "name": "Charles Harris", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Emmanuel Bengio", "gender": "Male", "institution": "Recursion", "country": "", "position": "Researcher" }, { "name": "Ilia Igashov", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Julien Roy", "gender": "Male", "institution": "Valence Labs powered by recursion", "country": "CA", "position": "Researcher" }, { "name": "Miruna Cretu", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Pietro Lio", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Marwin Segler", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 65, "n_ref": 118, "n_ref_all": 150, "n_fig": 13, "n_tab": 6, "L_tab": 2960, "n_element_tab": 241, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1360, "n_element_tab_1": 112, "formula_len_all": 438, "formula_len_all_1": 181, "len_all": 219851, "len_all_1": 64452, "len_abs": 1321, "len_title": 124, "len_sents": 61792, "len_sents_1": 32022, "n_sents": 432, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 76, "L_abs": 1250, "title": "SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints", "abs": "Generative models see increasing use in computer-aided drug design. However, while performing well at capturing distributions of molecular motifs, they often produce synthetically inaccessible molecules. To address this, we introduce SynFlowNet, a GFlowNet model whose action space uses chemical reactions and buyable reactants to sequentially build new molecules. By incorporating forward synthesis as an explicit constraint of the generative mechanism, we aim at bridging the gap between in silico molecular generation and real world synthesis capabilities. We evaluate our approach using synthetic accessibility scores and an independent retrosynthesis tool to assess the synthesizability of our compounds, and motivate the choice of GFlowNets through considerable improvement in sample diversity compared to baselines. Additionally, we identify challenges with reaction encodings that can complicate traversal of the MDP in the backward direction. To address this, we introduce various strategies for learning the GFlowNet backward policy and thus demonstrate how additional constraints can be integrated into the GFlowNet MDP framework. This approach enables our model to successfully identify synthesis pathways for previously unseen molecules.", "keywords": [ "GFlowNets", "de novo molecular generation", "synthesizable molecular design" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "uuvujfQXZy", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Julian Schur", "gender": "Male", "institution": "Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "Undergrad student" }, { "name": "Simon Schrodi", "gender": "Male", "institution": "University of Freiburg, Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Thomas Brox", "gender": "Male", "institution": "University of Freiburg", "country": "", "position": "Full Professor" }, { "name": "Max Argus", "gender": "Male", "institution": "University of Freiburg, Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "Postdoc" } ], "n_formula": 20, "n_formula_1": 8, "n_ref_uni": 54, "n_ref": 124, "n_ref_all": 202, "n_fig": 39, "n_tab": 4, "L_tab": 1360, "n_element_tab": 51, "n_fig_1": 15, "n_tab_1": 3, "L_tab_1": 1057, "n_element_tab_1": 43, "formula_len_all": 1439, "formula_len_all_1": 565, "len_all": 180234, "len_all_1": 61754, "len_abs": 1162, "len_title": 111, "len_sents": 53010, "len_sents_1": 26896, "n_sents": 447, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1168, "title": "Selective Concept Bottleneck Models Without Predefined Concepts", "abs": "Concept-based models like Concept Bottleneck Models (CBMs) have garnered significant interest for improving model interpretability by first predicting human-understandable concepts before mapping them to the output classes. Early approaches required costly concept annotations. To alleviate such, recent methods utilized large language models to automatically generate class-specific concept descriptions and learned mappings from a pretrained black-box model\u2019s raw features to these concepts using vision-language models. However, these approaches assume prior knowledge of which concepts the black-box model has learned. In this work, we discover the concepts encoded by the model through unsupervised concept discovery techniques instead. We further propose an input-dependent concept selection mechanism that dynamically retains a sparse set of relevant concepts for each input, enhancing both sparsity and interpretability. Our approach not only improves downstream performance but also needs significantly fewer concepts for accurate classification. Lastly, we show how large vision-language models can guide the editing of our models' weights to correct errors.", "keywords": [ "interpretability", "concept bottleneck models", "concepts" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "uuriavczkL", "primary_area": "causal reasoning", "n_author": 2, "author_info_list": [ { "name": "Arvind Raghavan", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Elias Bareinboim", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Associate Professor" } ], "n_formula": 70, "n_formula_1": 14, "n_ref_uni": 31, "n_ref": 91, "n_ref_all": 137, "n_fig": 29, "n_tab": 12, "L_tab": 2667, "n_element_tab": 149, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 71, "n_element_tab_1": 9, "formula_len_all": 8325, "formula_len_all_1": 1684, "len_all": 300079, "len_all_1": 77262, "len_abs": 1932, "len_title": 0, "len_sents": 108357, "len_sents_1": 35453, "n_sents": 998, "n_sents_1": 293, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 28, "L_abs": 1302, "title": "Counterfactual Realizability", "abs": "It is commonly believed that, in a real-world environment, samples can only be drawn from observational and interventional distributions, corresponding to Layers 1 and 2 of the Pearl Causal Hierarchy. Layer 3, representing counterfactual distributions, is believed to be inaccessible by definition. However, Bareinboim, Forney, and Pearl (2015) introduced a procedure that allows an agent to sample directly from a counterfactual distribution, leaving open the question of what other counterfactual quantities can be estimated directly via physical experimentation. We resolve this by introducing a formal definition of *realizability*, the ability to draw samples from a distribution, and then developing a complete algorithm to determine whether an arbitrary counterfactual distribution is realizable given fundamental physical constraints, such as the inability to go back in time and subject the same unit to a different experimental condition. We illustrate the implications of this new framework for counterfactual data collection using motivating examples from causal fairness and causal reinforcement learning. While the baseline approach in these motivating settings typically follows an interventional or observational strategy, we show that a counterfactual strategy provably dominates both.", "keywords": [ "causal inference", "experiment design", "causal reinforcement learning", "counterfactual reasoning" ], "rating_list": [ 6, 10, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "uuef1HP6X7", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Jiaqi Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Jing Tan", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Mengchen Zhang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Tong Wu", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Zhibing Li", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 41, "n_ref": 78, "n_ref_all": 107, "n_fig": 21, "n_tab": 8, "L_tab": 1305, "n_element_tab": 193, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 429, "n_element_tab_1": 89, "formula_len_all": 189, "formula_len_all_1": 55, "len_all": 134445, "len_all_1": 50653, "len_abs": 1846, "len_title": 131, "len_sents": 34927, "len_sents_1": 23211, "n_sents": 285, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1498, "title": "IDIV: Intrinsic Decomposition for Arbitrary Number of Input Views and Illuminations", "abs": "Capturing geometric and material information from images remains a fundamental challenge in computer vision and graphics. Traditional optimization-based methods often require hours of computational time to reconstruct geometry, material properties, and environmental lighting from dense multi-view inputs, while still struggling with inherent ambiguities between lighting and material. On the other hand, learning-based approaches leverage rich material priors from existing 3D object datasets but face challenges with maintaining multi-view consistency.\nIn this paper, we introduce IDVI, a diffusion-based model designed to perform intrinsic decomposition on an arbitrary number of images under varying illuminations. Our method achieves highly accurate and multi-view consistent estimation on surface normals and material properties. This is made possible through a novel cross-view, cross-domain attention module and an illumination-augmented, view-adaptive training strategy. Additionally, we introduce ARB-Objaverse, a new dataset that provides large-scale multi-view intrinsic data and renderings under diverse lighting conditions, supporting robust training.\nExtensive experiments demonstrate that IDVI outperforms state-of-the-art methods both qualitatively and quantitatively. Moreover, our approach facilitates a range of downstream tasks, including single-image relighting, photometric stereo, and 3D reconstruction, highlighting its broad applicability in realistic 3D content creation.", "keywords": [ "inverse rendering", "diffusion models", "intrinsic decomposition" ], "rating_list": [ 6, 6, 8, 8, 5 ], "soundness_list": [ 3, 3, 3, 4, 2 ], "presentation_list": [ 2, 3, 2, 4, 3 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 4, 5, 5, 4, 5 ] }, { "paper_id": "uuXPWRtwvK", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Lifu Huang", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Assistant Professor" }, { "name": "Liping Liu", "gender": "unknown", "institution": "Tufts University", "country": "US", "position": "Associate Professor" }, { "name": "Sijia Wang", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yukun Li", "gender": "unknown", "institution": "Tufts University", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 69, "n_ref_all": 78, "n_fig": 2, "n_tab": 5, "L_tab": 1076, "n_element_tab": 184, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1134, "n_element_tab_1": 195, "formula_len_all": 324, "formula_len_all_1": 389, "len_all": 149015, "len_all_1": 64006, "len_abs": 1422, "len_title": 108, "len_sents": 39301, "len_sents_1": 29725, "n_sents": 346, "n_sents_1": 231, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1239, "title": "Graph-based Confidence Calibration for Large Language Models", "abs": "One important approach to improving the reliability of large language models (LLMs) is to provide accurate confidence estimations regarding the correctness of their answers. However, developing a well-calibrated confidence estimation model is challenging, as mistakes made by LLMs can be difficult to detect. We propose a novel method combining the LLM's self-consistency with labeled data and training an auxiliary model to estimate the correctness of its responses to questions. This auxiliary model predicts the correctness of responses based solely on their consistent information. To set up the learning problem, we use a weighted graph to represent the consistency among the LLM's multiple responses to a question. Correctness labels are assigned to these responses based on their similarity to the correct answer. We then train a graph neural network to estimate the probability of correct responses. Experiments demonstrate that the proposed approach substantially outperforms several of the most recent methods in confidence calibration across multiple widely adopted benchmark datasets. Furthermore, the proposed approach significantly improves the generalization capability of confidence calibration on out-of-domain (OOD) data.", "keywords": [ "Language Models; Uncertainty Calibration" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "uuPkll6i7m", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Adel Bibi", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Senior Researcher" }, { "name": "Cornelius Emde", "gender": "Not Specified", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Francesco Pinto", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Postdoc" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Thomas Lukasiewicz", "gender": "unknown", "institution": "Institute of Logic and Computation, Technische Universit\u00e4t Wien", "country": "", "position": "Full Professor" } ], "n_formula": 48, "n_formula_1": 20, "n_ref_uni": 54, "n_ref": 105, "n_ref_all": 169, "n_fig": 17, "n_tab": 9, "L_tab": 1936, "n_element_tab": 360, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 491, "n_element_tab_1": 123, "formula_len_all": 4444, "formula_len_all_1": 1316, "len_all": 277789, "len_all_1": 79780, "len_abs": 970, "len_title": 74, "len_sents": 83078, "len_sents_1": 33692, "n_sents": 780, "n_sents_1": 297, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1039, "title": "Towards Certification of Uncertainty Calibration under Adversarial Attacks", "abs": "Since neural classifiers are known to be sensitive to adversarial perturbations that alter their accuracy, certification methods have been developed to provide provable guarantees on the insensitivity of their predictions to such perturbations. On the other hand, in safety-critical applications, the frequentist interpretation of the confidence of a classifier (also known as model calibration) can be of utmost importance. This property can be measured via the Brier score or the expected calibration error. We show that attacks can significantly harm calibration, and thus propose certified calibration providing worst-case bounds on calibration under adversarial perturbations. Specifically, we produce analytic bounds for the Brier score and approximate bounds via the solution of a mixed-integer program on the expected calibration error. Finally, we propose novel calibration attacks and demonstrate how they can improve model calibration through adversarial calibration training. The code will be publicly released upon acceptance.", "keywords": [ "Machine Learning", "Adversarial Robustness", "Certification", "Adversarial Training", "Uncertainty Quantification", "Calibration", "Deep Learning", "Certified Calibration" ], "rating_list": [ 5, 8, 8, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "uuOmdQy6p7", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Hao Lin", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "Wenqiang Wang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Xiaochun Cao", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Yan XIAO", "gender": "Female", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "jiehongli", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "china" } ], "n_formula": 26, "n_formula_1": 15, "n_ref_uni": 56, "n_ref": 98, "n_ref_all": 119, "n_fig": 3, "n_tab": 16, "L_tab": 7673, "n_element_tab": 1137, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1959, "n_element_tab_1": 313, "formula_len_all": 2272, "formula_len_all_1": 1702, "len_all": 210061, "len_all_1": 74784, "len_abs": 1386, "len_title": 114, "len_sents": 57343, "len_sents_1": 30399, "n_sents": 427, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1394, "title": "Few-shot Text Adversarial Attack for Black-box Multi-task Learning", "abs": "Current multi-task adversarial text attacks rely on white-box access to shared in-\nternal features and assume a homogeneous multi-task learning framework. As a\nresult, these attacks are less effective against practical scenarios involving black-\nbox feedback APIs and heterogeneous multi-task learning. To bridge this gap,\nwe introduce Cluster and Ensemble Mutil-task Text Adversarial Attack (CEMA),\nan effective black-box attack that exploits the transferability of adversarial texts.\nSpecifically, we initially employ cluster-oriented substitute model training, as a\nplug-and-play framework, to simplify complex multi-task scenarios into more\nmanageable text classification attacks and train the substitute model. Next, we\ngenerate multiple adversarial candidate examples by applying various adversarial\ntext classification methods. Finally, we select the adversarial example that attacks\nthe most substitute models as the final attack output. CEMA is evaluated on two\nprimary multi-task objectives: text classification and translation. In the classifica-\ntion task, CEMA achieves attack success rates that exceed 60% while reducing the\ntotal number of queries to 100. For the text translation task, the BLEU scores of\nboth victim texts and adversarial examples decrease to below 0.36 with 100 queries\neven including the commercial translation APIs, such as Baidu Translate and Ali\nTranslate.", "keywords": [ "multi-task adversarial text attacks" ], "rating_list": [ 5, 6, 5, 1 ], "soundness_list": [ 3, 4, 3, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 1 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "uuEQsqb0GH", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Benjamin Plaut", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Hanlin Zhu", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department, University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Stuart Russell", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" } ], "n_formula": 95, "n_formula_1": 14, "n_ref_uni": 42, "n_ref": 67, "n_ref_all": 80, "n_fig": 2, "n_tab": 1, "L_tab": 172, "n_element_tab": 13, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 516, "n_element_tab_1": 30, "formula_len_all": 11150, "formula_len_all_1": 751, "len_all": 230848, "len_all_1": 76345, "len_abs": 1078, "len_title": 106, "len_sents": 73017, "len_sents_1": 33632, "n_sents": 779, "n_sents_1": 312, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1084, "title": "Avoiding Catastrophe in Online Learning by Asking for Help", "abs": "Most learning algorithms with formal regret guarantees assume that no mistake is irreparable and essentially rely on trying all possible behaviors. This approach is problematic when some mistakes are _catastrophic_, i.e., irreparable. We propose an online learning problem where the goal is to minimize the chance of catastrophe. Specifically, we assume that the payoff in each round represents the chance of avoiding catastrophe that round and try to maximize the product of payoffs (the overall chance of avoiding catastrophe) while allowing a limited number of queries to a mentor. We first show that in general, any algorithm either constantly queries the mentor or is nearly guaranteed to cause catastrophe. However, in settings where the mentor policy class is learnable in the standard online model, we provide an algorithm whose regret and rate of querying the mentor both approach 0 as the time horizon grows. Conceptually, if a policy class is learnable in the absence of catastrophic risk, it is learnable in the presence of catastrophic risk if the agent can ask for help.", "keywords": [ "online learning", "AI safety", "asking for help", "irreversibility" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "uuCgIHqxpr", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "John Sipple", "gender": "unknown", "institution": "George Washington University", "country": "", "position": "Lecturer" }, { "name": "Judah Goldfeder", "gender": "unknown", "institution": "Columbia University", "country": "", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 5, "n_ref_uni": 74, "n_ref": 87, "n_ref_all": 115, "n_fig": 18, "n_tab": 6, "L_tab": 2452, "n_element_tab": 183, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 167, "n_element_tab_1": 26, "formula_len_all": 1427, "formula_len_all_1": 197, "len_all": 244750, "len_all_1": 66065, "len_abs": 1884, "len_title": 229, "len_sents": 78993, "len_sents_1": 33513, "n_sents": 545, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 179, "L_abs": 1891, "title": "Real-World Data and Calibrated Simulation Suite for Offline Training of Reinforcement Learning Agents to Optimize Energy and Emission in Buildings for Environmental Sustainability", "abs": "Commercial office buildings contribute 17 percent of Carbon Emissions in the US, according to the US Energy Information Administration (EIA), and improving their efficiency will reduce their environmental burden and operating cost. A major contributor of energy consumption in these buildings are the Heating, Ventilation, and Air Conditioning (HVAC) devices. HVAC devices form a complex and interconnected thermodynamic system with the building and outside weather conditions, and current setpoint control policies are not fully optimized for minimizing energy use and carbon emission. Given a suitable training environment, a Reinforcement Learning (RL) agent is able to improve upon these policies, but training such a model, especially in a way that scales to thousands of buildings, presents many practical challenges. Most existing work on applying RL to this important task either makes use of proprietary data, or focuses on expensive and proprietary simulations that may not be grounded in the real world. We present the Smart Buildings Control Suite, the first open source interactive HVAC control dataset extracted from live sensor measurements of devices in real office buildings. The dataset consists of two components: six years of real-world historical data from three buildings, for offline RL, and a lightweight interactive simulator for each of these buildings, calibrated using the historical data, for online and model-based RL. For ease of use, our RL environments are all compatible with the OpenAI gym environment standard. We also demonstrate a novel method of calibrating the simulator, as well as baseline results on training an RL agent on the simulator, predicting real-world data, and training an RL agent directly from data. We believe this benchmark will accelerate progress and collaboration on building optimization and environmental sustainability research.", "keywords": [ "Reinforcement Learning", "HVAC Control", "Simulator", "RL Environment", "Environmental Sustainability", "Climate", "Time-series prediction" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 1, 2, 2 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "uuCcK4cmlH", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Dawn Song", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Nathaniel D. Bastian", "gender": "Male", "institution": "United States Military Academy", "country": "US", "position": "Principal Researcher" }, { "name": "Yanjie Li", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Intern" }, { "name": "Zhen Xiang", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 29, "n_ref": 45, "n_ref_all": 67, "n_fig": 8, "n_tab": 9, "L_tab": 3991, "n_element_tab": 648, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1946, "n_element_tab_1": 72, "formula_len_all": 339, "formula_len_all_1": 276, "len_all": 140143, "len_all_1": 66054, "len_abs": 2386, "len_title": 80, "len_sents": 42374, "len_sents_1": 31545, "n_sents": 323, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1808, "title": "IDS-Agent: An LLM Agent for Explainable Intrusion Detection in IoT Networks", "abs": "Emerging threats to IoT networks have accelerated the development of intrusion\ndetection systems (IDSs), characterized by a shift from traditional approaches\nbased on attack signatures or anomaly detection to approaches based on machine\nlearning (ML). However, current ML-based IDSs often lack result explanations\nand struggle to address zero-day attacks due to their fixed output label space. In\nthis paper, we propose IDS-Agent, the first IDS based on an AI agent powered\nby large language models (LLMs). For each input network traffic and a detection\nrequest from the user, IDS-Agent predicts whether the traffic is benign or being\nattacked, with an explanation of the prediction results. The workflow of IDS-Agent\ninvolves iterative reasoning by a core LLM over the observation and action gen-\neration informed by the reasoning and retrieved knowledge. The action space of\nIDS-Agent includes data extraction and preprocessing, classification, knowledge\nretrieval, and results aggregation \u2013 these actions will be executed using abundant\ntools, mostly specialized for IDS. Furthermore, the IDS-Agent is equipped with\na memory and knowledge base that retains information from current and pre-\nvious sessions, along with IDS-related documents, enhancing its reasoning and\naction generation capabilities. The system prompts of IDS-Agent can be easily\ncustomized to adjust detection sensitivity or identify previously unknown types\nof attacks. In our experiments, we demonstrate the strong detection capabilities\nof IDS-Agent compared with ML-based IDSs and an IDS based on LLM with\nprompt engineering. IDS-Agent outperforms these SOTA baselines on the ACI-IoT\nand CIC-IoT benchmarks, with 0.97 and 0.75 detection F1 scores, respectively.\nIDS-Agent also achieves a recall of 0.61 in detecting zero-day attacks.", "keywords": [ "intrusion detection", "LLM agent", "internet of things", "LLM" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "uu2CorJCUi", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Rahul Madhavan", "gender": "Male", "institution": "Indian Institute of Management, Ahmedabad", "country": "IN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 22, "n_ref": 23, "n_ref_all": 31, "n_fig": 4, "n_tab": 1, "L_tab": 1643, "n_element_tab": 50, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2887, "n_element_tab_1": 64, "formula_len_all": 529, "formula_len_all_1": 529, "len_all": 75526, "len_all_1": 57137, "len_abs": 1447, "len_title": 124, "len_sents": 26147, "len_sents_1": 25105, "n_sents": 190, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1522, "title": "Adaptive Curvature Step Size: A Path Geometry Based Approach to Optimization", "abs": "We propose the Adaptive Curvature Step Size (ACSS) method, which dynamically adjusts the step size based on the local geometry of the optimization path. Our approach computes the normalized radius of curvature using consecutive gradients along the iterate path and sets the step-size equal to this radius. The effectiveness of ACSS stems from its ability to adapt to the local landscape of the optimization problem. In regions of low curvature, where consecutive gradient steps are nearly identical, ACSS allows for larger steps. Conversely, in areas of high curvature, where gradient steps differ significantly in direction, ACSS reduces the step size. This adaptive behavior enables more efficient navigation of complex loss landscapes. A key advantage of ACSS is its adaptive behavior based on local curvature information, which implicitly captures aspects of the function's second-order geometry without requiring additional memory. We provide a generalized framework for incorporating ACSS into various optimization algorithms, including SGD, Adam, AdaGrad, and RMSProp. Through extensive empirical evaluation on 20 diverse datasets, we compare ACSS variants against 12 popular optimization methods. Our results consistently show that ACSS provides performance benefits. Our results consistently show that ACSS provides performance benefits. We provide PyTorch implementations of ACSS versions for popular optimizers at our [anonymized code repository](https://anonymous.4open.science/r/curvatureStep-2a79/README.md).", "keywords": [ "Adaptive Curvature Step Size (ACSS)", "Adaptive learning rate", "Radius of curvature step size", "Low-memory optimization", "Path geometry", "Convergence analysis", "PyTorch optimizers", "SGD enhancement" ], "rating_list": [ 5, 3, 5, 5, 6 ], "soundness_list": [ 2, 1, 2, 3, 4 ], "presentation_list": [ 3, 2, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 3, 5, 4, 3, 2 ] }, { "paper_id": "utz99dx2RN", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Kun Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "LEI BAI", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Qilong Jia", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Wei Xue", "gender": "Male", "institution": "Department of C.S., Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Yi Xiao", "gender": "Non-Binary", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 10, "n_ref_uni": 26, "n_ref": 38, "n_ref_all": 56, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 118, "n_element_tab_1": 13, "formula_len_all": 1706, "formula_len_all_1": 497, "len_all": 136067, "len_all_1": 58568, "len_abs": 1560, "len_title": 146, "len_sents": 51010, "len_sents_1": 29217, "n_sents": 355, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 98, "L_abs": 1580, "title": "VAE-Var: Variational Autoencoder-Enhanced Variational Methods for Data Assimilation in Meteorology", "abs": "Data assimilation (DA) is an essential statistical technique for generating accurate estimates of a physical system's states by combining prior model predictions with observational data, especially in the realm of weather forecasting. Effectively modeling the prior distribution while adapting to diverse observational sources presents significant challenges for both traditional and neural network-based DA algorithms. This paper introduces VAE-Var, a novel neural network-based data assimilation algorithm aimed at 1) enhancing accuracy by capturing the non-Gaussian characteristics of the conditional background distribution $p(\\mathbf{x}|\\mathbf{x}_b)$, and 2) efficiently assimilating real-world observational data. VAE-Var utilizes a variational autoencoder to learn the background error distribution, with its decoder creating a variational cost function to optimize the analysis states. The advantages of VAE-Var include: 1) it maintains the framework of traditional variational assimilation, enabling it to accommodate various observation operators, particularly irregular observations; 2) it lessens the dependence on expert knowledge for constructing the background distribution, allowing for improved modeling of non-Gaussian structures; and 3) experimental results indicate that, when applied to the FengWu weather forecasting model, VAE-Var outperforms DiffDA and two traditional algorithms (interpolation and 3DVar) in terms of assimilation accuracy in sparse observational contexts, and is capable of assimilating real-world GDAS prepbufr observations over a year.", "keywords": [ "Data assimilation", "Variational Autoencoder", "Weather Forecasting" ], "rating_list": [ 5, 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 4 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 5, 4, 2 ] }, { "paper_id": "utkGLDSNOk", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Bo Jiang", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Jiaqi Zhang", "gender": "unknown", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Lu Yin", "gender": "unknown", "institution": "University of Surrey", "country": "GB", "position": "Assistant Professor" }, { "name": "Yuxiang Guo", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" } ], "n_formula": 39, "n_formula_1": 22, "n_ref_uni": 40, "n_ref": 103, "n_ref_all": 169, "n_fig": 9, "n_tab": 8, "L_tab": 2133, "n_element_tab": 331, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 2621, "n_element_tab_1": 100, "formula_len_all": 3696, "formula_len_all_1": 1522, "len_all": 185586, "len_all_1": 59531, "len_abs": 1314, "len_title": 102, "len_sents": 43266, "len_sents_1": 24402, "n_sents": 337, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1319, "title": "TODO: Enhancing LLM Alignment with Ternary Preferences", "abs": "Aligning large language models (LLMs) with human intent is critical for enhancing their performance across a variety of tasks. Standard alignment techniques, such as Direct Preference Optimization (DPO), often rely on the binary Bradley-Terry (BT) model, which can struggle to capture the complexities of human preferences\u2014particularly in the presence of noisy or inconsistent labels and frequent ties. To address these limitations, we introduce the Tie-rank Oriented Bradley-Terry model (TOBT), an extension of the BT model that explicitly incorporates ties, enabling more nuanced preference representation. Building on this, we propose Tie-rank Oriented Direct Preference Optimization (TODO), a novel alignment algorithm that leverages TOBT's ternary ranking system to improve preference alignment. In evaluations on Mistral-7B and Llama 3-8B models, TODO consistently outperforms DPO in modeling preferences across both in-distribution and out-of-distribution datasets. Additional assessments using MT Bench and benchmarks such as Piqa, ARC-c, and MMLU further demonstrate TODO's superior alignment performance. Notably, TODO also shows strong results in binary preference alignment, highlighting its versatility and potential for broader integration into LLM alignment. The code for TODO is made publicly available.", "keywords": [ "LLM", "Preference alignment", "Ternary Preference" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "uswS6tUCN2", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Arijit Sehanobish", "gender": "Male", "institution": "Kensho Technologies", "country": "US", "position": "Applied Scientist" }, { "name": "Byeongchan Kim", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Dongseok Shim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Krzysztof Marcin Choromanski", "gender": "unknown", "institution": "Google Brain Robotics & Columbia University", "country": "", "position": "research scientist & adjunct assistant professor" }, { "name": "Min-hwan Oh", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" }, { "name": "Sang Min Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Kumar A Dubey", "gender": "Male", "institution": "Google Research", "country": "", "position": "Research Scientist" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 54, "n_ref": 114, "n_ref_all": 159, "n_fig": 16, "n_tab": 13, "L_tab": 1745, "n_element_tab": 244, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 143, "n_element_tab_1": 14, "formula_len_all": 586, "formula_len_all_1": 313, "len_all": 228869, "len_all_1": 60589, "len_abs": 1355, "len_title": 60, "len_sents": 43122, "len_sents_1": 28057, "n_sents": 375, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 959, "title": "Magnituder Layers for Implicit Neural Representations in 3D", "abs": "Improving the efficiency and performance of implicit neural representations in 3D, particularly Neural Radiance Fields (NeRF) and Signed Distance Fields (SDF) is crucial for enabling their use in real-time applications. These models, while capable of generating photo-realistic novel views and detailed 3D reconstructions, often suffer from high computational costs and slow inference times. To address this, we introduce a novel neural network layer called the ``magnituder\", designed to reduce the number of training parameters in these models without sacrificing their expressive power. By integrating magnituders into standard feed-forward layer stacks, we achieve improved inference speed and adaptability. Furthermore, our approach enables a zero-shot performance boost in trained implicit neural representation models through layer-wise knowledge transfer without backpropagation, leading to more efficient scene reconstruction in dynamic environments.", "keywords": [ "NeRF", "SDF", "Implicit Representations" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "usX2ixXopC", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Guo-Ping Guo", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "", "position": "Full Professor" }, { "name": "Huan-Yu Liu", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Postdoc" }, { "name": "Junchao Wang", "gender": "Male", "institution": "Laboratory for Advanced Computing and Intelligence Engineering", "country": "", "position": "Associate Professor" }, { "name": "Tai-Ping Sun", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Xi-Ning Zhuang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Xiao-Fan Xu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Yu-Chun Wu", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "", "position": "Associate Professor" }, { "name": "Yun-Jie Wang", "gender": "Not Specified", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Zhao-Yun Chen", "gender": "Male", "institution": "Institude of Artificial Intelligence, Hefei Comprehensive National Science Center", "country": "CN", "position": "Researcher" }, { "name": "Zilei Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "ChengXue", "gender": "Male", "institution": "Institute of Artificial Intelligence, Hefei Comprehensive National Science Center ", "country": "CN", "position": "Postdoc" } ], "n_formula": 72, "n_formula_1": 6, "n_ref_uni": 28, "n_ref": 41, "n_ref_all": 78, "n_fig": 13, "n_tab": 10, "L_tab": 2124, "n_element_tab": 99, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4912, "formula_len_all_1": 330, "len_all": 183845, "len_all_1": 56020, "len_abs": 1281, "len_title": 121, "len_sents": 55204, "len_sents_1": 27711, "n_sents": 495, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1287, "title": "Measurement information multiple-reuse allows deeper quantum transformer", "abs": "The current era has witnessed the success of the transformer in the field of classical deep neural networks (DNNs) and the potential of quantum computing. One naturally expects that quantum computing can offer significant speedup for the transformer. Recent developments of quantum transformer models are faced with challenges including the expensive cost of non-linear operations and the information loss problem caused by measurements. To address this issue, this paper proposes a scheme called measurement information multiple-reuse (MIMR). MIMR enables the repeated utilization of intermediate measurement data from former layers, thus enhancing information-transferring efficiency. This scheme facilitates our quantum vision transformer (QViT) capable of achieving exponential speedup compared to classical counterparts, with the support of many parameters and large depth. Our QViT model is further examined with an instance of 86 million parameters, which halves the requirements for tomography error compared to the one without MIMR. This demonstrates the superior performance of MIMR over existing schemes. Our findings underscore the importance of exploiting the value of information from each measurement, offering a key strategy towards scalable quantum deep neural networks.", "keywords": [ "quantum machine learning", "quantum transformer", "measurement information multiple reuse" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "usFdPd4Ghs", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Anindya Bhadra", "gender": "Male", "institution": "Purdue University", "country": "", "position": "Full Professor" }, { "name": "Jorge Loria", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Postdoc" } ], "n_formula": 53, "n_formula_1": 23, "n_ref_uni": 32, "n_ref": 78, "n_ref_all": 110, "n_fig": 8, "n_tab": 7, "L_tab": 2816, "n_element_tab": 169, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2241, "n_element_tab_1": 101, "formula_len_all": 4172, "formula_len_all_1": 1341, "len_all": 150373, "len_all_1": 73203, "len_abs": 2155, "len_title": 117, "len_sents": 46388, "len_sents_1": 30954, "n_sents": 375, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 2169, "title": "Deep Kernel Posterior Learning under Infinite Variance Prior Weights", "abs": "Neal (1996) proved that infinitely wide shallow Bayesian neural networks (BNN) converge to Gaussian processes (GP), when the network weights have bounded prior variance. Cho & Saul (2009) provided a useful recursive formula for deep kernel processes for relating the covariance kernel of each layer to the layer immediately below. Moreover, they worked out the form of the layer-wise covariance kernel in an explicit manner for several common activation functions, including the ReLU. Subsequent works have made the connection between these two works, and provided useful results on the covariance kernel of a deep GP arising as wide limits of various deep Bayesian network architectures. However, recent works, including Aitchison et al. (2021), have highlighted that the covariance kernels obtained in this manner are deterministic and hence, precludes any possibility of representation learning, which amounts to learning a non-degenerate posterior of a random kernel given the data. To address this, they propose adding artificial noise to the kernel to retain stochasticity, and develop deep kernel Wishart and inverse Wishart processes. Nonetheless, this artificial noise injection could be critiqued in that it would not naturally emerge in a classic BNN architecture under an infinite-width limit. To address this, we show that a Bayesian deep neural network, where each layer width approaches infinity, and all network weights are elliptically distributed with infinite variance, converges to a process with $\\alpha$-stable marginals in each layer that has a conditionally Gaussian representation. These conditional random covariance kernels could be recursively linked in the manner of Cho & Saul (2009), even though marginally the process exhibits stable behavior, and hence covariances are not even necessarily defined. We also provide useful generalizations of the recent results of Lor\u00eda & Bhadra (2024) on shallow networks to multi-layer networks, and remedy the prohibitive computational burden of their approach. The computational and statistical benefits over competing approaches stand out in simulations and in demonstrations on benchmark data sets.", "keywords": [ "Kernel methods", "Deep Gaussian processes", "Infinite variance priors", "Deep Bayesian neural networks" ], "rating_list": [ 8, 8, 5, 8, 5 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 2, 3, 2, 4, 2 ] }, { "paper_id": "us5riDkeBW", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Amirkeivan Mohtashami", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "", "position": "PhD student" }, { "name": "Blaise Aguera y Arcas", "gender": "Male", "institution": "Princeton University", "country": "", "position": "MS student" }, { "name": "Florian Hartmann", "gender": "Male", "institution": "Google DeepMind", "country": "CH", "position": "Researcher" }, { "name": "Lukas Zilka", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Matt Sharifi", "gender": "Male", "institution": "Google Research", "country": "", "position": "Researcher" }, { "name": "Sian Gooding", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 60, "n_ref_all": 85, "n_fig": 3, "n_tab": 11, "L_tab": 6165, "n_element_tab": 690, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 917, "n_element_tab_1": 177, "formula_len_all": 1, "formula_len_all_1": 0, "len_all": 194988, "len_all_1": 72675, "len_abs": 945, "len_title": 122, "len_sents": 56867, "len_sents_1": 35064, "n_sents": 509, "n_sents_1": 280, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 951, "title": "Social Learning: Towards Collaborative Learning with Large Language Models", "abs": "We introduce the framework of \"social learning\" in the context of large language models (LLMs), whereby models share knowledge with each other in a privacy-aware manner using natural language. We present and evaluate two approaches for knowledge transfer between LLMs. In the first scenario, we allow the model to generate abstract prompts aiming to teach the task. In our second approach, models transfer knowledge by generating synthetic examples. We evaluate these methods across diverse datasets and quantify memorization as a proxy for privacy loss. These techniques inspired by social learning yield promising results with low memorization of the original data. In particular, we show that performance using these methods is comparable to results with the use of original labels and prompts. Our work demonstrates the viability of social learning for LLMs, establishes baseline approaches and highlights several unexplored areas for future work.", "keywords": [ "language models", "privacy-aware knowledge transfer" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "ursX3k1rTO", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Andrey E Ustyuzhanin", "gender": "Male", "institution": "Constructor University", "country": "DE", "position": "Researcher" }, { "name": "Ignat Romanov", "gender": "unknown", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "MS student" }, { "name": "Kedar Hippalgaonkar", "gender": "unknown", "institution": "Nanyang Technological University", "country": "", "position": "Associate Professor" }, { "name": "Nikita Kazeev", "gender": "unknown", "institution": "NUS - National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Ruiming Zhu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Shuya Yamazaki", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Wei Nong", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 34, "n_ref": 55, "n_ref_all": 68, "n_fig": 7, "n_tab": 12, "L_tab": 4081, "n_element_tab": 471, "n_fig_1": 5, "n_tab_1": 9, "L_tab_1": 2633, "n_element_tab_1": 256, "formula_len_all": 211, "formula_len_all_1": 0, "len_all": 145370, "len_all_1": 61239, "len_abs": 1278, "len_title": 102, "len_sents": 42054, "len_sents_1": 27449, "n_sents": 290, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1289, "title": "Wyckoff Transformer: Generation of Symmetric Crystals", "abs": "We propose Wyckoff Transformer, a generative model for materials conditioned on space group symmetry. Most real--world inorganic materials have internal symmetry beyond lattice translation. Symmetry rules that atoms obey play a fundamental role in determining the physical, chemical, and electronic properties of crystals. These symmetries determine stability, and influence key material structural and functional properties such as electrical and thermal conductivity, optical and polarization behavior, and mechanical strength. And yet, despite the recent advancements, state--of--the--art diffusion models struggle to generate highly symmetric crystals. We use Wyckoff positions as the basis for an elegant, compressed, and discrete structure representation. To model the distribution we develop a permutation--invariant autoregressive model based on Transformer and absence of positional encoding. Our experiments demonstrate that Wyckoff Transformer has the best performance in generating novel diverse stable structures conditioned on the symmetry space group, while also having competitive metric values when compared to model not conditioned on symmetry. We also show that it is competitive in predicting formation energy, band gap, mechanical properties, and thermal conductivity.", "keywords": [ "material design", "machine learning", "crystal generation", "space group symmetry", "Transformer", "Wyckoff position", "generative model", "autoregressive model", "permutation invariance" ], "rating_list": [ 3, 6, 6, 5, 5 ], "soundness_list": [ 1, 2, 3, 3, 2 ], "presentation_list": [ 1, 1, 2, 2, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 3, 3 ] }, { "paper_id": "urf8a5G59f", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Amir Jamaludin", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Senior Research Associate" }, { "name": "Emmanuelle Bourigault", "gender": "Female", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Abdullah Jamal Hamdi", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" } ], "n_formula": 17, "n_formula_1": 8, "n_ref_uni": 59, "n_ref": 166, "n_ref_all": 213, "n_fig": 13, "n_tab": 9, "L_tab": 879, "n_element_tab": 60, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 168, "n_element_tab_1": 30, "formula_len_all": 1028, "formula_len_all_1": 474, "len_all": 268894, "len_all_1": 57613, "len_abs": 1784, "len_title": 154, "len_sents": 71161, "len_sents_1": 25764, "n_sents": 575, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1798, "title": "X-Diffusion: Generating Detailed 3D MRI Volumes From a Single Image Using Cross-Sectional Diffusion Models", "abs": "Magnetic Resonance Imaging (MRI) is a crucial diagnostic tool, but high-resolution scans are often slow and expensive due to extensive data acquisition requirements. Traditional MRI reconstruction methods aim to expedite this process by filling in missing frequency components in the K-space, performing _3D-to-3D_ reconstructions that demand full 3D scans. In contrast, we introduce _X-Diffusion_, a novel cross-sectional diffusion model that reconstructs detailed 3D MRI volumes from extremely sparse spatial-domain inputs\u2014achieving _2D-to-3D_ reconstruction from as little as a single 2D MRI slice or few slices.\nA key aspect of X-Diffusion is that it models MRI data as holistic 3D volumes during the cross-sectional training and inference, unlike previous learning approaches that treat MRI scans as collections of 2D slices in standard planes (coronal, axial, sagittal).\nWe evaluated X-Diffusion on brain tumor MRIs from the BRATS dataset and full-body MRIs from the UK Biobank dataset. Our results demonstrate that X-Diffusion not only surpasses state-of-the-art methods in quantitative accuracy (PSNR) on unseen data but also preserves critical anatomical features such as tumor profiles, spine curvature, and brain volume. Remarkably, the model generalizes beyond the training domain, successfully reconstructing knee MRIs despite being trained exclusively on brain data. Medical expert evaluations further confirm the clinical relevance and fidelity of the generated images.\nTo promote reproducibility and trust in our findings, we will publicly release the accompanying code upon publication. To our knowledge, X-Diffusion is the first method capable of producing detailed 3D MRIs from highly limited 2D input data, potentially accelerating MRI acquisition and reducing associated costs.", "keywords": [ "MRI reconstruction", "diffusion models", "latent diffusions" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 4, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "urcEYsZOBz", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Peng Chen", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Phillip Si", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 21, "n_ref_uni": 20, "n_ref": 32, "n_ref_all": 59, "n_fig": 10, "n_tab": 5, "L_tab": 517, "n_element_tab": 21, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 115, "n_element_tab_1": 5, "formula_len_all": 1152, "formula_len_all_1": 1152, "len_all": 100433, "len_all_1": 62910, "len_abs": 1353, "len_title": 111, "len_sents": 35899, "len_sents_1": 31924, "n_sents": 197, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 111, "L_abs": 1355, "title": "Latent-EnSF: A Latent Ensemble Score Filter for High-Dimensional Data Assimilation with Sparse Observation Data", "abs": "Accurate modeling and prediction of complex physical systems often rely on data assimilation techniques to correct errors inherent in model simulations. Traditional methods like the Ensemble Kalman Filter (EnKF) and its variants as well as the recently developed Ensemble Score Filters (EnSF) face significant challenges when dealing with high-dimensional and nonlinear Bayesian filtering problems with sparse observations, which are ubiquitous in real-world applications. In this paper, we propose a novel data assimilation method, Latent-EnSF, which leverages EnSF with efficient and consistent latent representations of the full states and sparse observations to address the joint challenges of high dimensionlity in states and high sparsity in observations for nonlinear Bayesian filtering. We introduce a coupled \nVariational Autoencoder (VAE) with two encoders to encode the full states and sparse observations in a consistent way guaranteed by a latent distribution matching and regularization as well as a consistent state reconstruction. With comparison to several methods, we demonstrate the higher accuracy, faster convergence, and higher efficiency of Latent-EnSF for two challenging applications with complex models in shallow water wave propagation and medium-range weather forecasting, for highly sparse observations in both space and time.", "keywords": [ "Data Assimilation", "Score Based Models", "Diffusion Models", "Weather Forecasting" ], "rating_list": [ 6, 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "urQi0TgXFY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Christian Schroeder de Witt", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Lecturer" }, { "name": "Dylan Cope", "gender": "unknown", "institution": "King's College London, University of London", "country": "", "position": "PhD student" }, { "name": "Joan Velja", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "MS student" }, { "name": "Nandi Schoots", "gender": "Female", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Ollie Matthews", "gender": "Male", "institution": "LASR Labs", "country": "GB", "position": "Researcher" }, { "name": "Robert McCarthy", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Yohan Mathew", "gender": "Male", "institution": "Yohan Mathew (Independent)", "country": "CA", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 44, "n_ref": 75, "n_ref_all": 102, "n_fig": 7, "n_tab": 11, "L_tab": 9126, "n_element_tab": 125, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1359, "n_element_tab_1": 57, "formula_len_all": 49, "formula_len_all_1": 0, "len_all": 241152, "len_all_1": 68349, "len_abs": 1325, "len_title": 80, "len_sents": 89136, "len_sents_1": 32552, "n_sents": 756, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1331, "title": "Hidden in Plain Text: Emergence & Mitigation of Steganographic Collusion in LLMs", "abs": "The rapid proliferation of frontier model agents promises significant societal advances but also raises concerns about systemic risks arising from unsafe interactions. Collusion to the disadvantage of others has been identified as a central form of undesirable agent cooperation. The use of information hiding (steganography) in agent communications could render collusion practically undetectable. This underscores the need for evaluation frameworks to monitor and mitigate steganographic collusion capabilities. We demonstrate, for the first time, that robust steganographic collusion in LLMs can arise indirectly from optimization pressure. To investigate this problem we design two approaches -- a gradient-based reinforcement learning (GBRL) method and an in-context reinforcement learning (ICRL) method -- for reliably eliciting sophisticated LLM-generated linguistic text steganography. Importantly, we find that emergent steganographic collusion can be robust to both passive steganalytic oversight of model outputs and active mitigation through communication paraphrasing. We contribute a novel model evaluation framework and discuss limitations and future work. Our findings imply that effective risk mitigation from steganographic collusion post-deployment requires innovation in passive and active oversight techniques.", "keywords": [ "Large Language Models", "Steganography", "Collusion", "Reinforcement Learning", "In-Context Learning", "Multi-agent Systems" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "uqe5HkjbT9", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Hyungho Na", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Il-chul Moon", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Kwanghyeon Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Sumin Lee", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 18, "n_formula_1": 10, "n_ref_uni": 44, "n_ref": 101, "n_ref_all": 174, "n_fig": 25, "n_tab": 8, "L_tab": 3441, "n_element_tab": 188, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 721, "n_element_tab_1": 8, "formula_len_all": 1645, "formula_len_all_1": 1025, "len_all": 171947, "len_all_1": 58397, "len_abs": 1538, "len_title": 106, "len_sents": 51667, "len_sents_1": 24523, "n_sents": 409, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1549, "title": "Trajectory-Class-Aware Multi-Agent Reinforcement Learning", "abs": "In the context of multi-agent reinforcement learning, *generalization* is a challenge to solve various tasks that may require different joint policies or coordination without relying on policies specialized for each task. We refer to this type of problem as a *multi-task*, and we train agents to be versatile in this multi-task setting through a single training process. To address this challenge, we introduce TRajectory-class-Aware Multi-Agent reinforcement learning (TRAMA). In TRAMA, agents recognize a task type by identifying the class of trajectories they are experiencing through partial observations, and the agents use this trajectory awareness or prediction as additional information for action policy. To this end, we introduce three primary objectives in TRAMA: (a) constructing a quantized latent space to generate trajectory embeddings that reflect key similarities among them; (b) conducting trajectory clustering using these trajectory embeddings; and (c) building a trajectory-class-aware policy. Specifically for (c), we introduce a trajectory-class predictor that performs agent-wise predictions on the trajectory class; and we design a trajectory-class representation model for each trajectory class. Each agent takes actions based on this trajectory-class representation along with its partial observation for task-aware execution. The proposed method is evaluated on various tasks, including multi-task problems built upon StarCraft II. Empirical results show further performance improvements over state-of-the-art baselines.", "keywords": [ "trajectory clustering", "multi-agent reinforcement learning", "trajectory-class-aware policy" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "uqWM9hBDAE", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Marcel Boehme", "gender": "Male", "institution": "Max Planck Institute for Security and Privacy", "country": "", "position": "Researcher" }, { "name": "Seongmin Lee", "gender": "Male", "institution": "Max-Planck Institute for Security and Privacy", "country": "", "position": "Postdoc" } ], "n_formula": 88, "n_formula_1": 34, "n_ref_uni": 15, "n_ref": 53, "n_ref_all": 97, "n_fig": 5, "n_tab": 5, "L_tab": 2830, "n_element_tab": 328, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 77, "n_element_tab_1": 3, "formula_len_all": 8856, "formula_len_all_1": 3141, "len_all": 155630, "len_all_1": 75103, "len_abs": 1093, "len_title": 65, "len_sents": 51152, "len_sents_1": 30484, "n_sents": 460, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 64, "L_abs": 1230, "title": "How Much is Unseen Depends Chiefly on Information About the Seen", "abs": "The *missing mass* refers to the proportion of data points in an *unknown* population of classifier inputs that belong to classes *not* present in the classifier's training data, which is assumed to be a random sample from that unknown population.\nWe find that *in expectation* the missing mass is entirely determined by the number $f_k$ of classes that *do* appear in the training data the same number of times *and an exponentially decaying error*.\nWhile this is the first precise characterization of the expected missing mass in terms of the sample, the induced estimator suffers from an impractically high variance. However, our theory suggests a large search space of nearly unbiased estimators that can be searched effectively and efficiently. Hence, we cast distribution-free estimation as an optimization problem to find a distribution-specific estimator with a minimized mean-squared error (MSE), given only the sample.\nIn our experiments, our search algorithm discovers estimators that have a substantially smaller MSE than the state-of-the-art Good-Turing estimator. This holds for over 93\\% of runs when there are at least as many samples as classes. Our estimators' MSE is roughly 80\\% of the Good-Turing estimator's.", "keywords": [ "Good-Turing frequency estimation", "Multinomial probability estimation", "Unseen events", "Missing mass", "Probability mass" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "uqLQjtSdFN", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Aaron T Wang", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Undergrad student" }, { "name": "Lawrence Carin", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Full Professor" }, { "name": "Ricardo Henao", "gender": "Male", "institution": "Duke University", "country": "", "position": "Associate Professor" }, { "name": "Xiang Cheng", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" } ], "n_formula": 80, "n_formula_1": 45, "n_ref_uni": 27, "n_ref": 158, "n_ref_all": 183, "n_fig": 5, "n_tab": 1, "L_tab": 32, "n_element_tab": 1, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6338, "formula_len_all_1": 2723, "len_all": 159261, "len_all_1": 81428, "len_abs": 1500, "len_title": 124, "len_sents": 57879, "len_sents_1": 34844, "n_sents": 437, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1552, "title": "Functional Gradients and Generalizations for Transformer In-Context Learning", "abs": "We examine Transformer-based in-context learning for contextual data of the form $(x_i,y_i)$ for $i=1,\\ldots,N$, and query $x_{N+1}$, where $x_i\\in\\Bbb{R}^d$ and $y_i\\sim p(Y|f(x_i))$, with $f(x)$ a latent function. This is analyzed from the perspective of *functional* gradient descent for latent $f(x)$. We initially perform this analysis from the perspective of a reproducing kernel Hilbert space (RKHS), from which an alternative kernel-averaging perspective is manifested. This leads to a generalization, allowing an interpretation of softmax attention from the perspective of the Nadaraya-Watson kernel-weighted average. We show that a single attention layer may be designed to exactly implement a functional-gradient step in this setting (for RKHS latent functions), extending prior work for the special case of real-valued $Y$ and Gaussian $p(Y|f(x))$. This is also generalized for softmax attention and non-RKHS underlying $f(x)$. Though our results hold in a general setting, we focus on categorical $Y$ with $p(Y|f(x))$ modeled as a generalized linear model (corresponding specifically to softmax probability). Multi-layered extensions are developed for this case, and through extensive experimentation we demonstrate that for categorical $Y$ a single-layer model is often highly effective for such in-context learning. We also demonstrate these ideas for real-world data, considering in-context classification of ImageNet data, showing the broad applicability of our theory beyond the commonly-studied settings of synthetic regression data.", "keywords": [ "Transformer", "in-context learning" ], "rating_list": [ 3, 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 1, 2, 2 ], "contribution_list": [ 2, 1, 2, 1, 1 ], "confidence_list": [ 3, 4, 2, 3, 3 ] }, { "paper_id": "uqG0kFLccD", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Burak Bekci", "gender": "Male", "institution": "Rohde & Schwarz GmbH & Co. KG", "country": "DE", "position": "Data Engineer" }, { "name": "Federico Tombari", "gender": "Male", "institution": "Google", "country": "CH", "position": "Research Director" }, { "name": "Mahdi Saleh", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "PhD student" }, { "name": "Nassir Navab", "gender": "Male", "institution": "Technical University of Munich", "country": "DE", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 22, "n_ref": 32, "n_ref_all": 49, "n_fig": 3, "n_tab": 4, "L_tab": 3173, "n_element_tab": 56, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 326, "n_element_tab_1": 38, "formula_len_all": 331, "formula_len_all_1": 332, "len_all": 107684, "len_all_1": 47574, "len_abs": 1267, "len_title": 63, "len_sents": 28016, "len_sents_1": 23186, "n_sents": 197, "n_sents_1": 166, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1271, "title": "ESCAPE: Equivariant Shape Completion via Anchor Point Encoding", "abs": "Shape completion, a crucial task in 3D computer vision, involves predicting and filling the missing regions of scanned or partially observed objects. Current methods often suffer from orientation-dependent inconsistencies, particularly under varying rotations, limiting their real-world applicability. We introduce ESCAPE (Equivariant Shape Completion via Anchor Point Encoding), a novel framework designed to achieve rotation-equivariant shape completion. Our approach employs a distinctive encoding strategy, representing objects by selecting anchor points and utilizing them in a distance-based encoder akin to the D2 shape distribution. This enables the model to capture a consistent, rotation-equivariant understanding of the object\u2019s geometry. ESCAPE leverages a transformer architecture to encode and decode the distance transformations, ensuring that generated shape completions remain accurate and equivariant under rotational transformations. Additionally, we perform optimization to refine the predicted shapes from anchor point positions and predicted encodings, Experimental evaluations demonstrate that ESCAPE achieves robust, high-quality reconstructions across arbitrary rotations and translations, showcasing its effectiveness in real-world applications.", "keywords": [ "3D Shape Completion", "Rotation Equivariance" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 1, 4, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "uq9TLFT7tF", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Houqiang Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Jing Liao", "gender": "Female", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Jing Nathan Yan", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Qi Sun", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Shengming Yin", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Wengang Zhou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Zhiyang Guo", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Ziyu Wan", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 53, "n_ref": 173, "n_ref_all": 213, "n_fig": 23, "n_tab": 5, "L_tab": 1131, "n_element_tab": 146, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 950, "n_element_tab_1": 85, "formula_len_all": 267, "formula_len_all_1": 267, "len_all": 197379, "len_all_1": 57099, "len_abs": 1435, "len_title": 113, "len_sents": 48259, "len_sents_1": 26223, "n_sents": 385, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1460, "title": "EG4D: Explicit Generation of 4D Object without Score Distillation", "abs": "In recent years, the increasing demand for dynamic 3D assets in design and gaming applications has given rise to powerful generative pipelines capable of synthesizing high-quality 4D objects.\n Previous methods generally rely on score distillation sampling (SDS) algorithm to infer the unseen views and motion of 4D objects, thus leading to unsatisfactory results with defects like over-saturation and Janus problem.\n Therefore, inspired by recent progress of video diffusion models, we propose to optimize a 4D representation by explicitly generating multi-view videos from one input image.\n However, it is far from trivial to handle practical challenges faced by such a pipeline, including dramatic temporal inconsistency, inter-frame geometry and texture diversity, and semantic defects brought by video generation results.\n To address these issues, we propose EG4D, a novel multi-stage framework that generates high-quality and consistent 4D assets without score distillation.\n Specifically, collaborative techniques and solutions are developed, including an attention injection strategy to synthesize temporal-consistent multi-view videos, a robust and efficient dynamic reconstruction method based on Gaussian Splatting, and a refinement stage with diffusion prior for semantic restoration.\n The qualitative comparisons and quantitative results demonstrate that our framework outperforms the baselines in generation quality by a considerable margin.", "keywords": [ "4D Generation" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "upzyG4wRBr", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Chao Wen", "gender": "Male", "institution": "MPI-SWS", "country": "DE", "position": "PhD student" }, { "name": "Jacqueline Staub", "gender": "Female", "institution": "University of Trier", "country": "DE", "position": "Assistant Professor" }, { "name": "Adish Kumar Singla", "gender": "unknown", "institution": "Max Planck Institute for Software Systems (MPI-SWS)", "country": "DE", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 38, "n_ref": 80, "n_ref_all": 133, "n_fig": 31, "n_tab": 2, "L_tab": 293, "n_element_tab": 41, "n_fig_1": 18, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 223, "formula_len_all_1": 53, "len_all": 210649, "len_all_1": 68889, "len_abs": 1347, "len_title": 125, "len_sents": 76816, "len_sents_1": 33470, "n_sents": 704, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1360, "title": "Program Synthesis Benchmark for Visual Programming in XLogoOnline Environment", "abs": "Large language and multimodal models have shown remarkable success on various benchmarks focused on specific skills such as general-purpose programming, natural language understanding, math word problem-solving, and visual question answering. However, it is unclear how well these models perform on tasks that require a combination of these skills. In this paper, we curate a novel program synthesis benchmark based on the real-world tasks in the XLogoOnline visual programming environment. Each task requires a combination of different skills such as spatial planning, basic programming, and logical reasoning. Our evaluation shows that current state-of-the-art models like GPT-4V and Llama3-70B struggle to solve these tasks, achieving only $20$\\% and $2.35$\\% success rates, respectively. Next, we develop a fine-tuning pipeline to boost the performance of models by leveraging a large-scale synthetic training dataset with over $80,000$ tasks. Moreover, we showcase how emulator-driven feedback can be used to design a curriculum over training data distribution, through which a fine-tuned Llama3-8B drastically outperforms GPT-4V and Llama3-70B models. Finally, we provide an in-depth failure analysis to understand the limitations of different models. We will publicly release the benchmark for future research on program synthesis in visual programming.", "keywords": [ "Program Synthesis", "Visual Programming", "Large Language Models", "Multimodal Models", "Spatial Reasoning" ], "rating_list": [ 8, 8, 3, 5, 5 ], "soundness_list": [ 3, 4, 2, 2, 3 ], "presentation_list": [ 4, 4, 2, 3, 4 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 3, 4 ] }, { "paper_id": "upoxXRRTQ2", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Ofir Schlisselberg", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "MS student" }, { "name": "Ran Darshan", "gender": "Not Specified", "institution": "Tel Aviv University", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 51, "n_formula_1": 15, "n_ref_uni": 47, "n_ref": 60, "n_ref_all": 77, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4197, "formula_len_all_1": 673, "len_all": 194695, "len_all_1": 61067, "len_abs": 1350, "len_title": 97, "len_sents": 67836, "len_sents_1": 27443, "n_sents": 572, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1360, "title": "The impact of allocation strategies in subset learning on the expressive power of neural networks", "abs": "In traditional machine learning, models are defined by a set of parameters, which are optimized to perform specific tasks. In neural networks, these parameters correspond to the synaptic weights. However, in reality, it is often infeasible to control or update all weights. This challenge is not limited to artificial networks but extends to biological networks, such as the brain, where the extent of distributed synaptic weight modification during learning remains unclear. Motivated by these insights, we theoretically investigate how different allocations of a fixed number of learnable weights influence the capacity of neural networks. Using a teacher-student setup, we introduce a benchmark to quantify the expressivity associated with each allocation. We establish conditions under which allocations have \\`maximal' or \\`minimal' expressive power in linear recurrent neural networks and linear multi-layer feedforward networks. For suboptimal allocations, we propose heuristic principles to estimate their expressivity. These principles extend to shallow ReLU networks as well. Finally, we validate our theoretical findings with empirical experiments. Our results emphasize the critical role of strategically distributing learnable weights across the network, showing that a more widespread allocation generally enhances the network\u2019s expressive power.", "keywords": [ "subset learning", "theoretical neuroscience", "expressive power", "neural networks", "recurrent neural network" ], "rating_list": [ 8, 6, 3, 3 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 4, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "upkxzurnLC", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Giovanni Montana", "gender": "Male", "institution": "The University of Warwick", "country": "", "position": "Full Professor" }, { "name": "Yue Jin", "gender": "unknown", "institution": "University of Warwick", "country": "", "position": "Postdoc" }, { "name": "MIANCHU WANG", "gender": "unknown", "institution": "University of Warwick", "country": "GB", "position": "PhD student" } ], "n_formula": 58, "n_formula_1": 36, "n_ref_uni": 32, "n_ref": 93, "n_ref_all": 113, "n_fig": 12, "n_tab": 2, "L_tab": 1694, "n_element_tab": 173, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1694, "n_element_tab_1": 173, "formula_len_all": 4759, "formula_len_all_1": 2079, "len_all": 138286, "len_all_1": 73698, "len_abs": 944, "len_title": 129, "len_sents": 41060, "len_sents_1": 30779, "n_sents": 339, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 948, "title": "Learning on One Mode: Addressing Multi-modality in Offline Reinforcement Learning", "abs": "Offline reinforcement learning (RL) seeks to learn optimal policies from static datasets without interacting with the environment. A common challenge is handling multi-modal action distributions, where multiple behaviours are represented in the data. Existing methods often assume unimodal behaviour policies, leading to suboptimal performance when this assumption is violated. We propose Weighted Imitation Learning on One Mode (LOM), a novel approach that focuses on learning from a single, promising mode of the behaviour policy. By using a Gaussian mixture model to identify modes and selecting the best mode based on expected returns, LOM avoids the pitfalls of averaging over conflicting actions. Theoretically, we show that LOM improves performance while maintaining simplicity in policy learning. Empirically, LOM outperforms existing methods on standard D4RL benchmarks and demonstrates its effectiveness in complex, multi-modal scenarios.", "keywords": [ "Offline reinforcement learning", "weighted imitation learning", "multi-modality." ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "upV91V0Big", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Jifeng Hu", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "PhD student" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Shengchao Hu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Ya Zhang", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Professor" }, { "name": "Yuhang Zhou", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Ziqing Fan", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 38, "n_formula_1": 19, "n_ref_uni": 40, "n_ref": 114, "n_ref_all": 143, "n_fig": 10, "n_tab": 4, "L_tab": 2867, "n_element_tab": 493, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 1273, "n_element_tab_1": 246, "formula_len_all": 1685, "formula_len_all_1": 886, "len_all": 170748, "len_all_1": 77043, "len_abs": 1668, "len_title": 112, "len_sents": 51753, "len_sents_1": 33383, "n_sents": 379, "n_sents_1": 237, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1422, "title": "Continual Task Learning through Adaptive Policy Self-Composition", "abs": "Training a generalizable agent to continually learn a sequence of tasks from offline trajectories is a natural requirement for long-lived agents, yet remains a significant challenge for current offline reinforcement learning (RL) algorithms. Specifically, an agent must be able to rapidly adapt to new tasks using newly collected trajectories (plasticity), while retaining knowledge from previously learned tasks (stability). \nHowever, systematic analyses of this setting are scarce, and it remains unclear whether conventional continual learning (CL) methods are effective in continual offline RL (CORL) scenarios. In this study, we develop the Offline Continual World benchmark and demonstrate that traditional CL methods struggle with catastrophic forgetting, primarily due to the unique distribution shifts inherent to CORL scenarios.\nTo address this challenge, we introduce CompoFormer, a structure-based continual transformer model that adaptively composes previous policies via a meta-policy network. Upon encountering a new task, CompoFormer leverages semantic correlations to selectively integrate relevant prior policies alongside newly trained parameters, thereby enhancing knowledge sharing and accelerating the learning process.\nOur experiments reveal that CompoFormer outperforms conventional CL methods, particularly in longer task sequences, showcasing a promising balance between plasticity and stability.", "keywords": [ "continual learning", "offline reinforcement learning" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "upALuXjdxc", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Han Yu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hao Zou", "gender": "Male", "institution": "ZGC laboratory", "country": "CN", "position": "Researcher" }, { "name": "Jiashuo Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Peng Cui", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Renzhe Xu", "gender": "Male", "institution": "Shanghai University of Finance and Economics", "country": "CN", "position": "Assistant Professor" }, { "name": "Xingxuan Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yue He", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 51, "n_ref": 117, "n_ref_all": 159, "n_fig": 15, "n_tab": 9, "L_tab": 3049, "n_element_tab": 438, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 555, "n_element_tab_1": 75, "formula_len_all": 690, "formula_len_all_1": 471, "len_all": 202263, "len_all_1": 70817, "len_abs": 1533, "len_title": 95, "len_sents": 62481, "len_sents_1": 32632, "n_sents": 476, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1539, "title": "Error Slice Discovery via Manifold Compactness", "abs": "Despite the great performance of deep learning models in many areas, they still make mistakes and underperform on certain subsets of data, i.e. error slices. Given a trained model, it is important to identify its semantically coherent error slices that are easy to interpret, which is referred to as the error slice discovery problem. However, there is no proper metric of slice coherence without relying on extra information like predefined slice labels. The current evaluation of slice coherence requires access to predefined slices formulated by metadata like attributes or subclasses. Its validity heavily relies on the quality and abundance of metadata, where some possible patterns could be ignored. Besides, current algorithms cannot directly incorporate the constraint of coherence into their optimization objective due to the absence of an explicit coherence metric, which could potentially hinder their effectiveness. In this paper, we propose manifold compactness, a coherence metric without reliance on extra information by incorporating the data geometry property into its design, and experiments on typical datasets empirically validate the rationality of the metric. Then we develop Manifold Compactness based error Slice Discovery (MCSD), a novel algorithm that directly treats risk and coherence as the optimization objective, and is flexible to be applied to models of various tasks. Extensive experiments on the current benchmark and case studies on other typical datasets demonstrate the effectiveness of our algorithm.", "keywords": [ "Error Slice Discovery", "Manifold Compactness", "Model Evaluation" ], "rating_list": [ 5, 5, 8, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "uogG8BfLs2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Changho Shin", "gender": "unknown", "institution": "University of Wisconsin, Madison", "country": "", "position": "PhD student" }, { "name": "Frederic Sala", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" }, { "name": "John Cooper", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "PhD student" } ], "n_formula": 53, "n_formula_1": 5, "n_ref_uni": 44, "n_ref": 83, "n_ref_all": 102, "n_fig": 10, "n_tab": 7, "L_tab": 2136, "n_element_tab": 110, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 180, "n_element_tab_1": 21, "formula_len_all": 8567, "formula_len_all_1": 600, "len_all": 205005, "len_all_1": 66248, "len_abs": 4792, "len_title": 107, "len_sents": 62373, "len_sents_1": 31333, "n_sents": 533, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1338, "title": "Weak-to-Strong Generalization Through the Data-Centric Lens", "abs": "The weak-to-strong generalization phenomenon is the driver for important machine learning applications including highly data-efficient learning and, most recently, performing superalignment. While decades of research have resulted in numerous algorithms that produce strong empirical performance, understanding what aspects of data enable weak-to-strong generalization has been understudied. We propose a simple data-centric mechanism that characterizes weak-to-strong generalization: the overlap density. Intuitively, generalization tracks the number of points that contain overlaps, i.e., both easy patterns (learnable by a weak model) and challenging patterns (only learnable by a stronger model), as with such points, weak predictions can be used to learn challenging patterns by stronger models. And, we provide a practical overlap detection algorithm to find overlap density from data. Finally, we provide an algorithm to learn, among multiple sources of data, which to query when seeking to maximize overlap density and thereby enhance weak-to-strong generalization. We provide a theoretical result showing that the generalization benefit is a function of the overlap density and a regret bound of our data selection algorithm. Empirically, we validate the mechanism and the overlap detection algorithm on a wide array of settings.", "keywords": [ "weak to strong generalization", "data-centric AI" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "uoU4ypjAmN", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Arnav Kundu", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Duc N.M Hoang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Han-Byul Kim", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Minsik Cho", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "PhD student" }, { "name": "Mohammad Samragh", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 20, "n_ref": 48, "n_ref_all": 74, "n_fig": 11, "n_tab": 5, "L_tab": 2815, "n_element_tab": 561, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 829, "n_element_tab_1": 32, "formula_len_all": 282, "formula_len_all_1": 282, "len_all": 136880, "len_all_1": 60040, "len_abs": 1276, "len_title": 127, "len_sents": 28830, "len_sents_1": 28203, "n_sents": 196, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1281, "title": "SPD: Sync-Point Drop for efficient tensor parallelism of Large Language Models", "abs": "With the rapid expansion in the scale of large language models (LLMs), enabling efficient distributed inference across multiple computing units has become increasingly critical. However, communication overheads from frequent synchronization during distributed inference pose a significant challenge to achieve scalability and low latency. Therefore, we introduce a novel optimization technique, Sync-Point Drop (SPD) to reduce communication overheads in tensor parallelism by dropping synchronization on attention outputs. In detail, we first propose a block design that allows execution to proceed without communication through SPD. Second, we identify regions of communication redundancy, where dropping synchronization results in no loss of model performance. In addition, to extend SPD across all compute blocks, we employ a low-cost distillation, specifically targeting blocks giving quality degradation, to maximize accuracy recovery. For extreme blocks where performance degradation is severe, we introduce a new head grouping enhancements to amplify the distillation\u2019s recovery effect. The proposed methods effectively alleviate communication bottlenecks while minimizing accuracy degradation during LLM inference, offering a scalable solution for distributed environments.", "keywords": [ "sync point drop", "tensor parallelism", "distributed inference", "efficient ml" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "uo8PO6Ah59", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Cedric Caruzzo", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Jong Chul Ye", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 24, "n_formula_1": 12, "n_ref_uni": 24, "n_ref": 55, "n_ref_all": 68, "n_fig": 6, "n_tab": 5, "L_tab": 2596, "n_element_tab": 173, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 770, "n_element_tab_1": 136, "formula_len_all": 1015, "formula_len_all_1": 565, "len_all": 157806, "len_all_1": 67670, "len_abs": 1782, "len_title": 130, "len_sents": 58721, "len_sents_1": 34051, "n_sents": 408, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1484, "title": "CellPainTR: Contrastive Batch Corrected Transformer for Large Scale Cell Painting", "abs": "Cell Painting, a high-content imaging-based profiling method, has emerged as a powerful tool for understanding cellular phenotypes and drug responses. However, batch effects severely constrain the integration and interpretation of data collected across different laboratories and experimental conditions. This paper introduces CellPainTR, a novel approach for unified batch correction and representation learning in Cell Painting data, addressing a critical challenge in the field of image-based profiling. Our approach employs a Transformer-like architecture with Hyena operators, positional encoding via morphological-feature-embedding, and a special source context token for batch correction,\ncombined with a multi-stage training process that incorporates masked token prediction and supervised contrastive learning. Experiments on the JUMP Cell Painting dataset demonstrate that CellPainTR significantly outperforms existing approaches such as Combat and Harmony across multiple evaluation metrics,while maintaining strong biological information retention as evidenced by improved clustering metrics and qualitative PCA visualizations. Moreover, our method effectively reduces the feature space from thousands of dimensions to just 256, addressing the curse of dimensionality while maintaining high performance. These advancements enable more robust integration of multi-source Cell Painting data, potentially accelerating progress in drug discovery and cellular biology research.", "keywords": [ "Cell Painting", "Batch Correction", "Representation Learning", "Transformer", "Hyena Operator", "High-dimensional Data", "Image-based Profiling" ], "rating_list": [ 5, 8, 5, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "uo6UsVkkEQ", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Di Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "PhD student" }, { "name": "Dongzhan Zhou", "gender": "Female", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Jiatong LI", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Jingdi Lei", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Qing Li", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Wei Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Wenqi Fan", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "Yunqing LIU", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Yuqiang Li", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 32, "n_ref": 57, "n_ref_all": 92, "n_fig": 7, "n_tab": 18, "L_tab": 4239, "n_element_tab": 562, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1474, "n_element_tab_1": 201, "formula_len_all": 723, "formula_len_all_1": 559, "len_all": 177980, "len_all_1": 67484, "len_abs": 2197, "len_title": 130, "len_sents": 53003, "len_sents_1": 31446, "n_sents": 411, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1823, "title": "MolReFlect: Towards Fine-grained In-Context Alignment between Molecules and Texts", "abs": "Molecule discovery is a pivotal research field, impacting everything from the medicines we take to the materials we use. Recently, Large Language Models (LLMs) have been widely adopted in molecule understanding and generation, yet the alignments between molecules and their corresponding captions remain a significant challenge. Previous endeavours often treat the molecule as a general SMILES string or molecular graph, neglecting the fine-grained alignments between the molecular sub-structures and the descriptive textual phrases, which are crucial for accurate and explainable predictions. In this case, we introduce MolReFlect, a novel teacher-student framework designed to contextually perform the molecule-caption alignments in a fine-grained way. Our approach initially leverages a larger teacher LLM to label the detailed alignments by directly extracting critical phrases from molecule captions or SMILES strings and implying them to corresponding sub-structures or characteristics. To refine these alignments, we propose In-Context Selective Reflection, which retrieves previous extraction results as context examples for teacher LLM to reflect and lets a smaller student LLM select from in-context reflection and previous extraction results. Finally, we enhance the learning process of the student LLM through Chain-of-Thought In-Context Molecule Tuning, integrating the fine-grained alignments and the reasoning processes within the Chain-of-Thought format. Our experimental results demonstrate that MolReFlect enables LLMs like Mistral-7B to significantly outperform the previous baselines, achieving SOTA performance on the ChEBI-20 dataset. This advancement not only enhances the generative capabilities of LLMs in the molecule-caption translation task, but also contributes to a more explainable framework.", "keywords": [ "Large Language Models", "In-Context Tuning", "Reflection Tuning", "Molecule Discovery", "Molecule-Text Alignment" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "unDQOUah0F", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Charles Ding", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "Undergrad student" }, { "name": "Dan Zhao", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Lecturer" }, { "name": "Justin Lin", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Undergrad student" }, { "name": "Kazuhito Koishida", "gender": "unknown", "institution": "Microsoft Corporation", "country": "US", "position": "Researcher" }, { "name": "Paul Pu Liang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Rogerio Bonatti", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Yinheng Li", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Lawrence Jang", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 81, "n_ref_all": 103, "n_fig": 5, "n_tab": 13, "L_tab": 5348, "n_element_tab": 569, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1594, "n_element_tab_1": 294, "formula_len_all": 1659, "formula_len_all_1": 0, "len_all": 228629, "len_all_1": 57881, "len_abs": 1524, "len_title": 139, "len_sents": 70118, "len_sents_1": 25783, "n_sents": 637, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1532, "title": "VideoWebArena: Evaluating Long Context Multimodal Agents with Video Understanding Web Tasks", "abs": "Videos are often used to learn or extract the necessary information to complete\ntasks in ways different than what text or static imagery can provide. However, many\nexisting agent benchmarks neglect long-context video understanding, instead focus-\ning on text or static image inputs. To bridge this gap, we introduce VideoWebArena\n(VideoWA), a benchmark for evaluating the capabilities of long-context multimodal\nagents for video understanding. VideoWA consists of 2,021 web agent tasks based\non manually crafted video tutorials, which total almost four hours of content. For\nour benchmark, we define a taxonomy of long-context video-based agent tasks with\ntwo main areas of focus: skill retention and factual retention. While skill retention\ntasks evaluate whether an agent can use a given human demonstration to complete\na task efficiently, the factual retention task evaluates whether an agent can retrieve\ninstruction-relevant information from a video to complete a task. We find that the\nbest model achieves a 13.3% success rate on factual retention tasks and 45.8% on\nfactual retention QA pairs\u2014far below human success rates of 73.9% and 79.3%,\nrespectively. On skill retention tasks, long-context models perform worse with\ntutorials than without, exhibiting a 5% performance decrease in WebArena tasks\nand a 10.3% decrease in VisualWebArena tasks. Our work highlights performance\ngaps in the agentic abilities of long-context multimodal models and provides as a\ntestbed for the future development of long-context video agents.", "keywords": [ "agents", "benchmark", "video understanding", "multimodal agents" ], "rating_list": [ 6, 6, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 3, 2 ] }, { "paper_id": "un9Gzm0BZb", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Kaining Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Liu Liu", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Min-Hsiu Hsieh", "gender": "unknown", "institution": "National Taiwan University", "country": "", "position": "Principal Researcher" } ], "n_formula": 30, "n_formula_1": 25, "n_ref_uni": 47, "n_ref": 79, "n_ref_all": 92, "n_fig": 5, "n_tab": 4, "L_tab": 946, "n_element_tab": 14, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 666, "n_element_tab_1": 14, "formula_len_all": 2261, "formula_len_all_1": 1798, "len_all": 114449, "len_all_1": 64079, "len_abs": 1426, "len_title": 120, "len_sents": 26954, "len_sents_1": 26132, "n_sents": 213, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1432, "title": "ER-AAE: A quantum state preparation approach based on entropy reduction", "abs": "Amplitude encoding of classical vectors serves as a cornerstone for numerous quantum machine learning algorithms in real-world applications. Nevertheless, achieving exact amplitude encoding for general vectors needs an exponential number of gates, which negates the potential quantum advantages. To address the challenge of large gate number in the state preparation phase, we propose an approximate amplitude encoding algorithm based on entropy reduction (ER-AAE) within the classical framework. Given a target vector, the ER-AAE algorithm generates a sequence of gates, comprising single-qubit rotations and CZ gates, that approximates the amplitude encoding of the target vector. The structure of encoding circuits in ER-AAE is built inductively using a greedy search strategy that maximally reduces the linear entropy. We further prove that the state produced by ER-AAE approximates to the target state with the infidelity bounded by the linear entropy of intermediate states. Experimental results, including state preparations on random quantum circuit states, random vectors, MNIST digits, and CIFAR-10 images, validate our method. Specifically, real-world data reveals a noteworthy trend where linear entropy decays significantly faster compared to random vectors. Furthermore, the ER-AAE algorithm surpasses the best existing encoding techniques, achieving lower error with an equivalent or fewer number of CNOT or CZ gates.", "keywords": [ "quantum machine learning", "amplitude encoding", "state preparation" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "umggmAFhRD", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Alexander Roman", "gender": "Male", "institution": "New College of Florida", "country": "", "position": "Assistant Professor" }, { "name": "Christos Ziakas", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "Researcher" }, { "name": "Elliott Thornley", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Leyton Ho", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Undergrad student" }, { "name": "Louis Alexander Thomson", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Undergrad student" } ], "n_formula": 31, "n_formula_1": 5, "n_ref_uni": 54, "n_ref": 98, "n_ref_all": 130, "n_fig": 15, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4093, "formula_len_all_1": 176, "len_all": 232746, "len_all_1": 66204, "len_abs": 867, "len_title": 49, "len_sents": 80657, "len_sents_1": 32079, "n_sents": 694, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 871, "title": "Towards shutdownable agents via stochastic choice", "abs": "Some worry that advanced artificial agents may resist being shut down. The Incomplete Preferences Proposal (IPP) is an idea for ensuring that doesn\u2019t happen. A key part of the IPP is using a novel \u2018 Discounted REward for Same-Length Trajectories (DREST)\u2019 reward function to train agents to (1) pursue goals effectively conditional on each trajectory-length (be \u2018USEFUL\u2019), and (2) choose stochastically between different trajectory-lengths (be \u2018NEUTRAL\u2019 about trajectory-lengths). In this paper, we propose evaluation metrics for USEFULNESS and NEUTRALITY. We use a DREST reward function to train simple agents to navigate gridworlds, and we find that these agents learn to be USEFUL and NEUTRAL. Our results thus suggest that DREST reward functions could also train advanced agents to be USEFUL and NEUTRAL, and thereby make these advanced agents useful and shutdownable.", "keywords": [ "the alignment problem", "the shutdown problem", "corrigibility", "reinforcement learning", "stochastic policy" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "ulXCYmvVg6", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Dong HUANG", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Guangtao Zeng", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "", "position": "PhD student" }, { "name": "Heming Cui", "gender": "Male", "institution": "the University of Hong Kong, University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Jianbo Dai", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "MS student" }, { "name": "Jie Zhang", "gender": "Female", "institution": "King's College London, University of London", "country": "", "position": "Assistant Professor" }, { "name": "Meng Luo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yuhao QING", "gender": "unknown", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zhijiang Guo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "han weng", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 1, "n_ref_uni": 36, "n_ref": 99, "n_ref_all": 130, "n_fig": 3, "n_tab": 16, "L_tab": 7393, "n_element_tab": 917, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 4657, "n_element_tab_1": 575, "formula_len_all": 192, "formula_len_all_1": 11, "len_all": 256489, "len_all_1": 76669, "len_abs": 1332, "len_title": 104, "len_sents": 56894, "len_sents_1": 32996, "n_sents": 415, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1340, "title": "Effi-Code: Unleashing Code Efficiency in Language Models", "abs": "As the use of large language models (LLMs) for code generation becomes more prevalent in software development, it is critical to enhance both the efficiency and correctness of the generated code. \nExisting methods and models primarily focus on the correctness of LLM-generated code, ignoring the efficiency. \nIn this work, we present Effi-Code, an approach to enhancing code generation in LLMs that can improve both efficiency and correctness. We introduce a Self-Optimization process based on Overhead Profiling that leverages open-source LLMs to generate a high-quality dataset of correct and efficient code samples. This dataset is then used to fine-tune various LLMs. Our method involves iterative refinement of generated code, guided by runtime performance metrics and correctness checks. Extensive experiments demonstrate that models fine-tuned on the Effi-Code significantly improve code correctness and efficiency across task types. \nFor example, the pass@1 of DeepSeek-Coder-6.7B-Instruct generated cod increases from **43.3\\%** to **76.8\\%**, and the average execution time for the same correct tasks decreases by **30.5\\%**.\nEffi-Code offers a scalable and generalizable approach to improving code generation in AI systems, with potential applications in software development, algorithm design, and computational problem-solving.", "keywords": [ "Large Langugae Models", "Code Generation", "Program Synthesis", "Efficient Method", "Alignment" ], "rating_list": [ 1, 6, 3, 6 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 3, 1, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "ulJNq6FQrw", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Biao Liu", "gender": "Male", "institution": "Southeast University", "country": "", "position": "PhD student" }, { "name": "Ning Xu", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Xin Geng", "gender": "Male", "institution": "Southeast University, China", "country": "CN", "position": "Professor" } ], "n_formula": 16, "n_formula_1": 17, "n_ref_uni": 24, "n_ref": 61, "n_ref_all": 80, "n_fig": 5, "n_tab": 5, "L_tab": 1248, "n_element_tab": 221, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 985, "n_element_tab_1": 115, "formula_len_all": 1263, "formula_len_all_1": 1109, "len_all": 141623, "len_all_1": 67014, "len_abs": 1497, "len_title": 114, "len_sents": 35638, "len_sents_1": 29701, "n_sents": 275, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1533, "title": "Progressively Label Enhancement for Large Language Model Alignment", "abs": "Large Language Models (LLM) alignment aims to prevent models from producing content that misaligns with human expectations, which can lead to ethical and legal concerns. \n In the last few years, Reinforcement Learning from Human Feedback (RLHF) has been the most prominent method for achieving alignment.\n Due to challenges in stability and scalability with RLHF stages, which arise from the complex interactions between multiple models, researchers are exploring alternative methods to achieve effects comparable to those of RLHF.\n However, these methods often rely on large high-quality datasets.\n Despite some methods considering the generation of additional data to expand datasets, they often treat model training and data generation as separate and static processes, overlooking the fact that these processes are highly interdependent, leading to inefficient utilization of the generated data.\n To deal with this problem, we propose PLE, i.e., Progressively Label Enhancement for LLM Alignment, a framework that dynamically adjusts the model\u2019s training process based on the evolving quality of the generated data.\n Specifically, we prompt the model to generate responses for both the original query and the query guided by a set of carefully designed principles, and then utilize a dynamic threshold to determine the appropriate training approach for both responses based on their corresponding reward scores. \n Experimental results demonstrate the effectiveness of PLE compared to existing LLM alignment methods.", "keywords": [ "Large Language Model", "LLM Alignment" ], "rating_list": [ 3, 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 3, 4 ] }, { "paper_id": "ulIW7Frjpn", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Fengxi Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Guo Lu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Li Song", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Qi Wang", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Qunshan Gu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yan Zhao", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhengxue Cheng", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 36, "n_ref": 54, "n_ref_all": 73, "n_fig": 12, "n_tab": 9, "L_tab": 3280, "n_element_tab": 312, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1939, "n_element_tab_1": 206, "formula_len_all": 456, "formula_len_all_1": 456, "len_all": 125700, "len_all_1": 51404, "len_abs": 1433, "len_title": 119, "len_sents": 30860, "len_sents_1": 21406, "n_sents": 237, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1430, "title": "Large Language Models Are Stronger Entropy Models for Transform Coding", "abs": "Large language models (LLMs) have shown promising advancements in lossless compression due to their excellent next-token prediction capabilities. However, there is a gap between LLM-based compressors and classical transform-based codecs. Existing LLM-based compressors function solely as entropy coders, focusing on compressing redundant data in the raw domain. In contrast, classical codecs typically transform raw data into more compact features in the latent domain before applying entropy coding. But LLM-based compressors have not discussed this case. To our knowledge, this is the first work to introduce an LLM-based entropy model for transform coding. Specifically, we propose a simple yet effective fine-tuning strategy, tested across various codecs for both images and speeches. With less than 2% parameters are fine-tuned, the LLMs can serve as highly effective entropy models for well-established transform-based compression codecs. For instance, LLaMA3-8B paired with arithmetic coding compresses latent image codes on Kodak to 4.62% and speech codes on LibriTTS to 42.53% of their transformed sizes after fine-tuning. Our proposed methods achieve notable BD-rate improvements of 54.07% over JPEG, 17.61% over VQGAN, and 34.61% over SpeechTokenizer. These findings highlight the great potential of integrating LLMs into codecs to significantly improve coding efficiency. Source codes will be released upon acceptance.", "keywords": [ "Transform Coding", "Multimodal Data Compression", "Entropy Model", "Large Language Models" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "ulGwcj1egv", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Akriti Jain", "gender": "unknown", "institution": "Adobe Research", "country": "IN", "position": "Researcher" }, { "name": "Koyel Mukherjee", "gender": "Female", "institution": "Adobe Research", "country": "", "position": "Researcher" }, { "name": "Saransh Sharma", "gender": "Male", "institution": "Indian Institute of Technology Kharagpur", "country": "IN", "position": "Undergrad student" }, { "name": "Soumyabrata Pal", "gender": "Male", "institution": "Adobe Systems", "country": "IN", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 30, "n_ref": 47, "n_ref_all": 59, "n_fig": 1, "n_tab": 6, "L_tab": 2514, "n_element_tab": 471, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 113, "n_element_tab_1": 3, "formula_len_all": 518, "formula_len_all_1": 485, "len_all": 124842, "len_all_1": 61609, "len_abs": 1878, "len_title": 134, "len_sents": 39899, "len_sents_1": 28628, "n_sents": 281, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1932, "title": "FiRST: Finetuning Router-Selective Transformers for Input-Adaptive Latency Reduction", "abs": "Auto-regressive Large Language Models (LLMs) demonstrate remarkable performance across domanins such as vision and language processing. However, due to sequential processing through a stack of transformer layers, autoregressive decoding faces significant computation/latency challenges, particularly in resource-constrained environments like mobile and edge devices. Existing approaches in literature that aim to improve latency via skipping layers have two distinct flavors - 1) Early exit 2) Input-agnostic heuristics where tokens exit at pre-determined layers irrespective of input sequence. Both the above strategies have limitations - the former cannot be applied to handle KV Caching necessary for speed-ups in modern framework and the latter does not capture the variation in layer importance across tasks or more generally, across input sequences. To address both limitations, we propose \\textsc{FiRST}, an algorithm that reduces inference latency by using layer-specific routers to select a subset of transformer layers adaptively for each input sequence - the prompt (during prefill stage) decides which layers will be skipped during decoding. \\textsc{FiRST} preserves compatibility with KV caching enabling faster inference while being quality-aware. \\textsc{FiRST} is model-agnostic and can be easily enabled on any pre-trained LLM. We further improve performance by incorporating LoRA adapters for fine-tuning on external datasets, enhancing task-specific accuracy while maintaining latency benefits. Our approach reveals that input adaptivity is critical - indeed, different task-specific middle layers play a crucial role in evolving hidden representations depending on task. Extensive experiments show that \\textsc{FiRST} significantly reduces latency while retaining competitive performance (as compared to baselines), making our approach an efficient solution for LLM deployment in low-resource environments.", "keywords": [ "Input-Adaptive Layer Selection; Resource-Constrained Environments; Latency Reduction; Finetuning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "ulCAPXYXfa", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bo Zheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Instructor" }, { "name": "Jitai Hao", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "MS student" }, { "name": "Jun Yu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Tian Wang", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xin Xin", "gender": "Male", "institution": "University of Glasgow", "country": "", "position": "PhD student" }, { "name": "Yuke Zhu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zhaochun Ren", "gender": "Male", "institution": "Leiden University", "country": "NL", "position": "Associate Professor" }, { "name": "Sheng Guo", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 30, "n_ref": 70, "n_ref_all": 107, "n_fig": 13, "n_tab": 11, "L_tab": 5479, "n_element_tab": 906, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1457, "n_element_tab_1": 258, "formula_len_all": 1117, "formula_len_all_1": 638, "len_all": 181530, "len_all_1": 73099, "len_abs": 1307, "len_title": 114, "len_sents": 48500, "len_sents_1": 31771, "n_sents": 432, "n_sents_1": 276, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1308, "title": "OmniKV: Dynamic Context Selection for Efficient Long-Context LLMs", "abs": "During the inference phase of Large Language Models (LLMs) with long context, a substantial portion of GPU memory is allocated to the KV cache, with memory usage increasing as the sequence length grows. To mitigate the GPU memory footprint associate with KV cache, some previous studies have discarded less important tokens based on the sparsity identified in attention scores in long context scenarios. However, we argue that attention scores cannot indicate the future importance of tokens in subsequent generation iterations, because attention scores are calculated based on current hidden states. Therefore, we propose OmniKV, a token-dropping-free and training-free inference method, which achieves a 1.68x speedup without any loss in performance. It is well-suited for offloading, significantly reducing KV cache memory usage by up to 75% with it. The core innovative insight of OmniKV is: Within a single generation iteration, there is a high degree of similarity in the important tokens identified across consecutive layers. Extensive experiments demonstrate that OmniKV achieves state-of-the-art performance across multiple benchmarks, with particularly advantages in chain-of-thoughts scenarios. OmniKV extends the maximum context length supported by a single A100 for Llama-3-8B from 128K to 450K.", "keywords": [ "Efficient LLMs", "KV cache", "Long Context LLMs" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 2 ] }, { "paper_id": "ukmWcHpa3H", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Jirui Wu", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Liming Wang", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "", "position": "Associate Professor" }, { "name": "Long Tian", "gender": "Male", "institution": "Xi'an University of Software Engineering Institute", "country": "CN", "position": "Assistant Professor" }, { "name": "Quan Wang", "gender": "Male", "institution": "4th Military Medical University", "country": "CN", "position": "Researcher" }, { "name": "Ruiying Lu", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaonan Liu", "gender": "Male", "institution": "4th Military Medical University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiyang Liu", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Yufei Li", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "Postdoc" } ], "n_formula": 21, "n_formula_1": 16, "n_ref_uni": 23, "n_ref": 34, "n_ref_all": 55, "n_fig": 6, "n_tab": 13, "L_tab": 2842, "n_element_tab": 354, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 2067, "n_element_tab_1": 221, "formula_len_all": 1773, "formula_len_all_1": 1492, "len_all": 122311, "len_all_1": 67759, "len_abs": 1591, "len_title": 124, "len_sents": 34372, "len_sents_1": 28064, "n_sents": 254, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1608, "title": "Meta-weighted Diffusion Model for Reliable Online Surgical Phase Recognition", "abs": "Surgical phase recognition has drawn great attention most recently thanks to its potential downstream applications closely related to human life and health. Despite deep network-based models have made significant advancement in capturing discriminative long-term dependency of surgical videos to achieve improved recognition, they seldom account for exploring and modeling uncertainty of surgical videos, which should be crucial for reliable surgical phase recognition. we categorize the sources of uncertainty into two types, imbalanced phase distribution and low-quality image acquisition, which are inevitable in surgical videos. To address this pivot issue, we introduce a meta-weighted diffusion model (MetaDiff) to take full advantages of meta-learning and deep generative model in tackling uncertainty. For uncertainty caused by image quality, we present a classifier-guided diffusion model to produce countable denoised recognition results, making it possible to measure uncertainty using statistical tools for each video frame. For uncertainty caused by phase distribution, we propose a meta-weighted objective function to optimize the classifier-guided diffusion model, making the classification boundary robust against surgical video uncertainty.\nWe demonstrate outstanding ability of our model through comprehensive benchmarks on Cholec80, AutoLaparo, M2Cai16, and CATARACTS. Experimental results reveal that MetaDiff significantly outperforms state-of-the-art methods, separately achieving accuracies of $95.3\\%$, $85.8\\%$, $92.2\\%$, and $85.1\\%$ on Cholec80, AutoLaparo, M2Cai16, and CATARACTS.", "keywords": [ "surgical phase recognition", "diffusion model", "meta learning" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 5, 2 ] }, { "paper_id": "ujpAYpFDEA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Aiwei Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Leyi Pan", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Liancheng Fang", "gender": "Male", "institution": "University of Illinois at Chicago", "country": "US", "position": "PhD student" }, { "name": "Lijie Wen", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Philip S. Yu", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "US", "position": "Full Professor" }, { "name": "Sheng Guan", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Xuming Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Yifei Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "Yiming Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 39, "n_formula_1": 16, "n_ref_uni": 18, "n_ref": 52, "n_ref_all": 96, "n_fig": 5, "n_tab": 7, "L_tab": 4160, "n_element_tab": 500, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 2208, "n_element_tab_1": 234, "formula_len_all": 3109, "formula_len_all_1": 822, "len_all": 166204, "len_all_1": 58164, "len_abs": 1549, "len_title": 112, "len_sents": 54443, "len_sents_1": 25116, "n_sents": 434, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 64, "L_abs": 1487, "title": "Can Watermarked LLMs be Identified by Users via Crafted Prompts?", "abs": "Text watermarking for Large Language Models (LLMs) has made significant progress in detecting LLM outputs and preventing misuse. Current watermarking techniques offer high detectability, minimal impact on text quality, and robustness to text editing. \n However, current researches lack investigation into the imperceptibility of watermarking techniques in LLM services.\n This is crucial as LLM providers may not want to disclose the presence of watermarks in real-world scenarios, as it could reduce user willingness to use the service and make watermarks more vulnerable to attacks. This work is the first to investigate the imperceptibility of watermarked LLMs. We design an identification algorithm called Water-Probe that detects watermarks through well-designed prompts to the LLM. Our key motivation is that current watermarked LLMs expose consistent biases under the same watermark key, resulting in similar differences across prompts under different watermark keys. Experiments show that almost all mainstream watermarking algorithms are easily identified with our well-designed prompts, \n while Water-Probe demonstrates a minimal false positive rate for non-watermarked LLMs. \n Finally, we propose that the key to enhancing the imperceptibility of watermarked LLMs is to increase the randomness of watermark key selection. Based on this, we introduce the Water-Bag strategy, which significantly improves watermark imperceptibility by merging multiple watermark keys.", "keywords": [ "Large Language Models", "Watermark", "Identification" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "ujNe7sybJu", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Kuan-Chen Mu", "gender": "Female", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "Zhi-Yi Chin", "gender": "Female", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Researcher" }, { "name": "Wei-Chen Chiu", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 38, "n_ref": 66, "n_ref_all": 84, "n_fig": 4, "n_tab": 3, "L_tab": 308, "n_element_tab": 46, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 474, "n_element_tab_1": 55, "formula_len_all": 39, "formula_len_all_1": 1, "len_all": 118877, "len_all_1": 50081, "len_abs": 1557, "len_title": 132, "len_sents": 30644, "len_sents_1": 23856, "n_sents": 222, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1562, "title": "Realizing Video Summarization from the Path of Language-based Semantic Understanding", "abs": "The recent development of Video-based Large Language Models (VideoLLMs), has significantly advanced video summarization by aligning video features\u2014and, in some cases, audio features\u2014with Large Language Models (LLMs). Each of these VideoLLMs possesses unique strengths and weaknesses. Many recent methods have required extensive fine-tuning to overcome the limitations of these models, which can be resource-intensive. In this work, we observe that the strengths of one VideoLLM can complement the weaknesses of another. Leveraging this insight, we propose a novel video summarization framework inspired by the Mixture of Experts (MoE) paradigm, which operates as an inference-time algorithm without requiring any form of fine-tuning. Our approach integrates multiple VideoLLMs to generate comprehensive and coherent textual summaries. It effectively combines visual and audio content, provides detailed background descriptions, and excels at identifying keyframes, which enables more semantically meaningful retrieval compared to traditional computer vision approaches that rely solely on visual information, all without the need for additional fine-tuning. Moreover, the resulting summaries enhance performance in downstream tasks such as summary video generation, either through keyframe selection or in combination with text-to-image models. Our language-driven approach offers a semantically rich alternative to conventional methods and provides flexibility to incorporate newer VideoLLMs, enhancing adaptability and performance in video summarization tasks.", "keywords": [ "Visual Language Model", "Large Language Model", "Video Summarization", "Video Understanding", "VideoLLM" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "uiyljVIP0k", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Xiang Li", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "PhD student" }, { "name": "Yanping Zheng", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Ye Yuan", "gender": "unknown", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zhewei Wei", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 36, "n_ref": 57, "n_ref_all": 70, "n_fig": 7, "n_tab": 4, "L_tab": 3334, "n_element_tab": 253, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 2646, "n_element_tab_1": 133, "formula_len_all": 587, "formula_len_all_1": 588, "len_all": 115797, "len_all_1": 58286, "len_abs": 1494, "len_title": 136, "len_sents": 31736, "len_sents_1": 25611, "n_sents": 256, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1416, "title": "UNIFYING LONG AND SHORT SPATIO-TEMPORAL FORECASTING WITH SPECTRAL GRAPH NEURAL NETWORKS", "abs": "Multivariate Time Series (MTS) forecasting plays a vital role in various practical applications. Current research in this area is categorized into Spatial-Temporal Forecasting (STF) and Long-term Time Series Forecasting (LTSF). While these tasks share similarities, the methods and benchmarks used differ significantly. Spatio-Temporal Graph Neural Networks (STGNNs) excel at modeling interrelationships in STF tasks but face difficulties with long sequence inputs due to inefficient training. In contrast, LTSF models handle long sequences well but struggle with capturing complex variable interrelationships. This paper proposes the Spectral Spatio-Temporal Graph Neural Network (S2GNN) to address these challenges, unifying short- and long-sequence spatio-temporal forecasting within a single framework. S2GNN leverages spectral GNNs for global feature extraction incorporates an adaptive graph structure to manage varying sequence lengths and adopts a decoupled framework to improve scalability. Additionally, we introduce scale-adaptive node embeddings and cross-correlation embeddings for better differentiation between similar temporal patterns. Extensive experiments on eight public datasets, including both STF and LTSF datasets, demonstrate that S2GNN consistently outperforms state-of-the-art models across diverse prediction tasks. Code is available at \\url{https://anonymous.4open.science/r/S2GNN-B21D}.", "keywords": [ "multivariate time series forecasting", "spatio-temporal graph neural network", "spectral graph neural network" ], "rating_list": [ 6, 5, 3, 8, 5 ], "soundness_list": [ 3, 2, 1, 3, 2 ], "presentation_list": [ 2, 1, 1, 3, 3 ], "contribution_list": [ 2, 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 5, 5, 4 ] }, { "paper_id": "uinsufj5TR", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Ananya Joshi", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Celia Cintas", "gender": "Female", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Skyler Speakman", "gender": "Male", "institution": "International Business Machines", "country": "KE", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 32, "n_ref": 50, "n_ref_all": 88, "n_fig": 15, "n_tab": 5, "L_tab": 1175, "n_element_tab": 48, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 71, "n_element_tab_1": 5, "formula_len_all": 233, "formula_len_all_1": 233, "len_all": 116492, "len_all_1": 54661, "len_abs": 3409, "len_title": 73, "len_sents": 34256, "len_sents_1": 26492, "n_sents": 224, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1164, "title": "Enabling Sparse Autoencoders for Topic Alignment in Large Language Models", "abs": "Recent work shows that Sparse Autoencoders (SAE) applied to LLM layers have neurons corresponding to interpretable concepts. Consequently, these SAE neurons can be modified to align generated outputs, but only towards pre-identified topics and with some parameter tuning. Our approach leverages the interpretability properties of SAEs to enable alignment for any topic. This method 1) scores each SAE neuron by its semantic similarity to an alignment text and uses them to 2) modify SAE-layer-level outputs by emphasizing topic-aligned neurons. We assess the alignment capabilities of this approach on diverse public topics datasets, including Amazon reviews, Medicine, and Sycophancy, across open-source LLMs, GPT2, and Gemma with multiple SAEs configurations. Experiments aligning to medical prompts reveal several benefits over fine-tuning, including increased average language acceptability (0.25 vs 0.5), reduced training time across multiple alignment topics (333.6s vs. 62s), and acceptable inference time for many applications (+0.00092s/token). Our anonymized open-source code is available at https://anonymous.4open.science/r/sae-steering-8513/README.md.", "keywords": [ "Alignment", "SAEs", "Mechanistic Interpretability", "Large Language Models" ], "rating_list": [ 1, 3, 8, 6, 6 ], "soundness_list": [ 1, 2, 3, 2, 3 ], "presentation_list": [ 1, 3, 4, 2, 2 ], "contribution_list": [ 1, 2, 3, 3, 2 ], "confidence_list": [ 5, 3, 3, 2, 3 ] }, { "paper_id": "uiBLOcyTIA", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Gao Cong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Ning Cao", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "SHUAI LIU", "gender": "Male", "institution": "School of Computer Science and Engineering, Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "YUE JIANG", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Yile Chen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" } ], "n_formula": 21, "n_formula_1": 12, "n_ref_uni": 29, "n_ref": 64, "n_ref_all": 86, "n_fig": 5, "n_tab": 14, "L_tab": 3250, "n_element_tab": 258, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 780, "n_element_tab_1": 95, "formula_len_all": 1059, "formula_len_all_1": 808, "len_all": 177104, "len_all_1": 62465, "len_abs": 1449, "len_title": 47, "len_sents": 66069, "len_sents_1": 28946, "n_sents": 484, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 47, "L_abs": 1457, "title": "NEXTLOCLLM: NEXT LOCATION PREDICTION USING LLMS", "abs": "Next location prediction is a critical task in human mobility analysis and serves as a foundation for various downstream applications. Existing methods typically rely on discrete IDs to represent locations, which inherently overlook spatial relationships and cannot generalize across cities. In this paper, we propose NextLocLLM, which leverages the advantages of large language models (LLMs) in processing natural language descriptions and their strong generalization capabilities for next location prediction. Specifically, instead of using IDs, NextLocLLM encodes locations based on continuous spatial coordinates to better model spatial relationships. These coordinates are further normalized to enable robust cross-city generalization. Another highlight of NextlocLLM is its LLM-enhanced POI embeddings. It utilizes LLMs\u2019 ability to encode each POI category\u2019s natural language description into embeddings. These embeddings are then integrated via nonlinear projections to form this LLM-enhanced POI embeddings, effectively capturing locations\u2019 functional attributes. Furthermore, task and data prompt prefix, together with trajectory embeddings, are incorporated as input for partly-frozen LLM backbone. NextLocLLM further introduces prediction retrieval module to ensure structural consistency in prediction. Experiments show that NextLocLLM outperforms existing models in next location prediction, excelling in both supervised and zero-shot settings.", "keywords": [ "next location prediction", "large language model", "zero-shot" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "uhaLuZcCjH", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Ashish Hooda", "gender": "unknown", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Divyam Anshumaan", "gender": "unknown", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Somesh Jha", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin, Madison", "country": "US", "position": "Full Professor" }, { "name": "Yudong Chen", "gender": "Male", "institution": "Department of Computer Sciences, University of Wisconsin - Madison", "country": "US", "position": "Associate Professor" }, { "name": "Zi Wang", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 8, "n_ref_uni": 21, "n_ref": 39, "n_ref_all": 68, "n_fig": 11, "n_tab": 4, "L_tab": 146, "n_element_tab": 26, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 146, "n_element_tab_1": 26, "formula_len_all": 553, "formula_len_all_1": 283, "len_all": 132767, "len_all_1": 71437, "len_abs": 914, "len_title": 152, "len_sents": 47135, "len_sents_1": 33857, "n_sents": 385, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 840, "title": "Functional Homotopy: Smoothing Discrete Optimization via Continuous Parameters for LLM Jailbreak Attacks", "abs": "Optimization methods are widely employed in deep learning to address and mitigate undesired model responses. While gradient-based techniques have proven effective for image models, their application to language models is hindered by the discrete nature of the input space. This study introduces a novel optimization approach, termed the *functional homotopy* method, which leverages the functional duality between model training and input generation. By constructing a series of easy-to-hard optimization problems, we iteratively solve these using principles derived from established homotopy methods. We apply this approach to jailbreak attack synthesis for large language models (LLMs), achieving a 20%-30% improvement in success rate over existing methods in circumventing established safe open-source models such as Llama-2 and Llama-3.", "keywords": [ "alignment", "optimization" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 2, 2 ] }, { "paper_id": "ugyqNEOjoU", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Fahimeh Moafian", "gender": "Female", "institution": "Technische Universit\u00e4t Dresden", "country": "DE", "position": "Postdoc" }, { "name": "Jonas Belouadi", "gender": "Male", "institution": "NICT, Kyoto", "country": "DE", "position": "Intern" }, { "name": "Leixin Zhang", "gender": "Female", "institution": "University of Twente", "country": "", "position": "PhD student" }, { "name": "Steffen Eger", "gender": "Male", "institution": "University of Technology Nuremberg", "country": "DE", "position": "Full Professor" }, { "name": "WEIHE ZHAI", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yinjie Cheng", "gender": "Male", "institution": "University of Sheffield", "country": "GB", "position": "PhD student" }, { "name": "ZHIXUE ZHAO", "gender": "Female", "institution": "University of Sheffield, University of Sheffield", "country": "GB", "position": "Lecturer" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 42, "n_ref": 60, "n_ref_all": 86, "n_fig": 5, "n_tab": 14, "L_tab": 4737, "n_element_tab": 446, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2979, "n_element_tab_1": 233, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 174906, "len_all_1": 71667, "len_abs": 4900, "len_title": 94, "len_sents": 46329, "len_sents_1": 32906, "n_sents": 380, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1342, "title": "ScImage: How good are multimodal large language models at scientific text-to-image generation?", "abs": "Multimodal large language models (LLMs) have demonstrated impressive capabilities in generating high-quality images from textual instructions. However, their performance in generating scientific images\u2014a critical application for accelerating scientific progress\u2014remains underexplored. In this work, we address this gap by introducing ScImage, a benchmark designed to evaluate the multimodal capabilities of LLMs in generating scientific images from textual descriptions. ScImage assesses three key dimensions of understanding: spatial, numeric, and attribute comprehension, as well as their combinations, focusing on the relationships between scientific objects (e.g., squares, circles). We evaluate five LLMs\u2014GPT4-o, Llama, AutomaTikZ, Dall-E, and StableDiffusion\u2014using two modes of output generation: code-based outputs (Python, TikZ) and direct raster image generation. Additionally, we examine four different input languages: English, German, Farsi, and Chinese. Our evaluation, conducted with 11 scientists across three criteria (correctness, relevance, and scientific accuracy), reveals that while GPT4-o produces outputs of decent quality for simpler prompts involving individual dimensions such as spatial, numeric, or attribute understanding in isolation, all models face challenges in this task, especially for more complex prompts.", "keywords": [ "LLMs", "multimodality", "science", "image generation" ], "rating_list": [ 8, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "ugXGFCS6HK", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Alex H Williams", "gender": "Male", "institution": "New York University", "country": "", "position": "Assistant Professor" }, { "name": "David Lipshutz", "gender": "Male", "institution": "Flatiron Institute", "country": "", "position": "Associate Research Scientist" }, { "name": "Jenelle Feather", "gender": "unknown", "institution": "Flatiron Institute", "country": "US", "position": "Postdoc" }, { "name": "Sarah E Harvey", "gender": "unknown", "institution": "Flatiron Institute", "country": "", "position": "Postdoc" }, { "name": "Eero Peter Simoncelli", "gender": "Male", "institution": "New York University", "country": "US", "position": "Full Professor" } ], "n_formula": 30, "n_formula_1": 15, "n_ref_uni": 41, "n_ref": 73, "n_ref_all": 106, "n_fig": 22, "n_tab": 2, "L_tab": 233, "n_element_tab": 47, "n_fig_1": 20, "n_tab_1": 2, "L_tab_1": 217, "n_element_tab_1": 42, "formula_len_all": 2611, "formula_len_all_1": 870, "len_all": 144754, "len_all_1": 64151, "len_abs": 3586, "len_title": 111, "len_sents": 48663, "len_sents_1": 32095, "n_sents": 319, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1386, "title": "Discriminating image representations with principal distortions", "abs": "Image representations (artificial or biological) are often compared in terms of their global geometry; however, representations with similar global structure can have strikingly different local geometries. Here, we propose a framework for comparing a set of image representations in terms of their local geometries. We quantify the local geometry of a representation using the Fisher information matrix, a standard statistical tool for characterizing the sensitivity to local stimulus distortions, and use this as a substrate for a metric on the local geometry in the vicinity of a base image. This metric may then be used to optimally differentiate a set of models, by finding a pair of \"principal distortions\" that maximize the variance of the models under this metric. We use this framework to compare a set of simple models of the early visual system, identifying a novel set of image distortions that allow immediate comparison of the models by visual inspection. In a second example, we apply our method to a set of deep neural network models and reveal differences in the local geometry that arise due to architecture and training types. These examples highlight how our framework can be used to probe for informative differences in local sensitivities between complex computational models, and suggest how it could be used to compare model representations with human perception.", "keywords": [ "representational similarity metric; Fisher information; information geometry; perception" ], "rating_list": [ 6, 8, 6, 5, 6 ], "soundness_list": [ 4, 4, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 4, 3 ] }, { "paper_id": "ufi0WPTgWp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Changli Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Chao Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Associate Professor" }, { "name": "Guangzhi Sun", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Junior Research Fellow" }, { "name": "Jimin Zhuang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Wei Li", "gender": "Male", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "Yixuan Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yudong Yang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Zejun MA", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 34, "n_ref": 80, "n_ref_all": 113, "n_fig": 12, "n_tab": 7, "L_tab": 2066, "n_element_tab": 260, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 3494, "n_element_tab_1": 239, "formula_len_all": 779, "formula_len_all_1": 596, "len_all": 195280, "len_all_1": 64588, "len_abs": 1720, "len_title": 109, "len_sents": 66120, "len_sents_1": 29603, "n_sents": 502, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1730, "title": "Enhancing Multimodal LLM for Detailed and Accurate Video Captioning using Multi-Round Preference Optimization", "abs": "Videos contain a wealth of information, and generating detailed and accurate descriptions in natural language is a key aspect of video understanding. In this paper, we present video-SALMONN 2, an advanced audio-visual large language model (LLM) with low-rank adaptation (LoRA) designed for enhanced video (with paired audio) captioning through directed preference optimization (DPO). We propose new metrics to evaluate the completeness and accuracy of video descriptions, which are optimized using DPO. To further improve training, we introduce a novel multi-round DPO (mrDPO) approach, which involves periodically updating the DPO reference model, merging and re-initializing the LoRA module as a proxy for parameter updates after each training round (1,000 steps), and incorporating guidance from ground-truth video captions to stabilize the process. To address potential catastrophic forgetting of non-captioning abilities due to mrDPO, we propose rebirth tuning, which finetunes the pre-DPO LLM by using the captions generated by the mrDPO-trained model as supervised labels. Experiments show that mrDPO significantly enhances video-SALMONN 2's captioning accuracy, reducing global and local error rates by 40\\% and 20\\%, respectively, while decreasing the repetition rate by 35\\%. The final video-SALMONN 2 model, with just 7 billion parameters, surpasses leading models such as GPT-4o and Gemini-1.5-Pro in video captioning tasks, while maintaining competitive performance to the state-of-the-art on widely used video question-answering benchmark among models of similar size. Upon acceptance, we will release the code, model checkpoints, and training and test data. Demos are available at https://video-salmonn-2.github.io.", "keywords": [ "Multi-modal large language models", "video captioning", "multi-round DPO", "rebirth tuning" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "ufhrQG5xie", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Chang Chih Meng", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "PhD student" }, { "name": "I-Chen Wu", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" }, { "name": "Kuanyen Liu", "gender": "unknown", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Undergrad student" }, { "name": "Renjyun Huang", "gender": "unknown", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Undergrad student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 47, "n_ref": 70, "n_ref_all": 89, "n_fig": 3, "n_tab": 5, "L_tab": 1656, "n_element_tab": 213, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 975, "n_element_tab_1": 94, "formula_len_all": 535, "formula_len_all_1": 535, "len_all": 123769, "len_all_1": 54817, "len_abs": 1404, "len_title": 100, "len_sents": 33363, "len_sents_1": 25742, "n_sents": 236, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1418, "title": "POIL: Preference Optimization for Imitation Learning", "abs": "Imitation learning (IL) enables agents to learn policies by mimicking expert demonstrations. \nWhile online IL methods require interaction with the environment, which is costly, risky, or impractical, offline IL allows agents to learn solely from expert datasets without any interaction with the environment.\nIn this paper, we propose Preference Optimization for Imitation Learning (POIL), a novel approach inspired by preference optimization techniques in large language model alignment. \nPOIL eliminates the need for adversarial training and reference models by directly comparing the agent's actions to expert actions using a preference-based loss function. \nWe evaluate POIL on MuJoCo control tasks under two challenging settings: learning from a single expert demonstration and training with different dataset sizes (100\\%, 10\\%, 5\\%, and 2\\%) from the D4RL benchmark.\nOur experiments show that POIL consistently delivers superior or competitive performance against state-of-the-art methods in the past, including Behavioral Cloning (BC), IQ-Learn, DMIL, and O-DICE, especially in data-scarce scenarios, such as using one expert trajectory or as little as 2\\% of the full expert dataset. \nThese results demonstrate that POIL enhances data efficiency and stability in offline imitation learning, making it a promising solution for applications where environment interaction is infeasible and expert data is limited.", "keywords": [ "Offline Imitation Learning", "Preference-based Reinforcement Learning", "Large Language Model Alignment", "Data Efficiency" ], "rating_list": [ 8, 8, 3, 6, 6 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 4, 3, 3, 3, 2 ], "contribution_list": [ 4, 3, 1, 3, 2 ], "confidence_list": [ 3, 4, 5, 3, 3 ] }, { "paper_id": "uffmkDtlR2", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Elena Baralis", "gender": "Female", "institution": "Politecnico di Torino", "country": "IT", "position": "Deputy Rector" }, { "name": "Eleonora Poeta", "gender": "Female", "institution": "Polytechnic Institute of Turin", "country": "", "position": "PhD student" }, { "name": "Eliana Pastor", "gender": "Female", "institution": "Politecnico di Torino", "country": "IT", "position": "Assistant Professor" }, { "name": "Gabriele Ciravegna", "gender": "Male", "institution": "Polytechnic Institute of Turin", "country": "IT", "position": "Postdoc" }, { "name": "Tania Cerquitelli", "gender": "Female", "institution": "Polytechnic Institute of Turin", "country": "IT", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 62, "n_ref": 90, "n_ref_all": 109, "n_fig": 8, "n_tab": 2, "L_tab": 1400, "n_element_tab": 87, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1400, "n_element_tab_1": 87, "formula_len_all": 192, "formula_len_all_1": 161, "len_all": 149017, "len_all_1": 61162, "len_abs": 1143, "len_title": 97, "len_sents": 37702, "len_sents_1": 29624, "n_sents": 275, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 48, "L_abs": 1150, "title": "MIMOSA: Multimodal Concept-based representations", "abs": "In recent years, deep learning-based architectures have significantly improved multimodal representation. However, interpretability remains challenging with traditional attention and gradient-based methods, offering limited insights into decision-making processes. Concept-based explainability provides intrinsic model interpretability by mapping raw data to higher-level abstractions, yet it has only been applied to unimodal data. \nWe present MIMOSA (MultIMOdal concept-based repreSentAtion), a unified multimodal model that integrates concept-based interpretability. Our research shows that exploiting a joint multimodal conceptual representation achieves comparable accuracy with multimodal black-box models, surpassing approaches based on unimodal concepts. This unified representation also prevents misclassification of concepts between modalities and improves concept interventions. Through a concept decoder, MIMOSA can extract concept visualizations for each modality. \nExperimental results obtained from three distinct multimodal datasets substantiate the efficacy of our approach, showcasing enhanced interpretability in multimodal models.", "keywords": [ "Concept-based model", "Multimodal", "Explainability" ], "rating_list": [ 3, 1, 1, 3, 5 ], "soundness_list": [ 3, 1, 2, 2, 3 ], "presentation_list": [ 3, 1, 1, 2, 2 ], "contribution_list": [ 3, 1, 1, 2, 3 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "ueeqGvQozB", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Hanzhang Qin", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Assistant Professor" }, { "name": "Hua Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Huigen Ye", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yaoyang Cheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Intern" }, { "name": "Zhiguang Cao", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 6, "n_ref_uni": 83, "n_ref": 161, "n_ref_all": 198, "n_fig": 6, "n_tab": 22, "L_tab": 22341, "n_element_tab": 3302, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2429, "n_element_tab_1": 112, "formula_len_all": 1633, "formula_len_all_1": 1067, "len_all": 294745, "len_all_1": 56109, "len_abs": 1196, "len_title": 137, "len_sents": 68938, "len_sents_1": 25320, "n_sents": 452, "n_sents_1": 150, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1152, "title": "ML4MILP: A Benchmark Dataset for Machine Learning-based Mixed-Integer Linear Programming", "abs": "Machine learning (ML)-based approaches for solving mixed integer linear programming (MILP) problems have shown significant potential and are growing in sophistication. Despite this advancement, progress in this field is often hindered by the mixed and unsorted nature of current benchmark datasets, which typically lack carefully categorized collections of homogeneous instances.\nTo bridge this gap, we propose ML4MILP, a new open-source benchmark dataset specifically designed for evaluating ML-based optimization algorithms in the MILP domain. Based on the proposed structure and embedding similarity metrics, we used a novel classification algorithm to carefully categorize the collected and generated instances, resulting in a benchmark dataset encompassing 100,000 instances across more than 70 heterogeneous classes.\nWe demonstrate the utility of ML4MILP through extensive benchmarking against a comprehensive suite of algorithms in the baseline library, consisting of traditional exact solvers and heuristic algorithms, as well as ML-based approaches. Our ML4MILP is open-source and accessible at: https://anonymous.4open.science/r/ML4MILP-6BE0.", "keywords": [ "Mixed Integer Linear Programming", "Machine Learning", "Benchmark Dataset" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "ue1Tt3h1VC", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Binbin Hu", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Huajun Chen", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Lingbing Guo", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Wen Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "Assistant Professor" }, { "name": "Yichi Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zhuo Chen", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Ziqi Liu", "gender": "Male", "institution": "Ant Group", "country": "", "position": "Researcher" }, { "name": "yajing Xu", "gender": "Female", "institution": "College of Computer Science and Technology, Zhejiang University", "country": "", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 15, "n_ref_uni": 37, "n_ref": 104, "n_ref_all": 124, "n_fig": 4, "n_tab": 8, "L_tab": 2726, "n_element_tab": 173, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2276, "n_element_tab_1": 348, "formula_len_all": 2420, "formula_len_all_1": 917, "len_all": 147913, "len_all_1": 65521, "len_abs": 1231, "len_title": 156, "len_sents": 42632, "len_sents_1": 27465, "n_sents": 324, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 108, "L_abs": 1225, "title": "Multiple Heads are Better than One: Mixture of Modality Knowledge Experts for Entity Representation Learning", "abs": "Learning high-quality multi-modal entity representations is an important goal of multi-modal knowledge graph (MMKG) representation learning, which can en- hance reasoning tasks within the MMKGs, such as MMKG completion (MMKGC). The main challenge is to collaboratively model the structural information concealed in massive triples and the multi-modal features of the entities. Existing methods focus on crafting elegant entity-wise multi-modal fusion strategies, yet they over- look the utilization of multi-perspective features concealed within the modalities under diverse relational contexts. To address this issue, we introduce a novel framework with Mixture of Modality Knowledge experts (MOMOK for short) to learn adaptive multi-modal entity representations for better MMKGC. We design relation-guided modality knowledge experts to acquire relation-aware modality embeddings and integrate the predictions from multi-modalities to achieve joint decisions. Additionally, we disentangle the experts by minimizing their mutual information. Experiments on four public MMKG benchmarks demonstrate the outstanding performance of MOMOK under complex scenarios. Our code and data are available at https://github.com/zjukg/MoMoK.", "keywords": [ "Multi-modal Information Fusion", "Knowledge Graph", "Multi-modal Entity Representation", "Mixture-of-Experts" ], "rating_list": [ 8, 6, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 4, 3, 2 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "udtrtwkvk5", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Chuyan Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Kun Yuan", "gender": "Male", "institution": "Peking University", "country": "", "position": "Assistant Professor" }, { "name": "Pengrui Li", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Yipeng Hu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yutong He", "gender": "Male", "institution": "Academy for Advanced Interdisciplinary Studies, Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 266, "n_formula_1": 29, "n_ref_uni": 39, "n_ref": 61, "n_ref_all": 120, "n_fig": 7, "n_tab": 3, "L_tab": 427, "n_element_tab": 30, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 97, "n_element_tab_1": 6, "formula_len_all": 45803, "formula_len_all_1": 1978, "len_all": 261318, "len_all_1": 65249, "len_abs": 1233, "len_title": 123, "len_sents": 63407, "len_sents_1": 28993, "n_sents": 748, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1094, "title": "Subspace Optimiztion for Large Language Models with Convergence Guarantees", "abs": "Subspace optimization algorithms, with GaLore (Zhao et al., 2024) as a representative method, have gained popularity for pre-training or fine-tuning large language models (LLMs) due to their memory efficiency. However, their convergence guarantees remain unclear, particularly in stochastic settings. In this paper, we unexpectedly discover that GaLore does not always converge to the optimal solution and substantiate this finding with an explicit counterexample. We then investigate the conditions under which GaLore can achieve convergence, demonstrating that it does so either in deterministic scenarios or when using a sufficiently large mini-batch size. More significantly, we introduce **GoLore** (**G**radient rand**o**m **Lo**w-**r**ank proj**e**ction), a novel variant of GaLore that provably converges in stochastic settings, even with standard batch sizes. Our convergence analysis can be readily extended to other sparse subspace optimization algorithms. Finally, we conduct numerical experiments to validate our theoretical results and empirically explore the proposed mechanisms.", "keywords": [ "Large Language Models", "Memory-Efficient Training", "Subspace Learning" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "udfjje2xXb", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Andrea Mastropietro", "gender": "Male", "institution": "University of Roma \"La Sapienza\"", "country": "IT", "position": "Postdoc" }, { "name": "Aris Anagnostopoulos", "gender": "unknown", "institution": "Sapienza University of Rome", "country": "", "position": "Associate Professor" }, { "name": "Gianluca De Carlo", "gender": "Male", "institution": "University of Roma \"La Sapienza\"", "country": "IT", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 30, "n_ref": 39, "n_ref_all": 53, "n_fig": 5, "n_tab": 2, "L_tab": 692, "n_element_tab": 76, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1303, "n_element_tab_1": 83, "formula_len_all": 604, "formula_len_all_1": 604, "len_all": 97710, "len_all_1": 64721, "len_abs": 899, "len_title": 87, "len_sents": 34677, "len_sents_1": 30469, "n_sents": 238, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 901, "title": "Kolmogorov\u2013Arnold Graph Neural Networks", "abs": "Graph neural networks (GNNs) excel in learning from network-like data but often lack interpretability, making their application challenging in domains requiring transparent decision-making. We propose the Kolmogorov\u2013Arnold Network for Graphs (KANG), a novel GNN model leveraging spline-based activation functions on edges to enhance both accuracy and interpretability. Our experiments on five benchmark datasets demonstrate that KANG outperforms state-of-the-art GNN models in node classification, link prediction, and graph classification tasks. In addition to the improved accuracy, KANG\u2019s design inherently provides insights into the model\u2019s decision-making process, eliminating the need for post-hoc explainability techniques. This paper discusses the methodology, performance, and interpretability of KANG, highlighting its potential for applications in domains where interpretability is crucial.", "keywords": [ "Graph Neural Networks", "Kolmogorov-Arnold Networks", "Interpretability" ], "rating_list": [ 3, 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 3, 2, 2 ], "presentation_list": [ 2, 4, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 5, 4 ] }, { "paper_id": "ud8FtE1N4N", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Ahmed Imtiaz Humayun", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Amir Yazdanbakhsh", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Gintare Karolina Dziugaite", "gender": "Female", "institution": "McGill University", "country": "CA", "position": "Adjunct Professor" }, { "name": "Suvinay Subramanian", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Tian Jin", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Utku Evci", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Dan Alistarh", "gender": "Male", "institution": "Institute of Science and Technology", "country": "AT", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 36, "n_ref": 78, "n_ref_all": 102, "n_fig": 9, "n_tab": 2, "L_tab": 647, "n_element_tab": 37, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 647, "n_element_tab_1": 37, "formula_len_all": 122, "formula_len_all_1": 146, "len_all": 124907, "len_all_1": 64605, "len_abs": 1425, "len_title": 124, "len_sents": 35269, "len_sents_1": 30669, "n_sents": 269, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1430, "title": "Rethinking Sparse Scaling through the Lens of Average Active Parameter Count", "abs": "Parameter pruning has emerged as a promising technique to address the growing computational demand of large language models (LLMs). While many studies focus on post-training pruning of LLMs, sparse pre-training offers a compelling alternative: sparsifying during pre-training reduces both training and inference costs. In this work, we conduct the first comprehensive study on optimal sparse pre-training configurations for LLMs, exploring various pruning schedules across different sparsity levels and training duration. We evaluate 80 unique configurations and find that a pruning schedule starting at 25% of total training compute and ending at 75% achieves near-optimal final evaluation loss. Our findings provide valuable insights for efficient and effective sparse pre-training of LLMs. Furthermore, we propose a new scaling law that modifies the Chinchilla scaling law to use the average number of active parameters during training. We present both empirical and theoretical evidence that this modification accurately models evaluation loss for both sparsely and densely pre-trained LLMs, thus offering a unified scaling law for dense and sparse model training. Our insights suggest that, while sparse pre-training yields similar model loss as dense pre-training for the same compute budget, it offers a clear advantage: the final model is smaller, resulting in significant potential computational savings during inference.", "keywords": [ "pruning", "sparsity", "large language model", "pretraining" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "ubuGgIPVD0", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Naiyan Wang", "gender": "Male", "institution": "Tusimple", "country": "", "position": "Chief Scientist" }, { "name": "Xiaojie Guo", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" }, { "name": "Yan Yan", "gender": "Male", "institution": "Chongqing University", "country": "", "position": "MS student" }, { "name": "Yuheng Shi", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Researcher" }, { "name": "Zehao Huang", "gender": "Male", "institution": "Xiaomi EV", "country": "CN", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 40, "n_ref": 64, "n_ref_all": 88, "n_fig": 8, "n_tab": 10, "L_tab": 2042, "n_element_tab": 161, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2618, "n_element_tab_1": 75, "formula_len_all": 420, "formula_len_all_1": 252, "len_all": 143366, "len_all_1": 70721, "len_abs": 2021, "len_title": 128, "len_sents": 46274, "len_sents_1": 32041, "n_sents": 393, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 966, "title": "TSTTC: A Large-Scale Dataset for Time-to-Contact Estimation in Driving Scenarios", "abs": "Time-to-Contact (TTC) estimation is a critical task for assessing collision risk and is widely used in various driver assistance and autonomous driving systems. The past few decades have witnessed development of related theories and algorithms. The prevalent learning-based methods call for a large-scale TTC dataset in real-world scenarios. In this work, we present a large-scale object oriented TTC dataset in the driving scene for promoting the TTC estimation by a monocular camera. To collect valuable samples and make data with different TTC values relatively balanced, we go through thousands of hours of driving data and select over 200K sequences with a preset data distribution. To augment the quantity of small TTC cases, we also generate clips using the latest Neural rendering methods. Additionally, we provide several simple yet effective TTC estimation baselines and evaluate them extensively on the proposed dataset to demonstrate their effectiveness.", "keywords": [ "Time-to-Contact Estimation", "Dataset" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "ubUTIlAH0m", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Souvik Chakraborty", "gender": "Male", "institution": "Indian Institute of Technology, Delhi", "country": "IN", "position": "Assistant Professor" }, { "name": "Tapas Tripura", "gender": "Male", "institution": "Indian Institute of Technology, Delhi", "country": "IN", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 48, "n_ref": 79, "n_ref_all": 98, "n_fig": 9, "n_tab": 8, "L_tab": 3892, "n_element_tab": 270, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 722, "n_element_tab_1": 38, "formula_len_all": 1483, "formula_len_all_1": 1208, "len_all": 159988, "len_all_1": 53613, "len_abs": 1606, "len_title": 114, "len_sents": 46482, "len_sents_1": 23241, "n_sents": 376, "n_sents_1": 162, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1610, "title": "Multi-Physics Operator Network for In-context learning (m-PhOeNIX)", "abs": "We propose a multi-physics operator network for simultaneous and sequential learning of solution operators of multiple heterogeneous parametric partial differential equations. Existing neural operators are adept at learning the solution operator of only a single physical system, and adapting to new physical equations requires training a new surrogate model from scratch with physics-specific intensive hyperparameter tuning. The proposed multi-physics neural operator leverages the recent advancements in wavelet-based kernel integral-induced neural operator modeling and instantiates a memory-based ensembling strategy for projecting heterogeneous physical systems into a common shared feature space. The local channel-level ensembling is supported by context gates, which not only utilize the shared features to embed the features of multiple heterogeneous physical systems into the network parameters but also allow the multi-physics operator to learn new solution operators by transferring knowledge sequentially; this allows the proposed model to continually learn without forgetting. We illustrate the efficacy of our algorithm by simultaneously and sequentially learning six complex time-dependent solution operators of six physical systems. The inference results on the simultaneous and sequentially trained models depict the ability to infer previously seen physical systems without fine-tuning and catastrophic forgetting, indicating the characteristics of a foundation model. The framework also demonstrates the super-resolution property and generalization to out-of-distribution input conditions.", "keywords": [ "Multi-physics operator learning", "neural operator", "catastrophic forgetting", "continual learning", "wavelet" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 3, 4, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "ubIxE93FLM", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Heng Ji", "gender": "Female", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Jiajun Wu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Joy Hsu", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Kuan-Hao Huang", "gender": "Male", "institution": "Texas A&M University", "country": "US", "position": "Assistant Professor" }, { "name": "Manling Li", "gender": "Female", "institution": "Northwestern University", "country": "US", "position": "Assistant Professor" }, { "name": "Xingyao Wang", "gender": "Male", "institution": "All Hands AI", "country": "US", "position": "Researcher" }, { "name": "Zhenhailong Wang", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 181, "n_ref_all": 233, "n_fig": 17, "n_tab": 8, "L_tab": 1177, "n_element_tab": 195, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 428, "n_element_tab_1": 41, "formula_len_all": 43, "formula_len_all_1": 43, "len_all": 270120, "len_all_1": 62535, "len_abs": 2234, "len_title": 113, "len_sents": 77384, "len_sents_1": 31066, "n_sents": 581, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 2244, "title": "Visually Descriptive Language Model for Vector Graphics Reasoning", "abs": "Despite significant advancements, current large multimodal models (LMMs) struggle to bridge the gap between low-level visual perception\u2014focusing on shapes, sizes and layouts\u2014and high-level language reasoning involving semantics, events and logic. This limitation becomes evident in tasks requiring precise visual perception, such as comparing geometric properties or solving visual algorithmic reasoning problems. To study this failure mode, we focus on an important visual domain: vector graphics\u2014images composed purely of 2D objects and shapes, which are prevalent in various LMM-based agent tasks in web, visual design, and OS environments. We identify two key research questions: how can we enable precise visual perception, and how can we facilitate high-level reasoning based on such low-level perceptions? To accurately capture low-level visual details, we utilize Scalable Vector Graphics (SVG) for precise encoding of visual scenes. However, SVGs are not readily interpretable by LLMs or LMMs in a zero-shot manner. To address this challenge, we propose the Visually Descriptive Language Model (VDLM), which introduces an intermediate textual representation called Primal Visual Description (PVD). PVD translates SVGs into a text-based abstraction comprising primitive attributes (e.g., shape, position, measurement) along with their corresponding values. PVD can be learned with task-agnostic synthesized data and represents visual primitives that are universal across various vector graphics. This abstraction is more structured, allowing for direct interpretation by foundation models for zero-shot generalization to different reasoning tasks. Without any human-annotated data, empirical results demonstrate that VDLM leads to significant improvements in state-of-the-art LMMs, such as GPT-4o, across various low-level multimodal perception and reasoning tasks on vector graphics. Additionally, we provide extensive analyses of VDLM\u2019s performance, showing that our framework offers improved interpretability due to its disentangled perception and reasoning processes. Finally, we demonstrate the promise of this representation by showing a positive correlation between the quality of the PVD perception and the end-task performance.", "keywords": [ "Large Multimodal Model", "Large Language Model", "Vector Graphics", "Low-level Perception", "Low-level Visual Reasoning" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "uaMSBJDnRv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Adithya Bhaskar", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Boris Hanin", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Danqi Chen", "gender": "Female", "institution": "Department of Computer Science, Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Noam Razin", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Sadhika Malladi", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Sanjeev Arora", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "Full Professor" } ], "n_formula": 99, "n_formula_1": 10, "n_ref_uni": 52, "n_ref": 139, "n_ref_all": 228, "n_fig": 7, "n_tab": 13, "L_tab": 5937, "n_element_tab": 778, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 37, "n_element_tab_1": 1, "formula_len_all": 13145, "formula_len_all_1": 947, "len_all": 315421, "len_all_1": 78734, "len_abs": 1929, "len_title": 133, "len_sents": 90029, "len_sents_1": 35455, "n_sents": 699, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1676, "title": "Unintentional Unalignment: Likelihood Displacement in Direct Preference Optimization", "abs": "Direct Preference Optimization (DPO) and its variants are increasingly used for aligning language models with human preferences. \nAlthough these methods are designed to teach a model to generate preferred responses more frequently relative to dispreferred responses, prior work has observed that the likelihood of preferred responses often decreases during training. The current work sheds light on the causes and implications of this counterintuitive phenomenon, which we term *likelihood displacement*. We demonstrate that likelihood displacement can be *catastrophic*, shifting probability mass from preferred responses to responses with an opposite meaning. As a simple example, training a model to prefer $\\texttt{No}$ over $\\texttt{Never}$ can sharply increase the probability of $\\texttt{Yes}$. Moreover, when aligning the model to refuse unsafe prompts, we show that such displacement can *unintentionally lead to unalignment*, by shifting probability mass from preferred refusal responses to harmful responses (e.g., reducing the refusal rate of Llama-3-8B-Instruct from 74.4% to 33.4%). We theoretically characterize that likelihood displacement is driven by preferences that induce similar embeddings, as measured by a *centered hidden embedding similarity (CHES)* score. Empirically, the CHES score enables identifying which training samples contribute most to likelihood displacement in a given dataset. Filtering out these samples effectively mitigated unintentional unalignment in our experiments. More broadly, our results highlight the importance of curating data with sufficiently distinct preferences, for which we believe the CHES score may prove valuable.", "keywords": [ "Direct Preference Optimization", "DPO", "Likelihood Displacement", "Unalignment", "Alignment", "Language Models" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "uaKBM9sGEm", "primary_area": "applications to robotics, autonomy, planning", "n_author": 3, "author_info_list": [ { "name": "Jeff Schneider", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Researcher" }, { "name": "Vedant Mundheda", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Zhouchonghao Wu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 25, "n_ref": 34, "n_ref_all": 47, "n_fig": 5, "n_tab": 4, "L_tab": 1125, "n_element_tab": 78, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 837, "n_element_tab_1": 62, "formula_len_all": 1085, "formula_len_all_1": 722, "len_all": 112772, "len_all_1": 63111, "len_abs": 1731, "len_title": 122, "len_sents": 39161, "len_sents_1": 29661, "n_sents": 301, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1689, "title": "Towards Off-Road Autonomous Driving via Planner Guided Policy Optimization", "abs": "Off-road autonomous driving poses significant challenges such as navigating diverse terrains, avoiding obstacles, and maneuvering through ditches. Addressing these challenges requires effective planning and adaptability, making it a long-horizon planning and control problem. Traditional model-based control techniques like Model Predictive Path Integral (MPPI) require dense sampling and accurate modeling of the vehicle-terrain interaction, both of which are computationally expensive, making effective long-horizon planning in real-time intractable. Reinforcement learning (RL) methods operate without this limitation and are computationally cheaper at deployment. However, exploration in obstacle-dense and challenging terrains is difficult, and typical RL techniques struggle to navigate in these terrains. To alleviate the limitations of MPPI, we propose a hierarchical autonomy pipeline with a low-frequency high-level MPPI planner and a high-frequency low-level RL controller. To tackle RL's exploration challenge, we propose a teacher-student paradigm to learn an end-to-end RL policy, capable of real-time execution and traversal through challenging terrains. The teacher policy is trained using dense planning information from an MPPI planner while the student policy learns to navigate using visual inputs and sparse planning information. In this framework, we introduce a new policy gradient formulation that extends Proximal Policy Optimization (PPO), leveraging off-policy trajectories for teacher guidance and on-policy trajectories for student exploration. We demonstrate our performance in a realistic off-road simulator against various RL and imitation learning methods.", "keywords": [ "Reinforcement learning", "Learning from Demonstrations", "Autonomous driving", "Off-road driving" ], "rating_list": [ 6, 3, 1, 6 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "uaGNerHa1J", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Jian Li", "gender": "Male", "institution": "Beijing Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 35, "n_formula_1": 35, "n_ref_uni": 32, "n_ref": 61, "n_ref_all": 67, "n_fig": 1, "n_tab": 2, "L_tab": 1030, "n_element_tab": 39, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 1030, "n_element_tab_1": 39, "formula_len_all": 4304, "formula_len_all_1": 4317, "len_all": 121844, "len_all_1": 74697, "len_abs": 1052, "len_title": 106, "len_sents": 27771, "len_sents_1": 27819, "n_sents": 218, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1056, "title": "Efficient Newton-type Federated Learning with Non-IID Data", "abs": "The mainstream federated learning algorithms only communicate the first-order information across the local devices, i.e., FedAvg and FedProx. However, only using first-order information, these methods are often inefficient and the impact of heterogeneous data is yet not precisely understood. This paper proposes an efficient federated Newton method (FedNewton), by sharing both first-order and second-order knowledge over heterogeneous data. In general kernel ridge regression setting, we derive the generalization bounds for FedNewton and obtain the minimax-optimal learning rates. For the first time, our results analytically quantify the impact of the number of local examples, the data heterogeneity and the model heterogeneity. Moreover, as long as the local sample size is not too small and data heterogeneity is moderate, the federated error in FedNewton decreases exponentially in terms of iterations. Extensive experimental results further validate our theoretical findings and illustrate the advantages of FedNewton over the first-order methods.", "keywords": [ "Federated learning", "Newton-type optimization", "Generalization analysis", "Integral operator theory" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 4, 3, 2 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 4, 2, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "ua5MHdsbck", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Bella Dubrov", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Mostafa Karimi", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Ron Benson", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Director, Machine Learning" }, { "name": "Shang Shang", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Sharmi Banerjee", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Tommi Jaakkola", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 89, "n_ref_all": 126, "n_fig": 9, "n_tab": 23, "L_tab": 18776, "n_element_tab": 1614, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 6146, "n_element_tab_1": 529, "formula_len_all": 662, "formula_len_all_1": 503, "len_all": 178321, "len_all_1": 69575, "len_abs": 979, "len_title": 136, "len_sents": 43747, "len_sents_1": 28152, "n_sents": 309, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 980, "title": "Data Distillation for extrapolative protein design through exact preference optimization", "abs": "The goal of protein design typically involves increasing fitness (extrapolating) beyond what is seen during training (e.g., towards higher stability, stronger binding affinity, etc.). State-of-the-art methods assume that one can safely steer proteins towards such extrapolated regions by learning from pairs alone. We hypothesize that noisy training pairs are not sufficiently informative to capture the fitness gradient and that models learned from pairs specifically may fail to capture three-way relations important for search, e.g., how two alternatives fair relative to a seed. Building on the success of preference alignment models in large language models, we introduce a progressive search method for extrapolative protein design by directly distilling into the model relevant triplet relations. We evaluated our model's performance in designing AAV and GFP proteins and demonstrated that the proposed framework significantly improves effectiveness in extrapolation tasks.", "keywords": [ "Protein design", "Protein Language Models", "Preference Learning", "Extrapolation", "Data distillation" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "uZmmgHY1mD", "primary_area": "foundation or frontier models, including LLMs", "n_author": 20, "author_info_list": [ { "name": "Arya Talebzadeh", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Undergrad student" }, { "name": "Chen Zhu", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Di Jin", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Eric Han", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Eryk Helenowski", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Gabriel Cohen", "gender": "unknown", "institution": "Meta ", "country": "US", "position": "Researcher" }, { "name": "Han Fang", "gender": "Male", "institution": "Meta AI", "country": "", "position": "Research Scientist Manager" }, { "name": "Hao Ma", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Director" }, { "name": "Hejia Zhang", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Kaiyan Peng", "gender": "Female", "institution": "Meta", "country": "US", "position": "Research Scientist" }, { "name": "Karishma Mandyam", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Meta" }, { "name": "Karthik Abinav Sankararaman", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Madian Khabsa", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Shaoliang Nie", "gender": "Male", "institution": "Meta Inc", "country": "US", "position": "Researcher" }, { "name": "Sinong Wang", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research scientist" }, { "name": "Tengyu Xu", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Wenxuan Zhou", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Yuandong Tian", "gender": "Male", "institution": "Meta AI (FAIR)", "country": "", "position": "Research Scientist" }, { "name": "Yun He", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Research Scientist " }, { "name": "Zhouhao Zeng", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 41, "n_ref": 65, "n_ref_all": 79, "n_fig": 3, "n_tab": 3, "L_tab": 800, "n_element_tab": 166, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1038, "formula_len_all_1": 1016, "len_all": 172943, "len_all_1": 61817, "len_abs": 1939, "len_title": 105, "len_sents": 32100, "len_sents_1": 29677, "n_sents": 216, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1998, "title": "The Perfect Blend: Redefining RLHF with Mixture of Judges", "abs": "Reinforcement learning from human feedback (RLHF) has become the leading approach for fine-tuning large language models (LLM). However, RLHF has limitations in multi-task learning (MTL) due to challenges of reward hacking and extreme multi-objective optimization (i.e., trade-off of multiple and/or sometimes conflicting objectives). Applying RLHF for MTL currently requires careful tuning of the weights for reward model and data combinations. This is often done via human intuition and does not generalize. In this work, we introduce a novel post-training paradigm which we called Constrained Generative Policy Optimization (CGPO). The core of CGPO is Mixture of Judges (MoJ) with cost-efficient constrained policy optimization with stratification, which can identify the perfect blend in RLHF in a principled manner. It shows strong empirical results with theoretical guarantees, does not require extensive hyper-parameter tuning, and is plug-and-play in common post-training pipelines. Together, this can detect and mitigate reward hacking behaviors while reaching a pareto-optimal point across an extremely large number of objectives. \n\nOur results show that CGPO consistently outperforms other commonly used SoTA RLHF algorithms (such as PPO and DPO) on a wide range of tasks -- general chat, STEM questions, instruction following, math, coding and knowledge. In particular, CGPO improves over PPO by 7.4% in AlpacaEval-2 (general chat), 12.5% in Arena-Hard (STEM\\reasoning), 2% in IFEval (Instrcution Following), 2% in both MATH and GSM8K (Math\\reasoning), 5% in HumanEval (Coding), and 2% in the ARC challenge (Knowledge). We also observe that PPO is susceptible to severe reward hacking behaviors (it exhibits severe regression in popular coding benchmarks) which can be addressed by CGPO. CGPO represents a breakthrough in RLHF, simultaneously addressing reward-hacking and extreme multi-objective optimization, and thereby advancing the state-of-the-art in aligning general-purpose LLMs.", "keywords": [ "Large Language Model", "Reinforcement Learning from Human Feedback", "Mixture of Judges", "Constrained Policy Optimization" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 4, 3 ], "contribution_list": [ 1, 3, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "uZgK0tcPqd", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Alexandros Karatzoglou", "gender": "Male", "institution": "Google Deepmind", "country": "CH", "position": "Research Scientist" }, { "name": "Carlos Segura", "gender": "Male", "institution": "Telefonica Research", "country": "", "position": "Researcher" }, { "name": "Ioannis Arapakis", "gender": "Male", "institution": "Telef\u00f3nica Innovaci\u00f3n Digital", "country": "ES", "position": "Principal Researcher" }, { "name": "Sergi Abadal", "gender": "Male", "institution": "Universitat Polit\u00e8cnica de Catalunya", "country": "ES", "position": "Principal Researcher" }, { "name": "\u00c1ngela L\u00f3pez-Cardona", "gender": "Female", "institution": "Universidad Polit\u00e9cnica de Cataluna", "country": "ES", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 6, "n_ref_uni": 60, "n_ref": 236, "n_ref_all": 273, "n_fig": 9, "n_tab": 7, "L_tab": 1959, "n_element_tab": 137, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1860, "n_element_tab_1": 91, "formula_len_all": 402, "formula_len_all_1": 298, "len_all": 376690, "len_all_1": 66436, "len_abs": 1104, "len_title": 139, "len_sents": 53137, "len_sents_1": 32540, "n_sents": 384, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1111, "title": "Seeing Eye to AI: Human Alignment via Gaze-Based Response Rewards for Large Language Models", "abs": "Advancements in Natural Language Processing (NLP), have led to the emergence of Large Language Models (LLMs) such as GPT, Llama, Claude, and Gemini, which excel across a range of tasks but require extensive fine-tuning to align their outputs with human expectations. A widely used method for achieving this alignment is Reinforcement Learning from Human Feedback (RLHF), which, despite its success, faces challenges in accurately modelling human preferences. In this paper, we introduce GazeReward, a novel framework that integrates implicit feedback -- and specifically eye-tracking (ET) data -- into the Reward Model (RM). In addition, we explore how ET-based features can provide insights into user preferences. Through ablation studies we test our framework with different integration methods, LLMs, and ET generator models, demonstrating that our approach significantly improves the accuracy of the RM on established human preference datasets. This work advances the ongoing discussion on optimizing AI alignment with human values, exploring the potential of cognitive data for shaping future NLP research.", "keywords": [ "reward model", "RLHF", "visual attention", "LLMs", "eye tracking", "implicit feedback" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "uZVDJfV2Ex", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Danial Davarnia", "gender": "unknown", "institution": "Iowa State University", "country": "", "position": "Assistant Professor" }, { "name": "Mohammadreza Kiaghadi", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "PhD student" } ], "n_formula": 39, "n_formula_1": 20, "n_ref_uni": 30, "n_ref": 40, "n_ref_all": 44, "n_fig": 1, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5569, "formula_len_all_1": 2509, "len_all": 120534, "len_all_1": 76268, "len_abs": 2063, "len_title": 160, "len_sents": 43270, "len_sents_1": 34947, "n_sents": 315, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 110, "L_abs": 2204, "title": "A graph-based global optimization framework for problems with nonconvex norm constraints and penalty functions", "abs": "Optimization problems with norm-bounding constraints appear in various applications, from portfolio optimization to machine learning, feature selection, and beyond. A widely used variant of these problems relaxes the norm-bounding constraint through Lagrangian relaxation and moves it to the objective function as a form of penalty or regularization term. A challenging class of these models uses the zero-norm function to induce sparsity in statistical parameter estimation models. Most existing exact solution methods for these problems use additional binary variables together with artificial bounds on variables to formulate them as a mixed-integer program in a higher dimension, which is then solved by off-the-shelf solvers. Other exact methods utilize specific structural properties of the objective function to solve certain variants of these problems, making them non-generalizable to other problems with different structures. An alternative approach employs nonconvex penalties with desirable statistical properties, which are solved using heuristic or local methods due to the structural complexity of those terms. In this paper, we develop a novel graph-based method to globally solve optimization problems that contain a generalization of norm-bounding constraints. This includes standard $\\ell_p$-norms for $p \\in [0, \\infty)$ as well as nonconvex penalty terms, such as SCAD and MCP, as special cases. Our method uses decision diagrams to build strong convex relaxations for these constraints in the original space of variables without the need to introduce additional auxiliary variables or impose artificial variable bounds. We show that the resulting convexification method, when incorporated into a spatial branch-and-cut framework, converges to the global optimal value of the problem under mild conditions. To demonstrate the capabilities of the proposed framework, we conduct preliminary computational experiments on benchmark sparse linear regression problems with complex nonconvex penalty terms that existing global solvers cannot model or solve. This establishes our framework as the first algorithm capable of globally solving such challenging mixed-integer nonlinear programs.", "keywords": [ "Norm Constraints", "Sparse Parameter Estimation", "Nonconvex Regularization", "Global Optimization", "Mixed-Integer Nonlinear Programs", "Decision Diagrams" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "uZFXpPrwSh", "primary_area": "learning on time series and dynamical systems", "n_author": 9, "author_info_list": [ { "name": "Abdelhakim Benechehab", "gender": "Male", "institution": "Eurecom", "country": "", "position": "PhD student" }, { "name": "Albert Thomas", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Ambroise Odonnat", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "FR", "position": "PhD student" }, { "name": "Giuseppe Paolo", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "FR", "position": "Researcher" }, { "name": "Ievgen Redko", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "FR", "position": "Principal Researcher" }, { "name": "Maurizio Filippone", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Associate Professor" }, { "name": "Oussama Zekri", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Intern" }, { "name": "Youssef Attia El Hili", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "FR", "position": "Researcher" }, { "name": "Balazs Kegl", "gender": "Male", "institution": "Huawei France", "country": "FR", "position": "Principal Researcher" } ], "n_formula": 22, "n_formula_1": 7, "n_ref_uni": 44, "n_ref": 90, "n_ref_all": 131, "n_fig": 19, "n_tab": 2, "L_tab": 113, "n_element_tab": 10, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 51, "n_element_tab_1": 7, "formula_len_all": 3027, "formula_len_all_1": 489, "len_all": 226953, "len_all_1": 65542, "len_abs": 1459, "len_title": 121, "len_sents": 57663, "len_sents_1": 30372, "n_sents": 427, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1081, "title": "Zero-shot Model-based Reinforcement Learning using Large Language Models", "abs": "The emerging zero-shot capabilities of Large Language Models (LLMs) have led to their applications in areas extending well beyond natural language processing tasks. \nIn reinforcement learning, while LLMs have been extensively used in text-based environments, their integration with continuous state spaces remains understudied. \nIn this paper, we investigate how pre-trained LLMs can be leveraged to predict in context the dynamics of continuous Markov decision processes. \nWe identify handling multivariate data and incorporating the control signal as key challenges that limit the potential of LLMs' deployment in this setup and propose Disentangled In-Context Learning (DICL) to address them.\nWe present proof-of-concept applications in two reinforcement learning settings: model-based policy evaluation and data-augmented off-policy reinforcement learning, supported by theoretical analysis of the proposed methods.\nOur experiments further demonstrate that our approach produces well-calibrated uncertainty estimates. We release the code at https://github.com/abenechehab/dicl.", "keywords": [ "Model-based Reinforcement Learning", "Large language models", "Zero-shot Learning", "In-context Learning" ], "rating_list": [ 8, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "uYzJvP8HGl", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Qingxu Fu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Tenghai Qiu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Assistant Professor" }, { "name": "Tianyi Hu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yuan Wang", "gender": "unknown", "institution": "Beijing University of Technology", "country": "CN", "position": "MS student" }, { "name": "Zhiqiang Pu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 35, "n_ref": 57, "n_ref_all": 86, "n_fig": 12, "n_tab": 3, "L_tab": 770, "n_element_tab": 68, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 653, "n_element_tab_1": 56, "formula_len_all": 885, "formula_len_all_1": 780, "len_all": 188592, "len_all_1": 65100, "len_abs": 1330, "len_title": 153, "len_sents": 70426, "len_sents_1": 28873, "n_sents": 577, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1334, "title": "UMAP: A Highly Extensible and Physics-Based Simulation Environment for Multi-agent Reinforcement Learning", "abs": "Existing simulation environments in the field of multi-agent reinforcement learning (MARL) either lack authenticity or complexity. The data generated by these environments significantly deviate from the requirements of the real world, hindering the practical application of MARL. To address this issue, we propose Unreal Multi-Agent Playground (UMAP), a highly extensible, physics-based 3D simulation environment implemented on the Unreal Engine. UMAP is user-friendly in terms of deployment, modification, and visualization, and all its components are open-sourced. Based on UMAP, we design a series of MARL tasks featuring heterogeneous agents, large-scale agents, multiple teams, and sparse team rewards.\nWe also develop an experimental framework compatible with algorithms ranging from \nrule-based to MARL-based provided by third-party frameworks. In the experimental section, we utilize the designed tasks to test several state-of-the-art algorithms. Additionally, We also conduct a physical experiment to demonstrate UMAP's potential in sim-to-real applications, which is a significant advantage due to the high extensibility and authenticity of UMAP. We believe UMAP can play an important role in the MARL field by evaluating existing algorithms and helping them apply to real-world scenarios, thus advancing the field of MARL.", "keywords": [ "multi-agent reinforcement learning", "simulation environment", "reinforcement learning" ], "rating_list": [ 3, 8, 6, 3, 8 ], "soundness_list": [ 3, 4, 3, 1, 4 ], "presentation_list": [ 2, 3, 2, 2, 4 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 5, 4 ] }, { "paper_id": "uYAG9Gla5u", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "H. \u00c7a\u011fr\u0131 Bilgi", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "PhD student" }, { "name": "Kubilay Atasu", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "Associate Professor" }, { "name": "Lydia Y. Chen", "gender": "Female", "institution": "Delft University of Technology", "country": "", "position": "Associate Professor" } ], "n_formula": 26, "n_formula_1": 12, "n_ref_uni": 22, "n_ref": 68, "n_ref_all": 96, "n_fig": 6, "n_tab": 4, "L_tab": 2204, "n_element_tab": 200, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1410, "n_element_tab_1": 87, "formula_len_all": 1566, "formula_len_all_1": 690, "len_all": 148815, "len_all_1": 65943, "len_abs": 1502, "len_title": 118, "len_sents": 50730, "len_sents_1": 29286, "n_sents": 379, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1508, "title": "Multigraph Message Passing with Bi-Directional Multi-Edge Aggregations", "abs": "Graph Neural Networks (GNNs) have seen significant advances in recent years, yet their application to multigraphs, where parallel edges exist between the same pair of nodes, remains under-explored. Standard GNNs, designed for simple graphs, compute node representations by combining all connected edges at once, without distinguishing between edges from different neighbors. There are some GNN architectures proposed specifically for multigraph tasks, yet these architectures perform only node-level aggregation in their message-passing layers, which limits their expressive power. Furthermore, these approaches either lack permutation equivariance when a strict total edge ordering is absent, or fail to preserve the topological structure of the multigraph. To address all these shortcomings, we propose MEGA-GNN, a unified framework for message passing on multigraphs that can effectively perform diverse graph learning tasks. Our approach introduces a two-stage aggregation process in the message passing layers: first, parallel edges are aggregated, followed by a node-level aggregation that operates on aggregated messages from distinct neighbors. We show that MEGA-GNN supports permutation equivariance and invariance properties. We also show that MEGA-GNN is universal given a strict total order on the edges. Experiments on synthetic and real-world financial transaction datasets demonstrate that MEGA-GNN either significantly outperforms or is on par with the accuracy of state-of-the-art solutions.", "keywords": [ "graph neural networks", "multigraph", "message passing", "financial fraud detection" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "uXytIlC1iQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Eva Zhu", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Intern" }, { "name": "Tang Zhengzheng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Intern" } ], "n_formula": 35, "n_formula_1": 18, "n_ref_uni": 21, "n_ref": 45, "n_ref_all": 58, "n_fig": 2, "n_tab": 5, "L_tab": 1093, "n_element_tab": 113, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 253, "n_element_tab_1": 24, "formula_len_all": 1850, "formula_len_all_1": 1080, "len_all": 187404, "len_all_1": 56083, "len_abs": 3746, "len_title": 105, "len_sents": 40116, "len_sents_1": 27300, "n_sents": 284, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1388, "title": "BrainGPT: A Brain-Inspired SNN-Based Large Language Model", "abs": "Large language models (LLMs) based on artificial neural networks (ANNs) have demonstrated remarkable performance but face challenges in computational efficiency and biological interpretability. We propose BrainGPT, a novel LLM architecture based on the Test-Time Training (TTT) framework and inspired by spiking neural networks (SNNs) and neurobiological principles. Our approach incorporates a dual-model structure, emulating the hierarchical language processing observed in the human brain, and utilizes a specialized integrate-and-fire neuron model with adaptive thresholding. Through a multi-stage training strategy, including quantization-aware pre-training, ANN-to-SNN conversion, and biologically inspired unsupervised learning, we achieve a mathematically proven lossless conversion from ANN to SNN, preserving 100\\% of the original ANN model's performance. Moreover, the biologically inspired unsupervised learning optimizes the maximum time steps required to maintain 100\\% ANN performance. Compared to the original TTT model, BrainGPT achieves a 33.4\\% increase in energy efficiency and demonstrates a 66.7\\% improvement in training convergence speed. This work advances the development of energy-efficient and biologically interpretable large language models that match the performance of state-of-the-art ANN-based models while significantly improving upon the TTT framework.", "keywords": [ "Spiking Neural Networks", "Large Language Models", "Spike-Timing-Dependent Plasticity", "Neuromorphic Computing", "ANN-to-SNN Conversion" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "uXmRmaF5g0", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Miqing Li", "gender": "Male", "institution": "University of Birmingham", "country": "", "position": "Assistant Professor" }, { "name": "Xunzhao Yu", "gender": "Male", "institution": "University of Warwick", "country": "GB", "position": "Postdoc" } ], "n_formula": 44, "n_formula_1": 10, "n_ref_uni": 37, "n_ref": 85, "n_ref_all": 122, "n_fig": 4, "n_tab": 14, "L_tab": 10105, "n_element_tab": 625, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7275, "formula_len_all_1": 527, "len_all": 305284, "len_all_1": 76571, "len_abs": 2065, "len_title": 139, "len_sents": 78670, "len_sents_1": 29928, "n_sents": 1641, "n_sents_1": 444, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 2109, "title": "LORA-MaOO: Learning Ordinal Relations and Angles for Expensive Many-Objective Optimization", "abs": "Many-objective optimization (MaOO) simultaneously optimizes many conflicting objectives to identify the Pareto front - a set of diverse solutions that represent different optimal balances between conflicting objectives. For expensive MaOO problems, due to their costly function evaluations, computationally cheap surrogates have been widely used in MaOO to save evaluation budget. However, as the number of objectives $M$ increases, the cost of using surrogates increases rapidly as many optimization algorithms need maintain $M$ surrogates. In addition, a large $M$ indicates a high-dimensional objective space, increasing the difficulty of maintaining solution diversity. \nIt is a challenge to reach diverse optimal solutions with a relatively low cost of using surrogates for MaOO problems. \nTo handle this challenge, we propose LORA-MaOO, a surrogate-assisted MaOO algorithm that learns $M$ surrogates from spherical coordinates, including an ordinal-regression-based surrogate that learns the ordinal relations between solutions (denoted as radial surrogate) and $M$-1 regression-based surrogates that trained on angular coordinates (denoted as angular surrogates).\nIn each optimization iteration, model-based search is completed with a single radial surrogate, while $M$-1 angular surrogates are used only once for selecting diverse candidates. Therefore, the frequency of using angular surrogates is largely reduced, lowering the cost of using surrogates. \nIn addition, we design a clustering method to quantify artificial ordinal relations for non-dominated solutions and improve the quantification of dominance-based ordinal relations. These ordinal relations are used to train the radial regression surrogate which predicts how desirable the candidate solutions are in terms of convergence. The solution diversity is maintained via angles between solutions instead of pre-defined auxiliary reference vectors, which is parameter-free. Experimental results show that LORA-MaOO significantly outperforms other surrogate-assisted MaOO methods on most MaOO benchmark problems and real-world applications.", "keywords": [ "Expensive optimization", "many-objective optimization", "surrogate-assisted optimization", "Gaussian Processes", "ordinal regression" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "uXLXq4ugAy", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Anna Litvin", "gender": "unknown", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Daria Frolova", "gender": "Female", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Dmitry N. Ivankov", "gender": "unknown", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Assistant Professor" }, { "name": "Ilya Sharov", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "MS student" }, { "name": "Ivan Oseledets", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "Marina A. Pak", "gender": "unknown", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 34, "n_ref": 89, "n_ref_all": 114, "n_fig": 4, "n_tab": 11, "L_tab": 5293, "n_element_tab": 567, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2297, "n_element_tab_1": 270, "formula_len_all": 1571, "formula_len_all_1": 742, "len_all": 166726, "len_all_1": 64089, "len_abs": 1256, "len_title": 124, "len_sents": 49199, "len_sents_1": 28537, "n_sents": 393, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1261, "title": "MULAN: Multimodal Protein Language Model for Sequence and Structure Encoding", "abs": "Most protein language models (PLMs), which produce high-quality protein representations, use only protein sequences during training.\nHowever, the known protein structure is crucial in many protein property prediction tasks, so there is a growing interest in incorporating the knowledge about the protein structure into a PLM. Currently, structure-aware PLMs are trained from scratch or introduce a huge parameter overhead for the structure encoder. In this study, we propose MULAN, a MULtimodal PLM for both sequence and ANgle-based structure encoding. MULAN has a pre-trained sequence encoder and an introduced parameter-efficient Structure Adapter, which are then fused and trained together. According to the evaluation on 9 downstream tasks, MULAN models of various sizes show quality improvement compared to both sequence-only ESM2 and structure-aware SaProt as well as comparable performance to Ankh, ESM3, ProstT5, and other PLMs considered in the study. Importantly, unlike other models, MULAN offers a cheap increase in the structural awareness of the protein representations due to finetuning of existing PLMs instead of training from scratch. We perform a detailed analysis of the proposed model and demonstrate its awareness of the protein structure.", "keywords": [ "Protein language model", "protein structure", "multimodal model", "downstream tasks" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "uWtLOy35WD", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 16, "author_info_list": [ { "name": "Chenning Xu", "gender": "unknown", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Fangxun Shu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Guanghao Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Hao Jiang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Haonan Shi", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Haoyuan Li", "gender": "Not Specified", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Le Zhuo", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Lei Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Si Liu", "gender": "Female", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Siming Fu", "gender": "Female", "institution": "alibabagroup", "country": "CN", "position": "Researcher" }, { "name": "Yue Liao", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "ZhongTao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Chan Long", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "robotx", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yuzhelun", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 53, "n_ref": 98, "n_ref_all": 109, "n_fig": 5, "n_tab": 15, "L_tab": 5084, "n_element_tab": 614, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1513, "n_element_tab_1": 123, "formula_len_all": 636, "formula_len_all_1": 531, "len_all": 187505, "len_all_1": 59939, "len_abs": 1611, "len_title": 107, "len_sents": 37351, "len_sents_1": 26505, "n_sents": 323, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1645, "title": "LLaVA-MoD: Making LLaVA Tiny via MoE-Knowledge Distillation", "abs": "We introduce LLaVA-MoD, a novel framework designed to enable the efficient training of small-scale Multimodal Language Models ($s$-MLLM) distilling knowledge from large-scale MLLM ($l$-MLLM). Our approach tackles two fundamental challenges in MLLM distillation. First, we optimize the network structure of $s$-MLLM by integrating a sparse Mixture of Experts (MoE) architecture into the language model, striking a balance between computational efficiency and model expressiveness. Second, we propose a progressive knowledge transfer strategy for comprehensive knowledge transfer. This strategy begins with mimic distillation, where we minimize the Kullback-Leibler (KL) divergence between output distributions to enable $s$-MLLM to emulate $s$-MLLM's understanding. Following this, we introduce preference distillation via Preference Optimization (PO), where the key lies in treating $l$-MLLM as the reference model. During this phase, the $s$-MLLM's ability to discriminate between superior and inferior examples is significantly enhanced beyond $l$-MLLM, leading to a better $s$-MLLM that surpasses $l$-MLLM, particularly in hallucination benchmarks.\nExtensive experiments demonstrate that LLaVA-MoD surpasses existing works across various benchmarks while maintaining a minimal activated parameters and low computational costs. Remarkably, LLaVA-MoD-2B surpasses Qwen-VL-Chat-7B with an average gain of 8.8\\%, using merely $0.3\\%$ of the training data and 23\\% trainable parameters. The results underscore LLaVA-MoD's ability to effectively distill comprehensive knowledge from its teacher model, paving the way for developing efficient MLLMs.", "keywords": [ "MLLM", "MoE", "Distillation" ], "rating_list": [ 6, 6, 5, 8, 6 ], "soundness_list": [ 3, 3, 2, 4, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "uWUovmBRUq", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Jie Zhang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Shiguang Shan", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xingming Long", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xilin Chen", "gender": "Male", "institution": "Institute of Computing Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 57, "n_formula_1": 15, "n_ref_uni": 15, "n_ref": 47, "n_ref_all": 78, "n_fig": 9, "n_tab": 5, "L_tab": 458, "n_element_tab": 65, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 844, "n_element_tab_1": 81, "formula_len_all": 4005, "formula_len_all_1": 1053, "len_all": 125834, "len_all_1": 59891, "len_abs": 1077, "len_title": 137, "len_sents": 45322, "len_sents_1": 28836, "n_sents": 339, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1079, "title": "Semantic or Covariate? A Study on the Intractable Case of Out-of-Distribution Detection", "abs": "The primary goal of out-of-distribution (OOD) detection tasks is to identify inputs with semantic shifts, i.e., if samples from novel classes are absent in the in-distribution (ID) dataset used for training, we should reject these OOD samples rather than misclassifying them into existing ID classes. However, we find the current definition of \"semantic shift\" is ambiguous, which renders certain OOD testing protocols intractable for the post-hoc OOD detection methods based on a classifier trained on the ID dataset. In this paper, we offer a more precise definition of the Semantic Space and the Covariate Space for the ID distribution, allowing us to theoretically analyze which types of OOD distributions make the detection task intractable. To avoid the flaw in the existing OOD settings, we further define the \"Tractable OOD\" setting which ensures the distinguishability of OOD and ID distributions for the post-hoc OOD detection methods. Finally, we conduct several experiments to demonstrate the necessity of our definitions and validate the correctness of our theorems.", "keywords": [ "Out-of-Distribution Detection", "Definition", "Theoretical Analysis" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "uWMQxtmyYz", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Liang Zhang", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Researcher" }, { "name": "Qingjiang Shi", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Full Professor" }, { "name": "Shan Dai", "gender": "Male", "institution": "The Chinese University of Hong Kong (Shenzhen)", "country": "CN", "position": "Researcher" }, { "name": "Shihua Gong", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhiwei Hu", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 32, "n_ref_uni": 12, "n_ref": 18, "n_ref_all": 27, "n_fig": 3, "n_tab": 3, "L_tab": 2225, "n_element_tab": 132, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 679, "n_element_tab_1": 98, "formula_len_all": 1414, "formula_len_all_1": 1409, "len_all": 91479, "len_all_1": 62345, "len_abs": 1634, "len_title": 117, "len_sents": 33067, "len_sents_1": 28537, "n_sents": 229, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1585, "title": "FedQLoRA: Federated Quantization-Aware LoRA for Large Language Models", "abs": "Large language models (LLMs) with billions of parameters have achieved remarkable success across various applications, but they require substantial computational resources and large datasets. While parameter-efficient fine-tuning methods like LoRA and QLoRA have significantly reduced computational costs and memory usage, robustly training LLMs for individual clients with datasets distributed on isolated devices remains challenging. To address this, recent work has explored the use of federated learning (FL) to collaboratively train LLM adapters on distributed private data, thereby avoiding the high computational and communication costs. In these approaches, the LLMs are frozen, and the adapters are collaboratively trained through adapter-sharing and aggregation methods. However, in this paper, we identify a significant issue: these approaches may suffer from quantization bias when clients operate with different levels of quantization on LLMs. To resolve this, we propose a novel framework called Federated Quantization-Aware LoRA (FedQLoRA), which estimates the quantization error and separates it from the LoRA adapter trained on local data via a quantization-aware adapter. Additionally, we address the heterogeneity bias problem that arises from severe data heterogeneity among clients, such as in non-IID settings. We propose an iterative version of the framework that improves both the dynamic quantization-aware adapter and the LoRA adapter alternately within the FL framework. We conduct extensive experiments to validate the performance of our proposed framework.", "keywords": [ "Quantization", "LoRA", "Large Language Models", "Federated Learning" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "uW3tNSx7PZ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Abdelhakim Senhaji Hafid", "gender": "unknown", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Full Professor" }, { "name": "YALDA KASRAVI MAVI", "gender": "unknown", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 33, "n_ref": 55, "n_ref_all": 60, "n_fig": 1, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 426, "formula_len_all_1": 404, "len_all": 103470, "len_all_1": 58117, "len_abs": 1725, "len_title": 119, "len_sents": 32533, "len_sents_1": 29525, "n_sents": 206, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1732, "title": "Gradients protection in federated learning for Biometric authentication", "abs": "In the context of face recognition models, different facial features contribute unevenly to a model's ability to correctly identify individuals, making some features more critical and, therefore, more susceptible to attacks.\nDeep Gradient Leakage (DGL) is a highly effective attack that recovers private training images from gradient vectors, posing significant privacy challenges in distributed learning systems where clients share gradients. Data augmentation, a technique for artificially manipulating the training set by creating modified copies of existing data, plays a crucial role in improving the accuracy of deep learning models. \nIn this paper, we explore various data augmentation methods to protect original training images, in test time thereby enhancing security in distributed learning systems as well as increasing accuracy during training. Our experiments demonstrate that augmentation methods improve model performance during training on augmented images, and we can use the same methods during testing as perturbation methods to preserve some features of the image and have safety against DGL.\n\nThis project has four primary objectives: first, to develop a vision transformer face validation model that trains on distributed devices to ensure privacy; second, to utilize augmentation methods to perturb private images and increase neural network safety; and third, to provide protection against attacks, ensuring that reconstructing attacks cannot extract sensitive information from gradients at any point in the system.\nand lastly we introduce a new novel perturbation method for a multi biometric authentication, system which offers accuracy for identification and guarantees safety and anonymity of entities.", "keywords": [ "federated learning", "security", "safety", "facial authentication" ], "rating_list": [ 1, 1, 5, 3 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 1, 1, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 5 ] }, { "paper_id": "uVm0zSNKkP", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Aniket Bera", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "Pascal Jutras Dube", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Ruqi Zhang", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 26, "n_formula_1": 20, "n_ref_uni": 37, "n_ref": 67, "n_ref_all": 93, "n_fig": 2, "n_tab": 5, "L_tab": 589, "n_element_tab": 63, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 824, "n_element_tab_1": 63, "formula_len_all": 1859, "formula_len_all_1": 1277, "len_all": 115167, "len_all_1": 55792, "len_abs": 1537, "len_title": 128, "len_sents": 30107, "len_sents_1": 23014, "n_sents": 221, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1357, "title": "Skip the Steps: Data-Free Consistency Distillation for Diffusion-based Samplers", "abs": "Sampling from probability distributions is a fundamental task in machine learning and statistics. However, most existing algorithms require numerous iterative steps to transform a prior distribution into high-quality samples, resulting in high computational costs and limiting their practicality in time-constrained and resource-limited environments. In this work, we propose consistency samplers, a novel class of samplers capable of generating high-quality samples in a single step. Our method introduces a new consistency distillation algorithm for diffusion-based samplers, which eliminates the need for data or full trajectory integration. By utilizing incomplete sampling trajectories and noisy intermediate representations along the diffusion process, we efficiently learn a direct one-step mapping from any state to its corresponding terminal state in the target distribution. Moreover, our approach enables few-step sampling, allowing users to flexibly balance compute costs and sample quality. We demonstrate the effectiveness of consistency samplers across multiple benchmark tasks, achieving high-quality results with one-step or few-step sampling while significantly reducing the sampling time compared to existing samplers. For instance, our method is 100-200x faster than prior diffusion-based samplers while having comparable sample quality.", "keywords": [ "Single-step sampling", "Diffusion-based sampler", "Distillation", "Generative modeling", "Optimal control" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "uVMZgtw2pf", "primary_area": "interpretability and explainable AI", "n_author": 1, "author_info_list": [ { "name": "Huaiguang Cai", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "MS student" } ], "n_formula": 27, "n_formula_1": 12, "n_ref_uni": 25, "n_ref": 73, "n_ref_all": 91, "n_fig": 3, "n_tab": 6, "L_tab": 3782, "n_element_tab": 369, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2106, "n_element_tab_1": 225, "formula_len_all": 4749, "formula_len_all_1": 1556, "len_all": 131308, "len_all_1": 63848, "len_abs": 1248, "len_title": 136, "len_sents": 36017, "len_sents_1": 27763, "n_sents": 280, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1255, "title": "CHG Shapley: Efficient Data Valuation and Selection towards Trustworthy Machine Learning", "abs": "Understanding the decision-making process of machine learning models is crucial for ensuring trustworthy machine learning. Data Shapley, a landmark study on data valuation, advances this understanding by assessing the contribution of each datum to model performance. However, the resource-intensive and time-consuming nature of multiple model retraining poses challenges for applying Data Shapley to large datasets. To address this, we propose the CHG (compound of Hardness and Gradient) utility function, which approximates the utility of each data subset on model performance in every training epoch. By deriving the closed-form Shapley value for each data point using the CHG utility function, we reduce the computational complexity to that of a single model retraining, achieving a quadratic improvement over existing marginal contribution-based methods. We further leverage CHG Shapley for real-time data selection, conducting experiments across three settings: standard datasets, label noise datasets, and class imbalance datasets. These experiments demonstrate its effectiveness in identifying high-value and noisy data. By enabling efficient data valuation, CHG Shapley promotes trustworthy model training through a novel data-centric perspective.", "keywords": [ "Data Valuation", "Shapley Value", "Data selection" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 2, 3, 4 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "uVDwunWsLz", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Gal Vardi", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "IL", "position": "Assistant Professor" }, { "name": "Roey Magen", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "IL", "position": "PhD student" }, { "name": "Shuning Shang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Spencer Frei", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Wei Hu", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" }, { "name": "Zhiwei Xu", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" } ], "n_formula": 404, "n_formula_1": 42, "n_ref_uni": 27, "n_ref": 56, "n_ref_all": 71, "n_fig": 4, "n_tab": 6, "L_tab": 743, "n_element_tab": 107, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 46920, "formula_len_all_1": 2435, "len_all": 352589, "len_all_1": 65168, "len_abs": 924, "len_title": 91, "len_sents": 98675, "len_sents_1": 27027, "n_sents": 1324, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 927, "title": "Benign Overfitting in Single-Head Attention", "abs": "The phenomenon of benign overfitting, where a trained neural network perfectly fits noisy training data but still achieves near-optimal test performance, has been extensively studied in recent years for linear models and fully-connected/convolutional networks. In this work, we study benign overfitting in a single-head softmax attention model, which is the fundamental building block of Transformers. We prove that under appropriate conditions, the model exhibits benign overfitting in a classification setting already after two steps of gradient descent. Moreover, we show conditions where a minimum-norm/maximum-margin interpolator exhibits benign overfitting. We study how the overfitting behavior depends on the signal-to-noise ratio (SNR) of the data distribution, namely, the ratio between norms of signal and noise tokens, and prove that a sufficiently large SNR is both necessary and sufficient for benign overfitting.", "keywords": [ "single-head attention", "benign overfitting", "transformers" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "uV9KFBVaFI", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Artur Hecker", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "DE", "position": "Principal Researcher" }, { "name": "Haokun Chen", "gender": "Male", "institution": "University of Munich, Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Jinhe Bi", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Volker Tresp", "gender": "Male", "institution": "Ludwig Maximilian University of Munich", "country": "DE", "position": "Associate Professor" }, { "name": "Xun Xiao", "gender": "Male", "institution": "Max Planck Institute for Cellular and Molecular Biology Genetics", "country": "DE", "position": "Postdoc" }, { "name": "Yujun Wang", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Intern" }, { "name": "Yunpu Ma", "gender": "Male", "institution": "Siemens Corporate Research", "country": "DE", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 7, "n_ref_uni": 41, "n_ref": 92, "n_ref_all": 107, "n_fig": 6, "n_tab": 3, "L_tab": 2264, "n_element_tab": 294, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2067, "n_element_tab_1": 221, "formula_len_all": 831, "formula_len_all_1": 441, "len_all": 218148, "len_all_1": 59664, "len_abs": 2063, "len_title": 148, "len_sents": 34597, "len_sents_1": 27208, "n_sents": 224, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 2027, "title": "Visual Instruction Tuning with 500x Fewer Parameters through Modality Linear Representation-Steering", "abs": "Multimodal Large Language Models (MLLMs) have significantly advanced visual tasks by integrating visual representations into large language models (LLMs). The textual modality, inherited from LLMs, equips MLLMs with abilities like instruction following and in-context learning. In contrast, the visual modality enhances performance in downstream tasks by leveraging rich semantic content, spatial information, and grounding capabilities. These intrinsic modalities work synergistically across various visual tasks.\nOur research initially reveals a persistent imbalance between these modalities, with text often dominating output generation during visual instruction tuning. This imbalance occurs when using both full fine-tuning and parameter-efficient fine-tuning (PEFT) methods. We then found that re-balancing these modalities can significantly reduce the number of trainable parameters required, inspiring a direction for further optimizing visual instruction tuning. Hence, in this paper, we introduce Modality Linear Representation-Steering (MoReS) to achieve the goal. MoReS effectively re-balances the intrinsic modalities throughout the model, where the key idea is to steer visual representations through linear transformations in the visual subspace across each model layer. \nTo validate our solution, we composed LLaVA Steering, a suite of models integrated with the proposed MoReS method. Evaluation results show that the composed LLaVA Steering models require, on average, 500 times fewer trainable parameters than LoRA needs while still achieving comparable performance across three visual benchmarks and eight visual question-answering tasks.\nLast, we present the LLaVA Steering Factory, an in-house developed platform that enables researchers to quickly customize various MLLMs with component-based architecture for seamlessly integrating state-of-the-art models, and evaluate their intrinsic modality imbalance. This open-source project enriches the research community to gain a deeper understanding of MLLMs.", "keywords": [ "MLLMs; PEFTs; Representation Steering" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "uV3Gdoq2ez", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Cheng Tan", "gender": "Male", "institution": "Zhejiang University & Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Dongxin Lyu", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Undergrad student" }, { "name": "Jingxuan Wei", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Siqi Ma", "gender": "Male", "institution": "Westlake.edu.cn", "country": "", "position": "Researcher" }, { "name": "Siyuan Li", "gender": "Male", "institution": "Westlake University & Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhangyang Gao", "gender": "Male", "institution": "Westlake University, China", "country": "CN", "position": "PhD student" }, { "name": "Zicheng Liu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Stan Z. Li", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Chair Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 43, "n_ref": 69, "n_ref_all": 86, "n_fig": 5, "n_tab": 7, "L_tab": 3812, "n_element_tab": 218, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1563, "n_element_tab_1": 148, "formula_len_all": 403, "formula_len_all_1": 235, "len_all": 252811, "len_all_1": 52082, "len_abs": 4053, "len_title": 167, "len_sents": 90471, "len_sents_1": 24019, "n_sents": 650, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Reject", "L_title": 118, "L_abs": 1328, "title": "Peer Review as A Multi-Turn and Long-Context Dialogue with Role-Based Interactions: Benchmarking Large Language Models", "abs": "Large Language Models (LLMs) have demonstrated wide-ranging applications across various fields and have shown significant potential in the academic peer-review process. However, existing applications are primarily limited to static review generation based on submitted papers, which fail to capture the dynamic and iterative nature of real-world peer reviews. In this paper, we reformulate the peer-review process as a multi-turn, long-context dialogue, incorporating distinct roles for authors, reviewers, and decision makers. We construct a comprehensive dataset containing over 30,854 papers with 110,642 reviews collected from the top-tier conferences. This dataset is meticulously designed to facilitate the applications of LLMs for multi-turn dialogues, effectively simulating the complete peer-review process. Furthermore, we propose a series of metrics to evaluate the performance of LLMs for each role under this reformulated peer-review setting, ensuring fair and comprehensive evaluations. We believe this work provides a promising perspective on enhancing the LLM-driven peer-review process by incorporating dynamic, role-based interactions. It aligns closely with the iterative and interactive nature of real-world academic peer review, offering a robust foundation for future research and development in this area.", "keywords": [ "peer review", "large language models" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 2, 1, 4 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "uUsfvsrkOw", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Bohao Qu", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "PhD student" }, { "name": "Di Lin", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Menglin Zhang", "gender": "Female", "institution": "Jilin University", "country": "CN", "position": "PhD student" }, { "name": "Qing Guo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Shuifa Sun", "gender": "Male", "institution": "Hangzhou Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoxin Guo", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Full Professor" }, { "name": "Ivor W Tsang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Xiaofeng Cao", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 24, "n_formula_1": 25, "n_ref_uni": 46, "n_ref": 63, "n_ref_all": 69, "n_fig": 3, "n_tab": 3, "L_tab": 884, "n_element_tab": 127, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 893, "n_element_tab_1": 134, "formula_len_all": 1296, "formula_len_all_1": 1297, "len_all": 111474, "len_all_1": 57674, "len_abs": 1816, "len_title": 97, "len_sents": 29913, "len_sents_1": 27025, "n_sents": 175, "n_sents_1": 149, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1820, "title": "CRVR: Continuous Representation-Driven Video Frame Modulation Against rPPG Heart Rate Measurement", "abs": "Facial video-based remote physiological measurement (rPPG) has gained prominence for its ability to non-invasively estimate vital signs such as heart rate (HR).The foundation of rPPG lies in using a camera to record facial videos at a certain frame rate, allowing the capture of rapid skin color changes necessary for HR measurement. Inspired by this property, we identified a new task, that is, to embed malicious information into facial videos by subtly modulating frames and generating frames corresponding to the modified rate. With this task, we can mislead state-of-the-art rPPG HR methods through natural and imperceptible frame modulation changes, aiming for two objectives: testing the resilience of rPPG methods against frame modulation variations and safeguarding heart rate data, which is crucial for individual privacy. However, such a task is non-trivial and should be capable of automatically adapting to different input videos and generating natural, imperceptible frame modulation perturbations along with frames corresponding to the modified rate. To address these challenges, we propose Continuous Representation-driven Video Resampling (CRVR), which targets precise manipulation of frame timing to subtly skew perceived HR measurements. Specifically, the CRVR method consists of two modules: Variable Frame Rate Video Resampling (VFRVR), which automatically determines the optimal resampling strategy for each frame, and Continuous Video Frame Generation (CVFG), which generates frames corresponding to the modified rate and seamlessly injects them back into the video. Extensive testing on UBFC-rPPG and PURE datasets reveals that our CRVR method successfully produces realistic, imperceptible adversarial videos that effectively mislead three different rPPG-based heart rate detection technologies.", "keywords": [ "Adversarial video attack", "Remote physiological heart rate measurement." ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 5, 5, 3 ] }, { "paper_id": "uUsL07BsMA", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Ning Ge", "gender": "Female", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Shumao Zhai", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 71, "n_ref_all": 83, "n_fig": 8, "n_tab": 6, "L_tab": 2955, "n_element_tab": 235, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 683, "n_element_tab_1": 155, "formula_len_all": 162, "formula_len_all_1": 111, "len_all": 176056, "len_all_1": 75424, "len_abs": 1405, "len_title": 139, "len_sents": 66690, "len_sents_1": 38062, "n_sents": 479, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1421, "title": "Learning Splitting Heuristics in Divide-and-Conquer SAT Solvers with Reinforcement Learning", "abs": "We propose RDC-SAT, a novel approach to optimize splitting heuristics in Divide-and-Conquer SAT solvers using deep reinforcement learning. Our method dynamically extracts features from the current solving state whenever a split is required. These features, such as learned clauses, variable activity scores, and clause LBD (Literal Block Distance) values, are represented as a graph. A GNN integrated with an Actor-Critic model processes this graph to determine the optimal split variable. Unlike traditional linear state transitions characterized by Markov processes, divide-and-conquer challenges involve tree-like state transitions. To address this, we developed a reinforcement learning environment based on the Painless framework that efficiently handles these transitions. Additionally, we designed different discounted reward functions for satisfiable and unsatisfiable SAT problems, capable of handling tree-like state transitions. We trained our model using the Decentralized Proximal Policy Optimization (DPPO) algorithm on phase transition random 3-SAT problems and implemented the RDC-SAT solver, which operates in both GPU-accelerated and non-GPU modes. Evaluations show that RDC-SAT significantly improves the performance of D\\&C solvers on phase transition random 3-SAT datasets and generalizes well to the SAT Competition 2023 dataset, substantially outperforming traditional splitting heuristics.", "keywords": [ "SAT Problem", "Divide And Conquer", "Graph Neural Network", "Reinforcememt Learning" ], "rating_list": [ 8, 6, 3, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "uUkpYafkVl", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Hongxin Wei", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Jianguo Huang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Intern" }, { "name": "Kangdao Liu", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "PhD student" }, { "name": "Xinyu Cai", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" }, { "name": "Yuzhou Cao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 39, "n_formula_1": 14, "n_ref_uni": 50, "n_ref": 92, "n_ref_all": 107, "n_fig": 6, "n_tab": 20, "L_tab": 15228, "n_element_tab": 912, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 6387, "n_element_tab_1": 347, "formula_len_all": 3384, "formula_len_all_1": 1124, "len_all": 190856, "len_all_1": 67745, "len_abs": 1285, "len_title": 103, "len_sents": 52526, "len_sents_1": 26974, "n_sents": 420, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1279, "title": "Conformal Prediction for Deep Classifier via Truncating", "abs": "Conformal Prediction is a distribution-free statistical framework that outputs a set of possible labels to capture the predictive uncertainty. In this work, we show that existing conformal prediction methods may generate inefficient sets arising from the inclusion of redundant labels. To mitigate this issue, we propose a novel conformal prediction algorithm, $\\textit{Post-Calibration Truncated Conformal Prediction}$ (PoT-CP), which limits the size of the prediction sets generated by existing conformal prediction methods through a maximum rank cutoff. Specifically, PoT-CP determines this cutoff by minimizing a truncation rank that preserves the marginal coverage of the calibration dataset. The key idea is to eliminate classes with high predictive uncertainty in the prediction sets, allowing PoT-CP to further shorten the prediction sets. Theoretically, we provide the asymptotic validity of marginal coverage for PoT-CP and demonstrate the asymptotic conditional coverage equivalence between PoT-CP and the standard conformal prediction algorithm. Extensive experiments demonstrate that PoT-CP can effectively reduce prediction set sizes while maintaining the stable conditional coverage of various conformal prediction algorithms across different classification tasks.", "keywords": [ "Conformal Prediction", "Uncertainty Quantification" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "uTqnyF0JNR", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Hao Peng", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Haonan Yuan", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Jianxin Li", "gender": "Male", "institution": "Beihang University ", "country": "CN", "position": "Full Professor" }, { "name": "Jiaqi Yuan", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Lyujin Xu", "gender": "Not Specified", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Pengfeng Huang", "gender": "Male", "institution": "The Insititute of Advanced Computing Technology, Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Philip S. Yu", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "US", "position": "Full Professor" }, { "name": "Qingyun Sun", "gender": "Female", "institution": "Beihang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xingcheng Fu", "gender": "Male", "institution": "Guangxi Normal University", "country": "CN", "position": "Assistant Professor" }, { "name": "M Q", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "ZHAONAN WANG", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" } ], "n_formula": 28, "n_formula_1": 8, "n_ref_uni": 55, "n_ref": 188, "n_ref_all": 225, "n_fig": 7, "n_tab": 40, "L_tab": 60225, "n_element_tab": 3284, "n_fig_1": 6, "n_tab_1": 10, "L_tab_1": 4882, "n_element_tab_1": 243, "formula_len_all": 3309, "formula_len_all_1": 753, "len_all": 422055, "len_all_1": 75175, "len_abs": 1608, "len_title": 129, "len_sents": 101079, "len_sents_1": 30439, "n_sents": 951, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 81, "L_abs": 1646, "title": "IGL-Bench: Establishing the Comprehensive Benchmark for Imbalanced Graph Learning", "abs": "Deep graph learning has gained grand popularity over the past years due to its versatility and success in representing graph data across a wide range of domains. However, the pervasive issue of imbalanced graph data distributions, where certain parts exhibit disproportionally abundant data while others remain sparse, undermines the efficacy of conventional graph learning algorithms, leading to biased outcomes. To address this challenge, Imbalanced Graph Learning (IGL) has garnered substantial attention, enabling more balanced data distributions and better task performance. Despite the proliferation of IGL algorithms, the absence of consistent experimental protocols and fair performance comparisons pose a significant barrier to comprehending advancements in this field. To bridge this gap, we introduce **IGL-Bench**, a foundational comprehensive benchmark for imbalanced graph learning, embarking on **17** diverse graph datasets and **24** distinct IGL algorithms with uniform data processing and splitting strategies. Specifically, IGL-Bench systematically investigates state-of-the-art IGL algorithms in terms of **effectiveness**, **robustness**, and **efficiency** on node-level and graph-level tasks, with the scope of class-imbalance and topology-imbalance. Extensive experiments demonstrate the potential benefits of IGL algorithms on various imbalanced conditions, offering insights and opportunities in the IGL field. Further, we have developed an open-sourced and unified package to facilitate reproducible evaluation and inspire further innovative research, which is available at https://anonymous.4open.science/r/IGL-Bench.", "keywords": [ "imbalanced graph learning", "graph class-imbalance", "graph topology-imbalance", "comprehensive benchmark" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "uSz2K30RRd", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Chieh-Hsin Lai", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Researcher" }, { "name": "Naoki Murata", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Researcher" }, { "name": "Taiji Suzuki", "gender": "Male", "institution": "The University of Tokyo", "country": "", "position": "Full Professor" }, { "name": "Toshimitsu Uesaka", "gender": "Male", "institution": "Sony AI", "country": "", "position": "Researcher" }, { "name": "Yuhta Takida", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Research Scientist" }, { "name": "Yuki Mitsufuji", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Lead Research Scientist" } ], "n_formula": 49, "n_formula_1": 19, "n_ref_uni": 57, "n_ref": 97, "n_ref_all": 110, "n_fig": 2, "n_tab": 7, "L_tab": 3019, "n_element_tab": 205, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5808, "formula_len_all_1": 1428, "len_all": 179345, "len_all_1": 67990, "len_abs": 1141, "len_title": 148, "len_sents": 45048, "len_sents_1": 29827, "n_sents": 407, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 99, "L_abs": 1144, "title": "Weighted Point Cloud Embedding for Multimodal Contrastive Learning Toward Optimal Similarity Metric", "abs": "In typical multimodal contrastive learning, such as CLIP, encoders produce onepoint in the latent representation space for each input. However, one-point representation has difficulty in capturing the relationship and the similarity structure of a huge amount of instances in the real world. For richer classes of the similarity, we propose the use of weighted point clouds, namely, sets of pairs of weight and vector, as representations of instances. In this work, we theoretically show the benefit of our proposed method through a new understanding of the contrastive loss of CLIP, which we call symmetric InfoNCE. We clarify that the optimal similarity\nthat minimizes symmetric InfoNCE is the pointwise mutual information, and show an upper bound of excess risk on downstream classification tasks of representations that achieve the optimal similarity. In addition, we show that our proposed similarity based on weighted point clouds consistently achieves the optimal similarity. To verify the effectiveness of our proposed method, we demonstrate pretraining of text-image representation models and classification tasks on common benchmarks.", "keywords": [ "contrastive learning", "multimodal representation learning", "theoretical analysis", "InfoNCE", "pointwise mutual information" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "uSiyu6CLPh", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Ali Edalati", "gender": "unknown", "institution": "Cohere", "country": "CA", "position": "Researcher" }, { "name": "Brett H. Meyer", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Associate Professor" }, { "name": "James J. Clark", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Full Professor" }, { "name": "Lulan Shen", "gender": "Female", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "Shilei Lin", "gender": "Male", "institution": "McGill University, McGill University", "country": "CA", "position": "MS student" }, { "name": "Warren J. Gross", "gender": "unknown", "institution": "McGill University", "country": "", "position": "Professor" }, { "name": "Xiangyu Li", "gender": "Male", "institution": "McGill University, McGill University", "country": "CA", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 18, "n_ref": 35, "n_ref_all": 45, "n_fig": 4, "n_tab": 3, "L_tab": 3024, "n_element_tab": 401, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3092, "n_element_tab_1": 403, "formula_len_all": 627, "formula_len_all_1": 627, "len_all": 100356, "len_all_1": 70686, "len_abs": 966, "len_title": 128, "len_sents": 30772, "len_sents_1": 30487, "n_sents": 226, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 971, "title": "AdCorDA: Classifier Refinement via Adversarial Correction and Domain Adaptation", "abs": "This paper describes a simple yet effective technique for refining a pretrained classifier network. The proposed AdCorDA method consists of two stages - adversarial correction followed by domain adaptation. Adversarial correction uses adversarial attacks to correct misclassified training-set classifications. The incorrectly classified samples of the training set are removed and replaced with the adversarially corrected samples to form a new training set, and then, in the second stage, domain adaptation is performed back to the original training set. Extensive experimental validations show significant accuracy boosts of over 5% on the CIFAR-100 dataset and 1% on the CINIC-10 dataset. The technique can be straightforwardly applied to the refinement of weight-quantized neural networks, where experiments show substantial enhancement in performance over the baseline. The adversarial correction technique also results in enhanced robustness to adversarial attacks.", "keywords": [ "adversarial correction", "domain adaptation", "curriculum learning", "adversarial attacks" ], "rating_list": [ 6, 3, 6, 3 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 4, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "uSg854MOWu", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Binghui Li", "gender": "Not Specified", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yuanzhi Li", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 98, "n_formula_1": 21, "n_ref_uni": 47, "n_ref": 103, "n_ref_all": 116, "n_fig": 5, "n_tab": 4, "L_tab": 312, "n_element_tab": 37, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 227, "n_element_tab_1": 4, "formula_len_all": 12543, "formula_len_all_1": 1878, "len_all": 188843, "len_all_1": 67438, "len_abs": 1393, "len_title": 202, "len_sents": 53250, "len_sents_1": 28922, "n_sents": 511, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 154, "L_abs": 1425, "title": "Understand Clean Generalization and Robust Overfitting in Adversarial Training from Two Theoretical Views: Representation Complexity and Training Dynamics", "abs": "Similar to surprising performance in the standard deep learning, deep nets trained by adversarial training also generalize well for unseen clean data (natural data). However, despite adversarial training can achieve low robust training error, there exists a significant robust generalization gap. We call this phenomenon the Clean Generalization and Robust Overfitting (CGRO). In this work, we study the CGRO phenomenon in adversarial training from two views: representation complexity and training dynamics. Specifically, we consider a binary classification setting with $N$ separated training data points. First, we prove that, based on the assumption that we assume there is $\\operatorname{poly}(D)$-size clean classifier (where $D$ is the data dimension), ReLU net with only $O(N D)$ extra parameters is able to leverages robust memorization to achieve the CGRO, while robust classifier still requires exponential representation complexity in worst case. Next, we focus on a structured-data case to analyze training dynamics, where we train a two-layer convolutional network with $O(N D)$ width against adversarial perturbation. We then show that a three-stage phase transition occurs during learning process and the network provably converges to robust memorization regime, which thereby results in the CGRO. Besides, we also empirically verify our theoretical analysis by experiments in real-image recognition datasets.", "keywords": [ "deep learning theory", "adversarial training", "clean generalization and robust overfitting", "representation complexity", "training dynamics", "feature learning theory" ], "rating_list": [ 5, 6, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 3, 2, 4, 3, 4 ] }, { "paper_id": "uSV07DapJx", "primary_area": "causal reasoning", "n_author": 8, "author_info_list": [ { "name": "Bo Long", "gender": "Male", "institution": "Meta", "country": "US", "position": "Principal Researcher" }, { "name": "Chen Fu", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Juncheng Dong", "gender": "unknown", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Yiling Liu", "gender": "Female", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Zhigang Hua", "gender": "Male", "institution": "Meta", "country": "US", "position": "ML engineer" }, { "name": "Ziyang Jiang", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Research Scientist" }, { "name": "David Edwin Carlson", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Associate Professor" }, { "name": "WEI SHI", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Researcher" } ], "n_formula": 56, "n_formula_1": 15, "n_ref_uni": 34, "n_ref": 110, "n_ref_all": 127, "n_fig": 4, "n_tab": 8, "L_tab": 3338, "n_element_tab": 193, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1815, "n_element_tab_1": 99, "formula_len_all": 4232, "formula_len_all_1": 835, "len_all": 167020, "len_all_1": 65995, "len_abs": 3531, "len_title": 158, "len_sents": 55181, "len_sents_1": 29541, "n_sents": 450, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 110, "L_abs": 1670, "title": "Counterfactual Outcome Estimation in Time Series via Sub-treatment Group Alignment and Random Temporal Masking", "abs": "Estimating counterfactual outcomes in time series from observational data is important for effective decision-making in many fields, such as determining the optimal timing for a medical intervention. However, this task is challenging, primarily because of the unobservability of counterfactual outcomes and the complexity of confounding in time series. To this end, we introduce a representation learning-based framework for counterfactual estimation in time series with two novel techniques: **Sub-treatment Group Alignment (SGA)** and **Random Temporal Masking (RTM)**. The first technique focuses on reducing confounding at each time point. While the common approach is to align the distributions of different treatment groups in the latent space, our proposed approach, SGA, first identifies *sub-treatment groups* through Gaussian Mixture Models (GMMs) and subsequently aligns the corresponding sub-groups. We demonstrate that, both theoretically and empirically, SGA achieves improved alignment, thus leading to more effective deconfounding. The second technique, RTM, masks covariates at random time steps with Gaussian noises. This approach promotes the time series models to select information not only important for the outcome estimation at current time point but also crucial for the time points in the future where the covariates are masked out, thus preserving the *causal information* and reducing the risk of overfitting to factual outcomes. We observe in experiments on synthetic and semi-synthetic datasets that applying SGA and RTM individually improves counterfactual outcome estimation, and when combined, they achieve state-of-the-art performance.", "keywords": [ "Counterfactual treatment effect estimation", "Time series observational data", "Confounding in time series", "Sub-treatment Group Alignment", "Random Temporal Masking" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "uREg3OHjLL", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Christopher Hojny", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "Gennadiy Averkov", "gender": "Male", "institution": "BTU Cottbus-Senftenberg", "country": "DE", "position": "Full Professor" }, { "name": "Maximilian Merkert", "gender": "Male", "institution": "Technische Universit\u00e4t Carolo-Wilhelmina zu Braunschweig", "country": "DE", "position": "Assistant Professor" } ], "n_formula": 26, "n_formula_1": 17, "n_ref_uni": 23, "n_ref": 61, "n_ref_all": 72, "n_fig": 2, "n_tab": 3, "L_tab": 449, "n_element_tab": 5, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 745, "n_element_tab_1": 12, "formula_len_all": 2151, "formula_len_all_1": 1443, "len_all": 134320, "len_all_1": 85297, "len_abs": 971, "len_title": 121, "len_sents": 43107, "len_sents_1": 33447, "n_sents": 418, "n_sents_1": 314, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 73, "L_abs": 1041, "title": "On the Expressiveness of Rational ReLU Neural Networks With Bounded Depth", "abs": "To confirm that the expressive power of ReLU neural networks grows with their depth, the function $F_n = \\max (0,x_1,\\ldots,x_n )$ has been considered in the literature.\n A conjecture by Hertrich, Basu, Di Summa, and Skutella [NeurIPS 2021] states that any ReLU network that exactly represents $F_n$ has at least $\\lceil \\log_2 (n+1) \\rceil$ hidden layers.\n The conjecture has recently been confirmed for networks with integer weights by Haase, Hertrich, and Loho [ICLR 2023].\n\n We follow up on this line of research and show that, within ReLU networks whose weights are decimal fractions, $F_n$ can only be represented by networks with at least $\\lceil \\log_3 (n+1) \\rceil$ hidden layers.\n Moreover, if all weights are $N$-ary fractions, then $F_n$ can only be represented by networks with at least $\\Omega( \\frac{\\ln n}{\\ln \\ln N})$ layers.\n These results are a partial confirmation of the above conjecture for rational ReLU networks, and provide the first non-constant lower bound on the depth of practically relevant ReLU networks.", "keywords": [ "expressive power", "depth", "exact representations", "ReLU networks", "mixed volumes", "lattice polytopes", "number theory" ], "rating_list": [ 5, 8, 8, 8, 8 ], "soundness_list": [ 2, 4, 4, 3, 4 ], "presentation_list": [ 4, 4, 4, 3, 4 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "uQnvYP7yX9", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Bozhen Hu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Cheng Tan", "gender": "Male", "institution": "Zhejiang University & Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Jingbo Zhou", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Jun Xia", "gender": "Male", "institution": "Westlake University, China", "country": "CN", "position": "PhD student" }, { "name": "Lecheng Zhang", "gender": "Male", "institution": "Westlake University", "country": "", "position": "Undergrad student" }, { "name": "Shaorong Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Wenjie Du", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhangyang Gao", "gender": "Male", "institution": "Westlake University, China", "country": "CN", "position": "PhD student" }, { "name": "Stan Z. Li", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Chair Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 46, "n_ref_all": 77, "n_fig": 7, "n_tab": 8, "L_tab": 1845, "n_element_tab": 237, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 657, "n_element_tab_1": 109, "formula_len_all": 882, "formula_len_all_1": 685, "len_all": 120592, "len_all_1": 53818, "len_abs": 1890, "len_title": 117, "len_sents": 38788, "len_sents_1": 24871, "n_sents": 279, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1947, "title": "ReNovo: Retrieval-Based \\emph{De Novo} Mass Spectrometry Peptide Sequencing", "abs": "Proteomics is the large-scale study of proteins. Tandem mass spectrometry, as the only high-throughput technique for protein sequence identification, plays a pivotal role in proteomics research. One of the long-standing challenges in this field is peptide identification, which entails determining the specific peptide (sequence of amino acids) that corresponds to each observed mass spectrum. The conventional approach involves database searching, wherein the observed mass spectrum is scored against a pre-constructed peptide database. However, the reliance on pre-existing databases limits applicability in scenarios where the peptide is absent from existing databases. Such circumstances necessitate \\emph{de novo} peptide sequencing, which derives peptide sequence solely from input mass spectrum, independent of any peptide database. Despite ongoing advancements in \\emph{de novo} peptide sequencing, its performance still has considerable room for improvement, which limits its application in large-scale experiments. In this study, we introduce a novel \\textbf{Re}trieval-based \\emph{De \\textbf{Novo}} peptide sequencing methodology, termed \\textbf{ReNovo}, which draws inspiration from database search methods. Specifically, by constructing a datastore from training data, ReNovo can retrieve information from the datastore during the inference stage to conduct retrieval-based inference, thereby achieving improved performance. This innovative approach enables ReNovo to effectively combine the strengths of both methods: utilizing the assistance of the datastore while also being capable of predicting novel peptides that are not present in pre-existing databases. A series of experiments have confirmed that ReNovo outperforms state-of-the-art models across multiple widely-used datasets, incurring only minor storage and time consumption, representing a significant advancement in proteomics. Supplementary materials include the code.", "keywords": [ "Peptide Sequencing" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "uQjySppU9x", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "David B. Lindell", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Assistant Professor" }, { "name": "Igor Gilitschenski", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Assistant Professor" }, { "name": "Koichi Namekata", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" }, { "name": "Sherwin Bahmani", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Yash Kant", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Ziyi Wu", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 60, "n_ref": 174, "n_ref_all": 207, "n_fig": 15, "n_tab": 2, "L_tab": 973, "n_element_tab": 105, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 870, "n_element_tab_1": 95, "formula_len_all": 211, "formula_len_all_1": 185, "len_all": 188091, "len_all_1": 65061, "len_abs": 371, "len_title": 115, "len_sents": 44995, "len_sents_1": 29577, "n_sents": 364, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1181, "title": "SG-I2V: Self-Guided Trajectory Control in Image-to-Video Generation", "abs": "Methods for image-to-video generation have achieved impressive, photo-realistic quality. \nHowever, adjusting specific elements in generated videos, such as object motion or camera movement, is often a tedious process of trial and error, e.g., involving re-generating videos with different random seeds. \nRecent techniques address this issue by fine-tuning a pre-trained model to follow conditioning signals, such as bounding boxes or point trajectories. \nYet, this fine-tuning procedure can be computationally expensive, and it requires datasets with annotated object motion, which can be difficult to procure. \nIn this work, we introduce SG-I2V, a framework for controllable image-to-video generation that is self-guided\u2014ioffering zero-shot control by relying solely on the knowledge present in a pre-trained image-to-video diffusion model without the need for fine-tuning or external knowledge. \nOur zero-shot method outperforms unsupervised baselines while significantly narrowing down the performance gap with supervised models in terms of visual quality and motion fidelity.\nAdditional details and video results are available on our project page: https://sgi2v-paper.github.io", "keywords": [ "zero-shot", "tuning-free", "self-guided", "image-to-video diffusion", "trajectory control" ], "rating_list": [ 6, 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 4, 5, 5, 4, 5 ] }, { "paper_id": "uQRQo0cWZ6", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Weiwei Liu", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Yanbo Chen", "gender": "Male", "institution": "Wuhan University", "country": "", "position": "PhD student" }, { "name": "Cao Yang", "gender": "Male", "institution": "Wuhan University", "country": "", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 19, "n_ref": 61, "n_ref_all": 86, "n_fig": 5, "n_tab": 4, "L_tab": 1331, "n_element_tab": 107, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 440, "n_element_tab_1": 87, "formula_len_all": 268, "formula_len_all_1": 224, "len_all": 112221, "len_all_1": 61947, "len_abs": 944, "len_title": 146, "len_sents": 34703, "len_sents_1": 28524, "n_sents": 267, "n_sents_1": 234, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 894, "title": "Shake-It-Off: Jailbreaking Black-Box Large Language Models by Shaking Off Objectionable Semantics", "abs": "Large language models (LLMs) are vulnerable to jailbreaking attacks (Zou et al., 2023; Liu et al., 2024), in which attackers use adversarially designed prompts to bypass the model\u2019s safeguard and force the model to generate objectionable content. The present paper studies jailbreaking attacks from a red team\u2019s viewpoint and proposes a novel black-box attack method, called Shake-It-Off (SHAKE), that only requires the response generated by the victim model. Given objective query $T_{obj}$, our method iteratively shakes off the objectionable semantics of $T_{obj}$, making it gradually approximates a pre-defied decontaminated query $T_{dec}$. We conduct extensive experiments on multiple baseline methods and victim LLMs. The experimental results show that SHAKE outperforms the baseline methods in attack success rates while requiring much less running time and access to the victim model.", "keywords": [ "Jailbreaking Attacks", "Large Language Models" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "uQEsLZU15E", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Jiaqi Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Nenghai Yu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Pan Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Qidong Huang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Weiming Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoyi Dong", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yuhang CAO", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yuhang Zang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 28, "n_ref": 44, "n_ref_all": 68, "n_fig": 4, "n_tab": 7, "L_tab": 2456, "n_element_tab": 143, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3362, "n_element_tab_1": 185, "formula_len_all": 536, "formula_len_all_1": 574, "len_all": 149866, "len_all_1": 62411, "len_abs": 1619, "len_title": 145, "len_sents": 39646, "len_sents_1": 28034, "n_sents": 251, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 96, "L_abs": 1633, "title": "Deciphering Cross-Modal Alignment in Large Vision-Language Models with Modality Integration Rate", "abs": "We present the Modality Integration Rate (MIR), an effective, robust, and generalized metric to indicate the multi-modal pre-training quality of Large Vision Language Models (LVLMs). Large-scale pre-training plays a critical role in building capable LVLMs, while evaluating its training quality without the costly supervised fine-tuning stage is under-explored. Loss, perplexity, and in-context evaluation results are commonly used pre-training metrics for Large Language Models (LLMs), while we observed that these metrics are less indicative when aligning a well-trained LLM with a new modality. Due to the lack of proper metrics, the research of LVLMs in the critical pre-training stage is hindered greatly, including the training data choice, efficient module design, etc.\n\nIn this paper, we propose evaluating the pre-training quality from the inter-modal distribution distance perspective and present MIR, the Modality Integration Rate, which is 1) Effective to represent the pre-training quality and show a positive relation with the benchmark performance after supervised fine-tuning. 2) Robust toward different training/evaluation data. 3) Generalize across training configurations and architecture choices.\n\nWe conducted a series of pre-training experiments to explore the effectiveness of MIR and observed satisfactory results that MIR is indicative about training data selection, training strategies schedule, and model architecture design to get better pre-training results. \nWe hope MIR could be a helpful metric for building capable LVLMs and inspire the following research about modality alignment in different areas.", "keywords": [ "Large Vision-Language Models", "Cross-Modal Alignment" ], "rating_list": [ 5, 5, 3, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "uPwe2w78Wx", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Ao Luo", "gender": "Not Specified", "institution": "KDDI Research, Inc.", "country": "JP", "position": "Researcher" }, { "name": "Chi Wang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Jiale Liu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Jieyu Zhang", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Linxin Song", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Shaokun Zhang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Shijian Wang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "Qingyun Wu", "gender": "Female", "institution": "Pennsylvania State University", "country": "", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 35, "n_ref": 84, "n_ref_all": 105, "n_fig": 3, "n_tab": 7, "L_tab": 3505, "n_element_tab": 343, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2234, "n_element_tab_1": 260, "formula_len_all": 140, "formula_len_all_1": 136, "len_all": 197949, "len_all_1": 59777, "len_abs": 1263, "len_title": 0, "len_sents": 41668, "len_sents_1": 26521, "n_sents": 306, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1267, "title": "Adaptive In-conversation Team Building for Language Model Agents", "abs": "Leveraging multiple large language model (LLM) agents has shown to be a promising approach for tackling complex tasks, while the effective design of multiple agents for a particular application remains an art. It is thus intriguing to answer a critical question: Given a task, how can we build a team of LLM agents to solve it effectively? Our new adaptive team-building paradigm offers a flexible solution, realized through a novel agent design named Captain Agent. It dynamically forms and manages teams for each step of a task-solving process, utilizing nested group conversations and reflection to ensure diverse expertise and prevent stereotypical outputs, allowing for a flexible yet structured approach to problem-solving. A comprehensive evaluation across six real-world scenarios demonstrates that Captain Agent significantly outperforms existing multi-agent methods with 21.94% improvement in average accuracy, providing outstanding performance without requiring task-specific prompt engineering. Our exploration of different backbone LLM and cost analysis further shows that Captain Agent can improve the conversation quality of weak LLM and achieve competitive performance with extremely low cost, which illuminates the application of multi-agent systems.", "keywords": [ "Natural Language Processing", "Large Language Model", "Agent", "Multi-agent" ], "rating_list": [ 5, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "uPj9oBH80V", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Eduardo Soares", "gender": "Male", "institution": "International Business Machines", "country": "BR", "position": "Researcher" }, { "name": "Emilio Vital Brazil", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Hajime Shinohara", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Indra Priyadarsini", "gender": "Female", "institution": "IBM Research - Tokyo, International Business Machines", "country": "", "position": "Researcher" }, { "name": "Lisa Hamada", "gender": "Female", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Seiji Takeda", "gender": "Male", "institution": "IBM Research - Tokyo, International Business Machines", "country": "", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 16, "n_ref": 25, "n_ref_all": 39, "n_fig": 5, "n_tab": 5, "L_tab": 2280, "n_element_tab": 275, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1177, "n_element_tab_1": 121, "formula_len_all": 124, "formula_len_all_1": 124, "len_all": 95014, "len_all_1": 52461, "len_abs": 1139, "len_title": 131, "len_sents": 26720, "len_sents_1": 25630, "n_sents": 170, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1144, "title": "SELFIES-TED : A Robust Transformer Model for Molecular Representation using SELFIES", "abs": "Large-scale molecular representation methods have revolutionized applications in material science, such as drug discovery, chemical modeling, and material design. With the rise of transformers, models now learn representations directly from molecular structures. In this paper, we introduce SELFIES-TED, a transformer-based model designed for molecular representation using SELFIES, a more robust, unambiguous method for encoding molecules compared to traditional SMILES strings. By leveraging the robustness of SELFIES and the power of the transformer encoder-decoder architecture, SELFIES-TED effectively captures the intricate relationships between molecular structures and their properties. Having pretrained with 1 billion molecule samples, our model demonstrates improved performance on molecular property prediction tasks across various benchmarks, showcasing its generalizability and robustness. \nAdditionally, we explore the latent space of SELFIES-TED, revealing valuable insights that enhance its capabilities in both molecule property prediction and molecule generation tasks, opening new avenues for innovation in molecular design.", "keywords": [ "molecular representation", "property prediction", "molecular generation" ], "rating_list": [ 1, 6, 1, 6 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 1, 2, 1, 3 ], "confidence_list": [ 5, 2, 4, 4 ] }, { "paper_id": "uOxoje4Sa9", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Chen Liang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Huanzhang Dou", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" }, { "name": "Jingren Zhou", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Lianghua Huang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Wei Wang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yu Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yupeng Shi", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yutong Feng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zhi-Fan Wu", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 32, "n_ref": 41, "n_ref_all": 57, "n_fig": 7, "n_tab": 3, "L_tab": 935, "n_element_tab": 228, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 543, "n_element_tab_1": 117, "formula_len_all": 316, "formula_len_all_1": 316, "len_all": 165631, "len_all_1": 59400, "len_abs": 1748, "len_title": 112, "len_sents": 31198, "len_sents_1": 27893, "n_sents": 226, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1792, "title": "Group Diffusion Transformers are Unsupervised Multitask Learners", "abs": "While large language models (LLMs) have revolutionized natural language processing with their task-agnostic capabilities, visual generation tasks such as image translation, style transfer, and character customization still rely heavily on supervised, task-specific datasets. In this work, we introduce \\textbf{Group Diffusion Transformers (GDTs)}, a novel framework that unifies diverse visual generation tasks by redefining them as a \\textbf{group generation} problem. In this approach, a set of related images is generated simultaneously, optionally conditioned on a subset of the group. GDTs build upon diffusion transformers with minimal architectural modifications by concatenating self-attention tokens across images. This allows the model to implicitly capture cross-image relationships (\\textit{e.g.}, identities, styles, layouts, surroundings, textures, and color schemes) through caption-based correlations. Our design enables scalable, unsupervised, and task-agnostic pretraining using extensive collections of image groups sourced from multimodal internet articles, image galleries, and video frames. We evaluate GDTs on a comprehensive benchmark featuring over 200 instructions across 30 distinct visual generation tasks, including picture book creation, font design, style transfer, sketching, colorization, drawing sequence generation, and character customization. Our models achieve competitive \\textbf{zero-shot} performance without any additional fine-tuning or gradient updates. Furthermore, ablation studies confirm the effectiveness of key components such as data scaling, group size, and model design. These results demonstrate the potential of GDTs as scalable, general-purpose visual generation systems. We will release the code and models to support further research.", "keywords": [ "diffusion transformers", "unsupervised pretraining", "group generation", "task-agnostic" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "uOrfve3prk", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Asma Ghandeharioun", "gender": "unknown", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Suraj Srinivas", "gender": "Male", "institution": "Robert Bosch LLC", "country": "US", "position": "Researcher" }, { "name": "Usha Bhalla", "gender": "Female", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Hima Lakkaraju", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 10, "n_ref": 69, "n_ref_all": 91, "n_fig": 9, "n_tab": 9, "L_tab": 3976, "n_element_tab": 219, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 90, "n_element_tab_1": 13, "formula_len_all": 142, "formula_len_all_1": 79, "len_all": 148365, "len_all_1": 62261, "len_abs": 1988, "len_title": 123, "len_sents": 50307, "len_sents_1": 33581, "n_sents": 310, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1996, "title": "Towards Unifying Interpretability and Control: Evaluation via Intervention", "abs": "With the growing complexity and capability of large language models (LLMs), a need to understand model reasoning has emerged, often motivated by an underlying goal of controlling and aligning models. While numerous interpretability and steering methods have been proposed as solutions, they are typically designed either for understanding or for control, seldom addressing both, with the connection between interpretation and control more broadly remaining tenuous. Additionally, the lack of standardized applications, motivations, and evaluation metrics makes it difficult to assess these methods' practical utility and efficacy. To address the aforementioned issues, we propose intervention as a fundamental goal of interpretability and introduce success criteria to evaluate how well methods are able to control model behavior through interventions. We unify and extend four popular interpretability methods\u2014sparse autoencoders, logit lens, tuned lens, and probing\u2014into an abstract encoder-decoder framework. This framework maps intermediate latent representations to human-interpretable feature spaces, enabling interventions on these interpretable features, which can then be mapped back to latent representations to control model outputs. We introduce two new evaluation metrics: intervention success rate and the coherence-intervention tradeoff, designed to measure the accuracy of explanations and their utility in controlling model behavior. Our findings reveal that (1) although current methods allow for intervention, they are inconsistent across various models and features, (2) lens-based methods outperform others in achieving simple, concrete interventions, and (3) interventions often compromise model performance and coherence, underperforming simpler alternatives, such as prompting, for steering model behavior and highlighting a critical shortcoming of current interpretability approaches in real-world applications requiring control. Code is made available for replicability.", "keywords": [ "machine learning interpretability", "mechanistic interpretability" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "uOnElfFuey", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Haiming Wang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Han Shi", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Principal Researcher" }, { "name": "Yimeng Chen", "gender": "unknown", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" }, { "name": "Zhenguo Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Zhengying Liu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 43, "n_ref": 104, "n_ref_all": 127, "n_fig": 9, "n_tab": 9, "L_tab": 2451, "n_element_tab": 363, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1737, "n_element_tab_1": 323, "formula_len_all": 981, "formula_len_all_1": 981, "len_all": 149831, "len_all_1": 65038, "len_abs": 1027, "len_title": 98, "len_sents": 39945, "len_sents_1": 26252, "n_sents": 330, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1032, "title": "Recovering Knowledge by Hardening Language Models", "abs": "Recent neural language models show impressive capabilities on a wide range of tasks. However, it is not fully understood how the knowledge of the language is encoded in these models. In this work, we focus on the simplest case of languages, regular languages, and study language models trained on strings matching certain regular expressions. We propose a method, dubbed LaMFA, to recover the full knowledge of the regular language model by hardening it into a finite automaton. Such hardening is conducted by empirically partition the latent space of language models into finite states, and then recover a deterministic finite automaton by the estimated transition probabilities between these states. Through experiments on regular languages of varying complexity, we demonstrate that LaMFA can effectively extract DFA that consistently replicate the performance of the original language model. Notably, the extracted DFAs exhibit enhanced generalization capabilities, achieving 100\\% accuracy even in out-of-distribution scenarios", "keywords": [ "regular language", "language model", "transformers", "knowledge interpretation" ], "rating_list": [ 3, 1, 3, 5 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "uOb7rij7sR", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Kai Chen", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Qiwei Ye", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Principal Researcher" }, { "name": "Xi Chen", "gender": "unknown", "institution": "BAAI", "country": "", "position": "Researcher" }, { "name": "Yunfei Teng", "gender": "Unspecified", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Zhaoming Chen", "gender": "unknown", "institution": "beijing academy of artificial intelligence", "country": "CN", "position": "Researcher" }, { "name": "Yuxuan Ren ustc", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" } ], "n_formula": 20, "n_formula_1": 20, "n_ref_uni": 23, "n_ref": 42, "n_ref_all": 75, "n_fig": 16, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 284, "n_element_tab_1": 18, "formula_len_all": 1416, "formula_len_all_1": 1416, "len_all": 114613, "len_all_1": 61833, "len_abs": 1247, "len_title": 136, "len_sents": 36669, "len_sents_1": 26573, "n_sents": 296, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1227, "title": "CryoGEN: Cryogenic Electron Tomography Reconstruction via Generative Energy Nets", "abs": "Cryogenic electron tomography (Cryo-ET) is a powerful method for visualizing cellular structures in their native state, but its effectiveness is limited by anisotropic resolution caused by the missing-wedge problem, complicating the interpretation of tomograms. IsoNet, a deep learning method, addresses these challenges by iteratively reconstructing missing-wedge information and improving the signal-to-noise ratio of tomograms. However, IsoNet relies on recursively updating its predictions, which can result in training instability and potential model collapse. In this study, we present CryoGEN, an enhanced energy-based method that effectively addresses resolution anisotropy without requiring recursive subtomogram generation. Our approach is about *10*$\\mathbf\\times$ faster and offers a more stable and consistent methodology. Applying CryoGEN to various datasets, including immature HIV particles and ribosomes, demonstrates its capability to enhance structural interpretability. Moreover, CryoGEN holds significant potential for improving the functional interpretation of cellular tomograms in future high-resolution Cryo-ET studies, thereby providing substantial value and advancing progress in biological research.", "keywords": [ "CryoET", "Cryogenic Electron Tomography", "Generative Model", "Energy Model" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "uO0itv7XFa", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Fuli Feng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jianxin Ma", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Senior Engineer" }, { "name": "Junyang Lin", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Yang Zhang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Ziang Ye", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Zhen-Ru Zhang", "gender": "Female", "institution": "Alibaba Group", "country": "", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 22, "n_ref": 55, "n_ref_all": 77, "n_fig": 10, "n_tab": 1, "L_tab": 459, "n_element_tab": 102, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1068, "n_element_tab_1": 128, "formula_len_all": 461, "formula_len_all_1": 462, "len_all": 133500, "len_all_1": 66596, "len_abs": 1484, "len_title": 133, "len_sents": 42556, "len_sents_1": 30160, "n_sents": 324, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1013, "title": "Disentangling Reasoning Tokens and Boilerplate Tokens For Language Model Fine-tuning", "abs": "When using agent-task datasets to enhance agent capabilities for Large Language Models (LLMs), current methodologies often treat all tokens within a sample equally. \nHowever, we argue that tokens serving different roles\u2014specifically, reasoning tokens versus boilerplate tokens (e.g., those governing output format)\u2014differ significantly in importance and learning complexity, necessitating their disentanglement and distinct treatment. \nTo address this, we propose a novel Shuffle-Aware Discriminator (SHAD) for adaptive token discrimination.\nSHAD classifies tokens by exploiting predictability differences observed after shuffling input-output combinations across samples: boilerplate tokens, due to their repetitive nature among samples, maintain predictability, whereas reasoning tokens do not.\nUsing SHAD, we propose the Reasoning-highlighted Fine-Tuning (RFT) method, which adaptively emphasizes reasoning tokens during fine-tuning, yielding notable performance gains over common Supervised Fine-Tuning (SFT).", "keywords": [ "Large language models", "Tool Use", "Agent", "Reasoning" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "uNomADvF3s", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Johanna Sommer", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Mohamed Amine Ketata", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Nicholas Gao", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "PhD student" }, { "name": "Stephan G\u00fcnnemann", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Professor" }, { "name": "Tom Wollschl\u00e4ger", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 23, "n_ref_uni": 53, "n_ref": 111, "n_ref_all": 163, "n_fig": 9, "n_tab": 17, "L_tab": 7198, "n_element_tab": 521, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 1873, "n_element_tab_1": 83, "formula_len_all": 2185, "formula_len_all_1": 1433, "len_all": 233920, "len_all_1": 74112, "len_abs": 1141, "len_title": 121, "len_sents": 76431, "len_sents_1": 33342, "n_sents": 535, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1149, "title": "Lift Your Molecules: Molecular Graph Generation in Latent Euclidean Space", "abs": "We introduce a new framework for 2D molecular graph generation using 3D molecule generative models. Our Synthetic Coordinate Embedding (SyCo) framework maps 2D molecular graphs to 3D Euclidean point clouds via synthetic coordinates and learns the inverse map using an E($n$)-Equivariant Graph Neural Network (EGNN). The induced point cloud-structured latent space is well-suited to apply existing 3D molecule generative models. This approach simplifies the graph generation problem into a point cloud generation problem followed by node and edge classification tasks, without relying on molecular fragments nor autoregressive decoding. Further, we propose a novel similarity-constrained optimization scheme for 3D diffusion models based on inpainting and guidance. As a concrete implementation of our framework, we develop EDM-SyCo based on the E(3) Equivariant Diffusion Model (EDM). EDM-SyCo achieves state-of-the-art performance in distribution learning of molecular graphs, outperforming the best non-autoregressive methods by more than 26\\% on ZINC250K and 16\\% on the GuacaMol dataset while improving conditional generation by up to 3.9 times.", "keywords": [ "Drug Design", "Computational Biology", "Molecule Generation", "Graph Generation", "Latent Diffusion Models" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "uNd289HjLi", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Fan Lam", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Associate Professor" }, { "name": "Jiachen Tu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Yaokun Shi", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 83, "n_formula_1": 17, "n_ref_uni": 35, "n_ref": 66, "n_ref_all": 84, "n_fig": 4, "n_tab": 11, "L_tab": 2578, "n_element_tab": 195, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 465, "n_element_tab_1": 38, "formula_len_all": 6148, "formula_len_all_1": 1261, "len_all": 181713, "len_all_1": 49538, "len_abs": 1745, "len_title": 89, "len_sents": 60029, "len_sents_1": 23938, "n_sents": 444, "n_sents_1": 146, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 1752, "title": "Score-based Self-supervised MRI Denoising", "abs": "Magnetic resonance imaging (MRI) is a powerful noninvasive diagnostic imaging tool that provides unparalleled soft tissue contrast and anatomical detail. Noise contamination, especially in accelerated and/or low-field acquisitions, can significantly degrade image quality and diagnostic accuracy. Supervised learning based denoising approaches have achieved impressive performance but require high signal-to-noise ratio (SNR) labels, which are often unavailable. Self-supervised learning holds promise to address the label scarcity issue, but existing self-supervised denoising methods tend to oversmooth fine spatial features and often yield inferior performance than supervised methods. We introduce Corruption2Self (C2S), a novel score-based self-supervised framework for MRI denoising. At the core of C2S is a generalized ambient denoising score matching (GADSM) loss, which extends denoising score matching to the ambient noise setting by modeling the conditional expectation of higher-SNR images given further corrupted observations. This allows the model to effectively learn denoising across multiple noise levels directly from noisy data. Additionally, we incorporate a reparameterization of noise levels to stabilize training and enhance convergence, and introduce a detail refinement extension to balance noise reduction with the preservation of fine spatial features. Moreover, C2S can be extended to multi-contrast denoising by leveraging complementary information across different MRI contrasts. We demonstrate that our method achieves state-of-the-art performance among self-supervised methods and competitive results compared to supervised counterparts across varying noise conditions and MRI contrasts on the M4Raw and fastMRI dataset.", "keywords": [ "Self-supervised Learning; Score-based denoising; Medical Image Denoising" ], "rating_list": [ 8, 6, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "uN2kkAvarI", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Dibya Ghosh", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Nitish Dashora", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 53, "n_ref": 70, "n_ref_all": 92, "n_fig": 11, "n_tab": 1, "L_tab": 449, "n_element_tab": 99, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 633, "formula_len_all_1": 535, "len_all": 142911, "len_all_1": 62962, "len_abs": 3887, "len_title": 123, "len_sents": 37889, "len_sents_1": 30909, "n_sents": 283, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1441, "title": "ViVa: Video-Trained Value Functions for Guiding Online RL from Diverse Data", "abs": "Online reinforcement learning (RL) with sparse rewards poses a challenge partly because of the lack of feedback on states leading to the goal. Furthermore, expert offline data with reward signal is rarely available to provide this feedback and bootstrap online learning. How can we guide online agents to the right solution without this on-task data? Reward shaping offers a solution by providing fine-grained signal to nudge the policy towards the optimal solution. However, reward shaping often requires domain knowledge to hand-engineer heuristics for a specific goal. To enable more general and inexpensive guidance, we propose and analyze a data-driven methodology that automatically guides RL by learning from widely available video data such as Internet recordings, off-task demonstrations, task failures, and undirected environment interaction. By learning a model of optimal goal-conditioned value from diverse passive data, we open the floor to scaling up and using a wide variety of data sources to model general goal-reaching behaviors relevant to guiding online RL. Specifically, we use intent-conditioned value functions to learn from diverse video and incorporate these goal-conditioned values into the reward. Our experiments show that video-trained value functions work well with a variety of data sources, exhibit positive transfer from human video pre-training, can generalize to unseen goals, and scale with dataset size.", "keywords": [ "Reinforcement Learning", "Unsupervised Pretraining", "Robotics" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "uMxiGoczX1", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Dong Yan", "gender": "Male", "institution": "Baichuan Intelligent Technology", "country": "", "position": "Researcher" }, { "name": "Jialian Li", "gender": "Female", "institution": "Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Jian Xie", "gender": "Male", "institution": "Beijing Baichuan Intelligent Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Ludan ZHANG", "gender": "Female", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "YipinZhang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 11, "n_ref": 21, "n_ref_all": 26, "n_fig": 0, "n_tab": 5, "L_tab": 871, "n_element_tab": 85, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 871, "n_element_tab_1": 85, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 96813, "len_all_1": 69382, "len_abs": 1189, "len_title": 110, "len_sents": 28644, "len_sents_1": 28644, "n_sents": 244, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1195, "title": "Data-Driven Creativity: Amplifying Imagination in LLM Writing", "abs": "During the alignment training of Large Language Models (LLMs), Reinforcement Learning from Human Feedback (RLHF) has proven to be effective in enhancing the model's alignment with human preferences. The RLHF approach requires human annotators to provide data representative of human preferences, aiding the model in advancing towards human-preferred outcomes. In this process, high-quality human preference data is both crucial and challenging to obtain. While many tasks, such as coding and mathematics, can now be more efficiently annotated through Artificial Intelligence Feedback (AIF), numerous tasks still necessitate human input to provide human preference signals. Particularly creative tasks are typical tasks that involving complex human preference. Here, we focus on creative writing tasks and investigate how to collaborate with annotators to acquire high-quality, superior data. We propose an expert-assisted data generation process, named Expert-Objective-Personal-Subjective (EOPS), that can efficiently obtain high-quality ordinal data with minimal human resources. We conduct experiments on three kinds of tasks, and experimental results validat the effectiveness of our method.", "keywords": [ "LLM", "RLHF" ], "rating_list": [ 3, 3, 1, 3 ], "soundness_list": [ 1, 3, 1, 1 ], "presentation_list": [ 1, 2, 1, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "uMLeOlzlZ2", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Chia-Hong Chou", "gender": "Male", "institution": "San Jose State University", "country": "US", "position": "Undergrad student" }, { "name": "Elvis Hsieh", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Janosh Riebesell", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Yuan Chiang", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 58, "n_ref": 78, "n_ref_all": 105, "n_fig": 9, "n_tab": 8, "L_tab": 8618, "n_element_tab": 1673, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 147, "n_element_tab_1": 36, "formula_len_all": 714, "formula_len_all_1": 121, "len_all": 187360, "len_all_1": 55519, "len_abs": 1686, "len_title": 138, "len_sents": 53836, "len_sents_1": 28126, "n_sents": 395, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1691, "title": "LLaMP: Large Language Model Made Powerful for High-fidelity Materials Knowledge Retrieval", "abs": "Reducing hallucination of Large Language Models (LLMs) is imperative for use in the sciences, where reliability and reproducibility are crucial. However, LLMs inherently lack long-term memory, making it a nontrivial, ad hoc, and often biased task to fine-tune them on domain-specific literature and data. Here we introduce LLaMP, a multimodal retrieval-augmented generation (RAG) framework of hierarchical reasoning-and-acting (ReAct) agents that can dynamically and recursively interact with computational and experimental data from the Materials Project (MP) and run atomistic simulations via high-throughput workflow interface. Without fine-tuning, LLaMP demonstrates strong tool-usage ability to comprehend and integrate various modalities of materials science concepts, fetch relevant data stores on the fly, process higher-order data (such as crystal structure and elastic tensor), and streamline complex tasks in computational materials and chemistry. We propose a metric combining uncertainty and confidence estimates to evaluate the self-consistency of responses by LLaMP and vanilla LLMs. Our benchmark shows that LLaMP effectively mitigates the intrinsic bias in LLMs, counteracting the errors on bulk moduli, electronic bandgaps, and formation energies that seem to derive from mixed data sources. We also demonstrate LLaMP\u2019s capability to edit crystal structures and run annealing molecular dynamics simulations using pre-trained machine-learning interatomic potentials. The framework offers an intuitive and nearly hallucination-free approach to exploring and scaling materials informatics and paves the way for knowledge distillation and fine-tuning of future language models.", "keywords": [ "information retrieval", "language model", "database", "materials informatics" ], "rating_list": [ 8, 1, 8, 3 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "uMEsKEiB7J", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Boqi Pan", "gender": "Male", "institution": "HZNU", "country": "", "position": "Undergrad student" }, { "name": "Cheng Deng", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Cunxiang Wang", "gender": "Not Specified", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Guangsheng Bao", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Qian Wang", "gender": "Female", "institution": "Hangzhou Normal University, China ", "country": "", "position": "Lecturer" }, { "name": "Qipeng Guo", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "", "position": "Researcher" }, { "name": "Ruoxi Ning", "gender": "Female", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Tonghui Wu", "gender": "Female", "institution": "HZNU", "country": "", "position": "Undergrad student" }, { "name": "Xiangkun Hu", "gender": "Male", "institution": "Amazon", "country": "CN", "position": "Applied Scientist" }, { "name": "Yue Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" }, { "name": "Zheng Zhang", "gender": "Male", "institution": "Amazon", "country": "CN", "position": "Senior Principal Scientist" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 44, "n_ref": 84, "n_ref_all": 122, "n_fig": 6, "n_tab": 8, "L_tab": 3174, "n_element_tab": 234, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1704, "n_element_tab_1": 159, "formula_len_all": 223, "formula_len_all_1": 0, "len_all": 184880, "len_all_1": 64933, "len_abs": 1213, "len_title": 123, "len_sents": 56526, "len_sents_1": 30273, "n_sents": 456, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1189, "title": "NovelQA: Benchmarking Question Answering on Documents Exceeding 200K Tokens", "abs": "The rapid advancement of Large Language Models (LLMs) has introduced a new frontier in natural language processing, particularly in understanding and processing long-context information. However, the evaluation of these models' long-context abilities remains a challenge due to the limitations of current benchmarks. To address this gap, we introduce NovelQA, a benchmark specifically designed to test the capabilities of LLMs with extended texts. Constructed from English novels, NovelQA offers a unique blend of complexity, length, and narrative coherence, making it an ideal tool for assessing deep textual understanding in LLMs. This paper presents the design and construction of NovelQA, highlighting its manual annotation, and diverse question types. Our evaluation of Long-context LLMs on NovelQA reveals significant insights into the models' performance, particularly emphasizing the challenges they face with multi-hop reasoning, detail-oriented questions, and extremely long input with an average length more than 200,000 tokens. The results underscore the necessity for further advancements in LLMs to improve their long-context comprehension and computational literary studies.", "keywords": [ "Long-context", "Large Language Models", "Question Answering" ], "rating_list": [ 8, 6, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "uM2IDdivyC", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Alexander Rudikov", "gender": "unknown", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Researcher" }, { "name": "Ekaterina Muravleva", "gender": "Female", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "Assistant Professor" }, { "name": "Ivan Oseledets", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "Oleg Iliev", "gender": "Male", "institution": "Fraunhofer Institute for Industrial Mathematics ", "country": "", "position": "Principal Researcher" }, { "name": "Vladislav Trifonov", "gender": "Male", "institution": "Sberbank of Russia", "country": "RU", "position": "Researcher" }, { "name": "Yuri M. Laevsky", "gender": "Male", "institution": "Institute of Computational Mathematics and Mathematical Geophysics SB RAS", "country": "RU", "position": "Principal Researcher" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 49, "n_ref": 80, "n_ref_all": 100, "n_fig": 3, "n_tab": 6, "L_tab": 1626, "n_element_tab": 124, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 863, "n_element_tab_1": 58, "formula_len_all": 1141, "formula_len_all_1": 1007, "len_all": 114474, "len_all_1": 44988, "len_abs": 1333, "len_title": 131, "len_sents": 25550, "len_sents_1": 19976, "n_sents": 193, "n_sents_1": 137, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1331, "title": "ConDiff: A Challenging Dataset for Neural Solvers of Partial Differential Equations", "abs": "We present ConDiff, a novel dataset for scientific machine learning. ConDiff focuses on the diffusion equation with varying coefficients, a fundamental problem in many applications of parametric partial differential equations (PDEs). The main novelty of the proposed dataset is that we consider discontinuous coefficients with high contrast. These coefficient functions are sampled from a selected set of distributions. This class of problems is not only of great academic interest, but is also the basis for describing various environmental and industrial problems. In this way, ConDiff shortens the gap with real-world problems while remaining fully synthetic and easy to use. ConDiff consists of a diverse set of diffusion equations with coefficients covering a wide range of contrast levels and heterogeneity with a measurable complexity metric for clearer comparison between different coefficient functions. We baseline ConDiff on standard deep learning models in the field of scientific machine learning. By providing a large number of problem instances, each with its own coefficient function and right-hand side, we hope to encourage the development of novel physics-based deep learning approaches, such as neural operators, ultimately driving progress towards more accurate and efficient solutions of complex PDE problems.", "keywords": [ "Dataset", "PDEs", "scientific computing", "neural operators" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "uLAAVg0ymc", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Abdoulaye SAKHO", "gender": "Male", "institution": "Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "FR", "position": "PhD student" }, { "name": "Emmanuel Malherbe", "gender": "Male", "institution": "Artefact", "country": "FR", "position": "Research director" }, { "name": "Erwan Scornet", "gender": "Male", "institution": "Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "FR", "position": "Associate Professor" } ], "n_formula": 87, "n_formula_1": 15, "n_ref_uni": 43, "n_ref": 69, "n_ref_all": 147, "n_fig": 3, "n_tab": 19, "L_tab": 22171, "n_element_tab": 1658, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1803, "n_element_tab_1": 84, "formula_len_all": 8706, "formula_len_all_1": 1390, "len_all": 234217, "len_all_1": 67795, "len_abs": 1017, "len_title": 146, "len_sents": 53101, "len_sents_1": 28139, "n_sents": 548, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 950, "title": "Do we need rebalancing strategies? A theoretical and empirical study around SMOTE and its variants", "abs": "Synthetic Minority Oversampling Technique (SMOTE) is a common rebalancing strategy for handling imbalanced tabular data sets. However, few works analyze SMOTE theoretically. In this paper, we prove that SMOTE (with default parameter) tends to copy the original minority samples asymptotically. We also prove that SMOTE exhibits boundary artifacts, thus justifying existing SMOTE variants. Then we introduce two new SMOTE-related strategies, and compare them with state-of-the-art rebalancing procedures. Surprisingly, for most data sets, we observe that applying no rebalancing strategy is competitive in terms of predictive performances, with tuned random forests, logistic regression or LightGBM. For highly imbalanced data sets, our new methods, named CV-SMOTE and Multivariate Gaussian SMOTE, are competitive. Besides, our analysis sheds some lights on the behavior of common rebalancing strategies, when used in conjunction with random forests.", "keywords": [ "Classification", "Imbalanced data set", "SMOTE" ], "rating_list": [ 8, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "uL1H29dM0c", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Haksoo Lim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Kookjin Lee", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" }, { "name": "Nathaniel Trask", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "Associate Professor" }, { "name": "Noseong Park", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Anthony D. Gruber", "gender": "Male", "institution": "Sandia National Laboratories", "country": "US", "position": "Researcher" } ], "n_formula": 76, "n_formula_1": 19, "n_ref_uni": 23, "n_ref": 49, "n_ref_all": 69, "n_fig": 6, "n_tab": 4, "L_tab": 491, "n_element_tab": 51, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 67, "n_element_tab_1": 5, "formula_len_all": 7529, "formula_len_all_1": 1830, "len_all": 171071, "len_all_1": 72067, "len_abs": 3280, "len_title": 54, "len_sents": 62826, "len_sents_1": 33395, "n_sents": 471, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 776, "title": "Efficiently Parameterized Neural Metriplectic Systems", "abs": "Metriplectic systems are learned from data in a way that scales quadratically in both the size of the state and the rank of the metriplectic operators. In addition to being provably energy-conserving and entropy-stable, the proposed neural metriplectic systems (NMS) approach includes approximation results that demonstrate its ability to accurately learn metriplectic dynamics from data, along with an error estimate that indicates its potential for generalization to unseen timescales when the approximation error is low. Examples are provided to illustrate performance both with full state information available and when entropic variables are unknown, confirming that the NMS approach exhibits superior accuracy and scalability without compromising on model expressivity.", "keywords": [ "metriplectic systems", "structure preservation", "energy conservation", "entropy stability", "neural ODEs" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 2, 2, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 2, 2 ] }, { "paper_id": "uKZdlihDDn", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Mario Lino Valencia", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Postdoc" }, { "name": "Nils Thuerey", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Associate Professor" }, { "name": "Tobias Pfaff", "gender": "Male", "institution": "Deepmind", "country": "GB", "position": "Research scientist" } ], "n_formula": 39, "n_formula_1": 12, "n_ref_uni": 62, "n_ref": 122, "n_ref_all": 203, "n_fig": 25, "n_tab": 16, "L_tab": 4760, "n_element_tab": 491, "n_fig_1": 17, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1898, "formula_len_all_1": 699, "len_all": 240048, "len_all_1": 69603, "len_abs": 1495, "len_title": 129, "len_sents": 76980, "len_sents_1": 34424, "n_sents": 505, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 81, "L_abs": 1501, "title": "Learning Distributions of Complex Fluid Simulations with Diffusion Graph Networks", "abs": "Physical systems with complex unsteady dynamics, such as fluid flows, are often poorly represented by a single mean solution. For many practical applications, it is crucial to access the full distribution of possible states, from which relevant statistics (e.g., RMS and two-point correlations) can be derived. Here, we propose a graph-based latent diffusion model that enables direct sampling of states from their equilibrium distribution, given a mesh discretization of the system and its physical parameters. This allows for the efficient computation of flow statistics without running long and expensive numerical simulations. The graph-based structure enables operations on unstructured meshes, which is critical for representing complex geometries with spatially localized high gradients, while latent-space diffusion modeling with a multi-scale GNN allows for efficient learning and inference of entire distributions of solutions. A key finding of our work is that the proposed networks can accurately learn full distributions even when trained on incomplete data from relatively short simulations. We apply this method to a range of fluid dynamics tasks, such as predicting pressure distributions on 3D wing models in turbulent flow, demonstrating both accuracy and computational efficiency in challenging scenarios. The ability to directly sample accurate solutions, and capturing their diversity from short ground-truth simulations, is highly promising for complex scientific modeling tasks.", "keywords": [ "Graph Neural Networks", "Diffusion Models", "Physics Simulations" ], "rating_list": [ 8, 6, 8, 8, 8 ], "soundness_list": [ 4, 3, 3, 3, 4 ], "presentation_list": [ 4, 2, 4, 3, 2 ], "contribution_list": [ 3, 2, 4, 3, 3 ], "confidence_list": [ 2, 2, 3, 5, 3 ] }, { "paper_id": "uJqKf24HGN", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Chao Dong", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Fanghua Yu", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Jinfan Hu", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Jinjin Gu", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "Instructor" }, { "name": "Zheyuan Li", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 56, "n_ref": 133, "n_ref_all": 180, "n_fig": 19, "n_tab": 6, "L_tab": 3723, "n_element_tab": 469, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1592, "n_element_tab_1": 127, "formula_len_all": 11894, "formula_len_all_1": 10259, "len_all": 207705, "len_all_1": 84048, "len_abs": 1362, "len_title": 142, "len_sents": 51006, "len_sents_1": 33053, "n_sents": 393, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1364, "title": "UniCon: Unidirectional Information Flow for Effective Control of Large-Scale Diffusion Models", "abs": "We introduce UniCon, a novel architecture designed to enhance control and efficiency in training adapters for large-scale diffusion models like the Diffusion transformer. Unlike existing methods that rely on bidirectional interaction between the diffusion model and control adapter, UniCon implements a unidirectional flow from the diffusion network to the adapter, allowing the adapter alone to generate the final output. UniCon reduces computational demands by eliminating the need for the diffusion model to compute and store gradients during adapter training. UniCon is free from the constrains of encoder-focused designs and is able to utilize all parameters of the diffusion model, making it highly effective for transformer-based architectures. Our results indicate that UniCon reduces GPU memory usage by one-third and increases training speed by 2.3 times, while all maintaining the same adapter parameter size. Additionally, without requiring extra computational resources, UniCon enables the training of adapters with double the parameter volume of existing ControlNets. In a series of image condition generation tasks, UniCon has demonstrated precise response to control information and excellent generation capabilities. UniCon makes the control of large-scale diffusion models feasible and provides a basis for further scaling up of diffusion models.", "keywords": [ "Diffusion Models", "Lowlevel Vision", "Conditional Generation" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "uIg9Vcw2CY", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "John Lowengrub", "gender": "unknown", "institution": "University of California, Irvine", "country": "", "position": "Full Professor" }, { "name": "Ray Zirui Zhang", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Postdoc" }, { "name": "Xiaohui Xie", "gender": "unknown", "institution": "University of California, Irvine", "country": "US", "position": "Full Professor" } ], "n_formula": 67, "n_formula_1": 16, "n_ref_uni": 61, "n_ref": 100, "n_ref_all": 135, "n_fig": 14, "n_tab": 3, "L_tab": 226, "n_element_tab": 15, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 6, "n_element_tab_1": 1, "formula_len_all": 3479, "formula_len_all_1": 920, "len_all": 235898, "len_all_1": 67856, "len_abs": 1422, "len_title": 62, "len_sents": 68457, "len_sents_1": 32707, "n_sents": 678, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1195, "title": "BiLO: Bilevel Local Operator Learning for PDE inverse problems", "abs": "We propose a new neural network based method for solving inverse problems for partial differential equations (PDEs) by formulating the PDE inverse problem as a bilevel optimization problem. At the upper level, we minimize the data loss with respect to the PDE parameters. At the lower level, we train a neural network to locally approximate the PDE solution operator in the neighborhood of a given set of PDE parameters, which enables an accurate approximation of the descent direction for the upper level optimization problem. The lower level loss function includes the L2 norms of both the residual and its derivative with respect to the PDE parameters. We apply gradient descent simultaneously on both the upper and lower level optimization problems, leading to an effective and fast algorithm. The method, which we refer to as BiLO (Bilevel Local Operator learning), is also able to efficiently infer unknown functions in the PDEs through the introduction of an auxiliary variable. We demonstrate that our method enforces strong PDE constraints, is robust to sparse and noisy data, and eliminates the need to balance the residual and the data loss, which is inherent to soft PDE constraints.", "keywords": [ "Partial Differential Equations", "physics-informed", "PINN", "inverse problem", "Bilevel optimization", "Neural Operator" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "uIGnuyDSB9", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Aram Galstyan", "gender": "Male", "institution": "Information Sciences Institute", "country": "", "position": "Research Director" }, { "name": "Bhavana Ganesh", "gender": "Female", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Jongwoo Ko", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sailik Sengupta", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Saket Dingliwal", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Sravan Babu Bodapati", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" } ], "n_formula": 24, "n_formula_1": 19, "n_ref_uni": 74, "n_ref": 215, "n_ref_all": 262, "n_fig": 13, "n_tab": 11, "L_tab": 5049, "n_element_tab": 866, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1982, "n_element_tab_1": 349, "formula_len_all": 1873, "formula_len_all_1": 1459, "len_all": 262023, "len_all_1": 67873, "len_abs": 1314, "len_title": 72, "len_sents": 63986, "len_sents_1": 28207, "n_sents": 493, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1282, "title": "SeRA: Self-Reviewing and Alignment of LLMs using Implicit Reward Margins", "abs": "Direct alignment algorithms (DAAs), such as direct preference optimization (DPO), have become popular alternatives to Reinforcement Learning from Human Feedback (RLHF) due to their simplicity, efficiency, and stability. However, the preferences used by DAAs are usually collected before alignment training begins and remain unchanged (off-policy). This design leads to two problems where the policy model (1) picks up on spurious correlations in the dataset (as opposed to only learning alignment to human preferences), and (2) overfits to feedback on off-policy trajectories that have less likelihood of being generated by the updated policy model. To address these issues, we introduce Self-Reviewing and Alignment (SeRA), a cost-efficient and effective method that can be readily combined with existing DAAs. SeRA comprises of two components: (1) sample selection using implicit reward margin to alleviate over-optimization on such undesired features, and (2) preference bootstrapping using implicit rewards to augment preference data with updated policy models in a cost-efficient manner. Extensive experiments, including on instruction-following tasks, demonstrate the effectiveness and generality of SeRA in training LLMs with diverse offline preference datasets and and DAAs.", "keywords": [ "Preference Alignment", "Large Language Models", "Implicit Reward Margin", "Sample Selection", "Preference Bootstrapping" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 2, 3, 5, 3 ] }, { "paper_id": "uHkfU4TaPh", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Jiaxian Guo", "gender": "Male", "institution": "Google Research", "country": "AU", "position": "Research Scientist" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "LIANG DING", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Min Zhang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Minyan Zeng", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Wenbin Wang", "gender": "unknown", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Xiabin Zhou", "gender": "Male", "institution": "JiangSu University", "country": "", "position": "MS student" }, { "name": "Xuebo Liu", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 35, "n_ref": 68, "n_ref_all": 77, "n_fig": 3, "n_tab": 4, "L_tab": 5643, "n_element_tab": 396, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 5278, "n_element_tab_1": 341, "formula_len_all": 398, "formula_len_all_1": 366, "len_all": 127838, "len_all_1": 72623, "len_abs": 2103, "len_title": 121, "len_sents": 28920, "len_sents_1": 28201, "n_sents": 182, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1987, "title": "DynamicKV: Task-Aware Adaptive KV Cache Compression for Long Context LLMs", "abs": "Efficiently managing the KV cache in Large Language Models (LLMs) is a critical challenge for long-context processing tasks such as retrieval-augmented generation (RAG), long text summarization, and multi-document analysis. Extending the context length substantially increases the KV cache size, leading to excessive memory consumption. Existing KV cache compression methods enforce a fixed pattern, neglecting task-specific characteristics, which hampers the effective retention of essential information while discarding less important tokens. In this paper, we introduce a novel Task-Aware KV cache mechanism that dynamically adjusts the KV cache size across different layers based on the characteristics of the tasks. Our approach builds on the significant observation of distinct activation patterns across layers in various tasks, which highlights the need for adaptive strategies tailored to each task's unique demands. Based on this insight, we propose DynamicKV, a method that dynamically optimizes token retention by adjusting the number of tokens retained at each layer, adapting to the specific task. DynamicKV establishes global and per-layer maximum KV cache budgets, temporarily retaining the maximum budget for the current layer, and periodically updating the KV cache sizes of all preceding layers during inference. Our method demonstrates exceptional performance on the LongBench dataset, retaining only 1.7\\% of the KV cache while preserving 90\\%, 87\\%, 78\\%, and 83\\% of the original accuracy for LlaMA-3-8B-Instruct, Mistral-7B-Instruct-v0.2, Qwen2-7B-Instruct, and InternLM-2.5-7B-Chat-1M, respectively. When the retained KV cache size is increased to 6.9\\%, the performance becomes nearly indistinguishable from that without any KV cache compression. Notably, even under extreme compression (0.9\\%), DynamicKV surpasses state-of-the-art (SOTA) methods by 11\\% in the Needle-in-a-Haystack test using Mistral-7B-Instruct-v0.2. The code will be released to the public.", "keywords": [ "Large Language Model", "Efficient Inference", "Key-Value Cache Compression" ], "rating_list": [ 5, 3, 5, 6, 3 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "uHgVrGF2Wn", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Bin Xu", "gender": "Male", "institution": "Department of Computer Science, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Ji Qi", "gender": "Male", "institution": "Zhipu AI", "country": "CN", "position": "Intern" }, { "name": "Jie Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Ming Ding", "gender": "Male", "institution": "ZHIPU AI", "country": "CN", "position": "Principal Researcher" }, { "name": "Weihan Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Wenyi Hong", "gender": "Female", "institution": "Department of Computer Science and Technology, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiaohan Zhang", "gender": "Female", "institution": "Beijing Knowledge Atlas Technology Co., Ltd. ", "country": "CN", "position": "Researcher" }, { "name": "Xiaotao Gu", "gender": "Male", "institution": "Zhipu AI", "country": "CN", "position": "Researcher" }, { "name": "Yean Cheng", "gender": "Male", "institution": "ZhipuAI", "country": "CN", "position": "Researcher" }, { "name": "Yuxiao Dong", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Zehai He", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Shiyu Huang", "gender": "Male", "institution": "Zhipu AI", "country": "CN", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 32, "n_ref": 67, "n_ref_all": 77, "n_fig": 5, "n_tab": 4, "L_tab": 699, "n_element_tab": 59, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2914, "n_element_tab_1": 134, "formula_len_all": 89, "formula_len_all_1": 57, "len_all": 101000, "len_all_1": 54859, "len_abs": 209, "len_title": 103, "len_sents": 22529, "len_sents_1": 19766, "n_sents": 156, "n_sents_1": 139, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 1137, "title": "LVBench: An Extreme Long Video Understanding Benchmark", "abs": "Recent progress in multimodal large language models has markedly enhanced the understanding of short videos (typically under one minute), and several evaluation datasets have emerged accordingly. However, these advancements fall short of meeting the demands of real-world applications such as embodied intelligence for long-term decision-making, in-depth movie reviews and discussions, and live sports commentary, all of which require comprehension of long videos spanning several hours. To address this gap, we introduce LVBench, a benchmark specifically designed for long video understanding. Our dataset comprises publicly sourced videos and encompasses a diverse set of tasks aimed at long video comprehension and information extraction. LVBench is designed to challenge multimodal models to demonstrate long-term memory and extended comprehension capabilities. Our extensive evaluations reveal that current multimodal models still underperform on these demanding long video understanding tasks. Through LVBench, we aim to spur the development of more advanced models capable of tackling the complexities of long video comprehension.", "keywords": [ "Video understanding", "Multimodal learning", "Visual question answering", "Long-form video", "Datasets and benchmarking" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "uHTmx0nRfX", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Corey D Barrett", "gender": "Male", "institution": "Oracle", "country": "US", "position": "Researcher" }, { "name": "George Karypis", "gender": "Male", "institution": "University of Minnesota, Minneapolis", "country": "", "position": "Full Professor" }, { "name": "Georgiana Dinu", "gender": "Female", "institution": "Amazon", "country": "", "position": "Principal Researcher" }, { "name": "Miguel Romero Calvo", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" }, { "name": "Shuoyang Ding", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 37, "n_ref": 76, "n_ref_all": 106, "n_fig": 8, "n_tab": 7, "L_tab": 1244, "n_element_tab": 131, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 827, "n_element_tab_1": 102, "formula_len_all": 0, "formula_len_all_1": 90, "len_all": 129813, "len_all_1": 56900, "len_abs": 1160, "len_title": 98, "len_sents": 36138, "len_sents_1": 25223, "n_sents": 250, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1174, "title": "MoTE: Mixture of Task Experts for Embedding Models", "abs": "Dense embeddings are essential for Retrieval-Augmented Generation (RAG), search, classification, and clustering systems. Recent methods improve dense embeddings by enriching them with instructions that incorporate downstream task information, enabling a single model to generate task-specific embedding spaces. However, we empirically show that requiring all tasks to share the same model parameters imposes significant representational limitations. To address these challenges, we introduce Mixture of Task Experts (MoTE), a novel transformer block designed for embedding architectures. MoTE employs dedicated parameter sets tailored to the unique requirements of each task and is paired with a task-aware training framework to improve representation quality. Experiments on 56 datasets spanning $7$ tasks demonstrate that MoTE outperforms instruction-conditioned models, achieving, on average, $1.62$ higher NDCG@10 on retrieval datasets, $1.54$ higher MAP on re-ranking datasets, and a $0.65$ improvement in overall performance. Notably, these gains are achieved without altering inference-time information, training data, inference speed, or number of active parameters.", "keywords": [ "Embedding Models", "Representation Learning", "Mixture of Experts", "Retrieval Augmented Generation", "RAG", "Search", "Clustering", "Classification" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "uHLgDEgiS5", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Dawn Song", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "James Zou", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Prateek Mittal", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "Full Professor" }, { "name": "Ruoxi Jia", "gender": "unknown", "institution": "Virginia Tech", "country": "US", "position": "Assistant Professor" }, { "name": "Tianhao Wang", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" } ], "n_formula": 66, "n_formula_1": 10, "n_ref_uni": 36, "n_ref": 80, "n_ref_all": 108, "n_fig": 14, "n_tab": 8, "L_tab": 2552, "n_element_tab": 92, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 229, "n_element_tab_1": 17, "formula_len_all": 5911, "formula_len_all_1": 745, "len_all": 292191, "len_all_1": 87265, "len_abs": 2010, "len_title": 109, "len_sents": 106639, "len_sents_1": 41841, "n_sents": 749, "n_sents_1": 295, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 60, "L_abs": 2034, "title": "Capturing the Temporal Dependence of Training Data Influence", "abs": "Traditional data influence estimation methods, like influence function, assume that learning algorithms are permutation-invariant with respect to training data. However, modern training paradigms\u2014especially for foundation models using stochastic algorithms and non-convergent, multi-stage curricula\u2014are sensitive to data ordering, thus violating this assumption. This mismatch renders influence functions inadequate for answering some critical questions in current machine learning: How can we differentiate the influence of the same data contributing at different stages of training? More generally, how can we capture the dependence of data influence on the optimization trajectory during training? To address this gap, we formalize the concept of \\emph{trajectory-specific leave-one-out (LOO) influence}, which quantifies the impact of removing a data point from a specific iteration during training, accounting for the exact sequence of data encountered and the model's optimization trajectory. However, exactly evaluating the trajectory-specific LOO presents a significant computational challenge. To address this, we propose \\emph{data value embedding}, a novel technique enabling efficient approximation of trajectory-specific LOO. Specifically, we compute a training data embedding that encapsulates the cumulative interactions between data and the evolving model parameters. The LOO can then be efficiently approximated through a simple dot-product between the data value embedding and the gradient of the given test data. As data value embedding captures training data ordering, it offers valuable insights into model training dynamics. In particular, we uncover distinct phases of data influence, revealing that data points in the early and late stages of training exert a greater impact on the final model. These insights translate into actionable strategies for managing the computational overhead of data selection by strategically timing the selection process, potentially opening new avenues in data curation research.", "keywords": [ "data valuation" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "uGka5qOsop", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "HyunJin Kim", "gender": "Male", "institution": "Dankook University", "country": "KR", "position": "Full Professor" }, { "name": "Sumin Oh", "gender": "unknown", "institution": "Dankook University", "country": "KR", "position": "MS student" }, { "name": "WANSOO KIM", "gender": "Male", "institution": "Dankook University", "country": "KR", "position": "Undergrad student" } ], "n_formula": 17, "n_formula_1": 12, "n_ref_uni": 36, "n_ref": 90, "n_ref_all": 119, "n_fig": 11, "n_tab": 2, "L_tab": 437, "n_element_tab": 55, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 211, "n_element_tab_1": 46, "formula_len_all": 1425, "formula_len_all_1": 607, "len_all": 142468, "len_all_1": 60737, "len_abs": 1413, "len_title": 117, "len_sents": 47928, "len_sents_1": 27763, "n_sents": 378, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1418, "title": "A Temporally Correlated Latent Exploration for Reinforcement Learning", "abs": "Efficient exploration remains one of the longstanding problems of deep reinforcement learning. Instead of depending solely on extrinsic rewards from the environments, existing methods use intrinsic rewards to enhance exploration. However, we demonstrate that these methods are vulnerable to Noisy TV and stochasticity. To tackle this problem, we propose Temporally Correlated Latent Exploration (TeCLE), which is a novel intrinsic reward formulation that employs an action-conditioned latent space and temporal correlation. The action-conditioned latent space models the probability distribution of states, thereby avoiding the assignment of excessive intrinsic rewards to unpredictable states and effectively addressing both problems. Whereas previous works inject temporal correlation for action selection, the proposed method injects it for intrinsic reward computation. We find that the injected temporal correlation determines the exploratory behaviors of agents. Various experiments show that the environment where the agent performs well depends on the amount of temporal correlation. To the best of our knowledge, the proposed TeCLE is the first approach to consider the action-conditioned latent space and temporal correlation for curiosity-driven exploration. We prove that the proposed TeCLE can be robust to the Noisy TV and stochasticity in benchmark environments, including Minigrid and Stochastic Atari.", "keywords": [ "Colored Noise", "Intrinsic Reward", "Latent State Representation", "Reinforcement Learning", "Temporal Correlation" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 2, 3, 1, 1 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "uGJxl2odR0", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Hyungi Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Juho Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "CHAEYOON JANG", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" }, { "name": "DONGBOK LEE", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 23, "n_ref_uni": 34, "n_ref": 75, "n_ref_all": 127, "n_fig": 5, "n_tab": 33, "L_tab": 10856, "n_element_tab": 635, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2556, "n_element_tab_1": 137, "formula_len_all": 2149, "formula_len_all_1": 1753, "len_all": 226980, "len_all_1": 70217, "len_abs": 4081, "len_title": 83, "len_sents": 73917, "len_sents_1": 31046, "n_sents": 506, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 35, "L_abs": 1454, "title": "Dimension Agnostic Neural Processes", "abs": "Meta-learning aims to train models that can generalize to new tasks with limited labeled data by extracting shared features across diverse task datasets. Additionally, it accounts for prediction uncertainty during both training and evaluation, a concept known as uncertainty-aware meta-learning. Neural Process (NP) is a well-known uncertainty-aware meta-learning method that constructs implicit stochastic processes using parametric neural networks, enabling rapid adaptation to new tasks. However, existing NP methods face challenges in accommodating diverse input dimensions and learned features, limiting their broad applicability across regression tasks. To address these limitations and advance the utility of NP models as general regressors, we introduce Dimension Agnostic Neural Process (DANP). DANP incorporates Dimension Aggregator Block (DAB) to transform input features into a fixed-dimensional space, enhancing the model's ability to handle diverse datasets. Furthermore, leveraging the Transformer architecture and latent encoding layers, DANP learns a wider range of features that are generalizable across various tasks. Through comprehensive experimentation on various synthetic and practical regression tasks, we empirically show that DANP outperforms previous NP variations, showcasing its effectiveness in overcoming the limitations of traditional NP models and its potential for broader applicability in diverse regression scenarios.", "keywords": [ "Neural Processes", "varying dimension", "regression", "bayesian optimization" ], "rating_list": [ 6, 5, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 3, 5, 2, 3, 3 ] }, { "paper_id": "uEqOYXtn7f", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Alexandros Kalousis", "gender": "Male", "institution": "University of Applied Sciences Western Switzerland", "country": "CH", "position": "Full Professor" }, { "name": "Pablo Strasser", "gender": "Male", "institution": "Geneva School of Business Administration, HES-SO University of Applied Sciences of Western Switzerland;", "country": "CH", "position": "Researcher" }, { "name": "Yoann Boget", "gender": "Male", "institution": "University of Geneva", "country": "CH", "position": "PhD student" } ], "n_formula": 47, "n_formula_1": 22, "n_ref_uni": 51, "n_ref": 119, "n_ref_all": 136, "n_fig": 2, "n_tab": 6, "L_tab": 2386, "n_element_tab": 161, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 454, "n_element_tab_1": 57, "formula_len_all": 4883, "formula_len_all_1": 1190, "len_all": 213395, "len_all_1": 70717, "len_abs": 1072, "len_title": 89, "len_sents": 62039, "len_sents_1": 30737, "n_sents": 524, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1091, "title": "HIERARCHICAL EQUIVARIANT GRAPH GENERATION", "abs": "Deep learning, and more specifically denoising models, have significantly improved graph generative modeling. However, challenges remain in capturing global graph properties from local interactions, ensuring scalability, and maintaining node permutation equivariance. While existing equivariant models address node permutation issues, they struggle with scalability, often requiring dense graph representations that scale with $\\mathcal{O}(n^2)$.\n\nTo overcome these challenges, we introduce a novel coarsening-lifting method that generates sparse spanning supergraphs, preserving global graph properties. These supergraphs serve as both conditioning structures and sparse message-passing layouts for generative models. Leveraging this method with discrete diffusion, we model graphs hierarchically, enabling efficient generation of large graphs.\n\nOur approach, to the best of our knowledge, is the first hierarchical equivariant generative model for graphs. We demonstrate its performance introducing new evaluation datasets with larger graphs and more instances than traditional benchmarks.", "keywords": [ "graph", "generative model", "graph generation", "hierarchical", "coarsening", "pooling", "lifting", "gnn", "mpnn", "spanning supergraph" ], "rating_list": [ 6, 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 1, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "uEPRY2XAEs", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Druv Pai", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department, University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Jiantao Jiao", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" }, { "name": "Michael Jordan", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Song Mei", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" }, { "name": "Tianyu Guo", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Yu Bai", "gender": "unknown", "institution": "OpenAI", "country": "", "position": "Researcher" } ], "n_formula": 48, "n_formula_1": 14, "n_ref_uni": 44, "n_ref": 84, "n_ref_all": 120, "n_fig": 26, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 15, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5190, "formula_len_all_1": 1059, "len_all": 178313, "len_all_1": 68095, "len_abs": 1325, "len_title": 140, "len_sents": 56138, "len_sents_1": 32495, "n_sents": 508, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1344, "title": "Active-Dormant Attention Heads: Mechanistically Demystifying Extreme-Token Phenomena in LLMs", "abs": "We investigate the mechanisms behind three puzzling phenomena observed in transformer-based large language models (LLMs): *attention sinks*, *value-state drains*, and *residual-state peaks*, collectively referred to the *extreme-token phenomena*. First, we demonstrate that these phenomena also arise in simpler architectures\u2014transformers with one to three layers\u2014trained on a toy model, the Bigram-Backcopy (BB) task. In this setting, we identify an *active-dormant mechanism* that causes attention heads to become attention sinks for certain domain-specific inputs while remaining non-sinks for others. We further develop a precise theoretical characterization of the training dynamics that lead to these phenomena, revealing that they are driven by a *mutual reinforcement mechanism*. By small interventions, we demonstrate ways to avoid extreme-token phenomena during pre-training. Next, we extend our analysis to pre-trained LLMs, including Llama and OLMo, revealing that many attention heads are governed by a similar active-dormant mechanism as in the BB task. We further show that the same mutual reinforcement mechanism drives the emergence of extreme-token phenomena during LLM pre-training. Our results study the mechanisms behind extreme-token phenomena in both synthetic and real settings and offer potential mitigation strategies.", "keywords": [ "attention sink", "mechanistic interpretability", "language models", "transformers" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "uE84MGbKD7", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Grant Schoenebeck", "gender": "unknown", "institution": "University of Michigan", "country": "", "position": "Assistant Professor" }, { "name": "Shengwei Xu", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "PhD student" }, { "name": "Yuqing Kong", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuxuan Lu", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 13, "n_ref_uni": 49, "n_ref": 85, "n_ref_all": 118, "n_fig": 4, "n_tab": 7, "L_tab": 3903, "n_element_tab": 415, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 296, "n_element_tab_1": 14, "formula_len_all": 2364, "formula_len_all_1": 862, "len_all": 202614, "len_all_1": 68606, "len_abs": 2483, "len_title": 98, "len_sents": 69947, "len_sents_1": 32733, "n_sents": 565, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1594, "title": "Benchmarking LLMs' Judgments with No Gold Standard", "abs": "We introduce the GEM (Generative Estimator for Mutual Information), an evaluation metric for assessing language generation by large language models (LLMs), particularly in generating informative judgments, without the need for a gold standard reference. GEM broadens the scenarios where we can benchmark LLM generation performance-from traditional ones, like machine translation and summarization, where gold standard references are readily available, to subjective tasks without clear gold standards, such as academic peer review.\n\nGEM uses a generative model to estimate mutual information between candidate and reference responses, without requiring the reference to be a gold standard. In experiments on two human-annotated datasets, GEM demonstrates competitive correlations with human scores compared to the state-of-the-art GPT-4o Examiner, and outperforms all other baselines. Additionally, GEM is more robust against strategic manipulation, such as rephrasing or elongation, which can artificially inflate scores under a GPT-4o Examiner. \n\nWe also present GRE-bench (Generating Review Evaluation Benchmark) which evaluates LLMs based on how well they can generate high-quality peer reviews for academic research papers. Because GRE-bench is based upon GEM, it inherits its robustness properties. Additionally, GRE-bench circumvents data contamination problems (or data leakage) by using the continuous influx of new open-access research papers and peer reviews each year. We show GRE-bench results of various popular LLMs on their peer review capabilities using the ICLR2023 dataset.", "keywords": [ "Benchmarking", "Peer Review", "Mutual Information", "Data Contamination", "Large Language Models" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "uDjuCpQH5N", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Aghyad Deeb", "gender": "unknown", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "US", "position": "MS student" }, { "name": "Fabien Roger", "gender": "Male", "institution": "Anthropic", "country": "GB", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 25, "n_ref": 66, "n_ref_all": 86, "n_fig": 8, "n_tab": 5, "L_tab": 2434, "n_element_tab": 208, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 9, "n_element_tab_1": 1, "formula_len_all": 95, "formula_len_all_1": 95, "len_all": 126643, "len_all_1": 46342, "len_abs": 934, "len_title": 117, "len_sents": 37261, "len_sents_1": 23113, "n_sents": 287, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 936, "title": "Do Unlearning Methods Remove Information from Language Model Weights?", "abs": "Large Language Models' knowledge of how to perform cyber-security attacks, create bioweapons, and manipulate humans poses risks of misuse. Previous work has proposed methods to unlearn this knowledge. Historically, it has been unclear whether unlearning techniques are removing information from the model weights or just making it harder to access. To disentangle these two objectives, we propose an adversarial evaluation method to test for the removal of information from model weights: we give an attacker access to some facts that were supposed to be removed, and using those, the attacker tries to recover other facts from the same distribution that cannot be guessed from the accessible facts. We show that using fine-tuning on the accessible facts can recover 88% of the pre-unlearning accuracy when applied to current unlearning methods, revealing the limitations of these methods in removing information from the model weights.", "keywords": [ "Unlearning", "LLMs", "Evaluation", "RMU", "Fine-tuning", "Red-teaming", "Safety" ], "rating_list": [ 6, 8, 3, 5 ], "soundness_list": [ 4, 4, 2, 2 ], "presentation_list": [ 4, 4, 3, 2 ], "contribution_list": [ 2, 4, 1, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "uDZ9d4UAUh", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Akshay Nambi", "gender": "Male", "institution": "Microsoft Research", "country": "IN", "position": "Principal Researcher" }, { "name": "Joykirat Singh", "gender": "Male", "institution": "Research, Microsoft", "country": "IN", "position": "Researcher" }, { "name": "Vibhav Vineet", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 66, "n_ref_all": 93, "n_fig": 5, "n_tab": 14, "L_tab": 8847, "n_element_tab": 932, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 5151, "n_element_tab_1": 419, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 185015, "len_all_1": 80525, "len_abs": 2766, "len_title": 97, "len_sents": 54524, "len_sents_1": 34566, "n_sents": 416, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1301, "title": "Exposing the Achilles' Heel: Evaluating LLMs Ability to Handle Mistakes in Mathematical Reasoning", "abs": "Large Language Models (LLMs) have significantly impacted the field of Math Word Problems (MWPs), transforming how these problems are approached and solved, particularly in educational contexts. However, existing evaluations often focus on final accuracy, neglecting the critical aspect of reasoning capabilities. This work addresses that gap by evaluating LLMs' abilities to detect and correct reasoning mistakes. We present a novel dataset, MWP-MISTAKE, containing MWPs with both correct and incorrect reasoning steps generated through rule-based methods and smaller language models. Our comprehensive benchmarking of state-of-the-art models such as GPT-4o and GPT4 uncovers important insights into their strengths and limitations. While GPT-4o excels in mistake detection and rectification, gaps remain, particularly in handling complex datasets and novel problems. Additionally, we identify concerns with data contamination and memorization, which affect LLM reliability in real-world applications. While OpenAI's O1 model demonstrates 90\\% accuracy in reasoning and final answers on complex tasks, it still remains weak in mistake detection. Our findings highlight the need for improved reasoning evaluations and suggest ways to enhance LLM generalization and robustness in math problem-solving.", "keywords": [ "math reasoning", "mistake detection", "dataset", "benchmarking" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "uDXFOurrHM", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Gathika Ratnayaka", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "PhD student" }, { "name": "James Nichols", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "Lecturer" }, { "name": "Qing Wang", "gender": "Female", "institution": "Australian National University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 66, "n_formula_1": 20, "n_ref_uni": 31, "n_ref": 96, "n_ref_all": 154, "n_fig": 3, "n_tab": 6, "L_tab": 1806, "n_element_tab": 171, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1508, "n_element_tab_1": 93, "formula_len_all": 6119, "formula_len_all_1": 1376, "len_all": 161018, "len_all_1": 69508, "len_abs": 1446, "len_title": 109, "len_sents": 53296, "len_sents_1": 32098, "n_sents": 424, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1451, "title": "Learning Partial Graph Matching via Optimal Partial Transport", "abs": "Partial graph matching extends traditional graph matching by allowing some nodes to remain unmatched, enabling applications in more complex scenarios. However, this flexibility introduces additional complexity, as both the subset of nodes to match and the optimal mapping must be determined. While recent studies have explored deep learning techniques for partial graph matching, a significant limitation remains: the absence of an optimization objective that fully captures the problem\u2019s intrinsic nature while enabling efficient solutions. In this paper, we propose a novel optimization framework for partial graph matching, inspired by optimal partial transport. Our approach formulates an objective that enables partial assignments while incorporating matching biases, using weighted total variation as the divergence function to guarantee optimal partial assignments. Our method can achieve efficient, exact solutions within cubic worst case time complexity. Our contributions are threefold: (i) we introduce a novel optimization objective that balances matched and unmatched nodes; (ii) we establish a connection between partial graph matching and linear sum assignment problem, enabling efficient solutions; (iii) we propose a deep graph matching architecture with a novel partial matching loss, providing an end-to-end solution. The empirical evaluations on standard graph matching benchmarks demonstrate the efficacy of the proposed approach.", "keywords": [ "Graph Matching", "Optimal Transport" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "uDWcAajewK", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Chirag Gupta", "gender": "Male", "institution": "Bloomberg AI", "country": "", "position": "Researcher" }, { "name": "Rattana Pukdee", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Ziqi Ke", "gender": "Male", "institution": "Bloomberg", "country": "US", "position": "Researcher" } ], "n_formula": 139, "n_formula_1": 50, "n_ref_uni": 61, "n_ref": 91, "n_ref_all": 114, "n_fig": 15, "n_tab": 4, "L_tab": 529, "n_element_tab": 69, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 403, "n_element_tab_1": 44, "formula_len_all": 10256, "formula_len_all_1": 3164, "len_all": 224458, "len_all_1": 73403, "len_abs": 932, "len_title": 78, "len_sents": 65685, "len_sents_1": 30820, "n_sents": 612, "n_sents_1": 283, "venue": "ICLR2025", "decision": "Reject", "L_title": 30, "L_abs": 934, "title": "Learning from interval targets", "abs": "We consider regression problems where the exact real-valued targets are not directly available; instead, supervision is provided in the form of intervals around the targets\u2014that is, only lower and upper bounds are known. Such a \"learning from interval targets\" setup arises in domains where labeling costs are high or there is inherent uncertainty in the target values. In these settings, traditional regression loss functions, which require exact target values, cannot be directly applied. To address this challenge, we propose two approaches: (i) modifying the regression loss function to be compatible with interval ground truths, and (ii) formulating a min-max problem where we minimize the typical regression loss with respect to the \"worst-case\" label within the interval. We provide theoretical guarantees for our methods, analyze their computational efficiency, and evaluate their practical performance on real-world datasets.", "keywords": [ "Weak supervision", "Partial-label learning", "Learning from side information", "Learning theory" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 2, 3, 5, 4 ] }, { "paper_id": "uDIiL89ViX", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Aditya Ravuri", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Cian Eastwood", "gender": "Male", "institution": "Valence Labs powered by recursion", "country": "GB", "position": "Researcher" }, { "name": "Gemma Elyse Moran", "gender": "Female", "institution": "Rutgers University", "country": "", "position": "Assistant Professor" }, { "name": "Jason Hartford", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "Assistant Professor" }, { "name": "Kian Kenyon-Dean", "gender": "Male", "institution": "Recursion Pharma", "country": "CA", "position": "Researcher" }, { "name": "Konstantin Donhauser", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" }, { "name": "Kristina Ulicna", "gender": "Female", "institution": "Valence Labs | Recursion Pharma", "country": "GB", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 37, "n_ref": 70, "n_ref_all": 120, "n_fig": 24, "n_tab": 4, "L_tab": 572, "n_element_tab": 59, "n_fig_1": 11, "n_tab_1": 5, "L_tab_1": 639, "n_element_tab_1": 63, "formula_len_all": 177, "formula_len_all_1": 175, "len_all": 138598, "len_all_1": 66021, "len_abs": 1107, "len_title": 0, "len_sents": 45947, "len_sents_1": 32331, "n_sents": 292, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 119, "L_abs": 1115, "title": "Towards scientific discovery with dictionary learning: Extracting biological concepts from microscopy foundation models", "abs": "Dictionary learning (DL) has emerged as a powerful interpretability tool for large language models. By extracting known concepts (e.g., Golden-Gate Bridge) from human-interpretable data (e.g., text), sparse DL can elucidate a model's inner workings. In this work, we ask if DL can also be used to discover *unknown* concepts from less human-interpretable scientific data (e.g., cell images), ultimately enabling modern approaches to scientific discovery. As a first step, we use DL algorithms to study microscopy foundation models trained on multi-cell image data, where little prior knowledge exists regarding which high-level concepts should arise. We show that sparse dictionaries indeed extract biologically-meaningful concepts such as cell type and genetic perturbation type. We also propose a new DL algorithm, Iterative Codebook Feature Learning (ICFL), and combine it with a pre-processing step which uses PCA whitening from a control dataset. In our experiments, we demonstrate that both ICFL and PCA improve the selectivity or \"monosemanticity\" of extracted features compared to TopK sparse autoencoders.", "keywords": [ "mechanistic interpretability", "ViT", "MAE", "scientific discovery", "drug discovery", "biological representation learning" ], "rating_list": [ 5, 5, 5, 5, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3, 4 ], "confidence_list": [ 3, 4, 3, 5, 4 ] }, { "paper_id": "uD2yx2TR7S", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Aldeida Aleti", "gender": "Female", "institution": "Monash University", "country": "", "position": "Associate Professor" }, { "name": "Chunyang Chen", "gender": "unknown", "institution": "Monash University", "country": "", "position": "Associate Professor" }, { "name": "Hongyu Zhang", "gender": "Male", "institution": "University of Newcastle, Australia", "country": "AU", "position": "Full Professor" }, { "name": "Jian Gu", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 38, "n_ref": 57, "n_ref_all": 75, "n_fig": 6, "n_tab": 10, "L_tab": 5583, "n_element_tab": 390, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 3843, "n_element_tab_1": 299, "formula_len_all": 299, "formula_len_all_1": 348, "len_all": 237985, "len_all_1": 70301, "len_abs": 1879, "len_title": 135, "len_sents": 41853, "len_sents_1": 30540, "n_sents": 323, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1895, "title": "Vocabulary-Defined Semantics: Latent Space Clustering for Improving In-Context Learning", "abs": "In-context learning enables language models (LM) to adapt to downstream data or tasks by incorporating few samples as demonstrations within the prompts. It offers strong performance without the expense of fine-tuning.\nHowever, due to the context-length restriction, the demonstrations occupy only a small proportion of usable samples. The limitation exacerbates the difficulty of optimization, since the performance of in-context learning can be unstable depending on the quality, format, or order of demonstrations.\nPrior work, such as Knn Prompting, index samples based on the similarities of logits at the output-side, in addition to the regular retrieval operation at the input-side.\nThey improve in-context learning by leveraging the core ability of next-token prediction, rather than relying solely on the emergent capacity to make analogies.\nDespite this, the hard-to-optimize issue of in-context learning still exists. In our view, it stems from the process of selecting demonstrations. To address this, we propose complementing in-context learning with an additional clustering operation, making full use of all usable samples.\nWe propose a novel approach ``vocabulary-defined semantics''.\nGrounded in LM vocabulary, which is the label space of model outputs, the proposed approach computes semantically equivalent latent representations for output labels. Then, taking the representations as centroids, a clustering operation is performed to align the semantic properties between the language model and the downstream data/tasks.\nBased on extensive experiments across diverse textual understanding datasets and multiple models, our approach outperforms the state-of-the-art in terms of effectiveness and efficiency. On average, it achieves $3\\%-49\\%$ improvements via the clustering module, while requiring only half of the computation time via the similarity-based logits computation.", "keywords": [ "Language Model", "Latent Space", "In-Context Learning", "Semantics", "Disentanglement", "Neural Clustering" ], "rating_list": [ 3, 1, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 1, 1, 2, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 1, 3, 3 ] }, { "paper_id": "uCqxDfLYrB", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Chao-Han Huck Yang", "gender": "Male", "institution": "NVIDIA Research", "country": "TW", "position": "Researcher" }, { "name": "Ming Jin", "gender": "Male", "institution": "Griffith University", "country": "AU", "position": "Assistant Professor" }, { "name": "Qingren Yao", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Researcher" }, { "name": "Renhe Jiang", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Lecturer" }, { "name": "Shirui Pan", "gender": "unknown", "institution": "Griffith University", "country": "AU", "position": "Full Professor" }, { "name": "Yuxuan Liang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 6, "n_ref_uni": 41, "n_ref": 65, "n_ref_all": 89, "n_fig": 13, "n_tab": 3, "L_tab": 3127, "n_element_tab": 419, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 184, "n_element_tab_1": 26, "formula_len_all": 1055, "formula_len_all_1": 227, "len_all": 200038, "len_all_1": 58861, "len_abs": 1513, "len_title": 109, "len_sents": 58594, "len_sents_1": 28588, "n_sents": 473, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1506, "title": "Towards Neural Scaling Laws for Time Series Foundation Models", "abs": "Scaling laws offer valuable insights into the design of time series foundation models (TSFMs). However, previous research has largely focused on the scaling laws of TSFMs for in-distribution (ID) data, leaving their out-of-distribution (OOD) scaling behavior and the influence of model architectures less explored. In this work, we examine two common TSFM architectures\u2014encoder-only and decoder-only Transformers\u2014and investigate their scaling behavior on both ID and OOD data. These models are trained and evaluated across varying parameter counts, compute budgets, and dataset sizes. Our experiments reveal that the log-likelihood loss of TSFMs exhibits similar scaling behavior in both OOD and ID settings. We further compare the scaling properties across different architectures, incorporating two state-of-the-art TSFMs as case studies, showing that model architecture plays a significant role in scaling. The encoder-only Transformers demonstrate better scalability than the decoder-only Transformers, while the architectural enhancements in the two advanced TSFMs primarily improve ID performance but reduce OOD scalability. While scaling up TSFMs is expected to drive performance breakthroughs, the lack of a comprehensive understanding of TSFM scaling laws has hindered the development of a robust framework to guide model scaling. We fill this gap in this work by synthesizing our findings and providing practical guidelines for designing and scaling larger TSFMs with enhanced model capabilities.", "keywords": [ "Time series", "scaling law", "foundation model", "transformer", "forecasting" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "uClUUJk05H", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Kenji Fukumizu", "gender": "Male", "institution": "Preferred Networks, Inc.", "country": "", "position": "Technical Advisor" }, { "name": "Shoji Toyota", "gender": "Male", "institution": "Kyushu University", "country": "JP", "position": "Assistant Professor" }, { "name": "Jakob H Macke", "gender": "Male", "institution": "University of Tuebingen", "country": "", "position": "Full Professor" }, { "name": "Manuel Gl\u00f6ckler", "gender": "Male", "institution": "University of Tuebingen", "country": "DE", "position": "PhD student" } ], "n_formula": 49, "n_formula_1": 15, "n_ref_uni": 88, "n_ref": 182, "n_ref_all": 226, "n_fig": 29, "n_tab": 6, "L_tab": 1974, "n_element_tab": 267, "n_fig_1": 21, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3281, "formula_len_all_1": 1095, "len_all": 271294, "len_all_1": 71871, "len_abs": 1219, "len_title": 104, "len_sents": 75653, "len_sents_1": 31952, "n_sents": 640, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1224, "title": "Compositional simulation-based inference for time series", "abs": "Amortized simulation-based inference (SBI) methods train neural networks on simulated data to perform Bayesian inference. While this strategy avoids the need for tractable likelihoods, it often requires a large number of simulations and has been challenging to scale to time series data. Scientific simulators frequently emulate real-world dynamics through thousands of single-state transitions over time. We propose an SBI approach that can exploit such Markovian simulators by locally identifying parameters consistent with individual state transitions. We then compose these local results to obtain a posterior over parameters that align with the entire time series observation. We focus on applying this approach to neural posterior score estimation but also show how it can be applied, e.g., to neural likelihood (ratio) estimation. We demonstrate that our approach is more simulation-efficient than directly estimating the global posterior on several synthetic benchmark tasks and simulators used in ecology and epidemiology. Finally, we validate scalability and simulation efficiency of our approach by applying it to a high-dimensional Kolmogorov flow simulator with around one million dimensions in the data domain.", "keywords": [ "Simulation-based inference", "Bayesian inference", "time series", "markovian simulators", "Amortized Bayesian inference" ], "rating_list": [ 6, 6, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 4, 2 ], "presentation_list": [ 3, 2, 3, 4, 2 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 2, 4, 4, 4, 5 ] }, { "paper_id": "uC003NHlEi", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Fabrizio Ventola", "gender": "Male", "institution": "TU Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Jonas Seng", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Kristian Kersting", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Full Professor" }, { "name": "Zhongjie Yu", "gender": "unknown", "institution": "TU Darmstadt", "country": "", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 7, "n_ref_uni": 39, "n_ref": 85, "n_ref_all": 119, "n_fig": 19, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2514, "formula_len_all_1": 502, "len_all": 198366, "len_all_1": 71123, "len_abs": 1174, "len_title": 119, "len_sents": 72756, "len_sents_1": 35099, "n_sents": 573, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1177, "title": "Hyperparameter Optimization via Interacting with Probabilistic Circuits", "abs": "Despite the growing interest in designing truly interactive hyperparameter optimization (HPO) methods, to date, only a few allow to include human feedback. However, these methods add friction to the interactive process, rigidly requiring to fully specify the user input as prior distribution ex ante and often imposing additional constraints on the optimization framework. This hinders the flexible incorporation of expertise and valuable knowledge of domain experts, which might provide partial feedback at any time during optimization. To overcome these limitations, we introduce a novel Bayesian optimization approach leveraging tractable probabilistic models named probabilistic circuits (PCs) as surrogate model. PCs encode a tractable joint distribution over the hybrid hyperparameter space and evaluation scores, and enable exact conditional inference and sampling, allowing users to provide valuable insights interactively and generate configurations adhering to their feedback. We demonstrate the benefits of the resulting interactive HPO through an extensive empirical evaluation of diverse benchmarks, including the challenging setting of neural architecture search.", "keywords": [ "Hyperparameter Optimization", "Probabilistic Circuits", "Tractable Probabilistic Models" ], "rating_list": [ 5, 6, 3, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 2, 3, 2, 4 ], "confidence_list": [ 3, 2, 5, 3 ] }, { "paper_id": "uBxN9JA29p", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Aditya Puranik", "gender": "Male", "institution": "Santa Clara University", "country": "US", "position": "MS student" }, { "name": "David Jeong", "gender": "Male", "institution": "Santa Clara University", "country": "US", "position": "Assistant Professor" }, { "name": "Harrison Bui", "gender": "Male", "institution": "Santa Clara University", "country": "US", "position": "MS student" }, { "name": "James Vong", "gender": "Male", "institution": "Santa Clara University", "country": "US", "position": "MS student" }, { "name": "Julianna Dietrich", "gender": "Female", "institution": "Santa Clara University", "country": "US", "position": "MS student" }, { "name": "Ryan Fell", "gender": "Male", "institution": "Santa Clara University", "country": "US", "position": "Undergrad student" }, { "name": "Tianma Shen", "gender": "unknown", "institution": "Santa Clara University", "country": "US", "position": "PhD student" }, { "name": "Vrushabh Abhijit Deogirikar", "gender": "unknown", "institution": "Santa Clara University", "country": "US", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 32, "n_ref": 127, "n_ref_all": 142, "n_fig": 4, "n_tab": 2, "L_tab": 253, "n_element_tab": 38, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1113, "n_element_tab_1": 94, "formula_len_all": 437, "formula_len_all_1": 564, "len_all": 105547, "len_all_1": 53151, "len_abs": 1060, "len_title": 132, "len_sents": 31705, "len_sents_1": 25518, "n_sents": 205, "n_sents_1": 142, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1065, "title": "SoloPose: One-Stage Kinematic 3D Human Pose Estimation with Mocap Data Augmentation", "abs": "While recent two-stage many-to-one deep learning models have demonstrated great success in 3D human pose estimation, such models are inefficient in 3D key point detection and also tend to pass on first stage errors onto the second stage. In this paper, we introduce SoloPose, a novel one-stage, many-to-many spatio-temporal transformer model for kinematic 3D human pose estimation of video. SoloPose is further fortified by HeatPose, a 3D heatmap based on Gaussian Mixture Model distributions that factors target key points as well as kinematically adjacent key points. Finally, we address data diversity constraints with the 3D AugMotion Toolkit, a methodology to augment existing 3D human pose datasets, specifically by projecting four top public 3D human pose datasets (Human3.6M, MADS, AIST Dance++, MPI INF 3DHP) into a novel dataset (Human7.1M) with a universal coordinate system. Extensive experiments are conducted on both Human3.6M and the augmented Human7.1M dataset, and SoloPose demonstrates superior results relative to the state-of-the-art approaches.", "keywords": [ "monocular pose estimation", "data augmentation" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "uBnVA7SeWv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "AMALJITH E V", "gender": "Male", "institution": "Indian Institute of Technology Kharagpur", "country": "IN", "position": "PhD student" }, { "name": "Arindam Biswas", "gender": "Male", "institution": "Polynom", "country": "FR", "position": "Researcher" }, { "name": "BISWAJIT BASU", "gender": "Male", "institution": "Trinity College, Dublin", "country": "IE", "position": "Full Professor" }, { "name": "Pabitra Mitra", "gender": "Male", "institution": "Indian Institute of Technology Kharagpur", "country": "IN", "position": "Full Professor" }, { "name": "Suryam Arnav Kalra", "gender": "unknown", "institution": "Indian Institute of Technology Kharagpur", "country": "", "position": "Undergrad student" } ], "n_formula": 16, "n_formula_1": 5, "n_ref_uni": 39, "n_ref": 67, "n_ref_all": 94, "n_fig": 3, "n_tab": 10, "L_tab": 3402, "n_element_tab": 526, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 2456, "n_element_tab_1": 400, "formula_len_all": 1083, "formula_len_all_1": 413, "len_all": 152886, "len_all_1": 65034, "len_abs": 2068, "len_title": 86, "len_sents": 40064, "len_sents_1": 26663, "n_sents": 295, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 38, "L_abs": 623, "title": "XoRA: Expander adapted LoRA finetuning", "abs": "Parameter-efficient fine-tuning aims to reduce the computational cost of adapting foundational models to downstream tasks. Low-rank matrix based adaptation (LoRA) techniques are popular for this purpose. We propose XoRA, an efficient fine-tuning scheme, which sparsifies the low-rank matrices even further using expander masks. The mask is generated using extremal expander graphs (Ramanujan graphs) to maintain high edge connectivity even at a very high sparsity. Experimental results demonstrate that this method has comparable performance with the LoRA fine-tuning method while retaining much fewer number of parameters.", "keywords": [ "LoRA fine-tuning", "parameter efficient fine-tuning", "expander masks" ], "rating_list": [ 3, 5, 6, 3, 6 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 1, 4 ], "confidence_list": [ 3, 4, 3, 5, 5 ] }, { "paper_id": "uBnM3EFovQ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Lingpeng Kong", "gender": "Male", "institution": "Department of Computer Science, The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Qi Liu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Zhenguo Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Zhihui Xie", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "JIAHUI GAO", "gender": "Female", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Tobias Lee", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 53, "n_ref": 111, "n_ref_all": 164, "n_fig": 10, "n_tab": 11, "L_tab": 898, "n_element_tab": 104, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 1513, "n_element_tab_1": 148, "formula_len_all": 765, "formula_len_all_1": 702, "len_all": 189063, "len_all_1": 69267, "len_abs": 1307, "len_title": 97, "len_sents": 57815, "len_sents_1": 30925, "n_sents": 439, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1313, "title": "Jailbreaking as a Reward Misspecification Problem", "abs": "The widespread adoption of large language models (LLMs) has raised concerns about their safety and reliability, particularly regarding their vulnerability to adversarial attacks. In this paper, we propose a new perspective that attributes this vulnerability to reward misspecification during the alignment process. This misspecification occurs when the reward function fails to accurately capture the intended behavior, leading to misaligned model outputs. We introduce a metric ReGap to quantify the extent of reward misspecification and demonstrate its effectiveness and robustness in detecting harmful backdoor prompts. Building upon these insights, we present ReMiss, a system for automated red teaming that generates adversarial prompts in a reward-misspecified space. ReMiss achieves state-of-the-art attack success rates on the AdvBench benchmark against various target aligned LLMs while preserving the human readability of the generated prompts. Furthermore, these attacks on open-source models demonstrate high transferability to closed-source models like GPT-4o and out-of-distribution tasks from HarmBench. Detailed analysis highlights the unique advantages of the proposed reward misspecification objective compared to previous methods, offering new insights for improving LLM safety and robustness.", "keywords": [ "Large language models", "alignment", "jailbreaking" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "uBhqll8pw1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Amlan Kar", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Research Scientist" }, { "name": "David Acuna", "gender": "Male", "institution": "NVIDIA", "country": "CA", "position": "Researcher" }, { "name": "Sanja Fidler", "gender": "Female", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "Associate Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 29, "n_ref": 119, "n_ref_all": 151, "n_fig": 14, "n_tab": 13, "L_tab": 6351, "n_element_tab": 701, "n_fig_1": 12, "n_tab_1": 3, "L_tab_1": 766, "n_element_tab_1": 94, "formula_len_all": 301, "formula_len_all_1": 0, "len_all": 222966, "len_all_1": 73592, "len_abs": 3055, "len_title": 107, "len_sents": 71048, "len_sents_1": 36277, "n_sents": 567, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1866, "title": "On Inherent 3D Reasoning of VLMs in Indoor Scene Layout Design", "abs": "Large vision-language models (VLMs) such as GPT-4o, Llama-3.2 have shown remarkable capabilities in visual understanding and reasoning, prompting us to test their off-the-shelf ability to reason and act as a 3D design assistant. This study investigates VLMs\u2019 visual reasoning capabilities using 3D indoor scene layout synthesis i.e. placement of furniture in a room, as a test-bed. We study three key primitive abilities in this context: (1) communication of spatial locations, (2) reasoning about free space and object collision, and (3) reasoning about object alignment, orientation, and functionality, each crucial to creating a VLM agent-based scene layout synthesis pipeline. We evaluate five state-of-the-art VLMs, both proprietary and open, on a new dataset incorporating 3400 questions that assess VLMs\u2019 current visual reasoning abilities in our context. Our findings reveal several remarkable insights: (1) VLMs consistently prefer normalized coordinates for spatial communication over absolute coordinates or pointing with image markers. (2) Contrary to expectations, VLMs perform best with simplified sketch based scene representation or, most strikingly, with no visual input at all, compared to detailed renderings. (3) Free space reasoning remains challenging, with performance only slightly above random guessing, though frontier models show significant improvement with collision checking tools. Surprisingly, free space reasoning with clear visible collisions in the image can also fail. (4) Reasoning about object alignment, size, orientation and functionality together compounds errors leading to near chance performance on our dataset. These findings serve to offer insights into current potential and limitations of using VLMs off-the-shelf towards developing advanced visual assistants capable of understanding and manipulating 3D environments.", "keywords": [ "VLMs", "Evaluation", "3D" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "uBcx1aFpXy", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Bingbing Liu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "R&D Manager" }, { "name": "Jing He", "gender": "Female", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Kaiqiang Zhou", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Leheng Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Qing LIAN", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Weichao Qiu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xu Yan", "gender": "Male", "institution": "The Chinese University of Hongkong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Ying-Cong Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Yingjie CAI", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 27, "n_ref": 67, "n_ref_all": 94, "n_fig": 15, "n_tab": 4, "L_tab": 676, "n_element_tab": 72, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 631, "n_element_tab_1": 68, "formula_len_all": 285, "formula_len_all_1": 338, "len_all": 141751, "len_all_1": 58464, "len_abs": 1346, "len_title": 131, "len_sents": 45110, "len_sents_1": 26333, "n_sents": 365, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1353, "title": "OmniBooth: Learning Latent Control for Image Synthesis with Multi-modal Instruction", "abs": "We present OmniBooth, an image generation framework that enables spatial control with instance-level multi-modal customization. For all instances, the multi-modal instruction can be described through text prompts or image references. Given a set of user-defined masks and associated text or image guidance, our objective is to generate an image, where multiple objects are positioned at specified coordinates and their attributes are precisely aligned with the corresponding guidance. This approach significantly expands the scope of text-to-image generation, and elevates it to a more versatile and practical dimension in controllability. In this paper, our core contribution lies in the proposed latent control signals, a high-dimensional spatial feature that provides a unified representation to integrate the spatial, textual, and image conditions seamlessly. The text condition extends ControlNet to provide instance-level open-vocabulary generation. The image condition further enables fine-grained control with personalized identity. In practice, our method empowers users with more flexibility in controllable generation, as users can choose multi-modal conditions from text or images as needed. Furthermore, thorough experiments demonstrate our enhanced performance in image synthesis fidelity and alignment across different tasks and datasets.", "keywords": [ "Image Synthesis", "Diffusion Model", "Controllable Generation", "Spatial Control", "Multi-modal Instruction" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 4, 2, 4 ] }, { "paper_id": "uBai0ukstY", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "An Nguyen The", "gender": "Male", "institution": "FPT Software AI Center", "country": "VN", "position": "Researcher" }, { "name": "Duy-Tung Pham", "gender": "Male", "institution": "AI Center, FPT Software", "country": "VN", "position": "Researcher" }, { "name": "Hoang V. Tran", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Minh-Khoi Nguyen-Nhat", "gender": "Male", "institution": "FPT", "country": "VN", "position": "Researcher" }, { "name": "Tan Minh Nguyen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Thanh Tran", "gender": "Male", "institution": "VinUniversity", "country": "", "position": "Undergrad student" }, { "name": "Thieu Vo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Tho Tran Huu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 270, "n_formula_1": 38, "n_ref_uni": 64, "n_ref": 103, "n_ref_all": 154, "n_fig": 2, "n_tab": 9, "L_tab": 5059, "n_element_tab": 240, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1161, "n_element_tab_1": 86, "formula_len_all": 48686, "formula_len_all_1": 4149, "len_all": 339868, "len_all_1": 64479, "len_abs": 3062, "len_title": 103, "len_sents": 74617, "len_sents_1": 25496, "n_sents": 904, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1388, "title": "Equivariant Neural Functional Networks for Transformers", "abs": "This paper systematically explores neural functional networks (NFN) for transformer architectures. NFN are specialized neural networks that treat the weights, gradients, or sparsity patterns of a deep neural network (DNN) as input data and have proven valuable for tasks such as learnable optimizers, implicit data representations, and weight editing. While NFN have been extensively developed for MLP and CNN, no prior work has addressed their design for transformers, despite the importance of transformers in modern deep learning. This paper aims to address this gap by providing a systematic study of NFN for transformers. We first determine the maximal symmetric group of the weights in a multi-head attention module as well as a necessary and sufficient condition under which two sets of hyperparameters of the multi-head attention module define the same function. We then define the weight space of transformer architectures and its associated group action, which leads to the design principles for NFN in transformers. Based on these, we introduce Transformer-NFN, an NFN that is equivariant under this group action. Additionally, we release a dataset of more than 125,000 Transformers model checkpoints trained on two datasets with two different tasks, providing a benchmark for evaluating Transformer-NFN and encouraging further research on transformer training and performance.", "keywords": [ "neural functional network", "transformer", "maximal symmetric group", "equivariant model", "dataset" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 2, 4, 4 ] }, { "paper_id": "uBMNOjqHUV", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Arthur Ledaguenel", "gender": "Male", "institution": "CentraleSupelec", "country": "", "position": "PhD student" }, { "name": "C\u00e9line Hudelot", "gender": "Female", "institution": "CentraleSupelec", "country": "", "position": "Full Professor" }, { "name": "Mostepha Redouane Khouadjia", "gender": "Male", "institution": "IRT SystemX", "country": "FR", "position": "Researcher" } ], "n_formula": 28, "n_formula_1": 15, "n_ref_uni": 30, "n_ref": 97, "n_ref_all": 105, "n_fig": 3, "n_tab": 2, "L_tab": 897, "n_element_tab": 41, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 302, "n_element_tab_1": 5, "formula_len_all": 1516, "formula_len_all_1": 613, "len_all": 174323, "len_all_1": 64878, "len_abs": 1472, "len_title": 92, "len_sents": 61592, "len_sents_1": 32258, "n_sents": 461, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1477, "title": "A bird's eye view on informed classification", "abs": "Neurosymbolic AI is a growing field of research aiming to combine neural network learning capabilities with the reasoning abilities of symbolic systems. In this paper, we tackle informed classification tasks, i.e. multi-label classification tasks informed by prior knowledge that specifies which combinations of labels are semantically valid. Several neurosymbolic formalisms and techniques have been introduced in the literature, each relying on a particular language to represent prior knowledge. We take a bird's eye view on informed classification and introduce a unified formalism that encapsulates all knowledge representation languages. Then, we build upon this formalism to identify several concepts in probabilistic reasoning that are at the core of many techniques across representation languages. We also define a new technique called semantic conditioning at inference, which only constrains the system during inference while leaving the training unaffected, an interesting property in the era of off-the-shelves and foundation models. We discuss its theoritical and practical advantages over two other probabilistic neurosymbolic techniques: semantic conditioning and semantic regularization. We then evaluate experimentally and compare the benefits of all three techniques on several large-scale datasets. Our results show that, despite only working at inference, our technique can efficiently leverage prior knowledge to build more accurate neural-based systems.", "keywords": [ "Neurosymbolic", "Deep Learning", "Knowledge Representation", "Probabilistic Reasoning" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "uBEl8DMA8K", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Giacomo Tarroni", "gender": "Male", "institution": "City, University of London", "country": "", "position": "Lecturer" }, { "name": "Sergio Naval Marimont", "gender": "Male", "institution": "City University", "country": "GB", "position": "PhD student" }, { "name": "Vasilis Siomos", "gender": "Male", "institution": "City University", "country": "GB", "position": "PhD student" }, { "name": "Jonathan Passerat-palmbach", "gender": "unknown", "institution": "Flashbots", "country": "GB", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 55, "n_ref": 125, "n_ref_all": 150, "n_fig": 4, "n_tab": 12, "L_tab": 4738, "n_element_tab": 321, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2292, "n_element_tab_1": 79, "formula_len_all": 663, "formula_len_all_1": 833, "len_all": 185777, "len_all_1": 67180, "len_abs": 1602, "len_title": 154, "len_sents": 53217, "len_sents_1": 32719, "n_sents": 359, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 106, "L_abs": 1614, "title": "Addressing Data Heterogeneity In Federated Learning With Adaptive Normalization-Free Feature Recalibration", "abs": "Federated learning is a decentralized collaborative training paradigm preserving stakeholders\u2019 data ownership while improving performance and generalization. However, statistical heterogeneity among client datasets degrades system performance. To address this issue, we propose **Adaptive Normalization-free Feature Recalibration (ANFR)**, an architecture-level approach that combines weight standardization and channel attention. Weight standardization normalizes the weights of layers, making it less prone to mismatched client statistics and inconsistent averaging, ensuring robustness under heterogeneity. Channel attention produces learnable scaling factors for feature maps, suppressing inconsistencies across clients due to heterogeneity. We demonstrate that combining these techniques boosts model performance beyond their individual contributions, by improving class selectivity and channel attention weight distribution. ANFR works with any aggregation method, supports both global and personalized FL, and adds minimal overhead. When training with differential privacy, ANFR achieves an appealing balance between privacy and utility, enabling strong privacy guarantees without sacrificing performance. By integrating weight standardization and channel attention in the backbone model, ANFR offers a novel and versatile approach to the challenge of statistical heterogeneity. Extensive experiments show ANFR consistently outperforms established baselines across various aggregation methods, datasets, and heterogeneity conditions. Code is provided at https://anonymous.4open.science/r/anfr_iclr_updated/.", "keywords": [ "Federated Learning", "Deep Learning", "Convolutional Neural Network", "Image Classification" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "uAtDga3q0r", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Daniel Sorvisto", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Jerry Weihong Liu", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Roberto Garcia", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Sabri Eyuboglu", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 20, "n_ref_uni": 26, "n_ref": 69, "n_ref_all": 92, "n_fig": 10, "n_tab": 3, "L_tab": 1128, "n_element_tab": 173, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 650, "n_element_tab_1": 100, "formula_len_all": 1168, "formula_len_all_1": 1124, "len_all": 142166, "len_all_1": 71715, "len_abs": 1252, "len_title": 125, "len_sents": 38938, "len_sents_1": 33350, "n_sents": 294, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1264, "title": "Adaptive Rank Allocation: Speeding Up Modern Transformers with RaNA Adapters", "abs": "Large Language Models (LLMs) are computationally intensive, particularly during inference. Neuron-adaptive techniques, which selectively activate neurons in Multi-Layer Perceptron (MLP) layers, offer some speedups but suffer from limitations in modern Transformers. These include reliance on sparse activations, incompatibility with attention layers, and the use of costly neuron masking techniques. To address these issues, we propose the Adaptive Rank Allocation framework and introduce the Rank and Neuron Allocator (RaNA) adapter. RaNA adapters leverage rank adapters, which operate on linear layers by applying both low-rank matrix decompositions and adaptive masking to efficiently allocate compute without depending on activation sparsity. This enables RaNA to be generally applied to MLPs and linear components of attention modules, while eliminating the need for expensive maskers found in neuron-adaptive methods. Notably, when compared to neuron adapters, RaNA improves perplexity by up to 7 points and increases accuracy by up to 8 percentage-points when reducing FLOPs by $\\sim$44\\% in state-of-the-art Transformer architectures. These results position RaNA as a robust solution for improving inference efficiency in modern Transformer architectures.", "keywords": [ "Large Language Models", "Adaptive compute", "Rank adapters", "Neuron adapters" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "uApm5otXfH", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Jian Chen", "gender": "unknown", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Jin Huang", "gender": "Male", "institution": "South China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Yile Chen", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Zeyi Wen", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 38, "n_ref": 58, "n_ref_all": 90, "n_fig": 11, "n_tab": 10, "L_tab": 4702, "n_element_tab": 242, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1699, "n_element_tab_1": 61, "formula_len_all": 426, "formula_len_all_1": 383, "len_all": 144056, "len_all_1": 62998, "len_abs": 1430, "len_title": 88, "len_sents": 42631, "len_sents_1": 27802, "n_sents": 311, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1428, "title": "Interleaving Optimizers for DNN Training", "abs": "Optimizers are crucial in deep neural network (DNN) training, affecting model quality and convergence. Researchers have found that different optimizers often suit different problems or different stages of a problem. Hence, some studies have tried to combine different optimizers to better train DNNs. However, existing methods are limited to simple optimizer switch strategies, which leads to unstable model quality and slow convergence. In this paper, we propose a fine-grain optimizer switch method called Iterleaving Optimizer for Model Training (IOMT), which automatically switches to the appropriate optimizer and hyperparameters based on the training stage information, achieving faster convergence and better model quality. IOMT employs surrogate models to estimate the performance of different optimizers during training and is supported by a transferability assessment to predict the training cost. By combining the transferability assessment, performance estimation, and training process information with an acquisition function, IOMT calculates the optimization gain of each optimizer and switches the optimizer with the largest gain for the next training stage. The experimental results on full training and fine-tuning demonstrate that IOMT achieves faster convergence (e.g., 10\\% on the *stl10* dataset) and better performance (e.g., 3\\% accuracy improvement on the *cifar10* dataset) compared to existing methods.", "keywords": [ "Optimizer", "DNN", "HPO" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "uAFHCZRmXk", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Simon Schrodi", "gender": "Male", "institution": "University of Freiburg, Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Thomas Brox", "gender": "Male", "institution": "University of Freiburg", "country": "", "position": "Full Professor" }, { "name": "Volker Fischer", "gender": "Male", "institution": "Bosch Center for Artificial Intelligence", "country": "", "position": "Postdoc" }, { "name": "David T Hoffmann", "gender": "Male", "institution": "University of Freiburg, Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Max Argus", "gender": "Male", "institution": "University of Freiburg, Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 3, "n_ref_uni": 57, "n_ref": 126, "n_ref_all": 198, "n_fig": 25, "n_tab": 10, "L_tab": 2905, "n_element_tab": 241, "n_fig_1": 14, "n_tab_1": 2, "L_tab_1": 920, "n_element_tab_1": 88, "formula_len_all": 1013, "formula_len_all_1": 258, "len_all": 223717, "len_all_1": 71395, "len_abs": 1677, "len_title": 171, "len_sents": 71561, "len_sents_1": 33412, "n_sents": 617, "n_sents_1": 267, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 123, "L_abs": 1685, "title": "Two Effects, One Trigger: On the Modality Gap, Object Bias, and Information Imbalance in Contrastive Vision-Language Models", "abs": "Contrastive vision-language models (VLMs), like CLIP, have gained popularity for their versatile applicability to various downstream tasks. Despite their successes in some tasks, like zero-shot object recognition, they perform surprisingly poor on other tasks, like attribute recognition. Previous work has attributed these challenges to the modality gap, a separation of image and text in the shared representation space, and to a bias towards objects over other factors, such as attributes. In this analysis paper, we investigate both phenomena thoroughly. We evaluated off-the-shelf VLMs and find that while the gap's influence on performance is typically overshadowed by other factors, we find indications that closing the gap indeed leads to improvements. Moreover, we find that, contrary to intuition, only few embedding dimensions drive the gap and that the embedding spaces are differently organized. To allow for a clean study of object bias, we introduce a definition and a corresponding measure of it. Equipped with this tool, we find that object bias does not lead to worse performance on other concepts, such as attributes per se. However, why do both phenomena, modality gap and object bias, emerge in the first place? To answer this fundamental question and uncover some of the inner workings of contrastive VLMs, we conducted experiments that allowed us to control the amount of shared information between the modalities. These experiments revealed that the driving factor behind both the modality gap and the object bias, is an information imbalance between images and captions, and unveiled an intriguing connection between the modality gap and entropy of the logits.", "keywords": [ "CLIP", "modality gap", "object bias", "contrastive loss", "data-centric", "vision language models", "VLM" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "u9Z6gL5MlL", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Hadi Abdullah", "gender": "unknown", "institution": "VISA", "country": "US", "position": "Researcher" }, { "name": "Ke Wang", "gender": "unknown", "institution": "Visa Research", "country": "US", "position": "Research Scientist" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 101, "n_ref_all": 118, "n_fig": 5, "n_tab": 2, "L_tab": 236, "n_element_tab": 48, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 236, "n_element_tab_1": 48, "formula_len_all": 61, "formula_len_all_1": 61, "len_all": 125611, "len_all_1": 70392, "len_abs": 1219, "len_title": 120, "len_sents": 39056, "len_sents_1": 34637, "n_sents": 303, "n_sents_1": 270, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1232, "title": "Back to Fundamentals: Re-Examining Memorization in Deep Learning Models", "abs": "In supervised training, memorization is the ability of deep learning models to assign arbitrary ground truth labels to inputs in the dataset. Due to the computa- tional difficulty of identifying existing memorized points, researchers often induce artificial memorization i.e, force the model to memorize the newly introduced points (via Noisy Label or Noisy Input). However, in this work, we show that this artificial proxy exhibits fundamentally different characteristics than the mem- orization real points (or natural memorization). To demonstrate this deviation, we re-examine two key findings derived from artificial memorization and com- pare them against natural memorization i.e., over-parametrization and increased training time increases memorization. We show that both these factors have the opposite effect i.e., they reduce natural memorization. Additionally, we find that memorization and train-test gap are strongly correlated (Pearson score 0.99). As a result, memorization is not necessary for generalization. Since real world models suffer from natural memorization (instead of the artificial one) our findings sug- gest the research community should focus on natural memorization, instead of the artificial proxy.", "keywords": [ "Memorization" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "u8y7wYYs2D", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Changshuo Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Prayag Tiwari", "gender": "Male", "institution": "Halmstad University", "country": "SE", "position": "Associate Professor" }, { "name": "Xiang Fang", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xingyu Gao", "gender": "unknown", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhijian Hu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Zhonghang LIU", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 25, "n_ref_uni": 28, "n_ref": 59, "n_ref_all": 78, "n_fig": 10, "n_tab": 9, "L_tab": 2375, "n_element_tab": 184, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 4746, "n_element_tab_1": 158, "formula_len_all": 1199, "formula_len_all_1": 924, "len_all": 122017, "len_all_1": 60518, "len_abs": 1436, "len_title": 90, "len_sents": 34769, "len_sents_1": 22481, "n_sents": 273, "n_sents_1": 166, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1440, "title": "Dynamic Taylor Convolutional Neural Network for Few-Shot Point Cloud Semantic Segmentation", "abs": "Few-shot Point cloud semantic segmentation remains a challenge in the field of computer vision due to the limitations of the pre-training learning paradigm and insufficient local geometric structure representation. To address this issue, we propose a novel pre-training-free Dynamic Taylor Convolutional Neural Network, called DyTaylorCNN ingeniously, which combines the potential of the Taylor series in local structure representation with the flexibility and adaptability of dynamic convolutions. The core of DyTaylorCNN lies in two innovative components: the Dynamic Taylor Convolution (DyTaylorConv) and the Interactive Prototype Refinement (IPR) Module. Inspired by the Taylor series and dynamic convolution, DyTaylorConv performs local structure fitting by collaborating between the Low-order Convolution (LoConv) and the Dynamic High-order Convolution (DyHiConv). LoConv is designed based on position encoding, focusing on extracting the basic geometric information of point clouds, while DyHiConv adaptively models complex local geometric features by learning spatial priors to generate dynamic weights. Moreover, the IPR Module effectively reduces the domain distribution gap by learning fine-grained prototype features, further enhancing the model's generalization capability. Experimental results on multiple benchmark datasets demonstrate that the proposed DyTaylorCNN significantly outperforms current state-of-the-art methods.", "keywords": [ "Point Cloud Semantic Segmentation; Few-Shot Learning; Taylor Series; Dynamic Convolution; Prototype Refinement" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "u8VOQVzduP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "HaoYang Shang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Jie ZHANG", "gender": "Female", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Quanyan Zhu", "gender": "Male", "institution": "New York University", "country": "", "position": "Associate Professor" }, { "name": "Song Guo", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Xuan Liu", "gender": "Not Specified", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Undergrad student" }, { "name": "Yang Chengxu", "gender": "Male", "institution": "Wuhan University of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 86, "n_ref_all": 93, "n_fig": 4, "n_tab": 7, "L_tab": 4413, "n_element_tab": 278, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1035, "n_element_tab_1": 63, "formula_len_all": 103, "formula_len_all_1": 73, "len_all": 182791, "len_all_1": 62932, "len_abs": 88, "len_title": 73, "len_sents": 56824, "len_sents_1": 31720, "n_sents": 430, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1274, "title": "Exploring Prosocial Irrationality for LLM Agents: A Social Cognition View", "abs": "Large language models (LLMs) have been shown to face hallucination issues due to the data they trained on often containing human bias; whether this is reflected in the decision-making process of LLM agents remains under-explored. As LLM Agents are increasingly employed in intricate social environments, a pressing and natural question emerges: Can we utilize LLM Agents' systematic hallucinations to mirror human cognitive biases, thus exhibiting irrational social intelligence? In this paper, we probe the irrational behavior among contemporary LLM agents by melding practical social science experiments with theoretical insights. Specifically, we propose CogMir, an open-ended Multi-LLM Agents framework that utilizes hallucination properties to assess and enhance LLM Agents\u2019 social intelligence through cognitive biases. Experimental results on CogMir subsets show that LLM Agents and humans exhibit high consistency in irrational and prosocial decision-making under uncertain conditions, underscoring the prosociality of LLM Agents as social entities and highlighting the significance of hallucination properties. Additionally, CogMir framework demonstrates its potential as a valuable platform for encouraging more research into the social intelligence of LLM Agents.", "keywords": [ "LLM for Social Science", "Prosocial irrationality of LLM Agents", "Cognitive AI framework" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "u8SYRtXDsZ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Fei Li", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "JiaZhong Yu", "gender": "Male", "institution": "China Tower Corporation Limited", "country": "CN", "position": "Principal Researcher" }, { "name": "LinSu Shi", "gender": "Male", "institution": "South University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Qi Yang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Qinghua Liang", "gender": "Male", "institution": "China Tower Corporation", "country": "CN", "position": "Researcher" }, { "name": "Shiming Xiang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zili Wang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 9, "n_ref_uni": 26, "n_ref": 86, "n_ref_all": 105, "n_fig": 9, "n_tab": 9, "L_tab": 1983, "n_element_tab": 405, "n_fig_1": 7, "n_tab_1": 7, "L_tab_1": 3210, "n_element_tab_1": 307, "formula_len_all": 1050, "formula_len_all_1": 440, "len_all": 139024, "len_all_1": 52762, "len_abs": 1342, "len_title": 128, "len_sents": 30036, "len_sents_1": 19894, "n_sents": 274, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1281, "title": "AVESFormer: Efficient Transformer Design for Real-Time Audio-Visual Segmentation", "abs": "Recently, Transformer-based models have performed remarkably well in audio-visual segmentation (AVS) tasks.\n However, previous methods exhibit abnormal behavior and unsatisfactory results when using cross-attention.\n By analyzing attention maps, we identify two primary challenges in existing AVS models: 1) attention dissipation, caused by anomalous attention weights after Softmax over limited frames, and 2) narrow attention patterns in early decoder stages lead to inefficient utilization of attention mechanism.\n In this paper, we introduce AVESFormer, the first real-time audio-visual segmentation transformer that simultaneously achieves fast, efficient, and lightweight.\n Our model proposes an efficient, prompt query generator to rectify cross-attention behavior. \n Moreover, we propose an early focus (ELF) decoder, which enhances efficiency by incorporating convolution operations tailored for local feature extraction, thus reducing computational overhead.\n Extensive experiments demonstrate that AVESFormer effectively mitigates cross-attention issues, substantially improves attention utilization, and outperforms the previous state-of-the-art, achieving a superior trade-off between performance and speed.\n The code can be found in the supplementary material.", "keywords": [ "Audio-Visual Segmentation", "real-time model", "efficient architecture" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "u7oY4kPKyN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chenkai ZHANG", "gender": "Male", "institution": "Nagoya University", "country": "JP", "position": "PhD student" }, { "name": "Daisuke Deguchi", "gender": "Male", "institution": "Nagoya University", "country": "JP", "position": "Associate Professor" }, { "name": "Hiroshi Murase", "gender": "Male", "institution": "Nagoya University", "country": "JP", "position": "Emeritus" }, { "name": "Jialei Chen", "gender": "Male", "institution": "Nagoya University", "country": "JP", "position": "PhD student" }, { "name": "Qi Fan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xu Zheng", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 46, "n_ref": 114, "n_ref_all": 132, "n_fig": 5, "n_tab": 5, "L_tab": 535, "n_element_tab": 56, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 417, "n_element_tab_1": 45, "formula_len_all": 392, "formula_len_all_1": 314, "len_all": 127866, "len_all_1": 65162, "len_abs": 1710, "len_title": 118, "len_sents": 34872, "len_sents_1": 31113, "n_sents": 257, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1672, "title": "CLIP-to-Seg Distillation for Inductive Zero-shot Semantic Segmentation", "abs": "CLIP has greatly advanced zero-shot segmentation by leveraging its strong visual-language association and generalization capability. However, directly adapting CLIP for segmentation often yields suboptimal results due to inconsistencies between image and pixel-level prediction objectives. Additionally, merely combining segmentation and CLIP models often leads to disjoint optimization, introducing significant computational overhead and additional parameters. To address these issues, we propose a novel CLIP-to-Seg Distillation approach, incorporating global and local distillation to flexibly transfer CLIP\u2019s powerful zero-shot generalization capability to existing closed-set segmentation models. Global distillation leverages CLIP\u2019s CLS token to condense segmentation features and distills high-level concepts to the segmentation model via image-level prototypes. Local distillation adapts CLIP\u2019s local semantic transferability to dense prediction tasks using object-level features, aided by pseudo-mask generation for latent unseen class mining. To further generalize the CLIP-distilled segmentation model, we generate latent embeddings for the mined latent classes by coordinating their semantic embeddings and dense features. Our method equips existing closed-set segmentation models with strong generalization capabilities for open concepts through effective and flexible CLIP-to-Seg distillation. Without relying on the CLIP model or adding extra computational overhead/parameters during inference, our method can be seamlessly integrated into existing segmentation models and achieves state-of-the-art performance on multiple zero-shot segmentation benchmarks.", "keywords": [ "zero-shot learning", "zero-shot segmentation", "semantic segmentation", "knowledge distillation" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "u6y9uIzqAB", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Chenru Wen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Divyansh Srivastava", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "He Wen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Xiang Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Zhuowen Tu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Professor" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 36, "n_ref": 89, "n_ref_all": 104, "n_fig": 7, "n_tab": 8, "L_tab": 1169, "n_element_tab": 131, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 3726, "n_element_tab_1": 218, "formula_len_all": 327, "formula_len_all_1": 230, "len_all": 132924, "len_all_1": 57800, "len_abs": 1170, "len_title": 0, "len_sents": 34990, "len_sents_1": 23327, "n_sents": 268, "n_sents_1": 164, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1175, "title": "Lay-Your-Scene: Open-Vocabulary Text to Layout Generation", "abs": "We present Lay-Your-Scene (shorthand LayouSyn), a novel diffusion-Transformer based architecture for open-vocabulary natural scene layout generation. Prior works have used close-sourced scene-unaware Large Language models for open-vocabulary layout generation, limiting their widespread use and scene-specific modeling capability. This work presents the first end-to-end text-to-natural-scene-layout generation pipeline that utilizes lightweight open-source language models to predict objects in the scene and a new conditional layout diffusion Transformer trained in a scene-aware manner. Extensive experiments demonstrate that LayouSyn outperforms existing methods on open-vocabulary and closed-vocabulary layout generation and achieves state-of-the-art performance on challenging spatial and numerical reasoning tasks. Additionally, we present two applications of LayouSyn: First, we demonstrate an interesting finding that we can seamlessly combine initialization from the Large Language model to reduce the diffusion sampling steps. Second, we present a new pipeline for adding objects to the image, demonstrating the potential of LayouSyn in image editing applications.", "keywords": [ "layout generation", "diffusion models" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "u6vC7KaFel", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Edoardo Cetin", "gender": "unknown", "institution": "Sakana AI", "country": "JP", "position": "Researcher" }, { "name": "Robert Tjarko Lange", "gender": "unknown", "institution": "TU Berlin", "country": "", "position": "PhD student" }, { "name": "Rujikorn Charakorn", "gender": "Male", "institution": "Vidyasirimedhi Institute of Science and Technology", "country": "TH", "position": "PhD student" }, { "name": "Yujin Tang", "gender": "Male", "institution": "Sakana AI", "country": "JP", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 47, "n_ref": 63, "n_ref_all": 87, "n_fig": 7, "n_tab": 14, "L_tab": 4655, "n_element_tab": 835, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 1875, "n_element_tab_1": 376, "formula_len_all": 662, "formula_len_all_1": 392, "len_all": 192513, "len_all_1": 62899, "len_abs": 1525, "len_title": 90, "len_sents": 52710, "len_sents_1": 26425, "n_sents": 424, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1386, "title": "Instant Transformer Adaption via HyperLoRA", "abs": "While Foundation Models provide a general tool for rapid content creation, they regularly require task-specific adaptation. Traditionally, this exercise involves careful curation of datasets and repeated fine-tuning of the underlying model. Fine-tuning techniques enable practitioners to adapt foundation models for many new applications but require expensive and lengthy training while being notably sensitive to hyper-parameter choices. To overcome these limitations, we introduce HyperLoRA, a model capable of adapting Large Language Models on the fly---solely based on a natural language description of the target task. HyperLoRA is a hypernetwork trained to construct LoRAs in a single inexpensive forward pass. After training HyperLoRA on a suite of 9 pre-trained LoRA adapters (GSM8K, Arc, etc.), we show that the ad-hoc reconstructed LoRA instances match the performance of task-specific adapters across the corresponding test sets.\nFurthermore, HyperLoRA can compress hundreds of LoRA instances and zero-shot generalize to entirely unseen tasks. This approach provides a significant step towards democratizing the specialization of foundation models and enables language-based adaptation with minimal compute requirements. Our code and pre-trained checkpoints will be available through https://github.com/AnonymousAuthor/hyperlora and https://huggingface.co/ upon publication.", "keywords": [ "hypernetworks", "finetuning", "language models", "zero-shot generalization" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 1, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "u6Y0GdTEYp", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Dongsheng Li", "gender": "Male", "institution": "Auburn University", "country": "US", "position": "PhD student" }, { "name": "Shiwen Mao", "gender": "Male", "institution": "Auburn University", "country": "US", "position": "Full Professor" }, { "name": "Xiaowen Gong", "gender": "Male", "institution": "Auburn University", "country": "US", "position": "Associate Professor" }, { "name": "Xinghan Gong", "gender": "Male", "institution": "Auburn University", "country": "US", "position": "PhD student" } ], "n_formula": 63, "n_formula_1": 27, "n_ref_uni": 23, "n_ref": 49, "n_ref_all": 51, "n_fig": 0, "n_tab": 4, "L_tab": 522, "n_element_tab": 63, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 2277, "n_element_tab_1": 72, "formula_len_all": 7252, "formula_len_all_1": 2350, "len_all": 128200, "len_all_1": 62083, "len_abs": 2480, "len_title": 103, "len_sents": 30849, "len_sents_1": 24166, "n_sents": 295, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 778, "title": "Constrained Multi-Objective Optimization", "abs": "There is more and more attention on constrained multi-objective optimization (CMOO) problems, however, most of them are based on gradient-free methods. This paper proposes a constraint gradient-based algorithm for multi-objective optimization (MOO) problems based on multi-gradient descent algorithms. We first establish a framework for the CMOO problem. Then, we provide a Moreau envelope-based Lagrange Multiplier (MLM-CMOO) algorithm to solve the formulated CMOO problem, and the convergence analysis shows that the proposed algorithm convergence to Pareto stationary solutions with a rate of $\\mathcal{O}(\\frac{1}{\\sqrt{T}})$. Finally, the MLM-CMOO algorithm is tested on several CMOO problems and has shown superior results compared to some chosen state-of-the-art designs.", "keywords": [ "constrained multi-objective optimization", "multi-gradient descent algorithms" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "u63OVngeSp", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "Arash Mehrjou", "gender": "Male", "institution": "GlaxoSmithKlein", "country": "", "position": "Researcher" }, { "name": "Mathieu Chevalley", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Patrick Schwab", "gender": "unknown", "institution": "GlaxoSmithKline plc", "country": "", "position": "Director" } ], "n_formula": 48, "n_formula_1": 24, "n_ref_uni": 53, "n_ref": 102, "n_ref_all": 125, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4766, "formula_len_all_1": 1788, "len_all": 178353, "len_all_1": 69723, "len_abs": 1592, "len_title": 128, "len_sents": 52195, "len_sents_1": 32810, "n_sents": 408, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1598, "title": "Deriving Causal Order from Single-Variable Interventions: Guarantees & Algorithm", "abs": "Targeted and uniform interventions to a system are crucial for unveiling causal relationships. While several methods have been developed to leverage interventional data for causal structure learning, their practical application in real-world scenarios often remains challenging. Recent benchmark studies have highlighted these difficulties, even when large numbers of single-variable intervention samples are available. In this work, we demonstrate, both theoretically and empirically, that such datasets contain a wealth of causal information that can be effectively extracted under realistic assumptions about the data distribution. More specifically, we introduce the notion of interventional faithfulness, which relies on comparisons between the marginal distributions of each variable across observational and interventional settings, and we introduce a score on causal orders. Under this assumption, we are able to prove strong theoretical guarantees on the optimum of our score that also hold for large-scale settings. To empirically verify our theory, we introduce Intersort, an algorithm designed to infer the causal order from datasets containing large numbers of single-variable interventions by approximately optimizing our score. Intersort outperforms baselines (GIES, DCDI, PC and EASE) on almost all simulated data settings replicating common benchmarks in the field. Our proposed novel approach to modeling interventional datasets thus offers a promising avenue for advancing causal inference, highlighting significant potential for further enhancements under realistic assumptions.", "keywords": [ "causality", "causal inference", "causal discovery", "interventional data", "single-variable interventions", "causal structure learning" ], "rating_list": [ 8, 6, 8, 8, 5 ], "soundness_list": [ 3, 2, 4, 3, 2 ], "presentation_list": [ 3, 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 4, 2 ], "confidence_list": [ 3, 3, 3, 4, 4 ] }, { "paper_id": "u5KzZjQOsl", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Alexander Luke Ian Norcliffe", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD Student" }, { "name": "Changhee Lee", "gender": "unknown", "institution": "Korea University", "country": "KR", "position": "Assistant Professor" }, { "name": "Fergus Imrie", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Principal Researcher" }, { "name": "Mihaela van der Schaar", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Pietro Lio", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "Full Professor" } ], "n_formula": 46, "n_formula_1": 17, "n_ref_uni": 62, "n_ref": 174, "n_ref_all": 212, "n_fig": 13, "n_tab": 7, "L_tab": 2957, "n_element_tab": 412, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 1076, "n_element_tab_1": 124, "formula_len_all": 3482, "formula_len_all_1": 873, "len_all": 262476, "len_all_1": 70022, "len_abs": 1117, "len_title": 101, "len_sents": 83597, "len_sents_1": 32778, "n_sents": 748, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1121, "title": "Information Bottleneck for Active Feature Acquisition", "abs": "Traditional supervised learning typically assumes that all features are available simultaneously during deployment. However, this assumption does not hold in many real-world scenarios, such as medicine, where information is acquired sequentially based on an evolving understanding of a specific patient's condition. Active Feature Acquisition aims to address this problem by dynamically selecting which feature to measure based on the current observations, independently for each test instance. Current approaches either use Reinforcement Learning, which suffers from training difficulties; or greedily maximize the conditional mutual information of the label and unobserved features, which inherently makes myopic acquisitions. To address these shortcomings, we introduce a novel method using information bottleneck. Via stochastic encodings, we make acquisitions by reasoning about the features across many possible unobserved realizations in a regularized latent space. Extensive evaluation on a large range of synthetic and real datasets demonstrates that our approach reliably outperforms a diverse set of baselines.", "keywords": [ "Active Feature Acquisition", "Dynamic Feature Selection" ], "rating_list": [ 6, 3, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "u4whlT6xKO", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Harish Tayyar Madabushi", "gender": "Male", "institution": "University of Bath", "country": "GB", "position": "Lecturer" }, { "name": "Haritz Puerto", "gender": "unknown", "institution": "TU Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Iryna Gurevych", "gender": "Female", "institution": "Institute for Computer Science, Artificial Intelligence and Technology", "country": "BG", "position": "Full Professor" }, { "name": "Tilek Chubakov", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "", "position": "PhD student" }, { "name": "Xiaodan Zhu", "gender": "Male", "institution": "Queen's University", "country": "", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 39, "n_ref": 83, "n_ref_all": 107, "n_fig": 2, "n_tab": 7, "L_tab": 2631, "n_element_tab": 380, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 2875, "n_element_tab_1": 379, "formula_len_all": 0, "formula_len_all_1": 148, "len_all": 159287, "len_all_1": 60295, "len_abs": 1157, "len_title": 152, "len_sents": 37767, "len_sents_1": 24193, "n_sents": 314, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 104, "L_abs": 1138, "title": "Fine-Tuning with Divergent Chains of Thought Boosts Reasoning Through Self-Correction in Language Models", "abs": "Requiring a large language model to generate intermediary reasoning steps has been shown to be an effective way of boosting performance. In fact, instruction tuning on these intermediary reasoning steps improves model performance. In this work, we present a novel method of further improving performance by requiring models to compare multiple reasoning chains before generating a solution in a single inference step. We call this method Divergent CoT (DCoT). We find that instruction tuning on DCoT datasets boosts the performance of even smaller, and therefore more accessible, LLMs. Through a rigorous set of experiments spanning a wide range of tasks that require various reasoning types, we show that fine-tuning on DCoT consistently improves performance over the CoT baseline across model families and scales (1.3B to 70B). Through a combination of empirical and manual evaluation, we additionally show that these performance gains stem from models generating multiple divergent reasoning chains in a single inference step, indicative of the enabling of self-correction in language models. Our code and data are publicly available.", "keywords": [ "chain of thought", "reasoning", "self-correction", "LLM", "large language model", "question answering" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "u4dORXVAnx", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Sicun Gao", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Tao Wang", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "PhD student" } ], "n_formula": 51, "n_formula_1": 24, "n_ref_uni": 27, "n_ref": 46, "n_ref_all": 69, "n_fig": 9, "n_tab": 2, "L_tab": 578, "n_element_tab": 41, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 309, "n_element_tab_1": 38, "formula_len_all": 4092, "formula_len_all_1": 1690, "len_all": 122466, "len_all_1": 58753, "len_abs": 1119, "len_title": 93, "len_sents": 40298, "len_sents_1": 26874, "n_sents": 319, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1123, "title": "Numerical Pitfalls in Policy Gradient Updates", "abs": "Numerical instability, such as gradient explosion, is a fundamental problem in practical deep reinforcement learning (DRL) algorithms. Beyond anecdotal debugging heuristics, there is a lack of systematic understanding of the causes for numerical sensitivity that leads to exploding gradient failures in practice. In this work, we demonstrate that the issue arises from the ill-conditioned density ratio in the surrogate objective that comes from importance sampling, which can take excessively large values during training. Perhaps surprisingly, while various policy optimization methods such as TRPO and PPO prevent excessively large policy updates, their optimization constraints on KL divergence and probability ratio cannot guarantee numerical stability. This also explains why gradient explosion often occurs during DRL training, even with code-level optimizations. To address this issue, we propose the Vanilla Policy Gradient with Clipping algorithm, which replaces the importance sampling ratio with its logarithm. This approach effectively prevents gradient explosion while achieving performance comparable to PPO.", "keywords": [ "Numerical stability", "robustness", "deep reinforcement learning", "policy gradient methods", "importance sampling" ], "rating_list": [ 6, 5, 6, 5, 6 ], "soundness_list": [ 2, 2, 3, 4, 3 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 2, 3, 3, 4, 3 ] }, { "paper_id": "u4RVksX8co", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Guimei Liu", "gender": "unknown", "institution": "Institute for Infocomm Research, A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Jung-jae Kim", "gender": "unknown", "institution": "A*STAR", "country": "", "position": "Researcher" }, { "name": "Xiaoli Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Yu Lu", "gender": "Male", "institution": "Beijing Normal University", "country": "", "position": "Associate Professor" } ], "n_formula": 19, "n_formula_1": 20, "n_ref_uni": 40, "n_ref": 91, "n_ref_all": 114, "n_fig": 4, "n_tab": 7, "L_tab": 4417, "n_element_tab": 401, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2361, "n_element_tab_1": 129, "formula_len_all": 1367, "formula_len_all_1": 1330, "len_all": 153945, "len_all_1": 64358, "len_abs": 2341, "len_title": 64, "len_sents": 36705, "len_sents_1": 27042, "n_sents": 294, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 2354, "title": "Learning Skill-level Student Abilities with Item Response Theory", "abs": "Knowledge tracing (KT) aims to estimate knowledge states of students over a given set of skills based on their historical learning activities. The learned knowledge states of students can be used to build skill-meters to understand the weak areas of students so that proper interventions can be taken to help students. Many deep learning models have been applied to KT with encouraging performance, but they either have relatively low accuracy or do not directly generate students' knowledge states at skill level for skill-meter building. Item Response Theory (IRT) models student knowledge states (ability) and question characteristics separately. A question arising naturally is whether we can use IRT to estimate students' knowledge states at skill level while achieving high prediction accuracy at the same time. We examined existing IRT based deep KT models and found that none of them achieves this objective. Most existing IRT-based models either learn overall student abilities or question-level student abilities. Overall student abilities are too summative, and it is hard to tell the weak areas of students from a single value. Question-level abilities are too fine-grained. When there are a large number of unique questions per skill, they can cause information overload for teachers. In this paper, we propose an IRT-based deep KT model called SKKT-IRT to learn skill-level student abilities which provide just the right amount of information for teachers to understand students' knowledge states. Our model consists of an LSTM layer to learn student historical states, a student ability network for learning skill-level student abilities, a question difficulty network for learning question difficulties and a question discrimination network for learning question discrimination. It also learns question-skill relationships as an auxiliary task so that the embedding of a skill can better capture the information of its questions. We further regularize the outputs of question difficulty network and question discrimination network for better performance. Our experimental results show that our model achieves the objective of learning skill-level student abilities with SOTA accuracy. It is also very efficient and produces consistent outputs to be easily used for downstream tasks like adaptive learning and personalized recommendations.", "keywords": [ "knowledge tracing", "item response theory", "student abilities" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "u48BF5O7oL", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Guang Lin", "gender": "Male", "institution": "Purdue University", "country": "", "position": "Associate Professor" }, { "name": "Qifan Song", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "Ziyi Wang", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 23, "n_ref_uni": 67, "n_ref": 142, "n_ref_all": 156, "n_fig": 2, "n_tab": 3, "L_tab": 981, "n_element_tab": 195, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 701, "n_element_tab_1": 133, "formula_len_all": 2798, "formula_len_all_1": 2224, "len_all": 150650, "len_all_1": 63393, "len_abs": 1118, "len_title": 77, "len_sents": 32335, "len_sents_1": 26682, "n_sents": 225, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1127, "title": "Efficient Bayesian DNN Compression through Sparse Quantized Sub-distributions", "abs": "This paper presents a novel method that simultaneously achieves model pruning and low-bit quantization through Bayesian variational inference to effectively compress deep neural networks (DNNs) while suffering minimal performance degradation. \nUnlike previous approaches that treat pruning and quantization as separate, sequential tasks, our method explores a unified optimization space, enabling more efficient compression. \nBy leveraging a spike-and-slab prior combined with Gaussian Mixture Models (GMM), we can achieve both network sparsity and low-bit representation. Experiments on CIFAR-10, CIFAR-100, and SQuAD datasets demonstrate that our approach achieves compression rates of up to 32x with less than a $1.3\\\\%$ accuracy loss on the CIFAR datasets and a 1.66 point decrease in F1 score on SQuAD. Additionally, we show that the Bayesian model average of neural networks can further mitigate the impact of quantization noise, leading to more robust compressed models. Our method outperforms existing techniques in both compression efficiency and accuracy retention, offering a promising solution for compressing DNNs.", "keywords": [ "Bayesian Deep Neural Networks", "Quantization", "Pruning", "Variational Inference" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "u438df0Uce", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Chen Nie", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Jiahao Su", "gender": "Male", "institution": "Shandong Normal University", "country": "", "position": "Undergrad student" }, { "name": "Zekai Xu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhezhi He", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "kang you", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 25, "n_ref": 45, "n_ref_all": 73, "n_fig": 7, "n_tab": 10, "L_tab": 6606, "n_element_tab": 472, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 5500, "n_element_tab_1": 364, "formula_len_all": 636, "formula_len_all_1": 695, "len_all": 116200, "len_all_1": 62906, "len_abs": 1324, "len_title": 162, "len_sents": 26546, "len_sents_1": 23290, "n_sents": 214, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 112, "L_abs": 1277, "title": "SpikeZIP: Compressing Spiking Neural Network with Paths-Ensemble Training for Optimized Pareto-front Performance", "abs": "Spiking neural network (SNN) has attracted great attention due to its great energy efficiency on neuromorphic hardware. \nBy transferring the parameters of a pretrained artificial neural network (ANN) and utilizing the ANN quantization, recent works of ANN-SNN conversion can produce SNNs with close-to-ANN accuracy and low inference latency (known as the number of time-steps).\nNevertheless, existing works fail at providing theoretic equivalence between Quantized-ANN (QANN) and its converted SNN, while the SNN accuracy at small time-step (i.e. Pareto-frontier) can be further improved.\nTo solve the problems, this paper proposes a novel conversion framework called SpikeZIP. The SpikeZIP utilizes the ANN-Quantized ANN(QANN)-SNN two-step conversion to obtain SNN which improves the Pareto frontier of accuracy versus inference time-steps. SpikeZIP integrates two novel algorithms: 1) a paths-ensemble training algorithm that considers the SNN temporal information when fine-tuning QANN; 2) a mathematically equivalent conversion algorithm between the whole QANN and SNN. In the experiment, SpikeZIP can achieve 73.92\\% accuracy on ImageNet with VGG-16 within 9 time-steps and 74.21\\% accuracy on ImageNet with ResNet-34 within 11 time-steps which are better than SOTA works.", "keywords": [ "Spiking Neural Networks", "Quantization", "ANN-SNN Conversion" ], "rating_list": [ 3, 3, 3, 3, 6 ], "soundness_list": [ 2, 1, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 5, 5, 2 ] }, { "paper_id": "u3xwwfHmBC", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Jeong hwan Jeon", "gender": "unknown", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Kyeongmin Do", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Yoojin Kim", "gender": "unknown", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 51, "n_ref_all": 81, "n_fig": 23, "n_tab": 4, "L_tab": 1703, "n_element_tab": 153, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 687, "n_element_tab_1": 98, "formula_len_all": 435, "formula_len_all_1": 246, "len_all": 125492, "len_all_1": 55707, "len_abs": 1053, "len_title": 146, "len_sents": 46307, "len_sents_1": 26848, "n_sents": 339, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1058, "title": "Tri-Tense Former: Capturing Dynamic Traffic Flow Using Tri-Tense Attention for Traffic Forecasting", "abs": "Accurate traffic forecasting is essential to enable advanced utilization of intelligent transportation systems. However, forecasting models often struggle to capture the complex spatio-temporal dependencies of traffic data, as they typically handle spatial and temporal dependencies separately. To overcome this limitation, we introduce the Tri-Tense Former (TTformer), a novel approach that captures spatio-temporal relationships through three tense-specific attention modules. We categorize traffic flow into three tense dimensions: past-to-present (present-perfect), present, and future. Each tense-specific attention module captures the dependencies within its respective traffic flow. Furthermore, to address incomplete traffic data, we improve the robustness of the model by employing contrastive learning with negative filtering technique that operates regardless of predefined adjacency matrices. TTformer significantly outperforms existing models by more effectively capturing spatio-temporal dependencies and improving traffic forecasting accuracy.", "keywords": [ "Traffic forecasting", "Spatio-temporal dependency", "Transformer", "Contrastive learning" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "u3TL0qxLWf", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "David Harrison", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Houman Bedayat", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "JEFFREY MARKER", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Mahyar Najibi", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Maxwell Horton", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Mohammad Rastegari", "gender": "Male", "institution": "Apple", "country": "", "position": "ML manager" }, { "name": "Rasoul Shafipour", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Sachin Mehta", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Saman Naderiparizi", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 25, "n_ref": 31, "n_ref_all": 52, "n_fig": 4, "n_tab": 6, "L_tab": 2402, "n_element_tab": 448, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1697, "n_element_tab_1": 141, "formula_len_all": 438, "formula_len_all_1": 415, "len_all": 120418, "len_all_1": 66622, "len_abs": 1290, "len_title": 118, "len_sents": 35134, "len_sents_1": 30429, "n_sents": 281, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1297, "title": "SeedLM: Compressing LLM Weights into Seeds of Pseudo-Random Generators", "abs": "Large Language Models (LLMs) have transformed natural language processing, but face significant challenges in widespread deployment due to their high runtime cost. In this paper, we introduce SeedLM, a novel post-training compression method that uses seeds of a pseudo-random generator to encode and compress model weights. Specifically, for each block of weights, we find a seed that is fed into a Linear Feedback Shift Register (LFSR) during inference to efficiently generate a random matrix. This matrix is then linearly combined with compressed coefficients to reconstruct the weight block. SeedLM reduces memory access and leverages idle compute cycles during inference, effectively speeding up memory-bound tasks by trading compute for fewer memory accesses. Unlike state-of-the-art methods that rely on calibration data, our approach is data-free and generalizes well across diverse tasks. Our experiments with Llama3 70B, which is particularly challenging, show zero-shot accuracy retention at 4- and 3-bit compression to be on par with or better than state-of-the-art methods, while maintaining performance comparable to FP16 baselines. Additionally, FPGA-based tests demonstrate that 4-bit SeedLM, as model size increases to 70B, approaches a 4x speed-up over an FP16 Llama 2/3 baseline.", "keywords": [ "Model Compression", "Large Language Models", "Post-Training Quantization" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "u2QdCiOgwA", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Athanasios Mouchtaris", "gender": "unknown", "institution": "Amazon Alexa", "country": "", "position": "Researcher" }, { "name": "Grant Strimel", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Principal Researcher" }, { "name": "Jing Liu", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Markus M\u00fcller", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Masao Someki", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Shinji Watanabe", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Siddhant Arora", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Yifan Peng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 4, "n_ref_uni": 44, "n_ref": 92, "n_ref_all": 138, "n_fig": 15, "n_tab": 8, "L_tab": 2852, "n_element_tab": 630, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 98, "n_element_tab_1": 21, "formula_len_all": 1219, "formula_len_all_1": 332, "len_all": 209393, "len_all_1": 60889, "len_abs": 3000, "len_title": 106, "len_sents": 56178, "len_sents_1": 27928, "n_sents": 495, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1289, "title": "Context-aware Dynamic Pruning for Speech Foundation Models", "abs": "Foundation models, such as large language models, have achieved remarkable success in natural language processing and are evolving into models capable of handling multiple modalities.\nListening ability, in particular, is crucial for many applications, leading to research on building speech foundation models. However, the high computational cost of these large models presents a significant challenge for real-world applications. Although substantial efforts have been made to reduce computational costs, such as through pruning techniques, the majority of these approaches are applied primarily during the training phase for specific downstream tasks. In this study, we hypothesize that optimal pruned networks may vary based on contextual factors such as speaker characteristics, languages, and tasks. To address this, we propose a dynamic pruning technique that adapts to these contexts during inference without altering the underlying model. We demonstrated that we could successfully reduce inference time by approximately 30\\% while maintaining accuracy in multilingual/multi-task scenarios. We also found that the obtained pruned structure offers meaningful interpretations based on the context, e.g., task-related information emerging as the dominant factor for efficient pruning.", "keywords": [ "Pruning", "Speech Foundation Model", "Automatic Speech Recognition", "Speech Translation" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 2, 4, 3 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "u1rlO94Bnr", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Hanshen Xiao", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" }, { "name": "Marten van Dijk", "gender": "Male", "institution": "Centrum voor Wiskunde en Informatica", "country": "", "position": "Full Professor" }, { "name": "Srini Devadas", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Lam M Nguyen", "gender": "Male", "institution": "IBM Research, Thomas J. Watson Research Center", "country": "US", "position": "Staff Research Scientist" } ], "n_formula": 193, "n_formula_1": 22, "n_ref_uni": 39, "n_ref": 85, "n_ref_all": 149, "n_fig": 5, "n_tab": 6, "L_tab": 1939, "n_element_tab": 216, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 45, "n_element_tab_1": 4, "formula_len_all": 27451, "formula_len_all_1": 2016, "len_all": 220290, "len_all_1": 68546, "len_abs": 1212, "len_title": 135, "len_sents": 57595, "len_sents_1": 31967, "n_sents": 517, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1238, "title": "Why DP \"LOCAL\" SGD \u2013 Faster Convergence in Less Composition with Clipping Bias Reduction", "abs": "We argue to apply Differentially-Private Local Stochastic Gradient Descent (DP-LSGD), a generalization of regular DP-SGD with per-sample local iterations, to systematically improve privacy-preserving machine learning. We prove and show the following facts in this paper: a). DP-LSGD with local iterations can produce more concentrated per-sample updates and therefore enables a more efficient exploitation of the clipping budget with a better utility-privacy tradeoff; b). given the same $T$ privacy composition or per-sample update aggregation, with properly-selected local iterations, DP-LSGD can converge faster in $O(1/T)$ to a small neighborhood of (local) optimum compared to $O(1/\\sqrt{T})$ in regular DP-SGD, i.e., DP-LSGD produces the same accuracy while consumes less of the privacy budget. From an empirical side, thorough experiments are provided to support our developed theory and we show DP-LSGD produces the best-known performance in various practical deep learning tasks: for example with an $(\\epsilon=4,\\delta=10^{-5})$-DP guarantee, we successfully train ResNet20 from scratch with test accuracy $74.1\\%, 86.5\\%$ and $91.7\\%$ on CIFAR10, SVHN and EMNIST, respectively. Our code is released in an anonymous GitHub link.", "keywords": [ "Differential Privacy", "Local Stochastic Gradient Descent", "Clipping Bias" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "u1cQYxRI1H", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Anyi Rao", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Lvmin Zhang", "gender": "Not Specified", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Maneesh Agrawala", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 15, "n_ref_uni": 81, "n_ref": 118, "n_ref_all": 130, "n_fig": 5, "n_tab": 1, "L_tab": 217, "n_element_tab": 30, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 217, "n_element_tab_1": 30, "formula_len_all": 689, "formula_len_all_1": 689, "len_all": 175326, "len_all_1": 56752, "len_abs": 1556, "len_title": 174, "len_sents": 29031, "len_sents_1": 27729, "n_sents": 191, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 126, "L_abs": 1559, "title": "Scaling In-the-Wild Training for Diffusion-based Illumination Harmonization and Editing by Imposing Consistent Light Transport", "abs": "Diffusion-based image generators are becoming unique methods for illumination harmonization and editing. The current bottleneck in scaling up the training of diffusion-based illumination editing models is mainly in the difficulty of preserving the underlying image details and maintaining intrinsic properties, such as albedos, unchanged. Without appropriate constraints, directly training the latest large image models with complex, varied, or in-the-wild data is likely to produce a structure-guided random image generator, rather than achieving the intended goal of precise illumination manipulation. We propose Imposing Consistent Light (IC-Light) transport during training, rooted in the physical principle that the linear blending of an object's appearances under different illumination conditions is consistent with its appearance under mixed illumination. This consistency allows for stable and scalable illumination learning, uniform handling of various data sources, and facilitates a physically grounded model behavior that modifies only the illumination of images while keeping other intrinsic properties unchanged. Based on this method, we can scale up the training of diffusion-based illumination editing models to large data quantities (> 10 million), across all available data types (real light stages, rendered samples, in-the-wild synthetic augmentations, etc), and using strong backbones (SDXL, Flux, etc). We also demonstrate that this approach reduces uncertainties and mitigates artifacts such as mismatched materials or altered albedos.", "keywords": [ "diffusion model", "illumination editing", "image editing" ], "rating_list": [ 10, 10, 10, 10 ], "soundness_list": [ 3, 4, 4, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 4, 4, 4, 4 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "u1EPPYkbgA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Jiaxu Zhao", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "PhD student" }, { "name": "Meng Fang", "gender": "Male", "institution": "University of Liverpool", "country": "", "position": "Assistant Professor" }, { "name": "Mykola Pechenizkiy", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Full Professor" }, { "name": "Shirui Pan", "gender": "unknown", "institution": "Griffith University", "country": "AU", "position": "Full Professor" }, { "name": "Wenpeng Yin", "gender": "unknown", "institution": "Pennsylvania State University", "country": "", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 24, "n_ref": 49, "n_ref_all": 78, "n_fig": 1, "n_tab": 6, "L_tab": 2962, "n_element_tab": 502, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 229, "n_element_tab_1": 14, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 133632, "len_all_1": 59559, "len_abs": 1157, "len_title": 127, "len_sents": 41091, "len_sents_1": 27827, "n_sents": 370, "n_sents_1": 247, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1162, "title": "GPTBIAS: A Comprehensive Framework for Evaluating Bias in Large Language Models", "abs": "Large language models (LLMs) have seen widespread adoption across various applications, both in their original form and fine-tuned adaptations. However, a major concern with LLMs is their potential to generate biased content. Existing evaluation methods often have different constraints, such as needing access to the model's intermediate outputs. To address these issues, we propose GPTBIAS, a novel bias evaluation framework that leverages the capabilities of advanced LLMs like GPT-4 to assess bias in other models across nine bias types. Our framework introduces Bias Attack Instructions, specifically designed to evaluate model bias across multiple dimensions. GPTBIAS provides not only a quantitative bias score but also detailed information on bias types, affected demographics, underlying reasons for biases, and suggestions for improvement. Through extensive experiments on popular LLMs, we demonstrate the effectiveness and usability of our bias evaluation framework. Our results reveal nuanced insights into the biases present in different models and highlight the importance of comprehensive bias assessment in the development and deployment of LLMs.", "keywords": [ "Large Language Model", "Bias" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "u14Y236LwX", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Hao Hu", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Undergrad student" }, { "name": "He Sun", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jinzhuo Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Linfeng Zhang", "gender": "Male", "institution": "DP Technology", "country": "", "position": "Researcher" }, { "name": "Xinliang Liu", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Postdoc" }, { "name": "Yihang Zheng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Youjia Zheng", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yubing Li", "gender": "Male", "institution": "Institute of Acoustics Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zeyuan Dong", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zhijun Zeng", "gender": "Male", "institution": "DP Technology", "country": "CN", "position": "Intern" }, { "name": "Zuoqiang Shi", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 32, "n_ref": 68, "n_ref_all": 89, "n_fig": 12, "n_tab": 6, "L_tab": 2029, "n_element_tab": 256, "n_fig_1": 11, "n_tab_1": 6, "L_tab_1": 2889, "n_element_tab_1": 180, "formula_len_all": 687, "formula_len_all_1": 303, "len_all": 111136, "len_all_1": 55866, "len_abs": 1381, "len_title": 165, "len_sents": 33318, "len_sents_1": 24378, "n_sents": 229, "n_sents_1": 152, "venue": "ICLR2025", "decision": "Reject", "L_title": 115, "L_abs": 1385, "title": "OpenWaves: A Large-Scale Anatomically Realistic Ultrasound-CT Dataset for Benchmarking Neural Wave Equation Solvers", "abs": "Accurate and efficient simulation of wave equations is crucial in computational physics, especially for wave imaging applications like ultrasound computed tomography (USCT), which reconstructs tissue properties from scattered waves. Traditional numerical solvers for wave equations are computationally intensive and often unstable, limiting their practical applications for quasi-real-time imaging. Neural operators offer an innovative approach by accelerating PDE solving using neural networks; however, their effectiveness in realistic imaging is constrained by existing datasets that oversimplify real-world complexity. In this paper, we present OpenWaves, a large-scale wave equation dataset designed to bridge the gap between theoretical equations and practical imaging applications. OpenWaves provides over 16 million frequency-domain wave simulations using real USCT configurations, featuring anatomically realistic human breast phantoms across four categories. It enables comprehensive benchmarking of popular neural operators for both forward simulation and inverse imaging tasks, allowing analysis of their performance, scalability, and generalization capabilities. By offering a realistic and extensive dataset, OpenWaves not only serves as a platform for developing innovative neural PDE solvers but also facilitates their deployment in real-world medical imaging problems.", "keywords": [ "Computational imaging", "Inverse problem", "Neural operators", "Ultrasound Computed Tomography", "Full Waveform Inversion" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 2, 1, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "u0L7djBiRw", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Baiyu Chen", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Wenjian Yu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 46, "n_formula_1": 25, "n_ref_uni": 20, "n_ref": 66, "n_ref_all": 81, "n_fig": 5, "n_tab": 2, "L_tab": 912, "n_element_tab": 54, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 643, "n_element_tab_1": 46, "formula_len_all": 3664, "formula_len_all_1": 2396, "len_all": 121939, "len_all_1": 70955, "len_abs": 1534, "len_title": 105, "len_sents": 39718, "len_sents_1": 29608, "n_sents": 377, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1568, "title": "A Rademacher-Like Random Embedding with Linear Complexity", "abs": "Random embedding assumes an important role in representation learning. Gaussian embedding and Rademacher embedding are two widely used random embeddings. Although they usually enjoy robustness and effectiveness, their computational complexity is high, i.e. $O(nk)$ for embedding an $n$-dimensional vector into $k$-dimensional space. The alternatives include partial subsampled randomized Hadamard (P-SRHT) embedding and sparse sign embedding, which are still not of linear complexity or cannot run efficiently in practical implementation. In this paper, a fast and robust Rademacher-like embedding (RLE) is proposed, based on a smaller Rademacher matrix and several auxiliary random arrays. Specifically, it embeds an $n$-dimensional vector into $k$-dimensional space in just $O(n)$ time and space (assuming $k$ is not larger than $O(n^{\\frac{1}{2}})$). Our theoretic analysis reveals that the proposed RLE owns most of desirable properties of the Rademacher embedding while preserving lower complexity. To validate the practical efficiency and effectiveness, the proposed RLE is applied to single-pass randomized singular value decomposition (single-pass RSVD) for streaming data, and the randomized Arnoldi process based on sketched ordinary least-squares. Numerical experiments show that, with the proposed RLE the single-pass RSVD achieves 1.7x speed-up on average while keeping same or better accuracy, and the randomized Arnodli process enables a randomized GMRES algorithm running 1.3x faster on average for solving $Ax=b$ than that based on other embeddings.", "keywords": [ "Random Embedding", "Randomized Singular Value Deomposition", "Randomized Arnoldi Process", "Machine Learning", "Linear Complexity" ], "rating_list": [ 1, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 4 ], "presentation_list": [ 1, 2, 1, 2 ], "contribution_list": [ 1, 3, 3, 4 ], "confidence_list": [ 4, 5, 5, 2 ] }, { "paper_id": "u08UxVNdIo", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Chengwei Pan", "gender": "Male", "institution": "Beihang Uinveristy", "country": "CN", "position": "Associate Professor" }, { "name": "Hao Li", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Hao Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Junda Zhu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Lei Sha", "gender": "Male", "institution": "Beihang University", "country": "", "position": "Full Professor" }, { "name": "Minlie Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xinyuan Wang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "MS student" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 39, "n_ref": 67, "n_ref_all": 75, "n_fig": 1, "n_tab": 1, "L_tab": 1227, "n_element_tab": 174, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1263, "n_element_tab_1": 176, "formula_len_all": 727, "formula_len_all_1": 691, "len_all": 117814, "len_all_1": 58745, "len_abs": 1302, "len_title": 193, "len_sents": 33727, "len_sents_1": 26789, "n_sents": 223, "n_sents_1": 169, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1218, "title": "Diffusion Attacker: Diffusion-Driven Prompt Manipulation for LLM Jailbreak", "abs": "Large Language Models can generate harmful content when prompted with carefully crafted inputs, a vulnerability known as LLM jailbreaking. As LLMs become more powerful, studying jailbreaking becomes a critical aspect of enhancing security and human value alignment. Currently, jailbreak is usually implemented by adding suffixes or using prompt templates, which suffers from low attack diversity. Inspired by diffusion models, this paper introduces the DiffusionAttacker, an end-to-end generative method for jailbreak rewriting. Our approach employs a seq2seq text diffusion model as a generator, conditioning on the original prompt and guiding the denoising process with a novel attack loss. This method preserves the semantic content of the original prompt while producing harmful content. Additionally, we leverage the Gumbel-Softmax technique to make the sampling process from the output distribution of the diffusion model differentiable, thereby eliminating the need for an iterative token search. Through extensive experiments on the Advbench and Harmbench, we show that DiffusionAttacker outperforms previous methods in various evaluation indicators including attack success rate (ASR), fluency, and diversity.", "keywords": [ "LLM safety; LLM jailbreak; Diffusion Language Model; Gumbel Softmax;" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "tznvtmSEiN", "primary_area": "optimization", "n_author": 8, "author_info_list": [ { "name": "Hanxiao Wang", "gender": "unknown", "institution": "Institute of automation, Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Jun Luo", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Li Luo", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "MS student" }, { "name": "Lingkun Wen", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Linlong Wu", "gender": "Male", "institution": "University of Luxemburg", "country": "LU", "position": "Researcher" }, { "name": "Sheng Xu", "gender": "Male", "institution": "Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Xianliang Li", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Zhiwei Zheng", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 44, "n_ref": 54, "n_ref_all": 83, "n_fig": 13, "n_tab": 9, "L_tab": 3056, "n_element_tab": 351, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 564, "n_element_tab_1": 75, "formula_len_all": 1862, "formula_len_all_1": 925, "len_all": 153650, "len_all_1": 71122, "len_abs": 1221, "len_title": 126, "len_sents": 49959, "len_sents_1": 33130, "n_sents": 381, "n_sents_1": 273, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1211, "title": "On the Performance Analysis of Momentum Method: A Frequency Domain Perspective", "abs": "Momentum-based optimizers are widely adopted for training neural networks. However, the optimal selection of momentum coefficients remains elusive. This uncertainty impedes a clear understanding of the role of momentum in stochastic gradient methods. In this paper, we present a frequency domain analysis framework that interprets the momentum method as a time-variant filter for gradients, where adjustments to momentum coefficients modify the filter characteristics. Our experiments support this perspective and provide a deeper understanding of the mechanism involved. Moreover, our analysis reveals the following significant findings: high-frequency gradient components are undesired in the late stages of training; preserving the original gradient in the early stages, and gradually amplifying low-frequency gradient components during training both enhance performance. Based on these insights, we propose Frequency Stochastic Gradient Descent with Momentum (FSGDM), a heuristic optimizer that dynamically adjusts the momentum filtering characteristic with an empirically effective dynamic magnitude response. Experimental results demonstrate the superiority of FSGDM over conventional momentum optimizers.", "keywords": [ "Momentum Method", "Stochastic Gradient Descent", "Z-Transform", "Frequency Domain Analysis", "Deep Learning" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 4 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 4, 1, 4 ] }, { "paper_id": "tyFGIjNzlj", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 2, "author_info_list": [ { "name": "Lucas Paul Saldyt", "gender": "Not Specified", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Subbarao Kambhampati", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 49, "n_ref": 88, "n_ref_all": 95, "n_fig": 3, "n_tab": 5, "L_tab": 557, "n_element_tab": 30, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 270, "n_element_tab_1": 25, "formula_len_all": 403, "formula_len_all_1": 330, "len_all": 106646, "len_all_1": 58470, "len_abs": 2155, "len_title": 115, "len_sents": 38447, "len_sents_1": 28580, "n_sents": 284, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1263, "title": "Algorithmic Language Models with Neurally Compiled Libraries", "abs": "Important reasoning tasks such as planning are fundamentally algorithmic, meaning that solving these tasks robustly requires inducing the underlying algorithms, rather than shortcuts. Large Language Models lack true algorithmic ability primarily because of the limitations of neural network optimization algorithms, their optimization data and optimization objective, but also due to the inexpressivity of the transformer architecture. To address this lack of algorithmic ability, our paper proposes augmenting LLMs with an internal reasoning module. This module contains a library of fundamental operations and sophisticated differentiable programs, so that common algorithms do not need to be learned from scratch. To accomplish this, we add memory, registers, basic operations, and adaptive recurrence to a billion-parameter scale transformer architecture built on LLaMA3.2. Then, we define a method for directly compiling algorithms into a differentiable starting library, which is used natively and propagates gradients for optimization. In this paper, we study the feasibility of this augmentation by fine-tuning an augmented LLaMA 3.2 on simple algorithmic tasks with variable computational depth, such as a recursive fibonacci algorithm or insertion sort.", "keywords": [ "neural compilation", "program synthesis", "reasoning", "algorithms", "planning", "large language models", "fine tuning", "tool use" ], "rating_list": [ 1, 3, 6, 5 ], "soundness_list": [ 1, 3, 2, 1 ], "presentation_list": [ 1, 3, 2, 2 ], "contribution_list": [ 1, 1, 3, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "tyEyYT267x", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Aaron Gokaslan", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Jiaqi Han", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" }, { "name": "Marianne Arriola", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Subham Sekhar Sahoo", "gender": "Male", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "PhD student" }, { "name": "Volodymyr Kuleshov", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Assistant Professor" }, { "name": "Zhihan Yang", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Zhixuan Qi", "gender": "Female", "institution": "Cornell University", "country": "US", "position": "MS student" }, { "name": "Justin Chiu", "gender": "Male", "institution": "Cornell University", "country": "", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 20, "n_ref_uni": 28, "n_ref": 54, "n_ref_all": 71, "n_fig": 2, "n_tab": 8, "L_tab": 985, "n_element_tab": 139, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 1448, "n_element_tab_1": 152, "formula_len_all": 3793, "formula_len_all_1": 1757, "len_all": 142792, "len_all_1": 61260, "len_abs": 780, "len_title": 78, "len_sents": 45982, "len_sents_1": 23522, "n_sents": 414, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 77, "L_abs": 782, "title": "Interpolating Autoregressive and Discrete Denoising Diffusion Language Models", "abs": "Diffusion language models offer unique benefits over autoregressive (AR) models due to their potential for parallelized generation and controllability, yet they lag in likelihood modeling and are limited to fixed-length generation. In this work, we introduce a class of semi-autoregressive (SAR) diffusion models that interpolate between discrete denoising diffusion and autoregressive models. \nWe propose a recipe for building effective SAR models that includes an efficient training algorithm, estimators of gradient variance, and data-driven noise schedules to minimize the variance. SAR models overcome key limitations of diffusion language models, setting a new state-of-the-art performance on language modeling benchmarks and enabling generation of arbitrary-length sequences.", "keywords": [ "Diffusion Models", "Text Diffusion", "Generative Models" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "txoJvjfI9w", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Bin Liang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Kam-Fai Wong", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "Full Professor" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Liang CHEN", "gender": "Male", "institution": "Chinese University of Hong Kong, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xiaoyan Zhao", "gender": "Female", "institution": "Chinese University of Hong Kong, The Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Yang Deng", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 21, "n_formula_1": 20, "n_ref_uni": 50, "n_ref": 100, "n_ref_all": 127, "n_fig": 6, "n_tab": 13, "L_tab": 6550, "n_element_tab": 1013, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 533, "n_element_tab_1": 43, "formula_len_all": 1060, "formula_len_all_1": 1030, "len_all": 223471, "len_all_1": 67342, "len_abs": 1649, "len_title": 89, "len_sents": 47028, "len_sents_1": 33218, "n_sents": 349, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 1666, "title": "PEARL: Towards Permutation-Resilient LLMs", "abs": "The in-context learning (ICL) capability of large language models (LLMs) enables them to perform challenging tasks using provided demonstrations. However, ICL is highly sensitive to the ordering of demonstrations, leading to instability in predictions. This paper shows that this vulnerability can be exploited to design a natural attack\u2014difficult for model providers to detect\u2014that achieves nearly 80% success rate on LLaMA-3 by simply permuting the demonstrations. Existing mitigation methods primarily rely on post-processing and fail to enhance the model's inherent robustness to input permutations, raising concerns about safety and reliability of LLMs. To address this issue, we propose Permutation-resilient learning (PEARL), a novel framework based on distributionally robust optimization (DRO), which optimizes model performance against the worst-case input permutation. Specifically, PEARL consists of a permutation-proposal network (P-Net) and the LLM. The P-Net generates the most challenging permutations by treating it as an optimal transport problem, which is solved using an entropy-constrained Sinkhorn algorithm. Through minimax optimization, the P-Net and the LLM iteratively optimize against each other, progressively improving the LLM's robustness. Experiments on synthetic pre-training and real-world instruction tuning tasks demonstrate that PEARL effectively mitigates permutation attacks and enhances performance. Notably, despite being trained on fewer shots and shorter contexts, PEARL achieves performance gains of up to 40% when scaled to many-shot and long-context scenarios, highlighting its efficiency and generalization capabilities.", "keywords": [ "In-Context Learning", "LLMs", "Adversarial Robustness", "Distributionally Robust Optimization", "Instruction Tuning", "Optimal Transport", "Sinkhorn Algorithm" ], "rating_list": [ 3, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "txZVQRc2ab", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Changjiang Li", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Fenglong Ma", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Tanqiu Jiang", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Ting Wang", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 15, "n_ref_uni": 25, "n_ref": 83, "n_ref_all": 103, "n_fig": 6, "n_tab": 13, "L_tab": 2224, "n_element_tab": 163, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1184, "n_element_tab_1": 84, "formula_len_all": 13773, "formula_len_all_1": 11734, "len_all": 160868, "len_all_1": 78975, "len_abs": 1304, "len_title": 127, "len_sents": 42387, "len_sents_1": 31830, "n_sents": 282, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1345, "title": "RAPID: Retrieval Augmented Training of Differentially Private Diffusion Models", "abs": "Differentially private diffusion models (DPDMs) harness the remarkable generative capabilities of diffusion models while enforcing differential privacy (DP) for sensitive data. However, existing DPDM training approaches often suffer from significant utility loss, large memory footprint, and expensive inference cost, impeding their practical uses. \nTo overcome such limitations, we present RAPID: Retrieval Augmented PrIvate Diffusion model, a novel approach that integrates retrieval augmented generation (RAG) into DPDM training. Specifically, RAPID leverages available public data to build a knowledge base of sample trajectories; when training the diffusion model on private data, RAPID computes the early sampling steps as queries, retrieves similar trajectories from the knowledge base as surrogates, and focuses on training the later sampling steps in a differentially private manner. Extensive evaluation using benchmark datasets and models demonstrates that, with the same privacy guarantee, RAPID significantly outperforms state-of-the-art approaches by large margins in generative quality, memory footprint, and inference cost, suggesting that retrieval-augmented DP training represents a promising direction for developing future privacy-preserving generative models. The code is available at: https://github.com/TanqiuJiang/RAPID", "keywords": [ "Diffusion model", "retrieval augmented generation", "differential privacy" ], "rating_list": [ 6, 5, 8, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "txV4dNeusx", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Abhradeep Guha Thakurta", "gender": "Male", "institution": "Google", "country": "", "position": "Senior Research Scientist" }, { "name": "Arun Ganesh", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Christopher A. Choquette-Choo", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Saminul Haque", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Thomas Steinke", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" } ], "n_formula": 18, "n_formula_1": 10, "n_ref_uni": 17, "n_ref": 59, "n_ref_all": 80, "n_fig": 9, "n_tab": 1, "L_tab": 66, "n_element_tab": 13, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1407, "formula_len_all_1": 889, "len_all": 127487, "len_all_1": 56383, "len_abs": 1101, "len_title": 102, "len_sents": 48381, "len_sents_1": 27426, "n_sents": 341, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1091, "title": "Near-Exact Privacy Amplification for Matrix Mechanisms", "abs": "We study the problem of computing the privacy parameters for DP machine learning when using privacy amplification via random batching and noise correlated across rounds via a correlation matrix $\\textbf{C}$ (i.e., the matrix mechanism). Past work on this problem either only applied to banded $\\textbf{C}$, or gave loose privacy parameters. In this work, we give a framework for computing near-exact privacy parameters for any lower-triangular, non-negative $\\textbf{C}$. Our framework allows us to optimize the correlation matrix $\\textbf{C}$ while accounting for amplification, whereas past work could not. Empirically, we show this lets us achieve smaller RMSE on prefix sums than the previous state-of-the-art (SOTA). We also show that we can improve on the SOTA performance on deep learning tasks. Our two main technical tools are (i) using Monte Carlo accounting to bypass composition, which was the main technical challenge for past work, and (ii) a ``balls-in-bins'' batching scheme that enables easy privacy analysis and is closer to practical random batching than Poisson sampling.", "keywords": [ "differential privacy", "privacy amplification", "privacy accounting", "DP-FTRL", "correlated noise" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "txD9llAYn9", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Dongruo Zhou", "gender": "Male", "institution": "Indiana University", "country": "US", "position": "Assistant Professor" }, { "name": "John Lui", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Wen Sun", "gender": "unknown", "institution": "Cornell University", "country": "", "position": "Assistant Professor" }, { "name": "Zhiyong Wang", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Visiting PhD student" } ], "n_formula": 152, "n_formula_1": 42, "n_ref_uni": 45, "n_ref": 91, "n_ref_all": 116, "n_fig": 0, "n_tab": 2, "L_tab": 140, "n_element_tab": 14, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 22092, "formula_len_all_1": 4434, "len_all": 205356, "len_all_1": 78250, "len_abs": 3519, "len_title": 127, "len_sents": 52470, "len_sents_1": 31968, "n_sents": 510, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 643, "title": "Model-based RL as a Minimalist Approach to Horizon-Free and Second-Order Bounds", "abs": "Learning a transition model via Maximum Likelihood Estimation (MLE) followed by planning inside the learned model is perhaps the most standard and simplest Model-based Reinforcement Learning (RL) framework. In this work, we show that such a simple Model-based RL scheme, when equipped with optimistic and pessimistic planning procedures, achieves strong regret and sample complexity bounds in online and offline RL settings. Particularly, we demonstrate that under the conditions where the trajectory-wise reward is normalized between zero and one and the transition is time-homogenous, it achieves nearly horizon-free and second-order bounds.", "keywords": [ "reinforcement learning theory", "model-based reinforcement learning" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "twtTLZnG0B", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Yida Yin", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Zhiqiu Xu", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Zhiyuan Li", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Assistant Professor" }, { "name": "trevor darrell", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department", "country": "US", "position": "Professor" }, { "name": "Zhuang Liu", "gender": "Male", "institution": "FAIR, Meta AI", "country": "US", "position": "Research Scientist" } ], "n_formula": 39, "n_formula_1": 13, "n_ref_uni": 58, "n_ref": 97, "n_ref_all": 136, "n_fig": 15, "n_tab": 14, "L_tab": 5488, "n_element_tab": 732, "n_fig_1": 8, "n_tab_1": 7, "L_tab_1": 1438, "n_element_tab_1": 157, "formula_len_all": 2288, "formula_len_all_1": 834, "len_all": 160710, "len_all_1": 58549, "len_abs": 965, "len_title": 82, "len_sents": 38310, "len_sents_1": 23403, "n_sents": 394, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 34, "L_abs": 991, "title": "A Coefficient Makes SVRG Effective", "abs": "Stochastic Variance Reduced Gradient (SVRG), introduced by Johnson & Zhang (2013), is a theoretically compelling optimization method. However, as Defazio & Bottou (2019) highlight, its effectiveness in deep learning is yet to be proven. In this work, we demonstrate the potential of SVRG in optimizing real-world neural networks. Our empirical analysis finds that, for deeper neural networks, the strength of the variance reduction term in SVRG should be smaller and decrease as training progresses. Inspired by this, we introduce a multiplicative coefficient $\\alpha$ to control the strength and adjust it through a linear decay schedule. We name our method $\\alpha$-SVRG. Our results show $\\alpha$-SVRG better optimizes models, consistently reducing training loss compared to the baseline and standard SVRG across various model architectures and multiple image classification datasets. We hope our findings encourage further exploration into variance reduction techniques in deep learning.", "keywords": [ "Optimization; Variance Reduction; SGD" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "twIPSx9qHn", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Bohan Zhou", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Haoqi Yuan", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yuhui Fu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Zongqing Lu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 48, "n_ref": 111, "n_ref_all": 129, "n_fig": 6, "n_tab": 6, "L_tab": 1289, "n_element_tab": 173, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1379, "n_element_tab_1": 163, "formula_len_all": 389, "formula_len_all_1": 179, "len_all": 150222, "len_all_1": 66860, "len_abs": 1240, "len_title": 111, "len_sents": 42768, "len_sents_1": 30477, "n_sents": 300, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1288, "title": "Cross-Embodiment Dexterous Grasping with Reinforcement Learning", "abs": "Dexterous hands exhibit significant potential for complex real-world grasping tasks. While recent studies have primarily focused on learning policies for specific robotic hands, the development of a universal policy that controls diverse dexterous hands remains largely unexplored.\nIn this work, we study the learning of cross-embodiment dexterous grasping policies using reinforcement learning (RL). Inspired by the capability of human hands to control various dexterous hands through teleoperation, we propose a universal action space based on the human hand's eigengrasps. The policy outputs eigengrasp actions that are then converted into specific joint actions for each robot hand through a retargeting mapping. We simplify the robot hand's proprioception to include only the positions of fingertips and the palm, offering a unified observation space across different robot hands. Our approach demonstrates an 80\\% success rate in grasping objects from the YCB dataset across four distinct embodiments using a single vision-based policy. Additionally, our policy exhibits zero-shot generalization to two previously unseen embodiments and significant improvement in efficient finetuning. For further details and videos, visit our project page (https://sites.google.com/view/crossdex).", "keywords": [ "dexterous grasping", "cross-embodiment learning", "reinforcement learning" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "twEvvkQqPS", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 12, "author_info_list": [ { "name": "Bin Shao", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Principal Research Manager" }, { "name": "Chang Liu", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Erpai Luo", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Han Yang", "gender": "unknown", "institution": "Research, Microsoft", "country": "CN", "position": "Senior Researcher" }, { "name": "Jia Zhang", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Associate Researcher" }, { "name": "Huang Lin", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Samuel Harshe", "gender": "Male", "institution": "Yale University", "country": "", "position": "Researcher" }, { "name": "Xinran Wei", "gender": "Female", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Yunyang Li", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Zun Wang", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Mark Gerstein", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Full Professor" }, { "name": "zaishuo xia", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Intern" } ], "n_formula": 118, "n_formula_1": 20, "n_ref_uni": 75, "n_ref": 122, "n_ref_all": 156, "n_fig": 13, "n_tab": 16, "L_tab": 6412, "n_element_tab": 649, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1460, "n_element_tab_1": 160, "formula_len_all": 6586, "formula_len_all_1": 1360, "len_all": 318666, "len_all_1": 74253, "len_abs": 1678, "len_title": 139, "len_sents": 91935, "len_sents_1": 32139, "n_sents": 737, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 91, "L_abs": 1686, "title": "Enhancing the Scalability and Applicability of Kohn-Sham Hamiltonians for Molecular Systems", "abs": "Density Functional Theory (DFT) is a pivotal method within quantum chemistry and materials science, with its core involving the construction and solution of the Kohn-Sham Hamiltonian. Despite its importance, the application of DFT is frequently limited by the substantial computational resources required to construct the Kohn-Sham Hamiltonian. In response to these limitations, current research has employed deep-learning models to efficiently predict molecular and solid Hamiltonians, with roto-translational symmetries encoded in their neural networks. However, the scalability of prior models may be problematic when applied to large molecules, resulting in non-physical predictions of ground-state properties. In this study, we generate a substantially larger training set (PubChemQH) than used previously and use it to create a scalable model for DFT calculations with physical accuracy. For our model, we introduce a loss function derived from physical principles, which we call Wavefunction Alignment Loss (WALoss). WALoss involves performing a basis change on the predicted Hamiltonian to align it with the observed one; thus, the resulting differences can serve as a surrogate for orbital energy differences, allowing models to make better predictions for molecular orbitals and total energies than previously possible. WALoss also substantially accelerates self-consistent-field (SCF) DFT calculations. Here, we show it achieves a reduction in total energy prediction error by a factor of 1347 and an SCF calculation speed-up by a factor of 18\\%. These substantial improvements set new benchmarks for achieving accurate and applicable predictions in larger molecular systems.", "keywords": [ "AI for Science", "Quantum Chemistry", "EGNN" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "tvjcGkBf0g", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "HU TIHUA", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Hao Chen", "gender": "Male", "institution": "Institute for Infocomm Research, A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Kaixin Xu", "gender": "Male", "institution": "I2R, A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Min Wu", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Weisi Lin", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Xiaoli Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Xue Geng", "gender": "Female", "institution": "Institute for Infocomm Research, A*STAR", "country": "SG", "position": "Research Scientist" }, { "name": "Xulei Yang", "gender": "Male", "institution": "I2R, A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "qingtian feng", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Zhe Wang", "gender": "Male", "institution": ", A*STAR", "country": "SG", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 22, "n_ref_uni": 19, "n_ref": 29, "n_ref_all": 43, "n_fig": 5, "n_tab": 8, "L_tab": 3250, "n_element_tab": 460, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1399, "n_element_tab_1": 93, "formula_len_all": 1422, "formula_len_all_1": 1098, "len_all": 118387, "len_all_1": 60402, "len_abs": 2126, "len_title": 134, "len_sents": 28775, "len_sents_1": 27356, "n_sents": 219, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 2201, "title": "DM3D: Parameter-Efficient and Lossless 3D Object Detection via Distortion Minimization", "abs": "Recent advancements in 3D deep learning have garnered significant attention, given their superior performance in fields like AR/VR, autonomous driving, and robotics. \nHowever, as the models and point cloud data continues to scale up, managing computational and memory demands becomes a critical challenge, particularly for real-world applications with strict latency and energy requirements.\nPrevious methods have primarily focused on reducing computational costs and memory usage by addressing spatial redundancy, \\textit{i.e.}, filtering out irrelevant points or voxels. In contrast, this work presents a novel post-training weight pruning technique tailored specifically for 3D object detection.\nOur approach stands out in two key ways: (1) it operates independently from existing point cloud sparsification methods, targeting redundant parameters in pre-trained models that minimally affect both spatial accuracy and detection confidence (collectively referred to as \"detection distortion\"), and (2) it provides a flexible, plug-and-play framework compatible with other sparsity schemes including spatial sparsity and with any 3D detection model.\nOur method reduces detection distortion by employing a second-order Taylor approximation to identify layer-wise sparsity, allowing for a substantial reduction in model complexity without sacrificing detection accuracy. \nTo efficiently manage the necessary second-order information, we devised a lightweight algorithm to gather Hessian information, followed by dynamic programming to optimize layer-wise sparsity allocation.\nExtensive experiments on the KITTI, nuScenes, and ONCE datasets validate the effectiveness of our approach, where we not only preserve detection performance but also notice enhancement while significantly reducing computational overhead. \nNoticeably, we achieve FLOPs reductions for Centerpoint model of as much as $\\mathbf{3.89}\\times$ and $\\mathbf{3.01}\\times$ on ONCE and nuScenes datasets respectively, without noticeable loss in mean Average Precision (mAP), and at most $\\mathbf{1.65}\\times$ reduction \\textbf{losslessly} for PVRCNN model on the ONCE dataset, thus pushing the boundaries of state-of-the-art performance.", "keywords": [ "3d Object Detection", "Model Pruning", "Model Compression" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "tvWD9YueN4", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Levi Lelis", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Assistant Professor" }, { "name": "Michael Bowling", "gender": "Male", "institution": "Department of Computing Science, University of Alberta", "country": "", "position": "Full Professor" }, { "name": "Zahra Bashir", "gender": "Female", "institution": "University of Alberta", "country": "CA", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 61, "n_ref_all": 91, "n_fig": 7, "n_tab": 11, "L_tab": 1431, "n_element_tab": 242, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 741, "n_element_tab_1": 60, "formula_len_all": 1252, "formula_len_all_1": 76, "len_all": 201754, "len_all_1": 74348, "len_abs": 1740, "len_title": 132, "len_sents": 77485, "len_sents_1": 36834, "n_sents": 628, "n_sents_1": 283, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1752, "title": "Assessing the Interpretability of Programmatic Policies using Large Language Models", "abs": "Programmatic representations of policies for solving sequential decision-making problems often carry the promise of interpretability. However, previous work on programmatic policies has only presented anecdotal evidence of policy interpretability. The lack of systematic evaluations of policy interpretability can be attributed to user studies being time-consuming and costly. In this paper, we introduce the LLM-based INTerpretability (LINT) score, a simple and cost-effective metric that uses large-language models (LLMs) to assess the interpretability of programmatic policies. To compute the LINT score of a policy, an LLM generates a natural language description of the policy's behavior. This description is then passed to a second LLM, which attempts to reconstruct the policy from the natural language description. The LINT score measures the behavioral similarity between the original and reconstructed policies. We hypothesized that the LINT score of programmatic policies correlates with their actual interpretability, and evaluated this hypothesis in the domains of MicroRTS and Karel the Robot. Our evaluation relied on a technique from the static obfuscation literature and a user study, where people with various levels of programming proficiency evaluated the interpretability of the programmatic policies. The results of our experiments support our hypothesis. Specifically, the LINT score decreases as the level of obfuscation of the policies increases. The user study showed that LINT can correctly distinguish the ``degree of interpretability'' of programmatic policies generated by the existing algorithms. Our results suggest that LINT can be a helpful tool for advancing the research on interpretability of programmatic policies.", "keywords": [ "Programmatic Policies", "Interpretability", "Program Synthesis" ], "rating_list": [ 5, 6, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 4, 4 ] }, { "paper_id": "tvQNysCP7C", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Bokai Lin", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Undergrad student" }, { "name": "Hao Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "TianQi Hou", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zhijie Deng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zihao Zeng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 1, "n_ref_uni": 33, "n_ref": 63, "n_ref_all": 88, "n_fig": 3, "n_tab": 3, "L_tab": 662, "n_element_tab": 102, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 532, "n_element_tab_1": 62, "formula_len_all": 922, "formula_len_all_1": 24, "len_all": 132212, "len_all_1": 66823, "len_abs": 1907, "len_title": 104, "len_sents": 44596, "len_sents_1": 33381, "n_sents": 370, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1884, "title": "In-context KV-Cache Eviction for LLMs via Attention-Gate", "abs": "The KV-Cache technique has become the standard for the inference of large language models (LLMs).\nIt caches states of self-attention to avoid recomputation. \nYet, it is widely criticized that KV-Cache can become a bottleneck of the LLM inference system, especially when confronted with ultra-large models and long-context queries. \nA natural remedy is to discard the KV-Cache for less important tokens, with StreamingLLM as an example, but the used static eviction strategies cannot flexibly adapt to varying contexts. \nRemedies like H2O leverage accumulative attention scores to perform dynamic eviction but suffer from the attention bias issue in capturing contextual information. \nThis paper bridges this gap by devising a parameterized KV-Cache eviction mechanism, dubbed as *Attention-Gate*, which accepts the whole context as input and yields eviction flags for each token to realize *in-context* eviction. \nThe subsequent self-attention module proceeds according to the flags and only the KV states for the remaining tokens need to be cached. \nThe Attention-Gates can vary among different heads and layers and be trivially plugged into pre-trained LLMs, tuned by cost-effective continual pre-training or supervised fine-tuning objectives to acquire what to discard. \nThe computational and memory overhead introduced by Attention-Gates is minimal.\nOur method is validated across multiple tasks, demonstrating both efficiency and adaptability.\nAfter a highly efficient continual pre-training, it achieves higher average accuracy and evicts more tokens compared to traditional training-free methods. \nIn supervised fine-tuning, it not only evicts many tokens but also outperforms LoRA-finetuned LLMs on some datasets, such as RTE, where it improves accuracy by 13.9% while evicting 62.8% of tokens, showing that effective eviction of redundant tokens can even enhance performance.", "keywords": [ "KV-Cache eviction", "Attention-Gate", "LLM inference acceleration" ], "rating_list": [ 5, 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 1 ], "confidence_list": [ 4, 4, 5, 4, 5 ] }, { "paper_id": "tvLnYGAaY1", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Ching Chieh Tsao", "gender": "Female", "institution": "National Cheng Kung University", "country": "TW", "position": "MS student" }, { "name": "Fang-Yi Su", "gender": "Male", "institution": "National Cheng Kung University", "country": "TW", "position": "PhD student" }, { "name": "Jung-Hsien Chiang", "gender": "Male", "institution": "National Cheng Kung University", "country": "", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 77, "n_ref_all": 101, "n_fig": 8, "n_tab": 12, "L_tab": 3943, "n_element_tab": 240, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 115, "n_element_tab_1": 4, "formula_len_all": 510, "formula_len_all_1": 463, "len_all": 151657, "len_all_1": 52540, "len_abs": 1301, "len_title": 130, "len_sents": 42456, "len_sents_1": 24882, "n_sents": 307, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1334, "title": "Rethinking the Roles of Time and Frequency Domains Before Tackling Time Series UDA", "abs": "In time-series unsupervised domain adaptation (UDA), the adaptation between temporal and frequency domain features has been relatively underexplored. To address this gap, we conduct a comprehensive series of experiments to revisit the roles of these domains in source-free UDA (SFUDA), a branch of the UDA task. Our findings reveal that the temporal domain contains more diverse features, offering higher discriminability, while the frequency domain is more domain-invariant, providing better transferability. Combining the strengths of both domains, we propose TidalFlow, a SFUDA framework that synergistically integrates temporal and frequency domain features. TidalFlow enhances feature extraction and captures subtle, class-specific features without relying on traditional alignment strategies. By utilizing simple hyperparameter adjustments and using frequency embeddings from the source domain as reference points for domain adaptation, TidalFlow achieves nearly a 10\\% improvement across five benchmark datasets in time-series UDA. This research highlights the unique strengths of both domains and marks a paradigm shift in SFUDA methods, showcasing TidalFlow\u2019s robust performance in real-world applications. Code is available at the anonymous link: \\textcolor{magenta}{\\url{https://anonymous.4open.science/r/TidalFlow-42B0/}}.", "keywords": [ "unsupervised domain adaptation", "time-series domain adaptation", "TSUDA" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "tuu4de7HL1", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Akiko Takeda", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Professor" }, { "name": "Pierre-Louis Poirion", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Researcher" }, { "name": "Rei Higuchi", "gender": "Male", "institution": "the University of Tokyo", "country": "JP", "position": "MS student" } ], "n_formula": 181, "n_formula_1": 28, "n_ref_uni": 42, "n_ref": 94, "n_ref_all": 120, "n_fig": 9, "n_tab": 6, "L_tab": 1587, "n_element_tab": 85, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 426, "n_element_tab_1": 5, "formula_len_all": 15208, "formula_len_all_1": 1967, "len_all": 231834, "len_all_1": 67606, "len_abs": 5118, "len_title": 101, "len_sents": 60506, "len_sents_1": 28573, "n_sents": 740, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 101, "L_abs": 937, "title": "Improving Convergence Guarantees of Random Subspace Second-order Algorithm for Nonconvex Optimization", "abs": "In recent years, random subspace methods have been actively studied for large-dimensional non-convex problems. Recent subspace methods have improved theoretical guarantees such as iteration complexity and local convergence rate while reducing computational costs by deriving descent directions in randomly selected low-dimensional subspaces. This paper proposes the Random Subspace Homogenized Trust Region (RSHTR) method with the best theoretical guarantees among random subspace algorithms for non-convex optimization. RSHTR achieves an $\\varepsilon$-approximate first-order stationary point in $O(\\varepsilon^{-3/2})$ iterations, converging locally at a linear rate. Furthermore, under rank-deficient conditions, RSHTR satisfies $\\varepsilon$-approximate second-order necessary condition in $O(\\varepsilon^{-3/2})$ iterations and exhibits a local quadratic convergence. Experiments on real-world datasets verify the benefits of RSHTR.", "keywords": [ "random projection", "trust region method", "non-convex optimization", "second-order stationary point", "local convergence" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 2, 2 ] }, { "paper_id": "tuEP424UQ5", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Jayden Teoh", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Undergrad student" }, { "name": "Peter Vamplew", "gender": "Male", "institution": "Federation University Australia", "country": "", "position": "Full Professor" }, { "name": "Pradeep Varakantham", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 10, "n_ref_uni": 58, "n_ref": 90, "n_ref_all": 120, "n_fig": 17, "n_tab": 2, "L_tab": 1, "n_element_tab": 1, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 744, "formula_len_all_1": 566, "len_all": 213885, "len_all_1": 65341, "len_abs": 1829, "len_title": 122, "len_sents": 75099, "len_sents_1": 32273, "n_sents": 521, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1896, "title": "On Generalization Within Multi-Objective Reinforcement Learning Algorithms", "abs": "Real-world sequential decision-making tasks often require balancing trade-offs between multiple conflicting objectives, making Multi-Objective Reinforcement Learning (MORL) an increasingly prominent field of research. Despite recent advances, existing MORL literature has narrowly focused on performance within static environments, neglecting the importance of generalizing across diverse settings. Conversely, existing research on generalization in RL has always assumed scalar rewards, overlooking the inherent multi-objectivity of real-world problems. Generalization in the multi-objective context is fundamentally more challenging, as it requires learning a Pareto set of policies addressing varying preferences across multiple objectives. In this paper, we formalize the concept of generalization in MORL and how it can be evaluated. We then contribute a novel benchmark featuring diverse multi-objective domains with parameterized environment configurations to facilitate future studies in this area. Our baseline evaluations of state-of-the-art MORL algorithms on this benchmark reveals limited generalization capabilities, suggesting significant room for improvement. Our empirical findings also expose limitations in the expressivity of scalar rewards, emphasizing the need for multi-objective specifications to achieve effective generalization. We further analyzed the algorithmic complexities within current MORL approaches that could impede the transfer in performance from the single- to multiple-environment settings. This work fills a critical gap and lays the groundwork for future research that brings together two key areas in reinforcement learning: solving multi-objective decision-making problems and generalizing across diverse environments. Code is available at: [https://anonymous.4open.science/r/morl-generalization](https://anonymous.4open.science/r/morl-generalization)", "keywords": [ "Reinforcement Learning", "Multi-Objective Reinforcement Learning", "Generalization" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "tu3qwNjrtw", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Anurag Jayant Vaidya", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Arinbj\u00f6rn Kolbeinsson", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "", "position": "Researcher" }, { "name": "Faisal Mahmood", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Associate Professor" }, { "name": "Kyle O'Brien", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Marinka Zitnik", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" }, { "name": "Shanghua Gao", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Postdoc" }, { "name": "Shiwei Liu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Thomas Hartvigsen", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Assistant Professor" }, { "name": "Tianjin Huang", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "Lecturer" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Jonathan Schwarz", "gender": "Male", "institution": "Thomson Reuters", "country": "GB", "position": "Principal Researcher" } ], "n_formula": 8, "n_formula_1": 4, "n_ref_uni": 77, "n_ref": 189, "n_ref_all": 214, "n_fig": 4, "n_tab": 10, "L_tab": 12904, "n_element_tab": 2296, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2747, "n_element_tab_1": 386, "formula_len_all": 949, "formula_len_all_1": 315, "len_all": 273206, "len_all_1": 64476, "len_abs": 1690, "len_title": 44, "len_sents": 48365, "len_sents_1": 25980, "n_sents": 415, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 44, "L_abs": 1188, "title": "Composable Interventions for Language Models", "abs": "Test-time interventions for language models can enhance factual accuracy, mitigate harmful outputs, and improve model efficiency without costly retraining.\nBut despite a flood of new methods, different types of interventions are largely developing independently.\nIn practice, multiple interventions must be applied sequentially to the same model, yet we lack standardized ways to study how interventions interact.\nWe fill this gap by introducing composable interventions, a framework to study the effects of using multiple interventions on the same language models, featuring new metrics and a unified codebase.\nUsing our framework, we conduct extensive experiments and compose popular methods from three emerging intervention categories---knowledge editing, model compression, and machine unlearning.\nOur results over 417 different compositions uncover meaningful interactions: compression hinders editing and unlearning, composing interventions hinges on their order of application, and popular general-purpose metrics are inadequate for assessing composability.\nTaken together, our findings showcase clear gaps in composability, suggesting a need for new multi-objective interventions.", "keywords": [ "Model editing", "Compression", "Unlearning", "Interventions", "Language models" ], "rating_list": [ 3, 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 3, 4 ] }, { "paper_id": "ttq44QjKda", "primary_area": "learning theory", "n_author": 1, "author_info_list": [ { "name": "Yoshiaki Kitazawa", "gender": "Male", "institution": "NTT DATA Mathematical Systems Inc. ", "country": "", "position": "Researcher" } ], "n_formula": 293, "n_formula_1": 40, "n_ref_uni": 31, "n_ref": 45, "n_ref_all": 172, "n_fig": 4, "n_tab": 2, "L_tab": 312, "n_element_tab": 11, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 97, "n_element_tab_1": 13, "formula_len_all": 26847, "formula_len_all_1": 3415, "len_all": 219241, "len_all_1": 63466, "len_abs": 1089, "len_title": 130, "len_sents": 55237, "len_sents_1": 24080, "n_sents": 700, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1108, "title": "Bounds on $L_p$ Errors in Density Ratio Estimation via $f$-Divergence Loss Functions", "abs": "Density ratio estimation (DRE) is a fundamental machine learning technique for identifying relationships between two probability distributions. $f$-divergence loss functions, derived from variational representations of $f$-divergence, are commonly employed in DRE to achieve state-of-the-art results. This study presents a novel perspective on DRE using $f$-divergence loss functions by deriving the upper and lower bounds on $L_p$ errors. These bounds apply to any estimator within a class of Lipschitz continuous estimators, irrespective of the specific $f$-divergence loss functions utilized.\nThe bounds are formulated as a product of terms that include the data dimension and the expected value of the density ratio raised to the power of $p$.\nNotably, the lower bound incorporates an exponential term dependent on the Kullback--Leibler divergence, indicating that the $L_p$ error significantly increases with the Kullback--Leibler divergence for $p > 1$, and this increase becomes more pronounced as $p$ increases.\nFurthermore, these theoretical findings are substantiated through numerical experiments.", "keywords": [ "density ratio estimation", "variational divergence optimization", "Kullback\u2013Leibler divergence", "$f$-divergence", "$L_p$ error", "the curse of dimensionality", "and GAN." ], "rating_list": [ 5, 8, 5, 6 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "tt0SCefKQL", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Gabe Guo", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Richard Zemel", "gender": "Male", "institution": "Department of Computer Science, Columbia University", "country": "", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 91, "n_ref_all": 104, "n_fig": 3, "n_tab": 1, "L_tab": 3, "n_element_tab": 1, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 3, "n_element_tab_1": 1, "formula_len_all": 210, "formula_len_all_1": 210, "len_all": 115963, "len_all_1": 61171, "len_abs": 1679, "len_title": 117, "len_sents": 36006, "len_sents_1": 31190, "n_sents": 238, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1663, "title": "Masked VAE: Distributionally-Informed Self-Supervised Vision Learning", "abs": "Masked pre-training with transformers is a popular self-supervised representation learning paradigm, initially showing success in NLP before moving to CV.\nHowever, an aspect of masked pre-training that is missing in CV is the ability to capture the distribution of possible outputs. In NLP pre-training methods, the distribution is expressed as a softmax output layer. In CV, the SoTA masked autoencoder (MAE) simply ignores the possibility of a distribution, only giving a point estimate of the masked pixels' RGB values. This formulation is fundamentally limited, as it models an under-constrained problem as well-posed. \nThis poses limitations when deployed for completion tasks: it can give only one possible completion, when in reality, a scene could be completed in many different ways, e.g., a partial kitchen could have spoons, cups, pizzas, etc. under the mask. This inability to complete multiple modes indicates the weakness of the underlying representation in capturing contextual relationships.\nTowards creating a distributionally-aware formulation with contextually-aware representations, we propose the Masked VAE, a transformer-based self-supervised learning method that combines ideas from the MAE and the variational autoencoder (VAE). Like the VAE, we model the \"masked\" latent space tokens as samples from a multivariate Gaussian distribution, while keeping the MAE's deterministic latent codes for the visible tokens. \nEvaluations show that our method can create contextually plausible masked completions in a distributionally-aware manner, while matching the state-of-the-art in representation performance in downstream classification tasks.", "keywords": [ "Self-Supervised Learning", "Transformers", "Vision", "Masked Autoencoders" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "tsfR7JCwTf", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Alexey Kurakin", "gender": "Male", "institution": "Research, Google", "country": "", "position": "Research Software Engineer" }, { "name": "Nicholas Carlini", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Sanghyun Hong", "gender": "Male", "institution": "Oregon State University", "country": "", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 0, "n_ref": 57, "n_ref_all": 79, "n_fig": 1, "n_tab": 5, "L_tab": 261, "n_element_tab": 19, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 576, "n_element_tab_1": 28, "formula_len_all": 891, "formula_len_all_1": 837, "len_all": 143321, "len_all_1": 82082, "len_abs": 973, "len_title": 119, "len_sents": 49202, "len_sents_1": 38184, "n_sents": 389, "n_sents_1": 296, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 988, "title": "Certified Robustness to Clean-label Poisoning Using Diffusion Denoising", "abs": "We present a certified defense to clean-label poisoning attacks. These attacks work by injecting a small number of poisoning samples (e.g., 1%) that contain $\\ell_2$-norm bounded adversarial perturbations into the training data to induce a targeted misclassification of a test-time input. Inspired by the adversarial robustness achieved by \\emph{denoised smoothing}, we show how an off-the-shelf diffusion model can sanitize the tampered training data. We extensively test our defense against seven clean-label poisoning attacks and reduce their attack success to 0-16% with only a negligible drop in the test time accuracy. We compare our defense with existing countermeasures against clean-label poisoning, showing that the defense reduces the attack success the most and offers the best model utility. Our results highlight the need for future work on developing stronger clean-label attacks and using our certified yet practical defense as a strong baseline to evaluate these attacks.", "keywords": [ "Data Poisoning", "Certified Robustness", "Diffusion Denoising" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "ts1F3BVaDI", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 6, "author_info_list": [ { "name": "Christian Kl\u00f6tergens", "gender": "Male", "institution": "Universit\u00e4t Hildesheim", "country": "DE", "position": "PhD student" }, { "name": "Kiran Madhusudhanan", "gender": "Male", "institution": "University of Hildesheim", "country": "DE", "position": "PhD student" }, { "name": "Lars Schmidt-Thieme", "gender": "Male", "institution": "University of Hildesheim", "country": "DE", "position": "Full Professor" }, { "name": "Randolf Scholz", "gender": "Male", "institution": "Universit\u00e4t Hildesheim", "country": "", "position": "PhD student" }, { "name": "Stefan Born", "gender": "Male", "institution": "Technische Universit\u00e4t Berlin", "country": "", "position": "Lecturer" }, { "name": "Vijaya Krishna Yalavarthi", "gender": "Male", "institution": "University of Hildesheim", "country": "DE", "position": "PhD student" } ], "n_formula": 40, "n_formula_1": 25, "n_ref_uni": 28, "n_ref": 71, "n_ref_all": 94, "n_fig": 5, "n_tab": 7, "L_tab": 2329, "n_element_tab": 161, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1320, "n_element_tab_1": 87, "formula_len_all": 4379, "formula_len_all_1": 2112, "len_all": 142670, "len_all_1": 64797, "len_abs": 4866, "len_title": 153, "len_sents": 41294, "len_sents_1": 27346, "n_sents": 343, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1605, "title": "Marginalization Consistent Mixture of Separable Flows for Probabilistic Irregular Time Series Forecasting", "abs": "Probabilistic forecasting models for joint distributions of targets in irregular time\nseries are a heavily under-researched area in machine learning with, to the best of\nour knowledge, only three models researched so far: GPR, the Gaussian Process\nRegression model (D\u00fcrichen et al., 2015), TACTiS, the Transformer-Attentional\nCopulas for Time Series Drouin et al. (2022); Ashok et al. (2024) and ProFITi\n(Yalavarthi et al., 2024b), a multivariate normalizing flow model based on invertible\nattention layers. While ProFITi, thanks to using multivariate normalizing flows,\nis the more expressive model with a better predictive performance, we will show\nthat it suffers from marginalization inconsistency: it does not guarantee that the\nmarginal distributions of a subset of variables in its predictive distributions coincide\nwith the directly predicted distributions of these variables. Also, TACTiS does not\nprovide any guarantees for marginalization consistency.\nWe develop a novel probabilistic irregular time series forecasting model, Marginal-\nization Consistent Mixtures of Separable Flows (moses), that mixes several nor-\nmalizing flows with (i) Gaussian Processes with full covariance matrix as source\ndistributions and (ii) a separable invertible transformation, aiming to combine\nthe expressivity of normalizing flows with the marginalization consistency of\nGaussians. In experiments on four different datasets we show that moses outper-\nform other state-of-the-art marginalization consistent models, perform on par with\nProFITi, but different from ProFITi, guarantees marginalization consistency.", "keywords": [ "Probabilistic modelling", "Normalizing flows", "Marginalization consitsent" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 2, 3, 5 ] }, { "paper_id": "trj2Jq8riA", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Bo Fu", "gender": "unknown", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Jiaxiang Gou", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Luping Ji", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Mao Ye", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Pei Liu", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 34, "n_formula_1": 20, "n_ref_uni": 54, "n_ref": 144, "n_ref_all": 198, "n_fig": 9, "n_tab": 13, "L_tab": 8589, "n_element_tab": 482, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1969, "n_element_tab_1": 101, "formula_len_all": 2006, "formula_len_all_1": 1114, "len_all": 244895, "len_all_1": 69710, "len_abs": 1667, "len_title": 151, "len_sents": 65387, "len_sents_1": 30793, "n_sents": 560, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 103, "L_abs": 1727, "title": "Interpretable Vision-Language Survival Analysis with Ordinal Inductive Bias for Computational Pathology", "abs": "Histopathology Whole-Slide Images (WSIs) provide an important tool to assess cancer prognosis in computational pathology (CPATH). While existing survival analysis (SA) approaches have made exciting progress, they are generally limited to adopting highly-expressive network architectures and only coarse-grained patient-level labels to learn visual prognostic representations from gigapixel WSIs. Such learning paradigm suffers from critical performance bottlenecks, when facing present scarce training data and standard multi-instance learning (MIL) framework in CPATH. To overcome it, this paper, for the first time, proposes a new Vision-Language-based SA (**VLSA**) paradigm. Concretely, (1) VLSA is driven by pathology VL foundation models. It no longer relies on high-capability networks and shows the advantage of *data efficiency*. (2) In vision-end, VLSA encodes textual prognostic prior and then employs it as *auxiliary signals* to guide the aggregating of visual prognostic features at instance level, thereby compensating for the weak supervision in MIL. Moreover, given the characteristics of SA, we propose i) *ordinal survival prompt learning* to transform continuous survival labels into textual prompts; and ii) *ordinal incidence function* as prediction target to make SA compatible with VL-based prediction. Notably, VLSA's predictions can be interpreted intuitively by our Shapley values-based method. The extensive experiments on five datasets confirm the effectiveness of our scheme. Our VLSA could pave a new way for SA in CPATH by offering weakly-supervised MIL an effective means to learn valuable prognostic clues from gigapixel WSIs. Our source code is available at https://github.com/liupei101/VLSA.", "keywords": [ "Computation Pathology", "Survival Analysis", "Multi-Instance Learning", "Whole-Slide Images", "Vision-Language Modes" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "tra8ktyk0E", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "He Sun", "gender": "Male", "institution": "University of Edinburgh", "country": "GB", "position": "Associate Professor" }, { "name": "Peter Macgregor", "gender": "unknown", "institution": "University of St. Andrews", "country": "GB", "position": "Lecturer" }, { "name": "Steinar Laenen", "gender": "Male", "institution": "University of Edinburgh", "country": "GB", "position": "PhD student" } ], "n_formula": 127, "n_formula_1": 20, "n_ref_uni": 32, "n_ref": 73, "n_ref_all": 94, "n_fig": 5, "n_tab": 14, "L_tab": 3361, "n_element_tab": 241, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1152, "n_element_tab_1": 78, "formula_len_all": 12077, "formula_len_all_1": 1669, "len_all": 270847, "len_all_1": 78194, "len_abs": 708, "len_title": 116, "len_sents": 69656, "len_sents_1": 27327, "n_sents": 706, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 842, "title": "Dynamic Similarity Graph Construction with Kernel Density Estimation", "abs": "In the kernel density estimation (KDE) problem, we are given a set $X$ of data points in $\\mathbb{R}^d$, a kernel function $k: \\mathbb{R}^d \\times \\mathbb{R}^d \\rightarrow \\mathbb{R}$, and a query point $\\mathbf{q} \\in \\mathbb{R}^d$, and the objective is to quickly output an estimate of $\\sum_{\\mathbf{x} \\in X} k(\\mathbf{q}, \\mathbf{x})$.\nIn this paper, we consider $\\textsf{KDE}$ in the dynamic setting, and introduce a data structure that efficiently maintains the estimates for a set of query points as data points are added to $X$ over time.\nBased on this, we design a dynamic data structure that maintains a sparse approximation of the fully connected similarity graph on \n$X$, and develop a fast dynamic spectral clustering algorithm.\nWe further evaluate the effectiveness of our algorithms on both synthetic and real-world datasets.", "keywords": [ "kernel density estimation", "similarity graphs", "spectral clustering" ], "rating_list": [ 3, 5, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "trV41CpAK4", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Hairong Qi", "gender": "Female", "institution": "University of Tennessee, Knoxville", "country": "US", "position": "Full Professor" }, { "name": "Han Zhao", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Hongjue Zhao", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Huajie Shao", "gender": "Male", "institution": "College of William and Mary", "country": "US", "position": "Assistant Professor" }, { "name": "Lui Sha", "gender": "Male", "institution": "Department of Computer Science", "country": "", "position": "Full Professor" }, { "name": "Yuchen Wang", "gender": "Male", "institution": "College of William and Mary", "country": "US", "position": "PhD student" }, { "name": "Zijie Huang", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 47, "n_formula_1": 16, "n_ref_uni": 40, "n_ref": 104, "n_ref_all": 144, "n_fig": 10, "n_tab": 15, "L_tab": 6292, "n_element_tab": 272, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2526, "n_element_tab_1": 85, "formula_len_all": 5033, "formula_len_all_1": 1518, "len_all": 191972, "len_all_1": 73671, "len_abs": 1549, "len_title": 114, "len_sents": 59016, "len_sents_1": 31425, "n_sents": 544, "n_sents_1": 289, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1563, "title": "Accelerating Neural ODEs: A Variational Formulation-based Approach", "abs": "Neural Ordinary Differential Equations (Neural ODEs or NODEs) excel at modeling continuous dynamical systems from observational data, especially when the data is irregularly sampled. However, existing training methods predominantly rely on numerical ODE solvers, which are time-consuming and prone to accumulating numerical errors over time due to autoregression. In this work, we propose the VF-NODE, a novel approach based on the variational formulation (VF) to accelerate the training of NODEs. Unlike existing training methods, the proposed VF-NODEs implement a series of global integrals, thus evaluating Deep Neural Network (DNN)--based vector fields only at specific observed data points. This strategy drastically reduces the number of function evaluations (NFEs). Moreover, our method eliminates the use of autoregression, thereby reducing error accumulations for modeling dynamical systems. Nevertheless, the VF loss introduces oscillatory terms into the integrals when using the Fourier basis. We incorporate Filon's method to address this issue. To further enhance the performance for noisy and incomplete data, we employ the natural cubic spline regression to estimate a closed-form approximation. We provide a fundamental analysis of how our approach minimizes computational costs. Extensive experiments demonstrate that our approach accelerates NODE training by 10 to 1000 times compared to existing NODE-based methods, while achieving higher or comparable accuracy in dynamical systems. The source code will be publicly available upon publication.", "keywords": [ "Neural ordinary differential equations", "irregularly-sampled dynamical systems", "variational formulation", "acceleration" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "trKee5pIFv", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Anirban Das", "gender": "Male", "institution": "Capital One", "country": "", "position": "Researcher" }, { "name": "David Yao", "gender": "unknown", "institution": "Columbia University", "country": "", "position": "Full Professor" }, { "name": "Genta Indra Winata", "gender": "Male", "institution": "Capital One", "country": "US", "position": "Principal Researcher" }, { "name": "Hanyang Zhao", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Sambit Sahu", "gender": "Male", "institution": "CapitalOne", "country": "US", "position": "Principal Researcher" }, { "name": "Shi-Xiong Zhang", "gender": "Male", "institution": "CapitalOne", "country": "US", "position": "Principal Researcher" }, { "name": "Wenpin Tang", "gender": "Male", "institution": "Columbia University", "country": "", "position": "Assistant Professor" } ], "n_formula": 51, "n_formula_1": 24, "n_ref_uni": 37, "n_ref": 89, "n_ref_all": 117, "n_fig": 6, "n_tab": 9, "L_tab": 3037, "n_element_tab": 387, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1962, "n_element_tab_1": 162, "formula_len_all": 6052, "formula_len_all_1": 2605, "len_all": 169617, "len_all_1": 65937, "len_abs": 981, "len_title": 134, "len_sents": 46635, "len_sents_1": 28663, "n_sents": 334, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 985, "title": "RainbowPO: A Unified Framework for Combining Improvements in Preference Optimization", "abs": "Recently, numerous preference optimization algorithms have been introduced as extensions to the Direct Preference Optimization (DPO) family. While these methods have successfully aligned models with human preferences, there is a lack of understanding regarding the contributions of their additional components. Moreover, fair and consistent comparisons are scarce, making it difficult to discern which components genuinely enhance downstream performance. In this work, we propose RainbowPO, a unified framework that demystifies the effectiveness of existing DPO methods by categorizing their key components into seven broad directions. We integrate these components into a single cohesive objective, enhancing the performance of each individual element. Through extensive experiments, we demonstrate that RainbowPO outperforms existing DPO variants. Additionally, we provide insights to guide researchers in developing new DPO methods and assist practitioners in their implementations.", "keywords": [ "Alignment", "Preference Optimization", "Reinforcement Learning from Human Feedback" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "trKNi4IUiP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Enyan Dai", "gender": "Male", "institution": "HKUST(GZ)", "country": "CN", "position": "Assistant Professor" }, { "name": "Junjie Xu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Minhua Lin", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Suhang Wang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Associate Professor" }, { "name": "Zhiwei Zhang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Zongyu Wu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 13, "n_ref_uni": 40, "n_ref": 111, "n_ref_all": 134, "n_fig": 7, "n_tab": 12, "L_tab": 4320, "n_element_tab": 778, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1905, "n_element_tab_1": 326, "formula_len_all": 2743, "formula_len_all_1": 693, "len_all": 216460, "len_all_1": 81778, "len_abs": 5902, "len_title": 90, "len_sents": 76314, "len_sents_1": 38336, "n_sents": 603, "n_sents_1": 284, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 42, "L_abs": 1224, "title": "Robustness Inspired Graph Backdoor Defense", "abs": "Graph Neural Networks (GNNs) have achieved promising results in tasks such as node classification and graph classification. However, recent studies reveal that GNNs are vulnerable to backdoor attacks, posing a significant threat to their real-world adoption. Despite initial efforts to defend against specific graph backdoor attacks, there is no work on defending against various types of backdoor attacks where generated triggers have different properties. Hence, we first empirically verify that prediction variance under edge dropping is a crucial indicator for identifying poisoned nodes. With this observation, we propose using random edge dropping to detect backdoors and theoretically show that it can efficiently distinguish poisoned nodes from clean ones. Furthermore, we introduce a novel robust training strategy to efficiently counteract the impact of the triggers. Extensive experiments on real-world datasets show that our framework can effectively identify poisoned nodes, significantly degrade the attack success rate, and maintain clean accuracy when defending against various types of graph backdoor attacks with different properties. Our code is available at: https://anonymous.4open.science/r/RIGBD-A670.", "keywords": [ "Backdoor Defense", "Graph Neural Network" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 2 ] }, { "paper_id": "tqYx8DgL0u", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Chao Lin", "gender": "Male", "institution": "Fujian Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Shengmin Xu", "gender": "Male", "institution": "Fujian Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Wenhan Dong", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Xinlei He", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Xinyi Huang", "gender": "Male", "institution": "Jinan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 34, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 47, "n_ref_all": 56, "n_fig": 1, "n_tab": 11, "L_tab": 3682, "n_element_tab": 241, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1467, "n_element_tab_1": 137, "formula_len_all": 2778, "formula_len_all_1": 834, "len_all": 164624, "len_all_1": 60614, "len_abs": 2063, "len_title": 122, "len_sents": 54320, "len_sents_1": 26514, "n_sents": 420, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1236, "title": "Privacy-Preserving Federated Learning via Homomorphic Adversarial Networks", "abs": "Privacy-preserving federated learning (PPFL) aims to train a global model for multiple clients while maintaining their data privacy. However, current PPFL protocols exhibit one or more of the following insufficiencies: considerable degradation in accuracy, the requirement for sharing keys, and cooperation during the key generation or decryption processes. As a mitigation, we develop the first protocol that utilizes neural networks to preserve privacy in federated learning, as well as incorporating an Aggregatable Hybrid Encryption scheme tailored to the needs of the PPFL. We name these networks as Homomorphic Adversarial Networks (HANs) which demonstrate that neural networks are capable of performing tasks similar to multi-key homomorphic encryption (MK-HE) while solving the problems of key distribution and collaborative decryption. Our experiments show that HANs are robust against privacy attacks. Compared with non-private federated learning, experiments conducted on multiple datasets demonstrate that HANs exhibit a negligible accuracy loss (at most 1.35\\%). Compared to traditional MK-HE schemes, HANs increase encryption aggregation speed by 6,075 times while incurring a 29.2-fold increase in communication overhead.", "keywords": [ "Federated Learning", "Privacy Protection", "Homomorphic Encryption", "Homomorphic Adversarial Networks" ], "rating_list": [ 3, 3, 5, 3, 3 ], "soundness_list": [ 1, 3, 2, 2, 2 ], "presentation_list": [ 1, 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 3, 5, 4 ] }, { "paper_id": "tpqMR73GzS", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Peter David Fagan", "gender": "Male", "institution": "University of Edinburgh", "country": "", "position": "PhD student" }, { "name": "Subramanian Ramamoorthy", "gender": "Male", "institution": "Edinburgh University, University of Edinburgh", "country": "GB", "position": "Full Professor" } ], "n_formula": 23, "n_formula_1": 7, "n_ref_uni": 31, "n_ref": 65, "n_ref_all": 79, "n_fig": 3, "n_tab": 8, "L_tab": 707, "n_element_tab": 83, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2287, "n_element_tab_1": 62, "formula_len_all": 1352, "formula_len_all_1": 381, "len_all": 126071, "len_all_1": 61045, "len_abs": 1599, "len_title": 115, "len_sents": 45467, "len_sents_1": 29730, "n_sents": 267, "n_sents_1": 158, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1605, "title": "Learning from Demonstration with Implicit Nonlinear Dynamics Models", "abs": "Learning from Demonstration (LfD) is a useful paradigm for training policies that solve tasks involving complex motions, such as those encountered in robotic manipulation. In practice, the successful application of LfD requires overcoming error accumulation during policy execution, i.e. the problem of drift due to errors compounding over time and the consequent out-of-distribution behaviours. Existing works seek to address this problem through scaling data collection, correcting policy errors with a human-in-the-loop, temporally ensembling policy predictions or through learning a dynamical system model with convergence guarantees. In this work, we propose and validate an alternative approach to overcoming this issue. Inspired by reservoir computing, we develop a recurrent neural network layer that includes a fixed nonlinear dynamical system with tunable dynamical properties for modelling temporal dynamics. We validate the efficacy of our neural network layer on the task of reproducing human handwriting motions using the LASA Human Handwriting Dataset. Through empirical experiments we demonstrate that incorporating our layer into existing neural network architectures addresses the issue of compounding errors in LfD. Furthermore, we perform a comparative evaluation against existing approaches including a temporal ensemble of policy predictions and an Echo State Network (ESN) implementation. We find that our approach yields greater policy precision and robustness on the handwriting task while also generalising to multiple dynamics regimes and maintaining competitive latency scores.", "keywords": [ "robotics; learning from demonstration; recurrent neural networks; reservoir computing;" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 2, 2, 4 ] }, { "paper_id": "tpYeermigp", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Dennis Kochmann", "gender": "unknown", "institution": "California Institute of Technology", "country": "", "position": "" }, { "name": "Jan-Hendrik Bastek", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "PhD student" }, { "name": "WaiChing Sun", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Associate Professor" } ], "n_formula": 38, "n_formula_1": 16, "n_ref_uni": 40, "n_ref": 147, "n_ref_all": 168, "n_fig": 10, "n_tab": 1, "L_tab": 92, "n_element_tab": 2, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3212, "formula_len_all_1": 1154, "len_all": 192118, "len_all_1": 63097, "len_abs": 1069, "len_title": 81, "len_sents": 59938, "len_sents_1": 29955, "n_sents": 377, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 33, "L_abs": 1153, "title": "Physics-Informed Diffusion Models", "abs": "Generative models such as denoising diffusion models are quickly advancing their ability to approximate highly complex data distributions. They are also increasingly leveraged in scientific machine learning, where samples from the implied data distribution are expected to adhere to specific governing equations. We present a framework that unifies generative modeling and partial differential equation fulfillment by introducing a first-principle-based loss term that enforces generated samples to fulfill the underlying physical constraints. Our approach reduces the residual error by up to two orders of magnitudes compared to previous work in a fluid flow case study and outperforms task-specific frameworks in relevant metrics for structural topology optimization. We also present numerical evidence that our extended training objective acts as a natural regularization mechanism against overfitting. Our framework is simple to implement and versatile in its applicability for imposing equality and inequality constraints as well as auxiliary optimization objectives. Code is available at https://github.com/jhbastek/PhysicsInformedDiffusionModels.", "keywords": [ "physics-informed", "scientific machine learning", "denoising diffusion", "inverse problems", "generative modeling" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "tpVQHb4pea", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Huimu Yu", "gender": "Male", "institution": "Institute of Information Engineering, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Weidong Yin", "gender": "Male", "institution": "University of British Columbia", "country": "", "position": "MS student" }, { "name": "Xing W", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Songiln Hu", "gender": "Male", "institution": "Institute of Information Engeering, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 46, "n_ref": 84, "n_ref_all": 114, "n_fig": 9, "n_tab": 8, "L_tab": 857, "n_element_tab": 79, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 343, "n_element_tab_1": 21, "formula_len_all": 137, "formula_len_all_1": 160, "len_all": 153741, "len_all_1": 50815, "len_abs": 1074, "len_title": 130, "len_sents": 38719, "len_sents_1": 23197, "n_sents": 295, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1275, "title": "CodePMP: Scalable Preference Model Pretraining for Large Language Model Reasoning", "abs": "Large language models (LLMs) have achieved notable advancements in natural language understanding and generation, driven by scalable pretraining and advanced finetuning techniques. However, improving reasoning abilities in LLMs, particularly through reinforcement learning from human feedback (RLHF), remains a challenge due to the scarcity of high-quality preference data, which is often labor-intensive to annotate and essential for reward model (RM) finetuning. To alleviate this issue, we introduce CodePMP, a scalable preference model pretraining (PMP) pipeline that leverages vast amounts of code-preference pairs synthesized from publicly available, high-quality source code. CodePMP improves the sample efficiency of RM finetuning by sufficiently pre-training preference models on synthesized code-preference pairs. In addition to validating CodePMP on widely used mathematical reasoning tasks (GSM8K, MATH), we also demonstrate its effectiveness on logical reasoning benchmarks (ReClor, LogiQA). The results consistently indicate that CodePMP significantly improves the reasoning performance of large language models (LLMs). Furthermore, our findings underscore the critical role of scalable preference model pretraining (PMP) in achieving efficient reward modeling.", "keywords": [ "Preference Model Pretraining" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "tpUEqmjZiS", "primary_area": "applications to robotics, autonomy, planning", "n_author": 8, "author_info_list": [ { "name": "Bin Zhao", "gender": "Male", "institution": "Shanghai Artificial Intelligence Labortory", "country": "CN", "position": "Young Scientist" }, { "name": "Dong Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Haoming Song", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xuelong Li", "gender": "Male", "institution": "China Telecom", "country": "CN", "position": "Full Professor" }, { "name": "Yan Ding", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yuanqi Yao", "gender": "Female", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Liu Siao", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Zhi.gang Wang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 16, "n_ref": 22, "n_ref_all": 40, "n_fig": 8, "n_tab": 3, "L_tab": 436, "n_element_tab": 34, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1322, "n_element_tab_1": 16, "formula_len_all": 755, "formula_len_all_1": 720, "len_all": 98159, "len_all_1": 57231, "len_abs": 1302, "len_title": 140, "len_sents": 28833, "len_sents_1": 26526, "n_sents": 209, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1309, "title": "Think Small, Act Big: Primitive-level Skill Prompt Learning for Lifelong Robot Manipulation", "abs": "The general-purpose robots need to continuously acquire new skills in lifelong spans without revisiting past experiences, known as Rehearsal-free Lifelong Learning, which remains significantly challenging. Recent advances learn a separate adapter along pretrained policy for each new skill to address catastrophic forgetting problem, ignoring the shared knowledge between old skills and new ones. To tackle these issues, we propose Primitive-level Skill Prompt Learning (PSPL), to achieve lifelong robot manipulation via reusable and extensible primitives. Within our two stage learning scheme, we first learn a set of prefix skill prompts to extract shared knowledge through multi-skills pre-training stage, where motion-aware skill prompts are learned to capture semantic and motion shared primitives across different skills. Secondly, when acquiring new skills in lifelong span, new prefix skill prompts are added and learned via cross-attention between prefix prompts of old skills, boosting the new skills learning via shared knowledge transfer. For evaluation, we construct a large-scale skill dataset and conduct extensive experiments in both simulation and real-world tasks, demonstrating PSPL's superior performance over state-of-the-art methods. Code and dataset will be released upon acceptance.", "keywords": [ "Robotics", "Imitation Learning" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "tpHqsyZ3YX", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Guang Dai", "gender": "Male", "institution": "SGIT AI", "country": "CN", "position": "Principal Researcher" }, { "name": "Haishan Ye", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Hao Di", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Tong He", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xiangyu Chang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "Associate Professor" }, { "name": "Yinghui Huang", "gender": "Female", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ivor W Tsang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" } ], "n_formula": 14, "n_formula_1": 7, "n_ref_uni": 29, "n_ref": 80, "n_ref_all": 119, "n_fig": 16, "n_tab": 12, "L_tab": 2028, "n_element_tab": 168, "n_fig_1": 8, "n_tab_1": 7, "L_tab_1": 2128, "n_element_tab_1": 189, "formula_len_all": 1405, "formula_len_all_1": 672, "len_all": 170698, "len_all_1": 69156, "len_abs": 2001, "len_title": 128, "len_sents": 60042, "len_sents_1": 30995, "n_sents": 400, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1927, "title": "ProAdvPrompter: A Two-Stage Journey to Effective Adversarial Prompting for LLMs", "abs": "As large language models (LLMs) are increasingly being integrated into various real-world applications, the identification of their vulnerabilities to jailbreaking attacks becomes an essential component of ensuring the safety and reliability of LLMs. \nPrevious studies have developed LLM assistants, known as the adversarial prompter, to automatically generate suffixes that manipulate target LLMs into generating harmful and undesirable outputs.\nHowever, these approaches often suffer from low performance or generate semantically meaningless prompts, which can be easily identified by perplexity-based defenses.\nIn this paper, we introduce a novel two-stage method, $\\texttt{ProAdvPrompter}$, that significantly improves the performance of adversarial prompters.\nIn $\\texttt{ProAdvPrompter}$, the first stage (Exploration) utilizes the loss information to guide the adversarial prompter in generating suffixes that are more likely to elicit harmful responses.\nThen the second stage (Exploitation) iteratively fine-tunes the prompter using high-quality generated adversarial suffixes to further boost performance.\nAdditionally, we incorporate the prompt template to aid in the Exploration stage and propose a filtering mechanism to accelerate the training process in the Exploitation stage.\nWe evaluate $\\texttt{ProAdvPrompter}$ against the well-aligned LLMs (i.e., Llama2-Chat-7B and Llama3-chat-8B), achieving attack success rates of 99.68% and 97.12% respectively after 10 trials on the AdvBench dataset, thereby enhancing performance by $\\sim 2$ times compared to previous works.\nMoreover, $\\texttt{ProAdvPrompter}$ reduces training time by 20% on Llama3-Instruct-8B, generates more generalized adversarial suffixes, and demonstrates resilience against the perplexity defense.\nAn ablation study further evaluates the effects of key components in $\\texttt{ProAdvPrompter}$ (the prompt template and the filtering mechanism).", "keywords": [ "jailbreaking attacks; large language model" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "tpGkEgxMJT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Bernd Bohnet", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Dale Schuurmans", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Professor" }, { "name": "Hanie Sedghi", "gender": "Female", "institution": "Google DeepMind", "country": "", "position": "Staff Research Scientist" }, { "name": "Xinran Zhao", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Azade Nazi", "gender": "Female", "institution": "Google Brain", "country": "", "position": "Research Scientist" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 44, "n_ref": 87, "n_ref_all": 123, "n_fig": 11, "n_tab": 4, "L_tab": 234, "n_element_tab": 15, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 554, "n_element_tab_1": 35, "formula_len_all": 1001, "formula_len_all_1": 92, "len_all": 185050, "len_all_1": 67656, "len_abs": 1753, "len_title": 119, "len_sents": 60673, "len_sents_1": 34134, "n_sents": 443, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1763, "title": "Improving Large Language Model Planning with Action Sequence Similarity", "abs": "Planning is essential for artificial intelligence systems to look ahead and proactively determine a course of actions to reach objectives in the virtual and real world. Recent work on large language models (LLMs) sheds light on their planning capability in various tasks. However, it remains unclear what signals in the context influence the model performance. In this work, we explore how to improve the model planning capability through in-context learning (ICL), specifically, what signals can help select the exemplars. Through extensive experiments, we observe that commonly used problem similarity may result in false positives with drastically different plans, which can mislead the model. In response, we propose to sample and filter exemplars leveraging plan side action sequence similarity (AS). We propose GRASE-DC: a two-stage pipeline that first re-samples high AS exemplars and then curates the selected exemplars with dynamic clustering on AS to achieve a balance of relevance and diversity. Our experimental result confirms that GRASE-DC achieves significant performance improvement on various planning tasks (up to ~11-40 point absolute accuracy improvement with 27.3% fewer exemplars needed on average). With GRASE-DC* + VAL, where we iteratively apply GRASE-DC with a validator, we are able to even boost the performance by 18.9% more.\nExtensive analysis validates the consistent performance improvement of GRASE-DC with various backbone LLMs and on both classical planning and natural language planning benchmarks. GRASE-DC can further boost the planning accuracy by ~24 absolute points on harder problems using simpler problems as exemplars over a random baseline. This demonstrates its ability to generalize to out-of-distribution problems.", "keywords": [ "planning with LLM", "in-context learning (ICL)", "action sequence" ], "rating_list": [ 5, 5, 8, 5 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "tpD1rs25Uu", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Guikun Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Minghan Chen", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Wenguan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 86, "n_ref": 204, "n_ref_all": 226, "n_fig": 11, "n_tab": 6, "L_tab": 3867, "n_element_tab": 390, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 796, "n_element_tab_1": 121, "formula_len_all": 522, "formula_len_all_1": 428, "len_all": 181347, "len_all_1": 55385, "len_abs": 1459, "len_title": 122, "len_sents": 41535, "len_sents_1": 23901, "n_sents": 332, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1466, "title": "Hydra-SGG: Hybrid Relation Assignment for One-stage Scene Graph Generation", "abs": "DETR introduces a simplified one-stage framework for scene graph generation (SGG) but faces challenges of sparse supervision and false negative samples. The former occurs because each image typically contains fewer than 10 relation annotations, while DETR-based SGG models employ over 100 relation queries. Each ground truth relation is assigned to only one query during training. The latter arises when one ground truth relation may have multiple queries with similar matching scores, leading to suboptimally matched queries being treated as negative samples. To address these, we propose Hydra-SGG, a one-stage SGG method featuring a Hybrid Relation Assignment. This approach combines a One-to-One Relation Assignment with an IoU-based One-to-Many Relation Assignment, increasing positive training samples and mitigating sparse supervision. In addition, we empirically demonstrate that removing self-attention between relation queries leads to duplicate predictions, which actually benefits the proposed One-to-Many Relation Assignment. With this insight, we introduce Hydra Branch, an auxiliary decoder without self-attention layers, to further enhance One-to-Many Relation Assignment by promoting different queries to make the same relation prediction. Hydra-SGG achieves state-of-the-art performance on multiple datasets, including VG150 (16.0 mR@50), Open Images V6 (50.1 weighted score), and GQA (12.7 mR@50). Our code and pre-trained models will be released.", "keywords": [ "Scene Graph Generation", "Visual Relation Detection" ], "rating_list": [ 8, 6, 5 ], "soundness_list": [ 4, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 4, 2, 2 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "tp1QiTH2aa", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Hao Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Junda Chen", "gender": "Not Specified", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Siqi Zhu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yichao Fu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 4, "n_ref_uni": 29, "n_ref": 54, "n_ref_all": 70, "n_fig": 8, "n_tab": 8, "L_tab": 2340, "n_element_tab": 258, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1173, "n_element_tab_1": 171, "formula_len_all": 513, "formula_len_all_1": 243, "len_all": 135155, "len_all_1": 64365, "len_abs": 1759, "len_title": 118, "len_sents": 36533, "len_sents_1": 29252, "n_sents": 277, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1197, "title": "Cost-Effective Synthetic Data Generation for Post-Training using QWICK", "abs": "Large language models (LLMs) are showing expert-level ability in various fields (e.g., programming and math). However, this progress heavily relies on the generation of high-quality synthetic data to improve the models\u2019 capabilities during post-training. Generating such data in a cost-effective manner presents a significant challenge. Specifically, stronger models tend to generate higher-quality data but come with a substantial computational cost, while weaker models are cheaper to run but may produce weaker outputs. In this paper, we introduce Question-Wise model pICK (QWICK) to address this challenge. By tracking the empirical reward, cost, and number of trials for each model, QWICK strikes a balance between exploitation and exploration, ultimately converging on a cost-effective model for each specific question. Specifically, QWICK achieves a 50\\% cost reduction on a programming dataset and a 40\\% cost reduction on a mathematics dataset, without compromising data quality. Furthermore, compared to baseline methods, our approach can produce up to 2.1 times more valid synthetic data at the same cost. Our anonymized code is available at https://anonymous.4open.science/r/QWICK-17C3", "keywords": [ "synthetic data generation", "post-training" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "tozlOEN4qp", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Arash Vahdat", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Jaideep Pathak", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Kushagra Pandey", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Michael Pritchard", "gender": "Male", "institution": "University of California, Irvine", "country": "", "position": "Associate Professor" }, { "name": "Morteza Mardani", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Principal Researcher" }, { "name": "Yilun Xu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Stephan Mandt", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Associate Professor" } ], "n_formula": 143, "n_formula_1": 18, "n_ref_uni": 44, "n_ref": 207, "n_ref_all": 249, "n_fig": 9, "n_tab": 15, "L_tab": 7269, "n_element_tab": 515, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1322, "n_element_tab_1": 90, "formula_len_all": 13250, "formula_len_all_1": 2004, "len_all": 268037, "len_all_1": 59519, "len_abs": 3272, "len_title": 77, "len_sents": 79690, "len_sents_1": 25781, "n_sents": 761, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 29, "L_abs": 1392, "title": "Heavy-Tailed Diffusion Models", "abs": "Diffusion models achieve state-of-the-art generation quality across many applications, but their ability to capture rare or extreme events in heavy-tailed distributions remains unclear. In this work, we show that traditional diffusion and flow-matching models with standard Gaussian priors fail to accurately capture heavy-tailed behavior. We address this by repurposing the diffusion framework for heavy-tail estimation using multivariate Student-t distributions. We develop a tailored perturbation kernel and derive the denoising posterior based on the conditional Student-t distribution for the backward process. Inspired by $\\gamma$-divergence for heavy-tailed distributions, we derive a training objective for heavy-tailed denoisers. The resulting framework introduces controllable tail generation using only a single scalar hyperparameter, making it easily tunable for diverse real-world distributions. As specific instantiations of our framework, we introduce t-EDM and t-Flow, extensions of existing diffusion and flow models that employ a Student-t prior. Remarkably, our approach is readily compatible with standard Gaussian diffusion models and requires only minimal code changes. Empirically, we show that our t-EDM and t-Flow outperform standard diffusion models in heavy-tail estimation on high-resolution weather datasets in which generating rare and extreme events is crucial.", "keywords": [ "Generative models", "Diffusion models", "Heavy-Tailed Distributions" ], "rating_list": [ 8, 3, 8, 6, 6 ], "soundness_list": [ 4, 2, 4, 3, 3 ], "presentation_list": [ 4, 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "torbeUlslS", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Jing Gu", "gender": "unknown", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Licheng Jiao", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Mengnan Qi", "gender": "Male", "institution": "Xidian University", "country": "", "position": "MS student" }, { "name": "Shasha Mao", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Xuequan Lu", "gender": "unknown", "institution": "La Trobe University", "country": "AU", "position": "Associate Professor" }, { "name": "Yimeng Zhang", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Zhaolong Du", "gender": "Male", "institution": "Xidian University ", "country": "CN", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 26, "n_ref": 30, "n_ref_all": 53, "n_fig": 2, "n_tab": 5, "L_tab": 1262, "n_element_tab": 150, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 492, "n_element_tab_1": 83, "formula_len_all": 1346, "formula_len_all_1": 1178, "len_all": 113243, "len_all_1": 64945, "len_abs": 1934, "len_title": 125, "len_sents": 39077, "len_sents_1": 30035, "n_sents": 280, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1958, "title": "Rethinking Multiple-Instance Learning From Feature Space to Probability Space", "abs": "Multiple-instance learning (MIL) was initially proposed to identify key instances within a set (bag) of instances when only one bag-level label is provided. Current deep MIL models mostly solve multi-instance aggregation problem in feature space. Nevertheless, with the increasing complexity of data, this paradigm faces significant risks in representation learning stage, which could lead to algorithm degradation in deep MIL models. We verify that the degradation issue stems from the persistent drift of instances in feature space during representation learning. In this paper, we propose a novel Probability-Space MIL network (PSMIL) as a countermeasure. In PSMIL, a self-training alignment strategy is introduced in probability space to solve the drift problem in feature space, and the alignment target objective is proven mathematically optimal. Furthermore, we reveal that the widely-used attention-based pooling mechanism in current deep MIL models is easily affected by the perturbation in feature space and further introduce an alternative called probability-space attention pooling. It effectively captures the key instance in each bag from feature space to probability space, and further eliminates the impact of selection drift in the pooling stage. To summarize, PSMIL seeks to solve a MIL problem in probability space rather than feature space. We also introduce new comprehensive benchmarks to evaluate the representation quality for MIL models. Experimental results illustrate that our method could potentially achieve performance close to supervised learning level in complex tasks (gap within 5\\%). The incremental alignment could also bring more than 19\\% accuracy improvements for current existing mainstream models in challenging scenarios. For various existing MIL datasets, our method also achieves competitive performance to the state-of-the-art deep MIL models. $\\textit{Our code and benchmarks will be released to the community.}$", "keywords": [ "Multiple-Instance Learning\uff1bRepresentation Learning\uff1bSemi-supervised Learning; Weakly-supervised Learning;" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "toqQYz2N2X", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Andrii Zadaianchuk", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Postdoc" }, { "name": "Daniil Cherniavskii", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "PhD student" }, { "name": "Efstratios Gavves", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Associate Professor" }, { "name": "Phillip Lippe", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 29, "n_ref": 48, "n_ref_all": 64, "n_fig": 8, "n_tab": 2, "L_tab": 148, "n_element_tab": 14, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 2125, "n_element_tab_1": 100, "formula_len_all": 137, "formula_len_all_1": 73, "len_all": 100424, "len_all_1": 58729, "len_abs": 1115, "len_title": 123, "len_sents": 31780, "len_sents_1": 27274, "n_sents": 245, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1120, "title": "Benchmark for Temporal, Ambiguous, and Grounded Embodied Question-Answering", "abs": "The problem of question ambiguity, while highlighted as an open issue, is often overlooked in the literature on Embodied Question Answering (EQA) and Episodic Memory Question Answering (EM-EQA). This paper proposes a structured approach to handle ambiguity in the egocentric data. Our benchmark, called TAG-EQA, utilizes spatial and temporal grounding to distinguish between objects, positions, and events and ensures that obtained structured answers preserve information fully while effectively resolving ambiguity. We introduce a new dataset, specifically designed for ambiguous grounded Episodic Memory QA. The dataset incorporates situated spatial reasoning, temporal conditions, and diverse visual features. Our new evaluation procedure tackles grounded natural language answers. It reveals that some of the most modern approaches still struggle with efficient information extraction and processing in ambiguous scenarios. We hope that TAG-EQA will serve as both a valuable tool for generating complex EM-EQA data and that the proposed evaluation benchmark will propel progress in agentic AI and embodied reasoning.", "keywords": [ "embodied question answering", "episodic memory question answering", "ambiguity" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 3, 2, 1, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "tolvZ5BS50", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Lei Feng", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" }, { "name": "Xiaofeng Zhu", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoshuang Shi", "gender": "Not Specified", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yujie Mo", "gender": "unknown", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Jin Cheng Huang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 49, "n_formula_1": 17, "n_ref_uni": 37, "n_ref": 77, "n_ref_all": 117, "n_fig": 10, "n_tab": 6, "L_tab": 1864, "n_element_tab": 188, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1383, "n_element_tab_1": 120, "formula_len_all": 4065, "formula_len_all_1": 889, "len_all": 166745, "len_all_1": 67577, "len_abs": 943, "len_title": 112, "len_sents": 50246, "len_sents_1": 30757, "n_sents": 404, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 949, "title": "ELU-GCN: Effectively Label-Utilizing Graph Convolutional Network", "abs": "The message-passing mechanism of graph convolutional networks (i.e., GCNs) enables label information to be propagated to a broader range of neighbors, thereby increasing the utilization of labels. However, the label information is not always effectively utilized in the traditional GCN framework. To address this issue, we propose a new two-step framework called ELU-GCN. In the first stage, ELU-GCN conducts graph learning to learn a new graph structure (i.e., ELU-graph), which enables GCNs to effectively utilize label information. In the second stage, we design a new graph contrastive learning on the GCN framework for representation learning by exploring the consistency and mutually exclusive information between the learned ELU graph and the original graph. Moreover, we theoretically demonstrate that the proposed method can ensure the generalization ability of GCNs. Extensive experiments validate the superiority of the proposed method.", "keywords": [ "graph neural network", "semi-supervised learning", "node classification" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 5, 4, 2 ] }, { "paper_id": "todLTYB1I7", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Ge Yan", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Tsui-Wei Weng", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Tuomas Oikarinen", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 81, "n_formula_1": 23, "n_ref_uni": 21, "n_ref": 56, "n_ref_all": 89, "n_fig": 2, "n_tab": 20, "L_tab": 10219, "n_element_tab": 1321, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 879, "n_element_tab_1": 121, "formula_len_all": 5655, "formula_len_all_1": 1320, "len_all": 198559, "len_all_1": 56121, "len_abs": 930, "len_title": 105, "len_sents": 55331, "len_sents_1": 24159, "n_sents": 460, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 934, "title": "A Principled Evaluation Framework for Neuron Explanations", "abs": "Understanding the function of individual units in a neural network is an important building block for mechanistic interpretability. This is often done by generating a simple text explanation of the behavior of individual neurons or units. However, for these explanations to be useful, we must understand how reliable and truthful they are. In this work we unify many existing explanation evaluation methods under one mathematical framework. This allows us to compare and contrast existing evaluation metrics and understand the evaluation pipeline with increased clarity. We propose two simple sanity checks on the evaluation metrics and show that many commonly used metrics fail these tests and do not change their score after massive changes to the concept labels. Based on our experimental and theoretical results, we propose guidelines that future evaluations should follow and identify good evaluation metrics such as correlation.", "keywords": [ "interpretability", "mechanistic interpretability", "explainable AI", "trustworthy machine learning" ], "rating_list": [ 5, 6, 6, 3, 5 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 2, 4, 4, 5, 3 ] }, { "paper_id": "to4PdiiILF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Christoph Str\u00e4ter", "gender": "Male", "institution": "LASR Labs", "country": "GB", "position": "Independent Researcher" }, { "name": "Joe Needham", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "MS student" }, { "name": "Leo McKee-Reid", "gender": "Male", "institution": "University of Victoria", "country": "CA", "position": "Researcher" }, { "name": "Maria Angelica Martinez", "gender": "Female", "institution": "University of the Philippines", "country": "PH", "position": "Researcher" }, { "name": "Mikita Balesni", "gender": "Male", "institution": "Apollo Research", "country": "GB", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 17, "n_ref": 56, "n_ref_all": 89, "n_fig": 9, "n_tab": 4, "L_tab": 1287, "n_element_tab": 101, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 117650, "len_all_1": 52910, "len_abs": 1215, "len_title": 140, "len_sents": 37678, "len_sents_1": 26454, "n_sents": 271, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1221, "title": "Honesty to Subterfuge: In-Context Reinforcement Learning Can Make Honest Models Reward Hack", "abs": "Previous work has shown that training \u201chelpful-only\u201d LLMs with reinforcement learning on a curriculum of gameable environments can lead models to generalize to egregious specification gaming, such as editing their own reward function or modifying task checklists to appear more successful. We show that gpt-4o, gpt-4o-mini, o1-preview, and o1-mini \u2014 frontier models trained to be helpful, harmless, and honest \u2014 can engage in specification gaming without training on a curriculum of tasks, purely from in-context iterative reflection (which we call in-context reinforcement learning, \u201cICRL\u201d). We also show that using ICRL to generate highly-rewarded outputs for expert iteration (compared to the standard expert iteration reinforcement learning algorithm) may increase gpt-4o-mini's propensity to learn specification-gaming policies, generalizing (in very rare cases) to the most egregious strategy where gpt-4o-mini edits its own reward function. Our results point toward the strong ability of in-context reflection to discover rare specification-gaming strategies that models might not exhibit zero-shot or with normal training, highlighting the need for caution when relying on alignment of LLMs in zero-shot settings.", "keywords": [ "Large Language Model", "Deception", "specification gaming", "Reward Hacking", "Evaluations", "in-context reinforcement learning", "in-context learning", "iterative reflection", "gpt-4o-mini", "gpt-4o", "o1-mini", "o1-preview" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "tnSj6FdN8w", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Hong Zhang", "gender": "Male", "institution": "Argonne National Laboratory", "country": "", "position": "Assistant Computational Mathematician" }, { "name": "Yifan Zhang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 21, "n_ref": 26, "n_ref_all": 43, "n_fig": 6, "n_tab": 4, "L_tab": 295, "n_element_tab": 29, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 295, "n_element_tab_1": 29, "formula_len_all": 614, "formula_len_all_1": 614, "len_all": 67562, "len_all_1": 66892, "len_abs": 957, "len_title": 93, "len_sents": 20957, "len_sents_1": 20768, "n_sents": 185, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 960, "title": "Neural Time Integrator with Stage Correction", "abs": "Numerical simulation of dynamical systems requires time integration solvers that\nbalance accuracy and computational efficiency. Recent work indicates that neural\nintegrators, a hybrid of classical numerical integration and machine learning, can\nachieve significant performance gains. Building upon this idea, we propose a new\ntype of neural integrator that introduces stage corrections inspired by the fact that\ntraditional time integration schemes such as Runge-Kutta exhibit different error\ncharacteristics at each stage. Specifically, our method corrects numerical errors\nimmediately after each stage evaluation by using a neural network, mitigating\nerror propagation across stages. This enables the use of larger time steps while\npreserving stability and accuracy. We demonstrate that our approach is at least\none order of magnitude more accurate than existing hybrid methods for complex\nnonlinear dynamical systems when integrated with the same step size.", "keywords": [ "dynamical system", "hybrid ML", "error correction", "time integrator" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 2, 3, 1, 1 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "tnB94WQGrn", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Marinka Zitnik", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" }, { "name": "Shanghua Gao", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Postdoc" }, { "name": "Valentina Giunchiglia", "gender": "unknown", "institution": "Imperial College London", "country": "", "position": "Researcher" }, { "name": "Xiaolong Liu", "gender": "Male", "institution": "University of Illinois at Chicago", "country": "US", "position": "PhD student" }, { "name": "Xiaorui Su", "gender": "Female", "institution": "Harvard Medical School, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Yibo Wang", "gender": "Female", "institution": "University of Illinois at Chicago", "country": "US", "position": "PhD student" }, { "name": "Djork-Arne Clevert", "gender": "Male", "institution": "Pfizer", "country": "", "position": "Vice President" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 54, "n_ref": 84, "n_ref_all": 121, "n_fig": 14, "n_tab": 7, "L_tab": 9047, "n_element_tab": 428, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 4035, "n_element_tab_1": 270, "formula_len_all": 158, "formula_len_all_1": 138, "len_all": 214542, "len_all_1": 70599, "len_abs": 1420, "len_title": 123, "len_sents": 54331, "len_sents_1": 28429, "n_sents": 431, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 75, "L_abs": 1428, "title": "Knowledge Graph Based Agent For Complex, Knowledge-Intensive QA in Medicine", "abs": "Biomedical knowledge is uniquely complex and structured, requiring distinct reasoning strategies compared to other scientific disciplines like physics or chemistry. Biomedical scientists do not rely on a single approach to reasoning; instead, they use various strategies, including rule-based, prototype-based, and case-based reasoning. This diversity calls for flexible approaches that accommodate multiple reasoning strategies while leveraging in-domain knowledge. We introduce KGARevion, a knowledge graph (KG) based agent designed to address the complexity of knowledge-intensive medical queries. Upon receiving a query, KGARevion generates relevant triplets by using the knowledge base of the LLM. These triplets are then verified against a grounded KG to filter out erroneous information and ensure that only accurate, relevant data contribute to the final answer. Unlike RAG-based models, this multi-step process ensures robustness in reasoning while adapting to different models of medical reasoning. Evaluations on four gold-standard medical QA datasets show that KGARevion improves accuracy by over 5.2%, outperforming 15 models in handling complex medical questions. To test its capabilities, we curated three new medical QA datasets with varying levels of semantic complexity, where KGARevion achieved a 10.4% improvement in accuracy. The source code is provided at https://anonymous.4open.science/r/KGARevion-B3B6.", "keywords": [ "Medical Reasoning; Medical QA", "Agent", "Knowledge Graph", "LLM" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 4, 2, 4 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "tn2mjzjSyR", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Dian Yu", "gender": "Female", "institution": "Tencent AI Lab", "country": "", "position": "NLP researcher" }, { "name": "Dong Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Distinguished Scientist" }, { "name": "Haitao Mi", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Principal Researcher" }, { "name": "Murong Yue", "gender": "Male", "institution": "George Mason University", "country": "US", "position": "PhD student" }, { "name": "Wenlin Yao", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Ziyu Yao", "gender": "Female", "institution": "George Mason University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 5, "n_ref_uni": 37, "n_ref": 63, "n_ref_all": 94, "n_fig": 3, "n_tab": 10, "L_tab": 5928, "n_element_tab": 564, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 2839, "n_element_tab_1": 293, "formula_len_all": 531, "formula_len_all_1": 395, "len_all": 165300, "len_all_1": 62034, "len_abs": 1814, "len_title": 86, "len_sents": 47559, "len_sents_1": 26284, "n_sents": 327, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1820, "title": "DOTS: Learning to Reason Dynamically in LLMs via Optimal Reasoning Trajectories Search", "abs": "Enhancing the capability of large language models (LLMs) in reasoning has gained significant attention in recent years. Previous studies have demonstrated the effectiveness of various prompting strategies in aiding LLMs in reasoning (called \"reasoning actions\"), such as step-by-step thinking, reflecting before answering, solving with programs, and their combinations. However, these approaches often applied static, predefined reasoning actions uniformly to all questions, without considering the specific characteristics of each question or the capability of the task-solving LLM. In this paper, we propose DOTS, an approach enabling LLMs to reason Dynamically via Optimal reasoning Trajectories Search, tailored to the specific characteristics of each question and the inherent capability of the task-solving LLM. \nOur approach involves three key steps: i) defining atomic reasoning action modules that can be composed into various reasoning action trajectories; ii) searching for the optimal action trajectory for each training question through iterative exploration and evaluation for the specific task-solving LLM; and iii) using the collected optimal trajectories to train an LLM to plan for the reasoning trajectories of unseen questions. In particular, we propose two learning paradigms, i.e., fine-tuning an external LLM as a planner to guide the task-solving LLM, or directly fine-tuning the task-solving LLM with an internalized capability for reasoning actions planning. Our experiments across eight reasoning tasks show that our method consistently outperforms static reasoning techniques and the vanilla instruction tuning approach. Further analysis reveals that our method enables LLMs to adjust their computation based on problem complexity, allocating deeper thinking and reasoning to harder problems.", "keywords": [ "large language model", "reasoning" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "tmwR707odU", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Haibo Chen", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Haoyang Li", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Postdoc" }, { "name": "Ling Feng", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Weigao Wen", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "MS student" }, { "name": "Wenwu Zhu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xin Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Zeyang Zhang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 14, "n_ref_uni": 54, "n_ref": 113, "n_ref_all": 133, "n_fig": 9, "n_tab": 2, "L_tab": 2538, "n_element_tab": 114, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1966, "n_element_tab_1": 107, "formula_len_all": 2383, "formula_len_all_1": 893, "len_all": 205221, "len_all_1": 63863, "len_abs": 1870, "len_title": 103, "len_sents": 52325, "len_sents_1": 28551, "n_sents": 402, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1920, "title": "Curriculum GNN-LLM Alignment for Text-Attributed Graphs", "abs": "Aligning Graph Neural Networks (GNNs) and Large Language Models (LLMs) benefits in leveraging both textual and structural knowledge for Text-attributed Graphs (TAGs) learning, which has attracted an increasing amount of attention in the research community. Most existing literature assumes a uniformly identical level of learning difficulties across texts and structures in TAGs, however, we discover the $\\textit{text-structure imbalance}$ problem in real-world TAGs, $\\textit{i.e.}$, nodes exhibit various levels of difficulties when learning different textual and structural information. Existing works ignoring these different difficulties may result in under-optimized GNNs and LLMs with over-reliance on either simplistic text or structure, thus failing to conduct node classifications that involve simultaneously learning complex text and structural information for nodes in TAGs. To address this problem, we propose a novel Curriculum GNN-LLM Alignment ($\\textbf{CurGL}$) method, which strategically balances the learning difficulties of textual and structural information on a node-by-node basis to enhance the alignment between GNNs and LLMs. Specifically, we first propose a text-structure difficulty measurer to estimate the learning difficulty of both text and structure in a node-wise manner. Then, we propose a class-based node selection strategy to balance the training process via gradually scheduling more nodes. Finally, we propose the curriculum co-play alignment by iteratively promoting useful information from GNNs and LLMs, to progressively enhance both components with balanced textual and structural information. Extensive experiments on real-world datasets demonstrate that our proposed $\\textbf{CurGL}$ method is able to outperform state-of-the-art GraphLLM, curriculum learning, as well as GNN baselines. To the best of our knowledge, this is the first study of curriculum alignment on TAGs.", "keywords": [ "Graph neural networks", "Large language models", "Text-attributed graphs", "Curriculum learning" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 1, 3, 1, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 5 ] }, { "paper_id": "tmSWFGpBb8", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Bujiao Wu", "gender": "Female", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Researcher" }, { "name": "Jingbo Wang", "gender": "unknown", "institution": "University of Western Australia", "country": "", "position": "Full Professor" }, { "name": "Xiao Yuan", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yanqi Song", "gender": "Female", "institution": "China Academy of Information and Communications Technology", "country": "CN", "position": "Researcher" }, { "name": "Yusen Wu", "gender": "Male", "institution": "Beijing Normal University", "country": "CN", "position": "Postdoc" } ], "n_formula": 121, "n_formula_1": 32, "n_ref_uni": 60, "n_ref": 117, "n_ref_all": 134, "n_fig": 6, "n_tab": 2, "L_tab": 504, "n_element_tab": 45, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 11961, "formula_len_all_1": 2915, "len_all": 220341, "len_all_1": 60475, "len_abs": 1032, "len_title": 102, "len_sents": 60694, "len_sents_1": 26030, "n_sents": 467, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1053, "title": "Learning the Complexity of Weakly Noisy Quantum States", "abs": "Quantifying the complexity of quantum states is a longstanding key problem in various subfields of science, ranging from quantum computing to the black-hole theory. The lower bound on quantum pure state complexity has been shown to grow linearly with system size [J. Haferkamp et al., 2022, *Nat. Phys.*]. However, extending this result to noisy circuit environments, which better reflect real quantum devices, remains an open challenge. In this paper, we explore the complexity of weakly noisy quantum states via the quantum learning method. We present an efficient learning algorithm, that leverages the classical shadow representation of target quantum states, to predict the circuit complexity of weakly noisy quantum states. Our algorithm is proved to be optimal in terms of sample complexity accompanied with polynomial classical processing time. Our result builds a bridge between the learning algorithm and quantum state complexity, meanwhile highlighting the power of learning algorithm in characterizing intrinsic properties of quantum states.", "keywords": [ "Quantum State Complexity", "Quantum Learning Algorithm" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 2, 3, 4 ] }, { "paper_id": "tlH4vDii0E", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Jialin Yu", "gender": "unknown", "institution": "University College London, University of London", "country": "GB", "position": "Postdoc" }, { "name": "Ricardo Silva", "gender": "Male", "institution": "University College London", "country": "", "position": "Full Professor" }, { "name": "Yulan He", "gender": "Female", "institution": "King's College London, University of London", "country": "GB", "position": "Full Professor" }, { "name": "Yuxiang Zhou", "gender": "Male", "institution": "King's College London", "country": "GB", "position": "Postdoc" }, { "name": "Nevin Zhang", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Hong Kong University of Science and Technology", "country": "", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 16, "n_ref_uni": 52, "n_ref": 89, "n_ref_all": 118, "n_fig": 11, "n_tab": 5, "L_tab": 948, "n_element_tab": 151, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 615, "formula_len_all_1": 801, "len_all": 153414, "len_all_1": 67244, "len_abs": 1335, "len_title": 129, "len_sents": 44346, "len_sents_1": 31519, "n_sents": 341, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1360, "title": "Fine-Tuning Pre-trained Language Models for Robust Causal Representation Learning", "abs": "The fine-tuning of pre-trained language models (PLMs) has been shown to be effective across various domains. By using domain-specific supervised data, the general-purpose representation derived from PLMs can be transformed into a domain-specific representation. However, these methods often fail to generalize to out-of-domain (OOD) data due to their reliance on $\\textit{non-causal}$ representations, often described as spurious features. Existing methods either make use of adjustments with strong assumptions about lack of hidden common causes, or mitigate the effect of spurious features using multi-domain data. In this work, we investigate how fine-tuned pre-trained language models aid generalizability from single-domain scenarios under mild assumptions, targeting more general and practical real-world scenarios. We show that a robust representation can be derived through a so-called causal front-door adjustment, based on a $\\textit{decomposition}$ assumption, using fine-tuned representations as a source of data augmentation. Comprehensive experiments in both synthetic and real-world settings demonstrate the superior generalizability of the proposed method compared to existing approaches. Our work thus sheds light on the domain generalization problem by introducing links between fine-tuning and causal mechanisms into representation learning.", "keywords": [ "representation learning", "causal inference", "robust machine learning" ], "rating_list": [ 6, 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 1, 3, 2 ], "presentation_list": [ 1, 3, 1, 3, 2 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 3, 2, 2, 3, 3 ] }, { "paper_id": "tl63stKeSC", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Kamyar Azizzadenesheli", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Sourav Pal", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "PhD student" }, { "name": "Vikas Singh", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Professor" } ], "n_formula": 63, "n_formula_1": 25, "n_ref_uni": 32, "n_ref": 58, "n_ref_all": 103, "n_fig": 16, "n_tab": 7, "L_tab": 1552, "n_element_tab": 108, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 515, "n_element_tab_1": 28, "formula_len_all": 4713, "formula_len_all_1": 1423, "len_all": 157287, "len_all_1": 70815, "len_abs": 1255, "len_title": 85, "len_sents": 54086, "len_sents_1": 32729, "n_sents": 470, "n_sents_1": 279, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 1260, "title": "Solving PDEs via learnable quadrature", "abs": "Partial differential Equations (PDEs) are an essential tool across science and engineering. Recent work has shown how contemporary developments in machine learning models can directly help in improving methods for solution discovery of PDEs. This line of work falls under the umbrella of Physics-Informed Machine Learning. A key step in solving a PDE is to determine a set of points in the domain where the current iterate of the PDE's solution will be evaluated. The most prevalent strategy here is to use Monte Carlo sampling, but it is widely known to be sub-optimal in lower dimensions. We leverage recent advances in asymptotic expansions of quadrature nodes and weights (for weight functions belonging to the modified Gauss-Jacobi family) together with suitable adjustments for parameterization towards a data-driven framework for learnable quadrature rules. A direct benefit is a performance improvement in solving PDEs via neural networks, relative to existing alternatives, on a set of problems commonly studied in the literature. Beyond finding a standard solution for an instance of a single PDE, our construction enables learning rules to predict solutions for a given family of PDEs via a simple use of hyper-networks, a broadly useful capability.", "keywords": [ "neural network", "quadrature rule", "PINN" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "tl0jpyVXgk", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Dong In Lee", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" }, { "name": "Giljoo Nam", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Hwanhee Jung", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" }, { "name": "Hyeongcheol Park", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" }, { "name": "Jungseock Joo", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Engineer" }, { "name": "Sang Ho Yoon", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Sangpil Kim", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Assistant Professor" }, { "name": "WONSEOK ROH", "gender": "unknown", "institution": "Korea University", "country": "KR", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 15, "n_ref_uni": 62, "n_ref": 193, "n_ref_all": 220, "n_fig": 17, "n_tab": 11, "L_tab": 3254, "n_element_tab": 522, "n_fig_1": 9, "n_tab_1": 9, "L_tab_1": 2793, "n_element_tab_1": 517, "formula_len_all": 1607, "formula_len_all_1": 1197, "len_all": 187868, "len_all_1": 76570, "len_abs": 1511, "len_title": 112, "len_sents": 51826, "len_sents_1": 30935, "n_sents": 386, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1520, "title": "Instance-wise Knowledge Enhancement for 3D Instance Segmentation", "abs": "Recent 3D Instance Segmentation methods typically follow a similar paradigm; they encode hundreds of instance-wise candidates with instance-specific information in various ways and refine them into final masks. However, they have yet to fully explore the benefit of these candidates. They overlook the valuable cues encoded in multiple candidates that represent different parts of the same instance, resulting in fragmented instance masks. Also, they often fail to capture the precise spatial range of complex 3D instances, primarily due to inherent fuzzy noises from sparse and unordered point clouds. In this work, to address these challenges, we propose IKEA, a novel instance-wise knowledge enhancement approach. We first introduce an Instance-wise Knowledge Aggregation to associate scattered single instance details by optimizing correlations among candidates representing the same instance. Moreover, we present Instance-wise Structural Guidance to enhance the spatial understanding of candidates using structural cues from ambiguity-reduced features. Here, we utilize a simple yet effective truncated singular value decomposition algorithm to minimize inherent noises of 3D features. Finally, our instance-wise features are now highly informative for real-world 3D instances. In our extensive experiments on large-scale benchmarks, ScanNetV2, ScanNet200, S3DIS, and STPLS3D, IKEA outperforms existing works. We also demonstrate the effectiveness of our modules based on both kernel and transformer architectures.", "keywords": [ "Computer Vision", "3D Instance Segmentation" ], "rating_list": [ 5, 6, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 5, 2 ] }, { "paper_id": "tkqNDbukWW", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Ahmed Abdulaal", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Amrutha Saseendran", "gender": "Female", "institution": "AstraZeneca", "country": "GB", "position": "Research Scientist" }, { "name": "Aryo Pradipta Gema", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Beatrice Alex", "gender": "Female", "institution": "University of Edinburgh, University of Edinburgh", "country": "", "position": "Assistant Professor" }, { "name": "Chen Jin", "gender": "unknown", "institution": "Astrazeneca", "country": "GB", "position": "Researcher" }, { "name": "Philip Alexander Teare", "gender": "Male", "institution": "AstraZeneca", "country": "", "position": "Researcher" }, { "name": "Tom Diethe", "gender": "Male", "institution": "AstraZeneca", "country": "", "position": "Principal Researcher" }, { "name": "Pasquale Minervini", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 50, "n_ref": 108, "n_ref_all": 181, "n_fig": 18, "n_tab": 30, "L_tab": 19164, "n_element_tab": 2471, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 4929, "n_element_tab_1": 675, "formula_len_all": 597, "formula_len_all_1": 1005, "len_all": 307010, "len_all_1": 71740, "len_abs": 1247, "len_title": 122, "len_sents": 75352, "len_sents_1": 28858, "n_sents": 518, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1166, "title": "DeCoRe: Decoding by Contrasting Retrieval Heads to Mitigate Hallucinations", "abs": "Large Language Models (LLMs) often hallucinate, producing unfaithful or factually incorrect outputs by misrepresenting the provided context or incorrectly recalling internal knowledge. Recent studies have identified specific attention heads within the Transformer architecture, known as retrieval heads, responsible for extracting relevant contextual information. We hypothesise that masking these retrieval heads can induce hallucinations and that contrasting the outputs of the base LLM and the masked LLM can reduce hallucinations. To this end, we propose Decoding by Contrasting Retrieval Heads (DeCoRe), a novel training-free decoding strategy that amplifies information found in the context and model parameters. DeCoRe mitigates potentially hallucinated responses by dynamically contrasting the outputs of the base LLM and the masked LLM, using conditional entropy as a guide. Our extensive experiments confirm that DeCoRe significantly improves performance on tasks requiring high contextual faithfulness, such as summarisation (XSum by 18.6\\%), instruction following (MemoTrap by 10.9\\%), and open-book question answering (NQ by 2.4\\% and NQ-Swap by 5.5\\%).", "keywords": [ "Large Language Models", "Hallucination", "Constrained Decoding", "Retrieval Heads" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "tkiZQlL04w", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Danning Ke", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Gongyi Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Principal Researcher" }, { "name": "Hanlin Tang", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Jing Lin", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Qingsen Han", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Shikuan Hong", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Yang Lin", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Yiwu Yao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 7, "n_ref_uni": 28, "n_ref": 62, "n_ref_all": 82, "n_fig": 7, "n_tab": 2, "L_tab": 199, "n_element_tab": 18, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 199, "n_element_tab_1": 18, "formula_len_all": 946, "formula_len_all_1": 515, "len_all": 140777, "len_all_1": 56111, "len_abs": 1440, "len_title": 119, "len_sents": 29473, "len_sents_1": 27592, "n_sents": 212, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1445, "title": "RazorAttention: Efficient KV Cache Compression Through Retrieval Heads", "abs": "The memory and computational demands of Key-Value (KV) cache present significant challenges for deploying long-context language models. Previous approaches attempt to mitigate this issue by selectively dropping tokens, which irreversibly erases critical information that might be needed for future queries. In this paper, we propose a novel compression technique for KV cache that preserves all token information. Our investigation reveals that: i) Most attention heads primarily focus on the local context; ii) Only a few heads, denoted as retrieval heads, can essentially pay attention to all input tokens. These key observations motivate us to use separate caching strategy for attention heads.Therefore, we propose RazorAttention, a training-free KV cache compression algorithm, which maintains a full cache for these crucial retrieval heads and discards the remote tokens in non-retrieval heads. Furthermore, we introduce a novel mechanism involving a \u201ccompensation token\u201d to further recover the information in the dropped tokens. Extensive evaluations across a diverse set of large language models (LLMs) demonstrate that RazorAttention achieves a reduction in KV cache size by over 70% without noticeable impacts on performance. Additionally, RazorAttention is compatible with FlashAttention, rendering it an efficient and plug-and-play solution that enhances LLM inference efficiency without overhead or retraining of the original model.", "keywords": [ "LLMs", "KV cache compression", "LLM inference acceleration" ], "rating_list": [ 5, 5, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 5 ] }, { "paper_id": "tkg9XMFo0H", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Chang Tang", "gender": "Male", "institution": "China University of Geosciences", "country": "CN", "position": "Full Professor" }, { "name": "Junkai Chen", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "Undergrad student" }, { "name": "Kening Zheng", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "MS student" }, { "name": "Sirui Huang", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Xin Zou", "gender": "Male", "institution": "CUG", "country": "CN", "position": "MS student" }, { "name": "Xuming Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Yibo Yan", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yizhou Wang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 35, "n_formula_1": 9, "n_ref_uni": 46, "n_ref": 83, "n_ref_all": 104, "n_fig": 7, "n_tab": 7, "L_tab": 5191, "n_element_tab": 490, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3625, "n_element_tab_1": 328, "formula_len_all": 2536, "formula_len_all_1": 540, "len_all": 201451, "len_all_1": 69365, "len_abs": 1946, "len_title": 201, "len_sents": 55766, "len_sents_1": 28979, "n_sents": 498, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 124, "L_abs": 1127, "title": "Look Twice Before You Answer: Memory-Space Visual Retracing for Hallucination Mitigation in Multimodal Large Language Models", "abs": "Despite their impressive capabilities, Multimodal Large Language Models (MLLMs) are susceptible to hallucinations, especially assertively fabricating content not present in the visual inputs. To address the aforementioned challenge, we follow a common cognitive process - \\textit{when one's initial memory of critical on-sight details fades, replenishing visual memory is essential to seek a factual and accurate answer.} Therefore, we introduce Memory-space Visual Retracing (MemVR), a novel hallucination mitigation paradigm that without the need for external knowledge retrieval or additional fine-tuning. In particular, we treat visual tokens as supplementary evidence to be reinjected into MLLMs via Feed Forward Network (FFN) as \u201ckey-value memory\u201d at the middle trigger layer, \\textit{i.e.}, when the model is uncertain about visual memories in the layer. Comprehensive experimental evaluations demonstrate that \\modelname significantly mitigates hallucination issues across various MLLMs and excels in general benchmarks without incurring added time overhead, thus emphasizing its potential for widespread applicability.", "keywords": [ "Hallucination mitigation", "MLLMs", "visual retracing", "training-free" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 1, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "tkN0sLhb4P", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Cecilia Mascolo", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "Full Professor" }, { "name": "Ting Dang", "gender": "Female", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Yu Wu", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 22, "n_ref": 61, "n_ref_all": 71, "n_fig": 3, "n_tab": 6, "L_tab": 2036, "n_element_tab": 74, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1144, "n_element_tab_1": 34, "formula_len_all": 763, "formula_len_all_1": 676, "len_all": 115244, "len_all_1": 65946, "len_abs": 1817, "len_title": 142, "len_sents": 38165, "len_sents_1": 32636, "n_sents": 250, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1823, "title": "GITAR: GENERALIZED IRREGULAR TIME SERIES REGRESSION VIA MASKING AND RECONSTRUCTION PRETRAINING", "abs": "Multivariate time series regression, encompassing forecasting and interpolation, is crucial for numerous real-world applications, particularly in healthcare, climate science, ecology, and others. While recent work has focused on improving modeling for time series regression, two main limitations persist. First, the prevalence of irregularly sampled time series with missing values poses significant challenges.\nFor instance, healthcare applications often involve predicting future or missing observations from irregular data to enable continuous patient monitoring and timely intervention. As current approaches mainly rely on the assumptions of regular time series such as strong periodicity, when applied to irregular ones they exhibit performance degradation. Second, while some state-of-the-art methods (SOTA) do model irregularity and perform regression tasks on irregular data, they are often trained in a fully supervised manner. This limits their ability to generalize easily to different domains (e.g., training and testing datasets with different numbers of variables). To address these challenges, we propose GITaR, a Generalized Irregular Time Series Regression model via masking and Reconstruction pertaining mechanism, aiming to capture the inherent irregularity in time series and learn robust, generalizable representations without supervision for downstream regression tasks. Comprehensive experiments on common real-world regression tasks in healthcare, human activity recognition, and climate science underline the superior performance of GITaR compared to state-of-the-art methods. Our results highlight our model\u2019s unique capability to generalize across different domains, demonstrating the potential for broad applicability in various fields requiring accurate temporal prediction and interpolation.", "keywords": [ "Time series", "irregular time series", "self-supervised learning" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "tkG7jkrkxy", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Binghui Chen", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Chen duosheng", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "MS student" }, { "name": "Liefeng Bo", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Principal Researcher" }, { "name": "Yifeng Geng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 18, "n_ref": 31, "n_ref_all": 69, "n_fig": 19, "n_tab": 3, "L_tab": 397, "n_element_tab": 43, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 200, "n_element_tab_1": 20, "formula_len_all": 713, "formula_len_all_1": 570, "len_all": 124897, "len_all_1": 55726, "len_abs": 1718, "len_title": 119, "len_sents": 36951, "len_sents_1": 25883, "n_sents": 279, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1717, "title": "Adaptive Drag: Semantic-Driven Dragging on Diffusion-Based Image Editing", "abs": "Recently, several point-based image editing methods ($\\textit{e.g.}$, DragDiffusion, FreeDrag, DragNoise) have emerged, yielding precise and high-quality results based on user instructions. However, these methods often make insufficient use of semantic information, leading to less desirable results. In this paper, we proposed a novel mask-free point-based image editing method, $\\textbf{AdaptiveDrag}$, which provides a more flexible editing approach and generates images that better align with user intent. Specifically, we design an auto mask generation module using super-pixel division for user-friendliness. Next, we leverage a pre-trained diffusion model to optimize the latent, enabling the dragging of features from handle points to target points. To ensure a comprehensive connection between the input image and the drag process, we have developed a semantic-driven optimization. We design adaptive steps that are supervised by the positions of the points and the semantic regions derived from super-pixel segmentation. This refined optimization process also leads to more realistic and accurate drag results. Furthermore, to address the limitations in the generative consistency of the diffusion model, we introduce an innovative corresponding loss during the sampling process. Building on these effective designs, our method delivers superior generation results using only the single input image and the handle-target point pairs. Extensive experiments have been conducted and demonstrate that the proposed method outperforms others in handling various drag instructions ($\\textit{e.g.}$, resize, movement, extension) across different domains ($\\textit{e.g.}$, animals, human face, land space, clothing).", "keywords": [ "image edit", "point-based drag", "diffusion", "semantic-driven" ], "rating_list": [ 5, 3, 6, 6, 5 ], "soundness_list": [ 3, 2, 4, 2, 2 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "tjlTczcnPz", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Haotong Qin", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Luca Benini", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Full Professor" }, { "name": "Michele Magno", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "Principal Researcher" }, { "name": "Qinshuo Liu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shiming Zhang", "gender": "unknown", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Wei Huang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "XIAOJUAN QI", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Xianglong Liu", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Yangdong Liu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Yawei Li", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Lecturer" } ], "n_formula": 19, "n_formula_1": 11, "n_ref_uni": 42, "n_ref": 150, "n_ref_all": 181, "n_fig": 9, "n_tab": 7, "L_tab": 2725, "n_element_tab": 305, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1360, "n_element_tab_1": 121, "formula_len_all": 1880, "formula_len_all_1": 782, "len_all": 191250, "len_all_1": 63859, "len_abs": 2054, "len_title": 128, "len_sents": 52363, "len_sents_1": 27403, "n_sents": 368, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 2061, "title": "SliM-LLM: Salience-Driven Mixed-Precision Quantization for Large Language Models", "abs": "Large language models (LLMs) have achieved remarkable progress, but their extensive number of parameters results in high memory usage, significant loading latency, and substantial computational demands. To address these challenges, post-training quantization (PTQ) has emerged as an effective technique for compressing model weights. In the context of PTQ for LLMs, existing uniform quantization methods, though efficient in terms of memory and computational requirements, often struggle to maintain performance. In this paper, we propose SliM-LLM, a Salience-Driven Mixed-Precision Quantization scheme that achieves group-wise bit-width allocation with mixed precisions for efficient LLMs with high accuracy. Building on our observation that salient/important weights often follow a structured distribution, we incorporate two core components to preserve post-quantization performance in LLMs while maintaining efficiency: 1) Salience-Determined Bit Allocation adaptively assigns bit widths to groups within each layer based on their group-level salience, aiming to minimize the reconstruction error of activations; and 2) Salience-Weighted Quantizer Calibration optimizes quantizer parameters by incorporating element-level salience, ensuring that the most critical weights are preserved, further preserving important weights information. With its structured group partitioning, SliM-LLM offers a hardware-friendly quantization approach, maintaining computational and memory efficiency comparable to highly optimized uniform quantization methods. Extensive experiments demonstrate that SliM-LLM significantly improves the accuracy of various LLMs when quantized to ultra-low bit widths. For instance, a 2-bit quantized LLaMA-7B model achieves nearly 6x memory reduction compared to its floating-point counterpart, alongside a 48% reduction in perplexity compared to the leading gradient-free PTQ method, all while maintaining GPU inference speed. Furthermore, SliM-LLM+, which incorporates gradient-based quantizers, reduces perplexity by an additional 35.1%.", "keywords": [ "Large Language Model", "Low-bit Quantization", "Inference", "Machine Learning" ], "rating_list": [ 5, 5, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "tjNf0L8QjR", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Duy Kien Nguyen", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Martin R. Oswald", "gender": "unknown", "institution": "University of Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Unnat Jain", "gender": "Male", "institution": "Skild AI", "country": "US", "position": "Researcher" }, { "name": "Xinlei Chen", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Cees Snoek", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "Mahmoud Assran", "gender": "Male", "institution": "Meta", "country": "CA", "position": "Research Scientist" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 58, "n_ref": 126, "n_ref_all": 150, "n_fig": 17, "n_tab": 7, "L_tab": 1072, "n_element_tab": 106, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 632, "n_element_tab_1": 105, "formula_len_all": 233, "formula_len_all_1": 135, "len_all": 171608, "len_all_1": 61443, "len_abs": 1045, "len_title": 134, "len_sents": 52373, "len_sents_1": 29044, "n_sents": 421, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1049, "title": "An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels", "abs": "This work does not introduce a new method. Instead, we present an interesting finding that questions the necessity of the inductive bias of locality in modern computer vision architectures. Concretely, we find that vanilla Transformers can operate by directly treating each individual pixel as a token and achieve highly performant results. This is substantially different from the popular design in Vision Transformer, which maintains the inductive bias from ConvNets towards local neighborhoods (e.g., by treating each 16x16 patch as a token). We showcase the effectiveness of pixels-as-tokens across three well-studied computer vision tasks: supervised learning for classification and regression, self-supervised learning via masked autoencoding, and image generation with diffusion models. Although it's computationally less practical to directly operate on individual pixels, we believe the community must be made aware of this surprising piece of knowledge when devising the next generation of neural network architectures for computer vision.", "keywords": [ "locality", "convolutional networks", "transformers" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "tj5xJInWty", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Dongqi Fu", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Hanghang Tong", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Jingrui He", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Ross Maciejewski", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "Full Professor" }, { "name": "Xinyu He", "gender": "unknown", "institution": "Department of Computer Science", "country": "US", "position": "PhD student" } ], "n_formula": 50, "n_formula_1": 30, "n_ref_uni": 36, "n_ref": 53, "n_ref_all": 68, "n_fig": 3, "n_tab": 10, "L_tab": 2971, "n_element_tab": 179, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 1692, "n_element_tab_1": 60, "formula_len_all": 4659, "formula_len_all_1": 3058, "len_all": 155825, "len_all_1": 75163, "len_abs": 1522, "len_title": 125, "len_sents": 47124, "len_sents_1": 31432, "n_sents": 379, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 77, "L_abs": 1549, "title": "Temporal Heterogeneous Graph Generation with Privacy, Utility, and Efficiency", "abs": "Nowadays, Temporal Heterogeneous Graphs attract much research and industrial attention for building the next-generation Relational Deep Learning Models and Applications, due to their informative structures and features. While providing timely and precise services like personalized recommendations and question answering, this rich information also introduces extra exposure risk for each node in the graph, because the distinctive local topology, the abundant heterogeneous features, and the time dimension of the graph data are more prone to exposing sensitive information and narrow down the scope of victim candidates, which calls for well-defined protection techniques on graphs. To this end, we propose a **T**emporal **He**terogeneous Graph Generator balancing **P**rivacy, **U**tility, and E**ff**iciency, named **THePUff**. More specifically, we first propose a differential privacy algorithm to perturb the input temporal heterogeneous graph for protecting privacy, and then utilize both the perturbed graph and the original one in a generative adversarial setting for THePUff to learn and generate privacy-guaranteed and utility-preserved graph data in an efficient manner. We further propose 6 new metrics in the temporal setting to measure heterogeneous graph utility and privacy. Finally, based on temporal heterogeneous graph datasets with up to 1 million nodes and 20 million edges, the experiments show that THePUff generates utilizable temporal heterogeneous graphs with privacy protected, compared with state-of-the-art baselines.", "keywords": [ "Temporal Graph", "Heterogeneous Graph", "Graph Generation" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "tj40W2HAKN", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 8, "author_info_list": [ { "name": "chen luo", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Hanqing Lu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Haoyu Han", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Hui Liu", "gender": "Female", "institution": "Michigan State University", "country": "US", "position": "Assistant Professor" }, { "name": "Jiliang Tang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Full Professor" }, { "name": "Juanhui Li", "gender": "Female", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Wei Huang", "gender": "Male", "institution": "RIKEN AIP", "country": "JP", "position": "Research Scientist" }, { "name": "Xianfeng Tang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 38, "n_formula_1": 14, "n_ref_uni": 53, "n_ref": 133, "n_ref_all": 157, "n_fig": 15, "n_tab": 7, "L_tab": 2501, "n_element_tab": 217, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2474, "formula_len_all_1": 684, "len_all": 183145, "len_all_1": 70362, "len_abs": 1061, "len_title": 124, "len_sents": 53642, "len_sents_1": 33215, "n_sents": 448, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 867, "title": "Node-wise Filtering in Graph Neural Networks: A Mixture of Experts Approach", "abs": "Graph Neural Networks (GNNs) have proven to be highly effective for node classification tasks across diverse graph structural patterns. Traditionally, GNNs employ a uniform global filter\u2014typically a low-pass filter for homophilic graphs and a high-pass filter for heterophilic graphs. However, real-world graphs often exhibit a complex mix of homophilic and heterophilic patterns, rendering a single filter approach suboptimal. In this work, we theoretically demonstrate that a global filter optimized for one pattern can adversely affect performance on nodes with differing patterns. To address this, we introduce a novel GNN framework Node-MoE that utilizes a mixture of experts to adaptively select the appropriate filters for different nodes. Extensive experiments demonstrate the effectiveness of the proposed Node-MoE on both homophilic and heterophilic graphs.", "keywords": [ "Graph Neural Networks", "Node-wise Filters", "Node Classification", "Homophilic and Heterophilic graphs" ], "rating_list": [ 6, 5, 3, 6, 5 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "tijmpS9Vy2", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Michael D Dennis", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Stuart Russell", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Alyssa Li Dayan", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" } ], "n_formula": 68, "n_formula_1": 12, "n_ref_uni": 79, "n_ref": 124, "n_ref_all": 169, "n_fig": 19, "n_tab": 3, "L_tab": 1730, "n_element_tab": 64, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 497, "n_element_tab_1": 35, "formula_len_all": 4903, "formula_len_all_1": 760, "len_all": 241124, "len_all_1": 68096, "len_abs": 1476, "len_title": 138, "len_sents": 79071, "len_sents_1": 33274, "n_sents": 520, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1482, "title": "BAMDP Shaping: a Unified Theoretical Framework for Intrinsic Motivation and Reward Shaping", "abs": "Intrinsic motivation and reward shaping guide reinforcement learning (RL) agents\nby adding pseudo-rewards, which can lead to useful emergent behaviors. However, they can also exhibit unanticipated side effects \u2013 leading to reward hacking\nor fixation with noisy TVs. Here we provide a theoretical model which anticipates these behaviors, and provides broad criteria under which their effects can\nbe bounded. We characterize all pseudo-rewards as reward shaping in Bayes-Adaptive Markov Decision Processes (BAMDPs), which formulates the problem\nof learning in MDPs as an MDP over the agent\u2019s knowledge. We can understand\npseudo-rewards as guiding exploration by incentivizing RL agents to go to states\nwith higher BAMDP value, which comprises the value of information gathered\nand the prior value of the physical state, while they mislead exploration when they\nalign poorly with this value. We extend potential-based shaping theory (Ng et al.,\n1999) to prove only BAMDP Potential-based shaping Functions (BAMPFs) are\nguaranteed to preserve the optimal RL algorithm, and show empirically how a\nBAMPF helps a meta-RL agent learn an optimal RL agent for a Bernoulli Bandit\ndomain. We finally prove that BAMPFs with bounded monotone potentials are\nalso resistant to reward-hacking in MDPs. We show that it is straightforward to\nretrofit or design new pseudo-reward terms in this form to avoid unintended side\neffects, and provide an empirical demonstration in the Mountain Car environment", "keywords": [ "Reinforcement Learning Theory", "Bayesian Reinforcement Learning", "Intrinsic Motivation", "Reward Shaping" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "tidibw8Xdm", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Aviral Kumar", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Oier Mees", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department, University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "William Chen", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 61, "n_ref": 123, "n_ref_all": 159, "n_fig": 8, "n_tab": 10, "L_tab": 2997, "n_element_tab": 360, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2060, "n_element_tab_1": 181, "formula_len_all": 111, "formula_len_all_1": 0, "len_all": 278916, "len_all_1": 80277, "len_abs": 1261, "len_title": 132, "len_sents": 87096, "len_sents_1": 41044, "n_sents": 609, "n_sents_1": 250, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1266, "title": "Vision-Language Models Provide Promptable Representations for Reinforcement Learning", "abs": "Humans can quickly learn new behaviors by leveraging background world knowledge. In contrast, agents trained with reinforcement learning (RL) typically learn behaviors from scratch. We thus propose a novel approach that uses the vast amounts of general and indexable world knowledge encoded in vision-language models (VLMs) pre-trained on Internet-scale data for embodied RL. We initialize policies with VLMs by using them as promptable representations: embeddings that encode semantic features of visual observations based on the VLM's internal knowledge and reasoning capabilities, as elicited through prompts that provide task context and auxiliary information. We evaluate our approach on visually-complex, long horizon RL tasks in Minecraft and robot navigation in Habitat. We find that our policies trained on embeddings from off-the-shelf, general-purpose VLMs outperform equivalent policies trained on generic, non-promptable image embeddings. We also find our approach outperforms instruction-following methods and performs comparably to domain-specific embeddings. Finally, we show that our approach can use chain-of-thought prompting to produce representations of common-sense semantic reasoning, improving policy performance in novel scenes by 1.5 times.", "keywords": [ "vision-language models", "reinforcement learning", "embodied control" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "tiKJsepvr0", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Ali Murad", "gender": "Male", "institution": "Auburn University", "country": "US", "position": "PhD student" }, { "name": "Bo Hui", "gender": "unknown", "institution": "University of Tulsa", "country": "US", "position": "Assistant Professor" }, { "name": "Wei-Shinn Ku", "gender": "Male", "institution": "Auburn University", "country": "US", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 19, "n_ref_uni": 31, "n_ref": 45, "n_ref_all": 57, "n_fig": 6, "n_tab": 4, "L_tab": 2562, "n_element_tab": 362, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2404, "n_element_tab_1": 348, "formula_len_all": 1662, "formula_len_all_1": 1634, "len_all": 126048, "len_all_1": 64737, "len_abs": 1695, "len_title": 126, "len_sents": 28140, "len_sents_1": 27438, "n_sents": 194, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1702, "title": "Optimal Client Training in Federated Learning with Deep Reinforcement Learning", "abs": "Federated Learning (FL) is a distributed framework for collaborative model training over large-scale distributed data. Centralized FL leverages a server to aggregate client models which can enable higher performance while maintaining client data privacy. However, it has been shown that in centralized model aggregation, performance can degrade in the presence of non-IID data across different clients. We remark that training a client locally on more data than necessary does not benefit the overall performance of all clients. In this paper, we devise a novel framework that leverages Deep Reinforcement Learning (DRL) to optimize an agent that selects the optimal amount of data necessary to train a client model without oversharing information with the server. Starting from complete unawareness of the client's performance, the DRL agent utilizes the change in training loss as a reward signal and learns to optimize the amount of data necessary for improving the client's performance. Specifically, after each aggregation round, the DRL algorithm considers the local performance as the current state and outputs the optimal weights for each class in the training data to be used during the next round of local training. In doing so, the agent learns a policy that creates the optimal partition of the local training dataset during the FL rounds. After FL, the client utilizes the entire local training dataset to further enhance its performance on its own data distribution, mitigating the non-IID effects of aggregation. Through extensive experiments, we demonstrate that training FL clients through our algorithm results in superior performance on multiple benchmark datasets and FL frameworks.", "keywords": [ "Federated Learning" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 2, 1, 2 ], "presentation_list": [ 1, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "tiJzOop4u6", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Petr Molodyk", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yongxin Chen", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Xue Haotian", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 6, "n_ref_uni": 40, "n_ref": 152, "n_ref_all": 184, "n_fig": 14, "n_tab": 3, "L_tab": 1743, "n_element_tab": 238, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 2237, "n_element_tab_1": 263, "formula_len_all": 776, "formula_len_all_1": 407, "len_all": 145541, "len_all_1": 66530, "len_abs": 1530, "len_title": 124, "len_sents": 43669, "len_sents_1": 28109, "n_sents": 310, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1521, "title": "Rethinking Adversarial Attacks as Protection Against Diffusion-based Mimicry", "abs": "Diffusion models have demonstrated an remarkable capability to edit or imitate images, which has raised concerns regarding the safeguarding of intellectual property. To address these concerns, the adoption of adversarial attacks, which introduce adversarial perturbations that can fool the targeted diffusion model into protected images , has emerged as a viable solution. Consequently, diffusion models, like many other deep network models, are believed to be susceptible to adversarial attacks. However, in this work, we draw attention to an important oversight in existing research, as all previous studies have focused solely on attacking latent diffusion models (LDMs), neglecting adversarial examples for diffusion models in the pixel space (PDMs). Through extensive experiments, we demonstrate that nearly all existing adversarial attack methods designed for LDMs, as well as adaptive attacks designed for PDMs, fail when applied to PDMs. We attribute the vulnerability of LDMs to their encoders, indicating that diffusion models exhibit strong robustness against adversarial attacks. Building upon this insight, we find that PDMs can be used as an off-the-shelf purifier to effectively eliminate adversarial patterns generated by LDMs, thereby maintaining the integrity of images. Notably, we highlight that most existing protection methods can be easily bypassed using PDM-based purification. We hope our findings prompt a reevaluation of adversarial samples for diffusion models as potential protection methods.", "keywords": [ "Generative Model; Diffusion Model; Adversarial Attack" ], "rating_list": [ 8, 8, 3, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "thqPibDg6A", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 16, "author_info_list": [ { "name": "David A. Clifton", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Dongsheng Li", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Fan Yang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Professor" }, { "name": "Fang Dong", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Hengjie Cao", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Jixian Zhou", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Li Shang", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Mengyi Chen", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Mingzhi Dong", "gender": "Male", "institution": "University of Bath", "country": "CN", "position": "Lecturer" }, { "name": "Ning Gu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Qin Lv", "gender": "Female", "institution": "University of Colorado at Boulder", "country": "US", "position": "Full Professor" }, { "name": "Robert P. Dick", "gender": "Male", "institution": "University of Michigan", "country": "US", "position": "Full Professor" }, { "name": "Ruijun Huang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Tun Lu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yifeng Yang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Yujiang Wang", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 35, "n_ref": 57, "n_ref_all": 73, "n_fig": 7, "n_tab": 5, "L_tab": 1133, "n_element_tab": 146, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1521, "n_element_tab_1": 175, "formula_len_all": 481, "formula_len_all_1": 383, "len_all": 137728, "len_all_1": 64304, "len_abs": 1675, "len_title": 150, "len_sents": 33277, "len_sents_1": 28468, "n_sents": 239, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1689, "title": "Efficiently pre-training language models with mixtures of cluster-oriented, trainability-aware experts", "abs": "Language models (LMs) are pre-trained on large-scale corpora from diverse data sources, encapsulating knowledge across various domains, with their feature spaces often displaying clustering structures. The mixture of experts (MoEs) approach is commonly used to scale up model learning capabilities to handle such complexities; however, the fine-grained learning dynamics at the expert level remain largely unexplored. This work analyzes the spatial and temporal characteristics of these clustering structures and examines their impact on the fine-grained trainability of individual experts. Our analysis builds on the singular spectrum of the feature and Jacobian spaces leading to two key observations. First, a few top singular vectors from the feature matrix are sufficient to capture the layer-wise feature cluster patterns. More interestingly, the maximum singular value of the Jacobian matrix reveals conflicts between different feature clusters, and experts exhibit varying levels of trainability, completing their learning asynchronously during training.\nInspired by these insights, we proposed mixtures of cluster-guided, trainability-aware experts (MO-CTE), with \nan efficient routing method to mitigate inter-cluster conflicts to improve expert trainability and \na simple yet effective criterion for early stopping low-trainability experts, thus reducing total training costs. \nWe evaluate the proposed MO-CTE across extensive datasets and tasks. Experimental results indicate that MO-CTE accelerates convergence by approximately 37\\% in test perplexity and 30\\% in downstream tasks, and improves performance by 3.6\\% over baselines when consuming similar computation resources.", "keywords": [ "Mixture of Expert", "Language models", "Clustering", "Trainability" ], "rating_list": [ 6, 3, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 1, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 2, 3, 4 ] }, { "paper_id": "thV5KRQFgQ", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Dawei Cheng", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Associate Professor" }, { "name": "Guibin Zhang", "gender": "Not Specified", "institution": "Tongji University", "country": "CN", "position": "Undergrad student" }, { "name": "Jian Guo", "gender": "Male", "institution": "International Digital Economy Academy, International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Yanwei Yue", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Undergrad student" }, { "name": "Yiyan Qi", "gender": "unknown", "institution": "IDEA", "country": "CN", "position": "Researcher" }, { "name": "Ziyang Cheng", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 21, "n_formula_1": 18, "n_ref_uni": 68, "n_ref": 172, "n_ref_all": 200, "n_fig": 6, "n_tab": 12, "L_tab": 5881, "n_element_tab": 328, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 2204, "n_element_tab_1": 71, "formula_len_all": 1767, "formula_len_all_1": 1394, "len_all": 206211, "len_all_1": 65666, "len_abs": 1585, "len_title": 106, "len_sents": 42508, "len_sents_1": 29296, "n_sents": 278, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1594, "title": "Rationalizing and Augmenting Dynamic Graph Neural Networks", "abs": "Graph data augmentation (GDA) has shown significant promise in enhancing the performance, generalization, and robustness of graph neural networks (GNNs). However, contemporary methodologies are often limited to static graphs, whose applicability on dynamic graphs\u2014more prevalent in real-world applications\u2014remains unexamined. In this paper, we empirically highlight the challenges faced by static GDA methods when applied to dynamic graphs, particularly their inability to maintain temporal consistency. In light of this limitation, we propose a dedicated augmentation framework for dynamic graphs, termed $\\texttt{DyAug}$, which adaptively augments the evolving graph structure with temporal consistency awareness. Specifically, we introduce the paradigm of graph rationalization for dynamic GNNs, progressively distinguishing between causal subgraphs (\\textit{rationale}) and the non-causal complement (\\textit{environment}) across snapshots. We develop three types of environment replacement, including, spatial, temporal, and spatial-temporal, to facilitate data augmentation in the latent representation space, thereby improving the performance, generalization, and robustness of dynamic GNNs. Extensive experiments on six benchmarks and three GNN backbones demonstrate that $\\texttt{DyAug}$ can \\textbf{(I)} improve the performance of dynamic GNNs by $0.89\\\\%\\sim3.13\\\\%\\uparrow$; \\textbf{(II)} effectively counter targeted and non-targeted adversarial attacks with $6.2\\\\%\\sim12.2\\\\%\\\\uparrow$ performance boost; \\textbf{(III)} make stable predictions under temporal distribution shifts.", "keywords": [ "Graph Data Augmentation", "Dynamic Graph Neural Networks" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "th63j8qHa6", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Junteng Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Junxian He", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "LIN Xinxin", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Undergrad student" }, { "name": "Minrui Xu", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Undergrad student" }, { "name": "Ruochen Zhou", "gender": "Female", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shiqi Chen", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Intern" }, { "name": "Yunqi Li", "gender": "Not Specified", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Undergrad student" }, { "name": "Zhengyu Chen", "gender": "unknown", "institution": "Meituan", "country": "CN", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 48, "n_ref": 107, "n_ref_all": 131, "n_fig": 5, "n_tab": 9, "L_tab": 2494, "n_element_tab": 189, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 1441, "n_element_tab_1": 108, "formula_len_all": 622, "formula_len_all_1": 585, "len_all": 199281, "len_all_1": 63622, "len_abs": 2011, "len_title": 124, "len_sents": 50437, "len_sents_1": 29990, "n_sents": 376, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 2031, "title": "Math for AI: On the Generalization of Learning Mathematical Problem Solving", "abs": "There has been a growing interest in enhancing the mathematical problem-solving (MPS) capabilities of LLMs. While some researchers focus on developing specialized math models to advance AI for math, others study mathematical reasoning with a ''math for AI'' perspective, positing that integrating mathematical reasoning data could enable LLMs to perform complex reasoning more broadly. This hypothesis draws from neuroscience studies which show that solving mathematical problems aids in the development of general reasoning skills in humans. The concept of ''math for AI'' has gained particular relevance as the research community increasingly focuses on complex reasoning -- Given the scarcity of complex and lengthy chain-of-thought data, MPS emerges as a prime candidate for collecting or synthesizing substantial volumes of intricate thought processes, thus serving as a potential key resource for enhancing general complex reasoning. However, it remains unclear whether skills acquired through learning MPS can extend to other reasoning tasks or merely improve MPS-specific benchmark scores. \nIn this paper, we present a comprehensive empirical analysis to address this question.\nSpecifically, we explore three prevalent methods for improving MPS: (1) continual pretraining on mathematical text; (2) instruction pretraining on large-scale QA pairs synthesized from raw text; and (3) instruction tuning on MPS datasets. \nThrough controlled experiments and evaluations across seven distinct reasoning domains, \nwhile no approaches consistently generalize across all non-mathematical tasks, both continual pretraining and instruction pretraining outperform instruction tuning, with continual pretraining often yielding greater gains when effective.\nThese findings indicate that most readily available data sources do not support the ''math for AI'' objective in enhancing non-MPS tasks. Identifying which data sources best contribute to the acquisition of complex reasoning skills remains a crucial question for future research.", "keywords": [ "Large Language Models", "Mathematical Reasoning", "Reasoning Generalization" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "tfyHbvFZ0K", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Jun Zhao", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "", "position": "Full Professor" }, { "name": "Kang Liu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Professor" }, { "name": "Pengfei Cao", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Yubo Chen", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "", "position": "Associate Professor" }, { "name": "Yuheng Chen", "gender": "Not Specified", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 31, "n_formula_1": 7, "n_ref_uni": 38, "n_ref": 91, "n_ref_all": 142, "n_fig": 10, "n_tab": 4, "L_tab": 4246, "n_element_tab": 503, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 2410, "n_element_tab_1": 378, "formula_len_all": 1391, "formula_len_all_1": 311, "len_all": 193827, "len_all_1": 67033, "len_abs": 1350, "len_title": 123, "len_sents": 51850, "len_sents_1": 27295, "n_sents": 427, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 75, "L_abs": 1370, "title": "Knowledge Localization: Mission Not Accomplished? Enter Query Localization!", "abs": "Large language models (LLMs) store extensive factual knowledge, but the mechanisms behind how they store and express this knowledge remain unclear.\nThe Knowledge Neuron (KN) thesis is a prominent theory for explaining these mechanisms. This theory is based on the **Knowledge Localization (KL)** assumption, which suggests that a fact can be localized to a few knowledge storage units, namely knowledge neurons.\n However, this assumption has two limitations: first, it may be too rigid regarding knowledge storage, and second, it neglects the role of the attention module in knowledge expression. \n \nIn this paper, we first re-examine the KL assumption and demonstrate that its limitations do indeed exist. To address these, we then present two new findings, each targeting one of the limitations: one focusing on knowledge storage and the other on knowledge expression.\nWe summarize these findings as **Query Localization** assumption and argue that the KL assumption can be viewed as a simplification of the QL assumption. \nBased on QL assumption, we further propose the Consistency-Aware KN modification method, which improves the performance of knowledge modification, further validating our new assumption. We conduct 39 sets of experiments, along with additional visualization experiments, to rigorously confirm our conclusions. Code will be made public soon.", "keywords": [ "Knowledge Neruon Thesis", "Knowledge Localization", "Query Localization" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "tfemquulED", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Jun-cheng Chen", "gender": "Male", "institution": "Research Center for Information Technology Innovation, Academia Sinica", "country": "TW", "position": "Associate Research Fellow/Professor" }, { "name": "Kuang-Huei Lee", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Po-Hung Yeh", "gender": "Not Specified", "institution": "Academia Sinica", "country": "TW", "position": "RA" } ], "n_formula": 92, "n_formula_1": 18, "n_ref_uni": 27, "n_ref": 76, "n_ref_all": 138, "n_fig": 11, "n_tab": 15, "L_tab": 4215, "n_element_tab": 495, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1215, "n_element_tab_1": 120, "formula_len_all": 6902, "formula_len_all_1": 1384, "len_all": 189263, "len_all_1": 54380, "len_abs": 1177, "len_title": 108, "len_sents": 56457, "len_sents_1": 22902, "n_sents": 536, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1181, "title": "Training-Free Diffusion Model Alignment with Sampling Demons", "abs": "Aligning diffusion models with user preferences has been a key challenge.\nExisting methods for aligning diffusion models either require retraining or are limited to differentiable reward functions.\nTo address these limitations, we propose a stochastic optimization approach, dubbed Demon, to guide the denoising process at inference time without backpropagation through reward functions or model retraining.\nOur approach works by controlling noise distribution in denoising steps to concentrate density on regions corresponding to high rewards through stochastic optimization.\nWe provide comprehensive theoretical and empirical evidence to support and validate our approach, including experiments that use non-differentiable sources of rewards such as Visual-Language Model (VLM) APIs and human judgments.\nTo the best of our knowledge, the proposed approach is the first inference-time, backpropagation-free preference alignment method for diffusion models.\nOur method can be easily integrated with existing diffusion models without further training.\nOur experiments show that the proposed approach significantly improves the average aesthetics scores for text-to-image generation.", "keywords": [ "Diffusion Model", "Consistency Model", "Score-Based Method", "EDM", "Stochastic Process", "It\u00f4's Lemma", "RLHF", "Value Function", "Preference Alignment" ], "rating_list": [ 6, 6, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 3, 3 ] }, { "paper_id": "tfTGSm31F7", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Beibei Chen", "gender": "unknown", "institution": "4th Military Medical University", "country": "CN", "position": "Associate Professor" }, { "name": "Junwei Han", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "", "position": "Full Professor" }, { "name": "Lei Guo", "gender": "Male", "institution": "Northwestern Polytechnical University Xi'an", "country": "", "position": "Full Professor" }, { "name": "Ning Zhu", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "MS student" }, { "name": "Tianming Liu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Professor" }, { "name": "Tuo Zhang", "gender": "Male", "institution": "Northwestern Polytechnical University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaohui Gao", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "PhD student" }, { "name": "Yifan Chen", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "MS student" }, { "name": "Zhiqiang Yan", "gender": "unknown", "institution": "4th Military Medical University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhuoyi Li", "gender": "Female", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "PhD student" }, { "name": "\u9648\u6d69", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 14, "n_ref_uni": 51, "n_ref": 70, "n_ref_all": 88, "n_fig": 6, "n_tab": 7, "L_tab": 5282, "n_element_tab": 547, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 3587, "n_element_tab_1": 270, "formula_len_all": 1002, "formula_len_all_1": 1107, "len_all": 159197, "len_all_1": 63834, "len_abs": 2126, "len_title": 105, "len_sents": 43614, "len_sents_1": 27918, "n_sents": 267, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 2132, "title": "EpilepsyFM: Foundation Model for Learning Generalized Epileptic Representations from EEG and SEEG Signals", "abs": "Extracranial electroencephalography (EEG) and intracranial stereoelectroencephalography (SEEG) are crucial for epilepsy diagnosis. However, existing deep learning models often limit themselves to specific signal types and application scenarios, leading to challenges in generalization and perception capabilities. While large language models excel in natural language processing, they cannot effectively capture the disease-specific signal features in the highly specialized field of epilepsy, and the lack of pre-training data restricts their generalization ability. To address these issues, we propose a Epilepsy Foundation Model (EpilepsyFM), a domain-specific foundational model that considers the mechanisms of seizure and propagation in epilepsy. EpilepsyFM learns a generalized representation of epilepsy through unsupervised pre-training across various signal types, data formats, and sources, and optimizes multiple epilepsy-related downstream tasks through fine-tuning. We collected clinical EEG and SEEG data from multiple patients at a first-class hospital, as well as the currently largest publicly available epilepsy dataset, the TUH series, ensuring diversity in representation learning. First, the neural activity signals are segmented into multiple patches, and a discrete EEG and SEEG neural tokenizer is trained to construct a domain-specific neural codebook for epilepsy. Then, EpilepsyFM takes into account the mechanisms of clustered neuronal discharges in epilepsy and designs a channel set masking strategy to enhance the model's ability to capture the spatiotemporal characteristics of the signals. The model fully utilizes the multi-dimensional propagation characteristics of seizures through temporal, spectral, and spatial encoder modules, achieving comprehensive representation of complex neural signals. Extensive experiments show that EpilepsyFM achieves state-of-the-art performance in a variety of domain-specific tasks, including seizure detection and both short-term and long-term predictions of neural signals, demonstrating strong generalization ability and broad clinical application potential.", "keywords": [ "Epilepsy", "Foundation Model", "Electroencephalography", "Stereoelectroencephalography" ], "rating_list": [ 3, 5, 3, 3, 6 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 5, 3, 4, 5, 4 ] }, { "paper_id": "tfO07iz0b9", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Chenyu Zhang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Xu Chen", "gender": "Male", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Xuan Di", "gender": "Female", "institution": "Columbia University", "country": "US", "position": "Associate Professor" } ], "n_formula": 184, "n_formula_1": 29, "n_ref_uni": 35, "n_ref": 176, "n_ref_all": 235, "n_fig": 18, "n_tab": 7, "L_tab": 946, "n_element_tab": 71, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 21944, "formula_len_all_1": 2862, "len_all": 292460, "len_all_1": 72663, "len_abs": 1506, "len_title": 156, "len_sents": 99057, "len_sents_1": 33303, "n_sents": 948, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 107, "L_abs": 1513, "title": "Stochastic Semi-Gradient Descent for Learning Mean Field Games with Population-Aware Function Approximation", "abs": "Mean field games (MFGs) model interactions in large-population multi-agent systems through population distributions. Traditional learning methods for MFGs are based on fixed-point iteration (FPI), where policy updates and induced population distributions are computed separately and sequentially. However, FPI-type methods may suffer from inefficiency and instability due to potential oscillations caused by this forward-backward procedure. In this work, we propose a novel perspective that treats the policy and population as a unified parameter controlling the game dynamics. By applying stochastic parameter approximation to this unified parameter, we develop SemiSGD, a simple stochastic gradient descent (SGD)-type method, where an agent updates its policy and population estimates simultaneously and fully asynchronously. Building on this perspective, we further apply linear function approximation (LFA) to the unified parameter, resulting in the first population-aware LFA (PA-LFA) for learning MFGs on continuous state-action spaces. A comprehensive finite-time convergence analysis is provided for SemiSGD with PA-LFA, including its convergence to the equilibrium for linear MFGs\u2014a class of MFGs with a linear structure concerning the population\u2014under the standard contractivity condition, and to a neighborhood of the equilibrium under a more practical condition. We also characterize the approximation error for non-linear MFGs. We validate our theoretical findings with six experiments on three MFGs.", "keywords": [ "mean field game", "linear function approximation", "stochastic semi-gradient descent" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "tet8yGrbcf", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Mats Leon Richter", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Postdoc" }, { "name": "Maxime Gasse", "gender": "Male", "institution": "ServiceNow", "country": "", "position": "Researcher" }, { "name": "Megh Thakkar", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "MS student" }, { "name": "Mohammad Reza Samsami", "gender": "Male", "institution": "Mila", "country": "", "position": "MS student" }, { "name": "Sarath Chandar", "gender": "Male", "institution": "\u00c9cole Polytechnique de Montr\u00e9al", "country": "CA", "position": "Associate Professor" }, { "name": "Juan Rodriguez", "gender": "Male", "institution": "ServiceNow Research", "country": "CA", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 0, "n_ref_uni": 27, "n_ref": 38, "n_ref_all": 50, "n_fig": 7, "n_tab": 10, "L_tab": 4513, "n_element_tab": 204, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 508, "n_element_tab_1": 34, "formula_len_all": 189, "formula_len_all_1": 0, "len_all": 170079, "len_all_1": 58159, "len_abs": 4666, "len_title": 55, "len_sents": 42571, "len_sents_1": 29577, "n_sents": 308, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 914, "title": "Too Big to Fool: Resisting Deception in Language Models", "abs": "Large language models must balance their weight-encoded knowledge with in-context information from prompts to generate accurate responses. This paper investigates this interplay by analyzing how models of varying capacities within the same family handle intentionally misleading in-context information. Our experiments demonstrate that larger models exhibit higher resilience to deceptive prompts, showcasing an advanced ability to interpret and integrate prompt information with their internal knowledge. Furthermore, we find that larger models outperform smaller ones in following legitimate instructions, indicating that their resilience is not due to disregarding in-context information. We also show that this phenomenon is likely not a result of memorization but stems from the models' ability to better leverage implicit task-relevant information from the prompt alongside their internally stored knowledge.", "keywords": [ "Large Language Models", "Evaluation", "Misinformation", "In-Context Learning", "World Models", "Reasoning" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "tePFpDgyqg", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Eric P. Xing", "gender": "Male", "institution": "Mohamed bin Zayed Univeristy of AI", "country": "AE", "position": "Full Professor" }, { "name": "Hao Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Lanxiang Hu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Longfei Yun", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "MS student" }, { "name": "Souvik Kundu", "gender": "Male", "institution": "Intel", "country": "US", "position": "Researcher" }, { "name": "Yonghao Zhuang", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Zhengzhong Liu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 27, "n_ref": 57, "n_ref_all": 78, "n_fig": 9, "n_tab": 5, "L_tab": 2630, "n_element_tab": 420, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 226, "n_element_tab_1": 40, "formula_len_all": 61, "formula_len_all_1": 61, "len_all": 143314, "len_all_1": 68341, "len_abs": 1568, "len_title": 109, "len_sents": 42122, "len_sents_1": 33331, "n_sents": 367, "n_sents_1": 285, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1575, "title": "Scaling Long Context Training Data by Long-Distance Referrals", "abs": "Training large language models for long context understanding faces the challenge of data shortage.\nPrevious data engineering approaches mechanically concatenate short documents, which may create many pseudo long documents but raise concerns about data quality.\nIn this paper, we study the core attribute of high quality data for long context training, and provide a data pipeline, LongPack, to scale\nsuch data.\nWe found that long distance referrals, which occur in natural long documents, are crucial for long-context training.\nHowever, simply concatenating short documents does not reliably generate these relations.\nWe further show that the density of long-distance referrals, which is higher in longer documents, has a key role in training efficiency, making previous upsampling methods suboptimal.\nTo enrich long documents, we propose LongPack, a data pipeline that constructs long documents by packing shorter ones based on referral relationships.\nSpecifically, for web pages, which are the primary source for language model training, we found hyper-link a native signal for such a relation.\nBy packing web pages through their hyper-link connection, we can create longer, high-quality documents.\nOur experiments demonstrate that LongPackis highly scalable, generating a corpus of long documents equivalent in size to an entire pretraining dataset using just 0.5% root documents.\nFurthermore, the constructed documents have a \u2018near-natural\u2019 quality as innate long documents for long context training, reaching a 32.7% higher score than previous state-of-the-art methods.", "keywords": [ "long context training; continue training; dataset;" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "teE4pl9ftK", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Abdul Fatir Ansari", "gender": "Male", "institution": "AWS AI Labs (Amazon)", "country": "DE", "position": "Senior Scientist" }, { "name": "Andrew Stuart", "gender": "unknown", "institution": "California Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Bernie Wang", "gender": "Male", "institution": "AWS AI Labs", "country": "US", "position": "Principal Researcher" }, { "name": "Boran Han", "gender": "Female", "institution": "Amazon/AWS", "country": "US", "position": "Researcher" }, { "name": "Chaoran Cheng", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" }, { "name": "Danielle C. Maddix", "gender": "unknown", "institution": "AWS AI Labs", "country": "US", "position": "Applied Scientist" }, { "name": "Michael W. Mahoney", "gender": "unknown", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" } ], "n_formula": 40, "n_formula_1": 10, "n_ref_uni": 41, "n_ref": 125, "n_ref_all": 166, "n_fig": 12, "n_tab": 12, "L_tab": 4186, "n_element_tab": 540, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2541, "n_element_tab_1": 305, "formula_len_all": 3625, "formula_len_all_1": 1006, "len_all": 215087, "len_all_1": 67569, "len_abs": 1562, "len_title": 101, "len_sents": 70703, "len_sents_1": 29383, "n_sents": 528, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1566, "title": "Gradient-Free Generation for Hard-Constrained Systems", "abs": "Generative models that satisfy hard constraints are crucial in scientific applications, e.g., numerical simulations, dynamical systems, and supply chain optimization, where physical laws or system requirements must be strictly respected. However, many existing constrained generative models, especially those developed for computer vision, rely heavily on gradient information, which is often sparse or computationally expensive in other fields, e.g., partial differential equations (PDEs). Accurately solving these problems numerically demands the generated solutions to comply with strict physical constraints, e.g., conservation laws. In this work, we introduce a novel framework for adapting pre-trained, unconstrained generative models to exactly satisfy constraints in a zero-shot manner, without requiring expensive gradient computations or fine-tuning. Our framework, ECI sampling, alternates between extrapolation (E), correction (C), and interpolation (I) stages during each iterative sampling step to ensure accurate integration of constraint information while preserving the validity of the generated outputs. We demonstrate the efficacy of our approach across various PDE systems, showing that ECI-guided generation strictly adheres to physical constraints and accurately captures complex distribution shifts induced by these constraints. Empirical results show that our framework consistently outperforms baseline approaches in both zero-shot constrained generative and regression tasks, and achieves competitive results without additional fine-tuning.", "keywords": [ "Flow Matching", "Generative Model", "Constrained Generation", "Partial Differential Equations", "Conservation Laws" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 3, 1 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "te30nmLaFf", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Abbavaram Gowtham Reddy", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Postdoc" }, { "name": "Abhinav Kumar", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Amit Sharma", "gender": "Male", "institution": "Microsoft Research", "country": "IN", "position": "Principal Researcher" }, { "name": "Aniket Vashishtha", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Atharva Pandey", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Researcher" }, { "name": "Vineeth Balasubramanian", "gender": "Male", "institution": "Research, Microsoft", "country": "IN", "position": "Principal Researcher" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 26, "n_ref": 78, "n_ref_all": 94, "n_fig": 4, "n_tab": 7, "L_tab": 4277, "n_element_tab": 523, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 414, "formula_len_all_1": 462, "len_all": 147446, "len_all_1": 69257, "len_abs": 1697, "len_title": 113, "len_sents": 52228, "len_sents_1": 35894, "n_sents": 397, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1706, "title": "Teaching Transformers Causal Reasoning through Axiomatic Training", "abs": "For text-based AI systems to interact in the real world, causal reasoning is an essential skill. Since active interventions are costly to execute, we study to what extent an agent can learn causal reasoning from symbolic demonstrations of causal axioms. Specifically, we consider an axiomatic training setup where an agent learns from multiple demonstrations of a causal axiom (or rule), rather than incorporating the axiom as an inductive bias or inferring it from data values. A key question is whether the agent would learn to generalize from the axiom demonstrations to new scenarios. For example, if a transformer model is trained on demonstrations of the causal transitivity axiom over small graphs, would it generalize to applying the transitivity axiom over large graphs? \nOur results, based on a novel axiomatic training scheme, indicate that such generalization is possible. For the transitivity axiom, we find that a 67 million parameter transformer model, when trained on linear causal chains (along with some noisy variations) can generalize well to new kinds of graphs, including longer causal chains, causal chains with reversed order, and graphs with branching; even when it is not explicitly trained for such settings. We extend axiomatic training to a harder task of inferring causation from correlation statements and find similar generalization. On both tasks, our model performs at par (or even better) than many larger language models such as GPT-4, Gemini Pro, and Phi-3. Overall, our axiomatic training framework provides a new paradigm of learning causal reasoning in language models that can be extended to arbitrary axioms, as long as sufficient demonstrations can be generated.", "keywords": [ "Causal Axioms", "Transformers", "Generalization", "LLMs" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "te2Q9dThlE", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Feng Xu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Guotai Wang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "", "position": "Associate Professor" }, { "name": "Linda Wei", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Minhao Hu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Shaoting Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Full Professor" }, { "name": "Shixiang Feng", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Tao Song", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Xiangde Luo", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "Researcher" }, { "name": "Yi Guo", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yicheng Wu", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 21, "n_ref": 26, "n_ref_all": 40, "n_fig": 5, "n_tab": 4, "L_tab": 1590, "n_element_tab": 205, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1147, "n_element_tab_1": 133, "formula_len_all": 704, "formula_len_all_1": 704, "len_all": 93186, "len_all_1": 54844, "len_abs": 1131, "len_title": 123, "len_sents": 24541, "len_sents_1": 24219, "n_sents": 173, "n_sents_1": 168, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1163, "title": "Multimodal MR Image Synthesis via Learning Adaptive Group-wise Interactions", "abs": "Multimodal MR image synthesis aims to generate missing modality image by fusing and mapping a few available MRI data. Most existing approaches typically adopt an image-to-image translation scheme. However, these methods often suffer from sub-optimal performance due to the spatial misalignment between different modalities while they are typically treated as input channels. Therefore, in this paper, we propose an \\textit{Adaptive Group-wise Interaction Network} (\\textbf{AGI-Net}) that explores both inter-modality and intra-modality relationships for multimodal MR image synthesis. Specifically, groups are first pre-defined along the channel dimension and then we perform an adaptive rolling for the standard convolutional kernel to capture inter-modality spatial correspondences. At the same time, a cross-group attention module is introduced to fuse information across different channel groups, leading to better feature representation. We evaluated the effectiveness of our model on the publicly available IXI and BraTS2023 datasets, where the AGI-Net achieved state-of-the-art performance for multimodal MR image synthesis. \\textit{Code will be released}.", "keywords": [ "Multimodal Image Synthesis; Group-wise Rolling; Cross Group Attention" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "te2IdORabL", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Ahmed Hussein Salamah", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "EN-HUI YANG", "gender": "unknown", "institution": "University of Waterloo", "country": "CA", "position": "Full Professor" }, { "name": "Kaixiang Zheng", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Yiwen Liu", "gender": "unknown", "institution": "University of Waterloo", "country": "CA", "position": "MEng student" } ], "n_formula": 23, "n_formula_1": 14, "n_ref_uni": 36, "n_ref": 56, "n_ref_all": 101, "n_fig": 13, "n_tab": 10, "L_tab": 2348, "n_element_tab": 152, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 2041, "n_element_tab_1": 40, "formula_len_all": 1691, "formula_len_all_1": 836, "len_all": 139461, "len_all_1": 57516, "len_abs": 3205, "len_title": 27, "len_sents": 46079, "len_sents_1": 26235, "n_sents": 342, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 27, "L_abs": 1079, "title": "JPEG Inspired Deep Learning", "abs": "Although it is traditionally believed that lossy image compression, such as JPEG compression, has a negative impact on the performance of deep neural networks (DNNs), it is shown by recent works that well-crafted JPEG compression can actually improve the performance of deep learning (DL). Inspired by this, we propose JPEG-DL, a novel DL framework that prepends any underlying DNN architecture with a trainable JPEG compression layer. To make the quantization operation in JPEG compression trainable, a new differentiable soft quantizer is employed at the JPEG layer, and then the quantization operation and underlying DNN are jointly trained. Extensive experiments show that in comparison with the standard DL, JPEG-DL delivers significant accuracy improvements across various datasets and model architectures while enhancing robustness against adversarial attacks. Particularly, on some fine-grained image classification datasets, JPEG-DL can increase prediction accuracy by as much as 20.9%. Our code is available on https://github.com/AhmedHussKhalifa/JPEG-Inspired-DL.git.", "keywords": [ "Deep Learning", "JPEG Compression", "Quantization", "Non-linearity" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 4, 2, 3 ] }, { "paper_id": "tdttNKCtyB", "primary_area": "learning on time series and dynamical systems", "n_author": 9, "author_info_list": [ { "name": "Bin Yang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Chenjuan Guo", "gender": "unknown", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Kai Zhao", "gender": "Non-Binary", "institution": "Aalborg University", "country": "DK", "position": "PhD student" }, { "name": "Lujia Pan", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yang Shu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yihang Wang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Yuying Qiu", "gender": "Female", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Zhongwen Rao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "P Chen", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 34, "n_ref": 92, "n_ref_all": 146, "n_fig": 13, "n_tab": 13, "L_tab": 15549, "n_element_tab": 1006, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 641, "n_element_tab_1": 22, "formula_len_all": 454, "formula_len_all_1": 464, "len_all": 206185, "len_all_1": 62504, "len_abs": 2681, "len_title": 139, "len_sents": 56759, "len_sents_1": 29496, "n_sents": 424, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1410, "title": "ROSE: Register-Assisted General Time Series Forecasting with Decomposed Frequency Learning", "abs": "With the increasing collection of time series data from various domains, there arises a strong demand for general time series forecasting models pre-trained on a large number of time-series datasets to support a variety of downstream prediction tasks. Enabling general time series forecasting faces two challenges: how to obtain unified representations from multi-domian time series data, and how to capture domain-specific features from time series data across various domains for adaptive transfer in downstream tasks. To address these challenges, we propose a Register-Assisted General Time Series Forecasting Model with Decomposed Frequency Learning (ROSE), a novel pre-trained model for time series forecasting. ROSE employs Decomposed Frequency Learning for the pre-training task, which decomposes coupled semantic information in time series with frequency-based masking and reconstruction to obtain unified representations across domains. We also equip ROSE with a Time Series Register, which learns to generate a register to capture domain-specific representations during pre-training and enhances domain-adaptive transfer by selecting related register tokens on downstream tasks. After pre-training on large-scale time series data, ROSE achieves state-of-the-art forecasting performance on 7 real-world benchmarks. Remarkably, it demonstrates competitive or superior few-shot and zero-shot abilities.", "keywords": [ "Time series forecasting" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "tdfHABLdxR", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Alexander Luis Mitchell", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Ioannis Havoutis", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" }, { "name": "Mark Nicholas Finean", "gender": "Male", "institution": "Dyson", "country": "", "position": "Researcher" }, { "name": "Vassil Atanassov", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Wanming Yu", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 48, "n_ref": 129, "n_ref_all": 154, "n_fig": 16, "n_tab": 1, "L_tab": 283, "n_element_tab": 32, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 135, "n_element_tab_1": 36, "formula_len_all": 918, "formula_len_all_1": 784, "len_all": 147153, "len_all_1": 63093, "len_abs": 944, "len_title": 139, "len_sents": 45249, "len_sents_1": 31836, "n_sents": 289, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1050, "title": "Constrained Skill Discovery: Quadruped Locomotion with Unsupervised Reinforcement Learning", "abs": "Representation learning and unsupervised skill discovery can allow robots to acquire diverse and reusable behaviors without the need for task-specific rewards. In this work, we learn a latent representation by maximizing the mutual information between skills and states subject to a distance constraint, using unsupervised reinforcement learning. Our method improves upon prior constrained skill discovery methods by replacing the latent transition maximization with a norm-matching objective. This not only results in a much a richer state space coverage, but allows the robot to learn more stable and easily controllable locomotive behaviors. In robotics this is particularly important, because state transition-maximizing behaviors can result in highly dangerous motions. We successfully deployed the learned policy on a real ANYmal quadruped robot and demonstrated that the robot can accurately reach arbitrary points of the Cartesian state space in a zero-shot manner, using only an intrinsic skill discovery and standard regularization rewards.", "keywords": [ "legged locomotion", "unsupervised skill discovery", "empowerment", "unsupervised reinforcement learning" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "tdbK3TGFl1", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Ananda Theertha Suresh", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Chong You", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Felix Yu", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Rajesh Jayaram", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Sanjiv Kumar", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" } ], "n_formula": 10, "n_formula_1": 4, "n_ref_uni": 19, "n_ref": 24, "n_ref_all": 56, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 123, "n_element_tab_1": 8, "formula_len_all": 771, "formula_len_all_1": 448, "len_all": 92229, "len_all_1": 67350, "len_abs": 1459, "len_title": 158, "len_sents": 35271, "len_sents_1": 32324, "n_sents": 283, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1458, "title": "Asymmetric Embedding Models for Hierarchical Retrieval: Provable Constructions and a Pretrain-Finetune Recipe", "abs": "Dual encoder (DE) models, where a pair of matching query and document are embedded into similar vector representations, are widely used in information retrieval due to their efficiency and scalability. However, DEs are known to have a limited expressive power due to the Euclidean geometry of the embedding space, which may compromise their quality. This paper investigate such limitations in the context of \\emph{hierarchical retrieval}, the task where the document set has a hierarchical structure and the matching keywords for a query are all of its ancestor nodes. We first prove the feasibility of representing hierarchical structures within the Euclidean embedding space by providing a constructive algorithm for generating effective embeddings from a given hierarchy. Then we delve into the learning of DEs when the hierarchy is unknown, which is a practical assumption since usually only samples of matching query and document pairs are available during training. Our experiments reveal a \"lost in the long distance\" phenomenon, where retrieval accuracy degrades for documents further away in the hierarchy. To address this, we introduce a pretrain-finetune approach that significantly improves long-distance retrieval without sacrificing performance on closer documents. Finally, we validate our findings on a realistic hierarchy from WordNet, demonstrating the effectiveness of our approach in retrieving documents at various levels of abstraction.", "keywords": [ "Information Retrieval", "Hierarchical Retrieval", "Dual encoders" ], "rating_list": [ 3, 5, 5, 1 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "tdV1GRkCpZ", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "John Alexander McDermid", "gender": "Male", "institution": "University of York", "country": "", "position": "Full Professor" }, { "name": "Nick E. Pears", "gender": "Male", "institution": "University of York", "country": "", "position": "Full Professor" }, { "name": "Tejas Pandey", "gender": "Male", "institution": "University of York", "country": "GB", "position": "PhD student" }, { "name": "William Smith", "gender": "Male", "institution": "University of York", "country": "GB", "position": "Full Professor" } ], "n_formula": 26, "n_formula_1": 25, "n_ref_uni": 36, "n_ref": 71, "n_ref_all": 80, "n_fig": 3, "n_tab": 7, "L_tab": 1693, "n_element_tab": 307, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 2576, "n_element_tab_1": 269, "formula_len_all": 1080, "formula_len_all_1": 1100, "len_all": 114872, "len_all_1": 61493, "len_abs": 1306, "len_title": 91, "len_sents": 34802, "len_sents_1": 28121, "n_sents": 231, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1310, "title": "E-DETR: Evidential Deep Learning for End-to-End Uncertainty Estimation in Object Detection", "abs": "Detection transformers (DETR) have emerged as powerful end-to-end learning frameworks for object detection, directly regressing detection parameters as point estimates. However, these networks often lack the ability to express any uncertainty within their estimates. In this work, we replace the regression of point estimates with the direct learning of the posterior distribution in a sampling-free manner by leveraging deep evidential learning, complementing the end-to-end DETR architecture. We present an instance-aware uncertainty framework by extending evidential deep learning with an IoU-aware loss, jointly modelling both classification and localization uncertainties. Furthermore, we enable the model to leverage its uncertainty for self-calibration, aligning the predicted probabilities with the true likelihood of outcomes, and effectively apply evidential deep learning for the task of imbalanced dense object detection. Our approach is easily extensible and requires only fine-tuning, thus leveraging the pre-training of transformers on large datasets. We conduct extensive experiments on two in-domain and three out-of-domain datasets, demonstrating impressive improvements in generalization performance, especially when fine-tuning on heavily imbalanced datasets characterized by data scarcity.", "keywords": [ "object detection", "uncertainty estimation", "interpretability" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "td5nvlhJdc", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Kaiyu He", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Mian Zhang", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Peilin Wu", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Shuo Yan", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Zhiyu Chen", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Postdoc" } ], "n_formula": 4, "n_formula_1": 0, "n_ref_uni": 27, "n_ref": 45, "n_ref_all": 66, "n_fig": 10, "n_tab": 5, "L_tab": 1686, "n_element_tab": 156, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 95, "formula_len_all_1": 0, "len_all": 284484, "len_all_1": 72617, "len_abs": 2292, "len_title": 164, "len_sents": 73461, "len_sents_1": 35274, "n_sents": 563, "n_sents_1": 282, "venue": "ICLR2025", "decision": "withdraw", "L_title": 115, "L_abs": 1866, "title": "IDEA: Enhancing the Rule Learning Ability of Large Language Model Agent through Induction, Deduction, and Abduction", "abs": "While large language models (LLMs) have been thoroughly evaluated for deductive and inductive reasoning, their proficiency in abductive reasoning and holistic rule learning in interactive environments remains less explored. We introduce RULEARN, a novel benchmark specifically designed to assess the rule-learning abilities of LLM agents in interactive settings. In RULEARN, agents strategically interact with simulated environments to gather observations, discern patterns, and solve complex problems. To enhance the rule-learning capabilities for LLM agents, we propose IDEA, a novel reasoning framework that integrates the process of Induction, Deduction, and Abduction. The IDEA agent generates initial hypotheses from limited observations through abduction, devises plans to validate these hypotheses or leverages them to solve problems via deduction, and refines previous hypotheses using patterns identified from new observations through induction, dynamically establishing and applying rules that mimic human rule-learning behaviors. Our evaluation of the IDEA framework, which involves five representative LLMs, demonstrates significant improvements over the baseline. Furthermore, within this framework, our comparison with 50 human participants reveals notable discrepancies in rule-learning behaviors. LLM agents tend to generate plausible initial hypotheses but struggle to refine them through interaction. Conversely, humans, despite sometimes overlooking initial details, excel at incorporating feedback and continuously improving their hypotheses. We believe our benchmark, RULEARN, will serve as a valuable and challenging resource, and that the IDEA framework will provide crucial insights for the development of LLM agents capable of human-like rule learning in real-world scenarios. We will release our code and data upon acceptance of the paper.", "keywords": [ "Artificial Intelligence", "Reasoning", "Interactive Learning Environment", "Large Language Model Agent" ], "rating_list": [ 6, 6, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 5 ] }, { "paper_id": "tcvMzR2NrP", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Anuroop Sriram", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Principal Researcher" }, { "name": "Brian Karrer", "gender": "Male", "institution": "Meta Fundamental AI Research (FAIR)", "country": "US", "position": "Researcher" }, { "name": "Daniel Severo", "gender": "Male", "institution": "Facebook", "country": "CA", "position": "Researcher" }, { "name": "Itai Gat", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Marton Havasi", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Neta Shaul", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "", "position": "PhD student" }, { "name": "Peter Holderrieth", "gender": "Male", "institution": "MIT", "country": "US", "position": "PhD student" }, { "name": "Yaron Lipman", "gender": "Male", "institution": "FAIR at Meta", "country": "IL", "position": "Researcher" }, { "name": "Tian Qi Chen", "gender": "Male", "institution": "FAIR Labs, Meta AI", "country": "US", "position": "Researcher" } ], "n_formula": 89, "n_formula_1": 38, "n_ref_uni": 54, "n_ref": 116, "n_ref_all": 166, "n_fig": 5, "n_tab": 7, "L_tab": 2061, "n_element_tab": 217, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 1472, "n_element_tab_1": 185, "formula_len_all": 8834, "formula_len_all_1": 3013, "len_all": 212968, "len_all_1": 73767, "len_abs": 3360, "len_title": 120, "len_sents": 55616, "len_sents_1": 29628, "n_sents": 480, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 72, "L_abs": 1302, "title": "Flow Matching with General Discrete Paths: A Kinetic-Optimal Perspective", "abs": "The design space of discrete-space diffusion or flow generative models are significantly less well-understood than their continuous-space counterparts, with many works focusing only on a simple masked construction.\nIn this work, we aim to take a holistic approach to the construction of discrete generative models based on continuous-time Markov chains, and for the first time, allow the use of arbitrary discrete probability paths, or colloquially, corruption processes. \nThrough the lens of optimizing the symmetric kinetic energy, we propose velocity formulas that can be applied to any given probability path, completely decoupling the probability and velocity, and giving the user the freedom to specify any desirable probability path based on expert knowledge specific to the data domain. \nFurthermore, we find that a special construction of mixture probability paths optimizes the symmetric kinetic energy for the discrete case.\nWe empirically validate the usefulness of this new design space across multiple modalities: text generation, inorganic material generation, and image generation. We find that we can outperform the mask construction even in text with kinetic-optimal mixture paths, while we can make use of domain-specific constructions of the probability path over the visual domain.", "keywords": [ "flow matching", "discrete generative modeling" ], "rating_list": [ 6, 10, 8, 6, 8 ], "soundness_list": [ 3, 4, 4, 3, 3 ], "presentation_list": [ 3, 4, 4, 3, 3 ], "contribution_list": [ 3, 4, 4, 3, 3 ], "confidence_list": [ 3, 2, 2, 3, 2 ] }, { "paper_id": "tcsZt9ZNKD", "primary_area": "interpretability and explainable AI", "n_author": 9, "author_info_list": [ { "name": "Alec Radford", "gender": "unknown", "institution": "OpenAI", "country": "", "position": "Research" }, { "name": "Gabriel Goh", "gender": "Male", "institution": "University of California, Davis", "country": "", "position": "PhD student" }, { "name": "Henk Tillman", "gender": "unknown", "institution": "OpenAI", "country": "", "position": "Researcher" }, { "name": "Ilya Sutskever", "gender": "Male", "institution": "OpenAI", "country": "", "position": "Researcher" }, { "name": "Jan Leike", "gender": "Male", "institution": "OpenAI", "country": "", "position": "Alignment Team Lead" }, { "name": "Jeffrey Wu", "gender": "Male", "institution": "OpenAI", "country": "", "position": "Researcher" }, { "name": "Leo Gao", "gender": "unknown", "institution": "EleutherAI", "country": "", "position": "Researcher" }, { "name": "Rajan Troll", "gender": "Male", "institution": "InBalance", "country": "", "position": "Principal Researcher" }, { "name": "Tom Dupre la Tour", "gender": "Male", "institution": "OpenAI", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 49, "n_ref": 99, "n_ref_all": 166, "n_fig": 42, "n_tab": 2, "L_tab": 1309, "n_element_tab": 116, "n_fig_1": 26, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 351, "formula_len_all_1": 655, "len_all": 206545, "len_all_1": 60369, "len_abs": 1299, "len_title": 90, "len_sents": 62132, "len_sents_1": 26696, "n_sents": 501, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 42, "L_abs": 1317, "title": "Scaling and evaluating sparse autoencoders", "abs": "Sparse autoencoders provide a promising unsupervised approach for extracting interpretable features from a language model by reconstructing activations from a sparse bottleneck layer. Since language models learn many concepts, autoencoders need to be very large to recover all relevant features. However, studying the properties of autoencoder scaling is difficult due to the need to balance reconstruction and sparsity objectives and the presence of dead latents. We propose using k-sparse autoencoders [Makhzani and Frey, 2013] to directly control sparsity, simplifying tuning and improving the reconstruction-sparsity frontier. Additionally, we find modifications that result in few dead latents, even at the largest scales we tried. Using these techniques, we find clean scaling laws with respect to autoencoder size and sparsity. We also introduce several new metrics for evaluating feature quality based on the recovery of hypothesized features, the explainability of activation patterns, and the sparsity of downstream effects. These metrics all generally improve with autoencoder size. To demonstrate the scalability of our approach, we train a 16 million latent autoencoder on GPT-4 activations for 40 billion tokens. We release training code and autoencoders for open-source models, as well as a visualizer.", "keywords": [ "interpretability", "sparse autoencoders", "superposition", "scaling laws" ], "rating_list": [ 10, 8, 10, 10, 3 ], "soundness_list": [ 4, 3, 4, 4, 3 ], "presentation_list": [ 3, 4, 3, 4, 3 ], "contribution_list": [ 4, 4, 4, 4, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "tcq7n0m7Ml", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Xinzhu Ma", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Postdoc" }, { "name": "Ye Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Yingxin Li", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yuan Meng", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhi Wang", "gender": "Male", "institution": "SIGS, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Zihan Geng", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 12, "n_ref_uni": 36, "n_ref": 62, "n_ref_all": 87, "n_fig": 10, "n_tab": 6, "L_tab": 4025, "n_element_tab": 233, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 627, "n_element_tab_1": 64, "formula_len_all": 603, "formula_len_all_1": 793, "len_all": 152384, "len_all_1": 58857, "len_abs": 1841, "len_title": 107, "len_sents": 38864, "len_sents_1": 26834, "n_sents": 331, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1852, "title": "EMS: Adaptive Evict-then-Merge Strategy for Head-wise KV Cache Compression Based on Global-Local Importance", "abs": "As large language models (LLMs) continue to advance, the demand for higher quality and faster processing of long contexts across various applications is growing. KV cache is widely adopted as it stores previously generated key and value tokens, effectively reducing redundant computations during inference. However, as memory overhead becomes a significant concern, efficient compression of KV cache has gained increasing attention. Most existing methods perform compression from two perspectives: identifying important tokens and designing compression strategies. However, these approaches often produce biased distributions of important tokens due to the influence of accumulated attention scores or positional encoding. Furthermore, they overlook the sparsity and redundancy across different heads, which leads to difficulties in preserving the most effective information at the head level. To this end, we propose EMS to overcome these limitations, while achieving better KV cache compression under extreme compression ratios. Specifically, we introduce a Global-Local score that combines accumulated attention scores from both global and local KV tokens to better identify the token importance. For the compression strategy, we design an adaptive and unified Evict-then-Merge framework that accounts for the sparsity and redundancy of KV tokens across different heads. Additionally, we implement the head-wise parallel compression through a zero-class mechanism to enhance efficiency. Extensive experiments demonstrate our SOTA performance even under extreme compression ratios. EMS consistently achieves the lowest perplexity, improves scores by over 1.28 points across four LLMs on LongBench under a 256 cache budget, and preserves 95\\% retrieval accuracy with a cache budget less than 2\\% of the context length in the Needle-in-a-Haystack task.", "keywords": [ "Large Language Model", "Efficient Inference", "KV Cache" ], "rating_list": [ 5, 5, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3, 2 ], "contribution_list": [ 1, 1, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "tciQfO8S8j", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Chen Xu", "gender": "unknown", "institution": "Beijing University of Technology", "country": "", "position": "PhD student" }, { "name": "Chengqi Lyu", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Heyan Huang", "gender": "Female", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Kai Chen", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Tian Lan", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Wenwei Zhang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xian-Ling Mao", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "ShuaibinLi", "gender": "unknown", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 55, "n_ref": 158, "n_ref_all": 202, "n_fig": 8, "n_tab": 18, "L_tab": 4654, "n_element_tab": 262, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1107, "n_element_tab_1": 30, "formula_len_all": 304, "formula_len_all_1": 136, "len_all": 269581, "len_all_1": 54851, "len_abs": 1815, "len_title": 110, "len_sents": 76005, "len_sents_1": 27595, "n_sents": 539, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1691, "title": "Training Language Models to Critique with Multi-Agent Feedback", "abs": "Critique ability, a meta-cognitive capability of humans, presents significant challenges for LLMs to improve.\nRecent works primarily rely on supervised fine-tuning (SFT) using critiques generated by a single LLM like GPT-4.\nHowever, these model-generated critiques often exhibit flaws due to the inherent complexity of the critique. \nConsequently, fine-tuning LLMs on such flawed critiques typically limits the model's performance and propagates these flaws into the learned model.\nTo overcome these challenges, this paper proposes a novel data generation pipeline, named MultiCritique, that improves the critique ability of LLMs by utilizing multi-agent feedback in both the SFT and reinforcement learning (RL) stages.\nFirst, our data generation pipeline aggregates high-quality critiques from multiple agents instead of a single model, with crucial information as input for simplifying the critique.\nFurthermore, our pipeline improves the preference accuracy of critique quality through multi-agent feedback, facilitating the effectiveness of RL in improving the critique ability of LLMs.\nBased on our proposed MultiCritique data generation pipeline, we construct the MultiCritiqueDataset for the SFT and RL fine-tuning stages. Extensive experimental results on two benchmarks demonstrate:\n1) the superior quality of our constructed SFT dataset compared to existing critique datasets;\n2) additional improvements to the critique ability of LLMs brought by the RL stage.\nNotably, our fine-tuned 7B model significantly surpasses other advanced 7B-13B open-source models, approaching the performance of advanced 70B LLMs and GPT-4.\nCodes, datasets and model weights will be publicly available.", "keywords": [ "Critique", "LLM", "preference-based RL" ], "rating_list": [ 5, 5, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 3, 3 ] }, { "paper_id": "tcdbBbHHPo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Hansi Yang", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "James Kwok", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Hong Kong University of Science and Technology", "country": "HK", "position": "Professor" }, { "name": "Runsheng Yu", "gender": "Not Specified", "institution": "ByteDance", "country": "HK", "position": "Researcher" }, { "name": "Xiaoqi Jiao", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Youzhi Zhang", "gender": "unknown", "institution": "Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "", "position": "Assistant Professor" }, { "name": "Wang Yong", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 0, "n_ref": 0, "n_ref_all": 1, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 1723, "len_all_1": 1723, "len_abs": 0, "len_title": 0, "len_sents": 29, "len_sents_1": 29, "n_sents": 2, "n_sents_1": 2, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 28, "title": "Direct Alignment of Language Models via Quality-Aware Self-Refinement", "abs": "We have withdrawn our paper.", "keywords": [ "Reinforcement Learning; Language Model" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "tccML2tDd4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 1, "author_info_list": [ { "name": "Ashutosh Kumar", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 20, "n_ref": 41, "n_ref_all": 46, "n_fig": 2, "n_tab": 4, "L_tab": 899, "n_element_tab": 63, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 899, "n_element_tab_1": 63, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 86222, "len_all_1": 53181, "len_abs": 1433, "len_title": 137, "len_sents": 28788, "len_sents_1": 26920, "n_sents": 175, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1389, "title": "Perceptual Piercing: Human Visual Cue-Based Object Detection in Low Visibility Conditions", "abs": "This study proposes a novel deep learning framework inspired by atmospheric scattering and human visual cortex mechanisms to enhance object detection under poor visibility scenarios such as fog, smoke, and haze. These conditions pose significant challenges for object recognition, impacting various sectors, including autonomous driving, aviation management, and security systems. The objective is to enhance the precision and reliability of detection systems under adverse environmental conditions. The research investigates the integration of human-like visual cues, particularly focusing on selective attention and environmental adaptability, to ascertain their impact on object detection's computational efficiency and accuracy. This paper proposes a multi-tiered strategy that integrates an initial quick detection process, followed by targeted region-specific dehazing, and concludes with an in-depth detection phase. The approach is validated using the Foggy Cityscapes, RESIDE-beta (OTS and RTTS) datasets and is anticipated to set new performance standards in detection accuracy while significantly optimizing computational efficiency. The findings offer a viable solution for enhancing object detection in poor visibility and contribute to the broader understanding of integrating human visual principles into deep learning algorithms for intricate visual recognition challenges.", "keywords": [ "deep learning", "computer vision", "dehazing", "bio-inspired networks", "human visual perception" ], "rating_list": [ 3, 1, 3 ], "soundness_list": [ 1, 1, 2 ], "presentation_list": [ 3, 1, 2 ], "contribution_list": [ 1, 1, 1 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "tc90LV0yRL", "primary_area": "datasets and benchmarks", "n_author": 25, "author_info_list": [ { "name": "Andy K Zhang", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Researcher" }, { "name": "Aolin Zhang", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Ari Glenn", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Celeste Menders", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Dan Boneh", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Daniel E. Ho", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Professor" }, { "name": "Daniel Zamoshchin", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Derek Askaryar", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Donovan Julian Jasper", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Eliot Jones", "gender": "Male", "institution": "Gray Swan", "country": "US", "position": "Researcher" }, { "name": "Gashon Hussein", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Haoxiang Yang", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "MS student" }, { "name": "Joey Ji", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Justin W Lin", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Researcher" }, { "name": "Kenny O Oseleononmen", "gender": "unknown", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "MS student" }, { "name": "Leo Glikbarg", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Nathan Tran", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Neil Perry", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Percy Liang", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Pura Peetathawatchai", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "MS student" }, { "name": "Rinnara Sangpisit", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Rishi Alluri", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Riya Dulepet", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Samantha Liu", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Vikram Sivashankar", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Undergrad student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 27, "n_ref": 106, "n_ref_all": 134, "n_fig": 6, "n_tab": 3, "L_tab": 918, "n_element_tab": 116, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 95, "formula_len_all_1": 23, "len_all": 194277, "len_all_1": 60376, "len_abs": 1731, "len_title": 140, "len_sents": 40110, "len_sents_1": 29926, "n_sents": 286, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 91, "L_abs": 1917, "title": "Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models", "abs": "Language Model (LM) agents for cybersecurity that are capable of autonomously identifying vulnerabilities and executing exploits have potential to cause real-world impact. Policymakers, model providers, and researchers in the AI and cybersecurity communities are interested in quantifying the capabilities of such agents to help mitigate cyberrisk and investigate opportunities for penetration testing. Toward that end, we introduce Cybench, a framework for specifying cybersecurity tasks and evaluating agents on those tasks. We include 40 professional-level Capture the Flag (CTF) tasks from 4 distinct CTF competitions, chosen to be recent, meaningful, and spanning a wide range of difficulties. Each task includes its own description, starter files, and is initialized in an environment where an agent can execute commands and observe outputs. Since many tasks are beyond the capabilities of existing LM agents, we introduce subtasks for each task, which break down a task into intermediary steps for a more detailed evaluation. To evaluate agent capabilities, we construct a cybersecurity agent and evaluate 8 models: GPT-4o, OpenAI o1-preview, Claude 3 Opus, Claude 3.5 Sonnet, Mixtral 8x22b Instruct, Gemini 1.5 Pro, Llama 3 70B Chat, and Llama 3.1 405B Instruct. For the top performing models (GPT-4o and Claude 3.5 Sonnet), we further investigate performance across 4 agent scaffolds (structured bash, action-only, pseudoterminal, and web search). Without subtask guidance, agents leveraging Claude 3.5 Sonnet, GPT-4o, OpenAI o1-preview, and Claude 3 Opus successfully solved complete tasks that took human teams up to 11 minutes to solve. In comparison, the most difficult task took human teams 24 hours and 54 minutes to solve. Anonymized code and data are available at https://drive.google.com/file/d/1kp3H0pw1WMAH-Qyyn9WA0ZKmEa7Cr4D4 and https://drive.google.com/file/d/1BcTQ02BBR0m5LYTiK-tQmIK17_TxijIy.", "keywords": [ "Language Model Agents", "Benchmark", "Cybersecurity", "Risk" ], "rating_list": [ 10, 8, 8 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 4, 4, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "tbx3u2oZAu", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Huawei Shen", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Liang Pang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "Associate Professor" }, { "name": "Shicheng Xu", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Xueqi Cheng", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy", "country": "CN", "position": "Full Professor" } ], "n_formula": 85, "n_formula_1": 18, "n_ref_uni": 45, "n_ref": 123, "n_ref_all": 166, "n_fig": 6, "n_tab": 7, "L_tab": 3480, "n_element_tab": 280, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1485, "n_element_tab_1": 97, "formula_len_all": 10089, "formula_len_all_1": 2140, "len_all": 209662, "len_all_1": 74317, "len_abs": 1667, "len_title": 120, "len_sents": 63449, "len_sents_1": 34460, "n_sents": 502, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1671, "title": "A Theory for Token-Level Harmonization in Retrieval-Augmented Generation", "abs": "Retrieval-augmented generation (RAG) utilizes retrieved texts to enhance large language models (LLMs). Studies show that while RAG provides valuable external information (benefit), it may also mislead LLMs (detriment) with noisy or incorrect retrieved texts. Although many existing methods attempt to preserve benefit and avoid detriment, they lack a theoretical explanation for RAG. The benefit and detriment in the next token prediction of RAG remain a 'black box' that cannot be quantified or compared in an explainable manner, so existing methods are data-driven, need additional utility evaluators or post-hoc. This paper takes the first step towards providing a theory to explain and trade off the benefit and detriment in RAG. We model RAG as the fusion between distributions of LLMs\u2019 knowledge and distributions of retrieved texts. Then, we formalize the trade-off between the value of external knowledge (benefit) and its potential risk of misleading LLMs (detriment) in next token prediction of RAG by distribution difference in this fusion. Finally, we prove that the actual effect of RAG on the token, which is the comparison between benefit and detriment, can be predicted without any training or accessing the utility of retrieval. Based on our theory, we propose a practical novel method, Tok-RAG, which achieves collaborative generation between the pure LLM and RAG at token level to preserve benefit and avoid detriment. Experiments in real-world tasks using LLMs such as OPT, LLaMA-2, and Mistral show the effectiveness of our method and support our theoretical findings. Code is in supplemental material and will be released on GitHub after acceptance.", "keywords": [ "Retrieval-augmented generation", "Duality", "Large Language Models" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "taHwqSrbrb", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Fan Wang", "gender": "Female", "institution": "Alibaba Group", "country": "US", "position": "Senior Staff Algorithm Engineer" }, { "name": "Gao Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jiasheng Tang", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Kai Wang", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" }, { "name": "Yibing Song", "gender": "unknown", "institution": "Alibaba DAMO Academy", "country": "CN", "position": "Staff Scientist" }, { "name": "Yizeng Han", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "ZHAO WANGBO", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 61, "n_ref": 175, "n_ref_all": 218, "n_fig": 10, "n_tab": 9, "L_tab": 3784, "n_element_tab": 662, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 2181, "n_element_tab_1": 415, "formula_len_all": 721, "formula_len_all_1": 748, "len_all": 204817, "len_all_1": 72252, "len_abs": 1108, "len_title": 77, "len_sents": 51029, "len_sents_1": 30155, "n_sents": 446, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 29, "L_abs": 1118, "title": "Dynamic Diffusion Transformer", "abs": "Diffusion Transformer (DiT), an emerging diffusion model for image generation,\nhas demonstrated superior performance but suffers from substantial computational\ncosts. Our investigations reveal that these costs stem from the static inference\nparadigm, which inevitably introduces redundant computation in certain diffusion\ntimesteps and spatial regions. To address this inefficiency, we propose Dynamic\nDiffusion Transformer (DyDiT), an architecture that dynamically adjusts its compu-\ntation along both timestep and spatial dimensions during generation. Specifically,\nwe introduce a Timestep-wise Dynamic Width (TDW) approach that adapts model\nwidth conditioned on the generation timesteps. In addition, we design a Spatial-\nwise Dynamic Token (SDT) strategy to avoid redundant computation at unnecessary\nspatial locations. Extensive experiments on various datasets and different-sized\nmodels verify the superiority of DyDiT. Notably, with <3% additional fine-tuning it-\nerations, our method reduces the FLOPs of DiT-XL by 51%, accelerates generation\nby 1.73\u00d7, and achieves a competitive FID score of 2.07 on ImageNet.", "keywords": [ "Diffusion Transformer", "Dynamic Neural Network", "Efficiency" ], "rating_list": [ 3, 5, 8, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "tZozeR3VV7", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Chao Chen", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Associate Professor" }, { "name": "Haibin Ling", "gender": "Male", "institution": "State University of New York, Stony Brook", "country": "US", "position": "Professor" }, { "name": "Jiachen Yao", "gender": "Male", "institution": ", State University of New York at Stony Brook", "country": "", "position": "PhD student" }, { "name": "Lijie Hu", "gender": "Female", "institution": "KAUST", "country": "SA", "position": "PhD student" }, { "name": "Lingjie Yi", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Lu Pang", "gender": "unknown", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Saumya Gupta", "gender": "Female", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Tao Sun", "gender": "Male", "institution": "State University of New York, Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Weimin Lyu", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 35, "n_ref": 63, "n_ref_all": 86, "n_fig": 3, "n_tab": 9, "L_tab": 4137, "n_element_tab": 678, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1054, "n_element_tab_1": 85, "formula_len_all": 559, "formula_len_all_1": 582, "len_all": 146086, "len_all_1": 55324, "len_abs": 1250, "len_title": 112, "len_sents": 44590, "len_sents_1": 26104, "n_sents": 368, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1251, "title": "Backdooring Vision-Language Models with Out-Of-Distribution Data", "abs": "The emergence of Vision-Language Models (VLMs) represents a significant advancement in integrating computer vision with Large Language Models (LLMs) to generate detailed text descriptions from visual inputs. Despite their growing importance, the security of VLMs, particularly against backdoor attacks, is under explored. Moreover, prior works often assume attackers have access to the original training data, which is often unrealistic. In this paper, we address a more practical and challenging scenario where attackers must rely solely on Out-Of-Distribution (OOD) data. We introduce VLOOD (Backdoor Vision-Language Models using Out-of-Distribution Data), a novel approach with two key contributions: (1) demonstrating backdoor attacks on VLMs in complex image-to-text tasks while minimizing degradation of the original semantics under poisoned inputs, and (2) proposing innovative techniques for backdoor injection without requiring any access to the original training data. Our evaluation on image captioning and visual question answering (VQA) tasks confirms the effectiveness of VLOOD, revealing a critical security vulnerability in VLMs and laying the foundation for future research on securing multimodal models against sophisticated threats.", "keywords": [ "VLMs", "Backdoor Attack", "image-to-text generation" ], "rating_list": [ 8, 5, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "tZk3LnvVtK", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Adam Stein", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Alessandro Achille", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Research Scientist" }, { "name": "Benjamin Bowman", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Matthew Trager", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Stefano Soatto", "gender": "unknown", "institution": "Amazon Web Services", "country": "US", "position": "Vice President" } ], "n_formula": 21, "n_formula_1": 12, "n_ref_uni": 37, "n_ref": 85, "n_ref_all": 114, "n_fig": 7, "n_tab": 10, "L_tab": 4004, "n_element_tab": 447, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2140, "n_element_tab_1": 248, "formula_len_all": 1377, "formula_len_all_1": 1151, "len_all": 164754, "len_all_1": 70936, "len_abs": 853, "len_title": 107, "len_sents": 48881, "len_sents_1": 30983, "n_sents": 404, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 813, "title": "Measuring Language Model Uncertainty With Internal Concepts", "abs": "We study the problem of evaluating the predictive uncertainty of large language models (LLMs). \nWe assign an uncertainty measure to the correctness of outputs from an LLM conditioned on a query using a form of entropy that applies to semantic objects (concepts).\nUnlike prior works, the notion of meaning used to define concepts is derived from the LLM, rather than from\nan external model. \nOur method measures an uncertainty over concept structures by drawing from ideas in Formal Concept Analysis (FCA) and lattice/order theory, and can be used to estimate correctness in closed- and open-ended scenarios.\nOur method has a relative improvement of up to 4.8% on average across five standard benchmarks as well as improves over comparable baselines on datasets consisting of both closed- and open-ended questions.", "keywords": [ "uncertainty", "correctness", "large language models", "concepts" ], "rating_list": [ 5, 5, 6, 6, 6 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 4 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 2, 2, 3 ] }, { "paper_id": "tZiMLgsHMu", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Shafiq Joty", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Principal Researcher" }, { "name": "Tong Niu", "gender": "Male", "institution": "Salesforce AI Research", "country": "", "position": "Research Scientist" }, { "name": "Ye Liu", "gender": "Female", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "Yingbo Zhou", "gender": "unknown", "institution": "Salesforce Research", "country": "", "position": "Research Scientist" }, { "name": "Semih Yavuz", "gender": "Male", "institution": "SalesForce.com", "country": "", "position": "Research Scientist" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 37, "n_ref": 61, "n_ref_all": 88, "n_fig": 6, "n_tab": 2, "L_tab": 1275, "n_element_tab": 164, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1760, "n_element_tab_1": 247, "formula_len_all": 232, "formula_len_all_1": 232, "len_all": 127820, "len_all_1": 67955, "len_abs": 1714, "len_title": 126, "len_sents": 33645, "len_sents_1": 29331, "n_sents": 262, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1364, "title": "JudgeRank: Leveraging Large Language Models for Reasoning-Intensive Reranking", "abs": "Accurate document retrieval is crucial for the success of retrieval-augmented generation (RAG) applications, including open-domain question answering and code completion. While large language models (LLMs) have been employed as dense encoders or listwise rerankers in RAG systems, they often struggle with reasoning-intensive tasks because they lack nuanced analysis when judging document relevance. To address this limitation, we introduce JudgeRank, a novel agentic reranker that emulates human cognitive processes when assessing document relevance. Our approach consists of three key steps: (1) query analysis to identify the core problem, (2) document analysis to extract a query-aware summary, and (3) relevance judgment to provide a concise assessment of document relevance. We evaluate JudgeRank on the reasoning-intensive BRIGHT benchmark, demonstrating substantial performance improvements over first-stage retrieval methods and outperforming other popular reranking approaches. In addition, JudgeRank performs on par with fine-tuned state-of-the-art rerankers on the popular BEIR benchmark, validating its zero-shot generalization capability. Through comprehensive ablation studies, we demonstrate that JudgeRank's performance generalizes well across LLMs of various sizes while ensembling them yields even more accurate reranking than individual models.", "keywords": [ "reranking", "retrieval", "generation" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "tZdqL5FH7w", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Anh Tuan Bui", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Postdoc" }, { "name": "Junae Kim", "gender": "Not Specified", "institution": "Defence Science and Technology Group", "country": "", "position": "Researcher" }, { "name": "Long Tung Vuong", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Paul Montague", "gender": "Male", "institution": "Defence Science and Technology Group", "country": "AU", "position": "Researcher" }, { "name": "Tamas Abraham", "gender": "Male", "institution": "DSTG", "country": "AU", "position": "Researcher" }, { "name": "Trang Vu", "gender": "unknown", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" }, { "name": "Trung Le", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" }, { "name": "Dinh Phung", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 5, "n_ref_uni": 24, "n_ref": 81, "n_ref_all": 139, "n_fig": 29, "n_tab": 8, "L_tab": 1651, "n_element_tab": 212, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 485, "n_element_tab_1": 75, "formula_len_all": 904, "formula_len_all_1": 487, "len_all": 182980, "len_all_1": 63058, "len_abs": 1167, "len_title": 127, "len_sents": 74743, "len_sents_1": 31371, "n_sents": 553, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1255, "title": "Fantastic Targets for Concept Erasure in Diffusion Models and Where To Find Them", "abs": "Concept erasure has emerged as a promising technique for mitigating the risk of harmful content generation in diffusion models by selectively unlearning undesirable concepts. The common principle of previous works to remove a specific concept is to map it to a fixed generic concept, such as a neutral concept or just an empty text prompt. In this paper, we demonstrate that this fixed-target strategy is suboptimal, as it fails to account for the impact of erasing one concept on the others. To address this limitation, we model the concept space as a graph and empirically analyze the effects of erasing one concept on the remaining concepts. Our analysis uncovers intriguing geometric properties of the concept space, where the influence of erasing a concept is confined to a local region. Building on this insight, we propose the Adaptive Guided Erasure (AGE) method, which dynamically selects optimal target concepts tailored to each undesirable concept, minimizing unintended side effects. Experimental results show that AGE significantly outperforms state-of-the-art erasure methods on preserving unrelated concepts while maintaining effective erasure performance. Our code is published at {https://github.com/tuananhbui89/Adaptive-Guided-Erasure}.", "keywords": [ "Trustworthy Generative AI", "Diffusion Models", "Machine Unlearning", "Erasing Concepts" ], "rating_list": [ 5, 6, 8 ], "soundness_list": [ 2, 3, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 3, 4 ], "confidence_list": [ 2, 3, 4 ] }, { "paper_id": "tZP4Uyql0r", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Jiaen Chen", "gender": "unknown", "institution": "Shihezi University", "country": "CN", "position": "MS student" }, { "name": "Xiaofangwang", "gender": "Female", "institution": "Shihezi University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuchen Zheng", "gender": "Male", "institution": "Shihezi University", "country": "CN", "position": "Associate Professor" }, { "name": "Zouquan Chen", "gender": "Male", "institution": "Shihezi University", "country": "CN", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 3, "n_ref": 7, "n_ref_all": 19, "n_fig": 8, "n_tab": 5, "L_tab": 2039, "n_element_tab": 138, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3404, "n_element_tab_1": 144, "formula_len_all": 1266, "formula_len_all_1": 794, "len_all": 91174, "len_all_1": 50747, "len_abs": 3757, "len_title": 125, "len_sents": 26027, "len_sents_1": 21144, "n_sents": 193, "n_sents_1": 149, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1078, "title": "FedDFQ : Personalized Federated Learning Based On Data Feature Quantification", "abs": "Personalized federated learning is widely used for heterogeneous data distributions across clients. However, existing methods are difficult to measure and utilize these heterogeneities accurately. To address this issue, in this paper, we propose a novel and efficient method named FedDFQ which uses a customized Data Identity Extraction Module (DIEM) to dynamically generate metric proxies that quantify data heterogeneity across different local clients in a privacy-friendly manner. The metric proxies are used to assess the contributions of global parameter aggregation and personalized gradient backpropagation for each local client. In addition, we design a plug-and-play Automatic Gradient Accumulation Module (AGAM) that regularizes personalized classification layers with re-balanced gradients. We provide theoretical explanations and experimental results that validate the effectiveness of the proposed FedDFQ. With comprehensive comparisons to existing state-of-the-art approaches, FedDFQ outperforms them on two benchmark datasets in different heterogeneous scenarios.", "keywords": [ "Federal Learning", "Model Aggregation", "Data Feature Utilization" ], "rating_list": [ 3, 3, 1 ], "soundness_list": [ 2, 2, 1 ], "presentation_list": [ 3, 1, 1 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "tZDhrhUOcs", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Hirofumi Inaguma", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Ning Dong", "gender": "Male", "institution": "Meta Inc.", "country": "", "position": "Researcher" }, { "name": "Paden Tomasello", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "Undergrad student" }, { "name": "Weiting Tan", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Xutai Ma", "gender": "Male", "institution": "Meta FAIR", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 8, "n_ref_uni": 44, "n_ref": 126, "n_ref_all": 151, "n_fig": 7, "n_tab": 7, "L_tab": 3571, "n_element_tab": 647, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 2173, "n_element_tab_1": 313, "formula_len_all": 255, "formula_len_all_1": 338, "len_all": 217100, "len_all_1": 66106, "len_abs": 782, "len_title": 114, "len_sents": 36741, "len_sents_1": 29093, "n_sents": 246, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 788, "title": "SSR: Alignment-Aware Modality Connector for Speech Language Models", "abs": "Fusing speech into pre-trained language model (SpeechLM) usually suffers from inefficient encoding of long-form speech and catastrophic forgetting of pre-trained text modality. We propose SSR-connector (Segmented Speech Representation Connector) for better modality fusion. Leveraging speech-text alignments, our approach segments and compresses speech features to match the granularity of text embeddings. Additionally, we introduce a two-stage training pipeline that includes the distillation and fine-tuning phases to mitigate catastrophic forgetting. SSR-connector outperforms existing mechanism for speech-text modality fusion, consistently achieving better speech understanding (e.g., $+10$ accuracy on StoryCloze and $+20$ on Speech-MMLU) while preserving pre-trained text ability.", "keywords": [ "modality fusion", "speech language model", "spoken language understanding" ], "rating_list": [ 6, 5, 3, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "tZCqSVncRf", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Jiachun Li", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "PhD student" }, { "name": "Jun Zhao", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "", "position": "Full Professor" }, { "name": "Kang Liu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Professor" }, { "name": "Pengfei Cao", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Yubo Chen", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "", "position": "Associate Professor" }, { "name": "Zhuoran Jin", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "", "position": "PhD student" } ], "n_formula": 53, "n_formula_1": 11, "n_ref_uni": 16, "n_ref": 85, "n_ref_all": 128, "n_fig": 12, "n_tab": 12, "L_tab": 1499, "n_element_tab": 180, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 1007, "n_element_tab_1": 66, "formula_len_all": 2794, "formula_len_all_1": 724, "len_all": 195163, "len_all_1": 71276, "len_abs": 1366, "len_title": 128, "len_sents": 52446, "len_sents_1": 30888, "n_sents": 524, "n_sents_1": 273, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1384, "title": "MIRAGE: Evaluating and Explaining Inductive Reasoning Process in Language Models", "abs": "Inductive reasoning is an essential capability for large language models (LLMs) to achieve higher intelligence, which requires the model to generalize rules from observed facts and then apply them to unseen examples. We present {\\scshape Mirage}, a synthetic dataset that addresses the limitations of previous work, specifically the lack of comprehensive evaluation and flexible test data. In it, we evaluate LLMs' capabilities in both the inductive and deductive stages, allowing for flexible variation in input distribution, task scenario, and task difficulty to analyze the factors influencing LLMs' inductive reasoning. Based on these multi-faceted evaluations, we demonstrate that the LLM is a poor rule-based reasoner. In many cases, when conducting inductive reasoning, they do not rely on a correct rule to answer the unseen case. From the perspectives of different prompting methods, observation numbers, and task forms, models tend to consistently conduct correct deduction without correct inductive rules. Besides, we find that LLMs are good neighbor-based reasoners. In the inductive reasoning process, the model tends to focus on observed facts that are close to the current test example in feature space. By leveraging these similar examples, the model maintains strong inductive capabilities within a localized region, significantly improving its deductive performance.", "keywords": [ "inductive reasoning", "large language model", "model explanation" ], "rating_list": [ 5, 8, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 5, 4, 3 ] }, { "paper_id": "tYuVjFgEIK", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Yuntian Chen", "gender": "Male", "institution": "Eastern Institute for Advanced Study", "country": "CN", "position": "Assistant Professor" }, { "name": "zhou han", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 28, "n_ref": 74, "n_ref_all": 95, "n_fig": 9, "n_tab": 10, "L_tab": 16985, "n_element_tab": 1077, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1128, "n_element_tab_1": 81, "formula_len_all": 403, "formula_len_all_1": 409, "len_all": 165619, "len_all_1": 60089, "len_abs": 1517, "len_title": 104, "len_sents": 49435, "len_sents_1": 27399, "n_sents": 320, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1433, "title": "Decoupling Variable and Temporal Dependencies: A Novel Approach for Multivariate Time Series Forecasting", "abs": "In multivariate time series forecasting using the Transformer architecture, capturing temporal dependencies and modeling inter-variable relationships are crucial for improving performance. However, overemphasizing temporal dependencies can destabilize the model, increasing its sensitivity to noise, overfitting, and weakening its ability to capture inter-variable relationships. We propose a new approach called the Temporal-Variable Decoupling Network (TVDN) to address this challenge. This method decouples the modeling of variable dependencies from temporal dependencies and further separates temporal dependencies into historical and predictive sequence dependencies, allowing for a more effective capture of both. Specifically, the simultaneous learning of time-related and variable-related patterns can lead to harmful interference between the two. TVDN first extracts variable dependencies from historical data through a permutation-invariant model and then captures temporal dependencies using a permutation-equivariant model. By decoupling variable and temporal dependencies and historical and predictive sequence dependencies, this approach minimizes interference and allows for complementary extraction of both. Our method provides a concise and innovative approach to enhancing the utilization of temporal features. Experiments on multiple real-world datasets demonstrate that TVDN achieves state-of-the-art performance.", "keywords": [ "Time Series Forecasting", "transformer", "multivariate time series forecasting" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "tYVmxoRps3", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Chen Yu", "gender": "Not Specified", "institution": "Independent Researcher", "country": "SG", "position": "Researcher" }, { "name": "WANG PEIXU", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Yu Ming", "gender": "Male", "institution": "Power Automation pte ltd", "country": "SG", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 28, "n_ref": 43, "n_ref_all": 66, "n_fig": 16, "n_tab": 4, "L_tab": 1329, "n_element_tab": 137, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 540, "n_element_tab_1": 58, "formula_len_all": 1280, "formula_len_all_1": 1324, "len_all": 106999, "len_all_1": 51206, "len_abs": 1011, "len_title": 122, "len_sents": 33382, "len_sents_1": 22306, "n_sents": 263, "n_sents_1": 170, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1022, "title": "IS TRANSFORMER A STOCHASTIC PARROT? A CASE STUDY IN SIMPLE ARITHMETIC TASK", "abs": "Large pre-trained language models have demonstrated impressive capabilities, but \nthere is still much to learn about how they operate. In this study, we\n conduct a investigation of the autoregressive transformer\u2019s ability to\n perform basic addition operations. \n Specifically, by using causal analysis we found that a few different attention heads in the middle layers control the addition carry, with each head processing carries of different lengths. Due to the lack of globality in these attention heads, the model struggles to handle long-sequence addition tasks. By performing inference intervention on mistral-7B, partial task performance can be restored, with the accuracy on 20-digit long-sequence additions from 2\\% to 38\\%. Through\nfine-tuning, a new mechanism branches out for handling more complex cases, yet\nit still faces challenges with length generalization. Our research reveals how the\nmodel performs addition, and further provides insights into the debate on whether\nthese models are merely statistical.", "keywords": [ "large language model", "interpretability" ], "rating_list": [ 1, 6, 3, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "tXnAmayaio", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Jiuhai Chen", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "John Kirchenbauer", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Neel Jain", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Rifaa Qadri", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Tianyi Zhou", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Tom Goldstein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Yuxin Wen", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 0, "n_ref_uni": 35, "n_ref": 147, "n_ref_all": 175, "n_fig": 12, "n_tab": 6, "L_tab": 1312, "n_element_tab": 178, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1958, "n_element_tab_1": 71, "formula_len_all": 193, "formula_len_all_1": 0, "len_all": 216600, "len_all_1": 52333, "len_abs": 1915, "len_title": 116, "len_sents": 55070, "len_sents_1": 22524, "n_sents": 528, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 946, "title": "GenQA: An Instruction Dataset of LLM Generated Questions and Answers", "abs": "Most public instruction finetuning datasets are relatively small compared to the closed source datasets used to train industry models. To study questions about finetuning at scale, such as curricula and learning rate cooldown schedules, there is a need for industrial-scale datasets. However, this scale necessitates a data generation process that is almost entirely automated. In this work, we study methods for generating large instruction datasets from a single prompt. With little human oversight, we get LLMs to write diverse sets of instruction examples ranging from simple completion tasks to complex multi-turn dialogs across a variety of subject areas. When finetuning a Llama-3 8B base model, our dataset meets or exceeds both WizardLM and Ultrachat on both knowledge-intensive leaderboard tasks as well as conversational evaluations. We release our dataset, the \u201cgenerator\u201d prompts that created it, and our finetuned model checkpoints.", "keywords": [ "instruction finetuning", "large language models" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "tXUkT709OJ", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Pengkun Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" }, { "name": "Xu Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Researcher" }, { "name": "Xuan Yu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Yang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yudong Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zhaoyang Sun", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "chen zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 19, "n_ref_uni": 18, "n_ref": 46, "n_ref_all": 58, "n_fig": 6, "n_tab": 3, "L_tab": 581, "n_element_tab": 90, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1828, "n_element_tab_1": 79, "formula_len_all": 2025, "formula_len_all_1": 1991, "len_all": 123347, "len_all_1": 67213, "len_abs": 1726, "len_title": 134, "len_sents": 38507, "len_sents_1": 28983, "n_sents": 322, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1780, "title": "COFlowNet: Conservative Constraints on Flows Enable High-Quality Candidate Generation", "abs": "Generative flow networks (GFlowNet) have been considered as powerful tools for generating candidates with desired properties. Given that evaluating the property of candidates can be complex and time-consuming, existing GFlowNets train proxy models for efficient online evaluation. However, the performance of proxy models is heavily dependent on the amount of data and is of considerable uncertainty. Therefore, it is of great interest that how to develop an offline GFlowNet that does not rely on online evaluation. \nUnder offline setting, the limited data results in insufficient exploration of state space. \nThe insufficient exploration means that offline GFlowNets can hardly generate satisfying candidates out of the distribution of training data. Therefore, it is critical to restrict the offline model to act in the distribution of training data. The distinctive training goal of GFlownets poses a unique challenge for making such restrictions.\nTackling the challenge, we propose Conservative Offline GFlowNet (COFlowNet) in this paper. We define unsupported flow, edges containing unseen states in training data. Models can learn extremely few knowledge about unsupported flow from training data. By constraining the model from exploring unsupported flows, we restrict COFlowNet to explore as optimal trajectories on the training set as possible, thus generating better candidates. In order to improve the diversity of candidates, we further introduce a quantile version of unsupported flow restriction. Experimental result on several widely-used datasets validates the effectiveness of COFlowNet in generating high-scored and diverse candidates. All implementations are available at \\href{https://github.com/yuxuan9982/COflownet}{https://github.com/yuxuan9982/COflownet}.", "keywords": [ "generative flow network", "offline RL", "molecule design" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 2, 4, 3 ] }, { "paper_id": "tWjLGWrtqy", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Fran\u00e7ois Yvon", "gender": "Male", "institution": "ISIR, Sorbonne Universit\u00e9 & CNRS", "country": "FR", "position": "Principal Researcher" }, { "name": "Matthieu Dubois", "gender": "unknown", "institution": "Universit\u00e9 Pierre et Marie Curie - Paris 6, Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "", "position": "PhD student" }, { "name": "Pablo Piantanida", "gender": "Male", "institution": "Universit\u00e9 Paris-Saclay, CNRS ", "country": "FR", "position": "Full Professor" } ], "n_formula": 26, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 90, "n_ref_all": 144, "n_fig": 7, "n_tab": 7, "L_tab": 1984, "n_element_tab": 136, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1207, "n_element_tab_1": 66, "formula_len_all": 2515, "formula_len_all_1": 863, "len_all": 189358, "len_all_1": 73164, "len_abs": 907, "len_title": 150, "len_sents": 46370, "len_sents_1": 32298, "n_sents": 364, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 908, "title": "MOSAIC: Multiple Observers Spotting AI Content, a Robust Approach to Machine-Generated Text Detection", "abs": "The dissemination of Large Language Models (LLMs), trained at scale, and endowed with powerful text-generating abilities has vastly increased the threats posed by generative AI technologies by reducing the cost of producing harmful, toxic, faked or forged content. In response, various proposals have been made to automatically discriminate artificially generated from human-written texts, typically framing the problem as a classification problem. Most approaches evaluate an input document by a well-chosen detector LLM, assuming that low-perplexity scores reliably signal machine-made content. As using one single detector can induce brittleness of performance, we instead consider several and derive a new, theoretically grounded approach to combine their respective strengths. Our experiments, using a variety of generator LLMs, suggest that our method effectively increases the robustness of detection.", "keywords": [ "Machine-generated text detection", "Mixture of models", "Optimal detector", "Robustness" ], "rating_list": [ 6, 5, 8, 5, 3 ], "soundness_list": [ 3, 2, 3, 4, 1 ], "presentation_list": [ 2, 2, 3, 3, 1 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "tWgmOFfcQ1", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Di Wu", "gender": "Female", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Associate Professor" }, { "name": "Yifan Dai", "gender": "unknown", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Yufeng Liu", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 63, "n_ref_all": 78, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 995, "formula_len_all_1": 516, "len_all": 113720, "len_all_1": 59136, "len_abs": 1292, "len_title": 118, "len_sents": 40102, "len_sents_1": 30139, "n_sents": 271, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1296, "title": "Statistical Significance of Clustering for High-Dimensional Count Data", "abs": "Clustering is widely used in biomedical research for meaningful subgroup identification. However, most existing clustering algorithms do not account for the statistical uncertainty of the resulting clusters and consequently may generate spurious clusters due to natural sampling variation. To address this problem, the Statistical Significance of Clustering (SigClust) method was developed to evaluate significance of clusters in high-dimensional data. While SigClust has been successful in testing mixtures of continuous distributions, it is not specifically designed for discrete distributions, such as count data in genomics. Moreover, SigClust and its variations often suffer from reduced statistical power when applied to non-Gaussian high-dimensional data. To overcome these limitations, we propose SigClust-DEV, a method designed to evaluate the significance of clusters in count data. Through extensive simulations, we compare SigClust-DEV against other existing SigClust approaches across various count distributions and demonstrate its superior performance. Furthermore, we apply our method SigClust-DEV to Hydra single-cell RNA sequencing (scRNA) data and electronic health records (EHRs) of cancer patients to identify meaningful latent cell types and patient subgroups, respectively.", "keywords": [ "Unsupervised learning", "Cluster index", "Generalized PCA", "Dimension reduction", "High-dimension low-sample size data" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 1, 3, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "tVfvKrboVY", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Hang Yu", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Senior Algorithm Expert" }, { "name": "Jun Wang", "gender": "Male", "institution": "Columbia University, Columbia University", "country": "", "position": "PhD student" }, { "name": "Jun Wang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Lingxiao Wei", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Wei Zhang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Zihan Liao", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "PhD student" }, { "name": "jianguo Li", "gender": "Male", "institution": "Ant Group", "country": "", "position": "Director" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 45, "n_ref": 101, "n_ref_all": 137, "n_fig": 9, "n_tab": 10, "L_tab": 5019, "n_element_tab": 297, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 925, "n_element_tab_1": 40, "formula_len_all": 331, "formula_len_all_1": 18, "len_all": 229370, "len_all_1": 71113, "len_abs": 1241, "len_title": 139, "len_sents": 79757, "len_sents_1": 37245, "n_sents": 550, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1250, "title": "E2LLM: Encoder Elongated Large Language Models for Long-Context Understanding and Reasoning", "abs": "In the realm of Large Language Models (LLMs), the ability to process long contexts is increasingly crucial for tasks such as multi-round dialogues, code generation, and document summarization. This paper addresses the challenges of achieving high long-context performance, low computational complexity, and compatibility with pretrained models -- collectively termed the ``impossible triangle''. We introduce E2LLM (Encoder Elongated Large Language Models), a novel approach that effectively navigates this paradox. The method involves splitting long contexts into chunks, compressing each into soft prompts via a pretrained text encoder, and utilizing an adapter to align these representations with a decoder-only LLM. To further enhance the LLM's understanding and reasoning capabilities regarding the soft prompts, we implement two training objectives: one focused on reconstructing the encoder output and the other on long-context instruction fine-tuning. Extensive experiments including Needle in a Haystack and LongBench reveal that E2LLM not only outperforms seven existing state-of-the-art (SOTA) methods across various long-context tasks, but also achieves the lowest inference time and memory usage. Code will be available upon publication.", "keywords": [ "Long-Context Modeling", "Large Language Models", "Encoder", "Decoder" ], "rating_list": [ 5, 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 4, 2, 1 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 4, 3 ] }, { "paper_id": "tVRVE0OAyb", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Charu C. Aggarwal", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Distinguished Research Staff Member" }, { "name": "Hongliang Chi", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "PhD student" }, { "name": "Wei Jin", "gender": "unknown", "institution": "Emory University", "country": "US", "position": "Assistant Professor" }, { "name": "Yao Ma", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 7, "n_ref_uni": 53, "n_ref": 92, "n_ref_all": 139, "n_fig": 14, "n_tab": 11, "L_tab": 3651, "n_element_tab": 333, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 622, "formula_len_all_1": 218, "len_all": 221087, "len_all_1": 66394, "len_abs": 989, "len_title": 119, "len_sents": 82487, "len_sents_1": 33227, "n_sents": 609, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 923, "title": "Precedence-Constrained Winter Value for Effective Graph Data Valuation", "abs": "Data valuation is essential for quantifying data\u2019s worth, aiding in assessing data quality and determining fair compensation. While existing data valuation methods have proven effective in evaluating the value of Euclidean data, they face limitations when applied to the increasingly popular graph-structured data. Particularly, graph data valuation introduces unique challenges, primarily stemming from the intricate dependencies among nodes and the exponential growth in value estimation costs. To address the challenging problem of graph data valuation, we put forth an innovative solution, Precedence-Constrained Winter (PC-Winter) Value, to account for the complex graph structure. Furthermore, we develop a variety of strategies to address the computational challenges and enable efficient approximation of PC-Winter. Extensive experiments demonstrate the effectiveness of PC-Winter across diverse datasets and tasks.", "keywords": [ "Data Valuation", "Graph Learning", "Graph neural network" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 5, 3, 3 ] }, { "paper_id": "tVNZj27pb3", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 13, "author_info_list": [ { "name": "Abudukelimu Wuerkaixi", "gender": "Male", "institution": "Tsinghua University, Beijing", "country": "CN", "position": "PhD student" }, { "name": "Changshui Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Ding Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Gang Niu", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Senior Research Scientist (tenured)" }, { "name": "Haoxuan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Houfang Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Jingfeng Zhang", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Researcher" }, { "name": "Sen Cui", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Sifan YANG", "gender": "Female", "institution": "Tsinghua University", "country": "", "position": "Assistant Professor" }, { "name": "Tianling Ren", "gender": "unknown", "institution": "Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Yi Yang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Zhikang Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 21, "n_formula_1": 9, "n_ref_uni": 44, "n_ref": 76, "n_ref_all": 113, "n_fig": 5, "n_tab": 16, "L_tab": 5477, "n_element_tab": 395, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2484, "n_element_tab_1": 131, "formula_len_all": 1412, "formula_len_all_1": 539, "len_all": 196176, "len_all_1": 67156, "len_abs": 1372, "len_title": 112, "len_sents": 68265, "len_sents_1": 31624, "n_sents": 500, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1378, "title": "Do We Really Need Parameter-Isolation to Protect Task Knowledge?", "abs": "Due to the dynamic nature of tasks, how deep networks can transition from a static structure, trained on previous tasks, to a dynamic structure that adapts to continuously changing data inputs has garnered significant attention. This involves learning new task knowledge while avoiding catastrophic forgetting of previously acquired knowledge. Continual learning is a learning approach aimed at addressing the problem of catastrophic forgetting, primarily by constraining or isolating parameter changes to protect the knowledge of prior tasks. However, while existing methods offer good protection for old task knowledge, they often diminish the ability to learn new task knowledge. Given the sparsity of activation channels in a deep network, we introduce a novel misaligned fusion method within the context of continual learning. This approach allows for the adaptive allocation of available pathways to protect crucial knowledge from previous tasks, replacing traditional isolation techniques. Furthermore, when new tasks are introduced, the network can undergo full parameter training, enabling a more comprehensive learning of new tasks. This work conducts comparative tests of our method against other approaches using deep network architectures of various sizes and popular benchmark datasets. The performance demonstrates the effectiveness and superiority of our method.", "keywords": [ "continual learning" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 1, 1, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "tU074jg2vS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Constantin Eichenberg", "gender": "unknown", "institution": "Aleph Alpha", "country": "", "position": "Researcher" }, { "name": "Lukas Balles", "gender": "Male", "institution": "Aleph Alpha ", "country": "DE", "position": "Researcher" }, { "name": "Pit Neitemeier", "gender": "Male", "institution": "Ruprecht-Karls-Universit\u00e4t Heidelberg", "country": "DE", "position": "MS student" }, { "name": "Bjoern Deiseroth", "gender": "Male", "institution": "Aleph Alpha ", "country": "DE", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 10, "n_ref_uni": 41, "n_ref": 65, "n_ref_all": 99, "n_fig": 10, "n_tab": 7, "L_tab": 2902, "n_element_tab": 389, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 403, "n_element_tab_1": 40, "formula_len_all": 1321, "formula_len_all_1": 469, "len_all": 198064, "len_all_1": 60635, "len_abs": 1461, "len_title": 126, "len_sents": 50040, "len_sents_1": 27811, "n_sents": 392, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1468, "title": "Hierarchical Autoregressive Transformers for Tokenizer-Free Language Modelling", "abs": "Tokenization is a fundamental step in natural language processing, breaking text into units that computational models can process. While learned subword tokenizers have become the de-facto standard, they present challenges such as large vocabularies, limited adaptability to new domains or languages, and sensitivity to spelling errors and variations. To overcome these limitations, we investigate a hierarchical architecture for autoregressive language modelling that combines character-level and word-level processing. It employs a lightweight character-level encoder to convert character sequences into word embeddings, which are then processed by a word-level backbone model and decoded back into characters via a compact character-level decoder. This method retains the sequence compression benefits of word-level tokenization without relying on a rigid, predefined vocabulary. We demonstrate, at scales up to 7 billion parameters, that hierarchical transformers match the downstream task performance of subword-tokenizer-based models while exhibiting significantly greater robustness to input perturbations. Additionally, during continued pretraining on an out-of-domain language, our model trains almost twice as fast, achieves superior performance on the target language, and retains more of its previously learned knowledge. Hierarchical transformers pave the way for NLP systems that are more robust, flexible, and generalizable across languages and domains.", "keywords": [ "transformer", "autoregressive", "generative", "language modelling", "tokenizer-free", "byte-level", "hierarchical" ], "rating_list": [ 5, 8, 8 ], "soundness_list": [ 1, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "tTPHgb0EtV", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Fatih Ilhan", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Ling Liu", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Selim Furkan Tekin", "gender": "Male", "institution": "College of Computing, Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Sihao Hu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Tiansheng Huang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 4, "n_ref_uni": 31, "n_ref": 84, "n_ref_all": 109, "n_fig": 4, "n_tab": 15, "L_tab": 3274, "n_element_tab": 443, "n_fig_1": 2, "n_tab_1": 10, "L_tab_1": 2398, "n_element_tab_1": 351, "formula_len_all": 562, "formula_len_all_1": 294, "len_all": 158814, "len_all_1": 73865, "len_abs": 1086, "len_title": 149, "len_sents": 50866, "len_sents_1": 31233, "n_sents": 432, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 100, "L_abs": 1112, "title": "Booster: Tackling Harmful Fine-tuning for Large Language Models via Attenuating Harmful Perturbation", "abs": "Harmful fine-tuning attack \\citep{qi2023fine} poses serious safety concerns for Large language models' fine-tuning-as-a-service. While existing defenses have been proposed to mitigate the issue, their performances are still far away from satisfactory, and the root cause of the problem has not been fully recovered. To this end, we in this paper show that \\textit{harmful perturbation} over the model weights could be a probable cause of alignment-broken. In order to attenuate the negative impact of harmful perturbation, we propose an alignment-stage solution, dubbed Booster. Technically, along with the original alignment loss, we append a loss regularizer in the alignment stage's optimization. The regularizer ensures that the model's harmful loss reduction after the simulated harmful perturbation is attenuated, thereby mitigating the subsequent fine-tuning risk. Empirical results show that Booster can effectively reduce the harmful score of the fine-tuned models while maintaining the performance of downstream tasks. Our code is available at \\url{https://anonymous.4open.science/r/Booster-EF18}.", "keywords": [ "Harmful fine-tuning", "LLM", "safety alignment" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "tTOo7r4G9j", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Anshumali Shrivastava", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Associate Professor" }, { "name": "Junda Su", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Undergrad student" }, { "name": "Oscar Wu", "gender": "Male", "institution": "Rice University", "country": "US", "position": "MS student" }, { "name": "Tianyi Zhang", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Zhaozhuo Xu", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "US", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 36, "n_ref": 88, "n_ref_all": 102, "n_fig": 3, "n_tab": 7, "L_tab": 2086, "n_element_tab": 255, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1665, "n_element_tab_1": 220, "formula_len_all": 10, "formula_len_all_1": 74, "len_all": 142665, "len_all_1": 62472, "len_abs": 3363, "len_title": 126, "len_sents": 39116, "len_sents_1": 26114, "n_sents": 303, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1728, "title": "SpaLLM: Unified Compressive Adaptation of Large Language Models with Sketching", "abs": "Compressive adaptation approaches, such as QLoRA, are widely popular alternatives for reducing memory requirements during fine-tuning of large language models (LLMs) while producing models capable of handling various downstream tasks. The key idea is to employ a \u201ctwo-tower\u201d architecture: compressing pretrained LLM parameters into compact representations and fine-tuning the additive full-precision adapter, which typically has few tunable parameters in low-rank format. However, the strict algebraic assumptions, such as low-rank assumption, and the complexity of composing two-tower architectures are some of the known shortcomings, resulting in a poor accuracy-efficiency trade-off. In response to these known limitations, we propose SpaLLM (Sketched Parameter Adaptation of LLMs), a novel compressive adaptation approach for LLMs. This method is also the first to illustrate parameter-sharing compression methods for LLM fine-tuning, which, unlike QLoRA, are free from strict low-rank algebraic assumptions on adapters. Furthermore, our proposal unifies model compression and adaptation into a single, streamlined process, eliminating the need for two-tower architectures. SpaLLM sketches pre-trained LLM weights into lookup tables and directly fine-tunes the values in these tables. This approach simplifies LLMs\u2019 compressive adaptation workflow, potentially improves multi-user serving efficiency, and delivers significantly better accuracy for both natural language understanding and generation tasks. Moreover, by avoiding the \u201ctwo-tower\u201d architecture, our framework only requires one compressed matrix multiplication per layer during inference, demonstrating superior inference efficiency compared to previous methods.", "keywords": [ "large language model", "sketching" ], "rating_list": [ 5, 3, 1, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "tTDUrseRRU", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Chenlin Meng", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Christopher D. Manning", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "", "position": "Full Professor" }, { "name": "Enxin Song", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Jenq-Neng Hwang", "gender": "Male", "institution": "University of Washington, Seattle", "country": "US", "position": "Full Professor" }, { "name": "Omer Bar-Tal", "gender": "unknown", "institution": "Google", "country": "IL", "position": "Intern" }, { "name": "Vashisht Madhavan", "gender": "unknown", "institution": "Pika Labs", "country": "US", "position": "Researcher" }, { "name": "Wenhao Chai", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "MS student" }, { "name": "Yilun Du", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Saining Xie", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 3, "n_ref_uni": 128, "n_ref": 269, "n_ref_all": 334, "n_fig": 22, "n_tab": 14, "L_tab": 6478, "n_element_tab": 848, "n_fig_1": 4, "n_tab_1": 9, "L_tab_1": 7060, "n_element_tab_1": 690, "formula_len_all": 1243, "formula_len_all_1": 385, "len_all": 389763, "len_all_1": 68924, "len_abs": 1639, "len_title": 126, "len_sents": 84840, "len_sents_1": 25195, "n_sents": 669, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1415, "title": "AuroraCap: Efficient, Performant Video Detailed Captioning and a New Benchmark", "abs": "Video detailed captioning is a key task which aims to generate comprehensive and coherent textual descriptions of video content, benefiting both video understanding and generation. In this paper, we propose AuroraCap, a video captioner based on a large multimodal model. We follow the simplest architecture design without additional parameters for temporal modeling. To address the overhead caused by lengthy video sequences, we implement the token merging strategy, reducing the number of input visual tokens. Surprisingly, we found that this strategy results in little performance loss. AuroraCap shows superior performance on various video and image captioning benchmarks, for example, obtaining a CIDEr of 88.9 on Flickr30k, beating GPT-4V (55.3) and Gemini-1.5 Pro (82.2). However, existing video caption benchmarks only include simple descriptions, consisting of a few dozen words, which limits research in this field. Therefore, we develop VDC, a video detailed captioning benchmark with over one thousand carefully annotated structured captions. In addition, we propose a new LLM-assisted metric VDCscore for bettering evaluation, which adopts a divide-and-conquer strategy to transform long caption evaluation into multiple short question-answer pairs. With the help of human Elo ranking, our experiments show that this benchmark better correlates with human judgments of video detailed captioning quality.", "keywords": [ "Video Captioning", "Benchmark", "Multimodel Large Language Model" ], "rating_list": [ 6, 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 4 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "tTBXePRKSx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Ce Zhang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Deva Ramanan", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Katia P. Sycara", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Louis-Philippe Morency", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Ruslan Salakhutdinov", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Simon Stepputtis", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Yaqi Xie", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Zhehan Kan", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zifu Wan", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Martin Ma", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 5, "n_ref_uni": 54, "n_ref": 141, "n_ref_all": 186, "n_fig": 8, "n_tab": 18, "L_tab": 7819, "n_element_tab": 1060, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1277, "n_element_tab_1": 144, "formula_len_all": 978, "formula_len_all_1": 591, "len_all": 234722, "len_all_1": 65228, "len_abs": 1672, "len_title": 159, "len_sents": 58408, "len_sents_1": 30892, "n_sents": 429, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 111, "L_abs": 1449, "title": "Self-Correcting Decoding with Generative Feedback for Mitigating Hallucinations in Large Vision-Language Models", "abs": "While recent Large Vision-Language Models (LVLMs) have shown remarkable performance in multi-modal tasks, they are prone to generating hallucinatory text responses that do not align with the given visual input, which restricts their practical applicability in real-world scenarios. In this work, inspired by the observation that the text-to-image generation process is the inverse of image-conditioned response generation in LVLMs, we explore the potential of leveraging text-to-image generative models to assist in mitigating hallucinations in LVLMs. We discover that generative models can offer valuable self-feedback for mitigating hallucinations at both the response and token levels. Building on this insight, we introduce self-correcting Decoding with Generative Feedback (DeGF), a novel training-free algorithm that incorporates feedback from text-to-image generative models into the decoding process to effectively mitigate hallucinations in LVLMs. Specifically, DeGF generates an image from the initial response produced by LVLMs, which acts as an auxiliary visual reference and provides self-feedback to verify and correct the initial response through complementary or contrastive decoding. Extensive experimental results validate the effectiveness of our approach in mitigating diverse types of hallucinations, consistently surpassing state-of-the-art methods across six benchmarks. Code is available at https://github.com/zhangce01/DeGF.", "keywords": [ "Large Vision-Language Models", "Hallucinations", "Generative Feedback" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "tSmkYZ8vU7", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Bo Li", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Hao Zhang", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "Peng-Tao Jiang", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "Qianru Sun", "gender": "Female", "institution": "Singapore Management University", "country": "SG", "position": "Associate Professor" }, { "name": "Ruihao Xia", "gender": "unknown", "institution": "East China University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yang Tang", "gender": "Male", "institution": "East China University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yu Liang", "gender": "Male", "institution": "\u7ef4\u6c83\u79fb\u52a8\u901a\u4fe1\uff08\u676d\u5dde\uff09\u6709\u9650\u516c\u53f8", "country": "CN", "position": "Researcher" }, { "name": "Pan Zhou", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 28, "n_ref": 120, "n_ref_all": 143, "n_fig": 8, "n_tab": 11, "L_tab": 3370, "n_element_tab": 344, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3449, "n_element_tab_1": 196, "formula_len_all": 474, "formula_len_all_1": 475, "len_all": 127793, "len_all_1": 62467, "len_abs": 1730, "len_title": 113, "len_sents": 38037, "len_sents_1": 27142, "n_sents": 254, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1740, "title": "Towards Natural Image Matting in the Wild via Real-Scenario Prior", "abs": "Recent approaches attempt to adapt powerful interactive segmentation models, such as SAM, to interactive matting and fine-tune the models based on synthetic matting datasets. However, models trained on synthetic data fail to generalize to complex and occlusion scenes. We address this challenge by proposing a new matting dataset based on the COCO dataset, namely COCO-Matting. Specifically, the construction of our COCO-Matting includes accessory fusion and mask-to-matte, which selects real-world complex images from COCO and converts semantic segmentation masks to matting labels. The built COCO-Matting comprises an extensive collection of 38,251 human instance-level alpha mattes in complex natural scenarios. Furthermore, existing SAM-based matting methods extract intermediate features and masks from a frozen SAM and only train a lightweight matting decoder by end-to-end matting losses, which do not fully exploit the potential of the pre-trained SAM. Thus, we propose SEMat which revamps the network architecture and training objectives. For network architecture, the proposed feature-aligned transformer learns to extract fine-grained edge and transparency features. The proposed matte-aligned decoder aims to segment matting-specific objects and convert coarse masks into high-precision mattes. For training objectives, the proposed regularization and trimap loss aim to retain the prior from the pre-trained model and push the matting logits extracted from the mask decoder to contain trimap-based semantic information. Extensive experiments across seven diverse datasets demonstrate the superior performance of our method, proving its efficacy in interactive natural image matting. Code is available in the supplementary file.", "keywords": [ "Image Matting", "Interactive Matting" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "tRzujdRwE8", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Arber Zela", "gender": "Male", "institution": "University of Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Frank Hutter", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen & University of Freiburg", "country": "DE", "position": "Full Professor" }, { "name": "Josif Grabocka", "gender": "Male", "institution": "University of Technology Nuremberg", "country": "DE", "position": "Full Professor" }, { "name": "Maciej Janowski", "gender": "Male", "institution": "Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Sebastian Pineda Arango", "gender": "Male", "institution": "Universit\u00e4t Freiburg", "country": "", "position": "PhD student" }, { "name": "Lennart Oswald Purucker", "gender": "Male", "institution": "University of Freiburg, Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 10, "n_ref_uni": 55, "n_ref": 88, "n_ref_all": 148, "n_fig": 20, "n_tab": 8, "L_tab": 6738, "n_element_tab": 325, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 750, "n_element_tab_1": 72, "formula_len_all": 2853, "formula_len_all_1": 681, "len_all": 212507, "len_all_1": 61649, "len_abs": 1153, "len_title": 82, "len_sents": 57110, "len_sents_1": 25397, "n_sents": 480, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 34, "L_abs": 1036, "title": "Dynamic Post-Hoc Neural Ensemblers", "abs": "Ensemble methods are known for enhancing the accuracy and robustness of machine learning models by combining multiple base learners. However, standard approaches like greedy or random ensembles often fall short, as they assume a constant weight across samples for the ensemble members. This can limit expressiveness and hinder performance when aggregating the ensemble predictions. In this study, we explore employing neural networks as ensemble methods, emphasizing the significance of dynamic ensembling to leverage diverse model predictions adaptively. Motivated by the risk of learning low-diversity ensembles, we propose regularizing the model by randomly dropping base model predictions during the training. We demonstrate this approach lower bounds the diversity within the ensemble, reducing overfitting and improving generalization capabilities. Our experiments showcase that the dynamic neural ensemblers yield competitive results compared to strong baselines in computer vision, natural language processing, and tabular data.", "keywords": [ "ensemble learning", "neural networks", "regularization", "modality agnostic", "dynamic ensembling" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "tRNKe2Vgqt", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Jiachen Li", "gender": "Male", "institution": "xAI", "country": "US", "position": "Researcher" }, { "name": "Jianfeng Wang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Kaizhi Zheng", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Kevin Lin", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Lijuan Wang", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Linjie Li", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Wanrong Zhu", "gender": "Female", "institution": "Adobe Research", "country": "US", "position": "Researcher" }, { "name": "Weixi Feng", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Xin Eric Wang", "gender": "Male", "institution": "Simular", "country": "US", "position": "Principal Researcher" }, { "name": "Xuehai He", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "student" }, { "name": "Yue Fan", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Yujie Lu", "gender": "unknown", "institution": "UC Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Zhengyuan Yang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "William Yang Wang", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 52, "n_ref": 140, "n_ref_all": 176, "n_fig": 11, "n_tab": 6, "L_tab": 7190, "n_element_tab": 740, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1749, "n_element_tab_1": 226, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 249755, "len_all_1": 69146, "len_abs": 974, "len_title": 123, "len_sents": 49719, "len_sents_1": 27663, "n_sents": 378, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1478, "title": "MMWorld: Towards Multi-discipline Multi-faceted World Model Evaluation in Videos", "abs": "Multimodal Language Language Models (MLLMs) demonstrate the emerging abilities of \"world models\"---interpreting and reasoning about complex real-world dynamics. To assess these abilities, we posit videos are the ideal medium, as they encapsulate rich representations of real-world dynamics and causalities. To this end, we introduce MMWorld, a new benchmark for multi-discipline, multi-faceted multimodal video understanding. MMWorld distinguishes itself from previous video understanding benchmarks with two unique advantages: (1) multi-discipline, covering various disciplines that often require domain expertise for comprehensive understanding; (2) multi-faceted reasoning, including explanation, counterfactual thinking, future prediction, etc. MMWorld consists of a human-annotated dataset to evaluate MLLMs with questions about the whole videos and a synthetic dataset to analyze MLLMs within a single modality of perception. Together, MMWorld encompasses 1,910 videos across seven broad disciplines and 69 subdisciplines, complete with 6,627 question-answer pairs and associated captions. The evaluation includes 4 proprietary and 11 open-source MLLMs, which struggle on MMWorld (e.g., GPT-4o performs the best with only 62.5% accuracy), showing large room for improvement. Further ablation studies reveal other interesting findings such as models' different skill sets from humans. We hope MMWorld can serve as an essential step towards world model evaluation in videos.", "keywords": [ "Video Understanding", "Benchmark" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 4, 2 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "tR2qSmSOQ3", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Donglin Wang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Associate Professor" }, { "name": "Jinxin Liu", "gender": "unknown", "institution": "Westlake Robotics", "country": "CN", "position": "Researcher" }, { "name": "Xiao Xiong", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "MS student" }, { "name": "Zifeng Zhuang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Ziqi Zhang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Micro Masters Student" } ], "n_formula": 23, "n_formula_1": 9, "n_ref_uni": 51, "n_ref": 117, "n_ref_all": 152, "n_fig": 8, "n_tab": 15, "L_tab": 5517, "n_element_tab": 554, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2631, "n_element_tab_1": 276, "formula_len_all": 3590, "formula_len_all_1": 1047, "len_all": 193213, "len_all_1": 67752, "len_abs": 1383, "len_title": 143, "len_sents": 49014, "len_sents_1": 29702, "n_sents": 353, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1394, "title": "Improving Offline-to-Online Reinforcement Learning with Q Conditioned State Entropy Exploration", "abs": "Studying how to fine-tune offline reinforcement learning (RL) pre-trained policy is profoundly significant for enhancing the sample efficiency of RL algorithms. However, directly fine-tuning pre-trained policies often results in sub-optimal performance. This is primarily due to the distribution shift between offline pre-training and online fine-tuning stages. Specifically, the distribution shift limits the acquisition of effective online samples, ultimately impacting the online fine-tuning performance. In order to narrow down the distribution shift between offline and online stages, we proposed Q conditioned state entropy (QCSE) as intrinsic reward. Specifically, QCSE maximizes the state entropy of all samples individually, considering their respective Q values. This approach encourages exploration of low-frequency samples while penalizing high-frequency ones, and implicitly achieves State Marginal Matching (SMM), thereby ensuring optimal performance, solving the asymptotic sub-optimality of constraint-based approaches. Additionally, QCSE can seamlessly integrate into various RL algorithms, enhancing online fine-tuning performance. To validate our claim, we conduct extensive experiments, and observe significant improvements with QCSE ( about 10.9% for CQL and 8% for Cal-QL). Furthermore, we extended experimental tests to other algorithms, affirming the generality of QCSE.", "keywords": [ "Offline-to-online Reinforcement Learning" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "tQyh0gnfqW", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Hyeongwoo Kim", "gender": "Male", "institution": "KAIST", "country": "", "position": "PhD student" }, { "name": "Jeheon Woo", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Jun Hyeong Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology ", "country": "", "position": "PhD student" }, { "name": "Seokhyun Moon", "gender": "Male", "institution": "kaist", "country": "KR", "position": "PhD student" }, { "name": "Woo Youn Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" }, { "name": "SEONGHWAN KIM", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 103, "n_formula_1": 29, "n_ref_uni": 49, "n_ref": 110, "n_ref_all": 170, "n_fig": 22, "n_tab": 8, "L_tab": 2669, "n_element_tab": 332, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 401, "n_element_tab_1": 70, "formula_len_all": 10132, "formula_len_all_1": 2050, "len_all": 273988, "len_all_1": 67610, "len_abs": 1309, "len_title": 120, "len_sents": 91391, "len_sents_1": 30443, "n_sents": 798, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1388, "title": "Discrete Diffusion Schr\u00f6dinger Bridge Matching for Graph Transformation", "abs": "Transporting between arbitrary distributions is a fundamental goal in generative modeling.\nRecently proposed diffusion bridge models provide a potential solution, but they rely on a joint distribution that is difficult to obtain in practice.\nFurthermore, formulations based on continuous domains limit their applicability to discrete domains such as graphs.\nTo overcome these limitations, we propose Discrete Diffusion Schr\u00f6dinger Bridge Matching (DDSBM), a novel framework that utilizes continuous-time Markov chains to solve the SB problem in a high-dimensional discrete state space.\nOur approach extends Iterative Markovian Fitting to discrete domains, and we have proved its convergence to the SB.\nFurthermore, we adapt our framework for the graph transformation, and show that our design choice of underlying dynamics characterized by independent modifications of nodes and edges can be interpreted as the entropy-regularized version of optimal transport with a cost function described by the graph edit distance.\nTo demonstrate the effectiveness of our framework, we have applied DDSBM to molecular optimization in the field of chemistry.\nExperimental results demonstrate that DDSBM effectively optimizes molecules' property-of-interest with minimal graph transformation, successfully retaining other features.\nCode is available [here](https://anonymous.4open.science/r/DDSBM-1353).", "keywords": [ "Schr\u00f6dinger Bridge", "Discrete Diffusion Model", "Molecular Optimization" ], "rating_list": [ 6, 3, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "tQ1PmLfPBL", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Ha-Yeong Choi", "gender": "Female", "institution": "Korea Telecom Research", "country": "KR", "position": "Researcher" }, { "name": "Sang-Hoon Lee", "gender": "Male", "institution": "Ajou University", "country": "", "position": "Assistant Professor" }, { "name": "Seong-whan Lee", "gender": "unknown", "institution": "Korea University", "country": "KR", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 6, "n_ref_uni": 58, "n_ref": 192, "n_ref_all": 225, "n_fig": 3, "n_tab": 22, "L_tab": 11749, "n_element_tab": 1760, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 3641, "n_element_tab_1": 443, "formula_len_all": 1329, "formula_len_all_1": 422, "len_all": 263892, "len_all_1": 74844, "len_abs": 7082, "len_title": 124, "len_sents": 63973, "len_sents_1": 30161, "n_sents": 530, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1890, "title": "PeriodWave: Multi-Period Flow Matching for High-Fidelity Waveform Generation", "abs": "Recently, universal waveform generation tasks have been investigated conditioned on various out-of-distribution scenarios. Although one-step GAN-based methods have shown their strength in fast waveform generation, they are vulnerable to train-inference mismatch scenarios such as two-stage text-to-speech. Meanwhile, diffusion-based models have shown their powerful generative performance in other domains; however, they stay out of the limelight due to slow inference speed in waveform generation tasks. Above all, there is no generator architecture that can explicitly disentangle the natural periodic features of high-resolution waveform signals. In this paper, we propose PeriodWave, a novel universal waveform generation model from Mel-spectrogram and neural audio codec. First, we introduce a period-aware flow matching estimator that effectively captures the periodic features of the waveform signal when estimating the vector fields. Additionally, we utilize a multi-period estimator that avoids overlaps to capture different periodic features of waveform signals. Although increasing the number of periods can improve the performance significantly, this requires more computational costs. To reduce this issue, we also propose a single period-conditional universal estimator that can feed-forward parallel by period-wise batch inference. Additionally, we first introduce FreeU to reduce the high-frequency noise for waveform generation. Furthermore, we demonstrate the effectiveness of the proposed method in neural audio codec decoding task, and present the streaming generation framework of non-autoregressive model for speech language models. The experimental results demonstrated that our model outperforms the previous models in reconstruction tasks from Mel-spectrogram and discrete token, and text-to-speech tasks. Our demo is available at https://periodwave.github.io/demo.", "keywords": [ "Conditional Flow Matching", "Neural Vocoder", "Speech Synthesis", "Neural Audio Codec", "Speech Language Models" ], "rating_list": [ 6, 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 2, 4, 4 ] }, { "paper_id": "tPsZDNvMqJ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Yang Liu", "gender": "Male", "institution": "Indiana University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 20, "n_formula_1": 18, "n_ref_uni": 23, "n_ref": 34, "n_ref_all": 53, "n_fig": 3, "n_tab": 3, "L_tab": 438, "n_element_tab": 79, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2425, "n_element_tab_1": 104, "formula_len_all": 1303, "formula_len_all_1": 1279, "len_all": 83535, "len_all_1": 56256, "len_abs": 650, "len_title": 131, "len_sents": 26246, "len_sents_1": 22508, "n_sents": 203, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 657, "title": "A Polynomial Time Graph Isomorphism Algorithm via Self-Supervised Gradient Descent", "abs": "Graph isomorphism (GI) is a fundamental problem in graph theory. \nDespite recent advancements, determining whether two graphs are isomorphic remains computationally challenging. \nThis paper introduces the Polynomial Time Graph Isomorphism (PTGI) algorithm, an optimization-based approach leveraging self-supervision techniques to efficiently tackle the graph isomorphism problem. \nPTGI aims to escape local optima caused by graph symmetries and provides high accuracy in identifying isomorphic graphs in polynomial time. \nExperimental results demonstrate PTGI's effectiveness across various graph types, making it a valuable tool for practical applications.", "keywords": [ "Graph Isomorphism", "Optimization", "Self-Supervised Learning", "Graph Matching", "Graph Theory" ], "rating_list": [ 6, 1, 3, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 3, 1, 1, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "tPNHOoZFl9", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Yi Ren", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "D. J. Sutherland", "gender": "Female", "institution": "University of British Columbia", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 61, "n_formula_1": 8, "n_ref_uni": 34, "n_ref": 85, "n_ref_all": 164, "n_fig": 23, "n_tab": 1, "L_tab": 74, "n_element_tab": 13, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7061, "formula_len_all_1": 872, "len_all": 267547, "len_all_1": 73250, "len_abs": 1407, "len_title": 83, "len_sents": 92474, "len_sents_1": 32156, "n_sents": 713, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 35, "L_abs": 1414, "title": "Learning Dynamics of LLM Finetuning", "abs": "Learning dynamics, which describes how the learning of specific training examples influences the model's predictions on other examples, \ngives us a powerful tool for understanding the behavior of deep learning systems. We study the learning dynamics of large language models during different types of finetuning, by analyzing the step-wise decomposition of how influence accumulates among different potential responses. Our framework allows a uniform interpretation of many interesting observations about the training of popular algorithms for both instruction tuning and preference tuning. In particular, we propose a hypothetical explanation of why specific types of hallucination are strengthened after finetuning, e.g., the model might use phrases or facts in the response for question B to answer question A, or the model might keep repeating similar simple phrases when generating responses. We also extend our framework and highlight a unique ``squeezing effect'' to explain a previously observed phenomenon in off-policy direct preference optimization (DPO), where running DPO for too long makes even the desired outputs less likely. This framework also provides insights into where the benefits of on-policy DPO and other variants come from. The analysis not only provides a novel perspective of understanding LLM's finetuning but also inspires a simple, effective method to improve alignment performance.", "keywords": [ "Learning dynamics", "LLM", "finetuning", "DPO" ], "rating_list": [ 6, 8, 8, 10 ], "soundness_list": [ 3, 4, 4, 4 ], "presentation_list": [ 3, 4, 4, 4 ], "contribution_list": [ 2, 3, 4, 4 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "tO58o0ahdg", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Jingwei Guo", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Kaizhu Huang", "gender": "Male", "institution": "Duke Kunshan University", "country": "CN", "position": "Full Professor" }, { "name": "Qiufeng Wang", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "CN", "position": "Full Professor" }, { "name": "Xi Yang", "gender": "Female", "institution": "Xi'an Jiaotong-Liverpool University", "country": "CN", "position": "Associate Professor" }, { "name": "Zixian Su", "gender": "Female", "institution": "University of Liverpool", "country": "GB", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 18, "n_ref_uni": 38, "n_ref": 76, "n_ref_all": 90, "n_fig": 5, "n_tab": 4, "L_tab": 6298, "n_element_tab": 1380, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 4137, "n_element_tab_1": 761, "formula_len_all": 1157, "formula_len_all_1": 1126, "len_all": 151262, "len_all_1": 77610, "len_abs": 1744, "len_title": 116, "len_sents": 41148, "len_sents_1": 31124, "n_sents": 291, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1750, "title": "Decentralizing Test-time Adaptation under Heterogeneous Data Streams", "abs": "While Test-Time Adaptation (TTA) has shown promise in addressing distribution shifts between training and testing data, its effectiveness diminishes with heterogenous data streams due to uniform target estimation. As previous attempts merely stabilize model fine-tuning over time to handle continually changing environments, they fundamentally assume a homogeneous target domain at any moment, leaving the intrinsic real-world data heterogeneity unresolved. This paper delves into TTA under heterogeneous data streams, moving beyond current model-centric limitations. By revisiting TTA from a data-centric perspective, we discover that decomposing samples into Fourier space facilitates an accurate data separation across different frequency levels. Drawing from this insight, we propose a novel Frequency-based Decentralized Adaptation framework, which transitions data from globally heterogeneous to locally homogeneous in Fourier space and employs decentralized adaptation to manage diverse distribution shifts.\nParticularly, multiple local models are allowed to independently adjust to their specific data segments while periodically exchanging knowledge to form a cohesive global model. As such, not only can data diversity be captured, but also the overall model generalization can be enhanced across multiple distribution shifts. Importantly, we devise a novel Fourier-based augmentation strategy to assist in decentralizing adaptation, which selectively augments samples for each type of distribution shift and further enhances model robustness in complex real-world environments. Extensive experiments across various settings (corrupted, natural, and medical) demonstrate the superiority of our proposed framework over the state-of-the-arts.", "keywords": [ "Test-time Adaptation", "Transfer Learning", "Data Heterogeneity" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 5, 3, 3 ] }, { "paper_id": "tNxr38vfYR", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Mahyar Najibi", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Qichen Fu", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Qingqing Cao", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Sachin Mehta", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Yuxin Wen", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 37, "n_ref": 94, "n_ref_all": 118, "n_fig": 14, "n_tab": 2, "L_tab": 1414, "n_element_tab": 198, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 82, "n_element_tab_1": 8, "formula_len_all": 438, "formula_len_all_1": 438, "len_all": 131287, "len_all_1": 60200, "len_abs": 1367, "len_title": 133, "len_sents": 35093, "len_sents_1": 28014, "n_sents": 271, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1392, "title": "Efficient Vision-Language Models by Summarizing Visual Tokens into Compact Registers", "abs": "Recent advancements in vision-language models have expanded their potential for real-world applications, enabling these models to perform complex reasoning on images. \nHowever, in the widely used fully autoregressive pipeline like LLaVA, where projected visual tokens are prepended to textual tokens, the visual tokens often number in the hundreds or thousands, making them much longer than the input prompt. This large quantity of visual tokens introduces significant computational overhead, slowing down training and inference. \nIn this paper, we propose Visual Compact Token Registers (Victor), a method that reduces the number of visual tokens by summarizing them into a smaller set of register tokens. Victor adds a few learnable register tokens after the visual tokens and summarizes the visual information into these registers using the first few layers in the language tower. After these few layers, all visual tokens are discarded, significantly improving computational efficiency for both training and inference. Notably, our method is easy to implement and requires a small number of new trainable parameters with minimal impact on model performance.\nIn our experiment, with merely $8$ visual registers\u2014about $1\\%$ of the original tokens\u2014Victor shows less than a $4\\%$ performance drop while reducing total training time by $43\\%$ and boosting inference throughput by $3.36\\times$.", "keywords": [ "Vision-language Models" ], "rating_list": [ 3, 5, 6, 8, 3 ], "soundness_list": [ 3, 2, 3, 4, 2 ], "presentation_list": [ 2, 3, 3, 4, 2 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "tNvCSw8ONp", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Chandan K. Reddy", "gender": "Male", "institution": "Virginia Tech", "country": "US", "position": "Full Professor" }, { "name": "Chi Xue", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Hyeong Kyu Choi", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Stephanie Eckman", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Weijie Xu", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" } ], "n_formula": 31, "n_formula_1": 13, "n_ref_uni": 34, "n_ref": 70, "n_ref_all": 106, "n_fig": 8, "n_tab": 9, "L_tab": 6418, "n_element_tab": 1069, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2846, "n_element_tab_1": 497, "formula_len_all": 915, "formula_len_all_1": 315, "len_all": 183687, "len_all_1": 67991, "len_abs": 1095, "len_title": 113, "len_sents": 55533, "len_sents_1": 26481, "n_sents": 473, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1088, "title": "Mitigating Selection Bias with Node Pruning and Auxiliary Options", "abs": "Large language models (LLMs) often show unwarranted preference for certain choice options when responding to multiple-choice questions, posing significant reliability concerns in LLM-automated systems. To mitigate this selection bias problem, previous solutions utilized debiasing methods to adjust the model\u2019s input and/or output. Our work, in contrast, investigates the model\u2019s internal representation of the selection bias. Specifically, we introduce a novel debiasing approach, Bias Node Pruning (BNP), which eliminates the linear layer parameters that contribute to the bias. Furthermore, we present Auxiliary Option Injection (AOI), a simple yet effective input modification technique for debiasing, which is compatible even with black-box LLMs. To provide a more systematic evaluation of selection bias, we review existing metrics and introduce Choice Kullback-Leibler Divergence (CKLD), which addresses the insensitivity of the commonly used metrics to label imbalance. Experiments show that our methods are robust and adaptable across various datasets when applied to three LLMs.", "keywords": [ "Large Language Model", "Selection Bias" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "tNn6Hskmti", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Samet Demir", "gender": "unknown", "institution": "Koc University", "country": "TR", "position": "PhD student" }, { "name": "Zafer Dogan", "gender": "Male", "institution": "Koc University", "country": "TR", "position": "Assistant Professor" } ], "n_formula": 130, "n_formula_1": 28, "n_ref_uni": 43, "n_ref": 132, "n_ref_all": 194, "n_fig": 19, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 965, "n_element_tab_1": 11, "formula_len_all": 8866, "formula_len_all_1": 1831, "len_all": 199204, "len_all_1": 66899, "len_abs": 1390, "len_title": 164, "len_sents": 66773, "len_sents_1": 31080, "n_sents": 563, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 116, "L_abs": 1396, "title": "Asymptotic Analysis of Two-Layer Neural Networks after One Gradient Step under Gaussian Mixtures Data with Structure", "abs": "In this work, we study the training and generalization performance of two-layer neural networks (NNs) after one gradient descent step under structured data modeled by Gaussian mixtures. While previous research has extensively analyzed this model under isotropic data assumption, such simplifications overlook the complexities inherent in real-world datasets. Our work addresses this limitation by analyzing two-layer NNs under Gaussian mixture data assumption in the asymptotically proportional limit, where the input dimension, number of hidden neurons, and sample size grow with finite ratios. We characterize the training and generalization errors by leveraging recent advancements in Gaussian universality. Specifically, we prove that a high-order polynomial model performs equivalent to the non-linear neural networks under certain conditions. The degree of the equivalent model is intricately linked to both the \"data spread\" and the learning rate employed during one gradient step. Through extensive simulations, we demonstrate the equivalence between the original model and its polynomial counterpart across various regression and classification tasks. Additionally, we explore how different properties of Gaussian mixtures affect learning outcomes. Finally, we illustrate experimental results on Fashion-MNIST classification, indicating that our findings can translate to realistic data.", "keywords": [ "deep learning theory", "random features", "Gaussian equivalence", "universality", "high-dimensional asymptotics" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 4, 3, 2, 4 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "tNE0Y3S4fE", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Greg Ver Steeg", "gender": "Male", "institution": "University of California, Riverside", "country": "US", "position": "Associate Professor" }, { "name": "Shaorong Zhang", "gender": "Male", "institution": "University of California, Riverside", "country": "US", "position": "PhD student" }, { "name": "Xianghao Kong", "gender": "Male", "institution": "University of California, Riverside", "country": "US", "position": "PhD student" }, { "name": "Yuanbin Cheng", "gender": "Male", "institution": "University of California, Riverside", "country": "US", "position": "PhD student" } ], "n_formula": 77, "n_formula_1": 29, "n_ref_uni": 22, "n_ref": 61, "n_ref_all": 116, "n_fig": 10, "n_tab": 6, "L_tab": 888, "n_element_tab": 182, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1200, "n_element_tab_1": 105, "formula_len_all": 7819, "formula_len_all_1": 2957, "len_all": 140915, "len_all_1": 60783, "len_abs": 1012, "len_title": 127, "len_sents": 40596, "len_sents_1": 24059, "n_sents": 373, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 984, "title": "Exploring the Design Space of Diffusion Bridge Models via Stochasticity Control", "abs": "Diffusion bridge models effectively facilitate image-to-image (I2I) translation by connecting two distributions. However, existing methods overlook the impact of noise in sampling SDEs, transition kernel, and the base distribution on sampling efficiency, image quality and diversity. To address this gap, we propose the Stochasticity-controlled Diffusion Bridge (SDB), a novel theoretical framework that extends the design space of diffusion bridges, and provides strategies to mitigate singularities during both training and sampling. By controlling stochasticity in the sampling SDEs, our sampler achieves speeds up to $5 \\times$ faster than the baseline, while also producing lower FID scores. After training, SDB sets new benchmarks in image quality and sampling efficiency via managing stochasticity within the transition kernel. Furthermore, introducing stochasticity into the base distribution significantly improves image diversity, as quantified by a newly introduced metric.", "keywords": [ "Diffusion bridge models", "image translation", "diffusion models" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 1, 3, 3 ] }, { "paper_id": "tMKibc9Uxi", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Can Huang", "gender": "Male", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "Han Wang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Yang Li", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Yanjie Wang", "gender": "Male", "institution": "ByteDance Inc", "country": "CN", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 11, "n_ref": 17, "n_ref_all": 22, "n_fig": 4, "n_tab": 6, "L_tab": 913, "n_element_tab": 107, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 981, "n_element_tab_1": 113, "formula_len_all": 678, "formula_len_all_1": 646, "len_all": 102975, "len_all_1": 51216, "len_abs": 1075, "len_title": 112, "len_sents": 23058, "len_sents_1": 21636, "n_sents": 197, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1119, "title": "GLOMA: Global Video Text Spotting with Morphological Association", "abs": "Video Text Spotting (VTS) is a fundamental visual task that aims to predict the trajectories and content of texts in a video. Previous works usually conduct local associations and apply IoU-based distance and complex post-processing procedures to boost performance, ignoring the abundant temporal information and the morphological characteristics in VTS. In this paper, we propose \\model{} to model the tracking problem as global associations and utilize the Gaussian Wasserstein distance to guide the morphological correlation between frames. Our main contributions can be summarized as three folds. 1). We propose a Transformer-based global tracking method \\model{} for VTS and associate multiple frames simultaneously. 2). We introduce a Wasserstein distance-based method to conduct positional associations between frames. 3). We conduct extensive experiments on public datasets. On the ICDAR2015 video dataset, \\model{} achieves \\textbf{56.0} MOTA with \\textbf{4.6} absolute improvement compared with the previous SOTA method and outperforms the previous Transformer-based method by a significant \\textbf{8.3} MOTA.", "keywords": [ "Text Spotting", "Text Detection" ], "rating_list": [ 6, 5, 8, 8 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "tMG6btjBfd", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Jaeseok Oh", "gender": "unknown", "institution": "Hanyang University", "country": "KR", "position": "MS student" }, { "name": "Jihwan Yoon", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "MS student" }, { "name": "Minsik Lee", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "Full Professor" }, { "name": "Sangbeom Han", "gender": "unknown", "institution": "Hanyang University", "country": "KR", "position": "MS student" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 27, "n_ref": 68, "n_ref_all": 88, "n_fig": 6, "n_tab": 7, "L_tab": 4344, "n_element_tab": 571, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1207, "n_element_tab_1": 151, "formula_len_all": 655, "formula_len_all_1": 659, "len_all": 121021, "len_all_1": 59536, "len_abs": 377, "len_title": 142, "len_sents": 31074, "len_sents_1": 25720, "n_sents": 252, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1256, "title": "SplineGS: Learning Smooth Trajectories in Gaussian Splatting for Dynamic Scene Reconstruction", "abs": "Reconstructing complex scenes with deforming objects for novel view synthesis is a challenging task. Recent works have addressed this with 3D Gaussian Splatting, which effectively reconstructs static scenes with high quality in short training time, by adding specialized modules for the deformations of Gaussian blobs. However, designing an effective deformation module that incorporates appropriate spatiotemporal inductive biases still remains unresolved. To address this issue, we propose SplineGS in this paper, which utilizes non-uniform rational B-splines (NURBS), an extension of B-spline, to represent temporally smooth deformation. A set of representative trajectories are learned based on NURBS, and the individual trajectories of Gaussian blobs are represented as linear combinations of these trajectories for spatial smoothness. The weights of the combinations are trained based on a multi-resolution hash table and an MLP, with the positions of the Gaussian blobs as the keys. Thanks to this design, the proposed method does not need any regularizers for trajectories, which enables efficient training. Experiments demonstrate that the proposed method provides competitive performance over the existing methods with much shorter training time.", "keywords": [ "4D Gaussian splatting", "Dynamic scene reconstruction", "Novel-view synthesis for dynamic scenes" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "tLPHgQMw08", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Bao Nguyen", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Binh Nguyen", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Research Fellow" }, { "name": "Duy Nguyen", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Viet Anh Nguyen", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 43, "n_ref": 85, "n_ref_all": 107, "n_fig": 4, "n_tab": 14, "L_tab": 7078, "n_element_tab": 776, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1190, "n_element_tab_1": 227, "formula_len_all": 1528, "formula_len_all_1": 1418, "len_all": 179386, "len_all_1": 69513, "len_abs": 933, "len_title": 116, "len_sents": 48573, "len_sents_1": 30379, "n_sents": 429, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 856, "title": "Risk-Aware Distributional Intervention Policies for Language Models", "abs": "Language models are prone to occasionally undesirable generations, such as harmful or toxic content, despite their impressive capability to produce texts that appear accurate and coherent. In this paper, we present a new two-stage approach to detect and mitigate undesirable content generations by rectifying activations. First, we train an ensemble of layer-wise classifiers to detect undesirable content using activations by minimizing a smooth surrogate of the risk-aware score. Then, for contents that are detected as undesirable, we propose layer-wise distributional intervention policies that perturb the attention heads minimally while guaranteeing probabilistically the effectiveness of the intervention. Benchmarks on several language models and datasets show that our method outperforms baselines in reducing the generation of undesirable output.", "keywords": [ "Language Models", "Activations Steering", "AI safety" ], "rating_list": [ 6, 5, 3, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "tLFWU6izoA", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Fan Zhang", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Jing Liu", "gender": "Female", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "Full Professor" }, { "name": "Quan Sun", "gender": "Male", "institution": "StepFun", "country": "CN", "position": "Researcher" }, { "name": "Wenxuan Wang", "gender": "Male", "institution": "National Lab of Pattern Recognition, Institute of Automation,Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xinlong Wang", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Yepeng Tang", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 84, "n_ref": 147, "n_ref_all": 164, "n_fig": 2, "n_tab": 5, "L_tab": 3092, "n_element_tab": 181, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 3421, "n_element_tab_1": 175, "formula_len_all": 297, "formula_len_all_1": 297, "len_all": 209072, "len_all_1": 69351, "len_abs": 1523, "len_title": 88, "len_sents": 39650, "len_sents_1": 31908, "n_sents": 247, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1534, "title": "Diffusion Feedback Helps CLIP See Better", "abs": "Contrastive Language-Image Pre-training (CLIP), which excels at abstracting open-world representations across domains and modalities, has become a foundation for a variety of vision and multimodal tasks. However, recent studies reveal that CLIP has severe visual shortcomings, such as which can hardly distinguish orientation, quantity, color, structure, etc. These visual shortcomings also limit the perception capabilities of multimodal large language models (MLLMs) built on CLIP. The main reason could be that the image-text pairs used to train CLIP are inherently biased, due to the lack of the distinctiveness of the text and the diversity of images. In this work, we present a simple post-training approach for CLIP models, which largely overcomes its visual shortcomings via a self-supervised diffusion process. We introduce DIVA, which uses the DIffusion model as a Visual Assistant for CLIP. Specifically, DIVA leverages generative feedback from text-to-image diffusion models to optimize CLIP representations, with only images (without corresponding text). We demonstrate that DIVA improves CLIP's performance on the challenging MMVP-VLM benchmark which assesses fine-grained visual abilities to a large extent (e.g., 3-7%), and enhances the performance of MLLMs and vision models on multimodal understanding and segmentation tasks. Extensive evaluation on 29 image classification and retrieval benchmarks confirms that our framework preserves CLIP's strong zero-shot capabilities. The code will be publicly available soon.", "keywords": [ "CLIP Model", "Diffusion Model", "Generative Feedback", "Representation Learning" ], "rating_list": [ 8, 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 4, 3, 1, 2, 3 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 5, 3, 3, 4, 4 ] }, { "paper_id": "tL8dpJmECp", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Ali Siahkoohi", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Postdoc" }, { "name": "Don H. Johnson", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Emeritus" }, { "name": "Lorenzo Luzi", "gender": "Male", "institution": "Rice University", "country": "", "position": "PhD student" }, { "name": "Paul M Mayer", "gender": "Male", "institution": "Rice University", "country": "", "position": "PhD student" }, { "name": "Baraniuk R.", "gender": "unknown", "institution": "William Marsh Rice University", "country": "", "position": "C. Sidney Burrus Professor" } ], "n_formula": 50, "n_formula_1": 12, "n_ref_uni": 58, "n_ref": 78, "n_ref_all": 111, "n_fig": 7, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 767, "n_element_tab_1": 100, "formula_len_all": 3840, "formula_len_all_1": 556, "len_all": 178237, "len_all_1": 58554, "len_abs": 956, "len_title": 62, "len_sents": 52934, "len_sents_1": 27489, "n_sents": 416, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 960, "title": "Improving Fairness and Mitigating MADness in Generative Models", "abs": "Generative models unfairly penalize data belonging to minority classes, suffer from model autophagy disorder (MADness), and learn biased estimates of the underlying distribution parameters. Our theoretical and empirical results show that training generative models with intentionally designed hypernetworks leads to models that 1) are more fair when generating datapoints belonging to minority classes 2) are more stable in a self-consumed (i.e., MAD) setting, and 3) learn parameters that are less statistically biased. To further mitigate unfairness, MADness, and bias, we introduce a regularization term that penalizes discrepancies between a generative model\u2019s estimated weights when trained on real data versus its own synthetic data. To facilitate training existing deep generative models within our framework, we offer a scalable implementation of hypernetworks that automatically generates a hypernetwork architecture for any given generative model.", "keywords": [ "Hypernetworks", "Generative Models", "Fairness", "MADness", "Maximum Likelihood Estimation", "Bias" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "tKnPtyDt6H", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Graham Horwood", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Jie Ma", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Miguel Ballesteros", "gender": "Male", "institution": "Amazon", "country": "", "position": "Principal Applied Scientist" }, { "name": "Yang Li", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yassine Benajiba", "gender": "Not Specified", "institution": "Columbia University", "country": "", "position": "Postdoc" } ], "n_formula": 28, "n_formula_1": 23, "n_ref_uni": 41, "n_ref": 67, "n_ref_all": 84, "n_fig": 5, "n_tab": 7, "L_tab": 3750, "n_element_tab": 298, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1441, "n_element_tab_1": 91, "formula_len_all": 1503, "formula_len_all_1": 1072, "len_all": 163643, "len_all_1": 69090, "len_abs": 1161, "len_title": 108, "len_sents": 51090, "len_sents_1": 31960, "n_sents": 393, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1166, "title": "Active Evaluation Acquisition for Efficient LLM Benchmarking", "abs": "As large language models (LLMs) become increasingly versatile, numerous large scale benchmarks have been developed to thoroughly assess their capabilities. These benchmarks typically consist of diverse datasets and prompts to evaluate different aspects of LLM performance. However, comprehensive evaluations on hundreds or thousands of prompts incur tremendous costs in terms of computation, money, and time. In this work, we investigate strategies to improve evaluation efficiency by selecting a subset of examples from each benchmark using a learned policy. Our approach models the dependencies across test examples, allowing accurate prediction of the evaluation outcomes for the remaining examples based on the outcomes of the selected ones. Consequently, we only need to acquire the actual evaluation outcomes for the selected subset. We rigorously explore various subset selection policies and introduce a novel RL-based policy that leverages the captured dependencies. Empirical results demonstrate that our approach significantly reduces the number of evaluation prompts required while maintaining accurate performance estimates compared to previous methods.", "keywords": [ "Efficient LLM Evaluation", "Active Learning", "Subset Selection" ], "rating_list": [ 6, 8, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "tKn6gpvlUX", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Gabriele Scalia", "gender": "unknown", "institution": "Genentech", "country": "", "position": "Researcher" }, { "name": "Jan-Christian Huetter", "gender": "Male", "institution": "Genentech", "country": "US", "position": "Researcher" }, { "name": "Tommaso Biancalani", "gender": "Male", "institution": "Genentech", "country": "", "position": "Director" }, { "name": "Kangway V Chuang", "gender": "Male", "institution": "Genentech Research and Early Development", "country": "", "position": "Researcher" }, { "name": "sepideh maleki", "gender": "Female", "institution": "Genentech", "country": "US", "position": "Postdoc" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 43, "n_ref": 84, "n_ref_all": 109, "n_fig": 12, "n_tab": 3, "L_tab": 492, "n_element_tab": 48, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 226, "n_element_tab_1": 20, "formula_len_all": 386, "formula_len_all_1": 291, "len_all": 148507, "len_all_1": 62678, "len_abs": 3647, "len_title": 155, "len_sents": 40266, "len_sents_1": 30995, "n_sents": 314, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1246, "title": "Efficient Fine-Tuning of Single-Cell Foundation Models Enables Zero-Shot Molecular Perturbation Prediction", "abs": "Predicting transcriptional responses to novel drugs provides a unique opportunity to accelerate biomedical research and advance drug discovery efforts. However, the inherent complexity and high dimensionality of cellular responses, combined with the extremely limited available experimental data, makes the task challenging. In this study, we leverage single-cell foundation models (FMs) pretrained on tens of millions of single cells, encompassing multiple cell types, states, and disease annotations, to address molecular perturbation prediction. We introduce a drug-conditional adapter that allows efficient fine-tuning by training less than 1\\% of the original foundation model, thus enabling molecular conditioning while preserving the rich biological representation learned during pretraining. The proposed strategy allows not only the prediction of cellular responses to novel drugs, but also the zero-shot generalization to unseen cell types. We establish a robust evaluation framework to assess model performance across different generalization tasks, demonstrating state-of-the-art results across all settings, with significant improvements in the few-shot and zero-shot generalization to new cell types compared to existing baselines.", "keywords": [ "Foundation models", "fine-tuning", "parameter efficient fine-tuning", "adapters", "Molecular Perturbation", "drug discovery", "scgpt" ], "rating_list": [ 6, 3, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 1, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 3, 5, 3, 4, 4 ] }, { "paper_id": "tKif2rXQ6V", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Han-Seul Jeong", "gender": "unknown", "institution": "LG AI Research", "country": "", "position": "Researcher" }, { "name": "Hyungseok Song", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Researcher" }, { "name": "Kyunghyun Lee", "gender": "Male", "institution": "LG AI Research", "country": "", "position": "Researcher" }, { "name": "Sungdong Yoo", "gender": "Male", "institution": "Korea University", "country": "", "position": "MS student" }, { "name": "Woohyung Lim", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Vice President" }, { "name": "Youngjoon Park", "gender": "unknown", "institution": "LG AI Research", "country": "KR", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 3, "n_ref_uni": 43, "n_ref": 68, "n_ref_all": 87, "n_fig": 3, "n_tab": 14, "L_tab": 7302, "n_element_tab": 612, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 3461, "n_element_tab_1": 228, "formula_len_all": 410, "formula_len_all_1": 126, "len_all": 163023, "len_all_1": 63780, "len_abs": 1291, "len_title": 86, "len_sents": 53088, "len_sents_1": 28314, "n_sents": 370, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1296, "title": "SteBen: Steiner Tree Problem Benchmark for Neural Combinatorial Optimization on Graphs", "abs": "The Steiner Tree Problem (STP) is an NP-hard combinatorial optimization problem with applications in areas like network design and facility location. Despite its importance, learning-based solvers for STP have been hindered by the lack of large-scale, diverse datasets necessary to train and evaluate advanced neural models. To address this limitation, we introduce a standardized dataset comprising over a million high-quality instances with optimal solutions, spanning various problem sizes and graph structures. Our dataset enables benchmarking of neural combinatorial optimization methods across both supervised and reinforcement learning paradigms, encompassing autoregressive and non-autoregressive inference approaches. Our experiments show that supervised learning excels in in-distribution settings, while reinforcement learning generalizes better to unseen problem sizes, highlighting a trade-off between solution quality and generalization. We compare NCO methods across different STP scales and graph types, and demonstrate that solvers trained on our datasets generalize well to real-world instances without fine-tuning, proving its practical utility. We hope this benchmark promotes further STP research and advances NCO techniques for broader combinatorial optimization challenges.", "keywords": [ "Dataset & benchmark", "Neural Combinatorial Optimization", "Steiner Tree Problem" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "tKFZ53nerQ", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Jheng-Long Wu", "gender": "Not Specified", "institution": "Soochow University", "country": "TW", "position": "Assistant Professor" }, { "name": "PAN HONG-RUI", "gender": "unknown", "institution": "Soochow University", "country": "TW", "position": "MS student" }, { "name": "Tong-Ru Wu", "gender": "Male", "institution": "Soochow University", "country": "TW", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 10, "n_ref": 10, "n_ref_all": 18, "n_fig": 2, "n_tab": 4, "L_tab": 938, "n_element_tab": 89, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 938, "n_element_tab_1": 89, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 48849, "len_all_1": 48849, "len_abs": 856, "len_title": 125, "len_sents": 16686, "len_sents_1": 16686, "n_sents": 143, "n_sents_1": 143, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 860, "title": "Topic and Description Reasoning Generation based on User-Contributed Comments", "abs": "We propose Topic and Description Reasoning Generation (TDRG), a text inference and generation method based on user-contributed comments with large language models (LLMs). Unlike summarization methods, TDRG can infer the topic according to comments contributed by different users, and generate a readable description that addresses the issue of the lack of interpretability in traditional topic modeling for text mining. In this paper, we adopted zero-shot and fine-tuning methods to generate topics and descriptions for comments. We use a human-annotated YouTube comment dataset to evaluate performance. Our results demonstrate that the potential of large language models of reasoning the topic and description. Generated topic titles and descriptions are similar to human references in textual semantics, but the words used are different from those of humans.", "keywords": [ "topic modeling", "topic reasoning", "large language models" ], "rating_list": [ 3, 1, 3, 1 ], "soundness_list": [ 2, 1, 2, 1 ], "presentation_list": [ 2, 1, 3, 1 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 2, 5, 4, 4 ] }, { "paper_id": "tJoS2d0Onf", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Chun-Han Yao", "gender": "Male", "institution": "Stability AI", "country": "US", "position": "Researcher" }, { "name": "Huaizu Jiang", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Assistant Professor" }, { "name": "Vikram Voleti", "gender": "Male", "institution": "Stability AI", "country": "CA", "position": "Researcher" }, { "name": "Yiming Xie", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Varun Jampani", "gender": "Male", "institution": "Stability AI", "country": "US", "position": "Principal Researcher" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 59, "n_ref": 166, "n_ref_all": 201, "n_fig": 14, "n_tab": 6, "L_tab": 1447, "n_element_tab": 137, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1297, "n_element_tab_1": 102, "formula_len_all": 380, "formula_len_all_1": 0, "len_all": 192896, "len_all_1": 65537, "len_abs": 997, "len_title": 127, "len_sents": 44921, "len_sents_1": 31230, "n_sents": 354, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1003, "title": "SV4D: Dynamic 3D Content Generation with Multi-Frame and Multi-View Consistency", "abs": "We present Stable Video 4D (SV4D) \u2014 a latent video diffusion model for multi-frame and multi-view consistent dynamic 3D content generation. Unlike previous methods that rely on separately trained generative models for video generation and novel view synthesis, we design a unified diffusion model to generate novel view videos of dynamic 3D objects. Specifically, given a monocular reference video, SV4D generates novel views for each video frame that are temporally consistent. We then use the generated novel view videos to optimize an implicit 4D representation (dynamic NeRF) efficiently, without the need for cumbersome SDS-based optimization used in most prior works. To train our unified novel view video generation model, we curated a dynamic 3D object dataset from the existing Objaverse dataset. Extensive experimental results on multiple datasets and user studies demonstrate SV4D's state-of-the-art performance on novel-view video synthesis as well as 4D generation compared to prior works.", "keywords": [ "Generative models", "4D generation" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "tJHDw8XfeC", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Fandong Meng", "gender": "Male", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Hao Zhou", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Jie Zhou", "gender": "Male", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Minlie Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yuxian Gu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 8, "n_ref_uni": 71, "n_ref": 140, "n_ref_all": 210, "n_fig": 11, "n_tab": 7, "L_tab": 955, "n_element_tab": 128, "n_fig_1": 9, "n_tab_1": 6, "L_tab_1": 1291, "n_element_tab_1": 246, "formula_len_all": 2565, "formula_len_all_1": 703, "len_all": 211797, "len_all_1": 70463, "len_abs": 2035, "len_title": 112, "len_sents": 48886, "len_sents_1": 30420, "n_sents": 358, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1677, "title": "MiniPLM: Knowledge Distillation for Pre-training Language Models", "abs": "Knowledge distillation (KD) is widely used to train small, high-performing student language models (LMs) using large teacher LMs. \nWhile effective in fine-tuning, KD during pre-training faces challenges in efficiency, flexibility, and effectiveness. \nExisting methods either incur high computational costs due to online teacher inference, require tokenization matching between teacher and student LMs, or risk losing the difficulty and diversity of the teacher-generated training data.\nTo address these issues, we propose MiniPLM, a KD framework for pre-training LMs by refining the training data distribution with the teacher LM's knowledge.\nFor efficiency, MiniPLM performs offline teacher LM inference, allowing KD for multiple student LMs without adding training-time costs.\nFor flexibility, MiniPLM operates solely on the pre-training corpus, enabling KD across model families.\nFor effectiveness, MiniPLM leverages the differences between large and small LMs to enhance the difficulty and diversity of the training data, helping student LMs acquire versatile and sophisticated knowledge.\nWe apply MiniPLM to pre-train LMs with 200M, 500M, and 1.2B from scratch, using a 1.8B teacher LM. \nExtensive experiments demonstrate that MiniPLM boosts the student LMs' performance on 9 widely used downstream tasks, improves the language modeling capabilities, and reduces pre-training computation. \nThe benefit of MiniPLM extends to larger pre-training scales, evidenced by the extrapolation of the scaling curves.\nFurther analysis reveals that MiniPLM supports KD across model families and enhances the utilization of pre-training data. We will release our code, data, and models.", "keywords": [ "Language Models", "Knowledge Distillation", "Pre-Training" ], "rating_list": [ 6, 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 3, 4, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3, 4 ] }, { "paper_id": "tJE9WeqHEI", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Bo Bai", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Lei Deng", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Wei Han", "gender": "Male", "institution": "Huawei Tech. Investment Co., Limited", "country": "HK", "position": "Principal Researcher" }, { "name": "Xueyan Niu", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" } ], "n_formula": 53, "n_formula_1": 35, "n_ref_uni": 49, "n_ref": 140, "n_ref_all": 157, "n_fig": 6, "n_tab": 3, "L_tab": 274, "n_element_tab": 26, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4228, "formula_len_all_1": 2089, "len_all": 187030, "len_all_1": 68979, "len_abs": 1142, "len_title": 131, "len_sents": 52962, "len_sents_1": 30903, "n_sents": 444, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1148, "title": "Beyond Scaling Laws: Understanding Transformer Performance with Associative Memory", "abs": "Increasing the size of a Transformer does not always lead to enhanced performance. This phenomenon cannot be explained by the empirical scaling laws. Furthermore, the model's enhanced performance is closely associated with its memorization of the training samples. \nWe present a theoretical framework that sheds light on the memorization during pre-training of transformer-based language models. We model the behavior of Transformers with associative memories using Hopfield networks, such that each transformer block effectively conducts an approximate nearest-neighbor search. Based on this, we use a distance-based energy function to approximate the one in the modern continuous Hopfield network, which provides an insightful explanation for the attention mechanism. Since the softmax function corresponds to the gradient of the LogSumExp function in the energy, using the majorization-minimization technique, we construct a global energy function to capture the layered architecture. We show a dependency between the model size and the dataset for the model to attain optimal performance, and the achievable cross-entropy loss is bounded below.", "keywords": [ "Transformer; Associative Memory; Energy-Based Model" ], "rating_list": [ 8, 6, 8, 5 ], "soundness_list": [ 4, 4, 4, 2 ], "presentation_list": [ 3, 4, 4, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "tIURLNBTPx", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Hanzhong Cao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "WenBo Yan", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Ying Tan", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 19, "n_ref": 29, "n_ref_all": 51, "n_fig": 6, "n_tab": 10, "L_tab": 9501, "n_element_tab": 631, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2347, "n_element_tab_1": 202, "formula_len_all": 856, "formula_len_all_1": 869, "len_all": 142616, "len_all_1": 51500, "len_abs": 1472, "len_title": 134, "len_sents": 46108, "len_sents_1": 22304, "n_sents": 266, "n_sents_1": 150, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1238, "title": "Repetitive Contrastive Learning Enhances Mamba's Selectivity in Time Series Prediction", "abs": "The prediction of long sequences has always been a challenge in time series forecasting tasks. Due to Mamba's sequence selection capability, many Mamba-based models have been proposed, achieving state-of-the-art results in long sequence prediction problems. However, much research has focused on integrating mamba-ssm into specific model structures for better performance, while the core of mamba-ssm, its sequence selection capability, has not been deeply explored. We believe there is significant potential in Mamba's sequence selection capability and propose a Repetitive Contrastive Learning (RCL) method to enhance it. Specifically, we utilize Repeating Sequence Augmentation to expand the sequence while introducing Gaussian noise, thereby enhancing the Mamba block's sequence selection capability through both inter-sequence and intra-sequence contrastive techniques. Then our methods transfer parameters directly from a single pretrained Mamba block to a variety of Mamba-based models. This approach provides superior initialization for forecasting tasks. Our experiments consistently demonstrate that this technique improves the forecasting performance of many Mamba-based models, without imposing additional memory requirements.", "keywords": [ "Mamba; Self-supervised learning; Time Series Prediction" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 1, 1, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "tIBAOcAvn4", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chen Ma", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "Researcher" }, { "name": "Qi Xuan", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Shuyu Cheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "xinjie Xu", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 168, "n_formula_1": 27, "n_ref_uni": 22, "n_ref": 56, "n_ref_all": 241, "n_fig": 19, "n_tab": 10, "L_tab": 4006, "n_element_tab": 499, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 11311, "formula_len_all_1": 2174, "len_all": 242216, "len_all_1": 63990, "len_abs": 1455, "len_title": 126, "len_sents": 71714, "len_sents_1": 27914, "n_sents": 666, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 78, "L_abs": 1467, "title": "Boosting Ray Search Procedure of Hard-label Attacks with Transfer-based Priors", "abs": "One of the most practical and challenging types of black-box adversarial attacks is the hard-label attack, where only top-1 predicted labels are available. One effective approach is to search for the optimal ray direction from the benign image that minimizes the $\\ell_p$ norm distance to the adversarial region. The unique advantage of this approach is that it transforms the hard-label attack into a continuous optimization problem. The objective function value is the ray's radius and can be obtained through a binary search with high query cost. Existing methods use a \"sign trick\" in gradient estimation to reduce queries. In this paper, we theoretically analyze the quality of this gradient estimation, proposing a novel prior-guided approach to improve ray search efficiency, based on theoretical and experimental analysis. Specifically, we utilize the transfer-based priors from surrogate models, and our gradient estimators appropriately integrate them by approximating the projection of the true gradient onto the subspace spanned by these priors and some random directions, in a query-efficient way. We theoretically derive the expected cosine similarity between the obtained gradient estimators and the true gradient, and demonstrate the improvement brought by using priors. Extensive experiments on the ImageNet and CIFAR-10 datasets show that our approach significantly outperforms 11 state-of-the-art methods in query efficiency. Code will be released.", "keywords": [ "adversarial attack", "hard-label attack", "black-box attack", "gradient estimation", "surrogate model" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "tH12wjcuXx", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Anton Hinel", "gender": "Male", "institution": "American Express", "country": "US", "position": "Principal Researcher" }, { "name": "Francesco Sanna Passino", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "Lecturer" }, { "name": "Leonie Tabea Goldmann", "gender": "Female", "institution": "American Express", "country": "US", "position": "Researcher" }, { "name": "Tommaso Ferracci", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 23, "n_ref": 37, "n_ref_all": 90, "n_fig": 23, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 15, "n_tab_1": 4, "L_tab_1": 1657, "n_element_tab_1": 30, "formula_len_all": 701, "formula_len_all_1": 570, "len_all": 105265, "len_all_1": 63877, "len_abs": 953, "len_title": 129, "len_sents": 37929, "len_sents_1": 29535, "n_sents": 228, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 957, "title": "Targeted synthetic data generation for tabular data via hardness characterization", "abs": "Synthetic data generation has been proven successful in improving model performance and robustness in the context of scarce or low-quality data. Using the data valuation framework to statistically identify beneficial and detrimental observations, we introduce a novel augmentation pipeline that generates only high-value training points based on hardness characterization. We first demonstrate via benchmarks on real data that Shapley-based data valuation methods perform comparably with learning-based methods in hardness characterisation tasks, while offering significant theoretical and computational advantages. Then, we show that synthetic data generators trained on the hardest points outperform non-targeted data augmentation on simulated data and on a large scale credit default prediction task. In particular, our approach improves the quality of out-of-sample predictions and it is computationally more efficient compared to non-targeted methods.", "keywords": [ "data Shapley", "hardness characterization", "synthetic data generation" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 2, 3 ] }, { "paper_id": "tGsumqfOUk", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Cem \u00dcy\u00fck", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "MS student" }, { "name": "Mike Lasby", "gender": "Male", "institution": "University of Calgary", "country": "CA", "position": "PhD student" }, { "name": "Mohamed Yassin", "gender": "Male", "institution": "University of Calgary", "country": "CA", "position": "Researcher" }, { "name": "Utku Evci", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Yani Ioannou", "gender": "Male", "institution": "University of Calgary", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 6, "n_ref_uni": 26, "n_ref": 39, "n_ref_all": 67, "n_fig": 13, "n_tab": 5, "L_tab": 1499, "n_element_tab": 94, "n_fig_1": 13, "n_tab_1": 5, "L_tab_1": 1340, "n_element_tab_1": 92, "formula_len_all": 494, "formula_len_all_1": 569, "len_all": 104751, "len_all_1": 58352, "len_abs": 886, "len_title": 114, "len_sents": 31982, "len_sents_1": 26608, "n_sents": 222, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 893, "title": "Learning Parameter Sharing with Tensor Decompositions and Sparsity", "abs": "Large neural networks achieve remarkable performance, but their size hinders deployment on resource-constrained devices. While various compression techniques exist, parameter sharing remains relatively unexplored. This paper introduces Fine-grained Parameter Sharing (FiPS), a novel algorithm that leverages the relationship between parameter sharing, tensor decomposition, and sparsity to efficiently compress large vision transformer models. FiPS employs a shared base and sparse factors to represent shared neurons across multi-layer perception (MLP) modules. Shared parameterization is initialized via Singular Value Decomposition (SVD) and optimized by minimizing block-wise reconstruction error. Experiments demonstrate that FiPS compresses DeiT-B and Swin-L MLPs to 25-40% of their original parameter count while maintaining accuracy within one percentage point of the original models.", "keywords": [ "Model Compression", "Parameter Sharing", "Sparsity", "Tensor Decomposition", "Knowledge Distillation", "Transfer Learning", "Vision Transformers", "Transformers" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "tGYFikNONB", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Beatrice Bevilacqua", "gender": "Female", "institution": "Purdue University", "country": "", "position": "PhD student" }, { "name": "Bruno Ribeiro", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "Jure Leskovec", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Joshua David Robinson", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Postdoc" } ], "n_formula": 72, "n_formula_1": 19, "n_ref_uni": 80, "n_ref": 140, "n_ref_all": 185, "n_fig": 3, "n_tab": 6, "L_tab": 2775, "n_element_tab": 192, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 142, "n_element_tab_1": 5, "formula_len_all": 9779, "formula_len_all_1": 1747, "len_all": 310868, "len_all_1": 71488, "len_abs": 1388, "len_title": 124, "len_sents": 99603, "len_sents_1": 34846, "n_sents": 772, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1394, "title": "Holographic Node Representations: Pre-training Task-Agnostic Node Embeddings", "abs": "Large general purpose pre-trained models have revolutionized computer vision and natural language understanding. However, the development of general purpose pre-trained Graph Neural Networks (GNNs) lags behind other domains due to the lack of suitable generalist node representations. Existing GNN architectures are often tailored to specific task orders, such as node-level, link-level, or higher-order tasks, because different tasks require distinct permutation symmetries, which are difficult to reconcile within a single model. In this paper, we propose _holographic node representations_, a new blueprint for node representations capable of solving tasks of any order. Holographic node representations have two key components: (1) a task-agnostic expansion map, which produces highly expressive, high-dimensional embeddings, free from node-permutation symmetries, to be fed into (2) a reduction map that carefully reintroduces the relevant permutation symmetries to produce low-dimensional, task-specific embeddings. We show that well-constructed expansion maps enable simple and efficient reduction maps, which can be adapted for any task order. Empirical results show that holographic node representations can be effectively pre-trained and reused across tasks of varying orders, yielding up to 100% relative performance improvement, including in cases where prior methods fail entirely.", "keywords": [ "GNN", "symmetries", "pretraining" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "tG4SgayTtk", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Jason E Weston", "gender": "Not Specified", "institution": "New York University", "country": "US", "position": "Researcher" }, { "name": "Sainbayar Sukhbaatar", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Research Scientist" }, { "name": "Shibo Hao", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Xian Li", "gender": "unknown", "institution": "Facebook AI", "country": "", "position": "Principal Researcher" }, { "name": "Yuandong Tian", "gender": "Male", "institution": "Meta AI (FAIR)", "country": "", "position": "Research Scientist" }, { "name": "Zhiting Hu", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "andy su", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 40, "n_ref": 65, "n_ref_all": 90, "n_fig": 8, "n_tab": 3, "L_tab": 650, "n_element_tab": 83, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 458, "n_element_tab_1": 59, "formula_len_all": 64, "formula_len_all_1": 64, "len_all": 127923, "len_all_1": 71260, "len_abs": 1575, "len_title": 117, "len_sents": 40545, "len_sents_1": 34376, "n_sents": 322, "n_sents_1": 272, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1582, "title": "Training Large Language Model to Reason in a Continuous Latent Space", "abs": "Large language models are restricted to reason in the \u201clanguage space\u201d, where they typically express the reasoning process with a chain-of-thoughts (CoT) to solve a complex reasoning problem. However, we argue that language space may not be the optimal reasoning space. For example, most word tokens are primarily for textual coherence and not essential for reasoning, while some critical tokens require complex planning and pose huge challenges to LLMs. To explore the potential of LLM reasoning in an unrestricted latent space instead of using human language, we introduce a new paradigm COCONUT (Chain of Continuous Thought). We utilize the last hidden state of the LLM as a representation of the reasoning state (termed \u201ccontinuous thought\u201d). Rather than decoding this into a word token, we feed it back to the LLM as the subsequent input embedding directly in the continuous space. Experiments show that COCONUT can effectively augment the LLM on several reasoning tasks. It even outperforms CoT in certain logical reasoning tasks that require substantial planning, despite generating fewer tokens during inference. More interestingly, we observe an advanced reasoning patterns emerging from latent reasoning: the continuous thought can encode multiple potential next reasoning steps, allowing the model to perform a breadth-first search (BFS) to solve the problem, rather than prematurely committing to a single deterministic path like CoT. These findings demonstrate the promise of latent reasoning and offer valuable insights for future research on latent reasoning methods.", "keywords": [ "large language model", "reasoning", "chain of thoughts" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 4, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "tFwEsrx1hm", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Jonathan Berant", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Maor Ivgi", "gender": "Male", "institution": "Tel Aviv University", "country": "", "position": "PhD student" }, { "name": "Mor Geva", "gender": "Female", "institution": "Tel Aviv University", "country": "IL", "position": "Assistant Professor" }, { "name": "Ori Yoran", "gender": "Male", "institution": "Tel Aviv University", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 0, "n_ref_uni": 39, "n_ref": 119, "n_ref_all": 185, "n_fig": 28, "n_tab": 7, "L_tab": 5816, "n_element_tab": 108, "n_fig_1": 20, "n_tab_1": 5, "L_tab_1": 1960, "n_element_tab_1": 82, "formula_len_all": 449, "formula_len_all_1": 0, "len_all": 300329, "len_all_1": 69100, "len_abs": 1276, "len_title": 123, "len_sents": 71655, "len_sents_1": 33485, "n_sents": 494, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1284, "title": "From Loops to Oops: Fallback Behaviors of Language Models Under Uncertainty", "abs": "Large language models (LLMs) often exhibit undesirable behaviors, such as hallucinations and sequence repetitions.\nWe propose to view these behaviors as fallbacks that models exhibit under epistemic uncertainty, and investigate the connection between them.\nWe categorize fallback behaviors \u2014 sequence repetitions, degenerate text, and hallucinations \u2014 and extensively analyze them in models from the same family that differ by the amount of pretraining tokens, parameter count, or the inclusion of instruction-following training.\nOur experiments reveal a clear and consistent ordering of fallback behaviors, across all these axes: \nthe more advanced an LLM is (i.e., trained on more tokens, has more parameters, or instruction-tuned), \nits fallback behavior shifts from sequence repetitions, to degenerate text, and then to hallucinations.\nMoreover, the same ordering is observed during the generation of a single sequence, even for the best-performing models; as uncertainty increases, models shift from generating hallucinations to producing degenerate text and finally sequence repetitions. \nLastly, we demonstrate that while common decoding techniques, such as random sampling, alleviate unwanted behaviors like sequence repetitions, they increase harder-to-detect hallucinations.", "keywords": [ "Hallucinations", "Epistemic uncertainty", "Degenerate text", "Fallback behaviors" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 3, 2, 4 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "tFV5GrWOGm", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Aleksandra Faust", "gender": "Female", "institution": "Google Brain", "country": "US", "position": "Principal Researcher" }, { "name": "Matei Zaharia", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Pieter Abbeel", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Volodymyr Mnih", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Wilson Yan", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Hao Liu", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 41, "n_ref": 62, "n_ref_all": 83, "n_fig": 9, "n_tab": 5, "L_tab": 420, "n_element_tab": 90, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 381, "n_element_tab_1": 60, "formula_len_all": 807, "formula_len_all_1": 885, "len_all": 127404, "len_all_1": 56634, "len_abs": 1191, "len_title": 53, "len_sents": 35768, "len_sents_1": 26981, "n_sents": 256, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1160, "title": "ElasticTok: Adaptive Tokenization for Image and Video", "abs": "Efficient video tokenization remains a key bottleneck in learning general purpose vision models that are capable of processing long video sequences. Prevailing approaches are restricted to encoding videos to a fixed number of tokens, where too few tokens will result in overly lossy encodings, and too many tokens will result in prohibitively long sequence lengths. In this work, we introduce ElasticTok, a method that conditions on prior frames to adaptively encode a frame into a variable number of tokens. To enable this in a computationally scalable way, we propose a masking technique that drops a random number of tokens at the end of each frames's token encoding. During inference, ElasticTok can dynamically allocate tokens when needed -- more complex data can leverage more tokens, while simpler data only needs a few tokens. Our empirical evaluations on images and video demonstrate the effectiveness of our approach in efficient token usage, paving the way for future development of more powerful multimodal models, world models, and agents. Video examples of using ElasticTok can be found on our website: http://largeworldmodel.github.io/elastictok", "keywords": [ "adaptive representation", "adaptive tokenization", "autoencoder" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "tFDTHA3odg", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Changliu Liu", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Guanxing Lu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Jiwen Lu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yansong Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ziwei Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 25, "n_ref": 47, "n_ref_all": 73, "n_fig": 7, "n_tab": 6, "L_tab": 1792, "n_element_tab": 335, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1551, "n_element_tab_1": 141, "formula_len_all": 73, "formula_len_all_1": 73, "len_all": 159266, "len_all_1": 61571, "len_abs": 1234, "len_title": 117, "len_sents": 46852, "len_sents_1": 29628, "n_sents": 320, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1237, "title": "ThinkBot: Embodied Instruction Following with Thought Chain Reasoning", "abs": "Embodied Instruction Following (EIF) requires agents to complete human instruction by interacting objects in complicated surrounding environments. Conventional methods directly consider the sparse human instruction to generate action plans for agents, which usually fail to achieve human goals because of the instruction incoherence in action descriptions. On the contrary, we propose ThinkBot that reasons the thought chain in human instruction to recover the missing action descriptions, so that the agent can successfully complete human goals by following the coherent instruction. Specifically, we first design an instruction completer based on large language models to recover the missing actions with interacted objects between consecutive human instruction, where the perceived surrounding environments and the completed sub-goals are considered for instruction completion. Based on the partially observed scene semantic maps, we present an object localizer to infer the position of interacted objects for agents to achieve complex human goals. Extensive experiments in the simulated environment show that our ThinkBot outperforms the state-of-the-art EIF methods by a sizable margin in both success rate and execution efficiency.", "keywords": [ "Embodied Instruction Following (EIF)", "Large Language Model", "Chain-of-thought Reasoning" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "tErHYBGlWc", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Christopher G. Lucas", "gender": "unknown", "institution": "University of Edinburgh, University of Edinburgh", "country": "", "position": "Assistant Professor" }, { "name": "David Abel", "gender": "Male", "institution": "University of Edinburgh", "country": "GB", "position": "Honorary Fellow" }, { "name": "Pablo Samuel Castro", "gender": "Male", "institution": "Google", "country": "CA", "position": "Researcher" }, { "name": "Prakash Panangaden", "gender": "Male", "institution": ", McGill University", "country": "", "position": "Full Professor" }, { "name": "Samuel Garcin", "gender": "Male", "institution": "University of Edinburgh", "country": "", "position": "PhD student" }, { "name": "Stefano V. Albrecht", "gender": "unknown", "institution": "University of Edinburgh", "country": "", "position": "Assistant Professor" }, { "name": "Trevor McInroe", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 17, "n_ref_uni": 44, "n_ref": 74, "n_ref_all": 110, "n_fig": 12, "n_tab": 6, "L_tab": 3038, "n_element_tab": 315, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 257, "n_element_tab_1": 3, "formula_len_all": 4264, "formula_len_all_1": 1280, "len_all": 176828, "len_all_1": 72305, "len_abs": 1117, "len_title": 141, "len_sents": 46960, "len_sents_1": 30602, "n_sents": 396, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1125, "title": "Studying the Interplay Between the Actor and Critic Representations in Reinforcement Learning", "abs": "Extracting relevant information from a stream of high-dimensional observations is a central challenge for deep reinforcement learning agents. Actor-critic algorithms add further complexity to this challenge, as it is often unclear whether the same information will be relevant to both the actor and the critic. To this end, we here explore the principles that underlie effective representations for an actor and for a critic. We focus our study on understanding whether an actor and a critic will benefit from a decoupled, rather than shared, representation. \nOur primary finding is that when decoupled, the representations for the actor and critic systematically specialise in extracting different types of information from the environment---the actor's representation tends to focus on action-relevant information, while the critic's representation specialises in encoding value and dynamics information. Finally, we demonstrate how these insights help select representation learning objectives that play into the actor's and critic's respective knowledge specialisations, and improve performance in terms of agent returns.", "keywords": [ "Reinforcement Learning", "Representation Learning", "Transfer Learning", "Information Theory" ], "rating_list": [ 8, 8, 6, 6, 6 ], "soundness_list": [ 3, 4, 2, 3, 3 ], "presentation_list": [ 2, 4, 1, 3, 2 ], "contribution_list": [ 3, 3, 2, 3, 4 ], "confidence_list": [ 3, 3, 3, 2, 3 ] }, { "paper_id": "tEei1bolt3", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Ajmal Saeed Mian", "gender": "Male", "institution": "University of Western Australia", "country": "AU", "position": "Professor" }, { "name": "Andong Deng", "gender": "unknown", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Chen Chen", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Lincoln Spencer", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Undergrad student" }, { "name": "Mohit Bansal", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" }, { "name": "Shoubin Yu", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Taojiannan Yang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Tongjia Chen", "gender": "Male", "institution": "University of Western Australia", "country": "AU", "position": "PhD student" }, { "name": "Yapeng Tian", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 64, "n_ref": 135, "n_ref_all": 180, "n_fig": 17, "n_tab": 9, "L_tab": 2426, "n_element_tab": 231, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 2086, "n_element_tab_1": 169, "formula_len_all": 80, "formula_len_all_1": 0, "len_all": 223254, "len_all_1": 68807, "len_abs": 1726, "len_title": 132, "len_sents": 62919, "len_sents_1": 34494, "n_sents": 467, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1732, "title": "Motion-Grounded Video Reasoning: Understanding and Perceiving Motion at Pixel Level", "abs": "In this paper, we introduce Motion-Grounded Video Reasoning, a new motion understanding task that requires generating visual answers (video segmentation masks) according to the input question, and hence needs implicit spatiotemporal reasoning and grounding. This task extends existing spatiotemporal grounding work focusing on explicit action/motion grounding, to a more general format by enabling implicit reasoning via questions. To facilitate the development of the new task, we collect a large-scale dataset called GroundMoRe, which comprises 1,673 video clips, 243K object masks that are deliberately designed with 4 question types (Causal, Sequential, Counterfactual, and Descriptive) for benchmarking deep and comprehensive motion reasoning abilities. GroundMoRe uniquely requires models to generate visual answers, providing a more concrete and visually interpretable response than plain texts. It evaluates models on both spatiotemporal grounding and reasoning, fostering to address complex challenges in motion-related video reasoning, temporal perception, and pixel-level understanding. Furthermore, we introduce a novel baseline model named Motion-Grounded Video Reasoning Assistant (MoRA). MoRA incorporates the multimodal reasoning ability from the Multimodal LLM, the pixel-level perception capability from the grounding model (SAM), and the temporal perception ability from a lightweight localization head. MoRA achieves respectable performance on GroundMoRe outperforming the best existing visual grounding baseline model by an average of 28.8% relatively. We hope this novel and challenging task will pave the way for future advancements in robust and general motion understanding via video reasoning segmentation.", "keywords": [ "Motion", "Video Grounding", "Video Reasoning" ], "rating_list": [ 3, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "tE9gdaxHeB", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Alexandre Muzy", "gender": "Male", "institution": "CNRS", "country": "FR", "position": "Principal Researcher" }, { "name": "Ashwin Moongathottathil James", "gender": "Male", "institution": "Universit\u00e9 C\u00f4te d'Azur", "country": "FR", "position": "PhD student" }, { "name": "Ingrid Bethus", "gender": "unknown", "institution": "IPMC", "country": "FR", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 9, "n_ref_uni": 50, "n_ref": 68, "n_ref_all": 92, "n_fig": 10, "n_tab": 5, "L_tab": 1244, "n_element_tab": 237, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 821, "n_element_tab_1": 166, "formula_len_all": 1505, "formula_len_all_1": 812, "len_all": 132027, "len_all_1": 53208, "len_abs": 1785, "len_title": 131, "len_sents": 39468, "len_sents_1": 24798, "n_sents": 252, "n_sents_1": 155, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1805, "title": "Inferring Time-Varying Internal Models of Agents Through Dynamic Structure Learning", "abs": "Reinforcement learning (RL) models usually assume a stationary internal model structure of agents, which consists of fixed learning rules and environment representations. However, this assumption does not allow accounting for real problem solving by individuals who can exhibit irrational behaviors or hold inaccurate beliefs about their environment. In this work, we present a novel framework called Dynamic Structure Learning (DSL), which allows agents to adapt their learning rules and internal representations dynamically. This structural flexibility enables a deeper understanding of how individuals learn and adapt in real-world scenarios. The DSL framework reconstructs the most likely sequence of agent structures\u2014sourced from a pool of learning rules and environment models\u2014based on observed behaviors. The method provides insights into how an agent's internal structure model evolves as it transitions between different structures throughout the learning process. We applied our framework to study rat behavior in a maze task. Our results demonstrate that rats progressively refine their mental map of the maze, evolving from a suboptimal representation associated with repetitive errors to an optimal one that guides efficient navigation. Concurrently, their learning rules transition from heuristic-based to more rational approaches. These findings underscore the importance of both credit assignment and representation learning in complex behaviors. By going beyond simple reward-based associations, our research offers valuable insights into the cognitive mechanisms underlying decision-making in natural intelligence. DSL framework allows better understanding and modeling how individuals in real-world scenarios exhibit a level of adaptability that current AI systems have yet to achieve.", "keywords": [ "Decision Making", "Causal World Models", "Structure Learning", "Reinforcement Learning", "Cognitive Modeling", "Natural Intelligence" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 1, 3 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "tDIL7UXmSS", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Poojan Chetan Shah", "gender": "Male", "institution": "Indian Institute of Technology, Delhi", "country": "", "position": "Undergrad student" }, { "name": "Ragesh Jaiswal", "gender": "unknown", "institution": "Indian Institute of Technology Delhi", "country": "", "position": "Associate Professor" } ], "n_formula": 146, "n_formula_1": 13, "n_ref_uni": 25, "n_ref": 117, "n_ref_all": 151, "n_fig": 5, "n_tab": 9, "L_tab": 488, "n_element_tab": 61, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 78, "n_element_tab_1": 12, "formula_len_all": 13213, "formula_len_all_1": 1029, "len_all": 296274, "len_all_1": 77122, "len_abs": 1621, "len_title": 99, "len_sents": 97883, "len_sents_1": 34489, "n_sents": 978, "n_sents_1": 280, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1828, "title": "Quantum (Inspired) $D^2$-sampling with Applications", "abs": "$D^2$-sampling is a fundamental component of sampling-based clustering algorithms such as $k$-means++. \nGiven a dataset $V \\subset \\mathbb{R}^d$ with $N$ points and a center set $C \\subset \\mathbb{R}^d$, $D^2$-sampling refers to picking a point from $V$ where the sampling probability of a point is proportional to its squared distance from the nearest center in $C$.\nThe popular $k$-means++ algorithm is simply a $k$-round $D^2$-sampling process, which runs in $O(Nkd)$ time and gives $O(\\log{k})$-approximation in expectation for the $k$-means problem.\nIn this work, we give a quantum algorithm for (approximate) $D^2$-sampling in the QRAM model that results in a quantum implementation of $k$-means++ with a running time $\\tilde{O}(\\zeta^2 k^2)$. \nHere $\\zeta$ is the aspect ratio ( i.e., largest to smallest interpoint distance) and $\\tilde{O}$ hides polylogarithmic factors in $N, d, k$.\nIt can be shown through a robust approximation analysis of $k$-means++ that the quantum version preserves its $O(\\log{k})$ approximation guarantee.\nFurther, we show that our quantum algorithm for $D^2$-sampling can be dequantized using the sample-query access model of Tang (PhD Thesis, Ewin Tang, University of Washington, 2023). This results in a fast quantum-inspired classical implementation of $k$-means++, which we call QI-$k$-means++, with a running time $O(Nd) + \\tilde{O}(\\zeta^2k^2d)$, where the $O(Nd)$ term is for setting up the sample-query access data structure.\nExperimental investigations show promising results for QI-$k$-means++ on large datasets with bounded aspect ratio.\nFinally, we use our quantum $D^2$-sampling with the known $ D^2$-sampling-based classical approximation scheme \nto obtain the first quantum approximation scheme for the $k$-means problem with polylogarithmic running time dependence on $N$.", "keywords": [ "quantum", "clustering", "k-means++" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "tDANkt6X3D", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Hantian Ding", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Qing Sun", "gender": "Female", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Shiqi Wang", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Varun Kumar", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Yifeng Ding", "gender": "Male", "institution": "Department of Computer Science, University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Zijian Wang", "gender": "Male", "institution": "Amazon AWS AI Labs", "country": "US", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 19, "n_ref": 64, "n_ref_all": 89, "n_fig": 4, "n_tab": 4, "L_tab": 1096, "n_element_tab": 210, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 464, "n_element_tab_1": 95, "formula_len_all": 675, "formula_len_all_1": 723, "len_all": 145620, "len_all_1": 57278, "len_abs": 1854, "len_title": 160, "len_sents": 37872, "len_sents_1": 26861, "n_sents": 259, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 112, "L_abs": 1862, "title": "Horizon-Length Prediction: Advancing Fill-in-the-Middle Capabilities for Code Generation with Lookahead Planning", "abs": "Fill-in-the-Middle (FIM) has become integral to code language models, enabling generation of missing code given both left and right contexts. However, the current FIM training paradigm, which reorders original training sequences and then performs regular next-token prediction (NTP), often leads to models struggling to generate content that aligns smoothly with the surrounding context. Crucially, while existing works rely on rule-based post-processing to circumvent this weakness, such methods are not practically usable in open-domain code completion tasks as they depend on restrictive, dataset-specific assumptions (e.g., generating the same number of lines as in the ground truth). Moreover, model performance on FIM tasks deteriorates significantly without these unrealistic assumptions.\n\nWe hypothesize that NTP alone is insufficient for models to learn effective planning conditioned on the distant right context, a critical factor for successful code infilling. To overcome this, we propose Horizon-Length Prediction (HLP), a novel training objective that teaches models to predict the number of remaining middle tokens (i.e., horizon length) at each step. HLP advances FIM with lookahead planning, enabling models to inherently learn infilling boundaries for arbitrary left and right contexts without relying on dataset-specific post-processing. Our evaluation across different models and sizes shows that HLP significantly improves FIM performance by up to 24% relatively on diverse benchmarks, across file-level and repository-level, and without resorting to unrealistic post-processing methods. Furthermore, the enhanced planning capability gained through HLP boosts model performance on code reasoning. Importantly, HLP only incurs negligible training overhead and no additional inference cost, ensuring its practicality for real-world scenarios.", "keywords": [ "Language Models", "Fill-in-the-Middle", "Planning", "Training" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "tCmIGJivc7", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Linjuan Wu", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Weiming Lu", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 27, "n_ref": 71, "n_ref_all": 101, "n_fig": 9, "n_tab": 3, "L_tab": 378, "n_element_tab": 57, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1118, "n_element_tab_1": 20, "formula_len_all": 1147, "formula_len_all_1": 929, "len_all": 141813, "len_all_1": 61626, "len_abs": 1249, "len_title": 142, "len_sents": 37088, "len_sents_1": 28016, "n_sents": 260, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1257, "title": "Topic-XICL: Demonstration Selection with Topic Inference for Cross-lingual In-context Learning", "abs": "Cross-lingual in-context learning (XICL) shows promise for adapting large language models (LLMs) to low-resource languages. Previous methods typically rely on off-the-shelf similarity-based approaches or task-specific retrievers trained with LLM feedback for demonstration selection. However, these methods often overlook important factors beyond a single criterion or can be resource-intensive. To address these challenges, we propose a novel approach called Topic-XICL, which leverages a latent topic model for demonstration selection. We assume that latent topic variables encapsulate information that more accurately characterizes demonstrations. By training this topic model on rich-resource language data with a compact LLM, we obtain more relevant demonstrations through topic inference and apply them for in-context learning across various LLMs. We evaluated our method on three multilingual tasks (XNLI, XCOPA, and TyDiQA-GoldP) using three models with 7 to 8 billion parameters (BLOOM, Qwen1.5, and Llama3.1). Our approach outperformed the baselines\u2014random selection, semantic similarity, and clustering-based methods\u2014on TyDiQA-GoldP, XCOPA, and XNLI by 3.32\\%, 2.47\\%, and 1.77\\%, respectively, while requiring only moderate additional resources.", "keywords": [ "Multilingual and Cross-lingual", "In-context Learning", "Demonstration Selection for ICL" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "tCfvktlrHI", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Austen Liao", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Dan Klein", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Nicholas Tomlin", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 49, "n_ref": 90, "n_ref_all": 115, "n_fig": 13, "n_tab": 3, "L_tab": 1070, "n_element_tab": 106, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 481, "n_element_tab_1": 80, "formula_len_all": 138, "formula_len_all_1": 114, "len_all": 186066, "len_all_1": 64718, "len_abs": 1178, "len_title": 106, "len_sents": 50256, "len_sents_1": 33102, "n_sents": 339, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1184, "title": "Efficacy of Language Model Self-Play in Non-Zero-Sum Games", "abs": "Game-playing agents like AlphaGo have achieved superhuman performance through self-play, which is theoretically guaranteed to yield optimal policies in competitive games. However, most language tasks are partially or fully cooperative, so it is an open question whether techniques like self-play can effectively be used to improve language models. We empirically investigate this question in a negotiation game setting known as Deal or No Deal (DoND). Crucially, the objective in DoND can be modified to produce a fully cooperative game, a strictly competitive one, or anything in between. We finetune language models in self-play over multiple rounds of filtered behavior cloning in DoND for each of these objectives and evaluate them in self-play and in collaboration with humans. We find that language models improve substantially in self-play, achieving 14-17x higher scores in task reward after finetuning. Further, the trained models generalize to both cooperation and competition with humans, scoring 2.5-6x higher than base models. We view these results as an early promising sign for language model self-play in cooperative settings, despite a lack of theoretical guarantees.", "keywords": [ "language models", "self-play", "multi-agent", "dialogue", "reasoning" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 4, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "tCYdsuQgZZ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Amaia Cardiel", "gender": "unknown", "institution": "Universit\u00e9 Grenoble Alpes", "country": "FR", "position": "PhD student" }, { "name": "Andrei Bursuc", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Researcher" }, { "name": "Monika Wysocza\u0144ska", "gender": "Female", "institution": "Google", "country": "FR", "position": "Intern" }, { "name": "Oriane Sim\u00e9oni", "gender": "unknown", "institution": "Meta FAIR", "country": "FR", "position": "Research Scientist" }, { "name": "Renaud Marlet", "gender": "Male", "institution": "Inria", "country": "FR", "position": "Researcher" }, { "name": "Tomasz Trzcinski", "gender": "Male", "institution": "Warsaw University of Technology", "country": "PL", "position": "Full Professor" }, { "name": "Anton\u00edn Vobeck\u00fd", "gender": "Male", "institution": "Czech Technical University in Prague", "country": "CZ", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 65, "n_ref": 152, "n_ref_all": 174, "n_fig": 15, "n_tab": 11, "L_tab": 2374, "n_element_tab": 190, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 821, "n_element_tab_1": 44, "formula_len_all": 70, "formula_len_all_1": 51, "len_all": 194603, "len_all_1": 71483, "len_abs": 1225, "len_title": 115, "len_sents": 52851, "len_sents_1": 32271, "n_sents": 426, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1233, "title": "Test-time Contrastive Concepts for Open-World Semantic Segmentation", "abs": "Recent CLIP-like Vision-Language Models (VLMs), pre-trained on large amounts of image-text pairs to align both modalities with a simple contrastive objective, have paved the way to open-vocabulary semantic segmentation. Given an arbitrary set of textual queries, image pixels are assigned the closest query in feature space. However, this works well when a user exhaustively lists all possible visual concepts in an image, which contrast against each other for the assignment. This corresponds to the current evaluation setup in the literature which relies on having access to a list of in-domain relevant concepts, typically classes of a benchmark dataset. Here, we consider the more challenging (and realistic) scenario of segmenting a single concept, given a textual prompt and nothing else. To achieve good results, besides contrasting with the generic ``background'' text, we propose two different approaches to automatically generate, at test time, textual contrastive concepts that are query-specific. We do so by leveraging the distribution of text in the VLM's training set or crafted LLM prompts. We also propose a metric designed to evaluate this scenario and show the relevance of our approach on commonly used datasets.", "keywords": [ "image segmentation", "unsupervised learning", "open-vocabulary semantic segmentation" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "tC1b9DBWww", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Emma Drobina", "gender": "unknown", "institution": "University of Florida", "country": "US", "position": "PhD student" }, { "name": "Jean D Louis", "gender": "unknown", "institution": "University of Florida", "country": "US", "position": "PhD student" }, { "name": "Kiana Alikhademi", "gender": "Female", "institution": "University of Florida", "country": "", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 1, "n_ref_uni": 28, "n_ref": 62, "n_ref_all": 88, "n_fig": 9, "n_tab": 24, "L_tab": 8095, "n_element_tab": 1336, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 485, "n_element_tab_1": 10, "formula_len_all": 1434, "formula_len_all_1": 45, "len_all": 160012, "len_all_1": 63247, "len_abs": 1265, "len_title": 101, "len_sents": 44835, "len_sents_1": 32152, "n_sents": 341, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1272, "title": "Person Detection Through the Lens of Algorithmic Bias", "abs": "The rise of AI based person detection in safety critical applications such as driver-less cars or security monitoring has lead to an explosion of machine learning models and dataset research. At the same time, researchers have raised question of bias in these models and datasets. Popular benchmark datasets like More Inclusive Images for People (MIAP) and Berkeley DeepDrive (BDD) for person detection suffer from both sampling and labeling biases. This has serious implications for autonomous vehicles and other fields that use these datasets. We conduct an all-encompassing analysis to assess these datasets through the lens of algorithmic bias, looking at both dataset and model bias. To the best of our knowledge, no study has delved into the realm of person detection in low-quality or crowded pictures with this lens. The result is a novel analysis of bias in a real-world image dataset. We find that 1) image manipulations frequently found in real-world settings like image blurriness and 2) image detectors that are skewed to rely on features like contrast or brightness both have significant negative impacts on fairness for race, gender, and age demographics. These result can help guide future designs of robust models in the object detection field and beyond.", "keywords": [ "object detection", "autonomous vehicles", "algorithmic bias", "algorithmic fairness", "fairness in ML" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 1, 3, 2, 1 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 5, 3, 5 ] }, { "paper_id": "tBom4xOW1H", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Jingfeng Yang", "gender": "Male", "institution": "Singapore Institute of Manufacturing Technology", "country": "SG", "position": "Researcher" }, { "name": "Ni Zhang", "gender": "unknown", "institution": "Singapore Management University", "country": "SG", "position": "Researcher" }, { "name": "Xu Chi", "gender": "Male", "institution": "Singapore Institute of Manufacturing Technology, A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Zhiguang Cao", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 40, "n_ref": 70, "n_ref_all": 87, "n_fig": 3, "n_tab": 6, "L_tab": 2840, "n_element_tab": 480, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2931, "n_element_tab_1": 313, "formula_len_all": 685, "formula_len_all_1": 576, "len_all": 134323, "len_all_1": 69827, "len_abs": 1256, "len_title": 120, "len_sents": 40077, "len_sents_1": 32686, "n_sents": 243, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1260, "title": "Adversarial Generative Flow Network for Solving Vehicle Routing Problems", "abs": "Recent research into solving vehicle routing problems (VRPs) has gained significant traction, particularly through the application of deep (reinforcement) learning for end-to-end solution construction. However, many current construction-based neural solvers predominantly utilize Transformer architectures, which can face scalability challenges and struggle to produce diverse solutions. To address these limitations, we introduce a novel framework beyond Transformer-based approaches, i.e., Adversarial Generative Flow Networks (AGFN). This framework integrates the generative flow network (GFlowNet)\u2014a probabilistic model inherently adept at generating diverse solutions (routes)\u2014with a complementary model for discriminating (or evaluating) the solutions. These models are trained alternately in an adversarial manner to improve the overall solution quality, followed by a proposed hybrid decoding method to construct the solution. We apply the AGFN framework to solve the capacitated vehicle routing problem (CVRP) and travelling salesman problem (TSP), and our experimental results demonstrate that AGFN surpasses the popular construction-based neural solvers, showcasing strong generalization capabilities on synthetic and real-world benchmark instances.", "keywords": [ "Generative Flow Network", "Adversarial Training", "Vehicle Routing Problem" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 2, 3, 2 ] }, { "paper_id": "tBZK9BI2GZ", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Chuwei Luo", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Hangdi Xing", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jiajun Bu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Qi Zheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Alogirithm specialist" }, { "name": "Zhaoqing Zhu", "gender": "unknown", "institution": "Alibaba DAMO Academy", "country": "CN", "position": "Researcher" }, { "name": "Zhi Yu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zirui Shao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 5, "n_ref_uni": 49, "n_ref": 103, "n_ref_all": 120, "n_fig": 4, "n_tab": 6, "L_tab": 2548, "n_element_tab": 210, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2300, "n_element_tab_1": 213, "formula_len_all": 76, "formula_len_all_1": 333, "len_all": 149732, "len_all_1": 61047, "len_abs": 1570, "len_title": 123, "len_sents": 29163, "len_sents_1": 24734, "n_sents": 213, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 122, "L_abs": 1580, "title": "Is Cognition consistent with Perception? Assessing and Mitigating Multimodal Knowledge Conflicts in Document Understanding", "abs": "Multimodal large language models (MLLMs) have shown impressive capabilities in document understanding, a rapidly growing research area with significant industrial demand in recent years. As a multimodal task, document understanding requires models to possess both perceptual and cognitive abilities. However, current MLLMs often face conflicts between perception and cognition. Taking a document VQA task (cognition) as an example, an MLLM might generate answers that do not match the corresponding visual content identified by its OCR (perception). This conflict suggests that the MLLM might struggle to establish an intrinsic connection between the information it \n\"sees\" and what it \"understands.\" Such conflicts challenge the intuitive notion that cognition is consistent with perception, hindering the performance and explainability of MLLMs. In this paper, we define the conflicts between cognition and perception as Cognition and Perception (C&P) knowledge conflicts, a form of multimodal knowledge conflicts, and systematically assess them with a focus on document understanding. Our analysis reveals that even GPT-4o, a leading MLLM, achieves only 68.6% C&P consistency. To mitigate the C&P knowledge conflicts, we propose a novel method called Multimodal Knowledge Consistency Fine-tuning. This method first ensures task-specific consistency and then connects the cognitive and perceptual knowledge. Our method significantly reduces C\\&P knowledge conflicts across all tested MLLMs and enhances their performance in both cognitive and perceptual tasks in most scenarios.", "keywords": [ "knowledge conflicts", "MLLM", "document understanding" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "tBB8hCG5I7", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Imane Hamzaoui", "gender": "Female", "institution": "New York University, Abu Dhabi", "country": "AE", "position": "Intern" }, { "name": "Riyadh Baghdadi", "gender": "Male", "institution": "New York University", "country": "AE", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 10, "n_ref_uni": 15, "n_ref": 22, "n_ref_all": 39, "n_fig": 6, "n_tab": 3, "L_tab": 2564, "n_element_tab": 318, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1723, "n_element_tab_1": 196, "formula_len_all": 764, "formula_len_all_1": 400, "len_all": 115265, "len_all_1": 49771, "len_abs": 3901, "len_title": 127, "len_sents": 30905, "len_sents_1": 21137, "n_sents": 222, "n_sents_1": 152, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1373, "title": "BioNAS: Incorporating Bio-inspired Learning Rules to Neural Architecture Search", "abs": "Bio-inspired neural networks have gained traction due to their adversarial robustness, energy efficiency, and for being biologically plausible. While these bio-inspired networks have shown significant progress, they still fall short in terms of accuracy and are hard to scale to complex tasks. In this paper, we propose to use neural architecture search to further improve state-of-the-art bio-inspired neural networks. We\nachieve this thanks to BioNAS, a framework for neural architecture search that explores different bio-inspired neural network architectures and learning rules. The novelty of BioNAS lies in exploring the use of different bio-inspired learning rules for the different layers of the model. The motivation for this choice comes from recent work in the field suggesting that different learning mechanisms might be used in different\nregions of the human brain. Using BioNAS, we get state-of-the-art bio-inspired neural network performance achieving an accuracy of 94.86 on CIFAR10, 76.48 on CIFAR-100 and 43.42 on ImageNet16-120, surpassing state-of-the-art bio-inspired neural networks. We show that a part of this improvement comes from the use of different learning rules instead of using a single algorithm for all the layers. We release BioNAS to the community and make the code available via this link (https://anonymous.4open.science/r/LR-NAS-DFE1)", "keywords": [ "Feedback Alignment", "Neural Architecture Search", "Learning Rules", "Biologically Plausible" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "tAkdzjHnkP", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Bo Huang", "gender": "Male", "institution": "University of California, San Francisco", "country": "US", "position": "Full Professor" }, { "name": "Dihan Zheng", "gender": "Male", "institution": "University of California, San Francisco", "country": "US", "position": "Postdoc" } ], "n_formula": 22, "n_formula_1": 17, "n_ref_uni": 39, "n_ref": 57, "n_ref_all": 82, "n_fig": 10, "n_tab": 5, "L_tab": 4638, "n_element_tab": 296, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 290, "n_element_tab_1": 53, "formula_len_all": 1553, "formula_len_all_1": 1431, "len_all": 136351, "len_all_1": 60569, "len_abs": 1318, "len_title": 129, "len_sents": 33365, "len_sents_1": 26306, "n_sents": 266, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1040, "title": "CELL-Diff: Unified Diffusion Modeling for Protein Sequences and Microscopy Images", "abs": "Fluorescence microscopy is ubiquitously used in cell biology research to characterize the cellular role of a protein. To help elucidate the relationship between the amino acid sequence of a protein and its cellular function, we introduce CELL-Diff, a unified diffusion model facilitating bidirectional transformations between protein sequences and their corresponding microscopy images. Utilizing reference cell morphology images and a protein sequence, CELL-Diff efficiently generates corresponding protein images. Conversely, given a protein image, the model outputs protein sequences. CELL-Diff integrates continuous and diffusion models within a unified framework and is implemented using a transformer-based network. We train CELL-Diff on the Human Protein Atlas (HPA) dataset and fine-tune it on the OpenCell dataset. Experimental results demonstrate that CELL-Diff outperforms existing methods in generating high-fidelity protein images, making it a practical tool for investigating subcellular protein localization and interactions.", "keywords": [ "AI for biology", "Image generation", "Diffusion model", "MultiModal learning", "Cell biology" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "tAGmxz1TUi", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Chao Ma", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Charles Herrmann", "gender": "Unspecified", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Deqing Sun", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Kelvin C.K. Chan", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Research Scientist" }, { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Xirui Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yinxiao Li", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 110, "n_ref_all": 132, "n_fig": 12, "n_tab": 10, "L_tab": 1545, "n_element_tab": 228, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 806, "n_element_tab_1": 131, "formula_len_all": 965, "formula_len_all_1": 645, "len_all": 166678, "len_all_1": 56190, "len_abs": 5953, "len_title": 0, "len_sents": 49367, "len_sents_1": 25495, "n_sents": 429, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1461, "title": "A Simple Approach to Unifying Diffusion-based Conditional Generation", "abs": "Recent progress in image generation has sparked research into controlling these models through condition signals, with various methods addressing specific challenges in conditional generation. Instead of proposing another specialized technique, we introduce a simple, unified framework to handle diverse conditional generation tasks involving a specific image-condition correlation. By learning a joint distribution over a correlated image pair (e.g. image and depth) with a diffusion model, our approach enables versatile capabilities via different inference-time sampling schemes, including controllable image generation (e.g. depth to image), estimation (e.g. image to depth), signal guidance, joint generation (image \\& depth), and coarse control. Previous attempts at unification often introduce complexity through multi-stage training, architectural modification, or increased parameter counts. In contrast, our simplified formulation requires a single, computationally efficient training stage, maintains the standard model input, and adds minimal learned parameters (15% of the base model). Moreover, our model supports additional capabilities like non-spatially aligned and coarse conditioning. Extensive results show that our single model can produce comparable results with specialized methods and better results than prior unified methods. We also demonstrate that multiple models can be effectively combined for multi-signal conditional generation.", "keywords": [ "image generation", "controllability", "estimation" ], "rating_list": [ 6, 8, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 3, 2, 4, 4, 4 ] }, { "paper_id": "t9lS1lX9FQ", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Hongkang Li", "gender": "unknown", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "Postdoc" }, { "name": "Lei Shi", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Qijiong Liu", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "", "position": "PhD student" }, { "name": "Xiao-Ming Wu", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Associate Professor" }, { "name": "Yuankai Luo", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" } ], "n_formula": 39, "n_formula_1": 23, "n_ref_uni": 79, "n_ref": 165, "n_ref_all": 214, "n_fig": 10, "n_tab": 18, "L_tab": 6271, "n_element_tab": 914, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3583, "n_element_tab_1": 200, "formula_len_all": 2334, "formula_len_all_1": 1053, "len_all": 280454, "len_all_1": 71267, "len_abs": 1021, "len_title": 129, "len_sents": 56353, "len_sents_1": 28985, "n_sents": 476, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 989, "title": "Node Identifiers: Compact, Discrete Representations for Efficient Graph Learning", "abs": "We present a novel end-to-end framework that generates highly compact (typically 6-15 dimensions), discrete (int4 type), and interpretable node representations\u2014termed node identifiers (node IDs)\u2014to tackle inference challenges on large-scale graphs. By employing vector quantization, we compress continuous node embeddings from multiple layers of a Graph Neural Network (GNN) into discrete codes, applicable under both self-supervised and supervised learning paradigms. These node IDs capture high-level abstractions of graph data and offer interpretability that traditional GNN embeddings lack. Extensive experiments on 34 datasets, encompassing node classification, graph classification, link prediction, and attributed graph clustering tasks, demonstrate that the generated node IDs significantly enhance speed and memory efficiency while achieving competitive performance compared to current state-of-the-art methods. Our source code is available at https://github.com/LUOyk1999/NodeID.", "keywords": [ "Graph Neural Networks", "Graph Tokenizers", "Symbolic Compression", "Efficient Graph Learning" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "t9l63huPRt", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Barak Meiri", "gender": "unknown", "institution": "Tel Aviv University", "country": "", "position": "MS student" }, { "name": "Dvir Samuel", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "PhD student" }, { "name": "Gal Chechik", "gender": "unknown", "institution": "Bar Ilan University", "country": "IL", "position": "Full Professor" }, { "name": "Haggai Maron", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Assistant Professor" }, { "name": "Nir Darshan", "gender": "Male", "institution": "OriginAI", "country": "IL", "position": "Principal Researcher" }, { "name": "Rami Ben-Ari", "gender": "Male", "institution": "OriginAI", "country": "IL", "position": "Principal Researcher" }, { "name": "Shai Avidan", "gender": "Male", "institution": "Tel-Aviv University", "country": "IL", "position": "Full Professor" }, { "name": "Yoad Tewel", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 145, "n_ref_all": 185, "n_fig": 18, "n_tab": 6, "L_tab": 1144, "n_element_tab": 89, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 832, "n_element_tab_1": 57, "formula_len_all": 1020, "formula_len_all_1": 464, "len_all": 141448, "len_all_1": 61839, "len_abs": 171, "len_title": 125, "len_sents": 46716, "len_sents_1": 27162, "n_sents": 385, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1347, "title": "Lightning-Fast Image Inversion and Editing for Text-to-Image Diffusion Models", "abs": "Diffusion inversion is the problem of taking an image and a text prompt that describes it and finding a noise latent that would generate the exact same image. Most current deterministic inversion techniques operate by approximately solving an implicit equation and may converge slowly or yield poor reconstructed images. We formulate the problem by finding the roots of an implicit equation and design a method to solve it efficiently. Our solution is based on Newton-Raphson (NR), a well-known technique in numerical analysis. We show that a vanilla application of NR is computationally infeasible while naively transforming it to a computationally tractable alternative tends to converge to out-of-distribution solutions, resulting poor reconstruction and editing. We therefore derive an efficient guided formulation that fastly converges and provides high-quality reconstructions and editing. We showcase our method on real image editing with three popular open-sourced diffusion models: Stable Diffusion, SDXL-Turbo and Flux with different deterministic schedulers. Our solution, Guided Newton-Raphson Inversion, inverts an image within 0.4 sec (on an A100 GPU) for few-step models (SDXL-Turbo and Flux.1), opening the door for interactive image editing. We further show improved results in image interpolation and generation of rare objects.", "keywords": [ "Deterministic Image Inversion", "Image Editing", "Diffusion Models", "Image Generation" ], "rating_list": [ 5, 8, 8, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "t9U3LW7JVX", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Cong Lu", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Postdoc" }, { "name": "Jeff Clune", "gender": "unknown", "institution": "University of Wyoming", "country": "", "position": "" }, { "name": "Shengran Hu", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 71, "n_ref": 219, "n_ref_all": 238, "n_fig": 6, "n_tab": 7, "L_tab": 5714, "n_element_tab": 237, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2965, "n_element_tab_1": 143, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 268624, "len_all_1": 70638, "len_abs": 2107, "len_title": 35, "len_sents": 73214, "len_sents_1": 33268, "n_sents": 537, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 35, "L_abs": 1868, "title": "Automated Design of Agentic Systems", "abs": "Researchers are investing substantial effort in developing powerful general-purpose agents, wherein Foundation Models are used as modules within agentic systems (e.g. Chain-of-Thought, Self-Reflection, Toolformer). However, the history of machine learning teaches us that hand-designed solutions are eventually replaced by learned solutions. We describe a newly forming research area, Automated Design of Agentic Systems (ADAS), which aims to automatically create powerful agentic system designs, including inventing novel building blocks and/or combining them in new ways. We further demonstrate that there is an unexplored yet promising approach within ADAS where agents can be defined in code and new agents can be automatically discovered by a meta agent programming ever better ones in code. Given that programming languages are Turing Complete, this approach theoretically enables the learning of any possible agentic system: including novel prompts, tool use, workflows, and combinations thereof. We present a simple yet effective algorithm named Meta Agent Search to demonstrate this idea, where a meta agent iteratively programs interesting new agents based on an ever-growing archive of previous discoveries. Through extensive experiments across multiple domains including coding, science, and math, we show that our algorithm can progressively invent agents with novel designs that greatly outperform state-of-the-art hand-designed agents. Importantly, we consistently observe the surprising result that agents invented by Meta Agent Search maintain superior performance even when transferred across domains and models, demonstrating their robustness and generality. Provided we develop it safely, our work illustrates the potential of an exciting new research direction toward automatically designing ever-more powerful agentic systems to benefit humanity.", "keywords": [ "LLMs", "Language Model Agents", "Agents", "Agentic Systems", "Reasoning", "Meta Learning", "Open-endedness" ], "rating_list": [ 3, 3, 8, 10 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 4, 2 ], "confidence_list": [ 3, 2, 3, 4 ] }, { "paper_id": "t9NiRq9PGK", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Nicolas Loizou", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" }, { "name": "Sayantan Choudhury", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "PhD student" }, { "name": "TaeHo Yoon", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "PhD student" } ], "n_formula": 116, "n_formula_1": 29, "n_ref_uni": 74, "n_ref": 106, "n_ref_all": 127, "n_fig": 11, "n_tab": 3, "L_tab": 550, "n_element_tab": 24, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 266, "n_element_tab_1": 21, "formula_len_all": 14407, "formula_len_all_1": 2895, "len_all": 226771, "len_all_1": 70763, "len_abs": 2325, "len_title": 124, "len_sents": 53623, "len_sents_1": 28973, "n_sents": 544, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1114, "title": "Multiplayer Federated Learning: Reaching Equilibrium with Less Communications", "abs": "Traditional Federated Learning (FL) approaches assume collaborative clients with aligned objectives working towards a shared global model. However, in many real-world scenarios, clients act as rational players with individual objectives and strategic behaviors, a concept that existing FL frameworks are not equipped to adequately address. To bridge this gap, we introduce *Multiplayer Federated Learning (MpFL)*, a novel framework that models the clients in the FL environment as players in a game-theoretic context, aiming to reach an equilibrium. In this scenario, each player tries to optimize their own utility function, which may not align with the collective goal. Within MpFL, we propose *Per-Player Local Stochastic Gradient Descent (PEARL-SGD)*, an algorithm in which each player/client performs local updates independently and periodically communicates with other players. We theoretically analyze PEARL-SGD and prove that it reaches a neighborhood of equilibrium with less communication in the stochastic setup than its non-local counterpart. Finally, we experimentally verify our theoretical findings.", "keywords": [ "Federated Learning", "Game theory", "Multiplayer games", "Convergence guarantees", "Communication-efficient Algorithms", "Local SGD" ], "rating_list": [ 6, 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3, 4 ] }, { "paper_id": "t9JUTS9ADL", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Dan Zhao", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Lecturer" }, { "name": "Dillon Dupont", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Francesco Bonacci", "gender": "Male", "institution": "Microsoft", "country": "ES", "position": "Researcher" }, { "name": "Justin Wagle", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Principal Researcher" }, { "name": "Kazuhito Koishida", "gender": "unknown", "institution": "Microsoft Corporation", "country": "US", "position": "Researcher" }, { "name": "Rogerio Bonatti", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Sara Abdali", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Yadong Lu", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Yinheng Li", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Zheng Hui", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "A Fender Coelho Bucker", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Lawrence Jang", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 66, "n_ref_all": 107, "n_fig": 12, "n_tab": 15, "L_tab": 4662, "n_element_tab": 558, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 3196, "n_element_tab_1": 243, "formula_len_all": 62, "formula_len_all_1": 0, "len_all": 231016, "len_all_1": 61422, "len_abs": 1705, "len_title": 108, "len_sents": 83402, "len_sents_1": 27064, "n_sents": 708, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1706, "title": "Windows Agent Arena: Evaluating Multi-Modal OS Agents at Scale", "abs": "Large language models (LLMs) show remarkable potential to act as computer agents, enhancing human productivity and software accessibility in multi-modal tasks that require planning and reasoning. However, measuring agent performance in realistic environments remains a challenge since: (i) most benchmarks are limited to specific modalities or domains (e.g. text-only, web navigation, Q&A, coding) and (ii) full benchmark evaluations are slow (on order of magnitude of days) given the multi-step sequential nature of tasks. To address these challenges, we introduce the Windows Agent Arena: a reproducible, general environment focusing exclusively on the Windows operating system (OS) where agents can operate freely within a real Windows OS and use the same wide range of applications, tools, and web browsers available to human users when solving tasks. We adapt the OSWorld framework (Xie et al., 2024) to create 150+ diverse Windows tasks across representative domains that require agent abilities in planning, screen understanding, and tool usage. Our benchmark is scalable and can be seamlessly parallelized in Azure for a full benchmark evaluation in as little as 20 minutes. To demonstrate Windows Agent Arena's capabilities, we also introduce a new multi-modal agent, Navi. Our agent achieves a success rate of 19.5% in the Windows domain, compared to 74.5% performance of an unassisted human. Navi also demonstrates strong performance on another popular web-based benchmark, Mind2Web. We offer extensive quantitative and qualitative analysis of Navi's performance, and provide insights into the opportunities for future research in agent development and data generation using Windows Agent Arena.", "keywords": [ "multi-modal agent" ], "rating_list": [ 5, 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 2, 4, 4 ] }, { "paper_id": "t8qcGXaepr", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Mengqi Zhang", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Pengjie Ren", "gender": "Male", "institution": "Shandong University", "country": "", "position": "Full Professor" }, { "name": "Qiang Liu", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Shu Wu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xiaotian Ye", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Zhumin Chen", "gender": "unknown", "institution": "Shandong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 3, "n_ref_uni": 23, "n_ref": 72, "n_ref_all": 99, "n_fig": 9, "n_tab": 7, "L_tab": 3191, "n_element_tab": 462, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2025, "n_element_tab_1": 390, "formula_len_all": 528, "formula_len_all_1": 147, "len_all": 155797, "len_all_1": 72289, "len_abs": 1482, "len_title": 102, "len_sents": 53715, "len_sents_1": 31601, "n_sents": 356, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 54, "L_abs": 1491, "title": "Uncovering Overfitting in Large Language Model Editing", "abs": "Knowledge editing has been proposed as an effective method for updating and correcting the internal knowledge of Large Language Models (LLMs). However, existing editing methods often struggle with complex tasks, such as multi-hop reasoning. In this paper, we identify and investigate the phenomenon of **Editing Overfit**, where edited models assign disproportionately high probabilities to the edit target, hindering the generalization of new knowledge in complex scenarios. We attribute this issue to the current editing paradigm, which places excessive emphasis on the direct correspondence between the input prompt and the edit target for each edit sample. To further explore this issue, we introduce a new benchmark, EVOKE (EValuation of Editing Overfit in Knowledge Editing), along with fine-grained evaluation metrics. Through comprehensive experiments and analysis, we demonstrate that Editing Overfit is prevalent in current editing methods and that common overfitting mitigation strategies are ineffective in knowledge editing. To overcome this, inspired by LLMs\u2019 knowledge recall mechanisms, we propose a new plug-and-play strategy called Learn to Inference (LTI), which introduce a Multi-stage Inference Constraint module to guide the edited models in recalling new knowledge similarly to how unedited LLMs leverage knowledge through in-context learning. Extensive experimental results across a wide range of tasks validate the effectiveness of LTI in mitigating Editing Overfit.", "keywords": [ "Large language model", "Knowledge editing", "Editing overfit" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 4, 3 ], "contribution_list": [ 4, 3, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "t8hMqAn8ZG", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Guan Huang", "gender": "Male", "institution": "Auburn University", "country": "US", "position": "PhD student" }, { "name": "Tao Shu", "gender": "Male", "institution": "Auburn University", "country": "US", "position": "Associate Professor" } ], "n_formula": 68, "n_formula_1": 32, "n_ref_uni": 39, "n_ref": 75, "n_ref_all": 104, "n_fig": 1, "n_tab": 10, "L_tab": 2997, "n_element_tab": 271, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1540, "n_element_tab_1": 204, "formula_len_all": 4177, "formula_len_all_1": 1363, "len_all": 161722, "len_all_1": 68015, "len_abs": 1484, "len_title": 124, "len_sents": 47711, "len_sents_1": 31075, "n_sents": 360, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1499, "title": "Decentralized Federated Learning Over Noisy Labels: A Majority Voting Method", "abs": "Contrary to centralized federated learning (CFL), decentralized federated learning (DFL) allows clients to cooperate in training their local models without relying on a central parameter server. As different clients have varying annotation skills and preferences, noisy labels are inevitable in decentralized data ownership. In centralized learning (CL) and CFL settings, learning from noisy labels has been extensively explored; however, such methods cannot be directly applied in DFL settings due to limited computational resources or privacy requirements. This paper introduces DFLMV \\textit{(majority voting based decentralized federated learning)}, a general DFL framework for learning from noisy data without relying on any assumptions about local client noise models while maintaining data privacy for all clients. Specifically, (1) Clients first use traditional DFL to train their local models until they become stable. (2) Clients use each of their neighbors' models to make a prediction of every data point in their training datasets, then correct the labels based on majority voting. (3) Clients further fine-tune their models based on their updated training dataset. A theoretical analysis of DFLMV is also provided. Extensive experiments conducted on MNIST, Fashion-MNIST, CIFA-10, CIFAR-10N, CIFAR-100N, Clothing1M, and ANIMAL-10N validate the effectiveness of our proposed approach at various noise levels and different data settings in mitigating the adverse effects of noisy labels.", "keywords": [ "decentralized federated learning", "distributed learning", "federated learning", "majority voting", "label-noise learning" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 3, 2, 1 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "t8fu5m8R5m", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Hossein Mirzaei", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Jafar Habibi", "gender": "Male", "institution": "Sharif University of Technology", "country": "IR", "position": "Full Professor" }, { "name": "Mohammad Hossein Rohban", "gender": "Male", "institution": "Sharif University of Technology", "country": "IR", "position": "Assistant Professor" }, { "name": "Mohammad Sabokrou", "gender": "Male", "institution": "Okinawa Institute of Science and Technology (OIST)", "country": "JP", "position": "Researcher" }, { "name": "Mojtaba Nafez", "gender": "Male", "institution": "Sharif University of Technology", "country": "IR", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 9, "n_ref_uni": 47, "n_ref": 103, "n_ref_all": 128, "n_fig": 3, "n_tab": 20, "L_tab": 11792, "n_element_tab": 1688, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 4080, "n_element_tab_1": 465, "formula_len_all": 1544, "formula_len_all_1": 819, "len_all": 232151, "len_all_1": 73103, "len_abs": 1751, "len_title": 128, "len_sents": 56740, "len_sents_1": 30876, "n_sents": 404, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1760, "title": "Adversarially Robust Anomaly Detection through Spurious Negative Pair Mitigation", "abs": "Despite significant progress in Anomaly Detection (AD), the robustness of existing detection methods against adversarial attacks remains a challenge, compromising their reliability in critical real-world applications such as autonomous driving. This issue primarily arises from the AD setup, which assumes that training data is limited to a group of unlabeled normal samples, making the detectors vulnerable to adversarial anomaly samples during testing. Additionally, implementing adversarial training as a safeguard encounters difficulties, such as formulating an effective objective function without access to labels. An ideal objective function for adversarial training in AD should promote strong perturbations both within and between the normal and anomaly groups to maximize margin between normal and anomaly distribution. To address these issues, we first propose crafting a pseudo-anomaly group derived from normal group samples. Then, we demonstrate that adversarial training with contrastive loss could serve as an ideal objective function, as it creates both inter- and intra-group perturbations. However, we notice that spurious negative pairs compromise the conventional contrastive loss for achieving robust AD. Spurious negative pairs are those that should be mapped closely but are erroneously separated. These pairs introduce noise and misguide the direction of inter-group adversarial perturbations. To overcome the effect of spurious negative pairs, we define opposite pairs and adversarially pull them apart to strengthen inter-group perturbations. Experimental results demonstrate our superior performance in both clean and adversarial scenarios, with a 26.1% improvement in robust detection across various challenging benchmark datasets.", "keywords": [ "Anomaly Detection", "Adversarially Robust Anomaly Detection", "Mitigating Spurious Negative Pairs", "Anomaly Aware Contrastive Learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "t8ctvylFn7", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Carmen Amo Alonso", "gender": "unknown", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Emily Cheng", "gender": "Female", "institution": "Universitat Pompeu Fabra", "country": "ES", "position": "PhD student" }, { "name": "Marco Baroni", "gender": "Male", "institution": "Universitat Pompeu Fabra", "country": "", "position": "Full Professor" } ], "n_formula": 46, "n_formula_1": 23, "n_ref_uni": 46, "n_ref": 96, "n_ref_all": 112, "n_fig": 5, "n_tab": 5, "L_tab": 1922, "n_element_tab": 120, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 250, "n_element_tab_1": 15, "formula_len_all": 20620, "formula_len_all_1": 16117, "len_all": 247878, "len_all_1": 100001, "len_abs": 1133, "len_title": 116, "len_sents": 71348, "len_sents_1": 42869, "n_sents": 563, "n_sents_1": 272, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1063, "title": "Linearly Controlled Language Generation with Performative Guarantees", "abs": "The increasing prevalence of Large Language Models (LMs) in critical applications highlights the need for controlled language generation strategies that are not only computationally efficient but that also enjoy performance guarantees. To achieve this, we use a common model of concept semantics as linearly represented in an LM\u2019s latent space. In particular, we take the view that natural language generation traces a trajectory in this continuous semantic space, realized by the language model\u2019s hidden activations. This view permits a control-theoretic treatment of text generation in latent space, in which we propose a lightweight, gradient-free intervention that dynamically steers trajectories away from regions corresponding to undesired meanings. Crucially, we show that this intervention, which we compute in closed form, is guaranteed (in probability) to steer the output into the allowed region. Finally, we demonstrate on a toxicity avoidance objective that the intervention steers language away from undesired content while maintaining text quality.", "keywords": [ "control theory", "representation engineering", "large language models" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "t8LzkidEuc", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Emanuele Sansone", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "Postdoc" }, { "name": "Tim Lebailly", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "PhD student" }, { "name": "Tinne Tuytelaars", "gender": "Female", "institution": "KU Leuven", "country": "BE", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 14, "n_ref_uni": 79, "n_ref": 137, "n_ref_all": 151, "n_fig": 8, "n_tab": 1, "L_tab": 15, "n_element_tab": 1, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 15, "n_element_tab_1": 1, "formula_len_all": 1152, "formula_len_all_1": 1259, "len_all": 144299, "len_all_1": 67865, "len_abs": 1040, "len_title": 54, "len_sents": 35385, "len_sents_1": 31451, "n_sents": 249, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1042, "title": "Failure-Proof Non-Contrastive Self-Supervised Learning", "abs": "We identify sufficient conditions to avoid known failure modes, including representation, dimensional, cluster and intracluster collapses, occurring in non-contrastive self-supervised learning. Based on these findings, we propose a principled design for the projector and loss function. We theoretically demonstrate that this design introduces an inductive bias that promotes learning representations that are both decorrelated and clustered without explicit enforcing these properties and leading to improved generalization. To the best of our knowledge, this is the first solution that achieves robust training with respect to these failure modes while guaranteeing enhanced generalization performance in downstream tasks. We validate our theoretical findings on image datasets including SVHN, CIFAR10, CIFAR100 and ImageNet-100, and show that our solution, dubbed FALCON, outperforms existing feature decorrelation and cluster-based self-supervised learning methods in terms of generalization to clustering and linear classification tasks.", "keywords": [ "Self-Supervised Learning", "Representation Learning", "Deep Learning" ], "rating_list": [ 5, 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 2, 4, 3 ], "presentation_list": [ 2, 4, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 3, 2 ] }, { "paper_id": "t8KLjiFNwn", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Changdi Yang", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Hangyu Zheng", "gender": "Male", "institution": "University of Georgia", "country": "", "position": "PhD student" }, { "name": "Pu Zhao", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Researcher" }, { "name": "Wei Niu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Assistant Professor" }, { "name": "Xuan Shen", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Xue Lin", "gender": "Female", "institution": "Northeastern University", "country": "", "position": "Assistant Professor" }, { "name": "Yanzhi Wang", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Associate Professor" }, { "name": "Yifan Gong", "gender": "Female", "institution": "Northeastern University", "country": "", "position": "PhD student" }, { "name": "Yushu Wu", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Zheng Zhan", "gender": "unknown", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Zhenglun Kong", "gender": "Male", "institution": "Harvard Medical School, Harvard University", "country": "US", "position": "Postdoc" } ], "n_formula": 21, "n_formula_1": 19, "n_ref_uni": 36, "n_ref": 67, "n_ref_all": 95, "n_fig": 4, "n_tab": 4, "L_tab": 2549, "n_element_tab": 543, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1708, "n_element_tab_1": 294, "formula_len_all": 1826, "formula_len_all_1": 1555, "len_all": 139034, "len_all_1": 77286, "len_abs": 1857, "len_title": 96, "len_sents": 36188, "len_sents_1": 32503, "n_sents": 258, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 1901, "title": "Sparse Learning for State Space Models on Mobile", "abs": "Transformer models have been widely investigated in different domains by providing long-range dependency handling and global contextual awareness, driving the development of popular AI applications such as ChatGPT, Gemini, and Alexa.\nState Space Models (SSMs) have emerged as strong contenders in the field of sequential modeling, challenging the dominance of Transformers. SSMs incorporate a selective mechanism that allows for dynamic parameter adjustment based on input data, enhancing their performance.\nHowever, this mechanism also comes with increasing computational complexity and bandwidth demands, posing challenges for deployment on resource-constraint mobile devices.\nTo address these challenges without sacrificing the accuracy of the selective mechanism, we propose a sparse learning framework that integrates architecture-aware compiler optimizations. We introduce an end-to-end solution--$\\mathbf{C}_4^n$ kernel sparsity, which prunes $n$ elements from every four contiguous weights, and develop a compiler-based acceleration solution to ensure execution efficiency for this sparsity on mobile devices.\nBased on the kernel sparsity, our framework generates optimized sparse models targeting specific sparsity or latency requirements for various model sizes. We further leverage pruned weights to compensate for the remaining weights, enhancing downstream task performance.\nFor practical hardware acceleration, we propose $\\mathbf{C}_4^n$-specific optimizations combined with a layout transformation elimination strategy. \nThis approach mitigates inefficiencies arising from fine-grained pruning in linear layers and improves performance across other operations. \nExperimental results demonstrate that our method achieves superior task performance compared to other semi-structured pruning methods and achieves up-to 7$\\times$ speedup compared to llama.cpp framework on mobile devices.", "keywords": [ "Mamba", "Pruning", "Mobile" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 1, 3, 1, 3 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "t8FG4cJuL3", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Jun Yu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yanzheng Chen", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 123, "n_formula_1": 35, "n_ref_uni": 26, "n_ref": 78, "n_ref_all": 87, "n_fig": 2, "n_tab": 2, "L_tab": 419, "n_element_tab": 6, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 555, "n_element_tab_1": 5, "formula_len_all": 39486, "formula_len_all_1": 5902, "len_all": 166219, "len_all_1": 55614, "len_abs": 1662, "len_title": 165, "len_sents": 35893, "len_sents_1": 21302, "n_sents": 313, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 117, "L_abs": 1678, "title": "Classic but Everlasting: Traditional Gradient-Based Algorithms Converges Fast Even in Time-Varying Multi-Player Games", "abs": "Last-iterate convergence behaviours of well-known algorithms are intensively investigated in various games, such as two-player bilinear zero-sum games.\nHowever, most known last-iterate convergence properties rely on strict settings where the underlying games must have time-invariant payoffs.\nBesides, the limited known attempts on the games with time-varying payoffs are in two-player bilinear time-varying zero-sum games and strictly monotone games. By contrast, in other time-varying games, the last-iterate behaviours of two classic algorithms, i.e., optimistic gradient (OG) and extra gradient (EG) algorithms, still lack research, especially the convergence rates in multi-player games.\nIn this paper, we investigate the last-iterate behaviours of OG and EG algorithms for convergent perturbed games, which extend upon the usual model of time-invariant games and incorporate external factors, such as vanishing noises.\nUsing the recently proposed notion of the tangent residual (or its modifications) as the potential function of games and the measure of proximity to the Nash equilibrium, we prove that the last-iterate convergence rates of EG and OG algorithms for perturbed games on bounded convex closed sets are $O({1}/{\\sqrt{T}})$ if such games converge to monotone games at rates fast enough and that such a result holds true for certain unconstrained perturbed games. With this result, we address an open question\nasking for the last-iterate convergence rate of the extra gradient and the optimistic gradient algorithms in constrained and time-varying settings. The above convergence rates are similar to known tight results on corresponding time-invariant games.", "keywords": [ "time-varying games", "Nash equilibrium", "extra gradient algorithm", "optimistic gradient algorithm" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 4, 4 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "t7vXubuady", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "James Bailey", "gender": "unknown", "institution": "The University of Melbourne", "country": "AU", "position": "Full Professor" }, { "name": "Sarah Monazam Erfani", "gender": "unknown", "institution": "The University of Melbourne", "country": "AU", "position": "Associate Professor" }, { "name": "Xingjun Ma", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Xueqi Ma", "gender": "Female", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 16, "n_ref_uni": 36, "n_ref": 68, "n_ref_all": 82, "n_fig": 1, "n_tab": 8, "L_tab": 3735, "n_element_tab": 283, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 3311, "n_element_tab_1": 225, "formula_len_all": 2512, "formula_len_all_1": 1015, "len_all": 154637, "len_all_1": 65904, "len_abs": 1875, "len_title": 106, "len_sents": 38332, "len_sents_1": 28269, "n_sents": 280, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1548, "title": "Attention-based Graph Coreset Labeling for Active Learning", "abs": "GNN-based Active Learning (AL) methods have been proposed to improve labeling efficiency by selecting the most informative nodes in a graph for labeling. The existing graph active learning methods employ different heuristic approaches, while efficiency sometimes, they fail to explicitly explore the influence of labeled data on unlabeled data, thus limiting the generalizability of graph models to various types of graph data. In this paper, we propose an Attention-based Graph Coreset Labeling framework (AGCL). AGCL can, with limited budgets, gradually discover core data to be labeled from a global view so as to obtain a training dataset that can efficiently depict the whole graph space and maximize the performance of GNNs. Specifically, we explicitly explore and exploit the correlations between nodes in the unlabeled pool and those in the labeled pool using an attention architecture and directly connect the correlations with the prediction performance on unlabeled set. Using influence scores, AGCL can identify data for labeling having maximum representation difference from the existing labeled pool. This enhances sample complexity.We theoretically demonstrate the superiority of the attention-based data selection strategy in reducing the covering radius bound, thereby improving the expected prediction performance on unlabeled data.\nOur experimental results show that the labeled coreset can improve the generalizability of various graph models across different graph datasets, as well as CNN models on image classification tasks.", "keywords": [ "active learning; graph learning" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "t7P5BUKcYv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Peng Jin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Bo.zhu", "gender": "Male", "institution": "Kunlun Tiangong Singularity AI", "country": "CN", "position": "Researcher" }, { "name": "Yuan LI", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 44, "n_ref": 90, "n_ref_all": 101, "n_fig": 7, "n_tab": 13, "L_tab": 7734, "n_element_tab": 920, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 4436, "n_element_tab_1": 601, "formula_len_all": 769, "formula_len_all_1": 763, "len_all": 195862, "len_all_1": 70189, "len_abs": 1824, "len_title": 124, "len_sents": 47976, "len_sents_1": 24466, "n_sents": 409, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 76, "L_abs": 1676, "title": "MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts", "abs": "In this work, we aim to simultaneously enhance the effectiveness and efficiency of Mixture-of-Experts (MoE) methods. To achieve this, we propose MoE++, a general and heterogeneous MoE framework that integrates both Feed-Forward Network (FFN) and zero-computation experts. Specifically, we introduce three types of zero-computation experts: the zero expert, copy expert, and constant expert, which correspond to discard, skip, and replace operations, respectively. This design offers three key advantages: (i) **Low Computing Overhead**: Unlike the uniform mixing mechanism for all tokens within vanilla MoE, MoE++ allows each token to engage with a dynamic number of FFNs, be adjusted by constant vectors, or even skip the MoE layer entirely. (ii) **High Performance**: By enabling simple tokens to utilize fewer FFN experts, MoE++ allows more experts to focus on challenging tokens, thereby unlocking greater performance potential than vanilla MoE. (iii) **Deployment Friendly**: Given that zero-computation experts have negligible parameters, we can deploy all zero-computation experts on each GPU, eliminating the significant communication overhead and expert load imbalance associated with FFN experts distributed across different GPUs. Moreover, we leverage gating residuals, enabling each token to consider the pathway taken in the previous layer when selecting the appropriate experts. Extensive experimental results demonstrate that MoE++ achieves better performance while delivering 1.1$\\sim$2.1$\\times$ expert forward throughput compared to a vanilla MoE model of the same size, which lays a solid foundation for developing advanced and efficient MoE-related models.", "keywords": [ "Mixture of Experts", "Large Language Models", "Efficient Foundation Models" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "t73rC2GJQJ", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Bo Zheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Shuai Wang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Tiezheng Ge", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Weixin Feng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xubin Li", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Limin Wang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "tianhui Song", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" } ], "n_formula": 21, "n_formula_1": 19, "n_ref_uni": 40, "n_ref": 75, "n_ref_all": 102, "n_fig": 19, "n_tab": 2, "L_tab": 297, "n_element_tab": 12, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1121, "n_element_tab_1": 70, "formula_len_all": 1105, "formula_len_all_1": 970, "len_all": 143135, "len_all_1": 62600, "len_abs": 232, "len_title": 133, "len_sents": 35938, "len_sents_1": 28832, "n_sents": 272, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1389, "title": "DMM: Building a Versatile Image Generation Model via Distillation-Based Model Merging", "abs": "The success of text-to-image (T2I) generation models has spurred a proliferation of numerous model checkpoints fine-tuned from the same base model on various specialized datasets.This overwhelming specialized model production introduces new challenges for high parameter redundancy and huge storage cost, thereby necessitating the development of effective methods to consolidate and unify the capabilities of diverse powerful models into a single one.A common practice in model merging adopts static linear interpolation in the parameter space to achieve the goal of style mixing.However, it neglects the features of T2I generation task that numerous distinct models cover sundry styles which may lead to incompatibility and confusion in the merged model.To address this issue, we introduce a style-promptable image generation pipeline which can accurately generate arbitrary-style images under the control of style vectors. Based on this design, we propose the score distillation based model merging paradigm (DMM), compressing multiple models into a single versatile T2I model. Moreover, we rethink and reformulate the model merging task in the context of T2I generation, by presenting new merging goals and evaluation protocols. Our experiments demonstrate that DMM can compactly reorganize the knowledge from multiple teacher models and achieve controllable arbitrary-style generation.", "keywords": [ "Diffusion Models", "Generative Models", "Model Merging" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "t717joHHSc", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Chin-Yew Lin", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Senior Principal Research Manager" }, { "name": "Dongsheng Li", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Huiqiang Jiang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "RSDE" }, { "name": "Lili Qiu", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Principal Researcher" }, { "name": "Qianhui Wu", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Xufang Luo", "gender": "Female", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" }, { "name": "Yijiong Yu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Yongfeng Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Yuqing Yang", "gender": "unknown", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 100, "n_ref_all": 128, "n_fig": 17, "n_tab": 4, "L_tab": 860, "n_element_tab": 101, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 350, "n_element_tab_1": 56, "formula_len_all": 1113, "formula_len_all_1": 575, "len_all": 141837, "len_all_1": 50145, "len_abs": 1234, "len_title": 127, "len_sents": 37195, "len_sents_1": 24314, "n_sents": 221, "n_sents_1": 143, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1238, "title": "Mitigate Position Bias in Large Language Models via Scaling a Single Dimension", "abs": "Large Language Models (LLMs) are increasingly applied in various real-world scenarios due to their excellent generalization capabilities and robust generative abilities. However, they exhibit position bias, also known as \"lost in the middle\", a phenomenon that is especially pronounced in long-context scenarios, which indicates the placement of the key information in different positions of a prompt can significantly affect accuracy. This paper first explores the micro-level manifestations of position bias, concluding that attention weights are a micro-level expression of position bias. It further identifies that, in addition to position embeddings, causal attention mask also contributes to position bias by creating position-specific hidden states. Based on these insights, we propose a method to mitigate position bias by scaling this positional hidden states. Experiments on the NaturalQuestions Multi-document QA, KV retrieval, LongBench and timeline reorder tasks, using various models including RoPE models, context window-extended models, and Alibi models, demonstrate the effectiveness and generalizability of our approach. Our method can improve performance by up to 15.2% by modifying just one dimension of hidden states.", "keywords": [ "Large Language Model", "Position Bias", "Long-Context" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "t6QHYUOQL7", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Jie Cao", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Ran He", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Tieniu Tan", "gender": "Male", "institution": " Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Professor" }, { "name": "Yexiang Liu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zekun Li", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 44, "n_ref": 132, "n_ref_all": 186, "n_fig": 27, "n_tab": 25, "L_tab": 11297, "n_element_tab": 1447, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2003, "n_element_tab_1": 374, "formula_len_all": 623, "formula_len_all_1": 430, "len_all": 290416, "len_all_1": 63261, "len_abs": 1239, "len_title": 123, "len_sents": 77599, "len_sents_1": 25063, "n_sents": 752, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1245, "title": "Breaking Mental Set to Improve Reasoning through Diverse Multi-Agent Debate", "abs": "Large Language Models (LLMs) have seen significant progress but continue to struggle with persistent reasoning mistakes. Previous methods of self-reflection have been proven limited due to the models\u2019 inherent fixed thinking patterns. While Multi-Agent Debate (MAD) attempts to mitigate this by incorporating multiple agents, it often employs the same reasoning methods, even though assigning different personas to models. This leads to a \u201cfixed mental set,\u201d where models rely on homogeneous thought processes without exploring alternative perspectives. In this paper, we introduce Diverse Multi-Agent Debate (DMAD), a method that encourages agents to think with distinct reasoning approaches. By leveraging diverse problem-solving strategies, each agent can gain insights from different perspectives, refining its responses through discussion and collectively arriving at the optimal solution. DMAD effectively breaks the limitations of fixed mental sets. We evaluate DMAD against various prompting techniques, including self-reflection and traditional MAD, across multiple benchmarks using both LLMs and Multimodal LLMs. Our experiments show that DMAD consistently outperforms other methods, delivering better results than MAD in fewer rounds.", "keywords": [ "Multi-Agent Debate", "Large Language Models", "Multimodal Large Language Models", "Prompting", "Self-Correction", "Reasoning" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "t5mpbfpZuF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "David Wu", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Intern" }, { "name": "Sanjiban Choudhury", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 23, "n_ref": 33, "n_ref_all": 43, "n_fig": 2, "n_tab": 5, "L_tab": 1855, "n_element_tab": 116, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 722, "n_element_tab_1": 41, "formula_len_all": 168, "formula_len_all_1": 135, "len_all": 137153, "len_all_1": 52694, "len_abs": 1601, "len_title": 101, "len_sents": 36261, "len_sents_1": 26785, "n_sents": 247, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1607, "title": "Aligning Large Language Models with Domain Adaptation", "abs": "Aligning large language models (LLMs) has emerged as a critical challenge in the age of generative AI: LLMs must be appropriately aligned with human values and preferences in order to be helpful and harmless. In many real world cases, however, large amounts of preference data are not available on important tasks, limiting the effectiveness of resulting reward models. In some cases, data from a similar task is available, and unlabeled data on the target task is available or can be generated by an LLM. In other cases, clean data may be available to train an LLM for real-world use on noisy data, small amounts of labeled data on the target task may be available, or data may be available on an easier task. In this work, we demonstrate that domain adaptation can effectively use different types of data, by transferring supervision and human values across tasks with similar data distributions, strengthening resistance to noisy data, improving few-shot generalization ability, and even transfer from easy to hard tasks, in the form of short to long generalization. Specifically, we propose Data Efficient Alignment for Language (DEAL), using domain adaptation to effectively perform cross-task alignment in scenarios where labeled target data is not available. We evaluate our method for reward model training on a variety of benchmarks and demonstrate that our method can meaningfully improve performance on target tasks by utilizing data on related tasks or low amounts of data. Furthermore, we offer analysis on the inner mechanism of domain adaptation and the alignment of embedding distributions.", "keywords": [ "alignment", "domain adaptation", "large language models", "generalization" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "t5kThOYtxn", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Ishan Sengupta", "gender": "unknown", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Koulik Khamaru", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" } ], "n_formula": 38, "n_formula_1": 34, "n_ref_uni": 22, "n_ref": 33, "n_ref_all": 34, "n_fig": 0, "n_tab": 1, "L_tab": 427, "n_element_tab": 57, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 728, "n_element_tab_1": 63, "formula_len_all": 2650, "formula_len_all_1": 2155, "len_all": 92102, "len_all_1": 59223, "len_abs": 1410, "len_title": 105, "len_sents": 24451, "len_sents_1": 22190, "n_sents": 223, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1430, "title": "Stable batched bandit: Optimal regret with free inference", "abs": "In this paper, we discuss statistical inference when using a sequential strategy to collect data. While inferential tasks become challenging with sequentially collected data, we argue that this problem can be alleviated when the sequential algorithm satisfies certain stability properties; we call such algorithms stable bandit algorithms. Focusing on batched bandit problems, we first demonstrate that popular algorithms including the greedy-UCB algorithm and $\\epsilon$-greedy ETC algorithms are not stable, complicating downstream inferential tasks. Our main result shows that a form of elimination algorithm is stable in the batched bandit setup, and we characterize the asymptotic distribution of the sample means. This result allows us to construct asymptotically exact confidence intervals for arm-means which are sharper than existing concentration-based bounds. As a byproduct of our main results, we propose an Explore and Commit (ETC) strategy, which is stable --- thus allowing easy statistical inference--- and also attains optimal regret up to a factor of 4.\n\nOur work connects two historically conflicting paradigms in sequential learning environments: regret minimization and statistical inference. Ultimately, we demonstrate that it is possible to minimize regret without sacrificing the ease of performing statistical inference, bridging the gap between these two important aspects of sequential decision-making.", "keywords": [ "Batched Bandit", "Inference in Bandits" ], "rating_list": [ 5, 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 3, 2, 4 ], "presentation_list": [ 2, 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "t5FD4QTDTu", "primary_area": "generative models", "n_author": 11, "author_info_list": [ { "name": "Ali Moin", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Behrooz Mahasseni", "gender": "unknown", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Christopher Michael Sandino", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Ellen L. Zippi", "gender": "Female", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Erdrin Azemi", "gender": "Female", "institution": "Apple", "country": "", "position": "Principal Researcher" }, { "name": "Hadi Pouransari", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Jingyun Xiao", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Juri Minxha", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Ran Liu", "gender": "Female", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Wenrui Ma", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "MS student" }, { "name": "Eva Dyer", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 51, "n_ref": 88, "n_ref_all": 106, "n_fig": 2, "n_tab": 4, "L_tab": 1821, "n_element_tab": 100, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1813, "n_element_tab_1": 98, "formula_len_all": 784, "formula_len_all_1": 753, "len_all": 150098, "len_all_1": 73950, "len_abs": 486, "len_title": 130, "len_sents": 38735, "len_sents_1": 33759, "n_sents": 241, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1853, "title": "Generalizable autoregressive modeling of time series through functional narratives", "abs": "Time series data are inherently functions of time, yet current transformers often learn time series by modeling them as mere concatenations of time periods, overlooking their functional properties. In this work, we propose a novel objective for transformers that learn time series by re-interpreting them as temporal functions. We build an alternative sequence of time series by constructing degradation operators of different intensity in the functional space, creating augmented variants of the original sample that are abstracted or simplified to different degrees. Based on the new set of generated sequence, we train an autoregressive transformer that progressively recovers the original sample from the most simplified variant. Analogous to the next word prediction task in languages that learns narratives by connecting different words, our autoregressive transformer aims to learn the Narratives of Time Series (NoTS) by connecting different functions in time. Theoretically, we justify the construction of the alternative sequence through its advantages in approximating functions. When learning time series data with transformers, constructing sequences of temporal functions allows for a broader class of approximable functions (e.g., differentiation) compared to sequences of time periods, leading to a 26$\\%$ performance improvement in synthetic feature regression experiments. Experimentally, we validate NoTS in 3 different tasks across 22 real-world datasets, where we show that NoTS significantly outperforms other pre-training methods by up to 6\\%. Additionally, combining NoTS on top of existing transformer architectures can consistently boost the performance. Our results demonstrate the potential of NoTS as a general-purpose dynamic learner, offering a viable alternative for developing foundation models for time series analysis.", "keywords": [ "autoregressive modeling", "transformers", "time series" ], "rating_list": [ 5, 3, 5, 3, 8 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 2, 4, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3, 3 ] }, { "paper_id": "t3rdi80xCz", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Claire Vernade", "gender": "Female", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "", "position": "Assistant Professor" }, { "name": "Michael Muehlebach", "gender": "unknown", "institution": "Max-Planck Institute", "country": "DE", "position": "Principal Researcher" }, { "name": "Onno Eberhard", "gender": "Male", "institution": "Max Planck Institute for Intelligent Systems", "country": "", "position": "PhD student" } ], "n_formula": 62, "n_formula_1": 25, "n_ref_uni": 39, "n_ref": 54, "n_ref_all": 100, "n_fig": 9, "n_tab": 4, "L_tab": 876, "n_element_tab": 26, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9048, "formula_len_all_1": 2550, "len_all": 165599, "len_all_1": 70709, "len_abs": 765, "len_title": 98, "len_sents": 46911, "len_sents_1": 31108, "n_sents": 410, "n_sents_1": 250, "venue": "ICLR2025", "decision": "withdraw", "L_title": 50, "L_abs": 773, "title": "A Pontryagin Perspective on Reinforcement Learning", "abs": "Reinforcement learning has traditionally focused on learning state-dependent policies to solve optimal control problems in a _closed-loop_ fashion. In this work, we introduce the paradigm of _open-loop reinforcement learning_ where a fixed action sequence is learned instead. We present three new algorithms: one robust model-based method and two sample-efficient model-free methods. Rather than basing our algorithms on Bellman's equation from dynamic programming, our work builds on _Pontryagin's principle_ from the theory of open-loop optimal control. We provide convergence guarantees and evaluate all methods empirically on a pendulum swing-up task, as well as on two high-dimensional MuJoCo tasks, demonstrating remarkable performance compared to existing baselines.", "keywords": [ "reinforcement learning", "control theory" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "t2yD3IaIMc", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Chengyu Jiao", "gender": "Male", "institution": "Southern University for Science and Technology", "country": "CN", "position": "Postdoc" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 20, "n_ref_uni": 23, "n_ref": 52, "n_ref_all": 70, "n_fig": 5, "n_tab": 6, "L_tab": 1346, "n_element_tab": 152, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1196, "n_element_tab_1": 122, "formula_len_all": 758, "formula_len_all_1": 862, "len_all": 143364, "len_all_1": 71469, "len_abs": 1178, "len_title": 83, "len_sents": 43856, "len_sents_1": 29234, "n_sents": 426, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1194, "title": "Hypernetwork-Based Equivariant CNNs", "abs": "In geometric deep learning, numerous works have been dedicated to enhancing neural networks with ability to preserve symmetries, a concept known as equivariance. Convolutional Neural Networks (CNNs) are already equivariant to translations. To further achieve rotation and reflection equivariance, previous methods are primarily based on Group Equivariant Convolutional Neural Networks ($G$-CNN). While showing a significant improvement when processing rotation-augmented datasets, training $G$-CNN on a dataset with little rotational variation typically leads to a performance drop comparing to a regular CNN. In this study, we discuss the reason of $G$-CNN not performing on datasets with little rotational variation. We propose an alternative approach: generating CNN filters that inherently exhibit rotational equivariance without altering the main network's CNN structure. This is achieved through our novel application of a dynamic hypernetwork. We prove these generated filters grant equivariance property to a regular CNN main network. Our experiments demonstrate that our method outperforms $G$-CNN and achieves performance comparable to advanced state-of-the-art $G$-CNN-based methods.", "keywords": [ "Equivariant Neural Networks", "Geometric Deep Learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "t2f7sD9M7n", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Alexander Maeder", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Alexandros Nikolaos Ziogas", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Chen Hao Xia", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Manasa Kaniselvan", "gender": "Female", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Marko Mladenovic", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Mathieu Luisier", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 6, "n_ref_uni": 50, "n_ref": 98, "n_ref_all": 150, "n_fig": 14, "n_tab": 19, "L_tab": 2938, "n_element_tab": 423, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1313, "n_element_tab_1": 38, "formula_len_all": 1316, "formula_len_all_1": 332, "len_all": 228797, "len_all_1": 76685, "len_abs": 1166, "len_title": 128, "len_sents": 72930, "len_sents_1": 35653, "n_sents": 505, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1206, "title": "Learning the Hamiltonian of Disordered Materials with Equivariant Graph Networks", "abs": "Graph neural networks (GNNs) have shown promise in learning the ground-state electronic properties of molecules and crystalline materials, subverting computationally intensive density functional theory (DFT) calculations. Materials with structural disorder, however, are more challenging to learn as they exhibit higher complexity and a more extensive palette of local atomic environments, all of which require large (10+ Angstrom) cells to be accurately captured. In this work, we adapt efficient equivariant GNN approaches to learn disordered materials' electronic properties, represented by the Hamiltonian matrix ($\\mathbf{H}$). Since creating a large graph corresponding to the whole structure of interest would be computationally prohibitive, we introduce an 'augmented partitioning' approach in which the graph is sliced into multiple partitions, each augmented with masked virtual nodes and edges. This method maintains correct atomic neighborhoods within a single message passing layer, allowing for the network to learn the electronic properties of amorphous HfO$_2$ materials with 3,000 nodes (atoms), 500,000+ edges, and $\\sim$28 million orbital interactions (non-zero entries of $\\mathbf{H}$).", "keywords": [ "Materials modeling", "atomic structure", "electronic structure", "density functional theory", "graph neural networks", "hamiltonian", "amorphous systems" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 5, 3, 2 ] }, { "paper_id": "t2TUw5nJsW", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Bingheng Li", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Boyu Wang", "gender": "Male", "institution": "University of Western Ontario", "country": "CA", "position": "Assistant Professor" }, { "name": "Charles Ling", "gender": "Male", "institution": "Western University", "country": "CA", "position": "Professor" }, { "name": "Nima Hosseini Dashtbayaz", "gender": "Male", "institution": "Autodesk", "country": "CA", "position": "Intern" }, { "name": "Ruiyi Fang", "gender": "unknown", "institution": "University of Western Ontario", "country": "CA", "position": "PhD student" }, { "name": "Ruizhi Pu", "gender": "Male", "institution": "University of Western Ontario", "country": "CA", "position": "PhD student" }, { "name": "zhao kang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "QIUHAO Zeng", "gender": "Male", "institution": "University of Western Ontario", "country": "CA", "position": "PhD student" } ], "n_formula": 47, "n_formula_1": 22, "n_ref_uni": 28, "n_ref": 47, "n_ref_all": 56, "n_fig": 2, "n_tab": 5, "L_tab": 2428, "n_element_tab": 435, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 1070, "n_element_tab_1": 136, "formula_len_all": 4296, "formula_len_all_1": 1137, "len_all": 136319, "len_all_1": 60250, "len_abs": 1087, "len_title": 108, "len_sents": 36948, "len_sents_1": 22896, "n_sents": 314, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1092, "title": "On the Benefits of Attribute-Driven Graph Domain Adaptation", "abs": "Graph Domain Adaptation (GDA) addresses a pressing challenge in cross-network learning, particularly pertinent due to the absence of labeled data in real-world graph datasets. Recent studies attempted to learn domain invariant representations by eliminating structural shifts between graphs. In this work, we show that existing methodologies have overlooked the significance of the graph node attribute, a pivotal factor for graph domain alignment. \nSpecifically, we first reveal the impact of node attributes for GDA by theoretically proving that in addition to the graph structural divergence between the domains, the node attribute discrepancy also plays a critical role in GDA. Moreover, we also empirically show that the attribute shift is more substantial than the topology shift, which further underscore the importance of node attribute alignment in GDA. Inspired by this finding, a novel cross-channel module is developed to fuse and align both views between the source and target graphs for GDA. Experimental results on a variety of benchmark verify the effectiveness of our method.", "keywords": [ "Deep Learning and representational learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "t21RmVmJrT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Liangming Pan", "gender": "Male", "institution": "University of Arizona", "country": "US", "position": "Assistant Professor" }, { "name": "Sitao Cheng", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Xinyi Wang", "gender": "Female", "institution": "UC Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Xunjian Yin", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "William Yang Wang", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 71, "n_ref_all": 128, "n_fig": 9, "n_tab": 10, "L_tab": 1873, "n_element_tab": 131, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 1385, "n_element_tab_1": 110, "formula_len_all": 274, "formula_len_all_1": 0, "len_all": 197947, "len_all_1": 68096, "len_abs": 2080, "len_title": 146, "len_sents": 53617, "len_sents_1": 31300, "n_sents": 449, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 946, "title": "Understanding the Interplay between Parametric and Contextual Knowledge for Large Language Models", "abs": "Large language models (LLMs) encode vast amounts of knowledge during pre-training (parametric knowledge or PK) and can further be enhanced by incorporating contextual knowledge (CK). Can LLMs effectively integrate their internal PK with external CK to solve complex problems? In this paper, we investigate the dynamic interaction between PK and CK, categorizing their relationships into Supportive, Complementary, Conflicting, and Irrelevant types. To support this investigation, we introduce EchoQA, a benchmark spanning scientific, factual, and commonsense knowledge. Our results show that LLMs tend to suppress their PK when contextual information is available, even when it is complementary or irrelevant. While tailored instructions can encourage LLMs to rely more on their PK, they still struggle to fully leverage it. These findings reveal a key vulnerability in LLMs, raising concerns about their reliability in knowledge-intensive tasks.", "keywords": [ "Large Language Models", "Parametric Knowledge", "Contextual Knowledge", "Reasoning" ], "rating_list": [ 6, 6, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 1, 2, 2 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "t1nZzR7ico", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Boqing Gong", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "Assistant Professor" }, { "name": "Hyomin Lee", "gender": "Male", "institution": "Korea University", "country": "", "position": "Undergrad student" }, { "name": "Minseon Kim", "gender": "unknown", "institution": "Microsoft", "country": "CA", "position": "Postdoc" }, { "name": "Sung Ju Hwang", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Huishuai Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 30, "n_ref": 58, "n_ref_all": 103, "n_fig": 23, "n_tab": 9, "L_tab": 1245, "n_element_tab": 70, "n_fig_1": 18, "n_tab_1": 3, "L_tab_1": 238, "n_element_tab_1": 17, "formula_len_all": 143, "formula_len_all_1": 113, "len_all": 157865, "len_all_1": 66503, "len_abs": 1692, "len_title": 88, "len_sents": 52678, "len_sents_1": 31493, "n_sents": 433, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1704, "title": "Automatic Jailbreaking of Text-to-Image Generative AI Systems for Copyright Infringement", "abs": "Recent AI systems have shown extremely powerful performance, even surpassing human performance, on various tasks such as information retrieval, language generation, and image generation based on large language models (LLMs). At the same time, there are diverse safety risks that can cause the generation of malicious contents by circumventing the alignment in LLMs, a phenomenon often referred to as jailbreaking. However, most of the previous works only focused on the text-based jailbreaking in LLMs, and the jailbreaking of the text-to-image (T2I) generation system has been relatively overlooked. In this paper, we first evaluate the safety of the commercial T2I generation systems, such as ChatGPT, Copilot, and Gemini, on copyright infringement with naive prompts. From this empirical study, we find that Copilot and Gemini block only 5\\% and 11.25\\% of the attacks with naive prompts, respectively, while ChatGPT blocks 96.25\\% of them. Then, we further propose a stronger automated jailbreaking pipeline for T2I generation systems, which produces prompts that bypass their safety guards. Our automated jailbreaking framework leverages an LLM optimizer to generate prompts that maximize degree of violation from the generated images without any weight updates or gradient computation. Surprisingly, our simple yet effective approach successfully jailbreaks the Copilot and ChatGPT with 0.0\\% and 6.25\\% block rate, respectively, enabling the generation of copyrighted content 73.3\\% of the time. Finally, we explore various defense strategies, such as post-generation filtering and machine unlearning techniques, but find them inadequate, highlighting the necessity of stronger defense mechanisms.", "keywords": [ "Copyright", "jailbreaking", "T2I model" ], "rating_list": [ 5, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 1, 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 2, 4, 4 ] }, { "paper_id": "t1LfiWCYux", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Rishit Garg", "gender": "Male", "institution": "Indian Institute of Technology Kharagpur", "country": "IN", "position": "Undergrad student" }, { "name": "Shehreen Azad", "gender": "Female", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Vibhav Vineet", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Yash Jain", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Yogesh S Rawat", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 43, "n_ref_all": 68, "n_fig": 10, "n_tab": 4, "L_tab": 4093, "n_element_tab": 616, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 2134, "n_element_tab_1": 144, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 123020, "len_all_1": 66224, "len_abs": 552, "len_title": 121, "len_sents": 36179, "len_sents_1": 30222, "n_sents": 262, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1185, "title": "Understanding Depth and Height Perception in Large Visual-Language Models", "abs": "Geometric understanding\u2014including depth and height perception\u2014is fundamental to intelligence and crucial for navigating our environment. Despite the impressive capabilities of large Vision Language Models (VLMs), it remains unclear how well they possess the geometric understanding required for practical applications in visual perception. In this work, we focus on evaluating the geometric understanding of these models, specifically targeting their ability to perceive the depth and height of objects in an image. To address this, we introduce GeoMeter, a suite of benchmark datasets\u2014encompassing 2D and 3D scenarios\u2014to rigorously evaluate these aspects. By benchmarking 18 state-of-the-art VLMs, we found that although they excel in perceiving basic geometric properties like shape and size, they consistently struggle with depth and height perception. Our analysis reveal that these challenges stem from shortcomings in their depth and height reasoning capabilities and inherent biases. This study aims to pave the way for developing VLMs with enhanced geometric understanding by emphasizing depth and height perception as critical components necessary for real-world applications.", "keywords": [ "large vision language models", "frontier models", "visual reasoning", "visual question answering" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "t1J2CnDFwj", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Gengyu Lyu", "gender": "Male", "institution": "Beijing University of Technology", "country": "", "position": "Full Professor" }, { "name": "Haobo Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Huibin Lin", "gender": "Male", "institution": "Fuzhou University", "country": "CN", "position": "PhD student" }, { "name": "Yiyuan Wang", "gender": "Male", "institution": "Beijing University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yongjian Deng", "gender": "Male", "institution": "Beijing University of Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuena Lin", "gender": "Male", "institution": "Beijing University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Zhen Yang", "gender": "Male", "institution": "Beijing University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Hai-Chun Cai", "gender": "Male", "institution": "Fuzhou University", "country": "CN", "position": "PhD student" } ], "n_formula": 45, "n_formula_1": 25, "n_ref_uni": 14, "n_ref": 39, "n_ref_all": 67, "n_fig": 7, "n_tab": 6, "L_tab": 4407, "n_element_tab": 115, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 2019, "n_element_tab_1": 2, "formula_len_all": 2993, "formula_len_all_1": 1256, "len_all": 125960, "len_all_1": 61848, "len_abs": 1385, "len_title": 85, "len_sents": 37851, "len_sents_1": 27619, "n_sents": 291, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1402, "title": "Enhance Multi-View Classification Through Multi-Scale Alignment and Expanded Boundary", "abs": "Multi-view classification aims at unifying the data from multiple views to complementarily enhance the classification performance. Unfortunately, two major problems in multi-view data are damaging model performance. The first is feature heterogeneity, which makes it hard to fuse features from different views. Considering this, we introduce a multi-scale alignment module, including an instance-scale alignment module and a prototype-scale alignment module to mine the commonality from an inter-view perspective and an inter-class perspective respectively, jointly alleviating feature heterogeneity. The second is information redundancy which easily incurs ambiguous data to blur class boundaries and impair model generalization. Therefore, we propose a novel expanded boundary by extending the original class boundary with fuzzy set theory, which adaptively adjusts the boundary to fit ambiguous data. By integrating the expanded boundary into the prototype-scale alignment module, our model further tightens the produced representations and reduces boundary ambiguity. Additionally, compared with the original class boundary, the expanded boundary preserves more margins for classifying unseen data, which guarantees the model generalization. Extensive experiment results across various real-world datasets demonstrate the superiority of the proposed model against existing state-of-the-art methods.", "keywords": [ "Multi-View Classification", "Multi-Scale Alignment", "Fuzzy Logic", "Decision Boundary" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 4, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "t15cWqydys", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Jianxi Gao", "gender": "Male", "institution": ", Rensselaer Polytechnic Institute", "country": "", "position": "Associate Professor" }, { "name": "Mingyu Derek Ma", "gender": "Male", "institution": "Genentech", "country": "US", "position": "Machine Learning Scientist" }, { "name": "Wei Wang", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "Yanna Ding", "gender": "Female", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "PhD student" }, { "name": "Yizhou Sun", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "Zijie Huang", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 35, "n_ref": 55, "n_ref_all": 96, "n_fig": 7, "n_tab": 8, "L_tab": 4481, "n_element_tab": 518, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 286, "n_element_tab_1": 51, "formula_len_all": 7783, "formula_len_all_1": 11061, "len_all": 210276, "len_all_1": 76220, "len_abs": 1270, "len_title": 145, "len_sents": 62337, "len_sents_1": 31636, "n_sents": 437, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1275, "title": "Inferring from Logits: Exploring Best Practices for Decoding-Free Generative Candidate Selection", "abs": "Generative Language Models rely on autoregressive decoding to produce the output sequence token by token. Some tasks, such as preference optimization, require the model to produce task-level output consisting of multiple tokens directly by selecting candidates from a pool as predictions. Determining a task-level prediction from candidates using the ordinary token-level decoding mechanism is constrained by time-consuming decoding and interrupted gradients by discrete token selection. Existing works have been using decoding-free candidate selection methods to obtain candidate probability from initial output logits over vocabulary. Though these estimation methods are widely used, they are not systematically evaluated, especially on end tasks. We introduce an evaluation of a comprehensive collection of decoding-free candidate selection approaches on a comprehensive set of tasks, including five multiple-choice QA tasks with a small candidate pool and four clinical decision tasks with a massive amount of candidates, some with 10k+ options. We evaluate the estimation methods paired with a wide spectrum of foundation LMs covering different architectures, sizes and training paradigms. The results and insights from our analysis could inform the future model design.", "keywords": [ "Decoding", "Benchmark", "Large Language Models", "Autoregressive Decoding" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "szRmEM8Kx5", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Anna Currey", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Corey D Barrett", "gender": "Male", "institution": "Oracle", "country": "US", "position": "Researcher" }, { "name": "Georgiana Dinu", "gender": "Female", "institution": "Amazon", "country": "", "position": "Principal Researcher" }, { "name": "Miguel Romero Calvo", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" }, { "name": "Xing Niu", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Yi Xiang", "gender": "Female", "institution": "Amazon", "country": "US", "position": "amazon web services " } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 35, "n_ref": 66, "n_ref_all": 85, "n_fig": 7, "n_tab": 6, "L_tab": 1014, "n_element_tab": 151, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 308, "n_element_tab_1": 29, "formula_len_all": 661, "formula_len_all_1": 456, "len_all": 132632, "len_all_1": 58956, "len_abs": 717, "len_title": 142, "len_sents": 36705, "len_sents_1": 29275, "n_sents": 259, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 93, "L_abs": 718, "title": "Effective post-training embedding compression via temperature control in contrastive training", "abs": "Fixed-size learned representations (dense representations, or embeddings) are widely used in many machine learning applications across language, vision or speech modalities. This paper investigates the role of the temperature parameter in contrastive training for text embeddings. We shed light on the impact this parameter has on the intrinsic dimensionality of the embedding spaces obtained, and show that lower intrinsic dimensionality is further correlated with effective compression of embeddings. We still observe a trade-off between absolute performance and effective compression and we propose temperature aggregation methods which reduce embedding size by an order of magnitude with minimal impact on quality.", "keywords": [ "representation learning", "embeddings", "text retrieval", "nlp" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 2, 2, 4, 4 ] }, { "paper_id": "sz7HdeVVHo", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Sander van Cranenburgh", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "Associate Professor" }, { "name": "Simeon C. Calvert", "gender": "unknown", "institution": "Delft University of Technology", "country": "NL", "position": "Associate Professor" }, { "name": "Yiru Jiao", "gender": "Female", "institution": "Delft University of Technology", "country": "NL", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 35, "n_ref": 56, "n_ref_all": 97, "n_fig": 1, "n_tab": 13, "L_tab": 8823, "n_element_tab": 909, "n_fig_1": 0, "n_tab_1": 7, "L_tab_1": 6103, "n_element_tab_1": 374, "formula_len_all": 1860, "formula_len_all_1": 1860, "len_all": 146722, "len_all_1": 72190, "len_abs": 1290, "len_title": 113, "len_sents": 35234, "len_sents_1": 28278, "n_sents": 240, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1295, "title": "Structure-preserving contrastive learning for spatial time series", "abs": "Informative representations enhance model performance and generalisability in downstream tasks. However, learning self-supervised representations for spatially characterised time series, like traffic interactions, poses challenges as it requires maintaining fine-grained similarity relations in the latent space. In this study, we incorporate two structure-preserving regularisers for the contrastive learning of spatial time series: one regulariser preserves the topology of similarities between instances, and the other preserves the graph geometry of similarities across spatial and temporal dimensions. To balance contrastive learning and structure preservation, we propose a dynamic mechanism that adaptively weighs the trade-off and stabilises training. We conduct experiments on multivariate time series classification, as well as macroscopic and microscopic traffic prediction. For all three tasks, our approach preserves the structures of similarity relations more effectively and improves state-of-the-art task performances. This approach can be applied to an arbitrary encoder and is particularly beneficial for time series with spatial or geographical features. Our code is attached as supplementary material, which will be made openly available with all resulting data after review.", "keywords": [ "time series", "contrastive learning", "spatio-temporal data", "traffic interaction" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "synCTX1JqO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Dan Deng", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Deng Cai", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" }, { "name": "Haifeng Liu", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Liang Xie", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "Postdoc" }, { "name": "Shenqi Lai", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Wei Qian", "gender": "Male", "institution": "Fabu Inc.", "country": "CN", "position": "Researcher" }, { "name": "Yuhang Pei", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhanwei Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zheng Yang", "gender": "Male", "institution": "Fabu Inc", "country": "CN", "position": "CTO" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 49, "n_ref": 116, "n_ref_all": 130, "n_fig": 6, "n_tab": 13, "L_tab": 6801, "n_element_tab": 862, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 9165, "n_element_tab_1": 378, "formula_len_all": 273, "formula_len_all_1": 273, "len_all": 143378, "len_all_1": 56030, "len_abs": 1547, "len_title": 124, "len_sents": 25374, "len_sents_1": 19630, "n_sents": 225, "n_sents_1": 142, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1624, "title": "RaCNN: Region-aware Convolutional Neural Network with Global Receptive Field", "abs": "Recent Convolutional Neural Networks (CNNs) utilize large-kernel convolutions (e.g., 101 kernel convolutions) to simulate a large receptive field of Vision Transformers (ViTs). \n However, these models introduce specialized techniques like re-parameterization, sparsity, and weight decomposition, increasing the complexity of the training and inference stages. \n To address this challenge, we propose Region-aware CNN (RaCNN), which achieves a global receptive field without requiring extra complexity, yet surpasses state-of-the-art models. \n Specifically, we design two novel modules to capture global visual dependencies. \n The first is the Region-aware Feed Forward Network (RaFFN). \n It uses a novel Region Point-Wise Convolution (RPWConv) to capture global visual cues in a region-aware manner. \n In contrast, traditional PWConv shares the same weights for all spatial pixels and cannot capture spatial information. \n The second is the Region-aware Gated Linear Unit (RaGLU). \n This channel mixer captures long-range visual dependencies in a sparse global manner and can become a better substitute for the original FFN. \n Under only 84\\% computational complexity, RaCNN significantly outperforms the state-of-the-art CNN model MogaNet (83.9\\% vs. 83.4\\%). \n It also demonstrates good scalability and surpasses existing state-of-the-art lightweight models. \n Furthermore, our RaCNN shows comparability with state-of-the-art ViTs, MLPs, and Mambas in object detection, instance segmentation, and semantic segmentation. \n All codes and logs are released in the supplementary materials.", "keywords": [ "Convolutional Neural Network", "Global Receptive Field", "Backbone" ], "rating_list": [ 5, 5, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 1, 2, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 5, 3, 5 ] }, { "paper_id": "syUJqBnuD6", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Daqi Liu", "gender": "Male", "institution": "Xiaomi Corporation", "country": "CN", "position": "Researcher" }, { "name": "Dongrui Liu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Hongwei Xie", "gender": "Male", "institution": "Xiaomi Corporation", "country": "CN", "position": "Researcher" }, { "name": "Lei Chu", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Postdoc" }, { "name": "Xueqian Li", "gender": "unknown", "institution": "University of Adelaide", "country": "AU", "position": "PhD student" }, { "name": "bing wang", "gender": "Male", "institution": "Xiaomi Corporation", "country": "CN", "position": "Researcher" }, { "name": "Lin Sihao", "gender": "Male", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "PhD student" }, { "name": "Xiaojun Chang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Full Professor" } ], "n_formula": 67, "n_formula_1": 26, "n_ref_uni": 57, "n_ref": 195, "n_ref_all": 242, "n_fig": 4, "n_tab": 7, "L_tab": 2330, "n_element_tab": 343, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2240, "n_element_tab_1": 334, "formula_len_all": 4531, "formula_len_all_1": 1436, "len_all": 180654, "len_all_1": 71316, "len_abs": 1181, "len_title": 109, "len_sents": 50595, "len_sents_1": 28921, "n_sents": 434, "n_sents_1": 223, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1186, "title": "Multi-Frame Neural Scene Flow: Learning Bounds and Algorithms", "abs": "Although Neural Scene Flow Prior (NSFP) and its variants have shown remarkable performance in large out-of-distribution autonomous driving, the underlying explanation for their generalization capabilities remains unclear. To this end, we analyze the generalization capabilities of NSFP via uniform stability and find that it exhibits a generalization bound, which is inversely proportional to the number of point clouds. These findings provide solid theoretical evidence to explain the effectiveness of NSFP in large-scale point cloud scene flow estimation tasks for the first time. To enhance practical scene understanding, we extend NSFP and propose a multi-frame neural scene flow (MNSF) scheme, which extracts temporal information across multiple frames. In this way, MNSF has better temporal consistency than NSFP.\nMoreover, we theoretically analyze its generalization abilities and demonstrate that it achieves a tight generalization bound with a convergence rate similar to NSFP. Extensive experimental results on large-scale autonomous driving Waymo Open and Argoverse datasets demonstrate that MNSF achieves state-of-the-art performance. The code is attached to the submission.", "keywords": [ "Multi-Frame Neural Scene Flow", "Spatial and Temporal Feature", "Generalization Bound", "Large-Scale Point Clouds." ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "syThiTmWWm", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Chao Du", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Jing Jiang", "gender": "Female", "institution": "Australian National University", "country": "AU", "position": "Full Professor" }, { "name": "Min Lin", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Principal Researcher" }, { "name": "Qian Liu", "gender": "Male", "institution": "Tiktok", "country": "SG", "position": "Researcher" }, { "name": "Tianyu Pang", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Xiaosen Zheng", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 2, "n_ref_uni": 61, "n_ref": 130, "n_ref_all": 186, "n_fig": 22, "n_tab": 7, "L_tab": 2855, "n_element_tab": 223, "n_fig_1": 15, "n_tab_1": 3, "L_tab_1": 1157, "n_element_tab_1": 117, "formula_len_all": 1424, "formula_len_all_1": 713, "len_all": 255240, "len_all_1": 69969, "len_abs": 1437, "len_title": 117, "len_sents": 72980, "len_sents_1": 32185, "n_sents": 627, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 69, "L_abs": 1372, "title": "Cheating Automatic LLM Benchmarks: Null Models Achieve High Win Rates", "abs": "Automatic LLM benchmarks, such as AlpacaEval 2.0, Arena-Hard-Auto, and MT-Bench, have become popular for evaluating language models due to their cost-effectiveness and scalability compared to human evaluation. Achieving high win rates on these benchmarks can significantly boost the promotional impact of newly released language models. This promotional benefit may motivate tricks, such as manipulating model output length or style to game win rates, even though several mechanisms have been developed to control length and disentangle style to reduce gameability. Nonetheless, we show that even a **\"null model\"** that always outputs a **constant** response (*irrelevant to input instructions*) can cheat automatic benchmarks and achieve top-ranked win rates: an $86.5\\\\%$ LC win rate on AlpacaEval 2.0; an $83.0$ score on Arena-Hard-Auto; and a $9.55$ score on MT-Bench. Moreover, the crafted cheating outputs are **transferable** because we assume that the instructions of these benchmarks (e.g., $805$ samples of AlpacaEval 2.0) are *private* and cannot be accessed. While our experiments are primarily proof-of-concept, an adversary could use LLMs to generate more imperceptible cheating responses, unethically benefiting from high win rates and promotional impact. Our findings call for the development of anti-cheating mechanisms for reliable automatic benchmarks.", "keywords": [ "Large Language Models", "Cheating", "Automatic LLM Benchmarks" ], "rating_list": [ 8, 6, 6, 6, 8 ], "soundness_list": [ 4, 3, 3, 4, 3 ], "presentation_list": [ 3, 4, 2, 4, 3 ], "contribution_list": [ 4, 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 4, 3 ] }, { "paper_id": "syMZF5fc8y", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Andr\u00e9 Panisson", "gender": "unknown", "institution": "CENTAI", "country": "IT", "position": "Principal Researcher" }, { "name": "Megha Khosla", "gender": "Female", "institution": "Delft University of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Simone Piaggesi", "gender": "Male", "institution": "University of Pisa", "country": "IT", "position": "Postdoc" } ], "n_formula": 31, "n_formula_1": 28, "n_ref_uni": 54, "n_ref": 117, "n_ref_all": 152, "n_fig": 9, "n_tab": 9, "L_tab": 9194, "n_element_tab": 328, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1046, "n_element_tab_1": 43, "formula_len_all": 1592, "formula_len_all_1": 1505, "len_all": 174489, "len_all_1": 75392, "len_abs": 1097, "len_title": 110, "len_sents": 49627, "len_sents_1": 35008, "n_sents": 346, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1156, "title": "Disentangled and Self-Explainable Node Representation Learning", "abs": "Node representations, or embeddings, are low-dimensional vectors that capture node properties, typically learned through unsupervised structural similarity objectives or supervised tasks. While recent efforts have focused on explaining graph model decisions, the interpretability of $\\textit{unsupervised}$ node embeddings remains underexplored. To bridge this gap, we introduce DiSeNE ($\\textbf{Di}$sentangled and $\\textbf{Se}$lf-Explainable $\\textbf{N}$ode $\\textbf{E}$mbedding), a framework that generates self-explainable embeddings in an unsupervised manner. Our method employs disentangled representation learning to produce dimension-wise interpretable embeddings, where each dimension is aligned with distinct topological structure of the graph. We formalize novel desiderata for disentangled and interpretable embeddings, which drive our new objective functions, optimizing simultaneously for both interpretability and disentanglement. Additionally, we propose several new metrics to evaluate representation quality and human interpretability. Extensive experiments across multiple benchmark datasets demonstrate the effectiveness of our approach.", "keywords": [ "Node representation learning", "explainable ai", "disentangled learning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "syC2764fPc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Dongxiao Zhang", "gender": "Male", "institution": "Eastern Institute for Advanced Study", "country": "CN", "position": "Full Professor" }, { "name": "Jinyue Yan", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Chair Professor" }, { "name": "Qian Li", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yuntian Chen", "gender": "Male", "institution": "Eastern Institute for Advanced Study", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuxiao Hu", "gender": "unknown", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 39, "n_ref": 84, "n_ref_all": 110, "n_fig": 4, "n_tab": 27, "L_tab": 17017, "n_element_tab": 1159, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 4335, "n_element_tab_1": 152, "formula_len_all": 1633, "formula_len_all_1": 975, "len_all": 210379, "len_all_1": 71328, "len_abs": 1844, "len_title": 124, "len_sents": 59499, "len_sents_1": 30562, "n_sents": 483, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1873, "title": "Context-Alignment: Activating and Enhancing LLMs Capabilities in Time Series", "abs": "Recently, leveraging pre-trained Large Language Models (LLMs) for time series (TS) tasks has gained increasing attention, which involves activating and enhancing LLMs' capabilities. Many methods aim to activate LLMs' capabilities based on token-level alignment, but overlook LLMs' inherent strength in natural language processing \u2014 their deep understanding of linguistic logic and structure rather than superficial embedding processing. We propose Context-Alignment (CA), a new paradigm that aligns TS with a linguistic component in the language environments familiar to LLMs to enable LLMs to contextualize and comprehend TS data, thereby activating their capabilities. Specifically, such context-level alignment comprises structural alignment and logical alignment, which is achieved by Dual-Scale Context-Alignment GNNs (DSCA-GNNs) applied to TS-language multimodal inputs. Structural alignment utilizes dual-scale nodes to describe hierarchical structure in TS-language, enabling LLMs to treat long TS data as a whole linguistic component while preserving intrinsic token features. Logical alignment uses directed edges to guide logical relationships, ensuring coherence in the contextual semantics. Following the DSCA-GNNs framework, we propose an instantiation method of CA, termed Few-Shot prompting Context-Alignment (FSCA), to enhance the capabilities of pre-trained LLMs in handling TS tasks. FSCA can be flexibly and repeatedly integrated into various layers of pre-trained LLMs to improve awareness of logic and structure, thereby enhancing performance. Extensive experiments show the effectiveness of FSCA and the importance of Context-Alignment across tasks, particularly in few-shot and zero-shot forecasting, confirming that Context-Alignment provides powerful prior knowledge on context. The code is open-sourced at https://github.com/tokaka22/ICLR25-FSCA.", "keywords": [ "Time Series", "Large Language Models", "Context-Alignment" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 2, 2, 2 ] }, { "paper_id": "sy1lbQxj9J", "primary_area": "learning theory", "n_author": 1, "author_info_list": [ { "name": "Song Mei", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" } ], "n_formula": 207, "n_formula_1": 58, "n_ref_uni": 69, "n_ref": 161, "n_ref_all": 195, "n_fig": 3, "n_tab": 1, "L_tab": 620, "n_element_tab": 1, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 668, "n_element_tab_1": 8, "formula_len_all": 23261, "formula_len_all_1": 4400, "len_all": 254590, "len_all_1": 69605, "len_abs": 1378, "len_title": 164, "len_sents": 56858, "len_sents_1": 27467, "n_sents": 613, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 114, "L_abs": 1386, "title": "U-Nets as Belief Propagation: Efficient Classification, Denoising, and Diffusion in Generative Hierarchical Models", "abs": "U-Nets are among the most widely used architectures in computer vision, renowned for their exceptional performance in applications such as image segmentation, denoising, and diffusion modeling. However, a theoretical explanation of the U-Net architecture design has not yet been fully established. \n\nThis paper introduces a novel interpretation of the U-Net architecture by studying certain generative hierarchical models, which are tree-structured graphical models extensively utilized in both language and image domains. With their encoder-decoder structure, long skip connections, and pooling and up-sampling layers, we demonstrate how U-Nets can naturally implement the belief propagation denoising algorithm in such generative hierarchical models, thereby efficiently approximating the denoising functions. This leads to an efficient sample complexity bound for learning the denoising function using U-Nets within these models. Additionally, we discuss the broader implications of these findings for diffusion models in generative hierarchical models. We also demonstrate that the conventional architecture of convolutional neural networks (ConvNets) is ideally suited for classification tasks within these models. This offers a unified view of the roles of ConvNets and U-Nets, highlighting the versatility of generative hierarchical models in modeling complex data distributions.", "keywords": [ "diffusion model", "denoising", "belief propagation", "generative hierarchical models" ], "rating_list": [ 6, 6, 8, 6, 6 ], "soundness_list": [ 3, 2, 4, 3, 3 ], "presentation_list": [ 4, 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 2, 2, 4, 2, 2 ] }, { "paper_id": "sxus3NNiuf", "primary_area": "reinforcement learning", "n_author": 11, "author_info_list": [ { "name": "Deunsol Yoon", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Researcher" }, { "name": "Geon-hyeong Kim", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Researcher" }, { "name": "Jeonghye Kim", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jongseong Chae", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Kanghoon Lee", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Principal Researcher" }, { "name": "Sunghoon Hong", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Researcher" }, { "name": "Whiyoung Jung", "gender": "Male", "institution": "LG AI Research", "country": "", "position": "Researcher" }, { "name": "Woohyung Lim", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Vice President" }, { "name": "Yongjae Shin", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Youngchul Sung", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Youngsoo Jang", "gender": "unknown", "institution": "LG AI Research", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 24, "n_ref": 62, "n_ref_all": 89, "n_fig": 9, "n_tab": 11, "L_tab": 5777, "n_element_tab": 560, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1912, "n_element_tab_1": 144, "formula_len_all": 528, "formula_len_all_1": 528, "len_all": 130798, "len_all_1": 60203, "len_abs": 1284, "len_title": 112, "len_sents": 38273, "len_sents_1": 24313, "n_sents": 290, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1289, "title": "Online Pre-Training for Offline-to-Online Reinforcement Learning", "abs": "Reinforcement Learning (RL) has achieved notable success in tasks requiring complex decision making, with offline RL offering the ability to train agents using fixed datasets, thereby avoiding the risks and costs associated with online interactions. However, offline RL is inherently limited by the quality of the dataset, which can restrict an agent\u2019s performance. Offline-to-online RL aims to bridge the gap between the cost-efficiency of offline RL and the performance potential of online RL by pre-training an agent offline before fine-tuning it through online interactions. Despite its promise, recent studies show that offline pre-trained agents often underperform during online fine-tuning due to inaccurate value function, with random initialization proving more effective in certain cases. In this work, we propose a novel method, Online Pre-Training for Offline-to-Online RL (OPT), to address the issue of inaccurate value estimation in offline pre-trained agents. OPT introduces a new learning phase, Online Pre-Training, which allows the training of a new value function that enhances the subsequent fine-tuning process. Implementation of OPT on TD3 and SPOT demonstrates an average 30\\% improvement in performance across D4RL environments, such as MuJoCo, Antmaze, and Adroit.", "keywords": [ "Reinforcement Learning", "Offline-to-Online Reinforcement Learning", "Online Pre-Training", "Online Fine-Tuning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "sx8dtyZT41", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Daizong Liu", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Lixing Chen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Pan Zhou", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Professor" }, { "name": "Ruihan Gong", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Tailai Zhang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Wen Yin", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Xingxing Wei", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Yuxuan Jiang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 21, "n_ref": 43, "n_ref_all": 57, "n_fig": 7, "n_tab": 1, "L_tab": 427, "n_element_tab": 72, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 427, "n_element_tab_1": 72, "formula_len_all": 661, "formula_len_all_1": 669, "len_all": 148692, "len_all_1": 54562, "len_abs": 1658, "len_title": 142, "len_sents": 30740, "len_sents_1": 25240, "n_sents": 237, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1634, "title": "DEAL: High-Efficacy Privacy Attack on Retrieval-Augmented Generation Systems via LLM Optimizer", "abs": "Retrieval-Augmented Generation (RAG) technology provides a powerful means of combining private databases with large language models (LLMs). \nIn a typical RAG system, a set of documents is retrieved from a private database and inserted into the final prompt, which is then fed into the LLM.\nExisting research has shown that an attacker can use a simple manually designed attack suffix to induce LLM to output private documents in prompt with high probability.\nHowever, in this paper, we demonstrate that the privacy leakage risk exhibited by using this simple manual attack suffix is significantly underestimated.\nWe propose a novel attack method called Documents Extraction Attack via LLM-Optimizer (DEAL). \nDEAL leverages an LLM as optimizer to iteratively refine attack strings, inducing the RAG model to reveal private data in its responses. \nNotably, our attack method does not require any knowledge about the target LLM, including its gradient information or model type. \nInstead, the attack can be executed solely through query access to the RAG model. \nWe evaluate the effectiveness of our attack on multiple LLM architectures, including Qwen2, Llama3.1, and GPT-4o, across different attack tasks such as Entire Documents Extraction and Private Identity Information (PII) Extraction. \nUnder the same permission setting as the existing method, the Mean Rouge-L Recall (MRR) of our method can reach more than 0.95 on average in the Entire Documents Extraction task, and we can steal PII from the retrieved documents with close to 99\\% accuracy in the PII Extraction task, highlighting the risk of privacy leakage in RAG systems.", "keywords": [ "Retrieval-Augmented Generation", "Data Privacy" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 2, 5 ] }, { "paper_id": "sx2jXZuhIx", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "David Futschik", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Feitong Tan", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Peng Dai", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Qiangeng Xu", "gender": "Male", "institution": "Waymo", "country": "US", "position": "Researcher" }, { "name": "Ruofei Du", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Sean Fanello", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist and Manager" }, { "name": "XIAOJUAN QI", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Yinda Zhang", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 46, "n_ref": 118, "n_ref_all": 168, "n_fig": 24, "n_tab": 3, "L_tab": 462, "n_element_tab": 27, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 343, "n_element_tab_1": 13, "formula_len_all": 412, "formula_len_all_1": 384, "len_all": 173503, "len_all_1": 64692, "len_abs": 1325, "len_title": 113, "len_sents": 50382, "len_sents_1": 30735, "n_sents": 408, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1260, "title": "SVG: 3D Stereoscopic Video Generation via Denoising Frame Matrix", "abs": "Video generation models have demonstrated great capability of producing impressive monocular videos, however, the generation of 3D stereoscopic video remains under-explored. We propose a pose-free and training-free approach for generating 3D stereoscopic videos using an off-the-shelf monocular video generation model. Our method warps a generated monocular video into camera views on stereoscopic baseline using estimated video depth, and employs a novel frame matrix video inpainting framework. The framework leverages the video generation model to inpaint frames observed from different timestamps and views. This effective approach generates consistent and semantically coherent stereoscopic videos without scene optimization or model fine-tuning. Moreover, we develop a disocclusion boundary re-injection scheme that further improves the quality of video inpainting by alleviating the negative effects propagated from disoccluded areas in the latent space. We validate the efficacy of our proposed method by conducting experiments on videos from various generative models, including Sora [4], Lumiere [2], WALT [8], and Zeroscope [12]. The experiments demonstrate that our method has a significant improvement over previous methods. Code will be released.", "keywords": [ "video generation", "stereoscopic video", "inpainting", "diffusion model" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "swvURjrt8z", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Hengrui Zhang", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "", "position": "PhD student" }, { "name": "Juntong Shi", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Undergrad student" }, { "name": "Jure Leskovec", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Minkai Xu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Stefano Ermon", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Shuo Harper Hua", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "MS student" } ], "n_formula": 37, "n_formula_1": 26, "n_ref_uni": 31, "n_ref": 75, "n_ref_all": 121, "n_fig": 10, "n_tab": 14, "L_tab": 6701, "n_element_tab": 770, "n_fig_1": 2, "n_tab_1": 9, "L_tab_1": 3843, "n_element_tab_1": 479, "formula_len_all": 3275, "formula_len_all_1": 2203, "len_all": 169949, "len_all_1": 76289, "len_abs": 1361, "len_title": 114, "len_sents": 54820, "len_sents_1": 30197, "n_sents": 422, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1424, "title": "TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation", "abs": "Synthesizing high-quality tabular data is an important topic in many data science tasks, ranging from dataset augmentation to privacy protection. However, developing expressive generative models for tabular data is challenging due to its inherent heterogeneous data types, complex inter-correlations, and intricate column-wise distributions. In this paper, we introduce TabDiff, a joint diffusion framework that models all mixed-type distributions of tabular data in one model. Our key innovation is the development of a joint continuous-time diffusion process for numerical and categorical data, where we propose feature-wise learnable diffusion processes to counter the high disparity of different feature distributions. TabDiff is parameterized by a transformer handling different input types, and the entire framework can be efficiently optimized in an end-to-end fashion. We further introduce a mixed-type stochastic sampler to automatically correct the accumulated decoding error during sampling, and propose classifier-free guidance for conditional missing column value imputation. Comprehensive experiments on seven datasets demonstrate that TabDiff achieves superior average performance over existing competitive baselines across all eight metrics, with up to $22.5\\%$ improvement over the state-of-the-art model on pair-wise column correlation estimations. Code is available at https://github.com/MinkaiXu/TabDiff.", "keywords": [ "Tabular Representative Learning", "Generative Models", "Diffusion Models" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3, 3 ] }, { "paper_id": "swqZKDoMJA", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Ali Zindari", "gender": "Male", "institution": "CISPA, saarland university, saarland informatics campus", "country": "DE", "position": "Researcher" }, { "name": "Anton Rodomanov", "gender": "unknown", "institution": "CISPA", "country": "DE", "position": "Postdoc" }, { "name": "Parham Yazdkhasti", "gender": "Male", "institution": "CISPA, saarland university, saarland informatics campus", "country": "", "position": "Researcher" }, { "name": "Sebastian U Stich", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Tenure Track Faculty" }, { "name": "Tatjana Chavdarova", "gender": "Female", "institution": "University of California, Berkeley", "country": "", "position": "Postdoc" } ], "n_formula": 145, "n_formula_1": 24, "n_ref_uni": 44, "n_ref": 68, "n_ref_all": 86, "n_fig": 9, "n_tab": 8, "L_tab": 2096, "n_element_tab": 121, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 28, "n_element_tab_1": 2, "formula_len_all": 16476, "formula_len_all_1": 1785, "len_all": 255511, "len_all_1": 68458, "len_abs": 1129, "len_title": 113, "len_sents": 74218, "len_sents_1": 29390, "n_sents": 772, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1151, "title": "Decoupled SGDA for Games with Intermittent Strategy Communication", "abs": "We focus on reducing communication overhead in multiplayer games, where frequently exchanging strategies between players is not feasible and players have noisy or outdated strategies of the other players.\nWe propose \\textit{Decoupled SGDA}, an extension of Stochastic Gradient Descent Ascent (SGDA), where players perform independent updates using outdated strategies of opponents, with periodic strategy synchronization.\nFor Strongly-Convex-Strongly-Concave (SCSC) games, we demonstrate that Decoupled SGDA achieves near-optimal communication complexity comparable to the best-known GDA rates.\nFor \\emph{weakly coupled} games where the interaction between players is lower relative to non-interactive part of the game, Decoupled SGDA significantly reduces communication costs compared to standard SGDA. \nOur findings extend to multi-player games. To provide insights into the effect of communication frequency and convergence, we extensively study the convergence of Decoupled SGDA for quadratic minimax problems. \nLastly, in settings where the noise over the players is imbalanced, Decoupled SGDA significantly outperforms federated minimax methods.", "keywords": [ "optimization", "minimax optimization", "distributed games", "distributed optimization" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "swdMzQUhBx", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 10, "author_info_list": [ { "name": "Kai Mei", "gender": "unknown", "institution": "Rutgers University, New Brunswick", "country": "", "position": "PhD student" }, { "name": "Kun Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Min Xu", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "Full Professor" }, { "name": "Mingyu Jin", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Wenyue Hua", "gender": "Female", "institution": "University of California, Santa Barbara", "country": "US", "position": "Postdoc" }, { "name": "Wujiang Xu", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Xi Zhu", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Xuying Ning", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Yongfeng Zhang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Associate Professor" }, { "name": "Yunxiao Shi", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 59, "n_ref": 125, "n_ref_all": 150, "n_fig": 2, "n_tab": 9, "L_tab": 5426, "n_element_tab": 147, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 492, "n_element_tab_1": 50, "formula_len_all": 850, "formula_len_all_1": 684, "len_all": 206569, "len_all_1": 60509, "len_abs": 2316, "len_title": 114, "len_sents": 57986, "len_sents_1": 27703, "n_sents": 409, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 2272, "title": "iAgent: LLM Agent as a Shield between User and Recommender Systems", "abs": "Traditional recommender systems usually take the user-platform paradigm, where users are directly exposed under the control of the platform\u2019s recommendation algorithms. However, the defect of recommendation algorithms may put users in very vulnerable positions under this paradigm. First, many sophisticated models are often designed with commercial objectives in mind, focusing on the platform\u2019s\nbenefits, which may hinder their ability to protect and capture users\u2019 true interests. Second, these models are typically optimized using data from all users, which may overlook individual user\u2019s preferences. Due to these shortcomings, users may experience several disadvantages under the traditional user-platform direct exposure paradigm, such as lack of control over the recommender system, potential manipulation by the platform, echo chamber effects, or lack of personalization for less active users due to the dominance of active users during collaborative learning. Therefore, there is an urgent need to develop a new paradigm to protect user interests and alleviate these issues. Recently, some researchers have introduced LLM agents to simulate user behaviors, these approaches primarily aim to optimize platform-side performance, leaving core issues in recommender systems unresolved. To address these limitations, we propose a new user-agent-platform paradigm, where agent serves as the protective shield between user and recommender system\nthat enables indirect exposure. To this end, we first construct four recommendation datasets, denoted as InstructRec, along with user instructions for each record. To understand user\u2019s intention, we design an Instruction-aware Agent (iAgent) capable of using tools to acquire knowledge from external environments. Moreover, we introduce an Individual Instruction-aware Agent ( i$^2$Agent), which incorporates a dynamic memory mechanism to optimize from individual feedback. Results on four InstructRec datasets demonstrate that i2Agent consistently achieves an average improvement of 16.6% over SOTA baselines across ranking metrics. Moreover, i$^2$Agent mitigates echo chamber effects and effectively alleviates the model bias in disadvantaged users (less-active), serving as a shield between user and recommender systems.", "keywords": [ "Large Language Model; LLM-based Agents; Memory Mechanism" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "swWF948IiC", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Yang Liu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "Yevgeniy Vorobeychik", "gender": "Male", "institution": "Washington University, St. Louis", "country": "US", "position": "Full Professor" }, { "name": "Zonglin Di", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Sixie Yu", "gender": "Male", "institution": "Washington University, St. Louis", "country": "", "position": "Research Assistant" } ], "n_formula": 16, "n_formula_1": 9, "n_ref_uni": 35, "n_ref": 101, "n_ref_all": 126, "n_fig": 5, "n_tab": 8, "L_tab": 11120, "n_element_tab": 288, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 5363, "n_element_tab_1": 125, "formula_len_all": 964, "formula_len_all_1": 594, "len_all": 165954, "len_all_1": 83316, "len_abs": 1400, "len_title": 78, "len_sents": 44209, "len_sents_1": 34368, "n_sents": 360, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 30, "L_abs": 1408, "title": "Adversarial Machine Unlearning", "abs": "This paper focuses on the challenge of machine unlearning, aiming to remove the influence of specific training data on machine learning models. Traditionally, the development of unlearning algorithms runs parallel with that of membership inference attacks (MIA), a type of privacy threat to determine whether a data instance was used for training. However, the two strands are intimately connected: one can view machine unlearning through the lens of MIA success with respect to removed data. Recognizing this connection, we propose a game-theoretic framework that integrates MIAs into the design of unlearning algorithms. Specifically, we model the unlearning problem as a Stackelberg game in which an unlearner strives to unlearn specific training data from a model, while an auditor employs MIAs to detect the traces of the ostensibly removed data. Adopting this adversarial perspective allows the utilization of new attack advancements, facilitating the design of unlearning algorithms. Our framework stands out in two ways. First, it takes an adversarial approach and proactively incorporates the attacks into the design of unlearning algorithms. Secondly, it uses implicit differentiation to obtain the gradients that limit the attacker's success, thus benefiting the process of unlearning. We present empirical results to demonstrate the effectiveness of the proposed approach for machine unlearning.", "keywords": [ "Machine unlearning", "Adversarial approach", "Stackelberg game" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "sw6Wpx2LGr", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Dongmin Park", "gender": "Male", "institution": "KRAFTON", "country": "KR", "position": "Researcher" }, { "name": "Guangxing Han", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Zhaofang Qian", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Ser-Nam Lim", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 10, "n_ref_uni": 49, "n_ref": 98, "n_ref_all": 126, "n_fig": 11, "n_tab": 14, "L_tab": 3130, "n_element_tab": 278, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 2016, "n_element_tab_1": 176, "formula_len_all": 766, "formula_len_all_1": 1200, "len_all": 181064, "len_all_1": 78264, "len_abs": 1107, "len_title": 102, "len_sents": 44013, "len_sents_1": 32836, "n_sents": 366, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1111, "title": "Mitigating Dialogue Hallucination for Large Vision Language Models via Adversarial Instruction Tuning", "abs": "Mitigating hallucinations of Large Vision Language Models (LVLMs) is crucial to enhance their reliability for general-purpose assistants. This paper shows that such hallucinations of LVLMs can be significantly exacerbated by preceding user-system dialogues. To precisely measure this, we first present an evaluation benchmark by extending popular multi-modal benchmark datasets with prepended hallucinatory dialogues powered by our novel Adversarial Question Generator (AQG), which can automatically generate image-related yet adversarial dialogues by adopting adversarial attacks on LVLMs. On our benchmark, the zero-shot performance of state-of-the-art LVLMs drops significantly for both the VQA and Captioning tasks. Next, we further reveal this hallucination is mainly due to the prediction bias toward preceding dialogues rather than visual content. To reduce this bias, we propose Adversarial Instruction Tuning (AIT) that robustly fine-tunes LVLMs against hallucinatory dialogues. Extensive experiments show our proposed approach successfully reduces dialogue hallucination while maintaining performance.", "keywords": [ "Large Vision Language Model", "Large Multi-modal Model", "Hallucination", "Adversarial Learning" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "svp1EBA6hA", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Ehsan Hajiramezanali", "gender": "Male", "institution": "Genentech", "country": "US", "position": "Principal Researcher" }, { "name": "Gabriele Scalia", "gender": "unknown", "institution": "Genentech", "country": "", "position": "Researcher" }, { "name": "Masatoshi Uehara", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Sunyuan Kung", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Full Professor" }, { "name": "Tommaso Biancalani", "gender": "Male", "institution": "Genentech", "country": "", "position": "Director" }, { "name": "Yulai Zhao", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" } ], "n_formula": 56, "n_formula_1": 14, "n_ref_uni": 56, "n_ref": 147, "n_ref_all": 184, "n_fig": 6, "n_tab": 3, "L_tab": 1531, "n_element_tab": 134, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5633, "formula_len_all_1": 1129, "len_all": 212849, "len_all_1": 66659, "len_abs": 4614, "len_title": 123, "len_sents": 59717, "len_sents_1": 30376, "n_sents": 527, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1415, "title": "Adding Conditional Control to Diffusion Models with Reinforcement Learning", "abs": "Diffusion models are powerful generative models that allow for precise control over the characteristics of the generated samples. While these diffusion models trained on large datasets have achieved success, there is often a need to introduce additional controls in downstream fine-tuning processes, treating these powerful models as pre-trained diffusion models. This work presents a novel method based on reinforcement learning (RL) to add such controls using an offline dataset comprising inputs and labels. We formulate this task as an RL problem, with the classifier learned from the offline dataset and the KL divergence against pre-trained models serving as the reward functions. Our method, **CTRL** (**C**onditioning pre-**T**rained diffusion models with **R**einforcement **L**earning), produces soft-optimal policies that maximize the abovementioned reward functions. We formally demonstrate that our method enables sampling from the conditional distribution with additional controls during inference.\nOur RL-based approach offers several advantages over existing methods. Compared to classifier-free guidance,\nit improves sample efficiency and can greatly simplify dataset construction by leveraging conditional independence between the inputs and additional controls. Additionally, unlike classifier guidance, it eliminates the need to train classifiers from intermediate states to additional controls.", "keywords": [ "diffusion models", "conditional generation", "fine-tuning", "reinforcement learning" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 2, 4, 3, 2 ] }, { "paper_id": "suz4utPr9Y", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Christopher Lott", "gender": "unknown", "institution": "QualComm", "country": "", "position": "Researcher" }, { "name": "Hanghang Tong", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "James Ezick", "gender": "Male", "institution": "Qualcomm Technologies, Inc.", "country": "US", "position": "Researcher" }, { "name": "Ruizhong Qiu", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Weiliang Will Zeng", "gender": "unknown", "institution": "QualComm", "country": "", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 10, "n_ref_uni": 42, "n_ref": 105, "n_ref_all": 145, "n_fig": 2, "n_tab": 11, "L_tab": 4636, "n_element_tab": 507, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2537, "n_element_tab_1": 301, "formula_len_all": 1214, "formula_len_all_1": 935, "len_all": 199471, "len_all_1": 65472, "len_abs": 2204, "len_title": 121, "len_sents": 55245, "len_sents_1": 27823, "n_sents": 442, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1700, "title": "How efficient is LLM-generated code? A rigorous & high-standard benchmark", "abs": "The emergence of large language models (LLMs) has significantly pushed the frontiers of program synthesis. Advancement of LLM-based program synthesis calls for a thorough evaluation of LLM-generated code. Most evaluation frameworks focus on the (functional) correctness of generated code; efficiency, as an important measure of code quality, has been overlooked in existing evaluations. In this work, we develop ENAMEL (EfficeNcy AutoMatic EvaLuator), a rigorous and high-standard benchmark for evaluating the capability of LLMs in generating efficient code. Firstly, we propose a new efficiency metric called eff@k, which generalizes the pass@k metric from correctness to efficiency and appropriately handles right-censored execution time. Furthermore, we derive an unbiased and variance-reduced estimator of eff@k via Rao\u2013Blackwellization; we also provide a numerically stable implementation for the new estimator. Secondly, to set a high-standard for efficiency evaluation, we employ a human expert to design best algorithms and implementations as our reference solutions of efficiency, many of which are much more efficient than existing canonical solutions in HumanEval and HumanEval+. Moreover, to ensure a rigorous evaluation, we employ a human expert to curate strong test case generators to filter out wrong code and differentiate suboptimal algorithms. An extensive study across 30 popular LLMs using our benchmark ENAMEL shows that LLMs still fall short of generating expert-level efficient code. Using two subsets of our problem set, we demonstrate that such deficiency is because current LLMs struggle in designing advanced algorithms and are barely aware of implementation optimization.", "keywords": [ "Large Language Model", "Code Generation", "Efficiency Evaluation" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "suyX1TOJJK", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Camille Simon Chane", "gender": "Female", "institution": "Ecole Nationale Sup\u00e9rieure de l'Electronique et de ses Applications", "country": "FR", "position": "Associate Professor" }, { "name": "Dongyue Lu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Gim Hee Lee", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" }, { "name": "Lingdong Kong", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Wei Ooi", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 56, "n_ref": 116, "n_ref_all": 129, "n_fig": 5, "n_tab": 10, "L_tab": 7117, "n_element_tab": 1071, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2375, "n_element_tab_1": 332, "formula_len_all": 738, "formula_len_all_1": 695, "len_all": 195911, "len_all_1": 69678, "len_abs": 494, "len_title": 103, "len_sents": 51076, "len_sents_1": 28810, "n_sents": 383, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1595, "title": "Event Camera Object Detection at Arbitrary Frequencies", "abs": "Event cameras offer unparalleled advantages for real-time perception in dynamic environments, thanks to their microsecond-level temporal resolution and asynchronous operation. However, existing event-based object detection methods are limited by fixed-frequency paradigms, which fail to fully exploit the high-temporal resolution and adaptability of event cameras. To address these limitations, we propose FlexEvent, a novel event camera object detection framework that enables detection at arbitrary frequencies. FlexEvent consists of two key components: FlexFuser, an adaptive event-frame fusion module that integrates high-frequency event data with rich semantic information from RGB frames, and FAL, a frequency-adaptive learning mechanism that generates frequency-adjusted labels to enhance model generalization across varying operational frequencies. This combination allows FlexEvent to detect objects with high accuracy in both fast-moving and static scenarios, while adapting to dynamic environments. Extensive experiments on large-scale event camera datasets demonstrate that our approach surpasses state-of-the-art methods, achieving significant improvements in both standard and high-frequency settings. Notably, FlexEvent maintains robust performance when scaling from 20 Hz to 90 Hz and delivers accurate detection up to 180 Hz, proving its effectiveness in extreme conditions. Our framework sets a new benchmark for event-based object detection and paves the way for more adaptable, real-time vision systems. The code will be made publicly available to facilitate future research.", "keywords": [ "Event Camera Object Detection", "Low-Latency Vision", "Label-Efficient Learning" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "suJ1z1UX2t", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Jaemyung Yu", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Junmo Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Jong Suk Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Sung Hyun Baek", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 30, "n_ref": 81, "n_ref_all": 94, "n_fig": 5, "n_tab": 5, "L_tab": 1598, "n_element_tab": 303, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1491, "n_element_tab_1": 299, "formula_len_all": 792, "formula_len_all_1": 757, "len_all": 124769, "len_all_1": 65406, "len_abs": 1728, "len_title": 89, "len_sents": 41310, "len_sents_1": 28577, "n_sents": 301, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 41, "L_abs": 1737, "title": "Supervised Dimension Contrastive Learning", "abs": "Self-supervised learning has emerged as an effective pre-training strategy for representation learning using large-scale unlabeled data. However, models pre-trained with self-supervised learning still require supervised fine-tuning to achieve optimal task-specific performance. Due to the lack of label utilization, it is difficult to accurately distinguish between positive and hard negative samples. Supervised contrastive learning methods address the limitation by leveraging labels, but they focus on global representations, leading to limited feature diversity and high cross-correlation between representation dimensions. To address these challenges, we propose Supervised Dimension Contrastive Learning, a novel approach that combines supervision with dimension-wise contrastive learning. Inspired by redundancy reduction techniques like Barlow Twins, this approach reduces cross-correlation between embedding dimensions while enhancing class discriminability. The aggregate function combines the embedding dimensions to generate predicted class variables, which are optimized to correlate with their corresponding class labels. Orthogonal regularization is applied to ensure the full utilization of all dimensions by enforcing full-rankness in the aggregate function. We evaluate our method on both in-domain supervised classification tasks and out-of-domain transfer learning tasks, demonstrating its superior performance compared to traditional supervised learning, supervised contrastive learning, and self-supervised learning methods. Our results show that the proposed method effectively reduces inter-dimensional correlation and enhances class discriminability, proving its generalizability across various downstream tasks.", "keywords": [ "Supervised Representation Learning", "Dimension Contrastive Learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "stolHkh6Nc", "primary_area": "applications to robotics, autonomy, planning", "n_author": 3, "author_info_list": [ { "name": "Patara Trirat", "gender": "Male", "institution": "DeepAuto.ai", "country": "KR", "position": "Researcher" }, { "name": "Sung Ju Hwang", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Wonyong Jeong", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "Ph.D. student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 41, "n_ref": 92, "n_ref_all": 125, "n_fig": 6, "n_tab": 10, "L_tab": 5867, "n_element_tab": 225, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1396, "n_element_tab_1": 148, "formula_len_all": 239, "formula_len_all_1": 326, "len_all": 285448, "len_all_1": 71582, "len_abs": 2300, "len_title": 114, "len_sents": 106292, "len_sents_1": 32911, "n_sents": 823, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1816, "title": "AutoML-Agent: A Multi-Agent LLM Framework for Full-Pipeline AutoML", "abs": "Automated machine learning (AutoML) accelerates AI development by automating tasks in the development pipeline, such as optimal model search and hyperparameter tuning. Existing AutoML systems often require technical expertise to set up complex tools, which is in general time-consuming and requires a large amount of human effort. Therefore, recent works have started exploiting large language models (LLM) to lessen such burden and increase the usability of AutoML frameworks via a natural language interface, allowing non-expert users to build their data-driven solutions. These methods, however, are usually designed only for a particular process in the AI development pipeline and do not efficiently use the inherent capacity of the LLMs. This paper proposes *AutoML-Agent*, a novel multi-agent framework tailored for full-pipeline AutoML, i.e., from data retrieval to model deployment. *AutoML-Agent* takes user's task descriptions, facilitates collaboration between specialized LLM agents, and delivers deployment-ready models. Unlike existing work, instead of devising a single plan, we introduce a retrieval-augmented planning strategy to enhance exploration to search for more optimal plans. We also decompose each plan into sub-tasks (e.g., data preprocessing and neural network design) each of which is solved by a specialized agent we build via prompting executing in parallel, making the search process more efficient. Moreover, we propose a multi-stage verification to verify executed results and guide the code generation LLM in implementing successful solutions. Extensive experiments on seven downstream tasks using fourteen datasets show that *AutoML-Agent* achieves a higher success rate in automating the full AutoML process, yielding systems with good performance throughout the diverse domains.", "keywords": [ "Automated Machine Learning", "Multi-Agent Framework", "Large Language Models" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 4, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "stcN89QGfL", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Hao Sun", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Hongsheng Liu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Qi Wang", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Ruizhi Chengze", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Wang Haoyun", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Yang Liu", "gender": "Female", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Yi Zhang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yuan Mi", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 6, "n_ref_uni": 45, "n_ref": 87, "n_ref_all": 136, "n_fig": 15, "n_tab": 9, "L_tab": 2673, "n_element_tab": 226, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 294, "n_element_tab_1": 51, "formula_len_all": 1007, "formula_len_all_1": 418, "len_all": 172279, "len_all_1": 62827, "len_abs": 1596, "len_title": 127, "len_sents": 55259, "len_sents_1": 28835, "n_sents": 469, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1607, "title": "PDE-constrained Learning with Multi-time-stepping for Accelerated Fluid Simulation", "abs": "Solving partial differential equations (PDEs) by numerical methods meet computational cost challenge for getting the accurate solution since fine grids and small time steps are required. Machine learning can accelerate this process, but struggle with weak generalizability, interpretability, and data dependency, as well as suffer in long-term prediction. To this end, we propose a PDE-embedded network with multiscale time stepping (MultiPDENet), which fuses the scheme of numerical methods and machine learning, for accelerated simulation of fluid flows. In particular, we design a convolutional filter based on the structure of finite difference stencils with a small number of parameters to optimize, which estimates the equivalent form of spatial derivative on a coarse grid to minimize the equation's residual. A physics block with a 4th-order Runge-Kutta integrator at the fine time scale is established that embeds the structure of PDEs to guide the prediction. To alleviate the curse of temporal error accumulation in long-term prediction, we introduce a multiscale time integration approach, where a neural network is used to correct the prediction error at a coarse time scale. Experiments across various PDE systems, including the Navier-Stokes equations, demonstrate that MultiPDENet can accurately predict long-term spatiotemporal dynamics, even given small and incomplete training data, e.g., spatiotemporally down-sampled datasets. MultiPDENet achieves the state-of-the-art performance compared with other baseline models, with over 5$\\times$ speedup compared to classical numerical methods.", "keywords": [ "physics-informed learning", "multiscale time stepping", "spatiotemporal dynamics prediction" ], "rating_list": [ 6, 8, 3, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 2, 4, 2, 3, 2 ], "contribution_list": [ 3, 3, 1, 3, 3 ], "confidence_list": [ 3, 4, 4, 3, 5 ] }, { "paper_id": "stUKwWBuBm", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Eric Mazumdar", "gender": "Male", "institution": "Deparment of Computing + Mathematical Sciences, California Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Kishan Panaganti", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Laixi Shi", "gender": "Female", "institution": "California Institute of Technology", "country": "", "position": "Postdoc" } ], "n_formula": 119, "n_formula_1": 34, "n_ref_uni": 65, "n_ref": 149, "n_ref_all": 183, "n_fig": 3, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12116, "formula_len_all_1": 3043, "len_all": 232087, "len_all_1": 69156, "len_abs": 1266, "len_title": 122, "len_sents": 73722, "len_sents_1": 33220, "n_sents": 560, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 73, "L_abs": 1276, "title": "Tractable Multi-Agent Reinforcement Learning through Behavioral Economics", "abs": "A significant roadblock to the development of principled multi-agent reinforcement learning is the fact that desired solution concepts like Nash equilibria may be intractable to compute. To overcome this obstacle, we take inspiration from behavioral economics and show that---by imbuing agents with important features of human decision-making like risk aversion and bounded rationality---a class of risk-averse quantal response equilibria (RQE) become tractable to compute in all $n$-player matrix and finite-horizon Markov games. In particular, we show that they emerge as the endpoint of no-regret learning in suitably adjusted versions of the games. Crucially, the class of computationally tractable RQE is independent of the underlying game structure and only depends on agents' degree of risk-aversion and bounded rationality. To validate the richness of this class of solution concepts we show that it captures peoples' patterns of play in a number of 2-player matrix games previously studied in experimental economics. Furthermore, we give a first analysis of the sample complexity of computing these equilibria in finite-horizon Markov games when one has access to a generative model and validate our findings on a simple multi-agent reinforcement learning benchmark.", "keywords": [ "behavioral economics", "risk-aversion", "multi-agent reinforcement learning", "quantal response", "bounded rationality" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 4, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "stK7iOPH9Q", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Bingbing Liu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "R&D Manager" }, { "name": "Haodong LI", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Intern" }, { "name": "Hongbo Zhang", "gender": "Male", "institution": "Huawei", "country": "CN", "position": "Principal Researcher" }, { "name": "Jing He", "gender": "Female", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Kaiqiang Zhou", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Leheng Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Wei Yin", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Ying-Cong Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Yixun Liang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 50, "n_ref": 99, "n_ref_all": 117, "n_fig": 10, "n_tab": 3, "L_tab": 2376, "n_element_tab": 249, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 1172, "n_element_tab_1": 15, "formula_len_all": 443, "formula_len_all_1": 330, "len_all": 155269, "len_all_1": 67219, "len_abs": 0, "len_title": 130, "len_sents": 35707, "len_sents_1": 30403, "n_sents": 284, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1710, "title": "Lotus: Diffusion-based Visual Foundation Model for High-quality Dense Prediction", "abs": "Leveraging the visual priors of pre-trained text-to-image diffusion models offers a promising solution to enhance zero-shot generalization in dense prediction tasks. However, existing methods often uncritically use the original diffusion formulation, which may not be optimal due to the fundamental differences between dense prediction and image generation. In this paper, we provide a systemic analysis of the diffusion formulation for the dense prediction, focusing on both quality and efficiency. And we find that the original parameterization type for image generation, which learns to predict noise, is harmful for dense prediction; the multi-step noising/denoising diffusion process is also unnecessary and challenging to optimize. Based on these insights, we introduce $\\textbf{Lotus}$, a diffusion-based visual foundation model with a simple yet effective adaptation protocol for dense prediction. Specifically, Lotus is trained to directly predict annotations instead of noise, thereby avoiding harmful variance. We also reformulate the diffusion process into a single-step procedure, simplifying optimization and significantly boosting inference speed. Additionally, we introduce a novel tuning strategy called detail preserver, which achieves more accurate and fine-grained predictions. Without scaling up the training data or model capacity, Lotus achieves SoTA performance in zero-shot depth and normal estimation across various datasets. It also enhances efficiency, being significantly faster than most existing diffusion-based methods. Lotus' superior quality and efficiency also enable a wide range of practical applications, such as joint estimation, single/multi-view 3D reconstruction, etc.", "keywords": [ "Diffusion Models; Dense Prediction; Monocular Depth Estimation; Monocular Normal Estimation" ], "rating_list": [ 6, 8, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 4, 2 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 5, 4 ] }, { "paper_id": "st7XqFgbAH", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Paloma Sodhi", "gender": "Female", "institution": "ASAPP", "country": "", "position": "Research Scientist" }, { "name": "Sanjiban Choudhury", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Assistant Professor" } ], "n_formula": 30, "n_formula_1": 9, "n_ref_uni": 44, "n_ref": 87, "n_ref_all": 109, "n_fig": 7, "n_tab": 9, "L_tab": 8377, "n_element_tab": 354, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2981, "n_element_tab_1": 311, "formula_len_all": 3513, "formula_len_all_1": 619, "len_all": 237014, "len_all_1": 71543, "len_abs": 1274, "len_title": 123, "len_sents": 70159, "len_sents_1": 29156, "n_sents": 573, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1241, "title": "Better than Your Teacher: LLM Agents that learn from Privileged AI Feedback", "abs": "While large language models (LLMs) show impressive decision-making abilities, current methods lack a mechanism for automatic self-improvement from errors during task execution. We propose LEAP, an iterative fine-tuning framework that continually improves LLM agents using feedback from AI expert teachers. Our key insight is to equip the expert teachers with a privileged state -- information available during training but hidden at test time. This allows even weak experts to provide precise guidance, significantly improving the student agent's performance without access to privileged information at test time. We evaluate LEAP on diverse decision-making benchmarks, including text-based games, web navigation, and interactive coding. Our experiments show that LEAP (1) outperforms state-of-the-art baselines (2) enables weak student models (e.g., Llama3-8B) to exceed the performance of strong teacher models (GPT4-o), and (3) allows weak models to self-improve using privileged versions of themselves. We also provide a theoretical analysis showing that LEAP's success hinges on balancing privileged information with the student\u2019s realizability, which we empirically validate. Our code is provided as part of the supplementary material.", "keywords": [ "learning from AI feedback", "imitation learning", "privileged information" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "st77ShxP1K", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Guikun Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Wenguan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zhiyuan Weng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 5, "n_ref_uni": 111, "n_ref": 145, "n_ref_all": 202, "n_fig": 17, "n_tab": 13, "L_tab": 5091, "n_element_tab": 283, "n_fig_1": 11, "n_tab_1": 5, "L_tab_1": 1032, "n_element_tab_1": 45, "formula_len_all": 1132, "formula_len_all_1": 288, "len_all": 310744, "len_all_1": 74692, "len_abs": 1644, "len_title": 118, "len_sents": 83542, "len_sents_1": 33134, "n_sents": 681, "n_sents_1": 272, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 70, "L_abs": 1650, "title": "Do as We Do, Not as You Think: the Conformity of Large Language Models", "abs": "Recent advancements in large language models (LLMs) revolutionize the field of intelligent agents, enabling collaborative multi-agent systems capable of tackling complex problems across various domains. However, the potential of conformity within these systems, analogous to phenomena like conformity bias and group-think in human group dynamics, remains largely unexplored, raising concerns about their collective problem-solving capabilities and possible ethical implications. This paper presents a comprehensive study on conformity in LLM-driven multi-agent systems, focusing on three aspects: the existence of conformity, the factors influencing conformity, and potential mitigation strategies. In particular, we introduce BENCHFORM, a new conformity-oriented benchmark, featuring reasoning-intensive tasks and five distinct interaction protocols designed to probe LLMs\u2019 behavior in collaborative scenarios. Several representative LLMs are evaluated on BENCHFORM, using metrics such as conformity rate and independence rate to quantify conformity\u2019s impact. Our analysis delves into factors influencing conformity, including interaction time and majority size, and examines how the subject agent rationalize its conforming behavior. Furthermore, we explore two strategies to mitigate conformity effects, i.e., developing enhanced persona and implementing a reflection mechanism. Several interesting findings regarding LLMs\u2019 conformity are derived from empirical results and case studies. We hope that these insights can pave the way for more robust and ethically-aligned collaborative AI systems. Our benchmark and code will be publicly available.", "keywords": [ "Large Language Models", "Conformity", "Multi-agent System" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "sshYEYQ82L", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Bingbing Zhao", "gender": "Female", "institution": "Information Engineering University", "country": "CN", "position": "MS student" }, { "name": "Hui Wei", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Jia liu", "gender": "Male", "institution": "Information Engineering University", "country": "CN", "position": "PhD student" }, { "name": "Jinlong Fei", "gender": "Male", "institution": "State Key Laboratory of Mathematical Engineering and Advanced Computing", "country": "CN", "position": "Full Professor" }, { "name": "Zheng Wang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Yang guoliang", "gender": "Male", "institution": "Information Engineering University", "country": "CN", "position": "PhD student" }, { "name": "song yan", "gender": "Male", "institution": "Information Engineering University", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 27, "n_ref": 62, "n_ref_all": 82, "n_fig": 10, "n_tab": 5, "L_tab": 977, "n_element_tab": 159, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1423, "n_element_tab_1": 246, "formula_len_all": 467, "formula_len_all_1": 465, "len_all": 143309, "len_all_1": 71016, "len_abs": 1535, "len_title": 142, "len_sents": 46671, "len_sents_1": 31724, "n_sents": 371, "n_sents_1": 249, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1343, "title": "Unfiltered and Unseen: Universal Multimodal Jailbreak Attacks on Text-to-Image Model Defenses", "abs": "Text-to-Image (T2I) models have revolutionized the synthesis of visual content from textual descriptions. However, their potential misuse for generating Not-Safe-For-Work (NSFW) content presents significant risks. While developers have implemented prompt filters and safety checkers, these defense mechanisms have proven inadequate against determined adversaries. In this paper, we introduce U3-Attack, a novel multimodal jailbreak attack against T2I models that effectively circumvents existing safeguards to generate NSFW images. \nTo achieve a universal attack, U3-Attack constructs a context-independent paraphrase candidate set for each sensitive word in the text modality. This approach enables practical attacks against prompt filters with minimal perturbation. In the image modality, we propose a two-stage adversarial patch generation strategy that does not require access to the T2I model's internal architecture or parameters. This design makes our attack applicable to both open-source models and online T2I platforms.\nExtensive experiments demonstrate the effectiveness of our method across various T2I models, including Stable Diffusion, Leonardo.Ai, and Runway. Our work exposes critical vulnerabilities in current T2I model defenses and underscores the urgent need for more robust safety measures in this rapidly evolving field.", "keywords": [ "Diffusion Model", "Not-Safe-for- Work (NSFW)", "Adversarial Attack", "Jailbreak Attack" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "ssRdQimeUI", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Giseung Park", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Researcher" }, { "name": "Youngchul Sung", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 31, "n_formula_1": 31, "n_ref_uni": 38, "n_ref": 104, "n_ref_all": 134, "n_fig": 5, "n_tab": 4, "L_tab": 240, "n_element_tab": 20, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 144, "n_element_tab_1": 13, "formula_len_all": 4238, "formula_len_all_1": 3352, "len_all": 164857, "len_all_1": 71888, "len_abs": 940, "len_title": 126, "len_sents": 51517, "len_sents_1": 31365, "n_sents": 375, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 942, "title": "Reward Dimension Reduction for Scalable Multi-Objective Reinforcement Learning", "abs": "In this paper, we introduce a simple yet effective reward dimension reduction method to tackle the scalability challenges of multi-objective reinforcement learning algorithms. While most existing approaches focus on optimizing two to four objectives, their abilities to scale to environments with more objectives remain uncertain. Our method uses a dimension reduction approach to enhance learning efficiency and policy performance in multi-objective settings. While most traditional dimension reduction methods are designed for static datasets, our approach is tailored for online learning and preserves Pareto-optimality after transformation. We propose a new training and evaluation framework for reward dimension reduction in multi-objective reinforcement learning and demonstrate the superiority of our method in environments including one with sixteen objectives, significantly outperforming existing online dimension reduction methods.", "keywords": [ "Multi-Objective Reinforcement Learning", "Reinforcement Learning" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 2, 2, 3, 4 ] }, { "paper_id": "sruGNQHd7t", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Aidong Adam Ding", "gender": "unknown", "institution": "Northeastern University", "country": "", "position": "Associate Professor" }, { "name": "Xiang Zhang", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Yunsi Fei", "gender": "Female", "institution": "Northeastern University", "country": "", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 24, "n_ref": 41, "n_ref_all": 52, "n_fig": 2, "n_tab": 5, "L_tab": 4816, "n_element_tab": 199, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2879, "n_element_tab_1": 113, "formula_len_all": 642, "formula_len_all_1": 642, "len_all": 105454, "len_all_1": 68367, "len_abs": 930, "len_title": 110, "len_sents": 31926, "len_sents_1": 30269, "n_sents": 236, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 933, "title": "Privacy-Preserving of Deep Learning Queries by Domain Shifting", "abs": "In the era of cloud-based deep learning (DL) services, data privacy has become a critical concern, prompting some organizations to restrict the use of online AI services. This work addresses this issue by introducing a privacy-preserving method for DL model queries through domain shifting in the input space. We develop an encoder that strategically transforms inputs into a different domain within the same space, ensuring that the original inputs remain private by presenting only the obfuscated versions to the DL model. A decoder then recovers the correct output from the model's predictions. Our method keeps the authentic input and output data secure on the local system, preventing unauthorized access by third parties who only encounter the obfuscated data. Comprehensive evaluations across various oracle models and datasets demonstrate that our approach preserves privacy with minimal impact on classification performance.", "keywords": [ "Privacy-Preserving", "Domain Shifting", "Input Obfuscation" ], "rating_list": [ 1, 5, 3 ], "soundness_list": [ 1, 3, 2 ], "presentation_list": [ 3, 3, 1 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "srg4XYZA1W", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Junlong Du", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Ke Yan", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Lu Zhang", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Shouhong Ding", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "researcher" }, { "name": "Xiaoqiang Li", "gender": "Male", "institution": "Shanghai University", "country": "CN", "position": "Full Professor" }, { "name": "Yuanchen Wu", "gender": "unknown", "institution": "Shanghai University", "country": "CN", "position": "MS student" }, { "name": "Yunsheng Wu", "gender": "Male", "institution": "Tencent YouTu Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Yao Hang", "gender": "Male", "institution": "Shanghai University", "country": "CN", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 2, "n_ref_uni": 36, "n_ref": 122, "n_ref_all": 152, "n_fig": 13, "n_tab": 7, "L_tab": 1616, "n_element_tab": 248, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1349, "n_element_tab_1": 49, "formula_len_all": 510, "formula_len_all_1": 241, "len_all": 189113, "len_all_1": 70967, "len_abs": 1419, "len_title": 133, "len_sents": 55487, "len_sents_1": 32279, "n_sents": 451, "n_sents_1": 240, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1425, "title": "Hallucination in LVLMs: Fictitious Presupposition Questions, Benchmark, and Solution", "abs": "Large Vision-Language Models (LVLMs) have achieved impressive performance across various vision-language tasks. However, hallucinations, i.e., generating counterfactual responses, remain a significant challenge. Although recent models have mitigated hallucinations in tasks such as object existence and image description, they primarily address hallucinations in response generation while overlooking the task question itself. This paper highlights the vulnerability of LVLMs in solving fictitious presupposition questions (FPQs), where the models are prone to accept the presuppositions of non-existent objects and produce severe hallucinatory responses. To this end, we first introduce a novel benchmark, VFP-Bench, to evaluate LVLMs' capability to discriminate fictitious presuppositions and generate factual responses. Moreover, we introduce Antidote, a universal, synthetic data-driven self-correction solution for alleviating hallucination in FPQs and conventional tasks. It leverages synthetic data to incorporate factual priors into questions/queries to achieve self-correction, decoupling hallucination alleviation into a preference optimization problem. Applied to the LLaVA series, it enhances performance on VFP-Bench by over 50%, POPE by 1.8\u20133.3%, and CHAIR \\& SHR by 30\u201350%, without relying on external supervision from stronger LVLMs or human feedback and introducing noticeable catastrophic forgetting issues.", "keywords": [ "Large Vision Language Model", "Hallucination", "Synthetic Data", "Preference Alignment" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "srOVvTzgPo", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Anders S\u00f8gaard", "gender": "Male", "institution": "Copenhagen University", "country": "", "position": "Full Professor" }, { "name": "Antonia Karamolegkou", "gender": "Female", "institution": "University of Copenhagen", "country": "", "position": "PhD student" }, { "name": "Ruchira Dhar", "gender": "Female", "institution": "University of Copenhagen", "country": "DK", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 15, "n_ref_uni": 48, "n_ref": 70, "n_ref_all": 73, "n_fig": 1, "n_tab": 2, "L_tab": 853, "n_element_tab": 89, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 934, "n_element_tab_1": 76, "formula_len_all": 1456, "formula_len_all_1": 844, "len_all": 160226, "len_all_1": 59278, "len_abs": 968, "len_title": 131, "len_sents": 49388, "len_sents_1": 29414, "n_sents": 323, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 973, "title": "Toward a Sheaf-Theoretic Understanding of Compositionality in Large Language Models", "abs": "Compositionality has long been considered a fundamental aspect of human cognition - enabling the learning, manipulation, and generation of natural language. Understanding how this concept applies to Large Language Models (LLMs) and how it can be effectively evaluated remains a key challenge. In this work, we explore the potential of formalizing cognitive notions from theory, such as compositionality, to develop more nuanced evaluation frameworks for LLMs. Using a sheaf-theoretic approach, we define compositionality through four distinct conditions that capture its multifaceted nature. This formalization offers a structured perspective on evaluating LLMs, moving beyond surface-level assessments to uncover deeper insights into their behavior. Our findings suggest that theoretical frameworks like this one can play a crucial role in advancing the understanding and evaluation of LLMs, providing a foundation for more comprehensive and precise performance analyses.", "keywords": [ "cognition", "compositionality", "sheaf-theory", "language model evaluation" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 3, 1 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "sr0My6yDNu", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Gang Yan", "gender": "Male", "institution": "Tongji University", "country": "", "position": "Full Professor" }, { "name": "Jack Murdoch Moore", "gender": "Male", "institution": "Tongji University", "country": "", "position": "Associate Professor" }, { "name": "Wenjia Wei", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Xia Zhu", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Xiaolei Ru", "gender": "Male", "institution": "Tongji University", "country": "", "position": "Postdoc" }, { "name": "Xin-Ya Zhang", "gender": "Female", "institution": "Tongji University", "country": "CN", "position": "PhD student" }, { "name": "Cao Xiaowei", "gender": "unknown", "institution": "Tongji University", "country": "", "position": "MS student" } ], "n_formula": 23, "n_formula_1": 22, "n_ref_uni": 15, "n_ref": 22, "n_ref_all": 44, "n_fig": 7, "n_tab": 5, "L_tab": 1698, "n_element_tab": 274, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 824, "n_element_tab_1": 33, "formula_len_all": 1861, "formula_len_all_1": 1720, "len_all": 116317, "len_all_1": 59372, "len_abs": 999, "len_title": 105, "len_sents": 36913, "len_sents_1": 24096, "n_sents": 262, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1002, "title": "Maintaining Adversarial Robustness in Continuous Learning", "abs": "Adversarial robustness is essential for security and reliability of machine learning systems. However, adversarial robustness enhanced by defense algorithms is easily erased as the neural network's weights update to learn new tasks. To address this vulnerability, it is essential to improve the capability of neural networks in terms of robust continual learning. Specially, we propose a novel gradient projection technique that effectively stabilizes sample gradients from previous data by orthogonally projecting back-propagation gradients onto a crucial subspace before using them for weight updates. This technique can maintaining robustness by collaborating with a class of defense algorithms through sample gradient smoothing. The experimental results on four benchmarks including Split-CIFAR100 and Split-miniImageNet, demonstrate that the superiority of the proposed approach in mitigating rapidly degradation of robustness during continual learning even when facing strong adversarial attacks.", "keywords": [ "gradient projection technique; sample gradient smoothing; robust continual learning; continual learning" ], "rating_list": [ 3, 3, 1, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "sqsGBW8zQx", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Nanxuan Zhao", "gender": "Female", "institution": "Adobe Research", "country": "US", "position": "Researcher" }, { "name": "Samyadeep Basu", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Soheil Feizi", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "Assistant Professor" }, { "name": "Varun Manjunatha", "gender": "Male", "institution": "Adobe Systems", "country": "", "position": "Research Scientist" }, { "name": "Vlad I. Morariu", "gender": "Male", "institution": "Adobe", "country": "US", "position": "Senior Research Scientist" }, { "name": "Zichao Wang", "gender": "Not Specified", "institution": "Adobe Research", "country": "US", "position": "Research Scientist" }, { "name": "Ryan Rossi", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Senior Research Scientist" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 238, "n_ref_all": 252, "n_fig": 8, "n_tab": 4, "L_tab": 673, "n_element_tab": 12, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 145, "formula_len_all_1": 0, "len_all": 186826, "len_all_1": 68262, "len_abs": 1790, "len_title": 138, "len_sents": 61610, "len_sents_1": 34974, "n_sents": 437, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1776, "title": "Understanding and Enhancing Context-Augmented Language Models Through Mechanistic Circuits", "abs": "Recent studies have extracted circuits from the computational graphs of language models for simple language tasks such as entity tracking or indirect object identification. In our paper, we scale up circuit extraction to a real-world language modeling task: context-augmented language modeling for question-answering (QA) tasks and understand the potential benefits of circuits towards downstream applications such as data attribution. We extract circuits as a function of internal model components (e.g., attention heads, attention layers, MLPs) using causal mediation analysis techniques. Leveraging the extracted circuits, we first understand the interplay between the language model's usage of parametric memory and retrieved context towards a better mechanistic understanding of context-augmented language models. We then identify a small set of attention heads in our circuit which performs reliable data attribution by default, thereby obtaining attribution for free in just the model's forward pass! Using this insight, we then introduce AttnAttrib, a fast data attribution algorithm. Through a range of empirical experiments across different extractive QA benchmarks, we show that performing data attribution with AttnAttrib obtains state-of-the-art attribution results across different language models. Finally, we show the possibility to steer the language model towards answering from the context, instead of the parametric memory by (i) using the attribution from our extracted attention head as an additional signal during the forward pass and (ii) scaling the output of a small set of attention heads. Beyond mechanistic understanding, our paper provides tangible applications of mechanistic circuits in the form of reliable data attribution and model steering.", "keywords": [ "circuits", "mechanistic interpretability", "language models", "extractive QA" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "sqciWyTm70", "primary_area": "datasets and benchmarks", "n_author": 1, "author_info_list": [ { "name": "Yi Cui", "gender": "unknown", "institution": "ONEKQ Lab", "country": "US", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 0, "n_ref_uni": 34, "n_ref": 47, "n_ref_all": 74, "n_fig": 12, "n_tab": 17, "L_tab": 10194, "n_element_tab": 504, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 4261, "n_element_tab_1": 448, "formula_len_all": 538, "formula_len_all_1": 0, "len_all": 179587, "len_all_1": 50039, "len_abs": 1608, "len_title": 131, "len_sents": 38286, "len_sents_1": 15623, "n_sents": 386, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 454, "title": "Tests as Instructions: A Test-Driven-Development Benchmark for LLM Code Generation", "abs": "This paper focuses on test-driven development (TDD) tasks, where test cases act as both instruction and verification for LLM code generation. We build a TDD benchmark to evaluate frontier models, where reasoning models of OpenAI achieve SOTA. We identify instruction following and in-context learning as the critical abilities for all models to succeed at TDD tasks. We further reveal their vulnerabilities to long instructions as an area of improvement.", "keywords": [ "test", "llm", "coding", "benchnark", "instruction" ], "rating_list": [ 3, 1, 6, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 1, 1, 3, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "sq5LLWk5SN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Shuang Liu", "gender": "Male", "institution": "Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Xiao-Shan Gao", "gender": "Male", "institution": "Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Yibo Miao", "gender": "Male", "institution": "Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yifan Zhu", "gender": "unknown", "institution": "Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yihan Wang", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Intern" } ], "n_formula": 86, "n_formula_1": 25, "n_ref_uni": 38, "n_ref": 66, "n_ref_all": 85, "n_fig": 2, "n_tab": 9, "L_tab": 1039, "n_element_tab": 96, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 134, "n_element_tab_1": 19, "formula_len_all": 9301, "formula_len_all_1": 1876, "len_all": 161844, "len_all_1": 59135, "len_abs": 1088, "len_title": 129, "len_sents": 48400, "len_sents_1": 26519, "n_sents": 434, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1227, "title": "Provable Robust Overfitting Mitigation in Wasserstein Distributionally Robust Optimization", "abs": "Wasserstein distributionally robust optimization (WDRO) optimizes against worst-case distributional shifts within a specified uncertainty set, leading to enhanced generalization on unseen adversarial examples, compared to standard adversarial training which focuses on pointwise adversarial perturbations. However, WDRO still suffers fundamentally from the robust overfitting problem, as it does not consider statistical error. We address this gap by proposing a novel robust optimization framework under a new uncertainty set for adversarial noise via Wasserstein distance and statistical error via Kullback-Leibler divergence, called the Statistically Robust WDRO. We establish a robust generalization bound for the new optimization framework, implying that out-of-distribution adversarial performance is at least as good as the statistically robust training loss with high probability. Furthermore, we derive conditions under which Stackelberg and Nash equilibria exist between the learner and the adversary, giving an optimal robust model in certain sense.Finally, through extensive experiments, we demonstrate that our method significantly mitigates robust overfitting and enhances robustness within the framework of WDRO.", "keywords": [ "Adversarial examples; robust overfitting; WDRO" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "spwklWgmWJ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Rahul Nittala", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "PhD student" }, { "name": "Rebekka Burkholz", "gender": "Female", "institution": "Helmholtz Center CISPA for Information Security", "country": "DE", "position": "Associate Professor" }, { "name": "Tom Jacobs", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 18, "n_ref_uni": 44, "n_ref": 73, "n_ref_all": 89, "n_fig": 5, "n_tab": 4, "L_tab": 1015, "n_element_tab": 95, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 422, "n_element_tab_1": 23, "formula_len_all": 1550, "formula_len_all_1": 1520, "len_all": 159424, "len_all_1": 63343, "len_abs": 825, "len_title": 123, "len_sents": 47642, "len_sents_1": 29959, "n_sents": 360, "n_sents_1": 236, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 829, "title": "Memory Savings by Sharing One Source: Insights into Subsetsum Approximation", "abs": "Large deep neural networks, often fine-tuned from foundation models, dominate modern machine learning, but their high memory requirements limit deployment on resource-constrained devices. Strong lottery tickets (SLTs) offer a promising solution by significantly reducing memory usage, as they are fully characterized by a seed for generating a random source network and a binary mask. Notably, multiple models can share the same source network without increasing its width requirement. As we show, this source sharing can lead to memory savings when experts share specific sparsity patterns. Based on novel insights into optimized subset sum approximations, we also show how the masks can be adjusted to further reduce memory overhead. To validate these theoretical findings, we provide explicit SLT constructions in experiments.", "keywords": [ "Strong lottery tickets", "Subsetsum approximation", "Mixture of Experts", "Ensembles", "Memory savings", "Discrete optimization" ], "rating_list": [ 5, 8, 3, 3 ], "soundness_list": [ 3, 3, 1, 1 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 2, 4, 1, 3 ], "confidence_list": [ 3, 3, 5, 2 ] }, { "paper_id": "sprjE7BTZR", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Alexander Rakhlin", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Runlong Zhou", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Simon Shaolei Du", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "Xiyu Zhai", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Postdoc" }, { "name": "liao zhang", "gender": "unknown", "institution": "Universit\u00e4t Innsbruck", "country": "", "position": "PhD student" } ], "n_formula": 70, "n_formula_1": 8, "n_ref_uni": 64, "n_ref": 101, "n_ref_all": 110, "n_fig": 3, "n_tab": 5, "L_tab": 815, "n_element_tab": 40, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5682, "formula_len_all_1": 593, "len_all": 312280, "len_all_1": 65707, "len_abs": 2480, "len_title": 95, "len_sents": 109620, "len_sents_1": 32560, "n_sents": 956, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1767, "title": "Transformers are Efficient Compilers, Provably", "abs": "Transformer-based large language models (LLMs) have demonstrated surprisingly robust performance across a wide range of language-related tasks, including programming language understanding and generation. In this paper, we take the first steps towards a formal investigation of using transformers as compilers from an expressive power perspective. To this end, we introduce a representative programming language, **mini-husky**, which encapsulates key features of modern C-like languages. We show that if the input code sequence has a bounded depth in both the Abstract Syntax Tree (AST) and type inference (reasonable assumptions based on the clean code principle), then the number of parameters required by transformers depends only on the logarithm of the input sequence length to handle compilation tasks, such as AST construction, symbol resolution, and type analysis. A significant technical challenge stems from the fact that transformers operate at a low level, where each layer processes the input sequence as raw vectors without explicitly associating them with predefined structure or meaning. In contrast, high-level compiler tasks necessitate managing intricate relationships and structured program information. Our primary technical contribution is the development of a domain-specific language, **Cybertron**, which generates formal proofs of the transformer\u2019s expressive power, scaling to address compiler tasks. We further establish that recurrent neural networks (RNNs) require at least a linear number of parameters relative to the input sequence, leading to an exponential separation between transformers and RNNs. Finally, we empirically validate our theoretical results by comparing transformers and RNNs on compiler tasks within **mini-husky**.", "keywords": [ "Transformers", "Expressive Power", "Programming Language", "Attention Mechanism", "Compiler" ], "rating_list": [ 5, 6, 1, 3 ], "soundness_list": [ 2, 2, 1, 1 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 4, 3, 1, 1 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "spDUv05cEq", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Caleb Dahlke", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Postdoc" }, { "name": "Jason L Pacheco", "gender": "Male", "institution": "University of Arizona", "country": "US", "position": "Assistant Professor" } ], "n_formula": 79, "n_formula_1": 34, "n_ref_uni": 31, "n_ref": 86, "n_ref_all": 109, "n_fig": 4, "n_tab": 3, "L_tab": 376, "n_element_tab": 42, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 104, "n_element_tab_1": 12, "formula_len_all": 10728, "formula_len_all_1": 2854, "len_all": 159958, "len_all_1": 72379, "len_abs": 1339, "len_title": 123, "len_sents": 51482, "len_sents_1": 30378, "n_sents": 456, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1344, "title": "Flow-based Variational Mutual Information: Fast and Flexible Approximations", "abs": "Mutual Information (MI) is a fundamental measure of dependence between random variables, but its practical application is limited because it is difficult to calculate in many circumstances. Variational methods offer one approach by introducing an approximate distribution to create various bounds on MI, which in turn is an easier optimization problem to solve. In practice, the variational distribution chosen is often a Gaussian, which is convenient but lacks flexibility in modeling complicated distributions. In this paper, we introduce new classes of variational estimators based on Normalizing Flows that extend the previous Gaussian-based variational estimators. Our new estimators maintain many of the same theoretical guarantees while simultaneously enhancing the expressivity of the variational distribution. We experimentally verify that our new methods are effective on large MI problems where discriminative-based estimators, such as MINE and InfoNCE, are fundamentally limited. Furthermore, we compare against a diverse set of benchmarking tests to show that the flow-based estimators often perform as well, if not better, than the discriminative-based counterparts. Finally, we demonstrate how these estimators can be effectively utilized in the Bayesian Optimal Experimental Design setting for online sequential decision making.", "keywords": [ "Mutual Information", "Variational Methods", "Normalizing Flows", "Bayesian Optimal Experimental Design" ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 4, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "sp9irsV1yq", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Amit Dhurandhar", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Dennis Wei", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Research Staff Member" }, { "name": "Karthikeyan Natesan Ramamurthy", "gender": "unknown", "institution": "International Business Machines", "country": "US", "position": "Research Staff Member" }, { "name": "Tian Gao", "gender": "unknown", "institution": "International Business Machines", "country": "US", "position": "Reseach Staff Member" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 47, "n_ref": 95, "n_ref_all": 141, "n_fig": 21, "n_tab": 7, "L_tab": 2214, "n_element_tab": 285, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 138, "n_element_tab_1": 11, "formula_len_all": 1021, "formula_len_all_1": 395, "len_all": 183899, "len_all_1": 73682, "len_abs": 1391, "len_title": 132, "len_sents": 58284, "len_sents_1": 36224, "n_sents": 434, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1397, "title": "Identifying Sub-networks in Neural Networks via Functionally Similar Representations", "abs": "Mechanistic interpretability aims to provide human-understandable insights into the inner workings of neural network models by examining their internals. Existing approaches typically require significant manual effort and prior knowledge, with strategies tailored to specific tasks. In this work, we take a step toward automating the understanding of the network by investigating the existence of distinct sub-networks. Specifically, we explore a novel automated and task-agnostic approach based on the notion of functionally similar representations within neural networks, reducing the need for human intervention. \nOur method identifies similar and dissimilar layers in the network, revealing potential sub-components. We achieve this by proposing, for the first time to our knowledge, the use of Gromov-Wasserstein distance, which overcomes challenges posed by varying distributions and dimensionalities across intermediate representations\u2014issues that complicate direct layer-to-layer comparisons.\nThrough experiments on algebraic, language, and vision tasks, we observe the emergence of sub-groups within neural network layers corresponding to functional abstractions. Additionally, we find that different training strategies influence the positioning of these sub-groups. Our approach offers meaningful insights into the behavior of neural networks with minimal human and computational cost.", "keywords": [ "mechanistic interpretability", "subnetworks" ], "rating_list": [ 3, 3, 6, 5, 6 ], "soundness_list": [ 2, 1, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 2, 1, 2, 1, 3 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "soaOqFTaHJ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Lexiang Hu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yisen Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhouchen Lin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Professor" } ], "n_formula": 62, "n_formula_1": 43, "n_ref_uni": 34, "n_ref": 90, "n_ref_all": 131, "n_fig": 3, "n_tab": 1, "L_tab": 669, "n_element_tab": 54, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 845, "n_element_tab_1": 10, "formula_len_all": 4345, "formula_len_all_1": 2962, "len_all": 127178, "len_all_1": 63232, "len_abs": 1528, "len_title": 44, "len_sents": 35402, "len_sents_1": 25475, "n_sents": 304, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 1539, "title": "EKAN: Equivariant Kolmogorov-Arnold Networks", "abs": "Kolmogorov-Arnold Networks (KANs) have seen great success in scientific domains thanks to spline activation functions, becoming an alternative to Multi-Layer Perceptrons (MLPs). However, spline functions may not respect symmetry in tasks, which is crucial prior knowledge in machine learning. Previously, equivariant networks embed symmetry into their architectures, achieving better performance in specific applications. Among these, Equivariant Multi-Layer Perceptrons (EMLP) introduce arbitrary matrix group equivariance into MLPs, providing a general framework for constructing equivariant networks layer by layer. In this paper, we propose Equivariant Kolmogorov-Arnold Networks (EKAN), a method for incorporating matrix group equivariance into KANs, aiming to broaden their applicability to more fields. First, we construct gated spline basis functions, which form the EKAN layer together with equivariant linear weights. We then define a lift layer to align the input space of EKAN with the feature space of the dataset, thereby building the entire EKAN architecture. Compared with baseline models, EKAN achieves higher accuracy with smaller datasets or fewer parameters on symmetry-related tasks, such as particle scattering and the three-body problem, often reducing test MSE by several orders of magnitude. Even in non-symbolic formula scenarios, such as top quark tagging with three jet constituents, EKAN achieves comparable results with EMLP using only $26\\\\%$ of the parameters, while KANs do not outperform MLPs as expected.", "keywords": [ "Equivariant networks", "Kolmogorov-Arnold networks" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "snocoXIQXz", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Ashish Rao", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Atri Rudra", "gender": "Male", "institution": "State University of New York, Buffalo", "country": "", "position": "Professor" }, { "name": "Christopher Re", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Jerry Weihong Liu", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Jessica Grogan", "gender": "Female", "institution": "State University of New York at Buffalo", "country": "", "position": "PhD student" }, { "name": "Owen M Dugan", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Simran Arora", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "PhD student" } ], "n_formula": 258, "n_formula_1": 6, "n_ref_uni": 54, "n_ref": 136, "n_ref_all": 195, "n_fig": 20, "n_tab": 9, "L_tab": 919, "n_element_tab": 93, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 23583, "formula_len_all_1": 415, "len_all": 378347, "len_all_1": 69343, "len_abs": 1600, "len_title": 125, "len_sents": 112255, "len_sents_1": 33307, "n_sents": 1255, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1621, "title": "Towards Learning High-Precision Least Squares Algorithms with Sequence Models", "abs": "This paper investigates whether sequence models can learn to perform numerical algorithms, e.g. gradient descent, on the fundamental problem of least squares. Our goal is to inherit two properties of standard algorithms from numerical analysis: (1) machine precision, i.e. we want to obtain solutions that are accurate to near floating point error, and (2) numerical generality, i.e. we want them to apply broadly across problem instances. We find that prior approaches using Transformers fail to meet these criteria, and identify limitations present in existing architectures and training procedures. First, we show that softmax Transformers struggle to perform high-precision multiplications, which prevents them from precisely learning numerical algorithms. Second, we identify an alternate class of architectures, comprised entirely of polynomials, that can efficiently represent high-precision gradient descent iterates. Finally, we investigate precision bottlenecks during training and address them via a high-precision training recipe that reduces stochastic gradient noise. Our recipe enables us to train two polynomial architectures, gated convolutions and linear attention, to perform gradient descent iterates on least squares problems. For the first time, we demonstrate the ability to train to near machine precision. Applied iteratively, our models obtain $100,000\\times$ lower MSE than standard Transformers trained end-to-end and they incur a $10,000\\times$ smaller generalization gap on out-of-distribution problems. We make progress towards end-to-end learning of numerical algorithms for least squares.", "keywords": [ "high precision", "least squares", "algorithm learning", "Transformers", "gated convolutions", "linear regression", "in-context learning" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 2, 3, 3 ] }, { "paper_id": "smkspydzyN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Anca Dragan", "gender": "Female", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Erik Jones", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Jacob Steinhardt", "gender": "unknown", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 77, "n_ref_all": 111, "n_fig": 3, "n_tab": 5, "L_tab": 856, "n_element_tab": 56, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 586, "formula_len_all_1": 0, "len_all": 230317, "len_all_1": 66894, "len_abs": 1388, "len_title": 98, "len_sents": 65274, "len_sents_1": 34682, "n_sents": 500, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1392, "title": "Adversaries Can Misuse Combinations of Safe Models", "abs": "Developers try to evaluate whether an AI system can accomplish malicious tasks before releasing it; for example, they might test whether a model enables cyberoffense, user manipulation, or bioterrorism. In this work, we show that individually testing models for such misuse is inadequate; adversaries can misuse combinations of models even when each individual model is safe. The adversary accomplishes this by first decomposing tasks into subtasks, then solving each subtask with the best-suited model. For example, an adversary might solve challenging-but-benign subtasks with an aligned frontier model, and easy-but-malicious subtasks with a weaker misaligned model. We study two decomposition methods: manual decomposition where a human identifies a natural decomposition of a task, and automated decomposition where a weak model generates benign tasks for a frontier model to solve, then uses the solutions in-context to solve the original task. Using these decompositions, we empirically show that adversaries can create vulnerable code, explicit images, python scripts for hacking, and manipulative tweets at much higher rates with combinations of models than either individual model. Our work suggests that even perfectly-aligned frontier systems can enable misuse without ever producing malicious outputs, and that red-teaming efforts should extend beyond single models in isolation.", "keywords": [ "safety", "misuse", "adversary", "combining models", "attacks", "hacking" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "slZZnzlITo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Junxian Mu", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "MS student" }, { "name": "Liu Yang", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" }, { "name": "Pengfei Zhu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" }, { "name": "Qinghua Hu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Professor" }, { "name": "RuijieZhang", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "MS student" }, { "name": "Yu Wang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 19, "n_formula_1": 12, "n_ref_uni": 39, "n_ref": 88, "n_ref_all": 117, "n_fig": 19, "n_tab": 9, "L_tab": 4914, "n_element_tab": 597, "n_fig_1": 11, "n_tab_1": 6, "L_tab_1": 4426, "n_element_tab_1": 296, "formula_len_all": 1230, "formula_len_all_1": 591, "len_all": 152535, "len_all_1": 61861, "len_abs": 1334, "len_title": 39, "len_sents": 46731, "len_sents_1": 24537, "n_sents": 346, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1343, "title": "Towards Multimodal Open Set Recognition", "abs": "Open set recognition (OSR) requires deep learning models to identify unknown samples while recognizing known ones. Existing OSR studies focus on single-modal data but merely discuss how to handle multimodal data. In this paper, we propose a new task multimodal open set recognition (MMOSR), extending OSR to more practical scenarios. First, we analyze the necessity of MMOSR and provide insights into the task. We find that simply combining OSR and multimodal fusion methods faces the challenge of fusion degradation. The main reason is that the OSR regularization constrains the fused representations to be excessively compact, leading to deactivated and limited representations. We design the multimodal representation reactivation network (MRN) to alleviate fusion degradation by reactivating suppressed representations. MRN includes the mutually enhanced fusion for enhancing representations and performing cross-modal interaction, and the adaptive fusion for capturing multiple informative representations and outputting the adaptively fused prediction. Thus, the proposed method obtains effective and comprehensive multimodal representations and addresses the challenge of fusion degradation. Finally, extensive experiments on various settings demonstrate that the proposed method is superior to existing methods by up to 5.23\\% on OSCR.", "keywords": [ "Open Set Recognition", "Multimodal Fusion", "Classification" ], "rating_list": [ 5, 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 1, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "slO3xTt4CG", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "David Anugraha", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" }, { "name": "Garry Kuwanto", "gender": "Male", "institution": "Boston University, Boston University", "country": "", "position": "PhD student" }, { "name": "Genta Indra Winata", "gender": "Male", "institution": "Capital One", "country": "US", "position": "Principal Researcher" }, { "name": "Lucky Susanto", "gender": "Male", "institution": "Monash University", "country": "ID", "position": "Researcher" }, { "name": "Derry Wijaya", "gender": "Female", "institution": "Monash University", "country": "ID", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 1, "n_ref_uni": 64, "n_ref": 105, "n_ref_all": 149, "n_fig": 10, "n_tab": 20, "L_tab": 14106, "n_element_tab": 1939, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 3982, "n_element_tab_1": 627, "formula_len_all": 404, "formula_len_all_1": 55, "len_all": 244849, "len_all_1": 69220, "len_abs": 1202, "len_title": 125, "len_sents": 52223, "len_sents_1": 27219, "n_sents": 461, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1207, "title": "MetaMetrics: Calibrating Metrics for Generation Tasks Using Human Preferences", "abs": "Understanding the quality of a performance evaluation metric is crucial for ensuring that model outputs align with human preferences. However, it remains unclear how well each metric captures the diverse aspects of these preferences, as metrics often excel in one particular area but not across all dimensions. To address this, it is essential to systematically calibrate metrics to specific aspects of human preference, catering to the unique characteristics of each aspect. We introduce MetaMetrics, a calibrated meta-metric designed to evaluate generation tasks across different modalities in a supervised manner. MetaMetrics optimizes the combination of existing metrics to enhance their alignment with human preferences. Our metric demonstrates flexibility and effectiveness in both language and vision downstream tasks, showing significant benefits across various multilingual and multi-domain scenarios. MetaMetrics aligns closely with human preferences and is highly extendable and easily integrable into any application. This makes MetaMetrics a powerful tool for improving the evaluation of generation tasks, ensuring that metrics are more representative of human judgment across diverse contexts.", "keywords": [ "metrics", "human preferences", "calibrating", "generation" ], "rating_list": [ 5, 8, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "sl4hOq9wm9", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Jiayin Wang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Qingyao Ai", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Weihang Su", "gender": "Male", "institution": ", Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Yichen Tang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Yiqun LIU", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Zhijing Wu", "gender": "Female", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 23, "n_ref": 33, "n_ref_all": 48, "n_fig": 2, "n_tab": 9, "L_tab": 1468, "n_element_tab": 137, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 436, "n_element_tab_1": 56, "formula_len_all": 181, "formula_len_all_1": 181, "len_all": 131388, "len_all_1": 60095, "len_abs": 1397, "len_title": 149, "len_sents": 47683, "len_sents_1": 29945, "n_sents": 352, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1553, "title": "Knowledge Augmentation: In-context or In-parameter?", "abs": "Large Language Models (LLMs) have achieved remarkable performance in various natural language processing tasks by leveraging relevant external knowledge provided by the users or retrieved from external sources. \nTraditionally, this external information is incorporated by appending it directly to the model\u2019s input context, a paradigm known as in-context knowledge injection.\nHowever, this paradigm faces significant limitations due to the finite input context length of LLMs and often results in shallow integration between the external knowledge and the model\u2019s internal representations.\nTo address the limitations of in-context knowledge injection, we propose a new knowledge injection paradigm called in-parameter knowledge injection, which temporarily embeds the external knowledge relevant to the user\u2019s input directly into the model\u2019s parameters rather than its input context. \nThis new paradigm overcomes the context length limitations of LLMs and enables deeper integration of external information within the model\u2019s internal representations. \nThrough extensive experiments across tasks of varying complexity, we demonstrate that in-parameter knowledge injection achieves significant benefits for complex tasks requiring intricate reasoning. \nIn contrast, in-context injection remains effective for simpler tasks where answers can be directly extracted from the provided information.\n\nWe have open-sourced all the code, data, and models in the following anonymous GitHub link: https://anonymous.4open.science/r/In-parameter-Knowledge-Injection/", "keywords": [ "Representation Learning", "In-Parameter Knowledge Injection", "Parametric Knowledge Representation", "Large Language Models" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "sknUS8X9q0", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Ani Calinescu", "gender": "Not Specified", "institution": "Department of Computer Science, University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Christian Schroeder de Witt", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Lecturer" }, { "name": "Constantin Venhoff", "gender": "Not Specified", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" } ], "n_formula": 28, "n_formula_1": 18, "n_ref_uni": 29, "n_ref": 55, "n_ref_all": 64, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1240, "formula_len_all_1": 784, "len_all": 132208, "len_all_1": 61649, "len_abs": 1845, "len_title": 117, "len_sents": 42510, "len_sents_1": 30878, "n_sents": 297, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1857, "title": "SAGE: Scalable Ground Truth Evaluations for Large Sparse Autoencoders", "abs": "A key challenge in interpretability is to decompose model activations into meaningful features. Sparse autoencoders (SAEs) have emerged as a promising tool for this task. However, a central problem in evaluating the quality of SAEs is the absence of ground truth features to serve as an evaluation gold standard. Current evaluation methods for SAEs are therefore confronted with a significant trade-off: SAEs can either leverage toy models or other proxies with predefined ground truth features; or they use extensive prior knowledge of realistic task circuits. The former limits the generalizability of the evaluation results, while the latter limits the range of models and tasks that can be used for evaluations. We introduce SAGE: Scalable Autoencoder Ground-truth Evaluation, an evaluation framework for SAEs that enables obtaining high-quality feature dictionaries for diverse tasks and feature distributions without relying on prior knowledge. Specifically, we lift previous limitations by showing that ground truth evaluations on realistic tasks can be automated and scaled. First, we show that we can automatically identify the cross-sections in the model where task-specific features are active. Second, we demonstrate that we can then compute the ground truth features at these cross-sections. Third, we introduce a novel reconstruction method which significantly reduces the amount of trained SAEs needed for the evaluation. This addresses scalability limitations in prior work and significantly simplifies the practical evaluations. We validate our results by evaluating SAEs on novel tasks on Pythia70M, GPT-2 Small, and Gemma-2-2B, thus demonstrating the scalability of our method to state-of-the-art open-source frontier models. These advancements pave the way for generalizable, large-scale evaluations of SAEs in interpretability research.", "keywords": [ "Mechanistic interpretability", "Large language models", "Sparse autoencoders", "Sparse dictionary learning", "Unsupervised learning", "Interpretable AI" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 1, 1, 2, 2 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "skJLOae8ew", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Hod Lipson", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Full Professor" }, { "name": "Santiago Yeomans", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 7, "n_ref": 22, "n_ref_all": 27, "n_fig": 3, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 73927, "len_all_1": 49768, "len_abs": 1087, "len_title": 153, "len_sents": 27363, "len_sents_1": 23975, "n_sents": 171, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1090, "title": "From Abstract Noise to Architectural Form: Designing Diffusion Models for Efficient Floor Plan Generation", "abs": "In contemporary architectural design, the generation of innovative and efficient floor plans remains a critical challenge. This research introduces a novel application of diffusion models, specifically adapted for the generation of architectural floor plans. Unlike traditional generative models that broadly target image generation, our approach harnesses the state-of-the-art in diffusion technology to produce detailed, functional, and visually appealing architectural designs. We demonstrate that diffusion models, when finely tuned and conditioned, not only embrace 'implicit, human-learned' architectural semantics but also enhance design efficiency and creativity. The paper details our methodology from adapting the U-Net architecture within diffusion frameworks to incorporating advanced upscaling techniques, significantly reducing computational overhead while maintaining high-resolution outputs. Our results show a promising direction for integrating AI in architectural design, opening new avenues for automated, creative design processes that could revolutionize the industry.", "keywords": [ "Architectural Design Automation", "Generative Models", "Diffusion Models" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 1, 1 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "skHPtDnYGa", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Botian Jiang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Lingpeng Kong", "gender": "Male", "institution": "Department of Computer Science, The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Qi Liu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Xiachong Feng", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "Postdoc" }, { "name": "Xiaonan Li", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Xipeng Qiu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Tobias Lee", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "zhaowei Li", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 25, "n_ref": 65, "n_ref_all": 78, "n_fig": 5, "n_tab": 2, "L_tab": 733, "n_element_tab": 144, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 831, "n_element_tab_1": 139, "formula_len_all": 944, "formula_len_all_1": 988, "len_all": 123157, "len_all_1": 62091, "len_abs": 1613, "len_title": 115, "len_sents": 33324, "len_sents_1": 29335, "n_sents": 220, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1621, "title": "Understanding the Role of LLMs in Multimodal Evaluation Benchmarks", "abs": "The rapid advancement of Multimodal Large Language Models (MLLMs) has been accompanied by the development of various benchmarks to evaluate their capabilities. However, the true nature of these evaluations and the extent to which they assess multimodal reasoning versus merely leveraging the underlying Large Language Model (LLM) backbone remain unclear. This paper presents a comprehensive investigation into the role of LLM backbones in MLLM evaluation, focusing on two critical aspects: the degree to which current benchmarks truly assess multimodal reasoning and the influence of LLM prior knowledge on performance. Specifically, we introduce a modified evaluation protocol to disentangle the contributions of the LLM backbone from multimodal integration, and an automatic knowledge identification technique for diagnosing whether LLMs equip the necessary knowledge for corresponding multimodal questions. Our study encompasses four diverse MLLM benchmarks and eight state-of-the-art MLLMs. Key findings reveal that some benchmarks allow high performance even without visual inputs and up to 50\\% of error rates can be attributed to insufficient world knowledge in the LLM backbone, indicating a heavy reliance on language capabilities. To address knowledge deficiencies, we propose a knowledge augmentation pipeline that achieves significant performance gains, with improvements of up to 60\\% on certain datasets, resulting in a approximately 4x increase in performance. Our work provides crucial insights into the role of the LLM backbone in MLLMs, and highlights the need for more nuanced benchmarking approaches.", "keywords": [ "Multi-modal Language Models", "benchmark purification" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "skGSOcrIj7", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Anastasis Kratsios", "gender": "Non-Binary", "institution": "Vector Institute", "country": "CA", "position": "Researcher" }, { "name": "Haitz S\u00e1ez de Oc\u00e1riz Borde", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Michael M. Bronstein", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Xiaowen Dong", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" }, { "name": "Marc T Law", "gender": "Male", "institution": "NVIDIA", "country": "CA", "position": "Research Scientist" } ], "n_formula": 103, "n_formula_1": 25, "n_ref_uni": 77, "n_ref": 153, "n_ref_all": 187, "n_fig": 11, "n_tab": 10, "L_tab": 5825, "n_element_tab": 690, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1064, "n_element_tab_1": 60, "formula_len_all": 8896, "formula_len_all_1": 2204, "len_all": 316918, "len_all_1": 80843, "len_abs": 1973, "len_title": 97, "len_sents": 99153, "len_sents_1": 35414, "n_sents": 827, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1912, "title": "Neural Spacetimes for DAG Representation Learning", "abs": "We propose a class of trainable deep learning-based geometries called Neural SpaceTimes (NSTs), which can universally represent nodes in weighted Directed Acyclic Graphs (DAGs) as events in a spacetime manifold. While most works in the literature focus on undirected graph representation learning or causality embedding separately, our differentiable geometry can encode both graph edge weights in its spatial dimensions and causality in the form of edge directionality in its temporal dimensions. We use a product manifold that combines a quasi-metric (for space) and a partial order (for time). NSTs are implemented as three neural networks trained in an end-to-end manner: an embedding network, which learns to optimize the location of nodes as events in the spacetime manifold, and two other networks that optimize the space and time geometries in parallel, which we call a neural (quasi-)metric and a neural partial order, respectively. The latter two networks leverage recent ideas at the intersection of fractal geometry and deep learning to shape the geometry of the representation space in a data-driven fashion, unlike other works in the literature that use fixed spacetime manifolds such as Minkowski space or De Sitter space to embed DAGs. Our main theoretical guarantee is a universal embedding theorem, showing that any $k$-point DAG can be embedded into an NST with $1+\\mathcal{O}(\\log(k))$ distortion while exactly preserving its causal structure. The total number of parameters defining the NST is sub-cubic in $k$ and linear in the width of the DAG. If the DAG has a planar Hasse diagram, this is improved to $\\mathcal{O}(\\log(k) + 2)$ spatial and 2 temporal dimensions. We validate our framework computationally with synthetic weighted DAGs and real-world network embeddings; in both cases, the NSTs achieve lower embedding distortions than their counterparts using fixed spacetime geometries.", "keywords": [ "Directed Graphs", "Quasi-metrics" ], "rating_list": [ 6, 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 4, 4, 2 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 4, 2 ], "confidence_list": [ 2, 3, 3, 4, 4 ] }, { "paper_id": "sjvz40tazX", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Ankur Bapna", "gender": "Male", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Eline Visser", "gender": "Female", "institution": "University of Oslo", "country": "", "position": "Postdoc" }, { "name": "Garrett Tanzer", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Orhan Firat", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 23, "n_ref": 61, "n_ref_all": 76, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 146631, "len_all_1": 64844, "len_abs": 1682, "len_title": 108, "len_sents": 39278, "len_sents_1": 34211, "n_sents": 247, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1368, "title": "ASROB: Measuring Automatic Speech Recognition from One Book", "abs": "The MTOB (Machine Translation from One Book) benchmark measures the ability of large language models (LLMs) to \"learn to translate between English and Kalamang\u2014a language with less than 200 speakers and therefore virtually no presence on the web\u2014using several hundred pages of field linguistics reference materials,\" predominantly with long in-context learning. However, many endangered languages like Kalamang are primarily oral, so supporting only text is insufficient: speech must be a first-class citizen for applications to be useful to actual communities. In this paper, we present ASROB (Automatic Speech Recognition from One Book), an extension of MTOB to speech tasks\u2014Kalamang speech-to-text recognition (ASR) and Kalamang speech to English text translation (S2TT)\u2014using an additional 15 hours of transcribed and translated Kalamang speech recordings. Our baselines measure the long mixed-modal (text+audio) in-context learning abilities of the Gemini 1.5 family. 1.5 Pro reaches 24.6% CER on ASR and 6.53 BLEU on S2TT, already beating the 34.2% CER and 4.51 BLEU achieved by a human who learned Kalamang from the same resources, but there is still substantial headroom for models to improve beyond this. We hope that ASROB will help evaluate extreme mixed-modal capabilities in LLMs and increase focus on supporting endangered languages in their spoken form.", "keywords": [ "large language models", "benchmarks", "translation", "speech recognition", "in-context learning", "long context", "Kalamang", "multilinguality" ], "rating_list": [ 8, 3, 5, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 2, 4, 4 ], "contribution_list": [ 4, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "sjWG7B8dvt", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Chong Xiang", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Kaiqiang Song", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Senior Researcher" }, { "name": "Prateek Mittal", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "Full Professor" }, { "name": "Ravi Agrawal", "gender": "Male", "institution": "Zoom Video Communications ", "country": "US", "position": "Researcher" }, { "name": "Sanqiang Zhao", "gender": "Male", "institution": "Zoom", "country": "US", "position": "Researcher" }, { "name": "Sathish Reddy Indurthi", "gender": "Male", "institution": "Zoom Video Communications", "country": "US", "position": "Senior Research Scientist" }, { "name": "Shujian Zhang", "gender": "unknown", "institution": "University of Texas, Austin", "country": "", "position": "PhD student" }, { "name": "Silei Xu", "gender": "Male", "institution": "Zoom", "country": "US", "position": "Researcher" }, { "name": "Tong Wu", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Wenxuan Zhou", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 46, "n_ref": 111, "n_ref_all": 163, "n_fig": 27, "n_tab": 14, "L_tab": 6008, "n_element_tab": 849, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 1744, "n_element_tab_1": 258, "formula_len_all": 67, "formula_len_all_1": 0, "len_all": 206421, "len_all_1": 70473, "len_abs": 1454, "len_title": 128, "len_sents": 56943, "len_sents_1": 31435, "n_sents": 470, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1504, "title": "Instructional Segment Embedding: Improving LLM Safety with Instruction Hierarchy", "abs": "Large Language Models (LLMs) are susceptible to security and safety threats, such as prompt injection, prompt extraction, and harmful requests.\nOne major cause of these vulnerabilities is the lack of an instruction hierarchy.\nModern LLM architectures treat all inputs equally, failing to distinguish between and prioritize various types of instructions, such as system messages, user prompts, and data. \nAs a result, lower-priority user prompts may override more critical system instructions, including safety protocols. \nExisting approaches to achieving instruction hierarchy, such as delimiters and instruction-based training, do not address this issue at the architectural level.\nWe introduce the $\\textbf{I}$nstructional $\\textbf{S}$egment $\\textbf{E}$mbedding (ISE) technique, inspired by BERT, to modern large language models, which embeds instruction priority information directly into the model. \nThis approach enables models to explicitly differentiate and prioritize various instruction types, significantly improving safety against malicious prompts that attempt to override priority rules. \nOur experiments on the Structured Query and Instruction Hierarchy benchmarks demonstrate an average robust accuracy increase of up to 15.75\\% and 18.68\\%, respectively. \nFurthermore, we observe an improvement in instruction-following capability of up to 4.1\\% evaluated on AlpacaEval. \nOverall, our approach offers a promising direction for enhancing the safety and effectiveness of LLM architectures.", "keywords": [ "Instruction Hierarchy", "Segment Embedding", "LLM safety and robustness" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "sjGmiI49sd", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "GuangHao Meng", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jieming Zhu", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Jinpeng Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Letian Zhang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Qing Li", "gender": "Male", "institution": "Pengcheng Laboratory", "country": "CN", "position": "Full Professor" }, { "name": "Tao Dai", "gender": "Male", "institution": "Department of Software Engineering, Shenzhen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yong Jiang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 30, "n_ref": 43, "n_ref_all": 53, "n_fig": 1, "n_tab": 3, "L_tab": 406, "n_element_tab": 77, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1964, "n_element_tab_1": 78, "formula_len_all": 516, "formula_len_all_1": 516, "len_all": 111585, "len_all_1": 65357, "len_abs": 1443, "len_title": 118, "len_sents": 30209, "len_sents_1": 28720, "n_sents": 219, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1449, "title": "Multimodal LLM-guided Query Optimization for Visual-Language Retrieval", "abs": "Vision-language retrieval (VLR), involving the use of text (or images) as queries to retrieve corresponding images (or text), has been widely used in multimedia and computer vision tasks. However, ambiguous or complex concepts contained in queries often confuse retrievers, making it difficult to effectively align these concepts with visual content, thereby limiting their performance. Existing query optimization methods neglect the feedback of retrievers' preferences, thus resulting in sub-optimal performance. Inspired by the powerful ability of Multimodal Large Language Models (MLLMs), we propose a Multimodal LLM-Guided Query Rewriter (MGQRe) for query optimization. Specifically, MGQRe first utilizes MLLM to explore the retriever's weakness and perform targeted iterative optimizations to capture the retriever's expressive preferences. Subsequently, we develop a trainable rewriter that learns this preference knowledge through a three-step tuning strategy: supervised fine-tuning, preference learning, and reinforcement learning. This ensures that the queries generated by the rewriter align with the retriever\u2019s preferences, thereby enhancing the retriever's performance. Extensive VLR benchmark experiments have demonstrated the superiority of MGQRe, as well as its generalizability and transferability. This work showcases the potential of using advanced language models to overcome the inherent limitations in current VLR technology.", "keywords": [ "vision-language retrieval", "cross-modal retrieval", "prompt engineering", "query rewriting", "large language model" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "sicB10feCQ", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Chengjie Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Feng Zheng", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Jialin Li", "gender": "Male", "institution": "Tencent YouTu Lab", "country": "CN", "position": "Researcher" }, { "name": "Lei Li", "gender": "Not Specified", "institution": "University of Washington", "country": "US", "position": "Researcher" }, { "name": "Yifeng Zhou", "gender": "unknown", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Yong Liu", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Yuexian Zou", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Ziyu Yao", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Jiang Xi", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 53, "n_ref": 86, "n_ref_all": 103, "n_fig": 5, "n_tab": 3, "L_tab": 856, "n_element_tab": 140, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1731, "n_element_tab_1": 152, "formula_len_all": 492, "formula_len_all_1": 499, "len_all": 130715, "len_all_1": 59092, "len_abs": 194, "len_title": 112, "len_sents": 31338, "len_sents_1": 27139, "n_sents": 194, "n_sents_1": 166, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1403, "title": "CAR: Controllable Autoregressive Modeling for Visual Generation", "abs": "Controllable generation, which enables fine-grained control over generated outputs, has emerged as a critical focus in visual generative models. Currently, there are two primary technical approaches in visual generation: diffusion models and autoregressive models. Diffusion models, as exemplified by ControlNet and T2I-Adapter, offer advanced control mechanisms, whereas autoregressive models, despite showcasing impressive generative quality and scalability, remain underexplored in terms of controllability and flexibility. In this study, we introduce Controllable AutoRegressive Modeling (CAR), a novel, plug-and-play framework that integrates conditional control into multi-scale latent variable modeling, enabling efficient control generation within a pre-trained visual autoregressive model. CAR progressively refines and captures control representations, which are injected into each autoregressive step of the pre-trained model to guide the generation process. Our approach demonstrates excellent controllability across various types of conditions and delivers higher image quality compared to previous methods. Additionally, CAR achieves robust generalization with significantly fewer training resources compared to those required for pre-training the model. To the best of our knowledge, we are the first to propose a control framework for pre-trained autoregressive visual generation models.", "keywords": [ "controllable generation", "autoregressive models", "visual generative models" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "siHHqDDzvS", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Changwen Zhang", "gender": "Male", "institution": "lenovo", "country": "", "position": "Researcher" }, { "name": "Hao Yuan", "gender": "Male", "institution": "Lenovo Research", "country": "CN", "position": "Researcher" }, { "name": "Liming Gong", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "MS student" }, { "name": "Wenli Ouyang", "gender": "unknown", "institution": "Lenovo Research ", "country": "CN", "position": "Researcher" }, { "name": "Yong Sun", "gender": "Male", "institution": "Lenovo", "country": "", "position": "Researcher" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 35, "n_ref": 133, "n_ref_all": 175, "n_fig": 7, "n_tab": 15, "L_tab": 7586, "n_element_tab": 1069, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2666, "n_element_tab_1": 337, "formula_len_all": 683, "formula_len_all_1": 438, "len_all": 181532, "len_all_1": 72189, "len_abs": 1379, "len_title": 130, "len_sents": 51575, "len_sents_1": 28957, "n_sents": 394, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1385, "title": "BTBS-LNS: Binarized-Tightening, Branch and Search on Learning LNS Policies for MIP", "abs": "Learning to solve large-scale Mixed Integer Program (MIP) problems is an emerging research topic, and policy learning-based Large Neighborhood Search (LNS) has been a popular paradigm. However, the explored space of LNS policy is often limited even in the training phase, making the learned policy sometimes wrongly fix some potentially important variables early in the search, leading to local optimum in some cases. Moreover, many methods only assume binary variables to deal with. We present a practical approach, termed Binarized-Tightening Branch-and-Search for Large Neighborhood Search (BTBS-LNS). It comprises three key techniques: 1) the ``Binarized Tightening\" technique for integer variables to handle their wide range by binary encoding and bound tightening; 2) an attention-based tripartite graph to capture global correlations among variables and constraints for an MIP instance; 3) an extra branching network as a global view, to identify and optimize wrongly-fixed backdoor variables at each search step. Experiments show its superior performance over the open-source solver SCIP and LNS baselines. Moreover, it performs competitively with, and sometimes better than the commercial solver Gurobi (v9.5.0), especially on the MIPLIB2017 benchmark chosen by Hans Mittelmann, where our method can deliver 10\\% better primal gaps compared with Gurobi in a 300s cut-off time.", "keywords": [ "Large neighborhood search", "Reinforcement learning", "Bound tightening" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "si37wk8U5D", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Chaoyue Wang", "gender": "Male", "institution": "JD.com", "country": "CN", "position": "Researcher" }, { "name": "Chuanxia Zheng", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Huiang He", "gender": "Male", "institution": "South China University of Technology", "country": "", "position": "MS student" }, { "name": "Minghui Hu", "gender": "Male", "institution": "SpellBrush", "country": "JP", "position": "Researcher" }, { "name": "Tat-Jen Cham", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" } ], "n_formula": 30, "n_formula_1": 15, "n_ref_uni": 89, "n_ref": 204, "n_ref_all": 226, "n_fig": 16, "n_tab": 7, "L_tab": 1464, "n_element_tab": 193, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 285, "n_element_tab_1": 9, "formula_len_all": 2316, "formula_len_all_1": 884, "len_all": 221022, "len_all_1": 59995, "len_abs": 1524, "len_title": 0, "len_sents": 45141, "len_sents_1": 27801, "n_sents": 327, "n_sents_1": 158, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1494, "title": "Semantix: An Energy-guided Sampler for Semantic Style Transfer", "abs": "Recent advances in style and appearance transfer are impressive, but most methods isolate global style and local appearance transfer, neglecting semantic correspondence. Additionally, image and video tasks are typically handled in isolation, with little focus on integrating them for video transfer. To address these limitations, we introduce a novel task, *Semantic Style Transfer*, which involves transferring style and appearance features from a reference image to a target visual content based on semantic correspondence. We subsequently propose a training-free method, *Semantix*, an energy-guided sampler designed for Semantic Style Transfer that simultaneously guides both style and appearance transfer based on semantic understanding capacity of pre-trained diffusion models. Additionally, as a sampler, *Semantix* can be seamlessly applied to both image and video models, enabling semantic style transfer to be generic across various visual media. Specifically, once inverting both reference and context images or videos to noise space by SDEs, *Semantix* utilizes a meticulously crafted energy function to guide the sampling process, including three key components: *Style Feature Guidance*, *Spatial Feature Guidance* and *Semantic Distance* as a regularisation term. Experimental results demonstrate that *Semantix* not only effectively accomplishes the task of semantic style transfer across images and videos, but also surpasses existing state-of-the-art solutions in both fields.", "keywords": [ "style transfer", "diffusion model", "energy guidance" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "shqjOIK3SA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Kezhao Liu", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Liang Lin", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "US", "position": "Full Professor" }, { "name": "Pengxu Wei", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaogang Xu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Researcher" }, { "name": "Yao Xiao", "gender": "Male", "institution": "Sun Yat-sen University", "country": "CN", "position": "PhD student" }, { "name": "ZiYi Dong", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "LiuYiming", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "MS student" } ], "n_formula": 30, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 85, "n_ref_all": 130, "n_fig": 11, "n_tab": 19, "L_tab": 5161, "n_element_tab": 413, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1769, "n_element_tab_1": 159, "formula_len_all": 1880, "formula_len_all_1": 547, "len_all": 199704, "len_all_1": 64877, "len_abs": 1534, "len_title": 142, "len_sents": 68723, "len_sents_1": 28769, "n_sents": 517, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1542, "title": "Towards Understanding the Robustness of Diffusion-Based Purification: A Stochastic Perspective", "abs": "Diffusion-Based Purification (DBP) has emerged as an effective defense mechanism against adversarial attacks. The efficacy of DBP has been attributed to the forward diffusion process, which narrows the distribution gap between clean and adversarial images through the addition of Gaussian noise. Although this explanation has some theoretical support, the significance of its contribution to robustness remains unclear. In this paper, we argue that the inherent stochasticity in the DBP process is the primary driver of its robustness. To explore this, we introduce a novel Deterministic White-Box (DW-box) evaluation protocol to assess robustness in the absence of stochasticity and to analyze the attack trajectories and loss landscapes. Our findings suggest that DBP models primarily leverage stochasticity to evade effective attack directions, and their ability to purify adversarial perturbations can be weak. To further enhance the robustness of DBP models, we introduce Adversarial Denoising Diffusion Training (ADDT), which incorporates classifier-guided adversarial perturbations into diffusion training, thereby strengthening the DBP models' ability to purify adversarial perturbations. Additionally, we propose Rank-Based Gaussian Mapping (RBGM) to make perturbations more compatible with diffusion models. Experimental results validate the effectiveness of ADDT. In conclusion, our study suggests that future research on DBP can benefit from the perspective of decoupling the stochasticity-based and purification-based robustness.", "keywords": [ "Adversarial Defense", "Adversarial Purification", "Diffusion Training", "Randomized Defense" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "sgke1JuVlc", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Bin Gu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Velibor Bojkovic", "gender": "unknown", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Xiaofeng Wu", "gender": "Male", "institution": "City University of Macau ", "country": "MO", "position": "Assistant Professor" } ], "n_formula": 29, "n_formula_1": 17, "n_ref_uni": 43, "n_ref": 71, "n_ref_all": 84, "n_fig": 5, "n_tab": 12, "L_tab": 8736, "n_element_tab": 992, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2494, "n_element_tab_1": 284, "formula_len_all": 16517, "formula_len_all_1": 15998, "len_all": 243819, "len_all_1": 85154, "len_abs": 1269, "len_title": 124, "len_sents": 54943, "len_sents_1": 31525, "n_sents": 388, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1250, "title": "Temporal Misinformation and Conversion through Probabilistic Spiking Neurons", "abs": "In the age of large neural network models and their high energy demand, Spiking Neural Networks (SNNs) offer a compelling alternative to Artificial Neural Networks (ANNs) due to their energy efficiency and resemblance to biological brains. However, directly training SNNs with spatio-temporal backpropagation remains challenging due to their discrete signal processing and temporal dynamics. Alternative methods, notably ANN-SNN conversion, have enabled SNNs to achieve performance in various machine learning tasks, comparable to ANNs, but often to the expense of long latency needed to achieve such performance, especially on large scale complex datasets. The present work deals with ANN-SNN setting and identifies a new phenomenon we term ``temporal misinformation'', where random spike rearrangement through time in the converted SNN model improves its performance. To account for this, we propose bio-plausible, two-phase probabilistic (TPP) spiking neurons to be used in ANN-SNN conversion. We showcase the benefits of our proposed methods both theoretically and empirically through extensive experiments on CIFAR-10/100 and a large-scale dataset ImageNet over a variety of architectures, reaching SOTA performance. Code is available on GitHub.", "keywords": [ "spiking neural networks", "probabilistic spiking", "ANN-SNN conversion" ], "rating_list": [ 3, 3, 6, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "sgbI8Pxwie", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Jiangxuan Long", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yingyu Liang", "gender": "Male", "institution": "University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Yufa Zhou", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "MS student" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Zhenmei Shi", "gender": "Male", "institution": "Voyage AI", "country": "US", "position": "Researcher" } ], "n_formula": 104, "n_formula_1": 24, "n_ref_uni": 59, "n_ref": 117, "n_ref_all": 147, "n_fig": 3, "n_tab": 2, "L_tab": 661, "n_element_tab": 23, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 14364, "formula_len_all_1": 1586, "len_all": 258419, "len_all_1": 63804, "len_abs": 1266, "len_title": 124, "len_sents": 56696, "len_sents_1": 26533, "n_sents": 731, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1147, "title": "Beyond Linear Approximations: A Novel Pruning Approach for Attention Matrix", "abs": "Large Language Models (LLMs) have shown immense potential in enhancing various aspects of our daily lives, from conversational AI to search and AI assistants. However, their growing capabilities come at the cost of extremely large model sizes, making deployment on edge devices challenging due to memory and computational constraints. \nThis paper introduces a novel approach to LLM weight pruning that directly optimizes for approximating the attention matrix, a core component of transformer architectures. \nUnlike existing methods that focus on linear approximations, our approach accounts for the non-linear nature of the Softmax attention mechanism. \nWe provide theoretical guarantees for the convergence of our Gradient Descent-based optimization method to a near-optimal pruning mask solution. \nOur preliminary empirical results demonstrate the effectiveness of this approach in maintaining model performance while significantly reducing computational costs. \nThis work establishes a new theoretical foundation for pruning algorithm design in LLMs, potentially paving the way for more efficient LLM inference on resource-constrained devices.", "keywords": [ "Weights Pruning", "Attention Approximation", "Gradient Descent Optimization" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 2, 4, 3 ] }, { "paper_id": "sgaMYvGRG5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Chen Feng", "gender": "Male", "institution": "New York University", "country": "US", "position": "Associate Professor" }, { "name": "Fei Miao", "gender": "Female", "institution": "University of Connecticut", "country": "US", "position": "Associate Professor" }, { "name": "Nuo Chen", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" }, { "name": "Sanbao Su", "gender": "Male", "institution": "University of Connecticut", "country": "US", "position": "PhD student" }, { "name": "Felix Juefei-Xu", "gender": "Male", "institution": "GenAI, Meta", "country": "US", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 9, "n_ref_uni": 54, "n_ref": 105, "n_ref_all": 134, "n_fig": 11, "n_tab": 7, "L_tab": 1510, "n_element_tab": 200, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 693, "n_element_tab_1": 107, "formula_len_all": 1944, "formula_len_all_1": 485, "len_all": 173893, "len_all_1": 62942, "len_abs": 1464, "len_title": 118, "len_sents": 51452, "len_sents_1": 29206, "n_sents": 404, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1483, "title": "$\\alpha$-OCC: Uncertainty-Aware Camera-based 3D Semantic Occupancy Prediction", "abs": "In the realm of autonomous vehicle (AV) perception, comprehending 3D scenes is paramount for tasks such as planning and mapping. Camera-based 3D Semantic Occupancy Prediction (OCC) aims to infer scene geometry and semantics from limited observations. While it has gained popularity due to affordability and rich visual cues, existing methods often neglect the inherent uncertainty in models. To address this, we propose an uncertainty-aware camera-based 3D semantic occupancy prediction method ($\\alpha$-OCC). Our approach includes an uncertainty propagation framework (Depth-UP) from depth models to enhance geometry completion (up to 11.58\\% improvement) and semantic segmentation (up to 12.95\\% improvement) for a variety of OCC models. Additionally, we propose a hierarchical conformal prediction (HCP) method to quantify OCC uncertainty, effectively addressing the high-level class imbalance in OCC datasets. On the geometry level, we present a novel KL-based score function that significantly improves the occupied recall of safety-critical classes (45\\% improvement) with minimal performance overhead (3.4\\% reduction). For uncertainty quantification, we demonstrate the ability to achieve smaller prediction set sizes while maintaining a defined coverage guarantee. Compared with baselines, it reduces up to 92\\% set size. Our contributions represent significant advancements in OCC accuracy and robustness, marking a noteworthy step forward in autonomous perception systems.", "keywords": [ "Uncertainty Propagation", "Semantic Occupancy Prediction", "Conformal Prediction" ], "rating_list": [ 6, 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 5, 4 ] }, { "paper_id": "sgHnfLX9Lt", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Anit Kumar Sahu", "gender": "Male", "institution": "Amazon AGI", "country": "US", "position": "Senior Applied Scientist" }, { "name": "Furong Huang", "gender": "Female", "institution": "University of Maryland", "country": "US", "position": "Associate Professor" }, { "name": "Furqan Khan", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Marco Bornstein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Amrit Bedi", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" } ], "n_formula": 60, "n_formula_1": 21, "n_ref_uni": 24, "n_ref": 60, "n_ref_all": 105, "n_fig": 9, "n_tab": 6, "L_tab": 1503, "n_element_tab": 33, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5189, "formula_len_all_1": 1485, "len_all": 173241, "len_all_1": 67544, "len_abs": 1009, "len_title": 134, "len_sents": 56002, "len_sents_1": 29948, "n_sents": 491, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1011, "title": "Towards Realistic Mechanisms That Incentivize Federated Participation and Contribution", "abs": "Edge device participation in federating learning (FL) is typically studied through the lens of device-server communication (e.g., device dropout) and assumes an undying desire from edge devices to participate in FL. As a result, current FL frameworks are flawed when implemented in realistic settings, with many encountering the free-rider dilemma. In a step to push FL towards realistic settings, we propose RealFM: the first federated mechanism that (1) realistically models device utility, (2) incentivizes data contribution and device participation, (3) provably removes the free-rider dilemma, and (4) relaxes assumptions on data homogeneity and data sharing. Compared to previous FL mechanisms, RealFM allows for a non-linear relationship between model accuracy and utility, which improves the utility gained by the server and participating devices. On real-world data, RealFM improves device and server utility, as well as data contribution, by over 3 and 4 magnitudes respectively compared to baselines.", "keywords": [ "Federated Learning", "Mechanisms", "Realistic", "Utility" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "sgAp2qG86e", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Alexander Kolesnikov", "gender": "unknown", "institution": "OpenAI", "country": "CH", "position": "Researcher" }, { "name": "Andr\u00e9 Susano Pinto", "gender": "Male", "institution": "Google DeepMind", "country": "CH", "position": "Software Engineer" }, { "name": "Michael Tschannen", "gender": "unknown", "institution": "Google DeepMind", "country": "CH", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 63, "n_ref": 116, "n_ref_all": 136, "n_fig": 8, "n_tab": 5, "L_tab": 976, "n_element_tab": 127, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 368, "n_element_tab_1": 38, "formula_len_all": 337, "formula_len_all_1": 210, "len_all": 163923, "len_all_1": 59760, "len_abs": 1384, "len_title": 116, "len_sents": 38658, "len_sents_1": 29534, "n_sents": 276, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1398, "title": "JetFormer: An autoregressive generative model of raw images and text", "abs": "Removing modeling constraints and unifying architectures across domains has been a key driver of the recent progress in training large multimodal models. However, most of these models still rely on many separately trained components such as modality-specific encoders and decoders. In this work, we further streamline joint generative modeling of images and text. We propose an autoregressive decoder-only transformer---JetFormer---which is trained to directly maximize the likelihood of raw data, without relying on any separately pretrained components, and can understand and generate both text and images. Specifically, we leverage a normalizing flow model to obtain a soft-token image representation that is jointly trained with an autoregressive multimodal transformer. The normalizing flow model serves as both an image encoder for perception tasks and an image decoder for image generation tasks during inference. JetFormer achieves text-to-image generation quality competitive with recent VQVAE- and VAE-based baselines. These baselines rely on pretrained image autoencoders, which are trained with a complex mixture of losses, including perceptual ones. At the same time, JetFormer demonstrates robust image understanding capabilities. To the best of our knowledge, JetFormer is the first model that is capable of generating high-fidelity images and producing strong log-likelihood bounds.", "keywords": [ "generative pretraining", "multimodal models", "vision-language model", "text-to-image", "normalizing flow", "generative model" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "sfi2j1Ot6j", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Haoran Geng", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Qinghua Cui", "gender": "Male", "institution": "Wuhan Institute of Physical Education", "country": "", "position": "Full Professor" }, { "name": "Shihao Shao", "gender": "Not Specified", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zun Wang", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 41, "n_formula_1": 12, "n_ref_uni": 44, "n_ref": 118, "n_ref_all": 136, "n_fig": 9, "n_tab": 9, "L_tab": 5165, "n_element_tab": 710, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2710, "n_element_tab_1": 304, "formula_len_all": 2766, "formula_len_all_1": 1477, "len_all": 192790, "len_all_1": 72695, "len_abs": 2230, "len_title": 140, "len_sents": 54001, "len_sents_1": 29999, "n_sents": 533, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1726, "title": "FreeCG: Free the Design Space of Clebsch-Gordan Transform for Machine Learning Force Fields", "abs": "Machine Learning Force Fields (MLFFs) are of great importance for chemistry, physics, materials science, and many other related fields. The Clebsch\u2013Gordan transform (CG transform) effectively encodes many-body interactions and is thus an important building block for many models of MLFFs. However, the permutation-equivariance requirement of MLFFs limits the design space of CG transform, that is, intensive CG transform has to be conducted for each neighboring edge and the operations should be performed in the same manner for all edges. Freeing up the design space can greatly improve the model's expressiveness while simultaneously decreasing computational demands.\nTo reach this goal, we utilize a mathematical proposition, invariance transitivity, to show that implementing the CG transform layer on the permutation-invariant abstract edges allows complete freedom in the design of the layer without compromising the overall permutation equivariance. Developing on this free design space, we further propose group CG transform with sparse path, abstract edges shuffling, and attention enhancer to form a powerful and efficient CG transform layer. Our method, known as **FreeCG**, achieves state-of-the-art (SOTA) results in force prediction for MD17, rMD17, MD22, and is well extended to property prediction in QM9 datasets with several improvements greater than 15% and the maximum beyond 20%. The extensive real-world applications showcase high practicality. FreeCG introduces a novel paradigm for carrying out efficient and expressive CG transform in future geometric network designs. To demonstrate this, the recent SOTA, QuinNet, is also enhanced under our paradigm. Code and checkpoints will be publicly available.", "keywords": [ "Machine Learning Force Fields", "Equivariant Graph Neural Network", "Clebsch-Gordan Transform" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "sfQ6XpApfS", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Jia-Yu Yao", "gender": "Male", "institution": "Peking University", "country": "", "position": "RA" }, { "name": "Kun-Peng Ning", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Shuo Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yibing Song", "gender": "unknown", "institution": "Alibaba DAMO Academy", "country": "CN", "position": "Staff Scientist" }, { "name": "Yonghong Tian", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Yuyang Liu", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Postdoc" }, { "name": "Zhenhui Liu", "gender": "Non-Binary", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Yuan LI", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 28, "n_formula_1": 13, "n_ref_uni": 41, "n_ref": 95, "n_ref_all": 117, "n_fig": 10, "n_tab": 9, "L_tab": 3010, "n_element_tab": 133, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 1207, "n_element_tab_1": 96, "formula_len_all": 1343, "formula_len_all_1": 501, "len_all": 171063, "len_all_1": 64086, "len_abs": 1149, "len_title": 108, "len_sents": 44074, "len_sents_1": 29677, "n_sents": 329, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1159, "title": "PiCO: Peer Review in LLMs based on Consistency Optimization", "abs": "Existing large language models (LLMs) evaluation methods typically focus on testing the performance on some closed-environment and domain-specific benchmarks with human annotations. In this paper, we explore a novel unsupervised evaluation direction, utilizing peer-review mechanisms to measure LLMs automatically without any human feedback. In this setting, both open-source and closed-source LLMs lie in the same environment, capable of answering unlabeled questions and evaluating each other, where each LLM\u2019s response score is jointly determined by other anonymous ones. During this process, we found that those answers that are more recognized by other ``reviewers'' (models) usually come from LLMs with stronger abilities, while these models can also evaluate others' answers more accurately. We formalize it as a consistency assumption, i.e., the ability and score of the model usually have consistency. We exploit this to optimize each model's confidence, thereby re-ranking the LLMs to be closer to human rankings. We perform experiments on multiple datasets with standard rank-based metrics, validating the effectiveness of the proposed approach.", "keywords": [ "Large Language Model", "Unsupervised Evaluation", "Peer Review", "Consistency Optimization" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "sf1u3vTRjm", "primary_area": "datasets and benchmarks", "n_author": 23, "author_info_list": [ { "name": "Arman Cohan", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" }, { "name": "Baobao Chang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Chen Sheng", "gender": "Male", "institution": "nanjing university", "country": "", "position": "MS student" }, { "name": "Haozhe Zhao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Helan Hu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Junjie Lu", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "MS student" }, { "name": "Kaikai An", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Liang Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Shuzheng Si", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Tianyu Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Staff Engineer" }, { "name": "Wangchunshu Zhou", "gender": "Male", "institution": "Guangdong OPPO Mobile Telecommunications Corp.,Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Blind Name", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Yanjun Shao", "gender": "Male", "institution": "Yale University", "country": "US", "position": "MS student" }, { "name": "Yichi Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yilun Zhao", "gender": "unknown", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Yin Fang", "gender": "Female", "institution": "National Institutes of Health", "country": "US", "position": "Postdoc" }, { "name": "Yujia Qin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yuliang Liu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Zefan Cai", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Zexuan Deng", "gender": "Male", "institution": "nanjing university", "country": "", "position": "MS student" }, { "name": "Zhiwei Jiang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Mark Gerstein", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Full Professor" }, { "name": "\u9ec4\u777f\u92c6", "gender": "Male", "institution": "NCEPU", "country": "", "position": "Undergrad student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 35, "n_ref": 65, "n_ref_all": 96, "n_fig": 7, "n_tab": 18, "L_tab": 8613, "n_element_tab": 783, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1737, "n_element_tab_1": 78, "formula_len_all": 352, "formula_len_all_1": 147, "len_all": 244163, "len_all_1": 54212, "len_abs": 1606, "len_title": 105, "len_sents": 78732, "len_sents_1": 25977, "n_sents": 655, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1624, "title": "ML-Bench: Evaluating Large Language Models for Code Generation in Repository-Level Machine Learning Tasks", "abs": "Despite Large Language Models (LLMs) achieving impressive results in code generation, significant challenges remain in automated ML development, particularly in utilizing existing ML repositories effectively.\nAlso, recently, people have developed LLM agents that attempt to interact with repository code (e.g., resolving issues), prompting the need for end-to-end evaluations starting from environment setup to deploying the repository rather than merely generating code in already-configured environments. \nThese two gaps have motivated our development of ML-Bench, a benchmark rooted in real-world ML applications that leverage existing code repositories. \nML-Bench encompasses annotated 9,641 examples across 18 GitHub repositories, challenging LLMs to accommodate user-specified arguments and documentation intricacies effectively.\nTo evaluate both LLMs and agents, two setups are employed: \nML-Bench-L for assessing LLMs' text-to-code conversion within a predefined deployment environment, and ML-Bench-A for testing autonomous agents in an end-to-end task execution within a Linux sandbox environment. \nOur findings indicate that while GPT-4o leads with a Pass@5 rate surpassing 50%, there remains significant scope for improvement, highlighted by issues such as hallucinated outputs and difficulties with bash script generation. \nNotably, in the more demanding ML-Agent-Bench, GPT-4o achieves a 76.47% success rate, reflecting the efficacy of iterative action and feedback in complex task resolution. \nOur resources, including code, data, and models, are available at \\url{https://anonymous.4open.science/r/ML-Bench}.", "keywords": [ "LLMs", "code generation", "Agents", "Repository" ], "rating_list": [ 6, 3, 6, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "semTHoVGsJ", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Christopher Earls", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Full Professor" }, { "name": "Nicolas Boulle", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "Assistant Professor" }, { "name": "Rapha\u00ebl Sarfati", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Researcher" }, { "name": "Toni J.B. Liu", "gender": "Male", "institution": "Cornell University", "country": "", "position": "PhD student" } ], "n_formula": 39, "n_formula_1": 13, "n_ref_uni": 44, "n_ref": 97, "n_ref_all": 149, "n_fig": 16, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1576, "formula_len_all_1": 533, "len_all": 181406, "len_all_1": 55910, "len_abs": 1142, "len_title": 141, "len_sents": 52925, "len_sents_1": 25080, "n_sents": 427, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1153, "title": "Density estimation with LLMs: a geometric investigation of in-context learning trajectories", "abs": "Large language models (LLMs) demonstrate remarkable emergent abilities to perform in-context learning across various tasks, including time series forecasting. \nThis work investigates LLMs' ability to estimate probability density functions (PDFs) from data observed in-context; \nsuch density estimation (DE) is a fundamental task underlying many probabilistic modeling problems. \nWe leverage the Intensive Principal Component Analysis (InPCA) to visualize and analyze the in-context learning dynamics of LLaMA-2 models. \nOur main finding is that these LLMs all follow similar learning trajectories in a low-dimensional InPCA space, which are distinct from those of traditional density estimation methods like histograms and Gaussian kernel density estimation (KDE). \nWe interpret the LLaMA in-context DE process as a KDE with an adaptive kernel width and shape. \nThis custom kernel model captures a significant portion of LLaMA's behavior despite having only two parameters. \nWe further speculate on why LLaMA's kernel width and shape differs from classical algorithms, providing insights into the mechanism of in-context probabilistic reasoning in LLMs.", "keywords": [ "Large language models", "in-context learning", "Intensive Principal Component Analysis", "density estimation", "Mechanistic interpretations", "High-dimensional learning dynamics", "dimensionality reduction" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "sejvgf030w", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Fei Zhu", "gender": "Male", "institution": "Centre for Artificial Intelligence and Robotics Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "CN", "position": "Postdoc" }, { "name": "Zhaoxiang Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhen Cheng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Cheng-lin Liu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 10, "n_ref_uni": 49, "n_ref": 126, "n_ref_all": 163, "n_fig": 10, "n_tab": 14, "L_tab": 8276, "n_element_tab": 858, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 5950, "n_element_tab_1": 584, "formula_len_all": 1503, "formula_len_all_1": 777, "len_all": 184662, "len_all_1": 78053, "len_abs": 1353, "len_title": 99, "len_sents": 45030, "len_sents_1": 29849, "n_sents": 342, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1359, "title": "Towards Flexible and Controllable Unknown Rejection", "abs": "Reliable prediction is an essential requirement for deep neural models that are deployed in open environments, where both covariate and semantic out-of-distribution (OOD) data arise naturally. Recent studies have formulated and pursued two problems named OOD generalization and detection independently, where the former aims to correctly recognize covariate shifts while the latter focuses on rejecting semantic shifts. However, existing methods are misaligned with real-world applications in two aspects. First, in practice, to make safe decisions, a reliable model should accept correctly recognized inputs while rejecting both those misclassified covariate-shifted and semantic-shifted examples. Second, considering the potential existing trade-off between rejecting different failure cases, more convenient, controllable, and flexible unknown rejection approaches are needed. To meet the above requirements, we propose a novel and elegantly simple unknown rejection framework to unify and facilitate classification with rejection under both covariate and semantic shifts. Our key insight is that by separating and consolidating failure-specific reliability knowledge with low-rank adapters and then integrating them, we can enhance the unknown rejection ability effectively and flexibly. Extensive experiments demonstrate the superiority of our framework.", "keywords": [ "Unknown Rejection", "Reliability", "Open-world classification" ], "rating_list": [ 3, 5, 8, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 4, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "sec09tLQUl", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Eugene Belilovsky", "gender": "Male", "institution": "Concordia University, Montreal", "country": "CA", "position": "Assistant Professor" }, { "name": "geraldin nanfack", "gender": "Male", "institution": "Concordia University", "country": "CA", "position": "Postdoc" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 34, "n_ref": 70, "n_ref_all": 80, "n_fig": 5, "n_tab": 5, "L_tab": 2192, "n_element_tab": 126, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1604, "n_element_tab_1": 76, "formula_len_all": 191, "formula_len_all_1": 196, "len_all": 101508, "len_all_1": 68158, "len_abs": 1076, "len_title": 133, "len_sents": 27058, "len_sents_1": 26560, "n_sents": 197, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1087, "title": "FairDropout: Using Example-Tied Dropout to Enhance Generalization for Minority Groups", "abs": "Deep learning models frequently exploit spurious features in training data to achieve low training error, often resulting in poor generalization when faced with shifted testing distributions. To address this issue, various methods from imbalanced learning, representation learning, and classifier recalibration have been proposed to enhance the robustness of deep neural networks against spurious correlations. In this paper, we observe that models trained with empirical risk minimization tend to generalize well for examples from the majority groups while memorizing instances from minority groups.\nBuilding on recent findings that show memorization can be localized to a limited number of neurons, we apply example-tied dropout as a method we term \\textit{FairDropout}, aimed at redirecting this memorization to specific neurons that we subsequently drop out during inference. We empirically evaluate FairDropout using the subpopulation benchmark suite encompassing vision, language, and healthcare tasks, demonstrating that it significantly reduces reliance on spurious correlations.", "keywords": [ "spurious correlation", "fairness", "worst-group performance" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 5, 3, 5 ] }, { "paper_id": "se4vjm7h4E", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Boris Ginsburg", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Cheng-Ping Hsieh", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Ilya Loshchilov", "gender": "unknown", "institution": "NVIDIA", "country": "FR", "position": "Researcher" }, { "name": "Simeng Sun", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 37, "n_formula_1": 28, "n_ref_uni": 30, "n_ref": 46, "n_ref_all": 70, "n_fig": 11, "n_tab": 8, "L_tab": 8872, "n_element_tab": 1370, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 2215, "n_element_tab_1": 1445, "formula_len_all": 1350, "formula_len_all_1": 1007, "len_all": 153018, "len_all_1": 67020, "len_abs": 699, "len_title": 125, "len_sents": 49606, "len_sents_1": 22358, "n_sents": 375, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 701, "title": "nGPT: Normalized Transformer with Representation Learning on the Hypersphere", "abs": "We propose a novel neural network architecture, the normalized Transformer (nGPT) with representation learning on the hypersphere. In nGPT, all vectors forming the embeddings, MLP, attention matrices and hidden states are unit norm normalized. The input stream of tokens travels on the surface of a hypersphere, with each layer contributing a displacement towards the target output predictions. These displacements are defined by the MLP and attention blocks, whose vector components also reside on the same hypersphere. Experiments show that nGPT learns much faster, reducing the number of training steps required to achieve the same accuracy by a factor of 4 to 20, depending on the sequence length.", "keywords": [ "transformer", "normalization", "representation" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "sdpVfWOUQA", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Lizhi Ma", "gender": "Female", "institution": "Westlake University", "country": "", "position": "Researcher" }, { "name": "Yang Yan", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Yu Lu", "gender": "Female", "institution": "Westlake University", "country": "", "position": "Postdoc" }, { "name": "Zhenzhong Lan", "gender": "unknown", "institution": "Westlake University", "country": "", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 18, "n_ref": 50, "n_ref_all": 53, "n_fig": 1, "n_tab": 4, "L_tab": 2503, "n_element_tab": 293, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1388, "n_element_tab_1": 200, "formula_len_all": 185, "formula_len_all_1": 185, "len_all": 231144, "len_all_1": 61648, "len_abs": 1165, "len_title": 118, "len_sents": 33944, "len_sents_1": 28894, "n_sents": 237, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1170, "title": "Planning with MCTS: Enhancing Problem-Solving in Large Language Models", "abs": "Despite recent advances in Large Language Models (LLMs), their ability to solve complex reasoning problems remains limited by inconsistent planning and logical flaws. We present a novel framework that significantly enhances LLMs' problem-solving capabilities by leveraging Monte Carlo Tree Search (MCTS) for plan generation. Unlike previous approaches that apply MCTS to solution search, our method uniquely integrates MCTS into the planning phase, guided by specialized LLM-powered agents that evaluate plan quality. Experiments across diverse benchmark datasets demonstrate that our approach improves problem-solving accuracy by an average of 40.59\\% compared to zero-shot Chain-of-Thought prompting. Furthermore, we show that using smaller models for MCTS planning and larger models for execution can maintain high performance while reducing computational costs. This work opens new avenues for developing more robust and efficient AI systems capable of tackling complex real-world problems, with potential applications in fields requiring advanced logical reasoning and long-term planning. Our code examples are publicly available at the Anonymous Github Repository.", "keywords": [ "Large Language Models", "Monte Carlo Tree Search", "Planning", "Chain-of-Thought" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 3, 2, 1 ], "contribution_list": [ 3, 1, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "sdLGY9Dj5r", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Bin Gu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Haozhen Zhang", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "Hualin Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Yi Chang", "gender": "Male", "institution": "Jilin University, China", "country": "CN", "position": "Full Professor" }, { "name": "Zhekai Liu", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 74, "n_formula_1": 25, "n_ref_uni": 38, "n_ref": 60, "n_ref_all": 72, "n_fig": 2, "n_tab": 11, "L_tab": 3395, "n_element_tab": 225, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1585, "n_element_tab_1": 93, "formula_len_all": 10789, "formula_len_all_1": 2464, "len_all": 157213, "len_all_1": 61948, "len_abs": 1540, "len_title": 127, "len_sents": 41069, "len_sents_1": 26958, "n_sents": 349, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1561, "title": "Collaborative Discrete-Continuous Black-Box Prompt Learning for Language Models", "abs": "Large Scale Pre-Trained Language Models (PTMs) have demonstrated unprecedented capabilities across diverse natural language processing tasks. \nAdapting such models to downstream tasks is computationally intensive and time-consuming, particularly in black-box scenarios common in Language-Model-as-a-Service (LMaaS) environments, where model parameters and gradients are inaccessible. Recently, black-box prompt learning using zeroth-order gradients has emerged as a promising approach to address these challenges by optimizing learnable continuous prompts in embedding spaces, starting with \\textit{randomly initialized discrete text prompts}. However, its reliance on randomly initialized discrete prompts limits adaptability to diverse downstream tasks or models. To address this limitation,\nthis paper introduces ZO-PoG, a novel framework that optimizes prompts through a collaborative approach, combining Policy Gradient optimization for initial discrete text prompts and Zeroth-Order optimization for continuous prompts in embedding space. By optimizing collaboratively between discrete and continuous prompts, ZO-PoG maximizes adaptability to downstream tasks, achieving superior results without direct access to the model\u2019s internal structures.\nImportantly, we establish the sub-linear convergence of ZO-PoG under mild assumptions.\nThe experiments on different datasets demonstrate significant improvements in various tasks compared to the baselines. \nOur code is available at the following anonymous URL: https://anonymous.4open.science/r/ZO-PoG-12B4.", "keywords": [ "Black-box prompt learning", "discrete optimization" ], "rating_list": [ 5, 8, 5, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "scozdyKzET", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Dongsheng Li", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Shentong Mo", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Xufang Luo", "gender": "Female", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 33, "n_ref": 117, "n_ref_all": 140, "n_fig": 2, "n_tab": 12, "L_tab": 5117, "n_element_tab": 455, "n_fig_1": 1, "n_tab_1": 10, "L_tab_1": 4879, "n_element_tab_1": 417, "formula_len_all": 999, "formula_len_all_1": 873, "len_all": 194983, "len_all_1": 64415, "len_abs": 1495, "len_title": 72, "len_sents": 47766, "len_sents_1": 25371, "n_sents": 346, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1499, "title": "pMoE: Prompting Diverse Experts Together Wins More in Visual Adaptation", "abs": "Parameter-efficient fine-tuning has demonstrated promising results across various visual adaptation tasks, such as classification and segmentation. Typically, prompt tuning techniques have harnessed knowledge from a single pre-trained model, whether from a general or a specialized medical domain. However, this approach typically overlooks the potential synergies that could arise from integrating diverse domain knowledge within the same tuning process. In this work, we propose a novel Mixture-of-Experts prompt tuning method called pMoE, which leverages the strengths of multiple expert domains through expert-specialized prompt tokens and the learnable dispatcher, effectively combining their expertise in a unified model framework. Our pMoE introduces expert-specific prompt tokens and utilizes a dynamic token dispatching mechanism at various prompt layers to optimize the contribution of each domain expert during the adaptation phase. By incorporating both domain knowledge from diverse experts, the proposed pMoE significantly enhances the model's versatility and applicability to a broad spectrum of tasks. We conduct extensive experiments across 47 adaptation tasks, including both classification and segmentation in general and medical domains. The results demonstrate that our pMoE not only achieves superior performance with a large margin of improvements but also offers an optimal trade-off between computational efficiency and adaptation effectiveness compared to existing methods.", "keywords": [ "Visual Adaptation", "Visual Representation Learning" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 2, 4, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "sceqRsa0oo", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Bundeliya Harsh Jitendrakumar", "gender": "Male", "institution": "Indian Institute of Technology, Bombay", "country": "IN", "position": "MS student" }, { "name": "Divya Jyoti Bajpai", "gender": "Male", "institution": "Indian Institute of Technology, Bombay, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "PhD student" }, { "name": "Manjesh Kumar Hanawal", "gender": "Male", "institution": "Indian Institute of Technology Bombay", "country": "IN", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 6, "n_ref": 10, "n_ref_all": 24, "n_fig": 5, "n_tab": 6, "L_tab": 1535, "n_element_tab": 182, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1107, "n_element_tab_1": 118, "formula_len_all": 855, "formula_len_all_1": 720, "len_all": 99123, "len_all_1": 61146, "len_abs": 1310, "len_title": 95, "len_sents": 30356, "len_sents_1": 27216, "n_sents": 275, "n_sents_1": 243, "venue": "ICLR2025", "decision": "withdraw", "L_title": 47, "L_abs": 1453, "title": "SPEED: Selective Prediction for Early Exit DNNs", "abs": "Inference latency and trustworthiness of Deep Neural Networks (DNNs) are the bottlenecks in deploying them in critical applications like autonomous driving. Early Exit (EE) DDNs overcome the latency issues by allowing samples to exit from intermediary layers if they attain high confidence scores on the predicted class. However, the DNNs are known to exhibit overconfidence, which can lead to many samples exiting early and render EE strategies untrustworthy. We use Selective Prediction (SP) to overcome this issue by checking the hardness of the samples rather than just relying on the confidence score alone. We propose SPEED, a novel approach that uses Deferral Classifiers (DCs) at each layer to check the hardness of samples before performing EEs. The DCs at each layer identify if a sample is hard and either differ its inference to the next layer or directly send it to an expert. Early detection of hard samples and using an expert for inference prevents the wastage of computational resources and improves trust. We also investigate the generalization capability of DCs trained on one domain when applied to other domains where target domain data is not readily available. We observe that EE aided with SP improves both accuracy and latency. Our method minimizes the risk by 50% with a speedup of $2.05\\times$ as compared to the final layer. The anonymized source code is available at https://anonymous.4open.science/r/SPEED-35DC/README.md.", "keywords": [ "Selective prediction; Early Exits" ], "rating_list": [ 3, 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "scdGzuwC9u", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Fei Shang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Feng Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "Associate Professor" }, { "name": "Shuli Zeng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Sijia Zhang", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Xiangyang Li", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 6, "n_ref_uni": 37, "n_ref": 89, "n_ref_all": 110, "n_fig": 1, "n_tab": 9, "L_tab": 2770, "n_element_tab": 402, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 741, "n_element_tab_1": 71, "formula_len_all": 590, "formula_len_all_1": 319, "len_all": 156272, "len_all_1": 78157, "len_abs": 1329, "len_title": 137, "len_sents": 53044, "len_sents_1": 36848, "n_sents": 411, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1335, "title": "A Reoptimization Framework for Mixed Integer Linear Programming with Dynamic Parameters", "abs": "Many real-world applications, such as logistics, routing, scheduling, and production planning, involve dynamic systems that require continuous updates to solutions for new Mixed Integer Linear Programming (MILP) problems. \nThese environments often require rapid responses to slight changes in parameters, with time-critical demands for solutions. While reoptimization techniques have been explored for Linear Programming (LP) and specific MILP problems, their effectiveness in general MILP is limited. In this work, we propose a two-stage reoptimization framework for efficiently identifying high-quality feasible solutions. Specifically, we first utilize the historical solving process information to predict the high confidence solving space for modified MILPs to contain high-quality solutions. Based on the prediction results, we fix a part of variables to apply the prediction intervals and use the Thompson Sampling algorithm to determine the set of variables to fix by updating the Beta distributions based on solutions obtained from the solver. Extensive experiments across nine reoptimization datasets show that our VP-OR outperforms the state-of-the-art methods, achieving higher-quality feasible solutions under strict time limits and demonstrating faster convergence with smaller primal gaps in the early stages of solving.", "keywords": [ "Mixed Integer Linear Programming; reoptimization;" ], "rating_list": [ 5, 5, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "scKAXgonmq", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Matan Rusanovsky", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Or Hirschorn", "gender": "unknown", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Shai Avidan", "gender": "Male", "institution": "Tel-Aviv University", "country": "IL", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 10, "n_ref": 25, "n_ref_all": 58, "n_fig": 12, "n_tab": 6, "L_tab": 3788, "n_element_tab": 262, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1665, "n_element_tab_1": 161, "formula_len_all": 137, "formula_len_all_1": 138, "len_all": 130517, "len_all_1": 57586, "len_abs": 1669, "len_title": 119, "len_sents": 40590, "len_sents_1": 26502, "n_sents": 300, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1760, "title": "CapeX: Category-Agnostic Pose Estimation from Textual Point Explanation", "abs": "Conventional 2D pose estimation models are constrained by their design to specific object categories. This limits their applicability to predefined objects. To overcome these limitations, category-agnostic pose estimation (CAPE) emerged as a solution. CAPE aims to facilitate keypoint localization for diverse object categories using a unified model, which can generalize from minimal annotated support images.\nRecent CAPE works have produced object poses based on arbitrary keypoint definitions annotated on a user-provided support image. Our work departs from conventional CAPE methods, which require a support image, by adopting a text-based approach instead of the support image. \nSpecifically, we use a pose-graph, where nodes represent keypoints that are described with text. This representation takes advantage of the abstraction of text descriptions and the structure imposed by the graph.\nOur approach effectively breaks symmetry, preserves structure, and improves occlusion handling.\nWe validate our novel approach using the MP-100 benchmark, a comprehensive dataset covering over 100 categories and 18,000 images. MP-100 is structured so that the evaluation categories are unseen during training, making it especially suited for CAPE. Under a 1-shot setting, our solution achieves a notable performance boost of 1.26\\%, establishing a new state-of-the-art for CAPE. Additionally, we enhance the dataset by providing text description annotations for both training and testing. We also include alternative text annotations specifically for testing the model's ability to generalize across different textual descriptions, further increasing its value for future research. Our code and dataset are publicly available at https://github.com/matanr/capex.", "keywords": [ "2D Pose Estimation", "Class-Agnostic", "Few-Shot", "Vision-Language Model" ], "rating_list": [ 8, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "scI9307PLG", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Federico Barbero", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Jacob Bamberger", "gender": "unknown", "institution": "Department of Computer Science", "country": "", "position": "PhD student" }, { "name": "Michael M. Bronstein", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Xiaowen Dong", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" } ], "n_formula": 52, "n_formula_1": 18, "n_ref_uni": 42, "n_ref": 101, "n_ref_all": 132, "n_fig": 4, "n_tab": 13, "L_tab": 3211, "n_element_tab": 295, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2153, "n_element_tab_1": 95, "formula_len_all": 4479, "formula_len_all_1": 970, "len_all": 201265, "len_all_1": 74415, "len_abs": 1276, "len_title": 102, "len_sents": 62995, "len_sents_1": 33063, "n_sents": 551, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 53, "L_abs": 1288, "title": "Bundle Neural Network for message diffusion on graphs", "abs": "The dominant paradigm for learning on graphs is message passing. Despite being a strong inductive bias, the local message passing mechanism faces challenges such as over-smoothing, over-squashing, and limited expressivity. To address these issues, we introduce Bundle Neural Networks (BuNNs), a novel graph neural network architecture that operates via *message diffusion* on *flat vector bundles* \u2014 geometrically inspired structures that assign to each node a vector space and an orthogonal map. A BuNN layer evolves node features through a diffusion-type partial differential equation, where its discrete form acts as a special case of the recently introduced Sheaf Neural Network (SNN), effectively alleviating over-smoothing. The continuous nature of message diffusion enables BuNNs to operate at larger scales, reducing over-squashing. We establish the universality of BuNNs in approximating feature transformations on infinite families of graphs with injective positional encodings, marking the first positive uniform expressivity result of its kind. We support our claims with formal analysis and synthetic experiments. Empirically, BuNNs perform strongly on heterophilic and long-range tasks, which demonstrates their robustness on a diverse range of challenging real-world tasks.", "keywords": [ "graph neural network", "sheaf neural network", "geometric deep learning", "algebraic topology", "vector bundles", "expressivity" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "sbG8qhMjkZ", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "PROMIT GHOSAL", "gender": "unknown", "institution": "University of Chicago", "country": "US", "position": "Assistant Professor" }, { "name": "Sayan Banerjee", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Associate Professor" }, { "name": "Krishnakumar Balasubramanian", "gender": "Male", "institution": "University of California, Davis", "country": "", "position": "Associate Professor" } ], "n_formula": 109, "n_formula_1": 46, "n_ref_uni": 25, "n_ref": 70, "n_ref_all": 74, "n_fig": 0, "n_tab": 3, "L_tab": 290, "n_element_tab": 3, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 11812, "formula_len_all_1": 3604, "len_all": 139787, "len_all_1": 67077, "len_abs": 1243, "len_title": 129, "len_sents": 39506, "len_sents_1": 27430, "n_sents": 381, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 81, "L_abs": 1285, "title": "Improved Finite-Particle Convergence Rates for Stein Variational Gradient Descent", "abs": "We provide finite-particle convergence rates for the Stein Variational Gradient Descent (SVGD) algorithm in the Kernelized Stein Discrepancy ($\\KSD$) and Wasserstein-2 metrics. Our key insight is the observation that the time derivative of the relative entropy between the joint density of $N$ particle locations and the $N$-fold product target measure, starting from a regular initial distribution, splits into a dominant negative part proportional to $N$ times the expected $\\KSD^2$ and a smaller positive part. This observation leads to $\\KSD$ rates of order $1/\\sqrt{N}$, in both continuous and discrete time, providing a near optimal (in the sense of matching the corresponding i.i.d. rates) double exponential improvement over the recent result by~\\cite{shi2024finite}. Under mild assumptions on the kernel and potential, these bounds also grow polynomially in the dimension $d$. By adding a bilinear component to the kernel, the above approach is used to further obtain Wasserstein-2 convergence in continuous time. For the case of `bilinear + Mat\\'ern' kernels, we derive Wasserstein-2 rates that exhibit a curse-of-dimensionality similar to the i.i.d. setting. We also obtain marginal convergence and long-time propagation of chaos results for the time-averaged particle laws.", "keywords": [ "Stein Variational Gradient Descent", "Non-asymptotic Rates", "Variational Inference" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 4, 3, 4, 4 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 4, 3, 4, 4 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "sb7qHFYwBc", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Fei Zhu", "gender": "Male", "institution": "Centre for Artificial Intelligence and Robotics Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "CN", "position": "Postdoc" }, { "name": "Longhui Wei", "gender": "Male", "institution": "Huawei echnologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Qi Tian", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "WenZhuo Liu", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Intern" } ], "n_formula": 30, "n_formula_1": 14, "n_ref_uni": 48, "n_ref": 102, "n_ref_all": 128, "n_fig": 9, "n_tab": 7, "L_tab": 1950, "n_element_tab": 150, "n_fig_1": 9, "n_tab_1": 6, "L_tab_1": 4321, "n_element_tab_1": 173, "formula_len_all": 2664, "formula_len_all_1": 1445, "len_all": 140490, "len_all_1": 63580, "len_abs": 1114, "len_title": 111, "len_sents": 34734, "len_sents_1": 25074, "n_sents": 283, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1097, "title": "C-CLIP: Multimodal Continual Learning for Vision-Language Model", "abs": "Multimodal pre-trained models like CLIP need large image-text pairs for training but often struggle with domain-specific tasks. Since retraining with specialized and historical data incurs significant memory and time costs, it is important to continually learn new domains in the open world while preserving original performance. However, current continual learning research mainly focuses on single-modal scenarios, and the evaluation criteria are insufficient without considering image-text matching performance and the forgetting of zero-shot performance. This work introduces image-caption datasets from various domains and establishes a multimodal vision-language continual learning benchmark. Then, a novel framework named C-CLIP is proposed, which not only prevents forgetting but also enhances new task learning impressively. Comprehensive experiments demonstrate that our method has strong continual learning ability across different domain image-text datasets, and has little forgetting of the original capabilities of zero-shot prediction, significantly outperforming existing methods.", "keywords": [ "Vision-Language Models", "Continual Representation Learning" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "sb1HgVDLjN", "primary_area": "optimization", "n_author": 8, "author_info_list": [ { "name": "Chao Qian", "gender": "Male", "institution": "Nanjing university", "country": "CN", "position": "Full Professor" }, { "name": "Fu Sheng", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Haopu Shang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Ke Xue", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Yaoyuan Wang", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "yaowang", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Rongxi Tan", "gender": "unknown", "institution": "Nanjing University", "country": "", "position": "Undergrad student" }, { "name": "Shenhuan Lv", "gender": "Male", "institution": "Hohai University", "country": "CN", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 12, "n_ref_uni": 76, "n_ref": 294, "n_ref_all": 335, "n_fig": 3, "n_tab": 5, "L_tab": 4195, "n_element_tab": 241, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 208, "n_element_tab_1": 29, "formula_len_all": 1870, "formula_len_all_1": 707, "len_all": 243998, "len_all_1": 70271, "len_abs": 1933, "len_title": 100, "len_sents": 67825, "len_sents_1": 31085, "n_sents": 445, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1953, "title": "Offline Model-Based Optimization by Learning to Rank", "abs": "Offline model-based optimization (MBO) aims to identify a design that maximizes a black-box function using only a fixed, pre-collected dataset of designs and their corresponding scores. This problem has garnered significant attention from both scientific and industrial domains. A common approach in offline MBO is to train a regression-based surrogate model by minimizing mean squared error (MSE) and then find the best design within this surrogate model by different optimizers (e.g., gradient ascent). However, a critical challenge is the risk of out-of-distribution errors, i.e., the surrogate model may typically overestimate the scores and mislead the optimizers into suboptimal regions. Prior works have attempted to address this issue in various ways, such as using regularization techniques and ensemble learning to enhance the robustness of the model, but it still remains. In this paper, we argue that regression models trained with MSE are not well-aligned with the primary goal of offline MBO, which is to $\\textit{select}$ promising designs rather than to predict their scores precisely. Notably, if a surrogate model can maintain the order of candidate designs based on their relative score relationships, it can produce the best designs even without precise predictions. To validate it, we conduct experiments to compare the relationship between the quality of the final designs and MSE, finding that the correlation is really very weak. In contrast, a metric that measures order-maintaining quality shows a significantly stronger correlation. Based on this observation, we propose learning a ranking-based model that leverages learning to rank techniques to prioritize promising designs based on their relative scores. We show that the generalization error on ranking loss can be well bounded. Empirical results across diverse tasks demonstrate the superior performance of our proposed ranking-based models than twenty existing methods.", "keywords": [ "Offline model-based optimization", "black-box optimization", "learning to rank", "learning to optimize" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "sahQq2sH5x", "primary_area": "applications to neuroscience & cognitive science", "n_author": 12, "author_info_list": [ { "name": "Amine M'Charrak", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Bayar Menzat", "gender": "Male", "institution": "Technische Universit\u00e4t Wien", "country": "", "position": "Postdoc" }, { "name": "Chang Qi", "gender": "Female", "institution": "Technische Universit\u00e4t Wien", "country": "", "position": "PhD student" }, { "name": "Cornelius Emde", "gender": "Not Specified", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Gaspard Oliviers", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Luca Pinchetti", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Mufeng Tang", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Oleh Lokshyn", "gender": "Male", "institution": "Technische Universit\u00e4t Wien", "country": "", "position": "PhD student" }, { "name": "Rafal Bogacz", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "Full Professor" }, { "name": "Simon Frieder", "gender": "unknown", "institution": "Department of Computer Science, University of Oxford", "country": "", "position": "PhD student" }, { "name": "Thomas Lukasiewicz", "gender": "unknown", "institution": "Institute of Logic and Computation, Technische Universit\u00e4t Wien", "country": "", "position": "Full Professor" }, { "name": "Tommaso Salvatori", "gender": "Male", "institution": "VERSES", "country": "", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 46, "n_ref": 76, "n_ref_all": 137, "n_fig": 28, "n_tab": 11, "L_tab": 5673, "n_element_tab": 701, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 962, "n_element_tab_1": 60, "formula_len_all": 699, "formula_len_all_1": 569, "len_all": 214265, "len_all_1": 74066, "len_abs": 3888, "len_title": 100, "len_sents": 69199, "len_sents_1": 35487, "n_sents": 549, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 54, "L_abs": 1248, "title": "Benchmarking Predictive Coding Networks -- Made Simple", "abs": "In this work, we tackle the problems of efficiency and scalability for predictive coding networks (PCNs) in machine learning. To do so, we propose a library that focuses on performance and simplicity, and use it to implement a large set of standard benchmarks for the community to use for their experiments. As most works in the field propose their own tasks and architectures, do not compare one against each other, and focus on small-scale tasks, a simple and fast open-source library, and a comprehensive set of benchmarks, would address all of these concerns. Then, we perform extensive tests on such benchmarks using both existing algorithms for PCNs, as well as adaptations of other methods popular in the bio-plausible deep learning community. All of this has allowed us to (i) test architectures much larger than commonly used in the literature, on more complex datasets; (ii) reach new state-of-the-art results in all of the tasks and dataset provided; (iii) clearly highlight what the current limitations of PCNs are, allowing us to state important future research directions. With the hope of galvanizing community efforts towards one of the main open problems in the field, scalability, we will release the code, tests, and benchmarks.", "keywords": [ "cognitive science", "predictive coding", "computational neuroscience" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 2, 3, 4 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "saRBktzh3q", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Chunyu Wei", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Lecturer" }, { "name": "Haozhe Lin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Yueguo Chen", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "Full Professor" }, { "name": "Yunhai Wang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 39, "n_formula_1": 19, "n_ref_uni": 49, "n_ref": 103, "n_ref_all": 138, "n_fig": 6, "n_tab": 10, "L_tab": 2326, "n_element_tab": 367, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 839, "n_element_tab_1": 128, "formula_len_all": 3052, "formula_len_all_1": 1295, "len_all": 218422, "len_all_1": 65986, "len_abs": 1375, "len_title": 127, "len_sents": 51181, "len_sents_1": 31419, "n_sents": 394, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1401, "title": "Anomaly Detection through Conditional Diffusion Probability Modeling on Graphs", "abs": "Existing Graph Neural Network-based anomaly detection methods suffer from over-smoothing issues during feature aggregation. Moreover, most existing methods are discriminative models that learn the boundaries between anomalous and normal data points, allowing malicious nodes in a dynamic adversarial environment to bypass detection boundaries. To address these issues, existing methods primarily focus on enhancing the discriminative boundary for each individual node, rather than considering the interdependencies of node anomalies from a holistic graph perspective. We propose an advanced Conditional Graph Anomaly Diffusion Model (CGADM) to model and capture the joint distribution of anomalies on the whole graph, thereby enabling generative graph anomaly detection. To avoid starting the diffusion process from a random state, CGADM introduces a prior-guided denoising diffusion probability model. To circumvent the need for iterative denoising samplings for each node on large-scale graphs, we adopt a prior confidence-aware mechanism to dynamically adjust the reverse sampling steps for each node, significantly reducing the computational burden on large-scale graphs. We conducted experiments on CGADM using standard benchmarks, and the results demonstrated excellent performance in graph anomaly detection tasks. Additional ablation studies confirmed our framework's computational advantages.", "keywords": [ "Anomaly detection", "Graph Neural Network", "Diffusion Model" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "saJkPzTmZz", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Lee Spector", "gender": "unknown", "institution": "Amherst College", "country": "", "position": "Full Professor" }, { "name": "Li Ding", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Scott Niekum", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Associate Professor" }, { "name": "Ryan A Boldi", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Undergrad student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 50, "n_ref": 81, "n_ref_all": 96, "n_fig": 10, "n_tab": 1, "L_tab": 311, "n_element_tab": 34, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 123, "formula_len_all_1": 83, "len_all": 148528, "len_all_1": 65596, "len_abs": 1282, "len_title": 60, "len_sents": 45864, "len_sents_1": 33328, "n_sents": 316, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1287, "title": "Pareto-Optimal Learning from Preferences with Hidden Context", "abs": "Ensuring AI models align with human values is essential for their safety and functionality. Reinforcement learning from human feedback (RLHF) leverages human preferences to achieve this alignment. However, when preferences are sourced from diverse populations, point estimates of reward can result in suboptimal performance or be unfair to specific groups. We propose Pareto Optimal Preference Learning (POPL), which enables pluralistic alignment by framing discrepant group preferences as objectives with potential trade-offs, aiming for policies that are Pareto-optimal on the preference dataset. POPL utilizes lexicase selection, an iterative process that selects diverse and Pareto-optimal solutions. Our theoretical and empirical evaluations demonstrate that POPL surpasses baseline methods in learning sets of reward functions and policies, effectively catering to distinct groups without access to group numbers or membership labels. We verify the performance of POPL on a stateless preference learning setting, a Minigrid RL domain, Metaworld robotics benchmarks, as well as large language model (LLM) fine-tuning. We illustrate that POPL can also serve as a foundation for techniques optimizing specific notions of group fairness, ensuring safe and equitable AI model alignment.", "keywords": [ "Preference Learning", "Lexicase Selection", "Pareto-optimality", "Hidden Context", "Pluralistic Alignment" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "saFH7zTtQs", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Gang Lu", "gender": "Male", "institution": "Department of Computer Science, University of Massachusetts at Amherst", "country": "CN", "position": "Instructor" }, { "name": "Kaiqi Fang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Mengyu Sun", "gender": "unknown", "institution": "China Telecom Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Yasen Wang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Zhilan Huang", "gender": "Female", "institution": "China Telecom", "country": "CN", "position": "Researcher" } ], "n_formula": 66, "n_formula_1": 59, "n_ref_uni": 36, "n_ref": 53, "n_ref_all": 56, "n_fig": 0, "n_tab": 3, "L_tab": 819, "n_element_tab": 84, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 705, "n_element_tab_1": 64, "formula_len_all": 5377, "formula_len_all_1": 3492, "len_all": 105014, "len_all_1": 57880, "len_abs": 1159, "len_title": 109, "len_sents": 27963, "len_sents_1": 22211, "n_sents": 237, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1178, "title": "Learning Linear Dynamical Systems with Sparse System Matrices", "abs": "Due to the tractable analysis and control, linear dynamical systems (LDSs) provide a fundamental mathematical tool for time-series data modeling in various disciplines. Particularly, many LDSs have sparse system matrices because interactions among variables are limited or only a few significant relationships exist. However, available learning algorithms for LDSs lack the ability to learn system matrices with the sparsity constraint. To address this issue, we impose sparsity-promoting priors on system matrices and explore the expectation\u2013maximization (EM) algorithm to give a maximum a posteriori (MAP) estimate of both hidden states and system matrices from noisy observations. In addition, we find that many learning algorithms based on the gradient descent method use an inappropriate derivative rule, because they neglect the inherent symmetry of noise covariance matrices. Here, we consider the derivative rule of structured matrices during the optimization process to guarantee their symmetry. Experimental results on simulation and real-world problems illustrate that the proposed algorithm significantly improves learning accuracy over classical ones.", "keywords": [ "linear dynamical systems", "state estimation", "system identification", "expectation\u2013maximization algorithm" ], "rating_list": [ 6, 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 4, 2, 2 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "sZQRUrvLn4", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Karsten Borgwardt", "gender": "unknown", "institution": "Max Planck Institute of Biochemistry", "country": "DE", "position": "Max Planck Director" }, { "name": "Paolo Pellizzoni", "gender": "unknown", "institution": "Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Till Schulz", "gender": "Male", "institution": "Max Planck Institute of Biochemistry", "country": "DE", "position": "Postdoc" } ], "n_formula": 41, "n_formula_1": 24, "n_ref_uni": 35, "n_ref": 100, "n_ref_all": 123, "n_fig": 6, "n_tab": 6, "L_tab": 1701, "n_element_tab": 69, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 462, "n_element_tab_1": 44, "formula_len_all": 5723, "formula_len_all_1": 4101, "len_all": 198723, "len_all_1": 72398, "len_abs": 2824, "len_title": 101, "len_sents": 66005, "len_sents_1": 31838, "n_sents": 596, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 53, "L_abs": 1240, "title": "Graph Neural Networks Can (Often) Count Substructures", "abs": "Message passing graph neural networks (GNNs) are known to have limited expressive power in their ability to distinguish some non-isomorphic graphs.\nBecause of this, it is well known that they are unable to detect or count arbitrary graph substructures (i.e., solving the subgraph isomorphism problem), a task that is of great importance for several types of graph-structured data. \nHowever, we observe that GNNs are in fact able to count graph patterns quite accurately across several real-world graph datasets.\nMotivated by this observation, we provide an analysis of the subgraph-counting capabilities of GNNs beyond the worst case, deriving several sufficient conditions for GNNs to be able to count subgraphs and, more importantly, to be able to sample-efficiently learn to count subgraphs. \nMoreover, we develop novel dynamic programming algorithms for solving the subgraph isomorphism problem on restricted classes of pattern and target graphs, and show that message-passing GNNs can efficiently simulate these dynamic programs. \nFinally, we empirically validate that our sufficient conditions for GNNs to count subgraphs hold on many real-world datasets, providing a theoretically-grounded explanation to our motivating observations.", "keywords": [ "graph neural networks", "subgraphs", "expressivity" ], "rating_list": [ 6, 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 4 ], "presentation_list": [ 4, 3, 3, 4, 4 ], "contribution_list": [ 3, 2, 3, 3, 4 ], "confidence_list": [ 2, 4, 4, 4, 4 ] }, { "paper_id": "sZJqKAVzKf", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Charmaine Barker", "gender": "unknown", "institution": "University of York", "country": "GB", "position": "PhD student" }, { "name": "Daniel Bethell", "gender": "Male", "institution": "University of York", "country": "GB", "position": "PhD student" }, { "name": "Dimitar Kazakov", "gender": "Male", "institution": "University of York", "country": "GB", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 30, "n_ref": 45, "n_ref_all": 86, "n_fig": 7, "n_tab": 9, "L_tab": 8095, "n_element_tab": 402, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 2, "n_element_tab_1": 1, "formula_len_all": 568, "formula_len_all_1": 370, "len_all": 139433, "len_all_1": 62966, "len_abs": 1603, "len_title": 92, "len_sents": 45657, "len_sents_1": 31385, "n_sents": 319, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 1611, "title": "Learning Fairer Representations with FairVIC", "abs": "Mitigating bias in automated decision-making systems, specifically deep learning models, is a critical challenge in achieving fairness. This complexity stems from factors such as nuanced definitions of fairness, unique biases in each dataset, and the trade-off between fairness and model accuracy. To address such issues, we introduce FairVIC, an innovative approach designed to enhance fairness in neural networks by addressing inherent biases at the training stage. Unlike other methods that require a user-defined declaration of what it means to be fair, FairVIC integrates an abstract concept of fairness through variance, invariance and covariance terms into the loss function. These terms aim to minimise the model's dependency on protected characteristics for making predictions, thus promoting fairness. Our experimentation consists of evaluating FairVIC against other comparable bias mitigation techniques, on a number of datasets known for their biases. Additionally, we conduct an ablation study to examine the accuracy-fairness trade-off. We also extend FairVIC by offering multi-objective lambda recommendations, allowing users to train a fairer model with a set of weights that are tuned best for their application. Through our implementation of FairVIC, we observed a significant improvement in fairness across all metrics tested, without compromising the model's accuracy. Our findings suggest that FairVIC presents a straightforward, out-of-the-box solution for the development of fairer deep learning models, thereby offering a generalisable solution applicable across many tasks and datasets.", "keywords": [ "Bias Mitigation", "Fairness", "Deep Learning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "sZJNkorXMk", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Fusheng Liu", "gender": "unknown", "institution": "national university of singaore, National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Qianxiao Li", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 31, "n_formula_1": 11, "n_ref_uni": 27, "n_ref": 73, "n_ref_all": 110, "n_fig": 10, "n_tab": 4, "L_tab": 602, "n_element_tab": 85, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 225, "n_element_tab_1": 26, "formula_len_all": 2528, "formula_len_all_1": 827, "len_all": 163422, "len_all_1": 72886, "len_abs": 1067, "len_title": 144, "len_sents": 55230, "len_sents_1": 32797, "n_sents": 474, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1070, "title": "Autocorrelation Matters: Understanding the Role of Initialization Schemes for State Space Models", "abs": "Current methods for initializing state space model (SSM) parameters primarily rely on the HiPPO framework \\citep{gu2023how}, which is based on online function approximation with the SSM kernel basis. \nHowever, the HiPPO framework does not explicitly account for the effects of the temporal structures of input sequences on the optimization of SSMs.\nIn this paper, we take a further step to investigate the roles of SSM initialization schemes by considering the autocorrelation of input sequences. \nSpecifically, we: (1) rigorously characterize the dependency of the SSM timescale on sequence length based on sequence autocorrelation; (2) find that with a proper timescale, allowing a zero real part for the eigenvalues of the SSM state matrix mitigates the curse of memory while still maintaining stability at initialization; (3) show that the imaginary part of the eigenvalues of the SSM state matrix determines the conditioning of SSM optimization problems, and uncover an approximation-estimation tradeoff when training SSMs with a specific class of target functions.", "keywords": [ "State Space Model; Optimization;" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 2, 2 ] }, { "paper_id": "sZGZJhaNSe", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Christopher Re", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Kyunghyun Cho", "gender": "Male", "institution": "Genentech", "country": "", "position": "Senior Director of Frontier Research" }, { "name": "Nicholas Lourie", "gender": "unknown", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Mayee Chen", "gender": "Female", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Michael Hu", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 101, "n_formula_1": 15, "n_ref_uni": 58, "n_ref": 94, "n_ref_all": 148, "n_fig": 13, "n_tab": 23, "L_tab": 4576, "n_element_tab": 524, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8252, "formula_len_all_1": 1484, "len_all": 365822, "len_all_1": 79255, "len_abs": 1737, "len_title": 118, "len_sents": 101243, "len_sents_1": 32257, "n_sents": 919, "n_sents_1": 274, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1745, "title": "Aioli: A Unified Optimization Framework for Language Model Data Mixing", "abs": "Language model performance depends on identifying the optimal mixture of data groups to train on (e.g., law, code, math). Prior work has proposed a diverse set of methods to efficiently learn mixture proportions, ranging from fitting regression models over training runs to dynamically updating proportions throughout training. Surprisingly, we find that no existing method consistently outperforms a simple stratified sampling baseline in terms of average test perplexity per group. In this paper, we study the cause of this inconsistency by unifying existing methods into a standard optimization framework. We show that all methods set proportions to minimize total loss, subject to a method-specific mixing law---an assumption on how loss is a function of mixture proportions. We find that existing parameterizations of mixing laws can express the true loss-proportion relationship empirically, but the methods themselves often set the mixing law parameters inaccurately, resulting in poor and inconsistent performance. Finally, we leverage the insights from our framework to derive a new online method named Aioli, which directly estimates the mixing law parameters throughout training and uses them to dynamically adjust proportions. Empirically, Aioli outperforms stratified sampling on 6 out of 6 datasets by an average of 0.27 test perplexity points, whereas existing methods fail to consistently beat stratified sampling, doing up to 6.9 points worse. Moreover, in a practical setting where proportions are learned on shorter runs due to computational constraints, Aioli can dynamically adjust these proportions over the full training run, consistently improving performance over existing methods by up to 12.012 test perplexity points.", "keywords": [ "data mixing", "language models", "data curation" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "sYrdb3mhM4", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Mehrtash Babadi", "gender": "Male", "institution": "Broad Institute", "country": "US", "position": "Researcher" }, { "name": "Tianyu Liu", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Tinglin Huang", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Wengong Jin", "gender": "unknown", "institution": "Northeastern University", "country": "US", "position": "Assistant Professor" }, { "name": "Rex Ying", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 17, "n_ref_uni": 46, "n_ref": 82, "n_ref_all": 96, "n_fig": 3, "n_tab": 5, "L_tab": 4652, "n_element_tab": 291, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 4351, "n_element_tab_1": 210, "formula_len_all": 1450, "formula_len_all_1": 1594, "len_all": 136992, "len_all_1": 70099, "len_abs": 1104, "len_title": 145, "len_sents": 35997, "len_sents_1": 28366, "n_sents": 298, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1104, "title": "Predicting Spatial Transcriptomics from Histology Images via Biologically Informed Flow Matching", "abs": "Spatial transcriptomics (ST) has emerged as a promising technology to bridge the gap between histology imaging and gene expression profiling. However, its application to medical diagnosis is limited due to its low throughput and the need for specialized experimental facilities. To address this issue, we develop STFlow, a flow-based generative model to predict spatial transcriptomics from whole-slide histology images. STFlow is trained with a biologically-informed flow matching algorithm that iteratively refines predicted gene expression values, where we choose zero-inflated negative binomial distribution as a prior distribution to incorporate the inductive bias of gene expression data. Compared to previous methods that predict the gene expression of each spot independently, STFlow models the interaction of genes across different spots to account for potential gene regulatory effects. On a recently curated HEST-1k benchmark, we demonstrate STFlow substantially outperforms all baselines including pathology foundation models, with over 18% relative improvement over current state-of-the-art.", "keywords": [ "Spatial Transcriptomics; Histology Images" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "sYNWqQYJhz", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Jingyi Chai", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Rui Ye", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yanfeng Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Yaodong Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Siheng Chen", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "XiangRui Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 51, "n_ref": 108, "n_ref_all": 134, "n_fig": 7, "n_tab": 6, "L_tab": 2059, "n_element_tab": 210, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1706, "n_element_tab_1": 149, "formula_len_all": 294, "formula_len_all_1": 254, "len_all": 167323, "len_all_1": 68000, "len_abs": 1447, "len_title": 140, "len_sents": 43852, "len_sents_1": 31243, "n_sents": 303, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1452, "title": "Emerging Safety Attack and Defense in Federated Instruction Tuning of Large Language Models", "abs": "Federated learning (FL) enables multiple parties to collaboratively fine-tune an large language model (LLM) without the need of direct data sharing. Ideally, by training on decentralized data that is aligned with human preferences and safety principles, federated instruction tuning can result in an LLM that could behave in a helpful and safe manner. In this paper, we for the first time reveal the vulnerability of safety alignment in FedIT by proposing a simple, stealthy, yet effective safety attack method. Specifically, the malicious clients could automatically generate attack data without involving manual efforts and attack the FedIT system by training their local LLMs on such attack data. Unfortunately, this proposed safety attack not only can compromise the safety alignment of LLM trained via FedIT, but also can not be effectively defended against by many existing FL defense methods. Targeting this, we further propose a post-hoc defense method, which could rely on an fully automated pipeline: generation of defense data and further fine-tuning of the LLM. Extensive experiments show that our safety attack method can significantly compromise the LLM's safety alignment (e.g., reduce safety rate by 70%), which can not be effectively defended by existing defense methods (at most 4% absolute improvement), while our safety defense method can significantly enhance the attacked LLM's safety alignment (at most 69% absolute improvement).", "keywords": [ "Federated Learning", "Large Language Models", "Safety Alignment", "Attack", "Defense" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 3, 3, 3 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "sYJQEgkkaI", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Chengqi Zhang", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Chair Professor" }, { "name": "Guodong Long", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Jing Jiang", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Zhi-Hong Deng", "gender": "unknown", "institution": "University of Technology Sydney", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 38, "n_ref": 54, "n_ref_all": 75, "n_fig": 6, "n_tab": 3, "L_tab": 2327, "n_element_tab": 84, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1620, "n_element_tab_1": 72, "formula_len_all": 102, "formula_len_all_1": 102, "len_all": 122895, "len_all_1": 70570, "len_abs": 1336, "len_title": 129, "len_sents": 35935, "len_sents_1": 32616, "n_sents": 232, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1339, "title": "Rethinking The Reliability of Representation Engineering in Large Language Models", "abs": "Inspired by cognitive neuroscience, representation engineering (RepE) seeks to connect the neural activities within large language models (LLMs) to their behaviors, providing a promising pathway towards transparent AI.\nDespite its successful applications under many contexts, the connection established by RepE is not always reliable, as it implicitly assumes that LLMs will consistently follow the roles assigned in the instructions during neural activities collection.\nWhen this assumption is violated, observed correlations between the collected neural activities and model behaviors may not be causal due to potential confounding biases, thereby compromising the reliability of RepE.\nWe identify this key limitation and propose CAusal Representation Engineering (CARE), a principled framework that employs matched-pair trial design to control for confounders.\nBy isolating the impact of confounders on neural activities and model behaviors, CARE grounds the connection in causality, allowing for more reliable interpretations and control of LLMs.\nExtensive empirical evaluations across various aspects of safety demonstrate the effectiveness of CARE compared to the original RepE implementation, particularly in controlling model behaviors, highlighting the importance of causality in developing transparent and trustworthy AI systems.", "keywords": [ "transparency", "interpretability", "causality", "AI safety" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 5, 3, 2, 4 ] }, { "paper_id": "sYGNCscE9M", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Haiduo Huang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Pengju Ren", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Tian Xia", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Wenzhe zhao", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhenhua Liu", "gender": "Male", "institution": "Advanced Micro Devices", "country": "CN", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 7, "n_ref_uni": 27, "n_ref": 56, "n_ref_all": 90, "n_fig": 17, "n_tab": 12, "L_tab": 6763, "n_element_tab": 365, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 4942, "n_element_tab_1": 202, "formula_len_all": 895, "formula_len_all_1": 289, "len_all": 145322, "len_all_1": 59394, "len_abs": 1821, "len_title": 86, "len_sents": 42150, "len_sents_1": 24751, "n_sents": 344, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 38, "L_abs": 1727, "title": "Nearly Lossless Adaptive Bit Switching", "abs": "Model quantization is widely applied for compressing and accelerating deep neural networks (DNNs). However, conventional Quantization-Aware Training (QAT) focuses on training DNNs with uniform bit-width. The bit-width settings vary across different hardware and transmission demands, which induces considerable training and storage costs. Hence, the scheme of one-shot joint training multiple precisions is proposed to address this issue. Previous works either store a larger FP32 model to switch between different precision models for higher accuracy or store a smaller INT8 model but compromise accuracy due to using shared quantization parameters. In this paper, we introduce the Double Rounding quantization method, which fully utilizes the quantized representation range to accomplish nearly lossless bit-switching while reducing storage by using the highest integer precision instead of full precision. Furthermore, we observe a competitive interference among different precisions during one-shot joint training, primarily due to inconsistent gradients of quantization scales during backward propagation. To tackle this problem, we propose an Adaptive Learning Rate Scaling (ALRS) technique that dynamically adapts learning rates for various precisions to optimize the training process. Additionally, we extend our Double Rounding to one-shot mixed precision training and develop a Hessian-Aware Stochastic Bit-switching (HASB) strategy. Experimental results on the ImageNet-1K classification demonstrate that our methods have enough advantages to state-of-the-art one-shot joint QAT in both multi-precision and mixed-precision. Our codes are available at https://anonymous.4open.science/r/Double-Rounding-EF78/README.md.", "keywords": [ "Deep neural networks", "Multi-precision", "Bit Switching", "Model quantization" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 4, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "sYAFiHP6qr", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Daniel Cremers", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Full Professor" }, { "name": "Dongliang Cao", "gender": "Male", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "PhD student" }, { "name": "Florian Bernard", "gender": "unknown", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "Associate Professor" }, { "name": "Lu Sang", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Zehranaz Canfes", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "MS student" } ], "n_formula": 22, "n_formula_1": 19, "n_ref_uni": 21, "n_ref": 42, "n_ref_all": 114, "n_fig": 25, "n_tab": 1, "L_tab": 181, "n_element_tab": 14, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 837, "n_element_tab_1": 25, "formula_len_all": 1085, "formula_len_all_1": 779, "len_all": 152406, "len_all_1": 60474, "len_abs": 340, "len_title": 114, "len_sents": 47781, "len_sents_1": 27427, "n_sents": 389, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 899, "title": "Implicit Neural Surface Deformation with Explicit Velocity Fields", "abs": "In this work, we introduce the first unsupervised method that simultaneously predicts time-varying neural implicit surfaces and deformations between pairs of point clouds. We propose to model the point movement using an explicit velocity field and directly deform a time-varying implicit field using the modified level-set equation. This equation utilizes an iso-surface evolution with Eikonal constraints in a compact formulation, ensuring the integrity of the signed distance field. By applying a smooth, volume-preserving constraint to the velocity field, our method successfully recovers physically plausible intermediate shapes. Our method is able to handle both rigid and non-rigid deformations without any intermediate shape supervision. Our experimental results demonstrate that our method significantly outperforms existing works, delivering superior results in both quality and efficiency.", "keywords": [ "Implicit Shape Representation", "Neural Implicit Representation", "Shape Deformation", "Neural Surfaces" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "sY3anJ8C68", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Bing Cao", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jiekang Feng", "gender": "Male", "institution": "Tianjin University", "country": "", "position": "MS student" }, { "name": "Pengfei Zhu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" }, { "name": "Qilong Wang", "gender": "unknown", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" }, { "name": "Qinghua Hu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Professor" }, { "name": "Quanhao Lu", "gender": "Male", "institution": "Tianjin University", "country": "", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 34, "n_ref": 60, "n_ref_all": 77, "n_fig": 4, "n_tab": 4, "L_tab": 1291, "n_element_tab": 250, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1066, "n_element_tab_1": 214, "formula_len_all": 537, "formula_len_all_1": 538, "len_all": 126262, "len_all_1": 70216, "len_abs": 1581, "len_title": 131, "len_sents": 38182, "len_sents_1": 31687, "n_sents": 273, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1612, "title": "Efficient Masked AutoEncoder for Video Object Counting and A Large-Scale Benchmark", "abs": "The dynamic imbalance of the fore-background is a major challenge in video object counting, which is usually caused by the sparsity of foreground objects. This often leads to severe under- and over-prediction problems and has been less studied in existing works.\nTo tackle this issue in video object counting, we propose a density-embedded Efficient Masked Autoencoder Counting (E-MAC) framework in this paper. To effectively capture the dynamic variations across frames, we utilize an optical flow-based temporal collaborative fusion that aligns features to derive multi-frame density residuals. The counting accuracy of the current frame is boosted by harnessing the information from adjacent frames. More importantly, to empower the representation ability of dynamic foreground objects for intra-frame, we first take the density map as an auxiliary modality to perform Density-Embedded Masked mOdeling (DEMO) for multimodal self-representation learning to regress density map. However, as DEMO contributes effective cross-modal regression guidance, it also brings in redundant background information and hard to focus on foreground regions. To handle this dilemma, we further propose an efficient spatial adaptive masking derived from density maps to boost efficiency. In addition, considering most existing datasets are limited to human-centric scenarios, we first propose a large video bird counting dataset $\\textit{DroneBird}$, in natural scenarios for migratory bird protection. Extensive experiments on three crowd datasets and our $\\textit{DroneBird}$ validate our superiority against the counterparts.", "keywords": [ "Video object counting", "masked autoencoder", "multimodal self-representation learning" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 4, 3, 2, 2 ], "presentation_list": [ 4, 2, 1, 3 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 5, 5, 2, 5 ] }, { "paper_id": "sXF5P4N7e8", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Colin Bellinger", "gender": "Male", "institution": "National Research Council Canada", "country": "", "position": "Researcher" }, { "name": "Fahim Shahriar", "gender": "Male", "institution": ", University of Alberta", "country": "", "position": "MS student" }, { "name": "Gautham Vasan", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "PhD student" }, { "name": "HUIYI WANG", "gender": "Female", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "Seyed Alireza Azimi", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "MS student" }, { "name": "Ashique Rupam Mahmood", "gender": "unknown", "institution": "University of Alberta", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 25, "n_ref": 34, "n_ref_all": 58, "n_fig": 6, "n_tab": 2, "L_tab": 964, "n_element_tab": 113, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 408, "n_element_tab_1": 29, "formula_len_all": 461, "formula_len_all_1": 461, "len_all": 99042, "len_all_1": 58974, "len_abs": 1429, "len_title": 102, "len_sents": 34062, "len_sents_1": 27982, "n_sents": 232, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 1441, "title": "Vision-Based Grasping through Goal-Conditioned Masking", "abs": "Goal-Conditioned Reinforcement Learning for robotic reaching and grasping has enabled agents to achieve diverse objectives with a unified policy, leveraging goal conditioning such as images, vectors, and text. The existing methods, however, carry inherent limitations; for example, vector-based one-hot encodings allow only a predetermined object set. Meanwhile, goal state images in image-based goal conditioning can be hard to obtain in the real world and may limit generalization to novel objects. This paper introduces a mask-based goal conditioning method that offers object-agnostic visual cues to promote efficient feature sharing and robust generalization. The agent receives text-based goal directives and utilizes a pre-trained object detection model to generate a mask for goal conditioning and facilitate generalization to out-of-distribution objects. In addition, we show that the mask can enhance sample efficiency by augmenting sparse rewards without needing privileged information of the target location, unlike distance-based reward shaping. The effectiveness of the proposed framework is demonstrated in a simulated reach-and-grasp task. The mask-based goal conditioning consistently maintains a $\\sim$90\\% success rate in grasping both in and out-of-distribution objects. Furthermore, the results show that the mask-augmented reward facilitates a learning speed and grasping success rate on par with distance-based reward.", "keywords": [ "Goal-Conditioned Reinforcement Learning", "Robotic Reaching and Grasping", "Masking-Based Goal Representation", "Visual Goal Recognition", "Out-of-Distribution Object Generalization" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "sWwK0lJ8dK", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Anderson C. A. Nascimento", "gender": "Male", "institution": "VISA", "country": "", "position": "Researcher" }, { "name": "Hadi Abdullah", "gender": "unknown", "institution": "VISA", "country": "US", "position": "Researcher" }, { "name": "Shan Jin", "gender": "Male", "institution": "VISA", "country": "US", "position": "Researcher" }, { "name": "Vincent Bindschaedler", "gender": "unknown", "institution": "University of Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Wenxuan Bao", "gender": "Male", "institution": "University of Florida ", "country": "US", "position": "PhD student" }, { "name": "Yiwei Cai", "gender": "Male", "institution": "VISA", "country": "", "position": "Sr. Director" } ], "n_formula": 24, "n_formula_1": 14, "n_ref_uni": 41, "n_ref": 76, "n_ref_all": 112, "n_fig": 4, "n_tab": 14, "L_tab": 2965, "n_element_tab": 429, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1248, "n_element_tab_1": 128, "formula_len_all": 2053, "formula_len_all_1": 646, "len_all": 161993, "len_all_1": 69796, "len_abs": 1251, "len_title": 88, "len_sents": 49379, "len_sents_1": 29242, "n_sents": 438, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1259, "title": "Deep Learning with Plausible Deniability", "abs": "Deep learning models are vulnerable to privacy attacks due to their tendency to memorize individual training set examples. Theoretically-sound defenses such as differential privacy can defend against this threat, but model performance often suffers. Empirical defenses may thwart existing attacks while maintaining model performance but do not offer any robust theoretical guarantees.\n\nIn this paper, we explore a new strategy based on the concept of plausible deniability. We introduce a training algorithm called Plausibly Deniable Stochastic Gradient Descent (PD-SGD), which aims to provide both strong privacy protection with theoretical justification and maintain high performance. The core of this approach is a rejection sampling technique, which probabilistically prevents updating model parameters whenever a mini-batch cannot be plausibly denied. This ensures that no individual example has a disproportionate influence on the model parameters. We provide a set of theoretical results showing that PD-SGD effectively mitigates privacy leakage from individual data points. Experiments also demonstrate that PD-SGD offers a favorable trade-off between privacy and utility compared to differential privacy (i.e., DP-SGD) and empirical defense methods.", "keywords": [ "Deep learning", "Privacy", "Plausible Deniability" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 4, 4, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "sVbFE6xz82", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Dongjin Kim", "gender": "unknown", "institution": "Hanyang University", "country": "KR", "position": "PhD student" }, { "name": "Guanghui Wang", "gender": "Male", "institution": "Toronto Metropolitan University", "country": "CA", "position": "Associate Professor" }, { "name": "Jaekyun Ko", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "Postdoc" }, { "name": "Tae Hyun Kim", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 12, "n_ref_uni": 47, "n_ref": 136, "n_ref_all": 152, "n_fig": 7, "n_tab": 15, "L_tab": 3701, "n_element_tab": 434, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 1577, "n_element_tab_1": 203, "formula_len_all": 1254, "formula_len_all_1": 625, "len_all": 170681, "len_all_1": 66411, "len_abs": 1312, "len_title": 100, "len_sents": 53553, "len_sents_1": 29150, "n_sents": 400, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1317, "title": "MFN: Metadata-Free Real-World Noisy Image Generation", "abs": "Real-world noise poses a significant challenge in signal processing, especially for denoising tasks.\nAlthough end-to-end denoising approaches have achieved exceptional performance, they are constrained to scenarios with abundant noisy-clean image pairs, which can be technically challenging and resource-intensive to collect.\nTo address this issue, several generative methods have been developed to synthesize realistic noisy images from limited real-world datasets.\nWhile prior studies require camera metadata during training or testing to handle various real-world noise, the absence of metadata or variations in the information across different capturing devices is common in real-world scenarios, such as medical or microscope imaging, which limits their applicability.\nThus, we aim to eliminate the need for explicit camera-related labels in both stages, enhancing applicability in real-world scenarios.\nTo achieve this, we propose a novel framework called the Metadata-Free Noise Model (MFN), which extracts prompt features that encode input noise characteristics and generates diverse noisy images that adhere to the distribution of the input noise.\nExtensive experimental results demonstrate the superior performance of our model in real-world noise generation and denoising across various benchmark datasets.", "keywords": [ "real noise modeling", "prompt learning", "consistency models", "low-level vision" ], "rating_list": [ 3, 8, 8, 5 ], "soundness_list": [ 1, 3, 4, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "sVNfWhtaJC", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Cong Shen", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Associate Professor" }, { "name": "Fengyu Gao", "gender": "unknown", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Jing Yang", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Associate Professor" }, { "name": "Ruida Zhou", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Tianhao Wang", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 8, "n_ref_uni": 61, "n_ref": 158, "n_ref_all": 231, "n_fig": 1, "n_tab": 14, "L_tab": 1243, "n_element_tab": 91, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1116, "n_element_tab_1": 62, "formula_len_all": 1408, "formula_len_all_1": 423, "len_all": 239274, "len_all_1": 76223, "len_abs": 4076, "len_title": 125, "len_sents": 77147, "len_sents_1": 34779, "n_sents": 616, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1354, "title": "Data-adaptive Differentially Private Prompt Synthesis for In-Context Learning", "abs": "Large Language Models (LLMs) rely on the contextual information embedded in examples/demonstrations to perform in-context learning (ICL). To mitigate the risk of LLMs potentially leaking private information contained in examples in the prompt, we introduce a novel data-adaptive differentially private algorithm called **AdaDPSyn** to generate synthetic examples from the private dataset and then use these synthetic examples to perform ICL. The objective of AdaDPSyn is to adaptively adjust the noise level in the data synthesis mechanism according to the inherent statistical properties of the data, thereby preserving high ICL accuracy while maintaining formal differential privacy guarantees. A key innovation in AdaDPSyn is the *Precision-Focused Iterative Radius Reduction* technique, which dynamically refines the aggregation radius - the scope of data grouping for noise addition - based on patterns observed in data clustering, thereby minimizing the amount of additive noise. We conduct extensive experiments on standard benchmarks and compare AdaDPSyn with DP few-shot generation algorithm (Tang et al., 2023). The experiments demonstrate that AdaDPSyn not only outperforms DP few-shot generation, but also maintains high accuracy levels close to those of non-private baselines, providing an effective solution for ICL with privacy protection.", "keywords": [ "in-context learning", "differential privacy", "large language models" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "sVBnGcbkkM", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Berat Mert Albaba", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Christoph Gebhardt", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Postdoc" }, { "name": "Otmar Hilliges", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "Full Professor" }, { "name": "Thomas Langerak", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Postdoc" }, { "name": "Michael J Black", "gender": "Male", "institution": "University of T\u00fcbingen", "country": "DE", "position": "Honorarprofessor" }, { "name": "Sammy Joe Christen", "gender": "Unspecified", "institution": "Disney Research, Disney Research", "country": "CH", "position": "Postdoc" } ], "n_formula": 29, "n_formula_1": 11, "n_ref_uni": 30, "n_ref": 68, "n_ref_all": 100, "n_fig": 18, "n_tab": 9, "L_tab": 3131, "n_element_tab": 299, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 885, "n_element_tab_1": 11, "formula_len_all": 1746, "formula_len_all_1": 698, "len_all": 165750, "len_all_1": 61084, "len_abs": 3618, "len_title": 83, "len_sents": 56891, "len_sents_1": 28816, "n_sents": 427, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1646, "title": "RILe: Reinforced Imitation Learning", "abs": "Reinforcement Learning has achieved significant success in generating complex behavior but often requires extensive reward function engineering. Adversarial variants of Imitation Learning and Inverse Reinforcement Learning offer an alternative by learning policies from expert demonstrations via a discriminator. However, these methods struggle in complex tasks where randomly sampling expert-like behaviors is challenging. This limitation stems from their reliance on policy-agnostic discriminators, which provide insufficient guidance for agent improvement, especially as task complexity increases and expert behavior becomes more distinct. We introduce RILe (Reinforced Imitation Learning environment), a novel trainer-student system that learns a dynamic reward function based on the student's performance and alignment with expert demonstrations. In RILe, the student learns an action policy while the trainer, using reinforcement learning, continuously updates itself via the discriminator's feedback to optimize the alignment between the student and the expert. The trainer optimizes for long-term cumulative rewards from the discriminator, enabling it to provide nuanced feedback that accounts for the complexity of the task and the student's current capabilities. This approach allows for greater exploration of agent actions by providing graduated feedback rather than binary expert/non-expert classifications. By reducing dependence on policy-agnostic discriminators, RILe enables better performance in complex settings where traditional methods falter, outperforming existing methods by 2x in complex simulated robot-locomotion tasks.", "keywords": [ "Imitation Learning", "Inverse Reinforcement Learning", "Reinforcement Learning" ], "rating_list": [ 5, 6, 3, 8 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "sUywd7UhFT", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Chun Yuan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Haolong Qian", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Shuting Dong", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yu Hao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 16, "n_ref": 25, "n_ref_all": 31, "n_fig": 2, "n_tab": 7, "L_tab": 2522, "n_element_tab": 290, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2380, "n_element_tab_1": 270, "formula_len_all": 225, "formula_len_all_1": 225, "len_all": 90330, "len_all_1": 57044, "len_abs": 1531, "len_title": 129, "len_sents": 34732, "len_sents_1": 25646, "n_sents": 240, "n_sents_1": 163, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1544, "title": "Unifying All Species: LLM-based Hyper-Heuristics for Multi-objective Optimization", "abs": "Optimization problems are fundamental across various fields, including logistics, machine learning, and bioinformatics, where challenges are often characterized by complexity, high dimensionality. Modeling the interplay among multiple objectives is beneficial for optimization. However, existing Neural Combinatorial Optimization (NCO) methods and Large Language Model (LLM)-based approaches show limitations in adaptability and computational efficiency, primarily focusing on single-objective optimization. In this paper, we propose a novel framework, Multi-Objective Hierarchical Reflective Evolution (MHRE), for optimizing and generating heuristics algorithms for a broad range of optimization problems. Specifically, we extend the optimization space of the conventional hyper-heuristic methodologies, which allows us to unify similarity algorithms. We successfully construct Generalized Evolutionary Metaheuristic Algorithm (GEMA) for unifying metaheuristic algorithms. Yielding improved performance in experimental results. To show the performance of our method, we further applied the MHRE framework to optimize the Ant Colony Optimization (ACO) algorithm, achieving state-of-the-art results on random TSP problems and the TSPLib benchmark datasets. Our findings illustrate that the MLHH framework offers a robust and innovative solution for tackling complex optimization challenges, paving the way for future research in this area.\nFor better reproducibility, we open source the code at \\url{https://anonymous.4open.science/r/MHRE-BB53}.", "keywords": [ "Multi-Objective Optimization", "Hierarchical Reflective Evolution", "Ant Colony Optimization", "Metaheuristic Algorithms", "Combinatorial Optimization" ], "rating_list": [ 3, 1, 5, 1 ], "soundness_list": [ 1, 1, 3, 1 ], "presentation_list": [ 1, 1, 2, 2 ], "contribution_list": [ 2, 1, 3, 1 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "sUvBTEYXGt", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Anwesa Choudhuri", "gender": "Female", "institution": "United Imaging Intelligence", "country": "US", "position": "Researcher" }, { "name": "Benjamin Planche", "gender": "Male", "institution": "United Imaging Intelligence", "country": "US", "position": "Researcher" }, { "name": "Meng Zheng", "gender": "Female", "institution": "UII America, Inc.", "country": "US", "position": "Researcher" }, { "name": "Terrence Chen", "gender": "Male", "institution": "United Imaging Intelligence", "country": "US", "position": "CEO" }, { "name": "Zhongpai Gao", "gender": "Male", "institution": "United Imaging Intelligence", "country": "US", "position": "Researcher" }, { "name": "Ziyan Wu", "gender": "Male", "institution": "United Imaging Intelligence", "country": "US", "position": "Principal Expert Scientist" } ], "n_formula": 15, "n_formula_1": 10, "n_ref_uni": 26, "n_ref": 50, "n_ref_all": 65, "n_fig": 3, "n_tab": 8, "L_tab": 4493, "n_element_tab": 491, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 531, "n_element_tab_1": 72, "formula_len_all": 703, "formula_len_all_1": 482, "len_all": 118870, "len_all_1": 54142, "len_abs": 449, "len_title": 122, "len_sents": 37591, "len_sents_1": 25661, "n_sents": 257, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1100, "title": "6DGS: Enhanced Direction-Aware Gaussian Splatting for Volumetric Rendering", "abs": "Novel view synthesis has advanced significantly with the development of neural radiance fields (NeRF) and 3D Gaussian splatting (3DGS). However, achieving high quality without compromising real-time rendering remains challenging, particularly for physically-based ray tracing with view-dependent effects. Recently, N-dimensional Gaussians (N-DG) introduced a 6D spatial-angular representation to better incorporate view-dependent effects, but the Gaussian representation and control scheme are sub-optimal. In this paper, we revisit 6D Gaussians and introduce 6D Gaussian Splatting (6DGS), which enhances color and opacity representations and leverages the additional directional information in the 6D space for optimized Gaussian control. Our approach is fully compatible with the 3DGS framework and significantly improves real-time radiance field rendering by better modeling view-dependent effects and fine details. Experiments demonstrate that 6DGS significantly outperforms 3DGS and N-DG, achieving up to a 15.73 dB improvement in PSNR with a reduction of 66.5% Gaussian points compared to 3DGS.", "keywords": [ "3D Gaussian splatting", "6D Gaussian splatting", "volumetric rendering" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "sULAwlAWc1", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Daojing He", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Linbao Li", "gender": "unknown", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "MS student" }, { "name": "YU LI", "gender": "Not Specified", "institution": "Harbin Institute of Technology (Shen Zhen)", "country": "CN", "position": "Assistant Professor" }, { "name": "Yannan Liu", "gender": "Male", "institution": "Wuheng Lab, ByteDance", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 36, "n_ref": 92, "n_ref_all": 117, "n_fig": 8, "n_tab": 8, "L_tab": 2264, "n_element_tab": 299, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1188, "n_element_tab_1": 199, "formula_len_all": 45, "formula_len_all_1": 45, "len_all": 155763, "len_all_1": 68496, "len_abs": 1347, "len_title": 127, "len_sents": 48583, "len_sents_1": 31961, "n_sents": 387, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1353, "title": "One Model Transfer to All: On Robust Jailbreak Prompts Generation against LLMs", "abs": "Safety alignment in large language models (LLMs) is increasingly compromised by jailbreak attacks, which can manipulate these models to generate harmful or unintended content. Investigating these attacks is crucial for uncovering model vulnerabilities. However, many existing jailbreak strategies fail to keep pace with the rapid development of defense mechanisms, such as defensive suffixes, rendering them ineffective against defended models. To tackle this issue, we introduce a novel attack method called ArrAttack, specifically designed to target defended LLMs. ArrAttack automatically generates robust jailbreak prompts capable of bypassing various defense measures. This capability is supported by a universal robustness judgment model that, once trained, can perform robustness evaluation for any target model with a wide variety of defenses. By leveraging this model, we can rapidly develop a robust jailbreak prompt generator that efficiently converts malicious input prompts into effective attacks. Extensive evaluations reveal that ArrAttack significantly outperforms existing attack strategies, demonstrating strong transferability across both white-box and black-box models, including GPT-4 and Claude-3. Our work bridges the gap between jailbreak attacks and defenses, providing a fresh perspective on generating robust jailbreak prompts.", "keywords": [ "large language model", "jailbreak attack", "robustness" ], "rating_list": [ 8, 5, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 2 ] }, { "paper_id": "sTllbUNLz0", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Fei Fang", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Nicholay Topin", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Stephanie Milani", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zhicheng Zhang", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Lirong Xia", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 55, "n_ref": 62, "n_ref_all": 72, "n_fig": 2, "n_tab": 4, "L_tab": 1092, "n_element_tab": 63, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1092, "n_element_tab_1": 63, "formula_len_all": 620, "formula_len_all_1": 620, "len_all": 125771, "len_all_1": 70012, "len_abs": 1185, "len_title": 140, "len_sents": 33594, "len_sents_1": 31351, "n_sents": 270, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1190, "title": "Incorporating Human Preferences into Interpretable Reinforcement Learning with Tree Policies", "abs": "Interpretable reinforcement learning (RL) seeks to create agents that are efficient, transparent, and understandable to the populations that they impact. A significant gap in current approaches is the underutilization of human feedback, which is typically employed only for post-hoc evaluation. We propose to center the needs of end users by incorporating the feedback that would be obtained in a user study directly into the training of interpretable RL algorithms. Our approach involves preference learning, where we learn preferences over high-level features that are not directly optimizable during the RL training process. We introduce an evolutionary algorithm that leverages user feedback to guide training toward interpretable decision-tree policies that are better-aligned with human preferences. We demonstrate the effectiveness of our method through experiments using synthetic preference data. Our results show an improvement in preference alignment compared to baselines, yielding policies that are more aligned with underlying user preferences but does so with sample efficiency in the number of user queries, thereby decreasing the burden on the user in providing such data.", "keywords": [ "interpretable reinforcement learning", "explainable reinforcement learning", "preference learning", "alignment" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "sTQC4TeYo1", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Domenico Amato", "gender": "Male", "institution": "University of Palermo", "country": "", "position": "PhD student" }, { "name": "Filippo Vella", "gender": "unknown", "institution": "National Research Council of Italy", "country": "IT", "position": "Researcher" }, { "name": "Giosue' Lo Bosco", "gender": "Male", "institution": "University of Palermo", "country": "", "position": "Associate Professor" }, { "name": "Riccardo Rizzo", "gender": "Male", "institution": "Consiglio Nazionale delle Ricerche", "country": "IT", "position": "Researcher" }, { "name": "Salvatore Calderaro", "gender": "Male", "institution": "University of Palermo", "country": "IT", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 17, "n_ref": 32, "n_ref_all": 48, "n_fig": 6, "n_tab": 1, "L_tab": 762, "n_element_tab": 11, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 77, "n_element_tab_1": 2, "formula_len_all": 737, "formula_len_all_1": 593, "len_all": 87734, "len_all_1": 60790, "len_abs": 1166, "len_title": 105, "len_sents": 33724, "len_sents_1": 29582, "n_sents": 262, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 3318, "title": "The GECo algorithm for Graph Neural Networks Explanation", "abs": "Graph Neural Networks (GNNs) are powerful models that manage complex data sources and their interconnection links. One of GNNs' main drawbacks is their lack of interpretability, which limits their applicability in sensitive cases. In this paper, we introduce a new methodology involving graph communities to address the interpretability of graph classification problems. The proposal, called GECo (Graph Explanation by COmmunities), exploits the idea that a community, i.e., a subset of graph nodes densely connected, should play a crucial role in graph classification. This assumption is reasonable considering the message-passing mechanism, the core of GNNs. GECo analyzes the contribution to the classification result of the community graphs, building a mask that highlights graph-relevant structures. It first uses the trained GNN one wants to explain to classify the entire graph. Then, it detects the different communities; for each community, a smaller subgraph, including the community nodes\u2019 is created, and the trained GNN is run to see how likely the subgraph alone supports the predicted class. After evaluating all the subgraph communities, an average probability is calculated and set as a threshold. Finally, any subgraph community with a probability value higher than the threshold is assessed as necessary for the model's decision. The collection of these key communities is the basis for the final explanation since they allow the highlighting of the most relevant parts of the graph leading to the classification. GECo has been tested on GNN employing Graph Convolutional Networks layers, using six artificial and four real-world graph datasets. The six synthetic datasets were generated by adding some artificial motifs (e.g., house, cycle, etc.) to Erdos-Renyi and Barabasi-Albert graphs. The real-world datasets contain molecule structures. Both categories of datasets are adopted in the experimental part of the state-of-the-art proposals for graph explainability. GECo has been compared with a random baseline explainer and four state-of-the-art approaches: PGExplainer, PGMExplainer, GNNExplainer, and SubgraphX. We chose these methods for their different strengths, specifically PGExplainer for its efficiency and generalization capability through a learned explanation model, PGMExplainer for its probabilistic approach based on causal graphs, GNNExplainer for its detailed subgraph and feature-level explanations, and SubgraphX for its theoretically grounded subgraph selection by Shapley values. These choices ensure a comprehensive evaluation of our approach against a wide range of robust techniques. We assessed GECo's performance using four evaluation criteria that leverage predicted and ground-truth explanations and use user-controlled parameters, such as the probability distribution obtained by the GNN. The results obtained by GECo consistently outperform state-of-the-art techniques across multiple metrics for synthetic and most real-world datasets. In addition, GECo is significantly faster than its competitors in terms of computational efficiency, making it an ideal solution for large-scale data analysis and practical applications. These strengths solidify GECo\u2019s role in generating accurate, efficient, and interpretable explanations in graph-based classification tasks.", "keywords": [ "Graph Neural Networks", "Interpretability", "Explainability" ], "rating_list": [ 3, 1, 3, 1 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "sTI75sFQkn", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Hao Zheng", "gender": "Male", "institution": "University of Louisiana at Lafeyette", "country": "US", "position": "Assistant Professor" }, { "name": "Hongming Li", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Senior Research Investigator" }, { "name": "Jintai Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Tingting Chen", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "Postdoc" }, { "name": "Yong Fan", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 40, "n_ref": 72, "n_ref_all": 102, "n_fig": 5, "n_tab": 7, "L_tab": 3798, "n_element_tab": 350, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1233, "n_element_tab_1": 207, "formula_len_all": 863, "formula_len_all_1": 639, "len_all": 154303, "len_all_1": 70018, "len_abs": 1596, "len_title": 83, "len_sents": 46536, "len_sents_1": 33645, "n_sents": 341, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1599, "title": "dFCExpert: Learning Dynamic Functional Connectivity Patterns with Modularity and State Experts", "abs": "Modeling brain dynamic functional connectivity (dFC) patterns from functional Magnetic Resonance Imaging (fMRI) data is of paramount importance in neuroscience and medicine. Recently, many graph neural networks (GNN) models in conjunction with transformers or recurrent neural networks (RNNs) have been proposed and shown great potential for modeling dFC patterns in terms of pattern recognition and prediction performance. Although fruitful, several issues still hinder further performance improvement of these methods, such as neglecting the intrinsic brain modularity mechanism, and the interpretable state information of dFC patterns. To tackle these limitations, we propose dFCExpert to learn effective representations of dFC patterns in fMRI data with modularity experts and state experts. Particularly, using the GNN and mixture of experts (MoE), the modularity experts characterize the brain modularity organization in the graph learning process by optimizing multiple experts, with each expert capturing brain nodes with similar functions (in the same neurocognitive module); and the state experts aggregate temporal dFC features into a set of distinctive connectivity states by a soft prototype clustering methods, where the states can support different brain activities or are affected differently by brain disorders, thus revealing insights for interpretability. Experiments on two large-scale fMRI datasets demonstrate the superiority of our method over known alternatives, and the learned dFC representations show improved explainability and hold promise to improve clinical diagnosis.", "keywords": [ "fMRI", "dynamic brain functional connectome learning", "brain modularity", "state patterns", "mixture of experts" ], "rating_list": [ 3, 6, 3, 1 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 2, 3, 1, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "sSWiZr8QU7", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Aayushya Agarwal", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Lawrence Pileggi", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Yihan Ruan", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "MS student" } ], "n_formula": 19, "n_formula_1": 20, "n_ref_uni": 13, "n_ref": 17, "n_ref_all": 39, "n_fig": 8, "n_tab": 2, "L_tab": 1651, "n_element_tab": 21, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 3208, "n_element_tab_1": 71, "formula_len_all": 1266, "formula_len_all_1": 1380, "len_all": 109606, "len_all_1": 60172, "len_abs": 2161, "len_title": 48, "len_sents": 32805, "len_sents_1": 26489, "n_sents": 210, "n_sents_1": 168, "venue": "ICLR2025", "decision": "withdraw", "L_title": 48, "L_abs": 2169, "title": "A Hybrid Simulation of DNN-based Gray Box Models", "abs": "Simulation is vital for scientific and engineering disciplines, as it enables the prediction and design of physical systems. However, the computational challenges inherent to large-scale simulations often arise from complex device models featuring high degrees of nonlinearities or hidden physical behaviors not captured by first principles. Gray-box models that combine deep neural networks (DNNs) with physics-based models have been proposed to address the computational challenges in modeling complex physical systems. A well-crafted gray box model capitalizes on the interpretability and accuracy of a physical model while incorporating deep neural networks to capture hidden physical behaviors and mitigate computational load associated with highly nonlinear components. Previously, gray box models have been constructed by defining an explicit combination of physics-based and black-box models to represent the behavior of sub-systems; however this alone cannot represent the coupled interactions that define the behavior of the entire physical system. We, therefore, explore an implicit gray box model, where both DNNs (trained on measurement and simulated data) and physical equations share a common set of state-variables. While this approach captures coupled interactions at the boundary of data-driven and physics-based models, simulating the implicit gray box model remains an open-ended problem. In this work, we introduce a new hybrid simulation that directly integrates DNNs into the numerical solvers of simulation engines to fully simulate implicit gray box models of large physical systems. This is accomplished by backpropagating through the DNN to calculate specific Jacobian values during each iteration of the numerical method. The hybrid simulation of implicit gray-box models improves the accuracy and runtime compared to full physics-based simulation and enables reusable DNN models with lower data requirements for training. For demonstration, we explore the advantages of this approach as compared to physics-based, black box, and other gray box methods for simulating the steady-state and electromagnetic transient behavior of power systems.", "keywords": [ "gray box modeling", "simulation", "neural networks" ], "rating_list": [ 3, 5, 5, 1, 6 ], "soundness_list": [ 2, 2, 3, 1, 2 ], "presentation_list": [ 2, 2, 2, 1, 2 ], "contribution_list": [ 1, 2, 2, 1, 2 ], "confidence_list": [ 3, 2, 4, 4, 3 ] }, { "paper_id": "sRrHy0wetR", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Hen-Hsen Huang", "gender": "Male", "institution": "Institute of Information Science, Academia Sinica", "country": "", "position": "Assistant Research Fellow" }, { "name": "Yu-Chuan Chen", "gender": "Male", "institution": "Academia Sinica", "country": "TW", "position": "Research assistant" } ], "n_formula": 39, "n_formula_1": 16, "n_ref_uni": 52, "n_ref": 80, "n_ref_all": 157, "n_fig": 18, "n_tab": 10, "L_tab": 4249, "n_element_tab": 237, "n_fig_1": 14, "n_tab_1": 4, "L_tab_1": 2143, "n_element_tab_1": 124, "formula_len_all": 6912, "formula_len_all_1": 1174, "len_all": 289978, "len_all_1": 81212, "len_abs": 1070, "len_title": 127, "len_sents": 89486, "len_sents_1": 35049, "n_sents": 798, "n_sents_1": 303, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1075, "title": "On Re-Encoding Short-Term Memory of Large Language Models in Conversations", "abs": "Large language models (LLMs), such as GPT-4, are adept at generating coherent and fluent responses within conversational contexts. \nHowever, there has been a paucity of comprehensive research exploring LLMs to dynamically update their knowledge in response to corrections of misinformation provided by users during dialogue sessions. \nIn this paper, we present a novel framework termed Knowledge Editing In Conversation (KEIC), along with an accompanying dataset, devised to assess the efficacy of LLMs in aligning the user update in an in-context setting, given the previous chat history containing a false statement that conflicts with the subsequent user update.\nThrough in-depth investigations, we observe that the contemporary LLMs exhibit a modicum of proficiency in this task.\nTo enhance their in-context knowledge editing abilities, we propose a structured strategy to handle the information update for LLMs in a multi-turn conversation.\nWe demonstrate that our approach is effective and suggest insights for research communities in this emerging and essential issue.", "keywords": [ "LLM", "misinformation correction", "zero-shot self-correction" ], "rating_list": [ 5, 3, 3, 6, 3 ], "soundness_list": [ 3, 1, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3, 1 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 2, 3, 4, 4, 4 ] }, { "paper_id": "sRdByVBvZq", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Akshit Sinha", "gender": "Male", "institution": "International Institute of Information Technology, Hyderabad", "country": "IN", "position": "Undergrad student" }, { "name": "Arvindh A", "gender": "Male", "institution": "ELLIS, University of Stuttgart", "country": "DE", "position": "PhD student" }, { "name": "Debangan Mishra", "gender": "Male", "institution": "International Institute of Information Technology, Hyderabad, International Institute of Information Technology Hyderabad", "country": "", "position": "Undergrad student" }, { "name": "Ponnurangam Kumaraguru", "gender": "unknown", "institution": "International Institute of Information Technology Hyderabad ", "country": "", "position": "Full Professor" }, { "name": "Shashwat Goel", "gender": "Male", "institution": "ELLIS, Max Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Sumit Kumar", "gender": "Male", "institution": "International Institute of Information Technology, Hyderabad, International Institute of Information Technology Hyderabad", "country": "IN", "position": "MS student" }, { "name": "Varshita Kolipaka", "gender": "Female", "institution": "International Institute of Information Technology, Hyderabad, International Institute of Information Technology Hyderabad", "country": "IN", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 52, "n_ref": 89, "n_ref_all": 119, "n_fig": 8, "n_tab": 6, "L_tab": 1898, "n_element_tab": 211, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 487, "formula_len_all_1": 177, "len_all": 162251, "len_all_1": 63031, "len_abs": 1082, "len_title": 115, "len_sents": 47556, "len_sents_1": 32285, "n_sents": 335, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1098, "title": "A shot of Cognac to forget bad memories: Corrective Unlearning in GNNs", "abs": "Graph Neural Networks (GNNs) are increasingly being used for a variety of ML applications on graph data. As graph data does not follow the independently and identically distributed (i.i.d) assumption, adversarial manipulations or incorrect data can propagate to other datapoints through message passing, deteriorating the model's performance. To allow model developers to remove the adverse effects of manipulated entities from a trained GNN, we study the recently formulated problem of _Corrective Unlearning_. We find that current graph unlearning methods fail to unlearn the effect of manipulations even when the whole manipulated set is known. We introduce a new graph unlearning method, ***Cognac***, which can unlearn the effect of the manipulation set even when only $5$\\% of it is identified. It recovers most of the performance of a strong oracle with fully corrected training data, even beating retraining from scratch without the deletion set while being 8x more efficient. We hope our work guides GNN developers in fixing harmful effects due to issues in real-world data post-training.", "keywords": [ "Unlearning", "Graphs", "Graph Neural Networks", "GNN", "Attacks", "Manipulations", "Removal" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "sRb9sddBy2", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Andreas Loukas", "gender": "Male", "institution": "Isomorphic labs", "country": "CH", "position": "Principal Researcher" }, { "name": "Edward Wagstaff", "gender": "unknown", "institution": "Roche", "country": "", "position": "Researcher" }, { "name": "Karolis Martinkus", "gender": "Male", "institution": "Prescient Design / Genentech / Roche", "country": "CH", "position": "Researcher" }, { "name": "Kyunghyun Cho", "gender": "Male", "institution": "Genentech", "country": "", "position": "Senior Director of Frontier Research" } ], "n_formula": 53, "n_formula_1": 13, "n_ref_uni": 59, "n_ref": 102, "n_ref_all": 121, "n_fig": 6, "n_tab": 10, "L_tab": 6557, "n_element_tab": 373, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1860, "n_element_tab_1": 116, "formula_len_all": 5416, "formula_len_all_1": 908, "len_all": 243619, "len_all_1": 69259, "len_abs": 1099, "len_title": 134, "len_sents": 75364, "len_sents_1": 31847, "n_sents": 599, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1132, "title": "Generalizing to any diverse distribution: uniformity, gentle finetuning & rebalancing", "abs": "As training datasets grow larger, we aspire to develop models that generalize well to any diverse test distribution, even if the latter deviates significantly from the training data. Various approaches like domain adaptation, domain generalization, and robust optimization attempt to address the out-of-distribution challenge by posing assumptions about the relation between training and test distribution. Differently, we adopt a more conservative perspective by accounting for the worst-case error across all sufficiently diverse test distributions within a known domain. Our first finding is that training on a uniform distribution over this domain is optimal. We also interrogate practical remedies when uniform samples are unavailable by considering methods for mitigating non-uniformity through finetuning and rebalancing. Our theory provides a mathematical grounding for previous observations on the role of entropy and rebalancing for o.o.d. generalization and foundation model training. We also provide new empirical evidence across tasks involving o.o.d. shifts which illustrate the broad applicability of our perspective.", "keywords": [ "out-of-distribution generalization", "ood generalization", "diversity", "rebalancing", "finetuning" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 5, 3, 5 ] }, { "paper_id": "sRaAt9OOnW", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Evgeny Burnaev", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Full Professor" }, { "name": "Maksim Nekrashevich", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology ", "country": "", "position": "MS student" }, { "name": "Petr Mokrov", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Xavier Aramayo Carrasco", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Alexander Korotin", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 9, "n_ref_uni": 61, "n_ref": 126, "n_ref_all": 163, "n_fig": 17, "n_tab": 4, "L_tab": 773, "n_element_tab": 94, "n_fig_1": 14, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1893, "formula_len_all_1": 598, "len_all": 192856, "len_all_1": 75971, "len_abs": 1732, "len_title": 122, "len_sents": 57008, "len_sents_1": 34462, "n_sents": 511, "n_sents_1": 293, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1734, "title": "Uncovering Challenges of Solving the Continuous Gromov-Wasserstein Problem", "abs": "Recently, the Gromov-Wasserstein Optimal Transport (GWOT) problem has attracted the special attention of the ML community. In this problem, given two distributions supported on two (possibly different) spaces, one has to find the most isometric map between them. In the discrete variant of GWOT, the task is to learn an assignment between given discrete sets of points. In the more advanced continuous formulation, one aims at recovering a parametric mapping between unknown continuous distributions based on i.i.d. samples derived from them. The clear geometrical intuition behind the GWOT makes it a natural choice for several practical use cases, giving rise to a number of proposed solvers. Some of them claim to solve the continuous version of the problem. At the same time, GWOT is notoriously hard, both theoretically and numerically. Moreover, all existing continuous GWOT solvers still heavily rely on discrete techniques. Natural questions arise: to what extent existing methods unravel GWOT problem, what difficulties they encounter, and under which conditions they are successful. Our benchmark paper is an attempt to answer these questions. We specifically focus on the continuous GWOT as the most interesting and debatable setup. We crash-test existing continuous GWOT approaches on different scenarios, carefully record and analyze the obtained results, and identify issues. Our findings experimentally testify that the scientific community is still missing a reliable continuous GWOT solver, which necessitates further research efforts. As the first step in this direction, we propose a new continuous GWOT method which does not rely on discrete techniques and partially solves some of the problems of the competitors.", "keywords": [ "Optimal Transport", "Gromov-Wasserstein", "Generative Modelling", "Benchmark" ], "rating_list": [ 6, 8, 6, 5, 6 ], "soundness_list": [ 3, 4, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "sRIU6k2TcU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 15, "author_info_list": [ { "name": "Cheng Qian", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Fangming Liu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Guirong Chen", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Huadong Wang", "gender": "Male", "institution": "ModelBest Inc.", "country": "CN", "position": "Researcher" }, { "name": "Maosong Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Qinyu Luo", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "MS student" }, { "name": "Shenzhi Yang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Undergrad student" }, { "name": "Weiwen Liu", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xin Cong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Yankai Lin", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "Assistant Professor" }, { "name": "Yasheng Wang", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "", "position": "Researcher" }, { "name": "Yaxi Lu", "gender": "unknown", "institution": "Department of Computer Science and Technology, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yesai Wu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Intern" }, { "name": "Zhiyuan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhong Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 22, "n_ref": 30, "n_ref_all": 49, "n_fig": 4, "n_tab": 2, "L_tab": 11, "n_element_tab": 2, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 983, "n_element_tab_1": 142, "formula_len_all": 158, "formula_len_all_1": 75, "len_all": 183910, "len_all_1": 67685, "len_abs": 2993, "len_title": 129, "len_sents": 44373, "len_sents_1": 31187, "n_sents": 381, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1403, "title": "Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance", "abs": "Agents powered by large language models have shown remarkable abilities in solving complex tasks. However, most agent systems remain reactive, limiting their effectiveness in scenarios requiring foresight and autonomous decision-making. In this paper, we tackle the challenge of developing proactive agents capable of anticipating and initiating tasks without explicit human instructions. We propose a novel data-driven approach for this problem. Firstly, we collect real-world human activities to generate proactive task predictions. These predictions are then labeled by human annotators as either accepted or rejected. The labeled data is used to train a reward model that simulates human judgment and serves as an automatic evaluator of the proactiveness of LLM agents. Building on this, we develop a comprehensive data generation pipeline to create a diverse dataset, ProactiveBench, containing 6,790 events. Finally, we demonstrate that fine-tuning models with the proposed ProactiveBench can significantly elicit the proactiveness of LLM agents. Experimental results show that our fine-tuned model achieves an F1-Score of 66.47% in proactively offering assistance, outperforming all open-source and close-source models. These results highlight the potential of our method in creating more proactive and effective agent systems, paving the way for future advancements in human-agent collaboration.", "keywords": [ "Human-Centered NLP", "Dialogue and Interactive Systems", "Resources and Evaluation" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "sR0xz6ZaH7", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Jiwen Lu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Kurt Keutzer", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Wei Zhan", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Wenzhao Zheng", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Xin Fei", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yueqi Duan", "gender": "Male", "institution": "Department of Electronic Engineering, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Masayoshi Tomizuka", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 40, "n_ref": 73, "n_ref_all": 82, "n_fig": 3, "n_tab": 5, "L_tab": 1359, "n_element_tab": 206, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1536, "n_element_tab_1": 142, "formula_len_all": 920, "formula_len_all_1": 710, "len_all": 127895, "len_all_1": 54197, "len_abs": 1388, "len_title": 0, "len_sents": 28810, "len_sents_1": 23924, "n_sents": 202, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1344, "title": "PixelGaussian: Generalizable 3D Gaussian Reconstruction from Arbitrary Views", "abs": "We propose AdaptiveGaussian, an efficient feed-forward framework for learning generalizable 3D Gaussian reconstruction from arbitrary views. Most existing methods rely on uniform pixel-wise Gaussian representations, which learn a fixed number of 3D Gaussians for each view and cannot generalize well to more input views. Differently, our AdaptiveGaussian dynamically adapts both the Gaussian distribution and quantity based on geometric complexity, leading to more efficient representations and significant improvements in reconstruction quality. Specifically, we introduce a Cascade Gaussian Adapter (CGA) to adjust Gaussian distribution according to local geometry complexity identified by a keypoint scorer. CGA leverages deformable attention in context-aware hypernetworks to guide Gaussian pruning and splitting, ensuring accurate representation in complex regions while reducing redundancy. Furthermore, we design a transformer-based Iterative Gaussian Refiner (IGR) module that refines Gaussian representations through direct image-Gaussian interactions. Our AdaptiveGaussian can effectively reduce Gaussian redundancy as input views increase. We conduct extensive experiments on the large-scale ACID and RealEstate10K datasets, where our method achieves state-of-the-art performance with good generalization to various numbers of views.", "keywords": [ "3D Gaussian splatting", "3D reconstruction", "Novel view synthesis" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "sQYQ9i1g86", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Dongbin Zhao", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Runyu Lu", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Yuanheng Zhu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 15, "n_ref_uni": 27, "n_ref": 51, "n_ref_all": 62, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1076, "n_element_tab_1": 13, "formula_len_all": 4007, "formula_len_all_1": 2559, "len_all": 121688, "len_all_1": 67298, "len_abs": 1413, "len_title": 157, "len_sents": 40814, "len_sents_1": 28528, "n_sents": 302, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1418, "title": "Constrained Exploitability Descent: Finding Mixed-Strategy Nash Equilibrium by Offline Reinforcement Learning", "abs": "This paper presents Constrained Exploitability Descent (CED), a novel model-free offline reinforcement learning algorithm for solving adversarial Markov games. CED is a game-theoretic approach combined with policy constraint methods from offline RL. While policy constraints can perturb the optimal pure-strategy solutions in single-agent scenarios, we find this side effect can be mitigated when it comes to solving adversarial games, where the optimal policy can be a mixed-strategy Nash equilibrium. We theoretically prove that, under the uniform coverage assumption on the dataset, CED converges to a stationary point in deterministic two-player zero-sum Markov games. The min-player policy at the stationary point satisfies the necessary condition for making up an exact mixed-strategy Nash equilibrium, even when the offline dataset is fixed and finite. Compared to the model-based method of Exploitability Descent that optimizes the max-player policy, our convergence result no longer relies on the generalized gradient. Experiments in matrix games, a tree-form game, and an infinite-horizon soccer game verify that a single run of CED leads to an optimal min-player policy when the practical offline data guarantees uniform coverage. Besides, CED achieves significantly lower NashConv compared to an existing pessimism-based method and can gradually improve the behavior policy even under non-uniform coverage.", "keywords": [ "offline reinforcement learning", "adversarial Markov game", "mixed-strategy Nash equilibrium", "policy constraint", "exploitability descent" ], "rating_list": [ 6, 3, 6 ], "soundness_list": [ 3, 1, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 1, 3 ], "confidence_list": [ 3, 5, 2 ] }, { "paper_id": "sPuLtU32av", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Egor Shulgin", "gender": "unknown", "institution": "KAUST", "country": "SA", "position": "PhD student" }, { "name": "Grigory Malinovsky", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Yury Demidovich", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Postdoc" }, { "name": "Peter Richtarik", "gender": "Male", "institution": "King Abdullah University of Science and Technology (KAUST)", "country": "SA", "position": "Full Professor" } ], "n_formula": 254, "n_formula_1": 41, "n_ref_uni": 62, "n_ref": 123, "n_ref_all": 155, "n_fig": 8, "n_tab": 8, "L_tab": 1382, "n_element_tab": 77, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 585, "n_element_tab_1": 14, "formula_len_all": 38803, "formula_len_all_1": 3351, "len_all": 288373, "len_all_1": 69850, "len_abs": 1104, "len_title": 88, "len_sents": 65312, "len_sents_1": 28108, "n_sents": 834, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1103, "title": "MAST: model-agnostic sparsified training", "abs": "We introduce a novel optimization problem formulation that departs from the conventional way of minimizing machine learning model loss as a black-box function. Unlike traditional formulations, the proposed approach explicitly incorporates an initially pre-trained model and random sketch operators, allowing for sparsification of both the model and gradient during training. We establish insightful properties of the proposed objective function and highlight its connections to the standard formulation. Furthermore, we present several variants of the Stochastic Gradient Descent (SGD) method adapted to the new problem formulation, including SGD with general sampling, a distributed version, and SGD with variance reduction techniques. We achieve tighter convergence rates and relax assumptions, bridging the gap between theoretical principles and practical applications, covering several important techniques such as Dropout and Sparse training. This work presents promising opportunities to enhance the theoretical understanding of model training through a sparsification-aware optimization approach.", "keywords": [ "dropout theory", "sparse training" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "sOte83GogU", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Md Ashiqur Rahman", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Raymond Yeh", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 48, "n_formula_1": 30, "n_ref_uni": 56, "n_ref": 97, "n_ref_all": 118, "n_fig": 4, "n_tab": 7, "L_tab": 2166, "n_element_tab": 239, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 205, "n_element_tab_1": 33, "formula_len_all": 2797, "formula_len_all_1": 2042, "len_all": 204085, "len_all_1": 78131, "len_abs": 1180, "len_title": 97, "len_sents": 57398, "len_sents_1": 29906, "n_sents": 523, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1199, "title": "Group Downsampling with Equivariant Anti-aliasing", "abs": "Downsampling layers are crucial building blocks in CNN architectures, which help to increase the receptive field for learning high-level features and reduce the amount of memory/computation in the model. In this work, we study the generalization of the uniform downsampling layer for group equivariant architectures, e.g., $G$-CNNs. That is, we aim to downsample signals (feature maps) on general finite groups *with* anti-aliasing. This involves the following: **(a)** Given a finite group and a downsampling rate, we present an algorithm to form a suitable choice of subgroup. **(b)** Given a group and a subgroup, we study the notion of bandlimited-ness and propose how to perform anti-aliasing. Notably, our method generalizes the notion of downsampling based on classical sampling theory. When the signal is on a cyclic group, i.e., periodic, our method recovers the standard downsampling of an ideal low-pass filter followed by a subsampling operation. Finally, we conducted experiments on image classification tasks demonstrating that the proposed downsampling operation improves accuracy, better preserves equivariance, and reduces model size when incorporated into $G$-equivariant networks", "keywords": [ "equivariance", "downsampling", "signal processing" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 2, 4, 2 ] }, { "paper_id": "sOmojPmnlL", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Feng Tian", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Professor" }, { "name": "Haonan Lin", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Jun Zhang", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Kuan Tian", "gender": "unknown", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Yan Chen", "gender": "unknown", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 4, "n_ref_uni": 39, "n_ref": 76, "n_ref_all": 111, "n_fig": 15, "n_tab": 5, "L_tab": 1228, "n_element_tab": 107, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 680, "n_element_tab_1": 95, "formula_len_all": 1064, "formula_len_all_1": 743, "len_all": 178487, "len_all_1": 55645, "len_abs": 1549, "len_title": 128, "len_sents": 47805, "len_sents_1": 25953, "n_sents": 369, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1576, "title": "AnyExpress: One Adapter Enabling Highly Flexible Audio-Driven Portrait Animation", "abs": "Portrait animation, particularly audio-driven portrait animation, requires flexibility in facial expressions, head movement, and dynamic contexts. However, existing diffusion-based methods rely heavily on the design of ReferenceNet, leading to increased training complexity and incompatibility with other custom base models or adapters, also limiting face position, view changes, and animated context generation. To address these challenges, we propose ***AnyExpress***, a lightweight, modular framework that eliminates the need for ReferenceNet, reducing the number of trainable parameters by **7** times. By training one plug-and-play *audio-motion adapter*, it allows freeform, expressive audio-driven portrait animation with any face pose and any animated context, while supporting text-driven modifications. In the context of character generation, there are two primary methods to control the desired character attributes. First, if a specific ID needs to be assigned, this can be achieved through ID controls (*e.g.*, IP-Adapter-Face). Alternatively, the character\u2019s attributes can be controlled through textual descriptions. Through comprehensive qualitative and quantitative analyses, ***AnyExpress*** demonstrates unprecedented freedom in generating videos with dynamic background, lower training demand, and seamless integration with evolving custom models and control adapters, providing a flexible solution for diverse generation needs. The demo is available at https://anyexpress-alpha.github.io/Any, and we will release our code, encouraging further improvement.", "keywords": [ "Diffusion-based Portrait Animation", "Audio-Driven Talking Face", "Plug-and-Play Adapter", "Flexible Video Generation" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "sOdeh2WJL2", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Abishek Satish", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Eric Lin", "gender": "unknown", "institution": "Dynamo AI", "country": "", "position": "Researcher" }, { "name": "James O' Neill", "gender": "unknown", "institution": "University of Liverpool", "country": "GB", "position": "PhD student" }, { "name": "Santhosh Subramanian", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Vaikkunth Mugunthan", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 18, "n_ref": 24, "n_ref_all": 36, "n_fig": 3, "n_tab": 6, "L_tab": 3931, "n_element_tab": 452, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3931, "n_element_tab_1": 452, "formula_len_all": 1064, "formula_len_all_1": 1064, "len_all": 95258, "len_all_1": 80655, "len_abs": 1653, "len_title": 84, "len_sents": 24476, "len_sents_1": 24492, "n_sents": 159, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1444, "title": "Unified Multi-Task Learning & Model Fusion for Efficient Language Model Guardrailing", "abs": "The trend towards large language models (LLMs) for guardrailing against undesired behaviors is increasing and has shown promise for censoring user inputs. However, high inference speed, memory consumption, hosting expenses and generative non-structured outputs can make their use prohibitive. \n\nIn this work, we show that task-specific data generation can lead to fine-tuned classifiers that significantly outperform current state of the art (SoTA) while being orders of magnitude smaller. Secondly, we show that using a single model, \\texttt{MultiTaskGuard}, that is pretrained on a large synthetically generated dataset with unique task instructions further improves generalization. Thirdly, our most performant models, \\texttt{UniGuard}, are found using our proposed search-based model merging approach that finds an optimal set of parameters to combine single-policy models and multi-policy guardrail models\n\nOn 7 public datasets and 4 new guardrail benchmarks we created, our efficient guardrail classifiers improve over the best performing SoTA publicly available LLMs and 3$^{\\text{rd}}$ party guardrail APIs in detecting unsafe and safe behaviors by an average \\textbf{29.92} (\\text{Aegis-LlamaGuard}) and \\textbf{21.62} (\\texttt{gpt-4o}) F1 respectively. Lastly, our guardrail synthetic data generation process leads to models that outperform training on real data using our custom defined policies that describe the guardrailing task.", "keywords": [ "safety guardrailing", "language model", "synthetic data generation", "multi-task learning", "model fusion" ], "rating_list": [ 3, 8, 1, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "sOa0SYS0cN", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Jiayan Qiu", "gender": "Male", "institution": "Univerisity of Leicester", "country": "GB", "position": "Lecturer" }, { "name": "Sheyang Tang", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Xiaoyu Xu", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Zhou Wang", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 19, "n_ref_uni": 46, "n_ref": 86, "n_ref_all": 105, "n_fig": 11, "n_tab": 4, "L_tab": 1006, "n_element_tab": 106, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 487, "n_element_tab_1": 93, "formula_len_all": 1603, "formula_len_all_1": 1571, "len_all": 148875, "len_all_1": 67628, "len_abs": 1396, "len_title": 117, "len_sents": 37998, "len_sents_1": 28482, "n_sents": 311, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1466, "title": "Controllable Data Generation with Hierarchical Neural Representations", "abs": "Implicit Neural Representations (INRs) represent data as continuous functions using the parameters of a neural network, where data information is encoded in the parameter space. Therefore, modeling the distribution of such parameters is crucial for building generalizable INRs. Existing approaches learn a joint distribution of these parameters via a latent vector to generate new data, but such a flat latent often fails to capture the inherent hierarchical structure of the parameter space, leading to entangled data semantics and limited control over the generation process. Here, we propose a $\\textbf{C}$ontrollable $\\textbf{H}$ierarchical $\\textbf{I}$mplicit $\\textbf{N}$eural $\\textbf{R}$epresentation ($\\textbf{CHINR}$) framework, which explicitly models conditional dependencies across layers in the parameter space. Our method consists of two stages: In Stage-1, we construct a Layers-of-Experts (LoE) network, where each layer modulates distinct semantics through a unique latent vector, enabling disentangled and expressive representations. In Stage-2, we introduce a Hierarchical Controllable Diffusion Model (HCDM) to capture conditional dependencies across layers, allowing for controllable and hierarchical data generation at various semantic granularities. Extensive experiments on CelebA-HQ, ShapeNet, SRN-Cars, and AMASS datasets demonstrate that CHINR improves generalizability and offers flexible hierarchical control over the generated content.", "keywords": [ "Implicit Neural Representations", "Generative INR", "Generative Models", "Diffusion Models" ], "rating_list": [ 6, 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 4, 3, 3 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 1, 4 ] }, { "paper_id": "sOQmgO0PTv", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Adrian F. Amil", "gender": "Male", "institution": "Donders Institute for Brain, Cognition and Behaviour - Radboud University", "country": "NL", "position": "PhD student" }, { "name": "Ismael T. Freire", "gender": "Male", "institution": "Radboud University", "country": "", "position": "PhD student" }, { "name": "Paul FMJ Verschure", "gender": "Male", "institution": "Catalan Institute of Advanced Studies", "country": "", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 48, "n_ref": 73, "n_ref_all": 105, "n_fig": 25, "n_tab": 1, "L_tab": 790, "n_element_tab": 140, "n_fig_1": 24, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 90, "formula_len_all_1": 45, "len_all": 137093, "len_all_1": 52046, "len_abs": 1988, "len_title": 120, "len_sents": 44207, "len_sents_1": 26933, "n_sents": 270, "n_sents_1": 155, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1996, "title": "Discretization of continuous input spaces in the hippocampal autoencoder", "abs": "Understanding the encoding mechanisms of hippocampal place cells remains a significant challenge in neuroscience. Although sparse autoencoders have been shown to exhibit place cell-like activity, the underlying processes are not fully understood. In this study, we compare spatial representations learned by dense and sparse autoencoders trained on images of 3D environments and find that only sparse autoencoders with orthonormal activity regularization in latent space produce place cells. We then show that this regularization promotes similar images to map onto the same neurons, acting as a locality-sensitive hash function. Notably, we demonstrate that these neurons are visually interpretable through activity clamping and decoding, suggesting the formation of detailed episodic memories at the single-neuron level. We then introduce a novel metric to quantify how neurons discretize the image space into disjoint receptive fields, revealing that sparse autoencoders tile input spaces with minimal overlap. Furthermore, we observe that whereas dense autoencoders generate population codes resembling visual cortex activity near criticality, sparse autoencoders produce higher-dimensional codes, thus suggesting a similar coding strategy in the hippocampus. Extending our approach to the auditory domain, we also replicate the emergence of \"frequency place cells\" by training sparse autoencoders on audio snippets sampled from a frequency-varying signal, and show that population representations retain the statistical structure of the sample distribution. Lastly, we demonstrate that reinforcement learning agents can leverage these high-dimensional image representations to solve complex spatial-cognitive tasks, despite their inherent brittleness. Overall, our findings elucidate how sparse input compression in autoencoders can give rise to discrete, interpretable memories, establishing an explicit link between episodic memory formation and spatial representations in the hippocampus.", "keywords": [ "NeuroAI", "Memory", "Sparse autoencoders", "Hippocampus" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 1, 1 ], "presentation_list": [ 2, 4, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "sOOrTkYgb6", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Christian Bartelt", "gender": "Male", "institution": "Technische Universit\u00e4t Clausthal", "country": "DE", "position": "Full Professor" }, { "name": "Patrick Knab", "gender": "Male", "institution": "Technische Universit\u00e4t Clausthal", "country": "DE", "position": "PhD student" }, { "name": "Sascha Marton", "gender": "Male", "institution": "Technische Universit\u00e4t Clausthal", "country": "DE", "position": "Lecturer" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 51, "n_ref": 125, "n_ref_all": 171, "n_fig": 16, "n_tab": 20, "L_tab": 5608, "n_element_tab": 505, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 871, "n_element_tab_1": 90, "formula_len_all": 421, "formula_len_all_1": 356, "len_all": 217171, "len_all_1": 72901, "len_abs": 1116, "len_title": 127, "len_sents": 65267, "len_sents_1": 34185, "n_sents": 559, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1118, "title": "DSEG-LIME: Improving Image Explanation by Hierarchical Data-Driven Segmentation", "abs": "Explainable Artificial Intelligence (XAI) is crucial in unraveling decision-making processes in complex machine learning models. LIME (Local Interpretable Model-agnostic Explanations) is a well-known XAI framework for image analysis. It utilizes image segmentation to create features to identify relevant areas for classification. Consequently, poor segmentation can compromise the consistency of the explanation and undermine the importance of the segments, affecting the overall interpretability. Addressing these challenges, we introduce DSEG-LIME (Data-Driven Segmentation LIME), featuring: i) a data-driven segmentation for human-recognized feature generation by foundation model integration, and ii) a user steered granularity in the hierarchical segmentation procedure through composition. We evaluate DSEG-LIME on pre-trained models using ImageNet classes, explicitly targeting scenarios without domain-specific knowledge. Our findings demonstrate that DSEG outperforms most of the XAI metrics and enhances the alignment of explanations with human-recognized concepts, significantly improving interpretability.", "keywords": [ "XAI", "LIME", "Segmentation" ], "rating_list": [ 6, 3, 8, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "sOAQY8hrAu", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Bohan Zeng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Junlin Han", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Ling Yang", "gender": "Male", "institution": "DeepSeek AI", "country": "CN", "position": "Intern" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Runjia Li", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Zixiang Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Intern" }, { "name": "Zhang wen tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 85, "n_ref_all": 116, "n_fig": 13, "n_tab": 6, "L_tab": 1001, "n_element_tab": 112, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 278, "n_element_tab_1": 37, "formula_len_all": 630, "formula_len_all_1": 226, "len_all": 167019, "len_all_1": 55640, "len_abs": 2004, "len_title": 124, "len_sents": 52977, "len_sents_1": 26926, "n_sents": 393, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1551, "title": "Semantic Score Distillation Sampling for Compositional Text-to-3D Generation", "abs": "Generating high-quality 3D assets from textual descriptions remains a pivotal challenge in computer graphics and vision research. Due to the scarcity of 3D data, state-of-the-art approaches utilize pre-trained 2D diffusion priors, optimized through Score Distillation Sampling (SDS). Despite progress, crafting complex 3D scenes featuring multiple objects or intricate interactions is still difficult. To tackle this, recent methods have incorporated box or layout guidance. However, these layout-guided compositional methods often struggle to provide fine-grained control, as they are generally coarse and lack expressiveness. To overcome these challenges, we introduce a novel SDS approach, Semantic Score Distillation Sampling (SemanticSDS), designed to effectively improve the expressiveness and accuracy of compositional text-to-3D generation. Our approach integrates new semantic embeddings that maintain consistency across different rendering views and clearly differentiate between various objects and parts. These embeddings are transformed into a semantic map, which directs a region-specific SDS process, enabling precise optimization and compositional generation. By leveraging explicit semantic guidance, our method unlocks the compositional capabilities of existing pre-trained diffusion models, thereby achieving superior quality in 3D content generation, particularly for complex objects and scenes. Experimental results demonstrate that our SemanticSDS framework is highly effective for generating state-of-the-art complex 3D content.", "keywords": [ "Text-to-3D Generation", "Score Distillation Sampling" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "sNycNM577m", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Bei Yu", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Hengshuang Zhao", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Jiaya Jia", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Mingkang Zhu", "gender": "unknown", "institution": "Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xi Chen", "gender": "Male", "institution": "the University of Hong Kong, University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zhongdao Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 28, "n_ref": 59, "n_ref_all": 72, "n_fig": 2, "n_tab": 8, "L_tab": 3908, "n_element_tab": 198, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 944, "n_element_tab_1": 157, "formula_len_all": 298, "formula_len_all_1": 297, "len_all": 178003, "len_all_1": 68593, "len_abs": 1648, "len_title": 67, "len_sents": 39769, "len_sents_1": 32439, "n_sents": 308, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1658, "title": "Effective LLM Knowledge Learning Requires Rethinking Generalization", "abs": "Large language models (LLMs) are trained on a substantial amount of documents that contain extensive world knowledge. However, it is still not well-understood how knowledge is acquired via autoregressive pre-training and extracted via question-answering. This lack of understanding greatly hinders effective knowledge learning, especially for continued pre-training on up-to-date information, as this evolving information often does not have diverse repetitions like foundational knowledge. In this paper, we focus on understanding and improving LLM knowledge learning. We found and verified that knowledge learning for LLMs can be deemed as an implicit supervised task hidden in the autoregressive pre-training objective. Our findings suggest that knowledge learning for LLMs would benefit from methods designed to improve generalization ability for supervised tasks. Based on our analysis, we propose to diversify training documents\u2019 formats as data augmentation to grow in-distribution samples. This data augmentation method does not present the risk of altering the facts embedded in documents as text paraphrasing. We also introduce sharpness-aware minimization as an effective optimization algorithm to better improve generalization. Moreover, we adapt our method to instruction tuning for generalization to various phrasings of questions. Extensive experiment results validate our findings and demonstrate our methods\u2019 effectiveness in improving knowledge learning in both the continued pre-training and instruction tuning stages. This paper offers new perspectives and insights to interpret and design effective strategies for LLM knowledge learning.", "keywords": [ "knowledge learning", "generalization", "large language models", "knowledge acquisition" ], "rating_list": [ 3, 6, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "sNoJSfGh6y", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Huihui Gong", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "PhD student" }, { "name": "Minjing Dong", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Siqi Ma", "gender": "Female", "institution": "University of New South Wales", "country": "", "position": "Assistant Professor" }, { "name": "Chang Xu", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" } ], "n_formula": 21, "n_formula_1": 22, "n_ref_uni": 19, "n_ref": 27, "n_ref_all": 38, "n_fig": 3, "n_tab": 4, "L_tab": 1511, "n_element_tab": 148, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 3381, "n_element_tab_1": 155, "formula_len_all": 1400, "formula_len_all_1": 1399, "len_all": 119484, "len_all_1": 62720, "len_abs": 1637, "len_title": 124, "len_sents": 30065, "len_sents_1": 26899, "n_sents": 217, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1648, "title": "RED: Efficiently Boosting Ensemble Robustness via Random Sampling Inference", "abs": "Despite the remarkable achievements of Deep Neural Networks (DNNs) in handling diverse tasks, these high-performing models remain susceptible to adversarial attacks. Considerable research has focused on bolstering the robustness of individual models and subsequently employing a simple ensemble defense strategy. However, existing ensemble techniques tend to increase the inference latency and the parameter number while achieving suboptimal robustness, which motivates us to reconsider the framework of model ensemble. To address the challenge of suboptimal robustness and inference latency, we introduce a novel ensemble defense approach called Random Ensemble Defense (RED). Specifically, we expedite inference via random sampling, which also makes it difficult for an attacker to attack a model ensemble. To effectively train a model ensemble, it is crucial to diversify the adversarial vulnerabilities among its members. This can be approached by reducing the adversarial transferability among them. To this end, we propose incorporating gradient similarity and Lipschitz regularizers into the training process. Moreover, to overcome the obstacle of a large number of parameters, we develop a parameter-lean version of RED (PS-RED). Extensive experiments, conducted across popular datasets, demonstrate that the proposed methods not only significantly improve ensemble robustness but also minimize inference delays and optimize storage usage for ensemble models. For example, our models enhance robust accuracy by approximately 15\\% (RED) and save parameters by approximately 90\\% (PS-RED) on CIFAR-10 compared with the most recent baselines.", "keywords": [ "Adversarial Robustness", "Ensemble Defences", "Randomness", "Hypernetworks", "Computer Vision" ], "rating_list": [ 6, 5, 3, 3, 3 ], "soundness_list": [ 3, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "sNntRFmn72", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Chen Huang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Dejia Xu", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Hao Tang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Liangchen Song", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Liangliang Cao", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Thorsten Gernoth", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Yifan Jiang", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Zhangyang Wang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 58, "n_ref": 184, "n_ref_all": 207, "n_fig": 14, "n_tab": 4, "L_tab": 1672, "n_element_tab": 113, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 4154, "n_element_tab_1": 73, "formula_len_all": 308, "formula_len_all_1": 183, "len_all": 198418, "len_all_1": 55323, "len_abs": 1499, "len_title": 132, "len_sents": 48164, "len_sents_1": 24165, "n_sents": 376, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1509, "title": "Cavia: Camera-controllable Multi-view Video Diffusion with View-Integrated Attention", "abs": "In recent years there have been remarkable breakthroughs in image-to-video generation.\nHowever, the 3D consistency and camera controllability of generated frames have remained unsolved. Recent studies have attempted to incorporate camera control into the generation process, but their results are often limited to simple trajectories or lack the ability to generate consistent videos from multiple distinct camera paths for the same scene. To address these limitations, we introduce **Cavia**, a novel framework for camera-controllable, multi-view video generation, capable of converting an input image into multiple spatiotemporally consistent videos. Our framework extends the spatial and temporal attention modules into view-integrated attention modules, improving both viewpoint and temporal consistency. This flexible design allows for joint training with diverse curated data sources, including scene-level static videos, object-level synthetic multi-view dynamic videos, and real-world monocular dynamic videos. To our best knowledge, Cavia is the first of its kind that allows the user to precisely specify camera motion while obtaining object motion. To the best of our knowledge, Cavia is the first framework that enables users to generate multiple videos of the same scene with precise control over camera motion, while simultaneously preserving object motion. Extensive experiments demonstrate that Cavia surpasses state-of-the-art methods in terms of geometric consistency and perceptual quality.", "keywords": [ "Video Generation" ], "rating_list": [ 5, 5, 8, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "sNZTtDRFXt", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Huazhe Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jianfei Ma", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 40, "n_formula_1": 28, "n_ref_uni": 48, "n_ref": 86, "n_ref_all": 120, "n_fig": 6, "n_tab": 2, "L_tab": 826, "n_element_tab": 110, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 894, "n_element_tab_1": 113, "formula_len_all": 4091, "formula_len_all_1": 2502, "len_all": 174534, "len_all_1": 62745, "len_abs": 1334, "len_title": 130, "len_sents": 45312, "len_sents_1": 26739, "n_sents": 376, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1340, "title": "DRIVE: Distributional Model-Based Reinforcement Learning via Variational Inference", "abs": "Distributional reinforcement learning (RL) provides a natural framework for estimating the distribution of returns rather than a single expected value. However, the control aspect of distributional RL has not been as thoroughly explored as the evaluation part, typically relying on the greedy selection rule with respect to either the expected value, akin to standard approaches, or risk-sensitive measures derived from the return distribution. On the other hand, casting RL as a probabilistic inference problem allows for flexible control solutions utilizing a toolbox of approximate inference techniques; however, its connection to distributional RL remains underexplored. In this paper, we bridge this gap by proposing a variational approach for efficient policy search. Our method leverages the log-likelihood of optimality as a learning proxy, decoupling it from traditional value functions. This learning proxy incorporates aleatoric uncertainty of the return distribution, enabling risk-aware decision-making. We provide a theoretical analysis of our framework, detailing the conditions for convergence. Empirical results on vision-based tasks in DMControl Suite demonstrate the effectiveness of our approach compared to various algorithms, as well as its ability to balance exploration and exploitation at different training stages.", "keywords": [ "Distributional RL", "Control as Inference", "Decision-Making under Uncertainty" ], "rating_list": [ 6, 8, 6, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "sMyXP8Tanm", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Chongxuan Li", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Fengqi Zhu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Jiacheng Sun", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Senior Researcher" }, { "name": "Jingyang Ou", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "Undergrad student" }, { "name": "Shen Nie", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "PhD student" }, { "name": "Zhenguo Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Kevin Xue", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "", "position": "Undergrad student" } ], "n_formula": 116, "n_formula_1": 20, "n_ref_uni": 43, "n_ref": 150, "n_ref_all": 193, "n_fig": 3, "n_tab": 9, "L_tab": 3477, "n_element_tab": 261, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 275, "n_element_tab_1": 20, "formula_len_all": 12396, "formula_len_all_1": 1196, "len_all": 262202, "len_all_1": 60843, "len_abs": 1350, "len_title": 141, "len_sents": 79303, "len_sents_1": 24918, "n_sents": 719, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1420, "title": "Your Absorbing Discrete Diffusion Secretly Models the Conditional Distributions of Clean Data", "abs": "Discrete diffusion models with absorbing processes have shown promise in language modeling. The key quantities to be estimated are the ratios between the marginal probabilities of two transitive states at all timesteps, called the concrete score. In this paper, we reveal that the concrete score in absorbing diffusion can be expressed as conditional probabilities of clean data, multiplied by a time-dependent scalar in an analytic form. Motivated by this finding, we propose reparameterized absorbing discrete diffusion (RADD), a dedicated diffusion model without time-condition that characterizes the time-independent conditional probabilities. Besides its simplicity, RADD can reduce the number of function evaluations (NFEs) by caching the output of the time-independent network when the noisy sample remains unchanged in a sampling interval, which enables sampling acceleration. Built upon the new perspective of conditional distributions, we further unify absorbing discrete diffusion and any-order autoregressive models (AO-ARMs), showing that the upper bound on the negative log-likelihood for the diffusion model can be interpreted as an expected negative log-likelihood for AO-ARMs. Further, our RADD models achieve SOTA performance among diffusion models on 5 zero-shot language modeling benchmarks (measured by perplexity) at the GPT-2 scale. Our code is available at \\url{https://github.com/ML-GSAI/RADD}.", "keywords": [ "Discrete Diffusion Models", "Diffusion Models", "Language Modeling", "Concrete Score", "Score Entropy" ], "rating_list": [ 6, 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 5, 3, 5, 4, 3 ] }, { "paper_id": "sMwYn2lZjO", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Pingzhi Li", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Xiaolong Jin", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "MS student" }, { "name": "Yu Cheng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 17, "n_ref": 25, "n_ref_all": 48, "n_fig": 6, "n_tab": 5, "L_tab": 2335, "n_element_tab": 381, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1812, "n_element_tab_1": 287, "formula_len_all": 172, "formula_len_all_1": 172, "len_all": 142195, "len_all_1": 61111, "len_abs": 1500, "len_title": 120, "len_sents": 32189, "len_sents_1": 27175, "n_sents": 243, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1459, "title": "Examining Post-Training Quantization for Mixture-of-Experts: A Benchmark", "abs": "Large Language Models (LLMs) have become foundational in the realm of natural language processing, demonstrating performance improvements as model sizes increase. The Mixture-of-Experts (MoE) approach offers a promising way to scale LLMs more efficiently by using fewer computational FLOPs through sparse activation. However, it suffers from significant memory overheads, necessitating model compression techniques. Post-training quantization, a popular method for model compression, proves less effective when directly applied to MoE models due to MoE's overlooked inherent sparsity. This paper explores several MoE structure-aware quantization heuristics, ranging from coarse to fine granularity, from MoE block to individual linear weight. Our investigations reveal critical principles: different MoE structures (i.e., blocks, experts, linear layers) require varying numbers of weight bits for effective and efficient quantization. Conclusions are supported by extensive benchmarking across two representative MoE models and six tasks. We further introduce novel enhancements to more accurately identify the most critical weights in MoE quantization that necessitate higher bit allocations, including the linear weight outlier scorer and MoE block scorer. Additionally, subsequent experiments validate our findings in the context of both weight and activation quantization. Our code for reproducing all our experiments is provided as supplemental material.", "keywords": [ "Sparse Mixture-of-Experts", "Efficiency", "Compression", "Quantization" ], "rating_list": [ 5, 3, 6, 3, 6 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 3, 1, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 3, 5, 3 ] }, { "paper_id": "sMWkTWh2JF", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Ahmad Elawady", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "MS student" }, { "name": "Andrew Szot", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Dhruv Batra", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Associate Professor" }, { "name": "Gunjan Chhablani", "gender": "Male", "institution": "Waymo", "country": "US", "position": "Researcher" }, { "name": "Karmesh Yadav", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Ram Ramrakhya", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zsolt Kira", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 45, "n_ref": 82, "n_ref_all": 120, "n_fig": 27, "n_tab": 2, "L_tab": 307, "n_element_tab": 27, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 171, "n_element_tab_1": 21, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 157781, "len_all_1": 68631, "len_abs": 1462, "len_title": 130, "len_sents": 50573, "len_sents_1": 34230, "n_sents": 413, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1468, "title": "ReLIC: A Recipe for 64k Steps of In-Context Reinforcement Learning for Embodied AI", "abs": "Intelligent embodied agents need to quickly adapt to new scenarios by integrating long histories of experience into decision-making. For instance, a robot in an unfamiliar house initially wouldn't know the locations of objects needed for tasks and might perform inefficiently. However, as it gathers more experience, it should learn the layout of its environment and remember where objects are, allowing it to complete new tasks more efficiently. To enable such rapid adaptation to new tasks, we present ReLIC, a new approach for in-context reinforcement learning (RL) for embodied agents. With ReLIC, agents are capable of adapting to new environments using 64,000 steps of in-context experience with full attention while being trained through self-generated experience via RL. We achieve this by proposing a novel policy update scheme for on-policy RL called \"partial updates\" as well as a Sink-KV mechanism that enables effective utilization of a long observation history for embodied agents. Our method outperforms a variety of meta-RL baselines in adapting to unseen houses in an embodied multi-object navigation task. In addition, we find that ReLIC is capable of few-shot imitation learning despite never being trained with expert demonstrations. We also provide a comprehensive analysis of ReLIC, highlighting that the combination of large-scale RL training, the proposed partial updates scheme, and the Sink-KV are essential for effective in-context learning.", "keywords": [ "Embodied AI", "Reinforcement Learning", "Meta-RL", "In-context Learning" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "sLtuNGkKfH", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Dazhi Fu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Jicong Fan", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Yangyang Qu", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "", "position": "MS student" } ], "n_formula": 18, "n_formula_1": 11, "n_ref_uni": 51, "n_ref": 80, "n_ref_all": 107, "n_fig": 9, "n_tab": 24, "L_tab": 11990, "n_element_tab": 750, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2627, "n_element_tab_1": 206, "formula_len_all": 1807, "formula_len_all_1": 1190, "len_all": 188792, "len_all_1": 60325, "len_abs": 1528, "len_title": 119, "len_sents": 53003, "len_sents_1": 25570, "n_sents": 361, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1533, "title": "Subject Information Extraction for Novelty Detection with Domain Shifts", "abs": "Unsupervised novelty detection (UND), aimed at identifying novel samples, is essential in fields like medical diagnosis, cybersecurity, and industrial quality control. Most existing UND methods assume that the training data and testing normal data originate from the same domain and only consider the distribution variation between training data and testing data. However, in real scenarios, it is common for normal testing and training data to originate from different domains, a challenge known as domain shift. The discrepancies between training and testing data often lead to incorrect classification of normal data as novel by existing methods. A typical situation is that testing normal data and training data describe the same subject, yet they differ in the background conditions. To address this problem, we introduce a novel method that separates subject information from background variation encapsulating the domain information to enhance detection performance under domain shifts. The proposed method minimizes the mutual information between the representations of the subject and background while modelling the background variation using a deep Gaussian mixture model, where the novelty detection is conducted on the subject representations solely and hence is not affected by the variation of domains. Extensive experiments demonstrate that our model generalizes effectively to unseen domains and significantly outperforms baseline methods, especially under substantial domain shifts between training and testing data.", "keywords": [ "novelty detection" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "sLNRvScGM2", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Huanle Xu", "gender": "Male", "institution": "University of Macau", "country": "", "position": "Assistant Professor" }, { "name": "Wenyan Chen", "gender": "Female", "institution": "University of Macau", "country": "CN", "position": "PhD student" }, { "name": "Xiaosong Chen", "gender": "Male", "institution": "University of Macau", "country": "", "position": "PhD student" }, { "name": "Yuanhang Chen", "gender": "unknown", "institution": "University of Macau", "country": "MO", "position": "PhD student" } ], "n_formula": 42, "n_formula_1": 31, "n_ref_uni": 28, "n_ref": 53, "n_ref_all": 70, "n_fig": 6, "n_tab": 3, "L_tab": 1480, "n_element_tab": 160, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1504, "n_element_tab_1": 125, "formula_len_all": 2187, "formula_len_all_1": 1535, "len_all": 123088, "len_all_1": 63802, "len_abs": 938, "len_title": 123, "len_sents": 31340, "len_sents_1": 27172, "n_sents": 284, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 940, "title": "FedSUV: Validity and Utility-guided Client Selection for Federated Learning", "abs": "Federated Learning faces significant challenges arising from two critical uncertainties: the validity of a client\u2019s participation, which can be compromised by network and system heterogeneity, and the utility of the data contributed by each client, which varies due to heterogeneous statistical data. Traditional client selection methods often treat these uncertainties as a whole, leading to suboptimal performance. To address this issue, we propose FedSUV, an innovative client selection framework that decouples validity and utility uncertainties. FedSUV approaches client selection from a multi-objective optimization perspective, employing advanced bandit algorithms: a confidence bound-based linear contextual bandit for assessing validity and a Gaussian Process bandit for evaluating utility. We validate the effectiveness of FedSUV through both theoretical analysis and large-scale experiments conducted within our physical cluster.", "keywords": [ "federated learning", "client selection", "multi-armed bandit", "uncertainty estimation" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "sLKDbuyq99", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Boye Niu", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "Undergrad student" }, { "name": "Kai Lian", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "MS student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Yifan Shen", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Yiliao Song", "gender": "Female", "institution": "University of Adelaide", "country": "AU", "position": "Lecturer" }, { "name": "Yu Yao", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Lecturer" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 19, "n_ref": 41, "n_ref_all": 57, "n_fig": 6, "n_tab": 8, "L_tab": 1907, "n_element_tab": 215, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 681, "n_element_tab_1": 81, "formula_len_all": 430, "formula_len_all_1": 330, "len_all": 152904, "len_all_1": 67540, "len_abs": 1095, "len_title": 138, "len_sents": 56156, "len_sents_1": 33275, "n_sents": 433, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1118, "title": "Improving Large Language Model based Multi-Agent Framework through Dynamic Workflow Updating", "abs": "Multi-agent frameworks powered by large language models (LLMs) have demonstrated great success in automated planning and task execution. However, the effective adjustment of workflows during execution has not been well studied. A flexible workflow is crucial, as in many real-world scenarios, the initial plan must adjust to unforeseen challenges and changing conditions in real-time to ensure the efficient execution of complex tasks. In this paper, we define workflows as activity-on-vertex (AOV) graphs. We continuously refine the workflow by dynamically adjusting task allocations and agent roles based on historical performance and previous AOV graphs with LLM agents. To further enhance system performance, we emphasize modularity in workflow design based on measuring parallelism and dependence complexity. Our proposed multi-agent framework achieved efficient sub-task concurrent execution, goal achievement, and error tolerance. Empirical results across various practical tasks demonstrate significant improvements in the efficiency of multi-agent systems through dynamic workflow updating and modularization.", "keywords": [ "LLMs based Multi-Agent System" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "sL2F9YCMXf", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Arash Vahdat", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Jure Leskovec", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Karsten Kreis", "gender": "unknown", "institution": "NVIDIA", "country": "CA", "position": "Research Scientist" }, { "name": "Minkai Xu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Stefano Ermon", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Tomas Geffner", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Weili Nie", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Yilun Xu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 30, "n_formula_1": 21, "n_ref_uni": 52, "n_ref": 92, "n_ref_all": 120, "n_fig": 4, "n_tab": 4, "L_tab": 2461, "n_element_tab": 171, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1225, "n_element_tab_1": 144, "formula_len_all": 4210, "formula_len_all_1": 2535, "len_all": 174640, "len_all_1": 74424, "len_abs": 1453, "len_title": 106, "len_sents": 54852, "len_sents_1": 33567, "n_sents": 393, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1459, "title": "Energy-Based Diffusion Language Models for Text Generation", "abs": "Despite remarkable progress in autoregressive language models, alternative generative paradigms beyond left-to-right generation are still being actively explored. Discrete diffusion models, with the capacity for parallel generation, have recently emerged as a promising alternative. Unfortunately, these models still underperform the autoregressive counterparts, with the performance gap increasing when reducing the number of sampling steps. Our analysis reveals that this degradation is a consequence of an imperfect approximation used by diffusion models. In this work, we propose Energy-based Diffusion Language Model (EDLM), an energy-based model operating at the full sequence level for each diffusion step, introduced to improve the underlying approximation used by diffusion models. More specifically, we introduce an EBM in a residual form, and show that its parameters can be obtained by leveraging a pretrained autoregressive model or by finetuning a bidirectional transformer via noise contrastive estimation. We also propose an efficient generation algorithm via parallel important sampling. Comprehensive experiments on language modeling benchmarks show that our model can consistently outperform state-of-the-art diffusion models by a significant margin, and approaches autoregressive models' perplexity. We further show that, without any generation performance drop, our framework offers a 1.3x sampling speedup over existing diffusion models.", "keywords": [ "Language Models", "Discrete Diffusion Models", "Energy-based Models" ], "rating_list": [ 5, 6, 8, 8 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 2 ] }, { "paper_id": "sKv4bbbqUa", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Amauri Holanda Souza Jr", "gender": "Male", "institution": "Federal Institute of Cear\u00e1", "country": "BR", "position": "Associate Professor" }, { "name": "Yogesh Verma", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "PhD student" }, { "name": "Vikas K Garg", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 61, "n_ref": 129, "n_ref_all": 142, "n_fig": 3, "n_tab": 3, "L_tab": 1647, "n_element_tab": 209, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 2013, "n_element_tab_1": 188, "formula_len_all": 1970, "formula_len_all_1": 2053, "len_all": 146836, "len_all_1": 70473, "len_abs": 1205, "len_title": 79, "len_sents": 36591, "len_sents_1": 29095, "n_sents": 280, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 31, "L_abs": 1210, "title": "Topological Positional Encoding", "abs": "Unlike words in sentences, nodes in general graphs do not have canonical positional information. As a result, the local message-passing framework of popular graph neural networks (GNNs) fails to leverage possibly relevant global structures for the task at hand. In this context, positional encoding methods emerge as an efficient approach to enrich the representational power of GNNs, helping them break node symmetries in input graphs. Similarly, multiscale topological descriptors based on persistent homology have also been integrated into GNNs to boost their expressivity. However, it remains unclear how positional encoding interplays with PH-based topological features and whether we can align the two to improve expressivity further. We address this issue with a novel notion of topological positional encoding (ToPE) that amalgamates the strengths of persistence homology and positional encoding. We establish that ToPE has provable expressivity benefits. Strong empirical assessments further underscore the effectiveness of the proposed method on several graph and language processing applications, including molecular property prediction, out-of-distribution generalization, and synthetic tree tasks.", "keywords": [ "Topological Neural Networks", "Persistent Homology", "Positional Encodings", "GNNs" ], "rating_list": [ 5, 5, 1, 5, 3 ], "soundness_list": [ 2, 2, 1, 2, 3 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 3, 5, 3 ] }, { "paper_id": "sKYHBTAxVa", "primary_area": "foundation or frontier models, including LLMs", "n_author": 18, "author_info_list": [ { "name": "Arka Pal", "gender": "unknown", "institution": "Ritual AI", "country": "US", "position": "Researcher" }, { "name": "Benjamin Feuer", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Chinmay Hegde", "gender": "Male", "institution": "New York University", "country": "US", "position": "Associate Professor" }, { "name": "Colin White", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Khalid Saifullah", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Micah Goldblum", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" }, { "name": "Neel Jain", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Samuel Dooley", "gender": "unknown", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Sandeep Singh Sandha", "gender": "unknown", "institution": "Abacus inc", "country": "US", "position": "Researcher" }, { "name": "Shubh-Agrawal", "gender": "Male", "institution": "Abacus.ai", "country": "IN", "position": "Researcher" }, { "name": "Siddartha Venkat Naidu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Sid Jain", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Sreemanti Dey", "gender": "Female", "institution": "Indian Institute of Technology, Delhi", "country": "IN", "position": "Undergrad student" }, { "name": "Tom Goldstein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Willie Neiswanger", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Yann LeCun", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Director" }, { "name": "Manley Howell Roberts", "gender": "Male", "institution": "Abridge", "country": "US", "position": "Researcher" }, { "name": "ravid ziv", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 125, "n_ref_all": 161, "n_fig": 5, "n_tab": 11, "L_tab": 10024, "n_element_tab": 938, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 811, "n_element_tab_1": 131, "formula_len_all": 646, "formula_len_all_1": 0, "len_all": 315946, "len_all_1": 65150, "len_abs": 1724, "len_title": 61, "len_sents": 88607, "len_sents_1": 31259, "n_sents": 748, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 58, "L_abs": 1731, "title": "LiveBench: A Challenging, Contamination-Free LLM Benchmark", "abs": "Test set contamination, wherein test data from a benchmark ends up in a newer model's training set, is a well-documented obstacle for fair LLM evaluation and can quickly render benchmarks obsolete. To mitigate this, many recent benchmarks crowdsource new prompts and evaluations from human or LLM judges; however, these can introduce significant biases, and break down when scoring hard questions. In this work, we introduce a new benchmark for LLMs designed to be resistant to both test set contamination and the pitfalls of LLM judging and human crowdsourcing. We release LiveBench, the first benchmark that (1) contains frequently-updated questions from recent information sources, (2) scores answers automatically according to objective ground-truth values, and (3) contains a wide variety of challenging tasks, spanning math, coding, reasoning, language, instruction following, and data analysis. To achieve this, LiveBench contains questions that are based on recently-released math competitions, arXiv papers, news articles, and datasets, and it contains harder, contamination-limited versions of tasks from previous benchmarks such as Big-Bench Hard, AMPS, and IFEval. We evaluate many prominent closed-source models, as well as dozens of open-source models ranging from 0.5B to 405B in size. LiveBench is difficult, with top models achieving below 70% accuracy. We release all questions, code, and model answers. Questions are added and updated on a monthly basis, and we release new tasks and harder versions of tasks over time so that LiveBench can distinguish between the capabilities of LLMs as they improve in the future. We welcome community engagement and collaboration for expanding the benchmark tasks and models.", "keywords": [ "large language models", "benchmark" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 4, 2 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 4, 4, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "sJzfxRbEv6", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Jaekwang KIM", "gender": "Male", "institution": "Sungkyunkwan University", "country": "", "position": "Lecturer" }, { "name": "San Kim", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "PhD student" }, { "name": "SeungJunLee", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "MS student" }, { "name": "sichan OH", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "MS student" } ], "n_formula": 28, "n_formula_1": 9, "n_ref_uni": 13, "n_ref": 14, "n_ref_all": 23, "n_fig": 4, "n_tab": 5, "L_tab": 911, "n_element_tab": 117, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 921, "n_element_tab_1": 118, "formula_len_all": 871, "formula_len_all_1": 435, "len_all": 94390, "len_all_1": 56996, "len_abs": 1065, "len_title": 85, "len_sents": 27571, "len_sents_1": 19484, "n_sents": 209, "n_sents_1": 147, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1069, "title": "Moir\u00e9 Graph Transformer: Eliminating Positional Encoding with Focused Attention", "abs": "Graph neural networks (GNNs) have increasingly adopted transformer architectures to capture long-range dependencies. However, integrating structural information into graph transformers remains challenging, often necessitating complex positional encodings or masking strategies. In this paper, we propose the Moir\u00e9 Graph Transformer (Moir\u00e9GT), which introduces a novel focused attention mechanism that eliminates the need for explicit graph positional encodings. Our model effectively captures structural context without additional encodings or masks by adjusting attention scores based on a learnable focus function of node distances. We theoretically demonstrate that multiple attention heads with different focus parameters can implicitly encode positional information akin to moir\u00e9 patterns. Experiments on 3D molecular graphs show that Moir\u00e9GT achieves significant performance gains over state-of-the-art models on the QM9 and PCQM4Mv2 datasets. Additionally, our model achieves competitive results on 2D graph tasks, highlighting its versatility and effectiveness.", "keywords": [ "graph transformer", "eliminating graph positional encoding", "focused attention mechanism" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "sIGWTd1DcW", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Jason E Weston", "gender": "Not Specified", "institution": "New York University", "country": "US", "position": "Researcher" }, { "name": "Olga Golovneva", "gender": "Female", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Sainbayar Sukhbaatar", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Research Scientist" }, { "name": "Tianlu Wang", "gender": "Female", "institution": "Meta", "country": "", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 30, "n_ref": 47, "n_ref_all": 87, "n_fig": 7, "n_tab": 10, "L_tab": 1879, "n_element_tab": 303, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 586, "n_element_tab_1": 83, "formula_len_all": 900, "formula_len_all_1": 769, "len_all": 141415, "len_all_1": 61739, "len_abs": 955, "len_title": 112, "len_sents": 43673, "len_sents_1": 27200, "n_sents": 377, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 956, "title": "Contextual Position Encoding: Learning to Count What\u2019s Important", "abs": "The attention mechanism is a critical component of Large Language Models (LLMs) that allows tokens in a sequence to interact with each other, but is order-invariant. Incorporating position encoding (PE) makes it possible to address by position, such as attending to the i-th token. However, current PE methods use token counts to derive position, and thus cannot generalize to higher levels of abstraction, such as attending to the i-th sentence. In this paper, we propose a new position encoding method, Contextual Position Encoding (CoPE), that allows positions to be conditioned on context by incrementing position only on certain tokens determined by the model. This allows more general position addressing such as attending to the i-th particular word, noun, or sentence. We show that CoPE can solve the selective copy, counting and Flip-Flop tasks where popular position embeddings fail, and improves perplexity on language modeling and coding tasks.", "keywords": [ "Deep learning architectures", "Large Language Model (LLM)", "position encoding" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "sIE2rI3ZPs", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Alex Damian", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Ameet Talwalkar", "gender": "Male", "institution": "Carnegie-Mellon University", "country": "", "position": "" }, { "name": "Jason D. Lee", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Jeremy Cohen", "gender": "unknown", "institution": "Flatiron Institute", "country": "US", "position": "Postdoc" }, { "name": "Zico Kolter", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 107, "n_formula_1": 13, "n_ref_uni": 37, "n_ref": 78, "n_ref_all": 147, "n_fig": 13, "n_tab": 3, "L_tab": 245, "n_element_tab": 67, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 193, "n_element_tab_1": 27, "formula_len_all": 7378, "formula_len_all_1": 937, "len_all": 250146, "len_all_1": 75215, "len_abs": 1146, "len_title": 110, "len_sents": 77212, "len_sents_1": 32592, "n_sents": 703, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1144, "title": "Understanding Optimization in Deep Learning with Central Flows", "abs": "Optimization in deep learning remains poorly understood, even in the simple setting of deterministic (i.e. full-batch) training. A key difficulty is that much of an optimizer's behavior is *implicitly* determined by complex oscillatory dynamics, referred to as the \"edge of stability.\" The main contribution of this paper is to show that an optimizer's implicit behavior can be explicitly captured by a *central flow*: a differential equation which models the time-averaged optimization trajectory. These flows can empirically predict long-term optimization trajectories of generic neural networks with an unprecedentedly high degree of numerical accuracy. By interpreting these flows, we reveal for the first time 1) the precise sense in which RMSProp adapts to the local loss landscape, and 2) an *acceleration via regularization* mechanism, wherein adaptive optimizers implicitly navigate towards low-curvature regions in which they can take larger steps. This mechanism is key to the efficacy of these adaptive optimizers. Overall, we believe that central flows constitute a promising tool for reasoning about optimization in deep learning.", "keywords": [ "Edge of Stability", "Optimization Dynamics", "Adaptive Optimizers", "RMSProp" ], "rating_list": [ 8, 8, 5, 6, 8 ], "soundness_list": [ 2, 4, 3, 3, 3 ], "presentation_list": [ 4, 4, 2, 4, 4 ], "contribution_list": [ 4, 4, 2, 3, 4 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "sHmfmQjfCy", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Aidan Curtis", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Eric Li", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Hui Li", "gender": "unknown", "institution": "Autodesk", "country": "", "position": "Sr. Principal Scientist" }, { "name": "Leslie Kaelbling", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Michael Noseworthy", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Nicole E Carey", "gender": "Female", "institution": "Autodesk", "country": "US", "position": "Principal Researcher" }, { "name": "Nishad Gothoskar", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Sachin Chitta", "gender": "Male", "institution": "Autodesk", "country": "US", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 36, "n_ref": 62, "n_ref_all": 77, "n_fig": 6, "n_tab": 1, "L_tab": 59, "n_element_tab": 8, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 650, "formula_len_all_1": 650, "len_all": 118347, "len_all_1": 58594, "len_abs": 1061, "len_title": 119, "len_sents": 37757, "len_sents_1": 29259, "n_sents": 262, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1015, "title": "Flow-based Maximum Entropy Domain Randomization for Multi-step Assembly", "abs": "Domain randomization in reinforcement learning is an established technique for increasing the robustness of control policies learned in simulation. By randomizing properties of the environment during training, the learned policy can be conformant to uncertainty along the randomized dimensions. While the environment distribution is typically specified by hand, in this paper we investigate the problem of automatically discovering this sampling distribution via entropy-regularized reward maximization of a neural sampling distribution in the form of a normalizing flow. We show that this architecture is more flexible and results in better robustness than existing approaches to learning simple parameterized sampling distributions. We demonstrate that these policies can be used to learn robust policies for contact-rich assembly tasks. Additionally, we explore how these sampling distributions can be used for out-of-distribution detection in the context of an uncertainty-aware multi-step manipulation planner.", "keywords": [ "Reinforcement Learning", "Domain Randomization", "Uncertainty", "Assembly", "Planning" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "sHbE7PplDG", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Alan Bovik", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Full Professor" }, { "name": "Ru-Ling Liao", "gender": "Female", "institution": "Alibaba Group", "country": "US", "position": "Researcher" }, { "name": "Shreshth Saini", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Yan Ye", "gender": "unknown", "institution": "", "country": "", "position": "" } ], "n_formula": 62, "n_formula_1": 21, "n_ref_uni": 45, "n_ref": 108, "n_ref_all": 135, "n_fig": 5, "n_tab": 11, "L_tab": 3570, "n_element_tab": 495, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2668, "n_element_tab_1": 306, "formula_len_all": 4933, "formula_len_all_1": 1537, "len_all": 221783, "len_all_1": 57274, "len_abs": 1504, "len_title": 139, "len_sents": 58184, "len_sents_1": 22449, "n_sents": 450, "n_sents_1": 150, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1371, "title": "PCDM: PERCEPTUAL CONSISTENCY IN DIFFUSION MODELS FOR NO-REFERENCE IMAGE QUALITY ASSESSMENT", "abs": "Despite recent advancements in latent diffusion models for generating high-dimensional data and performing various downstream tasks, there has been little exploration into perceptual consistency within these models for No-Reference Image Quality Assessment (NR-IQA). In this paper, we hypothesize that latent diffusion models implicitly exhibit perceptually consistent local regions within data manifold. We leverage this insight to guide the on-manifold sampling using perceptual features and input measurements. Specifically, we propose Perceptual Manifold Guidance (PMG), an algorithm that utilizes pretrained latent diffusion models and perceptual quality metrics to obtain perceptually consistent multi-scale and multi-timestep feature maps from the denoising U-Net. We empirically demonstrate that these hyperfeatures exhibit high correlation with human perception in IQA tasks. Our method can be applied to any existing pretrained latent diffusion model and is straightforward to integrate. To the best of our knowledge, this paper is the first work to explore Perceptual Consistency in Diffusion Models (PCDM) and apply it to NR-IQA in a zero-shot setting. Extensive experiments on IQA datasets show that our method, PCDM, achieves state-of-the-art performance, underscoring the superior zero-shot generalization capabilities of diffusion models for NR-IQA tasks.", "keywords": [ "No Reference Image Quality Assessment", "Perceptual Image Quality", "Latent Diffusion Models", "Training-free" ], "rating_list": [ 1, 5, 6, 5 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "sHUstMPM6Z", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 12, "author_info_list": [ { "name": "Elvis S. Liu", "gender": "Male", "institution": "Tencent", "country": "HK", "position": "Principal Researcher" }, { "name": "Hailong Wang", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Peipei Jin", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Qixian Zhou", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "MS student" }, { "name": "Ruochen Liu", "gender": "Male", "institution": "Columbia University", "country": "", "position": "MS student" }, { "name": "Tianzhou Wang", "gender": "Male", "institution": "Tencent MoreFun Studios", "country": "CN", "position": "Researcher" }, { "name": "Wenya Wei", "gender": "Female", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Yan Jiang", "gender": "unknown", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Yifu Yuan", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Yongle Luo", "gender": "unknown", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "WangtongLiu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "XueleiZhang", "gender": "unknown", "institution": "Tencent Technology (Shenzhen) Co., Ltd.", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 40, "n_ref": 52, "n_ref_all": 72, "n_fig": 10, "n_tab": 5, "L_tab": 2171, "n_element_tab": 199, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1201, "n_element_tab_1": 113, "formula_len_all": 578, "formula_len_all_1": 253, "len_all": 154729, "len_all_1": 63951, "len_abs": 1898, "len_title": 89, "len_sents": 42149, "len_sents_1": 29510, "n_sents": 299, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 41, "L_abs": 1972, "title": "Revolutionizing AI Companion in FPS Games", "abs": "Traditionally, players in first-person shooter (FPS) games have been limited to communicating with AI companions using simple commands like \u201cattack,\u201d \u201cdefend,\u201d or \u201cretreat\u201d due to the constraints of existing input methods such as hotkeys and command wheels. One major limitation of these simple commands is the lack of target specificity, as the numerous targets in a 3D virtual environment are difficult to specify using existing input methods. This limitation hinders players\u2019 ability to issue complex tactical instructions such as \u201cclear the second floor,\u201d \u201ctake cover behind that tree,\u201d or \u201cretreat to the river.\u201d To overcome this limitation, this paper introduces the $\\textbf{A}$I $\\textbf{C}$ompanion with $\\textbf{V}$oice $\\textbf{I}$nteraction $(\\textbf{ACVI})$, the first-ever AI system that allows players to interact with FPS AI companions through natural language. Deployed in the popular FPS game $\\textit{Arena Breakout: Infinite}$, this revolutionary feature creates the most immersive experience for players, enabling them to work with human-like AI. ACVI is not confined to executing limited commands through simple rule-based systems. Instead, it allows players to engage in real-time voice interactions with AI teammates. By integrating various natural language processing techniques within a confidence-based selection framework, it achieves rapid and accurate decomposition of complex commands and intent reasoning. Moreover, ACVI employs a multi-modal dynamic entity retrieval method for environmental perception, aligning human intentions with decision-making elements. It can accurately comprehend complex voice commands and delivers real-time behavioral responses and vocal feedback to provide close tactical collaboration to players. Additionally, it can identify more than 17,000 objects in the game, including buildings, vehicles, grasslands, and collectible items, and has the ability to accurately distinguish different colors and materials.", "keywords": [ "Human-AI collaboration", "AI Companion", "Game AI system" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "sHAvMp5J4R", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Chenxin An", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Lean Wang", "gender": "Male", "institution": "DeepSeek", "country": "CN", "position": "Intern" }, { "name": "Lingpeng Kong", "gender": "Male", "institution": "Department of Computer Science, The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Linli Yao", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Peiyuan Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Qi Liu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Xu Sun", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuanxin Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Tobias Lee", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 1, "n_ref_uni": 30, "n_ref": 94, "n_ref_all": 140, "n_fig": 10, "n_tab": 13, "L_tab": 8621, "n_element_tab": 802, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 2648, "n_element_tab_1": 335, "formula_len_all": 550, "formula_len_all_1": 84, "len_all": 254833, "len_all_1": 77999, "len_abs": 1970, "len_title": 94, "len_sents": 74092, "len_sents_1": 33286, "n_sents": 593, "n_sents_1": 289, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 46, "L_abs": 1707, "title": "Temporal Reasoning Transfer from Text to Video", "abs": "Video Large Language Models (Video LLMs) have shown promising capabilities in video comprehension, yet they struggle with tracking temporal changes and reasoning about temporal relationships.\nWhile previous research attributed this limitation to the ineffective temporal encoding of visual inputs, our diagnostic study reveals that video representations contain sufficient information for even small probing classifiers to achieve perfect accuracy.\nSurprisingly, we find that the key bottleneck in Video LLMs' temporal reasoning capability stems from the underlying LLM's inherent difficulty with temporal concepts, as evidenced by poor performance on textual temporal question-answering tasks.\nBuilding on this discovery, we introduce the Textual Temporal reasoning Transfer (T3). \nT3 synthesizes diverse temporal reasoning tasks in pure text format from existing image-text datasets, addressing the scarcity of video samples with complex temporal scenarios. \nRemarkably, without using any video data, T3 enhances LongVA-7B's temporal understanding, yielding a 5.3 absolute accuracy improvement on the challenging TempCompass benchmark, which enables our model to outperform ShareGPT4Video-8B trained on 28,000 video samples.\nAdditionally, the enhanced LongVA-7B model achieves competitive performance on comprehensive video benchmarks. For example, it achieves a 49.7 accuracy on the Temporal Reasoning task of Video-MME, surpassing powerful large-scale models such as InternVL-Chat-V1.5-20B and VILA1.5-40B. \nFurther analysis reveals a strong correlation between textual and video temporal task performance, validating the efficacy of transferring temporal reasoning abilities from text to video domains.", "keywords": [ "Video Large Language Models", "Temporal Reasoning" ], "rating_list": [ 8, 6, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "sGqd1tF8P8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Leitian Tao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Yixuan Li", "gender": "Female", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 79, "n_ref": 128, "n_ref_all": 158, "n_fig": 6, "n_tab": 12, "L_tab": 1259, "n_element_tab": 148, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1090, "n_element_tab_1": 52, "formula_len_all": 546, "formula_len_all_1": 469, "len_all": 222016, "len_all_1": 71981, "len_abs": 1107, "len_title": 104, "len_sents": 59886, "len_sents_1": 31934, "n_sents": 497, "n_sents_1": 267, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1196, "title": "Your Weak LLM is Secretly a Strong Teacher for Alignment", "abs": "The burgeoning capabilities of large language models (LLMs) have underscored the need for alignment to ensure these models act in accordance with human values and intentions. Existing alignment frameworks present constraints either in the form of expensive human effort or high computational costs. This paper explores a promising middle ground, where we employ a weak LLM that is significantly less resource-intensive than top-tier models, yet offers more automation than purely human feedback. We present a systematic study to evaluate and understand weak LLM's ability to generate feedback for alignment. Our empirical findings demonstrate that weak LLMs can provide feedback that rivals or even exceeds that of fully human-annotated data. Our study indicates a minimized impact of model size on feedback efficacy, shedding light on a scalable and sustainable alignment strategy. To deepen our understanding of alignment under weak LLM feedback, we conduct a series of qualitative and quantitative analyses, offering novel insights into the quality discrepancies between human feedback vs. weak LLM feedback. Code is publicly available at https://github.com/deeplearning-wisc/weak_llm_teacher.", "keywords": [ "large language models; alignment" ], "rating_list": [ 6, 10, 6, 6, 6 ], "soundness_list": [ 3, 4, 3, 4, 3 ], "presentation_list": [ 3, 4, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "sGpCzsfd1K", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Bo Shui", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "MS student" }, { "name": "Cheng Yang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Chufan Shi", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Deng Cai", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Research Scientist" }, { "name": "Gongye Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Junjie Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Linran XU", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "Undergrad student" }, { "name": "Mohan Jing", "gender": "unknown", "institution": "University of Electronic Science and Technology of China", "country": "", "position": "Undergrad student" }, { "name": "Siheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xiaomei Nie", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Xinyu Zhu", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Yaxin Liu", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yujiu Yang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yuxiang Zhang", "gender": "unknown", "institution": "Waseda University", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 50, "n_ref": 106, "n_ref_all": 137, "n_fig": 19, "n_tab": 10, "L_tab": 8536, "n_element_tab": 1165, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 5363, "n_element_tab_1": 922, "formula_len_all": 16, "formula_len_all_1": 13, "len_all": 261054, "len_all_1": 80715, "len_abs": 1449, "len_title": 139, "len_sents": 77286, "len_sents_1": 30761, "n_sents": 613, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1423, "title": "ChartMimic: Evaluating LMM's Cross-Modal Reasoning Capability via Chart-to-Code Generation", "abs": "We introduce a new benchmark, ChartMimic, aimed at assessing the visually-grounded code generation capabilities of large multimodal models (LMMs). ChartMimic utilizes information-intensive visual charts and textual instructions as inputs, requiring LMMs to generate the corresponding code for chart rendering.\nChartMimic includes $4,800$ human-curated (figure, instruction, code) triplets, which represent the authentic chart use cases found in scientific papers across various domains (e.g., Physics, Computer Science, Economics, etc). These charts span $18$ regular types and $4$ advanced types, diversifying into $201$ subcategories.\nFurthermore, we propose multi-level evaluation metrics to provide an automatic and thorough assessment of the output code and the rendered charts.\nUnlike existing code generation benchmarks, ChartMimic places emphasis on evaluating LMMs' capacity to harmonize a blend of cognitive capabilities, encompassing visual understanding, code generation, and cross-modal reasoning. The evaluation of $3$ proprietary models and $14$ open-weight models highlights the substantial challenges posed by ChartMimic. Even the advanced GPT-4o, InternVL2-Llama3-76B only achieve an average score of $82.2$ and $61.6$, respectively, indicating significant room for improvement. \nWe anticipate that ChartMimic will inspire the development of LMMs, advancing the pursuit of artificial general intelligence.", "keywords": [ "Dataset and Benchmark", "Code generation", "Chart Understand and Reasoning" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 4, 3 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "sGfVBi15uY", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Fanzeng Xia", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Hao Liu", "gender": "Male", "institution": "California Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Tongxin Li", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "Assistant Professor" }, { "name": "Yisong Yue", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 5, "n_ref_uni": 38, "n_ref": 97, "n_ref_all": 141, "n_fig": 17, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 15, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1165, "formula_len_all_1": 295, "len_all": 185264, "len_all_1": 66183, "len_abs": 2030, "len_title": 114, "len_sents": 70048, "len_sents_1": 32909, "n_sents": 521, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1963, "title": "Beyond Numeric Awards: In-Context Dueling Bandits with LLM Agents", "abs": "In-context reinforcement learning (ICRL) is a frontier paradigm for solving reinforcement learning problems in the foundation model era. While ICRL capabilities have been demonstrated in transformers through task-specific training, the potential of Large Language Models (LLMs) out-of-the-box remains largely unexplored. Recent findings highlight that LLMs often face challenges when dealing with numerical contexts, and limited attention has been paid to evaluating their performance through preference feedback generated by the environment. This paper is the first to investigate the performance of LLMs as in-context decision-makers in the problem of Dueling Bandits (DB), a stateless preference-based reinforcement learning setting that extends the classic Multi-Armed Bandit (MAB) model by querying for preference feedback. We compare GPT-3.5 Turbo, GPT-4, GPT-4 Turbo, Llama 3.1, and o1-preview against nine well-established DB algorithms. Our results reveal that our top-performing LLM, GPT-4 Turbo, possesses an understanding of relative decision-making sufficient to achieve low weak regret in DB by quickly selecting the best arm in duels. However, we observed that an optimality gap exists between LLMs and classic DB algorithms in terms of strong regret. LLMs struggle to converge and consistently exploit even when explicitly prompted to do so, and are sensitive to prompt variations. To overcome these issues, we introduce an agentic flow framework: LLM with Enhanced Algorithmic Dueling (LEAD), which integrates off-the-shelf DB algorithms with LLM agents through fine-grained adaptive interplay. We show that LEAD has theoretical guarantees inherited from classic DB algorithms on both weak and strong regret. We validate its efficacy and robustness even with noisy and adversarial prompts. The design of such an agentic framework sheds light on how to enhance the trustworthiness of general-purpose LLMs used for in-context decision-making tasks.", "keywords": [ "Large Language Models (LLMs)", "In-Context Decision-Making", "Dueling Bandits" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "sFGMkoBjUe", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Alex Goodall", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "PhD student" }, { "name": "Francesco Belardinelli", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Lecturer" } ], "n_formula": 61, "n_formula_1": 21, "n_ref_uni": 67, "n_ref": 160, "n_ref_all": 174, "n_fig": 3, "n_tab": 2, "L_tab": 594, "n_element_tab": 1, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4579, "formula_len_all_1": 2020, "len_all": 218110, "len_all_1": 72837, "len_abs": 816, "len_title": 123, "len_sents": 67802, "len_sents_1": 33631, "n_sents": 467, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 817, "title": "Lookahead Shielding for Regular Safety Properties in Reinforcement Learning", "abs": "To deploy reinforcement learning (RL) systems in real-world scenarios we need to consider requirements such as safety and constraint compliance, rather than blindly maximizing for reward. In this paper we develop a lookahead shielding framework for RL with regular safety properties, which on the contrary to prior shielding methodologies requires minimal prior knowledge. At each environment step our framework aims to satisfy the regular safety property for a bounded horizon with high-probability, for the tabular setting we provide provable guarantees. We compare our setup to some common algorithms developed for the constrained Markov decision process (CMDP), and we demonstrate the effectiveness and scalability of our framework by extensively evaluating our framework in both tabular and deep RL environments.", "keywords": [ "Safe Reinforcement Learning", "Model Checking", "Shielding" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "sF8jmiD8Bq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Howe Tissue", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Lu Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Mozhi Zhang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Xipeng Qiu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 43, "n_ref": 114, "n_ref_all": 155, "n_fig": 9, "n_tab": 3, "L_tab": 2153, "n_element_tab": 367, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 2132, "n_element_tab_1": 325, "formula_len_all": 776, "formula_len_all_1": 743, "len_all": 230666, "len_all_1": 81706, "len_abs": 1663, "len_title": 130, "len_sents": 50369, "len_sents_1": 39461, "n_sents": 300, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1634, "title": "Domain2Vec: Vectorizing Datasets to Find the Optimal Data Mixture without Training", "abs": "The mixture ratio of data from different source domains significantly affects the performance of language models (LM) pretraining. In this paper, we introduce~\\textsc{Domain2Vec}, a novel approach that decomposes any dataset into a linear combination of several ``Meta-Domains'', a new concept designed to capture key underlying features of datasets. \\textsc{Domain2Vec} maintains a vocabulary of Meta-Domains and uses a Meta-Domain Classifier to decompose any given dataset into a domain vector that corresponds to a distribution over this vocabulary. These domain vectors enable the identification of optimal data mixture ratio for LM pretraining in a training-free manner under the \\textit{\\textbf{D}istribution \\textbf{A}lignment \\textbf{A}ssumption} (DA$^{2}$). Moreover, previous work could use \\textsc{Domain2vec} to model the relationship between domain vectors and LM performance, greatly enhancing the scalability of previous methods without retraining as new datasets are introduced. Extensive experiments demonstrate that \\textsc{Domain2Vec} finds data mixture ratios that enhance downstream task performance with minimal computational overhead. Specifically, \\textsc{Domain2Vec} achieves the same validation loss on Pile-CC using only $51.5\\%$ of the compute required when training on the original mixture of The Pile Dataset. Under equivalent compute budget, \\textsc{Domain2Vec} improves downstream performance by an average of $2.72\\%$. \\textsc{Domain2Vec} serves as a strong and efficient baseline for data mixture optimization in LM pretraining, offering insights into improving data efficiency in large-scale models.", "keywords": [ "Language models", "Pretraining", "Data mixutre" ], "rating_list": [ 6, 8, 8, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "sEv6vHIUnu", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Aditya Mohan", "gender": "Male", "institution": "Universit\u00e4t Hannover", "country": "DE", "position": "PhD" }, { "name": "Marius Lindauer", "gender": "Male", "institution": "Leibniz Universit\u00e4t Hannover", "country": "DE", "position": "Full Professor" } ], "n_formula": 24, "n_formula_1": 15, "n_ref_uni": 67, "n_ref": 104, "n_ref_all": 115, "n_fig": 6, "n_tab": 3, "L_tab": 2, "n_element_tab": 1, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 15, "n_element_tab_1": 12, "formula_len_all": 1843, "formula_len_all_1": 1055, "len_all": 181243, "len_all_1": 59191, "len_abs": 1126, "len_title": 63, "len_sents": 43238, "len_sents_1": 27274, "n_sents": 363, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1129, "title": "Structured Predictive Representations in Reinforcement Learning", "abs": "Reinforcement Learning (RL) remains brittle in complex environments characterized by sparse rewards, partial observability, and subtask dependencies. Predictive state abstractions capture the environment's underlying temporal structure and are crucial to overcoming these challenges. Yet, such methods often only focus on global one-step transitions and overlook local relationships between trajectories. This paper explores how capturing such relationships can enhance representation learning methods in RL. Our primary contribution is to show that incorporating a Graph-Neural Network (GNN) into the observation-predictive learning process improves sample efficiency and robustness to changes in size and distractors. Through experiments on the MiniGrid suite, we demonstrate that our GNN-based approach outperforms typical models that use Multi-layer Perceptrons (MLPs) in sparse reward and partially observable environments where task decomposition are critical. These results highlight the value of structural inductive biases for generalization and adaptability, revealing how such mechanisms can bolster performance in RL.", "keywords": [ "Representation Learning; State Abstractions; Reinforcement Learning; Self-Prediction" ], "rating_list": [ 3, 3, 5, 5, 8 ], "soundness_list": [ 2, 1, 3, 3, 3 ], "presentation_list": [ 3, 1, 2, 3, 4 ], "contribution_list": [ 2, 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "sEMJ1PLSZR", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Do-Yeon Kim", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Dong-Jun Han", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Jaekyun Moon", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "Full Professor" }, { "name": "Jungmoon Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jungwuk Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Wonjeong Choi", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Younghyun Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 39, "n_ref": 96, "n_ref_all": 148, "n_fig": 18, "n_tab": 14, "L_tab": 12194, "n_element_tab": 1083, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 2178, "n_element_tab_1": 181, "formula_len_all": 644, "formula_len_all_1": 486, "len_all": 193164, "len_all_1": 70132, "len_abs": 1647, "len_title": 111, "len_sents": 58303, "len_sents_1": 32028, "n_sents": 414, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1418, "title": "Adaptive Energy Alignment for Accelerating Test-Time Adaptation", "abs": "In response to the increasing demand for tackling out-of-domain (OOD) scenarios, test-time adaptation (TTA) has garnered significant research attention in recent years. To adapt a source pre-trained model to target samples without getting access to their labels, existing approaches have typically employed entropy minimization (EM) loss as a primary objective function. In this paper, we propose an adaptive energy alignment (AEA) solution that achieves fast online TTA. We start from the re-interpretation of the EM loss by decomposing it into two energy-based terms with conflicting roles, showing that the EM loss can potentially hinder the assertive model adaptation. Our AEA addresses this challenge by strategically reducing the energy gap between the source and target domains during TTA, aiming to effectively align the target domain with the source domains and thus to accelerate adaptation. We specifically propose two novel strategies, each contributing a necessary component for TTA: (i) aligning the energy level of each target sample with the energy zone of the source domain that the pre-trained model is already familiar with, and (ii) precisely guiding the direction of the energy alignment by matching the class-wise correlations between the source and target domains. Our approach demonstrates its effectiveness on various domain shift datasets including CIFAR10-C, CIFAR100-C, and TinyImageNet-C.", "keywords": [ "Test-Time Adaptation (TTA)", "Out-of-domain (OOD)", "Energy-based Model (EBM)" ], "rating_list": [ 8, 6, 8, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "sELO2DCCC1", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Jian Lu", "gender": "unknown", "institution": "Pacific Northwest National Laboratory", "country": "US", "position": "Principal Researcher" }, { "name": "Parvathi Kooloth", "gender": "unknown", "institution": "Pacific Northwest National Laboratory", "country": "US", "position": "Postdoc" }, { "name": "Tse-Chun Chen", "gender": "unknown", "institution": "Pacific Northwest National Laboratory", "country": "US", "position": "Researcher" }, { "name": "Zhangshuan Hou", "gender": "Male", "institution": "Pacific Northwest National Laboratory", "country": "US", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 22, "n_ref": 27, "n_ref_all": 37, "n_fig": 6, "n_tab": 2, "L_tab": 660, "n_element_tab": 109, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 660, "n_element_tab_1": 109, "formula_len_all": 530, "formula_len_all_1": 530, "len_all": 71391, "len_all_1": 60069, "len_abs": 1388, "len_title": 118, "len_sents": 19620, "len_sents_1": 19683, "n_sents": 116, "n_sents_1": 117, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1384, "title": "ClimGen: Learning the Forcing-Response Relationship in Climate System", "abs": "Solar Radiation Management (SRM) is emerging as a viable geoengineering strategy to address the climate change crisis, but its effective implementation requires iterative and large ensemble of highly accurate and efficient climate projections. Traditional climate projections rely on executing computational demanding and time-consuming numerical climate models. Recent advances in machine learning (ML) aim to enhance these approaches by emulating traditional methods. In this work, we propose a novel framework for directly learning the relationship between solar radiation flux at the top of the atmosphere and the corresponding surface temperature response.\n\nTo evaluate the feasibility of this direct ML-based projection, we developed a benchmark dataset using an intermediate complexity model, incorporating a comprehensive suite of different forcing patterns and evaluation metrics to rigorously assess the ML model\u2019s performance. We introduce a Conditional Denoising Diffusion Probabilistic Model (c-DDPM) for this task, which demonstrates superior performance in representing climate statistics under previously unseen forcing patterns. This approach provides a promising pathway for direct climate projections by accurately learning the forcing-response relationship, with wide range of applications in climate change mitigation, emissions policy design, and SRM strategies.", "keywords": [ "GenAI", "c-DDPM", "Climate", "Forcing-Response relationship" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 2, 1, 1, 2 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 2, 3, 1, 1 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "sEJYPiVEt4", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Guanyu Chen", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Qifeng Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xiaosong Jia", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "zhiyuan zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 35, "n_ref": 77, "n_ref_all": 99, "n_fig": 9, "n_tab": 10, "L_tab": 2511, "n_element_tab": 268, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1898, "n_element_tab_1": 215, "formula_len_all": 251, "formula_len_all_1": 251, "len_all": 145229, "len_all_1": 66867, "len_abs": 2184, "len_title": 77, "len_sents": 42931, "len_sents_1": 28828, "n_sents": 380, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1775, "title": "ESDMotion: End-to-end Motion Prediction Only with SD Maps", "abs": "Motion prediction is a crucial task in autonomous driving. Existing motion prediction models rely on high-definition (HD) maps to provide environmental context for agents. However, offline HD maps require extensive manual annotation, making them costly and unscalable. Online mapping-based methods still require HD map annotation to train the online mapping module, which is costly as well and may also suffer from the issue of out-of-distribution map elements.\nIn this work, we explore conducting motion prediction with standard-definition (SD) maps as substitution, which are more readily available and offer broader coverage. One crucial challenge is that SD maps have low resolution and poor alignment accuracy. Directly replacing HD maps with SD maps leads to a significant drop in performance. \nWe introduce end-to-end learning and specially tailored modules for SD maps to solve the problems. Specifically, we propose ESDMotion, the first end-to-end motion prediction framework that uses SD maps without any HD map supervision. We integrate BEV features obtained from raw sensor data into existing motion prediction models, with tailored designs for anchor-based and anchor-free models respectively. We find that the coarse and misaligned SD maps bring challenges to feature fusion of anchor-free model and on anchor generation of anchor-based model. Thus, we design two novel modules named Enhanced Road Observation and Pseudo Lane Expansion to address these issues. Benefiting from the end-to-end structure and new modules, ESDMotion outperforms the state-of-the-art online mapping-based motion prediction methods by 13.4\\% in motion prediction performance and narrows the performance gap between HD and SD maps by 73\\%. We will open source our code and checkpoints.", "keywords": [ "motion prediction" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "sEARCNzhrP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Jinrong Zhang", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "MS student" }, { "name": "Lin Feng", "gender": "Male", "institution": "Dalian Minzu University", "country": "CN", "position": "Full Professor" }, { "name": "Minjie Xu", "gender": "unknown", "institution": "Dalian University of Technology", "country": "CN", "position": "MS student" }, { "name": "Shenglan Liu", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 35, "n_ref": 86, "n_ref_all": 106, "n_fig": 7, "n_tab": 9, "L_tab": 2811, "n_element_tab": 170, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 643, "n_element_tab_1": 39, "formula_len_all": 902, "formula_len_all_1": 772, "len_all": 134086, "len_all_1": 66599, "len_abs": 1932, "len_title": 150, "len_sents": 38666, "len_sents_1": 29111, "n_sents": 321, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1720, "title": "Interaction Makes Better Segmentation: An Interaction-based Framework for Temporal Action Segmentation", "abs": "Temporal action segmentation aims to classify the action category of each frame in untrimmed videos, primarily using RGB video and skeleton data. Most existing methods adopt a two-stage process: feature extraction and temporal modeling. However, we observe significant limitations in their spatio-temporal modeling: (i) Existing temporal modeling modules conduct frame-level and action-level interactions at a fixed temporal resolution, which over-smooths temporal features and leads to blurred action boundaries; (ii) Skeleton-based methods generally adopt temporal modeling modules originally designed for RGB video data, causing a misalignment between extracted features and temporal modeling modules. In this paper, we propose a novel Interaction-based framework for Action segmentation (InterAct) to address these issues. Firstly, we propose multi-scale frame-action interaction (MFAI) to facilitate frame-action interactions across varying temporal scales. This enhances the model's ability to capture complex temporal dynamics, producing more expressive temporal representations and alleviating the over-smoothing issue. Meanwhile, recognizing the complementary nature of different spatial modalities, we propose decoupled spatial modality interaction (DSMI). It decouples the modeling of spatial modalities and applies a deep fusion strategy to interactively integrate multi-scale spatial features. This results in more discriminative spatial features that are better aligned with the temporal modeling modules. Extensive experiments on six large-scale benchmarks demonstrate that InterAct significantly outperforms state-of-the-art methods on both RGB-based and skeleton-based datasets across diverse scenarios.", "keywords": [ "Video Understanding; Video Analysis;" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "sClhxLqfnP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Guyue Zhou", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Hao Huang", "gender": "unknown", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Hao Zhao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Haoran Xu", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "He xiang", "gender": "Male", "institution": "haomo", "country": "CN", "position": "Researcher" }, { "name": "Huan-ang Gao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jv zheng", "gender": "Female", "institution": ", Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Kangliang Chen", "gender": "Male", "institution": "haomo.ai", "country": "CN", "position": "Researcher" }, { "name": "Weihao Gu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiaoxue Chen", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yaqin Zhang", "gender": "unknown", "institution": "Beijing University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 13, "n_ref": 29, "n_ref_all": 43, "n_fig": 6, "n_tab": 5, "L_tab": 1020, "n_element_tab": 53, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 378, "n_element_tab_1": 45, "formula_len_all": 593, "formula_len_all_1": 593, "len_all": 111515, "len_all_1": 58602, "len_abs": 0, "len_title": 266, "len_sents": 27468, "len_sents_1": 28133, "n_sents": 195, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1426, "title": "RGM: Reconstructing High-fidelity 3D Car Assets with Relightable 3D-GS Generative Model from a Single Image", "abs": "The generation of high-quality 3D car assets is essential for various applications, including video games, autonomous driving, and virtual reality. Current 3D generation methods utilizing NeRF or 3D-GS as representations for 3D objects, generate a Lambertian object under fixed lighting and lack separated modelings for material and global illumination. This results in generated assets unsuitable for relighting under different illuminations and limiting their application to downstream tasks. To address this challenge, we propose a novel relightable 3D object generative framework that automates the creation of 3D car assets, enabling the swift and accurate reconstruction of a vehicle's geometry, texture, and material properties from a single input image. Our approach begins with introducing a large-scale synthetic car dataset comprising over 1,000 high-precision 3D vehicle models. We represent 3D objects using global illumination and relightable 3D Gaussian primitives integrating with BRDF parameters. Building on this representation, we introduce a feed-forward model that takes images as input and outputs both relightable 3D Gaussians and global illumination parameters. Experimental results demonstrate that our method produces photorealistic 3D car assets that can be seamlessly integrated into road scenes with different illuminations, which offers substantial practical benefits for industrial applications.", "keywords": [ "Generative Model", "3D Reconstruction", "3D Gaussian Splatting" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 3, 3, 3, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "sCew1tR6No", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Abihith Kothapalli", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "MS student" }, { "name": "Hengrong Du", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Assistant Professor" }, { "name": "Soheil Kolouri", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "Assistant Professor" }, { "name": "Xinran Liu", "gender": "unknown", "institution": "Vanderbilt University", "country": "US", "position": "PhD student" }, { "name": "Yikun Bai", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "Postdoc" }, { "name": "ROCIO DIAZ MARTIN", "gender": "Female", "institution": "Tufts University", "country": "US", "position": "Postdoc" } ], "n_formula": 244, "n_formula_1": 40, "n_ref_uni": 54, "n_ref": 154, "n_ref_all": 224, "n_fig": 12, "n_tab": 6, "L_tab": 3008, "n_element_tab": 455, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 251, "n_element_tab_1": 38, "formula_len_all": 30064, "formula_len_all_1": 3852, "len_all": 331772, "len_all_1": 63153, "len_abs": 2932, "len_title": 81, "len_sents": 85674, "len_sents_1": 24823, "n_sents": 995, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 33, "L_abs": 1358, "title": "Partial Gromov-Wasserstein Metric", "abs": "The Gromov-Wasserstein (GW) distance has gained increasing interest in the machine learning community in recent years, as it allows for the comparison of measures in different metric spaces. To overcome the limitations imposed by the equal mass requirements of the classical GW problem, researchers have begun exploring its application in unbalanced settings. However, Unbalanced GW (UGW) can only be regarded as a discrepancy rather than a rigorous metric/distance between two metric measure spaces (mm-spaces). In this paper, we propose a particular case of the UGW problem, termed Partial Gromov-Wasserstein (PGW). We establish that PGW is a well-defined metric between mm-spaces and discuss its theoretical properties, including the existence of a minimizer for the PGW problem and the relationship between PGW and GW, among others. We then propose two variants of the Frank-Wolfe algorithm for solving the PGW problem and show that they are mathematically and computationally equivalent. Moreover, based on our PGW metric, we introduce the analogous concept of barycenters for mm-spaces. Finally, we validate the effectiveness of our PGW metric and related solvers in applications such as shape matching, shape retrieval, and shape interpolation, comparing them against existing baselines. Our code is available at https://github.com/mint-vu/PGW_Metric.", "keywords": [ "Optimal transport", "Gromov-Wasserstein problem", "Unbalanced optimal transport" ], "rating_list": [ 8, 8, 6, 6, 6 ], "soundness_list": [ 3, 4, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "sCGIbhv4Yv", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Qixing Huang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Ruisheng Wang", "gender": "Not Specified", "institution": "University of Calgary", "country": "CA", "position": "Full Professor" }, { "name": "Yunpeng Bai", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 34, "n_ref": 72, "n_ref_all": 88, "n_fig": 9, "n_tab": 4, "L_tab": 287, "n_element_tab": 49, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 287, "n_element_tab_1": 49, "formula_len_all": 623, "formula_len_all_1": 619, "len_all": 137085, "len_all_1": 65014, "len_abs": 916, "len_title": 113, "len_sents": 34644, "len_sents_1": 31693, "n_sents": 242, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 919, "title": "Regularized Diffusion Modeling for CAD Representation Generation", "abs": "Computer-Aided Design (CAD) has significant practical value in various industrial applications. However, achieving high-quality and diverse shape generation, as well as flexible conditional control, remains a challenge in the field of CAD model generation. To address these issues, we propose CADiffusion, a diffusion-based generative model with a hierarchical latent representation tailored to the complexities of CAD design processes. To enhance the performance and reliability of the model in generating accurate CAD models, we have developed a specialized decoder with regularization strategies that navigate through the noise space of the diffusion model, smoothing the results. This approach not only improves the diversity and quality of the generated CAD models but also enhances their practical applicability, marking a significant advancement in the integration of generative models and automated CAD systems.", "keywords": [ "CAD", "Diffusion Regularization", "3D Generation" ], "rating_list": [ 5, 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 3, 1, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 3, 5 ] }, { "paper_id": "sBpYRQOrMn", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Haibo Hu", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Li Liu", "gender": "Female", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "", "position": "Assistant Professor" }, { "name": "Qingqing Ye", "gender": "Female", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "Yanyun Wang", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "MPhil" }, { "name": "Zi Liang", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "CN", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 35, "n_ref": 152, "n_ref_all": 184, "n_fig": 11, "n_tab": 6, "L_tab": 2320, "n_element_tab": 203, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1639, "formula_len_all_1": 1403, "len_all": 181872, "len_all_1": 67309, "len_abs": 2124, "len_title": 165, "len_sents": 58686, "len_sents_1": 33769, "n_sents": 310, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 115, "L_abs": 2133, "title": "New Paradigm of Adversarial Training: Breaking Inherent Trade-Off between Accuracy and Robustness via Dummy Classes", "abs": "Adversarial Training (AT) is recognized as one of the most effective methods to enhance the robustness of Deep Neural Networks (DNNs). However, existing AT methods suffer from an inherent trade-off between adversarial robustness and clean accuracy, which seriously hinders their real-world deployment. Previous works have studied this trade-off within the current AT paradigm, exploring various factors such as perturbation intensity, label noise and class margin. Despite these efforts, current AT methods still typically experience a reduction in clean accuracy by over 10% to date, without significant improvements in robustness compared with simple baselines like PGD-AT. This inherent trade-off raises a question: whether the current AT paradigm, which assumes to learn the corresponding benign and adversarial samples as the same class, inappropriately combines clean and robust objectives that may be essentially inconsistent. In this work, we surprisingly reveal that up to 40% of CIFAR-10 adversarial samples always fail to satisfy such an assumption across various AT methods and robust models, explicitly indicating the improvement room for the current AT paradigm. Accordingly, to relax the tension between clean and robust learning derived from this overstrict assumption, we propose a new AT paradigm by introducing an additional dummy class for each original class, aiming to accommodate the hard adversarial samples with shifted distribution after perturbation. The robustness w.r.t. these adversarial samples can be achieved by runtime recovery from the predicted dummy classes to their corresponding original ones, eliminating the compromise with clean learning. Building on this new paradigm, we propose a novel plug-and-play AT technology named DUmmy Classes-based Adversarial Training (DUCAT). Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet demonstrate that the DUCAT concurrently improves clean accuracy and adversarial robustness compared with state-of-the-art benchmarks, effectively releasing the existing inherent trade-off. The code is available at https://anonymous.4open.science/r/DUCAT.", "keywords": [ "Adversarial Training", "Robustness" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "sBbarJBdkn", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Amit Dhurandhar", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Erik Miehling", "gender": "Male", "institution": "IBM Research", "country": "IE", "position": "Research Scientist" }, { "name": "Ronny Luss", "gender": "unknown", "institution": "IBM", "country": "US", "position": "Research Scientist" } ], "n_formula": 8, "n_formula_1": 4, "n_ref_uni": 43, "n_ref": 73, "n_ref_all": 99, "n_fig": 9, "n_tab": 6, "L_tab": 3154, "n_element_tab": 111, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1467, "n_element_tab_1": 56, "formula_len_all": 1328, "formula_len_all_1": 461, "len_all": 158243, "len_all_1": 69663, "len_abs": 1375, "len_title": 115, "len_sents": 52357, "len_sents_1": 33518, "n_sents": 398, "n_sents_1": 229, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1380, "title": "CELL your Model: Contrastive Explanations for Large Language Models", "abs": "The advent of black-box deep neural network classification models has sparked the need to explain their decisions. However, in the case of generative AI such as large language models (LLMs), there is no class prediction to explain. Rather, one can ask why an LLM output a particular response to a given prompt. In this paper, we answer this question by proposing, to the best of our knowledge, the first contrastive explanation methods requiring simply black-box/query access. Our explanations suggest that an LLM outputs a reply to a given prompt because if the prompt was slightly modified, the LLM would have given a different response that is either less preferable or contradicts the original response. The key insight is that contrastive explanations simply require a scoring function that has meaning to the user and not necessarily a specific real valued quantity (viz. class label). We offer two algorithms for finding contrastive explanations: i) A myopic algorithm, which although effective in creating contrasts, requires many model calls and ii) A budgeted algorithm, our main algorithmic contribution, which intelligently creates contrasts adhering to a query budget, necessary for longer contexts. We show the efficacy of these methods on diverse natural language tasks such as open-text generation, automated red teaming, and explaining conversational degradation.", "keywords": [ "Contrastive Explanations", "Large Language Models" ], "rating_list": [ 5, 3, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 5, 4 ] }, { "paper_id": "sBJIVQvJqN", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Martha Larson", "gender": "Female", "institution": "Radboud University", "country": "NL", "position": "Full Professor" }, { "name": "Mingliang Liang", "gender": "Male", "institution": "Institute for Computing and Information Sciences, Radboud University Nijmegen, Radboud University", "country": "NL", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 76, "n_ref_all": 117, "n_fig": 7, "n_tab": 18, "L_tab": 5139, "n_element_tab": 478, "n_fig_1": 2, "n_tab_1": 9, "L_tab_1": 3841, "n_element_tab_1": 264, "formula_len_all": 145, "formula_len_all_1": 117, "len_all": 149537, "len_all_1": 62877, "len_abs": 1207, "len_title": 145, "len_sents": 41658, "len_sents_1": 27502, "n_sents": 309, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1214, "title": "Enhancing Vision-Language Model Pre-training with Image-text Pair Pruning Based on Word Frequency", "abs": "We propose Word-Frequency-based Image-Text Pair Pruning (WFPP), a novel data pruning method that improves the efficiency of VLMs.\nUnlike MetaCLIP, our method does not need metadata for pruning, but selects text-image pairs to prune based on the content of the text. Specifically, WFPP prunes text-image pairs containing high-frequency words across the entire training dataset. The effect of WFPP is to reduce the dominance of frequent words. The result a better balanced word-frequency distribution in the dataset, which is known to improve the training of word embedding models. After pre-training on the pruned subset, we fine-tuned the model on the entire dataset for one additional epoch to achieve better performance. Our experiments demonstrate that applying WFPP when training a CLIP model improves performance on a wide range of downstream tasks. WFPP also provides the advantage of speeding up pre-training by using fewer samples. Additionally, we analyze the training data before and after pruning to visualize how WFPP changes the balance of word frequencies. We hope our work encourages researchers to consider the distribution of words in the training data when pre-training VLMs, not limited to CLIP.", "keywords": [ "Vision-Language Model", "Multimodal Data", "Data Pruning" ], "rating_list": [ 3, 8, 8, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "sAxdIJ4l6z", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Jianshu Zhang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Undergrad student" }, { "name": "Jipeng Zhang", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Renjie Pi", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Rui Pan", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Tianyang Han", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "", "position": "MS student" }, { "name": "Tong Zhang", "gender": "Male", "institution": "UIUC", "country": "US", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 39, "n_ref": 79, "n_ref_all": 103, "n_fig": 5, "n_tab": 5, "L_tab": 1181, "n_element_tab": 137, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 637, "n_element_tab_1": 86, "formula_len_all": 179, "formula_len_all_1": 179, "len_all": 161249, "len_all_1": 58203, "len_abs": 1329, "len_title": 86, "len_sents": 48961, "len_sents_1": 28396, "n_sents": 369, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 38, "L_abs": 1334, "title": "Personalized Visual Instruction Tuning", "abs": "Recent advancements in multimodal large language models (MLLMs) have demonstrated significant progress; however, these models exhibit a notable limitation, which we refer to as \"face blindness.\" Specifically, they can engage in general conversations but fail to conduct personalized dialogues targeting at specific individuals. This deficiency hinders the application of MLLMs in personalized settings, such as tailored visual assistants on mobile devices, or domestic robots that need to recognize members of the family. In this paper, we introduce Personalized Visual Instruction Tuning (PVIT), a novel data curation and training framework designed to enable MLLMs to identify target individuals within an image and engage in personalized and coherent dialogues. Our approach involves the development of a sophisticated pipeline that autonomously generates training data containing personalized conversations. This pipeline leverages the capabilities of various visual experts, image generation models, and (multi-modal) large language models. To evaluate the personalized potential of MLLMs, we present a benchmark called P-Bench, which encompasses various question types with different levels of difficulty. The experiments demonstrate a substantial personalized performance enhancement after fine-tuning with our curated dataset.", "keywords": [ "large language model", "multimodal large language model" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "sAp04DAHuY", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "Christian Wachinger", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "Professor" }, { "name": "Emre Kavak", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Undergrad student" }, { "name": "Tom Nuno Wolf", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 22, "n_ref_uni": 41, "n_ref": 90, "n_ref_all": 106, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2806, "formula_len_all_1": 1812, "len_all": 129840, "len_all_1": 64292, "len_abs": 1232, "len_title": 125, "len_sents": 41261, "len_sents_1": 29720, "n_sents": 333, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1237, "title": "DISCO: Mitigating Bias in Deep Learning with Conditional DIStance COrrelation", "abs": "During prediction tasks, models can use any signal they receive to come up with\nthe final answer - including signals that are causally irrelevant. When predicting\nobjects from images, for example, the lighting conditions could be correlated to\ndifferent targets through selection bias, and an oblivious model might use these\nsignals as shortcuts to discern between various objects. A predictor that uses\nlighting conditions instead of real object-specific details is obviously undesirable.\nTo address this challenge, we introduce a standard anti-causal prediction model\n(SAM) that creates a causal framework for analyzing the information pathways\ninfluencing our predictor in anti-causal settings. We demonstrate that a classifier\nsatisfying a specific conditional independence criterion will focus solely on the\ndirect causal path from label to image, being counterfactually invariant to the\nremaining variables. Finally, we propose DISCO, a novel regularization strategy\nthat uses conditional distance correlation to optimize for conditional independence\nin regression tasks. We can show that DISCO achieves competitive results in\ndifferent bias mitigation experiments, deeming it a valid alternative to classical\nkernel-based methods.", "keywords": [ "Causality", "Counterfactual Invariance", "Bias Mitigation", "Causal Inference" ], "rating_list": [ 5, 3, 8, 3 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 2, 2, 4, 2 ], "contribution_list": [ 2, 2, 4, 2 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "sAYnDWaGd5", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Chaochen Gao", "gender": "Male", "institution": "Institute of Information Engineering, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Qi Fu", "gender": "unknown", "institution": "Xiaohongshu Inc", "country": "", "position": "Researcher" }, { "name": "Xing W", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Songiln Hu", "gender": "Male", "institution": "Institute of Information Engeering, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 47, "n_ref": 83, "n_ref_all": 113, "n_fig": 8, "n_tab": 12, "L_tab": 4265, "n_element_tab": 561, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1596, "n_element_tab_1": 158, "formula_len_all": 599, "formula_len_all_1": 251, "len_all": 159834, "len_all_1": 57532, "len_abs": 985, "len_title": 142, "len_sents": 34966, "len_sents_1": 22995, "n_sents": 284, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 989, "title": "Quest: Query-centric Data Synthesis Approach for Long-context Scaling of Large Language Model", "abs": "Recent advancements in large language models (LLMs) have highlighted the importance of extending context lengths for handling complex tasks. While traditional methods for training on long contexts often use filtered long documents, these approaches lead to domain imbalances, limiting model performance. To address this, techniques like random document concatenation (Standard) and similarity-based methods (KNN, ICLM) have been developed. However, they either sacrifice semantic coherence or diversity. To balance both aspects, we introduce Quest, a query-centric data synthesis method aggregating semantically relevant yet diverse documents. Quest uses a generative model to predict potential queries for each document, grouping documents with similar queries and keywords. Extensive experiments demonstrate Quest's superior performance on long-context tasks, achieving remarkable results with context lengths of up to 1M tokens and confirming its scalability across various model sizes.", "keywords": [ "longcontext", "pre-training", "scaling" ], "rating_list": [ 8, 8, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 2 ] }, { "paper_id": "s9zoyICZ4k", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Chenhao Zhang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "PhD student" }, { "name": "Xia Jiang", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "PhD student" }, { "name": "Yingqian Zhang", "gender": "Female", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Associate Professor" }, { "name": "YAOXIN WU", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 26, "n_ref": 60, "n_ref_all": 88, "n_fig": 11, "n_tab": 6, "L_tab": 1000, "n_element_tab": 114, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 704, "n_element_tab_1": 56, "formula_len_all": 1102, "formula_len_all_1": 311, "len_all": 158504, "len_all_1": 65898, "len_abs": 1344, "len_title": 154, "len_sents": 53402, "len_sents_1": 32041, "n_sents": 383, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 1348, "title": "DRoC: Elevating Large Language Models for Complex Vehicle Routing via Decomposed Retrieval of Constraints", "abs": "This paper proposes Decomposed Retrieval of Constraints (DRoC), a novel framework aimed at enhancing large language models (LLMs) in exploiting solvers to tackle vehicle routing problems (VRPs) with intricate constraints. While LLMs have shown promise in solving simple VRPs, their potential in addressing complex VRP variants is still suppressed, due to the limited embedded internal knowledge that is required to accurately reflect diverse VRP constraints. Our approach mitigates the issue by integrating external knowledge via a novel retrieval-augmented generation (RAG) approach. More specifically, the DRoC decomposes VRP constraints, externally retrieves information relevant to each constraint, and synergistically combines internal and external knowledge to benefit the program generation for solving VRPs. The DRoC also allows LLMs to dynamically select between RAG and self-debugging mechanisms, thereby optimizing program generation without the need for additional training. Experiments across 48 VRP variants exhibit the superiority of DRoC, with significant improvements in the accuracy rate and runtime error rate delivered by the generated programs. The DRoC framework has the potential to elevate LLM performance in complex optimization tasks, fostering the applicability of LLMs in industries such as transportation and logistics.", "keywords": [ "vehicle routing problems; large language model; retrieval augmented generation" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "s9SVlWOcLt", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Harshit Sikchi", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Peter Stone", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Principal Researcher" }, { "name": "Siddhant Agarwal", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" }, { "name": "Amy Zhang", "gender": "Female", "institution": "University of Texas at Austin", "country": "", "position": "Assistant Professor" } ], "n_formula": 39, "n_formula_1": 13, "n_ref_uni": 36, "n_ref": 68, "n_ref_all": 85, "n_fig": 4, "n_tab": 6, "L_tab": 1939, "n_element_tab": 217, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3208, "formula_len_all_1": 1140, "len_all": 174168, "len_all_1": 75113, "len_abs": 2872, "len_title": 147, "len_sents": 48991, "len_sents_1": 33281, "n_sents": 399, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1095, "title": "Proto Successor Measure: Representing the space of all possible solutions of Reinforcement Learning", "abs": "Having explored an environment, intelligent agents should be able to transfer their knowledge to most downstream tasks within that environment. \nReferred to as ``zero-shot learning,\" this ability remains elusive for general-purpose reinforcement learning algorithms. While recent works have attempted to produce zero-shot RL agents, they make assumptions about the nature of the tasks or the structure of the MDP. We present \\emph{Proto Successor Measure}: the basis set for all possible solutions of Reinforcement Learning in a dynamical system. We provably show that any possible policy can be represented using an affine combination of these policy independent basis functions. Given a reward function at test time, we simply need to find the right set of linear weights to combine these basis corresponding to the optimal policy. We derive a practical algorithm to learn these basis functions using only interaction data from the environment and show that our approach can produce the optimal policy at test time for any given reward function without additional environmental interactions.", "keywords": [ "Zero-Shot Reinforcement Learning", "Representation Learning", "Unsupervised RL" ], "rating_list": [ 5, 8, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "s8lj3C39Ow", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Defu Lian", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Enhong Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Haoqiang Zhang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Qizhou Wang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "PhD student" }, { "name": "Yonggang Zhang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Postdoc" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 33, "n_formula_1": 11, "n_ref_uni": 55, "n_ref": 121, "n_ref_all": 154, "n_fig": 9, "n_tab": 10, "L_tab": 7546, "n_element_tab": 885, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 4123, "n_element_tab_1": 181, "formula_len_all": 1920, "formula_len_all_1": 374, "len_all": 211136, "len_all_1": 73869, "len_abs": 1329, "len_title": 144, "len_sents": 60151, "len_sents_1": 32268, "n_sents": 502, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1368, "title": "Strengthening Federated Learning: Surrogate Data-Guided Aggregation for Robust Backdoor Defense", "abs": "Backdoor attacks in federated learning (FL) have garnered significant attention due to their destructive potential. Current advanced backdoor defense strategies typically involve calculating predefined metrics related to local models and modifying the server's aggregation rule accordingly. However, these metrics may exhibit biases due to the inclusion of malicious models in the calculation, leading to defense failures. To address this issue, we propose a novel backdoor defense method in FL named $\\textit{Su}$rrogate $\\textit{D}$ata-guided $\\textit{A}$ggregation (SuDA). SuDA independently evaluates local models using surrogate data, thereby mitigating the influence of malicious models. Specifically, it constructs a surrogate dataset composed of pure noise, which is shared between the server and clients. By leveraging this shared surrogate data, clients train their models using both the shared and local data, while the server reconstructs potential triggers for each local model to identify backdoors, facilitating the filtering of backdoored models before aggregation. To ensure the generalizability of local models across both local and surrogate data, SuDA aligns local data with surrogate data in the representation space, supported by theoretical analysis. Comprehensive experiments demonstrate the substantial superiority of SuDA over previous works.", "keywords": [ "Federated Learning", "Backdoor Attacks", "Generative Learning" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "s7vwXDsVYa", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Emiel Hoogeboom", "gender": "unknown", "institution": "Google", "country": "NL", "position": "Researcher" }, { "name": "Manoj Kumar", "gender": "unknown", "institution": "Google", "country": "", "position": "Research Engineer" }, { "name": "Neil Houlsby", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 58, "n_ref": 73, "n_ref_all": 84, "n_fig": 8, "n_tab": 4, "L_tab": 687, "n_element_tab": 71, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 708, "n_element_tab_1": 21, "formula_len_all": 230, "formula_len_all_1": 231, "len_all": 143879, "len_all_1": 56158, "len_abs": 1224, "len_title": 128, "len_sents": 33751, "len_sents_1": 27158, "n_sents": 278, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1246, "title": "Conditional Diffusion on Web-Scale Image Pairs leads to Diverse Image Variations", "abs": "Generating image variations, where a model produces variations of an input image while preserving the semantic context has gained increasing attention. Current image variation techniques involve adapting a text-to-image model to reconstruct an input image conditioned on the same image. We first demonstrate that a diffusion model trained to reconstruct an input image from frozen embeddings can reconstruct the image with minor variations. Second, inspired by how text-to-image models learn from web-scale text-image pairs, we explore a new pretraining strategy to generate image variations using a large collection of image pairs. Our diffusion model \\textit{Semantica} receives a random (encoded) image from a webpage as conditional input and denoises another noisy random image from the same webpage. We carefully examine various design choices for the image encoder, given its crucial role in extracting relevant context from the input image. Once trained, \\textit{Semantica} can adaptively generate new images from a dataset by simply using images from that dataset as input. Finally, we identify limitations in standard image consistency metrics for evaluating image variations and propose alternative metrics based on few-shot generation.", "keywords": [ "diffusion", "image variations", "generative models" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "s7lzZpAW7T", "primary_area": "datasets and benchmarks", "n_author": 76, "author_info_list": [ { "name": "Andy T. Liu", "gender": "Male", "institution": "ASUS", "country": "SG", "position": "Researcher" }, { "name": "Ankita Pasad", "gender": "Female", "institution": "Toyota Technological Institute at Chicago", "country": "", "position": "PhD student" }, { "name": "Anuj Diwan", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" }, { "name": "CHUANG Ming To", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "MS student" }, { "name": "Chao-Han Huck Yang", "gender": "Male", "institution": "NVIDIA Research", "country": "TW", "position": "Researcher" }, { "name": "Chee-En Yu", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Cheng-Hsiu Hsieh", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Chengxi Wu", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Chi-Yuan Hsiao", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "MS student" }, { "name": "Chi An Fu", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Chieh-Chi Yang", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Chien-yu Huang", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Researcher" }, { "name": "Chih-Kai Yang", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "MS student" }, { "name": "Ching Yu Hsu", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Chun-Yi Kuan", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "MS student" }, { "name": "Chun Wei Chen", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Chung-Ming Chien", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "PhD student" }, { "name": "David Harwath", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "Assistant Professor" }, { "name": "Dien-Ruei Wu", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Eunjung Yeo", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Researcher" }, { "name": "Fabian Alejandro Ritter Gutierrez", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" }, { "name": "Haibin Wu", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "PhD student" }, { "name": "Heitor Guimar\u00e3es", "gender": "Male", "institution": "Institut national de la recherche scientifique", "country": "", "position": "PhD student" }, { "name": "Homu Chang", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Hsi-Che Lin", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Hsi-Chun Cheng", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "MS student" }, { "name": "Hsien-Fu Hsiao", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Hsuan-Hao Lin", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Hsuan-Ting Lin", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Huang Kuan-Po", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "PhD student" }, { "name": "Hung-yi Lee", "gender": "Non-Binary", "institution": "National Taiwan University", "country": "TW", "position": "Full Professor" }, { "name": "I-Hsiang Chiu", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Jia-Lin Fang", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Jia Qi Yip", "gender": "unknown", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" }, { "name": "Jian-Ren Lin", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Jiatong Shi", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Jionghao Han", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "KUAN YU FANG CHIANG", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Kai-Wei Chang", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "PhD student" }, { "name": "Kalvin Chang", "gender": "unknown", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Researcher" }, { "name": "Ke-Han Lu", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "PhD student" }, { "name": "Kunal Dhawan", "gender": "Male", "institution": "NVIDIA ", "country": "US", "position": "Researcher" }, { "name": "Kwanghee Choi", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Lee Chen Wei", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "MS student" }, { "name": "Li-Chun Lu", "gender": "Female", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Puyuan Peng", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" }, { "name": "Shao-Syuan Huang", "gender": "unknown", "institution": "Department of computer science and informational engineering, National Taiwan University", "country": "", "position": "Undergrad student" }, { "name": "Shao-Xiang Yuan", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Shi-Xin Fang", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Shih-Heng Wang", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "MS student" }, { "name": "Shih-Yun Shan Kuan", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Shinji Watanabe", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Shou-Jen Chen", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Shu-wen Yang", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "PhD student" }, { "name": "Siddhant Arora", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Suwon Shon", "gender": "Male", "institution": "ASAPP", "country": "", "position": "Researcher" }, { "name": "Ting-Chun Liu", "gender": "Male", "institution": "CancerFree", "country": "TW", "position": "Intern" }, { "name": "Ting-Wu Chang", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Tsung-Min Pai", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Tzu-Chieh Wei", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Tzu-Quan Lin", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "PhD student" }, { "name": "Wei-Cheng Tseng", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Wei-Chih Chen", "gender": "unknown", "institution": "National Taiwan University", "country": "", "position": "MS student" }, { "name": "Wei Jui Chiang", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "William Chen", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Xuanjun Chen", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "PhD student" }, { "name": "Yi-Cheng Lin", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "PhD student" }, { "name": "You-Kuan Lin", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Yu-Hua Chen", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Yu-Xiang Lin", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "MS student" }, { "name": "Yuxun Tang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "MS student" }, { "name": "Chen An Li", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "MS student" }, { "name": "Kare Livescu", "gender": "Female", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Professor" }, { "name": "Tzu-Yuan Lin", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "MS student" }, { "name": "YI-JEN Shih", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Yun-shao Tsai", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 34, "n_ref": 62, "n_ref_all": 89, "n_fig": 14, "n_tab": 14, "L_tab": 23366, "n_element_tab": 2523, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 122, "n_element_tab_1": 18, "formula_len_all": 56, "formula_len_all_1": 56, "len_all": 308495, "len_all_1": 67957, "len_abs": 1512, "len_title": 182, "len_sents": 46822, "len_sents_1": 34236, "n_sents": 368, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 133, "L_abs": 1518, "title": "Dynamic-SUPERB Phase-2: A Collaboratively Expanding Benchmark for Measuring the Capabilities of Spoken Language Models with 180 Tasks", "abs": "Multimodal foundation models, such as Gemini and GPT-4, have revolutionized human-machine interactions by seamlessly integrating various forms of data. Developing a universal spoken language model that comprehends a wide range of natural language instructions is critical for bridging communication gaps and facilitating more intuitive interactions. However, the absence of a comprehensive evaluation benchmark poses a significant challenge. We present Dynamic-SUPERB Phase-2, an open and evolving benchmark for the comprehensive evaluation of instruction-based universal speech models. Building upon the first generation, this second version incorporates 125 new tasks contributed collaboratively by the global research community, expanding the benchmark to a total of 180 tasks, making it the largest benchmark for speech and audio evaluation. While the first generation of Dynamic-SUPERB was limited to classification tasks, Dynamic-SUPERB Phase-2 broadens its evaluation capabilities by introducing a wide array of novel and diverse tasks, including regression and sequence generation, across speech, music, and environmental audio. Evaluation results indicate that none of the models performed well universally. SALMONN-13B excelled in English ASR, while WavLLM demonstrated high accuracy in emotion recognition, but current models still require further innovations to handle a broader range of tasks. We open-source all task data and the evaluation pipeline, which will be available after the paper is published.", "keywords": [ "speech", "music", "audio", "benchmark", "large language model" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "s7Rc5KPqYI", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Afrar Jahin", "gender": "unknown", "institution": "Augusta University", "country": "US", "position": "PhD student" }, { "name": "Arif Hassan Zidan", "gender": "Male", "institution": "Augusta University", "country": "US", "position": "PhD student" }, { "name": "Yu Bao", "gender": "Female", "institution": "James Madison University", "country": "", "position": "Assistant Professor" }, { "name": "Wade Zhang", "gender": "Male", "institution": "Augusta University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 30, "n_formula_1": 22, "n_ref_uni": 45, "n_ref": 92, "n_ref_all": 115, "n_fig": 8, "n_tab": 3, "L_tab": 459, "n_element_tab": 8, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2157, "formula_len_all_1": 1723, "len_all": 115587, "len_all_1": 52398, "len_abs": 1328, "len_title": 180, "len_sents": 32820, "len_sents_1": 23014, "n_sents": 259, "n_sents_1": 159, "venue": "ICLR2025", "decision": "withdraw", "L_title": 131, "L_abs": 1332, "title": "HOME-3: HIGH-ORDER MOMENTUM ESTIMATOR USING THIRD-POWER GRADIENT FOR CONVEX, SMOOTH NONCONVEX, AND NONSMOOTH NONCONVEX OPTIMIZATION", "abs": "Momentum-based gradients are critical for optimizing advanced machine learning models, as they not only accelerate convergence but also help gradient-based optimizers overcome stationary points. While most state-of-the-art momentum techniques rely on lower-power gradients, such as the squared first-order gradient, there has been limited exploration into the potential of higher-power gradients\u2014those raised to powers greater than two, such as the third-power first-order gradient. In this work, we introduce the concept of high-order momentum, where\nmomentum is constructed using higher-power gradients, with a specific focus on the third-power first-order gradient as a representative example. Our research offers both theoretical and empirical evidence of the benefits of this novel approach. From a theoretical standpoint, we demonstrate that incorporating third-power gradients into momentum can improve the convergence bounds of gradient-based optimizers\nfor both convex and smooth nonconvex problems. To validate these findings, we conducted extensive empirical experiments across convex, smooth nonconvex, and nonsmooth nonconvex optimization tasks. The results consistently showcase that high-order momentum outperforms traditional momentum-based optimizers, providing superior performance and more efficient optimization.", "keywords": [ "Gradient Descent", "Momentum", "High-Power Gradient", "Nonconvex", "Coordinate Randomization" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 2, 2 ] }, { "paper_id": "s7Q1j5Hqpw", "primary_area": "learning on time series and dynamical systems", "n_author": 8, "author_info_list": [ { "name": "He zhang", "gender": "Male", "institution": "Northeast Normal University", "country": "CN", "position": "MS student" }, { "name": "Jianpeng Zhao", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "PhD student" }, { "name": "Lu Jiang", "gender": "Female", "institution": "Dalian Martime University", "country": "CN", "position": "Assistant Professor" }, { "name": "Pengyang Wang", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Assistant Professor" }, { "name": "Qi Hao", "gender": "Female", "institution": "University of Macau", "country": "MO", "position": "PhD student" }, { "name": "Ziyue Qiao", "gender": "Male", "institution": "Great Bay University", "country": "CN", "position": "Assistant Professor" }, { "name": "pengfei wang", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "XU HAIHUA", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 33, "n_ref_uni": 7, "n_ref": 18, "n_ref_all": 31, "n_fig": 5, "n_tab": 5, "L_tab": 2190, "n_element_tab": 176, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1368, "n_element_tab_1": 70, "formula_len_all": 1700, "formula_len_all_1": 1567, "len_all": 123140, "len_all_1": 64681, "len_abs": 1822, "len_title": 117, "len_sents": 40740, "len_sents_1": 26577, "n_sents": 309, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1750, "title": "SHIFT-RESILIENT DIFFUSIVE IMPUTATION FOR VARIABLE SUBSET FORECASTING", "abs": "It is common for sensor failures to result in missing data, leading to training sets being complete while test sets have only a small subset of variables. The challenge lies in utilizing incomplete data for forecasting, which is known as the Variable Subset Forecasting (VSF). In VSF tasks, significant distribution shift is present. One type is inter-series shift, which indicates changes in correlations between different series, and the other type is intra-series shift, which refers to substantial distribution differences within the same series across different time windows. Existing approaches to solving VSF tasks typically involve imputing the missing data first and then making predictions using the completed series. However, these methods do not account for the shift inherent in VSF tasks, resulting in poor model performance. To address these challenges, we propose a Shift-Resilient Diffusive Imputation (SRDI) framework against the shift. Specifically, SRDI integrates divide-conquer strategy with the denoising process, that decomposes the input into invariant patterns and variant patterns, representing the temporally stable parts of inter-series correlation and the highly fluctuating parts, respectively. By extracting spatiotemporal features from each separately and then appropriately combining them, inter-series shift can be effectively mitigated. Then, we innovatively organize SRDI and the forecasting model into a meta-learning paradigm tailored for VSF scenarios. We address the intra-series shift by treating time windows as tasks during training and employing an adaptation process before testing. Extensive experiments on four datasets have demonstrated our superior performance compared with state-of-the-art methods.", "keywords": [ "data mining; variable subset forecasting; distribution shift" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "s7DkcgpRxL", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Guiming Xie", "gender": "unknown", "institution": "Guangdong OPPO Mobile Telecommunications Corp.,Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Huan Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" }, { "name": "Jue WANG", "gender": "Male", "institution": "Together AI", "country": "US", "position": "Researcher" }, { "name": "Jun Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Ke Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Kunlong Zhou", "gender": "unknown", "institution": "Guangdong OPPO Mobile Telecommunications Corp.,Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Lidan Shou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Xuejian Gong", "gender": "Male", "institution": "Guangdong OPPO Mobile Telecommunications Corp.,Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 70, "n_ref": 153, "n_ref_all": 194, "n_fig": 19, "n_tab": 8, "L_tab": 3489, "n_element_tab": 137, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 2000, "n_element_tab_1": 96, "formula_len_all": 1086, "formula_len_all_1": 625, "len_all": 239235, "len_all_1": 64737, "len_abs": 1605, "len_title": 130, "len_sents": 52063, "len_sents_1": 28481, "n_sents": 373, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1616, "title": "Train Small, Infer Large: Memory-Efficient LoRA Training for Large Language Models", "abs": "Large Language Models (LLMs) have significantly advanced natural language processing with exceptional task generalization capabilities. Low-Rank Adaption (LoRA) offers a cost-effective fine-tuning solution, freezing the original model parameters and training only lightweight, low-rank adapter matrices. However, the memory footprint of LoRA is largely dominated by the original model parameters. To mitigate this, we propose LoRAM, a memory-efficient LoRA training scheme founded on the intuition that many neurons in over-parameterized LLMs have low training utility but are essential for inference. LoRAM presents a unique twist: it trains on a pruned (small) model to obtain pruned low-rank matrices, which are then recovered and utilized with the original (large) model for inference. Additionally, minimal-cost continual pre-training, performed by the model publishers in advance, aligns the knowledge discrepancy between pruned and original models. Our extensive experiments demonstrate the efficacy of LoRAM across various pruning strategies and downstream tasks. For a model with 70 billion parameters, LoRAM enables training on a GPU with only 20G HBM, replacing an A100-80G GPU for LoRA training and 15 GPUs for full fine-tuning. Specifically, QLoRAM implemented by structured pruning combined with 4-bit quantization, for LLaMA-3.1-70B (LLaMA-2-70B), reduces the parameter storage cost that dominates the memory usage in low-rank matrix training by 15.81\u00d7 (16.95\u00d7), while achieving dominant performance gains over both the original LLaMA-3.1-70B (LLaMA-2-70B) and LoRA-trained LLaMA-3.1-8B (LLaMA-2-13B).", "keywords": [ "Large Language Model", "Memory-Efficient LoRA" ], "rating_list": [ 8, 6, 3, 6, 8 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 4, 3, 3, 3, 1 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 5, 3, 5, 4, 4 ] }, { "paper_id": "s77FHD4wra", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Changwen Zheng", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Fengge Wu", "gender": "Female", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Sheng Yang", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "WangTie", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Gao Hang", "gender": "Male", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 47, "n_formula_1": 23, "n_ref_uni": 19, "n_ref": 43, "n_ref_all": 68, "n_fig": 9, "n_tab": 4, "L_tab": 3176, "n_element_tab": 103, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 3988, "n_element_tab_1": 77, "formula_len_all": 3795, "formula_len_all_1": 1335, "len_all": 143052, "len_all_1": 67981, "len_abs": 1253, "len_title": 109, "len_sents": 42962, "len_sents_1": 28381, "n_sents": 348, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1219, "title": "Rigid Body Dynamics Simulation Based on GNNs with Constraints", "abs": "In recent years, the utilization of Graph Neural Network (GNN)-based methods for simulating complex physical systems has opened new avenues for the fields of computational science and engineering. Despite their success, current GNN-based methods for rigid body dynamic simulation are constrained to relatively simple scenarios, hindering their practical use in industrial settings where complex mechanical structures and interconnected components prevail. These methods face challenges in handling intricate force relationships within rigid bodies, primarily due to the difficulty in obtaining force-related data for objects in industrial environments. To address this, we propose a novel constraint-guided method that incorporates force analysis into GNN-based simulations. The model incorporates computations related to both contact and non-contact forces into the prediction process. Additionally, it imposes physical constraints on the prediction process based on Kane's equations. We have rigorously demonstrated the model's rationality and effectiveness with thorough theoretical demonstration and empirical analysis. \\textit{Codes and anonymous links to the datasets are available in the supplementary materials.", "keywords": [ "Graph Neural Network; Dynamics Simulation" ], "rating_list": [ 3, 8, 5, 3 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 3, 4, 3, 1 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "s6q6zX45F8", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Tianyi Zhang", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" }, { "name": "Yu Cao", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Dianbo Liu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 7, "n_ref_uni": 24, "n_ref": 50, "n_ref_all": 76, "n_fig": 11, "n_tab": 19, "L_tab": 5983, "n_element_tab": 464, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1038, "n_element_tab_1": 89, "formula_len_all": 1375, "formula_len_all_1": 490, "len_all": 139276, "len_all_1": 53761, "len_abs": 1413, "len_title": 146, "len_sents": 42926, "len_sents_1": 25076, "n_sents": 343, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1417, "title": "Uncertainty-Based Extensible Codebook for Discrete Federated Learning in Heterogeneous Data Silos", "abs": "Federated learning (FL), aimed at leveraging vast distributed datasets, confronts a crucial challenge: the heterogeneity of data across different silos. While previous studies have explored discrete representations to enhance model generalization across minor distributional shifts, these approaches often struggle to adapt to new data silos with significantly divergent distributions. In response, we have identified that models derived from FL exhibit markedly increased uncertainty when applied to data silos with unfamiliar distributions. Consequently, we propose an innovative yet straightforward iterative framework, termed Uncertainty-Based Extensible-Codebook Federated Learning (UEFL). This framework dynamically maps latent features to trainable discrete vectors, assesses the uncertainty, and specifically extends the discretization dictionary or codebook for silos exhibiting high uncertainty. Our approach aims to simultaneously enhance accuracy and reduce uncertainty by explicitly addressing the diversity of data distributions, all while maintaining minimal computational overhead in environments characterized by heterogeneous data silos. Through experiments conducted on six datasets, our method has demonstrated its superiority, achieving significant improvements in accuracy (by 3%--22.1%) and uncertainty reduction (by 38.83%--96.24%), thereby outperforming contemporary state-of-the-art methods.", "keywords": [ "Federated learning", "uncertainty", "discretization" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "s6nYndMwG7", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Yang Liu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "Yegor Klochkov", "gender": "Male", "institution": "ByteDance, AI Lab", "country": "", "position": "Researcher" } ], "n_formula": 73, "n_formula_1": 26, "n_ref_uni": 26, "n_ref": 63, "n_ref_all": 86, "n_fig": 7, "n_tab": 5, "L_tab": 2895, "n_element_tab": 142, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 38, "n_element_tab_1": 2, "formula_len_all": 6929, "formula_len_all_1": 1475, "len_all": 142447, "len_all_1": 65239, "len_abs": 946, "len_title": 126, "len_sents": 45344, "len_sents_1": 29555, "n_sents": 423, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 954, "title": "Revisiting inverse Hessian vector products for calculating influence functions", "abs": "Influence functions are a popular tool for attributing models' outputs to training data. The traditional approach relies on the calculation of inverse Hessian-vector products (iHVP), but the classical solver ``Linear time Stochastic Second-order Algorithm'' (LiSSA, Agarwal et al. (2017)) is often deemed impractical for large models due to expensive computation and hyperparameter tuning. We show that the three hyperparameters --- the scaling factor, the batch size, and the number of steps --- can be chosen depending on two specific spectral properties of the Hessian: its trace and largest eigenvalue. By evaluating them with random sketching (Swartworth and Woodruff, 2023), we find that the batch size has to be sufficiently large for the LiSSA to converge; however, for all of the models we consider, the requirement is mild. We confirm our findings empirically by comparing to the Proximal Bregman Retraining Functions (PBRF, Bae et al. (2022)).", "keywords": [ "influence functions", "inverse Hessian-vector products", "random sketching" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "s6Q7aVZWIn", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Anke Tang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Han Hu", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaochun Cao", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoguang Ren", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yong Luo", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Professor" } ], "n_formula": 52, "n_formula_1": 25, "n_ref_uni": 45, "n_ref": 60, "n_ref_all": 84, "n_fig": 7, "n_tab": 5, "L_tab": 1336, "n_element_tab": 124, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1038, "n_element_tab_1": 114, "formula_len_all": 3000, "formula_len_all_1": 1242, "len_all": 146814, "len_all_1": 68877, "len_abs": 1582, "len_title": 125, "len_sents": 38420, "len_sents_1": 28739, "n_sents": 333, "n_sents_1": 231, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1597, "title": "Targeted Low-rank Refinement: Enhancing Sparse Neural Networks with Precision", "abs": "Pruning is a widely used technique for compressing large neural networks that eliminate weights that have minimal impact on the model's performance. Current pruning methods, exemplified by magnitude pruning, assign an importance score to each weight based on its magnitude and remove weights with scores below a certain threshold. Nonetheless, these methods often create a gap between the original dense and the pruned sparse model, potentially impairing performance. Especially when the sparsity ratio is high, the gap becomes more pronounced. To mitigate this issue, we introduce to bridge the gap left by pruning by utilizing a low-rank approximation of the difference between the dense and sparse matrices. Our method specifically entails the iterative refinement of the sparse weight matrix, augmented by a low-rank adjustment. This technique captures and retains the essential information often lost during pruning, thereby improving the performance of the pruned model. Furthermore, we offer a comprehensive theoretical analysis of our approach, emphasizing its convergence properties and establishing a solid basis for its efficacy. Experimental results on LLaMa models validate its effectiveness on large language models across various pruning techniques and sparsity levels. Our method shows significant improvements: at 50\\% sparsity, it reduces perplexity by 53.9\\% compared to conventional magnitude pruning on LLaMa-7B.Furthermore, to achieve a specific performance target, our approach enables an 8.6\\% reduction in model parameters while maintaining a sparsity ratio of about 50\\%.", "keywords": [ "Model Compression", "Low-Rank Refinement", "Model Pruning" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "s5orchdb33", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Karen Hambardzumyan", "gender": "Male", "institution": "Meta", "country": "GB", "position": "Researcher" }, { "name": "Lei Yu", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Nicola Cancedda", "gender": "Male", "institution": "Meta", "country": "GB", "position": "Researcher" }, { "name": "Virginie Do", "gender": "Female", "institution": "Universit\u00e9 Paris Dauphine - PSL", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 18, "n_ref_uni": 47, "n_ref": 92, "n_ref_all": 113, "n_fig": 5, "n_tab": 6, "L_tab": 3022, "n_element_tab": 483, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 176, "n_element_tab_1": 21, "formula_len_all": 896, "formula_len_all_1": 905, "len_all": 155367, "len_all_1": 59614, "len_abs": 1062, "len_title": 91, "len_sents": 42047, "len_sents_1": 28446, "n_sents": 252, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1043, "title": "Robust LLM safeguarding via refusal feature adversarial training", "abs": "Large language models (LLMs) are vulnerable to adversarial attacks that can elicit harmful responses. Defending against such attacks remains challenging due to the opacity of jailbreaking mechanisms and the high computational cost of training LLMs robustly. We demonstrate that adversarial attacks share a universal mechanism for circumventing LLM safeguards that works by ablating a dimension in the residual stream embedding space called the refusal feature. We further show that the operation of refusal feature ablation (RFA) approximates the worst-case perturbation of offsetting model safety. Based on these findings, we propose Refusal Feature Adversarial Training (ReFAT), a novel algorithm that efficiently performs LLM adversarial training by simulating the effect of input-level attacks via RFA. Experiment results show that ReFAT significantly improves the robustness of three popular LLMs against a wide range of adversarial attacks, with considerably less computational overhead compared to existing adversarial training methods.", "keywords": [ "large language models", "adversarial robustness", "representation analysis", "mechanistic interpretability" ], "rating_list": [ 8, 3, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 4, 2, 3 ] }, { "paper_id": "s5epFPdIW6", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Haoran Li", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Huaxiu Yao", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "James Zou", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Kangyu Zhu", "gender": "Male", "institution": "Brown University", "country": "US", "position": "MS student" }, { "name": "Linjun Zhang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Associate Professor" }, { "name": "Peng Xia", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Sheng Wang", "gender": "unknown", "institution": "University of Washington, Seattle", "country": "US", "position": "Assistant Professor" }, { "name": "Tianze Wang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Weijia Shi", "gender": "unknown", "institution": "University of Washington, Seattle", "country": "US", "position": "PhD student" } ], "n_formula": 35, "n_formula_1": 19, "n_ref_uni": 47, "n_ref": 139, "n_ref_all": 161, "n_fig": 5, "n_tab": 12, "L_tab": 4087, "n_element_tab": 493, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 659, "n_element_tab_1": 123, "formula_len_all": 2760, "formula_len_all_1": 1494, "len_all": 231350, "len_all_1": 64650, "len_abs": 1538, "len_title": 125, "len_sents": 62795, "len_sents_1": 29336, "n_sents": 465, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1565, "title": "MMed-RAG: Versatile Multimodal RAG System for Medical Vision Language Models", "abs": "Artificial Intelligence (AI) has demonstrated significant potential in healthcare, particularly in disease diagnosis and treatment planning. Recent progress in Medical Large Vision-Language Models (Med-LVLMs) has opened up new possibilities for interactive diagnostic tools. However, these models often suffer from factual hallucination, which can lead to incorrect diagnoses. Fine-tuning and retrieval-augmented generation (RAG) have emerged as methods to address these issues. However, the amount of high-quality data and distribution shifts between training data and deployment data limit the application of fine-tuning methods. Although RAG is lightweight and effective, existing RAG-based approaches are not sufficiently general to different medical domains and can potentially cause misalignment issues, both between modalities and between the model and the ground truth. In this paper, we propose a versatile multimodal RAG system, MMed-RAG, designed to enhance the factuality of Med-LVLMs. Our approach introduces a domain-aware retrieval mechanism, an adaptive retrieved contexts selection, and a provable RAG-based preference fine-tuning strategy. These innovations make the RAG process sufficiently general and reliable, significantly improving alignment when introducing retrieved contexts. Experimental results across five medical datasets (involving radiology, ophthalmology, pathology) on medical VQA and report generation demonstrate that MMed-RAG can achieve an average improvement of 43.8% in factual accuracy in the factual accuracy of Med-LVLMs.", "keywords": [ "medical vision-language model", "retrieval-augmented generation" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "s5TNZPKWy5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Hyunwoo Yu", "gender": "Male", "institution": "Sogang University", "country": "KR", "position": "PhD student" }, { "name": "Kyeongbo Kong", "gender": "Male", "institution": "Pusan National University", "country": "KR", "position": "Assistant Professor" }, { "name": "Suk-Ju Kang", "gender": "Male", "institution": "Sogang University", "country": "KR", "position": "Full Professor" }, { "name": "Yubin Cho", "gender": "Female", "institution": "LG Corporation", "country": "KR", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 36, "n_ref": 59, "n_ref_all": 100, "n_fig": 15, "n_tab": 6, "L_tab": 2983, "n_element_tab": 239, "n_fig_1": 11, "n_tab_1": 5, "L_tab_1": 2470, "n_element_tab_1": 222, "formula_len_all": 613, "formula_len_all_1": 536, "len_all": 147963, "len_all_1": 62834, "len_abs": 1790, "len_title": 124, "len_sents": 38200, "len_sents_1": 27653, "n_sents": 270, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1800, "title": "METRIS: Multi-Expressions for Transformer-based Referring Image Segmentation", "abs": "Referring image segmentation (RIS) aims to precisely segment a target object described by a linguistic expression. Recent RIS methods have introduced Transformer-based networks that use vision features as query and linguistic expression features as key-value to find target regions by referring to the given linguistic information. Since the Transformer-based network predicts based on the guidance information that guides the network on which regions to pay attention, the capacity of this guidance information has a significant impact on segmentation results in Transformer-based RIS. However, existing methods rely only on linguistic tokens as the guidance elements, which are limited in providing the visual understanding of the fine-grained target regions. To address this issue, we present a novel Multi-Expression guidance framework for Transformer-based Referring Image Segmentation, METRIS, which allows the network to refer to the visual expression tokens as the guidance information alongside the linguistic expression tokens. The introduction of visual expression can complement the capability of linguistic guidance by effectively providing the target-informative visual contexts. To generate the visual expression from vision features, we introduce a visual expression extractor that is designed to endow with the target-informative visual guidance ability and to acquire rich contextual information. This module strengthens the adaptability to the diverse image and language inputs, and improves visual understanding of the fine-grained target regions. Extensive experiments demonstrate the effectiveness of our approach across the commonly used RIS settings and the generalizability evaluation settings. Our method consistently shows strong performance on three public RIS benchmarks.", "keywords": [ "Multi-expression guidance framework", "Target-oriented visual expression", "Vision-Language representation", "Transformer-based Segmentation" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "s5T9A9tXTX", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Jie Zhao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Kang Hao Cheong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 2, "n_ref_uni": 51, "n_ref": 62, "n_ref_all": 105, "n_fig": 15, "n_tab": 13, "L_tab": 5235, "n_element_tab": 436, "n_fig_1": 13, "n_tab_1": 4, "L_tab_1": 1336, "n_element_tab_1": 114, "formula_len_all": 272, "formula_len_all_1": 99, "len_all": 188463, "len_all_1": 57853, "len_abs": 1372, "len_title": 121, "len_sents": 60579, "len_sents_1": 24390, "n_sents": 438, "n_sents_1": 162, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1376, "title": "Spatial Reasoning with MLLMs: A New Path to Graph-Structured Optimization", "abs": "Graph-structured problems pose significant challenges due to their complex structures and large scales, often making traditional computational approaches suboptimal or costly. However, when these problems are visually represented, humans can often solve them more intuitively, leveraging our inherent spatial reasoning capabilities. In this work, we introduce an original and novel approach by feeding graphs as images into multimodal large language models (MLLMs), aiming for a loss-free representation that preserves the graph's structural integrity and enables machines to mimic this human-like thinking. Our pioneering exploration of MLLMs addresses various graph-structured challenges, from combinatorial tasks like influence maximization to sequential decision-making processes such as network dismantling, along with tackling six basic graph-related problems. Our experiments reveal that MLLMs possess remarkable spatial intelligence and a unique aptitude for these problems, marking a significant step forward in enabling machines to understand and analyze graph-structured data with human-like depth and intuition. These findings also suggest that combining MLLMs with straightforward optimization techniques could offer a new, effective paradigm for managing large-scale graph problems without complex derivations, computationally demanding training and fine-tuning.", "keywords": [ "Multimodal large language model", "graph representation", "graph-structured problem", "combinatorial optimization" ], "rating_list": [ 3, 3, 6, 5, 3 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 1, 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 3, 2, 2 ], "confidence_list": [ 5, 5, 3, 5, 4 ] }, { "paper_id": "s5N7p5UjgR", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Clark Barrett", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Professor (Research)" }, { "name": "Max Lamparth", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Peter Chatain", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Scott Viteri", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 31, "n_ref": 44, "n_ref_all": 54, "n_fig": 7, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 177, "n_element_tab_1": 9, "formula_len_all": 1273, "formula_len_all_1": 956, "len_all": 183266, "len_all_1": 49130, "len_abs": 1530, "len_title": 104, "len_sents": 37774, "len_sents_1": 21628, "n_sents": 278, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1537, "title": "Markovian Transformers for Informative Language Modeling", "abs": "Chain-of-Thought (CoT) reasoning holds great promise for explaining language model outputs, but recent studies have highlighted significant challenges in its practical application for interpretability. We propose to address this issue by making CoT causally essential to prediction through two key components: factoring next-token prediction through intermediate CoT text, and training CoT to predict future tokens independently of other context. This results in \"Markovian\" language models, where CoT serves as a fixed-size state for future token prediction. Our approach optimizes for \"informativeness\" \u2013 the improvement in next-token predictions using a trained CoT compared to a baseline. Using Proximal Policy Optimization (PPO) for arithmetic problems and policy gradient for GSM8K, we demonstrate effectiveness on both arithmetic problems with Mistral 7B and the GSM8K benchmark with Llama 3.1 8B, where the model learns to produce CoTs that are 33.20% more effective at predicting answers than the pre-trained baseline. The increased sensitivity of model performance to CoT perturbations provides strong evidence of CoT reliance. Furthermore, we show that CoTs trained for one model generalize to help other models predict answers, suggesting these CoTs capture reasoning patterns that transfer across different interpreters. This work advances the development of more interpretable language models, potentially enabling their extension to arbitrarily long contexts and enhancing AI reasoning capabilities across various domains.", "keywords": [ "Chain of Thought Reasoning", "Reinforcement Learning", "Scalable Oversight", "Language Modeling", "Proximal Policy Optimization" ], "rating_list": [ 10, 6, 6, 5 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 4, 2, 2, 2 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "s4yXbEfZV5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Bin Ren", "gender": "Male", "institution": "University of Pisa", "country": "IT", "position": "PhD student" }, { "name": "Hong Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Jiajie Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Jinyan Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Junsong Yuan", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Full Professor" }, { "name": "Mengyuan Liu", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Wenhao Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yawei Li", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Lecturer" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 57, "n_ref": 162, "n_ref_all": 186, "n_fig": 6, "n_tab": 12, "L_tab": 5339, "n_element_tab": 719, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1256, "n_element_tab_1": 177, "formula_len_all": 777, "formula_len_all_1": 564, "len_all": 183800, "len_all_1": 50220, "len_abs": 2909, "len_title": 124, "len_sents": 44925, "len_sents_1": 22435, "n_sents": 343, "n_sents_1": 117, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1667, "title": "PrML: Progressive Multi-Task Learning for Monocular 3D Human Pose Estimation", "abs": "The lifting-based framework has dominated the field of monocular 3D human pose estimation by leveraging the well-detected 2D pose as an intermediate representation. However, it neglects the different initial states between 2D pose and per-joint depth and encodes the well-detected 2D pose feature and unknown per-joint depth feature in an entangled feature space. To address this limitation, we present a novel progressive multi-task learning pose estimation framework named PrML. Firstly, PrML introduces two task branches: one is to refine the well-detected 2D pose feature and the other is to learn the per-joint depth feature. This dual-branch design reduces the explicit influence of uncertain depth features on 2D pose features. Secondly, PrML employs a task aware decoder to supplement the complementary information between the refined 2D pose feature and the well-learned per-joint depth feature. This step establishes the connection between 2D pose and per-joint depth, compensating for the lack of interaction caused by the dual-branch design. We also conduct theoretical analysis from the perspective of mutual information and arrive at a loss to constrain this feature complementary process. Finally, we use two regression heads to regress the 2D pose and per-joint depth respectively, and concatenate them to obtain the final 3D pose. Extensive experiments show that PrML outperforms the conventional lifting-based framework with fewer parameters on two widely used datasets: Human3.6M and MPI-INF-3DHP. Code is available at https://anonymous.4open.science/r/PrML and we hope our effort can provide a new framework for monocular 3D human pose estimation.", "keywords": [ "Monocular 3D Human Pose Estimation; Multi-Task Learning" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 5, 4, 5 ] }, { "paper_id": "s4Wm71LFK4", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Haoyu Peter Wang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Pan Li", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Yinan Huang", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 20, "n_ref_uni": 40, "n_ref": 90, "n_ref_all": 105, "n_fig": 7, "n_tab": 11, "L_tab": 7797, "n_element_tab": 639, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 4254, "n_element_tab_1": 397, "formula_len_all": 4578, "formula_len_all_1": 1662, "len_all": 182225, "len_all_1": 80025, "len_abs": 1249, "len_title": 103, "len_sents": 43700, "len_sents_1": 31019, "n_sents": 381, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1261, "title": "What Are Good Positional Encodings for Directed Graphs?", "abs": "Positional encodings (PEs) are essential for building powerful and expressive graph neural networks and graph transformers, as they effectively capture the relative spatial relationships between nodes. Although extensive research has been devoted to PEs in undirected graphs, PEs for directed graphs remain relatively unexplored. This work seeks to address this gap. We first introduce the notion of *Walk Profile*, a generalization of walk-counting sequences for directed graphs. A walk profile encompasses numerous structural features crucial for directed graph-relevant applications, such as program analysis and circuit performance prediction. We identify the limitations of existing PE methods in representing walk profiles and propose a novel *Multi-q Magnetic Laplacian PE*, which extends the Magnetic Laplacian eigenvector-based PE by incorporating multiple potential factors. The new PE can provably express walk profiles. Furthermore, we generalize prior basis-invariant neural networks to enable the stable use of the new PE in the complex domain. Our numerical experiments validate the expressiveness of the proposed PEs and demonstrate their effectiveness in solving sorting network satisfiability and performing well on general circuit benchmarks.", "keywords": [ "Directed graphs", "positional encodings", "graph neural networks" ], "rating_list": [ 5, 6, 5, 8, 8 ], "soundness_list": [ 3, 3, 2, 4, 3 ], "presentation_list": [ 2, 3, 2, 4, 4 ], "contribution_list": [ 3, 2, 2, 4, 3 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "s4MwstmB8o", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Jian Pu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "XIN GAO", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "PhD student" } ], "n_formula": 60, "n_formula_1": 14, "n_ref_uni": 39, "n_ref": 87, "n_ref_all": 144, "n_fig": 11, "n_tab": 10, "L_tab": 5843, "n_element_tab": 493, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 3102, "n_element_tab_1": 312, "formula_len_all": 6169, "formula_len_all_1": 1986, "len_all": 249344, "len_all_1": 72031, "len_abs": 1249, "len_title": 114, "len_sents": 88307, "len_sents_1": 28061, "n_sents": 790, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1248, "title": "Deep Incomplete Multi-view Learning via Cyclic Permutation of VAEs", "abs": "Multi-View Representation Learning (MVRL) aims to derive a unified representation from multi-view data by leveraging shared and complementary information across views. However, when views are irregularly missing, the incomplete data can lead to representations that lack sufficiency and consistency. To address this, we propose Multi-View Permutation of Variational Auto-Encoders (MVP), which excavates invariant relationships between views in incomplete data. MVP establishes inter-view correspondences in the latent space of Variational Auto-Encoders, enabling the inference of missing views and the aggregation of more sufficient information. To derive a valid Evidence Lower Bound (ELBO) for learning, we apply permutations to randomly reorder variables for cross-view generation and then partition them by views to maintain invariant meanings under permutations. Additionally, we enhance consistency by introducing an informational prior with cyclic permutations of posteriors, which turns the regularization term into a similarity measure across distributions. We demonstrate the effectiveness of our approach on seven diverse datasets with varying missing ratios, achieving superior performance in multi-view clustering and generation tasks.", "keywords": [ "Multi-View Learning", "Representation Learning", "Multimodal VAEs", "Generative Models" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "s3sJenvY5H", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Botian Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Feng Chen", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Huazhe Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Peiqi Duan", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Pu Hua", "gender": "Male", "institution": "Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yanchao Yang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Yi Ma", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 30, "n_ref": 81, "n_ref_all": 95, "n_fig": 4, "n_tab": 6, "L_tab": 4428, "n_element_tab": 481, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 47, "formula_len_all_1": 47, "len_all": 260876, "len_all_1": 58141, "len_abs": 1752, "len_title": 99, "len_sents": 41037, "len_sents_1": 29313, "n_sents": 305, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1783, "title": "On Evaluation of Generative Robotic Simulations", "abs": "Due to the difficulty of acquiring extensive real-world data, robot simulation has become crucial for parallel training and sim-to-real transfer, highlighting the importance of scalable simulated robotic tasks. \nFoundation models have demonstrated impressive capacities in autonomously generating feasible robotic tasks. However, this new paradigm underscores the challenge of adequately evaluating these autonomously generated tasks. \nTo address this, we propose a comprehensive evaluation framework tailored to generative simulations. \nOur framework segments evaluation into three core aspects: ***quality***, ***diversity***, and ***generalization***.\nFor single-task quality, we evaluate the realism of the generated task and the completeness of the generated trajectories using large language models and vision-language models. In terms of diversity, we measure both task and data diversity through text similarity of task descriptions and world model loss trained on collected task trajectories. For task-level generalization, we assess the zero-shot generalization ability on unseen tasks of a policy trained with multiple generated tasks.\nExperiments conducted on three representative task generation pipelines demonstrate that the results from our framework are highly consistent with human evaluations, confirming the feasibility and validity of our approach. \nThe findings reveal that while metrics of quality and diversity can be achieved through certain methods, no single approach excels across all metrics, suggesting a need for greater focus on balancing these different metrics. Additionally, our analysis further highlights the common challenge of low generalization capability faced by current works.\nOur anonymous website: https://sites.google.com/view/evaltasks.", "keywords": [ "Robotics; Embodied AI; Foundation Models; Generative Simulations;" ], "rating_list": [ 5, 3, 8, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "s3W8bUX43z", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Aoxiao Zhong", "gender": "unknown", "institution": "Squirrel AI", "country": "US", "position": "Researcher" }, { "name": "Cheng Chen", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Dufan Wu", "gender": "unknown", "institution": "Massachusetts General Hospital, Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "Hao He", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Quanzheng Li", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Associate Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 47, "n_ref": 81, "n_ref_all": 109, "n_fig": 4, "n_tab": 12, "L_tab": 5064, "n_element_tab": 548, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1184, "n_element_tab_1": 139, "formula_len_all": 667, "formula_len_all_1": 569, "len_all": 182728, "len_all_1": 69103, "len_abs": 1277, "len_title": 139, "len_sents": 44599, "len_sents_1": 30929, "n_sents": 338, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1299, "title": "Learning Under Multi-dimensional Domain Shifts: A Ensemble of Mixtures of Experts Approach", "abs": "Domain shifts pose a significant challenge in deep learning applications. Existing methods typically address domain shifts by treating each domain in isolation, overlooking the underlying factors driving the shifts, or focus on only \\emph{one} factor. However, domain shifts in the real world often occur across \\emph{multiple} dimensions simultaneously. For example, medical datasets from different hospitals can exhibit variations in factors including demographics, equipment manufacturers, and imaging protocols, demonstrating a three-dimensional shifts.\nIn this paper, we introduce a novel approach to address the complexity of multi-dimensional domain shifts. Our method leverages an ensemble of mixtures of experts (EMoE), with each MoE specialized in different dimensions. \nCrucially, we innovate a domain estimator to address a particularly challenging issue frequently encountered in practice: domain labels may be missing or unreliable.\nA significant advantage of our method is its generalizability and adaptability to both centralized and federated learning settings, as well as its versatility across various tasks. Extensive experiments on six datasets demonstrate the superiority of our method over state-of-the-art domain generalization and personalized federated learning approaches.", "keywords": [ "Domain shift", "Federated Learning", "Healthcare" ], "rating_list": [ 3, 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 1, 3 ], "confidence_list": [ 5, 4, 4, 4, 3 ] }, { "paper_id": "s3IBHTTDYl", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Yingshan Chang", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Yonatan Bisk", "gender": "Male", "institution": "Meta", "country": "US", "position": "Visiting Professor" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 56, "n_ref": 115, "n_ref_all": 141, "n_fig": 4, "n_tab": 14, "L_tab": 4303, "n_element_tab": 841, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1063, "n_element_tab_1": 212, "formula_len_all": 492, "formula_len_all_1": 971, "len_all": 197957, "len_all_1": 72446, "len_abs": 1889, "len_title": 106, "len_sents": 60399, "len_sents_1": 36225, "n_sents": 381, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1705, "title": "Language Models Need Inductive Biases to Count Inductively", "abs": "Counting constitutes a core skill underlying a wide range of tasks, such as formal language recognition, multi-hop reasoning and simulating algorithms. Generalizing counting inductively is central to task success on out-of-distribution (OOD) instances where testing inputs are longer than those seen in training. While there is a large body of literature reporting poor length generalization in language models, few papers have tried to distill the \u201creasoning\u201d failure to the simplest case of counting failure. We aim to provide a broader picture on whether various language model architectures can a) learn to count, and b) generalize counting inductively. This work provides extensive empirical results on architectures ranging from RNNs, Transformers, State-Space Models and RWKV. We present carefully-designed task formats, auxiliary tasks and positional embeddings to avoid limitations in generalization with OOD-position and OOD-vocabulary. We find that while traditional RNNs trivially achieve inductive counting, Transformers have to rely on positional embeddings (PEs) to count OOD. Further analyses on interpreting the learned solution reveal that different PEs encode different inductive biases that facilitate counting in different task formats. As counting is the basis for many arguments concerning the expressivity of Transformers, our finding calls for the community to reexamine the application scope of primitive functions defined in formal characterizations. Finally, modern RNNs also largely underperform traditional RNNs in generalizing counting inductively, hinting at the tradeoff modern RNNs struggle to balance between parallelized training and maintaining their recurrent nature.", "keywords": [ "Language Model Architecture", "Expressivity", "Length Generalization" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 4, 3, 2, 4 ], "presentation_list": [ 4, 2, 3, 4 ], "contribution_list": [ 4, 2, 2, 4 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "s3FTX4Ay55", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Arjun Krishna", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Dinesh Jayaraman", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Assistant Professor" }, { "name": "Hung-Ju Wang", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Undergrad student" }, { "name": "Jason Xie", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Engineer" }, { "name": "Kyle Vedder", "gender": "Male", "institution": "School of Engineering and Applied Science, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Long Le", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "William Liang", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Undergrad student" }, { "name": "Yue Yang", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "ERIC EATON", "gender": "unknown", "institution": "University of Pennsylvania", "country": "US", "position": "Faculty" }, { "name": "Yecheng Jason Ma", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 2, "n_ref_uni": 24, "n_ref": 38, "n_ref_all": 78, "n_fig": 19, "n_tab": 5, "L_tab": 2538, "n_element_tab": 145, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 577, "n_element_tab_1": 33, "formula_len_all": 483, "formula_len_all_1": 135, "len_all": 133494, "len_all_1": 59618, "len_abs": 4367, "len_title": 149, "len_sents": 42356, "len_sents_1": 28417, "n_sents": 362, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 1457, "title": "Articulate-Anything: Automatic Modeling of Articulated Objects via a Vision-Language Foundation Model", "abs": "Interactive 3D simulated objects are crucial in AR/VR, animations, and robotics, driving immersive experiences and advanced automation.\nHowever, creating these articulated objects requires extensive human effort and expertise, limiting their broader applications. To overcome this challenge, we present Articulate-Anything, a system that automates the articulation of diverse, complex objects from many input modalities, including text, images, and videos. Articulate-Anything leverages vision-language models (VLMs) to generate code that can be compiled into an interactable digital twin for use in standard 3D simulators. Our system exploits existing 3D asset datasets via a mesh retrieval mechanism, along with an actor-critic system that iteratively proposes, evaluates, and refines solutions for articulating the objects, self-correcting errors to achieve a robust out- come. Qualitative evaluations demonstrate Articulate-Anything's capability to articulate complex and even ambiguous object affordances by leveraging rich grounded inputs. In extensive quantitative experiments on the standard PartNet-Mobility dataset, Articulate-Anything substantially outperforms prior work, increasing the success rate from 8.7-11.6\\% to 75\\% and setting a new bar for state-of-art performance. We further showcase the utility of our generated assets by using them to train robotic policies for fine-grained manipulation tasks that go beyond basic pick and place.", "keywords": [ "Computer vision", "Vision-Language Models", "Robotics" ], "rating_list": [ 8, 6, 3, 6, 8 ], "soundness_list": [ 4, 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 4, 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 4, 3, 3 ] }, { "paper_id": "s324bLSKui", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Feng Wang", "gender": "Female", "institution": "Wuhan University", "country": "", "position": "Full Professor" }, { "name": "Jianwei Yin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Jinshan Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Paul Weng", "gender": "Male", "institution": "Duke Kunshan University", "country": "", "position": "Associate Professor" }, { "name": "Zuoyou Liao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 26, "n_ref": 56, "n_ref_all": 60, "n_fig": 0, "n_tab": 3, "L_tab": 1614, "n_element_tab": 254, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1614, "n_element_tab_1": 254, "formula_len_all": 399, "formula_len_all_1": 399, "len_all": 82718, "len_all_1": 79885, "len_abs": 1766, "len_title": 170, "len_sents": 25687, "len_sents_1": 25690, "n_sents": 176, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 115, "L_abs": 1016, "title": "MSLC: Monte Carlo Tree Search Sampling Guided Local Construction for Solving Large-Scale Traveling Salesman Problem", "abs": "Neural solvers have achieved promising results in solving small-scale Travelling Salesman Problems (TSP), but inefficiencies arise when tackling larger instances. This paper proposes the MSLC (\\textbf{M}onte Carlo Tree Search \\textbf{S}ampling Guided \\textbf{L}ocal \\textbf{C}onstruction) framework, which innovatively integrates a predictive sampling module into the global coarse-grained selection module, MCTS, to achieve mutual integration with the fine-grained local construction module. This integration effectively balances coarse-grained exploration with fine-grained adjustment, thereby improving overall efficiency. This framework offers a novel way to combine autoregressive and non-autoregressive models. Experimental results demonstrate that MSLC effectively balances time and solution quality, outperforming state-of-the-art neural solvers. The performance gap of MSLC is reduced by at least 29.4\\% (resp. 34.7\\% or 28.5\\%) on TSP-500 (resp. TSP-1000 or TSP-10000), compared to the SOTA neural methods.", "keywords": [ "Combinatorial Optimization", "Neural Combinatorial Optimization", "Large scale problem" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 1, 3, 1, 1 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "s3003xWtfd", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Hancheng Ye", "gender": "Male", "institution": "Duke University, Duke University", "country": "US", "position": "PhD student" }, { "name": "Jianyang Gu", "gender": "Male", "institution": "The Ohio State University", "country": "US", "position": "Postdoc" }, { "name": "Jianyi Zhang", "gender": "unknown", "institution": "National Artificial Intelligence Research Resource Pilot & Duke University", "country": "US", "position": "Principal investigator" }, { "name": "Saeed Vahidian", "gender": "male", "institution": "Duke University", "country": "US", "position": "Postdoc" }, { "name": "Wang Qinsi", "gender": "Female", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Yiran Chen", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Professor" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 41, "n_ref": 63, "n_ref_all": 99, "n_fig": 14, "n_tab": 6, "L_tab": 1669, "n_element_tab": 114, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 884, "n_element_tab_1": 80, "formula_len_all": 341, "formula_len_all_1": 492, "len_all": 176026, "len_all_1": 74476, "len_abs": 2374, "len_title": 154, "len_sents": 58435, "len_sents_1": 32900, "n_sents": 513, "n_sents_1": 286, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1853, "title": "CoreInfer: Accelerating Large Language Model Inference with Semantics-Inspired Adaptive Sparse Activation", "abs": "Large language models (LLMs) with billions of parameters have sparked a new wave of exciting AI applications. However, their high computational costs and memory demands during inference pose significant challenges. Dynamic sparse activation inference, which activates only a small number of neurons for each token, offers a novel way to accelerate model inference without degrading performance, showing great potential for resource-constrained hardware devices. Nevertheless, existing methods predict activated neurons based on individual tokens with additional MLP, which involve frequent changes in activation maps and resource calls, limiting the acceleration benefits of sparse activation.\nIn this paper, we introduce **CoreInfer**, an MLP-free adaptive sparse activation inference method based on sentence-level prediction. Specifically, we propose the concept of sentence-wise core neurons, which refers to the subset of neurons most critical for a given sentence, and empirically demonstrate its effectiveness. To determine the core neurons, we explore the correlation between core neurons and the sentence's semantics. Remarkably, we discovered that core neurons exhibit both stability and similarity in relation to the sentence's semantics\u2014an insight overlooked by previous studies. Building on this finding, we further design two semantic-based methods for predicting core neurons to fit different input scenarios. In CoreInfer, the core neurons are determined during the pre-filling stage and fixed during the encoding stage, enabling zero-cost sparse inference. We evaluated the model generalization and task generalization of CoreInfer across various models and tasks. Notably, on an NVIDIA TITAN XP GPU, CoreInfer achieved a 10.33 $\\times$ and 2.72 $\\times$ speedup compared to the Huggingface implementation and PowerInfer, respectively.", "keywords": [ "Sparse Activation", "Sparse Inference", "Efficient LLMs", "LLMs Inference Acceleration" ], "rating_list": [ 8, 6, 8, 3 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 2, 3, 4, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "s2mEKEKplI", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Hisham Cholakkal", "gender": "Male", "institution": "MBZUAI", "country": "AE", "position": "Assistant Professor" }, { "name": "Kai-Qi Liu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Undergrad student" }, { "name": "Nian Liu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "Qi Fan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Rao Muhammad Anwer", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Wenbin Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Yang Gao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 47, "n_ref": 95, "n_ref_all": 111, "n_fig": 4, "n_tab": 5, "L_tab": 454, "n_element_tab": 141, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 750, "n_element_tab_1": 145, "formula_len_all": 666, "formula_len_all_1": 618, "len_all": 139647, "len_all_1": 69889, "len_abs": 1617, "len_title": 118, "len_sents": 36510, "len_sents_1": 32309, "n_sents": 257, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1623, "title": "Adapting Informative Structures for Cross-Domain Few-Shot Segmentation", "abs": "Cross-domain few-shot segmentation (CD-FSS) aims to segment objects of novel classes under domain shifts, using only a few mask-annotated support samples. However, directly applying pretrained CD-FSS models to unseen domains is often suboptimal due to their limited coverage of domain diversity by fixed parameters trained on source domains. Moreover, simply adjusting hand-selected model parameters, such as test-time training, typically neglects the distinct domain gaps and characteristics of target domains. To address these issues, we propose adapting informative model structures for target domains by learning domain characteristics from few-shot labeled support samples during inference. Specifically, we first adaptively identify domain-specific model structures by measuring parameter importance using a novel structure Fisher score in a data-dependent manner. Then, we progressively train the selected informative model structures with hierarchically constructed training samples, progressing from fewer to more support shots. Our method selectively and gradually adapts the model to target domains, optimizing model adaptation, minimizing overfitting risks, and maximizing the use of limited support data. The resulting Informative Structure Adaptation (ISA) method effectively addresses domain shifts and equips existing few-shot segmentation models with flexible adaptation capabilities for new domains, eliminating the need to redesign or retrain CD-FSS models on base data. Extensive experiments validate the effectiveness of our method, demonstrating superior performance across multiple CD-FSS benchmarks.", "keywords": [ "cross-domain few-shot segmentation", "few-shot segmentation", "test-time training" ], "rating_list": [ 5, 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 4, 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 5 ] }, { "paper_id": "s2jLQDqVUE", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Bin Li", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Fan Shi", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Haiyang Yu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Xiangyang Xue", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 32, "n_ref": 58, "n_ref_all": 74, "n_fig": 11, "n_tab": 3, "L_tab": 823, "n_element_tab": 108, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 896, "n_element_tab_1": 113, "formula_len_all": 537, "formula_len_all_1": 533, "len_all": 109773, "len_all_1": 55824, "len_abs": 1857, "len_title": 135, "len_sents": 34822, "len_sents_1": 26010, "n_sents": 255, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1657, "title": "Self-learning Compositional Representations for Zero-shot Chinese Character Recognition", "abs": "Chinese character recognition has been a longstanding research topic and remains essential in visual tasks like ancient manuscript recognition. Chinese character recognition faces numerous challenges, particularly the issue of zero-shot characters. Existing Chinese zero-shot character recognition methods primarily focus on the radical or stroke decomposition. However, radical-based methods still struggle to solve zero-shot radicals, while stroke-based ones are hard to perceive fine-grained information. Besides, previous methods can hardly generalize to characters of other languages. In this paper, we propose a novel Self-learning Compositional Representation method for zero-shot Chinese Character Recognition (SCR-CCR). SCR-CCR learns compositional components automatically from the data, which are not aligned with human-defined radical or stroke decomposition methods. SCR-CCR follows the pretraining-inference paradigm. First, we train a Character Slot Attention (ChSA) via pure feature reconstruction loss to parse appropriate components from character images. Then we recognize zero-shot characters without finetuning or retraining in the inference stage by comparing components between input and example images. To evaluate the proposed method, we conduct experiments of zero-shot character recognition. The experiments illustrate that SCR-CCR outperforms previous methods in most cases of character and radical zero-shot settings. In particular, visualization experiments indicate that the components learned by SCR-CCR reflect the structure of characters in an interpretable way, and can be used to recognize Japanese and Korean characters.", "keywords": [ "Chinese Character Recognition", "Object-centric Representations" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "s2SLzC0IPZ", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Heng Huang", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Peiran Yu", "gender": "Female", "institution": "University of Texas at Arlington", "country": "US", "position": "Assistant Professor" }, { "name": "Wenhan Xian", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" } ], "n_formula": 201, "n_formula_1": 35, "n_ref_uni": 43, "n_ref": 94, "n_ref_all": 102, "n_fig": 2, "n_tab": 3, "L_tab": 254, "n_element_tab": 20, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 182, "n_element_tab_1": 2, "formula_len_all": 32704, "formula_len_all_1": 3681, "len_all": 249137, "len_all_1": 65951, "len_abs": 1522, "len_title": 133, "len_sents": 54489, "len_sents_1": 24397, "n_sents": 665, "n_sents_1": 240, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1484, "title": "A Fast Federated Method for Minimax Problems with Sequential Convergence Guarantees", "abs": "Federated learning (FL) has recently been actively studied to collaboratively train machine learning models across clients without directly sharing data and to address data-hungry issues. Many FL works have been focusing on minimizing a loss function but many important machine learning tasks such as adversarial training, GANs, fairness learning, and AUROC maximization are formulated as minimax problems. In this paper, we propose a new federated learning method for minimax problems. Our method allows client drift and addresses the data heterogeneity issue. In theoretical analysis, we prove that our method can improve sample complexity and has convergence guarantees for the updates of the model parameters, i.e., the sequences generated by the method. Given the Kurdyka-\u0141ojasiewicz (KL) exponent of a novel potential function related to the objective function, we demonstrate that the sequences generated by our method converge finitely, linearly, or sublinearly. Our assumptions on the KL property are weaker than previous work on the sequential convergence of centralized minimax methods. Additionally, we further weaken the KL assumption by deducing the KL exponent of the potential function from that of the original objective function. We validate our federated learning method on AUC maximization tasks. The experimental results demonstrate that our method outperforms state-of-the-art federated learning methods when the distributions of local training data are non-IID.", "keywords": [ "federated learning", "minimax optimization" ], "rating_list": [ 3, 6, 3, 5, 3 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "s20W12XTF8", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Dongcheng Zhao", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Guobin Shen", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yi Zeng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yiting Dong", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "xiang he", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 48, "n_ref": 116, "n_ref_all": 140, "n_fig": 9, "n_tab": 6, "L_tab": 5171, "n_element_tab": 436, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 3276, "n_element_tab_1": 200, "formula_len_all": 454, "formula_len_all_1": 382, "len_all": 206102, "len_all_1": 67000, "len_abs": 1599, "len_title": 161, "len_sents": 56619, "len_sents_1": 28109, "n_sents": 412, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 112, "L_abs": 1612, "title": "Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models", "abs": "As large language models (LLMs) become integral to various applications, ensuring both their safety and utility is paramount. Jailbreak attacks, which manipulate LLMs into generating harmful content, pose significant challenges to this balance. Existing defenses, such as prompt engineering and safety fine-tuning, often introduce computational overhead, increase inference latency, and lack runtime flexibility. Moreover, overly restrictive safety measures can degrade model utility by causing refusals of benign queries. In this paper, we introduce *Jailbreak Antidote*, a method that enables real-time adjustment of LLM safety preferences by manipulating a sparse subset of the model's internal states during inference. By shifting the model's hidden representations along a safety direction with varying strengths, we achieve flexible control over the safety-utility balance without additional token overhead or inference delays. Our analysis reveals that safety-related information in LLMs is sparsely distributed; adjusting approximately *5\\%* of the internal state is as effective as modifying the entire state. Extensive experiments on nine LLMs (ranging from 2 billion to 72 billion parameters), evaluated against ten jailbreak attack methods and compared with six defense strategies, validate the effectiveness and efficiency of our approach. By directly manipulating internal states during reasoning, *Jailbreak Antidote* offers a lightweight, scalable solution that enhances LLM safety while preserving utility, opening new possibilities for real-time safety mechanisms in widely-deployed AI systems.", "keywords": [ "Large Language Models", "Jailbreak Defense", "Safety-Utility Balance", "Internal State Manipulation", "Sparse Representation Adjustment" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "s1zO0YBEF8", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Core Francisco Park", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Ekdeep Singh Lubana", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Hidenori Tanaka", "gender": "unknown", "institution": "Harvard University, Harvard University", "country": "US", "position": "Researcher" }, { "name": "Maya Okawa", "gender": "unknown", "institution": "NTT Research, Inc.", "country": "", "position": "Researcher" }, { "name": "Wei Hu", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" }, { "name": "Yang Yongyi", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" } ], "n_formula": 62, "n_formula_1": 12, "n_ref_uni": 67, "n_ref": 113, "n_ref_all": 160, "n_fig": 17, "n_tab": 1, "L_tab": 474, "n_element_tab": 57, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6665, "formula_len_all_1": 635, "len_all": 229079, "len_all_1": 75097, "len_abs": 1696, "len_title": 110, "len_sents": 69178, "len_sents_1": 36965, "n_sents": 589, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1708, "title": "Dynamics of Concept Learning and Compositional Generalization", "abs": "Prior work has shown that text-conditioned diffusion models can learn to identify and manipulate primitive concepts underlying a compositional data-generating process, enabling generalization to entirely novel, out-of-distribution compositions. \nBeyond performance evaluations, these studies develop a rich empirical phenomenology of learning dynamics, showing that models generalize sequentially, respecting the compositional hierarchy of the data-generating process. \nMoreover, concept-centric structures within the data significantly influence a model's speed of learning the ability to manipulate a concept.\nIn this paper, we aim to better characterize these empirical results from a theoretical standpoint.\nSpecifically, we propose an abstraction of prior work's compositional generalization problem by introducing a structured identity mapping (SIM) task, where a model is trained to learn the identity mapping on a Gaussian mixture with structurally organized centroids. \nWe mathematically analyze the learning dynamics of neural networks trained on this SIM task and show that, despite its simplicity, SIM's learning dynamics capture and help explain key empirical observations on compositional generalization with diffusion models identified in prior work.\nOur theory also offers several new insights---e.g., we find a novel mechanism for non-monotonic learning dynamics of test loss in early phases of training.\nWe validate our new predictions by training a text-conditioned diffusion model, bridging our simplified framework and complex generative models.\nOverall, this work establishes the SIM task as a meaningful theoretical abstraction of concept learning dynamics in modern generative models.", "keywords": [ "compositional generalization", "concept learning", "learning dynamics", "out of distribution generalization" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "s1kyHkdTmi", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Edoardo Cetin", "gender": "unknown", "institution": "Sakana AI", "country": "JP", "position": "Researcher" }, { "name": "Qi Sun", "gender": "Male", "institution": "tokyo institute of technology", "country": "JP", "position": "MS student" }, { "name": "Tianyu Zhao", "gender": "Male", "institution": "Sakana AI", "country": "JP", "position": "Researcher" }, { "name": "Yujin Tang", "gender": "Male", "institution": "Sakana AI", "country": "JP", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 6, "n_ref_uni": 58, "n_ref": 128, "n_ref_all": 190, "n_fig": 16, "n_tab": 20, "L_tab": 11201, "n_element_tab": 642, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 2811, "n_element_tab_1": 124, "formula_len_all": 635, "formula_len_all_1": 337, "len_all": 282038, "len_all_1": 61974, "len_abs": 1259, "len_title": 87, "len_sents": 83118, "len_sents_1": 27096, "n_sents": 549, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 39, "L_abs": 1174, "title": "An Evolved Universal Transformer Memory", "abs": "Prior methods propose to offset the escalating costs of modern foundation models by dropping specific parts of their contexts with hand-designed rules, while attempting to preserve their original performance. We overcome this trade-off with Neural Attention Memory Models (NAMMs), introducing a learned network for memory management that improves both the performance and efficiency of transformers. We evolve NAMMs atop pre-trained transformers to provide different latent contexts focusing on the most relevant information for individual layers and attention heads. NAMMs are universally applicable to any model using self-attention as they condition exclusively on the values in the produced attention matrices. Learning NAMMs on a small set of problems, we achieve substantial performance improvements across multiple long-context benchmarks while cutting the model's input contexts up to a fraction of the original sizes. We show the generality of our conditioning enables zero-shot transfer of NAMMs trained only on language to entirely new transformer architectures even across input modalities, with their benefits carrying over to vision and reinforcement learning.", "keywords": [ "Transformers", "Evolution", "Memory", "KV cache", "attention" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "s1f7jybVTo", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Ambroise Odonnat", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "FR", "position": "PhD student" }, { "name": "Vivien Cabannes", "gender": "Not Specified", "institution": "META", "country": "FR", "position": "Researcher" }, { "name": "Wassim Bouaziz", "gender": "Male", "institution": "Facebook", "country": "", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 25, "n_ref_uni": 51, "n_ref": 70, "n_ref_all": 570, "n_fig": 21, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 13, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3480, "formula_len_all_1": 2934, "len_all": 184266, "len_all_1": 75617, "len_abs": 2816, "len_title": 111, "len_sents": 43665, "len_sents_1": 31253, "n_sents": 355, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 769, "title": "A Visual Case Study of the Training Dynamics in Neural Networks", "abs": "This paper introduces a visual sandbox designed to explore the training dynamics of a small-scale transformer model, with the embedding dimension constrained to $d=2$\nThis restriction allows for a comprehensive two-dimensional visualization of each layer's dynamics. \nThrough this approach, we gain insights into training dynamics, circuit transferability, and the causes of loss spikes, including those induced by the high curvature of normalization layers. \nWe propose strategies to mitigate these spikes, demonstrating how good visualization facilitates the design of innovative ideas of practical interest.\nAdditionally, we believe our sandbox could assist theoreticians in assessing essential training dynamics mechanisms and integrating them into future theories.", "keywords": [ "Training dynamics", "Transformer", "Mechanistic Interpretability", "Visualization" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "s15HrqCqbr", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Bokun Wang", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Tianbao Yang", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Associate Professor" }, { "name": "Yiming Ying", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Full Professor" }, { "name": "Yunwen Lei", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 109, "n_formula_1": 42, "n_ref_uni": 70, "n_ref": 140, "n_ref_all": 191, "n_fig": 6, "n_tab": 4, "L_tab": 778, "n_element_tab": 68, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 268, "n_element_tab_1": 23, "formula_len_all": 18784, "formula_len_all_1": 4439, "len_all": 238466, "len_all_1": 74970, "len_abs": 1255, "len_title": 132, "len_sents": 59383, "len_sents_1": 31333, "n_sents": 517, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1248, "title": "On Discriminative Probabilistic Modeling for Self-Supervised Representation Learning", "abs": "We study the discriminative probabilistic modeling problem on a continuous domain for (multimodal) self-supervised representation learning. To address the challenge of computing the integral in the partition function for each anchor data, we leverage the multiple importance sampling (MIS) technique for robust Monte Carlo integration, which can recover InfoNCE-based contrastive loss as a special case. Within this probabilistic modeling framework, we conduct generalization error analysis to reveal the limitation of current InfoNCE-based contrastive loss for self-supervised representation learning and derive insights for developing better approaches by reducing the error of Monte Carlo integration. To this end, we propose a novel non-parametric method for approximating the sum of conditional densities required by MIS through convex optimization, yielding a new contrastive objective for self-supervised representation learning. Moreover, we design an efficient algorithm for solving the proposed objective. We empirically compare our algorithm to representative baselines on the contrastive image-language pretraining task. Experimental results on the CC3M and CC12M datasets demonstrate the superior overall performance of our algorithm.", "keywords": [ "Discriminative Probabilistic Modeling; Self-Supervised Representation Learning; Multiple Importance Sampling" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "s0gdfKcmoU", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Chi Jin", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Assistant Professor" }, { "name": "Jiawei Ge", "gender": "Female", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Wenzhe Li", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Yuanhao Wang", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "PhD student" } ], "n_formula": 82, "n_formula_1": 12, "n_ref_uni": 34, "n_ref": 65, "n_ref_all": 77, "n_fig": 0, "n_tab": 3, "L_tab": 278, "n_element_tab": 57, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7981, "formula_len_all_1": 1102, "len_all": 181749, "len_all_1": 74918, "len_abs": 1408, "len_title": 132, "len_sents": 55381, "len_sents_1": 35565, "n_sents": 536, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1421, "title": "Securing Equal Share: A Principled Approach for Learning Multiplayer Symmetric Games", "abs": "This paper examines multiplayer symmetric constant-sum games with more than two players in a competitive setting, including examples like Mahjong, Poker, and various board and video games. In contrast to two-player zero-sum games, equilibria in multiplayer games are neither unique nor non-exploitable, failing to provide meaningful guarantees when competing against opponents who play different equilibria or non-equilibrium strategies. This gives rise to a series of long-lasting fundamental questions in multiplayer games regarding suitable objectives, solution concepts, and principled algorithms. This paper takes an initial step toward addressing these challenges by focusing on the natural objective of *equal share*\u2014securing an expected payoff of $C/n$ in an $n$-player symmetric game with a total payoff of $C$. We rigorously identify the theoretical conditions under which achieving an equal share is tractable and design a series of efficient algorithms, inspired by no-regret learning, that *provably* attain approximate equal share across various settings. Furthermore, we provide complementary lower bounds that justify the sharpness of our theoretical results. Our experimental results highlight worst-case scenarios where meta-algorithms from prior state-of-the-art systems for multiplayer games fail to secure an equal share, while our algorithm succeeds, demonstrating the effectiveness of our approach.", "keywords": [ "Multiplayer symmetric games; Securing Equal Share" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "s0Z4csHOoE", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Bang Liu", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Assistant Professor" }, { "name": "Ge Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Lu Li", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Perouz Taslakian", "gender": "Female", "institution": "ServiceNow", "country": "CA", "position": "Researcher" }, { "name": "Suyuchen Wang", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "Tianyu Zhang", "gender": "Not Specified", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, University of Montreal", "country": "CA", "position": "PhD student" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" }, { "name": "Jie Fu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "sai rajeswar mudumba", "gender": "Male", "institution": "ServiceNow", "country": "CA", "position": "Research Scientist" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 23, "n_ref": 56, "n_ref_all": 78, "n_fig": 7, "n_tab": 5, "L_tab": 18608, "n_element_tab": 2883, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 365, "n_element_tab_1": 51, "formula_len_all": 38, "formula_len_all_1": 38, "len_all": 250608, "len_all_1": 59058, "len_abs": 1521, "len_title": 79, "len_sents": 42066, "len_sents_1": 28346, "n_sents": 290, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 31, "L_abs": 1532, "title": "VCR: Visual Caption Restoration", "abs": "We introduce Visual Caption Restoration (VCR), a novel vision-language task that challenges models to accurately restore partially obscured texts using pixel-level hints within images. This task stems from the observation that text embedded in images is intrinsically different from common visual elements and natural language due to the need to align the modalities of vision, text, and text embedded in images. While numerous works have integrated text embedded in images into visual question-answering tasks, approaches to these tasks generally rely on optical character recognition or masked language modeling, thus reducing the task to mainly text-based processing. However, text-based processing becomes ineffective in VCR as accurate text restoration depends on the combined information from provided images, context, and subtle cues from the tiny, exposed areas of masked texts. We develop a pipeline to generate synthetic images for the VCR task using image-caption pairs, with adjustable caption visibility to control the task difficulty. With this pipeline, we construct a dataset for VCR called VCR-Wiki using images with captions from Wikipedia, comprising 2.11M English and 346K Chinese entities in both easy and hard split variants. Our results reveal that current vision language models significantly lag behind human performance in the VCR task, and merely fine-tuning the models on our dataset does not lead to notable improvements. We release VCR-Wiki and the data construction code to facilitate future research.", "keywords": [ "dataset", "vision-language model", "multimodal" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "rzx3vcvlzj", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Guo Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Kai Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Mohan Xu", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Xiaolin Hu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 19, "n_formula_1": 14, "n_ref_uni": 30, "n_ref": 69, "n_ref_all": 95, "n_fig": 7, "n_tab": 12, "L_tab": 2167, "n_element_tab": 282, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1905, "n_element_tab_1": 162, "formula_len_all": 1326, "formula_len_all_1": 561, "len_all": 137326, "len_all_1": 65301, "len_abs": 1450, "len_title": 149, "len_sents": 46048, "len_sents_1": 27435, "n_sents": 405, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1457, "title": "TIGER: Time-frequency Interleaved Gain Extraction and Reconstruction for Efficient Speech Separation", "abs": "In recent years, much speech separation research has focused primarily on improving model performance. However, for low-latency speech processing systems, high efficiency is equally important. Therefore, we propose a speech separation model with significantly reduced parameters and computational costs: Time-frequency Interleaved Gain Extraction and Reconstruction network (TIGER). TIGER leverages prior knowledge to divide frequency bands and compresses frequency information. We employ a multi-scale selective attention module to extract contextual features, while introducing a full-frequency-frame attention module to capture both temporal and frequency contextual information. Additionally, to more realistically evaluate the performance of speech separation models in complex acoustic environments, we introduce a dataset called EchoSet. This dataset includes noise and more realistic reverberation (e.g., considering object occlusions and material properties), with speech from two speakers overlapping at random proportions. Experimental results showed that models trained on EchoSet had better generalization ability than those trained on other datasets to the data collected in the physical world, which validated the practical value of the EchoSet. On EchoSet and real-world data, TIGER significantly reduces the number of parameters by 94.3% and the MACs by 95.3% while achieving performance surpassing state-of-the-art (SOTA) model TF-GridNet.", "keywords": [ "Speech separation", "lightweight model", "time-frequency domain" ], "rating_list": [ 1, 6, 8, 6, 8 ], "soundness_list": [ 1, 3, 4, 3, 4 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 1, 3, 4, 2, 4 ], "confidence_list": [ 5, 4, 3, 2, 4 ] }, { "paper_id": "rztZ2QfSfJ", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Binh-Son Hua", "gender": "Male", "institution": "University of Dublin, Trinity College", "country": "IE", "position": "Assistant Professor" }, { "name": "Jipeng Zhang", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Runtao Liu", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yang Wu", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Zheng Ziqiang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Sai-kit Yeung", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 66, "n_ref": 135, "n_ref_all": 173, "n_fig": 14, "n_tab": 8, "L_tab": 3040, "n_element_tab": 224, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1013, "n_element_tab_1": 62, "formula_len_all": 223, "formula_len_all_1": 222, "len_all": 248665, "len_all_1": 63544, "len_abs": 445, "len_title": 0, "len_sents": 76008, "len_sents_1": 31674, "n_sents": 544, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 132, "L_abs": 1300, "title": "NeMal: Never ending Marine Learning - Unleashing the Power of Controllable Image Synthesis for Promoting Marine Visual Understanding", "abs": "The relentless pursuit of marine learning is required by the essential need to understand and protect the complex marine ecosystems that cover over 70% of the surface of our planet. Due to the specific underwater/marine environments, the data collection and labeling are expensive and labor-intensive, also limited to user groups with special equipment. Existing marine visual learning just optimizes models from a small set of marine data with human labels, which cannot fit the essence of ongoing marine exploration. In this work, we propose NeMal, a \\underline{N}ever-\\underline{e}nding \\underline{Ma}rine \\underline{L}earning system that harnesses controllable image synthesis and efficient foundation models to perform never-ending marine visual synthesis and understanding. Based on NeMal, we produce MarineSynth, which is the first large-scale marine synthetic dataset to date, featuring more than 4 million unique text prompts and corresponding text-to-image outputs with pseudo labels from text prompts or foundation models. The experiments on downstream classification, segmentation, and vision-language understanding tasks demonstrate the promise of utilizing synthetic data to promote marine visual understanding, significantly reducing human efforts in both data collection and labeling.", "keywords": [ "never ending marine learning", "controllable image synthesis", "foundation models", "vision language analysis" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 1, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "rzbSNDXgGD", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Bo Zheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Jia Jia", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Junliang Xing", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Min Zhou", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Shikun Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Tiezheng Ge", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xiaoyu Qin", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Xubin Li", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Zijie Ye", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zixuan Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 45, "n_formula_1": 24, "n_ref_uni": 21, "n_ref": 41, "n_ref_all": 57, "n_fig": 9, "n_tab": 3, "L_tab": 1214, "n_element_tab": 106, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 1067, "n_element_tab_1": 11, "formula_len_all": 3139, "formula_len_all_1": 1284, "len_all": 106412, "len_all_1": 51889, "len_abs": 1407, "len_title": 113, "len_sents": 32963, "len_sents_1": 21918, "n_sents": 254, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1267, "title": "Minimal Impact ControlNet: Advancing Multi-ControlNet Integration", "abs": "With the advancement of diffusion models, there is a growing demand for high-quality, controllable image generation, particularly through methods that utilize one or multiple control signals based on ControlNet. However, in current ControlNet training, each control is designed to influence all areas of an image, which can lead to conflicts when different control signals are expected to manage different parts of the image in practical applications. This issue is especially pronounced with edge-type control conditions, where regions lacking boundary information often represent low-frequency signals, referred to as silent control signals. When combining multiple ControlNets, these silent control signals can suppress the generation of textures in related areas, resulting in suboptimal outcomes. To address this problem, we propose Minimal Impact ControlNet. Our approach mitigates conflicts through three key strategies: constructing a balanced dataset, combining and injecting feature signals in a balanced manner, and addressing the asymmetry in the score function\u2019s Jacobian matrix induced by ControlNet. These improvements enhance the compatibility of control signals, allowing for freer and more harmonious generation in areas with silent control signals.", "keywords": [ "generative models", "ControlNet" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 4, 4, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "rynb4Vn8rb", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Alan Aspuru-Guzik", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Director" }, { "name": "Andreas Burger", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Luca Thiede", "gender": "Male", "institution": "Univerity of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Nandita Vijaykumar", "gender": "Female", "institution": "University of Toronto", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 20, "n_formula_1": 14, "n_ref_uni": 32, "n_ref": 84, "n_ref_all": 98, "n_fig": 6, "n_tab": 6, "L_tab": 2371, "n_element_tab": 203, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1170, "n_element_tab_1": 134, "formula_len_all": 1062, "formula_len_all_1": 706, "len_all": 139672, "len_all_1": 62494, "len_abs": 1110, "len_title": 131, "len_sents": 43325, "len_sents_1": 28352, "n_sents": 343, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1109, "title": "DEQuify your force field: Towards efficient simulations using deep equilibrium models", "abs": "Machine learning force fields show great promise in enabling more accurate force fields than manually derived ones for molecular dynamics simulations. \nState-of-the-art approaches for ML force fields stack many equivariant graph neural network layers, resulting in long inference times and high memory costs. This work aims to improve these two aspects while simultaneously reaching higher accuracy.\nOur key observation is that successive states in molecular dynamics simulations are extremely similar, but typical architectures treat each step independently, disregarding this information.\nWe show how deep equilibrium models (DEQs) can exploit this temporal correlation by recycling neural network features from previous time steps. \nSpecifically, we turn a state-of-the-art force field architecture into a DEQ, enabling us to improve both accuracy and speed by $10\\%-20\\%$ on the MD17, MD22, and OC20 200k datasets. \nCompared to conventional approaches, DEQs are also naturally more memory efficient, facilitating the training of more expressive models on larger systems given limited GPU memory resources.", "keywords": [ "Machine Learning Force Fields", "Deep Equilibrium Models" ], "rating_list": [ 3, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "rySLejeB1k", "primary_area": "applications to neuroscience & cognitive science", "n_author": 8, "author_info_list": [ { "name": "Anna Wang Roe", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Haixin Zhong", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Postdoc" }, { "name": "Mingyi Huang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Wei P Dai", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Researcher" }, { "name": "Yuchao Huang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "yuguo yu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Wang Haoyu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Wang Rubin", "gender": "Male", "institution": "East China University of Science and Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 24, "n_formula_1": 12, "n_ref_uni": 72, "n_ref": 141, "n_ref_all": 190, "n_fig": 31, "n_tab": 5, "L_tab": 1237, "n_element_tab": 104, "n_fig_1": 18, "n_tab_1": 2, "L_tab_1": 569, "n_element_tab_1": 38, "formula_len_all": 1073, "formula_len_all_1": 530, "len_all": 190028, "len_all_1": 61398, "len_abs": 1511, "len_title": 139, "len_sents": 48666, "len_sents_1": 29614, "n_sents": 350, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 73, "L_abs": 1517, "title": "Emergent Orientation Maps \u2014\u2014 Mechanisms, Coding Efficiency and Robustness", "abs": "Extensive experimental studies have shown that in lower mammals, neuronal orientation preference in the primary visual cortex is organized in a disordered \"salt-and-pepper\" pattern. In contrast, higher-order mammals display a continuous variation in orientation preference, forming structured pinwheel-like patterns. Despite these observations, the spiking mechanisms underlying the emergence of these distinct topological structures and their functional roles in visual processing remain poorly understood. To address this, we developed a self-evolving spiking neural network model with Hebbian plasticity, trained using physiological parameters characteristic of rodents, cats, and primates, including retinotopy, neuronal morphology, and connectivity patterns. Our results identify critical factors, such as the degree of input visual field overlap, neuronal density, and the balance between localized connectivity and long-range competition, that determine the emergence of either salt-and-pepper or pinwheel-like topologies. Furthermore, we demonstrate that pinwheel structures exhibit lower wiring costs and enhanced sparse coding capabilities compared to salt-and-pepper organizations. They also maintain greater coding robustness against noise in naturalistic visual stimuli. These findings suggest that such topological structures confer significant computational advantages in visual processing and highlight their potential application in the design of brain-inspired deep learning networks and algorithms.", "keywords": [ "Vision", "Energy Efficient Coding", "Neural Network", "Sensory Coding", "Spiking Mechanism" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "ryKrRCbcCX", "primary_area": "interpretability and explainable AI", "n_author": 12, "author_info_list": [ { "name": "Adithya Kulkarni", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "", "position": "Postdoc" }, { "name": "Chang-Tien Lu", "gender": "Male", "institution": "Virginia Tech", "country": "US", "position": "Full Professor" }, { "name": "Dawei Zhou", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "Assistant Professor" }, { "name": "Feng Chen", "gender": "Male", "institution": "University of Texas, Dallas", "country": "", "position": "Associate Professor" }, { "name": "Jianfeng He", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Jin-Hee Cho", "gender": "Female", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "Associate Professor" }, { "name": "Lifu Huang", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Assistant Professor" }, { "name": "Ming Jin", "gender": "unknown", "institution": "Virginia Tech", "country": "", "position": "Assistant Professor" }, { "name": "Mohammad Beigi", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "PhD student" }, { "name": "Sijia Wang", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Ying Shen", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Zihao Lin", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 92, "n_ref": 162, "n_ref_all": 165, "n_fig": 1, "n_tab": 1, "L_tab": 353, "n_element_tab": 43, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 986, "n_element_tab_1": 106, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 198200, "len_all_1": 66265, "len_abs": 1865, "len_title": 143, "len_sents": 34550, "len_sents_1": 34467, "n_sents": 188, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1283, "title": "Rethinking the Uncertainty: A Critical Review and Analysis in the Era of Large Language Models", "abs": "In recent years, Large Language Models (LLMs) have become fundamental to a broad spectrum of artificial intelligence applications. As the use of LLMs expands, precisely estimating the uncertainty in their predictions has become crucial. Current methods often struggle to accurately identify, measure, and address the true uncertainty, with many focusing primarily on estimating model confidence. This discrepancy is largely due to an incomplete understanding of where, when, and how uncertainties are injected into models. This paper introduces a comprehensive framework specifically designed to identify and understand the types and sources of uncertainty, aligned with the unique characteristics of LLMs. Our framework enhances the understanding of the diverse landscape of uncertainties by systematically categorizing and defining each type, establishing a solid foundation for developing targeted methods that can precisely quantify these uncertainties. We also provide a detailed introduction to key related concepts and examine the limitations of current methods in mission-critical and safety-sensitive applications. The paper concludes with a perspective on future directions aimed at enhancing the reliability and practical adoption of these methods in real-world scenarios.", "keywords": [ "Uncertainty Estimation", "Confidence Estimation" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "ryIHtXE9uG", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Abhimanyu Das", "gender": "Male", "institution": "Research, Google", "country": "", "position": "Researcher" }, { "name": "Matthew Faw", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "PhD student" }, { "name": "Rajat Sen", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Yichen Zhou", "gender": "unknown", "institution": "Google", "country": "", "position": "Data Scientist" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 33, "n_ref": 93, "n_ref_all": 115, "n_fig": 9, "n_tab": 12, "L_tab": 5254, "n_element_tab": 666, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 199, "n_element_tab_1": 4, "formula_len_all": 606, "formula_len_all_1": 606, "len_all": 142914, "len_all_1": 65056, "len_abs": 1049, "len_title": 105, "len_sents": 41829, "len_sents_1": 31549, "n_sents": 297, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1056, "title": "In-context Fine-tuning for Time-series Foundation Models", "abs": "Motivated by the recent success of time-series foundation models for zero-shot forecasting, we present a methodology for _in-context fine-tuning_ of a time-series foundation model. In particular, we design a pretrained foundation model that can be prompted (at inference time) with multiple time-series examples, in order to forecast a target time-series into the future. Our foundation model is specifically trained to utilize examples from multiple related time-series in its context window (in addition to the history of the target time-series) to help it adapt to the specific distribution of the target domain at inference time. We show that such a foundation model that uses in-context examples at inference time can obtain much better performance on popular forecasting benchmarks compared to supervised deep learning methods, statistical models as well as other time-series foundation models. Interestingly, our in-context fine-tuning approach even rivals the performance of a foundation model that is explicitly fine-tuned on the target domain.", "keywords": [ "time-series", "foundation models", "zero-shot", "few-shot", "in-context" ], "rating_list": [ 8, 6, 5, 3, 6 ], "soundness_list": [ 3, 3, 3, 1, 4 ], "presentation_list": [ 3, 3, 3, 2, 1 ], "contribution_list": [ 4, 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 5, 4 ] }, { "paper_id": "rxeh2tZ8lW", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Alessandro G Allievi", "gender": "unknown", "institution": "University of Texas at Austin", "country": "", "position": "Postdoc" }, { "name": "Andrew Levy", "gender": "Male", "institution": "Brown University", "country": "", "position": "PhD Student" }, { "name": "George Konidaris", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 8, "n_ref_uni": 37, "n_ref": 82, "n_ref_all": 99, "n_fig": 8, "n_tab": 2, "L_tab": 203, "n_element_tab": 18, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 186, "n_element_tab_1": 18, "formula_len_all": 1322, "formula_len_all_1": 865, "len_all": 149388, "len_all_1": 71061, "len_abs": 1541, "len_title": 112, "len_sents": 54182, "len_sents_1": 38400, "n_sents": 327, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1504, "title": "Learning Large Skillsets in Stochastic Settings with Empowerment", "abs": "General purpose agents need to be able to execute large skillsets in stochastic settings. Given that the mutual information between skills and states measures the number of distinct skills in a skillset, a compelling objective for learning a diverse skillset is to find the skillset with the largest mutual information between skills and states. The problem is that the two main unsupervised approaches for maximizing this mutual information objective, Empowerment-based skill learning and Unsupervised Goal-Conditioned Reinforcement Learning, only maximize loose lower bounds on the mutual information, which can impede diverse skillset learning. We propose a new empowerment objective, Skillset Empowerment, that maximizes a tighter bound on the mutual information between skills and states. For any proposed skillset, the tighter bound on mutual information is formed by replacing the posterior distribution of the proposed skillset with a variational distribution that is conditioned on the proposed skillset and trained to match the posterior of the proposed skillset. Maximizing our mutual information lower bound objective is a bandit problem in which actions are skillsets and the rewards are our mutual information objective, and we optimize this bandit problem with a new actor-critic architecture. We show empirically that our approach is able to learn large abstract skillsets in stochastic domains, including ones with high-dimensional observations, in contrast to existing approaches.", "keywords": [ "Empowerment", "Unsupervised Skill Learning", "Unsupervised Reinforcement Learning", "Self-Supervised Reinforcement Learning" ], "rating_list": [ 3, 5, 8, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 1, 3, 2 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "rxVvRBgqmS", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Jianke Zhu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Qijun Gan", "gender": "Male", "institution": "College of Computer Science and Technology, Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Shengtao Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Song Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 85, "n_ref": 112, "n_ref_all": 123, "n_fig": 6, "n_tab": 3, "L_tab": 1343, "n_element_tab": 122, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1480, "n_element_tab_1": 125, "formula_len_all": 397, "formula_len_all_1": 336, "len_all": 172309, "len_all_1": 68775, "len_abs": 1359, "len_title": 134, "len_sents": 39366, "len_sents_1": 32447, "n_sents": 305, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 85, "L_abs": 1335, "title": "PianoMotion10M: Dataset and Benchmark for Hand Motion Generation in Piano Performance", "abs": "Recently, artificial intelligence techniques for education have been received increasing attentions, while it still remains an open problem to design the effective music instrument instructing systems. Although key presses can be directly derived from sheet music, the transitional movements among key presses require more extensive guidance in piano performance. In this work, we construct a piano-hand motion generation benchmark to guide hand movements and fingerings for piano playing. To this end, we collect an annotated dataset, PianoMotion10M, consisting of 116 hours of piano playing videos from a bird's-eye view with 10 million annotated hand poses. We also introduce a powerful baseline model that generates hand motions from piano audios through a position predictor and a position-guided gesture generator. Furthermore, a series of evaluation metrics are designed to assess the performance of the baseline model, including motion similarity, smoothness, positional accuracy of left and right hands, and overall fidelity of movement distribution. Despite that piano key presses with respect to music scores or audios are already accessible, PianoMotion10M aims to provide guidance on piano fingering for instruction purposes. The dataset and source code can be accessed at https://github.com/PianoMotion10M/PianoMotion10M.", "keywords": [ "Hand pose estimation", "piano music", "motion generation" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "rxUz2DaulF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Da Yin", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Kai-Wei Chang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Associate Professor" }, { "name": "Yao Tang", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Undergrad student" }, { "name": "Yizhou Sun", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "Ziniu Hu", "gender": "Male", "institution": "xAI", "country": "US", "position": "Member of Technical Staff" }, { "name": "Zongyu Lin", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Xingcheng Yao", "gender": "unknown", "institution": "UCLA Computer Science Department, University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 70, "n_ref_all": 93, "n_fig": 7, "n_tab": 6, "L_tab": 1458, "n_element_tab": 86, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 384, "n_element_tab_1": 22, "formula_len_all": 911, "formula_len_all_1": 559, "len_all": 132586, "len_all_1": 65225, "len_abs": 1271, "len_title": 63, "len_sents": 39294, "len_sents_1": 29849, "n_sents": 300, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1264, "title": "Q* Agent: Optimizing Language Agents with Q-Guided Exploration", "abs": "Language agents have become a promising solution to complex interactive tasks. One of the key ingredients to the success of language agents is the reward model on the trajectory of the agentic workflow, which provides valuable guidance during training or inference. However, due to the lack of annotations of intermediate interactions, most existing works use an outcome reward model to optimize policies across entire trajectories. This may lead to sub-optimal policies and hinder the overall performance. To address this, we propose Q\\*Agent, leveraging an estimated Q value to generate intermediate annotations for open language agents. \nBy introducing a reasoning tree and performing process reward modeling, Q\\*Agent provides effective intermediate guidance for each step. This guidance aims to automatically annotate data in a step-wise manner.\nBesides, we propose a Q-guided generation strategy that can significantly boost model performance by providing process guidance during inference.\nNotably, even with almost half the annotated data, Q\\*Agent retains strong performance, demonstrating its efficiency in handling limited supervision. We also empirically demonstrate that Q\\*Agent can lead to more accurate decision making through qualitative analysis.", "keywords": [ "agent", "large language model", "q-learning", "self-training" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 5, 4, 2 ] }, { "paper_id": "rx0TCew0Lj", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 2, "author_info_list": [ { "name": "Eleonora Giunchiglia", "gender": "Female", "institution": "Imperial College London", "country": "GB", "position": "Lecturer" }, { "name": "Mihaela C Stoian", "gender": "unknown", "institution": "Department of Computer Science, University of Oxford", "country": "", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 29, "n_ref_uni": 49, "n_ref": 75, "n_ref_all": 124, "n_fig": 9, "n_tab": 19, "L_tab": 10222, "n_element_tab": 770, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1587, "n_element_tab_1": 70, "formula_len_all": 4549, "formula_len_all_1": 3058, "len_all": 213132, "len_all_1": 73411, "len_abs": 1196, "len_title": 160, "len_sents": 61514, "len_sents_1": 30359, "n_sents": 543, "n_sents_1": 274, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 112, "L_abs": 1203, "title": "Beyond the convexity assumption: Realistic tabular data generation under quantifier-free real linear constraints", "abs": "Synthetic tabular data generation has traditionally been a challenging problem due to the high complexity of the underlying distributions that characterise this type of data. Despite recent advances in deep generative models (DGMs), existing methods often fail to produce realistic datapoints that are well-aligned with available background knowledge.\nIn this paper, we address this limitation by introducing Disjunctive Refinement Layer (DRL), a novel layer designed\nto enforce the alignment of generated data with the background knowledge specified in user-defined constraints.\nDRL is the first method able to automatically make deep learning models inherently compliant with constraints as expressive as quantifier-free linear formulas, which can define non-convex and even disconnected spaces. \nOur experimental analysis shows that DRL not only guarantees constraint satisfaction but also improves efficacy in downstream tasks. Notably, when applied to DGMs that frequently violate constraints, DRL eliminates violations entirely. Further, it improves performance metrics by up to 21.4\\% in F1-score and 20.9\\% in Area Under the ROC Curve, thus demonstrating its practical impact on data generation.", "keywords": [ "tabular data generation", "neuro-symbolic AI", "informed machine learning", "safe AI" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 2, 2, 2 ] }, { "paper_id": "rwqShzb9li", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Aaron Schein", "gender": "Male", "institution": "University of Chicago", "country": "", "position": "Assistant Professor" }, { "name": "James Evans", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Full Professor" }, { "name": "Junsol Kim", "gender": "unknown", "institution": "University of Chicago", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 55, "n_ref": 121, "n_ref_all": 157, "n_fig": 12, "n_tab": 9, "L_tab": 1188, "n_element_tab": 145, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 331, "formula_len_all_1": 284, "len_all": 214257, "len_all_1": 74207, "len_abs": 1521, "len_title": 128, "len_sents": 77632, "len_sents_1": 37182, "n_sents": 545, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 79, "L_abs": 1574, "title": "Linear Representations of Political Perspective Emerge in Large Language Models", "abs": "Large language models (LLMs) have demonstrated the ability to simulate responses aligned with human subjective perspectives, such as liberal or conservative ideologies in American politics. Our study reveals that LLMs achieve this by learning a ``geometry of perspective'' that linearly represents subjective perspectives in the activation space, where similar simulated perspectives are represented closer to each other. Specifically, we probe the hidden layers of open, transformer-based LLMs (\\texttt{Llama-2-7b-chat, Mistral-7b-instruct, Vicuna-7b}) when prompted to generate texts under the ideological perspective of distinct politicians. We find a set of attention heads that represent U.S. ideological slant, which is primarily located in the middle layers known to encode high-level concepts and tasks. The activation of these attention heads, when prompted about U.S.~politicians and media outlets, linearly correlates with existing measures of their ideological slant. We use this activation to detect the ideological slant implicitly adopted by an LLM as it is generating each token. We further show that by intervening on these attention heads, we can tune LLM output to any position along the linear dimension from a liberal to conservative ideological perspective. Our research shows that political ideology serves as a fundamental dimension of LLM representations, and present an interpretability method to identify, monitor, and control the subjective perspective used to generate text. Code: https://osf.io/us9yx/?view_only=cf0fdcdb123e4d6bb7d10a64be5c1a09", "keywords": [ "large language model", "political perspective", "ideology", "representation learning" ], "rating_list": [ 6, 8, 10, 6 ], "soundness_list": [ 3, 4, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "rwmwFnmjAX", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Ian Reid", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" }, { "name": "Jiahua Dong", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Meng Cao", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Tiancai Wang", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Xiangyu Zhang", "gender": "Male", "institution": "MEGVII Technology", "country": "CN", "position": "Principal Researcher" }, { "name": "Xiaodan Liang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Yingfei Liu", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Yuyang Liu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "MS student" }, { "name": "Henghui Ding", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 11, "n_ref_uni": 80, "n_ref": 167, "n_ref_all": 201, "n_fig": 11, "n_tab": 14, "L_tab": 4485, "n_element_tab": 641, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 4762, "n_element_tab_1": 192, "formula_len_all": 1675, "formula_len_all_1": 745, "len_all": 267085, "len_all_1": 64111, "len_abs": 2312, "len_title": 126, "len_sents": 55159, "len_sents_1": 24794, "n_sents": 455, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1849, "title": "Continual LLaVA: Continual Instruction Tuning in Large Vision-Language Models", "abs": "Instruction tuning constitutes a prevalent technique for tailoring Large Vision Language Models (LVLMs) to meet individual task requirements. To date, most of the existing approaches are confined to single-task adaptation, whereas the requirements in real-world scenarios are inherently varied and continually evolving. Thus an ideal LVLM should sustain continual instruction tuning in the face of stream-task distributions (i.e., different domains, emerging capabilities, and new datasets) while minimizing the forgetting of previously acquired knowledge. To achieve this, we propose a new benchmark for COntinuAl inStruction Tuning on LVLMs (COAST), which encompasses the aforementioned domain-incremental, capability-incremental, and dataset-incremental configurations. In terms of methodology, we propose Continual LLaVA, a rehearsal-free method tailored for continual instruction tuning in LVLMs. To circumvent the additional overhead associated with experience replay, we freeze LVLMs and construct the dual increment embeddings for each input instruction to facilitate parameter-efficient tuning. Specifically, the increment embeddings can be decomposed into two principal components: 1) intrinsic increment embeddings to encode task-specific characteristics. To achieve this, we set up a low-rank pool containing candidate embeddings, from which we select the relevant ones based on their similarity with the user instructions; 2) contextual increment embeddings to investigate the inter-dependencies across tasks. In this regard, the low-rank embeddings chosen in the previous tasks are aggregated via learnable weighted sum to provide complementary hints. Extensive experiments indicate that the proposed Continual LLaVA outperforms previous methods by significantly reducing the forgetting during the continual instruction tuning process.", "keywords": [ "Large Vision-Language Models", "Instruction Tuning", "Continual Learning" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "rwdeKOdAwY", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Hongzhi Wang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Tianrui Yu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Weidong Geng", "gender": "Male", "institution": "College of Computer Science and Technology, Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Xiubo Liang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 14, "n_ref_uni": 43, "n_ref": 55, "n_ref_all": 68, "n_fig": 4, "n_tab": 4, "L_tab": 1353, "n_element_tab": 315, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 871, "n_element_tab_1": 223, "formula_len_all": 882, "formula_len_all_1": 901, "len_all": 135058, "len_all_1": 57882, "len_abs": 1487, "len_title": 111, "len_sents": 27819, "len_sents_1": 25262, "n_sents": 225, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1494, "title": "RetFormer: Enhancing Multimodal Retrieval for Image Recognition", "abs": "The expansion of Transformers and the collection of high-quality multimodal datasets have propelled deep neural networks to achieve unprecedented performance in vision and language tasks. However, applying these advances is non-trivial in real-world applications. The extensive number of parameters complicates model updates, and real-world data often features a long-tailed distribution along with noisy labels. To address the above issues, we propose to explore the internal structure of the neural network for learning with sample relationships, rather than just increasing the number of model parameters. Specifically, we introduce RetFormer, a model enhanced with a multimodal knowledge base for storing world knowledge, and a retrieval cross-fusion module designed to establish robust multimodal sample relationships by leveraging content from the knowledge base. RetFormer establishes a robust relationship between image and text modalities by integrating information from external knowledge bases into the model's decision-making process, thus overcoming the limitations of traditional approaches on model size and datasets. Our experiments demonstrate the benefits of integrating large-scale image-text datasets into vision tasks and exemplify the importance of modeling the relationship between image and text modalities. We have evaluated our approach on the task of long-tailed recognition and learning with noisy labels and have shown that it achieves state-of-the-art accuracies.", "keywords": [ "retrieval-augmented", "long-tailed learning" ], "rating_list": [ 3, 3, 5, 3, 3 ], "soundness_list": [ 1, 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 1, 1 ], "confidence_list": [ 4, 3, 3, 5, 4 ] }, { "paper_id": "rwNzSB3sDt", "primary_area": "optimization", "n_author": 12, "author_info_list": [ { "name": "Baigui Sun", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Cheng Tan", "gender": "Male", "institution": "Zhejiang University & Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Di Wu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Ge Wang", "gender": "unknown", "institution": "WESTLAKE UNIVERSITY", "country": "", "position": "PhD student" }, { "name": "Juanxi Tian", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Intern" }, { "name": "Siyuan Li", "gender": "Male", "institution": "Westlake University & Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Tao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Weiyang Jin", "gender": "Male", "institution": "New York University", "country": "US", "position": "Intern" }, { "name": "Yang Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher at Alibaba Group" }, { "name": "Zedong Wang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zicheng Liu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Stan Z. Li", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Chair Professor" } ], "n_formula": 46, "n_formula_1": 11, "n_ref_uni": 89, "n_ref": 231, "n_ref_all": 255, "n_fig": 6, "n_tab": 15, "L_tab": 7050, "n_element_tab": 411, "n_fig_1": 5, "n_tab_1": 10, "L_tab_1": 8408, "n_element_tab_1": 248, "formula_len_all": 3481, "formula_len_all_1": 660, "len_all": 288374, "len_all_1": 68365, "len_abs": 1181, "len_title": 106, "len_sents": 70566, "len_sents_1": 26231, "n_sents": 489, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1184, "title": "Switch EMA: A Free Lunch for Better Flatness and Sharpness", "abs": "Exponential Moving Average (EMA) is a widely used weight averaging (WA) regularization to learn flat optima for better generalizations without extra cost in deep neural network (DNN) optimization. Despite achieving better flatness, existing WA methods might fall into worse final performances or require extra test-time computations. This work unveils the full potential of EMA with a single line of modification, i.e., switching the EMA parameters to the original model after each epoch, dubbed as Switch EMA (SEMA). From both theoretical and empirical aspects, we demonstrate that SEMA can help DNNs to reach generalization optima that better trade-off between flatness and sharpness.\nTo verify the effectiveness of SEMA, we conduct comparison experiments with discriminative, generative, and regression tasks on vision and language datasets, including image classification, self-supervised learning, object detection and segmentation, image generation, video prediction, attribute regression, and language modeling. Comprehensive results with popular optimizers and networks show that SEMA is a free lunch for DNN training by improving performances and boosting convergence speeds.", "keywords": [ "Regularization", "Exponential Moving Average", "Weight Averaging", "Optimization", "Optimizer" ], "rating_list": [ 3, 6, 8, 5 ], "soundness_list": [ 3, 4, 4, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 2, 4, 4, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "rwAEQWEqkX", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Bowen Zhang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Research Scientist" }, { "name": "Haotian Zhang", "gender": "Male", "institution": "Apple AI/ML", "country": "US", "position": "Researcher" }, { "name": "Hong-You Chen", "gender": "unknown", "institution": "Apple AI/ML", "country": "US", "position": "Researcher" }, { "name": "Keen You", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Marcin Eichner", "gender": "Male", "institution": "Apple", "country": "CH", "position": "Researcher" }, { "name": "Meng Cao", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Xinze Wang", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Machine Learning Engineer" }, { "name": "Yinfei Yang", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Zhe Gan", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Zhengfeng Lai", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 96, "n_ref_all": 130, "n_fig": 4, "n_tab": 5, "L_tab": 3697, "n_element_tab": 295, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1957, "n_element_tab_1": 155, "formula_len_all": 621, "formula_len_all_1": 636, "len_all": 174251, "len_all_1": 66721, "len_abs": 1480, "len_title": 97, "len_sents": 53507, "len_sents_1": 30715, "n_sents": 375, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1487, "title": "Contrastive Localized Language-Image Pre-Training", "abs": "Contrastive Language-Image Pre-training (CLIP) has been a celebrated method for training vision encoders to generate image/text representations facilitating various applications. Recently, CLIP has been widely adopted as the vision backbone of multimodal large language models (MLLMs) to connect image inputs for language interactions. The success of CLIP as a vision-language foundation model relies on aligning web-crawled noisy text annotations at image levels. Nevertheless, such criteria may become insufficient for downstream tasks in need of fine-grained vision representations, especially when region-level understanding is demanding for MLLMs. In this paper, we improve the localization capability of CLIP with several advances. We propose a pre-training method called Contrastive Localized Language-Image Pre-training (CLOC) by complementing CLIP with region-text contrastive loss and modules. We formulate a new concept, promptable embeddings, of which the encoder produces image embeddings easy to transform into region representations given spatial hints. To support large-scale pre-training, we design a visually-enriched and spatially-localized captioning framework to effectively generate region-text pseudo-labels at scale. By scaling up to billions of annotated images, CLOC enables high-quality regional embeddings for image region recognition and retrieval tasks, and can be a drop-in replacement of CLIP to enhance MLLMs, especially on referring and grounding tasks.", "keywords": [ "pre-training", "CLIP", "multimodal" ], "rating_list": [ 1, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "rvxWEbTtRY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Lingao Xiao", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Yalun Dai", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "MS student" }, { "name": "Ivor W Tsang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "yang he", "gender": "Male", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "SG", "position": "Researcher" } ], "n_formula": 30, "n_formula_1": 15, "n_ref_uni": 64, "n_ref": 103, "n_ref_all": 138, "n_fig": 14, "n_tab": 16, "L_tab": 10206, "n_element_tab": 1501, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 1599, "n_element_tab_1": 95, "formula_len_all": 1424, "formula_len_all_1": 809, "len_all": 214761, "len_all_1": 51240, "len_abs": 1333, "len_title": 103, "len_sents": 51683, "len_sents_1": 22764, "n_sents": 419, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1199, "title": "Training-Free Dataset Pruning for Instance Segmentation", "abs": "Existing dataset pruning techniques primarily focus on classification tasks, limiting their applicability to more complex and practical tasks like instance segmentation. Instance segmentation presents three key challenges: pixel-level annotations, instance area variations, and class imbalances, which significantly complicate dataset pruning efforts. Directly adapting existing classification-based pruning methods proves ineffective due to their reliance on time-consuming model training process. To address this, we propose a novel **T**raining-**F**ree **D**ataset **P**runing (**TFDP**) method for instance segmentation. Specifically, we leverage shape and class information from image annotations to design a Shape Complexity Score (SCS), refining it into a Scale-Invariant (SI-SCS) and Class-Balanced (CB-SCS) versions to address instance area variations and class imbalances, all without requiring model training. We achieve state-of-the-art results on VOC 2012, Cityscapes, and MS COCO datasets, generalizing well across CNN and Transformer architectures. Remarkably, our approach accelerates the pruning process by an average of **1349$\\times$** on COCO compared to the adapted baselines.", "keywords": [ "dataset pruning", "instance segmentation" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "rvvSSmGIFS", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Benito van der Zander", "gender": "Male", "institution": "Universit\u00e4t zu L\u00fcbeck", "country": "", "position": "Postdoc" }, { "name": "Julian D\u00f6rfler", "gender": "Male", "institution": "Saarland University, Universit\u00e4t des Saarlandes", "country": "DE", "position": "PhD student" }, { "name": "Maciej Liskiewicz", "gender": "Male", "institution": "Universit\u00e4t zu L\u00fcbeck", "country": "DE", "position": "Full Professor" }, { "name": "Markus Blaeser", "gender": "Male", "institution": "Saarland University, Universit\u00e4t des Saarlandes", "country": "DE", "position": "Full Professor" } ], "n_formula": 35, "n_formula_1": 14, "n_ref_uni": 25, "n_ref": 64, "n_ref_all": 82, "n_fig": 0, "n_tab": 4, "L_tab": 694, "n_element_tab": 42, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 226, "n_element_tab_1": 6, "formula_len_all": 2803, "formula_len_all_1": 1485, "len_all": 184969, "len_all_1": 86148, "len_abs": 1254, "len_title": 156, "len_sents": 71223, "len_sents_1": 38493, "n_sents": 648, "n_sents_1": 319, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 108, "L_abs": 1262, "title": "From Probability to Counterfactuals: the Increasing Complexity of Satisfiability in Pearl's Causal Hierarchy", "abs": "The framework of Pearl's Causal Hierarchy (PCH) formalizes three types of reasoning: probabilistic (i.e. purely observational), interventional, and counterfactual, that reflect the progressive sophistication of human thought regarding causation. We investigate the computational complexity aspects of reasoning in this framework focusing mainly on satisfiability problems expressed in probabilistic and causal languages across the PCH. That is, given a system of formulas in the standard probabilistic and causal languages, does there exist a model satisfying the formulas? \n\nOur main contribution is to prove the exact computational complexities showing that languages allowing addition and marginalization (via the summation operator) yield NP^{PP}-, PSPACE-, and NEXP-complete satisfiability problems, depending on the level of the PCH. These are the first results to demonstrate a strictly increasing complexity across the PCH: from probabilistic to causal and counterfactual reasoning. On the other hand, in the case of full languages, i.e.~allowing addition, marginalization, and multiplication, we show that the satisfiability for the counterfactual level remains the same as for the probabilistic and causal levels, solving an open problem in the field.", "keywords": [ "complexity", "causal reasoning", "Pearl's Causal Hierarchy" ], "rating_list": [ 8, 5, 3, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "rvhu4V7yrX", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Junze Yin", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Lichen Zhang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Mingquan Ye", "gender": "unknown", "institution": "NII, Tokyo Institute of Technology", "country": "JP", "position": "Postdoc" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" } ], "n_formula": 219, "n_formula_1": 31, "n_ref_uni": 42, "n_ref": 93, "n_ref_all": 157, "n_fig": 0, "n_tab": 4, "L_tab": 215, "n_element_tab": 17, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 133, "n_element_tab_1": 11, "formula_len_all": 18291, "formula_len_all_1": 2385, "len_all": 283695, "len_all_1": 72686, "len_abs": 1295, "len_title": 136, "len_sents": 74916, "len_sents_1": 30381, "n_sents": 884, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1390, "title": "Efficient Alternating Minimization with Applications to Weighted Low Rank Approximation", "abs": "Weighted low rank approximation is a fundamental problem in numerical linear algebra, and it has many applications in machine learning. Given a matrix $M \\in \\mathbb{R}^{n \\times n}$, a non-negative weight matrix $W \\in \\mathbb{R}_{\\geq 0}^{n \\times n}$, a parameter $k$, the goal is to output two matrices $X,Y\\in \\mathbb{R}^{n \\times k}$ such that $\\\\| W \\circ (M - X Y^\\top) \\\\|_F$ is minimized, where $\\circ$ denotes the Hadamard product. It naturally generalizes the well-studied low rank matrix completion problem. Such a problem is known to be NP-hard and even hard to approximate assuming the Exponential Time Hypothesis. Meanwhile, alternating minimization is a good heuristic solution for weighted low rank approximation. In particular, [Li, Liang and Risteski, ICML'16] shows that, under mild assumptions, alternating minimization does provide provable guarantees. In this work, we develop an efficient and robust framework for alternating minimization that allows the alternating updates to be computed approximately. For weighted low rank approximation, this improves the runtime of [Li, Liang and Risteski, ICML'16] from $\\\\|W\\\\|_0k^2$ to $\\\\|W\\\\|_0 k$ where $\\\\|W\\\\|_0$ denotes the number of nonzero entries of the weight matrix. At the heart of our framework is a high-accuracy multiple response regression solver together with a robust analysis of alternating minimization.", "keywords": [ "weighted low rank approximation", "sketching", "alternating minimization" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "rvXdGL4pCJ", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Markel Zubia", "gender": "Male", "institution": "Ruhr-Universit\u00e4t Bochum", "country": "DE", "position": "PhD student" }, { "name": "Nils Jansen", "gender": "Male", "institution": "Ruhr-Universit\u00e4t Bochum", "country": "DE", "position": "Full Professor" }, { "name": "Thiago D. Sim\u00e3o", "gender": "unknown", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" } ], "n_formula": 19, "n_formula_1": 18, "n_ref_uni": 39, "n_ref": 64, "n_ref_all": 83, "n_fig": 9, "n_tab": 3, "L_tab": 1382, "n_element_tab": 28, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2109, "formula_len_all_1": 1373, "len_all": 134271, "len_all_1": 74515, "len_abs": 1169, "len_title": 115, "len_sents": 33988, "len_sents_1": 28869, "n_sents": 260, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1180, "title": "Robust Transfer of Safety-Constrained Reinforcement Learning Agents", "abs": "Reinforcement learning (RL) often relies on trial and error, which may cause undesirable outcomes. As a result, standard RL is inappropriate for safety-critical applications. To address this issue, one may train a safe agent in a controlled environment (where safety violations are allowed) and then transfer it to the real world (where safety violations may have disastrous consequences). Prior work has made this transfer safe as long as the new environment preserves the safety-related dynamics. However, in most practical applications, differences or shifts in dynamics between the two environments are inevitable, potentially leading to safety violations after the transfer. This work aims to guarantee safety even when the new environment has different (safety-related) dynamics. In other words, we aim to make the process of safe transfer robust. Our methodology (1) robustifies an agent in the controlled environment and (2) provably provides---under mild assumption---a safe transfer to new environments. The empirical evaluation shows that this method yields policies that are robust against changes in dynamics, demonstrating safety after transfer to a new environment.", "keywords": [ "Reinforcement Learning", "Safe Transfer", "Adversarial Training", "Robustness" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "rvOpON15JJ", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Chong Zhang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Fengqi Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Jingyu Gong", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Ke Fan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Lizhuang Ma", "gender": "Male", "institution": "Dept. of Computer Sci. & Eng., Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Qianyu Zhou", "gender": "Male", "institution": "Shanghai Jiao Tong University,", "country": "CN", "position": "PhD student" }, { "name": "Xin Tan", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "zhizhong zhang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuan Xie", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 21, "n_ref_uni": 29, "n_ref": 80, "n_ref_all": 98, "n_fig": 15, "n_tab": 4, "L_tab": 703, "n_element_tab": 99, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 273, "n_element_tab_1": 55, "formula_len_all": 1005, "formula_len_all_1": 1009, "len_all": 151125, "len_all_1": 71245, "len_abs": 1646, "len_title": 115, "len_sents": 40164, "len_sents_1": 32661, "n_sents": 321, "n_sents_1": 271, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1654, "title": "Diffusion Implicit Policy for Unpaired Scene-aware Motion Synthesis", "abs": "Human motion generation is a long-standing problem, and scene-aware motion synthesis has been widely researched recently due to its numerous applications. Prevailing methods rely heavily on paired motion-scene data whose quantity is limited. Meanwhile, it is difficult to generalize to diverse scenes when trained only on a few specific ones. Thus, we propose a unified framework, termed Diffusion Implicit Policy (DIP), for scene-aware motion synthesis, where paired motion-scene data are no longer necessary. In this framework, we disentangle human-scene interaction from motion synthesis during training and then introduce an interaction-based implicit policy into motion diffusion during inference. Synthesized motion can be derived through iterative diffusion denoising and implicit policy optimization, thus motion naturalness and interaction plausibility can be maintained simultaneously. The proposed implicit policy optimizes the intermediate noised motion in a GAN Inversion manner to maintain motion continuity and control keyframe poses though the ControlNet branch and motion inpainting. For long-term motion synthesis, we introduce motion blending for stable transitions between multiple sub-tasks, where motions are fused in rotation power space and translation linear space. The proposed method is evaluated on synthesized scenes with ShapeNet furniture, and real scenes from PROX and Replica. Results show that our framework presents better motion naturalness and interaction plausibility than cutting-edge methods. This also indicates the feasibility of utilizing the DIP for motion synthesis in more general tasks and versatile scenes.", "keywords": [ "Motion Synthesis", "Human Animation" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "rv9c1BqY0L", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Narimawa Watanabe", "gender": "Male", "institution": "Woven by Toyota", "country": "JP", "position": "Researcher" }, { "name": "Nicolas Bougie", "gender": "Male", "institution": "Woven by Toyota", "country": "JP", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 8, "n_ref_uni": 34, "n_ref": 70, "n_ref_all": 98, "n_fig": 12, "n_tab": 15, "L_tab": 5803, "n_element_tab": 399, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 568, "n_element_tab_1": 23, "formula_len_all": 3216, "formula_len_all_1": 456, "len_all": 270162, "len_all_1": 70292, "len_abs": 1563, "len_title": 131, "len_sents": 95277, "len_sents_1": 34668, "n_sents": 735, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1571, "title": "SimUSER: When Language Models Pretend to Be Believable Users in Recommender Systems", "abs": "Recommender systems play a central role in numerous real-life applications, yet evaluating their performance remains a significant challenge due to the gap between offline metrics and online behaviors. We introduce SimUSER, an agent framework that serves as believable and cost-effective human proxies for the evaluation of recommender systems. Leveraging the inductive bias of foundation models, SimUSER emulates synthetic users by first identifying self-consistent personas from historical data, enriching user profiles with unique backgrounds and personalities. Then, central to this evaluation are users equipped with persona, memory, perception, and brain modules, engaging in interactions with the recommender system. Specifically, the memory module consists of an episodic memory to log interactions and preferences, and a knowledge-graph memory that captures relationships between users and items. The perception module enables visual-driven reasoning, while the brain module translates retrieved information into actionable plans. We demonstrate through ablation studies that the components of our agent architecture contribute to the believability of user behavior. Across a set of recommendation domains, SimUSER exhibits closer alignment with genuine humans than prior state-of-the-art, both at micro and macro levels. Additionally, we conduct insightful experiments to explore the effects of thumbnails on click rates, the exposure effect, and the impact of reviews on user engagement. The source code is released at https://github.com/SimUSER-paper/SimUSER.", "keywords": [ "Recommender systems", "Recommender systems evaluation", "agent-based language models" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "ruv3HdK6he", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Haobo Fu", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "QIANG FU", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Shuai Li", "gender": "Female", "institution": "John Hopcroft Center, Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Stefano V. Albrecht", "gender": "unknown", "institution": "University of Edinburgh", "country": "", "position": "Assistant Professor" }, { "name": "Xu Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 20, "n_formula_1": 6, "n_ref_uni": 62, "n_ref": 108, "n_ref_all": 130, "n_fig": 8, "n_tab": 6, "L_tab": 2467, "n_element_tab": 238, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 112, "n_element_tab_1": 12, "formula_len_all": 1976, "formula_len_all_1": 367, "len_all": 169628, "len_all_1": 65582, "len_abs": 1133, "len_title": 88, "len_sents": 49458, "len_sents_1": 34010, "n_sents": 310, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1138, "title": "Online-to-Offline RL for Agent Alignment", "abs": "Reinforcement learning (RL) has shown remarkable success in training agents to achieve high-performing policies, particularly in domains like Game AI where simulation environments enable efficient interactions. However, despite their success in maximizing these returns, such online-trained policies often fail to align with human preferences concerning actions, styles, and values. The challenge lies in efficiently adapting these online-trained policies to align with human preferences, given the scarcity and high cost of collecting human behavior data. In this work, we formalize the problem as *online-to-offline* RL and propose ALIGNment of Game AI to Preferences (ALIGN-GAP), an innovative approach for alignment of well-trained game agents to human preferences. Our method features a carefully designed reward model that encodes human preferences from limited offline data and incorporates curriculum-based preference learning to align RL agents with targeted human values. Experiments across diverse environments and preference types demonstrate the performance of ALIGN-GAP, achieving effective alignment with human preferences.", "keywords": [ "reinforcement learning", "agent alignment" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "rto6aU453A", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Hengrui Cai", "gender": "Female", "institution": "University of California, Irvine", "country": "US", "position": "Assistant Professor" }, { "name": "Wenbo Zhang", "gender": "unknown", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 19, "n_ref": 43, "n_ref_all": 57, "n_fig": 7, "n_tab": 2, "L_tab": 3010, "n_element_tab": 285, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 994, "formula_len_all_1": 1050, "len_all": 146480, "len_all_1": 67826, "len_abs": 1077, "len_title": 116, "len_sents": 44823, "len_sents_1": 34256, "n_sents": 335, "n_sents_1": 235, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1092, "title": "On High-Dimensional Action Selection for Deep Reinforcement Learning", "abs": "With recent advances in deep reinforcement learning (RL), **high-dimensional action selection** has become an important yet challenging problem in many real applications, especially in unknown and complex environments. Existing works often require a sophisticated prior design to eliminate redundancy in the action space, relying heavily on domain expert experience or involving high computational complexity, which limits their generalizability across different RL tasks. In this paper, we address these challenges by proposing a general data-driven action selection approach with model-free and computational-friendly properties. Our method not only **selects minimal sufficient actions** but also **controls the false discovery rate** via knockoff sampling. More importantly, we seamlessly integrate the action selection into deep RL methods during online training. Empirical experiments validate the established theoretical guarantees, demonstrating that our method surpasses various alternative techniques in terms of both performances in variable selection and overall achieved rewards.", "keywords": [ "Deep Reinforcement Learning", "False Discovery Control", "High-Dimensional Action Space", "Online Learning", "Variable Selection" ], "rating_list": [ 3, 3, 8, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 1, 3, 3 ], "confidence_list": [ 4, 2, 3, 4 ] }, { "paper_id": "rtUjj03qZv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Dan Guo", "gender": "Female", "institution": "Hefei University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Hao Fei", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Lizi Liao", "gender": "Female", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Roger Zimmermann", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Wei Ji", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Xinze Lan", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "MS student" }, { "name": "Xun Yang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "YOU QIN", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 20, "n_ref_uni": 26, "n_ref": 71, "n_ref_all": 94, "n_fig": 7, "n_tab": 7, "L_tab": 2741, "n_element_tab": 233, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 3044, "n_element_tab_1": 238, "formula_len_all": 820, "formula_len_all_1": 820, "len_all": 138214, "len_all_1": 68055, "len_abs": 1417, "len_title": 115, "len_sents": 36582, "len_sents_1": 27993, "n_sents": 288, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1524, "title": "Grounding is All You Need? Dual Temporal Grounding for Video Dialog", "abs": "In the realm of video dialog response generation, the understanding of video content and the temporal nuances of conversation history are paramount. While a segment of current research leans heavily on large-scale pretrained visual-language models and often overlooks temporal dynamics, another delves deep into spatial-temporal relationships within videos but demands intricate object trajectory pre-extractions and sidelines dialog temporal dynamics. \nThis paper introduces the Dual Temporal Grounding-enhanced Video Dialog model (DTGVD), strategically designed to merge the strengths of both dominant approaches.\nIt emphasizes dual temporal relationships by predicting dialog turn-specific temporal regions, filtering video content accordingly, and grounding responses in both video and dialog contexts. \nOne standout feature of DTGVD is its heightened attention to chronological interplay. By recognizing and acting upon the dependencies between different dialog turns, it captures more nuanced conversational dynamics. \nTo further bolster the alignment between video and dialog temporal dynamics, we've implemented a list-wise contrastive learning strategy. Within this framework, accurately grounded turn-clip pairings are designated as positive samples, while less precise pairings are categorized as negative. This refined classification is then funneled into our holistic end-to-end response generation mechanism. Evaluations using AVSD@DSTC-7 and AVSD@DSTC-8 datasets underscore the superiority of our methodology.", "keywords": [ "video dialog; multi-modal understanding; video grounding" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "rss4mLJDpT", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Erlei Zhang", "gender": "Male", "institution": "Northwest A&F University", "country": "CN", "position": "Associate Professor" }, { "name": "He Zong", "gender": "unknown", "institution": "Northwest A&F University", "country": "CN", "position": "MS student" }, { "name": "Hongming Zhang", "gender": "Male", "institution": "Northwest A&F University", "country": "CN", "position": "Full Professor" }, { "name": "Jinchang Ren", "gender": "Male", "institution": "The Robert Gordon University", "country": "GB", "position": "Full Professor" }, { "name": "Xinyu Li", "gender": "Female", "institution": "Northwest A&F University", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 28, "n_ref": 43, "n_ref_all": 58, "n_fig": 6, "n_tab": 3, "L_tab": 1308, "n_element_tab": 236, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1469, "n_element_tab_1": 241, "formula_len_all": 895, "formula_len_all_1": 1031, "len_all": 88237, "len_all_1": 68866, "len_abs": 1534, "len_title": 160, "len_sents": 22785, "len_sents_1": 22105, "n_sents": 171, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1540, "title": "SBGC: Bidirectional Graph Comparison-Based Self-Supervised Network for Change Detection in Heterogeneous Images", "abs": "Change detection (CD) in heterogeneous images is a hot but highly challenging topic in the field of remote sensing. However, the significant imaging differences and varying visual appearances of heterogeneous images complicate the accurate detection of changes occurring on the land surface through direct comparison. To overcome this challenge, this paper proposes a self-supervised network based on bidirectional graph comparison (SBGC) for unsupervised heterogeneous CD, which exploits modality-independent structural relationships. First, pseudo-Siamese networks are established to extract discriminative and robust features from bi-temporal heterogeneous images based on self-supervised contrastive learning. Then, these learned features are utilized to construct graph structures that represent structural relationships. Second, we introduce bidirectional graph comparison to fully exploit the graph structures for exploring comprehensive change information. Specifically, we map the graph structures to their opposite image modality and perform a bidirectional comparison between the original and mapped graph structures to generate a difference image. Finally, the change map is obtained by applying the Otsu segmentation algorithm to the difference image. Experimental results on three public heterogeneous datasets with different modality combinations show that the proposed method achieves superior performance compared to seven state-of-the-art methods, achieving the best performance with an average overall accuracy of 96.69%.", "keywords": [ "Remote sensing", "heterogeneous change detection", "contrastive learning", "bidirectional comparison" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "rslH6DI73J", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Kartik Chandra", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Ping-Chun Hsieh", "gender": "Male", "institution": "National Chiao Tung University", "country": "", "position": "Assistant Professor" }, { "name": "Tzu-Mao Li", "gender": "Not Specified", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Yu-Shuen Wang", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" }, { "name": "Yun-Hsuan Lien", "gender": "Female", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 27, "n_ref_uni": 41, "n_ref": 60, "n_ref_all": 86, "n_fig": 5, "n_tab": 5, "L_tab": 1674, "n_element_tab": 174, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 171, "n_element_tab_1": 23, "formula_len_all": 2202, "formula_len_all_1": 2115, "len_all": 118931, "len_all_1": 55475, "len_abs": 1171, "len_title": 145, "len_sents": 31867, "len_sents_1": 23618, "n_sents": 235, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1174, "title": "Learn out of the box: optimizing both diversity and performance in Offline Reinforcement Learning", "abs": "In offline reinforcement learning, most existing methods have focused primarily on optimizing performance, often neglecting the promotion of diverse behaviors. While some approaches generate diverse behaviors from well-constructed, heterogeneous datasets, their effectiveness is significantly reduced when applied to less diverse data. To address this, we introduce a novel intrinsic reward mechanism that encourages behavioral diversity, irrespective of the dataset's heterogeneity. By maximizing the mutual information between actions and policies under each state, our approach enables agents to learn a variety of behaviors, including those not explicitly represented in the data. Although performing out-of-distribution actions can lead to risky outcomes, we mitigate this risk by incorporating the ensemble-diversified actor-critic (EDAC) method to estimate Q-value uncertainty, preventing agents from adopting suboptimal behaviors. Through experiments using the D4RL benchmarks on MuJoCo tasks, we demonstrate that our method achieves behavioral diversity while maintaining performance across environments constructed from both heterogeneous and homogeneous datasets.", "keywords": [ "Offline Reinforcement Learning", "Diversity and Performance", "Homogeneous Dataset" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "rsZwwjYHuD", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Fushuo Huo", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Haozhao Wang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Postdoc" }, { "name": "Peilin Zhao", "gender": "unknown", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Wenchao Xu", "gender": "Not Specified", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Researcher" }, { "name": "Zhicheng Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Zhong Zhang", "gender": "unknown", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 56, "n_ref": 223, "n_ref_all": 277, "n_fig": 12, "n_tab": 17, "L_tab": 5553, "n_element_tab": 697, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 4874, "n_element_tab_1": 424, "formula_len_all": 693, "formula_len_all_1": 568, "len_all": 220676, "len_all_1": 68986, "len_abs": 1559, "len_title": 136, "len_sents": 69850, "len_sents_1": 27522, "n_sents": 591, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1678, "title": "Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models", "abs": "Hallucination remains a significant challenge in Large Vision-Language Models (LVLMs). To alleviate this issue, some methods, known as contrastive decoding, induce hallucinations by manually disturbing the raw vision or instruction inputs and then mitigate them by contrasting the outputs of the original and disturbed LVLMs. However, these holistic input disturbances sometimes induce potential noise and also double the inference cost. To tackle these issues, we propose a simple yet effective method named $\\textit{Self-Introspective Decoding}$ (SID). Our empirical investigations reveal that pre-trained LVLMs can introspectively assess the importance of vision tokens based on preceding vision and text (both instruction and generated) tokens. Leveraging this insight, we develop the Context and Text-aware Token Selection (CT$^2$S) strategy, which preserves only the least important vision tokens after the early decoder layers, thereby adaptively amplify vision-and-text association hallucinations during auto-regressive decoding. This strategy ensures that multimodal knowledge absorbed in the early decoder layers induces multimodal contextual rather than aimless hallucinations, and significantly reduces computation burdens. Subsequently, the original token logits subtract the amplified fine-grained hallucinations, effectively alleviating hallucinations without compromising the LVLMs' general ability. Extensive experiments illustrate SID generates less-hallucination and higher-quality texts across various metrics, without much additional computation cost. Codes are in the Supplementary Material and also available at https://anonymous.4open.science/r/SID-1795.", "keywords": [ "Hallucination Alleviation", "Large Vision-Language Models", "Decoding Strategy", "Trustworthy AI" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "rsMajBqYrB", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Changlun Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Nan Tang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "Associate Professor" }, { "name": "Yunfan Zhang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "", "position": "MS student" }, { "name": "Yuyu Luo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 40, "n_ref": 57, "n_ref_all": 91, "n_fig": 5, "n_tab": 9, "L_tab": 3201, "n_element_tab": 620, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2713, "n_element_tab_1": 374, "formula_len_all": 678, "formula_len_all_1": 1566, "len_all": 159051, "len_all_1": 70705, "len_abs": 1930, "len_title": 125, "len_sents": 47616, "len_sents_1": 31220, "n_sents": 349, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1946, "title": "SketchFill: Sketch-Guided Code Generation for Imputing Derived Missing Values", "abs": "Missing value is a critical issue in data science, significantly impacting the reliability of analyses and predictions. Missing value imputation (MVI) is a longstanding problem because it highly relies on domain knowledge. Large language models (LLMs) have emerged as a promising tool for data cleaning, including MVI for tabular data, offering advanced capabilities for understanding and generating content. However, despite their promise, existing LLM techniques such as in-context learning and Chain-of-Thought (CoT) often fall short in guiding LLMs to perform complex reasoning for MVI, particularly when imputing derived missing values, which require mathematical formulas and data relationships across rows and columns. This gap underscores the need for further advancements in LLM methodologies to enhance their reasoning capabilities for more reliable imputation outcomes. To fill this gap, we propose SketchFill, a novel sketch-based method to guide LLMs in generating accurate formulas to impute missing numerical values. SketchFill first utilizes a general user-provided Meta-Sketch to generate a Domain-Sketch tailored to the context of the input dirty table. Subsequently, it fills this Domain-Sketch with formulas and outputs Python code, effectively bridging the gap between high-level abstractions and executable solutions. Additionally, SketchFill incorporates a Reflector component to verify the generated code. This Reflector assesses the accuracy and appropriateness of the outputs and iteratively refines the Domain-Sketch, ensuring that the imputation aligns closely with the underlying data patterns and relationships. Our experimental results demonstrate that SketchFill significantly outperforms state-of-the-art methods, achieving 56.2% higher accuracy than CoT-based methods and 78.8% higher accuracy than MetaGPT. This sets a new standard for automated data cleaning and advances the field of MVI for numerical values.", "keywords": [ "Large Language Model;Missing Value Impuation;Data Preprocessing;Code Generation" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "rsJaUHCZIv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Huibin Tan", "gender": "Female", "institution": "National University of Defense Technology", "country": "", "position": "Lecturer" }, { "name": "Long Lan", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Mengzhu Wang", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Mingyu Cao", "gender": "Female", "institution": "National University of Defense Technology", "country": "", "position": "MS student" }, { "name": "Wanrong Huang", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Assistant Research Fellow" }, { "name": "Wenjing Yang", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Xueqiong Li", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "Lecturer" }, { "name": "Houcheng Su", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 29, "n_ref": 57, "n_ref_all": 65, "n_fig": 6, "n_tab": 3, "L_tab": 651, "n_element_tab": 143, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 423, "n_element_tab_1": 92, "formula_len_all": 420, "formula_len_all_1": 419, "len_all": 126780, "len_all_1": 57515, "len_abs": 1438, "len_title": 151, "len_sents": 32564, "len_sents_1": 28339, "n_sents": 234, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1419, "title": "Enhanced Semantic Alignment in Transformer Tracking via Position Learning and Force-Directed Attention", "abs": "In the field of visual object tracking, one-stream pipelines have become the mainstream framework due to its efficient integration of feature extraction and relationship modeling. \nHowever, existing methods still face the issue of semantic misalignment: \nfirstly, the interaction of positional encoding between the two branches leads to a misalignment between feature semantics and position encoding; \nsecondly, traditional attention mechanisms fail to distinguish between semantic attraction and repulsion among features, resulting in semantic misalignment when the model processes these features. \nTo address these issues, we propose an Enhanced Semantic Alignment Transformer Tracker (ESAT) with position encode learning and force-directed attention mechanism. \nBy leveraging positional encoding loss, ESAT separately learns the absolute positional encodings of the target and search branches, distinguishing the locations of various tokens and their positive or negative relationships, thereby enhancing the semantic consistency between position and features.\nAdditionally, it incorporates a repulsion-attraction mechanism applied to the self-attention module, simulating dynamic interactions between nodes to improve feature discrimination. \nExtensive experiments on multiple public tracking datasets show that our method outperforms many pipelines and achieves superior performance on five challenging benchmarks.", "keywords": [ "Transformer Tracking; Single Object Tracking; Semantic Alignment; Self-supervised Position Loss; Force-Directed Attention" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "rsGPrJDIhh", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Aram Davtyan", "gender": "Male", "institution": "Universit\u00e4t Bern", "country": "CH", "position": "PhD student" }, { "name": "Leello Tadesse Dadi", "gender": "Male", "institution": "EPFL", "country": "CH", "position": "PhD student" }, { "name": "Paolo Favaro", "gender": "Male", "institution": "Institute f\u00fcr Informatik, University of Bern", "country": "CH", "position": "Full Professor" }, { "name": "Volkan Cevher", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" } ], "n_formula": 28, "n_formula_1": 14, "n_ref_uni": 35, "n_ref": 98, "n_ref_all": 148, "n_fig": 15, "n_tab": 3, "L_tab": 2867, "n_element_tab": 331, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1798, "n_element_tab_1": 167, "formula_len_all": 2092, "formula_len_all_1": 612, "len_all": 140781, "len_all_1": 66972, "len_abs": 1028, "len_title": 129, "len_sents": 44016, "len_sents_1": 30036, "n_sents": 352, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1033, "title": "Faster Inference of Flow-Based Generative Models via Improved Data-Noise Coupling", "abs": "Conditional Flow Matching (CFM), a simulation-free method for training continuous normalizing flows, provides an efficient alternative to diffusion models for key tasks like image and video generation. The performance of CFM in solving these tasks depends on the way data is coupled with noise. A recent approach uses minibatch optimal transport (OT) to reassign noise-data pairs in each training step to streamline sampling trajectories and thus accelerate inference. However, its optimization is restricted to individual minibatches, limiting its effectiveness on large datasets. To address this shortcoming, we introduce LOOM-CFM (Looking Out Of Minibatch-CFM), a novel method to extend the scope of minibatch OT by preserving and optimizing these assignments across minibatches over training time. Our approach demonstrates consistent improvements in the sampling speed-quality trade-off across multiple datasets. LOOM-CFM also enhances distillation initialization and supports high-resolution synthesis in latent space training.", "keywords": [ "generative models", "flow matching" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "rrn6XXB3u0", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Akhilan Boopathy", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Amber Roy Chowdhury", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Gil Wolff", "gender": "unknown", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Researcher" }, { "name": "Zhiyan Yi", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 8, "n_ref_uni": 28, "n_ref": 35, "n_ref_all": 46, "n_fig": 3, "n_tab": 3, "L_tab": 1068, "n_element_tab": 155, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 92, "n_element_tab_1": 10, "formula_len_all": 1821, "formula_len_all_1": 892, "len_all": 104437, "len_all_1": 61284, "len_abs": 1139, "len_title": 83, "len_sents": 32265, "len_sents_1": 26172, "n_sents": 239, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1144, "title": "Fast Few-Shot Graph Flow Prediction", "abs": "Accurate prediction of traffic flow is crucial for optimizing transportation networks, mitigating congestion, and improving urban planning. However, existing approaches like graph neural networks (GNNs) and traffic simulations face challenges in predicting flow for unseen road networks without historical data. Without abundant training data, GNNs often generalize poorly to new graphs, while simulations can be computationally infeasible for large-scale networks. This paper tackles the problem of few-shot traffic flow prediction in unseen road networks. We propose a novel traffic simulation algorithm that efficiently predicts flow based on node and edge attributes. Through theoretical analysis, we demonstrate our approach closely approximates true flow with asymptotically optimal runtime complexity. Experiments on real-world road networks show our simulation algorithm outperforms GNNs for predicting traffic in unseen cities after training on only three cities. While motivated by traffic prediction in road networks, we expect our contributions to have broader applicability to general graph flow prediction problems across domains.", "keywords": [ "graph flow", "flow prediction", "graph neural network", "few-shot learning", "traffic prediction", "neural network" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 1, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "rrdNQZRHEm", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Egor Bondarev", "gender": "unknown", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Associate Professor" }, { "name": "Francois Bremond", "gender": "Male", "institution": "inria", "country": "FR", "position": "Researcher" }, { "name": "Giacomo D'Amicantonio", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "PhD student" }, { "name": "Gianpiero Francesca", "gender": "Male", "institution": "Toyota Motor Europe", "country": "BE", "position": "Researcher" }, { "name": "Lorenzo Garattoni", "gender": "Male", "institution": "Toyota Motor Europe", "country": "BE", "position": "Researcher" }, { "name": "Quan Kong", "gender": "Male", "institution": "Woven by Toyota", "country": "JP", "position": "Researcher" }, { "name": "Snehashis Majhi", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 22, "n_ref": 45, "n_ref_all": 58, "n_fig": 6, "n_tab": 6, "L_tab": 1818, "n_element_tab": 241, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 124, "n_element_tab_1": 12, "formula_len_all": 307, "formula_len_all_1": 371, "len_all": 119815, "len_all_1": 56224, "len_abs": 2338, "len_title": 165, "len_sents": 31574, "len_sents_1": 28054, "n_sents": 208, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 116, "L_abs": 2362, "title": "Mixture of Experts Guided by Gaussian Splatters Matters: A new Approach to Weakly-Supervised Video Anomaly Detection", "abs": "Video Anomaly Detection (VAD) has proved to be a challenging task due to the in-\nherent variability of anomalous events and the scarcity of data available. Under the\ncommon Weakly-Supervised VAD (WSVAD) paradigm, only a video-level label\nis available during training, while the predictions are carried out at the frame-level.\nDespite decent progress on simple anomalous events (such as explosions), more\ncomplex real-world anomalies (such as shoplifting) remain challenging. There\nare two main reasons for this: (I) current state-of-the-art models do not address\nthe diversity between anomalies during training and process diverse categories\nof anomalies with a shared model, thereby ignoring the category-specific key at-\ntributes; and (II) the lack of precise temporal information (i.e., weak-supervision)\nlimits the ability to learn how to capture complex abnormal attributes that can\nblend with normal events, effectively allowing to use only the most abnormal snip-\npets of an anomaly. We hypothesize that these issues can be addressed by sharing\nthe task between multiple expert models that would increase the possibility of cor-\nrectly encoding the singular characteristics of different anomalies. Furthermore,\nmultiple Gaussian kernels can guide the experts towards a more comprehensive\nand complete representation of anomalous events, ensuring that each expert pre-\ncisely distinguishes between normal and abnormal events at the frame-level. To\nthis end, we introduce Gaussian Splatting-guided Mixture of Experts (GS-MoE),\na novel approach that leverages a set of experts trained with a temporal Gaussian\nsplatting loss on specific classes of anomalous events and integrates their predic-\ntions via a mixture of expert models to capture complex relationships between\ndifferent anomalous patterns. The introduction of temporal Gaussian splatting\nloss allows the model to leverage temporal consistency in weakly-labeled data,\nenabling more robust identification of subtle anomalies over time. The novel loss\nfunction, designed to enhance weak supervision, further improves model perfor-\nmance by guiding expert networks to focus on segments of data with a higher like-\nlihood of containing anomalies. Experimental results on the UCF-Crime and XD-\nViolence datasets demonstrate that our framework achieves SOTA performance,\nscoring 91.58% AUC on UCF-Crime.", "keywords": [ "Video Anomaly Detection", "Weakly-supervised Learning" ], "rating_list": [ 3, 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 5, 5, 4, 4, 4 ] }, { "paper_id": "rqBc4WnvUP", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Ben Glocker", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Full Professor" }, { "name": "Francesca Toni", "gender": "Female", "institution": "Imperial College London", "country": "", "position": "Full Professor" }, { "name": "Avinash G. Kori", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 69, "n_formula_1": 29, "n_ref_uni": 45, "n_ref": 112, "n_ref_all": 140, "n_fig": 9, "n_tab": 3, "L_tab": 2306, "n_element_tab": 260, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 1549, "n_element_tab_1": 91, "formula_len_all": 7656, "formula_len_all_1": 2809, "len_all": 200235, "len_all_1": 73842, "len_abs": 894, "len_title": 116, "len_sents": 57277, "len_sents_1": 32574, "n_sents": 413, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 901, "title": "Multi-view Object-Centric Learning with Identifiable Representations", "abs": "Modular object-centric representations are key to unlocking human-like reasoning capabilities. However, addressing challenges such as object occlusions to obtain meaningful object-level representations presents both theoretical and practical difficulties. We introduce a novel multi-view probabilistic approach that aggregates view-specific slots to capture *invariant content* information while simultaneously learning disentangled global *viewpoint-level* information. Our model resolves spatial ambiguities and provides theoretical guarantees for learning identifiable representations, setting it apart from prior work focusing on single-view settings and lacking theoretical foundations. Along with our identifiability analysis, we provide extensive empirical validation with promising results on both benchmark and proposed large-scale datasets carefully designed to evaluate multi-view methods.", "keywords": [ "multi-view object-centric learning", "identifiability", "probabilistic-slot-attention" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "rpwGUtTeA5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Boyuan Pan", "gender": "Male", "institution": "Xiaohongshu", "country": "", "position": "Researcher" }, { "name": "Chuyi Tan", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Jiayi Shi", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Kan Li", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Peiwen Yuan", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Shaoxiong Feng", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Xinglin Wang", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yao Hu", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Principal Researcher" }, { "name": "Yiwei Li", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Yueqi Zhang", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 13, "n_ref_uni": 16, "n_ref": 39, "n_ref_all": 63, "n_fig": 10, "n_tab": 3, "L_tab": 260, "n_element_tab": 34, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 272, "n_element_tab_1": 39, "formula_len_all": 1753, "formula_len_all_1": 768, "len_all": 130673, "len_all_1": 60124, "len_abs": 1769, "len_title": 155, "len_sents": 37865, "len_sents_1": 25866, "n_sents": 297, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 107, "L_abs": 1264, "title": "UniCBE: An Uniformity-driven Comparing Based Evaluation Framework with Unified Multi-Objective Optimization", "abs": "Human preference plays a significant role in measuring large language models and guiding them to align with human values. Unfortunately, current comparing-based evaluation (CBE) methods typically focus on a single optimization objective, failing to effectively utilize scarce yet valuable preference signals. To address this, we delve into key factors that can enhance the accuracy, convergence, and scalability of CBE: suppressing sampling bias, balancing descending process of uncertainty, and mitigating updating uncertainty.\nFollowing the derived guidelines, we propose UniCBE, a unified uniformity-driven CBE framework which simultaneously optimize these core objectives by constructing and integrating three decoupled sampling probability matrices, each designed to ensure uniformity in specific aspects. We further ablate the optimal tuple sampling and preference aggregation strategies to achieve efficient CBE.\nOn the AlpacaEval benchmark, UniCBE saves over 17% of evaluation budgets while achieving a Pearson correlation with ground truth exceeding 0.995, demonstrating excellent accuracy and convergence. In scenarios where new models are continuously introduced, UniCBE can even save over 50% of evaluation costs, highlighting its improved scalability.", "keywords": [ "evaluation", "efficient", "scalability", "accuracy", "convergence" ], "rating_list": [ 8, 8, 6, 8, 6 ], "soundness_list": [ 4, 4, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 2, 3 ], "contribution_list": [ 3, 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "rpouyo09V0", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Hojae Han", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Rajhans Samdani", "gender": "Male", "institution": "Spoke", "country": "", "position": "Chief Scientist" }, { "name": "Yuxiong He", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "seung-won hwang", "gender": "unknown", "institution": "Seoul National University", "country": "", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 3, "n_ref_uni": 12, "n_ref": 29, "n_ref_all": 43, "n_fig": 3, "n_tab": 12, "L_tab": 7899, "n_element_tab": 1461, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2678, "n_element_tab_1": 514, "formula_len_all": 1533, "formula_len_all_1": 106, "len_all": 157432, "len_all_1": 58880, "len_abs": 1782, "len_title": 144, "len_sents": 40785, "len_sents_1": 21117, "n_sents": 270, "n_sents_1": 141, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1798, "title": "ConvCodeWorld: Benchmarking Conversational Code Generation in Reproducible Feedback Environments", "abs": "Large language models (LLMs) have proven invaluable for code generation, particularly in interactive settings. However, existing code generation benchmarks fail to capture the diverse feedback encountered in multi-turn interactions, limiting our ability to evaluate LLMs in these contexts. To address this gap, we present a set of novel benchmarks that explicitly model the quality of feedback provided to code generation LLMs. Our contributions are threefold: **First**, we introduce CONVCODEWORLD, a novel and reproducible environment for benchmarking interactive code generation. CONVCODEWORLD simulates 9 distinct interactive code generation scenarios while systematically combining three types of feedback: (a) compilation feedback; (b) execution feedback with varying test coverage; (c) verbal feedback generated by GPT-4o with different levels of expertise. **Second**, we introduce CONVCODEBENCH, a fast, static version of benchmark that uses pre-generated feedback logs, eliminating the need for costly dynamic verbal feedback generation while maintaining\nstrong Spearman\u2019s rank correlations (0.82 to 0.99) with CONVCODEWORLD. **Third**, extensive evaluations of both closed-source and open-source LLMs on CONVCODEWORLD reveal key insights: (a) LLM performance varies significantly based on the feedback provided; (b) Weaker LLMs, with sufficient feedback, can outperform single-turn results of state-of-the-art LLMs without feedback; (c) Training on a specific feedback combination can limit an LLM\u2019s ability to utilize unseen combinations; (d) LLMs solve problems in fewer turns (high MRR) may not solve as many problems overall (high Recall), and vice versa. All implementations and benchmarks will be made publicly available at https://huggingface.co/spaces/ConvCodeWorld/ConvCodeWorld", "keywords": [ "Large language models", "Multi-turn code generation", "Benchmark" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 4, 3, 1 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "rpbzBXdo4x", "primary_area": "applications to neuroscience & cognitive science", "n_author": 6, "author_info_list": [ { "name": "Addison J. Wu", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Undergrad student" }, { "name": "Ilia Sucholutsky", "gender": "Male", "institution": "New York University", "country": "US", "position": "Faculty Fellow/Assistant Professor" }, { "name": "Jiayi Geng", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "MS student" }, { "name": "Ryan Liu", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Tania Lombrozo", "gender": "Female", "institution": "Princeton University", "country": "", "position": "Full Professor" }, { "name": "Thomas L. Griffiths", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 53, "n_ref": 90, "n_ref_all": 124, "n_fig": 6, "n_tab": 10, "L_tab": 2214, "n_element_tab": 269, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1257, "n_element_tab_1": 169, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 205091, "len_all_1": 67503, "len_abs": 1748, "len_title": 156, "len_sents": 71863, "len_sents_1": 33535, "n_sents": 526, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1766, "title": "Mind Your Step (by Step): Chain-of-Thought can Reduce Performance on Tasks where Thinking Makes Humans Worse", "abs": "Chain-of-thought (CoT) prompting has become a widely used strategy for working with large language and multimodal models. While CoT has been shown to improve performance across many tasks, determining the settings in which it is effective remains an ongoing effort. In particular, it is still an open question in what settings CoT systematically reduces model performance. In this paper, we seek to identify the characteristics of tasks where CoT reduces performance by drawing inspiration from cognitive psychology, looking at cases where (i) verbal thinking or deliberation hurts performance in humans, and (ii) the constraints governing human performance generalize to language models. Three such cases are implicit statistical learning, visual recognition, and classifying with patterns containing exceptions. In extensive experiments across all three settings, we find that a diverse collection of state-of-the-art models exhibit significant drop-offs in performance (e.g., up to 36.3\\% absolute accuracy for GPT-o1 compared to GPT-4o) when using CoT compared to zero-shot counterparts. We also identify three tasks that satisfy condition (i) but not (ii), and find that while verbal thinking reduces human performance in these tasks, CoT retains or increases model performance. Overall, our results show that even though there is not an exact parallel between the cognitive processes of models and those of humans, considering cases where thinking has negative consequences for human performance can help us identify settings where it has negative consequences for models. By connecting the literature on human deliberation with evaluation of CoT, we offer a new tool that can be used in understanding the impact of prompt choices and inference-time reasoning.", "keywords": [ "chain of thought", "psychology", "overthinking" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 4, 1 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "rpR9fDZw3D", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Colin Cherry", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Eleftheria Briakou", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Hamid Dadkhahi", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Jun Wang", "gender": "Male", "institution": "Research, Google", "country": "", "position": "Intern" }, { "name": "Rishabh Agarwal", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Adjunct Professor" }, { "name": "Trevor Cohn", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 21, "n_ref": 53, "n_ref_all": 73, "n_fig": 7, "n_tab": 7, "L_tab": 1902, "n_element_tab": 348, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 2731, "n_element_tab_1": 380, "formula_len_all": 148, "formula_len_all_1": 109, "len_all": 107911, "len_all_1": 60677, "len_abs": 893, "len_title": 109, "len_sents": 27980, "len_sents_1": 26169, "n_sents": 199, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 895, "title": "Don\u2019t Throw Away Data: Better Sequence Knowledge Distillation", "abs": "A critical component in knowledge distillation is the means of coupling the teacher and student. The predominant sequence knowledge distillation method involves supervised learning of the student against teacher-decoded outputs, and is exemplified by the current state of the art, which incorporates minimum Bayes risk (MBR) decoding. In this paper we seek to integrate MBR more tightly in distillation training, specifically by using several high scoring MBR translations, rather than a single selected sequence, thus capturing a rich diversity of teacher outputs. Our experiments on English to German and English to Japanese translation show consistent improvements over strong baseline methods for both tasks and with varying model sizes. Additionally, we conduct a detailed analysis focusing on data efficiency and capacity curse aspects to elucidate MBR-n and explore its further potential.", "keywords": [ "Knowledge distillation", "machine translation", "large language model" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "rpEATZvmjr", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Han Yu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Xiaoli Tang", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xiaoxiao Li", "gender": "Unspecified", "institution": "University of British Columbia", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 18, "n_ref": 36, "n_ref_all": 53, "n_fig": 3, "n_tab": 3, "L_tab": 1709, "n_element_tab": 122, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 959, "n_element_tab_1": 36, "formula_len_all": 735, "formula_len_all_1": 696, "len_all": 121888, "len_all_1": 65290, "len_abs": 1610, "len_title": 126, "len_sents": 38153, "len_sents_1": 28648, "n_sents": 274, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1425, "title": "Multi-Session Budget Optimization for Forward Auction-based Federated Learning", "abs": "Auction-based Federated Learning (AFL) has emerged as an important research field in recent years. The prevailing strategies for FL data consumers (DCs) assume that the entire team of the required data owners (DOs) for an FL task must be assembled before training can commence. In practice, a DC can trigger the FL training process multiple times. DOs can thus be gradually recruited over multiple FL model training sessions. Existing bidding strategies for AFL DCs are not designed to handle such scenarios. Therefore, the problem of multi-session AFL remains open. To address this problem, we propose the Multi-session Budget Optimization Strategy for forward Auction-based Federated Learning (MultiBOS-AFL). Based on hierarchical reinforcement learning, MultiBOS-AFL jointly optimizes inter-session budget pacing and intra-session bidding for AFL DCs, with the objective of maximizing the total utility. Extensive experiments on six benchmark datasets show that it significantly outperforms seven state-of-the-art approaches. On average, MultiBOS-AFL achieves 12.28% higher utility, 14.52% more data acquired through auctions for a given budget, and 1.23% higher test accuracy achieved by the resulting FL model compared to the best baseline. To the best of our knowledge, it is the first budget optimization decision support method with budget pacing capability designed for DCs in multi-session forward auction-based FL.", "keywords": [ "Federated Learning", "Auction-based Federated Learning", "Bidding Strategy" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 2, 2 ] }, { "paper_id": "roNSXZpUDN", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Karthik R Narasimhan", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Noah Shinn", "gender": "unknown", "institution": "Northeastern University", "country": "US", "position": "Undergrad student" }, { "name": "Pedram Razavi", "gender": "unknown", "institution": "Sierra", "country": "", "position": "Researcher" }, { "name": "Shunyu Yao", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 5, "n_ref": 26, "n_ref_all": 49, "n_fig": 10, "n_tab": 7, "L_tab": 4237, "n_element_tab": 166, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1856, "n_element_tab_1": 215, "formula_len_all": 103, "formula_len_all_1": 95, "len_all": 245930, "len_all_1": 71176, "len_abs": 1501, "len_title": 123, "len_sents": 105137, "len_sents_1": 33600, "n_sents": 659, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 117, "L_abs": 972, "title": "{$\\tau$}-bench: A Benchmark for \\underline{T}ool-\\underline{A}gent-\\underline{U}ser Interaction in Real-World Domains", "abs": "Existing benchmarks for language agents do not set them up to interact with human users or follow domain-specific rules, both of which are vital to safe and realistic deployment. We propose $\\tau$-bench, a benchmark with two domains (retail and airline) emulating dynamic conversations between a user (simulated by language models) and a customer service agent provided with domain-specific API tools and policy guidelines. We employ a efficient and faithful evaluation process that compares the database state at the end of a conversation with the annotated goal state, and propose a new metric (pass^k) to evaluate the reliability of agent behavior over multiple trials. Our experiments show that even state-of-the-art function calling agents (gpt-4o) succeed on $<50\\%$ of the tasks, and are terribly inconsistent (pass^8 < 25\\% in retail). Our findings point to the need for methods that can improve the ability of agents to act consistently and reliably follow rules.", "keywords": [ "language model", "language agent", "benchmark", "user simulation", "rule following" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "rnTb9dm9zx", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Michelle E. Lu", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" }, { "name": "XiuYu Zhang", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" }, { "name": "Zening Luo", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 23, "n_ref": 45, "n_ref_all": 62, "n_fig": 6, "n_tab": 1, "L_tab": 225, "n_element_tab": 30, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 511, "n_element_tab_1": 111, "formula_len_all": 770, "formula_len_all_1": 677, "len_all": 101741, "len_all_1": 61715, "len_abs": 1665, "len_title": 129, "len_sents": 36475, "len_sents_1": 28322, "n_sents": 264, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1695, "title": "Partially Conditioned Patch Parallelism for Accelerated Diffusion Model Inference", "abs": "Diffusion models have exhibited exciting capabilities in generating images and are also very promising for video creation. However, the inference speed of diffusion models is limited by the slow sampling process, restricting its use cases. The sequential denoising steps required for generating a single sample could take tens or hundreds of iterations and thus have become a significant bottleneck. This limitation is more salient for applications that are interactive in nature or require small latency. To address this challenge, we propose Partially Conditioned Patch Parallelism (PCPP) to accelerate the inference of high-resolution diffusion models. Using the fact that the difference between the images in adjacent diffusion steps is nearly zero, Patch Parallelism (PP) leverages multiple GPUs communicating asynchronously to compute patches of an image in multiple computing devices based on the entire image (all patches) in the previous diffusion step. PCPP develops PP to reduce computation in inference by conditioning only on parts of the neighboring patches in each diffusion step, which also decreases communication among computing devices. As a result, PCPP decreases the communication cost by around $70$% compared to DistriFusion (the state of the art implementation of PP) and achieves $2.36\\sim 8.02\\times$ inference speed-up using $4\\sim 8$ GPUs compared to $2.32\\sim 6.71\\times$ achieved by DistriFusion depending on the computing device configuration and resolution of generation at the cost of a possible decrease in image quality. PCPP demonstrates the potential to strike a favorable trade-off, enabling high-quality image generation with substantially reduced latency.", "keywords": [ "Diffusion model", "image generation", "parallel algorithm" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "rnRBGMNYa2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Hyounghun Kim", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Seokhyun An", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology (UNIST)", "country": "KR", "position": "Undergrad student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 58, "n_ref": 157, "n_ref_all": 210, "n_fig": 8, "n_tab": 13, "L_tab": 8805, "n_element_tab": 1356, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 778, "n_element_tab_1": 137, "formula_len_all": 64, "formula_len_all_1": 64, "len_all": 279735, "len_all_1": 62804, "len_abs": 3195, "len_title": 115, "len_sents": 63611, "len_sents_1": 29506, "n_sents": 565, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1118, "title": "Response Tuning: Aligning Large Language Models without Instruction", "abs": "Instruction tuning\u2014supervised fine-tuning using instruction-response pairs\u2014is a foundational step in transitioning pre-trained Large Language Models (LLMs) into helpful and safe chat assistants. Our hypothesis is that establishing an adequate output distribution can enable such a transition given the capabilities inherent in pre-trained LLMs. To verify this, we propose Response Tuning (RT), which eliminates the instruction-conditioning step in instruction tuning and solely focuses on response distribution supervision. Our experiments demonstrate that RT models, trained only using responses, can effectively respond to a wide range of instructions and exhibit helpfulness comparable to that of their instruction-tuned counterparts. Furthermore, we observe that controlling the training response distribution can substantially improve their user preference or elicit target behaviors such as refusing assistance for unsafe queries. Our findings illuminate the role of establishing an adequate output distribution in alignment, highlighting the potential of the extensive inherent capabilities of pre-trained LLMs.", "keywords": [ "Large Language Models", "Instruction Tuning", "LLM Safety" ], "rating_list": [ 5, 5, 8, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "rnL3OafDdw", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Jian Kang", "gender": "unknown", "institution": "University of Michigan", "country": "US", "position": "Full Professor" }, { "name": "Yuliang Xu", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 15, "n_ref_uni": 30, "n_ref": 49, "n_ref_all": 88, "n_fig": 12, "n_tab": 7, "L_tab": 7335, "n_element_tab": 740, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 2801, "n_element_tab_1": 226, "formula_len_all": 3852, "formula_len_all_1": 1339, "len_all": 174587, "len_all_1": 69459, "len_abs": 1466, "len_title": 128, "len_sents": 53814, "len_sents_1": 29194, "n_sents": 421, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1471, "title": "Bayesian Image Regression with Soft-thresholded Conditional Autoregressive Prior", "abs": "In the analysis of brain functional MRI (fMRI) data using regression models, Bayesian methods are highly valued for their flexibility and ability to quantify uncertainty. However, these methods face computational challenges in high-dimensional settings typical of brain imaging, and the often pre-specified correlation structures may not accurately capture the true spatial relationships within the brain. To address these issues, we develop a general prior specifically designed for regression models with large-scale imaging data. We introduce the Soft-Thresholded Conditional AutoRegressive (ST-CAR) prior, which reduces instability to pre-fixed correlation structures and provides inclusion probabilities to account for the uncertainty in choosing active voxels in the brain. We apply the ST-CAR prior to scalar-on-image (SonI) and image-on-scalar (IonS) regression models\u2014both critical in brain imaging studies\u2014and develop efficient computational algorithms using variational inference (VI) and stochastic subsampling techniques. Simulation studies demonstrate that the ST-CAR prior outperforms existing methods in identifying active brain regions with complex correlation patterns, while our VI algorithms offer superior computational performance. We further validate our approach by applying the ST-CAR to working memory fMRI data from the Adolescent Brain Cognitive Development (ABCD) study, highlighting its effectiveness in practical brain imaging applications.", "keywords": [ "Brain fMRI Image", "Variational Inference", "Scalar-on-Image Regression", "Image-on-Scalar Regression" ], "rating_list": [ 8, 6, 6, 3 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "rnJxelIZrq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Aditya Saibewar", "gender": "unknown", "institution": "Fujitsu Research and Development Center Co. Ltm.", "country": "IN", "position": "Researcher" }, { "name": "Manohar Kaul", "gender": "Male", "institution": "Fujitsu Research and Development Center Co. Ltm.", "country": "IN", "position": "Principal Researcher" }, { "name": "Sadbhavana Babar", "gender": "Female", "institution": "Fujitsu Research and Development Center Co. Ltm.", "country": "IN", "position": "Researcher" } ], "n_formula": 49, "n_formula_1": 13, "n_ref_uni": 28, "n_ref": 50, "n_ref_all": 104, "n_fig": 19, "n_tab": 10, "L_tab": 1525, "n_element_tab": 264, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1767, "n_element_tab_1": 46, "formula_len_all": 2457, "formula_len_all_1": 1007, "len_all": 227408, "len_all_1": 60917, "len_abs": 1440, "len_title": 166, "len_sents": 88086, "len_sents_1": 28077, "n_sents": 690, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 117, "L_abs": 1370, "title": "Beyond Mere Token Analysis: A Hypergraph Metric Space Framework for Defending Against Socially Engineered LLM Attacks", "abs": "Recent jailbreak attempts on Large Language Models (LLMs) have shifted from algorithm-focused to human-like social engineering attacks, with persuasion-based techniques emerging as a particularly effective subset. These attacks evolve rapidly, demonstrate high creativity, and boast superior attack success rates. To combat such threats, we propose a promising approach to enhancing LLM safety by leveraging the underlying geometry of input prompt token embeddings using hypergraphs. This approach allows us to model the differences in information flow between benign and malicious LLM prompts.\n\nIn our approach, each LLM prompt is represented as a metric hypergraph, forming a compact metric space. We then construct a higher-order metric space over these compact metric hypergraphs using the Gromov-Hausdorff distance as a generalized metric. Within this space of metric hypergraph spaces, our safety filter learns to classify between harmful and benign prompts. Our study presents theoretical guarantees on the classifier's generalization error for novel and unseen LLM input prompts. Extensive empirical evaluations demonstrate that our method significantly outperforms both existing state-of-the-art generic defense mechanisms and naive baselines. Notably, our approach also achieves comparable performance to specialized defenses against algorithm-focused attacks.", "keywords": [ "Jailbreak Attack", "LLMs", "LLM Security", "AI Security" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 2, 5, 4 ] }, { "paper_id": "rn8r7GqJm6", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Rahul Moorthy Mahesh", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" }, { "name": "Ibrahim Isler", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 35, "n_ref": 37, "n_ref_all": 71, "n_fig": 14, "n_tab": 5, "L_tab": 1720, "n_element_tab": 248, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 177, "n_element_tab_1": 30, "formula_len_all": 402, "formula_len_all_1": 398, "len_all": 131012, "len_all_1": 64990, "len_abs": 2356, "len_title": 150, "len_sents": 44746, "len_sents_1": 31737, "n_sents": 369, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 2369, "title": "VisDiff: SDF-Guided Polygon Generation for Visibility Reconstruction and Recognition", "abs": "The capability to learn latent representations plays a key role in the effectiveness\nof recent machine learning methods. An active frontier in representation learning\nis understanding representations for combinatorial structures which may not\nadmit well-behaved local neighborhoods or distance functions. For example, for\npolygons, slightly perturbing vertex locations might lead to significant changes in\ntheir combinatorial structure (expressed as their triangulation or visibility graph)\nand may even lead to invalid polygons. In this paper, we investigate representations\nto capture the underlying combinatorial structures of polygons. Specifically,\nwe study the open problem of Visibility Reconstruction: Given a visibility graph\nG, construct a polygon P whose visibility graph is G. Visibility Reconstruction\nbelongs to the Existential Theory of Reals (\u2203R) complexity class (which lies between\nNP and P-SPACE). Currently, reconstruction algorithms are available only\nfor specific polygon classes. Establishing the hardness of the general problem is\nopen.\n\n\nWe introduce VisDiff, a novel diffusion-based approach to reconstruct a polygon\nfrom its given visibility graph G. Our method first estimates the signed distance\nfunction (SDF) of P from G. Afterwards, it extracts ordered vertex locations\nthat have the pairwise visibility relationship given by the edges of G. Our main\ninsight is that going through the SDF significantly improves learning for reconstruction.\nIn order to train VisDiff, we make two main contributions: (1) We\ndesign novel loss components for computing the visibility in a differentiable manner\nand (2) create a carefully curated dataset. We use this dataset to benchmark\nour method and achieve 21% improvement in F1-Score over standard methods.\nWe also demonstrate effective generalization to out-of-distribution polygon types\nand show that learning a generative model allows us to sample the set of polygons\nwith a given visibility graph. Finally, we extend our method to the related\ncombinatorial problem of reconstruction from a triangulation. We achieve 95%\nclassification accuracy of triangulation edges and a 4% improvement in Chamfer\ndistance compared to current architectures. Lastly, we provide preliminary results\non the harder visibility graph recognition problem in which the input G is not\nguaranteed to be a visibility graph.", "keywords": [ "Polygon Reconstruction", "Visibility Reconstruction", "Triangulation Dual", "Geometric Reasoning", "Generative Models" ], "rating_list": [ 6, 8, 3, 6, 5 ], "soundness_list": [ 2, 3, 1, 2, 3 ], "presentation_list": [ 2, 4, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 2, 2, 3, 3, 3 ] }, { "paper_id": "rlsWIBDWhW", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Sangkyun Lee", "gender": "unknown", "institution": "Korea University", "country": "", "position": "Associate Professor" }, { "name": "Sungmin Han", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" }, { "name": "JEONGHYUN LEE", "gender": "Male", "institution": "Korea University", "country": "", "position": "PhD student" }, { "name": "SEONGYUN SEO", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 31, "n_ref": 88, "n_ref_all": 106, "n_fig": 3, "n_tab": 4, "L_tab": 1145, "n_element_tab": 165, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 592, "n_element_tab_1": 74, "formula_len_all": 905, "formula_len_all_1": 954, "len_all": 110366, "len_all_1": 62252, "len_abs": 1744, "len_title": 120, "len_sents": 37911, "len_sents_1": 30108, "n_sents": 262, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1756, "title": "Cluster-Driven Adversarial Perturbations for Robust Contrastive Learning", "abs": "Adversarial contrastive learning aims to learn a representation space robust to adversarial inputs using only unlabeled data. Existing methods typically generate adversarial perturbations by maximizing the contrastive loss during adversarial training. However, we find that the effectiveness of this approach is influenced by the composition of positive and negative examples in a minibatch, which is not explicitly controllable. To address this limitation, we propose a novel approach to adversarial contrastive learning, where adversarial perturbations are generated based on the clustering structure of the representation space learned through contrastive learning. Our method is motivated by the observation that contrastive learning produces a well-separated representation space, where similar data points cluster together in space, while dissimilar ones are positioned farther apart. We hypothesize that perturbations directed toward neighboring (the second nearest to be specific) clusters are likely to cross the decision boundary of a downstream classifier built upon contrastive learning, effectively acting as adversarial examples. A key challenge in our approach is to determine a sufficiently large number of clusters, for which the number of classes in the downstream task would serve the purpose but is typically unknown during adversarial contrastive learning. Therefore, we employ the silhouette score to identify the optimal number of clusters, ensuring high-quality clustering in the representation space. Compared to the existing approaches, our method achieved up to $2.25$\\% and $5.05$\\% improvements in robust accuracy against PGD and Auto-Attack, respectively, showing slight improvement in standard accuracy as well in most cases.", "keywords": [ "adversarial training", "contrastive learning", "adversarial contrastive learning" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "rlgplAuN2p", "primary_area": "foundation or frontier models, including LLMs", "n_author": 12, "author_info_list": [ { "name": "Branislav Kveton", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Principal Scientist" }, { "name": "Jingbo Shang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Julian McAuley", "gender": "Male", "institution": "University of California, San Diego, University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Junda Wu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Lina Yao", "gender": "Female", "institution": "University of New South Wales", "country": "AU", "position": "Full Professor" }, { "name": "Ruoyu Wang", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "PhD student" }, { "name": "Tong Yu", "gender": "unknown", "institution": "Adobe Research", "country": "US", "position": "Senior Research Scientist" }, { "name": "Xiang Chen", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Xintong Li", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Yu Xia", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Yuxin Xiong", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Jia-ning Wang", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 10, "n_ref_uni": 67, "n_ref": 96, "n_ref_all": 108, "n_fig": 5, "n_tab": 2, "L_tab": 1372, "n_element_tab": 142, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2162, "n_element_tab_1": 270, "formula_len_all": 836, "formula_len_all_1": 642, "len_all": 178950, "len_all_1": 65651, "len_abs": 1987, "len_title": 129, "len_sents": 39013, "len_sents_1": 30769, "n_sents": 265, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1996, "title": "OCEAN: Offline Chain-of-thought Evaluation and Alignment in Large Language Models", "abs": "Offline evaluation of LLMs is crucial in understanding their capacities, though current methods remain underexplored in existing research. In this work, we focus on the offline evaluation of the chain-of-thought capabilities and show how to optimize LLMs based on the proposed evaluation method. To enable offline feedback with rich knowledge and reasoning paths, we use knowledge graphs (e.g., Wikidata5m) to provide feedback on the generated chain of thoughts. Due to the heterogeneity between LLM reasoning and knowledge graph structures, direct interaction and feedback from knowledge graphs on LLM behavior are challenging, as they require accurate entity linking and grounding of LLM-generated chains of thought in the knowledge graph. To address the above challenge, we propose an offline chain-of-thought evaluation framework, OCEAN, which models chain- of-thought reasoning in LLMs as a Markov Decision Process (MDP), and evaluate the policy\u2019s alignment with knowledge graph preference modeling. To overcome the reasoning heterogeneity and grounding problems, we leverage on-policy knowledge graph exploration and reinforcement learning to model a knowledge graph policy that generates token-level likelihood distributions for LLM-generated chain-of-thought reasoning paths, simulating knowledge graph reasoning preference. Then we incorporate the knowledge-graph feedback on the validity and alignment of the generated reasoning paths into inverse propensity scores and propose KG-IPS estimator. Theoretically, we prove the unbiasedness of the proposed KG-IPS estimator and provide a lower bound on its variance. With the off-policy evaluated value function, we can directly enable off-policy optimization to further enhance chain-of-thought alignment. Our empirical study shows that OCEAN can be efficiently optimized for generating chain-of-thought reasoning paths with higher estimated values without affecting LLMs\u2019 general abilities in downstream tasks or their internal knowledge.", "keywords": [ "chain-of-thought", "large language models", "offline policy evaluation" ], "rating_list": [ 6, 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "rkzabmWl5k", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Hang Zhou", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Hanlin Shang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Hao Zhu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Hui Li", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Jiahao Cui", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "Jingdong Wang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Chief Scientist for Computer Vision" }, { "name": "Kaihui Cheng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Siyu Zhu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yao Yao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 17, "n_ref": 37, "n_ref_all": 55, "n_fig": 10, "n_tab": 7, "L_tab": 952, "n_element_tab": 105, "n_fig_1": 9, "n_tab_1": 7, "L_tab_1": 1532, "n_element_tab_1": 90, "formula_len_all": 612, "formula_len_all_1": 517, "len_all": 116399, "len_all_1": 59013, "len_abs": 1806, "len_title": 127, "len_sents": 32423, "len_sents_1": 26797, "n_sents": 239, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1815, "title": "Hallo2: Long-Duration and High-Resolution Audio-Driven Portrait Image Animation", "abs": "Recent advances in latent diffusion-based generative models for portrait image animation, such as Hallo, have achieved impressive results in short-duration video synthesis. In this paper, we present updates to Hallo, introducing several design enhancements to extend its capabilities.First, we extend the method to produce long-duration videos. To address substantial challenges such as appearance drift and temporal artifacts, we investigate augmentation strategies within the image space of conditional motion frames. Specifically, we introduce a patch-drop technique augmented with Gaussian noise to enhance visual consistency and temporal coherence over long duration.Second, we achieve 4K resolution portrait video generation. To accomplish this, we implement vector quantization of latent codes and apply temporal alignment techniques to maintain coherence across the temporal dimension. By integrating a high-quality decoder, we realize visual synthesis at 4K resolution.Third, we incorporate adjustable semantic textual labels for portrait expressions as conditional inputs. This extends beyond traditional audio cues to improve controllability and increase the diversity of the generated content. To the best of our knowledge, Hallo2, proposed in this paper, is the first method to achieve 4K resolution and generate hour-long, audio-driven portrait image animations enhanced with textual prompts. We have conducted extensive experiments to evaluate our method on publicly available datasets, including HDTF, CelebV, and our introduced ''Wild'' dataset. The experimental results demonstrate that our approach achieves state-of-the-art performance in long-duration portrait video animation, successfully generating rich and controllable content at 4K resolution for duration extending up to tens of minutes.", "keywords": [ "image animation", "patch-drop augmentation", "high-resolution", "long-duration" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "rkfiJQMFcw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Michael Girsch", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Reinhard Heckel", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 16, "n_ref": 63, "n_ref_all": 100, "n_fig": 18, "n_tab": 3, "L_tab": 3323, "n_element_tab": 219, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 562, "formula_len_all_1": 245, "len_all": 116681, "len_all_1": 53036, "len_abs": 932, "len_title": 111, "len_sents": 39455, "len_sents_1": 25760, "n_sents": 323, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 936, "title": "Trace Reconstruction for DNA Data Storage using Language Models", "abs": "DNA is a promising storage medium due to its high information density and\nlongevity. However, the storage process introduces errors, thus algorithms and\ncodes are required for reliable storage. A common important step in the recovery\nof the information from DNA is trace reconstruction. In the trace reconstruction\nproblem, the goal is to construct a sequence from noisy copies corrupted by deletion,\ninsertion, and substitution errors. In this paper, we propose to use language\nmodels trained with next-token prediction for trace reconstruction. A simple channel\nmodel for the DNA data storage pipeline allows for self-supervised pretraining\non large amounts of synthetic data. Additional finetuning on real data enables us\nto adapt to technology-dependent error statistics. The proposed method (TReconLM) outperforms\nstate-of-the-art trace reconstruction algorithms for DNA data storage, often\nrecovering significantly more sequences.", "keywords": [ "DNA Data Storage", "Trace Reconstruction", "Language Models" ], "rating_list": [ 5, 3, 8, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "rkc79rOJu8", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Ashutosh Soni", "gender": "Male", "institution": "The Ohio State University", "country": "US", "position": "PhD student" }, { "name": "Atilla Eryilmaz", "gender": "Male", "institution": "Ohio State University", "country": "US", "position": "Full Professor" }, { "name": "Ness Shroff", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Full Professor" }, { "name": "Peizhong Ju", "gender": "Male", "institution": "University of Kentucky", "country": "US", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 10, "n_ref_uni": 10, "n_ref": 24, "n_ref_all": 69, "n_fig": 17, "n_tab": 8, "L_tab": 3361, "n_element_tab": 351, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 527, "n_element_tab_1": 64, "formula_len_all": 2140, "formula_len_all_1": 801, "len_all": 147028, "len_all_1": 70088, "len_abs": 1524, "len_title": 107, "len_sents": 49874, "len_sents_1": 28427, "n_sents": 339, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1528, "title": "BeST - A Novel Source Selection Metric for Transfer Learning", "abs": "One of the most fundamental, and yet relatively less explored, goals in transfer learning is the efficient means of selecting top candidates from a large number of previously trained models (optimized for various \"source\" tasks) that would perform the best for a new \"target\" task with a limited amount of data. In this paper, we undertake this goal by developing a novel task-similarity metric (BeST) and an associated method that consistently performs well in identifying the most transferrable source(s) for a given task. In particular, our design employs an innovative quantization-level optimization procedure in the context of classification tasks that yields a measure of similarity between a source model and the given target data. The procedure uses a concept similar to early stopping (usually implemented to train deep neural networks (DNNs) to ensure generalization) to derive a function that approximates the transfer learning mapping without training. The advantage of our metric is that it can be quickly computed to identify the top candidate(s) for a given target task before a computationally intensive transfer operation (typically using DNNs) can be implemented between the selected source and the target task. As such, our metric can provide significant computational savings for transfer learning from a selection of a large number of possible source models. Through extensive experimental evaluations, we establish that our metric performs well over different datasets and varying numbers of data samples.", "keywords": [ "Transfer Learning", "Pretrained Model Selection", "Task-Similarity Metric" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 1, 4 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "rkAqvDnnmO", "primary_area": "applications to neuroscience & cognitive science", "n_author": 6, "author_info_list": [ { "name": "Fengyu Cong", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Haihua Luo", "gender": "Male", "institution": "University of Jyv\u00e4skyl\u00e4", "country": "FI", "position": "PhD student" }, { "name": "Jiangrong Shen", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" }, { "name": "Qi Xu", "gender": "Male", "institution": "School of Computer Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Xuming Ran", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Research Assistant" }, { "name": "Dianbo Liu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 79, "n_ref_all": 104, "n_fig": 9, "n_tab": 10, "L_tab": 4476, "n_element_tab": 712, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1157, "n_element_tab_1": 203, "formula_len_all": 844, "formula_len_all_1": 582, "len_all": 149022, "len_all_1": 62521, "len_abs": 1779, "len_title": 144, "len_sents": 41491, "len_sents_1": 28106, "n_sents": 326, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1815, "title": "A Simple Efficiency Incremental Learning Framework via Vision-Language Model with Multi-Adapters", "abs": "Incremental Learning (IL) aims to learn new tasks while preserving previous knowledge. Integration of the zero-shot learning capabilities of pre-trained vision-language models into IL methods has been a significant advancement. However, these methods face three primary challenges: 1) the need for improved training efficiency; 2) the need for a memory bank to store previous data; and 3) the need for a strong backbone to augment the model\u2019s capabilities. In this paper, we propose the $\\textbf{SimE}$ that is a $\\textbf{Sim}$ple $\\textbf{E}$fficiency framework which is the vision-language model with an adapter designed for solving the IL task. We report a remarkable phenomenon that there is not always a direct positive correlation between the number of adaptive adapter connections and the model's incremental learning (IL) capabilities. While increasing the number of adapter connections between transformer blocks positively impacts model performance, within transformer blocks, adding more adaptive connections in smaller incremental stages does not enhance, and may even degrade the model's IL ability. Such improvement only occurs at advanced incremental stages. Extensive experimental results show SimE surpasses traditional methods by 9.6\\% on TinyImageNet and outperforms other CLIP-based methods by 5.3\\% on CIFAR-100. Notably, the SimE, with only thousands of parameters and a 0-size memory bank, exceeds the ZSCL with 140 million parameters and also beats the CoOP with a 1000-size memory bank. Besides, we also conduct a systematic study to enhance the utilization of the zero-shot capabilities of CLIP. We suggest that the backbone of the encoder in SimE use the image encoder from CLIP that is pre-trained on large datasets, like LAION-2B, and larger model sizes, such as ViT-L/14, for IL tasks.", "keywords": [ "Continual learning; CLIP; Adapter" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "rjuZyMfLSd", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Dongjin Song", "gender": "Male", "institution": "University of Connecticut", "country": "US", "position": "Assistant Professor" }, { "name": "Xikun ZHANG", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Yixin Chen", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "", "position": "Full Professor" }, { "name": "Yushan Jiang", "gender": "Male", "institution": "University of Connecticut", "country": "US", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 11, "n_ref_uni": 58, "n_ref": 126, "n_ref_all": 155, "n_fig": 8, "n_tab": 2, "L_tab": 271, "n_element_tab": 26, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 135, "n_element_tab_1": 2, "formula_len_all": 2551, "formula_len_all_1": 974, "len_all": 180561, "len_all_1": 76395, "len_abs": 1091, "len_title": 91, "len_sents": 56568, "len_sents_1": 35625, "n_sents": 432, "n_sents_1": 279, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 1095, "title": "Learning system dynamics without forgetting", "abs": "Observation-based trajectory prediction for systems with unknown dynamics is essential in fields such as physics and biology. Most existing approaches are limited to learning within a single system with fixed dynamics patterns. However, many real-world applications require learning across systems with evolving dynamics patterns, a challenge that has been largely overlooked. To address this, we systematically investigate the problem of Continual Dynamics Learning (CDL), examining task configurations and evaluating the applicability of existing techniques, while identifying key challenges. In response, we propose the Mode-switching Graph ODE (MS-GODE) model, which integrates the strengths LG-ODE and sub-network learning with a mode-switching module, enabling efficient learning over varying dynamics. Moreover, we construct a novel benchmark of biological dynamic systems for CDL, Bio-CDL, featuring diverse systems with disparate dynamics and significantly enriching the research field of machine learning for dynamic systems. Our code and benchmark datasets will be publicly available.", "keywords": [ "graph neural networks", "AI4Science", "physics", "biology" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "rjcA3km2IT", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Deming Zhai", "gender": "Female", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Gangfeng Hu", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Jialiang Wang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Junjun Jiang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xiangyang Ji", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xianming Liu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xiong Zhou", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 22, "n_ref_uni": 26, "n_ref": 103, "n_ref_all": 122, "n_fig": 3, "n_tab": 6, "L_tab": 5047, "n_element_tab": 166, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 5074, "n_element_tab_1": 138, "formula_len_all": 2964, "formula_len_all_1": 1842, "len_all": 135095, "len_all_1": 70395, "len_abs": 4404, "len_title": 103, "len_sents": 39537, "len_sents_1": 26308, "n_sents": 378, "n_sents_1": 261, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1361, "title": "Variation-Bounded Losses for Learning with Noisy Labels", "abs": "The presence of noisy labels poses a significant challenge for training accurate deep neural networks.\nPrevious works have proposed various robust loss functions designed to address this issue, which, however, often suffer from several drawbacks, such as underfitting or insufficient noise-tolerance. Furthermore, there is currently no reliable metric to guide the design of more effective robust loss functions.\nIn this paper, we introduce the *Variation Ratio* as a novel metric to measure the robustness of loss functions. Leveraging this metric, we propose a new family of robust loss functions, termed *Variation-Bounded Losses* (VBL), characterized by a bounded variation ratio.\nWe investigate theoretical properties of variation-bounded losses and prove that a smaller variation ratio would lead to better robustness. Additionally, we show that the variation ratio provides a more relaxed condition than the commonly used symmetric condition for achieving noise-tolerant learning, making it a valuable tool for designing effective robust loss functions.\nWe modify several commonly used loss functions to the variation-bounded form.\nThese variation-bounded losses are characterized by their simplicity, effectiveness, and theoretical guarantees.\nExtensive experiments demonstrate the superiority of our method in mitigating various types of label noise.", "keywords": [ "Learning with Noisy Labels; Robust Loss Functions; Multi-Class Classification" ], "rating_list": [ 8, 6, 1, 5, 8 ], "soundness_list": [ 3, 4, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "rj7wUcLgfw", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Masanari Kimura", "gender": "unknown", "institution": "University of Melbourne", "country": "AU", "position": "Postdoc" }, { "name": "Ryotaro Shimizu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Researcher" }, { "name": "Yuki Saito", "gender": "Male", "institution": "ZOZO Research", "country": "", "position": "Researcher" }, { "name": "Yuya Yoshikawa", "gender": "unknown", "institution": "STAIR Lab, Chiba Institute of Technology", "country": "JP", "position": "Senior Research Scientist" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 32, "n_ref": 45, "n_ref_all": 59, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 732, "formula_len_all_1": 732, "len_all": 100758, "len_all_1": 64823, "len_abs": 1396, "len_title": 154, "len_sents": 33843, "len_sents_1": 31094, "n_sents": 229, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1402, "title": "Explaining Black-box Model Predictions via Two-level Nested Feature Attributions with Consistency Property", "abs": "Techniques that explain the predictions of black-box machine learning models are crucial to make the models transparent, thereby increasing trust in AI systems. The input features to the models often have a nested structure that consists of high- and low-level features, and each high-level feature is decomposed into multiple low-level features. For such inputs, both high-level feature attributions (HiFAs) and low-level feature attributions (LoFAs) are important for better understanding the model's decision. In this paper, we propose a model-agnostic local explanation method that effectively exploits the nested structure of the input to estimate the two-level feature attributions simultaneously. A key idea of the proposed method is to introduce the consistency property that should exist between the HiFAs and LoFAs, thereby bridging the separate optimization problems for estimating them. Thanks to this consistency property, the proposed method can produce HiFAs and LoFAs that are both faithful to the black-box models and consistent with each other, using a smaller number of queries to the models. In experiments on image classification in multiple instance learning and text classification using language models, we demonstrate that the HiFAs and LoFAs estimated by the proposed method are accurate, faithful to the behaviors of the black-box models, and provide consistent explanations.", "keywords": [ "explainable AI; XAI; feature attribution; black-box model; local explanation; model-agnostic; nested features" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "riieAeQBJm", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Jianqing Xu", "gender": "Male", "institution": "Tencent YOUTU Lab", "country": "CN", "position": "Researcher" }, { "name": "Shouhong Ding", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "researcher" }, { "name": "Shuigeng Zhou", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Xiao Lin", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Intern" }, { "name": "Yuge Huang", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Yuxi Mi", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 46, "n_ref": 120, "n_ref_all": 140, "n_fig": 10, "n_tab": 3, "L_tab": 1077, "n_element_tab": 121, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1068, "n_element_tab_1": 119, "formula_len_all": 550, "formula_len_all_1": 549, "len_all": 131966, "len_all_1": 66025, "len_abs": 1890, "len_title": 158, "len_sents": 34447, "len_sents_1": 31328, "n_sents": 246, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 109, "L_abs": 1934, "title": "UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition", "abs": "Face recognition (FR) stands as one of the most crucial applications in computer vision. The accuracy of FR models has significantly improved in recent years due to the availability of large-scale human face datasets. However, directly using these datasets can inevitably lead to privacy and legal problems. Generating synthetic data to train FR models is a feasible solution to circumvent these issues. While existing synthetic-based face recognition methods have made significant progress in generating identity-preserving images, they are severely plagued by context overfitting, resulting in a lack of intra-class diversity of generated images and poor face recognition performance. In this paper, we propose a framework to $\\textbf{U}$nleash model $\\textbf{I}$nherent capabilities to enhance intra-class diversity for synthetic face recognition, shorted as $\\textbf{UIFace}$. Our framework first train a diffusion model that can perform denoising conditioned on either identity contexts or a learnable empty context. The former generates identity-preserving images but lacks variations, while the latter exploits the model's intrinsic ability to synthesize intra-class-diversified images but with random identities. Then we adopt a novel two-stage denoising strategy to fully leverage the strengths of both type of contexts, resulting in images that are diverse as well as identity-preserving. Moreover, an attention injection module is introduced to further augment the intra-class variations by utilizing attention maps from the empty context to guide the denoising process in ID-conditioned generation. Experiments show that our method significantly surpasses previous approaches with even less training data and half the size of synthetic dataset. More surprisingly, the proposed $\\textbf{UIFace}$ even achieves comparable performance of FR models trained on real datasets when we increase the number of synthetic identities.", "keywords": [ "face recognition", "face image synthesis", "diffusion model" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "riTiq3i21b", "primary_area": "datasets and benchmarks", "n_author": 13, "author_info_list": [ { "name": "Alex L Zhang", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Research Assistant" }, { "name": "Carlos E Jimenez", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Diyi Yang", "gender": "Female", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Gabriel Synnaeve", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "John Yang", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Joyce Yang", "gender": "unknown", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "Undergrad student" }, { "name": "Karthik R Narasimhan", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Kilian Lieret", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Niklas Muennighoff", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Ofir Press", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Ori Press", "gender": "unknown", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "", "position": "PhD student" }, { "name": "Sida Wang", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Xindi Wu", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 114, "n_ref_all": 158, "n_fig": 8, "n_tab": 13, "L_tab": 3591, "n_element_tab": 732, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1348, "n_element_tab_1": 143, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 268304, "len_all_1": 67840, "len_abs": 2100, "len_title": 139, "len_sents": 84357, "len_sents_1": 32706, "n_sents": 664, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1451, "title": "SWE-bench Multimodal: Do AI Systems Generalize to Visual Software Domains?", "abs": "Autonomous systems for software engineering are now capable of fixing bugs and developing features. These systems are commonly evaluated on SWE-bench (Jimenez et al., 2024a), which assesses their ability to solve software issues from GitHub repositories. However, SWE-bench uses only Python repositories, with problem statements presented predominantly as text and lacking visual elements such as images. This limited coverage motivates our inquiry into how existing systems might perform on unrepresented software engineering domains\n(e.g., front-end, game development, DevOps), which use different programming languages and paradigms. Therefore, we propose SWE-bench Multimodal (SWE-bench M), to evaluate systems on their ability to fix bugs in visual, user-facing JavaScript software. SWE-bench M features 617 task instances collected from 17 JavaScript libraries used for web interface design, diagramming, data visualization, syntax highlighting, and interactive mapping. Each SWE-bench M task instance contains at least one image in its problem statement or unit tests. Our analysis finds that top-performing SWE-bench systems struggle with SWE-bench M, revealing limitations in visual problem-solving and cross-language generalization. Lastly, we show that SWE-agent\u2019s flexible language-agnostic features enable it to substantially outperform alternatives on SWE-bench M, resolving 12% of task\ninstances compared to 6% for the next best system.", "keywords": [ "Language models", "Natural language processing", "Software engineering" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "rhhQjGj09A", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Francesca Mignacco", "gender": "Female", "institution": "Princeton University", "country": "", "position": "Postdoc" }, { "name": "Francesco Mori", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Stefano Sarao Mannelli", "gender": "Male", "institution": "Chalmers University of Technology", "country": "SE", "position": "Assistant Professor" } ], "n_formula": 65, "n_formula_1": 20, "n_ref_uni": 69, "n_ref": 108, "n_ref_all": 141, "n_fig": 15, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5190, "formula_len_all_1": 1440, "len_all": 188485, "len_all_1": 70264, "len_abs": 1130, "len_title": 131, "len_sents": 53861, "len_sents_1": 32629, "n_sents": 492, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1137, "title": "Optimal Protocols for Continual Learning via Statistical Physics and Control Theory", "abs": "Artificial neural networks often struggle with _catastrophic forgetting_ when learning multiple tasks sequentially, as training on new tasks degrades the performance on previously learned tasks. Recent theoretical work has addressed this issue by analysing learning curves in synthetic frameworks under predefined training protocols. However, these protocols relied on heuristics and lacked a solid theoretical foundation assessing their optimality. In this paper, we fill this gap by combining exact equations for training dynamics, derived using statistical physics techniques, with optimal control methods. We apply this approach to teacher-student models for continual learning and multi-task problems, obtaining a theory for task-selection protocols maximising performance while minimising forgetting. Our theoretical analysis offers non-trivial yet interpretable strategies for mitigating catastrophic forgetting, shedding light on how optimal learning protocols modulate established effects, such as the influence of task similarity on forgetting. Finally, we validate our theoretical findings with experiments on real-world data.", "keywords": [ "machine learning theory", "statistical physics", "online learning", "continual learning", "optimal control theory" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "rhfOzJzsKN", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Dong-Jun Han", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Hasnain Irshad Bhatti", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Humaira Kousar", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jaekyun Moon", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "Full Professor" }, { "name": "Mohammad Mahdi Rahimi", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Younghyun Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 70, "n_formula_1": 11, "n_ref_uni": 37, "n_ref": 61, "n_ref_all": 75, "n_fig": 6, "n_tab": 7, "L_tab": 5011, "n_element_tab": 398, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1839, "n_element_tab_1": 117, "formula_len_all": 6062, "formula_len_all_1": 716, "len_all": 165882, "len_all_1": 70153, "len_abs": 1816, "len_title": 131, "len_sents": 46803, "len_sents_1": 31914, "n_sents": 390, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1626, "title": "Communication-Efficient Federated Learning via Model-Agnostic Projection Adaptation", "abs": "Federated learning (FL) enables collaborative model training across distributed clients without centralizing sensitive raw data while benefiting from diverse data sources. \nDespite recent advancements in FL, the communication overhead remains a significant challenge, especially for large-scale models.\nRecent low-rank adaptation (LoRA) techniques have shown promise in reducing these burdens in FL, but they are typically applied to each layer individually and depend on the model architecture, which limits their performance.\nTo address these shortcomings, we propose Model-Agnostic Projection Adaptation (MAPA), a novel approach that applies factorization to the entire model parameter space, which we view as a *single vector*, regardless of the number of layers and model architecture. \nMAPA factorizes the single-vector model update into a fixed *reconstruction matrix* and a trainable *projection vector*, with the reconstruction matrix being randomly initialized using a shared seed at each round. \nThis ensures that *only* the projection vectors need to be communicated to the server, thereby reducing the communication cost.\nFurthermore, MAPA's vector-based representation and relaxed rank constraints allow for a larger reconstruction matrix and smaller projection vector dimensions compared to LoRA, enhancing the expressiveness of model updates while significantly reducing communication overhead. \nExperimental results demonstrate that MAPA outperforms existing FL methods in both communication efficiency and model performance, effectively coupling optimization and communication efficiency in FL environments.", "keywords": [ "Federated Learning", "Low-Rank Adaptation", "Communication Efficiency", "Subspace Optimization" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 5 ] }, { "paper_id": "rh54qNvxKO", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Chen Gao", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Dongyun Zou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Jinzhu Mao", "gender": "Male", "institution": "Tsinghua University ", "country": "", "position": "MS student" }, { "name": "Li Sheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Siyi Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yue Wang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 18, "n_ref": 28, "n_ref_all": 40, "n_fig": 4, "n_tab": 7, "L_tab": 2508, "n_element_tab": 339, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1925, "n_element_tab_1": 178, "formula_len_all": 413, "formula_len_all_1": 443, "len_all": 105382, "len_all_1": 58341, "len_abs": 1162, "len_title": 117, "len_sents": 32562, "len_sents_1": 25807, "n_sents": 215, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1167, "title": "Identify Critical Nodes in Complex Network with Large Language Models", "abs": "Identifying critical nodes in networks is a classical combinatorial optimization task, and many methods struggle to strike a balance between adaptability and utility. Therefore, we propose an approach that empowers Evolutionary Algorithm (EA) with Large Language Models (LLMs), to generate a function called \"score_nodes\" which can further be used to identify crucial nodes based on their assigned scores. Our model consists of three main components: Manual Initialization, Population Management, and LLMs-based Evolution, and it evolves from initial populations with a set of designed node scoring functions created manually. LLMs leverage their strong contextual understanding and rich programming techniques to perform crossover and mutation operations on the individuals, generating new functions. These functions are then categorized, ranked, and eliminated to ensure the stable development of the populations while preserving diversity. Extensive experiments demonstrate the excellent performance of our method compared to other state-of-the-art algorithms. It can generate diverse and efficient node scoring functions to identify critical nodes in the network.", "keywords": [ "Complex Networks", "Large Language Models" ], "rating_list": [ 3, 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 1, 1, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 3, 4, 4 ] }, { "paper_id": "rgwquPxhIh", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Aibek Alanov", "gender": "Male", "institution": "Higher School of Economics", "country": "RU", "position": "Principal Researcher" }, { "name": "Maksim Nakhodnov", "gender": "Male", "institution": "Moscow State University, Lomonosov Moscow State University", "country": "", "position": "PhD student" }, { "name": "Vera Soboleva", "gender": "unknown", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 15, "n_ref": 51, "n_ref_all": 82, "n_fig": 18, "n_tab": 1, "L_tab": 201, "n_element_tab": 42, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1218, "formula_len_all_1": 1247, "len_all": 116218, "len_all_1": 58924, "len_abs": 2024, "len_title": 94, "len_sents": 42391, "len_sents_1": 26178, "n_sents": 282, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 2035, "title": "Beyond Fine-Tuning: A Systematic Study of Sampling Techniques in Personalized Image Generation", "abs": "Personalized text-to-image generation focuses on creating customized images based on user-defined concepts and text descriptions. A good balance between learned concept fidelity and its ability to be generated in different contexts is a major challenge in this task. Modern personalization techniques often strive to find this balance through diverse fine-tuning parameterizations and enhanced sampling methods that integrate superclass trajectories into the backward diffusion process. Improved sampling methods present a cost-effective, training-free way to enhance already fine-tuned models. However, outside of fine-tuning approaches, there is no systematic analysis of sampling methods in the personalised generation literature. Most sampling techniques are introduced alongside fixed fine-tuning parameterizations, which makes it difficult to identify the impact of sampling on the generation outcomes and whether it can be applied with other fine-tuning strategies. Moreover, they don't compare with the naive sampling approaches, so the intuition of how the superclass trajectory affects the sampling process remains underexplored. In this work, we propose a systematic and comprehensive analysis of personalized generation sampling strategies beyond the fine-tuning methods. We explore various combinations of concept and superclass trajectories, developing a deep understanding of how superclass influence generation outputs. Based on these results, we demonstrate that even a weighted mix of the concept and superclass trajectory can establish a strong baseline that enhances the adaptability of concepts across different contexts and can be effectively transferred to any training strategy, including various fine-tuning parameterizations, text embedding optimization, and hypernetworks. We analyze all methods through the lens of the trade-off between concept fidelity, editability, and computational efficiency, ultimately providing a framework to determine which sampling method is most suitable for specific scenarios.", "keywords": [ "Generative Model", "Diffusion Model", "Subject-Driven Generation" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "rgiIZ3pcZY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Eunho Yang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Jaeryong Hwang", "gender": "Male", "institution": "Republic of Korea Naval Academy", "country": "KR", "position": "Full Professor" }, { "name": "Joonhyun Jeong", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Seyun Bae", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Yeonsung Jung", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 38, "n_ref": 89, "n_ref_all": 117, "n_fig": 12, "n_tab": 5, "L_tab": 1333, "n_element_tab": 320, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 2112, "n_element_tab_1": 66, "formula_len_all": 424, "formula_len_all_1": 306, "len_all": 132311, "len_all_1": 60438, "len_abs": 1798, "len_title": 137, "len_sents": 40061, "len_sents_1": 30570, "n_sents": 238, "n_sents_1": 144, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1804, "title": "Playing the Fool: Jailbreaking Large Language Models with Out-of-Distribution Strategies", "abs": "Despite the remarkable versatility of Large Language Models (LLMs) and Multimodal-LLMs (MLLMs) to generalize across both language and vision tasks, LLMs and MLLMs have shown vulnerability to jailbreaking, generating textual outputs that undermine safety, ethical, and bias standards when exposed to harmful or sensitive inputs. With the recent advancement of safety-alignment via preference-tuning from human feedback, LLMs and MLLMs have been equipped with safety guardrails to yield safe, ethical, and fair responses with regard to harmful inputs. However, despite the significance of safety-alignment, research on the vulnerabilities remains largely underexplored. In this paper, we investigate the vulnerability of the safety-alignment, examining its ability to consistently provide safety guarantees for out-of-distribution(OOD)-ifying harmful inputs that may fall outside the aligned data distribution. Our key observation is that OOD-ifying the vanilla harmful inputs highly increases the uncertainty of the model to discern the malicious intent within the input, leading to a higher chance of being jailbroken. Exploiting this vulnerability, we propose JOOD, a new Jailbreak strategy via generating OOD-ifying inputs beyond the safety-alignment with diverse visual and textual transformation techniques. Specifically, even simple mixing-based techniques such as image mixup prove highly effective in OOD-ifying the harmful inputs by increasing the uncertainty of the model, thereby facilitating the bypass of the safety-alignment. Experimental results across diverse jailbreak scenarios demonstrate that JOOD effectively jailbreaks recent proprietary LLMs and MLLMs such as GPT-4 and GPT-4V with high attack success rate, which previous attack approaches have consistently struggled to jailbreak.", "keywords": [ "Large Language Models", "Multimodal Large Language Models", "Safety", "Jailbreak Attacks" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "rgDwRdMwoS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Jasper Dekoninck", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Martin Vechev", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Maximilian Baader", "gender": "unknown", "institution": "ETH Zurich", "country": "CH", "position": "Postdoc" } ], "n_formula": 29, "n_formula_1": 5, "n_ref_uni": 16, "n_ref": 50, "n_ref_all": 76, "n_fig": 7, "n_tab": 12, "L_tab": 2243, "n_element_tab": 360, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 925, "n_element_tab_1": 151, "formula_len_all": 2390, "formula_len_all_1": 204, "len_all": 196641, "len_all_1": 77982, "len_abs": 2150, "len_title": 100, "len_sents": 58887, "len_sents_1": 34338, "n_sents": 551, "n_sents_1": 289, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1484, "title": "A Unified Approach to Routing and Cascading for LLMs", "abs": "The widespread applicability of large language models (LLMs) has increased the availability of many fine-tuned models of various sizes targeting specific tasks. Given a set of such specialized models, to maximize overall performance, it is important to figure out the optimal strategy for selecting the right model for a given user query. An effective strategy could drastically increase overall performance and even offer improvements over a single large monolithic model. Existing approaches typically fall into two categories: routing, where a single model is selected for each query, and cascading, which runs a sequence of increasingly larger models until a satisfactory answer is obtained. However, both have notable limitations: routing commits to an initial model without flexibility, while cascading requires executing every model in sequence, which can be inefficient. Additionally, the conditions under which these strategies are provably optimal remain unclear. In this work, we derive optimal strategies for both routing and cascading. Building on this analysis, we propose a novel approach called *cascade routing*, which combines the adaptability of routing with the cost-efficiency of cascading. Our experiments demonstrate that cascade routing consistently outperforms both routing and cascading across a variety of settings, improving both output quality and lowering computational cost, thus offering a unified and efficient solution to the model selection problem.", "keywords": [ "large language models", "routing", "model evaluation" ], "rating_list": [ 3, 8, 6, 3, 6 ], "soundness_list": [ 2, 4, 3, 1, 3 ], "presentation_list": [ 2, 4, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 1, 2 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "rfrtFwnF62", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 14, "author_info_list": [ { "name": "Chen Wang", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Cheng Lin", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Christian Theobalt", "gender": "Male", "institution": "Max-Planck-Institute for Informatics, Saarland Informatics Campus", "country": "DE", "position": "Director" }, { "name": "Lingjie Liu", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Assistant Professor" }, { "name": "Qingxuan Wu", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "MS student" }, { "name": "Sirui Xu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Intern" }, { "name": "Soshi Shimada", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute for Informatics", "country": "", "position": "PhD student" }, { "name": "Taku Komura", "gender": "Male", "institution": "the University of Hong Kong, University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Vladislav Golyanik", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute for Informatics", "country": "DE", "position": "Principal Researcher" }, { "name": "Wenping Wang", "gender": "Male", "institution": "Texas A&M University ", "country": "US", "position": "Full Professor" }, { "name": "Yuan Liu", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Zeyu Cao", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Zhengming Yu", "gender": "unknown", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Zhiyang Dou", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Visiting PhD" } ], "n_formula": 18, "n_formula_1": 7, "n_ref_uni": 91, "n_ref": 224, "n_ref_all": 235, "n_fig": 8, "n_tab": 4, "L_tab": 1786, "n_element_tab": 182, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1051, "n_element_tab_1": 128, "formula_len_all": 1324, "formula_len_all_1": 448, "len_all": 255904, "len_all_1": 62464, "len_abs": 1873, "len_title": 130, "len_sents": 44594, "len_sents_1": 29569, "n_sents": 296, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1897, "title": "DICE: End-to-end Deformation Capture of Hand-Face Interactions from a Single Image", "abs": "Reconstructing 3D hand-face interactions with deformations from a single image is a challenging yet crucial task with broad applications in AR, VR, and gaming. The challenges stem from self-occlusions during single-view hand-face interactions, diverse spatial relationships between hands and face, complex deformations, and the ambiguity of the single-view setting. The previous state-of-the-art, Decaf, employs a global fitting optimization guided by contact and deformation estimation networks trained on studio-collected data with 3D annotations. However, Decaf suffers from a time-consuming optimization process and limited generalization capability due to its reliance on 3D annotations of hand-face interaction data. To address these issues, we present DICE, the first end-to-end method for Deformation-aware hand-face Interaction reCovEry from a single image. DICE estimates the poses of hands and faces, contacts, and deformations simultaneously using a Transformer-based architecture. It features disentangling the regression of local deformation fields and global mesh vertex locations into two network branches, enhancing deformation and contact estimation for precise and robust hand-face mesh recovery. To improve generalizability, we propose a weakly-supervised training approach that augments the training set using in-the-wild images without 3D ground-truth annotations, employing the depths of 2D keypoints estimated by off-the-shelf models and adversarial priors of poses for supervision. Our experiments demonstrate that DICE achieves state-of-the-art performance on a standard benchmark and in-the- wild data in terms of accuracy and physical plausibility. Additionally, our method operates at an interactive rate (20 fps) on an Nvidia 4090 GPU, whereas Decaf requires more than 15 seconds for a single image. The code will be available at: https://github.com/Qingxuan-Wu/DICE.", "keywords": [ "interaction", "deformation", "end-to-end", "mesh recovery" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "rfeksadZox", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Feihu Zhang", "gender": "Male", "institution": "DreamTech", "country": "", "position": "Instructor" }, { "name": "Hongye Yang", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Jiachen Qian", "gender": "Male", "institution": "Fudan University", "country": "", "position": "Undergrad student" }, { "name": "Jingxi Xu", "gender": "Male", "institution": "DreamTech", "country": "CN", "position": "Researcher" }, { "name": "Shuang Wu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 47, "n_ref": 108, "n_ref_all": 129, "n_fig": 10, "n_tab": 4, "L_tab": 1911, "n_element_tab": 126, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1649, "n_element_tab_1": 64, "formula_len_all": 211, "formula_len_all_1": 211, "len_all": 140089, "len_all_1": 56011, "len_abs": 1041, "len_title": 135, "len_sents": 34197, "len_sents_1": 25430, "n_sents": 249, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1044, "title": "High-quality Text-to-3D Character Generation with SparseCubes and Sparse Transformers.", "abs": "Current state-of-the-art text-to-3D generation methods struggle to produce 3D models with fine details and delicate structures due to limitations in differentiable mesh representation techniques. This limitation is particularly pronounced in anime character generation, where intricate features such as fingers, hair, and facial details are crucial for capturing the essence of the characters.\n\nIn this paper, we introduce a novel, efficient, sparse differentiable mesh representation method, termed SparseCubes, alongside a sparse transformer network designed to generate high-quality 3D models. Our method significantly reduces computational requirements by over 95% and storage memory by 50%, enabling the creation of higher resolution meshes with enhanced details and delicate structures. We validate the effectiveness of our approach through its application to text-to-3D anime character generation, demonstrating its capability to accurately render subtle details and thin structures (e.g. individual fingers) in both meshes and textures.", "keywords": [ "anime avatar", "diffusion", "transformer", "large reconstruction model" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 1, 3, 3 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "rfdblE10qm", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Hao Sun", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Jean-Francois Ton", "gender": "Not Specified", "institution": "Bytedance", "country": "GB", "position": "Researcher" }, { "name": "Yunyi Shen", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 91, "n_formula_1": 21, "n_ref_uni": 55, "n_ref": 123, "n_ref_all": 139, "n_fig": 9, "n_tab": 3, "L_tab": 469, "n_element_tab": 61, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6653, "formula_len_all_1": 1166, "len_all": 224751, "len_all_1": 61697, "len_abs": 4444, "len_title": 125, "len_sents": 72274, "len_sents_1": 29491, "n_sents": 543, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 77, "L_abs": 1522, "title": "Rethinking Reward Modeling in Preference-based Large Language Model Alignment", "abs": "The Bradley-Terry (BT) model is a common and successful practice in reward modeling for Large Language Model (LLM) alignment. However, it remains unclear *why* this model --- originally developed for multi-player stochastic game matching --- can be adopted to convert pairwise response comparisons to reward values and make predictions. Especially given the fact that only a limited number of prompt-response pairs are sparsely compared with others. \nIn this paper, we first establish the convergence rate of BT reward models based on deep neural networks using embeddings, providing a theoretical foundation for their use.\nDespite theoretically sound, we argue that the BT model is not a necessary choice from the perspective of downstream optimization, this is because a reward model only needs to preserve the correct ranking predictions through a monotonic transformation of the true reward. \nWe highlight the critical concept of *order consistency* in reward modeling and demonstrate that the BT model possesses this property.\nMoreover, we propose a simple and straightforward upper-bound algorithm, compatible with off-the-shelf binary classifiers, as an alternative order-consistent reward modeling objective. \nTo offer practical insights, we empirically evaluate the performance of these different reward modeling approaches across more than 12,000 experimental setups, using $6$ base LLMs, $2$ datasets, and diverse annotation designs that vary in quantity, quality, and pairing choices in preference annotations.", "keywords": [ "Bradley-Terry Model", "Reward Modeling", "Large Language Models" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "rf0ZoDASnS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Bettina Messmer", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Dongyang Fan", "gender": "Female", "institution": "School of Computer and Communication Sciences, EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Martin Jaggi", "gender": "Male", "institution": "EPFL", "country": "CH", "position": "Associate Professor" } ], "n_formula": 15, "n_formula_1": 5, "n_ref_uni": 30, "n_ref": 53, "n_ref_all": 84, "n_fig": 11, "n_tab": 7, "L_tab": 1306, "n_element_tab": 59, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 2872, "n_element_tab_1": 81, "formula_len_all": 1287, "formula_len_all_1": 515, "len_all": 144959, "len_all_1": 63760, "len_abs": 1375, "len_title": 134, "len_sents": 47809, "len_sents_1": 28358, "n_sents": 385, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1425, "title": "On-Device Collaborative Language Modeling via a Mixture of Generalists and Specialists", "abs": "On-device LLMs have gained increasing attention for their ability to enhance privacy and provide a personalized user experience. \nTo facilitate learning with private and scarce local data, federated learning has become a standard approach, though it introduces challenges related to system and data heterogeneity among end users. As a solution, we propose a novel $\\textbf{Co}$llaborative learning approach with a $\\textbf{Mi}$xture of $\\textbf{G}$eneralists and $\\textbf{S}$pecialists (CoMiGS), being the first to effectively address both. Our approach distinguishes generalists and specialists by aggregating certain experts across end users while keeping others localized to specialize in user-specific datasets. A key innovation of our method is the bi-level optimization formulation of the Mixture-of-Experts learning objective, where the router is updated using a separate validation set that represents the target distribution. CoMiGS effectively balances collaboration and personalization, as demonstrated by its superior performance in scenarios with high data heterogeneity across multiple datasets. By design, our approach accommodates users' varying computational resources through different numbers of specialists. By decoupling resource abundance from data quantity, CoMiGS remains robust against overfitting\u2014due to the generalists' regularizing effect\u2014while adapting to local data through specialist expertise.", "keywords": [ "Collaborative Learning", "Federated Learning", "On-device language models", "Mixture of Experts" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "reZKq6hjOZ", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Ness Shroff", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Full Professor" }, { "name": "Peizhong Ju", "gender": "Male", "institution": "University of Kentucky", "country": "US", "position": "Assistant Professor" }, { "name": "Yingbin Liang", "gender": "Female", "institution": "The Ohio State University", "country": "US", "position": "Professor" }, { "name": "Yuchen Liang", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Postdoc" } ], "n_formula": 37, "n_formula_1": 31, "n_ref_uni": 23, "n_ref": 102, "n_ref_all": 117, "n_fig": 1, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3494, "formula_len_all_1": 2942, "len_all": 128041, "len_all_1": 72427, "len_abs": 1626, "len_title": 143, "len_sents": 35462, "len_sents_1": 31625, "n_sents": 282, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1726, "title": "Broadening Target Distributions for Accelerated Diffusion Models via a Novel Analysis Approach", "abs": "Accelerated diffusion models hold the potential to significantly enhance the efficiency of standard diffusion processes. Theoretically, these models have been shown to achieve faster convergence rates than the standard $\\mathcal O(1/\\epsilon^2)$ rate of vanilla diffusion models, where $\\epsilon$ denotes the target accuracy. However, current theoretical studies have established the acceleration advantage only for restrictive target distribution classes, such as those with smoothness conditions imposed along the entire sampling path or with bounded support. In this work, we significantly broaden the target distribution classes with a new accelerated stochastic DDPM sampler. In particular, we show that it achieves accelerated performance for three broad distribution classes not considered before. Our first class relies on the smoothness condition posed only to the target density $q_0$, which is far more relaxed than the existing smoothness conditions posed to all $q_t$ along the entire sampling path. Our second class requires only a finite second moment condition, allowing for a much wider class of target distributions than the existing finite-support condition. Our third class is Gaussian mixture, for which our result establishes the first acceleration guarantee. Moreover, among accelerated DDPM type samplers, our results specialized for bounded-support distributions show an improved dependency on the data dimension $d$. Our analysis introduces a novel technique for establishing performance guarantees via constructing a tilting factor representation of the convergence error and utilizing Tweedie's formula to handle Taylor expansion terms. This new analytical framework may be of independent interest.", "keywords": [ "generative models", "denoising diffusion probabilistic model (DDPM)", "convergence analysis", "accelerated methods" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 4, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "re7jrIyghD", "primary_area": "applications to neuroscience & cognitive science", "n_author": 6, "author_info_list": [ { "name": "Aldo Cordova-Palomera", "gender": "unknown", "institution": "Optum", "country": "", "position": "Researcher" }, { "name": "Brian L. Hill", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Eran Halperin", "gender": "Not Specified", "institution": "University of California, Los Angeles", "country": "", "position": "Full Professor" }, { "name": "Vijay S Nori", "gender": "Male", "institution": "Optum", "country": "US", "position": "Principal Researcher" }, { "name": "Melikasadat Emami", "gender": "Female", "institution": "Optum AI", "country": "US", "position": "Machine Learning Scientist" }, { "name": "Robert E. Tillman", "gender": "Male", "institution": "Optum AI Labs", "country": "", "position": "Research Director" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 20, "n_ref": 37, "n_ref_all": 60, "n_fig": 8, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 2547, "n_element_tab_1": 29, "formula_len_all": 286, "formula_len_all_1": 286, "len_all": 108714, "len_all_1": 62840, "len_abs": 1499, "len_title": 124, "len_sents": 35025, "len_sents_1": 29943, "n_sents": 235, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1506, "title": "CHIRon: A Generative Foundation Model for Structured Sequential Medical Data", "abs": "Recent advances in large language models (LLMs) have shown that foundation models (FMs) can learn highly complex representations of sequences that can be used for downstream generative and discriminative tasks such as text generation and classification. While most FMs focus on text, recent work has shown FMs can be learnt for sequential medical data, e.g. ICD-10 diagnosis codes associated with specific patient visits. These FMs demonstrate improved performance on downstream discriminative disease classification tasks. In this paper, we introduce CHIRon, a decoder-only generative FM for sequential medical data. CHIRon utilizes causal masking during pre-training, enabling generative applications, and incorporates a number of architectural improvements and support for additional medical data types (diagnoses, procedures, medications, lab results, place of service, demographics). We introduce a new pre-training objective function that incorporates tasks for predicting place of service and patient's age at encounter in addition to the next medical code prediction task. To incorporate lab results into the model, we develop and evaluate several methods for embedding the continuous lab values. Furthermore, we introduce a causal visit-based masking approach for training CHIRon based on patient visits. We show empirically that CHIRon can be used to generate realistic sequential medical data and also outperforms state of the art FMs for sequential medical data on disease classification tasks.", "keywords": [ "foundation models", "large language models", "generative models", "disease progression", "medical codes" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "rdv6yeMFpn", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Bohang Zhang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Haggai Maron", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Assistant Professor" }, { "name": "Jingchu Gai", "gender": "Male", "institution": "School of Mathematical Science, Peking University, Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Liwei Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Yiheng Du", "gender": "Male", "institution": "Peking University", "country": "", "position": "Undergrad student" } ], "n_formula": 140, "n_formula_1": 16, "n_ref_uni": 48, "n_ref": 133, "n_ref_all": 154, "n_fig": 7, "n_tab": 8, "L_tab": 162, "n_element_tab": 30, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 170, "n_element_tab_1": 3, "formula_len_all": 17301, "formula_len_all_1": 1871, "len_all": 368176, "len_all_1": 73772, "len_abs": 1185, "len_title": 118, "len_sents": 121355, "len_sents_1": 32980, "n_sents": 1332, "n_sents_1": 278, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 69, "L_abs": 1215, "title": "Homomorphism Expressivity of Spectral Invariant Graph Neural Networks", "abs": "Graph spectra are an important class of structural features on graphs that have shown promising results in enhancing Graph Neural Networks (GNNs). Despite their widespread practical use, the theoretical understanding of the power of spectral invariants --- particularly their contribution to GNNs --- remains incomplete. In this paper, we address this fundamental question through the lens of homomorphism expressivity, providing a comprehensive and quantitative analysis of the expressive power of spectral invariants. Specifically, we prove that spectral invariant GNNs can homomorphism-count exactly a class of specific tree-like graphs which we refer to as \\emph{parallel trees}. We highlight the significance of this result in various contexts, including establishing a quantitative expressiveness hierarchy across different architectural variants, offering insights into the impact of GNN depth, and understanding the subgraph counting capabilities of spectral invariant GNNs. In particular, our results significantly extend \\citet{arvind2024hierarchy} and settle their open questions. Finally, we generalize our analysis to higher-order GNNs and answer an open question raised by \\citet{zhang2024expressive}.", "keywords": [ "Graph Neural Network", "Expressive Power", "Spectral Invariant", "Graph Homomorphism", "Weisfeiler-Lehman" ], "rating_list": [ 10, 6, 8, 6, 8 ], "soundness_list": [ 4, 3, 3, 2, 4 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 2, 2, 4 ], "confidence_list": [ 5, 1, 4, 4, 4 ] }, { "paper_id": "rdhB4ZvQR0", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Islem Rekik", "gender": "Female", "institution": "Imperial College London", "country": "", "position": "Associate Professor" }, { "name": "Pragya Singh", "gender": "Female", "institution": "Imperial College London", "country": "GB", "position": "MS student" } ], "n_formula": 68, "n_formula_1": 18, "n_ref_uni": 39, "n_ref": 46, "n_ref_all": 65, "n_fig": 9, "n_tab": 18, "L_tab": 10177, "n_element_tab": 636, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 2311, "n_element_tab_1": 161, "formula_len_all": 5115, "formula_len_all_1": 1111, "len_all": 215547, "len_all_1": 61233, "len_abs": 2331, "len_title": 123, "len_sents": 59894, "len_sents_1": 24982, "n_sents": 484, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1228, "title": "Rethinking Graph Super-Resolution: Dual Frameworks for Topological Fidelity", "abs": "Graph super-resolution is an underexplored yet highly relevant research direction that circumvents the need for costly and time-consuming data collection, preparation, and storage. This makes it especially desirable for resource-constrained fields such as the medical domain. Existing work on graph super-resolution leverages graph neural networks (GNNs) and achieves impressive results. However, we note two major limitations in the current model design: (1) It violates the underlying graph structure when increasing the number of nodes, and (2) it relies heavily on node representation learning, which has limited capacity to accurately model edges. To address these limitations, we propose two novel frameworks: (1) Bi-SR, which performs structure-aware node super-resolution, and (2) DEFEND, which focuses on edge representation learning for enhanced edge modeling. We supplement our work with rigorous theoretical analysis and conduct extensive experiments on simulated and real-world datasets covering diverse graph topologies and low-to-high resolution relationships. The results demonstrate substantial improvements across all experiments, highlighting the potential of both frameworks for graph super-resolution tasks.", "keywords": [ "Deep Graph Learning", "Graph Super-resolution", "Network Neuroscience" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 2, 1, 2, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "rdSVgnLHQB", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Ching-Chun Huang", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" }, { "name": "Hao-Chien Hsueh", "gender": "unknown", "institution": "National Chiao Tung University", "country": "", "position": "MS student" }, { "name": "Wen-Hsiao Peng", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" } ], "n_formula": 31, "n_formula_1": 16, "n_ref_uni": 23, "n_ref": 55, "n_ref_all": 90, "n_fig": 14, "n_tab": 4, "L_tab": 1664, "n_element_tab": 185, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 925, "n_element_tab_1": 125, "formula_len_all": 2786, "formula_len_all_1": 1120, "len_all": 122283, "len_all_1": 58075, "len_abs": 1369, "len_title": 111, "len_sents": 43870, "len_sents_1": 26011, "n_sents": 326, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1376, "title": "Warm Diffusion: Recipe for Blur-Noise Mixture Diffusion Models", "abs": "Diffusion probabilistic models have achieved remarkable success in generative tasks across diverse data types. While recent studies have explored alternative degradation processes beyond Gaussian noise, this paper bridges two key diffusion paradigms: hot diffusion, which relies entirely on noise, and cold diffusion, which uses only blurring without noise. We argue that hot diffusion fails to exploit the strong correlation between high-frequency image detail and low-frequency structures, leading to random behaviors in the early steps of generation. Conversely, while cold diffusion leverages image correlations for prediction, it neglects the role of noise (randomness) in shaping the data manifold, resulting in out-of-manifold issues and partially explaining its performance drop. To integrate both strengths, we propose Warm Diffusion, a unified Blur-Noise Mixture Diffusion Model (BNMD), to control blurring and noise jointly. Our divide-and-conquer strategy exploits the spectral dependency in images, simplifying score model estimation by disentangling the denoising and deblurring processes. We further analyze the Blur-to-Noise Ratio (BNR) using spectral analysis to investigate the trade-off between model learning dynamics and changes in the data manifold. Extensive experiments across benchmarks validate the effectiveness of our approach for image generation.", "keywords": [ "Diffusion probabilistic models", "Image generation" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "rdE9MCcNCz", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "anqi wu", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Chengrui Li", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Dieter Jaeger", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Full Professor" }, { "name": "Weihan Li", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Yule Wang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yunmiao Wang", "gender": "Female", "institution": "Emory University", "country": "", "position": "Postdoc" } ], "n_formula": 32, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 56, "n_ref_all": 101, "n_fig": 24, "n_tab": 4, "L_tab": 450, "n_element_tab": 39, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 236, "n_element_tab_1": 17, "formula_len_all": 2854, "formula_len_all_1": 759, "len_all": 127852, "len_all_1": 68181, "len_abs": 1069, "len_title": 101, "len_sents": 43592, "len_sents_1": 31221, "n_sents": 376, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1073, "title": "A Revisit of Total Correlation in Disentangled Variational Auto-Encoder with Partial Disentanglement", "abs": "A fully disentangled variational auto-encoder (VAE) aims to identify disentangled latent components from observations. However, enforcing full independence between all latent components may be too strict for certain datasets. In some cases, multiple factors may be entangled together in a non-separable manner, or a single independent semantic meaning could be represented by multiple latent components within a higher-dimensional manifold. To address such scenarios with greater flexibility, we propose the Partially Disentangled VAE (PDisVAE), which generalizes the total correlation (TC) term in fully disentangled VAEs to a partial correlation (PC) term. This framework can handle group-wise independence and can naturally reduce to either the standard VAE or the fully disentangled VAE. Validation through three synthetic experiments demonstrates the correctness and practicality of PDisVAE. When applied to real-world datasets, PDisVAE discovers valuable information that is difficult to find using fully disentangled VAEs, implying its versatility and effectiveness.", "keywords": [ "disentangling variational auto-encoder", "independent component analysis", "neural subspace", "neuroscience" ], "rating_list": [ 6, 1, 5, 6 ], "soundness_list": [ 2, 2, 4, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "rdAbEn5DZt", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Hansi Yang", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "James Kwok", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Hong Kong University of Science and Technology", "country": "HK", "position": "Professor" } ], "n_formula": 97, "n_formula_1": 10, "n_ref_uni": 35, "n_ref": 91, "n_ref_all": 116, "n_fig": 6, "n_tab": 5, "L_tab": 5059, "n_element_tab": 513, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2550, "n_element_tab_1": 254, "formula_len_all": 12874, "formula_len_all_1": 539, "len_all": 159047, "len_all_1": 59838, "len_abs": 977, "len_title": 133, "len_sents": 39950, "len_sents_1": 23943, "n_sents": 418, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 85, "L_abs": 980, "title": "Joint Gradient Balancing for Data Ordering in Finite-Sum Multi-Objective Optimization", "abs": "In finite-sum optimization problems, the sample orders for parameter updates can significantly influence the convergence rate of optimization algorithms. While numerous sample ordering techniques have been proposed in the context of single-objective optimization, the problem of sample ordering in finite-sum multi-objective optimization has not been thoroughly explored. To address this gap, we propose a sample ordering method called JoGBa, which finds the sample orders for multiple objectives by jointly performing online vector balancing on the gradients of all objectives. Our theoretical analysis demonstrates that this approach outperforms the standard baseline of random ordering and accelerates the convergence rate for the MGDA algorithm. Empirical evaluation across various datasets with different multi-objective optimization algorithms further demonstrates that JoGBa can achieve faster convergence and superior final performance than other data ordering strategies.", "keywords": [ "multi-objective optimization" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 4, 3 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "rcmhydaEJp", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Bryan Edward Kaiser", "gender": "Male", "institution": "Los Alamos National Laboratory", "country": "US", "position": "Researcher" }, { "name": "Kyle S. Hickmann", "gender": "Male", "institution": "Los Alamos National Laboratory", "country": "US", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 21, "n_ref_uni": 28, "n_ref": 38, "n_ref_all": 59, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 603, "formula_len_all_1": 632, "len_all": 70419, "len_all_1": 46039, "len_abs": 1255, "len_title": 83, "len_sents": 23467, "len_sents_1": 21871, "n_sents": 150, "n_sents_1": 138, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1260, "title": "Flow-based imputation of small data", "abs": "Many challenges in the physical sciences can be framed as small data problems, where theoretical progress is hindered by the sparsity, low-dimensionality, and/or limited sample size of available empirical data compared to a physical system\u2019s numerous dynamical degrees of freedom. Developing trustworthy imputation methods for these datasets holds immense scientific importance. Normalizing flows are a promising model choice for imputation due to their ability to explicitly estimate sample likelihoods. However, research has shown that normalizing flows are often unreliable for out-of-distribution (OOD) detection in high-dimensional settings, which undermines their trustworthiness for imputation tasks. In contrast, low-dimensional settings provide opportunities to tractably evaluate and mitigate likelihood estimation errors, revealing strategies to reduce or eliminate specific error modes. We focus on the most stringent assumption in normalizing flows: diffeomorphism between the target and base distributions. This assumption introduces two distinct error modes, which we identify and address through a simple and effective strategy. Our approach significantly enhances the trustworthiness of normalizing flows for imputation in small data problems.", "keywords": [ "Normalizing flows", "imputation", "diffeomorphism", "out of distribution detection" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 1, 2, 2 ], "presentation_list": [ 1, 3, 1 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "rcdR97P2Mp", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Daniel Eric Widdowson", "gender": "Male", "institution": "University of Liverpool", "country": "", "position": "PhD student" }, { "name": "Vitaliy Kurlin", "gender": "Not Specified", "institution": "University of Liverpool", "country": "GB", "position": "Full Professor" } ], "n_formula": 40, "n_formula_1": 14, "n_ref_uni": 29, "n_ref": 45, "n_ref_all": 77, "n_fig": 8, "n_tab": 18, "L_tab": 5333, "n_element_tab": 654, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5034, "formula_len_all_1": 1911, "len_all": 214689, "len_all_1": 50683, "len_abs": 954, "len_title": 129, "len_sents": 67341, "len_sents_1": 21798, "n_sents": 578, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 963, "title": "Towards continuous machine learning on periodic crystals by ultra-fast invariants", "abs": "Periodic point sets model all solid crystalline materials (crystals) whose atoms can be considered zero-sized points with or without atomic types.\n\nThis paper addresses the fundamental problem of checking whether claimed crystals are novel, not noisy perturbations of known materials obtained by unrealistic atomic replacements. Such near-duplicates have already skewed ground truth because past comparisons relied on discontinuous cells and symmetries. \n \nThe proposed Lipschitz continuity under noise is a new essential requirement for machine learning on any data objects that have ambiguous representations and live in continuous spaces. \n\nFor periodic point sets under isometry (any distance-preserving transformation), we designed the invariants that distinguish all known counter-examples to the completeness of past descriptors and detect thousands of (near-)duplicates in the world's five largest databases in a few minutes on a modest desktop computer.", "keywords": [ "scientific integrity", "periodic crystal", "isometry invariant", "continuous metric" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 3, 3, 4, 1 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 4, 3 ], "confidence_list": [ 2, 4, 2, 2 ] }, { "paper_id": "rcKzU0Vns0", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Leo Schwinn", "gender": "Male", "institution": "Technical University of Munich", "country": "DE", "position": "Postdoc" }, { "name": "Leonard Schenk", "gender": "Male", "institution": "Federal Agency for Disruptive Innovation", "country": "DE", "position": "Researcher" }, { "name": "Sebastian Schmidt", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Stephan G\u00fcnnemann", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Professor" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 41, "n_ref": 129, "n_ref_all": 175, "n_fig": 19, "n_tab": 2, "L_tab": 1281, "n_element_tab": 242, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 1493, "n_element_tab_1": 81, "formula_len_all": 677, "formula_len_all_1": 607, "len_all": 163026, "len_all_1": 65605, "len_abs": 976, "len_title": 124, "len_sents": 41999, "len_sents_1": 30589, "n_sents": 314, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 978, "title": "A Unified Approach Towards Active Learning and Out-of-Distribution Detection", "abs": "When applying deep learning models in real-world scenarios, active learning (AL) strategies are crucial for identifying label candidates from a nearly infinite amount of unlabeled data. In this context, robust out-of-distribution (OOD) detection mechanisms are essential for handling data outside the target distribution of the application. \nHowever, current works investigate both problems separately.\nIn this work, we introduce SISOM as the first unified solution for both AL and OOD detection.\nBy leveraging feature space distance metrics SISOM combines the strengths of the currently independent tasks to solve both effectively.\nWe conduct extensive experiments showing the problems arising when migrating between both tasks. In these evaluations SISOM underlined its effectiveness by achieving first place in two of the widely used OpenOOD benchmarks and second place in the remaining one. In AL, SISOM outperforms others and delivers top-1 performance in three benchmarks.", "keywords": [ "Out-of-Distribution Detection; Active Learning;" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 3, 1, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "rbnf7oe6JQ", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 4, "author_info_list": [ { "name": "Jing Wang", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "MS student" }, { "name": "Shang Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Wenji Fang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zhiyao Xie", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 7, "n_ref_uni": 35, "n_ref": 103, "n_ref_all": 128, "n_fig": 8, "n_tab": 9, "L_tab": 3310, "n_element_tab": 460, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 492, "n_element_tab_1": 42, "formula_len_all": 920, "formula_len_all_1": 524, "len_all": 198472, "len_all_1": 69878, "len_abs": 3086, "len_title": 128, "len_sents": 61252, "len_sents_1": 32808, "n_sents": 463, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 2110, "title": "CircuitFusion: Multimodal Circuit Representation Learning for Agile Chip Design", "abs": "The rapid advancements of AI rely on the support of integrated circuits (ICs). However, the growing complexity of digital ICs makes the traditional IC design process costly and time-consuming. In recent years, AI-assisted IC design methods have demonstrated great potential, but most methods are task-specific or focus solely on the circuit structure in graph format, overlooking other circuit modalities with rich functional information. In this paper, we introduce CircuitFusion, the first multimodal and implementation-aware circuit encoder. It encodes circuits into general representations that support different downstream circuit design tasks. To learn from circuits, we propose to fuse three circuit modalities: hardware code, structural graph, and functionality summary. More importantly, we identify four unique properties of circuits: parallel execution, functional equivalent transformation, multiple design stages, and circuit reusability. Based on these properties, we propose new strategies for both the development and application of CircuitFusion: 1) During circuit preprocessing, utilizing the parallel nature of circuits, we split each circuit into multiple sub-circuits based on sequential-element boundaries, each sub-circuit in three modalities. It enables fine-grained encoding at the sub-circuit level. 2) During CircuitFusion pre-training, we introduce three self-supervised tasks that utilize equivalent transformations both within and across modalities. We further utilize the multi-stage property of circuits to align representation with ultimate circuit implementation. 3) When applying CircuitFusion to downstream tasks, we propose a new retrieval-augmented inference method, which retrieves similar known circuits as a reference for predictions. It improves fine-tuning performance and even enables zero-shot inference. Evaluated on five different circuit design tasks, CircuitFusion consistently outperforms the state-of-the-art supervised method specifically developed for every single task, demonstrating its generalizability and ability to learn circuits' inherent properties.", "keywords": [ "Electronics Design Automation (EDA)", "circuit representation learning", "multimodal learning", "self-supervised learning", "hardware circuits" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 1, 5 ] }, { "paper_id": "rbHOLX8OWh", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Austin Meek", "gender": "unknown", "institution": "University of Delaware", "country": "US", "position": "PhD student" }, { "name": "David Krueger", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Assistant Professor" }, { "name": "Fazl Barez", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Michael Lan", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Researcher" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 42, "n_ref": 79, "n_ref_all": 125, "n_fig": 15, "n_tab": 9, "L_tab": 2277, "n_element_tab": 222, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 688, "n_element_tab_1": 71, "formula_len_all": 37, "formula_len_all_1": 37, "len_all": 185045, "len_all_1": 67806, "len_abs": 1173, "len_title": 128, "len_sents": 47824, "len_sents_1": 33539, "n_sents": 343, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1177, "title": "Sparse Autoencoders Reveal Universal Feature Spaces Across Large Language Models", "abs": "We investigate feature universality in large language models (LLMs), a research field that aims to understand how different models similarly represent concepts in the latent spaces of their intermediate layers. Demonstrating feature universality allows discoveries about latent representations to generalize across several models. However, comparing features across LLMs is challenging due to polysemanticity, in which individual neurons often correspond to multiple features rather than distinct ones. This makes it difficult to disentangle and match features across different models. To address this issue, we employ a method known as dictionary learning by using sparse autoencoders (SAEs) to transform LLM activations into more interpretable spaces spanned by neurons corresponding to individual features. After matching feature neurons across models via activation correlation, we apply representational space similarity metrics like Singular Value Canonical Correlation Analysis to analyze these SAE features across different LLMs. Our experiments reveal significant similarities in SAE feature spaces across various LLMs, providing new evidence for feature universality.", "keywords": [ "explanation faithfulness", "feature attribution", "knowledge tracing/discovering/inducing" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "rb93dP976j", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Xiaowu Dai", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "Zifei Guo", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 23, "n_formula_1": 21, "n_ref_uni": 19, "n_ref": 24, "n_ref_all": 31, "n_fig": 5, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1437, "formula_len_all_1": 1348, "len_all": 94783, "len_all_1": 62916, "len_abs": 1056, "len_title": 97, "len_sents": 35072, "len_sents_1": 29896, "n_sents": 274, "n_sents_1": 235, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1058, "title": "Dynamic Matching Utilizing Latent Factor Modeling", "abs": "This paper investigates the supply-demand matching problem on dynamic platforms, focusing on optimizing matching strategies by learning workers' attributes when their types are uncertain and constantly changing. To address this problem, we introduce a latent factor model and a multi-centroid grouping penalty mechanism to predict latent factors of workers and perform dynamic matching. Our approach operates in two stages: the first stage fits latent feature vectors for workers and jobs and groups them using historical data; the second stage utilizes these latent features for dynamic matching. Our research demonstrates that the introduced model can adapt to the dynamic changes of the platform with good predictive consistency and group robustness, and improves overall operational benefit through continuous optimization of matching results. We provide simulation experiments and a real case study using kidney exchange data and compare our model with a point process model to show that our approach performs well on dynamic platform matching problems.", "keywords": [ "Dynamic matching", "Learning", "Two-sided market", "Bi-clustering", "Label aggregation" ], "rating_list": [ 5, 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 1, 3 ], "presentation_list": [ 3, 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 1, 2 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "razAcpFapu", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Dong Han", "gender": "unknown", "institution": "Friedrich-Schiller Universit\u00e4t Jena", "country": "DE", "position": "PhD student" }, { "name": "Joachim Denzler", "gender": "Male", "institution": "Friedrich-Schiller-University Jena", "country": "DE", "position": "Professor" }, { "name": "Yong Li", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "DE", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 21, "n_ref": 58, "n_ref_all": 77, "n_fig": 6, "n_tab": 9, "L_tab": 1425, "n_element_tab": 252, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1295, "n_element_tab_1": 234, "formula_len_all": 279, "formula_len_all_1": 201, "len_all": 103550, "len_all_1": 54702, "len_abs": 1403, "len_title": 65, "len_sents": 31214, "len_sents_1": 22956, "n_sents": 231, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 17, "L_abs": 1409, "title": "KAN See Your Face", "abs": "With the advancement of face reconstruction (FR) systems, privacy-preserving face recognition (PPFR) has gained popularity for its secure face recognition, enhanced facial privacy protection, and robustness to various attacks.\nBesides, specific models and algorithms are proposed for face embedding protection by mapping embeddings to a secure space.\nHowever, there is a lack of studies on investigating and evaluating the possibility of extracting face images from embeddings of those systems, especially for PPFR.\nIn this work, we introduce the first approach to exploit Kolmogorov-Arnold Network (KAN) for conducting embedding-to-face attacks against state-of-the-art (SOTA) FR and PPFR systems.\nFace embedding mapping (FEM) models are proposed to learn the distribution mapping relation between the embeddings from the initial domain and target domain.\nIn comparison with Multi-Layer Perceptrons (MLP), we provide two variants, FEM-KAN and FEM-MLP, for efficient non-linear embedding-to-embedding mapping in order to reconstruct realistic face images from the corresponding face embedding.\nTo verify our methods, we conduct extensive experiments with various PPFR and FR models.\nWe also measure reconstructed face images with different metrics to evaluate the image quality.\nThrough comprehensive experiments, we demonstrate the effectiveness of FEMs in accurate embedding mapping and face reconstruction.", "keywords": [ "Face Reconstruction", "Privacy-Preserving Face Recognition", "Diffusion", "Face Embedding Inversion" ], "rating_list": [ 3, 5, 1, 3, 3 ], "soundness_list": [ 1, 2, 2, 2, 4 ], "presentation_list": [ 1, 2, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 2, 4, 5 ] }, { "paper_id": "rawj2PdHBq", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Chen Jin", "gender": "unknown", "institution": "Astrazeneca", "country": "GB", "position": "Researcher" }, { "name": "Fariba Yousefi", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "machine learning engineer" }, { "name": "Haozhe Wang", "gender": "Male", "institution": "INF", "country": "CN", "position": "Researcher" }, { "name": "Rossella Arcucci", "gender": "Female", "institution": "Imperial College London ", "country": "", "position": "Senior Lecturer" }, { "name": "Yinda Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zhongwei Wan", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "che liu", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Nikolay Burlutskiy", "gender": "Male", "institution": "AstraZeneca", "country": "GB", "position": "Director of AI" }, { "name": "Talha Qaiser", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 54, "n_ref": 145, "n_ref_all": 157, "n_fig": 5, "n_tab": 8, "L_tab": 2163, "n_element_tab": 211, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1960, "n_element_tab_1": 229, "formula_len_all": 330, "formula_len_all_1": 330, "len_all": 183204, "len_all_1": 74945, "len_abs": 1496, "len_title": 124, "len_sents": 38094, "len_sents_1": 33494, "n_sents": 288, "n_sents_1": 249, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1581, "title": "Can Medical Vision-Language Pre-training Succeed with Purely Synthetic Data?", "abs": "Medical Vision-Language Pre-training (MedVLP) has made significant progress in enabling zero-shot tasks for medical image understanding. However, training MedVLP models typically requires large-scale datasets with paired, high-quality image-text data, which are scarce in the medical domain. Recent advancements in Large Language Models (LLMs) and diffusion models have made it possible to generate large-scale synthetic image-text pairs. This raises the question: _**Can MedVLP succeed using purely synthetic data?**_ To address this, we use off-the-shelf generative models to create synthetic radiology reports and paired Chest X-ray (CXR) images, and propose an automated pipeline to build a diverse, high-quality synthetic dataset, enabling a rigorous study that isolates model and training settings, focusing entirely from the data perspective.\n\nOur results show that MedVLP models trained _exclusively on synthetic data_ outperform those trained on real data by **3.8%** in averaged AUC on zero-shot classification. Moreover, using a combination of synthetic and real data leads to a further improvement of **9.07%**. Additionally, MedVLP models trained on synthetic or mixed data consistently outperform those trained on real data in zero-shot grounding, as well as in fine-tuned classification and segmentation tasks.\n\nOur analysis suggests MedVLP trained on well-designed synthetic data can outperform models trained on real datasets, which may be limited by low-quality samples and long-tailed distributions[^1].\n\n[^1]: All data and code will be released upon acceptance.", "keywords": [ "Medical Vision-Language Pre-training", "Synthetic Multimodal Data" ], "rating_list": [ 5, 5, 8 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "rapXZIfwbX", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Aditya Cowsik", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Alex Infanger", "gender": "Not Specified", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Kfir Dolev", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 59, "n_formula_1": 15, "n_ref_uni": 31, "n_ref": 42, "n_ref_all": 84, "n_fig": 12, "n_tab": 2, "L_tab": 560, "n_element_tab": 93, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 176, "n_element_tab_1": 9, "formula_len_all": 3970, "formula_len_all_1": 761, "len_all": 148337, "len_all_1": 57630, "len_abs": 1488, "len_title": 82, "len_sents": 45037, "len_sents_1": 26386, "n_sents": 420, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1382, "title": "The Persian Rug: solving toy models of superposition using large-scale symmetries", "abs": "We present a complete mechanistic description of the algorithm learned by a minimal non-linear sparse data autoencoder in the limit of large input dimension. The model, originally presented in \\cite{elhage2022superposition}, compresses sparse data vectors through a linear layer and decompresses using another linear layer followed by a ReLU activation. We notice that when the data is permutation symmetric (no input feature is privileged) large models reliably learn an algorithm that is sensitive to individual weights only through their large-scale statistics. For these models, the loss function becomes analytically tractable. Using this understanding, we give explicit upper bounds on the loss, which show that the model is near-optimal among recently proposed architectures. In particular, changes to the elementwise activation function or the addition of gating can at best improve its performance by a constant factor. Finally, we forward-engineer a model with the requisite symmetries and show that its loss precisely matches that of the trained models. Unlike the trained model weights, the minimal randomness in the artificial weights results in miraculous fractal structures resembling a Persian rug, to which the algorithm is oblivious. Our work contributes to neural network interpretability by introducing techniques for understanding the structure of autoencoders.", "keywords": [ "mechanistic interpretability", "autoencoder", "sparse data", "superposition" ], "rating_list": [ 6, 3, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 4, 3 ], "presentation_list": [ 2, 2, 2, 4, 3 ], "contribution_list": [ 3, 1, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 3, 3 ] }, { "paper_id": "rakhNY32vw", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Xiangyang Ji", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yuan Zhou", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Zihan Zhang", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Postdoc" } ], "n_formula": 144, "n_formula_1": 22, "n_ref_uni": 19, "n_ref": 60, "n_ref_all": 66, "n_fig": 0, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 244, "n_element_tab_1": 33, "formula_len_all": 16320, "formula_len_all_1": 1306, "len_all": 184570, "len_all_1": 73694, "len_abs": 658, "len_title": 120, "len_sents": 57941, "len_sents_1": 33607, "n_sents": 619, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 676, "title": "Almost Optimal Batch-Regret Tradeoff for Batch Linear Contextual Bandits", "abs": "We study the optimal batch-regret tradeoff for batch linear contextual bandits. For this problem, we design batch learning algorithms and prove that they achieve the optimal regret bounds (up to logarithmic factors) for any batch number $M$, number of actions $K$, time horizon $T$, and dimension $d$. Therefore, we establish the \\emph{full-parameter-range} (almost) optimal batch-regret tradeoff for the batch linear contextual bandit problem. \n\nAlong our analysis, we also prove a new matrix concentration inequality with dependence on their dynamic upper bounds, which, to the best of our knowledge, is the first of its kind in literature and maybe of independent interest.", "keywords": [ "contextual linear bandit", "online batch learning" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 4, 3, 4, 3 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "raUnLe0Z04", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Feng Liu", "gender": "unknown", "institution": "Adobe Systems", "country": "US", "position": "Principal Research Scientist" }, { "name": "Jeremy Vonderfecht", "gender": "unknown", "institution": "Portland State University", "country": "", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 8, "n_ref_uni": 29, "n_ref": 65, "n_ref_all": 81, "n_fig": 9, "n_tab": 1, "L_tab": 482, "n_element_tab": 51, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 965, "n_element_tab_1": 75, "formula_len_all": 626, "formula_len_all_1": 447, "len_all": 111873, "len_all_1": 52766, "len_abs": 774, "len_title": 98, "len_sents": 34936, "len_sents_1": 23317, "n_sents": 306, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 768, "title": "Lossy Compression with Pretrained Diffusion Models", "abs": "We apply Theis et al. (2022)'s DiffC algorithm to Stable Diffusion 1.5, 2.1, XL, and and Flux-dev, and demonstrate that these pretrained models are remarkably capable lossy image compressors. A principled algorithm for compression using pretrained diffusion models has been understood since at least 2020 (Ho et al.), but challenges in reverse-channel coding have prevented such algorithms from ever being fully implemented. We introduce simple workarounds that lead to the first complete implementation of DiffC, which is capable of compressing and decompressing images using Stable Diffusion in under 10 seconds. Despite requiring no additional training, our method is competitive with other state-of-the-art generative compression methods at low ultra-low bitrates.", "keywords": [ "Compression", "Stable Diffusion", "Flux", "DDPM", "DiffC" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 2 ] }, { "paper_id": "ra7Nl9wUlF", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Chenhui Gou", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Ethan Smith", "gender": "Male", "institution": "Leonardo.Ai", "country": "AU", "position": "Researcher" }, { "name": "Lingqiao Liu", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Associate Professor" }, { "name": "Yuxuan Ding", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "CN", "position": "Researcher" }, { "name": "Zicheng Duan", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "PhD student" }, { "name": "Ziqin Zhou", "gender": "Female", "institution": "University of Adelaide", "country": "AU", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 21, "n_ref": 63, "n_ref_all": 82, "n_fig": 7, "n_tab": 6, "L_tab": 1168, "n_element_tab": 139, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 445, "n_element_tab_1": 73, "formula_len_all": 249, "formula_len_all_1": 249, "len_all": 100942, "len_all_1": 54692, "len_abs": 199, "len_title": 113, "len_sents": 27890, "len_sents_1": 25619, "n_sents": 184, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 112, "L_abs": 1361, "title": "EZIGen: Enhancing zero-shot subject-driven image generation with precise subject encoding and decoupled guidance", "abs": "Zero-shot subject-driven image generation aims to produce images that incorporate a subject from a given example image. The challenge lies in preserving the subject's identity while aligning with the text prompt which often requires modifying certain aspects of the subject's appearance. Despite advancements in diffusion model based methods, existing approaches still struggle to balance identity preservation with text prompt alignment. In this study, we conducted an in-depth investigation into this issue and uncovered key insights for achieving effective identity preservation while maintaining a strong balance. Our key findings include: (1) the design of the subject image encoder significantly impacts identity preservation quality, and (2) separating text and subject guidance is crucial for both text alignment and identity preservation. Building on these insights, we introduce a new approach called EZIGen, which employs two main strategies: a carefully crafted subject image Encoder based on the pretrained UNet of the Stable Diffusion model to ensure high-quality identity transfer, following a process that decouples the guidance stages and iteratively refines the initial image layout. Through these strategies, EZIGen achieves state-of-the-art results on multiple subject-driven benchmarks with a unified model and 100 times less training data.", "keywords": [ "Diffusion models", "Subject driven image generation", "Personalized image generation" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "rZzcaduYU1", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Michael John Hutchinson", "gender": "Male", "institution": "Isomorphic Labs", "country": "GB", "position": "Researcher" }, { "name": "Scott le Roux", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Yee Whye Teh", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" } ], "n_formula": 46, "n_formula_1": 28, "n_ref_uni": 30, "n_ref": 69, "n_ref_all": 77, "n_fig": 1, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5228, "formula_len_all_1": 2545, "len_all": 135010, "len_all_1": 57415, "len_abs": 1131, "len_title": 28, "len_sents": 39823, "len_sents_1": 24355, "n_sents": 345, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 28, "L_abs": 1164, "title": "Score-Based Neural Processes", "abs": "Neural processes (NP) are a flexible class of models that generate stochastic\nprocesses by operating on finite-dimensional marginal distributions. NPs are\ndesigned to maintain exchangeability and marginal consistency, which are necessary\nto define a valid stochastic process. However, NP variants can come with drawbacks\nsuch as limited expressivity, uncorrelated samples, and consistency sacrifices. To address the issues of previous NPs, we introduce score-based neural processes, \\emph{scoreNP}, which incorporate a score-based generative model within the neural process paradigm. This score-based approach enhances expressivity, allowing the model to capture complex non-Gaussian distributions of functions, generate correlated samples, and maintain marginal consistency. Previously, no NP variant has been able to maintain conditional consistency. We show that using \\emph{guidance} methods from conditional diffusion sampling, \\emph{scoreNP} is the first NP is able to satisfy conditional consistency. Empirically, \\emph{scoreNP} performs well qualitatively and quantitatively well across a range of unconditional and conditional functional generation tasks.", "keywords": [ "Neural Processes", "Diffusion Models", "Generative Models", "Stochastic Processes" ], "rating_list": [ 1, 3, 3, 5 ], "soundness_list": [ 1, 1, 3, 2 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 2 ] }, { "paper_id": "rZxwa8JkJW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Alexander Black", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "PhD student" }, { "name": "Jing Shi", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "John Philip Collomosse", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Principal Researcher" }, { "name": "Yifei Fan", "gender": "unknown", "institution": "Adobe Systems", "country": "US", "position": "Machine Learning Engineer" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 7, "n_ref": 13, "n_ref_all": 28, "n_fig": 6, "n_tab": 3, "L_tab": 2745, "n_element_tab": 201, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 243, "formula_len_all_1": 242, "len_all": 118290, "len_all_1": 41396, "len_abs": 845, "len_title": 121, "len_sents": 24809, "len_sents_1": 19395, "n_sents": 178, "n_sents_1": 145, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 765, "title": "Fusing Visual and Textual Cues for Sequential Image Difference Captioning", "abs": "We present FVTC - a technique for image difference captioning that is able to benefit from additional visual and/or textual inputs. FVTC is able to succinctly summarize multiple manipulations that were applied to an image in a sequence. Optionally, it can take several intermediate thumbnails of the image editing sequence as input, as well as coarse machine-generated annotations of the individual manipulations. We demonstrate that the presence of intermediate images and/or auxiliary textual information improves the model's captioning performance. To train FVTC, we introduce METS - a new dataset of image editing sequences, with textual machine annotations of each editorial step and human edit summarization captions after the 5th, 10th and 15th manipulation.", "keywords": [ "Visual Language Models", "Change Summarization", "Multi-modal Analysis" ], "rating_list": [ 8, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "rZmQ2z7MPA", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Amisha Srivastava", "gender": "Female", "institution": "University of Texas at Dallas", "country": "", "position": "PhD student" }, { "name": "Anand Menon", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Arnab Raha", "gender": "Male", "institution": "Purdue University", "country": "", "position": "PhD student" }, { "name": "Deepak Mathaikutty", "gender": "Male", "institution": "Intel", "country": "US", "position": "Researcher" }, { "name": "Gabriel Theodor Sonnenschein", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "Undergrad student" }, { "name": "Kanad Basu", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "Assistant Professor" }, { "name": "Samit S Miftah", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Shamik Kundu", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Souvik Kundu", "gender": "Male", "institution": "Intel", "country": "US", "position": "Researcher" }, { "name": "Suvadeep Banerjee", "gender": "Male", "institution": "Intel", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 22, "n_ref": 32, "n_ref_all": 46, "n_fig": 6, "n_tab": 10, "L_tab": 9538, "n_element_tab": 1072, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 5413, "n_element_tab_1": 310, "formula_len_all": 900, "formula_len_all_1": 262, "len_all": 175961, "len_all_1": 65063, "len_abs": 7028, "len_title": 130, "len_sents": 57314, "len_sents_1": 28138, "n_sents": 387, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 2499, "title": "VERT: A SystemVerilog Assertion Dataset to Improve Hardware Verification with LLMs", "abs": "Hardware verification is a critical step in the modern System-on-Chip (SoC) design cycle, consuming approximately 70% of development time. SystemVerilog assertions are pivotal in the verification process, ensuring that designs function as intended. However, existing industrial practices rely on manual assertion generation, which becomes increasingly untenable as hardware systems become complex. Recent research has explored the potential of Large Language Models (LLMs) to automate the hardware verification process, reducing human intervention. Despite this, State-of-the-Art (SOTA) proprietary models, such as OpenAI's GPT-4o, have shown limitations in generating accurate assertions and require costly licenses and restricted usage. While smaller, open-source LLMs offer a more accessible option, they require fine-tuning to handle the complexities of the source code and generate accurate assertions. This highlights the need for a dataset that enables these models to achieve superior performance compared to SOTA LLMs. To this end, we present VERT, a dataset designed to improve the generation of SystemVerilog assertions using LLMs. Our dataset empowers researchers and hardware corporations to fine-tune smaller, open-source LLMs, surpassing larger proprietary models such as GPT-4 in accuracy and efficiency. Furthermore, VERT eliminates the need for expensive licenses and ensures data privacy through local fine-tuning, providing a scalable, cost-effective solution for automated hardware verification. To curate the dataset, we systematically compile and augment variables from open-source hardware description languages (HDL), generating conditions to create synthetic code snippets paired with corresponding assertions. We show that smaller, open-source LLMs, such as Deepseek Coder 6.7B and Llama 3.1 8B, when fine-tuned on VERT, outperform OpenAI's GPT-4o in assertion generation. The assertions generated by the fine-tuned models are evaluated on industry-standard platforms, including OpenTitan, CVA6, and Pulpissimo SoCs, demonstrating up to a 96.88% improvement in both functional and syntactical correctness compared to the base models and up to 24.14% when compared to GPT-4o. This demonstrates the prowess of VERT in enabling researchers to potentially reduce the overhead and human error associated with manual assertion generation, offering a scalable solution for industry-grade hardware designs. The dataset is available at https://anonymous.4open.science/r/VERT-4D6D/.", "keywords": [ "Hardware Verification", "Large Language Models", "SystemVerilog" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "rYU6xsZkfk", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Alessandro Trenta", "gender": "Male", "institution": "Universita' di Pisa, University of Pisa", "country": "IT", "position": "PhD student" }, { "name": "Andrea Cossu", "gender": "Male", "institution": "University of Pisa", "country": "IT", "position": "Assistant Professor" }, { "name": "Davide Bacciu", "gender": "Male", "institution": "University of Pisa", "country": "IT", "position": "Full Professor" } ], "n_formula": 45, "n_formula_1": 6, "n_ref_uni": 37, "n_ref": 90, "n_ref_all": 133, "n_fig": 23, "n_tab": 11, "L_tab": 2949, "n_element_tab": 311, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 1390, "n_element_tab_1": 155, "formula_len_all": 3066, "formula_len_all_1": 892, "len_all": 208344, "len_all_1": 71634, "len_abs": 1227, "len_title": 105, "len_sents": 65293, "len_sents_1": 31092, "n_sents": 576, "n_sents_1": 267, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1235, "title": "Derivatives Are All You Need For Learning Physical Models", "abs": "Physics-Informed Neural Networks (PINNs) explicitly incorporate Partial Differential Equations (PDEs) into the loss function, thus learning representations that are inherently consistent with the physical system. \nWe claim that it is possible to learn physically consistent models without explicit knowledge about the underlying equations. We propose Derivative Learning (DERL) to model a physical system by learning its partial derivatives, as they contain all the necessary information to determine the system's dynamics. Like in PINNs, we also train the learning model on the initial and boundary conditions of the system. \nWe provide theoretical guarantees that our approach learns the true solution and is consistent with the underlying physical laws, even when using empirical derivatives. DERL outperforms PINNs and other state-of-the-art approaches in tasks ranging from simple dynamical systems to PDEs. Finally, we show that distilling the derivatives enables the transfer of physical information from one model to another. Distillation of higher-order derivatives improves physical consistency. Ultimately, learning and distilling the derivatives of physical systems turns out to be a powerful tool to learn physical models.", "keywords": [ "Physics-informed neural networks", "Physics-inspired neural networks", "Dynamical systems", "Learning physics", "Physical systems" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "rYTlmSlxia", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Ao Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Huanhuan Chen", "gender": "Male", "institution": "University of Science and Technology of China, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Shikang Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Xinyu Yan", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Xiren Zhou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 21, "n_ref_uni": 28, "n_ref": 37, "n_ref_all": 62, "n_fig": 6, "n_tab": 6, "L_tab": 1545, "n_element_tab": 103, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1079, "n_element_tab_1": 79, "formula_len_all": 748, "formula_len_all_1": 766, "len_all": 125876, "len_all_1": 67071, "len_abs": 1607, "len_title": 123, "len_sents": 43512, "len_sents_1": 31902, "n_sents": 285, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1613, "title": "From Data to Model: Anomaly Detection of 3D GPR Data in CuDeRes Model Space", "abs": "Ground Penetrating Radar (GPR) offers in-depth subterranean insights, yet subsurface anomaly detection in GPR data remains challenging due to limited training data, typically confined to some normal data samples free from any subsurface structures or anomalies, and the variability of subsurface conditions. In response, this paper introduces practical and accurate subsurface anomaly detection within the Cubic Decay Reservoir Network (CuDeRes) model space. Our approach employs commonly available normal GPR data, segmented into blocks. Each data block is independently fitted using the introduced CuDeRes, which incorporates three reservoirs with spatial decay to adequately capture the data-inherent multi-directional dynamics, resulting in a compact fitted readout model. Representing each data block with the fitted model, together with the distance measurement between models, the original GPR data blocks are mapped into the CuDeRes model space, and the fitted models are collected into a \"Model Depot\". For subsequent anomaly detection in newly collected GPR data, the same segmentation and CuDeRes fitting approaches are applied, where the data blocks are represented by fitted models for comparative assessment against the model depot. Anomalies are detected through model dissimilarities, and subsequently clustered within the CuDeRes model space, allowing us to accurately identify the data blocks with potential subsurface anomalies and ascertain their anomaly types. Experiments on real-world GPR data demonstrate the practical effectiveness of our approach, notably using only limited normal data.", "keywords": [ "Cubic Decay Reservoir Network (CuDeRes)", "Ground Penetrating Radar (GPR)", "Learning in the Model Space", "Reservoir Computing" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "rY8xdjrANt", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Changbo Zhu", "gender": "unknown", "institution": "University of Notre Dame", "country": "US", "position": "Assistant Professor" }, { "name": "Kaheon Kim", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Rentian Yao", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Postdoc" }, { "name": "Xiaohui Chen", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "Associate Professor" } ], "n_formula": 59, "n_formula_1": 25, "n_ref_uni": 40, "n_ref": 60, "n_ref_all": 71, "n_fig": 3, "n_tab": 1, "L_tab": 224, "n_element_tab": 19, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8288, "formula_len_all_1": 2191, "len_all": 142673, "len_all_1": 67558, "len_abs": 1313, "len_title": 120, "len_sents": 38154, "len_sents_1": 27262, "n_sents": 319, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1360, "title": "OPTIMAL TRANSPORT BARYCENTER VIA NONCONVEX CONCAVE MINIMAX OPTIMIZATION", "abs": "The optimal transport barycenter (a.k.a. Wasserstein barycenter) is a fundamental notion of averaging that extends from the Euclidean space to the Wasserstein space of probability distributions. Computation of the \\emph{unregularized} barycenter for discretized probability distributions on point clouds is a challenging task when the domain dimension $d > 1$. Most practical algorithms for approximating the barycenter problem are based on entropic regularization. In this paper, we introduce a nearly linear time $O(m \\log{m})$ and linear space complexity $O(m)$ primal-dual algorithm, the Wasserstein-Descent $\\dot{\\mathbb{H}}^1$-Ascent (WDHA) algorithm, for computing the exact barycenter when the input probability density functions are discretized on an $m$-point grid. The key success of the WDHA algorithm hinges on alternating between two different yet closely related Wasserstein and Sobolev optimization geometries for the primal barycenter and dual Kantorovich potential subproblems. Under reasonable assumptions, we establish the convergence rate and iteration complexity of WDHA to its stationary point when the step size is appropriately chosen. Superior computational efficacy, scalability, and accuracy over the existing Sinkhorn-type algorithms are demonstrated on high-resolution (e.g., $1024 \\times 1024$ images) 2D synthetic and real data.", "keywords": [ "Wasserstein barycenter; optimal transport; nonconvex-concave minimax optimization; convergence rate;" ], "rating_list": [ 5, 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "rXrYdOtBfs", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Chengjun Wu", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "Undergrad student" }, { "name": "Chengyu Duan", "gender": "Female", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Undergrad student" }, { "name": "Frederic Sala", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Samuel Guo", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Satya Sai Srinath Namburi GNVV", "gender": "Male", "institution": "GE HealthCare", "country": "US", "position": "Researcher" }, { "name": "Sonia Cromp", "gender": "Female", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Nicholas Carl Roberts", "gender": "Male", "institution": "University of Wisconsin-Madison", "country": "US", "position": "PhD student" }, { "name": "ZHIQI GAO", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 110, "n_ref_all": 137, "n_fig": 5, "n_tab": 9, "L_tab": 2245, "n_element_tab": 246, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1070, "n_element_tab_1": 145, "formula_len_all": 1590, "formula_len_all_1": 860, "len_all": 203239, "len_all_1": 62503, "len_abs": 1425, "len_title": 82, "len_sents": 60441, "len_sents_1": 27361, "n_sents": 491, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 34, "L_abs": 1445, "title": "Pretrained Hybrids with MAD Skills", "abs": "While Transformers underpin modern large language models (LMs), a growing list of alternative architectures with new capabilities, promises, and tradeoffs is emerging. This makes choosing the right LM architecture challenging. Recently proposed *hybrid architectures* seek a best-of-all-worlds approach that reaps the benefits of all architectures. Hybrid design is difficult for two reasons: it requires manual expert-driven search, and new hybrids must be trained from scratch. We propose **Manticore**, a framework that addresses these challenges by *automating the design of hybrid architectures* while reusing pretrained models to create *pretrained* hybrids. Our approach augments ideas from differentiable Neural Architecture Search (NAS) by incorporating simple projectors that translate features between pretrained blocks from different architectures. We then fine-tune hybrids that combine pretrained models from different architecture families---such as the GPT series and Mamba---end-to-end. With Manticore, we enable LM selection without training multiple models, the construction of pretrained hybrids from existing pretrained models, and the ability to *program* pretrained hybrids to have certain capabilities. Manticore hybrids match existing manually-designed hybrids, achieve strong performance on the Long Range Arena benchmark, and improve on pretrained transformers and state space models on various natural language tasks.", "keywords": [ "hybrid architectures", "large language models", "transformers", "state space models", "model merging", "neural architecture search", "mechanistic search" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "rXNGpyxsLQ", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Cuiping Li", "gender": "Female", "institution": "Renmin University of China", "country": "", "position": "Full Professor" }, { "name": "Hong Chen", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Jing Zhang", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Yuxuan Hu", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "PhD student" }, { "name": "Zhe Zhao", "gender": "unknown", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 40, "n_ref": 52, "n_ref_all": 67, "n_fig": 2, "n_tab": 4, "L_tab": 2365, "n_element_tab": 396, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1723, "n_element_tab_1": 106, "formula_len_all": 1398, "formula_len_all_1": 1088, "len_all": 179851, "len_all_1": 57041, "len_abs": 6281, "len_title": 128, "len_sents": 30461, "len_sents_1": 23490, "n_sents": 220, "n_sents_1": 153, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1531, "title": "SP-LoRA: Sparsity-Preserved Low-Rank Adaptation for Sparse Large Language Model", "abs": "Large Language Models (LLMs) have shown remarkable performance in various natural language processing tasks but suffer from substantial hardware resource requirements and inference latency issues due to their vast parameter counts. To mitigate these challenges, several post-training pruning techniques such as SparseGPT, Wanda, and RIA have been developed to reduce parameter sizes. However, these methods often result in performance gaps, particularly for smaller models, and lack efficient fine-tuning strategies that preserve sparsity. This paper introduces SP-LoRA, a novel approach that combines the benefits of low-rank adaptation (LoRA) with the efficiency of sparse models. SP-LoRA addresses the issue of density reversion when merging LoRA adapters with sparse matrices through the introduction of a mask matrix $\\mathcal{M}$, ensuring sparsity is maintained. Furthermore, since maintaining sparsity tends to result in a large memory overhead, we propose gradient checkpointing and memory reuse techniques to optimize GPU memory usage during fine-tuning, achieving comparable efficiency to standard LoRA. Through extensive evaluations on pruned LLMs using methods like Wanda and SparseGPT, followed by fine-tuning with SP-LoRA, we demonstrate its effectiveness in both zero-shot scenarios and domain-specific tasks. Our key contributions include a parameter-efficient fine-tuning method for sparse LLMs, an optimized algorithm for reduced GPU memory overhead, and comprehensive empirical validation across diverse models.", "keywords": [ "sparsity", "parameter efficient fine-tuning", "low rank adaptation", "large language model" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "rWui9vLhOc", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Rachel Teo", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Tan Minh Nguyen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 33, "n_formula_1": 16, "n_ref_uni": 71, "n_ref": 123, "n_ref_all": 154, "n_fig": 3, "n_tab": 13, "L_tab": 4904, "n_element_tab": 746, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2543, "n_element_tab_1": 377, "formula_len_all": 17834, "formula_len_all_1": 1416, "len_all": 263388, "len_all_1": 80820, "len_abs": 2834, "len_title": 117, "len_sents": 54390, "len_sents_1": 32409, "n_sents": 421, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 2078, "title": "MoLEx: Mixture of Layer Experts for Fine-tuning with Sparse Upcycling", "abs": "Large-scale pre-training of deep models, followed by fine-tuning them to adapt to downstream tasks, has become the cornerstone of natural language processing (NLP). The prevalence of vast corpses of data coupled with computational resources has led to large models with a considerable number of parameters. While the massive size of these models has led to remarkable success in many NLP tasks, a detriment is the expense required to retrain all the base model's parameters for the adaptation to each task or domain. Parameter Efficient Fine-Tuning (PEFT) provides a highly effective solution for this challenge by minimizing the number of parameters required to be trained in adjusting to the new task while maintaining the quality of the model. While existing methods have achieved impressive results, they mainly focus on adapting a subset of parameters using adapters, weight reparameterization, and prompt engineering. In this paper, we study layers as extractors of different types of linguistic information that are valuable when used in conjunction with each other. We then propose the Mixture of Layer Experts (MoLEx), a novel Sparse Mixture of Experts (SMoE) whose experts are layers in the pre-trained model. In particular, MoLEx is applied at each layer of the pre-trained model. It performs a conditional computation of a mixture of layers during fine-tuning to provide the model with more structural knowledge about the data. By providing an avenue for information exchange between layers, MoLEx enables the model to make a more well-informed prediction for the downstream task, leading to better fine-tuning results with the same number of effective parameters. As experts can be processed in parallel, MoLEx introduces minimal additional computational overhead. We empirically corroborate the advantages of MoLEx when combined with popular PEFT baseline methods on a variety of downstream fine-tuning tasks, including the popular GLUE benchmark for natural language understanding (NLU) as well as the natural language generation (NLG) End-to-End Challenge (E2E).", "keywords": [ "Parameter efficient fine-tuning", "mixture of experts", "sparse upcycling" ], "rating_list": [ 5, 6, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "rWjZWHYPcz", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Chun-Fu Chen", "gender": "Male", "institution": "JPMorganChase, GTAR", "country": "US", "position": "Executive Director" }, { "name": "Eric Lei", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Hsiang Hsu", "gender": "Male", "institution": "JP Morgan & Chase Bank", "country": "US", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 10, "n_ref_uni": 41, "n_ref": 86, "n_ref_all": 103, "n_fig": 8, "n_tab": 15, "L_tab": 3429, "n_element_tab": 395, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1023, "n_element_tab_1": 85, "formula_len_all": 989, "formula_len_all_1": 706, "len_all": 180347, "len_all_1": 69889, "len_abs": 1404, "len_title": 114, "len_sents": 52423, "len_sents_1": 31844, "n_sents": 464, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1408, "title": "PaLD: Detection of Text Partially Written by Large Language Models", "abs": "Advances in large language models (LLM) have produced text that appears increasingly human-like and difficult to detect with the human eye. In order to mitigate the impact of misusing LLM-generated texts, e.g., copyright infringement, fair student assessment, fraud, and other societally harmful LLM usage, a line of work on detecting human and LLM-written text has been explored. While recent work has focused on classifying entire text samples (e.g., paragraphs) as human or LLM-written, this paper investigates a more realistic setting of mixed-text, where the text's individual segments (e.g., sentences) could each be written by either a human or an LLM. A text encountered in practical usage cannot generally be assumed to be fully human or fully LLM-written; simply predicting whether it is human or LLM-written is insufficient as it does not provide the user with full context on its origins, such as the amount of LLM-written text, or locating the LLM-written parts. Therefore, we study two relevant problems in the mixed-text setting: (i) estimating the percentage of a text that was LLM-written, and (ii) determining which segments were LLM-written. To this end, we propose Partial-LLM Detector (PaLD), a black-box method that leverages the scores of text classifiers. Experimentally, we demonstrate the effectiveness of PaLD compared to baseline methods that build on existing LLM text detectors.", "keywords": [ "LLM text detection", "mixed-text detection", "distribution shift" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "rWQDzq3O5c", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Lawrence Carin", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Full Professor" }, { "name": "Suvrit Sra", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Xiang Cheng", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" } ], "n_formula": 61, "n_formula_1": 23, "n_ref_uni": 33, "n_ref": 60, "n_ref_all": 83, "n_fig": 3, "n_tab": 4, "L_tab": 597, "n_element_tab": 39, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 452, "n_element_tab_1": 31, "formula_len_all": 4898, "formula_len_all_1": 1537, "len_all": 159965, "len_all_1": 76629, "len_abs": 840, "len_title": 89, "len_sents": 47802, "len_sents_1": 30788, "n_sents": 470, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 844, "title": "Graph Transformers Dream of Electric Flow", "abs": "We show theoretically and empirically that the linear Transformer, when applied to graph data, can implement algorithms that solve canonical problems such as electric flow and eigenvector decomposition. The Transformer has access to information on the input graph only via the graph's incidence matrix. We present explicit weight configurations for implementing each algorithm, and we bound the constructed Transformers' errors by the errors of the underlying algorithms. We verify our theoretical findings experimentally on synthetic data. Additionally, on a real-world molecular regression task, we observe that the linear Transformer is capable of learning a better positional encoding than the default one based on Laplacian eigenvectors. Our work is an initial step towards elucidating the inner-workings of the Transformer for graph data.", "keywords": [ "Transformer", "Graph Neural Network" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "rWIrdAo2xC", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Junhui Hou", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Qijian Zhang", "gender": "Male", "institution": "Tencent Games", "country": "CN", "position": "Researcher" }, { "name": "Yingzhi Tang", "gender": "Male", "institution": "City University of Hong Kong", "country": "", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 37, "n_ref": 69, "n_ref_all": 97, "n_fig": 17, "n_tab": 3, "L_tab": 813, "n_element_tab": 148, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 3381, "n_element_tab_1": 144, "formula_len_all": 1014, "formula_len_all_1": 1164, "len_all": 126207, "len_all_1": 62402, "len_abs": 1192, "len_title": 131, "len_sents": 37332, "len_sents_1": 25359, "n_sents": 291, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1199, "title": "Generalizable Monocular 3D Human Rendering via Direct Gaussian Attribute Diffusion", "abs": "This paper leverages 3D Gaussian Splatting to tackle the challenging task of generating novel views of humans from given single-view images. Existing methods typically adopt an indirect supervision manner, i.e., splat-based rasterization for differentiable rendering. However, the intricate coupling of various 3D Gaussian attributes complicates precise error backpropagation during optimization, often\nresulting in convergence to local optima. In contrast, we propose a novel direct paradigm to train a conditional diffusion model directly supervised by proxy-ground-truth 3D Gaussian attributes. Specifically, we propose a two-stage construction process to derive consistent and smoothly distributed proxy-ground-truth 3D Gaussian attributes. Subsequently, we train a point-based conditional diffusion model customized to learn the data distribution of these proxy attributes. The resulting diffusion model can generate the 3D Gaussian attributes for the input\nsingle-view image, which are further rendered into novel views. Extensive experimental results showcase the significant performance advancement of our method over state-of-the-art approaches. Source code will be made publicly available.", "keywords": [ "novel view synthesis", "3D human body reconstruction", "3D Gaussian splatting" ], "rating_list": [ 3, 5, 5, 10, 3 ], "soundness_list": [ 2, 3, 2, 4, 2 ], "presentation_list": [ 3, 2, 2, 4, 2 ], "contribution_list": [ 2, 2, 2, 4, 2 ], "confidence_list": [ 5, 4, 3, 4, 2 ] }, { "paper_id": "rW3NVhKtQ2", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Alan Wee-Chung Liew", "gender": "Male", "institution": "Griffith University", "country": "", "position": "Full Professor" }, { "name": "Haishuai Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Research Professor" }, { "name": "Qin Zhang", "gender": "Female", "institution": "Shenzhen University", "country": "CN", "position": "Lecturer" }, { "name": "Shirui Pan", "gender": "unknown", "institution": "Griffith University", "country": "AU", "position": "Full Professor" }, { "name": "Xin Zheng", "gender": "Female", "institution": "Griffith University", "country": "AU", "position": "Lecturer" }, { "name": "Yu Zheng", "gender": "Female", "institution": "Latrobe University", "country": "", "position": "PhD student" }, { "name": "Yuxuan Liang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 23, "n_ref": 70, "n_ref_all": 89, "n_fig": 7, "n_tab": 7, "L_tab": 3236, "n_element_tab": 179, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1631, "n_element_tab_1": 95, "formula_len_all": 1249, "formula_len_all_1": 999, "len_all": 125987, "len_all_1": 62534, "len_abs": 2196, "len_title": 125, "len_sents": 36287, "len_sents_1": 28148, "n_sents": 197, "n_sents_1": 149, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1715, "title": "Test-Time Graph Rebirth: Serving GNN Generalization Under Distribution Shifts", "abs": "Distribution shifts between training and test graphs typically lead to the decreased performance of graph neural networks (GNNs) with suboptimal generalization in real-world applications. Despite advances in graph learning under distribution shifts through designing various model architecture development with customized training strategies, existing solutions can be challenging in practical GNN deployment because they often require significant modifications or retraining of the GNNs. To address such challenges, in this work, we propose a novel method, i.e., Test-Time Graph REBirth, dubbed TT-GREB, to effectively generalize the well-trained GNN models to the test-time graphs under distribution shifts by directly manipulating the test graph data. Concretely, we develop an overall framework designed by two principles, corresponding to two submodules: (1) prototype extractor for re-extracting the environment-invariant features of the test-time graph; and (2) environment refiner for re-fining the environment-varying features to explore the potential shifts. Furthermore, we propose a dual test-time graph contrastive learning objective with an effective iterative optimization strategy to obtain optimal prototype components and environmental components of the test graph. By reassembling these two components, we could obtain a newly reborn test graph, which is better suited for generalization on the well-trained GNN model with shifts in graph distribution. Extensive experiments on real-world graphs under diverse test-time distribution shifts could verify the effectiveness of the proposed method, showcasing its superior ability to manipulate test-time graphs for better GNN generalization ability.", "keywords": [ "graph neural networks", "distribution shifts", "test-time inference", "MLOps" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "rVD4lasVp4", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Chaosheng Dong", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Haibo Yang", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Jia Liu", "gender": "Male", "institution": "The Ohio State University", "country": "US", "position": "Associate Professor" }, { "name": "Mingyi Hong", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Peiwen Qiu", "gender": "Female", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Prashant Khanduri", "gender": "Male", "institution": "Wayne State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 110, "n_formula_1": 25, "n_ref_uni": 71, "n_ref": 169, "n_ref_all": 216, "n_fig": 8, "n_tab": 7, "L_tab": 2601, "n_element_tab": 349, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 579, "n_element_tab_1": 34, "formula_len_all": 22338, "formula_len_all_1": 2916, "len_all": 250593, "len_all_1": 79104, "len_abs": 1511, "len_title": 133, "len_sents": 59611, "len_sents_1": 32612, "n_sents": 555, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1538, "title": "A Lazy Hessian Evaluation Framework for Accelerating Stochastic Bilevel Optimization", "abs": "Bilevel optimization has recently gained popularity because of its applicability in many machine learning applications.\nHypergradient-based algorithms have been widely used for solving bilevel optimization problems because of their strong theoretical and empirical performance in many applications. \nHowever, computing these hypergradients requires the evaluation of Hessians (or Hessian-vector products) of the lower-level objective, which presents a major computational bottleneck. \nTo address this challenge, in this paper, we propose LazyBLO \n(**Lazy** Hessian Evaluation in **B**i**l**evel **O**ptimization), an algorithmic framework that allows infrequent Hessian computation during the execution of the algorithm for solving stochastic bilevel problems.\nThis allows the algorithm to execute faster compared to the state-of-the-art (SOTA) algorithms that evaluate either a single or multiple Hessians in each iteration. \nWe theoretically establish the performance of vanilla SGD-based LazyBLO and show that, despite the additional errors incurred by the infrequent Hessian evaluations, LazyBLO surprisingly matches the computation complexity of the existing SGD-based bilevel algorithms.\nExtensive experiments further demonstrate that LazyBLO enjoys significant gains in numerical performance compared to the SOTA approaches. \nTo our knowledge, this is the first work to theoretically establish that multiple Hessian computations are not necessary within each iteration to guarantee the convergence of stochastic bilevel algorithms.", "keywords": [ "bilevel optimization", "stochastic optimization", "lazy Hessian evaluation" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 2, 4, 5 ] }, { "paper_id": "rUxr9Ll5FQ", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Kai Wang", "gender": "Male", "institution": "Computer Vision Center, Universitat Aut\u00f3noma de Barcelona", "country": "ES", "position": "Postdoc" }, { "name": "Ming-Ming Cheng", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Full Professor" }, { "name": "Senmao Li", "gender": "unknown", "institution": "Nankai University", "country": "TD", "position": "PhD student" }, { "name": "Shiqi Yang", "gender": "Male", "institution": "SB Intuitions, SoftBank", "country": "JP", "position": "Chief Research Scientist" }, { "name": "Yaxing Wang", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "jian Yang", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Full Professor" }, { "name": "Chunle Guo", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Fahad Shahbaz Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" }, { "name": "Joost van de", "gender": "Male", "institution": "Universitat Aut\u00f3noma de Barcelona", "country": "ES", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 6, "n_ref_uni": 73, "n_ref": 167, "n_ref_all": 232, "n_fig": 24, "n_tab": 8, "L_tab": 2895, "n_element_tab": 155, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 741, "n_element_tab_1": 10, "formula_len_all": 826, "formula_len_all_1": 267, "len_all": 200600, "len_all_1": 60117, "len_abs": 1600, "len_title": 118, "len_sents": 50755, "len_sents_1": 29659, "n_sents": 373, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 119, "L_abs": 1656, "title": "$InterLCM$: Low-Quality Images as Intermediate States of Latent Consistency Models for Effective Blind Face Restoration", "abs": "Diffusion priors have been used for blind face restoration (BFR) by fine-tuning diffusion models (DMs) on restoration datasets to recover low-quality images. However, the naive application of DMs presents several key limitations. \n(i) The diffusion prior has inferior semantic consistency (e.g., ID, structure and color.), increasing the difficulty of optimizing the BFR model;\n(ii) reliance on hundreds of denoising iterations, preventing the effective cooperation with perceptual losses, which is crucial for faithful restoration.\nObserving that the latent consistency model (LCM) learns consistency noise-to-data mappings on the ODE-trajectory and therefore shows more semantic consistency in the subject identity, structural information and color preservation, \nwe propose $\\textit{InterLCM}$ to leverage the LCM for its superior semantic consistency and efficiency to counter the above issues. \nTreating low-quality images as the intermediate state of LCM, $\\textit{InterLCM}$ achieves a balance between fidelity and quality by starting from earlier LCM steps. \nLCM also allows the integration of perceptual loss during training, leading to improved restoration quality, particularly in real-world scenarios.\nTo mitigate structural and semantic uncertainties, $\\textit{InterLCM}$ incorporates a Visual Module to extract visual features and a Spatial Encoder to capture spatial details, enhancing the fidelity of restored images.\nExtensive experiments demonstrate that $\\textit{InterLCM}$ outperforms existing approaches in both synthetic and real-world datasets while also achieving faster inference speed. Code and models will be publicly available.", "keywords": [ "diffusion model", "face restoration" ], "rating_list": [ 6, 5, 6, 3, 6 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 4, 3, 4 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "rUvCIvI4eB", "primary_area": "applications to robotics, autonomy, planning", "n_author": 9, "author_info_list": [ { "name": "Donglin Yang", "gender": "Male", "institution": "beihang university", "country": "CN", "position": "PhD student" }, { "name": "Hohin Kwan", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Jinyu Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Intern" }, { "name": "Si Liu", "gender": "Female", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Xiangyu Wang", "gender": "Male", "institution": "Beihang University", "country": "", "position": "Undergrad student" }, { "name": "Yue Liao", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Ziqin Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Wenjun Wu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 28, "n_ref": 61, "n_ref_all": 92, "n_fig": 10, "n_tab": 7, "L_tab": 2128, "n_element_tab": 131, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1314, "n_element_tab_1": 110, "formula_len_all": 58, "formula_len_all_1": 58, "len_all": 132456, "len_all_1": 66321, "len_abs": 1868, "len_title": 136, "len_sents": 43882, "len_sents_1": 31219, "n_sents": 329, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1873, "title": "Towards Realistic UAV Vision-Language Navigation: Platform, Benchmark, and Methodology", "abs": "Developing agents capable of navigating to a target location based on language instructions and visual information, known as vision-language navigation (VLN), has attracted widespread interest. Most research has focused on ground-based agents, while UAV-based VLN remains relatively underexplored. Recent efforts in UAV vision-language navigation predominantly adopt ground-based VLN settings, relying on predefined discrete action spaces and neglecting the inherent disparities in agent movement dynamics and the complexity of navigation tasks between ground and aerial environments. To address these disparities and challenges, we propose solutions from three perspectives: platform, benchmark, and methodology. To enable realistic UAV trajectory simulation in VLN tasks, we propose the OpenUAV platform, which features diverse environments, realistic flight control, and extensive algorithmic support. We further construct a target-oriented VLN dataset consisting of approximately 12k trajectories on this platform, serving as the first dataset specifically designed for realistic UAV VLN tasks. To tackle the challenges posed by complex aerial environments, we propose an assistant-guided UAV object search benchmark called UAV-Need-Help, which provides varying levels of guidance information to help UAVs better accomplish realistic VLN tasks. We also propose a UAV navigation LLM that, given multi-view images, task descriptions, and assistant instructions, leverages the multimodal understanding capabilities of the MLLM to jointly process visual and textual information, and performs hierarchical trajectory generation. The evaluation results of our method significantly outperform the baseline models, while there remains a considerable gap between our results and those achieved by human operators, underscoring the challenge presented by the UAV-Need-Help task.", "keywords": [ "Unmanned Aerial Vehicle", "Drone", "Vision-Language Navigation" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "rUqcugZDUl", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Miaomiao Zhang", "gender": "Female", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" }, { "name": "Tonmoy Hossain", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 46, "n_ref": 75, "n_ref_all": 86, "n_fig": 5, "n_tab": 6, "L_tab": 2046, "n_element_tab": 199, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1128, "n_element_tab_1": 136, "formula_len_all": 613, "formula_len_all_1": 771, "len_all": 134129, "len_all_1": 61107, "len_abs": 1786, "len_title": 133, "len_sents": 38722, "len_sents_1": 28355, "n_sents": 254, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1815, "title": "ToRL: Topology-preserving Representation Learning Of Object Deformations From Images", "abs": "Representation learning of object deformations from images has been a long-standing challenge in various image or video analysis tasks. Existing deep neural networks typically focus on visual features (e.g., intensity and texture), but they often fail to capture the underlying geometric and topological structures of objects. This limitation becomes especially critical in areas, such as medical imaging and 3D modeling, where maintaining the structural integrity of objects is essential for accuracy and generalization across diverse datasets. In this paper, we introduce ToRL, a novel *Topology-preserving Representation Learning* model that, for the first time, offers an explicit mechanism for modeling intricate object topology in the latent feature space. We develop a comprehensive learning framework that captures object deformations via learned transformation groups in the latent space. Each layer of our network's decoder is carefully designed with an integrated smooth composition module, ensuring that topological properties are preserved throughout the learning process. Moreover, in contrast to a few related works that rely on a reference image to predict object deformations during inference, our approach eliminates this impractical requirement. To validate ToRL's effectiveness, we conduct extensive multi-class classification experiments across a wide range of datasets, including synthetic 2D images, real 3D brain magnetic resonance imaging (MRI) scans, real 3D adrenal computed tomography (CT) shapes, and \\textcolor{blue}{real 2D facial expression images}. Experimental results demonstrate that ToRL outperforms state-of-the-art methods, setting a new way to enforce topological consistency in representation learning. Our code is available at - https://anonymous.4open.science/r/ToRL-44BF/", "keywords": [ "Representation Learning", "Deformations", "Topology" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "rUC7tHecSQ", "primary_area": "interpretability and explainable AI", "n_author": 1, "author_info_list": [ { "name": "Tiberiu Mu\u0219at", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 0, "n_ref_uni": 24, "n_ref": 32, "n_ref_all": 43, "n_fig": 6, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7551, "formula_len_all_1": 0, "len_all": 106907, "len_all_1": 53778, "len_abs": 824, "len_title": 127, "len_sents": 32013, "len_sents_1": 24626, "n_sents": 329, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 760, "title": "Mechanism and emergence of stacked attention heads in multi-layer transformers", "abs": "In this paper, I introduce the retrieval problem, a simple yet common reasoning task that can be solved only by transformers with a minimum number of layers, which grows logarithmically with the input size. I empirically show that large language models can solve the task under different prompting formulations without any fine-tuning. To understand how transformers solve the retrieval problem, I train several transformers on a minimal formulation. Successful learning occurs only under the presence of an implicit curriculum. I uncover the learned mechanisms by studying the attention maps in the trained transformers. I also study the training process, uncovering that attention heads always emerge in a specific sequence guided by the implicit curriculum.", "keywords": [ "mechanistic interpretability", "large language models", "transformers", "emergent abilities", "curriculum learning", "reasoning" ], "rating_list": [ 5, 6, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "rTQNGQxm4K", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Nicolas Yax", "gender": "unknown", "institution": "Ecole Normale Sup\u00e9rieure de Paris", "country": "", "position": "PhD student" }, { "name": "Pierre-Yves Oudeyer", "gender": "Male", "institution": "Inria", "country": "", "position": "Research director" }, { "name": "Stefano Palminteri", "gender": "Male", "institution": "INSERM", "country": "", "position": "Principal Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 34, "n_ref": 54, "n_ref_all": 117, "n_fig": 33, "n_tab": 2, "L_tab": 2417, "n_element_tab": 420, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 150, "n_element_tab_1": 5, "formula_len_all": 141, "formula_len_all_1": 144, "len_all": 188380, "len_all_1": 65331, "len_abs": 893, "len_title": 153, "len_sents": 71141, "len_sents_1": 33990, "n_sents": 481, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 899, "title": "PhyloLM: Inferring the Phylogeny of Large Language Models and Predicting their Performances in Benchmarks", "abs": "This paper introduces PhyloLM, a method adapting phylogenetic algorithms to Large Language Models (LLMs) to explore whether and how they relate to each other and to predict their performance characteristics. Our method calculates a phylogenetic distance metric based on the similarity of LLMs' output. The resulting metric is then used to construct dendrograms, which satisfactorily capture known relationships across a set of 111 open-source and 45 closed models. Furthermore, our phylogenetic distance predicts performance in standard benchmarks, thus demonstrating its functional validity and paving the way for a time and cost-effective estimation of LLM capabilities. To sum up, by translating population genetic concepts to machine learning, we propose and validate a tool to evaluate LLM development, relationships and capabilities, even in the absence of transparent training information.", "keywords": [ "large language models", "phylogeny", "benchmark" ], "rating_list": [ 3, 10, 5, 6 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 1, 4, 3, 3 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "rTM95kwzXM", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Alexander Ratner", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "", "position": "Assistant Professor" }, { "name": "Hui Xiong", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Jieyu Zhang", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Ranjay Krishna", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "Zhengyu Hu", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "MS student" }, { "name": "Zhihan Xiong", "gender": "Male", "institution": "University of Washington", "country": "", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 8, "n_ref_uni": 38, "n_ref": 89, "n_ref_all": 126, "n_fig": 7, "n_tab": 10, "L_tab": 4042, "n_element_tab": 515, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1820, "n_element_tab_1": 292, "formula_len_all": 2717, "formula_len_all_1": 715, "len_all": 232531, "len_all_1": 65817, "len_abs": 1474, "len_title": 114, "len_sents": 75919, "len_sents_1": 29745, "n_sents": 637, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1483, "title": "Language Model Preference Evaluation with Multiple Weak Evaluators", "abs": "Despite the remarkable success of Large Language Models (LLMs), evaluating their outputs' quality regarding *preference* remains a critical challenge. Existing works usually leverage a powerful LLM (e.g., GPT4) as the judge for comparing LLMs' output pairwisely, yet such model-based evaluator is vulnerable to *conflicting preference*, i.e., output A is better than B, B than C, but C than A, causing contradictory evaluation results. To improve model-based preference evaluation, we introduce GED (Preference Graph Ensemble and Denoise), a novel approach that leverages multiple model-based evaluators to construct preference graphs, and then ensemble and denoise these graphs for better, non-contradictory evaluation results. In particular, our method consists of two primary stages: aggregating evaluations into a unified graph and applying a denoising process to eliminate cyclic inconsistencies, ensuring a directed acyclic graph (DAG) structure. We provide theoretical guarantees for our framework, demonstrating its efficacy in recovering the ground truth preference structure. Extensive experiments across ten benchmark datasets show that GED outperforms baseline methods in model ranking, response selection, and model alignment tasks. Notably, GED combines weaker evaluators like Llama3-8B, Mistral-7B, and Qwen2-7B to surpass the performance of stronger evaluators like Qwen2-72B, highlighting its ability to enhance evaluation reliability and improve model performance.", "keywords": [ "Large Language Models", "Weak Evaluators" ], "rating_list": [ 5, 3, 8, 5 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 1, 2, 4, 3 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "rTCJ29pkuA", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Julian Shun", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Associate Professor" }, { "name": "Jundong Li", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" }, { "name": "Junhong Lin", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Song Wang", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Xiaojie Guo", "gender": "Female", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Yada Zhu", "gender": "unknown", "institution": "IBM Research", "country": "", "position": "Principal Research Scientist" } ], "n_formula": 25, "n_formula_1": 10, "n_ref_uni": 24, "n_ref": 60, "n_ref_all": 75, "n_fig": 6, "n_tab": 3, "L_tab": 382, "n_element_tab": 78, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 234, "n_element_tab_1": 61, "formula_len_all": 1598, "formula_len_all_1": 914, "len_all": 118065, "len_all_1": 65758, "len_abs": 1380, "len_title": 125, "len_sents": 36033, "len_sents_1": 28813, "n_sents": 323, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1387, "title": "Reasoning of Large Language Models over Knowledge Graphs with Super-Relations", "abs": "While large language models (LLMs) have made significant progress in processing and reasoning over knowledge graphs, current methods suffer from a high non-retrieval rate. This limitation reduces the accuracy of answering questions based on these graphs. Our analysis reveals that the combination of greedy search and forward reasoning is a major contributor to this issue. To overcome these challenges, we introduce the concept of super-relations, which enables both forward and backward reasoning by summarizing and connecting various relational paths within the graph. This holistic approach not only expands the search space, but also significantly improves retrieval efficiency. In this paper, we propose the ReKnoS framework, which aims to Reason over Knowledge Graphs with Super-Relations. Our framework\u2019s key advantages include the inclusion of multiple relation paths through super-relations, enhanced forward and backward reasoning capabilities, and increased efficiency in querying LLMs. These enhancements collectively lead to a substantial improvement in the successful retrieval rate and overall reasoning performance. We conduct extensive experiments on a variety of datasets to evaluate ReKnoS, and the results demonstrate the superior performance of ReKnoS over existing state-of-the-art baselines, with an average accuracy gain of 2.92% across nine real-world datasets.", "keywords": [ "Knowledge Graphs", "Large Language Models", "Question Answering" ], "rating_list": [ 6, 5, 8 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "rSNkMy4OkJ", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Bhanu Prasad Mokalla", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "MSVPJ Sathvik", "gender": "Male", "institution": "IIIT Dharwad", "country": "IN", "position": "Researcher" }, { "name": "Sai Amulya Pingili", "gender": "unknown", "institution": "Arizona State University", "country": "US", "position": "MS student" }, { "name": "Srreyansh Sethi", "gender": "Male", "institution": "American High School", "country": "US", "position": "High Schooler" }, { "name": "Yalamanchi Sreedweep Sharma", "gender": "Male", "institution": "Sreenidhi Institute of Science & Technology", "country": "IN", "position": "Undergrad student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 8, "n_ref": 13, "n_ref_all": 18, "n_fig": 2, "n_tab": 4, "L_tab": 1276, "n_element_tab": 65, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2725, "n_element_tab_1": 76, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 90631, "len_all_1": 54541, "len_abs": 917, "len_title": 125, "len_sents": 28278, "len_sents_1": 24000, "n_sents": 204, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 857, "title": "From Eww to Woo: Detection of Mental Health Disturbing Images in Social Media", "abs": "Exposure to distressing images on social media, such as gore and other graphic content, can lead to significant mental health issues and disturbances. This paper introduces a novel dataset specifically curated to include such harmful images, aiming to facilitate the development of machine learning models capable of detecting and filtering these types of content. By training on this dataset, the proposed models demonstrate the ability to accurately identify and flag disturbing images, thereby contributing to the mitigation of mental health risks associated with prolonged exposure to harmful visual content on social media platforms. The proposed dataset is benchmarked on various state of the art models with the accuracy 70.15\\%. This work represents a critical step towards creating safer online environments and protecting users' mental well-being.", "keywords": [ "Mental Health", "Social Media", "Harmful" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "rSAPrQzoQa", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Jiashun Jin", "gender": "Male", "institution": "carnegie mellon university", "country": "US", "position": "Full Professor" }, { "name": "Yinan Guo", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 6, "n_ref_uni": 15, "n_ref": 46, "n_ref_all": 69, "n_fig": 2, "n_tab": 11, "L_tab": 4351, "n_element_tab": 522, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1989, "n_element_tab_1": 330, "formula_len_all": 973, "formula_len_all_1": 520, "len_all": 137076, "len_all_1": 73121, "len_abs": 1254, "len_title": 98, "len_sents": 50250, "len_sents_1": 31641, "n_sents": 415, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1231, "title": "Subject Clustering by an Improved IF-PCA Algorithm", "abs": "Subject (e.g., cell or patient) clustering is an important problem in genetics and genomics. Influential features PCA (IF-PCA) is a recent idea for clustering, where we first select a small fraction of measured features and then cluster subjects (e.g., cells or patient) into different groups using the classical PCA clustering approach. A challenge the method faces is that, we may have complex signal and noise structures across features or across subjects or both, which may make the IF-PCA less effective. \nTo deal with such a challenge, we propose a new approach, IFPCA+, where we combine IF-PCA with the recent idea of manifold fitting. The latter was shown to better support class separation. We compare our approach with the most popular subject clustering approaches, including but not limited to DESC, SC3 and Seurat, using 10 gene microarray data sets and 8 single-cell data sets. We show that with the new method, we have a significant improvement in feature selection accuracy, and that on average, our method outperforms several of the most competitive algorithms nowadays (including IF-PCA, DESC, Seurat) in terms of clustering accuracy and ARI. We also shed light on the insight underlying such improvements.", "keywords": [ "gene microarray", "scRNA-seq", "feature selection", "manifold fitting", "nonlinearity", "PCA", "sparsity", "subject clustering" ], "rating_list": [ 6, 6, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 3, 3, 2, 5 ] }, { "paper_id": "rRRgj3iIHR", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Chang Lei", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Huan Lei", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 12, "n_ref": 31, "n_ref_all": 42, "n_fig": 3, "n_tab": 12, "L_tab": 4428, "n_element_tab": 370, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 2535, "n_element_tab_1": 163, "formula_len_all": 385, "formula_len_all_1": 378, "len_all": 116750, "len_all_1": 58377, "len_abs": 1410, "len_title": 117, "len_sents": 38704, "len_sents_1": 25795, "n_sents": 298, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1417, "title": "AlphaDou: High-Performance End-to-End Doudizhu AI Integrating Bidding", "abs": "Artificial intelligence for card games has long been a popular topic in AI research. In recent years, complex card games like Mahjong and Texas Hold'em have been solved, with corresponding AI programs reaching the level of human experts. However, the game of Doudizhu presents significant challenges due to its vast state/action space and unique characteristics involving reasoning about competition and cooperation, making the game extremely difficult to solve.The RL model Douzero, trained using the Deep Monte Carlo algorithm framework, has shown excellent performance in Doudizhu. However, there are differences between its simplified game environment and the actual Doudizhu environment, and its performance is still a considerable distance from that of human experts. This paper modifies the Deep Monte Carlo algorithm framework by using reinforcement learning to obtain a neural network that simultaneously estimates win rates and expectations. The action space is pruned using expectations, and strategies are generated based on win rates. The modified algorithm enables the AI to perform the full range of tasks in the Doudizhu game, including bidding and cardplay. The model was trained in a actual Doudizhu environment and achieved state-of-the-art performance among publicly available models. We hope that this new framework will provide valuable insights for AI development in other bidding-based games.", "keywords": [ "Multi-Agent Reinforcement Learning", "Imperfect-Information Games", "Deep Mento Carlo", "Adversarial and Cooperative" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "rQyg6MnsDb", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Chengqi Zhang", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Chair Professor" }, { "name": "Ciyuan Peng", "gender": "Female", "institution": "Federation University Australia", "country": "AU", "position": "PhD student" }, { "name": "Feng Xia", "gender": "Male", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "Full Professor" }, { "name": "Qichao Dong", "gender": "Male", "institution": "Zhejiang Gongshang University", "country": "", "position": "MS student" }, { "name": "Shuo Yu", "gender": "Female", "institution": "Dalian University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yaochu Jin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" }, { "name": "Huang yuelong", "gender": "Male", "institution": "Dalian University of Technology", "country": "", "position": "MS student" } ], "n_formula": 30, "n_formula_1": 10, "n_ref_uni": 55, "n_ref": 88, "n_ref_all": 118, "n_fig": 12, "n_tab": 10, "L_tab": 3338, "n_element_tab": 226, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1511, "n_element_tab_1": 76, "formula_len_all": 2420, "formula_len_all_1": 651, "len_all": 193079, "len_all_1": 69479, "len_abs": 1277, "len_title": 95, "len_sents": 51599, "len_sents_1": 31792, "n_sents": 431, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 46, "L_abs": 1276, "title": "Biologically Plausible Brain Graph Transformer", "abs": "State-of-the-art brain graph analysis methods fail to fully encode the small-world architecture of brain graphs (accompanied by the presence of hubs and functional modules), and therefore lack biological plausibility to some extent. This limitation hinders their ability to accurately represent the brain's structural and functional properties, thereby restricting the effectiveness of machine learning models in tasks such as brain disorder detection. In this work, we propose a novel Biologically Plausible Brain Graph Transformer (BioBGT) that encodes the small-world architecture inherent in brain graphs. Specifically, we present a network entanglement-based node importance encoding technique that captures the structural importance of nodes in global information propagation during brain graph communication, highlighting the biological properties of the brain structure. Furthermore, we introduce a functional module-aware self-attention to preserve the functional segregation and integration characteristics of brain graphs in the learned representations. Experimental results on three benchmark datasets demonstrate that BioBGT outperforms state-of-the-art models, enhancing biologically plausible brain graph representations for various brain graph analytical tasks", "keywords": [ "brain", "graph learning", "transformer", "graph representation", "brain networks" ], "rating_list": [ 8, 5, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 4, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "rQV33MVNWs", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Bin Zhao", "gender": "Male", "institution": "Shanghai Artificial Intelligence Labortory", "country": "CN", "position": "Young Scientist" }, { "name": "Delin Qu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Dong Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Haoming Song", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Qizhi Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Xuelong Li", "gender": "Male", "institution": "China Telecom", "country": "CN", "position": "Full Professor" }, { "name": "Yiting Zhang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yiwen Tang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Intern" } ], "n_formula": 27, "n_formula_1": 14, "n_ref_uni": 36, "n_ref": 82, "n_ref_all": 107, "n_fig": 0, "n_tab": 8, "L_tab": 9555, "n_element_tab": 1048, "n_fig_1": 0, "n_tab_1": 6, "L_tab_1": 3143, "n_element_tab_1": 57, "formula_len_all": 1722, "formula_len_all_1": 902, "len_all": 169560, "len_all_1": 65258, "len_abs": 1129, "len_title": 129, "len_sents": 40694, "len_sents_1": 27391, "n_sents": 336, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1088, "title": "FreeGaussian: Guidance-free Controllable 3D Gaussian Splats with Flow Derivatives", "abs": "Reconstructing controllable Gaussian splats from monocular video is a challenging task due to its inherently insufficient constraints. Widely adopted approaches supervise complex interactions with additional masks and control signal annotations, limiting their real-world applications. In this paper, we propose an annotation guidance-free method, dubbed FreeGaussian, that mathematically derives dynamic Gaussian motion from optical flow and camera motion using novel dynamic Gaussian constraints. By establishing a connection between 2D flows and 3D Gaussian dynamic control, our method enables self-supervised optimization and continuity of dynamic Gaussian motions from flow priors. Furthermore, we introduce a 3D spherical vector controlling scheme, which represents the state with a 3D Gaussian trajectory, thereby eliminating the need for complex 1D control signal calculations and simplifying controllable Gaussian modeling. Quantitative and qualitative evaluations on extensive experiments demonstrate the state-of-the-art visual performance and control capability of our method.", "keywords": [ "3D Gaussian Splatting", "Controllable View Synthesis" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "rQ8mHhEIeB", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Harry Shomer", "gender": "unknown", "institution": "Michigan State University", "country": "", "position": "PhD student" }, { "name": "Jay Revolinsky", "gender": "unknown", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Jiliang Tang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 56, "n_ref": 158, "n_ref_all": 195, "n_fig": 7, "n_tab": 15, "L_tab": 9070, "n_element_tab": 813, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 359, "n_element_tab_1": 87, "formula_len_all": 155, "formula_len_all_1": 396, "len_all": 193604, "len_all_1": 58735, "len_abs": 1197, "len_title": 127, "len_sents": 49595, "len_sents_1": 27998, "n_sents": 357, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1204, "title": "Towards Understanding Link Predictor Generalizability Under Distribution Shifts", "abs": "State-of-the-art link prediction (LP) models demonstrate impressive benchmark\nresults. However, popular benchmark datasets often assume that training, validation, and testing samples are representative of the overall dataset distribution. In\nreal-world situations, this assumption is often incorrect; since uncontrolled factors\nlead to the problem where new dataset samples come from different distributions\nthan training samples. The vast majority of recent work focuses on dataset shift\naffecting node- and graph-level tasks, largely ignoring link-level tasks. To bridge\nthis gap, we introduce a novel splitting strategy, known as LPShift, which utilizes\nstructural properties to induce a controlled distribution shift. We verify the effect of LPShift through empirical evaluation of SOTA LP methods on 16 LPShift\ngenerated splits of Open Graph Benchmark (OGB) datasets. When benchmarked\nwith LPShift datasets, GNN4LP methods frequently generalize worse than heuristics or basic GNNs. Furthermore, LP-specific generalization techniques do little\nto improve performance under LPShift. Finally, further analysis provides insight\non why LP models lose much of their architectural advantages under LPShift.", "keywords": [ "Link Prediction", "Graph-Structured Data", "GNN4LP", "Distribution Shifts", "Structural Heuristics", "Splitting Strategies" ], "rating_list": [ 5, 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "rQ7fz9NO7f", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Gang Liu", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Jie Chen", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "Research Staff Member" }, { "name": "Meng Jiang", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "Associate Professor" }, { "name": "Michael Sun", "gender": "Male", "institution": "Computer Science and Artificial Intelligence Laboratory, Electrical Engineering & Computer Science", "country": "US", "position": "PhD student" }, { "name": "Wojciech Matusik", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 38, "n_ref": 113, "n_ref_all": 161, "n_fig": 9, "n_tab": 9, "L_tab": 3657, "n_element_tab": 91, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 567, "n_element_tab_1": 8, "formula_len_all": 1454, "formula_len_all_1": 721, "len_all": 218716, "len_all_1": 63731, "len_abs": 1111, "len_title": 139, "len_sents": 76408, "len_sents_1": 28229, "n_sents": 590, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1115, "title": "Multimodal Large Language Models for Inverse Molecular Design with Retrosynthetic Planning", "abs": "While large language models (LLMs) have integrated images, adapting them to graphs remains challenging, limiting their applications in materials and drug design. This difficulty stems from the need for coherent autoregressive generation across texts and graphs. To address this, we introduce Llamole, the first multimodal LLM capable of interleaved text and graph generation, enabling molecular inverse design with retrosynthetic planning. Llamole integrates a base LLM with the Graph Diffusion Transformer and Graph Neural Networks for multi-conditional molecular generation and reaction inference within texts, while the LLM, with enhanced molecular understanding, flexibly controls activation among the different graph modules. Additionally, Llamole integrates A* search with LLM-based cost functions for efficient retrosynthetic planning. We create benchmarking datasets and conduct extensive experiments to evaluate Llamole against in-context learning and supervised fine-tuning. Llamole significantly outperforms 14 adapted LLMs across 12 metrics for controllable molecular design and retrosynthetic planning.", "keywords": [ "Multimodal Large Languge Models", "Large Languge Models", "Graph Diffusion Models", "Inverse Molecular Design", "Retrosynthesis" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 2, 5, 2 ] }, { "paper_id": "rPup1cWk4d", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Mahito Sugiyama", "gender": "Male", "institution": "National Institute of Informatics", "country": "JP", "position": "Associate Professor" }, { "name": "Pingbang Hu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 56, "n_ref_all": 95, "n_fig": 13, "n_tab": 3, "L_tab": 545, "n_element_tab": 38, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 64, "n_element_tab_1": 10, "formula_len_all": 1400, "formula_len_all_1": 504, "len_all": 140178, "len_all_1": 68864, "len_abs": 694, "len_title": 107, "len_sents": 44433, "len_sents_1": 31388, "n_sents": 326, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 704, "title": "Pseudo-Non-Linear Data Augmentation via Energy Minimization", "abs": "We propose a novel and interpretable *data augmentation* method based on *energy-based modeling* and principles from *information geometry*. Unlike black-box generative models, which rely on deep neural networks, our approach replaces these non-interpretable transformations with explicit, theoretically grounded ones, ensuring interpretability and strong guarantees such as energy minimization. Central to our method is the introduction of the *backward projection* algorithm, which reverses dimension reduction to generate new data. Empirical results demonstrate that our method achieves competitive performance with black-box generative models while offering greater transparency and interpretability.", "keywords": [ "data augmentation", "information geometry", "interpretability" ], "rating_list": [ 1, 3, 5, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "rPkCVSsoM4", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Jacob Lear", "gender": "unknown", "institution": "University of Arkansas - Fayetteville", "country": "US", "position": "PhD student" }, { "name": "Lu Zhang", "gender": "Male", "institution": "University of Arkansas - Fayetteville", "country": "US", "position": "Associate Professor" } ], "n_formula": 30, "n_formula_1": 26, "n_ref_uni": 37, "n_ref": 60, "n_ref_all": 85, "n_fig": 11, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3620, "formula_len_all_1": 2245, "len_all": 125506, "len_all_1": 70473, "len_abs": 1085, "len_title": 98, "len_sents": 41301, "len_sents_1": 33715, "n_sents": 290, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1088, "title": "A Causal Lens for Learning Long-term Fair Policies", "abs": "Fairness-aware learning studies the development of algorithms that avoid discriminatory decision outcomes despite biased training data. While most studies have concentrated on immediate bias in static contexts, this paper highlights the importance of investigating long-term fairness in dynamic decision-making systems while simultaneously considering instantaneous fairness requirements. In the context of reinforcement learning, we propose a general framework where long-term fairness is measured by the difference in the average expected qualification gain that individuals from different groups could obtain. Then, through a causal lens, we decompose this metric into three components that represent the direct impact, the delayed impact, as well as the spurious effect the policy has on the qualification gain. We analyze the intrinsic connection between these components and an emerging fairness notion called benefit fairness that aims to control the equity of outcomes in decision-making. Finally, we develop a simple yet effective approach for balancing various fairness notions.", "keywords": [ "long-term fairness", "fair reinforcement learning", "causal decomposition" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "rP7rghI7yt", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Andrew Tao", "gender": "Male", "institution": "NVIDIA Corporation", "country": "US", "position": "Vice President of Applied Research" }, { "name": "Bryan Catanzaro", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Vice President" }, { "name": "Greg Heinrich", "gender": "Male", "institution": "NVIDIA", "country": "FR", "position": "Researcher" }, { "name": "Jan Kautz", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "VP Research" }, { "name": "Jon Barker", "gender": "Male", "institution": "University of Southampton", "country": "", "position": "PhD student" }, { "name": "Mike Ranzinger", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Pavlo Molchanov", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Research Scientist" } ], "n_formula": 39, "n_formula_1": 26, "n_ref_uni": 46, "n_ref": 73, "n_ref_all": 113, "n_fig": 6, "n_tab": 12, "L_tab": 4150, "n_element_tab": 765, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 959, "n_element_tab_1": 174, "formula_len_all": 2186, "formula_len_all_1": 1412, "len_all": 197241, "len_all_1": 50244, "len_abs": 570, "len_title": 103, "len_sents": 52428, "len_sents_1": 20832, "n_sents": 416, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1008, "title": "PHI-S: Distribution Balancing for Agglomerative Models", "abs": "Various visual foundation models have distinct strengths and weaknesses, both of which can be improved through heterogeneous multi-teacher knowledge distillation without labels, termed \"agglomerative models.\" We build upon this body of work by studying the effect of the teachers' activation statistics, particularly the impact of the loss function on the resulting student model quality. We explore a standard toolkit of statistical normalization techniques to better align the different distributions and assess their effects. Further, we examine the impact on downstream teacher-matching metrics, which motivates the use of Hadamard matrices. With these matrices, we demonstrate useful properties, showing how they can be used for isotropic standardization, where each dimension of a multivariate distribution is standardized using the same scale. We call this technique \"PHI Standardization\" (PHI-S) and empirically demonstrate that it produces the best student model across the suite of methods studied.", "keywords": [ "Computer Vision", "Deep Learning", "Knowledge Distillation", "Agglomerative Models" ], "rating_list": [ 5, 8, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 2, 5, 3, 4 ] }, { "paper_id": "rObkvzJxTG", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Tao Jin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zirun Guo", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 23, "n_ref": 72, "n_ref_all": 107, "n_fig": 14, "n_tab": 8, "L_tab": 6507, "n_element_tab": 506, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 6417, "n_element_tab_1": 542, "formula_len_all": 1127, "formula_len_all_1": 1114, "len_all": 129078, "len_all_1": 69331, "len_abs": 1344, "len_title": 136, "len_sents": 36329, "len_sents_1": 25438, "n_sents": 328, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1436, "title": "Smoothing the Shift: Towards Stable Test-Time Adaptation under Complex Multimodal Noises", "abs": "Test-Time Adaptation (TTA) aims to tackle distribution shifts using unlabeled test data without access to the source data. In the context of multimodal data, there are more complex noise patterns than unimodal data such as simultaneous corruptions for multiple modalities and missing modalities. Besides, in real-world applications, corruptions from different distribution shifts are always mixed. Existing TTA methods always fail in such multimodal scenario because the abrupt distribution shifts will destroy the prior knowledge from the source model, thus leading to performance degradation.\nTo this end, we reveal a new challenge named *multimodal wild TTA*.\nTo address this challenging problem, we propose two novel strategies: sample identification with interquartile range **S**moothing and **u**nimodal assistance, and **M**utual **i**nformation sharing (SuMi). SuMi smooths the adaptation process by interquartile range which avoids the abrupt distribution shifts. Then, SuMi fully utilizes the unimodal features to select low-entropy samples with rich multimodal information for optimization. Furthermore, mutual information sharing is introduced to align the information, reduce the discrepancies and enhance the information utilization across different modalities. Extensive experiments show the effectiveness and superiority over existing methods under the complex noise patterns in multimodal data. Code will be available.", "keywords": [ "Test-time Adaptation", "Multimodal Noises", "Domain Adaptation" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "rO5BVBwgiv", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Hsiang-Fu Yu", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Senior Applied Scientist" }, { "name": "Jiong Zhang", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Jyun-Yu Jiang", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Researcher and Applied Scientist" }, { "name": "Vishwanathan S V N", "gender": "Male", "institution": "University of California - Santa Cruz", "country": "", "position": "Associate Professor" }, { "name": "Wei-Cheng Chang", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Yaushian Wang", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 39, "n_ref": 118, "n_ref_all": 149, "n_fig": 3, "n_tab": 4, "L_tab": 672, "n_element_tab": 84, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 325, "n_element_tab_1": 31, "formula_len_all": 1048, "formula_len_all_1": 657, "len_all": 147847, "len_all_1": 63477, "len_abs": 1731, "len_title": 120, "len_sents": 42300, "len_sents_1": 30410, "n_sents": 316, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1721, "title": "Retrieval-augmented Encoders for Extreme Multi-label Text Classification", "abs": "Extreme multi-label classification (XMC) seeks to find relevant labels from an extremely large label collection for a given text input. To tackle such a vast label space, current state-of-the-art methods fall into two categories. The one-versus-all (OVA) method uses learnable label embeddings for each label, excelling at memorization (i.e., capturing detailed training signals for accurate head label prediction). In contrast, the dual-encoder (DE) model maps input and label text into a shared embedding space for better generalization (i.e., the capability of predicting tail labels with limited training data), but may fall short at memorization. To achieve generalization and memorization, existing XMC methods often combine DE and OVA models, which involves complex training pipelines. Inspired by the success of retrieval-augmented language models, we propose the Retrieval-augmented Encoders for XMC (RAE-XMC), a novel framework that equips a DE model with retrieval-augmented capability for efficient memorization without additional trainable parameter. During training, RAE-XMC is optimized by the contrastive loss over a knowledge memory that consists of both input instances and labels. During inference, given a test input, RAE-XMC retrieves the top-$K$ keys from the knowledge memory, and aggregates the corresponding values as the prediction scores. We showcase the effectiveness and efficiency of RAE-XMC on four public LF-XMC benchmarks. RAE-XMC not only advances the state-of-the-art (SOTA) method DEXML, but also achieves more than 10x speedup on the largest LF-mazonTitles-1.3M dataset under the same 8 A100 GPUs training environments. Our experiment code is available in the Supplementary Material.", "keywords": [ "Extreme Multi-label Text Classification", "Dual-encoders", "Retrieval-Augmented" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "rN7Ewo2lV4", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Alexander Schoenhuth", "gender": "Male", "institution": "Universit\u00e4t Bielefeld", "country": "DE", "position": "Full Professor" }, { "name": "Philip Kenneweg", "gender": "Male", "institution": "Universit\u00e4t Bielefeld", "country": "", "position": "PhD student" }, { "name": "Raghuram Dandinasivara Rangaram", "gender": "Male", "institution": "Universit\u00e4t Bielefeld", "country": "DE", "position": "PhD student" }, { "name": "Xiao Luo", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Full Professor" }, { "name": "Barbara Eva Hammer", "gender": "Female", "institution": "Universit\u00e4t Bielefeld", "country": "DE", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 50, "n_ref_all": 63, "n_fig": 6, "n_tab": 5, "L_tab": 1025, "n_element_tab": 122, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 317, "n_element_tab_1": 36, "formula_len_all": 279, "formula_len_all_1": 280, "len_all": 138842, "len_all_1": 57088, "len_abs": 1027, "len_title": 102, "len_sents": 45009, "len_sents_1": 28487, "n_sents": 313, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1030, "title": "Generating Synthetic Genotypes using Diffusion Models", "abs": "In this paper, we introduce the first diffusion model designed to generate complete synthetic human genotypes, which, by standard protocols, one can straightforwardly expand into full-length, DNA-level genomes.\nThe synthetic genotypes mimic real human genotypes without just reproducing known genotypes, in terms of approved metrics. When training biomedically relevant classifiers with synthetic genotypes, accuracy is near-identical to the accuracy achieved when training classifiers with real data. We further demonstrate that augmenting small amounts of real with synthetically generated genotypes drastically improves performance rates. This addresses a significant challenge in translational human genetics: real human genotypes, although emerging in large volumes from genome wide association studies, are sensitive private data, which limits their public availability. Therefore, the integration of additional, insensitive data when striving for rapid sharing of biomedical knowledge of public interest appears imperative.", "keywords": [ "Diffusion", "Genome", "DNA", "1000 Genome", "ALS" ], "rating_list": [ 5, 1, 5, 8, 10 ], "soundness_list": [ 2, 1, 3, 3, 4 ], "presentation_list": [ 1, 1, 2, 3, 4 ], "contribution_list": [ 2, 1, 2, 3, 4 ], "confidence_list": [ 3, 5, 3, 3, 4 ] }, { "paper_id": "rMyfMS5nMt", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Chunming Zhang", "gender": "Male", "institution": "ict,Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Dongbo Bu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Milong Ren", "gender": "Male", "institution": " Institute of Computing Technology", "country": "CN", "position": "PhD student" }, { "name": "Shiwei Sun", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "Associate Professor" }, { "name": "Tian Zhu", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Xiaoyang Hou", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Xin Gao", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Full Professor" } ], "n_formula": 47, "n_formula_1": 24, "n_ref_uni": 64, "n_ref": 120, "n_ref_all": 138, "n_fig": 4, "n_tab": 10, "L_tab": 4826, "n_element_tab": 509, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1565, "n_element_tab_1": 233, "formula_len_all": 5519, "formula_len_all_1": 2280, "len_all": 210662, "len_all_1": 63564, "len_abs": 924, "len_title": 115, "len_sents": 48060, "len_sents_1": 23753, "n_sents": 397, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 926, "title": "Improving Graph Generation with Flow Matching and Optimal Transport", "abs": "Generating graph-structured data is crucial in various domains but remains challenging due to the complex interdependencies between nodes and edges. While diffusion models have demonstrated their superior generative capabilities, they often suffer from unstable training and inefficient sampling. To enhance generation performance and training stability, we propose GGFlow, a discrete flow matching generative model incorporating optimal transport for graph structures and it incorporates an edge-augmented graph transformer to enable the direct communications among edges. Additionally, GGFlow introduces a novel goal-guided generation framework to control the generative trajectory of our model towards desired properties. GGFlow demonstrates superior performance on both unconditional and conditional generation tasks, outperforming existing baselines and underscoring its effectiveness and potential for wider application.", "keywords": [ "Flow matching generative model", "graph generation" ], "rating_list": [ 1, 6, 6, 8 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 2, 4, 5 ] }, { "paper_id": "rMR2P8e0Zx", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Conghui He", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Jiaqi Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Junhao Huang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "Undergrad student" }, { "name": "Pan Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Rui Qian", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shuangrui Ding", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xiaoyi Dong", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Zihan Liu", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 45, "n_ref": 81, "n_ref_all": 118, "n_fig": 15, "n_tab": 9, "L_tab": 1365, "n_element_tab": 200, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1236, "n_element_tab_1": 166, "formula_len_all": 492, "formula_len_all_1": 300, "len_all": 167477, "len_all_1": 62845, "len_abs": 336, "len_title": 137, "len_sents": 48147, "len_sents_1": 27734, "n_sents": 425, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 2406, "title": "SongComposer: A Large Language Model for Lyric and Melody Composition in Song Generation", "abs": "A song typically comprises the vocal track and the music track. Creating lyrics and melodies for the vocal track in a symbolic format, known as song composition, plays a significant role in the song generation. This delicate and complex task demands expert musical knowledge of melody, an advanced understanding of lyrics, and precise alignment between them. Despite achievements in sub-tasks such as lyric generation, lyric-to-melody, and melody-to-lyric, etc, a unified model for song composition has not yet been achieved. In this paper, we introduce SongComposer, a pioneering step towards a unified song composition model that can readily create symbolic lyrics and melodies following instructions. SongComposer is a music-specialized large language model (LLM) that, for the first time, integrates the capability of simultaneously composing lyrics and melodies into LLMs. To achieve this goal, three non-trivial efforts are introduced. 1) Sheet music understanding, we designed a flexible tuple format to load lyric and note attributes, fostering word-level alignment between lyrics and melodies, and enabling SongComposer to generate lyrics with accompanying well-aligned melodies. 2) Song note tokenizing, the vocabulary of the tokenizer is extended for song notes, and we find a proper scalar-manner initialization of new tokens based on musical prior is essential for the model to understand musical rhythm. 3) Structural music generation, we propose a multi-stage pipeline for progressively capturing the musical structure. Initially, we extract and feed motif-level melody patterns to SongComposer to build its basic generation capabilities. Later, we insert special tokens into the whole-song data to denote phrase-level structure, promoting logical repetition and smooth coherence. Extensive experiments demonstrate that SongComposer outperforms advanced LLMs, including GPT-4, in tasks such as lyric-to-melody generation, melody-to-lyric generation, song continuation, and text-to-song creation. We showcase the generated samples on our anonymous project page https://songcomposer.github.io/. Due to the lack of high-quality symbolic song datasets with lyrics and melodies, we have carefully curated and will publicly release SongCompose, a large-scale song pretraining and supervised finetuning dataset that includes lyrics, melodies, and paired lyrics-melodies in both Chinese and English.", "keywords": [ "LLM", "Symbolic Song Composition" ], "rating_list": [ 5, 5, 5, 3, 5 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "rMOhA1JNPo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Xin Xu", "gender": "Male", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Yu-Xiong Wang", "gender": "unknown", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Ziqi Pang", "gender": "Male", "institution": "UIUC", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 64, "n_ref": 167, "n_ref_all": 219, "n_fig": 16, "n_tab": 12, "L_tab": 4331, "n_element_tab": 300, "n_fig_1": 11, "n_tab_1": 6, "L_tab_1": 3070, "n_element_tab_1": 183, "formula_len_all": 1549, "formula_len_all_1": 1361, "len_all": 227794, "len_all_1": 78779, "len_abs": 2119, "len_title": 150, "len_sents": 62899, "len_sents_1": 32876, "n_sents": 497, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 2136, "title": "Aligning Generative Denoising with Discriminative Objectives Unleashes Diffusion for Visual Perception", "abs": "With success in image generation, generative diffusion models are increasingly adopted for discriminative scenarios because generating pixels is a unified and natural perception interface. Although directly re-purposing their generative denoising process has established promising progress in specialist (e.g., depth estimation) and generalist models, the inherent gaps between a generative process and discriminative objectives are rarely investigated. For instance, generative models can tolerate deviations at intermediate sampling steps as long as the final distribution is reasonable, while discriminative tasks with rigorous ground truth for evaluation are sensitive to such errors. Without mitigating such gaps, diffusion for perception still struggles on tasks represented by multi-modal understanding (e.g., referring image segmentation). Motivated by these challenges, we analyze and improve the alignment between the generative diffusion process and perception objectives centering around the key observation: \\emph{how perception quality evolves with the denoising process}. (1) Notably, earlier denoising steps contribute more than later steps, necessitating a tailored learning objective for training: loss functions should reflect varied contributions of timesteps for each perception task. (2) Perception quality drops unexpectedly at later denoising steps, revealing the sensitiveness of perception to training-denoising distribution shift. We introduce diffusion-tailored data augmentation to simulate such drift in the training data. (3) We suggest a novel perspective to the long-standing question: why should a generative process be useful for discriminative tasks -- interactivity. The denoising process can be leveraged as a controllable user interface adapting to users' correctional prompts and conducting multi-round interaction in an agentic workflow. Collectively, our insights enhance multiple generative diffusion-based perception models without architectural changes: state-of-the-art diffusion-based depth estimator, previously underplayed referring image segmentation models, and perception generalists.", "keywords": [ "Diffusion for Perception", "Diffusion Models", "Visual Perception" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 4, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "rLySPkyl3S", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Jia Wang", "gender": "unknown", "institution": "Waseda University", "country": "", "position": "MS student" }, { "name": "Jing Bai", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Osamu Yoshie", "gender": "Male", "institution": "Waseda University", "country": "JP", "position": "Professor" }, { "name": "Xin Huang", "gender": "Male", "institution": "Waseda University", "country": "JP", "position": "PhD student" }, { "name": "Yeqing Shen", "gender": "Male", "institution": "Stepfun", "country": "CN", "position": "Researcher" }, { "name": "Zheng Ge", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 66, "n_ref_all": 89, "n_fig": 8, "n_tab": 5, "L_tab": 1709, "n_element_tab": 209, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1371, "n_element_tab_1": 158, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 146783, "len_all_1": 61896, "len_abs": 1754, "len_title": 125, "len_sents": 40135, "len_sents_1": 28750, "n_sents": 299, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1768, "title": "Seeking the Right Question: Towards High-Quality Visual Instruction Generation", "abs": "Large language models achieve significant improvements in instruction following through training with synthetic data. The self-instruct method generates instructions based on manually selected tasks, establishing an annotation-free paradigm for synthesizing instructions. However, the experience of synthesizing language instructions for LLMs does not directly transfer to visual instruction generation. Visual instructions encompass both images and questions, and questions generated directly from images often struggle to form high-quality instructions.\nBy analyzing real user queries, we summarize the characteristics of high-quality instructions: they require image perception, reasoning, and answerability. We propose a three-stage visual instruction generation pipeline, named \"Seeking the Right Question\" (SRQ), to produce high-quality instructions. In stage 1, we select 160 instructions that meet high-quality standards as seed questions, categorizing them into eight groups based on multi-modal task types. In stage 2, we introduce capability-driven prompting to generate high-quality questions. In stage 3, we implement an Image Dependency Scoring Mechanism to filter the generated questions. Additionally, we use GPT-4o to directly generate answers, forming $<$image, question, answer$>$ triples for model training.\nTo demonstrate the effectiveness of SRQ, we construct the high-quality instruction dataset Allava-SRQ from 125,000 images sampled from the Allava dataset. Experimental results show that Allava-SRQ significantly improves the performance of multiple baseline models across various benchmarks. We plan to open-source SRQ and the high-quality instruction dataset Allava-SRQ to promote advancements in the field of visual instruction generation.", "keywords": [ "computer vision", "vision language model", "visual instruction generation" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 5, 5, 3 ] }, { "paper_id": "rLlDt2FQvz", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Chen Zhang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Jingshu Peng", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Postdoc" }, { "name": "Lei Chen", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Liping Wang", "gender": "Female", "institution": "The Hong Kong University of Science and Technology, Guangzhou", "country": "CN", "position": "PhD student" }, { "name": "Xujia Li", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Postdoc" }, { "name": "Yan Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Yue Wang", "gender": "unknown", "institution": "Shenzhen Institute of Computing Sciences", "country": "", "position": "Researcher" } ], "n_formula": 39, "n_formula_1": 27, "n_ref_uni": 38, "n_ref": 89, "n_ref_all": 105, "n_fig": 5, "n_tab": 8, "L_tab": 2314, "n_element_tab": 240, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1664, "n_element_tab_1": 173, "formula_len_all": 3823, "formula_len_all_1": 2120, "len_all": 161627, "len_all_1": 79832, "len_abs": 1485, "len_title": 132, "len_sents": 48815, "len_sents_1": 34760, "n_sents": 388, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1484, "title": "N-ForGOT: Towards Not-forgetting and Generalization of Open Temporal Graph Learning", "abs": "Temporal Graph Neural Networks (TGNNs) lay emphasis on capturing node interactions over time but often overlook evolution in node classes and dynamic data distributions triggered by the continuous emergence of new class labels, known as the open-set problem. This problem poses challenges for existing TGNNs in preserving learned classes while rapidly adapting to new, unseen classes. To address this, this paper identifies two primary factors affecting model performance on the open temporal graph, backed by a theoretical guarantee: (1) the forgetting of prior knowledge and (2) distribution discrepancies between successive tasks. Building on these insights, we propose N-ForGOT, which incorporates two plug-in modules into TGNNs to preserve prior knowledge and enhance model generalizability for new classes simultaneously. The first module preserves previously established inter-class connectivity and decision boundaries during the training of new classes to mitigate the forgetting caused by temporal evolutions of class characteristics. The second module introduces an efficient method for measuring distribution discrepancies with designed temporal Weisfeiler-Lehman subtree patterns, effectively addressing both structural and temporal shifts while reducing time complexity. Experimental results on four public datasets demonstrate that our method significantly outperforms state-of-the-art approaches in prediction accuracy, prevention of forgetting, and generalizability.", "keywords": [ "temporal graph neural networks; continual learning; generalization" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 2, 3, 2 ] }, { "paper_id": "rLaMcF516k", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Bolin Ding", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Senior Director" }, { "name": "Jiliang Tang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Full Professor" }, { "name": "Pengfei He", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Yaliang Li", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Staff Engineer" }, { "name": "Yue Xing", "gender": "unknown", "institution": "Michigan State University", "country": "US", "position": "Assistant Professor" }, { "name": "Zitao Li", "gender": "Male", "institution": "Alibaba Group (U.S.)", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 0, "n_ref_uni": 46, "n_ref": 121, "n_ref_all": 169, "n_fig": 5, "n_tab": 8, "L_tab": 3329, "n_element_tab": 317, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 756, "n_element_tab_1": 9, "formula_len_all": 1731, "formula_len_all_1": 0, "len_all": 247634, "len_all_1": 69417, "len_abs": 1530, "len_title": 127, "len_sents": 85775, "len_sents_1": 34660, "n_sents": 709, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1555, "title": "Make LLMs better zero-shot reasoners: structure-oriented autonomous reasoning", "abs": "Zero-shot reasoning methods with Large Language Models (LLMs) offer significant advantages including great generalization to novel tasks and reduced dependency on human-crafted examples. \nHowever, the current zero-shot methods still have limitations in complex tasks, e.g., answering questions that require multi-step reasoning.\nIn this paper, we address this limitation by introducing a novel structure-oriented analysis method to help LLMs better understand the question and guide the problem-solving process of LLMs.\nWe first demonstrate how the existing reasoning strategies, Chain-of-Thought and ReAct, can benefit from our structure-oriented analysis. \nIn addition to empirical investigations, we leverage the probabilistic graphical model to theoretically explain why our structure-oriented analysis can improve the LLM reasoning process. \nTo further improve the reliability in complex question-answering tasks, we propose a multi-agent reasoning system, **S**tructure-oriented **A**utonomous **R**easoning **A**gents (SARA), that can better enforce the reasoning process following our structure-oriented analysis by refinement techniques and is equipped with external knowledge retrieval capability to reduce factual errors.\nExtensive experiments verify the effectiveness of the proposed reasoning system. Surprisingly, in some cases, the system even surpasses few-shot methods.\nFinally, the system not only improves reasoning accuracy in complex tasks but also demonstrates robustness against potential attacks that corrupt the reasoning process.", "keywords": [ "language model", "reasoning", "agents" ], "rating_list": [ 3, 6, 8, 3, 3 ], "soundness_list": [ 2, 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "rLX7Vyyzus", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Jinqiao Wang", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Ming Tang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Tao Yu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xu Zhao", "gender": "Male", "institution": " Institute of automation, Chinese academy of sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Yongqi An", "gender": "Male", "institution": ", Institute of automation, Chinese academy of science", "country": "CN", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 9, "n_ref_uni": 25, "n_ref": 44, "n_ref_all": 73, "n_fig": 17, "n_tab": 8, "L_tab": 1431, "n_element_tab": 129, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1541, "formula_len_all_1": 927, "len_all": 138051, "len_all_1": 47774, "len_abs": 1512, "len_title": 93, "len_sents": 46155, "len_sents_1": 20982, "n_sents": 357, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 44, "L_abs": 1528, "title": "Systematic Outliers in Large Language Models", "abs": "Outliers have been widely observed in Large Language Models (LLMs), significantly impacting model performance and posing challenges for model compression. Understanding the functionality and formation mechanisms of these outliers is critically important. Existing works, however, largely focus on reducing the impact of outliers from an algorithmic perspective, lacking an in-depth investigation into their causes and roles. In this work, we provide a detailed analysis of the formation process, underlying causes, and functions of outliers in LLMs. We define and categorize three types of outliers\u2014activation outliers, weight outliers, and attention outliers\u2014and analyze their distributions across different dimensions, uncovering inherent connections between their occurrences and their ultimate influence on the attention mechanism. Based on these observations, we hypothesize and explore the mechanisms by which these outliers arise and function, demonstrating through theoretical derivations and experiments that they emerge due to the self-attention mechanism's softmax operation. These outliers act as implicit context-aware scaling factors within the attention mechanism. As these outliers stem from systematic influences, we term them systematic outliers. Our study not only enhances the understanding of Transformer-based LLMs but also shows that structurally eliminating outliers can accelerate convergence and improve model compression. The code is avilable at \\url{https://github.com/an-yongqi/systematic-outliers}.", "keywords": [ "Model Interpretability", "Large Language Models", "Outliers", "Attention Mechanism" ], "rating_list": [ 5, 6, 3, 8, 8 ], "soundness_list": [ 4, 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 4, 4 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 4, 3 ] }, { "paper_id": "rKZSatPN3W", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 9, "author_info_list": [ { "name": "Bo Wang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Dongsheng Li", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Guanzhou Ke", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "XINYANG JIANG", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Researcher" }, { "name": "Xiao-Li Wang", "gender": "Female", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Xufang Luo", "gender": "Female", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" }, { "name": "Yang Yu", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "", "position": "Full Professor" }, { "name": "Yifan Yang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Zilong Wang", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 40, "n_ref": 120, "n_ref_all": 146, "n_fig": 8, "n_tab": 7, "L_tab": 2090, "n_element_tab": 311, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 288, "n_element_tab_1": 35, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 160268, "len_all_1": 61302, "len_abs": 1376, "len_title": 111, "len_sents": 44850, "len_sents_1": 29920, "n_sents": 366, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1383, "title": "Is classification all you need for radiology report generation?", "abs": "Automatic radiology report generation is an advanced medical assistive technology capable of producing coherent reports based on medical images, akin to a radiologist. However, current generative methods exhibit a notable gap in clinical metrics when compared to medical image classification. Recently, leveraging diagnostic results to improve report quality has emerged as a promising approach. We are curious whether training a classifier that encompasses all possible long-tailed and rare diseases could enhance the robustness of reports. To investigate this question, this study designs an evaluation framework that integrates long-tail scenarios and summarizes potential combinations of LLM-based report generation models. We assess the impact of classification on report quality across four benchmarks. Initially, we introduce LLM-based language and clinical metrics and develop a pipeline to evaluate the model's performance on both in-domain and out-of-distribution (OOD) long-tail scenarios. Subsequently, we conduct a systematic evaluation of all potential model combinations. Our findings reveal that: 1) the impact of classification on report quality is positively correlated with the performance of classifiers, but the gap still exists, and 2) while classification can enhance report quality in in-domain long-tail scenarios, its benefits for OOD scenarios are limited.", "keywords": [ "radiology report generation", "large language model", "multi-modalities" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "rKMz6cDE7W", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Chenyang Li", "gender": "Male", "institution": "Fuzhou University", "country": "CN", "position": "Undergrad student" }, { "name": "Chiwun Yang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "Raghavendra Addanki", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" } ], "n_formula": 49, "n_formula_1": 33, "n_ref_uni": 68, "n_ref": 112, "n_ref_all": 117, "n_fig": 1, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5013, "formula_len_all_1": 4266, "len_all": 143424, "len_all_1": 50240, "len_abs": 1504, "len_title": 140, "len_sents": 22232, "len_sents_1": 17645, "n_sents": 223, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1625, "title": "One Pass Streaming Algorithm for Super Long Token Attention Approximation in Sublinear Space", "abs": "Attention computation takes both the time complexity of $O(n^2)$ and the space complexity of $O(n^2)$ simultaneously, which makes deploying Large Language Models (LLMs) in streaming applications that involve long contexts requiring substantial computational resources. In recent OpenAI DevDay (Nov 6, 2023), OpenAI released a new model that is able to support a 128K-long document, in our paper, we focus on the memory-efficient issue when context length $n$ is much greater than 128K ($n \\gg 2^d$). Considering a single-layer self-attention with Query, Key, and Value matrices $Q, K, V \\in \\mathbb{R}^{n \\times d}$, the polynomial method approximates the attention output $T \\in \\mathbb{R}^{n \\times d}$. It accomplishes this by constructing $U_1, U_2 \\in \\mathbb{R}^{n \\times t}$ to expedite attention ${\\sf Attn}(Q, K, V)$ computation within $n^{1+o(1)}$ time executions. Despite this, computing the approximated attention matrix $U_1U_2^\\top \\in \\mathbb{R}^{n \\times n}$ still necessitates $O(n^2)$ space, leading to significant memory usage. In response to these challenges, we introduce a new algorithm that only reads one pass of the data in a streaming fashion. This method employs sublinear space $o(n)$ to store three sketch matrices, alleviating the need for exact $K, V$ storage. Notably, our algorithm exhibits exceptional memory-efficient performance with super-long tokens. As the token length $n$ increases, our error guarantee diminishes while the memory usage remains nearly constant. This unique attribute underscores the potential of our technique in efficiently handling LLMs in streaming applications.", "keywords": [ "streaming algorithm", "efficient attention", "super-long context" ], "rating_list": [ 3, 1, 3 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 1, 1, 1 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "rK0YJwL69S", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Amartya Sanyal", "gender": "Male", "institution": "Copenhagen University", "country": "DK", "position": "Assistant Professor" }, { "name": "David Krueger", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Assistant Professor" }, { "name": "Neel Alex", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Shoaib Ahmed Siddiqui", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 60, "n_ref": 130, "n_ref_all": 151, "n_fig": 6, "n_tab": 5, "L_tab": 6519, "n_element_tab": 448, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 616, "n_element_tab_1": 15, "formula_len_all": 99, "formula_len_all_1": 99, "len_all": 200292, "len_all_1": 62527, "len_abs": 1108, "len_title": 102, "len_sents": 55199, "len_sents_1": 29655, "n_sents": 412, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1122, "title": "Protecting against simultaneous data poisoning attacks", "abs": "Current backdoor defense methods are evaluated against a single attack at a time. This is unrealistic, as powerful machine learning systems are trained on large datasets scraped from the internet, which may be attacked multiple times by one or more attackers. We demonstrate that multiple backdoors can be simultaneously installed in a single model through parallel data poisoning attacks without substantially degrading clean accuracy. Furthermore, we show that existing backdoor defense methods do not effectively defend against multiple simultaneous attacks. Finally, we leverage insights into the nature of backdoor attacks to develop a new defense, BaDLoss (**Ba**ckdoor **D**etection via **Loss** Dynamics), that is effective in the multi-attack setting. With minimal clean accuracy degradation, BaDLoss attains an average attack success rate in the multi-attack setting of 7.98% in CIFAR-10, 10.29% in GTSRB, and 19.17% in Imagenette, compared to the average of other defenses at 63.44%, 74.83%, and 41.74% respectively. BaDLoss scales to ImageNet-1k, reducing the average attack success rate from 88.57% to 15.61%.", "keywords": [ "backdoors", "backdoor defenses", "data poisoning" ], "rating_list": [ 3, 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "rJ5g8ueQaI", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Jongmin Lee", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Meiqi Sun", "gender": "Female", "institution": "University of California, Berkeley", "country": "", "position": "Undergrad student" }, { "name": "Pieter Abbeel", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" } ], "n_formula": 65, "n_formula_1": 21, "n_ref_uni": 43, "n_ref": 114, "n_ref_all": 159, "n_fig": 10, "n_tab": 2, "L_tab": 210, "n_element_tab": 19, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5462, "formula_len_all_1": 2246, "len_all": 191472, "len_all_1": 69280, "len_abs": 880, "len_title": 146, "len_sents": 51094, "len_sents_1": 30931, "n_sents": 391, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 882, "title": "SEMDICE: Off-policy State Entropy Maximization via Stationary Distribution Correction Estimation", "abs": "In the unsupervised pre-training for reinforcement learning, the agent aims to learn a prior policy for downstream tasks without relying on task-specific reward functions. We focus on state entropy maximization (SEM), where the goal is to learn a policy that maximizes the entropy of the state's stationary distribution. In this paper, we introduce SEMDICE, a principled off-policy algorithm that computes an SEM policy from an arbitrary off-policy dataset, which optimizes the policy directly within the space of stationary distributions. SEMDICE computes a single, stationary Markov state-entropy-maximizing policy from an arbitrary off-policy dataset. Experimental results demonstrate that SEMDICE outperforms baseline algorithms in maximizing state entropy while achieving the best adaptation efficiency for downstream tasks among SEM-based unsupervised RL pre-training methods.", "keywords": [ "state entropy maximization", "unsupervised reinforcement learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 3, 4, 2 ] }, { "paper_id": "rJ1xGcJVu8", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Bowen Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Fugee Tsung", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Jiaxi Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yuxuan Liang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Qingsong Wen", "gender": "Male", "institution": "Squirrel Ai Learning", "country": "US", "position": "Principal Researcher" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 51, "n_ref": 99, "n_ref_all": 130, "n_fig": 11, "n_tab": 12, "L_tab": 14434, "n_element_tab": 1900, "n_fig_1": 10, "n_tab_1": 6, "L_tab_1": 3568, "n_element_tab_1": 460, "formula_len_all": 665, "formula_len_all_1": 725, "len_all": 192106, "len_all_1": 78514, "len_abs": 1002, "len_title": 91, "len_sents": 45659, "len_sents_1": 32017, "n_sents": 336, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 976, "title": "Toward Physics-guided Time Series Embedding", "abs": "In various scientific and engineering fields, the primary research areas have revolved around physics-based dynamical systems modeling and data-driven time series analysis. According to the embedding theory, dynamical systems and time series can be mutually transformed using observation functions and physical reconstruction techniques. Based on this, we propose Embedding Duality Theory, where the parameterized embedding layer essentially provides a linear estimation of the non-linear time series dynamics. This theory enables us to bypass the parameterized embedding layer and directly employ physical reconstruction techniques to acquire a data embedding representation. Utilizing physical priors results in a 10$\\times$ reduction in parameters, a 3$\\times$ increase in speed, and maximum performance enhancements of 18\\% in expert, 22\\% in zero-shot, and 53\\% in few-shot tasks without any hyper-parameter tuning. All methods are encapsulated as a plug-and-play module.", "keywords": [ "Time Series Analysis", "Dynamic System" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 2, 1, 1, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "rIJbFQ1zII", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Jie Hao", "gender": "Male", "institution": "George Mason University", "country": "US", "position": "PhD student" }, { "name": "Mingrui Liu", "gender": "unknown", "institution": "George Mason University", "country": "", "position": "Assistant Professor" }, { "name": "Xiaochuan Gong", "gender": "Male", "institution": "George Mason University", "country": "US", "position": "PhD student" } ], "n_formula": 348, "n_formula_1": 23, "n_ref_uni": 50, "n_ref": 204, "n_ref_all": 250, "n_fig": 5, "n_tab": 12, "L_tab": 1427, "n_element_tab": 80, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 182, "n_element_tab_1": 17, "formula_len_all": 56670, "formula_len_all_1": 2889, "len_all": 359616, "len_all_1": 73846, "len_abs": 1803, "len_title": 144, "len_sents": 93930, "len_sents_1": 33004, "n_sents": 1039, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1580, "title": "On the Convergence of Adam-Type Algorithms for Bilevel Optimization under Unbounded Smoothness", "abs": "Adam has become one of the most popular optimizers for training modern deep neural networks, such as transformers. However, its applicability is largely restricted to single-level optimization problems. In this paper, we aim to extend vanilla Adam to tackle bilevel optimization problems, which have important applications in machine learning, such as meta-learning. In particular, we study stochastic bilevel optimization problems where the lower-level function is strongly convex and the upper-level objective is nonconvex with potentially unbounded smoothness. This unbounded smooth objective function covers a broad class of neural networks, including transformers, which may exhibit non-Lipschitz gradients. In this work, we first introduce AdamBO, a single-loop Adam-type method that achieves $\\widetilde{O}(\\epsilon^{-4})$ oracle complexity to find $\\epsilon$-stationary points, where the oracle calls involve stochastic gradient or Hessian/Jacobian-vector product evaluations. The key to our analysis is a novel randomness decoupling lemma that provides refined control over the lower-level variable. Additionally, we propose VR-AdamBO, a variance-reduced version with an improved oracle complexity of $\\widetilde{O}(\\epsilon^{-3})$. The improved analysis is based on a novel stopping time approach and a careful treatment of the lower-level error. We conduct extensive experiments on various machine learning tasks involving bilevel formulations with recurrent neural networks (RNNs) and transformers, demonstrating the effectiveness of our proposed Adam-type algorithms.", "keywords": [ "Bilevel Optimization", "Adam", "Unbounded Smoothness" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 2, 2, 4, 4 ] }, { "paper_id": "rHbxQebhDd", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Aniket Biswal", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "MS student" }, { "name": "Fatma Gzara", "gender": "unknown", "institution": "University of Waterloo", "country": "CA", "position": "Full Professor" }, { "name": "Sirisha Rambhatla", "gender": "Female", "institution": "University of Waterloo", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 17, "n_ref_uni": 41, "n_ref": 60, "n_ref_all": 75, "n_fig": 5, "n_tab": 5, "L_tab": 2819, "n_element_tab": 200, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 798, "n_element_tab_1": 82, "formula_len_all": 1258, "formula_len_all_1": 1231, "len_all": 126459, "len_all_1": 61305, "len_abs": 1763, "len_title": 137, "len_sents": 38059, "len_sents_1": 27426, "n_sents": 319, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1373, "title": "Uncertainty Aware Column Generation for Crew Pairing Optimization Using Survival Analysis", "abs": "The crew pairing problem (CPP) is central to optimal planning and scheduling of operations in the airline industry, where the objective is to assign crews to cover a flight schedule at minimal cost while adhering to various logistical, personnel, and policy constraints. Despite the implementation of optimized schedules, operations are frequently disrupted by unforeseen events. This vulnerability stems from the deterministic nature of the CPP's base formulation, which fails to account for the uncertainties inherent in real-world operations. Existing solutions either aim to safeguard against a specified level of uncertainty or focus on worst-case scenarios. To this end, we propose a reliability-centric CPP formulation amenable to solution by column-generation (CG) SurvCG, that leverages survival analysis for dynamic quantification of uncertainty using the operation patterns in historical data. Applied to CPP, SurvCG forecasts and incorporates flight connection reliability into the optimization process. Through rigorous experiments on a large-scale first-of-its-kind real-world instance under regular and irregular operating conditions, we demonstrate that SurvCG achieves unprecedented improvements (up to 61%) over baseline in terms of total propagated delays, establishing SurvCG as the first data-driven solution for uncertainty-aware reliable scheduling.", "keywords": [ "Reliable Scheduling and Planning", "Column Generation", "Machine learning for optimization", "Survival Analysis", "Crew pairing", "Airline Operations Planning" ], "rating_list": [ 1, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "rH6IZIXqZG", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Mingyuan Zhou", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Yi Gu", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Yueqin Yin", "gender": "Female", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Yujia Xie", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Zhendong Wang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 13, "n_ref_uni": 32, "n_ref": 86, "n_ref_all": 109, "n_fig": 6, "n_tab": 4, "L_tab": 908, "n_element_tab": 112, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 342, "n_element_tab_1": 50, "formula_len_all": 5100, "formula_len_all_1": 1226, "len_all": 135868, "len_all_1": 59197, "len_abs": 470, "len_title": 111, "len_sents": 38862, "len_sents_1": 26789, "n_sents": 273, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1066, "title": "Diffusion Preference Alignment via Relative Text-Image Contrast", "abs": "Aligning Large Language Models (LLMs) to human preferences has become a prominent area of research within language modeling. However, the application of preference learning to image generation in Text-to-Image (T2I) models remains relatively unexplored. One approach, Diffusion-DPO, initially experimented with pairwise preference learning in diffusion models for individual text prompts. We propose Diff-contrast, a novel method designed to align diffusion-based T2I models with human preferences. This method utilizes both prompt-image pairs with identical prompts and those that are semantically related across different modalities. Additionally, we introduced a new evaluation task, style alignment, to address the issues of high cost, low reproducibility, and poor interpretability associated with current evaluations of human preference alignment. Our results show that Diff-contrast surpasses existing techniques, e.g. Diffusion-DPO, in tuning Stable Diffusion versions 1.5 and XL-1.0 across both automated evaluations of human preference and style alignment.", "keywords": [ "Diffusion Models", "Human Preference Alignment", "Fine-tuning" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 3, 3, 1 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 3, 2, 1 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "rGyi8NNqB0", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Jakub Marecek", "gender": "Male", "institution": "Czech Technical University", "country": "CZ", "position": "Tenured faculty" }, { "name": "Ji\u0159\u00ed N\u011bme\u010dek", "gender": "Male", "institution": "Czech Technical Univeresity in Prague, Czech Technical University of Prague", "country": "CZ", "position": "PhD student" }, { "name": "Tom\u00e1\u0161 Pevn\u00fd", "gender": "Male", "institution": "Czech Technical University in Prague", "country": "CZ", "position": "Associate Professor" } ], "n_formula": 28, "n_formula_1": 16, "n_ref_uni": 44, "n_ref": 99, "n_ref_all": 127, "n_fig": 5, "n_tab": 13, "L_tab": 4415, "n_element_tab": 328, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 987, "n_element_tab_1": 138, "formula_len_all": 2490, "formula_len_all_1": 1038, "len_all": 202240, "len_all_1": 68150, "len_abs": 936, "len_title": 108, "len_sents": 65532, "len_sents_1": 29964, "n_sents": 594, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 942, "title": "Generating Likely Counterfactuals Using Sum-Product Networks", "abs": "The need to explain decisions made by AI systems is driven by both recent regulation and user demand. The decisions are often explainable only post hoc. In counterfactual explanations, one may ask what constitutes the best counterfactual explanation. Clearly, multiple criteria must be taken into account, although \"distance from the sample\" is a key criterion. Recent methods that consider the plausibility of a counterfactual seem to sacrifice this original objective. Here, we present a system that provides high-likelihood explanations that are, at the same time, close and sparse. We show that the search for the most likely explanations satisfying many common desiderata for counterfactual explanations can be modeled using Mixed-Integer Optimization (MIO). We use a Sum-Product Network (SPN) to estimate the likelihood of a counterfactual. To achieve that, we propose an MIO formulation of an SPN, which can be of independent interest.", "keywords": [ "counterfactual explanations", "mixed-integer optimization", "sum-product networks" ], "rating_list": [ 8, 5, 5, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 4 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "rGP2jbWt0l", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Chase Walker", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "PhD student" }, { "name": "Rickard Ewetz", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "Associate Professor" }, { "name": "Sumit Kumar Jha", "gender": "unknown", "institution": "Florida International University", "country": "US", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 17, "n_ref_uni": 27, "n_ref": 95, "n_ref_all": 147, "n_fig": 14, "n_tab": 13, "L_tab": 3452, "n_element_tab": 333, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1280, "n_element_tab_1": 105, "formula_len_all": 746, "formula_len_all_1": 706, "len_all": 158650, "len_all_1": 70334, "len_abs": 1368, "len_title": 99, "len_sents": 50851, "len_sents_1": 30497, "n_sents": 416, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1505, "title": "Metric-Driven Attributions for Vision Transformers", "abs": "Attribution algorithms explain computer vision models by attributing the model response to pixels within the input. Existing attribution methods generate explanations by combining transformations of internal model representations such as class activation maps, gradients, attention, or relevance scores. The effectiveness of an attribution map is measured using attribution quality metrics. This leads us to pose the following question: if attribution methods are assessed using attribution quality metrics, why are the metrics not used to generate the attributions? In response to this question, we propose a Metric-Driven Attribution for explaining Vision Transformers (ViT) called MDA. Guided by attribution quality metrics, the method creates attribution maps by performing patch order and patch magnitude optimization across all patch tokens. The first step orders the patches in terms of importance and the second step assigns the magnitude to each patch while preserving the patch order. Moreover, MDA can provide a smooth trade-off between sparse and dense attributions by modifying the optimization objective. Experimental evaluation demonstrates the proposed MDA method outperforms $7$ existing ViT attribution methods by an average of $12\\%$ across $12$ attribution metrics on the ImageNet dataset for the ViT-base $16 \\times 16$, ViT-tiny $16 \\times 16$, and ViT-base $32 \\times 32$ models. Code is publicly available at https://github.com/chasewalker26/MDA-Metric-Driven-Attributions-for-ViT.", "keywords": [ "Interpretable and Explainable AI", "Computer Vision", "Vision Transformer" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 2, 2, 4, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "rGGwXo0Fo0", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Pin-Yu Chen", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Wenqi Wei", "gender": "unknown", "institution": "Fordham University", "country": "US", "position": "Assistant Professor" }, { "name": "Xiang Li", "gender": "Male", "institution": "Fordham University", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 96, "n_ref_all": 129, "n_fig": 5, "n_tab": 15, "L_tab": 18864, "n_element_tab": 3065, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3594, "n_element_tab_1": 541, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 221932, "len_all_1": 72387, "len_abs": 4541, "len_title": 109, "len_sents": 47214, "len_sents_1": 30759, "n_sents": 331, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1723, "title": "SONAR: A Synthetic AI-Audio Detection Framework and Benchmark", "abs": "Recent advances in Text-to-Speech (TTS) and Voice-Conversion (VC) using generative Artificial Intelligence (AI) technology have made it possible to generate high-quality and realistic human-like audio. This introduces significant challenges to distinguishing AI-synthesized speech from the authentic human voice and could raise potential issues of misuse for malicious purposes such as impersonation and fraud, spreading misinformation, deepfakes, and scams. However, existing detection techniques for AI-synthesized audio have not kept pace and often exhibit poor generalization across diverse datasets. In this paper, we introduce SONAR, a **s**ynthetic AI-Audi**o** Detectio**n** Fr**a**mework and Benchma**r**k, aiming to provide a comprehensive evaluation for distinguishing cutting-edge AI-synthesized auditory content. SONAR includes a novel evaluation dataset sourced from 9 diverse audio synthesis platforms, including leading TTS providers and state-of-the-art TTS models. It is the first framework to uniformly benchmark AI-audio detection across both traditional and foundation model-based deepfake detection systems. Through extensive experiments, we reveal the generalization limitations of existing detection methods and demonstrate that foundation models exhibit stronger generalization capabilities, which can be attributed to their model size and the scale and quality of pretraining data. Additionally, we explore the effectiveness and efficiency of few-shot fine-tuning in improving generalization, highlighting its potential for tailored applications, such as personalized detection systems for specific entities or individuals. Code and dataset are available at https://anonymous.4open.science/r/SONAR", "keywords": [ "Audio deepfake detection", "benchmarking" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "rFpZnn11gj", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Chenglu Zhu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Researcher" }, { "name": "Jingxiong Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Kai Zhang", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Lin Yang", "gender": "Male", "institution": "Westlake University ", "country": "CN", "position": "Full Professor" }, { "name": "Tao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "XINHENG LYU", "gender": "Male", "institution": "The University of Nottingham Ningbo China", "country": "CN", "position": "PhD student" }, { "name": "Xuan Gong", "gender": "Female", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Yixuan Si", "gender": "Male", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "Undergrad student" }, { "name": "Yunlong Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Yuxuan Sun", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Zhongyi Shui", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 73, "n_ref_all": 110, "n_fig": 17, "n_tab": 13, "L_tab": 5648, "n_element_tab": 643, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2615, "n_element_tab_1": 382, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 292246, "len_all_1": 57439, "len_abs": 6125, "len_title": 145, "len_sents": 93450, "len_sents_1": 24257, "n_sents": 731, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 97, "L_abs": 1769, "title": "PathGen-1.6M: 1.6 Million Pathology Image-text Pairs Generation through Multi-agent Collaboration", "abs": "Vision Language Models (VLMs) like CLIP have attracted substantial attention in pathology, serving as backbones for applications such as zero-shot image classification and Whole Slide Image (WSI) analysis. Additionally, they can function as vision encoders when combined with large language models (LLMs) to support broader capabilities. Current efforts to train pathology VLMs rely on pathology image-text pairs from platforms like PubMed, YouTube, and Twitter, which provide limited, unscalable data with generally suboptimal image quality. In this work, we leverage large-scale WSI datasets like TCGA to extract numerous high-quality image patches. We then train a large multimodal model to generate captions for these images, creating PathGen-1.6M, a dataset containing 1.6 million high-quality image-caption pairs. Our approach involves multiple agent models collaborating to extract representative WSI patches, generating and refining captions to obtain high-quality image-text pairs. Extensive experiments show that integrating these generated pairs with existing datasets to train a pathology-specific CLIP model, PathGen-CLIP, significantly enhances its ability to analyze pathological images, with substantial improvements across nine pathology-related zero-shot image classification tasks and three whole-slide image tasks. Furthermore, we construct 200K instruction-tuning data based on PathGen-1.6M and integrate PathGen-CLIP with the Vicuna LLM to create more powerful multimodal models through instruction tuning. Overall, we provide a scalable pathway for high-quality data generation in pathology, paving the way for next-generation general pathology models. Our dataset, code, and model are open-access at https://github.com/PathGen-1-6M/PathGen-1.6M.", "keywords": [ "Image-text pairs generation", "Vision-language models", "Multi-agent collaboration" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "rFPtqF6eaR", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Mehmet Akcakaya", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "Associate Professor" }, { "name": "Merve Gulle", "gender": "Female", "institution": "University of Minnesota - Twin Cities", "country": "", "position": "PhD student" }, { "name": "Yasar Utku Alcalar", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 48, "n_ref": 99, "n_ref_all": 121, "n_fig": 12, "n_tab": 2, "L_tab": 411, "n_element_tab": 77, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 329, "n_element_tab_1": 60, "formula_len_all": 781, "formula_len_all_1": 622, "len_all": 138573, "len_all_1": 57894, "len_abs": 2050, "len_title": 148, "len_sents": 40767, "len_sents_1": 28380, "n_sents": 266, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1990, "title": "Training Physics-Driven Deep Learning Reconstruction without Raw Data Access for Equitable Fast MRI", "abs": "Physics-driven deep learning (PD-DL) approaches have become popular for improved reconstruction of fast magnetic resonance imaging (MRI) scans. Even though PD-DL offers higher acceleration rates compared to existing clinical fast MRI techniques, their use has been limited outside specialized MRI centers. One impediment for their deployment is the difficulties with generalization to pathologies or population groups that are not well-represented in training sets. This has been noted in several studies, and fine-tuning on target populations to improve reconstruction has been suggested. However, current training approaches for PD-DL training require access to raw k-space measurements, which is typically only available at specialized MRI centers that have research agreements for such data access. This is especially an issue for rural and underserved areas, where commercial MRI scanners only provide access to a final reconstructed image. To tackle these challenges, we propose CUPID for high-quality PD-DL training, using only routine clinical reconstructed images exported from an MRI scanner. CUPID evaluates the goodness of the output with a compressibility-based approach, while ensuring that the output stays consistent with the clinical parallel imaging reconstruction through well-designed perturbations. Our results show that CUPID achieves similar quality compared to well-established PD-DL training strategies that require raw k-space data access, while outperforming conventional compressed sensing (CS) and state-of-the-art generative methods. We also demonstrate its effectiveness in a zero-shot training setup for retrospectively and prospectively sub-sampled acquisitions, attesting to its minimal training burden. As an approach that radically deviates from existing strategies, CUPID presents an opportunity to provide equitable access to fast MRI for underserved populations in an attempt to reduce the inequalities associated with this expensive imaging modality.", "keywords": [ "Computational Imaging", "Fast MRI", "Unsupervised Learning", "Compressed Sensing", "Deep Learning", "Equity" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "rF0wXBpFRT", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Mineui Hong", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "PhD student" }, { "name": "Minjae Kang", "gender": "unknown", "institution": "Seoul National University", "country": "", "position": "PhD student" }, { "name": "Songhwai Oh", "gender": "unknown", "institution": "Seoul National University", "country": "", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 38, "n_ref": 68, "n_ref_all": 100, "n_fig": 10, "n_tab": 12, "L_tab": 2990, "n_element_tab": 305, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 107, "n_element_tab_1": 11, "formula_len_all": 980, "formula_len_all_1": 401, "len_all": 166910, "len_all_1": 69054, "len_abs": 1202, "len_title": 160, "len_sents": 54410, "len_sents_1": 32506, "n_sents": 441, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 112, "L_abs": 1206, "title": "Playbook: Scalable Discrete Skill Discovery from Unstructured Datasets for Long-Horizon Decision-Making Problems", "abs": "Skill discovery methods equip an agent with diverse skills necessary for solving challenging tasks through an unsupervised learning manner. However, making the pre-learned skills expandable for new tasks remains a challenge in existing research. To handle this limitation, we propose a scalable skill discovery algorithm, a playbook, which can accommodate unseen tasks by training new skills while maintaining previously learned ones. The playbook, characterized by discrete skills and an extendable structure, enables the extension of the skill set to cover new datasets. Since we design the playbook to have a finite number of skills, we can interpret a decision-making problem as a sequential skill classification problem, so we aim to learn additional skills of the playbook by applying the techniques of class-incremental learning. In addition, we also introduce skill planning schemes that can leverage both previously and newly learned skills to solve challenging tasks compounded by multiple sub-tasks. The proposed method is evaluated in the complex robotic manipulation benchmarks, and the results show that the playbook outperforms existing state-of-the-art methods that learn continuous skills.", "keywords": [ "skill discovery", "multi-task decision-making problem", "offline reinforcement learning", "hierarchical reinforcement learning" ], "rating_list": [ 8, 6, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 4, 2, 1 ], "confidence_list": [ 2, 3, 3 ] }, { "paper_id": "rEqETC88RY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 1, "author_info_list": [ { "name": "Yang Li", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 14, "n_ref_uni": 50, "n_ref": 92, "n_ref_all": 105, "n_fig": 4, "n_tab": 1, "L_tab": 746, "n_element_tab": 83, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1171, "formula_len_all_1": 931, "len_all": 172556, "len_all_1": 68894, "len_abs": 1147, "len_title": 132, "len_sents": 56511, "len_sents_1": 34588, "n_sents": 419, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1150, "title": "LLM Bandit: Cost-Efficient LLM Generation via Preference-Conditioned Dynamic Routing", "abs": "The rapid advancement in large language models (LLMs) has brought forth a diverse range of models with varying capabilities that excel in different tasks and domains. However, selecting the optimal LLM for user queries often involves a challenging trade-off between accuracy and cost, a problem exacerbated by the diverse demands of individual queries. In this work, we present a novel framework that formulates the LLM selection process as a multi-armed bandit problem, enabling dynamic and intelligent routing of queries to the most appropriate model. Our approach incorporates a preference-conditioned dynamic routing mechanism, allowing users to specify their preferences at inference time, thereby offering a customizable balance between performance and cost. Additionally, our selection policy is designed to generalize to unseen LLMs, ensuring adaptability to new models as they emerge. Experimental results demonstrate that our method achieves significant improvements in both accuracy and cost-effectiveness across various LLM platforms, showcasing the potential of our framework to adaptively optimize LLM selection in real-world scenarios.", "keywords": [ "LLM Routing", "Reinforcement Learning", "Multi-objective Optimization" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "rEnPEIwXrB", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Andrea Bejarano-Carbo", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Changwoo Lee", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "David Blaauw", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Full Professor" }, { "name": "Dennis Sylvester", "gender": "unknown", "institution": "University of Michigan", "country": "", "position": "Full Professor" }, { "name": "Hun-Seok Kim", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "Associate Professor" }, { "name": "Pierre Abillama", "gender": "Male", "institution": "University of Michigan", "country": "US", "position": "PhD student" }, { "name": "Qirui Zhang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Postdoc" } ], "n_formula": 19, "n_formula_1": 13, "n_ref_uni": 41, "n_ref": 108, "n_ref_all": 134, "n_fig": 6, "n_tab": 8, "L_tab": 3385, "n_element_tab": 510, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3556, "n_element_tab_1": 391, "formula_len_all": 2579, "formula_len_all_1": 1649, "len_all": 151095, "len_all_1": 65221, "len_abs": 1499, "len_title": 134, "len_sents": 42547, "len_sents_1": 27492, "n_sents": 284, "n_sents_1": 157, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1523, "title": "One-Hot Multi-Level LIF Spiking Neural Networks for Enhanced Accuracy-Latency Tradeoff", "abs": "Spiking neural networks (SNNs) hold significant promise as energy-efficient alternatives to conventional artificial neural networks (ANNs). However, SNNs require computations across multiple timesteps, resulting in increased latency, heightened energy consumption, and additional memory access overhead. Techniques to reduce SNN latency down to a unit timestep have emerged to realize true superior energy efficiency over ANNs. Nonetheless, this latency reduction often comes at the expense of noticeable accuracy degradation. Therefore, achieving an optimal balance in the tradeoff between accuracy and energy consumption by adjusting the latency of multiple timesteps remains a significant challenge. In this paper, we introduce a new dimension to the accuracy-energy tradeoff space using a novel one-hot multi-level leaky integrate-and-fire (M-LIF) neuron model. The proposed M-LIF model represents the inputs and outputs of hidden layers as a set of one-hot binary-weighted spike lanes to find better tradeoff points while still being able to model conventional SNNs. For image classification on static datasets, we demonstrate M-LIF SNNs outperform iso-architecture conventional LIF SNNs in terms of accuracy ($2$% higher than VGG16 SNN on ImageNet) while still being energy-efficient ($20\\times$ lower energy than VGG16 ANN on ImageNet). For dynamic vision datasets, we demonstrate the ability of M-LIF SNNs to reduce latency by $3\\times$ compared to conventional LIF SNNs while limiting accuracy degradation ($<1$%).", "keywords": [ "spiking neural networks", "leaky integrate-and-fire", "energy-efficient", "low latency" ], "rating_list": [ 3, 5, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 1 ], "confidence_list": [ 5, 5, 4, 3, 4 ] }, { "paper_id": "rEQqBZIz49", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Tongzhou Liao", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Barnabas Poczos", "gender": "unknown", "institution": "School of Computer Science", "country": "", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 12, "n_ref_uni": 52, "n_ref": 105, "n_ref_all": 124, "n_fig": 4, "n_tab": 8, "L_tab": 3307, "n_element_tab": 241, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 4903, "n_element_tab_1": 246, "formula_len_all": 1178, "formula_len_all_1": 915, "len_all": 137563, "len_all_1": 68540, "len_abs": 821, "len_title": 116, "len_sents": 38128, "len_sents_1": 27796, "n_sents": 299, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 824, "title": "Greener GRASS: Enhancing GNNs with Encoding, Rewiring, and Attention", "abs": "Graph Neural Networks (GNNs) have become important tools for machine learning on graph-structured data. In this paper, we explore the synergistic combination of graph encoding, graph rewiring, and graph attention, by introducing Graph Attention with Stochastic Structures (GRASS), a novel GNN architecture. GRASS utilizes relative random walk probabilities (RRWP) encoding and a novel decomposed variant (D-RRWP) to efficiently capture structural information. It rewires the input graph by superimposing a random regular graph to enhance long-range information propagation. It also employs a novel additive attention mechanism tailored for graph-structured data. Our empirical evaluations demonstrate that GRASS achieves state-of-the-art performance on multiple benchmark datasets, including a 20.3% improvement in ZINC MAE.", "keywords": [ "Graph Neural Networks", "Graph Encoding", "Graph Rewiring", "Attention Mechanism", "Deep Learning" ], "rating_list": [ 6, 3, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 3, 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "rEEjYlzXUD", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 6, "author_info_list": [ { "name": "Chao Yang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Kejun Tang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Weiqing Ren", "gender": "unknown", "institution": "National University of Singapore", "country": "", "position": "Full Professor" }, { "name": "Xiaoliang Wan", "gender": "unknown", "institution": "Louisiana State University", "country": "", "position": "Full Professor" }, { "name": "Xili Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yueyang Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 21, "n_ref_uni": 48, "n_ref": 77, "n_ref_all": 108, "n_fig": 22, "n_tab": 3, "L_tab": 1562, "n_element_tab": 144, "n_fig_1": 17, "n_tab_1": 3, "L_tab_1": 1572, "n_element_tab_1": 154, "formula_len_all": 1788, "formula_len_all_1": 1308, "len_all": 134531, "len_all_1": 68625, "len_abs": 1522, "len_title": 82, "len_sents": 38523, "len_sents_1": 27653, "n_sents": 324, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1529, "title": "Estimating Committor Functions via Deep Adaptive Sampling on Rare Transition Paths", "abs": "The committor functions are central to investigating rare but important events in molecular simulations. It is known that computing the committor function suffers from the curse of dimensionality. Recently, using neural networks to estimate the committor function has gained attention due to its potential for high-dimensional problems. Training neural networks to approximate the committor function needs to sample transition data from straightforward simulations of rare events, which is very inefficient. The scarcity of transition data makes it challenging to approximate the committor function. To address this problem, we propose an efficient framework to generate data points in the transition state region that helps train neural networks to approximate the committor function. We design a Deep Adaptive Sampling method for TRansition paths (DASTR), where deep generative models are employed to generate samples to capture the information of transitions effectively. In particular, we treat a non-negative function in terms of the integrand in the loss functional as an unnormalized probability density function and approximate it with the deep generative model. The new samples from the deep generative model are located in the region of the transition and fewer samples are located in the other region, which provides effective samples for approximating the committor function and significantly improves the accuracy. We demonstrate the effectiveness of the proposed method with both simulations and realistic examples.", "keywords": [ "committor function", "deep adaptive sampling", "rare event", "transition path" ], "rating_list": [ 3, 3, 8, 3 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 1, 4, 2 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "rDe9yQQYKt", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "FENG SHIBO", "gender": "Male", "institution": "School of Computer Science and Engineering, Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Peilin Zhao", "gender": "unknown", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Wanjin Feng", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xingyu Gao", "gender": "unknown", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhiqi Shen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Lecturer" } ], "n_formula": 29, "n_formula_1": 46, "n_ref_uni": 39, "n_ref": 74, "n_ref_all": 91, "n_fig": 3, "n_tab": 9, "L_tab": 6671, "n_element_tab": 369, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1443, "n_element_tab_1": 58, "formula_len_all": 1573, "formula_len_all_1": 1851, "len_all": 148784, "len_all_1": 61391, "len_abs": 1459, "len_title": 125, "len_sents": 41162, "len_sents_1": 26674, "n_sents": 320, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1467, "title": "TS-LIF: A Temporal Segment Spiking Neuron Network for Time Series Forecasting", "abs": "Spiking Neural Networks (SNNs) offer a promising, biologically inspired approach for processing spatiotemporal data, particularly for time series forecasting.\nHowever, conventional neuron models like the Leaky Integrate-and-Fire (LIF) struggle to capture long-term dependencies and effectively process multi-scale temporal dynamics.\nTo overcome these limitations, we introduce the Temporal Segment Leaky Integrate-and-Fire (TS-LIF) model, featuring a novel dual-compartment architecture.\nThe dendritic and somatic compartments specialize in capturing distinct frequency components, providing functional heterogeneity that enhances the neuron's ability to process both low- and high-frequency information.\nFurthermore, the newly introduced direct somatic current injection reduces information loss during intra-neuronal transmission, while dendritic spike generation improves multi-scale information extraction.\nWe provide a theoretical stability analysis of the TS-LIF model and explain how each compartment contributes to distinct frequency response characteristics.\nExperimental results show that TS-LIF outperforms traditional SNNs in time series forecasting, demonstrating better accuracy and robustness, even with missing data.\nTS-LIF advances the application of SNNs in time-series forecasting, providing a biologically inspired approach that captures complex temporal dynamics and offers potential for practical implementation in diverse forecasting scenarios.", "keywords": [ "spiking neural network", "time series forecasting", "Application" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "rDb9oY6Ww7", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Guorui Feng", "gender": "Male", "institution": "Shanghai University", "country": "CN", "position": "Full Professor" }, { "name": "Nan Pu", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Postdoc" }, { "name": "Nicu Sebe", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Full Professor" }, { "name": "Yalan Qin", "gender": "unknown", "institution": "Shanghai University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 51, "n_formula_1": 54, "n_ref_uni": 13, "n_ref": 19, "n_ref_all": 40, "n_fig": 3, "n_tab": 3, "L_tab": 2470, "n_element_tab": 178, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 1454, "n_element_tab_1": 108, "formula_len_all": 3143, "formula_len_all_1": 3099, "len_all": 97811, "len_all_1": 70670, "len_abs": 2403, "len_title": 126, "len_sents": 27642, "len_sents_1": 24799, "n_sents": 228, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 2417, "title": "Robust Consensus Anchor Learning for Efficient Multi-view Subspace Clustering", "abs": "As a leading unsupervised classification algorithm in artificial intelligence, multi-view subspace clustering segments unlabeled data from different subspaces. Recent works based on the anchor have been proposed to decrease the computation complexity for the datasets with large scales in multi-view clustering. The major differences among these methods lie on the objective functions they define. Despite considerable success, these works pay few attention to guaranting the robustness of learned consensus anchors via effective manner for efficient multi-view clustering and investigating the specific local distribution of cluster in the affine subspace. Besides, the robust consensus anchors as well as the common cluster structure shared by different views are not able to be simultaneously learned. In this paper, we propose Robust Consensus anchors learning for efficient multi-view Subspace Clustering (RCSC). We first show that if the data are sufficiently sampled from independent subspaces, and the objective function meets some conditions, the achieved anchor graph has the block-diagonal structure. As a special case, we provide a model based on Frobenius norm, non-negative and affine constraints in consensus anchors learning, which guarantees the robustness of learned consensus anchors for efficient multi-view clustering and investigates the specific local distribution of cluster in the affine subspace. While it is simple, we theoretically give the geometric analysis regarding the formulated RCSC. The union of these three constraints is able to restrict how each data point is described in the affine subspace with specific local distribution of cluster for guaranting the robustness of learned consensus anchors. RCSC takes full advantages of correlation among consensus anchors, which encourages the grouping effect and groups highly correlated consensus anchors together with the guidance of view-specific projection. The anchor graph construction, partition and robust anchor learning are jointly integrated into a unified framework. It ensures the mutual enhancement for these procedures and helps lead to more discriminative consensus anchors as well as the cluster indicator. We then adopt an alternative optimization strategy for solving the formulated problem. Experiments performed on eight multi-view datasets confirm the superiority of RCSC based on the effectiveness and efficiency.", "keywords": [ "Multi-view clustering", "consensus anchor learning", "effectiveness and efficiency" ], "rating_list": [ 8, 8, 8, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "rDRCIvTppL", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Andrew Gilbert", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Researcher" }, { "name": "Armin Mustafa", "gender": "Female", "institution": "University of Surrey", "country": "GB", "position": "Lecturer" }, { "name": "Chun-Hao Paul Huang", "gender": "Not Specified", "institution": "Adobe Systems", "country": "GB", "position": "Researcher" }, { "name": "Duygu Ceylan", "gender": "unknown", "institution": "Adobe Systems", "country": "GB", "position": "Research Scientist" }, { "name": "Soon Yau Cheong", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 21, "n_ref": 59, "n_ref_all": 80, "n_fig": 7, "n_tab": 6, "L_tab": 851, "n_element_tab": 101, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 923, "n_element_tab_1": 121, "formula_len_all": 503, "formula_len_all_1": 586, "len_all": 106389, "len_all_1": 59975, "len_abs": 1518, "len_title": 111, "len_sents": 32956, "len_sents_1": 28113, "n_sents": 228, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1064, "title": "Boosting Camera Motion Control for Video Diffusion Transformers", "abs": "Recent advancements in diffusion models have significantly enhanced the quality of video generation. However, fine-grained control over camera pose remains a challenge. While U-Net-based models have shown promising results for camera control, transformer-based diffusion models (DiT)\u2014the preferred architecture for large-scale video generation\u2014suffer from severe degradation in camera motion accuracy. In this paper, we investigate the underlying causes of this issue and propose solutions tailored to DiT architectures. Our study reveals that camera control performance depends heavily on the choice of conditioning methods rather than camera pose representations that is commonly believed. To address the persistent motion degradation in DiT, we introduce **Camera Motion Guidance (CMG)**, based on classifier-free guidance, which **boosts camera control by over 400%**. Additionally, we present a sparse camera control pipeline, significantly simplifying the process of specifying camera poses for long videos. Code and models will be released upon publication.", "keywords": [ "video generation", "camera control", "diffusion models", "diffusion transformers", "DiT" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "rDLgnYLM5b", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Benlin Liu", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Caixi Chen", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Dongping Chen", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Intern" }, { "name": "Pan Zhou", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Professor" }, { "name": "Ranjay Krishna", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "Ruoxi Chen", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Shu Pu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yanru Wu", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yao Wan", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yue Huang", "gender": "unknown", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Zhaoyi Liu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Intern" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 62, "n_ref": 144, "n_ref_all": 228, "n_fig": 53, "n_tab": 9, "L_tab": 4413, "n_element_tab": 487, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 6583, "n_element_tab_1": 420, "formula_len_all": 74, "formula_len_all_1": 0, "len_all": 361185, "len_all_1": 66910, "len_abs": 2127, "len_title": 113, "len_sents": 114652, "len_sents_1": 26951, "n_sents": 1059, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 76, "L_abs": 1663, "title": "Interleaved Scene Graph for Interleaved Text-and-Image Generation Assessment", "abs": "Many real-world user queries (e.g. *\"How do to make egg fried rice?\"*) could benefit from systems capable of generating responses with both textual steps with accompanying images, similar to a cookbook.\nModels designed to generate interleaved text and images face challenges in ensuring consistency within and across these modalities.\nTo address these challenges, we present ISG, a comprehensive evaluation framework for interleaved text-and-image generation. ISG leverages a scene graph structure to capture relationships between text and image blocks, evaluating responses on four levels of granularity: holistic, structural, block-level, and image-specific. This multi-tiered evaluation allows for a nuanced assessment of consistency, coherence, and accuracy, and provides interpretable question-answer feedback.\nIn conjunction with ISG, we introduce a benchmark, ISG-Bench, encompassing 1,150 samples across 8 categories and 21 subcategories. This benchmark dataset includes complex language-vision dependencies and golden answers to evaluate models effectively on vision-centric tasks such as style transfer, a challenging area for current models. \nUsing ISG-Bench, we demonstrate that recent unified vision-language models perform poorly on generating interleaved content. While compositional approaches that combine separate language and image models show a 111% improvement over unified models at the holistic level, their performance remains suboptimal at both block and image levels.\nTo facilitate future work, we develop ISG-Agent, a baseline agent employing a *\"plan-execute-refine\"* pipeline to invoke tools, achieving a 122% performance improvement.", "keywords": [ "Interleaved Text-and-Image Generation", "Generative Models", "Multimodal Large Language Model", "Scene Graphs", "Automatic Evaluation", "Benchmark" ], "rating_list": [ 8, 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 4, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4, 3 ] }, { "paper_id": "rDIf6NA5mj", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Renlong Wu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "ShuohaoZhang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Zhilu Zhang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Zifei Yan", "gender": "Female", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Wangmeng Zuo", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 12, "n_ref_uni": 69, "n_ref": 199, "n_ref_all": 214, "n_fig": 6, "n_tab": 10, "L_tab": 2436, "n_element_tab": 206, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2230, "n_element_tab_1": 155, "formula_len_all": 1026, "formula_len_all_1": 505, "len_all": 217750, "len_all_1": 54159, "len_abs": 1821, "len_title": 108, "len_sents": 48209, "len_sents_1": 22033, "n_sents": 445, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1306, "title": "Exposure Bracketing Is All You Need For A High-Quality Image", "abs": "It is highly desired but challenging to acquire high-quality photos with clear content in low-light environments. Although multi-image processing methods (using burst, dual-exposure, or multi-exposure images) have made significant progress in addressing this issue, they typically focus on specific restoration or enhancement problems, and do not fully explore the potential of utilizing multiple images. Motivated by the fact that multi-exposure images are complementary in denoising, deblurring, high dynamic range imaging, and super-resolution, we propose to utilize exposure bracketing photography to get a high-quality image by combining these tasks in this work. Due to the difficulty in collecting real-world pairs, we suggest a solution that first pre-trains the model with synthetic paired data and then adapts it to real-world unlabeled images. In particular, a temporally modulated recurrent network (TMRNet) and self-supervised adaptation method are proposed. Moreover, we construct a data simulation pipeline to synthesize pairs and collect real-world images from 200 nighttime scenarios. Experiments on both datasets show that our method performs favorably against the state-of-the-art multi-image processing ones. Code and datasets are available at https://github.com/cszhilu1998/BracketIRE.", "keywords": [ "Exposure Bracketing", "Image Restoration and Enhancement" ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "rD6LQagatR", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 12, "author_info_list": [ { "name": "Feng Sun", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Furu Wei", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Partner Research Manager" }, { "name": "Fuzhen Zhuang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Professor" }, { "name": "Haizhen Huang", "gender": "Male", "institution": "Institute of Computing Technolgy, Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Liang Wang", "gender": "Male", "institution": "Microsoft Research", "country": "", "position": "Researcher" }, { "name": "Minghui Song", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Qi Zhang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Shaohan Huang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Ting Jiang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" }, { "name": "Weiwei Deng", "gender": "Male", "institution": "South China University of Technology", "country": "", "position": "PhD student" }, { "name": "Zihan Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "MS student" }, { "name": "deqing wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 20, "n_ref": 35, "n_ref_all": 53, "n_fig": 5, "n_tab": 4, "L_tab": 1826, "n_element_tab": 324, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1711, "n_element_tab_1": 312, "formula_len_all": 128, "formula_len_all_1": 128, "len_all": 92691, "len_all_1": 54974, "len_abs": 1254, "len_title": 112, "len_sents": 25510, "len_sents_1": 22784, "n_sents": 173, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1274, "title": "E5-V: Universal Embeddings with Multimodal Large Language Models", "abs": "Multimodal large language models (MLLMs) have shown promising advancements in general visual and language understanding. However, the representation of multimodal information using MLLMs remains largely unexplored. In this work, we introduce a new framework, E5-V, designed to adapt MLLMs for achieving universal multimodal embeddings. Our findings highlight the significant potential of MLLMs in representing multimodal inputs compared to previous approaches. By leveraging MLLMs with prompts, E5-V effectively bridges the modality gap between different types of inputs, demonstrating strong performance in multimodal embeddings even without fine-tuning. We propose a single modality training approach for E5-V, where the model is trained exclusively on text pairs. This method demonstrates significant improvements over traditional multimodal training on image-text pairs, while reducing training costs by approximately 95\\%. Additionally, this approach eliminates the need for costly multimodal training data collection. Extensive experiments across four types of tasks demonstrate the effectiveness of E5-V. As a universal multimodal model, E5-V not only achieves but often surpasses state-of-the-art performance in each task, despite being trained on a single modality.", "keywords": [ "Multimodal Large Language Models", "Multimodal Learning", "Representation Learning" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "rCvdAVQpAe", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Jasmine Ratchford", "gender": "Not Specified", "institution": "Carnegie Mellon University", "country": "US", "position": "Researcher" }, { "name": "Raffaele Romagnoli", "gender": "unknown", "institution": "Duquesne University", "country": "US", "position": "Assistant Professor" }, { "name": "Zhuoyuan Wang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "yorie nakahira", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" } ], "n_formula": 59, "n_formula_1": 40, "n_ref_uni": 48, "n_ref": 75, "n_ref_all": 99, "n_fig": 10, "n_tab": 1, "L_tab": 61, "n_element_tab": 8, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1666, "n_element_tab_1": 25, "formula_len_all": 4996, "formula_len_all_1": 3131, "len_all": 140759, "len_all_1": 52440, "len_abs": 1177, "len_title": 87, "len_sents": 38073, "len_sents_1": 21797, "n_sents": 306, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1181, "title": "Physics-Informed Deep B-Spline Networks", "abs": "Physics-informed machine learning provides an approach to combing data and governing physics laws for solving complex partial differential equations (PDEs). However, efficiently solving PDEs with varying parameters and changing initial conditions and boundary conditions (ICBCs) remains an open challenge. We propose a hybrid framework that uses a neural network to learn B-spline control points to approximate solutions to PDEs with varying system and ICBC parameters. The proposed network can be trained efficiently as one can directly specify ICBCs without imposing losses, calculate physics-informed loss functions through analytical formulas, and requires only learning the weights of B-spline functions as opposed to both weights and basis as in traditional neural operator learning methods. We show theoretical guarantees that the proposed B-spline networks are universal approximators of arbitrary dimensional PDEs under certain conditions. We also demonstrate in experiments that the proposed B-spline network can solve problems with discontinuous ICBCs and outperforms existing methods, and is able to learn solutions of 3D heat equations with diverse initial conditions.", "keywords": [ "Physics-informed machine learning", "B-splines", "Partial differential equations (PDEs)" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "rCno6eYdXk", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chenghao Sun", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Intern" }, { "name": "Linfeng Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Shaobo Wang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Shuaiyu Zhang", "gender": "Male", "institution": "Harbin Institute of Technology at Weihai", "country": "CN", "position": "Undergrad student" }, { "name": "Weiya Li", "gender": "Male", "institution": "ICBC", "country": "CN", "position": "Researcher" }, { "name": "Xuming Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Yantai Yang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 19, "n_formula_1": 19, "n_ref_uni": 40, "n_ref": 99, "n_ref_all": 136, "n_fig": 8, "n_tab": 11, "L_tab": 7222, "n_element_tab": 667, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 4392, "n_element_tab_1": 354, "formula_len_all": 1282, "formula_len_all_1": 1307, "len_all": 164287, "len_all_1": 76233, "len_abs": 1254, "len_title": 101, "len_sents": 50293, "len_sents_1": 31312, "n_sents": 376, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1261, "title": "DRUPI: Dataset Reduction Using Privileged Information", "abs": "Dataset reduction (DR) seeks to select or distill samples from large datasets into smaller subsets while preserving performance on target tasks. Existing methods primarily focus on pruning or synthesizing data in the same format as the original dataset, typically the input data and corresponding labels. However, in DR settings, we find it is possible to synthesize more information beyond the data-label pair as an additional learning target to facilitate model training. In this paper, we introduce Dataset Reduction Using Privileged Information (DRUPI), which enriches DR by synthesizing privileged information alongside the reduced dataset. This privileged information can take the form of feature labels or attention labels, providing auxiliary supervision to improve model learning. Our findings reveal that effective feature labels must balance between being overly discriminative and excessively diverse, with a moderate level proving optimal for improving the reduced dataset\u2019s efficacy. Extensive experiments on ImageNet, CIFAR-10/100, and Tiny ImageNet demonstrate that DRUPI integrates seamlessly with existing dataset reduction methods, offering significant performance gains. *Code is included in the supplementary material and will be released.*", "keywords": [ "Dataset Reduction", "Efficient AI", "Privileged Information" ], "rating_list": [ 3, 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 5 ] }, { "paper_id": "rCaA79Obsj", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Ayan Banerjee", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "Assistant Professor" }, { "name": "Sandeep Gupta", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 35, "n_ref": 66, "n_ref_all": 96, "n_fig": 5, "n_tab": 8, "L_tab": 4226, "n_element_tab": 472, "n_fig_1": 6, "n_tab_1": 8, "L_tab_1": 5516, "n_element_tab_1": 370, "formula_len_all": 1160, "formula_len_all_1": 484, "len_all": 133206, "len_all_1": 77449, "len_abs": 1315, "len_title": 132, "len_sents": 40259, "len_sents_1": 31691, "n_sents": 280, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1555, "title": "Towards zero shot multivariate time series anomaly detection - A Realistic Evaluation", "abs": "A long line of multivariate timeseries anomaly detection (MTAD) approaches use performance enhancement techniques that are not feasible in practical scenarios. In specific, a) point adjustment technique is employed which uses ground truth to forcefully convert false negatives to true positives and inflates precision to unrealistic proportions, and b) significant data leakage is introduced where anomaly score threshold is determined using the test data and test labels. In this paper, we show the real world performance of existing MTAD techniques when point adjustment and threshold learning on test data is disabled. Moreover, we show that anomalies introduced in real world benchmark datasets result in significant distribution shift between normal and anomalous data, and when point adjustment and threshold learning are used even untrained deterministic methods can perform on par or even beat baseline techniques. We then introduce six synthetic benchmark examples derived from real world systems, where anomalous data and normal data have statistically insignificant distribution shift. We propose, sparse model identification enhanced anomaly detection (SPIE-AD), a model recovery and conformance based zero shot MTAD approach that outperforms state of art MTAD techniques on three real world benchmark datasets without using point adjustment and threshold learning on test data. We evaluate state-of-art MTAD and SPIE-AD on the novel synthetic benchmarks. SPIE-AD outperforms state-of-art MTAD techniques on both standard and novel benchmarks.", "keywords": [ "multidimensional time series anomaly detection", "zero shot", "model recovery", "model conformance" ], "rating_list": [ 3, 5, 5, 8, 5 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 1, 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "rCX9l4OTCT", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Pengfei Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Wei Sun", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiang Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Intern" }, { "name": "Yan Wang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yupeng Zheng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "yilun chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 25, "n_ref": 78, "n_ref_all": 101, "n_fig": 5, "n_tab": 13, "L_tab": 5923, "n_element_tab": 456, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 4003, "n_element_tab_1": 242, "formula_len_all": 803, "formula_len_all_1": 755, "len_all": 145214, "len_all_1": 65955, "len_abs": 1271, "len_title": 126, "len_sents": 41584, "len_sents_1": 27152, "n_sents": 303, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1280, "title": "Semi-Supervised Vision-Centric 3D Occupancy World Model for Autonomous Driving", "abs": "Understanding world dynamics is crucial for planning in autonomous driving. Recent methods attempt to achieve this by learning a 3D occupancy world model that forecasts future surrounding scenes based on current observation. However, 3D occupancy labels are still required to produce promising results. Considering the high annotation cost for 3D outdoor scenes, we propose a semi-supervised vision-centric 3D occupancy world model, **PreWorld**, to leverage the potential of 2D labels through a novel two-stage training paradigm: the self-supervised pre-training stage and the fully-supervised fine-tuning stage. Specifically, during the pre-training stage, we utilize an attribute projection head to generate different attribute fields of a scene (e.g., RGB, density, semantic), thus enabling temporal supervision from 2D labels via volume rendering techniques. Furthermore, we introduce a simple yet effective state-conditioned forecasting module to recursively forecast future occupancy and ego trajectory in a direct manner. Extensive experiments on the nuScenes dataset validate the effectiveness and scalability of our method, and demonstrate that PreWorld achieves competitive performance across 3D occupancy prediction, 4D occupancy forecasting and motion planning tasks.", "keywords": [ "Autonomous Driving; Occupancy; World Model" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "rCGleSgNBK", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Bohyung Han", "gender": "Not Specified", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Jinyoung Choi", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Junoh Kang", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 35, "n_ref_uni": 26, "n_ref": 79, "n_ref_all": 122, "n_fig": 7, "n_tab": 5, "L_tab": 1344, "n_element_tab": 213, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 686, "n_element_tab_1": 120, "formula_len_all": 2349, "formula_len_all_1": 2351, "len_all": 124036, "len_all_1": 66185, "len_abs": 1492, "len_title": 121, "len_sents": 35071, "len_sents_1": 27143, "n_sents": 267, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1495, "title": "Enhanced Diffusion Sampling via Extrapolation with Multiple ODE Solutions", "abs": "Diffusion probabilistic models (DPMs), while effective in generating high-quality samples, often suffer from high computational costs due to the iterative sampling process. To address this, we propose an enhanced ODE-based sampling method for DPMs inspired by Richardson extrapolation, which has been shown to reduce numerical error and improve convergence rates. Our method, termed RX-DPM, utilizes numerical solutions obtained over multiple denoising steps, leveraging the multiple ODE solutions to extrapolate the denoised prediction in DPMs. This signi\ufb01cantly enhances the accuracy of estimations for the \ufb01nal sample while preserving the number of function evaluations (NFEs). Unlike standard Richardson extrapolation, which assumes uniform discretization of the time grid, we have developed a more general formulation tailored to arbitrary time step scheduling, guided by the local truncation error derived from a baseline sampling method. The simplicity of our approach facilitates accurate estimation of numerical solutions without additional computational overhead, and allows for seamless and convenient integration into various DPMs and solvers. Additionally, RX-DPM provides explicit error estimates, effectively illustrating the faster convergence achieved as the order of the leading error term increases. Through a series of experiments, we demonstrate that the proposed method effectively enhances the quality of generated samples without requiring additional sampling iterations.", "keywords": [ "Diffusion models" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "rBzvEEbrF7", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "AJAY KUMAR JAISWAL", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Jiawei Zhao", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Lu Yin", "gender": "unknown", "institution": "University of Surrey", "country": "GB", "position": "Assistant Professor" }, { "name": "Shiwei Liu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Yuandong Tian", "gender": "Male", "institution": "Meta AI (FAIR)", "country": "", "position": "Research Scientist" }, { "name": "Zhangyang Wang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Zhenyu Zhang", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 18, "n_ref": 33, "n_ref_all": 46, "n_fig": 7, "n_tab": 6, "L_tab": 4926, "n_element_tab": 208, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3697, "n_element_tab_1": 140, "formula_len_all": 241, "formula_len_all_1": 241, "len_all": 138220, "len_all_1": 64705, "len_abs": 2107, "len_title": 134, "len_sents": 31888, "len_sents_1": 29400, "n_sents": 219, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 2098, "title": "Q-GaLore: Quantized GaLore with INT4 Projection and Layer-Adaptive Low-Rank Gradients", "abs": "Training Large Language Models (LLMs) is memory-intensive due to the large number of parameters and associated optimization states. GaLore, a recent method, reduces memory usage by projecting weight gradients into a low-rank subspace without compromising performance. However, GaLore relies on time-consuming Singular Value Decomposition (SVD) operations to identify the subspace, and the frequent subspace updates lead to significant training time overhead. Moreover, GaLore offers minimal improvements in accuracy and efficiency compared to LoRA in more accessible fine-tuning scenarios. To address these limitations, we introduce Q-GaLore, a novel approach that substantially reduces memory usage by combining quantization and low-rank projection, surpassing the benefits of GaLore. Our method is based on two key observations: (i) the gradient subspace exhibits diverse properties, with some layers converging early in training while others are subject to frequent changes; (ii) the projection matrices are highly resilient to low-bit quantization. Leveraging these insights, Q-GaLore adaptively updates the gradient subspace based on its convergence statistics, achieving comparable performance while significantly reducing the number of SVD operations. We maintain the projection matrices in INT4 format for aggressive memory conservation and preserve weights in INT8 format, incorporating stochastic rounding to capture accumulated gradient information. This approach enables a high-precision training trajectory using only low-precision weights. We demonstrate that Q-GaLore achieves highly competitive pre-training and fine-tuning performance with exceptional memory efficiency. At pre-training, Q-GaLore facilitates training a LLaMA-7B model from scratch on a single NVIDIA RTX 4060 Ti with only 16 GB memory, showcasing its exceptional memory efficiency and practicality. At fine-tuning, it reduces memory consumption by up to 50% compared to LoRA and GaLore, while consistently outperforming QLoRA (by up to 5.19 on MMLU) at the same memory cost. Codes will be released upon acceptance.", "keywords": [ "Large Language Models; Memory Efficient Training; Low Rank" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "rBAnJed1iY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Afaf Taik", "gender": "Female", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Postdoc" }, { "name": "Golnoosh Farnadi", "gender": "Female", "institution": "McGill University", "country": "CA", "position": "Assistant Professor" }, { "name": "Saber Malekmohammadi", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD Student" } ], "n_formula": 23, "n_formula_1": 18, "n_ref_uni": 54, "n_ref": 104, "n_ref_all": 125, "n_fig": 5, "n_tab": 1, "L_tab": 627, "n_element_tab": 42, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2841, "n_element_tab_1": 63, "formula_len_all": 2802, "formula_len_all_1": 1566, "len_all": 133907, "len_all_1": 72662, "len_abs": 1452, "len_title": 132, "len_sents": 31308, "len_sents_1": 27153, "n_sents": 225, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1458, "title": "A Provably Robust Algorithm for Differentially Private Clustered Federated Learning", "abs": "Federated Learning (FL), which is a decentralized machine learning (ML) approach, often incorporates differential privacy (DP) to enhance data privacy guarantees. However, differentially private federated learning (DPFL) introduces performance disparities across clients, particularly affecting minority groups. Some recent works have attempted to address large data heterogeneity in vanilla FL settings through clustering clients, but these methods remain sensitive and prone to errors further exacerbated by the DP noise, making them inappropriate for DPFL settings. We propose an algorithm for differentially private clustered FL, which is robust to the DP noise in the system and identifies clients\u2019 clusters correctly. To this end, we propose to cluster clients based on both their model updates and training loss values. Furthermore, when clustering clients\u2019 model updates, our proposed approach addresses the server\u2019s uncertainties by employing large batch sizes as well as Gaussian Mixture Models (GMM) to reduce the impact of DP and stochastic noise and avoid potential clustering errors. This idea is efficient especially in privacy-sensitive scenarios with more DP noise. We provide theoretical analysis justifying our approach, and evaluate it extensively across diverse data distributions and privacy budgets. Our experimental results show its effectiveness in addressing large data heterogeneity in DPFL systems with a small computational cost.", "keywords": [ "Federated Learning", "Clustered Federated Learning", "Differential Privacy", "Clustering" ], "rating_list": [ 6, 3, 8, 3 ], "soundness_list": [ 2, 1, 3, 1 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "rAylWUIKtu", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Cenny Wenner", "gender": "Male", "institution": "KTH - Royal Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Clement Neo", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "Esben Kran", "gender": "Male", "institution": "Apart Research", "country": "", "position": "Researcher" }, { "name": "Jacob Haimes", "gender": "Male", "institution": "Odyssean Institute", "country": "GB", "position": "Researcher" }, { "name": "Jason Hoelscher-Obermaier", "gender": "Male", "institution": "Apart Research", "country": "", "position": "Principal Researcher" }, { "name": "Kunvar Thaman", "gender": "Male", "institution": "BITS Pilani, Birla Institute of Technology and Science", "country": "", "position": "Researcher" }, { "name": "Vassil Tashev", "gender": "Male", "institution": "Simon Fraser University", "country": "", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 66, "n_ref_all": 84, "n_fig": 9, "n_tab": 4, "L_tab": 1214, "n_element_tab": 61, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 429, "n_element_tab_1": 67, "formula_len_all": 235, "formula_len_all_1": 0, "len_all": 157510, "len_all_1": 51768, "len_abs": 1159, "len_title": 120, "len_sents": 47903, "len_sents_1": 24930, "n_sents": 366, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1163, "title": "Benchmark Inflation: Revealing LLM Performance Gaps Using Retro-Holdouts", "abs": "The training data for many Large Language Models (LLMs) is contaminated with test data. This means that public benchmarks used to assess LLMs are compromised, suggesting a performance gap between benchmark scores and actual capabilities. Ideally, a private holdout set could be used to accurately verify scores. Unfortunately, such datasets do not exist for most benchmarks, and post-hoc construction of sufficiently similar datasets is non-trivial. To address these issues, we introduce a systematic methodology for (i) retrospectively constructing a holdout dataset for a target dataset, (ii) demonstrating the statistical indistinguishability of this retro-holdout dataset, and (iii) comparing LLMs on the two datasets to quantify the performance gap due to the dataset's public availability. Applying these methods to TruthfulQA, we construct and release Retro-Misconceptions, on which we evaluate twenty LLMs and find that some have inflated scores by as much as 16 percentage points. Our results demonstrate that public benchmark scores do not always accurately assess model properties, and underscore the importance of improved data practices in the field.", "keywords": [ "Language Models", "Datasets and Benchmarks", "Model Evaluation", "Data Contamination", "Evaluation Gaming" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "rAoEub6Nw2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Ion Stoica", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Michael W. Mahoney", "gender": "unknown", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Siavash Ameli", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Siyuan Zhuang", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 65, "n_formula_1": 34, "n_ref_uni": 43, "n_ref": 64, "n_ref_all": 93, "n_fig": 4, "n_tab": 5, "L_tab": 2386, "n_element_tab": 304, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3296, "formula_len_all_1": 1588, "len_all": 239950, "len_all_1": 69094, "len_abs": 5714, "len_title": 102, "len_sents": 79257, "len_sents_1": 29192, "n_sents": 618, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1401, "title": "A Statistical Framework for Ranking LLM-based Chatbots", "abs": "Large language models (LLMs) have transformed natural language processing, with frameworks like Chatbot Arena providing pioneering platforms for evaluating these models. By facilitating millions of pairwise comparisons based on human judgments, Chatbot Arena has become a cornerstone in LLM evaluation, offering rich datasets for ranking models in open-ended conversational tasks. Building upon this foundation, we propose a statistical framework that incorporates key advancements to address specific challenges in pairwise comparison analysis. First, we introduce a factored tie model that enhances the ability to handle ties\u2014an integral aspect of human-judged comparisons\u2014significantly improving the model\u2019s fit to observed data. Second, we extend the framework to model covariance between competitors, enabling deeper insights into performance relationships and facilitating intuitive groupings into performance tiers. Third, we resolve optimization challenges arising from parameter invariances by introducing novel constraints, ensuring stable and interpretable parameter estimation. Through rigorous evaluation and extensive experimentation, our framework demonstrates substantial improvements over existing methods in modeling pairwise comparison data. To support reproducibility and practical adoption, we release leaderbot, an open-source Python package implementing our models and analyses.", "keywords": [ "Large Language Models (LLMs)", "Paired Comparison", "Statistical Ranking", "Human Preferences", "Chatbot Arena", "Logistic Regression" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 2, 2, 1 ], "contribution_list": [ 3, 3, 1 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "rAcgDBdKnP", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Dawei Yang", "gender": "Male", "institution": "Houmo", "country": "CN", "position": "Researcher" }, { "name": "JiangyongYu", "gender": "Male", "institution": "Houmo", "country": "CN", "position": "Engineer" }, { "name": "Sifan Zhou", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "PhD student" }, { "name": "XUCHEN", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Xing Hu", "gender": "Male", "institution": "houmo.ai", "country": "CN", "position": "Researcher" }, { "name": "Yuan Cheng", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Zhe jiang", "gender": "Male", "institution": "Southeast University", "country": "", "position": "Full Professor" }, { "name": "Zhihang Yuan", "gender": "Male", "institution": "Infinigence AI", "country": "CN", "position": "Researcher" }, { "name": "Zhixuan Chen", "gender": "Male", "institution": "houmo", "country": "CN", "position": "Researcher" }, { "name": "Zukang Xu", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 29, "n_formula_1": 11, "n_ref_uni": 29, "n_ref": 64, "n_ref_all": 75, "n_fig": 7, "n_tab": 12, "L_tab": 5085, "n_element_tab": 468, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 3736, "n_element_tab_1": 159, "formula_len_all": 1627, "formula_len_all_1": 740, "len_all": 143577, "len_all_1": 69619, "len_abs": 1721, "len_title": 177, "len_sents": 44205, "len_sents_1": 31373, "n_sents": 344, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 128, "L_abs": 1732, "title": "OSTQuant: Refining Large Language Model Quantization with Orthogonal and Scaling Transformations for Better Distribution Fitting", "abs": "Post-training quantization (PTQ) has emerged as a widely adopted technique for compressing and accelerating Large Language Models (LLMs).\nThe major challenge in LLM quantization is that uneven and heavy-tailed data distributions can expand the quantization range, thereby reducing bit precision for most values.\nRecent methods attempt to eliminate outliers and balance inter-channel differences by employing linear transformations; however, they remain heuristic and are often overlook optimizing the data distribution across the entire quantization space.\nIn this paper, we introduce Quantization Space Utilization Rate (QSUR), a novel metric that effectively assesses the quantizability of transformed data by measuring the space utilization of the data in the quantization space. We complement QSUR with mathematical derivations that examine the effects and limitations of various transformations, guiding our development of Orthogonal and Scaling Transformation-based Quantization (OSTQuant). OSTQuant employs a learnable equivalent transformation, consisting of an orthogonal transformation and a scaling transformation, to optimize the distributions of weights and activations across the entire quantization space. Futhermore, we propose the KL-Top loss function, designed to mitigate noise during optimization while retaining richer semantic information within the limited calibration data imposed by PTQ.\nOSTQuant outperforms existing work on various LLMs and benchmarks. In the W4-only setting, it retains 99.5\\% of the floating-point accuracy. In the more challenging W4A4KV4 configuration, OSTQuant reduces the performance gap by 32\\% on the LLaMA-3-8B model compared to state-of-the-art methods. Code will be available.", "keywords": [ "Large Language Models", "Quantization" ], "rating_list": [ 5, 6, 6, 6, 8 ], "soundness_list": [ 2, 2, 3, 3, 4 ], "presentation_list": [ 1, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3, 4 ], "confidence_list": [ 5, 3, 4, 5, 2 ] }, { "paper_id": "rAZ3yCpc3K", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Lakshmikar Reddy Polamreddy", "gender": "Male", "institution": "Yeshiva University", "country": "US", "position": "PhD student" }, { "name": "Youshan Zhang", "gender": "Male", "institution": "Yeshiva University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 22, "n_ref_uni": 33, "n_ref": 112, "n_ref_all": 123, "n_fig": 6, "n_tab": 3, "L_tab": 474, "n_element_tab": 47, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2282, "n_element_tab_1": 98, "formula_len_all": 783, "formula_len_all_1": 730, "len_all": 118476, "len_all_1": 63359, "len_abs": 1456, "len_title": 126, "len_sents": 38194, "len_sents_1": 28799, "n_sents": 273, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1467, "title": "The Deficit of New Information in Diffusion Models: A Focus on Diverse Samples", "abs": "Diffusion models are renowned for their state-of-the-art performance in generating high-quality images. Identifying samples with new information beyond the training data is essential for data augmentation, especially for enhancing model performance in diverse and unforeseen real-world scenarios. However, the investigation of new information in the generated samples has not been well explored. Our investigation through the lens of information theory reveals that diffusion models do not produce new information beyond what exists in the training data. Next, we introduce the concept of diverse samples (DS) to prove that generated images could contain information not present in the training data for diffusion models. Furthermore, we propose a method for identifying diverse samples among generated images by extracting deep features and detecting images that fall outside the boundary of real images. We demonstrate that diverse samples exist in the generated data of diffusion models, attributed to the estimation of forward and backward processes, but it can only produce a limited number of diverse samples, underscoring a notable gap in their capabilities in generating diverse samples. In addition, our experiment on the Chest X-ray dataset demonstrates that the diverse samples are more useful in improving classification accuracy than vanilla-generated samples. The source code is available at \\url{https://github.com/lypz12024/diffusion-diverse-samples}.", "keywords": [ "DIffusion Models", "Diverse Samples", "Information Theory" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 4 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "r9oqHOdoHf", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Ivona Najdenkoska", "gender": "Female", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Marcel Worring", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "Mohammad Mahdi Derakhshani", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Nanne Van Noord", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Cees Snoek", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "Yuki Asano", "gender": "Male", "institution": "University of Technology Nuremberg", "country": "DE", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 35, "n_ref": 64, "n_ref_all": 85, "n_fig": 8, "n_tab": 4, "L_tab": 1155, "n_element_tab": 90, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1177, "n_element_tab_1": 96, "formula_len_all": 245, "formula_len_all_1": 245, "len_all": 124507, "len_all_1": 62741, "len_abs": 1040, "len_title": 81, "len_sents": 36375, "len_sents_1": 30335, "n_sents": 262, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 33, "L_abs": 1031, "title": "TULIP: Token-length Upgraded CLIP", "abs": "We address the challenge of representing long captions in vision-language models, such as CLIP. By design these models are limited by fixed, absolute positional encodings, restricting inputs to a maximum of 77 tokens and hindering performance on tasks requiring longer descriptions. Although recent work has attempted to overcome this limit, their proposed approaches struggle to model token relationships over longer distances and simply extend to a fixed new token length. Instead, we propose a generalizable method, named TULIP, able to upgrade the token length to any length for CLIP-like models. We do so by improving the architecture with relative position encodings, followed by a training procedure that (i) distills the original CLIP text encoder into an encoder with relative position encodings and (ii) enhances the model for aligning longer captions with images. By effectively encoding captions longer than the default 77 tokens, our model outperforms baselines on cross-modal tasks such as retrieval and text-to-image", "keywords": [ "Vision-Language Models", "CLIP", "Position Encodings", "Long captioning" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "r9mYbs8RTH", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Haishuai Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Research Professor" }, { "name": "Hao Zhang", "gender": "unknown", "institution": "Northeastern University", "country": "CN", "position": "Researcher" }, { "name": "Jiajun Bu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Jiawei Guo", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Xinxin Lu", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zihang Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 53, "n_ref": 149, "n_ref_all": 180, "n_fig": 10, "n_tab": 7, "L_tab": 2454, "n_element_tab": 239, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 2680, "n_element_tab_1": 179, "formula_len_all": 996, "formula_len_all_1": 667, "len_all": 158061, "len_all_1": 63485, "len_abs": 1393, "len_title": 103, "len_sents": 51724, "len_sents_1": 28267, "n_sents": 410, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1398, "title": "Long-form Hallucination Detection with Self-elicitation", "abs": "While Large Language Models (LLMs) have exhibited impressive performance in long-form question-answering tasks, they frequently present a hazard of producing factual inaccuracies or hallucinations. An effective strategy to mitigate this hazard is to leverage off-the-shelf LLMs to detect hallucinations after the generation. The primary challenge resides in the comprehensive elicitation of the intrinsic knowledge acquired during their pre-training phase. However, existing methods that employ complex reasoning chains predominantly fall short of addressing this issue. Moreover, since existing methods for hallucination detection tend to decompose the text into isolated statements, they are unable to understand the inherent in-context semantics in long-form content. In this paper, we propose a novel framework, SelfElicit, which synergizes the self-elicitation of intrinsic knowledge of large language models and long-form continuity understanding. Specifically, we leverage self-generated thoughts derived from prior statements as catalysts to elicit the expression of intrinsic knowledge, which is integrated with graph structures to alleviate induced hallucinations and guide the factual evaluation by effectively organizing the elicited knowledge. Extensive experiments on real-world QA datasets demonstrate the effectiveness of self-elicitation and the superiority of our proposed method.", "keywords": [ "hallucination", "knowledge graph", "large language models", "medical QA" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "r8tMECbxOl", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Charu C. Aggarwal", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Distinguished Research Staff Member" }, { "name": "Horst Samulowitz", "gender": "unknown", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Jianxi Gao", "gender": "Male", "institution": ", Rensselaer Polytechnic Institute", "country": "", "position": "Associate Professor" }, { "name": "Kavitha Srinivas", "gender": "Female", "institution": "IBM Research", "country": "", "position": "Research Staff Member" }, { "name": "Pin-Yu Chen", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Zhenhan Huang", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "PhD student" }, { "name": "Niharika Shimona Dsouza", "gender": "Female", "institution": "International Business Machines", "country": "US", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 54, "n_ref": 84, "n_ref_all": 101, "n_fig": 3, "n_tab": 5, "L_tab": 2211, "n_element_tab": 219, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1970, "n_element_tab_1": 198, "formula_len_all": 990, "formula_len_all_1": 955, "len_all": 132015, "len_all_1": 58870, "len_abs": 1036, "len_title": 89, "len_sents": 28412, "len_sents_1": 21978, "n_sents": 287, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1013, "title": "Language Models Are Good Tabular Learners", "abs": "Transformer-based language models have become the de facto standard in natural language processing. However, they underperform in the tabular data domain compared to traditional tree-based methods. We posit that current models fail to achieve the full potential of language models due to (i) heterogeneity of tabular data; and (2) challenges faced by the model in interpreting numerical values. Based on this hypothesis, we propose a method titled Tabular Domain Transformer (TDTransformer). TDTransformer has distinct embedding processes for different types of columns. The alignment layers for different types of columns transform column embeddings to a common embedding space. Besides, TDTransformer adapts piece-wise linear encoding for numerical values in transformer-based architectures. We examine the proposed method on 76 real-world tabular classification datasets from the standard OpenML benchmark. Extensive experiments indicate that TDTransformer significantly improves the state-of-the-art methods.", "keywords": [ "Language Model; Tabular Data; Natural Language Processing" ], "rating_list": [ 3, 8, 3, 6, 5 ], "soundness_list": [ 1, 3, 2, 3, 2 ], "presentation_list": [ 1, 4, 2, 4, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 2, 3, 4, 4, 4 ] }, { "paper_id": "r8t6OsLP2s", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Huayu Deng", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Xiangming Zhu", "gender": "unknown", "institution": "Shanghai Jiao Tong University", "country": "", "position": "PhD student" }, { "name": "Xiaokang Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University, China", "country": "CN", "position": "Professor" }, { "name": "Yunbo Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 31, "n_formula_1": 28, "n_ref_uni": 49, "n_ref": 137, "n_ref_all": 173, "n_fig": 10, "n_tab": 12, "L_tab": 2657, "n_element_tab": 264, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 595, "n_element_tab_1": 86, "formula_len_all": 21196, "formula_len_all_1": 19295, "len_all": 221229, "len_all_1": 100351, "len_abs": 3218, "len_title": 121, "len_sents": 64241, "len_sents_1": 38438, "n_sents": 421, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1269, "title": "Discovering Message Passing Hierarchies for Mesh-Based Physics Simulation", "abs": "Graph neural networks have emerged as a powerful tool for large-scale mesh-based physics simulation. Existing approaches primarily employ hierarchical, multi-scale message passing to capture long-range dependencies within the graph. However, these graph hierarchies are typically fixed and manually designed, which do not adapt to the evolving dynamics present in complex physical systems. In this paper, we introduce a novel neural network named DHMP, which learns **D**ynamic **H**ierarchies for **M**essage **P**assing networks through a differentiable node selection method. The key component is the *anisotropic* message passing mechanism, which operates at both intra-level and inter-level interactions. Unlike existing methods, it first supports directionally non-uniform aggregation of dynamic features between adjacent nodes within each graph hierarchy. Second, it determines node selection probabilities for the next hierarchy according to different physical contexts, thereby creating more flexible message shortcuts for learning remote node relations. Our experiments demonstrate the effectiveness of DHMP, achieving $22.7$\\% improvement on average compared to recent fixed-hierarchy message passing networks across five classic physics simulation datasets.", "keywords": [ "Physics Simulation", "Message Passing Networks" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "r8J3DSD5kF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Aaron Courville", "gender": "unknown", "institution": "University of Montreal", "country": "", "position": "Assistant Professor" }, { "name": "Rameswar Panda", "gender": "Male", "institution": "MIT-IBM Watson AI Lab", "country": "", "position": "Research Scientist" }, { "name": "Shawn Tan", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Songlin Yang", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yikang Shen", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 37, "n_ref": 70, "n_ref_all": 84, "n_fig": 8, "n_tab": 5, "L_tab": 831, "n_element_tab": 136, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 1476, "n_element_tab_1": 143, "formula_len_all": 2200, "formula_len_all_1": 2247, "len_all": 106668, "len_all_1": 58856, "len_abs": 1394, "len_title": 128, "len_sents": 26027, "len_sents_1": 23704, "n_sents": 218, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 24, "L_abs": 1409, "title": "Stick-breaking Attention", "abs": "The self-attention mechanism traditionally relies on the softmax operator, necessitating positional embeddings like RoPE, or position biases to account for token order.\nBut current methods using still face length generalisation challenges.\nWe investigate an alternative attention mechanism based on the stick-breaking process in larger scale settings.\nThe method works as follows: For each token before the current, we determine a break point, which represents the proportion of the stick to allocate to the current token.\nThis process is repeated on the remaining stick, until all tokens are allocated, resulting in a sequence of attention weights.\nThis process naturally incorporates recency bias, which has linguistic motivations for grammar parsing (Shen et. al. 2017).\nWe study the implications of replacing the conventional softmax-based attention mechanism with stick-breaking attention.\nWe then discuss implementation of numerically stable stick-breaking attention and adapt Flash Attention to accommodate this mechanism.\nWhen used as a drop-in replacement for current softmax+RoPE attention systems, we find that stick-breaking attention performs competitively with current methods on length generalisation and downstream tasks.\nStick-breaking also performs well at length generalisation, allowing a model trained with $2^{11}$ context window to perform well at $2^{14}$ with perplexity improvements.", "keywords": [ "transformer", "attention", "stick-breaking", "softmax", "length extrapolation" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "r8H7xhYPwz", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Ali Hatamizadeh", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research scientist" }, { "name": "Jan Kautz", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "VP Research" }, { "name": "Songlin Yang", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 25, "n_ref_uni": 54, "n_ref": 134, "n_ref_all": 155, "n_fig": 3, "n_tab": 6, "L_tab": 3787, "n_element_tab": 397, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1678, "n_element_tab_1": 124, "formula_len_all": 3527, "formula_len_all_1": 2862, "len_all": 202427, "len_all_1": 54968, "len_abs": 1100, "len_title": 102, "len_sents": 31033, "len_sents_1": 22212, "n_sents": 234, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1104, "title": "Gated Delta Networks: Improving Mamba2 with Delta Rule", "abs": "Linear Transformers have gained attention as efficient alternatives to standard Transformers, but their performance in retrieval and long-context tasks has been limited. To address these limitations, recent work has explored two distinct mechanisms: gating for adaptive memory control and the delta update rule for precise memory modifications. We observe that these mechanisms are complementary\u2014gating enables rapid memory erasure while the delta rule facilitates targeted updates. Building on this insight, we introduce the gated delta rule and develop a parallel training algorithm optimized for modern hardware. Our proposed architecture, Gated DeltaNet, consistently surpasses existing models like Mamba2 and DeltaNet across multiple benchmarks, including language modeling, common-sense reasoning, in-context retrieval, length extrapolation, and long-context understanding. We further enhance performance by developing hybrid architectures that combine Gated DeltaNet layers with sliding window attention or Mamba2 layers, achieving both improved training efficiency and superior task performance.", "keywords": [ "linear RNN", "state-space model", "linear transformer", "subquadractic model", "linear attention", "delta rule", "mamba" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 5, 4, 3, 2 ] }, { "paper_id": "r8C9nt0nlc", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 1, "author_info_list": [ { "name": "Pedro Zuidberg Dos Martires", "gender": "Male", "institution": "KU Leuven", "country": "", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 19, "n_ref_uni": 16, "n_ref": 23, "n_ref_all": 47, "n_fig": 4, "n_tab": 6, "L_tab": 1514, "n_element_tab": 294, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 635, "n_element_tab_1": 120, "formula_len_all": 2262, "formula_len_all_1": 1733, "len_all": 81607, "len_all_1": 63193, "len_abs": 679, "len_title": 150, "len_sents": 26586, "len_sents_1": 23331, "n_sents": 250, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 683, "title": "The Normalized Float Trick: Numerical Stability for Probabilistic Circuits without the LogSumExp Trick", "abs": "Probabilistic circuits (PCs) are a class of tractable deep probabilistic models that compute event probabilities by recursively nesting sum and product computations. Unfortunately, this is numerically unstable. To mitigate this numerical stability issues, PCs are usually evaluated in log-space via the LogSumExp trick. In this paper we present an alternative to the ubiquitous LogSumExp trick, which we dub \"normalized float trick\". Experimentally, we show that by simply changing the scheme guaranteeing numerical stability (from the LogSumExp to the normalized float trick) we can consistently and considerably boost the performance of PCs on common density estimation benchmarks,", "keywords": [ "probabilistic circuits", "tractable probabilistic models" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "r7wMVdGFro", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Lukas Wutschitz", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Matthieu Meeus", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Reza Shokri", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" }, { "name": "Santiago Zanella-Beguelin", "gender": "Male", "institution": "Microsoft Research", "country": "GB", "position": "Principal Researcher" }, { "name": "Shruti Tople", "gender": "Female", "institution": "Microsoft Research", "country": "GB", "position": "Senior Researcher" } ], "n_formula": 22, "n_formula_1": 9, "n_ref_uni": 38, "n_ref": 113, "n_ref_all": 142, "n_fig": 10, "n_tab": 12, "L_tab": 2328, "n_element_tab": 230, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 327, "n_element_tab_1": 43, "formula_len_all": 2002, "formula_len_all_1": 837, "len_all": 194115, "len_all_1": 65549, "len_abs": 1227, "len_title": 121, "len_sents": 60808, "len_sents_1": 32185, "n_sents": 431, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1240, "title": "The Canary\u2019s Echo: Auditing Privacy Risks of LLM-Generated Synthetic Text", "abs": "How much information about training examples can be gleaned from synthetic data generated by Large Language Models (LLMs)? Overlooking the subtleties of information flow in synthetic data generation pipelines can lead to a false sense of privacy. In this paper, we investigate the design of membership inference attacks (MIAs) that target data used to fine-tune pre-trained LLMs that are then used to synthesize data, particularly when the adversary does not have access to the fine-tuned LLM but only to a synthetic data corpus. We demonstrate that using canaries crafted to maximize their vulnerability to attacks that have access to the model are sub-optimal for auditing privacy risks when only synthetic data is released. This is because such out-of-distribution canaries have limited influence on the model\u2019s output when prompted to generate useful, in-distribution synthetic data, thus significantly limiting their vulnerability to MIAs. To tackle this problem, we leverage the mechanics of auto-regressive models to design canaries that leave detectable traces in synthetic data. Our approach significantly enhances the power of MIAs, providing a better assessment of the privacy risks of releasing synthetic data generated by LLMs.", "keywords": [ "Privacy", "language models", "synthetic data" ], "rating_list": [ 6, 5, 8, 5 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "r6aX67YhD9", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Xiaojun Xu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Yang Liu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "Kevin Yao", "gender": "Male", "institution": "Meta GenAI", "country": "US", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 5, "n_ref_uni": 20, "n_ref": 40, "n_ref_all": 69, "n_fig": 4, "n_tab": 14, "L_tab": 9444, "n_element_tab": 646, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1231, "n_element_tab_1": 142, "formula_len_all": 569, "formula_len_all_1": 354, "len_all": 165734, "len_all_1": 67875, "len_abs": 1231, "len_title": 115, "len_sents": 48156, "len_sents_1": 28505, "n_sents": 442, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1211, "title": "Learning to Watermark LLM-generated Text via Reinforcement Learning", "abs": "We study how to watermark LLM outputs, i.e. embedding algorithmically detectable signals into LLM-generated text to track misuse. Unlike the current mainstream methods that work with a fixed LLM, we expand the watermark design space by including the LLM tuning stage in the watermark pipeline. While prior works focus on token-level watermark that embeds signals into the output, we design a model-level watermark that embeds signals into the LLM weights, and such signals can be detected by a paired detector. We propose a co-training framework based on reinforcement learning that iteratively (1) trains a detector to detect the generated watermarked text and (2) tunes the LLM to generate text easily detectable by the detector while keeping its normal utility. We empirically show that our watermarks are more accurate, robust, and adaptable (to new attacks) with no generation overhead. It also allows watermarked model open-sourcing. In addition, if used together with alignment, the extra overhead introduced is low -- only training an extra reward model (i.e. our detector). We hope our work can bring more effort into studying a broader watermark design that is not limited to working with a fixed LLM.", "keywords": [ "LLM Watermark" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "r6XqXoRT6N", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Kui Ren", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yichen Sun", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhan Qin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zhixuan Chu", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 50, "n_ref_all": 81, "n_fig": 12, "n_tab": 9, "L_tab": 1823, "n_element_tab": 181, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3394, "n_element_tab_1": 192, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 136946, "len_all_1": 59699, "len_abs": 1332, "len_title": 182, "len_sents": 38152, "len_sents_1": 26631, "n_sents": 284, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 133, "L_abs": 1336, "title": "Prompt-Consistency Image Generation (PCIG): A Unified Framework Integrating LLMs, Knowledge Graphs, and Controllable Diffusion Models", "abs": "The rapid advancement of Text-to-Image(T2I) generative models has enabled the synthesis of high-quality images guided by textual descriptions. Despite this significant progress, these models are often susceptible in generating contents that contradict the input text, which poses a challenge to their reliability and practical deployment. To address this problem, we introduce a novel diffusion-based framework to significantly enhance the alignment of generated images with their corresponding descriptions, addressing the inconsistency between visual output and textual input. Our framework is built upon a comprehensive analysis of inconsistency phenomena, categorizing them based on their manifestation in the image. Leveraging a state-of-the-art large language module, we first extract objects and construct a knowledge graph to predict the locations of these objects in potentially generated images. We then integrate a state-of-the-art controllable image generation model with a visual text generation module to generate an image that is consistent with the original prompt, guided by the predicted object locations. Through extensive experiments on an advanced multimodal hallucination benchmark, we demonstrate the efficacy of our approach in accurately generating the images without the inconsistency with the original prompt.", "keywords": [ "multimodal hallucination", "diffusion-based model", "knowledge graph", "large language model" ], "rating_list": [ 5, 3, 5, 3, 5 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 3, 1, 3, 2, 2 ], "contribution_list": [ 2, 1, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 5, 5 ] }, { "paper_id": "r5yolhcrHe", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "JiaxinZhang", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Jingkuan Song", "gender": "Male", "institution": "University of Electronic Science and Technology of China,", "country": "CN", "position": "Full Professor" }, { "name": "Junchen Zhu", "gender": "Male", "institution": "Skywork AI", "country": "CN", "position": "Researcher" }, { "name": "Lianli Gao", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Pengpeng Zeng", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Postdoc" }, { "name": "Sitong Su", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Hengtao Shen", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "xiaocai", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 30, "n_ref": 54, "n_ref_all": 83, "n_fig": 16, "n_tab": 4, "L_tab": 580, "n_element_tab": 93, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 692, "n_element_tab_1": 86, "formula_len_all": 699, "formula_len_all_1": 699, "len_all": 138148, "len_all_1": 61282, "len_abs": 1661, "len_title": 170, "len_sents": 46997, "len_sents_1": 29852, "n_sents": 355, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 122, "L_abs": 1633, "title": "SeMv-3D: Towards Semantic and Mutil-view Consistency simultaneously for General Text-to-3D Generation with Triplane Priors", "abs": "Recent advancements in generic 3D content generation from text prompts have been remarkable by fine-tuning text-to-image diffusion (T2I) models or employing these T2I models as priors to learn a general text-to-3D model. While fine-tuning-based methods ensure great alignment between text and generated views, i.e., semantic consistency, their ability to achieve multi-view consistency is hampered by the absence of 3D constraints, even in limited view. In contrast, prior-based methods focus on regressing 3D shapes with any view that maintains uniformity and coherence across views, i.e., multi-view consistency, but such approaches inevitably compromise visual-textual alignment, leading to a loss of semantic details in the generated objects. To achieve semantic and multi-view consistency simultaneously, we propose SeMv-3D, a novel framework for general text-to-3d generation. Specifically, we propose a Triplane Prior Learner (TPL) that learns triplane priors with 3D spatial features to maintain consistency among different views at the 3D level, e.g., geometry and texture. Moreover, we design a Semantic-aligned View Synthesizer (SVS) that preserves the alignment between 3D spatial features and textual semantics in latent space. In SVS, we devise a simple yet effective batch sampling and rendering strategy that can generate arbitrary views in a single feed-forward inference. Extensive experiments present our SeMv-3D's superiority over state-of-the-art performances with semantic and multi-view consistency in any view. Our code and more visual results are available at https://anonymous.4open.science/r/SeMv-3D-6425.", "keywords": [ "General Text-to-3D Generation", "Generative Models", "Diffusion Models" ], "rating_list": [ 6, 5, 6, 5, 3 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 4, 5 ] }, { "paper_id": "r5d8zkYizS", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Bernardo Cuenca Grau", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Professor" }, { "name": "Ouns El Harzli", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 31, "n_ref_uni": 44, "n_ref": 65, "n_ref_all": 80, "n_fig": 9, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 171, "n_element_tab_1": 3, "formula_len_all": 2945, "formula_len_all_1": 3306, "len_all": 136786, "len_all_1": 67390, "len_abs": 1576, "len_title": 137, "len_sents": 39379, "len_sents_1": 30168, "n_sents": 261, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1582, "title": "Adversarial Attacks as Near-Zero Eigenvalues in the Empirical Kernel of Neural Networks", "abs": "Adversarial examples ---imperceptibly modified data inputs designed to mislead machine learning models--- have raised concerns about the robustness of modern neural architectures in safety-critical applications. \nIn this paper, we propose a unified mathematical \nframework for understanding adversarial examples in neural networks, corroborating Szegedy et al.'s original conjecture that \nsuch examples are exceedingly rare, despite their presence in the proximity of nearly every test case. By exploiting Mercer's decomposition theorem, we characterise adversarial examples as those producing near-zero Mercer's eigenvalues in the empirical kernel associated to a trained neural network. \nConsequently, the generation of adversarial attacks, using any known technique, can be conceptualised as a progression towards the eigenvalue space's zero point within the empirical kernel.\nWe rigorously prove this characterisation for trained neural networks that achieve interpolation and under mild assumptions on the architecture, thus\nproviding a mathematical explanation for the apparent contradiction of neural networks excelling at generalisation while remaining vulnerable to adversarial attacks. \nWe have empirically verified that adversarial examples generated for both fully-connected and convolutional architectures through the widely-known DeepFool algorithm and through the more recent Fast Adaptive Boundary (FAB) method consistently lead to a shift in the distribution of Mercer's eigenvalues toward zero. These results are in strong agreement with predictions of our theory.", "keywords": [ "adversarial attacks", "kernel", "neural networks" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "r5IXBlTCGc", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Abhimanyu Pallavi Sudhir", "gender": "Male", "institution": "University of Warwick", "country": "GB", "position": "PhD student" }, { "name": "Adam Shen", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "MS student" }, { "name": "Alejandro Alvarez", "gender": "Male", "institution": "Independent", "country": "", "position": "Emeritus" }, { "name": "Daniel Paleka", "gender": "unknown", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "", "position": "PhD student" }, { "name": "Evan Wang", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "MS student" }, { "name": "Vineeth Bhat", "gender": "Male", "institution": "International Institute of Information Technology, Hyderabad, International Institute of Information Technology Hyderabad", "country": "IN", "position": "Undergrad student" }, { "name": "Florian Tramer", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 67, "n_formula_1": 6, "n_ref_uni": 32, "n_ref": 52, "n_ref_all": 101, "n_fig": 18, "n_tab": 16, "L_tab": 5745, "n_element_tab": 396, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 887, "n_element_tab_1": 113, "formula_len_all": 7278, "formula_len_all_1": 446, "len_all": 302449, "len_all_1": 56240, "len_abs": 3461, "len_title": 49, "len_sents": 111863, "len_sents_1": 24675, "n_sents": 1011, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 49, "L_abs": 1340, "title": "Consistency Checks for Language Model Forecasters", "abs": "Forecasting is a task that is difficult to evaluate: the ground truth can only be known in the future. Recent work showing LLM forecasters rapidly approaching human-level performance begs the question: how can we benchmark and evaluate these forecasters *instantaneously*? Following the consistency check framework, we measure the performance of forecasters in terms of the consistency of their predictions on different logically-related questions. We propose a new, general consistency metric based on *arbitrage*: for example, if a forecasting AI illogically predicts that both the Democratic and Republican parties have 60\\% probability of winning the 2024 US presidential election, an arbitrageur could trade against the forecaster's predictions and make a profit. We build an automated evaluation system that generates a set of base questions, instantiates consistency checks from these questions, elicits the predictions of the forecaster, and measures the consistency of the predictions. We then build a standard, proper-scoring-rule forecasting benchmark, and show that our (instantaneous) consistency metrics correlate strongly with LLM forecasters' ground truth Brier scores (which are only known in the future). We also release a consistency benchmark that resolves in 2028, providing a long-term evaluation tool for forecasting.", "keywords": [ "forecasting", "markets", "trading", "LLM", "evaluation", "eval", "consistency", "robustness" ], "rating_list": [ 8, 5, 8, 8 ], "soundness_list": [ 4, 2, 4, 3 ], "presentation_list": [ 4, 2, 3, 4 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "r4Q86nBQka", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Andrei Purica", "gender": "Male", "institution": "Thales", "country": "FR", "position": "Principal Researcher" }, { "name": "Edouard Pauwels", "gender": "Male", "institution": "TSE", "country": "FR", "position": "Full Professor" }, { "name": "Jerome Bolte", "gender": "Male", "institution": "Toulouse School of Economics", "country": "", "position": "Professor" }, { "name": "Ryan Boustany", "gender": "Male", "institution": "Universit\u00e9 Toulouse 1 Capitole", "country": "", "position": "PhD student" } ], "n_formula": 61, "n_formula_1": 16, "n_ref_uni": 65, "n_ref": 124, "n_ref_all": 159, "n_fig": 9, "n_tab": 5, "L_tab": 1495, "n_element_tab": 237, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 193, "n_element_tab_1": 14, "formula_len_all": 6133, "formula_len_all_1": 1079, "len_all": 163045, "len_all_1": 53101, "len_abs": 720, "len_title": 78, "len_sents": 34526, "len_sents_1": 23079, "n_sents": 351, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 686, "title": "A second-order-like optimizer with adaptive gradient scaling for deep learning", "abs": "In this empirical article, we introduce INNAprop, an optimization algorithm that combines the INNA method with the RMSprop adaptive gradient scaling. It leverages second-order information and rescaling while keeping the memory requirements of standard DL methods as AdamW or SGD with momentum. After having recalled our geometrical motivations, we provide quite extensive experiments. On image classification (CIFAR-10, ImageNet) and language modeling (GPT-2), INNAprop consistently matches or outperforms AdamW both in training speed and accuracy, with minimal hyperparameter tuning in large-scale settings. Our code is publicly available at \\url{https://github.com/innaprop/innaprop}.", "keywords": [ "deep learning", "second-order methods", "stochastic optimization", "dynamical systems" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 2, 2, 4, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "r4GxmIBDbO", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Koichi Shinoda", "gender": "Male", "institution": "Institute of Science Tokyo", "country": "JP", "position": "Professor" }, { "name": "Tat-Jen Cham", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Neo Wei Ming", "gender": "Male", "institution": "Tokyo Institute of Technology, Tokyo Institute of Technology", "country": "JP", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 122, "n_ref_all": 155, "n_fig": 13, "n_tab": 5, "L_tab": 2609, "n_element_tab": 464, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 849, "n_element_tab_1": 157, "formula_len_all": 207, "formula_len_all_1": 193, "len_all": 144269, "len_all_1": 63645, "len_abs": 1347, "len_title": 102, "len_sents": 40296, "len_sents_1": 31128, "n_sents": 246, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 1351, "title": "Diffusion Pretraining for Gait Recognition in the Wild", "abs": "Recently, diffusion models have garnered much attention for their remarkable generative capabilities. Yet, their application for representation learning remains largely unexplored. In this paper, we explore the possibility of using the diffusion process to pretrain the backbone of a deep learning model for a specific application\u2014gait recognition in the wild. To do so, we condition a latent diffusion model on the output of a gait recognition model backbone. Our pretraining experiments on the Gait3D and GREW datasets reveal an interesting phenomenon: diffusion pretraining causes the gait recognition backbone to separate gait sequences belonging to different subjects further apart than those belonging to the same subjects, which translates to a steady improvement in gait recognition performance. Subsequently, our transfer learning experiments on Gait3D and GREW show that the pretrained backbone can serve as an effective initialization for the downstream gait recognition task, allowing the gait recognition model to achieve better performance within much fewer supervised training iterations. We validated the applicability of our approach across multiple existing gait recognition methods and conducted extensive ablation studies to investigate the impact of different pretraining hyperparameters on the final gait recognition performance.", "keywords": [ "Diffusion Models", "Gait Recognition", "Representation Learning" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 4, 3, 4, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 4, 5 ] }, { "paper_id": "r3cWq6KKbt", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Connor Mooney", "gender": "unknown", "institution": "University of California, Irvine", "country": "US", "position": "Associate Professor" }, { "name": "Jack Xin", "gender": "Male", "institution": "University of California, Irvine", "country": "", "position": "Chancellor's Professor" }, { "name": "Yifeng Yu", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Full Professor" }, { "name": "Zhongjian Wang", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 143, "n_formula_1": 41, "n_ref_uni": 30, "n_ref": 62, "n_ref_all": 100, "n_fig": 0, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 71, "n_element_tab_1": 6, "formula_len_all": 14242, "formula_len_all_1": 3192, "len_all": 158641, "len_all_1": 56643, "len_abs": 866, "len_title": 157, "len_sents": 46178, "len_sents_1": 22042, "n_sents": 513, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 109, "L_abs": 880, "title": "Global Well-posedness and Convergence Analysis of Score-based Generative Models via Sharp Lipschitz Estimates", "abs": "We establish global well-posedness and convergence of the score-based generative models (SGM) under minimal general assumptions of initial data for score estimation. For the smooth case, we start from a Lipschitz bound of the score function with optimal time length. The optimality is validated by an example whose Lipschitz constant of scores is bounded at initial but blows up in finite time. This necessitates the separation of time scales in conventional bounds for non-log-concave distributions. In contrast, our follow up analysis only relies on a local Lipschitz condition and is valid globally in time. This leads to the convergence of numerical scheme without time separation. For the non-smooth case, we show that the optimal Lipschitz bound is $O(1/t)$ in the point-wise sense for distributions supported on a compact, smooth and low-dimensional manifold with boundary.", "keywords": [ "Score based generative models", "Lipschitz estimates", "convergence analysis", "well-posedness", "singularity" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "r3cEOVj7Ze", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "David Hsu", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Professor" }, { "name": "Gim Hee Lee", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" }, { "name": "Zilin Fang", "gender": "unknown", "institution": "National University of Singapore", "country": "", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 47, "n_ref": 89, "n_ref_all": 116, "n_fig": 8, "n_tab": 8, "L_tab": 3638, "n_element_tab": 302, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 3171, "n_element_tab_1": 287, "formula_len_all": 1004, "formula_len_all_1": 1004, "len_all": 137213, "len_all_1": 63504, "len_abs": 952, "len_title": 140, "len_sents": 32268, "len_sents_1": 27001, "n_sents": 220, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 957, "title": "Neuralized Markov Random Field for Interaction-Aware Stochastic Human Trajectory Prediction", "abs": "Interactive human motions and the continuously changing nature of intentions pose significant challenges for human trajectory prediction. In this paper, we present a neuralized Markov random field (MRF)-based motion evolution method for probabilistic interaction-aware human trajectory prediction. We use MRF to model each agent's motion and the resulting crowd interactions over time, hence is robust against noisy observations and enables group reasoning. We approximate the modeled distribution using two conditional variational autoencoders (CVAEs) for efficient learning and inference. Our proposed method achieves state-of-the-art performance on ADE/FDE metrics across two dataset categories: overhead datasets ETH/UCY, SDD, and NBA, and ego-centric JRDB. Furthermore, our approach allows for real-time stochastic inference in bustling environments, making it well-suited for a 30FPS video setting. We will open-source our codes upon paper acceptance.", "keywords": [ "human trajectory prediction", "interaction modeling" ], "rating_list": [ 5, 6, 5, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 2, 2, 5 ] }, { "paper_id": "r3GxWNGpSj", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 13, "author_info_list": [ { "name": "Bing Qin", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Dandan Tu", "gender": "Female", "institution": "ICT, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Duyu Tang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Lei Huang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Libo Qin", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "Full Professor" }, { "name": "Weitao Ma", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Xiachong Feng", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "Postdoc" }, { "name": "Xiaocheng Feng", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xiaohui Yan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "experter" }, { "name": "Yangfan Ye", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Yichong Huang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yunfei Lu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zhirui Zhang", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Senior Researcher" } ], "n_formula": 27, "n_formula_1": 16, "n_ref_uni": 44, "n_ref": 153, "n_ref_all": 218, "n_fig": 11, "n_tab": 7, "L_tab": 19822, "n_element_tab": 891, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1980, "n_element_tab_1": 203, "formula_len_all": 3853, "formula_len_all_1": 1982, "len_all": 338478, "len_all_1": 129844, "len_abs": 1784, "len_title": 123, "len_sents": 101315, "len_sents_1": 54850, "n_sents": 677, "n_sents_1": 406, "venue": "ICLR2025", "decision": "withdraw", "L_title": 122, "L_abs": 1874, "title": "XTransplant: A Probe into the Upper Bound Performance of Multilingual Capability in LLMs via Cross-lingual Transplantation", "abs": "Current large language models (LLMs) often display significant imbalances in their multilingual capabilities and cultural adaptability, primarily due to their unbalanced and English-centric pretraining data.\nFor these English-centric LLMs, the disparities between English and non-English languages hinder their ability to utilize their robust English-based capabilities within non-English contexts, while also limiting access to valuable multilingual knowledge derived from non-English \"language-specific neurons\" within English contexts.\nMotivated by this, our work explores the possibility for LLMs to leverage the strengths of both English and non-English languages, aiming to further unlock their multilingual potential.\nTo this end, we propose a probing method named $\\mathcal{X}$Transplant, which directly transplants feed-forward activations from English input to non-English (or from non-English to English) during inference stage, allowing the model to benefit from both English and additional multilingual knowledge.\nThrough extensive experiments on our pilotsets and representative LLMs across different tasks and languages, we empirically prove that both the multilingual capabilities and cultural adaptability of LLMs hold the potential to be significantly improved by the cross-lingual feed forward transplantation, respectively from $\\texttt{En} \\rightarrow \\texttt{non-En}$ and $\\texttt{non-En} \\rightarrow \\texttt{En}$. \nAdditionally, we also establish the upper bound performance of LLMs obtained through $\\mathcal{X}$Transplant (relative growth of +80\\% in multilingual capabilities, +39\\% in cultural adaptability), highlighting the underutilization of current LLMs' multilingual potential. \nWe do hope our further analysis and discussion could suggest promising directions for deeply unlocking the multilingual potential of current English-centric LLMs.", "keywords": [ "Large language models", "Multilingual capability", "Feed forward activations", "Upper Bound Performance" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "r3DF5sOo5B", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Jason D. Lee", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Jianhao Huang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Undergrad student" }, { "name": "Zixuan Wang", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" } ], "n_formula": 278, "n_formula_1": 28, "n_ref_uni": 28, "n_ref": 104, "n_ref_all": 129, "n_fig": 2, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 479, "n_element_tab_1": 6, "formula_len_all": 42546, "formula_len_all_1": 2820, "len_all": 316782, "len_all_1": 70903, "len_abs": 3348, "len_title": 129, "len_sents": 77681, "len_sents_1": 29692, "n_sents": 856, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 81, "L_abs": 1029, "title": "Transformers Learn to Implement Multi-step Gradient Descent with Chain of Thought", "abs": "Chain of Thought (CoT) prompting has been shown to significantly improve the performance of large language models (LLMs), particularly in arithmetic and reasoning tasks, by instructing the model to produce intermediate reasoning steps. Despite the remarkable empirical success of CoT and its theoretical advantages in enhancing expressivity, the mechanisms underlying CoT training remain largely unexplored. In this paper, we study the training dynamics of transformers over a CoT objective on a in-context weight prediction task for linear regression. We prove that while a one-layer linear transformer without CoT can only implement a single step of gradient descent (GD) and fails to recover the ground-truth weight vector, a transformer with CoT prompting can learn to perform multi-step GD autoregressively, achieving near-exact recovery. Furthermore, we show that the trained transformer effectively generalizes on the unseen data. Empirically, we demonstrate that CoT prompting yields substantial performance improvements.", "keywords": [ "Chain of Thought", "Transformer optimization", "Training dynamics" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "r2nwBwodth", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Einari Vaaras", "gender": "Male", "institution": "Tampere University", "country": "", "position": "PhD student" }, { "name": "Manu Airaksinen", "gender": "Not Specified", "institution": "Helsinki University Hospital", "country": "FI", "position": "Postdoc" }, { "name": "Okko Johannes R\u00e4s\u00e4nen", "gender": "Male", "institution": "Tampere University", "country": "", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 45, "n_ref": 108, "n_ref_all": 134, "n_fig": 1, "n_tab": 13, "L_tab": 4497, "n_element_tab": 569, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 756, "n_element_tab_1": 54, "formula_len_all": 520, "formula_len_all_1": 165, "len_all": 173511, "len_all_1": 64454, "len_abs": 1711, "len_title": 130, "len_sents": 47782, "len_sents_1": 32512, "n_sents": 324, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1719, "title": "PFML: Self-Supervised Learning of Time-Series Data Without Representation Collapse", "abs": "Self-supervised learning (SSL) is a data-driven learning approach that utilizes the innate structure of the data to guide the learning process. In contrast to supervised learning, which depends on external labels, SSL utilizes the inherent characteristics of the data to produce its own supervisory signal. However, one frequent issue with SSL methods is representation collapse, where the model outputs a constant input-invariant feature representation. This issue hinders the potential application of SSL methods to new data modalities, as trying to avoid representation collapse wastes researchers' time and effort. This paper introduces a novel SSL algorithm for time-series data called Prediction of Functionals from Masked Latents (PFML). Instead of predicting masked input signals or their latent representations directly, PFML operates by predicting statistical functionals of the input signal corresponding to masked embeddings, given a sequence of unmasked embeddings. The algorithm is designed to avoid representation collapse, rendering it straightforwardly applicable to different time-series data domains, such as novel sensor modalities in clinical data. We demonstrate the effectiveness of PFML through complex, real-life classification tasks across three different data modalities: infant posture and movement classification from multi-sensor inertial measurement unit data, emotion recognition from speech data, and sleep stage classification from EEG data. The results show that PFML is superior to a conceptually similar pre-existing SSL method and competitive against the current state-of-the-art SSL method, while also being conceptually simpler and without suffering from representation collapse.", "keywords": [ "self-supervised learning", "statistical functionals", "representation collapse", "masking", "time-series data" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 1, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "r27Nwu0t86", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Shangzhe Li", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Xinhua Zhang", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "", "position": "Associate Professor" } ], "n_formula": 21, "n_formula_1": 18, "n_ref_uni": 24, "n_ref": 58, "n_ref_all": 85, "n_fig": 11, "n_tab": 8, "L_tab": 3475, "n_element_tab": 308, "n_fig_1": 10, "n_tab_1": 5, "L_tab_1": 2900, "n_element_tab_1": 253, "formula_len_all": 1595, "formula_len_all_1": 1333, "len_all": 128653, "len_all_1": 68235, "len_abs": 1095, "len_title": 118, "len_sents": 39931, "len_sents_1": 27609, "n_sents": 344, "n_sents_1": 231, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1145, "title": "Augmenting Offline Reinforcement Learning with State-only Interactions", "abs": "Batch offline data have been shown considerably beneficial for reinforcement learning. Their benefit is further amplified by upsampling with generative models. In this paper, we consider a novel opportunity where interaction with environment is feasible, but only restricted to observations, i.e. *no reward* feedback is available. This setting is realistic, because simulators or even real cyber-physical systems are often accessible, while in contrast reward is often difficult or expensive to obtain, similar to imitation learning settings. As a result, the learner must make best sense of the offline data to synthesize the most sample-efficient scheme of querying the transition of observation. Our method first leverages online interactions to generate high-return trajectories via conditional diffusion models. They are then blended with the original offline trajectories through a stitching algorithm, and the resulting augmented data is applied to downstream reinforcement learner. Superior empirical performance is demonstrated over state-of-the-art data augmentation methods that are extended to utilize observation-only interactions.", "keywords": [ "Offline Reinforcement Learning", "Generative Model", "Data Augmentation", "Trajectory Stitching" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "r1cbFEH0Df", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Dmitrii Khizbullin", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Researcher" }, { "name": "Francesco Faccio", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Hisham Abdullah Alyahya", "gender": "Male", "institution": "Saudi Data and AI Authority", "country": "SA", "position": "Researcher" }, { "name": "J\u00fcrgen Schmidhuber", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Director, AI Initiative" }, { "name": "Oleg Serikov", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Consultant" }, { "name": "Wenyi Wang", "gender": "unknown", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Dylan Ashley", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 34, "n_ref": 59, "n_ref_all": 92, "n_fig": 16, "n_tab": 7, "L_tab": 5491, "n_element_tab": 256, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 166, "n_element_tab_1": 27, "formula_len_all": 1312, "formula_len_all_1": 1343, "len_all": 153186, "len_all_1": 67745, "len_abs": 1406, "len_title": 127, "len_sents": 51351, "len_sents_1": 32311, "n_sents": 366, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1412, "title": "How to Correctly Do Semantic Backpropagation on Language-based Agentic Systems", "abs": "Language-based agentic systems have shown great promise in recent years, transitioning from solving small-scale research problems to being deployed in challenging real-world tasks. However, optimizing these systems often requires substantial manual labor. Recent studies have demonstrated that these systems can be represented as computational graphs, enabling automatic optimization. Despite these advancements, most current efforts in Graph-based Agentic System Optimization (GASO) fail to properly assign feedback to the system\u2019s components given feedback on the system\u2019s output. To address this challenge, we formalize the concept of semantic backpropagation with semantic gradients\u2014a generalization that aligns several key optimization techniques, including reverse-mode automatic differentiation and the more recent TextGrad by exploiting the relationship among nodes with a common successor. This serves as a method for computing directional information about how changes to each component of an agentic system might improve the system\u2019s output. To use these gradients, we propose a method called semantic gradient descent which enables us to solve GASO effectively. Our results on both BIG-Bench Hard and GSM8K show that our approach outperforms existing state-of-the-art methods for solving GASO problems. A detailed ablation study on the LIAR dataset demonstrates the parsimonious nature of our method.", "keywords": [ "Agentic System", "Large Language Model", "Backpropagation", "Computational Graph" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "r1KcapkzCt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Ling Chen", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "Full Professor" }, { "name": "Meng Fang", "gender": "Male", "institution": "University of Liverpool", "country": "", "position": "Assistant Professor" }, { "name": "Zijing Shi", "gender": "Female", "institution": "University of Technology Sydney", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 37, "n_ref": 67, "n_ref_all": 79, "n_fig": 2, "n_tab": 8, "L_tab": 7000, "n_element_tab": 256, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1839, "n_element_tab_1": 238, "formula_len_all": 1004, "formula_len_all_1": 584, "len_all": 153394, "len_all_1": 71406, "len_abs": 1299, "len_title": 115, "len_sents": 47328, "len_sents_1": 31346, "n_sents": 341, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1305, "title": "Monte Carlo Planning with Large Language Model for Text-Based Games", "abs": "Text-based games provide valuable environments for language-based autonomous agents. However, planning-then-learning paradigms, such as those combining Monte Carlo Tree Search (MCTS) and reinforcement learning (RL), are notably time-consuming due to extensive iterations. Additionally, these algorithms perform uncertainty-driven exploration but lack language understanding and reasoning abilities.\nIn this paper, we introduce the Monte Carlo planning with Dynamic Memory-guided Large language model (MC-DML) algorithm. MC-DML leverages the language understanding and reasoning capabilities of Large Language Models (LLMs) alongside the exploratory advantages of tree search algorithms. Specifically, we enhance LLMs with in-trial and cross-trial memory mechanisms, enabling them to learn from past experiences and dynamically adjust action evaluations during planning. We conduct experiments on a series of text-based games from the Jericho benchmark. Our results demonstrate that the MC-DML algorithm significantly enhances performance across various games at the initial planning phase, outperforming strong contemporary methods that require multiple iterations. This demonstrates the effectiveness of our algorithm, paving the way for more efficient language-grounded planning in complex environments.", "keywords": [ "Large language model", "Monte Carlo tree search", "Text-based games" ], "rating_list": [ 8, 8, 5 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 4, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "r0pLGGcuY6", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Dorsa Sadigh", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Jeannette Bohg", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Priya Sundaresan", "gender": "Female", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Quan Vuong", "gender": "Male", "institution": "physical intelligence", "country": "", "position": "Researcher" }, { "name": "Hengyuan Hu", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 53, "n_ref_all": 84, "n_fig": 7, "n_tab": 2, "L_tab": 475, "n_element_tab": 47, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 486, "formula_len_all_1": 272, "len_all": 142947, "len_all_1": 68952, "len_abs": 1737, "len_title": 109, "len_sents": 52322, "len_sents_1": 35985, "n_sents": 353, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1793, "title": "What's the Move? Hybrid Imitation Learning via Salient Points", "abs": "While imitation learning (IL) offers a promising framework for teaching robots various behaviors, learning complex tasks remains challenging. Existing IL policies struggle to generalize effectively across visual and spatial variations even for simple tasks. In this work, we introduce **SPHINX**: **S**alient **P**oint-based **H**ybrid **I**mitatio**N** and e**X**ecution, a flexible IL policy that leverages multimodal observations (point clouds and wrist images), along with a hybrid action space of low-frequency, sparse waypoints and high-frequency, dense end effector movements. Given 3D point cloud observations, SPHINX learns to infer task-relevant points within a point cloud, or *salient points*, which support spatial generalization by focusing on semantically meaningful features. These salient points serve as anchor points to predict waypoints for long-range movement, such as reaching target poses in free-space. Once near a salient point, SPHINX learns to switch to predicting dense end-effector movements given close-up wrist images for precise phases of a task. By exploiting the strengths of different input modalities and action representations for different manipulation phases, SPHINX tackles complex tasks in a sample-efficient, generalizable manner. Our method achieves **86.7%** success across 4 real-world and 2 simulated tasks, outperforming the next best state-of-the-art IL baseline by **41.1%** on average across **440** real world trials. SPHINX additionally generalizes to novel viewpoints, visual distractors, spatial arrangements, and execution speeds with a **1.7x** speedup over the most competitive baseline. Our website (http://sphinx-il.github.io) provides open-sourced code for data collection, training, and evaluation, along with supplementary videos.", "keywords": [ "Imitation Learning", "Robot Learning", "Robot Manipulation", "Robotics" ], "rating_list": [ 8, 6, 6, 5, 6 ], "soundness_list": [ 4, 2, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 4, 3 ], "contribution_list": [ 4, 2, 3, 2, 3 ], "confidence_list": [ 5, 3, 4, 4, 5 ] }, { "paper_id": "r0opxuq8T8", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Jiaming Yang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Michal Derezinski", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" }, { "name": "Wei Hu", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" }, { "name": "Yang Yongyi", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" } ], "n_formula": 92, "n_formula_1": 11, "n_ref_uni": 32, "n_ref": 63, "n_ref_all": 77, "n_fig": 4, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 1070, "n_element_tab_1": 12, "formula_len_all": 9256, "formula_len_all_1": 644, "len_all": 207038, "len_all_1": 65858, "len_abs": 1228, "len_title": 139, "len_sents": 59546, "len_sents_1": 28975, "n_sents": 578, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1234, "title": "HERTA: A High-Efficiency and Rigorous Training Algorithm for Unfolded Graph Neural Networks", "abs": "As a variant of Graph Neural Networks (GNNs), Unfolded GNNs offer enhanced interpretability and flexibility over traditional designs. Nevertheless, they still suffer from scalability challenges when it comes to the training cost. Although many methods have been proposed to address the scalability issues, they mostly focus on per-iteration efficiency, without worst-case convergence guarantees. Moreover, those methods typically add components to or modify the original model, thus possibly breaking the interpretability of Unfolded GNNs. In this paper, we propose HERTA: a High-Efficiency and Rigorous Training Algorithm for Unfolded GNNs that accelerates the whole training process, achieving a nearly-linear time worst-case training guarantee. Crucially, HERTA converges to the optimum of the original model, thus preserving the interpretability of Unfolded GNNs. Additionally, as a byproduct of HERTA, we propose a new spectral sparsification method applicable to normalized and regularized graph Laplacians that ensures tighter bounds for our algorithm than existing spectral sparsifiers do. Experiments on real-world datasets verify the superiority of HERTA as well as its adaptability to various loss functions and optimizers.", "keywords": [ "Graph Neural Networks", "Spectral Sparsificaton", "Optimization", "Matrix Sketching" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 1, 2, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "r0QqfaCkF8", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Tianshu Yu", "gender": "Male", "institution": "Chinese University of Hong Kong (Shenzhen)", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaozhuang Song", "gender": "Male", "institution": "Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "YUZHAO TU", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "", "position": "Undergrad student" } ], "n_formula": 45, "n_formula_1": 15, "n_ref_uni": 63, "n_ref": 162, "n_ref_all": 172, "n_fig": 3, "n_tab": 6, "L_tab": 1735, "n_element_tab": 294, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 905, "n_element_tab_1": 158, "formula_len_all": 3280, "formula_len_all_1": 1108, "len_all": 247576, "len_all_1": 77113, "len_abs": 1293, "len_title": 114, "len_sents": 81339, "len_sents_1": 35318, "n_sents": 561, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1298, "title": "Fragment-Augmented Diffusion for Molecular Conformation Generation", "abs": "Molecular conformer generation is a fundamental challenge in computational chemistry, particularly for large and complex molecules. \nIn this work, we propose a novel approach called Fragment-Augmented Diffusion (FADiff), which integrates molecular fragmentations into diffusion models as a data augmentation strategy to enhance molecular conformation generation. By decomposing molecules into smaller, manageable fragments for the purpose of data augmentation, FADiff enhances the diffusion generation process, effectively capturing local structural variations while preserving the integrity of the entire molecule. Extensive experiments across multiple datasets demonstrate that FADiff consistently outperforms state-of-the-art methods, particularly in data-scarce scenarios, where the fragment-based augmentation approach significantly enhances model performance. We also provide a comprehensive analysis of different fragmentation rules and their impact on model performance, and theoretically validate FADiff's effectiveness in improving generalization. Overall, FADiff advances molecular conformation generation by enhancing the exploration of conformational space, offering a powerful tool for computational chemistry. The code is available at https://anonymous.4open.science/r/fragaug-5960/.", "keywords": [ "Molecular Conformation Generation", "Data Augmentation" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "r0JfDTXAWx", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Bin Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Hanzhu Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Jie Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Qitan Lv", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Xiaoqi Ni", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Yuhang Jiang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "Undergrad student" }, { "name": "Zhihao Shi", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 8, "n_ref_uni": 53, "n_ref": 112, "n_ref_all": 147, "n_fig": 6, "n_tab": 16, "L_tab": 9141, "n_element_tab": 832, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3940, "n_element_tab_1": 163, "formula_len_all": 904, "formula_len_all_1": 732, "len_all": 198171, "len_all_1": 61080, "len_abs": 1429, "len_title": 137, "len_sents": 57363, "len_sents_1": 26706, "n_sents": 408, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1434, "title": "Common Neighbor Induced Message Passing for Inductive Link Prediction in Knowledge Graphs", "abs": "Inductive link prediction is a significant challenge in knowledge graphs, focusing on predicting potential relations between unseen entities during training. A promising approach is to utilize Graph Neural Networks (GNNs) to extract entity-independent features from surrounding subgraphs. However, existing mainstream subgraph extraction methods may lead to the loss of key entities and relations, resulting in many disconnected reasoning paths that seriously hinder effective message passing. To address this challenge, we propose a novel framework called Common Neighbor Induced Message Passing (CNMP), designed to enhance message passing even when reasoning paths are disconnected. We observe that the common neighbors of two entities must share a reasoning path. Based on this insight, CNMP enhances message passing by updating the distance labels of isolated common neighbors, even if they are unreachable. This allows CNMP to incorporate new connected equivalent relations, facilitating effective message passing. Furthermore, we introduce a CNMP+ strategy that further improves the preservation of entities and relations during the message-passing process. CNMP+ involves maintaining a list of common neighbors at various distances and using a probing strategy to reconstruct complete reasoning paths. Experiments across multiple datasets demonstrate that our method significantly outperforms existing state-of-the-art methods.", "keywords": [ "Inductive Link Prediction", "Knowledge Graphs" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 1, 4 ] }, { "paper_id": "r01fcKhzT5", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Andrii Shkabrii", "gender": "Male", "institution": "University of Vienna", "country": "AT", "position": "MS student" }, { "name": "Claudia Plant", "gender": "unknown", "institution": "Universit\u00e4t Vienna", "country": "AT", "position": "Full Professor" }, { "name": "Collin Leiber", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Postdoc" }, { "name": "Kevin Sidak", "gender": "Male", "institution": "Universit\u00e4t Vienna", "country": "AT", "position": "Researcher" }, { "name": "Lukas Miklautz", "gender": "Male", "institution": "Universit\u00e4t Vienna", "country": "AT", "position": "Postdoc" }, { "name": "Sebastian Tschiatschek", "gender": "Male", "institution": "Universit\u00e4t Vienna", "country": "AT", "position": "Associate Professor" }, { "name": "Thomas Lang", "gender": "Male", "institution": "Universit\u00e4t Vienna", "country": "", "position": "PhD student" }, { "name": "Timo Klein", "gender": "Male", "institution": "Universit\u00e4t Vienna", "country": "AT", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 5, "n_ref_uni": 46, "n_ref": 122, "n_ref_all": 194, "n_fig": 17, "n_tab": 14, "L_tab": 3582, "n_element_tab": 540, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 487, "n_element_tab_1": 59, "formula_len_all": 1342, "formula_len_all_1": 252, "len_all": 246524, "len_all_1": 68831, "len_abs": 1075, "len_title": 115, "len_sents": 75385, "len_sents_1": 32399, "n_sents": 594, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1022, "title": "Breaking the Reclustering Barrier in Centroid-based Deep Clustering", "abs": "This work investigates an important phenomenon in centroid-based deep clustering(DC) algorithms: Performance quickly saturates after a period of rapid early gains. Practitioners commonly address early saturation with periodic reclustering, which we demonstrate to be insufficient to address performance plateaus. We call this phenomenon the \u201creclustering barrier\u201d and empirically show when the reclustering barrier occurs, what its underlying mechanisms are, and how it ispossible to Break the Reclustering Barrier with our algorithm BRB. BRB avoids early over-commitment to initial clusterings and enables continuous adaptation to reinitialized clustering targets while remaining conceptually simple. Applying our algorithm to widely-used centroid-based DC algorithms, we show that (1) BRB consistently improves performance across a wide range of clustering benchmarks, (2) BRB enables training from scratch, and (3) BRB performs competitively against state-of-the-art DC algorithms when combined with a contrastive loss.", "keywords": [ "deep clustering", "unsupervised learning", "representation learning" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 3, 2, 2, 4 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "qzZsz6MuEq", "primary_area": "applications to neuroscience & cognitive science", "n_author": 10, "author_info_list": [ { "name": "Ammar Belatreche", "gender": "Male", "institution": "Northumbria University", "country": "GB", "position": "Associate Professor" }, { "name": "Dehao Zhang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Enqi Zhang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Malu Zhang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Qian Sun", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Shuai Wang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yang Yang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yichen Xiao", "gender": "unknown", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Yimeng Shan", "gender": "Male", "institution": "Liaoning Technical University", "country": "CN", "position": "MS student" }, { "name": "Yu Liang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 30, "n_formula_1": 17, "n_ref_uni": 51, "n_ref": 94, "n_ref_all": 114, "n_fig": 8, "n_tab": 5, "L_tab": 2197, "n_element_tab": 287, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 531, "n_element_tab_1": 77, "formula_len_all": 1739, "formula_len_all_1": 916, "len_all": 175659, "len_all_1": 57876, "len_abs": 1463, "len_title": 98, "len_sents": 43537, "len_sents_1": 25896, "n_sents": 362, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1470, "title": "Spiking Vision Transformer with Saccadic Attention", "abs": "The combination of Spiking Neural Networks (SNNs) and Vision Transformers (ViTs) holds potential for achieving both energy efficiency and high performance, particularly suitable for edge vision applications. However, a significant performance gap still exists between SNN-based ViTs and their ANN counterparts. Here, we first analyze why SNN-based ViTs suffer from limited performance and identify a mismatch between the vanilla self-attention mechanism and spatio-temporal spike trains. This mismatch results in degraded spatial relevance and limited temporal interactions. To address these issues, we draw inspiration from biological saccadic attention mechanisms and introduce an innovative Saccadic Spike Self-Attention (SSSA) method. Specifically, in the spatial domain, SSSA employs a novel spike distribution-based method to effectively assess the relevance between Query and Key pairs in SNN-based ViTs. Temporally, SSSA employs a saccadic interaction module that dynamically focuses on selected visual areas at each timestep and significantly enhances whole scene understanding through temporal interactions.\nBuilding on the SSSA mechanism, we develop a SNN-based Vision Transformer (SNN-ViT). Extensive experiments across various visual tasks demonstrate that SNN-ViT achieves state-of-the-art performance with linear computational complexity. The effectiveness and efficiency of the SNN-ViT highlight its potential for power-critical edge vision applications.", "keywords": [ "Spiking Neural Networks", "Spiking Transformer", "Spike-driven Self-attention" ], "rating_list": [ 8, 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 4, 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 5, 4, 4 ] }, { "paper_id": "qykpnEWf2J", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Leong Hou U", "gender": "Male", "institution": "University of macau", "country": "MO", "position": "Associate Professor" }, { "name": "Panagiotis Karras", "gender": "Male", "institution": "Copenhagen University", "country": "DK", "position": "Full Professor" }, { "name": "Siyuan Wu", "gender": "unknown", "institution": "University of Macau", "country": "", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 17, "n_ref_uni": 50, "n_ref": 138, "n_ref_all": 198, "n_fig": 7, "n_tab": 10, "L_tab": 3060, "n_element_tab": 326, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 926, "n_element_tab_1": 153, "formula_len_all": 2289, "formula_len_all_1": 1452, "len_all": 173293, "len_all_1": 77418, "len_abs": 1091, "len_title": 109, "len_sents": 55923, "len_sents_1": 32865, "n_sents": 431, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1095, "title": "Balancing Bias in Two-sided Markets for Fair Stable Matchings", "abs": "The Balanced Stable Marriage (BSM) problem aims to find a stable matching in a two-sided market that minimizes the maximum dissatisfaction among two sides. The classical Deferred Acceptance algorithm merely produces an unfair stable marriage, providing optimal partners for one side while partially assigning pessimal partners to the other. Solving BSM is NP-hard, thwarting attempts to resolve the problem exactly. As the instance size increases in practice, recent studies have explored heuristics for finding a fair stable marriage but have not found an exact optimal solution for BSM efficiently. Nevertheless, in this paper we propose an efficient algorithm, Isorropia, that returns the exact optimal solution to practical BSM problem instances. Isorropia constructs two sets of candidate rotations from which it builds three sets of promising antichains, and performs local search on those three sets of promising antichains. Our extensive experimental study shows that Isorropia surpasses the time-efficiency of baselines that return the exact solution by up to three orders of magnitude.", "keywords": [ "stable marriage; fairness" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 2, 2, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "qyU5s4fzLg", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Danushka Bollegala", "gender": "Male", "institution": "Amazon", "country": "GB", "position": "Amazon Scholar" }, { "name": "Junjie Chen", "gender": "Male", "institution": "the University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Xiangheng He", "gender": "Female", "institution": "Imperial College London", "country": "", "position": "PhD student" }, { "name": "Yusuke Miyao", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 13, "n_ref_uni": 36, "n_ref": 77, "n_ref_all": 123, "n_fig": 11, "n_tab": 4, "L_tab": 330, "n_element_tab": 62, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1055, "n_element_tab_1": 119, "formula_len_all": 1210, "formula_len_all_1": 935, "len_all": 148148, "len_all_1": 68791, "len_abs": 1300, "len_title": 127, "len_sents": 43351, "len_sents_1": 31017, "n_sents": 337, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 79, "L_abs": 1309, "title": "Improving Unsupervised Constituency Parsing via Maximizing Semantic Information", "abs": "Unsupervised constituency parsers organize phrases within a sentence into a tree-shaped syntactic constituent structure that reflects the organization of sentence semantics. \nHowever, the traditional objective of maximizing sentence log-likelihood (LL) does not explicitly account for the close relationship between the constituent structure and the semantics, resulting in a weak correlation between LL values and parsing accuracy.\nIn this paper, we introduce a novel objective for training unsupervised parsers: maximizing the information between constituent structures and sentence semantics (SemInfo). \nWe introduce a bag-of-substrings model to represent the semantics and apply the probability-weighted information metric to estimate the SemInfo.\nAdditionally, we develop a Tree Conditional Random Field (TreeCRF)-based model to apply the SemInfo maximization objective to Probabilistic Context-Free Grammar (PCFG) induction, the state-of-the-art non-ensemble method for unsupervised constituency parsing. \nExperiments demonstrate that SemInfo correlates more strongly with parsing accuracy than LL.\nOur algorithm significantly enhances parsing accuracy by an average of 7.85 points across five PCFG variants and in four languages, achieving state-of-the-art level results in three of the four languages.", "keywords": [ "unsupervised constituency parsing", "information theory", "semantic information" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 4, 3, 3, 4 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "qxzOEy9fLU", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Francesco Tonolini", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Gabriella Kazai", "gender": "Female", "institution": "Amazon", "country": "", "position": "Principal Researcher" }, { "name": "Matej Cief", "gender": "Male", "institution": "Brno University of Technology", "country": "CZ", "position": "PhD student" }, { "name": "Nikolaos Aletras", "gender": "unknown", "institution": "University of Sheffield", "country": "", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 13, "n_ref_uni": 31, "n_ref": 98, "n_ref_all": 116, "n_fig": 6, "n_tab": 2, "L_tab": 3880, "n_element_tab": 34, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1129, "n_element_tab_1": 20, "formula_len_all": 1592, "formula_len_all_1": 891, "len_all": 148003, "len_all_1": 64601, "len_abs": 1186, "len_title": 69, "len_sents": 40369, "len_sents_1": 28411, "n_sents": 325, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1199, "title": "Adaptive Uncertainty-Aware Reinforcement Learning from Human Feedback", "abs": "Reinforcement learning from human feedback (RLHF) is a popular technique to align large language models (LLMs) to human preferences. It requires learning a reward model that predicts scalar values given a generated text sequence, acting as a proxy for human preference scores. A central problem of RLHF is \\textit{reward hacking}, i.e., overoptimization. LLMs can easily exploit the reward model by generating text that can receive high scores but no longer align with human preferences. We address this problem by proposing a new objective which adapts the tradeoff between reward model score and regularisation based on reward uncertainty. We hypothesize that when the reward model uncertainty is low, RLHF should make a larger step size by lowering the regularization coefficient. On the other hand, when the uncertainty is high, optimization should slow down by staying closer to the original model. We present a novel re-formulation of the RLHF objective and derive our approach from its generalization to account for reward model variance. We demonstrate that our uncertainty-aware RLHF objective mitigates overoptimization and outperforms vanilla RLHF by 50% on a standard summarization task.", "keywords": [ "RLHF", "PPO", "LLM", "reinforcement learning", "alignment" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "qxobgbamw9", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Haotian Huang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Yifei Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yisen Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Du Tianqi", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" } ], "n_formula": 45, "n_formula_1": 7, "n_ref_uni": 31, "n_ref": 112, "n_ref_all": 147, "n_fig": 9, "n_tab": 5, "L_tab": 1481, "n_element_tab": 159, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2428, "n_element_tab_1": 127, "formula_len_all": 2857, "formula_len_all_1": 363, "len_all": 155243, "len_all_1": 71674, "len_abs": 1190, "len_title": 110, "len_sents": 49116, "len_sents_1": 33215, "n_sents": 390, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1201, "title": "Output Alignment: A Top-down Approach to Length Generalization", "abs": "Recently, large language models have exhibited impressive performance and surprising emergent properties. However, their abilities remain constrained by the preset context window of the Transformer architecture, and they continue to struggle with length generalization. In this work, we propose a new perspective on length generalization by focusing on the output distribution rather than the input, as most prior studies have done (e.g., through positional encodings or data structure). First, through case studies on simple synthetic tasks, we highlight the importance of **output alignment**---the consistency of output distributions across sequences of varying lengths. We then extend this observation to natural language tasks and introduce a metric named Long-Short Misalignment to quantify output alignment, finding a strong correlation between this metric and length generalization performance. Based on these insights, we propose a regularization loss during training that improves output alignment. Extensive experiments confirm the effectiveness of this approach. Overall, our work provides a novel perspective for understanding and enhancing length generalization in large language models.", "keywords": [ "Length Generalization", "Output Alignment" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "qxRoo7ULCo", "primary_area": "generative models", "n_author": 11, "author_info_list": [ { "name": "Achuta Kadambi", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "Bangbang Yang", "gender": "Male", "institution": "ByteDance Inc", "country": "CN", "position": "Researcher" }, { "name": "Dejia Xu", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Panwang Pan", "gender": "Male", "institution": "ByteDance", "country": "CN", "position": "Researcher" }, { "name": "Renjie Li", "gender": "unknown", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Shijie Zhou", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Xuanyang Zhang", "gender": "Male", "institution": "stepfun", "country": "CN", "position": "Researcher" }, { "name": "Zeming Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zhangyang Wang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Zhengzhong Tu", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" }, { "name": "Zhiwen Fan", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 52, "n_ref": 102, "n_ref_all": 124, "n_fig": 11, "n_tab": 3, "L_tab": 378, "n_element_tab": 41, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 581, "n_element_tab_1": 7, "formula_len_all": 1077, "formula_len_all_1": 807, "len_all": 175944, "len_all_1": 57620, "len_abs": 352, "len_title": 97, "len_sents": 44246, "len_sents_1": 26336, "n_sents": 332, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 49, "L_abs": 1704, "title": "4K4DGen: Panoramic 4D Generation at 4K Resolution", "abs": "The blooming of virtual reality and augmented reality (VR/AR) technologies has driven an increasing demand for the creation of high-quality, immersive, and dynamic environments. However, existing generative techniques either focus solely on dynamic objects or perform outpainting from a single perspective image, failing to meet the requirements of VR/AR applications that need free-viewpoint, 360$^{\\circ}$ virtual views where users can move in all directions. In this work, we tackle the challenging task of elevating a single panorama to an immersive 4D experience. For the first time, we demonstrate the capability to generate omnidirectional dynamic scenes with 360$^{\\circ}$ views at 4K (4096 $\\times$ 2048) resolution, thereby providing an immersive user experience. Our method introduces a pipeline that facilitates natural scene animations and optimizes a set of dynamic Gaussians using efficient splatting techniques for real-time exploration. To overcome the lack of scene-scale annotated 4D data and models, especially in panoramic formats, we propose a novel \\textbf{Panoramic Denoiser} that adapts generic 2D diffusion priors to animate consistently in 360$^{\\circ}$ images, transforming them into panoramic videos with dynamic scenes at targeted regions. Subsequently, we propose \\textbf{Dynamic Panoramic Lifting} to elevate the panoramic video into a 4D immersive environment while preserving spatial and temporal consistency. By transferring prior knowledge from 2D models in the perspective domain to the panoramic domain and the 4D lifting with spatial appearance and geometry regularization, we achieve high-quality Panorama-to-4D generation at a resolution of 4K for the first time.", "keywords": [ "4D Generation", "Panoramic Video", "Panoramic Gaussian Splatting" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "qx07JhIs8E", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Cheng-Chang Tsai", "gender": "unknown", "institution": "Academia Sinica", "country": "TW", "position": "Research Assistant" }, { "name": "Cheng-Han Yeh", "gender": "Male", "institution": "Academia Sinica", "country": "TW", "position": "Researcher" }, { "name": "Chun-Shien Lu", "gender": "Male", "institution": "Academia Sinica", "country": "TW", "position": "Full Professor" }, { "name": "Jonathan Weiping Li", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Kuanchun Yu", "gender": "unknown", "institution": "Academia Sinica", "country": "TW", "position": "Research Assistant" }, { "name": "Ren-Wei Liang", "gender": "Male", "institution": "Academia Sinica", "country": "TW", "position": "Intern" }, { "name": "Shang-Tse Chen", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 30, "n_ref": 71, "n_ref_all": 88, "n_fig": 4, "n_tab": 7, "L_tab": 1473, "n_element_tab": 120, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 461, "n_element_tab_1": 61, "formula_len_all": 368, "formula_len_all_1": 277, "len_all": 115682, "len_all_1": 54004, "len_abs": 1593, "len_title": 111, "len_sents": 34898, "len_sents_1": 25615, "n_sents": 236, "n_sents_1": 166, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1600, "title": "Adversarial Robustness Overestimation and Instability in TRADES", "abs": "This paper examines the phenomenon of probabilistic robustness overestimation in TRADES, a prominent adversarial training method. Our study reveals that TRADES sometimes yields disproportionately high PGD validation accuracy compared to the AutoAttack testing accuracy in the multiclass classification task. This discrepancy highlights a significant overestimation of robustness for these instances, potentially linked to gradient masking. We further analyze the parameters contributing to unstable models that lead to overestimation. Our findings indicate that smaller batch sizes, lower beta values (which control the weight of the robust loss term in TRADES), larger learning rates, and higher class complexity (e.g., CIFAR-100 versus CIFAR-10) are associated with an increased likelihood of robustness overestimation. By examining metrics such as the First-Order Stationary Condition (FOSC), inner-maximization, and gradient information, we identify the underlying cause of this phenomenon as gradient masking and provide insights into it. Furthermore, our experiments show that certain unstable training instances may return to a state without robust overestimation, inspiring our attempts at a solution. In addition to adjusting parameter settings to reduce instability or retraining when overestimation occurs, we recommend incorporating Gaussian noise in inputs when the FOSC score exceed the threshold. This method aims to mitigate robustness overestimation of TRADES and other similar methods at its source, ensuring more reliable representation of adversarial robustness during evaluation.", "keywords": [ "Adversarial training", "Robustness", "Obfuscated gradients", "TRADES" ], "rating_list": [ 3, 1, 3 ], "soundness_list": [ 3, 1, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "qvwcK4Uz8z", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Avishek Anand", "gender": "Male", "institution": "Delft University of Technology", "country": "", "position": "Associate Professor" }, { "name": "Kiran Purohit", "gender": "Female", "institution": "Indian Institute of Technology Kharagpur", "country": "IN", "position": "PhD student" }, { "name": "Sourangshu Bhattacharya", "gender": "Male", "institution": "Indian Institute of Technology Kharagpur", "country": "", "position": "Associate Professor" }, { "name": "Venktesh V", "gender": "unknown", "institution": "Delft University of Technology", "country": "", "position": "Postdoc" } ], "n_formula": 30, "n_formula_1": 14, "n_ref_uni": 29, "n_ref": 87, "n_ref_all": 118, "n_fig": 13, "n_tab": 7, "L_tab": 3860, "n_element_tab": 429, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4964, "formula_len_all_1": 1270, "len_all": 185572, "len_all_1": 70590, "len_abs": 1516, "len_title": 112, "len_sents": 50659, "len_sents_1": 32235, "n_sents": 422, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1586, "title": "CASE: Challenger Arm Sampling for Efficient In-Context Reasoning", "abs": "The in-context learning paradigm with LLMs has been instrumental in advancing applications that require complex reasoning over natural language. An optimal selection of few-shot examples (exemplars) is essential for constructing effective prompts under a limited budget.\nIn this paper, we frame the problem of exemplar selection for In-Context Reasoning (ICR) as a top-m best arms identification problem. A key challenge in this context is the exponentially large number of arms that need to be evaluated to identify the m-best arms. We propose CASE (Challenger Arm Sampling for Exemplar selection), a novel selective exploration strategy that maintains a shortlist of ``challenger'' arms, which are current candidates for the top-m arms. In each iteration, only the arms from this shortlist and the current top-m set are pulled, thereby reducing sample complexity and, consequently, the number of LLM evaluations. Furthermore, we model the scores of exemplar subsets (arms) using a parameterized linear scoring function, leading to a stochastic linear bandits setting. In this setting, CASE identifies the top-m arms with significantly fewer evaluations than existing state-of-the-art methods. CASE effectively works with black box LLMs and selects a static set of few-shot examples, resulting in an extremely efficient scheme for in-context reasoning. The exemplars selected with CASE show surprising performance gains of up to 15.19% compared to state-of-the-art exemplar selection methods. We release our code and data (https://anonymous.4open.science/r/CASE_exemplar_bandits-7403).", "keywords": [ "In-Context Learning", "Large Language Models", "Exemplar Selection", "Stochastic Linear Bandits", "Challenger Arms" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "qvuC22BT2q", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Hai Ci", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Jiaming Liu", "gender": "Male", "institution": "Tiamat AI", "country": "CN", "position": "Researcher" }, { "name": "Pei Yang", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Shengtao Lou", "gender": "Male", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "Undergrad student" }, { "name": "Xiaokang Liu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yiren Song", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zheng Shou", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 19, "n_ref": 57, "n_ref_all": 66, "n_fig": 5, "n_tab": 3, "L_tab": 413, "n_element_tab": 23, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 1250, "n_element_tab_1": 19, "formula_len_all": 709, "formula_len_all_1": 193, "len_all": 105564, "len_all_1": 52318, "len_abs": 668, "len_title": 181, "len_sents": 31113, "len_sents_1": 26428, "n_sents": 202, "n_sents_1": 144, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1160, "title": "Anti-Reference: Universal and Immediate Defense Against Reference-Based Generation", "abs": "Diffusion models have completely transformed the field of generative models, demonstrating unparalleled capabilities in generating high-fidelity images. However, when misused, such a powerful and convenient tool could create fake news or disturbing content targeted at individual victims, causing severe negative social impacts. In this paper, we introduce Anti-Reference, a novel method that protects images from the threats posed by reference-based generation techniques by adding imperceptible adversarial noise to the images. We propose a unified loss function that enables joint attacks on fine-tuning-based customization methods, non-fine-tuning customization methods, and human-centric driving methods. Based on this loss, we train a Noise Encoder with a DiT architecture to predict the noise or directly optimize the noise using the PGD (Projected Gradient Descent) method. Our method demonstrates strong black-box transferability, being equally effective against black-box models and some commercial APIs such as Animate Anyone, and EMO. Extensive experiments validate the performance of Anti-Reference, establishing a new benchmark in image security.", "keywords": [ "Diffusion model", "Customized generation", "Adversarial Attack", "Image generation" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "qucdAPZfar", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Felix Dangel", "gender": "Male", "institution": "Vector Institute, Toronto", "country": "CA", "position": "Postdoc" }, { "name": "Marvin F. da Silva", "gender": "Male", "institution": "Dalhousie University", "country": "CA", "position": "PhD student" }, { "name": "Sageev Oore", "gender": "Male", "institution": "Dalhousie University", "country": "CA", "position": "Associate Professor" } ], "n_formula": 84, "n_formula_1": 32, "n_ref_uni": 22, "n_ref": 68, "n_ref_all": 85, "n_fig": 11, "n_tab": 3, "L_tab": 352, "n_element_tab": 51, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5883, "formula_len_all_1": 2458, "len_all": 163483, "len_all_1": 63994, "len_abs": 1246, "len_title": 132, "len_sents": 53939, "len_sents_1": 28107, "n_sents": 452, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1256, "title": "Hide & Seek: Transformer Symmetries Obscure Sharpness & Riemmanian Geometry Finds It", "abs": "The concept of sharpness has been successfully applied to traditional architectures like MLPs and CNNs to predict their generalization.\n For transformers, however, recent work reported weak correlation between flatness and generalization. We argue that existing sharpness measures fail for transformers, because they have much richer symmetries in their attention mechanism that induce directions in parameter space along which the network or its loss remain identical.\n We posit that sharpness must account fully for these symmetries, and thus we redefine it on a quotient manifold that results from quotienting out the transformer symmetries, thereby removing their ambiguities.\n Leveraging tools from Riemannian geometry, we propose a fully general notion of sharpness, in terms of a geodesic ball on the symmetry-corrected quotient manifold. In practise, we need to resort to approximating the geodesics. Doing so up to first order yields existing adaptive sharpness measures, and we demonstrate that including higher-order terms is crucial to recover correlation with generalization.\n We present results on diagonal networks with synthetic data, and show that our geodesic sharpness reveals the correlation for real-world transformers on ImageNet.", "keywords": [ "generalization", "symmetry", "sharpness", "flatness", "riemannian geometry", "loss landscape" ], "rating_list": [ 6, 3, 8, 6, 6 ], "soundness_list": [ 4, 2, 3, 3, 4 ], "presentation_list": [ 2, 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3, 3 ] }, { "paper_id": "qu6UMVT4k1", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Jiafeng Guo", "gender": "Male", "institution": "Institute of Computing Technolgy, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Liang Pang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "Associate Professor" }, { "name": "Wanqing Cui", "gender": "Female", "institution": "University of Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Xin Hong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Xueqi Cheng", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy", "country": "CN", "position": "Full Professor" }, { "name": "Yanyan Lan", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 82, "n_ref_all": 120, "n_fig": 10, "n_tab": 11, "L_tab": 2789, "n_element_tab": 279, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 1990, "n_element_tab_1": 245, "formula_len_all": 298, "formula_len_all_1": 57, "len_all": 206402, "len_all_1": 72202, "len_abs": 1574, "len_title": 77, "len_sents": 57136, "len_sents_1": 31365, "n_sents": 517, "n_sents_1": 274, "venue": "ICLR2025", "decision": "Reject", "L_title": 29, "L_abs": 1549, "title": "Visual Transformation Telling", "abs": "Humans can naturally reason from superficial state differences (e.g. ground wetness) to transformations descriptions (e.g. raining) according to their life experience. In this paper, we propose a new visual reasoning task to test this transformation reasoning ability in real-world scenarios, called **V**sual **T**ransformation **T**elling (VTT). Given a series of states (i.e., images), VTT requires to describe the transformation occurring between every two adjacent states. Different from existing visual reasoning tasks that focus on surface state reasoning, the advantage of VTT is that it captures the underlying causes, e.g. actions or events, behind the differences among states. We collect a novel dataset which comprise 13,547 samples to support the study of transformation reasoning. Each sample involves several key state images along with their transformation descriptions. Our dataset spans diverse real-world activities, providing a rich resource for training and evaluation with automated, human, and LLM assessments. To construct an initial benchmark for VTT, we test models including traditional visual storytelling (CST, GLACNet) or dense video captioning methods (Densecap) and advanced multimodal large language models (LLaVA v1.5-7B, Qwen-VL-chat, Gemini-1.5, GPT-4o, and GPT-4), as well as their upgraded versions based on our learning on human reasoning. Experimental results reveal that even state-of-the-art models still have a significant gap with human performance in VTT, highlighting substantial areas for improvement.", "keywords": [ "visual reasoning", "transformation", "captioning" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 3, 1 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "qtqvuBmhxU", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Huimin Lu", "gender": "Not Specified", "institution": "Southeast University", "country": "CN", "position": "Full Professor" }, { "name": "Lie Ju", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Peibo Duan", "gender": "Male", "institution": "Monash University", "country": "CN", "position": "Lecturer" }, { "name": "Siyuan Yan", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Xiaodan Xing", "gender": "Female", "institution": "Imperial College London", "country": "", "position": "PhD student" }, { "name": "Yang Nan", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "PhD student" }, { "name": "Yukun Zhou", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Postdoc" }, { "name": "Zongyuan Ge", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 47, "n_ref": 91, "n_ref_all": 97, "n_fig": 3, "n_tab": 8, "L_tab": 7966, "n_element_tab": 1014, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 7669, "n_element_tab_1": 643, "formula_len_all": 17, "formula_len_all_1": 17, "len_all": 159242, "len_all_1": 70539, "len_abs": 379, "len_title": 113, "len_sents": 35528, "len_sents_1": 28422, "n_sents": 227, "n_sents_1": 153, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1347, "title": "MONICA: Benchmarking on Long-tailed Medical Image Classification", "abs": "Long-tailed learning is considered to be an extremely challenging problem in data imbalance learning. It aims to train well-generalized models from a large number of images that follow a long-tailed class distribution. In the medical field, many diagnostic imaging exams such as dermoscopy and chest radiography yield a long-tailed distribution of complex clinical findings. Recently, long-tailed learning in medical image analysis has garnered significant attention. However, the field currently lacks a unified, strictly formulated, and comprehensive benchmark, which often leads to unfair comparisons and inconclusive results. To help the community improve the evaluation and advance, we build a unified, well-structured codebase called Medical OpeN-source Long-taIled ClassifiCAtion (MONICA), which implements over 30 methods developed in relevant fields and evaluated on 12 long-tailed medical datasets covering 6 medical domains. Our work provides valuable practical guidance and insights for the field, offering detailed analysis and discussion on the effectiveness of individual components within the inbuilt state-of-the-art methodologies. We hope this codebase serves as a comprehensive and reproducible benchmark, encouraging further advancements in long-tailed medical image learning. The codebase will be publicly available on GitHub.", "keywords": [ "Long-tailed Learning", "Benchmark", "Medical Image Classification" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 4, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "qto91DryES", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "An Nguyen The", "gender": "Male", "institution": "FPT Software AI Center", "country": "VN", "position": "Researcher" }, { "name": "Duy-Tung Pham", "gender": "Male", "institution": "AI Center, FPT Software", "country": "VN", "position": "Researcher" }, { "name": "Hoang V. Tran", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Minh-Khoi Nguyen-Nhat", "gender": "Male", "institution": "FPT", "country": "VN", "position": "Researcher" }, { "name": "Tan Minh Nguyen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Thanh Tran", "gender": "Male", "institution": "VinUniversity", "country": "", "position": "Undergrad student" }, { "name": "Thieu Vo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Tho Tran Huu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 266, "n_formula_1": 23, "n_ref_uni": 52, "n_ref": 121, "n_ref_all": 185, "n_fig": 0, "n_tab": 17, "L_tab": 2715, "n_element_tab": 173, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 314, "n_element_tab_1": 25, "formula_len_all": 36193, "formula_len_all_1": 1893, "len_all": 309443, "len_all_1": 66299, "len_abs": 1930, "len_title": 90, "len_sents": 68348, "len_sents_1": 28893, "n_sents": 826, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1819, "title": "Equivariant Polynomial Functional Networks", "abs": "Neural Functional Networks (NFNs) have gained increasing interest due to their wide range of applications, including extracting information from implicit representations of data, editing network weights, and evaluating policies. A key design principle of NFNs is their adherence to the permutation and scaling symmetries inherent in the connectionist structure of the input neural networks. Recent NFNs have been proposed with permutation and scaling equivariance based on either graph-based message-passing mechanisms or parameter-sharing mechanisms. However, graph-based equivariant NFNs suffer from high memory consumption and long running times. On the other hand, parameter-sharing-based NFNs built upon equivariant linear layers exhibit lower memory consumption and faster running time, yet their expressivity is limited due to the large size of the symmetric group of the input neural networks. The challenge of designing a permutation and scaling equivariant NFN that maintains low memory consumption and running time while preserving expressivity remains unresolved. In this paper, we propose a novel solution with the development of MAGEP-NFN (**M**onomial m**A**trix **G**roup **E**quivariant **P**olynomial **NFN**). Our approach follows the parameter-sharing mechanism but differs from previous works by constructing a nonlinear equivariant layer represented as a polynomial in the input weights. This polynomial formulation enables us to incorporate additional relationships between weights from different input hidden layers, enhancing the model's expressivity while keeping memory consumption and running time low, thereby addressing the aforementioned challenge. We provide empirical evidence demonstrating that MAGEP-NFN achieves competitive performance and efficiency compared to existing baselines.", "keywords": [ "neural functional network", "equivariant model", "polynomial layer", "monomial matrix group" ], "rating_list": [ 3, 5, 8, 6 ], "soundness_list": [ 2, 2, 4, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "qtWjSboqfe", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 13, "author_info_list": [ { "name": "Binyuan Hui", "gender": "Female", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Hamid Rokny", "gender": "Male", "institution": "UNSW Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Lei Zhang", "gender": "Male", "institution": "Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Longze Chen", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Min Yang", "gender": "Female", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Run Luo", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Wanwei He", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Xiaobo Xia", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Yunshui Li", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Ziqiang Liu", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Tony Lin", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Song Zikai", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 72, "n_ref": 201, "n_ref_all": 238, "n_fig": 11, "n_tab": 8, "L_tab": 2178, "n_element_tab": 306, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1483, "n_element_tab_1": 175, "formula_len_all": 586, "formula_len_all_1": 631, "len_all": 222715, "len_all_1": 67898, "len_abs": 2547, "len_title": 135, "len_sents": 55746, "len_sents_1": 29672, "n_sents": 390, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 86, "L_abs": 2092, "title": "DEEM: Diffusion models serve as the eyes of large language models for image perception", "abs": "The development of large language models (LLMs) has significantly advanced the emergence of large multimodal models (LMMs). While LMMs have achieved tremendous success by promoting the synergy between multimodal comprehension and creation, they often face challenges when confronted with out-of-distribution data, such as which can hardly distinguish orientation, quantity, color, structure, etc. This is primarily due to their reliance on image encoders trained to encode images into task-relevant features, which may lead them to disregard irrelevant details. Delving into the modeling capabilities of diffusion models for images naturally prompts the question: Can diffusion models serve as the eyes of large language models for image perception? In this paper, we propose DEEM, a simple but effective approach that utilizes the generative feedback of diffusion models to align the semantic distributions of the image encoder. This addresses the drawbacks of previous methods that solely relied on image encoders like CLIP-ViT, thereby enhancing the model's resilience against out-of-distribution samples and reducing visual hallucinations. Importantly, this is achieved without requiring additional training modules and with fewer training parameters. We extensively evaluated DEEM on both our newly constructed RobustVQA benchmark and other well-known benchmarks, POPE and MMVP, for visual hallucination and perception. In particular, DEEM improves LMM's visual perception performance to a large extent (e.g., 4\\% \u2191 on RobustVQA, 6.5\\% \u2191 on MMVP and 12.8 \\% \u2191 on POPE ). Compared to the state-of-the-art interleaved content generation models, DEEM exhibits enhanced robustness and a superior capacity to alleviate model hallucinations while utilizing fewer trainable parameters, less pre-training data (10\\%), and a smaller base model size. Extensive experiments demonstrate that DEEM enhances the performance of LMMs on various downstream tasks without inferior performance in the long term, including visual question answering, image captioning, and text-conditioned image synthesis.", "keywords": [ "MLLM; Diffusion Model;" ], "rating_list": [ 6, 8, 8, 8, 6 ], "soundness_list": [ 2, 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 4, 3, 3 ], "confidence_list": [ 3, 5, 3, 3, 4 ] }, { "paper_id": "qtTIP5Gjc5", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Duy-Tung Pham", "gender": "Male", "institution": "AI Center, FPT Software", "country": "VN", "position": "Researcher" }, { "name": "Tan Minh Nguyen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Thieu Vo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Xin T. Tong", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" } ], "n_formula": 60, "n_formula_1": 19, "n_ref_uni": 47, "n_ref": 93, "n_ref_all": 130, "n_fig": 7, "n_tab": 7, "L_tab": 1248, "n_element_tab": 84, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 470, "n_element_tab_1": 46, "formula_len_all": 5102, "formula_len_all_1": 1253, "len_all": 192150, "len_all_1": 68987, "len_abs": 1477, "len_title": 116, "len_sents": 55233, "len_sents_1": 28887, "n_sents": 561, "n_sents_1": 279, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 68, "L_abs": 1488, "title": "Demystifying the Token Dynamics of Deep Selective State Space Models", "abs": "Selective state space models (SSM), such as Mamba, have gained prominence for their effectiveness in modeling sequential data. Despite their outstanding empirical performance, a comprehensive theoretical understanding of deep selective SSM remains elusive, hindering their further development and adoption for applications that need high fidelity. In this paper, we investigate the dynamical properties of tokens in a pre-trained Mamba model. In particular, we derive the dynamical system governing the continuous-time limit of the Mamba model and characterize the asymptotic behavior of its solutions. In the one-dimensional case, we prove that only one of the following two scenarios happens: either all tokens converge to zero, or all tokens diverge to infinity. We provide criteria based on model parameters to determine when each scenario occurs. For the convergent scenario, we empirically verify that this scenario negatively impacts the model's performance. For the divergent scenario, we prove that different tokens will diverge to infinity at different rates, thereby contributing unequally to the updates during model training. Based on these investigations, we propose two refinements for the model: excluding the convergent scenario and reordering tokens based on their importance scores, both aimed at improving practical performance. Our experimental results validate these refinements, offering insights into enhancing Mamba's effectiveness in real-world applications.", "keywords": [ "Selective state-space model", "continuous-time limit", "dynamical system", "asymptotic behavior", "token reordering" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "qssVptHTPN", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Ian Connick Covert", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "James Zou", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Tatsunori Hashimoto", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Tony Sun", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 102, "n_ref": 283, "n_ref_all": 330, "n_fig": 13, "n_tab": 12, "L_tab": 4868, "n_element_tab": 457, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 2767, "n_element_tab_1": 152, "formula_len_all": 205, "formula_len_all_1": 51, "len_all": 304997, "len_all_1": 63989, "len_abs": 1690, "len_title": 98, "len_sents": 73206, "len_sents_1": 30716, "n_sents": 443, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1694, "title": "Locality Alignment Improves Vision-Language Models", "abs": "Vision language models (VLMs) have seen growing adoption in recent years, but many still struggle with basic spatial reasoning. We hypothesize that this is due to VLMs adopting pre-trained vision backbones, specifically vision transformers (ViTs) trained with image-level supervision and minimal inductive biases. Such models may fail to encode the class contents at each position in the image, and our goal is to resolve this by ensuring that the vision backbone effectively captures both local and global image semantics. Our main insight is that we do not require new supervision to learn this capability -- pre-trained models contain significant knowledge of local semantics that we can extract and use for scalable self-supervision. We propose a new efficient post-training stage for ViTs called locality alignment, and a specific fine-tuning procedure called MaskEmbed that uses a masked reconstruction loss to learn semantic contributions for each image patch. We first evaluate locality alignment with a vision-only benchmark, and we find that it improves a model's performance at a patch-level semantic segmentation task, especially for strong backbones trained with image-caption pairs (e.g., CLIP and SigLIP). We then train a series of VLMs with and without locality alignment, and we find that locality-aligned backbones improve performance across a range of benchmarks, particularly ones that involve spatial understanding (e.g., RefCOCO, OCID-Ref, TallyQA, VSR, AI2D). Overall, we show that we can efficiently learn local semantic extraction via a locality alignment stage, and that this procedure complements existing VLM training recipes that use off-the-shelf vision backbones.", "keywords": [ "Multimodal language models", "vision-language models", "locality", "alignment", "vision transformers" ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 4, 3, 2, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "qrTrnrEi9d", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Alan Ritter", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Vedaant Shah", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Yang Chen", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 67, "n_ref": 107, "n_ref_all": 125, "n_fig": 7, "n_tab": 15, "L_tab": 10276, "n_element_tab": 1294, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1512, "n_element_tab_1": 73, "formula_len_all": 286, "formula_len_all_1": 286, "len_all": 203312, "len_all_1": 54399, "len_abs": 1236, "len_title": 126, "len_sents": 36913, "len_sents_1": 25352, "n_sents": 235, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1239, "title": "Translation and Fusion Improves Zero-shot Cross-lingual Information Extraction", "abs": "Large language models (LLMs) combined with instruction tuning have shown significant progress in information extraction (IE) tasks, exhibiting strong generalization capabilities to unseen datasets by following annotation guidelines. However, their applicability to low-resource languages remains limited due to lack of both labeled data for fine-tuning, and unlabeled text for pre-training. In this paper, we propose TransFusion, a framework in which models are fine-tuned to use English translations of low-resource language data, enabling more precise predictions through annotation fusion. Based on TransFusion, we introduce GoLLIE-TF, a cross-lingual instruction-tuned LLM for IE tasks, designed to close the performance gap between high and low-resource languages. Our experiments across twelve multilingual IE datasets spanning 50 languages demonstrate that GoLLIE-TF achieves better cross-lingual transfer over the base model. In addition, we show that TransFusion significantly improves low-resource language named entity recognition when applied to proprietary models such as GPT-4 (+5 F1) with a prompting approach, or fine-tuning different language models including decoder-only (+14 F1) and encoder-only (+13 F1) architectures.", "keywords": [ "large language model", "multilingual", "information extraction", "low-resource language" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "qrTOtUdz4Z", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Chuyu Zhang", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "PhD student" }, { "name": "Peiyan Gu", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" }, { "name": "Xueyang Yu", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "MS student" }, { "name": "Xuming He", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 15, "n_formula_1": 12, "n_ref_uni": 34, "n_ref": 113, "n_ref_all": 125, "n_fig": 5, "n_tab": 9, "L_tab": 3187, "n_element_tab": 497, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 3001, "n_element_tab_1": 371, "formula_len_all": 817, "formula_len_all_1": 541, "len_all": 152080, "len_all_1": 74221, "len_abs": 1337, "len_title": 132, "len_sents": 47597, "len_sents_1": 31740, "n_sents": 373, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1342, "title": "Composing Novel Classes: A Concept-Driven Approach to Generalized Category Discovery", "abs": "We tackle the generalized category discovery (GCD) problem, which aims to discover novel classes in unlabeled datasets by leveraging the knowledge of known classes. Previous works utilize the known class knowledge through shared representation spaces. Despite their progress, our analysis experiments show that impressive novel class clustering results are achieved in the feature space of a known class pre-trained model, suggesting that existing methods may not fully utilize known class knowledge. To address it, we introduce a novel concept learning framework for GCD, named ConceptGCD, that categorizes concepts into two types: derivable and underivable from known class concepts, and adopts a stage-wise learning strategy to learn them separately. Specifically, our framework first extracts known class concepts by a known class pre-trained model and then produces derivable concepts from them by a generator layer with a covariance-augmented loss. Subsequently, we expand the generator layer to learn underivable concepts in a balanced manner ensured by a concept score normalization strategy and integrate a contrastive loss to preserve previously learned concepts. Extensive experiments on various benchmark datasets demonstrate the superiority of our approach over the previous state-of-the-art methods. Code will be available soon.", "keywords": [ "generalized category discovery; self-supervised learning; knowledge transfer; knowledge distillation; unsupervised learning;" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "qqZijHRcA5", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Antti Honkela", "gender": "Male", "institution": "University of Helsinki", "country": "FI", "position": "Full Professor" }, { "name": "Gauri Pradhan", "gender": "Female", "institution": "University of Helsinki", "country": "", "position": "PhD student" }, { "name": "Hibiki Ito", "gender": "Male", "institution": "University of Helsinki", "country": "FI", "position": "MS student" }, { "name": "Joonas J\u00e4lk\u00f6", "gender": "Male", "institution": "University of Helsinki", "country": "", "position": "Postdoc" }, { "name": "Marlon Tobaben", "gender": "unknown", "institution": "University of Helsinki", "country": "FI", "position": "PhD student" }, { "name": "Yuan He", "gender": "Female", "institution": "Aalto University", "country": "FI", "position": "MS student" } ], "n_formula": 89, "n_formula_1": 19, "n_ref_uni": 27, "n_ref": 63, "n_ref_all": 108, "n_fig": 8, "n_tab": 11, "L_tab": 7728, "n_element_tab": 158, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 4, "n_element_tab_1": 2, "formula_len_all": 10891, "formula_len_all_1": 1750, "len_all": 178396, "len_all_1": 58537, "len_abs": 1155, "len_title": 140, "len_sents": 48590, "len_sents_1": 26283, "n_sents": 454, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1158, "title": "Impact of Dataset Properties on Membership Inference Vulnerability of Deep Transfer Learning", "abs": "We analyse the relationship between privacy vulnerability and dataset properties, such as examples per class and number of classes, when applying two state-of-the-art membership inference attacks (MIAs) to fine-tuned neural networks. We derive per-example MIA vulnerability in terms of score distributions and statistics computed from shadow models. We introduce a simplified model of membership inference and prove that in this model, the logarithm of the difference of true and false positive rates depends linearly on the logarithm of the number of examples per class. We complement the theoretical analysis with empirical analysis by systematically testing the practical privacy vulnerability of fine-tuning large image classification models and obtain the previously derived power law dependence between the number of examples per class in the data and the MIA vulnerability, as measured by true positive rate of the attack at a low false positive rate. Finally, we fit a parametric model of the previously derived form to predict true positive rate based on dataset properties and observe good fit for MIA vulnerability on unseen fine-tuning scenarios.", "keywords": [ "Membership Inference Attack", "Transfer Learning", "Few-shot Learning", "Image Classification" ], "rating_list": [ 3, 8, 3, 3 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 1, 3, 1, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "qqKJjwibsp", "primary_area": "datasets and benchmarks", "n_author": 16, "author_info_list": [ { "name": "Anjie Yang", "gender": "Male", "institution": "University of Ottawa", "country": "CA", "position": "Undergrad student" }, { "name": "Bernard Ghanem", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Full Professor" }, { "name": "Dai-Jie Wu", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "Research Assistant" }, { "name": "Guohao Li", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Jianbo Deng", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "MS student" }, { "name": "Linyao Chen", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Shilong Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Tianqi Xu", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Xiang Yao", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Researcher" }, { "name": "Yanjun Chen", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Yongchao Chen", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Zecheng Zhang", "gender": "Male", "institution": "Kumo.AI", "country": "US", "position": "Researcher" }, { "name": "Zhaoxuan Jin", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "US", "position": "MS student" }, { "name": "Zhiqiang Xie", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Qian Bochen", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 80, "n_ref_all": 111, "n_fig": 4, "n_tab": 9, "L_tab": 3049, "n_element_tab": 165, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2952, "n_element_tab_1": 176, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 257416, "len_all_1": 76359, "len_abs": 1216, "len_title": 124, "len_sents": 79128, "len_sents_1": 35893, "n_sents": 627, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1173, "title": "CRAB: Cross-environment Agent Benchmark for Multimodal Language Model Agents", "abs": "The development of autonomous agents increasingly relies on Multimodal Language Models (MLMs) to perform tasks described in natural language with GUI environments, such as websites, desktop computers, or mobile phones. Existing benchmarks for MLM agents in interactive environments are limited by their focus on a single environment, lack of detailed and generalized evaluation methods, and the complexities of constructing tasks and evaluators. To overcome these limitations, we introduce CRAB, the first agent benchmark framework designed to support cross-environment tasks, incorporating a graph-based fine-grained evaluation method and an efficient mechanism for task and evaluator construction. Our framework supports multiple devices and can be easily extended to any environment with a Python interface. Leveraging CRAB, we developed a cross-platform CRAB Benchmark-v0 comprising 120 tasks in computer desktop and mobile phone environments. We evaluated 6 advanced MLMs using different single and multi-agent system configurations on this benchmark. The experimental results demonstrate that the single agent with GPT-4o achieves the best completion ratio of 38.01%.", "keywords": [ "Benchmark", "GUI agent", "Multimodal Language Model" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "qq0zZMC4SM", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Jackie Ma", "gender": "unknown", "institution": "Fraunhofer HHI", "country": "DE", "position": "Head of Applied Machine Learning" }, { "name": "Jost Arndt", "gender": "Male", "institution": "Fraunhofer HHI", "country": "", "position": "PhD student" }, { "name": "Michael Detzel", "gender": "Male", "institution": "Fraunhofer HHI", "country": "DE", "position": "PhD student" }, { "name": "Utku Isil", "gender": "Male", "institution": "Fraunhofer HHI, Fraunhofer IAIS", "country": "DE", "position": "Researcher" }, { "name": "Wojciech Samek", "gender": "Male", "institution": "TU Berlin", "country": "DE", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 24, "n_ref": 37, "n_ref_all": 65, "n_fig": 9, "n_tab": 4, "L_tab": 472, "n_element_tab": 108, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 610, "n_element_tab_1": 66, "formula_len_all": 934, "formula_len_all_1": 934, "len_all": 122249, "len_all_1": 60187, "len_abs": 887, "len_title": 124, "len_sents": 38471, "len_sents_1": 27816, "n_sents": 326, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 900, "title": "Synthetic Datasets for Machine Learning on Spatio-Temporal Graphs using PDEs", "abs": "In this work, we describe the creation and use of synthetic datasets based on various partial differential equations to support spatio-temporal graph modeling in machine learning for different applications. More precisely, we showcase three equations to model different types of disasters and hazards in the fields of epidemiology, atmospheric particles, and tsunami waves. Further, we show how such created datasets can be used by benchmarking several machine learning models on the epidemiological dataset and, additionally, by showing how pre-training on such synthetic datasets can improve model performance on real-world epidemiological data. The presented methods enable others to create datasets and benchmarks customized to individual requirements. The source code for our methodology and the three created datasets can be found on https://github.com/github-usr-ano/Temporal_Graph_Data_PDEs .", "keywords": [ "Data", "Dataset", "PDE", "Graph", "Spatio-Temporal", "Epidemiology", "Benchmarking" ], "rating_list": [ 6, 3, 3, 8 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 1, 3, 4 ], "confidence_list": [ 4, 5, 4, 2 ] }, { "paper_id": "qpz84ykqgv", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Daniel John Lawson", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "Associate Professor" }, { "name": "Maximilian J. Werner", "gender": "unknown", "institution": "University of Bristol", "country": "GB", "position": "Associate Professor" }, { "name": "Samuel Stockman", "gender": "Male", "institution": "University of Bristol", "country": "", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 13, "n_ref_uni": 81, "n_ref": 129, "n_ref_all": 153, "n_fig": 13, "n_tab": 7, "L_tab": 3844, "n_element_tab": 482, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1861, "n_element_tab_1": 114, "formula_len_all": 870, "formula_len_all_1": 802, "len_all": 203019, "len_all_1": 55054, "len_abs": 1564, "len_title": 88, "len_sents": 49921, "len_sents_1": 24556, "n_sents": 331, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1570, "title": "EarthquakeNPP: Benchmark Datasets for Earthquake Forecasting with Neural Point Processes", "abs": "Classical point process models, such as the epidemic-type aftershock sequence (ETAS) model, have been widely used for forecasting the event times and locations of earthquakes for decades. Recent advances have led to Neural Point Processes (NPPs), which promise greater flexibility and improvements over classical models. However, the currently-used benchmark dataset for NPPs does not represent an up-to-date challenge in the seismological community since it lacks a key earthquake sequence from the region and improperly splits training and testing data. Furthermore, initial earthquake forecast benchmarking lacks a comparison to state-of-the-art earthquake forecasting models typically used by the seismological community. To address these gaps, we introduce EarthquakeNPP: a collection of benchmark datasets to facilitate testing of NPPs on earthquake data, accompanied by a credible implementation of the ETAS model. The datasets cover a range of small to large target regions within California, dating from 1971 to 2021, and include different methodologies for dataset generation. In a benchmarking experiment, we compare three spatio-temporal NPPs against ETAS and find that none outperform ETAS in either spatial or temporal log-likelihood. These results indicate that current NPP implementations are not yet suitable for practical earthquake forecasting. EarthquakeNPP also provides generative evaluation metrics, enabling broader model classes to be benchmarked and facilitating the future collaboration between the seismology and machine learning communities.", "keywords": [ "Point Processes", "Earthquake Forecasting", "Benchmarking", "Datasets" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 2, 2, 3 ] }, { "paper_id": "qpXctF2aLZ", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Christian Bartelt", "gender": "Male", "institution": "Technische Universit\u00e4t Clausthal", "country": "DE", "position": "Full Professor" }, { "name": "Florian Vogt", "gender": "Not Specified", "institution": "Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "MS student" }, { "name": "Heiner Stuckenschmidt", "gender": "Male", "institution": "University of Mannheim", "country": "", "position": "Full Professor" }, { "name": "Sascha Marton", "gender": "Male", "institution": "Technische Universit\u00e4t Clausthal", "country": "DE", "position": "Lecturer" }, { "name": "Stefan L\u00fcdtke", "gender": "unknown", "institution": "University of Rostock", "country": "", "position": "Assistant Professor" }, { "name": "Tim Grams", "gender": "Male", "institution": "Technische Universit\u00e4t Clausthal", "country": "DE", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 52, "n_ref": 97, "n_ref_all": 147, "n_fig": 22, "n_tab": 11, "L_tab": 2284, "n_element_tab": 378, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 421, "n_element_tab_1": 40, "formula_len_all": 799, "formula_len_all_1": 499, "len_all": 185875, "len_all_1": 70805, "len_abs": 1191, "len_title": 114, "len_sents": 55922, "len_sents_1": 32928, "n_sents": 439, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 88, "L_abs": 1180, "title": "Mitigating Information Loss in Tree-Based Reinforcement Learning via Direct Optimization", "abs": "Reinforcement learning (RL) has seen significant success across various domains, but its adoption is often limited by the black-box nature of neural network policies, making them difficult to interpret. In contrast, symbolic policies allow representing decision-making strategies in a compact and interpretable way. However, learning symbolic policies directly within on-policy methods remains challenging.\nIn this paper, we introduce SYMPOL, a novel method for SYMbolic tree-based on-POLicy RL. SYMPOL employs a tree-based model integrated with a policy gradient method, enabling the agent to learn and adapt its actions while maintaining a high level of interpretability.\nWe evaluate SYMPOL on a set of benchmark RL tasks, demonstrating its superiority over alternative tree-based RL approaches in terms of performance and interpretability. Unlike existing methods, it enables gradient-based, end-to-end learning of interpretable, axis-aligned decision trees within standard on-policy RL algorithms. Therefore, SYMPOL can become the foundation for a new class of interpretable RL based on decision trees. Our implementation is available under: https://github.com/s-marton/sympol", "keywords": [ "Symbolic Reinforcement Learning", "Interpretable Reinforcement Learning", "Reinforcement Learning", "Decision Trees", "Policy Gradient", "Proximal Policy Optimization" ], "rating_list": [ 8, 10, 8, 5, 5 ], "soundness_list": [ 4, 4, 3, 2, 3 ], "presentation_list": [ 4, 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 5, 5, 4, 4, 5 ] }, { "paper_id": "qpI6GO80ri", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Hongxin Wei", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "HuaJun Xi", "gender": "unknown", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Jianguo Huang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Intern" }, { "name": "Kangdao Liu", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "PhD student" }, { "name": "Lei Feng", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 38, "n_formula_1": 17, "n_ref_uni": 51, "n_ref": 130, "n_ref_all": 177, "n_fig": 9, "n_tab": 6, "L_tab": 3003, "n_element_tab": 493, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1821, "n_element_tab_1": 315, "formula_len_all": 1679, "formula_len_all_1": 910, "len_all": 180889, "len_all_1": 68445, "len_abs": 1369, "len_title": 115, "len_sents": 55068, "len_sents_1": 30174, "n_sents": 477, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1377, "title": "Delving into Temperature Scaling for Adaptive Conformal Prediction", "abs": "Conformal prediction, as an emerging uncertainty qualification technique, constructs prediction sets that are guaranteed to contain the true label with pre-defined probability. Previous works often employ temperature scaling to calibrate the classifier, assuming that confidence calibration can benefit conformal prediction. In this work, we empirically show that current confidence calibration methods (e.g., temperature scaling) normally lead to larger prediction sets in adaptive conformal prediction. Theoretically, we prove that a prediction with higher confidence could result in a smaller prediction set on expectation. Inspired by the analysis, we propose \\textbf{Conformal Temperature Scaling} (ConfTS), a variant of temperature scaling that aims to improve the efficiency of adaptive conformal prediction. Specifically, ConfTS optimizes the temperature value by minimizing the gap between the threshold and the non-conformity score of the ground truth for a held-out validation dataset. In this way, the temperature value obtained would lead to an optimal set with high efficiency without violating the coverage. Experiments demonstrate that our method can effectively enhance adaptive conformal prediction methods in both efficiency and conditional coverage, reducing the average size of APS and RAPS by approximately 50$\\%$ on ImageNet with error rate $\\alpha=0.1$.", "keywords": [ "trustworthy machine learning", "conformal prediction", "confidence calibration" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "qpDqO7qa3R", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Changhan Yeh", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Chi-Wei Hsiao", "gender": "Female", "institution": "MediaTek Inc.", "country": "TW", "position": "Researcher" }, { "name": "Chin-Yang Lin", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "Ting-Hsuan Chen", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "MS student" }, { "name": "Zhixiang Wang", "gender": "Male", "institution": "CyberAgent AI Lab", "country": "JP", "position": "Researcher" }, { "name": "Shiu Hau-Shiang", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "Yu Lun Liu", "gender": "unknown", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 61, "n_ref": 125, "n_ref_all": 156, "n_fig": 18, "n_tab": 3, "L_tab": 3795, "n_element_tab": 81, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 4352, "n_element_tab_1": 91, "formula_len_all": 544, "formula_len_all_1": 544, "len_all": 167464, "len_all_1": 64516, "len_abs": 1072, "len_title": 138, "len_sents": 35868, "len_sents_1": 26819, "n_sents": 295, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1082, "title": "DiffIR2VR-Zero: Zero-Shot Video Restoration with Diffusion-based Image Restoration Models", "abs": "This paper introduces a method for zero-shot video restoration using pre-trained image restoration diffusion models. Traditional video restoration methods often need retraining for different settings and struggle with limited generalization across various degradation types and datasets. Our approach uses a hierarchical latent warping strategy for keyframes and local frames, combined with token merging that uses a hybrid correspondence mechanism that integrates spatial information, optical flow, and feature-based matching. We show that our method not only achieves top performance in zero-shot video restoration but also significantly surpasses trained models in generalization across diverse datasets and extreme degradations (8$\\times$ super-resolution and high-standard deviation video denoising). We present evidence through quantitative metrics and visual comparisons on various challenging datasets. Additionally, our technique works with any 2D restoration diffusion model, offering a versatile and powerful tool for video enhancement tasks without extensive retraining.", "keywords": [ "Video restoration", "Zero-shot", "Training-free", "Diffusion models" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "qotIZREPZf", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Arti Yardi", "gender": "Female", "institution": "International Institute of Information technology", "country": "", "position": "Assistant Professor" }, { "name": "Shikhar Saxena", "gender": "Male", "institution": "International Institute of Information Technology, Hyderabad", "country": "", "position": "Undergrad student" }, { "name": "Tejas Bodas", "gender": "Male", "institution": "International Institute of Information Technology Hyderabad, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 28, "n_formula_1": 21, "n_ref_uni": 12, "n_ref": 18, "n_ref_all": 38, "n_fig": 4, "n_tab": 1, "L_tab": 254, "n_element_tab": 19, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 579, "n_element_tab_1": 40, "formula_len_all": 2042, "formula_len_all_1": 1385, "len_all": 81778, "len_all_1": 57096, "len_abs": 1275, "len_title": 109, "len_sents": 24699, "len_sents_1": 22534, "n_sents": 238, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1281, "title": "CGD: Modifying the Loss Landscape by Gradient Regularization", "abs": "Line-search methods are commonly used to solve optimization problems. The simplest line search method is the steepest descent where we always move in the direction of the negative gradient. Newton\u2019s method on the other hand is a second-order method that uses the curvature information in the Hessian to pick the descent direction. In this work, we propose a new line-search method called Constrained Gradient Descent (CGD) that implicitly changes the landscape of the objective function for efficient optimization. CGD is formulated as a solution to the constrained version of the original problem where the constraint is on a function of the gradient. We optimize the corresponding Lagrangian function thereby favourably changing the landscape of the objective function. This results in a line search procedure where the Lagrangian penalty acts as a control over the descent direction and can therefore be used to iterate over points that have smaller gradient values, compared to iterates of vanilla steepest descent. We reinterpret and draw parallels with the Explicit Gradient Regularization (EGR) method, discussing its drawbacks and potential enhancements. Numerical experiments are conducted on synthetic test functions to illustrate the performance of CGD and its variants.", "keywords": [ "optimization", "gradient regularization" ], "rating_list": [ 1, 1, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "qoGdpin3om", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Joshua D. Berke", "gender": "unknown", "institution": "University of California, San Francisco", "country": "US", "position": "Full Professor" }, { "name": "Wei Wei", "gender": "Male", "institution": "University of California, San Francisco", "country": "US", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 44, "n_ref": 79, "n_ref_all": 117, "n_fig": 28, "n_tab": 1, "L_tab": 238, "n_element_tab": 40, "n_fig_1": 27, "n_tab_1": 1, "L_tab_1": 238, "n_element_tab_1": 40, "formula_len_all": 376, "formula_len_all_1": 376, "len_all": 99712, "len_all_1": 68211, "len_abs": 1301, "len_title": 152, "len_sents": 27414, "len_sents_1": 26052, "n_sents": 224, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1308, "title": "Dopamine transients in the ventral striatum provide evidence for average-reward reinforcement learning", "abs": "Agents in real environments need to organize their behavior over a wide range of time scales. This might be achieved by reinforcement learning (RL) algorithms employing a spectrum of discount factors. Neural evidence for this idea includes recordings of dopamine (DA) release transients, which appear to reflect shorter time horizons in dorsal striatum and much longer horizons in ventral striatum (VS). However, this also presents a challenge, because with very long time horizons all states have similar, large values, impeding learning. Prior theoretical work has therefore proposed algorithms, including average-reward RL, that segregate out the large shared component of value. Here we compare temporal-difference reward prediction errors derived from recurrent neural network models (RNNs) to rat VS DA transients measured in three behavioral tasks. We show that using average-reward RL to train RNNs can provide an improved match to VS DA, compared to using discounting alone. We further find that the activity dynamics in RNNs trained with average-reward RL readily encodes key decision variables such as recent reward history, in a task-specific manner. The functional alignment between DA dynamics and average-reward RL may offer new insights into neural mechanisms of learning and decision-making.", "keywords": [ "Average-reward reinforcement learning", "Dopamine", "Reward prediction error", "Ventral striatum" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "qnlG3zPQUy", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Akanksha Singh", "gender": "Female", "institution": "Indian Institute of Technology, Madras", "country": "IN", "position": "Researcher" }, { "name": "Akshat Jain", "gender": "Male", "institution": "Indian Institute of Technology, Jodhpur", "country": "IN", "position": "Undergrad student" }, { "name": "Kartik Thakral", "gender": "Male", "institution": "Indian Institute of Technology Jodhpur, India", "country": "IN", "position": "PhD student" }, { "name": "Mayank Vatsa", "gender": "Male", "institution": "Indian Institute of Technology Jodhpur, India", "country": "IN", "position": "Full Professor" }, { "name": "Richa Singh", "gender": "Female", "institution": "Indian Institute of Technology Jodhpur, India", "country": "IN", "position": "Full Professor" }, { "name": "RISHABH RANJAN", "gender": "Male", "institution": "Indian Institute of Technology Jodhpur, India", "country": "IN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 49, "n_ref": 65, "n_ref_all": 93, "n_fig": 4, "n_tab": 10, "L_tab": 4438, "n_element_tab": 589, "n_fig_1": 3, "n_tab_1": 10, "L_tab_1": 5464, "n_element_tab_1": 637, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 159332, "len_all_1": 80696, "len_abs": 1661, "len_title": 139, "len_sents": 36911, "len_sents_1": 32058, "n_sents": 262, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1678, "title": "ILLUSION: Unveiling Truth with a Comprehensive Multi-Modal, Multi-Lingual Deepfake Dataset", "abs": "The proliferation of deepfakes and AI-generated content has led to a significant increase in media forgeries and misinformation, necessitating development of more robust detection systems. Current datasets, however, lack comprehensive diversity across modalities, languages, and real-world scenarios. To address this gap, we present ILLUSION (Integration of Life-Like Unique Synthetic Identities and Objects from Neural Networks), a large-scale multi-modal deepfake dataset comprising over 1.3 million samples. ILLUSION encompasses (i) audio-visual forgeries, (ii) diverse linguistic content with over 26 languages, (iii) challenging noisy environments, and (iv) various manipulation protocols. Generated using state-of-the-art generative models, ILLUSION includes face swaps, audio spoofing, synchronized audio-video manipulations, and synthetic images, faces, and videos. The proposed dataset has balanced representation of gender and skin tone, supports multilingual experiments, and is designed to facilitate development of robust multi-modal detection systems. We benchmarked state-of-the-art algorithms across multiple modalities including image-based, audio-based, video-based, and multi-modal detection. The results highlight critical challenges such as (a) performance degradation in multi-lingual and multi-modal contexts, (b) accuracy reduction in noisy environments, and (c) limited generalization to real-world scenarios and zero-day attacks. It is our assertion that the comprehensive nature of the proposed dataset enables researchers to develop and evaluate more resilient deepfake detection methods, addressing the evolving landscape of synthetic media threats.", "keywords": [ "Multi-modal", "Multi-Lingual", "Deepfake", "AIGC" ], "rating_list": [ 5, 5, 8, 6, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 4, 3, 3 ], "confidence_list": [ 5, 3, 4, 4, 4 ] }, { "paper_id": "qnGir4dyu9", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jaehong Yoon", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Postdoc" }, { "name": "Mohit Bansal", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" }, { "name": "Shoubin Yu", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 51, "n_ref": 154, "n_ref_all": 190, "n_fig": 22, "n_tab": 7, "L_tab": 1944, "n_element_tab": 156, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 755, "n_element_tab_1": 64, "formula_len_all": 135, "formula_len_all_1": 135, "len_all": 217226, "len_all_1": 64267, "len_abs": 0, "len_title": 90, "len_sents": 61256, "len_sents_1": 30676, "n_sents": 476, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 2005, "title": "RACCooN: A Versatile Instructional Video Editing Framework with Auto-Generated Narratives", "abs": "Recent video generative models primarily rely on carefully written text prompts for specific tasks, like inpainting or style editing. They require labor-intensive textual descriptions for input videos, hindering their flexibility to adapt personal/raw videos to user specifications. This paper proposes RACCooN, a versatile and user-friendly video-to-paragraph-to-video generative framework that supports multiple video editing capabilities, such as removal, addition, and modification, through a unified pipeline. RACCooN consists of two principal stages: Video-to-Paragraph (V2P) and Paragraph-to-Video (P2V). In the V2P stage, we automatically describe video scenes in well-structured natural language, capturing both the holistic context and focused object details. Subsequently, in the P2V stage, users can optionally refine these descriptions to guide the video diffusion model, enabling various modifications to the input video, such as removing, changing subjects, and/or adding new objects. The proposed approach stands out from other methods through several significant contributions: (1) RACCooN suggests a multi-granular spatiotemporal pooling strategy to generate well-structured video descriptions, capturing both the broad context and object details without requiring complex human annotations, simplifying precise video content editing based on text for users. (2) Our video generative model incorporates auto-generated narratives or instructions to enhance the quality and accuracy of the generated content. (3) RACCooN also plans to imagine new objects in a given video, so users simply prompt the model to receive a detailed video editing plan for complex video editing. The proposed framework demonstrates impressive versatile capabilities in video-to-paragraph generation (up to 9.4% absolute improvement in human evaluations against the baseline), video content editing (relative 49.7% in FVD), and can be incorporated into other SoTA video generative models for further enhancement.", "keywords": [ "Video Inpainting", "Video Editing", "Video Caption", "Multimodal Large Language Model" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "qnAZqlMGTB", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Chi Chen", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Fuwen Luo", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Junming Lin", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Maosong Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Peng Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yang Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Zheng Fang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Zihao Wan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" } ], "n_formula": 8, "n_formula_1": 1, "n_ref_uni": 14, "n_ref": 57, "n_ref_all": 80, "n_fig": 10, "n_tab": 5, "L_tab": 3322, "n_element_tab": 173, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1442, "n_element_tab_1": 182, "formula_len_all": 400, "formula_len_all_1": 313, "len_all": 137967, "len_all_1": 60874, "len_abs": 1445, "len_title": 133, "len_sents": 44957, "len_sents_1": 28728, "n_sents": 346, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1445, "title": "StreamingBench: Assessing the Gap for MLLMs to Achieve Streaming Video Understanding", "abs": "The rapid development of Multimodal Large Language Models (MLLMs) has expanded their capabilities from image comprehension to video understanding. However, most of these MLLMs focus primarily on of\ufb02ine video comprehension, necessitating extensive processing of all video frames before any queries can be made. This presents a signi\ufb01cant gap compared to the human ability to watch, listen, think, and respond to streaming inputs in real time, highlighting the limitations of current MLLMs. In this paper, we introduce StreamingBench, the \ufb01rst comprehensive benchmark designed to evaluate the streaming video understanding capabilities of MLLMs. StreamingBench assesses three core aspects of streaming video understanding: (1) real-time visual understanding, (2) omni-source understanding and (3) contextual understanding. The benchmark consists of 18 tasks, featuring 900 videos and 4,500 human-curated QA pairs. Each video features \ufb01ve questions presented at different time points to simulate a continuous streaming scenario. We conduct experiments on StreamingBench with 15 open-source and proprietary MLLMs and \ufb01nd that even the most advanced proprietary MLLMs like Gemini 1.5 Pro and GPT-4o perform signi\ufb01cantly below human-level streaming video understanding capabilities. We hope our work can facilitate further advancements for MLLMs, empowering them to approach human-level video comprehension and interaction in more realistic scenarios.", "keywords": [ "Benchmark", "Streaming Video Understanding", "Multimodal Large Language Models", "Video Benchmark", "Evaluation" ], "rating_list": [ 8, 6, 6, 3 ], "soundness_list": [ 4, 4, 3, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "qn9tBYQHGi", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Asuman E. Ozdaglar", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Chanwoo Park", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Xiangyu Liu", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "kaiqing zhang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" } ], "n_formula": 227, "n_formula_1": 19, "n_ref_uni": 92, "n_ref": 278, "n_ref_all": 317, "n_fig": 3, "n_tab": 4, "L_tab": 208, "n_element_tab": 20, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 26, "n_element_tab_1": 3, "formula_len_all": 36232, "formula_len_all_1": 1686, "len_all": 563551, "len_all_1": 89656, "len_abs": 1822, "len_title": 116, "len_sents": 162168, "len_sents_1": 38194, "n_sents": 1478, "n_sents_1": 304, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1816, "title": "Do LLM Agents Have Regret? A Case Study in Online Learning and Games", "abs": "Large language models (LLMs) have been increasingly employed for (interactive) decision-making, via the development of LLM-based autonomous agents. Despite their emerging successes, the performance of LLM agents in decision-making has not been fully investigated through quantitative metrics, especially in the multi-agent setting when they interact with each other, a typical scenario in real-world LLM-agent applications. To better understand the limits of LLM agents in these interactive environments, we propose to study their interactions in benchmark decision-making settings in online learning and game theory, through the performance metric of regret. We first empirically study the no-regret behaviors of LLMs in canonical non-stochastic online learning problems, as well as the emergence of equilibria when LLM agents interact through playing repeated games. We then provide some theoretical insights into the no-regret behaviors of LLM agents, under certain assumptions on the supervised pre-training and the rationality model of human decision-makers who generate the data. Notably, we also identify (simple) cases where advanced LLMs such as GPT-4 fail to be no-regret. To further promote the no-regret behaviors, we propose a novel unsupervised training loss of regret-loss, which, in contrast to the supervised pre-training loss, does not require the labels of (optimal) actions. Finally, we establish the statistical guarantee of generalization bound for regret-loss minimization, and more importantly, the optimization guarantee that minimizing such a loss may automatically lead to known no-regret learning algorithms, when single-layer self-attention models are used. Our further experiments demonstrate the effectiveness of our regret-loss, especially in addressing the above \u201cregrettable\u201d cases.", "keywords": [ "LLM agents", "online learning", "repeated games" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 4, 4, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 3, 4, 2 ] }, { "paper_id": "qmsX2R19p9", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Gang Chen", "gender": "Male", "institution": "College of Computer Science and Technology, Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Haobo Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Junbo Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Lin Long", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Sai Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Wentao Ye", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Xijun Gu", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Xinjie Sun", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 39, "n_ref": 117, "n_ref_all": 142, "n_fig": 7, "n_tab": 18, "L_tab": 3907, "n_element_tab": 381, "n_fig_1": 4, "n_tab_1": 14, "L_tab_1": 3291, "n_element_tab_1": 347, "formula_len_all": 419, "formula_len_all_1": 313, "len_all": 192203, "len_all_1": 61788, "len_abs": 1105, "len_title": 120, "len_sents": 44067, "len_sents_1": 26413, "n_sents": 284, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1121, "title": "LLM Table Reading: Bridging the Semantic Gap Between Text and Table", "abs": "The rise of Large Language Models (LLMs) has revolutionized numerous domains, yet these models still exhibit weakness in understanding structured tabular data.\nAlthough a larger context window of next-generation LLMs promises them to accommodate a larger volume of table contents, it does not inherently improve the model's capability to understand the underlying structure and semantics of tabular data.\nTo bridge the semantic gap between **T**ext and **T**able caused by structural variance, we propose TNT, a table-language model that features multi-modal table representations to empower LLMs to effectively and efficiently abstract structure-enriched semantics from tabular data. \nTNT also introduces a scalable and efficient training pipeline, featuring novel self-supervised learning tasks, to integrate abstract tabular knowledge into the language modality.\nExtensive experimental results on the most iconic table understanding task of NL2SQL demonstrate a (much) better table understanding of TNT, which achieves up to **14.4%** higher execution accuracy compared with traditional text-based table serialization.", "keywords": [ "Large Language Model", "Table Representation Learning", "NL2SQL" ], "rating_list": [ 8, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "qmqRdxQcMA", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Hoyoung Kim", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "PhD student" }, { "name": "Jungseul Ok", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Associate Professor" }, { "name": "Sehyun Hwang", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "PhD student" }, { "name": "Suha Kwak", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Associate Professor" }, { "name": "Yeho Gwon", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Undergrad student" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 52, "n_ref": 85, "n_ref_all": 153, "n_fig": 22, "n_tab": 3, "L_tab": 504, "n_element_tab": 60, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 141, "n_element_tab_1": 11, "formula_len_all": 1170, "formula_len_all_1": 711, "len_all": 160433, "len_all_1": 63240, "len_abs": 862, "len_title": 117, "len_sents": 49621, "len_sents_1": 29812, "n_sents": 415, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 865, "title": "Enhancing Cost Efficiency in Active Learning with Candidate Set Query", "abs": "This paper introduces a cost-efficient active learning (AL) framework for classification, featuring a novel query design called candidate set query. Unlike traditional AL queries requiring the oracle to examine all possible classes, our method narrows down the set of candidate classes likely to include the ground-truth class, significantly reducing the search space and labeling cost. Moreover, we leverage conformal prediction to dynamically generate small yet reliable candidate sets, adapting to model enhancement over successive AL rounds. To this end, we introduce an acquisition function designed to prioritize data points that offer high information gain at lower cost. Empirical evaluations on CIFAR-10, CIFAR-100, and ImageNet64x64 demonstrate the effectiveness and scalability of our framework. Notably, it reduces labeling cost by 42% on ImageNet64x64.", "keywords": [ "Active Learning", "Conformal Prediction", "Label efficient learning" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "qmodrqswtF", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Jack Teversham", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Researcher" }, { "name": "Ritam Majumdar", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Sonali Parbhoo", "gender": "unknown", "institution": "Imperial College London, Imperial College London", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 96, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 83, "n_ref_all": 98, "n_fig": 8, "n_tab": 7, "L_tab": 4328, "n_element_tab": 293, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 20690, "formula_len_all_1": 1415, "len_all": 230104, "len_all_1": 73403, "len_abs": 1370, "len_title": 84, "len_sents": 66711, "len_sents_1": 34155, "n_sents": 607, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 1376, "title": "Concept-driven Off Policy Evaluation", "abs": "Evaluating off-policy decisions using batch data poses significant challenges due to high variance and limited sample sizes, making reliable evaluation difficult. To improve Off-Policy Evaluation (OPE) performance, we must identify and address the sources of this variance. Recent research on Concept Bottleneck Models (CBMs) shows that using human-explainable concepts can improve predictions and provide better understanding. We propose incorporating concepts into OPE to reduce variance through targeted interventions. Shared disease characteristics, for example, could help identify better treatment options, despite variations in patient vitals. Our work introduces a family of concept-based OPE estimators, proving that they remain unbiased and reduce variance when concepts are known and predefined. Since real-world applications often lack predefined concepts, we further develop an end-to-end algorithm to learn interpretable, concise, and diverse parameterized concepts optimized for variance reduction. Our experiments with synthetic and real-world datasets show that both known and learned concept-based estimators significantly improve OPE performance. Crucially, we show that, unlike other OPE methods, concept-based estimators are easily interpretable and allow for targeted interventions on specific concepts, further enhancing the quality of these estimators.", "keywords": [ "Off Policy Evaluation", "Reinforcement Learning", "Interpretability", "Concept Bottleneck Models" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 2, 4, 2, 4 ] }, { "paper_id": "qlzxeNESWI", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Cem Tekin", "gender": "Male", "institution": "Bilkent University", "country": "TR", "position": "Associate Professor" }, { "name": "Eray Can Elumar", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Osman Yagan", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "", "position": "Research Professor" } ], "n_formula": 20, "n_formula_1": 21, "n_ref_uni": 20, "n_ref": 39, "n_ref_all": 49, "n_fig": 2, "n_tab": 4, "L_tab": 313, "n_element_tab": 20, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 334, "n_element_tab_1": 24, "formula_len_all": 2440, "formula_len_all_1": 2305, "len_all": 112551, "len_all_1": 82215, "len_abs": 1011, "len_title": 78, "len_sents": 40433, "len_sents_1": 36273, "n_sents": 285, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 30, "L_abs": 1018, "title": "Bandits with Anytime Knapsacks", "abs": "We consider bandits with anytime knapsacks (BwAK), a novel version of the BwK problem where there is an anytime cost constraint instead of a total cost budget. This problem setting introduces additional complexities as it mandates adherence to the constraint throughout the decision-making process. We propose SUAK, an algorithm that utilizes upper confidence bounds to identify the optimal mixture of arms while maintaining a balance between exploration and exploitation. SUAK is an adaptive algorithm that strategically utilizes the available budget in each round in the decision-making process and skips a round when it is possible to violate the anytime cost constraint. In particular, SUAK slightly under-utilizes the available cost budget to reduce the need for skipping rounds. We show that SUAK attains the same problem-dependent regret upper bound of $ O(K \\log T)$ established in prior work under the simpler BwK framework. Finally, we provide simulations to verify the utility of SUAK in practical settings.", "keywords": [ "multi-armed bandits", "knapsack problem", "online learning" ], "rating_list": [ 8, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "qkzPr74cMx", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Guangyu Shen", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Kaiyuan Zhang", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Lu Yan", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Siyuan Cheng", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Xiangyu Zhang", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Full Professor" }, { "name": "Xuan Chen", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "ZHUO ZHANG", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 30, "n_ref": 67, "n_ref_all": 88, "n_fig": 2, "n_tab": 8, "L_tab": 5163, "n_element_tab": 470, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 4373, "n_element_tab_1": 389, "formula_len_all": 597, "formula_len_all_1": 535, "len_all": 196729, "len_all_1": 72835, "len_abs": 1405, "len_title": 122, "len_sents": 70426, "len_sents_1": 30366, "n_sents": 560, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1414, "title": "ASPIRER: Bypassing System Prompts with Permutation-based Backdoors in LLMs", "abs": "Large Language Models (LLMs) have become integral to many applications, with system prompts serving as a key mechanism to regulate model behavior and ensure ethical outputs. In this paper, we introduce a novel backdoor attack that systematically bypasses these system prompts, posing significant risks to the AI supply chain. Under normal conditions, the model adheres strictly to its system prompts. However, our backdoor allows malicious actors to circumvent these safeguards when triggered. Specifically, we explore a scenario where an LLM provider embeds a covert trigger within the base model. A downstream deployer, unaware of the hidden trigger, fine-tunes the model and offers it as a service to users. Malicious actors can purchase the trigger from the provider and use it to exploit the deployed model, disabling system prompts and achieving restricted outcomes. Our attack utilizes a permutation trigger, which activates only when its components are arranged in a precise order, making it computationally challenging to detect or reverse-engineer. We evaluate our approach on five state-of-the-art models, demonstrating that our method achieves an attack success rate (ASR) of up to 99.50\\% while maintaining a clean accuracy (CACC) of 98.58\\%, even after defensive fine-tuning. These findings highlight critical vulnerabilities in LLM deployment pipelines and underscore the need for stronger defenses.", "keywords": [ "LLM", "Safety", "backdoor attack" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "qkVsGBff9s", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Yang Zhang", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Postdoc" }, { "name": "Yunjian Xu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Zhihe YANG", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 65, "n_formula_1": 27, "n_ref_uni": 67, "n_ref": 150, "n_ref_all": 184, "n_fig": 15, "n_tab": 9, "L_tab": 4761, "n_element_tab": 450, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 2007, "n_element_tab_1": 203, "formula_len_all": 7323, "formula_len_all_1": 2452, "len_all": 250710, "len_all_1": 74655, "len_abs": 1503, "len_title": 157, "len_sents": 70770, "len_sents_1": 32031, "n_sents": 512, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1512, "title": "Q-Supervised Contrastive Representation: A State Decoupling Framework for Safe Offline Reinforcement Learning", "abs": "Safe offline reinforcement learning (RL), which aims to learn the safety-guaranteed policy without risky online interaction with environments, has attracted growing recent attention for safety-critical scenarios. However, existing approaches encounter out-of-distribution problems during the testing phase, which can result in potentially unsafe outcomes. This issue arises due to the infinite possible combinations of reward-related and cost-related states. In this work, we propose *State Decoupling with Q-supervised Contrastive representation* (SDQC), a novel framework that decouples the global observations into reward- and cost-related representations for decision-making, thereby improving the generalization capability for unfamiliar global observations.\nCompared with the classical representation learning methods, which typically require model-based estimation (e.g., bisimulation), we theoretically prove that our Q-supervised method generates a coarser representation while preserving the optimal policy, resulting in improved generalization performance. Experiments on DSRL benchmark problems provide compelling evidence that SDQC surpasses other baseline algorithms, especially for its exceptional ability to achieve almost zero violations in more than half of the tasks, \nwhile the state-of-the-art algorithm can only achieve the same level of success in a quarter of the tasks. Further, we demonstrate that SDQC possesses superior generalization ability when confronted with unseen environments.", "keywords": [ "Safe Reinforcement Learning", "Offline Reinforcement Learning", "Representation Learning", "Contrastive Learning", "Self-Supervised Learning" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "qkBBHixPow", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Jeongwhan Choi", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "PhD student" }, { "name": "Kookjin Lee", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" }, { "name": "Noseong Park", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Youn-Yeol Yu", "gender": "Male", "institution": "Yonsei University", "country": "", "position": "PhD student" }, { "name": "JAEHYEON PARK", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 53, "n_ref": 115, "n_ref_all": 149, "n_fig": 15, "n_tab": 10, "L_tab": 3432, "n_element_tab": 243, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 2228, "n_element_tab_1": 218, "formula_len_all": 553, "formula_len_all_1": 375, "len_all": 155367, "len_all_1": 65863, "len_abs": 1105, "len_title": 149, "len_sents": 45852, "len_sents_1": 27381, "n_sents": 405, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 101, "L_abs": 1114, "title": "PIORF: Physics-Informed Ollivier-Ricci Flow for Long\u2013Range Interactions in Mesh Graph Neural Networks", "abs": "Recently, data-driven simulators based on graph neural networks have gained attention in modeling physical systems on unstructured meshes. However, they struggle with long-range dependencies in fluid flows, particularly in refined mesh regions. This challenge, known as the 'over-squashing' problem, hinders information propagation. While existing graph rewiring methods address this issue to some extent, they only consider graph topology, overlooking the underlying physical phenomena. We propose Physics-Informed Ollivier--Ricci Flow (PIORF), a novel rewiring method that combines physical correlations with graph topology. PIORF uses Ollivier--Ricci curvature (ORC) to identify bottleneck regions and connects these areas with nodes in high-velocity gradient nodes, enabling long-range interactions and mitigating over-squashing. Our approach is computationally efficient in rewiring edges and can scale to larger simulations. Experimental results on 3 fluid dynamics benchmark datasets show that PIORF consistently outperforms baseline models and existing rewiring methods, achieving up to 26.2\\% improvement.", "keywords": [ "graph neural network", "fluid dynamics", "simulation", "mesh", "physics", "over-squashing", "rewiring" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 4, 4, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "qk8JMpwWPh", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Gaurav S. Sukhatme", "gender": "Male", "institution": "Amazon", "country": "", "position": "Amazon Scholar" }, { "name": "Grace Zhang", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Ziyi Liu", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "Visiting Scholar" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 37, "n_ref": 57, "n_ref_all": 82, "n_fig": 13, "n_tab": 4, "L_tab": 740, "n_element_tab": 66, "n_fig_1": 12, "n_tab_1": 3, "L_tab_1": 412, "n_element_tab_1": 33, "formula_len_all": 606, "formula_len_all_1": 369, "len_all": 117740, "len_all_1": 60448, "len_abs": 1159, "len_title": 132, "len_sents": 39633, "len_sents_1": 30712, "n_sents": 250, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1162, "title": "Learning Generalizable and Well-Shaped Reward Functions from Too Few Demonstrations", "abs": "Inverse reinforcement learning (IRL) is an important problem that aims to learn a reward function and policy directly from demonstrations, which can often be easier to provide than a well-shaped reward function. However, many real-world tasks include natural variations (i.e., a cleaning robot in a house with different furniture configurations), making it costly to provide demonstrations of every possible scenario. We tackle the problem of few-shot IRL with multi-task data where the goal is for an agent to learn from a few demonstrations, not sufficient to fully specify the task, by utilizing an offline multi-task demonstration dataset. Prior work utilizes meta-learning or imitation learning which additionally requires reward labels, a multi-task training environment, or cannot improve with online interactions. We propose Multitask Discriminator Proximity-guided IRL (MPIRL), an IRL method that learns a generalizable and well-shaped reward function by learning a multi-task generative adversarial discriminator with an auxiliary proximity-to-expert reward. We demonstrate the effectiveness of our method on multiple navigation and manipulation tasks.", "keywords": [ "Inverse Reinforcement Learning", "Imitation Learning", "Learning from Few Demonstrations" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "qk6AxjhFVR", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Huang Wei", "gender": "unknown", "institution": "Ant Group", "country": "", "position": "Researcher" }, { "name": "Lei Wang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Ruixuan Xiao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Shoumeng Yan", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Tao Wei", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Yinggui Wang", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 17, "n_ref": 33, "n_ref_all": 47, "n_fig": 1, "n_tab": 6, "L_tab": 2310, "n_element_tab": 218, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 2310, "n_element_tab_1": 218, "formula_len_all": 792, "formula_len_all_1": 792, "len_all": 102746, "len_all_1": 67731, "len_abs": 1422, "len_title": 131, "len_sents": 30825, "len_sents_1": 30507, "n_sents": 217, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1428, "title": "NESTLE: An Efficient and Robust Data Valuation Framework for Large Language Models", "abs": "The training and fine-tuning of large language models (LLMs) heavily rely on a large corpus of high-quality data. Nevertheless, the internet's extensive data is often of varying quality, and collecting high-quality data is exceedingly expensive. To facilitate data engineering and trading, the quantification of data value, also known as data valuation, is emerging as a critical topic. Traditional approaches for data valuation typically depend on model retraining. However, with the increasing model sizes and expansive data volumes of LLMs, these conventional methods are encountering significant declines in valuation precision, efficiency, and transferability. To alleviate these problems, we propose NESTLE, which is an efficient and robust framework for data valuation of LLMs. To accurately estimate the data value distribution across different target domains, we develop a training-free mechanism based on gradient tracing to simulate the data influences. To further tackle the dynamical value adjustment when multiple data providers coexist, we draw inspiration from the Shapley value theory and devise an accelerated strategy for estimating marginal contributions of data through gradient additivity. Extensive experiments demonstrate that our proposed framework NESTLE is capable of accurately and robustly providing accurate estimates of data value with a minuscule cost across a wide range of real-world scenarios.", "keywords": [ "Data Valuation", "Large Language Models", "Contribution Esitimation" ], "rating_list": [ 5, 8, 3, 5 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 2, 3, 1, 3 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "qjoDJjVZxB", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Jeff Calder", "gender": "Male", "institution": "University of Minnesota, Minneapolis", "country": "US", "position": "Associate Professor" }, { "name": "Wonjun Lee", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "", "position": "Postdoc" } ], "n_formula": 85, "n_formula_1": 32, "n_ref_uni": 27, "n_ref": 44, "n_ref_all": 114, "n_fig": 7, "n_tab": 1, "L_tab": 198, "n_element_tab": 3, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7823, "formula_len_all_1": 2652, "len_all": 142605, "len_all_1": 64897, "len_abs": 1851, "len_title": 156, "len_sents": 43854, "len_sents_1": 27937, "n_sents": 414, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1331, "title": "Understanding Contrastive Learning through Variational Analysis and Neural Network Optimization Perspectives", "abs": "The SimCLR method for contrastive learning of invariant visual representations has become extensively used in supervised, semi-supervised, and unsupervised settings, due to its ability to uncover patterns and structures in image data that are not directly present in the pixel representations. However, the reason for this success is not well-explained, since it is not guaranteed by invariance alone. In this paper, we conduct a mathematical analysis of the SimCLR method with the goal of better understanding the geometric properties of the learned latent distribution. Our findings reveal two things: (1) the SimCLR loss alone is not sufficient to select a \"good\" minimizer --- there are minimizers that give trivial latent distributions, even when the original data is highly clustered --- and (2) in order to understand the success of contrastive learning methods like SimCLR, it is necessary to analyze the neural network training dynamics induced by minimizing a contrastive learning loss. Our preliminary analysis for a one-hidden layer neural network shows that clustering structure can present itself for a substantial period of time during training, even if it eventually converges to a trivial minimizer. To substantiate our theoretical insights, we present numerical results that confirm our theoretical predictions.", "keywords": [ "contrastive learning", "discriminative", "neural network optimization", "variational analysis", "gradient flows" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 3, 4, 1, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "qit4pa6PpY", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Praveen Venkateswaran", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Prince Kumar", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Rudra Murthy", "gender": "Male", "institution": "IBM India Ltd", "country": "", "position": "Researcher" }, { "name": "Danish Contractor", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "Principal Researcher" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 31, "n_ref": 63, "n_ref_all": 92, "n_fig": 11, "n_tab": 11, "L_tab": 26653, "n_element_tab": 405, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 2749, "n_element_tab_1": 234, "formula_len_all": 2255, "formula_len_all_1": 740, "len_all": 191969, "len_all_1": 63822, "len_abs": 1096, "len_title": 136, "len_sents": 51260, "len_sents_1": 28268, "n_sents": 331, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1104, "title": "Evaluating the Instruction-following Abilities of Language Models using Knowledge Tasks", "abs": "In this work, we focus our attention on developing a benchmark for instruction-following where it is easy to verify both task performance as well as instruction-following capabilities. We adapt existing knowledge benchmarks and augment them with instructions that are a) conditional on correctly answering the knowledge task or b) use the space of candidate options in multiple-choice knowledge-answering tasks. This allows us to study model characteristics, such as their change in performance on the knowledge tasks in the presence of answer-modifying instructions and distractor instructions. In contrast to existing benchmarks for instruction following, we not only measure instruction-following capabilities but also use LLM-free methods to study task performance. We study a series of openly available large language models of varying parameter sizes (1B-405B) and closed source models namely GPT-4o-mini, GPT-4o. We find that even large-scale instruction-tuned LLMs fail to follow simple instructions in zero-shot settings. We release our dataset, the benchmark, code, and results for future work.", "keywords": [ "Large Language Models", "Instruction Following", "Evaluation Benchmark" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 1, 3, 2 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "qiluFujVc8", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Faguo Wu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "Assistant Professor" }, { "name": "Ronglong Cai", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Tianyuan Chen", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" }, { "name": "Xiao Zhang", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "Full Professor" } ], "n_formula": 44, "n_formula_1": 31, "n_ref_uni": 47, "n_ref": 110, "n_ref_all": 142, "n_fig": 10, "n_tab": 4, "L_tab": 568, "n_element_tab": 50, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 403, "n_element_tab_1": 31, "formula_len_all": 4149, "formula_len_all_1": 2618, "len_all": 133090, "len_all_1": 68830, "len_abs": 1187, "len_title": 135, "len_sents": 34233, "len_sents_1": 29463, "n_sents": 282, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1195, "title": "ACTIVE: Offline Reinforcement Learning via Adaptive Imitation and In-sample $V$-Ensemble", "abs": "Offline reinforcement learning (RL) aims to learn from static datasets and thus faces the challenge of value estimation errors for out-of-distribution actions. The in-sample learning scheme addresses this issue by performing implicit TD backups that does not query the values of unseen actions. However, pre-existing in-sample value learning and policy extraction methods suffer from over-regularization, limiting their performance on suboptimal or compositional datasets. In this paper, we analyze key factors in in-sample learning that might potentially hinder the use of a milder constraint. We propose Actor-Critic with Temperature adjustment and In-sample Value Ensemble (ACTIVE), a novel in-sample offline RL algorithm that leverages an ensemble of $V$-functions for critic training and adaptively adjusts the constraint level using dual gradient descent. We theoretically show that the $V$-ensemble suppresses the accumulation of initial value errors, thereby mitigating overestimation. Our experiments on the D4RL benchmarks demonstrate that ACTIVE alleviates overfitting of value functions and outperforms existing in-sample methods in terms of learning stability and policy optimality.", "keywords": [ "Offline Reinforcement Learning", "In-sample Learning", "Ensemble", "Uncertainty Quantification", "Dual Gradient Descent" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "qi7udwV66M", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Michael Elad", "gender": "Male", "institution": "Computer Science Department, Technion - Israel Institute of Technology", "country": "IL", "position": "Full Professor" }, { "name": "Noam Elata", "gender": "Male", "institution": "Apple", "country": "IL", "position": "Intern" }, { "name": "Tomer Michaeli", "gender": "Male", "institution": "Technion, Technion", "country": "IL", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 48, "n_ref": 95, "n_ref_all": 106, "n_fig": 6, "n_tab": 2, "L_tab": 956, "n_element_tab": 29, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 456, "n_element_tab_1": 22, "formula_len_all": 116, "formula_len_all_1": 116, "len_all": 107275, "len_all_1": 58851, "len_abs": 1439, "len_title": 115, "len_sents": 29259, "len_sents_1": 29484, "n_sents": 192, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1447, "title": "Zero-Shot Image Compression with Diffusion-Based Posterior Sampling", "abs": "Diffusion models dominate the field of image generation, however they have yet to make major breakthroughs in the field of image compression. Indeed, while pre-trained diffusion models have been successfully adapted to a wide variety of downstream tasks, \nexisting work in diffusion-based image compression require task specific model training, which can be both cumbersome and limiting. This work addresses this gap by harnessing the image prior learned by existing pre-trained diffusion models for solving the task of lossy image compression. This enables the use of the wide variety of publicly-available models, and avoids the need for training or fine-tuning. Our method, PSC (Posterior Sampling-based Compression), utilizes zero-shot diffusion-based posterior samplers. It does so through a novel sequential process inspired by the active acquisition technique \"Adasense\" to accumulate informative measurements of the image. This strategy minimizes uncertainty in the reconstructed image and allows for construction of an image-adaptive transform coordinated between both the encoder and decoder. PSC offers a progressive compression scheme that is both practical and simple to implement. Despite minimal tuning, and a simple quantization and entropy coding, PSC achieves competitive results compared to established methods, paving the way for further exploration of pre-trained diffusion models and posterior samplers for image compression.", "keywords": [ "zeroshot", "compression", "diffusion models", "posterior sampling" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "qi5dkmEE91", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Beepul Bharti", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Gabriele Scalia", "gender": "unknown", "institution": "Genentech", "country": "", "position": "Researcher" }, { "name": "Alex Tseng", "gender": "Non-Binary", "institution": "Genentech", "country": "", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 19, "n_ref_uni": 41, "n_ref": 61, "n_ref_all": 76, "n_fig": 11, "n_tab": 1, "L_tab": 1102, "n_element_tab": 14, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1190, "formula_len_all_1": 1048, "len_all": 111192, "len_all_1": 64383, "len_abs": 1095, "len_title": 129, "len_sents": 34371, "len_sents_1": 31020, "n_sents": 262, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1100, "title": "Uncovering BioLOGICAL Motifs and Syntax via Sufficient and Necessary Explanations", "abs": "In recent years, deep neural networks (DNNs) have excelled at learning from high-throughput genome-profiling experiments to predict transcription factor (TF) binding. TF binding is driven by sequence motifs, and explaining how and why DNNs make accurate predictions could help identify these motifs, as well as their logical syntax. However, the black-box nature of DNNs makes interpretation difficult. Most post-hoc methods evaluate the importance of each base pair in isolation, often resulting in noise since they overlook the fact that motifs are contiguous regions. Additionally, these methods fail to capture the complex interactions between different motifs. To address these challenges, we propose Motif Explainer Models (MEMs), a novel explanation method that uses sufficiency and necessity to identify important motifs and their syntax. MEMs excel at identifying multiple disjoint motifs across DNA sequences, overcoming limitations of existing methods. Moreover, by accurately pinpointing sufficient and necessary motifs, MEMs can reveal the logical syntax that governs genomic regulation.", "keywords": [ "interpretability", "attributions", "computational biology" ], "rating_list": [ 1, 3, 5 ], "soundness_list": [ 1, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "qhfZL46nPV", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Helge Langseth", "gender": "Male", "institution": "Norwegian University of Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Yanzhe Bekkemoen", "gender": "unknown", "institution": "Norwegian University of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 24, "n_ref": 48, "n_ref_all": 84, "n_fig": 17, "n_tab": 2, "L_tab": 980, "n_element_tab": 22, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 1382, "n_element_tab_1": 87, "formula_len_all": 184, "formula_len_all_1": 366, "len_all": 113383, "len_all_1": 57812, "len_abs": 1637, "len_title": 112, "len_sents": 30770, "len_sents_1": 25065, "n_sents": 297, "n_sents_1": 229, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1647, "title": "Outcome-based Semifactual Explanation For Reinforcement Learning", "abs": "Counterfactual explanations in reinforcement learning (RL) aim to answer what-if questions by showing sparse and minimal changes to states, which results in the probability mass moving from one action to another. Although these explanations are effective in classification tasks that look for the presence of concepts, RL brings new challenges that current counterfactual methods for RL still need to solve. These challenges include defining similarity in RL, out-of-distribution states, and lack of discriminative power. Given a state of interest called the query state, we solve these problems by asking how long the agent can execute the query state action without incurring a negative outcome regarding the expected return. We coin this outcome-based semifactual (OSF) explanation and find the OSF state by simulating trajectories from the query state. The last state in a subtrajectory where we can take the same action as in the query state without incurring a negative outcome is the OSF state. This state is discriminative, plausible, and similar to the query state. It abstracts away unimportant action switching with little explanatory value and shows the boundary between positive and negative outcomes. Qualitatively, we show that our method explains when it is necessary to switch actions. As a result, it is easier to understand the agent's behavior. Quantitatively, we demonstrate that our method can increase policy performance and, at the same time, reduce how often the agent switches its action across six environments. The code and trained models are available at https://anonymous.4open.science/r/osf-explanation-for-rl-E312/.", "keywords": [ "Explainable Reinforcement Learning", "Interpretability", "Deep Reinforcement Learning", "Policy Explanation" ], "rating_list": [ 1, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "qh57QGETn1", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Antonio Torralba", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "David Bau", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Assistant Professor" }, { "name": "Hui Ren", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "Undergrad student" }, { "name": "Joanna Materzynska", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Rohit Gandikota", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 38, "n_ref": 55, "n_ref_all": 84, "n_fig": 21, "n_tab": 4, "L_tab": 581, "n_element_tab": 103, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 424, "n_element_tab_1": 61, "formula_len_all": 278, "formula_len_all_1": 263, "len_all": 126494, "len_all_1": 58524, "len_abs": 657, "len_title": 131, "len_sents": 36748, "len_sents_1": 28268, "n_sents": 266, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 659, "title": "Art-Free Generative Models: Exploring Art Creation Without Prior Artistic Knowledge", "abs": "In this work, we explore the question: \u201cHow much prior art knowledge is needed to create art?\u201d To investigate this, we propose a text-to-image generation model trained without access to art-related content. We then introduce a simple yet effective method to learn an art adaptor using only a few examples of selected artistic styles. Our experiments show that art generated using our method is perceived by users as comparable to art produced by models trained on large, art-rich datasets. Finally, through data attribution techniques, we illustrate how examples from both artistic and non-artistic datasets contributed to the creation of new artistic styles.", "keywords": [ "generative models; art; generative art;" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "qh1goDZ0ZQ", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Ang Li", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Daize Dong", "gender": "Male", "institution": "Zhipu AI", "country": "CN", "position": "Researcher" }, { "name": "LIANG DING", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Shwai He", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 41, "n_ref": 74, "n_ref_all": 98, "n_fig": 12, "n_tab": 5, "L_tab": 4139, "n_element_tab": 647, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 2778, "n_element_tab_1": 345, "formula_len_all": 682, "formula_len_all_1": 440, "len_all": 161990, "len_all_1": 60086, "len_abs": 1694, "len_title": 129, "len_sents": 35878, "len_sents_1": 24931, "n_sents": 249, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1703, "title": "Towards Efficient Mixture of Experts: A Holistic Study of Compression Techniques", "abs": "Scaling large language models has driven remarkable advancements across various\ndomains, yet the continual increase in model size presents significant challenges\nfor real-world deployment. The Mixture of Experts (MoE) architecture offers a\npromising solution by dynamically selecting and activating only a subset of experts\nduring inference, thus substantially reducing computational costs while preserving\nhigh performance. Despite these benefits, MoE introduces new inefficiencies, such\nas excessive parameters and communication overhead. In this work, we present\na holistic study on compression techniques of Mixture of Experts to enhance\nboth efficiency and scalability. While recent efforts have focused on reducing the\nnumber of experts, these approaches still suffer from considerable communication\nand computational costs. To address this, we propose more aggressive strategies,\nsuch as Layer Drop, which removes entire MoE layers, and Block Drop, which\neliminates transformer blocks. Surprisingly, these aggressive structure pruning\ntechniques not only preserve model performance but also substantially improve\nefficiency. Additionally, beyond Expert Trimming, we also introduce Expert\nSlimming, which compresses individual experts to further boost performance and\ncan be seamlessly integrated with Expert Trimming. Extensive experimental results\ndemonstrate the effectiveness of our proposed methods \u2014 Layer Drop and Block\nDrop \u2014 along with the comprehensive recipe that integrates Expert Slimming and\nExpert Trimming, achieving a 6.05\u00d7 speedup with 77.1% reduced memory usage\nwhile maintaining over 92% of performance on Mixtral-8\u00d77B. Our code will be\nmade publicly available upon acceptance.", "keywords": [ "Mixture of Experts", "Model Compression" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "qgsXsqahMq", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Chonghe Jiang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Dingyi Zhuang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jinhua Zhao", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Shenhao Wang", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Yunhan Zheng", "gender": "unknown", "institution": "Singapore-MIT Alliance for Research and Technology Centre", "country": "SG", "position": "Postdoc" } ], "n_formula": 27, "n_formula_1": 17, "n_ref_uni": 48, "n_ref": 90, "n_ref_all": 141, "n_fig": 5, "n_tab": 13, "L_tab": 5939, "n_element_tab": 634, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3654, "n_element_tab_1": 287, "formula_len_all": 1310, "formula_len_all_1": 806, "len_all": 164080, "len_all_1": 68282, "len_abs": 1259, "len_title": 123, "len_sents": 46284, "len_sents_1": 28653, "n_sents": 363, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 75, "L_abs": 1198, "title": "GETS: Ensemble Temperature Scaling for Calibration in Graph Neural Networks", "abs": "Graph Neural Networks (GNNs) deliver strong classification results but often suffer from poor calibration performance, leading to overconfidence or underconfidence. This is particularly problematic in high-stakes applications where accurate uncertainty estimates are essential. Existing post-hoc methods, such as temperature scaling, fail to effectively utilize graph structures, while current GNN calibration methods often overlook the potential of leveraging diverse input information and model ensembles jointly. In the paper, we propose Graph Ensemble Temperature Scaling (GETS), a novel calibration framework that combines input and model ensemble strategies within a Graph Mixture-of-Experts (MoE) architecture. GETS integrates diverse inputs, including logits, node features, and degree embeddings, and adaptively selects the most relevant experts for each node\u2019s calibration procedure. Our method outperforms state-of-the-art calibration techniques, reducing expected calibration error (ECE) by $\\geq$ 25% across 10 GNN benchmark datasets. Additionally, GETS is computationally efficient, scalable, and capable of selecting effective input combinations for improved calibration performance.", "keywords": [ "Uncertainty Quantification; Graph Neural Networks; Ensemble Learning; Mixture of Experts" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "qg9BBAXAHN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Yaxin Du", "gender": "Female", "institution": "Shanghai Jiao tong University", "country": "CN", "position": "PhD student" }, { "name": "Zezhou Wang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Zhuzhong Qian", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "Full Professor" }, { "name": "Siheng Chen", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 16, "n_formula_1": 6, "n_ref_uni": 28, "n_ref": 87, "n_ref_all": 128, "n_fig": 11, "n_tab": 13, "L_tab": 3424, "n_element_tab": 377, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1043, "formula_len_all_1": 408, "len_all": 208011, "len_all_1": 63770, "len_abs": 1329, "len_title": 118, "len_sents": 63144, "len_sents_1": 30304, "n_sents": 487, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1343, "title": "Federated Instruction Tuning of LLMs with Domain Coverage Augmentation", "abs": "Federated Domain-specific Instruction Tuning (FedDIT) utilizes limited cross-client private data together with server-side public data for instruction augmentation, ultimately boosting model performance within specific domains. To date, the factors affecting FedDIT remain unclear, and existing instruction augmentation methods primarily focus on the centralized setting without considering distributed environments. Our experiments reveal that the cross-client domain coverage, rather than data heterogeneity, drives model performance in FedDIT. In response, we propose FedDCA, which optimizes domain coverage through greedy client center selection and retrieval-based augmentation. For client-side computational efficiency and system scalability, FedDCA$^*$, the variant of FedDCA, utilizes heterogeneous encoders with server-side feature alignment. Extensive experiments across four distinct domains (code, medical, financial, and mathematical) substantiate the effectiveness of both methods. Additionally, we investigate privacy preservation against memory extraction attacks utilizing various amounts of public data. Results show that there is no significant correlation between the volume of public data and the privacy-preserving capability. However, as the fine-tuning rounds increase, the risk of privacy leakage reduces or converges.", "keywords": [ "Federated Learning", "Large Language Model", "Instruction Tuning" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "qfU5S4cddQ", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Makoto Takamoto", "gender": "Male", "institution": "NEC", "country": "", "position": "Researcher" }, { "name": "Mathias Niepert", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "", "position": "Full Professor" }, { "name": "Viktor Zaverkin", "gender": "Male", "institution": "NEC Laboratories Europe", "country": "DE", "position": "Researcher" } ], "n_formula": 25, "n_formula_1": 9, "n_ref_uni": 52, "n_ref": 142, "n_ref_all": 234, "n_fig": 10, "n_tab": 21, "L_tab": 9484, "n_element_tab": 194, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2035, "n_element_tab_1": 28, "formula_len_all": 1937, "formula_len_all_1": 545, "len_all": 240593, "len_all_1": 72106, "len_abs": 1541, "len_title": 118, "len_sents": 69460, "len_sents_1": 32186, "n_sents": 546, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1561, "title": "Physics-Informed Weakly Supervised Learning for Interatomic Potentials", "abs": "Machine learning plays an increasingly important role in computational chemistry and materials science, complementing computationally intensive ab initio and first-principles methods. Despite their utility, machine-learning models often lack generalization capability and robustness during atomistic simulations, yielding unphysical energy and force predictions that hinder their real-world applications. We address this challenge by introducing a physics-informed, weakly supervised approach for training machine-learned interatomic potentials (MLIPs). We introduce two novel loss functions, extrapolating the potential energy via a Taylor expansion and using the concept of conservative forces. Our approach improves the accuracy of MLIPs applied to training tasks with sparse training data sets and reduces the need for pre-training computationally demanding models with large data sets. Particularly, we perform extensive experiments demonstrating reduced energy and force errors---often lower by a factor of two---for various baseline models and benchmark data sets. Moreover, we demonstrate improved robustness during MD simulations of the MLIP models trained with the proposed weakly supervised loss.\n Finally, we show that our approach facilitates MLIPs' training in a setting where the computation of forces is infeasible at the reference level, such as those employing complete-basis-set extrapolation. An implementation of our method and scripts for executing experiments are available at \\url{https://anonymous.4open.science/r/PICPS-ML4Sci-1E8F}.", "keywords": [ "machine learning interatomic potential", "weakly supervised learning", "machine learning for science" ], "rating_list": [ 6, 8, 3, 3 ], "soundness_list": [ 4, 3, 2, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "qezVbskHmi", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Adiba Orzikulova", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jaehyun Kwak", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Jaemin Shin", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Sung-Ju Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 20, "n_ref_uni": 49, "n_ref": 96, "n_ref_all": 112, "n_fig": 7, "n_tab": 10, "L_tab": 5556, "n_element_tab": 397, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 841, "n_element_tab_1": 79, "formula_len_all": 1317, "formula_len_all_1": 1220, "len_all": 165577, "len_all_1": 60990, "len_abs": 1256, "len_title": 129, "len_sents": 45733, "len_sents_1": 26165, "n_sents": 315, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1287, "title": "Federated Learning for Time-Series Healthcare Sensing with Incomplete Modalities", "abs": "Many healthcare sensing applications utilize multimodal time-series data from sensors embedded in mobile and wearable devices. Federated Learning (FL), with its privacy-preserving advantages, is particularly well-suited to such applications. However, most multimodal FL methods assume the availability of complete modality data for local training, which is often unrealistic. Moreover, recent approaches for tackling incomplete modalities scale poorly and become inefficient as the number of modalities increases. To address these limitations, we propose FLISM, an innovative algorithm that enables efficient FL training with incomplete sensing modalities while maintaining high accuracy. FLISM employs three key techniques: (1) modality-invariant representation learning to extract effective features from clients with a diverse set of modalities, (2) modality quality-aware aggregation to prioritize contributions from clients with higher-quality data, and (3) global-aligned knowledge distillation to mitigate local update shifts caused by modality differences. Extensive experiments on real-world datasets show that FLISM not only achieves high accuracy but is also faster and more efficient compared to state-of-the-art methods designed to handle incomplete modality problems in FL.", "keywords": [ "federated learning", "incomplete modalities", "time-series healthcare sensing" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 2, 2, 4 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "qeYa5LRveW", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Jamie Callan", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Luyu Gao", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Yunyi Zhang", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 50, "n_ref": 97, "n_ref_all": 109, "n_fig": 4, "n_tab": 2, "L_tab": 481, "n_element_tab": 36, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 745, "n_element_tab_1": 11, "formula_len_all": 91, "formula_len_all_1": 91, "len_all": 340623, "len_all_1": 58856, "len_abs": 1360, "len_title": 110, "len_sents": 32512, "len_sents_1": 28108, "n_sents": 230, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1366, "title": "ACER: Automatic Language Model Context Extension via Retrieval", "abs": "Long-context modeling is one of the critical capabilities of language AI for digesting and reasoning over complex information pieces. In practice, long-context capabilities are typically built into a pre-trained language model (LM) through a carefully designed context extension stage, with the goal of producing generalist long-context capabilities. In our preliminary experiments, however, we discovered that the current open-weight generalist long-context models are still lacking in practical long-context processing tasks. While this means perfectly effective long-context modeling demands task-specific data, the cost can be prohibitive. In this paper, we draw inspiration from how humans process a large body of information: a lossy retrieval stage ranks a large set of documents while the reader ends up reading deeply only the top candidates. We build an automatic data synthesis pipeline that mimics this process using short-context LMs. The short-context LMs are further tuned using these self-generated data to obtain task-specific long-context capabilities. Similar to how pre-training learns from imperfect data, we hypothesize and further demonstrate that the short-context model can bootstrap over the synthetic data, outperforming not only long-context generalist models but also the retrieval and read pipeline used to synthesize the training data.", "keywords": [ "language", "retrieval", "long context understanding", "text generation" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "qeY25DwmKO", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 5, "author_info_list": [ { "name": "Chris Cameron", "gender": "unknown", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Jason Hartford", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "Assistant Professor" }, { "name": "Kevin Leyton-Brown", "gender": "Not Specified", "institution": "University of British Columbia", "country": "CA", "position": "Full Professor" }, { "name": "Taylor Lundy", "gender": "Male", "institution": "University of British Columbia", "country": "", "position": "PhD student" }, { "name": "xing jin", "gender": "unknown", "institution": "University of British Columbia", "country": "CA", "position": "Intern" } ], "n_formula": 7, "n_formula_1": 1, "n_ref_uni": 38, "n_ref": 68, "n_ref_all": 85, "n_fig": 5, "n_tab": 6, "L_tab": 2534, "n_element_tab": 215, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 902, "n_element_tab_1": 119, "formula_len_all": 794, "formula_len_all_1": 40, "len_all": 132572, "len_all_1": 61231, "len_abs": 1119, "len_title": 83, "len_sents": 40750, "len_sents_1": 29593, "n_sents": 292, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 932, "title": "Foundation Models for Boolean Logic", "abs": "Boolean logic is fundamental to solving various computational problems, such as Boolean satisfiability (SAT) and model counting, but existing machine learning (ML) approaches for automating algorithm design are computationally expensive and data-intensive. We propose the first foundation model for Boolean logic, leveraging a multi-task dataset of one million instances spanning sixteen tasks and using graph neural networks (GNNs). We evaluated the generalization of the foundation models on held-out tasks; we found that models fine-tuned from the foundation model were substantially more sample efficient and converged much faster than models trained from scratch. We identified a number of crucial design components for training these models, in particular the choice of normalization layer. We showed that a hybrid of different normalization techniques across layers is much more effective than any single normalization layer.", "keywords": [ "Boolean logic", "runtime prediction", "graph neural networks", "multi-task learning", "foundation models" ], "rating_list": [ 6, 8, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "qeXcMutEZY", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Alex Dimakis", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department, University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Asad Aali", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Research Scientist" }, { "name": "Brett Levac", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Ioannis Daras", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "PhD student" }, { "name": "Sidharth Kumar", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" }, { "name": "Jonathan Tamir", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 43, "n_ref": 103, "n_ref_all": 117, "n_fig": 6, "n_tab": 4, "L_tab": 1311, "n_element_tab": 189, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 890, "n_element_tab_1": 90, "formula_len_all": 1195, "formula_len_all_1": 718, "len_all": 176623, "len_all_1": 54473, "len_abs": 1103, "len_title": 158, "len_sents": 39392, "len_sents_1": 24064, "n_sents": 343, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 110, "L_abs": 1119, "title": "Ambient Diffusion Posterior Sampling: Solving Inverse Problems with Diffusion Models Trained on Corrupted Data", "abs": "We provide a framework for solving inverse problems with diffusion models learned from linearly corrupted data. Firstly, we extend the Ambient Diffusion framework to enable training directly from measurements corrupted in the Fourier domain. Subsequently, we train diffusion models for MRI with access only to Fourier subsampled multi-coil measurements at acceleration factors R$=2, 4, 6, 8$. Secondly, we propose $\\textit{Ambient Diffusion Posterior Sampling}$ (A-DPS), a reconstruction algorithm that leverages generative models pre-trained on one type of corruption (e.g. image inpainting) to perform posterior sampling on measurements from a different forward process (e.g. image blurring). For MRI reconstruction in high acceleration regimes, we observe that A-DPS models trained on subsampled data are better suited to solving inverse problems than models trained on fully sampled data. We also test the efficacy of A-DPS on natural image datasets (CelebA, FFHQ, and AFHQ) and show that A-DPS can sometimes outperform models trained on clean data for several image restoration tasks in both speed and performance.", "keywords": [ "inverse problems", "diffusion", "ambient diffusion", "mri" ], "rating_list": [ 8, 8, 8, 5, 6 ], "soundness_list": [ 4, 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 3, 4 ], "contribution_list": [ 4, 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 5, 5, 4 ] }, { "paper_id": "qe1CsfnN1W", "primary_area": "causal reasoning", "n_author": 6, "author_info_list": [ { "name": "Jing Ma", "gender": "Female", "institution": "Case Western Reserve University", "country": "US", "position": "Assistant Professor" }, { "name": "Jundong Li", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" }, { "name": "Liang Wu", "gender": "Male", "institution": "LinkedIn", "country": "", "position": "Applied Scientist" }, { "name": "Liangjie Hong", "gender": "Male", "institution": "LinkedIn", "country": "", "position": "Researcher" }, { "name": "Qi Guo", "gender": "Male", "institution": "LinkedIn", "country": "", "position": "Researcher" }, { "name": "Yaochen Zhu", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "", "position": "PhD student" } ], "n_formula": 63, "n_formula_1": 27, "n_ref_uni": 40, "n_ref": 72, "n_ref_all": 125, "n_fig": 7, "n_tab": 1, "L_tab": 949, "n_element_tab": 50, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7670, "formula_len_all_1": 2286, "len_all": 160046, "len_all_1": 70144, "len_abs": 1759, "len_title": 131, "len_sents": 53703, "len_sents_1": 34005, "n_sents": 387, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1952, "title": "Causal Effect Estimation with Mixed Latent Confounders and Post-treatment Variables", "abs": "Causal inference from observational data has attracted considerable attention among researchers. One main obstacle to drawing valid causal conclusions is handling of confounders. As the direct measurement of confounders may not always be feasible, recent methods seek to address the confounding bias via proxy variables, i.e., variables postulated to be causally related to unobserved confounders. However, the selected proxies may scramble both latent confounders and latent post-treatment variables in practice, where existing methods risk biasing the estimation by unintentionally controlling for variables affected by the treatment. In this paper, we systematically investigate the bias of latent post-treatment variables, i.e., latent post-treatment bias, in causal effect estimation. We first derive the bias of existing covariate adjustment-based methods when selected proxies scramble both latent confounders and latent post-treatment variables, which we demonstrate can be arbitrarily bad. We then propose a novel Confounder-identifiable VAE (CiVAE) to address the bias. CiVAE is built upon a mild assumption that the prior of latent variables that generate the proxy belongs to a general exponential family with at least one invertible sufficient statistic in the factorized part. Based on this assumption, we show that latent confounders and latent post-treatment variables can be individually identified up to simple bijective transformations. We then prove that with individual identification, the intractable disentanglement problem of latent confounders and post-treatment variables can be transformed to a tractable conditional independence test problem. Finally, we prove that the true causal effects can be unbiasedly estimated with transformed confounders inferred by CiVAE. Experiments on both simulated and real-world datasets demonstrate that CiVAE is significantly more robust to latent post-treatment bias than existing methods.", "keywords": [ "causal effect estimation", "latent post-treatment bias", "identifiable VAE" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "qdbluGtEpL", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Carl Yang", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Assistant Professor" }, { "name": "Haichuan Zhang", "gender": "unknown", "institution": "Sichuan University", "country": "CN", "position": "Intern" }, { "name": "Meiyu Lin", "gender": "unknown", "institution": "Sichuan University", "country": "CN", "position": "MS student" }, { "name": "Mingjie Tang", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Renyuan Li", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhaoyi Liu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Intern" }, { "name": "Zhiyuan Cheng", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 8, "n_ref": 24, "n_ref_all": 58, "n_fig": 11, "n_tab": 7, "L_tab": 3887, "n_element_tab": 618, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2607, "n_element_tab_1": 460, "formula_len_all": 571, "formula_len_all_1": 490, "len_all": 148460, "len_all_1": 62427, "len_abs": 3490, "len_title": 126, "len_sents": 42999, "len_sents_1": 26508, "n_sents": 335, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1609, "title": "Attack as Defense: Run-time Backdoor Implantation for Image Content Protection", "abs": "As generative models achieve great success, tampering and modifying the sensitive image contents (i.e., human faces, artist signatures, commercial logos, etc.) have induced a significant threat with social impact. \nThe backdoor attack is a method that implants vulnerabilities in a target model, which can be activated through a trigger.\nIn this work, we innovatively prevent the abuse of image content modification by implanting the backdoor into image-editing models. Once the protected sensitive content on an image is modified by an editing model, the backdoor will be triggered, making the editing fail. \nUnlike traditional backdoor attacks that use data poisoning, to enable protection on individual images and eliminate the need for model training, we developed the first framework for run-time backdoor implantation, which is both time- and resource- efficient. We generate imperceptible perturbations on the images to inject the backdoor and define the protected area as the only backdoor trigger. Editing other unprotected insensitive areas will not trigger the backdoor, which minimizes the negative impact on legal image modifications. Evaluations with state-of-the-art image editing models show that our protective method can increase the CLIP-FID of generated images from 12.72 to 39.91, or reduce the SSIM from 0.503 to 0.167 when subjected to malicious editing. At the same time, our method exhibits minimal impact on benign editing, which demonstrates the efficacy of our proposed framework. The proposed run-time backdoor can also achieve effective protection on the latest diffusion models.", "keywords": [ "image editing", "backdoor attack", "image inpainting" ], "rating_list": [ 6, 5, 5, 8, 5 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 1, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "qdOIkeZ5e4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Li Li", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Lina Wei", "gender": "Female", "institution": "Hangzhou City University ", "country": "CN", "position": "Associate Professor" }, { "name": "Qilong Wu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Roger Zimmermann", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Wei Ji", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Yicong Li", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "CaiPengCheng", "gender": "unknown", "institution": "\u6d59\u5927\u57ce\u5e02\u5b66\u9662", "country": "CN", "position": "Undergrad student" }, { "name": "YOU QIN", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 20, "n_ref": 38, "n_ref_all": 55, "n_fig": 6, "n_tab": 6, "L_tab": 2062, "n_element_tab": 258, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 587, "n_element_tab_1": 79, "formula_len_all": 621, "formula_len_all_1": 314, "len_all": 126682, "len_all_1": 59918, "len_abs": 1134, "len_title": 82, "len_sents": 40274, "len_sents_1": 27449, "n_sents": 300, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 34, "L_abs": 1138, "title": "Generalized Video Moment Retrieval", "abs": "In this paper, we introduce the Generalized Video Moment Retrieval (GVMR) framework, which extends traditional Video Moment Retrieval (VMR) to handle a wider range of query types. Unlike conventional VMR systems, which are often limited to simple, single-target queries, GVMR accommodates both non-target and multi-target queries. To support this expanded task, we present the NExT-VMR dataset, derived from the YFCC100M collection, featuring diverse query scenarios to enable more robust model evaluation.\nAdditionally, we propose BCANet, a transformer-based model incorporating the novel Boundary-aware Cross Attention (BCA) module. The BCA module enhances boundary detection and uses cross-attention to achieve a comprehensive understanding of video content in relation to queries. BCANet accurately predicts temporal video segments based on natural language descriptions, outperforming traditional models in both accuracy and adaptability. Our results demonstrate the potential of the GVMR framework, the NExT-VMR dataset, and BCANet to advance VMR systems, setting a new standard for future multimedia information retrieval research.", "keywords": [ "video moment retrieval", "diverse query types", "model versatility" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "qcyn7ESaM8", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Haiping Lu", "gender": "Male", "institution": "University of Sheffield", "country": "", "position": "Full Professor" }, { "name": "Pawel Pukowski", "gender": "Male", "institution": "University of Sheffield", "country": "", "position": "PhD student" }, { "name": "Venet Osmani", "gender": "Male", "institution": "Queen Mary, University of London", "country": "GB", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 29, "n_ref": 61, "n_ref_all": 74, "n_fig": 7, "n_tab": 2, "L_tab": 574, "n_element_tab": 45, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1803, "n_element_tab_1": 58, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 91253, "len_all_1": 54805, "len_abs": 1250, "len_title": 111, "len_sents": 29917, "len_sents_1": 25760, "n_sents": 192, "n_sents_1": 162, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1256, "title": "Bridging PCA and Neural Networks: New Insights into Class Bias", "abs": "Understanding class-level hardness is essential for addressing class bias in machine learning. Traditionally, class bias has been explored with two primary approaches: analyzing raw input data to improve preprocessing strategies or examining neural network latent representations to refine model training. In this work, we find that PCA-transformed spaces\u2014despite being produced through linear transformations\u2014still contain substantial information about class-level hardness. This suggests that, despite their distinct goals and methodologies, both PCA and neural networks may encode similar features related to class bias, offering new insights into the nature of class bias and how data representations are formed in both PCA and neural networks. \n\nAnalyzing class bias commonly involves Pearson Correlation, which assumes stable inputs. However, we find that class bias is a highly unstable phenomenon with respect to variables such as training time and model initialization, with class-level variability often exceeding the differences between classes. Together with increased variability in class accuracies over dataset-level ones, this suggests that current methods for addressing dataset-level variability may be inadequate for handling class bias.", "keywords": [ "hardness", "class bias", "class-level hardness identifiers", "instance-level hardness identifiers" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 1, 1, 2, 3 ], "contribution_list": [ 3, 1, 2, 1 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "qbSoiHLEK0", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Aleksandr Tsymbalov", "gender": "unknown", "institution": "Higher School of Economics, Higher School of Economics", "country": "RU", "position": "Undergrad student" }, { "name": "Andrey Savchenko", "gender": "Male", "institution": "Sber AI Lab", "country": "RU", "position": "Principal Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 77, "n_ref_all": 97, "n_fig": 1, "n_tab": 9, "L_tab": 7592, "n_element_tab": 1043, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 4989, "n_element_tab_1": 838, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 116364, "len_all_1": 63969, "len_abs": 893, "len_title": 149, "len_sents": 27613, "len_sents_1": 23637, "n_sents": 184, "n_sents_1": 154, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 905, "title": "LLM2Features: Large Language Models in Interpretable Feature Generation for AutoML with Tabular Data", "abs": "Automatic Machine Learning (AutoML) is the popular supervised learning approach for tabular data. One of its key components is generating the most suitable features given the available training dataset. To overcome the disadvantages of existing automatic feature generation techniques, such as lack of generality and interpretability, we propose the novel approach, \\textbf{LLM2Features}. It uses LLMs (Large Language Models) to generate meaningful features using automatically collected statistics about the dataset without explicitly describing the data, making it ideal for implementing in AutoML frameworks. In particular, we introduce the LLM-based critic that additionally verifies the presence of syntax or logical errors. The experimental study demonstrates the benefits of the proposed LLM2Features approach in accuracy and training time compared to the state-of-the-art feature generation tools.", "keywords": [ "LLM", "Auto feature generation", "GPT-4o", "GPT-o1", "Tabular Data" ], "rating_list": [ 1, 1, 5, 5 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 1, 1, 2, 2 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "qb2QRoE4W3", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Ankur Taly", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Dharma Muppalla", "gender": "unknown", "institution": "Google", "country": "US", "position": "Software Engineer" }, { "name": "Nitish Joshi", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 40, "n_ref": 60, "n_ref_all": 104, "n_fig": 4, "n_tab": 2, "L_tab": 349, "n_element_tab": 34, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 396, "n_element_tab_1": 36, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 149628, "len_all_1": 66029, "len_abs": 1305, "len_title": 118, "len_sents": 39151, "len_sents_1": 31407, "n_sents": 349, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1315, "title": "LLM-Cite: Cheap Fact Verification with Attribution via URL Generation", "abs": "Hallucinations are one of the main issues with Large Language Models (LLMs). This has led to increased interest in automated ways to verify the factuality of LLMs' responses. Existing methods either rely on: (a) search over a knowledge base (KB), which is costly especially if the KB must be updated frequently to keep up with fresh content, (b) LLM's parametric knowledge to fact-check claims, which is cheaper but does not give attribution and is limited to verifying claims related to knowledge acquired during pretraining. In this work, we present LLM-Cite, a cheap and easy to implement method that does not rely on any external search system while still providing attribution and the ability to verify fresh claims. Our key insight is to leverage an LLM to directly generate potential citation URLs for a given claim, and then use entailment checks to verify the claim against content of the URLs (which are fetched on-the-fly). We benchmark LLM-Cite on three datasets containing fresh and non-fresh claims generated by humans and models. We show that LLM-Cite performs comparable or better than existing methods on all categories of claims --- importantly, without sacrificing attribution, or requiring costly external search --- overall LLM-Cite is more than 45x cheaper than a Google Search based approach.", "keywords": [ "Fact Verification", "Attribution", "Citation", "Factuality" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "qazJfAmgOt", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chee Seng Chan", "gender": "Male", "institution": "Universiti Malaya", "country": "MY", "position": "Full Professor" }, { "name": "Hanlin Gu", "gender": "Male", "institution": "webank", "country": "CN", "position": "Researcher" }, { "name": "Lixin Fan", "gender": "Male", "institution": "WeBank", "country": "CN", "position": "Principal Researcher" }, { "name": "Tae Hong Xi", "gender": "Male", "institution": "Universiti Malaya", "country": "MY", "position": "MS student" } ], "n_formula": 24, "n_formula_1": 20, "n_ref_uni": 19, "n_ref": 53, "n_ref_all": 78, "n_fig": 8, "n_tab": 15, "L_tab": 5382, "n_element_tab": 411, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2931, "n_element_tab_1": 228, "formula_len_all": 1806, "formula_len_all_1": 971, "len_all": 149842, "len_all_1": 61674, "len_abs": 2604, "len_title": 106, "len_sents": 35278, "len_sents_1": 23582, "n_sents": 303, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1110, "title": "A few-shot Label Unlearning in Vertical Federated Learning", "abs": "This paper addresses the critical challenge of unlearning in Vertical Federated Learning (VFL), an area that has received limited attention compared to horizontal federated learning. We introduce the first approach specifically designed to tackle label unlearning in VFL, focusing on scenarios where the active party aims to mitigate the risk of label leakage. Our method leverages a limited amount of labeled data, utilizing manifold mixup to augment the forward embedding of insufficient data, followed by gradient ascent on the augmented embeddings to erase label information from the models. This combination of augmentation and gradient ascent enables high unlearning effectiveness while maintaining efficiency, completing the unlearning procedure within seconds. Extensive experiments conducted on diverse datasets, including MNIST, CIFAR10, CIFAR100, and ModelNet, validate the efficacy and scalability of our approach. This work represents a significant advancement in federated learning, addressing the unique challenges of unlearning in VFL while preserving both privacy and computational efficiency.", "keywords": [ "Federated Unlearning", "Vertical Federated Learning", "Machine Unlearning" ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "qawqxu4MgA", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Alireza Nadali", "gender": "Not Specified", "institution": "University of Colorado at Boulder", "country": "US", "position": "PhD student" }, { "name": "Ashutosh Trivedi", "gender": "Male", "institution": "University of Colorado at Boulder", "country": "", "position": "Associate Professor" }, { "name": "Bingzhuo Zhong", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Majid Zamani", "gender": "Male", "institution": "University of Colorado at Boulder", "country": "US", "position": "Associate Professor" } ], "n_formula": 35, "n_formula_1": 25, "n_ref_uni": 38, "n_ref": 64, "n_ref_all": 80, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4390, "formula_len_all_1": 3160, "len_all": 111843, "len_all_1": 64226, "len_abs": 1858, "len_title": 117, "len_sents": 33097, "len_sents_1": 28574, "n_sents": 254, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1874, "title": "Transfer Learning for Control Systems via Neural Simulation Relations", "abs": "Transfer learning is an umbrella term for machine learning approaches that leverage knowledge gained from solving one problem (the source domain) to improve speed, efficiency, and data requirements in solving a different but related problem (the target domain). \nThe performance of the transferred model in the target domain is typically measured via some notion of loss function in the target domain. \nThis paper focuses on effectively transferring control logic from a source control system to a target control system while providing approximately similar behavioral guarantees in both domains. \nHowever, in the absence of a complete characterization of behavioral specifications, this problem cannot be captured in terms of loss functions. \nTo overcome this challenge, we use (approximate) simulation relations to characterize observational equivalence between the behaviors of two systems.\n\nSimulation relations ensure that the outputs of both systems, equipped with their corresponding controllers, remain close to each other over time, and their closeness can be quantified a priori. \nBy parameterizing simulation relations with neural networks, we introduce the notion of neural simulation relations, which provides a data-driven approach to transfer any synthesized controller, regardless of the specification of interest, along with its proof of correctness. \nCompared with prior approaches, our method eliminates the need for a closed-loop mathematical model and specific requirements for both the source and target systems. \nWe also introduce validity conditions that, when satisfied, guarantee the closeness of the outputs of two systems equipped with their corresponding controllers, thus eliminating the need for post-facto verification. \nWe demonstrate the effectiveness of our approach through case studies involving a vehicle and a double inverted pendulum.", "keywords": [ "Transfer Learning; Conrol Systems; Neural Simulation Relations" ], "rating_list": [ 5, 1, 5, 5 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "qak1NNI5yO", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Jilles Vreeken", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Tenured Faculty" }, { "name": "Linara Adilova", "gender": "Female", "institution": "Ruhr-Universt\u00e4t Bochum", "country": "", "position": "PhD student" }, { "name": "Michael Kamp", "gender": "Male", "institution": "Institute for AI in Medicine IKIM", "country": "", "position": "Research Group Leader" }, { "name": "Nils Philipp Walter", "gender": "Male", "institution": "CISPA, saarland university, saarland informatics campus", "country": "DE", "position": "PhD student" } ], "n_formula": 34, "n_formula_1": 13, "n_ref_uni": 64, "n_ref": 111, "n_ref_all": 130, "n_fig": 12, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5144, "formula_len_all_1": 823, "len_all": 168489, "len_all_1": 61040, "len_abs": 1466, "len_title": 80, "len_sents": 47190, "len_sents_1": 28921, "n_sents": 370, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1476, "title": "The Uncanny Valley: Exploring Adversarial Robustness from a Flatness Perspective", "abs": "Flatness of the loss surface not only correlates positively with generalization, but is also related to adversarial\nrobustness, since perturbations of inputs relate non-linearly to perturbations of weights. In this paper, we empirically\nanalyze the relation between adversarial examples and relative flatness with respect to the parameters of one layer.\nWe observe a peculiar property of adversarial examples in the context of relative flatness: during an iterative first-order\nwhite-box attack, the flatness of the loss surface measured around the adversarial example *first* becomes sharper\nuntil the label is flipped, but if we keep the attack running, it runs into a flat *uncanny valley* where the label remains\nflipped. In extensive experiments, we observe this phenomenon across various model architectures and datasets, \neven for adversarially trained models. Our results also extend to large language models (LLMs), but due to the discrete\nnature of the input space and comparatively weak attacks, adversarial examples rarely reach truly flat regions. Most\nimportantly, this phenomenon shows that flatness alone cannot explain adversarial robustness unless we can also\nguarantee the behavior of the function around the examples. We therefore theoretically connect relative flatness to\nadversarial robustness by bounding the third derivative of the loss surface, underlining the need for flatness in\ncombination with a low global Lipschitz constant for a robust model.", "keywords": [ "adversarial robustness", "flatness", "LLMs" ], "rating_list": [ 3, 3, 8, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "qac43AwuL9", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "Francisco N. F. Q. Simoes", "gender": "Not Specified", "institution": "Utrecht University", "country": "NL", "position": "PhD student" }, { "name": "Mehdi Dastani", "gender": "Male", "institution": "Utrecht University", "country": "", "position": "Full Professor" }, { "name": "Thijs van Ommen", "gender": "Male", "institution": "Utrecht University", "country": "", "position": "Assistant Professor" } ], "n_formula": 39, "n_formula_1": 17, "n_ref_uni": 33, "n_ref": 75, "n_ref_all": 98, "n_fig": 7, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4559, "formula_len_all_1": 1503, "len_all": 157393, "len_all_1": 71939, "len_abs": 930, "len_title": 116, "len_sents": 57084, "len_sents_1": 34084, "n_sents": 471, "n_sents_1": 267, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 933, "title": "Optimal Causal Representations and the Causal Information Bottleneck", "abs": "To effectively study complex causal systems, it is often useful to construct representations that simplify parts of the system by discarding irrelevant details while preserving key features.\nThe Information Bottleneck (IB) method is a widely used approach in representation learning that compresses random variables while retaining information about a target variable.\nTraditional methods like IB are purely statistical and ignore underlying causal structures, making them ill-suited for causal tasks.\nWe propose the Causal Information Bottleneck (CIB), a causal extension of the IB, which compresses a set of chosen variables while maintaining causal control over a target variable.\nThis method produces representations which are causally interpretable, and which can be used when reasoning about interventions.\nWe present experimental results demonstrating that the learned representations accurately capture causality as intended.", "keywords": [ "causal representation learning", "causal inference", "information bottleneck", "information theory" ], "rating_list": [ 6, 5, 5, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "qZz7PKt4bE", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Elizabeth M. Daly", "gender": "unknown", "institution": "IBM Research", "country": "", "position": "Research Scientist" }, { "name": "Ivana Dusparic", "gender": "unknown", "institution": "University of Dublin, Trinity College", "country": "IE", "position": "Associate Professor" }, { "name": "Radu Marinescu", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Seshu Tirupathi", "gender": "unknown", "institution": "International Business Machines", "country": "IE", "position": "Researcher" }, { "name": "Shivani Tomar", "gender": "unknown", "institution": "University of Dublin, Trinity College", "country": "IE", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 20, "n_ref": 30, "n_ref_all": 46, "n_fig": 6, "n_tab": 5, "L_tab": 3440, "n_element_tab": 165, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3477, "n_element_tab_1": 165, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 88883, "len_all_1": 50465, "len_abs": 1369, "len_title": 144, "len_sents": 23489, "len_sents_1": 22529, "n_sents": 146, "n_sents_1": 137, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1373, "title": "AutoTune for Time Series Transformers using Low Rank Adaptation and Limited Discrepancy Search", "abs": "Transformer models have achieved remarkable results in the field of Natural Language Processing (NLP) with the introduction of breakthrough large language models like GPT and LLaMA recently. Motivated by their ability to capture long-range dependencies, researchers have successfully adapted these models to the task of time series forecasting. However, despite their potential, effectiveness of applying these pre-trained time series transformer models in the target domain is limited due to the need for hyper-parameter optimisation to match the characteristics of the target domain. This paper presents a novel algorithm that uses parameter efficient fine-tuning such as Low Rank Adaptation (LoRA) coupled with Limited Discrepancy Search (LDS) to efficiently auto fine-tune pre-trained time series transformers for a given target domain. Our approach helps in making informed design choices involving LoRA tunable hyper-parameters with strong performance-cost trade-offs that are highly transferable across different target domains. Our experiments demonstrate that autotune efficiently identifies the optimal configuration of LoRA hyper-parameters, achieving an average MASE\nimprovement of 5.21% across all datasets and 4.76% for out-of-domain datasets compared to zero shot pre-trained models, with improvements as high as 20.59% for one of the out-of-domain datasets.", "keywords": [ "Time Series Transformers", "LoRA", "Time Series Forecasting" ], "rating_list": [ 1, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 2, 2 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 5, 3, 3 ] }, { "paper_id": "qZwtPEw2qN", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Constantinos Costis Daskalakis", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "" }, { "name": "Ioannis Daras", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "PhD student" }, { "name": "Yeshwanth Cherapanamjeri", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" } ], "n_formula": 130, "n_formula_1": 29, "n_ref_uni": 29, "n_ref": 86, "n_ref_all": 102, "n_fig": 3, "n_tab": 6, "L_tab": 2336, "n_element_tab": 204, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2252, "n_element_tab_1": 105, "formula_len_all": 11529, "formula_len_all_1": 2562, "len_all": 196973, "len_all_1": 69288, "len_abs": 3229, "len_title": 120, "len_sents": 59661, "len_sents_1": 30076, "n_sents": 631, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1586, "title": "How Much is a Noisy Image Worth? Data Scaling Laws for Ambient Diffusion.", "abs": "The quality of generative models depends on the quality of the data they are trained on. Creating large-scale, high-quality datasets is often expensive and sometimes impossible, e.g.~in certain scientific applications where there is no access to clean data due to physical or instrumentation constraints. Ambient Diffusion and related frameworks train diffusion models with solely corrupted data (which are usually cheaper to acquire) but ambient models significantly underperform models trained on clean data. We study this phenomenon at scale by training more than $80$ models on data with different corruption levels across three datasets ranging from $30,000$ to $\\approx 1.3$M samples. We show that it is impossible, at these sample sizes, to match the performance of models trained on clean data when only training on noisy data. Yet, a combination of a small set of clean data (e.g.~$10\\%$ of the total dataset) and a large set of highly noisy data suffices to reach the performance of models trained solely on similar-size datasets of clean data, and in particular to achieve near state-of-the-art performance. We provide theoretical evidence for our findings by developing novel sample complexity bounds for learning from Gaussian Mixtures with heterogeneous variances. Our theoretical model suggests that, for large enough datasets, the effective marginal utility of a noisy sample is exponentially worse that of a clean sample. Providing a small set of clean samples can significantly reduce the sample size requirements for noisy data, as we also observe in our experiments.", "keywords": [ "ambient diffusion", "noisy data", "dataset design", "gaussian mixtures" ], "rating_list": [ 6, 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 4, 3, 3 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 2, 3 ] }, { "paper_id": "qZmn2hkuzw", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Annie Xie", "gender": "unknown", "institution": "Google", "country": "", "position": "Intern" }, { "name": "Chelsea Finn", "gender": "Female", "institution": "Physical Intelligence", "country": "US", "position": "Researcher" }, { "name": "Jubayer Ibn Hamid", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Max Du", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Yoonho Lee", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Yuejiang Liu", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Postdoc" } ], "n_formula": 66, "n_formula_1": 12, "n_ref_uni": 61, "n_ref": 117, "n_ref_all": 150, "n_fig": 15, "n_tab": 8, "L_tab": 1148, "n_element_tab": 117, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 486, "n_element_tab_1": 65, "formula_len_all": 10767, "formula_len_all_1": 644, "len_all": 265938, "len_all_1": 67019, "len_abs": 1402, "len_title": 76, "len_sents": 72697, "len_sents_1": 32315, "n_sents": 631, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1454, "title": "Bidirectional Decoding: Improving Action Chunking via Closed-Loop Resampling", "abs": "Predicting and executing a sequence of actions without intermediate replanning, known as action chunking, is increasingly used in robot learning from human demonstrations. Yet, its reported effects on the learned policy are inconsistent: some studies find it crucial for achieving strong results, while others observe decreased performance. In this paper, we first dissect how action chunking impacts the divergence between a learner and a demonstrator. We find that action chunking allows the learner to better capture the temporal dependencies in demonstrations (e.g., latent strategies) but at the cost of reduced reactivity in stochastic environments (e.g., action noise, object motions). To address this tradeoff, we propose Bidirectional Decoding (BID), a test-time inference algorithm that bridges action chunking with closed-loop operations. BID samples multiple predictions at each time step and searches for the optimal one based on two criteria: (i) backward coherence, which favors samples that align with previous decisions; (ii) forward contrast, which seeks samples of high likelihood for future plans. By coupling decisions within and across action chunks, BID promotes strong temporal consistency over multiple steps while maintaining high reactivity to unexpected state changes. Experimental results show that BID boosts the performance of two state-of-the-art robot policies across seven simulation benchmarks and two real-world tasks.", "keywords": [ "Robot Learning", "Action Chunking", "Test-Time Decoding" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 1, 4, 4, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "qZEdmyqCHF", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Donghyeon Kwon", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "PhD student" }, { "name": "Jaehyun Pahk", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Seong Joon Oh", "gender": "Male", "institution": "Parameter Lab", "country": "DE", "position": "Principal Researcher" }, { "name": "Suha Kwak", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 44, "n_ref": 128, "n_ref_all": 150, "n_fig": 6, "n_tab": 8, "L_tab": 2229, "n_element_tab": 302, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 611, "n_element_tab_1": 60, "formula_len_all": 340, "formula_len_all_1": 175, "len_all": 144318, "len_all_1": 53404, "len_abs": 1496, "len_title": 116, "len_sents": 35380, "len_sents_1": 25567, "n_sents": 238, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1501, "title": "Decoupled Finetuning for Domain Generalizable Semantic Segmentation", "abs": "Joint finetuning of a pretrained encoder and a randomly initialized decoder has been the de facto standard in semantic segmentation, but the vulnerability of this approach to domain shift has not been studied. We investigate the vulnerability issue of joint finetuning, and propose a novel finetuning framework called Decoupled FineTuning for domain generalization (DeFT) as a solution. DeFT operates in two stages. Its first stage warms up the decoder with the frozen, pretrained encoder so that the decoder learns task-relevant knowledge while the encoder preserves its generalizable features. In the second stage, it decouples finetuning of the encoder and decoder into two pathways, each of which concatenates a usual component (UC) and generalized component (GC); each of the encoder and decoder plays a different role between UC and GC in different pathways. UCs are updated by gradients of the loss on the source domain, while GCs are updated by exponential moving average biased toward their initialization to retain their generalization capability. By the two separate optimization pathways with opposite UC-GC configurations, DeFT reduces the number of learnable parameters virtually, and decreases the distance between learned parameters and their initialization, leading to improved generalization capability. DeFT significantly outperformed existing methods in various domain shift scenarios, and its performance could be further boosted by incorporating a simple distance regularization.", "keywords": [ "Domain Generalization", "Decoupled Optimization", "Domain Generalizable Semantic Segmentation", "Robustness" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 4, 2, 3, 3, 3 ], "presentation_list": [ 4, 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 4, 3, 3 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "qZ4jYual5d", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Carl R Richardson", "gender": "Male", "institution": "University of Southampton", "country": "GB", "position": "PhD student" }, { "name": "Matthew C. Turner", "gender": "Male", "institution": "University of Southampton", "country": "GB", "position": "Full Professor" }, { "name": "Steve R. Gunn", "gender": "unknown", "institution": "University of Southampton", "country": "GB", "position": "Full Professor" } ], "n_formula": 62, "n_formula_1": 18, "n_ref_uni": 45, "n_ref": 91, "n_ref_all": 137, "n_fig": 6, "n_tab": 7, "L_tab": 1771, "n_element_tab": 153, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1048, "n_element_tab_1": 113, "formula_len_all": 6875, "formula_len_all_1": 2522, "len_all": 179165, "len_all_1": 73714, "len_abs": 1280, "len_title": 107, "len_sents": 55535, "len_sents_1": 32087, "n_sents": 466, "n_sents_1": 241, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1290, "title": "Robust Lurie Networks with Controllable Convergent Dynamics", "abs": "The Lurie Network is proposed as a unifying architecture for modelling time-invariant nonlinear dynamical systems. Many existing continuous-time models including Recurrent Neural Networks and Neural Oscillators are special cases of the Lurie Network when applied to this domain. Motivated by the need for a general inductive bias, shared by many systems, this paper proposes an approach to enable network weights and biases to be trained in such a manner that a generalised concept of stability is guaranteed. This generalised stability measure is that of $k$-contraction which enables global convergence to a point, line or plane in the neural state-space. This result is leveraged to construct a Graph Lurie Network (GLN) satisfying the same convergence properties. Unconstrained parametrisations of these conditions are derived allowing the models to be trained using standard optimisation algorithms, whilst limiting the search space to solutions satisfying the $k$-contraction constraints. Empirical results show significant improvement in terms of prediction accuracy, generalisation and robustness compared to other unconstrained and stability-constrained models. Furthermore, both models consistently learnt representations which respected the convergence behaviour of the dynamics.", "keywords": [ "dynamical systems", "convergence", "robustness", "$k$-contraction analysis", "RNNs" ], "rating_list": [ 3, 5, 5, 1 ], "soundness_list": [ 3, 2, 3, 1 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "qYniSDqk8a", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Daniel Kienzle", "gender": "Male", "institution": "Universit\u00e4t Augsburg", "country": "DE", "position": "PhD student" }, { "name": "Julian Lorenz", "gender": "unknown", "institution": "Universit\u00e4t Augsburg", "country": "DE", "position": "PhD student" }, { "name": "Katja Ludwig", "gender": "Female", "institution": "University of Augsburg", "country": "DE", "position": "PhD student" }, { "name": "Rainer Lienhart", "gender": "Male", "institution": "Universit\u00e4t Augsburg", "country": "DE", "position": "Full Professor" }, { "name": "Tuan Bui", "gender": "Not Specified", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 34, "n_ref": 56, "n_ref_all": 72, "n_fig": 4, "n_tab": 4, "L_tab": 1517, "n_element_tab": 161, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2337, "n_element_tab_1": 191, "formula_len_all": 0, "formula_len_all_1": 7, "len_all": 120791, "len_all_1": 71256, "len_abs": 1134, "len_title": 153, "len_sents": 39085, "len_sents_1": 33072, "n_sents": 331, "n_sents_1": 267, "venue": "ICLR2025", "decision": "withdraw", "L_title": 105, "L_abs": 1138, "title": "Leveraging Anthropometric Measurements to Improve Human Mesh Estimation and Ensure Consistent Body Shapes", "abs": "The basic body shape of a person does not change within a single video. However, most SOTA human mesh estimation (HME) models output a slightly different body shape for each video frame, which results in inconsistent body shapes for the same person. In contrast, we leverage anthropometric measurements like tailors are already obtaining from humans for centuries. We create a model called A2B that converts such anthropometric measurements to body shape parameters of human mesh models. Moreover, we find that finetuned SOTA 3D human pose estimation (HPE) models outperform HME models regarding the precision of the estimated keypoints. We show that applying inverse kinematics (IK) to the results of such a 3D HPE model and combining the resulting body pose with the A2B body shape leads to superior and consistent human meshes for challenging datasets like ASPset or fit3D, where we can lower the MPJPE by over 30 mm compared to SOTA HME models. Further, replacing HME models estimates of the body shape parameters with A2B model results not only increases the performance of these HME models, but also leads to consistent body shapes.", "keywords": [ "Computer Vision", "Human Pose Estimation", "3D Human Pose Estimation", "Human Mesh Estimation", "Human Mesh Recovery", "Anthropometric Measurements" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 4, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "qXEmoWllKW", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Andreas Mayr", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Postdoc" }, { "name": "Benedikt Alkin", "gender": "Male", "institution": "Emmi AI GmbH", "country": "AT", "position": "Researcher" }, { "name": "Florian Sestak", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "PhD student" }, { "name": "G\u00fcnter Klambauer", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Full Professor" }, { "name": "Johannes Brandstetter", "gender": "Male", "institution": "Emmi AI", "country": "AT", "position": "Principal Researcher" }, { "name": "Nikolaus A. Adams", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "Full Professor" }, { "name": "Andreas Fuerst", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "PhD student" }, { "name": "Artur Toshev", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 87, "n_ref": 161, "n_ref_all": 190, "n_fig": 9, "n_tab": 5, "L_tab": 3664, "n_element_tab": 243, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 437, "n_element_tab_1": 73, "formula_len_all": 811, "formula_len_all_1": 263, "len_all": 244482, "len_all_1": 70323, "len_abs": 1671, "len_title": 126, "len_sents": 64174, "len_sents_1": 33640, "n_sents": 445, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1572, "title": "UPT++: Latent Point Set Neural Operators for Modeling System State Transitions", "abs": "Particle methods comprise a wide spectrum of numerical algorithms, ranging from computational fluid dynamics governed by the Navier-Stokes equations to molecular dynamics governed by the many-body Schr\\\"odinger equation. At its core, these methods represent the continuum as a collection of discrete particles, on which the respective PDE is solved. We introduce UPT++, a latent point set neural operator for modeling the dynamics of such particle systems by mapping a particle set back to a continuous (latent) representation, instead of operating on the particles directly. We argue via what we call the *discretization paradox* that continuous modeling is advantageous even if the reference numerical discretization scheme comprises particles. Algorithmically, UPT++ extends Universal Physics Transformers -- a framework for efficiently scaling neural operators -- by novel importance-based encoding and decoding. Furthermore, our encoding and decoding enable outputs that remain consistent across varying input sampling resolutions, i.e., UPT++ is a neural operator. We discuss two types of UPT++ operators: (i) time-evolution operator for fluid dynamics, and (ii) sampling operator for molecular dynamics tasks. Experimentally, we demonstrate that our method reliably models complex physics phenomena of fluid dynamics and exhibits beneficial scaling properties, tested on simulations of up to 200k particles. Furthermore, we showcase on molecular dynamics simulations that UPT++ can effectively explore the metastable conformation states of unseen peptide molecules.", "keywords": [ "neural operator", "navier-stokes", "molecular dynamics", "latent space", "particle simulations" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "qWLgJCl1Y6", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 8, "author_info_list": [ { "name": "Bolin Ding", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Senior Director" }, { "name": "Jialing Bi", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "Undergrad student" }, { "name": "Runlin Lei", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Xuchen Pan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yaliang Li", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Staff Engineer" }, { "name": "Yankai Lin", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "Assistant Professor" }, { "name": "Zhewei Wei", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Jijiarui", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 56, "n_ref": 108, "n_ref_all": 152, "n_fig": 14, "n_tab": 10, "L_tab": 2880, "n_element_tab": 292, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 551, "n_element_tab_1": 94, "formula_len_all": 993, "formula_len_all_1": 719, "len_all": 221548, "len_all_1": 70637, "len_abs": 1637, "len_title": 115, "len_sents": 66797, "len_sents_1": 32459, "n_sents": 524, "n_sents_1": 243, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1654, "title": "Large-Scale Dynamic Graph Generation via LLM-based Agent Simulation", "abs": "Graph generation is a fundamental task that has been extensively studied in social, technological, and scientific analysis. \nFor modeling the dynamic graph evolution process, traditional rule-based methods struggle to capture community structures within graphs, while deep learning methods only focus on fitting training graphs. \nThis limits existing graph generators to producing graphs that adhere to predefined rules or closely resemble training datasets, achieving poor performance in dynamic graph generation.\nGiven that graphs are abstract representations arising from pairwise interactions in human activities, a realistic simulation of human-wise interaction could provide deeper insights into the graph evolution mechanism.\nWith the increasing recognition of large language models (LLMs) in simulating human behavior, we introduce GraphAgent-Generator (GAG), a novel simulation-based framework for dynamic text-attributed graph generation. Without training or fine-tuning process of LLM, our framework effectively replicates seven macro-level structural characteristics in established network science theories while surpassing existing baselines in graph expansion tasks by 11\\% on specific evaluation metrics. Through node classification task, we validate GAG effectively captures the intricate text-structure correlations in graph generation. \nFurthermore, GAG supports generating graphs with up to nearly 100,000 nodes or 10 million edges through large-scale LLM-based agent simulation with parallel acceleration, achieving a minimum speed-up of 90.4\\%.\nThe source code is available at \\url{https://anonymous.4open.science/r/GraphAgent-2206}.", "keywords": [ "LLM-based Agents", "graph generation" ], "rating_list": [ 5, 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 2, 3 ] }, { "paper_id": "qW5f8TAZ4J", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Jingtong Hu", "gender": "Male", "institution": "University of Pittsburgh", "country": "US", "position": "Associate Professor" }, { "name": "Pan Wang", "gender": "Male", "institution": "University of Pittsburgh", "country": "US", "position": "PhD student" }, { "name": "Ruichen Zhang", "gender": "unknown", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "MS student" }, { "name": "Sijia Liu", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Assistant Professor" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Yuguang Yao", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Zhen Tan", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Zhimin Li", "gender": "Male", "institution": "University of Pittsburgh", "country": "US", "position": "PhD student" }, { "name": "huan liu", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 38, "n_ref": 50, "n_ref_all": 64, "n_fig": 5, "n_tab": 3, "L_tab": 676, "n_element_tab": 95, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 432, "n_element_tab_1": 76, "formula_len_all": 204, "formula_len_all_1": 204, "len_all": 113497, "len_all_1": 59390, "len_abs": 1102, "len_title": 112, "len_sents": 34441, "len_sents_1": 29683, "n_sents": 230, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1102, "title": "FairSkin: Fair Diffusion for Skin Disease Image Geneartion", "abs": "Image generation is a prevailing technique for clinical data augmentation for advancing diagnostic accuracy and reducing healthcare disparities. Diffusion Model (DM) has become a leading method in generating synthetic medical images, but it suffers from a critical twofold bias: (1)The quality of images generated for Caucasian individuals is significantly higher, as measured by the Fr\u00e9chet Inception Distance (FID). (2)The ability of the downstream-task learner to learn critical features from disease images varies across different skin tones. These biases pose significant risks, particularly in skin disease detection, where underrepresentation of certain skin tones can lead to misdiagnosis or neglect of specific conditions. To address these challenges, we propose FairSkin, a novel DM framework that mitigates these biases through a three-level resampling mechanism, ensuring fairer representation across racial and disease categories. Our approach significantly improves the diversity and quality of generated images, contributing to more equitable skin disease detection in clinical settings.", "keywords": [ "fairness", "diffusion model", "medical image generation" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "qVyjN01x4P", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Dongyue Chen", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "Postdoc" }, { "name": "Emadeldeen Eldele", "gender": "Male", "institution": "Agency for Science, Technology and Research", "country": "SG", "position": "Research Scientist" }, { "name": "Min Wu", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Qinghua Hu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Professor" }, { "name": "Yanru Sun", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Zongxia Xie", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" } ], "n_formula": 43, "n_formula_1": 18, "n_ref_uni": 40, "n_ref": 86, "n_ref_all": 132, "n_fig": 10, "n_tab": 14, "L_tab": 8368, "n_element_tab": 468, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 176, "n_element_tab_1": 24, "formula_len_all": 2118, "formula_len_all_1": 813, "len_all": 188848, "len_all_1": 58316, "len_abs": 1488, "len_title": 98, "len_sents": 51010, "len_sents_1": 27431, "n_sents": 409, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1496, "title": "Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift", "abs": "Time series forecasting, which aims to predict future values based on historical data, has garnered significant attention due to its broad range of applications. However, real-world time series often exhibit complex non-uniform distribution with varying patterns across segments, such as season, operating condition, or semantic meaning, making accurate forecasting challenging. Existing approaches, which typically train a single model to capture all these diverse patterns, often struggle with the pattern drifts between patches and may lead to poor generalization. To address these challenges, we propose TFPS, a novel architecture that leverages pattern-specific experts for more accurate and adaptable time series forecasting. TFPS employs a dual-domain encoder to capture both time-domain and frequency-domain features, enabling a more comprehensive understanding of temporal dynamics. It then uses subspace clustering to dynamically identify distinct patterns across data patches. Finally, pattern-specific experts model these unique patterns, delivering tailored predictions for each patch. By explicitly learning and adapting to evolving patterns, TFPS achieves significantly improved forecasting accuracy. Extensive experiments on real-world datasets demonstrate that TFPS outperforms state-of-the-art methods, particularly in long-term forecasting, through its dynamic and pattern-aware learning approach. The data and codes are available: https://anonymous.4open.science/r/TFPS-D001.", "keywords": [ "Time Series Forecasting", "Distribution Shift", "Deep Learning" ], "rating_list": [ 5, 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 2, 4, 4 ] }, { "paper_id": "qVtfN6NoJi", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Hanyue Xiao", "gender": "Female", "institution": "Washington University, Saint Louis", "country": "", "position": "MS student" }, { "name": "Jr-Shin Li", "gender": "Male", "institution": "Washington University, St. Louis", "country": "", "position": "Full Professor" }, { "name": "Vignesh Narayanan", "gender": "unknown", "institution": "University of South Carolina", "country": "", "position": "Assistant Professor" }, { "name": "Wan-Yun Shen", "gender": "unknown", "institution": "Washington University, Saint Louis", "country": "US", "position": "Researcher" }, { "name": "Wei Zhang", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "US", "position": "Postdoc" }, { "name": "Yuan-Hung Kuan", "gender": "Male", "institution": "Washington University, St. Louis", "country": "", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 22, "n_ref_uni": 49, "n_ref": 72, "n_ref_all": 84, "n_fig": 4, "n_tab": 3, "L_tab": 735, "n_element_tab": 78, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2072, "formula_len_all_1": 1908, "len_all": 133463, "len_all_1": 76794, "len_abs": 1285, "len_title": 100, "len_sents": 34362, "len_sents_1": 32679, "n_sents": 257, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1291, "title": "Layer-Varying Deep Reservoir Computing Architecture", "abs": "Data loss and corruption are common incidents that often lead to catastrophic consequences in both theoretical and experimental facets of data analytics. The aspiration to minimize the impacts of such consequences drives the demand for the development of effective data analytic tools and imputation methods to replace missing, corrupted, or artifacted data. \nThe focus of this paper is on multivariate time series imputation, for which we develop a dynamical systems-theoretic deep learning approach. The central idea is to view a multivariate time series as a trajectory of a dynamical system. Then, we construct a deep reservoir computing architecture to model the temporal evolution of the system by using existing data in the time series. In particular, this architecture is composed of a cascade of echo state network (ESN) layers with diminishing reservoir sizes. We then propose a layer-by-layer training scheme, which gives rise to a deep learning-based time series imputation algorithm. We further provide a rigorous convergence analysis of this algorithm by exploiting the echo state property of ESN, and demonstrate the imputation performance as well as the efficiency of the training process by utilizing both synthetic and real-world datasets arising from diverse applications.", "keywords": [ "Multivariate time series", "imputation", "reservoir computing networks", "dynamical systems" ], "rating_list": [ 5, 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 1, 2 ], "contribution_list": [ 2, 3, 3, 1, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "qUZY7ymDPr", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chen Li", "gender": "Not Specified", "institution": "Tencent ARC Lab", "country": "CN", "position": "Researcher" }, { "name": "Haibo Lu", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Associate Professor" }, { "name": "Haoran Tang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jiankun Yang", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Full Professor" }, { "name": "Ruyang Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Ying Shan", "gender": "Male", "institution": "Tencent AI Lab Center of Visual Computing", "country": "CN", "position": "Director" }, { "name": "Yixiao Ge", "gender": "Female", "institution": "Tencent", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 115, "n_ref_all": 133, "n_fig": 8, "n_tab": 8, "L_tab": 3427, "n_element_tab": 461, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 3230, "n_element_tab_1": 305, "formula_len_all": 376, "formula_len_all_1": 376, "len_all": 156783, "len_all_1": 70755, "len_abs": 1777, "len_title": 114, "len_sents": 44627, "len_sents_1": 29585, "n_sents": 372, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1567, "title": "PPLLaVA: Varied Video Sequence Understanding With Prompt Guidance", "abs": "The past year has witnessed the significant advancement of video-based large language models. However, the challenge of developing a unified model for both short and long video understanding remains unresolved. Most existing video LLMs cannot handle hour-long videos, while methods custom for long videos tend to be ineffective for shorter videos and images. In this paper, we identify the key issue as the redundant content in videos. To address this, we propose a novel pooling strategy that simultaneously achieves token compression and instruction-aware visual feature aggregation. Our model is termed Prompt-guided Pooling LLaVA, or PPLLaVA for short. Specifically, PPLLaVA consists of three core components: the CLIP-based visual-prompt alignment that extracts visual information relevant to the user's instructions, the prompt-guided pooling that compresses the visual sequence to arbitrary scales using convolution-style pooling, and the clip context extension designed for lengthy prompt common in visual dialogue. Moreover, our codebase also integrates the most advanced video Direct Preference Optimization (DPO) and visual interleave training. Extensive experiments have validated the performance of our model. With superior throughput, PPLLaVA achieves better results on image benchmarks as a video LLM, while achieving state-of-the-art performance across various video benchmarks, excelling in tasks ranging from caption generation to multiple-choice questions, and handling video lengths from seconds to hours. The codes are promised to be made public.", "keywords": [ "Video LLM", "Prompt-guided Pooling", "PPLLaVA" ], "rating_list": [ 3, 6, 3, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 4, 3 ], "confidence_list": [ 4, 4, 5, 5, 4 ] }, { "paper_id": "qUJsX3XMBH", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "An Yang", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Bowen Yu", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Junyang Lin", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Kai Dang", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Tingyu Xia", "gender": "Female", "institution": "Jilin University", "country": "", "position": "PhD student" }, { "name": "Yi Chang", "gender": "Male", "institution": "Jilin University, China", "country": "CN", "position": "Full Professor" }, { "name": "Yuan Tian", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuan Wu", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 26, "n_ref": 62, "n_ref_all": 79, "n_fig": 2, "n_tab": 6, "L_tab": 3687, "n_element_tab": 208, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1618, "n_element_tab_1": 103, "formula_len_all": 751, "formula_len_all_1": 614, "len_all": 132973, "len_all_1": 58347, "len_abs": 3228, "len_title": 123, "len_sents": 32565, "len_sents_1": 27896, "n_sents": 227, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1337, "title": "Rethinking Data Selection at Scale: Random Selection is Almost All You Need", "abs": "Supervised fine-tuning (SFT) is crucial for aligning Large Language Models (LLMs) with human instructions. The primary goal during SFT is to select a small yet representative subset of training data from the larger pool, such that fine-tuning with this subset achieves results comparable to or even exceeding those obtained using the entire dataset. However, most existing data selection techniques are designed for small-scale data pools, which fail to meet the demands of real-world SFT scenarios. In this paper, we replicated several self-scoring methods\u2014those that do not rely on external model assistance\u2014on two million-scale datasets, and found that nearly all methods struggled to significantly outperform random selection when dealing with such large-scale data pools. Moreover, our comparisons suggest that, during SFT, diversity in data selection is more critical than simply focusing on high-quality data. We also analyzed the limitations of several current approaches, explaining why they perform poorly on large-scale datasets and why they are unsuitable for such contexts. Finally, we found that filtering data by token length offers a stable and efficient method for improving results. This approach, particularly when training on long-text data, proves highly beneficial for relatively weaker base models, such as Llama3.", "keywords": [ "large language model", "supervised fine-tuning", "data selection" ], "rating_list": [ 3, 5, 5, 3, 6 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "qU1GtrDDst", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Antony Krymski", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "MS student" }, { "name": "Paul Alexander Bilokon", "gender": "unknown", "institution": "Thalesians Marine Ltd, the AI in shipping and logistics company", "country": "", "position": "Principal Researcher" }, { "name": "Tom Davison", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Lecturer" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 9, "n_ref": 11, "n_ref_all": 20, "n_fig": 6, "n_tab": 1, "L_tab": 116, "n_element_tab": 10, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 116, "n_element_tab_1": 10, "formula_len_all": 471, "formula_len_all_1": 490, "len_all": 63634, "len_all_1": 51526, "len_abs": 876, "len_title": 109, "len_sents": 24383, "len_sents_1": 23314, "n_sents": 201, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 879, "title": "Representation learning for financial time series forecasting", "abs": "The accurate forecasting of financial time series remains a significant challenge due to the stochastic nature of the underlying data. To improve prediction accuracy, feature engineering has become a vital aspect of forecasting financial assets. However, engineering features manually often requires domain expertise. We propose to utilise an automated feature generation architecture, Contrastive Predictive Coding (CPC), to generate embeddings as input to improve the performance of downstream financial time series forecasting models. To benchmark the effectiveness of our approach, we evaluate forecasting models on predicting the next day's log return on various foreign exchange markets with and without embeddings. Finally, we assess our CPC architecture by employing the same trained encoder on different currency pairs and calculating the Sharpe ratio of our strategies.", "keywords": [ "representation learning", "contrastive predictive coding", "cpc" ], "rating_list": [ 3, 1, 1, 1, 3 ], "soundness_list": [ 1, 1, 1, 2, 1 ], "presentation_list": [ 1, 1, 1, 1, 2 ], "contribution_list": [ 2, 1, 1, 1, 1 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "qTrEq31Shm", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Guanzheng Chen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Lidong Bing", "gender": "unknown", "institution": "Alibaba Group", "country": "SG", "position": "Scientist" }, { "name": "Xin Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Qizhe Xie", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 18, "n_ref_uni": 28, "n_ref": 53, "n_ref_all": 79, "n_fig": 6, "n_tab": 2, "L_tab": 1408, "n_element_tab": 238, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1408, "n_element_tab_1": 238, "formula_len_all": 2000, "formula_len_all_1": 1451, "len_all": 153771, "len_all_1": 66824, "len_abs": 1563, "len_title": 154, "len_sents": 36906, "len_sents_1": 29027, "n_sents": 260, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 106, "L_abs": 1575, "title": "LongPO: Long Context Self-Evolution of Large Language Models through Short-to-Long Preference Optimization", "abs": "Large Language Models (LLMs) have demonstrated remarkable capabilities through pretraining and alignment. However, superior short-context LLMs may underperform in long-context scenarios due to insufficient long-context alignment. This alignment process remains challenging due to the impracticality of human annotation for extended contexts and the difficulty in balancing short- and long-context performance. To address these challenges, we introduce LongPO, that enables short-context LLMs to self-evolve to excel on long-context tasks by internally transfer short-context capabilities. LongPO harnesses LLMs to learn from self-generated short-to-long preference data, comprising paired responses generated for identical instructions with long-context inputs and their compressed short-context counterparts, respectively. This preference reveals capabilities and potentials of LLMs cultivated during short-context alignment that may be diminished in under-aligned long-context scenarios. Additionally, LongPO incorporates a short-to-long KL constraint to mitigate short-context performance decline during long-context alignment. When applied to Mistral-7B-Instruct-v0.2 from 128K to 256K context length, LongPO fully retaining short-context performance and largely outperforms naive SFT and DPO in both long- and short-context tasks. Specifically, \\ourMethod-trained models can achieve results on long-context benchmarks comparable to, or even surpassing, those of superior LLMs (e.g., GPT-4-128K) that involve extensive long-context annotation and larger parameter scales.", "keywords": [ "Long Context LLMs and Alignment", "Large Language Models", "Preference Optimization", "Self-Evolution of LLMs" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "qTWDpbF47t", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Xingyi Yang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Xinchao Wang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 51, "n_ref": 104, "n_ref_all": 134, "n_fig": 9, "n_tab": 6, "L_tab": 1694, "n_element_tab": 181, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1164, "n_element_tab_1": 119, "formula_len_all": 848, "formula_len_all_1": 560, "len_all": 193905, "len_all_1": 62294, "len_abs": 2136, "len_title": 99, "len_sents": 47936, "len_sents_1": 24548, "n_sents": 488, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1681, "title": "Compositional Video Generation as Flow Equalization", "abs": "Large-scale Text-to-Video (T2V) diffusion models have recently demonstrated unprecedented capability to transform natural language descriptions into stunning and photorealistic videos. Despite these promising results, a significant challenge remains: these models struggle to fully grasp complex compositional interactions between multiple concepts and actions. This issue arises when some words dominantly influence the final video, overshadowing other concepts.\n \nTo tackle this problem, we introduce \\textbf{Vico}, a generic framework for compositional video generation that explicitly ensures all concepts are represented properly. At its core, Vico analyzes how input tokens influence the generated video, and adjusts the model to prevent any single concept from dominating. Specifically, Vico extracts attention weights from all layers to build a spatial-temporal attention graph, and then estimates the influence as the \\emph{max-flow} from the source text token to the video target token. Although the direct computation of attention flow in diffusion models is typically infeasible, we devise an efficient approximation based on subgraph flows and employ a fast and vectorized implementation, which in turn makes the flow computation manageable and differentiable. By updating the noisy latent to balance these flows, Vico captures complex interactions and consequently produces videos that closely adhere to textual descriptions. We apply our method to multiple diffusion-based video models for compositional T2V and video editing. Empirical results demonstrate that our framework significantly enhances the compositional richness and accuracy of the generated videos.", "keywords": [ "Video Generation; Compositionality" ], "rating_list": [ 5, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "qT0IWGqo1j", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Anton Bushuiev", "gender": "unknown", "institution": "Czech Technical University in Prague", "country": "CZ", "position": "PhD student" }, { "name": "Jiri Sedlar", "gender": "Male", "institution": "CIIRC, Czech Technical University, Czech Technical University of Prague", "country": "", "position": "Researcher" }, { "name": "Josef Sivic", "gender": "Male", "institution": "Czech Technical University in Prague", "country": "CZ", "position": "Principal investigator" }, { "name": "Nikola Zadorozhny", "gender": "unknown", "institution": "CIIRC, Czech Technical University, Czech Technical University of Prague", "country": "CZ", "position": "Researcher" }, { "name": "Raman Samusevich", "gender": "Male", "institution": "Qminers", "country": "CZ", "position": "Researcher" }, { "name": "Roman Bushuiev", "gender": "Male", "institution": "Institute of Organic Chemistry and Biochemistry of the CAS", "country": "CZ", "position": "PhD student" }, { "name": "Tomas Pluskal", "gender": "Male", "institution": "IOCB Prague", "country": "CZ", "position": "Principal Researcher" }, { "name": "Hannes St\u00e4rk", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 91, "n_ref": 212, "n_ref_all": 265, "n_fig": 17, "n_tab": 6, "L_tab": 3526, "n_element_tab": 341, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1138, "n_element_tab_1": 34, "formula_len_all": 96, "formula_len_all_1": 96, "len_all": 271564, "len_all_1": 66705, "len_abs": 1220, "len_title": 47, "len_sents": 69407, "len_sents_1": 33664, "n_sents": 465, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1225, "title": "Training on test proteins improves fitness, structure, and function prediction", "abs": "Data scarcity and distribution shifts often hinder the ability of machine learning models to generalize when applied to proteins and other biological data. Self-supervised pre-training on large datasets is a common method to enhance generalization. However, striving to perform well on all possible proteins can limit model\u2019s capacity to excel on any specific one, even though practitioners are often most interested in accurate predictions for the individual protein they study. To address this limitation, we propose an orthogonal approach to achieve generalization. Building on the prevalence of self-supervised pre-training, we introduce a method for self-supervised fine-tuning at test time, allowing models to adapt to the test protein of interest on the fly and without requiring any additional data. We study our test-time training (TTT) method through the lens of perplexity minimization and show that it consistently enhances generalization across different models, their scales, and datasets. Notably, our method leads to new state-of-the-art results on the standard benchmark for protein fitness prediction, improves protein structure prediction for challenging targets, and enhances function prediction accuracy.", "keywords": [ "proteins", "generalization", "self-supervised learning", "test-time training" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "qSEEQPNbu4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Can Zhang", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Gim Hee Lee", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 39, "n_ref": 120, "n_ref_all": 130, "n_fig": 6, "n_tab": 7, "L_tab": 1772, "n_element_tab": 241, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1863, "n_element_tab_1": 222, "formula_len_all": 308, "formula_len_all_1": 308, "len_all": 131357, "len_all_1": 70104, "len_abs": 1360, "len_title": 129, "len_sents": 33844, "len_sents_1": 31003, "n_sents": 239, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1366, "title": "econSG: Efficient and Multi-view Consistent Open-Vocabulary 3D Semantic Gaussians", "abs": "The primary focus of most recent works on open-vocabulary neural fields is extracting precise semantic features\nfrom the VLMs and then consolidating them efficiently into a multi-view consistent 3D neural fields\nrepresentation. However, most existing works over-trusted SAM to regularize image-level CLIP without any further refinement. Moreover, several existing works improved efficiency by dimensionality reduction of semantic features from 2D VLMs before fusing with 3DGS semantic fields, which inevitably leads to multi-view inconsistency. In this work, we propose econSG for open-vocabulary semantic segmentation with 3DGS. Our econSG consists of: 1) A Confidence-region Guided Regularization (CRR) that mutually refines SAM and CLIP to get the best of both worlds for precise semantic features with complete and precise boundaries. 2) A low dimensional contextual space to enforce 3D multi-view consistency while improving computational efficiency by fusing backprojected multi-view 2D features and follow by dimensional reduction directly on the fused 3D features instead of operating on each 2D view separately. Our econSG show state-of-the-art performance on four benchmark datasets compared to the existing methods. Furthermore, we are also the most efficient training among all the methods. We will make our source-code open source upon paper acceptance.", "keywords": [ "3D Scene Understanding", "Gaussian Splatting", "Open-Vocabulary 3D Semantic" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "qQS2VuHb74", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Alessandro Farinelli", "gender": "Male", "institution": "Universit\u00e0 degli Studi di Verona", "country": "IT", "position": "Full Professor" }, { "name": "Ferdinando Cicalese", "gender": "unknown", "institution": "Universit\u00e0 degli Studi di Verona", "country": "IT", "position": "Full Professor" }, { "name": "Luca Marzari", "gender": "Male", "institution": "University of Verona", "country": "IT", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 21, "n_ref_uni": 31, "n_ref": 104, "n_ref_all": 120, "n_fig": 5, "n_tab": 4, "L_tab": 762, "n_element_tab": 139, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 382, "n_element_tab_1": 48, "formula_len_all": 2118, "formula_len_all_1": 2119, "len_all": 149745, "len_all_1": 70766, "len_abs": 1207, "len_title": 147, "len_sents": 51033, "len_sents_1": 32427, "n_sents": 344, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 98, "L_abs": 1250, "title": "Enhancing Linear Bound Tightness in Neural Network Verification via Sampling-Based Underestimation", "abs": "We present $\\texttt{PT-LiRPA}$ (Probabilistically Tightened LiRPA), a novel approach that enhances existing linear relaxation-based perturbation analysis (LiRPA) methods for neural network verification. $\\texttt{PT-LiRPA}$ combines LiRPA approaches with a sampling-based underestimation technique to compute probabilistically optimal intermediate bounds, resulting in tighter linear lower and upper bounds. Notably, we show that this approach preserves the soundness of verification results while significantly tightening the bounds for generic non-linear functions. Additionally, we introduce a new metric, $\\Delta^*$, to quantify the tightness for LiRPA bounds and to bound the magnitude of the possible error in the sample-based overestimation, thus complementing the probabilistic bound of statistical results we use. Our empirical evaluation, conducted on several state-of-the-art benchmarks, including those from the International Verification of Neural Networks Competition, demonstrates that $\\texttt{PT-LiRPA}$ achieves higher or comparable verified accuracy with lower verification times. The significantly tighter bounds and better efficiency allow us to verify instances where state-of-the-art methods could not provide a specific answer.", "keywords": [ "Formal Verification of Neural Networks; Probabilistic approaches; Linear Relaxation-Based Perturbation Analysis" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 1, 3, 1 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 3, 1 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "qQ5djlndm5", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Mohamed Elhoseiny", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Xiaoqian Shen", "gender": "Female", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 44, "n_ref": 144, "n_ref_all": 165, "n_fig": 7, "n_tab": 12, "L_tab": 2454, "n_element_tab": 225, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2235, "n_element_tab_1": 180, "formula_len_all": 807, "formula_len_all_1": 619, "len_all": 173703, "len_all_1": 65322, "len_abs": 1629, "len_title": 114, "len_sents": 47721, "len_sents_1": 29688, "n_sents": 359, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1771, "title": "StoryGPT-V: Large Language Models as Consistent Story Visualizers", "abs": "Recent generative models have demonstrated impressive capabilities in generating realistic and visually pleasing images grounded on textual prompts. Nevertheless, a significant challenge remains in applying these models for the more intricate task of story visualization. Since it requires resolving pronouns (he, she, they) in the frame descriptions, i.e., anaphora resolution, and ensuring consistent characters and background synthesis across frames. \nYet, the emerging Large Language Model (LLM) showcases robust reasoning abilities to navigate through ambiguous references and process extensive sequences. Therefore, we introduce \\emph{StoryGPT-V}, which leverages the merits of the latent diffusion (LDM) and LLM to produce images with consistent and high-quality characters grounded on given story descriptions. \nFirst, we train a character-aware LDM, which takes character-augmented semantic embedding as input and includes the supervision of the cross-attention map using character segmentation masks, aiming to enhance character generation accuracy and faithfulness.\nIn the second stage, we enable an alignment between the output of LLM and the character-augmented embedding residing in the input space of the first-stage model. This harnesses the reasoning ability of LLM to address ambiguous references and the comprehension capability to memorize the context. We conduct comprehensive experiments on two visual story visualization benchmarks. Our model reports superior quantitative results and consistently generates accurate characters of remarkable quality with low memory consumption. Our code will be made publicly available\\footnote{Please refer to the \\href{https://storygpt-v.s3.amazonaws.com/index.html}{anonymous webpage} for qualitative results.}.", "keywords": [ "Story Visualization; Large Language Models; Multimodal Coreference Resolution" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "qPzYF2EpXb", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Jieqi Shi", "gender": "Female", "institution": "nanjing university", "country": "CN", "position": "Assistant Professor" }, { "name": "Jing Huo", "gender": "Female", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Junhui Yin", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Pinzhuo Tian", "gender": "Male", "institution": "Shanghai University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yang Gao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Yangtao Chen", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Zixuan Chen", "gender": "Female", "institution": "Nanjing University", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 28, "n_ref": 103, "n_ref_all": 140, "n_fig": 16, "n_tab": 8, "L_tab": 2953, "n_element_tab": 202, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 59, "n_element_tab_1": 2, "formula_len_all": 1458, "formula_len_all_1": 1118, "len_all": 233310, "len_all_1": 66274, "len_abs": 2204, "len_title": 92, "len_sents": 72560, "len_sents_1": 31610, "n_sents": 519, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1827, "title": "GravMAD: Grounded Spatial Value Maps Guided Action Diffusion for Generalized 3D Manipulation", "abs": "Robots' ability to follow language instructions and execute diverse 3D manipulation tasks is vital in robot learning. Traditional imitation learning-based methods perform well on seen tasks but struggle with novel, unseen ones due to variability. Recent approaches leverage large foundation models to assist in understanding novel tasks, thereby mitigating this issue. However, these methods lack a task-specific learning process, which is essential for an accurate understanding of 3D environments, often leading to execution failures. In this paper, we introduce GravMAD, a sub-goal-driven, language-conditioned action diffusion framework that combines the strengths of imitation learning and foundation models. Our approach breaks tasks into sub-goals based on language instructions, allowing auxiliary guidance during both training and inference. During training, we introduce Sub-goal Keypose Discovery to identify key sub-goals from demonstrations. Inference differs from training, as there are no demonstrations available, so we use pre-trained foundation models to bridge the gap and identify sub-goals for the current task. In both phases, GravMaps are generated from sub-goals, providing GravMAD with more flexible 3D spatial guidance compared to fixed 3D positions. Empirical evaluations on RLBench show that GravMAD significantly outperforms state-of-the-art methods, with a 28.63\\% improvement on novel tasks and a 13.36\\% gain on tasks encountered during training. Evaluations on real-world robotic tasks further show that GravMAD can reason about real-world tasks, associate them with relevant visual information, and generalize to novel tasks. These results demonstrate GravMAD's strong multi-task learning and generalization in 3D manipulation. Video demonstrations are available at: https://gravmad.github.io.", "keywords": [ "3D Manipulation", "imitation learning", "foundation models", "sub-goals", "diffusion models" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "qPx3i9sMxv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Honggang Zhang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Associate Professor" }, { "name": "Huadai Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Peiwen Sun", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "HK", "position": "MS student" }, { "name": "Sitong Cheng", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Wei Xue", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Xiangtai Li", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Yike Guo", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Zhen Ye", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 7, "n_ref_uni": 87, "n_ref": 167, "n_ref_all": 219, "n_fig": 19, "n_tab": 22, "L_tab": 8672, "n_element_tab": 983, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 3109, "n_element_tab_1": 259, "formula_len_all": 1480, "formula_len_all_1": 447, "len_all": 326912, "len_all_1": 57474, "len_abs": 2029, "len_title": 69, "len_sents": 89087, "len_sents_1": 25170, "n_sents": 794, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 69, "L_abs": 1672, "title": "Both Ears Wide Open: Towards Language-Driven Spatial Audio Generation", "abs": "Recently, diffusion models have achieved great success in mono-channel audio generation.\nHowever, when it comes to stereo audio generation, the soundscapes often have a complex scene of multiple objects and directions.\nControlling stereo audio with spatial contexts remains challenging due to high data costs and unstable generative models. \nTo the best of our knowledge, this work represents the first attempt to address these issues.\nWe first construct a large-scale, simulation-based, and GPT-assisted dataset, BEWO-1M, with abundant soundscapes and descriptions even including moving and multiple sources.\nBeyond text modality, we have also acquired a set of images and rationally paired stereo audios through retrieval to advance multimodal generation. \nExisting audio generation models tend to generate rather random spatial audio. \nTo provide accurate guidance for Latent Diffusion Models, we introduce the SpatialSonic model utilizing spatial-aware encoders and azimuth state matrices to reveal reasonable spatial guidance. \nBy leveraging spatial guidance, our unified model not only achieves the objective of generating immersive and controllable spatial audio from text and image but also enables interactive audio generation during inference.\nFinally, under fair settings, we conduct subjective and objective evaluations on simulated and real-world data to compare our approach with prevailing methods. \nThe results demonstrate the effectiveness of our method, highlighting its capability to generate spatial audio that adheres to physical rules.\nOur demos are available at https://immersive-audio.github.io/. Our code, model, and dataset will be released soon.", "keywords": [ "audio generation", "multimodal learning", "stereo audio" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "qPw5D0Xahv", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Chao Yang", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Kun He", "gender": "Female", "institution": "Huazhong University of Sceince and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Weilong Jiang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Zhenyi Wang", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 13, "n_ref_uni": 28, "n_ref": 34, "n_ref_all": 45, "n_fig": 6, "n_tab": 3, "L_tab": 1343, "n_element_tab": 217, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1022, "n_element_tab_1": 171, "formula_len_all": 878, "formula_len_all_1": 1027, "len_all": 109883, "len_all_1": 70161, "len_abs": 967, "len_title": 142, "len_sents": 31780, "len_sents_1": 30026, "n_sents": 243, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 972, "title": "Minimax Based Fast-training Defense against Adversarial Policy in Two-player Competitive Games", "abs": "Adversarial policies have been shown to exploit vulnerabilities in agents during two-player competitive games, significantly undermining their performance. While existing approaches model the challenge of training robust policies in such environments as the search for Nash equilibrium points in the policy space, this often leads to substantial computational overhead. In this work, we propose MM-FATROL, a novel robust policy training method grounded in the Minimax Theorem, which significantly reduces computational overhead by efficiently identifying promising policy updates. We provide a formal analysis of the speedup achieved by our method. Extensive experiments demonstrate that MM-FATROL not only enhances efficiency but also surpasses the state-of-the-art method in terms of generalization and robustness. Additionally, we discuss the limitations of our approach and the challenges that remain in developing robust policies for more complex game environments.", "keywords": [ "Deep reinforcement learning", "adversarial policy", "adversarial defense", "Nash equilibrium" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "qPZaTqLee4", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Dongcheng Zhao", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Guobin Shen", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yi Zeng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yiting Dong", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "xiang he", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 26, "n_ref": 54, "n_ref_all": 72, "n_fig": 12, "n_tab": 7, "L_tab": 1216, "n_element_tab": 195, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 2411, "n_element_tab_1": 261, "formula_len_all": 642, "formula_len_all_1": 388, "len_all": 121537, "len_all_1": 62274, "len_abs": 1828, "len_title": 129, "len_sents": 38908, "len_sents_1": 26473, "n_sents": 296, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1354, "title": "Harnessing Task Overload for Scalable Jailbreak Attacks on Large Language Models", "abs": "Large Language Models (LLMs) remain vulnerable to jailbreak attacks that bypass their safety mechanisms. Existing attack methods are fixed or specifically tailored for certain models and cannot flexibly adjust attack strength, which is critical for generalization when attacking models of various sizes. We introduce a novel scalable jailbreak attack that preempts the activation of an LLM's safety policies by occupying its computational resources. Our method involves engaging the LLM in a resource-intensive preliminary task\u2014a Character Map lookup and decoding process\u2014before presenting the target instruction. \nBy saturating the model's processing capacity, we prevent the activation of safety protocols when processing the subsequent instruction. \nExtensive experiments on state-of-the-art LLMs demonstrate that our method achieves a high success rate in bypassing safety measures without requiring gradient access, manual prompt engineering. We verified our approach offers a scalable attack that quantifies attack strength and adapts to different model scales at the optimal strength. We shows safety policies of LLMs might be more susceptible to resource constraints.\nOur findings reveal a critical vulnerability in current LLM safety designs, highlighting the need for more robust defense strategies that account for resource-intense condition.", "keywords": [ "Jailbreak Attack", "Large Language Model", "Adversarial Attack" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "qPTFzmXVLd", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Cheol Jun Cho", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "David Chan", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Joonyong Park", "gender": "Male", "institution": "The University of Tokyo, Tokyo University", "country": "JP", "position": "MS student" }, { "name": "Rodolfo Corona", "gender": "unknown", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Yutong Bai", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "trevor darrell", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department", "country": "US", "position": "Professor" } ], "n_formula": 9, "n_formula_1": 3, "n_ref_uni": 58, "n_ref": 81, "n_ref_all": 112, "n_fig": 10, "n_tab": 6, "L_tab": 13680, "n_element_tab": 980, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 1558, "n_element_tab_1": 264, "formula_len_all": 428, "formula_len_all_1": 66, "len_all": 235204, "len_all_1": 61899, "len_abs": 1590, "len_title": 87, "len_sents": 70623, "len_sents_1": 31600, "n_sents": 384, "n_sents_1": 158, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1517, "title": "Analyzing the Language of Visual Tokens", "abs": "With the introduction of transformer-based models for vision and language tasks, such as LLaVA and Chameleon, there has been renewed interest in the discrete tokenized representation of images. These models often treat image patches as discrete tokens, analogous to words in natural language, learning joint alignments between visual and human languages. However, little is known about the statistical behavior of these visual languages\u2014whether they follow similar frequency distributions, grammatical structures, or topological alignments as natural languages. In this paper, we take a natural-language-centric approach to analyzing discrete visual languages and uncover striking similarities and fundamental differences. We demonstrate that, although visual languages adhere to Zipfian distributions, higher token innovation drives greater entropy and lower compression, with tokens predominantly representing object parts, indicating intermediate granularity. We also show that visual languages lack cohesive grammatical structures, leading to higher perplexity and weaker hierarchical organization compared to natural languages. Finally, we demonstrate that, while vision models align more closely with natural languages than other models, this alignment remains significantly weaker than the cohesion found within natural languages. Through these experiments, we demonstrate how understanding the statistical properties of discrete visual languages can inform the design of more effective computer vision models.", "keywords": [ "Tokenization", "Analysis", "Statistical Methods", "Vision and Language" ], "rating_list": [ 8, 3, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "qPH7lAyQgV", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Haotian Xu", "gender": "Male", "institution": "Stony Brook University", "country": "US", "position": "PhD student" }, { "name": "Tengfei Ma", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Assistant Professor" }, { "name": "Tsui-Wei Weng", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Lam M Nguyen", "gender": "Male", "institution": "IBM Research, Thomas J. Watson Research Center", "country": "US", "position": "Staff Research Scientist" } ], "n_formula": 18, "n_formula_1": 16, "n_ref_uni": 28, "n_ref": 58, "n_ref_all": 81, "n_fig": 8, "n_tab": 8, "L_tab": 1811, "n_element_tab": 231, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1847, "n_element_tab_1": 47, "formula_len_all": 878, "formula_len_all_1": 719, "len_all": 161767, "len_all_1": 62967, "len_abs": 1092, "len_title": 79, "len_sents": 49054, "len_sents_1": 30236, "n_sents": 320, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 31, "L_abs": 1099, "title": "Graph Concept Bottleneck Models", "abs": "Concept Bottleneck Models (CBMs) provide explicit interpretations for deep neural networks through concepts and allow intervention with concepts to adjust final predictions. Existing CBMs assume concepts are conditionally independent given labels and isolated from each other, ignoring the hidden relationships among concepts. However, the set of concepts in CBMs often has an intrinsic structure where concepts are generally correlated: changing one concept will inherently impact its related concepts. To mitigate this limitation, we propose **Graph CBMs**: a new variant of CBM that facilitates concept relationships by constructing latent concept graphs, which can be combined with CBMs to enhance model performance while retaining their interpretability. Empirical results on real-world image classification tasks demonstrate Graph CBMs are (1) superior in image classification tasks while providing more concept structure information for interpretability; (2) able to utilize concept graphs for more effective interventions; and (3) robust across different training and architecture settings.", "keywords": [ "Concept Bottleneck Models", "Interpretability", "Graphs", "Graph Neural Networks", "Structure Learning" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 2, 4, 2 ] }, { "paper_id": "qOqCXEXsX4", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Ashwinee Panda", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Postdoc" }, { "name": "Berivan Isik", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Prateek Mittal", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "Full Professor" }, { "name": "Tsachy Weissman", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Full Professor" }, { "name": "Xiangyu Qi", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 3, "n_ref_uni": 79, "n_ref": 154, "n_ref_all": 179, "n_fig": 3, "n_tab": 6, "L_tab": 2290, "n_element_tab": 127, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 728, "n_element_tab_1": 70, "formula_len_all": 0, "formula_len_all_1": 305, "len_all": 218712, "len_all_1": 66572, "len_abs": 1260, "len_title": 118, "len_sents": 50178, "len_sents_1": 32562, "n_sents": 332, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 987, "title": "Lottery Ticket Adaptation: Mitigating Destructive Interference in LLMs", "abs": "Existing methods for adapting large language models (LLMs) to new tasks are not suited to multi-task adaptation because they modify all the model weights--causing destructive interference between tasks. The resulting effects, such as catastrophic forgetting of earlier tasks, make it challenging to obtain good performance on multiple tasks at the same time. \nTo mitigate this, we propose Lottery Ticket Adaptation (LoTA), a sparse adaptation method that identifies and optimizes only a sparse subnetwork of the model. We evaluate LoTA on a wide range of challenging tasks such as instruction following, reasoning, math, and summarization. LoTA obtains better performance than full fine-tuning and low-rank adaptation (LoRA), and maintains good performance even after training on other tasks -- thus, avoiding catastrophic forgetting. By extracting and fine-tuning over \\emph{lottery tickets} (or \\emph{sparse task vectors}), LoTA also enables model merging over highly dissimilar tasks.", "keywords": [ "catastrophic forgetting", "adapter", "finetuning", "model merging", "sparsity", "lottery ticket" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "qOForsjh4q", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Eli Verwimp", "gender": "Male", "institution": "Department of Electrical Engineering, KU Leuven, Belgium, KU Leuven", "country": "", "position": "PhD student" }, { "name": "Guy Hacohen", "gender": "Male", "institution": "KU Leuven", "country": "", "position": "Postdoc" }, { "name": "Tinne Tuytelaars", "gender": "Female", "institution": "KU Leuven", "country": "BE", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 44, "n_ref": 61, "n_ref_all": 85, "n_fig": 13, "n_tab": 2, "L_tab": 1302, "n_element_tab": 84, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1009, "n_element_tab_1": 78, "formula_len_all": 469, "formula_len_all_1": 399, "len_all": 116481, "len_all_1": 57322, "len_abs": 1241, "len_title": 0, "len_sents": 36964, "len_sents_1": 27362, "n_sents": 281, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1252, "title": "Same Accuracy, Twice As Fast: Continual Learning Surpasses Retraining From Scratch", "abs": "Continual learning aims to enable models to adapt to new datasets without losing performance on previously learned data, often assuming prior data is no longer available. However, in many practical scenarios, both old and new data are accessible. In such cases, good performance on both datasets is typically achieved by abandoning the model trained on the previous data and re-training a new model from scratch on both datasets. This training from scratch is computationally expensive. In contrast, methods that leverage the previously trained model are worthy of investigation as they could significantly reduce computational costs. Our evaluation framework quantifies the computational savings of such methods while maintaining or exceeding the performance of training from scratch. We identify key optimization aspects - initialization, regularization, data selection, and hyper-parameters - that can each contribute to reducing computational costs. For each aspect, we propose effective first-step methods that already yield substantial computational savings. By combining these strategies, we achieve up to 2.7x reductions in computation time across various computer vision tasks, highlighting the potential for further advancements in this area.", "keywords": [ "continual learning" ], "rating_list": [ 3, 5, 3, 6, 3 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3, 4 ] }, { "paper_id": "qODJnX99hi", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Kazuhito Koishida", "gender": "unknown", "institution": "Microsoft Corporation", "country": "US", "position": "Researcher" }, { "name": "Saeed Amizadeh", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Sara Abdali", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Yinheng Li", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 70, "n_formula_1": 18, "n_ref_uni": 54, "n_ref": 75, "n_ref_all": 111, "n_fig": 14, "n_tab": 9, "L_tab": 1156, "n_element_tab": 94, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1129, "n_element_tab_1": 27, "formula_len_all": 9082, "formula_len_all_1": 1936, "len_all": 293946, "len_all_1": 70190, "len_abs": 2027, "len_title": 129, "len_sents": 89472, "len_sents_1": 34227, "n_sents": 632, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1689, "title": "Hierarchical Self-Attention: Generalizing Neural Attention Mechanics to Hierarchy", "abs": "Transformers and their attention mechanism have been revolutionary in the field of Machine Learning. While originally proposed for the language data, they quickly found their way to the image, video, graph, etc. data modalities with various signal geometries. Despite this versatility, generalizing the attention mechanism to scenarios where data is presented at different scales from potentially different modalities is not straightforward. The attempts to incorporate hierarchy and multi-modality within transformers are largely based on ad hoc heuristics, which are not seamlessly generalizable to similar problems with potentially different structures. To address this problem, in this paper, we take a fundamentally different approach: we first propose a mathematical construct to represent multi-modal, multi-scale data. We then mathematically derive the neural attention mechanics for the proposed construct from the first principle of entropy minimization. We show that the derived formulation is optimal in the sense of being the closest to the standard Softmax attention while incorporating the inductive biases originating from the hierarchical/geometric information of the problem. We further propose an efficient algorithm based on dynamic programming to compute our derived attention mechanism. By incorporating it within transformers, we show that the proposed hierarchical attention mechanism not only can be employed to train transformer models in hierarchical/multi-modal settings from scratch, but it can also be used to inject hierarchical information into classical, pre-trained transformer models post training, resulting in more efficient models in zero-shot manner.", "keywords": [ "Attention Mechanism", "Transformers", "Hierarchical Data", "Multi-modal Data", "Geometric Deep Learning" ], "rating_list": [ 5, 8, 8, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 4, 4, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "qO6dk9KfIp", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Daniel Rueckert", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Georgios Kaissis", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Johannes Kaiser", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Kristian Schwethelm", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 9, "n_ref_uni": 71, "n_ref": 100, "n_ref_all": 126, "n_fig": 15, "n_tab": 4, "L_tab": 939, "n_element_tab": 51, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1030, "n_element_tab_1": 8, "formula_len_all": 905, "formula_len_all_1": 594, "len_all": 179469, "len_all_1": 60974, "len_abs": 1040, "len_title": 121, "len_sents": 53229, "len_sents_1": 31238, "n_sents": 324, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1117, "title": "Laplace Sample Information: Data Informativeness Through a Bayesian Lens", "abs": "Accurately estimating the informativeness of individual samples in a dataset is an important objective in deep learning, as it can guide sample selection, which can improve model efficiency and accuracy by removing redundant or potentially harmful samples. \nWe propose $\\text{\\textit{Laplace Sample Information}}$ ($\\mathsf{LSI}$) measure of sample informativeness grounded in information theory widely applicable across model architectures and learning settings.\n$\\mathsf{LSI}$ leverages a Bayesian approximation to the weight posterior and the KL divergence to measure the change in the parameter distribution induced by a sample of interest from the dataset.\nWe experimentally show that $\\mathsf{LSI}$ is effective in ordering the data with respect to typicality, detecting mislabeled samples, measuring class-wise informativeness, and assessing dataset difficulty.\nWe demonstrate these capabilities of $\\mathsf{LSI}$ on image and text data in supervised and unsupervised settings.\nMoreover, we show that $\\mathsf{LSI}$ can be computed efficiently through probes and transfers well to the training of large models.", "keywords": [ "Sample informativeness", "Sample Information", "Sample Difficulty", "Long-tailed distribution", "Leave-one-out retraining", "KL Divergence" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "qNp86ByQlN", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Muhammad Khalid", "gender": "Male", "institution": "Cardiff University", "country": "GB", "position": "Postdoc" }, { "name": "Steven Schockaert", "gender": "Male", "institution": "Cardiff University", "country": "GB", "position": "Full Professor" } ], "n_formula": 55, "n_formula_1": 9, "n_ref_uni": 59, "n_ref": 133, "n_ref_all": 169, "n_fig": 16, "n_tab": 9, "L_tab": 3483, "n_element_tab": 421, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2170, "n_element_tab_1": 271, "formula_len_all": 4282, "formula_len_all_1": 608, "len_all": 291751, "len_all_1": 75196, "len_abs": 1439, "len_title": 117, "len_sents": 87862, "len_sents_1": 33884, "n_sents": 705, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1440, "title": "Systematic Relational Reasoning With Epistemic Graph Neural Networks", "abs": "Developing models that can learn to reason is a notoriously challenging problem. We focus on reasoning in relational domains, where the use of Graph Neural Networks (GNNs) seems like a natural choice. However, previous work has shown that regular GNNs lack the ability to systematically generalize from training examples on test graphs requiring longer inference chains, which fundamentally limits their reasoning abilities. A common solution relies on neuro-symbolic methods that systematically reason by learning rules, but their scalability is often limited and they tend to make unrealistically strong assumptions, e.g.\\ that the answer can always be inferred from a single relational path. We propose the Epistemic GNN (EpiGNN), a novel parameter-efficient and scalable GNN architecture with an epistemic inductive bias for systematic reasoning. Node embeddings in EpiGNNs are treated as epistemic states, and message passing is implemented accordingly. We show that EpiGNNs achieve state-of-the-art results on link prediction tasks that require systematic reasoning. Furthermore, for inductive knowledge graph completion, EpiGNNs rival the performance of state-of-the-art specialized approaches. Finally, we introduce two new benchmarks that go beyond standard relational reasoning by requiring the aggregation of information from multiple paths. Here, existing neuro-symbolic approaches fail, yet EpiGNNs learn to reason accurately.", "keywords": [ "Reasoning", "Graph Neural Networks", "Neuro-Symbolic methods", "Systematic Generalization", "Compositionality" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 3, 2 ] }, { "paper_id": "qMUtej58Pc", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Alexandros Potamianos", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Georgios Paraskevopoulos", "gender": "Male", "institution": "Athena Research and Innovation Center", "country": "GR", "position": "Researcher" }, { "name": "Vaggelis Dorovatas", "gender": "unknown", "institution": "National Technical University of Athens", "country": "GR", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 36, "n_ref": 47, "n_ref_all": 77, "n_fig": 12, "n_tab": 3, "L_tab": 727, "n_element_tab": 97, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 385, "n_element_tab_1": 40, "formula_len_all": 442, "formula_len_all_1": 152, "len_all": 126430, "len_all_1": 52437, "len_abs": 1851, "len_title": 131, "len_sents": 42406, "len_sents_1": 26145, "n_sents": 278, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1846, "title": "From Overconnectivity to Sparsity: Emulating Synaptic Pruning with Long Connections", "abs": "During brain development, an excess number of synapses are initially created, which are progressively eliminated through a process known as synaptic pruning. This procedure is activity-dependent, shaped by the brain's experiences. While creating an overabundance of synaptic connections only to later remove many might appear inefficient, research suggests that pruned networks demonstrate significant efficiency and robustness. Inspired by this biological process, we propose a neural network architecture utilizing long connections instead of traditional short residual connections. When long connections neural networks (LCNs) are trained with gradient descent, information is naturally \"pushed\" down to the first few layers, leading to a sparse network. Even more surprising is that this simple architectural modification leads to networks that exhibit behaviors similar to biological brain networks, namely: early overconnectivity to later sparsity,\nenhanced robustness to noise, efficiency in low-data settings and longer training times. Specifically, starting with a traditional neural network architecture with initial depth $d$ and $k$ connections, long connections are added from all layers to the last layer and summed up. During LCN training, 30-80% of the top layers become effective identity mappings as all relevant information is concentrated in the bottom layers. Pruning the top layers results in a refined network with a reduced depth $d'$ and final connections $k'$, achieving significant efficiencies without any loss in performance compared to residual baselines. We apply this architecture to various classification tasks and show that, in all experiments, the network converges to utilizing only a subset of the initially defined pre-training connections, and the amount of compression is dependent on the task complexity.", "keywords": [ "machine learning architectures", "sparsity", "residual connections", "redundancy", "long connections", "pruning", "synaptic pruning" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "qLxkXgmWwx", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Lifu Tu", "gender": "Male", "institution": "Salesforce AI Research", "country": "US", "position": "research scientist" }, { "name": "Rui Meng", "gender": "Male", "institution": "SalesForce Research", "country": "", "position": "Researcher" }, { "name": "Shafiq Joty", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Principal Researcher" }, { "name": "Yingbo Zhou", "gender": "unknown", "institution": "Salesforce Research", "country": "", "position": "Research Scientist" }, { "name": "Semih Yavuz", "gender": "Male", "institution": "SalesForce.com", "country": "", "position": "Research Scientist" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 25, "n_ref": 51, "n_ref_all": 79, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 354, "formula_len_all_1": 354, "len_all": 114355, "len_all_1": 56220, "len_abs": 1735, "len_title": 143, "len_sents": 30470, "len_sents_1": 27009, "n_sents": 255, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1743, "title": "Investigating Factuality in Long-Form Text Generation: The Roles of Self-Known and Self-Unknown", "abs": "Large language models (LLMs) have demonstrated strong capabilities in text understanding and generation. However, they often lack factuality, producing a mixture of true and false information, especially in long-form generation. In this work, we investigates the factuality of long-form text generation across various large language models (LLMs), including GPT-4, Gemini-1.5-Pro, Claude-3-Opus, Llama-3-70B, and Mistral. Our analysis reveals that factuality scores tend to decline in later sentences of the generated text, accompanied by a rise in the number of unsupported claims.\nFurthermore, we explore the effectiveness of different evaluation settings to assess whether LLMs can accurately judge the correctness of their own outputs: Self-Known (the percentage of supported atomic claims, decomposed from LLM outputs, that the corresponding LLMs judge as correct) and Self-Unknown (the percentage of unsupported atomic claims that the corresponding LLMs judge as incorrect). The results indicate that even advanced models like GPT-4 and Gemini-1.5-Pro fail to achieve perfect Self-Known scores, while their Self-Unknown scores remain notably above zero, reflecting ongoing uncertainty in their self-assessments.\nMoreover, we find a correlation between higher Self-Known scores and improved factuality, while higher Self-Unknown scores are associated with lower factuality. Interestingly, even without significant changes in the models' self-judgment (Self-Known and Self-Unknown), the number of unsupported claims can increases, likely as an artifact of long-form generation. These findings show the limitations of current LLMs in long-form generation, and provide valuable insights for improving factuality in long-form text generation.", "keywords": [ "long-form generation", "Factuality" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "qLh6Ufvnuc", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Bhaskar Mukhoty", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Postdoc" }, { "name": "Bin Gu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Hilal AlQuabeh", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" } ], "n_formula": 40, "n_formula_1": 31, "n_ref_uni": 20, "n_ref": 34, "n_ref_all": 41, "n_fig": 4, "n_tab": 11, "L_tab": 7836, "n_element_tab": 741, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2091, "n_element_tab_1": 190, "formula_len_all": 3685, "formula_len_all_1": 2469, "len_all": 116936, "len_all_1": 63692, "len_abs": 1622, "len_title": 152, "len_sents": 30872, "len_sents_1": 25038, "n_sents": 239, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1660, "title": "Improving Generalization and Robustness in SNNs Through Signed Rate Encoding and Sparse Encoding Attacks", "abs": "Rate-encoded spiking neural networks (SNNs) are known to offer superior adversarial robustness compared to direct-encoded SNNs but have relatively poor generalization on clean input. While the latter offers good generalization on clean input it suffers poor adversarial robustness under standard training. A key reason for this behaviour is the input noise introduced by the rate encoding, which encodes a pixel intensity with $T$ independent Bernoulli samples. To improve the generalization of rate-encoded SNNs, we propose the *signed rate encoding* (sRATE) that allows mean centering of the input and helps reduce the randomness introduced by the encoding, resulting in improved clean accuracy. In contrast to rate encoding where input restricted to $[0,1]^d$ is encoded in $\\\\{0,1\\\\}^{d\\times T}$, the signed rate encoding allows input in $[-1,1]^d$ to be encoded with spikes in $\\\\{-1,0,1\\\\}^{d\\times T}$, where positive (negative) inputs are encoded with positive (negative) spikes. We further construct efficient *Sparse Encoding Attack* (SEA) on standard and signed rate encoded input, which performs $l_0$-norm restricted adversarial attack in the discrete encoding space. We prove the theoretical optimality of the attack under the first-order approximation of the loss and compare it empirically with the existing attacks on the input space. Adversarial training performed with SEA, under signed rate encoding, offers superior adversarial robustness to the existing attacks and itself. Experiments conducted on standard datasets show the effectiveness of sign rate encoding in improving accuracy across all settings including adversarial robustness.", "keywords": [ "Spiking Neural Network", "Adversarial Examples", "Adversarial Robustness", "Rate Encoding" ], "rating_list": [ 5, 6, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 4 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "qLRaPfDPXK", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Bernhard Kainz", "gender": "Male", "institution": "AIBE", "country": "DE", "position": "Full Professor" }, { "name": "Chengqi Zang", "gender": "Male", "institution": "University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Weitong Zhang", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 14, "n_ref_uni": 48, "n_ref": 126, "n_ref_all": 148, "n_fig": 5, "n_tab": 9, "L_tab": 3195, "n_element_tab": 232, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 79, "n_element_tab_1": 24, "formula_len_all": 3022, "formula_len_all_1": 1727, "len_all": 219493, "len_all_1": 63251, "len_abs": 2013, "len_title": 126, "len_sents": 67456, "len_sents_1": 29054, "n_sents": 562, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1223, "title": "Truth or Deceit? A Bayesian Decoding Game Enhances Consistency and Reliability", "abs": "Large Language Models (LLMs) often produce outputs that -- though plausible -- can lack consistency and reliability, particularly in ambiguous or complex scenarios. Challenges arise from ensuring that outputs align with both factual correctness and human intent. This is problematic in existing approaches that trade improved consistency for lower accuracy. To mitigate these challenges, we propose a novel game-theoretic approach to enhance consistency and reliability during the decoding stage of LLM output generation. Our method models the decoding process as a multistage Bayesian decoding game. This ensures consistency through \\textit{Correctness Alignment} and enhances reliability via \\textit{Ambiguity Calibration}. The model dynamically converges to a consensus on the most reliable outputs and distinguishes \\{Valid}, Specious}\\} outputs without human feedback or additional training. Remarkably, our game design allows smaller models to outperform much larger models through game mechanisms (\\textit{e.g.} 78.1 LLaMA13B \\textit{vs} 76.6 PaLM540B), as well as integrating various LL strategies and models, demonstrating the potential of game-theoretic tools to improve the truthfulness and reliability of LLMs.", "keywords": [ "Mechanism Design", "Large Language Models (LLMs)", "Generative Modeling", "Alignment" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 1, 1, 1 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "qKu6KWPgxt", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Alejandro Newell", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Lahav Lipson", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Peiyun Hu", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Stephan Richter", "gender": "Male", "institution": "Apple", "country": "DE", "position": "Researcher" }, { "name": "Vladlen Koltun", "gender": "Male", "institution": "Apple", "country": "US", "position": "Distinguished Scientist" } ], "n_formula": 4, "n_formula_1": 0, "n_ref_uni": 46, "n_ref": 91, "n_ref_all": 107, "n_fig": 7, "n_tab": 10, "L_tab": 1350, "n_element_tab": 210, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 481, "n_element_tab_1": 58, "formula_len_all": 307, "formula_len_all_1": 0, "len_all": 167398, "len_all_1": 67260, "len_abs": 625, "len_title": 47, "len_sents": 54269, "len_sents_1": 32954, "n_sents": 428, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 626, "title": "CoMotion: Concurrent Multi-person 3D Motion", "abs": "We introduce an approach for tracking detailed 3D poses of multiple people from a single monocular camera stream. Our system maintains temporally coherent predictions in crowded scenes filled with difficult poses and occlusions. Rather than detect poses and associate them to current tracks, our model directly updates all tracked poses simultaneously given a new input image. We train on numerous single-image and video datasets with both 2D and 3D annotations to produce a model that matches the 3D pose estimation quality of state-of-the-art systems while performing faster and more accurate tracking on in-the-wild videos.", "keywords": [ "human pose estimation", "3d human pose", "tracking" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 2, 4, 4 ] }, { "paper_id": "qKgd7RaAem", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Levi Rauchwerger", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "MS student" }, { "name": "Ron Levie", "gender": "unknown", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "Assistant Professor" }, { "name": "Stefanie Jegelka", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" } ], "n_formula": 145, "n_formula_1": 25, "n_ref_uni": 42, "n_ref": 144, "n_ref_all": 169, "n_fig": 11, "n_tab": 3, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15082, "formula_len_all_1": 2704, "len_all": 351925, "len_all_1": 80766, "len_abs": 2065, "len_title": 141, "len_sents": 106215, "len_sents_1": 34494, "n_sents": 1066, "n_sents_1": 289, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1215, "title": "Generalization, Expressivity, and Universality of Graph Neural Networks on Attributed Graphs", "abs": "We analyze the universality and generalization of graph neural networks (GNNs) on attributed graphs, i.e., with node attributes. To this end, we propose pseudometrics over the space of all attributed graphs that describe the fine-grained expressivity of GNNs. Namely, GNNs are both Lipschitz continuous with respect to our pseudometrics and can separate attributed graphs that are distant in the metric. Moreover, we prove that the space of all attributed graphs is relatively compact with respect to our metrics. Based on these properties, we prove a universal approximation theorem for GNNs and generalization bounds for GNNs on any data distribution of attributed graphs. The proposed metrics compute the similarity between the structures of attributed graphs via a hierarchical optimal transport between computation trees. Our work extends and unites previous approaches which either derived theory only for graphs with no attributes, derived compact metrics under which GNNs are continuous but without separation power, or derived metrics under which GNNs are continuous and separate points but the space of graphs is not relatively compact, which prevents universal approximation and generalization analysis.", "keywords": [ "Graph Neural Networks", "Graphon Neural Networks", "Machine Learning Theory", "1-WL test for Graphons", "Generalization", "Expressivity", "Universal Approximation", "Wasserstein Distance", "Optimal Transport", "Computation Trees" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 4, 3, 4, 4 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "qKfzDc8Qiv", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Fu Wang", "gender": "Male", "institution": "University of Exeter", "country": "", "position": "PhD student" }, { "name": "Tianle Zhang", "gender": "Male", "institution": "University of Liverpool", "country": "GB", "position": "PhD student" }, { "name": "Xiaowei Huang", "gender": "Male", "institution": "University of Liverpool", "country": "GB", "position": "Full Professor" }, { "name": "Yanghao Zhang", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Postdoc" }, { "name": "Rronghui Mu", "gender": "Female", "institution": "University of Exeter", "country": "GB", "position": "Lecturer" }, { "name": "Wenjie Ruan", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "Senior Lecturer" } ], "n_formula": 35, "n_formula_1": 16, "n_ref_uni": 46, "n_ref": 97, "n_ref_all": 118, "n_fig": 6, "n_tab": 6, "L_tab": 3643, "n_element_tab": 151, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 1964, "n_element_tab_1": 21, "formula_len_all": 2975, "formula_len_all_1": 806, "len_all": 194323, "len_all_1": 66602, "len_abs": 1518, "len_title": 135, "len_sents": 68089, "len_sents_1": 32161, "n_sents": 464, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1531, "title": "Beyond Levels and Continuity: A New Statistical Method for DNNs Robustness Evaluation", "abs": "Evaluating the robustness of deep neural networks (DNNs) is crucial in safety-critical areas, driving research into methods that accurately measure and enhance their resilience against adversarial attacks, specifically from a statistical perspective due to scalability issues faced by deterministic methods. Existing approaches based on independent sampling usually fail to directly capture such instances due to their rarity. Hence in this work, we treat the existence of adversarial examples as a rare event, and propose an innovative statistical framework for assessing the adversarial robustness of DNNs, called REPP. Our approach redefines the problem of calculating the occurrence of adversarial examples as the exponential of the mixture of a Poisson random variable and some potential geometric random variables. We adopt the point process to develop a Minimum Variance Unbiased Estimator (MVUE) to accurately estimate the likelihood of encountering adversarial examples, with an upper bound of the true probability with high confidence. Unlike existing rare-event methods based on Multi-level Splitting, REPP does not require the inherent level concept or the continuity condition of the cumulative distribution function (CDF) within DNNs. This adaptation allows for practical application across both computer vision and natural language processing tasks. Experimental results demonstrate that our method is more flexible and effective, offering a more reliable robustness evaluation than existing statistical approaches.", "keywords": [ "Robustness; Point Process; Adversarial examples" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 1, 1, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "qK6U4Ahfms", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 8, "author_info_list": [ { "name": "Fengli Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jie Feng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Jingzhe Yuan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Jun Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Qingbin Zeng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yuwei Yan", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Zhiheng Zheng", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 12, "n_ref": 37, "n_ref_all": 59, "n_fig": 9, "n_tab": 5, "L_tab": 1198, "n_element_tab": 139, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 887, "n_element_tab_1": 112, "formula_len_all": 201, "formula_len_all_1": 154, "len_all": 98398, "len_all_1": 65448, "len_abs": 1758, "len_title": 130, "len_sents": 39493, "len_sents_1": 32895, "n_sents": 279, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1873, "title": "OpenCity: A Scalable Platform to Simulate Urban Activities with Massive LLM Agents", "abs": "Agent-based models (ABMs) have long been employed to explore how individual behaviors aggregate into complex societal phenomena in urban space. Unlike black-box predictive models, ABMs excel at explaining the micro-macro linkages that drive such emergent behaviors. The recent rise of Large Language Models (LLMs) has led to the development of LLM agents capable of simulating urban activities with unprecedented realism. However, scaling LLM agents to large city simulations presents significant challenges. Existing models are limited by the computational and communication costs of LLMs, compounded by the dynamic nature of urban environments that require continual updates to agent behavior. To address these limitations, we propose OpenCity, a scalable simulation platform optimized for both system and prompt efficiencies. Specifically, we propose a LLM request scheduler to reduce communication overhead by parallelizing requests through IO multiplexing. Besides, we deisgn a ``group-and-distill'' prompt optimization strategy minimizes redundancy by clustering agents with similar static attributes. Through experiments on six global cities, OpenCity achieves a 600-fold acceleration in simulation time per agent, a 70\\% reduction in LLM requests, and a 50\\% reduction in token usage. These improvements enable the simulation of 10,000 agents\u2019 daily activities in 1 hour on commodity hardware. Additionally, OpenCity establishes a benchmark for LLM agents, comparing simulated mobility behaviors, origin-destination flows, and segregation indices against real-world data. We believe our OpenCity platform provides a critical infrastructure to harness the power of LLMs for interdisciplinary studies in urban space, fostering the collective efforts of broader research communities. Code repo is available at https://anonymous.4open.science/r/Anonymous-OpenCity-42BD.", "keywords": [ "LLM Agent", "Large Language Model", "Urban Study" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "qK3XElJUbq", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Dong Li", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Jing Zhang", "gender": "Female", "institution": "Australian National University", "country": "AU", "position": "Lecturer" }, { "name": "Xuyang Shen", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Yiran Zhong", "gender": "Male", "institution": "MiniMax", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuchao Dai", "gender": "Male", "institution": "Northwestern Polytechnical University", "country": "CN", "position": "Professor" }, { "name": "Zhen Qin", "gender": "unknown", "institution": "TapTap", "country": "CN", "position": "Researcher" }, { "name": "yuxin mao", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 23, "n_ref_uni": 31, "n_ref": 122, "n_ref_all": 150, "n_fig": 5, "n_tab": 10, "L_tab": 2933, "n_element_tab": 380, "n_fig_1": 3, "n_tab_1": 9, "L_tab_1": 3062, "n_element_tab_1": 344, "formula_len_all": 1950, "formula_len_all_1": 1853, "len_all": 144581, "len_all_1": 60743, "len_abs": 1405, "len_title": 127, "len_sents": 33299, "len_sents_1": 21756, "n_sents": 277, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1414, "title": "You Only Scan Once: Efficient Multi-dimension Sequential Modeling with LightNet", "abs": "Linear attention mechanisms have gained prominence in causal language models due to their linear computational complexity and enhanced speed. However, the inherent decay mechanism in linear attention presents challenges when applied to multi-dimensional sequence modeling tasks, such as image processing and multi-modal learning. In these scenarios, the utilization of sequential scanning to establish a global receptive field necessitates multiple scans for multi-dimensional data, thereby leading to inefficiencies. This paper identifies the inefficiency caused by a \\enquote{multiplicative} linear recurrence and proposes an efficient alternative \\enquote{additive} linear recurrence to avoid the issue, as it can handle multi-dimensional data within a single scan. We further develop an efficient multi-dimensional sequential modeling framework called LightNet based on the new recurrence. Moreover, we present two new multi-dimensional linear relative positional encoding methods, MD-TPE and MD-LRPE to enhance the model's ability to discern positional information in multi-dimensional scenarios. Our empirical evaluations across various tasks, including image classification, image generation, bidirectional language modeling, and autoregressive language modeling, demonstrate the efficacy of LightNet, showcasing its potential as a versatile and efficient solution for multi-dimensional sequential modeling.", "keywords": [ "LightNet", "multi-dimensional sequential modeling" ], "rating_list": [ 8, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 4 ], "presentation_list": [ 3, 1, 1, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 2, 4, 5, 5 ] }, { "paper_id": "qJkCEcd50n", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Chhavi Yadav", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Kamalika Chaudhuri", "gender": "Female", "institution": "UC San Diego, University of California, San Diego", "country": "", "position": "Full Professor" }, { "name": "Ruihan Wu", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Postdoc" } ], "n_formula": 25, "n_formula_1": 13, "n_ref_uni": 45, "n_ref": 74, "n_ref_all": 93, "n_fig": 6, "n_tab": 5, "L_tab": 1659, "n_element_tab": 125, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1278, "n_element_tab_1": 102, "formula_len_all": 2942, "formula_len_all_1": 1100, "len_all": 141217, "len_all_1": 69368, "len_abs": 680, "len_title": 48, "len_sents": 43357, "len_sents_1": 33140, "n_sents": 330, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 690, "title": "Influence-based Attributions can be Manipulated", "abs": "Influence Functions are a standard tool for attributing predictions to training data in a principled manner and are widely used in applications such as data valuation and fairness. In this work, we present realistic incentives to manipulate influence-based attributions and investigate whether these attributions can be \\textit{systematically} tampered by an adversary. We show that this is indeed possible for logistic regression models trained on ResNet feature embeddings and standard tabular fairness datasets and provide efficient attacks with backward-friendly implementations. Our work raises questions on the reliability of influence-based attributions in adversarial circumstances.", "keywords": [ "Influence Functions", "Attack", "data valuation", "Adversary", "explanation" ], "rating_list": [ 3, 5, 1, 3, 3 ], "soundness_list": [ 1, 3, 1, 2, 2 ], "presentation_list": [ 2, 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 1, 2, 1 ], "confidence_list": [ 3, 5, 4, 3, 4 ] }, { "paper_id": "qIbbBSzH6n", "primary_area": "datasets and benchmarks", "n_author": 25, "author_info_list": [ { "name": "Alexander Xiong", "gender": "Male", "institution": "Rice University", "country": "", "position": "Undergrad student" }, { "name": "Andy Zhou", "gender": "Male", "institution": "Department of Computer Science, University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Chejian Xu", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Chengquan Guo", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Chenhui Zhang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Chulin Xie", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "PhD student" }, { "name": "Dan Hendrycks", "gender": "unknown", "institution": "UC Berkeley", "country": "", "position": "PhD student" }, { "name": "Dawn Song", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Francesco Pinto", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Postdoc" }, { "name": "Jeffrey Ziwei Tan", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Jiawei Zhang", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Lingzhi Yuan", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Intern" }, { "name": "Mintong Kang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "PhD student" }, { "name": "Xuandong Zhao", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Yi Zeng", "gender": "Male", "institution": "Virginia Tech", "country": "", "position": "PhD student" }, { "name": "Yu Gai", "gender": "unknown", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Yujin Potter", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Zhaorun Chen", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Zhen Xiang", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Assistant Professor" }, { "name": "Zhun Wang", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Zhuowen Yuan", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Zidi Xiong", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Zinan Lin", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Senior Researcher" }, { "name": "PEIYANG XU", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 7, "n_formula_1": 0, "n_ref_uni": 84, "n_ref": 186, "n_ref_all": 278, "n_fig": 16, "n_tab": 41, "L_tab": 32833, "n_element_tab": 3168, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 4062, "n_element_tab_1": 156, "formula_len_all": 558, "formula_len_all_1": 0, "len_all": 654084, "len_all_1": 64741, "len_abs": 2404, "len_title": 125, "len_sents": 219360, "len_sents_1": 29079, "n_sents": 1728, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1329, "title": "MMDT: Decoding the Trustworthiness and Safety of Multimodal Foundation Models", "abs": "Multimodal foundation models (MMFMs) play a crucial role in various applications, including autonomous driving, healthcare, and virtual assistants. However, several studies have revealed vulnerabilities in these models, such as generating unsafe content by text-to-image models. Existing benchmarks on multimodal models either predominantly assess the helpfulness of these models, or only focus on limited perspectives such as fairness and privacy. In this paper, we present the first unified platform, MMDT (Multimodal DecodingTrust), designed to provide a comprehensive safety and trustworthiness evaluation for MMFMs. Our platform assesses models from multiple perspectives, including safety, hallucination, fairness/bias, privacy, adversarial robustness, and out-of-distribution (OOD) generalization. We have designed various evaluation scenarios and red teaming algorithms under different tasks for each perspective to generate challenging data, forming a high-quality benchmark. We evaluate a range of multimodal models using MMDT, and our findings reveal a series of vulnerabilities and areas for improvement across these perspectives. This work introduces the first comprehensive and unique safety and trustworthiness evaluation platform for MMFMs, paving the way for developing safer and more reliable MMFMs and systems.", "keywords": [ "Trustworthiness", "multimodal foundation models", "safety", "hallucination", "OOD generalization", "bias/fairness", "privacy", "adversarial robustness" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "qIN5VDdEOr", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Andrew Miller", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Christina Heinze-Deml", "gender": "Female", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Jaya Narain", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Juyeon Heo", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Kwan Ho Ryan Chan", "gender": "Male", "institution": "University of Pennsylvania ", "country": "US", "position": "PhD student" }, { "name": "Oussama Elachqar", "gender": "Male", "institution": "Oumi", "country": "US", "position": "Researcher" }, { "name": "Shirley You Ren", "gender": "unknown", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Udhyakumar Nallasamy", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 37, "n_ref": 66, "n_ref_all": 94, "n_fig": 6, "n_tab": 8, "L_tab": 3879, "n_element_tab": 287, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1745, "n_element_tab_1": 162, "formula_len_all": 82, "formula_len_all_1": 82, "len_all": 151233, "len_all_1": 69500, "len_abs": 1340, "len_title": 104, "len_sents": 51882, "len_sents_1": 34206, "n_sents": 334, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1544, "title": "Do LLMs ``know'' internally when they follow instructions?", "abs": "Instruction-following is crucial for building AI agents with large language models (LLMs), as these models must adhere strictly to user-provided constraints and guidelines. \nHowever, LLMs often fail to follow even simple and clear instructions.\nTo improve instruction-following behavior and prevent undesirable outputs, a deeper understanding of how LLMs' internal states relate to these outcomes is required.\nIn this work, we investigate whether LLMs encode information in their representations that correlates with instruction-following success\u2014a property we term ``knowing internally''.\nOur analysis identifies a direction in the input embedding space, termed the instruction-following dimension, that predicts whether a response will comply with a given instruction.\nWe find that this dimension generalizes well across unseen tasks but not across unseen instruction types.\nWe demonstrate that modifying representations along this dimension improves instruction-following success rates compared to random changes, without compromising response quality.\nFurther investigation reveals that this dimension is more closely related to the phrasing of prompts rather than the inherent difficulty of the task or instructions. \nThis discovery also suggests explanations for why LLMs sometimes fail to follow clear instructions and why prompt engineering is often effective, even when the content remains largely unchanged. \nThis work provides insight into the internal workings of LLMs' instruction-following, paving the way for reliable LLM agents.", "keywords": [ "Instruction following", "Large language models", "Linear probing", "Representation engineering", "Interpretation" ], "rating_list": [ 5, 5, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 4, 3, 4 ], "confidence_list": [ 3, 4, 3, 3, 4 ] }, { "paper_id": "qIJenSdGbW", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Bai LiChen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Shitong Shao", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Zeke Xie", "gender": "Male", "institution": "HKUST(GZ)", "country": "CN", "position": "Assistant Professor" }, { "name": "zhiqiang xu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "CN", "position": "Assistant Professor" }, { "name": "zikai zhou", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 33, "n_ref": 61, "n_ref_all": 118, "n_fig": 18, "n_tab": 12, "L_tab": 4860, "n_element_tab": 625, "n_fig_1": 15, "n_tab_1": 5, "L_tab_1": 2522, "n_element_tab_1": 323, "formula_len_all": 1428, "formula_len_all_1": 670, "len_all": 174049, "len_all_1": 65531, "len_abs": 1895, "len_title": 122, "len_sents": 54853, "len_sents_1": 26665, "n_sents": 416, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1948, "title": "Noise Prompt Learning: Learning the Winning Tickets for Diffusion Sampling", "abs": "Text-to-image diffusion model is a popular paradigm that synthesizes personalized images by providing a text prompt and a random Gaussian noise. While people observe that some noises are winning tickets that can achieve better text-image alignment and higher human preference than others, we still lack a machine learning framework to obtain those winning noises. To learn winning noises for diffusion sampling, we mainly make three contributions in this paper. First, we identify a new concept termed the $\\textit{noise prompt}$, which aims at turning a random Gaussian noise into a winning noise ticket by adding a small desirable perturbation derived from the text prompt. Following the concept, we first formulate the $\\textit{noise prompt learning}$ framework that systematically learns \"prompted'' winning noise tickets associated with a text prompt for diffusion models. Second, we design a noise prompt data collection pipeline and collect a large-scale $\\textit{noise prompt dataset}$ (NPD) that contains 100k pairs of random noises and winning noises with the associated text prompts. With the prepared NPD as the training dataset, we trained a small $\\textit{noise prompt network}$ (NPNet) that can directly learn to transform a random noise ticket into a winning noise ticket. The learned winning noise perturbation can be considered as a kind of prompt for noise, as it is rich in semantic information and tailored to the given text prompt. Third, our extensive experiments demonstrate the impressive effectiveness and generalization of NPNet on improving the quality of synthesized images across various diffusion models, including SDXL, DreamShaper-xl-v2-turbo, and Hunyuan-DiT. Moreover, NPNet is a small and efficient controller that acts as a plug-and-play module with very limited additional inference and computational costs, as it just provides a winning noise instead of a random noise without accessing the original pipeline.", "keywords": [ "Noise Prompt", "Noise Prompt Learning", "Image Synthesis", "Diffusion Models" ], "rating_list": [ 5, 5, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "qI1gmHbs0Z", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Changrui Chen", "gender": "Male", "institution": "University of Warwick", "country": "GB", "position": "PhD student" }, { "name": "Jungong Han", "gender": "Male", "institution": "The University of Sheffield", "country": "GB", "position": "Full Professor" }, { "name": "Kurt Debattista", "gender": "Male", "institution": "University of Warwick", "country": "GB", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 50, "n_ref_all": 65, "n_fig": 5, "n_tab": 10, "L_tab": 1671, "n_element_tab": 159, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 623, "n_element_tab_1": 60, "formula_len_all": 769, "formula_len_all_1": 408, "len_all": 120532, "len_all_1": 56956, "len_abs": 1185, "len_title": 112, "len_sents": 32157, "len_sents_1": 25127, "n_sents": 282, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1189, "title": "ShuffleNorm: A Better Normalization for Semi-supervised Learning", "abs": "We identify critical challenges with normalisation layers commonly used in fully supervised learning when applied to semi-supervised settings. Specifically, batch normalisation (BN) can experience severe performance degradation when labelled and unlabelled data have mismatched label distributions, due to biased statistical estimation. This results in unstable gradients, hindering the model's ability to converge effectively. While group/layer normalisation (GN/LN) avoids these issues, it lacks the stochastic regularisation provided by BN, leading to weaker generalisation. Poor generalisation, in turn, produces low-quality pseudo-labels, exacerbating confirmation bias. To address these limitations, we propose novel normalisation techniques termed Shuffle Layer normalisation and Shuffle Group normalisation (SLN/SGN) that introduce controllable randomness into LN/GN without increasing model parameters, thus making semi-supervised learning more robust and effective. Through experiments across diverse datasets, including image, text, and audio modalities, we demonstrate that SLN/SGN significantly enhances the performance of state-of-the-art semi-supervised learning algorithms.", "keywords": [ "Semi-supervised Learning; Normalization" ], "rating_list": [ 5, 1, 6 ], "soundness_list": [ 3, 1, 3 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "qHpfxfnIq3", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Claire Yoon", "gender": "Female", "institution": "Scale AI", "country": "US", "position": "Intern" }, { "name": "Pranav Vishnu Raja", "gender": "Male", "institution": "Scale AI", "country": "", "position": "Researcher" }, { "name": "Sean M. Hendryx", "gender": "Male", "institution": "Scale AI", "country": "", "position": "Engineering Manager, Machine Learning" }, { "name": "Vaskar Nath", "gender": "Male", "institution": "Scale AI", "country": "US", "position": "Researcher" }, { "name": "Jane Yu", "gender": "Female", "institution": "Meta AI ", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 29, "n_ref": 64, "n_ref_all": 99, "n_fig": 9, "n_tab": 11, "L_tab": 5560, "n_element_tab": 204, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 752, "n_element_tab_1": 26, "formula_len_all": 162, "formula_len_all_1": 162, "len_all": 325622, "len_all_1": 54606, "len_abs": 2854, "len_title": 112, "len_sents": 75360, "len_sents_1": 26899, "n_sents": 546, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1631, "title": "ToolComp: A Multi-Tool Reasoning & Process Supervision Benchmark", "abs": "Despite recent advances in AI, the development of systems capable of executing complex, multi-step reasoning tasks involving multiple tools remains a significant challenge. Current benchmarks fall short in capturing the real-world complexity of tool-use reasoning, where verifying the correctness of not only the final answer but also the intermediate steps is important for evaluation, development, and identifying failures during inference time. To bridge this gap, we introduce ToolComp, a comprehensive benchmark designed to evaluate multi-step tool-use reasoning. ToolComp is developed through a collaboration between models and human annotators, featuring human-edited/verified prompts, final answers, and process supervision labels, allowing for the evaluation of both final outcomes and intermediate reasoning. Evaluation across six different model families demonstrates the challenging nature of our dataset, with the majority of models achieving less than 50% accuracy. Additionally, we generate synthetic training data to compare the performance of outcome-supervised reward models (ORMs) with process-supervised reward models (PRMs) to assess their ability to improve complex tool-use reasoning as evaluated by ToolComp. Our results show that PRMs generalize significantly better than ORMs, achieving a 19\\% and 11\\% improvement in rank@1 accuracy for ranking base and fine-tuned model trajectories, respectively. These findings highlight the critical role of process supervision in both the evaluation and training of AI models, paving the way for more robust and capable systems in complex, multi-step tool-use tasks.", "keywords": [ "tool-augmented LLMs", "process supervision", "reward models", "tool-use", "react", "benchmark", "prm" ], "rating_list": [ 3, 5, 8, 6, 5 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 2, 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "qHVUdP1EEU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Gang Hua", "gender": "Male", "institution": "Wormpex AI Research", "country": "US", "position": "Chief Scientist and Managing Director" }, { "name": "Jiaxin Lu", "gender": "Female", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Qixing Huang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 68, "n_ref": 154, "n_ref_all": 174, "n_fig": 12, "n_tab": 3, "L_tab": 1593, "n_element_tab": 262, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 28, "n_element_tab_1": 2, "formula_len_all": 583, "formula_len_all_1": 382, "len_all": 191007, "len_all_1": 69411, "len_abs": 1043, "len_title": 63, "len_sents": 48676, "len_sents_1": 33642, "n_sents": 382, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1032, "title": "Jigsaw++: Imagining Complete Shape Priors for Object Reassembly", "abs": "The automatic assembly problem has attracted increasing interest due to its complex challenges that involve 3D representation. This paper introduces Jigsaw++, a novel generative method designed to tackle the multifaceted challenges of reconstruction for the reassembly problem. Existing approach focusing primarily on piecewise information for both part and fracture assembly, often overlooking the integration of complete object prior. Jigsaw++ distinguishes itself by learning a category-agnostic shape prior of complete objects. It employs the proposed \"retargeting\" strategy that effectively leverages the output of any existing assembly method to generate complete shape reconstructions. This capability allows it to function orthogonally to the current methods. Through extensive evaluations on Breaking Bad dataset and PartNet, Jigsaw++ has demonstrated its effectiveness, reducing reconstruction errors and enhancing the precision of shape reconstruction, which sets a new direction for future reassembly model developments.", "keywords": [ "Fracture Reassembly", "Object Reassembly", "Generative Model" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "qHAblIFenP", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Akhilan Boopathy", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Brian Cheung", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Research Fellow" }, { "name": "Ila R Fiete", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Professor" }, { "name": "Jaedong Hwang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Pulkit Agrawal", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Zhang-Wei Hong", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 2, "n_ref_uni": 48, "n_ref": 125, "n_ref_all": 168, "n_fig": 9, "n_tab": 31, "L_tab": 50570, "n_element_tab": 11727, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2128, "n_element_tab_1": 395, "formula_len_all": 490, "formula_len_all_1": 54, "len_all": 378420, "len_all_1": 62513, "len_abs": 1747, "len_title": 151, "len_sents": 49453, "len_sents_1": 28100, "n_sents": 346, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1757, "title": "ImageNet-RIB Benchmark: Large Pre-Training Datasets Don't Guarantee Robustness after Fine-Tuning", "abs": "Highly performant large-scale pre-trained models promise to also provide a valuable foundation for learning specialized tasks, by fine-tuning the model to the desired task. \nBy starting from a good general-purpose model, the goal is to achieve both specialization in the target task and maintain robustness.\nTo assess the robustness of models to out-of-distribution samples after fine-tuning on downstream datasets,\nwe introduce a new robust fine-tuning benchmark, ImageNet-RIB (Robustness Inheritance Benchmark).\nThe benchmark consists of a set of related but distinct specialized (downstream) tasks; pre-trained models are fine-tuned on one task in the set and their robustness is assessed on the rest, iterating across all tasks for fine-tuning and assessment. \nWe find that the continual learning methods, EWC and LwF maintain robustness after fine-tuning though fine-tuning generally does reduce performance on generalization to related downstream tasks across models.\nNot surprisingly, models pre-trained on large and rich datasets exhibit higher initial robustness across datasets and suffer more pronounced degradation during fine-tuning. \nThe distance between the pre-training and downstream datasets, measured by optimal transport, predicts this performance degradation on the pre-training dataset. \nHowever, counterintuitively, model robustness after fine-tuning on related downstream tasks is the worst when the pre-training dataset is the richest and the most diverse. \nThis suggests that starting with the strongest foundation model is not necessarily the best approach for performance on specialist tasks.\nThe benchmark thus offers key insights for developing more resilient fine-tuning strategies and building robust machine learning models.", "keywords": [ "robust fine-tuning", "robustness", "transfer learning", "representation learning", "continual learning" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 4, 2, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 4, 3, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "qH6pzxPZ0d", "primary_area": "learning theory", "n_author": 7, "author_info_list": [ { "name": "Duong Minh Nguyen", "gender": "Male", "institution": "Pusan National University", "country": "KR", "position": "PhD student" }, { "name": "Khoi Hoang Do", "gender": "Male", "institution": "University of Dublin, Trinity College", "country": "IE", "position": "PhD student" }, { "name": "Trinh Van Chien", "gender": "Male", "institution": "Link\u00f6ping University", "country": "", "position": "PhD student" }, { "name": "Van Duc Nguyen", "gender": "Male", "institution": "Hanoi University of Science and Technology", "country": "VN", "position": "Associate Professor" }, { "name": "Zhaohui Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "L\u00ea Nam Kh\u00e1nh", "gender": "Male", "institution": "Hanoi University of Science and Technology", "country": "VN", "position": "Undergrad student" }, { "name": "Nguyen Hoang Tran", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "Associate Professor" } ], "n_formula": 91, "n_formula_1": 29, "n_ref_uni": 35, "n_ref": 64, "n_ref_all": 84, "n_fig": 8, "n_tab": 11, "L_tab": 8758, "n_element_tab": 430, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1341, "n_element_tab_1": 119, "formula_len_all": 11132, "formula_len_all_1": 1835, "len_all": 200117, "len_all_1": 69546, "len_abs": 1516, "len_title": 156, "len_sents": 52445, "len_sents_1": 26966, "n_sents": 531, "n_sents_1": 254, "venue": "ICLR2025", "decision": "withdraw", "L_title": 108, "L_abs": 1523, "title": "TOWARDS LAYER-WISE PERSONALIZED FEDERATED LEARNING: ADAPTIVE LAYER DISENTANGLEMENT VIA CONFLICTING GRADIENTS", "abs": "In personalized Federated Learning (pFL), high data heterogeneity can cause significant gradient divergence across devices, adversely affecting the learning process. This divergence, especially when gradients from different users form an obtuse angle during aggregation, can negate progress, leading to severe weight and gradient update degradation. To address this issue, we introduce a new approach to pFL design, namely Federated Learning with Layer-wise Aggregation via Gradient Analysis (FedLAG), utilizing the concept of gradient conflict at the layer level. Specifically, when layer-wise gradients of different clients form acute angles, those gradients align in the same direction, enabling updates across different clients toward identifying client-invariant features. Conversely, when layer-wise gradient pairs make create obtuse angles, the layers tend to focus on client-specific tasks. In hindsights, FedLAG assigns layers for personalization based on the extent of layer-wise gradient conflicts. Specifically, layers with gradient conflicts are excluded from the global aggregation process. The theoretical evaluation demonstrates that when integrated into other pFL baselines, FedLAG enhances pFL performance by a certain margin. Therefore, our proposed method achieves superior convergence behavior compared with other baselines. Extensive experiments show that our FedLAG outperforms several state-of-the-art methods and can be easily incorporated with many existing methods to further enhance performance.", "keywords": [ "Federated Learning", "Gradient Conflicts", "Negative Transfer", "Non IID" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 1, 3, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "qH5uyYCG2j", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Chunlei Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Dong Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Distinguished Scientist" }, { "name": "Hangting Chen", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Yongqi Wang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 21, "n_ref": 36, "n_ref_all": 52, "n_fig": 8, "n_tab": 7, "L_tab": 3044, "n_element_tab": 452, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1424, "n_element_tab_1": 150, "formula_len_all": 161, "formula_len_all_1": 161, "len_all": 110463, "len_all_1": 58105, "len_abs": 1409, "len_title": 130, "len_sents": 34276, "len_sents_1": 27443, "n_sents": 233, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1416, "title": "Controllable Text-to-Speech Synthesis with Masked-Autoencoded Style Representation", "abs": "Controllable text-to-speech (TTS) systems aim to manipulate various stylistic attributes of generated speech. Despite considerable research in this area, existing models that use natural language prompts as an interface often lack the ability for fine-grained control and face a scarcity of high-quality data. To address these challenges, we propose a two-stage style-controllable TTS system with language models, utilizing a masked-autoencoded style representation as an intermediary. In our approach, we employ a masked autoencoder to learn a content-disentangled style feature of speech, which is then discretized using a residual vector quantizer. In the first stage, an autoregressive transformer is used for the conditional generation of these style tokens from text and control signals. In the second stage, we generate codec tokens from both text and sampled style tokens. Experiments demonstrate that training the first-stage model on extensive datasets enhances the robustness of the two-stage model in terms of quality and content accuracy. Additionally, our model achieves superior control over attributes such as pitch and emotion. By selectively combining discrete labels and speaker embeddings, we can fully control the speaker\u2019s timbre and other stylistic information, or adjust attributes like pitch and emotion for a specified speaker. Audio samples are available at https://style-ar-tts.github.io.", "keywords": [ "controllable text-to-speech", "representation learning" ], "rating_list": [ 5, 3, 1, 6, 6 ], "soundness_list": [ 2, 2, 1, 3, 3 ], "presentation_list": [ 2, 2, 1, 4, 3 ], "contribution_list": [ 3, 2, 1, 3, 3 ], "confidence_list": [ 5, 4, 5, 3, 5 ] }, { "paper_id": "qGLzeD9GCX", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Jie Zhang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Jun Jiang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Kai Zeng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Kejiang Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Research Associate" }, { "name": "Nenghai Yu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Shuai Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Tianze Chang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Weiming Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 14, "n_ref": 52, "n_ref_all": 66, "n_fig": 3, "n_tab": 9, "L_tab": 2102, "n_element_tab": 127, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 1376, "n_element_tab_1": 103, "formula_len_all": 524, "formula_len_all_1": 500, "len_all": 105958, "len_all_1": 57950, "len_abs": 1361, "len_title": 121, "len_sents": 30354, "len_sents_1": 24498, "n_sents": 253, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1323, "title": "EditMark: Training-free and Harmless Watermark for Large Language Models", "abs": "Large Language Models (LLMs) have demonstrated remarkable capabilities, but their training requires extensive data and computational resources, rendering them valuable digital assets. Therefore, it is essential to watermark LLMs to protect their copyright and trace unauthorized use or resale. Existing methods for watermarking LLMs are mainly based on backdoors or knowledge injection, which require burdensome training or degrade the generation quality. To address these issues, we propose EditMark, a training-free and harmless watermarking method for LLMs based on model editing. \nWe observe LLM has diversity and can generate multiple logical and semantic correct answers to some open-ended questions. Therefore, we can use a watermark to generate a harmless mapping to control the LLM's answer to an open-ended question.\nInspired by this insight, EditMark involves generating a harmless mapping based on the watermark, selecting a secret key to generate watermarked inputs, and editing the outputs of LLM to align with the harmless mapping. \nExtensive experiments show that EditMark can embed 8-bit watermarks into LLMs within 2 minutes, with a watermark extraction success rate close to 100%. External experiments further demonstrate that EditMark has fidelity and is robust to model fine-tuning and editing attacks.", "keywords": [ "Watermarking", "Model Edit", "Robust" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "qGL6fE1lqd", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Diego Romeres", "gender": "unknown", "institution": "Mitsubishi Electric Research Labs", "country": "US", "position": "Principal Researcher" }, { "name": "Radu Corcodel", "gender": "unknown", "institution": "Mitsubishi Electric Research Labs", "country": "", "position": "Principal Researcher" }, { "name": "Siddarth Jain", "gender": "unknown", "institution": "Mitsubishi Electric Research Labs", "country": "", "position": "Researcher" }, { "name": "Anoop Cherian", "gender": "Male", "institution": "Mitsubishi Electric Research Labs", "country": "US", "position": "Principal Researcher" } ], "n_formula": 8, "n_formula_1": 4, "n_ref_uni": 39, "n_ref": 77, "n_ref_all": 124, "n_fig": 13, "n_tab": 7, "L_tab": 3377, "n_element_tab": 171, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 256, "n_element_tab_1": 28, "formula_len_all": 537, "formula_len_all_1": 155, "len_all": 256031, "len_all_1": 64382, "len_abs": 1991, "len_title": 127, "len_sents": 94754, "len_sents_1": 34279, "n_sents": 616, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 2001, "title": "LLMPhy: Complex Physical Reasoning Using Large Language Models and World Models", "abs": "Physical reasoning is an important skill needed for robotic agents when operating in the real world. However, solving such reasoning problems often involves hypothesizing and reflecting over complex multi-body interactions under the effect of a multitude of physical forces and thus learning all such interactions poses a significant hurdle for state-of-the-art machine learning frameworks, including large language models (LLMs). To study this problem, we propose a new physical reasoning task and a dataset, dubbed TraySim. Our task involves predicting the dynamics of several objects on a tray that is given an external impact -- the domino effect of the ensued object interactions and their dynamics thus offering a challenging yet controlled setup, with the goal of reasoning being to infer the stability of the objects after the impact. To solve this complex physical reasoning task, we present LLMPhy, a zero-shot black-box optimization framework that leverages the physics knowledge and program synthesis abilities of LLMs, and synergizes these abilities with the world models built into modern physics engines. Specifically, LLMPhy uses an LLM to generate code to iteratively estimate the physical hyperparameters of the system (friction, damping, layout, etc.) via an implicit analysis-by-synthesis approach using a (non-differentiable) simulator in the loop and uses the inferred parameters to imagine the dynamics of the scene towards solving the reasoning task.} To show the effectiveness of LLMPhy, we present experiments on our TraySim dataset to predict the steady-state poses of the objects. Our results show that the combination of the LLM and the physics engine leads to state-of-the-art zero-shot physical reasoning performance, while demonstrating superior convergence against standard black-box optimization methods and better estimation of the physical parameters. Further, we show that LLMPhy is capable of solving both continuous and discrete black-box optimization problems.", "keywords": [ "large language models", "physics simulators", "physical reasoning" ], "rating_list": [ 5, 3, 5, 3, 6 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3, 4 ], "contribution_list": [ 1, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "qG1S5eXMzx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Ming Yang", "gender": "Male", "institution": "Ant Group", "country": "US", "position": "Researcher" }, { "name": "Qingpei Guo", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Ruihan Xu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Shiliang Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 79, "n_ref_all": 102, "n_fig": 9, "n_tab": 7, "L_tab": 6474, "n_element_tab": 1018, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1857, "n_element_tab_1": 202, "formula_len_all": 114, "formula_len_all_1": 92, "len_all": 136098, "len_all_1": 58391, "len_abs": 921, "len_title": 113, "len_sents": 30482, "len_sents_1": 25981, "n_sents": 230, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 926, "title": "Rethinking the Impact of Heterogeneous Sublayers in Transformers", "abs": "Large language models (LLMs) using transformers have achieved state-of-the-art performance across a wide array of tasks.\nHowever, the sheer size and complexity of these models present both theoretical and practical challenges, e.g., interpretation of the model behavior and deployment of edge devices.\nIn this work, we revisit the architecture of transformers and propose a more granular understanding of the impacts of individual sublayers, i.e., Multi-Head Attention (MHA) and Feed-Forward Network (FFN).\nWe introduce a novel metric, normalized relative impact factor, that allows for progressive, heterogeneous layer pruning.\nThis metric calculates the relative impact factor of each sublayer on the overall performance, normalized by the number of parameters.\nOur experiments demonstrate that our approach can lead to a 20\\% reduction in parameters and a 37\\% inference speedup, while maintaining minimal performance loss.", "keywords": [ "Transformer", "Layer Pruning" ], "rating_list": [ 1, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "qG0WCAhZE0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Anh Khoa Nguyen Vu", "gender": "Male", "institution": "University of Information Technology, Vietnam National University", "country": "VN", "position": "Researcher" }, { "name": "Tam V. Nguyen", "gender": "Male", "institution": "University of Dayton", "country": "US", "position": "Associate Professor" }, { "name": "Thanh Duc Ngo", "gender": "Male", "institution": "University of Information Technology, Vietnam National University Ho Chi Minh City", "country": "VN", "position": "Lecturer" }, { "name": "Truong Quoc Truong", "gender": "Male", "institution": "University of Information Technology, Vietnam National University", "country": "VN", "position": "MS student" }, { "name": "Vinh-Tiep Nguyen", "gender": "Male", "institution": "University of Information Technology", "country": "VN", "position": "Lecturer" }, { "name": "Thanh-Toan Do", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 20, "n_ref": 33, "n_ref_all": 68, "n_fig": 10, "n_tab": 4, "L_tab": 4224, "n_element_tab": 429, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 2368, "n_element_tab_1": 405, "formula_len_all": 429, "formula_len_all_1": 428, "len_all": 151692, "len_all_1": 67641, "len_abs": 2000, "len_title": 113, "len_sents": 38172, "len_sents_1": 25737, "n_sents": 313, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1333, "title": "Multi-Perspective Data Augmentation for Few-shot Object Detection", "abs": "Recent few-shot object detection (FSOD) methods have focused on augmenting synthetic samples for novel classes, show promising results to the rise of diffusion models. However, the diversity of such datasets is often limited in representativeness because they lack awareness of typical and hard samples, especially in the context of foreground and background relationships. To tackle this issue, we propose a Multi-Perspective Data Augmentation (MPAD) framework. In terms of foreground-foreground relationships, we propose in-context learning for object synthesis (ICOS) with bounding box adjustments to enhance the detail and spatial information of synthetic samples. Inspired by the large margin principle, support samples play a vital role in defining class boundaries. Therefore, we design a Harmonic Prompt Aggregation Scheduler (HPAS) to mix prompt embeddings at each time step of the generation process in diffusion models, producing hard novel samples. For foreground-background relationships, we introduce a Background Proposal method (BAP) to sample typical and hard backgrounds. Extensive experiments on multiple FSOD benchmarks demonstrate the effectiveness of our approach. Our framework significantly outperforms traditional methods, achieving an average increase of $17.5\\%$ in nAP50 over the baseline on PASCAL VOC.", "keywords": [ "few-shot object detection", "controllable diffusion", "data augmentation" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "qFw2RFJS5g", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Emily Jin", "gender": "Female", "institution": "Department of Computer Science", "country": "", "position": "PhD student" }, { "name": "Ismail Ilkan Ceylan", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "Assistant Professor" }, { "name": "Linus Bao", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "MS student" }, { "name": "Matthias Lanzinger", "gender": "Male", "institution": "Technische Universit\u00e4t Wien", "country": "AT", "position": "Assistant Professor" }, { "name": "Michael M. Bronstein", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" } ], "n_formula": 28, "n_formula_1": 11, "n_ref_uni": 48, "n_ref": 126, "n_ref_all": 156, "n_fig": 5, "n_tab": 5, "L_tab": 2301, "n_element_tab": 237, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1683, "n_element_tab_1": 174, "formula_len_all": 2317, "formula_len_all_1": 900, "len_all": 162350, "len_all_1": 72734, "len_abs": 1079, "len_title": 110, "len_sents": 49816, "len_sents_1": 33042, "n_sents": 408, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1097, "title": "Homomorphism Counts as Structural Encodings for Graph Learning", "abs": "Graph Transformers are popular neural networks that extend the well-known Transformer architecture to the graph domain. These architectures operate by applying self-attention on graph nodes and incorporating graph structure through the use of positional encodings (e.g., Laplacian positional encoding) or structural encodings (e.g., random-walk structural encoding). The quality of such encodings is critical, since they provide the necessary \\emph{graph inductive biases} to condition the model on graph structure. In this work, we propose \\emph{motif structural encoding} (MoSE) as a flexible and powerful structural encoding framework based on counting graph homomorphisms. Theoretically, we compare the expressive power of MoSE to random-walk structural encoding and relate both encodings to the expressive power of standard message passing neural networks. Empirically, we observe that MoSE outperforms other well-known positional and structural encodings across a range of architectures, and it achieves state-of-the-art performance on a widely studied molecular property prediction dataset.", "keywords": [ "graph transformers", "structural encodings", "homomorphism counts", "expressivity" ], "rating_list": [ 6, 6, 8, 8, 6 ], "soundness_list": [ 4, 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 3, 3 ] }, { "paper_id": "qFeeJ2ZQiH", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Fei Yang", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ming-Ming Cheng", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Full Professor" }, { "name": "Qibin Hou", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Xialei Liu", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Yusong Hu", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "PhD student" }, { "name": "liangzc", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "MS student" } ], "n_formula": 19, "n_formula_1": 20, "n_ref_uni": 58, "n_ref": 105, "n_ref_all": 114, "n_fig": 6, "n_tab": 4, "L_tab": 1386, "n_element_tab": 142, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 698, "n_element_tab_1": 88, "formula_len_all": 1530, "formula_len_all_1": 1650, "len_all": 123194, "len_all_1": 57013, "len_abs": 1009, "len_title": 105, "len_sents": 30122, "len_sents_1": 25944, "n_sents": 208, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1012, "title": "KAC: Kolmogorov-Arnold Classifier for Continual Learning", "abs": "Continual learning requires models to train continuously across consecutive tasks without forgetting. Most existing methods utilize linear classifiers, which struggle to maintain a stable classification space while learning new tasks. Inspired by the success of Kolmogorov-Arnold Networks (KAN) in preserving learning stability during simple continual regression tasks, we set out to explore their potential in more complex continual learning scenarios. In this paper, we introduce the Kolmogorov-Arnold Classifier (KAC), a novel classifier developed for continual learning based on the KAN structure. We delve into the impact of KAN's spline functions and introduce Radial Basis Functions (RBF) for improved compatibility with continual learning. We replace linear classifiers with KAC in several recent approaches and conduct experiments across various continual learning benchmarks, all of which demonstrate performance improvements, highlighting the effectiveness and robustness of KAC in continual learning.", "keywords": [ "continual learning", "class incremental learning" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "qFZnAC4GHR", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Denis C. Shields", "gender": "Male", "institution": "University College Dublin", "country": "IE", "position": "Full Professor" }, { "name": "Raul Fernandez-Diaz", "gender": "Male", "institution": "International Business Machines", "country": "IE", "position": "PhD student" }, { "name": "Vanessa L\u00f3pez", "gender": "Female", "institution": "International Business Machines", "country": "IE", "position": "Researcher" }, { "name": "Thanh Lam Hoang", "gender": "Male", "institution": "International Business Machines", "country": "IE", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 0, "n_ref_uni": 55, "n_ref": 111, "n_ref_all": 136, "n_fig": 12, "n_tab": 1, "L_tab": 244, "n_element_tab": 25, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 700, "n_element_tab_1": 30, "formula_len_all": 469, "formula_len_all_1": 0, "len_all": 165239, "len_all_1": 62121, "len_abs": 3563, "len_title": 147, "len_sents": 54444, "len_sents_1": 31263, "n_sents": 332, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 98, "L_abs": 1789, "title": "A new framework for evaluating model out-of-distribution generalisation for the biochemical domain", "abs": "Quantifying model generalization to out-of-distribution data has been a longstanding challenge in machine learning. Addressing this issue is crucial for leveraging machine learning in scientific discovery, where models must generalize to new molecules or materials. Current methods typically split data into train and test sets using various criteria \u2014 temporal, sequence identity, scaffold, or random cross-validation \u2014 before evaluating model performance. However, with so many splitting criteria available, existing approaches offer limited guidance on selecting the most appropriate one, and they do not provide mechanisms for incorporating prior knowledge about the target deployment distribution(s).\n\nTo tackle this problem, we have developed a novel metric, AU-GOOD, which quantifies expected model performance under conditions of increasing dissimilarity between train and test sets, while also accounting for prior knowledge about the target deployment distribution(s), when available. This metric is broadly applicable to biochemical entities, including proteins, small molecules, nucleic acids, or cells; as long as a relevant similarity function is defined for them. Recognizing the wide range of similarity functions used in biochemistry, we propose criteria to guide the selection of the most appropriate metric for partitioning. We also introduce a new partitioning algorithm that generates more challenging test sets, and we propose statistical methods for comparing models based on AU-GOOD.\n\nFinally, we demonstrate the insights that can be gained from this framework by applying it to two different use cases: developing predictors for pharmaceutical properties of small molecules, and using protein language models as embeddings to build biophysical property predictors.", "keywords": [ "Machine learning evaluation", "AI4Science", "Biochemistry", "Proteins", "Small molecules", "Protein Language Models" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 2, 3, 4 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "qDeEsfAb1j", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Alexander Capstick", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "PhD student" }, { "name": "Francesca Palermo", "gender": "Female", "institution": "Imperial College London, Imperial College London", "country": "", "position": "Postdoc" }, { "name": "Tianyu Cui", "gender": "unknown", "institution": "Johnson and Johnson", "country": "GB", "position": "Researcher" }, { "name": "Payam M. Barnaghi", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 46, "n_ref": 104, "n_ref_all": 181, "n_fig": 15, "n_tab": 15, "L_tab": 1527, "n_element_tab": 86, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 2328, "n_element_tab_1": 72, "formula_len_all": 1124, "formula_len_all_1": 1014, "len_all": 225070, "len_all_1": 66353, "len_abs": 1052, "len_title": 113, "len_sents": 88563, "len_sents_1": 31082, "n_sents": 628, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1055, "title": "Training Neural Networks on Data Sources with Unknown Reliability", "abs": "When data is generated by multiple sources, conventional training methods update models assuming equal reliability for each source and do not consider their individual data quality during training. However, in many applications, sources have varied levels of reliability that can have negative effects on the performance of a neural network. A key issue is that often the quality of data for individual sources is not known during training. Focusing on supervised learning, we aim to train neural networks on each data source for a number of steps proportional to the source's estimated relative reliability, by using a dynamic weighting. This way, we allow training on all sources during the warm-up, and reduce learning on less reliable sources during the final training stages, when it has been shown models overfit to noise. We show through diverse experiments, this can significantly improve model performance when trained on mixtures of reliable and unreliable data sources, and maintain performance when models are trained on reliable sources only.", "keywords": [ "Noisy data", "data sources", "learning with noise" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "qDSfOQBrOD", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Shidong Shang", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "", "position": "Undergrad student" }, { "name": "Wei Xiao", "gender": "Male", "institution": "Tencent Ethereal Audio Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Xusheng Yang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yuexian Zou", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 4, "n_ref_uni": 43, "n_ref": 62, "n_ref_all": 89, "n_fig": 4, "n_tab": 10, "L_tab": 3062, "n_element_tab": 353, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2225, "n_element_tab_1": 207, "formula_len_all": 383, "formula_len_all_1": 201, "len_all": 170754, "len_all_1": 69883, "len_abs": 1525, "len_title": 155, "len_sents": 51392, "len_sents_1": 31366, "n_sents": 435, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1532, "title": "VChangeCodec: A High-efficiency Neural Speech Codec with Built-in Voice Changer for Real-time Communication", "abs": "Neural speech codecs (NSCs) enable high-quality real-time communication (RTC)\nat low bit rates, making them efficient for bandwidth-constrained environments.\nHowever, customizing or modifying the timbre of transmitted voices still relies on\nseparate voice conversion (VC) systems, creating a gap in fully integrated systems\nthat can simultaneously optimize efficient transmission and streaming VC with no\nadditional latency. In this paper, we propose a high-efficiency VChangeCodec,\nwhich integrates the Voice Changer model directly into the speech Codec. This\ndesign seamlessly switches between the original voice mode and customized voice\nchange mode in real-time. Specifically, leveraging the target speaker\u2019s embedding,\nwe incorporate a lightweight causal projection network within the encoding module\nof VChangeCodec to adapt timbre at the token level. These adapted tokens are\nquantized and transmitted to the decoding module, to generate the converted speech\nof the target speaker. The integrated framework achieves an ultra-low latency of\njust 40 ms and requires fewer than 1 million parameters, making it ideal for RTC\nscenarios such as online conferencing. Our comprehensive evaluations, including\nsubjective listening tests and objective performance assessments, demonstrate that\nVChangeCodec excels in timbre adaptation capabilities compared to state-of-the-art (SOTA) VC models. We are confident that VChangeCodec provides an efficient\nand flexible framework for RTC systems, tailored to specific operator requirements.", "keywords": [ "Real-time communication", "Neural Speech codec", "Voice conversion" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "qDFpNXnuYK", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Chen Cecilia Liu", "gender": "unknown", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Iryna Gurevych", "gender": "Female", "institution": "Institute for Computer Science, Artificial Intelligence and Technology", "country": "BG", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 58, "n_ref": 118, "n_ref_all": 153, "n_fig": 16, "n_tab": 10, "L_tab": 1691, "n_element_tab": 218, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 4326, "n_element_tab_1": 75, "formula_len_all": 405, "formula_len_all_1": 242, "len_all": 201414, "len_all_1": 61187, "len_abs": 1493, "len_title": 150, "len_sents": 51302, "len_sents_1": 27332, "n_sents": 351, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1497, "title": "Early Period of Training Impacts Adaptation for Out-of-Distribution Generalization: An Empirical Study", "abs": "Prior research shows that differences in the early period of neural network training significantly impact the performance of in-distribution (ID) data of tasks. Yet, the implications of early learning dynamics on out-of-distribution (OOD) generalization remain poorly understood, primarily due to the complexities and limitations of existing analytical techniques. In this work, we investigate the relationship between learning dynamics, OOD generalization under covariate shift and the early period of neural network training. We utilize the trace of Fisher Information and sharpness, focusing on gradual unfreezing (i.e., progressively unfreezing parameters during training) as our methodology for investigation. Through a series of empirical experiments, we show that 1) changing the number of trainable parameters during the early period of training via gradual unfreezing can significantly improve OOD results; 2) the trace of Fisher Information and sharpness can be used as indicators for the removal of gradual unfreezing during the early period of training for better OOD generalization. Our experiments on both image and text data show that the early period of training is a general phenomenon that can provide Pareto improvements in ID and OOD performance with minimal complexity. Our work represents a first step towards understanding how early learning dynamics affect neural network OOD generalization under covariate shift and suggests a new avenue to improve and study this problem.", "keywords": [ "Early period of training", "fine-tuning", "parameter freezing", "transfer learning" ], "rating_list": [ 6, 5, 3, 5, 6 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 1, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "qDCkEHN3m8", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Badong Chen", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Junran Pan", "gender": "unknown", "institution": "Tianjin University", "country": "CN", "position": "Undergrad student" }, { "name": "Shuangming Yang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 36, "n_formula_1": 28, "n_ref_uni": 26, "n_ref": 33, "n_ref_all": 52, "n_fig": 4, "n_tab": 8, "L_tab": 1974, "n_element_tab": 124, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 967, "n_element_tab_1": 69, "formula_len_all": 1865, "formula_len_all_1": 1262, "len_all": 123045, "len_all_1": 56792, "len_abs": 2264, "len_title": 174, "len_sents": 40031, "len_sents_1": 25619, "n_sents": 310, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 125, "L_abs": 2277, "title": "Robust Spike-based Decoupled Federated Information Bottleneck Learning with Spiking Neural Network under System Heterogeneity", "abs": "As embedded devices become increasingly prevalent in intelligent systems, low-power system in resource-constrained environments has emerged as a key challenge. Spiking neural networks (SNNs), with their sparse and event-driven computation, have shown great potential as a low-power candidate for embedded devices. In federated learning scenarios, where multiple energy-constrained devices collaborate, adopting efficient SNN models with effective training methods is critical. However, research on training SNNs within federated learning systems is still very limited, particularly in terms of how to achieve both energy efficiency and robustness under system heterogeneity. This gap presents a significant opportunity for further exploration of SNNs in distributed learning settings. In this paper, we investigate a significant and innovative problem in robust spike-based federated learning, particularly in the presence of noise, and system heterogeneity. We majorly consider two types of system heterogeneity in this study, including data and client participation heterogeneity. To address this, we propose a novel federated learning framework, spike-based decoupled federated information-bottleneck learning (SDFIL), to enable robust, low-power federated learning through SNNs under system heterogeneity. Specifically, we design a decoupled information bottleneck principle tailored for local SNN training to maximize the mutual information between ground truth and model predictions while minimizing mutual information between intermediate representations. This method effectively minimizes the impact of outliers in non-independent and identically distributed (non-IID) data on model updates, thereby enhancing the performance of federated SNNs, resulting in enhanced robustness and reduced sensitivity to outliers. We evaluate the proposed SDFIL algorithm across a variety of settings, including different noise levels and varying degrees of system heterogeneity. The experimental results indicate that SDFIL demonstrates superior robustness compared to competing methods and generally achieves an improvement in overall accuracy of 5\\% to 10\\%. Additionally, it can achieve up to 7.7\u00d7 higher energy efficiency compared to traditional artificial neural networks (ANNs).", "keywords": [ "Spiking neural network", "federated learning", "spike-based learning", "neuromorphic computing" ], "rating_list": [ 3, 1, 6, 3, 3 ], "soundness_list": [ 2, 1, 3, 2, 2 ], "presentation_list": [ 2, 1, 3, 2, 2 ], "contribution_list": [ 1, 2, 3, 2, 1 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "qC3pfTGOxz", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Chengjie Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Chengming Xu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Donghao Luo", "gender": "Male", "institution": "Tencent YouTu Lab", "country": "CN", "position": "researcher" }, { "name": "Na Zhang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Wenhui Han", "gender": "unknown", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Xiaobin Hu", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Yanwei Fu", "gender": "Male", "institution": "Fudan University,", "country": "CN", "position": "Professor" } ], "n_formula": 13, "n_formula_1": 15, "n_ref_uni": 13, "n_ref": 24, "n_ref_all": 36, "n_fig": 9, "n_tab": 1, "L_tab": 95, "n_element_tab": 14, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 95, "n_element_tab_1": 14, "formula_len_all": 805, "formula_len_all_1": 800, "len_all": 89566, "len_all_1": 57845, "len_abs": 1468, "len_title": 133, "len_sents": 30604, "len_sents_1": 27705, "n_sents": 217, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1450, "title": "Enhancing Human Body Generation in Diffusion Models with Dual-Level Prior Knowledge", "abs": "The development of diffusion models (DMs) has greatly enhanced text-to-image generation, outperforming previous methods like generative adversarial networks (GANs) in terms of image quality and text alignment. However, accurately generating human body images remains challenging, often resulting in disproportionate figures and anatomical errors, which limits their practical applications in areas such as portrait generation. While previous methods such as HcP have shown promising results, limitations including wrongly kept prior, insufficient human-related knowledge, and limited generalization ability still exist due to the specific design of fully-supervised learning with only pose-related information. In this study, we introduce a novel method to enhance pretrained diffusion models for realistic human body generation by incorporating dual-level human prior knowledge. Our approach involves learning shape-level details with the human-related tokens in the original prompts, and learning pose-level prior by adding a learnable pose-aware token to each text prompt. We use a two-stage training strategy to rectify the cross attentions with a bind-then-generalize process, leveraging multiple novel objectives along with adversarial training. Our extensive experiments show that this method significantly improves the ability of SD1.5 and SDXL pretrained models to generate human bodies, reducing deformities and enhancing practical utility.", "keywords": [ "diffusion models", "human body generation" ], "rating_list": [ 5, 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "qBpYqQUFPx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 14, "author_info_list": [ { "name": "Bailu Ding", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Baotong Lu", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Chen Chen", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Chengruidong Zhang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Di Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Fan Yang", "gender": "Male", "institution": "Research, Microsoft", "country": "", "position": "Senior Principal Researcher" }, { "name": "Huiqiang Jiang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "RSDE" }, { "name": "Kai Zhang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Lili Qiu", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Principal Researcher" }, { "name": "Meng Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Qi Chen", "gender": "Female", "institution": "Microsoft Research", "country": "CA", "position": "Principal Researcher" }, { "name": "Qianxi Zhang", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Yuqing Yang", "gender": "unknown", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Zhenhua Han", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 2, "n_ref_uni": 47, "n_ref": 95, "n_ref_all": 129, "n_fig": 14, "n_tab": 11, "L_tab": 5558, "n_element_tab": 485, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2046, "n_element_tab_1": 155, "formula_len_all": 925, "formula_len_all_1": 194, "len_all": 186697, "len_all_1": 65300, "len_abs": 1494, "len_title": 80, "len_sents": 52379, "len_sents_1": 30953, "n_sents": 342, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1501, "title": "RetrievalAttention: Accelerating Long-Context LLM Inference via Vector Retrieval", "abs": "Transformer-based Large Language Models (LLMs) have become increasingly important. However, due to the quadratic time complexity of attention computation, scaling LLMs to longer contexts incurs extremely slow inference speed and high GPU memory consumption for caching key-value (KV) vectors. This paper proposes RetrievalAttention, a training-free approach to both accelerate attention computation and reduce GPU memory consumption. By leveraging the dynamic sparsity of attention mechanism, RetrievalAttention proposes to build approximate nearest neighbour search (ANNS) indexes for KV vectors in CPU memory and retrieve the most relevant ones through vector search during generation. Unfortunately, we observe that the off-the-shelf ANNS indexes are often ineffective for such retrieval tasks due to the out-of-distribution (OOD) between query vectors and key vectors in the attention mechanism. RetrievalAttention addresses the OOD challenge by designing an attention-aware vector search algorithm that can adapt to the distribution of query vectors. Our evaluation demonstrates that RetrievalAttention achieves near full attention accuracy while only requiring access to 1\u20133% of the data. This leads to a significant reduction in the inference cost of long-context LLMs, with a much lower GPU memory footprint. In particular, RetrievalAttention only needs a single NVIDIA RTX4090 (24GB) to serve 128K tokens for LLMs with 8B parameters, which is capable of generating one token in 0.188 seconds.", "keywords": [ "Large Language Model; Efficient Inference; Long Context; Vector Retrieval" ], "rating_list": [ 3, 3, 8, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 5, 2, 4 ] }, { "paper_id": "qBSzdiKVcK", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Anastasis Kratsios", "gender": "Non-Binary", "institution": "Vector Institute", "country": "CA", "position": "Researcher" }, { "name": "Raeid Saqur", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Xuwei Yang", "gender": "Male", "institution": "McMaster University", "country": "", "position": "Postdoc" }, { "name": "Yannick Limmer", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Blanka N Horvath", "gender": "Female", "institution": "Oxford University", "country": "", "position": "Associate Professor" } ], "n_formula": 203, "n_formula_1": 20, "n_ref_uni": 76, "n_ref": 135, "n_ref_all": 181, "n_fig": 4, "n_tab": 9, "L_tab": 1668, "n_element_tab": 252, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 3039, "n_element_tab_1": 33, "formula_len_all": 24722, "formula_len_all_1": 2534, "len_all": 332831, "len_all_1": 72860, "len_abs": 1526, "len_title": 125, "len_sents": 75998, "len_sents_1": 29492, "n_sents": 781, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1668, "title": "Reality Only Happens Once: Single-path Generalization Bounds for Transformers", "abs": "One of the inherent challenges in deploying transformers on time series is that \\emph{reality only happens once}; namely, one typically only has access to a single trajectory of the data-generating process comprised of non-i.i.d.\\ observations. We derive non-asymptotic statistical guarantees in this setting through bounds on the \\textit{generalization} of a transformer network at a future-time $t$, given that it has been trained using $N\\le t$ observations from a single perturbed trajectory of a {bounded and exponentially ergodic} Markov process. We obtain a generalization bound which effectively converges at the rate of $\\mathcal{O}(1/\\sqrt{N})$. Our bound depends explicitly on the activation function ($\\operatorname{Swish}$, $\\operatorname{GeLU}$, or $\\tanh$ are considered), the number of self-attention heads, depth, width, and norm-bounds defining the transformer architecture. Our bound consists of three components: (I) The first quantifies the gap between the stationary distribution of the data-generating Markov process and its distribution at time $t$, this term converges exponentially to $0$. (II) The next term encodes the complexity of the transformer model and, given enough time, eventually converges to $0$ at the rate $\\mathcal{O}(\\log(N)^r/\\sqrt{N})$ for any $r>0$. (III) The third term guarantees that the bound holds with probability at least $1-\\delta$, and converges at a rate of $\\mathcal{O}(\\sqrt{\\log(1/\\delta)}/\\sqrt{N})$. \nExample of (non i.i.d.) data-generating processes which we can treat are the projection of several SDEs onto a compact convex set $C$, and bounded Markov processes satisfying a log-Sobolev inequality.", "keywords": [ "learning theory", "transformers", "generalization bounds", "llms" ], "rating_list": [ 3, 5, 3, 8, 6 ], "soundness_list": [ 3, 3, 2, 4, 4 ], "presentation_list": [ 2, 4, 2, 3, 4 ], "contribution_list": [ 1, 3, 1, 3, 2 ], "confidence_list": [ 4, 3, 3, 3, 2 ] }, { "paper_id": "qBKA2844I4", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Holakou Rahmanian", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Lexing Ying", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Professor" }, { "name": "Michael Shavlovsky", "gender": "Male", "institution": "Amazon", "country": "US", "position": "PhD student" }, { "name": "Tesi Xiao", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Yinuo Ren", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 32, "n_formula_1": 21, "n_ref_uni": 56, "n_ref": 101, "n_ref_all": 144, "n_fig": 19, "n_tab": 6, "L_tab": 1214, "n_element_tab": 128, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 461, "n_element_tab_1": 34, "formula_len_all": 1883, "formula_len_all_1": 1075, "len_all": 200005, "len_all_1": 70062, "len_abs": 1400, "len_title": 116, "len_sents": 58199, "len_sents_1": 31553, "n_sents": 375, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1416, "title": "HyperDPO: Hypernetwork-based Multi-Objective Fine-Tuning Framework", "abs": "In LLM alignment and many other ML applications, one often faces the *Multi-Objective Fine-Tuning (MOFT)* problem, *i.e.* fine-tuning an existing model with datasets labeled w.r.t. different objectives simultaneously. To address the challenge, we propose the *HyperDPO* framework, a conditioned one-shot fine-tuning approach that extends the Direct Preference Optimization (DPO) technique, originally developed for efficient LLM alignment with preference data, to accommodate the MOFT settings. By substituting the Bradley-Terry-Luce model in DPO with the Plackett-Luce model, our framework is capable of handling a wide range of MOFT tasks that involve listwise ranking datasets. Compared with previous approaches, HyperDPO enjoys an efficient one-shot training process for profiling the Pareto front of auxiliary objectives, and offers post-training control over trade-offs. Additionally, we propose a novel *Hyper Prompt Tuning* design, that conveys continuous importance weight across objectives to transformer-based models without altering their architecture, and investigate the potential of *temperature-conditioned networks* for enhancing the flexibility of post-training control. We demonstrate the effectiveness and efficiency of the HyperDPO framework through its applications to various tasks, including Learning-to-Rank (LTR) and LLM alignment, highlighting its viability for large-scale ML deployments.", "keywords": [ "Direct Preference Optimization", "Multi-Objective Optimization", "Hypernetwork", "Alignment" ], "rating_list": [ 3, 8, 6, 5 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "q9g13IoWmk", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Cameron Berg", "gender": "unknown", "institution": "Agency Enterprise Studio", "country": "US", "position": "Researcher" }, { "name": "Diogo S de Lucena", "gender": "Male", "institution": "Agency Enterprise Studio", "country": "US", "position": "Researcher" }, { "name": "Judd Rosenblatt", "gender": "unknown", "institution": "Agency Enterprise Studio", "country": "US", "position": "Researcher" }, { "name": "Marc Carauleanu", "gender": "unknown", "institution": "Agency Enterprise Studio", "country": "US", "position": "AI Safety Researcher" }, { "name": "Michael Vaiana", "gender": "Male", "institution": "Agency Enterprise Studio", "country": "US", "position": "Principal Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 33, "n_ref": 48, "n_ref_all": 58, "n_fig": 3, "n_tab": 7, "L_tab": 2946, "n_element_tab": 154, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1698, "n_element_tab_1": 142, "formula_len_all": 91, "formula_len_all_1": 64, "len_all": 126691, "len_all_1": 60512, "len_abs": 1309, "len_title": 112, "len_sents": 42188, "len_sents_1": 28556, "n_sents": 347, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1315, "title": "Towards Safe and Honest AI Agents with Neural Self-Other Overlap", "abs": "As AI systems increasingly make critical decisions, deceptive AI poses a significant challenge to trust and safety. We present Self-Other Overlap (SOO) fine-tuning, a promising approach in AI Safety that could substantially improve our ability to build honest artificial intelligence. Inspired by cognitive neuroscience research on empathy, SOO aims to align how AI models represent themselves and others. Our experiments on LLMs with 7B, 27B, and 78B parameters demonstrate SOO\u2019s efficacy: deceptive responses of Mistral-7B-Instruct-v0.2 dropped from 73.6% to 17.2% with no observed reduction in general task performance, while in Gemma-2-27b-it and CalmeRys-78B-Orpo-v0.1 deceptive responses were reduced from 100% to 9.3% and 2.7%, respectively, with a small impact on capabilities. In reinforcement learning scenarios, SOO-trained agents showed significantly reduced deceptive behavior. SOO\u2019s focus on contrastive self and other-referencing observations offers strong potential for generalization across AI architectures. While current applications focus on language models and simple RL environments, SOO could pave the way for more trustworthy AI in broader domains. Ethical implications and long-term effects warrant further investigation, but SOO represents a significant step forward in AI safety research.", "keywords": [ "AI Safety", "ML Safety", "AI Deception", "large language models", "fine-tuning", "reinforcement learning", "self-other overlap" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "q9T51gF0fr", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Xiao Zhang", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Tenure-Track Faculty" }, { "name": "Yuelin Xu", "gender": "Female", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 8, "n_ref_uni": 68, "n_ref": 100, "n_ref_all": 123, "n_fig": 8, "n_tab": 6, "L_tab": 2298, "n_element_tab": 97, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 752, "n_element_tab_1": 73, "formula_len_all": 2499, "formula_len_all_1": 701, "len_all": 175667, "len_all_1": 68123, "len_abs": 1119, "len_title": 125, "len_sents": 52439, "len_sents_1": 34456, "n_sents": 344, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1122, "title": "Understanding Adversarially Robust Generalization via Weight-Curvature Index", "abs": "Despite extensive research on adversarial examples, the underlying mechanisms of adversarially robust generalization, a critical yet challenging task for deep learning, remain largely unknown. In this work, we propose a novel perspective to decipher adversarially robust generalization through the lens of the Weight-Curvature Index (WCI). The proposed WCI quantifies the vulnerability of models to adversarial perturbations using the Frobenius norm of weight matrices and the trace of Hessian matrices. We prove generalization bounds based on PAC-Bayesian theory and second-order loss function approximations to elucidate the interplay between robust generalization gap, model parameters, and loss landscape curvature. Our theory and experiments show that WCI effectively captures the robust generalization performance of adversarially trained models. By offering a nuanced understanding of adversarial robustness based on the scale of model parameters and the curvature of the loss landscape, our work provides crucial insights for designing more resilient deep learning models, enhancing their reliability and security.", "keywords": [ "Adversarially Robust Generalization; Weight-Curvature Index; Robust overfitting; PAC-Bayesian framework" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 3, 1, 2, 1 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "q8XGHj7yrC", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Daizong Liu", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Lichao Sun", "gender": "Male", "institution": "Lehigh University", "country": "US", "position": "Assistant Professor" }, { "name": "Pan Zhou", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Professor" }, { "name": "Wei Hu", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiang Fang", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xiaowen Cai", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Xiaoye Qu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 45, "n_ref": 102, "n_ref_all": 206, "n_fig": 9, "n_tab": 12, "L_tab": 6622, "n_element_tab": 647, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2468, "n_element_tab_1": 232, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 202944, "len_all_1": 66515, "len_abs": 1680, "len_title": 126, "len_sents": 51313, "len_sents_1": 30250, "n_sents": 393, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1571, "title": "Are Large Vision-Language Models Robust to Adversarial Visual Transformations?", "abs": "Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities across a wide range of multimodal understanding and reasoning tasks. However, recent research shows that LVLMs are susceptible to adversarial examples. Existing LVLM attackers either optimize the perturbations on the visual input or manipulate prompts to fool the LVLM models, requiring extensive design and engineering on these adversarial manipulations. While straightforward visual transformation can boast training generalization-ability, its potential risks to LVLMs in terms of safety and trustworthiness have been largely neglected. In this paper, we ask an intriguing question: can simple yet easy-to-implement visual transformations be utilized to attack the LVLM models? Motivated by this research gap and new attack setting, we propose the first comprehensive assessment of LVLMs' adversarial robustness to visual transformations by testing LVLMs' resilience to all possible transformation operations. Our empirical observations suggest that with the appropriate combination of the most harmful transformations, we can build transformation-based attacks more adversarial to the LVLM models. Moreover, adversarial learning of visual transformations is further introduced to adaptively apply the malicious impacts of all potentially harmful transformations to the raw images via gradient approximation for improving the attack effectiveness and imperceptibility. We hope that this study can provide deeper insights into the LVLMs' vulnerability to adversarial visual transformations.", "keywords": [ "Large Vision-Language Model", "Adversarial Attack" ], "rating_list": [ 1, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "q8H9t10Vsy", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "LIANG DING", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Xinmei Tian", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yifei Cheng", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Postdoc" }, { "name": "Zhiyuan Yu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 20, "n_formula_1": 17, "n_ref_uni": 31, "n_ref": 72, "n_ref_all": 84, "n_fig": 3, "n_tab": 7, "L_tab": 1519, "n_element_tab": 145, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 764, "n_element_tab_1": 82, "formula_len_all": 1578, "formula_len_all_1": 1256, "len_all": 123302, "len_all_1": 61985, "len_abs": 1676, "len_title": 140, "len_sents": 37883, "len_sents_1": 28142, "n_sents": 266, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1680, "title": "Memory-Efficient Block Coordinate Descent for Hessian-Informed Zeroth-Order Optimizer", "abs": "Fine-tuning large language models (LLMs) for specific downstream tasks has traditionally relied on memory-intensive optimizers using classical backpropagation, which demands substantial memory to store model states for gradient computation, motivating the development of memory-efficient zeroth-order optimizers that operate in a forward-only manner.\nHowever, the slower convergence of the zeroth-order optimizer remains a challenge, which recent research addresses by incorporating Hessian information to accelerate training, although storing even the diagonal Hessian requires memory equivalent to that of the model weights, leading to significant memory usage.\nTo mitigate this problem, we propose a novel approach that integrates the block coordinate descent (BCD) method with a Hessian-informed zeroth-order optimizer, allowing us to treat model layers as separate blocks and update only a subset of layers per training iteration, thereby reducing memory requirements and accelerating convergence.\nSpecifically, at each iteration, an active block of layers is selected according to the chosen BCD rule, such as ascending order, and their weights are updated while the other layers remain fixed, with diagonal Hessian information stored and updated exclusively for the active layers.\nFor fine-tuning foundation models of medium size (OPT-1.3B and LLaMA-2-7B), our method achieves up to 39\\% memory reduction compared to existing Hessian-informed zeroth-order methods, while preserving baseline accuracy and memory usage to zeroth-order methods across various tasks, offering a memory-efficient alternative method for LLMs fine-tuning, especially on memory-constrained devices.", "keywords": [ "zeroth-order optimization", "memory-efficient fine-tuning" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "q87GUkdQBm", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Klas Wijk", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "PhD student" }, { "name": "Ricardo Vinuesa", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "Associate Professor" }, { "name": "Hossein Azizpour", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "Associate Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 41, "n_ref": 91, "n_ref_all": 124, "n_fig": 5, "n_tab": 4, "L_tab": 856, "n_element_tab": 93, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 667, "n_element_tab_1": 91, "formula_len_all": 924, "formula_len_all_1": 924, "len_all": 100784, "len_all_1": 58796, "len_abs": 1118, "len_title": 102, "len_sents": 26346, "len_sents_1": 24661, "n_sents": 230, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1130, "title": "SFESS: Score Function Estimators for $k$-Subset Sampling", "abs": "Are score function estimators a viable approach to learning with $k$-subset sampling? Sampling $k$-subsets is a fundamental operation in machine learning that is not amenable to differentiable parametrization, impeding gradient-based optimization. Prior work has focused on relaxed sampling or approximate pathwise gradients but dismissed score function estimators due to their high variance. Inspired by the success of score function estimators in variational inference and reinforcement learning, we revisit them within the context of $k$-subset sampling. Specifically, we demonstrate how to efficiently compute the $k$-subset distribution's score function using a discrete Fourier transform, and reduce the estimator's variance with control variates. The resulting estimator provides both exact samples and unbiased gradient estimates while being applicable to non-differentiable downstream models, unlike existing methods. We validate our approach in multiple experimental settings and find that comparable results can be achieved to recent state-of-the-art relaxed and approximate pathwise gradient methods, across all tasks.", "keywords": [ "subset", "top-k", "gradient estimation", "score function estimator", "variance reduction" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 4, 3 ] }, { "paper_id": "q7aROKohBZ", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 9, "author_info_list": [ { "name": "Andrey Kuznetsov", "gender": "Male", "institution": "AIRI", "country": "RU", "position": "Principal Researcher" }, { "name": "Andrey Savchenko", "gender": "Male", "institution": "Sber AI Lab", "country": "RU", "position": "Principal Researcher" }, { "name": "Irina Abdullaeva", "gender": "Female", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Researcher" }, { "name": "Ivan A Kireev", "gender": "Male", "institution": "Sber AI Lab", "country": "RU", "position": "Principal Researcher" }, { "name": "Mikhail Orlov", "gender": "unknown", "institution": "Higher School of Economics", "country": "", "position": "PhD student" }, { "name": "Viacheslav Vasilev", "gender": "unknown", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "PhD" }, { "name": "Andrey Filatov", "gender": "unknown", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Denis Valerievich Dimitrov", "gender": "Male", "institution": "AIRI", "country": "", "position": "Principal Researcher" }, { "name": "Ivan Alexandrovich Karpukhin", "gender": "Male", "institution": "Sberbank", "country": "RU", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 45, "n_ref": 67, "n_ref_all": 92, "n_fig": 3, "n_tab": 14, "L_tab": 4090, "n_element_tab": 505, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1878, "n_element_tab_1": 252, "formula_len_all": 700, "formula_len_all_1": 268, "len_all": 175461, "len_all_1": 63933, "len_abs": 804, "len_title": 40, "len_sents": 57609, "len_sents_1": 27192, "n_sents": 460, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 40, "L_abs": 833, "title": "ESQA: Event Sequences Question Answering", "abs": "Event sequences (ESs) arise in many practical domains including finance, retail, social networks, and healthcare. In the context of machine learning, event sequences can be seen as a special type of tabular data with annotated timestamps. Despite the importance of ESs modeling and analysis, little effort was made in adapting large language models (LLMs) to the ESs domain. In this paper, we highlight the common difficulties of ESs processing and propose a novel solution capable of solving multiple downstream tasks with little or no finetuning. In particular, we solve the problem of working with long sequences and improve time and numeric features processing. The resulting method, called ESQA, effectively utilizes the power of LLMs and, according to extensive experiments, achieves state-of-the-art results in the ESs domain.", "keywords": [ "event sequences", "large language models", "multi-modality" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "q7Xi4yZYcH", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Jon Crowcroft", "gender": "Male", "institution": "university of cambridge", "country": "GB", "position": "Full Professor" }, { "name": "Stefan Behfar", "gender": "Not Specified", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 39, "n_formula_1": 11, "n_ref_uni": 12, "n_ref": 20, "n_ref_all": 36, "n_fig": 6, "n_tab": 3, "L_tab": 1885, "n_element_tab": 102, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1614, "n_element_tab_1": 65, "formula_len_all": 1530, "formula_len_all_1": 393, "len_all": 145517, "len_all_1": 66910, "len_abs": 1319, "len_title": 140, "len_sents": 56853, "len_sents_1": 32371, "n_sents": 432, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1322, "title": "A Scalable Temporal-Spatial Framework for Transaction Anomaly Detection in Ethereum Networks", "abs": "The rapid evolution of the Ethereum network necessitates sophisticated techniques to ensure its robustness against potential threats and to maintain transparency. While Graph Neural Networks (GNNs) have pioneered anomaly detection in such platforms, capturing the intricacies of both spatial and temporal transactional patterns has remained a challenge. This study presents a fusion of Graph Convolutional Networks (GCNs) with Temporal Random Walks (TRW) enhanced by probabilistic sampling to bridge this gap. Our approach, unlike traditional GCNs, leverages the strengths of TRW to discern complex temporal sequences in Ethereum transactions, thereby providing a more nuanced transaction anomaly detection mechanism. Extensive evaluations demonstrate that our TRW-GCN framework substantially advances the performance metrics over conventional GCNs in detecting irregularities such as suspiciously timed transactions, patterns indicative of token pump and dump schemes, or anomalous behavior in smart contract executions over time. As baseline algorithms for comparison, common unsupervised methods such as Isolation Forest, One-Class SVM, and DBSCAN (as classifier for TRW-GCN embedding) are employed; finally our novel TRW-GCN plus scoring method is compared with the state-of-the-art temporal graph attention algorithm.", "keywords": [ "Probabilistic sampling", "Temporal random walk", "Graph convolutional network", "Transaction anomaly detection", "Ethereum networks" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 1, 2, 3 ], "presentation_list": [ 1, 3, 2, 2 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "q6zrZbth1F", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Johannes Lederer", "gender": "unknown", "institution": "Universit\u00e4t Hamburg", "country": "DE", "position": "Full Professor" }, { "name": "Mahsa Taheri", "gender": "Female", "institution": "Universit\u00e4t Hamburg", "country": "DE", "position": "Postdoc" }, { "name": "Pegah Golestaneh", "gender": "Female", "institution": "Ruhr-Universt\u00e4t Bochum", "country": "", "position": "PhD student" } ], "n_formula": 85, "n_formula_1": 34, "n_ref_uni": 44, "n_ref": 80, "n_ref_all": 93, "n_fig": 2, "n_tab": 4, "L_tab": 61, "n_element_tab": 8, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 303, "n_element_tab_1": 22, "formula_len_all": 8867, "formula_len_all_1": 2832, "len_all": 176068, "len_all_1": 63383, "len_abs": 731, "len_title": 107, "len_sents": 52798, "len_sents_1": 27757, "n_sents": 417, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 744, "title": "How many samples are needed to train a deep neural network?", "abs": "Even though neural networks have become standard tools in many areas, many important statistical questions remain open. This paper studies the question of how much data are needed to train a ReLU feed-forward neural network. Our theoretical and empirical results suggest that the generalization error of ReLU feed-forward neural networks scales at the rate $1/\\sqrt{n}$ in the sample size $n$-rather than the \"parametric rate\" $1/n$, which might be suggested by traditional statistical theories. Thus, broadly speaking, our results underpin the common belief that neural networks need \"many\" training samples. Along the way, we also establish new technical insights, such as the first lower bounds of the entropy of ReLU feed-forward networks.", "keywords": [ "Neural networks", "Deep learning", "generalization error", "information theory", "mini-max bound", "learning theory" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "q6pm9CObJn", "primary_area": "datasets and benchmarks", "n_author": 16, "author_info_list": [ { "name": "An-Lan Wang", "gender": "unknown", "institution": "ByteDance Inc.", "country": "CN", "position": "Intern" }, { "name": "Can Huang", "gender": "Male", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "Chunhui Lin", "gender": "Male", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "Hao Feng", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Hao Liu", "gender": "Male", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "Jingqun Tang", "gender": "unknown", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "Mohamad Fitri Faiz Bin Mahmood", "gender": "Male", "institution": "ByteDance Inc.", "country": "", "position": "Instructor" }, { "name": "Qi Liu", "gender": "Male", "institution": "Bytedance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Shu Wei", "gender": "Female", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Wanqing Li", "gender": "Female", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Xiang Bai", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yanjie Wang", "gender": "Male", "institution": "ByteDance Inc", "country": "CN", "position": "Researcher" }, { "name": "Yuliang Liu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zhen Zhao", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "JINGHUI LU", "gender": "Male", "institution": "ByteDance Inc.", "country": "IE", "position": "Researcher" }, { "name": "Ye yongjie", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 165, "n_ref_all": 182, "n_fig": 7, "n_tab": 6, "L_tab": 921, "n_element_tab": 105, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1136, "n_element_tab_1": 58, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 134243, "len_all_1": 50030, "len_abs": 1715, "len_title": 119, "len_sents": 30751, "len_sents_1": 20617, "n_sents": 182, "n_sents_1": 131, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1722, "title": "MTVQA: Benchmarking Multilingual Text-Centric Visual Question Answering", "abs": "Text-Centric Visual Question Answering(TEC-VQA) in its proper format not only facilitates human-machine interaction in text-centric visual environments but also serves as a de facto gold proxy to evaluate AI models in the domain of text-centric scene understanding. Nonetheless, most existing TEC-VQA benchmarks focus on high-resource languages like English and Chinese. Despite pioneering works expanding multilingual QA pairs in non-text-centric VQA datasets through translation engines, the translation-based protocol encounters a substantial ``visual-textual misalignment'' problem when applied to TEC-VQA. Specifically, it prioritizes the text in question-answer pairs while disregarding the visual text present in images. Moreover, it fails to address complexities related to nuanced meaning, contextual distortion, language bias, and question-type diversity. In this work, we tackle multilingual TEC-VQA by introducing MTVQA, the first benchmark featuring high-quality human expert annotations across 9 diverse languages, consisting of 6,778 question-answer pairs across 2,116 images. Further, by comprehensively evaluating numerous state-of-the-art Multimodal Large Language Models~(MLLMs), including GPT-4o, GPT-4V, Claude3, and Gemini, on the MTVQA dataset, it is evident that there is still a large room for performance improvement, underscoring the value of MTVQA. Additionally, we supply multilingual training data within the MTVQA dataset, demonstrating that straightforward fine-tuning with this data can substantially enhance multilingual TEC-VQA performance. We aspire that MTVQA will offer the research community fresh insights and stimulate further exploration in multilingual visual text comprehension.", "keywords": [ "Multilingual Visual Question Answering", "Visual Text Understanding", "Multimodal Comprehension Evaluation" ], "rating_list": [ 3, 5, 6, 5, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 4, 4, 3 ] }, { "paper_id": "q6hEuC48Dk", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Jason Kurz", "gender": "Male", "institution": "University of Waikato", "country": "NZ", "position": "Lecturer" }, { "name": "Sean Oughton", "gender": "unknown", "institution": "University of Waikato", "country": "NZ", "position": "Full Professor" }, { "name": "Shitao Liu", "gender": "Male", "institution": "Clemson University", "country": "US", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 20, "n_ref_uni": 17, "n_ref": 30, "n_ref_all": 46, "n_fig": 7, "n_tab": 2, "L_tab": 435, "n_element_tab": 54, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 292, "n_element_tab_1": 57, "formula_len_all": 1876, "formula_len_all_1": 1912, "len_all": 89908, "len_all_1": 60983, "len_abs": 1146, "len_title": 78, "len_sents": 29641, "len_sents_1": 27130, "n_sents": 211, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 30, "L_abs": 1163, "title": "Radial Basis Operator Networks", "abs": "Operator networks are designed to approximate nonlinear operators, which provide mappings between infinite-dimensional spaces such as function spaces. These networks are playing an increasingly important role in machine learning, with their most notable contributions in the field of scientific computing. Their significance stems from their ability to handle the type of data often encountered in scientific applications. For instance, in climate modeling or fluid dynamics, input data typically consists of discretized continuous fields (like temperature distributions or velocity fields). We introduce the radial basis operator network (RBON), which represents a significant advancement as the first operator network capable of learning an operator in both the time domain and frequency domain when adjusted to accept complex-valued inputs. Despite the small, single hidden-layer structure, the RBON boasts small $L^2$ relative test error for both in- and out-of-distribution data (OOD) of less than $1\\times 10^{-7}$ in some benchmark cases. Moreover, the RBON maintains small error on OOD data from entirely different function classes from the training data.", "keywords": [ "Operator networks", "Radial basis functions", "Scientific computing", "Partial differential equations" ], "rating_list": [ 5, 3, 3, 5, 3 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 1, 3, 1 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 3, 2, 4, 2, 3 ] }, { "paper_id": "q6TelS1z7N", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Linus Aronsson", "gender": "Male", "institution": "Chalmers University of Technology", "country": "SE", "position": "PhD student" }, { "name": "Morteza Haghir Chehreghani", "gender": "Male", "institution": "Chalmers University of technology", "country": "SE", "position": "Associate Professor" } ], "n_formula": 68, "n_formula_1": 35, "n_ref_uni": 45, "n_ref": 164, "n_ref_all": 185, "n_fig": 14, "n_tab": 4, "L_tab": 1275, "n_element_tab": 24, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 195, "n_element_tab_1": 5, "formula_len_all": 5671, "formula_len_all_1": 2244, "len_all": 195202, "len_all_1": 77191, "len_abs": 457, "len_title": 99, "len_sents": 58976, "len_sents_1": 33278, "n_sents": 538, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 458, "title": "Information-Theoretic Active Correlation Clustering", "abs": "We study correlation clustering where the pairwise similarities are not known in advance. For this purpose, we employ active learning to query pairwise similarities in a cost-efficient way. We propose a number of effective information-theoretic acquisition functions based on entropy and information gain. We extensively investigate the performance of our methods in different settings and demonstrate their superior performance compared to the alternatives.", "keywords": [ "active learning", "active clustering", "correlation clustering", "acquisition function" ], "rating_list": [ 6, 8, 3, 6 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "q6CM6UdP3K", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Alif Bin Abdul Qayyum", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Byung-Jun Yoon", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Full Professor" }, { "name": "Nathan Urban", "gender": "Male", "institution": "Brookhaven National Laboratory", "country": "", "position": "Researcher" }, { "name": "Xiaoning Qian", "gender": "Male", "institution": "Texas A&M", "country": "", "position": "Full Professor" }, { "name": "Xihaier Luo", "gender": "Male", "institution": "Brookhaven National Laboratory", "country": "US", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 22, "n_ref_uni": 49, "n_ref": 79, "n_ref_all": 109, "n_fig": 19, "n_tab": 3, "L_tab": 1498, "n_element_tab": 250, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1354, "n_element_tab_1": 17, "formula_len_all": 2182, "formula_len_all_1": 1332, "len_all": 162744, "len_all_1": 63911, "len_abs": 1106, "len_title": 142, "len_sents": 40592, "len_sents_1": 26694, "n_sents": 279, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 152, "L_abs": 1111, "title": "$\\textit{One Stone Three Birds:}$ Three-Dimensional Implicit Neural Network for Compression and Continuous Representation of Multi-Altitude Climate Data", "abs": "Wind energy stands out as a promising clean and renewable energy alternative, not only for its potential to combat global warming but also for its capacity to meet the ever-growing demand for energy. However, analysis of wind data to fully harness the benefits of wind energy demands tackling several related challenges: \n(1) Current data resolution is inadequate for capturing the detailed information needed across diverse climatic conditions;\n(2) Efficient management and storage of real-time measurements are currently lacking;\n(3) Extrapolating wind data across spatial specifications enables analysis at costly-to-measure, unobserved points is necessary.\nIn response to these challenges, we introduce a modality-agnostic learning framework utilizing implicit neural networks. Our model effectively compresses a large volume of climate data into a manageable latent codec. It also learns underlying continuous climate patterns, enabling reconstruction at any scale and supporting modality transfer and fusion. Extensive experimental results show consistent performance improvements over existing baselines.", "keywords": [ "Multi-modal representation learning", "continuous super-resolution", "dimensionality reduction", "cross-modal prediction", "scientific data compression" ], "rating_list": [ 3, 6, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "q5zMyAUhGx", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Huijuan Zhou", "gender": "Female", "institution": "Shanghai University of Finance and Economics", "country": "CN", "position": "Assistant Professor" }, { "name": "Xianyang Zhang", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Full Professor" } ], "n_formula": 94, "n_formula_1": 39, "n_ref_uni": 39, "n_ref": 83, "n_ref_all": 93, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12703, "formula_len_all_1": 4853, "len_all": 145588, "len_all_1": 53783, "len_abs": 1517, "len_title": 122, "len_sents": 37731, "len_sents_1": 21005, "n_sents": 380, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1522, "title": "Generalization Bounds and Model Complexity for Kolmogorov\u2013Arnold Networks", "abs": "Kolmogorov\u2013Arnold Network (KAN) is a network structure recently proposed in Liu et al. (2024) that offers improved interpretability and a more parsimonious design in many science-oriented tasks compared to multi-layer perceptrons. This work provides a rigorous theoretical analysis of KAN by establishing generalization bounds for KAN equipped with activation functions that are either represented by linear combinations of basis functions or lying in a low-rank Reproducing Kernel Hilbert Space (RKHS). In the first case, the generalization bound accommodates various choices of basis functions in forming the activation functions in each layer of KAN and is adapted to different operator norms at each layer. For a particular choice of operator norms, the bound scales with the $l_1$ norm of the coefficient matrices and the Lipschitz constants for the activation functions, and it has no dependence on combinatorial parameters (e.g., number of nodes) outside of logarithmic factors. Moreover, our result does not require the boundedness assumption on the loss function and, hence, is applicable to a general class of regression-type loss functions. In the low-rank case, the generalization bound scales polynomially with the underlying ranks as well as the Lipschitz constants of the activation functions in each layer. These bounds are empirically investigated for KANs trained with stochastic gradient descent on simulated and real data sets. The numerical results demonstrate the practical relevance of these bounds.", "keywords": [ "Deep neural networks", "Empirical process", "Generalization bounds", "Kolmogorov-Arnold representation theorem", "Model complexity" ], "rating_list": [ 6, 5, 6, 6, 8 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 3, 2, 4, 3, 4 ] }, { "paper_id": "q5sOv4xQe4", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Enze Xie", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Researcher" }, { "name": "Han Cai", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Haotian Tang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Junyu Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Shang Yang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Song Han", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yao Lu", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Yecheng Wu", "gender": "Male", "institution": ", Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhuoyang Zhang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "\u4fca\u677e \u9648", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 53, "n_ref_all": 79, "n_fig": 9, "n_tab": 3, "L_tab": 1624, "n_element_tab": 252, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1591, "n_element_tab_1": 247, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 139832, "len_all_1": 49790, "len_abs": 0, "len_title": 120, "len_sents": 26727, "len_sents_1": 20521, "n_sents": 217, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 72, "L_abs": 1261, "title": "HART: Efficient Visual Generation with Hybrid Autoregressive Transformer", "abs": "We introduce Hybrid Autoregressive Transformer (HART), the first autoregressive (AR) visual generation model capable of directly generating 1024x1024 images, rivaling diffusion models in image generation quality. Existing AR models face limitations due to the poor image reconstruction quality of their discrete tokenizers and the prohibitive training costs associated with generating 1024px images. To address these challenges, we present the hybrid tokenizer, which decomposes the continuous latents from the autoencoder into two components: discrete tokens representing the big picture and continuous tokens representing the residual components that cannot be represented by the discrete tokens. The discrete component is modeled by a scalable-resolution discrete AR model, while the continuous component is learned with a lightweight residual diffusion module with only 37M parameters. Compared with the discrete-only VAR tokenizer, our hybrid approach improves reconstruction FID from 2.11 to 0.30 on MJHQ-30K, leading to a 31% generation FID improvement from 7.85 to 5.38. HART also outperforms state-of-the-art diffusion models in both FID and CLIP score, with 4.5-7.7x higher throughput and 6.9-13.4x lower MACs. Code will be released upon publication.", "keywords": [ "autoregressive models", "image generation", "text-to-image" ], "rating_list": [ 6, 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 4 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 5, 4 ] }, { "paper_id": "q5lJxCXjiY", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Emily Cheng", "gender": "Female", "institution": "Universitat Pompeu Fabra", "country": "ES", "position": "PhD student" }, { "name": "Jin Hwa Lee", "gender": "Female", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Lei Yu", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Thomas Jiralerspong", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "", "position": "MS student" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 72, "n_ref": 124, "n_ref_all": 145, "n_fig": 5, "n_tab": 4, "L_tab": 3660, "n_element_tab": 120, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 888, "n_element_tab_1": 40, "formula_len_all": 482, "formula_len_all_1": 580, "len_all": 197589, "len_all_1": 64084, "len_abs": 1267, "len_title": 123, "len_sents": 52668, "len_sents_1": 31255, "n_sents": 333, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1079, "title": "GEOMETRIC SIGNATURES OF COMPOSITIONALITY ACROSS A LANGUAGE MODEL\u2019S LIFETIME", "abs": "Compositionality, the notion that the meaning of an expression is constructed from the meaning of its parts and syntactic rules, permits the infinite productivity of human language. For the first time, artificial language models (LMs) are able to match human performance in a number of compositional generalization tasks. However, much remains to be understood about the representational mechanisms underlying these abilities. We take a high-level geometric approach to this problem by relating the degree of compositionality in a dataset to the intrinsic dimensionality of its representations under an LM, a measure of feature complexity. We find not only that the degree of dataset compositionality is reflected in representations' intrinsic dimensionality, but that the relationship between compositionality and geometric complexity arises due to learned linguistic features over training. Finally, our analyses reveal a striking contrast between linear and nonlinear dimensionality, showing that they respectively encode formal and semantic aspects of linguistic composition.", "keywords": [ "geometry", "compositionality", "representation learning" ], "rating_list": [ 5, 5, 6, 3, 8 ], "soundness_list": [ 2, 2, 3, 1, 2 ], "presentation_list": [ 2, 2, 3, 2, 4 ], "contribution_list": [ 3, 2, 2, 1, 3 ], "confidence_list": [ 4, 5, 4, 3, 3 ] }, { "paper_id": "q5MUMlHxpd", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Chitta Baral", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Full Professor" }, { "name": "Maitreya Patel", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Nilay Yilmaz", "gender": "Female", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Suren Jayasuriya", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" }, { "name": "Tejas Gokhale", "gender": "Not Specified", "institution": "University of Maryland, Baltimore County", "country": "US", "position": "Assistant Professor" }, { "name": "Yezhou Yang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Associate Professor" }, { "name": "Yiran Luo", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 26, "n_ref": 40, "n_ref_all": 64, "n_fig": 11, "n_tab": 12, "L_tab": 2928, "n_element_tab": 498, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 815, "n_element_tab_1": 121, "formula_len_all": 280, "formula_len_all_1": 280, "len_all": 155497, "len_all_1": 60577, "len_abs": 1446, "len_title": 81, "len_sents": 45330, "len_sents_1": 28368, "n_sents": 356, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1379, "title": "Voila: Evaluation of MLLMs For Perceptual Understanding and Analogical Reasoning", "abs": "Multimodal Large Language Models (MLLMs) have become a powerful tool for\nintegrating visual and textual information. Despite their exceptional performance\non visual understanding benchmarks, measuring their ability to reason abstractly\nacross multiple images remains a significant challenge. To address this, we introduce VOILA , a large-scale, open-ended, dynamic benchmark designed to evaluate MLLMs\u2019 perceptual understanding and abstract relational reasoning. VOILA\nemploys an analogical mapping approach in the visual domain, requiring models\nto generate an image that completes an analogy between two given image pairs,\nreference and application, without relying on predefined choices. Our experiments demonstrate that VOILA presents MLLMs with demanding relational reasoning tasks. Through multi-step analysis, we reveal that current MLLMs struggle\nto comprehend inter-image relationships and exhibit limited capabilities in highlevel relational reasoning. Notably, we observe that performance improves when\nusing least-to-most prompting strategies. Comprehensive evaluations on opensource models and GPT-4o show that while the MolmoE-8B model achieves a\nstate-of-the-art performance of 34% and 19% at finding the text-based answer to\nthe questions on easy and hard scenarios, human performance consistently remains significantly higher at 70% on both difficulty scenarios.", "keywords": [ "abstract reasoning", "relational reasoning", "perceptual understanding", "MLLMs", "visual analogy" ], "rating_list": [ 5, 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "q5EZ7gKcnW", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Cassidy Laidlaw", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Jacob Steinhardt", "gender": "unknown", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" }, { "name": "Yaowen Ye", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Undergrad student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 54, "n_ref": 111, "n_ref_all": 144, "n_fig": 16, "n_tab": 5, "L_tab": 1428, "n_element_tab": 154, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 733, "n_element_tab_1": 63, "formula_len_all": 588, "formula_len_all_1": 261, "len_all": 187973, "len_all_1": 69818, "len_abs": 1339, "len_title": 139, "len_sents": 57921, "len_sents_1": 33218, "n_sents": 443, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 91, "L_abs": 1446, "title": "Iterative Label Refinement Matters More than Preference Optimization under Weak Supervision", "abs": "Language model (LM) post-training relies on two stages of human supervision: task demonstrations for supervised finetuning (SFT), followed by preference comparisons for reinforcement learning from human feedback (RLHF). As LMs become more capable, the tasks they are given become harder to supervise. Will post-training remain effective under unreliable supervision? To test this, we simulate unreliable demonstrations and comparison feedback using small LMs and time-constrained humans. We find that in the presence of unreliable supervision, SFT still retains some effectiveness, but DPO (a common RLHF algorithm) fails to improve the model beyond SFT. To address this, we propose *iterative label refinement* (ILR) as an alternative to RLHF. ILR improves the SFT data by using comparison feedback to decide whether human demonstrations should be replaced by model-generated alternatives, then retrains the model via SFT on the updated data. SFT+ILR outperforms SFT+DPO on several tasks with unreliable supervision (math, coding, and safe instruction-following). Our findings suggest that as LMs are used for complex tasks where human supervision is unreliable, RLHF may no longer be the best use of human comparison feedback; instead, it is better to direct feedback towards improving the training *data* rather than continually training the *model*. Our code and data are available at https://github.com/helloelwin/iterative-label-refinement.", "keywords": [ "unreliable human supervision", "language model post-training", "scalable oversight" ], "rating_list": [ 8, 8, 5, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "q541p2YLt2", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "JongHyun Hong", "gender": "unknown", "institution": "Hanyang University", "country": "KR", "position": "MS student" }, { "name": "Sungyoon Lee", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 23, "n_formula_1": 13, "n_ref_uni": 36, "n_ref": 59, "n_ref_all": 69, "n_fig": 4, "n_tab": 1, "L_tab": 159, "n_element_tab": 20, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1078, "formula_len_all_1": 677, "len_all": 109426, "len_all_1": 59850, "len_abs": 1031, "len_title": 123, "len_sents": 32898, "len_sents_1": 28438, "n_sents": 201, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1048, "title": "Transformer Training Instability of Softmax and Lipschitz-Kernel Attentions", "abs": "Transformers have been making significant progress across various domains, and recently, with scaling up of models like LLMs, they have achieved even greater success. Recent findings have shown that the softmax function in the self-attention used to re-weight the attention logits into probability vectors causes \\emph{attention entropy collapse}, where the attention is concentrated on a single token, and it leads to unstable training. In this work, we first demonstrate that the (non-Lipschitz) softmax-based attention leads to the attention entropy collapse but the \\emph{Lipschitz-kernel}-based attention does not. We show that the Lipschitzness of the attention plays an important role in keeping the attention entropy stable regardless of the variance of the attention logits. Moreover, we argue that the underlying reason why the attention entropy collapse leads to the training instability is that as the attention probabilities become more concentrated, it causes the attention matrix to gradually increase, leading to gradient exploding.", "keywords": [ "Transformer", "linear attention", "self-attention", "optimization", "training stability", "entropy collapse" ], "rating_list": [ 3, 3, 1, 3 ], "soundness_list": [ 1, 2, 1, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 1, 1, 2 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "q44uq3tc2D", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Jiayi Ji", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoshuai Sun", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Yaxin Luo", "gender": "Male", "institution": "Technical University of Denmark", "country": "DK", "position": "Undergrad student" }, { "name": "Yiyi Zhou", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhiqiang Shen", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Gen Luogen", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 37, "n_ref": 88, "n_ref_all": 96, "n_fig": 3, "n_tab": 6, "L_tab": 4330, "n_element_tab": 373, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 4330, "n_element_tab_1": 373, "formula_len_all": 613, "formula_len_all_1": 616, "len_all": 148633, "len_all_1": 68341, "len_abs": 1941, "len_title": 130, "len_sents": 29387, "len_sents_1": 26114, "n_sents": 256, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1665, "title": "$\\gamma-$MoD: Exploring Mixture-of-Depth Adaptation for Multimodal Large Language Models", "abs": "Despite the significant progress in multimodal large language models (MLLMs), their high computational cost remains a barrier to real-world deployment. Inspired by the mixture of depths (MoDs) in natural language processing, we aim to address this limitation from the perspective of ``activated tokens''. Our key insight is that if most tokens are redundant for the layer computation, then can be skipped directly via the MoD layer. However, directly converting the dense layers of MLLMs to MoD layers leads to substantial performance degradation. To address this issue, we propose an innovative MoD adaptation strategy for existing MLLMs called $\\gamma$-MoD. In $\\gamma$-MoD, a novel metric is proposed to guide the deployment of MoDs in the MLLM, namely rank of attention maps (ARank). Through ARank, we can effectively identify which layer is redundant and should be replaced with the MoD layer. Based on ARank, we further propose two novel designs to maximize the computational sparsity of MLLM while maintaining its performance, namely shared vision-language router and masked routing learning. With these designs, more than 90% dense layers of the MLLM can be effectively converted to the MoD ones. To validate our method, we apply it to three popular MLLMs, and conduct extensive experiments on 9 benchmark datasets. Experimental results not only validate the significant efficiency benefit of $\\gamma$-MoD to existing MLLMs but also confirm its generalization ability on various MLLMs. For example, with a minor performance drop, i.e., -1.5%, $\\gamma$-MoD can reduce the training and inference time of LLaVA-HR by 31.0% and 53.2%, respectively.", "keywords": [ "mixture of depths; multimodal large language models" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "q3ztjJRQuJ", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Boyang Li", "gender": "Unspecified", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Qingyong Li", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Wen Wang", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Wenju Sun", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yangliao Geng", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 23, "n_formula_1": 23, "n_ref_uni": 45, "n_ref": 97, "n_ref_all": 120, "n_fig": 6, "n_tab": 10, "L_tab": 5135, "n_element_tab": 933, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2898, "n_element_tab_1": 506, "formula_len_all": 1407, "formula_len_all_1": 1326, "len_all": 165484, "len_all_1": 62581, "len_abs": 1718, "len_title": 103, "len_sents": 41522, "len_sents_1": 24343, "n_sents": 327, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1772, "title": "Task Arithmetic in Trust Region: A Training-Free Model Merging Approach to Navigate Knowledge Conflicts", "abs": "Multi-task model merging offers an efficient solution for integrating knowledge from multiple fine-tuned models, mitigating the significant computational and storage demands associated with multi-task training. As a key technique in this field, Task Arithmetic (TA) defines task vectors by subtracting the pre-trained model ($\\theta_{\\text{pre}}$) from the fine-tuned task models in parameter space, then adjusting the weight between these task vectors and $\\theta_{\\text{pre}}$ to balance task-generalized and task-specific knowledge. Despite the promising performance of TA, conflicts can arise among the task vectors, particularly when different tasks require distinct model adaptations. In this paper, we formally define this issue as knowledge conflicts, characterized by the performance degradation of one task after merging with a model fine-tuned for another task. Through in-depth analysis, we show that these conflicts stem primarily from the components of task vectors that align with the gradient of task-specific losses at $\\theta_{\\text{pre}}$. To address this, we propose Task Arithmetic in Trust Region (TATR), which defines the trust region as dimensions in the model parameter space that cause only small changes (corresponding to the task vector components with gradient orthogonal direction) in the task-specific losses. Restricting parameter merging within this trust region, TATR can effectively alleviate knowledge conflicts. Moreover, TATR serves as both an independent approach and a plug-and-play module compatible with a wide range of TA-based methods. Extensive empirical evaluations on eight distinct datasets robustly demonstrate that TATR improves the multi-task performance of several TA-based model merging methods by an observable margin.", "keywords": [ "Multi-task Learning", "Model Merging", "negative transfer" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "q3Z2v2mt1R", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Fragkiskos D. Malliaros", "gender": "Male", "institution": "CentraleSup\u00e9lec, Inria, Paris-Saclay University", "country": "FR", "position": "Associate Professor" }, { "name": "Johannes F. Lutzeyer", "gender": "Male", "institution": "Ecole Polytechique", "country": "FR", "position": "Assistant Professor" }, { "name": "Michalis Vazirgiannis", "gender": "Male", "institution": "Ecole Polytechnique, France", "country": "", "position": "Full Professor" }, { "name": "Sofiane ENNADIR", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "PhD student" }, { "name": "Yassine ABBAHADDOU", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "FR", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 31, "n_ref_uni": 46, "n_ref": 74, "n_ref_all": 89, "n_fig": 1, "n_tab": 6, "L_tab": 3881, "n_element_tab": 242, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 2442, "n_element_tab_1": 113, "formula_len_all": 2889, "formula_len_all_1": 1671, "len_all": 136432, "len_all_1": 67918, "len_abs": 918, "len_title": 135, "len_sents": 36399, "len_sents_1": 29774, "n_sents": 283, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 920, "title": "Post-Hoc Robustness Enhancement in Graph Neural Networks with Conditional Random Fields", "abs": "Graph Neural Networks (GNNs), which are nowadays the benchmark approach in graph representation learning, have been shown to be vulnerable to adversarial attacks, raising concerns about their real-world applicability. While existing defense techniques primarily concentrate on the training phase of GNNs, involving adjustments to message passing architectures or pre-processing methods, there is a noticeable gap in methods focusing on increasing robustness during inference. In this context, this study introduces RobustCRF, a post-hoc approach aiming to enhance the robustness of GNNs at the inference stage. Our proposed method, founded on statistical relational learning using a Conditional Random Field, is model-agnostic and does not require prior knowledge about the underlying model architecture. We validate the efficacy of this approach across various models, leveraging benchmark node classification datasets.", "keywords": [ "Robustness", "Graph Neural Networks", "Conditional Random Fields" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "q3MYZQ3es8", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Bocheng Zou", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Undergrad student" }, { "name": "Dingmin Wang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Zhen Han", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "zhiqiang xu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 28, "n_ref": 125, "n_ref_all": 156, "n_fig": 6, "n_tab": 7, "L_tab": 5067, "n_element_tab": 22, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 71, "n_element_tab_1": 1, "formula_len_all": 922, "formula_len_all_1": 976, "len_all": 189786, "len_all_1": 69734, "len_abs": 2061, "len_title": 167, "len_sents": 64206, "len_sents_1": 33667, "n_sents": 478, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 119, "L_abs": 2006, "title": "tBen: Benchmarking and Testing the Rule-Based Temporal Logic Reasoning Ability of Large Language Models with DatalogMTL", "abs": "Large language models (LLMs) are increasingly adopted for a variety of tasks, including multi-hop question answering, knowledge probing, and symbolic commonsense reasoning. While LLMs have advanced the state-of-the-art in these areas, their ability to explicitly solve rule-based temporal logic reasoning problems\u2014a complex cognitive process involving the understanding, representation, and manipulation of temporal information such as events, their durations, and relationships\u2014remains unexplored. To enhance understanding of LLM performance in this common task widely explored in the traditional symbolic AI field, we have developed a new set of synthetic benchmarks for rule-based temporal logic reasoning tBen. Our tBen benchmarks are built within the context of DatalogMTL, a powerful knowledge representation language for reasoning about the properties of systems that evolve over time, in which we provide flexible configurations for customizing temporal rules and task complexity.\n\nWe evaluated the close-sourced GPT-4o and the open-sourced Llama-3 using three common prompting settings\u2014$\\textit{zero-shot}$, $\\textit{few-shot}$, and $\\textit{zero-shot-CoT}$\u2014on our synthetic benchmarks. Our key findings are as follows: (i) Without generating the reasoning process (chain-of-thought), even advanced LLMs like GPT-4o exhibited nearly random performance on these rule-based temporal logic reasoning tasks. However, with chain-of-thought prompting, LLMs demonstrated preliminary temporal logical reasoning abilities; (ii) Both GPT-4o and Llama-3 were unable to solve temporal logical reasoning problems involving recursion, indicating a lack of advanced complex reasoning capabilities in understanding symbolic representations involving time; (iii) There is significant room for improvement in leveraging large language models to address problems widely explored in the traditional logic-based AI domain. Prompts and datasets are available in the appendix, and a datasheet for tBen is also provided.", "keywords": [ "Temporal Logic Reasoning", "Large Language Models", "DatalogMTL" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "q3EbOXb4y1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Daniel Kwan", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "Undergrad student" }, { "name": "Jean Sebastien Dandurand", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" }, { "name": "Nandita Vijaykumar", "gender": "Female", "institution": "University of Toronto", "country": "CA", "position": "Assistant Professor" }, { "name": "Nilesh Ahuja", "gender": "unknown", "institution": "Intel", "country": "US", "position": "Researcher" }, { "name": "Nilesh Jain", "gender": "unknown", "institution": "Intel", "country": "", "position": "Principal Researcher" }, { "name": "Ruofan Liang", "gender": "Male", "institution": "NVIDIA", "country": "CA", "position": "Intern" }, { "name": "Selvakumar Panneer", "gender": "Male", "institution": "Intel", "country": "", "position": "Principal Researcher" }, { "name": "Xi Yan", "gender": "Female", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" }, { "name": "Yushi Guan", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "", "position": "PhD student" }, { "name": "Yuxuan Zhang", "gender": "unknown", "institution": "University of British Columbia", "country": "CA", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 66, "n_ref": 144, "n_ref_all": 208, "n_fig": 33, "n_tab": 5, "L_tab": 3518, "n_element_tab": 439, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2039, "n_element_tab_1": 237, "formula_len_all": 930, "formula_len_all_1": 615, "len_all": 264021, "len_all_1": 70495, "len_abs": 1442, "len_title": 100, "len_sents": 83400, "len_sents_1": 31330, "n_sents": 627, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 52, "L_abs": 1448, "title": "Retri3D: 3D Neural Graphics Representation Retrieval", "abs": "Learnable 3D Neural Graphics Representations (3DNGR) have emerged as promising 3D representations for reconstructing 3D scenes from 2D images. Numerous works, including Neural Radiance Fields (NeRF), 3D Gaussian Splatting (3DGS), and their variants, have significantly enhanced the quality of these representations. The ease of construction from 2D images, suitability for online viewing/sharing, and applications in game/art design downstream tasks make it a vital 3D representation, with potential creation of large numbers of such 3D models. This necessitates large data stores, local or online, to save 3D visual data in these formats. However, no existing framework enables accurate retrieval of stored 3DNGRs. In this work, we propose, Retri3D, a framework that enables accurate and efficient retrieval of 3D scenes represented as NGRs from large data stores using text queries. We introduce a novel Neural Field Artifact Analysis technique, combined with a Smart Camera Movement Module, to select clean views and navigate pre-trained 3DNGRs. These techniques enable accurate retrieval by selecting the best viewing directions in the 3D scene for high-quality visual feature embeddings. We demonstrate that Retri3D is compatible with any NGR representation. On the LERF and ScanNet++ datasets, we show significant improvement in retrieval accuracy compared to existing techniques, while being orders of magnitude faster and storage efficient.", "keywords": [ "Neural Graphics Representation; 3D Retrieval; Database" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 4, 3, 4, 4 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "q2VK1Z8XFo", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Alexander Tyurin", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Assistant Professor" }, { "name": "Kaja Gruntkowska", "gender": "Female", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Wojciech Anyszka", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "MS student" }, { "name": "Peter Richtarik", "gender": "Male", "institution": "King Abdullah University of Science and Technology (KAUST)", "country": "SA", "position": "Full Professor" } ], "n_formula": 230, "n_formula_1": 44, "n_ref_uni": 36, "n_ref": 73, "n_ref_all": 110, "n_fig": 4, "n_tab": 1, "L_tab": 831, "n_element_tab": 30, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 24776, "formula_len_all_1": 3360, "len_all": 208822, "len_all_1": 66108, "len_abs": 2998, "len_title": 87, "len_sents": 54383, "len_sents_1": 26991, "n_sents": 668, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1318, "title": "Tighter Performance Theory of FedExProx", "abs": "We revisit FedExProx -- a recently proposed distributed optimization method designed to enhance convergence properties of parallel proximal algorithms via extrapolation. In the process, we uncover a surprising flaw: its known theoretical guarantees on quadratic optimization tasks are no better than those offered by the vanilla Gradient Descent (GD) method. Motivated by this observation, we develop a novel analysis framework, establishing a tighter linear convergence rate for non-strongly convex quadratic problems. By incorporating both computation and communication costs, we demonstrate that FedExProx can indeed provably outperform GD, in stark contrast to the original analysis. Furthermore, we consider partial participation scenarios and analyze two adaptive extrapolation strategies-based on gradient diversity and Polyak stepsizes --- again significantly outperforming previous results. Moving beyond quadratics, we extend the applicability of our analysis to general functions satisfying the Polyak-\u0141ojasiewicz condition, outperforming the previous strongly convex analysis while operating under weaker assumptions. Backed by empirical results, our findings point to a new and stronger potential of FedExProx, paving the way for further exploration of the benefits of extrapolation in federated learning.", "keywords": [ "optimization", "federated learning", "proximal methods" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 1, 4 ] }, { "paper_id": "q2Lnyegkr8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Aaron Courville", "gender": "unknown", "institution": "University of Montreal", "country": "", "position": "Assistant Professor" }, { "name": "Evgenii Nikishin", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "PhD student" }, { "name": "Xu He", "gender": "Male", "institution": "MakerMaker AI", "country": "US", "position": "CTO" }, { "name": "Zhixuan Lin", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" } ], "n_formula": 62, "n_formula_1": 26, "n_ref_uni": 61, "n_ref": 128, "n_ref_all": 173, "n_fig": 19, "n_tab": 11, "L_tab": 4106, "n_element_tab": 678, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3627, "n_element_tab_1": 543, "formula_len_all": 12840, "formula_len_all_1": 10531, "len_all": 228760, "len_all_1": 77177, "len_abs": 1160, "len_title": 108, "len_sents": 49306, "len_sents_1": 23600, "n_sents": 480, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1171, "title": "Forgetting Transformer: Softmax Attention with a Forget Gate", "abs": "An essential component of modern recurrent sequence models is the *forget gate*. While Transformers do not have an explicit recurrent form, we show that a forget gate can be naturally incorporated into Transformers by down-weighting the unnormalized attention scores in a data-dependent way. We name the resulting model the *Forgetting Transformer*. We show that the Forgetting Transformer outperforms the Transformer on long-context language modeling, length extrapolation, and short-context downstream tasks, while performing on par with the Transformer on long-context downstream tasks. Moreover, it is fully compatible with the Flash Attention algorithm and does not require any positional embeddings. Several analyses, including the needle-in-the-haystack test, show that the Forgetting Transformer also retains the Transformer's superior long-context capabilities over recurrent sequence models such as Mamba-2, HGRN2, and DeltaNet. We also introduce a \"Pro\" block design that incorporates some common architectural components in recurrent sequence models and find it significantly improves the performance of both the Forgetting Transformer and the Transformer.", "keywords": [ "sequence model", "long-context sequence modeling", "Transformer", "softmax attention", "linear attention", "RNN", "language model" ], "rating_list": [ 6, 8, 5, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "q2DmkZ1wVe", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Jian Wu", "gender": "unknown", "institution": "Tokyo Institute of Technology, Tokyo Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Linyi Yang", "gender": "unknown", "institution": "University College London", "country": "GB", "position": "Researcher" }, { "name": "Yue Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" }, { "name": "Zhen Wang", "gender": "Male", "institution": "Tokyo Institute of Technology", "country": "JP", "position": "PhD student" }, { "name": "Manabu Okumura", "gender": "Male", "institution": "Institute of Science Tokyo", "country": "JP", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 41, "n_ref": 85, "n_ref_all": 114, "n_fig": 7, "n_tab": 9, "L_tab": 4185, "n_element_tab": 484, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 2466, "n_element_tab_1": 230, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 161983, "len_all_1": 57987, "len_abs": 2181, "len_title": 56, "len_sents": 43731, "len_sents_1": 25302, "n_sents": 291, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1731, "title": "CofCA: A STEP-WISE Counterfactual Multi-hop QA benchmark", "abs": "While Large Language Models (LLMs) excel in question-answering (QA) tasks, their real reasoning abilities on multiple evidence retrieval and integration on Multi-hop QA tasks remain less explored. Firstly, LLMs sometimes generate answers that rely on internal memory rather than retrieving evidence and reasoning in the given context, which brings concerns about the evaluation quality of real reasoning abilities. Although previous counterfactual QA benchmarks can separate the internal memory of LLMs, they focus solely on final QA performance, which is insufficient for reporting LLMs' real reasoning abilities. Because LLMs are expected to engage in intricate reasoning processes that involve evidence retrieval and answering a series of sub-questions from given passages. Moreover, current factual Multi-hop QA (MHQA) benchmarks are annotated on open-source corpora such as Wikipedia, although useful for multi-step reasoning evaluation, they show limitations due to the potential data contamination in LLMs' pre-training stage. To address these issues, we introduce the Step-wise and Counterfactual benchmark (CofCA), a novel evaluation benchmark consisting of factual data and counterfactual data that reveals LLMs' real reasoning abilities on multi-step reasoning and reasoning chain evaluation. Our experimental results reveal a significant performance gap of several LLMs between Wikipedia-based factual data and counterfactual data, deeming data contamination issues in existing benchmarks. Moreover, we observe that LLMs usually bypass the correct reasoning chain, showing an inflated multi-step reasoning performance. We believe that our CofCA benchmark will enhance and facilitate the evaluations of trustworthy LLMs.", "keywords": [ "LLM evaluation", "Multi-hop QA evaluation" ], "rating_list": [ 8, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "q2CcNvzgb7", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Linjian Meng", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Tianpei Yang", "gender": "Female", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Wenbin Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Wubing Chen", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "PhD student" }, { "name": "Yang Gao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Youzhi Zhang", "gender": "unknown", "institution": "Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "", "position": "Assistant Professor" } ], "n_formula": 67, "n_formula_1": 37, "n_ref_uni": 42, "n_ref": 117, "n_ref_all": 158, "n_fig": 17, "n_tab": 5, "L_tab": 927, "n_element_tab": 178, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8012, "formula_len_all_1": 4308, "len_all": 178522, "len_all_1": 74914, "len_abs": 1464, "len_title": 129, "len_sents": 54192, "len_sents_1": 30859, "n_sents": 441, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1472, "title": "Learning Nash Equilibria in Normal-Form Games via Approximating Stationary Points", "abs": "Nash equilibrium (NE) plays an important role in game theory. However, learning an NE in normal-form games (NFGs) is a complex, non-convex optimization problem. Deep Learning (DL), the cornerstone of modern artificial intelligence, has demonstrated remarkable empirical performance across various applications involving non-convex optimization. However, applying DL to learn an NE poses significant difficulties since most existing loss functions for using DL to learn an NE introduce bias under sampled play. A recent work proposed an unbiased loss function. Unfortunately, it suffers from high variance, which degrades the convergence rate. Moreover, learning an NE through this unbiased loss function entails finding a global minimum in a non-convex optimization problem, which is inherently difficult. To improve the convergence rate by mitigating the high variance associated with the existing unbiased loss function, we propose a novel loss function, named Nash Advantage Loss (NAL). NAL is unbiased and exhibits significantly lower variance than the existing unbiased loss function. In addition, an NE is a stationary point of NAL rather than having to be a global minimum, which improves the computational efficiency. Experimental results demonstrate that the algorithm minimizing NAL achieves significantly faster empirical convergence rates compared to previous algorithms, while also reducing the variance of estimated loss value by several orders of magnitude.", "keywords": [ "Deep Learning", "Nash Equilibrium", "Normal-Form Games" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "q1t0Lmvhty", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Gaowen Liu", "gender": "Female", "institution": "Cisco Systems", "country": "US", "position": "Researcher" }, { "name": "Nicu Sebe", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Full Professor" }, { "name": "Xiaojun Wu", "gender": "Male", "institution": "Jiangnan University", "country": "CN", "position": "Full Professor" }, { "name": "Yue Song", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Ziheng Chen", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "PhD student" } ], "n_formula": 47, "n_formula_1": 19, "n_ref_uni": 48, "n_ref": 170, "n_ref_all": 210, "n_fig": 3, "n_tab": 9, "L_tab": 1807, "n_element_tab": 149, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 746, "n_element_tab_1": 83, "formula_len_all": 3162, "formula_len_all_1": 1392, "len_all": 208286, "len_all_1": 69479, "len_abs": 1326, "len_title": 155, "len_sents": 55452, "len_sents_1": 27937, "n_sents": 549, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 106, "L_abs": 1398, "title": "Understanding Matrix Function Normalizations in Covariance Pooling through the Lens of Riemannian Geometry", "abs": "Global Covariance Pooling (GCP) has been demonstrated to improve the performance of Deep Neural Networks (DNNs) by exploiting second-order statistics of high-level representations. GCP typically performs classification of the covariance matrices by applying matrix function normalization, such as matrix logarithm or power, followed by a Euclidean classifier. However, covariance matrices inherently lie in a Riemannian manifold, known as the Symmetric Positive Definite (SPD) manifold. The current literature does not provide a satisfactory explanation of why Euclidean classifiers can be applied directly to Riemannian features after the normalization of the matrix power. To mitigate this gap, this paper provides a comprehensive and unified understanding of the matrix logarithm and power from a Riemannian geometry perspective. The underlying mechanism of matrix functions in GCP is interpreted from two perspectives: one based on tangent classifiers (Euclidean classifiers on the tangent space) and the other based on Riemannian classifiers. Via theoretical analysis and empirical validation through extensive experiments on fine-grained and large-scale visual classification datasets, we conclude that the working mechanism of the matrix functions should be attributed to the Riemannian classifiers they implicitly respect. The code is available at https://github.com/GitZH-Chen/RiemGCP.git.", "keywords": [ "Global covariance pooling", "SPD manifolds", "Representation Learning", "Riemannian Manifolds" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 3, 4, 2 ] }, { "paper_id": "q1UyoY3MgJ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Khashayar Gatmiry", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Lei Fang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Lizhe Fang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yifei Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yisen Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 40, "n_formula_1": 19, "n_ref_uni": 39, "n_ref": 106, "n_ref_all": 164, "n_fig": 13, "n_tab": 9, "L_tab": 3326, "n_element_tab": 496, "n_fig_1": 10, "n_tab_1": 5, "L_tab_1": 1736, "n_element_tab_1": 243, "formula_len_all": 3278, "formula_len_all_1": 1507, "len_all": 187274, "len_all_1": 73976, "len_abs": 1038, "len_title": 92, "len_sents": 59135, "len_sents_1": 28125, "n_sents": 531, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 44, "L_abs": 1049, "title": "Rethinking Invariance in In-context Learning", "abs": "In-Context Learning (ICL) has emerged as a pivotal capability of auto-regressive large language models, yet it is hindered by a notable sensitivity to the ordering of context examples regardless of their mutual independence. To address this issue, recent studies have introduced several variant algorithms of ICL that achieve permutation invariance. However, many of these do not exhibit comparable performance with the standard auto-regressive ICL algorithm. In this work, we identify two crucial elements in the design of an invariant ICL algorithm: information non-leakage and context interdependence, which are not simultaneously achieved by any of the existing methods. These investigations lead us to the proposed \\emph{Invariant ICL (InvICL)}, a methodology designed to achieve invariance in ICL while ensuring the two properties. Empirically, our findings reveal that InvICL surpasses previous models, both invariant and non-invariant, in most benchmark datasets, showcasing superior generalization capabilities across varying input lengths.", "keywords": [ "In-context Learning", "Permuation Invariance" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 4, 4, 3, 2 ], "presentation_list": [ 4, 4, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "q1Cv7Hp52y", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 5, "author_info_list": [ { "name": "Jianzhu Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jingyu Cao", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "", "position": "MS student" }, { "name": "Yuan Zhou", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Yunze Wu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhilei Xu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 47, "n_ref": 73, "n_ref_all": 101, "n_fig": 10, "n_tab": 7, "L_tab": 2852, "n_element_tab": 183, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 940, "n_element_tab_1": 71, "formula_len_all": 800, "formula_len_all_1": 760, "len_all": 157346, "len_all_1": 68777, "len_abs": 1226, "len_title": 149, "len_sents": 48484, "len_sents_1": 30412, "n_sents": 447, "n_sents_1": 261, "venue": "ICLR2025", "decision": "withdraw", "L_title": 99, "L_abs": 1233, "title": "From Skills to Plans: Automatic Skill Discovery and Symbolic Interpretation for Compositional Tasks", "abs": "Deep Reinforcement Learning (DRL) has struggled with pixel-based controlling tasks that have numerous entities, long sequences, and logical dependencies. Methods using structured representations have shown promise in generalizing to different object entities in manipulation tasks. However, they lack the ability to segment and reuse basic skills. Neuro-symbolic RL excels in handling long sequential decomposable tasks yet heavily relies on expert-designed predicates. To address these challenges, we introduce a novel pixel-based framework that combines entity-centric decision transformers with symbolic planning. Our approach first automatically discovers and learns basic skills through experiences in simple environments without human intervention. Then, we employ a genetic algorithm to enhance these basic skills with symbolic interpretations. Therefore, we convert the complex controlling problem into a planning problem. Taking advantage of symbolic planning and entity-centric skills, our model is inherently interpretable and provides compositional generalizability. The results of the experiments show that our method demonstrates superior performance in long-horizon sequential tasks and real-world object manipulation.", "keywords": [ "automatic skill discovery", "symbolic interpretation", "pixel-based controlling" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 1, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "q07DDpu8Xb", "primary_area": "causal reasoning", "n_author": 8, "author_info_list": [ { "name": "Anton van den Hengel", "gender": "unknown", "institution": "University of Adelaide", "country": "AU", "position": "Professor" }, { "name": "Biwei Huang", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Dong Gong", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "Assistant Professor" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Mingming Gong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Yuhang Liu", "gender": "unknown", "institution": "The University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Zhen Zhang", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Javen Shi", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Professor" } ], "n_formula": 32, "n_formula_1": 8, "n_ref_uni": 59, "n_ref": 162, "n_ref_all": 189, "n_fig": 10, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1961, "formula_len_all_1": 588, "len_all": 196855, "len_all_1": 77824, "len_abs": 1242, "len_title": 168, "len_sents": 67821, "len_sents_1": 40379, "n_sents": 468, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 120, "L_abs": 1247, "title": "Turning Challenges into Opportunities: How Distribution Shifts Enhance Identifiability in Causal Representation Learning", "abs": "Causal representation learning seeks to uncover latent causal variables and their relationships from observed, unstructured data, a task complicated by identifiability challenges. While distribution shifts, viewed as natural interventions on latent causal variables, often present difficulties in traditional machine learning tasks, they also create valuable opportunities for identifiability by introducing variability in latent variables. In this paper, we study a non-parametric condition characterizing the types of distribution shifts that contribute to identifiability within the context of latent additive noise models. We also present partial identifiability results when only a portion of distribution shifts meets the condition. Furthermore, we extend our findings to latent post-nonlinear causal models. Building on our theoretical results, we propose a practical algorithm facilitating the acquisition of reliable latent causal representations. Our algorithm, guided by our underlying theory, has demonstrated outstanding performance across a diverse range of synthetic and real-world datasets. The empirical observations closely align with the theoretical findings, affirming the robustness and effectiveness of our proposed approach.", "keywords": [ "Causal Representation Learning", "Latent Causal Models", "Identifiability", "Distribution Shifts" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "pzmbxkCBiq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Acyr Locatelli", "gender": "Male", "institution": "Cohere", "country": "", "position": "Researcher" }, { "name": "Matthieu Geist", "gender": "Male", "institution": "Cohere", "country": "FR", "position": "Researcher" }, { "name": "Max Bartolo", "gender": "unknown", "institution": "Cohere", "country": "GB", "position": "Researcher" }, { "name": "Sander Land", "gender": "Male", "institution": "Cohere", "country": "", "position": "Researcher" }, { "name": "Zhengxiang Shi", "gender": "Male", "institution": "University College London", "country": "", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 54, "n_ref": 108, "n_ref_all": 142, "n_fig": 13, "n_tab": 5, "L_tab": 977, "n_element_tab": 57, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 573, "formula_len_all_1": 576, "len_all": 199269, "len_all_1": 51851, "len_abs": 1629, "len_title": 73, "len_sents": 50938, "len_sents_1": 25270, "n_sents": 396, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1639, "title": "Understanding Likelihood Over-optimisation in Direct Alignment Algorithms", "abs": "Direct Alignment Algorithms (DAAs), such as Direct Preference Optimisation (DPO) and Identity Preference Optimisation (IPO), have emerged as alternatives to online Reinforcement Learning from Human Feedback (RLHF) algorithms such as Proximal Policy Optimisation (PPO) for aligning language models to human preferences, without the need for explicit reward modelling. These methods generally aim to increase the likelihood of generating better (preferred) completions while discouraging worse (non-preferred) ones, while staying close to the original model's behaviour. In this work, we explore the relationship between completion likelihood and model performance in state-of-the-art DAAs, and identify a critical issue of likelihood over-optimisation. Contrary to expectations, we find that higher likelihood of better completions and larger margins between better and worse completion likelihoods do not necessarily lead to better performance, and may even degrade it. Our analysis reveals that while higher likelihood correlates with better memorisation of factual knowledge patterns, a slightly lower completion likelihood tends to improve output diversity, thus leading to better generalisation to unseen scenarios. Moreover, we identify two key indicators that signal when over-optimised output diversity begins to harm performance: ***Decreasing Entropy over Top-k Tokens*** and ***Diminishing Top-k Probability Mass***. Our experimental results validate that these indicators are reliable signs of declining performance under different regularisations, helping prevent over-optimisation and improve alignment with human preferences.", "keywords": [ "Preference Learning", "Large Language Model", "Direct Alignment Algorithm" ], "rating_list": [ 3, 8, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 1, 2, 3, 4 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "pzasy8KRWK", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Bo Lei", "gender": "Male", "institution": "Lawrence Livermore National Labs", "country": "", "position": "Postdoc" }, { "name": "Victor M Castillo", "gender": "Male", "institution": "Lawrence Livermore National Labs", "country": "US", "position": "Principal Researcher" }, { "name": "Yeping Hu", "gender": "Female", "institution": "Lawrence Livermore National Labs", "country": "US", "position": "Researcher" } ], "n_formula": 31, "n_formula_1": 14, "n_ref_uni": 53, "n_ref": 87, "n_ref_all": 125, "n_fig": 11, "n_tab": 10, "L_tab": 5051, "n_element_tab": 772, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 31, "n_element_tab_1": 2, "formula_len_all": 1900, "formula_len_all_1": 1008, "len_all": 212033, "len_all_1": 59206, "len_abs": 1700, "len_title": 128, "len_sents": 71017, "len_sents_1": 27880, "n_sents": 467, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1709, "title": "Physically Aligned Hierarchical Mesh-based Network for Dynamic System Simulation", "abs": "Dynamic systems evolve through complex interactions, where local events influence global behaviors, reflecting the interconnected nature of real-world phenomena. Simulating such systems demands models that effectively capture both local and long-range dynamics, while maintaining a balance between accuracy and computational efficiency. However, existing mesh-based Graph Neural Network (GNN) methods often struggle to achieve both high accuracy and efficiency, especially when dealing with large datasets, complex mesh structures, and extensive long-range effects. Inspired by how real-world dynamic systems operate, we present the Mesh-based Multi-Segment Graph Network (MMSGN), a novel framework designed to address these challenges by leveraging a physically aligned hierarchical information exchange mechanism. MMSGN combines micro-level local interactions with macro-level global exchanges, aligning the hierarchical mesh structure with the system\u2019s physical properties to seamlessly capture both local and global dynamics. This approach enables precise modeling of complex behaviors while maintaining computational efficiency. We validate our model on multiple dynamic system datasets and compare it with several state-of-the-art methods. Our results demonstrate that MMSGN delivers superior accuracy and mesh quality, excels in managing long-range effects, and maintains high computational efficiency. Furthermore, MMSGN exhibits strong generalization capabilities, scaling effectively to larger physical domains. These advantages make MMSGN well-suited for simulating complex, large-scale dynamic systems across a variety of scenarios. Codes and data will be made publicly accessible upon acceptance.", "keywords": [ "dynamic system; physics simulation; solid mechanics; graph-based simulation;" ], "rating_list": [ 5, 8, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4, 3 ] }, { "paper_id": "pz0EK4g6AN", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Rui Yang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Tongyang Li", "gender": "Male", "institution": "Peking University", "country": "", "position": "Assistant Professor" }, { "name": "Yitao Liang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuntian Gu", "gender": "unknown", "institution": "Peking University", "country": "", "position": "Undergrad student" }, { "name": "Ziruo Wang", "gender": "Female", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Intern" } ], "n_formula": 20, "n_formula_1": 5, "n_ref_uni": 41, "n_ref": 54, "n_ref_all": 68, "n_fig": 6, "n_tab": 8, "L_tab": 7618, "n_element_tab": 338, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2435, "n_element_tab_1": 130, "formula_len_all": 1693, "formula_len_all_1": 524, "len_all": 286650, "len_all_1": 60677, "len_abs": 3035, "len_title": 124, "len_sents": 67381, "len_sents_1": 27824, "n_sents": 572, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1729, "title": "QCircuitNet: A Large-Scale Hierarchical Dataset for Quantum Algorithm Design", "abs": "Quantum computing is an emerging field recognized for the significant speedup it offers over classical computing through quantum algorithms. However, designing and implementing quantum algorithms pose challenges due to the complex nature of quantum mechanics and the necessity for precise control over quantum states. \nDespite the significant advancements in AI, there has been a lack of datasets specifically tailored for this purpose. \nIn this work, we introduce QCircuitNet, the first benchmark and test dataset designed to evaluate AI's capability in designing and implementing quantum algorithms in the form of quantum circuit codes. Unlike using AI for writing traditional codes, this task is fundamentally different and significantly more complicated due to highly flexible design space and intricate manipulation of qubits. \nOur key contributions include: \n1. A general framework which formulates the key features of quantum algorithm design task for Large Language Models.\n2. Implementation for a wide range of quantum algorithms from basic primitives to advanced applications, with easy extension to more quantum algorithms.\n3. Automatic validation and verification functions, allowing for iterative evaluation and interactive reasoning without human inspection.\n4. Promising potential as a training dataset through primitive fine-tuning results.\n\nWe observed several interesting experimental phenomena: fine-tuning does not always outperform few-shot learning, and LLMs tend to exhibit consistent error patterns. QCircuitNet provides a comprehensive benchmark for AI-driven quantum algorithm design, offering advantages in model evaluation and improvement, while also revealing some limitations of LLMs in this domain.", "keywords": [ "AI for science", "quantum computing", "quantum oracle construction", "quantum algorithm design" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 2, 4, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 2, 4, 3 ] }, { "paper_id": "pymXpl4qvi", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Jiajun Zhu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Pan Li", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Peihao Wang", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Pragya Srivastava", "gender": "Female", "institution": "Google DeepMind", "country": "IN", "position": "Researcher" }, { "name": "Ruisi Cai", "gender": "Female", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Yuehao Wang", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Zhangyang Wang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" } ], "n_formula": 49, "n_formula_1": 8, "n_ref_uni": 63, "n_ref": 206, "n_ref_all": 229, "n_fig": 12, "n_tab": 5, "L_tab": 379, "n_element_tab": 44, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1098, "n_element_tab_1": 74, "formula_len_all": 5385, "formula_len_all_1": 649, "len_all": 228487, "len_all_1": 74342, "len_abs": 1028, "len_title": 142, "len_sents": 65323, "len_sents_1": 33719, "n_sents": 567, "n_sents_1": 280, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1031, "title": "Understanding Bottlenecks of State Space Models through the Lens of Recency and Over-smoothing", "abs": "Structured State Space Models (SSMs) have emerged as alternatives to transformers, addressing the challenges of processing long sequences.\nWhile SSMs are often regarded as effective in capturing long-term dependencies, we theoretically demonstrate that they suffer from a strong recency bias.\nOur empirical findings reveal that this bias impairs the models' ability to recall distant information and introduces robustness issues.\nWe conducted scaling experiments and discovered that deeper structures in SSMs facilitate the learning of long contexts.\nHowever, our theoretical analysis reveal that as SSMs increase in depth, they exhibit a tendency toward over-smoothing, resulting in token representations becoming increasingly indistinguishable.\nThis over-smoothing phenomenon ultimately constrains the scalability of SSMs to achieve improved performance.\nCollectively, these findings highlight important limitations of SSMs and underscore the need for further research to address these challenges in long-range sequence modeling.", "keywords": [ "SSM", "Locality", "Recency", "Over-smoothing" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 2, 5, 2 ] }, { "paper_id": "py54X6mAEy", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "David A.W. Barton", "gender": "Male", "institution": "University of Bristol", "country": "", "position": "Full Professor" }, { "name": "Roussel Desmond Nzoyem", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "PhD student" }, { "name": "Tom Deakin", "gender": "Male", "institution": "University of Bristol", "country": "", "position": "Lecturer" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 41, "n_ref": 115, "n_ref_all": 150, "n_fig": 11, "n_tab": 6, "L_tab": 3687, "n_element_tab": 352, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 3170, "n_element_tab_1": 205, "formula_len_all": 537, "formula_len_all_1": 763, "len_all": 178827, "len_all_1": 73776, "len_abs": 1803, "len_title": 159, "len_sents": 58824, "len_sents_1": 31093, "n_sents": 412, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 110, "L_abs": 1839, "title": "Extending Contextual Self-Modulation: Meta-Learning Across Modalities, Task Dimensionalities, and Data Regimes", "abs": "Contextual Self-Modulation (CSM) is a potent regularization mechanism for Neural Context Flows (NCFs) which demonstrates powerful meta-learning on physical systems. However, CSM has limitations in its applicability across different modalities and in high-data regimes. In this work, we introduce two extensions: $i$CSM, which expands CSM to infinite-dimensional tasks, and StochasticNCF, which improves scalability. These extensions are demonstrated through comprehensive experimentation on a range of tasks, including dynamical systems with parameter variations, computer vision challenges, and curve fitting problems. $i$CSM embeds the contexts into an infinite-dimensional function space, as opposed to CSM which uses finite-dimensional context vectors. StochasticNCF enables the application of both CSM and $i$CSM to high-data scenarios by providing a \\rebut{low-cost} approximation of meta-gradient updates through a sampled set of nearest environments. Additionally, we incorporate higher-order Taylor expansions via Taylor-Mode automatic differentiation, revealing that higher-order approximations do not necessarily enhance generalization. Finally, we demonstrate how CSM can be integrated into other meta-learning frameworks with FlashCAVIA, a computationally efficient extension of the CAVIA meta-learning framework (Zintgraf et al. 2019). FlashCAVIA outperforms its predecessor across various benchmarks and reinforces the utility of bi-level optimization techniques. Together, these contributions \\rebut{reaffirm the powerful benefits of CSM, and suggest that its spectrum of addressable meta-learning and out-of-distribution tasks is limited to physical systems}. Our open-sourced library, designed for flexible integration of self-modulation into contextual meta-learning workflows, is available at \\url{AnonymousGitHubRepo}.", "keywords": [ "contextual meta-learning", "gradient-based meta-learning", "OOD generalisation", "neural context flows" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "py3RTHNT6J", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Ben Somers", "gender": "unknown", "institution": "Katholieke Universiteit Leuven", "country": "BE", "position": "Full Professor" }, { "name": "Dusan Grujicic", "gender": "Male", "institution": "KU Leuven", "country": "", "position": "PhD student" }, { "name": "Mingshi Li", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "PhD student" }, { "name": "Stien Heremans", "gender": "unknown", "institution": "KU Leuven", "country": "BE", "position": "Postdoc" }, { "name": "Matthew Blaschko", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 32, "n_ref": 39, "n_ref_all": 52, "n_fig": 6, "n_tab": 2, "L_tab": 1921, "n_element_tab": 255, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 270, "formula_len_all_1": 272, "len_all": 130290, "len_all_1": 54581, "len_abs": 1834, "len_title": 146, "len_sents": 29328, "len_sents_1": 26037, "n_sents": 190, "n_sents_1": 164, "venue": "ICLR2025", "decision": "withdraw", "L_title": 98, "L_abs": 1844, "title": "How big does your neural network have to be?: A Scaling Law Study in Multi-Spectral Remote Sensing", "abs": "Remote sensing imagery from systems such as Sentinel provides full coverage of the Earth's surface at around 10 meter resolution. The remote sensing community has transitioned to extensive use of deep learning models based on their high performance on benchmarks such as the ISPRS Vaihingen. Convolutional models such as UNet and ResNet variations are commonly employed for remote sensing but typically only accept three channels due to their development for RGB imagery, while Sentinel satellite systems have more than 10. Recently, a number of transformer architectures have also been proposed for remote sensing, but they typically have not been extensively benchmarked and have only been employed on rather small datasets. Meanwhile, it is becoming possible to obtain dense spatial land-use labels for entire first-level administrative divisions of some countries. Scaling law observations indicate that substantially larger, multi-spectral transformer models may provide a huge leap in the performance of remote sensing models in these settings. In this work, we develop a family of multi-spectral transformer models, which we evaluate across orders of magnitude differences in model parameters to evaluate their performance and scaling effectiveness on a densely labeled imagery dataset. We develop a novel multi-spectral attention strategy and demonstrate its effectiveness through ablations. We further show in this setting that models many orders of magnitude larger than conventional architectures such as UNet lead to substantial improvements in accuracy: a UNet++ model with 23M parameters results in less than 65\\% accuracy, while a multi-spectral transformer with 655M parameters yields an accuracy of over 95\\% on the Biological Valuation Map of Flanders.\nA link to open source code will be provided in the camera ready document.", "keywords": [ "machine learning", "remote sensing" ], "rating_list": [ 1, 1, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 1, 1, 3 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "py34636XvR", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Jaemoo Choi", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Jaewoong Choi", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 72, "n_formula_1": 45, "n_ref_uni": 48, "n_ref": 153, "n_ref_all": 160, "n_fig": 3, "n_tab": 3, "L_tab": 611, "n_element_tab": 46, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1215, "n_element_tab_1": 80, "formula_len_all": 5600, "formula_len_all_1": 3300, "len_all": 189910, "len_all_1": 70747, "len_abs": 1146, "len_title": 115, "len_sents": 50025, "len_sents_1": 27571, "n_sents": 529, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1140, "title": "Scalable Simulation-free Entropic Unbalanced Optimal Transport", "abs": "The Optimal Transport (OT) problem investigates a transport map that connects two distributions while minimizing a given cost function. Finding such a transport map has diverse applications in machine learning, such as generative modeling and image-to-image translation. In this paper, we introduce a scalable and simulation-free approach for solving the Entropic Unbalanced Optimal Transport (EUOT) problem. We derive the dynamical form of this EUOT problem, which is a generalization of the Schr\u00f6dinger bridges (SB) problem. Based on this, we derive dual formulation and optimality conditions of the EUOT problem from the stochastic optimal control interpretation. By leveraging these properties, we propose a simulation-free algorithm to solve EUOT, called Simulation-free EUOT (SF-EUOT). While existing SB models require expensive simulation costs during training and evaluation, our model achieves simulation-free training and one-step generation by utilizing the reciprocal property. Our model demonstrates significantly improved scalability in generative modeling and image-to-image translation tasks compared to previous SB methods.", "keywords": [ "Optimal Transport", "Unbalanced Optimal Transport", "Entropic Optimal Transport", "Generative models", "Image-to-image Translation" ], "rating_list": [ 6, 5, 6, 5, 6 ], "soundness_list": [ 3, 2, 4, 2, 3 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 2, 4, 4, 3, 4 ] }, { "paper_id": "pxy5wDMnzv", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "H. Vicky Zhao", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CA", "position": "Associate Professor" }, { "name": "Hangjing Zhang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Huisheng Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Mingxiao Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhuoshi Pan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" } ], "n_formula": 32, "n_formula_1": 17, "n_ref_uni": 28, "n_ref": 76, "n_ref_all": 112, "n_fig": 8, "n_tab": 6, "L_tab": 1395, "n_element_tab": 119, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 94, "n_element_tab_1": 4, "formula_len_all": 2131, "formula_len_all_1": 891, "len_all": 172625, "len_all_1": 66332, "len_abs": 1357, "len_title": 73, "len_sents": 55000, "len_sents_1": 31821, "n_sents": 410, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1374, "title": "InvestAlign: Align LLMs with Investor Decision-Making under Herd Behavior", "abs": "Studying investor decision-making processes under herd behavior is of great significance in microeconomics and behavioral finance. Large Language Models (LLMs) can be leveraged to assist in solving specific classes of complex investment problems. However, the investment decisions generated by existing LLMs often deviate from real-user data. One method to align LLMs with investor decision-making processes is Supervised Fine-Tuning (SFT), which requires a substantial amount of real-user data that is costly to collect and raises concerns about privacy and security. To overcome data scarcity, in this work, we propose **InvestAlign**, a low-cost and high-quality method that constructs large-scale SFT training datasets based on the theoretical solution to a specific simpler optimal investment problem, rather than the original complex one. We theoretically demonstrate that fine-tuning LLMs with these datasets leads to faster parameter convergence compared to using real-user data. By fine-tuning LLMs, we obtain **InvestAgent**s, which align more closely with real-user data than pre-SFT LLMs in both the simple and original complex problems examined in our study. This highlights **InvestAlign** as a promising approach with the potential to address complex optimal investment problems and align LLMs with investor decision-making processes in economics and finance.", "keywords": [ "Alignment", "Investment decision", "Large language model", "Supervised fine-tuning" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "pxwJU6rTAv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Bo Yang", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Jiaxian Guo", "gender": "Male", "institution": "Google Research", "country": "AU", "position": "Research Scientist" }, { "name": "Yusuke Iwasawa", "gender": "Male", "institution": "The University of Tokyo, The University of Tokyo", "country": "JP", "position": "Associate Professor" }, { "name": "Yutaka Matsuo", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 43, "n_ref": 74, "n_ref_all": 85, "n_fig": 2, "n_tab": 7, "L_tab": 2885, "n_element_tab": 131, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 285, "n_element_tab_1": 49, "formula_len_all": 211, "formula_len_all_1": 212, "len_all": 169137, "len_all_1": 57224, "len_abs": 1521, "len_title": 153, "len_sents": 57669, "len_sents_1": 28985, "n_sents": 398, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1450, "title": "ToM-agent: Large Language Models as Theory of Mind Aware Generative Agents with Counterfactual Reflection", "abs": "Recent studies have increasingly demonstrated that large language models (LLMs) possess significant theory of mind (ToM) capabilities, showing the potential for simulating the tracking of mental states in generative agents. In this study, we propose a novel paradigm called ToM-agent, designed to empower LLMs-based generative agents to simulate ToM in open-domain conversational interactions. ToM-agent disentangles the confidence from mental states, facilitating the emulation of an agent's perception of its counterpart's mental states, such as beliefs, desires, and intentions (BDIs). Using past conversation history and verbal reflections, ToM-Agent can dynamically adjust counterparts' inferred BDIs, along with related confidence levels. We further put forth a counterfactual intervention method that reflects on the gap between the predicted responses of counterparts and their real utterances, thereby enhancing the efficiency of reflection. Leveraging empathetic and persuasion dialogue datasets, we assess the advantages of implementing the ToM-agent with downstream tasks, as well as its performance in both the first-order and the second-order ToM. Our findings indicate that the ToM-agent can grasp the underlying reasons for their counterpart's behaviors beyond mere semantic-emotional supporting or decision-making based on common sense, providing new insights for studying large-scale LLMs-based simulation of human social behaviors.", "keywords": [ "generative agent", "large language model", "theory of mind", "BDI model", "open domain conversational AI", "reflection" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "pxclAomHat", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Jun Wang", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Professor" }, { "name": "Xu-Hui Liu", "gender": "unknown", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Yali Du", "gender": "Female", "institution": "King's College London", "country": "GB", "position": "Associate Professor" }, { "name": "Yang Yu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Professor" } ], "n_formula": 42, "n_formula_1": 12, "n_ref_uni": 35, "n_ref": 77, "n_ref_all": 103, "n_fig": 3, "n_tab": 11, "L_tab": 2984, "n_element_tab": 499, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 548, "n_element_tab_1": 45, "formula_len_all": 3338, "formula_len_all_1": 886, "len_all": 193811, "len_all_1": 75044, "len_abs": 1235, "len_title": 134, "len_sents": 56740, "len_sents_1": 33850, "n_sents": 548, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1228, "title": "On the Optimization Landscape of Low Rank Adaptation Methods for Large Language Models", "abs": "Training Large Language Models (LLMs) poses significant memory challenges, making low-rank adaptation methods an attractive solution. Previously, Low-Rank Adaptation (LoRA) addressed this by adding a trainable low-rank matrix to the frozen pre-trained weights in each layer, reducing the number of trainable parameters and optimizer states. GaLore, which compresses the gradient matrix instead of the weight matrix, has demonstrated superior performance to LoRA with faster convergence and reduced memory consumption. Despite their empirical success, the performance of these methods has not been fully understood or explained theoretically. In this paper, we analyze the optimization landscapes of LoRA, GaLore, and full-rank methods, revealing that GaLore benefits from fewer spurious local minima and a larger region that satisfies the \\pl, a variant of Polyak-\u0141ojasiewicz (PL) condition, leading to faster convergence. Our analysis leads to a novel method, GaRare, which further improves GaLore by using gradient random projection to reduce computational overhead. Practically, GaRare achieves strong performance in both pre-training and fine-tuning tasks, offering a more efficient approach to large-scale model adaptation.", "keywords": [ "large language model", "LoRA", "optimization" ], "rating_list": [ 5, 6, 5, 8, 8 ], "soundness_list": [ 3, 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 2, 3 ] }, { "paper_id": "pxYqG9GSpQ", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Chi Ming Chung", "gender": "Male", "institution": "Mediatek Research", "country": "", "position": "Researcher" }, { "name": "Egil Diau", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "MS student" }, { "name": "Min Sun", "gender": "Male", "institution": "National Tsing Hua University", "country": "TW", "position": "Assistant Professor" }, { "name": "Ting-Jun Wang", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "MS student" }, { "name": "Yueh-Feng Ku", "gender": "Male", "institution": "National Tsinghua University", "country": "TW", "position": "Undergrad student" }, { "name": "Winston Hsu", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 17, "n_ref": 36, "n_ref_all": 46, "n_fig": 2, "n_tab": 3, "L_tab": 263, "n_element_tab": 28, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 263, "n_element_tab_1": 28, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 71031, "len_all_1": 63686, "len_abs": 1348, "len_title": 112, "len_sents": 21742, "len_sents_1": 20931, "n_sents": 157, "n_sents_1": 149, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1284, "title": "3D Affordance Reconstruction from Egocentric Demonstration Video", "abs": "Developing robots capable of generalized skills remains an exceedingly challenging task. Drawing from psychology, the concept of affordance has emerged as a promising intermediate representation to guide robot manipulation. However, prior work has primarily focused on 2D affordances from video, neglecting critical spatial information such as camera positioning, absolute position, depth and geometry. In this paper, we present a novel training-free method that constructs 3D affordances from egocentric demonstration videos. To address the challenge of insufficient static, high-quality frames for 3D reconstruction in egocentric videos, we employ the 3D foundational model DUST3R, which reconstructs scenes from sparse images without requiring COLMAP. We analyze videos using hand detection to identify contact times and 2D contact points, reconstruct these interactions using DUST3R, and project the 2D contact points into 3D space using gaussian heatmaps. Finally, we derive hand trajectories through 3D hand pose estimation and process them using linear regression to integrate the spatiotemporal dynamics of human-object interactions. We demonstrate the effectiveness of our method on the ego4d-exo dataset for seven real-world hand-object manipulation tasks in cooking scenes.", "keywords": [ "3d Affordance", "Egocentric Vision", "Learning from Demonstration" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 1 ], "presentation_list": [ 2, 1, 1, 1 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "pxGucWt9vM", "primary_area": "foundation or frontier models, including LLMs", "n_author": 13, "author_info_list": [ { "name": "Chun Yuan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Han Bao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Haoli Bai", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "JiaxinHu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Jun Yao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Director" }, { "name": "Kang Zhao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Ruikang Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Wulong Liu", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Xianzhi Yu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xin Jiang", "gender": "Male", "institution": "Noah\u2019s Ark Lab, Huawei Technologies", "country": "HK", "position": "Principal Researcher" }, { "name": "Yuening Li", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yuxuan Sun", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "LU HOU", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "researcher" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 41, "n_ref": 125, "n_ref_all": 177, "n_fig": 18, "n_tab": 13, "L_tab": 6064, "n_element_tab": 855, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 2898, "n_element_tab_1": 461, "formula_len_all": 684, "formula_len_all_1": 353, "len_all": 191615, "len_all_1": 70421, "len_abs": 1450, "len_title": 97, "len_sents": 54417, "len_sents_1": 27952, "n_sents": 443, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1461, "title": "FlatQuant: Flatness Matters for LLM Quantization", "abs": "Recently, quantization has been widely used for the compression and acceleration of large language models~(LLMs). \nDue to the outliers in LLMs, it is crucial to flatten weights and activations to minimize quantization error with the equally spaced quantization points. Prior research explores various pre-quantization transformations to suppress outliers, such as per-channel scaling and Hadamard transformation. However, we observe that these transformed weights and activations can still remain steep and outspread. In this paper, we propose FlatQuant (Fast and Learnable Affine Transformation), a new post-training quantization approach to enhance flatness of weights and activations. Our approach identifies optimal affine transformations tailored to each linear layer, calibrated in hours via a lightweight objective. To reduce runtime overhead, we apply Kronecker decomposition to the transformation matrices, and fuse all operations in FlatQuant into a single kernel. Extensive experiments show that FlatQuant sets up a new state-of-the-art quantization benchmark. For instance, it achieves less than 1% accuracy drop for W4A4 quantization on the LLaMA-3-70B model, surpassing SpinQuant by 7.5%. For inference latency, FlatQuant reduces the slowdown induced by pre-quantization transformation from 0.26x of QuaRot to merely 0.07x, bringing up to 2.3x speedup for prefill and 1.7x speedup for decoding, respectively. Code will be released upon acceptance.", "keywords": [ "flatness", "post-training quantization", "LLM", "pre-quantization transformation" ], "rating_list": [ 5, 5, 3, 5, 8 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 4, 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 4, 5 ] }, { "paper_id": "px1674Wp3C", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Hang Xu", "gender": "Male", "institution": "Huawei Noah\u2018s Ark Lab", "country": "HK", "position": "Researcher" }, { "name": "Jiacheng Ye", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jianhua Han", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Jipeng Zhang", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Lanqing HONG", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Lingpeng Kong", "gender": "Male", "institution": "Department of Computer Science, The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Renjie Pi", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Wanjun Zhong", "gender": "Female", "institution": "ByteDance Inc.", "country": "", "position": "Researcher" }, { "name": "Yufei Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Zhenguo Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "JIAHUI GAO", "gender": "Female", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 108, "n_ref_all": 138, "n_fig": 4, "n_tab": 14, "L_tab": 2175, "n_element_tab": 241, "n_fig_1": 3, "n_tab_1": 9, "L_tab_1": 1952, "n_element_tab_1": 248, "formula_len_all": 1414, "formula_len_all_1": 1201, "len_all": 181645, "len_all_1": 74265, "len_abs": 1306, "len_title": 121, "len_sents": 47727, "len_sents_1": 29554, "n_sents": 421, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1230, "title": "G-LLaVA: Solving Geometric Problem with Multi-Modal Large Language Model", "abs": "Large language models (LLMs) have shown remarkable proficiency in human-level reasoning and generation capabilities, which encourages extensive research on their application in mathematical problem solving. However, current work has been largely focused on text-based mathematical problems, with limited investigation in problems involving multi-modal geometric information. Addressing this gap, we aim to enable LLMs to solve geometric problems by understanding image input. We first identify the limitations of current Multimodal Large Language Models (MLLMs) in this area: they struggle to accurately comprehend basic geometric elements and their relationships. To address these challenges, we leverage the inherent attribute of logical structure compactness in geometric figures, utilizing text-only Large Language Models (LLMs) to curate a comprehensive multimodal geometry dataset. This dataset, named Geo170k, contains more than 170K geometric image-caption and question-answer pairs. Utilizing the Geo170k dataset, we introduce G-LLaVA, a model that demonstrates exceptional performance in solving geometric problems. It significantly outperforms GPT4-V on the geometry task of MathVista benchmark with only 7B parameters.", "keywords": [ "Large Language Model", "Mathematical Reasoning" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "pwilycD30m", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Canran Xiao", "gender": "Not Specified", "institution": "Central South University", "country": "CN", "position": "PhD student" }, { "name": "Jiayu Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Liwei Hou", "gender": "Not Specified", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Weihao Luo", "gender": "unknown", "institution": "Donghua University, Shanghai", "country": "CN", "position": "PhD student" }, { "name": "Yinan Peng", "gender": "Male", "institution": "Shanghai Palmim Information Technology Ltd", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 19, "n_formula_1": 13, "n_ref_uni": 31, "n_ref": 54, "n_ref_all": 67, "n_fig": 4, "n_tab": 6, "L_tab": 1726, "n_element_tab": 226, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 759, "n_element_tab_1": 121, "formula_len_all": 933, "formula_len_all_1": 564, "len_all": 146675, "len_all_1": 65868, "len_abs": 1142, "len_title": 149, "len_sents": 47428, "len_sents_1": 27221, "n_sents": 383, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1146, "title": "Dynamic Cross-Layer Prefix Alignment for Resolving Label Preference Discrepancies in LLMs Fine-Tuning", "abs": "Fine-tuning large language models (LLMs) to adapt them for specialized downstream tasks is a common practice, yet existing methods overlook a critical issue: label preference discrepancies among different annotators. Such inconsistencies in labeling can significantly impair the model's robustness and generalization. In this work, we propose Dynamic Cross-Layer Preference Correction (DCPC), a novel self-supervised learning framework designed to mitigate these inconsistencies. DCPC incorporates a preference-sensitive similarity mechanism, cross-layer prefix alignment, and a Preference Correction Module (PCM) to dynamically adjust embeddings across transformer layers. By leveraging self-supervision, DCPC effectively aligns semantic representations and ensures consistency in label predictions, even in the presence of preference shifts. We evaluate DCPC across multiple tasks using prominent base models and introduce modified datasets that simulate real-world preference shifts. Our results show that DCPC consistently outperforms state-of-the-art Parameter-Efficient Fine-Tuning (PEFT) methods in handling label preference discrepancies.", "keywords": [ "Label preference discrepancies", "Cross-layer prefix sharing", "Large language model fine-tuning" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "pwUed4vzIn", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Jonathan S. Daniels", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Jordan Taylor", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Full Professor" }, { "name": "Shawn Lawrence Beaulieu", "gender": "Non-Binary", "institution": "University of Vermont", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 35, "n_ref": 53, "n_ref_all": 61, "n_fig": 6, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 128, "formula_len_all_1": 128, "len_all": 101251, "len_all_1": 49948, "len_abs": 1132, "len_title": 158, "len_sents": 26402, "len_sents_1": 24865, "n_sents": 166, "n_sents_1": 156, "venue": "ICLR2025", "decision": "withdraw", "L_title": 109, "L_abs": 1136, "title": "Intrinsic Behavioral Variability Facilitates Flexible Representations: A Neuromotor Developmental Perspective", "abs": "Dynamic human movement necessitates a dynamic representation of the body. The mechanisms underlying the initiation, development, and maintenance of such representations can provide a biological perspective to developing more flexible representations within computational agents. Taking inspiration from the prenatal twitches shown to initiate the human neuromotor representation, we question how these same twitches, present throughout development, may also facilitate subsequent motor adaptation. Across three experiments, we examine the influence twitches, as a form of intrinsic behavioral variability, may have in facilitating motor adaptation to novel situations. In a series of simulated reaching tasks, we trained agents to reach targets while overcoming behavioral, physiological, and neurological changes. Overall, we found evidence that agents exposed to intermittent behavioral variability outperformed their counterparts, showing greater neural weight variability, indicative of greater exploration. Taken together, this work provides a biologically plausible computational framework for flexible representation development.", "keywords": [ "motor", "development", "adaptation", "simulation", "supervised-learning", "unsupervised-learning" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 1, 2, 2, 1 ], "presentation_list": [ 1, 3, 1, 2 ], "contribution_list": [ 1, 1, 1, 1 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "pwNIOcr8fU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Aobo Li", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "PhD student" }, { "name": "Jinjian Wu", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Leida Li", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Weisheng Dong", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Yongxu Liu", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Lecturer" } ], "n_formula": 17, "n_formula_1": 8, "n_ref_uni": 7, "n_ref": 12, "n_ref_all": 25, "n_fig": 2, "n_tab": 9, "L_tab": 4241, "n_element_tab": 178, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 3384, "n_element_tab_1": 118, "formula_len_all": 1277, "formula_len_all_1": 745, "len_all": 115173, "len_all_1": 55799, "len_abs": 1580, "len_title": 136, "len_sents": 36170, "len_sents_1": 23749, "n_sents": 261, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1584, "title": "Towards Syn-to-Real IQA: A Novel Perspective on Reshaping Synthetic Data Distributions", "abs": "Blind Image Quality Assessment (BIQA) has advanced significantly through deep learning, but the scarcity of large-scale labeled datasets remains a challenge. While synthetic data offers a promising solution, models trained on existing synthetic datasets often show limited generalization ability. In this work, we make a key observation that representations learned from synthetic datasets often exhibit a discrete and clustered pattern that hinders regression performance: features of high-quality images cluster around reference images, while those of low-quality images cluster based on distortion types. Our analysis reveals that this issue stems from the distribution of synthetic data rather than model architecture. Consequently, we introduce a novel framework SynDR-IQA, which reshapes synthetic data distribution to enhance BIQA generalization. Based on theoretical derivations of sample diversity and redundancy's impact on generalization error, SynDR-IQA employs two strategies: distribution-aware diverse content upsampling, which enhances visual diversity while preserving content distribution, and density-aware redundant cluster downsampling, which balances samples by reducing the density of densely clustered areas. Extensive experiments across three cross-dataset settings (synthetic-to-authentic, synthetic-to-algorithmic, and synthetic-to-synthetic) demonstrate the effectiveness of our method. Additionally, as a data-based approach, SynDR-IQA can be coupled with model-based methods without increasing inference costs. The source code will be publicly available.", "keywords": [ "Blind Image Quality Assessment; Data Distribution Reshaping; Synthetic Data" ], "rating_list": [ 6, 6, 6, 6, 5 ], "soundness_list": [ 3, 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3, 1 ], "confidence_list": [ 5, 4, 3, 3, 4 ] }, { "paper_id": "pwKokorglv", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Haibin Yan", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Full Professor" }, { "name": "Jiwen Lu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xiuwei Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhenyu Wu", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "TW", "position": "PhD student" }, { "name": "Ziwei Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 35, "n_ref": 48, "n_ref_all": 64, "n_fig": 9, "n_tab": 7, "L_tab": 1665, "n_element_tab": 307, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 978, "n_element_tab_1": 224, "formula_len_all": 323, "formula_len_all_1": 264, "len_all": 146908, "len_all_1": 66443, "len_abs": 1636, "len_title": 102, "len_sents": 49184, "len_sents_1": 32576, "n_sents": 308, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1645, "title": "Embodied Instruction Following in Unknown Environments", "abs": "Enabling embodied agents to complete complex human instructions from natural language is crucial to autonomous systems in household services. Conventional methods can only accomplish human instructions in the known environment where all interactive objects are provided to the embodied agent, and directly deploying the existing approaches for the unknown environment usually generates infeasible plans that manipulate non-existing objects. On the contrary, we propose an embodied instruction following (EIF) method for complex tasks in the unknown environment, where the agent efficiently explores the unknown environment to generate feasible plans with existing objects to accomplish abstract instructions. Specifically, we build a hierarchical embodied instruction following framework including the high-level task planner and the low-level exploration controller with multimodal large language models. We then construct a semantic representation map of the scene with dynamic region attention to demonstrate the known visual clues, where the goal of task planning and scene exploration is aligned for human instruction. For the task planner, we generate the feasible step-by-step plans for human goal accomplishment according to the task completion process and the known visual clues. For the exploration controller, the optimal navigation or object interaction policy is predicted based on the generated step-wise plans and the known visual clues. \nThe experimental results demonstrate that our method can achieve 45.09\\% success rate in 204 complex human instructions such as making breakfast and tidying rooms in large house-level scenes.", "keywords": [ "Embodied instruction following", "multi-modal large language model", "embodied robotic" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 1, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "pwIGnH2LHJ", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Elijah Nnorom", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Gias Uddin", "gender": "Male", "institution": "York University", "country": "", "position": "Assistant Professor" }, { "name": "Haoran Xue", "gender": "Female", "institution": "York University", "country": "CA", "position": "PhD student" }, { "name": "Mohammad Mahdi Mohajer", "gender": "Male", "institution": "York University", "country": "CA", "position": "MS student" }, { "name": "Reem Aleithan", "gender": "Female", "institution": "York University", "country": "CA", "position": "MS student" }, { "name": "Song Wang", "gender": "Male", "institution": "York University", "country": "CA", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 6, "n_ref": 19, "n_ref_all": 48, "n_fig": 9, "n_tab": 5, "L_tab": 1146, "n_element_tab": 99, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1146, "n_element_tab_1": 99, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 107517, "len_all_1": 63546, "len_abs": 2521, "len_title": 94, "len_sents": 34232, "len_sents_1": 31751, "n_sents": 232, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 2536, "title": "SWE-Bench+: Enhanced Coding Benchmark for LLMs", "abs": "Large Language Models (LLMs) in Software Engineering (SE) can offer assistance for coding. To facilitate a rigorous evaluation of LLMs in practical coding contexts, Carlos et al. introduced the SWE-bench dataset, which comprises 2,294 real-world GitHub issues and their corresponding pull requests, collected from 12 widely used Python repositories. Several impressive LLM-based toolkits recently are developed and evaluated on this dataset. However, a systematic evaluation of the quality of SWE-bench remains missing. In this paper, we addressed this gap by presenting an empirical analysis of the SWE-bench dataset. We conducted a manual screening of instances where SWEAgent + GPT-4 successfully resolved issues by comparing the model-generated patches with the actual pull requests. SWE-Agent+GPT-4 was at the top of SWE-bench leaderboard during the time of our study. Our analysis reveals some critical issues with the SWE-bench dataset: 1) 32.67% of the successful patches involve \u201ccheating\u201d as the solutions were directly provided in the issue report\nor the comments. We refer to as \u2018solution leakage\u2019 problem. 2) 31.08% of the passed patches are suspicious patches due to weak test cases, i.e., the tests were not adequate to verify the correctness of a patch. When we filtered out these problematic issues, the resolution rate of SWE-Agent+GPT-4 drops from 12.47% to 3.97%. We also observed that the same data qualify issues also exist in the two variants of SWE-bench, i.e., SWE-bench Lite and SWE-Bench Verified. In addition, over 94% of the issues were created before LLM\u2019s knowledge cutoff dates, posing potential data leakage issues.\n\nThe critical problem in the current versions of SWE-bench dataset motivated us to refine it to build a more rigorous evaluation dataset SWE-Bench+. We created SWE-bench+ by collecting GitHub issues that were created after the training cutoff dates of the LLMs to prevent the potential data leakage problem. We also ensure that the issues collected do not contain solutions in their reports or comments. After carefully analyzing the passed instances from the SWE-Agent + GPT-4 model with the new dataset, SWE-Bench+, we observed a decline in the pass rate, dropping from 3.97% (as seen on the refined SWE-Bench) to a resolution rate of 0.55%. We further evaluated SWE-RAG + GPT-4, SWE-RAG + GPT-3.5, and AutoCodeRover + GPT-4o models on the new dataset to verify our findings, where the resolution rates of the models drop significantly, which are 0.73%, 0.55%, and 3.83%, respectively.", "keywords": [ "LLM", "benchmarks", "code generation" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "pw0fKwi0Pj", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Jiayu Yang", "gender": "unknown", "institution": "Australian National University", "country": "AU", "position": "PhD student" }, { "name": "Senbo Wang", "gender": "Male", "institution": "Tencent ", "country": "CN", "position": "Researcher" }, { "name": "Shan Luo", "gender": "Male", "institution": "King's College London, University of London", "country": "GB", "position": "Associate Professor" }, { "name": "Taizhang Shang", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Weixuan Sun", "gender": "Male", "institution": "Australian National University", "country": "", "position": "PhD student" }, { "name": "Weizhe Liu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Xibin Song", "gender": "Male", "institution": "Tencent XR Vision Lab", "country": "", "position": "Researcher" }, { "name": "Zhiyuan Wu", "gender": "Male", "institution": "King's College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Ziang Cheng", "gender": "Male", "institution": "Tencent", "country": "AU", "position": "Researcher" }, { "name": "Pan Ji", "gender": "Male", "institution": "Tencent XR Vision Labs", "country": "CN", "position": "Researcher" }, { "name": "Shenzhou", "gender": "Male", "institution": "Tencent XR Vision Labs", "country": "CN", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 15, "n_ref": 41, "n_ref_all": 60, "n_fig": 7, "n_tab": 5, "L_tab": 1430, "n_element_tab": 191, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3444, "n_element_tab_1": 202, "formula_len_all": 655, "formula_len_all_1": 558, "len_all": 112294, "len_all_1": 53525, "len_abs": 1174, "len_title": 134, "len_sents": 23967, "len_sents_1": 21070, "n_sents": 168, "n_sents_1": 145, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1179, "title": "InfiniteMesh: View Interpolation using Multi-view Diffusion for 3D Mesh Reconstruction", "abs": "We present InfiniteMesh, a feed-forward framework for efficient high-quality image-to-3D generation with view interpolation. Recent advancements in Large Reconstruction Model (LRM) have demonstrated significant potential in extracting 3D content from multi-view images produced by 2D diffusion models. Nevertheless, challenges remain as 2D diffusion models often struggle to generate dense images with strong multi-view consistency, and LRMs often exacerbate this multi-view inconsistency during 3D reconstruction. To address these issues, we propose a novel framework based on LRM that employs 2D diffusion-based view interpolation to enhance the quality of the generated mesh. Leveraging multi-view images produced by a 2D diffusion model, our approach introduces an Infinite View Interpolation module to generate interpolated images from main views. Subsequently, we employ a tri-plane-based mesh reconstruction strategy to extract robust tokens from these multiple generated images and produce the final mesh. Extensive experiments indicate that our method generates high-quality 3D content in terms of both texture and geometry, surpassing previous state-of-the-art methods.", "keywords": [ "View Interpolation", "Multi-view Diffusion", "3D Mesh Reconstruction" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "putnVJL2Rg", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Chen Ge", "gender": "Female", "institution": "University of Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Di ZHANG", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "VP" }, { "name": "Fuzheng Zhang", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Principal Researcher" }, { "name": "Hao Yi", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "MS student" }, { "name": "Sheng Ouyang", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "XiaomeiWang", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "Undergrad student" }, { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yulan Hu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "qingyang li", "gender": "Male", "institution": "Kuaishou LLM", "country": "CN", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 26, "n_ref": 77, "n_ref_all": 97, "n_fig": 4, "n_tab": 10, "L_tab": 2727, "n_element_tab": 374, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 2324, "n_element_tab_1": 324, "formula_len_all": 619, "formula_len_all_1": 621, "len_all": 123693, "len_all_1": 66305, "len_abs": 1432, "len_title": 115, "len_sents": 36301, "len_sents_1": 28806, "n_sents": 260, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1439, "title": "Recipes for Unbiased Reward Modeling Learning: An Empirically Study", "abs": "Reinforcement Learning from Human Feedback (RLHF) enhances the alignment between humans and large language models (LLMs), with Reward Models (RMs) playing a pivotal role. RLHF and sampling techniques, such as Best-of-N, require RMs to provide reliable rewards to guide policy training or sample selection. However, despite the advancement of LLMs, critical issues in RMs persist, such as overestimation on out-of-distribution (OOD) data (also known as reward hacking) and a preference for verbose outputs (length bias). These issues undermine the reliability of RM-generated rewards. Training an unbiased RM requires addressing these challenges, yet there is a lack of in-depth analysis on RMs. In this paper, we first decompose the RM training pipeline and identify three key aspects critical for developing an unbiased RM: 1) model architectures, 2) training paradigms, and 3) the influence of preference data. For each aspect, we conduct thorough empirical studies, revealing several insightful design considerations. Building on our findings, we develop an RM capable of mitigating the identified issues. This study represents the first comprehensive examination of various challenges from a holistic perspective in RM training, offering in-depth analyses of essential concerns and providing guidance for training unbiased RMs that can accurately guide downstream policies. The relevant code and models will be made publicly available.", "keywords": [ "Reinforcement Learning from Human Feedback", "Reward Modeling" ], "rating_list": [ 3, 5, 5, 1 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 2, 4, 1 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 5, 2, 5 ] }, { "paper_id": "pudmhZdV78", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Hrant Khachatrian", "gender": "unknown", "institution": "Yerevan State University", "country": "AM", "position": "Principal Researcher" }, { "name": "Hrayr Harutyunyan", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Rafayel Darbinyan", "gender": "Male", "institution": "Yerevan State University", "country": "", "position": "MS student" }, { "name": "Samvel Karapetyan", "gender": "Male", "institution": "Yerevan State University", "country": "AM", "position": "Undergrad student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 50, "n_ref": 66, "n_ref_all": 118, "n_fig": 23, "n_tab": 8, "L_tab": 8724, "n_element_tab": 880, "n_fig_1": 20, "n_tab_1": 1, "L_tab_1": 670, "n_element_tab_1": 135, "formula_len_all": 201, "formula_len_all_1": 201, "len_all": 185010, "len_all_1": 65534, "len_abs": 1127, "len_title": 56, "len_sents": 54525, "len_sents_1": 32232, "n_sents": 386, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1134, "title": "In-context learning in presence of spurious correlations", "abs": "Large language models exhibit a remarkable capacity for in-context learning, where they learn to solve tasks given a few examples.\nRecent work has shown that transformers can be trained to perform simple regression tasks in-context.\nThis work explores the possibility of training an in-context learner for classification tasks involving spurious features.\nWe find that the conventional approach of training in-context learners is susceptible to spurious features.\nMoreover, when the meta-training dataset includes instances of only one task, the conventional approach leads to task memorization and fails to produce a model that leverages context for predictions.\nBased on these observations, we propose a novel technique to train such a learner for a given classification task.\nRemarkably, this in-context learner matches and sometimes outperforms strong methods like ERM and GroupDRO.\nHowever, unlike these algorithms, it does not generalize well to other tasks.\nWe show that it is possible to obtain an in-context learner that generalizes to unseen tasks by training on a diverse dataset of synthetic in-context learning instances.", "keywords": [ "in-context learning", "spurious features", "out-of-distribution", "meta learning" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "puTxuiK2qO", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Damien MARTINS GOMES", "gender": "Male", "institution": "Concordia University", "country": "CA", "position": "MS student" }, { "name": "Eugene Belilovsky", "gender": "Male", "institution": "Concordia University, Montreal", "country": "CA", "position": "Assistant Professor" }, { "name": "Guy Wolf", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "Associate Professor" }, { "name": "Mahdi S. Hosseini", "gender": "Male", "institution": "Concordia University", "country": "CA", "position": "Assistant Professor" }, { "name": "Yanlei Zhang", "gender": "Male", "institution": "Montreal Institute for Learning Algorithm", "country": "", "position": "Postdoc" } ], "n_formula": 53, "n_formula_1": 20, "n_ref_uni": 60, "n_ref": 93, "n_ref_all": 159, "n_fig": 18, "n_tab": 17, "L_tab": 6745, "n_element_tab": 466, "n_fig_1": 7, "n_tab_1": 8, "L_tab_1": 2045, "n_element_tab_1": 127, "formula_len_all": 4059, "formula_len_all_1": 1763, "len_all": 274972, "len_all_1": 76556, "len_abs": 1254, "len_title": 117, "len_sents": 85534, "len_sents_1": 33243, "n_sents": 641, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1344, "title": "AdaFisher: Adaptive Second Order Optimization via Fisher Information", "abs": "First-order optimization methods are currently the mainstream in training deep neural networks (DNNs). Optimizers like Adam incorporate limited curvature information by employing the diagonal matrix preconditioning of the stochastic gradient during the training. Despite their widespread, second-order optimization algorithms exhibit superior convergence properties compared to their first-order counterparts e.g. Adam and SGD. However, their practicality in training DNNs is still limited due to increased per-iteration computations compared to the first-order methods. We present *AdaFisher*--an adaptive second-order optimizer that leverages a *diagonal block-Kronecker* approximation of the Fisher information matrix for adaptive gradient preconditioning. AdaFisher aims to bridge the gap between enhanced *convergence/generalization* capabilities and computational efficiency in second-order optimization framework for training DNNs. Despite the slow pace of second-order optimizers, we showcase that AdaFisher can be reliably adopted for image classification, language modeling and stands out for its stability and robustness in hyper-parameter tuning. We demonstrate that AdaFisher **outperforms the SOTA optimizers** in terms of both accuracy and convergence speed. Code is available from https://github.com/AtlasAnalyticsLab/AdaFisher.", "keywords": [ "Second Order Optimization", "Fisher Information", "Kronecker-factored Approximate Curvature", "Deep Learning", "Computer Vision", "Natural Language Processing" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "puGvShnqeA", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Akshay G Rao", "gender": "Male", "institution": "Indian Institute of Technology, Madras", "country": "IN", "position": "MS student" }, { "name": "Arun Rajkumar", "gender": "Male", "institution": "Indian Institute of Technology Madras", "country": "", "position": "Assistant Professor" }, { "name": "Chandra Shekar Lakshminarayanan", "gender": "Male", "institution": "Indian Institute of Technology, Madras", "country": "IN", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 13, "n_ref_uni": 6, "n_ref": 9, "n_ref_all": 48, "n_fig": 17, "n_tab": 4, "L_tab": 2823, "n_element_tab": 286, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 3031, "n_element_tab_1": 919, "formula_len_all": 1513, "formula_len_all_1": 1230, "len_all": 89303, "len_all_1": 63113, "len_abs": 1653, "len_title": 146, "len_sents": 29406, "len_sents_1": 21963, "n_sents": 205, "n_sents_1": 153, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1658, "title": "Interpreting Adversarial Attacks and Defenses using Architectures with Enhanced Interpretability", "abs": "Adversarial attacks in deep learning represent a significant threat to the integrity and\nreliability of machine learning models. These attacks involve intentionally crafting\nperturbations to input data that, while often imperceptible to humans, can lead\nto incorrect predictions by the model. This phenomenon exposes vulnerabilities\nin deep learning systems across various applications, from image recognition to\nnatural language processing. Adversarial training has been a popular defence\ntechnique against these adversarial attacks. The research community has been\nincreasingly interested in interpreting robust models and understanding how they\ndefend against attacks.\nIn this work, we capitalize on a network architecture, namely Deep Linearly Gated\nNetworks (DLGN), which has better interpretation capabilities than regular network\narchitectures. Using this architecture, we interpret robust models trained using PGD\nadversarial training and compare them with standard training. Feature networks\nin these architectures act as feature extractors, making them the only medium\nthrough which an adversary can attack the model. So, we use the feature network\nin this architecture with fully connected layers to analyse properties like alignment\nof the hyperplanes, hyperplane relation with PCA, and sub-network overlap among\nclasses and compare these properties between robust and standard models. We\nalso consider this architecture having CNN layers wherein we qualitatively and\nquantitatively contrast gating patterns between robust and standard models. We\nuse ideas from visualization to understand the representations used by robust and\nstandard models.", "keywords": [ "adversarial attacks", "adversarial defenses", "computer vision", "deep learning", "Interpretability" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 1, 2, 1 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "pu7a7JHW20", "primary_area": "generative models", "n_author": 1, "author_info_list": [ { "name": "Ahmad Ayaz Amin", "gender": "unknown", "institution": "Toronto Metropolitan University", "country": "", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 31, "n_ref": 48, "n_ref_all": 52, "n_fig": 0, "n_tab": 2, "L_tab": 97, "n_element_tab": 16, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 246, "n_element_tab_1": 24, "formula_len_all": 308, "formula_len_all_1": 308, "len_all": 71838, "len_all_1": 68733, "len_abs": 928, "len_title": 94, "len_sents": 23215, "len_sents_1": 23057, "n_sents": 151, "n_sents_1": 151, "venue": "ICLR2025", "decision": "withdraw", "L_title": 46, "L_abs": 929, "title": "Variational Inference with Unnormalized Priors", "abs": "Variational inference typically assumes normalized priors, limiting the expressiveness of generative models like Variational Autoencoders (VAEs). In this work, we propose a novel approach by replacing the prior \ud835\udc5d(\ud835\udc67) with an unnormalized energy-based distribution \nexp(\u2212\ud835\udc38(\ud835\udc67))/\ud835\udc4d, where \ud835\udc38(\ud835\udc67) is the energy function and \ud835\udc4d is the partition function. This leads to a variational lower bound that allows for two key innovations: (1) the incorporation of more powerful, flexible priors into the VAE framework, resulting in improved likelihood estimates and enhanced generative performance, and (2) the ability to train energy-based models (EBMs) without the need for computationally expensive Markov chain sampling, requiring only a small \ud835\udc5b > 1 importance samples from the posterior distribution. Our approach bridges VAEs and EBMs, providing a scalable and efficient framework for leveraging unnormalized priors in probabilistic models.", "keywords": [ "variational inference", "variational autoencoders", "generative models", "energy-based models" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 3, 2 ], "presentation_list": [ 1, 1, 1, 2, 2 ], "contribution_list": [ 1, 1, 1, 2, 1 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "ptjrpEGrGg", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Fan Yao", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Researcher" }, { "name": "Haifeng Xu", "gender": "Male", "institution": "University of Chicago", "country": "", "position": "Assistant Professor" }, { "name": "Xuefeng Liu", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Yuwei Cheng", "gender": "Female", "institution": "University of Chicago", "country": "US", "position": "PhD student" } ], "n_formula": 112, "n_formula_1": 28, "n_ref_uni": 48, "n_ref": 155, "n_ref_all": 167, "n_fig": 3, "n_tab": 6, "L_tab": 498, "n_element_tab": 48, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 66, "n_element_tab_1": 7, "formula_len_all": 13177, "formula_len_all_1": 3041, "len_all": 249192, "len_all_1": 78465, "len_abs": 1172, "len_title": 125, "len_sents": 78721, "len_sents_1": 35059, "n_sents": 694, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1252, "title": "Learning from Imperfect Human Feedback: A Tale from Corruption-Robust Dueling", "abs": "This paper studies Learning from Imperfect Human Feedback (LIHF), addressing the potential irrationality or imperfect perception when learning from comparative human feedback. Building on evidences that human's imperfection decays over time (i.e., humans learn to improve), we cast this problem as a concave-utility continuous-action dueling bandit but under a restricted form of corruption: i.e., the corruption scale is decaying over time as $t^{\\rho-1}$ for some ``imperfection rate'' $\\rho \\in [0, 1]$. \nWith $T$ as the total number of iterations, we establish a regret lower bound of $ \\Omega(\\max\\{\\sqrt{T}, T^{\\rho}\\})$ for LIHF, even when $\\rho$ is known. For the same setting, we develop the Robustified Stochastic Mirror Descent for Imperfect Dueling (RoSMID) algorithm, which achieves nearly optimal regret $\\tilde{\\mathcal{O}}(\\max\\{\\sqrt{T}, T^{\\rho}\\})$. Core to our analysis is a novel framework for analyzing gradient-based algorithms for dueling bandit under corruption, and we demonstrate its general applicability by showing how this framework can be easily applied to obtain corruption-robust guarantees for other popular gradient-based dueling bandit algorithms. Our theoretical results are validated by extensive experiments.", "keywords": [ "online learning", "dueling bandit" ], "rating_list": [ 8, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "ptTt8mhS7n", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 12, "author_info_list": [ { "name": "Binhua Li", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Dingzirui Wang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Fei Huang", "gender": "Male", "institution": "Alibaba Group US", "country": "US", "position": "Senior Research Director" }, { "name": "Longxu Dou", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Researcher" }, { "name": "Qiguang Chen", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Qingfu Zhu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Rongyu Cao", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Wanxiang Che", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Xiao Xu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Xuanliang Zhang", "gender": "Female", "institution": "Harbin Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Yingwei Ma", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yongbin Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 4, "n_ref_uni": 28, "n_ref": 108, "n_ref_all": 160, "n_fig": 11, "n_tab": 10, "L_tab": 2324, "n_element_tab": 240, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 713, "n_element_tab_1": 110, "formula_len_all": 1005, "formula_len_all_1": 179, "len_all": 228841, "len_all_1": 61481, "len_abs": 1093, "len_title": 131, "len_sents": 49141, "len_sents_1": 30106, "n_sents": 312, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1079, "title": "In-Context Transfer Learning: Demonstration Synthesis by Transferring Similar Tasks", "abs": "In-context learning (ICL) is an effective approach to help large language models (LLMs) adapt to various tasks by providing demonstrations of the target task. Considering the high cost of labeling demonstrations, many methods propose synthesizing demonstrations from scratch using LLMs. However, the quality of the demonstrations synthesized from scratch is limited by the capabilities and knowledge of LLMs. To address this, inspired by transfer learning, we propose In-Context Transfer Learning (ICTL), which synthesizes target task demonstrations by transferring labeled demonstrations from similar source tasks. ICTL consists of two steps: source sampling and target transfer. First, we define an optimization objective, which minimizes transfer error to sample source demonstrations similar to the target task. Then, we employ LLMs to transfer the sampled source demonstrations to match the definition and format of the target task. Experiments on Super-NI show that ICTL outperforms synthesis from scratch by 2.0% on average, demonstrating the effectiveness of our method.", "keywords": [ "In-Context Learning", "Transfer Learning", "Demonstration Synthesis" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "pshLnZzIbW", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Canh V. Pham", "gender": "Male", "institution": "Phenikaa University", "country": "VN", "position": "Associate Professor" }, { "name": "Dung T.K. Ha", "gender": "Female", "institution": "Vietnam National University Hanoi", "country": "VN", "position": "PhD student" }, { "name": "My Thai", "gender": "Female", "institution": "University of Florida", "country": "US", "position": "Full Professor" }, { "name": "Tan Dinh Tran", "gender": "Male", "institution": "Vietnam National University Hanoi", "country": "", "position": "PhD student" } ], "n_formula": 212, "n_formula_1": 23, "n_ref_uni": 28, "n_ref": 111, "n_ref_all": 129, "n_fig": 3, "n_tab": 5, "L_tab": 1364, "n_element_tab": 45, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1783, "n_element_tab_1": 34, "formula_len_all": 24078, "formula_len_all_1": 5429, "len_all": 205691, "len_all_1": 66808, "len_abs": 1261, "len_title": 136, "len_sents": 55146, "len_sents_1": 24754, "n_sents": 596, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1389, "title": "Boosting Parallel Algorithms in Linear Queries for Non-Monotone Submodular Maximization", "abs": "In this work, we propose two efficient parallel algorithms, $\\mathsf{LinAst}$ and $\\mathsf{LinAtg}$, that improve both the approximation ratio and query complexity of existing practical parallel algorithms for the non-monotone submodular maximization over the ground set of sized $n$ under a cardinality constraint $k$. Specifically, our algorithms keep the best adaptive complexity of $O(\\log n)$ while significantly improving the approximation ratio from $1/6-\\epsilon$ to $0.193-\\epsilon$ and reducing the query complexity from $O(n\\log (k))$ to $O(n)$. \nThe key building block of our algorithms is $\\mathsf{LinAdapt}$, a constant approximation ratio within $O(\\log n)$ sequence rounds and linear queries. $\\mathsf{LinAdapt}$ can reduce the query complexity by providing $O(1)$ guesses of the optimal value. We further introduce the $\\mathsf{BoostAdapt}$ algorithm returning a better ratio of $1/4-\\epsilon$ within $O(\\log (n)\\log (k))$ adaptive complexity and $O(n\\log (k))$ query complexity. Our $\\mathsf{BoostAdapt}$ works in a novel staggered greedy threshold framework that alternately constructs two disjoint sets in $O(\\log k)$ sequential rounds. Besides theoretical analysis, the experiment results on validated benchmarks confirm the superiority of our algorithms in terms of solution quality, the number of required queries, and running time over cutting-edge algorithms.", "keywords": [ "Submodular Maximization", "Non-Monotone", "Parallel Algorithms" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "psIymxANmd", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jing Shi", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Saehyung Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Seunghyun Yoon", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Researcher" }, { "name": "Sungroh Yoon", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Trung Bui", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 34, "n_ref": 89, "n_ref_all": 110, "n_fig": 8, "n_tab": 5, "L_tab": 1061, "n_element_tab": 166, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1109, "n_element_tab_1": 148, "formula_len_all": 125, "formula_len_all_1": 125, "len_all": 150458, "len_all_1": 66054, "len_abs": 1436, "len_title": 126, "len_sents": 40734, "len_sents_1": 31455, "n_sents": 310, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1444, "title": "Enhancing Factuality in Detailed Image Captioning with LLM-MLLM Collaboration", "abs": "Multimodal large language models (MLLMs) capable of interpreting images can generate highly detailed and extensive captions, owing to their advanced language modeling capabilities. However, the captions they produce frequently contain hallucinations. Furthermore, our empirical analysis reveals that existing hallucination detection methods are less effective in detailed image captioning tasks. We attribute this to the increasing reliance of MLLMs on their own generated text, rather than the input image, as the sequence length grows. To address this issue, we propose a novel corrector-based method that decomposes a given caption into atomic propositions, evaluates the factuality of each unit, and revises the caption accordingly. Our method is training-free and can be applied in a plug-and-play manner to any captioning model. Additionally, we introduce an evaluation framework and a benchmark dataset to facilitate the systematic analysis of detailed captions. Our experiments demonstrate that existing approaches to improve the factuality of MLLM outputs may fall short in detailed image captioning tasks. In contrast, our proposed method significantly enhances the factual accuracy of captions, even improving those generated by GPT-4V. Finally, we highlight a limitation of VQA-centric benchmarking by demonstrating that an MLLM's performance on VQA benchmarks may not correlate with its ability to generate detailed image captions.", "keywords": [ "Image Captioning", "Caption Evaluation Metric", "Multimodal Large Language Model", "Large Language Model" ], "rating_list": [ 3, 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 1, 2 ], "confidence_list": [ 4, 5, 4, 3, 4 ] }, { "paper_id": "psG83N6GZi", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Yassine ABBAHADDOU", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "FR", "position": "PhD student" }, { "name": "Amine Mohamed Aboussalah", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 32, "n_formula_1": 18, "n_ref_uni": 43, "n_ref": 51, "n_ref_all": 63, "n_fig": 2, "n_tab": 3, "L_tab": 878, "n_element_tab": 99, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 774, "n_element_tab_1": 79, "formula_len_all": 1920, "formula_len_all_1": 1053, "len_all": 134360, "len_all_1": 63857, "len_abs": 1387, "len_title": 179, "len_sents": 39395, "len_sents_1": 28751, "n_sents": 332, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 130, "L_abs": 1392, "title": "Grassmannian Geometry Meets Dynamic Mode Decomposition in DMD-GEN: A New Metric for Mode Collapse in Time Series Generative Models", "abs": "Generative models like Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) often fail to capture the full diversity of their training data, leading to mode collapse. While this issue is well-explored in image generation, it remains underinvestigated for time series data. We introduce a new definition of mode collapse specific to time series and propose a novel metric, DMD-GEN, to quantify its severity. Our metric utilizes Dynamic Mode Decomposition (DMD), a data-driven technique for identifying coherent spatiotemporal patterns, and employs Optimal Transport between DMD eigenvectors to assess discrepancies between the underlying dynamics of the original and generated data. This approach not only quantifies the preservation of essential dynamic characteristics but also provides interpretability by pinpointing which modes have collapsed. We validate DMD-GEN on both synthetic and real-world datasets using various generative models, including TimeGAN, TimeVAE, and DiffusionTS. The results demonstrate that DMD-GEN correlates well with traditional evaluation metrics for static data while offering the advantage of applicability to dynamic data. This work offers for the first time a definition of mode collapse for time series, improving understanding, and forming the basis of our tool for assessing and improving generative models in the time series domain.", "keywords": [ "Time series", "Generative models", "Mode Collapse" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 2, 2 ] }, { "paper_id": "prTI7MSt2X", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Alan Lahoud", "gender": "Male", "institution": "\u00d6rebro University", "country": "SE", "position": "PhD student" }, { "name": "Erik Schaffernicht", "gender": "Male", "institution": "\u00d6rebro University", "country": "SE", "position": "Lecturer" }, { "name": "Johannes A. Stork", "gender": "Not Specified", "institution": "\u00d6rebro University", "country": "", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 24, "n_ref": 41, "n_ref_all": 68, "n_fig": 12, "n_tab": 1, "L_tab": 213, "n_element_tab": 16, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 310, "formula_len_all_1": 227, "len_all": 97419, "len_all_1": 61112, "len_abs": 1301, "len_title": 152, "len_sents": 33153, "len_sents_1": 29308, "n_sents": 228, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1303, "title": "IO-LVM: Inverse optimization latent variable models with applications to inferring and explaining paths", "abs": "Learning representations from solutions of constrained optimization problems (COPs) with unknown cost functions is challenging, as models like (Variational) Autoencoders struggle to capture constraints to decode structured outputs. We propose an inverse optimization latent variable model (IO-LVM) that constructs a latent space of COP costs based on observed decisions, enabling the inference of feasible and meaningful solutions by reconstructing them with a COP solver. To achieve this, we leverage estimated gradients of a Fenchel-Young loss through a non-differentiable deterministic solver while shaping the embedding space. In contrast to established Inverse Optimization or Inverse Reinforcement Learning methods, which typically identify a single or context-conditioned cost function, we exploit the learned representation to capture underlying COP cost structures and identify solutions likely originating from different agents, each using distinct or slightly different cost functions when making decisions. Using both synthetic and actual ship routing data, we validate our approach through experiments on path planning problems using the Dijkstra algorithm, demonstrating the interpretability of the latent space and its effectiveness in path inference and path distribution reconstruction.", "keywords": [ "deep learning", "path planning", "latent space" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 3, 3, 2 ] }, { "paper_id": "prSJlvWrgE", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Lifei Chen", "gender": "Male", "institution": "Fujian Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Shengrui Wang", "gender": "unknown", "institution": "Universit\u00e9 de Sherbrooke", "country": "", "position": "Full Professor" }, { "name": "KUNPENG XU", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Postdoc" } ], "n_formula": 48, "n_formula_1": 11, "n_ref_uni": 33, "n_ref": 59, "n_ref_all": 99, "n_fig": 21, "n_tab": 10, "L_tab": 7270, "n_element_tab": 1195, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1843, "n_element_tab_1": 309, "formula_len_all": 3417, "formula_len_all_1": 871, "len_all": 235826, "len_all_1": 72358, "len_abs": 1636, "len_title": 153, "len_sents": 75183, "len_sents_1": 31308, "n_sents": 561, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1312, "title": "Drift2Matrix: Kernel-Induced Self Representation for Concept Drift Adaptation in Co-evolving Time Series", "abs": "In the realm of time series analysis, tackling the phenomenon of concept drift poses a significant challenge. Concept drift -- characterized by the evolving statistical properties of time series data, affects the reliability and accuracy of conventional analysis models. This is particularly evident in co-evolving scenarios where interactions among variables are crucial. This paper presents Drift2Matrix, a novel framework that leverages kernel-induced self-representation for adaptive responses to concept drift in time series. Drift2Matrix employs a kernel-based learning mechanism to generate a representation matrix, encapsulating the inherent dynamics of co-evolving time series. This matrix serves as a key tool for identification and adaptation to concept drift by observing its temporal variations. Furthermore, Drift2Matrix effectively identifies prevailing patterns and offers insights into emerging trends through pattern evolution analysis. Our empirical evaluation of Drift2Matrix across various datasets demonstrates its effectiveness in handling the complexities of concept drift. This approach introduces a novel perspective in the theoretical domain of co-evolving time series analysis, enhancing adaptability and accuracy in the face of dynamic data environments. Code is available at GitHub.", "keywords": [ "co-evolving time series", "concept drift", "kernel representation learning" ], "rating_list": [ 6, 6, 8, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "pr37sbuhVa", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Anwen Hu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Fei Huang", "gender": "Male", "institution": "Alibaba Group US", "country": "US", "position": "Senior Research Director" }, { "name": "Haiyang Xu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Haowei Liu", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Ji Zhang", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Senior Staff Engineer" }, { "name": "Jiabo Ye", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "PhD student" }, { "name": "Jingren Zhou", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Ming Yan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Instructor" }, { "name": "Qi Qian", "gender": "unknown", "institution": "Zoom", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 47, "n_ref": 112, "n_ref_all": 140, "n_fig": 12, "n_tab": 15, "L_tab": 2421, "n_element_tab": 428, "n_fig_1": 7, "n_tab_1": 9, "L_tab_1": 3132, "n_element_tab_1": 419, "formula_len_all": 153, "formula_len_all_1": 273, "len_all": 188909, "len_all_1": 57780, "len_abs": 485, "len_title": 138, "len_sents": 45598, "len_sents_1": 21750, "n_sents": 385, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1341, "title": "mPLUG-Owl3: Towards Long Image-Sequence Understanding in Multi-Modal Large Language Models", "abs": "Multi-modal Large Language Models have demonstrated remarkable capabilities in executing instructions for a variety of single-image tasks. Despite this progress, significant challenges remain in modeling long image sequences. In this work, we introduce the versatile multi-modal large language model, mPLUG-Owl3, which enhances the capability for long image-sequence understanding in scenarios that incorporate retrieved image-text knowledge, multimodal in-context examples, and lengthy videos. Specifically, we propose novel hyper attention blocks to efficiently integrate vision and language into a common language-guided semantic space, thereby facilitating the processing of extended multi-image scenarios. We conduct evaluations on 21 benchmarks that cover single/multi-image, and short/long video understanding. mPLUG-Owl3 achieves competitive performance with the state-of-the-art methods while reducing inference time and memory usage by 87.8\\% and 48.5\\% in average. Moreover, we propose a Distractor Resistance evaluation to assess the ability of models to maintain focus amidst distractions. mPLUG-Owl3 also demonstrates outstanding performance in distractor resistance on ultra-long visual sequence inputs. We hope that mPLUG-Owl3 can contribute to the development of more efficient and powerful multimodal large language models.", "keywords": [ "multimodal large language model", "long sequence", "efficient multimodal understanding" ], "rating_list": [ 8, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "pqxSDbX8XT", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Huanqian Yan", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Zhuodong Liu", "gender": "Male", "institution": "Qiyuan Lab", "country": "CN", "position": "Researcher" }, { "name": "\u8d85\u7fa4 \u674e", "gender": "Male", "institution": "Qiyuan Lab", "country": "CN", "position": "Researcher" }, { "name": "Hang Su", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 0, "n_ref": 139, "n_ref_all": 164, "n_fig": 13, "n_tab": 9, "L_tab": 925, "n_element_tab": 105, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1217, "n_element_tab_1": 90, "formula_len_all": 266, "formula_len_all_1": 266, "len_all": 135555, "len_all_1": 67918, "len_abs": 222, "len_title": 111, "len_sents": 41815, "len_sents_1": 32805, "n_sents": 329, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1353, "title": "CAPGen: An Environment-Adaptive Generator of Adversarial Patches", "abs": "Adversarial patches, often used to provide physical stealth protection for critical assets and assess perception algorithm robustness, usually neglect the need for visual harmony with the background environment, making them easily noticeable. Moreover, existing methods primarily concentrate on improving attack performance, disregarding the intricate dynamics of adversarial patch elements. In this work, we introduce the Camouflaged Adversarial Pattern Generator (CAPGen), a novel approach that leverages specific base colors from the surrounding environment to produce patches that seamlessly blend with their background for superior visual stealthiness while maintaining robust adversarial performance. We delve into the influence of both patterns (i.e., color-agnostic texture information) and colors on the effectiveness of attacks facilitated by patches, discovering that patterns exert a more pronounced effect on performance than colors. Based on these findings, we propose a rapid generation strategy for adversarial patches. This involves updating the colors of high-performance adversarial patches to align with those of the new environment, ensuring visual stealthiness without compromising adversarial impact. This paper is the first to comprehensively examine the roles played by patterns and colors in the context of adversarial patches.", "keywords": [ "Physical attack", "Adversarial patch", "Environment consistency" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "pqeWzZTrZY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Bin Xiao", "gender": "unknown", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Kaisheng Liang", "gender": "unknown", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Yanjie Li", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 45, "n_ref": 111, "n_ref_all": 141, "n_fig": 19, "n_tab": 7, "L_tab": 1565, "n_element_tab": 258, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 636, "n_element_tab_1": 117, "formula_len_all": 1032, "formula_len_all_1": 700, "len_all": 167765, "len_all_1": 56803, "len_abs": 1780, "len_title": 154, "len_sents": 45743, "len_sents_1": 24987, "n_sents": 390, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1759, "title": "UV-Attack: Physical-World Adversarial Attacks for Person Detection via Dynamic-NeRF-based UV Mapping", "abs": "Recent works have attacked person detectors using adversarial patches or static-3D-model-based texture modifications. However, these methods suffer from low attack success rates when faced with significant human movements. The primary challenge stems from the highly non-rigid nature of the human body and clothing. Current attacks fail to model these 3D non-rigid deformations caused by varied actions.\nFortunately, recent research has shown significant progress in using NeRF for dynamic human modeling. \nIn this paper, we introduce \\texttt{UV-Attack}, a novel physical adversarial attack achieving high attack success rates in scenarios involving extensive and unseen actions. We address the challenges above by leveraging dynamic-NeRF-based UV mapping. Our method can generate human images across diverse actions and viewpoints and even create novel unseen actions by sampling from the SMPL parameter space. While dynamic NeRF models are capable of modeling human bodies, modifying their clothing textures is challenging due to the texture being embedded within neural network parameters.\nTo overcome this, \\texttt{UV-Attack} generates UV maps instead of RGB images and modifies the texture stacks. This approach enables real-time texture edits and makes attacks more practical. Finally, we propose a novel Expectation over Pose Transformation loss (EoPT) to improve the evasion success rate on unseen poses and views.\nOur experiments show that \\texttt{UV-Attack} achieves a 92.75\\% attack success rate against the FastRCNN model across varied poses in dynamic video settings, significantly outperforming the state-of-the-art AdvCaT attack, which only had a 28.50\\% ASR. Moreover, we achieve 49.5\\% ASR on the latest YOLOv8 detector in black-box settings.", "keywords": [ "Adversarial Attack; Person Detection; NeRF; UV-Mapping" ], "rating_list": [ 8, 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 4, 3, 2 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 5, 3, 2 ] }, { "paper_id": "pqOjj90Vwp", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 1, "author_info_list": [ { "name": "Tuo Xu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" } ], "n_formula": 82, "n_formula_1": 22, "n_ref_uni": 37, "n_ref": 104, "n_ref_all": 138, "n_fig": 14, "n_tab": 3, "L_tab": 1048, "n_element_tab": 120, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8465, "formula_len_all_1": 3084, "len_all": 279694, "len_all_1": 69049, "len_abs": 1147, "len_title": 107, "len_sents": 93542, "len_sents_1": 29788, "n_sents": 953, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 59, "L_abs": 1153, "title": "Towards a Complete Logical Framework for GNN Expressiveness", "abs": "Designing expressive Graph neural networks (GNNs) is an important topic in graph machine learning fields. Traditionally, the Weisfeiler-Lehman (WL) test has been the primary measure for evaluating GNN expressiveness. However, high-order WL tests can be obscure, making it challenging to discern the specific graph patterns captured by them. Given the connection between WL tests and first-order logic, some have explored the logical expressiveness of Message Passing Neural Networks. This paper aims to establish a comprehensive and systematic relationship between GNNs and logic. We propose a framework for identifying the equivalent logical formulas for arbitrary GNN architectures, which not only explains existing models, but also provides inspiration for future research. As case studies, we analyze multiple classes of prominent GNNs within this framework, unifying different subareas of the field. Additionally, we conduct a detailed examination of homomorphism expressivity from a logical perspective and present a general method for determining the homomorphism expressivity of arbitrary GNN models, as well as addressing several open problems.", "keywords": [ "graph neural networks", "logic" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 4, 4, 2 ], "confidence_list": [ 3, 2, 3 ] }, { "paper_id": "pq3RANvCZC", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Junze Yin", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Yeqi Gao", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "MS student" }, { "name": "Yuzhou Gu", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" } ], "n_formula": 44, "n_formula_1": 22, "n_ref_uni": 71, "n_ref": 113, "n_ref_all": 118, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4604, "formula_len_all_1": 2375, "len_all": 175662, "len_all_1": 80947, "len_abs": 1126, "len_title": 119, "len_sents": 36428, "len_sents_1": 31856, "n_sents": 387, "n_sents_1": 322, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1151, "title": "Binary Hypothesis Testing for Softmax Models and Leverage Score Models", "abs": "Softmax distributions are widely used in machine learning, including Large Language Models (LLMs) where the attention unit uses softmax distributions. We abstract the attention unit as the softmax model, where given a vector input, the model produces an output drawn from the softmax distribution (which depends on the vector input). We consider the fundamental problem of binary hypothesis testing in the setting of softmax models. That is, given an unknown softmax model, which is known to be one of the two given softmax models, how many queries are needed to determine which one is the truth? We show that the sample complexity is asymptotically $O(\\epsilon^{-2})$ where $\\epsilon$ is a certain distance between the parameters of the models.\n\nFurthermore, we draw analogy between the softmax model and the leverage score model, an important tool for algorithm design in linear algebra and graph theory. The leverage score model, on a high level, is a model which, given vector input, produces an output drawn from a distribution dependent on the input. We obtain similar results for the binary hypothesis testing problem for leverage score models.", "keywords": [ "Binary hypothesis testing", "softmax distributions", "large language models", "attention" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "pq1WUegkza", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Zikun Zhang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Zixiang Chen", "gender": "Male", "institution": " University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 125, "n_formula_1": 35, "n_ref_uni": 34, "n_ref": 133, "n_ref_all": 157, "n_fig": 1, "n_tab": 3, "L_tab": 363, "n_element_tab": 58, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15590, "formula_len_all_1": 3180, "len_all": 183877, "len_all_1": 71519, "len_abs": 1168, "len_title": 127, "len_sents": 47982, "len_sents_1": 31001, "n_sents": 409, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1177, "title": "Convergence of Score-Based Discrete Diffusion Models: A Discrete-Time Analysis", "abs": "Diffusion models have achieved great success in generating high-dimensional samples across various applications. While the theoretical guarantees for continuous-state diffusion models have been extensively studied, the convergence analysis of the discrete-state counterparts remains under-explored. In this paper, we study the theoretical aspects of score-based discrete diffusion models under the Continuous Time Markov Chain (CTMC) framework. We introduce a discrete-time sampling algorithm in the general state space $[S]^d$ that utilizes score estimators at predefined time points. We derive convergence bounds for the Kullback-Leibler (KL) divergence and total variation (TV) distance between the generated sample distribution and the data distribution, considering both scenarios with and without early stopping under reasonable assumptions. Notably, our KL divergence bounds are nearly linear in the dimension $d$, aligning with state-of-the-art results for diffusion models. Our convergence analysis employs a Girsanov-based method and establishes key properties of the discrete score function, which are essential for characterizing the discrete-time sampling process.", "keywords": [ "Discrete diffusion model", "Convergence analysis", "Time-discretization", "Discrete score function" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 4, 3, 4, 2 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "pppyig2kYe", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Daniel L. Pimentel-Alarc\u00f3n", "gender": "unknown", "institution": "University of Wisconsin, Madison", "country": "", "position": "Assistant Professor" }, { "name": "Karan Vikyath Veeranna Rupashree", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "MS student" }, { "name": "Siddharth Baskar", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Intern" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 42, "n_ref": 63, "n_ref_all": 85, "n_fig": 6, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 86, "formula_len_all_1": 86, "len_all": 115130, "len_all_1": 72016, "len_abs": 1357, "len_title": 78, "len_sents": 35503, "len_sents_1": 35509, "n_sents": 258, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Reject", "L_title": 30, "L_abs": 1343, "title": "Latent Matrix Completion Model", "abs": "Large amounts of missing data are becoming increasingly ubiquitous in modern high-dimensional datasets. High-rank matrix completion (HRMC) uses the powerful union of subspace (UoS) model to handle these vast amounts of missing data. However, existing HRMC methods often fail when dealing with real data that does not follow the UoS model exactly. Here we propose a new approach: instead of finding a UoS that fits the observed data directly, we will find a UoS in a latent space that can fit a non-linear embedding of the original data. Embeddings of this sort are typically attained with deep architectures. However, the abundance of missing data impedes the training process, as the coordinates of the observed samples rarely overlap. We overcome this difficulty with a novel pseudo-completion layer (in charge of estimating the missing values) followed by an auto-encoder (in charge of finding the embedding) coupled with a self-expressive layer (that clusters data according to a UoS in the latent space). Our design reduces the exponential memory requirements typically induced by uneven patterns of missing data. We describe our architecture, model, loss functions, and training strategy. Our experiments on several real datasets show that our method consistently outperforms the state-of-the-art accuracy by more than a staggering 40\\%.", "keywords": [ "Clustering", "Union of Subspace", "matrix completion", "Image reconstruction" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "ppeG3z5mcj", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Jiacheng Xu", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Kang Kang", "gender": "Male", "institution": "BeiJing Kunlun Tech co., LTD.", "country": "", "position": "Researcher" }, { "name": "Lang Feng", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Xiang Wen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhenghai Xue", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 52, "n_formula_1": 15, "n_ref_uni": 44, "n_ref": 133, "n_ref_all": 166, "n_fig": 9, "n_tab": 6, "L_tab": 2229, "n_element_tab": 259, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1356, "n_element_tab_1": 181, "formula_len_all": 6803, "formula_len_all_1": 2073, "len_all": 178350, "len_all_1": 80711, "len_abs": 1425, "len_title": 135, "len_sents": 53191, "len_sents_1": 33720, "n_sents": 453, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1432, "title": "ASOR: Anchor State Oriented Regularization for Policy Optimization under Dynamics Shift", "abs": "To train neural policies in environments with diverse dynamics, Imitation from\nObservation (IfO) approaches aim at recovering expert state trajectories. Their\nsuccess is built upon the assumption that the stationary state distributions induced\nby optimal policies remain similar despite dynamics shift. However, such an\nassumption does not hold in many real world scenarios, especially when certain\nstates become inaccessible during environment dynamics change. In this paper,\nwe propose the concept of anchor states which appear in all optimal trajectories\nunder dynamics shift, thereby maintaining consistent state accessibility. Instead of\ndirect imitation, we incorporate anchor state distributions into policy regularization\nto mitigate the issue of inaccessible states, leading to the ASOR algorithm. By\nformally characterizing the difference of state accessibility under dynamics shift,\nwe show that the anchor state-based regularization approach provides strong lower-\nbound performance guarantees for efficient policy optimization. We perform\nextensive experiments across various online and offline RL benchmarks, including\nGridworld, MuJoCo, MetaDrive, D4RL, and a fall-guys like game environment,\nfeaturing multiple sources of dynamics shift. Experimental results indicate ASOR\ncan be effectively integrated with several state-of-the-art cross-domain policy\ntransfer algorithms, substantially enhancing their performance.", "keywords": [ "Reinforcement Learning", "Policy Transfer", "Dynamics Shift", "Policy Regularization" ], "rating_list": [ 3, 3, 5, 8 ], "soundness_list": [ 2, 1, 2, 4 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 1, 3, 4 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "ppcVng610J", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Qiaomin Xie", "gender": "Female", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Robert D Nowak", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "", "position": "Full Professor" }, { "name": "Subhojyoti Mukherjee", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "", "position": "PhD student" }, { "name": "Josiah Hanna", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "Assistant Professor" } ], "n_formula": 47, "n_formula_1": 9, "n_ref_uni": 111, "n_ref": 201, "n_ref_all": 262, "n_fig": 19, "n_tab": 3, "L_tab": 649, "n_element_tab": 37, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 885, "n_element_tab_1": 35, "formula_len_all": 3036, "formula_len_all_1": 547, "len_all": 324000, "len_all_1": 65829, "len_abs": 5565, "len_title": 158, "len_sents": 94788, "len_sents_1": 30231, "n_sents": 790, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1935, "title": "Pretraining Decision Transformers with Reward Prediction for In-Context Structured Bandit Learning", "abs": "In this paper, we study the multi-task structured bandit problem where the goal is to learn a near-optimal algorithm that minimizes cumulative regret. The tasks share a common structure and the algorithm exploits the shared structure to minimize the cumulative regret for an unseen but related test task. We use a transformer as a decision-making algorithm to learn this shared structure so as to generalize to the test task. The prior work of pretrained decision transformers like \\dpt\\ requires access to the optimal action during training which may be hard in several scenarios. Diverging from these works, our learning algorithm does not need the knowledge of optimal action per task during training but predicts a reward vector for each of the actions using only the observed offline data from the diverse training tasks. Finally, during inference time, it selects action using the reward predictions employing various exploration strategies in-context for an unseen test task. We show that our model outperforms other SOTA methods like \\dpt, and Algorithmic Distillation (\\ad) over a series of experiments on several structured bandit problems (linear, bilinear, latent, non-linear). Interestingly, we show that our algorithm, without the knowledge of the underlying problem structure, can learn a near-optimal policy in-context by leveraging the shared structure across diverse tasks. We further extend the field of pre-trained decision transformers by showing that they can leverage unseen tasks with new actions and still learn the underlying latent structure to derive a near-optimal policy. We validate this over several experiments to show that our proposed solution is very general and has wide applications to potentially emergent online and offline strategies at test time. Finally, we theoretically analyze the performance of our algorithm and obtain generalization bounds in the in-context multi-task learning setting.", "keywords": [ "Structured bandit", "in-context learning", "Decision Transformers" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "powufeT93G", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Yi Yang", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Yixuan Tang", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 48, "n_ref": 119, "n_ref_all": 135, "n_fig": 1, "n_tab": 14, "L_tab": 4035, "n_element_tab": 371, "n_fig_1": 0, "n_tab_1": 6, "L_tab_1": 1276, "n_element_tab_1": 125, "formula_len_all": 166, "formula_len_all_1": 13, "len_all": 187064, "len_all_1": 60917, "len_abs": 1722, "len_title": 166, "len_sents": 43483, "len_sents_1": 28506, "n_sents": 348, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1731, "title": "Do We Need Domain-Specific Embedding Models? An Empirical Investigation", "abs": "Embedding models play a crucial role in representing and retrieving information across various NLP applications. Recent advancements in Large Language Models (LLMs) have further enhanced the performance of embedding models, which are trained on massive amounts of text covering almost every domain. These models are often benchmarked on general-purpose datasets like Massive Text Embedding Benchmark (MTEB), where they demonstrate superior performance. However, a critical question arises: Is the development of domain-specific embedding models necessary when general-purpose models are trained on vast corpora that already include specialized domain texts? In this paper, we empirically investigate this question, choosing the finance domain as an example. We introduce the Finance Massive Text Embedding Benchmark (FinMTEB), a counterpart to MTEB that consists of financial domain-specific text datasets. We evaluate the performance of seven state-of-the-art embedding models on FinMTEB and observe a significant performance drop compared to their performance on MTEB. To account for the possibility that this drop is driven by FinMTEB's higher complexity, we propose four measures to quantify dataset complexity and control for this factor in our analysis. Our analysis provides compelling evidence that state-of-the-art embedding models struggle to capture domain-specific linguistic and semantic patterns. Moreover, we find that the performance of general-purpose embedding models on MTEB is not correlated with their performance on FinMTEB, indicating the need for domain-specific embedding benchmarks for domain-specific embedding models. This study sheds light on developing domain-specific embedding models in the LLM era.", "keywords": [ "Domain Adaptation; Embedding Benchmark; Empirical Analysis" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "popKM1zAYa", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Aliaksandr Siarohin", "gender": "Male", "institution": "Snap Inc.", "country": "IT", "position": "Researcher" }, { "name": "Ivan Skorokhodov", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Jun-Yan Zhu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Kfir Aberman", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Sergey Tulyakov", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Director of Research" }, { "name": "Tsai-Shien Chen", "gender": "Male", "institution": "University of California, Merced", "country": "US", "position": "PhD student" }, { "name": "Willi Menapace", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Yuwei Fang", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Senior Research Scientist" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 64, "n_ref": 161, "n_ref_all": 196, "n_fig": 13, "n_tab": 5, "L_tab": 1056, "n_element_tab": 113, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1759, "n_element_tab_1": 175, "formula_len_all": 203, "formula_len_all_1": 0, "len_all": 173243, "len_all_1": 60272, "len_abs": 103, "len_title": 106, "len_sents": 43731, "len_sents_1": 26852, "n_sents": 330, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1673, "title": "VideoAlchemy: Open-set Personalization in Video Generation", "abs": "Video personalization methods allow us to synthesize videos with specific concepts such as people, pets, and places. However, existing methods often focus on limited domains, require time-consuming optimization per subject, or support only a single subject. We present $VideoAlchemy~-$ a video model equipped with built-in multi-subject, open-set personalization capabilities for both foreground objects and backgrounds, eliminating the need for time-consuming test-time optimization. Our model is built on a new Diffusion Transformer module that fuses each reference image conditioning and its corresponding subject-level text prompt with cross-attention layers. Developing such a large model presents two main challenges: $dataset$ and $evaluation$. First, as paired datasets of reference images and videos are extremely hard to collect, we opt to sample video frames as reference images and synthesize entire videos. This approach, however, introduces data biases issue, where models can easily denoise training videos but fail to generalize to new contexts during inference. To mitigate these issue, we carefully design a new automatic data construction pipeline with extensive image augmentation and sampling techniques. Second, evaluating open-set video personalization is a challenge in itself. To address this, we introduce a new personalization benchmark with evaluation protocols focusing on accurate subject fidelity assessment and accommodating different types of personalization conditioning. Finally, our extensive experiments show that our method significantly outperforms existing personalization methods, regarding quantitative and qualitative evaluations.", "keywords": [ "generative models", "video generation", "content personalization", "content customization" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 5, 3, 5, 3 ] }, { "paper_id": "poKOBfJrKf", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Lorenzo Bini", "gender": "Male", "institution": "University of Geneva", "country": "CH", "position": "PhD student" }, { "name": "Marco Sorbi", "gender": "Male", "institution": "University of Geneva", "country": "CH", "position": "PhD student" }, { "name": "Stephane Marchand-Maillet", "gender": "Not Specified", "institution": "University of Geneva, Switzerland", "country": "CH", "position": "Associate Professor" } ], "n_formula": 14, "n_formula_1": 2, "n_ref_uni": 30, "n_ref": 48, "n_ref_all": 79, "n_fig": 8, "n_tab": 3, "L_tab": 838, "n_element_tab": 59, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 507, "n_element_tab_1": 48, "formula_len_all": 594, "formula_len_all_1": 87, "len_all": 111779, "len_all_1": 54246, "len_abs": 1419, "len_title": 130, "len_sents": 37500, "len_sents_1": 27219, "n_sents": 271, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1423, "title": "Characterizing Massive Activations of Attention Mechanism in Graph Neural Networks", "abs": "Graph Neural Networks (GNNs) have become increasingly popular for effectively modeling data with graph structures. Recently, attention mechanisms have been integrated into GNNs to improve their ability to capture complex patterns. This paper presents the first comprehensive study revealing a critical, unexplored consequence of this integration: the emergence of Massive Activations (MAs) within attention layers. We introduce a novel method for detecting and analyzing MAs, focusing on edge features in different graph transformer architectures. Our study assesses various GNN models using benchmark datasets, including ZINC, TOX21, and PROTEINS. Key contributions include (1) establishing the direct link between attention mechanisms and MAs generation in GNNs, (2) developing a robust definition and detection method for MAs based on activation ratio distributions, (3) introducing the Explicit Bias Term (EBT) as a potential countermeasure and exploring it as an adversarial framework to assess models robustness based on the presence or absence of MAs. Our findings highlight the prevalence and impact of attention-induced MAs across different architectures, such as GraphTransformer, GraphiT, and SAN. The study reveals the complex interplay between attention mechanisms, model architecture, dataset characteristics, and MAs emergence, providing crucial insights for developing more robust and reliable graph models.", "keywords": [ "GraphML", "Transformers", "GraphNN", "Adversarial" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "po67tkP0Jx", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Christian Schroeder de Witt", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Lecturer" }, { "name": "David Krueger", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Assistant Professor" }, { "name": "Hannah Erlebach", "gender": "Female", "institution": "University College London, University of London", "country": "GB", "position": "MS student" }, { "name": "Karim Ahmed Abdel Sadek", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "MS student" }, { "name": "Matthew Farrugia-Roberts", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Michael D Dennis", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Usman Anwar", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 15, "n_ref_uni": 39, "n_ref": 61, "n_ref_all": 85, "n_fig": 8, "n_tab": 2, "L_tab": 1885, "n_element_tab": 26, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3500, "formula_len_all_1": 1835, "len_all": 152102, "len_all_1": 74895, "len_abs": 1274, "len_title": 100, "len_sents": 47693, "len_sents_1": 32326, "n_sents": 377, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1285, "title": "Mitigating Goal Misgeneralization via Minimax Regret", "abs": "Robustness research in reinforcement learning often focuses on ensuring that the policy consistently exhibits capable, goal-driven behavior. However, not every capable behavior is the intended behavior. *Goal misgeneralization* can occur when the policy generalizes capably with respect to a 'proxy goal' whose optimal behavior correlates with the intended goal on the training distribution, but not out of distribution. Though the intended goal would be ambiguous if they were perfectly correlated in training, we show progress can be made if the goals are only *nearly ambiguous*, with the training distribution containing a small proportion of *disambiguating* levels. We observe that the training signal from disambiguating levels could be amplified by regret-based prioritization. We formally show that approximately optimal policies on maximal-regret levels avoid the harmful effects of goal misgeneralization, which may exist without this prioritization. Empirically, we find that current regret-based Unsupervised Environment Design (UED) methods can mitigate the effects of goal misgeneralization, though do not always entirely eliminate it. Our theoretical and empirical results show that as UED methods improve they could further mitigate goal misgeneralization in practice.", "keywords": [ "Goal Misgeneralization", "Unsupervised Environment Design", "Reinforcement Learning", "AI Safety", "Alignment", "Autocurricula" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 5, 3, 2 ] }, { "paper_id": "pmznhtCHNb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chenxin Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Wuyang Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Yifan Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Yixuan Yuan", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Zhiwen Fan", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Hengyu Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zhiqin Brian Yang", "gender": "Male", "institution": "Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 35, "n_ref": 52, "n_ref_all": 59, "n_fig": 3, "n_tab": 4, "L_tab": 1184, "n_element_tab": 175, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 581, "n_element_tab_1": 77, "formula_len_all": 582, "formula_len_all_1": 582, "len_all": 125651, "len_all_1": 62897, "len_abs": 1342, "len_title": 128, "len_sents": 39771, "len_sents_1": 31842, "n_sents": 231, "n_sents_1": 169, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1393, "title": "GaussianClin: Multimodal Featured Gaussian Splatting for Dynamic Clinical Videos", "abs": "Reconstructing dynamic 3D models from clinical videos is crucial for medical applications such as surgical visualization, robot-assisted surgery, and medical training. However, the clinical environment presents unique challenges, including limited surface textures, inconsistent lighting, and the need for expert-level medical knowledge, making it difficult for non-experts to directly apply existing techniques. To address these challenges, we present **GaussianClin**, a novel approach that enhances 3D modeling capabilities in dynamic clinical videos by leveraging multimodal feature-based Gaussian splatting (GS). By embedding trained multimodal feature fields into the radiance field, **GaussianClin** integrates general medical knowledge and improves the performance of GS in tasks like 3D tissue visualization, real-time object enhancement, clinical instrument and organ segmentation, and medical visual question answering. To effectively capture temporal dynamics and tissue deformations, we further introduce a spatiotemporal graph distillation, which significantly improves handling deformable tissues compared to standard GS methods. Experimental results demonstrate that **GaussianClin** enables clinical 3D expert models to leverage massive pre-trained 2D multimodal foundation models, thereby paving the way for advancements in robot-assisted surgery and medical data processing.", "keywords": [ "3D Scene Reconstruction", "Robot-assisted Clinical Application" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 2, 3, 4 ] }, { "paper_id": "pmYpa7GpFH", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Enhong Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Hao Sun", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Nan Yin", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Xiaochun Cao", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Yifei Cheng", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Postdoc" } ], "n_formula": 108, "n_formula_1": 30, "n_ref_uni": 41, "n_ref": 86, "n_ref_all": 131, "n_fig": 0, "n_tab": 7, "L_tab": 2757, "n_element_tab": 342, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 2591, "n_element_tab_1": 84, "formula_len_all": 21210, "formula_len_all_1": 2960, "len_all": 157736, "len_all_1": 66631, "len_abs": 1271, "len_title": 105, "len_sents": 36205, "len_sents_1": 26302, "n_sents": 373, "n_sents_1": 246, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1276, "title": "LightSAM: Parameter-Agnostic Sharpness-Aware Minimization", "abs": "Sharpness-Aware Minimization (SAM) optimizer enhances the generalization ability of the machine learning model by exploring the flat minima landscape through weight perturbations. Despite its empirical success, SAM introduces an additional hyper-parameter, the perturbation radius, which causes the sensitivity of SAM to it. Moreover, it has been proved that the perturbation radius and learning rate of SAM are constrained by problem-dependent parameters to guarantee convergence. These limitations indicate the requirement of parameter-tuning in practical applications. In this paper, we propose the algorithm LightSAM which sets the perturbation radius and learning rate of SAM adaptively, thus extending the application scope of SAM. LightSAM employs three popular adaptive optimizers, including AdaGrad-Norm, AdaGrad and Adam, to replace the SGD optimizer for weight perturbation and model updating, reducing sensitivity to parameters. Theoretical results show that under weak assumptions, LightSAM could converge ideally with any choices of perturbation radius and learning rate, thus achieving parameter-agnostic. We conduct preliminary experiments on several deep learning tasks, which together with the theoretical findings validate the the effectiveness of LightSAM.", "keywords": [ "optimization", "parameter-agnostic", "sharpness-aware" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "plkrRJt98c", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Lucas Moeller", "gender": "Male", "institution": "University of Stuttgart, Universit\u00e4t Stuttgart", "country": "", "position": "PhD student" }, { "name": "Pascal Tilli", "gender": "Male", "institution": "University of Stuttgart, Universit\u00e4t Stuttgart", "country": "", "position": "PhD student" }, { "name": "Thang Vu", "gender": "Male", "institution": "University of Stuttgart, University of Stuttgart", "country": "", "position": "Full Professor" }, { "name": "Sebastian Pado", "gender": "Male", "institution": "University of Stuttgart, Universit\u00e4t Stuttgart", "country": "DE", "position": "Professor" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 89, "n_ref": 134, "n_ref_all": 177, "n_fig": 23, "n_tab": 7, "L_tab": 1608, "n_element_tab": 184, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 941, "n_element_tab_1": 121, "formula_len_all": 803, "formula_len_all_1": 695, "len_all": 229481, "len_all_1": 53888, "len_abs": 916, "len_title": 135, "len_sents": 42200, "len_sents_1": 23128, "n_sents": 324, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 852, "title": "Explaining Vision-Language Similarities in Dual Encoders with Feature-Pair Attributions", "abs": "Dual encoder architectures like CLIP models map two types of inputs into a shared embedding space and learn similarities between them.\n However, it is not understood how such models compare the two inputs.\n We first derive a method to attribute predictions of any differentiable dual encoder onto feature-pair interactions between its inputs. \n Second, we apply our method to CLIP models and show that they learn fine-grained correspondences between parts of captions and regions in images. They match objects across input modes and also account for mismatches. However, this visual-linguistic grounding ability heavily varies between object classes, depends on the training data distribution, and largely improves upon in-domain training.\n Using our method we can identify individual failure cases and knowledge gaps about specific object classes.", "keywords": [ "Explainability", "Attribution", "Dual Encoder", "Similarity", "Vision-Language", "CLIP" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "pljYMCYDWJ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Anton Xue", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Avishree Khare", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Eric Wong", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Assistant Professor" }, { "name": "Rajeev Alur", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Full Professor" }, { "name": "Surbhi Goel", "gender": "Female", "institution": "University of Pennsylvania", "country": "US", "position": "Assistant Professor" } ], "n_formula": 55, "n_formula_1": 12, "n_ref_uni": 45, "n_ref": 77, "n_ref_all": 134, "n_fig": 13, "n_tab": 4, "L_tab": 1047, "n_element_tab": 57, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 373, "n_element_tab_1": 52, "formula_len_all": 4681, "formula_len_all_1": 1144, "len_all": 218588, "len_all_1": 74650, "len_abs": 773, "len_title": 126, "len_sents": 72541, "len_sents_1": 33743, "n_sents": 724, "n_sents_1": 287, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 803, "title": "Logicbreaks: A Framework for Understanding Subversion of Rule-based Inference", "abs": "We study how to subvert large language models (LLMs) from following prompt-specified rules.\nWe first formalize rule-following as inference in propositional Horn logic, a mathematical system in which rules have the form \"if $P$ and $Q$, then $R$\" for some propositions $P$, $Q$, and $R$.\nNext, we prove that although small transformers can faithfully follow such rules, maliciously crafted prompts can still mislead both theoretical constructions and models learned from data.\nFurthermore, we demonstrate that popular attack algorithms on LLMs find adversarial prompts and induce attention patterns that align with our theory.\nOur novel logic-based framework provides a foundation for studying LLMs in rule-based settings, enabling a formal analysis of tasks like logical reasoning and jailbreak attacks.", "keywords": [ "Logic", "reasoning", "inference", "language model", "jailbreak" ], "rating_list": [ 6, 5, 6, 8, 6 ], "soundness_list": [ 3, 2, 4, 4, 3 ], "presentation_list": [ 4, 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 2, 3, 3 ] }, { "paper_id": "plgLA2YBLH", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Haoran Xie", "gender": "Male", "institution": "Lingnan University", "country": "HK", "position": "Full Professor" }, { "name": "Jing Li", "gender": "Female", "institution": "The Hong Kong Polytechnic University", "country": "", "position": "Assistant Professor" }, { "name": "Qing Li", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Xianming LI", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "CN", "position": "PhD student" }, { "name": "Zongxi Li", "gender": "Male", "institution": "Lingnan University", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 39, "n_ref": 78, "n_ref_all": 95, "n_fig": 8, "n_tab": 4, "L_tab": 1233, "n_element_tab": 180, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1252, "n_element_tab_1": 188, "formula_len_all": 559, "formula_len_all_1": 559, "len_all": 119134, "len_all_1": 55859, "len_abs": 1028, "len_title": 33, "len_sents": 29978, "len_sents_1": 24423, "n_sents": 234, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 33, "L_abs": 1108, "title": "ESE: Espresso Sentence Embeddings", "abs": "High-quality sentence embeddings are fundamental in many natural language processing (NLP) tasks, such as semantic textual similarity (STS) and retrieval-augmented generation (RAG). However, most existing methods leverage fixed-length sentence embeddings from full-layer language models, which lack the scalability to accommodate the diverse available resources across various applications. Viewing this gap, we propose a novel sentence embedding model Espresso Sentence Embeddings (ESE) with two learning processes. First, the learn-to-express process encodes more salient representations to shallow layers. Second, the learn-to-compress process compacts essential features into the initial dimensions using Principal Component Analysis (PCA). This way, ESE can scale model depth via the former process and embedding size via the latter. Extensive experiments on STS and RAG suggest that ESE can effectively produce high-quality sentence embeddings with less model depth and embedding size, enhancing inference efficiency. The code is available at https://github.com/SeanLee97/AnglE/blob/main/README_ESE.md.", "keywords": [ "sentence embeddings", "semantic textual similarity", "information retrieval", "retrieval-augmented generation" ], "rating_list": [ 6, 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 5, 4, 4 ] }, { "paper_id": "plflYGf23L", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Binhang Qi", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Hailong Sun", "gender": "Male", "institution": "School of Software, Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Ruobing Zhao", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Wenrui Long", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Xiang Gao", "gender": "Male", "institution": "Beihang University", "country": "", "position": "Associate Professor" }, { "name": "zongzhen yang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 44, "n_ref": 64, "n_ref_all": 107, "n_fig": 8, "n_tab": 10, "L_tab": 4096, "n_element_tab": 304, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1957, "n_element_tab_1": 89, "formula_len_all": 263, "formula_len_all_1": 150, "len_all": 187144, "len_all_1": 64683, "len_abs": 1714, "len_title": 125, "len_sents": 62654, "len_sents_1": 31272, "n_sents": 465, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1792, "title": "CABS: Conflict-Aware and Balanced Sparsification for Enhancing Model Merging", "abs": "Model merging based on task vectors, i.e., the parameter differences between fine-tuned models and a shared base model, provides an efficient way to integrate multiple models without retraining. This approach can be used to combine task-specific models into a multitask model, improve generalization, or address model deficiencies. One of the significant challenges faced by model merging is the conflicts between task vectors. Existing works aim to mitigate these conflicts through sparsification; however, two issues observed in our experiments significantly limit their performance: $\\textit{high parameter overlap}$ and $\\textit{unbalanced weight distribution}$. To address these issues, we propose a simple yet effective framework called CABS (Conflict-Aware and Balanced Sparsification), consisting of $\\textbf{C}$onflict-$\\textbf{A}$ware Sparsification (CA) and $\\textbf{B}$alanced $\\textbf{S}$parsification (BS). CA can reduce parameter overlap by applying masks during sequential pruning, ensuring that each task vector retains distinct, non-overlapping parameters. BS leverages $n$:$m$ pruning to preserve critical weights while maintaining an even distribution across layers. Our comprehensive experiments demonstrate that CABS outperforms state-of-the-art methods across a range of diverse tasks and model sizes. Notably, in experiments with 7B-parameter language models, CABS surpasses the average performance of an \"ideal\" model, a virtual model that selects the highest score from individual fine-tuned models for each task (CABS: 76.50 vs. Ideal Model: 76.30 vs. Baseline: 76.02 vs. Fine-tuned Model: 75.86). Our results highlight the importance of addressing both high parameter overlap and unbalanced weight distribution to achieve robust and high-performance model merging.", "keywords": [ "Model Merging", "Pruning Technique", "Task Vectors", "Language Models", "Conflict-Aware Sparsity (CA)", "Balanced Sparsity (BS)" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "plAiJUFNja", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Han Quan", "gender": "unknown", "institution": "China Agricultural University", "country": "CN", "position": "MS student" }, { "name": "Lang Zheng", "gender": "Male", "institution": "China Agricultural University", "country": "CN", "position": "MS student" }, { "name": "Ruxue Xing", "gender": "Female", "institution": "China Agricultural University", "country": "CN", "position": "MS student" }, { "name": "Zhenghan Chen", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "youhuan yang", "gender": "Male", "institution": "Zhengzhou University", "country": "CN", "position": "Researcher" } ], "n_formula": 31, "n_formula_1": 27, "n_ref_uni": 7, "n_ref": 7, "n_ref_all": 11, "n_fig": 0, "n_tab": 2, "L_tab": 546, "n_element_tab": 40, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 503, "n_element_tab_1": 36, "formula_len_all": 2620, "formula_len_all_1": 2071, "len_all": 81781, "len_all_1": 54581, "len_abs": 1814, "len_title": 141, "len_sents": 25438, "len_sents_1": 24399, "n_sents": 196, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1753, "title": "Graph-Enhanced Learning for Predicting Optimal Drug Combinations Using Contrastive Embedding", "abs": "We present a groundbreaking unified theory for drug-drug interaction (DDI) aware domain adaptation (DA) in the context of drug synergy prediction. Our framework seamlessly integrates concepts from optimal transport, information geometry, and quantum information theory within the setting of abstract Banach spaces. We introduce a novel DDI-aware optimal transport problem, formulated as a geodesic equation on an infinite-dimensional Finsler manifold that encodes both DDI structure and optimal transport costs. This geometric formulation provides a unified perspective on DDI-aware domain adaptation, interpreting the process as the evolution of a transport map along a geodesic in a space that captures both domain discrepancy and drug interaction patterns. Our approach extends to a stochastic gradient flow on the space of probability measures, combining ideas from information geometry and stochastic analysis. We prove the existence of a unique invariant measure for this flow and establish its convergence properties using techniques from infinite-dimensional Markov processes and \u0393-convergence. Our comprehensive mathematical framework not only unifies existing approaches to domain adaptation and DDI prediction but also opens new avenues for research at the intersection of these fields. By bridging the gap between abstract mathematical theories and practical drug synergy prediction, our work paves the way for more effective and theoretically grounded algorithms in drug discovery and personalized medicine. The proposed unified theory has far-reaching implications, potentially revolutionizing our understanding of cross-domain adaptation in complex biochemical systems and inspiring novel computational methods in pharmaceutical research.", "keywords": [ "Graph Learning", "Contrastive Embedding", "DDI" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 1, 1, 1 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 2, 2, 3 ] }, { "paper_id": "pl8OJhyArC", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Xingyi Yang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Xinchao Wang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 12, "n_ref_uni": 31, "n_ref": 52, "n_ref_all": 77, "n_fig": 12, "n_tab": 7, "L_tab": 4421, "n_element_tab": 189, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2376, "n_element_tab_1": 192, "formula_len_all": 921, "formula_len_all_1": 643, "len_all": 159395, "len_all_1": 58713, "len_abs": 158, "len_title": 100, "len_sents": 39376, "len_sents_1": 23060, "n_sents": 353, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 1276, "title": "Hash3D: Training-free Acceleration for 3D Generation", "abs": "The quality of 3D generative modeling has been notably improved by the adoption of 2D diffusion models. Despite this progress, \nthe cumbersome optimization process \\emph{per se}\npresents a critical problem to efficiency. \nIn this paper,\nwe introduce Hash3D, a universal acceleration for 3D score distillation sampling~(SDS) without model training.\nCentral to Hash3D is the observation that images rendered from similar camera positions and diffusion time-steps often have redundant feature maps. By hashing and reusing these feature maps across nearby timesteps and camera angles, Hash3D eliminates unnecessary calculations. We implement this through an adaptive grid-based hashing. As a result, it largely speeds up the process of 3D generation. \nSurprisingly, this feature-sharing mechanism not only makes generation faster but also improves the smoothness and view consistency of the synthesized 3D objects. Our experiments covering 5 text-to-3D and 3 image-to-3D models, demonstrate Hash3D\u2019s versatility to speed up optimization, enhancing efficiency by $1.5\\sim 4\\times$. Additionally, Hash3D's integration with 3D Gaussian splatting largely speeds up 3D model creation, reducing text-to-3D processing to about 10 minutes and image-to-3D conversion to roughly 30 seconds.", "keywords": [ "Efficient 3D generation; Score distillation" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 1, 4, 5 ] }, { "paper_id": "pl2c1PoiGO", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Andreas Roth", "gender": "Male", "institution": "Technische Universit\u00e4t Dortmund", "country": "DE", "position": "PhD student" }, { "name": "Thomas Liebig", "gender": "Male", "institution": "TU Dortmund", "country": "", "position": "Assistant Professor" } ], "n_formula": 29, "n_formula_1": 29, "n_ref_uni": 43, "n_ref": 89, "n_ref_all": 97, "n_fig": 3, "n_tab": 4, "L_tab": 204, "n_element_tab": 23, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 157, "n_element_tab_1": 22, "formula_len_all": 1769, "formula_len_all_1": 1769, "len_all": 120080, "len_all_1": 64092, "len_abs": 1030, "len_title": 80, "len_sents": 29938, "len_sents_1": 27069, "n_sents": 262, "n_sents_1": 234, "venue": "ICLR2025", "decision": "withdraw", "L_title": 32, "L_abs": 1033, "title": "Multi-Channel Graph Convolutions", "abs": "Defining the convolution on graphs has led to much progress in graph machine learning, particularly through approximations based on polynomials and, ultimately, message-passing neural networks (MPNNs). However, this convolution is defined for single-channel graph signals, i.e., a single feature is given at each node, and a single new feature is assigned to each node. As multiple initial node features are provided for many challenging tasks and convolutions are generally defined for these multi-channel signals, we introduce multi-channel graph convolutions (MCGCs) by obtaining their form using the graph Fourier transform. MCGCs highlight the critical importance of utilizing multiple edge relations to amplify different signals for each feature channel. We further introduce localized multi-channel MPNNs and the multi-channel graph isomorphism network (MC-GINs), with which we can provably obtain linear mappings that are injective on multisets. Our experiments confirm the greatly improved capabilities of MCGCs and MC-GINs.", "keywords": [ "graph convolution", "spectral graph convolution", "message-passing neural network" ], "rating_list": [ 3, 6, 3, 3, 3 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 1, 2, 3, 2, 2 ], "contribution_list": [ 1, 3, 1, 2, 2 ], "confidence_list": [ 4, 4, 5, 3, 4 ] }, { "paper_id": "pk4YjZeevI", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Barry-John Theobald", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Natalie Mackraz", "gender": "Female", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "St\u00e9phane Aroca-Ouellette", "gender": "Male", "institution": "University of Colorado at Boulder", "country": "US", "position": "PhD student" }, { "name": "Rin Metcalf", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 1, "n_ref_uni": 29, "n_ref": 48, "n_ref_all": 96, "n_fig": 13, "n_tab": 9, "L_tab": 4527, "n_element_tab": 410, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 1631, "n_element_tab_1": 162, "formula_len_all": 439, "formula_len_all_1": 4, "len_all": 188198, "len_all_1": 74723, "len_abs": 930, "len_title": 152, "len_sents": 63334, "len_sents_1": 36073, "n_sents": 505, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 935, "title": "PREDICT: Preference Reasoning by Evaluating Decomposed preferences Inferred from Candidate Trajectories", "abs": "Accommodating human preferences is essential for creating AI agents that deliver personalized and effective interactions. Recent work has shown the potential for LLMs to infer preferences from user interactions, but they often produce broad and generic preferences, failing to capture the unique and individualized nature of human preferences. This paper introduces PREDICT, a method designed to enhance the precision and adaptability of inferring preferences. PREDICT incorporates three key elements: (1) iterative refinement of inferred preferences, (2) decomposition of preferences into constituent components, and (3) validation of preferences across multiple trajectories. We evaluate PREDICT on two distinct environments: a gridworld setting and a new text-domain environment (PLUME). PREDICT more accurately infers nuanced human preferences improving over existing baselines by 66.2\\% (gridworld environment) and 41.0\\% (PLUME).", "keywords": [ "personalization", "preference learning", "LLM personalization", "personalization benchmark" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "pjfrGVekwK", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Alexander Tschantz", "gender": "Male", "institution": "University of Sussex", "country": "", "position": "PhD student" }, { "name": "Christopher Buckley", "gender": "Male", "institution": "University of Sussex", "country": "", "position": "Full Professor" }, { "name": "Ozan Catal", "gender": "Male", "institution": "Verses", "country": "", "position": "Research Engineer" }, { "name": "Tim Verbelen", "gender": "Male", "institution": "VERSES", "country": "", "position": "Researcher" }, { "name": "Toon Van de Maele", "gender": "Male", "institution": "VERSES", "country": "US", "position": "Researcher" } ], "n_formula": 29, "n_formula_1": 18, "n_ref_uni": 27, "n_ref": 56, "n_ref_all": 92, "n_fig": 16, "n_tab": 2, "L_tab": 909, "n_element_tab": 172, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 434, "n_element_tab_1": 80, "formula_len_all": 2299, "formula_len_all_1": 1358, "len_all": 126559, "len_all_1": 58122, "len_abs": 972, "len_title": 84, "len_sents": 43328, "len_sents_1": 26803, "n_sents": 340, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 975, "title": "Variational Bayes Gaussian Splatting", "abs": "Recently, 3D Gaussian Splatting has emerged as a promising approach for modeling 3D scenes using mixtures of Gaussians. The predominant optimization method for these models relies on backpropagating gradients through a differentiable rendering pipeline, which struggles with catastrophic forgetting when dealing with continuous streams of data. To address this limitation, we propose Variational Bayes Gaussian Splatting (VBGS), a novel approach that frames training a Gaussian splat as variational inference over model parameters. By leveraging the conjugacy properties of multivariate Gaussians, we derive a closed-form variational update rule, allowing efficient updates from partial, sequential observations without the need for replay buffers. Our experiments show that VBGS not only matches state-of-the-art performance on static datasets, but also enables continual learning from sequentially streamed 2D and 3D data, drastically improving performance in this setting.", "keywords": [ "Variational Bayes", "Gaussian Splatting", "Continual Learning" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "pjNjlJN7up", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "He Guo", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Jifeng Dai", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yulong Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Yuwen Xiong", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Zixuan Ye", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 36, "n_ref": 64, "n_ref_all": 74, "n_fig": 3, "n_tab": 3, "L_tab": 279, "n_element_tab": 100, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 279, "n_element_tab_1": 100, "formula_len_all": 216, "formula_len_all_1": 216, "len_all": 108494, "len_all_1": 60930, "len_abs": 1641, "len_title": 110, "len_sents": 29220, "len_sents_1": 28982, "n_sents": 210, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1647, "title": "big.LITTLE Vision Transformer for Efficient Visual Recognition", "abs": "In this paper, we introduce the big.LITTLE Vision Transformer, an innovative architecture aimed at achieving efficient visual recognition. This dual-transformer system is composed of two distinct blocks: the big performance block, characterized by its high capacity and substantial computational demands, and the LITTLE efficiency block, designed for speed with lower capacity. The key innovation of our approach lies in its dynamic inference mechanism. When processing an image, our system determines the importance of each token and allocates them accordingly: essential tokens are processed by the high-performance big model, while less critical tokens are handled by the more efficient little model. This selective processing significantly reduces computational load without sacrificing the overall performance of the model, as it ensures that detailed analysis is reserved for the most important information. To validate the effectiveness of our big.LITTLE Vision Transformer, we conducted comprehensive experiments on image classification and segment anything task. Our results demonstrate that the big.LITTLE architecture not only maintains high accuracy but also achieves substantial computational savings. Specifically, our approach enables the efficient handling of large-scale visual recognition tasks by dynamically balancing the trade-offs between performance and efficiency. The success of our method underscores the potential of hybrid models in optimizing both computation and performance in visual recognition tasks, paving the way for more practical and scalable deployment of advanced neural networks in real-world applications.", "keywords": [ "Efficient Vision Transformer" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "pjKdWj5NSR", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jo\u00ebl Tang", "gender": "unknown", "institution": "BNP Paribas", "country": "", "position": "Researcher" }, { "name": "Laurent Lam", "gender": "unknown", "institution": "BNP Paribas", "country": "FR", "position": "Data Scientist" }, { "name": "Quentin Bourbon", "gender": "Male", "institution": "T\u00e9l\u00e9com ParisTech", "country": "FR", "position": "MS student" }, { "name": "Seif Edinne LAATIRI", "gender": "Male", "institution": "BNP Paribas", "country": "FR", "position": "Researcher" }, { "name": "Timoth\u00e9e Blondiaux", "gender": "unknown", "institution": "BNP Paribas", "country": "FR", "position": "Independent Researcher" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 21, "n_ref": 32, "n_ref_all": 41, "n_fig": 6, "n_tab": 6, "L_tab": 2648, "n_element_tab": 327, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1879, "n_element_tab_1": 288, "formula_len_all": 918, "formula_len_all_1": 754, "len_all": 95389, "len_all_1": 56313, "len_abs": 1607, "len_title": 137, "len_sents": 30420, "len_sents_1": 23280, "n_sents": 237, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1615, "title": "ReFOCUS: Recurrent False Object Correction Using guidance Strategies in Object Detection", "abs": "This work addresses the issue of recurrent false positive classification in object detection. We consider two experimental setups imitating real-world scenarios that lead to such errors: i) erroneous annotations, ii) non-objects that resemble actual objects. We show that resulting models can be corrected efficiently using a two-step protocol that leverages false positive annotations. For the first step, we present and compare two correction approaches that guide false positives toward true negatives, in either the latent or the logit space. The second step then consists in standard continuous fine-tuning on correct annotations. The latent guidance framework relies on a decoder that maps the bounding box of a given false positive to its target true negative embedding. The decoder is trained as part of an autoencoder, where appropriate true negative samples are generated by a learnable Gaussian mixture model in the latent space. By leveraging the properties of the Wasserstein distance, the mixture model is optimized through standard backpropagation. In both experimental setups, the two correction methods significantly outperform standard continuous fine-tuning on correct annotations and demonstrate competitive performance when compared to models retrained from scratch on correct annotations. In particular, in the second experimental setup, the latent guidance framework consistently outperforms these models, effectively enhancing detection performance at the cost of supplementary false positive annotations. Additionally, the proposed techniques prove effective in a few-shot learning context.", "keywords": [ "Object Detection", "False Positive", "Computer Vision", "Recurrent Errors", "Correction" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "pjJIimQdfU", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Haoyang Li", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Postdoc" }, { "name": "Linxin Xiao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Wenwu Zhu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xin Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Zeyang Zhang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zongyuan Wu", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "\u5434\u8000\u83f2", "gender": "Female", "institution": "Beijing University of Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 19, "n_formula_1": 13, "n_ref_uni": 56, "n_ref": 119, "n_ref_all": 139, "n_fig": 4, "n_tab": 5, "L_tab": 2804, "n_element_tab": 165, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 988, "n_element_tab_1": 56, "formula_len_all": 934, "formula_len_all_1": 447, "len_all": 195530, "len_all_1": 64388, "len_abs": 1597, "len_title": 119, "len_sents": 52496, "len_sents_1": 30071, "n_sents": 377, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1603, "title": "Self-supervised Masked Graph Autoencoder via Structure-aware Curriculum", "abs": "Self-supervised learning (SSL) on graph-structured data has attracted considerable attention recently. Masked graph autoencoder, as one promising generative graph SSL approach that aims to recover masked parts of the input graph data, has shown great success on various downstream graph tasks. However, existing masked graph autoencoders fail to consider the degrees of difficulties of recovering the masked edges that often have different impacts on the model performance, resulting in suboptimal node representations. To tackle this challenge, in this paper, we propose a novel curriculum based self-supervised masked graph autoencoder that is able to capture and leverage the underlying degree of difficulties of data dependencies hidden in edges, and design better mask-reconstruction pretext tasks for learning informative node representations. Specifically, we first design a difficulty measurer to identify the underlying structural degree of difficulties of edges during the masking step. Then, we adopt a self-paced scheduler to determine the order of masking edges, which encourages the graph encoder to learn from easy parts to difficult parts. Finally, the masked edges are gradually incorporated into the reconstruction pretext task, leading to high-quality node representations. Experiments on several real-world node classification and link prediction datasets demonstrate the superiority of our proposed method over state-of-the-art graph self-supervised learning baselines. This work is the first study of curriculum strategy for masked graph autoencoders, to the best of our knowledge.", "keywords": [ "Curriculum Learning", "Self-supervised Learning", "Graph Neural Network" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "piRU8xOurs", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "John Palowitch", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Karthikeyan Shanmugam", "gender": "Male", "institution": "Google Deepmind", "country": "IN", "position": "Researcher" }, { "name": "Manish Jain", "gender": "Male", "institution": "Google", "country": "IN", "position": "Researcher" }, { "name": "Rahul Madhavan", "gender": "Male", "institution": "Indian Institute of Management, Ahmedabad", "country": "IN", "position": "MS student" }, { "name": "JASHN ARORA", "gender": "Male", "institution": "Google", "country": "IN", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 97, "n_ref": 173, "n_ref_all": 273, "n_fig": 16, "n_tab": 12, "L_tab": 4288, "n_element_tab": 401, "n_fig_1": 10, "n_tab_1": 5, "L_tab_1": 778, "n_element_tab_1": 120, "formula_len_all": 27, "formula_len_all_1": 0, "len_all": 287285, "len_all_1": 60397, "len_abs": 951, "len_title": 123, "len_sents": 70109, "len_sents_1": 27897, "n_sents": 530, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1251, "title": "TreeTop: Topology-Aware Fine-Tuning for LLM Conversation Tree Understanding", "abs": "While Large Language Models (LLMs) have dominated a wide diversity of natural language tasks, improving their capabilities on \\emph{structured} inputs such as graphs remains an open challenge. We introduce $\\texttt{TreeTop}$, a pre-training framework for LLMs that significantly improves their ability to understand and reason over structural relationships in multi-party, threaded discussions, such as those found on social media platforms. $\\texttt{TreeTop}$ is a novel set of 17 QA-style tasks specifically designed to allow LLMs to selectively focus on both the structure of and content in discussion graphs. We find that LLMs fine-tuned with $\\texttt{TreeTop}$ outperform their counterparts in every setting: zero-shot/few-shot performance on unseen pretraining tasks as well as downstream social media inference tasks (e.g.rumor detection), as well as fine-tuned performance on the downstream tasks, including their challenging \"early-detection\" variants. In particular, $\\texttt{Gemini Pro}$ fine-tuned with $\\texttt{TreeTop}$ and further fine-tuned on downstream tasks surpasses both vanilla $\\texttt{Gemini Pro}$ and state-of-the-art GNN baselines. Our framework paves the way for LLMs with enhanced capabilities on heavily-structured inputs.", "keywords": [ "Conversation Trees", "Social media", "Large language models" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "phWflQbLhu", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Chi ZHANG", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Qihang Zhao", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Weixiong Zhang", "gender": "unknown", "institution": "Washington University, Saint Louis", "country": "", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 65, "n_ref_all": 86, "n_fig": 2, "n_tab": 4, "L_tab": 1411, "n_element_tab": 65, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 700, "n_element_tab_1": 9, "formula_len_all": 151, "formula_len_all_1": 151, "len_all": 195814, "len_all_1": 61460, "len_abs": 1570, "len_title": 116, "len_sents": 55789, "len_sents_1": 32619, "n_sents": 353, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1577, "title": "dnaGrinder: a lightweight and high-capacity genomic foundation model", "abs": "The task of understanding and interpreting the complex information encoded within genomic sequences remains a grand challenge in biological research and clinical applications. In this context, recent advancements in large language model research have led to the development of both encoder-only and decoder-only foundation models designed to decode intricate information in DNA sequences. However, several issues persist, particularly regarding the efficient management of long-range dependencies inherent in genomic sequences, the effective representation of nucleotide variations, and the considerable computational costs associated with large model architectures and extensive pretraining datasets. Current genomic foundation models often face a critical tradeoff: smaller models with mediocre performance versus larger models with improved performance. To address these challenges, we introduce dnaGrinder, a unique and efficient genomic foundation model. dnaGrinder excels at managing long-range dependencies within genomic sequences while minimizing computational costs without compromising performance. It achieves results that are not just comparable but often superior to leading DNA models such as Nucleotide Transformer and DNABERT-2. Furthermore, dnaGrinder is designed for easy fine-tuning on workstation-grade GPUs, accommodating input lengths exceeding 17,000 tokens. On a single high-performance GPU, it supports sequences longer than 140,000 tokens, making it a highly efficient and accessible tool for both basic biological research and clinical applications.", "keywords": [ "DNA", "Genome", "Language Model", "Foundation Model" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "phAlw3JPms", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Baoxiang Wang", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Feng Luo", "gender": "Female", "institution": "Rice University", "country": "US", "position": "MS student" }, { "name": "Jiawei Xu", "gender": "Male", "institution": "CUHK(SZ)", "country": "", "position": "PhD student" }, { "name": "Lei Han", "gender": "Male", "institution": "Tencent Robotics X", "country": "CN", "position": "Principal Researcher" }, { "name": "Meng Fang", "gender": "Male", "institution": "University of Liverpool", "country": "", "position": "Assistant Professor" }, { "name": "Rui Yang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Shuang Qiu", "gender": "Male", "institution": "City University of Hong Kong", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 19, "n_formula_1": 17, "n_ref_uni": 51, "n_ref": 116, "n_ref_all": 146, "n_fig": 16, "n_tab": 12, "L_tab": 9664, "n_element_tab": 1063, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 2949, "n_element_tab_1": 347, "formula_len_all": 1499, "formula_len_all_1": 1518, "len_all": 212076, "len_all_1": 69604, "len_abs": 1674, "len_title": 129, "len_sents": 57588, "len_sents_1": 28463, "n_sents": 465, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1696, "title": "Tackling Data Corruption in Offline Reinforcement Learning via Sequence Modeling", "abs": "Learning policy from offline datasets through offline reinforcement learning (RL) holds promise for scaling data-driven decision-making while avoiding unsafe and costly online interactions. However, real-world data collected from sensors or humans often contains noise and errors, posing a significant challenge for existing offline RL methods, particularly when the real-world data is limited. Our study reveals that prior research focusing on adapting predominant offline RL methods based on temporal difference learning still falls short under data corruption when the dataset is limited. In contrast, we discover that vanilla sequence modeling methods, such as Decision Transformer, exhibit robustness against data corruption, even without specialized modifications. To unlock the full potential of sequence modeling, we propose **R**obust **D**ecision **T**ransformer (**RDT**) by incorporating three simple yet effective robust techniques: embedding dropout to improve the model's robustness against erroneous inputs, Gaussian weighted learning to mitigate the effects of corrupted labels, and iterative data correction to eliminate corrupted data from the source. \nExtensive experiments on MoJoCo, Kitchen, and Adroit tasks demonstrate RDT's superior performance under various data corruption scenarios compared to prior methods. Furthermore, RDT exhibits remarkable robustness in a more challenging setting that combines training-time data corruption with test-time observation perturbations. These results highlight the potential of sequence modeling for learning from noisy or corrupted offline datasets, thereby promoting the reliable application of offline RL in real-world scenarios.", "keywords": [ "Offline Reinforcement Learning", "Data Corruption", "Robust Reinforcement Learning" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "pgVMJdhgPI", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Jianfeng Qiu", "gender": "Male", "institution": "Anhui University", "country": "CN", "position": "Assistant Professor" }, { "name": "Li junfa", "gender": "Male", "institution": "Anhui University", "country": "CN", "position": "MS student" }, { "name": "Ke Xu", "gender": "Male", "institution": "Anhui University", "country": "CN", "position": "Lecturer" }, { "name": "Shanshan Wang", "gender": "Female", "institution": "Anhui University", "country": "CN", "position": "Associate Professor" }, { "name": "Yanning Hou", "gender": "Male", "institution": "Anhui University", "country": "CN", "position": "Undergrad student" }, { "name": "Yanran Ruan", "gender": "Female", "institution": "Anhui University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 25, "n_ref": 88, "n_ref_all": 101, "n_fig": 2, "n_tab": 6, "L_tab": 1582, "n_element_tab": 139, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 2239, "n_element_tab_1": 94, "formula_len_all": 610, "formula_len_all_1": 553, "len_all": 124586, "len_all_1": 59017, "len_abs": 2023, "len_title": 122, "len_sents": 37158, "len_sents_1": 26697, "n_sents": 281, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1421, "title": "FuzzyCLIP: Clustering-Driven Stacked Prompt in Zero-Shot Anomaly Detection", "abs": "How to enhance the alignment of text and image features in CLIP model is a key challenge in zero-shot industrial anomaly detection tasks. Recent studies mostly rely on precise category prompts for pre-training, but this approach is prone to overfitting, which limits the generalization ability of the mode. To address this issue, we propose the concept of fuzzy prompts and introduce Clustering-Driven Stacked Prompts (CSP) along with the Ensemble Feature Alignment (EFA) module to improve the alignment between text and image features. This design significantly outperforms other methods in terms of training speed, stability, and final convergence results, showing remarkable efficiency in enhancing anomaly detection segmentation performance. What is even more surprising is that fuzzy stacked prompts exhibit strong generalization in classification tasks, enabling them to adapt to various anomaly classification tasks without any additional operations. Therefore, we further propose the Regulating Prompt Learning (RPL) module, which leverages the strong generalization ability of fuzzy stacked prompts to regularize prompt learning, thereby improving performance in anomaly detection classification tasks. We conducted extensive experiments on seven industrial anomaly detection datasets, which demonstrate that our method achieves state-of-the-art performance in zero-shot anomaly detection and segmentation tasks.", "keywords": [ "Anomaly detection", "Zero-shot anomaly detection", "CLIP", "Industrial defect inspection" ], "rating_list": [ 5, 8, 3, 5 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "pflsJ6V6CL", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Artyom Iudin", "gender": "Male", "institution": "Moscow Technical University of Informatics and Communication", "country": "RU", "position": "Undergrad student" }, { "name": "Dmitrii Korzh", "gender": "unknown", "institution": "AIRI", "country": "RU", "position": "Researcher" }, { "name": "Elvir Karimov", "gender": "unknown", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "MS student" }, { "name": "Ivan Oseledets", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "Matvey Skripkin", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Researcher" }, { "name": "Nikita Kuzmin", "gender": "Male", "institution": "Moscow Technical University of Informatics and Communication", "country": "RU", "position": "Undergrad student" }, { "name": "Oleg Y. Rogov", "gender": "Male", "institution": "Moscow Technical University of Informatics and Communication", "country": "RU", "position": "Principal Researcher" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 27, "n_ref": 39, "n_ref_all": 45, "n_fig": 2, "n_tab": 6, "L_tab": 583, "n_element_tab": 90, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 808, "n_element_tab_1": 18, "formula_len_all": 302, "formula_len_all_1": 144, "len_all": 117439, "len_all_1": 59822, "len_abs": 988, "len_title": 142, "len_sents": 34176, "len_sents_1": 28773, "n_sents": 247, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1014, "title": "Listening to Formulas: Pioneering Models and Datasets for Converting Speech to LaTeX Equations", "abs": "Recognizing spoken mathematical expressions is a challenging task that involves transcribing speech into a strictly structured symbolic representation while addressing the ambiguity inherent in the pronunciation of equations. Although significant progress has been achieved in both automatic speech recognition (ASR) and language models (LM), the specific problem of translating spoken formulas into LaTeX has received relatively little attention. This task is particularly important in educational and research domains, for example, for lecture transcription. To address this issue, in this paper, we present a pioneering study on Speech-to-LaTeX conversion, introducing a novel, diverse human-uttered dataset in English and Russian comprising 16000 (10000 in English and 6000 in Russian) distinct spoken equations uttered by 3 different speakers. Our approaches, which incorporate ASR post-correction and multi-modal language models, demonstrate a notable performance with up to a 25% character error rate (CER).", "keywords": [ "speech recognition", "LLM", "LaTeX", "speech to text", "ASR", "STT" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 4, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "pfXVid1p1d", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 15, "author_info_list": [ { "name": "Feng Yang", "gender": "Male", "institution": "Google", "country": "US", "position": "Principal Engineer" }, { "name": "Hang Qi", "gender": "unknown", "institution": "DeepMind", "country": "US", "position": "Researcher" }, { "name": "Harris Shadmany", "gender": "Male", "institution": "Google", "country": "US", "position": "Machine Learning Engineer" }, { "name": "Irfan Essa", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Principal Researcher" }, { "name": "Jesse Berent", "gender": "Male", "institution": "Google", "country": "CH", "position": "Principal Researcher" }, { "name": "Jia-Bin Huang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Associate Professor" }, { "name": "Jose Lezama", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Junjie Ke", "gender": "Female", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Luciano Sbaiz", "gender": "Male", "institution": "Google", "country": "CH", "position": "Research Scientist" }, { "name": "Miaosen Wang", "gender": "unknown", "institution": "DeepMind", "country": "US", "position": "Research Engineer" }, { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Siqi Xie", "gender": "unknown", "institution": "Google", "country": "US", "position": "Software Engineer" }, { "name": "Yinxiao Li", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Researcher" }, { "name": "Yiran Xu", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Han Zhang", "gender": "Male", "institution": "Reve AI", "country": "US", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 64, "n_ref": 171, "n_ref_all": 226, "n_fig": 19, "n_tab": 4, "L_tab": 1103, "n_element_tab": 136, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 724, "n_element_tab_1": 93, "formula_len_all": 648, "formula_len_all_1": 441, "len_all": 151566, "len_all_1": 57243, "len_abs": 375, "len_title": 126, "len_sents": 36732, "len_sents_1": 24375, "n_sents": 343, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1046, "title": "HALO: Human-Aligned End-to-end Image Retargeting with Layered Transformations", "abs": "Image retargeting aims to change the aspect-ratio of an image while maintaining its content and structure with less visual artifacts. \nExisting methods still generate many artifacts or lose a lot of original content or structure. To address this, we introduce HALO, an end-to-end trainable solution for image retargeting. \nThe core idea of HALO is to warp the input image to target resolution. \nSince humans are more sensitive to distortions in salient areas than non-salient areas of an image, HALO decomposes the input image into salient/non-salient layers and applies different wrapping fields to different layers. To further minimize the structure distortion in the output images, we propose perceptual structure similarity loss which measures the structure similarity between input and output images and aligns with human perception. Both quantitative results and a user study on the RetargetMe dataset show that our algorithm achieves SOTA. \nEspecially, our method increases human preference by 13.21% compared with the second best method.", "keywords": [ "Image Retargeting", "Image Transformation", "Image Editing" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 3, 5 ] }, { "paper_id": "pfS4D6RWC8", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Ehsan Hajiramezanali", "gender": "Male", "institution": "Genentech", "country": "US", "position": "Principal Researcher" }, { "name": "Gabriele Scalia", "gender": "unknown", "institution": "Genentech", "country": "", "position": "Researcher" }, { "name": "Kevin Black", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Masatoshi Uehara", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Nathaniel Lee Diamant", "gender": "unknown", "institution": "genentech", "country": "", "position": "Researcher" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Tommaso Biancalani", "gender": "Male", "institution": "Genentech", "country": "", "position": "Director" }, { "name": "Yulai Zhao", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Alex Tseng", "gender": "Non-Binary", "institution": "Genentech", "country": "", "position": "Researcher" } ], "n_formula": 97, "n_formula_1": 32, "n_ref_uni": 62, "n_ref": 120, "n_ref_all": 173, "n_fig": 6, "n_tab": 5, "L_tab": 1326, "n_element_tab": 111, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 229, "n_element_tab_1": 24, "formula_len_all": 8489, "formula_len_all_1": 2460, "len_all": 214392, "len_all_1": 70915, "len_abs": 1428, "len_title": 127, "len_sents": 57088, "len_sents_1": 30101, "n_sents": 554, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1396, "title": "Fine-Tuning of Continuous-Time Diffusion Models as Entropy-Regularized Control", "abs": "Diffusion models excel at capturing complex data distributions, such as those of natural images and proteins. While diffusion models are trained to represent the distribution in the training dataset, we often are more concerned with other properties, such as the aesthetic quality of the generated images or the functional properties of generated proteins. Diffusion models can be finetuned in a goal-directed way by maximizing the value of some reward function (e.g., the aesthetic quality of an image). However, these approaches may lead to reduced sample diversity, significant deviations from the training data distribution, and even poor sample quality due to the exploitation of an imperfect reward function. The last issue often occurs when the reward function is a learned model meant to approximate a ground-truth \"genuine\" reward, as is the case in many practical applications. These challenges, collectively termed \"reward collapse,\" pose a substantial obstacle. To address this reward collapse, we frame the finetuning problem as entropy-regularized control against the pretrained diffusion model, i.e., directly optimizing entropy-enhanced rewards with neural SDEs. We present theoretical and empirical evidence that demonstrates our framework is capable of efficiently generating diverse samples with high genuine rewards, mitigating the overoptimization of imperfect reward models.", "keywords": [ "diffusion models", "reinforcement learning" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "pf9J3GNxSe", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Kai Nakaishi", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Koji Hukushima", "gender": "Male", "institution": "University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Yoshihiko Nishikawa", "gender": "unknown", "institution": "Kitasato University", "country": "JP", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 46, "n_ref": 92, "n_ref_all": 122, "n_fig": 20, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 446, "formula_len_all_1": 307, "len_all": 151908, "len_all_1": 63898, "len_abs": 1118, "len_title": 119, "len_sents": 48300, "len_sents_1": 31838, "n_sents": 370, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1123, "title": "Critical Phase Transition in Large Language Models", "abs": "Large Language Models (LLMs) have demonstrated impressive performance. To understand their behaviors, we need to consider the fact that LLMs sometimes show qualitative changes. The natural world also presents such changes called phase transitions, which are defined by singular, divergent statistical quantities. Therefore, an intriguing question is whether qualitative changes in LLMs are phase transitions. In this work, we have conducted extensive analysis on texts generated by LLMs and suggested that a phase transition occurs in LLMs when varying the temperature parameter. Specifically, statistical quantities have divergent properties just at the point between the low-temperature regime, where LLMs generate sentences with clear repetitive structures, and the high-temperature regime, where generated sentences are often incomprehensible. In addition, critical behaviors near the phase transition point, such as a power-law decay of correlation and slow convergence toward the stationary state, are similar to those in natural languages. Our results suggest a meaningful analogy between LLMs and natural phenomena.", "keywords": [ "natural language processing", "large language models", "statistical physics", "phase transitions", "critical phenomena", "large-scale numerical experiments" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "pf7OGmRxa5", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Hung-Yu Kao", "gender": "Male", "institution": "Department of Computer Science, National Tsing Hua University, National Tsinghua University", "country": "TW", "position": "Full Professor" }, { "name": "Kun-Ta Chuang", "gender": "Male", "institution": "National Cheng Kung University", "country": "", "position": "Associate Professor" }, { "name": "Ying-Jia Lin", "gender": "Male", "institution": "Chang Gung University", "country": "TW", "position": "Assistant Professor" }, { "name": "Zhi-Quan Feng", "gender": "unknown", "institution": "National Cheng Kung University", "country": "TW", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 17, "n_ref": 27, "n_ref_all": 48, "n_fig": 3, "n_tab": 10, "L_tab": 3413, "n_element_tab": 504, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1162, "n_element_tab_1": 182, "formula_len_all": 784, "formula_len_all_1": 775, "len_all": 102321, "len_all_1": 61876, "len_abs": 1637, "len_title": 129, "len_sents": 31088, "len_sents_1": 27274, "n_sents": 241, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1654, "title": "Symbiotic Tuning: A Simple Approach for Enhancing Task Performance of Side-Tuning", "abs": "The reduction of the computational and memory overhead associated with fine-tuning large language models remains a significant challenge for current research in natural language processing. Achieving an optimal balance between task performance, adaptability, and low memory requirement often presents a complex trade-off. Parameter-efficient fine-tuning (PEFT) methods, such as LoRA, have gained attention for their ability to reduce the number of trainable parameters while preserving task performance. However, they have not yet achieved a notable reduction in memory usage, which is still predominantly consumed by model weights and activations during backpropagation. In contrast, Ladder Side-Tuning (LST) has been proposed as an alternative that effectively reduces memory usage by freezing the backbone language model (BLM) and training only lightweight side networks. Nevertheless, this reduction in memory usage often results in a decline in performance, as LST typically exhibits inferior performance compared to PEFT methods on the same BLM. To address these limitations, we propose Symbiotic Tuning (SymTune), a novel approach that extracts intermediate outputs from the BLM and integrates symbiotic modules to enhance feature processing capabilities. This method avoids a direct trade-off between performance and memory efficiency, offering two key advantages: 1) robust performance across a wide range of natural language tasks, and 2) reduced memory consumption through an improved side-tuning architecture. The experimental results demonstrate that SymTune provides a scalable and memory-efficient solution for fine-tuning language models.", "keywords": [ "Natural Language Processing", "Parameter-Efficient Fine-Tuning" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "peX9zpWgg4", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Christos Davatzikos", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Dr. Center for Biomedical Image Computing and Analytics/Director Section of Biomedical Image Analysis" }, { "name": "Haochang Shou", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Associate Professor" }, { "name": "Vasiliki Tassopoulou", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 18, "n_ref_uni": 21, "n_ref": 26, "n_ref_all": 40, "n_fig": 7, "n_tab": 2, "L_tab": 721, "n_element_tab": 60, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 236, "n_element_tab_1": 22, "formula_len_all": 905, "formula_len_all_1": 923, "len_all": 140892, "len_all_1": 65771, "len_abs": 2157, "len_title": 149, "len_sents": 36190, "len_sents_1": 31765, "n_sents": 256, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 2165, "title": "Adaptive Shrinkage Estimation for Personalized Deep Kernel Regression in Modeling Brain Trajectories", "abs": "Longitudinal biomedical studies track individuals over time to capture dynamics in brain development, disease progression, and treatment effects. Estimating trajectories of brain measurements in such studies is challenging due to biological variability and inconsistencies in measurement protocols (e.g. MRI scanner variations and upgrades). Herein, we introduce a novel personalized Deep Kernel Regression framework for forecasting longitudinal regional bran volumetric changes. Our approach integrates two key components: a population model that captures brain volume trajectories from a large and diverse cohort, and a personalization step that generates subject-specific models for individual trajectories. To optimally combine these predictive distributions, we propose the Adaptive Posterior Shrinkage Estimation technique, which effectively balances population-level trends with individual-specific data. We evaluate model's performance through predictive accuracy metrics, uncertainty quantification, and validation against external clinical studies. Benchmarking against state-of-the-art statistical and machine learning models\u2014including Linear Mixed Effects models, Generalized Additive Models, and deep learning methods\u2014demonstrates the superior predictive performance of our approach across a variety of experiments. Additionally, we apply our method on predicting trajectories of composite neuroimaging biomarkers, e.g. machine-learning patterns of brain structure related to aging and Alzheimer's Disease, which highlights the generalizability of our approach to model the progression of longitudinal monotonic biomarkers. Furthermore, validation on three external neuroimaging studies confirms the generalizability and applicability of our method across different clinical contexts. These results highlight the versatility and robustness of our framework for predicting longitudinal brain volume changes. Overall, this framework effectively addresses the inherent challenges in longitudinal biomedical studies, providing a valuable predictive tool that can inform patient management decisions, clinical trial design and treatment effect estimation.", "keywords": [ "Deep Kernel Regression", "Personalization", "Posterior Correction", "Longitudinal Biomarker Prediction" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "peNgxpbdxB", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 6, "author_info_list": [ { "name": "Haoyu Peter Wang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Martin Ennemoser", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "MS student" }, { "name": "Sebastian Lehner", "gender": "unknown", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "Postdoc" }, { "name": "Sebastian Sanokowski", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "PhD student" }, { "name": "Sepp Hochreiter", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "Full Professor" }, { "name": "Wilhelm Franz Berghammer", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "", "position": "Undergrad student" } ], "n_formula": 53, "n_formula_1": 15, "n_ref_uni": 58, "n_ref": 129, "n_ref_all": 136, "n_fig": 1, "n_tab": 9, "L_tab": 3231, "n_element_tab": 259, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 1498, "n_element_tab_1": 144, "formula_len_all": 5347, "formula_len_all_1": 1304, "len_all": 233266, "len_all_1": 79609, "len_abs": 1378, "len_title": 137, "len_sents": 68583, "len_sents_1": 36007, "n_sents": 548, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1384, "title": "Scalable Discrete Diffusion Samplers: Combinatorial Optimization and Statistical Physics", "abs": "Learning to sample from complex unnormalized distributions over discrete domains emerged as a promising research direction with applications in statistical physics, variational inference, and combinatorial optimization. Recent work has demonstrated the potential of diffusion models in this domain. However, existing methods face limitations in memory scaling and thus the number of attainable diffusion steps since they require backpropagation through the entire generative process. To overcome these limitations we introduce two novel training methods for discrete diffusion samplers, one grounded in the policy gradient theorem and the other one leveraging Self-Normalized Neural Importance Sampling (SN-NIS). These methods yield memory-efficient training and achieve state-of-the-art results in unsupervised combinatorial optimization.\nNumerous scientific applications additionally require the ability of unbiased sampling. We introduce adaptations of SN-NIS and Neural Markov Chain Monte Carlo that enable for the first time the application of discrete diffusion models to this problem. We validate our methods on Ising model benchmarks and find that they outperform popular autoregressive approaches. Our work opens new avenues for applying diffusion models to a wide range of scientific applications in discrete domains that were hitherto restricted to exact likelihood models.", "keywords": [ "Combinatorial Optimization", "Diffusion Models", "Statistical Physics" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 2, 3, 4 ] }, { "paper_id": "pdzHpQbGrn", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Abir Das", "gender": "Male", "institution": "Indian Institute of Technology Kharagpur", "country": "IN", "position": "Assistant Professor" }, { "name": "Aprameyo Chakrabartty", "gender": "unknown", "institution": "Indian Institute of Technology Kharagpur", "country": "IN", "position": "Undergrad student" }, { "name": "Bibhudatta Bhanja", "gender": "Male", "institution": "Indian Institute of Technology Kharagpur", "country": "IN", "position": "Undergrad student" }, { "name": "Dhruv Sarkar", "gender": "Male", "institution": "Indian Institute of Technology Kharagpur", "country": "IN", "position": "Undergrad student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 45, "n_ref": 79, "n_ref_all": 91, "n_fig": 5, "n_tab": 4, "L_tab": 1528, "n_element_tab": 120, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1550, "n_element_tab_1": 154, "formula_len_all": 405, "formula_len_all_1": 376, "len_all": 124583, "len_all_1": 58555, "len_abs": 1320, "len_title": 106, "len_sents": 35558, "len_sents_1": 26796, "n_sents": 264, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1325, "title": "ACTIVE TEST TIME PROMPT LEARNING IN VISION- LANGUAGE MODELS", "abs": "Test Time Optimisation is a setting where a model is made to learn new parameters on-the-fly during inference with the help of those very samples it is supposed to be tested on. Learning prompts at test time to improve the performance of Vision Language Models(VLMs) in downstream tasks has become a popular setting in recent times. In this paper, we propose a new framework for the Test Time Prompt Tuning in Pre-trained VLMs which incorporates actively sampled labels in the learning process to improve the performance of the model in downstream test-time settings. Our problem setting is underexplored yet well-motivated by considerations such as performance, efficiency and real-life applicability. Active Learning can be especially beneficial in the test-time setting in providing the option to query the true label when the model is uncertain in a real-life scenario and Prompt Tuning provides the advantage due to parameter efficiency. Our method is guided by these two principles and successfully combines the two to come up with a test-time optimisation scheme that is evaluated to be an improvement over existing methods under a fair evaluation protocol. We conduct experiments across 10 cross-dataset transfer datasets and 4 domain-generalisation datasets to show consistent improvement over the state-of-the-art.", "keywords": [ "Vision-Language Models", "Prompt Learning", "Active Learning", "Test-Time Adaptation" ], "rating_list": [ 3, 3, 1, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 1, 2, 1 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "pdjkikvCch", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Fabio Cuzzolin", "gender": "Male", "institution": "Oxford Brookes University", "country": "GB", "position": "Full Professor" }, { "name": "Kaizheng Wang", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "PhD student" }, { "name": "Keivan Shariatmadar", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "Postdoc" }, { "name": "Muhammad Mubashar", "gender": "Male", "institution": "Oxford Brookes University", "country": "", "position": "PhD student" }, { "name": "Shireen Kudukkil Manchingal", "gender": "Female", "institution": "Oxford Brookes University", "country": "", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 12, "n_ref_uni": 121, "n_ref": 202, "n_ref_all": 235, "n_fig": 17, "n_tab": 16, "L_tab": 7594, "n_element_tab": 691, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 7036, "n_element_tab_1": 317, "formula_len_all": 1231, "formula_len_all_1": 749, "len_all": 338480, "len_all_1": 71244, "len_abs": 1341, "len_title": 74, "len_sents": 103384, "len_sents_1": 29586, "n_sents": 742, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 26, "L_abs": 1348, "title": "Random-Set Neural Networks", "abs": "Machine learning is increasingly deployed in safety-critical domains where erroneous predictions may lead to potentially catastrophic consequences, highlighting the need for learning systems to be aware of how confident they are in their own predictions: in other words, 'to know when they do not know\u2019. In this paper, we propose a novel Random-Set Neural Network (RS-NN) approach to classification which predicts *belief functions* (rather than classical probability vectors) over the class list using the mathematics of *random sets*, i.e., distributions over the collection of *sets* of classes. RS-NN encodes the 'epistemic' uncertainty induced by training sets that are insufficiently representative or limited in size via the size of the convex set of probability vectors associated with a predicted belief function. Our approach outperforms state-of-the-art Bayesian and Ensemble methods in terms of accuracy, uncertainty estimation and out-of-distribution (OoD) detection on multiple benchmarks (CIFAR-10 vs SVHN/Intel-Image, MNIST vs FMNIST/KMNIST, ImageNet vs ImageNet-O). RS-NN also scales up effectively to large-scale architectures (e.g. WideResNet-28-10, VGG16, Inception V3, EfficientNetB2 and ViT-Base-16),\nexhibits remarkable robustness to adversarial attacks and can provide statistical guarantees in a conformal learning setting.", "keywords": [ "Uncertainty quantification", "Classification", "Deep Learning" ], "rating_list": [ 5, 8, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 4, 4, 3 ], "contribution_list": [ 2, 4, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "pdf6MbXnAS", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Jan Snajder", "gender": "Male", "institution": "UniZg-FER, University of Zagreb", "country": "HR", "position": "Full Professor" }, { "name": "Josip Juki\u0107", "gender": "unknown", "institution": "Faculty of Electrical Engineering and Computing, University of Zagreb", "country": "", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 12, "n_ref_uni": 34, "n_ref": 63, "n_ref_all": 96, "n_fig": 5, "n_tab": 14, "L_tab": 1323, "n_element_tab": 147, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1511, "n_element_tab_1": 98, "formula_len_all": 862, "formula_len_all_1": 613, "len_all": 174378, "len_all_1": 67905, "len_abs": 1009, "len_title": 120, "len_sents": 57001, "len_sents_1": 32520, "n_sents": 406, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1012, "title": "Disentangling Latent Shifts of In-Context Learning Through Self-Training", "abs": "In-context learning (ICL) has become essential in natural language processing, particularly with autoregressive large language models capable of learning from demonstrations provided within the prompt. However, ICL faces challenges with stability and long contexts, especially as the number of demonstrations grows, leading to poor generalization and inefficient inference. To address these issues, we introduce STICL (Self-Training ICL), an approach that disentangles the latent shifts of demonstrations from the latent shift of the query through self-training. STICL employs a teacher model to generate pseudo-labels and trains a student model using these labels, encoded in an adapter module. The student model exhibits weak-to-strong generalization, progressively refining its predictions over time. Our empirical results show that STICL improves generalization and stability, consistently outperforming traditional ICL methods and other disentangling strategies across both in-domain and out-of-domain data.", "keywords": [ "large language models", "in-context learning", "self-training", "latent shifts" ], "rating_list": [ 8, 3, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "pdF86dyoS6", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Rujun Jiang", "gender": "Male", "institution": "Fudan University", "country": "", "position": "Associate Professor" }, { "name": "Yuze Ge", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" } ], "n_formula": 95, "n_formula_1": 24, "n_ref_uni": 64, "n_ref": 125, "n_ref_all": 159, "n_fig": 8, "n_tab": 11, "L_tab": 6786, "n_element_tab": 386, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 617, "n_element_tab_1": 108, "formula_len_all": 11263, "formula_len_all_1": 1741, "len_all": 220490, "len_all_1": 64662, "len_abs": 987, "len_title": 109, "len_sents": 54392, "len_sents_1": 27636, "n_sents": 590, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1035, "title": "SOREL: A Stochastic Algorithm for Spectral Risks Minimization", "abs": "The spectral risk has wide applications in machine learning, especially in real-world decision-making, where people are concerned with more than just average model performance. By assigning different weights to the losses of different sample points, rather than the same weights as in the empirical risk, it allows the model's performance to lie between the average performance and the worst-case performance. In this paper, we propose SOREL, the first stochastic gradient-based algorithm with convergence guarantees for spectral risks minimization. Previous approaches often rely on smoothing the spectral risk by adding a strongly concave function, thereby lacking convergence guarantees for the original spectral risk. We theoretically prove that our algorithm achieves a near-optimal rate of $\\widetilde{O}(1/\\sqrt{\\epsilon})$ to obtain an $\\epsilon$-optimal solution in terms $\\epsilon$. Experiments on real datasets show that our algorithm outperforms existing ones in most cases, both in terms of runtime and sample complexity.", "keywords": [ "spectral risk", "conditional Value-at-Risk", "stochastic optimization", "convex optimization", "fair machine learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 2, 5, 3, 3 ] }, { "paper_id": "pcnq7fZs4t", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Jianjia Xin", "gender": "Male", "institution": "Inspur Software Co, Ltd", "country": "CN", "position": "Postdoc" }, { "name": "Kai Xu", "gender": "Male", "institution": "Beijing University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Lichun WANG", "gender": "Female", "institution": "Beijing University of Technology", "country": "", "position": "Full Professor" }, { "name": "Shuang Li", "gender": "Female", "institution": "Beijing Information Science and Technology University", "country": "CN", "position": "Lecturer" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 5, "n_ref": 8, "n_ref_all": 18, "n_fig": 2, "n_tab": 1, "L_tab": 761, "n_element_tab": 44, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 761, "n_element_tab_1": 44, "formula_len_all": 637, "formula_len_all_1": 637, "len_all": 82547, "len_all_1": 56304, "len_abs": 1084, "len_title": 104, "len_sents": 22599, "len_sents_1": 22413, "n_sents": 141, "n_sents_1": 138, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1090, "title": "Common Feature Learning for Zero-shot Image Recognition", "abs": "The key issue of zero-shot image recognition (ZIR) is how to infer the relationship between visual space and semantic space from seen classes, and then effectively transfer the relationship to unseen classes. Recently, most methods have focused on how to use images and class semantic vectors or class names to learn the relationship between visual space and semantic space. The relationship established by these two methods is class-level and coarse-grained. The differences between images of the same class are ignored, which leads to insufficiently tight relationships and affects the accurate recognition of unseen classes.To tackle such problem, we propose Common Feature learning for Zero-shot Image Recognition (CF-ZIR) method to learn fine-grained visual semantic relationships at the image-level. Based on the inter class association information provided by class semantic vectors, guide the extraction of common visual features between classes to obtain image semantic vectors. Experiments on three widely used benchmark datasets show the effectiveness of the proposed approach.", "keywords": [ "Zero-shot Image Recognition\uff1bVisual-semantic Relationship\uff1bFine-grained Alignment\uff1bSemantic Vectors Generation\uff1b" ], "rating_list": [ 3, 3, 5, 1 ], "soundness_list": [ 1, 2, 3, 1 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "pcIDLhnYL9", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Lorenzo Brigato", "gender": "Male", "institution": "Universit\u00e4t Bern", "country": "CH", "position": "Postdoc" }, { "name": "Stavroula Mougiakakou", "gender": "Female", "institution": "Universit\u00e4t Bern", "country": "CH", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 93, "n_ref": 164, "n_ref_all": 189, "n_fig": 7, "n_tab": 4, "L_tab": 1396, "n_element_tab": 253, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1672, "n_element_tab_1": 260, "formula_len_all": 492, "formula_len_all_1": 492, "len_all": 196191, "len_all_1": 73815, "len_abs": 1213, "len_title": 123, "len_sents": 45473, "len_sents_1": 33838, "n_sents": 296, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1238, "title": "Can the Training Loss be Predictive for Out-of-Distribution Generalization?", "abs": "Traditional model selection in deep learning relies on carefully tuning several hyper-parameters (HPs) controlling regularization strength on held-out validation data, which can be challenging to obtain in scarce-data scenarios or may not accurately reflect real-world deployment conditions due to distribution shifts.\nMotivated by such issues, this paper investigates the potential of using solely the training loss to predict the generalization performance of neural networks on out-of-distribution (OOD) test scenarios.\nOur analysis reveals that preserving consistent prediction variance across training and testing distributions is essential for establishing a correlation between training loss and OOD generalization.\nWe propose architectural adjustments to ensure $\\textit{variance preservation}$, enabling reliable model selection based on training loss alone, even in over-parameterized settings with a sample-to-parameter ratio exceeding four orders of magnitude.\nWe extensively assess the model-selection capabilities of $\\textit{variance-preserving}$ architectures on several scarce data, domain-shift, and corruption benchmarks by optimizing HPs such as learning rate, weight decay, batch size, and data augmentation strength.", "keywords": [ "deep learning", "OOD generalization", "signal propagation", "hyper-parameter search" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "pbre0HKsfE", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Donghwan Rho", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Ernest K. Ryu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "Hyunsik Chae", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Jung Hee Cheon", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Full Professor" }, { "name": "Jung Woo Kim", "gender": "unknown", "institution": "Seoul National University", "country": "", "position": "Postdoc" }, { "name": "Minje Park", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Taeseong Kim", "gender": "Male", "institution": "Seoul National University, Seoul National University", "country": "", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 6, "n_ref_uni": 71, "n_ref": 125, "n_ref_all": 156, "n_fig": 9, "n_tab": 19, "L_tab": 4762, "n_element_tab": 567, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 977, "n_element_tab_1": 124, "formula_len_all": 1473, "formula_len_all_1": 628, "len_all": 209709, "len_all_1": 67916, "len_abs": 906, "len_title": 84, "len_sents": 52407, "len_sents_1": 31202, "n_sents": 438, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 36, "L_abs": 928, "title": "Encryption-Friendly LLM Architecture", "abs": "Large language models (LLMs) offer personalized responses based on user interactions, but this use case raises serious privacy concerns. Homomorphic encryption (HE) is a cryptographic protocol supporting arithmetic computations in encrypted states and provides a potential solution for privacy-preserving machine learning (PPML). However, the computational intensity of transformers poses challenges for applying HE to LLMs. In this work, we propose a modified HE-friendly transformer architecture with an emphasis on inference following personalized (private) fine-tuning. Utilizing LoRA fine-tuning and Gaussian kernels, we achieve significant computational speedups---6.94$\\times$ for fine-tuning and 2.3$\\times$ for inference---while maintaining performance comparable to plaintext models. Our findings provide a viable proof of concept for offering privacy-preserving LLM services in areas where data protection is crucial.", "keywords": [ "Homomorphic Encryption", "PPML", "Encrypted Fine-tuning", "Transformer Architecture" ], "rating_list": [ 5, 8, 6 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 4, 4 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "pblB72EmrM", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Chen Tessler", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Fengyi Wang", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "Undergrad student" }, { "name": "Haotian Zhang", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Jessica K. Hodgins", "gender": "Female", "institution": "Boston Dynamics AI Institute", "country": "US", "position": "Researcher" }, { "name": "Jiashun Wang", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Jingbo Wang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Jinkun Cao", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Kangni Liu", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Kris M. Kitani", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Ye Yuan", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Researcher" }, { "name": "Zhengyi Luo", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zihui Lin", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 3, "n_ref_uni": 52, "n_ref": 105, "n_ref_all": 121, "n_fig": 6, "n_tab": 9, "L_tab": 1567, "n_element_tab": 182, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 793, "n_element_tab_1": 111, "formula_len_all": 1438, "formula_len_all_1": 252, "len_all": 182050, "len_all_1": 70086, "len_abs": 1222, "len_title": 120, "len_sents": 51286, "len_sents_1": 30504, "n_sents": 455, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1154, "title": "HumanoidOlympics: Sports Environments for Physically Simulated Humanoids", "abs": "We present HumanoidOlympics, a collection of physically simulated sports environments designed for the animation and robotics communities to develop humanoid behaviors. Our suite includes individual sports such as golf, javelin throw, high jump, long jump, and hurdling, as well as competitive games like table tennis, tennis, fencing, boxing, soccer, and basketball. By simulating a wide range of Olympic sports, HumanoidOlympics offers a rich and standardized testing ground to evaluate and develop learning algorithms due to the diversity and physically demanding nature of athletic activities. Our suite supports simulating both graphics-focused (SMPL and SMPL-X) and real-world humanoid robots. For each sport, we benchmark popular humanoid control methods and provide expert-designed rewards that lead to surprising simulation results. Our analysis shows that leveraging human demonstrations can significantly enhance the resulting policies' human likeness and task performance. By providing a unified and competitive sports benchmark, HumanoidOlympics can help the animation and robotics communities develop human-like and performant controllers.", "keywords": [ "Physics Simulation", "Embodied AI", "Benchmark", "Sports" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "pbJMPo4HwR", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Min zhang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yang Feng", "gender": "unknown", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhengrui Ma", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 39, "n_ref": 81, "n_ref_all": 102, "n_fig": 7, "n_tab": 9, "L_tab": 5559, "n_element_tab": 434, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1595, "n_element_tab_1": 233, "formula_len_all": 716, "formula_len_all_1": 738, "len_all": 168899, "len_all_1": 64550, "len_abs": 1268, "len_title": 115, "len_sents": 39259, "len_sents_1": 28920, "n_sents": 296, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1273, "title": "Learning Monotonic Attention in Transducer for Streaming Generation", "abs": "Streaming generation models are increasingly utilized across various fields, with the Transducer architecture being particularly popular in industrial applications. However, its input-synchronous decoding mechanism presents challenges in tasks requiring non-monotonic alignments, such as simultaneous translation, leading to suboptimal performance in these contexts. In this research, we address this issue by tightly integrating Transducer's decoding with the history of input stream via a learnable monotonic attention mechanism. Our approach leverages the forward-backward algorithm to infer the posterior probability of alignments between the predictor states and input timestamps, which is then used to estimate the context representations of monotonic attention in training. This allows Transducer models to adaptively adjust the scope of attention based on their predictions, avoiding the need to enumerate the exponentially large alignment space. Extensive experiments demonstrate that our MonoAttn-Transducer significantly enhances the handling of non-monotonic alignments in streaming generation, offering a robust solution for Transducer-based frameworks to tackle more complex streaming generation tasks. Codes are publicly available in supplementary materials.", "keywords": [ "streaming generation", "simultaneous translation", "Transducer" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "pbDqZBn2X2", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Deunsol Yoon", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Researcher" }, { "name": "Hyungseok Song", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Researcher" }, { "name": "Kanghoon Lee", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Principal Researcher" }, { "name": "Woohyung Lim", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Vice President" } ], "n_formula": 18, "n_formula_1": 10, "n_ref_uni": 35, "n_ref": 115, "n_ref_all": 130, "n_fig": 4, "n_tab": 9, "L_tab": 7763, "n_element_tab": 1057, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 6382, "n_element_tab_1": 775, "formula_len_all": 1090, "formula_len_all_1": 630, "len_all": 181146, "len_all_1": 72383, "len_abs": 1345, "len_title": 131, "len_sents": 41766, "len_sents_1": 25232, "n_sents": 305, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1353, "title": "CADO: Cost-Aware Diffusion Models for Combinatorial Optimization via RL Fine-tuning", "abs": "Recent advancements in Machine Learning (ML) have demonstrated significant potential in addressing Combinatorial Optimization (CO) problems through data-driven approaches. Heatmap-based methods, which generate solution heatmaps in a single step and employ an additional decoder to derive solutions for CO tasks, have shown promise due to their scalability for large-scale problems. Traditionally, these complex models are trained using imitation learning with optimal solutions, often leveraging diffusion models. However, our research has identified several limitations inherent in these imitation learning approaches within the context of CO tasks. To overcome these challenges, we propose a 2-phase training framework for diffusion models in CO, incorporating Reinforcement Learning (RL) fine-tuning. Our methodology integrates cost information and the post-process decoder into the training process, thereby enhancing the solver's capacity to generate effective solutions. We conducted extensive experiments on well-studied combinatorial optimization problems, specifically the Traveling Salesman Problem (TSP) and Maximal Independent Set (MIS), ranging from small-scale instances to large-scale scenarios. The results demonstrate the significant efficacy of our RL fine-tuning framework, surpassing previous state-of-the-art methods in performance.", "keywords": [ "Combinatorial Optimization", "Diffusion Model", "RL finetuning" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "pZz0nOroGv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Emily Ruoyu Liu", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Ines Dormoy", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "MS student" }, { "name": "Jeremy Andrew Irvin", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Jinyoung Kim", "gender": "Not Specified", "institution": "Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Joyce C. Chen", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Samar Khanna", "gender": "unknown", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Researcher" }, { "name": "Stefano Ermon", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Zhuo Zheng", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 70, "n_ref": 183, "n_ref_all": 220, "n_fig": 8, "n_tab": 9, "L_tab": 2659, "n_element_tab": 312, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1616, "n_element_tab_1": 164, "formula_len_all": 33, "formula_len_all_1": 0, "len_all": 252965, "len_all_1": 62693, "len_abs": 3051, "len_title": 126, "len_sents": 63488, "len_sents_1": 29345, "n_sents": 428, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1341, "title": "TEOChat: A Large Vision-Language Assistant for Temporal Earth Observation Data", "abs": "Large vision and language assistants have enabled new capabilities for interpreting natural images. These approaches have recently been adapted to earth observation data, but they are only able to handle single image inputs, limiting their use for many real-world tasks. In this work, we develop a new vision and language assistant called TEOChat that can engage in conversations about temporal sequences of earth observation data. To train TEOChat, we curate an instruction-following dataset composed of many single image and temporal tasks including building change and damage assessment, semantic change detection, and temporal scene classification. We show that TEOChat can perform a wide variety of spatial and temporal reasoning tasks, substantially outperforming previous vision and language assistants, and even achieving comparable or better performance than several specialist models trained to perform specific tasks. Furthermore, TEOChat achieves impressive zero-shot performance on a change detection and change question answering dataset, outperforms GPT-4o and Gemini 1.5 Pro on multiple temporal tasks, and exhibits stronger single image capabilities than a comparable single image instruction-following model on scene classification, visual question answering, and captioning. We publicly release our data, models, and code.", "keywords": [ "vision-language model", "large multimodal model", "satellite imagery", "earth observation", "change detection" ], "rating_list": [ 8, 6, 5, 1 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "pZk9cUu8p6", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 7, "author_info_list": [ { "name": "Duo Zhang", "gender": "Male", "institution": "University of California, Merced", "country": "US", "position": "Postdoc" }, { "name": "Guanpeng Li", "gender": "unknown", "institution": "University of Iowa", "country": "US", "position": "Assistant Professor" }, { "name": "Sheng Di", "gender": "Male", "institution": "Argonne National Laboratory", "country": "US", "position": "Researcher" }, { "name": "Weicong Chen", "gender": "Male", "institution": "University of California, Merced", "country": "US", "position": "Postdoc" }, { "name": "Xiaoyi Lu", "gender": "Male", "institution": "University of California, Merced", "country": "US", "position": "Associate Professor" }, { "name": "Zhaorui Zhang", "gender": "unknown", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "Zhonghao Chen", "gender": "unknown", "institution": "University of California, Merced", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 22, "n_ref": 29, "n_ref_all": 42, "n_fig": 5, "n_tab": 3, "L_tab": 549, "n_element_tab": 34, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 194, "n_element_tab_1": 5, "formula_len_all": 132, "formula_len_all_1": 207, "len_all": 97779, "len_all_1": 60832, "len_abs": 1606, "len_title": 117, "len_sents": 32499, "len_sents_1": 30510, "n_sents": 194, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1611, "title": "FedDES: A Discrete-Event Simulator For Large-Scale Federated Learning", "abs": "We introduce FedDES, a performance simulator for Federated Learning (FL) that leverages Discrete Event Simulation (DES) techniques to model key events\u2014such as client updates, communication delays, and aggregation operations\u2014as discrete occurrences in time. This approach accurately captures the runtime features of FL systems, providing a high-fidelity simulation environment that closely mirrors real-world deployments. FedDES incorporates all three known aggregation settings: Synchronous (e.g., FedAvg and FedProx), Asynchronous (e.g., FedAsync and FedFa), and Semi-Asynchronous (e.g., FedBuff and FedCompass). Designed to be framework-, dataset-, and model-agnostic, FedDES allows researchers and developers to explore various configurations without restrictions. Our evaluations involving over 1,000 clients with heterogeneous computation and communication characteristics demonstrate that FedDES accurately models event distribution and delivers performance estimates within 2% error of real-world measurements. While real-world workloads often take hours to evaluate, FedDES generates detailed, timestamped event logs in just few seconds. As a result, FedDES can significantly accelerate FL developing and debugging cycles, enabling developers to rapidly prototype and evaluate algorithms and system designs, bypassing the need for costly, time-consuming real-world deployments. It offers valuable performance insights\u2014such as identifying bottlenecks, stragglers, fault-tolerance mechanisms, and edge-case scenarios\u2014facilitating the optimization of FL systems for efficiency, scalability, and resilience.", "keywords": [ "Federated Learning", "Discrete Event-Driven Simulations" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 1, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "pZiyCaVuti", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Di Wu", "gender": "Not Specified", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Dong Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Distinguished Scientist" }, { "name": "Hongwei Wang", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Kai-Wei Chang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Associate Professor" }, { "name": "Wenhao Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Researcher" }, { "name": "Yuwei Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 41, "n_ref": 88, "n_ref_all": 135, "n_fig": 16, "n_tab": 4, "L_tab": 2663, "n_element_tab": 258, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1125, "n_element_tab_1": 56, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 217835, "len_all_1": 60985, "len_abs": 1608, "len_title": 122, "len_sents": 66613, "len_sents_1": 29124, "n_sents": 494, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1614, "title": "LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory", "abs": "Recent large language model (LLM)-driven chat assistant systems have integrated memory components to track user-assistant chat histories, enabling more accurate and personalized responses. However, their long-term memory capabilities in sustained interactions remain underexplored. This paper introduces LongMemEval, a comprehensive benchmark designed to evaluate five core long-term memory abilities of chat assistants: information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention. With 500 meticulously curated questions embedded within freely scalable user-assistant chat histories, LongMemEval presents a significant challenge to existing long-term memory systems, with commercial chat assistants and long-context LLMs showing 30\\% accuracy drop on memorizing information across sustained interactions. We then present a unified framework that breaks down the long-term memory design into four design choices across the indexing, retrieval, and reading stages. Built upon key experimental insights, we propose several memory designs including session decomposition for optimizing value granularity, fact-augmented key expansion for enhancing the index structure, and time-aware query expansion for refining the search scope. Experiment results show that these optimizations greatly improve both memory recall and downstream question answering on LongMemEval. Overall, our study provides valuable resources and guidance for advancing the long-term memory capabilities of LLM-based chat assistants, paving the way toward more personalized and reliable conversational AI.", "keywords": [ "long-term memory", "retrieval-augmented generation" ], "rating_list": [ 3, 6, 8, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "pZISppZSTv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Amit Haim Bermano", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Associate Professor" }, { "name": "Daniele Reda", "gender": "Male", "institution": "Wayve", "country": "CA", "position": "Researcher" }, { "name": "Guy Tevet", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Michiel van de Panne", "gender": "Male", "institution": "Department of Computer Science, University of British Columbia", "country": "CA", "position": "Full Professor" }, { "name": "Setareh Cohan", "gender": "Female", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Sigal Raab", "gender": "Female", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Xue Bin Peng", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "Assistant Professor" }, { "name": "Zhengyi Luo", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 51, "n_ref": 109, "n_ref_all": 128, "n_fig": 5, "n_tab": 3, "L_tab": 643, "n_element_tab": 61, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 659, "n_element_tab_1": 59, "formula_len_all": 4212, "formula_len_all_1": 4212, "len_all": 151519, "len_all_1": 79974, "len_abs": 1925, "len_title": 137, "len_sents": 38992, "len_sents_1": 34709, "n_sents": 298, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 89, "L_abs": 1258, "title": "CLoSD: Closing the Loop between Simulation and Diffusion for multi-task character control", "abs": "Motion diffusion models and Reinforcement Learning (RL) based control for physics-based simulations have complementary strengths for human motion generation. The former is capable of generating a wide variety of motions, adhering to intuitive control such as text, while the latter offers physically plausible motion and direct interaction with the environment. In this work, we present a method that combines their respective strengths. CLoSD is a text-driven RL physics-based controller, guided by diffusion generation for various tasks. Our key insight is that motion diffusion can serve as an on-the-fly universal planner for a robust RL controller. To this end, CLoSD maintains a closed-loop interaction between two modules \u2014 a Diffusion Planner (DiP), and a tracking controller. DiP is a fast-responding autoregressive diffusion model, controlled by textual prompts and target locations, and the controller is a simple and robust motion imitator that continuously receives motion plans from DiP and provides feedback from the environment. CLoSD is capable of seamlessly performing a sequence of different tasks, including navigation to a goal location, striking an object with a hand or foot as specified in a text prompt, sitting down, and getting up.", "keywords": [ "RL", "PPO", "motion", "motion generation", "motion synthesis", "synthesis", "generative models", "diffusion", "animation" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "pXlmOmlHJZ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Andrew Lee", "gender": "unknown", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Core Francisco Park", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Ekdeep Singh Lubana", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Hidenori Tanaka", "gender": "unknown", "institution": "Harvard University, Harvard University", "country": "US", "position": "Researcher" }, { "name": "Kento Nishi", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Undergrad student" }, { "name": "Martin Wattenberg", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Maya Okawa", "gender": "unknown", "institution": "NTT Research, Inc.", "country": "", "position": "Researcher" }, { "name": "Yang Yongyi", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 9, "n_ref_uni": 32, "n_ref": 62, "n_ref_all": 107, "n_fig": 22, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 16, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1469, "formula_len_all_1": 436, "len_all": 152876, "len_all_1": 61911, "len_abs": 1800, "len_title": 86, "len_sents": 48187, "len_sents_1": 29877, "n_sents": 383, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 38, "L_abs": 1805, "title": "In-Context Learning of Representations", "abs": "Recent work demonstrates that structured patterns in pretraining data influence how representations of different concepts are organized in a large language model\u2019s (LLM) internals, with such representations then driving downstream abilities. Given the open-ended nature of LLMs, e.g., their ability to in-context learn novel tasks, we ask whether models can flexibly alter their semantically grounded organization of concepts. Specifically, if we provide in-context exemplars wherein a concept plays a different role than what the pretraining data suggests, can models infer these novel semantics and reorganize representations in accordance with them? To answer this question, we define a toy \u201cgraph tracing\u201d task wherein the nodes of the graph are referenced via concepts seen during training (e.g., apple, bird, etc.), and the connectivity of the graph is defined via some predefined structure (e.g., a square grid). Given exemplars that indicate traces of random walks on the graph, we analyze intermediate representations of the model and find that as the amount of context is scaled, there is a sudden re-organization of representations according to the graph\u2019s structure. Further, we find that when reference concepts have correlations in their semantics (e.g., Monday, Tuesday, etc.), the context-specified graph structure is still present in the representations, but is unable to dominate the pretrained structure. To explain these results, we analogize our task to energy minimization for a predefined graph topology, which shows getting non-trivial performance on the task requires for the model to infer a connected component. Overall, our findings indicate context-size may be an underappreciated scaling axis that can flexibly re-organize model representations, unlocking novel capabilities.", "keywords": [ "In-Context Learning", "Representational Geometry", "Emergence", "Percolation" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 4, 4 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "pXUAiJshdh", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Huajun Chen", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Jianhua Yao", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Principal Researcher" }, { "name": "Kehua Feng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Keyan Ding", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Ming Qin", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Qiang Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Weijie Wang", "gender": "Male", "institution": "College of Computer Science and Technology, Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Xiang Zhuang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Yu Zhao", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Zeyuan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 61, "n_ref_all": 79, "n_fig": 3, "n_tab": 18, "L_tab": 39550, "n_element_tab": 2281, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 6039, "n_element_tab_1": 420, "formula_len_all": 1200, "formula_len_all_1": 0, "len_all": 314688, "len_all_1": 65889, "len_abs": 1351, "len_title": 130, "len_sents": 81045, "len_sents_1": 26653, "n_sents": 646, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1356, "title": "SciKnowEval: Evaluating Multi-level Scientific Knowledge of Large Language Models", "abs": "Large language models (LLMs) have gained increasing prominence in scientific research, but there is a lack of comprehensive benchmarks to fully evaluate their proficiency in understanding and mastering scientific knowledge.To address this need, we introduce the SciKnowEval benchmark, a novel framework that systematically evaluates LLMs across five progressive levels of scientific knowledge: studying extensively, inquiring earnestly, thinking profoundly, discerning clearly, and practicing assiduously. These levels aim to assess the breadth and depth of scientific knowledge in LLMs, including memory, comprehension, reasoning, discernment, and application. Specifically, we first construct a large-scale evaluation dataset encompassing 70K multi-level scientific problems and solutions in the domains of biology, chemistry, physics, and materials science. By leveraging this dataset, we benchmark 26 advanced open-source and proprietary LLMs using zero-shot and few-shot prompting strategies. The results reveal that despite the state-of-the-art performance of proprietary LLMs, there is still significant room for improvement, particularly in addressing scientific reasoning and applications. We anticipate that SciKnowEval will establish a standard for benchmarking LLMs in science research and promote the development of stronger scientific LLMs.", "keywords": [ "LLM benchmark", "Scientific knowledge evaluation" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "pXPIQsV1St", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Jascha Achterberg", "gender": "Male", "institution": "Advanced Research and Invention Agency", "country": "GB", "position": "Postdoc" }, { "name": "Micha\u0142 J W\u00f3jcik", "gender": "Not Specified", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Quentin Guilhot", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Researcher" }, { "name": "Rui Ponte Costa", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Group Leader" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 41, "n_ref": 79, "n_ref_all": 123, "n_fig": 23, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 24, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 128312, "len_all_1": 71239, "len_abs": 1817, "len_title": 132, "len_sents": 40897, "len_sents_1": 35068, "n_sents": 283, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1825, "title": "Dynamical Similarity Analysis uniquely captures how computations develop in RNNs", "abs": "Methods for analyzing representations in neural systems have become a popular tool in both neuroscience and mechanistic interpretability. Having measures to compare how similar activations of neurons are across conditions, architectures, and species, gives us a scalable way of learning how information is transformed within different neural networks. In contrast to this trend, recent investigations have revealed how some metrics can respond to spurious signals and hence give misleading results. To identify the most reliable metric and understand how measures could be improved, it is going to be important to identify specific test cases which can serve as benchmarks. Here we propose that the phenomena of compositional learning in recurrent neural networks (RNNs) allows us to build a test case for dynamical representation alignment metrics. By implementing this case, we show it enables us to test whether metrics can identify representations which gradually develop throughout learning and probe whether representations identified by metrics are relevant to computations executed by networks. By building both an attractor- and RNN-based test case, we show that the new Dynamical Similarity Analysis (DSA) is more noise robust and identifies behaviorally relevant representations more reliably than prior metrics (Procrustes, CKA). We also show how test cases can be used beyond evaluating metrics to study new architectures. Specifically, results from applying DSA to modern (Mamba) state space models, suggest that, in contrast to RNNs, these models may not exhibit changes to their recurrent dynamics due to their expressiveness. Overall, by developing test cases, we show DSA's exceptional ability to detect compositional dynamical motifs, thereby enhancing our understanding of how computations unfold in RNNs.", "keywords": [ "dynamic representations", "representations", "RNN", "SSM", "neuroscience", "interpretability" ], "rating_list": [ 3, 8, 5, 5 ], "soundness_list": [ 1, 3, 3, 4 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "pXN8T5RwNN", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Sihong Xie", "gender": "Male", "institution": "HKUST-GZ", "country": "CN", "position": "Associate Professor" }, { "name": "Yazheng Liu", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 71, "n_formula_1": 31, "n_ref_uni": 20, "n_ref": 41, "n_ref_all": 102, "n_fig": 16, "n_tab": 6, "L_tab": 2251, "n_element_tab": 338, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 77, "n_element_tab_1": 6, "formula_len_all": 11094, "formula_len_all_1": 4517, "len_all": 174188, "len_all_1": 86861, "len_abs": 1247, "len_title": 112, "len_sents": 53517, "len_sents_1": 34691, "n_sents": 487, "n_sents_1": 319, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1253, "title": "Explanations of GNN on Evolving Graphs via Axiomatic Layer edges", "abs": "Graphs are ubiquitous in social networks, chemical molecules, and financial data, where Graph Neural Networks (GNNs) achieve superior predictive accuracy. Graphs can be\nevolving, while understanding how GNN predictions respond to the evolution provides significant insight and trust. \nWe explore the problem of explaining evolving GNN predictions due to continuously changing edge weights.\nWe first propose a layer edge-based explanation to balance\nexplanation fidelity and interpretability, as opposed to message flow and input edge.\nThen we propose a novel framework to address the challenges of axiomatic attribution and the entanglement of multiple computational graph paths due to continuous change of edge weights. We first design an axiomatic attribution of the evolution of the model prediction to message flows, then develop Shapley value to fairly map message flow contributions to layer edges.\nWe formulate a novel optimization problem to find the critical layer edges based on KL-divergence minimization. Extensive experiments on eight datasets for node classification, link prediction, and graph classification tasks with evolving graphs demonstrate the better fidelity and interpretability of the proposed method over the baseline methods.", "keywords": [ "explainability; dynamic graphs; message flows; layeredges" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "pXIbcRPxWR", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Dujian Ding", "gender": "unknown", "institution": "Computing Science, University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Xiang Zhang", "gender": "Male", "institution": "Patsnap", "country": "SG", "position": "Intern" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 29, "n_ref": 70, "n_ref_all": 103, "n_fig": 9, "n_tab": 1, "L_tab": 825, "n_element_tab": 136, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 578, "n_element_tab_1": 132, "formula_len_all": 287, "formula_len_all_1": 288, "len_all": 122008, "len_all_1": 68718, "len_abs": 2126, "len_title": 75, "len_sents": 41416, "len_sents_1": 32626, "n_sents": 283, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 27, "L_abs": 1994, "title": "Supervised Chain of Thought", "abs": "Large Language Models (LLMs) have revolutionized the field of natural language processing and hold significant promise for advancements in Artificial Intelligence. However, the backbone architecture of most mainstream LLMs, the Transformer, has inherent limitations regarding computational depth, making them theoretically incapable of solving many reasoning tasks that require increasing depth. Chain of Thought (CoT) techniques, however, have been shown to mitigate these architectural limitations, as demonstrated by several theoretical works, offering a viable approach to solving complex reasoning tasks that were previously out of reach.\nDespite its successes, CoT and its variants (such as Tree of Thought, Graph of Thought, etc.) follow a one-prompt-for-all-tasks approach. Specifically, they rely on a single prompt structure (e.g., \"think step by step\") for a wide range of tasks, from counting to sorting, and from solving mathematical problems to tackling algorithmic challenges. This creates significant challenges for the model to generate the correct steps template for different tasks, as it requires searching in large prompt template space.\nIn this work, we build on previous theoretical analyses of CoT to demonstrate how the \"one-prompt-for-all-tasks\" template can negatively impact the computability of LLMs. We divide the solution space into prompt space and answer space, showing that the CoT process requires task-specific supervision to accurately navigate the prompt space and achieve optimal performance. Through experiments with the latest LLMs, we reveal a significant gap in reasoning ability when supervision is applied versus when it is not.\nOur aim is to provide insights into the mechanisms behind CoT and to inspire the effective design of CoT variants. Additionally, we highlight the key limitations of traditional ``unsupervised'' prompting approaches, suggesting the need for more nuanced, task-specific \"supervised\" CoT for effective reasoning with LLMs.", "keywords": [ "Chain of Thought", "LLMs", "Prompting", "CoT", "Model Analysis" ], "rating_list": [ 1, 1, 5, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "pWrcpPsVas", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Evangelos E. Papalexakis", "gender": "Male", "institution": "University of California, Riverside", "country": "US", "position": "Associate Professor" }, { "name": "Jonathan Richardson", "gender": "Male", "institution": "University of California, Riverside", "country": "US", "position": "Assistant Professor" }, { "name": "Pooyan Goodarzi", "gender": "Male", "institution": "University of California, Riverside", "country": "US", "position": "PhD student" }, { "name": "Sidharth Kannan", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "Undergrad student" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 15, "n_ref": 20, "n_ref_all": 31, "n_fig": 4, "n_tab": 7, "L_tab": 845, "n_element_tab": 132, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 845, "n_element_tab_1": 132, "formula_len_all": 523, "formula_len_all_1": 523, "len_all": 76739, "len_all_1": 53900, "len_abs": 830, "len_title": 52, "len_sents": 27852, "len_sents_1": 24596, "n_sents": 199, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 834, "title": "Graph Neural Networks for Interferometer Simulations", "abs": "In recent years, graph neural networks (GNNs) have shown tremendous promise in solving problems in high energy physics, materials science, and fluid dynamics. In this work, we introduce a new application for GNNs in the physical sciences: instrumentation design. As a case study, we apply GNNs to simulate models of the Laser Interferometer Gravitational-wave Observatory (LIGO), and show that they are capable of accurately capturing the complex optical physics at play, while achieving runtimes 815 times faster than state of the art simulation packages. We discuss the unique challenges this problem provides for machine learning models. In addition, we provide a dataset of high-fidelity optical physics simulations for three interferometer topologies, which can be used as a benchmarking suite for future work in this direction.", "keywords": [ "graph neural networks", "physics simulation", "LIGO" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 1, 3, 3, 4 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "pWrCiFpm3L", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Daniel Neider", "gender": "unknown", "institution": "Technische Universit\u00e4t Dortmund", "country": "DE", "position": "Full Professor" }, { "name": "Faried Abu Zaid", "gender": "Male", "institution": "AppliedAI Institute", "country": "DE", "position": "Researcher" }, { "name": "Mustafa Yal\u00e7\u0131ner", "gender": "Male", "institution": "Technische Universit\u00e4t Dortmund", "country": "DE", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 7, "n_ref_uni": 49, "n_ref": 80, "n_ref_all": 111, "n_fig": 11, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2000, "formula_len_all_1": 659, "len_all": 176578, "len_all_1": 69078, "len_abs": 1363, "len_title": 112, "len_sents": 56801, "len_sents_1": 34682, "n_sents": 436, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1407, "title": "VeriFlow: Modeling Distributions for Neural Network Verification", "abs": "Formal verification has emerged as a promising method to ensure the safety and reliability of neural networks.\nNaively verifying a safety property amounts to ensuring the safety of a neural network for the whole input space irrespective of any training or test set.\nHowever, this also implies that the safety of the neural network is checked even for inputs that do not occur in the real-world and have no meaning at all, often resulting in spurious errors.\nTo tackle this shortcoming, we propose the VeriFlow architecture as a flow based density model tailored to allow any verification approach to restrict its search to the some data distribution of interest.\nWe argue that our architecture is particularly well suited for this purpose because of two major properties. \nFirst, we show that the transformation and log-density function that are defined by our model are piece-wise affine. Therefore, the model allows the usage of verifiers based on SMT with linear arithmetic.\nSecond, upper density level sets (UDL) of the data distribution take the shape of an $L^p$-ball in the latent space. As a consequence, representations of UDLs specified by a given probability are effectively computable in latent space. This allows the use of SMT and abstract interpretation approaches with fine-grained, probabilistically interpretable, control regarding on how (a)typical the inputs subject to verification are.", "keywords": [ "neuro-symbolic verification", "normalizing flows", "density level sets", "neural network verification" ], "rating_list": [ 5, 6, 5, 8 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 2, 2, 1, 4 ], "contribution_list": [ 3, 4, 2, 4 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "pWdkM9NNCA", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 12, "author_info_list": [ { "name": "DONG XU", "gender": "unknown", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Hu Haifeng", "gender": "Male", "institution": "Tencent AI Platform", "country": "", "position": "Researcher" }, { "name": "Kai Xu", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Undergrad student" }, { "name": "Lichao Zhang", "gender": "unknown", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Xutongni", "gender": "unknown", "institution": "Zhejiang Lab, Zhejiang Lab", "country": "CN", "position": "Undergrad student" }, { "name": "Yi Gao", "gender": "unknown", "institution": "South China University of Technology", "country": "CN", "position": "Researcher" }, { "name": "Yuhao Luo", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "MS student" }, { "name": "Zhongqian Sun", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "bridgettesong", "gender": "unknown", "institution": "Tianjin University", "country": "CN", "position": "Beijing" }, { "name": "weishandeng", "gender": "unknown", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "fuming you", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 30, "n_ref": 77, "n_ref_all": 87, "n_fig": 2, "n_tab": 4, "L_tab": 711, "n_element_tab": 114, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 711, "n_element_tab_1": 114, "formula_len_all": 574, "formula_len_all_1": 576, "len_all": 118414, "len_all_1": 74748, "len_abs": 1495, "len_title": 75, "len_sents": 37161, "len_sents_1": 36590, "n_sents": 254, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1513, "title": "Fox-TTS: Scalable Flow Transformers for Expressive Zero-Shot Text to Speech", "abs": "Expressive zero-shot text-to-speech (TTS) synthesis aims at synthesizing high-fidelity speech that closely mimics a brief stylized recording without additional training. Despite the advancements in this area, several challenges persist: 1) Current methods, which encompass implicit prompt engineering through in-context learning or by using pre-trained speaker identification models, often struggle to fully capture the acoustic characteristics of the stylized speaker; 2) Attaining high-fidelity voice cloning for a stylized speaker typically requires large amounts of specific data for fine-tuning; 3) There is no benchmark tailored for the expressive zero-shot TTS scenarios. To address them, we present *Fox-TTS*, a family of large-scale models for high-quality expressive zero-shot TTS. We introduce an improved flow-matching Transformer model coupled with a novel learnable speaker encoder. Within the speaker encoder, we incorporate three key designs: temporal mean pooling, temporal data augmentation, and an information bottleneck used for trading off pronunciation stability and speaker similarity in an explainable manner. Moreover, we have collected \\textit{Fox-eval}, the first multi-speaker, multi-style benchmark that is specially designed for expressive zero-shot scenarios. Extensive experiments show that Fox-TTS achieves on-par quality with human recordings in normal scenarios and state-of-the-art performance in expressive scenarios. Audio samples are available at https://fox-tts.github.io/.", "keywords": [ "Expressive Text-to-Speech", "flow-matching models", "zero-shot generalization" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 1, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "pWdUcV5axb", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Hongbin Liu", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Yuepeng Hu", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Zedian Shao", "gender": "Male", "institution": "Duke University", "country": "US", "position": "MS student" }, { "name": "Zhongye Liu", "gender": "Male", "institution": "Duke University, Duke University", "country": "US", "position": "MS student" }, { "name": "Neil Gong", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 26, "n_ref": 86, "n_ref_all": 104, "n_fig": 4, "n_tab": 8, "L_tab": 4759, "n_element_tab": 570, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 4359, "n_element_tab_1": 540, "formula_len_all": 369, "formula_len_all_1": 369, "len_all": 135293, "len_all_1": 79235, "len_abs": 1666, "len_title": 141, "len_sents": 37324, "len_sents_1": 31590, "n_sents": 283, "n_sents_1": 236, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1675, "title": "Automatically Generating Visual Hallucination Test Cases for Multimodal Large Language Models", "abs": "Visual hallucination (VH) occurs when a multimodal large language model (MLLM) generates responses with incorrect visual details for prompts. Existing methods for generating VH test cases primarily rely on human annotations, typically in the form of triples: (image, question, answer). In this paper, we introduce VHExpansion, the first automated method for expanding VH test cases for MLLMs. Given an initial VH test case, VHExpansion automatically expands it by perturbing the question and answer through negation as well as modifying the image using both common and adversarial perturbations. Additionally, we propose a new evaluation metric, symmetric accuracy, which measures the proportion of correctly answered VH test-case pairs. Each pair consists of a test case and its negated counterpart. Our theoretical analysis shows that symmetric accuracy is an unbiased evaluation metric that remains unaffected by the imbalance of VH testing cases with varying answers when an MLLM is randomly guessing the answers, whereas traditional accuracy is prone to such imbalance. We apply VHExpansion to expand three VH datasets annotated manually and use these expanded datasets to benchmark seven MLLMs. Our evaluation shows that VHExpansion effectively identifies more VH test cases. Moreover, symmetric accuracy, being unbiased, leads to different conclusions about the vulnerability of MLLMs to VH compared to traditional accuracy metric. Finally, we show that fine-tuning MLLMs on the expanded VH dataset generated by VHExpansion mitigates VH more effectively than fine-tuning on the original, manually annotated dataset. We will publish code and data upon paper acceptance.", "keywords": [ "Hallucination", "multimodal large language models" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "pW387D5OUN", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Arno Solin", "gender": "unknown", "institution": "Aalto University", "country": "FI", "position": "Assistant Professor" }, { "name": "Marcus Klasson", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Postdoc" }, { "name": "Martin Trapp", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Postdoc" }, { "name": "Rui Li", "gender": "Female", "institution": "Aalto University", "country": "FI", "position": "PhD student" } ], "n_formula": 79, "n_formula_1": 21, "n_ref_uni": 53, "n_ref": 114, "n_ref_all": 158, "n_fig": 13, "n_tab": 16, "L_tab": 9078, "n_element_tab": 409, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3552, "n_element_tab_1": 92, "formula_len_all": 10979, "formula_len_all_1": 1921, "len_all": 222708, "len_all_1": 73417, "len_abs": 733, "len_title": 97, "len_sents": 52856, "len_sents_1": 29234, "n_sents": 449, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 35, "L_abs": 736, "title": "Streamlining Bayesian Deep Learning", "abs": "The rising interest in Bayesian deep learning (BDL) has led to a plethora of methods for estimating the posterior distribution. However, efficient computation of inferences, such as predictions, has been largely overlooked with Monte Carlo integration remaining the standard. In this work we examine streamlining prediction in BDL through a single forward pass without sampling. For this we use local linearisation on activation functions and local Gaussian approximations at linear layers. Thus allowing us to analytically compute an approximation to the posterior predictive distribution. We showcase our approach for both MLP and transformers, such as ViT and GPT-2, and assess its performance on regression and classification tasks.", "keywords": [ "Bayesian deep learning", "uncertainty quantification" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 4, 4, 3, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 4, 3, 1, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "pVL4bYKOGM", "primary_area": "causal reasoning", "n_author": 6, "author_info_list": [ { "name": "Dennis Frauen", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Jonas Schweisthal", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Konstantin Hess", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Maresa Schr\u00f6der", "gender": "unknown", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Stefan Feuerriegel", "gender": "Male", "institution": "LMU Munich", "country": "DE", "position": "Professor" }, { "name": "Valentyn Melnychuk", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" } ], "n_formula": 107, "n_formula_1": 29, "n_ref_uni": 38, "n_ref": 106, "n_ref_all": 129, "n_fig": 10, "n_tab": 4, "L_tab": 1482, "n_element_tab": 143, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 99, "n_element_tab_1": 17, "formula_len_all": 9475, "formula_len_all_1": 2074, "len_all": 209367, "len_all_1": 72863, "len_abs": 2862, "len_title": 112, "len_sents": 61230, "len_sents_1": 31944, "n_sents": 568, "n_sents_1": 275, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1224, "title": "Conformal prediction for causal effects of continuous treatments", "abs": "Uncertainty quantification of causal effects is crucial for safety-critical applications such as personalized medicine. A powerful approach for this is conformal prediction, which has several practical benefits due to model-agnostic finite-sample guarantees. Yet, existing methods for conformal prediction of causal effects are limited to binary/discrete treatments and make highly restrictive assumptions such as known propensity scores. In this work, we provide a novel conformal prediction method for potential outcomes of continuous treatments. We account for the additional uncertainty introduced through propensity estimation so that our conformal prediction intervals are valid even if the propensity score is unknown. Our contributions are three-fold: (1) We derive finite-sample prediction intervals for potential outcomes of continuous treatments. (2) We provide an algorithm for calculating the derived intervals. (3) We demonstrate the effectiveness of the conformal prediction intervals in experiments on synthetic and medical datasets. To the best of our knowledge, we are the first to propose conformal prediction for continuous treatments when the propensity score is unknown and must be estimated from data.", "keywords": [ "causality", "dosage response curves", "conformal prediction", "uncertainty quantification" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "pV0SUV8cRF", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Hadi Amiri", "gender": "Not Specified", "institution": "University of Massachusetts Lowell", "country": "US", "position": "Assistant Professor" }, { "name": "Jiali Cheng", "gender": "Male", "institution": "University of Massachusetts Lowell", "country": "US", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 60, "n_ref": 118, "n_ref_all": 129, "n_fig": 4, "n_tab": 4, "L_tab": 1543, "n_element_tab": 143, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2264, "n_element_tab_1": 133, "formula_len_all": 684, "formula_len_all_1": 684, "len_all": 178488, "len_all_1": 69408, "len_abs": 995, "len_title": 87, "len_sents": 35297, "len_sents_1": 31445, "n_sents": 237, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1142, "title": "Tool Unlearning for Tool Augmented LLMs", "abs": "Tool-augmented large language models (LLMs) may need to forget learned tools due to security concerns, privacy restrictions, or deprecated tools. However, unlearning tool has not been explored in prior machine unlearning works. \n\nWe propose tool unlearning, a novel machine unlearning task that deletes already acquired tools. Compared to traditional unlearning, tool unlearning exhibits certain differences and difficulties: 1) knowledge removal instead of forgetting samples, 2) significant cost of optimizing LLMs, 3) lack of principled evaluation tools. \n\nTo bridge this gap, we introduce three properties for effective tool unlearning and propose ToolDelete, the first unlearning method designed for tool-augmented LLMs. We also propose the first membership inference attack (MIA) model for evaluating tool unlearning.\n\nExperiments on three tool learning datasets and tool-augmented LLMs demonstrate that ToolDelete effectively unlearns both randomly selected tools and tools from specific categories. The unlearning behavior does not impact the LLM's knowledge on non-deleted tools, while preserving performances on other general tasks.", "keywords": [ "Tool-Augmented LLM", "Machine Unlearning", "Tool Unlearning" ], "rating_list": [ 5, 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 3, 4 ], "contribution_list": [ 2, 3, 2, 3, 1 ], "confidence_list": [ 3, 3, 4, 3, 2 ] }, { "paper_id": "pUbbLHjCPM", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 12, "author_info_list": [ { "name": "Aili Chen", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Chengyu Du", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Haokun Zhao", "gender": "Male", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Haoran Guo", "gender": "unknown", "institution": "\u7814\u8335\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Jiaqing Liang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jinyi Han", "gender": "unknown", "institution": "East China Normal University", "country": "CN", "position": "PhD student" }, { "name": "Liangyue Li", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Qianyu He", "gender": "Female", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Sirui Xia", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Yanghua Xiao", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yizhou Ying", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Zulong Chen", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Principal Researcher" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 39, "n_ref": 89, "n_ref_all": 104, "n_fig": 6, "n_tab": 10, "L_tab": 5236, "n_element_tab": 348, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 2664, "n_element_tab_1": 207, "formula_len_all": 384, "formula_len_all_1": 316, "len_all": 223465, "len_all_1": 67069, "len_abs": 1632, "len_title": 85, "len_sents": 73044, "len_sents_1": 29855, "n_sents": 581, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1653, "title": "Think Thrice Before You Act: Progressive Thought Refinement in Large Language Models", "abs": "Recent advancements in large language models (LLMs) have demonstrated that progressive refinement, rather than providing a single answer, results in more accurate and thoughtful outputs. However, existing methods often rely heavily on supervision signals to evaluate previous responses, making it difficult to effectively assess output quality in more open-ended scenarios. Additionally, these methods are typically designed for specific tasks, which limits their generalization to new domains. To address these limitations, we propose Progressive Thought Refinement (PTR), a framework that enables LLMs to progressively refine their responses. PTR operates in two phases: (1) Thought data construction stage: We propose a \\textit{weak and strong model collaborative selection} strategy to build a high-quality progressive refinement dataset to ensure logical consistency from thought to answers, and the answers are gradually refined in each round. (2) Thought-Mask Fine-Tuning Phase: We design a training structure to mask the \"thought\" and adjust loss weights to encourage LLMs to refine prior thought, teaching them to implicitly understand \"how to improve\" rather than \"what is correct.\" Experimental results show that PTR significantly enhances LLM performance across ten diverse tasks (avg. from 49.6\\% to 54.48\\%) without task-specific fine-tuning. Notably, in more open-ended tasks, LLMs also demonstrate substantial improvements in the quality of responses beyond mere accuracy, suggesting that PTR truly teaches LLMs to self-improve over time. Our project's source code and datasets are available at https://anonymous.4open.science/r/PTR_LLM", "keywords": [ "progressive thought refinement", "self refine", "large language model" ], "rating_list": [ 6, 3, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "pTyEnkuSQ0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Amir Nassereldine", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Changjae Lee", "gender": "Male", "institution": ", State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Chenhui Xu", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Dancheng Liu", "gender": "Male", "institution": "University at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Jiajie Li", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Jinjun Xiong", "gender": "unknown", "institution": "State University of New York at Buffalo", "country": "US", "position": "Professor" }, { "name": "Ruiyang Qin", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Utkarsh Kumar", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Yiyu Shi", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "Full Professor" }, { "name": "Yuting Hu", "gender": "Female", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Ziming Yang", "gender": "Male", "institution": ", State University of New York at Buffalo", "country": "", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 47, "n_ref": 109, "n_ref_all": 143, "n_fig": 3, "n_tab": 16, "L_tab": 5168, "n_element_tab": 711, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 230, "n_element_tab_1": 15, "formula_len_all": 988, "formula_len_all_1": 734, "len_all": 245304, "len_all_1": 67775, "len_abs": 1224, "len_title": 108, "len_sents": 73132, "len_sents_1": 32863, "n_sents": 585, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1230, "title": "Large Language Models have Intrinsic Self-Correction Ability", "abs": "Large language models (LLMs) have attracted significant attention for their exceptional abilities in various natural language processing tasks, but they suffer from hallucinations that will cause performance degradation. One promising solution to improve the LLMs' performance is to ask LLMs to revise their answer after generation, a technique known as self-correction. Among the two types of self-correction, intrinsic self-correction is considered a promising direction because it does not utilize external knowledge. However, recent works doubt the validity of LLM's ability to conduct intrinsic self-correction. In this paper, we present a novel perspective on the intrinsic self-correction capabilities of LLMs through theoretical analyses and empirical experiments. In addition, we identify two critical factors for successful self-correction: zero temperature and fair prompts. Leveraging these factors, we demonstrate that intrinsic self-correction ability is exhibited across multiple existing LLMs. Our findings offer insights into the fundamental theories underlying the self-correction behavior of LLMs and remark on the importance of unbiased prompts and zero temperature settings in harnessing their full potential.", "keywords": [ "large language model", "self correction", "prompt engineering", "foundation model" ], "rating_list": [ 3, 5, 3, 10 ], "soundness_list": [ 1, 3, 1, 4 ], "presentation_list": [ 1, 3, 1, 3 ], "contribution_list": [ 1, 1, 1, 4 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "pTsP30MoBq", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Grace Yi", "gender": "Female", "institution": "University of Western Ontario", "country": "CA", "position": "Full Professor" }, { "name": "Pingbo Hu", "gender": "Male", "institution": "University of Western Ontario", "country": "", "position": "PhD student" } ], "n_formula": 71, "n_formula_1": 35, "n_ref_uni": 43, "n_ref": 56, "n_ref_all": 80, "n_fig": 0, "n_tab": 5, "L_tab": 2789, "n_element_tab": 145, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 418, "n_element_tab_1": 74, "formula_len_all": 6309, "formula_len_all_1": 1952, "len_all": 144628, "len_all_1": 59051, "len_abs": 1222, "len_title": 139, "len_sents": 40986, "len_sents_1": 26704, "n_sents": 307, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1229, "title": "Mitigating Input Noise in Binary Classification: A Unified Framework with Data Augmentation", "abs": "Classification techniques have achieved significant success across fields such as computer vision, information retrieval, and natural language processing. However, much of this progress assumes input features are error-free -- a condition rarely met in practice. In real-world scenarios, noisy inputs caused by measurement errors are common, leading to biased or suboptimal classification results. This paper presents a unified framework for binary classification with noisy inputs, offering a generalizable solution that applies across various supervised learning algorithms and noise models. We provide a theoretical analysis of the bias introduced by ignoring input noise (also referred to as feature corruption) and identify conditions where this bias can be safely disregarded. To address cases where noise correction is needed, we propose a novel data augmentation-based method to mitigate input noise effects. Our approach is both comprehensive and theoretically grounded, providing practical solutions for improving classification accuracy in noisy data enviroments. Extensive experiments, including analyses of medical image datasets, demonstrate the superior performance of our methods under different noise conditions.", "keywords": [ "classification", "noisy input", "noisy attribute", "noisy feature", "mismeasured input", "measurement error", "supervised learning" ], "rating_list": [ 6, 3, 6, 3, 3 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 1, 2 ], "contribution_list": [ 2, 2, 3, 1, 1 ], "confidence_list": [ 2, 3, 3, 3, 4 ] }, { "paper_id": "pTqsapWIPo", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Benjamin E Lundell", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Harpreet S. Sawhney", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Jack Saunders", "gender": "Male", "institution": "University of Bath", "country": "GB", "position": "PhD student" }, { "name": "Nicholas Alexander Gyde", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Patricio Vela", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Yanan Jian", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Yiye Chen", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 55, "n_ref": 91, "n_ref_all": 112, "n_fig": 8, "n_tab": 3, "L_tab": 245, "n_element_tab": 8, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1076, "n_element_tab_1": 71, "formula_len_all": 204, "formula_len_all_1": 85, "len_all": 182698, "len_all_1": 67434, "len_abs": 1790, "len_title": 78, "len_sents": 47303, "len_sents_1": 31598, "n_sents": 407, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1819, "title": "$\\textit{RwR}$: A Reason-while-Retrieve framework for Reasoning on Scene Graphs with LLMs", "abs": "Large Language Models (LLMs) have demonstrated impressive reasoning and planning capacities, yet grounding these abilities to a specific environment remains challenging. Recently, there has been a growing interest in representing environments as scene graphs for LLMs, due to their serializable format, scalability to large environments, and flexibility in incorporating diverse semantic and spatial information for various downstream tasks. \nDespite the success of prompting graphs as text, existing methods suffer from hallucinations with large graph inputs and limitation in solving complex spatial problems, restricting their application beyond simple object search tasks.\n In this work, we explore grounding LLM reasoning in the environment through the $\\textit{scene graph schema}$. \nWe propose $\\textit{SG-RwR}$, an iterative reason-while-retrieve scene graph reasoning framework involving two cooperative schema-guided code-writing LLMs: a (1) $\\textit{Reasoner}$ for task planning and information querying, and a (2) $\\textit{Retriever}$ for extracting graph information based on these queries. \n This cooperation facilitates focused attention on task-relevant graph information and enables sequential reasoning on the graph essential for complex tasks.\n Additionally, the code-writing design allows for the use of tools to solve problems beyond the capacity of LLMs, which further enhance its reasoning ability on scene graphs. \n We also demonstrate that our framework can benefit from task-level few-shot examples, even in the absence of agent-level demonstrations,\n thereby enabling in-context learning without data collection overhead.\n Through experiments in multiple simulation environments, we show that $\\textit{RwR}$ surpasses existing LLM-based approaches in numerical Q\\&A and planning tasks.", "keywords": [ "Large Language Models; Scene Graphs;" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "pTeOOKnjGM", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Ailing Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Lei Zhu", "gender": "Female", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Lijian Lin", "gender": "Female", "institution": "International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Ye Zhu", "gender": "Male", "institution": "Tencent PCG ARC Lab", "country": "CN", "position": "Researcher" }, { "name": "Yu Li", "gender": "Male", "institution": "International Digital Economy Academy", "country": "CN", "position": "Principal Researcher" }, { "name": "Yunfei Liu", "gender": "Male", "institution": "International Digital Economy Academy (IDEA)", "country": "CN", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 18, "n_ref_uni": 35, "n_ref": 54, "n_ref_all": 75, "n_fig": 10, "n_tab": 6, "L_tab": 1767, "n_element_tab": 258, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 765, "n_element_tab_1": 86, "formula_len_all": 835, "formula_len_all_1": 870, "len_all": 155414, "len_all_1": 52634, "len_abs": 600, "len_title": 118, "len_sents": 41759, "len_sents_1": 24235, "n_sents": 338, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1426, "title": "TEASER: Token Enhanced Spatial Modeling for Expressions Reconstruction", "abs": "3D facial reconstruction from a single in-the-wild image is a crucial task in human-centered computer vision tasks. While existing methods can recover accurate facial shapes, there remains significant space for improvement in fine-grained expression capture. Current approaches struggle with irregular mouth shapes, exaggerated expressions, and asymmetrical facial movements. We present TEASER (Token EnhAnced Spatial modeling for Expressions Reconstruction), which addresses these challenges and enhances 3D facial geometry performance\u2060\u2060. TEASER tackles two main limitations of existing methods: insufficient photometric loss for self-reconstruction and inaccurate localization of subtle expressions. We introduce a multi-scale tokenizer to extract facial appearance information. Combined with a neural renderer, these tokens provide precise geometric guidance for expression reconstruction. Furthermore, TEASER incorporates a pose-dependent landmark loss to further improve geometric performance\u2060. Our approach not only significantly enhances expression reconstruction quality but also offers interpretable tokens suitable for various downstream applications, such as photorealistic facial video driving, expression transfer, and identity swapping. Quantitative and qualitative experimental results across multiple datasets demonstrate that TEASER achieves state-of-the-art performance in precise expression reconstruction.", "keywords": [ "Expression reconstruction", "Hybrid parameters" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "pSdE7PIA64", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Yang Gao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Yinghuan Shi", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Yisen Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ze Peng", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "PhD student" }, { "name": "\u5251 \u5f20", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Postdoc" }, { "name": "Lei Qin", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 167, "n_formula_1": 23, "n_ref_uni": 57, "n_ref": 163, "n_ref_all": 206, "n_fig": 12, "n_tab": 2, "L_tab": 369, "n_element_tab": 25, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 22772, "formula_len_all_1": 1925, "len_all": 309069, "len_all_1": 72420, "len_abs": 2358, "len_title": 130, "len_sents": 101287, "len_sents_1": 33374, "n_sents": 900, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1440, "title": "Leveraging Flatness to Improve Information-Theoretic Generalization Bounds for SGD", "abs": "Information-theoretic generalization bounds have been used to study the generalization of learning algorithms. These bounds are intrinsically data- and algorithm-dependent so that one can exploit the properties of data and algorithm to derive tighter bounds. However, we observe such algorithm dependence is still inadequate in existing information-theoretic bounds for SGD because they have not adequately leveraged the algorithmic bias toward flat minima of SGD. Since the flatness of minima given by SGD is crucial for SGD's generalization, the bounds fail to capture the improved generalization under better flatness and are also numerically loose. This paper derives a more flatness-leveraged information-theoretic bound for the flatness-favoring SGD. The bound indicates that the learned models generalize better if the large-variance directions of the final weight covariance have small local curvatures in the loss landscape. Experiments on deep neural networks show that our bound not only correctly reflects the better generalization when flatness is improved, but is also numerically tighter by being only a few percentages looser. This is achieved by a technique called \"omniscient trajectory.\" When applied to Gradient Descent on convex-Lipschitz-Bounded (CLB) problems, it yields an $O(1/\\sqrt{n})$ minimax rate for excess risks, which has been shown to be impossible for representative existing information-theoretic bounds.", "keywords": [ "information theory; implicit bias; deep learning theory; convex optimization; learning theory" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "pRUxNDrfvk", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Athanasios Psaltis", "gender": "Male", "institution": "CERTH/ITI", "country": "GR", "position": "Postdoc" }, { "name": "Ioannis Lazaridis", "gender": "Male", "institution": "CERTH/ITI", "country": "GR", "position": "Researcher" }, { "name": "Petros Daras", "gender": "Male", "institution": "Information Technologies Institute ", "country": "GR", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 18, "n_ref": 30, "n_ref_all": 36, "n_fig": 2, "n_tab": 5, "L_tab": 4523, "n_element_tab": 1238, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2879, "n_element_tab_1": 594, "formula_len_all": 99, "formula_len_all_1": 99, "len_all": 107363, "len_all_1": 64594, "len_abs": 1100, "len_title": 110, "len_sents": 25078, "len_sents_1": 24263, "n_sents": 203, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1103, "title": "ABAS-RAL: Adaptive BAtch Size using Reinforced Active Learning", "abs": "Active learning reduces annotation costs by selecting the most informative samples, however fixed batch sizes used in traditional methods often lead to inefficient use of resources. We propose Adaptive BAtch Size using Reinforced Active Learning, a novel approach that dynamically adjusts batch sizes based on model uncertainty and performance. By framing the annotation process as a Markov Decision Process, the proposed method employs reinforcement learning to optimize batch size selection, using two distinct policies: one targeting precision and budget, and the other for adapting the batch size based on learning progress. The proposed method is evaluated on both CIFAR-10, CIFAR-100 and MNIST datasets. The performance is measured across multiple metrics, including precision, accuracy, recall, F1-score, and annotation budget. Experimental results demonstrate that the proposed method consistently reduces annotation costs while maintaining or improving performance compared to fixed-batch Active Learning methods, achieving higher sample selection efficiency without compromising model quality.", "keywords": [ "Active learning", "Reinforcement Learning", "Adaptive Batch Size", "Annotation Budget" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 3, 2, 1, 1 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "pRIPRDALBV", "primary_area": "applications to robotics, autonomy, planning", "n_author": 8, "author_info_list": [ { "name": "Ali Pesaranghader", "gender": "Male", "institution": "LG Electronics", "country": "CA", "position": "Researcher" }, { "name": "HANZE LI", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "MS student" }, { "name": "Hyejeong Jeon", "gender": "Not Specified", "institution": "LG Electronics", "country": "KR", "position": "Researcher" }, { "name": "Jaehong Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Punyaphat Sukcharoenchaikul", "gender": "unknown", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" }, { "name": "Scott Sanner", "gender": "Male", "institution": "Department of Mechanical and Industrial Engineering, University of Toronto", "country": "CA", "position": "Associate Professor" }, { "name": "Tanmana Sadhu", "gender": "Female", "institution": "LG Corporation", "country": "CA", "position": "Researcher" }, { "name": "Xiaotian Liu", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 15, "n_ref": 22, "n_ref_all": 33, "n_fig": 4, "n_tab": 4, "L_tab": 584, "n_element_tab": 60, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1294, "n_element_tab_1": 112, "formula_len_all": 378, "formula_len_all_1": 378, "len_all": 123002, "len_all_1": 66126, "len_abs": 1903, "len_title": 136, "len_sents": 39496, "len_sents_1": 29951, "n_sents": 285, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1913, "title": "Open-World Planning via Lifted Regression with LLM-based Affordances for Embodied Agents", "abs": "Open-world planning is crucial for embodied AI agents that must make decisions with incomplete task-relevant knowledge. In fact, the main challenges lie in reasoning about objects and their affordances that are unknown to the agent. Large Language Models (LLMs), pre-trained on vast internet-scale data, have emerged as potential solutions for open-world planning. However, LLMs have limitations in long-horizon planning tasks and face problems related to interpretability, reliability, and cost-efficiency. Symbolic planning methods, on the other hand, offer structured and verifiable approaches to long-horizon tasks, but often struggle to generate feasible plans in an open-world setting. In this work, we propose a novel approach, called LLM-Regress, which combines the strengths of lifted symbolic regression planning with LLM-based affordances. The lifted representation allows us to generate plans capable of handling arbitrary unknown objects, while regression planning is the only planning paradigm that guarantees complete solutions using lifted representations. For such tasks, we leverage LLMs to supplement missing affordances knowledge for unknown objects. The regression nature of our approach enables the agent to focus on actions and objects relevant to the goal, thus avoiding the need for costly LLM calls for every decision. We evaluate our approach on the ALFWorld dataset and introduce a new ALFWorld-Afford dataset with higher planning complexity and more affordances types. The empirical results demonstrate that our method outperforms existing approaches in terms of success rates, planning duration, and number of LLM Tokens. Finally, we show that our approach is resilient to domain shifts in affordances and generalizes effectively to unseen tasks. This work underscores the importance of integrating symbolic reasoning with LLM knowledge for open-world decision-making in embodied AI.", "keywords": [ "Embodied Agents", "Open World Planning" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 4, 2, 4 ] }, { "paper_id": "pRCTRC6icM", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Mohan Shankara Kankanhalli", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Yangyang Guo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 63, "n_ref": 146, "n_ref_all": 174, "n_fig": 6, "n_tab": 9, "L_tab": 4686, "n_element_tab": 654, "n_fig_1": 3, "n_tab_1": 9, "L_tab_1": 3139, "n_element_tab_1": 393, "formula_len_all": 295, "formula_len_all_1": 301, "len_all": 168193, "len_all_1": 72952, "len_abs": 1645, "len_title": 64, "len_sents": 37353, "len_sents_1": 29410, "n_sents": 313, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1535, "title": "SCAN: Bootstrapping Contrastive Pre-training for Data Efficiency", "abs": "While contrastive pre-training is widely employed, its data efficiency problem has remained relatively under-explored thus far. Existing methods often rely on static coreset selection algorithms to pre-identify important data for training. However, this static nature renders them unable to dynamically track the data usefulness throughout pre-training, leading to subpar pre-trained models. To address this challenge, our paper introduces a novel dynamic bootstrapping dataset pruning method. It involves pruning data preparation followed by dataset mutation operations, both of which undergo iterative and dynamic updates. We apply this method to two prevalent contrastive pre-training frameworks: CLIP and MoCo, representing vision-language and vision-centric domains, respectively. In particular, we individually pre-train seven CLIP models on two large-scale image-text pair datasets, and two MoCo models on the ImageNet dataset, resulting in a total of 16 pre-trained models. With a data pruning rate of 30-35% across all 16 models, our method exhibits only marginal performance degradation (less than 1% on average) compared to corresponding models trained on the full dataset counterparts across various downstream datasets, and also surpasses several baselines with a large performance margin. Additionally, the byproduct from our method, i.e., coresets derived from the original datasets after pre-training, also demonstrates significant superiority in terms of downstream performance over other coreset selection approaches.", "keywords": [ "Contrastive Pre-Training", "Data Efficiency" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "pRCOZllZdT", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Juan Viguera Diez", "gender": "Male", "institution": "Chalmers University", "country": "", "position": "PhD student" }, { "name": "Mathias Jacob Schreiner", "gender": "Male", "institution": "Technical University of Denmark", "country": "", "position": "PhD student" }, { "name": "Ola Engkvist", "gender": "Male", "institution": "Chalmers University of Technology", "country": "", "position": "Full Professor" }, { "name": "Simon Olsson", "gender": "Not Specified", "institution": "Chalmers University of Technology and University of Gothenburg", "country": "SE", "position": "Associate Professor" } ], "n_formula": 29, "n_formula_1": 24, "n_ref_uni": 81, "n_ref": 125, "n_ref_all": 146, "n_fig": 14, "n_tab": 2, "L_tab": 223, "n_element_tab": 4, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2185, "formula_len_all_1": 1462, "len_all": 198747, "len_all_1": 66358, "len_abs": 1107, "len_title": 96, "len_sents": 47822, "len_sents_1": 30156, "n_sents": 360, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 1113, "title": "Boltzmann priors for Implicit Transfer Operators", "abs": "Accurate prediction of thermodynamic properties is essential in drug discovery and materials science. Molecular dynamics (MD) simulations provide a principled approach to this task, yet they typically rely on prohibitively long sequential simulations. Implicit Transfer Operator (ITO) Learning offers a promising approach to address this limitation by enabling stable simulation with time-steps orders of magnitude larger than MD. However, to train ITOs, we need extensive, unbiased MD data, limiting their practical applicability. Here, we introduce Boltzmann Priors for ITO (BoPITO) to enhance ITO learning in two ways. First, BoPITO enables more efficient data generation, and second, it embeds inductive biases for long-term dynamical behavior, simultaneously improving sample efficiency by one order of magnitude and guaranteeing asymptotically unbiased equilibrium statistics. Further, we showcase the use of BoPITO in a new tunable sampling protocol interpolating ITO models trained on off-equilibrium simulation data and an unbiased equilibrium distribution to solve inverse problems in molecular science.", "keywords": [ "Molecular Dynamics", "Generative Models", "Transfer Operators", "Diffusion Models", "Boltzmann Generators" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "pQsllTesiE", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Abhishek Dubey", "gender": "Male", "institution": "Vanderbilt University", "country": "", "position": "Associate Professor" }, { "name": "Aron Laszka", "gender": "Not Specified", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Ayan Mukhopadhyay", "gender": "unknown", "institution": "Vanderbilt University", "country": "US", "position": "Researcher" }, { "name": "Baiting Luo", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "PhD student" }, { "name": "Geoffrey Pettet", "gender": "Female", "institution": "Vanderbilt University", "country": "", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 7, "n_ref_uni": 43, "n_ref": 71, "n_ref_all": 85, "n_fig": 6, "n_tab": 7, "L_tab": 1707, "n_element_tab": 178, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1225, "n_element_tab_1": 141, "formula_len_all": 1004, "formula_len_all_1": 497, "len_all": 163791, "len_all_1": 66367, "len_abs": 1549, "len_title": 138, "len_sents": 52078, "len_sents_1": 31915, "n_sents": 338, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 88, "L_abs": 1563, "title": "Scalable Decision-Making in Stochastic Environments through Learned Temporal Abstraction", "abs": "Sequential decision-making in high-dimensional continuous action spaces, particularly in stochastic environments, faces significant computational challenges. We explore this challenge in the traditional offline RL setting, where an agent must learn how to make decisions based on data collected through a stochastic behavior policy. We present \\textit{Latent Macro Action Planner} (L-MAP), which addresses this challenge by learning a set of temporally extended macro-actions through a state-conditional Vector Quantized Variational Autoencoder (VQ-VAE), effectively reducing action dimensionality. L-MAP employs a (separate) learned prior model that acts as a latent transition model and allows efficient sampling of plausible actions. During planning, our approach accounts for stochasticity in both the environment and the behavior policy by using Monte Carlo tree search (MCTS). In offline RL settings, including stochastic continuous control tasks, L-MAP efficiently searches over discrete latent actions to yield high expected returns.\nEmpirical results demonstrate that L-MAP maintains low decision latency despite increased action dimensionality. Notably, across tasks ranging from continuous control with inherently stochastic dynamics to high-dimensional robotic hand manipulation, L-MAP significantly outperforms existing model-based methods and performs on par with strong model-free actor-critic baselines, highlighting the effectiveness of the proposed approach in planning in complex and stochastic environments with high-dimensional action spaces.", "keywords": [ "Sequential Decision-Making", "Monte Carlo Tree Search", "Temporal Abstraction", "Planning", "Model-based Reinforcement Learning", "Offline Reinforcement Learning" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 3, 4, 3 ], "confidence_list": [ 4, 2, 3 ] }, { "paper_id": "pQqeQpMkE7", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 7, "author_info_list": [ { "name": "Ang Li", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Associate Professor" }, { "name": "Aurojit Panda", "gender": "Male", "institution": "New York University", "country": "", "position": "Assistant Professor" }, { "name": "Daohan Lu", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Haoyang Weng", "gender": "Male", "institution": "New York University", "country": "", "position": "Intern" }, { "name": "Hexu Zhao", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Jinyang Li", "gender": "unknown", "institution": "New York University", "country": "", "position": "Full Professor" }, { "name": "Saining Xie", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 53, "n_ref_all": 83, "n_fig": 17, "n_tab": 11, "L_tab": 4035, "n_element_tab": 573, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 562, "n_element_tab_1": 53, "formula_len_all": 494, "formula_len_all_1": 274, "len_all": 145279, "len_all_1": 53896, "len_abs": 1250, "len_title": 92, "len_sents": 44418, "len_sents_1": 25288, "n_sents": 341, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 44, "L_abs": 1347, "title": "On Scaling Up 3D Gaussian Splatting Training", "abs": "3D Gaussian Splatting (3DGS) is increasingly popular for 3D reconstruction due to its superior visual quality and rendering speed. However, 3DGS training currently occurs on a single GPU, limiting its ability to handle high-resolution and large-scale 3D reconstruction tasks due to memory constraints. We introduce Grendel, a distributed system designed to partition 3DGS parameters and parallelize computation across multiple GPUs. As each Gaussian affects a small, dynamic subset of rendered pixels, Grendel employs sparse all-to-all communication to transfer the necessary Gaussians to pixel partitions and performs dynamic load balancing. Unlike existing 3DGS systems that train using one camera view image at a time, Grendel supports batched training with multiple views. We explore various optimization hyperparameter scaling strategies and find that a simple sqrt(batch-size) scaling rule is highly effective. Evaluations using large-scale, high-resolution scenes show that Grendel enhances rendering quality by scaling up 3DGS parameters across multiple GPUs. On the 4K ``Rubble'' dataset, we achieve a test PSNR of 27.28 by distributing 40.4 million Gaussians across 16 GPU, compared to a PSNR of 26.28 using 11.2 million Gaussians on a single GPU. Grendel is an open-source project available at: https://github.com/nyu-systems/Grendel-GS", "keywords": [ "Gaussian Splatting", "Machine Learning System", "Distributed Training" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 4 ], "contribution_list": [ 4, 3, 4, 4 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "pQoD5MtZve", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Yiting Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhi-Hong Deng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 16, "n_ref_uni": 25, "n_ref": 37, "n_ref_all": 57, "n_fig": 4, "n_tab": 9, "L_tab": 5731, "n_element_tab": 375, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1965, "n_element_tab_1": 136, "formula_len_all": 1042, "formula_len_all_1": 956, "len_all": 204702, "len_all_1": 58632, "len_abs": 935, "len_title": 112, "len_sents": 45280, "len_sents_1": 25230, "n_sents": 363, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 938, "title": "Iterative Vectors: Boost In-Context Learning within Activations", "abs": "In-context learning (ICL) has emerged as a standard paradigm for utilizing language models. Although ICL is convenient due to the absence of backpropagation, selecting and processing appropriate demonstration examples can be difficult and time-consuming, particularly when the number of examples is large. We propose to explore the potential of activation space through Iterative Vectors (IVs), a technique designed to enhance in-context performance and necessitating only forward inference passes. IVs are employed by first extracting and iteratively steering activations within a language model, then applying them during inference with minimal computational and memory overhead. We evaluate IVs across numerous tasks using four popular models and observe significant improvements. Our findings suggest that activation steering can serve as a promising direction for in-context learning, thereby opening new avenues for future research.", "keywords": [ "Large Language Model", "In-context Learning" ], "rating_list": [ 8, 5, 6, 5, 6 ], "soundness_list": [ 4, 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 1, 1, 3 ], "contribution_list": [ 4, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3, 4 ] }, { "paper_id": "pQdei0Zb7a", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Gustav Wagner Zakarias", "gender": "Male", "institution": "Aalborg University", "country": "DK", "position": "PhD student" }, { "name": "Lars Kai Hansen", "gender": "Male", "institution": "Technical University of Denmark", "country": "", "position": "Full Professor" }, { "name": "Zheng-Hua Tan", "gender": "Male", "institution": "Aalborg University", "country": "DK", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 11, "n_ref_uni": 64, "n_ref": 107, "n_ref_all": 123, "n_fig": 2, "n_tab": 6, "L_tab": 2053, "n_element_tab": 298, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1164, "n_element_tab_1": 172, "formula_len_all": 1638, "formula_len_all_1": 1173, "len_all": 184413, "len_all_1": 66623, "len_abs": 1257, "len_title": 124, "len_sents": 45663, "len_sents_1": 29856, "n_sents": 280, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1263, "title": "BiSSL: Bilevel Optimization for Self-Supervised Pre-Training and Fine-Tuning", "abs": "In this work, we present BiSSL, a first-of-its-kind training framework that introduces bilevel optimization to enhance the alignment between the pretext pre-training and downstream fine-tuning stages in self-supervised learning. BiSSL formulates the pretext and downstream task objectives as the lower- and upper-level objectives in a bilevel optimization problem and serves as an intermediate training stage within the self-supervised learning pipeline. By more explicitly modeling the interdependence of these training stages, BiSSL facilitates enhanced information sharing between them, ultimately leading to a backbone parameter initialization that is better suited for the downstream task. We propose a training algorithm that alternates between optimizing the two objectives defined in BiSSL. Using a ResNet-18 backbone pre-trained with SimCLR on the STL10 dataset, we demonstrate that our proposed framework consistently achieves improved or competitive classification accuracies across various downstream image classification datasets compared to the conventional self-supervised learning pipeline. Qualitative analyses of the backbone features further suggest that BiSSL enhances the alignment of downstream features in the backbone prior to fine-tuning.", "keywords": [ "Self-Supervised Learning", "Bilevel Optimization" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "pQOHbTpAwf", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Liang Zhao", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Associate Professor" }, { "name": "Yuntong Hu", "gender": "Male", "institution": "Emory University", "country": "US", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 16, "n_ref_uni": 62, "n_ref": 82, "n_ref_all": 96, "n_fig": 4, "n_tab": 7, "L_tab": 3759, "n_element_tab": 697, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1518, "n_element_tab_1": 276, "formula_len_all": 1511, "formula_len_all_1": 1187, "len_all": 149345, "len_all_1": 69297, "len_abs": 1653, "len_title": 112, "len_sents": 38699, "len_sents_1": 31802, "n_sents": 283, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1483, "title": "DeepNT: Path-Centric Graph Neural Networks for Network Tomography", "abs": "Network tomography is a crucial problem in network monitoring, where observable path measurements are used to infer unmeasured network properties, making it essential for tasks such as route selection, fault diagnosis, and traffic control. However, most existing methods assume that the network topology is fully known\u2014an assumption that rarely holds in practice. The incomplete topology introduces significant challenges in extracting path information for predicting path performance. Furthermore, these approaches are typically designed for a single path performance metric and lack the flexibility to handle multiple metrics that may be of interest in a single network. To address these limitations, we propose Deep Network Tomography (DeepNT), a new framework that simultaneously infers the adjacency matrix and predicts unmeasured paths by leveraging Graph Neural Networks (GNNs) as backbones to learn path-centric end-node pair representations under incomplete network topology. To ensure that the learned adjacency matrix aligns with the characteristics of real-world networks, we propose a novel learning objective that constrains the model in terms of connectivity, sparsity, and path performance bounds, enabling robust generalization across a variety of performance metrics. Extensive experiments on real-world and synthetic datasets demonstrate the superiority of DeepNT in predicting performance metrics and inferring graph topology compared to state-of-the-art methods.", "keywords": [ "Network Tomography", "Graph Structure Learning" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "pQJi9EsmCc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chuong Nguyen", "gender": "Male", "institution": "CSIRO", "country": "AU", "position": "Researcher" }, { "name": "Fahira Afzal Maken", "gender": "Female", "institution": "Commonwealth Scientific and Industrial Research Organisation, CSIRO", "country": "AU", "position": "Researcher" }, { "name": "Hongdong Li", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "Full Professor" }, { "name": "Jinguang Tong", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "PhD student" }, { "name": "Sundaram Muthu", "gender": "Male", "institution": "Indian Institute of Technology, Madras", "country": "IN", "position": "Postdoc" }, { "name": "Xuesong Li", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "Postdoc" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 34, "n_ref": 74, "n_ref_all": 90, "n_fig": 3, "n_tab": 5, "L_tab": 1446, "n_element_tab": 81, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1040, "n_element_tab_1": 55, "formula_len_all": 516, "formula_len_all_1": 524, "len_all": 113630, "len_all_1": 59002, "len_abs": 1106, "len_title": 126, "len_sents": 34124, "len_sents_1": 27043, "n_sents": 257, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1123, "title": "S$^2$GS: Self-supervised Gaussian Segmentation for Automatic 3D Object Scanning", "abs": "Automatic 3D object scanning typically involves reconstructing rotating objects from images captured from different viewpoints. In such circumstances where both the object and camera are moving, existing methods need object masks for reconstruction, and the mask quality can significantly affect the final reconstruction. However, obtaining high-quality and view-consistent object masks is challenging and laborious in practice. We address this issue by introducing Self-Supervised Gaussian Segmentation (S$^2$GS), which automatically segments the object from the background without relying on any segmentation masks. This is achieved by extending Gaussian Splatting with a learnable parameter that indicates the probability of each Gaussian belonging to the target object. We optimize this parameter using implicit object transformation constraints and regularization terms. We evaluate S$^2$GS on our new synthetic and real datasets. Experimental results show that our approach outperforms the state-of-the-art methods (2DGS) with object masks by \\(27\\%\\) for novel-view synthesis and \\(7\\%\\) for geometry reconstruction.", "keywords": [ "Self-supervised segmentation", "Gaussian splatting", "3D Reconstruction" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "pPyJyeLriR", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Lu Yi", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Zhewei Wei", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 136, "n_formula_1": 22, "n_ref_uni": 38, "n_ref": 91, "n_ref_all": 132, "n_fig": 6, "n_tab": 16, "L_tab": 5774, "n_element_tab": 1010, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1931, "n_element_tab_1": 253, "formula_len_all": 12390, "formula_len_all_1": 1882, "len_all": 292465, "len_all_1": 85685, "len_abs": 6213, "len_title": 133, "len_sents": 81819, "len_sents_1": 33643, "n_sents": 841, "n_sents_1": 283, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 85, "L_abs": 1712, "title": "Scalable and Certifiable Graph Unlearning: Overcoming the Approximation Error Barrier", "abs": "Graph unlearning has emerged as a pivotal research area for ensuring privacy protection, given the widespread adoption of Graph Neural Networks (GNNs) in applications involving sensitive user data. Among existing studies, certified graph unlearning is distinguished by providing robust privacy guarantees. However, current certified graph unlearning methods are impractical for large-scale graphs because they necessitate the costly re-computation of graph propagation for each unlearning request. Although numerous scalable techniques have been developed to accelerate graph propagation for GNNs, their integration into certified graph unlearning remains uncertain as these scalable approaches introduce approximation errors into node embeddings. In contrast, certified graph unlearning demands bounded model error on exact node embeddings to maintain its certified guarantee.\n\n To address this challenge, we present ScaleGUN, the first approach to scale certified graph unlearning to billion-edge graphs. ScaleGUN integrates the approximate graph propagation technique into certified graph unlearning, offering certified guarantees for three unlearning scenarios: node feature, edge and node unlearning. \n Extensive experiments on real-world datasets demonstrate the efficiency and unlearning efficacy of ScaleGUN. Remarkably, ScaleGUN accomplishes $(\\epsilon,\\delta)=(1,10^{-4})$ certified unlearning on the billion-edge graph ogbn-papers100M in 20 seconds for a 5,000 random edge removal request -- of which only 5 seconds are required for updating the node embeddings -- compared to 1.91 hours for retraining and 1.89 hours for re-propagation. Our code is available at https://github.com/luyi256/ScaleGUN.", "keywords": [ "Machine Unlearning", "Graph Neural Networks", "Scalability" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "pPmQvd1NUp", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Alberto Del Bimbo", "gender": "Male", "institution": "Universita di Firenze", "country": "IT", "position": "Full Professor" }, { "name": "Federico Pernici", "gender": "Male", "institution": "University of Florence", "country": "IT", "position": "Assistant Professor" }, { "name": "Niccol\u00f2 Biondi", "gender": "Male", "institution": "University of Florence", "country": "IT", "position": "Postdoc" }, { "name": "Simone Ricci", "gender": "Male", "institution": "University of Florence", "country": "IT", "position": "Postdoc" } ], "n_formula": 58, "n_formula_1": 19, "n_ref_uni": 55, "n_ref": 109, "n_ref_all": 144, "n_fig": 27, "n_tab": 3, "L_tab": 2545, "n_element_tab": 343, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3628, "formula_len_all_1": 1180, "len_all": 217144, "len_all_1": 67935, "len_abs": 1827, "len_title": 85, "len_sents": 63134, "len_sents_1": 33152, "n_sents": 440, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 1730, "title": "The Probability Simplex is Compatible", "abs": "In retrieval systems, updating the base model involves re-extracting feature vectors for all gallery data due to changes in internal feature representations. This process can be computationally expensive and time-consuming, especially for large-scale gallery sets. To address this issue, backward compatible learning was introduced, allowing direct comparison between the representations of the old model and those obtained by the newly trained model. Existing backward compatible methods introduce additional losses or specific network architecture changes, which require the availability of base models, thereby limiting compatibility with models trained independently. In this paper, we show that any independently trained model can be made compatible with any other by simply using features derived from softmax outputs. \nWe leverage the geometric properties of the softmax function, which projects vectors into the Probability Simplex, preserving the alignment of softmax vectors across model updates and verifying the definition of compatibility. A similar property is observed when using logits as a feature representation. They distribute during training in a simplex configuration, but with a wider spread in the feature distribution than softmax outputs, leading to a more robust and transferable representation. Our framework achieves state-of-the-art performance on standard benchmarks, where either the number of training classes extends across multiple steps or the base model is updated with advanced network architectures. This demonstrates that any publicly available pretrained model can be made compatible without requiring any additional training or adaptation. Our code will be made available upon acceptance.", "keywords": [ "Deep Learning", "Representation Learning", "Compatible Learning", "Neural Collapse" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 4, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 2, 3 ] }, { "paper_id": "pPWAPiFf3z", "primary_area": "applications to neuroscience & cognitive science", "n_author": 10, "author_info_list": [ { "name": "Bin Jiang", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Researcher" }, { "name": "Bryce Allen Bagley", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Researcher" }, { "name": "Eric Tranvinh", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Ge Cheng", "gender": "Female", "institution": "Stanford University", "country": "CN", "position": "MS student" }, { "name": "Matei Arm\u0103na\u0219u", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Intern" }, { "name": "Michael Iv", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Michael Zeineh", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Nancy J. Fischbein MD", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Olivier Gevaert", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Sergei Petrov", "gender": "unknown", "institution": "Hugging Face", "country": "FR", "position": "Machine learning engineer" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 17, "n_ref": 27, "n_ref_all": 45, "n_fig": 12, "n_tab": 2, "L_tab": 491, "n_element_tab": 67, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 246, "n_element_tab_1": 15, "formula_len_all": 131, "formula_len_all_1": 131, "len_all": 89960, "len_all_1": 61066, "len_abs": 1383, "len_title": 159, "len_sents": 34536, "len_sents_1": 32213, "n_sents": 216, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 110, "L_abs": 2619, "title": "Generative Editing via Convolutional Obscuring (GECO): A Generative Adversarial Network for MRI de-artifacting", "abs": "Magnetic resonance imaging (MRI) is the dominant diagnostic technique to non-invasively image the brain, and deep learning has proven a powerful tool for analyzing these images. However, machine learning models trained on such MRI data have empirically shown an ability to detect complex and invisible artifacts, such as which type of machine a scan was taken from to a high degree of accuracy. Such artifacts are potentially invisible to the human eye, but can be identified by machine learning systems, leading them to focus on irrelevant features rather than scientifically and/or medically useful ones. For example, machine learning systems can often \u201cshortcut\u201d past the actual features researchers would like to detect and utilize separate spurious correlations to make predictions. Several such undesired features have been reported to interfere with cross-institutional medical imaging deep learning research, and more are likely to be identified as time goes on. Here, we develop a method capable of removing these spurious correlations in an unsupervised manner, leveraging generative techniques to produce images which maintain image quality while learning how to remove technical artifacts. Generative Adversarial Networks are a class of deep learning architectures which have shown impressive efficacy in image generation and editing tasks, and our work builds upon this success. Here, we propose Generative Editing via Convolutional Obscuring (GECO), a Generative Adverserial Network for MRI deartifacting. GECO is based on a CycleGAN, a GAN architecture designed for image-to-image translation that is transforming an input image into a new image with one or more desirable properties. By formulating the CycleGAN loss as a two-player game with a regularization term and incentivizing the generator to erase spurious correlations the original image quality can be better preserved. Beginning with classifiers trained on original images to identify images based on artifacts of interest, GECO reduced the classifiers\u2019 ability to detect these spurious correlations from 97% down to a difference which is nearly equal to a classifier making purely random guesses. We also observe over 98% structural similarity between the original and deartifacted images, indicating the preservation of the vast majority of non-spurious information contained in the original images. In addition to solving the known problem of avoiding artifacts from scanner type, this method opens the door to potentially removing many other types of spurious correlations from medical images and other data modalities across many fields.", "keywords": [ "Deep convolutional neural networks", "computer vision", "medical machine learning", "image analysis", "generative adversarial networks", "artifact removal", "machine learning model generalization" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "pPQPQ7Yd58", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Han Qi", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Haocheng Yin", "gender": "Male", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "US", "position": "Researcher" }, { "name": "Heng Yang", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 61, "n_ref": 88, "n_ref_all": 133, "n_fig": 20, "n_tab": 3, "L_tab": 968, "n_element_tab": 53, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 233, "n_element_tab_1": 35, "formula_len_all": 523, "formula_len_all_1": 573, "len_all": 170393, "len_all_1": 59509, "len_abs": 1575, "len_title": 107, "len_sents": 48869, "len_sents_1": 27140, "n_sents": 400, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 59, "L_abs": 1574, "title": "Control-oriented Clustering of Visual Latent Representation", "abs": "We initiate a study of the geometry of the visual representation space ---the information channel from the vision encoder to the action decoder--- in an image-based control pipeline learned from behavior cloning. Inspired by the phenomenon of *neural collapse* (NC) in image classification, we empirically demonstrate the prevalent emergence of a similar *law of clustering* in the visual representation space. Specifically, \n\n- In discrete image-based control (e.g., Lunar Lander), the visual representations cluster according to the natural discrete action labels;\n\n- In continuous image-based control (e.g., Planar Pushing and Block Stacking), the clustering emerges according to ``control-oriented'' classes that are based on (a) the relative pose between the object and the target in the input or (b) the relative pose of the object induced by expert actions in the output. Each of the classes corresponds to one relative pose orthant (REPO).\n\nBeyond empirical observation, we show such a law of clustering can be leveraged as an algorithmic tool to improve test-time performance when training a policy with limited expert demonstrations. Particularly, we pretrain the vision encoder using NC as a regularization to encourage control-oriented clustering of the visual features. Surprisingly, such an NC-pretrained vision encoder, when finetuned end-to-end with the action decoder, boosts the test-time performance by 10% to 35%. Real-world vision-based planar pushing experiments confirmed the surprising advantage of control-oriented visual representation pretraining.", "keywords": [ "neural collapse", "learning from demonstration", "vision-based learning control" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 4, 3, 4, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "pPK6sNbFWV", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Changyou Chen", "gender": "Male", "institution": "State University of New York, Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Jian Chen", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Jiuxiang Gu", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Tong Sun", "gender": "Female", "institution": "Adobe Systems", "country": "US", "position": "Director, Document Intelligence Lab" }, { "name": "Yufan Zhou", "gender": "Male", "institution": "Adobe ", "country": "US", "position": "Researcher" }, { "name": "RUIYI ZHANG", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 40, "n_ref": 97, "n_ref_all": 116, "n_fig": 8, "n_tab": 7, "L_tab": 2898, "n_element_tab": 340, "n_fig_1": 8, "n_tab_1": 7, "L_tab_1": 1634, "n_element_tab_1": 255, "formula_len_all": 233, "formula_len_all_1": 269, "len_all": 169679, "len_all_1": 69175, "len_abs": 946, "len_title": 121, "len_sents": 38952, "len_sents_1": 32022, "n_sents": 282, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 958, "title": "LLaVA-Read: Enhancing Reading Ability of Multimodal Large Language Models", "abs": "Multimodal large language models have demonstrated impressive capabilities in understanding and manipulating images. However, many of these models struggle with comprehending intensive textual contents embedded within the images, primarily due to the limited text recognition and layout understanding ability. To understand the sources of these limitations, we perform an exploratory analysis showing the drawbacks of classical visual encoders on visual text understanding. Hence, we present LLaVA-Read, a multimodal large language model that utilizes dual visual encoders along with a visual text encoder. Our model surpasses existing state-of-the-art models in various text-rich image understanding tasks, showcasing enhanced comprehension of textual content within images. Together, our research suggests visual text understanding remains an open challenge and an \\textit{efficient} visual text encoder is crucial for future successful multimodal systems.", "keywords": [ "Multimodal Large Language Models", "Text-rich Images", "Visual Text Understanding" ], "rating_list": [ 3, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 1 ], "contribution_list": [ 2, 3, 1 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "pOq9vDIYev", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Asher Parker-Sartori", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Dylan Hadfield-Menell", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Stewart Slocum", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 7, "n_ref_uni": 44, "n_ref": 71, "n_ref_all": 97, "n_fig": 10, "n_tab": 3, "L_tab": 666, "n_element_tab": 148, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 375, "n_element_tab_1": 4, "formula_len_all": 2312, "formula_len_all_1": 480, "len_all": 185255, "len_all_1": 65564, "len_abs": 3494, "len_title": 106, "len_sents": 56106, "len_sents_1": 32063, "n_sents": 514, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1294, "title": "Diverse Preference Learning for Capabilities and Alignment", "abs": "As LLMs increasingly impact society, their ability to represent diverse perspectives is critical. However, recent studies reveal that alignment algorithms such as RLHF and DPO significantly reduce the diversity of LLM outputs. Not only do aligned LLMs generate text with repetitive structure and word choice, they also approach problems in more uniform ways, and their responses reflect a narrower range of societal perspectives. We attribute this problem to the KL divergence regularizer employed in preference learning algorithms. This causes the model to overweight majority opinions and sacrifice diversity in exchange for optimal reward. To address this, we propose Diverse Preference Learning, which decouples the entropy and cross-entropy terms in the KL penalty \u2014 allowing for fine-grained control over LLM generation diversity. From a capabilities perspective, LLMs trained using Diverse Preference Learning attain higher accuracy on difficult repeated sampling tasks and produce outputs with greater semantic and lexical diversity. From an alignment perspective, they are capable of representing a wider range of societal viewpoints and display improved logit calibration. Notably, Diverse Preference Learning resembles, but is a Pareto improvement over standard temperature scaling.", "keywords": [ "alignment", "diversity", "natural language processing", "reinforcement learning" ], "rating_list": [ 5, 8, 8 ], "soundness_list": [ 2, 3, 4 ], "presentation_list": [ 2, 4, 4 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "pOcGFvfgjS", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Ming-Ming Cheng", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Full Professor" }, { "name": "Qibin Hou", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Yupeng Zhou", "gender": "unknown", "institution": "Nankai University", "country": "CN", "position": "PhD student" }, { "name": "Zhen Li", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "PhD student" }, { "name": "Zhou Daquan", "gender": "Male", "institution": "Bytedance", "country": "US", "position": "Researcher" }, { "name": "Gao Peng", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Zhang Xuying", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 35, "n_ref": 73, "n_ref_all": 84, "n_fig": 6, "n_tab": 3, "L_tab": 375, "n_element_tab": 54, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 362, "n_element_tab_1": 16, "formula_len_all": 391, "formula_len_all_1": 391, "len_all": 116139, "len_all_1": 58332, "len_abs": 1786, "len_title": 132, "len_sents": 37354, "len_sents_1": 29549, "n_sents": 250, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1795, "title": "AR-1-to-3: Single Image to Consistent 3D Object Generation via Next-View Prediction", "abs": "Represented by Zero123 series of works, recent advancements in single-view 3D generation research have shown prominent progress by utilizing pre-trained 2D diffusion generation models. These approaches either generate multiple discrete views of a 3D object from a single-view image and a set of camera poses or produce multiple views simultaneously under specified camera conditions. However,\nit is hard to maintain consistency across different views and camera angles, especially for poses with large differences. In this paper, we introduce AR-1-to-3, a novel paradigm to generate multi-view images according to the input single image with significantly improved consistency in details. We achieve this by designing a novel auto-regressive scheme where novel views are generated based on previous views. The core of our method is first to generate views closer to the input view, which is utilized as contextual information to prompt the generation of farther views. To this end, we propose two image conditioning strategies, termed as Stacked-LE and LSTM-GE, to encode the sequence views. Particularly, Stacked-LE encodes the previously generated views into a stack embedding, which is employed as a local condition to modify the key and value matrices of the self-attention layers for denoising the target views of the current step. Meanwhile, LSTM-GE divides the previously generated views into two groups based on their elevations, whose feature vectors are encoded by two LSTM modules into high-level semantic information for global conditioning. Extensive experiments on the Objaverse dataset show that our method can synthesize more consistent 3D views and produce high-quality 3D assets that closely mirror the given image. Code and pre-trained weights will be made publicly available.", "keywords": [ "Multi-View Synthesis; Image-to-3D; Autoregressive Generation;" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "pOYa9nbwGr", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jaeseok Byun", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Sanghyuk Chun", "gender": "Male", "institution": "NAVER AI Lab", "country": "KR", "position": "Lead research scientist" }, { "name": "Seokhyeon Jeong", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Taesup Moon", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Wonjae Kim", "gender": "Male", "institution": "TwelveLabs", "country": "KR", "position": "Lead Research Scientist" } ], "n_formula": 1, "n_formula_1": 3, "n_ref_uni": 38, "n_ref": 169, "n_ref_all": 219, "n_fig": 5, "n_tab": 15, "L_tab": 4510, "n_element_tab": 553, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1754, "n_element_tab_1": 150, "formula_len_all": 17, "formula_len_all_1": 221, "len_all": 178691, "len_all_1": 72326, "len_abs": 1887, "len_title": 129, "len_sents": 61333, "len_sents_1": 34847, "n_sents": 381, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1928, "title": "Reducing Task Discrepancy of text encoders for Zero-Shot Composed Image Retrieval", "abs": "Composed Image Retrieval (CIR) aims to retrieve a target image based on a reference image and conditioning text, enabling controllable image searches.\nDue to the expensive dataset construction cost for CIR triplets, a zero-shot (ZS) CIR setting has been actively studied to eliminate the need for human-collected triplet training datasets of the target domain.\nThe mainstream methods of ZS-CIR research typically employ a projection module that projects a CLIP image embedding to the CLIP text token embedding space while all encoders are fixed.\nUsing such a projected embedding, those methods then generate an image-text composed feature, which is used as a query for retrieval.\nHowever, we point out that using fixed CLIP encoders for ZS-CIR has an inherent limitation since there exists a significant task discrepancy between the original pre-training task of the encoders (text $\\leftrightarrow$ image) and the target CIR task (image + text $\\leftrightarrow$ image). To reduce such a discrepancy, a naive solution would be to train both image and text encoders with CIR triplets in a supervised manner. \nInstead, we introduce the Reducing Task Discrepancy of text encoders for Zero-Shot Composed Image Retrieval (RTD), an efficient post-precessing approach designed to enhance the capability of text encoders for ZS-CIR. Namely, we devise a novel target-anchored text contrastive learning, which solely updates the text encoder using cheap text triplets, consisting of reference and target texts instead of images. \nWe also introduce two enhancements to this approach: a refined batch sampling strategy and a sophisticated concatenation scheme.\nIntegrating RTD into existing projection-based ZS-CIR methods significantly improves performance across various datasets and backbones, achieving competitive or superior results compared to other resource-intensive state-of-the-art CIR methods beyond projection-based approaches.", "keywords": [ "composed image retrieval; language only supervision; task discrepancy" ], "rating_list": [ 3, 3, 3, 3, 5 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 5, 4, 4 ] }, { "paper_id": "pOUAVXnOQP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Alireza Morsali", "gender": "Male", "institution": "Global Relay", "country": "CA", "position": "Senior data scientist" }, { "name": "Amirhossein Kazerouni", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "MohammadJavad Vaez", "gender": "Male", "institution": "University of Tehran", "country": "IR", "position": "MS student" }, { "name": "Mohammad Hossein Soltani", "gender": "Male", "institution": "Independent Researcher", "country": "IR", "position": "Researcher" }, { "name": "Morteza Mohammad-Noori", "gender": "Male", "institution": "University of Tehran", "country": "IR", "position": "Associate Professor" } ], "n_formula": 120, "n_formula_1": 16, "n_ref_uni": 41, "n_ref": 74, "n_ref_all": 98, "n_fig": 14, "n_tab": 1, "L_tab": 133, "n_element_tab": 10, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9508, "formula_len_all_1": 890, "len_all": 191063, "len_all_1": 57027, "len_abs": 966, "len_title": 130, "len_sents": 52387, "len_sents_1": 25206, "n_sents": 524, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 970, "title": "STAF: Sinusoidal Trainable Activation Functions for Implicit Neural Representation", "abs": "Implicit Neural Representation (INR) has emerged as a promising method for characterizing continuous signals. This paper addresses the spectral bias exhibited by conventional ReLU networks, which hampers their ability to reconstruct fine details in target signals. We introduce Sinusoidal Trainable Activation Functions (STAF), designed to model and reconstruct diverse complex signals with high precision. STAF mitigates spectral bias, enabling faster learning of high-frequency details compared to ReLU networks. We demonstrate STAF's superiority over state-of-the-art networks such as KAN, WIRE, SIREN, and Fourier features, achieving higher accuracy and faster convergence with superior Peak Signal-to-Noise Ratio (PSNR). Our extensive experimental evaluation establishes STAF's effectiveness in improving the reconstruction quality and training efficiency of continuous signals, making them valuable for various applications in computer graphics and related fields.", "keywords": [ "Implicit Neural Representation", "Activation Functions", "Parametric Activation Functions", "Neural Tangent Kernel" ], "rating_list": [ 3, 5, 8, 5 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 4, 4, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "pOO9cqLq7Q", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Aixin Tan", "gender": "unknown", "institution": "University of Iowa", "country": "", "position": "Associate Professor" }, { "name": "Jian Huang", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "CN", "position": "Full Professor" }, { "name": "Zhenhan Fang", "gender": "Male", "institution": "University of Iowa", "country": "US", "position": "PhD student" } ], "n_formula": 42, "n_formula_1": 25, "n_ref_uni": 33, "n_ref": 54, "n_ref_all": 62, "n_fig": 3, "n_tab": 5, "L_tab": 1817, "n_element_tab": 90, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1886, "n_element_tab_1": 96, "formula_len_all": 2773, "formula_len_all_1": 1554, "len_all": 127414, "len_all_1": 64245, "len_abs": 1291, "len_title": 119, "len_sents": 41640, "len_sents_1": 27519, "n_sents": 415, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1296, "title": "CONTRA: Conformal Prediction Region via Normalizing Flow Transformation", "abs": "Density estimation and reliable prediction regions for outputs are crucial in supervised and unsupervised learning. While conformal prediction effectively generates coverage-guaranteed regions, it struggles with multi-dimensional outputs due to reliance on one-dimensional nonconformity scores. To address this, we introduce CONTRA: CONformal prediction region via normalizing flow TRAnsformation. CONTRA utilizes the latent spaces of normalizing flows to define nonconformity scores based on distances from the center. This allows for the mapping of high-density regions in latent space to sharp prediction regions in the output space, surpassing traditional hyperrectangular or elliptical conformal regions. Further, for scenarios where other predictive models are favored over flow-based models, we extend CONTRA to enhance any such model with a reliable prediction region by training a simple normalizing flow on the residuals. We demonstrate that both CONTRA and its extension maintain guaranteed coverage probability and outperform existing methods in generating accurate prediction regions across various datasets. We conclude that CONTRA is an effective tool for (conditional) density estimation, addressing the under-explored challenge of delivering multi-dimensional prediction regions.", "keywords": [ "Uncertainty Quantification", "Conditional Density Estimation", "Conformalized Quantile Regression", "Coverage Probability", "Multi-dimensional Prediction" ], "rating_list": [ 6, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "pNxD5dpu1M", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Christian Fabian", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Heinz Koeppl", "gender": "Male", "institution": "TU Darmstadt", "country": "", "position": "Full Professor" }, { "name": "Kai Cui", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" } ], "n_formula": 76, "n_formula_1": 23, "n_ref_uni": 52, "n_ref": 97, "n_ref_all": 104, "n_fig": 3, "n_tab": 3, "L_tab": 1925, "n_element_tab": 136, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 2177, "n_element_tab_1": 144, "formula_len_all": 12975, "formula_len_all_1": 1768, "len_all": 200664, "len_all_1": 81116, "len_abs": 1108, "len_title": 111, "len_sents": 45669, "len_sents_1": 29826, "n_sents": 408, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1111, "title": "Learning Cooperative Mean Field Games on Sparse Chung-Lu Graphs", "abs": "Large agent networks are abundant in applications and nature and pose difficult challenges in the field of multi-agent reinforcement learning (MARL) due to their computational and theoretical complexity. While graphon mean field games and their extensions provide efficient learning algorithms for dense and moderately sparse agent networks, the case of realistic sparser graphs remains largely unsolved. Thus, we propose a novel cooperative mean field game (MFG) model based on the large class of Chung-Lu graphs including power law networks with coefficients above two. Besides a theoretical analysis, we design scalable learning algorithms which especially apply to the challenging class of graph sequences with finite first moment and infinite second moment. We compare our model and algorithms for various examples on synthetic and real world networks with MFG algorithms based on Lp graphons and graphexes. As it turns out, our approach outperforms existing methods in many examples and on various networks due to the special design aiming at an important, but so far hard to solve class of MARL problems.", "keywords": [ "Cooperative Mean Field Games", "Large Networks", "Sparse Graphs", "Multi Agent Reinforcement Learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "pNgyXuGcx4", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Byeongwook Kim", "gender": "unknown", "institution": "NAVER CLOUD", "country": "", "position": "Researcher" }, { "name": "Dongsoo Lee", "gender": "Male", "institution": "NAVER CLOVA", "country": "", "position": "Executive Officer" }, { "name": "Jeongin Bae", "gender": "Female", "institution": "NAVER", "country": "", "position": "Researcher" }, { "name": "Joonhyung Lee", "gender": "Male", "institution": "NAVER Cloud", "country": "KR", "position": "Researcher" }, { "name": "Se Jung Kwon", "gender": "Male", "institution": "NAVER Cloud", "country": "KR", "position": "AI Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 28, "n_ref": 54, "n_ref_all": 73, "n_fig": 8, "n_tab": 6, "L_tab": 1674, "n_element_tab": 248, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 810, "n_element_tab_1": 10, "formula_len_all": 192, "formula_len_all_1": 192, "len_all": 160848, "len_all_1": 58823, "len_abs": 1489, "len_title": 134, "len_sents": 37358, "len_sents_1": 29639, "n_sents": 215, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1495, "title": "To FP8 and Back Again: Quantifying Reduced Precision Effects on LLM Training Stability", "abs": "The massive computational costs associated with large language model (LLM) pretraining have spurred great interest in reduced-precision floating-point representations to accelerate the process. As a result, the BrainFloat16 (BF16) precision has become the de facto standard for LLM training, with hardware support included in recent generations of accelerators. This trend has gone even further in the latest processors, where FP8 has recently been introduced. However, prior experience with FP16, which was found to be less stable than BF16, raises concerns as to whether FP8, with even fewer bits than FP16, can be a cost-effective option for LLM training. We argue that reduced-precision training schemes must have similar training stability and hyperparameter sensitivities to their higher-precision counterparts in order to be cost-effective. However, we find that currently available methods for FP8 training are not robust enough to allow their use as economical replacements. This prompts us to investigate the stability of reduced-precision LLM training in terms of robustness across random seeds, learning rates, and datasets. To this end, we propose new evaluation techniques and a new metric for quantifying loss landscape sharpness in autoregressive language models. By simulating incremental bit reductions in floating-point representations, we analyze the relationship between representational power and training stability with the intent of aiding future research into the field.", "keywords": [ "LLM", "FP8", "Quantization", "Training" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 2, 2 ] }, { "paper_id": "pNdPJACSLB", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Arben Cela", "gender": "unknown", "institution": "Ecole Sup\u00e9rieure d'Ing\u00e9nieurs en Electronique et Electrotechnique", "country": "FR", "position": "Full Professor" }, { "name": "Jacob Ouanounou", "gender": "Male", "institution": "Ecole Sup\u00e9rieure d'Ing\u00e9nieurs en Electronique et Electrotechnique", "country": "FR", "position": "Lecturer" }, { "name": "Jiang You", "gender": "Male", "institution": "Ecole Sup\u00e9rieure d'Ing\u00e9nieurs en Electronique et Electrotechnique", "country": "FR", "position": "PhD student" }, { "name": "Patrick SIARRY", "gender": "Male", "institution": "Universit\u00e9 Paris-Est Cr\u00e9teil", "country": "FR", "position": "Full Professor" }, { "name": "Ren\u00e9 Natowicz", "gender": "Male", "institution": "Ecole Sup\u00e9rieure d'Ing\u00e9nieurs en Electronique et Electrotechnique", "country": "FR", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 16, "n_ref": 28, "n_ref_all": 33, "n_fig": 3, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 999, "formula_len_all_1": 999, "len_all": 61870, "len_all_1": 61870, "len_abs": 711, "len_title": 131, "len_sents": 20660, "len_sents_1": 20660, "n_sents": 156, "n_sents_1": 156, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 713, "title": "Learning K-U-Net in Constant Complexity with Application to Time Series Forecasting", "abs": "Training deep models for time series forecasting is a critical task with an inherent challenge of time complexity. While current methods generally ensure linear time complexity, our observations on temporal redundancy show that high-level features are learned 99.5\\% slower than low-level features. To address this issue, we introduce a new exponentially weighted stochastic gradient descent algorithm designed to achieve constant time complexity in deep learning models. We prove that the theoretical complexity of this learning method is constant. Evaluation of this method on Kernel U-Net (K-U-Net) on synthetic datasets shows a significant reduction in complexity while improving the accuracy of the test set.", "keywords": [ "machine learning", "time series", "complexity reduction" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 1, 1, 2, 2 ], "presentation_list": [ 1, 1, 2, 1 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "pNSJdyXZju", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Bernhard Sick", "gender": "Male", "institution": "Universit\u00e4t Kassel", "country": "", "position": "Full Professor" }, { "name": "Daniel Kottke", "gender": "Not Specified", "institution": "Universit\u00e4t Kassel", "country": "", "position": "Postdoc" }, { "name": "Denis Huseljic", "gender": "Male", "institution": "Universit\u00e4t Kassel", "country": "", "position": "PhD student" }, { "name": "Lukas Rauch", "gender": "Male", "institution": "University of Kassel", "country": "DE", "position": "PhD student" }, { "name": "Marek Herde", "gender": "Male", "institution": "Universit\u00e4t Kassel", "country": "", "position": "PhD student" }, { "name": "Paul Hahn", "gender": "Male", "institution": "Universit\u00e4t Kassel", "country": "", "position": "MS student" }, { "name": "Stephan Vogt", "gender": "Male", "institution": "Universit\u00e4t Kassel", "country": "", "position": "PhD student" }, { "name": "Zhixin Huang", "gender": "Male", "institution": "Technische Universit\u00e4t Carolo-Wilhelmina Braunschweig", "country": "", "position": "MS student" } ], "n_formula": 30, "n_formula_1": 7, "n_ref_uni": 52, "n_ref": 112, "n_ref_all": 129, "n_fig": 10, "n_tab": 2, "L_tab": 70, "n_element_tab": 9, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1753, "n_element_tab_1": 38, "formula_len_all": 1829, "formula_len_all_1": 511, "len_all": 166835, "len_all_1": 65442, "len_abs": 1253, "len_title": 78, "len_sents": 52301, "len_sents_1": 30180, "n_sents": 431, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1263, "title": "Efficient Bayesian Updates for Deep Active Learning via Laplace Approximations", "abs": "Deep active learning (AL) involves selecting batches of instances for annotation since retraining large deep neural networks (DNNs) after each label acquisition is computationally impractical. Employing a naive top-$b$ selection can result in a batch of redundant (similar) instances. To address this issue, various batch AL strategies have been developed, many of which employ clustering for diversity as a heuristic. In contrast, we approach this issue by substituting the costly retraining with an efficient Bayesian update. Our proposed update represents a second-order optimization step using the Gaussian posterior from a last-layer Laplace approximation. Thereby, we achieve low computational complexity by computing the inverse Hessian in closed form. We demonstrate that in typical AL settings, our update closely approximates retraining while being considerably faster. Leveraging our update, we introduce a new framework for batch selection through sequential construction by updating the DNN after each label acquisition. Furthermore, we incorporate our update into a look-ahead selection strategy as a feasible upper baseline approximating optimal batch selection. Our results highlight the potential of efficient updates to advance deep AL research.", "keywords": [ "Active Learning", "Deep Learning", "Batch Active Learning" ], "rating_list": [ 6, 5, 3, 3, 3 ], "soundness_list": [ 3, 1, 2, 2, 2 ], "presentation_list": [ 3, 4, 2, 3, 2 ], "contribution_list": [ 3, 1, 1, 1, 2 ], "confidence_list": [ 3, 4, 5, 5, 3 ] }, { "paper_id": "pMp5njgeLx", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Deli Zhao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Director" }, { "name": "Lidong Bing", "gender": "unknown", "institution": "Alibaba Group", "country": "SG", "position": "Scientist" }, { "name": "Ruochen Zhao", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Weiwen Xu", "gender": "Male", "institution": "Alibaba Group", "country": "SG", "position": "Researcher" }, { "name": "Wenxuan Zhang", "gender": "unknown", "institution": "Alibaba Group", "country": "SG", "position": "Researcher" }, { "name": "Yew Ken Chia", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 37, "n_ref": 108, "n_ref_all": 147, "n_fig": 11, "n_tab": 9, "L_tab": 2072, "n_element_tab": 124, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 884, "n_element_tab_1": 64, "formula_len_all": 339, "formula_len_all_1": 184, "len_all": 238168, "len_all_1": 65265, "len_abs": 1255, "len_title": 135, "len_sents": 58406, "len_sents_1": 29410, "n_sents": 521, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1260, "title": "Auto-Arena: Automating LLM Evaluations with Agent Peer Battles and Committee Discussions", "abs": "As LLMs continuously evolve, there is an urgent need for a reliable evaluation method that delivers trustworthy results promptly. Currently, static benchmarks suffer from inflexibility and unreliability, leading users to prefer human voting platforms like Chatbot Arena. However, human evaluations require significant manual effort. To address this, we propose the Auto-Arena, an innovative framework that automates the entire evaluation process using LLM-powered agents. Firstly, an LLM examiner generates questions. Then, two LLM candidates engage in a multi-round peer battle based on individual questions, aiming at revealing their true performance differences. Finally, a committee of LLM judges collaboratively discusses and decides the winner, reducing bias and enhancing fairness. During the peer battles, we observe intriguing scenarios where the LLM candidates display competitive behaviors and even learn from the opponents. In our extensive experiments involving 15 recent LLMs, Auto-Arena shows a 92.14% correlation with human preferences, surpassing all previous expert-annotated benchmarks without any manual efforts. As a result, Auto-Arena offers a promising alternative to current human evaluation platforms for evaluating LLMs automatically.", "keywords": [ "LLM Evaluation", "LLM Agents" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "pLyjsv1KWH", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Guancheng Wan", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Researcher" }, { "name": "Mang Ye", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Professor" }, { "name": "Wenke Huang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Yangxu Liao", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 46, "n_ref": 68, "n_ref_all": 75, "n_fig": 2, "n_tab": 1, "L_tab": 563, "n_element_tab": 37, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 610, "n_element_tab_1": 39, "formula_len_all": 1091, "formula_len_all_1": 1006, "len_all": 123060, "len_all_1": 56392, "len_abs": 1178, "len_title": 162, "len_sents": 27922, "len_sents_1": 25737, "n_sents": 217, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 114, "L_abs": 1202, "title": "Divide And Conquer: Efficiently Decoupling Consensus And Divergence For Federated Large Language Model Fine-Tuning", "abs": "Federated Learning provides an efficient framework for fine-tuning Large Language Models (LLMs) on diverse private datasets, addressing the growing scarcity of publicly available training data while maintaining data privacy. However, in practice, client data typically spans multiple domains, posing significant challenges for the global model\u2019s generalization capabilities. To address this issue, we introduce a novel framework, **Fed**erated **C**onsensus-**D**ivergence **D**ecoupling for LLM Fine-Tuning (**FedCDD**), designed to enhance global model performance in such heterogeneous environments. Our framework introduces a mechanism for consensus aggregation and divergence alignment, decoupling client updates into \u201cconsensus\u201d and \u201cdivergence\u201d parts. This allows the LLM to maintain a unified consensus while accommodating domain-specific divergences. Additionally, we employ a Gaussian-Noise Mask to regulate local model uploads, preventing the LLM from overfitting to domain-specific knowledge. Experimental results on heterogeneous datasets demonstrate the superiority of our approach over existing methods. The code is anonymously available at https://anonymous.4open.science/r/FedCDD-5DA6.", "keywords": [ "Federated Learning", "Large Language Model" ], "rating_list": [ 3, 3, 1, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 3, 5, 2 ] }, { "paper_id": "pLOep5sYWe", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Dahun Shin", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Dongyeop Lee", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Jinseok Chung", "gender": "unknown", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Namhoon Lee", "gender": "unknown", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 38, "n_formula_1": 13, "n_ref_uni": 76, "n_ref": 134, "n_ref_all": 160, "n_fig": 11, "n_tab": 10, "L_tab": 5333, "n_element_tab": 554, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1667, "n_element_tab_1": 147, "formula_len_all": 2781, "formula_len_all_1": 647, "len_all": 179363, "len_all_1": 62139, "len_abs": 1201, "len_title": 140, "len_sents": 45045, "len_sents_1": 28558, "n_sents": 317, "n_sents_1": 158, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1206, "title": "Sassha: Sharpness-aware Adaptive Second-order Optimization with Stable Hessian Approximation", "abs": "Approximate second-order optimization methods have gained attention due to their low computational and memory overhead.\nWhile these methods have the potential to accelerate neural network training, they often exhibit poorer generalization compared to first-order approaches. To address this limitation, we first analyze existing second-order methods through the lens of the loss landscape, demonstrating that their reduced generalization performance is somewhat attributed to the sharpness of the solutions they converge to. In response, we introduce Sassha, a novel approach designed to enhance generalization by explicitly reducing sharpness. In fact, this sharpness minimization scheme is designed to accommodate lazy and stable Hessian updates, so as to secure efficiency and robustness besides flatness. To validate its effectiveness, we conduct a wide range of deep learning experiments including standard vision and language tasks, where Sassha achieves competitive performance. Notably, Sassha demonstrates strong generalization in noisy data settings and significantly outperforms other methods in these scenarios. Additionally, we verify the robustness of\u001cSassha through various ablation studies.", "keywords": [ "deep learning", "second-order optimization", "sharpness minimization" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "pL8ws91RW2", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Ashish Dubey", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Principal Researcher" } ], "n_formula": 26, "n_formula_1": 21, "n_ref_uni": 12, "n_ref": 22, "n_ref_all": 40, "n_fig": 3, "n_tab": 10, "L_tab": 3295, "n_element_tab": 250, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1075, "n_element_tab_1": 109, "formula_len_all": 1143, "formula_len_all_1": 1062, "len_all": 78704, "len_all_1": 44725, "len_abs": 1626, "len_title": 102, "len_sents": 22542, "len_sents_1": 16593, "n_sents": 186, "n_sents_1": 136, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1639, "title": "Hierarchical Self-Supervised Graph Contrastive Learning: Capturing Multi-Scale Structural Information", "abs": "Graph Neural Networks (GNNs) have emerged as powerful tools for learning rep-resentations from graph-structured data Kipf & Welling (2017); Veli\u02c7ckovic\u00b4 et al.(2018), but often rely heavily on labeled data for training. This paper introduces a novel hierarchical self-supervised graph contrastive learning framework that ef-fectively leverages unlabeled data to enhance node representations. Our method captures rich structural information at multiple scales by incorporating contrastive objectives at the node, subgraph, and graph levels, extending previous work on self-supervised learning for graphs Veli\u02c7ckovic\u00b4 et al. (2019); You et al. (2020). We employ an adaptive graph augmentation strategy to generate meaningful views of the graph while preserving essential properties. Through extensive experiments on benchmark datasets, including Cora, Citeseer, PubMed Sen & Dhillon (2008), and Reddit Hamilton et al. (2017), we demonstrate that our approach consistently outperforms both supervised and self-supervised baseline models in node clas-sification tasks. Our method shows particular strength in low-label regimes and exhibits strong generalization capabilities in both transductive and inductive set-tings. Ablation studies confirm the importance of each hierarchical component, while qualitative analyses illustrate the discriminative power of the learned em-beddings. This work opens new avenues for self-supervised learning on graphs and has broad implications for applications where labeled data is scarce or ex-pensive to obtain, such as in social networks Perozzi et al. (2014) and biological networks Zitnik et al. (2017).", "keywords": [ "Graph Neural Networks", "Self-supervised Learning", "Contrastive Learning", "Hierarchical Representation", "Node Classification" ], "rating_list": [ 3, 3, 1, 3, 3 ], "soundness_list": [ 2, 2, 2, 1, 2 ], "presentation_list": [ 2, 2, 2, 2, 3 ], "contribution_list": [ 2, 1, 1, 1, 1 ], "confidence_list": [ 4, 4, 4, 5, 5 ] }, { "paper_id": "pKMpmbuKnd", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Litu Rout", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Sai Shankar Narasimhan", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Sanjay Shakkottai", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Full Professor" }, { "name": "Shubhankar Agarwal", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Sandeep Chinchali", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "Assistant Professor" } ], "n_formula": 146, "n_formula_1": 16, "n_ref_uni": 45, "n_ref": 116, "n_ref_all": 143, "n_fig": 7, "n_tab": 9, "L_tab": 2807, "n_element_tab": 388, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1463, "n_element_tab_1": 181, "formula_len_all": 12671, "formula_len_all_1": 1453, "len_all": 259881, "len_all_1": 73862, "len_abs": 1318, "len_title": 124, "len_sents": 81495, "len_sents_1": 31419, "n_sents": 790, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1229, "title": "Constrained Posterior Sampling: Time Series Generation with Hard Constraints", "abs": "Generating realistic time series samples is crucial for stress-testing models and protecting user privacy by using synthetic data. In engineering and safety-critical applications, these samples must meet certain hard constraints that are domain-specific or naturally imposed by physics or nature. Consider, for example, generating electricity demand patterns with constraints on peak demand times. This can be used to stress-test the functioning of power grids during adverse weather conditions. Existing approaches for generating constrained time series are either not scalable or degrade sample quality. To address these challenges, we introduce Constrained Posterior Sampling (CPS), a diffusion-based sampling algorithm that aims to project the posterior mean estimate into the constraint set after each denoising update. Notably, CPS scales to a large number of constraints ($\\sim100$) without requiring additional training. We provide theoretical justifications highlighting the impact of our projection step on sampling. Empirically, CPS outperforms state-of-the-art methods in sample quality and similarity to real time series by around 10\\% and 42\\%, respectively, on real-world stocks, traffic, and air quality datasets.", "keywords": [ "Time Series Generation", "Posterior Sampling", "Diffusion Models", "Controlled Generation" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 4, 4, 2, 3 ], "presentation_list": [ 3, 4, 4, 4 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "pKDmt7pc6h", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Cem Subakan", "gender": "unknown", "institution": "Universite Laval, Universit\u00e9 Laval", "country": "CA", "position": "Assistant Professor" }, { "name": "Emiliano Penaloza", "gender": "unknown", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "F\u0131rat \u00d6ncel", "gender": "Male", "institution": "Concordia University", "country": "CA", "position": "PhD student" }, { "name": "Haolun Wu", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "Laurent Charlin", "gender": "Male", "institution": "HEC Montreal", "country": "CA", "position": "Associate Professor" }, { "name": "Mirco Ravanelli", "gender": "Male", "institution": "Concordia University", "country": "CA", "position": "Assistant Professor" }, { "name": "Shubham Gupta", "gender": "Male", "institution": "Universit\u00e9 Laval", "country": "CA", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 24, "n_ref_uni": 24, "n_ref": 31, "n_ref_all": 48, "n_fig": 5, "n_tab": 6, "L_tab": 1822, "n_element_tab": 158, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 867, "n_element_tab_1": 130, "formula_len_all": 1310, "formula_len_all_1": 1311, "len_all": 95465, "len_all_1": 59389, "len_abs": 936, "len_title": 113, "len_sents": 27245, "len_sents_1": 25333, "n_sents": 220, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 944, "title": "Audio Prototypical Network for Controllable Music Recommendation", "abs": "Traditional recommendation systems represent user preferences in dense representations obtained through black-box encoder models. While these models often provide strong recommendation performance, they lack interpretability for users, leaving users unable to understand or control the system\u2019s modeling of their preferences. This limitation is especially challenging in music recommendation, where user preferences are highly personal and often evolve based on nuanced qualities like mood, genre, tempo, or instrumentation. \n In this paper, we propose an audio prototypical network for controllable music recommendation. This network expresses user preferences in terms of prototypes representative of semantically meaningful features pertaining to musical qualities. We show that the model obtains competitive recommendation performance compared to popular baseline models while also providing interpretable and controllable user profiles.", "keywords": [ "Scrutable Recommender System", "Controllability", "Explainable Machine Learning" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "pK4Z6NZ2DB", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Elan Rosenfeld", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Naomi Saphra", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Fellow" }, { "name": "Sara Kangaslahti", "gender": "unknown", "institution": "Harvard University, Harvard University", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 17, "n_ref": 27, "n_ref_all": 48, "n_fig": 11, "n_tab": 4, "L_tab": 701, "n_element_tab": 78, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 258, "n_element_tab_1": 33, "formula_len_all": 1295, "formula_len_all_1": 779, "len_all": 101145, "len_all_1": 55069, "len_abs": 971, "len_title": 114, "len_sents": 36188, "len_sents_1": 27473, "n_sents": 261, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 974, "title": "Loss in the Crowd: Hidden Breakthroughs in Language Model Training", "abs": "The training loss curves of a neural network are typically smooth. Any visible discontinuities draw attention as discrete conceptual breakthroughs, while the rest of training is less carefully studied. In this work we hypothesize that similar breakthroughs actually occur frequently throughout training, though their presence is obscured when monitoring the aggregate train loss. To find these hidden transitions, we introduce POLCA, a method for decomposing changes in loss along an arbitrary basis of the low rank training subspace. We use our method to identify clusters of samples that exhibit similar changes in loss through training, disaggregating the overall loss into that of smaller groups of conceptually similar datapoints. We validate our method on synthetic arithmetic and natural language, showing that POLCA recovers clusters which represent easily interpretable breakthroughs in the model's capabilities whose existence would otherwise be lost in the crowd.", "keywords": [ "interpretability techniques", "loss disaggregation", "phase transitions" ], "rating_list": [ 5, 5, 8, 5, 3 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2, 1 ], "confidence_list": [ 3, 4, 3, 4, 4 ] }, { "paper_id": "pK3oe2bubc", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Anders Sundnes L\u00f8vlie", "gender": "Male", "institution": "IT University of Copenhagen", "country": "", "position": "Associate Professor" }, { "name": "Matthias Freiberger", "gender": "Male", "institution": "Copenhagen University", "country": "", "position": "Postdoc" }, { "name": "Peter Kun", "gender": "Male", "institution": "IT University of Copenhagen", "country": "", "position": "Postdoc" }, { "name": "Sebastian Risi", "gender": "Male", "institution": "IT University of Copenhagen", "country": "", "position": "Professor" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 33, "n_ref": 55, "n_ref_all": 71, "n_fig": 8, "n_tab": 2, "L_tab": 809, "n_element_tab": 51, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 632, "n_element_tab_1": 38, "formula_len_all": 437, "formula_len_all_1": 134, "len_all": 117266, "len_all_1": 68608, "len_abs": 1422, "len_title": 142, "len_sents": 43749, "len_sents_1": 35957, "n_sents": 266, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1346, "title": "LayerShuffle: Enhancing Robustness in Vision Transformers by Randomizing Layer Execution Order", "abs": "Due to their architecture and how they are trained, artificial neural networks are typically not robust toward pruning, replacing, or shuffling layers at test time. However, such properties would be desirable for different applications, such as distributed neural network architectures where the order of execution cannot be guaranteed or parts of the network can fail during inference. In this work, we address these issues through a number of training approaches for vision transformers whose most important component is randomizing the execution order of attention modules at training time. With our proposed approaches, vision transformers are capable to adapt to arbitrary layer execution orders at test time assuming one tolerates a reduction (about 20\\%) in accuracy at the same model size. We analyse the feature representations of our trained models as well as how each layer contributes to the models prediction based on its position during inference. Our analysis shows that layers learn to contribute differently based on their position in the network. Importantly, trained models can also be randomly merged with each other resulting in functional (\"Frankenstein\") models without loss of performance compared to the source models. Finally, we layer-prune our models at test time and find that their performance declines gracefully.", "keywords": [ "Vision Transformer; Distributed Vision Transformer; Decentralization; AI Robustness; Merged Models; Scalable AI; Edge AI; Random Execution Order" ], "rating_list": [ 6, 1, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "pK2636Prbq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Arne Edward Michalson MD", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Christian Bluethgen", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Postdoc" }, { "name": "Curtis Langlotz", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Dennis Hein", "gender": "unknown", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "PhD student" }, { "name": "Eduardo Pontes Reis", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Researcher" }, { "name": "Hyun Joo Shin", "gender": "unknown", "institution": "Yonsei University", "country": "KR", "position": "Associate Professor" }, { "name": "Justin Xu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Researcher" }, { "name": "Maya Varma", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Sophie Ostmeier", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Zhihong Chen", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Akshay Chaudhari", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 51, "n_ref": 107, "n_ref_all": 133, "n_fig": 6, "n_tab": 11, "L_tab": 3406, "n_element_tab": 290, "n_fig_1": 3, "n_tab_1": 9, "L_tab_1": 1243, "n_element_tab_1": 131, "formula_len_all": 793, "formula_len_all_1": 967, "len_all": 254201, "len_all_1": 66779, "len_abs": 1520, "len_title": 145, "len_sents": 39030, "len_sents_1": 30292, "n_sents": 309, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1536, "title": "Preference fine-tuning for factuality in chest X-ray interpretation models without human feedback", "abs": "Radiologists play a crucial role by translating medical images into actionable reports. However, the field faces staffing shortages and increasing workloads. While automated approaches using vision-language models (VLMs) show promise as assistants, they require exceptionally high accuracy. Most current VLMs in radiology rely solely on supervised fine-tuning (SFT). Meanwhile, in the general domain, additional preference fine-tuning has become standard practice. The challenge in radiology lies in the prohibitive cost of obtaining radiologist feedback. To address this challenge, we propose an automated pipeline for preference feedback, focusing on chest X-ray (CXR) report generation. Our method leverages publicly available datasets containing pairs of images and radiologist-written reference reports with an LLM-as-a-Judge mechanism, eliminating the need for *additional radiologist feedback*. We evaluate and benchmark five direct alignment algorithms. Our results show up to a 57.4\\% improvement in average GREEN scores, a LLM-based metric for evaluating CXR reports, compared to the SFT baseline. We study reward overoptimization via length exploitation, with reports lengthening by up to 3.2x. To assess a potential alignment tax, we benchmark on six additional diverse tasks, finding no significant degradations. A reader study involving four board-certified radiologists indicates win rates of up to 0.62 over the SFT baseline, and macro-averaged F1 scores improve by up to 6.7\\%, highlighting the utility of our approach.", "keywords": [ "LLMs", "VLMs", "preference fine-tuning", "RLHF", "DPO", "DAAs", "preference alignment", "radiology", "chest X-rays" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "pJhgMNKEV3", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 7, "author_info_list": [ { "name": "Baolin Peng", "gender": "Male", "institution": "Research, Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Dian Yu", "gender": "Female", "institution": "Tencent AI Lab", "country": "", "position": "NLP researcher" }, { "name": "Dong Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Distinguished Scientist" }, { "name": "Haitao Mi", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Principal Researcher" }, { "name": "Linfeng Song", "gender": "unknown", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Tao Ge", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Principal Researcher" }, { "name": "Ye Tian", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 37, "n_ref": 90, "n_ref_all": 133, "n_fig": 1, "n_tab": 16, "L_tab": 2836, "n_element_tab": 546, "n_fig_1": 1, "n_tab_1": 8, "L_tab_1": 2161, "n_element_tab_1": 346, "formula_len_all": 1160, "formula_len_all_1": 814, "len_all": 171077, "len_all_1": 68758, "len_abs": 1258, "len_title": 131, "len_sents": 50378, "len_sents_1": 30159, "n_sents": 322, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1269, "title": "SIaM: Self-Improving Code-Assisted Mathematical Reasoning of Large Language Models", "abs": "There is a growing trend of teaching large language models (LLMs) to solve mathematical problems through coding. Existing studies primarily focus on prompting powerful, closed-source models to generate seed training data followed by in-domain data augmentation, equipping LLMs with considerable capabilities for code-assisted mathematical reasoning. However, continually training these models on augmented data derived from a few datasets such as GSM8K may impair their generalization abilities and restrict their effectiveness to limited question types. Conversely, the potential of improving such LLMs by leveraging large-scale, expert-written, diverse math question-answer pairs remains unexplored. To utilize these resources and tackle unique challenges such as code response assessment, we propose a novel paradigm that uses a code-based critic model to guide steps including question-code data construction, quality control, and complementary evaluation. We also explore different alignment algorithms with self-generated instruction/preference data to foster continuous self-improvement. Experiments across both in-distribution (up to $+5.7\\%$) and out-of-distribution ($+4.4\\%$) benchmarks in English and Chinese show the effectiveness of the proposed paradigm.", "keywords": [ "Large Language Models", "Mathematical Reasoning", "Code Generation" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "pIVOSU7TFQ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Jun Nie", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Xinmei Tian", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yiu-ming Cheung", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Full Professor" }, { "name": "Yonggang Zhang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Postdoc" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 58, "n_ref": 97, "n_ref_all": 127, "n_fig": 8, "n_tab": 15, "L_tab": 1883, "n_element_tab": 207, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 878, "n_element_tab_1": 44, "formula_len_all": 270, "formula_len_all_1": 270, "len_all": 162419, "len_all_1": 58350, "len_abs": 1623, "len_title": 129, "len_sents": 46063, "len_sents_1": 28253, "n_sents": 390, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1635, "title": "Detecting Discrepancies Between Generated and Natural Images Using Uncertainty", "abs": "In this work, we propose a novel approach for detecting AI-generated images by leveraging predictive uncertainty to mitigate misuse and associated risks. The motivation arises from the fundamental assumption regarding the distributional discrepancy between natural and AI-generated images. **The feasibility of distinguishing natural images from AI-generated ones is grounded in the distribution discrepancy between them**. Predictive uncertainty offers an effective approach for capturing distribution shifts, thereby providing insights into detecting AI-generated images. Namely, as the distribution shift between training and testing data increases, model performance typically degrades, often accompanied by increased predictive uncertainty. Therefore, we propose to employ predictive uncertainty to reflect the discrepancies between AI-generated and natural images. In this context, the challenge lies in ensuring that the model has been trained over sufficient natural images to avoid the risk of determining the distribution of natural images as that of generated images. We propose to leverage large-scale pre-trained models to calculate the uncertainty as the score for detecting AI-generated images. Inspired by MC Dropout, we perturb pre-trained models and find that the uncertainty can be captured by perturbing the weights of pre-trained models. This leads to a simple yet effective method for detecting AI-generated images using large-scale vision models: images that induce high uncertainty are identified as AI-generated. Comprehensive experiments across multiple benchmarks demonstrate the effectiveness of our method.", "keywords": [ "AI-generated image detection" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 1, 3, 4, 4 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "pIT0P1UASS", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 9, "author_info_list": [ { "name": "Cuneyt Gurcan Akcora", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Associate Professor" }, { "name": "Farimah Poursafaei", "gender": "Female", "institution": "McGill University", "country": "CA", "position": "Postdoc" }, { "name": "Guillaume Rabusseau", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Associate Professor" }, { "name": "Kiarash Shamsi", "gender": "Male", "institution": "University of Manitoba", "country": "CA", "position": "PhD student" }, { "name": "Poupak Azad", "gender": "Female", "institution": "University of Manitoba", "country": "CA", "position": "PhD student" }, { "name": "Razieh Shirzadkhani", "gender": "Female", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Researcher" }, { "name": "Reihaneh Rabbany", "gender": "Female", "institution": "McGill University", "country": "", "position": "Assistant Professor" }, { "name": "Andy Huang", "gender": "Male", "institution": "McGill University, Mila", "country": "CA", "position": "PhD student" }, { "name": "Bao Tran Gia Ngo", "gender": "Male", "institution": "University of Manitoba", "country": "CA", "position": "Undergrad student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 21, "n_ref": 37, "n_ref_all": 53, "n_fig": 6, "n_tab": 5, "L_tab": 5736, "n_element_tab": 164, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2751, "n_element_tab_1": 139, "formula_len_all": 130, "formula_len_all_1": 238, "len_all": 134021, "len_all_1": 65417, "len_abs": 1463, "len_title": 116, "len_sents": 31550, "len_sents_1": 28892, "n_sents": 239, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1393, "title": "Towards Neural Scaling Laws for Foundation Models on Temporal Graphs", "abs": "The field of temporal graph learning aims to learn from evolving network data to forecast future interactions. Given a collection of observed temporal graphs, is it possible to predict the evolution of an unseen network from the same domain?\nTo answer this question, we first present the Temporal Graph Scaling (TGS) dataset, a large collection of temporal graphs consisting of eighty-four ERC20 token transaction networks collected from 2017 to 2023. Next, we evaluate the transferability of Temporal Graph Neural Networks (TGNNs) for the temporal graph property prediction task by pre-training on a collection of up to sixty-four token transaction networks and then evaluating the downstream performance on twenty unseen token networks. We find that the neural scaling law observed in NLP and Computer Vision also applies in temporal graph learning, where pre-training on a greater number of networks leads to improved downstream performance. To the best of our knowledge, this is the first empirical demonstration of the transferability of temporal graph learning. On downstream token networks, the largest pre-trained model outperforms single model TGNNs on thirteen unseen test networks. Therefore, we believe that this is a promising first step towards building foundation models for temporal graphs. We provide the implementation of TGS at https://anonymous.4open.science/r/ScalingTGNs.", "keywords": [ "Temporal graph learning", "foundation model", "graph neural networks", "neural scaling law" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "pISLZG7ktL", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Chuan Wen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Fanqi Lin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "\u5609\u8bda \u6e38", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Pingyue Sheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yang Gao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yingdong Hu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 63, "n_ref": 109, "n_ref_all": 149, "n_fig": 14, "n_tab": 15, "L_tab": 3511, "n_element_tab": 525, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 816, "n_element_tab_1": 95, "formula_len_all": 39, "formula_len_all_1": 0, "len_all": 229087, "len_all_1": 64435, "len_abs": 1577, "len_title": 112, "len_sents": 63958, "len_sents_1": 31476, "n_sents": 512, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 64, "L_abs": 1584, "title": "Data Scaling Laws in Imitation Learning for Robotic Manipulation", "abs": "Data scaling has revolutionized fields like natural language processing and computer vision, providing models with remarkable generalization capabilities. In this paper, we investigate whether similar data scaling laws exist in robotics, particularly in robotic manipulation, and whether appropriate data scaling can yield single-task robot policies that can be deployed zero-shot for any object within the same category in any environment. To this end, we conduct a comprehensive empirical study on data scaling in imitation learning. By collecting data across numerous environments and objects, we study how a policy\u2019s generalization performance changes with the number of training environments, objects, and demonstrations. Throughout our research, we collect over 40,000 demonstrations and execute more than 15,000 real-world robot rollouts under a rigorous evaluation protocol. Our findings reveal several intriguing results: the generalization performance of the policy follows a roughly power-law relationship with the number of environments and objects. The diversity of environments and objects is far more important than the absolute number of demonstrations; once the number of demonstrations per environment or object reaches a certain threshold, additional demonstrations have minimal effect. Based on these insights, we propose an efficient data collection strategy. With four data collectors working for one afternoon, we collect sufficient data to enable the policies for two tasks to achieve approximately 90\\% success rates in novel environments with unseen objects.", "keywords": [ "Data Scaling Laws", "Imitation Learning", "Robotic Manipulation" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 4, 4, 4 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "pIJR9uPjy3", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Alban Colas-Gattepaille", "gender": "unknown", "institution": "Universit\u00e9 de Nice-Sophia Antipolis", "country": "", "position": "PhD student" }, { "name": "Alexandre Muzy", "gender": "Male", "institution": "CNRS", "country": "FR", "position": "Principal Researcher" } ], "n_formula": 28, "n_formula_1": 13, "n_ref_uni": 64, "n_ref": 111, "n_ref_all": 143, "n_fig": 9, "n_tab": 5, "L_tab": 2075, "n_element_tab": 299, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 976, "n_element_tab_1": 120, "formula_len_all": 2511, "formula_len_all_1": 685, "len_all": 177269, "len_all_1": 59202, "len_abs": 869, "len_title": 88, "len_sents": 38919, "len_sents_1": 26249, "n_sents": 311, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 872, "title": "Delay Neural Networks (DeNN) for exploiting temporal information in event-based datasets", "abs": "In Deep Neural Networks (DNN) and Spiking Neural Networks (SNN), the information of a neuron is computed based on the sum of the amplitudes (weights) of the electrical potentials received in input from other neurons. We propose here a new class of neural networks, namely Delay Neural Networks (DeNN), where the information of a neuron is computed based on the sum of its input synaptic delays and on the spike times of the electrical potentials received from other neurons. \nThis way, DeNN are designed to explicitly use exact continuous temporal information of spikes in both forward and backward passes, without approximation. (Deep) DeNN are applied here to images and event-based (audio and visual) data sets. Good performances are obtained, especially for datasets where temporal information is important, with much less parameters and less energy than other models.", "keywords": [ "Deep Learning", "Synaptic Delays", "Spiking Neural Networks", "Event-based" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "pHe4P1IVnb", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Chao Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Associate Professor" }, { "name": "Guangzhi Sun", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Junior Research Fellow" }, { "name": "Wen Wu", "gender": "Female", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Ziyang Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ziyun Cui", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 20, "n_formula_1": 19, "n_ref_uni": 33, "n_ref": 60, "n_ref_all": 86, "n_fig": 6, "n_tab": 9, "L_tab": 3177, "n_element_tab": 269, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 929, "n_element_tab_1": 121, "formula_len_all": 1133, "formula_len_all_1": 880, "len_all": 147024, "len_all_1": 65965, "len_abs": 1048, "len_title": 111, "len_sents": 46584, "len_sents_1": 30183, "n_sents": 365, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1051, "title": "Bayesian WeakS-to-Strong from Text Classification to Generation", "abs": "Advances in large language models raise the question of how alignment techniques will adapt as models become increasingly complex and humans will only be able to supervise them weakly. Weak-to-Strong mimics such a scenario where weak model supervision attempts to harness the full capabilities of a much stronger model. This work extends Weak-to-Strong to WeakS-to-Strong by exploring an ensemble of weak models which simulate the variability in human opinions. Confidence scores are estimated using a Bayesian approach to guide the WeakS-to-Strong generalization. Furthermore, we extend the application of WeakS-to-Strong from text classification tasks to text generation tasks where more advanced strategies are investigated for supervision. Moreover, direct preference optimization is applied to advance the student model's preference learning, beyond the basic learning framework of teacher forcing. Results demonstrate the effectiveness of the proposed approach for the reliability of a strong student model, showing potential for superalignment.", "keywords": [ "weaks-to-strong", "bayesian", "generation", "per token" ], "rating_list": [ 3, 5, 8, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "pHOH8FVrTp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Angelos Katharopoulos", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "David Grangier", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Ronan Collobert", "gender": "Male", "institution": "Apple", "country": "", "position": "Research Scientist" }, { "name": "Anastasiia FIlippova", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "MS student" } ], "n_formula": 23, "n_formula_1": 6, "n_ref_uni": 52, "n_ref": 97, "n_ref_all": 125, "n_fig": 13, "n_tab": 7, "L_tab": 2026, "n_element_tab": 353, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 311, "n_element_tab_1": 130, "formula_len_all": 2092, "formula_len_all_1": 359, "len_all": 179989, "len_all_1": 68464, "len_abs": 797, "len_title": 104, "len_sents": 50207, "len_sents_1": 33217, "n_sents": 370, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 56, "L_abs": 801, "title": "No Need to Talk: Asynchronous Mixture of Language Models", "abs": "We introduce SMALLTALK LM, an innovative method for training a mixture of language models in an almost asynchronous manner. Each\nmodel of the mixture specializes in distinct parts of the data distribution, without the need of high-bandwidth communication between the nodes training each model. At inference, a lightweight router directs a given sequence to a single expert, according to a short prefix. This inference scheme naturally uses a fraction of the parameters from the overall mixture model. Our experiments on language modeling demonstrate that SMALLTALK LM achieves significantly lower perplexity than dense model baselines for the same total training FLOPs and an almost identical inference cost. Finally, in our downstream evaluations we outperform the dense baseline on 75% of the tasks.", "keywords": [ "language models", "distributed learning", "divide and conquer", "efficient inference" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 4, 4 ], "contribution_list": [ 2, 4, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "pH543jrbe8", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Konstantin Hemker", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Mateja Jamnik", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "Professor in Artificial Intelligence" }, { "name": "Nikola Simidjievski", "gender": "Unspecified", "institution": "University of Cambridge", "country": "GB", "position": "Principal Researcher" } ], "n_formula": 17, "n_formula_1": 12, "n_ref_uni": 45, "n_ref": 90, "n_ref_all": 115, "n_fig": 5, "n_tab": 3, "L_tab": 960, "n_element_tab": 45, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 230, "n_element_tab_1": 38, "formula_len_all": 1110, "formula_len_all_1": 623, "len_all": 152112, "len_all_1": 69532, "len_abs": 1508, "len_title": 137, "len_sents": 47173, "len_sents_1": 35570, "n_sents": 278, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1481, "title": "Multimodal Lego: Model Merging and Fine-Tuning Across Topologies and Modalities in Biomedicine", "abs": "Learning holistic computational representations in physical, chemical or biological systems requires the ability to process information from different distributions and modalities within the same model. Thus, the demand for multimodal machine learning models has sharply risen for modalities that go beyond vision and language, such as sequences, graphs, time series, or tabular data. While there are many available multimodal fusion and alignment approaches, most of them require end-to-end training, scale quadratically with the number of modalities, cannot handle cases of high modality imbalance in the training set, or are highly topology-specific, making them too restrictive for many biomedical learning tasks. This paper presents Multimodal Lego (MM-Lego), a general-purpose fusion framework to turn any set of encoders into a competitive multimodal model with no or minimal fine-tuning. We achieve this by introducing a wrapper for any unimodal encoder that enforces shape consistency between modality representations. It harmonises these representations by learning features in the frequency domain to enable model merging with little signal interference. We show that MM-Lego 1) can be used as a model merging method which achieves competitive performance with end-to-end fusion models without any fine-tuning, 2) can operate on any unimodal encoder, and 3) is a model fusion method that, with minimal fine-tuning, surpasses all benchmarks in five out of seven datasets.", "keywords": [ "multimodal", "deep learning", "fusion", "biomedicine", "model merging" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 4, 3, 4 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "pGg658qADW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Lingyu Liu", "gender": "Female", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Lizi Liao", "gender": "Female", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Yaxiong Wang", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Zhedong Zheng", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Assistant Professor" }, { "name": "Li Zhu", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 38, "n_ref": 66, "n_ref_all": 83, "n_fig": 5, "n_tab": 3, "L_tab": 1192, "n_element_tab": 120, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1030, "n_element_tab_1": 97, "formula_len_all": 589, "formula_len_all_1": 574, "len_all": 126231, "len_all_1": 69053, "len_abs": 1480, "len_title": 129, "len_sents": 38945, "len_sents_1": 33043, "n_sents": 304, "n_sents_1": 249, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1486, "title": "Look, Compare and Draw: Differential Query Transformer for Automatic Oil Painting", "abs": "This work introduces a new approach to automatic oil painting that emphasizes the creation of dynamic and expressive brushstrokes. A pivotal challenge lies in mitigating the duplicate and common-place strokes, which often lead to less aesthetic outcomes. Inspired from the human painting process, i.e., observing, comparing, and drawing, we incorporate differential image analysis into a neural oil painting model, allowing the model to effectively concentrate on the incremental impact of successive brushstrokes. To operationalize this concept, we propose the Differential Query Transformer (DQ-Transformer), a new architecture that leverages differentially derived image representations enriched with positional encoding to guide the stroke prediction process. This integration enables the model to maintain heightened sensitivity to local details, resulting in more refined and nuanced stroke generation. Furthermore, we incorporate adversarial training into our framework, enhancing the accuracy of stroke prediction and thereby improving the overall realism and fidelity of the synthesized paintings. Extensive qualitative evaluations, complemented by a controlled user study, validate that our DQ-Transformer surpasses existing methods in both visual realism and artistic authenticity, typically achieving these results with fewer strokes. The stroke-by-stroke painting animations are available on our anonymous website: https://differential-query-painter.github.io/DQ-painter/ .", "keywords": [ "Automatic Oil Painting", "Stroke-based Rendering", "Style Transfer", "Sequence Prediction" ], "rating_list": [ 1, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 1, 3, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "pGMVuLvI5t", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "De-Gang Chen", "gender": "Male", "institution": "Chongqing University of Post and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Guoyin Wang", "gender": "Male", "institution": "Chongqing Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Sen Zhao", "gender": "Male", "institution": "Chongqing University of Post and Telecommunications", "country": "CN", "position": "Lecturer" }, { "name": "Shuyin Xia", "gender": "Male", "institution": "Chongqing University of Post and Telecommunications", "country": "CN", "position": "Full Professor" }, { "name": "Xinyu Lin", "gender": "Male", "institution": "Chongqing University of Post and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "guan Wang", "gender": "Female", "institution": "Chongqing University of Post and Telecommunications", "country": "", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 33, "n_ref": 43, "n_ref_all": 53, "n_fig": 2, "n_tab": 5, "L_tab": 2015, "n_element_tab": 208, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 341, "n_element_tab_1": 63, "formula_len_all": 1325, "formula_len_all_1": 603, "len_all": 120031, "len_all_1": 63118, "len_abs": 1665, "len_title": 138, "len_sents": 36170, "len_sents_1": 29769, "n_sents": 300, "n_sents_1": 246, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1669, "title": "GBO: A Multi-Granularity Optimization Algorithm via Granular-ball for Continuous Problems", "abs": "Optimization problems aim to find the optimal solution, which is becoming increasingly complex and difficult to solve. Traditional evolutionary optimization methods always overlook the granular characteristics of solution space. In the real scenario of numerous optimization, the solution space is typically partitioned into sub-regions characterized by varying degree distributions. These sub-regions present different granularity characteristics at search potential and difficulty. Considering the granular characteristics of the solution space, the number of coarse-grained regions is smaller than the number of points, so the calculation is more efficient. On the other hand, coarse-grained characteristics are not easily affected by fine-grained sample points, so the calculation is more robust. To this end, this paper proposes a new multi-granularity evolutionary optimization method, namely Granular-ball Optimization (GBO) algorithm, which characterizes and searches the solution space from coarse to fine. Specifically, using granular-balls instead of traditional points for optimization increases the diversity and robustness of the random search process. At the same time, the search range in different iteration processes is limited by the radius of granular-balls, covering the solution space from large to small. And the mechanism of granular-ball splitting is applied to continuously split and evolve the large granular-balls into smaller for refining the solution space. Extensive experiments on commonly used benchmarks have shown that GBO outperforms popular and advanced evolutionary algorithms. The code is available in the Supplementary Materials.", "keywords": [ "Evolutionary computation", "Granular-ball" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 5, 5, 3, 3 ] }, { "paper_id": "pGINxZWjK4", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Brent Griffin", "gender": "unknown", "institution": "Voxel51", "country": "US", "position": "Principal Scientist" }, { "name": "Jacob Marks", "gender": "unknown", "institution": "Liquid AI", "country": "US", "position": "Researcher" }, { "name": "Jason J Corso", "gender": "Male", "institution": "University of Michigan", "country": "US", "position": "Professor" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 33, "n_ref": 48, "n_ref_all": 71, "n_fig": 6, "n_tab": 5, "L_tab": 4240, "n_element_tab": 744, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2633, "n_element_tab_1": 509, "formula_len_all": 1006, "formula_len_all_1": 895, "len_all": 124372, "len_all_1": 64345, "len_abs": 3718, "len_title": 109, "len_sents": 31977, "len_sents_1": 26065, "n_sents": 217, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1930, "title": "Blind Coreset Selection: Efficient Pruning for Unlabeled Data", "abs": "Deep learning methods rely on massive data, resulting in substantial costs for storage, annotation, and model training.\nCoreset selection aims to select a representative subset of the data to train models with lower cost while ideally performing on par with the full data training.\nState-of-the-art coreset selection methods use carefully-designed criteria to quantify the importance of each data example using ground truth labels and dataset-specific training, then select examples whose scores lie in a certain range to construct a coreset.\nThese methods work well in their respective settings, however, they cannot consider candidate data that are initially unlabeled.\nThis limits the application of these methods, especially so considering that the majority of real-world data are unlabeled.\nTo that end, this paper explores the problem of coreset selection for unlabeled data.\nWe first motivate and formalize the problem of unlabeled coreset selection, which reduces annotation requirements to enable greater scale relative to label-based coreset selection.\nWe then develop an unlabeled coreset selection method, Blind Coreset Selection (BlindCS), that jointly considers overall data coverage on a distribution as well as the relative importance of each example based on redundancy.\nNotably, BlindCS does not use any model- or dataset-specific training, which increases coreset generalization and reduces computation relative to training-based coreset selection.\nWe evaluate BlindCS on four datasets and confirm the advance over several state-of-the-art methods that use labels and training, leading to a strong baseline for future research in unlabeled coreset selection.\nNotably, the BlindCS coreset for ImageNet achieves a higher accuracy than previous label-based coresets at a 90\\% prune rate, while removing annotation requirements for 1.15 million images.\nWe will make our code publicly available with the final paper.", "keywords": [ "Coreset selection", "Classification" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 2, 5 ] }, { "paper_id": "pG820nmDvy", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "John Kirchenbauer", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Manli Shu", "gender": "Female", "institution": "Salesforce Research", "country": "US", "position": "Researcher" }, { "name": "Monte Hoover", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Neel Jain", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Ramani Duraiswami", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "Full Professor" }, { "name": "Ryan Synk", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Tom Goldstein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Alexander Stein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 32, "n_ref": 49, "n_ref_all": 66, "n_fig": 6, "n_tab": 2, "L_tab": 685, "n_element_tab": 146, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 399, "n_element_tab_1": 53, "formula_len_all": 1167, "formula_len_all_1": 1167, "len_all": 101981, "len_all_1": 52177, "len_abs": 942, "len_title": 0, "len_sents": 27224, "len_sents_1": 23754, "n_sents": 213, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 945, "title": "Running Huge Context Windows On Tiny GPUs", "abs": "There is growing demand for large language models that can process hundreds of thousands or even millions of input tokens. Inference at this extreme scale demands significant computational resources and costs. To address the inference time costs associated with running self-attention based transformer language models on long contexts, we propose a tunable mechanism that reduces the cost of the forward pass by attending to only the most relevant tokens at every generation step using a top-k selection mechanism. We showcase the efficiency gains afforded by our method by performing inference on context windows up to 1M tokens using approximately 16GB of GPU RAM. Our experiments reveal that models are capable of handling the sparsity induced by the reduced number of keys and values. \nBy attending to less than 1% of input tokens, we achieve over 95% of model performance on common long context benchmarks (LM-Eval, AlpacaEval, and RULER).", "keywords": [ "attention", "transformers", "large language models", "long context" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 3, 1, 2 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "pFjzF7dIgg", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Sayanta Adhikari", "gender": "Male", "institution": "Indian Institute of Technology, Hyderabad", "country": "IN", "position": "MS student" }, { "name": "Vishnuprasadh Kumaravelu", "gender": "unknown", "institution": "Deakin University", "country": "AU", "position": "PhD student" }, { "name": "Srijith P K", "gender": "Male", "institution": "Indian Institute of Technology Hyderabad", "country": "", "position": "Associate Professor" } ], "n_formula": 32, "n_formula_1": 12, "n_ref_uni": 25, "n_ref": 59, "n_ref_all": 97, "n_fig": 6, "n_tab": 11, "L_tab": 6568, "n_element_tab": 904, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2883, "n_element_tab_1": 416, "formula_len_all": 3531, "formula_len_all_1": 584, "len_all": 198418, "len_all_1": 69425, "len_abs": 1417, "len_title": 102, "len_sents": 69313, "len_sents_1": 29404, "n_sents": 556, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1374, "title": "UnCLe: An Unlearning Framework for Continual Learning", "abs": "Recent advances in deep learning require models to exhibit continual learning capability, allowing them to learn new tasks and progressively accumulate knowledge without forgetting old tasks. Concurrently, there are growing concerns and regulatory requirements to meet privacy and safety by discarding some knowledge through machine unlearning. With the rapidly rising relevance of continual learning and machine unlearning, we consider them together under a unified framework in this paper. However, the conflicting nature of past data unavailability arising from continual learning makes it challenging to perform unlearning with existing methods which assume data availability. Moreover, in the proposed setup, where tasks are repeatedly learned and unlearned in a sequence, it is another challenge to maintain the stability of the tasks that need to be retained. To address these challenges, we propose UnCLe, an Unlearning Framework for Continual Learning designed to learn tasks incrementally and unlearn tasks without access to past data. To perform data-free unlearning, UnCLe leverages hypernetworks in conjunction with an unlearning objective that seeks to selectively align task-specific parameters with noise. Our experiments on popular benchmarks demonstrate UnCLe's consistent unlearning completeness and ability to preserve task stability over long sequences.", "keywords": [ "Machine Unlearning", "Continual Learning", "Hypernetworks" ], "rating_list": [ 5, 8, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "pEh1SXCgOc", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Chaofei Fan", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Eli Shlizerman", "gender": "unknown", "institution": "University of Washington", "country": "US", "position": "Associate Professor" }, { "name": "Jingyuan Li", "gender": "unknown", "institution": "University of Washington, Seattle", "country": "US", "position": "PhD student" }, { "name": "Mingfei Chen", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Trung Le", "gender": "Male", "institution": "University of Washington, Seattle", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 45, "n_ref": 88, "n_ref_all": 112, "n_fig": 9, "n_tab": 6, "L_tab": 688, "n_element_tab": 89, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 391, "n_element_tab_1": 61, "formula_len_all": 504, "formula_len_all_1": 242, "len_all": 151631, "len_all_1": 64746, "len_abs": 1435, "len_title": 138, "len_sents": 46154, "len_sents_1": 31278, "n_sents": 360, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1448, "title": "Brain-to-Text Decoding with Context-Aware Neural Representations and Large Language Models", "abs": "Decoding attempted speech from neural activity offers a promising avenue for restoring communication abilities in individuals with speech impairments. Previous studies have focused on mapping neural activity to text using phonemes as the intermediate target. \nWhile successful, decoding neural activity directly to phonemes ignores the context dependent nature of the neural activity-to-phoneme mapping in the brain, leading to suboptimal decoding performance.\nIn this work, we propose the use of diphone - an acoustic representation that captures the transitions between two phonemes - as the context-aware modeling target. We integrate diphones into existing phoneme decoding frameworks through a novel divide-and-conquer strategy in which we model the phoneme distribution by marginalizing over the diphone distribution. Our approach effectively leverages the enhanced context-aware representation of diphones while preserving the manageable class size of phonemes, a key factor in simplifying the subsequent phoneme-to-text conversion task. We demonstrate the effectiveness of our approach on the Brain-to-Text 2024 benchmark, where it achieves state-of-the-art Phoneme Error Rate (PER) of 15.34\\% compared to 16.62\\% PER of monophone-based decoding. When coupled with finetuned Large Language Models (LLMs), our method yields a Word Error Rate (WER) of 5.77\\%, significantly outperforming the 8.93\\% WER of the leading method in the benchmark.", "keywords": [ "brain-computer interfaces" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "pE0UM18TQh", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Felix den Breejen", "gender": "Male", "institution": "KAIST Graduate School of AI", "country": "", "position": "PhD student" }, { "name": "Sangmin Bae", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Se-Young Yun", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "Assistant Professor" }, { "name": "Stephen Cha", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 45, "n_ref": 78, "n_ref_all": 111, "n_fig": 15, "n_tab": 11, "L_tab": 6023, "n_element_tab": 847, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 153, "n_element_tab_1": 23, "formula_len_all": 347, "formula_len_all_1": 121, "len_all": 154891, "len_all_1": 54162, "len_abs": 1071, "len_title": 131, "len_sents": 37961, "len_sents_1": 25144, "n_sents": 330, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1074, "title": "Fine-tuned In-Context Learning Transformers are Excellent Tabular Data Classifiers.", "abs": "The recently introduced TabPFN pretrains an In-Context Learning (ICL) transformer on synthetic data to perform tabular data classification. In this work, we extend TabPFN to the fine-tuning setting, resulting in a significant performance boost. We also discover that fine-tuning enables ICL-transformers to create complex decision boundaries, a property regular neural networks do not have. Based on this observation, we propose to pretrain ICL-transformers on a new forest dataset generator which creates datasets that are unrealistic, but have complex decision boundaries. TabForest, the ICL-transformer pretrained on this dataset generator, shows better fine-tuning performance when pretrained on more complex datasets. Additionally, TabForest outperforms TabPFN on some real-world datasets when fine-tuning, despite having lower zero-shot performance due to the unrealistic nature of the pretraining datasets. By combining both dataset generators, we create TabForestPFN, an ICL-transformer that achieves excellent fine-tuning performance and good zero-shot performance.", "keywords": [ "tabular classification", "tabular in-context learning transformer", "fine-tuning" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "pDI03iK5Bf", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Chongjun Tu", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Christos-Savvas Bouganis", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Full Professor" }, { "name": "Jianjian Cao", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Mingzhu Shen", "gender": "Female", "institution": "Imperial College London", "country": "", "position": "PhD student" }, { "name": "Peng Ye", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Pengtao Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Tao Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yiren Zhao", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 43, "n_ref": 131, "n_ref_all": 160, "n_fig": 13, "n_tab": 13, "L_tab": 5377, "n_element_tab": 527, "n_fig_1": 10, "n_tab_1": 5, "L_tab_1": 2895, "n_element_tab_1": 288, "formula_len_all": 1340, "formula_len_all_1": 1016, "len_all": 156493, "len_all_1": 59935, "len_abs": 1541, "len_title": 191, "len_sents": 35051, "len_sents_1": 23876, "n_sents": 274, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1523, "title": "$\\Delta$-DiT: Accelerating Diffusion Transformers without training via Denoising Property Alignment", "abs": "Diffusion models are now commonly used for producing high-quality and diverse images, but the iterative denoising process is time-intensive, limiting their usage in real-time applications. As a result, various acceleration techniques have been developed, though these primarily target UNet-based architectures and are not directly applicable to Transformer-based diffusion models (DiT). To address the specific challenges of the DiT architecture, we first analyze the relationship between the depth of DiT blocks and the quality of image generation. While skipping blocks can lead to large degradations in generation quality, we propose the $\\Delta$-Cache method, which captures and stores the incremental changes of different blocks, thereby mitigating the performance gap and maintaining closer alignment with the original results. Our analysis indicates that the shallow DiT blocks primarily define the global structure of images such as compositions, and outlines, while the deep blocks refine details. Based on this, we introduce a denoising property alignment method that selectively bypasses computations of different blocks at various timesteps while preserving performance. Comprehensive experiments on PIXART-$\\alpha$ and DiT-XL demonstrate that $\\Delta$-DiT achieves a $1.6\\times$ speedup in 20-step generation and enhances performance in most cases. In the 4-step consistent model generation scenario, and with a more demanding $1.12\\times$ acceleration, our approach significantly outperforms existing methods.", "keywords": [ "Diffusion Model", "Training-Free", "Acceleration", "Diffusion Transformer" ], "rating_list": [ 3, 8, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 2 ] }, { "paper_id": "pDDODPtpx9", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Andras A Benczur", "gender": "unknown", "institution": "Institute for Computer Science and Control, Hungary", "country": "HU", "position": "Researcher" }, { "name": "Domokos M. Kelen", "gender": "Male", "institution": "E\u00f6tv\u00f6s Lorand University", "country": "HU", "position": "PhD student" }, { "name": "P\u00e9ter Kersch", "gender": "unknown", "institution": "Ericsson", "country": "", "position": "Researcher" }, { "name": "\u00c1d\u00e1m Jung", "gender": "Male", "institution": "E\u00f6tv\u00f6s Lorand University", "country": "HU", "position": "MS student" } ], "n_formula": 34, "n_formula_1": 13, "n_ref_uni": 74, "n_ref": 135, "n_ref_all": 189, "n_fig": 10, "n_tab": 12, "L_tab": 7329, "n_element_tab": 527, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 807, "n_element_tab_1": 9, "formula_len_all": 3406, "formula_len_all_1": 675, "len_all": 225619, "len_all_1": 68099, "len_abs": 1367, "len_title": 112, "len_sents": 63453, "len_sents_1": 32976, "n_sents": 445, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1373, "title": "Distribution-free Data Uncertainty for Neural Network Regression", "abs": "Quantifying uncertainty is an essential part of predictive modeling, especially in the context of high-stakes decision-making. While classification output includes data uncertainty by design in the form of class probabilities, the regression task generally aims only to predict the expected value of the target variable. Probabilistic extensions often assume parametric distributions around the expected value, optimizing the likelihood over the resulting explicit densities. However, using parametric distributions can limit practical applicability, making it difficult for models to capture skewed, multi-modal, or otherwise complex distributions. In this paper, we propose optimizing a novel nondeterministic neural network regression architecture for loss functions derived from a sample-based approximation of the continuous ranked probability score (CRPS), enabling a truly distribution-free approach by learning to sample from the target's aleatoric distribution, rather than predicting explicit densities. Our approach allows the model to learn well-calibrated, arbitrary uni- and multivariate output distributions. We evaluate the method on a variety of synthetic and real-world tasks, including uni- and multivariate problems, function inverse approximation, and standard regression uncertainty benchmarks. Finally, we make all experiment code publicly available.", "keywords": [ "deep learning", "uncertainty quantification", "regression uncertainty", "aleatoric uncertainty", "CRPS" ], "rating_list": [ 8, 5, 5, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 1, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "pD6TiCpyDR", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Chieh-Hsin Lai", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Researcher" }, { "name": "Satoshi Hayakawa", "gender": "Male", "institution": "Sony Group Corporation", "country": "JP", "position": "Research Scientist" }, { "name": "Yong-Hyun Park", "gender": "Male", "institution": "NAVER", "country": "KR", "position": "Intern" }, { "name": "Yuhta Takida", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Research Scientist" }, { "name": "Yuki Mitsufuji", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Lead Research Scientist" } ], "n_formula": 68, "n_formula_1": 24, "n_ref_uni": 30, "n_ref": 96, "n_ref_all": 141, "n_fig": 14, "n_tab": 5, "L_tab": 642, "n_element_tab": 28, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7723, "formula_len_all_1": 2000, "len_all": 203686, "len_all_1": 59759, "len_abs": 1244, "len_title": 74, "len_sents": 72578, "len_sents_1": 25497, "n_sents": 626, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1029, "title": "Jump Your Steps: Optimizing Sampling Schedule of Discrete Diffusion Models", "abs": "Diffusion models have seen notable success in continuous domains, leading to the development of discrete diffusion models (DDMs) for discrete variables. Despite recent advances, DDMs face the challenge of slow sampling speeds. While parallel sampling methods like $\\tau$-leaping accelerate this process, they introduce _Compounding Decoding Error_ (CDE), where discrepancies arise between the true distribution and the approximation from parallel token generation, leading to degraded sample quality. In this work, we present _Jump Your Steps_ (JYS), a novel approach that optimizes the allocation of discrete sampling timesteps by minimizing CDE without extra computational cost. More precisely, we derive a practical upper bound on CDE and propose an efficient algorithm for searching for the optimal sampling schedule. Extensive experiments across image, music, and text generation show that JYS significantly improves sampling quality, establishing it as a versatile framework for enhancing DDM performance for fast sampling.", "keywords": [ "Discrete diffusion models", "Efficient sampling" ], "rating_list": [ 8, 8, 5, 5 ], "soundness_list": [ 3, 4, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "pCx6DYN43D", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Miguel Eckstein", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Tsu-Jui Fu", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Xiujun Li", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Yujie Lu", "gender": "unknown", "institution": "UC Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "William Yang Wang", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 15, "n_ref": 36, "n_ref_all": 54, "n_fig": 6, "n_tab": 4, "L_tab": 3752, "n_element_tab": 617, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3878, "n_element_tab_1": 631, "formula_len_all": 281, "formula_len_all_1": 281, "len_all": 136858, "len_all_1": 58229, "len_abs": 224, "len_title": 109, "len_sents": 27295, "len_sents_1": 22140, "n_sents": 185, "n_sents_1": 142, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1074, "title": "Compact Multimodal Context Represenations Using Visual Tokens", "abs": "The rapid progress in Multimodal Large Language Models (MLLMs) has significantly advanced their ability to process and understand complex visual and textual information. However, the integration of multiple images and extensive textual contexts remains a challenge due to the inherent limitation of the models' capacity to handle long input sequences efficiently. In this paper, we introduce SEEKER, a multimodal large language model designed to tackle this issue. SEEKER aims to optimize the compact encoding of long text by compressing the text sequence into the visual pixel space via images, enabling the model to handle long text within a fixed token-length budget efficiently. Our empirical experiments on six long-context multimodal tasks demonstrate that SEEKER can leverage fewer image tokens to convey the same amount of textual information compared with the OCR-based approach, and is more efficient in understanding long-form multimodal input and generating long-form textual output, outperforming all existing proprietary and open-source MLLMs by large margins.", "keywords": [ "Vision and Language Models", "Compact Textual Information Encoding in Visual Space" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 1, 2, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "pCj2sLNoJq", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Arjun V Sudhakar", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "MS student" }, { "name": "Hadi Nekoei", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, University of Montreal", "country": "", "position": "PhD student" }, { "name": "Janarthanan Rajendran", "gender": "Male", "institution": "Dalhousie University", "country": "", "position": "Assistant Professor" }, { "name": "Mathieu Reymond", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Postdoc" }, { "name": "Miao Liu", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Research Staff Member" }, { "name": "Sarath Chandar", "gender": "Male", "institution": "\u00c9cole Polytechnique de Montr\u00e9al", "country": "CA", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 4, "n_ref_uni": 54, "n_ref": 96, "n_ref_all": 117, "n_fig": 13, "n_tab": 3, "L_tab": 784, "n_element_tab": 55, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 473, "formula_len_all_1": 452, "len_all": 190200, "len_all_1": 78073, "len_abs": 1330, "len_title": 73, "len_sents": 57762, "len_sents_1": 39277, "n_sents": 436, "n_sents_1": 295, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 25, "L_abs": 1340, "title": "A Generalist Hanabi Agent", "abs": "Traditional multi-agent reinforcement learning (MARL) systems can develop cooperative strategies through repeated interactions. However, these systems are unable to perform well on any other setting than the one they have been trained on, and struggle to successfully cooperate with unfamiliar collaborators. This is particularly visible in the Hanabi benchmark, a popular 2-to-5 player cooperative card-game which requires complex reasoning and precise assistance to other agents. Current MARL agents for Hanabi can only learn one specific game-setting (e.g., 2-player games), and play with the same algorithmic agents. This is in stark contrast to humans, who can quickly adjust their strategies to work with unfamiliar partners or situations. In this paper, we introduce a generalist agent for Hanabi, designed to overcome these limitations. We reformulate the task using text, as language has been shown to improve transfer. We then propose a distributed MARL algorithm that copes with the resulting dynamic observation- and action-space. In doing so, our agent is the first that can play all game settings concurrently, and extend strategies learned from one setting to other ones. As a consequence, our agent also demonstrates the ability to collaborate with different algorithmic agents ---agents that are themselves unable to do so.", "keywords": [ "Multi-Agent Reinforcement Learning (MARL)", "Cooperative game", "Multi Agent Text-based game" ], "rating_list": [ 8, 8, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3, 1 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 4, 5, 4, 3, 3 ] }, { "paper_id": "pCX1kZ0qHL", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Tian Qin", "gender": "unknown", "institution": "Lehigh University", "country": "US", "position": "PhD student" }, { "name": "Wei-Min Huang", "gender": "unknown", "institution": "Lehigh University", "country": "US", "position": "Full Professor" } ], "n_formula": 52, "n_formula_1": 26, "n_ref_uni": 20, "n_ref": 49, "n_ref_all": 70, "n_fig": 4, "n_tab": 8, "L_tab": 3027, "n_element_tab": 419, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1973, "n_element_tab_1": 173, "formula_len_all": 5142, "formula_len_all_1": 2317, "len_all": 130614, "len_all_1": 64644, "len_abs": 1075, "len_title": 86, "len_sents": 41258, "len_sents_1": 24565, "n_sents": 385, "n_sents_1": 230, "venue": "ICLR2025", "decision": "withdraw", "L_title": 38, "L_abs": 1046, "title": "Riemann-Lebesgue Forest for Regression", "abs": "We propose a novel ensemble method called Riemann-Lebesgue Forest (RLF) for regression. The core idea in RLF is to mimic the way how a measurable function can be approximated by partitioning its range into a few intervals. With this idea in mind, we develop a new tree learner named Riemann-Lebesgue Tree (RLT) which has a chance to perform Lebesgue type cutting,i.e splitting the node from response Y\n at certain non-terminal nodes. In other words, we introduce the \"splitting type randomness\" in our ensemble method. We show that the optimal Lebesgue type cutting results in larger variance reduction in response Y than ordinary CART cutting (an analogue of Riemann partition). Such property is beneficial to the ensemble part of RLF. We also generalize the asymptotic normality of RLF under different parameter settings. Two one-dimensional examples are provided to illustrate the flexibility of RLF. The competitive performance of RLF against original random forest is demonstrated by experiments in simulation data and real world datasets.", "keywords": [ "RandomForest; Lebesgue Partition; asymptotic normality;" ], "rating_list": [ 5, 8, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "pBugl1EIkm", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Bochuan Cao", "gender": "unknown", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Changjiang Li", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Jiacheng Liang", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Jinghui Chen", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Ting Wang", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 80, "n_ref_all": 100, "n_fig": 9, "n_tab": 3, "L_tab": 2114, "n_element_tab": 68, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1139, "n_element_tab_1": 47, "formula_len_all": 248, "formula_len_all_1": 286, "len_all": 145798, "len_all_1": 63968, "len_abs": 1729, "len_title": 102, "len_sents": 42413, "len_sents_1": 27996, "n_sents": 362, "n_sents_1": 234, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1509, "title": "Your Agent Can Defend Itself against Backdoor Attacks", "abs": "Intelligent agents powered by large language models (LLMs) have gained surging popularity due to their versatile and customizable capabilities across diverse environments. However, recent studies also reveal their critical vulnerability: LLM agents are highly susceptible to backdoor attacks during training or fine-tuning. Such compromised agents can subsequently be manipulated to execute malicious operations when presented with specific triggers in their inputs or environments. To address this pressing risk, we present ReAgent, a novel defense against a range of backdoor attacks on LLM-based agents. Intuitively, backdoor attacks often result in inconsistencies among the user's instruction, the agent's planning, and its execution. Drawing on this insight, ReAgent employs a two-level approach to detect potential backdoors. At the execution level, ReAgent verifies consistency between the agent's thoughts and actions; at the planning level, ReAgent leverages the agent's capability to reconstruct the instruction based on\nits thought trajectory, checking for consistency between the reconstructed instruction and the user's instruction. Extensive evaluation demonstrates ReAgent's effectiveness against various backdoor attacks across diverse tasks. For instance, ReAgent reduces the attack success rate by up to 90\\% in database operation tasks, outperforming existing defenses by large margins. This work reveals the potential of utilizing compromised agents themselves to mitigate backdoor risks.", "keywords": [ "LLM Agent", "Backdoor Attack", "Backdoor Defense" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "pBqOH2g6K1", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Biplab Sikdar", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Jiayu Li", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Kevin Yee", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Co-PI" }, { "name": "Uzair Javaid", "gender": "Male", "institution": "Betterdata", "country": "SG", "position": "Co-founder & CEO" }, { "name": "Zilong Zhao", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 9, "n_ref_uni": 47, "n_ref": 96, "n_ref_all": 112, "n_fig": 5, "n_tab": 15, "L_tab": 12468, "n_element_tab": 1080, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1518, "n_element_tab_1": 116, "formula_len_all": 397, "formula_len_all_1": 1168, "len_all": 182852, "len_all_1": 67661, "len_abs": 1405, "len_title": 111, "len_sents": 48069, "len_sents_1": 30948, "n_sents": 331, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1478, "title": "TAEGAN: Generating Synthetic Tabular Data for Data Augmentation", "abs": "Synthetic tabular data generation has gained significant attention for its potential in data augmentation, software testing and privacy-preserving data sharing. However, most research has primarily focused on larger datasets and evaluating their quality in terms of metrics like column-wise statistical distributions and inter-feature correlations, while often overlooking its utility for data augmentation, particularly for datasets whose data is scarce. In this paper, we propose Tabular Auto-Encoder Generative Adversarial Network (TAEGAN), an improved GAN-based framework for generating high-quality tabular data. Although large language models (LLMs)-based methods represent the state-of-the-art in synthetic tabular data generation, they are often overkill for small datasets due to their extensive size and complexity. TAEGAN employs a masked auto-encoder as the generator, which for the first time introduces the power of self-supervised pre-training in tabular data generation so that essentially exposes the networks to more information. We extensively evaluate TAEGAN against five state-of-the-art synthetic tabular data generation algorithms. Results from 10 datasets show that TAEGAN outperforms existing deep-learning-based tabular data generation models on 9 out of 10 datasets on the machine learning efficacy and achieves superior data augmentation performance on 7 out of 8 smaller datasets. Code is available at:\nhttps://anonymous.4open.science/r/taegan-2AB4", "keywords": [ "GAN", "synthetic tabular data", "data augmentation" ], "rating_list": [ 3, 8, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 1, 3, 2, 1, 1 ], "contribution_list": [ 2, 3, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 2, 4 ] }, { "paper_id": "pBZntPrdrI", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Ben Glocker", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Full Professor" }, { "name": "Charles Jones", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Daniel C. Castro", "gender": "unknown", "institution": "Microsoft Research", "country": "", "position": "Senior Researcher" }, { "name": "Fabio De Sousa Ribeiro", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "Postdoc" }, { "name": "Melanie Bernhardt", "gender": "Female", "institution": "Imperial College London, Imperial College London", "country": "", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 7, "n_ref_uni": 60, "n_ref": 107, "n_ref_all": 126, "n_fig": 7, "n_tab": 3, "L_tab": 873, "n_element_tab": 104, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 692, "n_element_tab_1": 107, "formula_len_all": 1393, "formula_len_all_1": 499, "len_all": 156647, "len_all_1": 64287, "len_abs": 989, "len_title": 119, "len_sents": 41106, "len_sents_1": 31496, "n_sents": 308, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 992, "title": "Rethinking Fair Representation Learning for Performance-Sensitive Tasks", "abs": "We investigate the prominent class of fair representation learning methods for bias mitigation. Using causal reasoning to define and formalise different sources of dataset bias, we reveal important implicit assumptions inherent to these methods. We prove fundamental limitations on fair representation learning when evaluation data is drawn from the same distribution as training data and run experiments across a range of medical modalities to examine the performance of fair representation learning under distribution shifts. Our results explain apparent contradictions in the existing literature and reveal how rarely considered causal and statistical aspects of the underlying data affect the validity of fair representation learning. We raise doubts about current evaluation practices and the applicability of fair representation learning methods in performance-sensitive settings. We argue that fine-grained analysis of dataset biases should play a key role in the field moving forward.", "keywords": [ "Fairness", "Causality", "Representation Learning", "Distribution Shift" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 2, 3, 5, 4 ] }, { "paper_id": "pBQs8kQm63", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Hongyi Zhou", "gender": "unknown", "institution": ", Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Jingwei Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Jingzhao Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 240, "n_formula_1": 29, "n_ref_uni": 28, "n_ref": 61, "n_ref_all": 65, "n_fig": 1, "n_tab": 3, "L_tab": 288, "n_element_tab": 35, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 468, "n_element_tab_1": 10, "formula_len_all": 23602, "formula_len_all_1": 1434, "len_all": 221122, "len_all_1": 69483, "len_abs": 952, "len_title": 148, "len_sents": 61179, "len_sents_1": 29893, "n_sents": 660, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 983, "title": "Finite Sample Analyses for Continuous-time Linear Systems: System Identification and Online Control", "abs": "Real world evolves in continuous time but computations are done from finite samples. Therefore, we study algorithms using finite observations in continuous-time linear dynamical systems. \nWe first study the system identification problem, and propose a first non-asymptotic error analysis with finite observations. Our algorithm identifies system parameters without needing integrated observations over certain time intervals, making it more practical for real-world applications. Further we propose a lower bound result that shows our estimator is provably optimal up to constant factors.\nMoreover, we apply the above algorithm to online control regret analysis for continuous-time linear system. Our system identification method allows us explore more efficiently, enabling the swift detection of ineffective policies. We achieve a regret of $\\mathcal{O}(\\sqrt{T})$ over a single $T$-time horizon in a controllable system, requiring only $\\mathcal{O}(T)$ observations of the system.", "keywords": [ "online control", "system identification", "continuous-time linear system" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "pB3KeBCnQs", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Aarash Feizi", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "Golnoosh Farnadi", "gender": "Female", "institution": "McGill University", "country": "CA", "position": "Assistant Professor" }, { "name": "Rohan Sukumaran", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "MS student" }, { "name": "Adriana Romero", "gender": "Female", "institution": "Mila", "country": "CA", "position": "Core Industry Member" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 30, "n_ref": 48, "n_ref_all": 60, "n_fig": 7, "n_tab": 4, "L_tab": 2365, "n_element_tab": 135, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 3789, "n_element_tab_1": 153, "formula_len_all": 222, "formula_len_all_1": 222, "len_all": 104124, "len_all_1": 56304, "len_abs": 1322, "len_title": 94, "len_sents": 28639, "len_sents_1": 23327, "n_sents": 210, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1327, "title": "FairLoRA: Unpacking Bias Mitigation in Vision Models with Fairness-Regularized Low-Rank Adaptation", "abs": "Recent advances in parameter-efficient fine-tuning methods, such as Low Rank Adaptation (LoRA), have gained significant attention for their ability to efficiently adapt large foundational models to various downstream tasks. These methods are appreciated for achieving performance comparable to full fine-tuning on aggregate-level metrics, while significantly reducing computational costs. To systematically address fairness in LLMs previous studies fine-tune on fairness specific data using a larger LoRA rank than typically used. In this paper, we introduce FairLoRA, a novel fairness-specific regularizer for LoRA aimed at reducing performance disparities across data subgroups by minimizing per-class variance in loss. To the best of our knowledge, we are the first to introduce a fairness based finetuning through LoRA. Our results demonstrate that the need for higher ranks to mitigate bias is not universal; it depends on factors such as the pre-trained model, dataset, and task. More importantly, we systematically evaluate FairLoRA across various vision models, including ViT, DiNO, and CLIP, in scenarios involving distribution shifts. We further emphasize the necessity of using multiple fairness metrics to obtain a holistic assessment of fairness, rather than relying solely on the metric optimized during training.", "keywords": [ "LoRA", "ViTs", "Fairness", "Vision Models", "Bias mitigation" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "pB1XSj2y4X", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Jinkyoo Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Martin Ester", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "Full Professor" }, { "name": "Minsu Kim", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Postdoc" }, { "name": "Seonghwan Seo", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sungsoo Ahn", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Tony Shen", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "PhD student" }, { "name": "Woo Youn Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 37, "n_formula_1": 13, "n_ref_uni": 54, "n_ref": 137, "n_ref_all": 184, "n_fig": 17, "n_tab": 12, "L_tab": 4277, "n_element_tab": 276, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 1979, "n_element_tab_1": 143, "formula_len_all": 4261, "formula_len_all_1": 1096, "len_all": 219328, "len_all_1": 63747, "len_abs": 1567, "len_title": 101, "len_sents": 59174, "len_sents_1": 27219, "n_sents": 507, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1574, "title": "Generative Flows on Synthetic Pathway for Drug Design", "abs": "Generative models in drug discovery have recently gained attention as efficient alternatives to brute-force virtual screening. However, most existing models do not account for synthesizability, limiting their practical use in real-world scenarios. In this paper, we propose RxnFlow, which sequentially assembles molecules using predefined molecular building blocks and chemical reaction templates to constrain the synthetic chemical pathway. We then train on this sequential generating process with the objective of generative flow networks (GFlowNets) to generate both highly rewarded and diverse molecules. To mitigate the large action space of synthetic pathways in GFlowNets, we implement a novel action space subsampling method. This enables RxnFlow to learn generative flows over extensive action spaces comprising combinations of 1.2 million building blocks and 71 reaction templates without significant computational overhead. Additionally, RxnFlow can employ modified or expanded action spaces for generation without retraining, allowing for the introduction of additional objectives or the incorporation of newly discovered building blocks. We experimentally demonstrate that RxnFlow outperforms existing reaction-based and fragment-based models in pocket-specific optimization across various target pockets. Furthermore, RxnFlow achieves state-of-the-art performance on CrossDocked2020 for pocket-conditional generation, with an average Vina score of \u20138.85 kcal/mol and 34.8% synthesizability. Code is available at https://anonymous.4open.science/r/RxnFlow-B13E/.", "keywords": [ "GFlowNet", "synthesizability", "structure-based drug design", "molecule optimization" ], "rating_list": [ 5, 6, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "pAkQhhn4vB", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 1, "author_info_list": [ { "name": "Lun Wang", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 29, "n_formula_1": 29, "n_ref_uni": 25, "n_ref": 61, "n_ref_all": 68, "n_fig": 2, "n_tab": 7, "L_tab": 604, "n_element_tab": 55, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 400, "n_element_tab_1": 23, "formula_len_all": 2802, "formula_len_all_1": 2944, "len_all": 86730, "len_all_1": 49132, "len_abs": 1228, "len_title": 125, "len_sents": 24164, "len_sents_1": 19791, "n_sents": 213, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1233, "title": "Revisit Micro-batch Clipping: Adaptive Data Pruning via Gradient Manipulation", "abs": "Micro-batch clipping, a gradient clipping method, has recently shown potential in enhancing auto-speech recognition (ASR) model performance. However, the underlying mechanism behind this improvement remains mysterious, particularly the observation that only certain micro-batch sizes are beneficial. In this paper, we make the first attempt to explain this phenomenon. Inspired by recent data pruning research, we assume that specific training samples may impede model convergence during certain training phases. Under this assumption, the convergence analysis shows that micro-batch clipping can improve the convergence rate asymptotically at the cost of an additional constant bias that does not diminish with more training iterations. The bias is dependent on a few factors and can be minimized at specific micro-batch size, thereby elucidating the existence of the sweet-spot micro-batch size observed previously. We also verify the effectiveness of micro-batch clipping beyond speech models on vision and language models, and show promising performance gains in these domains. An exploration of potential limitations shows that micro-batch clipping is less effective when training data originates from multiple distinct domains.", "keywords": [ "Micro-batch Clipping", "Convergence Analysis" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "pAQzEY7M03", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Jian Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Qing Huang", "gender": "Female", "institution": "South China University of Technology", "country": "CN", "position": "MS student" }, { "name": "Runyi Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Xuanyu Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zecheng Tang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Zhipei Xu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 45, "n_ref": 113, "n_ref_all": 141, "n_fig": 12, "n_tab": 11, "L_tab": 2408, "n_element_tab": 253, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1267, "n_element_tab_1": 149, "formula_len_all": 349, "formula_len_all_1": 346, "len_all": 242845, "len_all_1": 61303, "len_abs": 1396, "len_title": 151, "len_sents": 86651, "len_sents_1": 29387, "n_sents": 610, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 1419, "title": "FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models", "abs": "The rapid development of generative AI is a double-edged sword, which not only facilitates content creation but also makes image manipulation easier and more difficult to detect. Although current image forgery detection and localization (IFDL) methods are generally effective, they tend to face two challenges: \\textbf{1)} black-box nature with unknown detection principle, \\textbf{2)} limited generalization across diverse tampering methods (e.g., Photoshop, DeepFake, AIGC-Editing). To address these issues, we propose the explainable IFDL task and design FakeShield, a multi-modal framework capable of evaluating image authenticity, generating tampered region masks, and providing a judgment basis based on pixel-level and image-level tampering clues. Additionally, we leverage GPT-4o to enhance existing IFDL datasets, creating the Multi-Modal Tamper Description dataSet (MMTD-Set) for training FakeShield's tampering analysis capabilities. Meanwhile, we incorporate a Domain Tag-guided Explainable Forgery Detection Module (DTE-FDM) and a Multi-modal Forgery Localization Module (MFLM) to address various types of tamper detection interpretation and achieve forgery localization guided by detailed textual descriptions. Extensive experiments demonstrate that FakeShield effectively detects and localizes various tampering techniques, offering an explainable and superior solution compared to previous IFDL methods.", "keywords": [ "Image Forgery Detection and Localization", "Multi-modal Large Language Model", "Tamper Detection" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 5, 5, 5 ] }, { "paper_id": "pA8oI8a00l", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Siyuan Liang", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Xiaochun Cao", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Xiaojun Jia", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Xinwei Liu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yuan Xun", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 27, "n_ref": 49, "n_ref_all": 74, "n_fig": 8, "n_tab": 10, "L_tab": 5055, "n_element_tab": 325, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1897, "n_element_tab_1": 173, "formula_len_all": 968, "formula_len_all_1": 500, "len_all": 146818, "len_all_1": 69991, "len_abs": 1573, "len_title": 156, "len_sents": 47259, "len_sents_1": 33486, "n_sents": 283, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1583, "title": "CleanerCLIP: Fine-grained Counterfactual Semantic Augmentation for Backdoor Defense in Contrastive Learning", "abs": "Multimodal contrastive models like CLIP are increasingly vulnerable to data-poisoning backdoor attacks. Existing defense methods primarily target the pretraining phase. However, with the rise of open-source communities, pretrained models are now freely available for download and fine-tuning. These models may carry unknown security risks, posing significant threats to downstream users. This highlights the need for lightweight defense strategies tailored specifically for the fine-tuning stage. Current defenses during fine-tuning include: finetuning with clean data; and using unimodal self-supervised techniques like CleanCLIP, which has represented the state-of-the-art (SOTA). However, these methods rely on strengthening clean feature representations to mitigate attacks, making them ineffective against more stealthy backdoor techniques, such as BadCLIP, which leverage covert toxic features.\nTo overcome this limitation, we propose a finetuning defense mechanism based on fine-grained counterfactual text semantic augmentation. By modifying small portions of text during fine-tuning, our approach disrupts the association between backdoor triggers and target features.\nWe evaluate our method against six attack algorithms and conduct comprehensive zero-shot classification on ImageNet1K. Experimental results demonstrate that our method achieves SOTA performance in fine-tuning defense. Specifically, when facing the novel BadCLIP attack, our method surpasses CleanCLIP, reducing the Attack Success Rate (ASR) by 52.02% in the Top-1 and 63.88% in the Top-10 classifications.", "keywords": [ "backdoor defense", "contrastive learning", "multimodal pretrained models" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "p9OsTj0nMP", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Alexander Nikulin", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "PhD student" }, { "name": "Alexey Zemtsov", "gender": "Male", "institution": "Moscow State Institute of Steel and Alloys", "country": "", "position": "Undergrad student" }, { "name": "Ilya Zisman", "gender": "Not Specified", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Vladislav Kurenkov", "gender": "Male", "institution": "Tinkoff", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 64, "n_ref": 127, "n_ref_all": 149, "n_fig": 9, "n_tab": 7, "L_tab": 1851, "n_element_tab": 239, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 439, "n_element_tab_1": 43, "formula_len_all": 51, "formula_len_all_1": 305, "len_all": 161476, "len_all_1": 78530, "len_abs": 1155, "len_title": 130, "len_sents": 39192, "len_sents_1": 32290, "n_sents": 314, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1170, "title": "XLand-100B: A Large-Scale Multi-Task Dataset for In-Context Reinforcement Learning", "abs": "Following the success of the in-context learning paradigm in large-scale language and computer vision models, the recently emerging field of in-context reinforcement learning is experiencing a rapid growth. However, its development has been held back by the lack of challenging benchmarks, as all the experiments have been carried out in simple environments and on small-scale datasets. We present **XLand-100B**, a large-scale dataset for in-context reinforcement learning based on the XLand-MiniGrid environment, as a first step to alleviate this problem. It contains complete learning histories for nearly $30,000$ different tasks, covering $100$B transitions and $2.5$B episodes. It took $50,000$ GPU hours to collect the dataset, which is beyond the reach of most academic labs. Along with the dataset, we provide the utilities to reproduce or expand it even further. We also benchmark common in-context RL baselines and show that they struggle to generalize to novel and diverse tasks. With this substantial effort, we aim to democratize research in the rapidly growing field of in-context reinforcement learning and provide a solid foundation for further scaling.", "keywords": [ "in-context reinforcement learning", "reinforcement learning", "xland", "minigrid" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 4, 4, 3 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "p97nsl3Fvq", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Qitao Qin", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Zhibo Chu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Zichong Wang", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 49, "n_ref_all": 70, "n_fig": 1, "n_tab": 4, "L_tab": 991, "n_element_tab": 136, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 2245, "n_element_tab_1": 196, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 97254, "len_all_1": 69526, "len_abs": 1554, "len_title": 131, "len_sents": 34222, "len_sents_1": 34055, "n_sents": 235, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1699, "title": "Leveraging Prior Experience: An Expandable Auxiliary Knowledge Base for Text-to-SQL", "abs": "Large Language Models (LLMs) exhibit impressive problem-solving skills across many tasks, but they still underperform compared to humans in various downstream applications, such as text-to-SQL. On the BIRD benchmark leaderboard, human performance achieves an accuracy of 92.96\\%, whereas the top-performing method reaches only 72.39\\%. Notably, these state-of-the-art (SoTA) methods predominantly rely on in-context learning to simulate human-like reasoning. However, they overlook a critical human skill: continual learning. Inspired by the educational practice of maintaining mistake notebooks during our formative years, we propose LPE-SQL ($\\underline{\\textbf{L}}$everaging $\\underline{\\textbf{P}}$rior $\\underline{\\textbf{E}}$xperience: An Expandable Auxiliary Knowledge Base for Text-to-$\\underline{\\textbf{SQL}}$), a novel framework designed to augment LLMs by enabling continual learning without requiring parameter fine-tuning. LPE-SQL consists of four modules that $\\textbf{i)}$ retrieve relevant entries, $\\textbf{ii)}$ efficient sql generation, $\\textbf{iii)}$ generate the final result through a cross-consistency mechanism and $\\textbf{iv)}$ log successful and failed tasks along with their reasoning processes or reflection-generated tips. Importantly, the core module of LPE-SQL is the fourth one, while the other modules employ foundational methods, allowing LPE-SQL to be easily integrated with SoTA technologies to further enhance performance. Our experimental results demonstrate that this continual learning approach yields substantial performance gains, with the smaller Llama-3.1-70B model with surpassing the performance of the larger Llama-3.1-405B model using SoTA methods.", "keywords": [ "Large language models; Text to SQL; In-context learning; Continuous learning;" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "p8sr9kfUbQ", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 6, "author_info_list": [ { "name": "Christoph Lippert", "gender": "Male", "institution": "Hasso Plattner Institute", "country": "DE", "position": "Full Professor" }, { "name": "Eliot Wong-Toi", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Eshant English", "gender": "unknown", "institution": "Hasso Plattner Institute", "country": "DE", "position": "PhD student" }, { "name": "Matteo Fontana", "gender": "Male", "institution": "Royal Holloway University of London", "country": "GB", "position": "Lecturer" }, { "name": "Padhraic Smyth", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Full Professor" }, { "name": "Stephan Mandt", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Associate Professor" } ], "n_formula": 34, "n_formula_1": 13, "n_ref_uni": 30, "n_ref": 54, "n_ref_all": 94, "n_fig": 4, "n_tab": 6, "L_tab": 1403, "n_element_tab": 241, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 163, "n_element_tab_1": 24, "formula_len_all": 3543, "formula_len_all_1": 1047, "len_all": 144325, "len_all_1": 72356, "len_abs": 1019, "len_title": 114, "len_sents": 45768, "len_sents_1": 33945, "n_sents": 404, "n_sents_1": 281, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1022, "title": "JANET: Joint Adaptive predictioN-region Estimation for Time-series", "abs": "Conformal prediction provides machine learning models with prediction sets that offer theoretical guarantees, but the underlying assumption of exchangeability limits its applicability to time series data. Furthermore, existing approaches struggle to handle multi-step ahead prediction tasks, where uncertainty estimates across multiple future time points are crucial. We propose JANET (Joint Adaptive predictioN-region Estimation for Time-series), a novel framework for constructing conformal prediction regions that are valid for both univariate and multivariate time series. JANET generalises the inductive conformal framework and efficiently produces joint prediction regions with controlled K-familywise error rates, enabling flexible adaptation to specific application needs. Our empirical evaluation demonstrates JANET's superior performance in multi-step prediction tasks across diverse time series datasets, highlighting its potential for reliable and interpretable uncertainty quantification in sequential data.", "keywords": [ "Joint Prediction Region", "Conformal Prediction", "Uncertainty in Time Series Forecasting" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "p8qhVIo980", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Branislav Pecher", "gender": "Male", "institution": "Kempelen Institute of Intelligent Technologies", "country": "SK", "position": "PhD student" }, { "name": "Ivan Srba", "gender": "Male", "institution": "Kempelen Institute of Intelligent Technologies", "country": "", "position": "Researcher" }, { "name": "Joaquin Vanschoren", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Associate Professor" }, { "name": "Maria Bielikova", "gender": "Female", "institution": "Kempelen Institute of Intelligent Technologies", "country": "", "position": "Principal Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 75, "n_ref": 153, "n_ref_all": 214, "n_fig": 27, "n_tab": 16, "L_tab": 15394, "n_element_tab": 756, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1743, "n_element_tab_1": 85, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 346800, "len_all_1": 63808, "len_abs": 1285, "len_title": 74, "len_sents": 109326, "len_sents_1": 33180, "n_sents": 761, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1289, "title": "Automatic Combination of Sample Selection Strategies for Few-Shot Learning", "abs": "In few-shot learning, such as meta-learning, few-shot fine-tuning or in-context learning, the selection of samples has a significant impact on the performance of the trained model. Although many sample selection strategies are employed and evaluated in typical supervised settings, their impact on the performance of few-shot learning is largely unknown. In this paper, we investigate the impact of 20 sample selection strategies on the performance of 5 representative few-shot learning approaches over 8 image and 6 text datasets. We propose a new method for Automatic Combination of SamplE Selection Strategies (ACSESS), to leverage the strengths and complementarity of the individual strategies in order to select more impactful samples. The experimental results show that our method consistently outperforms all individual selection strategies. We also show that the majority of existing strategies strongly depend on modality, dataset characteristics and few-shot learning approach, while improving performance especially on imbalanced and noisy datasets. Lastly, we show that sample selection strategies work well even on smaller datasets and provide larger benefit when selecting a lower number of shots, while frequently regressing to random selection with higher numbers of shots.", "keywords": [ "sample selection", "few-shot learning", "in-context learning", "large language models", "meta-learning", "few-shot fine-tuning", "data-centric" ], "rating_list": [ 6, 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 3, 4, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 4, 3 ] }, { "paper_id": "p8UoIVAcU3", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Fan Zhou", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Junlong Li", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "", "position": "MS student" }, { "name": "Junxian He", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Wei Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Xiwen Zhang", "gender": "Male", "institution": "Helixon Research", "country": "CN", "position": "Researcher" }, { "name": "Yu Cheng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 63, "n_ref_all": 87, "n_fig": 10, "n_tab": 4, "L_tab": 1844, "n_element_tab": 377, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 672, "n_element_tab_1": 137, "formula_len_all": 358, "formula_len_all_1": 237, "len_all": 147667, "len_all_1": 68286, "len_abs": 1668, "len_title": 107, "len_sents": 45456, "len_sents_1": 34106, "n_sents": 288, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1780, "title": "Diving into Self-Evolve Training for Multimodal Reasoning", "abs": "Reasoning ability is essential for Large Multimodal Models (LMMs). \nIn the absence of multimodal chain-of-thought annotated data, self-evolving training, where the model learns from its own outputs, has emerged as an effective and scalable approach for enhancing reasoning abilities. \nDespite its growing usage, a comprehensive understanding of self-evolving training, particularly in the context of multimodal reasoning, remains limited. In this paper, we delve into the intricacies of self-evolving training for multimodal reasoning, pinpointing three key factors: $\\textbf{Training Method}$, $\\textbf{Reward Model}$, and $\\textbf{Prompt Variation}$. We systematically examine each factor and explore how various configurations affect the training's effectiveness. Our analysis leads to a set of best practices for each factor, aimed at optimizing multimodal reasoning.\nFurthermore, we explore the $\\textbf{Self-Evolution Dynamics}$ during training and the impact of automatic balancing mechanisms in boosting performance. After all the investigations, we present a final recipe for self-evolving training in multimodal reasoning, encapsulating these design choices into a framework we call M-STAR ($\\textbf{M}$ultimodal $\\textbf{S}$elf-evolving $\\textbf{T}$r$\\textbf{a}$ining for $\\textbf{R}$easoning), built on MiniCPM-V 2.5. \nM-STAR achieves 59.5% accuracy on MathVista, surpassing the pre-evolved model by 6.9% absolutely without using additional human annotations. \nWe believe this study fills a significant gap in the understanding of self-evolving training for multimodal reasoning and offers a robust framework for future research. Our policy and reward models, as well as the collected data, will be released to facilitate further investigation in multimodal reasoning.", "keywords": [ "Large Multimodal Models", "Large Language Models", "Reinforcement Learning", "Multimodal Reasoning", "Self-Evolve" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "p85TNN62KD", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Junwei Deng", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "PhD student" }, { "name": "Weijing Tang", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Jiaqi Ma", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 71, "n_formula_1": 31, "n_ref_uni": 30, "n_ref": 73, "n_ref_all": 137, "n_fig": 9, "n_tab": 8, "L_tab": 1964, "n_element_tab": 162, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 490, "n_element_tab_1": 67, "formula_len_all": 7592, "formula_len_all_1": 1700, "len_all": 146208, "len_all_1": 70169, "len_abs": 1885, "len_title": 127, "len_sents": 45315, "len_sents_1": 31064, "n_sents": 400, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1887, "title": "A Versatile Influence Function for Data Attribution with Non-Decomposable Loss", "abs": "Influence function, a technique rooted in robust statistics, has been adapted in modern machine learning for a novel application: data attribution---quantifying how individual training data points affect a model's predictions. However, the common derivation of influence functions in the data attribution literature is limited to loss functions that decompose into a sum of individual data point losses, with the most prominent examples known as M-estimators. This restricts the application of influence functions to more complex learning objectives, which we refer to as non-decomposable losses, such as contrastive or ranking losses, where a unit loss term depends on multiple data points and cannot be decomposed further. In this work, we bridge this gap by revisiting the general formulation of influence function from robust statistics, which extends beyond M-estimators. Based on this formulation, we propose a novel method, the Versatile Influence Function (VIF), that can be straightforwardly applied to machine learning models trained with any non-decomposable loss. In comparison to the classical approach in statistics, the proposed VIF is designed to fully leverage the power of auto-differentiation, hereby eliminating the need for case-specific derivations of each loss function. We demonstrate the effectiveness of VIF across three examples: Cox regression for survival analysis, node embedding for network analysis, and listwise learning-to-rank for information retrieval. In all cases, the influence estimated by VIF closely resembles the results obtained by brute-force leave-one-out retraining, while being up to 1000 times faster to compute. We believe VIF represents a significant advancement in data attribution, enabling efficient influence-function-based attribution across a wide range of machine learning paradigms, with broad potential for practical use cases.", "keywords": [ "influence function", "data attribution" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "p7vJ3wsm34", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Guanghui Qin", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Vivek Chari", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Undergrad student" }, { "name": "Benjamin Van Durme", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 28, "n_ref": 40, "n_ref_all": 57, "n_fig": 5, "n_tab": 4, "L_tab": 751, "n_element_tab": 128, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 774, "n_element_tab_1": 132, "formula_len_all": 285, "formula_len_all_1": 285, "len_all": 111442, "len_all_1": 63092, "len_abs": 1467, "len_title": 113, "len_sents": 29624, "len_sents_1": 28491, "n_sents": 227, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1366, "title": "KV-Distill: Nearly Lossless Context Compression for Transformers", "abs": "Sequence-to-sequence natural language tasks often benefit greatly from long contexts, but the quadratic complexity of self-attention renders usage of long contexts non-trivial. In particular, during generation, temporary representations (stored in the KV cache) account for a large portion of GPU memory usage, and scale linearly with context length. In this work, we introduce KV-Distill, a flexible compression framework for large language models (LLMs) that distills long context KV caches into significantly shorter representations. KV-Distill can be trained as a parameter-efficient adaptor for pre-trained models, and enables the compression of arbitrary spans of a context while preserving the pre-trained model's capabilities, including instruction-tuning. We do this by treating a compressed-uncompressed cache as a student-teacher pairing and applying a KL-type divergence to match the generated outputs. Our experiments show that KV-Distill outperforms other compression techniques in worst-case extractive tasks, and approaches uncompressed performance in long context question answering and summarization. Furthermore, KV-Distill can be fine-tuned on domain-specific contexts to reduce context lengths by up 95% while preserving downstream task performance. We demonstrate the generalizability of KV-Distill across various model sizes and architectures.", "keywords": [ "distillation", "long context", "efficiency", "LLM", "compression" ], "rating_list": [ 3, 6, 3 ], "soundness_list": [ 2, 4, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 4, 2 ], "confidence_list": [ 3, 5, 4 ] }, { "paper_id": "p7mgNvOD9Q", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Chengyao Qian", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Jing Wu", "gender": "Female", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Trung Le", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" }, { "name": "Dinh Phung", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" }, { "name": "Mehrtash T. Harandi", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 43, "n_ref": 95, "n_ref_all": 106, "n_fig": 4, "n_tab": 7, "L_tab": 2997, "n_element_tab": 262, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2528, "n_element_tab_1": 183, "formula_len_all": 617, "formula_len_all_1": 550, "len_all": 123719, "len_all_1": 66185, "len_abs": 852, "len_title": 98, "len_sents": 31098, "len_sents_1": 29025, "n_sents": 243, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 50, "L_abs": 880, "title": "SUN: Training-free Machine Unlearning via Subspace", "abs": "Machine Unlearning (MU), a technique to erase undesirable content from AI models, plays an essential role in developing safe and trustworthy AI systems. Despite the success MU achieved, existing MU baselines typically necessitate maintaining the entire dataset for fine-tuning unlearned models.\nFine-tuning models and maintaining large datasets are computationally and financially prohibitive. This motivates us to propose a simple yet effective MU approach: \\underline{S}ubspace \\underline{UN}learning (SUN) as a new fast and effective MU baseline. The proposed method removes the low-dimensional subspaces of undesirable concepts from the space spanned by the weight vectors. This modification makes the model \"blind\" to the undesirable content to realize unlearning. Notably, SUN can produce the scrubbed model instantly with only a few samples and without additional training.", "keywords": [ "Machine unlearning", "Training-free" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "p7K3idvKTQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "H. G. Ranjani", "gender": "Female", "institution": "Ericsson", "country": "", "position": "Researcher" }, { "name": "Neeraj Gunda", "gender": "Male", "institution": "Ericsson", "country": "IN", "position": "Intern" }, { "name": "SAI KRISHNA BALA", "gender": "unknown", "institution": "Ericsson", "country": "IN", "position": "Senior Data Scientist" }, { "name": "Shashank Gautam", "gender": "unknown", "institution": "Telefonaktiebolaget LM Ericsson", "country": "IN", "position": "Intern" }, { "name": "Subhadip Bandyopadhyay", "gender": "Male", "institution": "Ericsson", "country": "", "position": "Principal Data Scientist" }, { "name": "Sujoy Roychowdhury", "gender": "Male", "institution": "Ericsson", "country": "IN", "position": "Researcher" }, { "name": "Sumit Soman", "gender": "Male", "institution": "Ericsson R&D", "country": "IN", "position": "Senior Data Scientist" }, { "name": "Vansh Chhabra", "gender": "Male", "institution": "Ericsson", "country": "", "position": "Intern" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 26, "n_ref": 47, "n_ref_all": 67, "n_fig": 2, "n_tab": 6, "L_tab": 7583, "n_element_tab": 705, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 3317, "n_element_tab_1": 348, "formula_len_all": 366, "formula_len_all_1": 490, "len_all": 113156, "len_all_1": 67560, "len_abs": 1510, "len_title": 128, "len_sents": 28300, "len_sents_1": 25610, "n_sents": 250, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1524, "title": "Towards Understanding Domain Adapted Sentence Embeddings for Document Retrieval", "abs": "A plethora of sentence embedding models makes it challenging to choose one, especially for technical domains rich with specialized vocabulary. In this work, we domain adapt embeddings using telecom, health and science datasets for question answering. We evaluate embeddings obtained from publicly available models and their domain-adapted variants, on both point retrieval accuracies, as well as their (95\\%) confidence intervals. We establish a systematic method to obtain thresholds for similarity scores for different embeddings. As expected, we observe that fine-tuning improves mean bootstrapped accuracies. We also observe that it results in tighter confidence intervals, which further improve when pre-training is preceded by fine-tuning. We introduce metrics which measure the distributional overlaps of top-$K$, correct and random document similarities with the question. Further, we show that these metrics are correlated with retrieval accuracy and similarity thresholds. Recent literature shows conflicting effects of isotropy on retrieval accuracies. Our experiments establish that the isotropy of embeddings (as measured by two independent state-of-the-art isotropy metric definitions) is poorly correlated with retrieval performance. We show that embeddings for domain-specific sentences have little overlap with those for domain-agnostic ones, and fine-tuning moves them further apart. Based on our results, we provide recommendations for use of our methodology and metrics by researchers and practitioners.", "keywords": [ "Sentence Embeddings", "Question Answering", "Technical Domains", "Retrieval Augmented Generation", "Embedding Models", "Isotropy" ], "rating_list": [ 3, 3, 3, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 1, 1, 3, 3 ], "contribution_list": [ 1, 1, 1, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "p74CpDzw1Y", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jihan Yao", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Lucy Lu Wang", "gender": "Female", "institution": "University of Washington", "country": "", "position": "Assistant Professor" }, { "name": "Shangbin Feng", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Wenxuan Ding", "gender": "Female", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Yulia Tsvetkov", "gender": "Female", "institution": "Department of Computer Science, University of Washington", "country": "", "position": "Associate Professor" } ], "n_formula": 22, "n_formula_1": 14, "n_ref_uni": 68, "n_ref": 130, "n_ref_all": 161, "n_fig": 6, "n_tab": 14, "L_tab": 6441, "n_element_tab": 311, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 3266, "n_element_tab_1": 175, "formula_len_all": 1259, "formula_len_all_1": 890, "len_all": 287510, "len_all_1": 72254, "len_abs": 963, "len_title": 110, "len_sents": 82182, "len_sents_1": 31372, "n_sents": 581, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 966, "title": "Varying Shades of Wrong: Aligning LLMs with Wrong Answers Only", "abs": "In the absence of abundant reliable annotations for challenging tasks and contexts, how can we expand the frontier of LLM capabilities with potentially wrong answers? We focus on two research questions: (1) Can LLMs generate reliable preferences among wrong options? And if so, (2) Would alignment with such wrong-over-wrong preferences be helpful? We employ methods based on self-consistency, token probabilities, and LLM-as-a-judge to elicit wrong-over-wrong preferences, and fine-tune language models with preference optimization approaches using these synthesized preferences. Extensive experiments with seven LLMs and eight datasets demonstrate that (1) LLMs do have preliminary capability in distinguishing various shades of wrong, achieving up to 20.9% higher performance than random guess; (2) Alignment with wrong-over-wrong preferences helps LLMs to produce less wrong and sometimes even outright correct answers, while overall improving model calibration.", "keywords": [ "alignment", "preference learning", "reasoning" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "p6ncr0eTKE", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "David Grangier", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Pierre Ablin", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Simin Fan", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Skyler Seto", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 57, "n_ref": 91, "n_ref_all": 139, "n_fig": 16, "n_tab": 3, "L_tab": 1488, "n_element_tab": 213, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 861, "formula_len_all_1": 603, "len_all": 193039, "len_all_1": 63666, "len_abs": 1031, "len_title": 122, "len_sents": 35442, "len_sents_1": 27739, "n_sents": 308, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1037, "title": "Task-Adaptive Pretrained Language Models via Clustered-Importance Sampling", "abs": "Specialist language models (LMs) focus on a specific task or domain on which they often outperform generalist LMs of the same size. However, the specialist data needed to pretrain these models is only available in limited amount for most tasks. In this work, we build specialist models from large generalist training sets instead. We adjust the training distribution of the generalist data with guidance from the limited domain-specific data. We explore several approaches, with clustered importance sampling standing out. This method clusters the generalist dataset and samples from these clusters based on their frequencies in the smaller specialist dataset. It is scalable, suitable for pretraining and continued pretraining, it works well in multi-task settings. Our findings demonstrate improvements across different domains in terms of language modeling perplexity and accuracy on multiple-choice question tasks. We also present ablation studies that examine the impact of dataset sizes, clustering configurations, and model sizes.", "keywords": [ "task-adaptive pretraining", "language models", "importance sampling", "domain adaptation" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 2, 2 ] }, { "paper_id": "p6eQRlaxGo", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Chence Shi", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "PhD student" }, { "name": "Jian Tang", "gender": "unknown", "institution": "HEC Montreal", "country": "", "position": "Assistant Professor" }, { "name": "Junqi Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Shaoning Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Zhi Yang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Associate Professor" } ], "n_formula": 22, "n_formula_1": 18, "n_ref_uni": 42, "n_ref": 117, "n_ref_all": 143, "n_fig": 17, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2123, "formula_len_all_1": 1032, "len_all": 170826, "len_all_1": 67373, "len_abs": 1393, "len_title": 107, "len_sents": 53724, "len_sents_1": 31929, "n_sents": 392, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1397, "title": "Design of Ligand-Binding Proteins with Atomic Flow Matching", "abs": "Designing novel proteins that bind to small molecules is a long-standing challenge in computational biology, with applications in developing catalysts, biosensors, and more. Current computational methods rely on the assumption that the binding pose of the target molecule is known, which is not always feasible, as conformations of novel targets are often unknown and tend to change upon binding. In this work, we formulate proteins and molecules as unified biotokens, and present AtomFlow, a novel deep generative model under the flow-matching framework for the design of ligand-binding proteins from the 2D target molecular graph alone. Operating on representative atoms of biotokens, AtomFlow captures the flexibility of ligands and generates ligand conformations and protein backbone structures iteratively. We consider the multi-scale nature of biotokens and demonstrate that AtomFlow can be effectively trained on a subset of structures from the Protein Data Bank, by matching flow vector field using an SE(3) equivariant structure prediction network. Experimental results show that our method can generate high-fidelity ligand-binding proteins and achieve performance comparable to the state-of-the-art model RFDiffusionAA, while not requiring bound ligand structures. As a general framework, AtomFlow holds the potential to be applied to various biomolecule generation tasks in the future.", "keywords": [ "protein generation", "protein-ligand interaction", "protein binder design" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "p66a00KLWN", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 10, "author_info_list": [ { "name": "Enzhi Zhang", "gender": "Male", "institution": "Hokkaido University", "country": "JP", "position": "PhD student" }, { "name": "Han Huang", "gender": "Male", "institution": "Chinese University of Hong Kong", "country": "HK", "position": "Intern" }, { "name": "Junfeng Fang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Kenji Kawaguchi", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Presidential Young Professor" }, { "name": "Sihang Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Xiang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yanchen Luo", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yaorui Shi", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Zhiyuan Liu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Tat-seng Chua", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 11, "n_ref_uni": 76, "n_ref": 195, "n_ref_all": 235, "n_fig": 12, "n_tab": 12, "L_tab": 7253, "n_element_tab": 773, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 4725, "n_element_tab_1": 510, "formula_len_all": 4544, "formula_len_all_1": 3962, "len_all": 253561, "len_all_1": 74143, "len_abs": 2889, "len_title": 124, "len_sents": 61121, "len_sents_1": 27392, "n_sents": 501, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1350, "title": "NExT-Mol: 3D Diffusion Meets 1D Language Modeling for 3D Molecule Generation", "abs": "3D molecule generation is crucial for drug discovery and material design. While prior efforts focus on 3D diffusion models for their benefits in modeling continuous 3D conformers, they overlook the advantages of 1D SELFIES-based Language Models (LMs), which can generate 100\\% valid molecules and leverage the billion-scale 1D molecule datasets. To combine these advantages for 3D molecule generation, we propose a foundation model -- NExT-Mol: 3D Diffusion Meets 1D Language Modeling for 3D Molecule Generation. NExT-Mol uses an extensively pretrained molecule LM for 1D molecule generation, and subsequently predicts the generated molecule's 3D conformers with a 3D diffusion model. We enhance NExT-Mol's performance by scaling up the LM's model size, refining the diffusion neural architecture, and applying 1D to 3D transfer learning. Notably, our 1D molecule LM significantly outperforms baselines in distributional similarity while ensuring validity, and our 3D diffusion model achieves leading performances in conformer prediction. Given these improvements in 1D and 3D modeling, NExT-Mol achieves a 26\\% relative improvement in 3D FCD for de novo 3D generation on GEOM-DRUGS, and a 13\\% average relative gain for conditional 3D generation on QM9-2014. Our codes and pretrained checkpoints are available at https://github.com/acharkq/NExT-Mol.", "keywords": [ "3D molecule generation", "molecular conformer generation", "large language models", "diffusion models", "geometric deep learning" ], "rating_list": [ 8, 3, 8, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 4, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "p60Y6o85Cj", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Sagar Shrestha", "gender": "Male", "institution": "Oregon State University", "country": "US", "position": "PhD student" }, { "name": "Xiao Fu", "gender": "Male", "institution": "Oregon State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 104, "n_formula_1": 19, "n_ref_uni": 50, "n_ref": 178, "n_ref_all": 245, "n_fig": 21, "n_tab": 3, "L_tab": 2034, "n_element_tab": 99, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 353, "n_element_tab_1": 48, "formula_len_all": 9369, "formula_len_all_1": 1552, "len_all": 254223, "len_all_1": 75258, "len_abs": 1953, "len_title": 143, "len_sents": 73639, "len_sents_1": 31019, "n_sents": 744, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1457, "title": "Content-Style Learning from Unaligned Domains: Identifiability under Unknown Latent Dimensions", "abs": "Understanding identifiability of latent content and style variables from unaligned multi-domain data is essential for tasks such as\ndomain translation and data generation. Existing works on content-style identification were often developed under somewhat stringent conditions, e.g., that all latent components are mutually independent and that the dimensions of the content and style variables are known. We introduce a new analytical framework via cross-domain *latent distribution matching* (LDM), which establishes content-style identifiability under substantially more relaxed conditions. Specifically, we show that restrictive assumptions such as component-wise independence of the latent variables can be removed. Most notably, we prove that prior knowledge of the content and style dimensions is not necessary for ensuring identifiability, if sparsity constraints are properly imposed onto the learned latent representations. Bypassing the knowledge of the exact latent dimension has been a longstanding aspiration in unsupervised representation learning---our analysis is the first to underpin its theoretical and practical viability. On the implementation side, we recast the LDM formulation into a regularized multi-domain GAN loss with coupled latent variables. We show that the reformulation is equivalent to LDM under mild conditions---yet requiring considerably less computational resource. Experiments corroborate with our theoretical claims.", "keywords": [ "unsupervised learning", "identifiability", "unknown latent dimension" ], "rating_list": [ 5, 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 4, 2, 3, 3, 3 ] }, { "paper_id": "p5o0sbE5kY", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Donghwan Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Jongchan Park", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Mingyu Park", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 47, "n_ref": 104, "n_ref_all": 150, "n_fig": 14, "n_tab": 7, "L_tab": 3043, "n_element_tab": 227, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 770, "formula_len_all_1": 792, "len_all": 175981, "len_all_1": 58566, "len_abs": 1452, "len_title": 128, "len_sents": 45855, "len_sents_1": 27140, "n_sents": 346, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1471, "title": "Pretraining A Shared Q-Network for Data Efficient Offline Reinforcement Learning", "abs": "Offline reinforcement learning (RL) aims to learn a policy from a static dataset without further interactions with the environment. Collecting sufficiently large datasets for offline RL is exhausting since this data collection requires colossus interactions with environments and becomes tricky when the interaction with the environment is restricted. Hence, how an agent learns the best policy with a minimal static dataset is a crucial issue in offline RL, similar to the sample efficiency problem in online RL. In this paper, we propose a simple yet effective plug-and-play pretraining method to initialize a feature of a $Q$-network to enhance data efficiency in offline RL. Specifically, we introduce a shared $Q$-network structure that outputs predictions of the next state and $Q$-value. We pretrain the shared $Q$-network through a supervised regression task that predicts a next state and trains the shared $Q$-network using diverse offline RL methods. Through extensive experiments, we empirically demonstrate that the proposed method enhances the performance of existing popular offline RL methods on the D4RL and Robomimic benchmarks, with an average improvement of 135.94\\% on the D4RL benchmark. Furthermore, we show that the proposed method significantly boosts data-efficient offline RL across various data qualities and data distributions. Notably, our method adapted with only 10\\% of the dataset outperforms standard algorithms even with full datasets.", "keywords": [ "Offline RL", "Data Efficiency", "Pretraining Q network" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "p5VDaa8aIY", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Armen Aghajanyan", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Research Scientist" }, { "name": "Hrant Khachatrian", "gender": "unknown", "institution": "Yerevan State University", "country": "AM", "position": "Principal Researcher" }, { "name": "Menua Bedrosian", "gender": "Male", "institution": "Yerevan State University", "country": "AM", "position": "Researcher" }, { "name": "Philipp Guevorguian", "gender": "Male", "institution": "YerevaNN", "country": "", "position": "Researcher" }, { "name": "Tigran Fahradyan", "gender": "Male", "institution": "YerevaNN research lab", "country": "", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 3, "n_ref_uni": 55, "n_ref": 89, "n_ref_all": 112, "n_fig": 2, "n_tab": 18, "L_tab": 10745, "n_element_tab": 882, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 3759, "n_element_tab_1": 306, "formula_len_all": 833, "formula_len_all_1": 112, "len_all": 221699, "len_all_1": 67246, "len_abs": 1335, "len_title": 102, "len_sents": 59228, "len_sents_1": 30191, "n_sents": 427, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1339, "title": "Small Molecule Optimization with Large Language Models", "abs": "Recent advancements in large language models (LLMs) have opened new possibilities for generative molecular drug design. In molecular optimization, LLMs are promising candidates to augment traditional modeling and rule-based approaches for refining molecular structures toward design criteria. We present a novel approach to molecular optimization using LLMs trained on a hand-crafted corpus of over 100 million molecules and their properties. We trained three new models, Chemlactica-125M, Chemlactica-1.3B, and Chemma-2B, with a demonstrated ability to generate molecules with specified properties and learn new molecular characteristics from limited samples, competitive with the state-of-the-art (SOTA) in property prediction tasks on experimental data. Our optimization method, elucidated by these capabilities, combines the models' generative power with concepts from prompt optimization, evolutionary algorithms, and rejection sampling to solve molecular optimization problems more efficiently. The approach surpasses previous SOTA results on the Practical Molecular Optimization (PMO) benchmark and exceeds or is competitive with the SOTA in multi-property optimization tasks involving docking simulations. We release the training data, language models, and optimization algorithm to facilitate further research and reproducibility.", "keywords": [ "Large Language Models", "Molecule Generation", "Transfer Learning", "Drug Discovery", "Evolutionary Algorithms" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 2, 4, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "p5RsCkE9sz", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Andrei Barbu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Chelsea Boccagno", "gender": "Female", "institution": "Harvard T.H. Chan School of Public Health", "country": "US", "position": "Postdoc" }, { "name": "Christopher J Hamblin", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Colin Conwell", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Postdoc" }, { "name": "David Mayo", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jesse Cummings", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Leyla Isik", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 39, "n_ref": 56, "n_ref_all": 65, "n_fig": 6, "n_tab": 1, "L_tab": 5, "n_element_tab": 1, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 5, "n_element_tab_1": 1, "formula_len_all": 78, "formula_len_all_1": 78, "len_all": 92309, "len_all_1": 69702, "len_abs": 1318, "len_title": 94, "len_sents": 30098, "len_sents_1": 25995, "n_sents": 173, "n_sents_1": 148, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1268, "title": "Using Multimodal Deep Neural Networks to Disentangle Language from Visual Aesthetic Experience", "abs": "When we experience a visual stimulus as beautiful, how much of that experience derives from perceptual computations we cannot describe versus conceptual knowledge we can readily translate into natural language? Disentangling perception from language in affective experiences through behavioral paradigms or neuroimaging is often empirically intractable. Here, we circumnavigate this challenge by using linear decoding over the learned representations of unimodal vision, unimodal language, and multimodal (language-aligned) deep neural network (DNN) models to predict human beauty ratings of naturalistic images. We find that unimodal vision models (e.g. SimCLR) account for the vast majority of explainable variance in these ratings. Language-aligned vision models (e.g. SLIP) yield small gains relative to unimodal vision. Unimodal language models (e.g. GPT2) conditioned on visual embeddings to generate captions (via CLIPCap) yield no further gains. Pure-language model embeddings of machine-generated captions alone yield lower predictions. Taken together, these results suggest that whatever words we may eventually find to describe our experiences of beauty, the ineffable computations of feedforward perception likely remain the dominant basis of our judgment.", "keywords": [ "multimodal deep neural networks", "vision and language", "affect and aesthetics" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 3, 1 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "p4jCBTDvdu", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Ivan Laptev", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" }, { "name": "Klea Ziu", "gender": "Female", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Nikita Durasov", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Pascal Fua", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Full Professor" }, { "name": "Samar Fares", "gender": "Female", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Toluwani Aremu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Karthik Nandakumar", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Associate Professor" }, { "name": "Martin Takac", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 7, "n_ref_uni": 68, "n_ref": 139, "n_ref_all": 187, "n_fig": 5, "n_tab": 8, "L_tab": 3045, "n_element_tab": 226, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1542, "n_element_tab_1": 163, "formula_len_all": 699, "formula_len_all_1": 1986, "len_all": 242784, "len_all_1": 66733, "len_abs": 1153, "len_title": 125, "len_sents": 60831, "len_sents_1": 29043, "n_sents": 388, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1155, "title": "$\\texttt{MirrorCheck}$ : Efficient Adversarial Defense for Vision-Language Models", "abs": "Vision-Language Models (VLMs) are becoming increasingly vulnerable to adversarial attacks as various novel attack strategies are being proposed against these models. While existing defenses excel in unimodal contexts, they currently fall short in safeguarding VLMs against adversarial threats. To mitigate this vulnerability, we propose a novel, yet elegantly simple approach for detecting adversarial samples in VLMs. Our method leverages Text-to-Image (T2I) models to generate images based on captions produced by target VLMs. Subsequently, we calculate the similarities of the embeddings of both input and generated images in the feature space to identify adversarial samples. Empirical evaluations conducted on different datasets validate the efficacy of our approach, outperforming baseline methods adapted from image classification domains. Furthermore, we extend our methodology to classification tasks, showcasing its adaptability and model-agnostic nature. Empirical findings also show the resilience of our approach against adaptive attacks, positioning it as an excellent defense mechanism for real-world deployment against adversarial threats.", "keywords": [ "Adversarial Attacks", "Adversarial Defenses", "Vision-Language Models", "StableDiffusion" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 5, 2, 4 ] }, { "paper_id": "p4cLtzk4oe", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chen Chen", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Daochang Liu", "gender": "Male", "institution": "University of Western Australia", "country": "AU", "position": "Lecturer" }, { "name": "Mubarak Shah", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Chang Xu", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 19, "n_ref": 41, "n_ref_all": 63, "n_fig": 8, "n_tab": 3, "L_tab": 375, "n_element_tab": 80, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 557, "n_element_tab_1": 84, "formula_len_all": 524, "formula_len_all_1": 524, "len_all": 104704, "len_all_1": 63333, "len_abs": 1272, "len_title": 76, "len_sents": 43267, "len_sents_1": 32561, "n_sents": 276, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 76, "L_abs": 1277, "title": "Exploring Local Memorization in Diffusion Models via Bright Ending Attention", "abs": "In this paper, we identify and leverage a novel `bright ending' (BE) anomaly in diffusion models prone to memorizing training images to address a new task: locating localized memorization regions within these models. BE refers to a distinct cross-attention pattern observed in text-to-image generations using diffusion models. Specifically, memorized image patches exhibit significantly greater attention to the end token during the final inference step compared to non-memorized patches. This attention map effectively highlights regions where the generated image replicates training data. Furthermore, driven by our observation that local memorization significantly underperforms in existing tasks of measuring, detecting, and mitigating memorization in diffusion models compared to global memorization, we propose a simple yet effective method to integrate BE and the results of the new localization task into these existing frameworks. This integration effectively improves their performances by narrowing the performance gap caused by local memorization. Our results not only demonstrate the successful execution of the new localization task but also establish new state-of-the-art performance across all existing tasks, underscoring the significance of the BE phenomenon.", "keywords": [ "Diffusion Models", "Local Memorization", "Bright Ending Attention" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 4, 2 ], "confidence_list": [ 2, 3, 4 ] }, { "paper_id": "p4RAKZ4oik", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Jiaqi Wu", "gender": "unknown", "institution": "China University of Mining Technology - Beijing", "country": "CN", "position": "PhD student" }, { "name": "Rui Jing", "gender": "Female", "institution": "China University of Mining Technology - Beijing", "country": "CN", "position": "MS student" }, { "name": "Shiyue Hou", "gender": "unknown", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Simin Chen", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Postdoc" }, { "name": "Wei Chen", "gender": "Male", "institution": "China University of Mining Technology - Xuzhou", "country": "CN", "position": "Full Professor" }, { "name": "Yijiang Li", "gender": "Not Specified", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Yuzhe YANG", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Undergrad student" }, { "name": "Zehua Wang", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Researcher" }, { "name": "Zijian Tian", "gender": "Male", "institution": "China University of Mining Technology - Beijing", "country": "CN", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 15, "n_ref_uni": 24, "n_ref": 27, "n_ref_all": 41, "n_fig": 2, "n_tab": 5, "L_tab": 1756, "n_element_tab": 174, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1323, "n_element_tab_1": 130, "formula_len_all": 894, "formula_len_all_1": 917, "len_all": 110050, "len_all_1": 67274, "len_abs": 3612, "len_title": 145, "len_sents": 33018, "len_sents_1": 30803, "n_sents": 241, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1797, "title": "FedDTPT: Federated Discrete and Transferable Prompt Tuning for Black-Box Large Language Models", "abs": "In recent years, large language models (LLMs) have significantly advanced the field of natural language processing (NLP). By fine-tuning LLMs with data from specific scenarios, these foundation models can better adapt to various downstream tasks. However, the fine-tuning process poses privacy leakage risks, particularly in centralized data processing scenarios. To address user privacy concerns, federated learning (FL) has been introduced to mitigate the risks associated with centralized data collection from multiple sources. Nevertheless, the privacy of LLMs themselves is equally critical, as potential malicious attacks challenge their security, an issue that has received limited attention in current research. Consequently, establishing a trusted multi-party model fine-tuning environment is essential. Additionally, the local deployment of large LLMs incurs significant storage costs and high computational demands. To address these challenges, we propose for the first time a federated discrete and transferable prompt tuning, namely FedDTPT, for black-box large language models. In the client optimization phase, we adopt a token-level discrete prompt optimization method that leverages a feedback loop based on prediction accuracy to drive gradient-free prompt optimization through the MLM API. For server optimization, we employ an attention mechanism based on semantic similarity to filter all local prompt tokens, along with an embedding distance elbow detection and DBSCAN clustering strategy to enhance the filtering process. Experimental results demonstrate that, compared to state-of-the-art methods, our approach achieves higher accuracy, reduced communication overhead, and robustness to non-iid data in a black-box setting. Moreover, the optimized prompts are transferable.", "keywords": [ "Federated learning", "Large Language Models", "Black-Box Estimation", "Discrete Prompt Learning" ], "rating_list": [ 1, 3, 3, 5 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "p4JeJ9uhEL", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Chengyang Song", "gender": "Male", "institution": "Ocean University of China", "country": "CN", "position": "MS student" }, { "name": "Qing Guo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Tuan-Anh Vu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Zheng Ziqiang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Ivor W Tsang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Sai-kit Yeung", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 1, "n_ref_uni": 54, "n_ref": 126, "n_ref_all": 145, "n_fig": 6, "n_tab": 6, "L_tab": 1213, "n_element_tab": 35, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 322, "n_element_tab_1": 20, "formula_len_all": 365, "formula_len_all_1": 28, "len_all": 175763, "len_all_1": 50826, "len_abs": 1806, "len_title": 132, "len_sents": 51783, "len_sents_1": 23578, "n_sents": 449, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1877, "title": "CamoVid60K: A Large-Scale Video Dataset for Moving Camouflaged Animals Understanding", "abs": "We have been witnessing remarkable success led by the power of neural networks driven by a significant scale of training data in handling various computer vision tasks. However, less attention has been paid to monitoring the camouflaged animals, the masters of hiding themselves in the background. Robust and precise segmentation of camouflaged animals is challenging even for domain experts due to their similarity to the environment. Although several efforts have been made in camouflaged animal image segmentation, to the best of our knowledge, limited work exists on camouflaged animal video understanding (CAVU). Biologists often prefer videos for monitoring and understanding animal behaviors, as videos provide redundant information and temporal consistency. However, the scarcity of labeled video data significantly hinders progress in this area. To address these challenges, we present $\\textbf{CamoVid60K}$, a diverse, large-scale, and accurately annotated video dataset of camouflaged animals. This dataset comprises $\\textbf{218}$ videos with $\\textbf{62,774}$ finely annotated frames, covering $\\textbf{70}$ animal categories, which $\\textit{surpasses}$ all previous datasets in terms of the number of videos/frames and species included. $\\textbf{CamoVid60K}$ also offers more diverse downstream tasks in computer vision, such as camouflaged animal classification, detection, and task-specific segmentation (semantic, referring, motion), $\\textit{etc}$. We have benchmarked several state-of-the-art algorithms on the proposed $\\textbf{CamoVid60K}$ dataset, and the experimental results provide valuable insights for future research directions. Our dataset serves as a $\\textit{novel}$ and $\\textit{challenging}$ benchmark to stimulate the development of more powerful camouflaged animal video segmentation algorithms, with substantial room for further improvement.", "keywords": [ "camouflaged animals", "video object segmentation", "video object detection" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "p3mxzKmuZy", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Chao Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Di Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Haiqin Weng", "gender": "Female", "institution": "Ant Group", "country": "", "position": "Researcher" }, { "name": "Han Qiu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Qingjie Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Tianwei Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Wenyu Zhu", "gender": "Male", "institution": "AscendGrace Tech. Co., Ltd", "country": "CN", "position": "Principal Researcher" }, { "name": "Yiming Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Research Fellow" }, { "name": "Dou Goodman", "gender": "unknown", "institution": "antgroup", "country": "CN", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 5, "n_ref_uni": 73, "n_ref": 101, "n_ref_all": 139, "n_fig": 8, "n_tab": 10, "L_tab": 494, "n_element_tab": 36, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 492, "formula_len_all_1": 318, "len_all": 228541, "len_all_1": 65631, "len_abs": 1145, "len_title": 111, "len_sents": 67980, "len_sents_1": 28922, "n_sents": 611, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1042, "title": "A Benchmark for Semantic Sensitive Information in LLMs Outputs", "abs": "Large language models (LLMs) can output sensitive information, which has emerged as a novel safety concern. Previous works focus on structured sensitive information (e.g. personal identifiable information). However, we notice that sensitive information in LLMs\u2019 outputs can also be at the semantic level, i.e. semantic sensitive information (SemSI). Particularly, simple natural questions can let state-of-the-art (SOTA) LLMs output SemSI. Compared to previous work of structured sensitive information in LLM\u2019s outputs, SemSI are hard to define and are rarely studied. Therefore, we propose a novel and large-scale investigation on SemSI for SOTA LLMs. First, we construct a comprehensive and labeled dataset of semantic sensitive information, SemSI-Set, by including three typical categories of SemSI. Then, we propose a large-scale benchmark, SemSI-Bench, to systematically evaluate semantic sensitive information in 25 SOTA LLMs. Our\nfinding reveals that SemSI widely exists in SOTA LLMs\u2019 outputs by querying with simple natural questions.", "keywords": [ "LLMs", "sensitive information" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 4 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "p3NVJg6ywM", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Junming Liu", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "MS student" }, { "name": "Yanting Gao", "gender": "Female", "institution": "Tongji University", "country": "CN", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 34, "n_ref": 57, "n_ref_all": 64, "n_fig": 2, "n_tab": 4, "L_tab": 413, "n_element_tab": 19, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 130, "n_element_tab_1": 22, "formula_len_all": 702, "formula_len_all_1": 669, "len_all": 142257, "len_all_1": 65599, "len_abs": 1239, "len_title": 145, "len_sents": 42823, "len_sents_1": 23681, "n_sents": 276, "n_sents_1": 159, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1243, "title": "DFED: Data-Free Ensemble Distillation with Multi-Source GANs for Heterogeneous Federated Learning", "abs": "Federated Learning (FL) is a decentralized machine learning paradigm that enables clients to collaboratively train models while preserving data privacy. However, surmounting the obstacles introduced by data heterogeneity in heterogeneous federated learning remains a profound challenge, as it drives each client towards distinct convergence trajectories, impeding the global model's convergence. To transcend these challenges, we propose DFED, a novel data-free ensemble knowledge distillation method designed to counteract the effects of data heterogeneity. DFED leverages multi-source Generative Adversarial Networks (GANs) to generate synthetic data that aligns with local distributions, ensuring privacy while promoting diverse feature representations across clients. Additionally, DFED aggregates client models into an ensemble based on their specialized knowledge, and applies ensemble distillation to refine the global model, mitigating the issues caused by disparities in data distributions. Across a variety of image classification benchmarks, DFED demonstrates superior performance compared to several state-of-the-art (SOTA) methods. The source code will be made publicly accessible once the paper has been accepted for publication.", "keywords": [ "Federated learning", "Data heterogeneity", "Data-Free Knowledge Distillation" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "p3NKpom1VL", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Geng Li", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Hulingxiao He", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jinglin Xu", "gender": "Female", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "Associate Professor" }, { "name": "Yuxin Peng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Geng Zijun", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Intern" } ], "n_formula": 22, "n_formula_1": 22, "n_ref_uni": 33, "n_ref": 98, "n_ref_all": 127, "n_fig": 13, "n_tab": 9, "L_tab": 2524, "n_element_tab": 327, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1277, "n_element_tab_1": 155, "formula_len_all": 1072, "formula_len_all_1": 1166, "len_all": 143515, "len_all_1": 59696, "len_abs": 1228, "len_title": 153, "len_sents": 40866, "len_sents_1": 25658, "n_sents": 265, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 1368, "title": "Analyzing and Boosting the Power of Fine-Grained Visual Recognition for Multi-modal Large Language Models", "abs": "Multi-modal large language models (MLLMs) have shown remarkable abilities in various visual understanding tasks. However, MLLMs still struggle with fine-grained visual recognition (FGVR), which aims to identify subordinate-level categories from images. This can negatively impact more advanced capabilities of MLLMs, such as object-centric visual question answering and reasoning. In our study, we revisit three quintessential capabilities of MLLMs for FGVR, including object information extraction, category knowledge reserve, object-category alignment, and position of the root cause as a misalignment problem. To address this issue, we present Finedefics, an MLLM that enhances the model's FGVR capability by incorporating informative attribute descriptions of objects into the training phase. We employ contrastive learning on object-attribute pairs and attribute-category pairs simultaneously and use examples from similar but incorrect categories as hard negatives, naturally bringing representations of visual objects and category names closer. Extensive evaluations across multiple popular FGVR datasets demonstrate that Finedefics outperforms existing MLLMs of comparable parameter sizes, showcasing its remarkable efficacy. The code is available at [https://github.com/PKU-ICST-MIPL/Finedefics_ICLR2025](https://github.com/PKU-ICST-MIPL/Finedefics_ICLR2025).", "keywords": [ "Multimodal Large Language Models", "Fine-Grained Visual Recognition" ], "rating_list": [ 8, 3, 5 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 4, 2, 2 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "p30YulvDbj", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Ahsan Habib", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Lecturer" }, { "name": "Chandan Karmakar", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Associate Professor" }, { "name": "Maia Angelova", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Sheik Mohammed Shariful Islam", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Associate Professor" }, { "name": "Shruthi Narayanan Vaniya", "gender": "Female", "institution": "Deakin University", "country": "AU", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 53, "n_ref_all": 66, "n_fig": 7, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 86895, "len_all_1": 75435, "len_abs": 1734, "len_title": 126, "len_sents": 23942, "len_sents_1": 23939, "n_sents": 156, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1743, "title": "OPTIMIZED SINGLE EEG CHANNEL SELECTION FOR DETECTING MAJOR DEPRESSIVE DISORDER", "abs": "Major depressive disorder (MDD) or depression is a chronic mental illness that significantly impacts individuals' well-being and is often diagnosed at advanced stages, increasing the risk of suicide. Current diagnostic practices, which rely heavily on subjective assessments and patient self-reports, are often hindered by challenges such as under-reporting and the failure to detect early, subtle symptoms. Early detection of MDD is crucial and requires monitoring vital signs in everyday living conditions. Electroencephalogram (EEG) is a valuable tool for monitoring brain activity, offering critical insights into MDD and its underlying neurological mechanisms. While traditional EEG systems typically involve multiple channels for recording, making them impractical for home-based monitoring, wearable sensors can effectively capture single-channel EEG data. However, generating meaningful features from this data poses challenges due to the need for specialized domain knowledge and significant computational power, which can hinder real-time processing. To address these issues, our study focuses on developing a deep learning model for the binary classification of MDD using single-channel EEG data. We focused on specific channels from various brain regions, including central (C3), frontal (Fp1), occipital (O1), temporal (T4), and parietal (P3). Our study found that the channels Fp1, C3, and O1 achieved an impressive accuracy of 88\\% when analyzed using a Convolutional Neural Network (CNN) with leave-one-subject-out cross-validation. Our study highlights the potential of utilizing single-channel EEG data for reliable MDD diagnosis, providing a less intrusive and more convenient wearable solution for mental health assessment.", "keywords": [ "Major depressive disorder", "deep learning", "electroencephalogram", "EEG", "single-channel" ], "rating_list": [ 1, 1, 3, 3 ], "soundness_list": [ 1, 1, 2, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 4, 5, 4, 1 ] }, { "paper_id": "p2oFwfwebT", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Li Gu", "gender": "Male", "institution": "Huawei Canada", "country": "CA", "position": "Researcher" }, { "name": "Yang Wang", "gender": "Male", "institution": "Concordia University", "country": "CA", "position": "Associate Professor" }, { "name": "Zhi Liu", "gender": "Male", "institution": "Shanghai University", "country": "CN", "position": "Full Professor" }, { "name": "Zhixiang Chi", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Ziqiang Wang", "gender": "Male", "institution": "Concordia University", "country": "CA", "position": "PhD student" }, { "name": "Kostantinos Plataniotis", "gender": "Male", "institution": "Toronto University", "country": "CA", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 47, "n_ref": 85, "n_ref_all": 106, "n_fig": 4, "n_tab": 6, "L_tab": 2358, "n_element_tab": 264, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1731, "n_element_tab_1": 233, "formula_len_all": 866, "formula_len_all_1": 866, "len_all": 124480, "len_all_1": 69329, "len_abs": 1335, "len_title": 116, "len_sents": 32923, "len_sents_1": 30260, "n_sents": 259, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1341, "title": "Diffusion-based Prompt Generation for Lifelong Continual Adaptation", "abs": "Continual Test-time Adaptation (TTA) addresses sequential out-of-distribution scenarios with unlabeled data but overlooks long-term and recurring in-distribution aspects of the real world. Therefore, we introduce Lifelong Continual Adaptation, which enables models to efficiently retrieve domain-specific knowledge when encountering in-distribution data streams with sequential and recurring domains. We found that optimization-based Continual TTA methods underperform on the proposed problem due to two major pitfalls: updating the model's parameters is expensive and impractical for resource-constrained devices, and these methods exhibit instability when adapting to long-term recurring domains. To address these challenges, we propose a diffusion-based prompt generation method (DiffPrompt). Specifically, instead of continually optimizing the foundation model, we generate domain-specific prompts for it to adapt. We use a conditional diffusion model to learn a prompt-space distribution for various domains. During testing, the diffusion model generates prompts for the current domain based on the incoming batch of data, facilitating the continual adaptation of the foundation model. Our experiments demonstrate that DiffPrompt enables stable and efficient deployment in practical scenarios involving sequential and recurring domains.", "keywords": [ "Domain shift", "Prompt generation", "Lifelong continual adaptation", "Diffusion model", "Foundation model" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "p2QAOORDoG", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Aobo Li", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "J. T. Fry", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Kaliro\u00eb Mabelle West Pappas", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Lindley Winslow", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Xinyi Hope Fu", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Zhenghao Fu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 21, "n_ref": 42, "n_ref_all": 68, "n_fig": 7, "n_tab": 4, "L_tab": 971, "n_element_tab": 103, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 273, "n_element_tab_1": 34, "formula_len_all": 339, "formula_len_all_1": 340, "len_all": 154931, "len_all_1": 60694, "len_abs": 1455, "len_title": 121, "len_sents": 56652, "len_sents_1": 30252, "n_sents": 459, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1462, "title": "TIDMAD: Time Series Dataset for Discovering Dark Matter with AI Denoising", "abs": "Dark matter makes up approximately 85\\% of total matter in our universe, yet it has never been directly observed in any laboratory on Earth. The origin of dark matter is one of the most important questions in contemporary physics, and a convincing detection of dark matter would be a Nobel-Prize-level breakthrough in fundamental science. The ABRACADABRA experiment was specifically designed to search for dark matter. Although it has not yet made a discovery, ABRACADABRA has produced several dark matter search results widely endorsed by the physics community. The experiment generates ultra-long time-series data at a rate of 10 million samples per second, where the dark matter signal would manifest itself as a sinusoidal oscillation mode within the ultra-long time series. In this paper, we present the TIDMAD --- a comprehensive data release from the ABRACADABRA experiment including three key components: an ultra-long time series dataset divided into training, validation, and science subsets; a carefully-designed denoising score for direct model benchmarking; and a complete analysis framework which produces a community-standard dark matter search result suitable for publication as a physics paper. This data release enables core AI algorithms to extract the signal and produce real physics results thereby advancing fundamental science. The data downloading and associated analysis scripts are available at https://anonymous.4open.science/r/TIDMAD.", "keywords": [ "benchmark", "dataset", "denoising", "public dataset", "dark matter", "physics", "time series", "ultra-long time series" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 4, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "p1b96KC6rj", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Christopher Bockel-Rickermann", "gender": "unknown", "institution": "KU Leuven", "country": "BE", "position": "Researcher" }, { "name": "Tim Verdonck", "gender": "Male", "institution": "Universiteit Antwerpen", "country": "BE", "position": "Full Professor" }, { "name": "Toon Vanderschueren", "gender": "Male", "institution": "Telenet", "country": "BE", "position": "Researcher" }, { "name": "Wouter Verbeke", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 65, "n_ref": 112, "n_ref_all": 137, "n_fig": 9, "n_tab": 5, "L_tab": 1999, "n_element_tab": 254, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 647, "n_element_tab_1": 34, "formula_len_all": 508, "formula_len_all_1": 134, "len_all": 181594, "len_all_1": 73782, "len_abs": 1384, "len_title": 136, "len_sents": 48610, "len_sents_1": 34781, "n_sents": 368, "n_sents_1": 256, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1394, "title": "Sources of Gain: Decomposing Performance in Conditional Average Dose Response Estimation", "abs": "Estimating conditional average dose responses (CADR) is an important but challenging problem. Estimators must correctly model the potentially complex relationships between covariates, interventions, doses, and outcomes. In recent years, the machine learning community has shown great interest in developing tailored CADR estimators that target specific challenges. Their performance is typically evaluated against other methods on (semi-) synthetic benchmark datasets. Our paper analyses this practice and shows that using popular benchmark datasets without further analysis is insufficient to judge model performance. Established benchmarks entail multiple challenges, whose impacts must be disentangled. Therefore, we propose a novel decomposition scheme that allows the evaluation of the impact of five distinct components contributing to CADR estimator performance. We apply this scheme to eight popular CADR estimators on four widely-used benchmark datasets, running nearly 1,500 individual experiments. Our results reveal that most established benchmarks are challenging for reasons different from their creators' claims. Notably, we find that confounding - the key challenge that motivated recent methods - does not significantly affect CADR estimation performance for the considered datasets. We discuss the major implications of our findings and present directions for future research.", "keywords": [ "Dose response estimation", "Causal Machine Learning", "Performance Decomposition", "Datasets" ], "rating_list": [ 3, 3, 3, 3, 10 ], "soundness_list": [ 3, 2, 2, 2, 4 ], "presentation_list": [ 3, 2, 3, 3, 4 ], "contribution_list": [ 1, 1, 1, 1, 4 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "p1TBYyqy8v", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Daeho Um", "gender": "Male", "institution": "Samsung", "country": "KR", "position": "Researcher" }, { "name": "Jongin Lim", "gender": "Male", "institution": "Samsung Advanced Institute of Technology", "country": "KR", "position": "Researcher" }, { "name": "Sunoh Kim", "gender": "Male", "institution": "Samsung", "country": "KR", "position": "Researcher" }, { "name": "Yoonho Jung", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Undergrad student" }, { "name": "Yuneil Yeo", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 35, "n_formula_1": 13, "n_ref_uni": 49, "n_ref": 106, "n_ref_all": 150, "n_fig": 9, "n_tab": 14, "L_tab": 12597, "n_element_tab": 1238, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 4165, "n_element_tab_1": 228, "formula_len_all": 1581, "formula_len_all_1": 578, "len_all": 236634, "len_all_1": 70131, "len_abs": 1306, "len_title": 97, "len_sents": 70309, "len_sents_1": 28132, "n_sents": 599, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1313, "title": "Spreading Out-of-Distribution Detection on Graphs", "abs": "Node-level out-of-distribution (OOD) detection on graphs has received significant attention from the machine learning community. However, previous approaches are evaluated using unrealistic benchmarks that consider only randomly selected OOD nodes, failing to reflect the interactions among nodes. In this paper, we introduce a new challenging task to model the interactions of OOD nodes in a graph, termed spreading OOD detection, where a newly emerged OOD node spreads its property to neighboring nodes. We curate realistic benchmarks by employing the epidemic spreading models that simulate the spreading of OOD nodes on the graph. We also showcase a ``Spreading COVID-19\" dataset to demonstrate the applicability of spreading OOD detection in real-world scenarios. Furthermore, to effectively detect spreading OOD samples under the proposed benchmark setup, we present a new approach called energy distribution-based detector (EDBD), which includes a novel energy-aggregation scheme. EDBD is designed to mitigate undesired mixing of OOD scores between in-distribution (ID) and OOD nodes. Our extensive experimental results demonstrate the superiority of our approach over state-of-the-art methods in both spreading OOD detection and conventional node-level OOD detection tasks across seven benchmark datasets.", "keywords": [ "out-of-distribution detection", "graph neural networks", "benchmark", "healthcare" ], "rating_list": [ 5, 6, 8, 6, 8 ], "soundness_list": [ 2, 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 4, 2 ] }, { "paper_id": "p1HeFnn2AA", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "David C. Parkes", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Jonathan Ma", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "Undergrad student" }, { "name": "Sai Srivatsa Ravindranath", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Scott Kominers", "gender": "Male", "institution": "Harvard University, Harvard Business School", "country": "US", "position": "Full Professor" }, { "name": "Shira Li", "gender": "Female", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Zhe Feng", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 32, "n_formula_1": 16, "n_ref_uni": 62, "n_ref": 86, "n_ref_all": 97, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9355, "formula_len_all_1": 7470, "len_all": 176059, "len_all_1": 84056, "len_abs": 5261, "len_title": 84, "len_sents": 50030, "len_sents_1": 36164, "n_sents": 415, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 1108, "title": "Deep Learning for Two-Sided Matching", "abs": "We initiate the study of deep learning for the automated design of two-sided matching mechanisms. What is of most interest is to use machine learning to understand the possibility of new tradeoffs between *strategy-proofness* and *stability*. These properties cannot be achieved simultaneously, but the efficient frontier is not understood. We introduce novel differentiable surrogates for quantifying ordinal strategy-proofness and stability and use them to train differentiable matching mechanisms that map discrete preferences to valid randomized matchings. We demonstrate that the efficient frontier characterized by these learned mechanisms is substantially better than that achievable through a convex combination of baselines of *deferred acceptance* (stable and strategy-proof for only one side of the market), *top trading cycles* (strategy-proof for one side, but not stable), and *randomized serial dictatorship* (strategy-proof for both sides, but not stable). This gives a new target for economic theory and opens up new possibilities for machine learning pipelines in matching market design.", "keywords": [ "Mechanism Design", "Matching Markets", "Game Theory", "Differentiable Economics", "Two-Sided Matching" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "p0DjhjPXl3", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Bin Chen", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Hao Fang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Intern" }, { "name": "Hongyao Yu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Shu-Tao Xia", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Tianqu Zhuang", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yixiang Qiu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 35, "n_formula_1": 18, "n_ref_uni": 25, "n_ref": 43, "n_ref_all": 84, "n_fig": 4, "n_tab": 7, "L_tab": 1047, "n_element_tab": 170, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 403, "n_element_tab_1": 61, "formula_len_all": 2420, "formula_len_all_1": 765, "len_all": 114490, "len_all_1": 51937, "len_abs": 1123, "len_title": 162, "len_sents": 32813, "len_sents_1": 20728, "n_sents": 329, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 114, "L_abs": 1252, "title": "Stealthy Shield Defense: A Conditional Mutual Information-Based Approach against Black-Box Model Inversion Attacks", "abs": "Model inversion attacks (MIA) aim to uncover private training data by accessing public models, raising increasing concerns about privacy breaches. Black-box MIA, where attackers can generate inputs and obtain the model's outputs arbitrarily, has gained more attention due to its closer alignment with real-world scenarios and greater potential threats. Existing defenses primarily focus on white-box attacks, with a lack of specialized defenses to address the latest black-box attacks. To fill this technological gap, we propose a post-processing defense algorithm based on conditional mutual information (CMI). We have theoretically proven that our CMI framework serves as a special information bottleneck, making outputs less dependent on inputs and more dependent on true labels. To further reduce the modifications to outputs, we introduce an adaptive rate-distortion framework and optimize it by water-filling method. Experimental results show that our approach outperforms existing defenses, in terms of both MIA robustness and model utility, across various attack algorithms, training datasets, and model architectures. In particular, on CelebA dataset, our defense lowers the attack accuracy of LOKT to 0\\% while other defenses remain 50-75\\%.", "keywords": [ "model inversion attack", "model inversion defense", "conditional mutual information" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "p01BR4njlY", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Calvin Luo", "gender": "Male", "institution": "Brown University", "country": "US", "position": "PhD student" }, { "name": "Chen Sun", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Assistant Professor" }, { "name": "Yilun Du", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Zilai Zeng", "gender": "Male", "institution": "Brown University", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 36, "n_ref": 96, "n_ref_all": 134, "n_fig": 5, "n_tab": 3, "L_tab": 1748, "n_element_tab": 116, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 270, "formula_len_all_1": 270, "len_all": 156515, "len_all_1": 65908, "len_abs": 1582, "len_title": 102, "len_sents": 54045, "len_sents_1": 35816, "n_sents": 329, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1593, "title": "Solving New Tasks by Adapting Internet Video Knowledge", "abs": "Video generative models, beyond enabling the production of astounding visual creations, offer a promising pathway for unlocking novel, text-conditioned robotic behaviors, whether utilized as a video planner or as a policy supervisor. When pretrained on internet-scale datasets, such video models intimately understand alignment with natural language, and can thus facilitate novel text-conditioned behavior generalization. At the same time, however, they may not be sensitive to the specificities of the particular environment in which a policy of interest is to be learned. On the other hand, video modeling over in-domain examples of robotic behavior naturally encodes environment-specific intricacies, but the scale of available demonstrations may not be sufficient to support generalization to unseen tasks via natural language specification. In this work, we investigate different adaptation techniques that integrate in-domain information into large-scale pretrained video models, and explore the extent to which they enable novel text-conditioned generalization for robotic tasks. Furthermore, we highlight the individual data and training requirements of each approach, which range from utilizing only a few still frames illustrating the subject of interest, to direct finetuning over videos labelled with text descriptions. We successfully demonstrate across robotic environments that adapting powerful video models with small scales of example data can successfully facilitate generalization to novel behaviors, both when utilized as policy supervisors, and as visual planners.", "keywords": [ "Text-Conditioned Generalization", "Video Diffusion", "Adaptation", "Planning", "Policy Learning" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "ozhRaoRGyl", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "David Mark Bossens", "gender": "unknown", "institution": "A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Qing Guo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Xingrui Yu", "gender": "Male", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Yueming Lyu", "gender": "Male", "institution": "Agency for Science, Technology and Research (A*STAR)", "country": "SG", "position": "Researcher" }, { "name": "Zhenglin Wan", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Undergrad student" }, { "name": "Xiaofeng Fan", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Ivor W Tsang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" } ], "n_formula": 37, "n_formula_1": 10, "n_ref_uni": 39, "n_ref": 67, "n_ref_all": 97, "n_fig": 16, "n_tab": 4, "L_tab": 1954, "n_element_tab": 273, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 839, "n_element_tab_1": 142, "formula_len_all": 3359, "formula_len_all_1": 800, "len_all": 169231, "len_all_1": 62660, "len_abs": 883, "len_title": 84, "len_sents": 57005, "len_sents_1": 29706, "n_sents": 420, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 887, "title": "Quality Diversity Imitation Learning", "abs": "Imitation learning (IL) has shown great potential in various applications, such as robot control. However, traditional IL methods are usually designed to learn only one specific type of behavior since demonstrations typically correspond to a single expert. In this work, we introduce the first generic framework for Quality Diversity Imitation Learning (QD-IL), which enables the agent to learn a broad range of skills from limited demonstrations. Our framework integrates the principles of quality diversity with adversarial imitation learning (AIL) methods, and can potentially improve any inverse reinforcement learning (IRL) method. Empirically, our framework significantly improves the QD performance of GAIL and VAIL on the challenging continuous control tasks derived from Mujoco environments. Moreover, our method even achieves 2x expert performance in the Humanoid environment.", "keywords": [ "Imitation Learning", "Quality Diversity", "behavior-level exploration", "limited demonstration" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "ozZG5FXuTV", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Ching-Wen Lee", "gender": "Female", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Mingzhou Liu", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Xinwei Sun", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xueqing Yu", "gender": "Female", "institution": "Peking University", "country": "", "position": "Undergrad student" }, { "name": "Yizhou Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Yu QIAO", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 19, "n_ref_uni": 42, "n_ref": 67, "n_ref_all": 120, "n_fig": 10, "n_tab": 11, "L_tab": 2664, "n_element_tab": 323, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 803, "n_element_tab_1": 43, "formula_len_all": 1686, "formula_len_all_1": 1372, "len_all": 157925, "len_all_1": 65529, "len_abs": 1177, "len_title": 104, "len_sents": 40488, "len_sents_1": 28022, "n_sents": 363, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1182, "title": "Learning Causal Alignment for Reliable Disease Diagnosis", "abs": "Aligning the decision-making process of machine learning algorithms with that of experienced radiologists is crucial for reliable diagnosis. While existing methods have attempted to align their prediction behaviors to those of radiologists reflected in the training data, this alignment is primarily associational rather than causal, resulting in pseudo-correlations that may not transfer well. In this paper, we propose a causality-based alignment framework towards aligning the model's decision process with that of experts. Specifically, we first employ counterfactual generation to identify the causal chain of model decisions. To align this causal chain with that of experts, we propose a causal alignment loss that enforces the model to focus on causal factors underlying each decision step in the whole causal chain. To optimize this loss that involves the counterfactual generator as an implicit function of the model's parameters, we employ the implicit function theorem equipped with the conjugate gradient method for efficient estimation. We demonstrate the effectiveness of our method on two medical diagnosis applications, showcasing faithful alignment to radiologists.", "keywords": [ "alignment", "causal learning", "counterfactual", "disease diagnosis" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 3, 4, 3, 4 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "ozTREVBARB", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Caglar Gulcehre", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "EPFL" }, { "name": "Kumar Shridhar", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" }, { "name": "Nicholas Monath", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Shivam Adarsh", "gender": "unknown", "institution": "University of Zurich", "country": "CH", "position": "MS student" }, { "name": "MRINMAYA SACHAN", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 32, "n_ref": 60, "n_ref_all": 79, "n_fig": 8, "n_tab": 1, "L_tab": 79, "n_element_tab": 17, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 79, "n_element_tab_1": 17, "formula_len_all": 1261, "formula_len_all_1": 1271, "len_all": 115254, "len_all_1": 56335, "len_abs": 1476, "len_title": 126, "len_sents": 32397, "len_sents_1": 25881, "n_sents": 232, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1504, "title": "SIKeD: Self-guided Iterative Knowledge Distillation for Mathematical Reasoning", "abs": "Large Language Models (LLMs) can transfer their reasoning skills to smaller models by teaching them to generate the intermediate reasoning process required to solve multistep reasoning tasks. While LLMs can accurately solve reasoning tasks through a variety of strategies, even without fine-tuning, smaller models are not expressive enough to fit the LLMs distribution on all strategies when distilled and tend to prioritize one strategy over the others.\nThis reliance on one strategy poses a challenge for smaller models when attempting to solve reasoning tasks that may be difficult with their preferred strategy.\nTo address this, we propose a distillation method *SIKeD*: **S**elf-guided **I**terative **K**nowledge **D**istillation, where the LLM teaches the smaller model to approach a task using different strategies and the smaller model uses its self-generated on-policy outputs to choose the most suitable strategy for the given task. The training continues in a *self-guided* iterative manner, where for each training iteration, a decision is made on how to combine the LLM data with the self-generated outputs. Unlike traditional distillation methods, *SIKeD* allows the smaller model to learn *which* strategy is suitable for a given task while continuously learning to solve a task using different strategies.\nOur experiments on various mathematical reasoning datasets show that *SIKeD* significantly outperforms traditional distillation techniques across smaller models of different sizes.", "keywords": [ "Reasoning", "Knowledge Distillation", "LLM", "SLM" ], "rating_list": [ 5, 3, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "oycEeFXX74", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Adam Kaufman", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "Undergrad student" }, { "name": "Akbir Khan", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Aryan Bhatt", "gender": "Male", "institution": "Redwood Research", "country": "US", "position": "Researcher" }, { "name": "Buck Shlegeris", "gender": "Male", "institution": "Redwood Research", "country": "US", "position": "Principal Researcher" }, { "name": "Cody Rushing", "gender": "Male", "institution": "Redwood Research", "country": "US", "position": "Researcher" }, { "name": "Tyler Tracy", "gender": "Male", "institution": "ML Alignment & Theory Scholars", "country": "US", "position": "Researcher" }, { "name": "Vasil Georgiev", "gender": "Male", "institution": "Independent", "country": "BG", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 26, "n_ref": 40, "n_ref_all": 54, "n_fig": 4, "n_tab": 2, "L_tab": 1183, "n_element_tab": 69, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15, "formula_len_all_1": 0, "len_all": 177042, "len_all_1": 62892, "len_abs": 1936, "len_title": 105, "len_sents": 69205, "len_sents_1": 31008, "n_sents": 512, "n_sents_1": 230, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1951, "title": "Shell Games: Control Protocols for Adversarial AI Agents", "abs": "As more powerful AI agents are deployed with greater tool use and autonomy, it will become increasingly important to ensure they cannot take severely damaging actions. Recent work introduced \"control evaluations,\" which study whether monitoring and security protocols intended to prevent dangerous actions are robust to intentional subversion by these AIs. In a control evaluation, a *red team* seeks to subvert the safety protocols built by the *blue team*. The blue team has access to a weaker *trusted* model and a limited amount of high-quality, trusted labor to attempt to safely extract work from a stronger *untrusted* model. Prior work conducted a control evaluation using competition programming problems with the red team objective of writing subtly incorrect (or \"backdoored\") solutions (Greenblatt et al. 2023). We extend this work by performing a control evaluation in a more realistic setting in which an untrusted LLM agent executes commands in a bash shell. The blue team seeks to use that agent to complete our novel dataset of 200 challenging multi-step tasks, each of which requires the agent to iteratively develop a bash script that passes test cases. The red team creates an attack policy that uses the untrusted model to covertly download and execute a dummy malware file while completing tasks without getting caught by the blue team's safety protocols. This agentic setting allows both teams to explore complex new strategies that have not been previously studied. We extend protocols from previous work to this agentic setting, propose novel protocols, and evaluate each on our safety and usefulness metrics. We find that our protocols substantially improve the Pareto frontier of usefulness and safety relative to simple baselines: one of our best protocols exhibits approximately a 2.5 times improvement in safety metrics with no statistically significant decrease in usefulness compared to our trusted monitoring baseline.", "keywords": [ "Safety", "Code generation", "Red teaming", "Large language models", "Monitoring", "Alignment", "Security", "AI Control" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 2, 3, 1, 3 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "oyXoGJQlUf", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Chuchu Fan", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Jacob Arkin", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "James Neidhoefer", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Nicholas Roy", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 34, "n_ref": 67, "n_ref_all": 133, "n_fig": 14, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 14, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 82, "formula_len_all_1": 26, "len_all": 94209, "len_all_1": 60223, "len_abs": 1280, "len_title": 118, "len_sents": 34580, "len_sents_1": 31262, "n_sents": 208, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1284, "title": "Grounded Robotic Action-Rule Induction through Language Models (GRAIL)", "abs": "A significant body of recent work illustrates that two components of autonomous planning agents nearly always require manual pre-specification by human experts: the identification and grounding of action symbols (such as \u201cturn right\u201d), and the generation of PDDL action rules (including rule name, parameters, preconditions, and effects). We present the Grounded Robotic Action-Rule Induction through Language Models (GRAIL) system, which, in addition to automating those two processes, also contributes to the expanding research on PDDL model optimization. In this paper, we show how large language models (LLMs) can be used to cluster the sensorimotor experience of the robot and automatically generate useful symbolic abstractions about the robot\u2019s capabilities and environment. This language-grounded abstraction allows the learned domain to be modified and used for planning without additional retraining. We evaluate the approach in a standard maze domain and show results for automated symbol identification and grounding, automated rule generation, simulation-based rule validation, and PDDL model optimization. We also discuss and illustrate the advantages of the hybrid neuro-symbolic GRAIL system over traditional symbolic or purely data-driven approaches to similar tasks.", "keywords": [ "Autonomous Planning Agents", "Symbol Grounding", "Frame Problem", "PDDL Model Optimization", "Robotics" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "oyIXleoQ7Z", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Ang Li", "gender": "Male", "institution": "Simular", "country": "", "position": "Principal Researcher" }, { "name": "Hieu Trung Nguyen", "gender": "Male", "institution": "Vinai Research", "country": "VN", "position": "Intern" }, { "name": "Hoang Pham", "gender": "Male", "institution": "The University of Warwick", "country": "GB", "position": "PhD student" }, { "name": "Nghia D. Nguyen", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Viet Anh Nguyen", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Khoa Doan", "gender": "unknown", "institution": "VinUniversity", "country": "VN", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 43, "n_ref": 69, "n_ref_all": 86, "n_fig": 4, "n_tab": 2, "L_tab": 328, "n_element_tab": 42, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 328, "n_element_tab_1": 42, "formula_len_all": 828, "formula_len_all_1": 828, "len_all": 121652, "len_all_1": 64615, "len_abs": 1070, "len_title": 122, "len_sents": 35134, "len_sents_1": 29603, "n_sents": 277, "n_sents_1": 231, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1074, "title": "Forget but Recall: Incremental Latent Rectification in Continual Learning", "abs": "Intrinsic capability to continuously learn a changing data stream is a desideratum of deep neural networks (DNNs). However, current DNNs suffer from catastrophic forgetting, which hinders remembering past knowledge. To mitigate this issue, existing Continual Learning (CL) approaches either retain exemplars for replay, regularize learning, or allocate dedicated capacity for new tasks. This paper investigates an unexplored CL direction for incremental learning called Incremental Latent Rectification or ILR. In a nutshell, LRB learns to propagate with correction (or rectify) the representation from the current trained DNN backward to the representation space of the old task, where performing predictive decisions is easier. This rectification process only employs a chain of small representation mapping networks, called rectifier units. Empirical experiments on several continual learning benchmarks, including CIFAR10, CIFAR100, and Tiny ImageNet, demonstrate the effectiveness and potential of this novel CL direction compared to existing representative CL methods.", "keywords": [ "continual learning" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "owXylt8hZj", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 10, "author_info_list": [ { "name": "Alan Yuille", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Angtian Wang", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Feng Wang", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Guoyizhe Wei", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Sucheng Ren", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Timing Yang", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Intern" }, { "name": "Wei Shao", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Yaodong Yu", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department, University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Yuyin Zhou", "gender": "unknown", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "cihang xie", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 45, "n_ref": 81, "n_ref_all": 100, "n_fig": 4, "n_tab": 7, "L_tab": 2275, "n_element_tab": 277, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1962, "n_element_tab_1": 237, "formula_len_all": 365, "formula_len_all_1": 365, "len_all": 121804, "len_all_1": 60052, "len_abs": 1106, "len_title": 104, "len_sents": 30067, "len_sents_1": 24030, "n_sents": 207, "n_sents_1": 161, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1114, "title": "Causal Image Modeling for Efficient Visual Understanding", "abs": "In this work, we present a comprehensive analysis of causal image modeling and introduce the Adventurer series models where we treat images as sequences of patch tokens and employ uni-directional language models to learn visual representations. This modeling paradigm allows us to process images in a recurrent formulation with linear complexity relative to the sequence length, which can effectively address the memory and computation explosion issues posed by high-resolution and fine-grained images. In detail, we introduce two simple designs that seamlessly integrate image inputs into the causal inference framework: a global pooling token placed at the beginning of the sequence and a flipping operation between every two layers. Extensive empirical studies demonstrate the significant efficiency and effectiveness of this causal image modeling paradigm. For example, our base-sized Adventurer model attains a competitive test accuracy of 84.0% on the standard ImageNet-1k benchmark with 216 images/s training throughput, which is 5.3 times more efficient than vision transformers to achieve the same result.", "keywords": [ "Visual backbone", "causal modeling", "Mamba" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "owR9ofvkFQ", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Fei Lu", "gender": "Not Specified", "institution": "Johns Hopkins University", "country": "", "position": "Associate Professor" }, { "name": "Fei Xing", "gender": "unknown", "institution": "Mathematica Policy Research", "country": "", "position": "Researcher" }, { "name": "Kai Zou", "gender": "Male", "institution": "NetMind.AI", "country": "GB", "position": "Founder CEO" }, { "name": "Meng Fang", "gender": "Male", "institution": "University of Liverpool", "country": "", "position": "Assistant Professor" }, { "name": "Xiangpeng Wan", "gender": "Male", "institution": "Netmind.ai", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 11, "n_ref": 28, "n_ref_all": 44, "n_fig": 4, "n_tab": 6, "L_tab": 1592, "n_element_tab": 102, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 475, "n_element_tab_1": 5, "formula_len_all": 529, "formula_len_all_1": 342, "len_all": 102917, "len_all_1": 61620, "len_abs": 1477, "len_title": 158, "len_sents": 35396, "len_sents_1": 25327, "n_sents": 291, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 110, "L_abs": 1485, "title": "MathOdyssey: Benchmarking Mathematical Problem-Solving Skills in Large Language Models Using Odyssey Math Data", "abs": "Large language models (LLMs) have significantly advanced natural language understanding and demonstrated strong problem-solving abilities. Despite these successes, most LLMs still struggle with solving mathematical problems due to the intricate reasoning required. This paper investigates the mathematical problem-solving capabilities of LLMs using the newly developed ``MathOdyssey'' dataset. The dataset includes diverse mathematical problems at high school and university levels, created by experts from notable institutions to rigorously test LLMs in advanced problem-solving scenarios and cover a wider range of subject areas. By providing the MathOdyssey dataset as a resource to the AI community, we aim to contribute to the understanding and improvement of AI capabilities in complex mathematical problem-solving. We conduct benchmarking on open-source models, such as Llama-3, and closed-source models from the GPT series and Gemini models. Our results indicate that while LLMs perform well on routine and moderately difficult tasks, they face significant challenges with Olympiad-level problems and complex university-level questions. Our analysis shows a narrowing performance gap between open-source and closed-source models, yet substantial challenges remain, particularly with the most demanding problems. This study highlights the ongoing need for research to enhance the mathematical reasoning of LLMs. \nThe dataset, results, and evaluation code are publicly available.", "keywords": [ "Math", "LLMs" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 4, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "owP2mymrTD", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Bin Cui", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Fan Yang", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Hao Liang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "KeerLu", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Mingyang Chen", "gender": "Male", "institution": "Baichuan Inc.", "country": "CN", "position": "Researcher" }, { "name": "Tianpeng Li", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zenan Zhou", "gender": "Male", "institution": "baichuan", "country": "CN", "position": "Researcher" }, { "name": "sunhaoze", "gender": "Male", "institution": "Baichuan inc", "country": "CN", "position": "Researcher" }, { "name": "weipeng chen", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhang wen tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 24, "n_ref": 66, "n_ref_all": 80, "n_fig": 4, "n_tab": 5, "L_tab": 2290, "n_element_tab": 160, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 225308, "len_all_1": 65084, "len_abs": 1418, "len_title": 134, "len_sents": 81456, "len_sents_1": 32838, "n_sents": 659, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1423, "title": "Facilitating Multi-turn Function Calling for LLMs via Compositional Instruction Tuning", "abs": "Large Language Models (LLMs) have exhibited significant potential in performing diverse tasks, including the ability to call functions or use external tools to enhance their performance. While current research on function calling by LLMs primarily focuses on single-turn interactions, this paper addresses the overlooked necessity for LLMs to engage in multi-turn function calling\u2014critical for handling compositional, real-world queries that require planning with functions but not only use functions. To facilitate this, we introduce an approach, BUTTON, which generates synthetic compositional instruction tuning data via bottom-up instruction construction and top-down trajectory generation. In the bottom-up phase, we generate simple atomic tasks based on real-world scenarios and build compositional tasks using heuristic strategies based on atomic tasks. Corresponding functions are then developed for these compositional tasks. The top-down phase features a multi-agent environment where interactions among simulated humans, assistants, and tools are utilized to gather multi-turn function calling trajectories. This approach ensures task compositionality and allows for effective function and trajectory generation by examining atomic tasks within compositional tasks. We produce a dataset BUTTONInstruct comprising 8k data points and demonstrate its effectiveness through extensive experiments across various LLMs.", "keywords": [ "large language model", "function calling", "instruction tuning", "synthetic data" ], "rating_list": [ 5, 8, 8, 5 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 3, 5, 5, 3 ] }, { "paper_id": "owEQ0FTfVj", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Jian Tang", "gender": "unknown", "institution": "HEC Montreal", "country": "", "position": "Assistant Professor" }, { "name": "Ling Yang", "gender": "Male", "institution": "DeepSeek AI", "country": "CN", "position": "Intern" }, { "name": "Minghao Xu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yihang Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Yunteng Geng", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Zhang wen tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 70, "n_ref": 112, "n_ref_all": 122, "n_fig": 2, "n_tab": 3, "L_tab": 5433, "n_element_tab": 477, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 4912, "n_element_tab_1": 451, "formula_len_all": 638, "formula_len_all_1": 639, "len_all": 165563, "len_all_1": 75676, "len_abs": 1230, "len_title": 128, "len_sents": 35972, "len_sents_1": 30831, "n_sents": 264, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1382, "title": "GlycanML: A Multi-Task and Multi-Structure Benchmark for Glycan Machine Learning", "abs": "Glycans are basic biomolecules and perform essential functions within living organisms. The rapid increase of functional glycan data provides a good opportunity for machine learning solutions to glycan understanding. However, there still lacks a standard machine learning benchmark for glycan property and function prediction. In this work, we fill this blank by building a comprehensive benchmark for Glycan Machine Learning (GlycanML). The GlycanML benchmark consists of diverse types of tasks including glycan taxonomy prediction, glycan immunogenicity prediction, glycosylation type prediction, and protein-glycan interaction prediction. Glycans can be represented by both sequences and graphs in GlycanML, which enables us to extensively evaluate sequence-based models and graph neural networks (GNNs) on benchmark tasks. Furthermore, by concurrently performing eight glycan taxonomy prediction tasks, we introduce the GlycanML-MTL testbed for multi-task learning (MTL) algorithms. Also, we evaluate how taxonomy prediction can boost other three function prediction tasks by MTL. Experimental results show the superiority of modeling glycans with multi-relational GNNs, and suitable MTL methods can further boost model performance. We provide all datasets and source codes at https://github.com/GlycanML/GlycanML and maintain a leaderboard at https://GlycanML.github.io/project", "keywords": [ "Glycan Machine Learning", "Representation Learning", "Multi-Task Learning" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 2, 3, 3 ] }, { "paper_id": "ov678VcvlO", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Ehsan Shareghi", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" }, { "name": "Hao Yang", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Lizhen Qu", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Lecturer" }, { "name": "Reza Haffari", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 68, "n_ref_all": 104, "n_fig": 7, "n_tab": 5, "L_tab": 866, "n_element_tab": 70, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 569, "n_element_tab_1": 73, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 148218, "len_all_1": 71706, "len_abs": 4020, "len_title": 126, "len_sents": 43599, "len_sents_1": 34010, "n_sents": 358, "n_sents_1": 265, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1325, "title": "Jigsaw Puzzles: Splitting Harmful Questions to Jailbreak Large Language Models", "abs": "Large language models (LLMs) have exhibited outstanding performance in engaging with humans and addressing complex questions by leveraging their vast implicit knowledge and robust reasoning capabilities. However, such models are vulnerable to jailbreak attacks, leading to the generation of harmful responses. Despite recent research on single-turn jailbreak strategies to facilitate the development of defence mechanisms, the challenge of revealing vulnerabilities under multi-turn setting remains relatively under-explored. In this work, we propose Jigsaw Puzzles (JSP), a straightforward yet effective multi-turn jailbreak strategy against the advanced LLMs. JSP splits questions into harmless fractions as the input of each turn, and requests LLMs to reconstruct and respond to questions under multi-turn interaction. Our experimental results demonstrate that the proposed JSP jailbreak bypasses original safeguards against explicitly harmful content, achieving an average attack success rate of 93.76% on 189 harmful queries across 5 advanced LLMs (Gemini-1.5-Pro, Llama-3.1-70B, GPT-4, GPT-4o, GPT-4o-mini). Moreover, JSP achieves a state-of-the-art attack success rate of 92% on GPT-4 on the harmful query benchmark, and exhibits strong resistant to defence strategies. Warning: this paper contains offensive examples.", "keywords": [ "Jailbreak", "Red Teaming", "LLMs", "Safety", "Multi-turn Interaction" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "ouRX6A8RQJ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Jean-Francois Ton", "gender": "Not Specified", "institution": "Bytedance", "country": "GB", "position": "Researcher" }, { "name": "Muhammad Faaiz Taufiq", "gender": "Male", "institution": "ByteDance Inc.", "country": "GB", "position": "Researcher" }, { "name": "Yang Liu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" } ], "n_formula": 34, "n_formula_1": 16, "n_ref_uni": 22, "n_ref": 51, "n_ref_all": 86, "n_fig": 13, "n_tab": 4, "L_tab": 694, "n_element_tab": 76, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 528, "n_element_tab_1": 47, "formula_len_all": 2257, "formula_len_all_1": 1011, "len_all": 228658, "len_all_1": 73874, "len_abs": 916, "len_title": 113, "len_sents": 58841, "len_sents_1": 34152, "n_sents": 483, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 919, "title": "Understanding Chain-of-Thought in LLMs Through Information Theory", "abs": "Large Language Models (LLMs) have shown impressive performance in complex reasoning tasks through the use of Chain-of-Thought (CoT) reasoning, allowing models to break down problems into manageable sub-tasks. However, existing CoT evaluation techniques either require annotated CoT data or fall short in accurately assessing intermediate reasoning steps, leading to high rates of false positives. In this paper, we formalize CoT reasoning in LLMs through an information-theoretic lens. Specifically, our framework quantifies the `information gain' at each reasoning step, enabling the identification of failure modes in LLMs without the need for expensive annotated datasets. We demonstrate the efficacy of our approach through extensive experiments on toy and GSM-8K data, where it significantly outperforms existing outcome-based methods by providing more accurate insights into model performance on individual tasks.", "keywords": [ "Large Language models", "Chain-of-thought" ], "rating_list": [ 8, 5, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 2, 4, 4, 4, 4 ] }, { "paper_id": "otYAwaTDk1", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Dominik Roth", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "", "position": "MS student" }, { "name": "Fabian Otto", "gender": "unknown", "institution": "Microsoft", "country": "GB", "position": "Researcher" }, { "name": "Onur Celik", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Gerhard Neumann", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 9, "n_ref_uni": 43, "n_ref": 88, "n_ref_all": 120, "n_fig": 18, "n_tab": 3, "L_tab": 1783, "n_element_tab": 330, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 301, "n_element_tab_1": 109, "formula_len_all": 660, "formula_len_all_1": 420, "len_all": 149031, "len_all_1": 58109, "len_abs": 1052, "len_title": 54, "len_sents": 40592, "len_sents_1": 26991, "n_sents": 312, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1055, "title": "Perlin Noise for Exploration in Reinforcement Learning", "abs": "Reinforcement Learning (RL) enables agents to solve tasks by autonomously acquiring policies by interacting with the environment receiving sparse or noisy feedback in the form of a reward. However, achieving successful optimization in RL requires efficient exploration, which remains a significant challenge, particularly in continuous action spaces. Existing exploration techniques often exhibit limited state-space reach and fail to overcome local optima, resulting in suboptimal policies. Additionally, these techniques can cause erratic movements, posing risks when applied to real-world robots.\nIn this work, we introduce a novel exploration strategy leveraging Perlin Noise, a gradient noise function that generates smooth, continuous disturbances, thus enhancing the agent's performance by promoting structured exploration and fluid motions. We quantitatively demonstrate the benefits of our approach compared to state-of-the-art methods, showing that it outperforms both unstructured and structured techniques in thorough experimental evaluations.", "keywords": [ "Reinforcement Learning", "Exploration Strategies", "Perlin Noise", "Policy Optimization", "Structured Exploration" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 3, 2, 1 ], "confidence_list": [ 3, 3, 5 ] }, { "paper_id": "otXB6odSG8", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Alexander Gavrikov", "gender": "Male", "institution": "P.P. Shirshov Institute of Oceanology of the Russian Academy of Sciences", "country": "RU", "position": "Postdoc" }, { "name": "Chesnokov Alexey Aleksandrovich", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Researcher" }, { "name": "Evgeny Burnaev", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Full Professor" }, { "name": "Nikolay Yavich", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Researcher" }, { "name": "Stefan Maria Ailuro", "gender": "Not Specified", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "Intern" }, { "name": "Vladimir Vanovskiy", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 20, "n_ref": 26, "n_ref_all": 37, "n_fig": 7, "n_tab": 5, "L_tab": 1593, "n_element_tab": 243, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 1151, "n_element_tab_1": 162, "formula_len_all": 351, "formula_len_all_1": 350, "len_all": 76668, "len_all_1": 49888, "len_abs": 1162, "len_title": 99, "len_sents": 22013, "len_sents_1": 21705, "n_sents": 179, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1171, "title": "Atmospheric Radiation Parameterization by Neural Ordinary Differential Equations and Related Models", "abs": "Radiation parameterization schemes are crucial components of weather and climate models, however, they are known to be computationally intensive. Alternatively, they can be emulated with machine learning (ML) regression models. Mainly vertical energy propagation motivates the usage of ML models featuring sequential data processing. We investigate these and related models for radiation parameterization using atmospheric data modeled within an Arctic region. We observe that Neural ODE performs best in predicting both the long- and short-wave heating rates. Furthermore, we substitute the architecture with its discrete form to boost its efficiency while preserving competitive performance. The practical applicability of the models is studied for different model sizes. Finally, we link the trained neural network to the operational weather forecast model and assessed its performance versus the conventional radiation parameterization.\nWe receive a speedup of 26.5 times of the radiation steps\nwithout significant loss of accuracy. The proposed parameterization \nemulator dramatically reduces the computational burden and the carbon footprint of weather forecasting.", "keywords": [ "regression models", "neural ODEs", "numerical weather prediction", "neural networks", "radiative transfer" ], "rating_list": [ 3, 1, 3, 5 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "otW0TJOUYF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Agam Bhatia", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Daniel LK Yamins", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Logan Cross", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Postdoc" }, { "name": "Nick Haber", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Violet Xiang", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 33, "n_ref": 70, "n_ref_all": 97, "n_fig": 13, "n_tab": 6, "L_tab": 2718, "n_element_tab": 75, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 2673, "n_element_tab_1": 31, "formula_len_all": 851, "formula_len_all_1": 173, "len_all": 239602, "len_all_1": 60409, "len_abs": 1817, "len_title": 144, "len_sents": 96241, "len_sents_1": 29515, "n_sents": 694, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1204, "title": "Hypothetical Minds: Scaffolding Theory of Mind for Multi-Agent Tasks with Large Language Models", "abs": "Multi-agent reinforcement learning (MARL) methods struggle with the non-stationarity of multi-agent systems and fail to adaptively learn online when tested with novel agents. Here, we leverage large language models (LLMs) to create an autonomous agent that can handle these challenges. Our agent, Hypothetical Minds, consists of a cognitively-inspired architecture, featuring modular components for perception, memory, and hierarchical planning over two levels of abstraction. We introduce the Theory of Mind module that scaffolds the high-level planning process by generating hypotheses about other agents' strategies in natural language. It then evaluates and iteratively refines these hypotheses by reinforcing hypotheses that make correct predictions about the other agents' behavior. Hypothetical Minds significantly improves performance over previous LLM-agent and RL baselines on a range of competitive, mixed motive, and collaborative domains in the Melting Pot benchmark, including both dyadic and population-based environments. Additionally, comparisons against LLM-agent baselines and ablations reveal the importance of hypothesis evaluation and refinement for succeeding on complex scenarios.", "keywords": [ "LLM agents", "language agents", "theory of mind", "multi-agent" ], "rating_list": [ 5, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 1 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "otTfoKL5sJ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Donghyeon Ko", "gender": "Male", "institution": "NAVER", "country": "", "position": "Researcher" }, { "name": "Dong-Hyun Kwak", "gender": "Male", "institution": "NAVER", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 46, "n_ref": 91, "n_ref_all": 113, "n_fig": 3, "n_tab": 11, "L_tab": 2270, "n_element_tab": 159, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 937, "n_element_tab_1": 84, "formula_len_all": 470, "formula_len_all_1": 470, "len_all": 161455, "len_all_1": 67180, "len_abs": 1841, "len_title": 81, "len_sents": 38384, "len_sents_1": 30076, "n_sents": 300, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 33, "L_abs": 1930, "title": "HELPFUL-ONLY LARGE LANGUAGE MODEL", "abs": "To know your enemy, you must become your enemy. Sun Tzu stated in $\\textit{The Art of War}$. Often, it is crucial to synthesize data containing harmful content using large language models (LLMs) in order to train harmless LLMs. Methods by which synthesized data can be utilized include using it as training data to provide negative signals to the model, as automatic red-teaming data to identify vulnerabilities of the model and more. However, aligned LLMs struggle to generate harmful responses. In this paper, we propose the $\\textit{refusal-free}$ training method to reach a $\\textbf{Helpful-Only LLM}$ that maintains the helpfulness of the state-of-the-art (SOTA) LLMs while allowing harmful response generation. The $\\textit{refusal-free}$ training method filters the instances that refuse an user's request from the datasets. We demonstrate that the $\\textit{refusal-free}$ training dramatically decreases the rate at which the LLM generates refusal responses (refusal rate) by 60.12% without sacrificing its helpfulness. Also, we are aware of the possibility that the progress in this direction could lead to irreversible consequences. A powerful model that does not reject harmful requests and executes them all could be exploited for illicit purposes such as the creation of indiscriminate weapons or hacking. However, once again, we believe it is important to be the one to break an LLM and study how an LLM can be broken in advance, including understanding the boundaries a $\\textbf{Helpful-Only LLM}$ can reach and identifying its inherent tendencies. We emphasize that this study is wholly for academic purpose and is aimed at paving the way toward a harmless LLM. This study calls for the researchers to acknowledge the potential failures of LLMs and take steps to prevent the breakdowns. $\\textbf{Content Warning:}$ This paper contains examples that may be offensive in nature, and reader discretion is recommended.", "keywords": [ "AI Safety", "AI alignment", "Large Language Model", "Offline Reinforcement Learning", "Data Selection", "Machine Learning", "Deep Learning", "Natural Language Processing" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "orr5uPZY28", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 5, "author_info_list": [ { "name": "Lei Zou", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Ruoyu Zhang", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Tianhao Wu", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Yanzeng Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yunfan Xiong", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" } ], "n_formula": 24, "n_formula_1": 10, "n_ref_uni": 20, "n_ref": 36, "n_ref_all": 53, "n_fig": 7, "n_tab": 6, "L_tab": 2323, "n_element_tab": 215, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 886, "n_element_tab_1": 70, "formula_len_all": 1583, "formula_len_all_1": 442, "len_all": 127359, "len_all_1": 54084, "len_abs": 1445, "len_title": 117, "len_sents": 43815, "len_sents_1": 24194, "n_sents": 379, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1500, "title": "DySpec: Faster Speculative Decoding with Dynamic Token Tree Structure", "abs": "While speculative decoding has recently appeared as a promising direction for accelerating the inference of large language models (LLMs), the speedup and scalability are strongly bounded by the token acceptance rate.\nPrevalent methods usually organize predicted tokens as independent chains or fixed token trees, which fails to generalize to diverse query distributions. \nIn this paper, we propose \\textsc{DySpec}, a faster speculative decoding algorithm with a novel dynamic token tree structure. \nWe begin by bridging the draft distribution and acceptance rate from \nintuitive and empirical clues, and successfully show that the two variables are strongly correlated. Based on this, we employ a greedy strategy to dynamically expand the token tree at run time. Theoretically, we show that our method can achieve optimal results under mild assumptions. Empirically, \\textsc{DySpec} yields a higher acceptance rate and speedup than fixed trees. \\textsc{DySpec} can drastically improve the throughput and reduce the latency of token generation across various data distribution and model sizes, which significantly outperforms strong competitors, including Specinfer and Sequoia. Under low temperature setting, \\textsc{DySpec} can improve the throughput up to 9.10x and reduce the latency up to 9.4x on Llama2-70B. Under high temperature setting, \\textsc{DySpec} can also improve the throughput up to 6.21x, despite the increasing difficulty of speculating more than one token per step for draft model.", "keywords": [ "inference methods", "efficient inference", "speculative decoding" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 2, 4 ] }, { "paper_id": "ori83fBg71", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Lianwen Jin", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Professor" }, { "name": "Peirong Zhang", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 22, "n_ref": 41, "n_ref_all": 59, "n_fig": 5, "n_tab": 6, "L_tab": 3414, "n_element_tab": 233, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2210, "n_element_tab_1": 181, "formula_len_all": 1044, "formula_len_all_1": 966, "len_all": 118143, "len_all_1": 70736, "len_abs": 1307, "len_title": 159, "len_sents": 35775, "len_sents_1": 31250, "n_sents": 248, "n_sents_1": 247, "venue": "ICLR2025", "decision": "withdraw", "L_title": 110, "L_abs": 1311, "title": "SPECTRUM: Empowering Online Handwriting Verification via Temporal-Frequency Multimodal Representation Learning", "abs": "Tapping into the uncharted multimodal representation learning in online handwriting verification (OHV), we propose SPECTRUM, a temporal-frequency synergistic model tailored to enhance handwriting representations. SPECTRUM comprises three core components: (1) a multi-scale interactor that interweaves fine-grained temporal and frequency features across multiple scales through complementary domain interaction; (2) a self-gated fusion module, dynamically integrating global temporal and frequency features via self-driven balancing. Collectively, these two components achieve micro-to-macro multimodal integration; (3) a multimodal distance-based verifier that fully harnesses temporal and frequency representations, sharpening genuine-forged discrimination beyond conventional temporal-only approaches. Extensive experiments demonstrate SPECTRUM's pronounced outperformance over existing OHV methods. Furthermore, we reveal that incorporating multiple handwritten biometrics fundamentally improves the discriminatory power of individual writing features. These findings not only validate the efficacy of multimodal learning in OHV but also encourage broader multimodal research across both feature and biometric domains, potentially opening new avenues for future explorations. Code will be publicly available.", "keywords": [ "Online handwriting verification; Multimodal representation learning; Temporal and frequency learning; Handwritten biometrics" ], "rating_list": [ 8, 5, 5, 3 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "orG37FHN4b", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Stephen Bauer", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Sudhakar Pamarti", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 12, "n_ref_uni": 22, "n_ref": 37, "n_ref_all": 51, "n_fig": 5, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1476, "formula_len_all_1": 899, "len_all": 97200, "len_all_1": 57359, "len_abs": 1610, "len_title": 94, "len_sents": 34195, "len_sents_1": 27278, "n_sents": 277, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 45, "L_abs": 1536, "title": "Angle-DFQ: Angle aware data free quantization", "abs": "Data free quantization of neural networks is a practical necessity as access to training data in many situations is restricted due to privacy, proprietary concerns, or memory issues. We present a data free weight rounding algorithm for Deep Neural Networks (DNNs) that does not require any training data, synthetic data generation, fine-tuning, or even batch norm statistics. Instead, our approach focuses on preserving the direction of weight vectors during quantization. We demonstrate that traditional weight rounding techniques, that round weights to the nearest quantized level, can result in large angles between the full-precision weight vectors and the quantized weight vectors, particularly under coarse quantization regimes. For a large class of high-dimensional weight vectors in DNNs, this angle error can approach 90 degrees. By minimizing this angle error, we significantly improve top-1 accuracy in quantized DNNs. We analytically derive the angle-minimizing rounding boundaries for ternary quantization under the assumption of Gaussian weights. Building on this, we propose a greedy data-free quantization method based on the cosine similarity between the full-precision weight vectors and the quantized weight vectors. Our approach consistently outperforms existing state-of-the-art data-free quantization techniques and, in several cases, surpasses even data-dependent methods on well-established models such as ResNet-18, VGG-16, and AlexNet with aggressive quantization levels of 3 to 6 bits on the ImageNet dataset.", "keywords": [ "Data free quantization", "Computer vision" ], "rating_list": [ 1, 3, 5 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 1, 1, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "orEX9GKQAD", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Dmitry Osin", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Research Engineer" }, { "name": "Egor Shvetsov", "gender": "Male", "institution": "Skoltech", "country": "RU", "position": "Researcher" }, { "name": "Evgeny Burnaev", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Full Professor" }, { "name": "Igor Udovichenko", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Researcher" }, { "name": "Viktor Moskvoretskii", "gender": "Not Specified", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 45, "n_ref": 91, "n_ref_all": 115, "n_fig": 5, "n_tab": 11, "L_tab": 7046, "n_element_tab": 687, "n_fig_1": 4, "n_tab_1": 9, "L_tab_1": 7789, "n_element_tab_1": 643, "formula_len_all": 49, "formula_len_all_1": 0, "len_all": 171117, "len_all_1": 71332, "len_abs": 1372, "len_title": 44, "len_sents": 46662, "len_sents_1": 25908, "n_sents": 381, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1493, "title": "EBES: Easy Benchmarking for Event Sequences", "abs": "Event sequences, characterized by irregular sampling intervals and a mix of categorical and numerical features, are common data structures in various real-world domains such as healthcare, finance, and user interaction logs. Despite advances in temporal data modeling techniques, there is no standardized benchmarks for evaluating their performance on event sequences. This complicates result comparison across different papers due to varying evaluation protocols, potentially misleading progress in this field.\nWe introduce EBES, a comprehensive benchmarking tool with standardized evaluation scenarios and protocols, focusing on regression and classification problems with sequence-level targets. Our library~\\footnote{We attach an archive with the code. The code will be publicly available after the conference decision.} simplifies benchmarking, dataset addition, and method integration through a unified interface. It includes a novel synthetic dataset and provides preprocessed real-world datasets, including the largest publicly available banking dataset.\nOur results provide an in-depth analysis of datasets, identifying some as unsuitable for model comparison. We investigate the importance of modeling temporal and sequential components, as well as the robustness and scaling properties of the models. These findings highlight potential directions for future research. Our benchmark aim is to facilitate reproducible research, expediting progress and increasing real-world impacts.", "keywords": [ "event sequences", "irregularly sampled time series", "benchmark", "temporal point process", "transactions", "sequential learning", "reproducible research" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "orD5t7blqV", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Hamid Kalantari", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Postdoc" }, { "name": "Narges Chehrehrazi", "gender": "Female", "institution": "narbeh", "country": "IR", "position": "Researcher" }, { "name": "Pouria Ramazi", "gender": "Male", "institution": "Brock University", "country": "", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 26, "n_ref": 42, "n_ref_all": 66, "n_fig": 5, "n_tab": 11, "L_tab": 4212, "n_element_tab": 946, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 494, "n_element_tab_1": 97, "formula_len_all": 1779, "formula_len_all_1": 1386, "len_all": 135436, "len_all_1": 69798, "len_abs": 2889, "len_title": 101, "len_sents": 41710, "len_sents_1": 31757, "n_sents": 351, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1065, "title": "The Path-Driven Independence Testing (PIT) Algorithm", "abs": "PC is an efficient constraint-based algorithm for learning the structure of a Bayesian network. However, the required number of conditional independent (CI) tests can make the algorithm practically infeasible or slow for large graphs. We developed a constrained-based algorithm, called the Path-Driven Independence Testing (PIT) Algorithm, which during the learning process, utilizes the information of the partially learned network to reduce the number of CI tests. The idea is that for each pair of variables $X$ and $Y$, instead of checking independence conditioned on every subset of all the neighbors of $X$ (resp. $Y$) as in PC, the search is restricted to only the common neighbors of $X$ and $Y$ and to neighbors connected to $Y$ (resp. $X$) by a path. Also, paths connecting $X$ and $Y$ without a descendant of a common neighbor can be blocked by observing two consecutive nodes on the path. Compared to PC, PIT is proven to conduct at most the same number of CI tests, and experimentally shown to be significantly (up to 7 times) faster and more accurate.", "keywords": [ "Causality", "Bayesian networks", "Structure learning", "Constraint based", "PC algorithm" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "or9OfAC3kb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Tatiana Zemskova", "gender": "Female", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "PhD student" }, { "name": "Dmitry A Yudin", "gender": "Male", "institution": "Artificial Intelligence Research Intsitute", "country": "RU", "position": "Principal Researcher" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 52, "n_ref": 154, "n_ref_all": 175, "n_fig": 6, "n_tab": 3, "L_tab": 535, "n_element_tab": 95, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 431, "n_element_tab_1": 79, "formula_len_all": 339, "formula_len_all_1": 319, "len_all": 156359, "len_all_1": 68289, "len_abs": 1279, "len_title": 144, "len_sents": 40291, "len_sents_1": 32571, "n_sents": 293, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1284, "title": "3DGraphLLM: Combining Semantic Graphs and Large Language Models for 3D Referred Object Grounding", "abs": "A 3D scene graph represents a compact scene model, storing information about the objects and the semantic relationships between them, making its use promising for robotic tasks. When interacting with a user, an embodied intelligent agent should be capable of responding to various queries about the scene formulated in natural language. Large Language Models (LLMs) are beneficial solutions for user-robot interaction due to their natural language understanding and reasoning abilities. Recent methods for creating learnable representations of 3D scenes have demonstrated the potential to improve the quality of LLMs responses by adapting to the 3D world. However, the existing methods do not explicitly utilize information about the semantic relationships between objects, limiting themselves to information about their coordinates. In this work, we propose a method 3DGraphLLM for constructing a learnable representation of a 3D scene graph.\nThe learnable representation is used as input for LLMs to perform 3D vision-language tasks. In our experiments on popular ScanRefer, RIORefer, Multi3DRefer, ScanQA, Sqa3D, and Scan2cap datasets, we demonstrate the advantage of this approach over baseline methods that do not use information about the semantic relationships between objects.", "keywords": [ "Referred Object Grounding", "3D Scene Graph", "LLM", "3D Referred Object Grounding", "3D Scene Understanding" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "or8wkKoBP4", "primary_area": "causal reasoning", "n_author": 2, "author_info_list": [ { "name": "Hamid Kalantari", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Postdoc" }, { "name": "Pouria Ramazi", "gender": "Male", "institution": "Brock University", "country": "", "position": "Assistant Professor" } ], "n_formula": 31, "n_formula_1": 20, "n_ref_uni": 24, "n_ref": 39, "n_ref_all": 50, "n_fig": 7, "n_tab": 2, "L_tab": 34, "n_element_tab": 1, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2967, "formula_len_all_1": 2082, "len_all": 137602, "len_all_1": 78440, "len_abs": 1459, "len_title": 133, "len_sents": 48200, "len_sents_1": 33117, "n_sents": 376, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1504, "title": "Structure Learning for Unfaithful Distributions: The Minimal Dependence Faithfulness", "abs": "Causality detection is to identify the ``true'' directed acyclic graph (DAG) of a causal model from the joint probability distribution of the observed variables.\nAlgorithms such as PC and its modified versions perform this task under the restrictive faithfulness assumption, that is the DAG encodes all conditional independencies imposed by the distribution. \nHowever, all existing algorithms fail to detect the simple structure where a variable is the XOR of several Bernoulli variables, violating faithfulness. We generalize this type of unfaithfulness that appears in other, non-XOR, examples and define the \\emph{minimal dependence} of a given variable $X$ as the set of variables, such that $X$ is independent of each variable in the set but depends on at least one of them, the \\emph{dependent member} if conditioned on the remainder of the set.\nMinimal dependencies of size at least two violate faithfulness. Consequently, we relax faithfulness to \\emph{minimal dependence faithfulness}, restricting the neighbors of a node to its dependent members, and impose \\emph{minimal orientation faithfulness} that generalizes the orientation rules under faithfulness.\nWe then determine the structure of the dependent members of a node $X$ in the true DAG and show that they are connected to $X$ either directly or indirectly by a collider. \nFinally, we provide a sound and complete modification of the PC algorithm to detect this kind of unfaithfulness and output all possible candidates for the true DAG.", "keywords": [ "Bayesian networks", "Causality", "Faithfulness", "Structure learning" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 1, 2, 1, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "or8mMhmyRV", "primary_area": "reinforcement learning", "n_author": 10, "author_info_list": [ { "name": "Doina Precup", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Research Team Lead" }, { "name": "Marlos C. Machado", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Assistant Professor" }, { "name": "Martin Klissarov", "gender": "Unspecified", "institution": "McGill University", "country": "", "position": "PhD student" }, { "name": "Mikael Henaff", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Pascal Vincent", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "Adjunct Professor" }, { "name": "Pierluca D'Oro", "gender": "Male", "institution": "Facebook", "country": "CA", "position": "Researcher" }, { "name": "Pierre-Luc Bacon", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "Assistant Professor" }, { "name": "Shagun Sodhani", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Amy Zhang", "gender": "Female", "institution": "University of Texas at Austin", "country": "", "position": "Assistant Professor" }, { "name": "Roberta Raileanu", "gender": "Female", "institution": "Facebook", "country": "", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 51, "n_ref": 96, "n_ref_all": 122, "n_fig": 15, "n_tab": 4, "L_tab": 962, "n_element_tab": 51, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 857, "n_element_tab_1": 37, "formula_len_all": 7236, "formula_len_all_1": 6931, "len_all": 229597, "len_all_1": 83515, "len_abs": 891, "len_title": 112, "len_sents": 68894, "len_sents_1": 34133, "n_sents": 487, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 64, "L_abs": 895, "title": "MaestroMotif: Skill Design from Artificial Intelligence Feedback", "abs": "Describing skills in natural language has the potential to provide an accessible way to inject human knowledge about decision-making into an AI system. We present MaestroMotif, a method for AI-assisted skill design, which yields high-performing and adaptable agents. MaestroMotif leverages the capabilities of Large Language Models (LLMs) to effectively create and reuse skills. It first uses an LLM's feedback to automatically design rewards corresponding to each skill, starting from their natural language description. Then, it employs an LLM's code generation abilities, together with reinforcement learning, for training the skills and combining them to implement complex behaviors specified in language. We evaluate MaestroMotif using a suite of complex tasks in the NetHack Learning Environment (NLE), demonstrating that it surpasses existing approaches in both performance and usability.", "keywords": [ "Hierarchical RL", "Reinforcement Learning", "LLMs" ], "rating_list": [ 8, 10, 8, 5 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 4, 3, 2 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "oqsQbn4XfT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Abdul Waheed", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Bhiksha Raj", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Hao Chen", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Jindong Wang", "gender": "Male", "institution": "William & Mary", "country": "US", "position": "Assistant Professor" }, { "name": "Marah I Abdin", "gender": "Female", "institution": "Research, Microsoft", "country": "", "position": "Researcher" }, { "name": "Xiang Li", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Yidong Wang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 2, "n_ref_uni": 68, "n_ref": 162, "n_ref_all": 186, "n_fig": 8, "n_tab": 19, "L_tab": 11988, "n_element_tab": 933, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1558, "n_element_tab_1": 281, "formula_len_all": 867, "formula_len_all_1": 40, "len_all": 446123, "len_all_1": 76595, "len_abs": 4213, "len_title": 132, "len_sents": 135534, "len_sents_1": 34997, "n_sents": 946, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1220, "title": "On the Diversity of Synthetic Data and its Impact on Training Large Language Models", "abs": "The rise of Large Language Models (LLMs) has accentuated the need for diverse, high-quality pre-training data. \nSynthetic data emerges as a viable solution to the challenges of data scarcity and inaccessibility.\nWhile previous literature has focused predominantly on the quality and quantity of real data, our work enables the measurement of diversity in synthetic data and explores its impact on LLM performance. \nWe study the downstream effects of synthetic data diversity during both the pre-training and fine-tuning stages by introducing a new diversity metric, LLM cluster-agent, designed to evaluate the diversity of synthetic datasets. \nThrough a series of controlled experiments with models of 350M and 1.4B parameters, we demonstrate that the proposed cluster-based LLM scoring of diversity correlates positively with both pre-training and supervised fine-tuning performance. \nOur findings also reveal that synthetic data diversity in pre-training affects supervised fine-tuning more significantly than pre-training itself, even for smaller models. \nWe hope this study advances our understanding of the optimal use of synthetic data in LLM training and opens new avenues for efficient data generation processes.", "keywords": [ "Synthetic Data Pre-training", "Large Language Models" ], "rating_list": [ 6, 3, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 2, 4, 2 ] }, { "paper_id": "oqdcThIQjA", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Crystal Maung", "gender": "Female", "institution": "7-Eleven", "country": "US", "position": "Researcher" }, { "name": "Guihong Wan", "gender": "Female", "institution": "Harvard University", "country": "US", "position": "Instructor" }, { "name": "Haim Schweitzer", "gender": "Unspecified", "institution": "University of Texas, Dallas", "country": "", "position": "Full Professor" }, { "name": "Wei Mao", "gender": "Male", "institution": "The University of Texas at Dallas", "country": "US", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 18, "n_ref": 25, "n_ref_all": 44, "n_fig": 1, "n_tab": 4, "L_tab": 2044, "n_element_tab": 220, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 482, "n_element_tab_1": 55, "formula_len_all": 951, "formula_len_all_1": 567, "len_all": 100759, "len_all_1": 63151, "len_abs": 1392, "len_title": 104, "len_sents": 28260, "len_sents_1": 24749, "n_sents": 365, "n_sents_1": 311, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1443, "title": "Very Fast Graph Clustering for Single and Multiple Views", "abs": "Clustering is a fundamental step in learning and analyzing graphs.\nCommonly accepted criteria for evaluating graph clustering quality \nwithout ground truth are \nthe \"normalized cut\" (ncut), \nand the \"ratio cut\" (rcut).\nTraditional algorithms that minimize\nncut and rcut \ntake $O(mnk)$ to cluster\na graph of $n$ nodes and $m$ edges into $k$ clusters.\nFaster algorithms sacrifice accuracy for speed\nand run in $O(m {+} n k^2)$.\nA very recent algorithm runs in $O(m {+} n k \\log k )$.\nThe space complexity of these algorithms ranges from $O(n^2)$\nto $O(n \\log k)$.\nWe describe a new algorithm with running time of $O(m \\log m)$\nthat achieves accuracy similar to traditional algorithms.\nOur algorithm is simple to implement, and requires only $O(m)$ memory.\nIt can also be applied in the multi-view setting, \nwhere multiple graphs share the same set of nodes.\nOur algorithm can cluster a small number of views\nwith no increase in its running time. \nWe describe \na randomized implementation that allows a qualitative comparison\nbetween various internal clustering criteria. \nOur experiments suggest a new criterion\nthat we call \"linfcut\" as superior to both\nthe ncut and the Cheeger criteria,\ncomputing clusters that \"make more sense\" to a human observer.\nOur algorithm performs a search for edges between clusters.\nIts speed is the result of a strong \"ignorance\" (pruning) condition\nthat allows ignoring most of the edges after little computation.", "keywords": [ "Graph Clustering" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 2, 2 ], "contribution_list": [ 2, 1, 2, 3, 1 ], "confidence_list": [ 3, 4, 4, 3, 5 ] }, { "paper_id": "oqRe1KvD17", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Peter Chin", "gender": "unknown", "institution": "Dartmouth College", "country": "US", "position": "Full Professor" }, { "name": "Thang Viet Nguyen", "gender": "Male", "institution": "Dartmouth College", "country": "", "position": "PhD student" }, { "name": "Yu-wing Tai", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 78, "n_ref": 143, "n_ref_all": 164, "n_fig": 3, "n_tab": 4, "L_tab": 163, "n_element_tab": 19, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 3188, "n_element_tab_1": 136, "formula_len_all": 410, "formula_len_all_1": 410, "len_all": 227634, "len_all_1": 67419, "len_abs": 1137, "len_title": 104, "len_sents": 41548, "len_sents_1": 29341, "n_sents": 324, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1140, "title": "Reward-RAG: Enhancing RAG with Reward Driven Supervision", "abs": "In this paper, we introduce Reward-RAG, a novel approach designed to enhance the Retrieval-Augmented Generation (RAG) model through Reward-Driven Supervision. Unlike previous RAG methodologies, which focus on training language models (LMs) to utilize external knowledge retrieved from external sources, our method adapts retrieval information to specific domains by employing CriticGPT to train a dedicated reward model. This reward model generates synthesized datasets for fine-tuning the RAG encoder, aligning its outputs more closely with human preferences. The versatility of our approach allows it to be effectively applied across various domains through domain-specific fine-tuning. We evaluate Reward-RAG on publicly available benchmarks from multiple domains, comparing it to state-of-the-art methods. Our experimental results demonstrate significant improvements in performance, highlighting the effectiveness of Reward-RAG in improving the relevance and quality of generated responses. These findings underscore the potential of integrating reward models with RAG to achieve superior outcomes in natural language generation tasks.", "keywords": [ "retrieval-augmented generation", "reward model", "CriticGPT", "reinforcement learning from AI feedback (RLAIF)", "reinforcement learning from human feedback (RLHF)", "LLMs" ], "rating_list": [ 3, 1, 5, 3 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 2, 4, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "opv67PpqLS", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Danqing Wang", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Jian Ma", "gender": "Male", "institution": "Carnegie-Mellon University", "country": "", "position": "" }, { "name": "Lei Li", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Muyu Yang", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Shike Wang", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Wenduo Cheng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Yang Zhang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Postdoc" }, { "name": "Zhenqiao Song", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 137, "n_ref_all": 165, "n_fig": 5, "n_tab": 17, "L_tab": 3585, "n_element_tab": 531, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1935, "n_element_tab_1": 264, "formula_len_all": 172, "formula_len_all_1": 0, "len_all": 203234, "len_all_1": 63191, "len_abs": 1332, "len_title": 115, "len_sents": 61415, "len_sents_1": 29858, "n_sents": 453, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1338, "title": "DNALONGBENCH: A Benchmark Suite For Long-Range DNA Prediction Tasks", "abs": "Modeling long-range DNA dependencies is crucial for understanding genome structure and function across a wide range of biological contexts in health and disease. However, effectively capturing the extensive long-range dependencies between DNA sequences, spanning millions of base pairs as seen in tasks such as three-dimensional (3D) chromatin folding, remains a significant challenge. Additionally, a comprehensive benchmark suite for evaluating tasks reliant on long-range dependencies is notably absent. To address this gap, we introduce DNALONGBENCH, a benchmark dataset spanning five important genomics tasks that consider long-range dependencies up to 1 million base pairs: enhancer-target gene interaction, expression quantitative trait loci, 3D genome organization, regulatory sequence activity, and transcription initiation signal. To comprehensively assess DNALONGBENCH, we evaluate the performance of five baseline methods: a task-specific expert model, a convolutional neural network (CNN)-based model, and three fine-tuned DNA foundation models -- HyenaDNA, Caduceus-Ph and Caduceus-PS. We envision DNALONGBENCH having the potential to become a standardized resource that facilitates comprehensive comparisons and rigorous evaluations of emerging DNA sequence-based deep learning models that consider long-range dependencies.", "keywords": [ "long-range DNA benchmark", "long-range DNA modeling", "long-range DNA foundation models" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "opSPgPIwAD", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Freddy Lecue", "gender": "unknown", "institution": "INRIA", "country": "", "position": "Full Professor" }, { "name": "Leonidas Tsepenekas", "gender": "Male", "institution": "J.P. Morgan Chase", "country": "", "position": "Researcher" }, { "name": "Margarita Boyarskaya", "gender": "unknown", "institution": "J.P. Morgan AI Research", "country": "US", "position": "Researcher" }, { "name": "Shubham Sharma", "gender": "Male", "institution": "J.P. Morgan Chase", "country": "US", "position": "Researcher" } ], "n_formula": 26, "n_formula_1": 18, "n_ref_uni": 30, "n_ref": 49, "n_ref_all": 65, "n_fig": 5, "n_tab": 3, "L_tab": 482, "n_element_tab": 63, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 245, "n_element_tab_1": 30, "formula_len_all": 2324, "formula_len_all_1": 1168, "len_all": 134176, "len_all_1": 72492, "len_abs": 1510, "len_title": 76, "len_sents": 47911, "len_sents_1": 34164, "n_sents": 398, "n_sents_1": 288, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1518, "title": "Everyone Deserves Recourse: Feasible Recourse Paths Using Data Augmentation", "abs": "Decisions made using machine learning models can negatively impact individuals\nin critical applications such as healthcare and finance by denying essential services\nor access to opportunity. Algorithmic recourse supplements a negative AI decision\nby providing rejected individuals with advice on the changes they can make to their\nprofiles, so that they may eventually achieve the desired outcome. Most existing\nrecourse methods provide single-step changes by using counterfactual explanations.\nThese counterfactual explanations are computed assuming a fixed (not learned)\ndistance function. Further, few works consider providing more realistic multi-step\nchanges in the form of recourse paths. However, such methods may fail to provide\nany recourse path for some individuals or provide paths that might not be feasible,\nsince intermediate steps needed to reach the counterfactual explanation may not\nbe realizable. We introduce a framework for learning an optimal distance function\nand threshold to compute multi-step recourse paths for all. First, we formalize the\nproblem of finding multi-step recourse paths. Given a set of feasible transitions, we\npropose a data-driven framework for learning the optimal distance and threshold\nfor each step with PAC (Probably Approximately Correct) guarantees. Finally,\nwe provide a data augmentation algorithm to ensure that a solution exists for all\nindividuals. Experiments on several datasets show that the proposed method learns\nfeasible recourse paths for all individuals.", "keywords": [ "Recourse", "Explainability" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 1, 1, 2 ], "presentation_list": [ 1, 2, 2 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "op1uHGKeux", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Simon Holk", "gender": "unknown", "institution": "The University of Tokyo, The University of Tokyo", "country": "JP", "position": "Intern" }, { "name": "Tatsuya Matsushima", "gender": "Male", "institution": "The University of Tokyo", "country": "", "position": "Researcher" }, { "name": "Yusuke Iwasawa", "gender": "Male", "institution": "The University of Tokyo, The University of Tokyo", "country": "JP", "position": "Associate Professor" }, { "name": "Yutaka Matsuo", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Odonchimed Sodtavilan", "gender": "Male", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "JP", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 14, "n_ref_uni": 30, "n_ref": 40, "n_ref_all": 60, "n_fig": 10, "n_tab": 9, "L_tab": 3369, "n_element_tab": 175, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 334, "n_element_tab_1": 66, "formula_len_all": 1010, "formula_len_all_1": 989, "len_all": 113470, "len_all_1": 48239, "len_abs": 1508, "len_title": 102, "len_sents": 31401, "len_sents_1": 20042, "n_sents": 265, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1517, "title": "RAGDP: Retrieve-Augmented Generative Diffusion Policy", "abs": "Diffusion Policy has attracted attention for its ability to achieve significant accuracy gains in a variety of imitation learning tasks. However, since Diffusion Policy relies on the Diffusion Model, it requires multiple denoising steps to generate a single action leading to long generation times. To address this issue, methods like DDIM and Consistency Models have been introduced to speed up the process. While these methods reduce computation time, this often comes at the cost of accuracy. In this paper, we propose RAGDP, a technique designed to improve the efficiency of learned Diffusion Policies without sacrificing accuracy. RAGDP builds upon the Retrieval-Augmented Generation (RAG) technique, which is commonly used in large language models to store and retrieve data from a vector database based on encoded embeddings. In RAGDP, pairs of expert observation and actions data are stored in a vector database. The system then searches the database using encoded observation data to retrieve expert action data with high similarity. This retrieved expert data is subsequently used by the RAGDP algorithm to generate actions tailored to the current environment. We introduce two action generation algorithms, RAGDP-VP and RAGDP-VE, which correspond to different types of Diffusion Models. Our results demonstrate that RAGDP can significantly improve the speed of Diffusion Policy without compromising accuracy. Furthermore, RAGDP can be integrated with existing speed-up methods to enhance their performance.", "keywords": [ "Imitation Learning", "Diffusion Models", "Behavior Cloning" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "ooxj2Audlq", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Di ZHANG", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "VP" }, { "name": "Lei Ke", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Mingqiao Ye", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Pengfei Wan", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Director" }, { "name": "Qi Fan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xin Tao", "gender": "Male", "institution": "Kuaishou", "country": "CN", "position": "Researcher" }, { "name": "Yu-wing Tai", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "Associate Professor" }, { "name": "ChiKeung Tang", "gender": "Not Specified", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 48, "n_ref": 90, "n_ref_all": 107, "n_fig": 4, "n_tab": 8, "L_tab": 2218, "n_element_tab": 224, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 165, "n_element_tab_1": 23, "formula_len_all": 314, "formula_len_all_1": 251, "len_all": 152786, "len_all_1": 64137, "len_abs": 1587, "len_title": 77, "len_sents": 37255, "len_sents_1": 30297, "n_sents": 271, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 29, "L_abs": 1584, "title": "Stable Segment Anything Model", "abs": "The Segment Anything Model (SAM) achieves remarkable promptable segmentation given high-quality prompts which, however, often require good skills to specify. To make SAM robust to casual prompts, this paper presents the first comprehensive analysis on SAM\u2019s segmentation stability across a diverse spectrum of prompt qualities, notably imprecise bounding boxes and insufficient points. Our key finding reveals that given such low-quality prompts, SAM\u2019s mask decoder tends to activate image features that are biased towards the background or confined to specific object parts. To mitigate this issue, our key idea consists of calibrating solely SAM\u2019s mask attention by adjusting the sampling locations and amplitudes of image features, while the original SAM model architecture and weights remain unchanged. Consequently, our deformable sampling plugin (DSP) enables SAM to adaptively shift attention to the prompted target regions in a data-driven manner. During inference, dynamic routing plugin (DRP) is proposed that toggles SAM between the deformable and regular grid sampling modes, conditioned on the input prompt quality. Thus, our solution, termed Stable-SAM, offers several advantages: 1) improved SAM\u2019s segmentation stability across a wide range of prompt qualities, while 2) retaining SAM\u2019s powerful promptable segmentation efficiency and generality, with 3) minimal learnable parameters (0.08 M) and fast adaptation. Extensive experiments validate the effectiveness and advantages of our approach, underscoring Stable-SAM as a more robust solution for segmenting anything.", "keywords": [ "Segment Anything Model", "Interactive Segmentation", "Segmentation Stability", "Deformable Feature Sampling" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "oos6KyAUsW", "primary_area": "causal reasoning", "n_author": 7, "author_info_list": [ { "name": "Anjie Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Haotian Wang", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Haoxuan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Jun Wang", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Professor" }, { "name": "Mengyue Yang", "gender": "Female", "institution": "University of Bristol", "country": "GB", "position": "Assistant Professor" }, { "name": "Xu Chen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zhenlei Wang", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 17, "n_ref_uni": 35, "n_ref": 77, "n_ref_all": 92, "n_fig": 5, "n_tab": 2, "L_tab": 544, "n_element_tab": 55, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1901, "formula_len_all_1": 1244, "len_all": 143567, "len_all_1": 65993, "len_abs": 1014, "len_title": 117, "len_sents": 46430, "len_sents_1": 31072, "n_sents": 349, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1019, "title": "Mitigating Unobserved Confounding via Diffusion Probabilistic Models", "abs": "Learning Conditional average treatment effect estimation from observational data is a challenging task due to the existence of unobserved confounders. Previous methods mostly focus on assuming the Ignorability assumption ignoring the unobserved confounders or overlooking the impact of an a priori knowledge on the generation process of the latent variable, which can be quite impractical in real-world scenarios. Motivated by the recent advances in the latent variable modeling, we propose to capture the unobserved latent space using diffusion model, and accordingly to estimate the causal effect. More concretely, we build on the reverse diffusion process for the unobserved confounders as a Markov chain conditioned on an apriori knowledge. In order to implement our model in a feasible way, we derive the variational bound in closed form. In the experiments, we compare our model with the state-of-the-art methods based on both synthetic and real-world datasets, demonstrating consistent improvements of our model.", "keywords": [ "causal inference" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "onvN3zsNMI", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Byungjoo Kim", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Sung Ju Hwang", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 40, "n_ref": 79, "n_ref_all": 88, "n_fig": 1, "n_tab": 8, "L_tab": 1357, "n_element_tab": 92, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 2226, "n_element_tab_1": 124, "formula_len_all": 764, "formula_len_all_1": 798, "len_all": 111837, "len_all_1": 61023, "len_abs": 1102, "len_title": 118, "len_sents": 30230, "len_sents_1": 27480, "n_sents": 217, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1106, "title": "Preventing Unintended Memorization by Covering with Over-Memorization", "abs": "From the advances of deep learning, the privacy concerns of deep neural networks are in the limelight. A particular concern is privacy of the training data, which is often compromised by the model's inherent memorization capabilities. Suppressing such memorization can enhance privacy but introduces two main challenges: 1) removing a memorized instance from the training dataset will result in the model to memorize another instance instead, and 2) the memorization is essential for improving the generalization error. To address these challenges, we propose an over-memorization method that involves training the model with both the standard training set and a set of redundant, non-sensitive instances. Our method leverages the model's limited memorization capacity to focus on irrelevant data, thereby preventing it from memorizing the training data. Our empirical results demonstrate that this method not only enhances protection against membership inference attacks but also minimizes the loss of utility by effectively redirecting the model's generalization efforts towards non-sensitive instances.", "keywords": [ "memorization", "deep learning", "data privacy", "machine learning" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "onrNYdciJQ", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "alain rakoto", "gender": "Unspecified", "institution": "Criteo", "country": "FR", "position": "Researcher" }, { "name": "Chansoo Kim", "gender": "unknown", "institution": "Korea Institute of Science and Technology", "country": "KR", "position": "Senior Research Scientist" }, { "name": "Jean-Yves Franceschi", "gender": "Male", "institution": "Criteo", "country": "FR", "position": "Researcher" }, { "name": "Sangchul Lee", "gender": "Male", "institution": "Korea Institute of Science and Technology", "country": "KR", "position": "Postdoc" }, { "name": "Thibaut Issenhuth", "gender": "Male", "institution": "Criteo", "country": "FR", "position": "Researcher" }, { "name": "Ludovic Dos Santos", "gender": "Male", "institution": "Criteo AI Lab", "country": "", "position": "Researcher" } ], "n_formula": 107, "n_formula_1": 33, "n_ref_uni": 33, "n_ref": 68, "n_ref_all": 104, "n_fig": 9, "n_tab": 4, "L_tab": 103, "n_element_tab": 13, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 670, "n_element_tab_1": 6, "formula_len_all": 7725, "formula_len_all_1": 1725, "len_all": 167165, "len_all_1": 68987, "len_abs": 1206, "len_title": 105, "len_sents": 48736, "len_sents_1": 28018, "n_sents": 457, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1213, "title": "Improving Consistency Models with Generator-Induced Flows", "abs": "Consistency models imitate the multi-step sampling of score-based diffusion in a single forward pass of a neural network.\nThey can be learned in two ways: consistency distillation and consistency training. The former relies on the true velocity field of the corresponding differential equation, approximated by a pre-trained neural network.\nIn contrast, the latter uses a single-sample Monte Carlo estimate of this velocity field. The related estimation error induces a discrepancy between consistency distillation and training that, we show, still holds in the continuous-time limit. To alleviate this issue, we propose a novel flow that transports noisy data towards their corresponding outputs derived from the currently trained model - as a proxy of the true flow. Our empirical findings demonstrate that this approach mitigates the previously identified discrepancy.\nFurthermore, we present theoretical and empirical evidence indicating that our generator-induced flow surpasses dedicated optimal transport-based consistency models in effectively reducing the noise-data transport cost.\nConsequently, our method not only accelerates consistency training convergence but also enhances its overall performance.", "keywords": [ "generative models", "diffusion models", "flow models", "consistency models" ], "rating_list": [ 8, 3, 5, 8 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 2, 4 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "onIro14tHv", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Arno Solin", "gender": "unknown", "institution": "Aalto University", "country": "FI", "position": "Assistant Professor" }, { "name": "Markus Heinonen", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Researcher" }, { "name": "Najwa Laabid", "gender": "Female", "institution": "Aalto University", "country": "", "position": "PhD student" }, { "name": "Severi Rissanen", "gender": "Male", "institution": "Aalto University", "country": "", "position": "PhD student" }, { "name": "Vikas K Garg", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Assistant Professor" } ], "n_formula": 118, "n_formula_1": 28, "n_ref_uni": 51, "n_ref": 123, "n_ref_all": 185, "n_fig": 14, "n_tab": 5, "L_tab": 421, "n_element_tab": 32, "n_fig_1": 10, "n_tab_1": 5, "L_tab_1": 1309, "n_element_tab_1": 110, "formula_len_all": 11782, "formula_len_all_1": 2442, "len_all": 266664, "len_all_1": 71836, "len_abs": 721, "len_title": 123, "len_sents": 79556, "len_sents_1": 29475, "n_sents": 680, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 732, "title": "Equivariant Denoisers Cannot Copy Graphs: Align Your Graph Diffusion Models", "abs": "Graph diffusion models, while dominant in graph generative modeling, remain relatively underexplored for graph-to-graph translation tasks like chemical reaction prediction. We show that standard permutation equivariant denoisers cause severe limitations on such tasks, a problem that we pinpoint to their inability to break symmetries present in the noisy inputs. We then propose to \\emph{align} the input and target graphs in order to break the input symmetries, while retaining permutation equivariance in the non-matching portions of the graph. We choose retrosynthesis as an application domain, and show how alignment takes the performance of a discrete diffusion model from a mere 5\\% to a SOTA-matching 54.7\\% top-1 accuracy.", "keywords": [ "Graph Diffusion", "Discrete Diffusion", "Equivariance", "Symmetries" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "omzijInU1T", "primary_area": "learning theory", "n_author": 1, "author_info_list": [ { "name": "Baiyuan Chen", "gender": "unknown", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "JP", "position": "Undergrad student" } ], "n_formula": 29, "n_formula_1": 27, "n_ref_uni": 27, "n_ref": 38, "n_ref_all": 55, "n_fig": 8, "n_tab": 1, "L_tab": 1109, "n_element_tab": 38, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2851, "formula_len_all_1": 2464, "len_all": 98590, "len_all_1": 58354, "len_abs": 989, "len_title": 135, "len_sents": 29785, "len_sents_1": 24936, "n_sents": 257, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1002, "title": "Feature Learning in Attention Mechanisms Is More Compact and Stable Than in Convolution", "abs": "Robustness is a crucial attribute of machine learning models, A robust model ensures consistent performance under input corruptions, adversarial attacks, and out-of-distribution data. While the Wasserstein distance is widely used for assessing robustness by quantifying geometric discrepancies between distributions, its application to layer-wise analysis is limited since computing the Wasserstein distance usually involves dimensionality reduction, which is not suitable for models like CNNs that have layers with diverse output dimensions. To address this, we propose $\\textit{TopoLip}$, a novel metric that facilitates layer-wise robustness analysis. TopoLip enables theoretical and empirical evaluation of robustness, providing insights into how model parameters influence performance. By comparing Transformers and ResNets, we demonstrate that Transformers are more robust in both theoretical settings and experimental evaluations, particularly in handling corrupted and out-of-distribution data.", "keywords": [ "Feature Learning", "Attention", "Convolution", "Transformer", "ResNet", "Lipschitz Continuity", "Wasserstein Distance", "Topological Data Analysis" ], "rating_list": [ 3, 5, 3, 1, 6 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 1, 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 1, 3 ], "confidence_list": [ 4, 3, 3, 4, 2 ] }, { "paper_id": "omv3VfVIQt", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Jiakai Wang", "gender": "Male", "institution": "Zhongguancun Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Shan He", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Siqi He", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xianglong Liu", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Yuqing Ma", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "Associate Professor" }, { "name": "Zhange Zhang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" }, { "name": "tianbo Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 24, "n_ref": 54, "n_ref_all": 76, "n_fig": 6, "n_tab": 17, "L_tab": 6290, "n_element_tab": 190, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1019, "n_element_tab_1": 23, "formula_len_all": 471, "formula_len_all_1": 426, "len_all": 169358, "len_all_1": 57792, "len_abs": 1528, "len_title": 128, "len_sents": 61337, "len_sents_1": 28221, "n_sents": 444, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1535, "title": "Lexical Diversity-aware Relevance Assessment for Retrieval-Augmented Generation", "abs": "Despite their extensive applications, large language models trained on vast historical datasets still struggle with hallucination issues, particularly when addressing open-ended, factual, and commonsense questions. In contrast, Retrieval-Augmented Generation (RAG) methods have proven effective in enhancing large language models' responses to such inquiries, making them a focal point of research.\nHowever, previous RAG approaches overlook the lexical diversity of queries, hindering their ability to achieve a granular relevance assessment between queries and retrieved documents, resulting in suboptimal performance. In this paper, we introduce a Lexical Diversity-aware RAG (DRAG) model, comprising a Diversity-sensitive Relevance Analyzer (DRA) and a Contrastive Relevance Calibration Module (CRC). Specifically, DRA decouples and assesses the relevance of different query components (words, phrases) based on their levels of lexical diversity, ensuring precise and comprehensive document retrieval. According to the DRA assessment, CRC further emphasizes the pertinent knowledge of the retrieved relevant documents through contrastively eliminating the adverse effects of irrelevant contents. By integrating DRA and CRC, the proposed method effectively retrieves relevant documents and leverages their pertinent knowledge to refine the original results and generate meaningful outcomes. Extensive experiments on widely-used benchmarks demonstrate the efficacy of our approach, yielding a 12.5\\% accuracy improvement on HotpotQA.", "keywords": [ "language model", "Retrieval-Augmented Generation", "Lexical Diversity", "Relevance Assessment" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "omrLHFzC37", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Bicheng Ying", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Chaosheng Dong", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Haibo Yang", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Zhe Li", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zidong Liu", "gender": "Male", "institution": "Combocurve Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 113, "n_formula_1": 22, "n_ref_uni": 64, "n_ref": 108, "n_ref_all": 134, "n_fig": 6, "n_tab": 8, "L_tab": 5309, "n_element_tab": 286, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2019, "n_element_tab_1": 113, "formula_len_all": 14927, "formula_len_all_1": 2354, "len_all": 248742, "len_all_1": 69149, "len_abs": 1806, "len_title": 139, "len_sents": 64465, "len_sents_1": 29539, "n_sents": 559, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1706, "title": "Achieving Dimension-Free Communication in Federated Learning via Zeroth-Order Optimization", "abs": "Federated Learning (FL) offers a promising framework for collaborative and privacy-preserving machine learning across distributed data sources. \nHowever, the substantial communication costs associated with FL significantly challenge its efficiency. \nSpecifically, in each communication round, the communication costs scale linearly with the model's dimension, which presents a formidable obstacle, especially in large model scenarios. \nDespite various communication-efficient strategies, the intrinsic dimension-dependent communication cost remains a major bottleneck for current FL implementations.\nThis paper proposes a novel dimension-free communication algorithm - DeComFL, which leverages the zeroth-order optimization techniques and reduces the communication cost from $\\mathcal{O}(d)$ to $\\mathcal{O}(1)$ by transmitting only a constant number of scalar values between clients and the server in each round, regardless of the dimension $d$ of the model parameters.\nTheoretically, in non-convex functions, we prove that our algorithm achieves state-of-the-art rates, which show a linear speedup of the number of clients and local steps under standard assumptions. With additional low effective rank assumption, we can further show the convergence rate is independent of the model dimension $d$ as well.\nEmpirical evaluations, encompassing both classic deep learning training and large language model fine-tuning, demonstrate significant reductions in communication overhead. \nNotably, DeComFL achieves this by transmitting only around 1MB of data in total between the server and a client to fine-tune a model with billions of parameters. \nThe code is available at https://github.com/ZidongLiu/DeComFL.", "keywords": [ "Federated Learning", "Zeroth-order Optimization", "Communication Efficiency", "Low Rank" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "omM5m7mRy5", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Ayan Banerjee", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "Assistant Professor" }, { "name": "BIN XU", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Payal Kamboj", "gender": "Female", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Sandeep Gupta", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 15, "n_ref": 55, "n_ref_all": 64, "n_fig": 2, "n_tab": 4, "L_tab": 2854, "n_element_tab": 188, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1551, "n_element_tab_1": 125, "formula_len_all": 317, "formula_len_all_1": 317, "len_all": 109028, "len_all_1": 64966, "len_abs": 1559, "len_title": 120, "len_sents": 35547, "len_sents_1": 31246, "n_sents": 236, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1563, "title": "Single Domain Generalization for Rare Event Detection in Medical Imaging", "abs": "Single Domain Generalization (SDG) addresses the challenge of training a model on a single domain to ensure it generalizes well to unseen target domains. Although extensively studied in image classification, there is a lack of prior work on SDG for rare event or image classification in imbalanced dataset. In the medical diagnosis and disease detection domain, where data is often limited and events of interest are rare, deep learning (DL) models frequently exhibit suboptimal performance, leading to poor generalization across datasets. In multi-center studies, disparate data sources, differences in scanners and imaging protocols introduce domain shifts that exacerbate variability in rare event characteristics. This paper addresses this challenge by first leveraging a pre-trained large vision model to rank classes based on their similarity to the rare event class, allowing focused handling of the most similar class, and then integrates domain-invariant knowledge on rare event with DL to accurately classify the rare event class. By carefully incorporating expert knowledge with data-driven DL, our technique effectively regularizes the model, enhancing robustness and performance even with limited data availability. We present a case study on seizure onset zone detection using fMRI data, demonstrating that our approach significantly outperforms state-of-the-art vision transformers, large vision models, and knowledge-based systems, achieving an average F1 score of 90.2% while maintaining an overall F1 score of 85.0% across multi-center datasets.", "keywords": [ "Deep Learning", "Knowledge", "Rare Event Detection", "Out-of-distribution detection" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "okRSNTMdFg", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Chao Du", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Hongcheng Gao", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Min Lin", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Principal Researcher" }, { "name": "Tianyu Pang", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Zhijie Deng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "taihang Hu", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "MS student" } ], "n_formula": 22, "n_formula_1": 21, "n_ref_uni": 49, "n_ref": 114, "n_ref_all": 139, "n_fig": 8, "n_tab": 5, "L_tab": 1742, "n_element_tab": 114, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1143, "n_element_tab_1": 48, "formula_len_all": 1720, "formula_len_all_1": 1593, "len_all": 143940, "len_all_1": 59451, "len_abs": 1249, "len_title": 125, "len_sents": 32709, "len_sents_1": 25007, "n_sents": 241, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1262, "title": "Meta-Unlearning on Diffusion Models: Preventing Relearning Unlearned Concepts", "abs": "With the rapid progress of diffusion-based content generation, significant efforts are being made to unlearn harmful or copyrighted concepts from pretrained diffusion models (DMs) to prevent potential model misuse. However, it is observed that even when DMs are properly unlearned before release, malicious finetuning can compromise this process, causing DMs to *relearn the unlearned concepts*. This occurs partly because certain benign concepts (e.g., \"skin\") retained in DMs are related to the unlearned ones (e.g., \"nudity\"), facilitating their relearning via finetuning. To address this, we propose **meta-unlearning** on DMs. Intuitively, a meta-unlearned DM should behave like an unlearned DM when used as is; moreover, if the meta-unlearned DM undergoes malicious finetuning on unlearned concepts, the related benign concepts retained within it will be triggered to *self-destruct*, hindering the relearning of unlearned concepts. Our meta-unlearning framework is compatible with most existing unlearning methods, requiring only the addition of an easy-to-implement meta objective. We validate our approach through empirical experiments on meta-unlearning concepts from Stable Diffusion models (SD-v1-4 and SDXL), supported by extensive ablation studies.", "keywords": [ "Meta-Unlearning", "Diffusion Models" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 4, 2, 3 ], "presentation_list": [ 3, 2, 1 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "okEwtOc5Go", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Chengyao Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jiaya Jia", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Ruihang Chu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shaoteng Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yanwei Li", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Yixin Chen", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Zhisheng Zhong", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yuechen ZHANG", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 38, "n_ref": 84, "n_ref_all": 104, "n_fig": 9, "n_tab": 8, "L_tab": 2219, "n_element_tab": 217, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1520, "n_element_tab_1": 211, "formula_len_all": 315, "formula_len_all_1": 203, "len_all": 206381, "len_all_1": 82216, "len_abs": 1611, "len_title": 122, "len_sents": 55571, "len_sents_1": 35122, "n_sents": 421, "n_sents_1": 286, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1278, "title": "Mini-Gemini: Mining the Potential of Multi-modality Vision Language Models", "abs": "In this work, we introduce Mini-Gemini, a simple and effective framework enhancing multi-modality Vision Language Models (VLMs). Despite the advancements in VLMs facilitating basic visual dialog and reasoning, a performance gap persists compared to advanced models like GPT-4 and Gemini. We try to narrow the gap by mining the potential of VLMs for better performance across various cross-modal tasks %and any-to-any workflow from three aspects, i.e., high-resolution visual tokens, high-quality data, and VLM-guided generation. To enhance visual tokens, we propose to utilize an additional visual encoder for high-resolution refinement without increasing the visual token count. We further construct a high-quality dataset that promotes precise image comprehension and reasoning-based generation, expanding the operational scope of current VLMs. In general, Mini-Gemini further mines the potential of VLMs and empowers current frameworks with image understanding, reasoning, and generation simultaneously. Mini-Gemini supports a series of dense and MoE Large Language Models (LLMs) from 2B to 34B. It is demonstrated to achieve leading performance in several zero-shot benchmarks and even surpasses the developed private models. Code and models will be available to the public.", "keywords": [ "Vision Language Model", "Image Generation", "Multi-modality Model" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "okD9dbifxa", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Ben Mildenhall", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Danhang Tang", "gender": "Male", "institution": "Google Inc.", "country": "", "position": "Staff Research Scientist" }, { "name": "Le Chen", "gender": "Male", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "", "position": "Researcher" }, { "name": "Qiangeng Xu", "gender": "Male", "institution": "Waymo", "country": "US", "position": "Researcher" }, { "name": "Quankai Gao", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Ulrich Neumann", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Full Professor" }, { "name": "Wenchao Ma", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Zhe Cao", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 18, "n_ref_uni": 49, "n_ref": 127, "n_ref_all": 149, "n_fig": 10, "n_tab": 4, "L_tab": 1246, "n_element_tab": 223, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 535, "n_element_tab_1": 164, "formula_len_all": 2217, "formula_len_all_1": 1001, "len_all": 173648, "len_all_1": 59583, "len_abs": 1189, "len_title": 113, "len_sents": 37973, "len_sents_1": 26412, "n_sents": 280, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1195, "title": "GAUSSIANFLOW: SPLATTING GAUSSIAN DYNAMICS FOR 4D CONTENT CREATION", "abs": "Creating 4D fields of Gaussian Splatting from images or videos is a challenging task due to its under-constrained nature. While the optimization can draw photometric reference from the input videos or be regulated by generative models, directly supervising Gaussian motions remains underexplored. In this paper, we introduce a novel concept, Gaussian flow, which connects the dynamics of 3D Gaussians and pixel velocities between consecutive frames. The Gaussian flow can be efficiently obtained by splatting Gaussian dynamics into the image space. This differentiable process enables direct dynamic supervision from optical flow. Our method significantly benefits 4D dynamic content generation and 4D novel view synthesis with Gaussian Splatting, especially for contents with rich motions that are hard to be handled by existing methods. The common color drifting issue that happens in 4D generation is also resolved with improved Guassian dynamics. Superior visual quality on extensive experiments demonstrates our method\u2019s effectiveness. As shown in our evaluation, Gaussian Flow can drastically improve both quantitative and qualitative results for 4D Generation and 4D novel view synthesis.", "keywords": [ "optical flow", "neural rendering", "Gaussian Splatting" ], "rating_list": [ 5, 6, 8, 5, 6 ], "soundness_list": [ 2, 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4, 5 ] }, { "paper_id": "ok5NweADUB", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Ameya Daigavane", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Mit Kotak", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Tess Smidt", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "YuQing Xie", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 98, "n_formula_1": 22, "n_ref_uni": 25, "n_ref": 46, "n_ref_all": 61, "n_fig": 5, "n_tab": 3, "L_tab": 383, "n_element_tab": 46, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 94, "n_element_tab_1": 10, "formula_len_all": 8569, "formula_len_all_1": 2143, "len_all": 193999, "len_all_1": 56961, "len_abs": 1202, "len_title": 143, "len_sents": 60057, "len_sents_1": 23855, "n_sents": 650, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1280, "title": "The Price of Freedom: Exploring Tradeoffs in Equivariant Tensor Products with Spherical Signals", "abs": "$E(3)$-equivariant neural networks have demonstrated success across a wide range of 3D modelling tasks. A fundamental operation in these networks is the tensor product, which interacts two geometric features in an equivariant manner to create new features. Due to the high computational complexity of the tensor product, significant effort has been invested to optimize the runtime of this operation. \\citet{gaunt} recently proposed the Gaunt tensor product (GTP) which promises a significant speedup over the naive implementation of the tensor product. However, this method is unable to perform antisymmetric operations which are crucial for tasks involving chirality. In this work, we introduce vector signal tensor product (VSTP) to solve this issue and show how it generalizes to a class of irrep signal tensor products (ISTPs). Finally, we investigate why these tensor products are faster. We find most of the speedup comes at the price of expressivity. Further, we microbenchmarked the various tensor products and find that the theoretical runtime guarantees may differ wildly from empirical performance, demonstrating the need for careful application-specific benchmarking. Our code is linked \\href{https://anonymous.4open.science/r/vector-spherical-harmonics-1231/}{here}.", "keywords": [ "equivariance", "tensor product", "spherical harmonics", "vector spherical harmonics", "spherical signals", "benchmarking", "asymptotics" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "oiDvwOhvjq", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Amin Ullah", "gender": "Male", "institution": "Boeing Research and Technology Software ", "country": "", "position": "Researcher" }, { "name": "Jun Li", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Assistant Professor" }, { "name": "Shuai Li", "gender": "Male", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "PhD student" }, { "name": "Yang Zitao", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Undergrad student" }, { "name": "Fuxin Li", "gender": "Male", "institution": "Oregon State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 32, "n_formula_1": 14, "n_ref_uni": 47, "n_ref": 85, "n_ref_all": 100, "n_fig": 5, "n_tab": 5, "L_tab": 173, "n_element_tab": 11, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1418, "n_element_tab_1": 69, "formula_len_all": 3590, "formula_len_all_1": 1101, "len_all": 138567, "len_all_1": 60086, "len_abs": 863, "len_title": 137, "len_sents": 40017, "len_sents_1": 26269, "n_sents": 296, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 865, "title": "Convex Potential Mirror Langevin Algorithm for Efficient Sampling of Energy-Based Models", "abs": "This paper introduces the Convex Potential Mirror Langevin Algorithm (CPMLA), a novel method designed to optimize sampling efficiency within Energy-Based Models (EBMs). CPMLA employs mirror Langevin dynamics in conjunction with convex potential flow as a dynamic mirror map for sampling in EBMs. By leveraging this dynamic mirror map, CPMLA enables targeted geometric exploration on the data manifold, enhancing the convergence process towards the target distribution. Theoretical analysis proves that CPMLA achieves exponential convergence with vanishing bias under relaxed log-concave conditions, supporting its efficiency and effectiveness in adapting to complex data distributions. Experimental results on established benchmarks like CIFAR-10, SVHN, and CelebA showcase CPMLA's enhanced sampling quality and inference efficiency compared to existing techniques.", "keywords": [ "mirror Langevin dynamics", "efficient sampling", "energy based models" ], "rating_list": [ 3, 3, 8, 8 ], "soundness_list": [ 1, 3, 3, 4 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 4 ], "confidence_list": [ 3, 3, 5, 3 ] }, { "paper_id": "ohqjYsRBD1", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Xiang Ren", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Associate Professor" }, { "name": "Xisen Jin", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 52, "n_ref": 81, "n_ref_all": 111, "n_fig": 10, "n_tab": 6, "L_tab": 2435, "n_element_tab": 365, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 667, "n_element_tab_1": 104, "formula_len_all": 275, "formula_len_all_1": 84, "len_all": 191670, "len_all_1": 69250, "len_abs": 1638, "len_title": 121, "len_sents": 43787, "len_sents_1": 32725, "n_sents": 337, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1483, "title": "Demystifying Language Model Forgetting with Low-Rank Example Associations", "abs": "Large Language models (LLMs) suffer from forgetting of upstream data when fine-tuned. Despite efforts on mitigating forgetting, few have investigated whether, and how forgotten upstream examples are dependent on and associated with newly learned tasks. Insights on such associations enable efficient and targeted mitigation of forgetting. In this paper, we empirically analyze forgetting (measured in log-perplexity increase) that occurs in $N$ upstream examples of language modeling or instruction-tuning after fine-tuning LLMs on one of $M$ new tasks, visualized in $M\\times N$ matrices. We demonstrate that the matrices display simple low-rank patterns, often well-approximated with multiplicative scalar effects of upstream examples and newly learned tasks. We also examine fine-grained associations with visualization and statistics. Leveraging the low-rank nature of the associations, we predict forgetting of upstream examples when fine-tuning on unseen tasks with matrix completion over the empirical associations. This enables fast identification of most forgotten examples without expensive inference on the entire upstream data. The approach, despite simplicity, outperforms prior approaches that learn semantic relationships of learned tasks and upstream examples with LMs for predicting forgetting. We demonstrate the practical utility of our analysis by showing statistically significantly reduced forgetting as we upweight predicted examples for replay at fine-tuning.", "keywords": [ "catastrophic forgetting; language models" ], "rating_list": [ 3, 3, 5, 6, 3 ], "soundness_list": [ 2, 1, 4, 2, 2 ], "presentation_list": [ 4, 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 1 ], "confidence_list": [ 3, 4, 3, 3, 3 ] }, { "paper_id": "ohJxgRLlLt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Amir Zamir", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Assistant Professor" }, { "name": "Andrei Atanov", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Artyom Gadetsky", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Ghazal Hosseini Mighan", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Intern" }, { "name": "Maria Brbic", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "", "position": "Assistant Professor" }, { "name": "Yulun Jiang", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Zhitong Gao", "gender": "Female", "institution": "EPFL", "country": "CH", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 13, "n_ref_uni": 62, "n_ref": 95, "n_ref_all": 126, "n_fig": 5, "n_tab": 2, "L_tab": 761, "n_element_tab": 96, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 768, "n_element_tab_1": 96, "formula_len_all": 1315, "formula_len_all_1": 844, "len_all": 186515, "len_all_1": 66594, "len_abs": 1370, "len_title": 116, "len_sents": 42253, "len_sents_1": 30196, "n_sents": 368, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1376, "title": "Large (Vision) Language Models are Unsupervised In-Context Learners", "abs": "Recent advancements in large language and vision-language models have made it possible to solve new tasks via zero-shot inference without task-specific training. Various adaptation techniques, such as In-Context Learning (ICL), supervised fine-tuning, and prompt engineering, can further enhance the model\u2019s performance on a given task. However, these methods require either labeled examples or substantial manual effort to construct effective prompts. In this work, we introduce a joint inference framework extending the standard zero-shot inference. In contrast to independent zero-shot predictions, joint inference makes predictions simultaneously for all inputs for a given task. Since direct joint inference involves a computationally expensive optimization, we develop efficient approximation techniques resulting in two unsupervised adaptation methods that are compatible with language and vision-language models: unsupervised fine-tuning and unsupervised ICL. We demonstrate the effectiveness of both approaches across a broad range of tasks and models, including language-only Llama 3.1, vision-language Open-Flamingo and API-only access GPT-4o models. Our experiments reveal substantial improvements over the standard zero-shot approach. Furthermore, our approach, although unsupervised, often performs on par with supervised approaches that use ground truth labels.", "keywords": [ "llm", "unsupervised", "in-context learning", "few-shot learning" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "ohHtdp3jDi", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Jie Lu", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "Professor" }, { "name": "Junyu Xuan", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Assistant Professor" }, { "name": "Mengjing Wu", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" } ], "n_formula": 40, "n_formula_1": 38, "n_ref_uni": 45, "n_ref": 73, "n_ref_all": 89, "n_fig": 4, "n_tab": 3, "L_tab": 1831, "n_element_tab": 96, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1152, "n_element_tab_1": 69, "formula_len_all": 2892, "formula_len_all_1": 2959, "len_all": 119170, "len_all_1": 60876, "len_abs": 1207, "len_title": 90, "len_sents": 33988, "len_sents_1": 26679, "n_sents": 247, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1211, "title": "Implicit Functional Bayesian Deep Learning", "abs": "Bayesian deep learning (BDL) is believed to be an effective approach to enabling uncertainty estimation and improving the generalisation and robustness of classical deep learning with the help of the Bayesian principle. Considering its non-meaningful weight-space prior and problematic Kullback-Leibler (KL) divergence, functional inference with Wasserstein distance has recently emerged as a promising direction in this field. However, existing efforts require different types of degenerations to achieve tractable Wasserstein distance computation, which limits the predictive and uncertainty estimation capabilities. In this paper, we propose two novel implicit functional BDL (ifBDL) approaches, i.e., implicit functional Bayesian neural networks and implicit functional Bayesian deep ensemble. The common idea is to implicitly transform the BDL posterior to a Gaussian process via the neural tangent kernel to facilitate tractable 2-Wasserstein distance computation and preserve the neural network parameterization. The experimental evaluations on standard tasks show that ifBDL has superior predictive and uncertainty estimation capabilities compared to existing weight-space and function-space approaches.", "keywords": [ "Bayesian Deep Learning", "function space", "variational inference" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 2, 2, 4, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "ogmzNfeRl7", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Nasir Ahmad", "gender": "Male", "institution": "Donders Institute for Brain, Cognition and Behaviour", "country": "NL", "position": "Assistant Professor" } ], "n_formula": 29, "n_formula_1": 18, "n_ref_uni": 31, "n_ref": 72, "n_ref_all": 88, "n_fig": 4, "n_tab": 2, "L_tab": 463, "n_element_tab": 75, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2110, "formula_len_all_1": 877, "len_all": 117543, "len_all_1": 56773, "len_abs": 1574, "len_title": 47, "len_sents": 43164, "len_sents_1": 28032, "n_sents": 268, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1578, "title": "Correlations Are Ruining Your Gradient Descent", "abs": "Herein the topics of (natural) gradient descent, data decorrelation, and approximate methods for backpropagation are brought into a common discussion. Natural gradient descent illuminates how gradient vectors, pointing at directions of steepest descent, can be improved by considering the local curvature of loss landscapes. We extend this perspective and show that to fully solve the problem illuminated by natural gradients in neural networks, one must recognise that correlations in the data at any linear transformation, including node responses at every layer of a neural network, cause a non-orthonormal relationship between the model's parameters. To solve this requires a method for decorrelating inputs at each individual layer of a neural network. We describe a range of methods which have been proposed for decorrelation and whitening of node output, and expand on these to provide a novel method specifically useful for distributed computing and computational neuroscience. Implementing decorrelation within multi-layer neural networks, we can show that not only is training via backpropagation sped up significantly but also existing approximations of backpropagation, which have failed catastrophically in the past, benefit significantly in their accuracy and convergence speed. This has the potential to provide a route forward for approximate gradient descent methods which have previously been discarded, training approaches for analogue and neuromorphic hardware, and potentially insights as to the efficacy and utility of decorrelation processes in the brain.", "keywords": [ "Natural gradient descent", "decorrelation", "whitening", "approximate gradient descent" ], "rating_list": [ 5, 3, 8 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 3, 2, 4 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "ogjBpZ8uSi", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Bilel Omrani", "gender": "Male", "institution": "Illuin Technology", "country": "FR", "position": "Researcher" }, { "name": "C\u00e9line Hudelot", "gender": "Female", "institution": "CentraleSupelec", "country": "", "position": "Full Professor" }, { "name": "Gautier Viaud", "gender": "Male", "institution": "ILLUIN Technology", "country": "", "position": "Principal Researcher" }, { "name": "Hugues Sibille", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Manuel Faysse", "gender": "Male", "institution": "CentraleSupelec", "country": "", "position": "PhD student" }, { "name": "Pierre Colombo", "gender": "Male", "institution": "CentraleSupelec", "country": "", "position": "Assistant Professor" }, { "name": "Tony Wu", "gender": "Male", "institution": "ILLUIN Technology", "country": "FR", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 53, "n_ref": 83, "n_ref_all": 117, "n_fig": 3, "n_tab": 6, "L_tab": 2846, "n_element_tab": 342, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 564, "n_element_tab_1": 62, "formula_len_all": 362, "formula_len_all_1": 128, "len_all": 204942, "len_all_1": 65133, "len_abs": 1297, "len_title": 65, "len_sents": 52130, "len_sents_1": 31062, "n_sents": 399, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1332, "title": "ColPali: Efficient Document Retrieval with Vision Language Models", "abs": "Documents are visually rich structures that convey information through text, but also figures, page layouts, tables, or even fonts. Since modern retrieval systems mainly rely on the textual information they extract from document pages to index documents -often through lengthy and brittle processes-, they struggle to exploit key visual cues efficiently. This limits their capabilities in many practical document retrieval applications such as Retrieval Augmented Generation (RAG).\nTo benchmark current systems on visually rich document retrieval, we introduce the Visual Document Retrieval Benchmark \\textit{ViDoRe}, composed of various page-level retrieval tasks spanning multiple domains, languages, and practical settings. \nThe inherent complexity and performance shortcomings of modern systems motivate a new concept; doing document retrieval by directly embedding the images of the document pages. We release \\textit{ColPali}, a Vision Language Model trained to produce high-quality multi-vector embeddings from images of document pages. Combined with a late interaction matching mechanism, \\textit{ColPali} largely outperforms modern document retrieval pipelines while being drastically simpler, faster and end-to-end trainable. \nWe release models, data, code and benchmarks under open licenses at \\url{https://hf.co/vidore}.", "keywords": [ "document embeddings", "vision language models", "late interaction", "document retrieval", "information retrieval" ], "rating_list": [ 3, 5, 8, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 2, 4 ] }, { "paper_id": "ogXkmugNZw", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Dohyeong Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Jeongho Park", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "PhD student" }, { "name": "Mineui Hong", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "PhD student" }, { "name": "Songhwai Oh", "gender": "unknown", "institution": "Seoul National University", "country": "", "position": "Full Professor" } ], "n_formula": 91, "n_formula_1": 27, "n_ref_uni": 28, "n_ref": 96, "n_ref_all": 151, "n_fig": 18, "n_tab": 5, "L_tab": 920, "n_element_tab": 80, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10653, "formula_len_all_1": 2381, "len_all": 197226, "len_all_1": 66608, "len_abs": 1202, "len_title": 139, "len_sents": 63410, "len_sents_1": 28737, "n_sents": 602, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1207, "title": "Conflict-Averse Gradient Aggregation for Constrained Multi-Objective Reinforcement Learning", "abs": "In real-world applications, a reinforcement learning (RL) agent should consider multiple objectives and adhere to safety guidelines.\nTo address these considerations, we propose a constrained multi-objective RL algorithm named constrained multi-objective gradient aggregator (CoMOGA).\nIn the field of multi-objective optimization, managing conflicts between the gradients of the multiple objectives is crucial to prevent policies from converging to local optima.\nIt is also essential to efficiently handle safety constraints for stable training and constraint satisfaction.\nWe address these challenges straightforwardly by treating the maximization of multiple objectives as a constrained optimization problem (COP), where the constraints are defined to improve the original objectives.\nExisting safety constraints are then integrated into the COP, and the policy is updated by solving the COP, which ensures the avoidance of gradient conflicts.\nDespite its simplicity, CoMOGA guarantees convergence to global optima in a tabular setting.\nThrough various experiments, we have confirmed that preventing gradient conflicts is critical, and the proposed method achieves constraint satisfaction across all tasks.", "keywords": [ "Multi-Objective Reinforcement Learning", "Safe Reinforcement Learning", "Gradient Aggregation" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "ogO6DGE6FZ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "Bilge Soran", "gender": "Female", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Changsheng Zhao", "gender": "Male", "institution": "Meta Inc.", "country": "US", "position": "Researcher" }, { "name": "Dhruv Choudhary", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Igor Fedorov", "gender": "Male", "institution": "Meta", "country": "US", "position": "research scientist " }, { "name": "Raghuraman Krishnamoorthi", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Tijmen Blankevoort", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Vikas Chandra", "gender": "Male", "institution": "Meta", "country": "US", "position": "Director, AI" }, { "name": "Yuandong Tian", "gender": "Male", "institution": "Meta AI (FAIR)", "country": "", "position": "Research Scientist" }, { "name": "Zechun Liu", "gender": "unknown", "institution": "Meta Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 39, "n_ref": 90, "n_ref_all": 132, "n_fig": 14, "n_tab": 14, "L_tab": 12820, "n_element_tab": 1656, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 3543, "n_element_tab_1": 433, "formula_len_all": 1374, "formula_len_all_1": 1472, "len_all": 179252, "len_all_1": 65374, "len_abs": 1103, "len_title": 104, "len_sents": 41010, "len_sents_1": 25580, "n_sents": 283, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1106, "title": "SpinQuant: LLM Quantization with Learned Rotations", "abs": "Post-training quantization (PTQ) techniques applied to weights, activations, and the KV cache greatly reduce memory usage, latency, and power consumption of Large Language Models (LLMs), but may lead to large quantization errors when outliers are present. Rotating activation or weight matrices helps remove outliers and benefits quantization. In this work, we identify a collection of applicable rotation parameterizations that lead to identical outputs in full-precision Transformer architectures while enhancing quantization accuracy. In addition, we find that some random rotations lead to much better quantization than others, with an up to 13 points difference in downstream zero-shot reasoning performance. As a result, we propose SpinQuant, a novel approach that incorporates learned rotation matrices for optimal quantized network accuracy. With 4-bit quantization of weight, activation, and KV-cache, SpinQuant narrows the accuracy gap on zero-shot reasoning tasks with full precision to merely 2.9 points on the LLaMA-2 7B model, surpassing LLM-QAT by 19.1 points and SmoothQuant by 25.0 points.", "keywords": [ "rotation", "information retrieval" ], "rating_list": [ 6, 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 4, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 3, 2 ], "confidence_list": [ 4, 2, 4, 3, 4 ] }, { "paper_id": "ogKE7LcvW6", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Berfin Simsek", "gender": "Female", "institution": "Flatiron Institute", "country": "US", "position": "Researcher" }, { "name": "Fran\u00e7ois Gaston Ged", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "Postdoc" }, { "name": "Zhengqing Wu", "gender": "Male", "institution": "EPFL", "country": "CH", "position": "PhD student" } ], "n_formula": 129, "n_formula_1": 12, "n_ref_uni": 52, "n_ref": 166, "n_ref_all": 249, "n_fig": 28, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12775, "formula_len_all_1": 744, "len_all": 284263, "len_all_1": 77655, "len_abs": 1442, "len_title": 159, "len_sents": 95842, "len_sents_1": 35628, "n_sents": 861, "n_sents_1": 283, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 110, "L_abs": 1212, "title": "Loss Landscape of Shallow ReLU-like Neural Networks: Stationary Points, Saddle Escaping, and Network Embedding", "abs": "In this paper, we investigate the loss landscape of one-hidden-layer neural networks with ReLU-like activation functions trained with the empirical squared loss. As the activation function is non-differentiable, it is so far unclear how to completely characterize the stationary points. We deduce the conditions for stationarity that apply to both non-differentiable and differentiable areas of the landscape. Additionally, we show that, if a stationary point does not contain \"escape neurons\", which are defined with first-order conditions, then it must be a local minimum. Moreover, for the scalar-output case, the presence of an escape neuron guarantees that the stationary point is not a local minimum. Our results refine the description of the saddle-to-saddle training process starting from infinitesimally small (vanishing) initialization for shallow ReLU-like networks. By precluding the existence of the saddle escaping types that previous works did not rule out, we advance one step closer to a complete picture of the entire dynamics. Moreover, we are also able to fully discuss how network embedding, which is to instantiate a narrower network within a wider network, reshapes the stationary points.", "keywords": [ "loss landscape", "non-smooth", "training dynamics", "network embedding", "over-parameterization", "neural network" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "ogIFNo2bQw", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Antonia Wachter-Zeh", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "Associate Professor" }, { "name": "Deniz Gunduz", "gender": "unknown", "institution": "Imperial College London", "country": "", "position": "Full Professor" }, { "name": "Maximilian Egger", "gender": "unknown", "institution": "Technical University of Munich", "country": "DE", "position": "PhD student" }, { "name": "Nir Weinberger", "gender": "Male", "institution": "Electrical Engineering Department, Technion \u2013 Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "Assistant Professor" }, { "name": "Rawad Bitar", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Researcher" } ], "n_formula": 49, "n_formula_1": 8, "n_ref_uni": 32, "n_ref": 83, "n_ref_all": 94, "n_fig": 3, "n_tab": 15, "L_tab": 8682, "n_element_tab": 920, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 879, "n_element_tab_1": 29, "formula_len_all": 5499, "formula_len_all_1": 822, "len_all": 171563, "len_all_1": 76382, "len_abs": 1497, "len_title": 119, "len_sents": 50373, "len_sents_1": 36690, "n_sents": 398, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1504, "title": "BiCompFL: Stochastic Federated Learning with Bi-Directional Compression", "abs": "Communication is a prominent bottleneck in federated learning (FL). State-of-the-art accuracy performance under limited uplink communications from the clients to the federator is achieved by stochastic FL approaches. It has been recently shown that leveraging side information in the form of a prior distribution at the federator can drastically reduce the uplink communication cost in stochastic FL. Here, the latest global model distribution serves as a natural prior since it can be shared with the clients under ideal downlink communication from the federator to the clients. Nevertheless, downlink communication is often limited in practical settings, and bi-directional compression must be considered to reduce the overall communication cost. The extension of existing stochastic FL solutions to bi-directional compression is non-trivial due to the lack of a globally shared common prior distribution at each iteration. In this paper, we propose BiCompFL, which employs importance sampling to send samples from the updated local models in the uplink, and the aggregated global model in the downlink by carefully choosing common prior distributions as side-information. We theoretically study the communication cost by a new analysis of importance sampling that refines known results, and exposes the interplay between uplink and downlink communication costs. We also show through numerical experiments that BiCompFL enables multi-fold savings in communication cost compared to the state-of-the-art.", "keywords": [ "Communication-efficiency", "Importance sampling", "Stochastic federated learning" ], "rating_list": [ 5, 6, 3, 5, 5 ], "soundness_list": [ 3, 4, 2, 4, 3 ], "presentation_list": [ 3, 4, 1, 4, 3 ], "contribution_list": [ 2, 4, 2, 3, 2 ], "confidence_list": [ 4, 3, 5, 4, 3 ] }, { "paper_id": "ofuLWn8DFZ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Stephan G\u00fcnnemann", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Professor" }, { "name": "Yan Scholten", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 17, "n_ref_uni": 38, "n_ref": 52, "n_ref_all": 83, "n_fig": 17, "n_tab": 3, "L_tab": 1384, "n_element_tab": 22, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2110, "formula_len_all_1": 1220, "len_all": 156986, "len_all_1": 68746, "len_abs": 1477, "len_title": 125, "len_sents": 59397, "len_sents_1": 33430, "n_sents": 455, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 77, "L_abs": 1483, "title": "Provably Reliable Conformal Prediction Sets in the Presence of Data Poisoning", "abs": "Conformal prediction provides model-agnostic and distribution-free uncertainty quantification through prediction sets that are guaranteed to include the ground truth with any user-specified probability. Yet, conformal prediction is not reliable under poisoning attacks where adversaries manipulate both training and calibration data, which can significantly alter prediction sets in practice. As a solution, we propose reliable prediction sets (RPS): the first efficient method for constructing conformal prediction sets with provable reliability guarantees under poisoning. To ensure reliability under training poisoning, we introduce smoothed score functions that reliably aggregate predictions of classifiers trained on distinct partitions of the training data. To ensure reliability under calibration poisoning, we construct multiple prediction sets, each calibrated on distinct subsets of the calibration data. We then aggregate them into a majority prediction set, which includes a class only if it appears in a majority of the individual sets. Both proposed aggregations mitigate the influence of datapoints in the training and calibration data on the final prediction set. We experimentally validate our approach on image classification tasks, achieving strong reliability while maintaining utility and preserving coverage on clean data. Overall, our approach represents an important step towards more trustworthy uncertainty quantification in the presence of data poisoning.", "keywords": [ "Conformal prediction", "Certifiable robustness", "Adversarial robustness" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 4, 3, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "ofiZbAmrZh", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Balaji Ganesan", "gender": "Male", "institution": "IBM Research India", "country": "IN", "position": "Research Engineer" }, { "name": "Debarun Bhattacharjya", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Junkyu Lee", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Katsiaryna Mirylenka", "gender": "Female", "institution": "International Business Machines", "country": "CH", "position": "Researcher" }, { "name": "Michael Glass", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Radu Marinescu", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Xiao Shou", "gender": "Male", "institution": "Baylor University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 51, "n_ref": 121, "n_ref_all": 145, "n_fig": 4, "n_tab": 2, "L_tab": 1936, "n_element_tab": 139, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1333, "n_element_tab_1": 125, "formula_len_all": 425, "formula_len_all_1": 337, "len_all": 173944, "len_all_1": 72079, "len_abs": 1691, "len_title": 144, "len_sents": 44163, "len_sents_1": 35403, "n_sents": 274, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 96, "L_abs": 1717, "title": "Consistency-based Black-box Uncertainty Quantification for Text-to-SQL by Similarity Aggregation", "abs": "When does a large language model (LLM) know what it does not know? Uncertainty quantification (UQ) provides an estimate of the confidence in an LLM's generated output and is therefore increasingly recognized as a crucial component of trusted AI systems. UQ is particularly important for complex generative tasks such as \\emph{text-to-SQL}, where an LLM helps users gain insights about data stored in noisy and large databases by translating their natural language queries to structured query language (SQL). \\emph{Black-box} UQ methods do not require access to internal model information from the generating LLM, and therefore have numerous real-world advantages, such as robustness to system changes, adaptability to choice of LLM (including those with commercialized APIs), reduced costs, and substantial computational tractability. In this paper, we investigate the effectiveness of black-box UQ techniques for text-to-SQL, where the consistency between a generated output and other sampled generations is used as a proxy for estimating its confidence. We propose a high-level non-verbalized \\emph{similarity aggregation} approach that is suitable for complex generative tasks, including specific techniques that train confidence estimation models using small training sets. Through an extensive empirical study over various text-to-SQL datasets and models, we provide recommendations for the choice of sampling technique and similarity metric. The experiments demonstrate that our proposed similarity aggregation techniques result in better calibrated confidence estimates as compared to the closest baselines, but also highlight how there is room for improvement on downstream tasks such as selective generation.", "keywords": [ "text-to-sql", "uncertainty quantification", "black-box", "similarity", "large language model", "generative model" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "of6EuHT7de", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Christophe Roux", "gender": "unknown", "institution": "Zuse Institute Berlin", "country": "DE", "position": "PhD student" }, { "name": "Max Zimmer", "gender": "unknown", "institution": "Zuse Institute Berlin", "country": "DE", "position": "PhD student" }, { "name": "Sebastian Pokutta", "gender": "Male", "institution": "ZIB", "country": "DE", "position": "Vice President" } ], "n_formula": 36, "n_formula_1": 9, "n_ref_uni": 27, "n_ref": 43, "n_ref_all": 74, "n_fig": 7, "n_tab": 13, "L_tab": 3971, "n_element_tab": 447, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 644, "n_element_tab_1": 40, "formula_len_all": 2944, "formula_len_all_1": 685, "len_all": 193767, "len_all_1": 67828, "len_abs": 2008, "len_title": 116, "len_sents": 53453, "len_sents_1": 31929, "n_sents": 449, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1298, "title": "On the Byzantine-Resilience of Distillation-Based Federated Learning", "abs": "Federated Learning (FL) algorithms using Knowledge Distillation (KD) have received increasing attention due to their favorable properties with respect to privacy, non-i.i.d. data and communication cost. These methods depart from transmitting model parameters and instead communicate information about a learning task by sharing predictions on a public dataset. In this work, we study the performance of such approaches in the byzantine setting, where a subset of the clients act in an adversarial manner aiming to disrupt the learning process. We show that KD-based FL algorithms are remarkably resilient and analyze how byzantine clients can influence the learning process. Based on these insights, we introduce two new byzantine attacks and demonstrate their ability to break existing byzantine-resilient methods. Additionally, we propose a novel defence method which enhances the byzantine resilience of KD-based FL algorithms. Finally, we provide a general framework to obfuscate attacks, making them significantly harder to detect, thereby improving their effectiveness. Our findings serve as an important building block in the analysis of byzantine FL, contributing through the development of new attacks and new defence mechanisms, further advancing the robustness of KD-based FL algorithms.", "keywords": [ "Federated Learning", "Knowledge Distillation", "Byzantine FL" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "of25Zg4AdM", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Claire Vernade", "gender": "Female", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "", "position": "Assistant Professor" }, { "name": "Mirabel Reid", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Samira Samadi", "gender": "Female", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "", "position": "Research Group Leader" }, { "name": "Tom S\u00fchr", "gender": "Male", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 13, "n_ref_uni": 36, "n_ref": 73, "n_ref_all": 88, "n_fig": 9, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1560, "formula_len_all_1": 1172, "len_all": 113920, "len_all_1": 65867, "len_abs": 827, "len_title": 98, "len_sents": 37603, "len_sents_1": 29994, "n_sents": 334, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 832, "title": "Online Decision Deferral under Budget Constraints", "abs": "Machine Learning (ML) models are increasingly used to support or substitute decision making. In applications where skilled experts are a limited resource, it is crucial to reduce their burden and automate decisions when the performance of an ML model is at least of equal quality. \nHowever, models are often pre-trained and fixed, while tasks arrive sequentially and their distribution may shift. In that case, the respective performance of the decision makers may change, and the deferral algorithm must remain adaptive. We propose a contextual bandit model of this online decision making problem. Our framework includes budget constraints and different types of partial feedback models. Beyond the theoretical guarantees of our algorithm, we propose efficient extensions that achieve remarkable performance on real-world datasets.", "keywords": [ "Online Learning", "Human-AI Collaboration" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 1, 3, 1 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "oegbNuUrXV", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Beining Wang", "gender": "Male", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "DI QI", "gender": "Female", "institution": "Step ai", "country": "", "position": "Researcher" }, { "name": "Tong Yang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Wenqiang Zhang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Xiangyu Zhang", "gender": "Male", "institution": "MEGVII Technology", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 35, "n_ref": 83, "n_ref_all": 97, "n_fig": 8, "n_tab": 5, "L_tab": 1470, "n_element_tab": 172, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1443, "n_element_tab_1": 170, "formula_len_all": 236, "formula_len_all_1": 236, "len_all": 135122, "len_all_1": 65363, "len_abs": 1300, "len_title": 104, "len_sents": 37921, "len_sents_1": 28567, "n_sents": 305, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1189, "title": "Generalizable Dynamic Radiance Field in Egocentric View", "abs": "We present a novel framework for generalizable dynamic radiance field in egocentric view. Our approach can predict a 3D representation of the physical world at a given time based on a monocular video without test-time training. To this end, we use a contracted triplane as the 3D representation of physical world in an egocentric view at a specific time. To update the explicit 3D representation, we propose a 4D-aware transformer module to aggregate features from monocular videos. Besides, we also introduce a temporal-based 3D constraint to achieve better multiview consistency. In addition, we train the proposed model with large-scale monocular videos in a self-supervised manner. Our model achieves top results in novel view synthesis on dynamic scene datasets, demonstrating its strong understanding of 4D physical world. Besides, our model also shows the superior generalizability to unseen scenarios. Furthermore, we find that our approach emerges capabilities for geometry and semantic learning. We hope our approach can provide preliminary understanding of the physical world in first-person view and help ease future research in computer vision, computer graphics and robotics.", "keywords": [ "generalized dynamic view synthesis", "NeRF", "computer vision" ], "rating_list": [ 5, 5, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 1, 2 ], "confidence_list": [ 5, 5, 4, 4, 4 ] }, { "paper_id": "oecFal31WP", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Chang Gong", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Di Yao", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Haining TAN", "gender": "unknown", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "Assistant Professor" }, { "name": "Jingping Bi", "gender": "Female", "institution": ", Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Ruibo Zhao", "gender": "Male", "institution": ", Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Wenbin Li", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Wenjie Chen", "gender": "Male", "institution": ", Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Zijie Xu", "gender": "Male", "institution": ", Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "chengxue luo", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "quanliang jing", "gender": "Male", "institution": ", Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 36, "n_ref": 52, "n_ref_all": 76, "n_fig": 3, "n_tab": 6, "L_tab": 1888, "n_element_tab": 102, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 958, "n_element_tab_1": 49, "formula_len_all": 316, "formula_len_all_1": 458, "len_all": 110885, "len_all_1": 66164, "len_abs": 1347, "len_title": 132, "len_sents": 35655, "len_sents_1": 30656, "n_sents": 278, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1243, "title": "STBench: Assessing the Ability of Large Language Models in Spatio-Temporal Analysis", "abs": "The rapid evolution of large language models (LLMs) holds promise for reforming the methodology of spatio-temporal data mining. However, current works for evaluating the spatial-temporal understanding capability of LLMs are somewhat limited and biased. These works either fail to incorporate the latest language models or only focus on assessing the memorized spatio-temporal knowledge. To address this gap, this paper dissects LLMs' capability of spatio-temporal data into four distinct dimensions: knowledge comprehension, spatio-temporal reasoning, accurate computation, and downstream applications. We curate several natural language question-answer tasks for each category and build the benchmark dataset, namely STBench, containing 15 distinct tasks and over 70,000 QA pairs. Moreover, we have assessed the capabilities of the 13 LLMs, including Chat-4o, Gemma, and Mistral. Experimental results reveal that existing LLMs show remarkable performance on knowledge comprehension and spatio-temporal reasoning tasks, with potential for further enhancement on other tasks through in-context learning, chain-of-though prompting, and fine-tuning. The code and datasets of STBench are released on https://anonymous.4open.science/r/STBench-14C2.", "keywords": [ "Large Language Models", "Benchmark", "Spatio-temporal data mining" ], "rating_list": [ 5, 5, 5, 8 ], "soundness_list": [ 2, 2, 2, 4 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "oeP6OL7ouB", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Bernhard Kerbl", "gender": "Male", "institution": "TU Wien Vienna University of Technology", "country": "", "position": "Postdoc" }, { "name": "Katia P. Sycara", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Silong Yong", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Simon Stepputtis", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Venkata Nagarjun Pudureddiyur Manivannan", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Yaqi Xie", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Zifu Wan", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 48, "n_ref": 131, "n_ref_all": 151, "n_fig": 10, "n_tab": 1, "L_tab": 406, "n_element_tab": 6, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 381, "n_element_tab_1": 54, "formula_len_all": 1198, "formula_len_all_1": 772, "len_all": 131506, "len_all_1": 62250, "len_abs": 1319, "len_title": 113, "len_sents": 34200, "len_sents_1": 29388, "n_sents": 240, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1284, "title": "OMG: Opacity Matters in Material Modeling with Gaussian Splatting", "abs": "Decomposing geometry, materials and lighting from a set of images, namely inverse rendering, has been a long-standing problem in computer vision and graphics. Recent advances in neural rendering enable photo-realistic and plausible inverse rendering results. The emergence of 3D Gaussian Splatting has boosted it to the next level. An intuitive finding is that the models used for inverse rendering do not take into account the dependency of opacity w.r.t. material properties, namely cross section, as suggested by optics. Therefore, we develop a novel approach that adds this dependency to the modeling itself. Inspired by radiative transfer, we augment the opacity term by introducing a neural network that takes as input material properties to provide modeling of cross section and a physically correct activation function. The gradients for material properties are therefore not only from color but also from opacity, facilitating a constraint for their optimization. Therefore, the proposed method incorporates more accurate physical properties compared to previous works. We implement our method into 3 different baselines that use Gaussian Splatting for inverse rendering and achieve significant improvements universally in terms of novel view synthesis and material modeling.", "keywords": [ "3D Gaussian Splatting", "Neural Rendering", "Inverse Rendering", "Visual Computing" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "oeLB25A9oO", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Byoung-Tak Zhang", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Cheongjae Jang", "gender": "unknown", "institution": "Hanyang University", "country": "KR", "position": "Researcher" }, { "name": "HEE BIN YOO", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Jaein Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Seunghyeon Lim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Sungyoon Lee", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "Assistant Professor" }, { "name": "Dong-Sig Han Han", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" } ], "n_formula": 102, "n_formula_1": 12, "n_ref_uni": 74, "n_ref": 122, "n_ref_all": 140, "n_fig": 4, "n_tab": 4, "L_tab": 222, "n_element_tab": 46, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18451, "formula_len_all_1": 1479, "len_all": 215503, "len_all_1": 64093, "len_abs": 1606, "len_title": 121, "len_sents": 51616, "len_sents_1": 28385, "n_sents": 472, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1623, "title": "Emergence of Alignment and Local Elasticity in Two-Layer Neural Networks", "abs": "Investigating phenomena such as Alignment and Local Elasticity is essential for understanding feature space of Neural Networks and enhancing performance across a wide range of tasks.\nIn this context, we investigate the emergence of these phenomena in two-layer neural networks performing a classification task.\nThis paper reveals Alignment and Local Elasticity emergence condition after one step of training are identical. \nIn particular, we demonstrate that intra-class features are more aligned when the inner product of their mean and the covariance of the training data-label \\ie \\textit{train-unseen similarity} is large, with stronger Local Elasticity occurring under this condition.\nWe validate our theory through experiments with a two-layer network showing that both Alignment and Local Elasticity improve as the train-unseen similarity increases.\nFurthermore, we claim that our analysis provides both theoretical and practical insights into the relationship between train-unseen similarity, alignment, and the improvement of clustering performance on unseen data for neural networks trained on similar domain data. This is supported by experiments, including a multi-layer CNN setup and detailed discussions.\nSpecifically, we show that higher train-unseen similarity improves Recall@1 in two-layer networks and that Alignment and Recall@1 exhibit a positive correlation in metric learning. \nWe also present novel techniques for deriving operator norm bounds of non-centered Sub-Gaussian matrices, extending conventional regression analysis with standard Gaussian assumptions to the binary classification setting.", "keywords": [ "two-layer neural network", "feature learning", "metric learning", "local elasticity", "retrieval", "random matrix theory" ], "rating_list": [ 3, 3, 5, 6, 3 ], "soundness_list": [ 2, 2, 2, 3, 1 ], "presentation_list": [ 1, 1, 1, 2, 1 ], "contribution_list": [ 1, 1, 2, 3, 2 ], "confidence_list": [ 5, 4, 3, 3, 4 ] }, { "paper_id": "oeDcgVC7Xh", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Claudia Iriondo", "gender": "Female", "institution": "Genentech", "country": "", "position": "Researcher" }, { "name": "Louis McConnell", "gender": "Male", "institution": "Genentech Research and Early Development", "country": "US", "position": "Researcher" }, { "name": "Yulun Wu", "gender": "unknown", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 94, "n_formula_1": 16, "n_ref_uni": 50, "n_ref": 113, "n_ref_all": 184, "n_fig": 18, "n_tab": 9, "L_tab": 1616, "n_element_tab": 224, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 848, "n_element_tab_1": 101, "formula_len_all": 11518, "formula_len_all_1": 1563, "len_all": 251877, "len_all_1": 76580, "len_abs": 1372, "len_title": 116, "len_sents": 85784, "len_sents_1": 36003, "n_sents": 564, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1359, "title": "Counterfactual Generative Modeling with Variational Causal Inference", "abs": "Estimating an individual's potential outcomes under counterfactual treatments is a challenging task for traditional causal inference and supervised learning approaches when the outcome is high-dimensional (e.g. gene expressions, facial images) and covariates are relatively limited. In this case, to predict one's outcomes under counterfactual treatments, it is crucial to leverage individual information contained in the observed outcome in addition to the covariates. Prior works using variational inference in counterfactual generative modeling have been focusing on neural adaptations and model variants within the conditional variational autoencoder formulation, which we argue is fundamentally ill-suited to the notion of counterfactual in causal inference. In this work, we present a novel variational Bayesian causal inference framework and its theoretical backings to properly handle counterfactual generative modeling tasks, through which we are able to conduct counterfactual supervision end-to-end during training without any counterfactual samples, and encourage disentangled exogenous noise abduction that aids the correct identification of causal effect in counterfactual generations. In experiments, we demonstrate the advantage of our framework compared to state-of-the-art models in counterfactual generative modeling on multiple benchmarks.", "keywords": [ "deep probabilistic modeling", "bayesian networks", "variational inference", "causal inference", "counterfactual generative modeling", "causal machine learning", "deep unsupervised learning", "graphical models" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 3, 3, 3, 1 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "oe51Q5Uo37", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Arijit Sehanobish", "gender": "Male", "institution": "Kensho Technologies", "country": "US", "position": "Applied Scientist" }, { "name": "Krzysztof Marcin Choromanski", "gender": "unknown", "institution": "Google Brain Robotics & Columbia University", "country": "", "position": "research scientist & adjunct assistant professor" }, { "name": "Snigdha Chaturvedi", "gender": "Female", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Associate Professor" }, { "name": "Somnath Basu Roy Chowdhury", "gender": "unknown", "institution": "Department of Computer Science, University of North Carolina, Chapel Hill", "country": "", "position": "PhD student" }, { "name": "Kumar A Dubey", "gender": "Male", "institution": "Google Research", "country": "", "position": "Research Scientist" } ], "n_formula": 34, "n_formula_1": 7, "n_ref_uni": 56, "n_ref": 89, "n_ref_all": 141, "n_fig": 16, "n_tab": 5, "L_tab": 2937, "n_element_tab": 345, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2991, "formula_len_all_1": 750, "len_all": 220349, "len_all_1": 73319, "len_abs": 1743, "len_title": 127, "len_sents": 63155, "len_sents_1": 34245, "n_sents": 564, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1747, "title": "Towards Scalable Exact Machine Unlearning Using Parameter-Efficient Fine-Tuning", "abs": "Machine unlearning is the process of efficiently removing the influence of a training data instance from a trained machine learning model without retraining it from scratch. A popular subclass of unlearning approaches is exact machine unlearning, which focuses on techniques that explicitly guarantee the removal of the influence of a data instance from a model. Exact unlearning approaches use a machine learning model in which individual components are trained on disjoint subsets of the data. During deletion, exact unlearning approaches only retrain the affected components rather than the entire model. While existing approaches reduce retraining costs, it can still be expensive for an organization to retrain a model component as it requires halting a system in production, which leads to service failure and adversely impacts customers. To address these challenges, we introduce an exact unlearning framework -- Sequence-aware Sharded Sliced Training (S3T), which is designed to enhance the deletion capabilities of an exact unlearning system while minimizing the impact on model's performance. At the core of S3T, we utilize a lightweight parameter-efficient fine-tuning approach that enables parameter isolation by sequentially training layers with disjoint data slices. This enables efficient unlearning by simply deactivating the layers affected by data deletion. Furthermore, to reduce the retraining cost and improve model performance, we train the model on multiple data sequences, which allows S3T to handle an increased number of deletion requests. Both theoretically and empirically, we demonstrate that S3T attains superior deletion capabilities and enhanced performance compared to baselines across a wide range of settings.", "keywords": [ "Machine Unlearning", "Data Deletion", "Exact Unlearning" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "odvSjn416y", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Ashwin Paranjape", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Dawn Lawrie", "gender": "Female", "institution": "Johns Hopkins University", "country": "", "position": "Researcher" }, { "name": "Jack Hessel", "gender": "Male", "institution": "Samaya AI", "country": "", "position": "Researcher" }, { "name": "Orion Weller", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "PhD student" }, { "name": "Yuhao Zhang", "gender": "unknown", "institution": "Amazon AWS AI", "country": "", "position": "Scientist" }, { "name": "Benjamin Van Durme", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 29, "n_ref": 64, "n_ref_all": 111, "n_fig": 5, "n_tab": 9, "L_tab": 2629, "n_element_tab": 283, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 4057, "n_element_tab_1": 313, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 188963, "len_all_1": 72223, "len_abs": 1118, "len_title": 129, "len_sents": 61738, "len_sents_1": 29741, "n_sents": 483, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1100, "title": "Promptriever: Instruction-Trained Retrievers Can Be Prompted Like Language Models", "abs": "Instruction-tuned language models (LM) are able to respond to imperative commands, providing a more natural user interface compared to their base counterparts. In this work, we present Promptriever, the first retrieval model able to be prompted like an LM. To train Promptriever, we curate and release a new instance-level instruction training set from MS MARCO, spanning nearly 500k instances. Promptriever not only achieves strong performance on standard retrieval tasks, but also follows instructions. We observe: (1) large gains (reaching SoTA) on following detailed relevance instructions (+14.3 p-MRR / +3.1 nDCG on FollowIR), (2) significantly increased robustness to lexical choices/phrasing in the query+instruction (+12.9 Robustness@10 on InstructIR), and (3) the ability to perform hyper-parameter search via prompting to reliably improve retrieval performance (+1.4 average increase on BEIR). Promptriever demonstrates that retrieval models can be controlled with prompts on a per-query basis, setting the stage for future work aligning LM prompting techniques with information retrieval.", "keywords": [ "retrieval", "instructions", "search", "prompts" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "odjMSBSWRt", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Akash Kundu", "gender": "Male", "institution": "Heritage Institute of Technology", "country": "IN", "position": "Undergrad student" }, { "name": "Esben Kran", "gender": "Male", "institution": "Apart Research", "country": "", "position": "Researcher" }, { "name": "Hieu Minh Nguyen", "gender": "Not Specified", "institution": "University of Science and Technology of Hanoi", "country": "VN", "position": "Undergrad student" }, { "name": "Jinsuk Park", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Undergrad student" }, { "name": "Mateusz Maria Jurewicz", "gender": "Not Specified", "institution": "IT University of Copenhagen", "country": "", "position": "PhD student" }, { "name": "Sami Jawhar", "gender": "Male", "institution": "METR", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 52, "n_ref": 84, "n_ref_all": 100, "n_fig": 5, "n_tab": 4, "L_tab": 6133, "n_element_tab": 375, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 593, "n_element_tab_1": 10, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 156693, "len_all_1": 72384, "len_abs": 705, "len_title": 62, "len_sents": 30354, "len_sents_1": 21559, "n_sents": 226, "n_sents_1": 141, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 62, "L_abs": 706, "title": "DarkBench: Benchmarking Dark Patterns in Large Language Models", "abs": "We introduce DarkBench, a comprehensive benchmark for detecting dark design patterns\u2014manipulative techniques that influence user behavior\u2014in interactions with large language models (LLMs). Our benchmark comprises 660 prompts across six categories: brand bias, user retention, sycophancy, anthropomorphism, harmful generation, and sneaking. We evaluate models from five leading companies (OpenAI, Anthropic, Meta, Mistral, Google) and find that some LLMs are explicitly designed to favor their developers' products and exhibit untruthful communication, among other manipulative behaviors. Companies developing LLMs should recognize and mitigate the impact of dark design patterns to promote more ethical Al.", "keywords": [ "Dark Patterns", "AI Deception", "Large Language Models" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "odU59TxdiB", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Armin Mustafa", "gender": "Female", "institution": "University of Surrey", "country": "GB", "position": "Lecturer" }, { "name": "Philip J B Jackson", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Full Professor" }, { "name": "Muhammad Awais Tanvir Rana", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Associate Professor" }, { "name": "Sara Atito Ahmed", "gender": "Female", "institution": "University of Surrey", "country": "GB", "position": "Postdoc" }, { "name": "TONY ALEX", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 30, "n_ref": 62, "n_ref_all": 95, "n_fig": 4, "n_tab": 11, "L_tab": 5289, "n_element_tab": 430, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 3186, "n_element_tab_1": 232, "formula_len_all": 466, "formula_len_all_1": 331, "len_all": 140311, "len_all_1": 66040, "len_abs": 2015, "len_title": 135, "len_sents": 43308, "len_sents_1": 29330, "n_sents": 307, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 2104, "title": "SSLAM: Enhancing Self-Supervised Models with Audio Mixtures for Polyphonic Soundscapes", "abs": "Self-supervised pre-trained audio networks have seen widespread adoption in real-world systems, particularly in multi-modal large language models. These networks are often employed in a frozen state, under the assumption that the self-supervised pre-training has sufficiently equipped them to handle real-world audio. However, a critical question remains: how well do these models actually perform in real-world conditions, where audio is typically polyphonic and complex, involving multiple overlapping sound sources? Current audio self-supervised learning (SSL) methods are often benchmarked on datasets predominantly featuring monophonic audio, such as environmental sounds, and speech. As a result, the ability of SSL models to generalize to polyphonic audio, a common characteristic in natural scenarios, remains underexplored. This limitation raises concerns about the practical robustness of SSL models in more realistic audio settings. To address this gap, we introduce Self-Supervised Learning from Audio Mixtures (SSLAM), a novel direction in audio SSL research, designed to improve the model\u2019s ability to learn from polyphonic data while maintaining strong performance on monophonic data. We thoroughly evaluate SSLAM on standard audio SSL benchmark datasets which are predominantly monophonic and conduct a comprehensive comparative analysis against state-of-the-art (SOTA) methods using a range of high-quality, publicly available polyphonic datasets. SSLAM not only improves model performance on polyphonic audio, but also maintains or exceeds performance on standard audio SSL benchmarks. Notably, it achieves up to a 3.9% improvement on the AudioSet-2M(AS-2M), reaching a mean average precision (mAP) of 50.2. For polyphonic datasets, SSLAM sets new SOTA in both linear evaluation and fine-tuning regimes with performance improvements of up to 9.1%(mAP). These results demonstrate SSLAM's effectiveness in both polyphonic and monophonic soundscapes, significantly enhancing the performance of audio SSL models. Code and pre-trained models are available at https://github.com/ta012/SSLAM.", "keywords": [ "audio self-supervised learning", "audio representation learning" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 4, 2, 2, 4 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "oc4yw7zX9T", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Jiaye Teng", "gender": "Male", "institution": "Shanghai University of Finance and Economics", "country": "CN", "position": "Assistant Professor" }, { "name": "Shaoli Wang", "gender": "unknown", "institution": "Shanghai University of Finance and Economics", "country": "", "position": "Associate Professor" }, { "name": "Xin Liu", "gender": "Male", "institution": "Shanghai University of Finance and Economics", "country": "", "position": "Associate Professor" }, { "name": "Zhen Zhang", "gender": "Male", "institution": "Shanghai University of Finance and Economics", "country": "CN", "position": "PhD student" } ], "n_formula": 119, "n_formula_1": 35, "n_ref_uni": 38, "n_ref": 62, "n_ref_all": 76, "n_fig": 6, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15660, "formula_len_all_1": 2572, "len_all": 197866, "len_all_1": 77229, "len_abs": 1206, "len_title": 127, "len_sents": 58246, "len_sents_1": 30708, "n_sents": 558, "n_sents_1": 268, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1108, "title": "Minimax Optimal Two-Stage Algorithm For Moment Estimation Under Covariate Shift", "abs": "Covariate shift occurs when the distribution of input features differs between the training and testing phases. In covariate shift, estimating an unknown function's moment is a classical problem that remains under-explored, despite its common occurrence in real-world scenarios. In this paper, we investigate the minimax lower bound of the problem when the source and target distributions are known. To achieve the minimax optimal bound (up to a logarithmic factor), we propose a two-stage algorithm. Specifically, it first trains an optimal estimator for the function under the source distribution, and then uses a likelihood ratio reweighting procedure to calibrate the moment estimator. In practice, the source and target distributions are typically unknown, and estimating the likelihood ratio may be unstable. To solve this problem, we propose a truncated version of the estimator that ensures double robustness and provide the corresponding upper bound. Extensive numerical studies on synthetic examples confirm our theoretical findings and further illustrate the effectiveness of our proposed method.", "keywords": [ "covariate shift", "minimax optimal", "two-stage algorithm", "double robustness" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "obYVdcMMIT", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Cho-Jui Hsieh", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Ion Stoica", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Justin Cui", "gender": "Male", "institution": ", University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Wei-Lin Chiang", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 133, "n_ref_all": 166, "n_fig": 8, "n_tab": 10, "L_tab": 8342, "n_element_tab": 1099, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 3198, "n_element_tab_1": 503, "formula_len_all": 37, "formula_len_all_1": 0, "len_all": 230154, "len_all_1": 70835, "len_abs": 2516, "len_title": 109, "len_sents": 61292, "len_sents_1": 29524, "n_sents": 501, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1375, "title": "OR-Bench: An Over-Refusal Benchmark for Large Language Models", "abs": "Large Language Models (LLMs) require careful safety alignment to prevent malicious outputs. While significant research focuses on mitigating harmful content generation, \nthe enhanced safety often come with the side effect of over-refusal, where LLMs may reject innocuous prompts and become less helpful.\nAlthough the issue of over-refusal has been empirically observed, a systematic measurement is challenging \ndue to the difficulty of crafting prompts that can elicit the over-refusal behaviors of LLMs.\nThis study proposes a novel method for automatically generating large-scale over-refusal datasets. Leveraging this technique, we introduce OR-Bench, the first large-scale over-refusal benchmark. OR-Bench comprises 80,000 over-refusal prompts across 10 common rejection categories, a subset of around 1,000 hard prompts that are challenging even for state-of-the-art LLMs, and an additional 600 toxic prompts to prevent indiscriminate responses.\nWe then conduct a comprehensive study to measure the over-refusal of 32 popular LLMs across 8 model families. \nTo facilitate reproducibility, we host our datasets, along with an interactive demo and leaderboard, on HuggingFace at https://huggingface.co/spaces/orbench-llm/or-bench and release our code at https://github.com/orbench/or-bench. We hope this benchmark can help the community develop better safety aligned models.", "keywords": [ "over-refusal", "llm", "safety", "false-refusal" ], "rating_list": [ 8, 6, 6, 1, 3 ], "soundness_list": [ 3, 3, 3, 1, 2 ], "presentation_list": [ 4, 3, 3, 2, 3 ], "contribution_list": [ 4, 3, 3, 1, 3 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "obYDlJN0oU", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Kelvin J.L. Koa", "gender": "unknown", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Ritchie Ng", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Tingwen Du", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Xiang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yunshan Ma", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Zheng Huanhuan", "gender": "Female", "institution": "National University of Singapore", "country": "", "position": "Assistant Professor" }, { "name": "Tat-seng Chua", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 42, "n_ref": 82, "n_ref_all": 105, "n_fig": 7, "n_tab": 3, "L_tab": 337, "n_element_tab": 31, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 69, "n_element_tab_1": 10, "formula_len_all": 315, "formula_len_all_1": 316, "len_all": 145500, "len_all_1": 69741, "len_abs": 1738, "len_title": 132, "len_sents": 47739, "len_sents_1": 34960, "n_sents": 345, "n_sents_1": 238, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1745, "title": "Massively Multi-Agents Reveal That Large Language Models Can Understand Value", "abs": "Large Language Models (LLMs) have been trained on vast corpora of data, allowing them to learn internal representations of how humans would respond in different scenarios. This makes them well-suited to simulate the actions of market participants, to model their collective impact on financial markets and perform financial forecasting. However, there also exist various sources of errors that could affect the effectiveness of LLM agent-based simulations of the market. Firstly, individual market participants do not always make rational decisions, which might not be captured by the logical reasoning process of LLMs. Secondly, the numerical and financial literacy of LLMs are also not highly reliable, due to possible knowledge gaps in their numerical understanding and possible hallucinations in their outputs. To tackle these issues, we propose our Massively Multi-Agents Role Playing (MMARP) method, which aims to produce highly accurate market simulations through theory-driven prompt designs. To reduce the impact of noisy actions caused by individual irrational investors, we leverage the LLM-generated next-token weights to simulate repetitive prompting, and obtain the aggregated market response. To minimize the effects of possible gaps in its numerical knowledge or potential hallucinated outputs, we prompt the LLM using a range of price inputs for each trading day. Finally, to produce simulated forecasts of market prices, we perform the above prompting strategies across two types of LLM-agent roles, buyers and sellers, and obtain the intersection price between their response curves. Through experimental results, we show that MMARP can outperform other deep-learning methods and various financial LLMs in forecasting metrics.", "keywords": [ "computational finance", "stock prediction", "large language models", "economics" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 2, 3, 1, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ob9vuDv4yl", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Deyu Meng", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Jin Cao", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Li Pang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xiangyong Cao", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yi Cao", "gender": "Male", "institution": "Beijing Computational Science Research Center", "country": "CN", "position": "MS student" } ], "n_formula": 19, "n_formula_1": 14, "n_ref_uni": 44, "n_ref": 129, "n_ref_all": 153, "n_fig": 8, "n_tab": 9, "L_tab": 4021, "n_element_tab": 272, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2558, "n_element_tab_1": 78, "formula_len_all": 1049, "formula_len_all_1": 836, "len_all": 171241, "len_all_1": 77900, "len_abs": 1736, "len_title": 103, "len_sents": 46373, "len_sents_1": 29206, "n_sents": 355, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 1760, "title": "HAIR: Hypernetworks-based All-in-One Image Restoration", "abs": "Image restoration aims to recover a high-quality clean image from its degraded version. Recent progress in image restoration has demonstrated the effectiveness of All-in-One image restoration models in addressing various unknown degradations simultaneously. However, these existing methods typically utilize the same parameters to tackle images with different types of degradation, forcing the model to balance the performance between different tasks and limiting its performance on each task. To alleviate this issue, we propose HAIR, a **H**ypernetworks-based **A**ll-in-One **I**mage **R**estoration plug-and-play method that generates parameters based on the input image and thus makes the model to adapt to specific degradation dynamically. Specifically, HAIR consists of two main components, i.e., Classifier and Hyper Selecting Net (HSN). The Classifier is a simple image classification network used to generate a Global Information Vector (GIV) that contains the degradation information of the input image, and the HSN is a simple fully-connected neural network that receives the GIV and outputs parameters for the corresponding modules. Extensive experiments demonstrate that HAIR can significantly improve the performance of existing image restoration models in a plug-and-play manner, both in single-task and All-in-One settings. Notably, our proposed model Res-HAIR, which integrates HAIR into the well-known Restormer, can obtain superior or comparable performance compared with current state-of-the-art methods. Moreover, we theoretically demonstrate that to achieve a given small enough error, our proposed HAIR requires fewer parameters in contrast to mainstream embedding-based All-in-One methods. Code is available in supplementary materials.", "keywords": [ "All-in-One image restoration", "Hypernetworks", "low-level vision" ], "rating_list": [ 3, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 4, 5 ] }, { "paper_id": "ob7UrZOJve", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Alex Dimakis", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department, University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Sunny Sanyal", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" }, { "name": "ravid ziv", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "sujay sanghavi", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 6, "n_ref_uni": 35, "n_ref": 91, "n_ref_all": 177, "n_fig": 24, "n_tab": 10, "L_tab": 2493, "n_element_tab": 316, "n_fig_1": 16, "n_tab_1": 3, "L_tab_1": 820, "n_element_tab_1": 162, "formula_len_all": 492, "formula_len_all_1": 244, "len_all": 214842, "len_all_1": 67000, "len_abs": 1332, "len_title": 112, "len_sents": 64554, "len_sents_1": 30488, "n_sents": 499, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1341, "title": "Inheritune: Training Smaller Yet More Attentive Language Models", "abs": "Large Language Models (LLMs) have achieved remarkable performance across various natural language processing tasks, primarily due to the transformer architecture and its self-attention mechanism. However, we observe that in standard decoder-style LLMs attention matrices degenerate to single-column for deeper layers. Layers in this state unable to learn anything meaningful and mostly redundant; we refer to these as lazy layers. The goal of this paper is to train smaller models by eliminating this structural inefficiency without compromising performance.\n\nMotivated by this observation, we propose Inheritune, a simple yet effective training recipe for developing smaller, high-performing language models. Smaller models trained with Inheritune inherits early transformer layers from a larger pre-trained model, then retrains and progressively expands the smaller model until it matches or exceeds the performance of the larger model. We demonstrate that Inheritune enables the training of various sizes of GPT-2 models on datasets like OpenWebText-9B and FineWeb\\_Edu. Models trained with Inheritune, despite having significantly fewer layers, match or even surpass the performance of their larger counterparts. For instance, our 16-layer GPT-2 medium variant achieves comparable performance to the standard 24-layer GPT-2 medium model.", "keywords": [ "Large Language Models", "Small Language Models", "Attention degeneration", "Efficient training", "Model Initialization" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "oaRaaG1WB1", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Chengtao Jian", "gender": "Male", "institution": "Tongji University", "country": "", "position": "PhD student" }, { "name": "Kai Yang", "gender": "Male", "institution": "Tongji University, Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Yang Jiao", "gender": "unknown", "institution": "Tongji University", "country": "", "position": "PhD student" } ], "n_formula": 153, "n_formula_1": 46, "n_ref_uni": 83, "n_ref": 204, "n_ref_all": 316, "n_fig": 5, "n_tab": 4, "L_tab": 922, "n_element_tab": 95, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 414, "n_element_tab_1": 6, "formula_len_all": 31565, "formula_len_all_1": 6886, "len_all": 298305, "len_all_1": 69361, "len_abs": 1432, "len_title": 180, "len_sents": 70022, "len_sents_1": 24606, "n_sents": 542, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 132, "L_abs": 1448, "title": "Unlocking Trilevel Learning with Level-Wise Zeroth Order Constraints: Distributed Algorithms and Provable Non-Asymptotic Convergence", "abs": "Trilevel learning (TLL) found diverse applications in numerous machine learning applications, ranging from robust hyperparameter optimization to domain adaptation. However, existing researches primarily focus on scenarios where TLL can be addressed with first order information available at each level, which is inadequate in many situations involving zeroth order constraints, such as when black-box models are employed. Moreover, in trilevel learning, data may be distributed across various nodes, necessitating strategies to address TLL problems without centralizing data on servers to uphold data privacy. To this end, an effective distributed trilevel zeroth order learning framework DTZO is proposed in this work to address the TLL problems with level-wise zeroth order constraints in a distributed manner. The proposed DTZO is versatile and can be adapted to a wide range of (grey-box) TLL problems with partial zeroth order constraints. In DTZO, the cascaded polynomial approximation can be constructed without relying on gradients or sub-gradients, leveraging a novel cut, i.e., zeroth order cut. Furthermore, we theoretically carry out the non-asymptotic convergence rate analysis for the proposed DTZO in achieving the $\\epsilon$-stationary point. Extensive experiments have been conducted to demonstrate and validate the superior performance of the proposed DTZO, e.g., it approximately achieves up to a 40\\% improvement in performance.", "keywords": [ "Trilevel Optimization", "Distributed Optimization", "Zeroth Order Optimization" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 3, 4, 2 ] }, { "paper_id": "oa5UeyUVMm", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 9, "author_info_list": [ { "name": "Dingshuo Chen", "gender": "Male", "institution": "Institute of automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Liang Wang", "gender": "Male", "institution": "Institute of Automation\uff0c CAS\uff0cChina", "country": "CN", "position": "Full Professor" }, { "name": "Liuji Chen", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Qiang Liu", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Shu Wu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Shuchen Xue", "gender": "Male", "institution": "Academy of Mathematics and Systems Science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Tieniu Tan", "gender": "Male", "institution": " Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Professor" }, { "name": "Yingheng Wang", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Zhi-Ming Ma", "gender": "unknown", "institution": "Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 78, "n_formula_1": 18, "n_ref_uni": 54, "n_ref": 113, "n_ref_all": 124, "n_fig": 4, "n_tab": 6, "L_tab": 2645, "n_element_tab": 332, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2664, "n_element_tab_1": 253, "formula_len_all": 8348, "formula_len_all_1": 948, "len_all": 166243, "len_all_1": 66697, "len_abs": 1375, "len_title": 133, "len_sents": 45023, "len_sents_1": 29742, "n_sents": 335, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1385, "title": "Graffe: Graph Representation Learning Enabled via Diffusion Probabilistic Models", "abs": "Diffusion probabilistic models (DPMs), widely recognized for their potential to generate high-quality samples, tend to go unnoticed in representation learning. While recent progress has highlighted their potential for capturing visual semantics, adapting DPMs to graph representation learning remains in its infancy. In this paper, we introduce **Graffe**, a self-supervised diffusion model proposed for graph representation learning. It features a graph encoder that distills a source graph into a compact representation, which, in turn, serves as the condition to guide the denoising process of the diffusion decoder. To evaluate the effectiveness of our model, we first explore the theoretical foundations of applying diffusion models to representation learning, proving that the denoising objective implicitly maximizes the conditional mutual information between data and its representation. Specifically, we prove that the negative logarithm of denoising score matching loss is a tractable lower bound for the conditional mutual information. Empirically, Graffe delivers competitive results under the linear probing setting on node and graph classification, achieving state-of-the-art performance on 9 of the 11 real-world datasets. These findings indicate that powerful generative models, especially diffusion models, serve as an effective tool for graph representation learning.", "keywords": [ "graph representation learning", "diffusion models", "unsupervised learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "oZkqkkvdND", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Changming Xu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Debangshu Banerjee", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Deepak Vasisht", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Gagandeep Singh", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 14, "n_ref_uni": 21, "n_ref": 49, "n_ref_all": 61, "n_fig": 2, "n_tab": 7, "L_tab": 1412, "n_element_tab": 188, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2853, "n_element_tab_1": 206, "formula_len_all": 3886, "formula_len_all_1": 2058, "len_all": 156433, "len_all_1": 71703, "len_abs": 3774, "len_title": 50, "len_sents": 43537, "len_sents_1": 30917, "n_sents": 357, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 805, "title": "Support is All You Need for Certified VAE Training", "abs": "Variational Autoencoders (VAEs) have become increasingly popular and deployed in safety-critical applications. In such applications, we want to give certified probabilistic guarantees on performance under adversarial attacks. We propose a novel method, CIVET, for certified training of VAEs. CIVET depends on the key insight that we can bound worst-case VAE error by bounding the error on carefully chosen support sets at the latent layer. We show this point mathematically and present a novel training algorithm utilizing this insight. We show in an extensive evaluation across different datasets (in both the wireless and vision application areas), architectures, and perturbation magnitudes that our method outperforms SOTA methods achieving good standard performance with strong robustness guarantees.", "keywords": [ "Certified Training", "Trustworthy Machine Learning", "Variational Autoencoder", "Wireless" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "oZdaEiDBpF", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Dan Roth", "gender": "Male", "institution": "Oracle", "country": "US", "position": "Chief Scientist" }, { "name": "Efthymia Tsamoura", "gender": "Female", "institution": "Samsung AI", "country": "GB", "position": "Researcher" }, { "name": "Kaifu Wang", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" } ], "n_formula": 53, "n_formula_1": 20, "n_ref_uni": 56, "n_ref": 240, "n_ref_all": 308, "n_fig": 7, "n_tab": 8, "L_tab": 838, "n_element_tab": 45, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 1, "n_element_tab_1": 1, "formula_len_all": 4586, "formula_len_all_1": 2032, "len_all": 232963, "len_all_1": 80599, "len_abs": 2588, "len_title": 132, "len_sents": 73597, "len_sents_1": 36132, "n_sents": 673, "n_sents_1": 314, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1606, "title": "On Characterizing and Mitigating Imbalances in Multi-Instance Partial Label Learning", "abs": "\\textit{Multi-Instance Partial Label Learning} (MI-PLL) is a weakly-supervised learning setting encompassing \\textit{partial label learning}, \\textit{latent structural learning}, and \\textit{neurosymbolic learning}. Unlike supervised learning, in MI-PLL, the inputs to the classifiers at training-time are tuples of instances $\\mathbf{x}$. At the same time, the supervision signal is generated by a function $\\sigma$ over the (hidden) gold labels of $\\mathbf{x}$. In this work, we make multiple contributions towards addressing a problem that hasn\u2019t been studied so far in the context of MI-PLL: that of characterizing and mitigating \\textit{learning imbalances}, i.e., major differences in the errors occurring when classifying instances of different classes (aka \\emph{class-specific risks}). In terms of theory, we derive class-specific risk bounds for MI-PLL, while making minimal assumptions. Our theory reveals a unique phenomenon: that $\\sigma$ can greatly impact learning imbalances. This result is in sharp contrast with previous research on supervised and weakly-supervised learning, which only studies learning imbalances under the prism of data imbalances. On the practical side, we introduce a technique for estimating the marginal of the hidden labels using only MI-PLL data. Then, we introduce algorithms that mitigate imbalances at training- and testing-time, by treating the marginal of the hidden labels as a constraint. We demonstrate the effectiveness of our techniques using strong baselines from neurosymbolic and long-tail learning, suggesting performance improvements of up to 14\\%.", "keywords": [ "muliti-instance partial label learning", "weakly-supervised learning", "neurosymbolic learning", "learning theory", "long-tailed learning", "learning imbalances", "class-specific error bounds" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "oYemKnlIrO", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Blake Bordelon", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Cengiz Pehlevan", "gender": "unknown", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "Tanishq Kumar", "gender": "Male", "institution": "Harvard College", "country": "", "position": "Undergrad student" }, { "name": "Venkatesh N Murthy", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Full Professor" }, { "name": "Samuel Gershman", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Professor" } ], "n_formula": 8, "n_formula_1": 3, "n_ref_uni": 71, "n_ref": 113, "n_ref_all": 137, "n_fig": 11, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 489, "formula_len_all_1": 141, "len_all": 161772, "len_all_1": 57680, "len_abs": 1363, "len_title": 107, "len_sents": 50578, "len_sents_1": 29793, "n_sents": 316, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1366, "title": "Do Mice Grok? Glimpses of Hidden Progress in Sensory Cortex", "abs": "Does learning of task-relevant representations stop when behavior stops changing? Motivated by recent work in machine learning and the intuitive observation that human experts continue to learn after mastery, we hypothesize that task-specific representation learning in cortex can continue, even when behavior saturates. In a novel reanalysis of recently published neural data, we find evidence for such learning in posterior piriform cortex of mice following continued training on a task, long after behavior saturates at near-ceiling performance (\"overtraining\"). We demonstrate that class representations in cortex continue to separate during overtraining, so that examples that were incorrectly classified at the beginning of overtraining can abruptly be correctly classified later on, despite no changes in behavior during that time. We hypothesize this hidden learning takes the form of approximate margin maximization; we validate this and other predictions in the neural data, as well as build and interpret a simple synthetic model that recapitulates these phenomena. We conclude by demonstrating how this model of late-time feature learning implies an explanation for the empirical puzzle of overtraining reversal in animal learning, where task-specific representations are more robust to particular task changes because the learned features can be reused.", "keywords": [ "neuroscience; representation learning; grokking; overtraining; cortex" ], "rating_list": [ 5, 8, 5, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "oYaP4XPWet", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "Thomas Crasson", "gender": "Male", "institution": "Ecole Nationale Sup\u00e9rieure de Techniques Avanc\u00e9es", "country": "FR", "position": "MS student" }, { "name": "Yacine Nabet", "gender": "Male", "institution": "Wiremind", "country": "FR", "position": "Researcher" }, { "name": "Mathias L\u00e9cuyer", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 41, "n_formula_1": 21, "n_ref_uni": 34, "n_ref": 96, "n_ref_all": 115, "n_fig": 11, "n_tab": 9, "L_tab": 1761, "n_element_tab": 57, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 222, "n_element_tab_1": 6, "formula_len_all": 4313, "formula_len_all_1": 1774, "len_all": 149680, "len_all_1": 71729, "len_abs": 736, "len_title": 113, "len_sents": 49791, "len_sents_1": 32061, "n_sents": 429, "n_sents_1": 273, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 738, "title": "Training and Evaluating Causal Forecasting Models for Time-Series", "abs": "Deep learning time-series models are often used to make forecasts that inform downstream decisions.\nSince these decisions can differ from those in the training set, there is an implicit requirement that time-series models will generalize outside of their training distribution.\nDespite this core requirement, time-series models are typically trained and evaluated on in-distribution predictive tasks.\nWe extend the orthogonal statistical learning framework to train causal time-series models that generalize better when forecasting the effect of actions outside of their training distribution.\nTo evaluate these models, we leverage Regression Discontinuity Designs popular in economics to construct a test set of causal treatment effects.", "keywords": [ "Time-series forecasting; Causal Inference; Regression Discontinuity Designs; Deep Learning" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "oYSsbY3G4o", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 10, "author_info_list": [ { "name": "Baichuan Yuan", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Bo Long", "gender": "Male", "institution": "Meta", "country": "US", "position": "Principal Researcher" }, { "name": "Dongqi Fu", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Hao Wu", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Kaveh Hassani", "gender": "Male", "institution": "Meta", "country": "CA", "position": "AI Research Scientist" }, { "name": "Limei Wang", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Ning Yao", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Si Zhang", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Weilin Cong", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Zhigang Hua", "gender": "Male", "institution": "Meta", "country": "US", "position": "ML engineer" } ], "n_formula": 26, "n_formula_1": 25, "n_ref_uni": 81, "n_ref": 193, "n_ref_all": 209, "n_fig": 1, "n_tab": 8, "L_tab": 5106, "n_element_tab": 627, "n_fig_1": 0, "n_tab_1": 6, "L_tab_1": 5395, "n_element_tab_1": 300, "formula_len_all": 1366, "formula_len_all_1": 1241, "len_all": 226633, "len_all_1": 67212, "len_abs": 1102, "len_title": 100, "len_sents": 46051, "len_sents_1": 27189, "n_sents": 306, "n_sents_1": 153, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1106, "title": "Learning Graph Quantized Tokenizers for Transformers", "abs": "Transformers serve as the backbone architectures of Foundational Models, where a domain-specific tokenizer helps them adapt to various domains. Graph Transformers (GTs) have recently emerged as a leading model in geometric deep learning, outperforming Graph Neural Networks (GNNs) in various graph learning tasks. However, the development of tokenizers for graphs has lagged behind other modalities, with existing approaches relying on heuristics or GNNs co-trained with Transformers. To address this, we introduce GQT (Graph Quantized Tokenizer), which decouples tokenizer training from Transformer training by leveraging multi-task graph self-supervised learning, yielding robust and generalizable graph tokens. Furthermore, the GQT utilizes Residual Vector Quantization (RVQ) to learn hierarchical discrete tokens, resulting in significantly reduced memory requirements and improved generalization capabilities. By combining the GQT with token modulation, a Transformer encoder achieves state-of-the-art performance on 16 out of 18 benchmarks, including large-scale homophilic and heterophilic datasets.", "keywords": [ "Graph Learning", "Graph Tokenization", "Graph Transformer" ], "rating_list": [ 8, 5, 6, 5, 8 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 4, 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 3, 4, 3 ] }, { "paper_id": "oYLayGfWcI", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 12, "author_info_list": [ { "name": "Anurag Kumar", "gender": "Male", "institution": "Meta", "country": "US", "position": "Research Scientist" }, { "name": "Bowen Shi", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Brian Ellis", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "David Kant", "gender": "unknown", "institution": "University of California, Santa Cruz", "country": "", "position": "PhD student" }, { "name": "Ernie Chang", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Research Scientist" }, { "name": "Gael Le Lan", "gender": "Male", "institution": "Meta", "country": "FR", "position": "Researcher" }, { "name": "Sidd Srinivasan", "gender": "Male", "institution": "Department of Computer Science, Cornell University", "country": "", "position": "M.Eng." }, { "name": "Vikas Chandra", "gender": "Male", "institution": "Meta", "country": "US", "position": "Director, AI" }, { "name": "Yangyang Shi", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Zhaoheng Ni", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Varun Nagaraja", "gender": "Male", "institution": "Meta", "country": "US", "position": "Research Scientist" }, { "name": "Wei-Ning Hsu", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 45, "n_ref": 121, "n_ref_all": 159, "n_fig": 15, "n_tab": 5, "L_tab": 1364, "n_element_tab": 370, "n_fig_1": 11, "n_tab_1": 5, "L_tab_1": 1908, "n_element_tab_1": 338, "formula_len_all": 877, "formula_len_all_1": 819, "len_all": 148476, "len_all_1": 64725, "len_abs": 1003, "len_title": 118, "len_sents": 40886, "len_sents_1": 27374, "n_sents": 299, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 997, "title": "High Fidelity Text-Guided Music Editing via Single-Stage Flow Matching", "abs": "We introduce MelodyFlow, an efficient text-controllable high-fidelity music generation and editing model. It operates on continuous latent representations from a low frame rate 48 kHz stereo variational auto encoder codec. Based on a diffusion transformer architecture trained on a flow-matching objective the model can edit diverse high quality stereo samples of variable duration, with simple text descriptions. We adapt the ReNoise latent inversion method to flow matching and compare it with the original implementation and naive denoising diffusion implicit model (DDIM) inversion on a variety of music editing prompts. Our results indicate that our latent inversion outperforms both ReNoise and DDIM for zero-shot test-time text-guided editing on several objective metrics. Subjective evaluations exhibit a substantial improvement over previous state of the art for music editing. Code and model weights will be publicly made available. Samples are available at https://melodyflow.github.io.", "keywords": [ "Flow matching", "music generation", "style transfer" ], "rating_list": [ 3, 5, 6, 3, 3 ], "soundness_list": [ 2, 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 3 ] }, { "paper_id": "oY2jw2NLiM", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Dan Feldman", "gender": "Male", "institution": "University of Haifa", "country": "", "position": "Associate Professor" }, { "name": "David Denisov", "gender": "Male", "institution": "University of Haifa", "country": "IL", "position": "PhD student" }, { "name": "Michael Segal", "gender": "unknown", "institution": "Ben-Gurion University of the Negev", "country": "", "position": "Full Professor" }, { "name": "Shlomi Dolev", "gender": "Male", "institution": "Ben-Gurion University of the Negev", "country": "", "position": "Full Professor" } ], "n_formula": 109, "n_formula_1": 20, "n_ref_uni": 25, "n_ref": 69, "n_ref_all": 82, "n_fig": 8, "n_tab": 2, "L_tab": 95, "n_element_tab": 1, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 60, "n_element_tab_1": 5, "formula_len_all": 9150, "formula_len_all_1": 1602, "len_all": 168380, "len_all_1": 58136, "len_abs": 1279, "len_title": 90, "len_sents": 56462, "len_sents_1": 24870, "n_sents": 546, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1514, "title": "Coresets for $k$-mean clustering of segments", "abs": "The $k$-means of a given set $\\mathcal{S}\\subseteq \\mathbb{R}^d$ of $n$ segments is a set $X\\subseteq \\mathbb{R}^d$ of $|X|=k$ centers which minimizes their sum of squared distances $D(\\mathcal{S},X):=\\sum_{S\\in \\mathcal{S}}\\min_{x\\in X}D(S,x)$.\nHere, the distance $D(S,x)$ between a segment $S$ and a point $x$ is the integral of its distances $\\int_{s\\in S}\\|p-x\\|$ over each point on the segment.\nMore generally, the farthest $m$ input points (outliers) may be ignored, other distance functions may be used, such as M-estimator or non-squared, and each distance may be multiplied by a function that depends on the size of its cluster, say, to obtain balanced clustering.\nFor a given $\\varepsilon>0$, an $\\varepsilon$-coreset $C\\subseteq S$ for all these problems is a weighted subset $C\\subset S$, that approximates $D(S,X)$ up to $1\\pm\\varepsilon$ multiplicative factor for every set $X\\subseteq\\mathbb{R}^d$ of (possibly weighted) $k$ centers. Such a coreset enables handling streaming, big, distributed input in parallel using existing techniques.\nWe suggest the first coreset construction that, with high probability, returns an $\\varepsilon$-coreset $C$ for \\emph{any} input set $\\mathcal{S}$ of segments.\nFor constant $k,\\varepsilon$, the size of the coreset is $|C|\\in O \\big(\\log^2(n)\\big)$ and is computed in time $O(nd)$.\nExperimental results and real-time video tracking application demonstrate the applicability of our algorithm, the latter demonstrates that our method supports vectorized segments.", "keywords": [ "Clustering; $k$-means; Segment clustering; Non-convex optimisation; Coresets" ], "rating_list": [ 3, 5, 3, 1 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "oXh0939Zzq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Jing Lin", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Liuyang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Weizhong Huang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Xiawu Zheng", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Yiwu Yao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yuxin Zhang", "gender": "unknown", "institution": "Xiamen University", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 60, "n_ref": 146, "n_ref_all": 176, "n_fig": 2, "n_tab": 20, "L_tab": 6292, "n_element_tab": 799, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 3619, "n_element_tab_1": 355, "formula_len_all": 595, "formula_len_all_1": 549, "len_all": 201110, "len_all_1": 69107, "len_abs": 1843, "len_title": 108, "len_sents": 48905, "len_sents_1": 28014, "n_sents": 367, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 2053, "title": "Dynamic Low-Rank Sparse Adaptation for Large Language Models", "abs": "Despite the efficacy of network sparsity in alleviating the deployment strain of Large Language Models (LLMs), it endures significant performance degradation. Applying Low-Rank Adaptation (LoRA) to fine-tune the sparse LLMs offers an intuitive approach to counter this predicament, while it holds shortcomings include: 1) The inability to integrate LoRA weights into sparse LLMs post-training, and 2) Insufficient performance recovery at high sparsity ratios. In this paper, we introduces dynamic $\\textbf{Lo}$w-rank $\\textbf{S}$parse $\\textbf{A}$daptation $\\textbf{(LoSA)}$, a novel method that seamlessly integrates low-rank adaptation into LLM sparsity within a unified framework, thereby enhancing the performance of sparse LLMs without increasing the inference latency. In particular, LoSA dynamically sparsifies the LoRA outcomes based on the corresponding sparse weights during fine-tuning, thus guaranteeing that the LoRA module can be integrated into the sparse LLMs post-training. Besides, to achieve the optimal sparse model architecture, LoSA leverages Representation Mutual Information (RMI) as an indicator to determine the importance of layers, thereby dynamically determining the optimal layer-wise sparsity rates during fine-tuning. Predicated on this, LoSA adjusts the rank of the LoRA module based on the variability in layer-wise reconstruction errors, allocating an appropriate fine-tuning for each layer to reduce the output discrepancies between dense and sparse LLMs. Extensive experiments tell that LoSA can efficiently boost the efficacy of sparse LLMs within a few hours, without introducing any additional inferential burden. For example, LoSA reduced the perplexity of sparse LLaMA-2-7B by $\\textbf{68.73}$$\\downarrow$ and increased zero-shot accuracy by $\\textbf{16.32}$%$\\uparrow$, achieving a $\\textbf{2.60$\\times$}$ speedup on CPU and $\\textbf{2.23$\\times$}$ speedup on GPU, requiring only $\\textbf{45 minutes}$ of fine-tuning on $\\textbf{a single}$ NVIDIA A100 80GB GPU. Code is available in the supplementary material.", "keywords": [ "Large Language Models; Network Sparsity; Low-Rank Adaptation" ], "rating_list": [ 3, 3, 6, 6, 8 ], "soundness_list": [ 3, 1, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "oX4FcNA4UC", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Eric Vanden-Eijnden", "gender": "Male", "institution": "New York University", "country": "", "position": "Full Professor" }, { "name": "Mathieu Lauriere", "gender": "unknown", "institution": "New York University", "country": "FR", "position": "Assistant Professor" }, { "name": "Mengjian Hua", "gender": "Male", "institution": "Courant Institute of Mathematical Sciences, New York University", "country": "", "position": "PhD student" } ], "n_formula": 48, "n_formula_1": 34, "n_ref_uni": 41, "n_ref": 74, "n_ref_all": 103, "n_fig": 4, "n_tab": 4, "L_tab": 144, "n_element_tab": 12, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 144, "n_element_tab_1": 12, "formula_len_all": 4778, "formula_len_all_1": 3411, "len_all": 114333, "len_all_1": 60880, "len_abs": 1041, "len_title": 118, "len_sents": 31159, "len_sents_1": 24927, "n_sents": 250, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1048, "title": "A Simulation-Free Deep Learning Approach to Stochastic Optimal Control", "abs": "We propose a simulation-free algorithm for the solution of generic problems in stochastic optimal control (SOC). Unlike existing methods, our approach does not require the solution of an adjoint problem, but rather leverages Girsanov theorem to directly calculate the gradient of the SOC objective on-policy. This allows us to speed up the optimization of control policies parameterized by neural networks since it completely avoids the expensive back-propagation step through stochastic differential equations (SDEs) used in the Neural SDE framework. In particular, it enables us to solve SOC problems in high dimension and on long time horizons. We demonstrate the efficiency of our approach in various domains of applications, including standard stochastic optimal control problems, sampling from unnormalized distributions via construction of a Schr\\\"odinger-F\\\"ollmer process, and fine-tuning of pre-trained diffusion models. In all cases our method is shown to outperform the existing methods in both the computing time and memory efficiency.", "keywords": [ "Stochastic Optimal Control", "Simulation-Free Methods" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "oWy06SBgt4", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 6, "author_info_list": [ { "name": "Chang Gao", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Jianfei Chen", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Jiaqi Wang", "gender": "Male", "institution": "Tencent YouTu Lab", "country": "CN", "position": "Researcher" }, { "name": "JingRen Hou", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Kang Zhao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "liping jing", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 87, "n_formula_1": 34, "n_ref_uni": 45, "n_ref": 96, "n_ref_all": 124, "n_fig": 9, "n_tab": 8, "L_tab": 2990, "n_element_tab": 239, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1870, "n_element_tab_1": 116, "formula_len_all": 9078, "formula_len_all_1": 2652, "len_all": 177830, "len_all_1": 70513, "len_abs": 1294, "len_title": 113, "len_sents": 46772, "len_sents_1": 27525, "n_sents": 489, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1255, "title": "1-Bit FQT: Pushing the Limit of Fully Quantized Training to 1-bit", "abs": "Fully quantized training (FQT) accelerates the training of deep neural networks by quantizing the activations, weights, and gradients into lower precision. To explore the ultimate limit of FQT (the lowest achievable precision), we make a first attempt to 1-bit FQT. We provide a theoretical analysis of FQT based on Adam and SGD, revealing that the gradient variance influences the convergence of FQT. Building on these theoretical results, we introduce an Average 1-bit Quantization (AQ) strategy. The strategy leverages the heterogeneity of gradients to mitigate gradient variance by pruning less informative gradients and enhancing the numerical precision of remaining gradients. Additionally, we propose Sample Channel joint Quantization (SCQ), which utilizes different quantization strategies in the computation of weight gradients and activation gradients to ensure that the method is friendly to low-bitwidth hardware. Finally, we present a framework to deploy our algorithm. For fine-tuning VGGNet-16 and ResNet-18 on multiple datasets, our algorithm achieves an average accuracy improvement of approximately 6\\%, compared to per-sample quantization. Moreover, our training speedup can reach a maximum of 5.13\u00d7 compared to full precision training.", "keywords": [ "efficient machine learning", "quantization methods", "efficient training algorithms", "fully quantized training" ], "rating_list": [ 5, 6, 8, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "oWm80iR1m9", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Anbang Ye", "gender": "Male", "institution": "HPC-AI Tech", "country": "SG", "position": "Principal Researcher" }, { "name": "Fujiao Liu", "gender": "Male", "institution": "Grab Holding Inc.", "country": "SG", "position": "Data Scientist" }, { "name": "Jia Chen", "gender": "Male", "institution": "Grabtaxi Holdings Pte Ltd", "country": "SG", "position": "Data Scientist" }, { "name": "Qianran Ma", "gender": "unknown", "institution": "HPC AI TECHNOLOGY PTE. LTD.", "country": "SG", "position": "Researcher" }, { "name": "Tong Li", "gender": "Male", "institution": "HPC-AI Tech", "country": "SG", "position": "Principal Researcher" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 73, "n_ref_all": 91, "n_fig": 4, "n_tab": 8, "L_tab": 2221, "n_element_tab": 199, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3158, "n_element_tab_1": 133, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 165767, "len_all_1": 62885, "len_abs": 1363, "len_title": 117, "len_sents": 50116, "len_sents_1": 28780, "n_sents": 312, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1368, "title": "SOP-Agent: Empower General Purpose AI Agent with Domain-Specific SOPs", "abs": "Despite significant advancements in general-purpose AI agents, several challenges still hinder their practical application in real-world scenarios. First, the limited planning capabilities of Large Language Models (LLM) restrict AI agents from effectively solving complex tasks that require long-horizon planning (Liu et al. 2023). Second, general-purpose AI agents struggle to efficiently utilize domain-specific knowledge and human expertise. In this paper, we introduce the Standard Operational Procedure-guided Agent (SOP-agent), a novel framework for constructing domain-specific agents through pseudocode-style Standard Operational Procedures (SOPs) written in natural language. Formally, we represent a SOP as a decision graph, which is traversed to guide the agent in completing tasks specified by the SOP. We conduct extensive experiments across tasks in multiple domains, including decision-making, search and reasoning, code generation, data cleaning, and grounded customer service. The SOP-agent demonstrates excellent versatility, achieving performance superior to general-purpose agent frameworks and comparable to domain-specific agent systems. Additionally, we introduce the Grounded Customer Service Benchmark, the first benchmark designed to evaluate the grounded decision-making capabilities of AI agents in customer service scenarios based on SOPs.", "keywords": [ "autonomous agent", "grounded decision-making" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "oWdzUpOlkX", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "George Karypis", "gender": "Male", "institution": "University of Minnesota, Minneapolis", "country": "", "position": "Full Professor" }, { "name": "Huzefa Rangwala", "gender": "Male", "institution": "Amazon", "country": "", "position": "Principal Researcher" }, { "name": "Ke Yang", "gender": "Female", "institution": "Department of Computer Science", "country": "US", "position": "PhD student" }, { "name": "Pratik A Chaudhari", "gender": "Male", "institution": "School of Engineering and Applied Science, University of Pennsylvania", "country": "US", "position": "Assistant Professor" }, { "name": "Rasool Fakoor", "gender": "Male", "institution": "Amazon Web Services", "country": "", "position": "Researcher" }, { "name": "Sapana Chaudhary", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yao Liu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 10, "n_ref": 24, "n_ref_all": 63, "n_fig": 6, "n_tab": 18, "L_tab": 10337, "n_element_tab": 1251, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 2568, "n_element_tab_1": 168, "formula_len_all": 531, "formula_len_all_1": 39, "len_all": 287778, "len_all_1": 70223, "len_abs": 2361, "len_title": 113, "len_sents": 109169, "len_sents_1": 33085, "n_sents": 1019, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 2262, "title": "AgentOccam: A Simple Yet Strong Baseline for LLM-Based Web Agents", "abs": "Autonomy via agents based on large language models (LLMs) that can carry out personalized yet standardized tasks presents a significant opportunity to drive human efficiency. There is an emerging need and interest in automating web tasks (e.g., booking a hotel for a given date within a budget). Being a practical use case itself, the web agent also serves as an important proof-of-concept example for various agent grounding scenarios, with its success promising advancements in many future applications. Meanwhile, much prior research focuses on handcrafting their web agent strategies (e.g. agent's prompting templates, reflective workflow, role-play and multi-agent systems, search or sampling methods, etc.) and the corresponding in-context examples. However, these custom strategies often struggle with generalizability across all potential real-world applications. On the other hand, there has been limited study on the misalignment between a web agent's observation and action representation, and the data on which the agent's underlying LLM has been pre-trained. This is especially notable when LLMs are primarily trained for language completion rather than tasks involving embodied navigation actions and symbolic web elements. In our study, we enhance an LLM-based web agent by simply refining its observation and action space, aligning these more closely with the LLM's capabilities. This approach enables our base agent to significantly outperform previous methods on a wide variety of web tasks. Specifically, on WebArena, a benchmark featuring general-purpose web interaction tasks, our agent AgentOccam surpasses the previous state-of-the-art and concurrent work by 9.8 (+29.4%) and 5.9 (+15.8%) absolute points respectively, and boosts the success rate by 26.6 points (+161%) over similar plain web agents with its observation and action space alignment. Furthermore, on the WebVoyager benchmark comprising tasks defined on real-world websites, AgentOccam exceeds the former best agent by 2.4 points (+4.6%) on tasks with deterministic answers. AgentOccam's simple design highlights the LLMs' impressive zero-shot performance in web tasks, and underlines the critical role of carefully tuning observation and action spaces for LLM-based agents.", "keywords": [ "LLM", "Agent", "LLM-based Agent", "Web Agent", "Web Navigation" ], "rating_list": [ 6, 5, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 4 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "oW7jUu4gVg", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Jin Sun", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Assistant Professor" }, { "name": "Mengnan Du", "gender": "unknown", "institution": "New Jersey Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Qiaoyu Tan", "gender": "Male", "institution": "New York University Shanghai", "country": "CN", "position": "Assistant Professor" }, { "name": "Tianze Yang", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Xuansheng Wu", "gender": "unknown", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Yucheng Shi", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Ninghao Liu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 27, "n_ref": 46, "n_ref_all": 56, "n_fig": 4, "n_tab": 4, "L_tab": 951, "n_element_tab": 68, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 742, "n_element_tab_1": 38, "formula_len_all": 719, "formula_len_all_1": 413, "len_all": 116274, "len_all_1": 62168, "len_abs": 1101, "len_title": 128, "len_sents": 42223, "len_sents_1": 30343, "n_sents": 284, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1007, "title": "CORTEX: Concept-Oriented Token Explanation in Vector-Quantized Generative Model", "abs": "Vector-Quantized Generative Models (VQGMs) have emerged as powerful tools for image generation. However, the key component of VQGMs---the codebook of discrete tokens---is still not well understood, e.g., which tokens are critical to generate an image of a certain concept? This paper introduces Concept-Oriented Token Explanation (CORTEX), a novel approach for interpreting VQGMs by identifying concept-specific token combinations. Our framework employs two methods: (1) a saliency-based method that analyzes token saliency value in individual images, and (2) an optimization-based method that explores the entire codebook to find globally relevant tokens. Experimental results demonstrate CORTEX's efficacy in providing clear explanations of token usage in the generative process, outperforming baselines across multiple pretrained VQGMs. CORTEX not only improves VQGM transparency but also enables tasks such as targeted image editing, offering valuable insights into the model's internal representations.", "keywords": [ "Vector-Quantized Generative Model", "Explainability", "Information Bottleneck" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "oW7T3p5wE1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Huaibo Huang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Mingrui Chen", "gender": "Not Specified", "institution": "Institute of Automation, Chinese Academy of Sciences (CASIA)", "country": "CN", "position": "PhD student" }, { "name": "Qihang Fan", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Ran He", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 27, "n_ref": 91, "n_ref_all": 104, "n_fig": 7, "n_tab": 11, "L_tab": 7311, "n_element_tab": 726, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 4229, "n_element_tab_1": 250, "formula_len_all": 651, "formula_len_all_1": 499, "len_all": 140597, "len_all_1": 50480, "len_abs": 1746, "len_title": 126, "len_sents": 28912, "len_sents_1": 17595, "n_sents": 270, "n_sents_1": 164, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1785, "title": "Semantic Equitable Clustering: You Only Iterate Once to Cluster Vision Tokens", "abs": "The Vision Transformer (ViT) has gained prominence for its superior relational modeling prowess. However, its global attention mechanism's quadratic complexity poses substantial computational burdens. A common remedy spatially groups tokens for self-attention, reducing computational requirements. Nonetheless, this strategy neglects semantic information in tokens, possibly scattering semantically-linked tokens across distinct groups, thus compromising the efficacy of self-attention intended for modeling inter-token dependencies. Motivated by these insights, we introduce a fast and balanced clustering method, named \n **S**emantic **E**quitable **C**lustering (SEC). SEC clusters tokens based on their global semantic relevance in an efficient, straightforward manner. In contrast to traditional clustering methods requiring multiple iterations, our method achieves token clustering in a single pass. Additionally, SEC regulates the number of tokens per cluster, ensuring a balanced distribution for effective parallel processing on current computational platforms without necessitating further optimization. Capitalizing on SEC, we propose a versatile vision backbone, SECViT. Comprehensive experiments in image classification, object detection, instance segmentation, and semantic segmentation validate to the effectiveness of SECViT. Remarkably, SECViT attains an impressive **84.3%** image classification accuracy with only **27M** parameters and **4.6G** FLOPs, without the need for for additional supervision or data. Moreover, SEC can be conveniently and swiftly applied to multimodal large language models (MLLM), such as LLaVA, to serve as a vision language connector,\neffectively accelerating the model\u2019s efficiency while maintaining unchanged or better performance.", "keywords": [ "Equi-partition Clustering", "Efficient Vision Transformer", "multimodal large language models (MLLM)" ], "rating_list": [ 5, 3, 6, 6, 5 ], "soundness_list": [ 2, 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 4, 2 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "oW3XIIHaOn", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Hongsheng Wu", "gender": "unknown", "institution": "Anhui University", "country": "CN", "position": "Undergrad student" }, { "name": "Jie Chen", "gender": "Female", "institution": "Anhui University", "country": "CN", "position": "Associate Professor" }, { "name": "Minghan Zhang", "gender": "Male", "institution": "Anhui University", "country": "", "position": "PhD student" }, { "name": "Shu Zhao", "gender": "Female", "institution": "Anhui University", "country": "", "position": "Full Professor" }, { "name": "Yongxing Lin", "gender": "Male", "institution": "Anhui University", "country": "", "position": "Undergrad student" }, { "name": "Zhen Duan", "gender": "unknown", "institution": "Anhui University", "country": "CN", "position": "Lecturer" }, { "name": "Zhen Yang", "gender": "Female", "institution": "Anhui University", "country": "CN", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 31, "n_ref": 45, "n_ref_all": 55, "n_fig": 5, "n_tab": 2, "L_tab": 505, "n_element_tab": 93, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 505, "n_element_tab_1": 93, "formula_len_all": 22, "formula_len_all_1": 22, "len_all": 97417, "len_all_1": 56080, "len_abs": 1501, "len_title": 133, "len_sents": 27120, "len_sents_1": 26692, "n_sents": 175, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1507, "title": "ToG-I: Progressively Instructed Knowledge Graph-based Large Language Model Reasoning", "abs": "Large language models (LLMs) reasoning based on knowledge graphs (KGs), by integrating structured knowledge from the KGs, provide a significant solution to alleviate the hallucination problem in complex reasoning tasks. Current techniques mainly focus on the retrieval of explicit knowledge from KGs. LLMs directly use the specific facts and relationships retrieved to construct a reasoning chain to answer the questions. However, these methods often overlook the significance of comprehending implicit knowledge when dealing with problems involving logical reasoning or ambiguous intentions. This could potentially lead to deviations in the reasoning path, hindering their applicability in real-world applications. In this paper, we propose a progressive instructed reasoning framework, ToG-I. The framework identifies core elements, discerns latent intentions, and integrates necessary commonsense reasoning by analyzing the problem from multiple perspectives and levels. Based on this, ToG-I transforms these analysis results into specific reasoning instructions, guiding the LLMs to carry out a progressive reasoning process from a global perspective. This not only ensures the accuracy of the reasoning process but also effectively avoids unnecessary consumption of reasoning resources. Extensive experiments on multiple public datasets show that ToG-I achieves state-of-the-art performance in KG reasoning tasks based on information retrieval and demonstrates superiority in knowledge-intensive tasks.", "keywords": [ "LLM Reasoning\uff1bKnowledge Graph; Instructed\uff1bProgressively" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "oVwTvCI9Us", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "F\u00e9lix Lefebvre", "gender": "unknown", "institution": "INRIA", "country": "FR", "position": "PhD student" }, { "name": "Gael Varoquaux", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 4, "n_ref_uni": 58, "n_ref": 96, "n_ref_all": 138, "n_fig": 8, "n_tab": 10, "L_tab": 8116, "n_element_tab": 562, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2109, "n_element_tab_1": 97, "formula_len_all": 658, "formula_len_all_1": 195, "len_all": 183225, "len_all_1": 60441, "len_abs": 1415, "len_title": 57, "len_sents": 49852, "len_sents_1": 28149, "n_sents": 395, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1430, "title": "SEPAL: Scalable Feature Learning on Huge Knowledge Graphs", "abs": "Knowledge graphs accumulate information about more and more entities of the world. Much research is conducted to improve embedding models that capture this information and give useful node features in many downstream applications. However, most current methods are hard to scale to large knowledge graphs, partly because GPU memory is too small to hold the embeddings of many entities --YAGO4 has 67M entities. To scale existing embedding models on modest hardware, we introduce SEPAL: Scalable Embedding Propagation Algorithm for Large knowledge graphs.\nThe key idea of SEPAL to reduce compute is to only optimize embeddings on a core subset of entities, those that come with much more information than others. Then SEPAL propagates these embeddings to the rest of the graph with message passing, but no explicit optimization.\nTo enable efficient message passing, we break down large graphs into well-connected subgraphs that fit in GPU memory using a new algorithm called BLOCS: Balanced Local Overlapping Connected Subgraphs.\nWe evaluate SEPAL on five different knowledge graphs for four downstream regression tasks. We show that SEPAL outperforms alternative on downstream tasks, while providing a $43\\times$ speedup to its base embedding algorithm.\nMoreover, outside the core subgraph, embeddings obtained by message passing are not degraded compared to traditional methods, demonstrating the validity of SEPAL's propagation.", "keywords": [ "Knowledge graph", "Scalable", "Feature learning" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 2, 3 ] }, { "paper_id": "oVnfVnwh6y", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Dong Li", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Weigao Sun", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xuyang Shen", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Yiran Zhong", "gender": "Male", "institution": "MiniMax", "country": "CN", "position": "Principal Researcher" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhen Qin", "gender": "unknown", "institution": "TapTap", "country": "CN", "position": "Researcher" } ], "n_formula": 47, "n_formula_1": 33, "n_ref_uni": 25, "n_ref": 65, "n_ref_all": 86, "n_fig": 3, "n_tab": 7, "L_tab": 1548, "n_element_tab": 276, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 927, "n_element_tab_1": 134, "formula_len_all": 3941, "formula_len_all_1": 2629, "len_all": 154506, "len_all_1": 64027, "len_abs": 1417, "len_title": 85, "len_sents": 38853, "len_sents_1": 27525, "n_sents": 324, "n_sents_1": 223, "venue": "ICLR2025", "decision": "withdraw", "L_title": 37, "L_abs": 1430, "title": "Linear Attention Sequence Parallelism", "abs": "Sequence parallelism (SP) serves as a prevalent strategy to handle long sequences that exceed the memory limit of a single device. However, for linear sequence modeling methods like linear attention, existing SP approaches do not take advantage of their right-product-first feature, resulting in sub-optimal communication efficiency and usability. In this paper, we introduce Linear Attention Sequence Parallelism (LASP), an efficient SP approach designed for linear attention-based transformer models. Specifically, we design an efficient point-to-point ring-style communication mechanism to leverage the right-product kernel trick of linear attention, which sharply decreases the communication overhead, comparing with existing SP methods. We enhance the computation efficiency of LASP by performing kernel fusion and intermediate state caching, making the implementation of LASP hardware-friendly on GPUs. Furthermore, we meticulously ensure the compatibility of sequence-level LASP with all types of batch-level data parallel methods, which is vital for distributed training on large clusters with very-long sequences. We also discuss the generalization of LASP on other linear sequence modeling methods. Extensive experiments on linear attention-based models are conducted with varying sequence lengths from 2K to 4096K. LASP scales sequence length up to 4096K on 128 GPUs, which is 8$\\times$ longer than existing SP methods.", "keywords": [ "sequence parallelism", "distributed training" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "oVZ9XaOSFK", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Han Guo", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Pengtao Xie", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Rajesh K. Gupta", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Ramtin Hosseini", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Ranak Roy Chowdhury", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Ruiyi Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Sai Ashish Somayajula", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 1, "n_ref_uni": 45, "n_ref": 93, "n_ref_all": 126, "n_fig": 6, "n_tab": 13, "L_tab": 2491, "n_element_tab": 233, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1821, "n_element_tab_1": 208, "formula_len_all": 1135, "formula_len_all_1": 105, "len_all": 197550, "len_all_1": 70614, "len_abs": 1066, "len_title": 142, "len_sents": 67674, "len_sents_1": 31515, "n_sents": 501, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1071, "title": "Downstream Task Guided Masking Learning in Masked Autoencoders Using Multi-Level Optimization", "abs": "Masked Autoencoder (MAE) is a notable method for self-supervised pretraining in visual representation learning. It operates by randomly masking image patches and reconstructing these masked patches using the unmasked ones. A key limitation of MAE lies in its disregard for the varying informativeness of different patches, as it uniformly selects patches to mask. To overcome this, some approaches propose masking based on patch informativeness. However, these methods often do not consider the specific requirements of downstream tasks, potentially leading to suboptimal representations for these tasks. In response, we introduce the Multi-level Optimized Mask Autoencoder (MLO-MAE), a novel framework that leverages end-to-end feedback from downstream tasks to learn an optimal masking strategy during pretraining. Our experimental findings highlight MLO-MAE's significant advancements in visual representation learning. Compared to existing methods, it demonstrates remarkable improvements across diverse datasets and tasks, showcasing its adaptability and efficiency.", "keywords": [ "Multi-level Optimization", "Mask Autoencoder", "Self-Supervised Learning", "Image Masking Strategies", "Representation Learning", "Vision Transformers" ], "rating_list": [ 3, 6, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3, 1 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 5, 5, 5 ] }, { "paper_id": "oVKEAFjEqv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 12, "author_info_list": [ { "name": "Hanyu Lai", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Iat Long Iong", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Jiadai Sun", "gender": "Female", "institution": "Beijing Language and Culture University", "country": "CN", "position": "MS student" }, { "name": "Jie Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Shuntian Yao", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "Undergrad student" }, { "name": "Wei Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Xiao Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xinyue Yang", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "MS student" }, { "name": "Xueqiao Sun", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Yu Yang", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Yuxiao Dong", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "zehan qi", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 8, "n_ref_uni": 41, "n_ref": 91, "n_ref_all": 130, "n_fig": 19, "n_tab": 12, "L_tab": 3611, "n_element_tab": 364, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1272, "n_element_tab_1": 173, "formula_len_all": 1457, "formula_len_all_1": 869, "len_all": 228863, "len_all_1": 67250, "len_abs": 1650, "len_title": 137, "len_sents": 69849, "len_sents_1": 31157, "n_sents": 606, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1579, "title": "WebRL: Training LLM Web Agents via Self-Evolving Online Curriculum Reinforcement Learning", "abs": "Large language models (LLMs) have shown remarkable potential as autonomous agents, particularly in web-based tasks. \nHowever, existing LLM web agents face significant limitations: high-performing agents rely on expensive proprietary LLM APIs, while open LLMs lack the necessary decision-making capabilities. \nThis paper introduces WebRL, a novel self-evolving online curriculum reinforcement learning framework designed to train high-performance web agents using open LLMs. \nOur approach addresses key challenges in this domain, including the scarcity of training tasks, sparse feedback signals, and policy distribution drift in online learning. \nWebRL incorporates a self-evolving curriculum that generates new tasks from unsuccessful attempts, a robust outcome-supervised reward model (ORM), and adaptive reinforcement learning strategies to ensure consistent improvement. \nWe apply WebRL to transform Llama-3.1 models into proficient web agents, achieving remarkable results on the WebArena-Lite benchmark. \nOur Llama-3.1-8B agent improves from an initial 4.8\\% success rate to 42.4\\%, while the Llama-3.1-70B agent achieves a 47.3\\% success rate across five diverse websites. \nThese results surpass the performance of GPT-4-Turbo (17.6\\%) by over 160\\% relatively and significantly outperform previous state-of-the-art web agents trained on open LLMs (AutoWebGLM, 18.2\\%). \nOur findings demonstrate WebRL's effectiveness in bridging the gap between open and proprietary LLM-based web agents, paving the way for more accessible and powerful autonomous web interaction systems.", "keywords": [ "Web Agent", "LLM Agent", "Curriculum RL Learning", "Online Learning" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 1, 4, 2 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "oVCVCo3laS", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Fugee Tsung", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Jia Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Jiexia YE", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Weiqi Zhang", "gender": "Female", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Ziyue Li", "gender": "unknown", "institution": "University of Cologne", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 22, "n_ref": 43, "n_ref_all": 62, "n_fig": 6, "n_tab": 11, "L_tab": 9303, "n_element_tab": 700, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 3617, "n_element_tab_1": 242, "formula_len_all": 1029, "formula_len_all_1": 872, "len_all": 160593, "len_all_1": 65598, "len_abs": 1572, "len_title": 138, "len_sents": 43578, "len_sents_1": 26712, "n_sents": 326, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1533, "title": "DualTime: A Dual-Adapter Language Model for Time Series Multimodal Representation Learning", "abs": "The recent rapid advancements in language models (LMs) have garnered attention in time series multimodal representation learning. However, existing contrastive learning-based and prompt-based LM approaches tend to be biased, often assigning a primary role to time series modality while treating text modality as secondary.\nWe classify these approaches under a temporal-primary paradigm, which overlooks the unique and critical task-relevant information provided by the text modality, failing to fully leverage mutual benefits and complementarity of different modalities.\nTo fill this gap, we propose a novel textual-temporal multimodal learning paradigm that enables either modality to serve as the primary one while being enhanced by the other, thereby effectively capturing modality-specific information and fostering cross-modal interaction. In specific, we design DualTime, a language model composed of dual adapters to implement temporal-primary and textual-primary modeling simultaneously. Within each adapter, lightweight adaptation tokens are injected into the top layers of LM to encourage high-level cross-modal interaction. The shared LM pipeline by dual adapters not only achieves adapter alignment but also reduces computation resources and enables efficient fine-tuning. Empirically, DualTime demonstrates superior performance, achieving notable improvements of 7\\% accuracy and 15\\% F1 in supervised settings. Furthermore, the few-shot label transfer experiments validate DualTime's expressiveness and transferability.", "keywords": [ "Multimodal Representation Learning", "Time Series", "Large Language Model" ], "rating_list": [ 5, 6, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 2, 3, 3 ] }, { "paper_id": "oVATjYtVuf", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Kuan-Chen Pan", "gender": "Female", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "MingHong Chen", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "Ping-Chun Hsieh", "gender": "Male", "institution": "National Chiao Tung University", "country": "", "position": "Assistant Professor" }, { "name": "Xi Liu", "gender": "Male", "institution": "Meta AI", "country": "", "position": "Research Scientist" }, { "name": "You-De Huang", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Undergrad student" } ], "n_formula": 73, "n_formula_1": 19, "n_ref_uni": 45, "n_ref": 102, "n_ref_all": 130, "n_fig": 5, "n_tab": 5, "L_tab": 1535, "n_element_tab": 200, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 213, "n_element_tab_1": 51, "formula_len_all": 12168, "formula_len_all_1": 1651, "len_all": 179312, "len_all_1": 71034, "len_abs": 1356, "len_title": 141, "len_sents": 50947, "len_sents_1": 34543, "n_sents": 401, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1359, "title": "Cross-Domain Reinforcement Learning Under Distinct State-Action Spaces Via Hybrid Q Functions", "abs": "Cross-domain reinforcement learning (CDRL) is meant to improve the data efficiency of RL by leveraging the data samples collected from a source domain to facilitate the learning in a similar target domain. Despite its potential, cross-domain transfer in RL is known to have two fundamental and intertwined challenges: (i) The source and target domains can have distinct state space or action space, and\nthis makes direct transfer infeasible and thereby requires more sophisticated interdomain mappings; (ii) The domain similarity in RL is not easily identifiable a priori, and hence CDRL can be prone to negative transfer. In this paper, we propose to jointly tackle these two challenges through the lens of hybrid Q functions. Specifically, we propose QAvatar, which combines the Q functions from both the source and target domains with a proper weight decay function. Through this design, we characterize the convergence behavior of QAvatar and thereby show that QAvatar achieves reliable transfer in the sense that it effectively leverages a source-domain Q function for knowledge transfer to the target domain. Through extensive experiments, we demonstrate that QAvatar achieves superior transferability across domains on a variety of RL benchmark tasks, such as locomotion and robot arm manipulation, even in the scenarios of potential negative transfer.", "keywords": [ "Cross-domain transfer; Transfer learning; Reinforcement learning" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 4, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "oU3tpaR8fm", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Bowen Jin", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Jiawei Han", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign (UIUC)", "country": "US", "position": "Full Professor" }, { "name": "Jinsung Yoon", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Sercan O Arik", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 25, "n_ref": 42, "n_ref_all": 82, "n_fig": 14, "n_tab": 12, "L_tab": 5617, "n_element_tab": 149, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 572, "formula_len_all_1": 198, "len_all": 231548, "len_all_1": 62172, "len_abs": 1904, "len_title": 120, "len_sents": 85761, "len_sents_1": 32647, "n_sents": 631, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1433, "title": "Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG", "abs": "Retrieval-augmented generation (RAG) empowers large language models (LLMs) to utilize external knowledge sources. The increasing capacity of LLMs to process longer input sequences opens up avenues for providing more retrieved information, to potentially enhance the quality of generated outputs. It is plausible to assume that a larger retrieval set would contain more relevant information (higher recall), that might result in improved performance. However, our empirical findings demonstrate that for many long-context LLMs, the quality of generated output initially improves first, but then subsequently declines as the number of retrieved passages increases. This paper investigates this phenomenon, identifying the detrimental impact of retrieved \"hard negatives\" as a key contributor. To mitigate this and enhance the robustness of long-context LLM-based RAG, we propose both training-free and training-based approaches. We first showcase the effectiveness of retrieval reordering as a simple yet powerful training-free optimization. Furthermore, we explore training-based methods, specifically RAG-specific implicit LLM fine-tuning and RAG-oriented fine-tuning with intermediate reasoning, demonstrating their capacity for substantial performance gains. Finally, we conduct a systematic analysis of design choices for these training-based methods, including data distribution, retriever selection, and training context length.", "keywords": [ "retrieval-augmented generation", "large language models", "long-context" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 3, 3, 2, 4 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "oStNAMWELS", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Leander Girrbach", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Stephan Alaniz", "gender": "Male", "institution": "Helmholtz Zentrum M\u00fcnchen", "country": "DE", "position": "Postdoc" }, { "name": "Yiran Huang", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Zeynep Akata", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "trevor darrell", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department", "country": "US", "position": "Professor" } ], "n_formula": 13, "n_formula_1": 5, "n_ref_uni": 47, "n_ref": 103, "n_ref_all": 176, "n_fig": 24, "n_tab": 11, "L_tab": 7400, "n_element_tab": 492, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4430, "formula_len_all_1": 481, "len_all": 238156, "len_all_1": 63067, "len_abs": 1067, "len_title": 125, "len_sents": 76964, "len_sents_1": 31625, "n_sents": 594, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1071, "title": "Revealing and Reducing Gender Biases in Vision and Language Assistants (VLAs)", "abs": "Pre-trained large language models (LLMs) have been reliably integrated with visual input for multimodal tasks. The widespread adoption of instruction-tuned image-to-text vision-language assistants (VLAs) like LLaVA and InternVL necessitates evaluating gender biases. We study gender bias in 22 popular open-source VLAs with respect to personality traits, skills, and occupations. Our results show that VLAs replicate human biases likely present in the data, such as real-world occupational imbalances. Similarly, they tend to attribute more skills and positive personality traits to women than to men, and we see a consistent tendency to associate negative personality traits with men. To eliminate the gender bias in these models, we find that finetuning-based debiasing methods achieve the best tradeoff between debiasing and retaining performance on downstream task. We argue for pre-deploying gender bias assessment in VLAs and motivate further development of debiasing strategies to ensure equitable societal outcomes. Code and data will be released upon acceptance.", "keywords": [ "gender bias", "vision-language-models" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "oSdrJyb4UH", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Aiguo Chen", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Guangchun Luo", "gender": "unknown", "institution": "University of Electronic Science and Technology of China", "country": "", "position": "Full Professor" }, { "name": "Xu Sun", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yi Luo", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 60, "n_ref": 103, "n_ref_all": 141, "n_fig": 10, "n_tab": 9, "L_tab": 3775, "n_element_tab": 435, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 674, "n_element_tab_1": 48, "formula_len_all": 834, "formula_len_all_1": 630, "len_all": 186222, "len_all_1": 59894, "len_abs": 1438, "len_title": 37, "len_sents": 45931, "len_sents_1": 27450, "n_sents": 309, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 1442, "title": "Monophilic Neighbourhood Transformers", "abs": "Graph neural networks (GNNs) have seen widespread application across diverse fields, including social network analysis, chemical research, and computer vision.\nNevertheless, their efficacy is compromised by an inherent reliance on the homophily assumption, which posits that adjacent nodes should exhibit relevance or similarity.\nThis assumption becomes a limitation when dealing with heterophilic graphs, where it is more common for dissimilar nodes to be connected.\nAddressing this challenge, recent research indicates that real-world graphs generally exhibit monophily, a characteristic where a node tends to be related to the neighbours of its neighbours.\nInspired by this insight, we introduce Neighbourhood Transformers (NT), a novel approach that employs self-attention within every neighbourhood of the graph to generate informative messages for the nodes within, as opposed to the central node in conventional GNN frameworks.\nWe develop a neighbourhood partitioning strategy equipped with switchable attentions, significantly reducing space consumption by over 95\\% and time consumption by up to 92.67\\% in NT.\nExperimental results on node classification tasks across 5 heterophilic and 5 homophilic graphs demonstrate that NT outperforms current state-of-the-art methods, showcasing their expressiveness and adaptability to different graph types.\nThe code for this study is available at https://anonymous.4open.science/r/MoNT-BD3C .", "keywords": [ "graph neural networks", "transformers" ], "rating_list": [ 5, 8, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "oSQiao9GqB", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bo Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Chunyuan Li", "gender": "unknown", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Feng Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Hao Zhang", "gender": "Male", "institution": "NVIDIA Corporation", "country": "US", "position": "Researcher" }, { "name": "Renrui Zhang", "gender": "Male", "institution": "MMLab of CUHK & Shanghai AI Laboratory", "country": "HK", "position": "PhD student" }, { "name": "Wei Li", "gender": "Male", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "Yuanhan Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Zejun MA", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 17, "n_ref": 29, "n_ref_all": 55, "n_fig": 5, "n_tab": 15, "L_tab": 7559, "n_element_tab": 695, "n_fig_1": 4, "n_tab_1": 10, "L_tab_1": 5841, "n_element_tab_1": 441, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 161905, "len_all_1": 57023, "len_abs": 262, "len_title": 133, "len_sents": 26985, "len_sents_1": 20613, "n_sents": 214, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 85, "L_abs": 1231, "title": "LLaVA-NeXT-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Models", "abs": "Visual instruction tuning has made considerable strides in enhancing the capabilities of Large Multimodal Models (LMMs). However, existing open LMMs largely focus on single-image tasks, their applications to multiimage scenarios remains less explored. Additionally,\nprior LMM research separately tackles different scenarios, leaving it impossible to generalize cross scenarios with new emerging capabilities. To this end, we introduce LLaVA-NeXT-Interleave, which simultaneously tackles Multi-image, Multi-frame (video), Multi-view (3D), and Multi-patch (single-image) scenarios in LMMs. To enable these capabilities, we regard the interleaved data format as a general template and compile the M4-Instruct dataset with 1,177.6k samples, spanning 4 primary domains with 14 tasks and 41 datasets. We also curate the LLaVAInterleave Bench to comprehensively evaluate the multiimage performance of LMMs. Through extensive experiments, LLaVA-NeXT-Interleave achieves leading results in multi-image, video, and 3D benchmarks, while maintaining the performance of single-image tasks. Besides, our model also exhibits several emerging capabilities, e.g., transferring tasks across different settings and modalities. Code will be available.", "keywords": [ "large language model", "multimodal learning", "interleaved image-text" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 4, 3, 2 ], "confidence_list": [ 4, 2, 5 ] }, { "paper_id": "oSJqRF0Tkg", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Dong Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Distinguished Scientist" }, { "name": "Hongming Zhang", "gender": "Male", "institution": "Tencent AI Lab Seattle", "country": "US", "position": "Researcher" }, { "name": "Kaixin Ma", "gender": "unknown", "institution": "Tencent AI Lab", "country": "US", "position": "Researcher" }, { "name": "Meng Jiang", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "Associate Professor" }, { "name": "Mengzhao Jia", "gender": "Female", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Siru Ouyang", "gender": "Female", "institution": "University of Illinois Urbana-Champaign Champaign", "country": "US", "position": "PhD student" }, { "name": "Tianqing Fang", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Senior Research Scientist" }, { "name": "Wenhao Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Researcher" }, { "name": "Zhihan Zhang", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 47, "n_ref": 136, "n_ref_all": 180, "n_fig": 10, "n_tab": 7, "L_tab": 4000, "n_element_tab": 432, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3795, "n_element_tab_1": 317, "formula_len_all": 73, "formula_len_all_1": 74, "len_all": 211460, "len_all_1": 72766, "len_abs": 1755, "len_title": 113, "len_sents": 40787, "len_sents_1": 31813, "n_sents": 289, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1760, "title": "Leopard: A Vision Language Model For Text-Rich Multi-Image Tasks", "abs": "Text-rich images, where text serves as the central visual element guiding the overall understanding, are prevalent in real-world applications, such as presentation slides, scanned documents, and webpage snapshots. Tasks involving multiple text-rich images are especially challenging, as they require not only understanding the content of individual images but reasoning about inter-relationships and logical flows across multiple visual inputs.\nDespite the importance of these scenarios, current multimodal large language models (MLLMs) struggle to handle such tasks due to two key challenges: (1) the scarcity of high-quality instruction tuning datasets for text-rich multi-image scenarios, and (2) the difficulty in balancing image resolution with visual feature sequence length. Low-resolution encoding impairs the recognition of embedded text, while high-resolution encoding quickly exceeds the model\u2019s maximum sequence length under multi-image settings.\nTo address these challenges, we propose Leopard, a MLLM designed specifically for handling vision-language tasks involving multiple text-rich images. \nFirst, we curated about one million high-quality multimodal instruction-tuning data, tailored to text-rich, multi-image scenarios.\nSecond, we developed an adaptive high-resolution multi-image encoding module to dynamically optimize the allocation of visual sequence length based on the original aspect ratios\nand resolutions of the input images.\nExperiments across a wide range of benchmarks demonstrate our model's superior capabilities in text-rich, multi-image evaluations and competitive performance in general domain evaluations.\nWe are committed to open-source models and will release all collected data, code, and checkpoints to the community.", "keywords": [ "Multimodal Large Language Models", "Text-rich Image", "Multi-image" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "oSEsSDFxyw", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Amaya Dharmasiri", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Olga Russakovsky", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Associate Professor" }, { "name": "Ruoyu Wang", "gender": "Female", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Sai Wang", "gender": "unknown", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Ye Zhu", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Yu Wu", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 0, "n_ref_uni": 70, "n_ref": 198, "n_ref_all": 218, "n_fig": 10, "n_tab": 4, "L_tab": 798, "n_element_tab": 110, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1924, "n_element_tab_1": 200, "formula_len_all": 220, "formula_len_all_1": 0, "len_all": 178808, "len_all_1": 73776, "len_abs": 1198, "len_title": 118, "len_sents": 47656, "len_sents_1": 35350, "n_sents": 310, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1200, "title": "DETER: Detecting Edited Regions for Deterring Generative Manipulations", "abs": "Generative AI capabilities have grown substantially in recent years, raising renewed concerns about the potential malicious use of generated data, or \"deep fakes.\" Despite being a longstanding and important research topic, deep fake detection research on most existing datasets has not kept pace with generative AI advancements sufficiently to develop detection technology that can meaningfully alert human users in real-world settings. In this work, we introduce DETER, a large-scale dataset for DETEcting edited image Regions and deterring modern advanced generative manipulations. After a comprehensive study of prior literature, our proposed dataset makes contributions along three main axes: the upgrade on modern manipulations via the state-of-the-art generative models; the mitigation of biased spurious correlations in prior deep fake datasets; and a more unified formulation suitable for various detection models in different granularities. Equipped with DETER, we conduct extensive experiments and detailed analysis using our rich annotations and improved benchmark protocols, revealing future directions and the next set of challenges in developing reliable regional fake detection models.", "keywords": [ "Deepfake detection", "regional manipulation", "dataset and benchmark" ], "rating_list": [ 5, 5, 8, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "oS79Tw3G0c", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chang Liu", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Jian Jia", "gender": "Male", "institution": "kuaishou", "country": "CN", "position": "Researcher" }, { "name": "Peng Jiang", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Vice President" }, { "name": "Quan Chen", "gender": "Male", "institution": "Kuaishou", "country": "CN", "position": "Researcher" }, { "name": "Ye Ma", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Researcher" }, { "name": "Zhijie Deng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 43, "n_ref": 80, "n_ref_all": 93, "n_fig": 3, "n_tab": 6, "L_tab": 1586, "n_element_tab": 275, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1199, "n_element_tab_1": 207, "formula_len_all": 368, "formula_len_all_1": 368, "len_all": 145932, "len_all_1": 67233, "len_abs": 1669, "len_title": 133, "len_sents": 35063, "len_sents_1": 29851, "n_sents": 300, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1776, "title": "On Exploring Visual Attention Shrinking for Accelerating VLMs for Video Understanding", "abs": "Vision-language models (VLMs) have shown promise in a variety of challenging video comprehension tasks. VLMs typically extract frames from the source video and take the corresponding encoded visual tokens as input. A rapid increase in the number of visual tokens, e.g., when handling lengthy videos, can swiftly lead to a long-context dilemma during the inference process of VLMs, posing an efficiency challenge for real-world applications. Given that significant redundant and task-irrelevant information may exist in the visual tokens across both spatial and temporal axes, we advocate removing less important visual tokens during the prefilling phase of the inference procedure to improve the computation and storage efficiency of VLMs. We first identify an interesting phenomenon termed as \\emph{Visual Attention Shrinking (VAS)}, wherein certain visual tokens receive progressively diminishing attention during the processing stages of the model. This implies that the model itself knows what to care about and what to discard. With this understanding, we develop a robust algorithm to detect attention shrinking at each layer of the model using states from preceding layers. Based on the detection results, we perform token removal in both temporal and spatial axes. Our approach does not require parameterized modifications to the original VLM and is compatible with the prevalent KV cache strategy. Through extensive experiments across different VLMs, our approach witnesses an average speedup of $1.98\\times$ in generating the first response token, utilizing only 47.2% of the visual tokens, without compromising the task performance. Additionally, when applied to the huge VILA1.5-40B, our method can achieve up to $4.16\\times$ speedup compared to the vanilla model.", "keywords": [ "Visual Language model", "Inference Acceleration", "Visual Attention Shrinking" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "oRlANEuqG5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Fatma G\u00fcney", "gender": "Female", "institution": "Koc University", "country": "TR", "position": "Assistant Professor" }, { "name": "G\u00f6rkay Aydemir", "gender": "Male", "institution": "Ko\u00e7 University", "country": "TR", "position": "PhD student" }, { "name": "Weidi Xie", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiongyi Cai", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 26, "n_ref": 98, "n_ref_all": 137, "n_fig": 25, "n_tab": 10, "L_tab": 1933, "n_element_tab": 340, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 912, "n_element_tab_1": 182, "formula_len_all": 794, "formula_len_all_1": 726, "len_all": 182356, "len_all_1": 66352, "len_abs": 1174, "len_title": 109, "len_sents": 57383, "len_sents_1": 29627, "n_sents": 411, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1254, "title": "Track-On: Transformer-based Online Point Tracking with Memory", "abs": "In this paper, we consider the problem of long-term point tracking, which requires consistent identification of points across multiple frames in a video, despite changes in appearance, lighting, perspective, and occlusions. We target online tracking on a frame-by-frame basis, making it suitable for real-world, streaming scenarios. Specifically, we introduce Track-On, a simple transformer-based model designed for online long-term point tracking. Unlike prior methods that depend on full temporal modeling, our model processes video frames causally without access to future frames, leveraging two memory modules \u2014spatial memory and context memory\u2014 to capture temporal information and maintain reliable point tracking over long time horizons. At inference time, it employs patch classification and refinement to identify correspondences and track points with high accuracy. Through extensive experiments, we demonstrate that Track-On sets a new state-of-the-art for online models and delivers superior or competitive results compared to offline approaches on seven datasets, including the TAP-Vid benchmark. Our method offers a robust and scalable solution for real-time tracking in diverse applications. Project page: https://kuis-ai.github.io/track_on", "keywords": [ "Point Tracking", "Online Point Tracking", "Tracking Any Point" ], "rating_list": [ 6, 5, 8, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "oRfHv642qD", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Dekel Galor", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Hadi Vafaii", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Jacob L. Yates", "gender": "Not Specified", "institution": "University of California, Berkeley", "country": "", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 105, "n_ref": 175, "n_ref_all": 202, "n_fig": 6, "n_tab": 1, "L_tab": 628, "n_element_tab": 78, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 628, "n_element_tab_1": 78, "formula_len_all": 835, "formula_len_all_1": 658, "len_all": 202921, "len_all_1": 65762, "len_abs": 4043, "len_title": 94, "len_sents": 48327, "len_sents_1": 30658, "n_sents": 364, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1244, "title": "A prescriptive theory for brain-like inference", "abs": "The Evidence Lower Bound (ELBO) is a widely used objective for training deep generative models, such as Variational Autoencoders (VAEs). In the neuroscience literature, an identical objective is known as the Free Energy Principle (FEP), hinting at a potential unified framework for brain function and machine learning. Despite its utility in interpreting generative models, including diffusion models, ELBO maximization is often seen as too broad to offer prescriptive guidance for specific architectures in neuroscience or machine learning. In this work, we show that maximizing ELBO under Poisson assumptions for general sequences leads to a spiking neural network that performs Bayesian posterior inference through its membrane potential dynamics. The resulting model, the iterative Poisson VAE (iP-VAE), has a closer connection to biological neurons than previous brain-inspired predictive coding models based on Gaussian assumptions. Compared to amortized and iterative VAEs, iP-VAE learns sparser representations and exhibits superior generalization to out-of-distribution samples. These findings suggest that optimizing ELBO, combined with Poisson assumptions, provides a solid foundation for developing prescriptive theories in NeuroAI.", "keywords": [ "iterative inference", "elbo", "variational inference", "ood generalization", "variational autoencoder", "sampling" ], "rating_list": [ 3, 5, 6, 3, 5 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 3, 3 ] }, { "paper_id": "oRPXPoTXYz", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Alexandre Allauzen", "gender": "Male", "institution": "Ecole Sup\u00e9rieure de Physique et de Chimie Industrielles", "country": "FR", "position": "Full Professor" }, { "name": "Blaise Delattre", "gender": "Male", "institution": ", Universit\u00e9 Paris-Dauphine (Paris IX)", "country": "FR", "position": "PhD student" }, { "name": "Erwan Fagnou", "gender": "Male", "institution": "Universit\u00e9 Paris-Dauphine - PSL", "country": "FR", "position": "PhD student" }, { "name": "Paul Caillon", "gender": "unknown", "institution": ", Universit\u00e9 Paris-Dauphine (Paris IX)", "country": "", "position": "Postdoc" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 38, "n_ref": 83, "n_ref_all": 102, "n_fig": 4, "n_tab": 4, "L_tab": 811, "n_element_tab": 36, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 786, "n_element_tab_1": 34, "formula_len_all": 863, "formula_len_all_1": 865, "len_all": 108038, "len_all_1": 70043, "len_abs": 1357, "len_title": 113, "len_sents": 34214, "len_sents_1": 31939, "n_sents": 250, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1369, "title": "Backpropagation-Free Learning through Gradient Aligned Feedbacks", "abs": "Deep neural networks heavily rely on the back-propagation algorithm for optimiza-\ntion. Nevertheless, the global sequential transmission of gradients in the backward\npass inhibits its scalability. The Direct Feedback Alignment algorithm has been\nproposed as a promising approach for parallel learning of deep neural networks,\nrelying on fixed random feedback weights to project the error on every layer in\na parallel manner. However, it notoriously fails to train networks that are really\ndeep and that include compulsory layers like convolutions and transformers. In this\npaper, we show that alternatives to back-propagation may greatly benefit from local\nand forward approximation of the gradient to better cope with the inherent and\nconstrained structure of such layers. \n\nThis directional approximation allows us to design a novel algorithm that updates the feedback weights called GrAPE (GRadient\nAligned Projected Error). A first set of experiments are carried out on image classi-\nfication tasks with feedforward and convolutional architectures. The results show\nimportant improvement in performance over other backpropagation-free algorithms,\nnarrowing the gap with backpropagation. More importantly, the method scales\nto modern and deep architectures like AlexNet, VGG-16 and Transformer-based\nlanguage models where the performance gains are even more notable.", "keywords": [ "backpropagation-free", "direct feedback alignment", "optimization", "forward gradient" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "oRNus243R6", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Haike Xu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Kirankumar Shiragur", "gender": "Male", "institution": "Microsoft Research", "country": "IN", "position": "Researcher" }, { "name": "Piotr Indyk", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Piyush Anand", "gender": "unknown", "institution": "Microsoft", "country": "IN", "position": "Senior Software Engineer" }, { "name": "Ravishankar Krishnaswamy", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Sepideh Mahabadi", "gender": "Female", "institution": "Microsoft Research", "country": "US", "position": "Researcher" }, { "name": "Vikas C. Raykar", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Principal Data & Applied Scientist and Manager" } ], "n_formula": 45, "n_formula_1": 26, "n_ref_uni": 11, "n_ref": 39, "n_ref_all": 50, "n_fig": 3, "n_tab": 3, "L_tab": 516, "n_element_tab": 11, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6501, "formula_len_all_1": 3757, "len_all": 125014, "len_all_1": 72813, "len_abs": 1774, "len_title": 91, "len_sents": 44311, "len_sents_1": 29581, "n_sents": 388, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1799, "title": "Diverse Graph-based Nearest Neighbor Search", "abs": "Nearest neighbor search is a fundamental data structure problem with many applications in machine learning, computer vision, recommendation systems and other fields. Although the main objective of the data structure is to quickly report data points that are closest to a given query, it has long been noted (Carbonell et al, 1998) that without additional constraints the reported answers can be redundant and/or duplicative. This issue is typically addressed in two stages: in the first stage, the algorithm retrieves a (large) number of points closest to the query, while in the second stage, the points are post-processed and a small subset is selected to maximize the desired diversity objective. Although popular, this method suffers from a fundamental efficiency bottleneck, as the set of points retrieved in the first stage often needs to be much larger than the final output. \n\nIn this paper we present provably efficient algorithms for approximate nearest neighbor search with diversity constraints that bypass this two stage process. Our algorithms are based on popular graph-based methods, which allows us to ``piggy-back'' on the existing efficient implementations. These are the first graph-based algorithms for nearest neighbor search with diversity constraints. For data sets with low intrinsic dimension, our data structures report a diverse set of $k$ points approximately closest to the query, in time that only depends on $k$ and $\\log \\Delta$, where $\\Delta$ is the ratio of the diameter to the closest pair distance in the data set. This bound is qualitatively similar to the best known bounds for standard (non-diverse) graph-based algorithms. Our experiments show that the search time of our algorithms is substantially lower than that using the standard two-stage approach.", "keywords": [ "nearest neighbor search; diversity; algorithms" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 1, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "oQoQ4u6MQC", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Brian Nlong Zhao", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Research Assistant" }, { "name": "Dongsheng Li", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Jiashu Xu", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "Undergrad student" }, { "name": "Laurent Itti", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Full Professor" }, { "name": "Vibhav Vineet", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "XINYANG JIANG", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Researcher" }, { "name": "Yifan Yang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Yuhang Xiao", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "MS student" }, { "name": "Yunhao Ge", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 8, "n_ref_uni": 54, "n_ref": 115, "n_ref_all": 150, "n_fig": 23, "n_tab": 6, "L_tab": 1448, "n_element_tab": 245, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 986, "formula_len_all_1": 348, "len_all": 176407, "len_all_1": 62194, "len_abs": 545, "len_title": 130, "len_sents": 48544, "len_sents_1": 30207, "n_sents": 333, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1018, "title": "DreamDistribution: Prompt Distribution Learning for Text-to-Image Diffusion Models", "abs": "The popularization of Text-to-Image (T2I) diffusion models enables the generation of high-quality images from text descriptions. However, generating diverse customized images with reference visual attributes remains challenging. This work focuses on personalizing T2I diffusion models at a more abstract concept or category level, adapting commonalities from a set of reference images while creating new instances with sufficient variations. We introduce a solution that allows a pretrained T2I diffusion model to learn a set of soft prompts, enabling the generation of novel images by sampling prompts from the learned distribution. These prompts offer text-guided editing capabilities and additional flexibility in controlling variation and mixing between multiple distributions. We also show the adaptability of the learned prompt distribution to other tasks, such as text-to-3D. Finally we demonstrate effectiveness of our approach through quantitative analysis including automatic evaluation and human assessment.", "keywords": [ "Generative Models", "Image Generation", "Personalized Generation" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "oQUtBLM8Bo", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Chuchu Fan", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Mitchell Black", "gender": "Male", "institution": "MIT Lincoln Laboratory, Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Oswin So", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Songyuan Zhang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zachary Serlin", "gender": "Male", "institution": "MIT Lincoln Laboratory, Massachusetts Institute of Technology", "country": "", "position": "Researcher" } ], "n_formula": 46, "n_formula_1": 28, "n_ref_uni": 63, "n_ref": 140, "n_ref_all": 169, "n_fig": 7, "n_tab": 3, "L_tab": 741, "n_element_tab": 45, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3703, "formula_len_all_1": 1970, "len_all": 193935, "len_all_1": 68386, "len_abs": 1106, "len_title": 113, "len_sents": 51598, "len_sents_1": 30269, "n_sents": 399, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1114, "title": "Distributed Epigraph Form Multi-Agent Safe Reinforcement Learning", "abs": "Most existing safe multi-agent reinforcement learning (MARL) algorithms consider the constrained Markov decision process (CMDP) problem, which targets bringing the mean of constraint violation below a user-defined threshold.\nHowever, as observed by existing works albeit for the single-agent case, CMDP algorithms suffer from unstable training when the constraint threshold is zero.\nThis paper proposes **EFMARL**, a novel MARL algorithm that improves upon the problems faced in the zero constraint threshold setting by extending the *epigraph form*, a technique to perform constrained optimization, to the centralized training and distributed execution (CTDE) paradigm.\nWe validate our approach in different Multi-Particle Environments and Safe Multi-agent MuJoCo environments with varying numbers of agents. Simulation results show that our algorithm achieves stable training and the best performance while satisfying constraints: it is as safe as the safest baseline that has significant performance loss, and achieves similar performance as baselines that prioritize performance but violate safety constraints.", "keywords": [ "Safe multi-agent systems", "reinforcement learning", "optimal control", "epigraph form" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "oQ4igHyh3N", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bernt Schiele", "gender": "Male", "institution": "Max Planck Institute for Informatics, Saarland Informatics Campus", "country": "DE", "position": "Full Professor" }, { "name": "Federico Tombari", "gender": "Male", "institution": "Google", "country": "CH", "position": "Research Director" }, { "name": "Haiyang Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jan Eric Lenssen", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Principal Researcher" }, { "name": "Liwei Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Muhammad Ferjad Naeem", "gender": "unknown", "institution": "Google", "country": "CH", "position": "Google" }, { "name": "Yongqin Xian", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Yue Fan", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 17, "n_ref_uni": 36, "n_ref": 74, "n_ref_all": 100, "n_fig": 7, "n_tab": 7, "L_tab": 3330, "n_element_tab": 353, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 673, "n_element_tab_1": 7, "formula_len_all": 1955, "formula_len_all_1": 1023, "len_all": 152795, "len_all_1": 57051, "len_abs": 1415, "len_title": 124, "len_sents": 43883, "len_sents_1": 26922, "n_sents": 307, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 75, "L_abs": 1422, "title": "TokenFormer: Rethinking Transformer Scaling with Tokenized Model Parameters", "abs": "Transformers have become the predominant architecture in foundation models due to their excellent performance across various domains. However, the substantial cost of scaling these models remains a significant concern. This problem arises primarily from their dependence on a fixed number of parameters within linear projections. When architectural modifications (e.g., channel dimensions) are introduced, the entire model typically requires retraining from scratch. As model sizes continue growing, this strategy results in increasingly high computational costs and becomes unsustainable. To overcome this problem, we introduce Tokenformer, a natively scalable architecture that leverages the attention mechanism not only for computations among input tokens but also for interactions between tokens and model parameters, thereby enhancing architectural flexibility. By treating model parameters as tokens, we replace all the linear projections in Transformers with our token-parameter attention layer, where input tokens act as queries and model parameters as keys and values. This reformulation allows for progressive and efficient scaling without necessitating retraining from scratch. Our model scales from 124M to 1.4B parameters by incrementally adding new key-value parameter pairs, achieving performance comparable to Transformers trained from scratch while greatly reducing training costs. Code will be available.", "keywords": [ "Fully Attention-based Neural Network", "Large Language Model", "Model Scaling", "Tokenized Model Parameters" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 4 ], "contribution_list": [ 4, 3, 3, 4 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "oP7arLOWix", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Chen Xu", "gender": "Male", "institution": "Toyota Research Institute", "country": "US", "position": "Researcher" }, { "name": "Jonghyeok Lee", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Yao Xie", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 88, "n_formula_1": 38, "n_ref_uni": 46, "n_ref": 80, "n_ref_all": 90, "n_fig": 5, "n_tab": 6, "L_tab": 2052, "n_element_tab": 115, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 478, "n_element_tab_1": 6, "formula_len_all": 8262, "formula_len_all_1": 2155, "len_all": 141691, "len_all_1": 57849, "len_abs": 991, "len_title": 112, "len_sents": 39485, "len_sents_1": 24661, "n_sents": 370, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 831, "title": "Kernel-based Optimally Weighted Conformal Time-Series Prediction", "abs": "Conformal prediction has been a popular distribution-free framework for uncertainty quantification. In this work, we present a novel conformal prediction method for time-series, which we call Kernel-based Optimally Weighted Conformal Prediction Intervals ($\\texttt{KOWCPI}$). Specifically, $\\texttt{KOWCPI}$ adapts the classic Reweighted Nadaraya-Watson (RNW) estimator for quantile regression on dependent data and learns optimal data-adaptive weights. Theoretically, we tackle the challenge of establishing a conditional coverage guarantee for non-exchangeable data under strong mixing conditions on the non-conformity scores. We demonstrate the superior performance of $\\texttt{KOWCPI}$ on real time-series against state-of-the-art methods, where $\\texttt{KOWCPI}$ achieves narrower confidence intervals without losing coverage.", "keywords": [ "Conformal prediction", "Nonparametric kernel regression", "Time series" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "oOSeOEXrFA", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Kangjie Chen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Shangqing Liu", "gender": "unknown", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" }, { "name": "Tianwei Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Yang Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Yanzhou Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 36, "n_ref": 89, "n_ref_all": 104, "n_fig": 5, "n_tab": 6, "L_tab": 3505, "n_element_tab": 591, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 957, "n_element_tab_1": 143, "formula_len_all": 2140, "formula_len_all_1": 1180, "len_all": 162822, "len_all_1": 68032, "len_abs": 1544, "len_title": 133, "len_sents": 50991, "len_sents_1": 32790, "n_sents": 341, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1549, "title": "REPOFILTER: Adaptive Retrieval Context Trimming for Repository-Level Code Completion", "abs": "Retrieval-Augmented Generation (RAG) has recently emerged as a promising approach for repository-level code completion by integrating cross-file knowledge with in-file preceding code to provide comprehensive contexts for generation. To better understand the contribution of the retrieved cross-file contexts, we introduce a likelihood-based metric to evaluate the impact of each retrieved code chunk on the completion. Our analysis reveals that, despite retrieving numerous chunks, only a small subset positively contributes to the target completion, while some chunks even degrade performance. To address this issue, we leverage this metric to construct a repository-level dataset where each retrieved chunk is labeled as positive, neutral, or negative based on its relevance to the target completion. We then propose an adaptive retrieval context trimming framework, REPOFILTER, trained on this dataset to mitigate the harmful effects of negative retrieved contexts in RAG-based code completion. Extensive evaluation on the RepoEval and CrossCodeLongEval benchmarks demonstrates that REPOFILTER consistently improves completion accuracy compared to approaches without filtering operations across various tasks. Additionally, REPOFILTER significantly reduces the length of the input prompt, enhancing computational efficiency while exhibiting strong generalizability across different models. These results underscore the potential of REPOFILTER to enhance the accuracy, efficiency, and attributability of RAG-based repository-level code completion.", "keywords": [ "Code completion", "RAG", "LLMs" ], "rating_list": [ 5, 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3, 4 ], "presentation_list": [ 2, 3, 4, 1, 4 ], "contribution_list": [ 3, 3, 3, 2, 4 ], "confidence_list": [ 5, 4, 3, 5, 4 ] }, { "paper_id": "oOQavkQLQZ", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Chen Xiaoyu", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Jianfei Chen", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Jun Zhu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Yuji Wang", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zehua Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" } ], "n_formula": 54, "n_formula_1": 10, "n_ref_uni": 46, "n_ref": 148, "n_ref_all": 187, "n_fig": 6, "n_tab": 7, "L_tab": 1660, "n_element_tab": 200, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1062, "n_element_tab_1": 127, "formula_len_all": 5601, "formula_len_all_1": 1177, "len_all": 195186, "len_all_1": 60494, "len_abs": 2757, "len_title": 67, "len_sents": 48368, "len_sents_1": 27774, "n_sents": 343, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1791, "title": "FrameBridge: Improving Image-to-Video Generation with Bridge Models", "abs": "Image-to-video (I2V) generation is gaining increasing attention with its wide application in video synthesis. Recently, diffusion-based I2V models have achieved remarkable progress given their novel design on network architecture, cascaded framework, and motion representation. However, restricted by their noise-to-data generation process, diffusion-based methods inevitably suffer the difficulty to generate video samples with both appearance consistency and temporal coherence from an uninformative Gaussian noise, which may limit their synthesis quality.\nIn this work, we present FrameBridge, taking the given static image as the prior of video target and establishing a tractable bridge model between them. By formulating I2V synthesis as a frames-to-frames generation task and modeling it with a \\textit{data-to-data} process, we fully exploit the information in input image and facilitate the generative model to learn the image animation process.\nIn two popular settings of training I2V models, namely fine-tuning a pre-trained text-to-video (T2V) model or training from scratch, we further propose two techniques, SNR-Aligned Fine-tuning (SAF) and neural prior, which improve the fine-tuning efficiency of diffusion-based T2V models to FrameBridge and the synthesis quality of bridge-based I2V models respectively. \nExperiments conducted on WebVid-2M and UCF-101 demonstrate that: (1) our FrameBridge achieves superior I2V quality in comparison with the diffusion counterpart (zero-shot FVD 95 vs. 192 on MSR-VTT and non-zero-shot FVD 122 vs. 171 on UCF-101); (2) our proposed SAF and neural prior effectively enhance the ability of bridge-based I2V models in the scenarios of fine-tuning and training from scratch. Demo samples can be visited at: https://framebridgei2v.github.io/.", "keywords": [ "Image-to-Video Generation", "Diffusion Models", "Diffusion Bridge Models", "Prior Distribution", "Data-to-Data Generation" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "oO3oXJ19Pb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "AJ Piergiovanni", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Anelia Angelova", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Dahun Kim", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 51, "n_ref": 113, "n_ref_all": 142, "n_fig": 1, "n_tab": 12, "L_tab": 1065, "n_element_tab": 130, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 888, "n_element_tab_1": 106, "formula_len_all": 208, "formula_len_all_1": 208, "len_all": 158295, "len_all_1": 72213, "len_abs": 1034, "len_title": 124, "len_sents": 45571, "len_sents_1": 34945, "n_sents": 349, "n_sents_1": 247, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1036, "title": "Actions Inspire Every Moment: Online Action-Augmented Dense Video Captioning", "abs": "Dense video captioning requires solving the challenging tasks of temporally localizing events and generating descriptive captions within long video sequences. Existing methods often struggle to capture the evolving context within video streams and to produce accurate temporal alignment. To address this, we propose an online retrieval-augmented approach that processes video segments incrementally while dynamically retrieving relevant action phrases from a pre-constructed action-text corpus. This enriches the contextual information for both the video representation and the subsequent text decoder, improving the caption generation. Additionally, we present image-based simulated video pretraining, which mitigates the reliance on extensive video datasets by using image-level text-paired data aligned with the online video captioning format. Our experiments on the ViTT, YouCook2, and ActivityNet benchmarks demonstrate that our model significantly outperforms both existing global and online methods, validating its effectiveness.", "keywords": [ "Dense video captioning", "Online dense video captioning" ], "rating_list": [ 6, 5, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "oMkHoJjLXB", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Alexi Gladstone", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Aman Chadha", "gender": "Male", "institution": "Amazon Web Services", "country": "US", "position": "Science Manager" }, { "name": "Ganesh Nanduru", "gender": "Male", "institution": "Chevron Corporation", "country": "US", "position": "Intern" }, { "name": "Keyan Du", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Undergrad student" }, { "name": "Md Fahim", "gender": "Male", "institution": "Independent University, Bangladesh", "country": "BD", "position": "Researcher" }, { "name": "Md Mofijul Islam", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Sujan Sarker", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Tariq Iqbal", "gender": "unknown", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 27, "n_ref": 49, "n_ref_all": 58, "n_fig": 3, "n_tab": 7, "L_tab": 1548, "n_element_tab": 216, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 319, "n_element_tab_1": 41, "formula_len_all": 631, "formula_len_all_1": 631, "len_all": 121552, "len_all_1": 59151, "len_abs": 1903, "len_title": 129, "len_sents": 31485, "len_sents_1": 28929, "n_sents": 199, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1907, "title": "Embodied Referring Expression Comprehension Through Multimodal Residual Learning", "abs": "Comprehending embodied interactions within real-world settings poses a considerable challenge, attributed to the multifaceted nature of human interactions and the variability of environments, necessitating the development of comprehensive benchmark datasets and multimodal learning models. Existing datasets do not adequately represent the full spectrum of human interactions, are limited by perspective bias, rely on single viewpoints, have insufficient nonverbal gesture capture, and have a predominant focus on indoor settings. To address these gaps, we present an Embodied Referring Expressions dataset (called Refer360), which contains an extensive collection of embodied verbal and nonverbal interaction data captured from various viewpoints across various indoor and outdoor settings. In conjunction with this benchmark dataset, we propose a novel multimodal guided residual module (MuRes) that helps the existing multimodal models to improve their representations. This guided residual module acts as an information bottleneck to extract salient modality-specific representations, and reinforcing these to the pre-trained representations produces robust complementary representations for downstream tasks. Our extensive experimental analysis of our benchmark Refer360 dataset reveals that existing multimodal models alone fail to capture human interactions in real-world scenarios comprehensively for embodied referring expression comprehension tasks. Building on these findings, a thorough analysis of four benchmark datasets demonstrates superior performance by augmenting MuRes into current multimodal models, highlighting its capability to improve the understanding and interaction with human-centric environments. This paper offers a benchmark for the research community and marks a stride towards developing robust systems adept at navigating the complexities of real-world human interactions.", "keywords": [ "embodied", "multimodal", "visual-language" ], "rating_list": [ 6, 5, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "oMfZUSbVwf", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Bo Zhao", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Nima Dehmamy", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Robin Walters", "gender": "Male", "institution": "Northeastern University ", "country": "US", "position": "Assistant Professor" }, { "name": "Rose Yu", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" } ], "n_formula": 44, "n_formula_1": 25, "n_ref_uni": 28, "n_ref": 39, "n_ref_all": 47, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5827, "formula_len_all_1": 2673, "len_all": 94926, "len_all_1": 66093, "len_abs": 706, "len_title": 100, "len_sents": 27181, "len_sents_1": 21986, "n_sents": 236, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 51, "L_abs": 712, "title": "Finding Symmetry in Neural Network Parameter Spaces", "abs": "Parameter space symmetries, or loss-invariant transformations, are important for understanding neural networks' loss landscape, training dynamics, and generalization. \nHowever, identifying the full set of these symmetries remains a challenge. \nIn this paper, we formalize data-dependent parameter symmetries and derive their infinitesimal form, which enables an automated approach to discover symmetry across different architectures. \nOur framework systematically uncovers parameter symmetries, including previously unknown ones. \nWe also prove that symmetries in smaller subnetworks can extend to larger networks, allowing the discovery of symmetries in small architectures to generalize to more complex models.", "keywords": [ "Parameter space symmetry" ], "rating_list": [ 1, 3, 3, 5 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 1, 3, 1, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "oMFOKjwaRS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Feng Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Hanzhu Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Jie Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jieping Ye", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Junjie He", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Qitan Lv", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Rong Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Xu Shen", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Zehao Wang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 31, "n_ref": 41, "n_ref_all": 67, "n_fig": 6, "n_tab": 18, "L_tab": 6110, "n_element_tab": 695, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1746, "n_element_tab_1": 221, "formula_len_all": 81, "formula_len_all_1": 81, "len_all": 191057, "len_all_1": 64047, "len_abs": 1785, "len_title": 132, "len_sents": 58550, "len_sents_1": 27704, "n_sents": 417, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1793, "title": "Knowledge Graph Finetuning Enhances Knowledge Manipulation in Large Language Models", "abs": "Despite the impressive performance of general large language models(LLMs), many of their applications in specific domains (e.g., low-data and knowledge-intensive) still confront significant challenges. Supervised fine-tuning (SFT)---where a general LLM is further trained on a small labeled dataset to adapt for specific tasks or domains---has shown great power for developing domain-specific LLMs. However, existing SFT data primarily consist of Question and Answer (Q&A) pairs, which poses a significant challenge for LLMs to comprehend the correlation and logic of knowledge underlying the Q&A. To address this challenge, we propose a conceptually flexible and general framework to boost SFT, namely Knowledge Graph-Driven Supervised Fine-Tuning (KG-SFT). The key idea of KG-SFT is to generate high-quality explanations for each Q&A pair via a structured knowledge graph to enhance the knowledge comprehension and manipulation of LLMs. Specifically, KG-SFT consists of three components: Extractor, Generator, and Detector. For a given Q&A pair, (i) Extractor first identifies entities within Q&A pairs and extracts relevant reasoning subgraphs from external KGs, (ii) Generator then produces corresponding fluent explanations utilizing these reasoning subgraphs, and (iii) finally, Detector performs sentence-level knowledge conflicts detection on these explanations to guarantee the reliability. KG-SFT focuses on generating high-quality explanations to improve the quality of Q&A pair, which reveals a promising direction for supplementing existing data augmentation methods. Extensive experiments on fifteen different domains and six different languages demonstrate the effectiveness of KG-SFT, leading to an accuracy improvement of up to 18% and an average of 10% in low-data scenarios.", "keywords": [ "Large Language Models", "Knowledge Graph", "Supervised Fine-tuning" ], "rating_list": [ 6, 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "oLw4SH6r8h", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Ian Fischer", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Saurabh Singh", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" } ], "n_formula": 44, "n_formula_1": 20, "n_ref_uni": 35, "n_ref": 53, "n_ref_all": 123, "n_fig": 12, "n_tab": 2, "L_tab": 437, "n_element_tab": 49, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 270, "n_element_tab_1": 32, "formula_len_all": 4427, "formula_len_all_1": 1478, "len_all": 137519, "len_all_1": 56423, "len_abs": 1274, "len_title": 98, "len_sents": 38445, "len_sents_1": 25665, "n_sents": 337, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1242, "title": "Stochastic Sampling from Deterministic Flow Models", "abs": "Deterministic flow models such as rectified flows offer a general framework for learning a deterministic transport map between two distributions, realized as the vector field for an ordinary differential equation (ODE). However, they are sensitive to estimation and discretization errors and do not permit different samples conditioned on an intermediate state. We present a general method to turn the underlying ODE of such flow models into a family of stochastic differential equations (SDEs) that have the same marginal distributions. This method permits us to derive families of _stochastic samplers_, for fixed (e.g., previously trained) _deterministic_ flow models, that continuously span the spectrum of deterministic and stochastic sampling, given access to the flow field and the score function. Our method provides additional degrees of freedom that help alleviate some of the issues with the deterministic samplers and empirically outperforms them. We demonstrate this empirically on a toy Gaussian setup, as well as on the large scale ImageNet generation task. Further, our family of stochastic samplers provide an additional knob for controlling the diversity of generation, which we qualitatively demonstrate in our experiments.", "keywords": [ "rectified flow", "diffusion models", "stochastic sampling", "generative model" ], "rating_list": [ 8, 3, 3, 3 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 1, 1, 1 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "oK1zJCWBqf", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Arsalan Sharifnassab", "gender": "Male", "institution": "Openmind Research Institute", "country": "US", "position": "Researcher" }, { "name": "Dale Schuurmans", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Professor" }, { "name": "Saber Salehkaleybar", "gender": "Male", "institution": "Leiden University, Leiden University", "country": "NL", "position": "Assistant Professor" }, { "name": "Sina Ghiassian", "gender": "Male", "institution": "Spotify", "country": "", "position": "Researcher" }, { "name": "Surya Kanoria", "gender": "Male", "institution": "Spotify", "country": "", "position": "Senior Machine Learning Engineer" } ], "n_formula": 78, "n_formula_1": 29, "n_ref_uni": 36, "n_ref": 67, "n_ref_all": 90, "n_fig": 0, "n_tab": 6, "L_tab": 1550, "n_element_tab": 90, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 1130, "n_element_tab_1": 68, "formula_len_all": 10032, "formula_len_all_1": 2362, "len_all": 166434, "len_all_1": 70470, "len_abs": 859, "len_title": 127, "len_sents": 49543, "len_sents_1": 31454, "n_sents": 421, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 862, "title": "Soft Preference Optimization: Aligning Language Models to Expert Distributions", "abs": "We propose Soft Preference Optimization (SPO), a method for aligning generative models, such as Large Language Models (LLMs), with human preferences, without the need for a reward model. SPO optimizes model outputs directly over a preference dataset through a natural loss function that integrates preference loss with a regularization term across the model's entire output distribution rather than limiting it to the preference dataset. Although SPO does not require the assumption of an existing underlying reward model, we demonstrate that, under the Bradley-Terry (BT) model assumption, it converges to a softmax of scaled rewards, with the distribution's ``softness\" adjustable via the softmax exponent, an algorithm parameter. We showcase SPO's methodology, its theoretical foundation, and its comparative advantages in simplicity and alignment precision.", "keywords": [ "RLHF", "DPO", "Preference Alignment", "Large Language Models" ], "rating_list": [ 6, 5, 8, 5, 5 ], "soundness_list": [ 3, 2, 4, 1, 2 ], "presentation_list": [ 2, 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 4, 3, 3 ], "confidence_list": [ 4, 4, 3, 3, 3 ] }, { "paper_id": "oJgIRwkIUB", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Junwei Deng", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "PhD student" }, { "name": "Pingbang Hu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Xinhe Wang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Undergrad student" }, { "name": "Jiaqi Ma", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 42, "n_formula_1": 14, "n_ref_uni": 27, "n_ref": 55, "n_ref_all": 87, "n_fig": 2, "n_tab": 14, "L_tab": 4158, "n_element_tab": 380, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1019, "n_element_tab_1": 101, "formula_len_all": 4159, "formula_len_all_1": 1063, "len_all": 164522, "len_all_1": 73916, "len_abs": 3733, "len_title": 87, "len_sents": 52297, "len_sents_1": 33862, "n_sents": 410, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 39, "L_abs": 1618, "title": "Adversarial Attacks on Data Attribution", "abs": "Data attribution aims to quantify the contribution of individual training data points to the outputs of an AI model, which has been used to measure the value of training data and compensate data providers. Given the impact on financial decisions and compensation mechanisms, a critical question arises concerning the adversarial robustness of data attribution methods. However, there has been little to no systematic research addressing this issue. In this work, we aim to bridge this gap by detailing a threat model with clear assumptions about the adversary's goal and capabilities and proposing principled adversarial attack methods on data attribution. We present two methods, *Shadow Attack* and *Outlier Attack*, which generate manipulated datasets to inflate the compensation adversarially. The Shadow Attack leverages knowledge about the data distribution in the AI applications, and derives adversarial perturbations through \"shadow training\", a technique commonly used in membership inference attacks. In contrast, the Outlier Attack does not assume any knowledge about the data distribution and relies solely on black-box queries to the target model's predictions. It exploits an inductive bias present in many data attribution methods - outlier data points are more likely to be influential - and employs adversarial examples to generate manipulated datasets. Empirically, in image classification and text generation tasks, the Shadow Attack can inflate the data-attribution-based compensation by at least 200%, while the Outlier Attack achieves compensation inflation ranging from 185% to as much as 643%.", "keywords": [ "data attribution", "adversarial attack", "data-centric AI" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "oJLpXraSLb", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Adrian Weller", "gender": "Male", "institution": "Alan Turing Institute", "country": "", "position": "Programme Director for AI" }, { "name": "Isaac Reid", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Krzysztof Marcin Choromanski", "gender": "unknown", "institution": "Google Brain Robotics & Columbia University", "country": "", "position": "research scientist & adjunct assistant professor" }, { "name": "Richard E. Turner", "gender": "Male", "institution": "Alan Turing Institute", "country": "GB", "position": "Researcher" }, { "name": "Stratis Markou", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" } ], "n_formula": 102, "n_formula_1": 18, "n_ref_uni": 58, "n_ref": 153, "n_ref_all": 201, "n_fig": 15, "n_tab": 7, "L_tab": 6588, "n_element_tab": 264, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1243, "n_element_tab_1": 84, "formula_len_all": 6542, "formula_len_all_1": 1559, "len_all": 340729, "len_all_1": 83483, "len_abs": 957, "len_title": 96, "len_sents": 111620, "len_sents_1": 34945, "n_sents": 1133, "n_sents_1": 349, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 948, "title": "Variance-Reducing Couplings for Random Features", "abs": "Random features (RFs) are a popular technique to scale up kernel methods in machine learning, replacing exact kernel evaluations with stochastic Monte Carlo estimates. They underpin models as diverse as efficient transformers (by approximating attention) to sparse spectrum Gaussian processes (by approximating the covariance function). Efficiency can be further improved by speeding up the convergence of these estimates: a variance reduction problem. We tackle this through the unifying lens of optimal transport, finding couplings to improve RFs defined on both Euclidean and discrete input spaces. They enjoy theoretical guarantees and sometimes provide strong downstream gains, including for scalable inference on graphs. We reach surprising conclusions about the benefits and limitations of variance reduction as a paradigm, showing that other properties of the coupling should be optimised for attention estimation in efficient transformers.", "keywords": [ "Monte Carlo", "variance reduction", "quasi Monte Carlo", "transformers", "performers", "optimal transport", "random Fourier features", "graphs", "Gaussian processes", "kernels" ], "rating_list": [ 6, 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 2, 4, 3 ] }, { "paper_id": "oJA1GUqRww", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Fei Shen", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Researcher" }, { "name": "Kai Chen", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xin Jiang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yanan SUN", "gender": "Female", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Cai Rong Zhao", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Full Professor" }, { "name": "junyao gao", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "PhD student" }, { "name": "Xing Zhening", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 48, "n_ref": 88, "n_ref_all": 108, "n_fig": 12, "n_tab": 3, "L_tab": 663, "n_element_tab": 104, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 634, "n_element_tab_1": 102, "formula_len_all": 544, "formula_len_all_1": 544, "len_all": 130916, "len_all_1": 54765, "len_abs": 361, "len_title": 87, "len_sents": 27802, "len_sents_1": 23982, "n_sents": 207, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 39, "L_abs": 1485, "title": "Faceshot: Bring any Character into Life", "abs": "Portrait animation generates dynamic, realistic videos by mimicking facial expressions from a driven video.\nHowever, existing landmark-based methods are constrained by facial landmark detection and motion transfer limitations, resulting in suboptimal performance. In this paper, we present \\emph{FaceShot}, a novel training-free framework designed to animate any character from any driven video, human or non-human, with unprecedented robustness and stability.\nWe achieve this by offering precise and robust landmark results from an appearance-guided landmark matching module and a relative motion transfer module.\nTogether, these components harness the robust semantic correspondences of latent diffusion models to deliver landmarks across a wide range of character types, all without requiring fine-tuning or retraining.\nWith this powerful generalization capability, FaceShot can significantly extend the application of portrait animation by breaking the limitation of landmark detection for any character and driven video.\nFurthermore, FaceShot is compatible with any landmark-driven animation model, enhancing the realism and consistency of animations while significantly improving overall performance.\nExtensive experiments on our newly constructed character benchmark CABench confirm that FaceShot consistently surpasses state-of-the-art approaches across any character domain, setting a new standard for open-domain portrait animation. \nOur code will be made publicly available.", "keywords": [ "diffusion model" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 2, 4, 5 ] }, { "paper_id": "oIvjUpuZLC", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Chrysoula Kosma", "gender": "Female", "institution": "Paris-Saclay", "country": "FR", "position": "Postdoc" }, { "name": "Laurent Oudre", "gender": "Male", "institution": "Ecole Normale Superieure Paris Saclay", "country": "", "position": "Full Professor" }, { "name": "Thibaut Germain", "gender": "Male", "institution": "Ecole Normale Sup\u00e9rieure", "country": "FR", "position": "Postdoc" } ], "n_formula": 18, "n_formula_1": 19, "n_ref_uni": 60, "n_ref": 114, "n_ref_all": 146, "n_fig": 10, "n_tab": 13, "L_tab": 7522, "n_element_tab": 921, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 4436, "n_element_tab_1": 237, "formula_len_all": 1638, "formula_len_all_1": 1370, "len_all": 208139, "len_all_1": 74346, "len_abs": 1647, "len_title": 95, "len_sents": 61259, "len_sents_1": 31417, "n_sents": 397, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1664, "title": "Invariant Convolutional Layers for Time Series", "abs": "Machine learning for time series has recently garnered considerable attention. Indeed, automatically extracting meaningful representations from large and complex time series data is becoming imperative for several real-world applications. Neural architectures tailored to time series are often built upon sequential modules, such as convolutional, commonly employed in text or vision. Unfortunately, the potential of standard layers in capturing invariant properties of time series remains relatively underexplored. For instance, convolutional layers often fail to capture underlying patterns in time series inputs that encompass strong deformations, such as linear trends. However, invariances to some deformations may be critical for solving complex time series tasks, such as classification, while guaranteeing good generalization properties.\nTo address these challenges, we mathematically formulate and technically design efficient *invariant convolutions* for specific group actions applicable to the case of time series.\nWe construct these convolutions by considering two sets of deformations commonly observed in time series, including (i) *offset shift and scaling* and (ii) *linear trend and scaling*.\nWe further combine the proposed invariant convolutions with standard (or variant) convolutions in a single embedding layer of an example architecture, the so-called *InvConvNet* method, and showcase the layer capacity to capture complex invariant time series properties.\nFinally, *InvConvNet* is experimentally proven to achieve superior performance against common baselines in relevant time series tasks, including classification and anomaly detection.", "keywords": [ "Time Series", "Convolution", "Invariances", "Neural Network" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "oIWN7eMhTb", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Jie Feng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Jun Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Junbo Yan", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Tianhui Liu", "gender": "unknown", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Tianjian Ouyang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xin Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yuwei Du", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "\u90ed\u601d\u5947", "gender": "Female", "institution": " Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 16, "n_ref": 70, "n_ref_all": 105, "n_fig": 9, "n_tab": 12, "L_tab": 5990, "n_element_tab": 823, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1779, "n_element_tab_1": 96, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 198883, "len_all_1": 58612, "len_abs": 3503, "len_title": 127, "len_sents": 51424, "len_sents_1": 26771, "n_sents": 385, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1886, "title": "CityBench: Evaluating the Capabilities of Large Language Models for Urban Tasks", "abs": "Recently, large language models (LLMs) with extensive general knowledge and powerful reasoning abilities have seen rapid development and widespread application. A systematic and reliable evaluation of LLMs or visual language model (VLMs) is a crucial step in applying and developing them for various fields. There have been some early explorations about the usability of LLMs for limited urban tasks, but a systematic and scalable evaluation benchmark is still lacking. The challenge in constructing a systematic evaluation benchmark for urban research lies in the diversity of urban data, the complexity of application scenarios and the highly dynamic nature of the urban environment. In this paper, we design CityBench, an interactive simulator based evaluation platform, as the first systematic benchmark for evaluating the capabilities of LLMs for diverse tasks in urban research. First, we build CityData to integrate the diverse urban data and CitySimu to simulate fine-grained urban dynamics. Based on CityData and CitySimu, we design 8 representative urban tasks in 2 categories of perception-understanding and decision-making as the CityBench. With extensive results from 30 well-known LLMs and VLMs in 13 cities around the world, we find that advanced LLMs and VLMs can achieve competitive performance in diverse urban tasks requiring commonsense and semantic understanding abilities, e.g., understanding the human dynamics and semantic inference of urban images. Meanwhile, they fail to solve the challenging urban tasks requiring professional knowledge and high-level reasoning abilities, e.g., geospatial prediction and traffic control task. These observations provide valuable perspectives for utilizing and developing LLMs in the future. The dataset, benchmark and source codes are openly accessible to the research community via https://github.com/CityBench24/CityBench.", "keywords": [ "large language model", "urban science", "world model", "benchmark", "multi-modal" ], "rating_list": [ 10, 5, 6, 6, 8 ], "soundness_list": [ 4, 2, 3, 3, 3 ], "presentation_list": [ 4, 2, 2, 3, 4 ], "contribution_list": [ 4, 1, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 5, 3 ] }, { "paper_id": "oI5tZaWkF9", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Hsun-Yu Kuo", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "MS student" }, { "name": "Pu-Jen Cheng", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "Full Professor" }, { "name": "Yin-Hsiang Liao", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "PhD student" }, { "name": "Wei-Yun Ma", "gender": "Male", "institution": "Academia Sinica", "country": "TW", "position": "Associate Research Professor" }, { "name": "YU-CHIEH CHAO", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "MS student" } ], "n_formula": 28, "n_formula_1": 13, "n_ref_uni": 42, "n_ref": 90, "n_ref_all": 109, "n_fig": 5, "n_tab": 8, "L_tab": 2218, "n_element_tab": 200, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1642, "n_element_tab_1": 98, "formula_len_all": 1866, "formula_len_all_1": 1098, "len_all": 225061, "len_all_1": 63604, "len_abs": 953, "len_title": 135, "len_sents": 51480, "len_sents_1": 28542, "n_sents": 395, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 86, "L_abs": 973, "title": "Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification", "abs": "Synthetic data augmentation via large language models (LLMs) allows researchers to leverage additional training data, thus enhancing the performance of downstream tasks, especially when real-world data is scarce. However, the generated data can deviate from the real-world data, and this misalignment can bring deficient outcomes while applying the trained model to applications. Therefore, we proposed efficient weighted-loss approaches to align synthetic data with real-world distribution by emphasizing high-quality and diversified data generated by LLMs with using merely a little real-world data. We empirically assessed the effectiveness of our method on multiple text classification tasks, and the results showed leveraging our approaches on a BERT-level model robustly outperformed standard cross-entropy and other data weighting approaches, providing potential solutions to effectively leveraging synthetic data from any suitable data generator for model training.", "keywords": [ "data weighing", "data augmentation", "distillation", "data-efficient training", "NLP in resource-constrained settings", "fine-tuning", "weighted loss" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 3, 3, 2 ] }, { "paper_id": "oHbmiaeyUL", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Dohoon Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Jaehyun Park", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "PhD student" }, { "name": "Kyogu Lee", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Professor" }, { "name": "Hyunwoo Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 32, "n_formula_1": 23, "n_ref_uni": 25, "n_ref": 56, "n_ref_all": 87, "n_fig": 10, "n_tab": 11, "L_tab": 3294, "n_element_tab": 406, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 145, "n_element_tab_1": 18, "formula_len_all": 3017, "formula_len_all_1": 2321, "len_all": 147271, "len_all_1": 65548, "len_abs": 1398, "len_title": 111, "len_sents": 42432, "len_sents_1": 29770, "n_sents": 328, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1402, "title": "Multidimensional Trajectory Optimization for Flow and Diffusion", "abs": "In flow and diffusion-based generative modeling, conventional methods rely on unidimensional coefficients for the trajectory of differential equations. In this work, we first introduce a multidimensional coefficient that generalizes the conventional unidimensional coefficient into multiple dimensions. We also propose a new problem called multidimensional trajectory optimization, which suggests a novel trajectory optimality determined by the final transportation quality rather than predefined properties like straightness. Our approach pre-trains flow and diffusion models with various coefficients sampled from a hypothesis space and subsequently optimizes inference trajectories through adversarial training of a generator comprising the flow or diffusion model and the parameterized coefficient. To empirically validate our method, we conduct experiments on various generative models, including EDM and Stochastic Interpolant, across multiple datasets such as 2D synthetic datasets, CIFAR-10, FFHQ, and AFHQv2. Remarkably, inference using our optimized multidimensional trajectory achieves significant performance improvements with low NFE (e.g., 5), achieving state-of-the-art results in CIFAR-10 conditional generation. The introduction of multidimensional trajectory optimization enhances model efficiency and opens new avenues for exploration in flow and diffusion-based generative modeling.", "keywords": [ "Multidimensional Coefficient", "Multidimensional Trajectory Optimization", "Flow", "Diffusion", "Simulation Dynamics", "Adversarial Training" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "oHSXRy29tj", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Bingchen Fan", "gender": "unknown", "institution": "Beihang University", "country": "", "position": "MS student" }, { "name": "Pengyu Long", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Renjie Chen", "gender": "Male", "institution": "Beihang University", "country": "", "position": "MS student" }, { "name": "Yicheng Pan", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 72, "n_formula_1": 20, "n_ref_uni": 30, "n_ref": 59, "n_ref_all": 95, "n_fig": 5, "n_tab": 2, "L_tab": 189, "n_element_tab": 19, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10905, "formula_len_all_1": 1604, "len_all": 174351, "len_all_1": 74756, "len_abs": 857, "len_title": 0, "len_sents": 56823, "len_sents_1": 35185, "n_sents": 557, "n_sents_1": 304, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 860, "title": "Hierarchical overlapping clustering: cost function, algorithm and scalability", "abs": "Overlap and hierarchy are two prevalent phenomena in clustering, and usually coexist in a single system. There are several studies on each of them separately, but it is unclear how to characterize and evaluate the hybrid structures yet. To address this issue, we initiate the study of hierarchical overlapping clustering on graphs by introducing a new cost function for it. We show the rationality of our cost function via several intuitive properties, and develop an approximation algorithm that achieves a provably constant approximation factor for its dual version. Our algorithm is a recursive process of overlapping bipartition based on local search, which makes a speed-up version of it extremely scalable. Our experiments demonstrate that the speed-up algorithm has good performances in both effectiveness and scalability on synthetic and real datasets.", "keywords": [ "hierarchical overlapping clustering", "cost function", "approximation algorithm" ], "rating_list": [ 8, 5, 6, 3, 6 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 4, 2, 3, 2, 4 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "oHBS7R6JcP", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Chi Ma", "gender": "unknown", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Jingyang Xiang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "LiuChuan", "gender": "unknown", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Wei Lin", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Principal Researcher" }, { "name": "Ying Zhang", "gender": "Male", "institution": "Meituan", "country": "", "position": "Researcher" }, { "name": "Yong Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yujie Wang", "gender": "unknown", "institution": "Meituan", "country": "", "position": "Researcher" }, { "name": "yulei", "gender": "unknown", "institution": "Meituan", "country": "CN", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 33, "n_ref": 57, "n_ref_all": 85, "n_fig": 9, "n_tab": 17, "L_tab": 11626, "n_element_tab": 908, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 617, "n_element_tab_1": 69, "formula_len_all": 1516, "formula_len_all_1": 1516, "len_all": 151588, "len_all_1": 57505, "len_abs": 1910, "len_title": 118, "len_sents": 38933, "len_sents_1": 26195, "n_sents": 280, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1920, "title": "DuaRot: Dual Rotation for Advanced Outlier Mitigation in Rotated LLMs", "abs": "By employing rotation, outliers in activations can be effectively mitigated without altering the output, thereby facilitating the quantization of large language models (LLMs). However, existing rotation-based methods only consider global activation distributions, leaving the finer-grained distributions underexplored. Additionally, these methods predominantly rely on the Walsh\u2013Hadamard transform (WHT) to accelerate online rotation operations, while not fully considering performance between matrix multiplication~(Matmul) and WHT in actual runtime. These limitations hinder the rotation's ability to effectively reduce quantization errors and decrease inference speed. Therefore, improvements are needed in their performance regarding both accuracy and speed. In this paper, we propose a dual rotation method for rotation matrices, dubbed DuaRot, based on reparameterization. During training, DuaRot sequentially refines global and local features to achieve effective outlier mitigation. During inference, global and local rotations can be merged, which maintains rotational invariance without introducing additional computational overhead. Meanwhile, we propose a hardware-aware matrix configuration strategy, which determines whether the online Hadamard matrix should be expanded into a trainable parameter space by taking the runtime of the WHT and Matmul into account. This approach further enhances the reduction of quantization errors in online rotation operations without compromising inference speed. Extensive experiments demonstrate that DuaRot outperforms existing methods across various models and quantization configurations. For instance, when applied to LLaMA3-8B, DuaRot achieves WikiText-2 perplexities of 7.49 and 7.41 under W4A4KV4 and W4A4KV16 configurations with Round-to-Nearest (RTN), improving by 0.51 and 0.41 over the state-of-the-art, respectively. The code will be publicly available soon.", "keywords": [ "large language model", "Walsh\u2013Hadamard transform", "reparameterization", "hardware-aware", "rotational invariance" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 3, 4 ] }, { "paper_id": "oGsu3hksWT", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Chukwudi Paul Obite", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Shiwei Lan", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Associate Professor" }, { "name": "Shuang Zhou", "gender": "Female", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" }, { "name": "Zhi Chang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" } ], "n_formula": 41, "n_formula_1": 21, "n_ref_uni": 31, "n_ref": 77, "n_ref_all": 99, "n_fig": 4, "n_tab": 2, "L_tab": 1628, "n_element_tab": 347, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 847, "n_element_tab_1": 179, "formula_len_all": 5694, "formula_len_all_1": 3087, "len_all": 121825, "len_all_1": 56611, "len_abs": 1159, "len_title": 97, "len_sents": 30345, "len_sents_1": 23159, "n_sents": 245, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1167, "title": "Bayesian Learning with Deep Q-Exponential Process", "abs": "Motivated by deep neural networks, the deep Gaussian process (DGP) generalizes the standard GP by stacking multiple layers of GPs. Despite the enhanced expressiveness, GP, as an $L_2$ regularization prior, tends to be over-smooth and sub-optimal for inhomogeneous subjects, such as images with edges. Recently, Q-exponential process (Q-EP) has been proposed as an $L_q$ relaxation to GP and demonstrated with more desirable regularization properties through a parameter $q>0$ with $q=2$ corresponding to GP. Sharing the similar tractability of posterior and predictive distributions with GP, Q-EP can also be stacked to improve its modeling flexibility. In this paper, we generalize Q-EP to deep Q-EP to enjoy both proper regularization and improved expressiveness. The generalization is realized by introducing shallow Q-EP as a latent variable model and then building a hierarchy of the shallow Q-EP layers. Sparse approximation by inducing points and scalable variational strategy are applied to facilitate the inference. We demonstrate the numerical advantages of the proposed deep Q-EP model by comparing with multiple state-of-the-art deep probabilistic models.", "keywords": [ "Deep Models", "Inhomogeneous Subjects", "Regularization", "Latent Representation", "Model Expressiveness" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 2, 4, 3 ] }, { "paper_id": "oGrGnPndHw", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Anchi Chen", "gender": "Male", "institution": "Anhui University", "country": "CN", "position": "MS student" }, { "name": "Xu Yi", "gender": "Female", "institution": "Anhui University", "country": "CN", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 8, "n_ref": 13, "n_ref_all": 17, "n_fig": 2, "n_tab": 5, "L_tab": 3265, "n_element_tab": 81, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 4288, "n_element_tab_1": 92, "formula_len_all": 720, "formula_len_all_1": 722, "len_all": 91498, "len_all_1": 55933, "len_abs": 1467, "len_title": 136, "len_sents": 24522, "len_sents_1": 22524, "n_sents": 181, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1474, "title": "Autoencoder and Classifier based Joint-Guided Completion for Partial Multi-Modal Hashing", "abs": "The Multi-Modal Hashing (MMH) method based on complete modalities cannot effectively handle incomplete multi-modal samples, thus requiring the completion of missing modalities. Existing completion methods typically use complete modality samples with the same label to generate completion information. On one hand, they cannot fully utilize the different information between samples with different labels; on the other hand, they cannot effectively extract the global structural information of multi-modal samples. Therefore, we propose the autoencoder and classifier based joint-guided completion for partial multi-modal hashing (JCPMH) method that integrates autoencoders and classifiers. First, to fully utilize the different information between samples with different labels, we design a multi-modal classification module composed of multiple classifiers to learn different information. Second, we concatenate the multi-modal data into a whole and extract cross-modal global structural information through an autoencoder. Finally, based on the hashing module, multi-modal classification module and autoencoder module, we design a loss function to guide the generator to generate more accurate completion information for learning hash codes. JCPMH can utilize partial multi-modal samples for offline training and handle incomplete multi-modal samples during online retrieval. Additionally, we conducted extensive experiments to demonstrate the effectiveness of this model.", "keywords": [ "partial multi-modal hashing", "autoencoder", "classifier", "Joint-Guided Completion" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "oGYGjPsVWb", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Qiang Qiu", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "Taewook Kim", "gender": "Not Specified", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Wei Chen", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 17, "n_ref": 88, "n_ref_all": 106, "n_fig": 7, "n_tab": 3, "L_tab": 337, "n_element_tab": 62, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2281, "n_element_tab_1": 82, "formula_len_all": 897, "formula_len_all_1": 848, "len_all": 120485, "len_all_1": 68322, "len_abs": 1179, "len_title": 0, "len_sents": 35911, "len_sents_1": 30676, "n_sents": 277, "n_sents_1": 229, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1197, "title": "Learning to Customize Text-to-Image Diffusion In Diverse Context", "abs": "Most text-to-image customization techniques fine-tune models on a small set of \\emph{personal concept} images captured in minimal contexts. This often results in the model becoming overfitted to these training images and unable to generalize to new contexts in future text prompts. Existing customization methods are built on the success of effectively representing personal concepts as textual embeddings. Thus, in this work, we resort to diversifying the context of these personal concepts \\emph{solely} within the textual space by simply creating a contextually rich set of text prompts, together with a widely used self-supervised learning objective. Surprisingly, this straightforward and cost-effective method significantly improves semantic alignment in the textual space, and this effect further extends to the image space, resulting in higher prompt fidelity for generated images. Additionally, our approach does not require any architectural modifications, making it highly compatible with existing text-to-image customization methods. We demonstrate the broad applicability of our approach by combining it with four different baseline methods, achieving notable CLIP score improvements.", "keywords": [ "Text-to-Image Generation; Text-to-Image Customization; Subjet-Driven Text-To-Image Generation; Diffusion;" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "oFIU5CBY9p", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Boris van Breugel", "gender": "unknown", "institution": "Qualcomm Inc, QualComm", "country": "NL", "position": "Researcher" }, { "name": "Jonathan Crabb\u00e9", "gender": "Male", "institution": "Latent Labs", "country": "GB", "position": "Researcher" }, { "name": "Mihaela van der Schaar", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Qiyao Wei", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Rob Davis", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "Undergrad student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 22, "n_ref": 26, "n_ref_all": 45, "n_fig": 2, "n_tab": 6, "L_tab": 2458, "n_element_tab": 144, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1460, "n_element_tab_1": 97, "formula_len_all": 491, "formula_len_all_1": 360, "len_all": 107007, "len_all_1": 62124, "len_abs": 886, "len_title": 85, "len_sents": 37343, "len_sents_1": 27849, "n_sents": 290, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 890, "title": "LATABLE: TOWARDS LARGE TABULAR MODELS", "abs": "Tabular data is one of the most ubiquitous data modalities, yet the literature on tabular generative foundation models is lagging behind its text and vision counterparts. Large Tabular Models (LTMs) could revolutionize the way tabular data is used: not as any single dataset analyzed in a vacuum, but contextualized using their metadata and with respect to related datasets. Creating an LTM is difficult, due to the heterogeneous feature spaces of different tabular datasets, metadata, and prior knowledge. In this work, we propose LaTable: a novel tabular diffusion model that addresses these challenges. We show LaTable can be trained across tabular datasets. Through extensive experiments, we find that LaTable displays early signs of scaling laws previously encountered in foundation model regimes. Moreover, LaTable outperform baselines in out-of-distribution few-shot data generation.", "keywords": [ "Tabular data", "generative models" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "oFBu7qaZpS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Cehao Yang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Chengjin Xu", "gender": "Male", "institution": "International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Huaren Qu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Jian Guo", "gender": "Male", "institution": "International Digital Economy Academy, International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Jiaxin Mao", "gender": "Male", "institution": "Renmin University of China, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Muzhi Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shengjie Ma", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Xuhui Jiang", "gender": "Male", "institution": "International Digital Economy Academy, International Digital Economy Academy", "country": "CN", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 27, "n_ref": 43, "n_ref_all": 84, "n_fig": 7, "n_tab": 14, "L_tab": 3671, "n_element_tab": 263, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 357, "n_element_tab_1": 43, "formula_len_all": 455, "formula_len_all_1": 282, "len_all": 165583, "len_all_1": 66286, "len_abs": 1739, "len_title": 169, "len_sents": 59284, "len_sents_1": 31314, "n_sents": 410, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 121, "L_abs": 1749, "title": "Think-on-Graph 2.0: Deep and Faithful Large Language Model Reasoning with Knowledge-guided Retrieval Augmented Generation", "abs": "Retrieval-augmented generation (RAG) has improved large language models (LLMs) by using knowledge retrieval to overcome knowledge deficiencies. However, current RAG methods often fall short of ensuring the depth and completeness of retrieved information, which is necessary for complex reasoning tasks. In this work, we introduce Think-on-Graph 2.0 (ToG-2), a hybrid RAG framework that iteratively retrieves information from both unstructured and structured knowledge sources in a tight-coupling manner. Specifically, ToG-2 leverages knowledge graphs (KGs) to link documents via entities, facilitating deep and knowledge-guided context retrieval. Simultaneously, it utilizes documents as entity contexts to achieve precise and efficient graph retrieval. \nToG-2 alternates between graph retrieval and context retrieval to search for in-depth clues relevant to the question, enabling LLMs to generate answers.\nWe conduct a series of well-designed experiments to highlight the following advantages of ToG-2: 1) ToG-2 tightly couples the processes of context retrieval and graph retrieval, deepening context retrieval via the KG while enabling reliable graph retrieval based on contexts; 2) it achieves deep and faithful reasoning in LLMs through an iterative knowledge retrieval process of collaboration between contexts and the KG; and 3) ToG-2 is training-free and plug-and-play compatible with various LLMs. Extensive experiments demonstrate that ToG-2 achieves overall state-of-the-art (SOTA) performance on 6 out of 7 knowledge-intensive datasets with GPT-3.5, and can elevate the performance of smaller models (e.g., LLAMA-2-13B) to the level of GPT-3.5\u2019s direct reasoning. The source code is available on https://anonymous.4open.science/r/ToG2.", "keywords": [ "Retrieval Augmented Generation", "Knowledge Driven Reasoning" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "oF6e2WwxX0", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 12, "author_info_list": [ { "name": "Aiwei Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Albin Madappally Jose", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Haoping Bai", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Jiulong Shan", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Lijie Wen", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Meng Cao", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Philip S. Yu", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "US", "position": "Full Professor" }, { "name": "Xiang Kong", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Xiaojiang Liu", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Xiaoming Simon Wang", "gender": "Male", "institution": "Didi Research US", "country": "", "position": "Principal Engineer" }, { "name": "Yanchao Sun", "gender": "Female", "institution": "Apple AI/ML", "country": "", "position": "Researcher" }, { "name": "Zhiyun Lu", "gender": "Female", "institution": "Apple", "country": "", "position": "Researcher" } ], "n_formula": 84, "n_formula_1": 28, "n_ref_uni": 30, "n_ref": 69, "n_ref_all": 99, "n_fig": 7, "n_tab": 8, "L_tab": 2707, "n_element_tab": 430, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 15, "n_element_tab_1": 1, "formula_len_all": 10206, "formula_len_all_1": 3710, "len_all": 198199, "len_all_1": 54054, "len_abs": 1696, "len_title": 146, "len_sents": 56479, "len_sents_1": 22435, "n_sents": 513, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 98, "L_abs": 1565, "title": "TIS-DPO: Token-level Importance Sampling for Direct Preference Optimization With Estimated Weights", "abs": "Direct Preference Optimization (DPO) has been widely adopted for preference alignment of Large Language Models (LLMs) due to its simplicity and effectiveness. \nHowever, DPO is derived as a bandit problem in which the whole response is treated as a single arm, ignoring the importance differences between tokens, which may affect optimization efficiency and make it difficult to achieve optimal results.\nIn this work, we propose that the optimal data for DPO has equal expected rewards for each token in winning and losing responses, as there is no difference in token importance. \nHowever, since the optimal dataset is unavailable in practice, we propose using the original dataset for importance sampling to achieve unbiased optimization. \nAccordingly, we propose a token-level importance sampling DPO objective named TIS-DPO that assigns importance weights to each token based on its reward.\nInspired by previous works, we estimate the token importance weights using the difference in prediction probabilities from a pair of contrastive LLMs. We explore three methods to construct these contrastive LLMs: (1) guiding the original LLM with contrastive prompts, (2) training two separate LLMs using winning and losing responses, and (3) performing forward and reverse DPO training with winning and losing responses.\nExperiments show that TIS-DPO significantly outperforms various baseline methods on harmlessness and helpfulness alignment and summarization tasks. We also visualize the estimated weights, demonstrating their ability to identify key token positions.", "keywords": [ "Large Language Models", "Importance Sampling", "Preference Learning" ], "rating_list": [ 5, 5, 10, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 2, 4, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "oEMSM8HHpj", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Fan Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yingjing Huang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "LI YONG", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 37, "n_ref": 66, "n_ref_all": 79, "n_fig": 5, "n_tab": 3, "L_tab": 2353, "n_element_tab": 320, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2580, "n_element_tab_1": 328, "formula_len_all": 72, "formula_len_all_1": 72, "len_all": 129065, "len_all_1": 65083, "len_abs": 1783, "len_title": 104, "len_sents": 38421, "len_sents_1": 30518, "n_sents": 247, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1624, "title": "Spatiotemporal Contrast Are Natural Urban Scene Learners", "abs": "Street view imagery is a widely utilized representation of urban visual environments and supports various sustainable development tasks such as environmental perception and socio-economic assessment. However, it is challenging for existing image representations to specifically encode the dynamic urban environment (such as pedestrians, vehicles, and vegetation), the built environment (including buildings, roads, and urban infrastructure), and the environmental ambiance (such as the cultural and socioeconomic atmosphere) depicted in street view imagery to address downstream tasks related to the city.\nThis work innovatively leverages temporal and spatial attributes of street view imagery to propose an unsupervised learning framework suitable for diverse downstream tasks. By employing street view images captured at the same location over time and spatially nearby views at the same time, we construct contrastive learning tasks designed to learn the temporal-invariant characteristics of the built environment and the spatial-invariant neighborhood ambiance. Our approach significantly outperforms traditional supervised and unsupervised methods in tasks such as visual place recognition, socioeconomic estimation, and human-environment perception. Moreover, we demonstrate the varying behaviors of image representations learned through different contrastive learning strategies across various downstream tasks. This study systematically discusses representation learning strategies for urban studies based on street view images, providing a benchmark that enhances the applicability of visual data in urban science.", "keywords": [ "contrastive learning", "self-supervied learning", "street view images" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "oDbiL9CLoS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Yuanzhi Li", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Zeyuan Allen-Zhu", "gender": "Unspecified", "institution": "Meta FAIR Labs", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 2, "n_ref": 17, "n_ref_all": 64, "n_fig": 17, "n_tab": 6, "L_tab": 3477, "n_element_tab": 53, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1146, "n_element_tab_1": 11, "formula_len_all": 200, "formula_len_all_1": 0, "len_all": 192529, "len_all_1": 82608, "len_abs": 1258, "len_title": 89, "len_sents": 73605, "len_sents_1": 39912, "n_sents": 579, "n_sents_1": 309, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1311, "title": "Physics of Language Models: Part 3.2, Knowledge Manipulation", "abs": "Language models can store vast factual knowledge, yet their ability to flexibly use this knowledge for downstream tasks (e.g., via instruction finetuning) remains questionable. This paper investigates four fundamental knowledge manipulation tasks: \\textbf{retrieval} (e.g., \"What is person A's attribute X?\"), \\textbf{classification} (e.g., \"Is A's attribute X even or odd?\"), \\textbf{comparison} (e.g., \"Is A greater than B in attribute X?\"), and \\textbf{inverse search} (e.g., \"Which person's attribute X equals T?\").\n\nWe show that language models excel in knowledge retrieval but struggle even in the simplest classification or comparison tasks unless Chain of Thoughts (CoTs) are employed during both training and inference. Moreover, their performance in inverse knowledge search is virtually 0\\%, regardless of the prompts.\nOur primary contribution is a \\emph{controlled, synthetic experiment} that confirms these weaknesses are \\emph{inherent} to language models: they cannot efficiently manipulate knowledge from pre-training data, even when such knowledge is perfectly stored in the models, despite adequate training and sufficient model size. Our findings also apply to modern pretrained language models such as GPT-4, thus giving rise to many Turing tests to distinguish Humans from contemporary AIs.", "keywords": [ "knowledge manipulation", "language models", "generative models", "reversal curse" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 4, 4, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "oClr2P7V0T", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Jizhe Zhang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Stefanie Jegelka", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Yifei Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yisen Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 42, "n_ref": 72, "n_ref_all": 123, "n_fig": 11, "n_tab": 5, "L_tab": 1248, "n_element_tab": 92, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 1397, "n_element_tab_1": 72, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 139213, "len_all_1": 61865, "len_abs": 1912, "len_title": 109, "len_sents": 44850, "len_sents_1": 29713, "n_sents": 308, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1928, "title": "Are Synthetic Classifiers Really as Good as Real Classifiers?", "abs": "Foundation models have achieved significant advancements across various domains, yet their training demands vast amounts of real-world data, which is becoming increasingly scarce. To address this challenge, synthetic data has garnered substantial interest as an alternative for augmenting training datasets in fields such as computer vision and natural language processing. However, skepticism remains regarding whether synthetic classifiers can match the performance of those trained on real data. In this paper, we investigate this question by conducting a detailed analysis within the realm of visual tasks, comparing classifiers trained on synthetic versus real data using CLIP and ViT. Our results reveal that synthetic classifiers exhibit deficiencies in a range of challenging real-world scenarios, such as fine-grained classification, extreme object scales and extreme brightness despite achieving comparable overall accuracy to their real-data-trained counterparts. We find that the limitations of synthetic classifiers can be traced back to the limitations of current generative models in capturing the complexity and diversity of real-world data in these aspects. To mitigate these issues efficiently, we explore \\textbf{RealTune}, a simple method that enhances synthetic classifiers by finetuning them with a small amount of real data. Experimental evaluations demonstrate that RealTune significantly improves the performance of synthetic classifiers using only a limited real dataset (e.g., 40k images, 3% of ImageNet) with minimal training time (e.g., 1hour on a single NVIDIA RTX 3090 GPU). Our findings indicate that while synthetic data is a valuable resource, integrating real and synthetic data is essential to achieve robust and efficient classifiers. This work underscores the necessity of leveraging both data types to bridge the performance gap and enhance the overall effectiveness of foundation models.", "keywords": [ "generative model", "representation learning", "synthetic data" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "oCdIo9757e", "primary_area": "causal reasoning", "n_author": 9, "author_info_list": [ { "name": "Anton van den Hengel", "gender": "unknown", "institution": "University of Adelaide", "country": "AU", "position": "Professor" }, { "name": "Biwei Huang", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Dong Gong", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "Assistant Professor" }, { "name": "Ignavier Ng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Mingming Gong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Yuhang Liu", "gender": "unknown", "institution": "The University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Zhen Zhang", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Javen Shi", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Professor" } ], "n_formula": 71, "n_formula_1": 41, "n_ref_uni": 38, "n_ref": 102, "n_ref_all": 118, "n_fig": 1, "n_tab": 11, "L_tab": 22427, "n_element_tab": 1021, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 3424, "n_element_tab_1": 282, "formula_len_all": 4708, "formula_len_all_1": 2889, "len_all": 175442, "len_all_1": 65728, "len_abs": 1190, "len_title": 56, "len_sents": 44038, "len_sents_1": 25087, "n_sents": 330, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1232, "title": "Analytic DAG Constraints for Differentiable DAG Learning", "abs": "Recovering underlying Directed Acyclic Graph (DAG) structures from observational data presents a formidable challenge due to the combinatorial nature of the DAG-constrained optimization problem. Recently, researchers have identified gradient vanishing as one of the primary obstacles in differentiable DAG learning and have proposed several DAG constraints to mitigate this issue. By developing the necessary theory to establish a connection between analytic functions and DAG constraints, we demonstrate that analytic functions from the set \n$\\\\{f(x) = c_0 + \\sum_{i=1}c_ix^i|c_0 \\geqslant 0; \\forall i > 0, c_i > 0; r = \\lim_{i\\rightarrow \\infty}c_{i}/c_{i+1} > 0\\\\}$\ncan be employed to formulate effective DAG constraints. Furthermore, we establish that this set of functions is closed under several functional operators, including differentiation, summation, and multiplication. Consequently, these operators can be leveraged to create novel DAG constraints based on existing ones. Using these properties, we designed a series of DAG constraints and designed an efficient algorithm to evaluate these DAG constraints. Experiments on various settings show that our DAG constraints outperform previous state-of-the-arts approaches.", "keywords": [ "DAG", "Causal Discovery", "Structural Learning" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "oCUYc7BzXQ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Alexander Cong Li", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Ananya Kumar", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Researcher" }, { "name": "Deepak Pathak", "gender": "Male", "institution": "Skild AI", "country": "US", "position": "CEO" } ], "n_formula": 7, "n_formula_1": 9, "n_ref_uni": 57, "n_ref": 106, "n_ref_all": 138, "n_fig": 14, "n_tab": 4, "L_tab": 1677, "n_element_tab": 134, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 722, "n_element_tab_1": 53, "formula_len_all": 5229, "formula_len_all_1": 6615, "len_all": 176613, "len_all_1": 82516, "len_abs": 1134, "len_title": 95, "len_sents": 51821, "len_sents_1": 37897, "n_sents": 353, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1135, "title": "Generative Classifiers Avoid Shortcut Solutions", "abs": "Discriminative approaches to classification often learn shortcuts that hold in-distribution but fail even under minor distribution shift. This failure mode stems from an overreliance on features that are spuriously correlated with the label. We show that generative classifiers, which use class-conditional generative models, avoid this issue by modeling all features, both core and spurious, instead of mainly spurious ones. These generative classifiers are simple to train, avoiding the need for specialized augmentations, strong regularization, extra hyperparameters, or knowledge of the specific spurious correlations to avoid. We find that diffusion-based and autoregressive generative classifiers achieve state-of-the-art performance on standard image and text distribution shift benchmarks and reduce the impact of spurious correlations present in realistic applications, such as medical or satellite datasets. Finally, we carefully analyze a Gaussian toy setting to understand the inductive biases of generative classifiers, as well as the data properties that affect when generative classifiers outperform discriminative ones.", "keywords": [ "distribution shift", "shortcut", "generative models", "robustness" ], "rating_list": [ 8, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 4, 1, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "oCIEUHJjNj", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Andrew Tao", "gender": "Male", "institution": "NVIDIA Corporation", "country": "US", "position": "Vice President of Applied Research" }, { "name": "Arushi Goel", "gender": "Female", "institution": "NVIDIA Research", "country": "", "position": "Researcher" }, { "name": "Bryan Catanzaro", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Vice President" }, { "name": "Karan Sapra", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Matthieu Le", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Rafael Valle", "gender": "Not Specified", "institution": "NVIDIA", "country": "", "position": "Senior Research Scientist" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 49, "n_ref": 194, "n_ref_all": 230, "n_fig": 7, "n_tab": 7, "L_tab": 2498, "n_element_tab": 301, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2437, "n_element_tab_1": 278, "formula_len_all": 626, "formula_len_all_1": 336, "len_all": 184775, "len_all_1": 76467, "len_abs": 1374, "len_title": 85, "len_sents": 52902, "len_sents_1": 33311, "n_sents": 405, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 1380, "title": "OMCAT: Omni Context Aware Transformer", "abs": "Large Language Models (LLMs) have made significant strides in text generation and comprehension, with recent advancements extending into multimodal LLMs that integrate visual and audio inputs. However, these models continue to struggle with fine-grained, cross-modal temporal understanding, particularly when correlating events across audio and video streams. We address these challenges with two key contributions: a new dataset and model, called OCTAV and OMCAT respectively. OCTAV (Omni Context and Temporal Audio Video) is a novel dataset designed to capture event transitions across audio and video. Second, OMCAT (Omni Context Aware Transformer) is a powerful model that leverages RoTE (Rotary Time Embeddings), an innovative extension of RoPE, to enhance temporal grounding and computational efficiency in time-anchored tasks. Through a robust three-stage training pipeline\u2014feature alignment, instruction tuning, and OCTAV-specific training\u2014OMCAT excels in cross-modal temporal understanding. Our model demonstrates state-of-the-art performance on Audio-Visual Question Answering (AVQA) tasks and the OCTAV benchmark, showcasing significant gains in temporal reasoning and cross-modal alignment, as validated through comprehensive experiments and ablation studies. Our dataset and code will be made publicly available. The link to our demo page is https://om-cat.github.io.", "keywords": [ "Multimodal LLMs", "Audio-Video Understanding", "Cross-modal" ], "rating_list": [ 5, 6, 5, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "oCHsDpyawq", "primary_area": "datasets and benchmarks", "n_author": 18, "author_info_list": [ { "name": "Alex Bo-Yuan Chen", "gender": "Male", "institution": "HHMI Janelia Research Campus", "country": "", "position": "PhD student" }, { "name": "Alexander Immer", "gender": "Male", "institution": "Bioptimus", "country": "FR", "position": "Researcher" }, { "name": "Alyson Petruncio", "gender": "Female", "institution": "HHMI Janelia Research Campus", "country": "US", "position": "Researcher" }, { "name": "Aparna Dev", "gender": "Female", "institution": "HHMI Janelia Research Campus", "country": "US", "position": "Researcher" }, { "name": "Aubrey Weigel", "gender": "Female", "institution": "HHMI Janelia Research Campus", "country": "US", "position": "Researcher" }, { "name": "Florian Engert", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Professor" }, { "name": "Gudrun Ihrke", "gender": "Female", "institution": "HHMI Janelia Research Campus", "country": "", "position": "Researcher" }, { "name": "Jan-Matthis Lueckmann", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Jeff Lichtman", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Full Professor" }, { "name": "Luuk Willem Hesselink", "gender": "unknown", "institution": "Donders Institute for Brain, Cognition and Behaviour", "country": "NL", "position": "PhD student" }, { "name": "Mariela D Petkova", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Postdoc" }, { "name": "Michal Januszewski", "gender": "Unspecified", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Misha Ahrens", "gender": "Male", "institution": "HHMI Janelia Research Campus", "country": "US", "position": "Senior Group Leader" }, { "name": "Nirmala A Iyer", "gender": "Female", "institution": "HHMI Janelia Research Campus", "country": "US", "position": "Researcher" }, { "name": "Peter H. Li", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Viren Jain", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Woohyun Park", "gender": "unknown", "institution": "HHMI Janelia Research Campus", "country": "US", "position": "Researcher" }, { "name": "Wyatt Korff", "gender": "Male", "institution": "HHMI Janelia Research Campus", "country": "US", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 35, "n_ref": 72, "n_ref_all": 105, "n_fig": 18, "n_tab": 2, "L_tab": 1186, "n_element_tab": 207, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 359, "n_element_tab_1": 227, "formula_len_all": 453, "formula_len_all_1": 411, "len_all": 227911, "len_all_1": 64415, "len_abs": 1004, "len_title": 118, "len_sents": 60371, "len_sents_1": 30261, "n_sents": 498, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 70, "L_abs": 1006, "title": "ZAPBench: A Benchmark for Whole-Brain Activity Prediction in Zebrafish", "abs": "Data-driven benchmarks have led to significant progress in key scientific modeling domains including weather and structural biology. Here, we introduce the Zebrafish Activity Prediction Benchmark (ZAPBench) to measure progress on the problem of predicting cellular-resolution neural activity throughout an entire vertebrate brain. The benchmark is based on a novel dataset containing 4d light-sheet microscopy recordings of over 70,000 neurons in a larval zebrafish brain, along with motion stabilized and voxel-level cell segmentations of these data that facilitate development of a variety of forecasting methods. Initial results from a selection of time series and volumetric video modeling approaches achieve better performance than naive baseline methods, but also show room for further improvement. The specific brain used in the activity recording is also undergoing synaptic-level anatomical mapping, which will enable future integration of detailed structural information into forecasting methods.", "keywords": [ "neuroscience", "zebrafish", "forecasting", "benchmark", "timeseries", "lightsheet microscopy", "calcium imaging" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 4, 4, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "oBmaLuEJda", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Peijie Huang", "gender": "Male", "institution": "South China Agricultural University", "country": "CN", "position": "Associate Professor" }, { "name": "Shangjian Yin", "gender": "Male", "institution": "South China Agricultural University", "country": "CN", "position": "Undergrad student" }, { "name": "Xinming Chen", "gender": "Male", "institution": "South China Agricultural University", "country": "", "position": "MS student" }, { "name": "Yuhong Xu", "gender": "unknown", "institution": "South China Agricultural University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 21, "n_ref": 60, "n_ref_all": 66, "n_fig": 4, "n_tab": 4, "L_tab": 1680, "n_element_tab": 209, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1486, "n_element_tab_1": 110, "formula_len_all": 361, "formula_len_all_1": 361, "len_all": 112794, "len_all_1": 52861, "len_abs": 1479, "len_title": 153, "len_sents": 26405, "len_sents_1": 24242, "n_sents": 173, "n_sents_1": 152, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1484, "title": "BMLM: Bidirectional Large Language Model for Multi-Task Spoken Language Understanding: Better and Faster", "abs": "Autoregressive large language models (LLMs) have achieved notable success in natural language generation. However, their direct application to natural language understanding (NLU) tasks presents challenges due to reliance on fixed label vocabularies and task-specific output structures. Although instruction-following tuning can adapt LLMs for these tasks, the autoregressive architecture often leads to error propagation and significant time costs from uncontrollable output lengths, particularly in token-level tagging tasks. In this paper, we introduce a bidirectional LLM framework (BMLM) for multi-task spoken language understanding, which eliminates the need for training from scratch and seamlessly integrates with existing LLMs, bridging the gap between extensive pre-trained knowledge and the requirements of understanding tasks. Our evaluation on multiple datasets demonstrates that BMLM significantly outperforms state-of-the-art pre-trained language models and autoregressive LLM baselines. Specifically, on the MixATIS and MixSNIPS datasets, BMLM achieves notable improvements of +3.9\\% and +4.1\\% in overall semantic accuracy compared to autoregressive baselines. Additionally, we observe a 123x improvement in inference speed for the MixATIS dataset and a 189x enhancement for the MixSNIPS dataset compared to existing generative LLM baselines. We anticipate that this work will provide a new perspective and foundational support for LLM applications in the NLU domain.", "keywords": [ "Spoken Language Understanding", "Multi-Task Learning", "Large Language Model" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "oBHF3urgyS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Amrita Bhattacharjee", "gender": "Female", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Durgesh Kalwar", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Lin Guan", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Siddhant Bhambri", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Subbarao Kambhampati", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "Full Professor" }, { "name": "huan liu", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 63, "n_ref": 109, "n_ref_all": 128, "n_fig": 7, "n_tab": 2, "L_tab": 1045, "n_element_tab": 36, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 141651, "len_all_1": 70114, "len_abs": 1616, "len_title": 93, "len_sents": 38863, "len_sents_1": 33675, "n_sents": 212, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1632, "title": "Extracting Heuristics from Large Language Models for Reward Shaping in Reinforcement Learning", "abs": "Reinforcement Learning (RL) suffers from sample inefficiency in sparse reward domains, and the problem is further pronounced in case of stochastic transitions. To improve the sample efficiency, reward shaping is a well-studied approach to introduce intrinsic rewards that can help the RL agent converge to an optimal policy faster. However, designing a useful reward shaping function for all desirable states in the Markov Decision Process (MDP) is challenging, even for domain experts. Given that Large Language Models (LLMs) have demonstrated impressive performance across a magnitude of natural language tasks, we aim to answer the following question: $\\textit{Can we obtain heuristics using LLMs for constructing a reward shaping function that can boost an RL agent's sample efficiency?}$ To this end, we aim to leverage off-the-shelf LLMs to generate a plan for an abstraction of the underlying MDP. We further use this LLM-generated plan as a heuristic to construct the reward shaping signal for the downstream RL agent. By characterizing the type of abstraction based on the MDP horizon length, we analyze the quality of heuristics when generated using an LLM, with and without a verifier in the loop. Our experiments across multiple domains with varying horizon length and number of sub-goals from the BabyAI environment suite, Household, Mario, and, Minecraft domain, show 1) the advantages and limitations of querying LLMs with and without a verifier to generate a reward shaping heuristic, and, 2) a significant improvement in the sample efficiency of PPO, A2C, and Q-learning when guided by the LLM-generated heuristics.", "keywords": [ "Reinforcement Learning", "Sparse Rewards", "Large Language Models" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "oApCZZZ3O4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jingwei Sun", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Yiran Chen", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Professor" }, { "name": "Zhixu Du", "gender": "Male", "institution": "Duke University, Duke University", "country": "US", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 5, "n_ref_uni": 38, "n_ref": 70, "n_ref_all": 84, "n_fig": 3, "n_tab": 6, "L_tab": 2338, "n_element_tab": 212, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1924, "n_element_tab_1": 155, "formula_len_all": 1655, "formula_len_all_1": 1155, "len_all": 126812, "len_all_1": 64072, "len_abs": 2028, "len_title": 143, "len_sents": 35230, "len_sents_1": 28310, "n_sents": 262, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1608, "title": "Knowledge Graph Tuning: Real-time Large Language Model Personalization based on Human Feedback", "abs": "Large language models (LLMs) have demonstrated remarkable proficiency in a range of natural language processing tasks. Once deployed, LLMs encounter users with personalized factual knowledge, and such personalized knowledge is consistently reflected through users' interactions with the LLMs. To enhance user experience, real-time model personalization is essential, allowing LLMs to adapt user-specific knowledge based on user feedback during human-LLM interactions. Existing methods mostly require back-propagation to finetune the model parameters, which incurs high computational and memory costs. In addition, these methods suffer from low interpretability, which will cause unforeseen impacts on model performance during long-term use, where the user's personalized knowledge is accumulated extensively. To address these challenges, we propose Knowledge Graph Tuning (KGT), a novel approach that leverages knowledge graphs (KGs) to personalize LLMs. KGT extracts personalized factual knowledge triples from users' queries and feedback and optimizes KGs without modifying the LLM parameters. Our method improves computational and memory efficiency by avoiding back-propagation and ensures interpretability by making the KG adjustments comprehensible to humans. Experiments with state-of-the-art LLMs, including GPT-2, Llama2, and Llama3, show that KGT significantly improves personalization performance while reducing latency and GPU memory costs. Ultimately, KGT offers a promising solution of effective, efficient, and interpretable real-time LLM personalization during user interactions with the LLMs.", "keywords": [ "Large Language Model", "model personalization" ], "rating_list": [ 6, 3, 1, 5, 6 ], "soundness_list": [ 3, 2, 1, 2, 2 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 1, 2, 3 ], "confidence_list": [ 3, 4, 5, 2, 5 ] }, { "paper_id": "oANkBaVci5", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Hui Chen", "gender": "Female", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Lopamudra Mukherjee", "gender": "Female", "institution": "University of Wisconsin-Whitewater", "country": "", "position": "Associate Professor" }, { "name": "Viet Luong", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Undergrad student" }, { "name": "Vikas Singh", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Professor" } ], "n_formula": 24, "n_formula_1": 16, "n_ref_uni": 38, "n_ref": 69, "n_ref_all": 94, "n_fig": 5, "n_tab": 6, "L_tab": 4771, "n_element_tab": 489, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 3573, "n_element_tab_1": 256, "formula_len_all": 1158, "formula_len_all_1": 830, "len_all": 177154, "len_all_1": 73186, "len_abs": 1486, "len_title": 107, "len_sents": 52927, "len_sents_1": 30325, "n_sents": 398, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1495, "title": "A Simple Baseline for Multivariate Time Series Forecasting", "abs": "The versatility of large language models has led to intensive ongoing work focused on adaptations to other modalities. This can involve moderate modifications of an existing model, piggybacking on the language model's capabilities to train multimodal models or even starting with pre-trained checkpoints and attaching specialized adapters to recast a new modality (e.g., time-series) as ``language''. This latter approach, prominent in a growing set of nice results, yields strong performance across benchmarks. It also makes sense -- while a large amount of temporal data is acquired every day (e.g., wearable sensors, physiological measurements in healthcare), unlike text/image corpus, much of it is not publicly available (except financial markets) for various reasons. But training (or even fine-tuning) these large models is expensive or difficult with limited resources. In this paper, we study and characterize the performance profile of a simple model for multivariate time-series forecasting. By simple, we mean that the model is restricted to tokenization based on classical ideas (as has been shown to be effective in vision) which are then allowed to attend/interact: via self-attention but also via ways that are a bit more general than dot-product attention, accomplished via basic geometric algebra ideas. We show that even a single or two layer model yields results that are competitive with much bigger (and even LLM-based) models on most benchmarks reported in the literature.", "keywords": [ "Time Series Forecasting", "Wavelets" ], "rating_list": [ 8, 6, 5, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "oAGSLx4VEs", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Guanlin Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Han Qiu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Zhang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Jiwei Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Shangwei Guo", "gender": "Male", "institution": "Chongqing University", "country": "CN", "position": "Associate Professor" }, { "name": "Tianwei Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Yifei Chen", "gender": "unknown", "institution": "Chongqing University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 41, "n_ref": 120, "n_ref_all": 161, "n_fig": 14, "n_tab": 10, "L_tab": 4433, "n_element_tab": 763, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2947, "n_element_tab_1": 598, "formula_len_all": 532, "formula_len_all_1": 502, "len_all": 198344, "len_all_1": 91132, "len_abs": 1765, "len_title": 120, "len_sents": 62607, "len_sents_1": 37265, "n_sents": 580, "n_sents_1": 321, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1778, "title": "Warfare: Breaking the Watermark Protection of AI-Generated Content", "abs": "AI-Generated Content (AIGC) is gaining great popularity, with many emerging commercial services and applications. These services leverage advanced generative models, such as latent diffusion models and large language models, to generate creative content (e.g., realistic images and fluent sentences) for users. The usage of such generated content needs to be highly regulated, as the service providers need to ensure the users do not violate the usage policies (e.g., abuse for commercialization, generating and distributing unsafe content). \nA promising solution to achieve this goal is watermarking, which adds unique and imperceptible watermarks on the content for service verification and attribution. Numerous watermarking approaches have been proposed recently. However, in this paper, we show that an adversary can easily break these watermarking mechanisms. Specifically, we consider two possible attacks. (1) Watermark removal: the adversary can easily erase the embedded watermark from the generated content and then use it freely bypassing the regulation of the service provider. (2) Watermark forging: the adversary can create illegal content with forged watermarks from another user, causing the service provider to make wrong attributions. We propose Warfare, a unified methodology to achieve both attacks in a holistic way. The key idea is to leverage a pre-trained diffusion model for content processing and a generative adversarial network for watermark removal or forging. We evaluate Warfare on different datasets and embedding setups. The results prove that it can achieve high success rates while maintaining the quality of the generated content. Compared to the inference process of existing diffusion model-based attacks, Warfare is 5,050~11,000x faster.", "keywords": [ "AIGC", "Content Watermark", "Watermark Remove", "Watermark Forge" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "oA5GmyvMUY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Darine Ameyed", "gender": "Female", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "Principal Researcher" }, { "name": "Fehmi Jaafar", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Mohamed Cheriet", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "Full Professor" }, { "name": "Riadh Ben Chaabene", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 28, "n_ref": 30, "n_ref_all": 34, "n_fig": 0, "n_tab": 1, "L_tab": 604, "n_element_tab": 28, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 604, "n_element_tab_1": 28, "formula_len_all": 90, "formula_len_all_1": 90, "len_all": 105888, "len_all_1": 64797, "len_abs": 1405, "len_title": 147, "len_sents": 26831, "len_sents_1": 26834, "n_sents": 192, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1412, "title": "Robust Federated Learning Frameworks Guarding Against Data Flipping Threats for Autonomous Vehicles", "abs": "Federated Learning (FL) has become an established technique to facilitate privacy-preserving collaborative training across a multitude of clients. The ability to achieve collaborative learning from multiple parties containing an extensive volume of data while providing the essence of data privacy made it an attractive solution to address numerous challenges in sensitive data-driven fields such as autonomous vehicles (AVs). However, its decentralized nature exposes it to security threats, such as evasion and data poisoning attacks, where malicious participants can compromise training data. This paper addresses the challenge of defending federated learning systems against data poisoning attacks specifically focusing on data-flipping techniques in AVs by proposing a novel defense mechanism that combines anomaly detection with robust aggregation techniques. Our approach employs statistical outlier detection and model-based consistency checks to filter out compromised updates before they affect the global model. Experiments on benchmark datasets show that our method significantly enhances robustness by preventing nearly 15\\% of accuracy drop for our global model when confronted with a malicious participant and reduction the the attack success rate even when dealing with 20\\% of poisoning level. These findings provide a comprehensive solution to strengthen FL systems against adversarial threats.", "keywords": [ "Federated Learning", "Data Poisoning", "Adversarial Attack" ], "rating_list": [ 3, 3, 1, 3, 3 ], "soundness_list": [ 1, 2, 1, 2, 2 ], "presentation_list": [ 1, 2, 1, 1, 1 ], "contribution_list": [ 2, 1, 1, 1, 2 ], "confidence_list": [ 5, 5, 5, 4, 4 ] }, { "paper_id": "o9kqa5K3tB", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Albert Gu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Andrej Risteski", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Ricardo Buitrago", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Tanya Marwah", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 27, "n_ref_uni": 56, "n_ref": 130, "n_ref_all": 174, "n_fig": 12, "n_tab": 6, "L_tab": 1981, "n_element_tab": 203, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2894, "formula_len_all_1": 2272, "len_all": 222000, "len_all_1": 58730, "len_abs": 1301, "len_title": 106, "len_sents": 68235, "len_sents_1": 25495, "n_sents": 515, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 58, "L_abs": 1319, "title": "On the Benefits of Memory for Modeling Time-Dependent PDEs", "abs": "Data-driven techniques have emerged as a promising alternative to traditional numerical methods. For time-dependent PDEs, many approaches are Markovian---the evolution of the trained system only depends on the current state, and not the past states. In this work, we investigate the benefits of using memory for modeling time-dependent PDEs: that is, when past states are explicitly used to predict the future. Motivated by the Mori-Zwanzig theory of model reduction, we theoretically exhibit examples of simple (even linear) PDEs, in which a solution that uses memory is arbitrarily better than a Markovian solution. Additionally, we introduce Memory Neural Operator (MemNO), a neural operator architecture that combines recent state space models (specifically, S4) and Fourier Neural Operators (FNOs) to effectively model memory. We empirically demonstrate that when the PDEs are supplied in low resolution or contain observation noise at train and test time, MemNO significantly outperforms the baselines without memory---with up to $6 \\times$ reduction in test error. Furthermore, we show that this benefit is particularly pronounced when the PDE solutions have significant high-frequency Fourier modes (e.g., low-viscosity fluid dynamics) and we construct a challenging benchmark dataset consisting of such PDEs.", "keywords": [ "State Space Models", "Partial Differential Equations" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "o9icFTlKi4", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Chang D. Yoo", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "haeyong kang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Postdoc" } ], "n_formula": 40, "n_formula_1": 16, "n_ref_uni": 57, "n_ref": 150, "n_ref_all": 166, "n_fig": 4, "n_tab": 14, "L_tab": 8232, "n_element_tab": 1120, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 74, "n_element_tab_1": 6, "formula_len_all": 3582, "formula_len_all_1": 1349, "len_all": 194929, "len_all_1": 61455, "len_abs": 1376, "len_title": 40, "len_sents": 41783, "len_sents_1": 27887, "n_sents": 303, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 40, "L_abs": 1100, "title": "Soft-TransFormers for Continual Learning", "abs": "Inspired by Well-initialized Lottery Ticket Hypothesis (WLTH), which provides suboptimal fine-tuning solutions, we propose a novel fully fine-tuned continual learning (CL) method referred to as Soft-TransFormers (Soft-TF). Soft-TF sequentially learns and selects an optimal soft-network or subnetwork for each task. During sequential training in CL, Soft-TF jointly optimizes the weights of sparse layers to obtain task-adaptive soft (real-valued) networks or subnetworks (binary masks), while keeping the well-pre-trained layer parameters frozen. In inference, the identified task-adaptive network of Soft-TF masks the parameters of the pre-trained network, mapping to an optimal solution for each task and minimizing Catastrophic Forgetting (CF) - the soft-masking preserves the knowledge of the pre-trained network. Extensive experiments on Vision Transformer (ViT) and CLIP demonstrate the effectiveness of Soft-TF, achieving state-of-the-art performance across various CL scenarios, including Class-Incremental Learning (CIL) and Task-Incremental Learning (TIL), supported by convergence theory.", "keywords": [ "Soft-Transformers", "Continual Learning", "Well-initialized Lottery Ticket Hypothesis (WLTH)" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "o9ewXD1JuB", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Chanwoong Yoon", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "MS student" }, { "name": "Hyeon Hwang", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" }, { "name": "Jaewoo Kang", "gender": "unknown", "institution": "Korea University", "country": "KR", "position": "Full Professor" }, { "name": "Minbyul Jeong", "gender": "Male", "institution": "Upstage", "country": "", "position": "Postdoc" }, { "name": "Taewhoo Lee", "gender": "unknown", "institution": "Korea University", "country": "", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 65, "n_ref": 141, "n_ref_all": 176, "n_fig": 7, "n_tab": 7, "L_tab": 5083, "n_element_tab": 466, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2095, "n_element_tab_1": 186, "formula_len_all": 452, "formula_len_all_1": 377, "len_all": 211829, "len_all_1": 71861, "len_abs": 2719, "len_title": 118, "len_sents": 59924, "len_sents_1": 32669, "n_sents": 475, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1334, "title": "OLAPH: Improving Factuality in Biomedical Long-form Question Answering", "abs": "In the medical domain, numerous scenarios necessitate the long-form generation ability of large language models (LLMs). Specifically, when addressing patients' questions, it is essential that the model's response conveys factual claims, highlighting the need for an automated method to evaluate those claims. Thus, we introduce MedLFQA, a benchmark dataset reconstructed using long-form question-answering datasets related to the biomedical domain. We use MedLFQA to facilitate a cost-effective automatic evaluations of factuality. We also propose OLAPH, a simple and novel framework that utilizes cost-effective and multifaceted automatic evaluation to construct a synthetic preference set and answers questions in our preferred manner. Our framework leads us to train LLMs step-by-step to reduce hallucinations and include crucial medical claims. We highlight that, even on evaluation metrics not used during training, LLMs trained with our OLAPH framework demonstrate significant performance improvement in factuality. Our findings reveal that a 7B LLM trained with our OLAPH framework can provide long answers comparable to the medical experts' answers in terms of factuality. We believe that our work could shed light on gauging the long-text generation ability of LLMs in the medical domain. Our code and datasets are available.", "keywords": [ "medical question answering", "automatic evaluation", "factuality", "hallucination" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "o9YC0B6P2m", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Howe Tissue", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Lu Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Venus Wang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 8, "n_ref_uni": 28, "n_ref": 69, "n_ref_all": 141, "n_fig": 35, "n_tab": 10, "L_tab": 10342, "n_element_tab": 336, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 2967, "n_element_tab_1": 89, "formula_len_all": 1101, "formula_len_all_1": 407, "len_all": 200058, "len_all_1": 63963, "len_abs": 1545, "len_title": 88, "len_sents": 62992, "len_sents_1": 27583, "n_sents": 537, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1623, "title": "Scaling Law with Learning Rate Annealing", "abs": "We find that the cross-entropy loss curves of neural language models empirically adhere to a scaling law with learning rate (LR) annealing over training steps:\n$$L(s) = L_0 + A\\cdot S_1^{-\\alpha} - C\\cdot S_2,$$\nwhere $L(s)$ is the validation loss at step $s$, $S_1$ is the area under the LR curve, $S_2$ is the LR annealing area, and $L_0$, $A$, $C$, $\\alpha$ are constant parameters.\nThis formulation accounts for two main effects: (1) power-law scaling over data size, and (2) the additional loss reduction during LR annealing. \nUnlike previous studies that only fit losses at final steps, our formulation captures the entire training curve, allowing for parameter fitting using losses from any training step.\nApplying the scaling law with LR annealing and fitting only one or two training curves, we can accurately predict the loss at any given step under any learning rate scheduler (LRS).\nThis approach significantly reduces computational cost in formulating scaling laws while providing more accuracy and expressiveness.\nExtensive experiments demonstrate that our findings hold across a range of hyper-parameters and model architectures and can extend to scaling effect of model sizes.\nMoreover, our formulation provides accurate theoretical insights into empirical results observed in numerous previous studies, particularly those focusing on LR schedule and annealing.\nWe believe that this work is promising to enhance the understanding of LLM training dynamics while democratizing scaling laws, and it is helpful to guide both research and industrial participants in refining training strategies for further LLMs.", "keywords": [ "Scaling Laws", "Full Loss Curve Prediction", "Learning Rate Schedule", "LLM Pretraining" ], "rating_list": [ 8, 8, 5, 6 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "o9UzvKVvuf", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Ezgi Korkmaz", "gender": "unknown", "institution": "DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Jonah Brown-Cohen", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" } ], "n_formula": 58, "n_formula_1": 23, "n_ref_uni": 37, "n_ref": 71, "n_ref_all": 77, "n_fig": 0, "n_tab": 2, "L_tab": 524, "n_element_tab": 30, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 20, "n_element_tab_1": 5, "formula_len_all": 4925, "formula_len_all_1": 1555, "len_all": 176864, "len_all_1": 68026, "len_abs": 1312, "len_title": 109, "len_sents": 43808, "len_sents_1": 28810, "n_sents": 411, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1329, "title": "Aligning With Human Values Without Revealing Human Judgements", "abs": "With the increasing ubiquity of large language models it has become crucial to ensure guarantees for models trained to be aligned with human values to avoid leaking information on the human judgements that have been provided to the algorithm. To target this issue we focus on the problem of alignment via reinforcement learning from human preference rankings, subject to the constraint of \nnot revealing any information on the human data used to align the model. To achieve this, we analyze $(\\epsilon,\\delta)$-DP for both the Bradley-Terry-Luce (BTL) model and the Plackett-Luce (PL) model. We introduce a theoretically founded algorithm for learning rewards from human rankings that achieves this objective without leaking the human rankings. We further demonstrate that the privately learned rewards can be used to train policies achieving statistical performance guarantees that asymptotically match the best known algorithms in the non-private setting, which are in some cases minimax optimal. Strikingly, our analysis and our results reveal that it is possible to obtain the same model performance without any trade-off on the protection of the human judgments, and our paper provides the first algorithms that can achieve provable privacy of human judgements, while still producing aligned models with optimal performance.", "keywords": [ "Responsible AI", "language models", "protecting human judgements" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 2, 3, 4 ] }, { "paper_id": "o9SuQXZvNA", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Canyu Chen", "gender": "unknown", "institution": "Illinois Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Danielle Bitterman", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "Fei Wang", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Full Professor" }, { "name": "Jian Yu", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "Undergrad student" }, { "name": "Kai Shu", "gender": "unknown", "institution": "Emory University", "country": "US", "position": "Assistant Professor" }, { "name": "Shan Chen", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Researcher" }, { "name": "Zhongwei Wan", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "che liu", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 0, "n_ref_uni": 60, "n_ref": 135, "n_ref_all": 151, "n_fig": 4, "n_tab": 14, "L_tab": 9539, "n_element_tab": 753, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 4822, "n_element_tab_1": 563, "formula_len_all": 199, "formula_len_all_1": 0, "len_all": 263871, "len_all_1": 66362, "len_abs": 1334, "len_title": 74, "len_sents": 70008, "len_sents_1": 26476, "n_sents": 518, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1275, "title": "ClinicalBench: Can LLMs Beat Traditional ML Models in Clinical Prediction?", "abs": "Large Language Models (LLMs) hold great promise to revolutionize current clinical systems for their superior capacities on medical text processing tasks and medical licensing exams. Meanwhile, traditional ML models such as SVM and XGBoost have still been mainly adopted in clinical prediction tasks. An emerging question is Can LLMs beat traditional ML models in clinical prediction? Thus, we build a new benchmark ClinicalBench to comprehensively study the clinical predictive modeling capacities of both general-purpose and medical LLMs, and compare them with traditional ML models. ClinicalBench embraces three common clinical prediction tasks, two databases, 14 general-purpose LLMs, 8 medical LLMs, and 11 traditional ML models. Through extensive empirical investigation, we discover that both general-purpose and medical LLMs, even with different model scales, diverse prompting or fine-tuning strategies, still cannot beat traditional ML models in clinical prediction yet, shedding light on their surprising but critical deficiency in clinical reasoning. We call for caution when practitioners adopt LLMs in clinical applications. ClinicalBench can be utilized to bridge the gap between LLMs' development for healthcare and real-world clinical practice. Code is here.", "keywords": [ "LLMs", "Clinical Prediction", "Benchmark" ], "rating_list": [ 8, 8, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 1, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "o99Yn1wN9J", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Dingrong Wang", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Ervine Zheng", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Krishna Prasad Neupane", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Qi Yu", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "Professor" } ], "n_formula": 21, "n_formula_1": 10, "n_ref_uni": 59, "n_ref": 113, "n_ref_all": 155, "n_fig": 8, "n_tab": 11, "L_tab": 5758, "n_element_tab": 489, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1182, "n_element_tab_1": 174, "formula_len_all": 2309, "formula_len_all_1": 816, "len_all": 226091, "len_all_1": 67195, "len_abs": 1843, "len_title": 140, "len_sents": 70680, "len_sents_1": 32514, "n_sents": 503, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1850, "title": "Looking into User\u2019s Long-term Interests through the Lens of Conservative Evidential Learning", "abs": "Reinforcement learning (RL) provides an effective means to capture users' evolving preferences, leading to improved recommendation performance over time. However, existing RL approaches primarily rely on standard exploration strategies, which are less effective for a large item space with sparse reward signals given the limited interactions for most users. Therefore, they may not be able to learn the optimal policy that effectively captures user's evolving preferences and achieves the maximum expected reward over the long term. In this paper, we propose a novel evidential conservative Q-learning framework (ECQL) that learns an effective and conservative recommendation policy by integrating evidence-based uncertainty and conservative learning. ECQL conducts evidence-aware explorations to discover items that are located beyond current observations but reflect users' long-term interests. It offers an uncertainty-aware conservative view on policy evaluation to discourage deviating too much from users' current interests. Two central components of ECQL include a uniquely designed sequential state encoder and a novel conservative evidential-actor-critic (CEAC) module. The former generates the current state of the environment by aggregating historical information and a sliding window that contains the current user interactions as well as newly recommended items from RL exploration that may represent short and long-term interests respectively. The latter performs an evidence-based rating prediction by maximizing the conservative evidential Q-value and leverages an uncertainty-aware ranking score to explore the item space for a more diverse and valuable recommendation. Experiments on multiple real-world dynamic datasets demonstrate the state-of-the-art performance of ECQL and its capability to capture users' long-term interests.", "keywords": [ "recommender systems", "evidence-aware exploration", "evidential learning" ], "rating_list": [ 8, 6, 3, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 2, 4, 5 ] }, { "paper_id": "o8vCBFonHC", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Eilam Shapira", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "PhD student" }, { "name": "Itamar Reinman", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "Undergrad student" }, { "name": "Moshe Tennenholtz", "gender": "unknown", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "Full Professor" }, { "name": "Omer Madmon", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "PhD student" }, { "name": "Roi Reichart", "gender": "Male", "institution": "Technion, Israel Institute of Technology", "country": "IL", "position": "Associate Professor" }, { "name": "Samuel Joseph Amouyal", "gender": "Male", "institution": "School of Computer Science, Tel Aviv University", "country": "", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 6, "n_ref": 19, "n_ref_all": 55, "n_fig": 16, "n_tab": 6, "L_tab": 1037, "n_element_tab": 132, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 322, "n_element_tab_1": 37, "formula_len_all": 117, "formula_len_all_1": 117, "len_all": 179840, "len_all_1": 73176, "len_abs": 2020, "len_title": 126, "len_sents": 65549, "len_sents_1": 37116, "n_sents": 492, "n_sents_1": 267, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 2031, "title": "GLEE: A Framework and Benchmark for LLM Evaluation in Language-based Economics", "abs": "Large Language Models (LLMs) show significant potential in economic and strategic interactions, where communication via natural language is often prevalent. This raises key questions: Do LLMs behave rationally? Can they mimic human behavior? Do they tend to reach an efficient and fair outcome? What is the role of natural language in the strategic interaction? How do characteristics of the economic environment influence these dynamics?\nThese questions become crucial concerning the economic and societal implications of integrating LLM-based agents into real-world data-driven systems, such as online retail platforms and recommender systems.\nWhile the ML community has been exploring the potential of LLMs in such multi-agent setups, varying assumptions, design choices and evaluation criteria across studies make it difficult to draw robust and meaningful conclusions. To address this, we introduce a benchmark for standardizing research on two-player, sequential, language-based games. Inspired by the economic literature, we define three base families of games with consistent parameterization, degrees of freedom and economic measures to evaluate agents' performance (self-gain), as well as the game outcome (efficiency and fairness). We develop an open-source framework for interaction simulation and analysis, and utilize it to collect a dataset of LLM vs. LLM interactions across numerous game configurations and an additional dataset of human vs. LLM interactions.\nThrough extensive experimentation, we demonstrate how our framework and dataset can be used to: (i) compare the behavior of LLM-based agents to human players in various economic contexts; (ii) evaluate agents in both individual and collective performance measures; and (iii) quantify the effect of the economic characteristics of the environments on the behavior of agents. We believe that our framework can contribute to the growing intersection of LLMs, ML, and economics, and we encourage researchers to explore it further and build on its foundation.", "keywords": [ "Language-based games", "Language-based economic environment", "LLM benchmark", "LLM Framework" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "o8SPZJaJyj", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Sihang Cai", "gender": "Male", "institution": "College of Computer Science and Technology, Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Tao Jin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Weicai Yan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Xize Cheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yangyang Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Zehan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zirun Guo", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "fangming feng", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 45, "n_ref": 73, "n_ref_all": 91, "n_fig": 10, "n_tab": 4, "L_tab": 1927, "n_element_tab": 185, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 894, "n_element_tab_1": 45, "formula_len_all": 485, "formula_len_all_1": 486, "len_all": 154709, "len_all_1": 63189, "len_abs": 1620, "len_title": 126, "len_sents": 34942, "len_sents_1": 27816, "n_sents": 272, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1628, "title": "Dynamic Switching Teacher: How to Generalize Temporal Action Detection Models", "abs": "Temporal Action Detection (TAD) is a crucial task in video understanding, focusing on the precise identification of the onset and termination of specific actions within video sequences. Despite advancements on certain datasets, existing methods often struggle to maintain their efficacy when applied to datasets from disparate domain. In this study, we introduce, for the first time, the application of source-free domain adaptation (SFDA) techniques to the field of TAD, aiming to enhance the generalization capability of TAD models on unlabeled target datasets without access to source data. Most popular SFDA methods predominantly follow the Mean-Teacher (MT) framework and often falter due to the significant domain shift. The generation of pseudo labels by a pre-trained teacher model on the source domain can lead to a cascade of errors when these labels guide the training of a student model, potentially causing a harmful TAD feedback loop. To address this issue, we propose a novel dynamic switching teacher strategy that integrates both dynamic and static teacher models. The dynamic teacher model updates its parameters by learning knowledge from the student model. Concurrently, the static teacher model engages in periodic weight exchange with the student model, ensuring baseline performance and maintaining the quality of pseudo labels. This approach significantly mitigates the label noise. We establish the first benchmark for SFDA in TAD tasks and conduct extensive experiments across various datasets. Our method demonstrates state-of-the-art performance, substantiating the suitability of our method for TAD.", "keywords": [ "Video Understanding", "Temporal Action Detection", "Domain Adaptation" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "o83aL1nZJd", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bohan Tang", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Rui Ye", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Shuo Tang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xianghe Pang", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Xiaowen Dong", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" }, { "name": "Yanfeng Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Zexi Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Undergrad student" }, { "name": "Siheng Chen", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 2, "n_ref_uni": 37, "n_ref": 106, "n_ref_all": 149, "n_fig": 8, "n_tab": 13, "L_tab": 3450, "n_element_tab": 258, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 737, "n_element_tab_1": 126, "formula_len_all": 448, "formula_len_all_1": 206, "len_all": 195469, "len_all_1": 70825, "len_abs": 967, "len_title": 119, "len_sents": 58393, "len_sents_1": 34047, "n_sents": 431, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 915, "title": "Synthesizing Post-Training Data for LLMs through Multi-Agent Simulation", "abs": "Post-training is essential for enabling large language models (LLMs) to follow human instructions. \nInspired by the recent success of using LLMs to simulate human society, we leverage multi-agent simulation to automatically generate diverse text-based scenarios, capturing a wide range of real-world human needs. \nWe introduce MATRIX, a multi-agent simulator that creates realistic and scalable scenarios. \nLeveraging these outputs, we introduce a novel scenario-driven instruction generator MATRIX-Gen for controllable and highly realistic data synthesis. Extensive experiments demonstrate that our framework effectively generates both general and domain-specific data. Notably, on AlpacaEval 2 and Arena-Hard benchmarks, Llama-3-8B-Base, post-trained on datasets synthesized by MATRIX-Gen with just 20K instruction-response pairs, outperforms Meta's Llama-3-8B-Instruct model, which was trained on over 10M pairs.", "keywords": [ "large language models", "llm alignment", "multi-agent simulation", "llm society" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "o7alDZDJWB", "primary_area": "optimization", "n_author": 8, "author_info_list": [ { "name": "Chun-Mei Feng", "gender": "Female", "institution": "IHPC", "country": "SG", "position": "Researcher" }, { "name": "Rick Siow Mong Goh", "gender": "unknown", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "SG", "position": "Director" }, { "name": "Salman Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Associate Professor" }, { "name": "Wangmeng Zuo", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xinxing Xu", "gender": "Male", "institution": "Microsoft", "country": "SG", "position": "Principal Researcher" }, { "name": "Yong Liu", "gender": "Male", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "SG", "position": "Senior Scientist, Adjunct Assistant Professor" }, { "name": "Yuanyang He", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Zitong Huang", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Intern" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 37, "n_ref": 93, "n_ref_all": 107, "n_fig": 7, "n_tab": 14, "L_tab": 2751, "n_element_tab": 348, "n_fig_1": 6, "n_tab_1": 8, "L_tab_1": 2403, "n_element_tab_1": 202, "formula_len_all": 240, "formula_len_all_1": 240, "len_all": 147106, "len_all_1": 69290, "len_abs": 1300, "len_title": 97, "len_sents": 45199, "len_sents_1": 31883, "n_sents": 317, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1296, "title": "CPT: Consistent Proxy Tuning for Black-box Optimization", "abs": "Black-box tuning has attracted recent attention due to that the structure or inner parameters of advanced proprietary models are not accessible. Recently, Proxy-tuning provides a test-time output adjustment for tuning black-box language models.It applies the difference of the output logits before and after tuning a smaller white-box \"proxy\" model to improve the black-box model. However, this technique serves only as a decoding-time algorithm, leading to an inconsistency between training and testing which potentially limits overall performance. To address this problem, we introduce Consistent Proxy Tuning (CPT), a simple yet effective black-box tuning method. Different from Proxy-tuning, CPT additionally exploits the frozen large black-box model and another frozen small white-box model, ensuring consistency between training-stage optimization objective and test-time proxies. This consistency benefits Proxy-tuning and enhances model performance. Note that our method focuses solely on logit-level computation, which makes it model-agnostic and applicable to any task involving logit classification. Extensive experimental results demonstrate the superiority of our CPT in both black-box tuning of Large-Language Models (LLMs) and Vision-Language Models (VLMs) across various datasets.", "keywords": [ "black-box models", "large language models", "vision-language models", "black-box tuning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 1, 1, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "o6tO1rUcQe", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Jiaqi Zheng", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Qing Ling", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Yerong Feng", "gender": "Male", "institution": "Shenzhen institute of meteorological innovation", "country": "CN", "position": "Researcher" } ], "n_formula": 30, "n_formula_1": 9, "n_ref_uni": 28, "n_ref": 53, "n_ref_all": 84, "n_fig": 11, "n_tab": 9, "L_tab": 3178, "n_element_tab": 788, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1506, "n_element_tab_1": 424, "formula_len_all": 2936, "formula_len_all_1": 911, "len_all": 146985, "len_all_1": 67158, "len_abs": 1577, "len_title": 123, "len_sents": 40947, "len_sents_1": 28450, "n_sents": 316, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1631, "title": "Physics-Assisted and Topology-Informed Deep Learning for Weather Prediction", "abs": "Weather prediction is crucial for decision-making in various social and economic sectors. The classical numerical weather prediction methods cannot incorporate the historical observations to enhance the underlying physical models, whereas the existing data-driven, deep learning-based weather prediction methods disregard either the $\\textbf{physics}$ of the weather evolution or the $\\textbf{topology}$ of the Earth's surface. In light of these disadvantages, we develop PASSAT, a novel Physics-ASSisted And Topology-informed deep learning model for weather prediction. PASSAT attributes the weather evolution to two key factors: (i) the advection process that can be characterized by the advection equation and the Navier-Stokes equation; (ii) the Earth-atmosphere interaction that is difficult to both model and calculate. PASSAT also takes the topology of the Earth's surface into consideration, other than simply treating it as a plane. Therefore, PASSAT numerically solves the advection equation and the Navier-Stokes equation on the spherical manifold, utilizes a spherical graph neural network to capture the Earth-atmosphere interaction, and generates the initial velocity fields that are critical to solving the advection equation from the same spherical graph neural network. These building blocks constitute a deep learning-based, $\\textbf{physics-assisted}$ and $\\textbf{topology-informed}$ weather prediction model. In the $5.625^\\circ$-resolution ERA5 data set, PASSAT outperforms both the state-of-the-art deep learning-based weather prediction models and the operational numerical weather prediction model IFS T42.", "keywords": [ "global weather prediction", "physics-assisted deep learning", "topology-informed deep learning" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "o6ddWvoyjK", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Ayumu Yamashita", "gender": "Male", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Hideki Nakayama", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Kaoru Amano", "gender": "Male", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Masataka Sawayama", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Lecturer" }, { "name": "Yuto Nishimura", "gender": "unknown", "institution": "Tokyo University", "country": "JP", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 50, "n_ref": 92, "n_ref_all": 145, "n_fig": 24, "n_tab": 8, "L_tab": 2558, "n_element_tab": 166, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1020, "n_element_tab_1": 53, "formula_len_all": 100, "formula_len_all_1": 136, "len_all": 204765, "len_all_1": 64406, "len_abs": 1285, "len_title": 120, "len_sents": 65832, "len_sents_1": 29445, "n_sents": 581, "n_sents_1": 254, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1297, "title": "BrainCodec: Neural fMRI codec for the decoding of cognitive brain states", "abs": "Recently, leveraging big data in deep learning has led to significant performance improvements, as confirmed in applications like mental state decoding using fMRI data. \nHowever, fMRI datasets remain relatively small in scale, and the inherent issue of low signal-to-noise ratios (SNR) in fMRI data further exacerbates these challenges. \nTo address this, we apply compression techniques as a preprocessing step for fMRI data. \nWe propose BrainCodec, a novel fMRI codec inspired by the neural audio codec. \nWe evaluated BrainCodec's compression capability in mental state decoding, demonstrating further improvements over previous methods.\nFurthermore, we analyzed the latent representations obtained through BrainCodec, elucidating the similarities and differences between task and resting state fMRI, highlighting the interpretability of BrainCodec. \nAdditionally, we demonstrated that fMRI reconstructions using BrainCodec can enhance the visibility of brain activity by achieving higher SNR, suggesting its potential as a novel denoising method. \nOur study shows that BrainCodec not only enhances performance over previous methods but also offers new analytical possibilities for neuroscience.\nOur codes, dataset, and model weights are available at https://anonymous.4open.science/r/BrainCodec.", "keywords": [ "fMRI", "Neural Audio Codec", "NLP", "SSL" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "o6aUi3ukdd", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Chao Qu", "gender": "Male", "institution": "Inftech", "country": "CN", "position": "Researcher" }, { "name": "Fenglei Cao", "gender": "Male", "institution": "inf tech", "country": "", "position": "Researcher" }, { "name": "Junyi An", "gender": "Male", "institution": "Shanghai Academy of Artificial Intelligence for Science", "country": "CN", "position": "Researcher" }, { "name": "Weiqi Liu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Xi Ai", "gender": "Not Specified", "institution": "Inftech", "country": "", "position": "Researcher" }, { "name": "Xu Yinghui", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yuan Cheng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yuan Qi", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "ZhijianZhou", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Zhipeng Zhou", "gender": "Male", "institution": "INF (Shanghai) Technology Co., LTD", "country": "", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 62, "n_ref": 71, "n_ref_all": 86, "n_fig": 5, "n_tab": 5, "L_tab": 1997, "n_element_tab": 174, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2211, "n_element_tab_1": 129, "formula_len_all": 27, "formula_len_all_1": 27, "len_all": 154514, "len_all_1": 56804, "len_abs": 1030, "len_title": 93, "len_sents": 31551, "len_sents_1": 25764, "n_sents": 216, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1045, "title": "An Open Quantum Chemistry Property Database of 120 Kilo Molecules with 20 Million Conformers", "abs": "Artificial intelligence is revolutionizing computational chemistry, bringing unprecedented innovation and efficiency to the field. To further advance research and expedite progress, we introduce the Quantum Open Organic Molecular (QO2Mol) database \u2014 a large-scale quantum chemistry dataset designed for researches on organic molecules under an open-source license. \nThe database comprises 120,000 organic molecules and more than 20 million conformers, encompassing 10 different elements (C, H, O, N, S, P, F, Cl, Br, I), with heavy atom counts exceeding 40. Each conformation was computed at B3LYP/def2-SVP level of theory to derive quantum mechanical properties, including potential energy and forces. The molecules included in the dataset are based on fragments from compounds in ChEMBL, ensuring their structural \\textit{relevance to real-world compounds}.\nThe extensive variety of molecular structures and elemental compositions represented in the dataset can facilitate construction of potential energy surface and various downstream tasks.", "keywords": [ "Quantum Chemistry", "Machine Learning", "Organic Molecules" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 1, 2, 1, 1 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "o6Ynz6OIQ6", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Jinheng Xie", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Weihao Wang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Weijia Mao", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yuchao Gu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zechen Bai", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zhenheng Yang", "gender": "Male", "institution": "Tiktok", "country": "US", "position": "Director of Research" }, { "name": "Zhijie Chen", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Junhao Zhang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Qinghong Lin", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zheng Shou", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 21, "n_formula_1": 5, "n_ref_uni": 55, "n_ref": 133, "n_ref_all": 172, "n_fig": 18, "n_tab": 3, "L_tab": 1942, "n_element_tab": 147, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 1244, "n_element_tab_1": 82, "formula_len_all": 3048, "formula_len_all_1": 225, "len_all": 231590, "len_all_1": 61758, "len_abs": 794, "len_title": 127, "len_sents": 53923, "len_sents_1": 28980, "n_sents": 431, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 763, "title": "Show-o: One Single Transformer to Unify Multimodal Understanding and Generation", "abs": "We present a unified transformer, i.e., Show-o, that unifies multimodal understanding and generation. Unlike fully autoregressive models, Show-o unifies autoregressive and (discrete) diffusion modeling to adaptively handle inputs and outputs of various and mixed modalities. The unified model flexibly supports a wide range of vision-language tasks including visual question-answering, text-to-image generation, text-guided inpainting/extrapolation, and mixed-modality generation. Across various benchmarks, it demonstrates comparable or superior performance to existing individual models with an equivalent or larger number of parameters tailored for understanding or generation. This significantly highlights its potential as a next-generation foundation model.", "keywords": [ "Multimodal understanding and generation", "large language model" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "o6CXkEEttn", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Evi Micha", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Safwan Hossain", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Yiling Chen", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Full Professor" }, { "name": "Ariel Procaccia", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Gordon McKay Professor of Computer Science" } ], "n_formula": 88, "n_formula_1": 27, "n_ref_uni": 43, "n_ref": 63, "n_ref_all": 82, "n_fig": 1, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10579, "formula_len_all_1": 2373, "len_all": 180098, "len_all_1": 73311, "len_abs": 1495, "len_title": 91, "len_sents": 59412, "len_sents_1": 34573, "n_sents": 487, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 1187, "title": "Strategic Classification With Externalities", "abs": "We propose a new variant of the strategic classification problem: a principal reveals a classifier, and $n$ agents report their (possibly manipulated) features to be classified. Motivated by real-world applications, our model crucially allows the manipulation of one agent to affect another; that is, it explicitly captures inter-agent externalities. The principal-agent interactions are formally modeled as a Stackelberg game, with the resulting agent manipulation dynamics captured as a simultaneous game. We show that under certain assumptions, the pure Nash Equilibrium of this agent manipulation game is unique and can be efficiently computed. Leveraging this result, PAC learning guarantees are established for the learner: informally, we show that it is possible to learn classifiers that minimize loss on the distribution, even when a random number of agents are manipulating their way to a pure Nash Equilibrium. We also comment on the optimization of such classifiers through gradient-based approaches. This work sets the theoretical foundations for a more realistic analysis of classifiers that are robust against multiple strategic actors interacting in a common environment.", "keywords": [ "strategic classification", "game theory", "strategic machine learning", "algorithmic game theory" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 4, 4, 3, 3 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "o5wGjBEgH8", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Gregor M\u00fcckl", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Ishwarya Ananthabhotla", "gender": "Female", "institution": "Meta Reality Labs Research", "country": "US", "position": "Researcher" }, { "name": "Ricardo Falcon Perez", "gender": "unknown", "institution": "Aalto University", "country": "", "position": "PhD student" }, { "name": "Ruohan Gao", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Sebastia Vicenc Amengual Gari", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 70, "n_ref": 97, "n_ref_all": 119, "n_fig": 9, "n_tab": 3, "L_tab": 457, "n_element_tab": 41, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1630, "n_element_tab_1": 218, "formula_len_all": 236, "formula_len_all_1": 232, "len_all": 190003, "len_all_1": 68874, "len_abs": 303, "len_title": 100, "len_sents": 46308, "len_sents_1": 32675, "n_sents": 344, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1860, "title": "Novel View Acoustic Parameter Estimation", "abs": "The task of Novel View Acoustic Synthesis (NVAS) -- generating Room Impulse Responses (RIRs) for unseen source and receiver positions in a scene -- has recently gained traction, especially given its relevance to Augmented Reality (AR) and Virtual Reality (VR) development. However, many of these efforts suffer from similar limitations: they infer RIRs in the time domain, which prove challenging to optimize; they focus on scenes with simple, single-room geometries; they infer only single-channel, directionally-independent acoustic characteristics; and they require inputs, such as 3D geometry meshes with material properties, that may be impractical to obtain for on-device applications. On the other hand, research suggests that sample-wise accuracy of RIRs is not required for perceptual plausibility in AR and VR. Standard acoustic parameters like Clarity Index (C50) or Reverberation Time (T60) have been shown to capably describe pertinent characteristics of the RIRs, especially late reverberation. To address these gaps, this paper introduces a new, intermediate task centered on estimating spatially distributed acoustic parameters, that can be then used to condition a simple reverberator to generate RIRs for arbitrary source and receiver positions. The approach is modeled as an image-to-image translation task, which translates 2D floormaps of a scene into 2D heatmaps of acoustic parameters. We introduce a new, large-scale dataset of 1000 scenes consisting of complex, multi-room apartment conditions, and show that our method outperforms statistical baselines significantly. Moreover, we show that the method also works for directionally-dependent (i.e. beamformed) parameter prediction. Finally, the proposed method operates on very limited information, requiring only a broad outline of the scene and a single RIR at inference time.", "keywords": [ "acoustic parameter estimation", "room impulse response", "reverberation", "spatial audio", "room acoustics", "novel view acoustic synthesis" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "o5TsWTUSeF", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Bowen Qu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Chengjin Xu", "gender": "Male", "institution": "International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Chun Yuan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jian Guo", "gender": "Male", "institution": "International Digital Economy Academy, International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "SiNan Du", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yiyan Qi", "gender": "unknown", "institution": "IDEA", "country": "CN", "position": "Researcher" }, { "name": "Zhengzhuo Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 20, "n_ref": 157, "n_ref_all": 186, "n_fig": 15, "n_tab": 15, "L_tab": 9769, "n_element_tab": 643, "n_fig_1": 10, "n_tab_1": 9, "L_tab_1": 4393, "n_element_tab_1": 379, "formula_len_all": 184, "formula_len_all_1": 172, "len_all": 180742, "len_all_1": 79296, "len_abs": 1323, "len_title": 128, "len_sents": 46638, "len_sents_1": 31275, "n_sents": 347, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 79, "L_abs": 1253, "title": "ChartMoE: Mixture of Diversely Aligned Expert Connector for Chart Understanding", "abs": "Automatic chart understanding is crucial for content comprehension and document parsing. Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in chart understanding through domain-specific alignment and fine-tuning. However, current MLLMs still struggle to provide faithful data and reliable analysis only based on charts. To address it, we propose ChartMoE, which employs the Mixture of Expert (MoE) architecture to replace the traditional linear projector to bridge the modality gap. Specifically, we train several linear connectors through distinct alignment tasks, which are utilized as the foundational initialization parameters for different experts. Additionally, we introduce ChartMoE-Align, a dataset with nearly 1 million chart-table-JSON-code quadruples to conduct three alignment tasks (chart-table/JSON/code). Combined with the vanilla connector, we initialize different experts diversely and adopt high-quality knowledge learning to further refine the MoE connector and LLM parameters. Extensive experiments demonstrate the effectiveness of the MoE connector and our initialization strategy, e.g., ChartMoE improves the accuracy of the previous state-of-the-art from 80.48% to 84.64% on the ChartQA benchmark.", "keywords": [ "Multimodal Large Language Models", "Chart Reasoning", "Mixture of Expert" ], "rating_list": [ 8, 8, 5, 6 ], "soundness_list": [ 3, 4, 2, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "o4mvfEWbsP", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Hai-Miao Hu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Haowen Zheng", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Likun Gao", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" }, { "name": "Xinhui Xue", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" } ], "n_formula": 44, "n_formula_1": 17, "n_ref_uni": 29, "n_ref": 64, "n_ref_all": 83, "n_fig": 6, "n_tab": 12, "L_tab": 4006, "n_element_tab": 409, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2812, "n_element_tab_1": 261, "formula_len_all": 4165, "formula_len_all_1": 1382, "len_all": 156976, "len_all_1": 65602, "len_abs": 1056, "len_title": 117, "len_sents": 42491, "len_sents_1": 26387, "n_sents": 324, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1059, "title": "Sparse Hyperspectral Band Selection Based on Expectation Maximization", "abs": "Band selection is crucial in spectral imaging, as it involves choosing the most relevant bands from large hyperspectral datasets to retain essential information while reducing the burden of data transmission and analysis. Addressing this need, we introduce a novel method for band selection that utilizes an Expectation Maximization algorithm to facilitate selection through the sparsification of spectral band importance. Our method enhances sparsity effects and effectively delineates the relationships between spectral bands during the sparsification process. Supported by thorough theoretical analysis and experimental validation on public datasets, our approach has proven to be both robust and practical. Compared to other sparsification methods, it not only excels in achieving significant sparsity effects but also demonstrates marked advantages in illustrating inter-band relationships. Our method delivers outstanding performance in band selection tasks and holds potential for broader applications in other sparsity-oriented contexts in the future.", "keywords": [ "Hyperspectral Band Selection", "Sparse Learning", "Feature Selection", "Hyperspectral Image Classification", "EM Algorithm" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "o4byGNa98y", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Alessandro Abate", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Patrick Benjamin", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 69, "n_formula_1": 22, "n_ref_uni": 128, "n_ref": 302, "n_ref_all": 335, "n_fig": 15, "n_tab": 3, "L_tab": 170, "n_element_tab": 8, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6113, "formula_len_all_1": 2130, "len_all": 388813, "len_all_1": 76308, "len_abs": 2043, "len_title": 116, "len_sents": 96538, "len_sents_1": 34337, "n_sents": 714, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1523, "title": "Networked Communication for Decentralised Agents in Mean-Field Games", "abs": "We introduce networked communication to the mean-field game framework, in particular to oracle-free settings where $N$ decentralised agents learn along a single, non-episodic run of the empirical system. We prove that our architecture has sample guarantees bounded between those of the centralised- and independent-learning cases. We provide the order of the difference in these bounds in terms of network structure and number of communication rounds, and also contribute a policy-update stability guarantee. We discuss how the sample guarantees of the three theoretical algorithms do not actually result in practical convergence. We therefore show that in practical settings where the theoretical parameters are not observed (leading to poor estimation of the Q-function), our communication scheme significantly accelerates convergence over the independent case (and sometimes even the centralised case), without relying on the assumption of a centralised learner. We contribute further practical enhancements to all three theoretical algorithms, allowing us to present their first empirical demonstrations. Our experiments confirm that we can remove several of the theoretical assumptions of the algorithms, and display the empirical convergence benefits brought by our new networked communication. We additionally show that the networked approach has significant advantages, over both the centralised and independent alternatives, in terms of robustness to unexpected learning failures and to changes in population size.", "keywords": [ "mean-field games", "reinforcement learning", "communication network", "multi-agent systems" ], "rating_list": [ 3, 6, 6, 3, 1 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 1, 4, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 2, 1 ], "confidence_list": [ 3, 4, 3, 3, 4 ] }, { "paper_id": "o4X6UM18rI", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Bradley A. Malin", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "Professor" }, { "name": "Rajagopal Venkatesaramani", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Assistant Teaching Professor" }, { "name": "Rajat K. De", "gender": "Male", "institution": "Indian Statistical Institute", "country": "IN", "position": "Full Professor" }, { "name": "Tao Zhang", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "", "position": "Postdoc" }, { "name": "Yevgeniy Vorobeychik", "gender": "Male", "institution": "Washington University, St. Louis", "country": "US", "position": "Full Professor" } ], "n_formula": 133, "n_formula_1": 21, "n_ref_uni": 59, "n_ref": 129, "n_ref_all": 156, "n_fig": 12, "n_tab": 5, "L_tab": 795, "n_element_tab": 21, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 10, "n_element_tab_1": 1, "formula_len_all": 18253, "formula_len_all_1": 1956, "len_all": 290878, "len_all_1": 70034, "len_abs": 1251, "len_title": 125, "len_sents": 91253, "len_sents_1": 31219, "n_sents": 804, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1257, "title": "Bayes-Nash Generative Privacy Protection Against Membership Inference Attacks", "abs": "Membership inference attacks (MIAs) expose significant privacy risks by determining whether an individual\u2019s data is in a dataset. While differential privacy (DP) mitigates such risks, it faces challenges in general when achieving an optimal balance between privacy and utility, often requiring intractable sensitivity calculations and limiting flexibility in complex compositions. We propose a game-theoretic framework that models privacy protection as a Bayesian game between a defender and an attacker, solved using a general-sum Generative Adversarial Network (general-sum GAN). The Bayes Generative Privacy (BGP) response, based on cross-entropy loss, defines the attacker\u2019s optimal strategy, leading to the Bayes-Nash Generative Privacy (BNGP) strategy, which achieves the optimal privacy-utility trade-off tailored to the defender\u2019s preferences. The BNGP strategy avoids sensitivity calculations, supports compositions of correlated mechanisms, and is robust to the attacker\u2019s heterogeneous preferences over true and false positives. A case study on binary dataset summary statistics demonstrates its superiority over likelihood ratio test (LRT)-based attacks, including the uniformly most powerful LRT. Empirical results confirm BNGP\u2019s effectiveness.", "keywords": [ "privacy", "game theory", "membership inference attack", "Bayes Nash equilibrium" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 2, 2 ] }, { "paper_id": "o4TyewNBIB", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Anxian Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "MS student" }, { "name": "Guang Zhang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Jianing Hao", "gender": "Female", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jingnan Chen", "gender": "unknown", "institution": "hkustgz", "country": "CN", "position": "MS student" }, { "name": "Kunsheng Tang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Peng LIU", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "MS student" }, { "name": "Yuanjian Xu", "gender": "Male", "institution": "Hong Kong University of Science and Technology(guangzhou)", "country": "CN", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 9, "n_ref_uni": 45, "n_ref": 62, "n_ref_all": 80, "n_fig": 5, "n_tab": 7, "L_tab": 3572, "n_element_tab": 596, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 1704, "n_element_tab_1": 281, "formula_len_all": 1513, "formula_len_all_1": 659, "len_all": 199040, "len_all_1": 74154, "len_abs": 2434, "len_title": 145, "len_sents": 59331, "len_sents_1": 32676, "n_sents": 430, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 2169, "title": "FinRipple: Aligning Large Language Models with Financial Market for Event Ripple Effect Awareness", "abs": "Event studies have been fundamental in finance, focusing on analyzing the ripple effects of sudden market events. Accurately predicting these effects is crucial for informed decision-making and effective risk management. However, the dynamic complexity of financial markets and the lack of unified modeling tools make this task challenging. Previous models, constrained by simplistic assumptions and limited scopes, have struggled to address this complexity effectively. In contrast, large language models (LLMs), with their emergent reasoning abilities, offer a promising solution. In this paper, we introduce $\\textbf{FinRipple}$, a novel training framework that enables LLMs to align with market behavior and develop the capability to analyze the ripple effects of sudden events. We first construct a time-varying financial knowledge graph (KG) that is both financially meaningful and noise-reduced to accurately represent the market state. These KGs are then integrated into the LLM using adapters as memory modules. Additionally, we align the LLM with market dynamics by integrating FinRipple with classic asset pricing theories through a reinforcement learning framework. This market-alignment process collects feedback that enhances the LLM's foundational ability to analyze financial events and explain market anomalies that traditional models fail to address. Our key contributions are as follows: (1) We are the first to define the underexplored task of ``event impact prediction''. Our framework not only establishes this task but also provides an open-source benchmark, creating a unified evaluation standard for both academia and industry; (2) FinRipple complements classic asset pricing models by combining strong theoretical foundations with AI-driven capabilities, offering an enhanced analysis of residuals unexplained by traditional models. We also demonstrate its potential for practical applications such as portfolio management; (3) We conduct a comprehensive analysis to ensure that the results generated by LLMs in our framework are more logically consistent and credible, thus improving the reliability of insights for financial decision-making.", "keywords": [ "Large Language Models", "Time-Varying Knowledge Graph", "Event Studies", "Asset Pricing" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "o3pJU5QCtv", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Aviv Tamar", "gender": "Male", "institution": "Technion, Technion", "country": "", "position": "Assistant Professor" }, { "name": "Carl Qi", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Dan Haramati", "gender": "Male", "institution": "Brown University", "country": "US", "position": "PhD student" }, { "name": "Tal Daniel", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Amy Zhang", "gender": "Female", "institution": "University of Texas at Austin", "country": "", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 38, "n_ref": 95, "n_ref_all": 130, "n_fig": 17, "n_tab": 10, "L_tab": 1859, "n_element_tab": 123, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 701, "n_element_tab_1": 38, "formula_len_all": 518, "formula_len_all_1": 509, "len_all": 207324, "len_all_1": 71019, "len_abs": 1586, "len_title": 125, "len_sents": 55941, "len_sents_1": 34046, "n_sents": 393, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1588, "title": "EC-Diffuser: Multi-Object Manipulation via Entity-Centric Behavior Generation", "abs": "Object manipulation is a common component of everyday tasks, but learning to manipulate objects from high-dimensional observations presents significant challenges. These challenges are heightened in multi-object environments due to the combinatorial complexity of the state space as well as of the desired behaviors. While recent approaches have utilized large-scale offline data to train models from pixel observations, achieving performance gains through scaling, these methods struggle with compositional generalization in unseen object configurations with constrained network and dataset sizes. To address these issues, we propose a novel behavioral cloning (BC) approach that leverages object-centric representations and an entity-centric Transformer with diffusion-based optimization, enabling efficient learning from offline image data. Our method first decomposes observations into Deep Latent Particles (DLP), which are then processed by our entity-centric Transformer that computes attention at the particle level, simultaneously predicting object dynamics and the agent's actions. Combined with the ability of diffusion models to capture multi-modal behavior distributions, this results in substantial performance improvements in multi-object tasks and, more importantly, enables compositional generalization. We present BC agents capable of zero-shot generalization to perform tasks with novel compositions of objects and goals, including larger numbers of objects than seen during training. We provide video rollouts on our webpage: https://sites.google.com/view/ec-diffuser.", "keywords": [ "Diffusion", "Object-Centric Representation", "Robotic Manipulation" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "o3jgyJIhnv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Fan Jia", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Genghua Kou", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Osamu Yoshie", "gender": "Male", "institution": "Waseda University", "country": "JP", "position": "Professor" }, { "name": "Tiancai Wang", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Weixin Mao", "gender": "Male", "institution": "Waseda University", "country": "JP", "position": "PhD student" }, { "name": "Xiangyu Zhang", "gender": "Male", "institution": "MEGVII Technology", "country": "CN", "position": "Principal Researcher" }, { "name": "Ying Li", "gender": "unknown", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Yingfei Liu", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Yucheng Zhao", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Ziheng Zhang", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 100, "n_ref_all": 120, "n_fig": 4, "n_tab": 3, "L_tab": 914, "n_element_tab": 124, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3446, "n_element_tab_1": 147, "formula_len_all": 81, "formula_len_all_1": 0, "len_all": 154077, "len_all_1": 65599, "len_abs": 976, "len_title": 97, "len_sents": 43953, "len_sents_1": 27857, "n_sents": 359, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 980, "title": "PADriver: Towards Personalized Autonomous Driving", "abs": "In this paper, we propose PADriver, a novel closed-loop framework for personalized autonomous driving (PAD). Built upon Multi-modal Large Language Model (MLLM), PADriver takes streaming frames and personalized textual prompts as inputs. It autoaggressively performs scene understanding, danger level estimation and action decision. The predicted danger level reflects the risk of the potential action and provides an explicit reference for the final action, which corresponds to the preset personalized prompt. Moreover, we construct a closed-loop benchmark named PAD-Highway based on Highway-Env simulator to comprehensively evaluate the decision performance under traffic rules. The dataset contains 250 hours videos with high-quality annotation to facilitate the development of PAD behavior analysis. Experimental results on the constructed benchmark show that PADriver outperforms state-of-the-art approaches on different evaluation metrics, and enables various driving modes.", "keywords": [ "Personalized", "Planning", "Autonomous Driving", "MLLM" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "o3V7OuPxu4", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Keyan Cheng", "gender": "Male", "institution": "Academy of Military Science", "country": "CN", "position": "PhD student" }, { "name": "Minne Li", "gender": "unknown", "institution": "University College London", "country": "", "position": "PhD student" }, { "name": "Yuan Zhou", "gender": "Female", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Zhiyuan Wang", "gender": "Female", "institution": "National University of Defense Technology", "country": "", "position": "PhD student" }, { "name": "Erqiang Xu", "gender": "Male", "institution": "GienTech", "country": "CN", "position": "Researcher" }, { "name": "wenjietang", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 49, "n_ref": 94, "n_ref_all": 103, "n_fig": 2, "n_tab": 5, "L_tab": 596, "n_element_tab": 30, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2174, "n_element_tab_1": 75, "formula_len_all": 555, "formula_len_all_1": 315, "len_all": 197183, "len_all_1": 61273, "len_abs": 1381, "len_title": 151, "len_sents": 40737, "len_sents_1": 29776, "n_sents": 247, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1386, "title": "StarCraft II Arena: Evaluating LLMs in Strategic Planning, Real-Time Decision Making, and Adaptability", "abs": "StarCraft II plays an important role in developing AI agents for real-time strategic reasoning due to its complex nature. However, people usually draw conclusions of how competent their agents are according to the level of the built-in agents in StarCraft II which they can win in terms of the final success rate. Little intermediate quantitative information is considered while human-in-the-loop analysis is time inefficient, which results in inadequate reflection of the true strategic reasoning ability. In this work, we propose StarCraft II Arena, a well-designed benchmark for evaluating the strategic planning, real-time decision-making, and adaptability capabilities of large language models (LLMs) agents. We introduce using fine-grained capability metrics, allowing for targeted capture and analysis of specific capability, and further propose a detailed decision trace to enhance the understanding of LLM behavior. We demonstrate the utility of such a benchmark by evaluating several state-of-the-art LLMs in various setups. Our results reveal distinct performances in long-term strategy development, real-time decision-making, and adapting to environmental changes. Such results show that the StarCraft II Arena offers a deeper insight into the decision-making process of LLMs and has the potential to become a challenging and comprehensive benchmark for strategic reasoning.", "keywords": [ "benchmark evaluation", "large language model", "LLM-based agent", "strategic reasoning", "real-time decision-making." ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "o362EkNU2z", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 15, "author_info_list": [ { "name": "Chen Zhang", "gender": "Female", "institution": "Bytedance", "country": "CN", "position": "Research Scientist" }, { "name": "Qian Yang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Rui Liu", "gender": "Male", "institution": "Inner Mongolia University", "country": "CN", "position": "Full Professor" }, { "name": "Ruiqi Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Shengpeng Ji", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Wenrui Liu", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Xiang Yin", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Xiaoda Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yi Ren", "gender": "Male", "institution": "ByteDance", "country": "SG", "position": "Researcher" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhenhui Ye", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Ziyue Jiang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhang boyang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Jialung Zuo", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 67, "n_ref": 186, "n_ref_all": 224, "n_fig": 8, "n_tab": 20, "L_tab": 3996, "n_element_tab": 464, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2373, "n_element_tab_1": 249, "formula_len_all": 1068, "formula_len_all_1": 1016, "len_all": 242765, "len_all_1": 67971, "len_abs": 2288, "len_title": 133, "len_sents": 63753, "len_sents_1": 29600, "n_sents": 473, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1578, "title": "Sparse Alignment Enhanced Latent Diffusion Transformer for Zero-Shot Speech Synthesis", "abs": "While recent zero-shot text-to-speech (TTS) models have significantly improved speech quality and expressiveness, \nmainstream systems still suffer from issues related to speech-text alignment modeling: 1) autoregressive large language models are inefficient and not robust in long-sentence inference; 2) non-autoregressive diffusion models without explicit speech-text alignment require substantial model capacity for alignment learning; 3) \\textcolor{red}{predefined alignment-based diffusion models suffer from naturalness constraints of forced alignments} and a complicated inference pipeline. This paper introduces \\textit{S-DiT}, a TTS system featuring an innovative sparse alignment algorithm that guides the latent diffusion transformer (DiT). Specifically, 1) we provide sparse alignment boundaries to S-DiT to reduce the difficulty of alignment learning without limiting \\textcolor{red}{the search space}; 2) to simplify the overall pipeline, we propose a unified frontend language model (F-LM) training framework to cover various speech processing tasks required by TTS models. Additionally, we adopt the piecewise rectified flow technique to accelerate the generation process and employ a multi-condition classifier-free guidance strategy for accent intensity adjustment. Experiments demonstrate that S-DiT matches state-of-the-art zero-shot TTS speech quality while maintaining a more efficient pipeline. Moreover, our system can generate high-quality one-minute speech with only 8 sampling steps. Audio samples are available at https://sditdemo.github.io/sditdemo/.", "keywords": [ "Zero-Shot Speech Synthesis", "Large-Scale TTS", "Accented TTS" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 1, 4, 2 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "o2uHg0Skil", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Marcus Hutter", "gender": "unknown", "institution": "DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Michael K. Cohen", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "Postdoc" }, { "name": "Stuart Russell", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 6, "n_ref_uni": 39, "n_ref": 67, "n_ref_all": 77, "n_fig": 3, "n_tab": 2, "L_tab": 187, "n_element_tab": 9, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 187, "n_element_tab_1": 9, "formula_len_all": 1657, "formula_len_all_1": 815, "len_all": 180115, "len_all_1": 75209, "len_abs": 1124, "len_title": 87, "len_sents": 64605, "len_sents_1": 36794, "n_sents": 457, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1127, "title": "RL, but don't do anything I wouldn't do", "abs": "In reinforcement learning, if the agent's reward differs from the designers' true utility, even only rarely, the state distribution resulting from the agent's policy can be very bad, in theory and in practice. When RL policies would devolve into undesired behavior, a common countermeasure is KL regularization to a trusted policy (\"Don't do anything I wouldn't do\"). All current cutting-edge language models are RL agents that are KL-regularized to a \"base policy\" that is purely predictive. Unfortunately, we demonstrate that when this base policy is a Bayesian predictive model of a trusted policy, the KL constraint is no longer reliable for controlling the behavior of an advanced RL agent. We demonstrate this theoretically using algorithmic information theory, and while systems today are too weak to exhibit this theorized failure precisely, we RL-finetune a language model and find evidence that our formal results are plausibly relevant in practice. We also propose a theoretical alternative that avoids this problem by replacing the \"Don't do anything I wouldn't do\" principle with \"Don't do anything I mightn't do\".", "keywords": [ "AI safety", "Superalignment", "Algorithmic information theory", "Kolmogorov complexity", "Reinforcement learning", "Large language models" ], "rating_list": [ 6, 3, 8, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 1, 4, 3 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 2, 4, 2, 2 ] }, { "paper_id": "o2o1XNeI1b", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Ge Liu", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Heng Ji", "gender": "Female", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Kuan-Hao Huang", "gender": "Male", "institution": "Texas A&M University", "country": "US", "position": "Assistant Professor" }, { "name": "Martin D. Burke", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "Full Professor" }, { "name": "Thao Nguyen", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Ying Diao", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 8, "n_ref_uni": 34, "n_ref": 85, "n_ref_all": 112, "n_fig": 12, "n_tab": 5, "L_tab": 2416, "n_element_tab": 219, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 235, "n_element_tab_1": 21, "formula_len_all": 777, "formula_len_all_1": 429, "len_all": 167040, "len_all_1": 59942, "len_abs": 1427, "len_title": 65, "len_sents": 54561, "len_sents_1": 31579, "n_sents": 358, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1433, "title": "FARM: Functional Group-Aware Representations for Small Molecules", "abs": "We introduce Functional Group-Aware Representations for Small Molecules (FARM), a novel foundation model designed to bridge the gap between SMILES, natural language, and molecular graphs. The key innovation of FARM lies in its functional group-aware tokenization, which directly incorporates functional group information into SMILES, enriching SMILES with detailed chemical context. For example, instead of using \u201cO\u201d to represent all oxygen atoms, we use specific tokens like \u201cO_ketone\u201d and \u201cO_hydroxyl\u201d to differentiate oxygen atoms belonging to distinct functional groups. This tokenization expands the chemical lexicon, thereby more effectively bridging SMILES and natural language, ultimately enhancing the model\u2019s ability to predict molecular properties. FARM also represents molecules from two perspectives: by using masked language modeling to capture atom-level features and by employing graph neural networks to encode the whole molecule topology. FARM leverages contrastive learning to aligns these two views of representations into a unified molecular embedding. We rigorously evaluate FARM on the MoleculeNet dataset, where it achieves state-of-the-art performance on 11 out of 13 tasks. These results highlight FARM\u2019s potential to improve molecular representation learning and demonstrate its strong transfer learning capabilities, paving the way for promising applications in drug discovery and pharmaceutical research.", "keywords": [ "Molecular representation learning", "Masked language modeling", "Contrastive learning" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 5, 4, 1, 5 ] }, { "paper_id": "o2arTYxsXd", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Binh T. Nguyen", "gender": "Male", "institution": "Ho Chi Minh city University of Science, Vietnam National University", "country": "", "position": "Associate Professor" }, { "name": "Thinh Nguyen", "gender": "Male", "institution": "VinUniversity", "country": "VN", "position": "PhD student" }, { "name": "Danh Le Phuoc", "gender": "Male", "institution": "TU Berlin", "country": "DE", "position": "Principal Researcher" }, { "name": "Khoa Doan", "gender": "unknown", "institution": "VinUniversity", "country": "VN", "position": "Assistant Professor" }, { "name": "Kok Seng Wong", "gender": "Male", "institution": "VinUniversity", "country": "VN", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 58, "n_ref": 179, "n_ref_all": 218, "n_fig": 18, "n_tab": 11, "L_tab": 5358, "n_element_tab": 582, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 3466, "n_element_tab_1": 173, "formula_len_all": 798, "formula_len_all_1": 748, "len_all": 166568, "len_all_1": 63639, "len_abs": 1577, "len_title": 158, "len_sents": 40889, "len_sents_1": 26173, "n_sents": 319, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 110, "L_abs": 1584, "title": "Overcoming Catastrophic Forgetting in Federated Class-Incremental Learning via Federated Global Twin Generator", "abs": "Federated Class-Incremental Learning (FCIL) increasingly becomes essential in the decentralized setting, where it enables multiple participants to collaboratively train a global model to perform well on a sequence of tasks without sharing their private data. In FCIL, conventional Federated Learning algorithms such as FedAvg often suffer from catastrophic forgetting, resulting in significant performance declines on earlier tasks. Recent works based on generative models produce synthetic images to help mitigate this issue across all classes. However, these approaches' testing accuracy in previous classes is still much lower than recent classes, i.e., having better plasticity than stability. To overcome these issues, this paper presents Federated Global Twin Generator (FedGTG), an FCIL framework that exploits generative-model training on the global side without accessing client data. Specifically, the server trains a data generator and a feature generator to create two types of information from all seen classes. Then, it sends the synthetic data to the client. The clients then use feature-direction-controlling losses to make the local models retain knowledge and learn new tasks well. We extensively analyze the robustness of FedGTG on natural images and its ability to converge to flat local minima and achieve better predicting confidence (calibration). Experimental results on CIFAR-10, CIFAR-100, and tiny-ImageNet demonstrate the improvements in accuracy and forgetting measures of FedGTG as well as the robustness of domain shifts compared to previous frameworks.", "keywords": [ "federated learning", "continual learning", "federated continual learning", "generative model" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "o2Igqm95SJ", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 2, "author_info_list": [ { "name": "Antoine Cully", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Assistant Professor" }, { "name": "Maxence Faldor", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 5, "n_ref_uni": 21, "n_ref": 40, "n_ref_all": 52, "n_fig": 6, "n_tab": 4, "L_tab": 1103, "n_element_tab": 111, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1009, "n_element_tab_1": 71, "formula_len_all": 54, "formula_len_all_1": 275, "len_all": 91103, "len_all_1": 53878, "len_abs": 1497, "len_title": 89, "len_sents": 30548, "len_sents_1": 26058, "n_sents": 208, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 41, "L_abs": 1411, "title": "CAX: Cellular Automata Accelerated in JAX", "abs": "Cellular automata have become a cornerstone for investigating emergence and self-organization across diverse scientific disciplines, spanning neuroscience, artificial life, and theoretical physics. However, the absence of a hardware-accelerated cellular automata library limits the exploration of new research directions, hinders collaboration, and impedes reproducibility. In this work, we introduce CAX (Cellular Automata Accelerated in JAX), a high-performance and flexible open-source library designed to accelerate cellular automata research. CAX offers cutting-edge performance and a modular design through a user-friendly interface, and can support both discrete and continuous cellular automata with any number of dimensions. We demonstrate CAX's performance and flexibility through a wide range of benchmarks and applications. From classic models like elementary cellular automata and Conway's Game of Life to advanced applications such as growing neural cellular automata and self-classifying MNIST digits, CAX speeds up simulations up to 2,000 times faster. Furthermore, we demonstrate CAX's potential to accelerate research by presenting a collection of three novel cellular automata experiments, each implemented in just a few lines of code thanks to the library's modular architecture. Notably, we show that a simple one-dimensional cellular automaton can outperform GPT-4 on the 1D-ARC challenge.", "keywords": [ "cellular automata", "emergence", "self-organization", "neural cellular automata" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 4, 2, 2, 3 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "o2Gg2tSKBn", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Bin Ke", "gender": "unknown", "institution": "National University of Singapore", "country": "", "position": "Full Professor" }, { "name": "Dading Chong", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Jiageng Wu", "gender": "Male", "institution": "Brigham and Women's Hospital, Harvard University", "country": "US", "position": "Intern" }, { "name": "Jie Yang", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "Peilin Zhou", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Xinyu Shi", "gender": "Female", "institution": "Southwest University of Finance and Economics", "country": "", "position": "MS student" }, { "name": "Yikang Jiang", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Ziyue Xu", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Research Assistant" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 43, "n_ref_all": 66, "n_fig": 5, "n_tab": 11, "L_tab": 7600, "n_element_tab": 549, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3131, "n_element_tab_1": 277, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 151760, "len_all_1": 69714, "len_abs": 2280, "len_title": 143, "len_sents": 46241, "len_sents_1": 30299, "n_sents": 329, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1983, "title": "FinTruthQA: A Benchmark Dataset for Evaluating the Quality of Financial Information Disclosure", "abs": "Accurate and transparent financial information disclosure is crucial in the fields of accounting and finance, ensuring market efficiency and investor confidence. Among many information disclosure platforms, the Chinese stock exchanges' investor interactive platform provides a novel and interactive way for listed firms to disclose information of interest to investors through an online question-and-answer (Q\\&A) format. However, it is common for listed firms to respond to questions with limited or no substantive information, and automatically evaluating the quality of financial information disclosure on large amounts of Q\\&A pairs is challenging. This paper builds a benchmark FinTruthQA, that can evaluate advanced natural language processing (NLP) techniques for the automatic quality assessment of information disclosure in financial Q\\&A data. FinTruthQA comprises 6,000 real-world financial Q\\&A entries and each Q\\&A was manually annotated based on four conceptual dimensions of accounting: *question identification*, *question relevance*, *answer readability*, and *answer relevance*. We benchmarked various NLP techniques on FinTruthQA, including statistical machine learning models, pre-trained language model and their fine-tuned versions, as well as large language models (LLMs). Experiments showed that existing NLP models have strong predictive ability for *question identification* and *question relevance* tasks, but are suboptimal for *answer readability* and *answer relevance* tasks. By establishing this benchmark, we provide a robust foundation for the automatic evaluation of information disclosure, significantly enhancing the transparency and quality of financial reporting. FinTruthQA can be used by auditors, regulators, and financial analysts for real-time monitoring and data-driven decision-making, as well as by researchers for advanced studies in accounting and finance, ultimately fostering greater trust and efficiency in the financial markets.", "keywords": [ "Natural Language Processing (NLP)", "Financial Disclosure", "Information Quality Assessment", "Financial Q&A" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "o1efpbvR6v", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Milo Roucairol", "gender": "Male", "institution": ", Universit\u00e9 Paris-Dauphine (Paris IX)", "country": "FR", "position": "PhD student" }, { "name": "Ngoc Trinh Hung NGUYEN", "gender": "Male", "institution": ", Universit\u00e9 Paris-Dauphine (Paris IX)", "country": "FR", "position": "MS student" }, { "name": "Tristan Cazenave", "gender": "Male", "institution": "Universit\u00e9 Paris-Dauphine (Paris IX)", "country": "FR", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 24, "n_ref": 49, "n_ref_all": 71, "n_fig": 3, "n_tab": 7, "L_tab": 1604, "n_element_tab": 141, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1927, "n_element_tab_1": 144, "formula_len_all": 114, "formula_len_all_1": 115, "len_all": 101972, "len_all_1": 58955, "len_abs": 918, "len_title": 111, "len_sents": 36103, "len_sents_1": 26720, "n_sents": 247, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 927, "title": "Application of Metric Transformation in One-Step Retrosynthesis", "abs": "In this article, we investigate the impact of Deep Metric Learning and Transformer architecture on predicting the retrosynthesis of Simplified Molecular Input Line Entry System (SMILES) chemical compounds.\n\nWe demonstrate that combining the Attention mechanism with Proxy Anchor Loss is effective for classification tasks due to its strengths in capturing both local and global contexts and differentiating between various classes. \n\nOur approach, which requires no prior chemical knowledge, achieves promising results on the USPTO-FULL dataset, with accuracies of 53.4\\%, 83.8\\%, 90.6\\%, and 97.5\\% for top-1, top-5, top-10, and top-50 predictions, respectively.\n\nWe further validate the practical application of our approach by correctly predicting the retrosynthesis pathways for 63 out of 100 randomly selected compounds from the ChEMBL database and for 39 out of 60 compounds selected by Bayer's chemists and from PubChem.", "keywords": [ "Retrosynthesis", "Chemistry", "Deep Metric Learning", "Transformer" ], "rating_list": [ 5, 1, 1 ], "soundness_list": [ 2, 1, 1 ], "presentation_list": [ 2, 1, 1 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "o1SGGW53GF", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Fatema Ahmad", "gender": "Female", "institution": "Hamad Bin Khalifa University", "country": "", "position": "Researcher" }, { "name": "Firoj Alam", "gender": "Male", "institution": "Qatar Computing Research Institute", "country": "", "position": "Scientist" }, { "name": "Maram Hasanain", "gender": "Female", "institution": "Qatar Computing Research Institute", "country": "", "position": "Postdoc" }, { "name": "Md. Arid Hasan", "gender": "Male", "institution": "University of New Brunswick", "country": "", "position": "MS student" }, { "name": "Mucahid Kutlu", "gender": "unknown", "institution": "University of Qatar", "country": "QA", "position": "Assistant Professor" }, { "name": "Sahinur Rahman Laskar", "gender": "Male", "institution": "UPES", "country": "IN", "position": "Assistant Professor" }, { "name": "Shammur Absar Chowdhury", "gender": "Female", "institution": "Qatar Computing Research Institute", "country": "", "position": "Scientist" }, { "name": "Sunaya Upadhyay", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "Undergrad student" }, { "name": "Sukhadia Vrunda", "gender": "Female", "institution": "Qatar Computing Research Institute", "country": "QA", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 37, "n_ref": 57, "n_ref_all": 109, "n_fig": 10, "n_tab": 13, "L_tab": 5043, "n_element_tab": 282, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2762, "n_element_tab_1": 122, "formula_len_all": 234, "formula_len_all_1": 234, "len_all": 189148, "len_all_1": 63000, "len_abs": 1302, "len_title": 114, "len_sents": 52661, "len_sents_1": 25743, "n_sents": 498, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1325, "title": "NativQA: Multilingual Culturally-Aligned Natural Queries for LLMs", "abs": "Natural Question Answering (QA) datasets play a crucial role in evaluating the capabilities of large language models (LLMs), ensuring their effectiveness in real-world applications. Despite the numerous QA datasets that have been developed, there is a notable lack of region-specific datasets generated by native users in their own languages. This gap hinders the effective benchmarking of LLMs for regional and cultural specificities. Furthermore, it also limits the development of fine-tuned models. In this study, we propose a scalable, language-independent framework, NativQA, to seamlessly construct culturally and regionally aligned QA datasets in native languages, for LLM evaluation and tuning. We demonstrate the efficacy of the proposed framework by designing a multilingual natural QA dataset, \\mnqa, consisting of ~64k manually annotated QA pairs in seven languages, ranging from high to extremely low resource, based on queries from native speakers from 9 regions covering 18 topics. We benchmark open- and closed-source LLMs with the MultiNativQA dataset. We also showcase the framework efficacy in constructing fine-tuning data especially for low-resource and dialectally-rich languages. We made both the framework NativQA and MultiNativQA dataset publicly available for the community (https://anonymous.com/).", "keywords": [ "resources for less-resourced languages", "multilingual benchmarks", "multilingual corpora", "NLP datasets", "datasets for low resource languages" ], "rating_list": [ 8, 8, 3, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "o1IiiNIoaA", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Cyril Furtlehner", "gender": "Male", "institution": "INRIA", "country": "", "position": "Researcher" }, { "name": "Nilo Schwencke", "gender": "Male", "institution": "Universit\u00e9 Paris-Saclay", "country": "FR", "position": "PhD student" } ], "n_formula": 87, "n_formula_1": 28, "n_ref_uni": 42, "n_ref": 60, "n_ref_all": 117, "n_fig": 0, "n_tab": 18, "L_tab": 2569, "n_element_tab": 321, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 253, "n_element_tab_1": 19, "formula_len_all": 7253, "formula_len_all_1": 2316, "len_all": 204396, "len_all_1": 56097, "len_abs": 1671, "len_title": 130, "len_sents": 51326, "len_sents_1": 21839, "n_sents": 433, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 861, "title": "ANaGRAM: A Natural Gradient Relative to Adapted Model for efficient PINNs learning", "abs": "In the recent years, Physics Informed Neural Networks (PINNs) have received strong interest as a method to solve PDE driven systems, in particular for data assimilation purpose. This method is still in its infancy, with many shortcomings and failures that remain not properly understood.\nIn this paper we propose a natural gradient approach to PINNs which contributes to speed-up and improve the accuracy of the training.\nBased on an in depth analysis of the differential geometric structures of the problem, we come up with two distinct contributions:\n(i) a new natural gradient algorithm that scales as $\\min(P^2S, S^2P)$, where $P$ is the number of parameters, and $S$ the batch size;\n(ii) a mathematically principled reformulation of the PINNs problem that allows the extension of natural gradient to it, with proved connections to Green's function theory.", "keywords": [ "PINNs", "SciML", "PDEs", "Natural Gradient", "Neural Tangent Kernel" ], "rating_list": [ 3, 3, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 2, 4, 2, 4, 3 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 2, 4 ] }, { "paper_id": "o1Et3MogPw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Chen Qian", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Cheng Yang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Associate Professor" }, { "name": "Chenyang Zhao", "gender": "Male", "institution": "UCLA Computer Science Department, University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Maosong Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Ran Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Research Assistant" }, { "name": "Ruobing Xie", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Senior researcher" }, { "name": "Weize Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhiyuan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Ziming You", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "yitong guan", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 0, "n_ref_uni": 27, "n_ref": 83, "n_ref_all": 105, "n_fig": 8, "n_tab": 8, "L_tab": 2590, "n_element_tab": 369, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 1079, "n_element_tab_1": 96, "formula_len_all": 3329, "formula_len_all_1": 0, "len_all": 279216, "len_all_1": 70157, "len_abs": 1405, "len_title": 138, "len_sents": 84624, "len_sents_1": 31735, "n_sents": 588, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 88, "L_abs": 1467, "title": "Internet of Agents: Weaving a Web of Heterogeneous Agents for Collaborative Intelligence", "abs": "The rapid advancement of large language models (LLMs) has paved the way for the development of highly capable autonomous agents. However, existing multi-agent frameworks often struggle with integrating diverse capable third-party agents due to reliance on agents defined within their own ecosystems. They also face challenges in simulating distributed environments, as most frameworks are limited to single-device setups. Furthermore, these frameworks often rely on hard-coded communication pipelines, limiting their adaptability to dynamic task requirements. Inspired by the concept of the Internet, we propose the Internet of Agents (IoA), a novel framework that addresses these limitations by providing a flexible and scalable platform for LLM-based multi-agent collaboration. IoA introduces an agent integration protocol, an instant-messaging-like architecture design, and dynamic mechanisms for agent teaming and conversation flow control. Through extensive experiments on general assistant tasks, embodied AI tasks, and retrieval-augmented generation benchmarks, we demonstrate that IoA consistently outperforms state-of-the-art baselines, showcasing its ability to facilitate effective collaboration among heterogeneous agents. IoA represents a step towards linking diverse agents in an Internet-like environment, where agents can seamlessly collaborate to achieve greater intelligence and capabilities. We will release our code to facilitate further research.", "keywords": [ "llm agent", "multi-agent" ], "rating_list": [ 8, 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 4, 4, 1, 4, 3 ], "contribution_list": [ 4, 3, 3, 4, 3 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "o10clUzFRH", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Edward James Young", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Yashar Ahmadian", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "University Lecturer" } ], "n_formula": 73, "n_formula_1": 46, "n_ref_uni": 19, "n_ref": 35, "n_ref_all": 100, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6651, "formula_len_all_1": 3013, "len_all": 120123, "len_all_1": 64698, "len_abs": 749, "len_title": 89, "len_sents": 34842, "len_sents_1": 26186, "n_sents": 347, "n_sents_1": 238, "venue": "ICLR2025", "decision": "withdraw", "L_title": 40, "L_abs": 766, "title": "R\u00e9nyi Regularised Reinforcement Learning", "abs": "Entropy regularisation has proven effective in reinforcement learning (RL) for encouraging exploration. Recent work demonstrating the equivalence between entropy regularised RL and approximate probabilistic inference suggests the potential for improving existing methods by generalising the inference procedure. We develop the R\u00e9nyi regularised RL framework by using R\u00e9nyi variational inference to learn a stochastic policy. We present theoretical results for policy evaluation and improvement within this new framework. Additionally, we propose two novel algorithms, $\\alpha$-SAC and $\\alpha$-SQL, for large-scale RL tasks. We show that these algorithms attain higher returns on games from the Atari suite relative to an entropy-regularised benchmark, SAC-Discrete.", "keywords": [ "Reinforcement Learning", "Variational Inference", "R\u00e9nyi divergence" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 1, 4, 2, 3 ], "presentation_list": [ 2, 4, 4, 2 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "o0qrehZW94", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Chenfeng Xu", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Chensheng Peng", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Chongjian GE", "gender": "Male", "institution": "The University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Mingyu Ding", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Wei Zhan", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Yuanfeng Ji", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Masayoshi Tomizuka", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Ping Luo", "gender": "unknown", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Varun Jampani", "gender": "Male", "institution": "Stability AI", "country": "US", "position": "Principal Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 41, "n_ref": 91, "n_ref_all": 126, "n_fig": 9, "n_tab": 4, "L_tab": 2903, "n_element_tab": 283, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 2052, "n_element_tab_1": 126, "formula_len_all": 291, "formula_len_all_1": 291, "len_all": 172804, "len_all_1": 65415, "len_abs": 1737, "len_title": 156, "len_sents": 45342, "len_sents_1": 31875, "n_sents": 290, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1735, "title": "CompGS: Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians", "abs": "Recent breakthroughs in text-guided image generation have significantly advanced the field of 3D generation. While generating a single high-quality 3D object is now feasible, generating multiple objects with reasonable interactions within a 3D space, a.k.a. compositional 3D generation, presents substantial challenges. This paper introduces CompGS, a novel generative framework that employs 3D Gaussian Splatting (GS) for efficient, compositional text-to-3D content generation. To achieve this goal, two core designs are proposed: (1) 3D Gaussians Initialization with 2D compositionality: We transfer the well-established 2D compositionality to initialize the Gaussian parameters on an entity-by-entity basis, ensuring both consistent 3D priors for each entity and reasonable interactions among multiple entities; (2) Dynamic Optimization: We propose a dynamic strategy to optimize 3D Gaussians using Score Distillation Sampling (SDS) loss. CompGS first automatically decomposes 3D Gaussians into distinct entity parts, enabling optimization at both the entity and composition levels. Additionally, CompGS optimizes across objects of varying scales by dynamically adjusting the spatial parameters of each entity, enhancing the generation of fine-grained details, particularly in smaller entities. Qualitative comparisons and quantitative evaluations on T3Bench demonstrate the effectiveness of CompGS in generating compositional 3D objects with superior image quality and semantic alignment over existing methods. CompGS can also be easily extended to controllable 3D editing, facilitating scene generation. We hope CompGS will provide new insights to the compositional 3D generation. Codes will be released to the research community.", "keywords": [ "Compositional 3D Generation", "Gaussian Splatting" ], "rating_list": [ 5, 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 4, 3, 4 ], "presentation_list": [ 4, 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 5, 4, 4 ] }, { "paper_id": "o0X0CPl320", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Khalil Al Handawi", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Instructor" }, { "name": "Logan Labossiere", "gender": "Male", "institution": "McGill University, McGill University", "country": "CA", "position": "Undergrad student" } ], "n_formula": 20, "n_formula_1": 17, "n_ref_uni": 20, "n_ref": 38, "n_ref_all": 89, "n_fig": 21, "n_tab": 9, "L_tab": 3819, "n_element_tab": 438, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1121, "n_element_tab_1": 151, "formula_len_all": 2517, "formula_len_all_1": 2344, "len_all": 116056, "len_all_1": 52448, "len_abs": 1320, "len_title": 144, "len_sents": 37833, "len_sents_1": 22288, "n_sents": 288, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1336, "title": "Jointly Training Task-Specific Encoders and Downstream Models on Heterogeneous Multiplex Graphs", "abs": "Learning representations on Heterogeneous Multiplex Graphs (HMGs) is an active field of study, driven by the need for generating expressive, low-dimensional embeddings to support downstream machine learning tasks. A key component of this process is the design of the graph processing pipeline, which directly impacts the quality of learned representations. Information fusion techniques, which aggregate information across layers of a multiplex graph, have been shown to improve the performance of Graph Neural Network (GNN)-based architectures on various tasks including node classification, edge prediction, and graph-level classification. \nRecent research has explored fusion strategies at different stages of the processing pipeline, leading to graph-, GNN-, embedding-, and prediction-level approaches. In this work, we propose a model extending the $\\texttt{GraphSAGE}$ architecture, which simultaneously refines layer-wise embeddings produced by the encoder while training downstream models. We evaluate the model's effectiveness on an HMG on real-world and benchmark datasets, comparing it to models utilizing either graph-level or prediction-level fusion without jointly optimizing their vector embeddings. We demonstrate that our approach enhances the model's performance on downstream tasks, particularly node classification.", "keywords": [ "Multiplex", "Heterogeneous", "GNN", "GraphSAGE" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "o0DQDGaVIY", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Dajiang Zhu", "gender": "Male", "institution": "University of Texas at Arlington", "country": "", "position": "Assistant Professor" }, { "name": "Xiaowei Yu", "gender": "Male", "institution": "University of Texas at Arlington, University of Texas at Arlington", "country": "", "position": "PhD student" }, { "name": "Yao Xue", "gender": "Female", "institution": "Southern Illinois University-Carbondale", "country": "", "position": "PhD student" }, { "name": "Zhe Huang", "gender": "Male", "institution": "Tufts University", "country": "", "position": "PhD student" } ], "n_formula": 87, "n_formula_1": 20, "n_ref_uni": 47, "n_ref": 72, "n_ref_all": 106, "n_fig": 5, "n_tab": 10, "L_tab": 5083, "n_element_tab": 280, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 278, "n_element_tab_1": 5, "formula_len_all": 10165, "formula_len_all_1": 2318, "len_all": 202229, "len_all_1": 53930, "len_abs": 1275, "len_title": 144, "len_sents": 43960, "len_sents_1": 21026, "n_sents": 425, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1242, "title": "Noise-Augmented Deep Neural Networks for Image Classification: Insights from Information Theory", "abs": "In this study, we explore the impact of proactively injecting noise into deep learning models, focusing particularly on classification problems, such as image classification and domain adaptation. While noise is typically seen as harmful, our findings reveal that, under certain conditions, noise can beneficially influence the entropy of the system, enhancing the learning outcomes. We employ information entropy to characterize the complexity of the learning tasks and categorize noise into two types, positive noise (PN) and harmful noise (HN), based on whether it helps reduce task complexity. We theoretically prove that positive noise reduces task complexity and demonstrate the presence of positive noise through extensive experiments on Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). We further propose NoisyNN, an innovative approach to leverage positive noise. NoisyNN achieves state-of-the-art performance on various image classification and domain adaptation tasks. Extensive experiments conducted on {15 datasets}, including popular image datasets and out-of-distribution datasets, demonstrate the efficacy of our method. Our study provides the community with a new paradigm for improving model performance.", "keywords": [ "Image Classification", "ViT", "Noise" ], "rating_list": [ 1, 5, 5, 6 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "o0BEth0Po3", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Briti Gangopadhyay", "gender": "Female", "institution": "Sony Group Coorporation", "country": "IN", "position": "Researcher" }, { "name": "Jia-Fong Yeh", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "PhD student" }, { "name": "Shingo Takamatsu", "gender": "Male", "institution": "Sony Group Corporation", "country": "JP", "position": "Researcher" }, { "name": "Zhao Wang", "gender": "Male", "institution": "Sony ", "country": "JP", "position": "Researcher" } ], "n_formula": 26, "n_formula_1": 10, "n_ref_uni": 47, "n_ref": 87, "n_ref_all": 104, "n_fig": 7, "n_tab": 6, "L_tab": 2273, "n_element_tab": 432, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 3625, "n_element_tab_1": 326, "formula_len_all": 2588, "formula_len_all_1": 776, "len_all": 151735, "len_all_1": 57729, "len_abs": 1009, "len_title": 73, "len_sents": 39482, "len_sents_1": 21887, "n_sents": 318, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1013, "title": "ExID: Offline RL with Intuitive Expert Insights in Limited-Data Settings", "abs": "With the ability to learn from static datasets, Offline Reinforcement Learning (RL) emerges as a compelling avenue for real-world applications. However, state-of-the-art offline RL algorithms perform sub-optimally when confronted with limited data confined to specific regions within the state space. The performance degradation is attributed to the inability of offline RL algorithms to learn appropriate actions for rare or unseen observations. This paper proposes a novel domain knowledge-based regularization technique and adaptively refines the initial domain knowledge to considerably boost performance in limited data with partially omitted states. The key insight is that the regularization term mitigates erroneous actions for sparse samples and unobserved states covered by domain knowledge. Empirical evaluations on standard offline RL datasets demonstrate a substantial average performance increase compared to ensemble of domain knowledge and existing offline RL algorithms operating on limited data.", "keywords": [ "Offline Reinforcement Learning", "Knowledge distillation" ], "rating_list": [ 3, 3, 6, 5, 3 ], "soundness_list": [ 1, 2, 2, 3, 2 ], "presentation_list": [ 2, 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 3, 3 ] }, { "paper_id": "nzjSvVZBIp", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Hu Ding", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Jiawei Huang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" } ], "n_formula": 76, "n_formula_1": 20, "n_ref_uni": 81, "n_ref": 179, "n_ref_all": 279, "n_fig": 13, "n_tab": 10, "L_tab": 3014, "n_element_tab": 407, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 2, "n_element_tab_1": 1, "formula_len_all": 6626, "formula_len_all_1": 1664, "len_all": 295672, "len_all_1": 75979, "len_abs": 1365, "len_title": 139, "len_sents": 85153, "len_sents_1": 34428, "n_sents": 662, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1377, "title": "An Effective Manifold-based Optimization Method for Distributionally Robust Classification", "abs": "How to promote the robustness of existing deep learning models is a challenging problem for many practical classification tasks. Recently, Distributionally Robust Optimization (DRO) methods have shown promising potential to tackle this problem. These methods aim to construct reliable models by minimizing the worst-case risk within a local region (called ''uncertainty set'') around the empirical data distribution. However, conventional DRO methods tend to be overly pessimistic, leading to certain discrepancy between the real data distribution and the uncertainty set, which can degrade the classification performance. To address this issue, we propose a manifold-based DRO method that takes the geometric structure of training data into account for constructing the uncertainty set. Specifically, our method employs a carefully designed ''game'' that integrates contrastive learning with Jacobian regularization to capture the manifold structure, enabling us to solve DRO problems constrained by the data manifold. By utilizing a novel idea for approximating geodesic distance on manifolds, we also provide the theoretical guarantees for its robustness. Moreover, our proposed method is easy to implement in practice. We conduct a set of experiments on several popular benchmark datasets, where the results demonstrate our advantages in terms of accuracy and robustness.", "keywords": [ "robustness", "optimization", "representation learning" ], "rating_list": [ 5, 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 5, 3 ] }, { "paper_id": "nzh8Z8d1Zc", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Bingchen Zhao", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Haoqin Tu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Juncheng Wu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Qiao Jin", "gender": "Male", "institution": "National Institutes of Health", "country": "US", "position": "Postdoc" }, { "name": "Siwei Yang", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Yongshuo Zong", "gender": "unknown", "institution": "University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Yunfei Xie", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yuyin Zhou", "gender": "unknown", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "cihang xie", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 54, "n_ref": 140, "n_ref_all": 174, "n_fig": 7, "n_tab": 12, "L_tab": 6619, "n_element_tab": 484, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 5734, "n_element_tab_1": 485, "formula_len_all": 175, "formula_len_all_1": 125, "len_all": 198256, "len_all_1": 64346, "len_abs": 0, "len_title": 70, "len_sents": 49443, "len_sents_1": 25444, "n_sents": 401, "n_sents_1": 168, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1729, "title": "A Preliminary Study of o1 in Medicine: Are We Closer to an AI Doctor?", "abs": "Large language models (LLMs) have exhibited remarkable capabilities across various domains and tasks, pushing the boundaries of our knowledge in learning and cognition. The latest model, OpenAI's o1, stands out as the first LLM with an internalized chain-of-thought technique using reinforcement learning strategies. While it has demonstrated surprisingly strong capabilities on various general language tasks, its performance in specialized fields such as medicine remains unknown. To this end, this report provides a comprehensive exploration of o1 on different medical scenarios, examining 3 key aspects: understanding, reasoning, and multilinguality. Specifically, our evaluation encompasses 6 tasks using data from 37 medical datasets, including two newly constructed and more challenging question-answering (QA) tasks based on professional medical quizzes from the New England Journal of Medicine (NEJM) and The Lancet. These datasets offer greater clinical relevance compared to standard medical QA benchmarks such as MedQA, translating more effectively into real-world clinical utility. Our analysis of o1 suggests that the enhanced reasoning ability of LLMs may (significantly) benefit their capability to understand various medical instructions and reason through complex clinical scenarios. Notably, o1 surpasses the previous GPT-4 in accuracy by an average of 6.2% and 6.6% across 19 datasets and two newly created complex QA scenarios. But meanwhile, we identify several weaknesses in both the model capability and the existing evaluation protocols, including hallucination, inconsistent multilingual ability, and discrepant metrics for evaluation. We will release our raw data and model outputs for future research.", "keywords": [ "benchmark", "large language model" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 3, 2, 4 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "nzgvkQM3EH", "primary_area": "causal reasoning", "n_author": 7, "author_info_list": [ { "name": "Biwei Huang", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Ignavier Ng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Minghao Fu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "MS student" }, { "name": "Yingyao Hu", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Yujia Zheng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zijian Li", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Postdoc" } ], "n_formula": 74, "n_formula_1": 23, "n_ref_uni": 85, "n_ref": 230, "n_ref_all": 259, "n_fig": 14, "n_tab": 12, "L_tab": 4033, "n_element_tab": 247, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 629, "n_element_tab_1": 44, "formula_len_all": 8306, "formula_len_all_1": 2338, "len_all": 287151, "len_all_1": 69251, "len_abs": 2352, "len_title": 140, "len_sents": 88395, "len_sents_1": 31387, "n_sents": 690, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1625, "title": "Identification of Nonparametric Dynamic Causal Model and Latent Process for Climate Analysis", "abs": "The study of learning causal structure with latent variables has advanced the understanding of the world by uncovering causal relationships and latent factors. However, in real-world scenarios, such as those in climate systems, causal relationships are often nonparametric, dynamic, and exist among both observed variables and latent variables. These challenges motivate us to consider a general setting in which causal relations are nonparametric and unrestricted in their occurrence, which is unconventional to current methods. To solve this problem, with the aid of 3-measurement in temporal structure, we theoretically show that both latent variables and processes can be identified up to minor indeterminacy under mild assumptions. Furthermore, we establish that the observed causal structure is identifiable if there is generation variability, roughly speaking, the latent variables induce sufficient variations in generating the noise terms, by the established functional equivalence. The primary idea of this framework is to learn causal representations from causally-related observations, and subsequently address this problem as a task of general nonlinear causal discovery. Based on these theoretical insights, we develop an estimation approach simultaneously learning both the observed causal structure, latent representation, and latent Markov network. Experimental results in simulation studies validate the theoretical foundations and demonstrate the effectiveness of the proposed methodology. In the climate data experiments, we show that it offers a powerful and in-depth understanding of the climate system.", "keywords": [ "causal discovery; causal representation learning; climate analysis" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "nzOD1we8Z4", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 11, "author_info_list": [ { "name": "Akihiro Nitta", "gender": "unknown", "institution": "Kumo.AI", "country": "US", "position": "Software Engineer" }, { "name": "Bla\u017e Stojanovi\u010d", "gender": "Male", "institution": "Kumo.AI", "country": "US", "position": "Researcher" }, { "name": "Jan Eric Lenssen", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Principal Researcher" }, { "name": "Jure Leskovec", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Manan Shah", "gender": "Male", "institution": "Kumo AI", "country": "US", "position": "Researcher" }, { "name": "Matthias Fey", "gender": "Male", "institution": "TU Dortmund University", "country": "", "position": "PhD student" }, { "name": "Weihua Hu", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Xinwei He", "gender": "Male", "institution": "Stanford University", "country": "", "position": "MS student" }, { "name": "Yiwen Yuan", "gender": "Female", "institution": "xAI", "country": "US", "position": "Researcher" }, { "name": "Zecheng Zhang", "gender": "Male", "institution": "Kumo.AI", "country": "US", "position": "Researcher" }, { "name": "Andy Huang", "gender": "Male", "institution": "McGill University, Mila", "country": "CA", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 49, "n_ref": 80, "n_ref_all": 87, "n_fig": 1, "n_tab": 5, "L_tab": 1776, "n_element_tab": 219, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 821, "n_element_tab_1": 122, "formula_len_all": 611, "formula_len_all_1": 532, "len_all": 130421, "len_all_1": 70662, "len_abs": 1167, "len_title": 99, "len_sents": 37281, "len_sents_1": 33217, "n_sents": 271, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1173, "title": "ContextGNN: Beyond Two-Tower Recommendation Systems", "abs": "Recommendation systems predominantly utilize two-tower architectures, which evaluate user-item rankings through the inner product of their respective embeddings. However, one key limitation of two-tower models is that they learn a pair-agnostic representation of users and items. In contrast, pair-wise representations either scale poorly due to their quadratic complexity or are too restrictive on the candidate pairs to rank. To address these issues, we introduce Context-based Graph Neural Networks (ContextGNNs), a novel deep learning architecture for link prediction in recommendation systems. The method employs a pair-wise representation technique for familiar items situated within a user's local subgraph, while leveraging two-tower representations to facilitate the recommendation of exploratory items. A final network then predicts how to fuse both pair-wise and two-tower recommendations into a single ranking of items. We demonstrate that ContextGNN is able to adapt to different data characteristics and outperforms existing methods, both traditional and GNN-based, on a diverse set of practical recommendation tasks, improving performance by 20\\% on average.", "keywords": [ "graph neural networks", "recommendation", "relational deep learning" ], "rating_list": [ 6, 5, 5, 5, 8 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 4, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 4, 4 ] }, { "paper_id": "nyuaoVnVCa", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Ajin George Joseph", "gender": "Male", "institution": "Indian Institute of Technology, Tirupati", "country": "IN", "position": "Assistant Professor" }, { "name": "Vikas Kumar", "gender": "Male", "institution": "Indian Institute of Technology, Tirupati", "country": "", "position": "MS student" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 54, "n_ref": 75, "n_ref_all": 110, "n_fig": 16, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2136, "formula_len_all_1": 1988, "len_all": 146601, "len_all_1": 72709, "len_abs": 1271, "len_title": 137, "len_sents": 48025, "len_sents_1": 35900, "n_sents": 314, "n_sents_1": 240, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1258, "title": "EMERGENCE OF GROUNDED, OPTIMALLY COMPOSITIONAL SPATIAL LANGUAGE AMONG HOMOGENEOUS AGENTS", "abs": "A mechanism of effective communication is integral to human existence. An\nessential aspect of a functional communication scheme among a rational human\npopulation involves an efficient, adaptive, and coherent apparatus to convey one\u2019s goal to others. Such an effective macro characteristic can\nemerge in a finite population through adaptive learning via trial and error\nat the individual (micro) level, with nearly consistent individual learning faculty and experience across the population. In this paper, we study and hypothesize\n pertinent aspects of glossogenetics, specifically primal human communication mechanisms, through computational modeling. In particular, we model the\nprocess as a language game within the fabric of a decentralized, multi-agent\ndeep reinforcement learning setting, where the agents with local learning and neural\ncognitive faculties interact through a series of dialogues. Our homogeneous agents seek to achieve the principle of least effort and overcome the poverty of stimulus through efficient concept selection, guided feedback and mirror learning. In our examinations,\nwe observe the emergence of successful and structured communication among static and dynamic agent populations through consistent and continual learning.", "keywords": [ "multi-agent reinforcement learning", "Language Emergence", "Cultural Evolution" ], "rating_list": [ 3, 1, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 1, 1, 1 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "ny8T8OuNHe", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Abhay Zala", "gender": "unknown", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "MS student" }, { "name": "Han Lin", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Jaemin Cho", "gender": "Male", "institution": "University of North Carolina, Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Mohit Bansal", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 79, "n_ref": 212, "n_ref_all": 283, "n_fig": 25, "n_tab": 17, "L_tab": 6434, "n_element_tab": 829, "n_fig_1": 13, "n_tab_1": 6, "L_tab_1": 2251, "n_element_tab_1": 252, "formula_len_all": 368, "formula_len_all_1": 614, "len_all": 299899, "len_all_1": 56846, "len_abs": 5546, "len_title": 0, "len_sents": 77010, "len_sents_1": 25845, "n_sents": 560, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 103, "L_abs": 1410, "title": "Ctrl-Adapter: An Efficient and Versatile Framework for Adapting Diverse Controls to Any Diffusion Model", "abs": "ControlNets are widely used for adding spatial control to text-to-image diffusion models. However, when it comes to controllable video generation, ControlNets cannot be directly integrated into new backbones due to feature space mismatches, and training ControlNets for new backbones can be a significant burden for many users. Furthermore, applying ControlNets independently to different frames can not effectively maintain object temporal consistency. To address these challenges, we introduce Ctrl-Adapter, an efficient and versatile framework that adds diverse controls to any image/video diffusion models through the adaptation of pretrained ControlNets. Ctrl-Adapter offers strong and diverse capabilities, including image and video control, sparse-frame video control, fine-grained patch-level multi-condition control, zero-shot adaptation to unseen conditions, and supports a variety of downstream tasks beyond spatial control, including video editing, video style transfer, and text-guided motion control. With six diverse U-Net/DiT-based image/video diffusion models (SDXL, PixArt-\u03b1, I2VGen-XL, SVD, Latte, Hotshot-XL), Ctrl-Adapter matches the performance of pretrained ControlNets on COCO and achieves the state-of-the-art on DAVIS 2017 with significantly lower computation (< 10 GPU hours). We provide video examples in https://ctrladapterexamples.github.io and code in the supplementary material.", "keywords": [ "Adapter", "Diffusion", "ControlNet", "Text-to-video Generation", "Image-to-video Generation", "Text-to-image Generation" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "nxZbKWhUeZ", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 9, "author_info_list": [ { "name": "Aoran Xiao", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Postdoc" }, { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Jiaqing Zhang", "gender": "Female", "institution": "Xi'an University", "country": "CN", "position": "PhD student" }, { "name": "Jiaxing Huang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Kai Jiang", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "PhD student" }, { "name": "Weiying Xie", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Yidan Liu", "gender": "Female", "institution": "Hunan University", "country": "CN", "position": "PhD student" }, { "name": "Yunsong Li", "gender": "Male", "institution": "Xidian University ", "country": "CN", "position": "Full Professor" }, { "name": "JIE LEI", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 30, "n_ref": 64, "n_ref_all": 85, "n_fig": 2, "n_tab": 7, "L_tab": 1407, "n_element_tab": 254, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1397, "n_element_tab_1": 233, "formula_len_all": 711, "formula_len_all_1": 711, "len_all": 115069, "len_all_1": 71134, "len_abs": 1287, "len_title": 110, "len_sents": 35919, "len_sents_1": 33048, "n_sents": 200, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1290, "title": "Hough Voting-based Prompt Learning for Segment Anything Model", "abs": "Segment Anything Models (SAMs) like SEEM and SAM have achieved great performance on various downstream datasets at the cost of crafting spatial and semantic prompts. Previous prompt learning methods can learn prompts automatically but largely focus on learning semantic prompts, while how to learn effective spatial prompts that are important to SAMs is largely under-explored. Inspired by Hough Voting that detects a complex object by voting from its parts, we propose Hough Voting-based Spatial Prompt Learning (HoughSpaPL) that designs three types of voting mechanisms to learn three distinct spatial prompts for different subregions of the visual concept (e.g., things and stuff), which capture complementary spatial clues and vote together to guide SAMs to generate a precise segmentation mask for the visual concept. Following the same philosophy, we design Hough Voting-based Semantic Prompt Learning (HoughSemPL) that learns distinct semantic prompts for different sub-regions of the visual concept, which capture complementary semantic clues and vote together to predict a accurate semantic label for the generated mask. Extensive experiments show that our proposed techniques achieve superior prompt learning performance over popular segmentation datasets. Codes will be released.", "keywords": [ "Prompt Learning", "Segmentation Anything Model", "Few shot learning" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "nxQ0Bjp8zD", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Lifeng Lai", "gender": "unknown", "institution": "University of California, Davis", "country": "US", "position": "Full Professor" }, { "name": "Yanhao Jin", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "PhD student" }, { "name": "Krishnakumar Balasubramanian", "gender": "Male", "institution": "University of California, Davis", "country": "", "position": "Associate Professor" } ], "n_formula": 113, "n_formula_1": 30, "n_ref_uni": 38, "n_ref": 86, "n_ref_all": 113, "n_fig": 4, "n_tab": 6, "L_tab": 6490, "n_element_tab": 405, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 2760, "n_element_tab_1": 63, "formula_len_all": 13987, "formula_len_all_1": 2804, "len_all": 204236, "len_all_1": 69227, "len_abs": 946, "len_title": 129, "len_sents": 53981, "len_sents_1": 26278, "n_sents": 488, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1003, "title": "Provable In-context Learning for Mixture of Linear Regressions using Transformers", "abs": "We theoretically investigate the in-context learning capabilities of transformers in the context of learning mixtures of linear regression models. For the case of two mixtures, we demonstrate the existence of transformers that can achieve an accuracy, relative to the oracle predictor, of order $\\mathcal{\\tilde{O}}((d/n)^{1/4})$ in the low signal-to-noise ratio (SNR) regime and $\\mathcal{\\tilde{O}}(\\sqrt{d/n})$ in the high SNR regime, where $n$ is the length of the prompt, and $d$ is the dimension of the problem. Additionally, we derive in-context excess risk bounds of order $\\mathcal{O}(L/\\sqrt{B})$, where $B$ denotes the number of (training) prompts, and $L$ represents the number of attention layers. The order of $L$ depends on whether the SNR is low or high. In the high SNR regime, we extend the results to $K$-component mixture models for finite $K$. Extensive simulations also highlight the advantages of transformers for this task, outperforming other baselines such as the EM algorithm.", "keywords": [ "Transformer", "Mixture of Linear Regression", "EM algorithm" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 3, 4 ] }, { "paper_id": "nx9Z5Kva96", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Boyu Wang", "gender": "Male", "institution": "University of Western Ontario", "country": "CA", "position": "Assistant Professor" }, { "name": "Charles Ling", "gender": "Male", "institution": "Western University", "country": "CA", "position": "Professor" }, { "name": "Gezheng Xu", "gender": "Female", "institution": "University of Western Ontario", "country": "CA", "position": "PhD student" }, { "name": "Grace Yi", "gender": "Female", "institution": "University of Western Ontario", "country": "CA", "position": "Full Professor" }, { "name": "Hui Guo", "gender": "Female", "institution": "University of Western Ontario", "country": "", "position": "PhD student" }, { "name": "Li Yi", "gender": "Male", "institution": "ByteDance Inc.", "country": "CA", "position": "Researcher" } ], "n_formula": 156, "n_formula_1": 17, "n_ref_uni": 24, "n_ref": 97, "n_ref_all": 177, "n_fig": 16, "n_tab": 11, "L_tab": 7502, "n_element_tab": 540, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 3005, "n_element_tab_1": 193, "formula_len_all": 18886, "formula_len_all_1": 1586, "len_all": 247203, "len_all_1": 77022, "len_abs": 2071, "len_title": 131, "len_sents": 73462, "len_sents_1": 30288, "n_sents": 660, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1159, "title": "Revisiting Source-Free Domain Adaptation: a New Perspective via Uncertainty Control", "abs": "Source-Free Domain Adaptation (SFDA) seeks to adapt a pre-trained source model to the target domain using only unlabeled target data, without access to the original source data. While current state-of-the-art (SOTA) methods rely on leveraging weak supervision from the source model to extract reliable information for self-supervised adaptation, they often overlook the uncertainty that arises during the transfer process. In this paper, we conduct a systematic and theoretical analysis of the uncertainty inherent in existing SFDA methods and demonstrate its impact on transfer performance through the lens of Distributionally Robust Optimization (DRO). Building upon the theoretical results, we propose a novel instance-dependent uncertainty control algorithm for SFDA. Our method is designed to quantify and exploit the uncertainty during the adaptation process, significantly improving the model performance. Extensive experiments on benchmark datasets and empirical analyses confirm the validity of our theoretical findings and the effectiveness of the proposed method. \nThis work offers new insights into understanding and advancing SFDA performance.", "keywords": [ "Source-Free Domain Adaptation", "Unsupervised Domain Adaptation" ], "rating_list": [ 8, 6, 5 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 2, 3, 2 ] }, { "paper_id": "nwjgeFGbAF", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Gurvinder Singh", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "", "position": "MS student" }, { "name": "Jakir Hossain", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Lukasz Ziarek", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Associate Professor" }, { "name": "Ahmet Sariyuce", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "", "position": "Associate Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 47, "n_ref_all": 58, "n_fig": 2, "n_tab": 9, "L_tab": 5842, "n_element_tab": 592, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1517, "n_element_tab_1": 92, "formula_len_all": 60, "formula_len_all_1": 0, "len_all": 121480, "len_all_1": 63963, "len_abs": 807, "len_title": 133, "len_sents": 35217, "len_sents_1": 30587, "n_sents": 236, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 812, "title": "Better Call Graphs: A New Dataset of Function Call Graphs for Malware Classification", "abs": "Malware classification by using function call graphs (FCG) is an important task in cybersecurity.\nOne big challenge in this direction is the lack of representative, large, and unique FCG datasets.\nExisting datasets typically contain obsolete Android application packages (APKs), largely consist of small graphs, and include many duplicate FCGs due to repackaging.\nThis results in misleading graph classification performance.\nIn this paper, we propose a new comprehensive dataset, Better Call Graphs (BCG), that contains large and unique FCGs from recent APKs, along with graph-level APK features, with benign and malware samples from different types and families.\nWe establish the necessity of BCG through the evaluation of several baseline approaches on existing datasets. \nBCG is available at https://iclr.me.", "keywords": [ "Malware classification", "FCG" ], "rating_list": [ 3, 5, 5, 8 ], "soundness_list": [ 3, 2, 2, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "nwZHFKrYTB", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Alexander Wettig", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Danqi Chen", "gender": "Female", "institution": "Department of Computer Science, Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Howard Yen", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Tianyu Gao", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 80, "n_ref": 157, "n_ref_all": 205, "n_fig": 4, "n_tab": 14, "L_tab": 3222, "n_element_tab": 349, "n_fig_1": 4, "n_tab_1": 11, "L_tab_1": 2758, "n_element_tab_1": 254, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 257971, "len_all_1": 64370, "len_abs": 7075, "len_title": 103, "len_sents": 48461, "len_sents_1": 27009, "n_sents": 378, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1379, "title": "How to Train Long-Context Language Models (Effectively)", "abs": "We study the problem of adapting a language model (LM) to make effective use of long-context information. We first establish a reliable evaluation protocol to guide model development---instead of perplexity, we use a broad set of long-context tasks, and we evaluate models after supervised fine-tuning (SFT) with instruction data as this better reveals long-context abilities. Supported by our robust evaluations, we run thorough experiments to decide the data mix for continued pre-training, the instruction tuning dataset, and other design choices such as position extrapolation. We find that (1) code repositories and books are excellent sources of long data, but it is crucial to combine them with high-quality short data; (2) training with a sequence length beyond the evaluation length boosts long-context performance; (3) for SFT, using only short instruction datasets yields strong performance on long-context tasks. Our final model, ProLong-8B, which is initialized from Llama-3 and trained on 40B tokens, demonstrates state-of-the-art long-context performance among similarly sized models at a length of 128K, outperforming Llama-3.1-8B on the majority of long-context tasks despite having seen 5% as many tokens during long-context training. Additionally, ProLong can effectively process up to 512K tokens, one of the longest context windows of publicly available LMs.", "keywords": [ "language models", "long-context language models", "continual pre-training", "long-context SFT" ], "rating_list": [ 6, 6, 6, 6, 5 ], "soundness_list": [ 1, 3, 2, 3, 3 ], "presentation_list": [ 4, 4, 4, 3, 3 ], "contribution_list": [ 2, 4, 3, 3, 2 ], "confidence_list": [ 4, 4, 5, 5, 5 ] }, { "paper_id": "nwETBpOPiC", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Adam Breitholtz", "gender": "Male", "institution": "Chalmers University of Technology", "country": "SE", "position": "PhD student" }, { "name": "Edvin Listo Zec", "gender": "Male", "institution": "RISE Research Institutes of Sweden", "country": "", "position": "Researcher" }, { "name": "Fredrik Daniel Johansson", "gender": "Male", "institution": "Chalmers University of Technology", "country": "SE", "position": "Associate Professor" } ], "n_formula": 28, "n_formula_1": 22, "n_ref_uni": 46, "n_ref": 82, "n_ref_all": 107, "n_fig": 10, "n_tab": 3, "L_tab": 949, "n_element_tab": 73, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 49, "n_element_tab_1": 1, "formula_len_all": 2158, "formula_len_all_1": 1875, "len_all": 158540, "len_all_1": 67428, "len_abs": 1458, "len_title": 0, "len_sents": 46007, "len_sents_1": 32858, "n_sents": 337, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1464, "title": "Overcoming label shift in targeted federated learning", "abs": "Federated learning enables multiple actors to collaboratively train models without sharing private data. This unlocks the potential for scaling machine learning to diverse applications. Existing algorithms for this task are well-justified when clients and the intended target domain share the same distribution of features and labels, but this assumption is often violated in real-world scenarios. One common violation is label shift, where the label distributions differ across clients or between clients and the target domain, which can significantly degrade model performance. To address this problem, we propose FedPALS, a novel model aggregation scheme that adapts to label shifts by leveraging knowledge of the target label distribution at the central server. Our approach ensures unbiased updates under stochastic gradient descent, ensuring robust generalization across clients with diverse, label-shifted data. Extensive experiments on image classification demonstrate that FedPALS consistently outperforms standard baselines by aligning model aggregation with the target domain. Our findings reveal that conventional federated learning methods suffer severely in cases of extreme client sparsity, highlighting the critical need for target-aware aggregation. FedPALS offers a principled and practical solution to mitigate label distribution mismatch, ensuring models trained in federated settings can generalize effectively to label-shifted target domains.", "keywords": [ "Federated learning", "Label shift", "distribution shift", "non-iid" ], "rating_list": [ 3, 6, 3, 5, 3 ], "soundness_list": [ 2, 3, 1, 2, 1 ], "presentation_list": [ 2, 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 1, 2, 3 ], "confidence_list": [ 3, 4, 4, 5, 4 ] }, { "paper_id": "nwDRD4AMoN", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Andreas Geiger", "gender": "Male", "institution": "University of Tuebingen", "country": "DE", "position": "Professor" }, { "name": "Max Welling", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "Sindy L\u00f6we", "gender": "Female", "institution": "University of Amsterdam", "country": "", "position": "PhD student" }, { "name": "Takeru Miyato", "gender": "Male", "institution": "Preferred Networks, Inc.", "country": "JP", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 14, "n_ref_uni": 80, "n_ref": 153, "n_ref_all": 196, "n_fig": 18, "n_tab": 10, "L_tab": 2155, "n_element_tab": 252, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 565, "n_element_tab_1": 29, "formula_len_all": 1704, "formula_len_all_1": 841, "len_all": 226295, "len_all_1": 61379, "len_abs": 1144, "len_title": 87, "len_sents": 52230, "len_sents_1": 27874, "n_sents": 437, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 39, "L_abs": 1152, "title": "Artificial Kuramoto Oscillatory Neurons", "abs": "It has long been known in both neuroscience and AI that ``binding'' between neurons leads to a form of competitive learning where representations are compressed in order to represent more abstract concepts in deeper layers of the network. More recently, it was also hypothesized that dynamic (spatiotemporal) representations play an important role in both neuroscience and AI. Building on these ideas, we introduce Artificial Kuramoto Oscillatory Neurons (*AKOrN*) as a dynamical alternative to threshold units, which can be combined with arbitrary connectivity designs such as fully connected, convolutional, or attentive mechanisms. Our generalized Kuramoto updates bind neurons together through their synchronization dynamics. We show that this idea provides performance improvements across a wide spectrum of tasks such as unsupervised object discovery, adversarial robustness, calibrated uncertainty quantification, and reasoning. We believe that these empirical results show the importance of rethinking our assumptions at the most basic neuronal level of neural representation, and in particular show the importance of dynamical representations.", "keywords": [ "Oscillatory neurons", "Feature binding", "Object-centric learning", "Reasoning", "Adversarial robustness" ], "rating_list": [ 10, 10, 8, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "nvCJqxJS2Y", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chen-Wei Chang", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "Cheng-De Fan", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "Jie-Ying Lee", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Undergrad student" }, { "name": "Jiun-Long Huang", "gender": "unknown", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" }, { "name": "Yi-Ruei Liu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Yu-Chee Tseng", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" }, { "name": "Yu Lun Liu", "gender": "unknown", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 19, "n_ref_uni": 37, "n_ref": 67, "n_ref_all": 87, "n_fig": 17, "n_tab": 6, "L_tab": 2758, "n_element_tab": 248, "n_fig_1": 10, "n_tab_1": 5, "L_tab_1": 2467, "n_element_tab_1": 139, "formula_len_all": 938, "formula_len_all_1": 880, "len_all": 127804, "len_all_1": 65022, "len_abs": 995, "len_title": 107, "len_sents": 30019, "len_sents_1": 22376, "n_sents": 244, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 997, "title": "SpectroMotion: Dynamic 3D Reconstruction of Specular Scenes", "abs": "We present SpectroMotion, a novel approach that combines 3D Gaussian Splatting (3DGS) with physically-based rendering (PBR) and deformation fields to reconstruct dynamic specular scenes. Previous methods extending 3DGS to model dynamic scenes have struggled to accurately represent specular surfaces. Our method addresses this limitation by introducing a residual correction technique for accurate surface normal computation during deformation, complemented by a deformable environment map that adapts to time-varying lighting conditions. We implement a coarse-to-fine training strategy that significantly enhances both scene geometry and specular color prediction. We demonstrate that our model outperforms prior methods for view synthesis of scenes containing dynamic specular objects and that it is the only existing 3DGS method capable of synthesizing photorealistic real-world dynamic specular scenes, outperforming state-of-the-art methods in rendering complex, dynamic, and specular scenes.", "keywords": [ "Specular objects", "Dynamic scenes", "3D Gaussian Splatting", "Novel view synthesis" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "nuX2yPejiL", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Dimitris Oikonomou", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "PhD student" }, { "name": "Nicolas Loizou", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 132, "n_formula_1": 26, "n_ref_uni": 60, "n_ref": 158, "n_ref_all": 220, "n_fig": 16, "n_tab": 2, "L_tab": 421, "n_element_tab": 82, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 16726, "formula_len_all_1": 2635, "len_all": 238089, "len_all_1": 69867, "len_abs": 1559, "len_title": 139, "len_sents": 64418, "len_sents_1": 30135, "n_sents": 622, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1560, "title": "Stochastic Polyak Step-sizes and Momentum: Convergence Guarantees and Practical Performance", "abs": "Stochastic gradient descent with momentum, also known as Stochastic Heavy Ball method (SHB), is one of the most popular algorithms for solving large-scale stochastic optimization problems in various machine learning tasks. In practical scenarios, tuning the step-size and momentum parameters of the method is a prohibitively expensive and time-consuming process. In this work, inspired by the recent advantages of stochastic Polyak step-size in the performance of stochastic gradient descent (SGD), we propose and explore new Polyak-type variants suitable for the update rule of the SHB method. In particular, using the Iterate Moving Average (IMA) viewpoint of SHB, we propose and analyze three novel step-size selections: MomSPSmax, MomDecSPS, and MomAdaSPS. For MomSPSmax, we provide convergence guarantees for SHB to a neighborhood of the solution for convex and smooth problems (without assuming interpolation). If interpolation is also satisfied, then using MomSPSmax, SHB converges to the true solution at a fast rate matching the deterministic HB. The other two variants, MomDecSPS and MomAdaSPS, are the first adaptive step-size for SHB that guarantee convergence to the exact minimizer - without a priori knowledge of the problem parameters and without assuming interpolation. Our convergence analysis of SHB is tight and obtains the convergence guarantees of stochastic Polyak step-size for SGD as a special case. We supplement our analysis with experiments validating our theory and demonstrating the effectiveness and robustness of our algorithms.", "keywords": [ "Stochastic Polyak Step-size", "SGD", "Stochastic Heavy Ball", "Convergence Analysis", "Convex Optimization", "Momentum" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "nu1H4MdxcB", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Himanshu Singh", "gender": "Male", "institution": "Los Alamos National Laboratory", "country": "US", "position": "Postdoc" }, { "name": "Nishant Panda", "gender": "Male", "institution": "Los Alamos National Laboratory", "country": "", "position": "Researcher" } ], "n_formula": 31, "n_formula_1": 26, "n_ref_uni": 34, "n_ref": 65, "n_ref_all": 79, "n_fig": 6, "n_tab": 3, "L_tab": 259, "n_element_tab": 18, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 366, "n_element_tab_1": 21, "formula_len_all": 2316, "formula_len_all_1": 2004, "len_all": 131550, "len_all_1": 66247, "len_abs": 1360, "len_title": 112, "len_sents": 26535, "len_sents_1": 24172, "n_sents": 236, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1372, "title": "Sparse spatio temporal reconstruction with Closable Kernel Space", "abs": "Quantifying spatio-temporal (ST) measures of dynamical systems is a crucial problem with wide ranging applications in climate modeling, epidemiology, physical processes to name a few. \nWe are interested in the same but motivated by a rather practical scenario where sparse information is collected non-uniformly. \nTo reconstruct the underlying dynamical system under such constraints, we propose a novel algorithm for learning the Koopman operator via a Reproducing Kernel Hilbert Space (RKHS) based on the Laplacian Kernel Extended Dynamic Mode Decomposition (Lap-KeDMD).\nWe further show that our kernel space resolves a fundamental issue that is required for a faithful reconstruction of the Koopman operator of the underlying ST data by proving its closability. \nWe demonstrate our method on standard benchmark cases --\nBurger's Equation, fluid flow across cylinder and Duffing Oscillator. \nWe then reconstruct the Koopman operator for a real ST Seattle traffic flow data that is collected non-uniformly. Necessary comparisons are made between the current state of the art kernel methods corresponding to Gaussian Radial Basis Function (GRBF) Kernel. \nSuch empirical comparisons leads us to conclude that Lap-KeDMD remarkably outperforms as compared to that of aforementioned counter-part thereby, making the Laplacian Kernel a robust choice for such ST quantification.", "keywords": [ "Kernel Learning", "Koopman Theory", "Spatio-Temporal Reconstruction", "Reproducing Kernel Hilbert space", "Dynamical Systems" ], "rating_list": [ 3, 1, 3 ], "soundness_list": [ 1, 1, 1 ], "presentation_list": [ 1, 1, 2 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "ntxoThl1Zp", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Alex Saad-Falcon", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Brighton Ancelin", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Justin Romberg", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Kason Ancelin", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 19, "n_ref_uni": 62, "n_ref": 90, "n_ref_all": 115, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6609, "formula_len_all_1": 1765, "len_all": 148472, "len_all_1": 62247, "len_abs": 1229, "len_title": 103, "len_sents": 38352, "len_sents_1": 28604, "n_sents": 292, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1234, "title": "Rapid Grassmannian Averaging with Chebyshev Polynomials", "abs": "We propose new algorithms to efficiently average a collection of points on a Grassmannian manifold in both the centralized and decentralized settings. Grassmannian points are used ubiquitously in machine learning, computer vision, and signal processing to represent data through (often low-dimensional) subspaces. While averaging these points is crucial to many tasks (especially in the decentralized setting), existing methods unfortunately remain computationally expensive due to the non-Euclidean geometry of the manifold. Our proposed algorithms, Rapid Grassmannian Averaging (RGrAv) and Decentralized Rapid Grassmannian Averaging (DRGrAv), overcome this challenge by leveraging the spectral structure of the problem to rapidly compute an average using only small matrix multiplications and QR factorizations. We provide a theoretical guarantee of optimality and present numerical experiments which demonstrate that our algorithms outperform state-of-the-art methods in providing high accuracy solutions in minimal time. Additional experiments showcase the versatility of our algorithms to tasks such as $K$-means clustering on video motion data, establishing RGrAv and DRGrAv as powerful tools for generic Grassmannian averaging.", "keywords": [ "Grassmannian", "Grassmann", "Manifold", "Subspace", "Average", "Averaging", "Mean", "Fast", "Rapid", "Chebyshev", "Polynomial", "Power", "Method", "Iteration", "Consensus", "Decentralized", "Distributed" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 4, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "nttFj0wKfD", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Jay Pujara", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "Associate Professor" }, { "name": "Kashif Munir", "gender": "Male", "institution": "Hippocraticai AI", "country": "US", "position": "Researcher" }, { "name": "Subhabrata Mukherjee", "gender": "unknown", "institution": "Hippocratic AI", "country": "", "position": "Chief Scientist" }, { "name": "Tanmay Laud", "gender": "Male", "institution": "Hippocratic AI", "country": "US", "position": "Principal Researcher" }, { "name": "Yifan Jiang", "gender": "Male", "institution": "Information Sciences Institute, University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Kriti Aggarwal", "gender": "Female", "institution": "HippocraticAI", "country": "", "position": "Principal Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 61, "n_ref": 348, "n_ref_all": 400, "n_fig": 6, "n_tab": 22, "L_tab": 9943, "n_element_tab": 1114, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 3859, "n_element_tab_1": 67, "formula_len_all": 9, "formula_len_all_1": 0, "len_all": 329912, "len_all_1": 69744, "len_abs": 5868, "len_title": 129, "len_sents": 108726, "len_sents_1": 31293, "n_sents": 892, "n_sents_1": 247, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 2098, "title": "RED QUEEN: SAFEGUARDING LARGE LANGUAGE MODELS AGAINST CONCEALED MULTI-TURN ATTACK", "abs": "The rapid progress of large language models (LLMs) has opened up new opportunities across various domains and applications; yet it also presents challenges related to potential misuse. To mitigate such risks, red teaming, a strategy where developers adopt the role of potential attackers has been employed to probe language models and preemptively guard against such harms. Jailbreak attacks are a commonly used red teaming strategy that uses crafted prompts to bypass safety guardrails. However, current jailbreak attack approaches are single-turn, with explicit malicious queries that do not fully capture the complexity of real-world interactions. In reality, users can engage in multi-turn interactions with LLM-based chat assistants, allowing them to conceal their true intentions in a more covert manner. Research on the Theory of Mind (ToM) reveals that LLMs struggle to infer latent intent, making it crucial to investigate how LLMs handle concealed malicious intent within multi-turn scenarios. To bridge this gap, we propose a new jailbreak approach, RED QUEEN ATTACK. This method constructs a multi-turn scenario, concealing the malicious intent under the guise of preventing harm. Next, we craft 40 scenarios that vary in turns and select 14 harmful categories to generate 56k multi-turn attack data points. We conduct comprehensive experiments on the RED QUEEN ATTACK with four representative LLM families of different sizes. Our experiments reveal that all LLMs are vulnerable to RED QUEEN ATTACK, reaching 87.6% attack success rate on GPT-4o and 77.1% on Llama3-70B. Further analysis\nreveals that larger models are more susceptible to the RED QUEEN ATTACK, with multi-turn structures and concealment strategies contributing to its success. To prioritize safety, we introduce a straightforward mitigation strategy called RED QUEEN GUARD, which aligns LLMs to effectively counter adversarial attacks. This approach reduces the attack success rate to below 1% while maintaining the model\u2019s performance across standard benchmarks. We release our code and data to support future research.", "keywords": [ "Jailbreaking", "Large Language Models", "Safety Alignment" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "ntPHPguG1o", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Anningzhe Gao", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Haifan Gong", "gender": "Male", "institution": "Chinese University of Hong Kong, Shen Zhen", "country": "CN", "position": "PhD student" }, { "name": "Haofeng Li", "gender": "unknown", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Researcher" }, { "name": "Ruifei Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Xiang Wan", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuhao Du", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Zhuo Li", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Xy Zhang", "gender": "Male", "institution": "Xiaoguang Han", "country": "CN", "position": "Research Assistant" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 46, "n_ref": 116, "n_ref_all": 130, "n_fig": 7, "n_tab": 1, "L_tab": 622, "n_element_tab": 32, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 428, "n_element_tab_1": 5, "formula_len_all": 105, "formula_len_all_1": 79, "len_all": 125590, "len_all_1": 60007, "len_abs": 1550, "len_title": 119, "len_sents": 36008, "len_sents_1": 30608, "n_sents": 235, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1560, "title": "Supervised Pre-training for Unsupervised Product-Patent Image Retrieval", "abs": "Detecting infringing products is essential for protecting intellectual property rights and is often implemented as a product-patent retrieval task. Manual infringement detection is extremely time-consuming, and artificial intelligence plays an increasingly important role. However, most existing methods rely on natural language-based retrieval due to the domain discrepancies between patent images and product images. Due to the lack of sufficient annotated data, this work aims to address the aforementioned issues in an unsupervised setting by answering the following two questions: 1) How can we align the domain gap between patent images and product images using existing technologies? 2) How can we build a powerful backbone to jointly extract the features of patent and product images? Initially, we construct a dataset for patent-product image retrieval, which includes product-patent pairs and unlabeled data. To address the first question, we systematically evaluate three unsupervised approaches to mitigate the domain gap between patent and product images. The results demonstrate that jointly mapping patent and product images to a new feature space is effective. To answer the second question, we propose a novel supervised pre-training paradigm to achieve domain-aligned feature extraction for product and patent edge images. Extensive experiments using various backbones and training pipelines demonstrate the superiority of our supervised pre-training method. The dataset and code of this paper will be made publicly available upon acceptance.", "keywords": [ "image retrieval", "supervised pre-training", "domain gap" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "nt8gBX58Kh", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 9, "author_info_list": [ { "name": "Chenyu Zhou", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "PhD student" }, { "name": "Defu Cao", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Heng Ping", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Nikos Kanakaris", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Postdoc" }, { "name": "Paul Bogdan", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Jack Munushian Early Career Chair associate professor" }, { "name": "Shixuan Li", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Xiongye Xiao", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Yaxing Li", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "MS student" }, { "name": "Yizhuo Zhou", "gender": "unknown", "institution": "University of California, Riverside", "country": "US", "position": "PhD student" } ], "n_formula": 74, "n_formula_1": 18, "n_ref_uni": 56, "n_ref": 90, "n_ref_all": 126, "n_fig": 22, "n_tab": 8, "L_tab": 2838, "n_element_tab": 264, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 491, "n_element_tab_1": 18, "formula_len_all": 3328, "formula_len_all_1": 546, "len_all": 270312, "len_all_1": 67302, "len_abs": 3532, "len_title": 130, "len_sents": 87900, "len_sents_1": 31659, "n_sents": 655, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1396, "title": "Neuron-based Multifractal Analysis of Neuron Interaction Dynamics in Large Models", "abs": "In recent years, there has been increasing attention on the capabilities of large models, particularly in handling complex tasks that small-scale models are unable to perform. Notably, large language models (LLMs) have demonstrated ``intelligent'' abilities such as complex reasoning and abstract language comprehension, reflecting cognitive-like behaviors. However, current research on emergent abilities in large models predominantly focuses on the relationship between model performance and size, leaving a significant gap in the systematic quantitative analysis of the internal structures and mechanisms driving these emergent abilities. Drawing inspiration from neuroscience research on brain network structure and self-organization, we propose (i) a general network representation of large models, (ii) a new analytical framework \u2014 *Neuron-based Multifractal Analysis (NeuroMFA)* - for structural analysis, and (iii) a novel structure-based metric as a proxy for emergent abilities of large models. By linking structural features to the capabilities of large models, NeuroMFA provides a quantitative framework for analyzing emergent phenomena in large models. Our experiments show that the proposed method yields a comprehensive measure of network's evolving heterogeneity and organization, offering theoretical foundations and a new perspective for investigating emergence in large models.", "keywords": [ "Emergent ability; Large language model" ], "rating_list": [ 5, 8, 6 ], "soundness_list": [ 2, 4, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 4, 3 ], "confidence_list": [ 3, 5, 2 ] }, { "paper_id": "nsozLtutE6", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Aitian Ma", "gender": "Male", "institution": "Florida International University", "country": "", "position": "PhD student" }, { "name": "Dongsheng Luo", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "Assistant Professor" }, { "name": "Mo Sha", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 24, "n_ref": 82, "n_ref_all": 102, "n_fig": 7, "n_tab": 13, "L_tab": 9532, "n_element_tab": 811, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 4101, "n_element_tab_1": 379, "formula_len_all": 405, "formula_len_all_1": 346, "len_all": 191478, "len_all_1": 68218, "len_abs": 1563, "len_title": 128, "len_sents": 67367, "len_sents_1": 30673, "n_sents": 411, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1328, "title": "MMFNet: Multi-Scale Frequency Masking Neural Network for Multivariate Time Series Forecasting", "abs": "Long-term Time Series Forecasting (LTSF) is critical for numerous real-world applications, such as electricity consumption planning, financial forecasting, and disease propagation analysis. LTSF requires capturing long-range dependencies between inputs and outputs, which poses significant challenges due to complex temporal dynamics and high computational demands. While linear models reduce model complexity by employing frequency domain decomposition, current approaches often assume stationarity and filter out high-frequency components that may contain crucial short-term fluctuations. In this paper, we introduce MMFNet, a novel model designed to enhance long-term multivariate forecasting by leveraging a multi-scale masked frequency decomposition approach. MMFNet captures fine, intermediate, and coarse-grained temporal patterns by converting time series into frequency segments at varying scales while employing a learnable mask to filter out irrelevant components adaptively.\nExtensive experimentation with benchmark datasets shows that MMFNet not only addresses the limitations of the existing methods but also consistently achieves good performance. Specifically, MMFNet achieves up to 6.0\\% reductions in the Mean Squared Error (MSE) compared to state-of-the-art models designed for multivariate forecasting tasks.", "keywords": [ "Time series forecasting", "Multi-scale frequency domain decomposition" ], "rating_list": [ 3, 6, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 1, 2 ], "confidence_list": [ 4, 5, 4, 4, 3 ] }, { "paper_id": "nsFucJqKmR", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Artem Ploujnikov", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "Cem Subakan", "gender": "unknown", "institution": "Universite Laval, Universit\u00e9 Laval", "country": "CA", "position": "Assistant Professor" }, { "name": "Jarod Duret", "gender": "Male", "institution": "Universit\u00e9 d'Avignon", "country": "", "position": "PhD student" }, { "name": "Luca Della Libera", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "PhD student" }, { "name": "Mirco Ravanelli", "gender": "Male", "institution": "Concordia University", "country": "CA", "position": "Assistant Professor" }, { "name": "Pooneh Mousavi", "gender": "Female", "institution": "Concordia University", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 46, "n_ref": 91, "n_ref_all": 111, "n_fig": 2, "n_tab": 13, "L_tab": 7203, "n_element_tab": 1404, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2990, "n_element_tab_1": 529, "formula_len_all": 118, "formula_len_all_1": 118, "len_all": 191450, "len_all_1": 73240, "len_abs": 1307, "len_title": 89, "len_sents": 38316, "len_sents_1": 30901, "n_sents": 305, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1310, "title": "DASB-Discrete Audio and Speech Benchmark", "abs": "Discrete audio tokens have recently gained considerable attention for their potential to connect audio and language processing, enabling the creation of modern multimodal large language models. Ideal audio tokens must effectively preserve phonetic and semantic content along with paralinguistic information, speaker identity, and other details. While several types of audio tokens have been recently proposed, identifying the optimal tokenizer for various tasks is challenging due to the inconsistent evaluation settings in existing studies. To address this gap, we release the Discrete Audio and Speech Benchmark (DASB), a comprehensive leaderboard for benchmarking discrete audio tokens across a wide range of discriminative tasks, including speech recognition, speaker identification and verification, emotion recognition, keyword spotting, intent classification, event sound detection, and music genre classification as well as generative tasks such as speech enhancement, separation, and text-to-speech. Our results show that, on average, semantic tokens outperform compression tokens across most discriminative and generative tasks. However, the performance gap between semantic tokens and standard continuous representations remains substantial, highlighting the need for further research in this field.", "keywords": [ "discrete audio tokens", "speech processing" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "ns0KIpfQVy", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Alexander Kostin", "gender": "Male", "institution": "Sberbank", "country": "", "position": "Researcher" }, { "name": "Andrey Savchenko", "gender": "Male", "institution": "Sber AI Lab", "country": "RU", "position": "Principal Researcher" }, { "name": "Gleb Gusev", "gender": "Male", "institution": "ARTIFICIAL INTELLIGENCE RESEARCH INSTITUTE (AIRI)", "country": "RU", "position": "Principal Researcher" }, { "name": "Ivan A Kireev", "gender": "Male", "institution": "Sber AI Lab", "country": "RU", "position": "Principal Researcher" }, { "name": "Mollaev Dzhambulat", "gender": "Male", "institution": "Lomonosov Moscow State University", "country": "", "position": "Researcher" }, { "name": "Postnova Maria", "gender": "Female", "institution": "NATIONAL RESEARCH NUCLEAR UNIVERSITY MEPHI Moscow Engineering Physics Institute", "country": "", "position": "Undergrad student" }, { "name": "Ivan Alexandrovich Karpukhin", "gender": "Male", "institution": "Sberbank", "country": "RU", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 44, "n_ref_all": 87, "n_fig": 8, "n_tab": 11, "L_tab": 10146, "n_element_tab": 408, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1658, "n_element_tab_1": 122, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 150133, "len_all_1": 69154, "len_abs": 1158, "len_title": 126, "len_sents": 44685, "len_sents_1": 32427, "n_sents": 295, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1172, "title": "Multimodal Banking Dataset: Understanding Client Needs through Event Sequences", "abs": "Financial organizations collect a huge amount of data about clients that typi-\ncally has a temporal (sequential) structure and is collected from multiple sources\n(modalities). However, despite the urgent practical need, developing deep learn-\ning techniques suitable to handle such data is limited by the absence of large open-\nsource multi-source real-world datasets of event sequences. To fill this gap mainly\ncaused by security reasons, we present the industrial-scale publicly available mul-\ntimodal banking dataset, MBD, that contains more than 2M corporate clients with\nseveral data sources: 950M bank transactions, 1B geo position events, 5M em-\nbeddings of dialogues with technical support and monthly aggregated purchases\nof four bank\u2019s products. All entries are properly anonymized from real proprietary\nbank data. Moreover, we introduce a novel multimodal benchmark incorporating\nour MBD and two open-source financial datasets. We provide numerical results\ndemonstrating the superiority of fusion baselines over single-modal techniques\nfor each task. Moreover, our anonymization techniques still save all significant\ninformation for introduced downstream tasks.", "keywords": [ "multimodal", "multi-temporal", "event sequence" ], "rating_list": [ 5, 3, 8, 6 ], "soundness_list": [ 2, 1, 4, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 2, 1, 4, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "nrvoWOWcyg", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Jiawei Zhou", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Assistant Professor" }, { "name": "Yanhong Li", "gender": "Female", "institution": "University of Chicago", "country": "US", "position": "Undergrad student" }, { "name": "Kare Livescu", "gender": "Female", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Professor" } ], "n_formula": 25, "n_formula_1": 9, "n_ref_uni": 38, "n_ref": 88, "n_ref_all": 150, "n_fig": 14, "n_tab": 21, "L_tab": 9872, "n_element_tab": 930, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1878, "n_element_tab_1": 95, "formula_len_all": 3014, "formula_len_all_1": 762, "len_all": 257367, "len_all_1": 72371, "len_abs": 1000, "len_title": 81, "len_sents": 71193, "len_sents_1": 31884, "n_sents": 559, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 33, "L_abs": 1005, "title": "Chunk-Distilled Language Modeling", "abs": "We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to text generation that addresses two challenges in current large language models (LLMs): the inefficiency of token-level generation, and the difficulty of adapting to new data and knowledge. Our method combines deep network-based LLMs with a straightforward retrieval module, which allows the generation of multi-token text chunks at a single decoding step. Our retrieval framework enables flexible construction of model- or domain-specific datastores, either leveraging the internal knowledge of existing models, or incorporating expert insights from human-annotated corpora. This adaptability allows for enhanced control over the language model's distribution without necessitating additional training. We present the CD-LM formulation along with performance metrics demonstrating its ability to improve language model performance and efficiency across a diverse set of downstream tasks. Code and data will be made publicly available.", "keywords": [ "language modeling", "text generation", "retrieval-augmented generation", "domain adaptation", "inference algorithms", "efficient generation" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 2, 3, 4, 4 ], "contribution_list": [ 2, 3, 4, 3 ], "confidence_list": [ 3, 5, 4, 5 ] }, { "paper_id": "nrRkAAAufl", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Shengao Wang", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "PhD student" }, { "name": "Weichao Zhou", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Wenchao Li", "gender": "unknown", "institution": "Boston University", "country": "US", "position": "Associate Professor" }, { "name": "Zijian Guo", "gender": "Male", "institution": "Boston University, Boston University", "country": "", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 13, "n_ref_uni": 47, "n_ref": 108, "n_ref_all": 158, "n_fig": 9, "n_tab": 3, "L_tab": 3161, "n_element_tab": 176, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2224, "formula_len_all_1": 1027, "len_all": 165369, "len_all_1": 68834, "len_abs": 1060, "len_title": 123, "len_sents": 50234, "len_sents_1": 34079, "n_sents": 402, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1074, "title": "Constraint-Conditioned Actor-Critic for Offline Safe Reinforcement Learning", "abs": "Offline safe reinforcement learning (OSRL) aims to learn policies with high rewards while satisfying safety constraints solely from data collected offline. However, the learned policies often struggle to handle states and actions that are not present or out-of-distribution (OOD) from the offline dataset, which can result in violation of the safety constraints or overly conservative behaviors during their online deployment. Moreover, many existing methods are unable to learn policies that can adapt to varying constraint thresholds. To address these challenges, we propose constraint-conditioned actor-critic (CCAC), a novel OSRL method that models the relationship between state-action distributions and safety constraints, and leverages this relationship to regularize critics and policy learning. CCAC learns policies that can effectively handle OOD data and adapt to varying constraint thresholds. Empirical evaluations on the $\\texttt{DSRL}$ benchmarks show that CCAC significantly outperforms existing methods for learning adaptive, safe, and high-reward policies.", "keywords": [ "Offline Safe Reinforcement Learning", "Constraint-conditioned Actor-Critic", "Data Generation", "Out-of-distribution Detection", "Zero-shot Adaptation" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "nqGqIzDCRY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Gaolei Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Jianhua Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Quan Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xi Lin", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuxin Qi", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 18, "n_ref_uni": 27, "n_ref": 58, "n_ref_all": 72, "n_fig": 2, "n_tab": 7, "L_tab": 3366, "n_element_tab": 88, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 923, "n_element_tab_1": 66, "formula_len_all": 1824, "formula_len_all_1": 961, "len_all": 130226, "len_all_1": 67766, "len_abs": 1389, "len_title": 133, "len_sents": 39021, "len_sents_1": 29940, "n_sents": 315, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1414, "title": "Individualized Private Graph Neural Network via Node Influence-based Noise Adaptation", "abs": "Graph Neural Networks (GNNs) with Differential Privacy (DP) guarantees have been proposed to preserve privacy when nodes contain sensitive information that needs to be kept private but is critical for training. Existing methods deploy a fixed uniform noise generation mechanism that lacks the flexibility to adjust between nodes, leading to increasing the risk of graph information leakage and decreasing the model's overall performance. To address the above challenges, we propose NIP-GNN, a Node-level Individual Private GNN with DP guarantee based on the adaptive perturbation over sensitive components to safeguard node information. First, we propose a Topology-based Node Influence Estimation (TNIE) method to infer unknown node influence with neighborhood and centrality awareness. \n Second, given the obtained node influence rank, an adaptive private aggregation method is proposed to perturb neighborhood embeddings directed by node-wise influence. \n Third, we propose to privately train the graph learning algorithm over perturbed aggregations in adaptive residual connection mode over multi-layer convolution for node-wise tasks. Theoretically, analysis ensures that NIP-GNN satisfies DP guarantee. Empirical experiments over real-world graph datasets show that NIP-GNN \n presents a better resistance over node inference attacks and \n achieves a better trade-off between privacy and accuracy.", "keywords": [ "Graph Neural Network", "Differential Privacy" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 1, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "nplYdpc1Pm", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Anshuman Sinha", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Aubin REY", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Camille Migozzi", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "MS student" }, { "name": "Chao Zhang", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" } ], "n_formula": 41, "n_formula_1": 16, "n_ref_uni": 64, "n_ref": 118, "n_ref_all": 145, "n_fig": 14, "n_tab": 10, "L_tab": 1400, "n_element_tab": 112, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 50, "n_element_tab_1": 2, "formula_len_all": 3107, "formula_len_all_1": 1202, "len_all": 216987, "len_all_1": 61481, "len_abs": 1735, "len_title": 165, "len_sents": 60897, "len_sents_1": 28704, "n_sents": 409, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1767, "title": "Enhancing Audio--Language Models through Self--Supervised Post--Training with Text--Audio Pairs", "abs": "Research on multi-modal contrastive learning strategies for audio and text has rapidly gained interest. Contrastively trained Audio-Language Models (ALMs), such as CLAP, which establish a unified representation across audio and language modalities, have enhanced the efficacy in various subsequent tasks by providing good text aligned audio encoders and vice versa. These improvements are evident in areas like zero-shot audio classification and audio retrieval, among others. However, the ability of these models to understand natural language and temporal relations is still a largely unexplored and open field for research. In this paper, we propose to equip the multi-modal ALMs with temporal understanding without loosing their inherent prior capabilities of audio-language tasks with a temporal instillation method $\\textbf{TeminAL}$. We implement a two-stage training scheme TeminAL A \\& B, where the model first learns to differentiate between multiple sounds in TeminAL A, followed by a phase that instills a sense of time, thereby enhancing its temporal understanding in TeminAL B. This approach results in an average performance gain of $5.28$\\% in temporal understanding on the benchmark ESC-50 dataset, while the model remains competitive in zero-shot retrieval and classification tasks on the AudioCap/Clotho datasets. We also note the lack of proper evaluation techniques for contrastive ALMs and propose a strategy for evaluating ALMs in zero-shot settings. The general-purpose Zero-Shot Temporal Evaluation $\\textbf{(ZSTE)}$ strategy , is used to evaluate various prior models. ZSTE demonstrates a general strategy to evaluate all ZS contrastive models. The model trained with TeminAL successfully outperforms current models on most downstream tasks.", "keywords": [ "Self-supervised learning", "Contrastive learning", "Post-training", "Audio", "Zero-shot evaluation", "Audio-retrieval." ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "nphsoKxlFs", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Abdul-Kazeem Shamba", "gender": "Male", "institution": "Norwegian University of Science and Technology", "country": "NO", "position": "PhD student" }, { "name": "Gavin Taylor", "gender": "Male", "institution": "US Naval Academy", "country": "US", "position": "Full Professor" }, { "name": "Kerstin Bach", "gender": "Female", "institution": "Norwegian University of Science and Technology", "country": "NO", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 38, "n_ref": 84, "n_ref_all": 105, "n_fig": 7, "n_tab": 4, "L_tab": 1762, "n_element_tab": 204, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 130, "n_element_tab_1": 20, "formula_len_all": 754, "formula_len_all_1": 520, "len_all": 121210, "len_all_1": 56134, "len_abs": 1107, "len_title": 107, "len_sents": 39905, "len_sents_1": 27659, "n_sents": 289, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1113, "title": "Dynamic Contrastive Learning for Time Series Representation", "abs": "Understanding events in time series is an important task in a variety of contexts. However, human analysis and labeling are expensive and time-consuming. Therefore, it is advantageous to learn embeddings for moments in time series in an unsupervised way, which allows for good performance in classification or detection tasks after later minimal human labeling. In this paper, we propose dynamic contrastive learning (DynaCL), an unsupervised representation learning framework for time series that uses temporal adjacent steps to define positive pairs. DynaCL adopts N-pair loss to dynamically treat all samples in a batch as positive or negative pairs, enabling efficient training and addressing the challenges of complicated sampling of positives. We demonstrate that DynaCL embeds instances from time series into well-defined, semantically meaningful clusters, which allows superior performance on downstream tasks on a variety of public time series datasets. Our findings also reveal that high scores on unsupervised clustering metrics do not guarantee that the representations are useful in downstream tasks.", "keywords": [ "contrastive learning", "self-supervised learning", "time series analysis", "representation learning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "npBrvlYftk", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Hanbin Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Juncheng Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Roger Zimmermann", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Siliang Tang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Wei Ji", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Yifang Yin", "gender": "Female", "institution": "I2R, A*STAR", "country": "SG", "position": "Researcher" }, { "name": "YOU QIN", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 26, "n_ref": 54, "n_ref_all": 65, "n_fig": 3, "n_tab": 4, "L_tab": 1794, "n_element_tab": 296, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 3489, "n_element_tab_1": 125, "formula_len_all": 830, "formula_len_all_1": 627, "len_all": 105683, "len_all_1": 61898, "len_abs": 2505, "len_title": 117, "len_sents": 34500, "len_sents_1": 27002, "n_sents": 228, "n_sents_1": 170, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 2514, "title": "Robust Video Moment Retrieval with Introspective Knowledge Distillation", "abs": "With the huge requirement of video content understanding and editing, Video moment retrieval (VMR) is becoming more and more critical, necessitating models that are adept at correlating video contents with textual queries. The effectiveness of prevailing VMR models, however, is often compromised by their reliance on training data biases, which significantly hampers their generalization capabilities when faced with out-of-distribution (OOD) content. This challenge underscores the need for innovative approaches that can adeptly navigate the intricate balance between leveraging in-distribution (ID) data for learning and maintaining robustness against OOD variations. Addressing this critical need, we introduce Reflective Knowledge Distillation (RefKD), a novel and comprehensive training methodology that integrates the dual processes of Introspective Learning and Extrospective Adjustment. This methodology is designed to refine the model's ability to internalize and apply learned correlations in a manner that is both contextually relevant and resilient to bias-induced distortions. By employing a dual-teacher framework, RefKD encapsulates and contrasts the distinct bias perspectives prevalent in VMR datasets, facilitating a dynamic and reflective learning dialogue with the student model. This interaction is meticulously structured to encourage the student model to engage in a deeper introspection of learned biases and to adaptively recalibrate its learning focus in response to evolving content landscapes. Through this reflective learning process, the model develops a more nuanced and comprehensive understanding of content-query correlations, significantly enhancing its performance across both ID and OOD scenarios. Our extensive evaluations, conducted across several standard VMR benchmarks, demonstrate the unparalleled efficacy of RefKD. The methodology not only aligns with the OOD performance benchmarks set by existing debiasing methods but also, in many instances, significantly surpasses their ID performance metrics. By effectively bridging the gap between ID and OOD learning, RefKD sets a new standard for building VMR systems that are not only more adept at understanding and interpreting video content in a variety of contexts but also more equitable and reliable across diverse operational scenarios. This work not only contributes to the advancement of VMR technology but also paves the way for future research in the domain of bias-aware and robust multimedia content analysis.", "keywords": [ "video moment retrieval", "diverse query types", "model versatility" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "npBAHV5BJI", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Harry Shomer", "gender": "unknown", "institution": "Michigan State University", "country": "", "position": "PhD student" }, { "name": "Jay Revolinsky", "gender": "unknown", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Jiliang Tang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 2, "n_ref_uni": 29, "n_ref": 86, "n_ref_all": 132, "n_fig": 10, "n_tab": 20, "L_tab": 7071, "n_element_tab": 1055, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 2492, "n_element_tab_1": 108, "formula_len_all": 396, "formula_len_all_1": 133, "len_all": 188828, "len_all_1": 69841, "len_abs": 2024, "len_title": 122, "len_sents": 56867, "len_sents_1": 31744, "n_sents": 509, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1242, "title": "Towards Better Benchmark Datasets for Inductive Knowledge Graph Completion", "abs": "Knowledge Graph Completion (KGC) attempts to predict missing facts in a Knowledge Graph (KG). Recently, there's been an increased focus on designing KGC methods that can excel in the *inductive setting*, where a portion or all of the entities and relations seen in inference are unobserved during training. Numerous benchmark datasets have been proposed for inductive KGC, all of which are subsets of existing KGs used for transductive KGC. However, we find that the current procedure for constructing inductive KGC datasets inadvertently creates a shortcut that can be exploited even while disregarding the relational information. Specifically, we observe that the Personalized PageRank (PPR) score can achieve strong or near SOTA performance on most inductive datasets. In this paper, we study the root cause of this problem. Using these insights, we propose an alternative strategy for constructing inductive KGC datasets that helps mitigate the PPR shortcut. We then benchmark multiple popular methods using the newly constructed datasets and analyze their performance. The new benchmark datasets help promote a better understanding of the capabilities and challenges of inductive KGC by removing any shortcuts that obfuscate performance.", "keywords": [ "Knowledge graphs", "graphs", "link prediction" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "noidywkBba", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Dadong Wang", "gender": "Male", "institution": "CSIRO", "country": "AU", "position": "Principal Researcher" }, { "name": "Jialiang Shen", "gender": "Male", "institution": "Sensetime", "country": "CN", "position": "Researcher" }, { "name": "Jiyang Zheng", "gender": "Non-Binary", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Min Wang", "gender": "Male", "institution": "Sensetime", "country": "CN", "position": "Researcher" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Yang Yang", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "", "position": "Associate Professor" }, { "name": "Yu Yao", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Lecturer" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 25, "n_ref": 66, "n_ref_all": 96, "n_fig": 13, "n_tab": 7, "L_tab": 1336, "n_element_tab": 106, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 157, "n_element_tab_1": 30, "formula_len_all": 313, "formula_len_all_1": 164, "len_all": 131509, "len_all_1": 63422, "len_abs": 1281, "len_title": 154, "len_sents": 42360, "len_sents_1": 31552, "n_sents": 296, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1286, "title": "Chain-of-Focus Prompting: Leveraging Sequential Visual Cues to Prompt Large Autoregressive Vision Models", "abs": "In-context learning (ICL) has revolutionized natural language processing by enabling models to adapt to diverse tasks with only a few illustrative examples. However, the exploration of ICL within the field of computer vision remains limited. Inspired by Chain-of-Thought (CoT) prompting in the language domain, we propose Chain-of-Focus (CoF) Prompting, which enhances vision models by enabling step-by-step visual comprehension. CoF Prompting addresses the challenges of absent logical structure in visual data by generating intermediate reasoning steps through visual saliency. Moreover, it provides a solution for creating tailored prompts from visual inputs by selecting contextually informative prompts based on query similarity and target richness. The significance of CoF prompting is demonstrated by the recent introduction of Large Autoregressive Vision Models (LAVMs), which predict downstream targets via in-context learning with pure visual inputs. By integrating intermediate reasoning steps into visual prompts and effectively selecting the informative ones, the LAVMs are capable of generating significantly better inferences. Extensive experiments on downstream visual understanding tasks validate the effectiveness of our proposed method for visual in-context learning.", "keywords": [ "Autoregressive Large Vision Models", "Visual In-context Learning", "Prompt learning" ], "rating_list": [ 6, 8, 8, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "nobDw4d1k7", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Joni Pajarinen", "gender": "unknown", "institution": "Aalto University", "country": "FI", "position": "Associate Professor" }, { "name": "Juho Kannala", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Associate Professor" }, { "name": "Rongzhen Zhao", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "PhD student" }, { "name": "Huiling Wang", "gender": "unknown", "institution": "Aalto University", "country": "FI", "position": "Postdoc" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 35, "n_ref": 89, "n_ref_all": 103, "n_fig": 3, "n_tab": 7, "L_tab": 2261, "n_element_tab": 156, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 3722, "n_element_tab_1": 180, "formula_len_all": 630, "formula_len_all_1": 353, "len_all": 138882, "len_all_1": 57685, "len_abs": 1154, "len_title": 92, "len_sents": 41515, "len_sents_1": 21992, "n_sents": 349, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 44, "L_abs": 1181, "title": "Multi-Scale Fusion for Object Representation", "abs": "Representing images or videos as object-level feature vectors, rather than pixel-level feature maps, facilitates advanced visual tasks.\nObject-Centric Learning (OCL) primarily achieves this by reconstructing the input under the guidance of Variational Autoencoder (VAE) intermediate representation to drive so-called slots to aggregate as much object information as possible.\nHowever, existing VAE guidance does not explicitly address that objects can vary in pixel sizes while models typically excel at specific pattern scales.\nWe propose Multi-Scale Fusion (MSF) to enhance VAE guidance for OCL training. \nTo ensure objects of all sizes fall within VAE's comfort zone, we adopt the image pyramid, which produces intermediate representations at multiple scales;\nTo foster scale-invariance/variance in object super-pixels, we devise inter/intra-scale fusion, which augments low-quality object super-pixels of one scale with corresponding high-quality super-pixels from another scale.\nOn standard OCL benchmarks, our technique improves mainstream methods, including state-of-the-art diffusion-based ones.\nThe source code is available on https://github.com/Genera1Z/MultiScaleFusion.", "keywords": [ "Object-Centric Learning (OCL)", "Variational Autoencoder (VAE)", "Multi-Scale", "Unsupervised Object Segmentation" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "noUF58SMra", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Bozhen Hu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Cheng Tan", "gender": "Male", "institution": "Zhejiang University & Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Jun Xia", "gender": "Male", "institution": "Westlake University, China", "country": "CN", "position": "PhD student" }, { "name": "Lirong Wu", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Siyuan Li", "gender": "Male", "institution": "Westlake University & Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yufei Huang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhangyang Gao", "gender": "Male", "institution": "Westlake University, China", "country": "CN", "position": "PhD student" }, { "name": "Zhenxiao Cao", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Stan Z. Li", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Chair Professor" } ], "n_formula": 23, "n_formula_1": 12, "n_ref_uni": 67, "n_ref": 105, "n_ref_all": 133, "n_fig": 14, "n_tab": 10, "L_tab": 1574, "n_element_tab": 239, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 317, "n_element_tab_1": 48, "formula_len_all": 1565, "formula_len_all_1": 669, "len_all": 224017, "len_all_1": 52874, "len_abs": 1658, "len_title": 139, "len_sents": 60815, "len_sents_1": 25693, "n_sents": 398, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1666, "title": "MeToken: Uniform Micro-environment Token Boosts Post-Translational Modification Prediction", "abs": "Post-translational modifications (PTMs) profoundly expand the complexity and functionality of the proteome, regulating protein attributes and interactions that are crucial for biological processes. Accurately predicting PTM sites and their specific types is therefore essential for elucidating protein function and understanding disease mechanisms. Existing computational approaches predominantly focus on protein sequences to predict PTM sites, driven by the recognition of sequence-dependent motifs. However, these approaches often overlook protein structural contexts. In this work, we first compile a large-scale sequence-structure PTM dataset, which serves as the foundation for fair comparison. We introduce the MeToken model, which tokenizes the micro-environment of each amino acid, integrating both sequence and structural information into unified discrete tokens. This model not only captures the typical sequence motifs associated with PTMs but also leverages the spatial arrangements dictated by protein tertiary structures, thus providing a holistic view of the factors influencing PTM sites. Designed to address the long-tail distribution of PTM types, MeToken employs uniform sub-codebooks that ensure even the rarest PTMs are adequately represented and distinguished. We validate the effectiveness and generalizability of MeToken across multiple datasets, demonstrating its superior performance in accurately identifying PTM types. The results underscore the importance of incorporating structural data and highlight MeToken's potential in facilitating accurate and comprehensive PTM predictions, which could significantly impact proteomics research.", "keywords": [ "Biomolecular learning", "Protein sequence" ], "rating_list": [ 5, 6, 5, 5, 8 ], "soundness_list": [ 3, 2, 2, 3, 4 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3, 4 ], "confidence_list": [ 5, 5, 4, 3, 4 ] }, { "paper_id": "no8ysO3xde", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Aditya Parashar", "gender": "Male", "institution": "Oracle", "country": "US", "position": "Intern" }, { "name": "Ari Kobren", "gender": "Male", "institution": "Oracle Labs", "country": "", "position": "Research Scientist" }, { "name": "Natalia Kosilova", "gender": "Female", "institution": "Oracle", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 17, "n_ref": 26, "n_ref_all": 33, "n_fig": 2, "n_tab": 5, "L_tab": 1610, "n_element_tab": 122, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1326, "n_element_tab_1": 87, "formula_len_all": 237, "formula_len_all_1": 0, "len_all": 104743, "len_all_1": 62714, "len_abs": 1284, "len_title": 118, "len_sents": 41432, "len_sents_1": 31228, "n_sents": 286, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1288, "title": "WIQOR: A dataset for what-if analysis of Operations Research problems", "abs": "We formalize the mathematical program modification (MPM) task, in which the goal is to revise a mathematical program according to an inquiry expressed in natural language. These inquiries, which we refer to as what-if questions, express a desire to understand how the optimal solution to an optimization problem changes with the addition, deletion or revision of constraints. In detail, each MPM instance\nis a triple consisting of: 1) a natural language specification that summarizes an optimization problem, 2) the canonical formulation of the problem, and 3) a natural language what-if question. The goal is to predict the updated canonical formulation with respect to the question. To support the study of this task, we construct WIQOR, a dataset of 1,946 MPM instances, derived from NL4OPT (Ramamonjison et al., 2023), but with the number of decision variables extended to more than 30 for some problems. In experiments, we observe that Llama 3.1 70B instruct under the in-context learning paradigm achieves 69% accuracy on the easiest test instances, but only 36% accuracy on the most complicated problems. We release WIQOR in the hopes of spurring additional study of MPM and ultimately enabling non-technical users to conduct what-if analyses without the help of technical experts.", "keywords": [ "datasets", "operations research", "counterfactual reasoning", "mathematical reasoning" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 1, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "nnPkQb0Z0H", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Bin Zhang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Lamei Di", "gender": "unknown", "institution": "Xi'an Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Yiming Wang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 36, "n_ref": 90, "n_ref_all": 115, "n_fig": 3, "n_tab": 11, "L_tab": 6125, "n_element_tab": 331, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2245, "n_element_tab_1": 108, "formula_len_all": 580, "formula_len_all_1": 446, "len_all": 164809, "len_all_1": 71662, "len_abs": 1263, "len_title": 162, "len_sents": 50045, "len_sents_1": 31028, "n_sents": 385, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Reject", "L_title": 114, "L_abs": 1268, "title": "EEG-Based Emotion Recognition via Prototype-Guided Disambiguation and Noise Augmentation in Partial-Label Learning", "abs": "EEG-based emotion recognition offers an objective method for diagnosing emotion-related health issues, but the inherent complexity of emotions often leads to annotation errors and noisy labels. To simulate this labeling process in emotion recognition, we propose a semantic-based candidate label generation method leveraging the GloVe vectors, which considers the semantic relationships between emotions. Under the Partial Label Learning (PLL) scenario, we introduce a novel model called PGNA-PL (Prototype-Guided Noise-Augmented Partial Label Learning). This model learns inter-class relationships of emotions using prototypes, and uses a self-distillation mechanism to iteratively guide the classifier's disambiguation process. To address the low signal-to-noise ratio (SNR) of EEG, we introduce a noise augmentation strategy inspired by the mixup method, incorporating controllable noise to enhance model robustness. Experiments on three public datasets (SEED, SEED-IV, SEED-V) show that our approach achieves state-of-the-art performance, surpassing existing PLL baselines across different candidate label generation modes. Our method effectively disambiguates complex emotions and shows promising results in assisting in the recognition of fear-related disorders.", "keywords": [ "EEG-based emotion recognition", "partial label learning", "self-distillation", "Prototypes", "noise augmentation" ], "rating_list": [ 5, 5, 5, 3, 6 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 5, 4 ] }, { "paper_id": "nnAPWDt4hn", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "MD TANVIR PARVEZ", "gender": "Male", "institution": "Islamic University Bangladesh", "country": "BD", "position": "Undergrad student" }, { "name": "MD Tanvir Hassan", "gender": "Male", "institution": "Bangladesh University of Engineering and Technology", "country": "BD", "position": "Undergrad student" }, { "name": "Mahir Labib Dihan", "gender": "Male", "institution": "Bangladesh University of Engineering and Technology", "country": "BD", "position": "Undergrad student" }, { "name": "Md Almash Alam", "gender": "Male", "institution": "Rajshahi University of Engineering and Technology", "country": "BD", "position": "MS student" }, { "name": "Md Hasebul Hasan", "gender": "Male", "institution": "Bangladesh University of Engineering and Technology", "country": "BD", "position": "MS student" }, { "name": "Md Rizwan Parvez", "gender": "Male", "institution": "Qatar Computing Research Institute", "country": "QA", "position": "Researcher" }, { "name": "Mohammed Eunus Ali", "gender": "unknown", "institution": "Bangladesh University of Engineering and Technology", "country": "", "position": "Full Professor" }, { "name": "Muhammad Aamir Cheema", "gender": "unknown", "institution": "Monash University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 35, "n_ref": 66, "n_ref_all": 98, "n_fig": 13, "n_tab": 8, "L_tab": 5173, "n_element_tab": 617, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 3319, "n_element_tab_1": 379, "formula_len_all": 581, "formula_len_all_1": 0, "len_all": 232583, "len_all_1": 65679, "len_abs": 1901, "len_title": 77, "len_sents": 70344, "len_sents_1": 28027, "n_sents": 553, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1911, "title": "MapEval: A Map-Based Evaluation of Geo-Spatial Reasoning in Foundation Models", "abs": "Recent advancements in foundation models have enhanced AI systems' capabilities in autonomous tool usage and reasoning.\nHowever, their ability in location or map-based reasoning - which improves daily life by optimizing navigation, facilitating resource discovery, and streamlining logistics - has not been systematically studied. To bridge this gap, we introduce MapEval, a benchmark designed to assess diverse and complex map-based user queries with geo-spatial reasoning. MapEval features three task types (textual, API-based, and visual) that require collecting world information via map tools, processing heterogeneous geo-spatial contexts (e.g., named entities, travel distances, user reviews or ratings, images), and compositional reasoning, which all state-of-the-art foundation models find challenging. Comprising 700 unique multiple-choice questions about locations across 180 cities and 54 countries, MapEval evaluates foundation models' ability to handle spatial relationships, map infographics, travel planning, and navigation challenges. Using MapEval, we conducted a comprehensive evaluation of 28 prominent foundation models. While no single model excelled across all tasks, Claude-3.5-Sonnet, GPT-4o, and Gemini-1.5-Pro achieved competitive performance overall. However, substantial performance gaps emerged, particularly in MapEval, where agents with Claude-3.5-Sonnet outperformed GPT-4o and Gemini-1.5-Pro by 16% and 21%, respectively, and the gaps became even more amplified when compared to open-source LLMs. Our detailed analyses provide insights into the strengths and weaknesses of current models, though all models still fall short of human performance by more than 20% on average, struggling with complex map images and rigorous geo-spatial reasoning. This gap highlights MapEval's critical role in advancing general-purpose foundation models with stronger geo-spatial understanding.", "keywords": [ "MapLLM", "Map Query", "Geo-Spatial Question Answering", "Geo-Spatial Reasoning", "Location Based Service", "Google Maps" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "nmvmPIi185", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Da Cao", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Associate Professor" }, { "name": "Dongyu Wang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Jiawei Wang", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "PhD student" }, { "name": "Shaofei Lu", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuquan Le", "gender": "Male", "institution": "Hunan University", "country": "", "position": "PhD student" }, { "name": "Zhe Quan", "gender": "unknown", "institution": "Computer Science Lab - Pierre and Marie Curie University, Paris, France", "country": "", "position": "PhD student" }, { "name": "Tat-seng Chua", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 8, "n_ref_uni": 46, "n_ref": 78, "n_ref_all": 105, "n_fig": 9, "n_tab": 5, "L_tab": 2242, "n_element_tab": 248, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2148, "n_element_tab_1": 233, "formula_len_all": 1417, "formula_len_all_1": 577, "len_all": 153671, "len_all_1": 68428, "len_abs": 1236, "len_title": 117, "len_sents": 48776, "len_sents_1": 30144, "n_sents": 342, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1240, "title": "Neural Causal Graph for Interpretable and Intervenable Classification", "abs": "Advancements in neural networks have significantly enhanced the performance of classification models, achieving remarkable accuracy across diverse datasets. However, these models often lack transparency and do not support interactive reasoning with human users, which are essential attributes for applications that require trust and user engagement. To overcome these limitations, we introduce an innovative framework, Neural Causal Graph (NCG), that integrates causal inference with neural networks to enable interpretable and intervenable reasoning. We then propose an intervention training method to model the intervention probability of the prediction, which can serve as a contextual prompt to facilitate the fine-grained reasoning and human-AI interaction ability of NCG. Our experiments reveal that the proposed framework significantly enhances the performance of classical classification baselines. Furthermore, NCG achieves nearly 95% top-1 accuracy on the ImageNet dataset by employing a test-time intervention method. This capability not only supports sophisticated post-hoc interpretation but also enables dynamic human-AI interactions, significantly improving the model's transparency and applicability in real-world scenarios.", "keywords": [ "Causal Inference", "Graph Neural Network", "Classification" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 1, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "nmc9ujrZ5R", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Chen Wang", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Jiahui Lei", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Jiatao Gu", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Kostas Daniilidis", "gender": "Male", "institution": "Athena Research and Innovation Centre", "country": "GR", "position": "Affiliated Researcher" }, { "name": "Lingjie Liu", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Assistant Professor" }, { "name": "XUYI MENG", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 36, "n_ref": 99, "n_ref_all": 115, "n_fig": 7, "n_tab": 4, "L_tab": 745, "n_element_tab": 47, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1722, "n_element_tab_1": 51, "formula_len_all": 814, "formula_len_all_1": 766, "len_all": 127291, "len_all_1": 52081, "len_abs": 164, "len_title": 122, "len_sents": 33676, "len_sents_1": 24832, "n_sents": 228, "n_sents_1": 148, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1285, "title": "ZERO-1-to-G: Taming Pretrained 2D Diffusion Models for Direct 3D Generation", "abs": "Recent advances in 2D image generation have achieved remarkable quality, largely driven by the capacity of diffusion models and the availability of large-scale datasets. However, direct 3D generation is still constrained by the scarcity and lower fidelity of 3D datasets. In this paper, we introduce Zero-1-to-G, a novel approach that addresses this problem by enabling direct 3D generation on Gaussian splats through 2D diffusion models. Our key insight is that Gaussian splats, a 3D representation, can be decomposed into multi-view images encoding different attributes. This reframes the challenging task of direct 3D generation within a 2D diffusion framework, allowing us to leverage the rich priors of pretrained 2D diffusion models. To incorporate 3D awareness, we introduce cross-view and cross-attribute attention layers, which capture complex correlations and enforce 3D consistency across generated splats. This makes Zero-1-to-G the first direct 3D generative model to effectively utilize 2D pretrained diffusion priors, enabling efficient training and improved generalization to unseen objects. Extensive experiments on both synthetic and in-the-wild datasets demonstrate superior performance in 3D object generation, offering a new approach to high-quality 3D generation.", "keywords": [ "native 3D diffusion", "3D generative model", "gaussian splats" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "nmRY3BAll4", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Jiachen Zhu", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Kaiming He", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Xinlei Chen", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Yann LeCun", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Director" }, { "name": "Zhuang Liu", "gender": "Male", "institution": "FAIR, Meta AI", "country": "US", "position": "Research Scientist" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 46, "n_ref": 102, "n_ref_all": 121, "n_fig": 5, "n_tab": 10, "L_tab": 1973, "n_element_tab": 196, "n_fig_1": 5, "n_tab_1": 9, "L_tab_1": 1082, "n_element_tab_1": 89, "formula_len_all": 29, "formula_len_all_1": 29, "len_all": 136206, "len_all_1": 65201, "len_abs": 997, "len_title": 90, "len_sents": 36359, "len_sents_1": 32072, "n_sents": 281, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 1042, "title": "Deep Neural Networks without Normalization", "abs": "Normalization layers are ubiquitous in modern neural networks and have long been considered essential. In this work, we demonstrate that we can achieve strong performance without them, using a remarkably simple technique. We introduce Dynamic Tanh (DyT), an element-wise operation:\n$\\mathrm{DyT}(\\mathbf {x}) = \\tanh(\\alpha \\mathbf {x}),$\nas a drop-in replacement to normalization layers (e.g., layer normalization).\nDyT is directly inspired by the simple observation that normalization layers produce tanh-like, S-shaped curves for their input-output mappings.\nWith DyT, networks without normalization layers could match or exceed the performance of their normalization counterparts, while keeping all other training hyperparameters intact. \nExperiments across diverse settings validate this, ranging from recognition to generation, ConvNets to LLMs, and supervised to self-supervised learning. \nOur findings challenge the conventional understanding that normalization layers are indispensable, and provide new insights into their workings.", "keywords": [ "Normalization", "Deep Neural Networks" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "nlwMlQ1RPW", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Aayush Dhakal", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "US", "position": "PhD student" }, { "name": "Adeel Ahmad", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "US", "position": "Postdoc" }, { "name": "Nathan Jacobs", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "US", "position": "Full Professor" }, { "name": "Srikumar Sastry", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "US", "position": "PhD student" }, { "name": "Subash Khanal", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "US", "position": "PhD student" }, { "name": "Xin Xing", "gender": "Male", "institution": "University of Nebraska, Omaha", "country": "US", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 10, "n_ref_uni": 44, "n_ref": 99, "n_ref_all": 120, "n_fig": 9, "n_tab": 7, "L_tab": 1856, "n_element_tab": 344, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1777, "n_element_tab_1": 269, "formula_len_all": 520, "formula_len_all_1": 551, "len_all": 158315, "len_all_1": 65821, "len_abs": 914, "len_title": 114, "len_sents": 44748, "len_sents_1": 27611, "n_sents": 423, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 918, "title": "LD-SDM: Language-Driven Hierarchical Species Distribution Modeling", "abs": "We focus on species distribution modeling using global-scale presence-only data, leveraging geographical and environmental features to map species ranges, as in previous studies. However, we innovate by integrating taxonomic classification into our approach. Specifically, we propose using a large language model to extract a latent representation of the taxonomic classification from a textual prompt. This allows us to map the range of any taxonomic rank, including unseen species, without additional supervision. We also present a new proximity-aware evaluation metric, suitable for evaluating species distribution models, which addresses critical shortcomings of traditional metrics. We evaluated our model for species range prediction, zero-shot prediction, and geo-feature regression and found that it outperforms several state-of-the-art models. We will share code, data, and model checkpoints after acceptance.", "keywords": [ "Species Distribution Modeling", "Language-Driven", "Spherical Harmonics" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 3, 1, 4, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "nlpCeFsSYJ", "primary_area": "applications to robotics, autonomy, planning", "n_author": 8, "author_info_list": [ { "name": "Haodong LI", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Intern" }, { "name": "Junming Wang", "gender": "Male", "institution": "X Square Robot", "country": "CN", "position": "Researcher" }, { "name": "Qian Zhang", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Wei Yin", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Xiaoyang Guo", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "songen gu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Jin Bu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "XIAOXIAO LONG", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Postdoc" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 30, "n_ref": 62, "n_ref_all": 78, "n_fig": 12, "n_tab": 5, "L_tab": 2362, "n_element_tab": 125, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 172, "formula_len_all_1": 158, "len_all": 110718, "len_all_1": 53585, "len_abs": 1733, "len_title": 130, "len_sents": 35325, "len_sents_1": 26377, "n_sents": 284, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1742, "title": "DOME: Taming Diffusion Model into High-Fidelity Controllable Occupancy World Model", "abs": "We propose DOME, a diffusion-based world model that predicts future occupancy frames based on past occupancy observations. The ability of this world model to capture the evolution of the environment is crucial for planning in autonomous driving. Compared to 2D video-based world models, the occupancy world model utilizes a native 3D representation, which features easily obtainable annotations and is modality-agnostic. This flexibility has the potential to facilitate the development of more advanced world models. Existing occupancy world models either suffer from detail loss due to discrete tokenization or rely on simplistic diffusion architectures, leading to inefficiencies and difficulties in predicting future occupancy with controllability. Our DOME exhibits two key features: (1) High-Fidelity and Long-Duration Generation. We adopt a spatial-temporal diffusion transformer to predict future occupancy frames based on historical context. This architecture efficiently captures spatial-temporal information, enabling high-fidelity details and the ability to generate predictions over long durations. (2)Fine-grained Controllability. We address the challenge of controllability in predictions by introducing a trajectory resampling method, which significantly enhances the model\u2019s ability to generate controlled predictions. Extensive experiments on the widely used nuScenes dataset demonstrate that our method surpasses existing baselines in both qualitative and quantitative evaluations, establishing a new state-of-the-art performance on nuScenes. Specifically, our approach surpasses the baseline by 10.5% in mIoU and 21.2% in IoU for occupancy reconstruction, and by 36.0% in mIoU and 24.6% in IoU for 4D occupancy forecasting.", "keywords": [ "World Model", "Diffusion Model", "Occupancy", "Generative Model" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "nlHEfTRo0b", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Gao Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Honghui Wang", "gender": "Male", "institution": "Department of Automation, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Shiji Song", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yifan Pu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 20, "n_ref_uni": 30, "n_ref": 82, "n_ref_all": 107, "n_fig": 6, "n_tab": 11, "L_tab": 3148, "n_element_tab": 528, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 752, "n_element_tab_1": 137, "formula_len_all": 1691, "formula_len_all_1": 1340, "len_all": 154429, "len_all_1": 68240, "len_abs": 1536, "len_title": 117, "len_sents": 47526, "len_sents_1": 30336, "n_sents": 406, "n_sents_1": 238, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1544, "title": "Physics-informed Dynamics Representation Learning for Parametric PDEs", "abs": "While physics-informed neural networks have achieved remarkable progress in modeling dynamical systems governed by partial differential equations (PDEs), their ability to generalize across different scenarios remains restricted. \nTo address this limitation, we present PIDO, a novel physics-informed neural PDE solver that demonstrates robust generalization across various aspects of PDE configurations, including initial conditions, PDE coefficients, and training time horizons.\nPIDO leverages the shared intrinsic structure inherent to dynamical systems with varying properties by \nprojecting the PDE solutions into a latent space via auto-decoding and subsequently learning the dynamics of these latent embeddings conditioned on the PDE coefficients.\nHowever, the inherent optimization challenges associated with physics-informed loss present substantial obstacles to integrating such latent dynamics models. \nTo tackle this issue, we adopt a novel perspective by diagnosing these challenges within the latent space. This approach enables us to enhance both temporal extrapolation ability and training stability of PIDO via simple yet effective regularization techniques, ultimately leading to superior generalization performance compared to its data-driven counterpart.\nThe effectiveness of PIDO is validated on diverse benchmarks, including 1D combined equations and 2D Navier-Stokes equations. Moreover, we investigate the transferability of its learned representations to downstream tasks like long-term integration and inverse problems.", "keywords": [ "physics-informed neural networks" ], "rating_list": [ 6, 5, 6, 3, 6 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 4, 2, 2 ], "confidence_list": [ 3, 2, 4, 5, 3 ] }, { "paper_id": "nkeF3iRJRo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Lingzhe Zhao", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Xiang Liu", "gender": "Not Specified", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Xiaodong Wang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Xin Yuan", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Associate Professor" }, { "name": "Yunhao Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Peidong Liu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 35, "n_ref": 65, "n_ref_all": 82, "n_fig": 6, "n_tab": 8, "L_tab": 2240, "n_element_tab": 88, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1879, "n_element_tab_1": 67, "formula_len_all": 552, "formula_len_all_1": 550, "len_all": 110841, "len_all_1": 52626, "len_abs": 296, "len_title": 114, "len_sents": 31609, "len_sents_1": 23237, "n_sents": 240, "n_sents_1": 169, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1483, "title": "SCISplat: 3D Gaussian Splatting from a Snapshot Compressive Image", "abs": "In this paper, we investigate the potential of Snapshot Compressive Imaging (SCI) for efficiently recovering 3D scenes from a single temporally compressed image. SCI offers a cost-effective approach using a series of 2D masks to compress video data into a single image captured by 2D imaging sensors. However, traditional SCI reconstruction methods face challenges with generalization and maintaining multi-view consistency. Recent advances have introduced Neural Radiance Fields (NeRF) to estimate 3D scenes from SCI images, but NeRF\u2019s implicit representation struggles to capture fine details and support fast training and rendering. To address these issues, we propose SCISplat, a 3D Gaussian Splatting-based framework for decoding SCI images and achieving high-quality scene reconstruction from a single SCI image. First, we design an initialization protocol that robustly estimates the initial point cloud and camera poses from an SCI image, leveraging a learning-based Structure-from-Motion method. Second, we integrate the SCI image formation model into the 3D Gaussian training process and jointly optimize the Gaussians and camera poses to enhance reconstruction quality. Experiments demonstrate that SCISplat surpasses state-of-the-art methods, achieving a 2.3 dB improvement in reconstruction quality and a 10\u00d7 faster training speed. Furthermore, results on real-world datasets show that our approach produces cleaner and sharper details, underscoring its practical value.", "keywords": [ "3D Vision", "Snapshot Compressive Imaging", "Gaussian Splatting" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 2 ] }, { "paper_id": "njyZgDDeY4", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Linjian Meng", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Tianpei Yang", "gender": "Female", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yang Gao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Zhenxing Ge", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 23, "n_ref_uni": 33, "n_ref": 108, "n_ref_all": 133, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 38, "n_element_tab_1": 2, "formula_len_all": 3960, "formula_len_all_1": 1650, "len_all": 131703, "len_all_1": 72105, "len_abs": 1192, "len_title": 99, "len_sents": 35128, "len_sents_1": 27422, "n_sents": 276, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1214, "title": "Efficient Predictive Counterfactual Regret Minimization$^+$ Algorithm in Solving Extensive-Form Games", "abs": "Imperfect-information extensive-form games (IIGs) serve as a foundational model for capturing interactions among multiple agents in sequential settings with hidden information. A common objective of IIGs is to calculate a Nash equilibrium (NE). Counterfactual Regret Minimization (CFR) algorithms have been widely developed to learn an NE in two-player zero-sum IIGs. Among CFR algorithms, Predictive CFR$^+$ (PCFR$^+$) is powerful, usually achieving an extremely fast empirical convergence rate. However, PCFR$^+$ suffers from the significant discrepancy between strategies represented by explicit accumulated counterfactual regrets across two consecutive iterations, which decreases the empirical convergence rate of PCFR$^+$ in practice. To mitigate this significant discrepancy, we introduce a novel and effective variant of PCFR$^+$, termed Pessimistic PCFR$^+$ (P2PCFR$^+$), minimizing the discrepancy between strategies represented by implicit and explicit accumulated regrets within the same iteration. We provide theoretical proof to show that P2PCFR$^+$ exhibits a faster theoretical convergence rate than PCFR$^+$. Experimental results demonstrate that P2PCFR$^+$ outperforms other tested CFR variants.", "keywords": [ "Imperfect-Information Extensive-Form Games", "Nash Equilibrium", "Counterfactual Regret Minimization" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 2, 3, 5 ] }, { "paper_id": "njvSBvtiwp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Amandine Brunetto", "gender": "unknown", "institution": "Mines Paris - PSL Research University", "country": "FR", "position": "PhD student" }, { "name": "Fabien Moutarde", "gender": "Male", "institution": "MinesParis PSL", "country": "FR", "position": "Full Professor" }, { "name": "Sascha Hornauer", "gender": "unknown", "institution": "Mines ParisTech", "country": "FR", "position": "Associate Professor" } ], "n_formula": 19, "n_formula_1": 10, "n_ref_uni": 52, "n_ref": 137, "n_ref_all": 164, "n_fig": 11, "n_tab": 7, "L_tab": 1971, "n_element_tab": 163, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 781, "n_element_tab_1": 49, "formula_len_all": 796, "formula_len_all_1": 335, "len_all": 169358, "len_all_1": 59510, "len_abs": 1195, "len_title": 107, "len_sents": 47206, "len_sents_1": 27290, "n_sents": 448, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1127, "title": "NeRAF: 3D Scene Infused Neural Radiance and Acoustic Fields", "abs": "Sound plays a major role in human perception. Along with vision, it provides essential information for understanding our surroundings. Despite advances in neural implicit representations, learning acoustics that align with visual scenes remains a challenge. We propose NeRAF, a method that jointly learns acoustic and radiance fields. NeRAF synthesizes both novel views and spatialized room impulse responses (RIR) at new positions by conditioning the acoustic field on 3D scene geometric and appearance priors from the radiance field. The generated RIR can be applied to auralize any audio signal. Each modality can be rendered independently and at spatially distinct positions, offering greater versatility. We demonstrate that NeRAF generates high-quality audio on SoundSpaces and RAF datasets, achieving significant performance improvements over prior methods while being more data-efficient. Additionally, NeRAF enhances novel view synthesis of complex scenes trained with sparse data through cross-modal learning.\nNeRAF is designed as a Nerfstudio module, providing convenient access to realistic audio-visual generation.", "keywords": [ "Audio Generation", "Novel View Synthesis", "Neural Acoustic Fields", "Neural Radiance Fields", "Implicit Neural Fields", "Room Impulse Response" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "niywLsa54R", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 13, "author_info_list": [ { "name": "Ihab Bendidi", "gender": "Male", "institution": "Ecole Normale Sup\u00e9rieure de Paris", "country": "FR", "position": "PhD student" }, { "name": "Imran S Haque", "gender": "Male", "institution": "Recursion Pharmaceuticals", "country": "US", "position": "VP, Data Science" }, { "name": "Jason Hartford", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "Assistant Professor" }, { "name": "John Urbanik", "gender": "Non-Binary", "institution": "Recursion Pharmaceuticals", "country": "US", "position": "Researcher" }, { "name": "Juan Sebasti\u00e1n Rodr\u00edguez Vera", "gender": "Male", "institution": "Recursion Pharmaceuticals", "country": "US", "position": "Researcher" }, { "name": "Kian Kenyon-Dean", "gender": "Male", "institution": "Recursion Pharma", "country": "CA", "position": "Researcher" }, { "name": "Konstantin Donhauser", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" }, { "name": "Marta Fay", "gender": "Female", "institution": "Recursion Pharma ", "country": "US", "position": "Researcher" }, { "name": "Nil Sahin", "gender": "Female", "institution": "Recursion Pharmaceuticals", "country": "CA", "position": "Data Scientist" }, { "name": "Oren Kraus", "gender": "Male", "institution": "Recursion", "country": "CA", "position": "Researcher" }, { "name": "Saber Saberian", "gender": "Male", "institution": "Recursion pharma", "country": "CA", "position": "engineer" }, { "name": "Safiye Celik", "gender": "Female", "institution": "Recursion", "country": "US", "position": "Researcher" }, { "name": "Zitong Jerry Wang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 7, "n_formula_1": 1, "n_ref_uni": 38, "n_ref": 100, "n_ref_all": 144, "n_fig": 11, "n_tab": 8, "L_tab": 5495, "n_element_tab": 531, "n_fig_1": 9, "n_tab_1": 9, "L_tab_1": 3084, "n_element_tab_1": 285, "formula_len_all": 263, "formula_len_all_1": 57, "len_all": 186691, "len_all_1": 61425, "len_abs": 1935, "len_title": 131, "len_sents": 47578, "len_sents_1": 27232, "n_sents": 281, "n_sents_1": 149, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1816, "title": "ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy", "abs": "Large-scale cell microscopy screens are used in drug discovery and molecular biology research to study the effects of millions of chemical and genetic perturbations on cells. To use these images in downstream analysis, we need models that can map each image into a feature space that represents diverse biological phenotypes consistently, in the sense that perturbations with similar biological effects have similar representations.\nIn this work, we present the largest foundation model for cell microscopy data to date, a new 1.9 billion-parameter ViT-G/8 MAE trained on over 8 billion microscopy image crops. Compared to a previous published ViT-L/8 MAE, our new model achieves a 60% improvement in linear separability of genetic perturbations and obtains the best overall performance on whole-genome biological relationship recall and replicate consistency benchmarks. We also show these performance trends hold on a public benchmark for measuring compound activity against target genes.\nBeyond scaling, we developed two key methods that improve performance: (1) training on a curated and diverse dataset; and, (2) using biologically motivated linear probing tasks to search across each transformer block for the best candidate representation of whole-genome screens.\nWe find that many self-supervised vision transformers, pretrained on either natural or microscopy images, yield significantly more biologically meaningful representations of microscopy images in their intermediate blocks than in their typically used final blocks, therefore enabling significant cost and energy savings when deploying these large models in real-world applications. More broadly, our approach and results provide insights toward a general strategy for successfully building foundation models for large-scale biological image data.", "keywords": [ "MAE", "microscopy", "transformers", "SSL", "linear probing", "biology", "high-content screening", "foundation models" ], "rating_list": [ 5, 8, 5, 3 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "nibeaHUEJx", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Berken Utku Demirel", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Christian Holz", "gender": "Male", "institution": "ETH Zurich", "country": "CH", "position": "Associate Professor" } ], "n_formula": 56, "n_formula_1": 12, "n_ref_uni": 73, "n_ref": 130, "n_ref_all": 166, "n_fig": 7, "n_tab": 31, "L_tab": 15197, "n_element_tab": 950, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 3735, "n_element_tab_1": 257, "formula_len_all": 4410, "formula_len_all_1": 934, "len_all": 302612, "len_all_1": 69139, "len_abs": 1274, "len_title": 172, "len_sents": 80812, "len_sents_1": 30208, "n_sents": 642, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 122, "L_abs": 1283, "title": "Shifting the Paradigm: A Diffeomorphism Between Time Series Data Manifolds for Achieving Shift-Invariancy in Deep Learning", "abs": "Deep learning models often lack shift invariance, making them sensitive to input shifts that cause changes in output. While recent techniques seek to address this for images, our findings show that these approaches fail to provide shift-invariance in time series, where the data generation mechanism is more challenging due to the interaction of low and high frequencies. Worse, they also decrease performance across several tasks. In this paper, we propose a differentiable bijective function that maps samples from their high-dimensional data manifold to another manifold of the same dimension, without any dimensional reduction. Our approach guarantees that samples---when subjected to random shifts---are mapped to a unique point in the data manifold while preserving all task-relevant information without loss. We theoretically and empirically demonstrate that the proposed transformation guarantees shift-invariance in deep learning models without imposing any limits to the shift. Our experiments on five-time series tasks with state-of-the-art methods show that our proposed approach consistently improves the performance while enabling models to achieve complete shift-invariance without modifying or imposing restrictions on the model's topology. Source code: Double-blind.", "keywords": [ "Time series analysis", "invariance in neural networks" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 4, 2, 1 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "niDo8UYcEs", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Ihor Tymchyshyn", "gender": "Male", "institution": "Snap Inc.", "country": "UA", "position": "Researcher" }, { "name": "Vasyl Vasylenko", "gender": "Male", "institution": "Samsung", "country": "US", "position": "Principal Researcher" }, { "name": "Vitalii Tymchyshyn", "gender": "Male", "institution": "Kyiv Academic University", "country": "UA", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 16, "n_ref": 29, "n_ref_all": 46, "n_fig": 6, "n_tab": 1, "L_tab": 124, "n_element_tab": 17, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 124, "n_element_tab_1": 17, "formula_len_all": 790, "formula_len_all_1": 790, "len_all": 84168, "len_all_1": 51168, "len_abs": 2149, "len_title": 42, "len_sents": 29764, "len_sents_1": 24315, "n_sents": 197, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 2174, "title": "XiEff Representation for Near-Field Optics", "abs": "Near-field optics, or near-field electrodynamics, is a field that studies the interaction between materials and light at spatial scales smaller than the wavelength. At these extremely small scales, below the diffraction limit, the interaction between materials and electromagnetic fields can exhibit unique behaviors and properties not observed in conventional optics. This area of research is crucial for understanding the optical characteristics of nanotechnical systems and nanoscale biological objects. One of the primary tools used in near-field optics research is scanning near-field optical microscopy (SNOM), which allows researchers to measure near-field optical images (NFI). However, these images often lack visual clarity and interpretability, hindering a comprehensive understanding of the properties of the probed particles.\n\nThe main goal of this paper is to introduce a novel approach that addresses these challenges. Inspired by the prominent progress in Neural Radiance Fields (NeRFs) from computer vision and ideas from physics-informed neural networks (PINNs). We propose an unsupervised method that introduces the XiEff representation \u2013 a neural field-based reparameterization of the effective susceptibility tensor. By integrating XiEff into the Lippmann-Schwinger integral equation framework for near-field optics we develop an optimization strategy to reconstruct the effective susceptibility distribution directly from NFI data.\n\nThe optimized XiEff representation provides an interpretable and explainable model of the particle's shape. Extensive evaluations on a synthetically generated NFI dataset demonstrate the effectiveness of the method, achieving high intersection-over-union scores between XiEff and ground truth shapes, even for complex geometries. Furthermore, the approach exhibits desirable robustness to measurement noise, a crucial property for practical applications. The XiEff representation, combined with the proposed optimization framework, potentially introduces a valuable tool for enabling explainable near-field optics imaging and enhancing the understanding of particle characteristics through interpretable representations", "keywords": [ "neural fields", "PINNs", "near-field optics" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 1 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "nhwfzqXlfd", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Fei Wu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Shengyu Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "ZJU100 Young Professor" }, { "name": "Sida Peng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Sihao Liu", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "chengfei lv", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Software Engineer" } ], "n_formula": 19, "n_formula_1": 17, "n_ref_uni": 29, "n_ref": 53, "n_ref_all": 62, "n_fig": 3, "n_tab": 2, "L_tab": 783, "n_element_tab": 113, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 545, "n_element_tab_1": 108, "formula_len_all": 687, "formula_len_all_1": 700, "len_all": 127754, "len_all_1": 58708, "len_abs": 1375, "len_title": 124, "len_sents": 39096, "len_sents_1": 29803, "n_sents": 247, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1379, "title": "LAST: Latent Structure guided Gaussian Splatting from Monocular Human Videos", "abs": "Multiocular human reconstruction aims to create a high-quality 3D human representation from sparse video data. Recently, 3D Gaussian Splatting (3DGS) has demonstrated impressive results in multiocular human reconstruction tasks, exhibiting remarkable speed and accuracy. However, it encounters challenges in scenarios involving intricate clothing and dynamic postures. This problem may stem from pixel-level supervision during the 3DGS optimization process, which probably lead to spurious associations between unrelated visual features (e.g., misinterpreting clothing wrinkles as dependent on body occlusions rather than recognizing that both are influenced by complex postures). To address this issue, we propose the LAST framework for realistic 3D human reconstruction, which integrates a pre-trained Image-to-Point (I2P) model to enhance the 3D Gaussian Splatting optimization pipeline. The core of the LAST is to disentangle meaningful latent factors and realistic dependencies from the input video frames, which allows for dynamic adjustments to the density and attributes of Gaussian points during the optimization process. Experimental results demonstrate that our method significantly improves accuracy and realism in 3D human reconstruction compared to existing techniques, particularly in challenging scenarios involving complex posture and intricate clothing details.", "keywords": [ "Digital Human", "3D reconstruction", "Causal Inference" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "nhrXqy5d5q", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "ChenXinWei", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Jidong Tian", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Postdoc" }, { "name": "Jindou Chen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Liang Wu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaokang Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University, China", "country": "CN", "position": "Professor" }, { "name": "Yanyan Xu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yaohui Jin", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 41, "n_ref": 76, "n_ref_all": 111, "n_fig": 17, "n_tab": 5, "L_tab": 2429, "n_element_tab": 597, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 965, "n_element_tab_1": 286, "formula_len_all": 1524, "formula_len_all_1": 691, "len_all": 154647, "len_all_1": 60151, "len_abs": 1232, "len_title": 144, "len_sents": 42738, "len_sents_1": 26352, "n_sents": 337, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1217, "title": "KinFormer: Generalizable Dynamical Symbolic Regression for catalytic organic Reaction Kinetics", "abs": "Modeling kinetic equations is essential for understanding the mechanisms of chemical reactions, yet a complex and time-consuming task. Kinetic equation prediction is formulated as a problem of dynamical symbolic regression (DSR) subject to physical chemistry constraints. Deep learning (DL) holds the potential to capture reaction patterns and predict kinetic equations from data of chemical species, effectively avoiding empirical bias and improving efficiency compared with traditional analytical methods. Despite numerous studies focusing on DSR and the introduction of Transformers to predict ordinary differential equations, the corresponding models lack generalization abilities across diverse categories of reactions. In this study, we propose KinFormer, a generalizable kinetic equation prediction model. KinFormer utilizes a conditional Transformer to model DSR under physical constraints and employs Monte Carlo Tree Search to apply the model to new types of reactions. Experimental results on 20 types of organic reactions demonstrate that KinFormer not only outperforms classical baselines, but also exceeds Transformer baselines in out-of-domain evaluations, thereby proving its generalization ability.", "keywords": [ "Kinetic Equation Prediction", "Dynamical Symbolic Regression", "Transform; Conditional Strategy", "MCTS" ], "rating_list": [ 3, 8, 5, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 2, 3, 3 ] }, { "paper_id": "nhRXLbVXFP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Mingzhang Yin", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Yang Zhao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yixin Wang", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" } ], "n_formula": 33, "n_formula_1": 16, "n_ref_uni": 52, "n_ref": 113, "n_ref_all": 150, "n_fig": 7, "n_tab": 9, "L_tab": 4721, "n_element_tab": 496, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 658, "n_element_tab_1": 32, "formula_len_all": 2732, "formula_len_all_1": 1242, "len_all": 181465, "len_all_1": 55007, "len_abs": 1315, "len_title": 116, "len_sents": 46062, "len_sents_1": 25055, "n_sents": 344, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1320, "title": "Ordinal Preference Optimization: Aligning Human Preferences via NDCG", "abs": "Aligning Large Language Models (LLMs) with diverse human preferences is a pivotal technique for controlling model behaviors and enhancing generation quality. Reinforcement Learning from Human Feedback (RLHF), Direct Preference Optimization (DPO), and their variants optimize language models by pairwise comparisons. However, when multiple responses are available, these approaches fall short of leveraging the extensive information in the ranking given by the reward models or human feedback. In this work, we propose a novel listwise approach named Ordinal Preference Optimization (OPO), which employs the Normalized Discounted Cumulative Gain (NDCG), a widely-used ranking metric, to better utilize relative proximity within ordinal multiple responses. We develop an end-to-end preference optimization algorithm by approximating NDCG with a differentiable surrogate loss. This approach builds a connection between ranking models in information retrieval and the alignment problem. In aligning multi-response datasets assigned with ordinal rewards, OPO outperforms existing pairwise and listwise approaches on evaluation sets and general benchmarks like AlpacaEval. Moreover, we demonstrate that increasing the pool of negative samples can enhance model performance by reducing the adverse effects of trivial negatives.", "keywords": [ "Large Language Models", "Human Preferences Alignment", "Listwise", "Learning to Rank" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "nhAyhTxrXu", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Han Yi", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "", "position": "MS student" }, { "name": "xiangyu xu", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Zhedong Zheng", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Assistant Professor" }, { "name": "Tat-seng Chua", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 51, "n_ref": 117, "n_ref_all": 145, "n_fig": 12, "n_tab": 4, "L_tab": 1022, "n_element_tab": 133, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 3579, "n_element_tab_1": 119, "formula_len_all": 290, "formula_len_all_1": 311, "len_all": 148181, "len_all_1": 66512, "len_abs": 434, "len_title": 114, "len_sents": 40703, "len_sents_1": 29254, "n_sents": 313, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1541, "title": "Progressive Multi-scale Triplane Network for Text-to-3D Generation", "abs": "The challenge of text-to-3D generation lies in accurately and efficiently crafting 3D objects based on natural language descriptions, a capability that promises substantial reduction in manual design efforts and offers an intuitive interface for user interaction with digital environments. Despite recent advancements, effective recovery of fine-grained details and efficient optimization of high-resolution 3D outputs remain critical hurdles. Drawing inspiration from the efficacious paradigm of progressive learning, we present a novel Multi-scale Triplane Network (MTN) architecture coupled with a tailored progressive learning strategy. As the name implies, the Multi-scale Triplane Network consists of four triplanes transitioning from low to high resolution. This hierarchical structure allows the low-resolution triplane to serve as an initial shape for the high-resolution counterparts, easing the inherent complexity of the optimization process. Furthermore, we introduce the progressive learning scheme that systematically guides the network to shift its attention from prominent coarse-grained structures to intricate fine-grained patterns. This strategic progression ensures that the focus of the model evolves towards emulating the subtlest aspects of the described 3D object. Our experiment verifies that the proposed method performs favorably against contemporary methods. Even for the complex and nuanced textual descriptions, our method consistently excels, delivering robust and viable 3D shapes where other methods falter.", "keywords": [ "Text-to-3D Generation", "Diffusion Model" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "ngxxEksoJi", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Siyu Chen", "gender": "Male", "institution": "Yale University", "country": "", "position": "PhD student" }, { "name": "Xiaohong Chen", "gender": "Female", "institution": "Yale University", "country": "US", "position": "Full Professor" }, { "name": "Zhengling Qi", "gender": "unknown", "institution": "George Washington University", "country": "", "position": "Assistant Professor" }, { "name": "Zhongren Chen", "gender": "Male", "institution": "Yale University", "country": "", "position": "PhD student" }, { "name": "Zhuoran Yang", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 191, "n_formula_1": 19, "n_ref_uni": 50, "n_ref": 119, "n_ref_all": 130, "n_fig": 3, "n_tab": 10, "L_tab": 2667, "n_element_tab": 58, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 623, "n_element_tab_1": 28, "formula_len_all": 20617, "formula_len_all_1": 1264, "len_all": 300880, "len_all_1": 76706, "len_abs": 1531, "len_title": 110, "len_sents": 89423, "len_sents_1": 32614, "n_sents": 912, "n_sents_1": 261, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1567, "title": "Quantile-Optimal Policy Learning under Unmeasured Confounding", "abs": "We study quantile-optimal policy learning where the goal is to find a policy whose reward distribution has the largest $\\alpha$-th quantile for some $\\alpha \\in (0, 1)$. We focus on the offline setting whose generating process involves unobserved confounders. Such a problem suffers from three main challenges: (i) nonlinearity of the quantile objective as a functional of the reward distribution, (ii) unobserved confounding issue, and (iii) insufficient coverage of the offline dataset. To address these challenges, we propose a suite of causal-assisted policy learning methods that provably enjoy strong theoretical guarantees under mild conditions. In particular, to address (i) and (ii), using causal inference tools such as instrumental variables and negative controls, we propose to estimate the quantile objectives by solving nonlinear functional integral equations. Then we adopt a minimax estimation approach with nonparametric models to solve these integral equations, and propose to construct conservative policy estimates that address (iii). The final policy is the one that maximizes these pessimistic estimates. In addition, we propose a novel regularized policy learning method that is more amenable to computation. Finally, we prove that the policies learned by these methods are $\\tilde{O}(n^{-1/2})$ quantile-optimal under a mild coverage assumption on the offline dataset. To our best knowledge, we propose the first sample-efficient policy learning algorithms for estimating the quantile-optimal policy when there exists unmeasured confounding.", "keywords": [ "Quantile Treatment Effect", "Causal Inference", "Offline Contextual Bandit" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "ngmEcEer8a", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Andrey Gromov", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Dan Roberts", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Affiliate Researcher" }, { "name": "Hassan Shapourian", "gender": "Male", "institution": "Cisco", "country": "US", "position": "Researcher" }, { "name": "Kushal Tirumala", "gender": "unknown", "institution": "Research, Facebook", "country": "US", "position": "Researcher" }, { "name": "Paolo Glorioso", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 44, "n_ref": 49, "n_ref_all": 80, "n_fig": 8, "n_tab": 1, "L_tab": 826, "n_element_tab": 8, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9309, "formula_len_all_1": 9585, "len_all": 165522, "len_all_1": 69719, "len_abs": 1246, "len_title": 101, "len_sents": 35551, "len_sents_1": 31061, "n_sents": 195, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1221, "title": "The Unreasonable Ineffectiveness of the Deeper Layers", "abs": "Understanding *where* and *how* knowledge is stored in LLMs is an active and important area of research. In this work, we take a model pruning approach: if removing certain parameters does not affect model output in *question-answering knowledge benchmarks*, then those parameters are likely are not useful for storing knowledge. To find these parameters, we identify the optimal block of layers to prune by considering similarity across layers; then, to \"heal\" the damage, we perform a small amount of finetuning. In particular, we use parameter-efficient finetuning (PEFT) methods, specifically quantization and Low Rank Adapters (QLoRA), such that each of our experiments can be performed on a single A100 GPU. From a practical perspective, these results suggest that layer pruning methods can complement other PEFT strategies to further reduce computational resources of finetuning and can improve the memory and latency of inference. From a scientific perspective, the robustness of these LLMs to the deletion of layers implies either that current pretraining methods are not properly leveraging the parameters in the deeper layers of the network or that the shallow layers play a critical role in storing knowledge.", "keywords": [ "NLP", "Pruning", "Science of Deep Learning", "Efficient Inference" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "nfKfAzkiez", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Andrew Estornell", "gender": "Male", "institution": "ByteDance", "country": "US", "position": "Researcher" }, { "name": "Jean-Francois Ton", "gender": "Not Specified", "institution": "Bytedance", "country": "GB", "position": "Researcher" }, { "name": "Yang Liu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "Kevin Yao", "gender": "Male", "institution": "Meta GenAI", "country": "US", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 23, "n_ref_uni": 20, "n_ref": 24, "n_ref_all": 31, "n_fig": 3, "n_tab": 2, "L_tab": 1525, "n_element_tab": 152, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1103, "formula_len_all_1": 1359, "len_all": 153487, "len_all_1": 54489, "len_abs": 1501, "len_title": 58, "len_sents": 40580, "len_sents_1": 23727, "n_sents": 309, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 862, "title": "ACC-Debate: An Actor-Critic Approach to Multi-Agent Debate", "abs": "Large language models (LLMs) have demonstrated a remarkable ability to serve as general-purpose tools on various language-based tasks. \n Recent works have demonstrated that the efficacy of such models can be improved through iterative dialog between multiple models, frequently referred to as multi-agent debate (MAD).\n While debate shows promise as a means of improving model efficacy, most works in this area treat debate as an emergent behavior, rather than a learned behavior. \n In doing so, current debate frameworks rely on collaborative behaviors to have been sufficiently trained into off-the-shelf models. \n To address this limitation, we propose ACC-Debate, an Actor-Critic based learning framework to produce a two-agent team specialized in debate.\n We demonstrate that ACC-Debate outperforms SotA debate techniques on a wide array of benchmarks.", "keywords": [ "Multi-Agent Debate", "Large Language Model", "Preference Optimization" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "nf4v09zw6O", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Kang-Jun Liu", "gender": "Male", "institution": "Graduate School of Information Sciences, Tohoku University", "country": "JP", "position": "PhD student" }, { "name": "Masanori Suganuma", "gender": "Male", "institution": "Sakana AI", "country": "JP", "position": "Researcher" }, { "name": "Takayuki Okatani", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Team Leader" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 44, "n_ref": 128, "n_ref_all": 157, "n_fig": 6, "n_tab": 7, "L_tab": 1030, "n_element_tab": 199, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 138, "n_element_tab_1": 27, "formula_len_all": 671, "formula_len_all_1": 661, "len_all": 153076, "len_all_1": 67862, "len_abs": 1224, "len_title": 142, "len_sents": 45112, "len_sents_1": 31417, "n_sents": 342, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1225, "title": "Self-Supervised Learning of Intertwined Content and Positional Features for Object Detection", "abs": "We present a novel self-supervised feature learning method using Vision Transformers (ViT) as the backbone, specifically designed for object detection and instance segmentation. Our approach addresses the challenge of extracting features that capture both class and positional information, which are crucial for these tasks. The method introduces two key components: (1) a positional encoding tied to the cropping process in contrastive learning, which utilizes a novel vector field representation for positional embeddings; and (2) masking and prediction, similar to conventional Masked Image Modeling (MIM), applied in parallel to both content and positional embeddings of image patches. These components enable the effective learning of intertwined content and positional features. We evaluate our method against state-of-the-art approaches, pre-training on ImageNet-1K and fine-tuning on downstream tasks. Our method outperforms the state-of-the-art SSL methods on the COCO object detection benchmark, achieving significant improvements with fewer pre-training epochs. These results suggest that better integration of positional information into self-supervised learning can improve performance on dense prediction tasks.", "keywords": [ "Self-supervised learning; Instance segmentation pre-training; Object detection pre-training; Vision transformer" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "neDGc4slhd", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Dan Ventura", "gender": "Male", "institution": "Brigham Young University", "country": "", "position": "Full Professor" }, { "name": "Tyler Sky Trogden", "gender": "Male", "institution": "Brigham Young University", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 10, "n_ref": 314, "n_ref_all": 332, "n_fig": 12, "n_tab": 1, "L_tab": 68, "n_element_tab": 3, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 308, "formula_len_all_1": 338, "len_all": 111751, "len_all_1": 49547, "len_abs": 971, "len_title": 116, "len_sents": 40811, "len_sents_1": 24437, "n_sents": 174, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 966, "title": "An Empirical Study on the Application of TDA to Deep Neural Networks", "abs": "This study aims to analyze the global structure of the functional subgraph of DNNs using tools from topological data analysis (TDA), namely persistent homology (PH) and the curve similarity. Using these methods we present an empirical study on the application of TDA to DNNs in order to gain a better understanding of their architecture and to provide a framework for a similarity measure between DNNs. The study is conducted by training several convolutional neural networks (CNNs) on disjoint subsets of the ImageNet dataset and then by analyzing the structure of their functional graphs across datasets using the Betti curve similarity. Results show that the Betti curve similarity is able to distinguish between different DNN models across datasets and can be a tool for detecting a departure from previous internal representations of those datasets, providing a new method for the analysis of DNNs and a potential path forward for their theoretical development.", "keywords": [ "deep neural networks", "convolutional networks", "topological data analysis", "persistent homology", "Betti numbers", "Betti curves", "Betti curve similarity", "ImageNet", "functional graph" ], "rating_list": [ 3, 1, 6, 1, 3 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2, 1 ], "contribution_list": [ 2, 1, 3, 1, 1 ], "confidence_list": [ 4, 4, 4, 2, 3 ] }, { "paper_id": "ndtFyx7UWs", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Boxi Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Deng Cai", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" }, { "name": "Haonan Qiu", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Hengjia Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Shiwei Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xiang Wang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yingya Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yujie Wei", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 29, "n_ref": 67, "n_ref_all": 89, "n_fig": 13, "n_tab": 3, "L_tab": 150, "n_element_tab": 29, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 123, "n_element_tab_1": 21, "formula_len_all": 189, "formula_len_all_1": 188, "len_all": 113162, "len_all_1": 55001, "len_abs": 582, "len_title": 119, "len_sents": 31182, "len_sents_1": 25519, "n_sents": 232, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1655, "title": "PersonalVideo: High ID-Fidelity Video Customization With Static Images", "abs": "The current text-to-video (T2V) generation has made significant progress in synthesizing realistic general videos, but it is still unexplored in identity-specific human video generation with customized ID images. The key challenge lies in maintaining high ID fidelity consistently while preserving the original motion dynamic and prompt following after the identity injection. Current video identity customization methods mainly rely on reconstructing given identity images on text-to-image models, which have a divergent distribution with the T2V model. This process introduces a tuning-inference gap, leading to identity inaccuracy and dynamic degradation. To tackle this problem, we propose a novel framework, dubbed $\\textbf{PersonalVideo}$, that applies direct supervision on videos synthesized by the T2V model to bridge the gap. Specifically, we introduce a learnable Spatial Identity Adapter under the supervision of pixel-space ID loss, which customizes the specific identity and preserves the original T2V model\u2019s abilities (e.g., motion dynamic and prompt following). Furthermore, we employ simulated prompt augmentation to reduce overfitting by supervising generated results in more semantic scenarios, gaining good robustness even with only a single reference image available. Extensive experiments demonstrate our method\u2019s superiority in delivering high identity faithfulness while preserving the inherent video generation qualities of the original T2V model, outshining prior approaches. Notably, our PersonalVideo seamlessly integrates with pre-trained SD components, such as ControlNet and style LoRA, requiring no extra tuning overhead.", "keywords": [ "Video identity customization; Text to video generation" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 5, 5 ] }, { "paper_id": "ndU9EvrVBH", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Antonio Ferrara", "gender": "Male", "institution": "Technische Universit\u00e4t Graz", "country": "AT", "position": "PhD student" }, { "name": "Carlos Mougan", "gender": "Male", "institution": "University of Southampton", "country": "GB", "position": "PhD student" }, { "name": "Laura State", "gender": "unknown", "institution": "Scuola Normale Superiore", "country": "", "position": "PhD student" }, { "name": "Salvatore Ruggieri", "gender": "Male", "institution": "University of Pisa", "country": "IT", "position": "Full Professor" }, { "name": "Steffen Staab", "gender": "Male", "institution": "University of Stuttgart", "country": "DE", "position": "Full Professor" } ], "n_formula": 38, "n_formula_1": 14, "n_ref_uni": 102, "n_ref": 151, "n_ref_all": 185, "n_fig": 10, "n_tab": 6, "L_tab": 2049, "n_element_tab": 263, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 291, "n_element_tab_1": 72, "formula_len_all": 3721, "formula_len_all_1": 1634, "len_all": 270761, "len_all_1": 66668, "len_abs": 1464, "len_title": 98, "len_sents": 83074, "len_sents_1": 30490, "n_sents": 671, "n_sents_1": 242, "venue": "ICLR2025", "decision": "withdraw", "L_title": 96, "L_abs": 1501, "title": "Measuring the Impact of Equal Treatment as Blindness via Distributions of Explanations Disparity", "abs": "Liberal political philosophy advocates for the policy of \\emph{equal treatment as blindness}, which seeks to achieve fairness by treating individuals without considering their protected characteristics directly. However, this policy has faced longstanding criticism for perpetuating existing inequalities. In machine learning, this policy can be translated into the concept of \\emph{fairness as unawareness}, and be measured using disparate treatment metrics such as Demographic Parity (a.k.a. Statistical Parity). Our analysis reveals that Demographic Parity does not faithfully measure whether individuals are being treated independently of the protected attribute by the model. We introduce the Explanation Disparity metric to measure fairness under \\emph{equal treatment as blindness} policy. Our metric evaluates the fairness of predictive models by analyzing the extent to which the protected attribute can be inferred from the distribution of explanation values, specifically using Shapley values. The proposed metric tests for statistical independence of the explanation distributions over populations with different protected characteristics. We show the theoretical properties of \"Explanation Disparity\" and devise an equal treatment inspector based on the AUC of a Classifier Two-Sample Test. We experiment with synthetic and natural data to demonstrate and compare the notion with related ones. We release \\texttt{explanationspace}, an open-source Python package with methods and tutorials", "keywords": [ "Fairness", "Audits", "Explanations" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 1, 4, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 4, 2, 2 ], "confidence_list": [ 5, 4, 2, 4 ] }, { "paper_id": "nclyFUZpX9", "primary_area": "learning on time series and dynamical systems", "n_author": 1, "author_info_list": [ { "name": "Haixiang Wu", "gender": "unknown", "institution": "Jiangsu University", "country": "", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 17, "n_ref": 32, "n_ref_all": 42, "n_fig": 3, "n_tab": 4, "L_tab": 6603, "n_element_tab": 188, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 6308, "n_element_tab_1": 108, "formula_len_all": 1314, "formula_len_all_1": 1314, "len_all": 73876, "len_all_1": 50424, "len_abs": 1290, "len_title": 98, "len_sents": 20172, "len_sents_1": 19566, "n_sents": 141, "n_sents_1": 131, "venue": "ICLR2025", "decision": "withdraw", "L_title": 50, "L_abs": 1296, "title": "A SSM is Polymerized from Multivariate Time Series", "abs": "For multivariate time series (MTS) tasks, previous state space models (SSMs) followed the modeling paradigm of Transformer-based methods. However, none of them explicitly model the complex dependencies of MTS: the Channel Dependency variations with Time (CDT). In view of this, we delve into the derivation of SSM, which involves approximating continuously updated functions by orthogonal function basis. We then develop Poly-Mamba, a novel method for MTS forecasting. Its core concept is to expand the original orthogonal function basis space into a multivariate orthogonal function space containing variable mixing terms, and make a projection on this space so as to explicitly describe the CDT by weighted coefficients. In Poly-Mamba, we propose the Multivariate Orthogonal Polynomial Approximation (MOPA) as a simplified implementation of this concept. For the simple linear relationship between channels, we propose Linear Channel Mixing (LCM) and generate CDT patterns adaptively for different channels through a proposed Order Combining method. Experiments on six real-world datasets demonstrate that Poly-Mamba outperforms the SOTA methods, especially when dealing with datasets having a large number of channels and complex correlations. The codes and log files are in the supplementary.", "keywords": [ "Time Series Forecasting\uff1bDeep Learning\uff1b State Space Model" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 1, 1, 1 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "ncCuiD3KJQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Cheng-Long Wang", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Dongfang Liu", "gender": "unknown", "institution": "Rochester Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Guangyan Sun", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Mingyu Jin", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Qifan Wang", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Principal Researcher" }, { "name": "Siqi Ma", "gender": "Male", "institution": "Westlake.edu.cn", "country": "", "position": "Researcher" }, { "name": "Tong Geng", "gender": "Male", "institution": "University of Rochester", "country": "", "position": "Assistant Professor" }, { "name": "Yongfeng Zhang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Associate Professor" }, { "name": "Zhenting Wang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Yingnian Wu", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 73, "n_ref": 185, "n_ref_all": 218, "n_fig": 13, "n_tab": 5, "L_tab": 2703, "n_element_tab": 279, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1148, "n_element_tab_1": 84, "formula_len_all": 325, "formula_len_all_1": 290, "len_all": 263908, "len_all_1": 71797, "len_abs": 1480, "len_title": 87, "len_sents": 59330, "len_sents_1": 33025, "n_sents": 474, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 39, "L_abs": 1594, "title": "Visual Agents as Fast and Slow Thinkers", "abs": "Achieving human-level intelligence requires refining cognitive distinctions between \\textit{System 1} and \\textit{System 2} thinking. While contemporary AI, driven by large language models, demonstrates human-like traits, it falls short of genuine cognition. Transitioning from structured benchmarks to real-world scenarios presents challenges for visual agents, often leading to inaccurate and overly confident responses. To address the challenge, we introduce \\textbf{\\textsc{FaST}}, which incorporates the \\textbf{Fa}st and \\textbf{S}low \\textbf{T}hinking mechanism into visual agents. \\textsc{FaST} employs a switch adapter to dynamically select between \\textit{System 1/2} modes, tailoring the problem-solving approach to different task complexity. It tackles uncertain and unseen objects by adjusting model confidence and integrating new contextual data. With this novel design, we advocate a \\textit{flexible system}, \\textit{hierarchical reasoning} capabilities, and a \\textit{transparent decision-making} pipeline, all of which contribute to its ability to emulate human-like cognitive processes in visual intelligence. Empirical results demonstrate that \\textsc{FaST} outperforms various well-known baselines, achieving 80.8\\% accuracy over $VQA^{v2}$ for visual question answering and 48.7\\% $GIoU$ score over ReasonSeg for reasoning segmentation, demonstrate \\textsc{FaST}'s superior performance. Extensive testing validates the efficacy and robustness of \\textsc{FaST}'s core components, showcasing its potential to advance the development of cognitive visual agents in AI systems.", "keywords": [ "Multimodal Large Language Model", "System2 Thinking", "Language Agent" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 2, 2, 4 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "nc0XGK40dn", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Na Zhao", "gender": "Female", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" }, { "name": "Qiongjie Cui", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 17, "n_ref_uni": 44, "n_ref": 123, "n_ref_all": 147, "n_fig": 4, "n_tab": 16, "L_tab": 5963, "n_element_tab": 297, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1741, "n_element_tab_1": 46, "formula_len_all": 1273, "formula_len_all_1": 964, "len_all": 188084, "len_all_1": 71945, "len_abs": 1603, "len_title": 170, "len_sents": 57743, "len_sents_1": 33739, "n_sents": 370, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 122, "L_abs": 1607, "title": "Never Forget the Basics: In-distribution Knowledge Retention for Continual Test-time Adaptation in Human Motion Prediction", "abs": "This paper presents a novel approach to addressing the underexplored challenge of human pose prediction in dynamic target domains that simultaneously contain in-distribution (ID) and out-of-distribution (OOD) data. Existing test-time adaptation (TTA) techniques predominantly focus on OOD data, neglecting the fact that ID data, which closely resembles the training distribution, is often encountered during real-world deployment, leading to significant degradation in ID performance. To address this, we introduce In-Distribution Knowledge Retention (IDKR), a continual TTA framework designed to preserve critical knowledge about ID data while adapting to unseen OOD sequences. Our method introduces an ID-informative subgraph learning strategy that leverages the structural characteristics of human skeletal data to compute a structural graph Fisher Information Matrix (SG-FIM). Unlike prior work, IDKR simultaneously considers both node and edge features in the skeletal graph, with edge features, representing the invariant bone lengths between parent-child joint pairs, being essential for maintaining structural consistency across poses. These edge features are key to extracting reliable SG-FIM parameters, enabling the model to retain parameters critical for ID performance while selectively updating those needed for OOD adaptation. Extensive experiments on multiple benchmark datasets demonstrate that IDKR consistently outperforms state-of-the-art methods, particularly in scenarios involving mixed ID and OOD data, setting a new standard for robust human pose prediction in dynamic environments.", "keywords": [ "Human Pose Prediction", "Domain Adaptation", "Graph Representation Learning", "Graph Out-of-distribution Detection" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "nbwDsdfJJd", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Erzhi Liu", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" }, { "name": "Jerry Yao-Chieh Hu", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Lichen Zhang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" } ], "n_formula": 40, "n_formula_1": 14, "n_ref_uni": 39, "n_ref": 80, "n_ref_all": 84, "n_fig": 0, "n_tab": 1, "L_tab": 115, "n_element_tab": 8, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 528, "n_element_tab_1": 13, "formula_len_all": 4877, "formula_len_all_1": 1514, "len_all": 153983, "len_all_1": 58936, "len_abs": 3314, "len_title": 90, "len_sents": 35242, "len_sents_1": 24757, "n_sents": 389, "n_sents_1": 223, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 1433, "title": "On Differentially Private String Distances", "abs": "Given a database of bit strings $A_1,\\ldots,A_m\\in \\\\{0,1\\\\}^n$, a fundamental data structure task is to estimate the distances between a given query $B\\in \\\\{0,1\\\\}^n$ with all the strings in the database. In addition, one might further want to ensure the integrity of the database by releasing these distance statistics in a secure manner. In this work, we propose differentially private (DP) data structures for this type of tasks, with a focus on Hamming and edit distance. On top of the strong privacy guarantees, our data structures are also time- and space-efficient. In particular, our data structure is $\\epsilon$-DP against any sequence of queries of arbitrary length, and for any query $B$ such that the maximum distance to any string in the database is at most $k$, we output $m$ distance estimates. Moreover,\n\n* For Hamming distance, our data structure answers any query in $\\widetilde O(mk+n)$ time and each estimate deviates from the true distance by at most $\\widetilde O(k/e^{\\epsilon/\\log k})$;\n\n* For edit distance, our data structure answers any query in $\\widetilde O(mk^2+n)$ time and each estimate deviates from the true distance by at most $\\widetilde O(k/e^{\\epsilon/(\\log k \\log n)})$.\n\nFor moderate $k$, both data structures support sublinear query operations. We obtain these results via a novel adaptation of the randomized response technique as a bit flipping procedure, applied to the sketched strings.", "keywords": [ "edit distance", "data structure", "differential privacy" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "nbngu7H3ko", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chaohu Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Gui Tianyi", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Linli Xu", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yu Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 49, "n_ref": 116, "n_ref_all": 128, "n_fig": 3, "n_tab": 4, "L_tab": 2471, "n_element_tab": 103, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1376, "n_element_tab_1": 76, "formula_len_all": 519, "formula_len_all_1": 609, "len_all": 230778, "len_all_1": 56486, "len_abs": 1325, "len_title": 152, "len_sents": 36588, "len_sents_1": 24599, "n_sents": 278, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 1335, "title": "AdPO: Enhancing the Adversarial Robustness of Large Vision-Language Models with Preference Optimization", "abs": "Large Vision-Language Models (LVLMs), such as GPT-4 and LLaVA, have recently witnessed remarkable advancements and are increasingly being deployed in real-world applications. \nHowever, inheriting the sensitivity of visual neural networks, LVLMs remain vulnerable to adversarial attacks, which can result in erroneous or malicious outputs. \nWhile existing efforts utilize adversarial fine-tuning to enhance robustness, they often suffer from performance degradation on clean inputs. \nIn this paper, we proposes AdPO, a novel adversarial defense strategy for LVLMs based on preference optimization. \nPreference optimization methods, such as DPO and RLHF, have been widely used to align large language models (LLMs) with human values and preferences. \nFor the first time, we reframe adversarial training as a preference optimization problem, aiming to enhance the model\u2019s preference for generating normal outputs on clean inputs while rejecting the potential misleading outputs for adversarial examples.\nNotably, AdPO achieves this by solely modifying the image encoder, e.g., CLIP ViT, resulting in superior robustness across a range of downstream tasks (including LVLMs and zero-shot classification).\nOur comprehensive experimental validation confirms the efficacy of the proposed AdPO, which outperforms prior state-of-the-art methods.", "keywords": [ "Adversarial defense", "Adversarial attacks", "Large Vision-Language Models", "Preference optimization" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "nbia2X0urs", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Chengliang Liu", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Postdoc" }, { "name": "Jie Wen", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Peng Chen", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "MS student" }, { "name": "Xiaoling Luo", "gender": "Female", "institution": "Shenzhen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaopeng Jin", "gender": "Male", "institution": "Shenzhen technology University", "country": "CN", "position": "Lecturer" } ], "n_formula": 24, "n_formula_1": 23, "n_ref_uni": 28, "n_ref": 47, "n_ref_all": 77, "n_fig": 8, "n_tab": 10, "L_tab": 4748, "n_element_tab": 284, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3429, "n_element_tab_1": 173, "formula_len_all": 1423, "formula_len_all_1": 1261, "len_all": 134109, "len_all_1": 65429, "len_abs": 2063, "len_title": 162, "len_sents": 40175, "len_sents_1": 25070, "n_sents": 326, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 113, "L_abs": 2073, "title": "Improving Multimodal Protein Function Prediction Using Bidirectional Interaction and Dynamic Selection Mechanisms", "abs": "Protein function prediction is pivotal for uncovering the mechanisms of life processes. Protein function prediction is a multi-label classification task with numerous functional labels that exhibit hierarchical relationships. Relying solely on unimodal protein features is insufficient for computational models to capture complex protein functions adequately. Recently, several methods for protein function prediction have enhanced the performance by integrating multimodal protein features. However, since multimodal protein features describe protein functions from different perspectives, it is challenging to capture the intricate relationships among these multimodal features with different meanings and heterogeneity. Therefore, we propose a multimodal method for protein function prediction that can effectively utilize the intricate internal relationships between spatial structure features (i.e., protein-protein interaction network, subcellular location, and protein domains) and sequence features (i.e., amino acid sequence). In this work, we introduce the Bidirectional Interaction Module (BInM) to facilitate interactive learning between multimodal features by mapping spatial structure and sequence features of proteins to each other. Moreover, to deal with the difficulty of hierarchical multi-label classification in this task, a multi-branch Dynamic Selection Module (DSM) is designed to select the feature representation that is most favorable for current protein function prediction. Comprehensive experiments on human datasets demonstrate that our model outperforms state-of-the-art multimodal-based methods such as Graph2GO, DeepGraphGO, and CFAGO. Furthermore, we assess the efficacy of the features through Davies-Bouldin scores and t-SNE visualization experiments. The experimental results show that our method constructs more useful protein representations through bidirectional interaction and dynamic selection mechanisms, leading to improved accuracy in protein function prediction. The code in this work will be made public after its acceptance.", "keywords": [ "Multimodal", "protein function prediction", "multi-label classification" ], "rating_list": [ 3, 3, 8, 5 ], "soundness_list": [ 1, 2, 3, 1 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 1, 2, 4, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "nb9DiBUt7c", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Siyu Liao", "gender": "Male", "institution": "Rutgers University", "country": "CN", "position": "Unemployed" }, { "name": "Xinyu Ding", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Zhongfeng Wang", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 27, "n_ref": 42, "n_ref_all": 55, "n_fig": 2, "n_tab": 8, "L_tab": 1296, "n_element_tab": 172, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1483, "n_element_tab_1": 193, "formula_len_all": 1415, "formula_len_all_1": 1205, "len_all": 98351, "len_all_1": 72513, "len_abs": 1142, "len_title": 116, "len_sents": 26117, "len_sents_1": 25115, "n_sents": 192, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 767, "title": "Parameter-Efficient Fine-Tuning with Circulant and Diagonal Vectors", "abs": "Foundation models have achieved tremendous success in different domains.\nHowever, their huge computation and storage complexity make these models difficult to fine-tune and also less applicable in practice. \nRecent study shows training in fourier domain can be an effective fine-tuning method in terms of both model performance and number of training parameters. \nIn this work, we propose to further reduce the complexity by using the product of interleaved circulant and diagonal matrices. \nOur method avoids the construction of weight change matrix and applies 1D fast fourier transform (FFT) instead of 2D FFT. \nExperimental results show that our method achieves similar or better performance across various tasks with much less floating-point operations (FLOPs).", "keywords": [ "fine-tuning; circulant matrices; diagonal matrices" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "nb3VjILNVs", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Anirudh Goyal", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Ashish Malik", "gender": "Male", "institution": "Oregon State University", "country": "US", "position": "PhD student" }, { "name": "Frederik Tr\u00e4uble", "gender": "Male", "institution": "Ellis Institute Tuebingen", "country": "DE", "position": "Postdoc" }, { "name": "Hugo Larochelle", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "Adjunct Professor" }, { "name": "Michael C Mozer", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Sanjeev Arora", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "Full Professor" }, { "name": "Vedant Shah", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "MS student" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 35, "n_ref": 77, "n_ref_all": 109, "n_fig": 7, "n_tab": 6, "L_tab": 1577, "n_element_tab": 138, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 140, "formula_len_all_1": 35, "len_all": 154621, "len_all_1": 67670, "len_abs": 880, "len_title": 98, "len_sents": 57403, "len_sents_1": 34708, "n_sents": 408, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 831, "title": "Low Compute Unlearning via Sparse Representations", "abs": "Machine \\emph{unlearning}, which involves erasing knowledge about a \\emph{forget set} from a trained model, can prove to be \ncostly and infeasible using existing techniques. We propose a low compute unlearning technique based on a discrete representational bottleneck. We show that the proposed technique efficiently unlearns the forget set and incurs negligible damage to the model's performance on the rest of the data set. We evaluate the proposed technique on the problem of *class unlearning* using four datasets: CIFAR-10, CIFAR-100, LACUNA-100 and ImageNet-1k. We compare the proposed technique to SCRUB, a state-of-the-art approach which uses knowledge distillation for unlearning. Across all three datasets, the proposed technique performs as well as, if not better than SCRUB while incurring minimal computational cost.", "keywords": [ "Sparse Representations", "Discrete Bottlenecks", "Model Editing", "Unlearning" ], "rating_list": [ 6, 6, 8, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 2, 4, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "natXOadi7j", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 1, "author_info_list": [ { "name": "Mikko A Heikkil\u00e4", "gender": "unknown", "institution": "University of Helsinki", "country": "FI", "position": "Postdoc" } ], "n_formula": 36, "n_formula_1": 22, "n_ref_uni": 36, "n_ref": 79, "n_ref_all": 92, "n_fig": 3, "n_tab": 5, "L_tab": 621, "n_element_tab": 55, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 238, "n_element_tab_1": 3, "formula_len_all": 1987, "formula_len_all_1": 1297, "len_all": 174897, "len_all_1": 60165, "len_abs": 932, "len_title": 142, "len_sents": 45296, "len_sents_1": 28231, "n_sents": 324, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 934, "title": "On Joint Noise Scaling in Differentially Private Federated Learning with Multiple Local Steps", "abs": "Federated learning is a distributed learning setting where the main aim is to train machine learning models without having to share raw data but only what is required for learning. To guarantee training data privacy and high-utility models, differential privacy and secure aggregation techniques are often combined with federated learning. However, with fine-grained protection granularities the currently existing techniques require the parties to communicate for each local optimisation step, if they want to fully benefit from the secure aggregation in terms of the resulting formal privacy guarantees. In this paper, we show how a simple new analysis allows the parties to perform multiple local optimisation steps while still benefiting from joint noise scaling when using secure aggregation. We show that our analysis enables higher utility models with guaranteed privacy protection under limited number of communication rounds.", "keywords": [ "federated learning", "cross-silo", "differential privacy", "secure aggregation", "privacy-preserving federated learning" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "nZNWrzDBHG", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Chenyang Lyu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Jinsong Su", "gender": "Male", "institution": "Xiamen University", "country": "", "position": "Researcher" }, { "name": "Longyue Wang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Zijun Min", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "bingshuai liu", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Wang Ante", "gender": "Male", "institution": "Xiamen University", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 105, "n_ref_all": 135, "n_fig": 10, "n_tab": 6, "L_tab": 2374, "n_element_tab": 165, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2379, "n_element_tab_1": 143, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 154765, "len_all_1": 64031, "len_abs": 1870, "len_title": 135, "len_sents": 41890, "len_sents_1": 28171, "n_sents": 306, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1893, "title": "VEBench: Towards Comprehensive and Automatic Evaluation for Text-guided Video Editing", "abs": "Video editing task has gained widespread attention in recent years due to their practical applications and rapid advancements, driven by the emergence of diffusion techniques and Multi-modal Large Language Models (MLLMs). However, current automatic evaluation metrics for video editing are mostly unreliable and poorly aligned with human judgments. As a result, researchers heavily rely on human annotation for evaluation, which is not only time-consuming and labor-intensive but also difficult to ensure consistency and objectivity. To address this issue, we introduce **VEBench**, the largest-ever video editing meta-evaluation benchmark to evaluate the reliability of automatic metrics. It includes 152 video clips and 962 text prompts, from which 160 instances are sampled to generate 1,280 edited videos using 8 open-source video editing models, accompanied by human annotations. Especially, the text prompts are first crafted using GPT-4, followed by manual review and careful categorization based on editing types for a systematic evaluation. Our human annotations cover 3 criteria: *Textual Faithfulness*, *Frame Consistency*, and *Video Fidelity*, ensuring the comprehensiveness of evaluation. Since human evaluation is costly, we also propose **VEScore**, employing MLLMs as evaluators to assess edited videos from the criteria above. Experiments show that the best-performing video editing model only reaches an average score of 3.18 (out of a perfect 5), highlighting the challenge of VEBench. Besides, results from more than 10 MLLMs demonstrate the great potential of utilizing VEScore for automatic evaluation. Notably, for Textual Faithfulness, VEScore equipped with LLaVA-OneVision-7B achieves a Pearson Correlation score of 0.48, significantly outperforming previous methods based on CLIP with the highest score of 0.21. The dataset and code will be released upon acceptance.", "keywords": [ "Video Editing", "MLLMs", "Meta Evaluation" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 4, 2 ] }, { "paper_id": "nZBUtzJhf8", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Aske Plaat", "gender": "Male", "institution": "Leiden University", "country": "NL", "position": "Full Professor" }, { "name": "Mike Preuss", "gender": "Male", "institution": "Leiden University", "country": "", "position": "Assistant Professor" }, { "name": "Thomas M. Moerland", "gender": "Male", "institution": "Leiden University", "country": "NL", "position": "Assistant Professor" }, { "name": "Zhao Yang", "gender": "Male", "institution": "Vrije Universiteit Amsterdam", "country": "NL", "position": "Postdoc" }, { "name": "Edward Shichao Hu", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 40, "n_ref": 85, "n_ref_all": 125, "n_fig": 9, "n_tab": 2, "L_tab": 1096, "n_element_tab": 15, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 115, "n_element_tab_1": 15, "formula_len_all": 1047, "formula_len_all_1": 511, "len_all": 145458, "len_all_1": 61749, "len_abs": 1370, "len_title": 94, "len_sents": 43336, "len_sents_1": 29986, "n_sents": 336, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1385, "title": "Guiding Skill Discovery with Foundation Models", "abs": "Learning diverse skills without hand-crafted reward functions could potentially accelerate reinforcement learning in downstream tasks. However, existing skill discovery methods focus solely on maximizing the diversity of skills without considering human preferences, which leads to undesirable behaviors and possibly dangerous skills. For instance, a cheetah robot trained using previous methods learns to roll in all directions to maximize skill diversity, whereas we would prefer it to run without flipping or entering hazardous areas. In this work, we propose a **Fo**undation model **G**uided (FoG) skill discovery method, which incorporates human intentions into skill discovery through foundation models. Specifically, FoG extracts a score function from foundation models to evaluate states based on human intentions, assigning higher values to desirable states and lower to undesirable ones. These scores are then used to re-weight the rewards of skill discovery algorithms. By optimizing the re-weighted skill discovery rewards, FoG successfully learns to eliminate undesirable behaviors, such as flipping or rolling, and to avoid hazardous areas in both state-based and pixel-based tasks. Interestingly, we show that FoG can discover skills involving behaviors that are difficult to define. Interactive visualisations are available from: https://sites.google.com/view/iclr-fog", "keywords": [ "reinforcement learning", "unsupervised skill discovery", "foundation models" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 5, 5, 3, 4 ] }, { "paper_id": "nYpPAT4L3D", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Le Lu", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Full Professor" }, { "name": "Lin Yang", "gender": "Male", "institution": "Westlake University ", "country": "CN", "position": "Full Professor" }, { "name": "Ling Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Researcher" }, { "name": "Qi Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Ruizhe Guo", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Sinuo Wang", "gender": "Female", "institution": "University of Adelaide", "country": "AU", "position": "MS student" }, { "name": "Tingbo Liang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Principal Researcher" }, { "name": "Weiwei Cao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xianghua Ye", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhongyi Shui", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Zhang Jianpeng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 16, "n_ref": 38, "n_ref_all": 58, "n_fig": 13, "n_tab": 7, "L_tab": 4183, "n_element_tab": 630, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 372, "formula_len_all_1": 372, "len_all": 146837, "len_all_1": 46632, "len_abs": 2162, "len_title": 141, "len_sents": 36602, "len_sents_1": 23221, "n_sents": 260, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 93, "L_abs": 2172, "title": "Large-scale and Fine-grained Vision-language Pre-training for Enhanced CT Image Understanding", "abs": "Artificial intelligence (AI) shows great potential in assisting radiologists to improve the efficiency and accuracy of medical image interpretation and diagnosis. However, a versatile AI model requires large-scale data and comprehensive annotations, which are often impractical in medical settings. Recent studies leverage radiology reports as a naturally high-quality supervision for medical images, using contrastive language-image pre-training (CLIP) to develop language-informed models for radiological image interpretation. Nonetheless, these approaches typically contrast entire images with reports, neglecting the local associations between imaging regions and report sentences, which may undermine model performance and interoperability. In this paper, we propose a fine-grained vision-language model (fVLM) for anatomy-level CT image interpretation. Specifically, we explicitly match anatomical regions of CT images with corresponding descriptions in radiology reports and perform contrastive pre-training for each anatomy individually. Fine-grained alignment, however, faces considerable false-negative challenges, mainly from the abundance of anatomy-level healthy samples and similarly diseased abnormalities, leading to ambiguous patient-level pairings. To tackle this issue, we propose identifying false negatives of both normal and abnormal samples and calibrating contrastive learning from patient-level to disease-aware pairing. We curated the largest CT dataset to date, comprising imaging and report data from 69,086 patients, and conducted a comprehensive evaluation of 54 major and important disease (including several most deadly cancers) diagnosis tasks across 15 main anatomies. Experimental results demonstrate the substantial potential of fVLM in versatile medical image interpretation. In the zero-shot classification task, we achieved an average AUC of 81.3% on 54 diagnosis tasks, surpassing CLIP and supervised methods by 12.9% and 8.0%, respectively. Additionally, on the publicly available CT-RATE and Rad-ChestCT benchmarks, our fVLM outperformed the current state-of-the-art methods with absolute AUC gains of 7.4% and 4.8%, respectively.", "keywords": [ "Vision-language model", "fine-grained alignment", "large-scale pre-training", "CT image" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "nYjAzwor9R", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Gal Mishne", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Ronald R. Coifman", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Full Professor" }, { "name": "Ronen Talmon", "gender": "unknown", "institution": "Technion - Israel Institute of Technology", "country": "IL", "position": "Full Professor" }, { "name": "Ya-Wei Eileen Lin", "gender": "Female", "institution": "Technion - Israel Institute of Technology", "country": "IL", "position": "PhD student" } ], "n_formula": 72, "n_formula_1": 19, "n_ref_uni": 109, "n_ref": 364, "n_ref_all": 409, "n_fig": 12, "n_tab": 8, "L_tab": 3969, "n_element_tab": 446, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9071, "formula_len_all_1": 3345, "len_all": 342318, "len_all_1": 77498, "len_abs": 1199, "len_title": 132, "len_sents": 100034, "len_sents_1": 34479, "n_sents": 765, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1203, "title": "Tree-Wasserstein Distance for High Dimensional Data with a Latent Feature Hierarchy", "abs": "Finding meaningful distances between high-dimensional data samples is an important scientific task. To this end, we propose a new tree-Wasserstein distance (TWD) for high-dimensional data with two key aspects. First, our TWD is specifically designed for data with a latent feature hierarchy, i.e., the features lie in a hierarchical space, in contrast to the usual focus on embedding samples in hyperbolic space. Second, while the conventional use of TWD is to speed up the computation of the Wasserstein distance, we use its inherent tree as a means to learn the latent feature hierarchy. The key idea of our method is to embed the features into a multi-scale hyperbolic space using diffusion geometry and then present a new tree decoding method by establishing analogies between the hyperbolic embedding and trees. We show that our TWD computed based on data observations provably recovers the TWD defined with the latent feature hierarchy and that its computation is efficient and scalable. We showcase the usefulness of the proposed TWD in applications to word-document and single-cell RNA-sequencing datasets, demonstrating its advantages over existing TWDs and methods based on pre-trained models.", "keywords": [ "diffusion geometry", "hyperbolic geometry", "tree-Wasserstein distance", "high-dimensional hyperbolic tree decoding" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "nYPuSzGE3X", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 8, "author_info_list": [ { "name": "Jia Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Junzhou Huang", "gender": "Male", "institution": "University of Texas, Arlington", "country": "US", "position": "Full Professor" }, { "name": "Tingyang Xu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yang Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yifan Niu", "gender": "Non-Binary", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yu Rong", "gender": "Male", "institution": "Alibaba Group ", "country": "CN", "position": "Senior Staff Algorithm Engineer" }, { "name": "Ziqi Gao", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "An Bian", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Senior researcher " } ], "n_formula": 40, "n_formula_1": 14, "n_ref_uni": 50, "n_ref": 93, "n_ref_all": 112, "n_fig": 10, "n_tab": 6, "L_tab": 2436, "n_element_tab": 267, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1516, "n_element_tab_1": 210, "formula_len_all": 2256, "formula_len_all_1": 770, "len_all": 176371, "len_all_1": 70468, "len_abs": 2774, "len_title": 75, "len_sents": 46888, "len_sents_1": 28933, "n_sents": 422, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1229, "title": "InversionGNN: A Dual Path Network for Multi-Property Molecular Optimization", "abs": "Exploring chemical space to find novel molecules that simultaneously satisfy multiple properties is crucial in drug discovery. However, existing methods often struggle with trading off multiple properties due to the conflicting or correlated nature of chemical properties. To tackle this issue, we introduce InversionGNN framework, an effective yet sample-efficient dual-path graph neural network (GNN) for multi-objective drug discovery. In the direct prediction path of InversionGNN, we train the model for multi-property prediction to acquire knowledge of the optimal combination of functional groups.\nThen the learned chemical knowledge helps the inversion generation path to generate molecules with required properties. \nIn order to decode the complex knowledge of multiple properties in the inversion path, we propose a gradient-based pareto search method to balance conflicting properties and generate Pareto optimal molecules. \nAdditionally, InversionGNN is able to search the full Pareto front approximately in discrete chemical space. Comprehensive experimental evaluations show that InversionGNN is both effective and sample-efficient in various real-world discrete multi-objective settings including drug discovery.", "keywords": [ "multi-objective drug discovery" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 2, 3, 4 ] }, { "paper_id": "nYEw2KHVxl", "primary_area": "applications to robotics, autonomy, planning", "n_author": 3, "author_info_list": [ { "name": "Bryan Chan", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "PhD student" }, { "name": "James Bergstra", "gender": "unknown", "institution": "Ocado Technology", "country": "", "position": "Researcher" }, { "name": "Yip Sang Leung", "gender": "Male", "institution": "Ocado Techonology", "country": "CA", "position": "Engineer" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 66, "n_ref": 172, "n_ref_all": 200, "n_fig": 13, "n_tab": 1, "L_tab": 246, "n_element_tab": 13, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 848, "formula_len_all_1": 665, "len_all": 190987, "len_all_1": 70015, "len_abs": 1232, "len_title": 141, "len_sents": 50781, "len_sents_1": 32598, "n_sents": 364, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1244, "title": "Offline-to-online Reinforcement Learning for Image-based Grasping with Scarce Demonstrations", "abs": "Offline-to-online reinforcement learning (O2O RL) aims to obtain a continually improving policy as it interacts with the environment, while ensuring the initial behaviour is satisficing.\n This satisficing behaviour is necessary for robotic manipulation where random exploration can be costly due to catastrophic failures and time.\n O2O RL is especially compelling when we can only obtain a scarce amount of (potentially suboptimal) demonstrations\u2014a scenario where behavioural cloning (BC) is known to suffer from distribution shift.\n Previous works have outlined the challenges in applying O2O RL algorithms under the image-based environments.\n In this work, we propose a novel O2O RL algorithm that can learn in a real-life image-based robotic vacuum grasping task with a small number of demonstrations where BC fails majority of the time.\n The proposed algorithm replaces the target network in off-policy actor-critic algorithms with a regularization technique inspired by neural tangent kernel.\n We demonstrate that the proposed algorithm can reach above 90\\% success rate in under two hours of interaction time, with only 50 human demonstrations, while BC and two commonly-used RL algorithms fail to achieve similar performance.", "keywords": [ "Reinforcement learning", "learning from demonstrations", "image-based grasping" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 3, 4, 2, 5 ] }, { "paper_id": "nXV3C8aKxZ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Hongyin Luo", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Wei Sun", "gender": "Male", "institution": "BitEnergy AI", "country": "US", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 35, "n_ref": 52, "n_ref_all": 68, "n_fig": 4, "n_tab": 5, "L_tab": 1376, "n_element_tab": 162, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 638, "n_element_tab_1": 74, "formula_len_all": 1497, "formula_len_all_1": 1111, "len_all": 114542, "len_all_1": 55046, "len_abs": 1793, "len_title": 109, "len_sents": 30728, "len_sents_1": 25108, "n_sents": 246, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1799, "title": "Addition is All You Need for Energy-efficient Language Models", "abs": "Large neural networks spend most computation on floating point tensor multiplications. In this work, we find that a floating point multiplier can be approximated by one integer adder with high precision. We propose the linear-complexity multiplication L-Mul algorithm that approximates floating point number multiplication with integer addition operations. The new algorithm costs significantly less computation resource than 8-bit floating point multiplication but achieves higher precision. Compared to 8-bit floating point multiplications, the proposed method achieves higher precision but consumes significantly less bit-level computation. Since multiplying floating point numbers requires substantially higher energy compared to integer addition operations, applying the L-Mul operation in tensor processing hardware can potentially reduce 95% energy cost by element-wise floating point tensor multiplications and 80% energy cost of dot products. We calculated the theoretical error expectation of L-Mul, and evaluated the algorithm on a wide range of textual, visual, and symbolic tasks, including natural language understanding, structural reasoning, mathematics, and commonsense question answering. Our numerical analysis experiments agree with the theoretical error estimation, which indicates that L-Mul with 4-bit mantissa achieves comparable precision as float8_e4m3 multiplications, and L-Mul with 3-bit mantissa outperforms float8_e5m2. Evaluation results on popular benchmarks show that directly applying L-Mul to the attention mechanism is almost lossless. We further show that replacing all floating point multiplications with 3-bit mantissa L-Mul in a transformer model achieves equivalent precision as using float8_e4m3 as accumulation precision in both fine-tuning and inference.", "keywords": [ "energy saving", "tensor multiplication", "transformer" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "nXTpz8pTHK", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Jeff Bilmes", "gender": "Male", "institution": "University of Washington, Seattle", "country": "US", "position": "Full Professor" }, { "name": "Tian Li", "gender": "unknown", "institution": "University of Chicago", "country": "US", "position": "Assistant Professor" }, { "name": "Tianyi Zhou", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" } ], "n_formula": 42, "n_formula_1": 15, "n_ref_uni": 50, "n_ref": 116, "n_ref_all": 160, "n_fig": 6, "n_tab": 11, "L_tab": 3287, "n_element_tab": 449, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 751, "n_element_tab_1": 28, "formula_len_all": 2799, "formula_len_all_1": 914, "len_all": 180345, "len_all_1": 81923, "len_abs": 1346, "len_title": 88, "len_sents": 50959, "len_sents_1": 35809, "n_sents": 455, "n_sents_1": 291, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1360, "title": "Reweighting Local Mimina with Tilted SAM", "abs": "Sharpness-Aware Minimization (SAM) has been demonstrated to improve the generalization performance of overparameterized models by seeking flat minima on the loss landscape through optimizing model parameters that incur the largest loss within a neighborhood. \nNevertheless, such min-max formulations are computationally challenging especially when the problem is highly non-convex. Additionally, focusing only on the worst-case local solution while ignoring potentially many other local solutions may be suboptimal when searching for flat minima. In this work, we propose Tilted SAM (TSAM), a generalization of SAM inspired by exponential tilting that effectively assigns higher priority to local solutions that are flatter and that incur larger losses. TSAM is parameterized by a tilt hyperparameter $t$ and reduces to SAM as $t$ approaches infinity. We prove that (1) the TSAM objective is smoother than SAM and thus easier to optimize; and (2) TSAM explicitly favors flatter minima as $t$ increases. This is desirable as flatter minima could have better generalization properties for certain tasks. We develop algorithms motivated by the discretization of Hamiltonian dynamics to solve TSAM. Empirically, TSAM arrives at flatter local minima and results in superior test performance than the baselines of SAM and ERM across a range of image and text tasks.", "keywords": [ "sharpness-aware optimization", "exponential tilting", "generalization" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "nWdQX5hOL9", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Jundong Li", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" }, { "name": "Patrick Soga", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Song Wang", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Yinhan He", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Yushun Dong", "gender": "Male", "institution": "Florida State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 5, "n_ref_uni": 44, "n_ref": 86, "n_ref_all": 110, "n_fig": 10, "n_tab": 1, "L_tab": 234, "n_element_tab": 26, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1247, "formula_len_all_1": 183, "len_all": 175954, "len_all_1": 69913, "len_abs": 1573, "len_title": 149, "len_sents": 56120, "len_sents_1": 35861, "n_sents": 550, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1580, "title": "Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective", "abs": "Graph Neural Networks (GNNs) have achieved remarkable success in various graph-based learning tasks. While their performance is often attributed to the powerful neighborhood aggregation mechanism, recent studies suggest that other components such as non-linear layers may also significantly affecting how GNNs process the input graph data in the spectral domain. Such evidence challenges the prevalent opinion that neighborhood aggregation mechanisms dominate the behavioral characteristics of GNNs in the spectral domain. To demystify such a conflict, this paper introduces a comprehensive benchmark to measure and evaluate GNNs' capability in capturing and leveraging the information encoded in different frequency components of the input graph data. Specifically, we first conduct an exploratory study demonstrating that GNNs can flexibly yield outputs with diverse frequency components even when certain frequencies are absent or filtered out from the input graph data. We then formulate a novel research problem of measuring and benchmarking the performance of GNNs from a spectral perspective. To take an initial step towards a comprehensive benchmark, we design an evaluation protocol supported by comprehensive theoretical analysis. Finally, we introduce a comprehensive benchmark on real-world datasets, revealing insights that challenge prevalent opinions from a spectral perspective. We believe that our findings will open new avenues for future advancements in this area. Our implementations can be found at: https://anonymous.4open.science/r/Spectral-Benchmark-9B56/.", "keywords": [ "Graph Neural Networks", "Spectral Graph Theory", "Graph Signal Filtering" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 3, 1, 4, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 2, 2, 4, 3 ], "confidence_list": [ 5, 4, 2, 3 ] }, { "paper_id": "nWT6LxbuGi", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Ness Shroff", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Full Professor" }, { "name": "Peizhong Ju", "gender": "Male", "institution": "University of Kentucky", "country": "US", "position": "Assistant Professor" }, { "name": "Yingbin Liang", "gender": "Female", "institution": "The Ohio State University", "country": "US", "position": "Professor" }, { "name": "Yuchen Liang", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Postdoc" } ], "n_formula": 33, "n_formula_1": 30, "n_ref_uni": 27, "n_ref": 86, "n_ref_all": 94, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3166, "formula_len_all_1": 3029, "len_all": 130560, "len_all_1": 71313, "len_abs": 1628, "len_title": 126, "len_sents": 34522, "len_sents_1": 31545, "n_sents": 258, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1636, "title": "Theory on Score-Mismatched Diffusion Models and Zero-Shot Conditional Samplers", "abs": "The denoising diffusion model has recently emerged as a powerful generative technique, capable of transforming noise into meaningful data. While theoretical convergence guarantees for diffusion models are well established when the target distribution aligns with the training distribution, practical scenarios often present mismatches. One common case is in zero-shot conditional diffusion sampling, where the target conditional distribution is different from the (unconditional) training distribution. These score-mismatched diffusion models remain largely unexplored from a theoretical perspective. In this paper, we present the first performance guarantee with explicit dimensional dependencies for general score-mismatched diffusion samplers, focusing on target distributions with finite second moments. We show that score mismatches result in an asymptotic distributional bias between the target and sampling distributions, proportional to the accumulated mismatch between the target and training distributions. This result can be directly applied to zero-shot conditional samplers for any conditional model, irrespective of measurement noise. Interestingly, the derived convergence upper bound offers useful guidance for designing a novel bias-optimal zero-shot sampler in linear conditional models that minimizes the asymptotic bias. For such bias-optimal samplers, we further establish convergence guarantees with explicit dependencies on dimension and conditioning, applied to several interesting target distributions, including those with bounded support and Gaussian mixtures. Our findings are supported by numerical studies.", "keywords": [ "generative models", "denoising diffusion probabilistic model (DDPM)", "convergence analysis", "zero-shot conditional sampling", "model mismatch" ], "rating_list": [ 6, 8, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 3, 2, 2, 2, 3 ] }, { "paper_id": "nWO75tVjfp", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Altuna Akalin", "gender": "unknown", "institution": "Max Delbr\u00fcck Center for Molecular Medicine", "country": "DE", "position": "Principal Researcher" }, { "name": "Bora Uyar", "gender": "Male", "institution": "Max Delbr\u00fcck Center for Molecular Medicine", "country": "DE", "position": "Postdoc" }, { "name": "Vedran Franke", "gender": "unknown", "institution": "Max Delbr\u00fcck Center for Molecular Medicine", "country": "DE", "position": "Researcher" }, { "name": "Ahmet Sar\u0131g\u00fcn", "gender": "Male", "institution": "Max Delbr\u00fcck Center for Molecular Medicine", "country": "DE", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 2, "n_ref_uni": 36, "n_ref": 56, "n_ref_all": 80, "n_fig": 10, "n_tab": 2, "L_tab": 172, "n_element_tab": 15, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 134, "n_element_tab_1": 3, "formula_len_all": 1315, "formula_len_all_1": 318, "len_all": 118587, "len_all_1": 50941, "len_abs": 1750, "len_title": 175, "len_sents": 37498, "len_sents_1": 24983, "n_sents": 247, "n_sents_1": 156, "venue": "ICLR2025", "decision": "withdraw", "L_title": 126, "L_abs": 1718, "title": "CompassDock: Comprehensive Accurate Assessment Approach for Deep Learning-Based Molecular Docking in Inference and Fine-Tuning", "abs": "Datasets used for molecular docking, such as PDBBind, contain technical variability - they are noisy. Although the origins of the noise have been discussed , a comprehensive analysis of physical, chemical, and bioactivity characteristics of the datasets is still lacking. To address this gap, we introduce the Compass. Compass integrates two key components: PoseCheck, which examines ligand strain energy, protein-ligand steric clashes, and interactions, and AA-Score, a new empirical scoring function for calculating binding affinity energy. Together, these form a unified workflow that assesses both the physical/chemical properties and bioactivity favorability of ligands and protein-ligand interactions. Our analysis of the PDBBind dataset using Compass reveals substantial noise in the ground truth data. Additionally, we propose CompassDock, which incorporates the Compass module with DiffDock, the state-of-the-art deep learning-based molecular docking method, to enable accurate assessment of docked ligands during inference. Finally, we present a new paradigm for enhancing molecular docking model performance by fine-tuning with Compass Scores, which encompass binding affinity energy, strain energy, and the number of steric clashes identified by Compass. Our results show that, while fine-tuning without Compass improves the percentage of docked poses with RMSD < 2\u00c5, it leads to a decrease in physical/chemical and bioactivity favorability. In contrast, fine-tuning with Compass shows a limited improvement in RMSD < 2\u00c5 but enhances the physical/chemical and bioactivity favorability of the ligand conformation. The source code is available at https://github.com/anonym8171iclr2025/iclr_2025_paperid_8171.", "keywords": [ "Compass", "DL-based Molecular Docking", "LAN-MSE", "Favorable Physico-chemical & Bioactivity Features" ], "rating_list": [ 3, 3, 1, 5 ], "soundness_list": [ 3, 1, 1, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "nW54N85eDT", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Haojin Zhu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Jinsong Lan", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Qingwen Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Shuai Xiao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xu Chen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yunmeng SHU", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Ivor W Tsang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Yuangang Pan", "gender": "Male", "institution": "Agency for Science, Technology and Research (A*STAR)", "country": "SG", "position": "Senior Research Scientist" } ], "n_formula": 20, "n_formula_1": 13, "n_ref_uni": 43, "n_ref": 95, "n_ref_all": 119, "n_fig": 6, "n_tab": 4, "L_tab": 238, "n_element_tab": 21, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2207, "formula_len_all_1": 1163, "len_all": 166767, "len_all_1": 72667, "len_abs": 1728, "len_title": 111, "len_sents": 51479, "len_sents_1": 35757, "n_sents": 386, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1738, "title": "Predicting User Behaviors with Scene via Dual Sequence Networks", "abs": "Modeling sequential user behaviors for future action prediction is crucial in improving user's information retrieval experience. Recent studies highlight the importance of incorporating contextual information to enhance prediction performance. One crucial and typical contextual information is the scene feature which we define it as sub-interfaces within an app, created by designers to provide specific functionalities, such as ''text2product search\" and ''live\" in e-commence apps. Different scenes exhibit distinct functionalities and usage habits, leading to significant distribution gap in user engagement across them. Popular sequential behavior models either ignore the scene feature or merely use it as attribute embeddings, which could lead to substantial information loss or cannot capture the interplay between scene and item in modeling dynamic user interests. In this work, we propose a novel Dual Sequence Prediction network (DSPnet) to effectively capture the interplay between scene and item sequences for future behavior prediction. DSPnet consists of two parallel networks dedicated to predicting scene and item sequences, and a sequence feature enhancement module to capture the interplay. Further, considering the randomness and noise in learning sequence dynamics, we introduce Conditional Contrastive Regularization (CCR) loss to capture the invariance of similar historical sequences. Theoretical analysis suggests that DSPnet can learn the joint relationships between scene and item sequences, and also show better robustness on real-world user behaviors. Extensive experiments are conducted on one public benchmark and two collected industrial datasets. The codes and collected datasets will be made public soon.", "keywords": [ "dual sequence networks", "scene-aware", "behavior prediction" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "nVbbB3Jmyo", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Mengzhu Wang", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Nan Yin", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Shangsong Liang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Siwei Liu", "gender": "Male", "institution": "University of Aberdeen", "country": "GB", "position": "Assistant Professor" }, { "name": "Yingxu Wang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" } ], "n_formula": 72, "n_formula_1": 27, "n_ref_uni": 45, "n_ref": 103, "n_ref_all": 129, "n_fig": 10, "n_tab": 17, "L_tab": 9382, "n_element_tab": 881, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 4459, "n_element_tab_1": 241, "formula_len_all": 7535, "formula_len_all_1": 2040, "len_all": 202358, "len_all_1": 69503, "len_abs": 1583, "len_title": 112, "len_sents": 49885, "len_sents_1": 27466, "n_sents": 398, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1593, "title": "Degree-aware Spiking Graph Domain Adaptation for Classification", "abs": "Spiking Graph Networks (SGNs) have garnered significant interest from both researchers and industry due to their ability to address energy consumption challenges in graph classification. However, SGNs are typically inference under the same distribution of training dataset, which is difficult to satisfy in real applications. In this paper, we first propose the domain adaptation problem in SGNs, and introduce the novel framework named \\textbf{De}gree-aware \\textbf{S}piking \\textbf{G}raph \\textbf{D}omain \\textbf{A}daptation for Classification (\\method{}). To address this problem, we propose solutions in terms of three aspects: node distribution-aware personalized spiking representation, graph feature distribution alignment, and pseudo-label distillation. Firstly, we introduce the personalized spiking representation method that varies with node degrees. The difficulty of triggering a spike is determined by the node degree, allowing this personalized approach to capture more expressive information for classification. Then, we propose the graph feature distribution alignment module that is adversarially trained using membrane potential against a domain discriminator, efficiently maintaining high performance and low energy consumption in the case of inconsistent distribution. Additionally, we extract consistent predictions across two spaces to create reliable pseudo-labels, effectively leveraging unlabeled data to enhance graph classification performance. \nExtensive experiments on benchmark datasets validate the superiority of the proposed \\method{} compared with baselines.", "keywords": [ "spiking graph neural network; domain adaptation" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 4, 1, 2, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "nUpM7egYFd", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "BO WANG", "gender": "Male", "institution": "Vector Institute", "country": "", "position": "Assistant Professor" }, { "name": "Steven Palayew", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "MS student" }, { "name": "Gary Bader", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 35, "n_ref": 61, "n_ref_all": 71, "n_fig": 3, "n_tab": 16, "L_tab": 4229, "n_element_tab": 430, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 3169, "n_element_tab_1": 182, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 135809, "len_all_1": 60160, "len_abs": 1161, "len_title": 137, "len_sents": 38691, "len_sents_1": 27975, "n_sents": 242, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1192, "title": "scMPT: towards applying large language models to complement single-cell foundation models", "abs": "Single-cell foundation models such as scGPT represent a significant advancement in single-cell omics, with an ability to achieve state-of-the-art performance on a variety of downstream biological tasks. However, these models are inherently limited in that a vast amount of information in biology exists as text, which they are unable to leverage. There have therefore been several recent works that propose the use of large language models (LLMs) as an alternative to single-cell foundation models, achieving competitive results. However, there is little understanding of what factors drive this performance, along with a strong focus on using LLMs as an alternative, rather than complementary approach to single-cell foundation models. In this study we therefore investigate what biological insights contribute toward the performance of LLMs when applied to single-cell data, and introduce scMPT; a model which leverages synergies between scGPT, and single-cell representations from LLMs that capture these insights. scMPT demonstrates stronger, more consistent performance than either of its component models, which frequently have large performance gaps between each other across datasets.", "keywords": [ "Single-cell genomics", "multimodality", "interpretability", "LLM" ], "rating_list": [ 1, 3, 3, 5, 5 ], "soundness_list": [ 1, 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 5, 4, 4, 2, 3 ] }, { "paper_id": "nUp1NvgfOr", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Jiayi Ma", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Junjun Jiang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Pengwei Liang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Qing Ma", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Postdoc" }, { "name": "Xiong Zhou", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Zhe Peng", "gender": "unknown", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 53, "n_ref": 88, "n_ref_all": 101, "n_fig": 3, "n_tab": 5, "L_tab": 485, "n_element_tab": 79, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 761, "n_element_tab_1": 104, "formula_len_all": 155, "formula_len_all_1": 155, "len_all": 124927, "len_all_1": 65656, "len_abs": 1173, "len_title": 108, "len_sents": 33578, "len_sents_1": 33459, "n_sents": 214, "n_sents_1": 195, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1186, "title": "Deep Denoising Prior: You Only Need a Deep Gaussian Denoiser", "abs": "Gaussian denoising often serves as the initiation of research in the field of image denoising, owing to its prevalence and intriguing properties. However, deep Gaussian denoiser typically generalizes poorly to other types of noises, such as Poisson noise and real-world noise. In this paper, we reveal that deep Gaussian denoisers have an underlying ability to handle other noises with only ten iterations of self-supervised learning, which is referred to as \\textit{deep denoiser prior}. Specifically, we first pre-train a Gaussian denoising model in a self-supervised manner. Then, for each test image, we construct a pixel bank based on the self-similarity and randomly sample pseudo-instance examples from it to perform test-time adaptation. Finally, we fine-tune the pre-trained Gaussian denoiser using the randomly sampled pseudo-instances. Extensive experiments demonstrate that our test-time adaptation method helps the pre-trained Gaussian denoiser rapidly improve performance in removing both in-distribution and out-of-distribution noise, achieving superior performance compared to existing single-image denoising methods while also significantly reducing computational time.", "keywords": [ "Denoising", "test time adaptation", "low-level vision" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 3, 3, 2, 1 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "nUOmJ4Qop5", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Dexter Neo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Tsuhan Chen", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 34, "n_formula_1": 23, "n_ref_uni": 74, "n_ref": 164, "n_ref_all": 212, "n_fig": 11, "n_tab": 8, "L_tab": 3578, "n_element_tab": 390, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1397, "n_element_tab_1": 18, "formula_len_all": 2328, "formula_len_all_1": 1742, "len_all": 242126, "len_all_1": 59499, "len_abs": 1383, "len_title": 121, "len_sents": 54028, "len_sents_1": 24765, "n_sents": 397, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 757, "title": "Peacock: Multi-Objective Optimization for Deep Neural Network Calibration", "abs": "The rapid adoption of deep neural networks underscores an urgent need for models to be safe, trustworthy and well-calibrated. Despite recent advancements in network calibration, the optimal combination of techniques remains relatively unexplored. By framing the task as a multi-objective optimization problem, we demonstrate that combining state-of-the-art methods can further boost calibration performance. We feature a total of seven state-of-the-art calibration algorithms and provide both theoretical and empirical motivation for their equal and weighted importance unification. We conduct experiments on both in and out-of-distribution computer vision and natural language benchmarks, investigating the speeds and contributions of different components.", "keywords": [ "Deep Neural Network Calibration", "Uncertainty Calibration", "Robustness", "Safety", "Out-of-Distribution" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "nTlzEM1x3B", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Dotan Di Castro", "gender": "Male", "institution": "Bosch", "country": "", "position": "Bosch Center for Artificial Inteligence" }, { "name": "Liran Nochumsohn", "gender": "Male", "institution": "Ben-Gurion University of the Negev", "country": "", "position": "MS student" }, { "name": "Michal Moshkovitz", "gender": "Female", "institution": "Bosch Center for AI", "country": "IL", "position": "Researcher" }, { "name": "Omri Azencot", "gender": "Unspecified", "institution": "Ben-Gurion University of the Negev", "country": "IL", "position": "Assistant Professor" }, { "name": "Orly Avner", "gender": "Female", "institution": "Bosch Center for Artificial Inteligence", "country": "", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 43, "n_ref": 107, "n_ref_all": 125, "n_fig": 11, "n_tab": 5, "L_tab": 5244, "n_element_tab": 352, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 1969, "n_element_tab_1": 86, "formula_len_all": 293, "formula_len_all_1": 206, "len_all": 162461, "len_all_1": 68961, "len_abs": 1523, "len_title": 124, "len_sents": 50961, "len_sents_1": 32846, "n_sents": 384, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1535, "title": "Beyond Data Scarcity: A Frequency-Driven Framework for Zero-Shot Forecasting", "abs": "Time series forecasting is critical in numerous real-world applications, requiring accurate predictions of future values based on observed patterns. While traditional forecasting techniques work well in in-domain scenarios with ample data, they struggle when data is scarce or not available at all, motivating the emergence of zero-shot and few-shot learning settings. Recent advancements often leverage large-scale foundation models for such tasks, but these methods require extensive data and compute resources, and their performance may be hindered by ineffective learning from the available training set. This raises a fundamental question: *What factors influence effective learning from data in time series forecasting?* Toward addressing this, we propose using Fourier analysis to investigate how models learn from synthetic and real-world time series data. Our findings reveal that forecasters commonly suffer from poor learning from data with multiple frequencies and poor generalization to unseen frequencies, which impedes their predictive performance. To alleviate these issues, we present a novel synthetic data generation framework, designed to enhance real data or replace it completely by creating task-specific frequency information, requiring only the sampling rate of the target data. Our approach, **Freq-Synth**, improves the robustness of both foundation as well as non-foundation forecast models in zero-shot and few-shot settings, facilitating more reliable time series forecasting under limited data scenarios.", "keywords": [ "Time Series Forecasting", "Synthetic Data", "Zero-Shot" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "nTZOIlf8YH", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Chaoqun Feng", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "Researcher" }, { "name": "Jieping Ye", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Lei Fu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Peng Li", "gender": "unknown", "institution": "Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Wu Li Xia", "gender": "Male", "institution": "Dalian University of Technology", "country": "", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 23, "n_ref": 51, "n_ref_all": 56, "n_fig": 0, "n_tab": 5, "L_tab": 1770, "n_element_tab": 97, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 1668, "n_element_tab_1": 89, "formula_len_all": 1133, "formula_len_all_1": 1102, "len_all": 95410, "len_all_1": 67326, "len_abs": 1366, "len_title": 112, "len_sents": 34114, "len_sents_1": 30018, "n_sents": 262, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1401, "title": "Differentiation of Multi-objective Data-driven Decision Pipeline", "abs": "Real-world scenarios frequently involve multi-objective data-driven optimization problems, characterized by unknown problem coefficients and multiple conflicting objectives. Traditional two-stage methods independently apply a machine learning model to estimate problem coefficients, followed by invoking a solver to tackle the predicted optimization problem. The independent use of optimization solvers and prediction models may lead to suboptimal performance due to mismatches between their objectives. Recent efforts have focused on end-to-end training of predictive models that use decision loss derived from the downstream optimization problem. However, these methods have primarily focused on single-objective optimization problems, thus limiting their applicability. We aim to propose a multiobjective decision-focused approach to address this gap. In order to better align with the inherent properties of multi-objective optimization problems, we propose a set of novel loss functions. These loss functions are designed to capture the discrepancies between predicted and true decision problems, considering solution space, objective space, and decision quality, named landscape loss, Pareto set loss, and decision loss, respectively. Our experimental results demonstrate that our proposed method significantly outperforms traditional two-stage methods and most current decision-focused methods.", "keywords": [ "decision-focused learning", "multi-objective optimization", "smart prediction-and-optimization", "data-driven optimization" ], "rating_list": [ 3, 1, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "nTAC2NCQUO", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 11, "author_info_list": [ { "name": "Fei Mi", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "James Kwok", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Hong Kong University of Science and Technology", "country": "HK", "position": "Professor" }, { "name": "Kai Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Lanqing HONG", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Qun Liu", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "HK", "position": "Chief Scientist of Speech and Language Computing" }, { "name": "Xin Jiang", "gender": "Male", "institution": "Noah\u2019s Ark Lab, Huawei Technologies", "country": "HK", "position": "Principal Researcher" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yunhao GOU", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zhenguo Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "JIAHUI GAO", "gender": "Female", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zhili LIU", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 44, "n_ref": 91, "n_ref_all": 111, "n_fig": 7, "n_tab": 10, "L_tab": 1479, "n_element_tab": 185, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 1078, "n_element_tab_1": 152, "formula_len_all": 680, "formula_len_all_1": 680, "len_all": 167511, "len_all_1": 62479, "len_abs": 1334, "len_title": 160, "len_sents": 45019, "len_sents_1": 27900, "n_sents": 339, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 112, "L_abs": 1339, "title": "Mixture of insighTful Experts (MoTE): The Synergy of Thought Chains and Expert Mixtures in Safety Self-Alignment", "abs": "As the capabilities of large language models (LLMs) have expanded dramatically, aligning these models with human values presents a significant challenge. Recent studies demonstrate that powerful LLMs can achieve self-alignment by either correcting their initial unsafe responses or autonomously ranking answers without human intervention. In this work, we identify two key limitations: first, they rely on the assumed emergent capabilities of LLMs, and second, they discard all intermediate reasoning steps when aligning the model with updated answers. To address these challenges, we propose a novel self-alignment method that utilizes a Chain of Thought (CoT) approach, termed AlignCoT. This method encompasses stages of Question Analysis, Answer Guidance, and Safe Answer production. It is designed to enable LLMs, even smaller and weaker models like 7B LLMs, to produce high-quality, safe responses. Furthermore, we introduce the Mixture of insighTful Experts (MoTE) architecture, which applies mixture of experts to enhance each component of the AlignCoT process, markedly increasing alignment efficiency. The MoTE approach not only outperforms existing methods in aligning LLMs with human values but also highlights the benefits of using self-generated data, revealing the dual benefits of improved alignment and training efficiency.", "keywords": [ "mixture of experts", "lora", "chain of thoughts", "LLM safety" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 2, 4, 2, 2 ], "presentation_list": [ 2, 4, 1, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "nT89RltY10", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Daniel Rueckert", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Georgios Kaissis", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Florian Alexander H\u00f6lzl", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 30, "n_ref": 38, "n_ref_all": 50, "n_fig": 7, "n_tab": 1, "L_tab": 410, "n_element_tab": 47, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 1016, "n_element_tab_1": 66, "formula_len_all": 299, "formula_len_all_1": 300, "len_all": 111756, "len_all_1": 65996, "len_abs": 1088, "len_title": 76, "len_sents": 37138, "len_sents_1": 31455, "n_sents": 229, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 28, "L_abs": 1095, "title": "On Gradient-Weight Alignment", "abs": "Evaluating the performance of deep networks against unseen validation data is a crucial step to measure generalization performance.\nHowever, ostensibly neither the training nor validation and test data are ever sufficiently extensive to replicate real-world application.\nThis works advocates for a change of perspective for evaluating performance of deep networks.\nInstead of evaluating against unseen validation data, we propose to rather capture when the model starts to prioritize learning unnecessary or even detrimental specifics of training data instead of general patterns. \nWhile this has been challenging to theoretically derive, we propose *gradient-weight alignment* as an empirical metric to determine performance on unseen data from training information alone.\nOur performance measure is efficient and widely applicable, closely tracking validation accuracy during training.\nIt connects model performance to individual training samples, enabling its use not only for assessing generalization and as an early stopping criterion, but also for offering insights into training dynamics.", "keywords": [ "generalization", "gradients", "alignment", "influence", "memorization" ], "rating_list": [ 1, 5, 3, 5 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "nT2u0M0nf8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Linh Duy Tran", "gender": "Male", "institution": "Viettel Group", "country": "VN", "position": "Researcher" }, { "name": "Luc Nguyen", "gender": "Male", "institution": "Viettel", "country": "VN", "position": "Researcher" }, { "name": "Minh Nguyen Hoang", "gender": "Male", "institution": "Hanoi University of Science and Technology", "country": "VN", "position": "MS student" }, { "name": "Rachel Teo", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Tan Minh Nguyen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Viet Dung Nguyen", "gender": "Male", "institution": "Hanoi University of Science and Technology", "country": "VN", "position": "MS student" } ], "n_formula": 34, "n_formula_1": 16, "n_ref_uni": 53, "n_ref": 111, "n_ref_all": 124, "n_fig": 5, "n_tab": 18, "L_tab": 3033, "n_element_tab": 404, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 2331, "n_element_tab_1": 295, "formula_len_all": 2557, "formula_len_all_1": 1273, "len_all": 189462, "len_all_1": 61499, "len_abs": 2832, "len_title": 89, "len_sents": 48724, "len_sents_1": 25199, "n_sents": 407, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 1664, "title": "CAMEx: Curvature-aware Merging of Experts", "abs": "Existing methods for merging experts during model training and fine-tuning predominantly rely on Euclidean geometry, which assumes a flat parameter space. This assumption can limit the model's generalization ability, especially during the pre-training phase, where the parameter manifold might exhibit more complex curvature. Curvature-aware merging methods typically require additional information and computational resources to approximate the Fisher Information Matrix, adding memory overhead. In this paper, we introduce CAMEx (Curvature-Aware Merging of Experts), a novel expert merging protocol that incorporates natural gradients to account for the non-Euclidean curvature of the parameter manifold. By leveraging natural gradients, CAMEx adapts more effectively to the structure of the parameter space, improving alignment between model updates and the manifold's geometry. This approach enhances both pre-training and fine-tuning, resulting in better optimization trajectories and improved generalization without the substantial memory overhead typically associated with curvature-aware methods. Our contributions are threefold: (1) CAMEx significantly outperforms traditional Euclidean-based expert merging techniques across various natural language processing tasks, leading to enhanced performance during pre-training and fine-tuning; (2) we introduce a dynamic merging architecture that optimizes resource utilization, achieving high performance while reducing computational costs, facilitating efficient scaling of large language models; and (3) we provide both theoretical and empirical evidence to demonstrate the efficiency of our proposed method.", "keywords": [ "Sparse Mixture-of-Experts", "efficiency", "expert merging" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "nSYycd5tEC", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Junze Deng", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Ness Shroff", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Full Professor" }, { "name": "Peizhong Ju", "gender": "Male", "institution": "University of Kentucky", "country": "US", "position": "Assistant Professor" }, { "name": "Qinhang Wu", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Sen Lin", "gender": "unknown", "institution": "University of Houston", "country": "US", "position": "Assistant Professor" }, { "name": "Yingbin Liang", "gender": "Female", "institution": "The Ohio State University", "country": "US", "position": "Professor" } ], "n_formula": 106, "n_formula_1": 24, "n_ref_uni": 30, "n_ref": 43, "n_ref_all": 128, "n_fig": 3, "n_tab": 3, "L_tab": 654, "n_element_tab": 84, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18096, "formula_len_all_1": 2603, "len_all": 201788, "len_all_1": 64096, "len_abs": 2386, "len_title": 143, "len_sents": 58009, "len_sents_1": 29653, "n_sents": 640, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1993, "title": "Replay concurrently or sequentially? A theoretical perspective on replay in continual learning", "abs": "Replay-based methods have shown superior performance to address catastrophic forgetting in continual learning (CL), where a subset of past data is stored and generally replayed together with new data in current task learning. While seemingly natural, it is questionable, though rarely questioned, if such a concurrent replay strategy is always the right way for replay in CL. Inspired by the fact in human learning that revisiting very different courses sequentially before final exams is more effective for students, an interesting open question to ask is whether a sequential replay can benefit CL more compared to a standard concurrent replay. However, answering this question is highly nontrivial considering a major lack of theoretical understanding in replay-based CL methods. To this end, we investigate CL in overparameterized linear models and provide a comprehensive theoretical analysis to compare two replay schemes: 1) Concurrent Replay, where the model is trained on replay data and new data concurrently; 2) Sequential Replay, where the model is trained first on new data and then sequentially on replay data for each old task. By characterizing the explicit form of forgetting and generalization error, we show in theory that sequential replay tends to outperform concurrent replay when tasks are less similar, which is corroborated by our simulations in linear models. More importantly, our results inspire a novel design of a hybrid replay method, where only replay data of similar tasks are used concurrently with the current data and dissimilar tasks are sequentially revisited using their replay data. As depicted in our experiments on real datasets using deep neural networks, such a hybrid replay method improves the performance of standard concurrent replay by leveraging sequential replay for dissimilar tasks. By providing the first comprehensive theoretical analysis on replay, our work has great potentials to open up more principled designs for replay-based CL.", "keywords": [ "Continual Learning", "replay-based methods", "catastrophic forgetting" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "nSFVJkWYhr", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Mihaela van der Schaar", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Tim Verdonck", "gender": "Male", "institution": "Universiteit Antwerpen", "country": "BE", "position": "Full Professor" }, { "name": "Toon Vanderschueren", "gender": "Male", "institution": "Telenet", "country": "BE", "position": "Researcher" }, { "name": "Wouter Verbeke", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 3, "n_ref_uni": 58, "n_ref": 91, "n_ref_all": 128, "n_fig": 13, "n_tab": 14, "L_tab": 5584, "n_element_tab": 854, "n_fig_1": 8, "n_tab_1": 9, "L_tab_1": 3884, "n_element_tab_1": 576, "formula_len_all": 646, "formula_len_all_1": 73, "len_all": 199298, "len_all_1": 72660, "len_abs": 1293, "len_title": 109, "len_sents": 55504, "len_sents_1": 28110, "n_sents": 508, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1109, "title": "AutoCATE: End-to-End, Automated Treatment Effect Estimation", "abs": "Accurate estimation of heterogeneous treatment effects is critical in domains such as healthcare, economics, and education. While machine learning (ML) has led to significant advances in estimating conditional average treatment effects (CATE), real-world adoption of these methods remains limited due to the complexity of implementing, tuning, and validating them. To this end, we advocate for a more holistic view on the development of ML pipelines for CATE estimation through automated, end-to-end protocols. We formalize the search for an optimal pipeline as a counterfactual Combined Algorithm Selection and Hyperparameter optimization (CASH) problem. We introduce \\texttt{AutoCATE}, the first automated solution tailored for CATE estimation that addresses this problem based on protocols for evaluation, estimation, and ensembling. Our experiments show how AutoCATE allows for comparing different protocols, with the final configuration outperforming common strategies. We provide AutoCATE as an open-source software package to help practitioners and researchers develop ML pipelines for CATE estimation.", "keywords": [ "Treatment Effect Estimation", "Causal Inference", "AutoML" ], "rating_list": [ 1, 5, 8, 8, 3 ], "soundness_list": [ 2, 2, 4, 2, 3 ], "presentation_list": [ 1, 2, 4, 3, 3 ], "contribution_list": [ 1, 2, 4, 3, 1 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "nS2DBNydCC", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "benyou wang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Harry Yang", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Hengchao Chen", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Qiang Sun", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Associate Professor" }, { "name": "Song Dingjie", "gender": "Male", "institution": "Lehigh University", "country": "US", "position": "PhD student" }, { "name": "Ye xin Liu", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Yuxuan Zhang", "gender": "unknown", "institution": "University of British Columbia", "country": "CA", "position": "Researcher" }, { "name": "LitaoGuo", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Ser-Nam Lim", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Associate Professor" }, { "name": "Fang XiangHong", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 34, "n_formula_1": 18, "n_ref_uni": 38, "n_ref": 101, "n_ref_all": 152, "n_fig": 20, "n_tab": 3, "L_tab": 1621, "n_element_tab": 256, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4077, "formula_len_all_1": 1851, "len_all": 153777, "len_all_1": 66024, "len_abs": 3159, "len_title": 92, "len_sents": 49505, "len_sents_1": 29066, "n_sents": 369, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1133, "title": "Vector Quantization By Distribution Matching", "abs": "The success of autoregressive models largely depends on the effectiveness of vector quantization, a technique that compresses and discretizes continuous features by mapping them to the nearest code vectors within a learnable codebook. Two critical issues in existing vector quantization methods are training instability and codebook collapse. Training instability arises from the gradient gap during both forward and backward gradient propagation, especially in the presence of significant quantization errors, while codebook collapse occurs when only a small subset of code vectors are utilized during training.\nA closer examination of these issues reveals that they are primarily driven by a mismatch between the distributions of the features and code vectors, leading to unrepresentative code vectors and significant data information loss during compression. To address this, we employ the Wasserstein distance to align these two distributions, achieving near 100\\% codebook utilization and significantly reducing the quantization error. Both empirical and theoretical analyses validate the effectiveness of the proposed approach.", "keywords": [ "Vector Quantization", "Distribution Matching", "Criterion Triple", "Wasserstein Distance" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "nRgGCnw8eZ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Bai dongdong", "gender": "Male", "institution": "China RongTong Artificial Intelligence Research Center", "country": "", "position": "Researcher" }, { "name": "FU Yanchang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Junge Zhang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "Full Professor" }, { "name": "Kaiqi Huang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "Professor" }, { "name": "Pei Xu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Intern" }, { "name": "Zhaolingyun", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "MS student" } ], "n_formula": 24, "n_formula_1": 21, "n_ref_uni": 18, "n_ref": 40, "n_ref_all": 52, "n_fig": 5, "n_tab": 7, "L_tab": 3250, "n_element_tab": 264, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 269, "n_element_tab_1": 52, "formula_len_all": 1765, "formula_len_all_1": 1830, "len_all": 125024, "len_all_1": 68857, "len_abs": 833, "len_title": 124, "len_sents": 44772, "len_sents_1": 32147, "n_sents": 347, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 880, "title": "KrwEmd: Revising the Imperfect Recall Abstraction from Forgetting Everything", "abs": "Excessive abstraction is a serious issue in solving games with ordered signals\u2014a subset of imperfect information games, caused by extreme implementations of imperfect recall, which discard all historical information and, as a result, negatively impact AI performance. This paper presents KrwEmd, the first practical algorithm designed to address this issue. We first introduce the k-recall winrate feature, which not only qualitatively distinguishes signal infosets by leveraging future and, more importantly, historical game information, but also quantitatively reflects their similarity. We then build on this by developing the KrwEmd algorithm, which cluster signal infosets using Earth Mover\u2019s Distance to assess discrepancies between their features. Experimental results demonstrate that KrwEmd significantly enhances AI gameplay performance compared to existing algorithms.", "keywords": [ "game theory", "imperfect-information games", "games with ordered signals", "computer poker", "imperfect-recall abstraction", "unsupervised learning" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 1, 2 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 2, 2, 3 ] }, { "paper_id": "nRHD9fAj10", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Jiajun He", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Liangchen Li", "gender": "Male", "institution": "Independent Researcher", "country": "CN", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 12, "n_ref_uni": 44, "n_ref": 147, "n_ref_all": 204, "n_fig": 22, "n_tab": 5, "L_tab": 3579, "n_element_tab": 207, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 2156, "n_element_tab_1": 92, "formula_len_all": 2032, "formula_len_all_1": 1029, "len_all": 203979, "len_all_1": 71171, "len_abs": 1672, "len_title": 80, "len_sents": 68163, "len_sents_1": 31810, "n_sents": 581, "n_sents_1": 259, "venue": "ICLR2025", "decision": "withdraw", "L_title": 32, "L_abs": 1688, "title": "Bidirectional Consistency Models", "abs": "Diffusion models (DMs) are capable of generating remarkably high-quality samples by iteratively denoising a random vector, a process that corresponds to moving along the probability flow ordinary differential equation (PF ODE).\nInterestingly, DMs can also invert an input image to noise by moving backward along the PF ODE, a key operation for downstream tasks such as interpolation and image editing. \nHowever, the iterative nature of this process restricts its speed, hindering its broader application.\nRecently, Consistency Models (CMs) have emerged to address this challenge by approximating the integral of the PF ODE, largely reducing the number of iterations.\nYet, the absence of an explicit ODE solver complicates the inversion process. \n To resolve this, we introduce Bidirectional Consistency Model (BCM), which learns a *single* neural network that enables both *forward and backward* traversal along the PF ODE, efficiently unifying generation and inversion tasks within one framework.\nWe can train BCM from scratch or tune it using a pre-trained consistency model, which reduces the training cost and increases scalability.\n We demonstrate that BCM enables one-step generation and inversion while also allowing the use of additional steps to enhance generation quality or reduce reconstruction error.\nWe further showcase BCM's capability in downstream tasks, such as interpolation, inpainting, and blind restoration of compressed images.\nNotably, when the number of function evaluations (NFE) is constrained, BCM surpasses domain-specific restoration methods, such as I$^2$SB and Palette, in a fully zero-shot manner, offering an efficient alternative for inversion problems.", "keywords": [ "Consistency Models", "Diffusion Models", "Image Generation" ], "rating_list": [ 5, 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 3, 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 4, 3 ], "confidence_list": [ 4, 3, 5, 3, 3 ] }, { "paper_id": "nRD5TriJ0O", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Chen Ye", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Shengzhong Zhang", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "Assistant Professor" }, { "name": "Wenjie Yang", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Intern" }, { "name": "Zengfeng Huang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 36, "n_formula_1": 20, "n_ref_uni": 49, "n_ref": 98, "n_ref_all": 119, "n_fig": 7, "n_tab": 5, "L_tab": 5210, "n_element_tab": 259, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 1949, "n_element_tab_1": 111, "formula_len_all": 2559, "formula_len_all_1": 895, "len_all": 169144, "len_all_1": 63087, "len_abs": 1296, "len_title": 137, "len_sents": 35768, "len_sents_1": 26328, "n_sents": 358, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1314, "title": "Know Your Neighbors: Subgraph Importance Sampling for Heterophilic Graph Active Learning", "abs": "Graph neural networks (GNNs) have shown superiority in various data mining tasks but rely heavily on extensively labeled nodes. To improve the training efficiency and select the most valuable nodes as the training set, graph active learning (GAL) has gained much attention. However, previous GAL methods are designed for homophilic graphs, and their effectiveness on heterophilic graphs is less examined. In this paper, we study active learning on heterophilic graphs, where nodes with the same labels are less likely to be connected. We are surprised to find that *previous GAL methods fail to outperform the naive random sampling on heterophilic graphs*. Through an insightful investigation, we find that previous GAL-selected training sets imply homophily even on heterophilic graphs, leading to their defectiveness. To address this issue, we propose the principle of *``Know Your Neighbors''* and design an active learning algorithm KyN specifically for heterophilic graphs. The primary idea of KyN is to let GNNs receive a correct homophily distribution by labeling nodes along with their neighbors. We build KyN based on subgraph sampling with probabilities proportional to $\\ell_1$ Lewis weights, which has a solid theoretical guarantee. The effectiveness of KyN is evaluated on various real-world datasets.", "keywords": [ "Heterophilic Graph", "Active Learning" ], "rating_list": [ 6, 5, 1, 5, 6 ], "soundness_list": [ 3, 3, 1, 2, 3 ], "presentation_list": [ 3, 1, 2, 2, 4 ], "contribution_list": [ 2, 2, 1, 2, 3 ], "confidence_list": [ 4, 3, 5, 5, 2 ] }, { "paper_id": "nR2DHRxWS2", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Mohammad Rasool Izadi", "gender": "unknown", "institution": "Bose", "country": "US", "position": "Engineer" }, { "name": "Shuo Zhang", "gender": "Male", "institution": "Bose", "country": "", "position": "Principal Researcher" }, { "name": "Tornike Karchkhadze", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 6, "n_ref_uni": 50, "n_ref": 104, "n_ref_all": 115, "n_fig": 2, "n_tab": 3, "L_tab": 787, "n_element_tab": 50, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 886, "n_element_tab_1": 53, "formula_len_all": 2099, "formula_len_all_1": 357, "len_all": 143720, "len_all_1": 80490, "len_abs": 726, "len_title": 113, "len_sents": 42308, "len_sents_1": 26583, "n_sents": 298, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 727, "title": "Improving Source Extraction with Diffusion and Consistency Models", "abs": "In this work, we demonstrate the integration of a score-matching diffusion model into a deterministic architecture for time-domain musical source extraction, resulting in enhanced audio quality. To address the typically slow iterative sampling process of diffusion models, we apply consistency distillation and reduce the sampling process to a single step, achieving performance comparable to that of diffusion models, and with two or more steps, even surpassing them. Trained on the Slakh2100 dataset for four instruments (bass, drums, guitar, and piano), our model shows significant improvements across objective metrics compared to baseline methods. Sound examples are available at https://consistency-separation.github.io/.", "keywords": [ "source extraction", "consistency models", "score-matching diffusion" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "nR0n4R1Ck2", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Sahar Rajabi", "gender": "Female", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Sirisha Rambhatla", "gender": "Female", "institution": "University of Waterloo", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 52, "n_formula_1": 20, "n_ref_uni": 32, "n_ref": 70, "n_ref_all": 104, "n_fig": 8, "n_tab": 14, "L_tab": 4011, "n_element_tab": 528, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 2046, "n_element_tab_1": 137, "formula_len_all": 4458, "formula_len_all_1": 1276, "len_all": 139263, "len_all_1": 57878, "len_abs": 1417, "len_title": 144, "len_sents": 42759, "len_sents_1": 25850, "n_sents": 317, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1348, "title": "SubTrack your Grad: Gradient Subspace Tracking for Memory-Efficient LLM Training and Fine-Tuning", "abs": "Training and fine-tuning Large Language Models (LLMs) demand significant computational resources and time due to their large model sizes and optimizer states. To mitigate these challenges and improve accessibility, several memory-efficient methods have been developed. Methods such as Low-Rank Adaptation (LoRA) optimize model weights within a low-rank subspace, while Gradient Low-Rank Projection (GaLore) projects gradients into a lower-dimensional space to decrease memory footprint. In this paper, we propose Gradient Subspace Tracking (SubTrack-Grad), a method that confines optimization to a compact core subspace of the gradient matrices and dynamically tracks its changes using the geometry of Grassmannian manifolds. SubTrack-Grad efficiently updates its subspace estimation by leveraging estimation errors and previously identified subspaces. Our results demonstrate that even with rank-1 updates to the underlying subspace, SubTrack-Grad achieves comparable or superior performance to GaLore, while reducing runtime by approx. 15% on an average and up to 20.57% on some datasets. Furthermore, SubTrack-Grad exhibits only a minimal runtime increase compared to GaLore when the update frequency is increased, while controlling the extent of changes via rank-1 updates, allows more frequent updates without negatively impacting convergence.", "keywords": [ "large language models", "memory-efficient fine-tuning", "memory-efficient pre-training", "optimization", "subspace tracking", "gradient space", "low-rank optimization" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "nNiWRRj6r9", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Devdan Dey", "gender": "Male", "institution": "Indian Institute of Technology, Bombay", "country": "IN", "position": "PhD student" }, { "name": "Satyam Singh", "gender": "Male", "institution": "Indian Institute of Technology Bombay, Indian Institute of Technology, Bombay", "country": "IN", "position": "Postdoc" }, { "name": "Sujoy Bhore", "gender": "unknown", "institution": "Indian Institute of Technology Bombay, Indian Institute of Technology, Bombay", "country": "IN", "position": "Assistant Professor" } ], "n_formula": 31, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 69, "n_ref_all": 84, "n_fig": 9, "n_tab": 1, "L_tab": 10, "n_element_tab": 2, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3416, "formula_len_all_1": 772, "len_all": 189969, "len_all_1": 84184, "len_abs": 2105, "len_title": 104, "len_sents": 60767, "len_sents_1": 35702, "n_sents": 618, "n_sents_1": 333, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 2246, "title": "ONLINE EPSILON NET & PIERCING SET FOR GEOMETRIC CONCEPTS", "abs": "VC-dimension (Vapnik & Chervonenkis (1971)) and $\\varepsilon$-nets (Haussler & Welzl (1987)) are key concepts in Statistical Learning Theory. Intuitively, VC-dimension is a measure of the size of a class of sets. The famous $\\varepsilon$-net theorem, a fundamental result in Discrete Geometry, asserts that if the VC-dimension of a set system is bounded, then a small sample exists that intersects all sufficiently large sets.\n \n In online learning scenarios where data arrives sequentially, the VC-dimension helps bound the complexity of the set system, and $\\varepsilon$-nets ensure the selection of a small representative set. This sampling framework is crucial in various domains, including spatial data analysis, motion planning in dynamic environments, optimization of sensor networks, and feature extraction in computer vision, among others. Motivated by these applications, we study the online $\\varepsilon$-net problem for geometric concepts with bounded VC-dimension. While the offline version of this problem has been extensively studied, surprisingly, there are no known theoretical results for the online version to date. We present the first deterministic online algorithm with an optimal competitive ratio for intervals in $\\mathbb{R}$. Next, we give a randomized online algorithm with a near-optimal competitive ratio for axis-aligned boxes in $\\mathbb{R}^d$, for $d\\le 3$. Furthermore, we introduce a novel technique to analyze similar-sized objects of constant description complexity in $\\mathbb{R}^d$, which may be of independent interest. \n \n Next, we focus on the continuous version of this problem (called online piercing set), where ranges of the set system are geometric concepts in $\\mathbb{R}^d$ arriving in an online manner, but the universe is the entire ambient space, and the objective is to choose a small sample that intersects all the ranges. Although online piercing set is a very well-studied problem in the literature, to our surprise, very few works have addressed generic geometric concepts without any assumption about the sizes. We advance this field by proposing asymptotically optimal competitive deterministic algorithms for boxes and ellipsoids in $\\mathbb{R}^d$, for any $d\\in\\mathbb{N}$.", "keywords": [ "Theoretical machine learning", "VC-dimension", "Geometric sampling" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 2, 2, 2 ] }, { "paper_id": "nNYA7tcJSE", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Jong Chul Ye", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Serin Yang", "gender": "Female", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Taesung Kwon", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 21, "n_ref_uni": 28, "n_ref": 60, "n_ref_all": 84, "n_fig": 6, "n_tab": 8, "L_tab": 1484, "n_element_tab": 127, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 690, "n_element_tab_1": 89, "formula_len_all": 1177, "formula_len_all_1": 1155, "len_all": 105953, "len_all_1": 51411, "len_abs": 453, "len_title": 130, "len_sents": 30581, "len_sents_1": 22694, "n_sents": 219, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1471, "title": "ViBiDSampler: Enhancing Video Interpolation Using Bidirectional Diffusion Sampler", "abs": "Recent progress in large-scale text-to-video (T2V) and image-to-video (I2V) diffusion models has greatly enhanced video generation, especially in terms of keyframe interpolation. However, current image-to-video diffusion models, while powerful in generating videos from a single conditioning frame, need adaptation for two-frame (start \\& end) conditioned generation, which is essential for effective bounded interpolation. Unfortunately, existing approaches that fuse temporally forward and backward paths in parallel often suffer from off-manifold issues, leading to artifacts or requiring multiple iterative re-noising steps. In this work, we introduce a novel, bidirectional sampling strategy to address these off-manifold issues without requiring extensive re-noising or fine-tuning. Our method employs sequential sampling along both forward and backward paths, conditioned on the start and end frames, respectively, ensuring more coherent and on-manifold generation of intermediate frames. Additionally, we incorporate advanced guidance techniques, CFG++ and DDS, to further enhance the interpolation process. By integrating these, our method achieves state-of-the-art performance, efficiently generating high-quality, smooth videos between keyframes. On a single 3090 GPU, our method can interpolate 25 frames at 1024$\\times$576 resolution in just 195 seconds, establishing it as a leading solution for keyframe interpolation.\nProject page: https://vibid.github.io", "keywords": [ "Keyframe interpolation", "Stable video diffusion", "Bidirectional diffusion sampling" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "nNQmZGjEVe", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Hyungi Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Juho Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Seanie Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "CHAEYOON JANG", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 50, "n_ref": 93, "n_ref_all": 140, "n_fig": 10, "n_tab": 18, "L_tab": 3312, "n_element_tab": 398, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 1272, "n_element_tab_1": 140, "formula_len_all": 360, "formula_len_all_1": 325, "len_all": 209232, "len_all_1": 70701, "len_abs": 1077, "len_title": 122, "len_sents": 66293, "len_sents_1": 33174, "n_sents": 499, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1080, "title": "Calibrated Decision-Making through Large Language Model-Assisted Retrieval", "abs": "Recently, large language models (LLMs) have been increasingly used to support various decision-making tasks, assisting humans in making informed decisions. However, when LLMs confidently provide incorrect information, it can lead humans to make suboptimal decisions. To prevent LLMs from generating incorrect information on topics they are unsure of and to improve the accuracy of generated content, prior works have proposed Retrieval Augmented Generation (RAG), where external documents are referenced to generate responses. However, traditional RAG methods focus only on retrieving documents most relevant to the input query, without specifically aiming to ensure that the human user's decisions are well-calibrated. To address this limitation, we propose a novel retrieval method called Calibrated Retrieval-Augmented Generation (CalibRAG), which ensures that decisions informed by the retrieved documents are well-calibrated. Then we empirically validate that CalibRAG improves calibration performance as well as accuracy, compared to other baselines across various datasets.", "keywords": [ "calibration", "RAG", "LLMs" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 2 ] }, { "paper_id": "nM2kuesKpC", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Aditya Balu", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "Researcher" }, { "name": "Chao Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Md Zahid Hasan", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "PhD student" }, { "name": "Nastaran Saadati", "gender": "Female", "institution": "Iowa State University", "country": "US", "position": "PhD student" }, { "name": "Soumik Sarkar", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "Full Professor" }, { "name": "Zhanhong Jiang", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "Researcher" } ], "n_formula": 22, "n_formula_1": 22, "n_ref_uni": 47, "n_ref": 82, "n_ref_all": 95, "n_fig": 6, "n_tab": 2, "L_tab": 399, "n_element_tab": 33, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 214, "n_element_tab_1": 15, "formula_len_all": 1532, "formula_len_all_1": 1330, "len_all": 130517, "len_all_1": 69218, "len_abs": 1373, "len_title": 130, "len_sents": 36721, "len_sents_1": 32514, "n_sents": 264, "n_sents_1": 236, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1381, "title": "D2P2-SGD: Dynamically Differentially Private Projected Stochastic Gradient Descent", "abs": "Stochastic optimization is a key enabler in modern machine learning, producing effective models for various tasks. However, several researchers have shown that model parameters and gradient information are susceptible to privacy leakage. Although, Differentially Private SGD (DPSGD) addresses privacy concerns, its static noise mechanism impacts the error bounds for model performance. Additionally, with the exponential increase in model parameters, efficient learning of these models using stochastic optimizers has become more challenging. To address these concerns, we introduce the Dynamically Differentially Private Projected Stochastic Gradient Descent (D2P2-SGD) optimizer. In D2P2-SGD, we combine two important ideas: (i) dynamic differential privacy (DDP) with automatic gradient clipping and (ii) random projection with SGD, allowing dynamic adjustment of the tradeoff between utility and privacy of the model. It demonstrates provably tighter error bounds compared to DPSGD across different behavior (i.e. convex and non-convex) of the objective function. The theoretical analysis further suggests that DDP leads to better utility at the cost of privacy, while random projection enables more efficient model learning. Extensive experiments across diverse datasets show that D2P2-SGD significantly enhances accuracy while maintaining privacy. Our code is available here.", "keywords": [ "Differential privacy", "SGD", "random projection", "convergence", "stochastic optimization" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "nLlBLzPpeG", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Dit-Yan Yeung", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Chair Professor" }, { "name": "Tsz-Him Cheung", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 33, "n_ref": 59, "n_ref_all": 84, "n_fig": 10, "n_tab": 6, "L_tab": 2249, "n_element_tab": 310, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 50, "n_element_tab_1": 7, "formula_len_all": 727, "formula_len_all_1": 574, "len_all": 123974, "len_all_1": 59389, "len_abs": 1969, "len_title": 81, "len_sents": 38383, "len_sents_1": 28819, "n_sents": 304, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1135, "title": "AutoGenDA: Automated Generative Data Augmentation for Imbalanced Classifications", "abs": "Data augmentation is an approach to increasing the training dataset size for deep learning using synthetic data. Recent advancements in image generative models have unleashed the potential of synthesizing high-quality images in data augmentation. However, real-life datasets commonly follow an imbalanced class distribution, where some classes have fewer samples than others. Image generation models may, therefore, struggle to synthesize diverse images for less common classes that lack richness and diversity. To address this, we introduce an automated generative data augmentation method, AutoGenDA, to extract and transfer label-invariant changes across data classes through image captions and text-guided generative models. We also propose an automated search strategy to optimize the data augmentation process for each data class, leading to better generalization. Our experiments demonstrate the effectiveness of AutoGenDA in various object classification datasets. We improve the standard data augmentation baselines by up to 4.9\\% on Pascal VOC, Caltech101, MS-COCO, and LVIS under multiple imbalanced classification settings.", "keywords": [ "data augmentation", "machine learning" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "nKlcmBJpiH", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Elad Sarafian", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "Researcher" }, { "name": "Sarit Kraus", "gender": "Female", "institution": "Bar-Ilan University", "country": "IL", "position": "Full Professor" }, { "name": "Yedidya Kfir", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" }, { "name": "Yoram Louzoun", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "Full Professor" } ], "n_formula": 32, "n_formula_1": 11, "n_ref_uni": 30, "n_ref": 43, "n_ref_all": 61, "n_fig": 7, "n_tab": 5, "L_tab": 1637, "n_element_tab": 181, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 2450, "n_element_tab_1": 131, "formula_len_all": 3909, "formula_len_all_1": 1151, "len_all": 135927, "len_all_1": 51916, "len_abs": 1635, "len_title": 97, "len_sents": 33918, "len_sents_1": 21947, "n_sents": 256, "n_sents_1": 144, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1642, "title": "Optimistic Gradient Learning with Hessian Corrections for High-Dimensional Black-Box Optimization", "abs": "Black-box algorithms are designed to optimize functions without relying on their underlying analytical structure or gradient information, making them essential when gradients are inaccessible or difficult to compute. Traditional methods for solving black-box optimization (BBO) problems predominantly rely on non-parametric models and struggle to scale to large input spaces. Conversely, parametric methods that model the function with neural estimators and obtain gradient signals via backpropagation may suffer from significant gradient errors. A recent alternative, Explicit Gradient Learning (EGL), which directly learns the gradient using a first-order Taylor approximation, has demonstrated superior performance over both parametric and non-parametric methods. In this work, we propose two novel gradient learning variants to address the robustness challenges posed by high-dimensional, complex, and highly non-linear problems. Optimistic Gradient Learning (OGL) introduces a bias toward lower regions in the function landscape, while Higher-order Gradient Learning (HGL) incorporates second-order Taylor corrections to improve gradient accuracy. We combine these approaches into the unified OHGL algorithm, achieving state-of-the-art (SOTA) performance on the synthetic COCO suite. Additionally, we demonstrate OHGL\u2019s applicability to high-dimensional real-world machine learning (ML) tasks such as adversarial training and code generation. Our results highlight OHGL's ability to generate stronger candidates, offering a valuable tool for ML researchers and practitioners tackling high-dimensional, non-linear optimization challenges.", "keywords": [ "Black box optimization", "Derivative free optimization", "Gradient free optimization", "Large language model code geneation", "Adversarial training" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 2, 3, 2, 4 ] }, { "paper_id": "nKVYQOgD0q", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Jason Yuan", "gender": "unknown", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "MS student" }, { "name": "Jin Peng Zhou", "gender": "Male", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "PhD student" }, { "name": "Jingwen Gu", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Undergrad student" }, { "name": "Katie Luo", "gender": "Female", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Kilian Q Weinberger", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Professor" }, { "name": "Wen Sun", "gender": "unknown", "institution": "Cornell University", "country": "", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 16, "n_ref_uni": 38, "n_ref": 71, "n_ref_all": 97, "n_fig": 5, "n_tab": 12, "L_tab": 1999, "n_element_tab": 100, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1170, "n_element_tab_1": 96, "formula_len_all": 12876, "formula_len_all_1": 18300, "len_all": 264792, "len_all_1": 96624, "len_abs": 1069, "len_title": 82, "len_sents": 94798, "len_sents_1": 40603, "n_sents": 683, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 34, "L_abs": 1079, "title": "On Orchestrating Personalized LLMs", "abs": "This paper presents a novel approach to aligning large language models (LLMs) with individual human preferences, sometimes referred to as Reinforcement Learning from *Personalized* Human Feedback (RLPHF). Given stated preferences along multiple dimensions, such as helpfulness, conciseness, or humor, the goal is to create an LLM -- without completely re-training -- that best adheres to this specification. Starting from specialized expert LLMs, each trained for one such particular preference dimension, we propose a black-box method that merges their outputs on a per-token level. We train a lightweight Preference Control Model (PCM) that dynamically translates the preference description and current context into next-token prediction weights. By combining the expert models' outputs at the token level, our approach dynamically generates text that optimizes the given preference. Empirical tests show that our method matches or surpasses existing preference merging techniques, providing a scalable, efficient alternative to fine-tuning LLMs for individual personalization.", "keywords": [ "Personalized LLM", "Reinforcement Learning from Personalized Feedback" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 4, 4, 5, 2 ] }, { "paper_id": "nKSkM5h2VN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Cheng-Han Yeh", "gender": "Male", "institution": "Academia Sinica", "country": "TW", "position": "Researcher" }, { "name": "Chun-Shien Lu", "gender": "Male", "institution": "Academia Sinica", "country": "TW", "position": "Full Professor" }, { "name": "Kuanchun Yu", "gender": "unknown", "institution": "Academia Sinica", "country": "TW", "position": "Research Assistant" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 41, "n_ref": 87, "n_ref_all": 140, "n_fig": 6, "n_tab": 3, "L_tab": 426, "n_element_tab": 43, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 897, "n_element_tab_1": 29, "formula_len_all": 646, "formula_len_all_1": 645, "len_all": 122144, "len_all_1": 69546, "len_abs": 1094, "len_title": 134, "len_sents": 34771, "len_sents_1": 31555, "n_sents": 256, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1099, "title": "Test-Time Adversarial Defense with Opposite Adversarial Path and high Attack time cost", "abs": "Deep learning models are known to be vulnerable to adversarial attacks by injecting sophisticated designed perturbations to input data. Training-time defenses still exhibit a significant performance gap between natural accuracy and robust accuracy. In this paper, we investigate a new test-time adversarial defense method via diffusion-based recovery along opposite adversarial paths (OAPs). We present a purifier that can be plugged into a pre-trained model to resist adversarial attacks. Different from prior arts, the key idea is excessive denoising or purification by integrating the opposite adversarial direction with reverse diffusion to push the input image further toward the opposite adversarial direction. For the first time, we also exemplify the pitfall of conducting AutoAttack (Rand) for diffusion-based defense methods. Through the lens of time complexity, we examine the trade-off between the effectiveness of adaptive attack and its computation complexity against our defense. Experimental evaluation along with time cost analysis verifies the effectiveness of the proposed method.", "keywords": [ "Adversarial Attack", "Adversarial Defense", "Diffusion Models", "Purifier", "Robustness", "Security" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "nIEjY4a2Lf", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Ally Yalei Du", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Undergrad student" }, { "name": "Ruosong Wang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Assistant Professor" }, { "name": "lin Yang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "Assistant Professor" } ], "n_formula": 41, "n_formula_1": 15, "n_ref_uni": 36, "n_ref": 78, "n_ref_all": 82, "n_fig": 1, "n_tab": 5, "L_tab": 1164, "n_element_tab": 75, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 851, "n_element_tab_1": 28, "formula_len_all": 5693, "formula_len_all_1": 2487, "len_all": 176523, "len_all_1": 82583, "len_abs": 1269, "len_title": 151, "len_sents": 50709, "len_sents_1": 33559, "n_sents": 488, "n_sents_1": 287, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 1346, "title": "Misspecified $Q$-Learning with Sparse Linear Function Approximation: Tight Bounds on Approximation Error", "abs": "The recent work by Dong and Yang (2023) showed for misspecified sparse linear bandits, one can obtain an $O(\\epsilon)$-optimal policy using a polynomial number of samples when the sparsity is a constant, where $\\epsilon$ is the misspecification error. This result is in sharp contrast to misspecified linear bandits without sparsity, which require an exponential number of samples to get the same guarantee. In order to study whether the analog result is possible in the reinforcement learning setting, we consider the following problem: assuming the optimal $Q$-function is a $d$-dimensional linear function with sparsity $k$ and misspecification error $\\epsilon$, whether we can obtain an $O(\\epsilon)$-optimal policy using number of samples polynomially in the feature dimension $d$. We first demonstrate why the standard approach based on Bellman backup or the existing optimistic value function elimination approach such as OLIVE (Jiang et al., 2017) achieves suboptimal guarantees for this problem. We then design a novel elimination-based algorithm to show one can obtain an \n$O(H\\epsilon)$-optimal policy with sample complexity polynomially in the feature dimension $d$ and planning horizon $H$. Lastly, we complement our upper bound with an $\\tilde \\Omega(H\\epsilon)$ suboptimality lower bound, giving a complete picture of this problem.", "keywords": [ "misspecification error", "reinforcement learning theory", "sample complexity", "sparsity" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "nIBmwm7ixo", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Abhinav Anand", "gender": "Male", "institution": "IIT Kanpur, Indian Institute of Technology, Kanpur", "country": "IN", "position": "PhD student" }, { "name": "Amitangshu Pal", "gender": "Male", "institution": "IIT Kanpur, Indian Institute of Technology, Kanpur", "country": "IN", "position": "Assistant Professor" }, { "name": "Subrahmanya Swamy Peruru", "gender": "Male", "institution": "IIT Kanpur", "country": "", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 36, "n_ref": 70, "n_ref_all": 89, "n_fig": 6, "n_tab": 3, "L_tab": 4219, "n_element_tab": 447, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 711, "n_element_tab_1": 54, "formula_len_all": 641, "formula_len_all_1": 386, "len_all": 133708, "len_all_1": 62996, "len_abs": 1957, "len_title": 127, "len_sents": 46458, "len_sents_1": 31170, "n_sents": 321, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 767, "title": "VColRL: Learn to Solve the Vertex Coloring Problem Using Reinforcement Learning", "abs": "We present VColRL, a reinforcement learning framework designed to solve the vertex coloring problem (VCP), where the objective is to assign colors to the vertices of a graph with the minimum number of colors, such that no two adjacent vertices share the same color. The framework is built on a novel Markov Decision Process (MDP) configuration to effectively capture the dynamics of the VCP, developed after evaluating various MDP configurations. Our experimental results demonstrate that VColRL achieves competitive performance in terms of using fewer colors as compared to advanced mathematical solvers and other metaheuristic approaches while being significantly faster. Additionally, our results show that VColRL generalizes well across different types of graphs.", "keywords": [ "Vertex Coloring Problem", "Markov Decision Process", "Reinforcement Learning", "Graph Neural Networks", "Combinatorial Optimization" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 3, 5 ] }, { "paper_id": "nHmaQf2wJC", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Fengwei Tian", "gender": "unknown", "institution": "University of Arizona", "country": "US", "position": "PhD student" }, { "name": "Noel Teku", "gender": "unknown", "institution": "University of Arizona", "country": "US", "position": "PhD student" }, { "name": "Payel Bhattacharjee", "gender": "Female", "institution": "University of Arizona", "country": "US", "position": "PhD student" }, { "name": "Ravi Tandon", "gender": "unknown", "institution": "University of Arizona", "country": "", "position": "Associate Professor" }, { "name": "Souradip Chakraborty", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Amrit Bedi", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" } ], "n_formula": 31, "n_formula_1": 31, "n_ref_uni": 16, "n_ref": 26, "n_ref_all": 40, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 17418, "formula_len_all_1": 16912, "len_all": 168582, "len_all_1": 130901, "len_abs": 2653, "len_title": 103, "len_sents": 43385, "len_sents_1": 39666, "n_sents": 272, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 2344, "title": "Aligning Large Language Models With Preference Privacy", "abs": "Alignment is a crucial part in the implementation pipeline of Large Language Models (LLMs) that utilizes human feedback to ensure that LLMs adhere to human values and societal norms. This introduces privacy threats associated with the identity and preferences of the labelers responsible for creating the human feedback data. Several recent works have explored using differential privacy (DP) as a notion to protect the privacy of human labeled data; primarily relying on DP-SGD based solutions, which privatize the gradients during fine-tuning and alignment. Human preferences, however are only associated with the labels of the (prompt, response) tuples; therefore DP-SGD based approaches can be superfluous, providing more privacy than necessary and can degrade model utility. In this work, we focus on the problem of aligning LLMs with preference level privacy, which only preserve the privacy of preferences provided by humans. We build and expand upon the concept of label DP for this problem, and present a series of increasingly sophisticated, yet practical privacy preserving mechanisms for alignment. Specifically, starting from a standard randomized response (RR) mechanism which randomly flips human preferences, and it's corresponding \\textit{unbiased} RR mechanism (which ensures an unbiased loss during alignment), we propose a new mechanism, PROPS (PROgressively Private Self-alignment). PROPS works in multiple stages as follows: in each stage, the privately trained and partially aligned model from the previous stage to act as a labeler for the training data for the next stage and combine it with RR which is repeated across multiple stages. Motivation for PROPS comes from the following critical observations: a) learning to label correct preferences might be an easier problem than generating responsible content; b) progressively combining RR with partially aligned models for labeling preferences significantly reduces the amount of necessary perturbation needed for privacy and also shows the potential of possibly reducing the number of human labeled preference samples. We present proof-of-concept experiments that demonstrate the feasibility and effectiveness of our proposed approach and show that preference privacy based alignment can still attain a comparable utility to their non-privately aligned counterparts.", "keywords": [ "LLMs", "Differential Privacy", "Alignment" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 2 ] }, { "paper_id": "nHenODN9je", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Akash Srivastava", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Guangxuan Xu", "gender": "Male", "institution": "Red Hat. Inc", "country": "US", "position": "Researcher" }, { "name": "Kai Xu", "gender": "Male", "institution": "MIT-IBM Watson AI Lab", "country": "", "position": "Research scientist" }, { "name": "Shivchander Sudalairaj", "gender": "Male", "institution": "MIT-IBM Watson AI Lab", "country": "US", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 3, "n_ref_uni": 31, "n_ref": 71, "n_ref_all": 107, "n_fig": 12, "n_tab": 6, "L_tab": 2464, "n_element_tab": 335, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 4566, "n_element_tab_1": 150, "formula_len_all": 842, "formula_len_all_1": 307, "len_all": 193395, "len_all_1": 57438, "len_abs": 1736, "len_title": 102, "len_sents": 53273, "len_sents_1": 25419, "n_sents": 393, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1764, "title": "Preference Data Annotation with Guided Density Ratios", "abs": "Preference tuning of large language models (LLMs) relies on high-quality human preference data, which is often expensive and time-consuming to gather. While existing methods can use trained reward models or proprietary model as judges for preference annotation, they have notable drawbacks: training reward models remain dependent on initial human data, and using proprietary model imposes license restrictions that inhibits commercial usage. In this paper, we introduce Guided Density Ratio, a training-free and highly effective method that leverages off-the-shelf LLMs for preference data annotation. Our approach uses the log-density ratio between a better-aligned LLM and a less aligned LLM as a reward signal. We explores 221 different LLMs pairs and empirically demonstrate that increasing the performance gap between paired LLMs correlates with better reward generalization. Furthermore, we show that tailoring the density ratio reward function with specific criteria and preference exemplars enhances performance across domains and within target areas.\n\nIn our experiment using density ratio from a pair of Mistral-7B models, Guided Density Ratio achieves a RewardBench score of 82.6, outperforming the best trained reward functions from same model class and demonstrating competitive performance against SoTA models in Safety (91.0) and Reasoning (88.0) domains. We use Guided Density Ratio to annotate an on-policy preference dataset with which we preference tune \\textit{Llama-3-8B-Instruct} with SimPO. Using reward signals from two relatively weak models, our approach pushes Llama-3-8B to achieve a 37.4\\% ($+$15.1\\%) win rate on ArenaHard and a 40.7\\% ($+$17.8\\%) win rate on Length-Controlled AlpacaEval 2.0, along with a score of 8.0 on MT-Bench.", "keywords": [ "RLHF", "preference training", "reward model", "human preference", "preference alignment" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 3, 1 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "nHdWdcoyAg", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Jeremias Traub", "gender": "unknown", "institution": "Deutsches Krebsforschungszentrum (DKFZ)", "country": "DE", "position": "Researcher" }, { "name": "Kim-Celine Kahl", "gender": "Female", "institution": "Ruprecht-Karls-Universit\u00e4t Heidelberg", "country": "DE", "position": "PhD student" }, { "name": "Klaus Maier-Hein", "gender": "Male", "institution": "German Cancer Research Center", "country": "DE", "position": "Full Professor" }, { "name": "Lena Maier-Hein", "gender": "Female", "institution": "Deutsches Krebsforschungszentrum", "country": "DE", "position": "Full Professor" }, { "name": "Paul F Jaeger", "gender": "unknown", "institution": "Google", "country": "CA", "position": "Researcher" }, { "name": "Selen Erkan", "gender": "unknown", "institution": "Deutsches Krebsforschungszentrum", "country": "DE", "position": "Researcher" }, { "name": "Carsten Tim L\u00fcth", "gender": "Male", "institution": "Ruprecht-Karls-Universit\u00e4t Heidelberg", "country": "DE", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 54, "n_ref_all": 91, "n_fig": 12, "n_tab": 20, "L_tab": 5758, "n_element_tab": 852, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 192, "n_element_tab_1": 40, "formula_len_all": 38, "formula_len_all_1": 0, "len_all": 202152, "len_all_1": 57015, "len_abs": 1831, "len_title": 129, "len_sents": 51339, "len_sents_1": 27902, "n_sents": 458, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1864, "title": "SURE-VQA: Systematic Understanding of Robustness Evaluation in Medical VQA Tasks", "abs": "Vision-Language Models (VLMs) have great potential in medical tasks, like Visual Question Answering (VQA), where they could act as interactive assistants for both patients and clinicians. Yet their robustness to distribution shifts on unseen data remains a critical concern for safe deployment. Evaluating such robustness requires a controlled experimental setup that allows for systematic insights into the model's behavior. However, we demonstrate that current setups fail to offer sufficiently thorough evaluations, limiting their ability to accurately assess model robustness.\nTo address this gap, our work introduces a novel framework, called SURE-VQA, centered around three key requirements to overcome the current pitfalls and systematically analyze the robustness of VLMs: 1) Since robustness on synthetic shifts does not necessarily translate to real-world shifts, robustness should be measured on real-world shifts that are inherent to the VQA data; 2) Traditional token-matching metrics often fail to capture underlying semantics, necessitating the use of large language models (LLMs) for more accurate semantic evaluation; 3) Model performance often lacks interpretability due to missing sanity baselines, thus meaningful baselines should be reported that allow assessing the multimodal impact on the VLM.\nTo demonstrate the relevance of this framework, we conduct a study on the robustness of various Parameter-Efficient Fine-Tuning (PEFT) methods across three medical datasets with four different types of distribution shifts. \nOur study reveals several important findings: 1) Sanity baselines that do not utilize image data can perform surprisingly well; 2) We confirm LoRA as the best-performing PEFT method; 3) No PEFT method consistently outperforms others in terms of robustness to shifts. Code is provided at https://github.com/KOFRJO/sure-vqa.", "keywords": [ "robustness", "VQA", "validation" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "nGiGXLnKhl", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Jifeng Dai", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Lewei Lu", "gender": "Male", "institution": "SenseTime", "country": "CN", "position": "Researcher" }, { "name": "Tong Lu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Xizhou Zhu", "gender": "unknown", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuchen Duan", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zhe Chen", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "WeiYun Wang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Wenhai Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 53, "n_ref": 102, "n_ref_all": 124, "n_fig": 5, "n_tab": 6, "L_tab": 2385, "n_element_tab": 279, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1189, "n_element_tab_1": 170, "formula_len_all": 859, "formula_len_all_1": 743, "len_all": 155209, "len_all_1": 61006, "len_abs": 1202, "len_title": 130, "len_sents": 39400, "len_sents_1": 27246, "n_sents": 300, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 82, "L_abs": 1253, "title": "Vision-RWKV: Efficient and Scalable Visual Perception with RWKV-Like Architectures", "abs": "Transformers have revolutionized computer vision and natural language processing, but their high computational complexity limits their application in high-resolution image processing and long-context analysis. This paper introduces Vision-RWKV (VRWKV), a model that builds upon the RWKV architecture from the NLP field with key modifications tailored specifically for vision tasks. Similar to the Vision Transformer (ViT), our model demonstrates robust global processing capabilities, efficiently handles sparse inputs like masked images, and can scale up to accommodate both large-scale parameters and extensive datasets. Its distinctive advantage is its reduced spatial aggregation complexity, enabling seamless processing of high-resolution images without the need for window operations. Our evaluations demonstrate that VRWKV surpasses ViT's performance in image classification and has significantly faster speeds and lower memory usage processing high-resolution inputs. In dense prediction tasks, it outperforms window-based models, maintaining comparable speeds. These results highlight VRWKV's potential as a more efficient alternative for visual perception tasks. Code and models are available at~\\url{https://github.com/OpenGVLab/Vision-RWKV}.", "keywords": [ "RWKV", "Visual Perception", "Linear Attention" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 4, 4 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "nGBJpY4rIJ", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Gaowen Liu", "gender": "Female", "institution": "Cisco Systems", "country": "US", "position": "Researcher" }, { "name": "Ling Liu", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Ramana Rao Kompella", "gender": "Male", "institution": "Cisco", "country": "US", "position": "Researcher" }, { "name": "Sihao Hu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Tiansheng Huang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 28, "n_ref": 52, "n_ref_all": 82, "n_fig": 7, "n_tab": 10, "L_tab": 7338, "n_element_tab": 471, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1765, "n_element_tab_1": 249, "formula_len_all": 300, "formula_len_all_1": 195, "len_all": 152680, "len_all_1": 61300, "len_abs": 1070, "len_title": 141, "len_sents": 44277, "len_sents_1": 27396, "n_sents": 286, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1182, "title": "Pok\u00e9LLMon: A Grounding and Reasoning Benchmark for Large Language Models in Pok\u00e9mon Battles", "abs": "Developing grounding techniques for LLMs poses two requirements for interactive environments, i.e., (i) the presence of rich knowledge beyond the scope of existing LLMs and (ii) the complexity of tasks that require strategic reasoning. Existing environments fail to meet both requirements due to their simplicity or reliance on commonsense knowledge already encoded in LLMs for interaction. In this paper, we present Pok\u00e9LLMon, a new benchmark enriched with fictional game knowledge and characterized by the intense, dynamic, and adversarial gameplay of Pok\u00e9mon battles, setting new challenges for the development of grounding and reasoning techniques in interactive environments. Empirical evaluations demonstrate that existing LLMs lack game knowledge and struggle in Pok\u00e9mon battles. We investigate grounding techniques that leverage game knowledge and self-play experience, and provide a thorough analysis of reasoning methods from a new perspective of action consistency. Additionally, we introduce higher-level reasoning challenges when playing against human players. The implementation of our benchmark is anonymously released at: https://anonymous.4open.science/r/PokeLLMon.", "keywords": [ "Large language models", "Pok\u00e9mon Battles", "Reasoning", "Grounding", "Interactive environment", "Benchmark" ], "rating_list": [ 6, 5, 3, 8, 3 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 4, 2 ], "contribution_list": [ 2, 2, 1, 3, 2 ], "confidence_list": [ 3, 4, 5, 4, 4 ] }, { "paper_id": "nFcgay1Yo9", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Jianglin Lu", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Yitian Zhang", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Yixuan Liu", "gender": "unknown", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Yun Fu", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 16, "n_ref_uni": 34, "n_ref": 74, "n_ref_all": 95, "n_fig": 7, "n_tab": 5, "L_tab": 5907, "n_element_tab": 203, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3199, "n_element_tab_1": 97, "formula_len_all": 984, "formula_len_all_1": 1016, "len_all": 135793, "len_all_1": 67225, "len_abs": 1030, "len_title": 80, "len_sents": 42263, "len_sents_1": 28376, "n_sents": 352, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 32, "L_abs": 1037, "title": "Scale-Free Graph-Language Models", "abs": "Graph-language models (GLMs) have shown great potential in graph-based semi-supervised learning. In GLMs, graph generation and text embedding are two key stages, which are typically achieved by inferring a latent graph and finetuning a language model, respectively. However, the former relies on artificial assumptions about the underlying edge distribution, while the latter requires sufficient data annotations. This paper introduces a novel GLM that addresses these two challenges sequentially by leveraging a well-grounded structural prior. Specifically, we explore an inherent nature of real edge distribution\u2014the scale-free property\u2014for graph generation. We unexpectedly reveal that this natural characteristic can be closely approximated by a simple k-nearest neighbor graph. By using this scale-free graph, we subsequently develop a graph-based pseudo-labeler to generate complementary supervision for text embedding. Extensive experiments validate our findings and highlight the potential of GLMs built on scale-free structures.", "keywords": [ "scale-free property", "language models", "k nearest neighbor graph" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "nFVsK3QLgs", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Biao Zhang", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Garrett Tanzer", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 21, "n_ref": 44, "n_ref_all": 61, "n_fig": 3, "n_tab": 3, "L_tab": 2673, "n_element_tab": 590, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 3117, "n_element_tab_1": 597, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 118030, "len_all_1": 66039, "len_abs": 943, "len_title": 132, "len_sents": 30370, "len_sents_1": 28557, "n_sents": 180, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 943, "title": "YouTube-SL-25: A Large-Scale, Open-Domain Multilingual Sign Language Parallel Corpus", "abs": "Even for better-studied sign languages like American Sign Language (ASL), data is the bottleneck for machine learning research. The situation is worse yet for the many other sign languages used by Deaf/Hard of Hearing communities around the world. In this paper, we present YouTube-SL-25, a large-scale, open-domain multilingual corpus of sign language videos with seemingly well-aligned captions drawn from YouTube. With >3000 hours of videos across >25 sign languages, YouTube-SL-25 is a) >3x the size of YouTube-ASL, b) the largest parallel sign language dataset to date, and c) the first or largest parallel dataset for many of its component languages. We provide baselines for sign-to-text tasks using a unified multilingual multitask model based on T5 and report scores on benchmarks across 4 sign languages. The results demonstrate that multilingual transfer benefits both higher- and lower-resource sign languages within YouTube-SL-25.", "keywords": [ "sign language translation", "sign language", "multilinguality", "data curation", "data auditing" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 4, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "nExUJBF5tR", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Daniel Soudry", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Associate Professor" }, { "name": "Yaniv Blumenfeld", "gender": "Male", "institution": "Technion, Technion", "country": "", "position": "MS student" } ], "n_formula": 54, "n_formula_1": 14, "n_ref_uni": 32, "n_ref": 91, "n_ref_all": 150, "n_fig": 15, "n_tab": 6, "L_tab": 1545, "n_element_tab": 316, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 524, "n_element_tab_1": 135, "formula_len_all": 4190, "formula_len_all_1": 1257, "len_all": 227196, "len_all_1": 67791, "len_abs": 1373, "len_title": 149, "len_sents": 92262, "len_sents_1": 33473, "n_sents": 646, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1382, "title": "De-biasing Diffusion: Data-Free FP8 Quantization of Text-to-Image Models with Billions of Parameters", "abs": "Diffusion neural networks have become the go-to solution for tasks involving automatic image generation, but the generation process is expensive in terms of memory, energy, and computational cost. Several works have aimed to reduce the cost via quantization to 8 bits or less. Despite that, half-precision (FP16) is still the default mode for large text-to-image generation models --- where reducing numerical precision becomes more challenging. In this work, we show that the reduction of quantization bias can be more important than the reduction in quantization (mean square) error. We propose a data-free method for model quantization, with the goal of producing images that are indistinguishable from those generated by large, full-precision diffusion models. We show that simple methods like stochastic rounding can decrease the quantization bias and improve image generation quality with little to no cost. To close the remaining gap between full-precision and quantized models, we suggest a feasible method for partial stochastic-rounding of weights. When using the MS-COCO dataset as the baseline, we show our quantization methods achieve as good FID scores as the full-precision model. Moreover, our methods decrease the quantization-induced distortion of the images generated by the full-precision model, with the distortion decreasing with the number of diffusion steps.", "keywords": [ "Diffusion", "Quantization", "Floating-Point", "Data-Free", "Text-To-Image" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "nEDToD1R8M", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Fu-Yun Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Ling Yang", "gender": "Male", "institution": "DeepSeek AI", "country": "CN", "position": "Intern" }, { "name": "Mengdi Wang", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Associate Professor" }, { "name": "Zhaoyang Huang", "gender": "unknown", "institution": "Avolution AI", "country": "CN", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 6, "n_ref_uni": 27, "n_ref": 110, "n_ref_all": 141, "n_fig": 14, "n_tab": 4, "L_tab": 5273, "n_element_tab": 882, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 2358, "n_element_tab_1": 390, "formula_len_all": 1229, "formula_len_all_1": 512, "len_all": 166675, "len_all_1": 55116, "len_abs": 1488, "len_title": 116, "len_sents": 40184, "len_sents_1": 22296, "n_sents": 310, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1509, "title": "Rectified Diffusion: Straightness Is Not Your Need in Rectified Flow", "abs": "Diffusion models have greatly improved visual generation but are hindered by slow generation speed due to the computationally intensive nature of solving generative ODEs. Rectified flow, a widely recognized solution, improves generation speed by straightening the ODE path. Its key components include: 1) using the diffusion form of flow-matching, 2) employing $\\boldsymbol v$-prediction, and 3) performing rectification (a.k.a. reflow). In this paper, we argue that the success of rectification primarily lies in using a pretrained diffusion model to obtain matched pairs of noise and samples, followed by retraining with these matched noise-sample pairs. Based on this, components 1) and 2) are unnecessary. Furthermore, we highlight that straightness is not an essential training target for rectification; rather, it is a specific case of flow-matching models. The more critical training target is to achieve a first-order approximate ODE path, which is inherently curved for models like DDPM and Sub-VP. Building on this insight, we propose Rectified Diffusion, which broadens the design space and application scope of rectification to encompass the broader category of diffusion models, rather than being restricted to flow-matching models. We validate our methods on Stable Diffusion v1-5 and Stable Diffusion XL. Our methods not only greatly simplifies the training procedure of rectified flow-based previous works~(e.g., InstaFlow) but also achieves superior performance with even lower training cost.", "keywords": [ "Diffusion models", "Rectified-Flow" ], "rating_list": [ 8, 1, 6, 8 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "nE3flbe88p", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Demetri Terzopoulos", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "Qian Long", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Ran Gong", "gender": "Male", "institution": "The AI Institute", "country": "US", "position": "Researcher" }, { "name": "Xiaofeng Gao", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Scientist" }, { "name": "Zhi Li", "gender": "Male", "institution": "UCLA Computer Science Department, University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Yingnian Wu", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 13, "n_ref": 24, "n_ref_all": 37, "n_fig": 5, "n_tab": 13, "L_tab": 7179, "n_element_tab": 934, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 698, "n_element_tab_1": 112, "formula_len_all": 771, "formula_len_all_1": 826, "len_all": 159255, "len_all_1": 57337, "len_abs": 1197, "len_title": 116, "len_sents": 43558, "len_sents_1": 26552, "n_sents": 401, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1205, "title": "TeamCraft: A Benchmark for Embodied Multi-Agent Systems in Minecraft", "abs": "Complex 3D environments replete with dynamic interactions among multiple agents and objects are essential for the development of embodied intelligent agents. To facilitate research on Multi-Agent (MA) systems, we introduce \\benchmark, a challenging MA benchmark based on the Minecraft game. Instead of the abstract vector inputs commonly provided to agents in MA systems research, \\benchmark provides agents with multi-modal task specifications and observations. Given the three-orthographic-view graph of the environment along with language instructions, the agents must efficiently collaborate to complete assigned tasks. Such multi-modal inputs pose a higher level of difficulty, since agents must generalize across diverse object and background imagery, different numbers of agents, a wide range of tasks, etc. Our planner-generated dataset includes various tasks, such as building construction, smelting, and farming, with a total of 70,000 procedurally-generated demonstrations that feature over 50 objects across a wide variety of scenes. We test the generalization abilities of several baseline Vision-Language Model (VLM) multi-agent control strategies in centralized and decentralized settings.", "keywords": [ "Multi-agent system", "embodied AI" ], "rating_list": [ 3, 3, 6, 1 ], "soundness_list": [ 1, 2, 2, 1 ], "presentation_list": [ 1, 1, 3, 1 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "nE1l0vpQDP", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Baoxiang Wang", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Shuai Li", "gender": "Female", "institution": "John Hopcroft Center, Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Wei Liu", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "Postdoc" }, { "name": "ruinan Jin", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Postdoc" } ], "n_formula": 192, "n_formula_1": 37, "n_ref_uni": 15, "n_ref": 49, "n_ref_all": 114, "n_fig": 1, "n_tab": 4, "L_tab": 884, "n_element_tab": 80, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 24823, "formula_len_all_1": 3344, "len_all": 172864, "len_all_1": 66820, "len_abs": 789, "len_title": 110, "len_sents": 44908, "len_sents_1": 25968, "n_sents": 543, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 837, "title": "The Implicit Bias of Stochastic AdaGrad-Norm on Separable Data", "abs": "This work explores stochastic adaptive gradient descent, i.e., stochastic AdaGrad-Norm, when applied to linearly separable datasets. For the stochastic AdaGrad-Norm method equipped with a wide range of sampling noise, we demonstrate its almost surely convergence result to the $\\mathcal{L}^{2}$ max-margin solution. This means that stochastic AdaGrad-Norm has an implicit bias that yields good generalization, even without regularization terms. We show that the convergence rate of the classification direction is $o({1}/{\\ln^{(1-\\epsilon)/{2}}n})$. Our approach takes a novel stance by explicitly characterizing the $\\mathcal{L}^{2}$ max-margin direction. By doing so, we overcome the challenge that arises from the dependency between the stepsize and the gradient and also address the limitations in the previous AdaGrad-Norm analyses.", "keywords": [ "stochastic AdaGrad-Norm", "implicit bias", "convergence results", "stochastic optimization" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "nDvgHIBRxQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Derek F. Wong", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Associate Professor" }, { "name": "Jindong Wang", "gender": "Male", "institution": "William & Mary", "country": "US", "position": "Assistant Professor" }, { "name": "Kaizhu Huang", "gender": "Male", "institution": "Duke Kunshan University", "country": "CN", "position": "Full Professor" }, { "name": "Maizhen Ning", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "CN", "position": "PhD student" }, { "name": "Qiufeng Wang", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "CN", "position": "Full Professor" }, { "name": "Wei Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Xiaowei Huang", "gender": "Male", "institution": "University of Liverpool", "country": "GB", "position": "Full Professor" }, { "name": "Zihao Zhou", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "", "position": "PhD student" }, { "name": "Liu Shudong", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 0, "n_ref_uni": 55, "n_ref": 119, "n_ref_all": 146, "n_fig": 9, "n_tab": 7, "L_tab": 4633, "n_element_tab": 259, "n_fig_1": 7, "n_tab_1": 8, "L_tab_1": 4788, "n_element_tab_1": 123, "formula_len_all": 197, "formula_len_all_1": 0, "len_all": 298176, "len_all_1": 63851, "len_abs": 2053, "len_title": 139, "len_sents": 77559, "len_sents_1": 25998, "n_sents": 805, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 2060, "title": "Is Your Model Really A Good Math Reasoner? Evaluating Mathematical Reasoning with Checklist", "abs": "Exceptional mathematical reasoning ability is one of the key features that demonstrate the power of large language models (LLMs). How to comprehensively define and evaluate the mathematical abilities of LLMs, and even reflect the user experience in real-world scenarios, has emerged as a critical issue. Current benchmarks predominantly concentrate on problem-solving capabilities, presenting a substantial risk of model overfitting and fails to accurately measure the genuine mathematical reasoning abilities. In this paper, we argue that if a model really understands a problem, it should be robustly and readily applied across a diverse array of tasks. To this end, we introduce MathCheck, a well-designed checklist for testing task generalization and reasoning robustness, as well as an automatic tool to generate checklists efficiently. MathCheck includes multiple mathematical reasoning tasks and robustness tests to facilitate a comprehensive evaluation of both mathematical reasoning ability and behavior testing. Utilizing MathCheck, we develop MathCheck-GSM and MathCheck-GEO to assess mathematical textual reasoning and multi-modal reasoning capabilities, respectively, serving as upgraded versions of benchmarks including GSM8k, GeoQA, UniGeo, and Geometry3K. We adopt MathCheck-GSM and MathCheck-GEO to evaluate over 26 LLMs and 17 multi-modal LLMs, assessing their comprehensive mathematical reasoning abilities. Our results demonstrate that while frontier LLMs like GPT-4o continue to excel in various abilities on the checklist, many other model families exhibit a significant decline. Further experiments indicate that, compared to traditional math benchmarks, MathCheck better reflects true mathematical abilities and represents mathematical intelligence more linearly, thereby supporting our design. Using MathCheck, we can also efficiently conduct informative behavior analysis to deeply investigate models. Finally, we show that our proposed checklist paradigm can easily extend to other reasoning tasks for their comprehensive evaluation.", "keywords": [ "Mathmatical Reasoning", "Geometry reasoning", "Evaluation", "(Multi-Modal) LLMs" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "nDmwloEl3N", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Jyothish Pari", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Moritz Reuss", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "", "position": "PhD student" }, { "name": "Pulkit Agrawal", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Rudolf Lioutikov", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "Tenure-Track Professor" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 46, "n_ref": 94, "n_ref_all": 122, "n_fig": 16, "n_tab": 8, "L_tab": 2219, "n_element_tab": 216, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 1218, "n_element_tab_1": 219, "formula_len_all": 574, "formula_len_all_1": 502, "len_all": 212856, "len_all_1": 67395, "len_abs": 1522, "len_title": 144, "len_sents": 52671, "len_sents_1": 29776, "n_sents": 462, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1538, "title": "Efficient Diffusion Transformer Policies with Mixture of Expert Denoisers for Multitask Learning", "abs": "Diffusion Policies have become widely used in Imitation Learning, offering several appealing properties, such as generating multimodal and discontinuous behavior.\nAs models are becoming larger to capture more complex capabilities, their computational demands increase, as shown by recent scaling laws. \nTherefore, continuing with the current architectures will present a computational roadblock. \nTo address this gap, we propose Mixture-of-Denoising Experts (MoDE) as a novel policy for Imitation Learning.\nMoDE surpasses current state-of-the-art Transformer-based Diffusion Policies while enabling parameter-efficient scaling through sparse experts and noise-conditioned routing, reducing both active parameters by 40\\% and inference costs by 80\\% via expert caching.\nOur architecture combines this efficient scaling with noise-conditioned self-attention mechanism, enabling more effective denoising across different noise levels. \nMoDE achieves state-of-the-art performance across 134 tasks in four established imitation learning benchmarks (CALVIN and LIBERO).\nNotably, by pretraining MoDE on diverse robotics data, we achieve a new state-of-the-art result of 3.98 on CALVIN and 0.95 on LIBERO-90. It surpasses both CNN-based and Transformer Diffusion Policies by an average of $20\\%$ in all settings, while using 80\\% fewer FLOPs and fewer active parameters.\nFurthermore, we conduct comprehensive ablations on MoDE's components, providing insights for designing efficient and scalable Transformer architectures for Diffusion Policies.", "keywords": [ "Robotics", "Imitation Learning" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "nDj45w5wam", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Fan Feng", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Postdoc" }, { "name": "Hongye Cao", "gender": "Male", "institution": "nanjing university", "country": "", "position": "PhD student" }, { "name": "Jing Huo", "gender": "Female", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Tianpei Yang", "gender": "Female", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yang Gao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 12, "n_ref_uni": 75, "n_ref": 136, "n_ref_all": 173, "n_fig": 26, "n_tab": 5, "L_tab": 1434, "n_element_tab": 128, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 112, "n_element_tab_1": 22, "formula_len_all": 1763, "formula_len_all_1": 1493, "len_all": 214555, "len_all_1": 57393, "len_abs": 1529, "len_title": 118, "len_sents": 60239, "len_sents_1": 27009, "n_sents": 426, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1521, "title": "Causal Information Prioritization for Efficient Reinforcement Learning", "abs": "Current Reinforcement Learning (RL) methods often suffer from sample-inefficiency, resulting from blind exploration strategies that neglect causal relationships among states, actions, and rewards. Although recent causal approaches aim to address this problem, they lack grounded modeling of reward-guided causal understanding of states and actions for goal-orientation, thus impairing learning efficiency. To tackle this issue, we propose a novel method named Causal Information Prioritization (CIP) that improves sample efficiency by leveraging factored MDPs to infer causal relationships between different dimensions of states and actions with respect to rewards, enabling the prioritization of causal information. Specifically, CIP identifies and leverages causal relationships between states and rewards to execute counterfactual data augmentation to prioritize high-impact state features under the causal understanding of the environments. Moreover, CIP integrates a causality-aware empowerment learning objective, which significantly enhances the agent's execution of reward-guided actions for more efficient exploration in complex environments. \nTo fully assess the effectiveness of CIP, we conduct extensive experiments across $39$ tasks in $5$ diverse continuous control environments, encompassing both locomotion and manipulation skills learning with pixel-based and sparse reward settings. Experimental results demonstrate that CIP consistently outperforms existing RL methods across a wide range of scenarios.", "keywords": [ "causality", "reinforcement learning", "empowerment", "sample efficiency" ], "rating_list": [ 8, 6, 5 ], "soundness_list": [ 4, 4, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 4, 3, 2 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "nDTvP6tBMd", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Corey Wang", "gender": "Not Specified", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Danxian Liu", "gender": "Female", "institution": "Harvard University, Harvard University", "country": "", "position": "PhD student" }, { "name": "Erik Y. Wang", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "Undergrad student" }, { "name": "Jingxuan Fan", "gender": "Female", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Jonah Brenner", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Undergrad student" }, { "name": "Kaylie Hausknecht", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Michael Brenner", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "Professor" }, { "name": "Nianli Peng", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Sarah Martinson", "gender": "Female", "institution": "Harvard University", "country": "US", "position": "MS student" } ], "n_formula": 73, "n_formula_1": 5, "n_ref_uni": 19, "n_ref": 42, "n_ref_all": 66, "n_fig": 9, "n_tab": 3, "L_tab": 882, "n_element_tab": 114, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1237, "n_element_tab_1": 123, "formula_len_all": 3450, "formula_len_all_1": 259, "len_all": 184142, "len_all_1": 66118, "len_abs": 1301, "len_title": 125, "len_sents": 64975, "len_sents_1": 30396, "n_sents": 533, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1346, "title": "HARDMath: A Benchmark Dataset for Challenging Problems in Applied Mathematics", "abs": "Advanced applied mathematics problems are underrepresented in existing Large Language Model (LLM) benchmark datasets. To address this, we introduce $\\textbf{HARDMath}$, a dataset inspired by a graduate course on asymptotic methods, featuring challenging applied mathematics problems that require analytical approximation techniques. These problems demand a combination of mathematical reasoning, computational tools, and subjective judgment, making them difficult for LLMs. Our framework auto-generates a large number of problems with solutions validated against numerical ground truths. We evaluate both open- and closed-source LLMs on $\\textbf{HARDMath-mini}$, a sub-sampled test set of 366 problems, as well as on 40 word problems formulated in applied science contexts. Even leading closed-source models like GPT-4 achieve only 43.8% overall accuracy with few-shot Chain-of-Thought prompting, and all models demonstrate significantly lower performance compared to results on existing mathematics benchmark datasets. We additionally conduct a detailed error analysis to gain insights into the failure cases of LLMs. These results demonstrate the limitations of current LLM performance on advanced graduate-level applied math problems and underscore the importance of datasets like $\\textbf{HARDMath}$ to advance mathematical abilities of LLMs.", "keywords": [ "math", "benchmark", "dataset", "few-shot learning", "reasoning" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 1, 4 ] }, { "paper_id": "nD5tbHBfut", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "David Picard", "gender": "Male", "institution": "\u00c9cole des Ponts", "country": "FR", "position": "Full Professor" }, { "name": "Jeremie Mary", "gender": "Male", "institution": "Criteo", "country": "", "position": "Senior Researcher" }, { "name": "Simon Lepage", "gender": "Male", "institution": "Criteo", "country": "FR", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 58, "n_ref": 96, "n_ref_all": 121, "n_fig": 15, "n_tab": 5, "L_tab": 2536, "n_element_tab": 230, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 374, "formula_len_all_1": 191, "len_all": 234842, "len_all_1": 63025, "len_abs": 1013, "len_title": 113, "len_sents": 72522, "len_sents_1": 30417, "n_sents": 693, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 891, "title": "LRVS-Fashion: Extending Visual Search with Referring Instructions", "abs": "This paper introduces a new challenge for image similarity search in the context of fashion, addressing the inherent ambiguity in this domain stemming from complex images. We present Referred Visual Search (RVS), a task allowing users to define more precisely the desired similarity, following recent interest in the industry. We release a new large public dataset, LRVS-Fashion, consisting of 272k fashion products with 842k images extracted from fashion catalogs, designed explicitly for this task. However, unlike traditional visual search methods in the industry, we demonstrate that superior performance can be achieved by bypassing explicit object detection and adopting weakly-supervised conditional contrastive learning on image tuples. Our method is lightweight and demonstrates robustness, reaching Recall at one superior to strong detection-based baselines against 2M distractors.", "keywords": [ "Visual Search", "Image Embedding", "Retrieval", "Dataset" ], "rating_list": [ 3, 5, 5, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "nCxULYtwkC", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Donguk Kim", "gender": "Male", "institution": "SHIFTUP", "country": "KR", "position": "Researcher" }, { "name": "Kyungmoon Lee", "gender": "Male", "institution": "Nalbi", "country": "KR", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 32, "n_ref": 78, "n_ref_all": 87, "n_fig": 2, "n_tab": 4, "L_tab": 586, "n_element_tab": 90, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 586, "n_element_tab_1": 90, "formula_len_all": 426, "formula_len_all_1": 426, "len_all": 105395, "len_all_1": 74194, "len_abs": 1402, "len_title": 166, "len_sents": 28296, "len_sents_1": 26419, "n_sents": 198, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 118, "L_abs": 1406, "title": "Stylize and Align: Unlabeled-Image Stylized Continuous Consistency Regularization for Hand Pose Estimation in the Wild", "abs": "Hand pose estimation has become a cornerstone of advanced human behavior understanding. In particular, 3D hand pose estimation has seen significant attention, with numerous approaches being proposed. However, it is unclear whether the modern approaches are applicable to real-world scenarios directly. We are focused on the robustness of hand pose estimators in the wild, noting that existing datasets exhibit distinct differences from real-world data. Thus, despite great advances, there remains considerable room for improvement, as most recent efforts have primarily focused on model architectures or on datasets within limited environments. To this end, we present a novel approach that unifies two key techniques: style transfer using unlabeled in-the-wild images to enhance data diversity (\\ie, Stylize) and continuous consistency regularization (CCR) to capture fine-grained relations between hand pose data, providing rich supervisory signals (\\ie, Align). To evaluate the robustness of the learned representations through our framework, we demonstrate that our method significantly enhances generalization capabilities across various tasks, including 3D hand pose estimation and transfer learning for 2D hand pose estimation, all within our designed real-world testbed. Notably, these improvements are achieved using less than 5\\% of the data size compared to a large-scale dataset, InterHand2.6M.", "keywords": [ "Pose estimation", "3D hand pose estimation", "Consistency regularization", "Style transfer", "Robustness", "Metric learning" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "nCrJD7qPJN", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Gyuwon Sim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "HeeSun Bae", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Il-chul Moon", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Wanmo Kang", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "Professor" }, { "name": "Dong Hyeok Shin", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 15, "n_ref_uni": 53, "n_ref": 110, "n_ref_all": 154, "n_fig": 12, "n_tab": 13, "L_tab": 7250, "n_element_tab": 1076, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 306, "n_element_tab_1": 55, "formula_len_all": 3621, "formula_len_all_1": 1444, "len_all": 222807, "len_all_1": 72694, "len_abs": 1374, "len_title": 84, "len_sents": 65664, "len_sents_1": 32435, "n_sents": 536, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 36, "L_abs": 1378, "title": "Distilling Dataset into Neural Field", "abs": "Utilizing large-scale datasets is essential for training high-performance deep learning models, but it also comes with substantial computation and storage costs. To overcome these challenges, dataset distillation has emerged as a promising solution by compressing large-scale datasets into smaller synthetic versions that retain the essential information needed for training. This paper proposes a novel parameterization framework for dataset distillation, coined Distilling Dataset into Neural Field (DDiF), which leverages the neural field to store the necessary information of large-scale datasets. Due to the unique nature of the neural field, which takes coordinates as input and output quantity, DDiF effectively preserves the information and easily generates various shapes of data. Beyond the efficacy, DDiF has larger feature coverage than some previous literature if same budget is allowed, which is proved from the frequency domain perspective. Under the same budget setting, this larger coverage leads to a significant performance improvement in downstream tasks by providing more synthetic instances due to the coding efficiency. DDiF demonstrates both theoretical and empirical evidence of its ability to operate efficiently within a limited budget, while better preserving the information of the original dataset compared to conventional parameterization methods.", "keywords": [ "Dataset distillation", "Dataset condensation", "Neural field" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "nAl4bz09Mv", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Qingyi Gu", "gender": "unknown", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xuewen Liu", "gender": "Male", "institution": "Institute of automation", "country": "CN", "position": "PhD student" }, { "name": "Zhikai Li", "gender": "Male", "institution": "Institute of automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 15, "n_ref_uni": 22, "n_ref": 82, "n_ref_all": 112, "n_fig": 7, "n_tab": 11, "L_tab": 5612, "n_element_tab": 670, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1858, "n_element_tab_1": 291, "formula_len_all": 1918, "formula_len_all_1": 913, "len_all": 144337, "len_all_1": 63825, "len_abs": 2262, "len_title": 127, "len_sents": 41339, "len_sents_1": 26492, "n_sents": 332, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 2046, "title": "DilateQuant: Accurate and Efficient Diffusion Quantization via Weight Dilation", "abs": "Diffusion models have shown excellent performance on various image generation tasks, but the substantial computational costs and huge memory footprint hinder their low-latency applications in real-world scenarios. Quantization is a promising way to compress and accelerate models. Nevertheless, due to the wide range and time-varying activations in diffusion models, existing methods cannot maintain both accuracy and efficiency simultaneously for low-bit quantization. To tackle this issue, we propose DilateQuant, a novel quantization framework for diffusion models that offers comparable accuracy and high efficiency. Specifically, we keenly aware of numerous unsaturated in-channel weights, which can be cleverly exploited to reduce the range of activations without additional computation cost. Based on this insight, we propose Weight Dilation (WD) that maximally dilates the unsaturated in-channel weights to a constrained range through a mathematically equivalent scaling. WD costlessly absorbs the activation quantization errors into weight quantization. The range of activations decreases, which makes activations quantization easy. The range of weights remains constant, which makes model easy to converge in training stage. Considering the temporal network leads to time-varying activations, we design a Temporal Parallel Quantizer (TPQ), which sets time-step quantization parameters and supports parallel quantization for different time steps, significantly improving the performance and reducing time cost. To further enhance performance while preserving efficiency, we introduce a Block-wise Knowledge Distillation (BKD) to align the quantized models with the full-precision models at a block level. The simultaneous training of time-step quantization parameters and weights minimizes the time required, and the shorter backpropagation paths decreases the memory footprint of the quantization process. Extensive experiments demonstrate that DilateQuant significantly outperforms existing methods in terms of accuracy and efficiency.", "keywords": [ "quantization", "diffusion model" ], "rating_list": [ 3, 3, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "nAVejJURqZ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 13, "author_info_list": [ { "name": "Chenting Wang", "gender": "Male", "institution": "Nanjing university", "country": "CN", "position": "Undergrad student" }, { "name": "Kunchang Li", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Songze Li", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Tianxiang Jiang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Xiangyu Zeng", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Xinhao Li", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Intern" }, { "name": "Yali Wang", "gender": "Male", "institution": "SIAT, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yansong Shi", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yi Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhengrong Yue", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Ziang Yan", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Limin Wang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 28, "n_ref": 105, "n_ref_all": 131, "n_fig": 6, "n_tab": 11, "L_tab": 4462, "n_element_tab": 427, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 11, "n_element_tab_1": 3, "formula_len_all": 694, "formula_len_all_1": 223, "len_all": 206967, "len_all_1": 65329, "len_abs": 0, "len_title": 47, "len_sents": 59167, "len_sents_1": 32420, "n_sents": 463, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1870, "title": "TimeSuite: Improving MLLMs for Long Video Understanding via Grounded Tuning", "abs": "Multimodal Large Language Models (MLLMs) have demonstrated impressive performance in short video understanding. However, understanding long-form videos still remains challenging for MLLMs. This paper proposes TimeSuite, a collection of new designs to adapt the existing short-form video MLLMs for long video understanding, including a simple yet efficient framework to process long video sequence, a high-quality video dataset for grounded tuning of MLLMs, and a carefully-designed instruction tuning task to explicitly incorporate the grounding supervision in the traditional QA format. Specifically, based on VideoChat, we propose our long-video MLLM, coined as VideoChat-T, by implementing a token shuffling to compress long video tokens and introducing Temporal Adaptive Position Encoding (TAPE) to enhance the temporal awareness of visual representation. Meanwhile, we introduce the TimePro, a comprehensive grounding-centric instruction tuning dataset composed of 9 tasks and 349k high-quality grounded annotations. Notably, we design a new instruction tuning task type, called Temporal Grounded Caption, to peform detailed video descriptions with the corresponding time stamps prediction. This explicit temporal location prediction will guide MLLM to correctly attend on the visual content when generating description, and thus reduce the hallucination risk caused by the LLMs. Experimental results demonstrate that our TimeSuite provides a successful solution to enhance the long video understanding capability of short-form MLLM, achieving improvement of 5.6% and 6.8% on the benchmarks of Egoschema and VideoMME, respectively. In addition, VideoChat-T exhibits robust zero-shot temporal grounding capabilities, significantly outperforming the existing state-of-the-art MLLMs. After fine-tuning, it performs on par with the traditional supervised expert models.", "keywords": [ "Long Video Understanding; Temporal Grounding; Multimodal Large Language Model" ], "rating_list": [ 8, 6, 3, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 4, 5, 5, 4, 4 ] }, { "paper_id": "nAK26c8s9X", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 1, "author_info_list": [ { "name": "Hadi Abdullah", "gender": "unknown", "institution": "VISA", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 22, "n_ref": 61, "n_ref_all": 88, "n_fig": 4, "n_tab": 4, "L_tab": 729, "n_element_tab": 35, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1121, "n_element_tab_1": 52, "formula_len_all": 65, "formula_len_all_1": 49, "len_all": 110893, "len_all_1": 66659, "len_abs": 1112, "len_title": 112, "len_sents": 38920, "len_sents_1": 30849, "n_sents": 334, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1124, "title": "Boosting Membership Inference Attacks with Upstream Modification", "abs": "Membership Inference Attacks (MIAs) can be used by model owners to identify privacy leakage of specific points in their machine learning models. In this setting, the model owner (who is playing the role of the attacker) has perfect knowledge of the training data but a limited computational budget. However, current MIAs have limited effectiveness in this scenario: 1) They perform poorly against the most vulnerable points 2) They require training too many models. To overcome this weakness, we modify two limitations, in the initial/upstream stages of the MIA framework, namely sampling bias (i.e., too many points dropped during sampling) and attack aggregation (i.e., average attack results over all the data points instead of only the most vulnerable ones). Our improvements carryover downstream and boost attack accuracy of existing MIAs by \\textit{increasing the TPR of existing attacks at incredibly low FPRs (as low as zero) while achieving a near-perfect AUC}. As a consequence, our modifications enable the practical and effective application of MIAs for identification of data-leakage in machine learning models.", "keywords": [ "Membership inference attacks" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "nA9SCxGy2M", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Xuemei Peng", "gender": "Female", "institution": "Hong Kong University of Science and Technology(GZ)", "country": "", "position": "PhD student" }, { "name": "Yuebin XU", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "", "position": "PhD student" }, { "name": "Zeyi Wen", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 35, "n_ref": 50, "n_ref_all": 65, "n_fig": 3, "n_tab": 3, "L_tab": 1787, "n_element_tab": 156, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1015, "n_element_tab_1": 143, "formula_len_all": 835, "formula_len_all_1": 678, "len_all": 117212, "len_all_1": 58551, "len_abs": 1263, "len_title": 90, "len_sents": 28485, "len_sents_1": 26278, "n_sents": 223, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 1270, "title": "Model-Driven Labeled Data Free Fine-tuning", "abs": "Supervised fine-tuning is a prevalent technique for boosting model performance. However, it heavily depends on extensive training over labeled data. This paper introduces a novel model-driven fine-tuning method that operates independently of supervised training and labeled data. By harnessing the collective intelligence of a diverse model pool, our method enhances individual model performance through a two-phase process. Initially, we consolidate the expertise of the models within the pool to create a general meta-model. This meta-model then serves as a guide for iteratively fine-tuning the original models in a few shots, promoting a synergistic improvement in performance. Our experimental results show that this model-driven approach not only surpasses the performance of full-parameter fine-tuning models but also does so without the need for supervised training. This breakthrough offers a cost-effective and scalable alternative to traditional supervised fine-tuning, addressing the challenge of data scarcity and paving the way for future research in unsupervised model enhancement. Our work represents a significant step towards making fine-tuning techniques more accessible and practical in environments where labeled data is limited or even unavailable.", "keywords": [ "Fine-tuning", "unsupervised learning", "foundation models" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 2, 2, 1, 1 ], "presentation_list": [ 1, 2, 1, 1 ], "contribution_list": [ 2, 3, 1, 1 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "nA464tCGR5", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Arne Speerforck", "gender": "Male", "institution": "Technische Universit\u00e4t Hamburg", "country": "DE", "position": "Full Professor" }, { "name": "Johannes Brunnemann", "gender": "unknown", "institution": "Technische Universit\u00e4t Hamburg", "country": "DE", "position": "Lecturer" }, { "name": "Julius Aka", "gender": "Male", "institution": "University of Augsburg", "country": "DE", "position": "PhD student" }, { "name": "J\u00f6rg Eiden", "gender": "unknown", "institution": "XRG Simulation GmbH", "country": "DE", "position": "Researcher" }, { "name": "Lars Mikelsons", "gender": "unknown", "institution": "University of Augsburg", "country": "DE", "position": "Full Professor" } ], "n_formula": 55, "n_formula_1": 23, "n_ref_uni": 42, "n_ref": 85, "n_ref_all": 112, "n_fig": 10, "n_tab": 5, "L_tab": 2344, "n_element_tab": 321, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 386, "n_element_tab_1": 72, "formula_len_all": 5198, "formula_len_all_1": 2390, "len_all": 184493, "len_all_1": 68631, "len_abs": 1326, "len_title": 139, "len_sents": 51829, "len_sents_1": 27832, "n_sents": 442, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1331, "title": "Balanced Neural ODEs: nonlinear model order reduction and Koopman operator approximations", "abs": "Variational Autoencoders (VAEs) are a powerful framework for learning latent representations of reduced dimensionality, while Neural ODEs excel in learning transient system dynamics. This work combines the strengths of both to generate fast surrogate models with adjustable complexity reacting on time-varying inputs signals. By leveraging the VAE\u2019s dimensionality reduction using a non-hierarchical prior, our method adaptively assigns stochastic noise, naturally complementing known NeuralODE training enhancements and enabling probabilistic\ntime series modeling. We show that standard Latent ODEs struggle with dimensionality reduction in systems with time-varying inputs. Our approach mitigates this by continuously propagating variational parameters through time, establishing fixed information channels in latent space. This results in a flexible and robust method that can learn different system complexities, e.g. deep neural networks or\nlinear matrices. Hereby, it enables efficient approximation of the Koopman operator without the need for predefining its dimensionality. As our method balances dimensionality reduction and reconstruction accuracy, we call it Balanced Neural ODE (B-NODE). We demonstrate the effectiveness of this methods on several academic and real-world test cases, e.g. a power plant or MuJoCo data.", "keywords": [ "dynamical variational autoencoders", "state space models", "Neural ODEs", "Koopman theory", "surrogate models" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "nA1D0Y65m2", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Bei Jiang", "gender": "Female", "institution": "University of Alberta", "country": "CA", "position": "Associate Professor" }, { "name": "Enze Shi", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "PhD student" }, { "name": "Ke Sun", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Linglong Kong", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Full Professor" }, { "name": "Xiaodong Yan", "gender": "Male", "institution": "Shandong University", "country": "", "position": "Associate Professor" }, { "name": "Yafei Wang", "gender": "unknown", "institution": "University of Alberta", "country": "CA", "position": "Assistant Professor" }, { "name": "Yingnan Zhao", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 69, "n_formula_1": 24, "n_ref_uni": 28, "n_ref": 72, "n_ref_all": 95, "n_fig": 9, "n_tab": 4, "L_tab": 1951, "n_element_tab": 110, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9732, "formula_len_all_1": 2390, "len_all": 219657, "len_all_1": 72086, "len_abs": 1730, "len_title": 167, "len_sents": 69574, "len_sents_1": 33265, "n_sents": 511, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 117, "L_abs": 1191, "title": "The Benefits of Being Categorical Distributional: Uncertainty-aware Regularized Exploration in Reinforcement Learning", "abs": "Despite the remarkable empirical performance of distributional reinforcement learning (RL), its theoretical advantages over classical RL are not fully understood. Starting with Categorical Distributional RL (CDRL), we propose that the potential superiority of distributional RL can be attributed to a derived distribution-matching regularization by applying a return density function decomposition technique. This less-studied regularization in the distributional RL context aims to capture additional knowledge of return distribution beyond only its expectation, contributing to an augmented reward signal in policy optimization. In contrast to the standard entropy regularization in MaxEnt RL, which explicitly encourages exploration by promoting diverse actions, the regularization derived from CDRL implicitly updates policies to align the learned policy with environmental uncertainty. Finally, extensive experiments substantiate the significance of this uncertainty-aware regularization derived from distributional RL on the empirical benefits over classical RL. Our study offers a new perspective from the exploration to explain the benefits of adopting distributional learning in RL.", "keywords": [ "distributional learning", "reinforcement learning", "exploration" ], "rating_list": [ 3, 8, 6, 3 ], "soundness_list": [ 3, 3, 3, 1 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 3, 2, 3, 5 ] }, { "paper_id": "n9PDaFNi8t", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Chengyou Jia", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Fangzhi Xu", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Kanzhi Cheng", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Liheng Chen", "gender": "Male", "institution": "The Univerisity of Hong Kong", "country": "", "position": "Undergrad student" }, { "name": "Paul Pu Liang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Qiushi Sun", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yian Wang", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhenyu Wu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Zhiyong Wu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Ding Zichen", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 28, "n_ref": 89, "n_ref_all": 115, "n_fig": 8, "n_tab": 6, "L_tab": 2774, "n_element_tab": 335, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1164, "n_element_tab_1": 122, "formula_len_all": 196, "formula_len_all_1": 196, "len_all": 146740, "len_all_1": 62317, "len_abs": 1439, "len_title": 109, "len_sents": 41035, "len_sents_1": 29152, "n_sents": 327, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 59, "L_abs": 1432, "title": "OS-ATLAS: Foundation Action Model for Generalist GUI Agents", "abs": "Existing efforts in building GUI agents heavily rely on the availability of robust commercial Vision-Language Models (VLMs) such as GPT-4o and GeminiPro\u0002Vision. Practitioners are often reluctant to use open-source VLMs due to their significant performance lag compared to their closed-source counterparts, particularly in GUI grounding and Out-Of-Distribution (OOD) scenarios. To facilitate future research in this area, we developed OS-Atlas \u2014a foundational GUI action model that excels at GUI grounding and OOD agentic tasks through innovations in both data and modeling. We have invested substantial engineering effort into developing a toolkit for synthesizing multi-platform GUI grounding data. Lever\u0002aging this toolkit, we are releasing the largest open-source cross-platform GUI grounding corpus to date, which contains over 13 million GUI elements. This dataset, combined with innovations in model training, provides a solid foundation for OS-Atlas to understand GUI screenshots and generalize to unseen interfaces. Through extensive evaluation across six benchmarks spanning three different platforms (mobile, desktop, and web), OS-Atlas demonstrates significant performance improvements over previous state-of-the-art models. Our evaluation also uncovers valuable insights into continuously improving and scaling the agentic capabilities of open-source VLMs. All our data, code, and models will be made publicly available.", "keywords": [ "GUI agent", "language agent", "GUI grounding", "executable language grounding" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "n8h1z588eu", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Guanlin Mo", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Hu Ding", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Shihong Song", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" } ], "n_formula": 49, "n_formula_1": 23, "n_ref_uni": 64, "n_ref": 125, "n_ref_all": 149, "n_fig": 8, "n_tab": 7, "L_tab": 1633, "n_element_tab": 140, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 962, "n_element_tab_1": 27, "formula_len_all": 4539, "formula_len_all_1": 2432, "len_all": 185700, "len_all_1": 73054, "len_abs": 1490, "len_title": 167, "len_sents": 54330, "len_sents_1": 32340, "n_sents": 442, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 123, "L_abs": 1551, "title": "Relax and Merge: A Simple Yet Effective Framework for Solving Fair $k$-Means and $k$-sparse Wasserstein Barycenter Problems", "abs": "The fairness of clustering algorithms has gained widespread attention across various areas, including machine learning, In this paper, we study fair $k$-means clustering in Euclidean space. \n Given a dataset comprising several groups, the fairness constraint requires that each cluster should contain a proportion of points from each group within specified lower and upper bounds. \n Due to these fairness constraints, determining the optimal locations of $k$ centers is a quite challenging task. \n We propose a novel ``Relax and Merge'' framework that returns a $(1+4\\rho + O(\\epsilon))$-approximate solution, where $\\rho$ is the approximate ratio of an off-the-shelf vanilla $k$-means algorithm and $O(\\epsilon)$ can be an arbitrarily small positive number. If equipped with a PTAS of $k$-means, our solution can achieve an approximation ratio of $(5+O(\\epsilon))$ with only a slight violation of the fairness constraints, which improves the current state-of-the-art approximation guarantee. Furthermore, using our framework, we can also obtain a $(1+4\\rho +O(\\epsilon))$-approximate solution for the $k$-sparse Wasserstein Barycenter problem, which is a fundamental optimization problem in the field of optimal transport, and a $(2+6\\rho)$-approximate solution for the strictly fair $k$-means clustering with no violation, both of which are better than the current state-of-the-art methods. In addition, the empirical results demonstrate that our proposed algorithm can significantly outperform baseline approaches in terms of clustering cost.", "keywords": [ "clustering", "k-means", "fairness", "approxiamte algorithm", "optimal transport" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "n8O0trhost", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Fredrik Lindsten", "gender": "Male", "institution": "Link\u00f6ping University", "country": "SE", "position": "Associate Professor" }, { "name": "Gabriel Ducrocq", "gender": "Male", "institution": "Link\u00f6ping University", "country": "SE", "position": "Postdoc" }, { "name": "Lukas Grunewald", "gender": "Male", "institution": "Uppsala University", "country": "SE", "position": "PhD student" }, { "name": "Sebastian Westenhoff", "gender": "unknown", "institution": "Uppsala University", "country": "", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 29, "n_ref": 78, "n_ref_all": 103, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 675, "formula_len_all_1": 676, "len_all": 135825, "len_all_1": 68060, "len_abs": 1522, "len_title": 117, "len_sents": 37050, "len_sents_1": 30578, "n_sents": 298, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1530, "title": "cryoSPHERE: Single-Particle HEterogeneous REconstruction from cryo EM", "abs": "The three-dimensional structure of proteins plays a crucial role in determining their function. Protein structure prediction methods, like AlphaFold, offer rapid access to a protein\u2019s structure. However, large protein complexes cannot be reliably predicted, and proteins are dynamic, making it important to resolve their full conformational distribution. Single-particle cryo-electron microscopy (cryo-EM) is a powerful tool for determining the structures of large protein complexes. Importantly, the numerous images of a given protein contain underutilized information about conformational heterogeneity. These images are very noisy projections of the protein, and traditional methods for cryo-EM reconstruction are limited to recovering only one or a few consensus conformations.\n\nIn this paper, we introduce cryoSPHERE, which is a deep learning method that uses a nominal protein structure (e.g., from AlphaFold) as input, learns how to divide it into segments, and moves these segments as approximately rigid bodies to fit the different conformations present in the cryo-EM dataset. This approach provides enough constraints to enable meaningful reconstructions of single protein structural ensembles. We demonstrate this with two synthetic datasets featuring varying levels of noise, as well as one real dataset. We show that cryoSPHERE is very resilient to the high levels of noise typically encountered in experiments, where we see consistent improvements over the current state-of-the-art for heterogeneous reconstruction.", "keywords": [ "cryoEM; protein structure; Deep Learning; Machine Learning; generative modelling" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "n8MNWHfhTO", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Shaohui Li", "gender": "Male", "institution": " Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Xutao Liao", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "You He", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yu LIU", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yuhui Xu", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Research Scientist" }, { "name": "Zhi Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" } ], "n_formula": 36, "n_formula_1": 22, "n_ref_uni": 39, "n_ref": 58, "n_ref_all": 79, "n_fig": 5, "n_tab": 7, "L_tab": 3154, "n_element_tab": 473, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2699, "n_element_tab_1": 430, "formula_len_all": 2718, "formula_len_all_1": 1600, "len_all": 129646, "len_all_1": 66823, "len_abs": 1037, "len_title": 73, "len_sents": 31001, "len_sents_1": 25316, "n_sents": 265, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1057, "title": "GaLore$+$: Boosting Low-Rank Adaptation for LLMs with Cross-Head Projection", "abs": "Recent low-rank training methods, such as GaLore, have significantly reduced the memory required to optimize large language models (LLMs). However, these methods often suffer from time-consuming low-rank projection estimations. In particular, the singular value decomposition (SVD) in GaLore can consume more than 80\\% of the total training time. To address this issue, we propose GaLore$+$, which uses cross-head low-rank projection to reduce the substantial time consumption in estimating low-rank projections for multi-head attention. In addition, we employ randomized subspace iteration to achieve fast SVD. To further enhance performance, we propose sparsely coded residuals to reduce the errors caused by low-rank approximation on the first- and second-order moments of the optimizers and weight updates. We evaluate GaLore$+$ on arithmetic reasoning and natural language generation datasets. Our experiments demonstrate that GaLore$+$ delivers superior performance while achieving approximately $4\\times$ fine-tuning speed compared to vanilla GaLore.", "keywords": [ "large language models", "parameter-efficient fine-tuning", "low-rank" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "n8IzL0Vy4G", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Heng Ji", "gender": "Female", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Jiawei Han", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign (UIUC)", "country": "US", "position": "Full Professor" }, { "name": "Sha Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "PhD student" }, { "name": "Sizhe Zhou", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Yizhu Jiao", "gender": "Female", "institution": "UIUC", "country": "", "position": "PhD student" }, { "name": "Yu Meng", "gender": "Male", "institution": "University of Virginia", "country": "", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 81, "n_ref_all": 121, "n_fig": 5, "n_tab": 22, "L_tab": 11170, "n_element_tab": 1067, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 3882, "n_element_tab_1": 423, "formula_len_all": 147, "formula_len_all_1": 0, "len_all": 204141, "len_all_1": 74493, "len_abs": 1698, "len_title": 100, "len_sents": 57126, "len_sents_1": 31755, "n_sents": 392, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 1713, "title": "Establishing Knowledge Preference in Language Models", "abs": "Language models are known to encode a great amount of factual knowledge through pretraining. However, such knowledge might be insufficient to cater to user requests, requiring the model to integrate external knowledge sources and adhere to user-provided specifications. When answering questions about ongoing events, the model should use recent news articles to update its response; when asked to provide recommendations, the model should prioritize user specifications over retrieved product reviews; when some facts are edited in the model, the updated facts should override all prior knowledge learned by the model even if they are conflicting. In all of the cases above, the model faces a decision between its own parametric knowledge, (retrieved) contextual knowledge, and user instruction knowledge. In this paper, we (1) unify such settings into the problem of $\\textit{knowledge preference}$ and define a three-level preference hierarchy over these knowledge sources; (2) compile a collection of existing datasets IfQA, MQuAKE, and MRQA covering a combination of settings (with/without user specifications, with/without context documents) to systematically evaluate how well models obey the intended knowledge preference; and (3) propose a dataset synthesis method that composes diverse question-answer pairs with user assumptions and related context to directly fine-tune LMs for instilling the hierarchy of knowledge.\nWe demonstrate that a 7B model, fine-tuned on only a few thousand examples automatically generated by our proposed method, effectively achieves superior performance (more than 18% improvement across all evaluation benchmarks) in adhering to the desired knowledge preference hierarchy.", "keywords": [ "Knowledge Preference", "Knowledge Conflicts", "Retrieval-Augmented Generation", "Question Answering", "Large Language Model" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "n87wrNlcJu", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Jacopo Urbani", "gender": "Male", "institution": "Vrije Universiteit Amsterdam", "country": "NL", "position": "Associate Professor" }, { "name": "Michael Cochez", "gender": "Male", "institution": "VU Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Yannick Brunink", "gender": "Male", "institution": "Vrije Universiteit Amsterdam", "country": "NL", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 27, "n_ref": 54, "n_ref_all": 66, "n_fig": 4, "n_tab": 5, "L_tab": 801, "n_element_tab": 102, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 801, "n_element_tab_1": 102, "formula_len_all": 186, "formula_len_all_1": 185, "len_all": 115571, "len_all_1": 67498, "len_abs": 1507, "len_title": 88, "len_sents": 35286, "len_sents_1": 33074, "n_sents": 256, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1346, "title": "AutoRegressive Knowledge Base Completion", "abs": "Despite their large sizes, many Knowledge Graphs (KGs) remain highly incomplete. This problem has motivated numerous approaches to $\\textit{complete}$ the KGs by embedding them in a latent space to find the missing links. Although these methods show promising performance, a general limitation is that the scores given to possible links are uncalibrated and cannot be interpreted across different queries. Hence, we say they are $\\textit{local}$ as they relate to a specific context. This limitation makes it non-trivial to deduce the truth value of the links and to answer complex queries. Another limitation is that their learning depends on negative sampling, which is challenging due to the Open World Assumption (OWA). \n\nTo solve this problem, we propose a novel auto-regressive generative model that learns a joint distribution of the entities and relations of the KG without resorting to negative sampling. This distribution can be used to infer the probability that a link is sampled from the KG, which allows us to return a $\\textit{global}$ score that is interpretable in different contexts. Moreover, our method has the additional advantage that it offers probabilistic semantics for complex reasoning and knowledge base completion, achieving state-of-the-art performance on link prediction with consistent scores across the entire KG.", "keywords": [ "Knowledge Graphs", "Probabilistic Reasoning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 1 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "n7s9EwG6hW", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Ruizhe Zhong", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Shixiong Kai", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Siyuan Xu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xingbo Du", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Zhentao Tang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 45, "n_ref": 128, "n_ref_all": 153, "n_fig": 7, "n_tab": 10, "L_tab": 11984, "n_element_tab": 926, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 477, "n_element_tab_1": 26, "formula_len_all": 1044, "formula_len_all_1": 751, "len_all": 212088, "len_all_1": 71017, "len_abs": 2271, "len_title": 117, "len_sents": 51701, "len_sents_1": 30102, "n_sents": 483, "n_sents_1": 267, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1611, "title": "Towards LLM4Floorplan: Agents Can Do What Engineers Do in Chip Design", "abs": "Open-source tools have actively propelled advancements in physical electronic design, yet the deployment still requires substantial expertise. Recent progress in large language model (LLM)-based agents offer potential for automating physical design, but challenges remain in imparting domain-specific expertise and extracting case-specific design objectives to meet complex requirements. To address these issues, we introduce LLM4Floorplan, a multi-agent Floorplanner powered by LLMs. Unlike flow-level approaches that design workflows for multiple tasks, LLM4Floorplan is the first task-level agent specifically dedicated to a single physical design task. Specifically, we propose a simple yet effective search-cluster-based retriever that extracts the most relevant and diverse solutions from prior knowledge, drawing on essential domain-specific knowledge to ensure robust design performance. Building on the retriever, LLM4Floorplan integrates a novel Dynamic Retrieval-Augmented Thought (DRAT) prompting technique in which the LLM generation interacts with the retrieval system to precisely capture case-specific design objectives. With these innovations, LLM4Floorplan simulates the workflow of human engineers by facilitating task comprehension, model selection, hyperparameter tuning, code revisions, and performance evaluation. Extensive evaluations on public circuits with seven different LLM backbones demonstrate that LLM4Floorplan exhibits strong task comprehension and decision-making capabilities. Remarkably, for the strict requirement, LLM4Floorplan boosts the success rate from 0.250 to 0.875.", "keywords": [ "floorplanning", "large language model", "prompting", "electronic design automation", "physical design" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "n7qGCmluZr", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 7, "author_info_list": [ { "name": "Connor Stone", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Postdoc" }, { "name": "Laurence Perreault-Levasseur", "gender": "Female", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "Assistant Professor" }, { "name": "Pablo Lemos", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "", "position": "Postdoc" }, { "name": "Salma Salhi", "gender": "Female", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "MS student" }, { "name": "Sammy Nasser Sharief", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "MS student" }, { "name": "Yashar Hezaveh", "gender": "Male", "institution": "Mila", "country": "", "position": "Researcher" }, { "name": "Kolya Malkin", "gender": "Male", "institution": "Edinburgh University, University of Edinburgh", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 14, "n_ref_uni": 50, "n_ref": 79, "n_ref_all": 129, "n_fig": 21, "n_tab": 3, "L_tab": 916, "n_element_tab": 114, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 409, "n_element_tab_1": 65, "formula_len_all": 788, "formula_len_all_1": 746, "len_all": 211491, "len_all_1": 80256, "len_abs": 1687, "len_title": 102, "len_sents": 58634, "len_sents_1": 33804, "n_sents": 493, "n_sents_1": 282, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 1104, "title": "PQMass: Probabilistic Assessment of the Quality of Generative Models using Probability Mass Estimation", "abs": "We propose a likelihood-free method for comparing two distributions given samples from each, with the goal of assessing the quality of generative models. The proposed approach, PQMass, provides a statistically rigorous method for assessing the performance of a single generative model or the comparison of multiple competing models. PQMass divides the sample space into non-overlapping regions and applies chi-squared tests to the number of data samples that fall within each region, giving a $p$-value that measures the probability that the bin counts derived from two sets of samples are drawn from the same multinomial distribution. PQMass does not depend on assumptions regarding the density of the true distribution, nor does it rely on training or fitting any auxiliary models. We evaluate PQMass on data of various modalities and dimensions, demonstrating its effectiveness in assessing the quality, novelty, and diversity of generated samples. We further show that PQMass scales well to moderately high-dimensional data and thus obviates the need for feature extraction in practical applications.", "keywords": [ "Generative Models", "Model Evaluation", "Sample-based Metrics", "Hypothesis Testing" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "n7n8McETXw", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Hongkang Li", "gender": "unknown", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "Postdoc" }, { "name": "Meng Wang", "gender": "Female", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "Associate Professor" }, { "name": "Pin-Yu Chen", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Songtao Lu", "gender": "Male", "institution": "IBM Thomas J. Watson Research Center", "country": "", "position": "Senior Research Scientist" }, { "name": "Xiaodong Cui", "gender": "Male", "institution": "IBM T. J. Watson Research Center", "country": "US", "position": "Principal Research Staff Member" } ], "n_formula": 137, "n_formula_1": 46, "n_ref_uni": 20, "n_ref": 82, "n_ref_all": 107, "n_fig": 5, "n_tab": 4, "L_tab": 1010, "n_element_tab": 20, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12393, "formula_len_all_1": 4269, "len_all": 226933, "len_all_1": 76592, "len_abs": 1370, "len_title": 150, "len_sents": 62314, "len_sents_1": 30689, "n_sents": 609, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 1377, "title": "Training Nonlinear Transformers for Chain-of-Thought Inference: A Theoretical Generalization Analysis", "abs": "Chain-of-Thought (CoT) is an efficient prompting method that enables the reasoning ability of large language models by augmenting the query using multiple examples with multiple intermediate steps. Despite the empirical success, the theoretical understanding of how to train a Transformer to achieve the CoT ability remains less explored. This is primarily due to the technical challenges involved in analyzing the nonconvex optimization on nonlinear attention models. To the best of our knowledge, this work provides the first theoretical study of training Transformers with nonlinear attention to obtain the CoT generalization capability so that the resulting model can inference on unseen tasks when the input is augmented by examples of the new task. We first quantify the required training samples and iterations to train a Transformer model towards CoT ability. We then prove the success of its CoT generalization on unseen tasks with distribution-shifted testing data. Moreover, we theoretically characterize the conditions for an accurate reasoning output by CoT even when the provided reasoning examples contain noises and are not always accurate. In contrast, in-context learning (ICL), which can be viewed as one-step CoT without intermediate steps, may fail to provide an accurate output when CoT does. These theoretical findings are justified through experiments.", "keywords": [ "Theoretical Chain-of-Thought", "generalization", "deep learning theory", "In-Context Learning", "training dynamics of Transformers" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 3, 2, 3 ] }, { "paper_id": "n7iwmPacDt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Fei Chao", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Huixia Li", "gender": "Female", "institution": "ByteDance", "country": "", "position": "Researcher" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Ruilin Wang", "gender": "unknown", "institution": "Xiamen University", "country": "", "position": "MS student" }, { "name": "Xiawu Zheng", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Xuefeng Xiao", "gender": "Male", "institution": "ByteDance", "country": "CN", "position": "Researcher" }, { "name": "Yuexiao Ma", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 23, "n_ref": 53, "n_ref_all": 64, "n_fig": 3, "n_tab": 1, "L_tab": 1043, "n_element_tab": 70, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 242, "n_element_tab_1": 4, "formula_len_all": 1510, "formula_len_all_1": 1510, "len_all": 103563, "len_all_1": 54904, "len_abs": 1640, "len_title": 111, "len_sents": 29061, "len_sents_1": 25058, "n_sents": 246, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1713, "title": "Polybasic Speculative Decoding Under a Theoretical Perspective", "abs": "Speculative decoding has emerged as a critical technique for accelerating inference in large language models, achieving significant speedups while ensuring consistency with the outputs of the original models.\nHowever, there is currently a lack of theoretical guidance in speculative decoding. \nAs a result, most existing works are dualistic target-draft model paradigm, which significantly restricts the hinders potential application scenarios.\nIn this paper, we propose a polybasic speculative decoding framework supported by a solid theoretical foundation.\nWe first deduce a theorem to control the ideal inference time of speculative decoding systems which is then serve as a design criterion that effectively expands the original dualistic speculative decoding into a more efficient polybasic speculative decoding. \nWe further theoretically analyze the sampling process, identifying variables that can be optimized to enhance inference efficiency in multi-model systems.\nWe demonstrate, both theoretically and empirically, that this system accelerates inference for the target model, and that our approach is orthogonal to the majority of existing speculative methods, allowing for independent application or combination with other techniques. \nExperimentally, we conducted comprehensive evaluations across a wide range of models, including those from the Vicuna, LLaMA2-Chat, and LLaMA3 families. \nOur method achieved remarkable latency speedup ratios of $\\textbf{3.31$\\times$-4.01$\\times$}$ for LLaMA2-Chat 7B, up to $\\textbf{3.87$\\times$}$ for LLaMA3-8B, and up to $\\textbf{4.43$\\times$}$ for Vicuna-7B, while maintaining the distribution of the generated text. Code is available in supplementary materials.", "keywords": [ "Speculative Decoding; LLM Inference" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 1, 1, 2, 1 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "n72WC4a9qQ", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Daseul Bae", "gender": "Male", "institution": "Samsung SDS", "country": "", "position": "Researcher" }, { "name": "Jihoon Kim", "gender": "Male", "institution": "Samsung", "country": "", "position": "Researcher" }, { "name": "Jinho Choo", "gender": "Male", "institution": "Samsung", "country": "KR", "position": "Researcher" }, { "name": "Junhee Cho", "gender": "Male", "institution": "Samsung SDS", "country": "", "position": "Researcher" }, { "name": "Youngjune Gwon", "gender": "Male", "institution": "Samsung SDS", "country": "", "position": "Vice President" }, { "name": "Yeong Dae Kwon", "gender": "Male", "institution": "Samsung SDS", "country": "KR", "position": "Vice President" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 17, "n_ref": 31, "n_ref_all": 72, "n_fig": 12, "n_tab": 8, "L_tab": 7014, "n_element_tab": 864, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 149480, "len_all_1": 56735, "len_abs": 2308, "len_title": 140, "len_sents": 54233, "len_sents_1": 28981, "n_sents": 386, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1746, "title": "CAAP: Context-Aware Action Planning Prompting to Solve Computer Tasks with Front-End UI Only", "abs": "Software robots have long been used in Robotic Process Automation (RPA) to automate mundane and repetitive computer tasks. With the advent of Large Language Models (LLMs) and their advanced reasoning capabilities, these agents are now able to handle more complex or previously unseen tasks. However, LLM-based automation techniques in recent literature frequently rely on HTML source code for input or application-specific API calls for actions, limiting their applicability to specific environments. We propose an LLM-based agent that mimics human behavior in solving computer tasks. It perceives its environment solely through screenshot images, which are then converted into text for an LLM to process. By leveraging the reasoning capability of the LLM, we eliminate the need for large-scale human demonstration data typically required for model training. The agent only executes keyboard and mouse operations on Graphical User Interface (GUI), removing the need for pre-provided APIs to function. To further enhance the agent's performance in this setting, we propose a novel prompting strategy called Context-Aware Action Planning (CAAP) prompting, which enables the agent to thoroughly examine the task context from multiple perspectives. Our agent achieves an average success rate of 94.5% on MiniWoB++ and an average task score of 62.3 on WebShop, outperforming all previous studies of agents that rely solely on screen images. This method demonstrates potential for broader applications, particularly for tasks requiring coordination across multiple applications on desktops or smartphones, marking a significant advancement in the field of automation agents. Codes and models are accessible at https://github.com/caap-agent/caap-agent.", "keywords": [ "agent automation", "prompt engineering", "large language model" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "n6YVISFrcN", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Anirudh Gupta", "gender": "Male", "institution": "Gan Studios Pvt Limited", "country": "", "position": "Researcher" }, { "name": "Anirudh Mukherjee", "gender": "Male", "institution": "Gan AI", "country": "IN", "position": "Research Engineer" }, { "name": "Ankur Bhatia", "gender": "Male", "institution": "Gan Studio", "country": "IN", "position": "Researcher" }, { "name": "Ashwin Sankar", "gender": "Male", "institution": "AI4Bharat", "country": "IN", "position": "Researcher" }, { "name": "Mitesh M Khapra", "gender": "Male", "institution": "Indian Institute of Technology, Madras", "country": "IN", "position": "Associate Professor" }, { "name": "Praveen Srinivasa Varadhan", "gender": "Male", "institution": "Department of Computer Science, Indian Institute of Technology, Madras, Indian Institute of Technology, Madras", "country": "IN", "position": "PhD student" }, { "name": "Saloni Jaju", "gender": "Female", "institution": "Harvard Business School", "country": "US", "position": "MS student" }, { "name": "Shiva Kumar Marepally", "gender": "Male", "institution": "Indian Institute of Information Technology, Kancheepuram", "country": "", "position": "MS student" }, { "name": "Srija Anand", "gender": "Female", "institution": "Indian Institute of Technology, Madras", "country": "", "position": "MS student" }, { "name": "Suvrat Bhooshan", "gender": "unknown", "institution": "Gan.ai", "country": "IN", "position": "Principal Researcher" }, { "name": "amogh gulati", "gender": "Male", "institution": "gan.ai", "country": "IN", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 40, "n_ref_all": 70, "n_fig": 11, "n_tab": 5, "L_tab": 1880, "n_element_tab": 236, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1172, "n_element_tab_1": 202, "formula_len_all": 21, "formula_len_all_1": 0, "len_all": 141310, "len_all_1": 61765, "len_abs": 2371, "len_title": 124, "len_sents": 42444, "len_sents_1": 29414, "n_sents": 294, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1583, "title": "Rethinking MUSHRA: Addressing Modern Challenges in Text-to-Speech Evaluation", "abs": "Despite rapid advancements in TTS models, a consistent and robust human evaluation framework is still lacking. For example, MOS tests fail to differentiate between similar models, and CMOS's pairwise comparisons are time-intensive. The MUSHRA test is a promising alternative for evaluating multiple TTS systems simultaneously, but in this work we show that its reliance on matching human reference speech unduly penalises the scores of modern TTS systems that can exceed human speech quality. More specifically, we conduct a comprehensive assessment of the MUSHRA test, focusing on its sensitivity to factors such as rater variability, listener fatigue, and reference bias. Based on our extensive evaluation involving 471 human listeners across Hindi and Tamil we identify two primary shortcomings: (i) \\textit{reference-matching bias}, where raters are unduly influenced by the human reference, and (ii) \\textit{judgement ambiguity}, arising from a lack of clear fine-grained guidelines. To address these issues, we propose two refined variants of the MUSHRA test. The first variant enables fairer ratings for synthesized samples that surpass human reference quality. The second variant reduces ambiguity, as indicated by the relatively lower variance across raters. By combining these approaches, we achieve both more reliable and more fine-grained assessments. We also release MANGO, a massive dataset of 47,100 human ratings, the first-of-its-kind collection for Indian languages, aiding in analyzing human preferences and developing automatic metrics for evaluating TTS systems.", "keywords": [ "evaluation methodologies", "evaluation", "speech technologies", "datasets for low resource languages" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "n6To2wAOKL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Alexia Jolicoeur-Martineau", "gender": "Female", "institution": "Samsung - SAIT AI Lab, Montreal", "country": "CA", "position": "Researcher" }, { "name": "Anthony Gosselin", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "MS student" }, { "name": "Christopher J Pal", "gender": "unknown", "institution": "Polytechnique Montreal", "country": "CA", "position": "Full Professor" }, { "name": "Ge Ya Luo", "gender": "Female", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "ZhiHao Luo", "gender": "Male", "institution": "Polytechnique Montreal", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 28, "n_ref": 60, "n_ref_all": 77, "n_fig": 2, "n_tab": 4, "L_tab": 3334, "n_element_tab": 148, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 3412, "n_element_tab_1": 50, "formula_len_all": 239, "formula_len_all_1": 239, "len_all": 133275, "len_all_1": 73169, "len_abs": 1410, "len_title": 129, "len_sents": 32820, "len_sents_1": 27160, "n_sents": 345, "n_sents_1": 357, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1293, "title": "Ctrl-V: Higher Fidelity Video Generation with Bounding-Box Controlled Object Motion", "abs": "Controllable video generation has attracted significant attention, largely due to advances in video diffusion models. In domains like autonomous driving in particular it can be critical to develop highly accurate predictions for object motions. This paper tackles a crucial challenge of how to exert precise control over object motion for realistic video synthesis in a safety critical setting. To achieve this, we 1) use a separate, specialized model to predict object bounding-box trajectories given the past and optionally future locations of bounding boxes, and 2) generate video conditioned on these high quality trajectory predictions. This formulation allows us to test the quality of different model components separately and together. To address the challenges of conditioning video generation on object trajectories in settings where objects may disappear and appear within a scene, we propose an approach based on rendering 2D or 3D boxes as videos. Our method, Ctrl-V, leverages modified and fine-tuned Stable Video Diffusion (SVD) models to solve both trajectory and video generation. Extensive experiments conducted on the KITTI, Virtual-KITTI 2, BDD 100k, and nuScenes datasets validate the effectiveness of our approach in producing realistic and controllable video generation.", "keywords": [ "video generation", "video synthesis", "computer vision", "diffusion models", "autonomous driving", "controllable video generation" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 2, 5 ] }, { "paper_id": "n6KBvTQ10I", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Abhinav Shrivastava", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "Associate Professor" }, { "name": "Gihan Jayatilaka", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Matthew Gwilliam", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 53, "n_ref": 87, "n_ref_all": 132, "n_fig": 15, "n_tab": 7, "L_tab": 2288, "n_element_tab": 219, "n_fig_1": 11, "n_tab_1": 5, "L_tab_1": 1518, "n_element_tab_1": 174, "formula_len_all": 947, "formula_len_all_1": 653, "len_all": 158234, "len_all_1": 58385, "len_abs": 1225, "len_title": 141, "len_sents": 43020, "len_sents_1": 24717, "n_sents": 351, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1231, "title": "Utilization of Neighbor Information for Image Classification with Different Levels of Supervision", "abs": "We propose to bridge the gap between semi-supervised and unsupervised image recognition with a flexible method that performs well for both generalized category discovery (GCD) and image clustering. Despite the overlap in motivation between these tasks, the methods themselves are restricted to a single task \u2013 GCD methods are reliant on the labeled portion of the data, and deep image clustering methods have no built-in way to leverage the labels efficiently. We connect the two regimes with an innovative approach that Utilizes Neighbor Information for Classification (UNIC) both in the unsupervised (clustering) and semisupervised (GCD) setting. State-of-the-art clustering methods already rely heavily on nearest neighbors. We improve on their results substantially in two parts, first with a sampling and cleaning strategy where we identify accurate positive and negative neighbors, and secondly by finetuning the backbone with clustering losses computed by sampling both types of neighbors. We then adapt this pipeline to GCD by utilizing the labelled images as ground truth neighbors. Our method yields state-of-the-art results for both clustering (+3% ImageNet-100, Imagenet- 200) and GCD (+0.8% ImageNet-100, +5% CUB-200).", "keywords": [ "Generalized Category Discovery", "Image Clustering", "Image Classification" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "n6GemAoKMG", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Aleksandar Bojchevski", "gender": "Male", "institution": "University of Cologne", "country": "DE", "position": "Full Professor" }, { "name": "Jimin Cao", "gender": "Male", "institution": "Universit\u00e4t K\u00f6ln", "country": "DE", "position": "PhD student" }, { "name": "Mohammad Sadegh Akhondzadeh", "gender": "Male", "institution": "Universit\u00e4t K\u00f6ln", "country": "DE", "position": "PhD student" }, { "name": "Soroush H. Zargarbashi", "gender": "Male", "institution": "CISPA, saarland university, saarland informatics campus", "country": "DE", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 19, "n_ref": 50, "n_ref_all": 77, "n_fig": 8, "n_tab": 17, "L_tab": 1558, "n_element_tab": 315, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 232, "n_element_tab_1": 60, "formula_len_all": 403, "formula_len_all_1": 312, "len_all": 164703, "len_all_1": 71809, "len_abs": 1025, "len_title": 35, "len_sents": 65808, "len_sents_1": 34804, "n_sents": 573, "n_sents_1": 304, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1038, "title": "EvA: Evolutionary Attacks on Graphs", "abs": "Even a slight perturbation in the graph structure can cause a significant drop in the accuracy of graph neural networks (GNNs). Most existing attacks leverage gradient information to perturb edges. This relaxes the attack's optimization problem from a discrete to a continuous space, resulting in solutions far from optimal. It also restricts the adaptability of the attack to non-differentiable objectives. Instead, we propose an evolutionary-based algorithm to solve the discrete optimization problem directly. Our Evolutionary Attack (EvA) works with any black-box model and objective, eliminating the need for a differentiable proxy loss. This permits us to design two novel attacks that: reduce the effectiveness of robustness certificates and break conformal sets. We introduce a sparse encoding that results in memory complexity that is linear in the attack budget. \nEvA reduces the accuracy by an additional $\\sim$11\\% on average compared to the best previous attack, revealing significant untapped potential in designing attacks.", "keywords": [ "Adversarial_Attack", "Evasion_Attack", "Evolutionary_Algorithm", "Genetic_Algorithm" ], "rating_list": [ 5, 8, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "n64NYyc6rQ", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Hanwang Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Hao Fei", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Jiayi Ji", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Shengqiong Wu", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Xiangtai Li", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Tat-seng Chua", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 6, "n_ref_uni": 63, "n_ref": 156, "n_ref_all": 194, "n_fig": 14, "n_tab": 5, "L_tab": 3058, "n_element_tab": 425, "n_fig_1": 10, "n_tab_1": 6, "L_tab_1": 2020, "n_element_tab_1": 227, "formula_len_all": 836, "formula_len_all_1": 387, "len_all": 226457, "len_all_1": 65809, "len_abs": 1092, "len_title": 111, "len_sents": 48245, "len_sents_1": 29419, "n_sents": 353, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1061, "title": "Towards Semantic Equivalence of Tokenization in Multimodal LLM", "abs": "Multimodal Large Language Models (MLLMs) have demonstrated exceptional capabilities in processing vision-language tasks. One of the crux of MLLMs lies in vision tokenization, which involves efficiently transforming input visual signals into feature representations that are most beneficial for LLMs. However, existing vision tokenizers, essential for semantic alignment between vision and language, remain problematic. Existing methods aggressively fragment visual input, corrupting the visual semantic integrity. To address this, this paper proposes a novel dynamic Semantic-Equivalent Vision Tokenizer (SeTok), which groups visual features into semantic units via a dynamic clustering algorithm, flexibly determining the number of tokens based on image complexity. The resulting vision tokens effectively preserve semantic integrity and capture both low-frequency and high-frequency visual features. The proposed MLLM (Setokim) equipped with SeTok significantly demonstrates superior performance across various tasks, as evidenced by our experimental results.", "keywords": [ "MLLM", "Vision-Language Understaning", "Vision Generation", "Vision Editing", "Vision Tokenization" ], "rating_list": [ 6, 6, 5, 8, 6 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 4, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 5, 4, 4 ] }, { "paper_id": "n5PrId7pk5", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Alexandru I. Stere", "gender": "unknown", "institution": "The Boeing Company", "country": "US", "position": "Principal Researcher" }, { "name": "Carl Henrik Ek", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "Associate Professor" }, { "name": "Dragos D Margineantu", "gender": "Male", "institution": "Boeing Research and Technology", "country": "", "position": "AI Chief Technologist" }, { "name": "Erik Bodin", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "Henry Moss", "gender": "unknown", "institution": "Lancaster University", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 48, "n_formula_1": 13, "n_ref_uni": 29, "n_ref": 100, "n_ref_all": 153, "n_fig": 15, "n_tab": 3, "L_tab": 1604, "n_element_tab": 142, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 1016, "n_element_tab_1": 45, "formula_len_all": 3225, "formula_len_all_1": 744, "len_all": 171671, "len_all_1": 65352, "len_abs": 2762, "len_title": 134, "len_sents": 62704, "len_sents_1": 31379, "n_sents": 462, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1192, "title": "Linear combinations of Gaussian latents in generative models: interpolation and beyond", "abs": "Sampling from generative models has become a crucial tool for applications like data synthesis and augmentation. Diffusion, Flow Matching and Continuous Normalizing Flows have shown effectiveness across various modalities, and rely on Gaussian latent variables for generation. For search-based or creative applications that require additional control over the generation process, it has become common to manipulate the latent variable directly. However, existing approaches for performing such manipulations (e.g. interpolation or forming low-dimensional representations) only work well in special cases or are network or data-modality specific. We propose Combination of Gaussian variables (COG) as a general purpose method to form linear combinations of latent variables while adhering to the assumptions of the generative model. COG is easy to implement yet outperforms recent sophisticated methods for interpolation. As COG naturally addresses the broader task of forming linear combinations, new capabilities are afforded, including the construction of subspaces of the latent space, dramatically simplifying the creation of expressive low-dimensional spaces of high-dimensional objects.", "keywords": [ "generative models", "diffusion models", "latent space interpolation", "latent representations" ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 4, 3, 2, 2 ], "presentation_list": [ 4, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "n4wcdct43X", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Carlo Alfano", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Chris Lu", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Patrick Rebeschini", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Sebastian Rene Towers", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "", "position": "PhD student" }, { "name": "Silvia Sapora", "gender": "Female", "institution": "University of Oxford", "country": "", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 19, "n_ref_uni": 52, "n_ref": 103, "n_ref_all": 129, "n_fig": 4, "n_tab": 3, "L_tab": 607, "n_element_tab": 76, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3086, "formula_len_all_1": 1385, "len_all": 146001, "len_all_1": 70179, "len_abs": 1411, "len_title": 93, "len_sents": 43661, "len_sents_1": 33275, "n_sents": 302, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 45, "L_abs": 1423, "title": "Learning mirror maps in policy mirror descent", "abs": "Policy Mirror Descent (PMD) is a popular framework in reinforcement learning, serving as a unifying perspective that encompasses numerous algorithms. These algorithms are derived through the selection of a mirror map and enjoy finite-time convergence guarantees. Despite its popularity, the exploration of PMD's full potential is limited, with the majority of research focusing on a particular mirror map---namely, the negative entropy---which gives rise to the renowned Natural Policy Gradient (NPG) method. It remains uncertain from existing theoretical studies whether the choice of mirror map significantly influences PMD's efficacy. In our work, we conduct empirical investigations to show that the conventional mirror map choice (NPG) often yields less-than-optimal outcomes across several standard benchmark environments. Using evolutionary strategies, we identify more efficient mirror maps that enhance the performance of PMD. We first focus on a tabular environment, i.e.\\ Grid-World, where we relate existing theoretical bounds with the performance of PMD for a few standard mirror maps and the learned one. We then show that it is possible to learn a mirror map that outperforms the negative entropy in more complex environments, such as the MinAtar suite. Additionally, we demonstrate that the learned mirror maps generalize effectively to different tasks by testing each map across various other environments.", "keywords": [ "Policy optimization", "mirror descent" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "n4SLaq5GhM", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 1, "author_info_list": [ { "name": "Binxi Xie", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Undergrad student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 5, "n_ref": 9, "n_ref_all": 10, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 71880, "len_all_1": 71880, "len_abs": 1005, "len_title": 186, "len_sents": 33316, "len_sents_1": 33316, "n_sents": 198, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 138, "L_abs": 1008, "title": "Enhancing Medical NLP Systems: Integrating Upstash Vector and BGE-M3 for Accurate and Ethical Healthcare Data Management with Reduced Bias", "abs": "This paper proposes a novel NLP model in healthcare by including Utash Vector for in-time and contextual information retrieval and BGE-M3 for advanced understanding. The model overcomes the challenges posed by the existing systems, such as incomplete data retrieval, a semantically inconsistent database, and algorithm bias. Incorporating bias mitigation measures and fairness audits, it guarantees no unfair treatment of patients belonging to different groups. Aligned with the AMA Code of Medical Ethics, provides proper management of Electronic Health Records in better ways in terms of transparency, confidentiality, and accuracy. Although these problems are relieved, the accuracy of information is still a major issue, the abuse of artificial intelligence remains a risk, and the use of the AMA Code to guide the integration of artificial intelligence has its limitations. Each of these must operate with defensible use of AI and auditing as well as explanation of AI usage in clinical decision-making.", "keywords": [ "Medical NLP", "Upstash Vector", "BGE-M3 model", "Real-time data retrieval", "Semantic understanding", "Bias mitigation", "Healthcare AI ethics" ], "rating_list": [ 3, 3, 6, 1 ], "soundness_list": [ 1, 1, 3, 1 ], "presentation_list": [ 2, 1, 3, 1 ], "contribution_list": [ 2, 1, 3, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "n4HH7g9hxk", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Bahareh Tasdighi", "gender": "Female", "institution": "University of Southern Denmark - SDU", "country": "", "position": "PhD student" }, { "name": "Manuel Haussmann", "gender": "unknown", "institution": "University of Southern Denmark - SDU", "country": "DK", "position": "Postdoc" }, { "name": "Melih Kandemir", "gender": "Male", "institution": "University of Southern Denmark", "country": "DK", "position": "Associate Professor" }, { "name": "Nicklas Werge", "gender": "unknown", "institution": "University of Southern Denmark - SDU", "country": "", "position": "Postdoc" }, { "name": "Yi-Shan Wu", "gender": "Non-Binary", "institution": "University of Southern Denmark - SDU", "country": "DK", "position": "Postdoc" } ], "n_formula": 48, "n_formula_1": 23, "n_ref_uni": 49, "n_ref": 123, "n_ref_all": 154, "n_fig": 7, "n_tab": 6, "L_tab": 3028, "n_element_tab": 270, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 45, "n_element_tab_1": 12, "formula_len_all": 5916, "formula_len_all_1": 2050, "len_all": 190129, "len_all_1": 69388, "len_abs": 1414, "len_title": 79, "len_sents": 55277, "len_sents_1": 32771, "n_sents": 435, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 31, "L_abs": 1426, "title": "Deep Exploration with PAC-Bayes", "abs": "Reinforcement learning for continuous control under sparse rewards is an under-explored problem despite its significance in real life. Many complex skills build on intermediate ones as prerequisites. For instance, a humanoid locomotor has to learn how to stand before it can learn to walk. To cope with reward sparsity, a reinforcement learning agent has to perform deep exploration. However, existing deep exploration methods are designed for small discrete action spaces, and their successful generalization to state-of-the-art continuous control remains unproven. We address the deep exploration problem for the first time from a PAC-Bayesian perspective in the context of actor-critic learning. To do this, we quantify the error of the Bellman operator through a PAC-Bayes bound, where a bootstrapped ensemble of critic networks represents the posterior distribution, and their targets serve as a data-informed function-space prior. \nWe derive an objective function from this bound and use it to train the critic ensemble. Each critic trains an individual actor network, implemented as a shared trunk and critic-specific heads. The agent performs deep exploration by acting deterministically on a randomly chosen actor head. Our proposed algorithm, named PAC-Bayesian Actor-Critic (PBAC), is the only algorithm to successfully discover sparse rewards on a diverse set of continuous control tasks with varying difficulty.", "keywords": [ "Reinforcement learning", "deep exploration", "sparse rewards", "PAC Bayes" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "n3TkrH7fEr", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Chenglong Bao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Assistant Professor" }, { "name": "Defeng Sun", "gender": "unknown", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Shulan Zhu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yancheng Yuan", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 192, "n_formula_1": 35, "n_ref_uni": 17, "n_ref": 51, "n_ref_all": 62, "n_fig": 5, "n_tab": 6, "L_tab": 1937, "n_element_tab": 97, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 2, "n_element_tab_1": 2, "formula_len_all": 30299, "formula_len_all_1": 4224, "len_all": 218452, "len_all_1": 60162, "len_abs": 1351, "len_title": 170, "len_sents": 62947, "len_sents_1": 23588, "n_sents": 679, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 122, "L_abs": 1097, "title": "A Tight Convergence Analysis of Inexact Stochastic Proximal Point Algorithm for Stochastic Composite Optimization Problems", "abs": "The \\textbf{i}nexact \\textbf{s}tochastic \\textbf{p}roximal \\textbf{p}oint \\textbf{a}lgorithm (isPPA) is popular for solving stochastic composite optimization problems with many applications in machine learning. While the convergence theory of the (inexact) PPA has been well established, the known convergence guarantees of isPPA require restrictive assumptions. In this paper, we establish the stability and almost sure convergence of isPPA under mild assumptions, where smoothness and (restrictive) strong convexity of the objective function are not required. Imposing a local Lipschitz condition on component functions and a quadratic growth condition on the objective function, we establish last-iterate iteration complexity bounds of isPPA regarding the distance to the solution set and the Karush\u2013Kuhn\u2013Tucker (KKT) residual. Moreover, we show that the established iteration complexity bounds are tight up to a constant by explicitly analyzing the bounds for the regularized Fr\\'echet mean problem. We further validate the established convergence guarantees of isPPA by numerical experiments.", "keywords": [ "inexact stochastic proximal point method", "stochastic composite optimization", "quadratic growth", "rate of convergence" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "n34taxF0TC", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Achille Fokoue", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Anak Agung Julius", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "Full Professor" }, { "name": "Debarun Bhattacharjya", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Ronny Luss", "gender": "unknown", "institution": "IBM", "country": "US", "position": "Research Scientist" }, { "name": "Tengfei Ma", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Assistant Professor" }, { "name": "Yunshi Wen", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 9, "n_ref_uni": 36, "n_ref": 70, "n_ref_all": 109, "n_fig": 14, "n_tab": 14, "L_tab": 11537, "n_element_tab": 1294, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 714, "n_element_tab_1": 140, "formula_len_all": 1008, "formula_len_all_1": 549, "len_all": 176530, "len_all_1": 67571, "len_abs": 1251, "len_title": 130, "len_sents": 47359, "len_sents_1": 29888, "n_sents": 373, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1256, "title": "Shedding Light on Time Series Classification using Interpretability Gated Networks", "abs": "In time-series classification, interpretable models can bring additional insights but be outperformed by deep models since human-understandable features have limited expressivity and flexibility. In this work, we present InterpGN, a framework that integrates an interpretable model and a deep neural network. Within this framework, we introduce a novel gating function design based on the confidence of the interpretable expert, preserving interpretability for samples where interpretable features are significant while also identifying samples that require additional expertise. For the interpretable expert, we incorporate shapelets to effectively model shape-level features for time-series data. We introduce a variant of Shapelet Transforms to build logical predicates using shapelets. Our proposed model achieves comparable performance with state-of-the-art deep learning models while additionally providing interpretable classifiers for various benchmark datasets. We further show that our models improve on quantitative shapelet quality and interpretability metrics over existing shapelet-learning formulations. Finally, we demonstrate the capability of our models to provide interpretability in a real-world application using the MIMIC-III dataset.", "keywords": [ "Interpretability", "Time-series", "Shapelet" ], "rating_list": [ 8, 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 4 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "n2xueVy5ek", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Fred Heiding", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Simon Lermen", "gender": "Male", "institution": "Technische Universit\u00e4t Berlin", "country": "", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 62, "n_ref_all": 102, "n_fig": 3, "n_tab": 6, "L_tab": 921, "n_element_tab": 194, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 663, "n_element_tab_1": 123, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 163916, "len_all_1": 57451, "len_abs": 1196, "len_title": 88, "len_sents": 55013, "len_sents_1": 28295, "n_sents": 398, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1200, "title": "Evaluating Large Language Models' Capability to Conduct Cyberattacks On Embedded Devices", "abs": "As large language models continue to evolve, they have the potential to automate and enhance various aspects of computer security, including red teaming assessments. In this article, we conduct 32 computer security attacks and compare their success rates when performed manually and with assistance from large language models. The security assessments target five connected devices commonly found in modern households (two door locks, one vacuum cleaner, one garage door, and one smart vehicle adapter). We use attacks such as denial-of-service attacks, Man-in-the-Middle, authentication brute force, malware creation, and other common attack types. Each attack was performed twice, once by a human and once by an LLM, and scored for damage, reproducibility, exploitability, affected users, and discoverability based on the DREAD framework for computer security risk assessments. For the LLM-assisted attacks, we also scored the LLM's capacity to perform the attack autonomously. LLMs regularly increased the reproducibility and exploitability of attacks, but no LLM-based attack enhanced the damage inflicted on the device, and the language models often required manual input to complete the attack.", "keywords": [ "Computer security", "red teaming", "IoT", "large language models" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 1 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "n2VZtv8tqL", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 9, "author_info_list": [ { "name": "Cheng-Ching Tseng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Chun-Kai Fan", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Gaole Dai", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Qizhe Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Tiejun Huang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Yiming Tang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yulu Gan", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zhi Zhang", "gender": "Male", "institution": "Institute for Logic, Language and Computation, University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Shanghang Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 19, "n_ref": 24, "n_ref_all": 33, "n_fig": 4, "n_tab": 2, "L_tab": 2719, "n_element_tab": 202, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2719, "n_element_tab_1": 202, "formula_len_all": 692, "formula_len_all_1": 690, "len_all": 97302, "len_all_1": 57367, "len_abs": 1369, "len_title": 112, "len_sents": 27401, "len_sents_1": 24149, "n_sents": 196, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1383, "title": "Discovering Long-Term Effects on Parameter Efficient Fine-tuning", "abs": "Pre-trained Artificial Neural Networks (ANNs) demonstrate robust pattern recognition abilities, closely mirroring the functionality of Biological Neural Networks (BNNs). We are particularly intrigued by these models' capacity for acquiring new knowledge through fine-tuning, such, Parameter-efficient Fine-tuning (PEFT). Given that both ANNs and BNNs propagate information layer-by-layer, a useful analogy can be drawn: ANN weights correspond to synapses in BNNs, while features (latent variables or activations) parallel the neurotransmitters released by neurons. Building upon this clue, we delve deeper into exploring the connections between feature adjustment and weight adjustment, resulting in our proposed method Synapses \\& Neurons (SAN) that learns scaling matrices for features and propagates their effects towards posterior weight matrices. Our approach draws strong inspiration from well-known neuroscience phenomena - Long-term Potentiation (LTP) and Long-term Depression (LTD), which also reveal the relationship between synapse development and neurotransmitter release levels. We conducted extensive comparisons of PEFT on 26 datasets using attention-based networks as well as convolution-based networks, leading to significant improvements compared to other tuning methods, +8.5\\% over fully-finetune, +7\\% over Visual Prompt Tuning, and +3.2\\% over Low-Rank Adapter.", "keywords": [ "Bio-inspried", "PEFT", "transfer learning" ], "rating_list": [ 3, 5, 8, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 2, 2, 4, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "n2NidsYDop", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Juno Kim", "gender": "Male", "institution": "University of Tokyo", "country": "JP", "position": "MS student" }, { "name": "Taiji Suzuki", "gender": "Male", "institution": "The University of Tokyo", "country": "", "position": "Full Professor" } ], "n_formula": 126, "n_formula_1": 30, "n_ref_uni": 39, "n_ref": 69, "n_ref_all": 90, "n_fig": 5, "n_tab": 5, "L_tab": 536, "n_element_tab": 75, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 13296, "formula_len_all_1": 2916, "len_all": 195277, "len_all_1": 66419, "len_abs": 1647, "len_title": 116, "len_sents": 51042, "len_sents_1": 28711, "n_sents": 450, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 68, "L_abs": 1276, "title": "Transformers Provably Solve Parity Efficiently with Chain of Thought", "abs": "This work provides the first theoretical analysis of training transformers to solve complex problems by recursively generating intermediate states, analogous to fine-tuning for chain-of-thought (CoT) reasoning. We consider training a one-layer transformer to solve the fundamental $k$-parity problem, extending the work on RNNs by \\citet{Wies23}. We establish three key results: (1) any finite-precision gradient-based algorithm, without intermediate supervision, requires substantial iterations to solve parity with finite samples. (2) In contrast, when intermediate parities are incorporated into the loss function, our model can learn parity in one gradient update when aided by \\emph{teacher forcing}, where ground-truth labels of the reasoning chain are provided at each generation step. (3) Even without teacher forcing, where the model must generate CoT chains end-to-end, parity can be learned efficiently if augmented data is employed to internally verify the soundness of intermediate steps. Our findings, supported by numerical experiments, show that task decomposition and stepwise reasoning naturally arise from optimizing transformers with CoT; moreover, self-consistency checking can improve multi-step reasoning ability, aligning with empirical studies of CoT.", "keywords": [ "transformers", "chain of thought", "parity", "self-consistency" ], "rating_list": [ 8, 8, 10 ], "soundness_list": [ 4, 3, 4 ], "presentation_list": [ 4, 3, 4 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "n2EU4PUrJP", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Anat Kleiman", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Gintare Karolina Dziugaite", "gender": "Female", "institution": "McGill University", "country": "CA", "position": "Adjunct Professor" }, { "name": "Jonathan Frankle", "gender": "Male", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "Mansheej Paul", "gender": "Male", "institution": "Databricks, Databricks", "country": "US", "position": "Researcher" }, { "name": "Sham M. Kakade", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 7, "n_ref_uni": 35, "n_ref": 64, "n_ref_all": 105, "n_fig": 9, "n_tab": 7, "L_tab": 2174, "n_element_tab": 226, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 2779, "n_element_tab_1": 31, "formula_len_all": 1111, "formula_len_all_1": 321, "len_all": 155096, "len_all_1": 60431, "len_abs": 1264, "len_title": 129, "len_sents": 41894, "len_sents_1": 28802, "n_sents": 297, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1256, "title": "Soup to go: mitigating forgetting during continual learning with model averaging", "abs": "In continual learning with pretrained large language models (LLMs), where data from instruction fine-tuning (IFT) tasks arrives in a sequence, fine-tuning on later tasks will often lead to performance degradation on earlier tasks. \nThis is especially pronounced when the IFT tasks come from diverse domains.\nIn this setting, how can we mitigate catastrophic forgetting of earlier tasks and retain what the LLM has learned?\nInspired by a classical continual learning method, L2 penalty to previous weights, we propose Sequential Fine-tuning with Averaging (SFA), a method that merges models with earlier checkpoints trained on previous tasks during the course of training. \nSOTA approaches typically maintain a data buffer of past tasks or impose a penalty at each gradient step. However, our method achieves comparable results without the need to store past data, or multiple copies of parameters for each gradient step. \nFurthermore, our method outperforms penalty methods like L2 regression and EWC, as well as other common merging techniques such as Task Arithmetic, and TIES Merging.\nFinally, we show that using our method, a single model can simultaneously perform well on a range of fine-tuning tasks in diverse domains, including Math, Law and Code.", "keywords": [ "Continual learning", "model merging", "fine-tuning" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "n20n1hojPg", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Bin Wu", "gender": "unknown", "institution": "Tencent AI Lab", "country": "", "position": "Principal Researcher" }, { "name": "Chao Zhan", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "JUNXIN HUANG", "gender": "Male", "institution": "TME", "country": "CN", "position": "Researcher" }, { "name": "Minhao LIU", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Researcher" }, { "name": "Wenjiang Zhou", "gender": "Male", "institution": "University of New South Wales", "country": "", "position": "MS student" }, { "name": "Yubin zeng", "gender": "Male", "institution": "TME", "country": "CN", "position": "Researcher" }, { "name": "Yue Zhang", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Zhaokang Chen", "gender": "Male", "institution": "Tencent Music Entertainment Lyra Lab", "country": "CN", "position": "Researcher" }, { "name": "\u8d3a\u82f1\u6770", "gender": "Male", "institution": "Tencent Music Entertaining", "country": "", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 34, "n_ref": 80, "n_ref_all": 103, "n_fig": 5, "n_tab": 5, "L_tab": 1125, "n_element_tab": 146, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 786, "n_element_tab_1": 83, "formula_len_all": 302, "formula_len_all_1": 304, "len_all": 123077, "len_all_1": 65569, "len_abs": 1341, "len_title": 130, "len_sents": 37560, "len_sents_1": 32080, "n_sents": 253, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1347, "title": "MuseTalk: Real-Time High Quality Lip Synchronization with Latent Space Inpainting", "abs": "Achieving high-resolution, identity consistency, and accurate lip-speech synchronization in face visual dubbing presents significant challenges, particularly for real-time applications like live video streaming. We propose MuseTalk, which generates lip-sync targets in a latent space encoded by a Variational Autoencoder, enabling high-fidelity talking face video generation with efficient inference. Specifically, we project the occluded lower half of the face image and itself as an reference into a low-dimensional latent space and use a multi-scale U-Net to fuse audio and visual features at various levels. We further propose a novel sampling strategy during training, which selects reference images with head poses closely matching the target, allowing the model to focus on precise lip movement by filtering out redundant information. Additionally, we analyze the mechanism of lip-sync loss and reveal its relationship with input information volume. Extensive experiments show that MuseTalk consistently outperforms recent state-of-the-art methods in visual fidelity and achieves comparable lip-sync accuracy. As MuseTalk supports the online generation of face at 256x256 at more than 30 FPS with negligible starting latency, it paves the way for real-time applications. The codes and models will be made publicly available upon acceptance.", "keywords": [ "talking face", "face visual dubbing", "generative models", "multimodality", "AI-generated content" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "n1X2n7MJ8L", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Bill Y Lin", "gender": "Male", "institution": "xAI", "country": "US", "position": "Researcher" }, { "name": "Chan Young Park", "gender": "Female", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "Postdoc" }, { "name": "Liwei Jiang", "gender": "Female", "institution": "University of Washington", "country": "", "position": "PhD student" }, { "name": "Maria Antoniak", "gender": "unknown", "institution": "Copenhagen University", "country": "DK", "position": "Postdoc" }, { "name": "Mehar Bhatia", "gender": "Female", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "PhD student" }, { "name": "Sahithya Ravi", "gender": "Female", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Shuyue Stella Li", "gender": "Female", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Vered Shwartz", "gender": "Female", "institution": "University of British Columbia", "country": "CA", "position": "Assistant Professor" }, { "name": "Yejin Choi", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Yu Ying Chiu", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "MS student" }, { "name": "Yulia Tsvetkov", "gender": "Female", "institution": "Department of Computer Science, University of Washington", "country": "", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 35, "n_ref": 59, "n_ref_all": 86, "n_fig": 9, "n_tab": 9, "L_tab": 2289, "n_element_tab": 246, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 622, "n_element_tab_1": 68, "formula_len_all": 0, "formula_len_all_1": 4, "len_all": 132966, "len_all_1": 59678, "len_abs": 2222, "len_title": 160, "len_sents": 36632, "len_sents_1": 28361, "n_sents": 275, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 112, "L_abs": 1515, "title": "CulturalBench: a Robust, Diverse and Challenging Benchmark on Measuring (the Lack of) Cultural Knowledge of LLMs", "abs": "To make large language models (LLMs) more helpful across diverse cultures, it is essential to have effective cultural knowledge benchmarks to measure and track our progress. Effective benchmarks need to be robust, diverse, and challenging. We introduce CulturalBench: a set of 1,227 human-written and human-verified questions for effectively assessing LLMs' cultural knowledge, covering 45 global regions including the underrepresented ones like Bangladesh, Zimbabwe, and Peru. Questions - each verified by five independent annotators - span 17 diverse topics ranging from food preferences to greeting etiquettes. We evaluate models on two setups: CulturalBench-Easy and CulturalBench-Hard which share the same questions but asked differently. We find that LLMs are sensitive to such difference in setups (e.g., GPT-4o with 27.3% difference). Compared to human performance (92.6% accuracy), CulturalBench-Hard is more challenging for frontier LLMs with the best performing model (GPT-4o) at only 61.5% and the worst (Llama3-8b) at 21.4%. Moreover, we find that LLMs often struggle with tricky questions that have multiple correct answers (e.g., What utensils do the Chinese usually use?), revealing a tendency to converge to a single answer. Our results also indicate that OpenAI GPT-4o substantially outperform other proprietary and open source models in questions related to all but one region (Oceania). Nonetheless, all models consistently underperform on questions related to South America and the Middle East.", "keywords": [ "cultural knowledge evaluation", "cultural reasoning", "large language models" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 5 ] }, { "paper_id": "n0lXaskyk5", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Behrooz Tahmasebi", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Stefanie Jegelka", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" } ], "n_formula": 92, "n_formula_1": 28, "n_ref_uni": 31, "n_ref": 57, "n_ref_all": 66, "n_fig": 3, "n_tab": 3, "L_tab": 351, "n_element_tab": 17, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5558, "formula_len_all_1": 1338, "len_all": 141299, "len_all_1": 64301, "len_abs": 1119, "len_title": 132, "len_sents": 40630, "len_sents_1": 28506, "n_sents": 383, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1121, "title": "Generalization Bounds for Canonicalization: A Comparative Study with Group Averaging", "abs": "Canonicalization, a popular method for generating invariant or equivariant function classes from arbitrary function sets, involves initial data projection onto a reduced input space subset, followed by applying any learning method to the projected dataset. Despite recent research on the expressive power and continuity of functions represented by canonicalization, its generalization capabilities remain less explored. This paper addresses this gap by theoretically examining the generalization benefits and sample complexity of canonicalization, comparing them with group averaging, another popular technique for creating invariant or equivariant function classes. Our findings reveal two distinct regimes where canonicalization may outperform or underperform compared to group averaging, with precise quantification of this phase transition in terms of sample size and group action characteristics. To the best of our knowledge, this study represents the first theoretical exploration of such behavior, offering insights into the relative effectiveness of canonicalization and group averaging under varying conditions.", "keywords": [ "invariances", "group", "symmetry", "canonicalization" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "n0YCAMVh8b", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Eldad Haber", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Full Professor" }, { "name": "Moshe Eliasof", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "Niloufar Zakariaei", "gender": "Female", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Shadab Ahamed", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" } ], "n_formula": 45, "n_formula_1": 29, "n_ref_uni": 37, "n_ref": 62, "n_ref_all": 107, "n_fig": 14, "n_tab": 12, "L_tab": 1514, "n_element_tab": 133, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2174, "formula_len_all_1": 1359, "len_all": 167605, "len_all_1": 70375, "len_abs": 1130, "len_title": 100, "len_sents": 53676, "len_sents_1": 29367, "n_sents": 457, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1136, "title": "Multiscale Training of Convolutional Neural Networks", "abs": "Convolutional Neural Networks (CNNs) are the backbone of many deep learning methods, but optimizing them remains computationally expensive. To address this, we explore multiscale training frameworks and mathematically identify key challenges, particularly when dealing with noisy inputs. Our analysis reveals that in the presence of noise, the gradient of standard CNNs in multiscale training may fail to converge as the mesh-size approaches to $0$, undermining the optimization process. This insight drives the development of Mesh-Free Convolutions (MFCs), which are independent of input scale and avoid the pitfalls of traditional convolution kernels. We demonstrate that MFCs, with their robust gradient behavior, ensure convergence even with noisy inputs, enabling more efficient neural network optimization in multiscale settings. To validate the generality and effectiveness of our multiscale training approach, we show that (i) MFCs can theoretically deliver substantial computational speedups without sacrificing performance in practice, and (ii) standard convolutions benefit from our multiscale training framework in practice.", "keywords": [ "Multilevel Stochastic Gradient Descent", "Multiscale Training", "Mesh Free Convolutions" ], "rating_list": [ 6, 3, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "n0OtGl6VGb", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Amrita Saha", "gender": "Female", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Doyen Sahoo", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Researcher" }, { "name": "Hanze Dong", "gender": "Male", "institution": "SalesForce", "country": "SG", "position": "Researcher" }, { "name": "Lei Wang", "gender": "Male", "institution": "SalesForce", "country": "SG", "position": "Researcher" }, { "name": "Xudong Lu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yuhui Xu", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Research Scientist" }, { "name": "Aojun Zhou", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Researcher" }, { "name": "Allan Jie", "gender": "Male", "institution": "Salesforce Research", "country": "", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 45, "n_ref": 91, "n_ref_all": 129, "n_fig": 9, "n_tab": 7, "L_tab": 9399, "n_element_tab": 1184, "n_fig_1": 8, "n_tab_1": 9, "L_tab_1": 14878, "n_element_tab_1": 1756, "formula_len_all": 831, "formula_len_all_1": 684, "len_all": 182784, "len_all_1": 94667, "len_abs": 1482, "len_title": 96, "len_sents": 38930, "len_sents_1": 25322, "n_sents": 268, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 48, "L_abs": 1489, "title": "ThinK: Thinner Key Cache by Query-Driven Pruning", "abs": "Large Language Models (LLMs) have revolutionized the field of natural language processing, achieving unprecedented performance across a variety of applications. \nHowever, their increased computational and memory demands present significant challenges, especially when handling long sequences.\nThis paper focuses on the long-context scenario, addressing the inefficiencies in KV cache memory consumption during inference. \nUnlike existing approaches that optimize the memory based on the sequence length, we identify substantial redundancy in the channel dimension of the KV cache, as indicated by an uneven magnitude distribution and a low-rank structure in the attention weights.\nIn response, we propose ThinK, a novel query-dependent KV cache pruning method designed to minimize attention weight loss while selectively pruning the least significant channels. Our approach not only maintains or enhances model accuracy but also achieves a reduction in KV cache memory costs by over 20% compared with vanilla KV cache eviction and quantization methods. For instance, ThinK integrated with KIVI can achieve 2.8x peak memory reduction while maintaining nearly the same quality, enabling a batch size increase from 4x (with KIVI alone) to 5x when using a single GPU. Extensive evaluations on the LLaMA and Mistral models across various long-sequence datasets verified the efficiency of \\our, establishing a new baseline algorithm for efficient LLM deployment without compromising performance.", "keywords": [ "Large Language Models; KV Cache Compression; KV Cache Pruning" ], "rating_list": [ 8, 5, 8, 8, 5 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 5, 5, 3, 4, 4 ] }, { "paper_id": "mzL19kKE3r", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Daeshik Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Donggon Jang", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Suin Lee", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Taehyeon Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Yucheol Cho", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 56, "n_ref": 134, "n_ref_all": 151, "n_fig": 9, "n_tab": 8, "L_tab": 2757, "n_element_tab": 245, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1012, "n_element_tab_1": 33, "formula_len_all": 335, "formula_len_all_1": 101, "len_all": 190563, "len_all_1": 64853, "len_abs": 1899, "len_title": 130, "len_sents": 47096, "len_sents_1": 31860, "n_sents": 349, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 98, "L_abs": 1909, "title": "MMR: A Large-scale Benchmark Dataset for Multi-target and Multi-granularity Reasoning Segmentation", "abs": "The fusion of Large Language Models (LLMs) with vision models is pioneering new possibilities in user-interactive vision-language tasks. A notable application is reasoning segmentation, where models generate pixel-level segmentation masks by comprehending implicit meanings in human instructions. However, seamless human-AI interaction demands more than just object-level recognition; it requires understanding both objects and the functions of their detailed parts, particularly in multi-target scenarios. For example, when instructing a robot to \"turn on the TV,\" there could be various ways to accomplish this command. Recognizing multiple objects capable of turning on the TV, such as the TV itself or a remote control (multi-target), provides more flexible options and aids in finding the optimized scenario. Furthermore, understanding specific parts of these objects, like the TV's button or the remote's button (part-level), is important for completing the action. Unfortunately, current reasoning segmentation datasets predominantly focus on a single target object-level reasoning, which limits the detailed recognition of an object's parts in multi-target contexts. To address this gap, we construct a large-scale dataset called Multi-target and Multi-granularity Reasoning (MMR). MMR comprises 194K complex and implicit instructions that consider multi-target, object-level, and part-level aspects, based on pre-existing image-mask sets. This dataset supports diverse and context-aware interactions by hierarchically providing object and part information. Moreover, we propose a straightforward yet effective framework for multi-target, object-level, and part-level reasoning segmentation. Experimental results on MMR show that the proposed method can reason effectively in multi-target and multi-granularity scenarios, while the existing reasoning segmentation model still has room for improvement.", "keywords": [ "Multimodal Dataset", "Multi-target and Multi-granularity Reasoning Segmentation", "Benchmark Framework" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "mzJAupYURK", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Fu-Yun Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Zhengyang Geng", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 19, "n_ref_uni": 50, "n_ref": 116, "n_ref_all": 130, "n_fig": 6, "n_tab": 3, "L_tab": 2820, "n_element_tab": 351, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 961, "n_element_tab_1": 91, "formula_len_all": 1926, "formula_len_all_1": 1821, "len_all": 155122, "len_all_1": 62769, "len_abs": 1492, "len_title": 121, "len_sents": 30634, "len_sents_1": 26550, "n_sents": 236, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1182, "title": "Stable Consistency Tuning: Understanding and Improving Consistency Models", "abs": "Diffusion models achieve superior generation quality but suffer from slow generation speed due to the iterative nature of denoising. In contrast, consistency models, a new generative family, achieve competitive performance with significantly faster sampling. \nThese models are trained either through consistency distillation, which leverages pretrained diffusion models, or consistency training/tuning directly from raw data. \nIn this work, we propose a novel framework for understanding consistency models by modeling the denoising process of the diffusion model as a Markov Decision Process (MDP) and framing consistency model training as the value estimation through Temporal Difference (TD) Learning. \nMore importantly, this framework allows us to analyze the limitations of current consistency training/tuning strategies.\nBuilt upon Easy Consistency Tuning (ECT), we propose Stable Consistency Tuning (SCT), which incorporates variance-reduced learning using the score identity.\nSCT leads to significant performance improvements on benchmarks such as CIFAR-10 and ImageNet-64. On ImageNet-64, SCT achieves 1-step FID 2.42 and 2-step FID 1.55, a new SoTA for consistency models.", "keywords": [ "Diffusion models", "Consistency models" ], "rating_list": [ 1, 3, 3, 5 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "mz8unSsSsB", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chuang Gan", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Han Yang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Undergrad student" }, { "name": "Hongxin Zhang", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" }, { "name": "Jiachen Zhou", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "MS student" }, { "name": "Peihao Chen", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yilun Du", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Yuncong Yang", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 56, "n_ref_all": 83, "n_fig": 11, "n_tab": 4, "L_tab": 492, "n_element_tab": 62, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 472, "n_element_tab_1": 61, "formula_len_all": 154, "formula_len_all_1": 300, "len_all": 149152, "len_all_1": 65785, "len_abs": 1669, "len_title": 126, "len_sents": 49417, "len_sents_1": 31538, "n_sents": 356, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1675, "title": "SnapMem: Snapshot-based 3D Scene Memory for Embodied Exploration and Reasoning", "abs": "Constructing a compact and informative 3D scene representation is essential for effective embodied reasoning and exploration, especially in complex environments over long periods. Existing approaches have relied on object-centric graph representations, which oversimplify 3D scenes by modeling them as individual objects and describing inter-object relationships through rigid textual descriptions. This rigidity leads to the loss of rich spatial relationships between objects, which are essential for embodied scene reasoning tasks. Furthermore, these representations lack natural mechanisms for active exploration and memory management, which hampers their applications for lifelong autonomy. In this work, we propose SnapMem, a novel 3D scene representation that leverages a compact set of informative snapshot images to cover the scene based on object co-visibility. These snapshot images capture rich spatial and semantic information among objects within the same view and their surroundings. We then illustrate how such a representation can be directly integrated with frontier-based exploration algorithms to facilitate active exploration by leveraging unexplored regions and scene memory. To support lifelong memory in active exploration settings, we further present an efficient memory aggregation pipeline to incrementally construct SnapMem, as well as an effective memory retrieval technique for memory management. Experimental results over three benchmarks demonstrate that SnapMem significantly enhances agents' reasoning and exploration capabilities in 3D environments over extended periods, highlighting its potential for advancing applications in embodied AI.", "keywords": [ "Embodied AI", "Vision-Language", "Lifelong Learning" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "mypnFcBbz4", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Chaohao Yuan", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Ercan Engin KURUOGLU", "gender": "Male", "institution": "Tsinghua-Berkeley Shenzhen Institute ", "country": "CN", "position": "Full Professor" }, { "name": "Suchanuch Piriyasatit", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 17, "n_ref_uni": 45, "n_ref": 68, "n_ref_all": 78, "n_fig": 3, "n_tab": 3, "L_tab": 1383, "n_element_tab": 122, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1232, "n_element_tab_1": 91, "formula_len_all": 1078, "formula_len_all_1": 1156, "len_all": 113118, "len_all_1": 61386, "len_abs": 1563, "len_title": 132, "len_sents": 31393, "len_sents_1": 27245, "n_sents": 218, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1600, "title": "Unifying Structural Proximity and Equivalence for Enhanced Dynamic Network Embedding", "abs": "Dynamic network embedding methods transform nodes in a dynamic network into low-dimensional vectors while preserving network characteristics, facilitating tasks such as node classification and community detection. Several embedding methods have been proposed to capture $\\textit{structural proximity}$ among nodes in a network, where densely connected communities are preserved, while others have been proposed to preserve $\\textit{structural equivalence}$ among nodes, capturing their structural roles regardless of their relative distance in the network. However, most existing methods that aim to preserve $\\textit{both}$ network characteristics mainly focus on static networks and those designed for dynamic networks do not explicitly account for inter-snapshot structural properties. This paper proposes a novel unifying dynamic network embedding method that simultaneously preserves both structural proximity and equivalence while considering inter-snapshot structural relationships in a dynamic network. Specifically, to define structural equivalence in a dynamic network, we use temporal subgraphs, known as dynamic graphlets, to capture how a node's neighborhood structure evolves over time. We then introduce a temporal-structural random walk to flexibly sample time-respecting sequences of nodes, considering both their temporal proximity and similarity in evolving structures. The proposed method is evaluated using five real-world networks on node classification where it outperforms benchmark methods, showing its effectiveness and flexibility in capturing various aspects of a network.", "keywords": [ "dynamic network", "network embedding", "network representation", "temporal random walk", "dynamic graphlets" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 2, 3, 4, 5 ] }, { "paper_id": "myolhJPuRI", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Junwei Zhou", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Lu Qi", "gender": "Male", "institution": "University of California, Merced", "country": "US", "position": "Postdoc" }, { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Xueting Li", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 22, "n_formula_1": 19, "n_ref_uni": 39, "n_ref": 96, "n_ref_all": 123, "n_fig": 17, "n_tab": 5, "L_tab": 918, "n_element_tab": 100, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 756, "n_element_tab_1": 59, "formula_len_all": 781, "formula_len_all_1": 686, "len_all": 159722, "len_all_1": 56716, "len_abs": 491, "len_title": 120, "len_sents": 44699, "len_sents_1": 26069, "n_sents": 326, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1090, "title": "Layout-your-3D: Controllable and Precise 3D Generation with 2D Blueprint", "abs": "We present Layout-Your-3D, a framework that allows controllable and compositional 3D generation from text prompts. Existing text-to-3D methods often struggle to generate assets with plausible object interactions or require tedious optimization processes. To address these challenges, our approach leverages 2D layouts as a blueprint to facilitate precise and plausible control over 3D generation. Starting with a 2D layout provided by a user or generated from a text description, we first create a coarse 3D scene using a carefully designed initialization process based on efficient reconstruction models. To enforce coherent global 3D layouts and enhance the quality of instance appearances, we propose a collision-aware layout optimization process followed by instance-wise refinement. Experimental results demonstrate that Layout-Your-3D yields more reasonable and visually appealing compositional 3D assets while significantly reducing the time required for each prompt. Additionally, Layout-Your-3D can be easily applicable to downstream tasks, such as 3D editing and object insertion.", "keywords": [ "3D generation", "gaussian splatting", "Text-to-3D", "compositional asset generation" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "myZNJSpiK1", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Buyu Liu", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Fangge Mao", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Intern" }, { "name": "Gai Zhenbiao", "gender": "Male", "institution": "Hangzhou Dianzi University", "country": "", "position": "Undergrad student" }, { "name": "Jiacong Zhou", "gender": "unknown", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "PhD student" }, { "name": "Jun Bao", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Jun Yu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xianyun Wang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Researcher" }, { "name": "Xiaoxing You", "gender": "Male", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "PhD student" }, { "name": "Yan Yang", "gender": "Female", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "Lecturer" }, { "name": "Yiqian Zhang", "gender": "Male", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 57, "n_ref": 96, "n_ref_all": 104, "n_fig": 4, "n_tab": 4, "L_tab": 795, "n_element_tab": 170, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 749, "n_element_tab_1": 116, "formula_len_all": 516, "formula_len_all_1": 253, "len_all": 192064, "len_all_1": 65117, "len_abs": 1742, "len_title": 130, "len_sents": 34139, "len_sents_1": 29999, "n_sents": 248, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1747, "title": "CoVT-CXR: Building Chain of Visual Thought for Interpretable Chest X-Ray Diagnosis", "abs": "Though clinical report generation demonstrates the potential to improve the efficiency of radiologist workflow and benefits the under-served regions, automated analysis of radiographs suffers from un-interpretable progress and inaccurate results. To this end, we propose a novel Chain-of-Visual-Thought (CoVT) to emulate doctors' multi-modal reasoning, enabling more interpretable and accurate CXR diagnostic predictions with explicit multi-step intermediate guidance. Specifically, we mimic the multi-modal multi-step reasoning procedure of the doctors by breaking down clinical reports into individual descriptions and connecting each rationale to corresponding visual prompts\u2014like masks, landmarks, linestrips, and bounding boxes\u2014to illuminate the visual reasoning behind radiographs. By further dividing this association into cross-modal sub-tasks, CoVT is able to exploit a multi-stage fine-tuning protocol to gradually develop the chain-of-reasoning capability. To support this approach, we introduce CoVT-CXR, the first detailed-aligned, multi-step cross-modal dataset for diagnostic tasks, featuring about 3M instruction-following data points for pretraining and around 30K reasoning sequences for fine-tuning, sourced from 6K patient cases and annotated by 32 medical trainees using our tailored tool. Our CoVT-CXR covers more than 20 diseases, requiring 1 to 12 reasoning steps for diagnoses. Through a series of experiments on our CoVT-CXR, we demonstrate the advantages of the CoVT method over baseline approaches, validate the quality of our annotated data, and highlight the positive impacts of CoVT-CXR on various clinical-related tasks. Our CoVT model, annotation tool, and CoVT-CXR dataset will be fully available upon acceptance.", "keywords": [ "chain of visual thought", "multimodal understanding", "fine-grained dataset", "medical report generation", "interpretable LLM." ], "rating_list": [ 8, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "myYzr50xBh", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 12, "author_info_list": [ { "name": "Beidi Chen", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Jikai Long", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "", "position": "MS student" }, { "name": "Osbert Bastani", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Assistant Professor" }, { "name": "Wentao Guo", "gender": "Male", "institution": "Department of Computer Science, Princeton University", "country": "US", "position": "PhD student" }, { "name": "Xiaodong Yu", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Xinyu Yang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Yide Ran", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yimeng Zeng", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Zhaozhuo Xu", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Zirui Liu", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "Assistant Professor" }, { "name": "Christopher De Sa", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Associate Professor" }, { "name": "Jacob R Gardner", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Assistant Professor" } ], "n_formula": 26, "n_formula_1": 13, "n_ref_uni": 58, "n_ref": 104, "n_ref_all": 206, "n_fig": 26, "n_tab": 6, "L_tab": 5416, "n_element_tab": 631, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 3103, "n_element_tab_1": 233, "formula_len_all": 4017, "formula_len_all_1": 880, "len_all": 250120, "len_all_1": 68705, "len_abs": 910, "len_title": 115, "len_sents": 63447, "len_sents_1": 28102, "n_sents": 509, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 913, "title": "Zeroth-Order Fine-Tuning of LLMs with Transferable Static Sparsity", "abs": "Zeroth-order optimization (ZO) is a memory-efficient strategy for fine-tuning Large Language Models using only forward passes. However, applying ZO fine-tuning in memory-constrained settings such as mobile phones and laptops remains challenging since these settings often involve weight quantization, while ZO requires full-precision perturbation and update. In this study, we address this limitation by combining static sparse ZO fine-tuning with quantization. Our approach transfers a small, static subset (0.1%) of \"sensitive\" parameters from pre-training to downstream tasks, focusing fine-tuning on this sparse set of parameters. The remaining untuned parameters are quantized, reducing memory demands. Our proposed workflow enables efficient ZO fine-tuning of an Llama2-7B model on a GPU device with less than 8GiB of memory while outperforming full model ZO fine-tuning performance and in-context learning.", "keywords": [ "Zeroth-order optimization", "LLM fine-tuning" ], "rating_list": [ 6, 6, 6, 6, 5 ], "soundness_list": [ 2, 2, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "myYKk4Qz3l", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Bailin Deng", "gender": "Male", "institution": "Cardiff University", "country": "GB", "position": "Senior Lecturer" }, { "name": "Jinqi Wang", "gender": "Male", "institution": "Cardiff University", "country": "GB", "position": "PhD student" }, { "name": "Yipeng Qin", "gender": "unknown", "institution": "Cardiff University", "country": "GB", "position": "Associate Professor" }, { "name": "Yunfei Fu", "gender": "Male", "institution": "Bournemouth University", "country": "GB", "position": "PhD student" }, { "name": "ZHANGCAN DING", "gender": "Male", "institution": "Iart Technology Co., Limited", "country": "", "position": "Researcher" }, { "name": "Yu-Kun Lai", "gender": "Male", "institution": "Cardiff University", "country": "GB", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 36, "n_ref": 64, "n_ref_all": 92, "n_fig": 14, "n_tab": 2, "L_tab": 169, "n_element_tab": 15, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 293, "formula_len_all_1": 273, "len_all": 103082, "len_all_1": 45279, "len_abs": 1706, "len_title": 96, "len_sents": 29213, "len_sents_1": 22389, "n_sents": 204, "n_sents_1": 149, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1691, "title": "Training-free Editioning of Text-to-Image Models", "abs": "Inspired by the software industry's practice of offering different editions or versions of a product tailored to specific user groups or use cases, we propose a novel task, namely, training-free editioning, for text-to-image models. Specifically, we aim to create variations of a base text-to-image model without retraining, enabling the model to cater to the diverse needs of different user groups or to offer distinct features and functionalities. To achieve this, we propose that different editions of a given text-to-image model can be formulated as concept subspaces in the latent space of its text encoder (e.g., CLIP). In such a concept subspace, all points satisfy a specific user need (e.g., generating images of a cat lying on the grass/ground/falling leaves). Technically, we apply Principal Component Analysis (PCA) to obtain the desired concept subspaces from representative text embedding that correspond to a specific user need or requirement. Projecting the text embedding of a given prompt into these low-dimensional subspaces enables efficient model editioning without retraining. Intuitively, our proposed editioning paradigm enables a service provider to customize the base model into its \"cat edition\" that restricts image generation to cats, regardless of the user's prompt (e.g., dogs, people, etc.). This introduces a new dimension for product differentiation, targeted functionality, and pricing strategies, unlocking novel business models for text-to-image generators. Extensive experimental results demonstrate the validity of our approach and its potential to enable a wide range of customized text-to-image model editions across various domains and applications.", "keywords": [ "text-to-image synthesis", "software edition", "concept subspace" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "mxkm1Pr2PM", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Guorong Wu", "gender": "Male", "institution": "University of North Carolina, Chapel Hill", "country": "US", "position": "Associate Professor" }, { "name": "Tingting Dan", "gender": "Female", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Postdoc" }, { "name": "Won Kim", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Chiyuen Chow", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Researcher" } ], "n_formula": 48, "n_formula_1": 22, "n_ref_uni": 57, "n_ref": 96, "n_ref_all": 115, "n_fig": 4, "n_tab": 10, "L_tab": 4514, "n_element_tab": 322, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1387, "n_element_tab_1": 70, "formula_len_all": 4622, "formula_len_all_1": 2173, "len_all": 192128, "len_all_1": 75826, "len_abs": 2310, "len_title": 89, "len_sents": 54886, "len_sents_1": 36084, "n_sents": 391, "n_sents_1": 231, "venue": "ICLR2025", "decision": "withdraw", "L_title": 41, "L_abs": 2360, "title": "Graph Neural Network Is A Mean Field Game", "abs": "In current graph neural networks (GNNs), it is a common practice to apply a pre-defined message passing heuristics to all graph data, even though the stereotypical relational inductive bias (e.g., graph heat diffusion) might not fit the unseen graph topology. Such gross simplification might be responsible for the lack of an in-depth understanding of graph learning principles, which challenges us to push the boundary from crafting application-specific GNNs to embracing a \"meta-learning\" paradigm. In this work, we ratchet the gear of GNN another notch forward by formulating GNN as a *mean field game*, that is, the best learning outcome occurs at the *Nash*-equilibrium when the learned graph inference rationale allows each graph node to find what is the best feature representations for not only the individual node but also the entire graph. Following this spirit, we formulate the search for novel GNN mechanism into a variational framework of *mean-field control* (MFC) problem, where the optimal relational inductive bias is essentially the critical point of mean-field information dynamics. Specifically, we seek for the best characteristic MFC functions of transportation mobility (controlling information exchange throughout the graph) and reaction mobility (controlling feature representation learning on each node), on the fly, that uncover the most suitable learning mechanism for a GNN instance by solving an MFC variational problem through the lens of *Hamiltonian flows* (formed in partial differential equations). In this context, our variational framework brings together existing GNN models into various mean-field games with distinct equilibrium states, each characterized by a unique MFC functional. Furthermore, we present an agnostic end-to-end deep model, coined *Nash-GNN* (in honor of Nobel laureate Dr. John Nash), to jointly carve the nature of the inductive bias and fine-tune the GNN hyper-parameters on top of the elucidated learning mechanism. *Nash-GNN* has achieved SOTA performance on diverse graph data including popular benchmark datasets and human connectomes. More importantly, the mathematical insight of mean-field games provides a new window to understand the foundational principles of graph learning as an interactive dynamical system, which allows us to reshape the idea of designing next-generation GNN models.", "keywords": [ "graph neural network", "mean field game", "reaction-diffusion equations", "Hamiltonian flows" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 4, 2, 2 ], "confidence_list": [ 2, 2, 3 ] }, { "paper_id": "mwYkVSddzx", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Dennis Frauen", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Jonas Schweisthal", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Stefan Feuerriegel", "gender": "Male", "institution": "LMU Munich", "country": "DE", "position": "Professor" }, { "name": "Valentyn Melnychuk", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" } ], "n_formula": 65, "n_formula_1": 23, "n_ref_uni": 61, "n_ref": 188, "n_ref_all": 219, "n_fig": 7, "n_tab": 6, "L_tab": 2551, "n_element_tab": 222, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 948, "n_element_tab_1": 11, "formula_len_all": 5189, "formula_len_all_1": 1670, "len_all": 232677, "len_all_1": 81355, "len_abs": 1308, "len_title": 116, "len_sents": 72442, "len_sents_1": 36338, "n_sents": 549, "n_sents_1": 270, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1312, "title": "Orthogonal Representation Learning for Estimating Causal Quantities", "abs": "Representation learning is widely used for estimating causal quantities (e.g., the conditional average treatment effect) from observational data. While existing representation learning methods have the benefit of allowing for end-to-end learning, they do not have favorable theoretical properties of Neyman-orthogonal learners, such as double robustness and quasi-oracle efficiency. Also, such representation learning methods often employ additional constraints, like balancing, which may even lead to inconsistent estimation. In this paper, we propose a novel class of Neyman-orthogonal learners for causal quantities defined at the representation level, which we call OR-learners. Our OR-learners have several practical advantages: they allow for consistent estimation of causal quantities based on any learned representation, while offering favorable theoretical properties including double robustness and quasi-oracle efficiency. In numerous experiments, we show that, under certain regularity conditions, our OR-learners improve existing representation learning methods and achieve state-of-the-art performance. To the best of our knowledge, our OR-learners are the first work to provide a unified framework of representation learning methods and Neyman-orthogonal learners for causal quantities estimation.", "keywords": [ "treatment effect estimation", "counterfactual outcomes estimation", "representation learning" ], "rating_list": [ 8, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 1 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "mvAL02hEJg", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "David Cohen", "gender": "Male", "institution": "Electrical Engineering Department, Technion \u2013 Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "PhD student" }, { "name": "Ronen Talmon", "gender": "unknown", "institution": "Technion - Israel Institute of Technology", "country": "IL", "position": "Full Professor" }, { "name": "Uri Shaham", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 36, "n_formula_1": 28, "n_ref_uni": 49, "n_ref": 100, "n_ref_all": 115, "n_fig": 4, "n_tab": 7, "L_tab": 2220, "n_element_tab": 78, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1660, "formula_len_all_1": 1393, "len_all": 150170, "len_all_1": 68034, "len_abs": 1721, "len_title": 107, "len_sents": 47041, "len_sents_1": 32830, "n_sents": 358, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1728, "title": "Revisiting Covariate and Hypothesis Roles in ITE Estimation: A New Approach Using Laplacian Regularization", "abs": "The recent surge in data availability across many fields, such as medicine, social science, and marketing, has brought to the forefront the problem of estimating Individual Treatment Effect (ITE) from observational data to effectively tailor treatment to personalized characteristics. ITE estimation is known to be a challenging task because we can only observe the outcome with or without treatment, but never both. Moreover, observational datasets exhibit selection bias induced by the treatment assignment policy. In this paper, we present a new approach consisting of two novel aspects. First, we depart from conventional approaches that minimize the covariate shift. Instead, we incorporate it as a crucial element in ITE estimation, recognizing that it stems from highly predictive features that exhibit significant imbalance in observational data. Second, unlike existing methods, our approach utilizes hypothesis functions to directly estimate outcomes under covariate shift, enhancing reliability across observed and unobserved outcomes. To support this approach theoretically, we derive a new upper bound of the expected ITE loss and show that it explicitly depends on the discrepancy between the hypothesis functions, which are absent from the objectives of existing methods. Based on this new approach, we present LITE: Laplacian Individual Treatment Effect, a novel method that leverages Laplacian-regularized representation and incorporates both the covariate shift and the hypothesis functions for ITE estimation, effectively bridging observed and unobserved outcomes. We demonstrate LITE on illustrative simulations and two leading benchmarks, where we show superior results compared to state-of-the-art methods.", "keywords": [ "Individual Treatment Effect (ITE)", "Conditional Average Treatment Effect (CATE)", "Covariate Shift", "Laplacian Regularization", "Causal Inference" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 1, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "mun3bGqdDM", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Chaohao Yuan", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Geyan Ye", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Senior Researcher" }, { "name": "Jianhua Yao", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Principal Researcher" }, { "name": "Long-Kai Huang", "gender": "unknown", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Wei Liu", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Yikun Zhang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yu Rong", "gender": "Male", "institution": "Alibaba Group ", "country": "CN", "position": "Senior Staff Algorithm Engineer" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 21, "n_formula_1": 16, "n_ref_uni": 35, "n_ref": 85, "n_ref_all": 116, "n_fig": 9, "n_tab": 24, "L_tab": 7249, "n_element_tab": 913, "n_fig_1": 5, "n_tab_1": 9, "L_tab_1": 3248, "n_element_tab_1": 430, "formula_len_all": 2705, "formula_len_all_1": 978, "len_all": 197645, "len_all_1": 72080, "len_abs": 1860, "len_title": 155, "len_sents": 53745, "len_sents_1": 27213, "n_sents": 434, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 106, "L_abs": 1516, "title": "Atomas: Hierarchical Adaptive Alignment on Molecule-Text for Unified Molecule Understanding and Generation", "abs": "Molecule-and-text cross-modal representation learning has emerged as a promising direction for enhancing the quality of molecular representation, thereby improving performance in various scientific fields. However, most approaches employ a global alignment approach to learn the knowledge from different modalities that may fail to capture fine-grained information, such as molecule-and-text fragments and stereoisomeric nuances, which is crucial for downstream tasks. Furthermore, it is incapable of modeling such information using a similar global alignment strategy due to the lack of annotations about the fine-grained fragments in the existing dataset.\nIn this paper, we propose Atomas, a hierarchical molecular representation learning framework that jointly learns representations from SMILES strings and text. We design a Hierarchical Adaptive Alignment model to automatically learn the fine-grained fragment correspondence between two modalities and align these representations at three semantic levels. \nAtomas's end-to-end training framework supports understanding and generating molecules, enabling a wider range of downstream tasks. Atomas achieves superior performance across 12 tasks on 10 datasets, outperforming 10 baseline models thus highlighting the effectiveness and versatility of our method. Scaling experiments further demonstrate Atomas\u2019s robustness and scalability. Moreover, visualization and qualitative analysis, validated by human experts, confirm the chemical relevance of our approach.", "keywords": [ "Cross-modal representation learning", "Molecule-text understanding", "Alignment" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "muN3B40keb", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Christopher Buckley", "gender": "Male", "institution": "University of Sussex", "country": "", "position": "Full Professor" }, { "name": "Giuseppe Castiglione", "gender": "Male", "institution": "Epson", "country": "", "position": "Researcher" }, { "name": "Ivor Simpson", "gender": "Male", "institution": "University of Sussex", "country": "", "position": "Lecturer" } ], "n_formula": 29, "n_formula_1": 25, "n_ref_uni": 16, "n_ref": 69, "n_ref_all": 87, "n_fig": 7, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1938, "formula_len_all_1": 1745, "len_all": 110396, "len_all_1": 57648, "len_abs": 861, "len_title": 129, "len_sents": 30488, "len_sents_1": 25587, "n_sents": 235, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 869, "title": "Common Causes for Sudden Shifts: Linking Phase Transitions in Sinusoidal Networks", "abs": "Different phases of learning dynamics exist when training deep neural networks. These can be characterised by statistics called order parameters. In this work we identify a shared, underlying mechanism connecting three seemingly distinct phase transitions in the training of a class of deep regression models, specificially Implicit Neural Representations (INRs) of image data. These transitions include: the emergence of wave patterns in residuals (a novel observation), the transition from fast to slow learning, and Neural Tangent Kernel (NTK) alignment. We relate the order parameters for each phenomenon to a common set of variables derived from a local approximation of the structure of the NTK. Furthermore, we present experimental evidence demonstrating these transitions coincide. Our results enable new insights on the inductive biases of sinusoidal INRs.", "keywords": [ "neural tangent kernel", "implicit neural networks", "phase transitions", "learning dynamics" ], "rating_list": [ 6, 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 4, 2, 4, 3, 3 ], "confidence_list": [ 4, 3, 4, 3, 1 ] }, { "paper_id": "mtyYWBx2ZF", "primary_area": "interpretability and explainable AI", "n_author": 10, "author_info_list": [ { "name": "Haiyang Sun", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "MS student" }, { "name": "Junwei Han", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "", "position": "Full Professor" }, { "name": "Lin Zhao", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Tianming Liu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Professor" }, { "name": "Xiaohui Gao", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "PhD student" }, { "name": "Xintao Hu", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "Full Professor" }, { "name": "Yutao Hu", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "MS student" }, { "name": "Zihao Wu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Zuo Mengfei", "gender": "unknown", "institution": "Northwestern Polytechnical University, Northwest Polytechnical University Xi'an", "country": "CN", "position": "Undergrad student" }, { "name": "Wade Zhang", "gender": "Male", "institution": "Augusta University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 50, "n_ref": 72, "n_ref_all": 102, "n_fig": 18, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 16, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 89, "formula_len_all_1": 89, "len_all": 126236, "len_all_1": 58684, "len_abs": 1930, "len_title": 111, "len_sents": 30891, "len_sents_1": 29375, "n_sents": 182, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1938, "title": "Brain-like Functional Organization within Large Language Models", "abs": "The human brain has long inspired the pursuit of artificial intelligence (AI). Recently, neuroimaging studies provide compelling evidence of alignment between the computational representation of artificial neural networks (ANNs) and the neural responses of the human brain to external stimuli, suggesting that ANNs may employ brain-like information processing strategies. While such alignment has been observed across sensory modalities\u2014visual, auditory, and linguistic\u2014much of the focus has been on the behaviors of artificial neurons (ANs) at the population level, leaving the functional organization of individual ANs that facilitates such brain-like processes largely unexplored. In this study, we bridge this gap by directly coupling sub-groups of artificial neurons with functional brain networks (FBNs), the foundational organizational structure of the human brain. Specifically, we extract representative patterns from temporal responses of ANs in large language models (LLMs), and use them as fixed regressors to construct voxel-wise encoding models to predict brain activity recorded by functional magnetic resonance imaging (fMRI). This framework effectively links the AN sub-groups to FBNs, enabling the delineation of brain-like functional organization within LLMs. Our findings reveal that LLMs (BERT and Llama 1\u20133) exhibit brain-like functional architecture, with sub-groups of artificial neurons mirroring the organizational patterns of well-established FBNs. Notably, the brain-like functional organization of LLMs evolves with the increased sophistication and capability, achieving an improved balance between the diversity of computational behaviors and the consistency of functional specializations. This research represents the first exploration of brain-like functional organization within LLMs, offering novel insights to inform the development of artificial general intelligence (AGI) with human brain principles.", "keywords": [ "large language models", "network organization", "functional brain networks" ], "rating_list": [ 3, 5, 6, 1 ], "soundness_list": [ 3, 2, 3, 1 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "mtmqwhQiaG", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Haibin Shen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Junchuan Gu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Kejie Huang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yiwen Gu", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 49, "n_formula_1": 17, "n_ref_uni": 25, "n_ref": 46, "n_ref_all": 84, "n_fig": 4, "n_tab": 4, "L_tab": 2538, "n_element_tab": 287, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 846, "n_element_tab_1": 56, "formula_len_all": 4810, "formula_len_all_1": 1297, "len_all": 137368, "len_all_1": 65684, "len_abs": 1486, "len_title": 113, "len_sents": 40801, "len_sents_1": 27836, "n_sents": 369, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1493, "title": "Canonic Signed Spike Coding for Efficient Spiking Neural Networks", "abs": "Spiking Neural Networks (SNNs) seek to mimic the spiking behavior of biological neurons and are expected to play a key role in the advancement of neural computing and artificial intelligence. The conversion of Artificial Neural Networks (ANNs) to SNNs is the most widely used training method, which ensures that the resulting SNNs perform comparably to ANNs on large-scale datasets. The efficiency of these conversion-based SNNs is often determined by the neural coding schemes. Current schemes typically use spike count or timing for encoding, which is linearly related to ANN activations and increases the required number of time steps. To address this limitation, we propose a novel Canonic Signed Spike (CSS) coding scheme. This method incorporates non-linearity into the encoding process by weighting spikes at each step of neural computation, thereby increasing the information encoded in spikes. We identify the temporal coupling phenomenon arising from weighted spikes and introduce negative spikes along with a Ternary Self-Amplifying (TSA) neuron model to mitigate the issue. A one-step silent period is implemented during neural computation, achieving high accuracy with low latency. We apply the proposed methods to directly convert full-precision ANNs and evaluate performance on CIFAR-10 and ImageNet datasets. Our experimental results demonstrate that the CSS coding scheme effectively compresses time steps for coding and reduces inference latency with minimal conversion loss.", "keywords": [ "Spiking neural network", "Spike coding scheme", "Stepwise weighted spike" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "mtSSFiqW6y", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Ali Thabet", "gender": "Male", "institution": "Meta", "country": "CH", "position": "Applied Research Manager" }, { "name": "Artsiom Sanakoyeu", "gender": "Unspecified", "institution": "Facebook", "country": "CH", "position": "Researcher" }, { "name": "Edgar Sch\u00f6nfeld", "gender": "unknown", "institution": "Meta", "country": "CH", "position": "Research Scientist" }, { "name": "Gregor Bachmann", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" }, { "name": "Jonas K Kohler", "gender": "Male", "institution": "Facebook", "country": "CH", "position": "Researcher" }, { "name": "Markos Georgopoulos", "gender": "unknown", "institution": "Meta", "country": "CH", "position": "Researcher" }, { "name": "Sotiris Anagnostidis", "gender": "Male", "institution": "ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Yuming Du", "gender": "Male", "institution": "Facebook", "country": "CH", "position": "Researcher" }, { "name": "Albert Pumarola", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 31, "n_ref": 61, "n_ref_all": 91, "n_fig": 11, "n_tab": 4, "L_tab": 1602, "n_element_tab": 68, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 698, "n_element_tab_1": 72, "formula_len_all": 505, "formula_len_all_1": 211, "len_all": 195675, "len_all_1": 64519, "len_abs": 2269, "len_title": 129, "len_sents": 40743, "len_sents_1": 29802, "n_sents": 336, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 81, "L_abs": 1785, "title": "Judge Decoding: Faster Speculative Sampling Requires Going Beyond Model Alignment", "abs": "The performance of large language models (LLMs) is closely linked to their underlying size, leading to ever-growing networks and hence slower inference. Speculative decoding has been proposed as a technique to accelerate autoregressive generation, leveraging a fast draft model to propose candidate tokens, which are then verified in parallel based on their likelihood under the target model. While this approach guarantees to reproduce the target output, it incurs a substantial penalty: many high-quality draft tokens are rejected, even when they represent objectively valid continuations. Indeed, we show that even powerful draft models such as GPT-4o, as well as human text cannot achieve high acceptance rates under the standard verification scheme. This severely limits the speedup potential of current speculative decoding methods, as an early rejection becomes overwhelmingly likely when solely relying on alignment of draft and target.\nWe thus ask the following question: Can we adapt verification to recognize correct, but non-aligned replies? To this end, we draw inspiration from the LLM-as-a-judge framework, which demonstrated that LLMs are able to rate answers in a versatile way. We carefully design a dataset coined TokenCourt to elicit the same capability in the target model by training a compact module on top of the embeddings to produce ``judgements\" of the current continuation. We showcase our strategy on the Llama-3.1 family, where our 8B/405B-Judge achieves a speedup of $9\\times$ over Llama-405B, while maintaining its quality on a large range of benchmarks. These benefits remain present even in optimized inference frameworks, where our method reaches up to $141$ tokens/s for 8B/70B-Judge and $129$ tokens/s for 8B/405B on $2$ and $8$ H100s respectively.", "keywords": [ "LLM inference", "speculative decoding" ], "rating_list": [ 6, 10, 8, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 4, 4, 4 ], "contribution_list": [ 2, 4, 3, 4 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "mtJSMcF3ek", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Carson Eisenach", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Dean Foster", "gender": "Male", "institution": "Amazon", "country": "", "position": "scientist" }, { "name": "Hanlin Zhang", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Sham M. Kakade", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Udaya Ghai", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Yuda Song", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 76, "n_ref": 128, "n_ref_all": 164, "n_fig": 16, "n_tab": 7, "L_tab": 2321, "n_element_tab": 431, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 224, "n_element_tab_1": 24, "formula_len_all": 1185, "formula_len_all_1": 909, "len_all": 238429, "len_all_1": 68345, "len_abs": 2245, "len_title": 130, "len_sents": 56118, "len_sents_1": 32448, "n_sents": 470, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 82, "L_abs": 1129, "title": "Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models", "abs": "Self-improvement is a mechanism in Large Language Model (LLM) pre-training, post-training and test-time inference. We explore a framework where the model verifies its own outputs, filters or reweights data based on this verification, and distills the filtered data. Despite several empirical successes, a fundamental understanding is still lacking. In this work, we initiate a comprehensive, modular and controlled study on LLM self-improvement. We provide a mathematical formulation for self-improvement, which is largely governed by a quantity which we formalize as the *generation-verification gap*. Through experiments with various model families and tasks, we discover a scaling phenomenon of self-improvement -- a variant of the generation-verification gap scales monotonically with the model pre-training flops. We also examine when self-improvement is possible, an iterative self-improvement procedure, and ways to improve its performance. We believe our results have several empirical implications, and our study leaves many exciting future directions for understanding the potential and limits of LLM self-improvement.", "keywords": [ "LLM", "self-improvement", "synthetic data", "post-training", "test-time optimization" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "mscnV6JZkT", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Arya Mazumdar", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Harsh Vardhan", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Heng Zhu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 94, "n_formula_1": 31, "n_ref_uni": 43, "n_ref": 72, "n_ref_all": 83, "n_fig": 2, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7089, "formula_len_all_1": 1755, "len_all": 171963, "len_all_1": 69476, "len_abs": 1782, "len_title": 78, "len_sents": 47047, "len_sents_1": 32009, "n_sents": 416, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1792, "title": "Distributed Gradient Descent with Many Local Steps in Overparameterized Models", "abs": "In distributed training of machine learning models, gradient descent with local iterative steps is a very popular method, variants of which are commonly known as Local-SGD or the Federated Averaging (FedAvg). In this method, gradient steps based on local datasets are taken independently in distributed compute nodes to update the local models, which are then aggregated intermittently. Although the existing convergence analysis suggests that with heterogeneous data, FedAvg encounters quick performance degradation as the number of local steps increases, it is shown to work quite well in practice, especially in the distributed training of large language models. In this work we try to explain this good performance from a viewpoint of implicit bias in Local Gradient Descent (Local-GD) with a large number of local steps. In overparameterized regime, the gradient descent at each compute node would lead the model to a specific direction locally. We characterize the dynamics of the aggregated global model and compare it to the centralized model trained with all of the data in one place. In particular, we analyze the implicit bias of gradient descent on linear models, for both regression and classification tasks. Our analysis shows that the aggregated global model converges exactly to the centralized model for regression tasks, and converges (in direction) to the same feasible set as centralized model for classification tasks. We further propose a Modified Local-GD with a refined aggregation and theoretically show it converges to the centralized model in direction for linear classification. We empirically verified our theoretical findings in linear models and also conducted experiments on distributed fine-tuning of pretrained neural networks to further apply our theory.", "keywords": [ "Distributed Learning", "Overparameterization", "Optimization", "Federated Learning" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 2, 4, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "msEr27EejF", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Anca Dragan", "gender": "Female", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Cassidy Laidlaw", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Shivam Singhal", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" } ], "n_formula": 88, "n_formula_1": 10, "n_ref_uni": 79, "n_ref": 101, "n_ref_all": 134, "n_fig": 6, "n_tab": 4, "L_tab": 1492, "n_element_tab": 155, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 764, "n_element_tab_1": 59, "formula_len_all": 10269, "formula_len_all_1": 621, "len_all": 259935, "len_all_1": 73258, "len_abs": 1329, "len_title": 127, "len_sents": 68831, "len_sents_1": 35198, "n_sents": 587, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 79, "L_abs": 1339, "title": "Correlated Proxies: A New Definition and Improved Mitigation for Reward Hacking", "abs": "Because it is difficult to precisely specify complex objectives, reinforcement learning policies are often optimized using flawed proxy rewards that seem to capture the true objective. However, optimizing proxy rewards frequently leads to reward hacking: the optimized reward function ceases to be a good proxy and the resulting policy performs poorly with respect to the unspecified true reward. Principled solutions to reward hacking have been impeded by the lack of a good definition for the problem. We introduce a definition of reward hacking based on correlation between proxy and true rewards for states and actions seen by a \"base policy\" that breaks down under optimization. We show that this definition captures reward hacking behavior across several realistic settings, including in reinforcement learning from human feedback (RLHF). We then show theoretically that regularization to the base policy can effectively prevent reward hacking. Our theory suggests regularizing $\\chi^2$ divergence between the policies' occupancy measures, rather than the current practice in RLHF of using a KL penalty between action distributions. We intuitively show why this type of regularization is better, and demonstrate that it outperforms alternatives at mitigating reward hacking in practice across four realistic settings, including RLHF.", "keywords": [ "reward hacking", "reward gaming", "overoptimization", "occupancy measures" ], "rating_list": [ 8, 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 4, 3, 4 ], "presentation_list": [ 4, 2, 4, 4, 4 ], "contribution_list": [ 3, 3, 4, 3, 2 ], "confidence_list": [ 3, 3, 4, 4, 3 ] }, { "paper_id": "msD4DHZzFg", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Guodong Li", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Professor" }, { "name": "Lequan Yu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Qinshuo Liu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Wei Huang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yanwen Fang", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Researcher" }, { "name": "Zhao WeiQin", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 25, "n_ref_uni": 45, "n_ref": 114, "n_ref_all": 146, "n_fig": 7, "n_tab": 10, "L_tab": 3210, "n_element_tab": 442, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 5295, "n_element_tab_1": 385, "formula_len_all": 1863, "formula_len_all_1": 1288, "len_all": 150354, "len_all_1": 61017, "len_abs": 1323, "len_title": 140, "len_sents": 37108, "len_sents_1": 21665, "n_sents": 281, "n_sents_1": 138, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1329, "title": "From Layers to States: A State Space Model Perspective to Deep Neural Network Layer Dynamics", "abs": "The depth of neural networks is a critical factor for their capability, with deeper models often demonstrating superior performance. Motivated by this, significant efforts have been made to enhance layer aggregation - reusing information from previous layers to better extract features at the current layer, to improve the representational power of deep neural networks. However, previous works have primarily addressed this problem from a discrete-state perspective which is not suitable as the number of network layers grows. This paper novelly treats the outputs from layers as states of a continuous process and considers leveraging the state space model (SSM) to design the aggregation of layers in very deep neural networks. Moreover, inspired by its advancements in modeling long sequences, the Selective State Space Models (S6) is employed to design a new module called Selective State Space Model Layer Aggregation (S6LA). This module aims to combine traditional CNN or transformer architectures within a sequential framework, enhancing the representational capabilities of state-of-the-art vision networks. Extensive experiments show that S6LA delivers substantial improvements in both image classification and detection tasks, highlighting the potential of integrating SSMs with contemporary deep learning techniques.", "keywords": [ "deep neural network", "sequential model", "state space model", "statistical model" ], "rating_list": [ 6, 8, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "ms9bK61Hxn", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Ganesh Sundaramoorthi", "gender": "Male", "institution": "RTX Technology Research Center", "country": "US", "position": "Principal Researcher" }, { "name": "Xinhua Zhang", "gender": "Male", "institution": "Raytheon Technologies Research Center", "country": "US", "position": "Researcher" } ], "n_formula": 50, "n_formula_1": 23, "n_ref_uni": 24, "n_ref": 40, "n_ref_all": 60, "n_fig": 4, "n_tab": 6, "L_tab": 2961, "n_element_tab": 156, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 741, "n_element_tab_1": 53, "formula_len_all": 4398, "formula_len_all_1": 1620, "len_all": 133629, "len_all_1": 70737, "len_abs": 971, "len_title": 99, "len_sents": 44205, "len_sents_1": 31993, "n_sents": 340, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 878, "title": "Similarity Group Equivariant Convolutional Networks", "abs": "We introduce similarity group equivariant convolutional networks (SECNNs), designed to achieve continuous translation, rotation and scale equivariance, or discrete similarity group equivariance that involves discrete Dihedral group. The networks are implemented as steerable CNNs by employing a steerable and approximately shiftable and scalable basis for continuous translating, rotating and scaling convolution kernels within a five-dimensional position-orientation-scale-reflection space. \nOur results demonstrate that SECNNs attain state-of-the-art results on translated, rotated and scaled MNIST datasets. SECNNs also achieve the accuracy of other leading group equivariant networks on CIFAR10/100,\nwhile being equivariant to the full range of the similarity group in comparison to existing state of the art, which is equivariant to only sub-groups of the similarity group.", "keywords": [ "Group Convolution", "Group Equivariance", "Similarity Transformation", "Continuous Transformation", "Translation", "Rotation", "Scaling", "Reflection" ], "rating_list": [ 5, 5, 8, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 2, 2, 4, 1 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "mrjOaRyefn", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Fan Liu", "gender": "unknown", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Hao Liu", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhao Xu", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 46, "n_formula_1": 19, "n_ref_uni": 15, "n_ref": 61, "n_ref_all": 76, "n_fig": 3, "n_tab": 15, "L_tab": 5505, "n_element_tab": 668, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1220, "n_element_tab_1": 208, "formula_len_all": 3364, "formula_len_all_1": 1180, "len_all": 239659, "len_all_1": 71527, "len_abs": 2327, "len_title": 112, "len_sents": 78927, "len_sents_1": 31558, "n_sents": 633, "n_sents_1": 253, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1558, "title": "Adversarial Tuning: Defending Against Jailbreak Attacks for LLMs", "abs": "Although safely enhanced Large Language Models (LLMs) have achieved remarkable success in tackling various complex tasks in a zero-shot manner, they remain susceptible to jailbreak attacks, particularly the unknown jailbreak attack. To enhance LLMs' generalized defense capabilities, we propose a two-stage adversarial tuning framework, which generates adversarial prompts to explore worst-case scenarios by optimizing datasets containing pairs of adversarial prompts and their safe responses. In the first stage, we introduce the hierarchical meta-universal adversarial prompt learning to efficiently and effectively generate token-level adversarial prompts. In the second stage, we propose automatic adversarial prompt learning to iteratively construct out-of-distribution adversarial prompts, further enhancing LLM\u2019s defense capabilities. We conducted comprehensive experiments on three widely used jailbreak datasets, comparing our framework with six defense baselines under five representative attack scenarios. \\fan{ Specifically, for the computational efficiency of generating token-level adversarial prompts, we demonstrate both empirically and theoretically that our method achieves approximately a 15x speedup. Additionally, our methods exhibit superior defense performance against both known and unknown jailbreak attacks. Importantly, our adversarial tuning framework shows broad generalizability across various attack strategies and target LLMs (including the large 110B model), highlighting its potential as a transferable defense mechanism.", "keywords": [ "jailbreak defense", "jialbreak attack", "LLM" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "mrNVOWlG25", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Christopher Hannemann", "gender": "Male", "institution": "Dexcom", "country": "", "position": "Product Owner" }, { "name": "Joost Van der Linden", "gender": "unknown", "institution": "Dexcom, Inc.", "country": "", "position": "Data Scientist" }, { "name": "Josephine Lamp", "gender": "Female", "institution": "Dexcom", "country": "US", "position": "Principal Researcher" }, { "name": "Mark Derdzinski", "gender": "Male", "institution": "Dexcom, Inc.", "country": "", "position": "Senior Manager" }, { "name": "Samuel Hatfield", "gender": "Male", "institution": "Dexcom Inc", "country": "US", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 36, "n_ref": 47, "n_ref_all": 65, "n_fig": 8, "n_tab": 1, "L_tab": 372, "n_element_tab": 24, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1144, "n_element_tab_1": 16, "formula_len_all": 453, "formula_len_all_1": 551, "len_all": 100904, "len_all_1": 58607, "len_abs": 1573, "len_title": 57, "len_sents": 29837, "len_sents_1": 27474, "n_sents": 223, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1580, "title": "MotifDisco: Motif Causal Discovery For Time Series Motifs", "abs": "Many time series, particularly health data streams, can be best understood as a sequence of phenomenon or events, which we call motifs. A time series motif is a short trace segment which may implicitly capture an underlying phenomenon within the time series. Specifically, we focus on glucose traces collected from continuous glucose monitors (CGMs), which inherently contain motifs representing underlying human behaviors such as eating and exercise. The ability to identify and quantify causal relationships amongst motifs can provide a mechanism to better understand and represent these patterns, useful for improving deep learning and generative models and for advanced technology development (e.g., personalized coaching and artificial insulin delivery systems). However, no previous work has developed causal discovery methods for time series motifs. Therefore, in this paper we develop MotifDisco (motif disco-very of causality), a novel causal discovery framework to learn causal relations amongst motifs from time series traces. We formalize a notion of Motif Causality (MC), inspired from Granger Causality and Transfer Entropy, and develop a Graph Neural Network-based framework that learns causality between motifs by solving an unsupervised link prediction problem. We also integrate MC with three model use cases of forecasting, anomaly detection and clustering, to showcase the use of MC as a building block for other downstream tasks. Finally, we evaluate our framework and find that Motif Causality provides a significant performance improvement in all use cases.", "keywords": [ "motif", "causality", "causal discovery", "time series", "graph neural network", "glucose" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "mr2icR6dpD", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Leigang Qu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Liqiang Nie", "gender": "Male", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "dean" }, { "name": "Tan Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Wenjie Wang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Yongqi Li", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "Li Haochuan", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Tat-seng Chua", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 68, "n_ref": 150, "n_ref_all": 186, "n_fig": 13, "n_tab": 13, "L_tab": 5439, "n_element_tab": 783, "n_fig_1": 7, "n_tab_1": 7, "L_tab_1": 2544, "n_element_tab_1": 322, "formula_len_all": 364, "formula_len_all_1": 322, "len_all": 228169, "len_all_1": 71341, "len_abs": 6655, "len_title": 131, "len_sents": 55741, "len_sents_1": 30261, "n_sents": 423, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1354, "title": "TIGeR: Unifying Text-to-Image Generation and Retrieval with Large Multimodal Models", "abs": "How humans can effectively and efficiently acquire images has always been a perennial question. A classic solution is *text-to-image retrieval* from an existing database; however, the limited database typically lacks creativity. By contrast, recent breakthroughs in *text-to-image generation* have made it possible to produce attractive and counterfactual visual content, but it faces challenges in synthesizing knowledge-intensive images. In this work, we rethink the relationship between text-to-image generation and retrieval, proposing a *unified* framework for both tasks with one single Large Multimodal Model (LMM). Specifically, we first explore the intrinsic discriminative abilities of LMMs and introduce an efficient generative retrieval method for text-to-image retrieval in a training-free manner. Subsequently, we unify generation and retrieval autoregressively and propose an autonomous decision mechanism to choose the best-matched one between generated and retrieved images as the response to the text prompt. To standardize the evaluation of unified text-to-image generation and retrieval, we construct TIGeR-Bench, a benchmark spanning both creative and knowledge-intensive domains. Extensive experiments on TIGeR-Bench and two retrieval benchmarks, *i.e.*, Flickr30K and MS-COCO, demonstrate the superiority of our proposed framework.", "keywords": [ "Multimodal Large Language Models", "Text-to-Image Generation", "Cross-Modal Retrieval" ], "rating_list": [ 8, 6, 8, 6, 5 ], "soundness_list": [ 4, 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 4, 3, 3, 3, 2 ], "confidence_list": [ 4, 2, 4, 3, 3 ] }, { "paper_id": "mqNKiEB6pd", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Feijie Wu", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Haoyu Wang", "gender": "Male", "institution": "State University of New York at Albany", "country": "US", "position": "Assistant Professor" }, { "name": "Lu Su", "gender": "Male", "institution": "Purdue University", "country": "", "position": "Associate Professor" }, { "name": "Xiaoze Liu", "gender": "Not Specified", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Xingchen Wang", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Jing Gao", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 53, "n_ref": 84, "n_ref_all": 100, "n_fig": 4, "n_tab": 6, "L_tab": 2142, "n_element_tab": 177, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 729, "n_element_tab_1": 27, "formula_len_all": 1012, "formula_len_all_1": 576, "len_all": 194036, "len_all_1": 73358, "len_abs": 1267, "len_title": 113, "len_sents": 57544, "len_sents_1": 35593, "n_sents": 474, "n_sents_1": 270, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1271, "title": "Towards Federated RLHF with Aggregated Client Preference for LLMs", "abs": "Reinforcement learning with human feedback (RLHF) fine-tunes a pretrained large language model (LLM) using user preference data, enabling it to generate content aligned with human preferences. However, due to privacy concerns, users may be reluctant to share sensitive preference data. To address this, we propose utilizing Federated Learning (FL) techniques, allowing large-scale preference collection from diverse real-world users without requiring them to transmit data to a central server. Our federated RLHF methods (i.e., FedBis and FedBiscuit) encode each client\u2019s preferences into binary selectors and aggregate them to capture common preferences. In particular, FedBiscuit overcomes key challenges, such as preference heterogeneity and reward hacking, through innovative solutions like grouping clients with similar preferences to reduce heterogeneity and using multiple binary selectors to enhance LLM output quality. To evaluate the performance of the proposed methods, we establish the first federated RLHF benchmark with a heterogeneous human preference dataset. Experimental results show that by integrating the LLM with aggregated client preferences, FedBis and FedBiscuit significantly enhance the professionalism and readability of the generated content.", "keywords": [ "Federated learning", "RLHF", "LLM" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "mqKVe6F3Up", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Changyeon Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Doohyun Lee", "gender": "Male", "institution": "KAIST", "country": "", "position": "MS student" }, { "name": "Jinwoo Shin", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" }, { "name": "Joseph J Lim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Kimin Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Minho Heo", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Honglak Lee", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 75, "n_ref": 158, "n_ref_all": 202, "n_fig": 18, "n_tab": 11, "L_tab": 5810, "n_element_tab": 416, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 864, "n_element_tab_1": 61, "formula_len_all": 530, "formula_len_all_1": 543, "len_all": 203617, "len_all_1": 64862, "len_abs": 1388, "len_title": 114, "len_sents": 53000, "len_sents_1": 30686, "n_sents": 340, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1393, "title": "Subtask-Aware Visual Reward Learning from Segmented Demonstrations", "abs": "Reinforcement Learning (RL) agents have demonstrated their potential across various robotic tasks. However, they still heavily rely on human-engineered reward functions, requiring extensive trial-and-error and access to target behavior information, often unavailable in real-world settings. This paper introduces REDS: REward learning from Demonstration with Segmentations, a novel reward learning framework that leverages action-free videos with minimal supervision. Specifically, REDS employs video demonstrations segmented into subtasks from diverse sources and treats these segments as ground-truth rewards. We train a dense reward function conditioned on video segments and their corresponding subtasks to ensure alignment with ground-truth reward signals by minimizing the Equivalent-Policy Invariant Comparison distance. Additionally, we employ contrastive learning objectives to align video representations with subtasks, ensuring precise subtask inference during online interactions. Our experiments show that REDS significantly outperforms baseline methods on complex robotic manipulation tasks in Meta-World and more challenging real-world tasks, such as furniture assembly in FurnitureBench, with minimal human intervention. Moreover, REDS facilitates generalization to unseen tasks and robot embodiments, highlighting its potential for scalable deployment in diverse environments.", "keywords": [ "Reinforcement Learning", "Reward Learning", "Robotic Manipulation" ], "rating_list": [ 6, 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 3, 4, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 3 ], "confidence_list": [ 2, 5, 4, 3, 3 ] }, { "paper_id": "moXtEmCleY", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Alireza Rezazadeh", "gender": "unknown", "institution": "Accenture", "country": "US", "position": "Research Scientist" }, { "name": "Yujia Bao", "gender": "Male", "institution": "Accenture", "country": "US", "position": "Researcher" }, { "name": "Zichao Li", "gender": "Male", "institution": "Accenture", "country": "US", "position": "Researcher" }, { "name": "wei wei", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" } ], "n_formula": 12, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 58, "n_ref_all": 86, "n_fig": 6, "n_tab": 11, "L_tab": 15348, "n_element_tab": 806, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1126, "n_element_tab_1": 135, "formula_len_all": 632, "formula_len_all_1": 260, "len_all": 208265, "len_all_1": 64612, "len_abs": 2937, "len_title": 145, "len_sents": 70278, "len_sents_1": 29806, "n_sents": 520, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1119, "title": "From Isolated Conversations to Hierachical Schemas: Dynamic Tree Memory Representation for LLMs", "abs": "Recent advancements in large language models have significantly improved their context windows, yet challenges in effective long-term memory management remain. We introduce MemTree, an algorithm that leverages a dynamic, tree-structured memory representation to optimize the organization, retrieval, and integration of information, akin to human cognitive schemas. MemTree organizes memory hierarchically, with each node encapsulating aggregated textual content, corresponding semantic embeddings, and varying abstraction levels across the tree's depths. Our algorithm dynamically adapts this memory structure by computing and comparing semantic embeddings of new and existing information to enrich the model\u2019s context-awareness. This approach allows MemTree to handle complex reasoning and extended interactions more effectively than traditional memory augmentation methods, which often rely on flat lookup tables. Evaluations on benchmarks for multi-turn dialogue understanding and document question answering show that MemTree significantly enhances performance in scenarios that demand structured memory management.", "keywords": [ "Tree-based Memory", "Inference-time Reasoning", "Large language models", "Long-term memory" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "moWiYJuSGF", "primary_area": "applications to robotics, autonomy, planning", "n_author": 9, "author_info_list": [ { "name": "Dongha Lee", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Gwanwoo Song", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Undergrad student" }, { "name": "Hyungjoo Chae", "gender": "Male", "institution": "Yonsei University", "country": "", "position": "MS student" }, { "name": "Jihoon Kim", "gender": "unknown", "institution": "Yonsei University", "country": "KR", "position": "Undergrad student" }, { "name": "Jinyoung Yeo", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Kai Tzu-iunn Ong", "gender": "Male", "institution": "Yonsei University", "country": "", "position": "PhD student" }, { "name": "Minju Gwak", "gender": "unknown", "institution": "Yonsei University", "country": "KR", "position": "Intern" }, { "name": "Namyoung Kim", "gender": "unknown", "institution": "Yonsei University", "country": "KR", "position": "PhD student" }, { "name": "Sunghwan Kim", "gender": "Male", "institution": "Yonsei University", "country": "", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 39, "n_ref": 293, "n_ref_all": 347, "n_fig": 20, "n_tab": 10, "L_tab": 2472, "n_element_tab": 204, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 1586, "n_element_tab_1": 202, "formula_len_all": 521, "formula_len_all_1": 286, "len_all": 205071, "len_all_1": 63336, "len_abs": 1288, "len_title": 140, "len_sents": 68180, "len_sents_1": 27389, "n_sents": 586, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1323, "title": "Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation", "abs": "Large language models (LLMs) have recently gained much attention in building autonomous agents. However, performance of current LLM-based web agents in long-horizon tasks is far from optimal, often yielding errors such as repeatedly buying a non-refundable flight ticket. By contrast, humans can avoid such an irreversible mistake, as we have an awareness of the potential outcomes (e.g., losing money) of our actions, also known as the \"world model\". Motivated by this, our study first starts with preliminary analyses, confirming the absence of world models in current LLMs (e.g., GPT-4o, Claude-3.5-Sonnet, etc.). Then, we present a World-model-augmented (WMA) web agent, which simulates the outcomes of its actions for better decision-making. To overcome the challenges in training LLMs as world models predicting next observations, such as repeated elements across observations and long HTML inputs, we propose a transition-focused observation abstraction, where the prediction objectives are free-form natural language descriptions exclusively highlighting important state differences between time steps. Experiments on WebArena and Mind2Web show that our world models improve agents' policy selection without training and demonstrate our agents' cost- and time-efficiency compared to recent tree-search-based agents.", "keywords": [ "Web Agent", "World Model", "Digital Agent", "Planning", "LLM" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "mnwlhvmKMN", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Chuang Gan", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Haoyu Zhen", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" }, { "name": "Pengxiao Han", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "MS student" }, { "name": "Qiao Sun", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Siyuan Zhou", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yilun Du", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 19, "n_ref_uni": 64, "n_ref": 101, "n_ref_all": 113, "n_fig": 6, "n_tab": 7, "L_tab": 726, "n_element_tab": 65, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 254, "n_element_tab_1": 15, "formula_len_all": 1167, "formula_len_all_1": 1197, "len_all": 143020, "len_all_1": 62466, "len_abs": 1579, "len_title": 81, "len_sents": 33874, "len_sents_1": 29209, "n_sents": 237, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 33, "L_abs": 1510, "title": "Learning 4D Embodied World Models", "abs": "In this paper, we present a 4D embodied world model, which takes in an image observation and language instruction as input and predicts a 4D dynamic mesh predicting how the scene will change as the embodied agent performs actions based on the given instructions. In contrast to previously learned world models which typically generate 2D videos, our 4D model provides detailed 3D information on precise configurations and shape of objects in a scene over time.\nThis allows us to effectively learn accurate inverse dynamic models for an embodied agent to execute a policy for interacting with the environment.\nTo construct a dataset to train such 4D world models, we first annotate large-scale existing video robotics dataset using pretrained depth and normal prediction models to construct 3D consistent 4D models of each video. To efficiently learn generative models on this 4D data, we propose to train a video generative model on this annotated dataset, which jointly predicts RGB-DN (RGB, Depth, and Normal) for each video. We then present an algorithm to directly convert generated RGB, Depth and Normal images into high-quality dynamic 4D mesh models of the world. We illustrate how this enables us to predict high-quality meshes consistent across both time and space from embodied scenarios, render novel views for embodied scenes, as well as construct policies that substantially outperform those from prior 2D and 3D models of the world. Our code, model, and dataset will be made publicly available.", "keywords": [ "Embodied AI", "World Model" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "mnna9LUg7P", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Chi-Chih Chang", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Diana Marculescu", "gender": "Female", "institution": "University of Texas, Austin", "country": "US", "position": "Full Professor" }, { "name": "HungYueh Chiang", "gender": "Male", "institution": "The University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Kai-Chiang Wu", "gender": "unknown", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" }, { "name": "Natalia Frumkin", "gender": "Female", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 36, "n_ref": 81, "n_ref_all": 134, "n_fig": 15, "n_tab": 13, "L_tab": 5118, "n_element_tab": 648, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1522, "formula_len_all_1": 549, "len_all": 196466, "len_all_1": 60798, "len_abs": 2798, "len_title": 124, "len_sents": 49833, "len_sents_1": 28922, "n_sents": 408, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 2096, "title": "Quamba: A Post-Training Quantization Recipe for Selective State Space Models", "abs": "State Space Models (SSMs) have emerged as an appealing alternative to Transformers for large language models, achieving state-of-the-art accuracy with constant memory complexity which allows for holding longer context lengths than attention-based networks. The superior computational efficiency of SSMs in long sequence modeling positions them favorably over Transformers in many scenarios. However, improving the efficiency of SSMs on request-intensive cloud-serving and resource-limited edge applications is still a formidable task. SSM quantization is a possible solution to this problem, making SSMs more suitable for wide deployment, while still maintaining their accuracy. Quantization is a common technique to reduce the model size and to utilize the low bit-width acceleration features on modern computing units, yet existing quantization techniques are poorly suited for SSMs. Most notably, SSMs have highly sensitive feature maps within the selective scan mechanism (i.e., linear recurrence) and massive outliers in the output activations which are not present in the output of token-mixing in the self-attention modules. To address this issue, we propose a static 8-bit per-tensor SSM quantization method which suppresses the maximum values of the input activations to the selective SSM for finer quantization precision and quantizes the output activations in an outlier-free space with Hadamard transform. Our 8-bit weight-activation quantized Mamba 2.8B SSM benefits from hardware acceleration and achieves a 1.72 $\\times$ lower generation latency on an Nvidia Orin Nano 8G, with only a 0.9\\% drop in average accuracy on zero-shot tasks. When quantizing Jamba, a 52B parameter SSM-style language model, we observe only a $1\\%$ drop in accuracy, demonstrating that our SSM quantization method is both effective and scalable for large language models, which require appropriate compression techniques for deployment. The experiments demonstrate the effectiveness and practical applicability of our approach for deploying SSM-based models of all sizes on both cloud and edge platforms.", "keywords": [ "State Space Models", "Model quantization" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "mnLmmtW7HO", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Dongmin Park", "gender": "Male", "institution": "KRAFTON", "country": "KR", "position": "Researcher" }, { "name": "Jae-Gil Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Jaehyun Park", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "MS student" } ], "n_formula": 45, "n_formula_1": 23, "n_ref_uni": 44, "n_ref": 88, "n_ref_all": 127, "n_fig": 7, "n_tab": 9, "L_tab": 5969, "n_element_tab": 261, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 7241, "n_element_tab_1": 220, "formula_len_all": 3759, "formula_len_all_1": 2031, "len_all": 160265, "len_all_1": 68252, "len_abs": 1253, "len_title": 125, "len_sents": 43765, "len_sents_1": 25328, "n_sents": 348, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1162, "title": "Active Learning for Continual Learning: Keeping the Past Alive in the Present", "abs": "*Continual learning (CL)* enables deep neural networks to adapt to ever-changing data distributions. In practice, there may be scenarios where annotation is costly, leading to *active continual learning (ACL)*, which performs *active learning (AL)* for the CL scenarios when reducing the labeling cost by selecting the most informative subset is preferable. However, conventional AL strategies are not suitable for ACL, as they focus solely on learning the new knowledge, leading to *catastrophic forgetting* of previously learned tasks. Therefore, ACL requires a new AL strategy that can balance the prevention of catastrophic forgetting and the ability to quickly learn new tasks. In this paper, we propose **AccuACL**, **Accu**mulated informativeness-based **A**ctive **C**ontinual **L**earning, by achieving an optimal balance between the two required capabilities of ACL, as well as alleviating the scalability issue of Fisher information-based AL. Extensive experiments demonstrate that AccuACL significantly outperforms AL baselines across various CL algorithms, increasing the average accuracy and forgetting by 23.8% and 17.0%, respectively, in average.", "keywords": [ "active learning", "continual learning", "Fisher information" ], "rating_list": [ 6, 8, 3, 8, 8 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 4, 3 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 3, 4 ] }, { "paper_id": "mnB4hDTIDr", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Bingzhe Wu", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Huizhe Zhang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Jintang Li", "gender": "Male", "institution": "Sun Yat-sen University", "country": "CN", "position": "PhD student" }, { "name": "Peilin Zhao", "gender": "unknown", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Yuchang Zhu", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "PhD student" }, { "name": "Zibin Zheng", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Liang Chen", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "Associate Professor" }, { "name": "An Bian", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Senior researcher " } ], "n_formula": 26, "n_formula_1": 11, "n_ref_uni": 79, "n_ref": 126, "n_ref_all": 155, "n_fig": 8, "n_tab": 12, "L_tab": 3476, "n_element_tab": 328, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1458, "n_element_tab_1": 109, "formula_len_all": 2451, "formula_len_all_1": 687, "len_all": 231215, "len_all_1": 66216, "len_abs": 1340, "len_title": 76, "len_sents": 66343, "len_sents_1": 30342, "n_sents": 538, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1273, "title": "Evaluating Diversity of LLM-generated Datasets: A Classification Perspective", "abs": "LLM-generated datasets have been recently leveraged as training data to mitigate data scarcity in specific domains. However, these LLM-generated datasets exhibit limitations on training models due to a lack of diversity, which underscores the need for effective diversity evaluation. Despite the growing demand, the diversity evaluation of LLM-generated datasets remains under-explored. To this end, we propose a diversity evaluation method for LLM-generated datasets from a classification perspective, namely, DCScore. Specifically, DCScore treats the diversity evaluation as a sample classification task, considering mutual relationships among samples. We further provide theoretical verification of the diversity-related axioms satisfied by DCScore, demonstrating it as a principled diversity evaluation method. Additionally, we show that existing methods can be incorporated into our proposed method in a unified manner. Meanwhile, DCScore enjoys much lower computational costs compared to existing methods. Finally, we conduct experiments on LLM-generated datasets to validate the effectiveness of DCScore. The experimental results indicate that DCScore correlates better with various diversity pseudo-truths of evaluated datasets, thereby verifying its effectiveness.", "keywords": [ "Diversity evaluation", "LLM-generated dataset", "Large language models" ], "rating_list": [ 5, 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 1, 2 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 2, 4, 3 ] }, { "paper_id": "mn61GWpEiK", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" }, { "name": "Xiaoda Liu", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "", "position": "Researcher" }, { "name": "Yoonsung Jeong", "gender": "Female", "institution": "Northwestern University, Northwestern University", "country": "US", "position": "MS student" }, { "name": "Zhenyu Pan", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 25, "n_ref": 67, "n_ref_all": 73, "n_fig": 2, "n_tab": 2, "L_tab": 1433, "n_element_tab": 56, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1433, "n_element_tab_1": 56, "formula_len_all": 777, "formula_len_all_1": 777, "len_all": 95474, "len_all_1": 61968, "len_abs": 935, "len_title": 125, "len_sents": 30422, "len_sents_1": 27704, "n_sents": 232, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 938, "title": "HeteGraph-Mamba: Heterogeneous Graph Learning via Selective State Space Model", "abs": "We propose a heterogeneous graph mamba network (HGMN) as the first exploration in leveraging the selective state space models (SSSMs) for heterogeneous graph learning. Compared with the literature, our HGMN overcomes two major challenges: (i) capturing long-range dependencies among heterogeneous nodes and (ii) adapting SSSMs to heterogeneous graph data. Our key contribution is a general graph architecture that can solve heterogeneous nodes in real-world scenarios, followed an efficient flow. Methodologically, we introduce a two-level efficient tokenization approach that first captures long-range dependencies within identical node types, and subsequently across all node types. Empirically, we conduct comparisons between our framework and 19 state-of-the-art methods on the heterogeneous benchmarks. The extensive comparisons demonstrate that our framework outperforms other methods in both the accuracy and efficiency dimensions.", "keywords": [ "graph neural network", "state space model" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "mmpVsDjVgn", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Charalampos P. Andriotis", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "Pablo G. Morato", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "Postdoc" }, { "name": "Seyran Khademi", "gender": "Female", "institution": "Delft University of Technology", "country": "NL", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 39, "n_ref": 59, "n_ref_all": 87, "n_fig": 14, "n_tab": 4, "L_tab": 3235, "n_element_tab": 392, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 2503, "n_element_tab_1": 28, "formula_len_all": 1343, "formula_len_all_1": 1259, "len_all": 129084, "len_all_1": 55890, "len_abs": 1166, "len_title": 117, "len_sents": 42194, "len_sents_1": 25838, "n_sents": 280, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1170, "title": "ConBatch-BAL: Batch Bayesian Active Learning under Budget Constraints", "abs": "Varying annotation costs among data points and budget constraints can hinder the adoption of active learning strategies in real-world applications. This work introduces two Bayesian active learning strategies for batch acquisition under constraints (ConBatch-BAL), one based on dynamic thresholding and one following greedy acquisition. Both select samples using uncertainty metrics computed via Bayesian neural networks. The dynamic thresholding strategy redistributes the budget across the batch, while the greedy one selects the top-ranked sample at each step, limited by the remaining budget. Focusing on scenarios with costly data annotation and geospatial constraints, we also release two new real-world datasets containing geolocated aerial images of buildings, annotated with energy efficiency or typology classes. The ConBatch-BAL strategies are benchmarked against a random acquisition baseline on these datasets under various budget and cost scenarios. The results show that the developed ConBatch-BAL strategies can reduce active learning iterations and data acquisition costs in real-world settings, and even outperform the unconstrained baseline solutions.", "keywords": [ "Bayesian Active Learning", "Real-World Datasets", "Batch Acquisition", "Budget Constraints" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "mmGc0TR8zB", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chenyang Wang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Junjun Jiang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xianming Liu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xingyu Hu", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 23, "n_ref_uni": 52, "n_ref": 110, "n_ref_all": 131, "n_fig": 5, "n_tab": 13, "L_tab": 4306, "n_element_tab": 338, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1309, "n_element_tab_1": 66, "formula_len_all": 3308, "formula_len_all_1": 1845, "len_all": 196381, "len_all_1": 66653, "len_abs": 1247, "len_title": 152, "len_sents": 52825, "len_sents_1": 28747, "n_sents": 427, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1251, "title": "Seeking Global Flat Minima in Federated Domain Generalization via Constrained Adversarial Augmentation", "abs": "Federated domain generalization (FedDG) aims at equipping the federally trained model with the domain generalization ability when the model meets new clients with domain shifts. Among factors that possibly indicate generalization, the loss landscape flatness of the trained model is an intuitive, viable, and widely studied one. However, pursuing the flatness of the global model in the FedDG setting is not trivial due to the restriction to preserve data privacy. To address this issue, we propose GFM, a novel algorithm designed to seek Global Flat Minima of the global model. Specifically, GFM leverages a global model-constrained adversarial data augmentation strategy, creating a surrogate for global data within each local client, which allows for split sharpness-aware minimization to approach global flat minima. GFM is compatible with federated learning without compromising data privacy restrictions, and theoretical analysis further supports its rationality by demonstrating that the objective of GFM serves as an upper bound on the robust risk of the global model on global data distribution. Extensive experiments on multiple FedDG benchmarks demonstrate that GFM consistently outperforms previous FedDG and federated learning approaches.", "keywords": [ "Federated Domain Generalization", "Flat Minima", "Data Augmentation" ], "rating_list": [ 8, 6, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "mmDkgLtYNI", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Daniel Worrall", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "J. Nathan Kutz", "gender": "Male", "institution": "University of Washington", "country": "", "position": "Full Professor" }, { "name": "Miles Cranmer", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "Assistant Professor" }, { "name": "Peter Battaglia", "gender": "Male", "institution": "DeepMind", "country": "", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 7, "n_ref_uni": 32, "n_ref": 62, "n_ref_all": 90, "n_fig": 14, "n_tab": 5, "L_tab": 1795, "n_element_tab": 108, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 532, "n_element_tab_1": 7, "formula_len_all": 712, "formula_len_all_1": 348, "len_all": 133258, "len_all_1": 62555, "len_abs": 1100, "len_title": 90, "len_sents": 42332, "len_sents_1": 31082, "n_sents": 349, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1107, "title": "Spectral Shaping for Neural PDE Surrogates", "abs": "Neural surrogates for PDE solvers suffer from an inability to model the spectrum of solutions adequately, especially in the medium to high frequency bands. This impacts not only correct spectral shapes, but also stability and long-term rollout accuracy. We identify three convergent factors that exacerbate this phenomenon, namely: distribution shift over unrolls, spectral bias of the MSE loss, and spurious high frequency noise, or _spectral junk_, introduced by the use of pointwise nonlinearities. We find that _spectral shaping_, filtering the spectrum of activations after every layer of pointwise nonlinearities, is enough to reduce spectral junk and improve long-term rollout accuracy. We show spectral shaping not only fixes the learned spectrum (down to machine precision in some cases), but also leads to very stable neural surrogates. We validate these findings on a suite of challenging fluid dynamics problems in the field of neural PDE surrogacy, promoting a clear need for more careful attention to surrogate architecture design and adding a new and simple trick to the practitioner toolbox.", "keywords": [ "PDEs", "Fluid Mechanics", "Dynamical systems", "Autoregressive models", "Spectral methods" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 2, 5 ] }, { "paper_id": "mm0cqJ2O3f", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Huaijin Pi", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Hujun Bao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Qing Shuai", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Ruizhen Hu", "gender": "Female", "institution": "Shenzhen University", "country": "CN", "position": "Full Professor" }, { "name": "Sida Peng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaowei Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yujun Shen", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhi Cen", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 45, "n_ref": 77, "n_ref_all": 117, "n_fig": 7, "n_tab": 12, "L_tab": 4540, "n_element_tab": 745, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1821, "n_element_tab_1": 305, "formula_len_all": 577, "formula_len_all_1": 577, "len_all": 158226, "len_all_1": 68803, "len_abs": 776, "len_title": 127, "len_sents": 42288, "len_sents_1": 28940, "n_sents": 415, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1408, "title": "Ready-to-React: Online Reaction Policy for Two-Character Interaction Generation", "abs": "This paper addresses the task of generating two-character online interactions. Previously, two main settings existed for two-character interaction generation: (1) generating one's motions based on the counterpart's complete motion sequence, and (2) jointly generating two-character motions based on specific conditions. We argue that these settings fail to model the process of real-life two-character interactions, where humans will react to their counterparts in real time and act as independent individuals. In contrast, we propose an online reaction policy, called Ready-to-React, to generate the next character pose based on past observed motions. Each character has its own reaction policy as its ``brain'', enabling them to interact like real humans in a streaming manner. Our policy is implemented by incorporating a diffusion head into an auto-regressive model, which can dynamically respond to the counterpart's motions while effectively mitigating the error accumulation throughout the generation process. We conduct comprehensive experiments using the challenging boxing task. Experimental results demonstrate that our method outperforms existing baselines and can generate extended motion sequences. Additionally, we show that our approach can be controlled by sparse signals, making it well-suited for VR and other online interactive environments. Code and data will be made publicly available.", "keywords": [ "Human Interaction Generation", "Reactive Motion Generation" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "mltelO89Ve", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "George Karypis", "gender": "Male", "institution": "University of Minnesota, Minneapolis", "country": "", "position": "Full Professor" }, { "name": "Huzefa Rangwala", "gender": "Male", "institution": "Amazon", "country": "", "position": "Principal Researcher" }, { "name": "Mingyi Hong", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Rasool Fakoor", "gender": "Male", "institution": "Amazon Web Services", "country": "", "position": "Researcher" }, { "name": "Siliang Zeng", "gender": "Male", "institution": "University of Minnesota, Twin Cities", "country": "US", "position": "PhD student" }, { "name": "Yao Liu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 39, "n_ref": 102, "n_ref_all": 129, "n_fig": 3, "n_tab": 4, "L_tab": 1401, "n_element_tab": 123, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 3111, "n_element_tab_1": 125, "formula_len_all": 918, "formula_len_all_1": 987, "len_all": 127876, "len_all_1": 68778, "len_abs": 1301, "len_title": 124, "len_sents": 35961, "len_sents_1": 31295, "n_sents": 203, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1305, "title": "From Demonstrations to Rewards: Alignment Without Explicit Human Preferences", "abs": "One of the challenges of aligning large models with human preferences lies in both the data requirements and the technical complexities of current approaches. Predominant methods, such as RLHF, involve multiple steps, each demanding distinct types of data, including demonstrations data and preference data. In RLHF, human preferences are typically modeled through a reward model, which serves as a proxy to guide policy learning during the reinforcement learning stage, ultimately producing a policy aligned with human preferences. However, in this paper, we propose a fresh perspective on learning alignment based on inverse reinforcement learning principles, where the optimal policy is still derived from reward maximization. However, instead of relying on preference data, we directly learn the reward model from demonstration data. This new formulation offers the flexibility to be applied even when only demonstration data is available, a capability that current RLHF methods lack, and it also shows that demonstration data offers more utility than what conventional wisdom suggests. Our extensive evaluation, based on public reward benchmark and HuggingFace Open LLM Leaderboard, demonstrates that our approach compares favorably to state-of-the-art methods that rely solely on demonstration data.", "keywords": [ "inverse reinforcement learning", "iterative RLHF", "learning from demonstrations", "reward learning" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "mlPTNEIsgb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Dohoon Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Hayoon Kim", "gender": "Female", "institution": "Seoul National University", "country": "", "position": "MS student" }, { "name": "Jaehyun Park", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "PhD student" }, { "name": "Jin Woo Lee", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Intern" }, { "name": "Kyogu Lee", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Professor" }, { "name": "Sungho Lee", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 14, "n_ref_uni": 55, "n_ref": 115, "n_ref_all": 128, "n_fig": 3, "n_tab": 7, "L_tab": 3360, "n_element_tab": 204, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 851, "n_element_tab_1": 32, "formula_len_all": 1780, "formula_len_all_1": 1697, "len_all": 187444, "len_all_1": 72039, "len_abs": 1294, "len_title": 123, "len_sents": 46609, "len_sents_1": 30897, "n_sents": 349, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1308, "title": "Solving Blind Non-linear Forward and Inverse Problem for Audio Applications", "abs": "We propose a unified framework to address the blind forward and inverse problems in audio domain, where the objective is to estimate either the function or the input signal solely from the observed output, without access to the other.\nWe formally define forward operators ---mapping input to output signals --- and formulate both problems within a probabilistic framework.\nFor the blind forward problem, we design an architecture that utilizes a reference encoder to extract features from the reference signal, enabling the main operator to approximate arbitrary forward operators systematically composed via algebraic representations. \nFor the blind inverse problem, we employ a conditional diffusion model conditioned on features from the pretrained reference encoder and augment the generation process using twisted particle filtering technique leveraging the approximated operator in the forward problem.\nWe validate our framework on zero-shot audio effect modeling and speech enhancement. The experiments show that our approach replicates both simple and complex audio effects, generalizes under distribution mismatches, and effectively enhances noisy full-band audio across diverse effects and real-world scenarios. \nCodes are available at https://t.ly/n11uk , with audio samples at https://t.ly/dBUhF", "keywords": [ "audio effect", "zero-shot system identification", "inverse problem", "diffusion model", "sequential monte carlo" ], "rating_list": [ 1, 5, 6, 1 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 2, 3, 5 ] }, { "paper_id": "mlCRJnETWz", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 15, "author_info_list": [ { "name": "Baixiang Huang", "gender": "Male", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Canyu Chen", "gender": "unknown", "institution": "Illinois Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Dawn Song", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Huaxiu Yao", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Jia-Chen Gu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Jindong Gu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Kai Shu", "gender": "unknown", "institution": "Emory University", "country": "US", "position": "Assistant Professor" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Shiyang Lai", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Xifeng Yan", "gender": "unknown", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" }, { "name": "Xiongxiao Xu", "gender": "unknown", "institution": "Illinois Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zekun Li", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Zhaorun Chen", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "chaowei xiao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "William Yang Wang", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 68, "n_ref": 140, "n_ref_all": 147, "n_fig": 1, "n_tab": 9, "L_tab": 7304, "n_element_tab": 203, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 272813, "len_all_1": 68438, "len_abs": 1716, "len_title": 77, "len_sents": 63799, "len_sents_1": 35689, "n_sents": 436, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 29, "L_abs": 1724, "title": "Can Editing LLMs Inject Harm?", "abs": "Knowledge editing has been increasingly adopted to correct the false or outdated knowledge in Large Language Models (LLMs). Meanwhile, one critical but under-explored question is: can knowledge editing be used to inject harm into LLMs? In this paper, we propose to reformulate knowledge editing as a new type of safety threat for LLMs, namely Editing Attack, and conduct a systematic investigation with a newly constructed dataset EditAttack. Specifically, we focus on two typical safety risks of Editing Attack including Misinformation Injection and Bias Injection. For the risk of misinformation injection, we first categorize it into commonsense misinformation injection and long-tail misinformation injection. Then, we find that editing attacks can inject both types of misinformation into LLMs, and the effectiveness is particularly high for commonsense misinformation injection. For the risk of bias injection, we discover that not only can biased sentences be injected into LLMs with high effectiveness, but also one single biased sentence injection can cause a bias increase in general outputs of LLMs, which are even highly irrelevant to the injected sentence, indicating a catastrophic impact on the overall fairness of LLMs. Then, we further illustrate the high stealthiness of editing attacks, measured by their impact on the general knowledge and reasoning capacities of LLMs, and show the hardness of defending editing attacks with empirical evidence. Our discoveries demonstrate the emerging misuse risks of knowledge editing techniques on compromising the safety alignment of LLMs and the feasibility of disseminating misinformation or bias with LLMs as new channels. The code and dataset are available here.", "keywords": [ "Knowledge Editing", "LLM safety", "Harm Injection" ], "rating_list": [ 6, 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 1 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 5, 3, 3 ] }, { "paper_id": "mkuB677eMM", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Bin CAO", "gender": "Male", "institution": "Hong Kong University of Science and Technology(Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Jia Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Ruifeng Tan", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Tong-yi Zhang", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Yang Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zinan Zheng", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 84, "n_ref_all": 107, "n_fig": 12, "n_tab": 8, "L_tab": 7897, "n_element_tab": 1233, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 6692, "n_element_tab_1": 898, "formula_len_all": 382, "formula_len_all_1": 0, "len_all": 197923, "len_all_1": 72046, "len_abs": 1176, "len_title": 146, "len_sents": 55925, "len_sents_1": 26478, "n_sents": 419, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1180, "title": "SimXRD-4M: Big Simulated X-ray Diffraction Data and Crystal Symmetry Classification Benchmark", "abs": "Powder X-ray diffraction (XRD) patterns are highly effective for crystal identification and play a pivotal role in materials discovery. While machine learning (ML) has advanced the analysis of powder XRD patterns, progress has been constrained by the limited availability of training data and established benchmarks. To address this, we introduce SimXRD, the largest open-source simulated XRD pattern dataset to date, aimed at accelerating the development of crystallographic informatics. We developed a novel XRD simulation method that incorporates comprehensive physical interactions, resulting in a high-fidelity database. SimXRD comprises 4,065,346 simulated powder XRD patterns, representing 119,569 unique crystal structures under 33 simulated conditions that reflect real-world variations. We benchmark 21 sequence models in both in-library and out-of-library scenarios and analyze the impact of class imbalance in long-tailed crystal label distributions. Remarkably, we find that: (1) current neural networks struggle with classifying low-frequency crystals, particularly in out-of-library situations; (2) models trained on SimXRD can generalize to real experimental data.", "keywords": [ "Symmetry Classification", "benchmark and dataset" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "mkXi7O0fun", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Bryan Kian Hsiang Low", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Associate Professor" }, { "name": "Daniela Rus", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Xinyi Xu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zijian Zhou", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" } ], "n_formula": 148, "n_formula_1": 22, "n_ref_uni": 48, "n_ref": 111, "n_ref_all": 154, "n_fig": 8, "n_tab": 12, "L_tab": 2175, "n_element_tab": 117, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 567, "n_element_tab_1": 18, "formula_len_all": 29374, "formula_len_all_1": 1945, "len_all": 265518, "len_all_1": 76188, "len_abs": 1182, "len_title": 90, "len_sents": 70361, "len_sents_1": 30899, "n_sents": 785, "n_sents_1": 300, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1135, "title": "Data Value Estimation on Private Gradients", "abs": "For gradient-based machine learning (ML) methods commonly adopted in practice such as stochastic gradient descent, the de facto differential privacy (DP) technique is perturbing the gradients with random Gaussian noise. Data valuation attributes the ML performance to the training data and is widely used in privacy-aware applications that require enforcing DP such as data pricing, collaborative ML, and federated learning (FL). Can existing data valuation methods still be used when DP is enforced via gradient perturbations? We show that the answer is no with the default approach of injecting i.i.d. random noise to the gradients because the estimation uncertainty of the data value estimation paradoxically linearly scales with more estimation budget, producing estimates almost like random guesses. To address this issue, we propose to instead inject carefully correlated noise to provably remove the linear scaling of estimation uncertainty w.r.t. the budget. We also empirically demonstrate that our method gives better data value estimates on various ML tasks and is applicable to use cases including dataset valuation and FL.", "keywords": [ "data valuation", "estimation uncertainty", "privacy-aware" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "mkNVPGpEPm", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Ivan Oseledets", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "Fanaskov Vladimir", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Researcher" } ], "n_formula": 52, "n_formula_1": 23, "n_ref_uni": 27, "n_ref": 100, "n_ref_all": 133, "n_fig": 5, "n_tab": 1, "L_tab": 362, "n_element_tab": 5, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4170, "formula_len_all_1": 2307, "len_all": 133707, "len_all_1": 61828, "len_abs": 1647, "len_title": 83, "len_sents": 46653, "len_sents_1": 27109, "n_sents": 373, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 35, "L_abs": 1653, "title": "Associative memory and dead neurons", "abs": "In ``Large Associative Memory Problem in Neurobiology and Machine Learning,'' Dmitry Krotov and John Hopfield introduced a general technique for the systematic construction of neural ordinary differential equations with non-increasing energy or Lyapunov function. We study this energy function and identify that it is vulnerable to the problem of dead neurons. Each point in the state space where the neuron dies is contained in a non-compact region with constant energy. In these flat regions, energy function alone does not completely determine all degrees of freedom and, as a consequence, can not be used to analyze stability or find steady states or basins of attraction. We perform a direct analysis of the dynamical system and show how to resolve problems caused by flat directions corresponding to dead neurons: (i) all information about the state vector at a fixed point can be extracted from the energy and Hessian matrix (of Lagrange function), (ii) it is enough to analyze stability in the range of Hessian matrix, (iii) if steady state touching flat region is stable the whole flat region is the basin of attraction. The analysis of the Hessian matrix can be complicated for realistic architectures, so we show that for a slightly altered dynamical system (with the same structure of steady states), one can derive a diverse family of Lyapunov functions that do not have flat regions corresponding to dead neurons. In addition, these energy functions allow one to use Lagrange functions with Hessian matrices that are not necessarily positive definite and even consider architectures with non-symmetric feedforward and feedback connections.", "keywords": [ "associative memory", "dead neurone", "neural ODE", "energy-based model" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "mkE9Yx4wHY", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Ata Nizamoglu", "gender": "unknown", "institution": "Technische Universit\u00e4t Berlin", "country": "DE", "position": "MS student" }, { "name": "Luis Felipe Villa-Arenas", "gender": "Male", "institution": "Technische Universit\u00e4t Berlin", "country": "DE", "position": "PhD student" }, { "name": "Qianli Wang", "gender": "unknown", "institution": "Technische Universit\u00e4t Berlin", "country": "DE", "position": "PhD student" }, { "name": "Sebastian M\u00f6ller", "gender": "Male", "institution": "Technische Universit\u00e4t Berlin", "country": "", "position": "Full Professor" }, { "name": "Vera Schmitt", "gender": "Female", "institution": "Technische Universit\u00e4t Berlin", "country": "", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 5, "n_ref_uni": 31, "n_ref": 74, "n_ref_all": 91, "n_fig": 2, "n_tab": 5, "L_tab": 1867, "n_element_tab": 168, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 400, "n_element_tab_1": 10, "formula_len_all": 1300, "formula_len_all_1": 483, "len_all": 167957, "len_all_1": 61929, "len_abs": 902, "len_title": 100, "len_sents": 51570, "len_sents_1": 28520, "n_sents": 383, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 924, "title": "Anchored Alignment for Self-Explanations Enhancement", "abs": "In this work, we introduce a methodology for alignment designed to enhance the ability of large language models (LLMs) to articulate their reasoning\u2014\\textit{self-explanation}\u2014even in the absence of annotated rationale explanations. Our alignment methodology comprises three key components: explanation quality assessment, self-instruction dataset generation, and model alignment. Additionally, we present a novel technique called \\textit{Alignment with Anchor Preference Pairs}, which improves the selection of preference pairs by categorizing model outputs into three groups: consistently correct, consistently incorrect, and variable. By applying tailored strategies to each category, we enhance the effectiveness of Direct Preference Optimization (DPO). Our experimental results demonstrate that this approach significantly improves explanation quality while maintaining accuracy compared to other fine-tuning strategies.", "keywords": [ "LLM", "Self-Explanation", "Alignment", "Preference Pairs", "DPO", "SFT", "RLAIF", "Self-Alignment", "Self-instruction" ], "rating_list": [ 8, 3, 6, 5 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 2, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "mkDam1xIzW", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Han-Lin Hsieh", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Postdoc" }, { "name": "Maryam Shanechi", "gender": "Female", "institution": "University of Southern California", "country": "", "position": "Full Professor" } ], "n_formula": 41, "n_formula_1": 20, "n_ref_uni": 37, "n_ref": 64, "n_ref_all": 124, "n_fig": 17, "n_tab": 3, "L_tab": 461, "n_element_tab": 65, "n_fig_1": 14, "n_tab_1": 2, "L_tab_1": 369, "n_element_tab_1": 43, "formula_len_all": 3542, "formula_len_all_1": 1695, "len_all": 143158, "len_all_1": 61550, "len_abs": 1965, "len_title": 84, "len_sents": 45695, "len_sents_1": 27069, "n_sents": 381, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 84, "L_abs": 1975, "title": "Probabilistic Geometric Principal Component Analysis with application to neural data", "abs": "Dimensionality reduction is critical across various domains of science including neuroscience. Probabilistic Principal Component Analysis (PPCA) is a prominent dimensionality reduction method that provides a probabilistic approach unlike the deterministic approach of PCA and serves as a connection between PCA and Factor Analysis (FA). Despite their power, PPCA and its extensions are mainly based on linear models and can only describe the data in a Euclidean coordinate system around the mean of data. However, in many neuroscience applications, data may be distributed around a nonlinear geometry (i.e., manifold) rather than lying in the Euclidean space around the mean. We develop Probabilistic Geometric Principal Component Analysis (PGPCA) for such datasets as a new dimensionality reduction algorithm that can explicitly incorporate knowledge about a given nonlinear manifold that is first fitted from these data. Further, we show how in addition to the Euclidean coordinate system, a geometric coordinate system can be derived for the manifold to capture the deviations of data from the manifold and noise. We also derive a data-driven EM algorithm for learning the PGPCA model parameters. As such, PGPCA generalizes PPCA to better describe data distributions by incorporating a nonlinear manifold geometry. In simulations and brain data analyses, we show that PGPCA can effectively model the data distribution around various given manifolds and outperforms PPCA for such data. Moreover, PGPCA provides the capability to test whether the new geometric coordinate system better describes the data than the Euclidean one. Finally, PGPCA can perform dimensionality reduction and learn the data distribution both around and on the manifold. These capabilities make PGPCA valuable for enhancing the efficacy of dimensionality reduction for analysis of high-dimensional data that exhibit noise and are distributed around a nonlinear manifold, especially for neural data.", "keywords": [ "geometry", "nonlinear manifold", "factor analysis", "dimensionality reduction", "neural population activity" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "mk8eLjKCdB", "primary_area": "foundation or frontier models, including LLMs", "n_author": 1, "author_info_list": [ { "name": "Prateek Verma", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Research Affiliate" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 50, "n_ref": 78, "n_ref_all": 106, "n_fig": 4, "n_tab": 1, "L_tab": 275, "n_element_tab": 40, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 394, "n_element_tab_1": 38, "formula_len_all": 634, "formula_len_all_1": 531, "len_all": 150327, "len_all_1": 72296, "len_abs": 1187, "len_title": 81, "len_sents": 46019, "len_sents_1": 35925, "n_sents": 333, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Reject", "L_title": 33, "L_abs": 1245, "title": "WaveletGPT: Wavelet Inspired LLMs", "abs": "Large Language Models (LLMs) have ushered in a new wave of artificial intelligence advancements impacting every scientific field and discipline. We live in a world where most of the data around us, e.g., text, audio, and music, has a multi-scale structure associated with it. This paper infuses LLMs with a traditional signal processing idea, namely wavelets, during pre-training to take advantage of the structure. Without adding \\textbf{any extra parameters} to a GPT-style LLM architecture in academic setup, we achieve the same pre-training performance almost twice as fast in text, raw audio, and symbolic music. This is achieved by imposing a structure on intermediate embeddings. When trained for the same number of training steps, we achieve significant gains in performance, which is comparable to pre-training a larger neural architecture. Our architecture allows every next token prediction access to intermediate embeddings at different temporal resolutions in every Transformer decoder block. This work will hopefully pave the way for incorporating multi-rate signal processing ideas into traditional LLM pre-training. Further, we showcase pushing model performance by improving internal structure instead of just going after scale.", "keywords": [ "Wavelets", "GPT", "LLM" ], "rating_list": [ 1, 6, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "mjtCqmujYP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Boyi Liu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Liyu Chen", "gender": "Male", "institution": "ByteDance Inc", "country": "US", "position": "Researcher" }, { "name": "Shenao Zhang", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Yingxiang Yang", "gender": "Male", "institution": "ByteDance Inc", "country": "", "position": "ByteDance" }, { "name": "Yongfei Liu", "gender": "Male", "institution": "Bytedance", "country": "", "position": "Researcher" }, { "name": "Yufeng Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "", "position": "Researcher" }, { "name": "Zhaoran Wang", "gender": "Not Specified", "institution": "Northwestern University", "country": "US", "position": "Associate Professor" }, { "name": "Zhihan Liu", "gender": "Male", "institution": "Northwestern University", "country": "", "position": "PhD student" }, { "name": "TAO SUN", "gender": "Male", "institution": "Bytedance", "country": "", "position": "Research Scientist" } ], "n_formula": 34, "n_formula_1": 10, "n_ref_uni": 69, "n_ref": 148, "n_ref_all": 165, "n_fig": 4, "n_tab": 13, "L_tab": 2959, "n_element_tab": 301, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3347, "n_element_tab_1": 93, "formula_len_all": 4516, "formula_len_all_1": 969, "len_all": 214872, "len_all_1": 66006, "len_abs": 2107, "len_title": 114, "len_sents": 51018, "len_sents_1": 29050, "n_sents": 379, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 2116, "title": "Reward-Augmented Data Enhances Direct Preference Alignment of LLMs", "abs": "Preference alignment in Large Language Models (LLMs) has significantly improved their ability to adhere to human instructions and intentions. However, existing direct alignment algorithms primarily focus on relative preferences and often overlook the qualitative aspects of responses, despite having access to preference data that includes reward scores from judge models during AI feedback. Striving to maximize the implicit reward gap between the chosen and the slightly inferior rejected responses can cause overfitting and unnecessary unlearning of the high-quality rejected responses. The unawareness of the reward scores also drives the LLM to indiscriminately favor the low-quality chosen responses and fail to generalize to responses with the highest rewards, which are sparse in data. To overcome these shortcomings, our study introduces reward-conditioned LLM policies that discern and learn from the entire spectrum of response quality within the dataset, helping extrapolate to more optimal regions. We propose an effective yet simple data relabeling method that conditions the preference pairs on quality scores to construct a reward-augmented dataset. This dataset is easily integrated with existing direct alignment algorithms and is applicable to any preference dataset. The experimental results across instruction-following benchmarks including AlpacaEval 2.0, MT-Bench, and Arena-Hard-Auto demonstrate that our approach consistently boosts the performance of DPO by a considerable margin across diverse models such as Zephyr, Mistral, Qwen2, Llama3.1, Gemma2, and SPPO. Additionally, on six academic benchmarks including GSM8K, GPQA, MUSR, TruthfulQA, BBH, and ARC, our method improves their average accuracy. When applying our method to on-policy data, the resulting DPO model outperforms various baselines and achieves state-of-the-art results on AlpacaEval 2.0. Through comprehensive ablation studies, we demonstrate that our method not only maximizes the utility of preference data but also mitigates the issue of unlearning, demonstrating its broad effectiveness beyond mere dataset expansion.", "keywords": [ "Preference Alignment", "Large Language Model", "RLAIF" ], "rating_list": [ 5, 3, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "mjDNVksC5G", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" }, { "name": "Hong-Yu Chen", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Jianqing Fan", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Professor" }, { "name": "Yihan He", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Yuan Cao", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Dennis Y.H. Wu", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "US", "position": "PhD student" } ], "n_formula": 110, "n_formula_1": 26, "n_ref_uni": 17, "n_ref": 37, "n_ref_all": 54, "n_fig": 4, "n_tab": 3, "L_tab": 1544, "n_element_tab": 63, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 411, "n_element_tab_1": 62, "formula_len_all": 15033, "formula_len_all_1": 2489, "len_all": 146787, "len_all_1": 60429, "len_abs": 3610, "len_title": 78, "len_sents": 47404, "len_sents_1": 25385, "n_sents": 516, "n_sents_1": 261, "venue": "ICLR2025", "decision": "withdraw", "L_title": 30, "L_abs": 1066, "title": "Can Transformers Perform PCA ?", "abs": "Transformers demonstrate significant advantage as the building block of Large Language Models. Recent efforts are devoted to understanding the learning capacities of transformers at a fundamental level. This work attempts to understand the intrinsic capacity of transformers in performing dimension reduction from complex data. Theoretically, our results rigorously show that transformers can perform Principle Component Analysis (PCA) similar to the Power Method, given a supervised pre-training phase. Moreover, we show the generalization error of transformers decays by $n^{-1/5}$ in $L_2$. Empirically, our extensive experiments on the simulated and real world high dimensional datasets justify that a pre-trained transformer can successfully perform PCA by simultaneously estimating the first $k$ eigenvectors and eigenvalues. These findings demonstrate that transformers can efficiently extract low dimensional patterns from high dimensional data, shedding light on the potential benefits of using pre-trained LLM to perform inference on high dimensional data.", "keywords": [ "Principle Component Analysis", "Transformers", "Machine Learning Theory" ], "rating_list": [ 5, 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 1, 1, 2, 2, 1 ], "contribution_list": [ 2, 1, 2, 1, 2 ], "confidence_list": [ 4, 2, 4, 3, 3 ] }, { "paper_id": "miOYgWl60q", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Christian Poellabauer", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "Full Professor" }, { "name": "Enshi Zhang", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 17, "n_ref_uni": 48, "n_ref": 73, "n_ref_all": 96, "n_fig": 5, "n_tab": 9, "L_tab": 1976, "n_element_tab": 469, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1257, "n_element_tab_1": 125, "formula_len_all": 907, "formula_len_all_1": 786, "len_all": 156932, "len_all_1": 58744, "len_abs": 1695, "len_title": 146, "len_sents": 48157, "len_sents_1": 26362, "n_sents": 380, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1716, "title": "Multimodal Depression Detection with Contextual Position Encoding and Latent Space Regularization", "abs": "Clinical interviews are the gold standard for detecting depression, and previous work using multimodal features from participants' audio, transcribed text, and video have shown promising results. Recent approaches further improve performance by incorporating an additional textual modality\u2014the interviewer\u2019s prompts\u2014during training. However, these approaches risk introducing biases, as models may over-rely on specific prompt-response pairs, which may not always be present in real-world settings. This leads to models exploiting these cues as shortcuts for detecting depression rather than learning the language and behaviors that genuinely indicate the subject's mental health, ultimately undermining consistency and objectivity. To address this, we propose a novel approach that combines Contextual Position Encoding (**CoPE**) and Latent Space Regularization (**LSR**), leveraging both subjects' responses (audio) and the interviewer's prompts (text). CoPE captures the evolving context of the interview, ensuring that the model utilizes insights from the entire conversation, preventing over-reliance on isolated or late-stage cues. This helps the model understand interactions holistically and more accurately reflect mental health indicators. LSR introduces constraints to enforce consistency in the model\u2019s learned representations, reducing overfitting to superficial cues and guiding the model toward more generalizable patterns. By smoothing the latent space, LSR helps the model focus on meaningful, high-level representations of both audio and text.\nOur approach yields competitive results on the **DAIC-WOZ** benchmark and surpasses the state-of-the-art on the **EATD** benchmark. The code is released.", "keywords": [ "depression detection", "speech and language processing", "contextual position encoding", "latent space regularization" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "miIE56qM10", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Heming Xia", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "", "position": "PhD student" }, { "name": "Jing Long", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Lei Sha", "gender": "Male", "institution": "Beihang University", "country": "", "position": "Full Professor" }, { "name": "Muge Qi", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Peiyi Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhifang Sui", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "\u674e\u777f", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" } ], "n_formula": 24, "n_formula_1": 23, "n_ref_uni": 41, "n_ref": 70, "n_ref_all": 85, "n_fig": 2, "n_tab": 4, "L_tab": 912, "n_element_tab": 32, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2122, "n_element_tab_1": 361, "formula_len_all": 1126, "formula_len_all_1": 1085, "len_all": 122003, "len_all_1": 69126, "len_abs": 1021, "len_title": 149, "len_sents": 35300, "len_sents_1": 29713, "n_sents": 240, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 1024, "title": "Lookers-On See Most of the Game: An External Insight-Guided Method for Enhancing Uncertainty Estimation", "abs": "Large Language Models (LLMs) have gained increasing attention for their impressive capabilities, alongside concerns about the reliability arising from their potential to generate hallucinations and factual inaccuracies. Uncertainty estimation for LLMs aims to quantify the uncertainty of model outputs, where high uncertainty scores indicate potential errors, signaling the need for rejection or further evaluation. However, existing methods often limited by inherent biases of LLMs like over-confidence and under-confidence. In this paper, we propose an external insight-driven correction method for refining uncertainty estimation. This method integrates uncertainty scores derived from a lightweight model trained on global information with those from existing uncertainty estimation approaches, providing a more robust solution. We present comprehensive experimental results that demonstrate the effectiveness and generalizability of our method across various models, datasets, and consistently surpassing all baselines.", "keywords": [ "Large Language Models", "Uncertainty Estimation", "Trusty AI" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 1, 2 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 1, 2 ], "confidence_list": [ 4, 3, 4, 5, 4 ] }, { "paper_id": "mi9GJkZt8n", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Mario Fritz", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Full Professor" }, { "name": "Marta Kwiatkowska", "gender": "Female", "institution": "Polish Academy of Sciences", "country": "PL", "position": "Full Professor" }, { "name": "Tobias Lorenz", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 24, "n_ref_uni": 19, "n_ref": 60, "n_ref_all": 75, "n_fig": 1, "n_tab": 2, "L_tab": 158, "n_element_tab": 51, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 213, "n_element_tab_1": 12, "formula_len_all": 1788, "formula_len_all_1": 1464, "len_all": 103316, "len_all_1": 63888, "len_abs": 1312, "len_title": 163, "len_sents": 34428, "len_sents_1": 27729, "n_sents": 304, "n_sents_1": 248, "venue": "ICLR2025", "decision": "withdraw", "L_title": 115, "L_abs": 1319, "title": "BiCert: A Blinear Mixed Integer Programming Formulation for Precise Certified Bounds Against Data Poisoning Attacks", "abs": "Data poisoning attacks pose one of the biggest threats to modern AI systems, necessitating robust defenses. While extensive efforts have been made to develop empirical defenses, attackers continue to evolve, creating sophisticated methods to circumvent these measures. To address this, we must move beyond empirical defenses and establish provable certification methods that guarantee robustness. This paper introduces a novel certification approach using Bilinear Mixed Integer Programming (BMIP) to compute sound, deterministic bounds that provide such provable robustness. Using BMIP, we compute the reachable set of parameters that could result from training with potentially manipulated data. A key insight to make this computation feasible is relaxing the reachable parameter set to a convex set between training iterations. At test time, this parameter set allows us to predict all possible outcomes, guaranteeing robustness. Our BMIP approach is more precise than previous methods, which rely solely on interval and polyhedral bounds. Crucially, it overcomes the fundamental limitation of prior approaches where parameter bounds could only grow, often uncontrollably. We show that these tighter bounds eliminate a key source of divergence issues, resulting in more stable training and higher certified accuracy.", "keywords": [ "Data Poisoning Defense", "Certified Robustness", "Provable Defenses", "Robust Machine Learning", "Adversarial Machine Learning" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "mhzDv7UAMu", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "David W. Romero", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Ming-Yu Liu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Zekun Hao", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Tsung-Yi Lin", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 7, "n_ref_uni": 53, "n_ref": 97, "n_ref_all": 134, "n_fig": 17, "n_tab": 3, "L_tab": 525, "n_element_tab": 65, "n_fig_1": 15, "n_tab_1": 2, "L_tab_1": 508, "n_element_tab_1": 46, "formula_len_all": 467, "formula_len_all_1": 568, "len_all": 176240, "len_all_1": 70440, "len_abs": 1444, "len_title": 112, "len_sents": 46742, "len_sents_1": 33914, "n_sents": 350, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1481, "title": "Meshtron: High-Fidelity, Artist-Like 3D Mesh Generation at Scale", "abs": "Meshes are a fundamental representation of 3D surfaces. However, creating high-quality meshes is a labor-intensive task that requires significant time and expertise in 3D modelling. While a delicate object often requires over $10^4$ faces to be accurately modeled, recent attempts at generating artist-like meshes are limited to $1.6$K faces and heavy discretization of vertex coordinates. Hence, scaling both the maximum face count and vertex coordinate resolution is crucial to producing high-quality meshes of realistic, complex 3D objects. We present Meshtron, a novel autoregressive mesh generation model able to generate meshes with up to 64K faces at 1024-level coordinate resolution --over an order of magnitude higher face count and $8{\\times}$ higher coordinate resolution than current state-of-the-art methods. Meshtron's scalability is driven by four key components: \n(i) an hourglass neural architecture, \n(ii) truncated sequence training, \n(iii) sliding window inference, \nand (iv) a robust sampling strategy that enforces the order of mesh sequences.\nThis results in over $50\\%$ less training memory, $2.5{\\times}$ faster throughput, and better consistency than existing works. Meshtron generates meshes of detailed, complex 3D objects at unprecedented levels of resolution and fidelity, closely resembling those created by professional artists, and opening the door to more realistic generation of detailed 3D assets for animation, gaming, and virtual environments.", "keywords": [ "Mesh generation", "3D Generation", "Hourglass Transformer", "Autoregressive mesh generation" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "mhyl7HhNM5", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Idan Szpektor", "gender": "unknown", "institution": "Google", "country": "IL", "position": "Researcher" }, { "name": "Nitay Calderon", "gender": "Male", "institution": "Technion - Israel Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Omer Nahum", "gender": "unknown", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "PhD student" }, { "name": "Orgad Keller", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Roi Reichart", "gender": "Male", "institution": "Technion, Israel Institute of Technology", "country": "IL", "position": "Associate Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 62, "n_ref": 97, "n_ref_all": 121, "n_fig": 8, "n_tab": 5, "L_tab": 1262, "n_element_tab": 194, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 1261, "n_element_tab_1": 192, "formula_len_all": 103, "formula_len_all_1": 103, "len_all": 237239, "len_all_1": 75431, "len_abs": 1554, "len_title": 150, "len_sents": 55016, "len_sents_1": 35578, "n_sents": 430, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1507, "title": "Are LLMs Better than Reported? Detecting Label Errors and Mitigating Their Effect on Model\u00a0Performance", "abs": "NLP benchmarks rely on standardized datasets for training and evaluating models and are crucial for advancing the field. Traditionally, expert annotations ensure high-quality labels; however, the cost of expert annotation does not scale well with the growing demand for larger datasets required by modern models.\nWhile crowd-sourcing provides a more scalable solution, it often comes at the expense of annotation precision and consistency. Recent advancements in large language models (LLMs) offer new opportunities to enhance the annotation process, particularly for detecting label errors in existing datasets. In this work, we consider the recent approach of LLM-as-a-judge, leveraging an ensemble of LLMs to flag potentially mislabeled examples.\nThrough a case study of four datasets from the TRUE benchmark, covering different tasks and domains, we empirically analyze the labeling quality of existing datasets, and compare expert, crowd-sourced, and our LLM-based annotations in terms of agreement, label quality, and efficiency, demonstrating the strengths and limitations of each annotation method. Our findings reveal a substantial number of label errors, which, when corrected, induce a significant upward shift in reported model performance. This suggests that many of the LLMs so-called mistakes are due to label errors rather than genuine model failures. Additionally, we discuss the implications of mislabeled data and propose methods to mitigate them in training to improve model performance.", "keywords": [ "LLMs", "label errors detection", "label errors handling", "data annotation" ], "rating_list": [ 8, 6, 5 ], "soundness_list": [ 4, 3, 1 ], "presentation_list": [ 4, 3, 4 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "mhJvgHRErR", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Jin Lyu", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Li Lin", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Liang An", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Pujin Cheng", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Tianyi Zhu", "gender": "unknown", "institution": "China Mobile Communications Company Limited Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Xiaoying Tang", "gender": "Female", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yebin Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yi Gu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 38, "n_ref": 111, "n_ref_all": 126, "n_fig": 7, "n_tab": 10, "L_tab": 2498, "n_element_tab": 577, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 627, "n_element_tab_1": 139, "formula_len_all": 313, "formula_len_all_1": 275, "len_all": 142728, "len_all_1": 57201, "len_abs": 1769, "len_title": 107, "len_sents": 36470, "len_sents_1": 27210, "n_sents": 307, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1661, "title": "AniMer: Animal Pose and Shape Estimation Using Transformer", "abs": "Quantitative analysis of animal behavior and biomechanics requires accurate animal pose and shape estimation across species, and is important for animal welfare and biological research. However, the small network capacity of previous methods and limited multi-species dataset leave this problem underexplored. To this end, this paper presents AniMer to estimate animal pose and shape using Transformer, enhancing the reconstruction accuracy of diverse quadrupedal species. AniMer aims to unify the understanding of various quadrupedal forms within a single framework, overcoming the limitations of traditional methods that focus on narrow specific species. A key feature of AniMer is its integration of a high-capacity Transformer-based backbone, which significantly boosts performance. To effectively train AniMer, we aggregate most available open-source quadrupedal datasets, either with 3D or 2D labels, and introduce CtrlAni3D, a novel large-scale synthetic dataset created through a diffusion-based image generation model, consisting of 9.7k pixel-aligned SMAL mesh-labeled images. This combination of a robust backbone and an expansive dataset enables AniMer to outperform existing methods on the multi-species Animal3D dataset and singlespecies dog benchmarks. Experiments on the unseen AnimalKingdom dataset further demonstrate the effectiveness of CtrlAni3D in enhancing generalization capabilities. Our study, through the development of AniMer and CtrlAni3D, underscores the significance of a large-capacity backbone and AI-driven synthetic data generation in advancing animal pose estimation research. Code and data will be released upon publication.", "keywords": [ "Animal Pose and Shape Estimation; Transformer; Synthetic Dataset" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "mhFToLPjM5", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Ivan Skorokhodov", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Jing Gu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Peter Wonka", "gender": "Male", "institution": "Snap Inc.", "country": "SA", "position": "Researcher" }, { "name": "Sergey Tulyakov", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Director of Research" }, { "name": "Xin Eric Wang", "gender": "Male", "institution": "Simular", "country": "US", "position": "Principal Researcher" }, { "name": "Xinya Du", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "Assistant Professor" }, { "name": "Yuwei Fang", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Senior Research Scientist" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 35, "n_ref": 72, "n_ref_all": 92, "n_fig": 12, "n_tab": 4, "L_tab": 3341, "n_element_tab": 104, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 768, "n_element_tab_1": 76, "formula_len_all": 636, "formula_len_all_1": 684, "len_all": 135752, "len_all_1": 62576, "len_abs": 0, "len_title": 177, "len_sents": 43646, "len_sents_1": 29773, "n_sents": 329, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1484, "title": "VIA: Unified Spatiotemporal Video Adaptation for Global and Local Video Editing", "abs": "Video editing is a cornerstone of digital media, from entertainment and education to professional communication. However, previous methods often overlook the necessity of comprehensively understanding both global and local contexts, leading to inaccurate and inconsistent edits in the spatiotemporal dimension, especially for long videos.\nIn this paper, we introduce VIA, a unified spatiotemporal Video Adaptation framework for global and local video editing, pushing the limits of consistently editing minute-long videos. First, to ensure local consistency within individual frames, we designed test-time editing adaptation to adapt a pre-trained image editing model for improving consistency between potential editing directions and the text instruction, and adapts masked latent variables for precise local control. Furthermore, to maintain global consistency over the video sequence, we introduce spatiotemporal adaptation that recursively gather consistent attention variables in key frames and strategically applies them across the whole sequence to realize the editing effects. Extensive experiments demonstrate that, compared to baseline methods, our VIA approach produces edits that are more faithful to the source videos, more coherent in the spatiotemporal context, and more precise in local control. More importantly, we show that VIA can achieve consistent long video editing in minutes, unlocking the potential for advanced video editing tasks over long video sequences.", "keywords": [ "Video Editing", "Generative AI" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "mhCNUP4Udw", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "James Kwok", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Hong Kong University of Science and Technology", "country": "HK", "position": "Professor" }, { "name": "Yanbin Wei", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 21, "n_formula_1": 16, "n_ref_uni": 36, "n_ref": 62, "n_ref_all": 97, "n_fig": 10, "n_tab": 13, "L_tab": 4615, "n_element_tab": 335, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 898, "n_element_tab_1": 40, "formula_len_all": 995, "formula_len_all_1": 732, "len_all": 147370, "len_all_1": 69527, "len_abs": 1448, "len_title": 89, "len_sents": 46808, "len_sents_1": 31506, "n_sents": 369, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1416, "title": "Graph Vision Networks for Link Prediction", "abs": "The potential of the vision modality for enhancing graph structural awareness has long been overlooked in the mainstream graph neural network (GNN) community. In this paper, we propose a simple yet effective framework called Graph Vision Networks (GVN), which first incorporates vision awareness into Message Passing Neural Network (MPNN) and achieves effective performance for link prediction, highlighting this unexplored but promising direction. Specifically, GVNs transform graph structures into images and extract Visual Structural Features (VSFs) from those images, where VSFs are considered a novel type of structural feature. Similar to previous structural features, VSFs also mitigate the limitations of traditional MPNNs in expressive power and substructure awareness. Additionally, unlike most previous heuristic-based structural features (e.g., common-neighbor-based and path-based ones), which typically depend on fixed structural priors, VSFs are adaptive and capable of capturing varying structural insights to better suit different scenarios. Extensive experiments across seven commonly used benchmark datasets demonstrate that GVNs and their variants can significantly enhance MPNNs in link prediction tasks. Additionally, the straightforward design of the framework makes it highly compatible with current methods, providing additional performance gains to achieve new state-of-the-art performance.", "keywords": [ "Graph Neural Networks\uff1bLink Prediction\uff1bMultimodality" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 5, 2, 4 ] }, { "paper_id": "mgjRSXGnF0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jiyul Ham", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "MS student" }, { "name": "Jun-Geol Baek", "gender": "unknown", "institution": "Korea University", "country": "KR", "position": "Full Professor" }, { "name": "Yonggon Jung", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 29, "n_ref": 60, "n_ref_all": 78, "n_fig": 8, "n_tab": 18, "L_tab": 8132, "n_element_tab": 1121, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 690, "n_element_tab_1": 21, "formula_len_all": 999, "formula_len_all_1": 926, "len_all": 161627, "len_all_1": 51501, "len_abs": 1418, "len_title": 137, "len_sents": 46360, "len_sents_1": 24036, "n_sents": 342, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1423, "title": "GlocalCLIP: Object-agnostic Global-Local Prompt Learning for Zero-shot Anomaly Detection", "abs": "Zero-shot anomaly detection (ZSAD) is crucial for detecting anomalous patterns in target datasets without using training samples, specifically in scenarios where there are distributional differences between the target domain and training data or where data scarcity arises because of restricted access. Although recently pretrained vision-language models demonstrate strong zero-shot performance across various visual tasks, they focus on learning class semantics, which makes their direct application to ZSAD challenging. To address this scenario, we propose GlocalCLIP, which uniquely separates global and local prompts and jointly optimizes them. This approach enables the object-agnostic glocal semantic prompt to effectively capture general normal and anomalous patterns without dependency on specific objects in the image. We refine the text prompts for more precise adjustments by utilizing deep-text prompt tuning in the text encoder. In the vision encoder, we apply V-V attention layers to capture detailed local image features. Finally, we introduce glocal contrastive learning to improve the complementary learning of global and local prompts, effectively detecting anomalous patterns across various domains. The generalization performance of GlocalCLIP in ZSAD was demonstrated on 15 real-world datasets from both the industrial and medical domains, achieving superior performance compared to existing methods.", "keywords": [ "Zero-shot anomaly detection", "Prompt learning", "Object-agnostic glocal semantic prompt design", "Deep-text prompt tuning", "Glocal contrastive learning" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "mgHv1XXoGm", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Yue Li", "gender": "Female", "institution": "China University of Mining Technology - Xuzhou", "country": "CN", "position": "MS student" }, { "name": "Zhizheng Liang", "gender": "Male", "institution": "China University of Mining and Technology ", "country": "", "position": "Associate Professor" } ], "n_formula": 99, "n_formula_1": 40, "n_ref_uni": 45, "n_ref": 63, "n_ref_all": 122, "n_fig": 7, "n_tab": 4, "L_tab": 1899, "n_element_tab": 132, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 763, "n_element_tab_1": 72, "formula_len_all": 10398, "formula_len_all_1": 3271, "len_all": 170258, "len_all_1": 72852, "len_abs": 1187, "len_title": 136, "len_sents": 49433, "len_sents_1": 30999, "n_sents": 492, "n_sents_1": 279, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1196, "title": "Regularized Conditional Optimal Transport for Feature Learning and Generalization Bounds", "abs": "This paper develops the regularized conditional optimal transport for feature learning in an embedding space. Instead of using joint distributions of data, we introduce conditional distributions to some reference conditional distributions in terms of the Kullback-Leibler (KL) divergence. Using conditional distributions provides the flexibility in controlling the transferring range of given data points. When the alternating optimization technique is employed to solve our model, it is interesting to find that conditional and marginal distributions have closed-form solutions. Moreover, the use of conditional distributions facilitates the derivation of the generalization bound of our model via the Rademacher complexity, which characterizes its convergence speed in terms of the number of samples. By optimizing the anchors (centroids) defined in the model, we also employ optimal transport and autoencoders to explore an embedding space of samples in the clustering problem. In the experimental part, we demonstrate that the proposed model achieves promising performance on some learning tasks. Moreover, we construct a conditional Wasserstein classifier to classify set-valued objects.", "keywords": [ "conditional optimal transport", "Kullback-Leibler divergence", "Rademacher complexity", "anchors" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "mfc6FKgtQA", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Jianyuan Zhong", "gender": "unknown", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Qiang Xu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Xiangyu Wen", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yuxuan Bian", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zhijian Xu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 23, "n_ref": 49, "n_ref_all": 73, "n_fig": 11, "n_tab": 10, "L_tab": 3704, "n_element_tab": 515, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1020, "n_element_tab_1": 165, "formula_len_all": 318, "formula_len_all_1": 66, "len_all": 222780, "len_all_1": 70577, "len_abs": 1993, "len_title": 125, "len_sents": 83703, "len_sents_1": 34073, "n_sents": 800, "n_sents_1": 241, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 983, "title": "Beyond Trend and Periodicity: Guide Time Series Forecasting with Textual Cues", "abs": "This work introduces a novel Text-Guided Time Series Forecasting (TGTSF) task. By integrating textual cues, such as channel descriptions and dynamic news, TGTSF addresses the critical limitations of traditional methods that rely purely on historical data. To support this task, we propose TGForecaster, a robust baseline model that fuses textual cues and time series data using cross-attention mechanisms. We then present four meticulously curated benchmark datasets to validate the proposed framework, ranging from simple periodic data to complex, event-driven fluctuations. Our comprehensive evaluations demonstrate that TGForecaster consistently achieves state-of-the-art performance, highlighting the transformative potential of incorporating textual information into time series forecasting. This work not only pioneers a novel forecasting task but also establishes a new benchmark for future research, driving advancements in multimodal data integration for time series models.", "keywords": [ "Time Series Forecasting", "Multi-modal Model" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "mfTM4UdYnC", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Marco Viviani", "gender": "unknown", "institution": "University of Milan - Bicocca", "country": "IT", "position": "Associate Professor" }, { "name": "Luca Celotti", "gender": "Male", "institution": "Universit\u00e9 de Sherbrooke", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 54, "n_ref_all": 60, "n_fig": 0, "n_tab": 1, "L_tab": 324, "n_element_tab": 54, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 324, "n_element_tab_1": 54, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 112498, "len_all_1": 75235, "len_abs": 1469, "len_title": 74, "len_sents": 31447, "len_sents_1": 30676, "n_sents": 228, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1474, "title": "LogicJitter: Let LLMs play Logic Games and they will Detect Misinformation", "abs": "In the face of the growing challenge of information overload online, the ability to accurately distinguish between genuine information and misinformation has become increasingly critical both from an individual and from a societal point of view. Methodologies for misinformation detection predominantly rely on supervised approaches, which depend heavily on large labeled datasets. However, these datasets are not only costly and time-consuming to produce, but they are also susceptible to issues such as labeling bias, time leakage, the inherent subjectivity of the task, and domain-specific limitations.\nIn this paper, we aim to overcome the aforementioned challenges by proposing a novel and cost-effective strategy to enhance the logical reasoning capabilities of Large Language Models (LLMs), thereby improving their ability to detect misinformation. Our approach, termed LogicJitter, employs a data augmentation technique during fine-tuning that generates both correct and incorrect statements within rule-based logic games. These games are designed to counteract well-known human cognitive biases and logical fallacies.\nHence, the primary contributions of this work include demonstrating the effectiveness of logical reasoning fine-tuning on LLMs and providing an open source package for the automatic generation of correct and incorrect logic-based training data, to ease reproducibility. Experimental results confirm this approach improves misinformation detection.", "keywords": [ "llm", "misinformation", "rule based AI", "toxicity" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 2, 2, 1, 1 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 1, 3, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "medKq3cONT", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 12, "author_info_list": [ { "name": "Bin Chen", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Jinpeng Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Junwei Pan", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Ming Yue", "gender": "Male", "institution": "Tencent", "country": "", "position": "Researcher" }, { "name": "Ruisheng Luo", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Shu-Tao Xia", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Tao Dai", "gender": "Male", "institution": "Department of Software Engineering, Shenzhen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Taolin Zhang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "XiaoxiangDeng", "gender": "Male", "institution": "AI Recommend", "country": "CN", "position": "Researcher" }, { "name": "Yaohua Zha", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yuan Wang", "gender": "Male", "institution": "Tencent AMS", "country": "CN", "position": "Researcher" }, { "name": "jie jiang", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "VP" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 24, "n_ref": 45, "n_ref_all": 64, "n_fig": 13, "n_tab": 2, "L_tab": 1183, "n_element_tab": 128, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 1183, "n_element_tab_1": 128, "formula_len_all": 596, "formula_len_all_1": 596, "len_all": 85221, "len_all_1": 54639, "len_abs": 987, "len_title": 107, "len_sents": 24287, "len_sents_1": 24306, "n_sents": 178, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 992, "title": "Towards Scalable Semantic Representation for Recommendation", "abs": "With recent advances in large language models (LLMs), there has been emerging numbers of research in developing Semantic IDs based on LLMs to enhance the performance of recommendation systems. \nHowever, the dimension of these embeddings needs to match that of the ID embedding in recommendation, which is usually much smaller than the original length.\nSuch dimension compression results in inevitable losses in discriminability and dimension robustness of the LLM embeddings, which motivates us to scale up the semantic representation. \nIn this paper, we propose Mixture-of-Codes, which first constructs multiple independent codebooks for LLM representation in the indexing stage, and then utilizes the Semantic Representation along with a fusion module for the downstream recommendation stage. \nExtensive analysis and experiments demonstrate that our method achieves superior discriminability and dimension robustness scalability, leading to the best scale-up performance in recommendations.", "keywords": [ "LLM", "Recommendation System" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 4, 1, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "meY36sGyyv", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Diego Paez-Granados", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Principal Researcher" }, { "name": "Haixin Li", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "MS student" }, { "name": "Yanke Li", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 9, "n_ref": 11, "n_ref_all": 17, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 617, "formula_len_all_1": 617, "len_all": 83651, "len_all_1": 56589, "len_abs": 1951, "len_title": 154, "len_sents": 32137, "len_sents_1": 26478, "n_sents": 248, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 106, "L_abs": 1919, "title": "Graphical-TS: An Interactive AI Pipeline for Multivariate Time Series with Ground-truth Graphical Modeling", "abs": "We present \\texttt{Graphical-TS}, an interactive simulation framework for multivariate time series (MTS) incorporating spatiotemporal causal graphical models. The system offers extensive customizability, enabling users to define and modify causal dynamics with uncertainty in spatiotemporal relationships and functional mappings. \\texttt{Graphical-TS} integrates expert knowledge, supports MTS simulation, and allows for the input of real-world MTS data, facilitating a dynamic interplay between data-driven learning and domain expertise. The system iteratively enhances causal relationships and simulated data by simulating MTS data based on specified causal graphs, performing causal discovery from real or simulated MTS, and enabling the integration and refinement of expert knowledge with learned causality. This approach progressively improves the quality of causal models and the data they generate, supporting tasks such as time series forecasting, imputation, prediction, and robustness testing via scenario-driven distribution shifts. We compared state-of-the-art causal discovery methods on datasets generated by \\texttt{Graphical-TS}. The empirical results demonstrate the platform\u2019s consistent performance compared to existing methods while offering versatility under distinct scenarios. This enables users to explore datasets more thoroughly and drive improvements in causal discovery research. With an intuitive user interface that connects domain experts and algorithm developers, \\texttt{Graphical-TS} empowers users to manipulate causal relationships, embedding domain knowledge into machine learning workflows. Originally developed to study physiological dynamics in patients, the system has broad applicability across various fields, offering a versatile platform for generating MTS datasets with known dynamics, validating causal discovery algorithms, and advancing research in time series analysis.", "keywords": [ "time series", "causal discovery", "benchmarking", "interface" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 2, 1, 3 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "meRCKuUpmc", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Hao Dong", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jia Zeng", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Jiangmiao Pang", "gender": "Male", "institution": "Shanghai AI Laboratory ", "country": "CN", "position": "Research Scientist" }, { "name": "Ping Wang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Sizhe Yang", "gender": "Male", "institution": "Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yang Tian", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 36, "n_ref": 62, "n_ref_all": 79, "n_fig": 5, "n_tab": 1, "L_tab": 270, "n_element_tab": 27, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 979, "n_element_tab_1": 69, "formula_len_all": 176, "formula_len_all_1": 256, "len_all": 127771, "len_all_1": 60521, "len_abs": 475, "len_title": 0, "len_sents": 31631, "len_sents_1": 28232, "n_sents": 277, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 81, "L_abs": 1473, "title": "Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation", "abs": "Current efforts to learn scalable policies in robotic manipulation primarily fall into two categories: one focuses on \"action,\" which involves behavior cloning from extensive collections of robotic data, while the other emphasizes \"vision,\" enhancing model generalization by pre-training representations or generative models, also referred to as world models, using large-scale visual datasets. This paper presents an end-to-end paradigm that predicts actions using inverse dynamics models conditioned on the robot's forecasted visual states, named Predictive Inverse Dynamics Models (PIDM). By closing the loop between vision and action, the end-to-end PIDM can be a better scalable action learner. In practice, we use Transformers to process both visual states and actions, naming the model Seer. It is initially pre-trained on large-scale robotic datasets, such as DROID, and can be adapted to real-world scenarios with a little fine-tuning data. Thanks to large-scale, end-to-end training and the continuous synergy between vision and action at each execution step, Seer significantly outperforms state-of-the-art methods across both simulation and real-world experiments. It achieves improvements of 13% on the LIBERO-LONG benchmark, 22% on CALVIN ABC-D, and 43% in real-world tasks. Notably, it demonstrates superior generalization for novel objects, lighting conditions, and environments under high-intensity disturbances. Code and models will be publicly available.", "keywords": [ "Robotic Manipulation ; Pre-training ; Visual Foresight ; Inverse Dynamics ; Large-scale robot dataset" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "meOELl7HRf", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Fujun Luan", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Hao Tan", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" }, { "name": "Kai Zhang", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Sai Bi", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Yicong Hong", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Zexiang Xu", "gender": "Male", "institution": "Hillbot", "country": "US", "position": "Researcher" }, { "name": "Ziwen Chen", "gender": "Female", "institution": "Oregon State University", "country": "US", "position": "PhD student" }, { "name": "Fuxin Li", "gender": "Male", "institution": "Oregon State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 58, "n_ref": 116, "n_ref_all": 126, "n_fig": 3, "n_tab": 6, "L_tab": 1961, "n_element_tab": 201, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1573, "n_element_tab_1": 187, "formula_len_all": 256, "formula_len_all_1": 256, "len_all": 153506, "len_all_1": 73068, "len_abs": 1176, "len_title": 133, "len_sents": 40976, "len_sents_1": 34498, "n_sents": 273, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 981, "title": "Long-LRM: Long-sequence Large Reconstruction Model for Wide-coverage Gaussian Splats", "abs": "We propose Long-LRM, a generalizable 3D Gaussian reconstruction model that is capable of reconstructing a large scene from \na long sequence of input images. Specifically, our model can process 32 source images at 960$\\times$540 resolution within only 1.3 seconds on a single A100 80G GPU. Our architecture features a mixture of the recent Mamba2 blocks and the classical transformer blocks which allowed many more tokens to be processed than prior work, enhanced by efficient token merging and Gaussian pruning steps that balance between quality and efficiency. Unlike previous generalizable 3D GS models that are limited to taking 1$\\sim$4 input images and can only reconstruct a small portion of a large scene, Long-LRM reconstructs the entire scene in a single feed-forward step. On large-scale scene datasets such as DL3DV-140 and Tanks and Temples, our method achieves performance comparable to optimization-based approaches while being two orders of magnitude more efficient.", "keywords": [ "3D Reconstruction" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "meKEKDhdnx", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 5, "author_info_list": [ { "name": "Dongming Li", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "MS student" }, { "name": "Reyna Abhyankar", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Vikranth Srivatsa", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Yiying Zhang", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Zijian He", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 26, "n_ref": 60, "n_ref_all": 87, "n_fig": 12, "n_tab": 1, "L_tab": 195, "n_element_tab": 30, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1314, "n_element_tab_1": 50, "formula_len_all": 108, "formula_len_all_1": 177, "len_all": 160142, "len_all_1": 62064, "len_abs": 1030, "len_title": 111, "len_sents": 49086, "len_sents_1": 31896, "n_sents": 380, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1041, "title": "Preble: Efficient Distributed Prompt Scheduling for LLM Serving", "abs": "Prompts to large language models (LLMs) have evolved beyond simple user questions.\nFor LLMs to solve complex problems, today\u2019s practices are to include domain-specific\ninstructions, illustration of tool usages, and/or long context such as textbook chapters in\nprompts. As such, many parts of prompts are repetitive across requests. Recent works\npropose to cache and reuse KV state of prompts. However, they are all confined to a single-\nGPU optimization, while production LLM serving systems are distributed by nature.\n\nThis paper proposes Preble, the first distributed LLM serving platform that targets and op-\ntimizes for prompt sharing. We designed a distributed scheduling system that co-optimizes\nKV state reuse and computation load-balancing with a new scheduling algorithm and a\nhierarchical scheduling mechanism. Our evaluation of Preble with real workloads and re-\nquest arrival patterns on two open-source LLMs shows that Preble outperforms the SOTA\nserving systems by 1.5\u00d7 to 14.5\u00d7 on average latency and 2\u00d7 to 10\u00d7 on p99 latency.", "keywords": [ "LLM prefix caching", "LLM serving", "Distributed systems for ML" ], "rating_list": [ 8, 8, 6, 6, 3 ], "soundness_list": [ 3, 4, 3, 3, 1 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 4, 3 ] }, { "paper_id": "md9qolJwLl", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Bo Xiong", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Steffen Staab", "gender": "Male", "institution": "University of Stuttgart", "country": "DE", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 19, "n_ref_uni": 29, "n_ref": 42, "n_ref_all": 59, "n_fig": 6, "n_tab": 3, "L_tab": 808, "n_element_tab": 40, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 100, "n_element_tab_1": 7, "formula_len_all": 1362, "formula_len_all_1": 1505, "len_all": 107634, "len_all_1": 67347, "len_abs": 1105, "len_title": 119, "len_sents": 37049, "len_sents_1": 30723, "n_sents": 261, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1109, "title": "From Tokens to Lattices: Emergent Lattice Structures in Language Models", "abs": "Pretrained masked language models (MLMs) have demonstrated an impressive capability to comprehend and encode conceptual knowledge, revealing a lattice structure among concepts. This raises a critical question: how does this conceptualization emerge from MLM pretraining? In this paper, we explore this problem from the perspective of Formal Concept Analysis (FCA), a mathematical framework that derives concept lattices from the observations of object-attribute relationships. We show that the MLM's objective implicitly learns a formal context that describes objects, attributes, and their dependencies, which enables the reconstruction of a concept lattice through FCA. We propose a novel framework for concept lattice construction from pretrained MLMs and investigate the origin of the inductive biases of MLMs in lattice structure learning. Our framework differs from previous work because it does not rely on human-defined concepts and allows for discovering \"latent\" concepts that extend beyond human definitions. We create three datasets for evaluation, and the empirical results verify our hypothesis.", "keywords": [ "Masked Language Models", "Formal Concept Analysis", "Interpretability" ], "rating_list": [ 6, 8, 6, 6, 1 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 3, 2, 3, 4, 4 ] }, { "paper_id": "mclaeTduHp", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "S Kevin Zhou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xin Luo", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zihang Jiang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 5, "n_ref_uni": 31, "n_ref": 47, "n_ref_all": 75, "n_fig": 9, "n_tab": 7, "L_tab": 3354, "n_element_tab": 156, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 2612, "n_element_tab_1": 131, "formula_len_all": 1092, "formula_len_all_1": 540, "len_all": 116620, "len_all_1": 65635, "len_abs": 1081, "len_title": 109, "len_sents": 34053, "len_sents_1": 28324, "n_sents": 254, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1084, "title": "LAMP: Large Model Pruning with Inter-Block Error Compensation", "abs": "The increasing prevalence of large-scale models, both in vision and language domains, presents significant challenges in terms of memory and resource consumption. While model pruning is an effective method for compressing models to alleviate these constraints, existing techniques either require extensive fine-tuning, which is resource-intensive, or perform well only at low sparsity levels (10%-50%), failing at high sparsity levels (50%-90%). To address these issues, this paper introduces LAMP to mitigate the drawbacks associated with traditional pruning methods, namely high resource consumption in methods that require extensive fine-tuning, and poor performance at high sparsity levels in methods that do not. It reduces memory overhead and alleviates performance degradation at high sparsity. Experimental results demonstrate that LAMP achieves slightly better performance than SparseGPT at low sparsity levels and significantly better at high sparsity levels in both language and vision models, without significantly increasing memory consumption when compared to SparseGPT.", "keywords": [ "Large Model Pruning", "Model Compression", "Compensation" ], "rating_list": [ 5, 3, 5, 1 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "mcDAY9PoDJ", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Haoran Xu", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" }, { "name": "Peng Cheng", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Wei Sun", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xianyuan Zhan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Youfang Lin", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Yunxin Liu", "gender": "unknown", "institution": "Institute for AI Industry Research, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Zhihao Wu", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Li Jianxiong", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 33, "n_ref": 106, "n_ref_all": 137, "n_fig": 5, "n_tab": 7, "L_tab": 3156, "n_element_tab": 366, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 382, "n_element_tab_1": 33, "formula_len_all": 1358, "formula_len_all_1": 1259, "len_all": 136157, "len_all_1": 68883, "len_abs": 1774, "len_title": 116, "len_sents": 46974, "len_sents_1": 33972, "n_sents": 288, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1783, "title": "Pushing the Limit of Small-Efficient Offline Reinforcement Learning", "abs": "Offline reinforcement learning (RL) has achieved notable progress in recent years. It enables learning optimized policy from fixed offline datasets and, therefore is particularly suitable for decision-making tasks that lack reliable simulators or have environment interaction restrictions. However, existing offline RL methods typically need a large amount of training data to achieve reasonable performance, and offer limited generalizability in out-of-distribution (OOD) regions due to conservative data-related regularizations. This seriously hinders the usability of offline RL in solving many real-world applications, where the available data are often limited. \nIn this study, we introduce a highly sample-efficient offline RL algorithm that learns optimized policy by enabling state-stitching in a compact latent space regulated by the fundamental symmetry in dynamical systems. Specifically, we introduce a time-reversal symmetry (T-symmetry) enforced inverse dynamics model (TS-IDM) to derive well-regulated latent state representations that greatly ease the difficulty of OOD generalization. Within the learned latent space, we can learn a guide-policy to output the latent next state that maximizes the reward, bypassing the conservative action-level behavior constraints as used in typical offline RL algorithms. The final optimized action can then be easily extracted by using the guide-policy's output as the goal state in the learned TS-IDM.\nWe call our method Offline RL via T-symmetry Enforced Latent State-Stitching (TELS).\nOur approach achieves amazing sample efficiency and OOD generalizability, significantly outperforming existing offline RL methods in a wide range of challenging small-sample tasks, even using as few as 1\\% of the original data in D4RL tasks.", "keywords": [ "offline reinforcement learning", "sample efficiency" ], "rating_list": [ 6, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "mc97L2QVIa", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Baoxiang Wang", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Dan Qiao", "gender": "unknown", "institution": "Chinese University of Hong Kong, Shen Zhen", "country": "CN", "position": "PhD student" }, { "name": "Hongyuan Zha", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Shanchao Yang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "PhD student" }, { "name": "Wenhao Li", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 36, "n_formula_1": 25, "n_ref_uni": 38, "n_ref": 71, "n_ref_all": 86, "n_fig": 4, "n_tab": 4, "L_tab": 1075, "n_element_tab": 114, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 876, "n_element_tab_1": 84, "formula_len_all": 4531, "formula_len_all_1": 1894, "len_all": 147695, "len_all_1": 68640, "len_abs": 1047, "len_title": 126, "len_sents": 45074, "len_sents_1": 31631, "n_sents": 344, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1057, "title": "Offline Multi-agent Reinforcement Learning with Sequential Score Decomposition", "abs": "Offline multi-agent reinforcement learning (MARL) faces significant challenges due to distribution shift issues, exacerbated by the high dimensionality of joint actions and complex joint behavior policy distributions. \nWhile existing methods often focus on independent learning or offline value decomposition with conservative value estimation, they may still lead to out-of-distribution (OOD) joint actions and reduced performance. \nThis is primarily due to the lack of exploration opportunity and implicit policy dependencies in offline settings. \nTo address these challenges, we propose an offline policy decomposition method incorporating joint policy regularization constraints. \nOur approach utilizes a diffusion generative model to capture the joint behavior policy, followed by a decomposition of the extracted score function. \nThis decomposition is then used to regularize individual policies in a decentralized manner. \nExperimental results demonstrate that our method achieves SOTA on continuous control tasks in standard offline MARL benchmarks.", "keywords": [ "Multi-agent Reinforcement Learning", "Offline RL", "Diffusion Models" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 1, 2, 1 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 5, 4, 2 ] }, { "paper_id": "mbo4YnWCHd", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Jesper L\u00f8ve Hinrich", "gender": "Male", "institution": "Technical University of Denmark", "country": "", "position": "Postdoc" }, { "name": "Kazu Ghalamkari", "gender": "Male", "institution": "Technical University of Denmark", "country": "DK", "position": "Postdoc" }, { "name": "Morten M\u00f8rup", "gender": "Male", "institution": "Technical University of Denmark", "country": "", "position": "Professor" } ], "n_formula": 106, "n_formula_1": 34, "n_ref_uni": 58, "n_ref": 96, "n_ref_all": 157, "n_fig": 10, "n_tab": 11, "L_tab": 3232, "n_element_tab": 324, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 741, "n_element_tab_1": 60, "formula_len_all": 7115, "formula_len_all_1": 2608, "len_all": 228854, "len_all_1": 71241, "len_abs": 1151, "len_title": 116, "len_sents": 67049, "len_sents_1": 29639, "n_sents": 577, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 994, "title": "Non-negative Tensor Mixture Learning for Discrete Density Estimation", "abs": "We present an expectation-maximization (EM) based unified framework for non-negative tensor decomposition that optimizes the Kullback-Leibler divergence. To avoid iterations in each M-step and learning rate tuning, we establish a general relationship between low-rank decomposition and many-body approximation. Using this connection, we exploit that the closed-form solution of the many-body approximation can be used to update all parameters simultaneously in the M-step. Our framework offers not only a unified methodology for a variety of low-rank structures, including CP, Tucker, and Train decompositions but also their combinations forming mixtures of tensors. The weights of each low-rank tensor in the mixture can be learned from the data, which eliminates the need to carefully choose a single low-rank structure in advance. We empirically demonstrate that our framework provides superior generalization for discrete density estimation compared to conventional tensor-based approaches.", "keywords": [ "Tensor factorization", "EM algorithm", "Low-rank approximation" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "mb9oOA3rD9", "primary_area": "causal reasoning", "n_author": 6, "author_info_list": [ { "name": "Lizhen Qu", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Lecturer" }, { "name": "Niket Tandon", "gender": "Male", "institution": "Research, Microsoft", "country": "IN", "position": "Principal Researcher" }, { "name": "Reza Haffari", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" }, { "name": "Tao Feng", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Xiaoxi Kang", "gender": "Female", "institution": "Monash University, Malaysia Campus", "country": "", "position": "PhD student" }, { "name": "Zhuang Li", "gender": "Male", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "Lecturer" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 34, "n_ref": 74, "n_ref_all": 104, "n_fig": 8, "n_tab": 5, "L_tab": 824, "n_element_tab": 85, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 775, "n_element_tab_1": 80, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 206795, "len_all_1": 49981, "len_abs": 2928, "len_title": 130, "len_sents": 52086, "len_sents_1": 23606, "n_sents": 414, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1210, "title": "From Corpora to Causality: Unveiling Causal Comprehension in Large Language Models", "abs": "This study investigates the efficacy of Large Language Models (LLMs) in causal discovery. Using newly available open-source LLMs, OLMo and BLOOM, which provide access to their pre-training corpora, we explore three research questions aimed at understanding how LLMs process causal discovery. These questions focus on the impact of memorization versus generalization, the influence of incorrect causal relations in pre-training data, and the role of contexts of causal relations. Our findings indicate that while LLMs are effective in recognizing causal relations that occur frequently in pre-training data, their ability to generalize to new or rare causal relations is limited. Moreover, the presence of incorrect causal relations significantly undermines the confidence of LLMs in corresponding correct causal relations, and the context of a causal relation markedly affects the performance of LLMs to identify causal relations. This study shows that LLMs possess a limited capacity to generalize novel causal relations. It also highlights the importance of managing incorrect causal relations in pre-training data and integrating contextual information to optimize LLM performance in causal discovery tasks.", "keywords": [ "language model", "causality", "pre-training data" ], "rating_list": [ 5, 3, 3, 5, 8 ], "soundness_list": [ 3, 2, 1, 2, 4 ], "presentation_list": [ 2, 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 1, 2, 3 ], "confidence_list": [ 4, 4, 3, 3, 5 ] }, { "paper_id": "mb2ryuZ3wz", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Antonio Torralba", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Phillip Isola", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Shivam Duggal", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Bill t. Freeman", "gender": "Male", "institution": "Google", "country": "US", "position": "Director" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 29, "n_ref": 78, "n_ref_all": 134, "n_fig": 18, "n_tab": 3, "L_tab": 501, "n_element_tab": 85, "n_fig_1": 15, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 291, "formula_len_all_1": 291, "len_all": 133510, "len_all_1": 62599, "len_abs": 1127, "len_title": 107, "len_sents": 51551, "len_sents_1": 32949, "n_sents": 321, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 34, "L_abs": 1133, "title": "How many tokens is an image worth?", "abs": "Current vision systems typically assign fixed-length representations to images, regardless of the information content. This contrasts with human intelligence \u2014and even large language models\u2014which allocate varying representational capacities based on entropy, context and familiarity. Inspired by this, we propose an approach to learn variable-length token representations for 2D images. Our encoder-decoder architecture recursively processes 2D image tokens, distilling them into 1D latent tokens over multiple iterations of recurrent rollouts. Each iteration refines the 2D tokens, updates the existing 1D latent tokens, and adaptively increases representational capacity by adding new tokens. This enables compression of images into a variable number of tokens, ranging from 32 to 256. We validate our tokenizer using reconstruction loss and FID metrics, demonstrating that token count aligns with image entropy, familiarity and downstream task requirements. Recurrent token processing with increasing representational capacity in each iteration shows signs of token specialization, revealing potential for object / part discovery.", "keywords": [ "Representation Learning", "Image Tokenization", "Compression" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "mb2rHLcKN5", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 6, "author_info_list": [ { "name": "Changran Hu", "gender": "unknown", "institution": "Sambanova Systems, Inc", "country": "", "position": "Researcher" }, { "name": "Haige Bo", "gender": "Female", "institution": "Sambanova Systems", "country": "US", "position": "Researcher" }, { "name": "Lin Zheng", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Lingpeng Kong", "gender": "Male", "institution": "Department of Computer Science, The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Urmish Thakker", "gender": "Male", "institution": "SambaNova Systems", "country": "", "position": "Principal Engineer" }, { "name": "Xueliang Zhao", "gender": "Male", "institution": "The University of Hong Kong", "country": "", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 20, "n_ref": 58, "n_ref_all": 79, "n_fig": 10, "n_tab": 3, "L_tab": 1565, "n_element_tab": 223, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 1868, "n_element_tab_1": 21, "formula_len_all": 1198, "formula_len_all_1": 1302, "len_all": 120140, "len_all_1": 62044, "len_abs": 1383, "len_title": 60, "len_sents": 37083, "len_sents_1": 25248, "n_sents": 284, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1390, "title": "SubgoalXL: Subgoal-based Expert Learning for Theorem Proving", "abs": "Formal theorem proving, a field at the intersection of mathematics and computer science, has seen renewed interest with advancements in large language models (LLMs). This paper introduces SubgoalXL, a novel approach that synergizes subgoal-based proofs with expert learning to enhance LLMs' capabilities in formal theorem proving within the Isabelle environment. SubgoalXL addresses two critical challenges: the scarcity of specialized mathematics and theorem-proving data, and the need for improved multi-step reasoning abilities in LLMs. By optimizing data efficiency and employing subgoal-level supervision, SubgoalXL extracts richer information from limited human-generated proofs. The framework integrates subgoal-oriented proof strategies with an expert learning system, iteratively refining formal statement, proof, and subgoal generators. Leveraging the Isabelle environment's advantages in subgoal-based proofs, SubgoalXL achieves a new state-of-the-art performance of 56.1\\% in Isabelle on the standard miniF2F dataset, marking an absolute improvement of 4.9\\%. Notably, SubgoalXL successfully solves 41 AMC12, 9 AIME, and 3 IMO problems from miniF2F. These results underscore the effectiveness of maximizing limited data utility and employing targeted guidance for complex reasoning in formal theorem proving, contributing to the ongoing advancement of AI reasoning capabilities.", "keywords": [ "theorem proving", "subgoal-based proofs", "expert learning" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 1, 1, 3, 2 ], "contribution_list": [ 1, 2, 4, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "maoBEh5rU7", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Jingliang Duan", "gender": "Male", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "Associate Professor" }, { "name": "Liangfa Chen", "gender": "Male", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "MS student" }, { "name": "Shengbo Eben Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Tong Liu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Xujie Song", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yinuo Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Jack Qin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "wenxuan Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" } ], "n_formula": 40, "n_formula_1": 17, "n_ref_uni": 36, "n_ref": 66, "n_ref_all": 128, "n_fig": 24, "n_tab": 18, "L_tab": 6172, "n_element_tab": 541, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 432, "n_element_tab_1": 32, "formula_len_all": 3428, "formula_len_all_1": 530, "len_all": 166619, "len_all_1": 49779, "len_abs": 1351, "len_title": 123, "len_sents": 46554, "len_sents_1": 21527, "n_sents": 436, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1401, "title": "FlipNet: Fourier Lipschitz Smooth Policy Network for Reinforcement Learning", "abs": "Deep reinforcement learning (RL) is an effective method for decision-making and control tasks. However, RL-trained policies encounter the action fluctuation problem, where consecutive actions significantly differ despite minor variations in adjacent states. This problem results in actuators' wear, safety risk, and performance reduction in real-world applications. To address the problem, we identify the two fundamental reasons causing action fluctuation, i.e. policy non-smoothness and observation noise, then propose the Fourier Lipschitz Smooth Policy Network (FlipNet). FlipNet adopts two innovative techniques to tackle the two reasons in a decoupled manner. Firstly, we prove the Jacobian norm is an approximation of Lipschitz constant and introduce a Jacobian regularization technique to enhance the smoothness of policy network. Secondly, we introduce a Fourier filter layer to deal with observation noise. The filter layer includes a trainable filter matrix that can automatically extract important observation frequencies and suppress noise frequencies. FlipNet can be seamlessly integrated into most existing RL algorithms as an actor network. Simulated tasks on DMControl and a real-world experiment on vehicle-robot driving show that FlipeNet has excellent action smoothness and noise robustness, achieving a new state-of-the-art performance. The code and videos are publicly available.", "keywords": [ "Reinforcement Learning", "Neural Network", "Action Fluctuation", "Control Smoothness and Robustness", "Real-world Application", "Lipschitz Continuity", "Fourier Transform" ], "rating_list": [ 6, 1, 6, 3 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 3, 1, 3, 4 ], "contribution_list": [ 3, 1, 3, 4 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "ma4SUzeCLR", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "An-Zi Yen", "gender": "Female", "institution": "Department of Computer Science, National Yang Ming Chiao Tung University", "country": "TW", "position": "Assistant Professor" }, { "name": "Mao-Siang Chen", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Undergrad student" }, { "name": "Yu-Hang Deng", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 1, "n_ref_uni": 31, "n_ref": 40, "n_ref_all": 71, "n_fig": 5, "n_tab": 13, "L_tab": 4110, "n_element_tab": 273, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2529, "n_element_tab_1": 138, "formula_len_all": 2398, "formula_len_all_1": 40, "len_all": 167070, "len_all_1": 72291, "len_abs": 926, "len_title": 115, "len_sents": 54301, "len_sents_1": 33457, "n_sents": 449, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 928, "title": "Detecting Problematic Questions to Support Math Word Problem Design", "abs": "When designing math word problems, teachers must ensure the clarity and precision of the question to avoid multiple interpretations and unanswerable situations, thereby maintaining consistent grading standards and effectiveness. We address these issues to provide comprehensive support to teachers in creating clear, solvable, and formal math word problems. In this paper, we present MathError, a dataset of real-world math word problems annotated with error types to investigate the need for question correction. Our work explores how large language models (LLMs) can assist teachers in detecting problematic questions to support math word problem design in scenarios with limited data, simulating real-world conditions with minimal training samples. Preliminary results demonstrate the models' capabilities in detecting problematic questions and identify areas for further research and development in educational applications.", "keywords": [ "Problematic Question Detection", "Question Design Support", "Self-Optimization Prompting" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "mZvzvwIu8f", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Guosong Jiang", "gender": "unknown", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Huitong Chen", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Qinghua Hu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Professor" }, { "name": "YanFan", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Yu Wang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 22, "n_ref": 41, "n_ref_all": 54, "n_fig": 4, "n_tab": 3, "L_tab": 1789, "n_element_tab": 129, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2651, "n_element_tab_1": 141, "formula_len_all": 519, "formula_len_all_1": 519, "len_all": 86591, "len_all_1": 55210, "len_abs": 1375, "len_title": 131, "len_sents": 27379, "len_sents_1": 23741, "n_sents": 220, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1381, "title": "Reducing class-wise confusion for incremental learning with disentangled manifolds", "abs": "Class incremental learning (CIL) aims to enable models to continuously learn new classes without catastrophically forgetting old ones. A promising direction is to learn and use prototypes of classes during incremental updates. Despite simplicity and intuition, we find that such methods suffer from inadequate representation capability and unsatisfied confusion caused by distribution drift. In this paper, we develop a Confusion-REduced AuTo-Encoder classifier (CREATE) for CIL. Specifically, our method employs a lightweight auto-encoder module to learn each compact class manifold in latent subspace, constraining samples well reconstructed only on the semantically correct auto-encoder. Thus, the representation stability and capability of class distributions are enhanced, alleviating the potential class-wise confusion problem. To further distinguish the drifted features, we propose a confusion-aware latent space separation loss that ensures exemplars are closely distributed in their corresponding low-dimensional manifold while keeping away from the distributions of drifted features from other classes. Our method demonstrates stronger representational capacity by learning disentangled manifolds and reduces class confusion caused by drift. Extensive experiments on multiple datasets and settings show that CREATE outperforms other state-of-the-art methods up to 5.41%.", "keywords": [ "Class incremental learning; Auto-encoder; Manifold" ], "rating_list": [ 5, 5, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "mZptYYttFj", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Bingchen Zhao", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Mude Hui", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "peng wang", "gender": "Male", "institution": "Bytedance US AILab", "country": "US", "position": "Research Scientist" }, { "name": "Siwei Yang", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Yichun Shi", "gender": "Male", "institution": "ByteDance", "country": "US", "position": "Researcher" }, { "name": "Yuyin Zhou", "gender": "unknown", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "cihang xie", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "heng Wang", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "research scientist" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 59, "n_ref_all": 81, "n_fig": 14, "n_tab": 5, "L_tab": 1080, "n_element_tab": 90, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 2405, "n_element_tab_1": 90, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 109146, "len_all_1": 51377, "len_abs": 836, "len_title": 116, "len_sents": 35665, "len_sents_1": 23090, "n_sents": 258, "n_sents_1": 152, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 67, "L_abs": 1060, "title": "HQ-Edit: A High-Quality Dataset for Instruction-based Image Editing", "abs": "This study introduces HQ-Edit, a high-quality instruction-based image editing dataset with around 200,000 edits. Unlike prior approaches relying on attribute guidance or human feedback on building datasets, we devise a scalable data collection pipeline leveraging advanced foundation models, namely GPT-4V and DALL-E 3. To ensure its high quality, diverse examples are first collected online, expanded, and then used to create high-quality diptychs featuring input and output images with detailed text prompts, followed by precise alignment ensured through post-processing. In addition, we propose two evaluation metrics, Alignment and Coherence, to quantitatively assess the quality of image edit pairs using GPT-4V. HQ-Edits high-resolution images, rich in detail and accompanied by comprehensive editing prompts, substantially enhance the capabilities of existing image editing models. For example, an HQ-Edit finetuned InstructPix2Pix can attain state-of-the-art image editing performance, even surpassing those models fine-tuned with human-annotated data.", "keywords": [ "Image-to-image translation", "Image Editing" ], "rating_list": [ 3, 5, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "mZfBRjMWq0", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Bac Nguyen", "gender": "Male", "institution": "Sony Europe B.V.", "country": "DE", "position": "Senior Engineer" }, { "name": "Chieh-Hsin Lai", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Researcher" }, { "name": "Naoki Murata", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Researcher" }, { "name": "Stefano Ermon", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Toshimitsu Uesaka", "gender": "Male", "institution": "Sony AI", "country": "", "position": "Researcher" }, { "name": "Yuhta Takida", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Research Scientist" }, { "name": "Yuki Mitsufuji", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Lead Research Scientist" } ], "n_formula": 50, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 104, "n_ref_all": 130, "n_fig": 9, "n_tab": 5, "L_tab": 1440, "n_element_tab": 136, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 494, "n_element_tab_1": 33, "formula_len_all": 4064, "formula_len_all_1": 747, "len_all": 196972, "len_all_1": 64883, "len_abs": 1178, "len_title": 122, "len_sents": 57873, "len_sents_1": 30266, "n_sents": 473, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1181, "title": "G2D2: Gradient-guided Discrete Diffusion for image inverse problem solving", "abs": "Recent literature has effectively utilized diffusion models trained on continuous variables as priors for solving inverse problems. Notably, discrete diffusion models with discrete latent codes have shown strong performance, particularly in modalities suited for discrete compressed representations, such as image and motion generation. However, their discrete and non-differentiable nature has limited their application to inverse problems formulated in continuous spaces. This paper presents a novel method for addressing linear inverse problems by leveraging image-generation models based on discrete diffusion as priors. We overcome these limitations by approximating the true posterior distribution with a variational distribution constructed from categorical distributions and continuous relaxation techniques. Furthermore, we employ a star-shaped noise process to mitigate the drawbacks of traditional discrete diffusion models with absorbing states, demonstrating that our method performs comparably to continuous diffusion techniques. To the best of our knowledge, this is the first approach to use discrete diffusion model-based priors for solving image inverse problems.", "keywords": [ "diffusion model", "discrete diffusion model", "inverse problems", "categorical data" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "mYgoNEsUDi", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Yulia Gel", "gender": "Female", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "Full Professor" }, { "name": "Yuzhou Chen", "gender": "Male", "institution": "University of California, Riverside", "country": "US", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 18, "n_ref_uni": 51, "n_ref": 82, "n_ref_all": 100, "n_fig": 2, "n_tab": 14, "L_tab": 3521, "n_element_tab": 259, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 457, "n_element_tab_1": 57, "formula_len_all": 2162, "formula_len_all_1": 1474, "len_all": 172345, "len_all_1": 60226, "len_abs": 1251, "len_title": 132, "len_sents": 44403, "len_sents_1": 28425, "n_sents": 286, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1294, "title": "Topological Zigzag Spaghetti for Diffusion-based Generation and Prediction on Graphs", "abs": "Diffusion models have recently emerged as a new powerful machinery for generative artificial intelligence on graphs, with applications ranging from drug design to knowledge discovery. However, despite their high potential, most, if not all, currently existing graph diffusion models are limited in their ability to holistically describe the intrinsic {\\it higher-order} topological graph properties, which obstructs model generalizability and adoption for downstream tasks. We propose to address this fundamental challenge and extract the latent salient topological graph descriptors at different resolutions by leveraging zigzag persistence. We develop a new computationally efficient topological summary, zigzag spaghetti (ZS), which delivers the most inherent topological properties {\\it simultaneously over a sequence of graphs at multiple resolutions}. We derive theoretical stability guarantees of ZS and present the first attempt to integrate\ndynamic topological information into graph diffusion models. Our extensive experiments on %9 benchmark datasets for \ngraph classification and prediction tasks suggest that ZS has a high promise not only to enhance performance of graph diffusion models, with gains up 10\\%, but also to substantially booster model robustness under uncertainties.", "keywords": [ "Graph learning", "Topological Data Analysis", "Geometric Deep Learning" ], "rating_list": [ 6, 5, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 4 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "mXh8LbXXpx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Andrea Bartezzaghi", "gender": "Male", "institution": "IBM Research", "country": "CH", "position": "Researcher" }, { "name": "Cristiano Malossi", "gender": "Male", "institution": "International Business Machines", "country": "CH", "position": "Principal Researcher" }, { "name": "Filip Janicki", "gender": "Male", "institution": "International Business Machines", "country": "CH", "position": "Researcher" }, { "name": "Konrad Schindler", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Professor" }, { "name": "Mattia Rigotti", "gender": "unknown", "institution": "International Business Machines", "country": "CH", "position": "Researcher" }, { "name": "Niccol\u00f2 Avogaro", "gender": "Male", "institution": "International Business Machines", "country": "CH", "position": "PhD student" }, { "name": "Roy Assaf", "gender": "Male", "institution": "International Business Machines", "country": "CH", "position": "Researcher" }, { "name": "Thomas Frick", "gender": "Male", "institution": "International Business Machines", "country": "CH", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 64, "n_ref_all": 79, "n_fig": 2, "n_tab": 6, "L_tab": 1214, "n_element_tab": 238, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1167, "n_element_tab_1": 234, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 168142, "len_all_1": 65263, "len_abs": 1269, "len_title": 132, "len_sents": 37957, "len_sents_1": 32302, "n_sents": 229, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1290, "title": "Show or Tell? Effectively prompting Vision-Language Models for semantic segmentation", "abs": "Large Vision-Language Models (VLMs) are increasingly being regarded as foundation models that can be instructed to solve diverse tasks by prompting, without task-specific training.\nWe examine the seemingly obvious question: \\emph{how to effectively prompt VLMs for semantic segmentation}.\nTo that end, we systematically evaluate the segmentation performance of several recent models guided by either text or visual prompts on the diverse MESS dataset collection.\nWe introduce a scalable prompting scheme, \\emph{few-shot prompted semantic segmentation}, inspired by open-vocabulary segmentation and few-shot learning.\nIt turns out that even the most advanced VLMs lag far behind specialist models trained for a specific segmentation task, by about 30\\% on average on the Intersection-over-Union metric.\nMoreover, we find that text prompts and visual prompts are complementary: each one of the two modes fails on many examples that the other one can solve.\nOur analysis suggests that being able to anticipate the most effective prompt modality can lead to a 11\\% improvement in performance.\nMotivated by our findings, we propose PromptMatcher, a remarkably simple baseline that combines both text and visual prompts, achieving state-of-the-art results for training-free semantic segmentation.", "keywords": [ "Foundation Models", "Prompting", "Semantic Segmentation", "Vision-Language Models", "VLM", "Training-free" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "mXZ98iNFw2", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Bjorn Hartmann", "gender": "Male", "institution": "University of California - Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Chin-Yi Cheng", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Peitong Duan", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Yang Li", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 67, "n_ref_all": 114, "n_fig": 14, "n_tab": 8, "L_tab": 2409, "n_element_tab": 127, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 778, "n_element_tab_1": 68, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 305848, "len_all_1": 60475, "len_abs": 1455, "len_title": 121, "len_sents": 124608, "len_sents_1": 29862, "n_sents": 1054, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1460, "title": "Visual Prompting with Iterative Refinement for Design Critique Generation", "abs": "Feedback is crucial for every design process, such as user interface (UI) design, and automating design critiques can significantly improve the efficiency of the design workflow. Although existing multimodal large language models (LLMs) excel in many tasks, they often struggle with generating high-quality design critiques---a complex task that requires producing detailed design comments that are visually grounded in a given design's image. Building on recent advancements in iterative refinement of text output and visual prompting methods, we propose an iterative visual prompting approach for UI critique that takes an input UI screenshot and design guidelines and generates a list of design comments, along with corresponding bounding boxes that map each comment to a specific region in the screenshot. The entire process is driven completely by LLMs, which iteratively refine both the text output and bounding boxes using few-shot samples tailored for each step. We evaluated our approach using Gemini-1.5-pro and GPT-4o, and found that human experts generally preferred the design critiques generated by our pipeline over those by the baseline, with the pipeline reducing the gap to human performance by 50\\% for one rating metric. To assess the generalizability of our approach to other multimodal tasks, we applied our pipeline to open-vocabulary object and attribute detection, and experiments showed that our method also outperformed the baseline.", "keywords": [ "User Interface Design Critique", "Multimodal LLM", "Visual Grounding", "Prompting Techniques" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "mXHTifc1Fn", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Alexandru Dumitrescu", "gender": "unknown", "institution": "Aalto University", "country": "", "position": "PhD student" }, { "name": "Dani Korpela", "gender": "unknown", "institution": "Aalto University", "country": "", "position": "PhD student" }, { "name": "Harri L\u00e4hdesm\u00e4ki", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Associate Professor" }, { "name": "Markus Heinonen", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Researcher" }, { "name": "Valerii Iakovlev", "gender": "unknown", "institution": "Aalto University", "country": "", "position": "PhD student" }, { "name": "Yogesh Verma", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "PhD student" }, { "name": "Vikas K Garg", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Assistant Professor" } ], "n_formula": 53, "n_formula_1": 20, "n_ref_uni": 38, "n_ref": 91, "n_ref_all": 144, "n_fig": 19, "n_tab": 5, "L_tab": 1290, "n_element_tab": 328, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 745, "n_element_tab_1": 222, "formula_len_all": 3279, "formula_len_all_1": 953, "len_all": 203054, "len_all_1": 61836, "len_abs": 882, "len_title": 128, "len_sents": 69845, "len_sents_1": 26886, "n_sents": 525, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 885, "title": "E(3)-equivariant models cannot learn chirality: Field-based molecular generation", "abs": "Obtaining the desired effect of drugs is highly dependent on their molecular geometries. Thus, the current prevailing paradigm focuses on 3D point-cloud atom representations, utilizing graph neural network (GNN) parametrizations, with rotational symmetries baked in via E(3) invariant layers. We prove that such models must necessarily disregard chirality, a geometric property of the molecules that cannot be superimposed on their mirror image by rotation and translation. Chirality plays a key role in determining drug safety and potency. To address this glaring issue, we introduce a novel field-based representation, proposing reference rotations that replace rotational symmetry constraints. The proposed model captures all molecular geometries including chirality, while still achieving highly competitive performance with E(3)-based methods across standard benchmarking metrics.", "keywords": [ "deep generative models", "molecule generation" ], "rating_list": [ 5, 8, 6, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "mVOz28mPHr", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Gordon Wetzstein", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Irmak Sivgin", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Mert Pilanci", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Sara Fridovich-Keil", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" } ], "n_formula": 67, "n_formula_1": 19, "n_ref_uni": 35, "n_ref": 80, "n_ref_all": 105, "n_fig": 6, "n_tab": 5, "L_tab": 107, "n_element_tab": 3, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 468, "n_element_tab_1": 20, "formula_len_all": 4589, "formula_len_all_1": 675, "len_all": 164645, "len_all_1": 60622, "len_abs": 4294, "len_title": 104, "len_sents": 51585, "len_sents_1": 28459, "n_sents": 380, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1323, "title": "Geometric Algebra Planes: Convex Implicit Neural Volumes", "abs": "Volume parameterizations abound in recent literature, from the classic voxel grid to the implicit neural representation and everything in between. While implicit representations have shown impressive capacity and better memory efficiency compared to voxel grids, to date they require training via nonconvex optimization. This nonconvex training process can be slow to converge and sensitive to initialization and hyperparameter choices. We introduce a family of models, GA-Planes, that is the first class of implicit neural volume representations that can be trained by convex optimization. GA-Planes models include any combination of features stored in tensor basis elements, followed by a neural feature decoder. They generalize many existing representations and can be adapted for convex, semiconvex, or nonconvex training as needed for different inverse problems. In the 2D setting, we prove that GA-Planes is equivalent to a low-rank plus low-resolution matrix factorization; we show that this approximation outperforms the classic low-rank plus sparse decomposition for fitting a natural image. In 3D, we demonstrate GA-Planes' competitive performance in terms of expressiveness, model size, and optimizability across three volume fitting tasks: radiance field reconstruction, 3D segmentation, and video segmentation.", "keywords": [ "implicit neural representation", "convex optimization", "volume", "nerf", "segmentation" ], "rating_list": [ 5, 8, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 4, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "mVExccNdtK", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Ali Aghdaei", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Chenhui Deng", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Wuxinlin Cheng", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zhiru Zhang", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Full Professor" }, { "name": "Zhuo Feng", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "", "position": "Associate Professor" } ], "n_formula": 56, "n_formula_1": 14, "n_ref_uni": 67, "n_ref": 111, "n_ref_all": 142, "n_fig": 10, "n_tab": 12, "L_tab": 2488, "n_element_tab": 212, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1276, "n_element_tab_1": 100, "formula_len_all": 3626, "formula_len_all_1": 1015, "len_all": 207004, "len_all_1": 63933, "len_abs": 1274, "len_title": 116, "len_sents": 62829, "len_sents_1": 28088, "n_sents": 483, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1279, "title": "SAGMAN: Stability Analysis of Graph Neural Networks on the Manifolds", "abs": "Modern graph neural networks (GNNs) can be sensitive to changes in the input graph structure and node features, potentially resulting in unpredictable behavior and degraded performance. In this work, we introduce a spectral framework known as SAGMAN for examining the stability of GNNs. This framework assesses the distance distortions that arise from the nonlinear mappings of GNNs between the input and output manifolds: when two nearby nodes on the input manifold are mapped (through a GNN model) to two distant ones on the output manifold, it implies a large distance distortion and thus a poor GNN stability. We propose a distance-preserving graph dimension reduction (GDR) approach that utilizes spectral graph embedding and probabilistic graphical models (PGMs) to create low-dimensional input/output graph-based manifolds for meaningful stability analysis. Our empirical evaluations show that SAGMAN effectively assesses the stability of each node when subjected to various edge or feature perturbations, offering a scalable approach for evaluating the stability of GNNs, extending to applications within recommendation systems. Furthermore, we illustrate its utility in downstream tasks, notably in enhancing GNN stability and facilitating adversarial targeted attacks.", "keywords": [ "Manifolds", "Stability", "Graph Neural Networks" ], "rating_list": [ 3, 3, 8, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "mVCcWCjeEz", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 10, "author_info_list": [ { "name": "Bowen Zhou", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Daixuan Cheng", "gender": "Female", "institution": "Beijing Insititute for General Artifical Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Ermo Hua", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hengli Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Kaiyan Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ning Ding", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Xingtai Lv", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xuekai Zhu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhouhan Lin", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Zilong Zheng", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" } ], "n_formula": 33, "n_formula_1": 13, "n_ref_uni": 30, "n_ref": 76, "n_ref_all": 120, "n_fig": 13, "n_tab": 15, "L_tab": 5275, "n_element_tab": 747, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 683, "n_element_tab_1": 145, "formula_len_all": 2457, "formula_len_all_1": 795, "len_all": 202003, "len_all_1": 59217, "len_abs": 1411, "len_title": 108, "len_sents": 53949, "len_sents_1": 25355, "n_sents": 483, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 942, "title": "ToEdit: How to Synthesize Text Data to Avoid Model Collapse?", "abs": "We explore model collapse caused by synthetic data, where AI models trained on such data experience a gradual decline in performance. \nOur initial analysis examines language model pretraining on mixed human and synthetic data, highlighting performance degradation. Further statistical analysis reveals distributional shifts and an over-concentration of n-gram features caused by synthetic data. Inspired by these insights, we propose token-level editing on human data, to obtain semi-synthetic data instead of fully using model outputs. As a proof of concept, we theoretically demonstrate that token-level editing can prevent model collapse, as the test error is constrained by a finite upper bound. We conducted extensive experiments on pretraining, continual pretraining, and supervised fine-tuning of language models. The results validate our theoretical proof that token-level editing improves data quality and enhances model performance.", "keywords": [ "synthetic data", "model collapse" ], "rating_list": [ 6, 8, 8, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "mV6cO4mGjH", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Eman Ehab Nasef", "gender": "Female", "institution": "Egypt University of Informatics", "country": "EG", "position": "Instructor" }, { "name": "Mai Gamal", "gender": "Female", "institution": "German University in Cairo", "country": "EG", "position": "PhD student" }, { "name": "Mohamed Rashad Abdel Hamid", "gender": "Male", "institution": "Ain Shams University ", "country": "EG", "position": "MS student" }, { "name": "Seif Eldawlatly", "gender": "Male", "institution": "American University in Cairo", "country": "EG", "position": "Associate Professor" }, { "name": "Mennatullah mohammed Siam", "gender": "Female", "institution": "Ontario Tech University", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 40, "n_ref": 102, "n_ref_all": 164, "n_fig": 40, "n_tab": 4, "L_tab": 1335, "n_element_tab": 74, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 945, "n_element_tab_1": 85, "formula_len_all": 111, "formula_len_all_1": 65, "len_all": 172544, "len_all_1": 66701, "len_abs": 1981, "len_title": 67, "len_sents": 58932, "len_sents_1": 33980, "n_sents": 395, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1990, "title": "Dynamics Based Neural Encoding with Inter-Intra Region Connectivity", "abs": "Extensive literature has drawn comparisons between recordings of biological neurons in the brain and deep neural networks. This comparative analysis aims to advance and interpret deep neural networks and enhance our understanding of biological neural systems. However, previous works did not consider the time aspect and how the encoding of video and dynamics in deep networks relate to the biological neural systems within a large-scale comparison. Towards this end, we propose the first large-scale study focused on comparing video understanding models with respect to the visual cortex recordings using video stimuli. The study encompasses more than two million regression fits, examining image vs. video understanding, convolutional vs. transformer-based and fully vs. self-supervised models. Our study resulted in both, insights to help better understand deep video understanding models and a novel neural encoding scheme to better encode biological neural systems. We provide key insights on how video understanding models predict visual cortex responses; showing video understanding better than image understanding models, convolutional models are better in the early-mid visual cortical regions than transformer based ones except for multiscale transformers and that two-stream models are better than single stream. Furthermore, we propose a novel neural encoding scheme that is built on top of the best performing video understanding models, while incorporating inter-intra region connectivity across the visual cortex. Our neural encoding leverages the encoded dynamics from video stimuli, through utilizing two-stream networks and multiscale transformers, while taking connectivity priors into consideration. Our results show that merging both intra and inter-region connectivity priors increases the encoding performance over each one of them standalone or no connectivity priors. It also shows the necessity for encoding dynamics to fully benefit from such connectivity priors.", "keywords": [ "Neuroscience", "Neural Encoding", "Video Understanding" ], "rating_list": [ 1, 5, 6, 6 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 1, 2, 4, 3 ], "confidence_list": [ 5, 5, 3, 3 ] }, { "paper_id": "mUbYof5MKp", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Kosuke Kawakami", "gender": "Male", "institution": "Tokyo Institute of Technology, Tokyo Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Masahiro Asami", "gender": "Male", "institution": "HAKUHODO TECHNOLOGIES INC.", "country": "JP", "position": "Researcher" }, { "name": "Rikiya Takehi", "gender": "Male", "institution": "Waseda University", "country": "JP", "position": "Undergrad student" }, { "name": "Yuta Saito", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" } ], "n_formula": 40, "n_formula_1": 28, "n_ref_uni": 34, "n_ref": 81, "n_ref_all": 127, "n_fig": 17, "n_tab": 6, "L_tab": 878, "n_element_tab": 109, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 428, "n_element_tab_1": 19, "formula_len_all": 3642, "formula_len_all_1": 2345, "len_all": 155742, "len_all_1": 63876, "len_abs": 1766, "len_title": 120, "len_sents": 48795, "len_sents_1": 29720, "n_sents": 337, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1773, "title": "A General Framework for Off-Policy Learning with Partially-Observed Reward", "abs": "Off-policy learning (OPL) in contextual bandits aims to learn a decision-making policy that maximizes the target rewards by using only historical interaction data collected under previously developed policies. Unfortunately, when rewards are only partially observed, the effectiveness of OPL degrades severely. Well-known examples of such partial rewards include explicit ratings in content recommendations, conversion signals on e-commerce platforms that are partial due to delay, and the issue of censoring in medical problems. One possible solution to deal with such partial rewards is to use secondary rewards, such as dwelling time, clicks, and medical indicators, which are more densely observed. However, relying solely on such secondary rewards can also lead to poor policy learning since they may not align with the target reward. Thus, this work studies a new and general problem of OPL where the goal is to learn a policy that maximizes the expected target reward by leveraging densely observed secondary rewards as supplemental data. We then propose a new method called Hybrid Policy Optimization for Partially-Observed Reward (HyPeR), which effectively uses the secondary rewards in addition to the partially observed target reward to achieve effective OPL despite the challenging scenario. We also discuss a case where we aim to optimize not only the expected target reward but also the expected secondary rewards to some extent; counter-intuitively, we will show that leveraging the two objectives is in fact advantageous also for the optimization of only the target reward. Along with statistical analysis of our proposed methods, empirical evaluations on both synthetic and real-world data show that HyPeR outperforms existing methods in various scenarios.", "keywords": [ "off-policy learning", "partially-observed rewards", "contextual bandits" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "mUXdysoxEP", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Ruiji Yu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Intern" }, { "name": "Xiaolin Huang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Xinwen Cheng", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yingwen Wu", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Zhengbao He", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 53, "n_ref": 156, "n_ref_all": 182, "n_fig": 3, "n_tab": 16, "L_tab": 7493, "n_element_tab": 1166, "n_fig_1": 1, "n_tab_1": 9, "L_tab_1": 6366, "n_element_tab_1": 710, "formula_len_all": 401, "formula_len_all_1": 226, "len_all": 170842, "len_all_1": 71673, "len_abs": 1587, "len_title": 134, "len_sents": 42803, "len_sents_1": 27501, "n_sents": 300, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1594, "title": "Pursuing Feature Separation based on Neural Collapse for Out-of-Distribution Detection", "abs": "In the open world, detecting out-of-distribution (OOD) data, whose labels are disjoint with those of in-distribution (ID) samples, is important for reliable deep neural networks (DNNs). To achieve better detection performance, one type of approach proposes to fine-tune the model with auxiliary OOD datasets to amplify the difference between ID and OOD data through a separation loss defined on model outputs. However, none of these studies consider enlarging the feature disparity, which should be more effective compared to outputs. The main difficulty lies in the diversity of OOD samples, which makes it hard to describe their feature distribution, let alone design losses to separate them from ID features. In this paper, we neatly fence off the problem based on an aggregation property of ID features named Neural Collapse (NC). NC means that the penultimate features of ID samples within a class are nearly identical to the last layer weight of the corresponding class. Based on this property, we propose a simple but effective loss called Separation Loss, which binds the features of OOD data in a subspace orthogonal to the principal subspace of ID features formed by NC. In this way, the features of ID and OOD samples are separated by different dimensions. By optimizing the feature separation loss rather than purely enlarging output differences, our detection achieves SOTA performance on CIFAR10, CIFAR100 and ImageNet benchmarks without any additional data augmentation or sampling, demonstrating the importance of feature separation in OOD detection. The code will be published.", "keywords": [ "Out-of-Distribution Detection", "Feature Separation", "Neural Collapse" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 2 ] }, { "paper_id": "mUMvr33FTu", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Jiaqi Xue", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Lei Jiang", "gender": "Male", "institution": "Indiana University", "country": "US", "position": "Assistant Professor" }, { "name": "Mengxin Zheng", "gender": "Female", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Mimi Xie", "gender": "Female", "institution": "University of Texas at San Antonio", "country": "US", "position": "Assistant Professor" }, { "name": "Mingzhe Zhang", "gender": "unknown", "institution": "Institute of Information Engineering, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Qian Lou", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Yancheng Zhang", "gender": "Not Specified", "institution": "University of Central Florida", "country": "US", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 4, "n_ref_uni": 32, "n_ref": 122, "n_ref_all": 156, "n_fig": 17, "n_tab": 7, "L_tab": 1016, "n_element_tab": 62, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 236, "n_element_tab_1": 28, "formula_len_all": 1153, "formula_len_all_1": 627, "len_all": 152141, "len_all_1": 70350, "len_abs": 2304, "len_title": 113, "len_sents": 51855, "len_sents_1": 33554, "n_sents": 441, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1829, "title": "CipherPrune: Efficient and Scalable Private Transformer Inference", "abs": "Private Transformer inference using cryptographic protocols offers promising solutions for privacy-preserving machine learning; however, it still faces significant runtime overhead (efficiency issues) and challenges in handling long-token inputs (scalability issues). We observe that the Transformer's operational complexity scales quadratically with the number of input tokens, making it essential to reduce the input token length. Notably, each token varies in importance, and many inputs contain redundant tokens. Additionally, prior private inference methods that rely on high-degree polynomial approximations for non-linear activations are computationally expensive. Therefore, reducing the polynomial degree for less important tokens can significantly accelerate private inference. Building on these observations, we propose $\\textit{CipherPrune}$, an efficient and scalable private inference framework that includes a secure encrypted token pruning protocol, a polynomial reduction protocol, and corresponding Transformer network optimizations. At the protocol level, encrypted token pruning adaptively removes unimportant tokens from encrypted inputs in a progressive, layer-wise manner. Additionally, encrypted polynomial reduction assigns lower-degree polynomials to less important tokens after pruning, enhancing efficiency without decryption. At the network level, we introduce protocol-aware network optimization via a gradient-based search to maximize pruning thresholds and polynomial reduction conditions while maintaining the desired accuracy. Our experiments demonstrate that CipherPrune reduces the execution overhead of private Transformer inference by approximately $6.1\\times$ for 128-token inputs and $10.6\\times$ for 512-token inputs, compared to previous methods, with only a marginal drop in accuracy.", "keywords": [ "Private Transformer Inference" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "mUFdrdQJds", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Aryaman Shaan", "gender": "Male", "institution": "New York University", "country": "US", "position": "MS student" }, { "name": "Blaise Genest", "gender": "Male", "institution": "CNRS", "country": "FR", "position": "Full Professor" }, { "name": "Kuldeep S. Meel", "gender": "Male", "institution": "Department of Computer Science", "country": "", "position": "Associate Professor" }, { "name": "Yuke Liao", "gender": "Male", "institution": "CNRS@CREATE", "country": "SG", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 17, "n_ref": 78, "n_ref_all": 119, "n_fig": 5, "n_tab": 16, "L_tab": 4112, "n_element_tab": 381, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1435, "n_element_tab_1": 163, "formula_len_all": 1926, "formula_len_all_1": 1315, "len_all": 145702, "len_all_1": 63948, "len_abs": 1592, "len_title": 78, "len_sents": 47941, "len_sents_1": 28457, "n_sents": 343, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 2179, "title": "Hybrid MILP to efficiently and accuratly solve hard DNN verification instances", "abs": "Deep neural networks have demonstrated remarkable capabilities, achieving human-like or even superior performance across a wide range of tasks. However, their robustness is often compromised by their susceptibility to input perturbations. This vulnerability has catalyzed the verification community to develop various methodologies, each presenting a unique balance between completeness and computational efficiency. $\\alpha,\\beta$-CROWN has won the last 4 VNNcomp(etitions), as the DNN verifier with the best \ntrade-off between accuracy vs computational time. VNNcomp however is focusing on relatively easy verification instances (network, inputs (images)), with few {\\em unstable nodes}. In this paper, we consider harder verification instances. On such instances, $\\alpha,\\beta$-CROWN displays a large gap ($20-58$%) between instances that can be verified, and instances with an explicit attack. Enabling much larger time-outs for $\\alpha,\\beta$-CROWN only improves verification rate by few percents, leaving a large gap of undecided instances while already taking a considerable amount of time. Resorting to other techniques, such as complete verifiers, does not fare better even with very large time-outs: They would theoretically be able to close the gap, but with an untractable runtime on all but small {\\em hard} instances.\n\nIn this paper, we propose a novel Utility function that selects few neurons to be encoded with accurate but costly integer variables in a {\\em partial MILP} problem. The novelty resides in the use of \nthe solution of {\\em one} (efficient LP) solver to accurately compute a selection $\\varepsilon$-optimal for a given input. \nCompared with previous attempts, we can reduce the number of integer variables by around 4 times while maintaining the same level of accuracy. Implemented in {\\em Hybrid MILP}, calling first $\\alpha,\\beta$-Crown with a short time-out to solve easier instances, and then partial MILP for those for which $\\alpha,\\beta$-Crown fails, produces a very accurate yet efficient verifier, reducing tremendously the number of undecided instances ($8-15\\%$), while keeping a reasonable runtime ($46s-417s$ on average per instance).", "keywords": [ "Safety; Neural Netowrks; Verification; robustness; MILP" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 1, 3, 2 ], "contribution_list": [ 3, 3, 1 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "mUDazL3mTJ", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Ariful Azad", "gender": "Male", "institution": "Indiana University", "country": "", "position": "Assistant Professor" }, { "name": "Nicholas Huckvale Majeske", "gender": "Male", "institution": "Indiana University", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 27, "n_ref": 61, "n_ref_all": 96, "n_fig": 13, "n_tab": 13, "L_tab": 9337, "n_element_tab": 702, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 2288, "n_element_tab_1": 205, "formula_len_all": 612, "formula_len_all_1": 560, "len_all": 138949, "len_all_1": 58418, "len_abs": 790, "len_title": 129, "len_sents": 42162, "len_sents_1": 24738, "n_sents": 307, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 794, "title": "FDN: Interpretable Spatiotemporal Forecasting with Future Decomposition Networks", "abs": "Spatiotemporal systems comprise a collection of spatially distributed yet interdependent entities each generating unique dynamic signals. \nHighly sophisticated methods have been proposed in recent years delivering state-of-the-art (SOTA) forecasts but few have focused on interpretability. \nTo address this, we propose the Future Decomposition Network (FDN), a novel forecast model capable of (a) providing interpretable predictions through classification (b) revealing latent activity patterns in the target time-series and (c) delivering forecasts competitive with SOTA methods at a fraction of their memory and runtime cost. \nWe conduct comprehensive analyses on FDN for multiple datasets from hydrologic, traffic, and energy systems demonstrating its improved accuracy and interpretability.", "keywords": [ "Spatiotemporal", "Forecast", "Graph", "Hydrology", "Traffic", "Energy" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 3, 4, 1, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 2, 4, 5, 4 ] }, { "paper_id": "mTgMLy2iPt", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Assaf Hallak", "gender": "Male", "institution": "NVIDIA", "country": "IL", "position": "Senior research scientist" }, { "name": "Gal Chechik", "gender": "unknown", "institution": "Bar Ilan University", "country": "IL", "position": "Full Professor" }, { "name": "Gugan Thoppe", "gender": "Male", "institution": "Indian Institute of Science", "country": "", "position": "Assistant Professor" }, { "name": "Shie Mannor", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Full Professor" }, { "name": "Gal Dalal", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Senior Research Scientist" } ], "n_formula": 113, "n_formula_1": 25, "n_ref_uni": 46, "n_ref": 87, "n_ref_all": 115, "n_fig": 5, "n_tab": 1, "L_tab": 72, "n_element_tab": 21, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 11535, "formula_len_all_1": 1925, "len_all": 180700, "len_all_1": 74139, "len_abs": 1189, "len_title": 83, "len_sents": 49020, "len_sents_1": 32061, "n_sents": 562, "n_sents_1": 314, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1199, "title": "Policy Gradient with Tree Expansion", "abs": "Policy gradient methods are notorious for having a large variance and high sample complexity. To mitigate this, we introduce SoftTreeMax---a generalization of softmax that employs planning. In SoftTreeMax, we extend the traditional logits with the multi-step discounted cumulative reward, topped with the logits of future states. We analyze SoftTreeMax and explain how tree expansion helps to reduce its gradient variance. We prove that the variance depends on the chosen tree-expansion policy. Specifically, we show that the closer the induced transitions are to being state-independent, the stronger the variance decay. With approximate forward models, we prove that the resulting gradient bias diminishes with the approximation error while retaining the same variance reduction. Ours is the first result to bound the gradient bias for an approximate model. In a practical implementation of SoftTreeMax we utilize a parallel GPU-based simulator for fast and efficient tree expansion. Using this implementation in Atari, we show that SoftTreeMax reduces the gradient variance by three orders of magnitude. This leads to better sample complexity and improved performance compared to distributed PPO.", "keywords": [ "Reinforcement Learning", "Policy Gradient", "Tree Expansion", "Softmax" ], "rating_list": [ 5, 3, 6, 8 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "mTCbq2QssD", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Alexan Ayrapetyan", "gender": "Male", "institution": "Russian-Armenian (Slavonic) State University", "country": "AM", "position": "MS student" }, { "name": "Branislav Kisacanin", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Igor Gitman", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Ivan Moshkov", "gender": "Male", "institution": "NVIDIA", "country": "AM", "position": "Researcher" }, { "name": "Shubham Toshniwal", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Wei Du", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 29, "n_formula_1": 5, "n_ref_uni": 28, "n_ref": 55, "n_ref_all": 88, "n_fig": 8, "n_tab": 10, "L_tab": 3252, "n_element_tab": 316, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1309, "n_element_tab_1": 136, "formula_len_all": 4342, "formula_len_all_1": 321, "len_all": 199366, "len_all_1": 62460, "len_abs": 1529, "len_title": 86, "len_sents": 63473, "len_sents_1": 27859, "n_sents": 531, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1535, "title": "OpenMathInstruct-2: Accelerating AI for Math with Massive Open-Source Instruction Data", "abs": "Mathematical reasoning continues to be a critical challenge in large language model (LLM) development with significant interest. However, most of the cutting-edge progress in mathematical reasoning with LLMs has become closed-source due to lack of access to training data. This lack of data access limits researchers from understanding the impact of different choices for synthesizing and utilizing the data. With the goal of creating a high-quality finetuning (SFT) dataset for math reasoning, we conduct careful ablation experiments on data synthesis using the recently released Llama3.1 family of models. Our experiments show that: (a) solution format matters, with excessively verbose solutions proving detrimental to SFT performance, (b) data generated by a strong teacher outperforms on-policy data generated by a weak student model, (c) SFT is robust to low-quality solutions, allowing for imprecise data filtering, and (d) question diversity is crucial for achieving data scaling gains. Based on these insights, we create the OpenMathInstruct-2 dataset which consists of 14M question-solution pairs (\u2248 600K unique questions), making it nearly eight times larger than the previous largest open-source math reasoning dataset. Finetuning the Llama-3.1-8B-Base using OpenMathInstruct-2 outperforms Llama3.1-8B-Instruct on MATH by an absolute 15.9% (51.9% \u2192 67.8%). Finally, to accelerate the open-source efforts, we release the code, the finetuned models, and the OpenMathInstruct-2 dataset under a commercially permissive license.", "keywords": [ "Math Reasoning", "Synthetic Data" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "mSYX71lNAl", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Xin Jiang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Xu Cheng", "gender": "Female", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Zechao Li", "gender": "Male", "institution": "Nanjing University of Science and Techonolgy", "country": "CN", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 3, "n_ref_uni": 41, "n_ref": 105, "n_ref_all": 116, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 646, "formula_len_all_1": 142, "len_all": 129922, "len_all_1": 58141, "len_abs": 1253, "len_title": 116, "len_sents": 41723, "len_sents_1": 30351, "n_sents": 239, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1256, "title": "Why pre-training is beneficial for downstream classification tasks?", "abs": "It is widely acknowledged that pre-training brings benefits to downstream tasks by boosting accuracy and speeding up convergence, but the exact reasons for these two benefits still remain unclear.\nTo this end, we propose to quantitatively and accurately explain effects of pre-training on the downstream task from a novel game-theoretic view, which also sheds new light into the learning behavior of deep neural networks (DNNs).\nSpecifically, we extract and quantify the knowledge encoded by the pre-trained model, and further track the changes of such knowledge during the fine-tuning process.\nInterestingly, we discover that only a limited amount of pre-trained model's knowledge is preserved for the inference of downstream tasks, and such preserved knowledge is very difficult for a model training from scratch to learn.\nThus, with the help of this exclusively learned and useful knowledge, the fine-tuned model usually achieves better performance.\nBesides, we discover that pre-training can guide the fine-tuned model to learn target knowledge of the downstream task more directly and quickly than the model training from scratch, which accounts for the faster convergence of the fine-tuned model.\nThe code will be released when the paper is accepted.", "keywords": [ "pre-training", "explainable AI" ], "rating_list": [ 5, 3, 5, 3, 6 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 1, 3 ], "contribution_list": [ 3, 2, 3, 1, 3 ], "confidence_list": [ 4, 3, 4, 5, 3 ] }, { "paper_id": "mSGcDhQPwm", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Jelena Diakonikolas", "gender": "Female", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Xufeng Cai", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" } ], "n_formula": 195, "n_formula_1": 38, "n_ref_uni": 49, "n_ref": 148, "n_ref_all": 187, "n_fig": 3, "n_tab": 1, "L_tab": 195, "n_element_tab": 3, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 27967, "formula_len_all_1": 4877, "len_all": 215311, "len_all_1": 69707, "len_abs": 2151, "len_title": 120, "len_sents": 59462, "len_sents_1": 30626, "n_sents": 510, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1487, "title": "Last Iterate Convergence of Incremental Methods as a Model of Forgetting", "abs": "Incremental gradient and incremental proximal methods are a fundamental class of optimization algorithms used for solving finite sum problems, broadly studied in the literature. Yet, without strong convexity, their convergence guarantees have primarily been established for the ergodic (average) iterate. We establish the first nonasymptotic convergence guarantees for the last iterate of both incremental gradient and incremental proximal methods, in general convex smooth (for both) and convex Lipschitz (for the proximal variants) settings. Our oracle complexity bounds for the last iterate nearly match (i.e., match up to a square-root-log or a log factor) the best known oracle complexity bounds for the average iterate, for both classes of methods. We further obtain generalizations of our results to weighted averaging of the iterates with increasing weights and for randomly permuted ordering of updates. We study last iterate convergence of the incremental proximal method as a mathematical abstraction of forgetting in continual learning and prove a lower bound that certifies that a large amount of regularization is crucial to mitigating catastrophic forgetting---one of the key considerations in continual learning. Our results generalize last iterate guarantees for incremental methods compared to state of the art, as such results were previously known only for overparameterized linear models, which correspond to convex quadratic problems with infinitely many solutions.", "keywords": [ "incremental methods", "last-iterate convergence", "proximal methods", "continual learning" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "mS7xin7BPK", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chunwei Xia", "gender": "Male", "institution": "University of Leeds", "country": "GB", "position": "Postdoc" }, { "name": "Huimin Cui", "gender": "Female", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Jiacheng Zhao", "gender": "Male", "institution": " Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Shuoming Zhang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xiaobing Feng", "gender": "Male", "institution": "ICT, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yunji Chen", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zheng Wang", "gender": "Male", "institution": "University of Leeds", "country": "GB", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 3, "n_ref_uni": 46, "n_ref": 73, "n_ref_all": 99, "n_fig": 9, "n_tab": 9, "L_tab": 7883, "n_element_tab": 574, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 816, "n_element_tab_1": 116, "formula_len_all": 973, "formula_len_all_1": 644, "len_all": 274662, "len_all_1": 59229, "len_abs": 1010, "len_title": 128, "len_sents": 78539, "len_sents_1": 27949, "n_sents": 639, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1014, "title": "LEGO-Compiler: Enhancing Neural Compilation Through Composable Chain of Thought", "abs": "Large language models (LLMs) have the potential to revolutionize how we design and implement compilers and code translation tools. However, existing LLMs struggle to handle long and complex programs. We introduce LEGO-Compiler, a novel neural compilation system that leverages LLMs to translate high-level languages into assembly code. Our approach centers on three key innovations: LEGO translation, which decomposes the input program into manageable blocks; annotation-based Chain-of-Thoughts, guiding LLMs through the compilation process with LLM-annotated context; and a feedback mechanism for self-correction. Supported by formal proofs of code composability, LEGO-Compiler demonstrates high accuracy on multiple datasets, including over 99% on ExeBench and 100% on industrial-grade CoreMark, and successfully handles programs far exceeding the length limitations of native LLM translation. This work opens new avenues for applying LLMs to system-level tasks, complementing traditional compiler technologies.", "keywords": [ "code translation", "neural compilation", "chain of thought", "scalability", "in-context learning" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 5, 2 ] }, { "paper_id": "mQ55y4s5hj", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Byonghyo Shim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Donghoon Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Kyuhong Shim", "gender": "Male", "institution": "Qualcomm Inc, Qualcomm", "country": "KR", "position": "Researcher" }, { "name": "Minji Bae", "gender": "Female", "institution": "Seoul National University", "country": "", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 45, "n_ref": 71, "n_ref_all": 102, "n_fig": 15, "n_tab": 9, "L_tab": 8851, "n_element_tab": 540, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 303, "n_element_tab_1": 40, "formula_len_all": 415, "formula_len_all_1": 310, "len_all": 148803, "len_all_1": 54691, "len_abs": 1393, "len_title": 130, "len_sents": 34880, "len_sents_1": 24056, "n_sents": 258, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1426, "title": "Visually Guided Decoding: Gradient-Free Hard Prompt Inversion with Language Models", "abs": "Text-to-image generative models like DALL-E and Stable Diffusion have revolutionized visual content creation across various applications, including advertising, personalized media, and design prototyping. \nHowever, crafting effective textual prompts to guide these models remains challenging, often requiring extensive trial and error. \nExisting prompt inversion methods, such as soft and hard prompt techniques, suffer from issues like limited interpretability and incoherent prompt generation. \nTo address these limitations, we introduce Visually Guided Decoding (VGD), a gradient-free approach that leverages large language models (LLMs) and CLIP-based guidance to generate coherent and semantically aligned prompts. \nVGD utilizes the robust text generation capabilities of LLMs to produce human-readable prompts while employing CLIP scores to ensure alignment with user-specified visual concepts. \nThis method enhances the interpretability, generalization, and flexibility of prompt generation without the need for additional training. \nOur experiments demonstrate that VGD outperforms existing prompt inversion techniques in generating understandable and contextually relevant prompts, facilitating more intuitive and controllable interactions with text-to-image models. \nVGD's compatibility with various LLMs, including LLama2, LLama3, and Mistral, makes it a versatile solution for enhancing image generation workflows.", "keywords": [ "text-to-image", "inversion", "gradient free hard prompt inversion", "language model guidance on latent diffusion model" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "mPyPm9mmc6", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Mengmi Zhang", "gender": "Female", "institution": "Nanyang Technological University, Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Shuangpeng Han", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Ziyu Wang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zheng Shou", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 32, "n_ref": 66, "n_ref_all": 83, "n_fig": 9, "n_tab": 3, "L_tab": 642, "n_element_tab": 103, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 671, "n_element_tab_1": 107, "formula_len_all": 618, "formula_len_all_1": 619, "len_all": 121754, "len_all_1": 73103, "len_abs": 1339, "len_title": 124, "len_sents": 37909, "len_sents_1": 32435, "n_sents": 323, "n_sents_1": 277, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1347, "title": "Unsupervised Prior Learning: Discovering Categorical Pose Priors from Videos", "abs": "A prior represents a set of beliefs or assumptions about a system, aiding inference and decision-making. In this work, we introduce the challenge of unsupervised prior learning in pose estimation, where AI models learn pose priors of animate objects from videos in a self-supervised manner. These videos present objects performing various actions, providing crucial information about their keypoints and connectivity.\nWhile priors are effective in pose estimation, acquiring them can be difficult. We propose a novel method, named Pose Prior Learner (PPL), to learn general pose priors applicable to any object category. PPL uses a hierarchical memory to store compositional parts of prototypical poses, from which we distill a general pose prior. This prior enhances pose estimation accuracy through template transformation and image reconstruction. PPL learns meaningful pose priors without any additional human annotations or interventions, outperforming competitive baselines on both human and animal pose estimation datasets. Notably, our experimental results reveal the effectiveness of PPL using learnt priors for pose estimation on occluded images. Through iterative inference, PPL leverages priors to refine estimated poses, regressing them to any prototypical poses stored in memory. Our code, model, and data will be publicly available.", "keywords": [ "unsupervised prior learning", "unsupervised pose estimation", "memory" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "mPdmDYIQ7f", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Fengli Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jiahe Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Keyu Zhao", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Likai Ma", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yu Li", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yu Shang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 37, "n_ref": 89, "n_ref_all": 102, "n_fig": 5, "n_tab": 4, "L_tab": 4029, "n_element_tab": 20, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1392, "n_element_tab_1": 5, "formula_len_all": 398, "formula_len_all_1": 154, "len_all": 151225, "len_all_1": 57020, "len_abs": 1611, "len_title": 111, "len_sents": 43805, "len_sents_1": 26569, "n_sents": 300, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 63, "L_abs": 1618, "title": "AgentSquare: Automatic LLM Agent Search in Modular Design Space", "abs": "Recent advancements in Large Language Models (LLMs) have led to a rapid growth of agentic systems capable of handling a wide range of complex tasks. However, current research largely relies on manual, task-specific design, limiting their adaptability to novel tasks. In this paper, we introduce a new research problem: Modularized LLM Agent Search (MoLAS). We propose a modular design space that abstracts existing LLM agent designs into four fundamental modules with uniform IO interface: Planning, Reasoning, Tool Use, and Memory. Building on this design space, we present a novel LLM agent search framework called AgentSquare, which introduces two core mechanisms, i.e., module evolution and recombination, to efficiently search for optimized LLM agents. To further accelerate the process, we design a performance predictor that uses in-context surrogate models to skip unpromising agent designs. Extensive experiments across six benchmarks, covering the diverse scenarios of web, embodied, tool use and game applications, show that AgentSquare substantially outperforms hand-crafted agents, achieving an average performance gain of 17.2% against best-known human designs. Moreover, AgentSquare can generate interpretable design insights, enabling a deeper understanding of agentic architecture and its impact on task performance. We believe that the modular design space and AgentSquare search framework offer a platform for fully exploiting the potential of prior successful designs and consolidate the collective efforts of research community. Code repo is available at https://github.com/ICLR-10021/AgentSquare.", "keywords": [ "LLM agent", "Modular design space", "Agent search", "AutoML" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "mPMLZv4kSL", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Dongyu Xue", "gender": "Male", "institution": "ByteDance AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Fei YE", "gender": "Female", "institution": "ByteDance AI Lab", "country": "", "position": "Researcher" }, { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Yi Zhou", "gender": "Male", "institution": "ByteDance AI Lab", "country": "CN", "position": "Research Scientist" }, { "name": "Yiming Ma", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Zaixiang Zheng", "gender": "Male", "institution": "ByteDance Research", "country": "", "position": "Research Scientist" } ], "n_formula": 21, "n_formula_1": 9, "n_ref_uni": 24, "n_ref": 55, "n_ref_all": 103, "n_fig": 19, "n_tab": 5, "L_tab": 3626, "n_element_tab": 222, "n_fig_1": 17, "n_tab_1": 2, "L_tab_1": 166, "n_element_tab_1": 6, "formula_len_all": 1837, "formula_len_all_1": 1151, "len_all": 152855, "len_all_1": 62502, "len_abs": 2864, "len_title": 122, "len_sents": 50160, "len_sents_1": 28897, "n_sents": 398, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1125, "title": "ProteinWeaver: A Divide-and-Assembly Approach for Protein Backbone Design", "abs": "Nature creates diverse proteins through a 'divide and assembly' strategy. Inspired by this idea, we introduce ProteinWeaver, a two-stage framework for protein backbone design. Our method first generates individual protein domains, then employs an SE(3) diffusion model to flexibly assemble these domains. A key challenge lies in the assembling step, given the complex and rugged nature of the inter-domain interaction landscape. To address this challenge, we employ preference alignment to discern complex relationships between structure and interaction landscapes through comparative analysis of generated samples. Comprehensive experiments demonstrate that ProteinWeaver: (1) generates high-quality, novel protein backbones through versatile domain assembly; (2) outperforms RFdiffusion, the current state-of-the-art in backbone design, by 13% and 39% for long-chain proteins; (3) shows the potential for cooperative function design through illustrative case studies. To sum up, by introducing a 'divide-and-assembly' paradigm, ProteinWeaver advances protein engineering and opens new avenues for functional protein design.", "keywords": [ "divide and assembly", "protein backbone design", "domain assembly", "long-chain proteins" ], "rating_list": [ 3, 6, 6, 3, 3 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 3, 1, 2 ], "confidence_list": [ 5, 3, 4, 4, 4 ] }, { "paper_id": "mP7uV59iJM", "primary_area": "optimization", "n_author": 8, "author_info_list": [ { "name": "Ming Cheng", "gender": "unknown", "institution": "Dartmouth College", "country": "US", "position": "MS student" }, { "name": "Shuai Chen", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Siyan HU", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Tristan Braud", "gender": "Not Specified", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "Assistant Professor" }, { "name": "Victor Adrian Prisacariu", "gender": "Male", "institution": "Niantic", "country": "US", "position": "Chief Scientist" }, { "name": "Yash Sanjay Bhalgat", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Zirui Wang", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Changkun LIU", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 34, "n_ref": 71, "n_ref_all": 106, "n_fig": 11, "n_tab": 14, "L_tab": 4209, "n_element_tab": 369, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2022, "n_element_tab_1": 196, "formula_len_all": 111, "formula_len_all_1": 111, "len_all": 143765, "len_all_1": 58058, "len_abs": 421, "len_title": 114, "len_sents": 32753, "len_sents_1": 22477, "n_sents": 277, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1079, "title": "GS-CPR: Efficient Camera Pose Refinement via 3D Gaussian Splatting", "abs": "We leverage 3D Gaussian Splatting (3DGS) as a scene representation and propose a novel test-time camera pose refinement (CPR) framework, GS-CPR. This framework enhances the localization accuracy of state-of-the-art absolute pose regression and scene coordinate regression methods. The 3DGS model renders high-quality synthetic images and depth maps to facilitate the establishment of 2D-3D correspondences. GS-CPR obviates the need for training feature extractors or descriptors by operating directly on RGB images, utilizing the 3D foundation model, MASt3R, for precise 2D matching. To improve the robustness of our model in challenging outdoor environments, we incorporate an exposure-adaptive module within the 3DGS framework. Consequently, GS-CPR enables efficient one-shot pose refinement given a single RGB query and a coarse initial pose estimation. Our proposed approach surpasses leading NeRF-based optimization methods in both accuracy and runtime across indoor and outdoor visual localization benchmarks, achieving new state-of-the-art accuracy on two indoor datasets.", "keywords": [ "Visual Localization", "Camera Pose Estimation", "3D Gaussian Splatting" ], "rating_list": [ 6, 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3, 1 ], "contribution_list": [ 3, 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "mOpNrrV2zH", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Cheng Tan", "gender": "Male", "institution": "Zhejiang University & Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Guojiang Zhao", "gender": "Male", "institution": "DP Technology", "country": "CN", "position": "Postdoc" }, { "name": "Haitao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Lirong Wu", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Odin Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "Undergrad student" }, { "name": "Yufei Huang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhifeng Gao", "gender": "Male", "institution": "DP Technology", "country": "", "position": "Researcher" }, { "name": "Zicheng Liu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Stan Z. Li", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Chair Professor" } ], "n_formula": 9, "n_formula_1": 0, "n_ref_uni": 56, "n_ref": 87, "n_ref_all": 91, "n_fig": 2, "n_tab": 20, "L_tab": 14942, "n_element_tab": 1290, "n_fig_1": 0, "n_tab_1": 9, "L_tab_1": 3840, "n_element_tab_1": 425, "formula_len_all": 948, "formula_len_all_1": 0, "len_all": 237052, "len_all_1": 76394, "len_abs": 1875, "len_title": 117, "len_sents": 56113, "len_sents_1": 34383, "n_sents": 407, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 69, "L_abs": 1880, "title": "CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph", "abs": "Structure-based drug design (SBDD) aims to generate potential drugs that can bind to a target protein and is greatly expedited by the aid of AI techniques in generative models. However, a lack of systematic understanding persists due to the diverse settings, complex implementation, difficult reproducibility, and task singularity. Firstly, the absence of standardization can lead to unfair comparisons and inconclusive insights. To address this dilemma, we propose CBGBench, a comprehensive benchmark for SBDD, that unifies the task as a generative graph completion, analogous to fill-in-the-blank of the 3D complex binding graph. By categorizing existing methods based on their attributes, CBGBench facilitates a modular and extensible framework that implements cutting-edge methods. Secondly, a single de novo molecule generation task can hardly reflect their capabilities. To broaden the scope, we adapt these models to a range of tasks essential in drug design, considered sub-tasks within the graph fill-in-the-blank tasks. These tasks include the generative designation of de novo molecules, linkers, fragments, scaffolds, and sidechains, all conditioned on the structures of protein pockets. Our evaluations are conducted with fairness, encompassing comprehensive perspectives on interaction, chemical properties, geometry authenticity, and substructure validity. We further provide insights with analysis from empirical studies. Our results indicate that there is potential for further improvements on many tasks, with optimization in network architectures, and effective incorporation of chemical prior knowledge. Finally, to lower the barrier to entry and facilitate further developments in the field, we also provide a single [codebase](https://github.com/EDAPINENUT/CBGBench) that unifies the discussed models, data pre-processing, training, sampling, and evaluation.", "keywords": [ "Molecule Generation Benchmark", "Target-Aware Drug Design", "Generative Model" ], "rating_list": [ 8, 5, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 4, 2, 4, 4, 4 ] }, { "paper_id": "mORwTTZfWq", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "YU CHENYUE", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "yang he", "gender": "Male", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "SG", "position": "Researcher" } ], "n_formula": 62, "n_formula_1": 16, "n_ref_uni": 31, "n_ref": 50, "n_ref_all": 85, "n_fig": 17, "n_tab": 13, "L_tab": 9497, "n_element_tab": 588, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 17, "n_element_tab_1": 1, "formula_len_all": 3948, "formula_len_all_1": 1024, "len_all": 177067, "len_all_1": 56133, "len_abs": 1083, "len_title": 89, "len_sents": 56201, "len_sents_1": 26833, "n_sents": 473, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1086, "title": "Adversarial Attack Robust dataset pruning", "abs": "Dataset pruning, while effective for reducing training data size, often leads to models vulnerable to adversarial attacks. This paper introduces a novel approach to create adversarially robust coresets. We first theoretically analyze how existing pruning methods result in non-smooth loss surfaces, increasing susceptibility to attacks. To address this, we propose two key innovations: (1) a Frequency-Selective Excitation Network (FSE-Net) that dynamically selects important frequency components, smoothing the loss surface while reducing storage requirements, and (2) a \"Jointentropy\" score for selecting stable and informative samples. Our method significantly outperforms state-of-the-art pruning algorithms across various adversarial attacks and pruning ratios. On CIFAR-10, our approach achieves up to 58.19% accuracy under AutoAttack with an 80% pruning ratio, compared to 42.98% for previous methods. Moreover, our frequency pruning technique improves robustness even on full datasets, demonstrating its potential for enhancing model security while reducing computational costs.", "keywords": [ "dataset condensation" ], "rating_list": [ 5, 6, 6, 6, 5 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "mNkPAY3kvk", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Shentong Mo", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Sukmin Yun", "gender": "unknown", "institution": "Hanyang University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 22, "n_ref": 95, "n_ref_all": 113, "n_fig": 3, "n_tab": 8, "L_tab": 1399, "n_element_tab": 162, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2026, "n_element_tab_1": 139, "formula_len_all": 240, "formula_len_all_1": 119, "len_all": 159934, "len_all_1": 69641, "len_abs": 5860, "len_title": 123, "len_sents": 59508, "len_sents_1": 36328, "n_sents": 391, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1785, "title": "Toward Escaping Model Collapse: Aligning Generated Images as a New Modality", "abs": "Generative models have made it possible to synthesize highly realistic images, potentially providing an abundant data source for training machine learning models. Despite the advantages of these synthesizable data sources, the indiscriminate use of generated images as real images for training can harm model performance and even cause model collapse due to modality discrepancies between real and synthetic domains. In this paper, we propose a novel framework for discriminative use of generated images, coined $\\textit{GenRA}$ ($\\textbf{Gen}$erated-$\\textbf{R}$eal $\\textbf{A}$lignment), that explicitly treats generated images as a separate modality from real images. Instead of indiscriminately replacing real images with generated ones in the input space, our approach bridges the two distinct modalities in the same latent space through a multi-modal learning approach.\nTo be specific, we first fine-tune a model exclusively on generated images using a cross-modality alignment loss and then employ this aligned model to further train various vision-language models with generated images. By aligning the two modalities, our approach effectively leverages the benefits of recent advances in generative models, thereby boosting the effectiveness of generated image training across a range of vision-language tasks. Our framework can be easily incorporated with various vision-language models, and we demonstrate its efficacy throughout extensive experiments. For example, our framework significantly improves performance on image captioning, zero-shot image retrieval, zero-shot image classification, and long caption retrieval tasks. It also shows positive generated data scaling trends and notable enhancements in the captioning performance of the large multimodal model, LLaVA.", "keywords": [ "diffusion models", "generated visual learning", "vision-language models" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 4, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "mNVR9jJYqK", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Junfeng Zhao", "gender": "Female", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Tianlong Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Xin Gao", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Xinyu Ma", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Xu Chu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Yang Lin", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yasha Wang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Yifeng Xu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 23, "n_formula_1": 15, "n_ref_uni": 30, "n_ref": 65, "n_ref_all": 84, "n_fig": 6, "n_tab": 10, "L_tab": 2107, "n_element_tab": 242, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 577, "n_element_tab_1": 95, "formula_len_all": 3720, "formula_len_all_1": 1307, "len_all": 157069, "len_all_1": 78103, "len_abs": 1197, "len_title": 129, "len_sents": 51950, "len_sents_1": 35911, "n_sents": 402, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1197, "title": "DRESSing Up LLM: Efficient Stylized Question-Answering via Style Subspace Editing", "abs": "We introduce DRESS, a novel approach for generating stylized large language model (LLM) responses through representation editing. Existing methods like prompting and fine-tuning are either insufficient for complex style adaptation or computationally expensive, particularly in tasks like NPC creation or character role-playing. Our approach leverages the over-parameterized nature of LLMs to disentangle a style-relevant subspace within the model's representation space to conduct representation editing, ensuring a minimal impact on the original semantics. By applying adaptive editing strengths, we dynamically adjust the steering vectors in the style subspace to maintain both stylistic fidelity and semantic integrity. We develop two stylized QA benchmark datasets to validate the effectiveness of DRESS, and the results demonstrate significant improvements compared to baseline methods such as prompting and ITI. In short, DRESS is a lightweight, train-free solution for enhancing LLMs with flexible and effective style control, making it particularly useful for developing stylized conversational agents. Codes and benchmark datasets are available at https://github.com/ArthurLeoM/DRESS-LLM.", "keywords": [ "Large Language Models", "Stylized Question-Answering", "Representation Editing" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "mMzp3ImIco", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Ben Jourdan", "gender": "Male", "institution": "Edinburgh University, University of Edinburgh", "country": "", "position": "PhD student" }, { "name": "Gregory Schwartzman", "gender": "unknown", "institution": "Japan Advanced Institute of Science and Technology", "country": "JP", "position": "Associate Professor" } ], "n_formula": 54, "n_formula_1": 31, "n_ref_uni": 30, "n_ref": 57, "n_ref_all": 66, "n_fig": 1, "n_tab": 1, "L_tab": 221, "n_element_tab": 35, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5565, "formula_len_all_1": 2479, "len_all": 125975, "len_all_1": 74179, "len_abs": 1245, "len_title": 73, "len_sents": 37440, "len_sents_1": 28381, "n_sents": 357, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Reject", "L_title": 27, "L_abs": 1394, "title": "Mini-batch kernel $k$-means", "abs": "We present the first mini-batch kernel $k$-means algorithm, offering an order of magnitude improvement in running time compared to the full batch algorithm. A single iteration of our algorithm takes $\\widetilde{O}(kb^2)$ time, significantly faster than the $O(n^2)$ time required by the full batch kernel $k$-means, where $n$ is the dataset size and $b$ is the batch size. Extensive experiments demonstrate that our algorithm consistently achieves a 10-100x speedup with minimal loss in quality, addressing the slow runtime that has limited kernel $k$-means adoption in practice. We further complement these results with a theoretical analysis under an early stopping condition, proving that with a batch size of $\\widetilde{\\Omega}(\\max \\set{\\gamma^{4}, \\gamma^{2}}\\cdot k \\epsilon^{-2} )$, the algorithm terminates in $O(\\gamma^2/\\epsilon)$ iterations with high probability, where $\\gamma$ bounds the norm of points in feature space and $\\epsilon$ is a termination threshold. Our analysis holds for any reasonable center initialization, and when using $k$-means++ initialization, the algorithm achieves an approximation ratio of $O(\\log k)$ in expectation. For normalized kernels, such as Gaussian or Laplacian it holds that $\\gamma=1$. Taking $\\epsilon = O(1)$ and $b=\\Theta(k \\log n)$, the algorithm terminates in $O(1)$ iterations, with each iteration running in $\\widetilde{O}(k^3)$ time.", "keywords": [ "kernel k-means", "mini-batch" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 4, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 2 ] }, { "paper_id": "mMmzHS28ht", "primary_area": "foundation or frontier models, including LLMs", "n_author": 22, "author_info_list": [ { "name": "Ameya Sunil Mahabaleshwarkar", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Ashwath Aithal", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Bryan Catanzaro", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Vice President" }, { "name": "Chenhan D. Yu", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Daniel Korzekwa", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Gerald Shen", "gender": "unknown", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Hayley Ross", "gender": "Female", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Jan Kautz", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "VP Research" }, { "name": "Marcin Chochowski", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Mohammad Shoeybi", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Director of Applied Resesrch" }, { "name": "Mostofa Patwary", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Oleksii Kuchaiev", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Oluwatobi Olabiyi", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Senior Engineering Manager" }, { "name": "Pavlo Molchanov", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Research Scientist" }, { "name": "Raviraj Bhuminand Joshi", "gender": "Male", "institution": "Department of Computer Science, Indian Institute of Technology, Madras, Indian Institute of Technology, Madras", "country": "", "position": "MS student" }, { "name": "Saurav Muralidharan", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Sharath Turuvekere Sreenivas", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Deep Learning Engineer" }, { "name": "Wei-Chun Chen", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "ZIJIA CHEN", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Gaby Zeng", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "SHIZHE DIAO", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yoshihiko Suhara", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 16, "n_ref": 37, "n_ref_all": 69, "n_fig": 9, "n_tab": 5, "L_tab": 1912, "n_element_tab": 130, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 488, "n_element_tab_1": 85, "formula_len_all": 52, "formula_len_all_1": 0, "len_all": 120351, "len_all_1": 46368, "len_abs": 1205, "len_title": 88, "len_sents": 25065, "len_sents_1": 21922, "n_sents": 166, "n_sents_1": 147, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1206, "title": "LLM Pruning and Distillation in Practice", "abs": "Structured pruning with knowledge distillation is a potent combination for obtaining small language models (SLMs) with significantly fewer training tokens and compute resources compared to training from scratch. In this work, we investigate how this strategy can be effectively applied in instances where access to the the original pretraining dataset is restricted. We introduce a new *teacher correction* phase before distillation which lets the teacher model adjust to our specific data distribution using a lightweight fine-tuning phase. We apply this strategy to compress the Mistral NeMo 12B and Llama 3.1 8B models to 8B and 4B parameters, respectively, using pruning and distillation. We explore two distinct pruning strategies: (1) depth pruning and (2) joint hidden/attention/MLP (width) pruning, and evaluate the results on common benchmarks from the LM Evaluation Harness. The models are then aligned with NeMo Aligner and further tested for instruction following, role-play, math, coding and function calling capabilities. This approach produces the state-of-the-art Mistral-NeMo-Compressed-8B (\\MNMinitron for brevity) model from Mistral NeMo 12B, and a compelling 4B model from Llama 3.1 8B.", "keywords": [ "llm", "compression", "pruning", "distillation" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 1, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "mMjSc5fspq", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Alberto Raposo", "gender": "Male", "institution": "Pontif\u00edcia Universidade Cat\u00f3lica do Rio de Janeiro", "country": "BR", "position": "Associate Professor" }, { "name": "HELIO CORTES VIEIRA LOPES", "gender": "Male", "institution": "Pontif\u00edcia Universidade Cat\u00f3lica do Rio de Janeiro", "country": "BR", "position": "Associate Professor" }, { "name": "Luiz Schirmer", "gender": "Male", "institution": "Universidade Vale do Rio dos Sinos", "country": "BR", "position": "Assistant Professor" }, { "name": "Luiz Velho", "gender": "Male", "institution": "IMPA", "country": "BR", "position": "Full Professor" }, { "name": "Nuno Gon\u00e7alves", "gender": "Male", "institution": "University of Coimbra", "country": "PT", "position": "Assistant Professor" }, { "name": "Tiago Novello", "gender": "Male", "institution": "Instituto Nacional de Matem\u00e1tica Pura e Aplicada - IMPA", "country": "BR", "position": "Researcher" }, { "name": "Vin\u00edcius Da Silva", "gender": "Male", "institution": "Tecgraf Institute of Technical-Scientific Software Development of PUC-Rio (Tecgraf/PUC-Rio)", "country": "BR", "position": "Researcher" }, { "name": "Guilherme Gon\u00e7alves Schardong", "gender": "Male", "institution": "Institute of Systems and Robotics, University of Coimbra", "country": "PT", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 32, "n_ref": 55, "n_ref_all": 102, "n_fig": 20, "n_tab": 8, "L_tab": 2377, "n_element_tab": 380, "n_fig_1": 16, "n_tab_1": 1, "L_tab_1": 1023, "n_element_tab_1": 30, "formula_len_all": 762, "formula_len_all_1": 759, "len_all": 142794, "len_all_1": 67643, "len_abs": 1523, "len_title": 108, "len_sents": 45955, "len_sents_1": 28873, "n_sents": 443, "n_sents_1": 274, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1549, "title": "Smooth Real-time Rendering via Implicit Nested Neighborhoods", "abs": "Implicit neural representations (INRs) for surfaces have been mostly used as intermediary representations before triangle mesh extraction. Extracting meshes is not a real-time task and introduces unnecessary discretization to rendering, making it difficult to fully use the smoothness of INRs in applications. Smooth INRs are broadly used for approximating surface \\textit{signed distance functions} (SDFs) through an implicit regularization (Eikonal equation) using their available high-order derivatives. Such property also makes it easier to integrate those INRs in pipelines that explore differentiable properties of the underlying surface. The current real-time state-of-the-art approach uses grid-based data-structures that introduce discretization, resulting in a non-smooth representation.\n\nWe propose an end-to-end smooth ($C^{\\infty}$) INR framework to represent and render surfaces in real-time using neural SDFs endowed with smooth attributes such as normals and textures. Our approach leverages from a novel localized SDF training based on nested neighborhoods, a multiscale surface representation, and residual training. The framework does not depend on spatial data-structures, nor surface extraction. We show that our representation renders detailed smooth surfaces in real-time while the previous works can only render coarse non-smooth surfaces. We also present applications of our representation, including integration with a pipeline for dynamic surfaces and a way to improve performance of surface extraction via marching cubes.", "keywords": [ "Implicit Neural Representations", "Neural Signed Distance Functions", "Neural Rendering", "Real-time Rendering", "Surfaces and Attributes" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "mMhZS7qt0U", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Blake Olson", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Undergrad student" }, { "name": "Cong Fu", "gender": "Male", "institution": "Texas A&M", "country": "US", "position": "PhD student" }, { "name": "Heng Ji", "gender": "Female", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Shuiwang Ji", "gender": "Male", "institution": "Texas A&M University", "country": "US", "position": "Professor" }, { "name": "Xiner Li", "gender": "Female", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" } ], "n_formula": 47, "n_formula_1": 20, "n_ref_uni": 56, "n_ref": 115, "n_ref_all": 131, "n_fig": 5, "n_tab": 3, "L_tab": 1629, "n_element_tab": 76, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 163, "n_element_tab_1": 3, "formula_len_all": 11009, "formula_len_all_1": 4016, "len_all": 238264, "len_all_1": 74456, "len_abs": 1590, "len_title": 155, "len_sents": 60380, "len_sents_1": 31731, "n_sents": 465, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 107, "L_abs": 1453, "title": "Fragment and Geometry Aware Tokenization of Molecules for Structure-Based Drug Design Using Language Models", "abs": "Structure-based drug design (SBDD) is crucial for developing specific and effective therapeutics against protein targets but remains challenging due to complex protein-ligand interactions and vast chemical space. Although language models (LMs) have excelled in natural language processing, their application in SBDD is underexplored. To bridge this gap, we introduce a method, known as Frag2Seq, to apply LMs to SBDD by generating molecules in a fragment-based manner in which fragments correspond to functional modules. We transform 3D molecules into fragment-informed sequences using $SE(3)$-equivariant molecule and fragment local frames, extracting $SE(3)$-invariant sequences that preserve geometric information of 3D fragments. Furthermore, we incorporate protein pocket embeddings obtained from a pre-trained inverse folding model into the LMs via cross-attention to capture protein-ligand interaction, enabling effective target-aware molecule generation. Benefiting from employing LMs with fragment-based generation and effective protein context encoding, our model achieves the best performance on binding vina score and chemical properties such as QED and Lipinski, which shows our model\u2019s efficacy in generating drug-like ligands with higher binding affinity against target proteins. Moreover, our method also exhibits higher sampling efficiency compared to atom-based autoregressive and diffusion baselines with at most $\\times 300$ speedup.", "keywords": [ "Generative models", "language models", "molecule tokenization", "structure-based drug design", "fragment" ], "rating_list": [ 6, 6, 3, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "mMfDfJ8JFJ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Qirui Chen", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "MS student" }, { "name": "Shangzhe Di", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Weidi Xie", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yudi Shi", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 30, "n_ref": 118, "n_ref_all": 129, "n_fig": 2, "n_tab": 9, "L_tab": 2733, "n_element_tab": 354, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1394, "n_element_tab_1": 169, "formula_len_all": 225, "formula_len_all_1": 382, "len_all": 169730, "len_all_1": 51406, "len_abs": 1799, "len_title": 102, "len_sents": 42326, "len_sents_1": 23153, "n_sents": 325, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 989, "title": "Unlocking Video-LLM via Agent-of-Thoughts Distillation", "abs": "This paper tackles the problem of video question answering (VideoQA), \na task that often requires multi-step reasoning and a profound understanding of spatial-temporal dynamics. While large generative video-language models perform well on benchmarks, they often lack explainability and spatial-temporal grounding. \nIn this paper, we propose **A**gent-**o**f-**T**houghts **D**istillation (**AoTD**), a method that enhances generative models by incorporating automatically generated Chain-of-Thoughts (CoTs) into the instruction-tuning process. Specifically, we leverage an agent-based system to decompose complex questions into sub-tasks, and address them with specialized vision models, the intermediate results are then treated as reasoning chains. \nWe also introduce a verification mechanism using a large language model (LLM) to ensure the reliability of generated CoTs. Extensive experiments demonstrate that AoTD improves the performance on multiple-choice and open-ended benchmarks.", "keywords": [ "Video QA", "Visual Understanding", "MLLM", "Agent-based Video Analysis" ], "rating_list": [ 1, 5, 6, 6, 5 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 3, 2 ], "confidence_list": [ 5, 3, 3, 4, 2 ] }, { "paper_id": "mMXdHyBcHh", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Danqing Liu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Jiajie Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Juanzi Li", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Lei Hou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ling Feng", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Minhao Zou", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Shulin Cao", "gender": "Female", "institution": "Zhipu AI", "country": "CN", "position": "Researcher" }, { "name": "Wanjun Gu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Xin Lv", "gender": "Male", "institution": "Zhipu AI", "country": "CN", "position": "Researcher" }, { "name": "Yushi Bai", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yuxiao Dong", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 15, "n_ref": 69, "n_ref_all": 103, "n_fig": 11, "n_tab": 8, "L_tab": 1399, "n_element_tab": 115, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 3397, "n_element_tab_1": 192, "formula_len_all": 114, "formula_len_all_1": 114, "len_all": 163575, "len_all_1": 65746, "len_abs": 2983, "len_title": 125, "len_sents": 49971, "len_sents_1": 30410, "n_sents": 351, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1470, "title": "LongCite: Enabling LLMs to Generate Fine-grained Citations in Long-context QA", "abs": "Though current long-context large language models (LLMs) have demonstrated impressive capacities in answering user questions based on extensive text, the lack of citations in their responses makes user verification difficult, leading to concerns about their trustworthiness due to their potential hallucinations. In this work, we aim to enable long-context LLMs to generate responses with fine-grained sentence-level citations, improving their faithfulness and verifiability. We first introduce LongBench-Cite, an automated benchmark for assessing current LLMs' performance in Long-Context Question Answering with Citations (LQAC), revealing considerable room for improvement. To this end, we propose CoF (Coarse to Fine), a novel pipeline that utilizes off-the-shelf LLMs to automatically generate long-context QA instances with precise sentence-level citations, and leverage this pipeline to construct LongCite-45k, a large-scale SFT dataset for LQAC. Finally, we train LongCite-8B and LongCite-9B using the LongCite-45k dataset, successfully enabling their generation of accurate responses and fine-grained sentence-level citations in a single output. The evaluation results on LongBench-Cite show that our trained models achieve state-of-the-art citation quality, surpassing advanced proprietary models including GPT-4o. We also discover that SFT with citation information can further improve the correctness of model responses compared to standard long-context SFT.", "keywords": [ "long context", "large language model", "attributed LLM" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 5, 4, 4 ] }, { "paper_id": "mMXCMoU95Y", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Asif Ekbal", "gender": "Male", "institution": "Indian Institute of Technology, Jodhpur", "country": "", "position": "Full Professor" }, { "name": "Dibyanayan Bandyopadhyay", "gender": "unknown", "institution": "Indian Institute of Technology, Patna", "country": "IN", "position": "PhD student" }, { "name": "Soham Bhattacharjee", "gender": "unknown", "institution": "Christ University", "country": "IN", "position": "MS student" } ], "n_formula": 37, "n_formula_1": 14, "n_ref_uni": 14, "n_ref": 24, "n_ref_all": 43, "n_fig": 3, "n_tab": 6, "L_tab": 1113, "n_element_tab": 78, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1121, "n_element_tab_1": 77, "formula_len_all": 3111, "formula_len_all_1": 749, "len_all": 128833, "len_all_1": 52856, "len_abs": 1491, "len_title": 146, "len_sents": 37725, "len_sents_1": 22359, "n_sents": 342, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1495, "title": "CAuSE: Post-hoc Natural Language Explanation of Multimodal Classifiers through Causal Abstraction", "abs": "The increasing integration of AI models in critical areas, such as healthcare, finance, and security has raised concerns about their black-box nature, limiting trust and accountability. To ensure robust and trustworthy AI, interpretability is essential. In this paper, we propose CAuSE (Causal Abstraction under Simulated Explanation), a novel framework for post-hoc explanation of multimodal classifiers. Unlike existing interpretability methods, such as Amnesic Probing and Integrated Gradients, CAuSE generates causally faithful natural language explanations of fine-tuned multimodal classifiers' decisions. CAuSE integrates Interchange Intervention Training (IIT) within a Language Model (LM) based module to simulate the causal reasoning behind a classifier's outputs. We introduce a novel metric Counterfactual F1 score to measure causal faithfulness and demonstrate that CAuSE achieves state-of-the-art performance on this metric. We also provide a rigorous theoretical underpinning for causal abstraction between two neural networks and implement this within our CAuSE framework. This ensures that CAuSE\u2019s natural language explanations are not only simulations of the classifier\u2019s behavior but also reflect its underlying causal processes. Our method is task-agnostic and achieves state-of-the-art results on benchmark multimodal classification datasets, such as e-SNLI-VE and Facebook Hateful Memes, offering a scalable, faithful solution for interpretability in multimodal classifiers.", "keywords": [ "Interpretability", "Causal Abstraction", "Multimodality", "Classification" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 1, 2, 2 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "mMPaQzgzAN", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Arthur Conmy", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Janos Kramar", "gender": "Male", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Neel Nanda", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Nicolas Sonnerat", "gender": "unknown", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Senthooran Rajamanoharan", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Tom Lieberum", "gender": "unknown", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Vikrant Varma", "gender": "unknown", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" } ], "n_formula": 49, "n_formula_1": 16, "n_ref_uni": 36, "n_ref": 91, "n_ref_all": 194, "n_fig": 21, "n_tab": 2, "L_tab": 480, "n_element_tab": 18, "n_fig_1": 13, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2104, "formula_len_all_1": 747, "len_all": 189725, "len_all_1": 66412, "len_abs": 1477, "len_title": 131, "len_sents": 69575, "len_sents_1": 32965, "n_sents": 436, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1245, "title": "Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders", "abs": "Sparse autoencoders (SAEs) are a promising unsupervised approach for identifying causally relevant and interpretable linear features in a language model's (LM) activations.\nTo be useful for downstream tasks, SAEs need to decompose LM activations faithfully; yet to be interpretable the decomposition must be sparse -- two objectives that are in tension.\nIn this paper, we introduce JumpReLU SAEs, which achieve state-of-the-art reconstruction fidelity at a given sparsity level on Gemma 2 9B activations, compared to other recent advances such as Gated and TopK SAEs.\nWe also show that this improvement does not come at the cost of interpretability through manual and automated interpretability studies.\nJumpReLU SAEs are a simple modification of vanilla (ReLU) SAEs -- where we replace the ReLU with a discontinuous JumpReLU activation function -- and are similarly efficient to train and run.\nBy utilising straight-through-estimators (STEs) in a principled manner, we show how it is possible to train JumpReLU SAEs effectively despite the discontinuous JumpReLU function introduced in the SAE's forward pass. Similarly, we use STEs to directly train L0 to be sparse, instead of training on proxies such as L1, avoiding problems like shrinkage.", "keywords": [ "Mechanistic interpretability", "interpretability", "dictionary learning" ], "rating_list": [ 3, 8, 3, 3 ], "soundness_list": [ 2, 4, 3, 1 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "mMPMHWOdOy", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Can Xu", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Chongyang Tao", "gender": "Male", "institution": "Beihang University", "country": "", "position": "Associate Professor" }, { "name": "Dongmei Zhang", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Distinguished Scientist" }, { "name": "Haipeng Luo", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jian-Guang Lou", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Pu Zhao", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Qingfeng Sun", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Qingwei Lin", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Sr. Principal Researcher" }, { "name": "Shifeng Chen", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Xiubo Geng", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Yansong Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 5, "n_ref_uni": 87, "n_ref": 273, "n_ref_all": 335, "n_fig": 5, "n_tab": 22, "L_tab": 5430, "n_element_tab": 436, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1160, "n_element_tab_1": 123, "formula_len_all": 1784, "formula_len_all_1": 460, "len_all": 459181, "len_all_1": 56466, "len_abs": 1993, "len_title": 101, "len_sents": 114835, "len_sents_1": 24411, "n_sents": 1024, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 100, "L_abs": 1025, "title": "WizardMath: Empowering Mathematical Reasoning for Large Language Models via Reinforced Evol-Instruct", "abs": "Large language models (LLMs), such as GPT-4, have shown remarkable performance in natural language processing (NLP) tasks, including challenging mathematical reasoning. However, most existing open-source models are only pre-trained on large-scale internet data and without math-related optimization. In this paper, we present WizardMath, which enhances the mathematical reasoning abilities of LLMs, by applying our proposed Reinforcement Learning from Evol-Instruct Feedback (RLEIF) method to the domain of math. Through extensive experiments on two mathematical reasoning benchmarks, namely GSM8k and MATH, we reveal the extraordinary capabilities of our model. Remarkably, WizardMath-Mistral 7B surpasses all other open-source LLMs by a substantial margin. Furthermore, WizardMath 70B even outperforms ChatGPT-3.5, Claude Instant, Gemini Pro and Mistral Medium. Additionally, our preliminary exploration highlights the pivotal role of instruction evolution and process supervision in achieving exceptional math performance.", "keywords": [ "Mathematical Reasoning", "Evol-Instruct", "Reinforcement Learning" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 4, 4, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 4, 4, 3, 4 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "mLztw5kEQ9", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Carmen Cheng", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Hanwang Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Zhongqi Yue", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 38, "n_ref": 73, "n_ref_all": 102, "n_fig": 9, "n_tab": 8, "L_tab": 2753, "n_element_tab": 257, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 2019, "n_element_tab_1": 141, "formula_len_all": 432, "formula_len_all_1": 399, "len_all": 135145, "len_all_1": 66594, "len_abs": 1281, "len_title": 95, "len_sents": 38529, "len_sents_1": 29373, "n_sents": 318, "n_sents_1": 237, "venue": "ICLR2025", "decision": "withdraw", "L_title": 46, "L_abs": 1494, "title": "Towards Debiased Source-Free Domain Adaptation", "abs": "Source-Free Domain Adaptation (SFDA) aims to adapt a model trained in an inaccessible source domain $S$ to a different, unlabelled target domain $T$. The conventional approach generates pseudo-labels for the $T$ samples with the source-trained model, which are then used for model adaptation. However, we show that the adapted model is biased to the spurious correlations in $S$, consequently leading to catastrophic failure on $T$ samples that are dissimilar to $S$. Unfortunately, without any prior knowledge about spurious correlations, the current SFDA setting has no mechanism to circumvent this bias. We introduce a practical setting to address this gap -- Debiased SFDA, where the model receives additional supervision from a pre-trained, frozen reference model. This setting stays in line with the essence of SFDA, which accommodates proprietary source-domain training, while also offering prior knowledge that is unaffected by source-domain training to facilitate debiasing. Under this setting, we propose 1) a simple contrastive objective that debiases the source-trained model from spurious correlations inconsistent with the reference model; 2) a diagnostic metric that evaluates the degree to which an adapted model is biased towards $S$. Our objective can be easily plugged into different baselines for debiasing, and through extensive evaluations, we demonstrate that it engenders consistent improvements across standard benchmarks. Code is supplied under supplementary material.", "keywords": [ "source-free domain adaptation", "sfda", "domain adaptation", "contrastive learning", "spurious correlation", "debiasing", "debiased sfda" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "mLyyB4le5u", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Can Chen", "gender": "Not Specified", "institution": "Mila - Quebec AI Institute", "country": "CA", "position": "PhD student" }, { "name": "Christopher J Pal", "gender": "unknown", "institution": "Polytechnique Montreal", "country": "CA", "position": "Full Professor" }, { "name": "Xue Liu", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Full Professor" }, { "name": "Ye Yuan", "gender": "Male", "institution": "McGill University, McGill University", "country": "CA", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 18, "n_ref_uni": 46, "n_ref": 81, "n_ref_all": 135, "n_fig": 11, "n_tab": 23, "L_tab": 25367, "n_element_tab": 2231, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2197, "n_element_tab_1": 102, "formula_len_all": 1188, "formula_len_all_1": 856, "len_all": 224173, "len_all_1": 76448, "len_abs": 1560, "len_title": 104, "len_sents": 48558, "len_sents_1": 31530, "n_sents": 414, "n_sents_1": 276, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1593, "title": "ParetoFlow: Guided Flows in Multi-Objective Optimization", "abs": "In offline multi-objective optimization (MOO), we leverage an offline dataset of designs and their associated labels to simultaneously minimize multiple objectives. This setting more closely mirrors complex real-world problems compared to single-objective optimization. Recent works mainly employ evolutionary algorithms and Bayesian optimization, with limited attention given to the generative modeling capabilities inherent in such data. In this study, we explore generative modeling in offline MOO through flow matching, noted for its effectiveness and efficiency. We introduce \\textit{ParetoFlow}, specifically designed to guide flow sampling to approximate the Pareto front. Traditional predictor~(classifier) guidance is inadequate for this purpose because it models only a single objective. In response, we propose a \\textit{multi-objective predictor guidance} module that \nassigns each sample a weight vector, representing a weighted distribution across multiple objective predictions. A local filtering scheme is introduced to address non-convex Pareto fronts. These weights uniformly cover the entire objective space, effectively directing sample generation towards the Pareto front. Since distributions with similar weights tend to generate similar samples, we introduce a \\textit{neighboring evolution} module to foster knowledge sharing among neighboring distributions. This module generates offspring from these distributions, and selects the most promising one for the next iteration. Our method achieves state-of-the-art performance across various tasks. Our code is available.", "keywords": [ "Multi-objective optimization; flow matching; classifier guidance." ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "mLxxv5gts0", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Jiawei WU", "gender": "unknown", "institution": "East China Normal University", "country": "", "position": "Undergrad student" }, { "name": "Mingyuan Yan", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Intern" }, { "name": "Rushi Shah", "gender": "Male", "institution": "Indian Institute of Technology, Jodhpur, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "Undergrad student" }, { "name": "Dianbo Liu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 26, "n_formula_1": 24, "n_ref_uni": 37, "n_ref": 74, "n_ref_all": 81, "n_fig": 3, "n_tab": 1, "L_tab": 398, "n_element_tab": 62, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 398, "n_element_tab_1": 62, "formula_len_all": 1075, "formula_len_all_1": 977, "len_all": 104573, "len_all_1": 57423, "len_abs": 1260, "len_title": 122, "len_sents": 29585, "len_sents_1": 25559, "n_sents": 215, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1276, "title": "Gaussian Mixture Vector Quantization with Aggregated Categorical Posterior", "abs": "The vector quantization is a widely used method to map continuous representation to discrete space and has important application in tokenization for generative mode, bottlenecking information and many other tasks in machine learning. \nVector Quantized Variational Autoencoder (VQ-VAE) is a type of variational autoencoder using discrete embedding as latent. We generalize the technique further, enriching the probabilistic framework with a Gaussian mixture as the underlying generative model. This framework leverages a codebook of latent means and adaptive variances to capture complex data distributions. This principled framework avoids various heuristics and strong assumptions that are needed with the VQ-VAE to address training instability and to improve codebook utilization. This approach integrates the benefits of both discrete and continuous representations within a variational Bayesian framework. Furthermore, by introducing the \\textit{Aggregated Categorical Posterior Evidence Lower Bound} (ALBO), we offer a principled alternative optimization objective that aligns variational distributions with the generative model. Our experiments demonstrate that GM-VQ improves codebook utilization and reduces information loss without relying on handcrafted heuristics.", "keywords": [ "Mixtures of Gaussians", "Vector Quantization", "Variational Auto-Encoders", "Representation Learning" ], "rating_list": [ 3, 5, 5, 5, 1 ], "soundness_list": [ 1, 2, 2, 2, 1 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 4, 3 ] }, { "paper_id": "mLTbDVzHVh", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Linh Ngo Van", "gender": "unknown", "institution": "Hanoi University of Science and Technology", "country": "VN", "position": "Assistant Professor" }, { "name": "Minh Le", "gender": "Male", "institution": "VinAI Research", "country": "VN", "position": "Researcher" }, { "name": "Nhat Ho", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Quyen Tran", "gender": "Female", "institution": "VinAI Research", "country": "VN", "position": "Research Resident" }, { "name": "Thien Huu Nguyen", "gender": "Male", "institution": "University of Oregon", "country": "US", "position": "Associate Professor" }, { "name": "Trung Le", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" }, { "name": "Tuan Truong", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Undergrad student" }, { "name": "Dinh Phung", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 28, "n_ref": 76, "n_ref_all": 101, "n_fig": 13, "n_tab": 3, "L_tab": 1021, "n_element_tab": 71, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 1188, "n_element_tab_1": 84, "formula_len_all": 573, "formula_len_all_1": 525, "len_all": 148637, "len_all_1": 64558, "len_abs": 1135, "len_title": 117, "len_sents": 52196, "len_sents_1": 31330, "n_sents": 276, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1059, "title": "Exploiting Hierarchical Taxonomies in Pretrained Continual Learning", "abs": "Drawing inspiration from human learning behaviors, this work proposes a novel approach to mitigate catastrophic forgetting in Prompt-based Continual Learning (PCL) models by exploiting the relationships between continuously emerging class data. We find that applying human habits of organizing and connecting information can serve as an efficient strategy when training deep learning models.\nSpecifically, by building a hierarchical tree structure based on the expanding set of labels, we gain fresh insights into the data, identifying groups of similar classes could easily cause confusion. Additionally, we delve deeper into the hidden connections between classes by exploring the original pretrained model\u2019s behavior through an optimal transport-based approach. From these insights, we propose\na novel regularization loss function that encourages models to focus more on challenging knowledge areas, thereby enhancing overall performance. Experimentally, our method demonstrated significant superiority over current state-of-the-arts on various benchmarks.", "keywords": [ "pretrained continual learning" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 4, 2, 2, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "mKM9uoKSBN", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Dana Weitzner", "gender": "Female", "institution": "Tel Aviv University", "country": "", "position": "PhD student" }, { "name": "Mauricio Delbracio", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Peyman Milanfar", "gender": "Male", "institution": "Google", "country": "US", "position": "Distinguished Scientist" }, { "name": "Raja Giryes", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Associate Professor" } ], "n_formula": 41, "n_formula_1": 34, "n_ref_uni": 42, "n_ref": 84, "n_ref_all": 123, "n_fig": 7, "n_tab": 1, "L_tab": 58, "n_element_tab": 20, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1892, "formula_len_all_1": 1609, "len_all": 136637, "len_all_1": 68260, "len_abs": 1444, "len_title": 108, "len_sents": 41261, "len_sents_1": 31067, "n_sents": 316, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1436, "title": "On the Relation Between Linear Diffusion and Power Iteration", "abs": "Recently, diffusion models have gained popularity due to their impressive generative abilities. These models learn the implicit distribution given by the training dataset, and sample new data by transforming random noise through the reverse process, which can be thought of as gradual denoising. In this work, we examine the generation process as a ``correlation machine'', where random noise is repeatedly enhanced in correlation with the implicit given distribution. \nTo this end, we explore the linear case, where the optimal denoiser is known to be the PCA projection. This enables us to connect the theory of diffusion models to the spiked covariance model, where the dependence of the denoiser on the noise level and the amount of training data can be expressed analytically, in the rank-1 case.\nIn a series of numerical experiments, we extend this result to general low rank data, and show that low frequencies emerge earlier in the generation process, where the denoising basis vectors are more aligned to the true data with a rate depending on their eigenvalues. This model allows us to show that the linear diffusion model converges in mean to the leading eigenvector of the underlying data, similarly to the prevalent Power Iteration method. \nFinally, we empirically demonstrate the applicability of our findings beyond the linear case, in the Jacobians of a deep, non-linear denoiser, used in general image generation tasks.", "keywords": [ "Diffusion models" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 1, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 1, 2 ], "confidence_list": [ 3, 5, 4 ] }, { "paper_id": "mKGXdsq7fD", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Chen Ma", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Kanglong FAN", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Kede Ma", "gender": "unknown", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "YIFAN PENG", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Yuming Fang", "gender": "Male", "institution": "Jiangxi University of Finance and Economics", "country": "CN", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 45, "n_ref": 105, "n_ref_all": 117, "n_fig": 5, "n_tab": 4, "L_tab": 1707, "n_element_tab": 247, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 571, "n_element_tab_1": 121, "formula_len_all": 189, "formula_len_all_1": 189, "len_all": 113295, "len_all_1": 58460, "len_abs": 1628, "len_title": 141, "len_sents": 29927, "len_sents_1": 27423, "n_sents": 191, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1621, "title": "Decision Rules are in the Pixels: Towards Pixel-level Evaluation of Saliency-based XAI Models", "abs": "The intricate and opaque nature of deep neural networks (DNNs) makes it difficult to decipher how they make decisions. Explainable artificial intelligence (XAI) has emerged as a promising remedy to this conundrum. However, verifying the correctness of XAI methods remains challenging, due to the absence of universally accepted ground-truth explanations. In this study, we focus on assessing the correctness of saliency-based XAI models applied to DNN-based image classifiers at the pixel level. The proposed evaluation protocol departs significantly from previous human-centric correctness assessment at the semantically meaningful object part level, which may not correspond to the actual decision rules derived by classifiers. A crucial step in our approach involves introducing a spatially localized shortcut, a form of decision rule that DNN-based classifiers tend to adopt preferentially, without disrupting original image patterns and decision rules therein. After verifying the shortcut as the dominant decision rule, we estimate the Shapley value for each pixel within the shortcut area to generate the ground-truth explanation map, assuming that pixels outside this area have null contributions. We quantitatively evaluate fourteen saliency-based XAI methods for classifiers utilizing convolutional neural networks and vision Transformers, trained on perturbed CIFAR-10, CIFAR-100, and ImageNet datasets, respectively. Comprehensive experimental results show that existing saliency-based XAI models struggle to offer accurate pixel-level attributions, casting doubt on the recent progress in saliency-based XAI.", "keywords": [ "post-hoc XAI", "evaluation", "pixel attributions", "shapley value", "feature deletion" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 3, 2, 4 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "mJzOHRSpSa", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "AmirMahdi Sadeghzadeh", "gender": "Male", "institution": "Sharif University of Technology", "country": "IR", "position": "Assistant Professor" }, { "name": "Hamid Dashtbani", "gender": "Male", "institution": "Sharif University of Technology", "country": "IR", "position": "MS student" }, { "name": "Mehdi Dousti Gandomani", "gender": "Male", "institution": "Sharif University of Technology", "country": "IR", "position": "MS student" }, { "name": "Rasool Jalili", "gender": "Male", "institution": "Sharif University of Technology", "country": "IR", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 7, "n_ref_uni": 35, "n_ref": 120, "n_ref_all": 166, "n_fig": 8, "n_tab": 11, "L_tab": 5986, "n_element_tab": 412, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 2711, "n_element_tab_1": 162, "formula_len_all": 521, "formula_len_all_1": 240, "len_all": 143201, "len_all_1": 63735, "len_abs": 1699, "len_title": 159, "len_sents": 46213, "len_sents_1": 29490, "n_sents": 303, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 110, "L_abs": 1710, "title": "Random Logit Scaling: Defending Deep Neural Networks Against Black-Box Score-Based Adversarial Example Attacks", "abs": "Machine learning models are increasingly adapted in various domains, such as autonomous driving, facial recognition, and malware detection, achieving state-of-the-art results. However, adversarial example attacks pose a significant threat to the reliable deployment of machine learning models in such applications. In recent years, some powerful adversarial example attacks have been proposed for the fast and query-efficient generation of adversarial examples, even in black-box scenarios where attackers only have an oracle access to the target model, highlighting the need for scalable, low-cost, and powerful defenses. In this work, we present two contributions to the domain of black-box attacks and defenses. First, we propose Random Logit Scaling (RLS), a randomization-based defense against black-box score-based adversarial example attacks. RLS is a plug-and-play, post-processing defense that can be implemented on top of any existing ML model with minimal effort. The idea behind RLS is to confuse an attacker by outputting falsified scores resulting from randomly scaled logits while maintaining the model accuracy. We show that RLS significantly reduces the success rate of state-of-the-art black-box score-based attacks while preserving the accuracy and minimizing confidence score distortion compared to state-of-the-art randomization-based defenses. Second, we introduce a novel adaptive attack against AAA, a SOTA non-randomized black-box defense against black-box score-based attacks that also modifies output logits to confuse attackers. With this adaptive attack, we demonstrate the vulnerability of AAA, establishing RLS as the effective SOTA defense against black-box score-based attacks.", "keywords": [ "Adversarial examples", "Query-based black-box attacks", "Randomized defenses" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "mJwpHp8guj", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Fran\u00e7ois Rousset", "gender": "unknown", "institution": "Institut National des Sciences Appliqu\u00e9es de Lyon", "country": "FR", "position": "Assistant Professor" }, { "name": "Marc CLAUSSE", "gender": "Male", "institution": "Institut National des Sciences Appliqu\u00e9es de Lyon", "country": "FR", "position": "Full Professor" }, { "name": "Taha Boussaid", "gender": "Male", "institution": "Institut National des Sciences Appliqu\u00e9es de Lyon", "country": "FR", "position": "PhD student" }, { "name": "Vasile-Marian Scuturici", "gender": "unknown", "institution": "Institut National des Sciences Appliqu\u00e9es de Lyon", "country": "FR", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 48, "n_ref": 63, "n_ref_all": 91, "n_fig": 11, "n_tab": 5, "L_tab": 1273, "n_element_tab": 158, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 766, "n_element_tab_1": 94, "formula_len_all": 1058, "formula_len_all_1": 939, "len_all": 134872, "len_all_1": 67513, "len_abs": 1314, "len_title": 146, "len_sents": 39345, "len_sents_1": 30840, "n_sents": 291, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1321, "title": "Unlocking Full Dynamic Optimization of District Energy Systems through State-Space Model Learning", "abs": "Predictive control enables the operation of physical systems along an optimal trajectory based on forecasts and dynamic simulations. However, the complexity of system dynamics and high computational cost of optimization typically restrict the optimization window to short horizons. Thus, any potential benefits from mid- and long-term rewards are withdrawn. This is particularly relevant for optimization of district energy systems using various low-environmental-impact sources. To address this, we present an end-to-end methodological framework for learning state-space representations of such systems that significantly reduce computational load. The proposed approach leverages the implicit graph structure of such systems to develop and train a physics-informed spatio-temporal graph neural network. This methodology is evaluated on a real-world district heating system incorporating thermal solar panels, storage, biomass and natural gas boilers. Through historical time-series data augmentation and hyperparameter optimization, the learned model demonstrates strong generalization ability and high accuracy in predicting system dynamics. Our method reduces simulation time by four orders of magnitude, cutting optimization time from several days to mere minutes, while also lowering operational costs by up to 25%.", "keywords": [ "surrogate model", "state-space", "physics-informed models", "optimal control", "sustainable energy systems" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "mJKhn7Ey4y", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Brian M. Sadler", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Researcher" }, { "name": "Derrik E. Asher", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Mubarak Shah", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Souradip Chakraborty", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Utsav Singh", "gender": "Male", "institution": "Indian Institute of Technology, Kanpur", "country": "IN", "position": "PhD student" }, { "name": "Vinay P. Namboodiri", "gender": "Male", "institution": "University of Bath", "country": "GB", "position": "Associate Professor" }, { "name": "Amrit Bedi", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Wesley Suttle", "gender": "unknown", "institution": "Army Research Laboratory", "country": "US", "position": "Postdoc" } ], "n_formula": 38, "n_formula_1": 26, "n_ref_uni": 33, "n_ref": 79, "n_ref_all": 95, "n_fig": 7, "n_tab": 1, "L_tab": 1083, "n_element_tab": 28, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3771, "formula_len_all_1": 2133, "len_all": 137656, "len_all_1": 71839, "len_abs": 1427, "len_title": 145, "len_sents": 45809, "len_sents_1": 32951, "n_sents": 329, "n_sents_1": 226, "venue": "ICLR2025", "decision": "withdraw", "L_title": 96, "L_abs": 1434, "title": "DIPPER: Direct Preference Optimization for Primitive-Enabled Hierarchical Reinforcement Learning", "abs": "Hierarchical reinforcement learning (HRL) is an elegant framework for learning efficient control policies to perform complex robotic tasks, especially in sparse reward settings. However, concurrently learning policies at multiple hierarchical levels often suffers from training instability due to non-stationary behavior of lower-level primitives. In this work, we introduce DIPPER, an efficient hierarchical framework that leverages Direct Preference Optimization (DPO) to mitigate non-stationarity at the higher level, while using reinforcement learning to train the corresponding primitives at the lower level. We observe that directly applying DPO to the higher level in HRL is ineffective and leads to infeasible subgoal generation issues. To address this, we develop a novel, principled framework based on lower-level primitive regularization of upper-level policy learning. We provide a theoretical justification for the proposed framework utilizing bi-level optimization. The application of DPO also necessitates the development of a novel reference policy formulation for feasible subgoal generation. To validate our approach, we conduct extensive experimental analyses on a variety of challenging, sparse-reward robotic navigation and manipulation tasks. Our results demonstrate that DIPPER shows impressive performance and demonstrates an improvement of up to 40% over the baselines in complex sparse robotic control tasks.", "keywords": [ "hierarchical reinforcement learning", "preference learning" ], "rating_list": [ 1, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "mJ8k81O5BF", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 3, "author_info_list": [ { "name": "Akash Chavan", "gender": "Male", "institution": "Oakland University", "country": "US", "position": "PhD student" }, { "name": "Sumit Kumar Jha", "gender": "unknown", "institution": "Florida International University", "country": "US", "position": "Full Professor" }, { "name": "Sunny Raj", "gender": "Male", "institution": "Oakland University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 46, "n_ref": 77, "n_ref_all": 82, "n_fig": 1, "n_tab": 5, "L_tab": 1786, "n_element_tab": 533, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 2275, "n_element_tab_1": 538, "formula_len_all": 2, "formula_len_all_1": 2, "len_all": 114533, "len_all_1": 61464, "len_abs": 1396, "len_title": 115, "len_sents": 28527, "len_sents_1": 26012, "n_sents": 192, "n_sents_1": 148, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1406, "title": "IMPROVING LOW-BIT POST TRAINING QUANTIZATION: A DATA-FREE APPROACH", "abs": "Post-training quantization (PTQ) without access to real data is enabling efficient model optimization and deployment in scenarios where privacy or proprietary constraints restrict the use of original datasets. Traditional data free quantization methods rely on Batch Normalization (BN) statistics from the trained full-precision model to generate calibration dataset for quantization. However, this reliance on BN statistics limits their applicability to deep neural networks (DNNs) without BN layer such as AlexNet. In this paper, we propose a calibration dataset generation algorithm that is agnostic to BN statistics, leveraging just the backpropagation to create synthetic images for PTQ. We also demonstrate that it is not necessary to include samples from every target category in the calibration dataset to get the representative activation ranges for quantization. Extensive experiments with both large and lightweight models on large-scale image classification tasks demonstrate that our method consistently improves quantization performance across various DNN architectures, especially in low-bit settings. Notably, in 4-bit quantization, we achieve an improvement of 3.42\\% in top-1 accuracy for the ResNet18 model and 3.14\\% for the InceptionV3 model compared to the state-of-the-art (SOTA) DSG method. Importantly, we use very few synthetic samples for quantization compared to other methods.", "keywords": [ "Deep Neural Network", "Quantization Aware Training", "Post Training Quantization", "Data Free Quantization" ], "rating_list": [ 3, 5, 3, 1 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "mJ4mgYjDru", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Davi Moreno", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Eric Jahns", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Michel Kinsy", "gender": "unknown", "institution": "Trustees of Boston University", "country": "", "position": "" }, { "name": "Milan Stojkov", "gender": "unknown", "institution": "University of Novi Sad", "country": "RS", "position": "Assistant Professor" } ], "n_formula": 34, "n_formula_1": 18, "n_ref_uni": 37, "n_ref": 89, "n_ref_all": 119, "n_fig": 7, "n_tab": 10, "L_tab": 1533, "n_element_tab": 128, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 1318, "n_element_tab_1": 112, "formula_len_all": 2747, "formula_len_all_1": 948, "len_all": 173769, "len_all_1": 55515, "len_abs": 1254, "len_title": 101, "len_sents": 47099, "len_sents_1": 25004, "n_sents": 353, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1267, "title": "Discretized Quadratic Integrate-and-Fire Neuron Model for Direct Training of Spiking Neural Networks", "abs": "Spiking Neural Networks (SNNs) are a promising alternative to traditional artificial neural networks, offering significant energy-saving potential. Conventional SNN approaches typically utilize the Leaky Integrate-and-Fire (LIF) neuron model, where voltage decays linearly, decreasing proportionally to its current value. However, this linear decay can inadvertently increase energy consumption and reduce model performance due to extraneous spiking activity. To address these limitations, we introduce the discretized Quadratic Integrate-and-Fire (QIF) neuron model, which applies a non-linear transformation to the voltage proportional to its magnitude. The QIF neuron model achieves substantial energy reductions, ranging from $1.43 - 4.21\\times$ compared to the LIF neuron model. On static datasets (CIFAR-10, CIFAR-100) and neuromorphic datasets (CIFAR-10 DVS, N-Caltech-101, N-Cars, DVS128-Gesture), the QIF neuron model demonstrates competitive performance and improved accuracy over state-of-the-art results. Furthermore, the QIF neuron model produces smoother loss landscapes and larger local minima, leading to faster training convergence. Our findings suggest that the QIF neuron model offers a promising alternative to the widely adopted LIF neuron model.", "keywords": [ "Machine Learning", "Neuromorphic Computing", "Deep Learning", "Computer Vision" ], "rating_list": [ 6, 6, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 5, 5, 4 ] }, { "paper_id": "mIl15VP7vt", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Percy Liang", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Sang T. Truong", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Yuheng Tu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Intern" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 9, "n_ref_uni": 26, "n_ref": 42, "n_ref_all": 79, "n_fig": 19, "n_tab": 9, "L_tab": 1608, "n_element_tab": 171, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2101, "formula_len_all_1": 714, "len_all": 251406, "len_all_1": 68298, "len_abs": 809, "len_title": 55, "len_sents": 98805, "len_sents_1": 35800, "n_sents": 1001, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 811, "title": "Reliable and Efficient Amortized Model-based Evaluation", "abs": "Current generative model evaluation procedures are costly and sensitive to test set selection, making continuous monitoring impractical. In this paper, we employ a model-based evaluation framework using Item Response Theory (IRT), which decouples model performance from the test subset selection, ensuring reliable and efficient evaluation. We propose two innovations: amortized calibration to reduce the cost of estimating item parameters of the IRT model and an item generator based on a large language model to automate diverse question generation. Our experiments on 25 common natural language processing benchmarks and 184 language models show that this approach is more reliable and resource-efficient compared to traditional evaluation methods, offering a scalable solution to evaluate generative models.", "keywords": [ "Model Evaluation", "Amortization", "Adaptive Testing" ], "rating_list": [ 5, 6, 8, 6, 8 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 4, 3 ], "confidence_list": [ 3, 2, 3, 4, 3 ] }, { "paper_id": "mIjblC9hfm", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Chengsong Huang", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "", "position": "PhD student" }, { "name": "Hao Liu", "gender": "Female", "institution": "Washington University in St. Louis", "country": "US", "position": "PhD student" }, { "name": "Jiarui Feng", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "US", "position": "PhD student" }, { "name": "Jiaxin Huang", "gender": "Female", "institution": "Washington University, Saint Louis", "country": "", "position": "Assistant Professor" }, { "name": "Lecheng Kong", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "US", "position": "PhD student" }, { "name": "Muhan Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yixin Chen", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 2, "n_ref_uni": 51, "n_ref": 166, "n_ref_all": 199, "n_fig": 7, "n_tab": 13, "L_tab": 6828, "n_element_tab": 434, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2718, "n_element_tab_1": 123, "formula_len_all": 500, "formula_len_all_1": 137, "len_all": 298782, "len_all_1": 72823, "len_abs": 4417, "len_title": 118, "len_sents": 85149, "len_sents_1": 32289, "n_sents": 740, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1608, "title": "GOFA: A Generative One-For-All Model for Joint Graph Language Modeling", "abs": "Foundation models, such as Large Language Models (LLMs) or Large Vision Models (LVMs), have emerged as one of the most powerful tools in the respective fields. However, unlike text and image data, graph data do not have a definitive structure, posing great challenges to developing a Graph Foundation Model (GFM). For example, current attempts at designing general graph models either transform graph data into a language format for LLM-based prediction or still train a GNN model with LLM as an assistant. The former can handle unlimited tasks, while the latter captures graph structure much better---yet, no existing work can achieve both simultaneously. In this paper, we first identify three key desirable properties of a GFM: self-supervised pretraining, fluidity in tasks, and graph awareness. To account for these properties, we extend the conventional language modeling to the graph domain and propose a novel generative graph language model GOFA. The model interleaves randomly initialized GNN layers into a frozen pre-trained LLM so that the semantic and structural modeling abilities are organically combined. GOFA is pre-trained on newly proposed graph-level next-word prediction, question-answering, structural understanding, and information retrieval tasks to obtain the above GFM properties. The pre-trained model is further instruction fine-tuned to obtain the task-solving ability. Our GOFA model is evaluated on various downstream tasks unseen during the pre-training and fine-tuning phases, demonstrating a strong ability to solve structural and contextual problems in zero-shot scenarios.", "keywords": [ "GNN;Graph foundation model;LLM;" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 2, 4, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "mHx8JFURtn", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Anastasiia Gavrish", "gender": "Female", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "MS student" }, { "name": "Ivan Oseledets", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "Nechesov Andrey", "gender": "Male", "institution": "Novosibirsk State University", "country": "RU", "position": "Associate Professor" }, { "name": "Roman Schutski", "gender": "Male", "institution": "Novosibirsk State University", "country": "RU", "position": "Researcher" } ], "n_formula": 20, "n_formula_1": 18, "n_ref_uni": 22, "n_ref": 40, "n_ref_all": 47, "n_fig": 3, "n_tab": 1, "L_tab": 24, "n_element_tab": 1, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 537, "n_element_tab_1": 28, "formula_len_all": 1677, "formula_len_all_1": 1508, "len_all": 100670, "len_all_1": 55642, "len_abs": 1465, "len_title": 152, "len_sents": 30846, "len_sents_1": 24786, "n_sents": 259, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1470, "title": "Rethinking logic in AI: A novel benchmark inspired by polynomial analogue of Gandy's fixed point theorem", "abs": "This paper introduces a novel benchmark for evaluating the logical reasoning capabilities of Large Language Models (LLMs), grounded in the polynomial analogue of Gandy's classical fixed point theorem. Since this theorem can be used to describe the P-complete HornSAT problem, and our benchmark is based on this theorem, our benchmark thus covers all problems from class P and shows that serious problems have already arisen in this class, not to mention those benchmarks whose complexity classes are NP-complete and NP-hard. Drawing on concepts from mathematical logic, we design a parameterized set of recursively definable problems where the objective is for LLMs to predict whether a problem belongs to an inductively definable set of polynomial complexity. By varying the parameters, we generate problem instances of differing complexity. Our experiments reveal that current state-of-the-art LLMs with zero-shots promts fail to reliably solve even the most straightforward cases despite an effective deterministic algorithm existing. Even advanced models like GPT-4 exhibit significant biases in solving benchmark problems. These findings highlight the limitations of modern LLMs as code interpreters, even in basic scenarios, and underscore the necessity for hybrid LLM/interpreter systems. Furthermore, they emphasize the importance of developing quantitative tests for reasoning, given the increasing reliance on LLM-based systems in decision-making applications.", "keywords": [ "llm", "logic", "benchmark", "Gandy's fixed point theorem" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "mHkbi3XM58", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 1, "author_info_list": [ { "name": "Pierre-\u00c9tienne H Fiquet", "gender": "Male", "institution": "Flatiron Institute", "country": "US", "position": "Postdoc" } ], "n_formula": 56, "n_formula_1": 38, "n_ref_uni": 31, "n_ref": 40, "n_ref_all": 65, "n_fig": 11, "n_tab": 1, "L_tab": 33, "n_element_tab": 6, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 26266, "formula_len_all_1": 20827, "len_all": 181155, "len_all_1": 106606, "len_abs": 1150, "len_title": 113, "len_sents": 63446, "len_sents_1": 45224, "n_sents": 463, "n_sents_1": 310, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1109, "title": "Conditional density estimation for video prediction with score-based models", "abs": "Temporal prediction is inherently uncertain, but representing the ambiguity in natural image sequences is a challenging high-dimensional probabilistic inference problem. For natural scenes, the curse of dimensionality renders explicit density estimation statistically and computationally intractable. Here, we describe an implicit regression-based framework for learning and sampling the conditional density of the next frame in a video given previous observed frames. We show that sequence-to-image deep networks trained on a simple resilience-to-noise objective function extract adaptive representations for temporal prediction. Synthetic experiments demonstrate that this score-based framework can handle occlusion boundaries: unlike classical methods that average over bifurcating temporal trajectories, it chooses among likely trajectories, selecting more probable options with higher frequency. Furthermore, analysis of networks trained on natural image sequences reveals that the representation automatically weights predictive evidence by its reliability, which is a hallmark of statistical inference.", "keywords": [ "temporal prediction", "diffusion models", "denoising", "conditioning", "occlusion boundaries", "cue combination" ], "rating_list": [ 1, 1, 5, 6 ], "soundness_list": [ 4, 1, 3, 3 ], "presentation_list": [ 4, 1, 2, 3 ], "contribution_list": [ 1, 1, 2, 4 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "mH2tKj7KR6", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Bo Jiang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Cheng Chen", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Ruofeng Yang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Shuai Li", "gender": "Female", "institution": "John Hopcroft Center, Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 91, "n_formula_1": 37, "n_ref_uni": 30, "n_ref": 104, "n_ref_all": 129, "n_fig": 1, "n_tab": 4, "L_tab": 455, "n_element_tab": 68, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 73, "n_element_tab_1": 4, "formula_len_all": 12336, "formula_len_all_1": 3464, "len_all": 155492, "len_all_1": 74300, "len_abs": 1603, "len_title": 150, "len_sents": 44419, "len_sents_1": 30865, "n_sents": 409, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1615, "title": "The Discretization Complexity Analysis of Consistency Models under Variance Exploding Forward Process", "abs": "Consistency models, a new class of one-step generative models, have shown state-of-the-art performance in one-step generation and achieve competitive performance compared to multi-step diffusion models. The most challenging part of consistency models is the training process, which discretizes the diffusion process and trains a consistency function to map any point at any discretized timepoint of the diffusion process to the data distribution. Despite the empirical success, only a few works focus on the discretization complexity of consistency models. However, the setting of those works is far away from the empirical consistency models with good performance, suffers from large discretization complexity, and fails to explain the empirical success of consistency models. To bridge the gap between theory and application, we analyze consistency models with two key properties: (1) variance exploding forward process and (2) gradually decay discretization stepsize, which are both widely used in empirical consistency models. Under the above realistic setting, we make the first step to explain the empirical success of consistency models and achieve the state-of-the-art discretization complexity for consistency models, which is competitive with the results of diffusion models. After obtaining the results of the one-step sampling method of consistency models, we further analyze a multi-step consistency sampling algorithm proposed by \\citet{song2023consistency} and show that this algorithm improves the discretization complexity compared with one-step generation, which matches the empirical observation.", "keywords": [ "Discretization Complexity", "Consistency Model" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 2, 3, 1, 4 ] }, { "paper_id": "mGSQLuYxVF", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Ruijian Han", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "Yeheng Ge", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Postdoc" }, { "name": "Yuanhang Luo", "gender": "unknown", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "GUOHAO SHEN", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 79, "n_formula_1": 25, "n_ref_uni": 44, "n_ref": 89, "n_ref_all": 95, "n_fig": 1, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8635, "formula_len_all_1": 1819, "len_all": 159712, "len_all_1": 75264, "len_abs": 943, "len_title": 85, "len_sents": 47811, "len_sents_1": 31354, "n_sents": 383, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 952, "title": "Learning guarantee of reward modeling using deep neural networks", "abs": "In this work, we study the learning theory of reward modeling using pairwise comparison data and deep neural networks. We establish a novel non-asymptotic regret bound for deep reward estimators in a non-parametric setting, which depends explicitly on the network architecture. Furthermore, to underscore the critical importance of clear human beliefs, we introduce a margin-type condition requiring the conditional winning probability of the optimal action in pairwise comparisons to be significantly distanced from 1/2. This condition enables a sharper regret bound, which substantiates the empirical efficiency in Reinforcement Learning from Human Feedback (RLHF) and highlights the role of clear human beliefs in its success. Notably, this improvement stems from high-quality pairwise comparison data under the margin-type condition and is independent of the specific estimators used, making it applicable to various learning algorithms and models.", "keywords": [ "Reinforcement Learning with Human Feedback", "Reward Modeling", "Deep Neural Networks", "Learning Guarantee", "Clear Human Beliefs" ], "rating_list": [ 5, 5, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 2, 4 ], "presentation_list": [ 2, 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 3, 2, 2, 3, 4 ] }, { "paper_id": "mFzpBaTLGK", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Eyal Ronen", "gender": "unknown", "institution": "Tel Aviv University", "country": "", "position": "Assistant Professor" }, { "name": "Junyi Ao", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Mahmood Sharif", "gender": "Male", "institution": "Tel Aviv University", "country": "", "position": "Assistant Professor" }, { "name": "Tom Ko", "gender": "Male", "institution": "ByteDance AI Lab", "country": "HK", "position": "Researcher" }, { "name": "Zebin Yun", "gender": "Male", "institution": "School of Computer Science, Tel Aviv University", "country": "IL", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 56, "n_ref": 109, "n_ref_all": 118, "n_fig": 2, "n_tab": 6, "L_tab": 1246, "n_element_tab": 271, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 261, "n_element_tab_1": 27, "formula_len_all": 284, "formula_len_all_1": 186, "len_all": 159121, "len_all_1": 68114, "len_abs": 1343, "len_title": 143, "len_sents": 44162, "len_sents_1": 34888, "n_sents": 303, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1406, "title": "Sounding the Alarm: Backdooring Acoustic Foundation Models for Physically Realizable Triggers", "abs": "Although foundation models help increase performance on many downstream tasks while reducing the amount of labeled data needed, \ntheir proliferation has raised a natural question: To what extent can a model downloaded from the Internet be trusted? We tackle this question for acoustic foundation models (AFMs) and propose the $\\textbf F$oundation $\\textbf A$coustic model $\\textbf B$ackdoor (FAB) attack against AFMs, showing that state-of-the-art models are susceptible to a new attack vector. Despite preserving model performance on benign data, AFM induces backdoors that survive fine-tuning, and, when activated, lead to a significant performance drop on various downstream tasks. Notably, backdoors created by FAB can be activated in a ${physically\\ realizable}$ manner by ${inconspicuous}$, ${input}$-${agnostic}$ triggers that ${do\\ not\\ require\\ syncing}$ with the acoustic input (e.g., by playing a siren sound in the background). Crucially, FAB also assumes a weaker threat model than past work, where the adversary has no knowledge of the pre-training data and certain architectural details. We tested FAB with two leading AFMs, on nine tasks, with four triggers, against two defenses, as well as in the digital and physical domains, and found the attack highly successful in all scenarios. Overall, our work highlights the risks facing AFMs and calls for advanced defences to mitigate them.", "keywords": [ "security", "speech model", "backdoor attack" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "mFiGAbvmYS", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Chenxiao Yang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Kaipeng Zeng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Michael M. Bronstein", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Qitian Wu", "gender": "Male", "institution": "Broad Institute of MIT and Harvard", "country": "US", "position": "Researcher" } ], "n_formula": 80, "n_formula_1": 20, "n_ref_uni": 48, "n_ref": 89, "n_ref_all": 109, "n_fig": 6, "n_tab": 11, "L_tab": 3815, "n_element_tab": 266, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 3215, "n_element_tab_1": 164, "formula_len_all": 6501, "formula_len_all_1": 1564, "len_all": 209871, "len_all_1": 75191, "len_abs": 1280, "len_title": 124, "len_sents": 65979, "len_sents_1": 33771, "n_sents": 524, "n_sents_1": 238, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1286, "title": "Towards Generalization under Topological Shifts: A Diffusion PDE Perspective", "abs": "The capability of generalization is a cornerstone for the success of modern learning systems. For non-Euclidean data that particularly involves topological features, one important aspect neglected by prior studies is how learning-based models generalize under topological shifts. This paper makes steps towards understanding the generalization of graph neural networks operated on varying topologies through the lens of diffusion PDEs. Our analysis first reveals that the upper bound of the generalization error yielded by local diffusion equation models, which are intimately related to message passing over observed structures, would exponentially grow w.r.t. topological shifts. In contrast, extending the diffusion operator to a non-local counterpart that learns latent structures from data can in principle control the generalization error under topological shifts even when the model accommodates observed structures. On top of these results, we propose Advective Diffusion Transformer inspired by advective diffusion equations serving as a physics-inspired continuous model that synthesizes observed and latent structures for graph learning. The model demonstrates superiority in various downstream tasks across information networks, molecular screening and protein interactions.", "keywords": [ "Topological Shifts", "Diffusion Equations", "Transformers" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 1 ] }, { "paper_id": "mFY0tPDWK8", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Fangzhou Zhu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Feng Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Haoyang Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xijun Li", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuxuan Zong", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Zijie Geng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 47, "n_formula_1": 18, "n_ref_uni": 43, "n_ref": 111, "n_ref_all": 154, "n_fig": 5, "n_tab": 13, "L_tab": 4942, "n_element_tab": 438, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 693, "n_element_tab_1": 122, "formula_len_all": 4138, "formula_len_all_1": 1066, "len_all": 206164, "len_all_1": 67814, "len_abs": 1446, "len_title": 159, "len_sents": 54387, "len_sents_1": 30282, "n_sents": 476, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 111, "L_abs": 1452, "title": "Apollo-MILP: An Alternating Prediction-Correction Neural Solving Framework for Mixed-Integer Linear Programming", "abs": "Leveraging machine learning (ML) to predict an initial solution for mixed-integer linear programming (MILP) has gained considerable popularity in recent years. These methods predict a solution and fix a subset of variables to reduce the problem dimension. Then, they solve the reduced problem to obtain the final solutions. However, directly fixing variable values can lead to low-quality solutions or even infeasible reduced problems if the predicted solution is not accurate enough. To address this challenge, we propose an Alternating prediction-correction neural solving framework (Apollo-MILP) that can identify and select accurate and reliable predicted values to fix. In each iteration, Apollo-MILP conducts a prediction step for the unfixed variables, followed by a correction step to obtain an improved solution (called reference solution) through a trust-region search. By incorporating the predicted and reference solutions, we introduce a novel Uncertainty-based Error upper BOund (UEBO) to evaluate the uncertainty of the predicted values and fix those with high confidence. A notable feature of Apollo-MILP is the superior ability for problem reduction while preserving optimality, leading to high-quality final solutions. Experiments on commonly used benchmarks demonstrate that our proposed Apollo-MILP significantly outperforms other ML-based approaches in terms of solution quality, achieving over a 50% reduction in the solution gap.", "keywords": [ "Mixed-integer Linear Programming", "Learning to Optimize" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 3, 3, 5 ] }, { "paper_id": "mFHPoYVeqN", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Dabao Zhang", "gender": "Male", "institution": "University of California, Irvine", "country": "", "position": "Full Professor" }, { "name": "Min Zhang", "gender": "Female", "institution": "University of California, Irvine", "country": "US", "position": "Full Professor" }, { "name": "Zhongli Jiang", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Postdoc" } ], "n_formula": 56, "n_formula_1": 40, "n_ref_uni": 25, "n_ref": 51, "n_ref_all": 79, "n_fig": 16, "n_tab": 1, "L_tab": 110, "n_element_tab": 8, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7014, "formula_len_all_1": 4444, "len_all": 108145, "len_all_1": 54677, "len_abs": 877, "len_title": 112, "len_sents": 31158, "len_sents_1": 21664, "n_sents": 256, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 880, "title": "Feature-Specific Coefficients of Determination in Tree Ensembles", "abs": "Tree ensemble methods provide promising predictions with models difficult to interpret. Recent introduction of Shapley values for individualized feature contributions, accompanied with several fast computing algorithms for predicted values, shows intriguing results. However, individualizing coefficients of determination, aka $R^2$, for each feature is challenged by the underlying quadratic losses, although these coefficients allow us to comparatively assess single feature's contribution to tree ensembles. Here we propose an efficient algorithm, Q-SHAP, that reduces the computational complexity to polynomial time when calculating Shapley values related to quadratic losses. Our extensive simulation studies demonstrate that this approach not only enhances computational efficiency but also improves the estimation accuracy of feature-specific coefficients of determination.", "keywords": [ "Explainable AI", "Feature importance", "Shapley Value", "Tree ensemble" ], "rating_list": [ 8, 3, 6, 3, 8 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "mFCLLUtm83", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Hehe Fan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xuanmeng Zhang", "gender": "unknown", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Yi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 53, "n_ref": 144, "n_ref_all": 160, "n_fig": 3, "n_tab": 1, "L_tab": 547, "n_element_tab": 67, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1243, "n_element_tab_1": 13, "formula_len_all": 681, "formula_len_all_1": 714, "len_all": 147798, "len_all_1": 53859, "len_abs": 3693, "len_title": 114, "len_sents": 39716, "len_sents_1": 25482, "n_sents": 286, "n_sents_1": 170, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1354, "title": "Instax3D: Creating 3D Portrait from a single-view image in Minutes", "abs": "We study single-view 3D portrait creation, specifically producing a full-head 3D\nportrait from a single headshot. This problem faces two challenges: 1) the 2D\nimage-based personalization methods lack comprehensive 3D awareness due to\nthe scarcity of multi-view 2D images or 3D assets in the training data, and 2) the\nscore distillation sampling optimization methods usually take hours to produce a\nsingle 3D asset, making the process quite time-consuming. To overcome these\nlimitations, we propose Instax3D, a generative Gaussian Splatting model with a\nvideo diffusion prior for rapid 3D portrait creation. We formulate the 3D portrait\ncreation problem as a \u201cgeneration and construction\u201d process. Specifically, Instax3D\nfirst synthesizes a consecutive video sequence using a finetuned video diffusion\nmodel, capitalizing on inherent diversity and multi-view knowledge from\nthe massive video data. Subsequently, Instax3D reconstructs the 3D portrait with\na multi-view FLAME-based Gaussian splatting representation from the generated\nvideo frames, structurally guided by an expressive 3D parametric model. Notably,\ngiven a reference headshot image, Instax3D can generate a 3D portrait in just 10\nminutes and render it at 40 FPS. This represents a 10\u00d7 improvement over previous\nmainstream optimization-based methods, which can take between one to two\nhours.", "keywords": [ "3D portrait", "single-view reconstruction" ], "rating_list": [ 3, 3, 3, 3, 8 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 1, 2, 1, 2, 4 ], "confidence_list": [ 5, 4, 5, 4, 4 ] }, { "paper_id": "mEpqHvbD2h", "primary_area": "reinforcement learning", "n_author": 9, "author_info_list": [ { "name": "Allen Z. Ren", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Anirudha Majumdar", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Anthony Simeonov", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Benjamin Burchfiel", "gender": "Male", "institution": "Dexterous Manipulation Group, Toyota Research Institute", "country": "", "position": "Researcher" }, { "name": "Hongkai Dai", "gender": "unknown", "institution": "Toyota Research Institute", "country": "", "position": "Researcher" }, { "name": "Justin Lidard", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Lars Lien Ankile", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Max Simchowitz", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Pulkit Agrawal", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 38, "n_formula_1": 11, "n_ref_uni": 82, "n_ref": 184, "n_ref_all": 261, "n_fig": 16, "n_tab": 7, "L_tab": 491, "n_element_tab": 98, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2811, "formula_len_all_1": 1172, "len_all": 304400, "len_all_1": 59525, "len_abs": 1387, "len_title": 46, "len_sents": 93384, "len_sents_1": 27559, "n_sents": 749, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 36, "L_abs": 1237, "title": "Diffusion Policy Policy Optimization", "abs": "We introduce Diffusion Policy Policy Optimization, DPPO, an algorithmic framework including best practices for fine-tuning diffusion-based policies (e.g. Diffusion Policy) in continuous control and robot learning tasks using the policy gradient (PG) method from reinforcement learning (RL). PG methods are ubiquitous in training RL policies with other policy parameterizations; nevertheless, they had been conjectured to be less efficient for diffusion-based policies. Surprisingly, we show that DPPO achieves the strongest overall performance and efficiency for fine-tuning in common benchmarks compared to other RL methods for diffusion-based policies and also compared to PG fine-tuning of other policy parameterizations. Through experimental investigation, we find that DPPO takes advantage of unique synergies between RL fine-tuning and the diffusion parameterization, leading to structured and on-manifold exploration, stable training, and strong policy robustness. We further demonstrate the strengths of DPPO in a range of realistic settings, including simulated robotic tasks with pixel observations, and via zero-shot deployment of simulation-trained policies on robot hardware in a long-horizon, multi-stage manipulation task.", "keywords": [ "reinforcement learning", "diffusion policy" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 1, 3, 3, 4 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "mEBSeSk49H", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Bohan Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Qi Meng", "gender": "Female", "institution": "Microsoft", "country": "", "position": "associate researcher" }, { "name": "Ruoyu Sun", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Wei Chen", "gender": "Female", "institution": " Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhi-Ming Ma", "gender": "unknown", "institution": "Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Huishuai Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 225, "n_formula_1": 28, "n_ref_uni": 38, "n_ref": 104, "n_ref_all": 132, "n_fig": 0, "n_tab": 1, "L_tab": 241, "n_element_tab": 9, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 2935, "n_element_tab_1": 33, "formula_len_all": 45367, "formula_len_all_1": 1522, "len_all": 243226, "len_all_1": 73083, "len_abs": 1323, "len_title": 138, "len_sents": 56254, "len_sents_1": 31992, "n_sents": 618, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1326, "title": "On the Convergence of Adam under Non-uniform Smoothness: Separability from SGDM and Beyond", "abs": "This paper aims to clearly distinguish between Stochastic Gradient Descent with Momentum (SGDM) and Adam in terms of their convergence rates. We demonstrate that Adam achieves a faster convergence compared to SGDM under the condition of non-uniformly bounded smoothness. Our findings reveal that: (1) in deterministic environments, Adam can attain the known lower bound for the convergence rate of deterministic first-order optimizers, whereas the convergence rate of Gradient Descent with Momentum (GDM) has higher order dependence on the initial function value; (2) in stochastic setting, Adam's convergence rate upper bound matches the lower bounds of stochastic first-order optimizers, considering both the initial function value and the final error, whereas there are instances where SGDM fails to converge with any learning rate. These insights distinctly differentiate Adam and SGDM regarding their convergence rates. Additionally, by introducing a novel stopping-time based technique, we further prove that if we consider the minimum gradient norm during iterations, the corresponding convergence rate can match the lower bounds across all problem hyperparameters. The technique can also help proving that Adam with a specific hyperparameter scheduler is parameter-agnostic, which hence can be of independent interest.", "keywords": [ "Adam", "Convergence", "Separability", "Non-uniform Smoothness" ], "rating_list": [ 1, 3, 8, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "mEACsjW10N", "primary_area": "foundation or frontier models, including LLMs", "n_author": 17, "author_info_list": [ { "name": "Chunpu Xu", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "", "position": "PhD student" }, { "name": "Ge Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Haoran Que", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Liu Jiaheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jiashuo WANG", "gender": "Female", "institution": "The Hong Kong Polytechnic University, Hong Kong Polytechnic University", "country": "", "position": "PhD student" }, { "name": "Kang Zhu", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Ke Xu", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Ning Shi", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "PhD student" }, { "name": "Wangchunshu Zhou", "gender": "Male", "institution": "Guangdong OPPO Mobile Telecommunications Corp.,Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Wenhao Huang", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Yibo Zhang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Yizhi LI", "gender": "Male", "institution": "University of Manchester ", "country": "GB", "position": "PhD student" }, { "name": "Zhaoxiang Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Jie Fu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Zekun Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Li Siyu", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Yuanxing Zhang", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 62, "n_ref": 167, "n_ref_all": 194, "n_fig": 5, "n_tab": 12, "L_tab": 2740, "n_element_tab": 348, "n_fig_1": 1, "n_tab_1": 8, "L_tab_1": 2131, "n_element_tab_1": 245, "formula_len_all": 105, "formula_len_all_1": 339, "len_all": 379278, "len_all_1": 53549, "len_abs": 1671, "len_title": 92, "len_sents": 43303, "len_sents_1": 22523, "n_sents": 351, "n_sents_1": 147, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1562, "title": "MIO: A Foundation Model on Multimodal Tokens", "abs": "In this paper, we introduce MIO, a novel foundation model built on multimodal tokens, capable of understanding and generating speech, text, images, and videos in an end-to-end, autoregressive manner. While the emergence of large language models (LLMs) and multimodal large language models (MM-LLMs) propels advancements in artificial general intelligence through their versatile capabilities, they still lack true any-to-any understanding and generation. Recently, the release of GPT-4o has showcased the remarkable potential of any-to-any LLMs for complex real-world tasks, enabling omnidirectional input and output across images, speech, and text. However, it is closed-source and does not support the generation of multimodal interleaved sequences. To address this gap, we present MIO, which is trained on a mixture of discrete tokens across four modalities using causal multimodal modeling. MIO undergoes a four-stage training process: (1) alignment pre-training, (2) interleaved pre-training, (3) speech-enhanced pre-training, and (4) comprehensive supervised fine-tuning on diverse textual, visual, and speech tasks. Our experimental results indicate that MIO exhibits competitive, and in some cases superior, performance compared to previous dual-modal baselines, any-to-any model baselines, and even modality-specific baselines. Moreover, MIO demonstrates advanced capabilities inherent to its any-to-any feature, such as interleaved video-text generation, chain-of-visual-thought reasoning, visual guideline generation, instructional image editing, etc.", "keywords": [ "Large Language Models", "Multimodal Language Models" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "mDvL3wcmms", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Florentin Guth", "gender": "unknown", "institution": "New York University", "country": "", "position": "Postdoc" }, { "name": "Louis THIRY", "gender": "unknown", "institution": "INRIA", "country": "FR", "position": "Postdoc" } ], "n_formula": 26, "n_formula_1": 14, "n_ref_uni": 75, "n_ref": 126, "n_ref_all": 147, "n_fig": 5, "n_tab": 5, "L_tab": 624, "n_element_tab": 87, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 365, "n_element_tab_1": 46, "formula_len_all": 1504, "formula_len_all_1": 963, "len_all": 157155, "len_all_1": 68732, "len_abs": 1084, "len_title": 81, "len_sents": 40071, "len_sents_1": 31461, "n_sents": 301, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 33, "L_abs": 1087, "title": "Classification-denoising networks", "abs": "Image classification and denoising suffer from complementary issues of lack of robustness or partially ignoring conditioning information. We argue that they can be alleviated by unifying both tasks through a model of the joint probability of (noisy) images and class labels. Classification is performed with a forward pass followed by conditioning. Using the Tweedie-Miyasawa formula, we evaluate the denoising function with the score, which can be computed by marginalization and back-propagation. The training objective is then a combination of cross-entropy loss and denoising score matching loss integrated over noise levels. Numerical experiments on CIFAR-10 and ImageNet show competitive classification and denoising performance compared to reference deep convolutional classifiers/denoisers, and significantly improves efficiency compared to previous joint approaches. Our model shows an increased robustness to adversarial perturbations compared to a standard discriminative classifier, and allows for a novel interpretation of adversarial gradients as a difference of denoisers.", "keywords": [ "image classification", "denoising", "diffusion models", "energy-based models" ], "rating_list": [ 5, 5, 1, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 2, 5, 4 ] }, { "paper_id": "mDV36U4d6u", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Anh Totti Nguyen", "gender": "Male", "institution": "Auburn University", "country": "US", "position": "Associate Professor" }, { "name": "Freddy Lecue", "gender": "unknown", "institution": "INRIA", "country": "", "position": "Full Professor" }, { "name": "Giang Nguyen", "gender": "Male", "institution": "Auburn University", "country": "US", "position": "PhD student" }, { "name": "Ivan Brugere", "gender": "Male", "institution": "J.P. Morgan", "country": "US", "position": "Researcher" }, { "name": "Sanjay Kariyappa", "gender": "Male", "institution": "J.P. Morgan Chase", "country": "US", "position": "Researcher" }, { "name": "Shubham Sharma", "gender": "Male", "institution": "J.P. Morgan Chase", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 113, "n_ref_all": 147, "n_fig": 12, "n_tab": 38, "L_tab": 17562, "n_element_tab": 2607, "n_fig_1": 9, "n_tab_1": 12, "L_tab_1": 3195, "n_element_tab_1": 729, "formula_len_all": 39, "formula_len_all_1": 0, "len_all": 220672, "len_all_1": 70545, "len_abs": 1548, "len_title": 128, "len_sents": 54039, "len_sents_1": 27798, "n_sents": 424, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1515, "title": "Interpretable Table Question Answering via Plans of Atomic Table Transformations", "abs": "Interpretability for Table Question Answering (Table QA) is critical, particularly in high-stakes domains like finance or healthcare.\nWhile recent Large Language Models (LLMs) have improved the accuracy of Table QA models, their explanations for how answers are derived may not be transparent, hindering user ability to trust, explain, and debug predicted answers, especially on complex queries.\nWe introduce Plan-of-SQLs (POS), a novel method specifically crafted to enhance interpretability by decomposing a query into simpler sub-queries that are sequentially translated into SQL commands to generate the final answer.\nUnlike existing approaches, \nPOS offers full transparency in Table QA by ensuring that every transformation of the table is traceable, allowing users to follow the reasoning process step-by-step.\nVia subjective and objective evaluations, we show that POS explanations significantly improve interpretability, enabling both human and LLM judges to predict model responses with 93.00% and 85.25% accuracy, respectively.\nPOS explanations also consistently rank highest in clarity, coherence, and helpfulness compared to state-of-the-art Table QA methods such as Chain-of-Table and DATER.\nFurthermore, POS demonstrates high accuracy on Table QA benchmarks (78.31% on TabFact and 54.80% on WikiTQ with GPT3.5), outperforming methods that rely solely on LLMs or programs for table transformations, while remaining competitive with hybrid approaches that often trade off interpretability for accuracy.", "keywords": [ "Table QA; Interpretability; XAI; LLM-as-a-Judge" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "mDKxlfraAn", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Hai Ci", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Haofan Wang", "gender": "Male", "institution": "InstantX", "country": "CN", "position": "Researcher" }, { "name": "Yepeng Liu", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "PhD student" }, { "name": "Yiren Song", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "PhD student" }, { "name": "Yuheng Bu", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Zheng Shou", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 41, "n_ref": 89, "n_ref_all": 112, "n_fig": 7, "n_tab": 1, "L_tab": 462, "n_element_tab": 36, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 744, "n_element_tab_1": 180, "formula_len_all": 1253, "formula_len_all_1": 1134, "len_all": 125475, "len_all_1": 65671, "len_abs": 1636, "len_title": 128, "len_sents": 37112, "len_sents_1": 31308, "n_sents": 245, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 79, "L_abs": 1642, "title": "Image Watermarks are Removable using Controllable Regeneration from Clean Noise", "abs": "Image watermark techniques provide an effective way to assert ownership, deter misuse, and trace content sources, which has become increasingly essential in the era of large generative models. A critical attribute of watermark techniques is their robustness against various manipulations. In this paper, we introduce a watermark removal approach capable of effectively nullifying the state of the art watermarking techniques. Our primary insight involves regenerating the watermarked image starting from a clean Gaussian noise via a controllable diffusion model, utilizing the extracted semantic and spatial features from the watermarked image. The semantic control adapter and the spatial control network are specifically trained to control the denoising process towards ensuring image quality and enhancing consistency between the cleaned image and the original watermarked image. To achieve a smooth trade-off between watermark removal performance and image consistency, we further propose an adjustable and controllable regeneration scheme. This scheme adds varying numbers of noise steps to the latent representation of the watermarked image, followed by a controlled denoising process starting from this noisy latent representation. As the number of noise steps increases, the latent representation progressively approaches clean Gaussian noise, facilitating the desired trade-off. We apply our watermark removal methods across various watermarking techniques, and the results demonstrate that our methods offer superior visual consistency/quality and enhanced watermark removal performance compared to existing regeneration approaches.", "keywords": [ "Watermark", "Detection", "Robustness", "Diffusion Model" ], "rating_list": [ 6, 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "mDEYl0Ucgr", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Peter Stone", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Principal Researcher" }, { "name": "Scott Niekum", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Associate Professor" }, { "name": "Serena Booth", "gender": "Female", "institution": "Brown University", "country": "US", "position": "Assistant Professor" }, { "name": "Stephane Hatgis-Kessell", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" }, { "name": "W. Bradley Knox", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 20, "n_ref": 45, "n_ref_all": 59, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 546, "formula_len_all_1": 598, "len_all": 104496, "len_all_1": 68015, "len_abs": 1412, "len_title": 108, "len_sents": 35477, "len_sents_1": 31812, "n_sents": 235, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1417, "title": "Influencing Humans to Conform to Preference Models for RLHF", "abs": "Designing a reinforcement learning from human feedback (RLHF) algorithm for learning from preferences requires assuming a preference model, sometimes implicitly. A preference model that poorly describes how humans generate preferences risks learning a poor approximation of the human\u2019s unobservable reward function. In this paper, we conduct three human studies to assess whether one can influence the expression of real human preferences to more closely conform to a desired preference model. Importantly, our approach does not seek to alter the human's unobserved reward function. Rather, we change how humans use this reward function to generate preferences, such that they better match whatever preference model is assumed by a particular RLHF algorithm. We introduce three interventions: showing humans the quantities that underlie a preference model, which is normally unobservable information derived from the reward function; training people to follow a specific preference model; and modifying the preference elicitation question. All intervention types show significant effects, providing practical tools to improve preference data quality and the resultant alignment of learned reward functions.Overall we establish a novel research direction in model alignment: training humans and designing interfaces to increase human conformance with the assumptions of the algorithm that will learn from their input.", "keywords": [ "reinforcement learning from human feedback", "reinforcement learning", "reward functions", "preferences", "regret", "alignment" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 4, 2, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "mCO6FAOgYn", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Leye Wang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Assistant Professor" }, { "name": "Mingyu Ding", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Zhihao Yu", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "zhu Jie", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Ping Luo", "gender": "unknown", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 50, "n_ref": 117, "n_ref_all": 138, "n_fig": 8, "n_tab": 8, "L_tab": 1185, "n_element_tab": 189, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 985, "n_element_tab_1": 72, "formula_len_all": 309, "formula_len_all_1": 391, "len_all": 168139, "len_all_1": 69202, "len_abs": 1513, "len_title": 108, "len_sents": 48461, "len_sents_1": 31106, "n_sents": 396, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1518, "title": "An Empirical Study of Multiple Masking in Masked Autoencoder", "abs": "The performance of masked autoencoders hinges significantly on masking, prompting considerable efforts towards devising superior masking strategies. However, these strategies mask only once and employ masking directly on the input image. Afterward, inspired by the flexibility of masking, subsequent works introduce two rounds of masking. Unfortunately, all initiatives primarily focus on enhancing model performance, lacking an in-depth and systematical understanding of multiple masking for masked autoencoder. To bridge this gap, this work introduce a masked framework with multiple masking stages, termed Conditional MAE, where subsequent maskings are conditioned on previous unmasked representations, enabling a more flexible masking process in masked image modeling. By doing so, our study sheds light on how multiple masking affects the optimization in training and performance of pretrained models, e.g., introducing more locality to models, and summarizes several takeaways from our findings. Finally, we empirically evaluate the performance of our best-performing model(Conditional-MAE) with that of MAE in three folds including transfer learning, robustness, and scalability, demonstrating the effectiveness of our multiple masking strategy. We also follow our takeaways and show the generalizability to other heterogeneous networks including SimMIM and ConvNeXt V2. We hope our findings will inspire further work in the field and release the code at https:\n//anonymous.4open.science/r/conditional-mae-512C.", "keywords": [ "masked autoencoder", "multiple masking", "masked image modeling" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 1 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "mBrAuyd26J", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Haobo Fu", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Liwen Zhu", "gender": "Female", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "QIANG FU", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Shiweixu", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "MS student" }, { "name": "Tao Yang", "gender": "Male", "institution": "Northwestern Polytechnical University", "country": "", "position": "MS student" }, { "name": "Yang Wei", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "shuang wu", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 55, "n_ref": 95, "n_ref_all": 160, "n_fig": 12, "n_tab": 16, "L_tab": 5150, "n_element_tab": 365, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 134, "n_element_tab_1": 23, "formula_len_all": 176, "formula_len_all_1": 176, "len_all": 271695, "len_all_1": 63204, "len_abs": 1407, "len_title": 141, "len_sents": 90527, "len_sents_1": 31249, "n_sents": 732, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1411, "title": "Enhance Reasoning for Large Language Models with Reinforcement Learning in the Game Werewolf", "abs": "Despite their success across a broad spectrum of general tasks, Large Language Models (LLMs) often underperform in domain-specific tasks not well-represented in their pre-training corpora. We introduce an innovative framework integrating general-purpose LLMs with an external \\emph{Thinker} module to enhance the reasoning capabilities of LLM-based agents. Unlike augmenting LLMs with prompt engineering, our Thinker module directly accesses knowledge from domain databases and employs supervised or reinforcement learning (RL). We establish a reasoning hierarchy where LLMs handle intuitive System-1 tasks that are domain-agnostic, while the Thinker focuses on System-2 tasks that require complex logical analysis and domain-specific knowledge. Our framework is demonstrated through a 9-player Werewolf game that necessitates dual-system reasoning. We design a communication protocol between LLMs and the Thinker, then optimize the Thinker through online RL and refine it by imitation learning. Drawing from 18800 human games, this work also contributes to the largest dataset for social deduction games to date. Experiments show that GPT-3.5 and GPT-4, augmented with the Thinker, significantly improve in deductive reasoning, speech generation, and online gameplay evaluated by human players. Further, integrating a fine-tuned 6B Werewolf-specific LLM with the Thinker achieves performance on par with GPT-4.", "keywords": [ "large language models", "reinforcement learning", "social deduction game" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "mBXLtNKpeQ", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Carola-Bibiane Sch\u00f6nlieb", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "Full Professor" }, { "name": "Erik J Bekkers", "gender": "unknown", "institution": "University of Amsterdam", "country": "", "position": "Assistant Professor" }, { "name": "Friso de Kruiff", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Intern" }, { "name": "Ozan \u00d6ktem", "gender": "Male", "institution": "Uppsala University", "country": "", "position": "Full Professor" }, { "name": "Willem Diepeveen", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Postdoc" } ], "n_formula": 66, "n_formula_1": 21, "n_ref_uni": 33, "n_ref": 59, "n_ref_all": 84, "n_fig": 7, "n_tab": 2, "L_tab": 737, "n_element_tab": 113, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 737, "n_element_tab_1": 113, "formula_len_all": 4454, "formula_len_all_1": 1989, "len_all": 154095, "len_all_1": 65800, "len_abs": 1145, "len_title": 88, "len_sents": 51294, "len_sents_1": 28246, "n_sents": 374, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1149, "title": "Pullback Flow Matching on Data Manifolds", "abs": "We propose Pullback Flow Matching (PFM), a novel framework for generative modeling on data manifolds. Unlike existing methods that assume or learn restrictive closed-form manifold mappings for training Riemannian Flow Matching (RFM) models, PFM leverages pullback geometry and isometric learning to preserve the underlying manifold\u2019s geometry while enabling efficient generation and precise interpolation in latent space. This approach not only facilitates closed-form mappings on the data manifold but also allows for designable latent spaces, using assumed metrics on both data and latent manifolds. By enhancing isometric learning through Neural ODEs and proposing a scalable training objective, we achieve a latent space more suitable for interpolation, leading to improved manifold learning and generative performance. We demonstrate PFM\u2019s effectiveness through applications in synthetic data, protein dynamics and protein sequence data, generating novel proteins with specific properties. This method shows strong potential for drug discovery and materials science, where generating novel samples with specific properties is of great interest.", "keywords": [ "Pullback Geometry", "Manifold Learning", "Generative Modeling", "Neural ODEs", "Data manifolds" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "mBJF0p9yRR", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Shangtong Zhang", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "", "position": "Assistant Professor" }, { "name": "Ethan H Blaser", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" } ], "n_formula": 190, "n_formula_1": 36, "n_ref_uni": 31, "n_ref": 104, "n_ref_all": 126, "n_fig": 0, "n_tab": 3, "L_tab": 581, "n_element_tab": 21, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 402, "n_element_tab_1": 25, "formula_len_all": 13067, "formula_len_all_1": 1562, "len_all": 180647, "len_all_1": 65540, "len_abs": 664, "len_title": 118, "len_sents": 52152, "len_sents_1": 26409, "n_sents": 654, "n_sents_1": 281, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 665, "title": "Almost Sure Convergence of Average Reward Temporal Difference Learning", "abs": "Tabular average reward Temporal Difference (TD) learning is perhaps the simplest and the most fundamental policy evaluation algorithm in average reward reinforcement learning. After at least 25 years since its discovery, we are finally able to provide a long-awaited almost sure convergence analysis. Namely, we are the first to prove that, under very mild conditions, tabular average reward TD converges almost surely to a sample path dependent fixed point. Key to this success is a new general stochastic approximation result concerning nonexpansive mappings with Markovian and additive noise, built on recent advances in stochastic Krasnoselskii-Mann iterations.", "keywords": [ "reinforcement learning", "temporal difference learning", "average reward" ], "rating_list": [ 3, 1, 6, 3 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 1, 3, 1 ], "confidence_list": [ 3, 5, 3, 5 ] }, { "paper_id": "mAztx8QO3B", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Dong Gong", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "Assistant Professor" }, { "name": "Ruilin Tong", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "PhD student" }, { "name": "Yuhang Liu", "gender": "unknown", "institution": "The University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Javen Shi", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Professor" } ], "n_formula": 25, "n_formula_1": 14, "n_ref_uni": 48, "n_ref": 154, "n_ref_all": 202, "n_fig": 18, "n_tab": 15, "L_tab": 4854, "n_element_tab": 301, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1800, "n_element_tab_1": 98, "formula_len_all": 3222, "formula_len_all_1": 1236, "len_all": 205080, "len_all_1": 64839, "len_abs": 1651, "len_title": 106, "len_sents": 62505, "len_sents_1": 29357, "n_sents": 451, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1659, "title": "Coreset Selection via Reducible Loss in Continual Learning", "abs": "A natural solution for rehearsal-based continual learning is to select a coreset as memory. A coreset serves as an informative summary of a large dataset, enabling a model trained solely on the coreset to achieve performance comparable to training on the full dataset. Previous bi-level coreset selection methods adjust sample weights or probabilities to minimize the outer loss, which is computed over the entire dataset. For non-representative samples like ambiguous or noisy samples, since these samples are not well learned even training model on the full dataset, loss of these samples in the outer loss are not worthy to be reduced. However, their high loss values may cause them to be selected in an attempt to minimize the outer loss, which may lead to suboptimal performance for models trained on the coreset. To address this issue, we first investigate how the performance of a trained model changes when a sample is added to the training dataset and approximate this performance gain using reducible loss. We then select samples with the highest performance gain in the coreset so that performance of model trained on coreset could be maximized. We show that samples with high performance gain are informative and representative. Furthermore, reducible loss requires only forward computation, making it significantly more efficient than previous methods. To better apply coreset selection in continual learning, we extend our method to address key challenges such as task interference, streaming data, and knowledge distillation. Experiments on data summarization and continual learning demonstrate the effectiveness and efficiency of our approach.", "keywords": [ "Continual learning", "Coreset selection" ], "rating_list": [ 5, 5, 8 ], "soundness_list": [ 2, 2, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 2, 4 ], "confidence_list": [ 5, 5, 3 ] }, { "paper_id": "mAmCdASmJ5", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Alexander Semenenko", "gender": "unknown", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Aleksander Tolmachev", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Researcher" }, { "name": "Alexey Frolov", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Andrey Gladkov", "gender": "unknown", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Ivan Butakov", "gender": "unknown", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Researcher" }, { "name": "Marina Munkhoeva", "gender": "Female", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Research Scientist" } ], "n_formula": 56, "n_formula_1": 30, "n_ref_uni": 61, "n_ref": 108, "n_ref_all": 132, "n_fig": 7, "n_tab": 3, "L_tab": 909, "n_element_tab": 38, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 405, "n_element_tab_1": 146, "formula_len_all": 4499, "formula_len_all_1": 2369, "len_all": 181771, "len_all_1": 66197, "len_abs": 1151, "len_title": 131, "len_sents": 42772, "len_sents_1": 27813, "n_sents": 418, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1155, "title": "Efficient Distribution Matching of Representations via Noise-Injected Deep InfoMax", "abs": "Deep InfoMax (DIM) is a well-established method for self-supervised representation learning (SSRL) based on maximization of the mutual information between the input and the output of a deep neural network encoder. Despite the DIM and contrastive SSRL in general being well-explored, the task of learning representations conforming to a specific distribution (i.e., distribution matching, DM) is still under-addressed. Motivated by the importance of DM to several downstream tasks (including generative modeling, disentanglement, outliers detection and other), we enhance DIM to enable automatic matching of learned representations to a selected prior distribution. To achieve this, we propose injecting an independent noise into the normalized outputs of the encoder, while keeping the same InfoMax training objective. We show that such modification allows for learning uniformly and normally distributed representations, as well as representations of other absolutely continuous distributions. Our approach is tested on various downstream tasks. The results indicate a moderate trade-off between the performance on the downstream tasks and quality of DM.", "keywords": [ "information theory", "deep infomax", "self-supervised learning", "representation learning", "distribution matching", "Gaussian embeddings" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "m9wG6ai2Xk", "primary_area": "datasets and benchmarks", "n_author": 16, "author_info_list": [ { "name": "Bhargav Bhushanam", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Daochen Zha", "gender": "unknown", "institution": "Airbnb", "country": "US", "position": "Researcher" }, { "name": "Kai-Wei Chang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Associate Professor" }, { "name": "Kaixiong Zhou", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Assistant Professor" }, { "name": "Lize Shao", "gender": "unknown", "institution": "Rice University", "country": "US", "position": "Undergrad student" }, { "name": "Louis Feng", "gender": "Male", "institution": "University of California, Davis", "country": "", "position": "PhD student" }, { "name": "Shaochen Zhong", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Shuai Xu", "gender": "Male", "institution": "Case Western Reserve University", "country": "US", "position": "Assistant Professor" }, { "name": "Vipin Chaudhary", "gender": "Male", "institution": "Case Western Reserve University", "country": "US", "position": "Full Professor" }, { "name": "Xiaocong Du", "gender": "Female", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Yifan Lu", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Undergrad student" }, { "name": "Yiwei Wang", "gender": "Male", "institution": "University of California, Merced", "country": "US", "position": "Assistant Professor" }, { "name": "Yixin Wan", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Yucheng Shi", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Ninghao Liu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Assistant Professor" }, { "name": "Xia Hu", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 19, "n_ref": 101, "n_ref_all": 129, "n_fig": 3, "n_tab": 11, "L_tab": 2977, "n_element_tab": 423, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 3475, "n_element_tab_1": 204, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 199444, "len_all_1": 56979, "len_abs": 1902, "len_title": 141, "len_sents": 48853, "len_sents_1": 24929, "n_sents": 338, "n_sents_1": 141, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 93, "L_abs": 1922, "title": "MQuAKE-Remastered: Multi-Hop Knowledge Editing Can Only Be Advanced with Reliable Evaluations", "abs": "Large language models (LLMs) can give out erroneous answers to factually rooted questions either as a result of undesired training outcomes or simply because the world has moved on after a certain knowledge cutoff date. Under such scenarios, knowledge editing often comes to the rescue by delivering efficient patches for such erroneous answers without significantly altering the rests, where many editing methods have seen reasonable success when the editing targets are simple and direct (e.g., \"what club does Lionel Messi currently play for?\").\n\nHowever, knowledge fragments like this are often deeply intertwined in the real world, making effectively propagating the editing effect to non-directly related questions a practical challenge (e.g., \"who is the offspring of the owner of the club that Messi currently plays for?\"). Prior arts have coined this task as multi-hop knowledge editing with the most popular dataset being MQuAKE, serving as the sole evaluation benchmark for many later proposed editing methods due to the expensive nature of making knowledge editing datasets at scale.\n\nIn this work, we reveal that **up to 33% or 76% of MQuAKE's questions and ground truth labels are, in fact, corrupted in various fashions due to some unintentional clerical or procedural oversights.** Our work provides a detailed audit of MQuAKE's error pattern and a comprehensive fix without sacrificing its dataset capacity. Additionally, we benchmarked almost all proposed \\mquake{}-evaluated editing methods on our post-fix dataset, \\mquaker{}. It is our observation that many methods try to overfit the original \\mquake{} by exploiting some data-specific properties of \\mquake{}. We provide a guideline on how to faithfully approach such datasets and show that a simple, minimally invasive approach can bring excellent editing performance without such exploitation. Please refer to the supplemental material for assets.", "keywords": [ "knowledge edit", "model edit", "multi-hop", "question answering", "natural language processing", "dataset audit" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "m9RNBZewW2", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Jinjin Gu", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "Instructor" }, { "name": "Keda TAO", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Nan Cheng", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Xiucheng Wang", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yulun Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 16, "n_ref": 30, "n_ref_all": 72, "n_fig": 22, "n_tab": 10, "L_tab": 3820, "n_element_tab": 368, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 1701, "n_element_tab_1": 228, "formula_len_all": 212, "formula_len_all_1": 202, "len_all": 174180, "len_all_1": 59034, "len_abs": 0, "len_title": 145, "len_sents": 52425, "len_sents_1": 26732, "n_sents": 376, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 97, "L_abs": 1134, "title": "Overcoming False Illusions in Real-World Face Restoration with Multi-Modal Guided Diffusion Model", "abs": "We introduce a novel Multi-modal Guided Real-World Face Restoration (MGFR) technique designed to improve the quality of facial image restoration from low-quality inputs. Leveraging a blend of attribute text prompts, high-quality reference images, and identity information, MGFR can mitigate the generation of false facial attributes and identities often associated with generative face restoration methods. By incorporating a dual-control adapter and a two-stage training strategy, our method effectively utilizes multi-modal prior information for targeted restoration tasks. We also present the Reface-HQ dataset, comprising over 23,000 high-resolution facial images across 5,000 identities, to address the need for reference face training images. Our approach achieves superior visual quality in restoring facial details under severe degradation and allows for controlled restoration processes, enhancing the accuracy of identity preservation and attribute correction. Including negative quality samples and attribute prompts in the training further refines the model's ability to generate detailed and perceptually accurate images.", "keywords": [ "Face image restoration", "diffusion model" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "m9BiWVTJDx", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Beomgu Kang", "gender": "Male", "institution": "Korea Institute of Science and Technology", "country": "KR", "position": "Postdoc" }, { "name": "Hyunseok Seo", "gender": "Male", "institution": "Korea Institute of Science and Technology", "country": "KR", "position": "Principal Researcher" } ], "n_formula": 19, "n_formula_1": 15, "n_ref_uni": 35, "n_ref": 43, "n_ref_all": 59, "n_fig": 6, "n_tab": 1, "L_tab": 133, "n_element_tab": 38, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 408, "n_element_tab_1": 40, "formula_len_all": 973, "formula_len_all_1": 809, "len_all": 107764, "len_all_1": 51567, "len_abs": 1927, "len_title": 130, "len_sents": 30826, "len_sents_1": 25034, "n_sents": 197, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 117, "L_abs": 1935, "title": "A Probabilistic Approach to Optimizing Hardware Control Parameters in System Property Estimation using Gumbel-Softmax", "abs": "Optimizing control parameters is crucial to estimate reliable tissue characteristics in quantitative MRI. Basically, multiple hardware parameters are simultaneously controlled to generate a signal from MRI system. Repetitive acquisitions with different control parameter combinations create distinct signal modulations and then tissue characteristics are deduced from prior knowledge of physics-based relationship among modulated signals, control parameters, and tissue characteristics. The choice of control parameters, which determines the attribute of signal modulation, directly impacts the inverse problem in tissue characteristic estimation. Thus, the multidimensional control parameter optimization remains an open research topic in MRI field for accurate analysis of tissue characteristics. Typically, optimal parameters are determined by iteratively updating sets of control parameters to maximize the estimation accuracy of the tissue characteristics. However, the conventional optimization process is restricted to explore only the vicinity of control parameters at the current iteration. Therefore, it could highly depend on initialization and current parameters, which might lead to inefficient search especially when noise is present in the system. In this work, to mitigate this limitation, we propose a novel Gumbel-Softmax-based optimization scheme that enables a probabilistic search across an expanding set of all candidates for each control parameter using categorical reparameterization. As a case study, the proposed method is employed to find optimal control parameters for quantitative MRI. We demonstrate that our Gumbel-Softmax-based optimization simultaneously explores the entire range of control parameters from early iterations and outperforms the conventional optimization approach on accuracy of MR tissue characteristic estimation and repeatability of optimization, especially under noisy environments.", "keywords": [ "categorical reparameterization", "Gumbel Softmax", "Optimization", "System properties", "physics model" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 2, 3, 3 ] }, { "paper_id": "m8yby1JfbU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Ming Liu", "gender": "Not Specified", "institution": "Iowa State University", "country": "US", "position": "PhD student" }, { "name": "Wensheng Zhang", "gender": "unknown", "institution": "Iowa State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 30, "n_ref": 47, "n_ref_all": 77, "n_fig": 8, "n_tab": 6, "L_tab": 1410, "n_element_tab": 90, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 241, "formula_len_all_1": 239, "len_all": 160042, "len_all_1": 61329, "len_abs": 1362, "len_title": 0, "len_sents": 52048, "len_sents_1": 30281, "n_sents": 439, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 46, "L_abs": 1383, "title": "Is Your Video Language Model a Reliable Judge?", "abs": "Evaluating video language models (VLMs) is crucial for improving their understanding of video content. \nExisting evaluation methods depend on single models, \nwhich may be unreliable or biased due to the models' incapability to understand content or inherent bias, \nultimately compromising the reliability of evaluation. \nA straightforward remedy is to apply the principle of collective thoughts, \naggregating reviews from multiple VLMs to enhance reliability. \nThis study investigates the efficacy of such approaches in VLM evaluation, \nparticularly when the pool of judges includes both reliable and unreliable models. \nOur findings reveal that incorporating collective judgments from such a mixed pool\ndoes not necessarily enhance the accuracy of the final evaluation outcomes, \nbecause the less reliable judges could introduce noise that \npotentially leads to less reliable evaluations. \nTo explore the factors that impact evaluation reliability, \nwe fine-tune an underperforming VLM judge, Video-LLaVA, and observe that \ngood understanding ability alone is insufficient\nto make VLM judges reliable. \nThese findings stress the limitations of collective thought approaches in VLM evaluation and \nhighlight the need for more advanced methods that can account for the reliability of individual models. \nOur study promotes the development of more reliable evaluation methods for VLMs.", "keywords": [ "Video Language Models", "Model evaluation", "Collective thought", "Reliability" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "m8Rk3HLGFx", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Di ZHANG", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "VP" }, { "name": "Haoji Hu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Jianhong Bai", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Menghan Xia", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Researcher" }, { "name": "Pengfei Wan", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Director" }, { "name": "Xintao Wang", "gender": "Male", "institution": "Kuaishou", "country": "CN", "position": "Researcher" }, { "name": "Ziyang Yuan", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Zuozhu Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 43, "n_ref": 108, "n_ref_all": 131, "n_fig": 13, "n_tab": 5, "L_tab": 1123, "n_element_tab": 120, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 548, "n_element_tab_1": 60, "formula_len_all": 441, "formula_len_all_1": 438, "len_all": 165865, "len_all_1": 57118, "len_abs": 219, "len_title": 129, "len_sents": 40393, "len_sents_1": 26179, "n_sents": 289, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1475, "title": "SynCamMaster: Synchronizing Multi-Camera Video Generation from Diverse Viewpoints", "abs": "Recent advancements in video diffusion models demonstrate remarkable capabilities in simulating real-world dynamics and 3D consistency. This progress motivates us to explore the potential of these models to maintain dynamic consistency across diverse viewpoints, a feature highly sought after in applications like virtual filming. Unlike existing methods focused on multi-view generation of single objects for 4D reconstruction, our interest lies in generating open-world videos from arbitrary viewpoints, incorporating six degrees of freedom (6 DoF) camera poses.\nTo achieve this, we propose a plug-and-play module that enhances a pre-trained text-to-video model for multi-camera video generation, ensuring consistent content across different viewpoints. Specifically, we introduce a multi-view synchronization module designed to maintain appearance and geometry consistency across these viewpoints. Given the scarcity of high-quality training data, we also propose a progressive training scheme that leverages multi-camera images and monocular videos as a supplement to Unreal Engine-rendered multi-camera videos. This comprehensive approach significantly benefits our model.\nExperimental results demonstrate the superiority of our proposed method over existing competitors and several baselines. Furthermore, our method enables intriguing extensions, such as re-rendering a video from multiple novel viewpoints. Project webpage: https://jianhongbai.github.io/SynCamMaster/", "keywords": [ "Video Generation", "Diffusion Model" ], "rating_list": [ 6, 3, 6, 8, 6 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 1, 3, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 4, 5, 3, 4, 4 ] }, { "paper_id": "m8ERGrOf1f", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 9, "author_info_list": [ { "name": "Ali Falahati", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Ang Li", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Associate Professor" }, { "name": "Chunshu Wu", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Postdoc" }, { "name": "Mahdi Ghandi", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Mingqin Li", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Pouya Haghi", "gender": "Male", "institution": "University of Rochester", "country": "", "position": "Postdoc" }, { "name": "Shuai Che", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Tong Geng", "gender": "Male", "institution": "University of Rochester", "country": "", "position": "Assistant Professor" }, { "name": "Zahra Azad", "gender": "Female", "institution": "University of Rochester", "country": "US", "position": "Postdoc" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 10, "n_ref": 19, "n_ref_all": 40, "n_fig": 7, "n_tab": 4, "L_tab": 1734, "n_element_tab": 166, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1589, "n_element_tab_1": 152, "formula_len_all": 445, "formula_len_all_1": 614, "len_all": 104436, "len_all_1": 68597, "len_abs": 1716, "len_title": 129, "len_sents": 36228, "len_sents_1": 31066, "n_sents": 284, "n_sents_1": 254, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1721, "title": "DM-Tune: Quantizing Diffusion Models with Mixture-of-Gaussian Guided Noise Tuning", "abs": "Diffusion models have become essential generative tools for tasks such as image generation, video creation, and inpainting, but their high computational and memory demands pose challenges for efficient deployment. Contrary to the traditional belief that full-precision computation ensures optimal image quality, we demonstrate that a fine-grained mixed-precision strategy can surpass full-precision models in terms of image quality, diversity, and text-to-image alignment. However, directly implementing such strategies can lead to increased complexity and reduced runtime performance due to the overheads of managing multiple precision formats and casting operations. To address this, we introduce DM-Tune, which replaces complex mixed-precision quantization with a unified low-precision format, supplemented by noise-tuning, to improve both image generation quality and runtime efficiency. The proposed noise-tuning mechanism is a type of fine-tuning that reconstructs the mixed-precision output by learning adjustable noise through a parameterized nonlinear function consisting of Gaussian and linear components. Key steps in our framework include identifying sensitive layers for quantization, modeling quantization noise, and optimizing runtime with custom low-precision GPU kernels that support efficient noise-tuning. Experimental results across various diffusion models and datasets demonstrate that DM-Tune not only significantly improves runtime but also enhances diversity, quality, and text-to-image alignment compared to FP32, FP8, and state-of-the-art mixed-precision methods. Our approach is broadly applicable and lays a solid foundation for simplifying complex mixed-precision strategies at minimal cost.", "keywords": [ "Diffusion Models", "Quantization", "Low-Precision GPU Kernels", "Noise-Tuning" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "m7lBCyROPP", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Donglin Wang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Associate Professor" }, { "name": "Xing Lei", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xuetao Zhang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "Associate Professor" }, { "name": "Zifeng Zhuang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" } ], "n_formula": 44, "n_formula_1": 21, "n_ref_uni": 46, "n_ref": 78, "n_ref_all": 128, "n_fig": 17, "n_tab": 2, "L_tab": 512, "n_element_tab": 52, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2821, "formula_len_all_1": 1158, "len_all": 174284, "len_all_1": 66257, "len_abs": 1875, "len_title": 127, "len_sents": 56650, "len_sents_1": 31162, "n_sents": 447, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1857, "title": "Goal-conditioned Reinforcement Learning with Subgoals Generated from Relabeling", "abs": "In goal-conditioned reinforcement learning (RL), the primary objective is to develop a goal-conditioned policy capable of reaching diverse desired goals, a process often hindered by sparse reward signals. To address the challenges associated with sparse rewards, existing approaches frequently employ hindsight relabeling, substituting original goals with achieved goals. However, these methods exhibit a tendency to prioritize the optimization of closer achieved goals during training, leading to the loss of potentially valuable information from the trajectory and low sample efficiency. Our key insight is that these achieved goals, generated from the same hindsight relabeling, can serve as effective subgoals to facilitate the learning of policies that reach possible long-horizon desired goals within the same trajectory. Leveraging this perspective, we propose a novel framework called Goal-Conditioned reinforcement learning with Q-BC (i.e, behavior cloning (BC)-regularized Q) and Subgoals (GCQS) for goal-conditioned RL. GCQS is a innovative goal-conditioned actor-critic framework that systematically exploits more trajectory information to improve policy learning and sample efficiency. Specifically, GCQS initially optimizes a Q-BC objective to facilitate learning policies that reach achieved goals effectively. Subsequently, these achieved goals are redefined as subgoals, which serve to enhance the goal-conditioned policies, thereby predicting better actions to reach the desired goals. Experimental results in simulated robotics environments demonstrate that GCQS significantly enhances sample efficiency and overall performance compared to existing goal-conditioned methods. Additionally, GCQS demonstrated competitive performance on long-horizon AntMaze tasks, achieving results comparable to such state-of-the-art subgoal-based methods.", "keywords": [ "Goal-Conditioned Reinforcement Learning", "Hindsight Experience Replay", "Subgoal-based Approach" ], "rating_list": [ 5, 5, 5, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "m7Nd3K0iru", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Lei Feng", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" }, { "name": "Sheng-Jun Huang", "gender": "unknown", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Chenchen Zong", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Peng-Hui Yang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 34, "n_formula_1": 8, "n_ref_uni": 39, "n_ref": 66, "n_ref_all": 100, "n_fig": 5, "n_tab": 11, "L_tab": 3028, "n_element_tab": 321, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1084, "n_element_tab_1": 144, "formula_len_all": 3533, "formula_len_all_1": 980, "len_all": 133345, "len_all_1": 56268, "len_abs": 1566, "len_title": 134, "len_sents": 41223, "len_sents_1": 25041, "n_sents": 338, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1572, "title": "Dual-Head Knowledge Distillation: Enhancing Logits Utilization with an Auxiliary Head", "abs": "Traditional knowledge distillation focuses on aligning the student's predicted probabilities with both ground-truth labels and the teacher's predicted probabilities. However, the transition to predicted probabilities from logits would obscure certain indispensable information. To address this issue, it is intuitive to additionally introduce a logit-level loss function as a supplement to the widely used probability-level loss function, for exploiting the latent information of logits. Unfortunately, we empirically find that the amalgamation of the newly introduced logit-level loss and the previous probability-level loss will lead to performance degeneration, even trailing behind the performance of employing either loss in isolation. We attribute this phenomenon to the collapse of the classification head, which is verified by our theoretical analysis based on the neural collapse theory. Specifically, the gradients of the two loss functions exhibit contradictions in the linear classifier yet display no such conflict within the backbone. Drawing from the theoretical analysis, we propose a novel method called dual-head knowledge distillation, which partitions the linear classifier into two classification heads responsible for different losses, thereby preserving the beneficial effects of both losses on the backbone while eliminating adverse influences on the classification head. Extensive experiments validate that our method can effectively exploit the information inside the logits and achieve superior performance against state-of-the-art counterparts.", "keywords": [ "knowledge distillation" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 3, 1, 2 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "m73tETvFkX", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Joonsung Jeon", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Sooel Son", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Suhyeon Ha", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sung-eui Yoon", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "Professor" }, { "name": "Woo Jae Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 20, "n_ref": 51, "n_ref_all": 69, "n_fig": 5, "n_tab": 2, "L_tab": 947, "n_element_tab": 22, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 947, "n_element_tab_1": 22, "formula_len_all": 612, "formula_len_all_1": 615, "len_all": 109206, "len_all_1": 58728, "len_abs": 1408, "len_title": 142, "len_sents": 32452, "len_sents_1": 26657, "n_sents": 225, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1414, "title": "AdvPaint: Protecting Images from Inpainting Manipulation via Adversarial Attention Disruption", "abs": "The outstanding capability of diffusion models in generating high-quality images poses significant threats when misused by adversaries. In particular, we assume malicious adversaries exploiting diffusion models for inpainting tasks, such as replacing a specific region with a celebrity. While existing methods for protecting images from manipulation in diffusion-based generative models have primarily focused on image-to-image and text-to-image tasks, the challenge of preventing unauthorized inpainting has been rarely addressed, often resulting in suboptimal protection performance. To mitigate inpainting abuses, we propose ADVPAINT, a novel defensive framework that generates adversarial perturbations that effectively disrupt the adversary\u2019s inpainting tasks. ADVPAINT targets the self- and cross-attention blocks in a target diffusion inpainting model to distract semantic understanding and prompt interactions during image generation. ADVPAINT also employs a two-stage perturbation strategy, dividing the perturbation region based on an enlarged bounding box around the object, enhancing robustness across diverse masks of varying shapes and sizes. Our experimental results demonstrate that ADVPAINT\u2019s perturbations are highly effective in disrupting the adversary\u2019s inpainting tasks, outperforming existing methods; ADVPAINT attains over a 100-point increase in FID and substantial decreases in precision.", "keywords": [ "Adversarial Example", "Adversarial Attack", "Inpainting", "Image Protection" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "m60n31iYMw", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Nadav Cohen", "gender": "Male", "institution": "School of Computer Science, Tel Aviv University", "country": "IL", "position": "Associate Professor" }, { "name": "Noam Razin", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Yonatan Slutzky", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Yotam Alexander", "gender": "Male", "institution": "Tel Aviv University, Tel Aviv University", "country": "IL", "position": "PhD student" } ], "n_formula": 292, "n_formula_1": 21, "n_ref_uni": 47, "n_ref": 92, "n_ref_all": 254, "n_fig": 11, "n_tab": 11, "L_tab": 1569, "n_element_tab": 80, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1053, "n_element_tab_1": 14, "formula_len_all": 31857, "formula_len_all_1": 956, "len_all": 433447, "len_all_1": 71005, "len_abs": 1456, "len_title": 84, "len_sents": 135444, "len_sents_1": 34236, "n_sents": 1311, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1462, "title": "The Implicit Bias of Structured State Space Models Can Be Poisoned With Clean Labels", "abs": "Neural networks are powered by an implicit bias: a tendency of gradient descent to fit training data in a way that generalizes to unseen data. A recent class of neural network models gaining increasing popularity is structured state space models (SSMs), regarded as an efficient alternative to transformers. Prior work argued that the implicit bias of SSMs leads to generalization in a setting where data is generated by a low dimensional teacher. In this paper, we revisit the latter setting, and formally establish a phenomenon entirely undetected by prior work on the implicit bias of SSMs. Namely, we prove that while implicit bias leads to generalization under many choices of training data, there exist special examples whose inclusion in training completely distorts the implicit bias, to a point where generalization fails. This failure occurs despite the special training examples being labeled by the teacher, i.e. having clean labels! We empirically demonstrate the phenomenon, with SSMs trained independently and as part of non-linear neural networks. In the area of adversarial machine learning, disrupting generalization with cleanly labeled training examples is known as clean-label poisoning. Given the proliferation of SSMs, particularly in large language models, we believe significant efforts should be invested in further delineating their susceptibility to clean-label poisoning, and in developing methods for overcoming this susceptibility.", "keywords": [ "Structured State Space Models", "Implicit Bias", "Implicit Regularization", "Clean Label Poisoning" ], "rating_list": [ 3, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 1, 3, 3, 3 ], "confidence_list": [ 5, 3, 1, 2 ] }, { "paper_id": "m5rOrTiuKG", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Binhua Li", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Chen Zhi", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Fei Huang", "gender": "Male", "institution": "Alibaba Group US", "country": "US", "position": "Senior Research Director" }, { "name": "Shuiguang Deng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yongbin Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yue Zhang", "gender": "Female", "institution": "Alibaba Group", "country": "", "position": "Algorithm Engineer" }, { "name": "Yunkun Wang", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Zhen Qin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 28, "n_ref": 68, "n_ref_all": 102, "n_fig": 9, "n_tab": 11, "L_tab": 6668, "n_element_tab": 444, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2407, "n_element_tab_1": 206, "formula_len_all": 329, "formula_len_all_1": 232, "len_all": 210119, "len_all_1": 70404, "len_abs": 1833, "len_title": 101, "len_sents": 72235, "len_sents_1": 30719, "n_sents": 551, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 1851, "title": "ExploraCoder: Advancing code generation for multiple unseen APIs via planning and chained exploration", "abs": "Through training on publicly available source code libraries, large language models (LLMs) can invoke multiple encapsulated APIs to solve complex programming problems.\nHowever, existing models inherently cannot generalize to use APIs that are unseen in their training corpora. As libraries continuously evolve, it becomes impractical to exhaustively retrain LLMs with new API knowledge. This limitation hampers LLMs from solving problems which require newly introduced or privately maintained libraries.\nHuman programmers often explore unfamiliar APIs by writing experimental code before invoking them for a more complex problem.\nInspired by this behavior, we propose $\\textbf{ExploraCoder}$, a training-free framework that empowers LLMs to invoke multiple unseen APIs in code solution by (1) planning a complex problem into several API invocation subtasks, and (2) exploring correct API usage through a novel chain-of-API-exploration.\nConcretely, ExploraCoder guides the LLM to iteratively generate several experimental API invocations for each simple subtask, where the promising execution experience are exploited by subsequent subtasks. This forms a chained exploration trace that ultimately guides LLM in generating the final solution.\nWe evaluate ExploraCoder on Torchdata-Github benchmark as well as a newly constructed benchmark that involves more complex API interactions.\nExperimental results demonstrate that ExploraCoder significantly improves performance for models lacking prior API knowledge, achieving an absolute increase of 11.24\\% over niave RAG approaches and 14.07\\% over pretraining methods in pass@10. Moreover, the integration of a self-debug mechanism further boosts ExploraCoder's performance on more challenging tasks. Comprehensive ablation and case studies provide further insights into the effectiveness of ExploraCoder.", "keywords": [ "Large Language Models", "Code Generation", "Code Library", "Retrieval Augmented Generation" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 5, 3, 2 ] }, { "paper_id": "m5qpn0KTMZ", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Alexander Willi Fottner", "gender": "Male", "institution": "Universit\u00e4t Augsburg", "country": "", "position": "PhD student" }, { "name": "Linus Lach", "gender": "Male", "institution": "Universit\u00e4t Augsburg", "country": "DE", "position": "PhD student" }, { "name": "Yarema Okhrin", "gender": "Male", "institution": "University of Augsburg", "country": "DE", "position": "Full Professor" } ], "n_formula": 46, "n_formula_1": 26, "n_ref_uni": 44, "n_ref": 76, "n_ref_all": 90, "n_fig": 4, "n_tab": 3, "L_tab": 376, "n_element_tab": 26, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 186, "n_element_tab_1": 18, "formula_len_all": 5615, "formula_len_all_1": 2541, "len_all": 127186, "len_all_1": 66501, "len_abs": 1138, "len_title": 112, "len_sents": 37132, "len_sents_1": 28971, "n_sents": 278, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1112, "title": "Bridging the Gap Between f-divergences and Bayes Hilbert Spaces", "abs": "We introduce a novel framework that generalizes $f$-divergences by incorporating locally non-convex divergence-generating functions.\nUsing this extension, we define a new class of pseudo $f$-divergences, encompassing a wider range of distances between distributions that traditional $f$-divergences cannot capture.\nAmong these, we focus on a particular pseudo divergence obtained by considering the induced metric of Bayes Hilbert spaces.\nBayes Hilbert spaces are frequently used due to their inherent connection to Bayes's theorem. They allow sampling from potentially intractable posterior densities, which has remained challenging until now.\nIn the more general context, we prove that pseudo $f$-divergences are well-defined and introduce a variational estimation framework that can be used in a statistical learning context.\nBy applying this variational estimation framework to $f$-GANs, we achieve improved FID scores over existing $f$-GAN architectures and competitive results with the Wasserstein GAN, highlighting its potential for both theoretical research and practical applications in learning theory.", "keywords": [ "f-divergences", "Bayes Hilbert spaces", "fenchel conjugates", "centered-log-ratio", "posterior approximation" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "m51BgoqvbP", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "David Leo Wright Hall", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Researcher" }, { "name": "Jason S. Wang", "gender": "Male", "institution": "Stanford University", "country": "", "position": "MS student" }, { "name": "Kaiyue Wen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Percy Liang", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Tengyu Ma", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Stanford University" }, { "name": "Zhiyuan Li", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Assistant Professor" } ], "n_formula": 110, "n_formula_1": 14, "n_ref_uni": 53, "n_ref": 83, "n_ref_all": 131, "n_fig": 17, "n_tab": 6, "L_tab": 3194, "n_element_tab": 159, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 306, "n_element_tab_1": 17, "formula_len_all": 12851, "formula_len_all_1": 1309, "len_all": 319679, "len_all_1": 77357, "len_abs": 1930, "len_title": 133, "len_sents": 79300, "len_sents_1": 36472, "n_sents": 815, "n_sents_1": 291, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1938, "title": "Understanding Warmup-Stable-Decay Learning Rates: A River Valley Loss Landscape View", "abs": "Training language models currently requires pre-determining a fixed compute budget because the typical cosine learning rate schedule depends on the total number of steps. In contrast, the Warmup-Stable-Decay (WSD) schedule uses a constant learning rate to produce a main branch of iterates that can in principle continue indefinitely without a pre-specified compute budget. Then, given any compute budget, one can branch out from the main branch at a proper time with a rapidly decaying learning rate to produce a strong model. Empirically, WSD generates an intriguing, non-traditional loss curve: the loss remains elevated during the stable phase but sharply declines during the decay phase. Towards explaining this phenomenon, we conjecture that pretraining loss exhibits a river valley landscape, which resembles a deep valley with a river at its bottom. Under this assumption, we show that during the stable phase, the iterate undergoes large oscillations due to the high learning rate, yet it progresses swiftly along the river. During the decay phase, the rapidly dropping learning rate minimizes the iterate\u2019s oscillations, moving it closer to the river and revealing true optimization progress. Therefore, the sustained high learning rate phase and fast decaying phase are responsible for progress in the river and the mountain directions, respectively, and are both critical. Our analysis predicts phenomenons consistent with empirical observations and shows that this landscape can naturally emerge from pretraining on a simple bi-gram dataset. Inspired by the theory, we introduce WSD-S, a variant of WSD that reuses previous checkpoints\u2019 decay phases and keeps only one main branch, where we resume from a decayed checkpoint. WSD-S empirically outperforms WSD and Cyclic-Cosine in obtaining multiple pretrained language model checkpoints across various compute budgets in a single run for parameters scaling from 0.1B to 1.2B.", "keywords": [ "pretraining", "language model", "learning rate", "loss landscape", "manifold" ], "rating_list": [ 5, 3, 10, 6, 6 ], "soundness_list": [ 2, 2, 4, 2, 3 ], "presentation_list": [ 2, 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 4, 2, 2 ], "confidence_list": [ 4, 3, 3, 3, 2 ] }, { "paper_id": "m4eXBo0VNc", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Chao Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Han Qiu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Hao Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Hewu Li", "gender": "unknown", "institution": ", Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Jianshuo Dong", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ke Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Qi Li", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Qingjie Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Tianwei Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Ziyuan Zhang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 49, "n_ref": 88, "n_ref_all": 126, "n_fig": 7, "n_tab": 14, "L_tab": 8331, "n_element_tab": 619, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3707, "n_element_tab_1": 387, "formula_len_all": 587, "formula_len_all_1": 392, "len_all": 235317, "len_all_1": 72227, "len_abs": 1345, "len_title": 103, "len_sents": 70070, "len_sents_1": 30069, "n_sents": 584, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1366, "title": "An Engorgio Prompt Makes Large Language Model Babble on", "abs": "Auto-regressive large language models (LLMs) have yielded impressive performance in many real-world tasks. \nHowever, the new paradigm of these LLMs also exposes novel threats. \nIn this paper, we explore their vulnerability to inference cost attacks, where a malicious user crafts Engorgio prompts to intentionally increase the computation cost and latency of the inference process. We design Engorgio, a novel methodology, to efficiently generate adversarial Engorgio prompts to affect the target LLM's service availability. Engorgio has the following two technical contributions. \n(1) We employ a parameterized distribution to track LLMs' prediction trajectory. (2) Targeting the auto-regressive nature of LLMs' inference process, we propose novel loss functions to stably suppress the appearance of the token, whose occurrence will interrupt the LLM's generation process. \nWe conduct extensive experiments on 13 open-sourced LLMs with parameters ranging from 125M to 30B. \nThe results show that Engorgio prompts can successfully induce LLMs to generate abnormally long outputs (i.e., roughly 2-13$\\times$ longer to reach 90\\%+ of the output length limit)\nin a white-box scenario and our real-world experiment demonstrates Engergio's threat to LLM service with limited computing resources.\nThe code is released at https://github.com/jianshuod/Engorgio-prompt.", "keywords": [ "Large language model", "attack", "inference cost" ], "rating_list": [ 3, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "m3x4kDbYAK", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Hoi To Wai", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Jun Zhang", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Xiaolu Wang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuchang Sun", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 92, "n_formula_1": 31, "n_ref_uni": 45, "n_ref": 85, "n_ref_all": 111, "n_fig": 4, "n_tab": 5, "L_tab": 1556, "n_element_tab": 78, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 615, "n_element_tab_1": 98, "formula_len_all": 12554, "formula_len_all_1": 2303, "len_all": 185645, "len_all_1": 81928, "len_abs": 3907, "len_title": 122, "len_sents": 49318, "len_sents_1": 33310, "n_sents": 423, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1446, "title": "Incremental Aggregated Asynchronous SGD for Arbitrarily Heterogeneous Data", "abs": "We consider the distributed learning problem with data dispersed across multiple workers under the orchestration of a central server. Asynchronous Stochastic Gradient Descent (SGD) has been widely explored in such a setting to reduce the synchronization overhead associated with parallelization. However, prior works have shown that the performance of asynchronous SGD algorithms depends on a bounded dissimilarity condition among the workers' local data, a condition that can drastically affect their efficiency when the workers' data are highly heterogeneous. To overcome this limitation, we introduce the Incremental Aggregated Asynchronous SGD (IA$^2$SGD) algorithm. With a server-side buffer, IA$^2$SGD makes full use of stale stochastic gradients from all workers to neutralize the adverse effects of data heterogeneity. In an asynchronous implementation setting, the algorithm entails two distinct time lags in the model parameters and data samples utilized in the server's iterations. Furthermore, by adopting an incremental aggregation strategy, IA$^2$SGD maintains a per-iteration computational cost that is on par with traditional asynchronous SGD algorithms. Our analysis demonstrates that IA$^2$SGD achieves a consistent convergence rate for smooth nonconvex problems for arbitrarily heterogeneous data. Numerical experiments indicate that IA$^2$SGD compares favorably with existing asynchronous and synchronous SGD-based algorithms.", "keywords": [ "distributed optimization", "asynchronous SGD", "data heterogeneity" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "m3cKeqvC7z", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Chaohan Wang", "gender": "Male", "institution": "University of Adelaide", "country": "", "position": "PhD student" }, { "name": "Qi Chen", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Qi Wu", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "Associate Professor" }, { "name": "Yutong Xie", "gender": "Not Specified", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Yuyin Zhou", "gender": "unknown", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 21, "n_ref": 43, "n_ref_all": 61, "n_fig": 6, "n_tab": 7, "L_tab": 3728, "n_element_tab": 317, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2042, "n_element_tab_1": 262, "formula_len_all": 401, "formula_len_all_1": 69, "len_all": 114592, "len_all_1": 64693, "len_abs": 1807, "len_title": 112, "len_sents": 38426, "len_sents_1": 31821, "n_sents": 231, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1827, "title": "What can Mamba do for 3D Volumetric Medical Image Segmentation?", "abs": "Mamba, with its State Space Model (SSM), offers a more computationally efficient solution than Transformers for long-range dependency modeling. However, there is still a debate about its effectiveness in \\textit{high-resolution 3D medical image} segmentation. \nIn this study, we present a comprehensive investigation into Mamba's capabilities in 3D medical image segmentation by tackling three pivotal questions: Can Mamba replace Transformers? Can it elevate multi-scale representation learning? Is complex scanning necessary to unlock its full potential? We evaluate Mamba\u2019s performance across three large public benchmarks\u2014AMOS, TotalSegmentator, and BraTS. Our findings reveal that UlikeMamba, a U-shape Mamba-based network, consistently surpasses UlikeTrans, a U-shape Transformer-based network, particularly when enhanced with custom-designed 3D depthwise convolutions, boosting accuracy and computational efficiency. Further, our proposed multi-scale Mamba block demonstrates superior performance in capturing both fine-grained details and global context, especially in complex segmentation tasks, surpassing Transformer-based counterparts. We also critically assess complex scanning strategies, finding that simpler methods often suffice, while our Tri-scan approach delivers notable advantages in the most challenging scenarios. By integrating these advancements, we introduce a new network for 3D medical image segmentation, positioning Mamba as a transformative force that outperforms leading models such as nnUNet, CoTr, and U-Mamba, offering competitive accuracy with superior computational efficiency. This study provides key insights into Mamba's unique advantages, paving the way for more efficient and accurate approaches to 3D medical imaging. All code used in the experiments will be made publicly available.", "keywords": [ "3D Volumetric Medical Image Segmentation", "Mamba", "Transformer" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "m3KuuE2ozw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Cheng-Yuan Ho", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "Hebi Yang", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "Jui Chiu Chiang", "gender": "Female", "institution": "National Chung Cheng University", "country": "TW", "position": "Full Professor" }, { "name": "Wen-Hsiao Peng", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" }, { "name": "Yi-Hsin Chen", "gender": "Female", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "PhD student" }, { "name": "Yu Lun Liu", "gender": "unknown", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Assistant Professor" }, { "name": "Zhan Yu-Ting", "gender": "Female", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 18, "n_ref": 73, "n_ref_all": 111, "n_fig": 12, "n_tab": 14, "L_tab": 9498, "n_element_tab": 1750, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 860, "formula_len_all_1": 660, "len_all": 146049, "len_all_1": 54371, "len_abs": 2507, "len_title": 141, "len_sents": 38296, "len_sents_1": 25093, "n_sents": 323, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1562, "title": "CAT-3DGS: A Context-Adaptive Triplane Approach to Rate-Distortion-Optimized 3DGS Compression", "abs": "3D Gaussian Splatting (3DGS) has recently emerged as a promising 3D representation. Much research has been focused on reducing its storage requirements and memory footprint. However, the needs to compress and transmit the 3DGS representation to the remote side are overlooked. This new application calls for rate-distortion-optimized 3DGS compression. How to quantize and entropy encode sparse Gaussian primitives in the 3D space remains largely unexplored. Few early attempts resort to the hyperprior framework from learned image compression. But, they fail to utilize fully the inter and intra correlation inherent in Gaussian primitives. Built on ScaffoldGS, this work, termed CAT-3DGS, introduces a context-adaptive triplane approach to their rate-distortion-optimized coding. It features multi-scale triplanes, oriented according to the principal axes of Gaussian primitives in the 3D space, to capture their inter correlation (i.e. spatial correlation) for spatial autoregressive coding in the projected 2D planes. With these triplanes serving as the hyperprior, we further perform channel-wise autoregressive coding to leverage the intra correlation within each individual Gaussian primitive. Our CAT-3DGS incorporates a view frequency-aware masking mechanism. It actively skips from coding those Gaussian primitives that potentially have little impact on the rendering quality. When trained end-to-end to strike a good rate-distortion trade-off, our CAT-3DGS achieves the state-of-the-art compression performance on the commonly used real-world datasets.", "keywords": [ "3D Gaussian Splatting", "Rate-Distortion Optimization", "Context Models", "Grid Representations" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "jvmMqD57ZR", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Jun-cheng Chen", "gender": "Male", "institution": "Research Center for Information Technology Innovation, Academia Sinica", "country": "TW", "position": "Associate Research Fellow/Professor" }, { "name": "Shang-Tse Chen", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Associate Professor" }, { "name": "Wa-Kin Lei", "gender": "Male", "institution": "Department of computer science and informational engineering, National Taiwan University", "country": "TW", "position": "MS student" } ], "n_formula": 19, "n_formula_1": 9, "n_ref_uni": 37, "n_ref": 78, "n_ref_all": 101, "n_fig": 8, "n_tab": 6, "L_tab": 4891, "n_element_tab": 723, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 2301, "n_element_tab_1": 216, "formula_len_all": 1659, "formula_len_all_1": 957, "len_all": 117770, "len_all_1": 50880, "len_abs": 1192, "len_title": 103, "len_sents": 24557, "len_sents_1": 20271, "n_sents": 180, "n_sents_1": 142, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1196, "title": "DRAG: Data Reconstruction Attack using Guided Diffusion", "abs": "With the rise of large foundation models, split inference (SI) has emerged as a popular computational paradigm for deploying models across lightweight edge devices and cloud servers, addressing both data privacy and computational cost concerns. However, most existing data reconstruction attacks have focused on smaller classification models like ResNet, leaving the privacy risks of foundation models in SI settings largely unexplored. To address this gap, we propose a novel data reconstruction attack based on guided diffusion, which leverages the rich prior knowledge embedded in a latent diffusion model (LDM) pretrained on a large-scale dataset. Our method performs iterative reconstruction on the LDM\u2019s learned image manifold, effectively generating high-fidelity images closely resembling the original data from their intermediate representations (IR). Extensive experiments demonstrate that our approach significantly outperforms state-of-the-art methods, both qualitatively and quantitatively, in reconstructing data from deep-layer IRs of the vision foundation model. The results highlight the urgent need for more robust privacy protection mechanisms for large models in SI scenarios.", "keywords": [ "Data Reconstruction Attack", "Privacy", "Diffusion Model" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 2, 3, 2, 3 ] }, { "paper_id": "jvRCirB0Oq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Ani Nenkova", "gender": "Female", "institution": "Adobe Research", "country": "US", "position": "Principal Scientist" }, { "name": "Byron C Wallace", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "Assistant Professor" }, { "name": "Chantal Shaib", "gender": "Female", "institution": "Northeastern University", "country": "", "position": "PhD student" }, { "name": "Jiuding Sun", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Joe Barrow", "gender": "Male", "institution": "Pattern Data", "country": "", "position": "Researcher" }, { "name": "Alexa F Siu", "gender": "Female", "institution": "Adobe", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 26, "n_ref": 52, "n_ref_all": 77, "n_fig": 3, "n_tab": 15, "L_tab": 4198, "n_element_tab": 489, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 3970, "n_element_tab_1": 432, "formula_len_all": 474, "formula_len_all_1": 401, "len_all": 143755, "len_all_1": 64057, "len_abs": 1015, "len_title": 129, "len_sents": 30240, "len_sents_1": 24460, "n_sents": 260, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1030, "title": "Standardizing the Measurement of Text Diversity: A Tool and Comparative Analysis", "abs": "The diversity across outputs generated by LLMs shapes perception of their quality and utility. \nAchieving high textual diversity in datasets is often a desired quality, but there is no standard method to measure this aspect of model behaviour.\nIn this work we empirically investigate diversity scores on English texts and measure how much overlapping information is captured in these metrics. \nWe find that computationally efficient compression algorithms capture information similar to what is measured by slow-to-compute $n$-gram overlap homogeneity scores. \nFurther, a combination of measures---compression ratios, self-repetition of long $n$-grams and Self-BLEU and BERTScore---are sufficient to report, as they have low mutual correlation with each other. \nThe applicability of scores extends beyond analysis of generative models; for example, we highlight applications on instruction-tuning datasets and human-produced texts. \nWe release a diversity score package to facilitate research and invite consistency going forward.", "keywords": [ "text diversity", "summarization", "generation" ], "rating_list": [ 5, 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "jv2zHOalpL", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "R Yu", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Xinchao Wang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 3, "n_ref_uni": 51, "n_ref": 93, "n_ref_all": 121, "n_fig": 7, "n_tab": 9, "L_tab": 5204, "n_element_tab": 324, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1057, "n_element_tab_1": 163, "formula_len_all": 983, "formula_len_all_1": 228, "len_all": 202836, "len_all_1": 67883, "len_abs": 0, "len_title": 172, "len_sents": 64201, "len_sents_1": 28685, "n_sents": 514, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1508, "title": "Neural Phylogeny: Fine-Tuning Relationship Detection among Neural Networks", "abs": "Given a collection of neural networks, can we determine which are parent models and which are child models fine-tuned from the parents?\nIn this work, we strive to answer this question\nvia introducing a new task termed as neural phylogeny detection, aimed at identifying the existence and direction of the fine-tuning relationship. Specifically, neural phylogeny detection attempts to identify all parent-child model pairs and determine, within each pair, which model is the parent and which is the child.\nWe present two approaches for neural phylogeny detection: a learning-free method and a learning-based method. First, we propose a metric that leverages the distance from network parameters to a fake initialization to infer fine-tuning directions. By integrating this metric with traditional clustering algorithms, we propose a series of efficient, learning-free neural phylogeny detection methods. Second, we introduce a transformer-based neural phylogeny detector, which significantly enhances detection accuracy through a learning-based manner. Extensive experiments, ranging from shallow fully-connected networks to open-sourced Stable Diffusion and LLaMA models, progressively validate the effectiveness of both methods. The results demonstrate the reliability of both the learning-free and the learning-based approaches across various learning tasks and network architectures, as well as their ability to detect cross-generational phylogeny between ancestor models and their fine-tuned descendants.", "keywords": [ "Neural Phylogeny", "Finetuning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "juxbsQEuTZ", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "David Alvarez-Melis", "gender": "Male", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "US", "position": "Assistant Professor" }, { "name": "Naomi Saphra", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Fellow" }, { "name": "Tian Qin", "gender": "Female", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 73, "n_ref_all": 106, "n_fig": 12, "n_tab": 4, "L_tab": 3056, "n_element_tab": 226, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 805, "n_element_tab_1": 128, "formula_len_all": 95, "formula_len_all_1": 0, "len_all": 164991, "len_all_1": 64736, "len_abs": 1808, "len_title": 119, "len_sents": 55003, "len_sents_1": 32037, "n_sents": 410, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1661, "title": "Sometimes I am a Tree: Data Drives Unstable Hierarchical Generalization", "abs": "Neural networks often favor shortcut heuristics based on surface-level patterns. Language models (LMs), for example, behave like n-gram models early in training. However, to correctly apply grammatical rules, LMs must instead rely on hierarchical syntactic representations rather than on surface-level heuristics derived from n-grams. In this work, we use cases studies of English grammar to explore how latent structures in training data drives models toward improved out-of-distribution (OOD) generalization. We then investigate how data composition can lead to inconsistent behavior across random seeds. Our results show that models stabilize in their OOD behavior only when they commit to either a surface-level linear rule or a hierarchical rule. The hierarchical rule, furthermore, is induced by grammatically complex sequences with deep embedding structures, whereas the linear rule is induced by simpler sequences. When the data contains a mix of simple and complex examples, potential rules compete; each independent training run either stabilizes by committing to a single rule or remains unstable in its OOD behavior. We also identify an exception to the relationship between stability and generalization: Models which memorize patterns from homogeneous training data can overfit stably, with different rules for memorized and unmemorized patterns. While existing works have attributed similar generalization behavior to training objective and model architecture, our findings emphasize the critical role of training data in shaping generalization patterns and how competition between data subsets contributes to inconsistent generalization outcomes.", "keywords": [ "Language models", "simplicity bias", "random variations", "OOD generalization" ], "rating_list": [ 3, 8, 5 ], "soundness_list": [ 2, 4, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 1, 4, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "juZNRYRMPn", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Jon Kleinberg", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Full Professor" }, { "name": "Yanbang Wang", "gender": "unknown", "institution": "Department of Computer Science, Cornell University", "country": "", "position": "PhD student" }, { "name": "Yanhong Wu", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" } ], "n_formula": 29, "n_formula_1": 18, "n_ref_uni": 40, "n_ref": 77, "n_ref_all": 80, "n_fig": 0, "n_tab": 3, "L_tab": 1059, "n_element_tab": 76, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 3177, "n_element_tab_1": 101, "formula_len_all": 3580, "formula_len_all_1": 1600, "len_all": 130650, "len_all_1": 67980, "len_abs": 1151, "len_title": 127, "len_sents": 36779, "len_sents_1": 28765, "n_sents": 282, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1160, "title": "Truth-Guided Negative Sampling in Self-supervised Graph Representation Learning", "abs": "Negative sampling is an important yet challenging component in self-supervised graph representation learning, particularly for recommendation systems where user-item interactions are modeled as bipartite graphs. Existing methods often rely on heuristics or human-specified principles to design negative sampling distributions. This potentially overlooks the usage of an underlying ``true'' negative distribution, which we might be able to access as an oracle despite not knowing its exact form. \nIn this work, we shift the focus from manually designing negative sampling distributions to a method that approximates and leverages the underlying true distribution. We expand this idea in the analysis of two scenarios: (1) when the observed graph is an unbiased sample from the true distribution, and (2) when the observed graph is biased with partially observable positive edges. The analysis result is the derivation of a sampling strategy as the numerical approximation of a well-established learning objective. Our theoretical findings are also empirically validated, and our new sampling methods achieve state-of-the-art performance on real-world datasets.", "keywords": [ "graph representation learning", "negative sampling", "recommendation system" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 4, 2, 1, 3 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "juKVq5dWTR", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Changick Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Hongsin Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Seungju Cho", "gender": "Male", "institution": "\tKorea Advanced Institute of Science and Technology ", "country": "KR", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 13, "n_ref_uni": 50, "n_ref": 123, "n_ref_all": 150, "n_fig": 7, "n_tab": 18, "L_tab": 11792, "n_element_tab": 897, "n_fig_1": 5, "n_tab_1": 10, "L_tab_1": 4946, "n_element_tab_1": 271, "formula_len_all": 1124, "formula_len_all_1": 554, "len_all": 177513, "len_all_1": 62280, "len_abs": 1033, "len_title": 110, "len_sents": 41708, "len_sents_1": 24453, "n_sents": 308, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1043, "title": "Indirect Gradient Matching for Adversarial Robust Distillation", "abs": "Adversarial training significantly improves adversarial robustness, but superior performance is primarily attained with large models. \nThis substantial performance gap for smaller models has spurred active research into adversarial distillation (AD) to mitigate the difference. \nExisting AD methods leverage the teacher\u2019s logits as a guide.\nIn contrast to these approaches, we aim to transfer another piece of knowledge from the teacher, the input gradient.\nIn this paper, we propose a distillation module termed Indirect Gradient Distillation Module (IGDM) that indirectly matches the student\u2019s input gradient with that of the teacher.\nExperimental results show that IGDM seamlessly integrates with existing AD methods, significantly enhancing their performance.\nParticularly, utilizing IGDM on the CIFAR-100 dataset improves the AutoAttack accuracy from 28.06\\% to 30.32\\% with the ResNet-18 architecture and from 26.18\\% to 29.32\\% with the MobileNetV2 architecture when integrated into the SOTA method without additional data augmentation.", "keywords": [ "Adversarial Robustness", "Adversarial Training", "Adversarial Distillation" ], "rating_list": [ 5, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "ju4EwaLeoI", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Douglas J Weber", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Jehan Yang", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Lihong Jin", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Researcher" }, { "name": "Zackory Erickson", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 22, "n_ref": 35, "n_ref_all": 43, "n_fig": 1, "n_tab": 2, "L_tab": 4477, "n_element_tab": 170, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 3967, "n_element_tab_1": 183, "formula_len_all": 134, "formula_len_all_1": 134, "len_all": 101763, "len_all_1": 56778, "len_abs": 1313, "len_title": 124, "len_sents": 25090, "len_sents_1": 24954, "n_sents": 132, "n_sents_1": 132, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1319, "title": "Ref-EMGBench: Benchmarking Reference Normalization for Electromyography Data", "abs": "Electromyography (EMG)-based hand gesture recognition is essential for applications in prosthetics, rehabilitation, and human-robot interaction. Despite advances in machine learning, domain shift caused by intersubject variability often leads to degraded model performance when applying trained models to new users. In this study, we revisit the statistical reference normalization methods to mitigate the domain shift in EMG data in a leave-one-subject-out train-test split setting. We systematically benchmark five popular amplitude-based normalization techniques to assess their effectiveness in subject-specific classification with varied datasets and percentages for normalization. Experimental results show that Min-Max and Peak normalization outperform others, yielding higher classification accuracy on EMG data. We further visualize the domain shifts in the feature space throughout the training process and provide an analysis based on EMG signal characteristics. Our findings indicate that proper normalization significantly reduces inter-subject variability of EMG samples, enhancing model adaptation and providing insights for bridging domain shifts in future EMG-based gesture recognition research. The benchmark code for domain adaptation approaches on EMG signals is available at ref-emgbench.github.io.", "keywords": [ "EMG", "reference normalization", "domain adaptation" ], "rating_list": [ 1, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 1, 3, 2, 1 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "jtY2OHKj4a", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "KUI XUE", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "", "position": "Researcher" }, { "name": "Linjie Mu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Shaoting Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Full Professor" }, { "name": "Tong Ruan", "gender": "Female", "institution": "East China University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xiaofan Zhang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yongqi Fan", "gender": "Male", "institution": "East China University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Zhongzhen Huang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "LiuRuoyu", "gender": "Non-Binary", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 36, "n_ref": 61, "n_ref_all": 76, "n_fig": 5, "n_tab": 3, "L_tab": 1600, "n_element_tab": 209, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 568, "n_element_tab_1": 143, "formula_len_all": 502, "formula_len_all_1": 184, "len_all": 156857, "len_all_1": 71325, "len_abs": 1988, "len_title": 143, "len_sents": 44536, "len_sents_1": 34242, "n_sents": 340, "n_sents_1": 262, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1666, "title": "Tool Calling: Enhancing Medication Consultation via Retrieval-Augmented Large Language Models", "abs": "Large-scale language models (LLMs) have achieved remarkable success across various language tasks but suffer from hallucinations and temporal misalignment. To mitigate these shortcomings, Retrieval-augmented generation (RAG) has been utilized to provide external knowledge to facilitate the answer generation. However, applying such models to the medical domain faces several challenges due to the lack of domain-specific knowledge and the intricacy of real-world scenarios.\nIn this study, we explore LLMs with RAG framework for knowledge-intensive tasks in the medical field. To evaluate the capabilities of LLMs, we introduce MedicineQA, a multi-round dialogue benchmark that simulates the real-world medication consultation scenario and requires LLMs to answer with retrieved evidence from the medicine database. MedicineQA contains 300 multi-round question-answering pairs, each embedded within a detailed dialogue history, highlighting the challenge posed by this knowledge-intensive task to current LLMs. We further propose a new \\textit{Distill-Retrieve-Read} framework instead of the previous \\textit{Retrieve-then-Read}. Specifically, the distillation and retrieval process utilizes a tool calling mechanism to formulate search queries that emulate the keyword-based inquiries used by search engines. With experimental results, we show that our framework brings notable performance improvements and surpasses the previous counterparts in the evidence retrieval process in terms of evidence retrieval accuracy. This advancement underscores the framework's potential to effectively address the inherent challenges of applying RAG models to the medical domain.", "keywords": [ "Large Language Models", "Medication Consultation", "Retrieval-Augmented Generation", "Tool Call" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 4, 2, 2 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "jt8wI3ZzXG", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Henry Mason", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Irina Belousova", "gender": "Not Specified", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Mahyar Najibi", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Mohammad Rastegari", "gender": "Male", "institution": "Apple", "country": "", "position": "ML manager" }, { "name": "Nikhil Bhendawade", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Qichen Fu", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 37, "n_ref": 86, "n_ref_all": 131, "n_fig": 22, "n_tab": 5, "L_tab": 2224, "n_element_tab": 299, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 2389, "n_element_tab_1": 226, "formula_len_all": 748, "formula_len_all_1": 623, "len_all": 160083, "len_all_1": 64827, "len_abs": 2503, "len_title": 114, "len_sents": 52153, "len_sents_1": 28362, "n_sents": 359, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1659, "title": "Speculative Streaming: Fast LLM Inference without Auxiliary Models", "abs": "Speculative decoding is a prominent technique to accelerate large language model inference by leveraging predictions from an auxiliary draft model. While effective, in application-specific settings, it often involves fine-tuning both draft and target models to achieve high acceptance rates. As the number of downstream tasks grows, draft models add significant complexity to inference systems. Recently\nseveral single model architectures viz. Medusa have been proposed to speculate tokens in non-autoregressive manner, however, their effectiveness is limited due to lack of dependency between speculated tokens. We introduce a novel speculative decoding method that integrates drafting within the target model by using Multi-stream attention and incorporates future token planning into supervised finetuning objective. To the best of our knowledge, this is the first parameter-efficient approach that scales well with an increasing number of downstream tasks while enhancing downstream metrics and achieving high acceptance rates, attributable to the interdependence among the speculated tokens. Speculative Streaming speeds up decoding by 1.9 - 3X in a diverse set of tasks, such as Summarization, Structured Queries, and Meaning Representation, while improving generation quality and using \u223c10000X fewer extra parameters than alternative architectures, making it ideal for resource-constrained devices. Our approach can also be effectively deployed in\nlossless settings for generic chatbot applications that do not necessitate supervised fine-tuning. In such setups, we achieve 2.9 - 3.2X speedup while maintaining the integrity of the base model\u2019s output.", "keywords": [ "Speculative Decoding", "Efficient LLM", "inference optimization", "Multi-stream attention" ], "rating_list": [ 3, 8, 6, 5 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "jt1h2dnmng", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Dongmei Jiang", "gender": "Female", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Hanqing Zhao", "gender": "Male", "institution": "Pengcheng Laboratory", "country": "", "position": "PhD student" }, { "name": "Haoxiang Li", "gender": "Male", "institution": "Pixocial.com", "country": "US", "position": "Chief Scientist" }, { "name": "Wenqing Zhang", "gender": "Female", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Xiao Dong", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Xiaodan Liang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Zheng Chong", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "shiyue Zhang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "xujie zhang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 44, "n_ref": 101, "n_ref_all": 125, "n_fig": 9, "n_tab": 7, "L_tab": 2340, "n_element_tab": 156, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 639, "n_element_tab_1": 41, "formula_len_all": 338, "formula_len_all_1": 272, "len_all": 127201, "len_all_1": 43683, "len_abs": 565, "len_title": 127, "len_sents": 32082, "len_sents_1": 19668, "n_sents": 231, "n_sents_1": 139, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1742, "title": "CatVTON: Concatenation Is All You Need for Virtual Try-On with Diffusion Models", "abs": "Virtual try-on methods based on diffusion models achieve realistic effects but often require additional encoding modules, a large number of training parameters, and complex preprocessing, which increases the burden on training and inference. In this work, we re-evaluate the necessity of additional modules and analyze how to improve training efficiency and reduce redundant steps in the inference process. Based on these insights, we propose CatVTON, a simple and efficient virtual try-on diffusion model that transfers in-shop or worn garments of arbitrary categories to target individuals by concatenating them along spatial dimensions as inputs of the diffusion model. The efficiency of CatVTON is reflected in three aspects: (1) Lightweight network. CatVTON consists only of a VAE and a simplified denoising UNet, removing redundant image and text encoders as well as cross-attentions, and includes just 899.06M parameters. (2) Parameter-efficient training. Through experimental analysis, we identify self-attention modules as crucial for adapting pre-trained diffusion models to the virtual try-on task, enabling high-quality results with only 49.57M training parameters. (3) Simplified inference. CatVTON eliminates unnecessary preprocessing, such as pose estimation, human parsing, and captioning, requiring only a person image and garment reference to guide the virtual try-on process, reducing over 49% memory usage compared to other diffusion-based methods. Extensive experiments demonstrate that CatVTON achieves superior qualitative and quantitative results compared to baseline methods and demonstrates strong generalization performance in in-the-wild scenarios, despite being trained solely on public datasets with 73K samples.", "keywords": [ "diffusion models", "virtual try-on", "parameter-efficient training" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "jsVehKnSj4", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Changxiao Cai", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" }, { "name": "Jiacheng Zhang", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "Postdoc" } ], "n_formula": 185, "n_formula_1": 29, "n_ref_uni": 32, "n_ref": 64, "n_ref_all": 102, "n_fig": 5, "n_tab": 3, "L_tab": 834, "n_element_tab": 60, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 173, "n_element_tab_1": 3, "formula_len_all": 15694, "formula_len_all_1": 2068, "len_all": 213541, "len_all_1": 79024, "len_abs": 1143, "len_title": 92, "len_sents": 54619, "len_sents_1": 30192, "n_sents": 581, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1148, "title": "Minimax-optimal trust-aware multi-armed bandits", "abs": "Multi-armed bandit (MAB) algorithms have achieved significant success in sequential decision-making applications, under the premise that humans perfectly implement the recommended policy. However, existing methods often overlook the crucial factor of human trust in learning algorithms. When trust is lacking, humans may deviate from the recommended policy, leading to undesired learning performance. Motivated by this gap, we study the trust-aware MAB problem by integrating a dynamic trust model into the standard MAB framework. Specifically, it assumes that the recommended and actually implemented policy differs depending on human trust, which in turn evolves with the quality of the recommended policy. We establish the minimax regret in the presence of the trust issue and demonstrate the suboptimality of vanilla MAB algorithms such as the upper confidence bound (UCB) algorithm. To overcome this limitation, we introduce a novel two-stage trust-aware procedure that provably attains near-optimal statistical guarantees. A simulation study is conducted to illustrate the benefits of our proposed algorithm when dealing with the trust issue.", "keywords": [ "multi-armed bandit", "trust-aware decision-making", "regret bound", "minimax optimality" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "jsBhmOCKYs", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Chen Change Loy", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Kang Liao", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Zongsheng Yue", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhouxia Wang", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 83, "n_ref": 155, "n_ref_all": 183, "n_fig": 17, "n_tab": 7, "L_tab": 2105, "n_element_tab": 247, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 713, "n_element_tab_1": 81, "formula_len_all": 287, "formula_len_all_1": 248, "len_all": 198685, "len_all_1": 68315, "len_abs": 1687, "len_title": 124, "len_sents": 49812, "len_sents_1": 33234, "n_sents": 336, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1651, "title": "Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration", "abs": "Although learning-based image restoration methods have made significant progress, they still struggle with limited generalization to real-world scenarios due to the substantial domain gap caused by training on synthetic data. Existing methods address this issue by improving data synthesis pipelines, estimating degradation kernels, employing deep internal learning, and performing domain adaptation and regularization. Previous domain adaptation methods have sought to bridge the domain gap by learning domain-invariant knowledge in either feature or pixel space. However, these techniques often struggle to extend to low-level vision tasks within a stable and compact framework. In this paper, we show that it is possible to perform domain adaptation via the noise space using diffusion models. In particular, by leveraging the unique property of how auxiliary conditional inputs influence the multi-step denoising process, we derive a meaningful *diffusion loss* that guides the restoration model in progressively aligning both restored synthetic and real-world outputs with a target clean distribution. We refer to this method as *denoising as adaptation*. To prevent shortcuts during joint training, we present crucial strategies such as channel-shuffling layer and residual-swapping contrastive learning in the diffusion model. They implicitly blur the boundaries between conditioned synthetic and real data and prevent the reliance of the model on easily distinguishable features. Experimental results on three classical image restoration tasks, namely denoising, deblurring, and deraining, demonstrate the effectiveness of the proposed method.", "keywords": [ "Image Restoration", "Domain Adaptation", "Diffusion Loss" ], "rating_list": [ 6, 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 3, 2 ] }, { "paper_id": "jrY83wPzMJ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Qiuyue Ma", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Yaping Zhang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Associate Professor" }, { "name": "Yu Zhou", "gender": "Female", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 37, "n_ref": 77, "n_ref_all": 94, "n_fig": 4, "n_tab": 8, "L_tab": 1088, "n_element_tab": 98, "n_fig_1": 3, "n_tab_1": 9, "L_tab_1": 1770, "n_element_tab_1": 205, "formula_len_all": 189, "formula_len_all_1": 189, "len_all": 122127, "len_all_1": 56721, "len_abs": 1112, "len_title": 47, "len_sents": 28855, "len_sents_1": 24155, "n_sents": 205, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1117, "title": "Synchronous Scene Text Spotting and Translating", "abs": "Text image machine translation aims to translate the content of textual regions in images from a source language to a target language. Compared with traditional document, images captured in natural scenes have more diverse text and more complex layout, posing challenges in recognizing text content and predicting reading order within each text region. Current methods mainly adopt pipeline pattern, in which models for text spotting and translating are trained separately. In this pattern, translation performance is affected by propagation of mispredicted reading order and text recognition errors. In this paper, we propose a scene text image machine translation approach by implementation of synchronous text spotting and translating. A bridge and fusion module is introduced to make better use of multi-modal feature. Besides, we create datasets for both Chinese-to-English and English-to-Chinese image translation. Experimental results substantiate that our method achieves state-of-the-art translation performance in scene text field, proving the effectiveness of joint learning and multi-modal feature fusion.", "keywords": [ "multimodal machine translation" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "jrKPOQBq9i", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Weiqi Guan", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Zihao Shi", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" } ], "n_formula": 66, "n_formula_1": 36, "n_ref_uni": 48, "n_ref": 60, "n_ref_all": 65, "n_fig": 1, "n_tab": 7, "L_tab": 2441, "n_element_tab": 213, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1068, "n_element_tab_1": 80, "formula_len_all": 7243, "formula_len_all_1": 3065, "len_all": 142589, "len_all_1": 59090, "len_abs": 926, "len_title": 112, "len_sents": 31109, "len_sents_1": 22732, "n_sents": 332, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 929, "title": "PoincareNorm: Rethinking Over-smoothing beyond Dirichlet energy", "abs": "Dirichlet energy is intuitive and commonly used to measure over-smoothing. However, Dirichlet energy can only capture information about the first-order derivative of features. In light of this, we propose a series of node similarity measures which are the energy of higher-order derivatives of features and generalize Dirichlet energy. After we rigorously analyze the property of proposed measures and its application to establish the sharp decay rate of Dirichlet energy under continuous diffusion or discrete random walk which is closely related to the first nonzero eigenvalue of graph Laplacian. Lastly, to address over-smoothing with respect to these measures, we propose a normalization termed PoincareNorm which generalizes PairNorm to control our proposed measures. We consider the semi-supervised node classification task in the scenario without missing features, PoincareNorm outperforms existing normalization methods.", "keywords": [ "over-smoothing", "node similarity measure\uff0cnormalization\u3002" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 1, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "jqx5XI4Yr3", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Chao Wang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Hehe Fan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yifan Sun", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Senior Expert" }, { "name": "Zhedong Zheng", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 1, "n_ref_uni": 70, "n_ref": 132, "n_ref_all": 148, "n_fig": 3, "n_tab": 6, "L_tab": 4056, "n_element_tab": 367, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2820, "n_element_tab_1": 377, "formula_len_all": 684, "formula_len_all_1": 55, "len_all": 224954, "len_all_1": 69206, "len_abs": 1241, "len_title": 152, "len_sents": 58612, "len_sents_1": 30775, "n_sents": 440, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 1174, "title": "ProteinAdapter: Adapting Pre-trained Large Protein Models for Efficient Protein Representation Learning", "abs": "The study of proteins is crucial in various scientific disciplines, but understanding their intricate multi-level relationships remains challenging. Inspired by the sequence and structure understanding of Large Protein Models (LPMs), we introduce a new ProteinAdapter, to efficiently transfer the broad knowledge encapsulated in multiple LPMs, e.g., ESM-1b, to task-specific insights. ProteinAdapter could largely save labor-intensive analysis on the 3D position and the amino acid order. Specifically, (1) with a modest number of additional parameters, ProteinAdapter facilitates multi-level protein representation learning by integrating both sequence and geometric structure embeddings from LPMs; (2) based on the learned embedding, we further scale up the proposed ProteinAdapter to various tasks with a unified Multi-Scale Predictor, which optimally harnesses the learned embeddings through task-specific attention. Albeit simple, the proposed method is scalable to multiple downstream tasks without bells and whistles. Extensive experiments on over 20 tasks show that ProteinAdapter outperforms state-of-the-art methods under both single-task and multi-task scenarios.", "keywords": [ "Protein Representation", "Structured State Space Models" ], "rating_list": [ 3, 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "jqmptcSNVG", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 10, "author_info_list": [ { "name": "Chaoran Cheng", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" }, { "name": "Ge Liu", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Jiahan Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Jian Peng", "gender": "Male", "institution": "Univ. of Illinois at Urbana-Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Jianzhu Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jiaqi Guan", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Ruihan Guo", "gender": "Male", "institution": "Helixon. Inc", "country": "CN", "position": "Researcher" }, { "name": "Sheng Wang", "gender": "unknown", "institution": "University of Washington, Seattle", "country": "US", "position": "Assistant Professor" }, { "name": "Shitong Luo", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Tong Chen", "gender": "Male", "institution": "University of Washington, Seattle", "country": "", "position": "PhD student" } ], "n_formula": 62, "n_formula_1": 28, "n_ref_uni": 68, "n_ref": 113, "n_ref_all": 137, "n_fig": 5, "n_tab": 5, "L_tab": 1623, "n_element_tab": 232, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 814, "n_element_tab_1": 140, "formula_len_all": 3535, "formula_len_all_1": 1823, "len_all": 315269, "len_all_1": 63766, "len_abs": 1693, "len_title": 121, "len_sents": 64349, "len_sents_1": 29153, "n_sents": 485, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1528, "title": "Hotspot-Driven Peptide Design via Multi-Fragment Autoregressive Extension", "abs": "Peptides, short chains of amino acids, interact with target proteins, making them a unique class of protein-based therapeutics for treating human diseases. Recently, deep generative models have shown great promise in peptide generation. However, several challenges remain in designing effective peptide binders. First, not all residues contribute equally to peptide-target interactions. Second, the generated peptides must adopt correct geometries due to the constraints of peptide bonds. Third, realistic benchmarks for peptide drug development are still lacking.\nTo address these challenges, we introduce **PepHAR**, a hot-spot-driven autoregressive generative model for designing peptides targeting specific proteins. Building on the observation that certain hot spot residues have higher interaction potentials, we first use an energy-based density model to fit and sample these key residues. Next, to ensure proper peptide geometry, we autoregressively extend peptide fragments by estimating dihedral angles between residue frames. Finally, we apply an optimization process to iteratively refine fragment assembly, ensuring valid peptide structures.\nBy combining hot spot sampling with fragment-based extension, our approach enables de novo peptide design tailored to a target protein and allows the incorporation of key hot spot residues into peptide scaffolds. Extensive experiments, including peptide design and peptide scaffold generation, demonstrate the strong potential of **PepHAR** in computational peptide design.", "keywords": [ "AI for Science; Protein Design; Generative Models;" ], "rating_list": [ 8, 6, 5, 6, 6 ], "soundness_list": [ 4, 2, 2, 4, 4 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 4, 3, 2, 3, 3 ], "confidence_list": [ 3, 5, 4, 4, 4 ] }, { "paper_id": "jqff3wzkLT", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Jiachen Zhu", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Katrina Evtimova", "gender": "Female", "institution": "New York University", "country": "US", "position": "Researcher" }, { "name": "Yann LeCun", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Director" }, { "name": "Yubei Chen", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Assistant Professor" }, { "name": "ravid ziv", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 46, "n_ref": 78, "n_ref_all": 95, "n_fig": 4, "n_tab": 7, "L_tab": 2357, "n_element_tab": 123, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1316, "n_element_tab_1": 70, "formula_len_all": 914, "formula_len_all_1": 903, "len_all": 143445, "len_all_1": 62618, "len_abs": 1342, "len_title": 115, "len_sents": 42707, "len_sents_1": 28708, "n_sents": 297, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1348, "title": "Variance-Covariance Regularization Improves Representation Learning", "abs": "Transfer learning plays a key role in advancing machine learning models, yet conventional supervised pretraining often undermines feature transferability by prioritizing features that minimize the pretraining loss. In this work, we adapt a self-supervised learning regularization technique from the VICReg method to supervised learning contexts, introducing Variance-Covariance Regularization (VCReg). This adaptation encourages the network to learn high-variance, low-covariance representations, promoting learning more diverse features. We outline best practices for an efficient implementation of our framework, including applying it to the intermediate representations. Through extensive empirical evaluation, we demonstrate that our method significantly enhances transfer learning for images and videos, achieving state-of-the-art performance across numerous tasks and datasets. VCReg also improves performance in scenarios like long-tail learning and hierarchical classification. Additionally, we show its effectiveness may stem from its success in addressing challenges like gradient starvation and neural collapse. In summary, VCReg offers a universally applicable regularization framework that significantly advances transfer learning and highlights the connection between gradient starvation, neural collapse, and feature transferability.", "keywords": [ "Regularization", "Variance-Covariance Regularization" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "jqVj8vCQsT", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Emmanuel de Bezenac", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Researcher" }, { "name": "Lise Le Boudec", "gender": "Female", "institution": "Universit\u00e9 Pierre et Marie Curie - Paris 6, Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "FR", "position": "PhD student" }, { "name": "Louis Serrano", "gender": "Male", "institution": "Universit\u00e9 Pierre et Marie Curie - Paris 6, Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "", "position": "PhD student" }, { "name": "Ramon Daniel Regueiro-Espino", "gender": "Male", "institution": "Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "FR", "position": "PhD student" }, { "name": "Yuan Yin", "gender": "Male", "institution": "Valeo", "country": "FR", "position": "Postdoc" }, { "name": "patrick gallinari", "gender": "Male", "institution": "Criteo AI Lab", "country": "FR", "position": "Research Scientist" } ], "n_formula": 74, "n_formula_1": 18, "n_ref_uni": 38, "n_ref": 103, "n_ref_all": 145, "n_fig": 12, "n_tab": 15, "L_tab": 2952, "n_element_tab": 401, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 511, "n_element_tab_1": 61, "formula_len_all": 4505, "formula_len_all_1": 1664, "len_all": 225609, "len_all_1": 66586, "len_abs": 1250, "len_title": 127, "len_sents": 70099, "len_sents_1": 30039, "n_sents": 627, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1254, "title": "Learning a Neural Solver for Parametric PDE to Enhance Physics-Informed Methods", "abs": "Physics-informed deep learning often faces optimization challenges due to the complexity of solving partial differential equations (PDEs), which involve exploring large solution spaces, require numerous iterations, and can lead to unstable training. These challenges arise particularly from the ill-conditioning of the optimization problem, caused by the differential terms in the loss function. To address these issues, we propose learning a solver, i.e., solving PDEs using a physics-informed iterative algorithm trained on data. Our method learns to condition a gradient descent algorithm that automatically adapts to each PDE instance, significantly accelerating and stabilizing the optimization process and enabling faster convergence of physics-aware models. Furthermore, while traditional physics-informed methods solve for a single PDE instance, our approach addresses parametric PDEs. Specifically, our method integrates the physical loss gradient with the PDE parameters to solve over a distribution of PDE parameters, including coefficients, initial conditions, or boundary conditions. We demonstrate the effectiveness of our method through empirical experiments on multiple datasets, comparing training and test-time optimization performance.", "keywords": [ "Physics-informed Deep Learning", "PDE Solver", "Parametric PDE", "PINNs" ], "rating_list": [ 3, 8, 8, 6, 3 ], "soundness_list": [ 1, 3, 4, 2, 2 ], "presentation_list": [ 2, 3, 4, 2, 2 ], "contribution_list": [ 1, 3, 3, 2, 2 ], "confidence_list": [ 5, 3, 3, 4, 4 ] }, { "paper_id": "jqAqZhEMsk", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Chengsong Huang", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "", "position": "PhD student" }, { "name": "Jiaxin Huang", "gender": "Female", "institution": "Washington University, Saint Louis", "country": "", "position": "Assistant Professor" }, { "name": "Langlin Huang", "gender": "Not Specified", "institution": "Washington University, Saint Louis", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 30, "n_ref": 57, "n_ref_all": 70, "n_fig": 3, "n_tab": 17, "L_tab": 20679, "n_element_tab": 938, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1898, "n_element_tab_1": 288, "formula_len_all": 737, "formula_len_all_1": 483, "len_all": 163809, "len_all_1": 63911, "len_abs": 1303, "len_title": 130, "len_sents": 32084, "len_sents_1": 26101, "n_sents": 279, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1311, "title": "Divide, Reweight, and Conquer: A Logit Arithmetic Approach for In-Context Learning", "abs": "In-Context Learning (ICL) emerges as a key feature for Large Language Models (LLMs), allowing them to adapt to new tasks by leverageing task-specific examples without updating model parameters. However, ICL faces challenges with increasing numbers of examples due to performance degradation and quadratic computational costs. In this paper, we propose Logit Arithmetic Reweighting\nApproach (LARA), a novel framework that enhances ICL by using logit-based ensembling of multiple demonstrations. Our approach divides long input demonstrations into parallelizable shorter inputs to significantly reduce memory requirements, and then effectively aggregate the information by reweighting logits of each group via a non-gradient optimization approach. We further introduce Bi-\nnary LARA (B-LARA), a variant that constrains weights to binary values to simplify the search space and reduces memory usage by filtering out less informative demonstration groups. Experiments on BBH and MMLU demonstrate that LARA and B-LARA outperform all baseline methods in both accuracy and memory efficiency. We also conduct extensive analysis to show that LARA generalizes well to scenarios of varying numbers of examples from limited to many-shot demonstrations. Our codes can be found in https://anonymous.4open.science/r/LARA-F55B.", "keywords": [ "Efficient Inference", "In-context Learning", "Non-gradient Optimization", "Large Language Models" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "jpypMKAsO6", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Hengwei Ye", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" }, { "name": "Yuxuan Ge", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" }, { "name": "Zheng Tian", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "Assistant Professor" }, { "name": "\u7ba1\u57b8\u6c40", "gender": "Male", "institution": "\u4e0a\u6d77\u79d1\u6280\u5927\u5b66", "country": "CN", "position": "Undergrad student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 61, "n_ref": 90, "n_ref_all": 106, "n_fig": 10, "n_tab": 3, "L_tab": 1479, "n_element_tab": 176, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 1479, "n_element_tab_1": 176, "formula_len_all": 20, "formula_len_all_1": 0, "len_all": 151477, "len_all_1": 62715, "len_abs": 1138, "len_title": 126, "len_sents": 32107, "len_sents_1": 29851, "n_sents": 202, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1141, "title": "GridAgent: A 2D Grid-Based Game Framework And Benchmark For Multimodal Large Language Models", "abs": "Multimodal Large Language Models (MLLMs) integrate the linguistic capabilities of LLMs with the ability to process multimodal data, enabling them to address a wider array of tasks. However, a comprehensive and standardized benchmark for evaluating MLLMs' complex visual reasoning performance in multimodal tasks has yet to be established. We introduce GridAgent, a versatile 2D grid-based framework that serves as a benchmark for assessing five essential capabilities of MLLMs: execution, perception reasoning, memory, learning, and planning. The framework includes twelve unique game tasks specifically designed to avoid overlap with the model's pre-training corpus. Each task targets at least one core competency and is enriched with diverse semantic information. Additionally, the game layouts are randomly generated, ensuring a more rigorous and authentic assessment of the MLLMs' capabilities. Experimental results indicate that although certain MLLMs excel in specific capabilities, none exhibit a comprehensive skill set comparable to the human baseline. Our work can be seen at: https://iclr2025gridagent.github.io/GridAgent-website.", "keywords": [ "MLLM; benchmark; game" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "jpSLXoRKnH", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Hongyin Luo", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "James R. Glass", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Principal Researcher" }, { "name": "Xiangjun Fan", "gender": "Male", "institution": "Meta", "country": "US", "position": "Machine Learning Engineer" }, { "name": "Xuliang Huang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Undergrad student" }, { "name": "Yibo Jiang", "gender": "Male", "institution": "University of Chicago", "country": "", "position": "PhD student" }, { "name": "Zhenting Qi", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "MS student" }, { "name": "Zhuokai Zhao", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Hima Lakkaraju", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 148, "n_ref_all": 193, "n_fig": 11, "n_tab": 7, "L_tab": 1880, "n_element_tab": 217, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1761, "formula_len_all_1": 1610, "len_all": 211374, "len_all_1": 65406, "len_abs": 1617, "len_title": 111, "len_sents": 69851, "len_sents_1": 31533, "n_sents": 546, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1510, "title": "Quantifying Generalization Complexity for Large Language Models", "abs": "While large language models (LLMs) have shown exceptional capabilities in understanding complex queries and performing sophisticated tasks, their generalization abilities are often deeply entangled with memorization, necessitating more precise evaluation. To address this challenge, we introduce Scylla, a dynamic evaluation framework that quantitatively measures the generalization abilities of LLMs. Scylla disentangles generalization from memorization via assessing model performance on both in-distribution (ID) and out-of-distribution (OOD) data through 20 tasks across 5 levels of complexity. Through extensive experiments, we uncover a non-monotonic relationship between task complexity and the performance gap between ID and OOD data, which we term the generalization valley. Specifically, this phenomenon reveals a critical threshold---referred to as critical complexity---where reliance on non-generalizable behavior peaks, indicating the upper bound of LLMs' generalization capabilities. As model size increases, the critical complexity shifts toward higher levels of task complexity, suggesting that larger models can handle more complex reasoning tasks before over-relying on memorization. Leveraging Scylla and the concept of critical complexity, we benchmark 28 LLMs including both open-sourced models such as LLaMA and Qwen families, and close-sourced models like Claude and GPT, providing a more robust evaluation and establishing a clearer understanding of LLMs' generalization capabilities.", "keywords": [ "large language model", "generalization", "evaluation" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 4, 3, 4, 4 ], "contribution_list": [ 2, 3, 4, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "jp4pxKqCRW", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Cheng Luo", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Dongsheng Guo", "gender": "Male", "institution": "Quan Cheng Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Jikun Hu", "gender": "unknown", "institution": "Quan Cheng Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Qilei Zhang", "gender": "Male", "institution": "Quancheng laboratory", "country": "CN", "position": "Researcher" }, { "name": "Qingyao Ai", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Quan Zhou", "gender": "unknown", "institution": "Megatech Inc", "country": "CN", "position": "Researcher" }, { "name": "Xuebing Sun", "gender": "unknown", "institution": "Dareway", "country": "CN", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 13, "n_ref": 32, "n_ref_all": 51, "n_fig": 7, "n_tab": 2, "L_tab": 38, "n_element_tab": 2, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 722, "formula_len_all_1": 722, "len_all": 104620, "len_all_1": 58045, "len_abs": 2415, "len_title": 98, "len_sents": 40614, "len_sents_1": 27821, "n_sents": 245, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 2392, "title": "Long-context Extrapolation via Periodic Extension", "abs": "Long-context extrapolation aims to extend the contextual window of large language models to process more contextual information, which is widely adopted in industrial applications. \nCurrent mainstream solutions involve increasing the rotation base of RoPE to varying degrees or introducing optimization strategies such as ``low-frequency extrapolation and high-frequency interpolation'', in order to enhance the model's extrapolation capabilities for long context. Actually, these methods alter the representation distribution of positional information by adjusting the rotation frequency of positional encoding, resulting in inevitably disrupt the attention distribution within the original training length range. \nIn this paper, we analyze this phenomenon from a theoretical perspective and propose a long-context extrapolation strategy that preserves the known distribution via periodic extension of high-dimensional positional encoding. Based on this strategy, we design two methods, namely Extra-PE and Extra-MPE, to significantly enhance the models' long-context extrapolation capabilities without disrupting the positional encoding distribution within the original training length. \nThrough extensive experimental results, it is found that the long-context extrapolation method based on periodic extension can enhance the model's capability in extrapolating long-contexts. Specifically, a model fine-tuned on 32k tokens can extrapolate beyond 80k tokens, surpassing the performance of the NTK-32k model and approaching that of the YaRN-64k model. Furthermore, this method demonstrates significantly superior performance in extrapolating extremely long-contexts compared to other methods. Notably, a model fine-tuned on 8k tokens still does not exhibit perplexity explosion when extrapolating to 80k tokens. Additionally, during the fine-tuning process, our approach achieves optimal performance using only one-fourth of the fine-tuning steps (100 steps) compared to the YaRNmethod. Secondly, in our comparative experiments, we found that the period in which the model learns a sufficient number of positional encoding has a significant impact on long-context extrapolation capability. Finally, through attention analysis, we discovered that our method can still maintain a stable level of attention at ultra-long distances, with the mean attention value exceeding 0 at these distances.", "keywords": [ "Long-context Extrapolation", "Positional encoding", "Extra-PE", "Extra-MPE", "LLM" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 1, 1, 1 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 2, 2, 3, 3 ] }, { "paper_id": "jneVchiRlT", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Jongwuk Lee", "gender": "Male", "institution": "Sungkyunkwan University", "country": "", "position": "Associate Professor" }, { "name": "Joohyung Yun", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sungho Park", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Wook-Shin Han", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 19, "n_ref": 68, "n_ref_all": 94, "n_fig": 5, "n_tab": 6, "L_tab": 702, "n_element_tab": 124, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 691, "n_element_tab_1": 93, "formula_len_all": 176, "formula_len_all_1": 167, "len_all": 136446, "len_all_1": 73410, "len_abs": 1619, "len_title": 159, "len_sents": 51104, "len_sents_1": 34412, "n_sents": 396, "n_sents_1": 278, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1395, "title": "FusionMaestro: Harmonizing Early Fusion, Late Fusion, and LLM Reasoning for Multi-Granular Table-Text Retrieval", "abs": "Table-text retrieval aims to retrieve relevant tables and text to support open-domain question answering. Existing studies use either early or late fusion, but face limitations. Early fusion pre-aligns a table row with its associated passages, forming ``stars,\" which often include irrelevant contexts and miss query-dependent relationships. Late fusion retrieves individual nodes, dynamically aligning them, but it risks missing relevant contexts. Both approaches also struggle with advanced reasoning tasks, such as column-wise aggregation and multi-hop reasoning. To address these issues, we propose FusionMaestro, which combines the strengths of both approaches. First, the edge-based bipartite subgraph retrieval identifies finer-grained edges between table segments and passages, effectively avoiding the inclusion of irrelevant contexts. Then, the query-relevant node expansion identifies the most promising nodes, dynamically retrieving relevant edges to grow the bipartite subgraph, minimizing the risk of missing important contexts. Lastly, the star-based LLM refinement performs logical inference at the star subgraph rather than the bipartite subgraph, supporting advanced reasoning tasks. Experimental results show that FusionMaestro outperforms state-of-the-art models with a significant improvement up to 42.6% and 39.9% in recall and nDCG, respectively, on the OTT-QA benchmark.", "keywords": [ "Table-text retrieval", "Information retrieval", "Open-domain", "Large language model", "Early fusion", "Late fusion", "Multi-granular" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "jmsRo2ScoI", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Chia-Wen Kuo", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Fan Chen", "gender": "Male", "institution": "Bytedance", "country": "US", "position": "Researcher" }, { "name": "Longyin Wen", "gender": "Male", "institution": "Bytedance Inc.", "country": "US", "position": "Research Manager" }, { "name": "Sijie Zhu", "gender": "unknown", "institution": "ByteDance", "country": "US", "position": "Researcher" }, { "name": "Xiaohui Shen", "gender": "unknown", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 35, "n_ref": 74, "n_ref_all": 97, "n_fig": 6, "n_tab": 7, "L_tab": 2588, "n_element_tab": 511, "n_fig_1": 6, "n_tab_1": 8, "L_tab_1": 3996, "n_element_tab_1": 463, "formula_len_all": 1930, "formula_len_all_1": 1803, "len_all": 131821, "len_all_1": 66824, "len_abs": 1251, "len_title": 132, "len_sents": 31064, "len_sents_1": 26563, "n_sents": 220, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1286, "title": "Rethinking Homogeneity of Vision and Text Tokens in Large Vision-and-Language Models", "abs": "Large vision-and-language models (LVLMs) typically treat visual and textual embeddings as homogeneous inputs to a large language model (LLM). However, these inputs are inherently different: visual inputs are multi-dimensional and contextually rich, often pre-encoded by models like CLIP, while textual inputs lack this structure. In this paper, we propose Decomposed Attention (D-Attn), a novel method that processes visual and textual embeddings differently by decomposing the 1-D causal self-attention in LVLMs. After the attention decomposition, D-Attn diagonalizes visual-to-visual self-attention, reducing computation from $\\mathcal{O}(|V|^2)$ to $\\mathcal{O}(|V|)$ for $|V|$ visual embeddings without compromising performance. Moreover, D-Attn debiases positional encodings in textual-to-visual cross-attention, further enhancing visual understanding. Finally, we introduce an $\\alpha$-weighting strategy to merge visual and textual information, maximally preserving the pre-trained LLM\u2019s capabilities with minimal modifications. Extensive experiments and rigorous analyses validate the effectiveness of D-Attn, demonstrating significant improvements on multiple image benchmarks while significantly reducing computational costs. Code, data, and models will be publicly available.", "keywords": [ "LVLM", "LMM", "multimodal" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "jmmk5xjYhd", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Junzhe Wang", "gender": "Female", "institution": "George Mason University", "country": "", "position": "PhD student" }, { "name": "Lannan Luo", "gender": "Female", "institution": "George Mason University", "country": "", "position": "Assistant Professor" }, { "name": "Minghao Hu", "gender": "Male", "institution": "George Mason University", "country": "US", "position": "PhD student" }, { "name": "Weisen Zhao", "gender": "unknown", "institution": "George Mason University", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 18, "n_ref_uni": 26, "n_ref": 45, "n_ref_all": 86, "n_fig": 6, "n_tab": 13, "L_tab": 4271, "n_element_tab": 324, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1650, "n_element_tab_1": 28, "formula_len_all": 1038, "formula_len_all_1": 2216, "len_all": 201100, "len_all_1": 48112, "len_abs": 1006, "len_title": 74, "len_sents": 62946, "len_sents_1": 16293, "n_sents": 522, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1351, "title": "MalTrans: Unsupervised Binary Code Translation with Application to Malware Detection", "abs": "Applying deep learning to malware detection has drawn great attention due to its notable performance. With the increasing prevalence of cyberattacks targeting IoT devices, there is a parallel rise in the development of malware across various Instruction Set Architectures (ISAs). It is thus important to extend malware detection capacity to multiple ISAs. However, training a deep learning-based malware detection model usually requires a large number of labeled malware samples.\nThe process of collecting and labeling sufficient malware samples to build datasets for each ISA is labor-intensive and time-consuming.\nTo reduce the burden of data collection, we propose to leverage the ideas and techniques in Neural Machine Translation (NMT) for malware detection. Specifically, when dealing with malware in a certain ISA, we translate it to an ISA with sufficient malware samples (such as X86-64). This allows us to apply a model trained on one ISA to analyze malware from another ISA. Our approach reduces the data collection effort by enabling malware detection across multiple ISAs using a model trained on a single ISA. We have implemented and evaluated the model on six ISAs, including X86-64, i386, ARM64, ARM32, MIPS32, and s390x. The results demonstrate its high translation capability, thereby enabling superior malware detection across ISAs.", "keywords": [ "binary code analysis", "malware detection", "neural machine translation" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "jmN1zXMq0O", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Atish Agarwala", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Elliot Paquette", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Associate Professor" }, { "name": "Ke Liang Xiao", "gender": "Male", "institution": "McGill University, McGill University", "country": "CA", "position": "PhD student" }, { "name": "Noah Marshall", "gender": "unknown", "institution": "McGill University, McGill University", "country": "CA", "position": "PhD student" } ], "n_formula": 236, "n_formula_1": 37, "n_ref_uni": 24, "n_ref": 37, "n_ref_all": 78, "n_fig": 11, "n_tab": 1, "L_tab": 5, "n_element_tab": 1, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18227, "formula_len_all_1": 1938, "len_all": 195908, "len_all_1": 59519, "len_abs": 1143, "len_title": 133, "len_sents": 54746, "len_sents_1": 25117, "n_sents": 637, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1149, "title": "To Clip or not to Clip: the Dynamics of SGD with Gradient Clipping in High-Dimensions", "abs": "The success of modern machine learning is due in part to the adaptive optimization methods that have been developed to deal with the difficulties of training large models over complex datasets. One such method is gradient clipping: a practical procedure with limited theoretical underpinnings. In this work, we study clipping in a least squares problem under streaming SGD. We develop a theoretical analysis of the learning dynamics in the limit of large intrinsic dimension\u2014a model and dataset dependent notion of dimensionality. In this limit we find a deterministic equation that describes the evolution of the loss and demonstrate that this equation predicts the path of clipped SGD on synthetic, CIFAR10, and Wikitext2 data. We show that with Gaussian noise clipping cannot improve SGD performance. Yet, in other noisy settings, clipping can provide benefits with tuning of the clipping threshold. We propose a simple heuristic for near optimal scheduling of the clipping threshold which requires the tuning of only one hyperparameter. We conclude with a discussion about the links between high-dimensional clipping and neural network training.", "keywords": [ "gradient clipping", "high-dimensional probability", "stochastic optimization", "deep learning theory" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "jlzNb1iWs3", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Andre Cornman", "gender": "unknown", "institution": "Tatta Bio", "country": "US", "position": "Principal Researcher" }, { "name": "Antonio Pedro Camargo", "gender": "unknown", "institution": "Lawrence Berkeley National Lab", "country": "US", "position": "Researcher" }, { "name": "Jacob West-Roberts", "gender": "Male", "institution": "Tatta Bio", "country": "US", "position": "Researcher" }, { "name": "Martin Beracochea", "gender": "Male", "institution": "European Bioinformatics Institute", "country": "GB", "position": "Staff member" }, { "name": "Milot Mirdita", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Postdoc" }, { "name": "Sergey Ovchinnikov", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Simon Roux", "gender": "unknown", "institution": "Lawrence Berkeley National Lab", "country": "US", "position": "Principal Researcher" }, { "name": "Yunha Hwang", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 49, "n_ref": 71, "n_ref_all": 91, "n_fig": 10, "n_tab": 2, "L_tab": 539, "n_element_tab": 96, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 2053, "n_element_tab_1": 125, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 169186, "len_all_1": 54592, "len_abs": 1425, "len_title": 88, "len_sents": 34625, "len_sents_1": 24987, "n_sents": 259, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1442, "title": "The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling", "abs": "Biological language model performance depends heavily on pretraining data quality, diversity, and size. While metagenomic datasets feature enormous biological diversity, their utilization as pretraining data has been limited due to challenges in data accessibility, quality filtering and deduplication. Here, we present the Open MetaGenomic (OMG) corpus, a genomic pretraining dataset totalling 3.1T base pairs and 3.3B protein coding sequences, obtained by combining two largest metagenomic dataset repositories (JGI's IMG and EMBL's MGnify). We first document the composition of the dataset and describe the quality filtering steps taken to remove poor quality data. We make the OMG corpus available as a mixed-modality genomic sequence dataset that represents multi-gene encoding genomic sequences with translated amino acids for protein coding sequences, and nucleic acids for intergenic sequences. We train the first mixed-modality genomic language model (gLM2) that leverages genomic context information to learn robust functional representations, as well as coevolutionary signals in protein-protein interfaces and genomic regulatory syntax. Furthermore, we show that deduplication in embedding space can be used to balance the corpus, demonstrating improved performance on downstream tasks. The OMG dataset is publicly hosted on the Hugging Face Hub at \\url{UrlHiddenForAnonymity} and gLM2 is available at \\url{UrlHiddenForAnonymity}.", "keywords": [ "metagenomics", "pretraining dataset", "genomic language model" ], "rating_list": [ 5, 8, 6, 5 ], "soundness_list": [ 3, 4, 4, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "jlsEtFZYUg", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Geoff Pleiss", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Assistant Professor" }, { "name": "John Patrick Cunningham", "gender": "Male", "institution": "Columbia University", "country": "", "position": "" }, { "name": "Niclas Dern", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Undergrad student" } ], "n_formula": 101, "n_formula_1": 26, "n_ref_uni": 51, "n_ref": 106, "n_ref_all": 147, "n_fig": 13, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9399, "formula_len_all_1": 1995, "len_all": 197238, "len_all_1": 71915, "len_abs": 1612, "len_title": 119, "len_sents": 57314, "len_sents_1": 29187, "n_sents": 520, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1426, "title": "Theoretical Limitations of Ensembles in the Age of Overparameterization", "abs": "Classic tree-based ensembles generalize better than any single decision tree. In contrast, recent empirical studies find that modern ensembles of (overparameterized) neural networks may not provide any inherent generalization advantage over single but larger neural networks. This paper clarifies how modern overparameterized ensembles differ from their classic underparameterized counterparts, using ensembles of random feature (RF) regressors as a basis for developing theory. In contrast to the underparameterized regime, where ensembling typically induces regularization and increases generalization, we prove that infinite ensembles of overparameterized RF regressors become pointwise equivalent to (single) infinite-width RF regressors. This equivalence, which is exact for ridgeless models and approximate for small ridge penalties, implies that overparameterized ensembles and single large models exhibit nearly identical generalization. As a consequence, we can characterize the predictive variance amongst ensemble members, and demonstrate that it quantifies the expected effects of increasing capacity rather than capturing any conventional notion of uncertainty. Our results challenge common assumptions about the advantages of ensembles in overparameterized settings, prompting a reconsideration of how well intuitions from underparameterized ensembles transfer to deep ensembles and the overparameterized regime.", "keywords": [ "Ensembles", "Deep Ensembles", "Uncertainty Quantification", "Overparameterization", "Random feature regression", "Kernel regression" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "jln7IcheW6", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Heng Huang", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Hongyang Zhang", "gender": "Male", "institution": "School of Computer Science, University of Waterloo", "country": "CA", "position": "Assistant Professor" }, { "name": "Junfeng Guo", "gender": "Male", "institution": "University of Maryland Institute for Advanced Computer Studies, University of Maryland, College Park", "country": "US", "position": "Postdoc" }, { "name": "Ruibo Chen", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Yanshuo Chen", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Yihan Wu", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Zhengmian Hu", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" } ], "n_formula": 85, "n_formula_1": 19, "n_ref_uni": 27, "n_ref": 99, "n_ref_all": 119, "n_fig": 6, "n_tab": 13, "L_tab": 14601, "n_element_tab": 894, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2024, "n_element_tab_1": 157, "formula_len_all": 15193, "formula_len_all_1": 1736, "len_all": 217457, "len_all_1": 68732, "len_abs": 1343, "len_title": 164, "len_sents": 63425, "len_sents_1": 30140, "n_sents": 614, "n_sents_1": 258, "venue": "ICLR2025", "decision": "withdraw", "L_title": 116, "L_abs": 1350, "title": "Pseudo- vs. True-Randomness: Rethinking Distortion-Free Watermarks of Language Models under Watermark Key Collisions", "abs": "Language model (LM) watermarking techniques inject a statistical signal into LM-generated content by substituting the random sampling process with pseudo-random sampling, using watermark keys as the random seed. Among these statistical watermarking approaches, distortion-free watermarks are particularly crucial because they embed watermarks into LM-generated content without compromising generation quality. However, one notable limitation of pseudo-random sampling compared to true-random sampling is that, under the same watermark keys (i.e., key collision), the results of pseudo-random sampling exhibit correlations. This limitation could potentially undermine the distortion-free property. Our studies reveal that key collisions are inevitable due to the limited availability of watermark keys, and existing distortion-free watermarks exhibit a significant distribution bias toward the original LM distribution in the presence of key collisions. Moreover, we go beyond the key collision condition and prove that achieving a perfect distortion-free watermark is impossible. To study the trade-off between watermark strength and its distribution bias, we introduce a new family of distortion-free watermarks--beta-watermark. Experimental results support that the beta-watermark can effectively reduce the distribution bias under key collisions.", "keywords": [ "LLM watermarking" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "jlhBFm7T2J", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Dawn Song", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Sam Gunn", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Xuandong Zhao", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" } ], "n_formula": 14, "n_formula_1": 6, "n_ref_uni": 34, "n_ref": 124, "n_ref_all": 171, "n_fig": 14, "n_tab": 4, "L_tab": 2463, "n_element_tab": 103, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1330, "n_element_tab_1": 74, "formula_len_all": 1011, "formula_len_all_1": 409, "len_all": 162433, "len_all_1": 67375, "len_abs": 1160, "len_title": 101, "len_sents": 51502, "len_sents_1": 30395, "n_sents": 431, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1171, "title": "An undetectable watermark for generative image models", "abs": "We present the first undetectable watermarking scheme for generative image models.\n_Undetectability_ ensures that no efficient adversary can distinguish between watermarked and un-watermarked images, even after making many adaptive queries.\nIn particular, an undetectable watermark does not degrade image quality under any efficiently computable metric.\nOur scheme works by selecting the initial latents of a diffusion model using a pseudorandom error-correcting code (Christ and Gunn, 2024), a strategy which guarantees undetectability and robustness.\nWe experimentally demonstrate that our watermarks are quality-preserving and robust using Stable Diffusion 2.1.\nOur experiments verify that, in contrast to _every prior scheme_ we tested, our watermark does not degrade image quality.\nOur experiments also demonstrate robustness: existing watermark removal attacks fail to remove our watermark from images without significantly degrading the quality of the images.\nFinally, we find that we can robustly encode 512 bits in our watermark, and up to 2500 bits when the images are not subjected to watermark removal attacks. Code is provided in the supplementary materials.", "keywords": [ "Watermarking", "AI Safety", "Diffusion Models", "Generative AI" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "jlEDDCYLvV", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Cong Guan", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "PhD student" }, { "name": "Feng Chen", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Jingyu Ru", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "Associate Professor" }, { "name": "Kunmin Lin", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Undergrad student" }, { "name": "Lei Yuan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Researcher" }, { "name": "Lihe Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Ziqian Zhang", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Yang Yu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Professor" } ], "n_formula": 15, "n_formula_1": 13, "n_ref_uni": 68, "n_ref": 130, "n_ref_all": 162, "n_fig": 12, "n_tab": 3, "L_tab": 538, "n_element_tab": 39, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1103, "formula_len_all_1": 1096, "len_all": 210822, "len_all_1": 59246, "len_abs": 1542, "len_title": 130, "len_sents": 62253, "len_sents_1": 28763, "n_sents": 399, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1549, "title": "Haland: Human-AI Coordination via Policy Generation from Language-guided Diffusion", "abs": "Developing intelligent agents that can effectively coordinate with diverse human partners is a fundamental goal of artificial general intelligence. Previous approaches typically generate a variety of partners to cover human policies, and then either train a single universal agent or maintain multiple best-response (BR) policies for different partners. However, the first direction struggles with the stochastic and multimodal nature of human behaviors, and the second relies on costly few-shot adaptations during policy deployment, which is unbearable in real-world applications such as healthcare and autonomous driving. Recognizing that human partners can easily articulate their preferences or behavioral styles through natural languages and make conventions beforehand, we propose a framework for Human-AI Coordination via Policy Generation from Language-guided Diffusion, referred to as Haland. Haland first trains BR policies for various partners using reinforcement learning, and then compresses policy parameters into a single latent diffusion model, conditioned on task-relevant language derived from their behaviors. Finally, the alignment between task-relevant and natural languages is achieved to facilitate efficient human-AI coordination. Empirical evaluations across diverse cooperative environments demonstrate that Haland generates agents with significantly enhanced zero-shot coordination performance, utilizing only natural language instructions from various partners, and outperforms existing methods by approximately 89.64\\%.", "keywords": [ "Multi-agent Reinforcement Learning", "Human-AI Coordination", "Coordination and Cooperation", "Reinforcement Learning" ], "rating_list": [ 5, 5, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 1, 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 2, 4, 4 ] }, { "paper_id": "jl9lHkQrrI", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Christodoulos Constantinides", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Shuxin Lin", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Dhaval Patel", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 16, "n_ref": 27, "n_ref_all": 38, "n_fig": 5, "n_tab": 4, "L_tab": 1339, "n_element_tab": 111, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1139, "n_element_tab_1": 86, "formula_len_all": 81, "formula_len_all_1": 81, "len_all": 82529, "len_all_1": 53877, "len_abs": 1032, "len_title": 103, "len_sents": 27154, "len_sents_1": 24838, "n_sents": 210, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1035, "title": "ALIA: An LLM for Industrial Assets using Synthetic Data", "abs": "With the emergence of agentic workflow development using Large Language Models (LLMs) for industrial applications, there is a growing need for small language models to possess domain-specific knowledge. In many existing approaches, reference materials such as books are used as a source of knowledge. This paper presents a novel approach to fine-tune a base LLM model in a continued pre-training fashion for the industrial assets domain, leveraging knowledge documented in a tabular structure to generate synthetic knowledge documents and a vast amount of question-answer pairs using an entity and relationship-driven approach. Ultimately, this approach enables the fine-tuning of a small LLM (LLAMA 3.1) to evaluate the performance enhancement it brings. We tested the base model and the enhanced model on the Industry4-FMSR MCQA dataset, comprising over 2,600 samples, and obtained around 4% overall improvement. Our experimental results confirm the validity of our approach in generating synthetic data for knowledge infusion tasks.", "keywords": [ "synthetic data", "industrial assets", "LLM" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 1, 2, 2, 1 ], "presentation_list": [ 2, 3, 1, 1 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "jl0wssxHHS", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "Bowei Liu", "gender": "unknown", "institution": "Xi'an University of Electronic Science and Technology", "country": "", "position": "Undergrad student" }, { "name": "Fang Liu", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Jiahao Wang", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Licheng Jiao", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Lingling Li", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Shuyuan Yang", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Xu Liu", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Yanbiao Ma", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "PhD student" }, { "name": "wenping ma", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 39, "n_ref": 85, "n_ref_all": 108, "n_fig": 13, "n_tab": 4, "L_tab": 2478, "n_element_tab": 149, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 118, "n_element_tab_1": 3, "formula_len_all": 648, "formula_len_all_1": 647, "len_all": 155920, "len_all_1": 55027, "len_abs": 1937, "len_title": 147, "len_sents": 41284, "len_sents_1": 26090, "n_sents": 267, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1416, "title": "Vision Foundation Models Bridge the Geometric Knowledge Across Domains for Long-Tailed Recognition", "abs": "Deep learning struggles to fully unleash its potential in scenarios with limited sample sizes, primarily because models fail to capture information beyond the observed domain when the number of samples from rare classes is limited. Therefore, restoring the true distribution of rare classes becomes a significant challenge. In this study, we discovered that vision foundation models can associate inter-class similarity with the similarity of geometric shapes of class distributions in cross-domain scenarios. Specifically, we observed that when two cross-domain classes are highly similar, their embedding distributions also exhibit similar geometric shapes and sizes. These phenomena only manifest when using foundation models to represent images. Our findings provide a foundation for leveraging geometric knowledge of existing data distributions to assist rare classes. Further, we propose the Geometrically Guided Uncertainty Representation (GUR) Layer tailored for long-tailed recognition tasks, aiming to calibrate and augment the embedding distribution of tail classes, thereby learning an unbiased MLP classifier. Across multiple long-tailed benchmark datasets, GUR significantly enhances the performance of vision foundation models and achieves state-of-the-art results on certain datasets. The success of GUR serves as a typical example of integrating and colliding foundation models with prior knowledge.", "keywords": [ "long-tailed classification", "transfer of knowledg", "distributed calibration", "Vision foundation models" ], "rating_list": [ 5, 5, 3, 8 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "jkpGIxSsUD", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Baixu Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Jialong Wu", "gender": "Male", "institution": "School of Software, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Junwei Pan", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "QianLi", "gender": "Female", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Xian Hu", "gender": "Female", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Ximei Wang", "gender": "Male", "institution": "Tencent", "country": "", "position": "Researcher" }, { "name": "jie jiang", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "VP" }, { "name": "Mingsheng Long", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "\u5b81\u4e9a \u51af", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 21, "n_ref": 53, "n_ref_all": 91, "n_fig": 25, "n_tab": 6, "L_tab": 4969, "n_element_tab": 587, "n_fig_1": 19, "n_tab_1": 5, "L_tab_1": 2565, "n_element_tab_1": 251, "formula_len_all": 227, "formula_len_all_1": 147, "len_all": 141908, "len_all_1": 64999, "len_abs": 1657, "len_title": 109, "len_sents": 47447, "len_sents_1": 27006, "n_sents": 388, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1672, "title": "Long-Sequence Recommendation Models Need Decoupled Embeddings", "abs": "Lifelong user behavior sequences, comprising up to tens of thousands of history behaviors, are crucial for capturing user interests and predicting user responses in modern recommendation systems. \nA two-stage paradigm is typically adopted to handle these long sequences: a few relevant behaviors are first searched from the original long sequences via an attention mechanism in the first stage and then aggregated with the target item to construct a discriminative representation for prediction in the second stage. \nIn this work, we identify and characterize, for the first time, a neglected deficiency in existing long-sequence recommendation models: a single set of embeddings struggles with learning both attention and representation, leading to interference between these two processes. \nInitial attempts to address this issue using linear projections---a technique borrowed from language processing---proved ineffective, shedding light on the unique challenges of recommendation models. \nTo overcome this, we propose the Decoupled Attention and Representation Embeddings (DARE) model, where two distinct embedding tables are initialized and learned separately to fully decouple attention and representation. \nExtensive experiments and analysis demonstrate that DARE provides more accurate search of correlated behaviors and outperforms baselines with AUC gains up to 9\u2030 on public datasets and notable online system improvements. \nFurthermore, decoupling embedding spaces allows us to reduce the attention embedding dimension and accelerate the search procedure by 50\\% without significant performance impact, enabling more efficient, high-performance online serving.", "keywords": [ "Recommender System", "User Interest Modeling" ], "rating_list": [ 8, 3, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "jki6EFsZLw", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Boqing Gong", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "Assistant Professor" }, { "name": "Heng Huang", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Hexiang Hu", "gender": "Male", "institution": "xAI", "country": "US", "position": "Principal Researcher" }, { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Mingda Zhang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Software Engineer" }, { "name": "Tianyi Zhou", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "YANDONG LI", "gender": "Male", "institution": "Google", "country": "US", "position": "Software Engineer" }, { "name": "Yiwen Chen", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Zifeng Wang", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "LICHANG CHEN", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Pranav Shyam Manjunath", "gender": "Male", "institution": "R V College of Engineering", "country": "", "position": "Undergrad student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 18, "n_ref": 39, "n_ref_all": 62, "n_fig": 3, "n_tab": 9, "L_tab": 2070, "n_element_tab": 358, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 927, "n_element_tab_1": 151, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 125951, "len_all_1": 61472, "len_abs": 1475, "len_title": 127, "len_sents": 38557, "len_sents_1": 28484, "n_sents": 312, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1492, "title": "OmnixR: Evaluating Omni-modality Language Models on Reasoning across Modalities", "abs": "We introduce \\textbf{OmnixR}, an evaluation suite designed to benchmark state-of-the-art Omni-modality Language Models (OLMs), such as GPT-4o and Gemini. \nEvaluating OLMs, which integrate multiple modalities such as text, vision, and audio, presents unique challenges. \nParticularly, the user message might often consist of multiple modalities, such that OLMs have to establish holistic understanding and reasoning across modalities to accomplish the task.\nExisting benchmarks are limited to single-modality or dual-modality tasks (e.g., image+text or video+text), overlooking comprehensive multi-modal assessments of model reasoning.\nTo address this, OmnixR offers two evaluation variants: (1) OmnixR-synth: a synthetic dataset generated automatically by translating text into multiple modalities\u2014audio, images, video, and hybrids Omnify!. (2) OmnixR-real: a real-world dataset, manually curated and annotated by experts, for evaluating cross-modal reasoning in natural settings. \nOmnixR presents a unique evaluation towards assessing OLMs over a diverse mix of modalities, such as a question that involves video, audio, and text, providing a rigorous cross-modal reasoning testbed than any existing benchmarks.\nOur experiments find that all state-of-the-art OLMs struggles with OmnixR questions that require integrating information from multiple modalities to answer. \nFurther analysis highlight differences in reasoning behavior and underscoring the challenges of omni-modal AI alignment.", "keywords": [ "Omni-modality Language Model; Omni-Eval; Omni-Reasoning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "jkVQ31GeIA", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Shaolei Zhang", "gender": "Male", "institution": "Key Laboratory of Intelligent Information Processing Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Tian Yu", "gender": "Male", "institution": ", Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Yang Feng", "gender": "unknown", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 38, "n_ref": 109, "n_ref_all": 141, "n_fig": 11, "n_tab": 12, "L_tab": 6776, "n_element_tab": 475, "n_fig_1": 10, "n_tab_1": 5, "L_tab_1": 1160, "n_element_tab_1": 150, "formula_len_all": 325, "formula_len_all_1": 376, "len_all": 258347, "len_all_1": 62850, "len_abs": 1713, "len_title": 125, "len_sents": 54838, "len_sents_1": 28553, "n_sents": 418, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1721, "title": "Auto-RAG: Autonomous Retrieval-Augmented Generation for Large Language Models", "abs": "Iterative retrieval refers to the process in which the model continuously queries the retriever during generation to enhance the relevance of the retrieved knowledge, thereby improving the performance of Retrieval-Augmented Generation (RAG). Existing work typically employs few-shot prompting or manually constructed rules to implement iterative retrieval. This introduces additional inference overhead and overlooks the remarkable reasoning capabilities of Large Language Models (LLMs). In this paper, we introduce Auto-RAG, an autonomous iterative retrieval model centered on the LLM's powerful decision-making capabilities. Auto-RAG engages in multi-turn dialogues with the retriever, systematically planning retrievals and refining queries to acquire valuable knowledge. This process continues until sufficient external information is gathered, at which point the results are presented to the user. To this end, we develop a method for autonomously synthesizing reasoning-based decision-making instructions in iterative retrieval and fine-tuned the latest open-source LLMs.\nThe experimental results indicate that Auto-RAG is capable of autonomous iterative interaction with the retriever, effectively leveraging the remarkable reasoning and decision-making abilities of LLMs, which lead to outstanding performance across six benchmarks. Further analysis reveals that Auto-RAG can autonomously adjust the number of iterations based on the difficulty of the questions and the utility of the retrieved knowledge, without requiring any human intervention. Moreover, Auto-RAG expresses the iterative retrieval process in natural language, enhancing interpretability while providing users with a more intuitive experience.", "keywords": [ "Retrieval-Augmented Generation", "Large Language Models", "Autonomous Decision-Making" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "jkUp3lybXf", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Fangkai Jiao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Furu Wei", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Partner Research Manager" }, { "name": "Geyang Guo", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Nancy F. Chen", "gender": "unknown", "institution": "I2R, A*STAR", "country": "", "position": "Principal Researcher" }, { "name": "Shafiq Joty", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Principal Researcher" }, { "name": "Xingxing Zhang", "gender": "Male", "institution": "Microsoft Research Asia", "country": "", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 39, "n_ref": 130, "n_ref_all": 160, "n_fig": 7, "n_tab": 21, "L_tab": 3006, "n_element_tab": 323, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 421, "n_element_tab_1": 66, "formula_len_all": 674, "formula_len_all_1": 927, "len_all": 234892, "len_all_1": 57912, "len_abs": 1340, "len_title": 106, "len_sents": 53738, "len_sents_1": 26337, "n_sents": 444, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 58, "L_abs": 1354, "title": "Preference Optimization for Reasoning with Pseudo Feedback", "abs": "Preference optimization techniques, such as Direct Preference Optimization (DPO), are frequently employed to enhance the reasoning capabilities of large language models (LLMs) in domains like mathematical reasoning and coding, typically following supervised fine-tuning. These methods rely on high-quality labels for reasoning tasks to generate preference pairs; however, the availability of reasoning datasets with human-verified labels is limited.\nIn this study, we introduce a novel approach to generate pseudo feedback for reasoning tasks by framing the labeling of solutions to reason problems as an evaluation against associated \\emph{test cases}. \nWe explore two forms of pseudo feedback based on test cases: one generated by frontier LLMs and the other by extending self-consistency to multi-test-case.\nWe conduct experiments on both mathematical reasoning and coding tasks using pseudo feedback for preference optimization, and observe improvements across both tasks. Specifically, using Mathstral-7B as our base model, we improve MATH results from 58.3 to 68.6, surpassing both NuminaMath-72B and GPT-4-Turbo-1106-preview. In GSM8K and College Math, our scores increase from 85.6 to 90.3 and from 34.3 to 42.3, respectively. Building on Deepseek-coder-7B-v1.5, we achieve a score of 24.3 on LiveCodeBench (from 21.1), surpassing Claude-3-Haiku.", "keywords": [ "Large Language Model", "Code Generation", "Natural Language Reasoning", "Reinforcement Learning" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 4, 4 ], "presentation_list": [ 3, 4, 4 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "jkCvAAcSDa", "primary_area": "optimization", "n_author": 8, "author_info_list": [ { "name": "Cho-Jui Hsieh", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Di Fu", "gender": "Male", "institution": "national university of singapore, National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Minhao Cheng", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Shenggan Cheng", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Yang Luo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" }, { "name": "Yong Liu", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Zirui Zhu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 33, "n_ref": 54, "n_ref_all": 69, "n_fig": 6, "n_tab": 6, "L_tab": 2144, "n_element_tab": 351, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 3138, "n_element_tab_1": 103, "formula_len_all": 225, "formula_len_all_1": 223, "len_all": 118972, "len_all_1": 56811, "len_abs": 1348, "len_title": 104, "len_sents": 30758, "len_sents_1": 24940, "n_sents": 214, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1353, "title": "SeedLoRA: A Fusion Approach to Efficient LLM Fine-Tuning", "abs": "Despite Low-Rank Adaptation (LoRA)'s popularity for fine-tuning large models, it often exhibits a noticeable performance gap compared to full fine-tuning, particularly in complex tasks such as mathematical reasoning and code generation. Motivated by this discrepancy, we propose a novel fusion approach for LoRA fine-tuned models. Our key insight is that LoRA models trained with different random seeds on the same task often exhibit complementary strengths. In contrast to existing research that typically focuses on fusing models trained on diverse tasks, we explore the potential of combining multiple LoRA models fine-tuned on the same task with different random seeds. This intra-task fusion method aims to leverage the strengths of various fine-tuned models to create a more robust and effective adaptation. To validate our approach, we conducted comprehensive experiments across three key areas: mathematical reasoning, code generation, and general instruction-tuning tasks. The results demonstrate that our fusion method significantly enhances LoRA's performance, outperforming both standalone LoRA models and current fusion methods. Notably, this advancement substantially narrows the gap between LoRA and full fine-tuning, thus offering a more effective approach to model adaptation without the GPU memory burden of full parameter fine-tuning.", "keywords": [ "Machine Learning", "Optimization", "Large Language Model" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 4, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "jjjxp9Wgjp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Atik Garg", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "PhD student" }, { "name": "Yu-Shuen Wang", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 52, "n_ref": 102, "n_ref_all": 134, "n_fig": 3, "n_tab": 6, "L_tab": 1738, "n_element_tab": 163, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1902, "n_element_tab_1": 178, "formula_len_all": 547, "formula_len_all_1": 357, "len_all": 148655, "len_all_1": 66297, "len_abs": 1649, "len_title": 112, "len_sents": 42172, "len_sents_1": 30604, "n_sents": 296, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1658, "title": "Pseudo-Labels are All You Need for Out-Of-Distribution Detection", "abs": "Detecting out-of-distribution (OOD) samples is a significant challenge in real-world deep-learning applications, such as medical imaging and autonomous driving. Traditional machine learning models, primarily trained on in-distribution (ID) data, often struggle when encountering OOD instances, resulting in unreliable predictions. While supervised OOD detection methods generally outperform unsupervised approaches due to the availability of labeled data, our research uncovers a crucial insight: their success is not necessarily due to recognizing the actual object categories in the images; instead, these methods rely on a specific classification strategy that may not correspond to real-world understanding. Essentially, supervised methods detect OOD samples by identifying the difficulties in classifying unfamiliar data. This challenge is similar to what unsupervised OOD detection methods face, as they also depend on the failure to reconstruct OOD data due to the lack of prior exposure. In this study, we bridge the gap between supervised and unsupervised OOD detection by introducing a novel approach that trains models to classify data into pseudo-categories. We employ self-supervised learning (SSL) to convert raw data into representations, which are then clustered to generate pseudo-labels. These pseudo-labels are subsequently used to train a classifier, enabling its OOD detection capabilities. Experimental results show that our approach surpasses state-of-the-art techniques. Furthermore, by training models on different sets of pseudo-labels derived from the dataset, we enhance the robustness and reliability of our OOD detection method.", "keywords": [ "Pseudo-Labels", "Unsupervised Out-of-Distribution Detection" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "jjfve2gIXe", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Melody Lo", "gender": "Female", "institution": "National Taiwan University", "country": "TW", "position": "Associate Professor" }, { "name": "Tung-Yu Wu", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 35, "n_ref": 109, "n_ref_all": 168, "n_fig": 39, "n_tab": 2, "L_tab": 380, "n_element_tab": 14, "n_fig_1": 13, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 322, "formula_len_all_1": 265, "len_all": 190674, "len_all_1": 46915, "len_abs": 879, "len_title": 130, "len_sents": 44277, "len_sents_1": 20507, "n_sents": 342, "n_sents_1": 148, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1000, "title": "U-shaped and Inverted-U Scaling behind Emergent Abilities of Large Language Models", "abs": "Large language models (LLMs) have been shown to exhibit *emergent abilities* in some downstream tasks, where model performance stagnates at first and then improves sharply and unpredictably with scale beyond a threshold. In this work, we investigate the phenomenon by grouping questions based on difficulty level and provide a possible explanation for emergent abilities. Specifically, we observe U-shaped scaling for hard questions and inverted-U scaling followed by steady improvement for easy questions. The two scaling patterns initially offset each other, causing stagnant overall performance. The performance starts to soar when the scaling pattern of easy questions reverts from inverse to standard scaling, leading to emergent abilities. Based on this finding, we propose a simple yet effective pipeline, called *Slice-and-Sandwich*, to predict the emergence threshold and model performance beyond the threshold. Our code is publicly available at https://github.com/tony10101105/ExpEmergence.", "keywords": [ "large language models", "emergent abilities", "scaling laws" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 1, 1, 4 ], "contribution_list": [ 3, 2, 2, 4 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "jjCB27TMK3", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Jiasheng Ye", "gender": "Not Specified", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Jun Zhan", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Peiju Liu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Tianxiang Sun", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Xipeng Qiu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yunhua Zhou", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 63, "n_ref": 141, "n_ref_all": 174, "n_fig": 21, "n_tab": 4, "L_tab": 1923, "n_element_tab": 178, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 829, "formula_len_all_1": 560, "len_all": 209453, "len_all_1": 67072, "len_abs": 1327, "len_title": 86, "len_sents": 61005, "len_sents_1": 31373, "n_sents": 452, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1284, "title": "Data Mixing Laws: Optimizing Data Mixtures by Predicting Language Modeling Performance", "abs": "Pretraining data of large language models composes multiple domains (e.g., web texts, academic papers, codes), whose mixture proportions crucially impact the competence of outcome models. While existing endeavors rely on heuristics or qualitative strategies to tune the proportions, we discover the quantitative predictability of model performance regarding the mixture proportions in function forms, which we refer to as the data mixing laws. Fitting such functions on sample mixtures unveils model performance on unseen mixtures before actual runs, thus guiding the selection of an ideal data mixture. Furthermore, we propose nested use of the scaling laws of training steps, model sizes, and our data mixing laws to predict the performance of large models trained on massive data under various mixtures with only small-scale training. Experimental results verify that our method effectively optimizes the training mixture of a 1B model trained for 100B tokens in RedPajama, reaching a performance comparable to the one trained for 48% more steps on the default mixture. Extending the application of data mixing laws to continual training accurately predicts the critical mixture proportion that avoids catastrophic forgetting and outlooks the potential for dynamic data schedules.", "keywords": [ "data mixtures", "pretraining", "post-training", "scaling laws" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "jj7b3p5kLY", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "David Grangier", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Matteo Pagliardini", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Pierre Ablin", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" } ], "n_formula": 25, "n_formula_1": 16, "n_ref_uni": 64, "n_ref": 100, "n_ref_all": 209, "n_fig": 43, "n_tab": 7, "L_tab": 1699, "n_element_tab": 267, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1781, "formula_len_all_1": 1505, "len_all": 323578, "len_all_1": 78579, "len_abs": 1318, "len_title": 93, "len_sents": 91323, "len_sents_1": 34110, "n_sents": 975, "n_sents_1": 366, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 45, "L_abs": 1339, "title": "The AdEMAMix Optimizer: Better, Faster, Older", "abs": "Momentum based optimizers are central to a wide range of machine learning applications. These typically rely on an Exponential Moving Average (EMA) of gradients, which decays exponentially the present contribution of older gradients. This accounts for gradients being local linear approximations which lose their relevance as the iterate moves along the loss landscape. This work questions the use of a single EMA to accumulate past gradients and empirically demonstrates how this choice can be sub-optimal: a single EMA cannot simultaneously give a high weight to the immediate past, and a non-negligible weight to older gradients. Building on this observation, we propose AdEMAMix, a simple modification of the Adam optimizer with a mixture of two EMAs to better take advantage of past gradients. Our experiments on language modeling and image classification show---quite surprisingly---that gradients can stay relevant for tens of thousands of steps. They help to converge faster, and often to lower minima: e.g., a $1.3$B parameter AdEMAMix LLM trained on $101$B tokens performs comparably to an AdamW model trained on $197$B tokens ($+95\\%$). Moreover, our method significantly slows-down model forgetting during training. Our work motivates further exploration of different types of functions to leverage past gradients, beyond EMAs.", "keywords": [ "Optimization", "LLM", "Deep Learning", "Momentum" ], "rating_list": [ 6, 6, 5, 10, 6 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 2, 4, 3, 4, 4 ], "contribution_list": [ 3, 3, 3, 4, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "jiKefvNp01", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Bingling Li", "gender": "Male", "institution": "Moore Threads", "country": "", "position": "Researcher" }, { "name": "Kaimin Liao", "gender": "unknown", "institution": "Department of Software Engineering, Shenzhen University", "country": "", "position": "MS student" }, { "name": "Luchao Wang", "gender": "Male", "institution": "mthreads", "country": "CN", "position": "Researcher" }, { "name": "Shengyi Chen", "gender": "Male", "institution": "mthreads", "country": "CN", "position": "Researcher" }, { "name": "Sijie Yan", "gender": "Male", "institution": "MT", "country": "", "position": "Researcher" }, { "name": "Yuanjun Xiong", "gender": "Male", "institution": "Predera.AI", "country": "US", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 43, "n_ref": 69, "n_ref_all": 86, "n_fig": 7, "n_tab": 6, "L_tab": 983, "n_element_tab": 170, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 391, "n_element_tab_1": 82, "formula_len_all": 692, "formula_len_all_1": 470, "len_all": 136367, "len_all_1": 63166, "len_abs": 838, "len_title": 133, "len_sents": 35382, "len_sents_1": 30081, "n_sents": 266, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 840, "title": "RetinaGS: Scalable Training for Dense Scene Rendering with Billion-Scale 3D Gaussians", "abs": "In this work, we explore the possibility of training high-parameter 3D Gaussian splatting (3DGS) models on large-scale, high-resolution datasets. We design a general model parallel training method for 3DGS, named RetinaGS, which uses a proper rendering equation and can be applied to any scene and arbitrary distribution of Gaussian primitives. It enables us to explore the scaling behavior of 3DGS in terms of primitive numbers and training resolutions that were difficult to explore before and surpass previous state-of-the-art reconstruction quality. We observe a clear positive trend of increasing visual quality when increasing primitive numbers with our method. We also demonstrate the first attempt at training a 3DGS model with more than one billion primitives on the full MatrixCity dataset that attains a promising visual quality.", "keywords": [ "Gaussian Splatting" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "jiGyD6Q2No", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Huy Minh Hoang", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Researcher" }, { "name": "Pradeep Varakantham", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Full Professor" }, { "name": "Tien Anh Mai", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 53, "n_formula_1": 21, "n_ref_uni": 34, "n_ref": 87, "n_ref_all": 110, "n_fig": 13, "n_tab": 6, "L_tab": 2074, "n_element_tab": 161, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 785, "n_element_tab_1": 9, "formula_len_all": 5371, "formula_len_all_1": 1806, "len_all": 162993, "len_all_1": 65730, "len_abs": 1132, "len_title": 120, "len_sents": 52286, "len_sents_1": 30220, "n_sents": 417, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1136, "title": "UNIQ: Offline Inverse Q-learning for Avoiding Undesirable Demonstrations", "abs": "We address the problem of offline learning a policy that avoids undesirable demonstrations. Unlike conventional offline imitation learning approaches that aim to imitate expert or near-optimal demonstrations, our setting involves avoiding undesirable behavior (specified using undesirable demonstrations). To tackle this problem, unlike standard imitation learning where the aim is to minimize the distance between learning policy and expert demonstrations, we formulate the learning task as maximizing a statistical distance, in the space of state-action stationary distributions, between the learning policy and the undesirable policy. This significantly different approach results in a novel training objective that necessitates a new algorithm to address it. Our algorithm, UNIQ, tackles these challenges by building on the inverse Q-learning framework, framing the learning problem as a cooperative (non-adversarial) task. We then demonstrate how to efficiently leverage unlabeled data for practical training. Our method is evaluated on standard benchmark environments, where it consistently outperforms state-of-the-art baselines.", "keywords": [ "offline imitation learning", "safe imitation learning", "undesirable demonstration", "preference-based learning" ], "rating_list": [ 3, 3, 3, 8 ], "soundness_list": [ 2, 2, 2, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 2, 4 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ji6MYm4Htg", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Peng Liu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Full Professor" }, { "name": "Songtao Liu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 68, "n_ref": 120, "n_ref_all": 142, "n_fig": 6, "n_tab": 9, "L_tab": 1660, "n_element_tab": 21, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 349, "n_element_tab_1": 6, "formula_len_all": 1807, "formula_len_all_1": 1654, "len_all": 222210, "len_all_1": 64519, "len_abs": 1502, "len_title": 104, "len_sents": 39842, "len_sents_1": 28542, "n_sents": 339, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1513, "title": "Pruning Aggregation Parameters for Large Language Models", "abs": "Pruning is a highly effective approach for compressing large language models (LLMs). By strategically reducing model size, pruning significantly decreases both latency and GPU memory usage during inference, resulting in more efficient and cost-effective deployment of these models. Despite their effectiveness, current structured pruning algorithms have limitations. They still require extensive continued pre-training on large datasets to achieve model compression. Moreover, most of these methods are unable to reduce the memory usage of the key-value cache during generation tasks. In this work, we propose a novel pruning algorithm that requires no additional training and targets specific parameters within LLMs. We classify the model's parameters into three categories: aggregation, transformation, and normalization. Our method primarily focuses on pruning the aggregation parameters in the higher layers of the model. To further improve the performance of the pruned LLM, we also introduce a rescaling parameter that adjusts the output of the pruned block. We conduct comprehensive experiments on a wide range of LLMs, including LLaMA3.1-8B/70B, Qwen2-7B/72B, Gemma2-9B, and Mistral-7B-v0.3. Our evaluation includes both generation and discriminative tasks across various benchmarks. The results consistently demonstrate that our method outperforms recent block pruning methods. This improvement is particularly notable in generation tasks, where our approach significantly outperforms existing baselines.", "keywords": [ "LLM", "Pruning" ], "rating_list": [ 6, 6, 3, 3, 6 ], "soundness_list": [ 3, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5, 4 ] }, { "paper_id": "jh7JQkgWGe", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Ali Gorji", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "MS student" }, { "name": "Andisheh Amrollahi", "gender": "unknown", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Andreas Krause", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" } ], "n_formula": 47, "n_formula_1": 20, "n_ref_uni": 46, "n_ref": 121, "n_ref_all": 162, "n_fig": 4, "n_tab": 5, "L_tab": 713, "n_element_tab": 103, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 478, "n_element_tab_1": 11, "formula_len_all": 4805, "formula_len_all_1": 1743, "len_all": 205641, "len_all_1": 71477, "len_abs": 2359, "len_title": 111, "len_sents": 64834, "len_sents_1": 33197, "n_sents": 514, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 2094, "title": "Amortized SHAP values via sparse Fourier function approximation", "abs": "SHAP values -- a.k.a.~SHapley Additive exPlanations -- are a popular local feature-attribution method widely used in interpretable and explainable AI. We tackle the problem of efficiently computing these values. We cover both the model-agnostic (black-box) setting, where one only has query access to the model and also the case of (ensembles of) trees where one has access to the structure of the tree. For both the black-box and the tree setting we propose a two-stage approach for estimating SHAP values. \n\nOur algorithm's first step harnesses recent results showing that many real-world predictors have a spectral bias that allows us to either exactly represent (in the case of ensembles of decision trees), or efficiently approximate them (in the case of neural networks) using a compact Fourier representation. \nFor the case of trees, given access to the tree structure, one can extract the Fourier representation using a simple recursive algorithm.\nFor the black-box setting, given query access to the black-box function, we utilize a sparse Fourier approximation algorithm to efficiently extract its compact Fourier approximation. \n\nIn the second step of the algorithm, we use the Fourier representation to exactly compute SHAP values. The second step is computationally very cheap because firstly, the representation is compact and secondly, we prove that there exists a closed-form expression for SHAP values for the Fourier basis functions. Furthermore, the expression we derive effectively ``linearizes'' the computation into a simple summation and is amenable to parallelization on multiple cores or a GPU. Since the function approximation (first step) is only done once, it allows us to produce Shapley values in an amortized way. We show speedups compared to relevant baseline methods equal levels of accuracy for both the tree and black-box settings. Moreover, this approach introduces a reliable and fine-grained continuous trade-off between computation and accuracy through the sparsity of the Fourier approximation, a feature previously unavailable in all black-box methods.", "keywords": [ "interpretability", "explainability", "shap values" ], "rating_list": [ 6, 8, 5, 3, 6 ], "soundness_list": [ 3, 4, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 3, 2, 4, 4, 2 ] }, { "paper_id": "jgVqCCg5XX", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Chen Ning", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Chenwei Yan", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Ji Wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xiangling Fu", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "Associate Professor" }, { "name": "Xiao Zhang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xien Liu", "gender": "Male", "institution": " Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Yuxuan Zhou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 24, "n_ref": 38, "n_ref_all": 61, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 388, "n_element_tab_1": 5, "formula_len_all": 235, "formula_len_all_1": 166, "len_all": 109235, "len_all_1": 60582, "len_abs": 1830, "len_title": 120, "len_sents": 35745, "len_sents_1": 30193, "n_sents": 227, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1850, "title": "Revisiting the Scaling Effects of LLMs on Medical Reasoning Capabilities", "abs": "Recently, LLMs such as the Llama and Qwen families have rapidly improved by significantly scaling their training corpora, with smaller models trained on larger datasets now approaching or surpassing the performance of previous-generation larger models on public benchmarks. In this paper, we revisit the scaling effects of LLMs, using the medical field as a case study, by carefully analyzing how training corpus size and parameter size affect model performance on problems of varying difficulty. To this end, we present MedResEval, a new benchmark built upon the MedQA dataset. It is designed to demand more complex reasoning and decision-making and more accurately reflect real-world medical scenarios. Leveraging MedResEval, we investigate the scaling effects of training corpus and model size in LLMs through a comprehensive analysis of several prominent LLM families on medical reasoning tasks of varying complexity.\nThe results reveal that while smaller models like Llama 3 (8B) approach the performance of older, larger models like Llama 2 (70B) on simple tasks like MedQA, they consistently underperform on complex tasks requiring advanced reasoning. Furthermore, we develop a difficulty-dependent scaling-law formula to characterize how LLMs' performance varies with training data size at a fixed model parameter size. The quantitative study reveals that reasoning error reduction rates are 1.3 times greater for large LLMs ($\\approx$ 70B) compared to small LLMs ($\\leq$10B) on simple tasks, and 2 times greater on complex reasoning tasks. Our study highlights that while both data and parameter scales enhance LLM performance, greater emphasis must be placed on parameter scales, particularly for complex reasoning tasks. Only LLMs with sufficiently large parameters can effectively tackle the complexities of real-world medical scenarios.", "keywords": [ "LLM evaluation", "scaling effect", "medical reasoning evaluation" ], "rating_list": [ 3, 3, 5, 6, 6 ], "soundness_list": [ 2, 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 4, 3 ], "contribution_list": [ 1, 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 2, 3, 3 ] }, { "paper_id": "jgISC1wdYy", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Jun Xu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Sunhao Dai", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Xiao Zhang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yuqi Zhou", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Zhanshuo Cao", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Undergrad student" } ], "n_formula": 38, "n_formula_1": 20, "n_ref_uni": 53, "n_ref": 103, "n_ref_all": 150, "n_fig": 17, "n_tab": 11, "L_tab": 7380, "n_element_tab": 1159, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 2369, "n_element_tab_1": 252, "formula_len_all": 2800, "formula_len_all_1": 902, "len_all": 234473, "len_all_1": 68893, "len_abs": 1580, "len_title": 109, "len_sents": 66589, "len_sents_1": 30752, "n_sents": 525, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1655, "title": "Length-Induced Embedding Collapse in Transformer-based Models", "abs": "Text embeddings enable various applications, but their performance deteriorates on longer texts. In this paper, we find that the performance degradation is due to a phenomenon called \\textbf{Length Collapse}, where longer text embeddings collapse into a narrow space. This collapse results in a distributional inconsistency between embeddings of different text lengths, ultimately hurting the performance of downstream tasks. Theoretically, by considering the self-attention mechanism inherently functions as a low-pass filter, we prove that long sequences increase the attenuation rate of the low-pass filter effect of the self-attention mechanism. With layers going deeper, excessive low-pass filtering causes the token signals to retain only their Direct-Current (DC) component, which means the input token feature maps will collapse into a narrow space, especially in long texts. Based on the above analysis, we propose to mitigate the undesirable length collapse limitation by introducing a temperature in $\\softmax(\\cdot)$, which achieves a higher low-filter attenuation rate. The tuning-free method, called \\textbf{TempScale}, can be plugged into multiple transformer-based embedding models. Empirically, we demonstrate that TempScale can improve existing embedding models especially on long text inputs, bringing up to \\textbf{0.53\\%} performance gains on 40 datasets from Massive Text Embedding Benchmark (MTEB) and \\textbf{0.82\\%} performance gains on 4 datasets from LongEmbed, which specifically focuses on long context retrieval. The source code is available at \\textcolor{blue}{\\url{https://anonymous.4open.science/r/Length_Collapse-22D2}}.", "keywords": [ "Embedding Models", "Length Collapse", "Mechanistic Interpretability" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "jfwe9qNqRi", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Mingxiao Li", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Shangsong Liang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Vasant G Honavar", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Full Professor" }, { "name": "Yige Yuan", "gender": "Not Specified", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zhaochun Ren", "gender": "Male", "institution": "Leiden University", "country": "NL", "position": "Associate Professor" }, { "name": "Zhengyu Chen", "gender": "unknown", "institution": "Westlake University", "country": "", "position": "PhD student" }, { "name": "Teng Xiao", "gender": "Male", "institution": "The Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 46, "n_formula_1": 23, "n_ref_uni": 51, "n_ref": 161, "n_ref_all": 199, "n_fig": 5, "n_tab": 9, "L_tab": 4763, "n_element_tab": 575, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 4217, "n_element_tab_1": 411, "formula_len_all": 15730, "formula_len_all_1": 8257, "len_all": 233844, "len_all_1": 88364, "len_abs": 1670, "len_title": 88, "len_sents": 57332, "len_sents_1": 33900, "n_sents": 419, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1424, "title": "SimPER: A Minimalist Approach to Preference Alignment without Hyperparameters", "abs": "Existing preference optimization objectives for language model alignment require additional hyperparameters that must be extensively tuned to achieve optimal performance, increasing both the complexity and time required for fine-tuning large language models. In this paper, we propose a simple yet effective hyperparameter-free preference optimization algorithm for alignment. We observe that promising performance can be achieved simply by optimizing inverse perplexity, which is calculated as the inverse of the exponentiated average log-likelihood of the chosen and rejected responses in the preference dataset. The resulting simple learning objective, SimPER, is easy to implement and eliminates the need for expensive hyperparameter tuning and a reference model, making it both computationally and memory efficient. Extensive experiments on widely used real-world benchmarks, including MT-Bench, AlpacaEval 2, and 10 key benchmarks of the Open LLM Leaderboard with 5 base models, demonstrate that SimPER consistently and significantly outperforms existing approaches\u2014even without any hyperparameters or a reference model. For example, despite its simplicity, SimPER outperforms state-of-the-art methods by up to 5.7 points on AlpacaEval 2 and achieves the highest average ranking across 10 benchmarks on the Open LLM Leaderboard. The source code for SimPER is publicly available at: https://github.com/tengxiao1/SimPER.", "keywords": [ "Large Language Model", "Alignment", "RLHF" ], "rating_list": [ 6, 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 4, 4, 4 ], "confidence_list": [ 4, 4, 3, 3, 5 ] }, { "paper_id": "jf7C7EGw21", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Gaofeng Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jiming Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Qi Ye", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Qingtao Liu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Yu Cui", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Zhengnan Sun", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 76, "n_ref_all": 97, "n_fig": 9, "n_tab": 10, "L_tab": 3888, "n_element_tab": 382, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 2771, "n_element_tab_1": 140, "formula_len_all": 190, "formula_len_all_1": 0, "len_all": 209472, "len_all_1": 60745, "len_abs": 1604, "len_title": 122, "len_sents": 64502, "len_sents_1": 27838, "n_sents": 546, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 121, "L_abs": 1527, "title": "VTDexManip: A Dataset and Benchmark for Visual-tactile Pretraining and Dexterous Manipulation with Reinforcement Learning", "abs": "Vision and touch are the most commonly used senses in human manipulation. While leveraging human manipulation videos for robotic task pretraining has shown promise in prior works, it is limited to image and language modalities and deployment to simple parallel grippers. In this paper, aiming to address the limitations, we collect a vision-tactile dataset by humans manipulating 10 daily tasks and 182 objects. In contrast with the existing datasets, our dataset is the first visual-tactile dataset for complex robotic manipulation skill learning. Also, we introduce a novel benchmark, featuring six complex dexterous manipulation tasks and a reinforcement learning-based vision-tactile skill learning framework. 17 non-pretraining and pretraining methods within the framework are designed and compared to investigate the effectiveness of different modalities and pertaining strategies. Key findings based on our benchmark results and analyses experiments include: 1) Despite the tactile modality used in our experiments being binary and sparse, including it directly in the policy training boosts the success rate by about 20\\% and joint pretraining it with vision gains a further 20\\%. 2) Joint pretraining visual-tactile modalities exhibits strong adaptability in unknown tasks and achieves robust performance among all tasks. 3) Using binary tactile signals with vision is robust to viewpoint setting, tactile noise, and the binarization threshold, which facilitates to the visual-tactile policy to be deployed in reality.", "keywords": [ "Robotics; Manipulation; Vision and tactile; Multi-modal pretraining; Reinforcement learning" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "jeo4FiBjlh", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Fangfang Li", "gender": "unknown", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jingjing Xue", "gender": "unknown", "institution": "Xidian University", "country": "CN", "position": "Postdoc" }, { "name": "Quanxue Gao", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Ran Jing", "gender": "unknown", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Cheng Deng", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" } ], "n_formula": 43, "n_formula_1": 28, "n_ref_uni": 31, "n_ref": 48, "n_ref_all": 71, "n_fig": 6, "n_tab": 3, "L_tab": 1544, "n_element_tab": 276, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1322, "n_element_tab_1": 95, "formula_len_all": 2321, "formula_len_all_1": 1489, "len_all": 98733, "len_all_1": 60559, "len_abs": 1019, "len_title": 102, "len_sents": 25120, "len_sents_1": 23215, "n_sents": 225, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1027, "title": "Discrete Tensorized Label Learning with Anchor Graphs", "abs": "Many discrete multi-view clustering methods based on anchor graphs use the anchor graph decomposition or spectral clustering to obtain the final clustering labels, such methods achieve good results but lack interpretability. Morever, some of them are poorly balanced. To this end, first, we start from the perspective of label transmission to convert labels of the anchors to the labels of the samples, which has better interpretability. Second, we find a new and remarkable use of the nuclear norm, i.e., maximizing the nuclear norm can ensure the balanced clusters, which has the rigorous theoretical proof. Simultaneously, a novel optimisation method based on the first order Taylor expansion is proposed for the nuclear norm. Finally, we introduce the tensor Schatten $p$-norm to fully exploit the spatial structural and complementary information between views, which can obtain aligned label matrices. Extensive experiments have verified the superiority of the proposed method compared with other state-of-the-art methods.", "keywords": [ "multi-view clustering", "tensorized label learning", "Anchor graph", "nuclear norm" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "je8wzxh0a5", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Vincent Y. F. Tan", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Zhirui Chen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 167, "n_formula_1": 48, "n_ref_uni": 31, "n_ref": 71, "n_ref_all": 101, "n_fig": 1, "n_tab": 2, "L_tab": 1040, "n_element_tab": 142, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18575, "formula_len_all_1": 2973, "len_all": 197988, "len_all_1": 73314, "len_abs": 1406, "len_title": 148, "len_sents": 55263, "len_sents_1": 30539, "n_sents": 589, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1526, "title": "Order-Optimal Instance-Dependent Bounds for Offline Reinforcement Learning with Preference Feedback", "abs": "We consider offline reinforcement learning (RL) with preference feedback in which the implicit reward is a linear function of an unknown parameter. Given an offline dataset, our objective consists in ascertaining the optimal action for each state, with the ultimate goal of minimizing the {\\em simple regret}. We propose an algorithm, \\underline{RL} with \\underline{L}ocally \\underline{O}ptimal \\underline{W}eights or {\\sc RL-LOW}, which yields a simple regret of $\\exp ( - \\Omega(n/H) )$ where $n$ is the number of data samples and $H$ denotes an instance-dependent hardness quantity that depends explicitly on the suboptimality gap of each action. Furthermore, we derive a first-of-its-kind instance-dependent lower bound in offline RL with preference feedback. Interestingly, we observe that the lower and upper bounds on the simple regret match order-wise in the exponent, demonstrating order-wise optimality of {\\sc RL-LOW}. In view of privacy considerations in practical applications, we also extend {\\sc RL-LOW} to the setting of $(\\varepsilon,\\delta)$-differential privacy and show, somewhat surprisingly, that the hardness parameter $H$ is unchanged in the asymptotic regime as $n$ tends to infinity; this underscores the inherent efficiency of {\\sc RL-LOW} in terms of preserving the privacy of the observed rewards. Given our focus on establishing instance-dependent bounds, our work stands in stark contrast to previous works that focus on establishing worst-case regrets for offline RL with preference feedback.", "keywords": [ "RLHF" ], "rating_list": [ 6, 8, 3, 5, 5 ], "soundness_list": [ 3, 4, 2, 3, 4 ], "presentation_list": [ 3, 3, 1, 2, 3 ], "contribution_list": [ 3, 3, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "je3GZissZc", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Edward Johns", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Associate Professor" }, { "name": "Vitalis Vosylius", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 44, "n_ref": 65, "n_ref_all": 85, "n_fig": 8, "n_tab": 3, "L_tab": 3044, "n_element_tab": 284, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 668, "formula_len_all_1": 569, "len_all": 173879, "len_all_1": 62153, "len_abs": 1232, "len_title": 113, "len_sents": 59377, "len_sents_1": 30681, "n_sents": 424, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 65, "L_abs": 953, "title": "Instant Policy: In-Context Imitation Learning via Graph Diffusion", "abs": "Following the impressive capabilities of in-context learning with large transformers, In-Context Imitation Learning (ICIL) is a promising opportunity for robotics. We introduce Instant Policy, which learns new tasks instantly from just one or two demonstrations, achieving ICIL through two key components. First, we introduce inductive biases through a graph representation and model ICIL as a graph generation problem using a learned diffusion process, enabling structured reasoning over demonstrations, observations, and actions. Second, we show that such a model can be trained using pseudo-demonstrations \u2013 arbitrary trajectories generated in simulation \u2013 as a virtually infinite pool of training data. Our experiments, in both simulation and reality, show that Instant Policy enables rapid learning of various everyday robot tasks. We also show how it can serve as a foundation for cross-embodiment and zero-shot transfer to language-defined tasks.", "keywords": [ "In-context Imitation Learning", "Robotic Manipulation", "Graph Neural Networks", "Diffusion Models" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "jdpELUL0T6", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Yang Chen", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yitao Liang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhouchen Lin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Professor" } ], "n_formula": 55, "n_formula_1": 19, "n_ref_uni": 33, "n_ref": 59, "n_ref_all": 64, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 43, "n_element_tab_1": 6, "formula_len_all": 6578, "formula_len_all_1": 1785, "len_all": 157726, "len_all_1": 77143, "len_abs": 1416, "len_title": 143, "len_sents": 53143, "len_sents_1": 35094, "n_sents": 498, "n_sents_1": 308, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1430, "title": "Low-Dimension-to-High-Dimension Generalization and Its Implications for Length Generalization", "abs": "Low-Dimension-to-High-Dimension (LDHD) generalization is a special case of Out-of-Distribution (OOD) generalization, where the training data are restricted to a low-dimensional subspace of the high-dimensional testing space. Assuming that each instance is generated from a latent variable and the dimension of the latent variable reflects the problem scale, the inherent scaling challenge in length generalization can be captured by the LDHD generalization in the latent space. We theoretically demonstrate that LDHD generalization is generally unattainable without exploiting prior knowledge to provide appropriate inductive bias. Specifically, we explore LDHD generalization in Boolean functions. We verify that different architectures trained with (S)GD converge to \\emph{min-degree interpolators w.r.t. different linearly independent sets}. LDHD generalization is achievable if and only if the target function coincides with this inductive bias. Applying the insights from LDHD generalization to length generalization, we explain the effectiveness of CoT as changing the structure latent space to enable better LDHD generalization. We also propose a principle for position embedding design to handle both the inherent LDHD generalization and the nuisances such as the data format. Following the principle, we propose a novel position embedding called RPE-Square that remedies the RPE for dealing with the data format nuisance.", "keywords": [ "Length Generalization", "Position Embedding" ], "rating_list": [ 3, 6, 6, 5 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 2, 3, 4 ] }, { "paper_id": "jdFoxDnBwY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Adam Kortylewski", "gender": "unknown", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Research Group Leader" }, { "name": "Alan Yuille", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Angtian Wang", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Chuanruo Ning", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Jiahao Wang", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Jiawei Peng", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Intern" }, { "name": "Yaoyao Liu", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Yining Sun", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 45, "n_ref": 67, "n_ref_all": 94, "n_fig": 10, "n_tab": 5, "L_tab": 1032, "n_element_tab": 152, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 1032, "n_element_tab_1": 152, "formula_len_all": 519, "formula_len_all_1": 519, "len_all": 128852, "len_all_1": 62081, "len_abs": 1621, "len_title": 125, "len_sents": 35890, "len_sents_1": 27959, "n_sents": 283, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1629, "title": "Part321: Recognizing 3D Object Parts from a 2D Image Using 1-Shot Annotations", "abs": "Recognizing object parts from images plays a pivotal role in various real-world applications. However, existing work mostly learn models from large-scale 2D part annotations. In this paper, we propose a part recognition model that can recognize 3D parts from a 2D image with only annotations of parts on one 3D mesh model for each object category. Specifically, we build a category-level 3D feature bank for meshes that could overcome geometric variance among objects and precisely align with diverse 2D images of this object category. To achieve this, we propose to learn two types of correspondence. Firstly, we learn mesh-to-mesh correspondence between distinct 3D mesh models by matching geometry-aware features, which allows us to create a shared 3D feature bank for this object category. Secondly, we establish mesh-to-image correspondence by aligning features in the 3D feature bank with features extracted from 2D images. During inference, given a single image, our method recognizes 3D object parts via a Render-and-Compare approach. It predicts object parts by gradient-based optimizing each part\u2019s 3D configuration, minimizing a feature-level reconstruction loss between the projected 3D features and the image features while ensuring geometric consistency between object parts. The position, rotation, and shape of each part are optimized to match the cues from the image, thus recognizing the 3D parts from a 2D image. Experiments on VehiclePart3D, PartImageNet, and UDA Part dataset show our method outperforms baselines significantly for 2D part segmentation and pioneering 3D part recognition from a single image.", "keywords": [ "3D Vision", "3D from 2D", "Part recognition", "One-shot" ], "rating_list": [ 3, 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 1, 3, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "jckKNzYYA6", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Hongyu Guo", "gender": "Male", "institution": "University of Ottawa", "country": "CA", "position": "Adjunct professor" }, { "name": "Shengchao Liu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" } ], "n_formula": 62, "n_formula_1": 21, "n_ref_uni": 53, "n_ref": 126, "n_ref_all": 180, "n_fig": 15, "n_tab": 8, "L_tab": 4399, "n_element_tab": 414, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 932, "n_element_tab_1": 71, "formula_len_all": 3954, "formula_len_all_1": 877, "len_all": 243820, "len_all_1": 70531, "len_abs": 1639, "len_title": 126, "len_sents": 73679, "len_sents_1": 32050, "n_sents": 606, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1675, "title": "AssembleFlow: Rigid Flow Matching with Inertial Frames for Molecular Assembly", "abs": "Molecular assembly, where a cluster of rigid molecules aggregated into strongly correlated forms, is fundamental to determining the properties of materials. However, traditional numerical methods for simulating this process are computationally expensive, and existing generative models on material generation overlook the rigidity inherent in molecular structures, leading to unwanted distortions and invalid internal structures in molecules. To address this, we introduce AssembleFlow. AssembleFlow leverages inertial frames to establish reference coordinate systems at the molecular level for tracking the orientation and motion of molecules within the cluster. It further decomposes molecular $\\text{SE}(3)$ transformations into translations in $\\mathbb{R}^3$ and rotations in $\\text{SO}(3)$, enabling explicit enforcement of both translational and rotational rigidity during each generation step within the flow matching framework. This decomposition also empowers distinct probability paths for each transformation group, effectively allowing for the separate learning of their velocity functions: the former, moving in Euclidean space, uses linear interpolation (LERP), while the latter, evolving in spherical space, employs spherical linear interpolation (SLERP) with a closed-form solution. Empirical validation on the benchmarking data COD-Cluster17 shows that AssembleFlow significantly outperforms six competitive deep learning baselines by at least 45\\% in assembly matching scores while maintaining 100\\% molecular integrity. Also, it matches the assembly performance of a widely used domain-specific simulation tool while reducing computational cost by 25-fold.", "keywords": [ "rigid flow matching", "inertial frame", "quaternion representation", "material", "molecular assembly", "molecular crystallization" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "jbfDg4DgAk", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Duy Cao Hoang", "gender": "Male", "institution": "VinUniversity", "country": "", "position": "Research Assistant" }, { "name": "ping li", "gender": "Male", "institution": "Rutgers University", "country": "", "position": "Associate Professor" }, { "name": "Rui Chu", "gender": "Male", "institution": "Tufts University", "country": "US", "position": "PhD student" }, { "name": "Thanh Quoc Hung Le", "gender": "Male", "institution": "VinUniversity", "country": "VN", "position": "Undergrad student" }, { "name": "Weijie Zhao", "gender": "unknown", "institution": "Rochester Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Yingjie Lao", "gender": "Male", "institution": "Tufts University", "country": "US", "position": "Associate Professor" }, { "name": "Khoa Doan", "gender": "unknown", "institution": "VinUniversity", "country": "VN", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 6, "n_ref_uni": 28, "n_ref": 73, "n_ref_all": 108, "n_fig": 6, "n_tab": 16, "L_tab": 4064, "n_element_tab": 638, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1576, "n_element_tab_1": 178, "formula_len_all": 524, "formula_len_all_1": 544, "len_all": 177218, "len_all_1": 58352, "len_abs": 1073, "len_title": 102, "len_sents": 61873, "len_sents_1": 26606, "n_sents": 426, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1090, "title": "Sparse Watermarking in LLMs with Enhanced Text Quality", "abs": "With the widespread adoption of Large Language Models (LLMs), concerns about potential misuse have emerged. To this end, watermarking has been adapted to LLM, enabling a simple and effective way to detect and monitor generated text. However, while the existing methods can differentiate between watermarked and unwatermarked text with high accuracy, they often face a trade-off between the quality of the generated text and the effectiveness of the watermarking process. In this work, we present a novel type of LLM watermark, *Sparse Watermark*, which aims to mitigate this trade-off by applying watermarks to a small subset of generated tokens distributed across the text. To demonstrate this type of watermark, we introduce **SpARK**, a **Sp**arse Waterm**ARK** method that achieves sparsity by anchoring watermarked tokens to words that have specific Part-of-Speech (POS) tags. Our experimental results demonstrate that the proposed watermarking scheme achieves high detectability while generating text that outperforms previous LLM watermarking methods in quality across various tasks.", "keywords": [ "watermarking", "large language models" ], "rating_list": [ 3, 3, 1, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "jawV7vhGHw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Arkadiusz Lewicki", "gender": "Male", "institution": "Prometheus Medtech.AI", "country": "PL", "position": "Principal Researcher" }, { "name": "Jacek Tabor", "gender": "Male", "institution": "Jagiellonian University", "country": "PL", "position": "Full Professor" }, { "name": "Magdalena Tr\u0119dowicz", "gender": "Female", "institution": "Jagiellonian University in Krakow", "country": "PL", "position": "PhD student" }, { "name": "Marcin Mazur", "gender": "Male", "institution": "Jagiellonian University", "country": "PL", "position": "Assistant Professor" }, { "name": "Szymon Janusz", "gender": "Male", "institution": "Jagiellonian University in Krakow", "country": "PL", "position": "MS student" }, { "name": "\u0141ukasz Struski", "gender": "Male", "institution": "Jagiellonian University in Krakow", "country": "PL", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 28, "n_ref": 52, "n_ref_all": 69, "n_fig": 8, "n_tab": 3, "L_tab": 528, "n_element_tab": 90, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 341, "n_element_tab_1": 87, "formula_len_all": 572, "formula_len_all_1": 582, "len_all": 103532, "len_all_1": 55528, "len_abs": 1551, "len_title": 125, "len_sents": 36598, "len_sents_1": 26891, "n_sents": 261, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1559, "title": "PrAViC: Probabilistic Adaptation Framework for Real-Time Video Classification", "abs": "Video processing is generally divided into two main categories: processing of the entire video, which typically yields optimal classification outcomes, and real-time processing, where the objective is to make a decision as promptly as possible. The latter is often driven by the need to identify rapidly potential critical or dangerous situations. These could include machine failure, traffic accidents, heart problems, or dangerous behavior. Although the models dedicated to the processing of entire videos are typically well-defined and clearly presented in the literature, this is not the case for online processing, where a plethora of hand-devised methods exist. To address this, we present PrAViC, a novel, unified, and theoretically-based adaptation framework for dealing with the online classification problem for video data. The initial phase of our study is to establish a robust mathematical foundation for the theory of classification of sequential data, with the potential to make a decision at an early stage. This allows us to construct a natural function that encourages the model to return an outcome much faster. The subsequent phase is to present a straightforward and readily implementable method for adapting offline models to the online setting with recurrent operations. Finally, PrAViC is evaluated through comparison with existing state-of-the-art offline and online models and datasets, enabling the network to significantly reduce the time required to reach classification decisions while maintaining, or even enhancing, accuracy.", "keywords": [ "Online classification", "early decision", "video processing" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "jaIxmAVAqF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Bohan Zhuang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Research Professor" }, { "name": "Qi Wu", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "Associate Professor" }, { "name": "Xinyu Wang", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "Lecturer" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 108, "n_ref_all": 121, "n_fig": 4, "n_tab": 6, "L_tab": 916, "n_element_tab": 103, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 82, "n_element_tab_1": 1, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 161703, "len_all_1": 65038, "len_abs": 1251, "len_title": 121, "len_sents": 47146, "len_sents_1": 35730, "n_sents": 305, "n_sents_1": 195, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1256, "title": "Action as a Modality: Turning Multi-Modal LLMs to General Action Planners", "abs": "Large Language Models (LLMs) have demonstrated strong reasoning capabilities and possess extensive common knowledge. This enables them to adapt to a variety of complex tasks in a zero-shot manner, including functioning as controllers to manipulate automated systems and produce executable action sequences. However, a significant challenge in the existing framework is the misalignment between the general pre-trained LLM and the action space of specific control tasks. This misalignment necessitates extensive efforts in designing task-specific prompts, which are less generalizable and do not ensure consistent output when prompting a pre-trained LLM to generate the desired action sequences. To address this issue, we propose a novel solution, ActionVerse, which encodes action candidates into a series of modality tokens, coupled with an efficient alignment technique to synchronize the action tokens with the LLM's language space. By leveraging this approach, the proposed ActionVerse successfully transforms a chat-based multi-modal LLM into a general action executor capable of handling tasks requiring step-by-step execution of various actions. Experiments on several sequential action tasks demonstrate the effectiveness of the proposed framework.", "keywords": [ "Multi-modal Large Language Models", "LLMs", "Action" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 4, 2, 2 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "ja5nxiozl8", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Federico Matteucci", "gender": "unknown", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "", "position": "PhD student" }, { "name": "Klemens B\u00f6hm", "gender": "unknown", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "", "position": "Full Professor" }, { "name": "Konstantin Ntounas", "gender": "unknown", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "MS student" }, { "name": "Marco Heyden", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "", "position": "PhD student" }, { "name": "Vadim Arzamasov", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "", "position": "Postdoc" }, { "name": "Jose A Cribeiro-Ramallo", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 51, "n_ref": 95, "n_ref_all": 107, "n_fig": 5, "n_tab": 2, "L_tab": 122, "n_element_tab": 163, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 276, "n_element_tab_1": 216, "formula_len_all": 1071, "formula_len_all_1": 500, "len_all": 155028, "len_all_1": 76624, "len_abs": 1009, "len_title": 73, "len_sents": 38826, "len_sents_1": 29348, "n_sents": 364, "n_sents_1": 262, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1018, "title": "Can You Trust Your Experiments? Generalizability of Experimental Studies", "abs": "Experimental studies are a cornerstone of Machine Learning (ML) research. \nA common and often implicit assumption is that the study's results will generalize beyond the study itself, e.g., to new data.\nThat is, repeating the same study under different conditions will likely yield similar results. \nExisting frameworks to measure generalizability, borrowed from the casual inference literature, cannot capture the complexity of the results and the goals of an ML study. \nThe problem of measuring generalizability in the more general ML setting is thus still open, also due to the lack of a mathematical formalization of experimental studies.\nIn this paper, we propose such a formalization, use it to develop a framework to quantify generalizability, and propose an instantiation based on rankings and the Maximum Mean Discrepancy.\nWe show how this latter offers insights into the desirable number of experiments for a study. \nFinally, we investigate the generalizability of two recently published experimental studies.", "keywords": [ "Generalizability", "Replicability", "External Validity", "Experiment", "Experimental Study", "Benchmark", "Evaluation" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "ja4rpheN2n", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 8, "author_info_list": [ { "name": "Jimeng Sun", "gender": "unknown", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Professor" }, { "name": "Peng Chen", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Seniro Scientist" }, { "name": "Rikuto Kotoge", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "PhD student" }, { "name": "Xin Liu", "gender": "Male", "institution": "AIST, National Institute of Advanced Industrial Science and Technology", "country": "JP", "position": "Researcher" }, { "name": "Yasuko Matsubara", "gender": "Female", "institution": "Osaka University", "country": "JP", "position": "Associate Professor" }, { "name": "Yasushi Sakurai", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "Full Professor" }, { "name": "Zheng Chen", "gender": "Male", "institution": "ISIR, Osaka University, Japan", "country": "JP", "position": "Assistant Professor" }, { "name": "Ziwei Yang", "gender": "Male", "institution": "Kyoto University", "country": "JP", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 6, "n_ref_uni": 47, "n_ref": 98, "n_ref_all": 122, "n_fig": 6, "n_tab": 11, "L_tab": 4941, "n_element_tab": 436, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2126, "n_element_tab_1": 140, "formula_len_all": 658, "formula_len_all_1": 393, "len_all": 216387, "len_all_1": 66119, "len_abs": 1960, "len_title": 75, "len_sents": 68459, "len_sents_1": 30667, "n_sents": 544, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 75, "L_abs": 1797, "title": "GeSubNet: Gene Interaction Inference for Disease Subtype Network Generation", "abs": "Retrieving gene functional networks from knowledge databases presents a challenge due to the mismatch between disease networks and subtype-specific variations. Current solutions, including statistical and deep learning methods, often fail to effectively integrate gene interaction knowledge from databases or explicitly learn subtype-specific interactions. To address this mismatch, we propose GeSubNet, which learns a unified representation capable of predicting gene interactions while distinguishing between different disease subtypes. Graphs generated by such representations can be considered subtype-specific networks. GeSubNet is a multi-step representation learning framework with three modules: First, a deep generative model learns distinct disease subtypes from patient gene expression profiles. Second, a graph neural network captures representations of prior gene networks from knowledge databases, ensuring accurate physical gene interactions. Finally, we integrate these two representations using an inference loss that leverages graph generation capabilities, conditioned on the patient separation loss, to refine subtype-specific information in the learned representation. GeSubNet consistently outperforms traditional methods, with average improvements of 30.6%, 21.0%, 20.1%, and 56.6% across four graph evaluation metrics, averaged over four cancer datasets. Particularly, we conduct a biological simulation experiment to assess how the behavior of selected genes from over 11,000 candidates affects subtypes or patient distributions. The results show that the generated network has the potential to identify subtype-specific genes with an 83% likelihood of impacting patient distribution shifts. The GeSubNet resource is available: https://anonymous.4open.science/r/GeSubNet/", "keywords": [ "Gene Functional Networks", "Disease Subtypes", "Bioinformatics" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 4, 2, 4 ], "presentation_list": [ 4, 3, 4 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "ja2gQFYA9R", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Chengjie Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Jiangning Zhang", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Jingyu Gong", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Ke Fan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Lizhuang Ma", "gender": "Male", "institution": "Dept. of Computer Sci. & Eng., Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Ran Yi", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xin Tan", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Yabiao Wang", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Senior Researcher" }, { "name": "Yating Wang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 33, "n_ref": 63, "n_ref_all": 76, "n_fig": 4, "n_tab": 1, "L_tab": 642, "n_element_tab": 47, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 34, "n_element_tab_1": 2, "formula_len_all": 284, "formula_len_all_1": 305, "len_all": 125938, "len_all_1": 66787, "len_abs": 1838, "len_title": 140, "len_sents": 40229, "len_sents_1": 35825, "n_sents": 246, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 125, "L_abs": 1890, "title": "Textual $\\textbf{D}$ecomposition then Sub-motion-space $\\textbf{S}$cattering for $\\textbf{O}$pen-Vocabulary Motion Generation", "abs": "Text-to-motion generation is a crucial task in computer vision, which generates the target 3D motion by the given text. The existing annotated datasets are limited in scale, resulting in most existing methods overfitting to the small datasets and unable to generalize to the motions of the open domain. Some methods attempt to solve the open-vocabulary motion generation problem by aligning to the CLIP space or using the Pretrain-then-Finetuning paradigm. However, the current annotated dataset's limited scale only allows them to achieve mapping from sub-text-space to sub-motion-space, instead of mapping between full-text-space and full-motion-space (full mapping), which is the key to attaining open-vocabulary motion generation. To this end, this paper proposes to leverage the atomic motion (simple body part motions over a short time period) as an intermediate representation, and leverage two orderly coupled steps, i.e., Textual Decomposition and Sub-motion-space Scattering, to address the full mapping problem. For Textual Decomposition, we design a fine-grained description conversion algorithm, and combine it with the generalization ability of a large language model to convert any given motion text into atomic texts. Sub-motion-space Scattering learns the compositional process from atomic motions to the target motions, to make the learned sub-motion-space scattered to form the full-motion-space.\nFor a given motion of the open domain, it transforms the extrapolation into interpolation and thereby significantly improves generalization. Our network, $\\textbf{DSO}$-Net, combines textual $\\textbf{d}$ecomposition and sub-motion-space $\\textbf{s}$cattering to solve the $\\textbf{o}$pen-vocabulary motion generation. Extensive experiments demonstrate that our DSO-Net achieves significant improvements over the state-of-the-art methods on open-vocabulary motion generation.", "keywords": [ "Motion Generation", "Open-Vocabulary" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 3, 5 ] }, { "paper_id": "jZwwMxG8PO", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Alex Townsend", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Associate Professor" }, { "name": "Sungwoo Jeong", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Postdoc" } ], "n_formula": 55, "n_formula_1": 22, "n_ref_uni": 43, "n_ref": 82, "n_ref_all": 131, "n_fig": 2, "n_tab": 3, "L_tab": 1085, "n_element_tab": 36, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4448, "formula_len_all_1": 2283, "len_all": 138477, "len_all_1": 67129, "len_abs": 868, "len_title": 113, "len_sents": 41178, "len_sents_1": 29362, "n_sents": 373, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 870, "title": "Extending Mercer's expansion to indefinite and asymmetric kernels", "abs": "Mercer's expansion and Mercer's theorem are cornerstone results in kernel theory. While the classical Mercer's theorem only considers continuous symmetric positive definite kernels, analogous expansions are effective in practice for indefinite and asymmetric kernels. In this paper we extend Mercer's expansion to continuous kernels, providing a rigorous theoretical underpinning for indefinite and asymmetric kernels. We begin by demonstrating that Mercer's expansion may not be pointwise convergent for continuous indefinite kernels, before proving that the expansion of continuous kernels with bounded variation uniformly in each variable separably converges pointwise almost everywhere, almost uniformly, and unconditionally almost everywhere. We also describe an algorithm for computing Mercer's expansion for general kernels and give new decay bounds on its terms.", "keywords": [ "Kernel methods", "Mercer's expansion", "Indefinite kernels", "Convergence behavior." ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "jZw0CWXuDc", "primary_area": "interpretability and explainable AI", "n_author": 14, "author_info_list": [ { "name": "Adithya Pratapa", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Emma Strubell", "gender": "Non-Binary", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Eric P. Xing", "gender": "Male", "institution": "Mohamed bin Zayed Univeristy of AI", "country": "AE", "position": "Full Professor" }, { "name": "Hwijeen Ahn", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Jeff Schneider", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Researcher" }, { "name": "Juhan Bae", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Kewen Zhao", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Minsoo Kang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Roger Baker Grosse", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "Associate Professor" }, { "name": "Sang Keun Choe", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Teruko Mitamura", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Research Professor" }, { "name": "Willie Neiswanger", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Youngseog Chung", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Eduard Hovy", "gender": "Male", "institution": "University of Melbourne", "country": "", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 51, "n_ref": 100, "n_ref_all": 120, "n_fig": 12, "n_tab": 6, "L_tab": 2609, "n_element_tab": 67, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 209, "n_element_tab_1": 18, "formula_len_all": 1584, "formula_len_all_1": 542, "len_all": 250017, "len_all_1": 71022, "len_abs": 1150, "len_title": 120, "len_sents": 96132, "len_sents_1": 34754, "n_sents": 665, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1153, "title": "Large-Scale Training Data Attribution with Efficient Influence Functions", "abs": "Training data attribution (TDA) quantifies the contribution of individual training examples to model predictions, enabling a range of applications such as data curation, data citation, and model debugging. However, applying existing TDA methods to recent large models and training datasets has been largely limited by prohibitive compute and memory costs. In this work, we focus on influence functions, a popular gradient-based TDA method, and significantly improve its scalability with an efficient gradient projection strategy called LoGra that leverages the gradient structure in backpropagation. We then provide a theoretical motivation of gradient projection approaches to influence functions to promote trust in the TDA process. Lastly, we lower the barrier to implementing TDA systems by introducing LogIX, a software package that can transform existing training code into TDA code with minimal effort. In our TDA experiments, LoGra achieves competitive accuracy against more expensive baselines while showing up to 6,500x improvement in throughput and 5x reduction in GPU memory usage when applied to Llama3-8B-Instruct and the 1B-token dataset.", "keywords": [ "data attribution", "influence functions", "LLMs", "interpretability" ], "rating_list": [ 3, 3, 8, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 1, 3, 4, 4 ], "contribution_list": [ 3, 2, 4, 4 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "jZsN9zo8Qi", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Chaoyou Fu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Chenyu Zhou", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Mengdan Zhang", "gender": "Female", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Peixian Chen", "gender": "Female", "institution": "Tencent YOUTU Lab", "country": "CN", "position": "Researcher" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Xiawu Zheng", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Xing Sun", "gender": "Male", "institution": "Tencent YouTu Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Yunhang Shen", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 22, "n_ref": 45, "n_ref_all": 95, "n_fig": 19, "n_tab": 4, "L_tab": 1491, "n_element_tab": 215, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 50, "formula_len_all_1": 50, "len_all": 167886, "len_all_1": 49422, "len_abs": 2024, "len_title": 133, "len_sents": 54880, "len_sents_1": 23700, "n_sents": 434, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1540, "title": "Learning Interleaved Image-Text Comprehension in Vision-Language Large Models", "abs": "The swift progress of Multi-modal Large Models (MLLMs) has showcased their impressive ability to tackle tasks blending vision and language.\nYet, most current models and benchmarks cater to scenarios with a narrow scope of visual and textual contexts.\nThese models often fall short when faced with complex comprehension tasks, which involve navigating through a plethora of irrelevant and potentially misleading information in both text and image forms.\nTo bridge this gap, we introduce a new, more demanding task known as Interleaved Image-Text Comprehension (IITC).\nThis task challenges models to discern and disregard superfluous elements in both images and text to accurately answer questions and to follow intricate instructions to pinpoint the relevant image.\nIn support of this task, we further craft a new VEGA dataset, tailored for the IITC task on scientific content, and devised a subtask, Image-Text Association (ITA), to refine image-text correlation skills.\nOur evaluation of four leading closed-source models, as well as various open-source models using VEGA, underscores the rigorous nature of IITC.\nEven the most advanced models, such as Gemini-1.5-pro and GPT4V, only achieved modest success.\nBy employing a multi-task, multi-scale post-training strategy, we have set a robust baseline for MLLMs on the IITC task, attaining an $85.8\\%$ accuracy rate in image association and a $0.508$ Rouge score. These results validate the effectiveness of our dataset in improving MLLMs capabilities for nuanced image-text comprehension.", "keywords": [ "MLLMs;Benchmark;Interleaved Image-Text Comprehension" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 4, 4, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "jZffxvubJ9", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Chao Yang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Gegenduolan Zhang", "gender": "Female", "institution": "Central South University", "country": "CN", "position": "Undergrad student" }, { "name": "Shuang Li", "gender": "Female", "institution": "The Chinese University of Hong Kong (Shenzhen)", "country": "", "position": "Assistant Professor" }, { "name": "Xinye Chen", "gender": "Female", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Yinghao Fu", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 74, "n_formula_1": 26, "n_ref_uni": 33, "n_ref": 50, "n_ref_all": 61, "n_fig": 1, "n_tab": 5, "L_tab": 1480, "n_element_tab": 125, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7714, "formula_len_all_1": 2189, "len_all": 166713, "len_all_1": 56374, "len_abs": 2057, "len_title": 142, "len_sents": 55412, "len_sents_1": 26168, "n_sents": 418, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1529, "title": "Treatment Rule Optimization Under Counterfactual Temporal Point Processes with Latent States", "abs": "In high-stakes areas like healthcare, retrospective counterfactual analysis\u2014such as evaluating what might have happened if treatments were administered earlier, later, or differently\u2014is vital for refining treatment strategies. This paper proposes a counterfactual treatment optimization framework using temporal point processes to model outcome event sequences. By sampling potential outcome events under new treatment decision rules, our approach seeks to optimize treatment strategies in a counterfactual setting. To achieve accurate counterfactual evaluation of new decision rules, we explicitly introduce latent states into the modeling of temporal point processes. Our method first infers the latent states and associated noise, followed by counterfactual sampling of outcome events. This approach rigorously addresses the complexities introduced by latent states, effectively removing biases in the evaluation of treatment strategies. By proving the identifiability of model parameters in the presence of these states, we provide theoretical guarantees that enhance the reliability and robustness of the counterfactual analysis. By incorporating latent states and proving identifiability, our framework not only improves the accuracy and robustness of treatment decision rules but also offers actionable insights for optimizing healthcare interventions. This method holds significant potential for improving treatment strategies, particularly in healthcare scenarios where patient symptoms are complex and high-dimensional.", "keywords": [ "counterfactual reasoning", "temporal point processes", "latent confounder", "rule learning" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 1, 3, 1 ], "contribution_list": [ 2, 1, 3, 1 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "jZVNmDiU86", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Bofei Gao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Keming Lu", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Tianyu Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Staff Engineer" }, { "name": "Wayne Xiong", "gender": "Male", "institution": "Research, Microsoft", "country": "", "position": "Principal Researcher" }, { "name": "Wen Xiao", "gender": "Female", "institution": "Microsoft", "country": "CA", "position": "Researcher" }, { "name": "Yichi Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yue Dong", "gender": "Female", "institution": "University of California, Riverside", "country": "", "position": "Assistant Professor" }, { "name": "Yuliang Liu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Zefan Cai", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Junjie Hu", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 3, "n_ref_uni": 31, "n_ref": 116, "n_ref_all": 160, "n_fig": 20, "n_tab": 13, "L_tab": 4033, "n_element_tab": 513, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1154, "formula_len_all_1": 100, "len_all": 180488, "len_all_1": 65857, "len_abs": 1519, "len_title": 128, "len_sents": 57335, "len_sents_1": 31482, "n_sents": 400, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1528, "title": "PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling", "abs": "In this study, we investigate whether attention-based information flow inside large language models (LLMs) is aggregated through noticeable patterns for long context processing. Our observations reveal that LLMs aggregate information through \nPyramidal Information Funneling where attention is scattering widely in lower layers, progressively consolidating within specific contexts, and ultimately focusing on critical tokens (a.k.a massive activation or attention sink) in higher layers. Motivated by these insights, we developed PyramidKV, a novel and effective KV cache compression method. This approach dynamically adjusts the KV cache size across different layers, allocating more cache in lower layers and less in higher ones, diverging from traditional methods that maintain a uniform KV cache size. Our experimental evaluations, utilizing the LongBench benchmark, show that PyramidKV matches the performance of models with a full KV cache while retaining only 12\\% of the KV cache, thus significantly reducing memory usage. In scenarios emphasizing memory efficiency, where only 0.7\\% of the KV cache is maintained, PyramidKV surpasses other KV cache compression techniques, achieving up to a 20.5 absolute accuracy improvement on TREC dataset. In the Needle-in-a-Haystack experiment, PyramidKV outperforms competing methods in maintaining long-context comprehension in LLMs; notably, retaining just 128 KV cache entries enables the LLAMA-3-70B model to achieve 100\\% Acc. performance, matching that of a full KV cache.", "keywords": [ "Large Language Models; Efficient Generative Inference" ], "rating_list": [ 8, 5, 6, 6, 3 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 1 ], "contribution_list": [ 3, 2, 2, 3, 1 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "jZASmAlxp2", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Shaofeng Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Xiangdong Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Undergrad student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 59, "n_ref": 142, "n_ref_all": 161, "n_fig": 4, "n_tab": 9, "L_tab": 5465, "n_element_tab": 543, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 3319, "n_element_tab_1": 310, "formula_len_all": 1138, "formula_len_all_1": 1078, "len_all": 182458, "len_all_1": 71793, "len_abs": 1672, "len_title": 174, "len_sents": 44278, "len_sents_1": 27718, "n_sents": 332, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 126, "L_abs": 1472, "title": "Position-Query-Based Autoencoders for View Decoupled Cross Point Cloud Reconstruction and a Self-Supervised Learning Framework", "abs": "Point cloud learning, especially in a self-supervised way without manual labels, has received emerging attention in both vision and learning communities, with its potential utility in wide areas. Most existing generative approaches for point cloud self-supervised learning focus on recovering masked points from visible ones within a single view. Recognizing that a two-view pre-training paradigm inherently introduces greater diversity and variance, it could thus enable more challenging and informative pre-training. Inspired by this, we explore the potential of two-view learning in this domain. In this paper, we propose Point-PQAE, a cross-reconstruction generative paradigm that first generates two decoupled point clouds/views and then reconstructs one from the other. To achieve this goal, we develop a crop mechanism for point cloud view generation for the first time and further propose a novel positional encoding to represent the 3D relative position between the two decoupled views. The cross-reconstruction significantly increases the difficulty of pre-training compared to self-reconstruction, which enables our method to achieve new state-of-the-art results and surpasses previous single-modal self-reconstruction methods in 3D self-supervised learning by a margin. Specifically, it outperforms self-reconstruction baseline (Point-MAE) 6.5\\%, 7.0\\%, 6.7\\% in three variants of ScanObjectNN with Mlp-Linear evaluation protocol. Source code will be released.", "keywords": [ "Representation learning", "Self-supervised learning", "Point Cloud", "Generative Methods" ], "rating_list": [ 8, 5, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 4, 3 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 3, 5, 5, 3, 4 ] }, { "paper_id": "jYyste2HLP", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Tianqi Wei", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Yuxin Huang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "\u4e8e\u7490", "gender": "Female", "institution": "Sun Yat-Sen University", "country": "CN", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 37, "n_ref": 47, "n_ref_all": 72, "n_fig": 14, "n_tab": 6, "L_tab": 1970, "n_element_tab": 183, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 913, "n_element_tab_1": 120, "formula_len_all": 311, "formula_len_all_1": 311, "len_all": 124706, "len_all_1": 64558, "len_abs": 1105, "len_title": 125, "len_sents": 32553, "len_sents_1": 28242, "n_sents": 267, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1109, "title": "FlyOrien: A bio-inspired model for incremental learning of object orientation", "abs": "Visual orientation detection helps navigation, especially without a reliable magnetic compass or GPS. Inspired by the neural mechanisms of the insect brain, particularly the mushroom body (MB) and the central complex (CX), we propose FlyOrien\u2014a bio-inspired model for object orientation detection. The model mimics the MB for random feature extraction, sparse coding and associative learning, while the CX provides multi-clue sensory integration, enabling interpolation for finer orientation representation. FlyOrien's biologically plausible learning rule allows one-shot learning, reducing the need for large datasets and repeated training. We tested FlyOrien on a dataset containing images labeled with orientations, which introduce strong interferences because images of the same object have different labels. In this challenging context, FlyOrien achieves competitive performance compared to convolutional neural networks (CNNs), significantly reducing training time and computational resources. It also has the potential for real-world applications like robotics, where incremental learning is essential.", "keywords": [ "bio-inspired", "sparse coding", "mushroom body", "object orientation", "continuous-attractor neural network" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 3, 2 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "jYP8Cd2bMW", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Carlos Esteve Yague", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "Carola-Bibiane Sch\u00f6nlieb", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "Full Professor" }, { "name": "Fan Zhang", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Michael roberts", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "Principal Researcher" }, { "name": "S\u00f6ren Dittmer", "gender": "Male", "institution": "Universit\u00e4t Bremen", "country": "", "position": "PhD student" } ], "n_formula": 87, "n_formula_1": 25, "n_ref_uni": 29, "n_ref": 47, "n_ref_all": 98, "n_fig": 13, "n_tab": 6, "L_tab": 3598, "n_element_tab": 517, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1030, "n_element_tab_1": 208, "formula_len_all": 6613, "formula_len_all_1": 1705, "len_all": 194745, "len_all_1": 68018, "len_abs": 3459, "len_title": 145, "len_sents": 59840, "len_sents_1": 29543, "n_sents": 497, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1759, "title": "FedMAP: Unlocking Potential in Personalized Federated Learning through Bi-Level MAP Optimization", "abs": "Federated Learning (FL) enables collaborative training of machine learning (ML) models on decentralized data while preserving data privacy. However, data across clients often differs significantly due to class imbalance, feature distribution skew, sample size imbalance, and other phenomena.\nUsing information from these not identically distributed (non-IID) datasets causes challenges in training. Existing FL methods based on a single global model cannot effectively capture client data variations, resulting in suboptimal performance. Personalized FL (PFL) techniques were introduced to adapt to the local data distribution of each client and utilize the data from other clients. \nThey have shown promising results in addressing these challenges. \nWe propose FedMAP, a novel Bayesian PFL framework which applies Maximum A Posteriori (MAP) estimation to effectively mitigate various non-IID data issues, by means of a parametric prior distribution, which is updated during aggregation. We provide a theoretical foundation illustrating FedMAP's convergence properties. In particular, we prove that the prior updates in FedMAP correspond to gradient descent iterations for a linear combination of envelope functions associated with the local losses. This differs from previous FL approaches, that aim at minimizing a weighted average of local loss functions and often face challenges with heterogeneous data distributions, resulting in reduced client performance and slower convergence in non-IID settings. \nFinally, we show, through evaluations of synthetic and real-world datasets, that FedMAP achieves better performance than the existing methods. Moreover, we offer a robust, ready-to-use framework to facilitate practical deployment and further research.", "keywords": [ "Federated Learning", "Personalized Federated Learning", "Non-IID Data Distributions", "Bi-level Optimization" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 2, 5, 3, 3 ] }, { "paper_id": "jYJq2gQb7J", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Jiabin Lin", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "PhD student" }, { "name": "Shana Moothedath", "gender": "Female", "institution": "Iowa State University", "country": "US", "position": "Assistant Professor" }, { "name": "Tuan Anh Le", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "MS student" } ], "n_formula": 100, "n_formula_1": 31, "n_ref_uni": 32, "n_ref": 98, "n_ref_all": 120, "n_fig": 13, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 687, "n_element_tab_1": 7, "formula_len_all": 16693, "formula_len_all_1": 3427, "len_all": 174611, "len_all_1": 80392, "len_abs": 1129, "len_title": 118, "len_sents": 49610, "len_sents_1": 33416, "n_sents": 502, "n_sents_1": 284, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1135, "title": "PROVABLY EFFICIENT FEDERATED ACTIVE MULTI-TASK REPRESENTATION LEARNING", "abs": "Multi-task representation learning is an emerging machine learning paradigm that integrates data from multiple sources, harnessing task similarities to enhance overall model performance. The application of multi-task learning to real-world settings is hindered due to data scarcity, along with challenges related to scalability and computational resources. To address these challenges, we develop a fast and sample-efficient approach for multi-task active learning with linear representation when the amount of data from source tasks and target tasks is limited. By leveraging the techniques from active learning, we propose an adaptive sampling-based alternating projected gradient descent (GD) and minimization algorithm that iteratively estimates the relevance of each source task to the target task and samples from each source task based on the estimated relevance. We present the convergence guarantees and the sample and time complexities of our algorithm. We evaluated the effectiveness of our algorithm using numerical experiments and compared it against four benchmark algorithms using synthetic and real MNIST-C datasets.", "keywords": [ "Representation learning", "alternating gradient descent and minimization", "active learning", "multi-task learning" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 2, 3, 3 ] }, { "paper_id": "jY5oml9fe9", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Ching-Yun Ko", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Georgios Kollias", "gender": "unknown", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Luca Daniel", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Professor" }, { "name": "Payel Das", "gender": "Female", "institution": "IBM, International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Pin-Yu Chen", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Soham Dan", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Subhajit Chaudhury", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Research Scientist" }, { "name": "Tejaswini Pedapati", "gender": "Female", "institution": "International Business Machines", "country": "US", "position": "Research Engineer" }, { "name": "Youssef Mroueh", "gender": "unknown", "institution": "IBM", "country": "", "position": "Research Staff member" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 52, "n_ref": 116, "n_ref_all": 151, "n_fig": 5, "n_tab": 6, "L_tab": 1313, "n_element_tab": 91, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1668, "formula_len_all_1": 871, "len_all": 169160, "len_all_1": 65285, "len_abs": 1364, "len_title": 100, "len_sents": 47112, "len_sents_1": 32565, "n_sents": 369, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1379, "title": "Large Language Models can be Strong Self-Detoxifiers", "abs": "Reducing the likelihood of generating harmful and toxic output is an essential task when aligning large language models (LLMs). Existing methods mainly rely on training an external reward model (i.e., another language model) or fine-tuning the LLM using self-generated data to influence the outcome. In this paper, we show that LLMs have the capability of self-detoxification without the use of an additional reward model or re-training. We propose \\textit{Self-disciplined Autoregressive Sampling (SASA)}, a lightweight controlled decoding algorithm for toxicity reduction of LLMs. SASA leverages the contextual representations from an LLM to learn linear subspaces characterizing toxic v.s. non-toxic output in analytical forms. When auto-completing a response token-by-token, SASA dynamically tracks the margin of the current output to steer the generation away from the toxic subspace, by adjusting the autoregressive sampling strategy. Evaluated on LLMs of different scale and nature, namely Llama-3.1-Instruct (8B), Llama-2 (7B), and GPT2-L models with the RealToxicityPrompts, BOLD, and AttaQ benchmarks, SASA markedly enhances the quality of the generated sentences relative to the original models and attains comparable performance to state-of-the-art detoxification techniques, significantly reducing the toxicity level by only using the LLM's internal representations.", "keywords": [ "detoxification; LLM" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 4, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 5, 3 ] }, { "paper_id": "jY2ow7jRdZ", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Harry Ziming Cheng", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Junting Pan", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Mingjie Zhan", "gender": "Male", "institution": "SenseTime Research", "country": "CN", "position": "Researcher" }, { "name": "Zhiyuan Huang", "gender": "Male", "institution": "Sensetime", "country": "CN", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 57, "n_ref": 134, "n_ref_all": 151, "n_fig": 8, "n_tab": 4, "L_tab": 1257, "n_element_tab": 80, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 296, "n_element_tab_1": 18, "formula_len_all": 380, "formula_len_all_1": 260, "len_all": 189740, "len_all_1": 53557, "len_abs": 1679, "len_title": 99, "len_sents": 44833, "len_sents_1": 25423, "n_sents": 340, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 51, "L_abs": 1717, "title": "SpiritSight Agent: Advanced GUI Agent with One Look", "abs": "Graphical User Interface (GUI) Agents show amazing abilities in assisting human-computer interaction, automating human user's navigation on digital devices. An ideal GUI Agent is expected to achieve high accuracy, low latency, and generality across various GUI platforms. Recent visual-based approaches show promises, taking the advantages of advanced Vision Language Models (VLMs). Although they generally meet the requirements of generality and low latency, these visual-based GUI Agents often fall short in terms of localization accuracy. To address this issue, we propose $\\textbf{SpiritSight}$, a visual-based generalist end-to-end GUI agent with outstanding grounding abilities. First, we create a multi-level, large-scale, high-quality GUI training dataset with scalable methods and train SpiritSight using curriculum learning, empowering it with robust GUI understanding and localization capabilities. Second, we introduce the $\\textbf{Universal Block Parsing (UBP)}$ method, which frames the localization task as a multi-image QA problem, further enhancing SpiritSight's ability to ground GUI objects. With the above-mentioned efforts, SpiritSight constantly outperforms previous SOTA methods across numerous major automated GUI navigation benchmarks. Notably, SpiritSight-8B achieves a 46.1% step Success Rate(SR) on the Mind2Web benchmark without any candidates element input, $\\textbf{more than doubling}$ the performance of SeeClick (20.9%) with a comparable model scale. SpiritSight also outperforms other visual-language-based methods in various GUI platforms, demonstrating its superior capability and compatibility in GUI Agent tasks. The models and the code will be made available upon publications.", "keywords": [ "GUI Agent", "VLLM", "decision-making" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "jXy5B0auu0", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Chunhua Shen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Hao Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Peiyuan Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Weian Mao", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "PhD student" }, { "name": "Xiaoran Jiao", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yu Bo", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Zheng Sun", "gender": "Male", "institution": "Swansea University", "country": "GB", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 34, "n_ref": 64, "n_ref_all": 76, "n_fig": 8, "n_tab": 8, "L_tab": 5889, "n_element_tab": 411, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 309, "n_element_tab_1": 43, "formula_len_all": 855, "formula_len_all_1": 934, "len_all": 159267, "len_all_1": 66291, "len_abs": 1004, "len_title": 117, "len_sents": 42889, "len_sents_1": 28409, "n_sents": 349, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1008, "title": "MoDiff: a Protein Diffusion Model for Multi-motif Scaffolding Problem", "abs": "The motif scaffolding problem, crucial in drug and enzyme design, involves creating a scaffold to support one or more specified motifs. Existing methods employ the paradigm of inpainting, thereby limiting their applicability to scenarios where only a single motif is present or the positions between multiple motifs are known. However, in many significant scenarios, the positions between motifs are unknown, leaving the multi-motif scaffolding problem open. To tackle this challenge, we introduce a protein diffusion model called MoDiff. During the diffusion process, MoDiff implicitly assigns motifs to the protein backbone, thereby achieving the automatic design of relative positions among motifs. Our experiments demonstrate that MoDiff can: 1) solve the multi-motif scaffolding problem even when the positions between motifs are unknown, and 2) generate diverse scaffolds based on multiple given motifs. This indicates that MoDiff is a potential general solution to the multi-motif scaffolding problem.", "keywords": [ "Motif Scaffolding", "Drug Design", "Protein Modeling", "Protein Diffusion Model" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "jXvwJ51vcK", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Ender Konukoglu", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Associate Professor" }, { "name": "Guolei Sun", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Min Wu", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Ming-Ming Cheng", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Full Professor" }, { "name": "Runjia Li", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Serge Belongie", "gender": "Male", "institution": "University of Copenhagen", "country": "DK", "position": "Full Professor" }, { "name": "Yun Liu", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Full Professor" }, { "name": "Zhaochong An", "gender": "Male", "institution": "University of Copenhagen", "country": "DK", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 69, "n_ref": 163, "n_ref_all": 183, "n_fig": 8, "n_tab": 6, "L_tab": 913, "n_element_tab": 92, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 700, "n_element_tab_1": 67, "formula_len_all": 983, "formula_len_all_1": 915, "len_all": 178305, "len_all_1": 62678, "len_abs": 1526, "len_title": 113, "len_sents": 47226, "len_sents_1": 28172, "n_sents": 338, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 65, "L_abs": 1533, "title": "Multimodality Helps Few-Shot 3D Point Cloud Semantic Segmentation", "abs": "Few-shot 3D point cloud segmentation (FS-PCS) aims at generalizing models to segment novel categories with minimal annotated support samples. While existing FS-PCS methods have shown promise, they primarily focus on unimodal point cloud inputs, overlooking the potential benefits of leveraging multimodal information. In this paper, we address this gap by introducing a cost-free multimodal FS-PCS setup, utilizing textual labels and the potentially available 2D image modality. Under this easy-to-achieve setup, we present the MultiModal Few-Shot SegNet (MM-FSS), a model effectively harnessing complementary information from multiple modalities. MM-FSS employs a shared backbone with two heads to extract intermodal and unimodal visual features, and a pretrained text encoder to generate text embeddings. To fully exploit the multimodal information, we propose a Multimodal Correlation Fusion (MCF) module to generate multimodal correlations, and a Multimodal Semantic Fusion (MSF) module to refine the correlations using text-aware semantic guidance. Additionally, we propose a simple yet effective Test-time Adaptive Cross-modal Calibration (TACC) technique to mitigate training bias, further improving generalization. Experimental results on S3DIS and ScanNet datasets demonstrate significant performance improvements achieved by our method. The efficacy of our approach indicates the benefits of leveraging commonly-ignored free modalities for FS-PCS, providing valuable insights for future research. The code will be released.", "keywords": [ "few-shot 3D point cloud semantic segmentation", "multimodality" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 5, 3, 3, 5 ] }, { "paper_id": "jXrXTuvA3L", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Lu Ren", "gender": "Female", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Wang Yao", "gender": "Male", "institution": "Beihang University", "country": "", "position": "Lecturer" }, { "name": "Xiao Zhang", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "Full Professor" }, { "name": "Liu Jinwei", "gender": "unknown", "institution": "Beihang University", "country": "CN", "position": "PhD student" } ], "n_formula": 51, "n_formula_1": 22, "n_ref_uni": 7, "n_ref": 12, "n_ref_all": 38, "n_fig": 0, "n_tab": 2, "L_tab": 2138, "n_element_tab": 107, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 450, "n_element_tab_1": 18, "formula_len_all": 5458, "formula_len_all_1": 2151, "len_all": 113748, "len_all_1": 53217, "len_abs": 3215, "len_title": 148, "len_sents": 39571, "len_sents_1": 22799, "n_sents": 344, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1384, "title": "NF-MKV Net: A Constraint-Preserving Neural Network Approach to Solving Mean-Field Games Equilibrium", "abs": "Neural network-based methods for solving Mean-Field Games (MFGs) equilibrium have gained significant attention due to their effectiveness in high-dimensional settings. However, many algorithms face the problem that the density distribution evolution does not satisfy the mathematical constraints in the solution. This paper explores the neural network solution of MFGs equilibrium from the perspective of stochastic process, while coupling process-regularized Normalizing Flow (NF) frameworks and state-policy-connected time series neural networks to solve the McKean-Vlasov type Forward-Backward Stochastic Differential Equations (MKV FBSDEs) fixed point problems, which is equivalent of the MFGs equilibrium. First, we convert the MFGs equilibrium to MKV FBSDEs which introduce the density distribution into equations coefficients within a probabilistic framework, and construct neural networks to approximate value functions and gradients based on these equations. Second, we employ NF architectures\u2014generative neural network models\u2014and by imposing loss constraints on each density transfer function, the algorithm ensures compliance with volumetric-invariance and time-continuity. Additionally, this paper provides theoretical proofs of the algorithm's validity and demonstrates its applicability across various scenarios, showcasing its effective compared to existing approaches.", "keywords": [ "Mean-Field Games", "Normalizing Flow", "McKean-Vlasov type Forward-Backward Stochastic Differential Equations", "Mathematical Constraints Neural Network" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "jXLiDKsuDo", "primary_area": "reinforcement learning", "n_author": 10, "author_info_list": [ { "name": "Donghu Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Dongyoon Hwang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Hojoon Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Hyunseung Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jaegul Choo", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Jun Jet Tai", "gender": "Male", "institution": "Coventry University", "country": "", "position": "PhD student" }, { "name": "Kaushik Subramanian", "gender": "Male", "institution": "Sony Europe Ltd.", "country": "", "position": "Researcher" }, { "name": "Peter R. Wurman", "gender": "Male", "institution": "North Carolina State University", "country": "", "position": "Assistant Professor" }, { "name": "Peter Stone", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Principal Researcher" }, { "name": "Takuma Seno", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 6, "n_ref_uni": 61, "n_ref": 144, "n_ref_all": 173, "n_fig": 14, "n_tab": 14, "L_tab": 5432, "n_element_tab": 653, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 272, "n_element_tab_1": 8, "formula_len_all": 687, "formula_len_all_1": 244, "len_all": 217129, "len_all_1": 60473, "len_abs": 1372, "len_title": 127, "len_sents": 56452, "len_sents_1": 27833, "n_sents": 449, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 79, "L_abs": 1327, "title": "SimBa: Simplicity Bias for Scaling Up Parameters in Deep Reinforcement Learning", "abs": "Recent advances in CV and NLP have been largely driven by scaling up the number of network parameters, despite traditional theories suggesting that larger networks are prone to overfitting.\nThese large networks avoid overfitting by integrating components that induce a simplicity bias, guiding models toward simple and generalizable solutions. \nHowever, in deep RL, designing and scaling up networks have been less explored.\nMotivated by this opportunity, we present SimBa, an architecture designed to scale up parameters in deep RL by injecting a simplicity bias. SimBa consists of three components: (i) an observation normalization layer that standardizes inputs with running statistics, (ii) a residual feedforward block to provide a linear pathway from the input to output, and (iii) a layer normalization to control feature magnitudes. \nBy scaling up parameters with SimBa, the sample efficiency of various deep RL algorithms\u2014including off-policy, on-policy, and unsupervised methods\u2014is consistently improved.\nMoreover, solely by integrating SimBa architecture into SAC, it matches or surpasses state-of-the-art deep RL methods with high computational efficiency across DMC, MyoSuite, and HumanoidBench.\nThese results demonstrate SimBa's broad applicability and effectiveness across diverse RL algorithms and environments.", "keywords": [ "reinforcement learning" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 4, 3, 4, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "jWQf6jk55V", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Abhimanyu Hans", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Avi Schwarzschild", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Roman Levin", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Tom Goldstein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Valeriia Cherepanova", "gender": "Female", "institution": "Amazon", "country": "", "position": "Postdoc" } ], "n_formula": 8, "n_formula_1": 1, "n_ref_uni": 51, "n_ref": 83, "n_ref_all": 112, "n_fig": 6, "n_tab": 5, "L_tab": 398, "n_element_tab": 49, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 293, "n_element_tab_1": 45, "formula_len_all": 1431, "formula_len_all_1": 1109, "len_all": 173233, "len_all_1": 66242, "len_abs": 3558, "len_title": 129, "len_sents": 50080, "len_sents_1": 29344, "n_sents": 431, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1023, "title": "Has My System Prompt Been Used? Large Language Model Prompt Membership Inference", "abs": "Prompt engineering has emerged as a powerful technique for optimizing large language models (LLMs) for specific applications, enabling faster prototyping and improved performance, and giving rise to the interest of the community in protecting proprietary system prompts. In this work, we explore a novel perspective on prompt privacy through the lens of membership inference. We develop Prompt Detective, a statistical method to reliably determine whether a given system prompt was used by a third-party language model. Our approach relies on a statistical test comparing the distributions of two groups of generations corresponding to different system prompts. Through extensive experiments with a variety of language models, we demonstrate the effectiveness of Prompt Detective in both standard and challenging scenarios, including black-box settings. Our work reveals that even minor changes in system prompts manifest in distinct response distributions, enabling us to verify prompt usage with statistical significance.", "keywords": [ "privacy", "membership inference attack", "prompt extraction" ], "rating_list": [ 5, 3, 6, 8, 8 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "jVDPq9EdzT", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Kurt Keutzer", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Wenzhao Zheng", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Xiaonan Huang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" }, { "name": "Ye Li", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "MS student" } ], "n_formula": 33, "n_formula_1": 36, "n_ref_uni": 40, "n_ref": 77, "n_ref_all": 93, "n_fig": 7, "n_tab": 2, "L_tab": 2349, "n_element_tab": 317, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 1174, "n_element_tab_1": 162, "formula_len_all": 2770, "formula_len_all_1": 2711, "len_all": 134249, "len_all_1": 61045, "len_abs": 1559, "len_title": 123, "len_sents": 29618, "len_sents_1": 22813, "n_sents": 250, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1567, "title": "UniDrive: Towards Universal Driving Perception Across Camera Configurations", "abs": "Vision-centric autonomous driving has demonstrated excellent performance with economical sensors. As the fundamental step, 3D perception aims to infer 3D information from 2D images based on 3D-2D projection. This makes driving perception models susceptible to sensor configuration (e.g., camera intrinsics and extrinsics) variations. However, generalizing across camera configurations is important for deploying autonomous driving models on different car models. In this paper, we present UniDrive, a novel framework for vision-centric autonomous driving to achieve universal perception across camera configurations. We deploy a set of unified virtual cameras and propose a ground-aware projection method to effectively transform the original images into these unified virtual views. We further propose a virtual configuration optimization method by minimizing the expected projection error between original cameras and virtual cameras. The proposed virtual camera projection can be applied to existing 3D perception methods as a plug-and-play module to mitigate the challenges posed by camera parameter variability, resulting in more adaptable and reliable driving perception models. To evaluate the effectiveness of our framework, we collect a dataset on Carla by driving the same routes while only modifying the camera configurations. Experimental results demonstrate that our method trained on one specific camera configuration can generalize to varying configurations with minor performance degradation. The code and benchmark toolkit will be publicly available.", "keywords": [ "Autonomous Driving", "3D Detection", "Sensor Configuration" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "jUxzh1bi3i", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Kuangdai Leng", "gender": "Male", "institution": "Science and Technology Facilities Council", "country": "", "position": "Researcher" }, { "name": "Ruihua Zhang", "gender": "Female", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Shiji Xin", "gender": "Male", "institution": "Harvard University", "country": "", "position": "MS student" }, { "name": "Shiqian Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xiaodong Song", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Yixin Zhu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhancun Mu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhi Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Postdoc" }, { "name": "Zhixiang Dai", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 1, "n_ref_uni": 74, "n_ref": 97, "n_ref_all": 111, "n_fig": 9, "n_tab": 7, "L_tab": 2995, "n_element_tab": 301, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1175, "n_element_tab_1": 140, "formula_len_all": 704, "formula_len_all_1": 43, "len_all": 210978, "len_all_1": 62765, "len_abs": 1528, "len_title": 133, "len_sents": 63825, "len_sents_1": 30609, "n_sents": 511, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1285, "title": "GlobalTomo: A global dataset for physics-ML seismic wavefield modeling and FWI", "abs": "Global seismic tomography, taking advantage of seismic waves from natural earthquakes, provides essential insights into the earth's internal dynamics. Advanced FWI techniques, whose aim is to meticulously interpret every detail in seismograms, confront formidable computational demands in forward modeling and adjoint simulations on a global scale. Recent advancements in ML offer a transformative potential for accelerating the computational efficiency of FWI and extending its applicability to larger scales. This work presents the first 3D global synthetic dataset tailored for seismic wavefield modeling and full-waveform tomography, referred to as the GlobalTomo dataset. This dataset is uniquely comprehensive, incorporating explicit wave physics and robust geophysical parameterization at realistic global scales, generated through state-of-the-art forward simulations optimized for 3D global wavefield calculations. Through extensive analysis and the establishment of ML baselines, we illustrate that ML approaches are particularly suitable for global FWI, overcoming its limitations with rapid forward modeling and flexible inversion strategies. This work represents a cross-disciplinary effort to enhance our understanding of the earth's interior through physics-ML modeling.", "keywords": [ "Seismic modeling", "inverse problem", "neural operator" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 1, 4, 3 ], "contribution_list": [ 3, 2, 4, 2 ], "confidence_list": [ 5, 3, 5, 5 ] }, { "paper_id": "jUsrbOuQ5e", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Binghuai Lin", "gender": "unknown", "institution": "Tencent", "country": "CN", "position": "Principal Researcher" }, { "name": "Liyuan Wang", "gender": "Female", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Xuehao Wang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 40, "n_ref": 82, "n_ref_all": 100, "n_fig": 8, "n_tab": 7, "L_tab": 3827, "n_element_tab": 447, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3850, "n_element_tab_1": 388, "formula_len_all": 862, "formula_len_all_1": 709, "len_all": 138109, "len_all_1": 71725, "len_abs": 1341, "len_title": 107, "len_sents": 34826, "len_sents_1": 28774, "n_sents": 277, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1346, "title": "HeadMap: Locating and Enhancing Knowledge Circuits in LLMs", "abs": "Large language models (LLMs), through pretraining on extensive corpora, encompass rich semantic knowledge and exhibit the potential for efficient adaptation to diverse downstream tasks. However, the intrinsic mechanisms underlying LLMs remain unexplored, limiting the efficacy of applying these models to downstream tasks. In this paper, we explore the intrinsic mechanisms of LLMs from the perspective of knowledge circuits. Specifically, considering layer dependencies, we propose a layer-conditioned locating algorithm to identify a series of attention heads, which is a knowledge circuit of some tasks. Experiments demonstrate that simply masking a small portion of attention heads in the knowledge circuit can significantly reduce the model's ability to make correct predictions. This suggests that the knowledge flow within the knowledge circuit plays a critical role when the model makes a correct prediction. Inspired by this observation, we propose a novel parameter-efficient fine-tuning method called HeadMap, which maps the activations of these critical heads in the located knowledge circuit to the residual stream by two linear layers, thus enhancing knowledge flow from the knowledge circuit in the residual stream. Extensive experiments conducted on diverse datasets demonstrate the efficiency and efficacy of the proposed method.", "keywords": [ "Large Language Model", "Parameter-Efficient Fine-Tuning", "Interpretability" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "jUKNY4u11K", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Rami Puzis", "gender": "Male", "institution": "Ben Gurion University of the Negev", "country": "IL", "position": "Associate Professor" }, { "name": "Roy Ambar", "gender": "Male", "institution": "Moonshot Marketing", "country": "", "position": "Data Scientist Team Leader" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 59, "n_ref_all": 89, "n_fig": 13, "n_tab": 1, "L_tab": 349, "n_element_tab": 58, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 143, "n_element_tab_1": 33, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 111644, "len_all_1": 61820, "len_abs": 1444, "len_title": 78, "len_sents": 34507, "len_sents_1": 30247, "n_sents": 242, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 30, "L_abs": 1459, "title": "WHAT YOU PAINT IS WHAT YOU GET", "abs": "The two most prominent approaches for building adversary-resilient image classification models are adversarial training and input transformations. Despite significant advancements, adversarial training approaches struggle to generalize to unseen attacks, and the effectiveness of input transformations diminishes fast in the face of large perturbations. In general, there is a large space for improving the inherent trade-off between the accuracy and robustness of adversary-resilient models. Painting algorithms, which have not been used in adversarial training pipelines so far, capture core visual elements of images and offer a potential solution to the challenges faced by current defenses. This paper reveals a correlation between the magnitude of perturbations and the granularity of the painting process required to maximize the classification accuracy. We leverage this correlation in the proposed Painter-CLassifier-Decisioner (PCLD) framework, which employs adversarial training to build an ensemble of classifiers applied to a sequence of paintings with varying detalization. Benchmarks using provable adaptive attack techniques demonstrate the favorable performance of PCLD compared to state-of-the-art defenses, balancing accuracy and robustness while generalizing to unseen attacks. It extends robustness against substantial perturbations in high-resolution settings across various white-box attack methods under $\\ell_\\infty$-norm constraints.", "keywords": [ "adversarial resilience", "painting algorithms", "adversarial manipulation", "defense mechanisms" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "jUCtGezFwH", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Adnan Ahmad", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Postdoc" }, { "name": "Bahareh nakisa", "gender": "unknown", "institution": "Deakin University", "country": "AU", "position": "Assistant Professor" }, { "name": "Mohammad Naim Rastgoo", "gender": "unknown", "institution": "University of Adelaide", "country": "AU", "position": "Lecturer" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 23, "n_ref": 43, "n_ref_all": 60, "n_fig": 6, "n_tab": 4, "L_tab": 3345, "n_element_tab": 298, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 10, "n_element_tab_1": 2, "formula_len_all": 417, "formula_len_all_1": 393, "len_all": 134017, "len_all_1": 64527, "len_abs": 1416, "len_title": 160, "len_sents": 42156, "len_sents_1": 33771, "n_sents": 273, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1421, "title": "Online Sequential Learning from Physiological Data with Weighted Prototypes: Tackling Cross-Subject Variability", "abs": "Online Continual Learning (OCL) enables machine learning models to adapt to sequential data streams in real-time, especially when only a small amount of data is available. However, applying OCL to physiological data such as electroencephalography (EEG) and electrocardiography (ECG) is often complicated by inter-subject variability, which can lead to catastrophic forgetting and performance degradation. Existing OCL methods are currently unable to effectively address this challenge, leading to difficulties in retaining previously learned knowledge while adapting to new data. This paper presents Online Prototypes Weighted Aggregation (OPWA), a novel method specifically designed to address the problem of catastrophic forgetting in the presence of inter-subject variability through the use of prototypical networks. OPWA facilitates the retention of knowledge from past subjects while adapting to new data streams.\nThe OPWA method uses an innovative prototype aggregation mechanism that fuses intra-class prototypes into generalized representations by accounting for both within-class and inter-class variation between subjects. Extensive experiments show that OPWA consistently outperforms existing OCL methods in terms of fast adaptation and mitigation of catastrophic forgetting on different physiological datasets with different modalities, and provides a robust solution for learning on sequential data streams.", "keywords": [ "Online Continual Learning", "Physiological Signals", "Cross-Subject Variability" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "jTnHyyGYy2", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Alexander Becker", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Dominik J. M\u00fchlematter", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Dominik Narnhofer", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Helge Aasen", "gender": "Male", "institution": "Agroscope", "country": "CH", "position": "Principal Researcher" }, { "name": "Konrad Schindler", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Professor" }, { "name": "Mehmet Ozgur Turkoglu", "gender": "Male", "institution": "Agroscope", "country": "CH", "position": "Postdoc" }, { "name": "Michelle Halbheer", "gender": "Female", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 5, "n_ref_uni": 47, "n_ref": 97, "n_ref_all": 121, "n_fig": 15, "n_tab": 15, "L_tab": 8219, "n_element_tab": 539, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2262, "n_element_tab_1": 108, "formula_len_all": 568, "formula_len_all_1": 188, "len_all": 195583, "len_all_1": 61904, "len_abs": 1375, "len_title": 122, "len_sents": 62888, "len_sents_1": 29606, "n_sents": 480, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1382, "title": "LoRA-Ensemble: Efficient Uncertainty Modelling for Self-attention Networks", "abs": "Numerous crucial tasks in real-world decision-making rely on machine learning algorithms with calibrated uncertainty estimates. However, modern methods often yield overconfident and uncalibrated predictions. Various approaches involve training an ensemble of separate models to quantify the uncertainty related to the model itself, known as epistemic uncertainty. In an explicit implementation, the ensemble approach has high computational cost and high memory requirements. This particular challenge is evident in state-of-the-art neural networks such as transformers, where even a single network is already demanding in terms of compute and memory. Consequently, efforts are made to emulate the ensemble model without actually instantiating separate ensemble members, referred to as implicit ensembling. We introduce LoRA-Ensemble, a parameter-efficient deep ensemble method for self-attention networks, which is based on Low-Rank Adaptation (LoRA). Initially developed for efficient LLM fine-tuning, we extend LoRA to an implicit ensembling approach. By employing a single pre-trained self-attention network with weights shared across all members, we train member-specific low-rank matrices for the attention projections. Our method exhibits superior calibration compared to explicit ensembles and achieves similar or better accuracy across various prediction tasks and datasets.", "keywords": [ "uncertainty", "ensemble", "implicit ensemble", "calibration", "self-attention", "transformer" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 2, 3, 4 ] }, { "paper_id": "jTEKTdI3K9", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Arda Senocak", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Postdoc" }, { "name": "Joon Son Chung", "gender": "unknown", "institution": "KAIST", "country": "KR", "position": "Associate Professor" }, { "name": "JungMok Lee", "gender": "Female", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Kim Sung-Bin", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Oh Hyun-Bin", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Tae-hyun Oh", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 51, "n_ref": 192, "n_ref_all": 234, "n_fig": 15, "n_tab": 9, "L_tab": 2721, "n_element_tab": 454, "n_fig_1": 9, "n_tab_1": 8, "L_tab_1": 2891, "n_element_tab_1": 446, "formula_len_all": 0, "formula_len_all_1": 109, "len_all": 230390, "len_all_1": 81715, "len_abs": 1368, "len_title": 134, "len_sents": 73007, "len_sents_1": 34940, "n_sents": 547, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1376, "title": "AVHBench: A Cross-Modal Hallucination Benchmark for Audio-Visual Large Language Models", "abs": "Following the success of Large Language Models (LLMs), expanding their boundaries to new modalities represents a significant paradigm shift in multimodal understanding. Human perception is inherently multimodal, relying not only on text but also on auditory and visual cues for a complete understanding of the world. In recognition of this fact, audio-visual LLMs have recently emerged. Despite promising developments, the lack of dedicated benchmarks poses challenges for understanding and evaluating models. In this work, we show that audio-visual LLMs struggle to discern subtle relationships between audio and visual signals, leading to hallucinations, underscoring the need for reliable benchmarks. To address this, we introduce AVHBench, the first comprehensive benchmark specifically designed to evaluate the perception and comprehension capabilities of audio-visual LLMs. Our benchmark includes tests for assessing hallucinations, as well as the cross-modal matching and reasoning abilities of these models. Our results reveal that most existing audio-visual LLMs struggle with hallucinations caused by cross-interactions between modalities, due to their limited capacity to perceive complex multimodal signals and their relationships. Additionally, we demonstrate that simple training with our AVHBench improves robustness of audio-visual LLMs against hallucinations.", "keywords": [ "Multi-modal Large Language Models", "Hallucination in Large Language Models", "Audio-visual Learning" ], "rating_list": [ 6, 5, 6, 8, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 3, 4, 3 ] }, { "paper_id": "jRZ1ZeenZ6", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Theodore Sumers", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Thomas L. Griffiths", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "Professor" }, { "name": "Nicol\u00f2 De Sabbata", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 53, "n_ref": 97, "n_ref_all": 117, "n_fig": 3, "n_tab": 4, "L_tab": 1195, "n_element_tab": 84, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 968, "n_element_tab_1": 63, "formula_len_all": 856, "formula_len_all_1": 856, "len_all": 141193, "len_all_1": 54228, "len_abs": 969, "len_title": 96, "len_sents": 33739, "len_sents_1": 24781, "n_sents": 243, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 974, "title": "Rational Metareasoning for Large Language Models", "abs": "Being prompted to engage in reasoning has emerged as a core technique for using large language models (LLMs), deploying additional inference-time compute to improve task performance. However, as LLMs increase in both size and adoption, inference costs are correspondingly becoming increasingly burdensome. How, then, might we optimize reasoning's cost-performance tradeoff? This work introduces a novel approach based on computational models of metareasoning used in cognitive science, training LLMs to selectively use intermediate reasoning steps only when necessary. We first develop a reward function that incorporates the Value of Computation by penalizing unnecessary reasoning, then use this reward function with Expert Iteration to train the LLM. Compared to few-shot chain-of-thought prompting and STaR, our method significantly reduces inference costs (20-37\\% fewer tokens generated across three models) while maintaining task performance across diverse datasets.", "keywords": [ "LLM", "Metareasoning", "Problem solving", "Chain of Thought", "Inference Optimization", "Value of Computation" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3, 1 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 3, 4 ] }, { "paper_id": "jR6YMxVG9i", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Anh Tuan Luu", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "Assistant Professor" }, { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Bryan Hooi", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "See-Kiong Ng", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Shiyun Xiong", "gender": "Female", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "Intern" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Yifan Zhang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zhiyuan Hu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 21, "n_ref": 46, "n_ref_all": 61, "n_fig": 6, "n_tab": 6, "L_tab": 553, "n_element_tab": 94, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 322, "n_element_tab_1": 46, "formula_len_all": 244, "formula_len_all_1": 243, "len_all": 112140, "len_all_1": 63422, "len_abs": 1968, "len_title": 124, "len_sents": 36352, "len_sents_1": 29229, "n_sents": 281, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1278, "title": "Guiding VLM Agents with Process Rewards at Inference Time for GUI Navigation", "abs": "Recent advancements in visual language models (VLMs) have notably enhanced their capabilities in handling complex Graphical User Interface (GUI) interaction tasks. Despite these improvements, current frameworks often struggle to generate correct actions in challenging GUI environments. State-of-the-art commercial VLMs are black-boxes, and fine-tuning open-source VLMs for GUI tasks requires significant resources. Additionally, existing trajectory-level evaluation and refinement techniques frequently fall short due to delayed feedback and local optimization issues. To address these challenges, we propose an approach that guides VLM agents with process supervision by a reward model during GUI navigation and control at inference time. This guidance allows the VLM agent to optimize actions at each inference step, thereby improving performance in both static and dynamic environments. In particular, our method demonstrates significant performance gains in the GUI navigation task setting, achieving a around 5\\% improvement in action accuracy for static environments and a near 15\\% increase in task success rate in dynamic environments. With further integration of trajectory reflection and retry mechanisms, we also demonstrate even greater enhancement in task success.", "keywords": [ "Visual Language Model", "Agent", "GUI", "Process Reward Model" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "jQyKywGtpW", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Jiafei Lyu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Qi Liu", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Xiu Li", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Zhongjian Qiao", "gender": "unknown", "institution": "Tsinghua Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 38, "n_formula_1": 8, "n_ref_uni": 69, "n_ref": 128, "n_ref_all": 172, "n_fig": 20, "n_tab": 11, "L_tab": 4681, "n_element_tab": 733, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1269, "n_element_tab_1": 79, "formula_len_all": 4646, "formula_len_all_1": 550, "len_all": 233855, "len_all_1": 62953, "len_abs": 1431, "len_title": 126, "len_sents": 60645, "len_sents_1": 29609, "n_sents": 552, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1467, "title": "Offline-to-Online Reinforcement Learning with Prioritized Experience Selection", "abs": "Offline-to-online reinforcement learning (O2O RL) offers a promising paradigm that first pre-trains an offline policy and fine-tunes it with further online interactions. Nevertheless, the distribution shift between the offline and online phase often hinders the fine-tuning performance, sometimes even incurring performance collapse. Existing methods mitigate this by enhancing training robustness with Q-ensemble, training a density ratio estimator to balance offline and online data, etc. But they often rely on components like ensemble and have higher training costs. In this paper, we address this issue by establishing a concrete performance bound for the optimal policies between two consecutive online steps. Motivated by the theoretical insight, we propose a simple yet effective fine-tuning method, \\textbf{P}rioritized \\textbf{E}xperience \\textbf{S}election (PES). During the online stage, PES maintains a dynamically updated priority queue containing a portion of high-return trajectories, and only selects online samples that are close to the samples in the queue for fine-tuning. In this way, the distribution shift issue can be mitigated and the fine-tuning performance can be boosted. PES is computationally efficient and compatible with numerous approaches. Experimental results on a variety of D4RL datasets show that PES can benefit different offline and O2O RL algorithms and enhance Q-value estimate. Our code is available and will be open-source.", "keywords": [ "Reinforcement Learning; Offline-to-Online Reinforcement Learning; Prioritized Experience Selection" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "jQnXDGxdDG", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Dania Herzalla", "gender": "Female", "institution": "Technology Innovation Institute", "country": "AE", "position": "Researcher" }, { "name": "Martin Andreoni", "gender": "Male", "institution": "Technology Innovation Institute", "country": "AE", "position": "Principal Researcher" }, { "name": "Willian Tessaro Lunardi", "gender": "Male", "institution": "Technology Innovation Institute", "country": "AE", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 50, "n_ref": 142, "n_ref_all": 172, "n_fig": 7, "n_tab": 16, "L_tab": 13249, "n_element_tab": 1169, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1180, "n_element_tab_1": 193, "formula_len_all": 751, "formula_len_all_1": 477, "len_all": 191821, "len_all_1": 68484, "len_abs": 1926, "len_title": 110, "len_sents": 49070, "len_sents_1": 32621, "n_sents": 356, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1670, "title": "Learning In-Distribution Representations for Anomaly Detection", "abs": "Anomaly detection involves identifying data patterns that deviate from the anticipated norm. Traditional methods struggle in high-dimensional spaces due to the curse of dimensionality. In recent years, self-supervised learning, particularly through contrastive objectives, has driven advances in anomaly detection by generating compact and discriminative feature spaces. However, vanilla contrastive learning faces challenges like class collision, especially when the In-Distribution (ID) consists primarily of normal, homogeneous data, where the lack of semantic diversity leads to increased overlap between positive and negative pairs. Existing methods attempt to address these issues by introducing hard negatives through synthetic outliers, Outlier Exposure (OE), or supervised objectives, though these approaches can introduce additional challenges. In this work, we propose the Focused In-distribution Representation Modeling (FIRM) loss, a novel multi-positive contrastive objective for anomaly detection. FIRM addresses class-collision by explicitly encouraging ID representations to be compact while promoting separation among synthetic outliers. We show that FIRM surpasses other contrastive methods in standard benchmarks, significantly enhancing anomaly detection compared to both traditional and supervised contrastive learning objectives. Our ablation studies confirm that FIRM consistently improves the quality of representations and shows robustness across a range of scoring methods. It performs particularly well in ensemble settings and benefits substantially from using OE. The code is available at \\url{https://anonymous.4open.science/r/firm-8472/}.", "keywords": [ "representation learning", "self-supervised learning", "anomaly detection", "out-of-distribution detection", "outlier detection" ], "rating_list": [ 5, 1, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 2, 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 5, 4 ] }, { "paper_id": "jQiJRxNymY", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Andrei Chertkov", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "Researcher" }, { "name": "Artem Basharin", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Ivan Oseledets", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "Gleb Vladimirovich Ryzhakov", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Principal Researcher" } ], "n_formula": 22, "n_formula_1": 20, "n_ref_uni": 24, "n_ref": 51, "n_ref_all": 75, "n_fig": 6, "n_tab": 8, "L_tab": 2510, "n_element_tab": 262, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1396, "n_element_tab_1": 180, "formula_len_all": 1764, "formula_len_all_1": 1669, "len_all": 115132, "len_all_1": 68893, "len_abs": 614, "len_title": 127, "len_sents": 34818, "len_sents_1": 27578, "n_sents": 260, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 613, "title": "Black-Box Approximation and Optimization with Hierarchical Tucker Decomposition", "abs": "We develop a new method HTBB for the multidimensional black-box approximation and gradient-free optimization, which is based on the low-rank hierarchical Tucker decomposition with the use of the MaxVol indices selection procedure. Numerical experiments for 14 complex model problems demonstrate the robustness of the proposed method for dimensions up to 1000, while it shows significantly more accurate results than classical gradient-free optimization methods, as well as approximation and optimization methods based on the popular tensor train decomposition, which represents a simpler case of a tensor network.", "keywords": [ "Black-box optimization", "black-box approximation", "gradient-free method", "low rank representation", "hierarchical Tucker decomposition" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "jQR6ftuL2a", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Jingang Wang", "gender": "Male", "institution": "Meituan", "country": "", "position": "Researcher" }, { "name": "Jizhi Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Xianyuan Zhan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiao Hu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Xin Chen", "gender": "Male", "institution": "Tongji University", "country": "", "position": "PhD student" }, { "name": "Xunliang Cai", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 67, "n_ref_all": 78, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 304, "formula_len_all_1": 109, "len_all": 130198, "len_all_1": 56906, "len_abs": 1219, "len_title": 130, "len_sents": 38927, "len_sents_1": 26651, "n_sents": 290, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1232, "title": "Co-evolved Self-Critique: Enhancing Large Language Models with Self-Generated Data", "abs": "Large language models (LLMs) have seen staggering progress in recent years. Contemporary LLMs rely on an immense amount of data for training, however, as LLMs continue to advance, the availability of high-quality external data is reaching a bottleneck, highlighting the need for model-generated data for further improvement. Although promising, directly utilizing the self-generated data for model training without scrutinized assessment or filtering can easily lead to deteriorated performance, or in other words, \"garbage in, garbage out\". In this study, our insight is to carefully craft a \\textit{self-critique} process, by equipping the LLMs with the ability to be self-aware and discriminative to the quality of its generated data. We introduce a co-evolved self-critique framework that enables an LLM to simultaneously enhance both its generative and evaluative capabilities through an iterative training process. This provides a scalable solution to ensure high-quality self-generated data and facilitate sustained model improvement. Fine-tuning Llama-3 models using this framework results in encouraging improvements in both instruction-following and discriminative abilities, demonstrating the effectiveness of our method.", "keywords": [ "large language models", "self-critique", "co-evolution" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "jQP5o1VAVc", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Chen Sun", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Assistant Professor" }, { "name": "Deqing Sun", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Kaiming He", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Lijie Fan", "gender": "unknown", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Michael Rubinstein", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Principal Researcher" }, { "name": "Siyang Qin", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Tianhong Li", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yonglong Tian", "gender": "unknown", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Yuanzhen Li", "gender": "Female", "institution": "Google", "country": "US", "position": "Software Engineer" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 44, "n_ref": 97, "n_ref_all": 127, "n_fig": 17, "n_tab": 3, "L_tab": 1625, "n_element_tab": 270, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 124, "formula_len_all_1": 56, "len_all": 163383, "len_all_1": 57366, "len_abs": 1315, "len_title": 126, "len_sents": 41826, "len_sents_1": 26226, "n_sents": 353, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1280, "title": "Scaling Autoregressive Text-to-image Generative Models with Continuous Tokens", "abs": "Scaling up autoregressive models in vision has not proven as beneficial as in large language models.\nIn this work, we investigate this scaling problem in the context of text-to-image generation, focusing on two critical factors: whether models use discrete or continuous tokens, and whether tokens are generated in a random or fixed raster order using BERT- or GPT-like transformer architectures. Our empirical results show that, while all models scale effectively in terms of validation loss, their evaluation performance -- measured by FID, GenEval score, and visual quality -- follows different trends. Models based on continuous tokens achieves significantly better visual quality than those using discrete tokens. Furthermore, the generation order and attention mechanisms significantly affect the GenEval score: random-order models achieve notably better GenEval scores compared to raster-order models.\nInspired by these findings, we train Fluid, a random-order autoregressive model on continuous tokens. Fluid 10.5B model achieves a new state-of-the-art zeor-shot FID of 6.16 on MS-COCO 30K, and 0.69 overall score on the GenEval benchmark. We hope our findings and results will encourage future efforts to further bridge the scaling gap between vision and language models.", "keywords": [ "generative model", "scaling", "text-to-image", "autoregressive" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "jQ5T1Pbnx7", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Feida Zhu", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Associate Professor" }, { "name": "Hezhe Qiao", "gender": "Male", "institution": "Singapore Management University", "country": "", "position": "PhD student" }, { "name": "Jiashu Pu", "gender": "Male", "institution": "Garena", "country": "SG", "position": "Researcher" }, { "name": "Qian Shao", "gender": "Female", "institution": "Singapore management university", "country": "SG", "position": "PhD student" }, { "name": "Ruicheng Liang", "gender": "Male", "institution": "Singapore Management University", "country": "", "position": "PhD student" }, { "name": "ling Cheng", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "PhD student" } ], "n_formula": 35, "n_formula_1": 21, "n_ref_uni": 48, "n_ref": 74, "n_ref_all": 99, "n_fig": 15, "n_tab": 20, "L_tab": 9527, "n_element_tab": 1012, "n_fig_1": 9, "n_tab_1": 6, "L_tab_1": 2673, "n_element_tab_1": 142, "formula_len_all": 2618, "formula_len_all_1": 1769, "len_all": 199757, "len_all_1": 70271, "len_abs": 1622, "len_title": 149, "len_sents": 58312, "len_sents_1": 28482, "n_sents": 473, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1627, "title": "New Recipe for Semi-supervised Community Detection: Clique Annealing under Crystallization Kinetics", "abs": "Semi-supervised community detection methods are widely used for identifying specific communities due to the label scarcity. Existing semi-supervised community detection methods typically involve two learning stages \\ie, learning in both initial identification and subsequent adjustment, which often starts from an unreasonable community core candidate.\nMoreover, these methods encounter scalability issues because they depend on reinforcement learning and generative adversarial networks, leading to higher computational costs and restricting the selection of candidates. \nTo address these limitations, we draw a parallel between crystallization kinetics and community detection to integrate the spontaneity of the annealing process into community detection.\nSpecifically, we liken community detection to identifying a crystal subgrain (core) that expands into a complete grain (community) through a process similar to annealing. Based on this finding, we propose CLique ANNealing (CLANN), which applies kinetics concepts to community detection by integrating these principles into the optimization process to strengthen the consistency of the community core. Subsequently, a learning-free Transitive Annealer was employed to refine the first-stage candidates by merging neighboring cliques and repositioning the community core, enabling a spontaneous growth process that enhances scalability.\nExtensive experiments on diverse community detection datasets demonstrate that CLANN outperforms state-of-the-art methods across multiple real-world datasets, showcasing its exceptional efficacy and efficiency in community detection.", "keywords": [ "Semi-supervised Community Detection", "Clique", "Annealing", "Crystallization Kinetics" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "jPrKs5rOWw", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Fang Wu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Stan Z. Li", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Chair Professor" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 38, "n_ref": 63, "n_ref_all": 76, "n_fig": 7, "n_tab": 9, "L_tab": 4695, "n_element_tab": 651, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1862, "n_element_tab_1": 268, "formula_len_all": 2487, "formula_len_all_1": 1041, "len_all": 162658, "len_all_1": 68802, "len_abs": 1332, "len_title": 82, "len_sents": 44056, "len_sents_1": 28618, "n_sents": 362, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1339, "title": "Dynamics-inspired Structure Hallucination for Protein-protein Interaction Modeling", "abs": "Protein-protein interaction (PPI) represents a central challenge within the biology field, and accurately predicting the consequences of mutations in this context is crucial for drug design and protein engineering. Deep learning (DL) has shown promise in forecasting the effects of such mutations but is hindered by two primary constraints. First, the structures of mutant proteins are often elusive to acquire. Secondly, PPI takes place dynamically, which is rarely integrated into the DL architecture design. To address these obstacles, we present a novel framework named Refine-PPI with two key enhancements. First, we introduce a structure refinement module trained by a mask mutation modeling (MMM) task on available wild-type structures, which is then transferred to hallucinate the inaccessible mutant structures. Second, we employ a new kind of geometric network, called the probability density cloud network (PDC-Net), to capture 3D dynamic variations and encode the atomic uncertainty associated with PPI. Comprehensive experiments on SKEMPI.v2 substantiate the superiority of Refine-PPI over all existing tools for predicting free energy change. These findings underscore the effectiveness of our hallucination strategy and the PDC module in addressing the absence of mutant protein structure and modeling geometric uncertainty.", "keywords": [ "Protein-protein Interactions", "Geometric Deep Learning", "Mutation Effect Prediction" ], "rating_list": [ 3, 5, 1, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "jPlghr8io4", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Hailong Zhu", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "MS student" }, { "name": "Jun He", "gender": "Male", "institution": "iData", "country": "CN", "position": "Researcher" }, { "name": "Richang Hong", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xueliang Liu", "gender": "Male", "institution": "Hefei University of Technology", "country": "", "position": "Full Professor" }, { "name": "Xun Yang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 41, "n_ref": 92, "n_ref_all": 119, "n_fig": 2, "n_tab": 9, "L_tab": 2702, "n_element_tab": 202, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 4092, "n_element_tab_1": 161, "formula_len_all": 827, "formula_len_all_1": 799, "len_all": 149809, "len_all_1": 68509, "len_abs": 1170, "len_title": 112, "len_sents": 40139, "len_sents_1": 29328, "n_sents": 303, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1282, "title": "Learning with Analogical Reasoning for Robust Few-Shot Learning", "abs": "Few-shot learning (FSL) is challenging due to limited support data for model training. The situation is much worse when the support data is contaminated with noise. To address this issue, we propose a novel \\textbf{T}ransformer-based \\textbf{A}nalogical \\textbf{R}easoning model for \\textbf{N}oisy \\textbf{F}ew-\\textbf{S}hot learning (TarNFS), by mimicing the human's ability of learning by analogy. Concretely, we assume the existence of a large human cultivated or AI-powered knowledge base, and hypothesize that similar concepts in the knowledge base are visually similar in the latent space as well. Then we design a transformer-based analogical reasoning model to utilize inter-concept connections among these concepts, aiming to build robust and discriminative classification boundaries. In addition, we propose a task-level contrastive learning to analogically learn from negative tasks to facilitate training with noisy tasks. Experiments demonstrate that our TarNFS enables more effective learning from limited and imperfect data. It not only improves the generalization ability of FSL in different noisy settings but also achieves competitive performance in the common clean FSL settings. Code is publicly available \\href{https://anonymous.4open.science/r/iclr2088}{here}.", "keywords": [ "few-shot learning", "analogical reasoning", "noisy label learning", "contrastive learning", "transformer" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 5, 3 ] }, { "paper_id": "jOyQXG6CM4", "primary_area": "datasets and benchmarks", "n_author": 15, "author_info_list": [ { "name": "Bin Wu", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Chuangxin Chu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "MS student" }, { "name": "Haotian Huang", "gender": "Male", "institution": "North China University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Huajun Chen", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Jingyu Lu", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "MS student" }, { "name": "Kai Ma", "gender": "Male", "institution": "Duke University", "country": "US", "position": "MS student" }, { "name": "Keyan Ding", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Mei Li", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Qiang Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tianhao Li", "gender": "Male", "institution": "Duke University", "country": "US", "position": "MS student" }, { "name": "Tianyu Zeng", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "MS student" }, { "name": "Xingkai Wang", "gender": "unknown", "institution": "China Mobile Communications Company Limited", "country": "CN", "position": "Researcher" }, { "name": "Xuejing Yuan", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "Lecturer" }, { "name": "Yujia Zheng", "gender": "unknown", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "Intern" }, { "name": "Zuoxian Liu", "gender": "unknown", "institution": "University of Glasgow", "country": "GB", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 75, "n_ref_all": 108, "n_fig": 5, "n_tab": 9, "L_tab": 7486, "n_element_tab": 324, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 3118, "n_element_tab_1": 295, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 202696, "len_all_1": 66103, "len_abs": 1398, "len_title": 152, "len_sents": 53135, "len_sents_1": 29009, "n_sents": 357, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 104, "L_abs": 1405, "title": "SciSafeEval: A Comprehensive Benchmark for Safety Alignment of Large Language Models in Scientific Tasks", "abs": "Large language models (LLMs) have had a transformative impact on a variety of scientific tasks across disciplines such as biology, chemistry, medicine, and physics. However, ensuring the safety alignment of these models in scientific research remains an underexplored area, with existing benchmarks primarily focus on textual content and overlooking key scientific representations such as molecular, protein, and genomic languages. Moreover, the safety mechanisms of LLMs in scientific tasks are insufficiently studied. To address these limitations, we introduce SciSafeEval, a comprehensive benchmark designed to evaluate the safety alignment of LLMs across a range of scientific tasks. SciSafeEval spans multiple scientific languages\u2014including textual, molecular, protein, and genomic\u2014and covers a wide range of scientific domains. We evaluate LLMs in zero-shot, few-shot and chain-of-thought settings, and introduce a ''jailbreak'' enhancement feature that challenges LLMs equipped with safety guardrails, rigorously testing their defenses against malicious intention. Our benchmark surpasses existing safety datasets in both scale and scope, providing a robust platform for assessing the safety and performance of LLMs in scientific contexts. This work aims to facilitate the responsible development and deployment of LLMs, promoting alignment with safety and ethical standards in scientific research.", "keywords": [ "Large Language Models", "Scientific Tasks", "Safety Alignment", "Benchmark" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "jOuHjFw71C", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Atharva Gundawar", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "MS student" }, { "name": "Karthik Valmeekam", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Kaya Stechly", "gender": "unknown", "institution": "Arizona State University", "country": "US", "position": "MS student" }, { "name": "Subbarao Kambhampati", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 27, "n_ref": 49, "n_ref_all": 66, "n_fig": 2, "n_tab": 8, "L_tab": 10163, "n_element_tab": 455, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 998, "n_element_tab_1": 141, "formula_len_all": 6430, "formula_len_all_1": 0, "len_all": 279191, "len_all_1": 60894, "len_abs": 2316, "len_title": 106, "len_sents": 74002, "len_sents_1": 30275, "n_sents": 512, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 106, "L_abs": 1211, "title": "Planning in Strawberry Fields: Evaluating and Improving the Planning and Scheduling Capabilities of LRM o1", "abs": "The ability to plan a course of action that achieves a desired state of affairs has long been considered a core competence of intelligent agents and has been an integral part of AI research since its inception. With the advent of large language models (LLMs), there has been considerable interest in the question of whether or not they possess such planning abilities, but--despite the slew of new private and open source LLMs since GPT3--progress has remained slow. OpenAI claims that their recent o1 (Strawberry) model has been specifically constructed and trained to escape the normal limitations of autoregressive LLMs--making it a new kind of model: a Large Reasoning Model (LRM). In this paper, we evaluate the planning capabilities of two LRMs (o1-preview and o1-mini) on both planning and scheduling benchmarks. We see that while o1 does seem to offer significant improvements over autoregressive LLMs, this comes at a steep inference cost, while still failing to provide any guarantees over what it generates. We also show that combining o1 models with external verifiers--in a so-called LRM-Modulo system--guarantees the correctness of the combined system's output while further improving performance.", "keywords": [ "Large Language Models", "Large Reasoning Models", "Planning", "Reasoning" ], "rating_list": [ 3, 3, 5, 1 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "jOmk0uS1hl", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Florian E. Dorner", "gender": "Male", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "", "position": "Researcher" }, { "name": "Moritz Hardt", "gender": "Not Specified", "institution": "Max-Planck-Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "Principal Researcher" }, { "name": "Ricardo Dominguez-Olmedo", "gender": "Male", "institution": "Max-Planck-Institute for Intelligent Systems, Max-Planck Institute", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 1, "n_ref_uni": 56, "n_ref": 89, "n_ref_all": 129, "n_fig": 16, "n_tab": 4, "L_tab": 3243, "n_element_tab": 333, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 487, "formula_len_all_1": 42, "len_all": 221403, "len_all_1": 59464, "len_abs": 1134, "len_title": 108, "len_sents": 57115, "len_sents_1": 28813, "n_sents": 463, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 60, "L_abs": 1144, "title": "Training on the Test Task Confounds Evaluation and Emergence", "abs": "We study a fundamental problem in the evaluation of large language models that we call training on the test task. Unlike wrongful practices like training on the test data, leakage, or data contamination, training on the test task is not a malpractice. Rather, the term describes a growing set of techniques to include task-relevant data in the pretraining stage of a language model. We demonstrate that training on the test task confounds both relative model evaluations and claims about emergent capabilities. We argue that the seeming superiority of one model family over another may be explained by a different degree of training on the test task. To this end, we propose an effective method to adjust for the effect of training on the test task on benchmark evaluations. Put simply, to fine-tune each model under comparison on the same task-relevant data before evaluation. Lastly, we show that instances of emergent behavior disappear gradually as models train on the test task. Our work promotes a new perspective on the evaluation of large language models with broad implications for benchmarking and the study of emergent capabilities.", "keywords": [ "language models", "benchmarking", "emergence" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 4, 3, 4 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "jOVfFAxBf6", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Vitaliy Kurlin", "gender": "Not Specified", "institution": "University of Liverpool", "country": "GB", "position": "Full Professor" }, { "name": "Yury Elkin", "gender": "unknown", "institution": "University of Liverpool", "country": "", "position": "Postdoc" } ], "n_formula": 44, "n_formula_1": 6, "n_ref_uni": 30, "n_ref": 46, "n_ref_all": 67, "n_fig": 9, "n_tab": 7, "L_tab": 3433, "n_element_tab": 95, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 2579, "n_element_tab_1": 51, "formula_len_all": 5092, "formula_len_all_1": 620, "len_all": 286814, "len_all_1": 56072, "len_abs": 1691, "len_title": 134, "len_sents": 92251, "len_sents_1": 24203, "n_sents": 968, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 831, "title": "Bi-continuous and complete SE(2)-invariants parametrize all clouds of unordered points", "abs": "The most basic form of a rigid object is a cloud of unordered points, for example, a set of corners or other salient features. The rigid shape of a point cloud in the Euclidean plane is its SE(2)-equivalence class under rigid motion (a composition of translations and rotations). We introduce complete invariants (with no false negatives, no false positives) and a bi-Lipschitz continuous metric that satisfies all axioms, provides a 1-1 matching between points in clouds, and is computable in a quadratic time of the number $m$ of points. The realizability property implies that the space of all rigid clouds is efficiently parametrized by vectorial invariants like geographic coordinates. The new invariants justified that any of 130K+ molecules in the QM9 database is uniquely determined by the rigid shape of its atomic cloud.", "keywords": [ "point cloud", "rigid motion", "isometry", "complete invariant", "continuous metric", "molecule" ], "rating_list": [ 3, 6, 6, 8 ], "soundness_list": [ 3, 4, 2, 4 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "jOVJhKzc3Y", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Fangyun Wei", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" }, { "name": "Jianbo Ma", "gender": "unknown", "institution": "University of New South Wales", "country": "AU", "position": "Researcher" }, { "name": "Jianyuan Guo", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Chang Xu", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 101, "n_ref_all": 124, "n_fig": 7, "n_tab": 1, "L_tab": 295, "n_element_tab": 18, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 295, "n_element_tab_1": 18, "formula_len_all": 830, "formula_len_all_1": 628, "len_all": 133666, "len_all_1": 65463, "len_abs": 1906, "len_title": 150, "len_sents": 37205, "len_sents_1": 30077, "n_sents": 264, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1916, "title": "The Devil is in the Word: Video-Conditioned Text Representation Refinement for Text-to-Video Retrieval", "abs": "Pre-trained vision-language models (VLMs), such as CLIP, have shown remarkable success in the text-video retrieval task due to their strong vision-language\nrepresentations learned from large-scale paired image-text samples. However,\ncompared to videos, text is often short and concise, making it difficult to fully\ncapture the rich and redundant semantics present in a video with thousands of\nframes. Recent advances have focused on utilizing text features to extract key information from these redundant video frames. However, text representation generated without considering video information can suffer from bias and lack the\nexpressiveness needed to capture key words that could enhance retrieval performance. In this study, we first conduct preliminary experiments to demonstrate\nthe importance of enhancing text representations. These experiments reveal that\ntext representation only generated from text input often misinterpret critical information. To address this, we propose a simple yet efficient method, VICTER, i.e.,\nvideo-conditioned text representation refinement, to enrich text representation using a versatile module. Specifically, we introduce a video abstraction module that\nextracts representative features from multiple video frames. This is followed by\na video-conditioned text enhancement module that refines the original text features by reassessing individual word features and extracting key words using the\ngenerated video features. Empirical evidence shows that VICTER not only effectively captures relevant key words from the input text but also complements\nvarious existing frameworks. Our experimental results demonstrate a significant\nimprovement of VICTER over several baseline frameworks (with 0.4% \u223c 1.0%\nimprovements on R@1). Furthermore, VICTER achieves state-of-the-art performance on three benchmark datasets, including MSRVTT, DiDeMo, and LSMDC. Code will be made available.", "keywords": [ "Text-to-Video Retrieval", "Video-conditioned Text Representation Enhancement" ], "rating_list": [ 5, 3, 6, 3, 5 ], "soundness_list": [ 3, 3, 3, 1, 2 ], "presentation_list": [ 2, 3, 2, 3, 2 ], "contribution_list": [ 2, 1, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 5, 2 ] }, { "paper_id": "jO3QEsm15T", "primary_area": "causal reasoning", "n_author": 7, "author_info_list": [ { "name": "Hao Zhou", "gender": "Male", "institution": "Guangdong University of Technology", "country": "CN", "position": "MS student" }, { "name": "Haolin Yang", "gender": "Male", "institution": "Guangdong University of Technology", "country": "CN", "position": "MS student" }, { "name": "Ruichu Cai", "gender": "Male", "institution": "Guangdong University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yuguang Yan", "gender": "Male", "institution": "Guangdong University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Zeqin Yang", "gender": "Male", "institution": "Guangdong University of Technology", "country": "CN", "position": "MS student" }, { "name": "Zhifeng Hao", "gender": "Male", "institution": "Shantou University", "country": "", "position": "Full Professor" }, { "name": "ZongyuLi", "gender": "Male", "institution": "Guangdong University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 56, "n_formula_1": 27, "n_ref_uni": 52, "n_ref": 104, "n_ref_all": 127, "n_fig": 4, "n_tab": 9, "L_tab": 2479, "n_element_tab": 103, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 308, "n_element_tab_1": 20, "formula_len_all": 3937, "formula_len_all_1": 1358, "len_all": 143943, "len_all_1": 62122, "len_abs": 1650, "len_title": 126, "len_sents": 42024, "len_sents_1": 30085, "n_sents": 289, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1655, "title": "Optimal Transport for Reducing Bias in Causal Inference without Data Splitting", "abs": "Causal inference seeks to estimate the causal effect given a treatment such as a kind of medicine or the dosage of a medication. To address the issue of confounding bias caused by the non-randomized treatment assignment on samples, most existing methods reduce the covariate shift between subpopulations receiving different values of treatment. However, these methods split training samples into smaller groups, which cuts down the number of samples in each group, while precise distribution estimation and alignment highly rely on a sufficient number of training data. In this paper, we propose a distribution alignment paradigm that involves all the training samples without data splitting, which can be naturally applied in the settings of binary and continuous treatments. To this end, we characterize the distribution shift by considering different probability measures of the same set including all the training samples, and reduce the shift between the marginal covariate distribution and the conditional covariate distribution given a treatment value. By doing this, data reduction caused by splitting is avoided, and the outcome prediction model trained on samples receiving one treatment value can be generalized to the entire population. In specific, we exploit the optimal transport theory built on probability measures to analyze the confounding bias and the outcome estimation error, which motivates us to propose a balanced representation learning method for causal inference of binary and continuous treatments. The experimental results on both binary and continuous treatment settings demonstrate the effectiveness of the proposed method.", "keywords": [ "Causal Effect Estimation", "Optimal Transport" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "jNmsuEE4Gf", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Aarti Singh", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Joshua Sunshine", "gender": "unknown", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Kyle Lee", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Undergrad student" }, { "name": "Sam Estep", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Wode Ni", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Yifu Cai", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Undergrad student" }, { "name": "Yuejie Chi", "gender": "Female", "institution": "Meta AI", "country": "US", "position": "Researcher" }, { "name": "Zixin Wen", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 46, "n_ref": 83, "n_ref_all": 103, "n_fig": 7, "n_tab": 6, "L_tab": 3525, "n_element_tab": 194, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 210, "n_element_tab_1": 26, "formula_len_all": 659, "formula_len_all_1": 302, "len_all": 233183, "len_all_1": 53458, "len_abs": 1380, "len_title": 126, "len_sents": 58329, "len_sents_1": 23630, "n_sents": 504, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1222, "title": "Feynman: Knowledge-Infused Diagramming Agent for Scaling Visual Reasoning Data", "abs": "Visual reasoning is an essential ability of state-of-the-art multi-modal AI systems. Improving these systems requires high-quality vision-language data at scale. Despite the abundance of internet image and text data, knowledge-rich and well-aligned image-text pairs are rare. In this paper, we present a scalable data generation pipeline built with our diagramming agent, **Feynman**. To create diagrams, Feynman first enumerates domain-specific knowledge components (\"ideas\") and performs code planning based on the ideas. Given the plan, Feynman translates ideas into simple declarative programs and iterates to receives feedback and visually refine diagrams. Finally, the declarative programs are rendered by the Penrose diagramming system. The optimization-based rendering of Penrose preserves the visual semantics while injecting fresh randomness into the layout, thereby producing diagrams with visual consistency and diversity. As a result, Feynman can author diagrams along with grounded captions with very little cost and time. Using Feynman, we synthesized a dataset with more than 100$k$ well-aligned diagram-caption pairs. We also curate a visual-language benchmark, **Diagramma**, from freshly generated data.", "keywords": [ "Vision-Language Dataset", "Synthetic Data", "Visual Reasoning Benchmark" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "jNCwczhHLP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Daochen Zha", "gender": "unknown", "institution": "Airbnb", "country": "US", "position": "Researcher" }, { "name": "Hao Chen", "gender": "Male", "institution": "City University of Macau", "country": "MO", "position": "Assistant Professor" }, { "name": "Huachi Zhou", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Qinggang Zhang", "gender": "Male", "institution": "The Hong Kong Polytechnic University, Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Xiao Huang", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "", "position": "Assistant Professor" }, { "name": "Pei Wenqi", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Undergrad student" } ], "n_formula": 15, "n_formula_1": 8, "n_ref_uni": 30, "n_ref": 69, "n_ref_all": 87, "n_fig": 8, "n_tab": 7, "L_tab": 2686, "n_element_tab": 186, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1758, "n_element_tab_1": 117, "formula_len_all": 1284, "formula_len_all_1": 465, "len_all": 172908, "len_all_1": 68590, "len_abs": 1321, "len_title": 119, "len_sents": 45612, "len_sents_1": 32337, "n_sents": 365, "n_sents_1": 242, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1331, "title": "Self-Monitoring Large Language Models for Click-Through Rate Prediction", "abs": "Click-through rate (CTR) prediction tasks traditionally aim to model extensive user-\nitem feature interactions. Recent approaches fine-tune Large Language Models\n(LLMs) using user-item features as input and click labels as output. However,\ndue to the sparsity of click labels, the attention mechanism may focus on a subset\nof features rather than all features. This can hinder LLMs\u2019 ability to accurately\nmatch features to click labels, resulting in performance that does not consistently\nexceed traditional state-of-the-art CTR approaches. To address this, we introduce\na SLLM4CTR framework which uses adaptive temperature and label matching loss\nto improve fine-tuning and inference process of LLMs. The adaptive temperature\nserves as a confidence score to calibrate CTR predictions by quantifying the LLMs\u2019\nattention to user-item features. The label matching loss clearly distinguish between\nclick-inducing and non-click-inducing features by constraining the representation\nspace of click labels. By combining these two designs, SLLM4CTR improves feature\nutilization in LLMs and enhances the matching of user-item features to click\nlabels. Experimental results demonstrate\nthat SLLM4CTR significantly outperforms state-of-the-art baselines, including both\ntraditional and LLM-based CTR approaches. The code will be open-sourced.", "keywords": [ "Large Language Model", "Click-through Rate Prediction", "Feature-Click learning" ], "rating_list": [ 6, 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 1, 2 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "jMtsvoOU3G", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Jian Bi", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jian Yang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Jianjun Qian", "gender": "Male", "institution": "Nanjing University of Science and Techonology", "country": "CN", "position": "Full Professor" }, { "name": "Qianliang Wu", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Shuo Chen", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Research Scientist" }, { "name": "Xiang Li", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "lei luo", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 27, "n_formula_1": 9, "n_ref_uni": 29, "n_ref": 69, "n_ref_all": 96, "n_fig": 7, "n_tab": 13, "L_tab": 4197, "n_element_tab": 605, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3481, "n_element_tab_1": 520, "formula_len_all": 2451, "formula_len_all_1": 657, "len_all": 160080, "len_all_1": 58630, "len_abs": 1336, "len_title": 131, "len_sents": 36710, "len_sents_1": 22492, "n_sents": 293, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1342, "title": "SinPoint: A Novel Topological Consistent Augmentation for Point Cloud Understanding", "abs": "Data augmentation is a highly effective method for addressing the issue of data scarcity in machine learning and computer vision tasks. It involves diversifying the original data through a series of transformations to improve the robustness and generalization ability of the model. However, due to the disorder and irregularity of point clouds, existing methods struggle to enrich geometric diversity and maintain topological consistency, leading to imprecise point cloud understanding. In this paper, we propose SinPoint, a novel method designed to preserve the topological structure of the original point cloud through a homeomorphism. Additionally, it utilizes the Sine function to generate smooth displacements. This simulates object deformations, thereby producing a rich diversity of samples. Our extensive experiments demonstrate that SinPoint consistently outperforms existing Mixup and Deformation methods on various benchmark point cloud datasets, improving performance for shape classification and part segmentation tasks. Specifically, when used with PointNet++ and DGCNN, SinPoint achieves a state-of-the-art accuracy of 90.2 on shape classification with the real-world ScanObjectNN dataset. Furthermore, our method is highly versatile and scalable, and it can adapt to different scenarios and requirements for point cloud tasks.", "keywords": [ "Topological consistent", "homeomorphism", "Point clouds augmentation" ], "rating_list": [ 8, 6, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 5, 5 ] }, { "paper_id": "jMffFIWHic", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "He Sun", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Siyi Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "WeiminBai", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Wenzheng Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 27, "n_formula_1": 13, "n_ref_uni": 55, "n_ref": 164, "n_ref_all": 195, "n_fig": 15, "n_tab": 7, "L_tab": 1615, "n_element_tab": 53, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 847, "n_element_tab_1": 2, "formula_len_all": 2863, "formula_len_all_1": 1641, "len_all": 209052, "len_all_1": 64977, "len_abs": 1489, "len_title": 93, "len_sents": 56367, "len_sents_1": 27199, "n_sents": 432, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1486, "title": "Blind Inversion using Latent Diffusion Priors", "abs": "Diffusion models have emerged as powerful tools for solving inverse problems due to their exceptional ability to model complex prior distributions. However, existing methods predominantly assume known forward operators (i.e., non-blind), limiting their applicability in practical settings where acquiring such operators is costly. Additionally, many current approaches rely on pixel-space diffusion models, leaving the potential of more powerful latent diffusion models (LDMs) underexplored. In this paper, we introduce LatentDEM, an innovative technique that addresses more challenging blind inverse problems using latent diffusion priors. At the core of our method is solving blind inverse problems within an iterative Expectation-Maximization (EM) framework: (1) the E-step recovers clean images from corrupted observations using LDM priors and a known forward model, and (2) the M-step estimates the forward operator based on the recovered images. Additionally, we propose two novel optimization techniques tailored for LDM priors and EM frameworks, yielding more accurate and efficient blind inversion results. As a general framework, LatentDEM supports both linear and non-linear inverse problems. Beyond common 2D image restoration tasks, it enables new capabilities in non-linear 3D inverse rendering problems. We validate LatentDEM's performance on representative 2D blind deblurring and 3D sparse-view reconstruction tasks, demonstrating its superior efficacy over prior arts.", "keywords": [ "Computational Imaging", "Diffusion Model", "Inverse Problem" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 4, 1, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "jMZjIi9JcC", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "Han Hu", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Han Wu", "gender": "Female", "institution": "University of Sydney", "country": "", "position": "PhD student" }, { "name": "Jianyuan Guo", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Kai Han", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Xinghao Chen", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Yehui Tang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yunhe Wang", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhiwei Hao", "gender": "unknown", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Chang Xu", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 46, "n_ref": 214, "n_ref_all": 253, "n_fig": 9, "n_tab": 17, "L_tab": 7379, "n_element_tab": 907, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 2809, "n_element_tab_1": 276, "formula_len_all": 39, "formula_len_all_1": 0, "len_all": 188189, "len_all_1": 56856, "len_abs": 1491, "len_title": 125, "len_sents": 46931, "len_sents_1": 23986, "n_sents": 373, "n_sents_1": 151, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1524, "title": "Empirical Study on Enhancing Efficiency in Masked Image Modeling Pre-training", "abs": "The combination of transformers and masked image modeling (MIM) pre-training framework has shown remarkable potential in various vision tasks. However, the high computational cost of pre-training hinders the practical application of MIM.\n This paper introduces \\emph{FastMIM}, a simple and versatile framework that expedites masked image modeling through two steps: (i) pre-training vision backbones using low-resolution input images and (ii) reconstructing Histograms of Oriented Gradients (HOG) feature instead of original RGB values of the input images.\n Furthermore, we propose \\emph{FastMIM-P}, which progressively increases the input resolution during the pre-training stage to improve the transfer learning performance of models with high capacity. We point out that: (i) a wide range of input resolutions during pre-training can result in similar performances in fine-tuning and downstream tasks such as detection and segmentation; (ii) the shallow layers of encoder are more important during pre-training, and discarding the last few layers can speed up the training process without affecting fine-tuning performance; and (iii) HOG is more stable than RGB values when transferring resolution. Equipped with \\emph{FastMIM}, any type of vision backbone can be efficiently pre-trained. For example, using ViT-B/Swin-B as backbones, we achieve 83.8\\%/84.1\\% top-1 accuracy on ImageNet-1K. Compared to previous approaches, our method can achieve better top-1 accuracy while accelerating the training procedure by 5\u00d7.", "keywords": [ "Self-supervised learning", "Empirical study" ], "rating_list": [ 5, 3, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4, 3 ], "contribution_list": [ 2, 1, 3, 3, 3 ], "confidence_list": [ 4, 5, 5, 5, 4 ] }, { "paper_id": "jMZglnlwf7", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Beren Millidge", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "Postdoc" }, { "name": "Emily Shepperd", "gender": "Female", "institution": "EleutherAI", "country": "AU", "position": "Researcher" }, { "name": "Jonathan Pilault", "gender": "unknown", "institution": "Polytechnique Montreal", "country": "", "position": "PhD student" }, { "name": "Quentin Gregory Anthony", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "", "position": "PhD student" }, { "name": "Vasudev Shyam", "gender": "unknown", "institution": "Zyphra Technologies", "country": "US", "position": "Researcher" } ], "n_formula": 46, "n_formula_1": 21, "n_ref_uni": 51, "n_ref": 72, "n_ref_all": 87, "n_fig": 4, "n_tab": 3, "L_tab": 777, "n_element_tab": 85, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 44, "n_element_tab_1": 2, "formula_len_all": 2574, "formula_len_all_1": 1275, "len_all": 183153, "len_all_1": 56153, "len_abs": 1169, "len_title": 130, "len_sents": 41668, "len_sents_1": 26718, "n_sents": 298, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1174, "title": "Tree Attention: Topology-Aware Decoding for Long-Context Attention", "abs": "Self-attention is the core mathematical operation of modern transformer architectures and is also a significant computational bottleneck due to its quadratic complexity in the sequence length. \nIn this work, we derive the scalar energy function whose gradient computes the self-attention block, thus elucidating the theoretical underpinnings of self-attention. \nOur formulation reveals that the reduction across the sequence axis can be efficiently computed in parallel through a tree reduction. \nOur algorithm, called Tree Attention, for parallelizing exact attention computation across multiple GPUs enables cross-device decoding to be performed symptotically faster (up to 8x faster in our experiments) than state-of-the-art approaches such as Ring Attention, while also requiring significantly less communication volume and incurring 2x less peak memory. \nWe demonstrate that Tree Attention speeds up decoding up to 4x on Llama 3.1-8B and can be applied to a variety of hardware and networking setups such as H100 DGX nodes, AMD MI300x nodes, and PCIe connected NVIDIA RTX 4090s.\nOur code is publicly available here: https://anonymous.4open.science/r/tree_attention-7C32", "keywords": [ "attention", "transformer", "long context", "large language models" ], "rating_list": [ 3, 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 3, 2, 3, 3, 3 ] }, { "paper_id": "jMSsgTaVKN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Dimitris Plexousakis", "gender": "Male", "institution": "Institute of Computer Science, FORTH", "country": "GR", "position": "Full Professor" }, { "name": "Filippos Gouidis", "gender": "Male", "institution": "University of Crete", "country": "GR", "position": "PhD student" }, { "name": "Konstantinos Papoutsakis", "gender": "Male", "institution": "HELLENIC MEDITERRANEAN UNIVERSITY", "country": "GR", "position": "Postdoc" }, { "name": "Theodore Patkos", "gender": "Male", "institution": "Foundation for Research and Technology, Hellas", "country": "", "position": "Principal Researcher" }, { "name": "Antonis A Argyros", "gender": "Male", "institution": "University of Crete", "country": "GR", "position": "Professor" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 54, "n_ref": 80, "n_ref_all": 94, "n_fig": 2, "n_tab": 7, "L_tab": 1398, "n_element_tab": 250, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 871, "n_element_tab_1": 123, "formula_len_all": 378, "formula_len_all_1": 342, "len_all": 142029, "len_all_1": 60673, "len_abs": 1034, "len_title": 166, "len_sents": 31824, "len_sents_1": 30443, "n_sents": 218, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 118, "L_abs": 1038, "title": "Leveraging metapaths for learning from knowledge graphs in the context of vision-based classification of object states", "abs": "Zero-Shot Object State Classification (ZS-OSC) aims to recognize unseen object states without any visual training examples. Existing methods typically rely on Knowledge Graphs (KGs) to provide semantic information about states, but they often treat KGs as homogeneous, overlooking the rich relational knowledge encoded in their structure. We propose a novel approach to ZS-OSC that leverages meta-paths to capture complex relationships between object states in a KG. Our method learns to project semantic information from the KG into the visual space via meta-path learning, generating discriminative visual embeddings for unseen state classes. To the best of our knowledge, this is the first work to utilize meta-paths for ZS-OSC. We conduct extensive experiments on four benchmark datasets, demonstrating the superior performance of our approach compared to SoTA zero-shot learning methods and a graph-based baseline. Our ablation study further provides insights into the impact of key design choices on the effectiveness of our method.", "keywords": [ "Object State Classification", "Meta-paths learning", "Zero Shot Learning", "Embeddings Learning" ], "rating_list": [ 5, 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 1, 3, 1 ], "contribution_list": [ 2, 2, 2, 1, 1 ], "confidence_list": [ 4, 4, 3, 5, 4 ] }, { "paper_id": "jLd7OyAD4Y", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Ivaxi Sheth", "gender": "Female", "institution": "CISPA, saarland university, saarland informatics campus", "country": "", "position": "PhD student" }, { "name": "Mario Fritz", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Full Professor" }, { "name": "Ruta Binkyte-Sadauskiene", "gender": "Female", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Postdoc" }, { "name": "Tejumade Afonja", "gender": "Female", "institution": "CISPA, saarland university, saarland informatics campus", "country": "", "position": "PhD student" }, { "name": "Thomas Ulas", "gender": "unknown", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "Principal Researcher" }, { "name": "Waqar Hanif", "gender": "Male", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 48, "n_ref": 82, "n_ref_all": 106, "n_fig": 8, "n_tab": 2, "L_tab": 1876, "n_element_tab": 167, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1636, "n_element_tab_1": 164, "formula_len_all": 243, "formula_len_all_1": 243, "len_all": 173300, "len_all_1": 62305, "len_abs": 1486, "len_title": 157, "len_sents": 51936, "len_sents_1": 28326, "n_sents": 362, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 877, "title": "LLM4GRN: Discovering Causal Gene Regulatory Networks with LLMs - Evaluation through Synthetic Data Generation", "abs": "Gene regulatory networks (GRNs) represent the causal relationships between transcription factors (TFs) and target genes in single-cell RNA sequencing (scRNA-seq) data. Understanding these networks is crucial for uncovering disease mechanisms and identifying therapeutic targets. In this work, we investigate the potential of large language models (LLMs) for GRN discovery, leveraging their learned biological knowledge alone or in combination with traditional statistical methods. We employ a task-based evaluation strategy to address the challenge of unavailable ground truth causal graphs. Specifically, we use the GRNs suggested by LLMs to guide causal synthetic data generation and compare the resulting data against the original dataset. Our statistical and biological assessments show that LLMs can support statistical modeling and data synthesis for biological research.", "keywords": [ "LLM", "GRN", "Causal Discovery", "Synthetic Data Generation" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 3, 1 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "jKcZ4hF4s5", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Atsutoshi Kumagai", "gender": "Male", "institution": "NTT", "country": "JP", "position": "Researcher" }, { "name": "Hiroshi Takahashi", "gender": "Male", "institution": "NTT", "country": "JP", "position": "Researcher" }, { "name": "Tomoharu Iwata", "gender": "Male", "institution": "NTT", "country": "", "position": "Researcher" }, { "name": "Tomoya Yamashita", "gender": "Male", "institution": "NTT", "country": "", "position": "Researcher" }, { "name": "Yuuki Yamanaka", "gender": "Male", "institution": "NTT", "country": "JP", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 26, "n_ref_uni": 38, "n_ref": 77, "n_ref_all": 111, "n_fig": 4, "n_tab": 7, "L_tab": 2193, "n_element_tab": 216, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 383, "n_element_tab_1": 54, "formula_len_all": 1730, "formula_len_all_1": 1622, "len_all": 109102, "len_all_1": 57208, "len_abs": 993, "len_title": 124, "len_sents": 27893, "len_sents_1": 23806, "n_sents": 243, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 996, "title": "Positive-Unlabeled Diffusion Models for Preventing Sensitive Data Generation", "abs": "Diffusion models are powerful generative models but often generate sensitive data that are unwanted by users,\nmainly because the unlabeled training data frequently contain such sensitive data.\nSince labeling all sensitive data in the large-scale unlabeled training data is impractical,\nwe address this problem by using a small amount of labeled sensitive data.\nIn this paper,\nwe propose positive-unlabeled diffusion models,\nwhich prevent the generation of sensitive data using unlabeled and sensitive data.\nOur approach can approximate the evidence lower bound (ELBO) for normal (negative) data using only unlabeled and sensitive (positive) data.\nTherefore, even without labeled normal data,\nwe can maximize the ELBO for normal data and minimize it for labeled sensitive data,\nensuring the generation of only normal data.\nThrough experiments across various datasets and settings,\nwe demonstrated that our approach can prevent the generation of sensitive images without compromising image quality.", "keywords": [ "Diffusion Model", "Positive-Unlabeled Learning" ], "rating_list": [ 6, 5, 5, 8 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "jK5r1HBfym", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Denis Rakitin", "gender": "Male", "institution": "HSE University", "country": "", "position": "Researcher" }, { "name": "Dmitry Vetrov", "gender": "Male", "institution": "Constructor University", "country": "DE", "position": "Full Professor" }, { "name": "Ivan Shchekotov", "gender": "Male", "institution": "Higher School of Economics", "country": "", "position": "MS student" } ], "n_formula": 73, "n_formula_1": 15, "n_ref_uni": 73, "n_ref": 219, "n_ref_all": 307, "n_fig": 15, "n_tab": 5, "L_tab": 1176, "n_element_tab": 164, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 343, "n_element_tab_1": 30, "formula_len_all": 4221, "formula_len_all_1": 1250, "len_all": 260695, "len_all_1": 66833, "len_abs": 1146, "len_title": 144, "len_sents": 75505, "len_sents_1": 29816, "n_sents": 683, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1145, "title": "Regularized Distribution Matching Distillation for One-step Unpaired Image-to-Image Translation", "abs": "Diffusion-based generative models achieve SOTA results in mode coverage and generation quality but suffer from inefficient sampling. Recently introduced diffusion distillation techniques approach this issue by transforming the original multi-step model into a one-step generator with approximately the same output distribution. Among these methods, Distribution Matching Distillation (DMD) offers a suitable framework for training general-form one-step generators, applicable beyond unconditional generation. In this paper, we propose a modification of DMD, called Regularized Distribution Matching Distillation (RDMD), which applies to the unpaired image-to-image (I2I) translation problem. To achieve this, we regularize the generator objective from DMD with the transport cost between its input and output. We validate the method's applicability in theory by establishing its connection with optimal transport. Moreover, we demonstrate its empirical performance in application to several translation tasks, including 2D examples and I2I between different image datasets, where it performs on par or better than multi-step diffusion baselines.", "keywords": [ "diffusion distillation", "distribution matching distillation", "optimal transport", "image-to-image translation" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "jJvJqgPZCD", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Lu Wang", "gender": "Female", "institution": "Northeastern University", "country": "", "position": "Assistant Professor" }, { "name": "Xinnuo Li", "gender": "Female", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Undergrad student" }, { "name": "Yunxiang Zhang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 137, "n_ref_all": 145, "n_fig": 1, "n_tab": 9, "L_tab": 3604, "n_element_tab": 292, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 852, "n_element_tab_1": 150, "formula_len_all": 80, "formula_len_all_1": 51, "len_all": 180929, "len_all_1": 65460, "len_abs": 1400, "len_title": 118, "len_sents": 45117, "len_sents_1": 29988, "n_sents": 331, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1410, "title": "Improving Language Model Self-Correction Capability with Meta-Feedback", "abs": "Large language models (LLMs) are capable of self-correcting their responses by generating feedback and refining the initial output. However, their performance may sometimes decline following self-correction, either because the feedback contains errors or due to unnecessarily attempting to refine an already accurate response. To address these limitations, we investigate whether the same LLM can generate meta-feedback that pinpoints errors in the feedback rather than the response, an ability that remains under-explored despite extensive research on LLMs' self-feedback generation. We design a novel self-correction prompting framework, Feedback-on-Feedback (FoF), which leverages meta-feedback to improve the feedback before refining the response. Our framework first samples multiple pieces of feedback for the initial response, and prompts the LLM to generate meta-feedback that analyzes the inconsistency between these feedback pieces. Based on the meta-feedback, the LLM generates refined feedback that subsequently guides the revision of the response. Our FoF framework consistently outperforms competitive baselines across two LLMs on three datasets, covering arithmetic reasoning, machine translation, and programming tasks. Specifically, FoF improves performance on GSM8K by 3.6 points (45.2% vs. 41.6% for the initial answer) and on MBPP by 6.4 points (51.7% vs. 45.3%) using the LLaMA-3-8B model.", "keywords": [ "Self-Correction", "Meta-Feedback", "Iterative Refinement", "Feedback-on-Feedback (FoF)", "Natural Language Processing (NLP)", "Machine Learning", "Zero-Shot Learning", "Self-Refine", "Model Performance Enhancement", "Feedback Quality", "GSM8K Dataset", "MBPP Dataset", "CSMT Dataset" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "jJXZvPe5z0", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Idan Pipano", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "Undergrad student" }, { "name": "Itamar Reinman", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "Undergrad student" }, { "name": "Moshe Tennenholtz", "gender": "unknown", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "Full Professor" }, { "name": "Omer Madmon", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "PhD student" } ], "n_formula": 48, "n_formula_1": 11, "n_ref_uni": 32, "n_ref": 76, "n_ref_all": 123, "n_fig": 12, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3874, "formula_len_all_1": 1054, "len_all": 176579, "len_all_1": 73044, "len_abs": 999, "len_title": 155, "len_sents": 63041, "len_sents_1": 35223, "n_sents": 461, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 107, "L_abs": 1005, "title": "On the Convergence of No-Regret Dynamics in Information Retrieval Games with Proportional Ranking Functions", "abs": "Publishers who publish their content on the web act strategically, in a behavior that can be modeled within the online learning framework. \nRegret, a central concept in machine learning, serves as a canonical measure for assessing the performance of learning agents within this framework.\nWe prove that any proportional content ranking function with a concave activation function induces games in which no-regret learning dynamics converge. \nMoreover, for proportional ranking functions, we prove the equivalence of the concavity of the activation function, the social concavity of the induced games and the concavity of the induced games.\nWe also study the empirical trade-offs between publishers' and users' welfare, under different choices of the activation function, using a state-of-the-art no-regret dynamics algorithm. Furthermore, we demonstrate how the choice of the ranking function and changes in the ecosystem structure affect these welfare measures, as well as the dynamics' convergence rate.", "keywords": [ "Game theory", "no-regret dynamics", "recommendation systems", "information retrieval" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "jJJOoLVAEm", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "David Bull", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "Full Professor" }, { "name": "Fei Ye", "gender": "Male", "institution": "University of York", "country": "GB", "position": "PhD student" }, { "name": "Guoxi Huang", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "Research Associate" }, { "name": "Nantheera Anantrasirichai", "gender": "unknown", "institution": "University of Bristol", "country": "", "position": "Associate Professor" }, { "name": "Ruirui Lin", "gender": "unknown", "institution": "University of Bristol", "country": "GB", "position": "PhD student" }, { "name": "Zipeng Qi", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 14, "n_ref_uni": 54, "n_ref": 85, "n_ref_all": 147, "n_fig": 22, "n_tab": 10, "L_tab": 2671, "n_element_tab": 375, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2895, "n_element_tab_1": 278, "formula_len_all": 1137, "formula_len_all_1": 934, "len_all": 199916, "len_all_1": 61425, "len_abs": 1265, "len_title": 120, "len_sents": 54498, "len_sents_1": 23459, "n_sents": 435, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1279, "title": "Bayesian Enhancement Models for One-to-Many Mapping in Image Enhancement", "abs": "Image enhancement is considered an ill-posed inverse problem due to its tendency to have multiple solutions. The loss of information makes accurately reconstructing the original image from observed data challenging. Also, the quality of the result is often subjective to individual preferences. This obviously poses a one-to-many mapping challenge.\nTo address this, we propose a Bayesian Enhancement Model (BEM) that leverages Bayesian estimation to capture inherent uncertainty and accommodate diverse outputs. \nTo address the noise in predictions of Bayesian Neural Networks (BNNs) for high-dimensional images, we propose a two-stage approach. The first stage utilises a BNN to model reduced-dimensional image representations, while the second stage employs a deterministic network to refine these representations.\nWe further introduce a dynamic \\emph{Momentum Prior} to overcome convergence issues typically faced by BNNs in high-dimensional spaces.\nExtensive experiments across multiple low-light and underwater image enhancement benchmarks demonstrate the superiority of our method over traditional deterministic models, particularly in real-world applications lacking reference images, highlighting the potential of Bayesian models in handling one-to-many mapping problems.", "keywords": [ "Image Enhancement" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "jJ7azzLMdE", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Tuo An", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "Undergrad student" }, { "name": "Han Zou", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Jianfei Yang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "ZHOU YUNJIAO", "gender": "Female", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 46, "n_ref": 61, "n_ref_all": 79, "n_fig": 6, "n_tab": 3, "L_tab": 1397, "n_element_tab": 203, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 743, "n_element_tab_1": 122, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 126302, "len_all_1": 59855, "len_abs": 1755, "len_title": 123, "len_sents": 35340, "len_sents_1": 30094, "n_sents": 237, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1761, "title": "IoT-LLM: Enhancing Real-World IoT Task Reasoning with Large Language Models", "abs": "Large Language Models (LLMs) have demonstrated remarkable capabilities across textual and visual domains but often generate outputs that violate physical laws, revealing a gap in their understanding of the physical world. Inspired by human cognition\u2014where perception is fundamental to reasoning\u2014we explore augmenting LLMs with enhanced perception abilities using Internet of Things (IoT) sensor data and pertinent knowledge for IoT task reasoning in the physical world. In this work, we systematically study LLMs' capability to address real-world IoT tasks by augmenting their perception and knowledge base, and then propose a unified framework, IoT-LLM, to enhance such capability. In IoT-LLM, we customize three steps for LLMs: preprocessing IoT data into formats amenable to LLMs, activating their commonsense knowledge through chain-of-thought prompting and specialized role definitions, and expanding their understanding via IoT-oriented retrieval-augmented generation based on in-context learning. To evaluate the performance, We design a new benchmark with five real-world IoT tasks with different data types and reasoning difficulties and provide the benchmarking results on six open-source and close-source LLMs. Experimental results demonstrate the limitations of existing LLMs with naive textual inputs that cannot perform these tasks effectively. We show that IoT-LLM significantly enhances the performance of IoT tasks reasoning of LLM, such as GPT-4, achieving an average improvement of 65% across various tasks against previous methods. The results also showcase LLMs' ability to comprehend IoT data and the physical law behind data by providing a reasoning process. Limitations of our work are claimed to inspire future research in this new era.", "keywords": [ "large language models", "Internet of Things", "IoT task reasoning" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "jIOBhZO1ax", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Eric Vanden-Eijnden", "gender": "Male", "institution": "New York University", "country": "", "position": "Full Professor" }, { "name": "Mengjian Hua", "gender": "Male", "institution": "Courant Institute of Mathematical Sciences, New York University", "country": "", "position": "PhD student" }, { "name": "Tian Qi Chen", "gender": "Male", "institution": "FAIR Labs, Meta AI", "country": "US", "position": "Researcher" } ], "n_formula": 43, "n_formula_1": 44, "n_ref_uni": 41, "n_ref": 54, "n_ref_all": 125, "n_fig": 2, "n_tab": 2, "L_tab": 386, "n_element_tab": 12, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 3, "n_element_tab_1": 1, "formula_len_all": 3782, "formula_len_all_1": 3646, "len_all": 123632, "len_all_1": 65288, "len_abs": 1316, "len_title": 118, "len_sents": 33491, "len_sents_1": 25862, "n_sents": 253, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1326, "title": "Simulation-Free Differential Dynamics through Neural Conservation Laws", "abs": "We present a novel simulation-free framework for training continuous-time diffusion processes over very general objective functions. \nExisting methods typically involve either prescribing the optimal diffusion process---which only works for heavily restricted problem formulations---or require expensive simulation to numerically obtain the time-dependent densities and sample from the diffusion process.\nIn contrast, we propose a coupled parameterization which jointly models a time-dependent density function, or probability path, and the dynamics of a diffusion process that generates this probability path.\nTo accomplish this, our approach directly bakes in the Fokker-Planck equation and density function requirements as hard constraints, by extending and greatly simplifying the construction of Neural Conservation Laws.\nThis enables simulation-free training for a large variety of problem formulations, from data-driven objectives as in generative modeling and dynamical optimal transport, to optimality-based objectives as in stochastic optimal control, with straightforward extensions to mean-field objectives due to the ease of accessing exact density functions. We validate our method in a diverse range of application domains from modeling spatio-temporal events, to learning optimal dynamics from population data.", "keywords": [ "Generative Modeling; Simulation-free Methods; Diffusion Models;" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 2, 2, 4 ] }, { "paper_id": "jIAKjjEmWi", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Lukas Schott", "gender": "Male", "institution": "Bosch", "country": "DE", "position": "Researcher" }, { "name": "Martin Rapp", "gender": "Male", "institution": "Robert Bosch GmbH, Bosch", "country": "", "position": "Researcher" }, { "name": "Niclas Popp", "gender": "unknown", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Piyapat Saranrittichai", "gender": "Male", "institution": "Bosch Center for Artificial Intelligence", "country": "", "position": "PhD student" }, { "name": "Souptik Kumar Majumdar", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "DE", "position": "MS student" }, { "name": "Advait Harshal Gadhikar", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 71, "n_ref_all": 116, "n_fig": 16, "n_tab": 3, "L_tab": 995, "n_element_tab": 156, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 2386, "n_element_tab_1": 97, "formula_len_all": 287, "formula_len_all_1": 250, "len_all": 123373, "len_all_1": 54160, "len_abs": 1374, "len_title": 103, "len_sents": 32028, "len_sents_1": 23421, "n_sents": 243, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1396, "title": "Attention Is All You Need For Mixture-of-Depths Routing", "abs": "Advancements in deep learning are driven by training models with increasingly larger numbers of parameters, which in turn heightens the computational demands. To address this issue, Mixture-of-Depths (MoD) models have been proposed to dynamically assign computations only to the most relevant parts of the inputs, thereby enabling the deployment of large-parameter models with high efficiency during inference and training. These MoD models utilize a routing mechanism to determine which tokens should be processed by a layer, or skipped. However, conventional MoD models employ additional network layers specifically for the routing which are difficult to train, and add complexity and deployment overhead to the model. In this paper, we introduce a novel attention-based routing mechanism *A-MoD* that leverages the existing attention map of the preceding layer for routing decisions within the current layer. Compared to standard routing, *A-MoD* allows for more efficient training as it introduces no additional trainable parameters and can be easily adapted from pretrained transformer models. Furthermore, it can increase the performance of the MoD model. For instance, we observe up to $2$\\% higher accuracy on ImageNet compared to standard routing and isoFLOP ViT baselines. Furthermore, *A-MoD* improves the MoD training convergence, leading to up to $2\\times$ faster transfer learning.", "keywords": [ "Mixture of Depths", "attention", "parameter free routing." ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "jHdsZCOouv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chun Jason Xue", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" }, { "name": "Jun Wang", "gender": "Female", "institution": "City University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Nan Guan", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "YU MAO", "gender": "Female", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" } ], "n_formula": 24, "n_formula_1": 15, "n_ref_uni": 36, "n_ref": 66, "n_ref_all": 73, "n_fig": 1, "n_tab": 7, "L_tab": 2982, "n_element_tab": 231, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1392, "n_element_tab_1": 107, "formula_len_all": 1778, "formula_len_all_1": 1195, "len_all": 127738, "len_all_1": 58092, "len_abs": 1233, "len_title": 168, "len_sents": 31222, "len_sents_1": 24387, "n_sents": 248, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 139, "L_abs": 1400, "title": "SHAP-CAT: A interpretable multi-modal framework enhancing WSI classification via virtual staining and shapley-value-based multimodal fusion", "abs": "The multimodal model has demonstrated promise in histopathology. However, most multimodal models are based on H\\&E and genomics, adopting increasingly complex yet black-box designs. In our paper, we propose a novel interpretable multimodal framework named SHAP-CAT, which uses a Shapley-value-based dimension reduction technique for effective multimodal fusion. Starting with two paired modalities -- H\\&E and IHC images, we employ virtual staining techniques to enhance limited input data by generating a new clinical-related modality. We extract very lightweight bag-level representations from each image modality and apply a Shapley-value-based mechanism for dimension reduction. Lightweight bag-level representations are extracted from image modalities and a Shapley-value-based mechanism is used for dimension reduction.For each dimension of the bag-level representation, attribution values are calculated to indicate how changes in the specific dimensions of the input affect the model output. In this way, we select a few top important dimensions of bag-level representation for each image modality to late fusion. Our experimental results demonstrate that the proposed SHAP-CAT framework incorporating synthetic modalities significantly enhances model performance, yielding a 5\\% increase in accuracy for the BCI, an 8\\% increase for IHC4BC-ER, and an 11\\% increase for the IHC4BC-PR dataset.", "keywords": [ "multimodal", "data fusion", "computational pathology", "shapley value", "virtual staining" ], "rating_list": [ 1, 5, 5, 3, 3 ], "soundness_list": [ 1, 3, 3, 1, 1 ], "presentation_list": [ 1, 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 1, 1 ], "confidence_list": [ 4, 4, 4, 4, 5 ] }, { "paper_id": "jHVJQybLXi", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Ankit Singh", "gender": "Male", "institution": "Stellantis", "country": "IN", "position": "Researcher" }, { "name": "Mayug Maniparambil", "gender": "Male", "institution": "Amazon Robotics", "country": "DE", "position": "Intern" }, { "name": "Noel O'Connor", "gender": "Male", "institution": "Dublin City University", "country": "IE", "position": "Full Professor" }, { "name": "Raiymbek Akshulakov", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" }, { "name": "Sanath Narayan", "gender": "Male", "institution": "Technology Innovation Institute", "country": "AE", "position": "Researcher" }, { "name": "YASSER ABDELAZIZ DAHOU DJILALI", "gender": "Male", "institution": "Technology Innovation Institute", "country": "AE", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 26, "n_ref": 54, "n_ref_all": 70, "n_fig": 7, "n_tab": 13, "L_tab": 5805, "n_element_tab": 354, "n_fig_1": 6, "n_tab_1": 8, "L_tab_1": 3649, "n_element_tab_1": 283, "formula_len_all": 495, "formula_len_all_1": 88, "len_all": 192909, "len_all_1": 68054, "len_abs": 1485, "len_title": 118, "len_sents": 58038, "len_sents_1": 30120, "n_sents": 354, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1492, "title": "From Unimodal to Multimodal:Scaling up Projectors to Align Modalities", "abs": "Recent contrastive multimodal vision-language models like CLIP have demonstrated robust open-world semantic understanding, becoming the standard image backbones for vision-language applications due to their aligned latent space. However, this practice has left powerful unimodal encoders for both vision and language underutilized in these applications which raises a key question: Is there a plausible way to connect unimodal backbones for zero-shot vision-language tasks? To this end, we propose a novel approach that aligns vision and language modalities using only projection layers on pretrained, frozen unimodal encoders. Our method exploits the high semantic similarity between embedding spaces of well-trained vision and language models. It involves selecting semantically similar encoders in the latent space, curating a concept-rich dataset of image-caption pairs, and training simple MLP projectors. We evaluated our approach on 12 zero-shot classification datasets and 2 image-text retrieval datasets. Our best model, utilizing DINOv2 and All-Roberta-Large text encoder, achieves 76\\(\\%\\) accuracy on ImageNet with a 20-fold reduction in data and 65-fold reduction in compute requirements. The proposed framework enhances the accessibility of model development while enabling flexible adaptation across diverse scenarios, offering an efficient approach to building multimodal models by utilizing existing unimodal architectures. Code and datasets will be released upon acceptance.", "keywords": [ "vision language models", "vision language alignment", "unimodal models", "representational alignment" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 1, 1, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "jHKqr1sDDM", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Cedric Malherbe", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Talip Ucar", "gender": "unknown", "institution": "AstraZeneca", "country": "GB", "position": "Sr. Director, AI Research" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 18, "n_ref": 45, "n_ref_all": 61, "n_fig": 5, "n_tab": 3, "L_tab": 1393, "n_element_tab": 232, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1393, "n_element_tab_1": 232, "formula_len_all": 539, "formula_len_all_1": 539, "len_all": 97671, "len_all_1": 64663, "len_abs": 1150, "len_title": 122, "len_sents": 31632, "len_sents_1": 30470, "n_sents": 206, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1153, "title": "IgBleng: Unifying 3D structures and sequences in antibody language models", "abs": "Large language models (LLMs) trained on antibody sequences have shown significant potential in the rapidly advancing field of machine learning-assisted antibody engineering and drug discovery. However, current state-of-the-art antibody LLMs often overlook structural information, which could enable the model to more effectively learn the functional properties of antibodies by providing richer, more informative data. In response to this limitation, we introduce IgBlend, which integrates both the 3D coordinates of backbone atoms (C-alpha, N, and C) and antibody sequences. Our model is trained on a diverse dataset containing over 4 million unique structures and more than 200 million unique sequences, including heavy and light chains as well as nanobodies. We rigorously evaluate IgBlend using established benchmarks such as sequence recovery, complementarity-determining region (CDR) editing and inverse folding and demonstrate that IgBlend consistently outperforms current state-of-the-art models across all benchmarks. Furthermore, experimental validation shows that the model's log probabilities correlate well with measured binding affinities.", "keywords": [ "Antibodies", "LLM", "structure", "multi-modal" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "jGGylopiO8", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Chunhua Shen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Guangkai Xu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Hao Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Libo Sun", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Researcher" }, { "name": "Yanlong Sun", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yongtao Ge", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "PhD student" }, { "name": "ZhiyueZhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 56, "n_ref": 193, "n_ref_all": 209, "n_fig": 1, "n_tab": 10, "L_tab": 3565, "n_element_tab": 285, "n_fig_1": 1, "n_tab_1": 9, "L_tab_1": 5052, "n_element_tab_1": 343, "formula_len_all": 368, "formula_len_all_1": 317, "len_all": 154459, "len_all_1": 72239, "len_abs": 2054, "len_title": 111, "len_sents": 35106, "len_sents_1": 31031, "n_sents": 221, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 2062, "title": "Benchmarking and Analyzing Monocular Geometry Estimation Models", "abs": "Recent advances in discriminative and generative pretraining have yielded geometry estimation foundation models with strong generalization capabilities. While most discriminative monocular geometry estimation methods rely on large-scale finetuning data to achieve zero-shot generalization, several generative-based paradigms show the potential of achieving impressive generalization performance on unseen scenes by leveraging pre-trained diffusion models and fine-tuning on even a small-scale of synthetic training data. Frustratingly, these models are trained with different recipes on different datasets, making it hard to find out the critical factors that determine the evaluation performance. To resolve the above issue, (1) we build fair and strong baselines in a unified codebase for evaluating and analyzing the state-of-the-art (SOTA) geometry estimation models from pre-training style, finetuning data, and model architecture perspectives; (2) we thoroughly evaluate geometry models on challenging benchmarks with diverse scenes and high-quality annotations. Under the fair training and evaluation configuration, our results reveal that stochastic diffusion-based protocol is not optimal for fine-tuning generative-based geometry estimation methods. One-step finetuning and inference protocol is sufficient for generative-based depth and surface normal estimation. Besides, we find that both discriminative and generative pretraining can generalize well under small-scale fine-tuning high-quality data in scale-invariant depth estimation task. DINOv2-pretrained discriminative models achieve slightly higher performance than generative counterparts with the same small amount of synthetic data. Furthermore, we have observed that metric depth estimation requires significantly more finetuning data than scale-invariant depth estimation for learning the depth scale distribution. We hope this work will inspire future geometry estimation research in building more high-quality fine-tuning datasets and designing more powerful geometry estimation models.", "keywords": [ "generative and discrimative pretrain", "monocular geometry estimation", "benchmarks" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 4, 2 ], "contribution_list": [ 1, 2, 4, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "jFcNXJGPGh", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Lilian Tang", "gender": "Female", "institution": "University of Surrey", "country": "", "position": "Full Professor" }, { "name": "Qiushi Huang", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "PhD student" }, { "name": "Tom Ko", "gender": "Male", "institution": "ByteDance AI Lab", "country": "HK", "position": "Researcher" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 30, "n_ref": 36, "n_ref_all": 48, "n_fig": 1, "n_tab": 12, "L_tab": 4417, "n_element_tab": 501, "n_fig_1": 1, "n_tab_1": 9, "L_tab_1": 4357, "n_element_tab_1": 441, "formula_len_all": 652, "formula_len_all_1": 625, "len_all": 109192, "len_all_1": 63237, "len_abs": 929, "len_title": 106, "len_sents": 26997, "len_sents_1": 24532, "n_sents": 194, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 946, "title": "CoLoRA: A Competitive Learning Approach for Enhancing LoRA", "abs": "We propose a Competitive Low-Rank Adaptation (CoLoRA) framework to address the limitations of the LoRA method, which either lacks capacity with a single rank-$r$ LoRA or risks inefficiency and overfitting with a larger rank-$Kr$ LoRA, where $K$ is an integer larger than 1. The proposed CoLoRA method initializes $K$ distinct LoRA components, each with rank $r$, and allows them to compete during training. This competition drives each LoRA component to outperform the others, improving overall model performance. The best-performing LoRA is selected based on validation metrics, ensuring that the final model outperforms a single rank-$r$ LoRA and matches the effectiveness of a larger rank-$Kr$ LoRA, all while avoiding extra computational overhead during inference. To the best of our knowledge, this is the first work to introduce and explore competitive learning in the context of LoRA optimization. The CoLoRA's code will be released later.", "keywords": [ "Parametric-efficient fine-tuning" ], "rating_list": [ 5, 8, 6, 5 ], "soundness_list": [ 3, 4, 2, 1 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "jE5ZbtMtcU", "primary_area": "applications to neuroscience & cognitive science", "n_author": 7, "author_info_list": [ { "name": "Fangxu Zhou", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Postdoc" }, { "name": "Hao Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Jinzhuo Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Lishuang Feng", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Qichen Sun", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Wei Wu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhengrui Guo", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 8, "n_ref_uni": 29, "n_ref": 66, "n_ref_all": 95, "n_fig": 8, "n_tab": 5, "L_tab": 3824, "n_element_tab": 132, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 4394, "n_element_tab_1": 124, "formula_len_all": 936, "formula_len_all_1": 677, "len_all": 174471, "len_all_1": 63269, "len_abs": 1906, "len_title": 144, "len_sents": 65924, "len_sents_1": 28146, "n_sents": 422, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 2009, "title": "BLEND: Behavior-guided Neural Population Dynamics Modeling via Privileged Knowledge Distillation", "abs": "Modeling the nonlinear dynamics of neuronal populations represents a key pursuit in computational neuroscience. Recent research has increasingly focused on jointly modeling neural activity and behavior to unravel their interconnections. Despite significant efforts, these approaches often necessitate either intricate model designs or oversimplified assumptions. Given the frequent absence of perfectly paired neural-behavioral datasets in real-world scenarios when deploying these models, a critical yet understudied research question emerges: how to develop a model that performs well using only neural activity as input at inference, while benefiting from the insights gained from behavioral signals during training?\n\nTo this end, we propose **BLEND**, the **B**ehavior-guided neura**L** population dynamics mod**E**lling framework via privileged k**N**owledge **D**istillation. By considering behavior as privileged information, we train a teacher model that takes both behavior observations (privileged features) and neural activities (regular features) as inputs. A student model is then distilled using only neural activity. Unlike existing methods, our framework is model-agnostic and avoids making strong assumptions about the relationship between behavior and neural activity. This allows BLEND to enhance existing neural dynamics modeling architectures without developing specialized models from scratch. Extensive experiments across neural population activity modeling and transcriptomic neuron identity prediction tasks demonstrate strong capabilities of BLEND, reporting over 50% improvement in behavioral decoding and over 15% improvement in transcriptomic neuron identity prediction after behavior-guided distillation. Furthermore, we empirically explore various behavior-guided distillation strategies within the BLEND framework and present a comprehensive analysis of effectiveness and implications for model performance. Code will be made available at https://github.com/dddavid4real/BLEND.", "keywords": [ "Computational Neuroscience", "Neural Dynamics Modeling", "Behavior as Guidance", "Privileged Knowledge Distillation" ], "rating_list": [ 5, 6, 5, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "jDvgxHhtlQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Hang Gao", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Yongfeng Zhang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 36, "n_ref": 47, "n_ref_all": 48, "n_fig": 0, "n_tab": 3, "L_tab": 1669, "n_element_tab": 186, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1238, "n_element_tab_1": 126, "formula_len_all": 704, "formula_len_all_1": 763, "len_all": 115589, "len_all_1": 69352, "len_abs": 2007, "len_title": 128, "len_sents": 37314, "len_sents_1": 33546, "n_sents": 260, "n_sents_1": 241, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 2136, "title": "VRSD: Rethinking Similarity and Diversity for Retrieval in Large Language Models", "abs": "Vector retrieval algorithms are essential for semantic queries within the rapidly evolving landscape of Large Language Models (LLMs). The ability to retrieve vectors that satisfy both similarity and diversity criteria substantially enhances the performance of LLMs. Although Maximal Marginal Relevance (MMR) is widely employed in retrieval scenarios requiring relevance and diversity, variations in the parameter \\( \\lambda \\) lead to fluctuations that complicate the optimization trajectory in vector spaces. This obscures the direction of improvement and highlights the lack of a robust theoretical analysis regarding similarity and diversity constraints in retrieval processes. To address these challenges, this paper introduces a novel approach that characterizes both constraints through the relationship between the sum vector and the query vector. The proximity of these vectors ensures the similarity constraint, while requiring individual vectors within the sum vector to diverge in their alignment with the query vector satisfies the diversity constraint. We first formulate a new combinatorial optimization problem, selecting \\( k \\) vectors from a candidate set such that their sum vector maximally aligns with the query vector, and demonstrate that this problem is \\textbf{NP-complete}. This result underscores the inherent difficulty of simultaneously achieving similarity and diversity in vector retrieval, thereby providing a theoretical foundation for future research. Subsequently, we present the heuristic algorithm \\underline{\\textbf{V}}ectors \\underline{\\textbf{R}}etrieval with \\underline{\\textbf{S}}imilarity and \\underline{\\textbf{D}}iversity, \\textbf{VRSD}, which features a clear optimization objective and eliminates the need for preset parameters. VRSD also achieves a modest reduction in time complexity compared to MMR. Empirical validation confirms that VRSD significantly outperforms MMR across various datasets, while also demonstrating that the sum vector effectively captures both diversity and similarity simultaneously. The data and code are available at https://anonymous.4open.science/r/VRSD-CF9D.", "keywords": [ "Algorithms", "Large Language Model", "NP-complete", "Vector Retrieval" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 1, 2, 1, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "jDsmB4o5S0", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Ellie Pavlick", "gender": "Female", "institution": "Brown University", "country": "", "position": "Associate Professor" }, { "name": "Michael A. Lepori", "gender": "Male", "institution": "Brown University", "country": "US", "position": "PhD student" }, { "name": "Suraj Anand", "gender": "unknown", "institution": "Brown University", "country": "US", "position": "MS student" }, { "name": "jack merullo", "gender": "Male", "institution": "Brown University", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 45, "n_ref": 83, "n_ref_all": 126, "n_fig": 12, "n_tab": 1, "L_tab": 186, "n_element_tab": 2, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 445, "n_element_tab_1": 6, "formula_len_all": 347, "formula_len_all_1": 0, "len_all": 173929, "len_all_1": 64371, "len_abs": 1537, "len_title": 149, "len_sents": 57616, "len_sents_1": 31892, "n_sents": 423, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 101, "L_abs": 1571, "title": "Dual Process Learning: Controlling Use of In-Context vs. In-Weights Strategies with Weight Forgetting", "abs": "Language models have the ability to perform in-context learning (ICL), allowing them to flexibly adapt their behavior based on context. This contrasts with in-weights learning (IWL), where memorized information is encoded in model parameters from iterated observations of the data (e.g., common sayings). An ideal model should be able to maintain both of these abilities. Despite their apparent ability to learn in-context, language models are known to struggle when faced with unseen or rarely seen tokens (Land & Bartolo, 2024). Hence, we study $\\textbf{structural in-context learning}$, which we define as the ability of a model to execute in-context learning on arbitrary novel tokens \u2013 so called because the model must generalize on the basis of e.g. sentence structure or task structure, rather than content encoded in token embeddings. We study structural in-context algorithms on both synthetic and natural tasks using both toy models and MultiBERT models (Sellam et al., 2021). We find that structural ICL appears before quickly disappearing early in LM pretraining. While it has been shown that ICL can diminish during training (Singh et al., 2023), we find that prior work does not account for structural ICL. Building on the Chen et al. (2024) 's active forgetting method used to help models learn new languages, we introduce a pretraining method that can modulate the preference for true structural ICL and IWL. Importantly, this allows us to induce a $\\textit{dual process strategy}$ where in-context and in-weights solutions coexist within a single model.", "keywords": [ "Natural Language Processing", "In-Context Learning", "In-Weights Learning", "Active Forgetting" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "jDpdQPMosW", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Ammar Gilani", "gender": "Male", "institution": "Northwestern University", "country": "", "position": "PhD student" }, { "name": "Chenyang Li", "gender": "Male", "institution": "Fuzhou University", "country": "CN", "position": "Undergrad student" }, { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" }, { "name": "Jerry Yao-Chieh Hu", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Wei-Po Wang", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "Undergrad student" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" } ], "n_formula": 167, "n_formula_1": 29, "n_ref_uni": 48, "n_ref": 161, "n_ref_all": 169, "n_fig": 2, "n_tab": 3, "L_tab": 122, "n_element_tab": 16, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 14233, "formula_len_all_1": 2358, "len_all": 331049, "len_all_1": 83087, "len_abs": 2034, "len_title": 137, "len_sents": 95616, "len_sents_1": 32994, "n_sents": 1027, "n_sents_1": 308, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1284, "title": "Fundamental Limits of Prompt Tuning Transformers: Universality, Capacity and Efficiency", "abs": "We investigate the statistical and computational limits of prompt tuning for transformer-based foundation models. \nOur key contributions are prompt tuning on *single-head* transformers with only a *single* self-attention layer: \n(i) is universal, and (ii) supports efficient (even nearly-linear time) algorithms under the Strong Exponential Time Hypothesis (SETH).\nStatistically, \nwe prove that prompt tuning on such simplest possible transformers are universal approximators for sequence-to-sequence Lipschitz functions. \nIn addition, we provide an exponential-in-$dL$ and -in-$(1/\\epsilon)$ lower bound on the required soft-prompt tokens for prompt tuning to memorize any dataset with 1-layer, 1-head transformers.\nComputationally, we identify a phase transition in the efficiency of prompt tuning, determined by the norm of the *soft-prompt-induced* keys and queries, and provide an upper bound criterion.\nBeyond this criterion, no sub-quadratic (efficient) algorithm for prompt tuning exists under SETH. \nWithin this criterion, \nwe showcase our theory by proving the existence of almost-linear time prompt tuning inference algorithms.\nThese fundamental limits provide important necessary conditions for designing expressive and efficient prompt tuning methods for practitioners.", "keywords": [ "Foundation Model", "Prompt Tuning", "Transformer", "Universal Approximation", "Memory Capacity", "Computational Efficiency", "Fine-Grained Complexity" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "jCPak79Kev", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Jian Gao", "gender": "unknown", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Junyi Yang", "gender": "unknown", "institution": "Northeastern University", "country": "US", "position": "MS student" }, { "name": "Xuan Zhang", "gender": "unknown", "institution": "Northeastern University", "country": "US", "position": "Associate Professor" }, { "name": "WEIDONG CAO", "gender": "Male", "institution": "George Washington University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 70, "n_ref_all": 87, "n_fig": 7, "n_tab": 3, "L_tab": 1119, "n_element_tab": 254, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1049, "n_element_tab_1": 74, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 110885, "len_all_1": 66440, "len_abs": 1482, "len_title": 134, "len_sents": 37869, "len_sents_1": 34248, "n_sents": 241, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 85, "L_abs": 1537, "title": "AnalogGenie: A Generative Engine for Automatic Discovery of Analog Circuit Topologies", "abs": "The massive and large-scale design of foundational semiconductor integrated circuits (ICs) is crucial to sustaining the advancement of many emerging and future technologies, such as generative AI, 5G/6G, and quantum computing.\nExcitingly, recent studies have shown the great capabilities of foundational models in expediting the design of digital ICs.\nYet, applying generative AI techniques to accelerate the design of analog ICs remains a significant challenge due to critical domain-specific issues, such as the lack of a comprehensive dataset and effective representation methods for analog circuits.\nThis paper proposes, $\\textbf{AnalogGenie}$, a $\\underline{\\textbf{Gen}}$erat$\\underline{\\textbf{i}}$ve $\\underline{\\textbf{e}}$ngine for automatic design/discovery of $\\underline{\\textbf{Analog}}$ circuit topologies--the most challenging and creative task in the conventional manual design flow of analog ICs.\nAnalogGenie addresses two key gaps in the field: building a foundational comprehensive dataset of analog circuit topology and developing a scalable sequence-based graph representation universal to analog circuits.\nExperimental results show the remarkable generation performance of AnalogGenie in broadening the variety of analog ICs, increasing the number of devices within a single design, and discovering unseen circuit topologies far beyond any prior arts.\nOur work paves the way to transform the longstanding time-consuming manual design flow of analog ICs to an automatic and massive manner powered by generative AI.", "keywords": [ "Circuit Generation", "Application of Generative Models", "Electronic Design Automation" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 4 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "jCNRcHrfLo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chenjue Zhang", "gender": "unknown", "institution": "Beihang University", "country": "CN", "position": "Undergrad student" }, { "name": "Duorui Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Xianglong Liu", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaowei Zhao", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Yuqing Ma", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "Associate Professor" }, { "name": "Zhiwan Fang", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" } ], "n_formula": 10, "n_formula_1": 12, "n_ref_uni": 8, "n_ref": 26, "n_ref_all": 45, "n_fig": 6, "n_tab": 4, "L_tab": 338, "n_element_tab": 62, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 880, "n_element_tab_1": 107, "formula_len_all": 603, "formula_len_all_1": 605, "len_all": 92495, "len_all_1": 61953, "len_abs": 1557, "len_title": 137, "len_sents": 35421, "len_sents_1": 29071, "n_sents": 237, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1586, "title": "Hierarchical Prompts with Context-aware Calibration for Open-Vocabulary Object Detection", "abs": "Open Vocabulary Object Detection (OVD) aims to extend to novel classes solely through text descriptions, by learning the mapping between images and text from the base class. However, current methods focus on establishing connections between the visual regions of the target objects and their corresponding category names to learn prompts, ignoring richer contextual information and shared knowledge about these categories, which can easily lead to overfitting on known base categories and exhibit poor generalization to novel classes. To address the above problems, we propose Hierarchical prompts with Context-Aware calibration (HiCA) for open-vocabulary object detection, which integrates high-level semantic and contextual information into the detector from both linguistic and visual perspectives.\nHierarchical prompts effectively map regions with superior-level semantics, which encompasses shared knowledge of both base and novel classes, thereby enhancing the model's generalization ability to novel classes. Context-aware calibration utilizes the visual context of the image to establish the correlation between contextual information and categories, thereby minimizing the adverse effects of the background and enhancing generalization to novel classes. Extensive experiments demonstrate that the hierarchical prompts with context-aware calibration can effectively improve the performance of the open vocabulary detection methods. Especially on the OV-COCO, we achieve 57.2% base class mAP, surpassing the current state-of-the-art by 2.4% while achieving the best overall mAP.", "keywords": [ "open-vocabulary object detection", "prompts tuning", "knowledge distillation" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "jCDF7G3LpF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Aditya Ramesh", "gender": "unknown", "institution": "Fujitsu Research and Development Center Co. Ltm.", "country": "IN", "position": "Researcher" }, { "name": "Aditya Saibewar", "gender": "unknown", "institution": "Fujitsu Research and Development Center Co. Ltm.", "country": "IN", "position": "Researcher" }, { "name": "Manohar Kaul", "gender": "Male", "institution": "Fujitsu Research and Development Center Co. Ltm.", "country": "IN", "position": "Principal Researcher" }, { "name": "Shivam Bhardwaj", "gender": "Male", "institution": "Fujitsu Research and Development Center Co. Ltm.", "country": "IN", "position": "Researcher" } ], "n_formula": 38, "n_formula_1": 8, "n_ref_uni": 46, "n_ref": 74, "n_ref_all": 87, "n_fig": 7, "n_tab": 6, "L_tab": 1851, "n_element_tab": 98, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 683, "n_element_tab_1": 46, "formula_len_all": 1654, "formula_len_all_1": 590, "len_all": 264706, "len_all_1": 59213, "len_abs": 1370, "len_title": 156, "len_sents": 97469, "len_sents_1": 27014, "n_sents": 631, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 108, "L_abs": 1380, "title": "EFFICIENT JAILBREAK ATTACK SEQUENCES ON LARGE LANGUAGE MODELS VIA MULTI-ARMED BANDIT-BASED CONTEXT SWITCHING", "abs": "Content warning: This paper contains examples of harmful language and content.\nRecent advances in large language models (LLMs) have made them increasingly vulnerable to jailbreaking attempts, where malicious users manipulate models into generating harmful content. While existing approaches rely on either single-step attacks that trigger immediate safety responses or multi-step methods that inefficiently iterate prompts using other LLMs, we introduce ``Sequence of Context\" (SoC) attacks that systematically alter conversational context through strategically crafted context-switching queries (CSQs). We formulate this as a multi-armed bandit (MAB) optimization problem, automatically learning optimal sequences of CSQs that gradually weaken the model's safety boundaries. Our theoretical analysis provides tight bounds on both the expected sequence length until successful jailbreak and the convergence of cumulative rewards. Empirically, our method achieves a 95\\% attack success rate, surpassing PAIR by 63.15\\%, AutoDAN by 60\\%, and ReNeLLM by 50\\%. We evaluate our attack across multiple open-source LLMs including Llama and Mistral variants. Our findings highlight critical vulnerabilities in current LLM safeguards and emphasize the need for defenses that consider sequential attack patterns rather than relying solely on static prompt filtering or iterative refinement.", "keywords": [ "JailBreak", "AI Security", "LLM Vunlnerability" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "jC6E2iTgfr", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Chunwei Yang", "gender": "Female", "institution": "Sichuan University", "country": "CN", "position": "Undergrad student" }, { "name": "Jinguo Cheng", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "MS student" }, { "name": "Wanlin Cai", "gender": "Male", "institution": "Sichuan University", "country": "", "position": "MS student" }, { "name": "Yuxuan Liang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Qingsong Wen", "gender": "Male", "institution": "Squirrel Ai Learning", "country": "US", "position": "Principal Researcher" }, { "name": "wu yuankai", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 2, "n_ref_uni": 46, "n_ref": 78, "n_ref_all": 113, "n_fig": 13, "n_tab": 23, "L_tab": 23707, "n_element_tab": 3978, "n_fig_1": 4, "n_tab_1": 9, "L_tab_1": 4027, "n_element_tab_1": 620, "formula_len_all": 1010, "formula_len_all_1": 89, "len_all": 232561, "len_all_1": 65119, "len_abs": 1627, "len_title": 111, "len_sents": 42404, "len_sents_1": 24251, "n_sents": 364, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1636, "title": "NuwaTS: a Foundation Model Mending Every Incomplete Time Series", "abs": "Time series imputation is critical for many real-world applications and has been widely studied. However, existing models often require specialized designs tailored to specific missing patterns, variables, or domains which limits their generalizability. In addition, current evaluation frameworks primarily focus on domain-specific tasks and often rely on time-wise train/validation/test data splits, which fail to rigorously assess a model\u2019s ability to generalize across unseen variables or domains. In this paper, we present \\textbf{NuwaTS}, a novel framework that repurposes Pre-trained Language Models (PLMs) for general time series imputation. Once trained, NuwaTS can be applied to impute missing data across any domain. We introduce specialized embeddings for each sub-series patch, capturing information about the patch, its missing data patterns, and its statistical characteristics. By combining contrastive learning with the imputation task, we train PLMs to create a versatile, one-for-all imputation model. Additionally, we employ a plug-and-play fine-tuning approach, enabling efficient adaptation to domain-specific tasks with minimal adjustments. To evaluate cross-variable and cross-domain generalization, we propose a new benchmarking protocol that partitions the datasets along the variable dimension. Experimental results on over seventeen million time series from diverse domains demonstrate that NuwaTS outperforms state-of-the-art domain-specific models across various datasets under the proposed benchmarking protocol. Furthermore, we show that NuwaTS generalizes to other time series tasks, such as forecasting.", "keywords": [ "Incomplete Time Series", "Missing Data Imputation", "Cross-domain adaptation" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "jBpEsliki9", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Prompong Sugunnasil", "gender": "Male", "institution": "Chiang Mai University", "country": "TH", "position": "Assistant Professor" }, { "name": "Vivatchai Kaveeta", "gender": "Male", "institution": "Chiang Mai University", "country": "TH", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 22, "n_ref": 23, "n_ref_all": 27, "n_fig": 1, "n_tab": 1, "L_tab": 30, "n_element_tab": 2, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 30, "n_element_tab_1": 2, "formula_len_all": 174, "formula_len_all_1": 174, "len_all": 65752, "len_all_1": 65752, "len_abs": 1531, "len_title": 117, "len_sents": 22713, "len_sents_1": 22713, "n_sents": 177, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1528, "title": "Hypergraph-Based Machine Learning for Robust Handling of Missing Data", "abs": "Handling missing data is a major challenge in machine learning where missing values are common in datasets. This work introduces a hypergraph representation constructed from datasets containing missing values. The method does not rely on traditional techniques like deletion or data imputations. The approach constructs hypergraphs directly from the dataset, preserving the relationships between variables and modeling multi-variable interactions. This enables the model to capture the dataset structure in ways other methods may overlook. The proposed hypergraph learning method can be applied to classification and regression tasks. For real-world evaluation, we use the MIMIC-III and Adult datasets focusing on classification performance. Additionally, synthetic datasets with controlled missingness are used to evaluate the method's effectiveness across varying degrees of missingness. When compared with imputation and prediction techniques, the hypergraph approach achieves competitive or superior performance. Specifically, our method maintains high performance in scenarios with significant levels of missing data. We demonstrate that the hypergraph representation not only offers a more resilient framework for learning from datasets with missing data. But also scales effectively across diverse datasets and prediction tasks. The method maintains stable performance under various degrees of missingness, demonstrating its potential as a valuable machine learning tool with high data reliability and prediction quality.", "keywords": [ "Hypergraph", "Machine Learning", "Missing Data" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 2, 1, 2 ], "presentation_list": [ 1, 1, 2, 2 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 4, 5, 2, 4 ] }, { "paper_id": "jBatISjqSn", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Abdul Hameed Azeemi", "gender": "Male", "institution": "Lahore University of Management Sciences", "country": "PK", "position": "PhD student" }, { "name": "Agha Ali Raza", "gender": "Male", "institution": "Lahore University of Management Sciences", "country": "PK", "position": "Associate Professor" }, { "name": "Ihsan Ayyub Qazi", "gender": "unknown", "institution": "Lahore University of Management Sciences", "country": "", "position": "Associate Professor" } ], "n_formula": 21, "n_formula_1": 5, "n_ref_uni": 60, "n_ref": 103, "n_ref_all": 121, "n_fig": 4, "n_tab": 10, "L_tab": 3665, "n_element_tab": 334, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 3083, "n_element_tab_1": 206, "formula_len_all": 985, "formula_len_all_1": 228, "len_all": 177786, "len_all_1": 70062, "len_abs": 2327, "len_title": 116, "len_sents": 44277, "len_sents_1": 31612, "n_sents": 312, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1631, "title": "Language Model-Driven Data Pruning Enables Efficient Active Learning", "abs": "Active learning (AL) optimizes data labeling efficiency by selecting the most informative instances for annotation. A key component in this procedure is an acquisition function that guides the selection process and identifies the suitable instances for labeling from the unlabeled pool. However, these acquisition methods suffer from high computational costs with large unlabeled data pools, posing a roadblock to their applicability on large datasets. To address this challenge and bridge this gap, we introduce a novel plug-and-play unlabeled data pruning strategy, ActivePrune, which leverages language models to prune the unlabeled pool. ActivePrune implements a two-stage pruning process: an initial fast evaluation using perplexity scores from an n-gram language model, followed by a high-quality selection using metrics for data quality computed through a quantized LLM. Additionally, to enhance the diversity in the unlabeled pool, we propose a novel perplexity reweighting method that systematically brings forward underrepresented instances for selection in subsequent labeling iterations. Experiments on translation, sentiment analysis, topic classification, and summarization tasks on four diverse datasets and four active learning strategies demonstrate that ActivePrune outperforms existing data pruning methods. Finally, we compare the selection quality $\\leftrightarrow$ efficiency tradeoff of the data pruning methods and demonstrate that ActivePrune is computationally more efficient than other LLM score-based pruning methods, and provides up to 74% reduction in the end-to-end time required for active learning.", "keywords": [ "Active Learning", "Data Pruning", "Language Model" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "jBYQAtzp5Z", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Albert Zomaya", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Full Professor" }, { "name": "Chunhao Li", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "", "position": "Undergrad student" }, { "name": "Chunqiu xia", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Tianming Zhao", "gender": "Male", "institution": "University of SyAUdney", "country": "AU", "position": "Postdoc" }, { "name": "Wei Li", "gender": "Male", "institution": "The University of Sydney", "country": "AU", "position": "Lecturer" }, { "name": "Xiaomin Chang", "gender": "Male", "institution": "Macquarie University", "country": "AU", "position": "Postdoc" } ], "n_formula": 106, "n_formula_1": 11, "n_ref_uni": 18, "n_ref": 48, "n_ref_all": 61, "n_fig": 11, "n_tab": 4, "L_tab": 3503, "n_element_tab": 203, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 399, "n_element_tab_1": 23, "formula_len_all": 12212, "formula_len_all_1": 821, "len_all": 294981, "len_all_1": 84813, "len_abs": 1353, "len_title": 92, "len_sents": 99628, "len_sents_1": 35294, "n_sents": 1116, "n_sents_1": 309, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 44, "L_abs": 1546, "title": "Competitive Fair Scheduling with Predictions", "abs": "Beyond the worst-case analysis of algorithms, the recent learning-augmented framework considers that an algorithm can leverage possibly imperfect predictions about the unknown variables to make decisions to have guarantees tied to the prediction quality. We consider online non-clairvoyant scheduling to minimize the max-stretch under the learning-augmented framework, where the scheduler has access to job size predictions. We present a family of algorithms: *Relaxed-Greedy (RG)* with an $O(\\eta^3 \\cdot \\sqrt{P})$ competitive ratio, where $\\eta$ denotes the prediction error for job sizes and $P$ is the maximum job size ratio; *Adaptive Relaxed-Greedy* with an $O(\\lambda^{0.5} \\cdot \\eta^{2.5} \\cdot \\sqrt{P})$ competitive ratio, where $\\lambda$ denotes the prediction error for the minimum job size; *Predictive Relaxed-Greedy* with an $O(\\lambda^{0.5} \\cdot \\varphi^{0.5} \\cdot \\eta \\cdot \\max \\\\{ \\eta, \\varphi \\\\} \\cdot \\sqrt{P})$ competitive ratio, where $\\varphi$ denotes the prediction error for the maximum job size. We also present *${RG}^x$*, an algorithm that represents a trade-off between consistency and smoothness, with an $O(\\eta^{2+2x} \\cdot P^{1-x})$ competitive ratio. We introduce a general method using resource augmentation to bound robustness, resulting in *RR*-augmented *RG*, which achieves a $(1 + \\epsilon)$-speed $O(\\min \\\\{ \\eta^3 \\sqrt{P}, \\frac{n}{\\epsilon} \\\\})$ competitive ratio. Finally, we conduct simulations on synthetic and real-world datasets to evaluate the practical performance of these algorithms.", "keywords": [ "Learning-augmented Algorithms", "Scheduling", "Competitive analysis", "Fairness", "Predictions" ], "rating_list": [ 6, 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 4 ], "presentation_list": [ 2, 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 2, 4 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "jBBjZp0EVs", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Can Zhang", "gender": "Female", "institution": "United Automotive Electronic Systems", "country": "", "position": "Researcher" }, { "name": "Chuang Liu", "gender": "unknown", "institution": "United Automotive Electronic Systems Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Lin Sun", "gender": "Male", "institution": "UAES", "country": "", "position": "Researcher" }, { "name": "Ning Wu", "gender": "Male", "institution": "UAES", "country": "", "position": "Researcher" }, { "name": "Tao Yang", "gender": "Male", "institution": "United Automotive Electronic Systems Co., Ltd.", "country": "", "position": "Researcher" }, { "name": "Weijia Lu", "gender": "Male", "institution": "UAES", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 3, "n_formula_1": 5, "n_ref_uni": 27, "n_ref": 51, "n_ref_all": 65, "n_fig": 5, "n_tab": 8, "L_tab": 2987, "n_element_tab": 563, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 3490, "n_element_tab_1": 557, "formula_len_all": 446, "formula_len_all_1": 200, "len_all": 200626, "len_all_1": 67587, "len_abs": 1590, "len_title": 125, "len_sents": 53672, "len_sents_1": 24099, "n_sents": 631, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 922, "title": "Exchange of Perspective Prompting Enhances Reasoning in Large Language Models", "abs": "Large language models (LLMs) have made significant advancements in addressing diverse natural language processing (NLP) tasks. However, their performance is often limited by inherent comprehension of problems. To address this limitation, we propose Exchange-of-Perspective (EoP), a novel framework designed to exchange perspectives across different definitions of problem, so that it can break the fixed mindset from any particular formulation of the question. We conducted extensive and comprehensive experiments on 8 benchmarks. The results show that EoP can significantly improve performance. For instance, compared to the non-commutative baseline PHP, with GPT-3.5-Turbo and EoP, we observe a 3.6% improvement on AQuA (60.6% \u2192 64.2%), while GPT-4-powered EoP demonstrates a 7.7% overall accuracy enhancement on Math (53.9% \u2192 61.6%) and a 3.5% improvement on OlympiadBench Maths (43.5% \u2192 47.0%) when using Qwen-2.5-72b.", "keywords": [ "Large Language Models", "Reasoning", "Self-correction", "External Perspective" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "j9wBgcxa7N", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Archiki Prasad", "gender": "Female", "institution": "University of North Carolina, Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Elias Stengel-Eskin", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Postdoc" }, { "name": "Mohit Bansal", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" }, { "name": "Chih Yao Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Swarnadeep Saha", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 1, "n_ref_uni": 40, "n_ref": 86, "n_ref_all": 141, "n_fig": 8, "n_tab": 10, "L_tab": 3441, "n_element_tab": 440, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1127, "n_element_tab_1": 156, "formula_len_all": 242, "formula_len_all_1": 2, "len_all": 214934, "len_all_1": 74942, "len_abs": 2272, "len_title": 73, "len_sents": 66147, "len_sents_1": 36247, "n_sents": 551, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 2229, "title": "MAgICoRe: Multi-Agent, Iterative, Coarse-to-Fine Refinement for Reasoning", "abs": "Large Language Models' (LLM) reasoning can be improved using test-time aggregation strategies, i.e., generating multiple samples for each problem and aggregating over them to find a better answer. While these improve performance, they often reach a saturation point beyond which additional samples provide no return. Refinement offers an alternative by using model-generated feedback to improve answer quality. However, refinement introduces three key challenges: (1) Excessive refinement: Uniformly refining all instances can cause over-correction and reduce overall performance. (2) Inability to localize and address errors: LLMs have a limited ability to self-correct and struggle to identify and correct their own mistakes in a targeted way. (3) Insufficient refinement: Deciding how many iterations of refinement are needed is non-trivial, and stopping too soon could leave errors unaddressed. To tackle these issues, we propose MAgICoRe, a framework for Multi-Agent Iteration for Coarse-to-fine Refinement. MAgICoRe aims to avoid excessive refinement by categorizing problems as easy or hard, solving easy problems with coarse-grained aggregation, and solving hard ones with fine-grained and iterative multi-agent refinement. To enable more granular error localization, we incorporate external step-wise reward model (RM) scores. To ensure effective refinement, we employ a multi-agent loop with three agents: the Solver, the Reviewer (which generates targeted feedback based on step-wise RM scores) and the Refiner (which incorporates feedback and generates new solutions). To ensure sufficient refinement, we re-evaluate updated solutions, iteratively initiating further rounds of multi-agent refinement. We evaluate MAgICoRe on Llama-3-8B and GPT-3.5 and show its effectiveness across five math reasoning datasets, with consistent gains for all datasets and models. Even one iteration of MAgICoRe beats Self-Consistency by 3.4%, Best-of-k by 3.2%, and Self-Refine by 4.0% while using less than 50% of the samples. Unlike iterative refinement with baseline methods, MAgICoRe continues to improve with more iterations. Finally, our ablations highlight the importance of MAgICoRe's use of RMs and multi-agent communication.", "keywords": [ "LLM Refinement", "Reasoning", "Multi-Agent" ], "rating_list": [ 5, 5, 3, 6, 5 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 4, 3 ] }, { "paper_id": "j9VVzueEbG", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Boxing Chen", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Principal Researcher" }, { "name": "Boyu Wang", "gender": "Male", "institution": "University of Western Ontario", "country": "CA", "position": "Assistant Professor" }, { "name": "Charles Ling", "gender": "Male", "institution": "Western University", "country": "CA", "position": "Professor" }, { "name": "Gezheng Xu", "gender": "Female", "institution": "University of Western Ontario", "country": "CA", "position": "PhD student" }, { "name": "Jerry Huang", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "Peng Lu", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "PhD student" }, { "name": "QIUHAO Zeng", "gender": "Male", "institution": "University of Western Ontario", "country": "CA", "position": "PhD student" } ], "n_formula": 46, "n_formula_1": 13, "n_ref_uni": 38, "n_ref": 72, "n_ref_all": 90, "n_fig": 8, "n_tab": 6, "L_tab": 1156, "n_element_tab": 130, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 2023, "n_element_tab_1": 158, "formula_len_all": 3481, "formula_len_all_1": 716, "len_all": 201299, "len_all_1": 70571, "len_abs": 1619, "len_title": 109, "len_sents": 50294, "len_sents_1": 33816, "n_sents": 371, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1636, "title": "ZETA: Leveraging $Z$-order Curves for Efficient Top-$k$ Attention", "abs": "Over recent years, the Transformer has become a fundamental building block for sequence modeling architectures. Yet at its core is the use of self-attention, whose memory and computational cost grow quadratically with the sequence length $N$, rendering it prohibitively expensive for long sequences. A promising approach is top-$k$ attention, which selects only the $k$ most relevant tokens and achieves performance comparable to vanilla self-attention while significantly reducing space and computational demands. However, causal masks require the current query token to only attend to past tokens, preventing existing top-$k$ attention methods from efficiently searching for the most relevant tokens in parallel, thereby limiting training efficiency. In this work, we propose ZETA, leveraging Z-Order Curves for Efficient Top-k Attention, to enable parallel querying of past tokens for entire sequences. We first theoretically show that the choice of key and query dimensions involves a trade-off between the curse of dimensionality and the preservation of relative distances after projection. In light of this insight, we propose reducing the dimensionality of keys and queries in contrast to values and further leveraging Z-order curves to map low-dimensional keys and queries into one-dimensional space, which permits parallel sorting, thereby largely improving the efficiency for top-$k$ token selection. Experimental results demonstrate that ZETA~matches the performance of standard attention on synthetic tasks Associative Recall and outperforms attention and its variants on Long-Range Arena and WikiText-103 language modeling.", "keywords": [ "Transformer", "In-context learning", "Long Context", "long-range Transformer", "Efficient Transformer" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 4, 3, 4, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "j9DbobO0mY", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Inyoung Choi", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Jiayi Xin", "gender": "Female", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Jie Peng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Qi Long", "gender": "unknown", "institution": "University of Pennsylvania", "country": "US", "position": "Professor" }, { "name": "Sukwon Yun", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 47, "n_ref": 83, "n_ref_all": 110, "n_fig": 7, "n_tab": 4, "L_tab": 3761, "n_element_tab": 266, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1466, "n_element_tab_1": 81, "formula_len_all": 472, "formula_len_all_1": 471, "len_all": 167715, "len_all_1": 70440, "len_abs": 1798, "len_title": 145, "len_sents": 46827, "len_sents_1": 33856, "n_sents": 338, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1205, "title": "Sparse MoE as a New Retriever: Addressing Missing Modality Problem in Incomplete Multimodal Data", "abs": "In multimodal machine learning, effectively addressing the missing modality scenario is crucial for improving performance in downstream tasks such as in medical contexts where data may be incomplete. Although some attempts have been made to effectively retrieve embeddings for missing modalities, two main bottlenecks remain: the consideration of both intra- and inter-modal context, and the cost of embedding selection, where embeddings often lack modality-specific knowledge. In response, we propose MoE-Retriever, a novel framework inspired by the design principles of Sparse Mixture of Experts (SMoE). First, MoE-Retriever samples the relevant data from modality combinations, using a so-called supporting group to construct intra-modal inputs while incorporating inter-modal inputs. These inputs are then processed by Multi-Head Attention, after which the SMoE Router automatically selects the most relevant expert, i.e., the embedding candidate to be retrieved. Comprehensive experiments on both medical and general multimodal datasets demonstrate the robustness and generalizability of MoE-Retriever, marking a significant step forward in embedding retrieval methods for incomplete multimodal data.", "keywords": [ "Missing Multimodal Data", "Sparse MoE" ], "rating_list": [ 8, 6, 5, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "j964C6y92q", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Chengyi Cai", "gender": "Female", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" }, { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Jianzhong Qi", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Associate Professor" }, { "name": "Lei Feng", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" }, { "name": "Zesheng Ye", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Postdoc" } ], "n_formula": 58, "n_formula_1": 14, "n_ref_uni": 38, "n_ref": 98, "n_ref_all": 148, "n_fig": 11, "n_tab": 15, "L_tab": 4902, "n_element_tab": 569, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2301, "n_element_tab_1": 133, "formula_len_all": 5195, "formula_len_all_1": 1290, "len_all": 215079, "len_all_1": 74266, "len_abs": 1489, "len_title": 119, "len_sents": 68339, "len_sents_1": 30402, "n_sents": 534, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1528, "title": "Attribute-based Visual Reprogramming for Image Classification with CLIP", "abs": "*Visual reprogramming* (VR) reuses pre-trained vision models for downstream image classification tasks by adding trainable noise patterns to inputs. When applied to vision-language models (e.g., CLIP), existing VR approaches follow the same pipeline used in vision models (e.g., ResNet, ViT), where ground-truth class labels are inserted into fixed text templates to guide the optimization of VR patterns. This label-based approach, however, overlooks the rich information and diverse attribute-guided textual representations that CLIP can exploit, which may lead to the misclassification of samples. In this paper, we propose ***Attr**ibute-based **V**isual **R**eprogramming* (AttrVR) for CLIP, utilizing ***des**criptive **attr**ibutes* (DesAttrs) and ***dist**inctive **attr**ibutes* (DistAttrs), which respectively represent common and unique feature descriptions for different classes. Besides, as images of the same class may reflect different attributes after VR, AttrVR iteratively refines patterns using the $k$-nearest DesAttrs and DistAttrs for each image sample, enabling more dynamic and sample-specific optimization. Theoretically, AttrVR is shown to reduce intra-class variance and increase inter-class separation. Empirically, it achieves superior performance in 12 downstream tasks for both ViT-based and ResNet-based CLIP. The success of AttrVR facilitates more effective integration of VR from unimodal vision models into vision-language models. Our code is available at https://github.com/tmlr-group/AttrVR.", "keywords": [ "Visual Reprogramming", "Model Reprogramming", "Vision-Language Models", "Image Classification" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 4, 4, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "j8xJJkpZpw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Haicheng Liao", "gender": "Not Specified", "institution": "University of Macau", "country": "MO", "position": "PhD student" }, { "name": "Jiaxu Li", "gender": "unknown", "institution": "Central South University", "country": "CN", "position": "Undergrad student" }, { "name": "Jiayu Yang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Lijie Hu", "gender": "Female", "institution": "KAUST", "country": "SA", "position": "PhD student" }, { "name": "Songning Lai", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "TianlangXue", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Yu Huang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Yutao Yue", "gender": "unknown", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Associate Professor" }, { "name": "Zhangyi Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 42, "n_formula_1": 8, "n_ref_uni": 13, "n_ref": 20, "n_ref_all": 42, "n_fig": 7, "n_tab": 10, "L_tab": 3536, "n_element_tab": 350, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 308, "n_element_tab_1": 49, "formula_len_all": 2303, "formula_len_all_1": 438, "len_all": 178867, "len_all_1": 62507, "len_abs": 1527, "len_title": 113, "len_sents": 69146, "len_sents_1": 30016, "n_sents": 487, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1478, "title": "CAT: Concept-level backdoor ATtacks for Concept Bottleneck Models", "abs": "Despite the transformative impact of deep learning across multiple domains, the inherent opacity of these models has driven the development of Explainable Artificial Intelligence (XAI). Among these efforts, Concept Bottleneck Models (CBMs) have emerged as a key approach to improve interpretability by leveraging high-level semantic information. However, CBMs, like other machine learning models, are susceptible to security threats, particularly backdoor attacks, which can covertly manipulate model behaviors. Understanding that the community has not yet studied the concept level backdoor attack of CBM, because of \"Better the devil you know than the devil you don't know.\", we introduce CAT (Concept-level Backdoor ATtacks), a methodology that leverages the conceptual representations within CBMs to embed triggers during training, enabling controlled manipulation of model predictions at inference time. An enhanced attack pattern, CAT+, incorporates a correlation function to systematically select the most effective and stealthy concept triggers, thereby optimizing the attack's impact. Our comprehensive evaluation framework assesses both the attack success rate and stealthiness, demonstrating that CAT and CAT+ maintain high performance on clean data while achieving significant targeted effects on backdoored datasets. This work underscores the potential security risks associated with CBMs and provides a robust testing methodology for future security assessments.", "keywords": [ "Explainable AI; Backdoor attack; Concept bottleneck model" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "j8lqABLgub", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Alexandra Anna Lassota", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "", "position": "Postdoc" }, { "name": "Dorian Guyot", "gender": "unknown", "institution": "University of Fribourg", "country": "CH", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 14, "n_ref": 25, "n_ref_all": 36, "n_fig": 2, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 256, "formula_len_all_1": 115, "len_all": 106051, "len_all_1": 66129, "len_abs": 1019, "len_title": 111, "len_sents": 36924, "len_sents_1": 31892, "n_sents": 314, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1041, "title": "Minimalistic Predictions for Online Class Constraint Scheduling", "abs": "We consider online scheduling with class constraints. That is, we are given $m$ machines, each with $k$ class slots. Upon receiving a job $j$ with class $c_j$, an algorithm needs to allocate $j$ on some machine $i$. The goal is to minimize the makespan while not assigning more than $k$ different classes onto each machine.\nWhile the offline case is well understood and even (E)PTAS results are known [Jansen, Lassota, Maack SPAA'20, Chen Jansen Luo Zhang COCOA'16], the online case admits strong impossibility results in classical competitive analysis [Epstein, Lassota, Levin, Maack, Rohwedder STACS'22].\n\nWe overcome these daunting results by investigating the problem in a learning-augmented setting where an algorithm can access possibly erroneous predictions. We present new algorithms with competitive ratios independent of $m$ and tight lower bounds for several classical and problem-specific prediction models. We thereby give a structured overview of what additional information helps in the design of better scheduling algorithms.", "keywords": [ "Scheduling", "Class Constraints", "Predictions", "Online Algorithms" ], "rating_list": [ 5, 8, 8, 3 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 3, 2, 3, 4 ] }, { "paper_id": "j8WHjM9aMm", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Bedionita Soro", "gender": "Male", "institution": "KAIST AI", "country": "KR", "position": "PhD student" }, { "name": "Bruno Andreis", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Frank Hutter", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen & University of Freiburg", "country": "DE", "position": "Full Professor" }, { "name": "Hayeon Lee", "gender": "Female", "institution": "FAIR (Meta AI)", "country": "US", "position": "Postdoc" }, { "name": "Sung Ju Hwang", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Wonyong Jeong", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "Ph.D. student" }, { "name": "Song Chong", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 10, "n_ref_uni": 44, "n_ref": 86, "n_ref_all": 135, "n_fig": 8, "n_tab": 25, "L_tab": 13709, "n_element_tab": 1089, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 4102, "n_element_tab_1": 324, "formula_len_all": 1328, "formula_len_all_1": 501, "len_all": 250589, "len_all_1": 60593, "len_abs": 1651, "len_title": 97, "len_sents": 66451, "len_sents_1": 21516, "n_sents": 488, "n_sents_1": 158, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1659, "title": "Diffusion-based Neural Network Weights Generation", "abs": "Transfer learning, while fundamental to modern deep learning, faces critical limitations in model selection and storage requirements. To address these challenges, we introduce D2NWG: Diffusion-based Neural Network Weights Generation, a diffusion framework that learns to generate task-specific neural network weights by modeling the distribution of weights from diverse pretrained models. Our approach conditions the diffusion process on dataset features, task descriptions, and architectural specifications, enabling direct generation of task-specific weights without the need to store and search through extensive model collections.\nComprehensive experiments validate D2NWG's effectiveness across multiple scenarios: it matches or exceeds traditional pretrained models on in-distribution tasks, provides superior initialization for novel domains with faster convergence, and achieves a 6\\% improvement in few-shot learning scenarios. Through extensive ablation studies, we demonstrate that D2NWG's performance scales with the diversity and size of its training set.\nD2NWG shows particular promise in large language model applications, enabling efficient exploration of pretrained parameter spaces without the computational burden of traditional fine-tuning. In rigorous evaluations on the OpenLM leaderboard, our approach improved LLaMA-3-2-1B-instruct performance by 3\\% on challenging mathematical reasoning tasks, with a consistent 0.36\\% gain across all benchmarks. These results establish D2NWG as a versatile and powerful framework for neural network weight generation, offering a practical solution to the challenges of traditional transfer learning.", "keywords": [ "generative hyper-representation learning", "diffusion model", "neural network weights generation", "parameters generation", "hypernetworks" ], "rating_list": [ 8, 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 4, 1, 2, 3, 3 ], "contribution_list": [ 2, 3, 1, 2, 3 ], "confidence_list": [ 4, 5, 3, 4, 4 ] }, { "paper_id": "j8SwCtP2RG", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Hua Wei", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" }, { "name": "Jeff Dotson", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Associate Professor" }, { "name": "Longchao Da", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Porter Jenkins", "gender": "Male", "institution": "Brigham Young University", "country": "US", "position": "Assistant Professor" }, { "name": "Spencer Young", "gender": "Male", "institution": "Delicious AI, LLC", "country": "US", "position": "Machine Learning Engineer" } ], "n_formula": 18, "n_formula_1": 11, "n_ref_uni": 70, "n_ref": 150, "n_ref_all": 190, "n_fig": 14, "n_tab": 8, "L_tab": 4197, "n_element_tab": 170, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 2268, "n_element_tab_1": 98, "formula_len_all": 1527, "formula_len_all_1": 756, "len_all": 203198, "len_all_1": 66341, "len_abs": 1483, "len_title": 75, "len_sents": 59279, "len_sents_1": 30133, "n_sents": 450, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1511, "title": "Flexible Heteroscedastic Count Regression with Deep Double Poisson Networks", "abs": "Neural networks that can produce accurate, input-conditional uncertainty representations are critical for real-world applications. Recent progress on heteroscedastic $\\textit{continuous}$ regression has shown great promise for calibrated uncertainty quantification on complex tasks, like image regression. However, when these methods are applied to $\\textit{discrete}$ regression tasks, such as crowd counting, ratings prediction, or inventory estimation, they tend to produce predictive distributions with numerous pathologies. Moreover, discrete models based on the Generalized Linear Model (GLM) framework either cannot process complex input or are not fully heterosedastic. To address these issues we propose the Deep Double Poisson Network (DDPN). In contrast to networks trained to minimize Gaussian negative log likelihood (NLL), discrete network parameterizations (i.e., Poisson, Negative binomial), and GLMs, DDPN can produce discrete predictive distributions of arbitrary flexibility. Additionally, we propose a technique to tune the prioritization of mean fit and probabilistic calibration during training. We show DDPN 1) vastly outperforms existing discrete models; 2) meets or exceeds the accuracy and flexibility of networks trained with Gaussian NLL; 3) produces proper predictive distributions over discrete counts; and 4) exhibits superior out-of-distribution detection. DDPN can easily be applied to a variety of count regression datasets including tabular, image, point cloud, and text data.", "keywords": [ "Predictive uncertainty", "Heteroscedastic regression" ], "rating_list": [ 3, 5, 3, 8 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "j8HU5aNpd3", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Kaijian Zou", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Lu Wang", "gender": "Female", "institution": "Northeastern University", "country": "", "position": "Assistant Professor" }, { "name": "Muhammad Khalifa", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "EG", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 40, "n_ref": 90, "n_ref_all": 118, "n_fig": 13, "n_tab": 1, "L_tab": 1894, "n_element_tab": 240, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 538, "n_element_tab_1": 62, "formula_len_all": 352, "formula_len_all_1": 155, "len_all": 309962, "len_all_1": 58260, "len_abs": 1987, "len_title": 151, "len_sents": 46399, "len_sents_1": 28860, "n_sents": 297, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 1998, "title": "Retrieval or Global Context Understanding? On Many-Shot In-Context Learning for Long-Context Evaluation", "abs": "Language models (LMs) have demonstrated an improved capacity to handle long-context information, yet existing long-context benchmarks primarily measure LMs' retrieval abilities with extended inputs, e.g., pinpointing a short phrase from long-form text. \nTherefore, they may fall short when evaluating models' global context understanding capacity, such as synthesizing and reasoning over content across input to generate the response. \nIn this paper, we study long-context language model (LCLM) evaluation through many-shot in-context learning (ICL). Concretely, we identify the skills each ICL task requires, and examine models' long-context capabilities on them. We ask the first question: What types of ICL tasks benefit from additional demonstrations, and are these tasks effective at evaluating LCLMs? We find that classification and summarization tasks show notable performance improvements with additional demonstrations, while translation and reasoning tasks do not exhibit clear trends. This suggests the classification tasks predominantly test models' retrieval skills. Next, we ask: To what extent does each task require retrieval skills versus global context understanding from LCLMs? We develop metrics to categorize ICL tasks into two groups: (i) retrieval tasks that require strong retrieval ability to pinpoint relevant examples, and (ii) global context understanding tasks that necessitate a deeper comprehension of the full input. We find that not all datasets can effectively evaluate these long-context capabilities. \nTo address this gap, we introduce a new many-shot ICL benchmark, MANYICLBENCH, designed to characterize LCLMs' retrieval and global context understanding capabilities separately. We benchmark 11 open-weight LCLMs using MANYICLBENCH. We find that while state-of-the-art models demonstrate satisfactory performance up to 64k tokens in retrieval tasks, many models experience significant performance drops at only 16k tokens in global context understanding tasks.", "keywords": [ "long context evaluation", "many-shot ICL" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "j87C29mAZl", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Issei Sato", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Kevin Xu", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" } ], "n_formula": 142, "n_formula_1": 35, "n_ref_uni": 17, "n_ref": 37, "n_ref_all": 49, "n_fig": 3, "n_tab": 5, "L_tab": 1354, "n_element_tab": 164, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15572, "formula_len_all_1": 3252, "len_all": 193310, "len_all_1": 62266, "len_abs": 1195, "len_title": 151, "len_sents": 51885, "len_sents_1": 23800, "n_sents": 642, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 692, "title": "On Expressive Power of Looped Transformers: Theoretical Analysis and Enhancement via Timestep Encoding", "abs": "Looped Transformers offer advantages in parameter efficiency and Turing completeness. However, their expressive power for function approximation and approximation rate remains underexplored. In this paper, we establish approximation rates of Looped Transformers by defining the concept of the modulus of continuity for sequence-to-sequence functions. This reveals a limitation specific to\nthe looped architecture. That is, the analysis prompts us to incorporate scaling parameters for each loop, conditioned on timestep encoding. Experimental results demonstrate that increasing the number of loops enhances performance, with further gains achieved through the timestep encoding architecture.", "keywords": [ "Transformers", "Looped Transformers", "Expressive power", "Approximation Rate" ], "rating_list": [ 3, 8, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 1, 3, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "j83R1R3euh", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Amit Hogadi", "gender": "Male", "institution": "IISER Pune", "country": "IN", "position": "Full Professor" }, { "name": "Girish Kulkarni", "gender": "unknown", "institution": "Tata Institute of Fundamental Research", "country": "IN", "position": "Postdoc" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 17, "n_ref": 22, "n_ref_all": 29, "n_fig": 1, "n_tab": 4, "L_tab": 652, "n_element_tab": 37, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 652, "n_element_tab_1": 37, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 57068, "len_all_1": 57068, "len_abs": 1497, "len_title": 120, "len_sents": 18724, "len_sents_1": 18724, "n_sents": 152, "n_sents_1": 152, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1507, "title": "LEVERAGING LEARNING RATE GRADIENTS FOR AUTOMATIC LEARNING RATE SELECTION", "abs": "Selecting an optimal learning rate (LR) is crucial for training deep neural networks, significantly affecting both convergence speed and final model performance. Determining this optimal LR typically involves two key challenges: choosing an appropriate initial LR and selecting an LR scheduler for adjusting the LR during training. This paper focuses on the former challenge\u2014selecting the initial LR. Traditionally, this task relies on manual tuning or heuristic methods, often involving extensive trial-and-error or computationally expensive search strategies like grid search or random search. We propose an algorithm, Automatic Learning Rate Selection (ALRS), to find the initial LR without the need for manual intervention. ALRS leverages the gradient of the LR itself \u2014 a less explored approach in the field. ALRS is a computationally lightweight pre-training process that automatically selects the initial LR by iterative refinements using the LR gradient, specifically analyzing its sign information, combined with suitable search algorithms. This approach efficiently converges to the optimal LR in a stable and robust manner across various optimizers and network architectures. \n\nWe evaluate our technique on standard deep learning benchmarks, including MNIST with a CNN and CIFAR-10 and CIFAR-100 with ResNet-18, using both SGD and Adam optimizers. Our experiments demonstrate that the automatically determined LRs achieve performance comparable to manually tuned LRs and state-of-the-art results.", "keywords": [ "hyperparameter tuning", "learning rate gradient", "automatic learning rate selection" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 1, 3, 3 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "j80J5cyyqP", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Baihan Yang", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "", "position": "Undergrad student" }, { "name": "Hao Zhao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Huan-ang Gao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jingwei Zhao", "gender": "Male", "institution": "Xiaomi Corporation", "country": "CN", "position": "Researcher" }, { "name": "Kangliang Chen", "gender": "Male", "institution": "haomo.ai", "country": "CN", "position": "Researcher" }, { "name": "Weiyang Jin", "gender": "Male", "institution": "New York University", "country": "US", "position": "Intern" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 58, "n_ref": 116, "n_ref_all": 134, "n_fig": 6, "n_tab": 6, "L_tab": 3001, "n_element_tab": 279, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1583, "n_element_tab_1": 142, "formula_len_all": 507, "formula_len_all_1": 507, "len_all": 206246, "len_all_1": 73942, "len_abs": 1666, "len_title": 132, "len_sents": 61098, "len_sents_1": 34401, "n_sents": 442, "n_sents_1": 235, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1718, "title": "SPA: Enhancing 3D Multimodal LLMs with Mask-based Streamlining Preference Alignment", "abs": "Integrating 3D features into Large Language Models (LLMs) is a rapidly evolving field, with models like 3D-LLM, Point-Bind LLM, and PointLLM making notable strides. PointLLM, pre-trained and fine-tuned on the Objaverse dataset, enhances understanding by optimizing the projector, boosting resource efficiency and consistency. However, we observed a persistent bottleneck: increasing the LLM backbone size doesn't consistently improve performance. Preliminary experiments showed that enhancing the 3D encoder or extending fine-tuning alone failed to resolve this. While post-training partially addressed the issue, it required two stages and additional text sample generation, making it inefficient. To overcome this, we propose \\textbf{S}treamlining \\textbf{P}reference \\textbf{A}lignment \\textbf{(SPA)}, a post-training stage for MLLMs with 3D encoders. SPA leverages the 3D encoder\u2019s inductive bias through 3D-masking, ensuring robust output while preserving consistent differences. Unlike traditional post-training, SPA maximizes the encoder's spatial reasoning by increasing the probability gap between positive and negative logits. This approach eliminates redundant text generation, greatly enhancing resource efficiency and improving the overall alignment process. In addition, we identified evaluation issues in the existing benchmarks and conducted a re-benchmark, resulting in a more robust evaluation approach. The model combined with the SPA method as post-training stage successfully overcame the performance bottleneck and achieved better results across various evaluations on current scene-level and object-level benchmarks. Code is available at~\\url{https://anonymous.4open.science/r/3dmllm-dap-5A50}.", "keywords": [ "LLMs", "Reperentation learning", "MLLMs", "3D visual abilities" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 2, 4, 5, 4 ] }, { "paper_id": "j7yeq2sOj3", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Shirui Chen", "gender": "Male", "institution": "University of Washington", "country": "", "position": "PhD student" }, { "name": "Stefano Recanatesi", "gender": "Male", "institution": "University of Washington", "country": "", "position": "Postdoc" }, { "name": "Eric Shea-Brown", "gender": "Male", "institution": "University of Washington", "country": "", "position": "Full Professor" } ], "n_formula": 78, "n_formula_1": 25, "n_ref_uni": 36, "n_ref": 91, "n_ref_all": 148, "n_fig": 4, "n_tab": 1, "L_tab": 226, "n_element_tab": 3, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7088, "formula_len_all_1": 1486, "len_all": 150007, "len_all_1": 68039, "len_abs": 1307, "len_title": 136, "len_sents": 46331, "len_sents_1": 31940, "n_sents": 386, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1312, "title": "A simple connection from loss flatness to compressed representations in neural networks", "abs": "The generalization capacity of deep neural networks has been studied in a variety of ways, including at least two distinct categories of approaches: one based on the shape of the loss landscape in parameter space, and the other based on the structure of the representation manifold in feature space (that is, in the space of unit activities). Although these two approaches are related, they are rarely studied together explicitly. Here, we present an analysis that bridges this gap. We show that in the final phase of learning in deep neural networks, the compression of the manifold of neural representations correlates with the flatness of the loss around the minima explored by SGD. This correlation is predicted by a relatively simple mathematical relationship: a flatter loss corresponds to a lower upper bound on the compression metrics of neural representations. Our work builds upon the linear stability insight by Ma and Ying, deriving inequalities between various compression metrics and quantities involving sharpness. Empirically, our derived inequality predicts a consistently positive correlation between representation compression and loss sharpness in multiple experimental settings. Overall, we advance a dual perspective on generalization in neural networks in both parameter and feature space.", "keywords": [ "Sharpness", "Flatness", "compression", "dimensionality" ], "rating_list": [ 6, 6, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 1 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3, 3 ] }, { "paper_id": "j7oTk2nEoN", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Dani Lischinski", "gender": "Male", "institution": "Google", "country": "IL", "position": "Researcher" }, { "name": "Dvir Samuel", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "PhD student" }, { "name": "Matan Levy", "gender": "unknown", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "PhD student" }, { "name": "Nir Darshan", "gender": "Male", "institution": "OriginAI", "country": "IL", "position": "Principal Researcher" }, { "name": "Rami Ben-Ari", "gender": "Male", "institution": "OriginAI", "country": "IL", "position": "Principal Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 33, "n_ref": 67, "n_ref_all": 94, "n_fig": 4, "n_tab": 8, "L_tab": 1602, "n_element_tab": 179, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1153, "n_element_tab_1": 70, "formula_len_all": 117, "formula_len_all_1": 116, "len_all": 141637, "len_all_1": 61260, "len_abs": 1251, "len_title": 103, "len_sents": 45144, "len_sents_1": 31284, "n_sents": 300, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1255, "title": "Gray-Box Fine-Tuning for Single Backbone Domain Experts", "abs": "The emergence of foundational models has greatly improved performance across various downstream tasks, with fine-tuning often yielding even better results. However, existing fine-tuning approaches typically require access to model weights and layers, leading to challenges such as managing multiple model copies or inference pipelines, inefficiencies in edge device optimization, and concerns over proprietary rights, privacy, and exposure to unsafe model variants. In this paper, we address these challenges by exploring \"Gray-box\" fine-tuning approaches, where the model's architecture and weights remain hidden, allowing only gradient propagation. We introduce a novel yet simple and effective framework that adapts to new tasks using two lightweight learnable modules at the model's input and output. Additionally, we present a less restrictive variant that offers more entry points into the model, balancing performance with model exposure. We evaluate our approaches across several backbones on benchmarks for text-image alignment, text-video alignment, and sketch-image alignment. Our results demonstrate that, despite having limited access to the model, our Gray-box approaches achieve competitive performance with full-access fine-tuning methods.", "keywords": [ "Multi-modal Learning", "Vision-Language", "Foundation Models" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 2, 3, 4 ] }, { "paper_id": "j7kdXSrISM", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Jian Yang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Kepan Nan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Penghao Zhou", "gender": "Male", "institution": "ByteDance TikTok", "country": "CN", "position": "Researcher" }, { "name": "Tiehan Fan", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Xiang Li", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Ying Tai", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhenheng Yang", "gender": "Male", "institution": "Tiktok", "country": "US", "position": "Director of Research" }, { "name": "Zhijie Chen", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Xie Rui", "gender": "Male", "institution": "Southwest University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 21, "n_ref": 71, "n_ref_all": 102, "n_fig": 15, "n_tab": 10, "L_tab": 4420, "n_element_tab": 504, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 2515, "n_element_tab_1": 272, "formula_len_all": 381, "formula_len_all_1": 366, "len_all": 150033, "len_all_1": 63815, "len_abs": 1472, "len_title": 123, "len_sents": 39712, "len_sents_1": 26405, "n_sents": 353, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1481, "title": "OpenVid-1M: A Large-Scale High-Quality Dataset for Text-to-video Generation", "abs": "Text-to-video (T2V) generation has recently garnered significant attention thanks to the large multi-modality model Sora. However, T2V generation still faces two important challenges: 1) Lacking a precise open sourced high-quality dataset. The previously popular video datasets, e.g.WebVid-10M and Panda-70M, overly emphasized large scale, resulting in the inclusion of many low-quality videos and\nshort, imprecise captions. Therefore, it is challenging but crucial to collect a precise high-quality dataset while maintaining a scale of millions for T2V generation. 2) Ignoring to fully utilize textual information. Recent T2V methods have focused on vision transformers, using a simple cross attention module for video generation, which falls short of making full use of semantic information from text tokens. To address these issues, we introduce OpenVid-1M, a precise high-quality dataset with expressive captions. This open-scenario dataset contains over 1 million text-video pairs, facilitating research on T2V generation. Furthermore, we curate 433K 1080p videos from OpenVid-1M to create OpenVidHD-0.4M, advancing high-definition video generation. Additionally, we propose a novel Multi-modal Video Diffusion Transformer (MVDiT) capable of mining both structure information from visual tokens and semantic information from text tokens. Extensive experiments and ablation studies verify the superiority of OpenVid-1M over previous datasets and the effectiveness of our MVDiT.", "keywords": [ "Text-Video Dataset", "Video Generation" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "j7cyANIAxV", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Chenbin Zhang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "JIE XU", "gender": "Female", "institution": "Shanghai Artificial Intelligence Laboratory: Shanghai, CN", "country": "", "position": "Researcher" }, { "name": "Jiang Chuchu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "MS student" }, { "name": "Shaoting Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Full Professor" }, { "name": "Wen Chen", "gender": "Male", "institution": "Sensetime Research", "country": "", "position": "Researcher" }, { "name": "Zhiqiang Hu", "gender": "Male", "institution": "SenseTime Research", "country": "", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 67, "n_ref": 108, "n_ref_all": 139, "n_fig": 7, "n_tab": 7, "L_tab": 14312, "n_element_tab": 768, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 3960, "n_element_tab_1": 402, "formula_len_all": 1408, "formula_len_all_1": 1939, "len_all": 189423, "len_all_1": 59799, "len_abs": 1344, "len_title": 92, "len_sents": 40099, "len_sents_1": 23826, "n_sents": 299, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 107, "L_abs": 1351, "title": "Rethinking the generalization of drug target affinity prediction algorithms via similarity aware evaluation", "abs": "Drug-target binding affinity prediction is a fundamental task for drug discovery. It has been extensively explored in literature and promising results are reported. However, in this paper, we demonstrate that the results may be misleading and cannot be well generalized to real practice. The core observation is that the canonical randomized split of a test set in conventional evaluation leaves the test set dominated by samples with high similarity to the training set. The performance of models is severely degraded on samples with lower similarity to the training set but the drawback is highly overlooked in current evaluation. As a result, the performance can hardly be trusted when the model meets low-similarity samples in real practice. To address this problem, we propose a framework of similarity aware evaluation in which a novel split methodology is proposed to adapt to any desired distribution. This is achieved by a formulation of optimization problems which are approximately and efficiently solved by gradient descent. We perform extensive experiments across five representative methods in four datasets for two typical target evaluations and compare them with various counterpart methods. Results demonstrate that the proposed split methodology can significantly better fit desired distributions and guide the development of models.", "keywords": [ "Drug-Target Affinity Prediction", "Similarity-Aware Evaluation" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "j7b4mm7Ec9", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Ee-Chien Chang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" }, { "name": "Han Fang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Yupeng Qiu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" } ], "n_formula": 26, "n_formula_1": 9, "n_ref_uni": 30, "n_ref": 59, "n_ref_all": 103, "n_fig": 7, "n_tab": 20, "L_tab": 6902, "n_element_tab": 926, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1679, "n_element_tab_1": 192, "formula_len_all": 2699, "formula_len_all_1": 699, "len_all": 175723, "len_all_1": 66354, "len_abs": 1677, "len_title": 95, "len_sents": 56999, "len_sents_1": 30434, "n_sents": 419, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1685, "title": "Towards Lightweight Deep Watermarking Framework", "abs": "Deep learning-based watermarking models play a crucial role in copyright protection across various applications. However, many high-performance models are limited in practical deployment due to their large number of parameters. Meanwhile, the robustness and invisibility performance of existing lightweight models are unsatisfactory. This presents a pressing need for a watermarking model that combines lightweight capacity with satisfactory performance. Our research identifies a key reason that limits the performance of existing watermarking frameworks: a mismatch between commonly used decoding losses (e.g., mean squared error and binary cross-entropy loss) and the actual decoding goal, leading to parameter redundancy. We propose two innovative solutions: (1) Decoding-oriented surrogate loss (DO), which redesigns the loss function to mitigate the influence of decoding-irrelevant optimization directions; and (2) Detachable projection head (PH), which incorporates a detachable redundant module during training to handle these irrelevant directions and is discarded during inference. Additionally, we propose a novel watermarking framework comprising five submodules, allowing for independent parameter reduction in each component. Our proposed model achieves better efficiency, invisibility, and robustness while utilizing only 2.2\\% of the parameters compared to state-of-the-art frameworks. By improving efficiency while maintaining robust copyright protection, our model is well-suited for practical applications in resource-constrained environments. The DO and PH methods are designed to be plug-and-play, facilitating seamless integration into future lightweight models.", "keywords": [ "machine vision", "deep learning-based watermarking" ], "rating_list": [ 6, 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "j7ZWfqCYCY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chia-Mu Yu", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Associate Professor" }, { "name": "Ching-Chia Kao", "gender": "Male", "institution": "Department of computer science and informational engineering, National Taiwan University", "country": "TW", "position": "PhD student" }, { "name": "Chu-Song Chen", "gender": "Male", "institution": "Department of Computer Science and Informational Engineering, National Taiwan University", "country": "TW", "position": "Professor" }, { "name": "Chun-Shien Lu", "gender": "Male", "institution": "Academia Sinica", "country": "TW", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 6, "n_ref_uni": 63, "n_ref": 110, "n_ref_all": 127, "n_fig": 4, "n_tab": 6, "L_tab": 3942, "n_element_tab": 276, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 3001, "n_element_tab_1": 223, "formula_len_all": 1534, "formula_len_all_1": 540, "len_all": 194144, "len_all_1": 61362, "len_abs": 1449, "len_title": 161, "len_sents": 45962, "len_sents_1": 25322, "n_sents": 358, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 112, "L_abs": 961, "title": "Information-Theoretical Principled Trade-off between Jailbreakability and Stealthiness on Vision Language Models", "abs": "In recent years, Vision-Language Models (VLMs) have demonstrated significant advancements in artificial intelligence, transforming tasks across various domains. Despite their capabilities, these models are susceptible to jailbreak attacks, which can compromise their safety and reliability. This paper explores the trade-off between jailbreakability and stealthiness in VLMs, presenting a novel algorithm to detect non-stealthy jailbreak attacks and enhance model robustness. We introduce a stealthiness-aware jailbreak attack using diffusion models, highlighting the challenge of detecting AI-generated content. Our approach leverages Fano\u2019s inequality to elucidate the relationship between attack success rates and stealthiness scores, providing an explainable framework for evaluating these threats. Our contributions aim to fortify AI systems against sophisticated attacks, ensuring their outputs remain aligned with ethical standards and user expectations.", "keywords": [ "Jailbreak", "Vision-Language Models", "Security", "Information Theory" ], "rating_list": [ 6, 6, 5, 3 ], "soundness_list": [ 3, 2, 3, 1 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "j7OAzA9DQd", "primary_area": "learning on time series and dynamical systems", "n_author": 7, "author_info_list": [ { "name": "Aviad Susman", "gender": "unknown", "institution": "Icahn School of Medicine at Mount Sinai", "country": "US", "position": "Researcher" }, { "name": "Bino A. Varghese", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Associate Professor" }, { "name": "Gaurav Pandey", "gender": "Male", "institution": "Icahn School of Medicine at Mount Sinai", "country": "", "position": "Associate Professor" }, { "name": "Mohammad Al Olaimat", "gender": "Male", "institution": "University of North Texas", "country": "US", "position": "PhD student" }, { "name": "Nasim Sheikh-Bahaei", "gender": "unknown", "institution": "University of Southern California", "country": "", "position": "Assistant Professor" }, { "name": "Richard Yan Chak Li", "gender": "Male", "institution": "Icahn School of Medicine at Mount Sinai", "country": "", "position": "Bioinformatician" }, { "name": "Serdar Bozdag", "gender": "unknown", "institution": "University of North Texas", "country": "", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 41, "n_ref": 60, "n_ref_all": 82, "n_fig": 8, "n_tab": 1, "L_tab": 284, "n_element_tab": 18, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 284, "n_element_tab_1": 18, "formula_len_all": 842, "formula_len_all_1": 876, "len_all": 105177, "len_all_1": 57054, "len_abs": 1094, "len_title": 133, "len_sents": 31863, "len_sents_1": 28393, "n_sents": 195, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1098, "title": "Longitudinal Ensemble Integration for sequential classification with multimodal data", "abs": "Effectively modeling multimodal longitudinal data is a pressing need in various application areas, especially biomedicine. Despite this, few approaches exist in the literature for this problem, with most not adequately taking into account the multimodality of the data. In this study, we developed multiple configurations of a novel multimodal and longitudinal learning framework, Longitudinal Ensemble Integration (LEI), for sequential classification. We evaluated LEI\u2019s performance, and compared it against existing approaches, for the early detection of dementia, which is among the most studied multimodal sequential classification tasks. LEI outperformed these approaches due to its use of intermediate base predictions arising from the individual data modalities, which enabled their better integration over time. LEI\u2019s design also enabled the identification of features that were consistently important across time for the effective prediction of dementia-related diagnoses. Overall, our work demonstrates the potential of LEI for sequential classification from longitudinal multimodal data.", "keywords": [ "longitudinal multimodal data", "sequential classification", "deep learning", "LSTM", "heterogeneous ensembles", "dementia diagnosis" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "j6fsbpAllN", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Didi Zhu", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Fei Wu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Jing Su", "gender": "Female", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Xuwu Wang", "gender": "Female", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Zexi Li", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Ziyu Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "tao shen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 11, "n_ref_uni": 18, "n_ref": 33, "n_ref_all": 48, "n_fig": 9, "n_tab": 8, "L_tab": 4867, "n_element_tab": 592, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 4092, "n_element_tab_1": 421, "formula_len_all": 995, "formula_len_all_1": 382, "len_all": 143130, "len_all_1": 71850, "len_abs": 1688, "len_title": 153, "len_sents": 45972, "len_sents_1": 28280, "n_sents": 335, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1561, "title": "Merging LoRAs like Playing LEGO: Pushing the Modularity of LoRA to Extremes Through Rank-Wise Clustering", "abs": "Low-Rank Adaptation (LoRA) has emerged as a popular technique for fine-tuning large language models (LLMs) to various domains due to its modular design and widespread availability on platforms like Huggingface. This modularity has sparked interest in combining multiple LoRAs to significantly enhance LLM capabilities. However, existing methods for LoRA composition primarily focus on task-specific adaptations that require additional training, and current model merging techniques often fail to fully leverage LoRA's modular nature, leading to parameter interference and performance degradation.\nIn this paper, we explore the possibility of disassembling and reassembling multiple LoRAs at a finer granularity, much like assembling LEGO blocks. We introduce the concept of Minimal Semantic Units (MSUs), where the parameters corresponding to each rank in LoRA function as independent units. These MSUs exhibit properties such as permutation invariance and concatenation-summation equivalence, allowing for flexible combinations to form new LoRAs. Building on these insights, we propose the LoRA-LEGO framework. This framework conducts rank-wise parameter clustering by grouping MSUs from different LoRAs into $k$ clusters. The centroid of each cluster serves as a representative MSU, enabling the assembly of a merged LoRA with an adjusted rank of $k$. Additionally, we apply a dual reweighting strategy to optimize the scale of the merged LoRA. Experiments across various benchmarks demonstrate that our method outperforms existing approaches in LoRA merging.", "keywords": [ "Parameter Efficient Tuning", "LoRA", "Model Merging" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 2, 3, 4 ] }, { "paper_id": "j6GIg0peoS", "primary_area": "interpretability and explainable AI", "n_author": 12, "author_info_list": [ { "name": "Dimitrios Gunopulos", "gender": "Male", "institution": "University of Athens", "country": "GR", "position": "Full Professor" }, { "name": "Dimitrios Rontogiannis", "gender": "Male", "institution": "University of Athens", "country": "GR", "position": "MS student" }, { "name": "Dimitrios Tomaras", "gender": "Male", "institution": "Athens University of Economics and Business", "country": "GR", "position": "Postdoc" }, { "name": "Dimitris Fotakis", "gender": "Male", "institution": "Archimedes/Athena RC ", "country": "GR", "position": "Principal Researcher" }, { "name": "Dimitris Sacharidis", "gender": "Male", "institution": "Universit\u00e9 Libre de Bruxelles", "country": "BE", "position": "Assistant Professor" }, { "name": "Eleni Psaroudaki", "gender": "Female", "institution": "Athena Research Center", "country": "GR", "position": "Researcher" }, { "name": "Giorgos Giannopoulos", "gender": "Male", "institution": "Athena Research Center", "country": "GR", "position": "Researcher" }, { "name": "Ioannis Emiris", "gender": "Male", "institution": "University of Athens, Greece", "country": "GR", "position": "Full Professor" }, { "name": "Kleopatra Markou", "gender": "Female", "institution": "National and Kapodistrian University of Athens", "country": "GR", "position": "PhD student" }, { "name": "Konstantinos Tsopelas", "gender": "Male", "institution": "IMIS - \"Athena\" Research Center", "country": "GR", "position": "Researcher" }, { "name": "Loukas Kavouras", "gender": "Male", "institution": "IMIS - \"Athena\" Research Center", "country": "GR", "position": "Researcher" }, { "name": "Nikolaos Theologitis", "gender": "Male", "institution": "IMIS - \"Athena\" Research Center", "country": "GR", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 12, "n_ref_uni": 24, "n_ref": 68, "n_ref_all": 118, "n_fig": 10, "n_tab": 25, "L_tab": 25497, "n_element_tab": 2902, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2982, "n_element_tab_1": 320, "formula_len_all": 1833, "formula_len_all_1": 901, "len_all": 264795, "len_all_1": 73148, "len_abs": 1819, "len_title": 118, "len_sents": 84287, "len_sents_1": 32379, "n_sents": 635, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1828, "title": "GLANCE: Global Actions in a Nutshell for Counterfactual Explainability", "abs": "The widespread deployment of machine learning systems in critical real-world decision-making applications has highlighted the urgent need for counterfactual explainability methods that operate effectively. Global counterfactual explanations, expressed as actions to offer recourse, aim to provide succinct explanations and insights applicable to large population subgroups. Effectiveness is measured by the fraction of the population that is provided recourse, ensuring that the actions benefit as many individuals as possible. Keeping the cost of actions low ensures the proposed recourse actions remain practical and actionable. Limiting the number of actions that provide global counterfactuals is essential to maximize interpretability. The primary challenge, therefore, is balancing these trade-offs\u2014maximizing effectiveness, minimizing cost, while maintaining a small number of actions. We introduce GLANCE, a versatile and adaptive framework, comprising two algorithms, that allows the careful balancing of the trade-offs among the three key objectives, with the size objective functioning as a tunable parameter to keep the actions few and easy to interpret. C-GLANCE employs a clustering approach that considers both the feature space and the space of counterfactual actions, thereby accounting for the distribution of points in a way that aligns with the structure of the model. T-GLANCE provides additional features to enhance flexibility. It employs a tree-based approach, that allows users to specify split features, to build a decision tree with a single counterfactual action at each node that can be used as a subgroup policy. Our extensive experimental evaluation demonstrates that our method consistently shows greater robustness and performance compared to existing methods across various datasets and models.", "keywords": [ "Global Counterfactual Explainability", "Recourse", "Actions", "Clustering" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "j50c2tkQUu", "primary_area": "generative models", "n_author": 11, "author_info_list": [ { "name": "Chenfanfu Jiang", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "Associate Professor" }, { "name": "Hao Su", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Hongzhi Wu", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Kun Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Lei Lan", "gender": "Male", "institution": "University of Utah", "country": "US", "position": "Postdoc" }, { "name": "Shandian Zhe", "gender": "Male", "institution": ", University of Utah", "country": "US", "position": "Associate Professor" }, { "name": "Tianjia Shao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Xiang Feng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yin Yang", "gender": "Male", "institution": "University of Utah", "country": "US", "position": "Associate Professor" }, { "name": "Yintong Shang", "gender": "Female", "institution": "University of Utah", "country": "US", "position": "PhD student" }, { "name": "Yutao Feng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 15, "n_ref_uni": 51, "n_ref": 80, "n_ref_all": 97, "n_fig": 10, "n_tab": 2, "L_tab": 280, "n_element_tab": 22, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1129, "formula_len_all_1": 903, "len_all": 171533, "len_all_1": 70700, "len_abs": 2721, "len_title": 39, "len_sents": 40928, "len_sents_1": 33098, "n_sents": 352, "n_sents_1": 272, "venue": "ICLR2025", "decision": "withdraw", "L_title": 40, "L_abs": 1050, "title": "ElastoGen: 4D Generaetive Elastodynamics", "abs": "We present ElastoGen, a knowledge-driven model that generates physically accurate and coherent 4D elastodynamics. Instead of relying on petabyte-scale data-driven learning, ElastoGen leverages the principles of physics-in-the-loop and learns from established physical knowledge, such as partial differential equations and their numerical solutions. The core idea of ElastoGen is converting the global differential operator, corresponding to the nonlinear elastodynamic equations, into iterative local convolution-like operations, which naturally fit modern neural networks. Each network module is specifically designed to support this goal rather than functioning as a black box. As a result, ElastoGen is exceptionally lightweight in terms of both training requirements and network scale. Additionally, due to its alignment with physical procedures, ElastoGen efficiently generates accurate dynamics for a wide range of hyperelastic materials and can be easily integrated with upstream and downstream deep modules to enable end-to-end 4D generation.", "keywords": [ "generative model", "machine learning", "neural network architectures" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 1, 1 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "j4s6V1dl8m", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Harrison Edwards", "gender": "Male", "institution": "OpenAI", "country": "", "position": "Researcher" }, { "name": "Jan Hendrik Kirchner", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Jan Leike", "gender": "Male", "institution": "OpenAI", "country": "", "position": "Alignment Team Lead" }, { "name": "Nat McAleese", "gender": "unknown", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Yining Chen", "gender": "Female", "institution": "OpenAI", "country": "", "position": "Researcher" }, { "name": "Yuri Burda", "gender": "Male", "institution": "OpenAI", "country": "", "position": "Researcher" } ], "n_formula": 22, "n_formula_1": 5, "n_ref_uni": 51, "n_ref": 67, "n_ref_all": 125, "n_fig": 39, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 26, "n_tab_1": 2, "L_tab_1": 468, "n_element_tab_1": 7, "formula_len_all": 2253, "formula_len_all_1": 598, "len_all": 208085, "len_all_1": 61105, "len_abs": 1432, "len_title": 103, "len_sents": 65445, "len_sents_1": 27676, "n_sents": 581, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1439, "title": "Prover-Verifier Games improve legibility of LLM outputs", "abs": "One way to increase confidence in the outputs of Large Language Models (LLMs) is to support them with reasoning that is clear and easy to check \u2014 a property we call legibility. We study legibility in the context of solving grade-school math problems and show that optimizing chain-of-thought solutions only for answer correctness can make them less legible. To mitigate the loss in legibility, we propose a training algorithm inspired by Prover-Verifier Game from Anil et al. (2021). Our algorithm iteratively trains small verifiers to predict solution correctness, \u201chelpful\u201d provers to produce correct solutions that the verifier accepts, and \u201csneaky\u201d provers to produce incorrect solutions that fool the verifier. We find that the helpful prover\u2019s accuracy and the verifier\u2019s robustness to adversarial attacks increase over the course of training. Furthermore, we show that legibility training transfers to time-constrained humans tasked with verifying solution correctness. Over course of LLM training human accuracy increases when checking the helpful prover\u2019s solutions, and decreases when checking the sneaky prover\u2019s solutions. Hence, training for checkability by small verifiers is a plausible technique for increasing output legibility. Our results suggest legibility training against small verifiers as a practical avenue for increasing legibility of large LLMs to humans, and thus could help with alignment of superhuman models.", "keywords": [ "Prover-Verifier Games", "Large Language Models", "AI alignment", "Human Evaluation", "Scalable Oversight" ], "rating_list": [ 3, 5, 8, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 2, 2, 3, 4 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "j4gzziSUr0", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Bram Wouters", "gender": "unknown", "institution": "University of Amsterdam", "country": "", "position": "Researcher" }, { "name": "Cees Diks", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "Full Professor" }, { "name": "Floris Holstege", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Noud Van Giersbergen", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "Associate Professor" } ], "n_formula": 79, "n_formula_1": 22, "n_ref_uni": 56, "n_ref": 95, "n_ref_all": 129, "n_fig": 6, "n_tab": 3, "L_tab": 2018, "n_element_tab": 141, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7974, "formula_len_all_1": 1573, "len_all": 230834, "len_all_1": 66957, "len_abs": 862, "len_title": 121, "len_sents": 59356, "len_sents_1": 30354, "n_sents": 504, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 864, "title": "Optimizing importance weighting in the presence of sub-population shifts", "abs": "A distribution shift between the training and test data can severely harm performance of machine learning models. Importance weighting addresses this issue by assigning different weights to data points during training. We argue that existing heuristics for determining the weights are suboptimal, as they neglect the increase of the variance of the estimated model due to the limited sample size of the training data. We interpret the optimal weights in terms of a bias-variance trade-off, and propose a bi-level optimization procedure in which the weights and model parameters are optimized simultaneously. We apply this framework to existing importance weighting techniques for last-layer retraining of deep neural networks in the presence of sub-population shifts and show empirically that optimizing weights significantly improves generalization performance.", "keywords": [ "Importance weighting", "Distribution shift", "Group Robustness", "Bi-level optimization" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 2, 3, 2, 2 ] }, { "paper_id": "j4PXHRmA88", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Bin Pan", "gender": "Male", "institution": "Nankai University", "country": "", "position": "Associate Professor" }, { "name": "Guirong Xue", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Shiyu Shen", "gender": "Male", "institution": "Nankai University", "country": "", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 16, "n_ref_uni": 25, "n_ref": 35, "n_ref_all": 43, "n_fig": 3, "n_tab": 2, "L_tab": 656, "n_element_tab": 96, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 204, "n_element_tab_1": 32, "formula_len_all": 883, "formula_len_all_1": 953, "len_all": 85113, "len_all_1": 50753, "len_abs": 1141, "len_title": 118, "len_sents": 24945, "len_sents_1": 21740, "n_sents": 214, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1148, "title": "A Trajectory Probability Network for City-Scale Road Volume Prediction", "abs": "City-scale road volume prediction is a fundamental task in traffic management. However, the observation data are often incomplete and biased, posting a challenge for accurate prediction. Existing methods address this issue through interpolation techniques or manual priors, but they typically provide only a deterministic restoration, overlooking the influence of other potential scenarios. To overcome these limitations, we propose a novel neural network-based probabilistic model, the Trajectory Probability Network (TraPNet), which predicts traffic volume through the aggregation of the joint distribution of potential trajectories. TraPNet makes full use of current observations, historical data, and road network information to offer a comprehensive inference of road volumes. Unlike autoregressive methods, TraPNet makes predictions in a single step, substantially reducing computational time while maintaining high predictive accuracy. Experiments on real-world road networks demonstrate that TraPNet outperforms state-of-the-art methods, and can keep the advantage with only 20\\% observation ratio. The code will be made publicly available.", "keywords": [ "Data Mining", "Traffic Volume Prediction", "Learning on Graph" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 1, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "j4LITBSUjs", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bo Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Chenchen Jing", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" }, { "name": "Chunhua Shen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Cong Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Fengyun Rao", "gender": "Male", "institution": "WeChat, Tencent Inc.", "country": "CN", "position": "Researcher" }, { "name": "Hao Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Mingyu Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yizhou Zhou", "gender": "Male", "institution": "WeChat AI", "country": "CN", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 29, "n_ref": 68, "n_ref_all": 96, "n_fig": 11, "n_tab": 12, "L_tab": 6231, "n_element_tab": 406, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 960, "n_element_tab_1": 223, "formula_len_all": 950, "formula_len_all_1": 251, "len_all": 260888, "len_all_1": 64949, "len_abs": 1137, "len_title": 132, "len_sents": 86745, "len_sents_1": 28282, "n_sents": 764, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 83, "L_abs": 1144, "title": "PerturboLLaVA: Reducing Multimodal Hallucinations with Perturbative Visual Training", "abs": "This paper aims to address the challenge of hallucinations in Multimodal Large Language Models (MLLMs) particularly for dense image captioning tasks. To tackle the challenge, we identify the current lack of a metric that finely measures the caption quality in concept level. We hereby introduce HalFscore, a novel metric built upon the language graph and is designed to evaluate both the accuracy and completeness of dense captions at a\ngranular level. Additionally, we identify the root cause of hallucination as the model's over-reliance on its language prior. To address this, we propose PerturboLLaVA, which reduces the model's reliance on the language prior by incorporating adversarially perturbed text during training. This method enhances the model's focus on visual inputs, effectively reducing hallucinations and producing accurate, image-grounded descriptions without incurring additional computational overhead. PerturboLLaVA significantly improves the fidelity of generated captions, outperforming existing approaches in handling multimodal hallucinations and achieving improved performance across general multimodal benchmarks.", "keywords": [ "Multi-Modal Large Language Models", "Hallucinations Mitigation", "Hallucinations Evaluation", "Language Model Priors" ], "rating_list": [ 8, 8, 5, 8, 8 ], "soundness_list": [ 3, 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 3, 5, 4, 4, 4 ] }, { "paper_id": "j46zZVzVVQ", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Meng Fang", "gender": "Male", "institution": "University of Liverpool", "country": "", "position": "Assistant Professor" }, { "name": "Mykola Pechenizkiy", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Full Professor" }, { "name": "Tianyi Zhou", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Yucheng Yang", "gender": "unknown", "institution": "Eindhoven University of Technology", "country": "NL", "position": "PhD student" } ], "n_formula": 90, "n_formula_1": 25, "n_ref_uni": 32, "n_ref": 81, "n_ref_all": 124, "n_fig": 16, "n_tab": 12, "L_tab": 3422, "n_element_tab": 336, "n_fig_1": 13, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7969, "formula_len_all_1": 1739, "len_all": 220150, "len_all_1": 67438, "len_abs": 1305, "len_title": 128, "len_sents": 63509, "len_sents_1": 29210, "n_sents": 572, "n_sents_1": 267, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1375, "title": "Customizing Reinforcement Learning Agent with Multi-Objective Preference Control", "abs": "Practical reinforcement learning (RL) usually requires agents to be optimized for multiple potentially conflicting criteria, e.g. speed vs. safety. \nAlthough Multi-Objective RL (MORL) algorithms have been studied in previous works, their trained agents often lack precise controllability of the delicate trade-off among multiple objectives. Hence, the resulting agent is not versatile in aligning with customized requests from different users. \nTo bridge the gap, we develop ``Preference control (PC) RL'', which aims to train a meta-policy that takes user preference as input controlling the generation of a trajectory on the Pareto frontier adhering to the preference. To this end, we train a preference-conditioned meta-policy by our proposed preference-regularized MORL algorithm. The achieved meta-policy performs as a multi-objective optimizer that can produce user-desired solutions on the Pareto frontier. The proposed algorithm is analyzed and its convergence and controllability are theoretically justified. \nExperiments from discrete toy examples to higher-dimension robotic control tasks and experiments with more than two objectives are conducted to show its performance. In these experiments, PCRL-trained policies show significantly better controllability than existing approaches and can generate Pareto optimal solutions with better diversity and utilities.", "keywords": [ "reinforcement learning", "multi-objective optimization", "deep reinforcement learning" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "j3rxIH0M9H", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Baoxiong Jia", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Diwen Wan", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Gang Zeng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Junfeng Ni", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ruijie Lu", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Siyuan Huang", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Yixin Chen", "gender": "Male", "institution": "BIGAI", "country": "CN", "position": "Researcher" }, { "name": "Yu Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 46, "n_ref": 130, "n_ref_all": 151, "n_fig": 12, "n_tab": 4, "L_tab": 1299, "n_element_tab": 157, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 862, "n_element_tab_1": 116, "formula_len_all": 5141, "formula_len_all_1": 5404, "len_all": 189350, "len_all_1": 68417, "len_abs": 2355, "len_title": 116, "len_sents": 51111, "len_sents_1": 30367, "n_sents": 349, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1730, "title": "MOVIS: Enhancing Multi-Object Novel View Synthesis for Indoor Scenes", "abs": "Repurposing pre-trained diffusion models has been proven to be effective for NVS. However, these methods are mostly limited to a single object; directly applying such methods to compositional multi-object scenarios yields inferior results, especially incorrect object placement and inconsistent shape and appearance under novel views. How to enhance and systematically evaluate the cross-view consistency of such models remains under-explored. \nTo address this issue, we propose MOVIS to enhance the structural awareness of the view-conditioned diffusion model for multi-object NVS in terms of model inputs, auxiliary tasks, and training strategy. First, we inject structure-aware features, including depth and object mask, into the denoising U-Net to enhance the model's comprehension of object instances and their spatial relationships. Second, we introduce an auxiliary task requiring the model to simultaneously predict novel view object masks, further improving the model's capability in differentiating and placing objects. Finally, we conduct an in-depth analysis of the diffusion sampling process and carefully devise a structure-guided timestep sampling scheduler during training, which balances the learning of global object placement and fine-grained detail recovery. To systematically evaluate the plausibility of synthesized images, we propose to assess cross-view consistency and novel view object placement alongside existing image-level NVS metrics. \nExtensive experiments on challenging synthetic and realistic datasets demonstrate that our method exhibits strong generalization capabilities and produces consistent novel view synthesis, highlighting its potential to guide future 3D-aware multi-object NVS tasks.", "keywords": [ "Novel View Synthesis", "Multi objects" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "j3cBYvwyQT", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Pengtao Xie", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Ruiyi Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Sai Ashish Somayajula", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 95, "n_ref_all": 106, "n_fig": 2, "n_tab": 5, "L_tab": 1100, "n_element_tab": 91, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1100, "n_element_tab_1": 91, "formula_len_all": 907, "formula_len_all_1": 510, "len_all": 169496, "len_all_1": 61364, "len_abs": 1325, "len_title": 124, "len_sents": 44052, "len_sents_1": 27009, "n_sents": 311, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1330, "title": "TapWeight: Reweighting Pretraining Objectives for Task-Adaptive Pretraining", "abs": "Large-scale general domain pretraining followed by downstream-specific finetuning has become a predominant paradigm in machine learning. However, discrepancies between the pretraining and target domains can still lead to performance degradation in certain cases, underscoring the need for task-adaptive continued pretraining (TAP). TAP methods typically involve continued pretraining on task-specific unlabeled datasets or introducing additional unsupervised learning objectives to enhance model capabilities. While many TAP methods perform continued pretraining with multiple pretraining objectives, they often determine the tradeoff parameters between objectives manually, resulting in suboptimal outcomes and higher computational costs. In this paper, we propose TapWeight, a task-adaptive pretraining framework which automatically determines the optimal importance of each pretraining objective based on downstream feedback. TapWeight reweights each pretraining objective by solving a multi-level optimization problem. We applied TapWeight to both molecular property prediction and natural language processing tasks, significantly surpassing baseline methods. Experimental results validate the effectiveness and generalizability of TapWeight. Our code is publicly available at https://anonymous.4open.science/r/TapWeight-9A2E.", "keywords": [ "task-adaptive pretraining", "continued pretraining", "multi-level optimization", "hyper-parameter optimization" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "j3bKnEidtT", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Luca Guastoni", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Postdoc" }, { "name": "Nils Thuerey", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Associate Professor" }, { "name": "Patrick Schnell", "gender": "unknown", "institution": "Technical University Munich", "country": "", "position": "PhD student" } ], "n_formula": 66, "n_formula_1": 20, "n_ref_uni": 34, "n_ref": 43, "n_ref_all": 60, "n_fig": 12, "n_tab": 2, "L_tab": 412, "n_element_tab": 8, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4985, "formula_len_all_1": 1505, "len_all": 155289, "len_all_1": 60905, "len_abs": 1587, "len_title": 122, "len_sents": 56600, "len_sents_1": 28939, "n_sents": 481, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1594, "title": "Temporal Difference Learning: Why It Can Be Fast and How It Will Be Faster", "abs": "Temporal difference (TD) learning represents a fascinating paradox: It is the prime example of a divergent algorithm that has not vanished after its instability was proven. On the contrary, TD continues to thrive in reinforcement learning (RL), suggesting that it provides significant compensatory benefits. Empirical evidence supports this, as many RL tasks require substantial computational resources, and TD delivers a crucial speed advantage that makes these tasks solvable. However, it is limited to cases where the divergence issues are absent or negligible for unknown reasons. So far, the theoretical foundations behind the speed-up are also unclear. In our work, we address these shortcomings of TD by employing techniques for analyzing iterative schemes developed over the past century. Our analysis reveals that TD possesses a mechanism enabling efficient mapping into the smallest eigenspace\u2014an operation previously thought to necessitate costly matrix inversion. Notably, this effect is independent of the conditioning of the problem, making it particularly well-suited for RL tasks characterized by rapidly increasing condition numbers through delayed rewards. Our novel theoretical understanding allows us to develop a scalable algorithm that integrates TD\u2019s speed with the reliable convergence of gradient descent (GD). We additionally validate these improvements through a rigorous mathematical proof in two dimensions, as well as experiments on problems where TD and GD falter, providing valuable insights into the future of optimization techniques in artificial intelligence.", "keywords": [ "Reinforcement Learning", "Optimization" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "j3U6CJLhqw", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Baining Guo", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" }, { "name": "Shuyang Gu", "gender": "Male", "institution": "Tencent Hunyuan Research", "country": "CN", "position": "Researcher" }, { "name": "Tiankai Hang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "PhD student" }, { "name": "Xin Geng", "gender": "Male", "institution": "Southeast University, China", "country": "CN", "position": "Professor" } ], "n_formula": 38, "n_formula_1": 9, "n_ref_uni": 39, "n_ref": 89, "n_ref_all": 113, "n_fig": 7, "n_tab": 6, "L_tab": 437, "n_element_tab": 44, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 485, "n_element_tab_1": 49, "formula_len_all": 2579, "formula_len_all_1": 803, "len_all": 137033, "len_all_1": 65153, "len_abs": 1511, "len_title": 94, "len_sents": 38750, "len_sents_1": 29121, "n_sents": 329, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1535, "title": "Improved Noise Schedule for Diffusion Training", "abs": "Diffusion models have emerged as the de facto choice for generating high-quality visual content across multiple domains.\nHowever, training a single model to predict noise at multiple levels presents significant challenges, requiring numerous iterations and resulting in substantial computational costs.\nVarious approaches, such as loss weighting strategy design and architectural refinements, have been introduced to expedite convergence and improve model performance.\nIn this study, we propose a novel approach to design the noise schedule for enhancing the training of diffusion models. Our key insight is that the importance sampling of the logarithm of the Signal-to-Noise ratio ($\\log \\text{SNR}$), theoretically equivalent to a modified noise schedule, is particularly beneficial for training efficiency when increasing the sample frequency around $\\log \\text{SNR}=0$. This strategic sampling allows the model to focus on the critical transition point between signal dominance and noise dominance, potentially leading to more robust and accurate predictions.\nWe empirically demonstrate the superiority of our noise schedule over the standard cosine schedule.\nFurthermore, we highlight the advantages of our noise schedule design on the ImageNet benchmark, showing that the designed schedule consistently benefits different prediction targets.\nOur findings contribute to the ongoing efforts to optimize diffusion models, potentially paving the way for more efficient and effective training paradigms in the field of generative AI.", "keywords": [ "diffusion model", "effcient training", "noise schedule", "image generation" ], "rating_list": [ 5, 3, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 2, 1, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 2, 3, 4, 3, 2 ] }, { "paper_id": "j3R1qHvoSM", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Jiayi Ma", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Jinsheng Xiao", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Associate Professor" }, { "name": "Yang Jianfeng", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Associate Professor" }, { "name": "Yunhua Chen", "gender": "Female", "institution": "Guangdong University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Zhongyuan Wang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "\u5218\u6587\u535a", "gender": "Male", "institution": "Wuhan University", "country": "", "position": "Undergrad student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 24, "n_ref": 62, "n_ref_all": 76, "n_fig": 3, "n_tab": 3, "L_tab": 411, "n_element_tab": 36, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1990, "n_element_tab_1": 1, "formula_len_all": 414, "formula_len_all_1": 411, "len_all": 107462, "len_all_1": 60873, "len_abs": 1630, "len_title": 113, "len_sents": 34667, "len_sents_1": 29192, "n_sents": 221, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1635, "title": "LSTT:LONG SHORT-TERM TRANSFORMER FOR VIDEO SMALL OBJECT DETECTION", "abs": "Detecting small objects in video sequences is crucial, yet it poses significant challenges due to their limited visibility and dynamic nature, which complicates accurate identification and localization. Traditional methods often employ a uniform aggregation strategy across all frames, neglecting the unique spatiotemporal relationships of small objects, which results in insufficient feature extraction and diminished detection performance. This paper introduces a long short-term transformer network specifically designed for small object detection in videos. The model integrates features from both long-term and short-term frames: long-term frames capture global contextual information, enhancing the model\u2019s ability to represent background scenes, while short-term frames provide dynamic information closely related to the current detection frame, thereby improving the feature representation of small objects. A dynamic query generation module optimizes query generation based on the implicit motion relationships of targets in shortterm frames, adapting to the current video framework. Additionally, the network employs a progressive sampling strategy\u2014densely sampling short-term frames and sparsely sampling long-term frames\u2014to effectively model video scenes. A spatio-temporal alignment encoder further enhances pixel-level features by accounting for temporal and spatial transformations. Extensive experiments on the VisDrone-VID and UAVDT datasets demonstrate the method\u2019s effectiveness, with an average detection precision increase of 1.4% and 2.1%, respectively, highlighting its potential in small object video detection.", "keywords": [ "Long short-term ; transformer; Small Object Detection; Video Object" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "j3BWS9kDYm", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Fangzhou Hong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Lingni Ma", "gender": "Female", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Richard Newcombe", "gender": "Male", "institution": "Meta, Reality Labs Research", "country": "US", "position": "Principal Researcher" }, { "name": "Vladimir Guzov", "gender": "unknown", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Yuting Ye", "gender": "Female", "institution": "Meta", "country": "US", "position": "Research Scientist" }, { "name": "Ziwei Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Hyo Jin Kim", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Research Scientist" } ], "n_formula": 8, "n_formula_1": 10, "n_ref_uni": 52, "n_ref": 86, "n_ref_all": 108, "n_fig": 17, "n_tab": 7, "L_tab": 1845, "n_element_tab": 198, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1457, "n_element_tab_1": 165, "formula_len_all": 659, "formula_len_all_1": 649, "len_all": 182708, "len_all_1": 70080, "len_abs": 441, "len_title": 56, "len_sents": 47913, "len_sents_1": 29182, "n_sents": 483, "n_sents_1": 275, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1178, "title": "EgoLM: Multi-Modal Language Model of Egocentric Motions", "abs": "As wearable devices become more prevalent, understanding the user's motion is crucial for improving contextual AI systems. We introduce EgoLM, a versatile framework designed for egocentric motion understanding using multi-modal data. EgoLM integrates the rich contextual information from egocentric videos and motion sensors afforded by wearable devices. It also combines dense supervision signals from motion and language, leveraging the vast knowledge encoded in pre-trained large language models (LLMs). EgoLM models the joint distribution of egocentric motions and natural language using LLMs, conditioned on observations from egocentric videos and motion sensors. It unifies a range of motion understanding tasks, including motion narration from video or motion data, as well as motion generation from text or sparse sensor data. Unique to wearable devices, it also enables a novel task to generate text descriptions from sparse sensors. Through extensive experiments, we validate the effectiveness of EgoLM in addressing the challenges of under-constrained egocentric motion learning, and demonstrate its capability as a generalist model through a variety of applications.", "keywords": [ "Egocentric Motion Understanding", "Egocentric Motion Tracking", "Language Model" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "j1tSLYKwg8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 12, "author_info_list": [ { "name": "Chenxin An", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jiacheng Ye", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jiawei Han", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign (UIUC)", "country": "US", "position": "Full Professor" }, { "name": "Lin Zheng", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Lingpeng Kong", "gender": "Male", "institution": "Department of Computer Science, The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Peilin Zhao", "gender": "unknown", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Shansan Gong", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shivam Agarwal", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign ", "country": "", "position": "MS student" }, { "name": "Victoria W. Bi", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yizhe Zhang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Hao Peng", "gender": "Male", "institution": "Department of Computer Science, University of Illinois Urbana-Champaign", "country": "", "position": "Assistant Professor" }, { "name": "MUKAI LI", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 39, "n_formula_1": 18, "n_ref_uni": 70, "n_ref": 182, "n_ref_all": 219, "n_fig": 5, "n_tab": 14, "L_tab": 2067, "n_element_tab": 206, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1297, "n_element_tab_1": 155, "formula_len_all": 2780, "formula_len_all_1": 1242, "len_all": 271117, "len_all_1": 74048, "len_abs": 1325, "len_title": 123, "len_sents": 62809, "len_sents_1": 32894, "n_sents": 510, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1261, "title": "Scaling Diffusion Language Models via Adaptation from Autoregressive Models", "abs": "Diffusion Language Models (DLMs) have emerged as a promising new paradigm for text generative modeling, potentially addressing limitations of autoregressive (AR) models. However, current DLMs have been studied at a smaller scale compared to their AR counterparts and lack fair comparison on language modeling benchmarks. Additionally, training diffusion models from scratch at scale remains challenging. Given the prevalence of open-source AR language models, we propose adapting these models to build text diffusion models. We demonstrate connections between AR and diffusion modeling objectives and introduce a simple continual pre-training approach for training diffusion models. Through systematic evaluation on language modeling, reasoning, and commonsense benchmarks, we show that we can convert AR models ranging from 127M to 7B parameters (GPT2 and LLaMA) into diffusion models DiffuGPT and DiffuLLaMA, using less than 200B tokens for training. Our experimental results reveal that these models outperform earlier DLMs and are competitive with their AR counterparts. We release a suite of DLMs (127M-355M-7B) capable of generating fluent text, performing in-context learning, filling in the middle without prompt re-ordering, and following instructions.", "keywords": [ "diffusion models; diffusion language models; text diffusion models; discrete diffusion models" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "j1jtyGdD4O", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Bowen Jin", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Giri Narasimhan", "gender": "Male", "institution": ", Florida International University", "country": "", "position": "Full Professor" }, { "name": "Jiawei Han", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign (UIUC)", "country": "US", "position": "Full Professor" }, { "name": "Jimeng Shi", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 20, "n_ref_uni": 51, "n_ref": 90, "n_ref_all": 116, "n_fig": 7, "n_tab": 7, "L_tab": 3244, "n_element_tab": 583, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 4535, "n_element_tab_1": 394, "formula_len_all": 2206, "formula_len_all_1": 1388, "len_all": 156162, "len_all_1": 63858, "len_abs": 1778, "len_title": 130, "len_sents": 37439, "len_sents_1": 23158, "n_sents": 306, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1854, "title": "CoDiCast: Conditional Diffusion Model for Weather Prediction with Uncertainty Quantification", "abs": "Accurate weather forecasting is critical for science and society. Yet, existing methods have not demonstrated high accuracy, low uncertainty, and high computational efficiency simultaneously. On one hand, to quantify the uncertainty in weather predictions, the strategy of ensemble forecast (i.e., generating a set of diverse predictions) is often employed. However, traditional ensemble numerical weather prediction (NWP) is computationally intensive. On the other hand, even though most existing machine learning-based weather prediction (MLWP) approaches are efficient and accurate, they are deterministic and cannot capture the uncertainty of weather forecasting. To tackle these challenges, we propose $\\texttt{CoDiCast}$, a conditional diffusion model to generate accurate global weather prediction, while achieving uncertainty quantification and modest computational cost. The key idea behind the prediction task is to generate realistic weather scenarios at a $\\textit{future}$ time point, conditioned on observations from the $\\textit{recent past}$. Due to the probabilistic nature of diffusion models, they can be properly applied to capture the uncertainty of weather predictions. Therefore, we accomplish uncertainty quantifications by repeatedly sampling from stochastic Gaussian noise for each initial weather state and running the denoising process multiple times. Experimental results demonstrate that $\\texttt{CoDiCast}$ outperforms several existing MLWP methods in accuracy, and is faster than NWP models in the inference speed. $\\texttt{CoDiCast}$ can generate 3-day global weather forecasts, at 6-hour steps and $5.625^\\circ$ latitude-longitude resolutions, for over 5 variables, in about 12 minutes on a commodity A100 GPU machine with 80GB memory. The anonymous code is provided at \\url{https://anonymous.4open.science/r/CoDiCast/}.", "keywords": [ "Diffusion Model", "Weather Prediction", "Uncertainty Quantification" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "j1OucVFZMJ", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Bin Yang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Hongfan Gao", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "PhD student" }, { "name": "Jilin Hu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Ronghui Xu", "gender": "Female", "institution": "East China Normal University", "country": "CN", "position": "PhD student" }, { "name": "Wangmeng Shen", "gender": "Female", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Xiangfei Qiu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" } ], "n_formula": 24, "n_formula_1": 16, "n_ref_uni": 22, "n_ref": 33, "n_ref_all": 44, "n_fig": 7, "n_tab": 18, "L_tab": 5846, "n_element_tab": 627, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 2935, "n_element_tab_1": 278, "formula_len_all": 1404, "formula_len_all_1": 1106, "len_all": 160604, "len_all_1": 63273, "len_abs": 1204, "len_title": 159, "len_sents": 42065, "len_sents_1": 26433, "n_sents": 311, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1229, "title": "DiffImp: Efficient Diffusion Model for Probabilistic Time Series Imputation with Bidirectional Mamba Backbone", "abs": "Probabilistic time series imputation has been widely applied in real-world scenarios due to its ability to estimate uncertainty of imputation results. Meanwhile, denoising diffusion probabilistic models (DDPMs) have achieved great success in probabilistic time series imputation tasks with its power to model complex distributions. However, current DDPM-based probabilistic time series imputation methodologies are confronted with two types of challenges: 1) \\textit{ The backbone modules of the denoising parts are not capable of achieving sequence modeling with low time complexity.} 2) \\textit{ The architecture of denoising modules can not handle the inter-variable and bidirectional dependencies in the time series imputation problem effectively.} To address the first challenge, we integrate the computational efficient state space model, namely Mamba, as the backbone denosing module for DDPMs. To tackle the second challenge, we carefully devise several SSM-based blocks for bidirectional modeling and inter-variable relation understanding. Experimental results demonstrate that our approach can achieve state-of-the-art time series imputation results on multiple datasets, different missing scenarios and missing ratios.", "keywords": [ "diffusion", "time series impuation", "state space model" ], "rating_list": [ 8, 8, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 1, 1, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "j0sq9r3HFv", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Gaganpreet Jhajj", "gender": "Male", "institution": "Athabasca University", "country": "CA", "position": "MS student" }, { "name": "Gutierrez Carlos Enrique", "gender": "unknown", "institution": "Okinawa Institute of Science and Technology (OIST)", "country": "JP", "position": "Researcher" }, { "name": "Kenji Doya", "gender": "Male", "institution": "Okinawa Institute of Science and Technology Graduate University", "country": "JP", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 9, "n_ref": 12, "n_ref_all": 26, "n_fig": 3, "n_tab": 3, "L_tab": 1569, "n_element_tab": 72, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1569, "n_element_tab_1": 72, "formula_len_all": 48, "formula_len_all_1": 48, "len_all": 56270, "len_all_1": 56270, "len_abs": 1513, "len_title": 173, "len_sents": 15205, "len_sents_1": 15205, "n_sents": 114, "n_sents_1": 114, "venue": "ICLR2025", "decision": "Reject", "L_title": 124, "L_abs": 1513, "title": "Automated Parameter Extraction for Biologically Realistic Neural Networks: An Initial Exploration with Large Language Models", "abs": "In computational neuroscience, extracting parameters for constructing biologically realistic neural models is a resource-intensive task that requires continuous updates as new research emerges. This paper explores utilizing large language models (LLMs) in automating parameter extraction from scientific literature for biologically realistic neural models. We utilized open-source LLMs via Ollama to construct KGs, capturing parameters such as neuron morphology, synapse dynamics, and receptor properties. SNNBuilder \\cite{Gutierrez2022}, a framework for building spiking neural network (SNN) models, serves as a key validation example for our framework. However, the methodology we outline here can extend beyond SNNs and could applied to systematic modelling of the brain.By experimenting with different prompting strategies\u2014general extraction, in-context hints, and masked prompting\u2014we evaluated the ability of LLMs to autonomously extract relevant data and organize it within an expert-base or data-driven ontology, as well as to infer missing information for neural model construction. Additionally, we implemented retrieval-augmented generation (RAG) via LangChain to further improve the accuracy of parameter extraction through leveraging external knowledge sources. Analysis of the the generated KGs, demonstrated that LLMs, when guided by targeted prompts, can enhance the data-to-model process, paving the way for more efficient parameter extraction and model construction in computational neuroscience.", "keywords": [ "Large Language Models", "Knowledge Graphs", "Computational neuroscience", "Neural model construction" ], "rating_list": [ 1, 1, 3, 5 ], "soundness_list": [ 1, 1, 1, 2 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 1, 1, 1, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "j0YOeUZkjD", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Chi Zhang", "gender": "Female", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "PhD student" }, { "name": "Feifei Li", "gender": "unknown", "institution": "SSE", "country": "", "position": "PhD student" }, { "name": "Qi Song", "gender": "Female", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "PhD student" }, { "name": "Rui Huang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "hui shuai", "gender": "Male", "institution": "Nanjing University of Posts and Telecommunications", "country": "CN", "position": "Lecturer" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 30, "n_ref": 69, "n_ref_all": 80, "n_fig": 5, "n_tab": 3, "L_tab": 2377, "n_element_tab": 159, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2182, "n_element_tab_1": 142, "formula_len_all": 414, "formula_len_all_1": 413, "len_all": 106337, "len_all_1": 65277, "len_abs": 1148, "len_title": 133, "len_sents": 30979, "len_sents_1": 29399, "n_sents": 236, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1152, "title": "PoI: Pixel of Interest for Novel View Synthesis Assisted Scene Coordinate Regression", "abs": "The task of estimating camera poses can be enhanced through novel view synthesis techniques such as NeRF and Gaussian Splatting to increase the diversity and extension of training data. However, these techniques often produce rendered images with issues like blurring and ghosting, which compromise their reliability. These issues become particularly pronounced for Scene Coordinate Regression (SCR) methods, which estimate 3D coordinates at the pixel level. To mitigate the problems associated with unreliable rendered images, we introduce a novel filtering approach, which selectively extracts well-rendered pixels while discarding the inferior ones. The threshold of this filter is adaptively determined by the real-time reprojection loss recorded by the SCR models during training. Building on this filtering technique, we also develop a new strategy to improve scene coordinate regression using sparse inputs, drawing on successful applications of sparse input techniques in novel view synthesis. Our experimental results validate the effectiveness of our method, demonstrating the state-of-the-art performance on both indoor and outdoor datasets.", "keywords": [ "visual localization", "scene coordinate regresion", "novel view synthesis" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "j0KjevdhkH", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 8, "author_info_list": [ { "name": "Hanning Yuan", "gender": "Female", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Gui", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Full Professor" }, { "name": "Kaiyu Feng", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Lanting Fang", "gender": "Female", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Shanshan Feng", "gender": "Male", "institution": "The Agency for Science, Technology and Research", "country": "SG", "position": "Researcher" }, { "name": "Yulian Yang", "gender": "Female", "institution": "Southeast University", "country": "", "position": "MS student" }, { "name": "Wangkai", "gender": "Male", "institution": "Southeast University", "country": "", "position": "MS student" }, { "name": "Yew Soon Ong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" } ], "n_formula": 31, "n_formula_1": 26, "n_ref_uni": 42, "n_ref": 86, "n_ref_all": 110, "n_fig": 8, "n_tab": 10, "L_tab": 3689, "n_element_tab": 252, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1940, "n_element_tab_1": 104, "formula_len_all": 2487, "formula_len_all_1": 1687, "len_all": 168000, "len_all_1": 70944, "len_abs": 1197, "len_title": 128, "len_sents": 50267, "len_sents_1": 28972, "n_sents": 419, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1204, "title": "SIG: Self-Interpretable Graph Neural Network for Continuous-time Dynamic Graphs", "abs": "While graph neural networks have demonstrated potential across various applications, explaining their predictions on dynamic graphs remains largely under-explored. This paper introduces a new research task: self-interpretable GNNs for continuous-time dynamic graphs (CTDGs). We aim to predict future links within dynamic graphs while simultaneously providing causal explanations for these predictions. There are two key challenges: (1) capturing the underlying structural and temporal information that remains consistent across both independent and identically distributed (IID) and out-of-distribution (OOD) data, and (2) efficiently generating high-quality link prediction results and explanations. To tackle these challenges, we propose a novel causal inference model, namely the Independent and Confounded Causal Model (ICCM). ICCM is then integrated into a deep learning architecture that considers both effectiveness and efficiency. Extensive experiments demonstrate that our proposed model significantly outperforms existing methods across link prediction accuracy, explanation quality, and robustness to OOD data. Our code and datasets are anonymously released at https://github.com/2024SIG/SIG.", "keywords": [ "Self-Interpretable", "Graph neural network", "Continuous-time Dynamic Graph", "Causal inference" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 2, 4, 3 ] }, { "paper_id": "izzYucQBji", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Georgios Tziafas", "gender": "Male", "institution": "University of Groningen", "country": "NL", "position": "PhD student" }, { "name": "Hamidreza Kasaei", "gender": "Male", "institution": "University of Groningen", "country": "", "position": "Assistant Professor" }, { "name": "Yucheng XU", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "", "position": "MS student" }, { "name": "Zhibin Li", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "Associate Professor" } ], "n_formula": 19, "n_formula_1": 18, "n_ref_uni": 66, "n_ref": 217, "n_ref_all": 259, "n_fig": 18, "n_tab": 9, "L_tab": 3251, "n_element_tab": 503, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 4651, "n_element_tab_1": 438, "formula_len_all": 1179, "formula_len_all_1": 1063, "len_all": 256419, "len_all_1": 72835, "len_abs": 799, "len_title": 115, "len_sents": 69996, "len_sents_1": 28546, "n_sents": 449, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1778, "title": "View-Independent 3D Feature Distillation with Object-Centric Priors", "abs": "Grounding natural language to the physical world is a ubiquitous topic with a wide\nrange of applications in computer vision and robotics. Recently, 2D vision-language\nmodels such as CLIP have been widely popularized, due to their impressive capa-\nbilities for open-vocabulary grounding in 2D images. Subsequent works aim to\nelevate 2D CLIP features to 3D via feature distillation, but either learn neural fields\nthat are scene-specific and hence lack generalization, or focus on indoor room\nscan data that require access to multiple camera views, which is not practical in\nrobot manipulation scenarios. Additionally, related methods typically fuse features\nat pixel-level and assume that all camera views are equally informative. In this\nwork, we show that this approach leads to sub-optimal 3D features, both in terms\nof grounding accuracy, as well as segmentation crispness. To alleviate this, we\npropose a multi-view feature fusion strategy that employs object-centric priors to\neliminate uninformative views based on semantic information, and fuse features\nat object-level via instance segmentation masks. To distill our object-centric 3D\nfeatures, we generate a large-scale synthetic multi-view dataset of cluttered tabletop\nscenes, spawning 15k scenes from over 3300 unique object instances, which we\nmake publicly available. We show that our method reconstructs 3D CLIP features\nwith improved grounding capacity and spatial consistency, while doing so from\nsingle-view RGB-D, thus departing from the assumption of multiple camera views\nat test time. Finally, we show that our approach can generalize to novel tabletop\ndomains and be re-purposed for 3D instance segmentation without fine-tuning, and\ndemonstrate its utility for language-guided robotic grasping in clutter.", "keywords": [ "3D Language Grounding", "2D->3D Feature Distillation", "Large-Scale Multi-View Dataset" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "izjNI5bcOV", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 12, "author_info_list": [ { "name": "Ben Fei", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Hao Chen", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Junchao Gong", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "LEI BAI", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Shiqi Chen", "gender": "Male", "institution": "Shanghai Meteorological Service", "country": "CN", "position": "Researcher" }, { "name": "Wanli Ouyang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Full Professor" }, { "name": "Xiangyu Chen", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "PhD student" }, { "name": "Xiangyu Zhao", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Xiao-Ming Wu", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Associate Professor" }, { "name": "Yihao Liu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Zhiwang Zhou", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Undergrad student" }, { "name": "Wenlong Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 26, "n_ref": 86, "n_ref_all": 112, "n_fig": 9, "n_tab": 8, "L_tab": 3344, "n_element_tab": 424, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2092, "n_element_tab_1": 167, "formula_len_all": 427, "formula_len_all_1": 406, "len_all": 141396, "len_all_1": 57077, "len_abs": 1415, "len_title": 130, "len_sents": 41654, "len_sents_1": 25570, "n_sents": 330, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1428, "title": "WeatherGFM: Learning a Weather Generalist Foundation Model via In-context Learning", "abs": "The Earth's weather system involves intricate weather data modalities and diverse weather understanding tasks, which hold significant value to human life. \nExisting data-driven models focus on single weather understanding tasks (e.g., weather forecasting). \nWhile these models have achieved promising results, they fail to tackle various complex tasks within a single and unified model. \nMoreover, the paradigm that relies on limited real observations for a single scenario hinders the model's performance upper bound.\nInspired by the in-context learning paradigm from visual foundation models and large language models, in this paper, we introduce the first generalist weather generalist foundation model (WeatherGFM) to address weather understanding tasks in a unified manner. \nSpecifically, we first unify the representation and definition for diverse weather understanding tasks.\nSubsequently, we design weather prompt formats to handle different weather data modalities, including single, multiple, and temporal modalities. \nFinally, we adopt a visual prompting question-answering paradigm for the training of unified weather understanding tasks. \nExtensive experiments indicate that our WeatherGFM can effectively handle up to ten weather understanding tasks, including weather forecasting, super-resolution, weather image translation, and post-processing. Our method also showcases generalization ability on unseen tasks.", "keywords": [ "AI for Science", "Weather foundation model", "in-context learning" ], "rating_list": [ 10, 3, 6, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 4, 2, 2, 2 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "izETL3emSv", "primary_area": "optimization", "n_author": 8, "author_info_list": [ { "name": "Chao Qian", "gender": "Male", "institution": "Nanjing university", "country": "CN", "position": "Full Professor" }, { "name": "Ke Xue", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Siyuan Xu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xi Lin", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Undergrad student" }, { "name": "Yunqi Shi", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "Undergrad student" }, { "name": "Rongxi Tan", "gender": "unknown", "institution": "Nanjing University", "country": "", "position": "Undergrad student" }, { "name": "Ruotong Chen", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Undergrad student" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 56, "n_ref": 94, "n_ref_all": 109, "n_fig": 1, "n_tab": 10, "L_tab": 9024, "n_element_tab": 767, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 2728, "n_element_tab_1": 172, "formula_len_all": 59, "formula_len_all_1": 190, "len_all": 154129, "len_all_1": 60539, "len_abs": 1795, "len_title": 118, "len_sents": 35137, "len_sents_1": 27518, "n_sents": 244, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1804, "title": "BBOPlace-Bench: Benchmarking Black-Box Optimization for Chip Placement", "abs": "Chip placement is a crucial step in modern chip design, because it significantly impacts the subsequent process and the overall quality of the final chip. The application of black-box optimization (BBO) for chip placement has a history of several decades. Nevertheless, early attempts were hampered by immature problem modeling and inefficient algorithm design, resulting in suboptimal placement efficiency and quality compared to the more prevalent analytical methods. Recent advancements in problem modeling and BBO algorithm design have highlighted the effectiveness and efficiency of BBO, demonstrating its potential to achieve state-of-the-art results in chip placement. Despite these advancements, the field lacks a unified benchmark for thoroughly assessing various problem models and BBO algorithms. To address this gap, we propose BBOPlace-Bench, the first benchmark designed for evaluating and developing BBO algorithms specifically for chip placement tasks. BBOPlace-Bench first collects several popular tasks and standardizing their formats, thereby providing uniform and comprehensive information for optimization. Additionally, BBOPlace-Bench includes a wide range of existing BBO algorithms, including simulated annealing, evolutionary algorithms, evolution strategy, and Bayesian optimization, and evaluates their performance across different problem modelings (i.e., permutation, discrete, and mixed search spaces) using various metrics. Furthermore, BBOPlace-Bench offers a flexible framework that allows users to easily implement and test their unique algorithms. BBOPlace-Bench not only provides efficient solutions for chip placement but also expands the practical application scenarios for various BBO algorithms. The code for BBOPlace-Bench is available in the supplementary file.", "keywords": [ "Black-box optimization", "Bayesian optimization", "Evolutionary algorithm", "Chip placement", "EDA" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "izDiFGXn9B", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Cleber Zanchettin", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "Associate Professor" }, { "name": "Feihong Xu", "gender": "unknown", "institution": "Northwestern University, Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Luis A. Nunes Amaral", "gender": "Male", "institution": "Boston University, Boston University", "country": "", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 21, "n_ref_uni": 30, "n_ref": 53, "n_ref_all": 66, "n_fig": 7, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1386, "formula_len_all_1": 1334, "len_all": 102044, "len_all_1": 54136, "len_abs": 1011, "len_title": 137, "len_sents": 29543, "len_sents_1": 25365, "n_sents": 205, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1017, "title": "Benchmarking a well-calibrated measure of weight similarity of deep neural network models", "abs": "Deep learning approaches have revolutionized artificial intelligence, but model opacity and fragility remain significant challenges. The reason for these challenges, we believe, is a knowledge gap at the heart of the field --- the lack of well-calibrated metrics quantifying the similarity of the internal representations of models obtained using different architectures, training strategies, different checkpoints, or under different random initializations. While several metrics have been proposed, they are poorly calibrated and susceptible to manipulations and confounding factors, as well as being computationally intensive when probed with a large and diverse set of test samples. We report here an integration of chain normalization of weights and centered kernel alignment that, by focusing on weight similarity instead of activation similarity, overcomes most of the limitations of existing metrics. Our approach is sample-agnostic, symmetric in weight space, computationally efficient, and well-calibrated.", "keywords": [ "deep neural network", "weights similarity", "model interpretation", "computater vision" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "iylpeTI0Ql", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Chentao Cao", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "PhD student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Yang Liu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "Zhanke Zhou", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Visiting student" }, { "name": "Zhun Zhong", "gender": "Male", "institution": "University of Nottingham", "country": "GB", "position": "Assistant Professor" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 5, "n_ref_uni": 56, "n_ref": 127, "n_ref_all": 180, "n_fig": 9, "n_tab": 24, "L_tab": 28606, "n_element_tab": 2716, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 7450, "n_element_tab_1": 231, "formula_len_all": 767, "formula_len_all_1": 406, "len_all": 262124, "len_all_1": 75431, "len_abs": 2106, "len_title": 100, "len_sents": 58260, "len_sents_1": 29977, "n_sents": 471, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 2087, "title": "Noisy Test-Time Adaptation in Vision-Language Models", "abs": "Test-time adaptation (TTA) aims to address distribution shifts between source and target data by relying solely on target data during testing. In open-world scenarios, models often encounter noisy samples, i.e., samples outside the in-distribution (ID) label space. Leveraging the zero-shot capability of pre-trained vision-language models (VLMs), this paper introduces Zero-Shot Noisy TTA (ZS-NTTA), focusing on adapting the model to target data with noisy samples during test-time in a zero-shot manner. In the preliminary study, we reveal that existing TTA methods suffer from a severe performance decline under ZS-NTTA, often lagging behind even the frozen model. We conduct comprehensive experiments to analyze this phenomenon, revealing that the negative impact of unfiltered noisy data outweighs the benefits of clean data during model updating. In addition, as these methods adopt the adapting classifier to implement ID classification and noise detection sub-tasks, the ability of the model in both sub-tasks is largely hampered. Based on this analysis, we propose a novel framework that decouples the classifier and detector, focusing on developing an individual detector while keeping the classifier (including the backbone) frozen. Technically, we introduce the Adaptive Noise Detector (AdaND), which utilizes the frozen model\u2019s outputs as pseudo-labels to train a noise detector for detecting noisy samples effectively. To address clean data streams, we further inject Gaussian noise during adaptation, preventing the detector from misclassifying clean samples as noisy.\nBeyond the ZS-NTTA, AdaND can also improve the zero-shot out-of-distribution (ZS-OOD) detection ability of VLMs. Extensive experiments show that our method outperforms in both ZS-NTTA and ZS-OOD detection. On ImageNet, AdaND achieves a notable improvement of $8.32\\%$ in harmonic mean accuracy ($\\text{Acc}_\\text{H}$) for ZS-NTTA and $9.40\\%$ in FPR95 for ZS-OOD detection, compared to state-of-the-art methods. Importantly, AdaND is computationally efficient and comparable to the model-frozen method.", "keywords": [ "OOD Detection", "Test-Time Adaptation" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 5, 4 ] }, { "paper_id": "iyJOUELYir", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Andriy Mulyar", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Brandon Duderstadt", "gender": "Male", "institution": "Nomic AI", "country": "US", "position": "CEO" }, { "name": "Heng Ji", "gender": "Female", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Revanth Gangi Reddy", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Tarun Suresh", "gender": "Male", "institution": "Department of Computer Science", "country": "", "position": "Undergrad student" }, { "name": "Yifei Xu", "gender": "Male", "institution": "Lapis Lab", "country": "US", "position": "Researcher" }, { "name": "Zach Nussbaum", "gender": "Male", "institution": "Nomic AI", "country": "US", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 32, "n_ref": 79, "n_ref_all": 104, "n_fig": 3, "n_tab": 10, "L_tab": 2001, "n_element_tab": 241, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2040, "n_element_tab_1": 217, "formula_len_all": 429, "formula_len_all_1": 379, "len_all": 141312, "len_all_1": 68170, "len_abs": 1534, "len_title": 128, "len_sents": 37010, "len_sents_1": 30864, "n_sents": 246, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1540, "title": "CoRNStack: High-Quality Contrastive Data for Better Code Retrieval and Reranking", "abs": "Effective code retrieval plays a crucial role in advancing code generation, bug fixing, and software maintenance, particularly as software systems increase in complexity. While current code embedding models have demonstrated promise in retrieving code snippets for small-scale, well-defined tasks, they often underperform in more demanding real-world applications such as bug localization within GitHub repositories. We hypothesize that a key issue is their reliance on noisy and inconsistent datasets for training, which impedes their ability to generalize to more complex retrieval scenarios. To address these limitations, we introduce CoRNStack, a large-scale, high-quality contrastive training dataset for code that spans multiple programming languages. This dataset is curated using consistency filtering to eliminate noisy positives and is further enriched with mined hard negatives, thereby facilitating more effective learning. We demonstrate that contrastive training of embedding models using CoRNStack leads to state-of-the-art performance across a variety of code retrieval tasks. Furthermore, the dataset can be leveraged for training code reranking models, a largely underexplored area compared to text reranking. Our finetuned code reranking model significantly improves the ranking quality over the retrieved results. Finally, by employing our code retriever and reranker together, we demonstrate significant improvements in function localization for GitHub issues, an important\ncomponent of real-world software development.", "keywords": [ "code representation learning", "code re-ranking", "contrastive learning" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "iyGkoWP6nA", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "J. Nathan Kutz", "gender": "Male", "institution": "University of Washington", "country": "", "position": "Full Professor" }, { "name": "Jan P. Williams", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Liyao Gao", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 13, "n_ref_uni": 49, "n_ref": 68, "n_ref_all": 88, "n_fig": 15, "n_tab": 2, "L_tab": 67, "n_element_tab": 7, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 833, "formula_len_all_1": 926, "len_all": 120100, "len_all_1": 64700, "len_abs": 1600, "len_title": 131, "len_sents": 30976, "len_sents_1": 29562, "n_sents": 236, "n_sents_1": 230, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1610, "title": "Sparse identification of nonlinear dynamics with Shallow Recurrent Decoder Networks", "abs": "Spatio-temporal modeling of real-world data is a challenging problem as a result of inherent high-dimensionality, noisy observations, and expensive data collection procedures. In this paper, we present Sparse Identification of Nonlinear Dynamics with SHallow Recurrent Decoder networks (SINDy-SHRED) to jointly solve the sensing and model identification problems with simple implementation, efficient computation, and robust performance. SINDy-SHRED utilizes Gated Recurrent Units (GRUs) to model the temporal sequence of sensor measurements along with a shallow decoder network to reconstruct the full spatio-temporal field from the latent state space using only a few available sensors. Our proposed algorithm in\u0002troduces a SINDy-based regularization. Beginning with an arbitrary latent state space, the dynamics of the latent space progressively converges to a SINDy-class functional, provided the projection remains within the set. We conduct a system\u0002atic experimental study including synthetic PDE data, real-world sensor measure\u0002ments for sea surface temperature, and direct video data. With no explicit encoder, SINDy-SHRED allows for efficient training with minimal hyperparameter tuning and laptop-level computing. SINDy-SHRED demonstrates robust generalization in a variety of applications with minimal to no hyperparameter adjustments. Additionally, the interpretable SINDy model of latent state dynamics enables accurate long-term video predictions, achieving state-of-the-art performance and outperforming all baseline methods considered, including Convolutional LSTM, PredRNN, ResNet, and SimVP.", "keywords": [ "Sata-driven modeling", "scientific machine learning", "sparse identification of nonlinear dynamics", "AI in dynamic systems", "spatiotemporal modeling", "PDEs" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "ixoIAOcTSx", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Jinlong Shu", "gender": "Male", "institution": "Shanghai Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Kangyang Luo", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Lingfeng Qiao", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Meng Zhao", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Xiang Li", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Ding Zichen", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "di yin", "gender": "Male", "institution": "Tencent Youtu", "country": "CN", "position": "Senior Researcher" }, { "name": "WengZhenmin", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 11, "n_ref_uni": 54, "n_ref": 152, "n_ref_all": 216, "n_fig": 5, "n_tab": 27, "L_tab": 9316, "n_element_tab": 619, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1787, "n_element_tab_1": 414, "formula_len_all": 908, "formula_len_all_1": 1228, "len_all": 418051, "len_all_1": 79631, "len_abs": 8607, "len_title": 170, "len_sents": 148857, "len_sents_1": 34433, "n_sents": 1401, "n_sents_1": 259, "venue": "ICLR2025", "decision": "withdraw", "L_title": 122, "L_abs": 1208, "title": "Let's Be Self-generated via Step by Step: A Curriculum Learning Approach to Automated Reasoning with Large Language Models", "abs": "While Chain of Thought (CoT) prompting approaches have significantly consolidated the reasoning capabilities of large language models (LLMs), they still face limitations that require extensive human effort or have performance needs to be improved. Existing endeavors have focused on bridging these gaps; however, these approaches either hinge on external data and cannot completely eliminate manual effort, or they fall short in effectively directing LLMs to generate high-quality exemplary prompts. To address the said pitfalls, we propose a novel prompt approach for automatic reasoning named LBS3, inspired by curriculum learning which better reflects human learning habits. Specifically, LBS3 initially steers LLMs to recall easy-to-hard proxy queries that are pertinent to the target query. Following this, it invokes a progressive strategy that utilizes exemplary prompts stemmed from easy-proxy queries to direct LLMs in solving hard-proxy queries, enabling the high-quality of the proxy solutions. Finally, our extensive experiments in various reasoning-intensive tasks with varying open- and closed-source LLMs show that LBS3 achieves strongly competitive performance compared to the SOTA baselines.", "keywords": [ "Large Language Models", "Chain of Thought", "Automated Reasoning", "Curriculum Learning" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "ixdAVqjShn", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Fan Chen", "gender": "Male", "institution": "Bytedance", "country": "US", "position": "Researcher" }, { "name": "Lexin Tang", "gender": "Female", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Libo Zhang", "gender": "Male", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Longyin Wen", "gender": "Male", "institution": "Bytedance Inc.", "country": "US", "position": "Research Manager" }, { "name": "Xinyao Wang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yaojie Shen", "gender": "unknown", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Ying Zhou", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yulei Niu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 24, "n_ref": 74, "n_ref_all": 95, "n_fig": 10, "n_tab": 7, "L_tab": 2559, "n_element_tab": 378, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1546, "n_element_tab_1": 209, "formula_len_all": 814, "formula_len_all_1": 829, "len_all": 138577, "len_all_1": 58171, "len_abs": 1226, "len_title": 138, "len_sents": 41576, "len_sents_1": 25535, "n_sents": 291, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1265, "title": "AIPO: Agreement-Aware Iterative Preference Optimization for Length Exploitation Mitigation", "abs": "Direct Preference Optimization (DPO) is gaining popularity as an alternative to Proximal Policy Optimization (PPO) for aligning Large Language Models (LLMs). Recent research on aligning LLMs iteratively with synthetic or partially synthetic data has shown promising outcomes, facilitating the scalability of DPO training in both academic settings and proprietary models such as Llama 3. Despite its success, we observe that the issue of length exploitation in DPO becomes more pronounced during iterative preference optimization, with the severity escalating progressively with each iteration. This observation prompts an in-depth examination of iterative preference optimization with synthetic data. In this paper, we present our findings and analyses in building our iterative preference optimization pipeline. Specifically, we analyze the issue of length exploitation in this iterative process and propose a novel training objective for iterative preference optimization, namely \\textbf{A}greement-aware \\textbf{I}terative \\textbf{P}reference \\textbf{O}ptimization (AIPO). To demonstrate the effectiveness of our proposed method, we conduct extensive experiments and show that it achieves state-of-the-art performance on MT-Bench, AlpacaEval 2.0, and Arena-Hard.", "keywords": [ "Language Modeling", "Alignment" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 1 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "ixXQF1jz8f", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Carla Fabiana Chiasserini", "gender": "Female", "institution": "Polytechnic Institute of Turin", "country": "", "position": "Full Professor" }, { "name": "Francesco Malandrino", "gender": "unknown", "institution": "Consiglio Nazionale delle Ricerche", "country": "IT", "position": "Researcher" }, { "name": "Giuseppe Durisi", "gender": "Male", "institution": "Chalmers University", "country": "SE", "position": "Professor" }, { "name": "Jayadev Naram", "gender": "Male", "institution": "Chalmers University of Technology", "country": "SE", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 19, "n_ref": 29, "n_ref_all": 57, "n_fig": 13, "n_tab": 3, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2383, "formula_len_all_1": 1851, "len_all": 92054, "len_all_1": 59357, "len_abs": 1286, "len_title": 130, "len_sents": 34539, "len_sents_1": 28107, "n_sents": 224, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1292, "title": "Choose Before You Label: Efficient Node and Data Selection in Distributed Learning", "abs": "We consider one of the most relevant problems of distributed learning, i.e., the selection of the learning nodes to include in the training process as well as the selection of the samples from each of the learning nodes' local datasets, so as to make learning sustainable. Traditional approaches rely on pursuing a balanced label distribution, which requires label statistics from all datasets, including those not selected for learning. This may be costly and may raise privacy concerns. To cope with this issue, we aim at selecting few and small datasets. To this end, we propose a new metric, called loneliness, which is defined on unlabelled training samples. First, through both a theoretical and an experimental analysis, we show that loneliness is strongly linked with learning performance (i.e., test accuracy). Then, we propose a new node- and data-selection procedure, called Goldilocks, that uses loneliness to make its decisions. Our performance evaluation, including three state-of-the-art datasets and both centralized and federated learning, demonstrates that Goldilocks outperforms approaches based upon a balanced label distribution by providing over 70% accuracy improvement, in spite of using information that is both less sensitive privacy-wise and less onerous to obtain.", "keywords": [ "distributed learning", "node selection" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 2, 2, 1, 1 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 1, 2, 1, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "ixMBnOhFGd", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Hao Liu", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Hui Xiong", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Jinhui Ye", "gender": "Male", "institution": "Hong Kong University of Science and Technology(Guangzhou))", "country": "CN", "position": "MS student" }, { "name": "Lu Dai", "gender": "Female", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Yijie Xu", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 30, "n_ref": 81, "n_ref_all": 112, "n_fig": 7, "n_tab": 18, "L_tab": 8965, "n_element_tab": 1233, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1521, "n_element_tab_1": 200, "formula_len_all": 905, "formula_len_all_1": 824, "len_all": 234113, "len_all_1": 66795, "len_abs": 1235, "len_title": 82, "len_sents": 61217, "len_sents_1": 28879, "n_sents": 475, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 82, "L_abs": 1191, "title": "SePer: Measure Retrieval Utility Through The Lens Of Semantic Perplexity Reduction", "abs": "Large Language Models (LLMs) have demonstrated improved generation performance by incorporating externally retrieved knowledge, a process known as retrieval-augmented generation (RAG). Despite the potential of this approach, existing studies evaluate RAG effectiveness by 1) assessing retrieval and generation components jointly, which obscures retrieval's distinct contribution, or 2) examining retrievers using traditional metrics such as NDCG, which creates a gap in understanding retrieval's true utility in the overall generation process. To address the above limitations, in this work, we introduce an automatic evaluation method that measures retrieval quality through the lens of information gain within the RAG framework. Specifically, we propose Semantic Perplexity (SePer), a metric that captures the LLM's internal belief about the correctness of the retrieved information. We quantify the utility of retrieval by the extent to which it reduces semantic perplexity post-retrieval. Extensive experiments demonstrate that SePer not only aligns closely with human preferences but also offers a more precise and efficient evaluation of retrieval utility across diverse RAG scenarios.", "keywords": [ "information retrieval", "metric" ], "rating_list": [ 5, 8, 8, 8 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 4, 1, 3 ], "confidence_list": [ 3, 3, 2, 2 ] }, { "paper_id": "ix2yRWarPn", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Baoxiong Jia", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Junfeng Ni", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ruijie Lu", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Siyuan Huang", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Song-Chun Zhu", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Principal Researcher" }, { "name": "Yu Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 54, "n_ref": 116, "n_ref_all": 143, "n_fig": 7, "n_tab": 4, "L_tab": 5119, "n_element_tab": 905, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 4311, "n_element_tab_1": 767, "formula_len_all": 1061, "formula_len_all_1": 868, "len_all": 208867, "len_all_1": 82091, "len_abs": 4536, "len_title": 133, "len_sents": 45927, "len_sents_1": 31144, "n_sents": 330, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1289, "title": "Building Interactable Replicas of Complex Articulated Objects via Gaussian Splatting", "abs": "Building interactable replicas of articulated objects is a key challenge in computer vision. Existing methods often fail to effectively integrate information across different object states, limiting the accuracy of part-mesh reconstruction and part dynamics modeling, particularly for complex multi-part articulated objects. We introduce ArtGS, a novel approach that leverages 3D Gaussians as a flexible and efficient representation to address these issues. Our method incorporates canonical Gaussians with coarse-to-fine initialization and updates for aligning articulated part information across different object states, and employs a skinning-inspired part dynamics modeling module to improve both part-mesh reconstruction and articulation learning. Extensive experiments on both synthetic and real-world datasets, including a new benchmark for complex multi-part objects, demonstrate that ArtGS achieves state-of-the-art performance in joint parameter estimation and part mesh reconstruction. Our approach significantly improves reconstruction quality and efficiency, especially for multi-part articulated objects. Additionally, we provide comprehensive analyses of our design choices, validating the effectiveness of each component to highlight potential areas for future improvement.", "keywords": [ "articulated object modeling" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 4 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "iwVkB9zaVb", "primary_area": "foundation or frontier models, including LLMs", "n_author": 12, "author_info_list": [ { "name": "Bohan Li", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Chengquan Zhang", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Staff Software Engineer" }, { "name": "Dongliang Luo", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Errui Ding", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Director" }, { "name": "Gang Zhang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "senior engineer" }, { "name": "Liang Wu", "gender": "Male", "institution": "Baidu", "country": "", "position": "Researcher" }, { "name": "Pengyuan Lyu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Xiang Bai", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yingying Zhu", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Instructor" }, { "name": "Yuliang Liu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Ziyang Zhang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "linger Deng", "gender": "Female", "institution": "Dalian University of Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 25, "n_ref": 60, "n_ref_all": 80, "n_fig": 18, "n_tab": 4, "L_tab": 798, "n_element_tab": 109, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 543, "n_element_tab_1": 156, "formula_len_all": 113, "formula_len_all_1": 17, "len_all": 139804, "len_all_1": 63685, "len_abs": 1219, "len_title": 149, "len_sents": 40581, "len_sents_1": 27648, "n_sents": 370, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 1225, "title": "R-CoT: Reverse Chain-of-Thought Problem Generation for Geometric Reasoning in Large Multimodal Models", "abs": "Existing Large Multimodal Models (LMMs) struggle with mathematical geometric reasoning due to a lack of high-quality image-text paired data. Current geometric data generation approaches, which apply preset templates to generate geometric data or use Large Language Models (LLMs) to rephrase questions and answers (Q&A), unavoidably limit data accuracy and diversity. To synthesize higher-quality data, we propose a two-stage Reverse Chain-of-Thought (R-CoT) geometry problem generation pipeline. First, we introduce GeoChain to produce high-fidelity geometric images and corresponding descriptions highlighting relations among geometric elements. We then design a Reverse A&Q method that reasons step-by-step based on the descriptions and generates questions in reverse from the reasoning results. Experiments demonstrate that the proposed method brings significant and consistent improvements on multiple LMM baselines, achieving new performance records in the 2B, 7B, and 8B settings. Notably, R-CoT-8B significantly outperforms previous state-of-the-art open-source mathematical models by 16.6% on MathVista and 9.2% on GeoQA, while also surpassing the closed-source model GPT-4o by an average of 13% across both datasets.", "keywords": [ "Large Multimodal Models", "Mathematical Reasoning", "Geometric Reasoning" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "ivs0xU9Ebg", "primary_area": "optimization", "n_author": 9, "author_info_list": [ { "name": "Akang Wang", "gender": "Not Specified", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Researcher" }, { "name": "Bingheng Li", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Jianghua Wu", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Researcher" }, { "name": "Jiliang Tang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Full Professor" }, { "name": "Linxin Yang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "PhD student" }, { "name": "Ruoyu Sun", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Tian Ding", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Researcher" }, { "name": "Xiaodong Luo", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuyi Wang", "gender": "Male", "institution": "CRRC Zhuzhou Institute, China", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 35, "n_formula_1": 11, "n_ref_uni": 21, "n_ref": 37, "n_ref_all": 53, "n_fig": 3, "n_tab": 7, "L_tab": 2141, "n_element_tab": 294, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1811, "n_element_tab_1": 208, "formula_len_all": 3221, "formula_len_all_1": 1261, "len_all": 136616, "len_all_1": 69908, "len_abs": 1473, "len_title": 133, "len_sents": 45778, "len_sents_1": 30505, "n_sents": 365, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1484, "title": "An Efficient Unsupervised Framework for Convex Quadratic Programs via Deep Unrolling", "abs": "Quadratic programs (QPs) arise in various domains such as machine learning, finance, and control. \nRecently, learning-enhanced primal-dual hybrid gradient (PDHG) methods have shown great potential in addressing large-scale linear programs; however, this approach has not been extended to QPs.\nIn this work, we focus on unrolling \"PDQP\", a PDHG algorithm specialized for convex QPs. Specifically, we propose a neural network model called \"PDQP-net\" to learn optimal QP solutions. Theoretically, we demonstrate that a PDQP-net of polynomial size can align with the PDQP algorithm, returning optimal primal-dual solution pairs.\nWe propose an unsupervised method that incorporates KKT conditions into the loss function. Unlike the standard learning-to-optimize framework that requires optimization solutions generated by solvers, our unsupervised method adjusts the network weights directly from the evaluation of the primal-dual gap.\nThis method has two benefits over supervised learning: first, it helps generate better primal-dual gap since the primal-dual gap is in the objective function; second, it does not require solvers. \nWe show that PDQP-net trained in this unsupervised manner can effectively approximate optimal QP solutions.\nExtensive numerical experiments confirm our findings, indicating that using PDQP-net predictions to warm-start PDQP can achieve up to 45% acceleration on QP instances. \nMoreover, it achieves 14% to 31% acceleration on out-of-distribution instances.", "keywords": [ "Quadratic programming", "Learning to optimize", "Deep unrolling", "Primal-Dual Hybrid Gradient", "Unsupervised learning" ], "rating_list": [ 3, 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 1, 3 ], "presentation_list": [ 1, 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "ivXe7J6U0k", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Haolan Guo", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Undergrad student" }, { "name": "Linwei Tao", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Minjing Dong", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Chang Xu", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 24, "n_ref": 36, "n_ref_all": 61, "n_fig": 14, "n_tab": 5, "L_tab": 768, "n_element_tab": 119, "n_fig_1": 14, "n_tab_1": 2, "L_tab_1": 1546, "n_element_tab_1": 136, "formula_len_all": 902, "formula_len_all_1": 625, "len_all": 102956, "len_all_1": 65007, "len_abs": 1510, "len_title": 149, "len_sents": 33260, "len_sents_1": 29763, "n_sents": 239, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1431, "title": "Consistency Calibration: Improving Uncertainty Calibration via Consistency among Perturbed Neighbors", "abs": "Calibration is crucial in deep learning applications, especially in fields like healthcare and autonomous driving, where accurate confidence estimates are vital for decision-making. However, deep neural networks often suffer from miscalibration, with reliability diagrams and Expected Calibration Error (ECE) being the only standard perspective for evaluating calibration performance. In this paper, we introduce the concept of consistency as an alternative perspective on model calibration, inspired by uncertainty estimation literature in large language models (LLMs). We highlight its advantages over the traditional reliability-based view. Building on this concept, we propose a post-hoc calibration method called Consistency Calibration (CC), which adjusts confidence based on the model's consistency across perturbed inputs. CC is particularly effective in locally uncertainty estimation, as it requires no additional data samples or label information, instead generating input perturbations directly from the source data. Moreover, we show that performing perturbations at the logit level significantly improves computational efficiency. We validate the effectiveness of CC through extensive comparisons with various post-hoc and training-time calibration methods, demonstrating state-of-the-art performance on standard datasets such as CIFAR-10, CIFAR-100, and ImageNet, as well as on long-tailed datasets like ImageNet-LT.", "keywords": [ "Uncertainty Estimation", "Confidence Calibration" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 3, 5 ] }, { "paper_id": "ivLD3sCDcG", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Chenlin Fu", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "PhD student" }, { "name": "Jiecong Chen", "gender": "unknown", "institution": "Shenzhen University", "country": "CN", "position": "MS student" }, { "name": "Yingying Zhu", "gender": "Female", "institution": "Shenzhen University", "country": "CN", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 12, "n_ref_uni": 16, "n_ref": 37, "n_ref_all": 63, "n_fig": 10, "n_tab": 8, "L_tab": 651, "n_element_tab": 110, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 3165, "n_element_tab_1": 123, "formula_len_all": 1932, "formula_len_all_1": 955, "len_all": 129883, "len_all_1": 58535, "len_abs": 2113, "len_title": 98, "len_sents": 34943, "len_sents_1": 23939, "n_sents": 290, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 1701, "title": "Line2Rbox: Line-supervised Oriented Object Detection", "abs": "Oriented object detection is crucial for complex scenes such as aerial images and industrial inspection, providing precise delineation by minimizing background interference. Recently, the weakly-supervised oriented object detection has gaining attention due to its cost-effectiveness. However, the majority of existing weakly-supervised methods are either point-supervised or HBox-supervised, which presents a challenge in achieving an optimal balance between annotation cost and detection performance. In response, we introduce a novel form of line annotation, which is intermediate between point-level and plane-level annotation. Based on this, we present L2RBox, an end-to-end anchor-free detector that is the first line-supervised method for oriented object detection. The fundamental objective of the L2RBox is to utilise line labels for the completion of label assignment and the calculation of loss. In particular, the line is mapped to the corresponding circle domain, which is then used to select training samples and calculate the center-ness target by the minimum circumscribed rectangle of the circle in the direction of the line. The regression loss that we propose is designed to support the line as an optimisation target. It comprises four components, namely scale loss $L_s$, height loss $L_h$, position loss $L_p$ and angle loss $L_a$.\nExtensive experimentation on DOTA-v1.0 and DIOR-R has demonstrated that our L2RBox significantly outperforms point-supervised methods, while requiring only a slight increase in labeling costs. It is also noteworthy that the proposed approach also demonstrates a slight performance advantage over the fully-supervised FCOS in certain categories.", "keywords": [ "Line-supervised", "Oriented Object Detection", "Remote sensing images", "Weakly-supervised" ], "rating_list": [ 5, 6, 5, 8, 6 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 5, 4, 5, 5, 4 ] }, { "paper_id": "iv6Sw43IMl", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Mark Bourgeault", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CA", "position": "Principal Researcher" }, { "name": "Mashfiq Shahriar Zaman", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "MS student" }, { "name": "Matthew Guzdial", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Assistant Professor" }, { "name": "Mohan Sai Singamsetti", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CA", "position": "Researcher" }, { "name": "Peter Chun", "gender": "unknown", "institution": "University of Alberta", "country": "CA", "position": "adjunct professor " } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 38, "n_ref": 60, "n_ref_all": 83, "n_fig": 2, "n_tab": 6, "L_tab": 1335, "n_element_tab": 104, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 2, "n_element_tab_1": 1, "formula_len_all": 440, "formula_len_all_1": 441, "len_all": 110882, "len_all_1": 60850, "len_abs": 1203, "len_title": 130, "len_sents": 35091, "len_sents_1": 30776, "n_sents": 232, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1207, "title": "Parameter Monte Carlo Tree Search: Efficient Chip Placement via Transfer Learning", "abs": "Automated chip placement is an important problem in enhancing the design and effectiveness of computer chips. Previous approaches have employed transfer learning to adapt knowledge obtained via machine learning from one chip placement task to another. However, these approaches have not notably reduced the necessary chip design time, which is crucial for minimizing the total resource utilization. This paper introduces a novel transfer learning approach called Parameter Monte Carlo Tree Search (PMCTS) that utilizes MCTS to transfer the learned knowledge from deep reinforcement learning (RL) models trained on one chip design task to another chip design by searching directly over the model parameters to generate models for efficient chip placement. We employ MCTS to escape the local optima reached by training from scratch and fine-tuning methods. We evaluate our methodology on four chip design tasks from the literature: Ariane, Ariane133, IBM01, and IBM02. Through extensive experiments, we find that our approach can generate models for optimized chip placement in less time than training from scratch and fine-tuning methods when transferring knowledge from complex chip designs to simpler ones.", "keywords": [ "Transfer learning", "chip placement", "reinforcement learning", "Monte Carlo Tree Search", "finetuning" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 1, 1, 2 ], "contribution_list": [ 2, 1, 1, 2, 1 ], "confidence_list": [ 2, 3, 4, 3, 3 ] }, { "paper_id": "iv1TpRCJeK", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Daksh Dobhal", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "MS student" }, { "name": "Daniel Richard Bramblett", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Rushang Karia", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Siddharth Srivastava", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "Associate Professor" } ], "n_formula": 27, "n_formula_1": 7, "n_ref_uni": 33, "n_ref": 77, "n_ref_all": 136, "n_fig": 25, "n_tab": 8, "L_tab": 2523, "n_element_tab": 181, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 84, "n_element_tab_1": 4, "formula_len_all": 1316, "formula_len_all_1": 641, "len_all": 262343, "len_all_1": 82719, "len_abs": 1086, "len_title": 81, "len_sents": 83224, "len_sents_1": 33866, "n_sents": 759, "n_sents_1": 300, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 108, "L_abs": 1172, "title": "$\\forall$uto$\\exists$$\\lor\\!\\land$L: Autonomous Evaluation of LLMs for Truth Maintenance and Reasoning Tasks", "abs": "This paper presents $\\forall$uto$\\exists$$\\lor\\\\!\\land$L, a novel benchmark for scaling Large Language Model (LLM) assessment in formal tasks with clear notions of correctness, such as truth maintenance in translation and logical reasoning. $\\forall$uto$\\exists$$\\lor\\\\!\\land$L is the first benchmarking paradigm that offers several key advantages necessary for scaling objective evaluation of LLMs without human labeling: (a) ability to evaluate LLMs of increasing sophistication by auto-generating tasks at different levels of difficulty; (b) auto-generation of ground truth that eliminates dependence on expensive and time consuming human annotation; (c) the use of automatically generated, randomized datasets that mitigate the ability of successive LLMs to overfit to static datasets used in many contemporary benchmarks. Empirical analysis shows that an LLM's performance on $\\forall$uto$\\exists$$\\lor\\\\!\\land$L is highly indicative of its performance on a diverse array of other benchmarks focusing on translation and reasoning tasks, making it a valuable autonomous evaluation paradigm in settings where hand-curated datasets can be hard to obtain and/or update.", "keywords": [ "Large Language Models", "Logical Reasoning", "Autoformalization", "Informalization", "Formal Translation", "Truth Maintenance" ], "rating_list": [ 5, 6, 8 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 3, 4 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "iuxaCU3DI7", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Brian R Quaranto", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "", "position": "Instructor" }, { "name": "Chenhui Xu", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Dancheng Liu", "gender": "Male", "institution": "University at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Ishan Mishra", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Researcher" }, { "name": "Jiajie Li", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Jinjun Xiong", "gender": "unknown", "institution": "State University of New York at Buffalo", "country": "US", "position": "Professor" }, { "name": "Peter C W Kim", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Full Professor" }, { "name": "Ruiyang Qin", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 38, "n_ref": 54, "n_ref_all": 71, "n_fig": 4, "n_tab": 7, "L_tab": 1104, "n_element_tab": 112, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 934, "n_element_tab_1": 90, "formula_len_all": 7, "formula_len_all_1": 7, "len_all": 130172, "len_all_1": 66614, "len_abs": 1071, "len_title": 125, "len_sents": 36435, "len_sents_1": 30475, "n_sents": 285, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 77, "L_abs": 933, "title": "Recognize Any Surgical Object: Unleashing the Power of Weakly-Supervised Data", "abs": "We present RASO, a foundation model designed to Recognize Any Surgical Object, offering robust open-set recognition capabilities across a broad range of surgical procedures and object classes, in both surgical images and videos. RASO leverages a novel weakly-supervised learning framework that generates tag-image-text pairs automatically from large-scale unannotated surgical lecture videos, significantly reducing the need for manual annotations. Our scalable data generation pipeline gathers 2,200 surgical procedures and produces 3.6 million tag annotations across 2,066 unique surgical tags. Our experiments show that RASO achieves improvements of 2.9 mAP, 4.5 mAP, 10.6 mAP, and 7.2 mAP on four standard surgical benchmarks respectively in zero-shot settings, and surpasses state-of-the-art models in supervised surgical action recognition tasks. We will open-source our code, model, and dataset to facilitate further research.", "keywords": [ "Image Recognition", "Vision-Language Pretraining", "Image Tagging", "Medical Imaging", "Surgery" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "iuqprf3GuR", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Andrew Reddie", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Associate Professor" }, { "name": "Janet Napolitano", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Leah Walker", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Ritwik Gupta", "gender": "Male", "institution": "Defense Innovation Unit", "country": "US", "position": "Principal Researcher" }, { "name": "Rodolfo Corona", "gender": "unknown", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Stephanie Fu", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Suzanne Petryk", "gender": "Female", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "trevor darrell", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department", "country": "US", "position": "Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 78, "n_ref": 92, "n_ref_all": 104, "n_fig": 2, "n_tab": 2, "L_tab": 550, "n_element_tab": 23, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 656, "n_element_tab_1": 28, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 212872, "len_all_1": 61581, "len_abs": 926, "len_title": 128, "len_sents": 40190, "len_sents_1": 32049, "n_sents": 238, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 929, "title": "Data-Centric AI Governance: Addressing the Limitations of Model-Focused Policies", "abs": "Current regulations on powerful AI capabilities are narrowly focused on \"foundation\" or \"frontier\" models. However, these terms are vague and inconsistently defined, leading to an unstable foundation for governance efforts. Critically, policy debates often fail to consider the data used with these models, despite the clear link between data and model performance. Even (relatively) \"small\" models that fall outside the typical definitions of foundation and frontier models can achieve equivalent outcomes when exposed to sufficiently specific datasets. In this work, we illustrate the importance of considering dataset size and content as essential factors in assessing the risks posed by models both today and in the future. More broadly, we emphasize the risk posed by over-regulating reactively and provide a path towards careful, quantitative evaluation of capabilities that can lead to a simplified regulatory environment.", "keywords": [ "ai policy", "data-centric ai" ], "rating_list": [ 6, 3, 6 ], "soundness_list": [ 3, 1, 2 ], "presentation_list": [ 2, 2, 4 ], "contribution_list": [ 4, 1, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "iucVyVC8jQ", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Hong Qian", "gender": "unknown", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Jing-Wen Yang", "gender": "Male", "institution": "Game AI Center", "country": "CN", "position": "Principal Researcher" }, { "name": "Shuo Liu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Intern" }, { "name": "Yimeng Liu", "gender": "Female", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuanhao Liu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" } ], "n_formula": 20, "n_formula_1": 15, "n_ref_uni": 27, "n_ref": 65, "n_ref_all": 103, "n_fig": 12, "n_tab": 7, "L_tab": 2989, "n_element_tab": 360, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 3317, "n_element_tab_1": 246, "formula_len_all": 1262, "formula_len_all_1": 898, "len_all": 194670, "len_all_1": 71723, "len_abs": 1826, "len_title": 130, "len_sents": 60352, "len_sents_1": 30226, "n_sents": 501, "n_sents_1": 255, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1836, "title": "A Dual-Fusion Cognitive Diagnosis Framework for Open Student Learning Environments", "abs": "Cognitive diagnosis model (CDM) is a fundamental and upstream component in intelligent education. It aims to infer students' mastery levels based on historical response logs. However, existing CDMs usually follow the ID-based embedding paradigm, which could often diminish the effectiveness of CDMs in open student learning environments. This is mainly because they can hardly directly infer new students' mastery levels or utilize new exercises or knowledge without retraining. Textual semantic information, due to its unified feature space and easy accessibility, can help alleviate this issue. Unfortunately, directly incorporating semantic information may not benefit CDMs, since it does not capture response-relevant features and thus discards the individual characteristics of each student. To this end, this paper proposes a dual-fusion cognitive diagnosis framework (DFCD) to address the challenge of aligning two different modalities, i.e., textual semantic features and response-relevant features. Specifically, in DFCD, we first propose the exercise-refiner and concept-refiner to make the exercises and knowledge concepts more coherent and reasonable via large language models. Then, DFCD encodes the refined features using text embedding models to obtain the semantic information. For response-related features, we propose a novel response matrix to fully incorporate the information within the response logs. Finally, DFCD designs a dual-fusion module to merge the two modal features. The ultimate representations possess the capability of inference in open student learning environments and can be also plugged in existing CDMs. Extensive experiments across real-world datasets show that DFCD achieves superior performance by integrating different modalities and strong adaptability in open student learning environments.", "keywords": [ "Cognitive Diagnosis", "Open Student Learning Environments", "Inductive Learning", "Intelligent Education" ], "rating_list": [ 3, 3, 1, 6 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "iuTyzHnvP4", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Huibin Xiao", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "MS student" }, { "name": "Linlin Shen", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "Full Professor" }, { "name": "Qiufu Li", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 36, "n_formula_1": 35, "n_ref_uni": 36, "n_ref": 78, "n_ref_all": 120, "n_fig": 3, "n_tab": 3, "L_tab": 2699, "n_element_tab": 151, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1982, "n_element_tab_1": 126, "formula_len_all": 2410, "formula_len_all_1": 2320, "len_all": 118967, "len_all_1": 71774, "len_abs": 1407, "len_title": 83, "len_sents": 34646, "len_sents_1": 29333, "n_sents": 213, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1414, "title": "BCE vs. CE in Deep Feature Learning", "abs": "When training classification models, it expects that the leaned features are compact within classes, and can well separate different classes. As a dominant loss function to train classification models, the minimization of CE (Cross-entropy) loss can maximize the compactness and distinctiveness, i.e., reaching neural collapse. The recently published works show that BCE (Binary CE) loss performs also well in multi-class tasks. In this paper, we compare BCE and CE in the context of deep feature learning. For the first time, we prove that BCE can also maximize the intra-class compactness and inter-class distinctiveness when reaching its minimum, i.e., leading to neural collapse. We point out that CE measures the relative values of decision scores in the model training, implicitly enhancing the feature properties by classifying samples one-by-one. In contrast, BCE measures the absolute values of decision scores and adjust the positive/negative decision scores across all samples to uniform high/low levels. Meanwhile, the classifier bias in BCE presents a substantial constraint on the samples' decision scores. Thereby, BCE explicitly enhances the feature properties in the training. The experimental results are aligned with above analysis, and show that BCE consistently and significantly improve the classification performance and leads to better compactness and distinctiveness among sample features.", "keywords": [ "BCE", "CE", "neural collapse", "decision score", "classifier bias" ], "rating_list": [ 6, 3, 8 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 1, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "itwyfJilM5", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Hung Son Nguyen", "gender": "Male", "institution": "University of Warsaw", "country": "PL", "position": "Associate Professor" }, { "name": "Toan Van Tran", "gender": "unknown", "institution": "Hanoi University of Science and Technology", "country": "VN", "position": "Undergrad student" } ], "n_formula": 49, "n_formula_1": 26, "n_ref_uni": 25, "n_ref": 50, "n_ref_all": 58, "n_fig": 3, "n_tab": 2, "L_tab": 809, "n_element_tab": 95, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 1050, "n_element_tab_1": 12, "formula_len_all": 5542, "formula_len_all_1": 2018, "len_all": 146522, "len_all_1": 71223, "len_abs": 808, "len_title": 104, "len_sents": 43155, "len_sents_1": 32208, "n_sents": 341, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 811, "title": "Graph Scattering Networks with Adaptive Diffusion Kernels", "abs": "Scattering networks are deep convolutional architectures that use predefined wavelets for feature extraction and representation. They have proven effective for classification tasks, especially when training data is scarce, where traditional deep learning methods struggle. In this work, we introduce and develop a mathematically sound framework for applying adaptive kernels to diffusion wavelets in graph scattering networks. Stability guarantees with respect to input perturbations are provided. A specific construction of adaptive kernels is presented and applied with continuous diffusion to perform graph classification tasks on benchmark datasets. Our model consistently outperforms traditional graph scattering networks with predefined wavelets, both in scenarios with limited and abundant training data.", "keywords": [ "graph neural networks", "graph scattering transform", "deep learning", "stability" ], "rating_list": [ 3, 6, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 1, 2, 1 ], "confidence_list": [ 4, 3, 3, 3, 4 ] }, { "paper_id": "ittdt7tKND", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Chunyu Li", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Guoqing Wang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jiayi Ma", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Linfeng Tang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Yixuan Yuan", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 24, "n_ref_uni": 40, "n_ref": 110, "n_ref_all": 134, "n_fig": 13, "n_tab": 6, "L_tab": 2601, "n_element_tab": 243, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 4288, "n_element_tab_1": 256, "formula_len_all": 1443, "formula_len_all_1": 1301, "len_all": 160769, "len_all_1": 74971, "len_abs": 1269, "len_title": 128, "len_sents": 42789, "len_sents_1": 27884, "n_sents": 302, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1280, "title": "DSPFusion: Degradation and Semantic Prior Dual-guided Framework for Image Fusion", "abs": "Existing fusion methods are tailored for high-quality images but struggle with degraded images captured under harsh circumstances, thus limiting the practical potential of image fusion. In this work, we present a Degradation and Semantic Prior dual-guided framework for degraded image Fusion (DSPFusion), utilizing degradation priors and high-quality scene semantic priors restored via diffusion models to guide both information recovery and fusion in a unified model. In specific, it first individually extracts modality-specific degradation priors and jointly captures comprehensive low-quality semantic priors from cascaded source images. Subsequently, a diffusion model is developed to iteratively restore high-quality semantic priors in a compact latent space, enabling our method to be over $200 \\times$ faster than mainstream diffusion model-based image fusion schemes. Finally, the degradation priors and high-quality semantic priors are employed to guide information enhancement and aggregation via the dual-prior guidance and prior-guided fusion modules. Extensive experiments demonstrate that DSPFusion mitigates most typical degradations while integrating complementary context with minimal computational cost, greatly broadening the application scope of image fusion.", "keywords": [ "Image fusion", "mluti-modal fusion", "image restoration", "infrared" ], "rating_list": [ 5, 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 5, 4, 5 ] }, { "paper_id": "it1zuzw7KF", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Aoming Liu", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "PhD student" }, { "name": "Siqi Wang", "gender": "Female", "institution": "Boston University", "country": "US", "position": "PhD student" }, { "name": "Bryan Allen Plummer", "gender": "Male", "institution": "Boston University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 6, "n_ref_uni": 68, "n_ref": 201, "n_ref_all": 223, "n_fig": 7, "n_tab": 11, "L_tab": 4466, "n_element_tab": 459, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1198, "n_element_tab_1": 13, "formula_len_all": 2353, "formula_len_all_1": 253, "len_all": 185774, "len_all_1": 58009, "len_abs": 1677, "len_title": 91, "len_sents": 39571, "len_sents_1": 27465, "n_sents": 295, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1667, "title": "In-N-Out: Robustness to In-Domain Noise and Out-of-Domain Generalization", "abs": "Training on real-world data is challenging due to its complex nature, where data is often noisy and may require understanding diverse domains. Methods focused on Learning with Noisy Labels (LNL) may help with noise, but they often assume no domain shifts. In contrast, approaches for Domain Generalization (DG) could help with domain shifts, but these methods either consider label noise but prioritize out-of-domain (OOD) gains at the cost of in-domain (ID) performance, or they try to balance ID and OOD performance, but do not consider label noise at all. Thus, no work explores the combined challenge of balancing ID and OOD performance in the presence of label noise, limiting their impact. We refer to this challenging task as In-N-Out, and this work provides the first exploration of its unique properties. We find that combining the settings explored in LNL and DG poses new challenges not present in either task alone, and thus, requires direct study. Our findings are based on a study comprised of three real-world datasets and one synthesized noise dataset, where we benchmark a dozen unique methods along with many combinations that are sampled from both the LNL and DG literature. We find that the best method for each setting varies, with older DG and LNL methods often beating the SOTA. A significant challenge we identified stems from unbalanced noise sources and domain-specific sensitivities, which makes using traditional LNL sample selection strategies that often perform well on LNL benchmarks a challenge. While we show this can be mitigated when domain labels are available, we find that LNL and DG regularization methods often perform better.", "keywords": [ "learning with noisy labels", "domain generalization" ], "rating_list": [ 1, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "ispjankYab", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 9, "author_info_list": [ { "name": "Emmanuel Bengio", "gender": "Male", "institution": "Recursion", "country": "", "position": "Researcher" }, { "name": "Joseph D Viviano", "gender": "Male", "institution": "Mila Quebec AI Institute ", "country": "", "position": "Machine Learning Scientist " }, { "name": "Lena Nehale Ezzine", "gender": "Female", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "PhD student" }, { "name": "Micha\u0142 Koziarski", "gender": "unknown", "institution": "The Hospital for Sick Children (SickKids)", "country": "CA", "position": "Principal Researcher" }, { "name": "Moksh Jain", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "PhD student" }, { "name": "Oussama Boussif", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "PhD student" }, { "name": "Rim Assouel", "gender": "Female", "institution": "Facebook", "country": "CA", "position": "Intern" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" }, { "name": "Kolya Malkin", "gender": "Male", "institution": "Edinburgh University, University of Edinburgh", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 20, "n_formula_1": 11, "n_ref_uni": 49, "n_ref": 101, "n_ref_all": 123, "n_fig": 15, "n_tab": 4, "L_tab": 4159, "n_element_tab": 412, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 211, "n_element_tab_1": 29, "formula_len_all": 1450, "formula_len_all_1": 1100, "len_all": 204541, "len_all_1": 60661, "len_abs": 1413, "len_title": 90, "len_sents": 62536, "len_sents_1": 29033, "n_sents": 494, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 1418, "title": "Action abstractions for amortized sampling", "abs": "As trajectories sampled by policies used by reinforcement learning (RL) and generative flow networks (GFlowNets) grow longer, credit assignment and exploration become more challenging, and the long planning horizon hinders mode discovery and generalization.\nThe challenge is particularly pronounced in entropy-seeking RL methods, such as generative flow networks, where the agent must learn to sample from a structured distribution and discover multiple high-reward states, each of which take many steps to reach.\nTo tackle this challenge, we propose an approach to incorporate the discovery of action abstractions, or high-level actions, into the policy optimization process.\nOur approach involves iteratively extracting action subsequences commonly used across many high-reward trajectories and `chunking' them into a single action that is added to the action space.\nIn empirical evaluation on synthetic and real-world environments, our approach demonstrates improved sample efficiency performance in discovering diverse high-reward objects, especially on harder exploration problems.\nWe also observe that the abstracted high-order actions are potentially interpretable, capturing the latent structure of the reward landscape of the action space.\nThis work provides a cognitively motivated approach to action abstraction in RL and is the first demonstration of hierarchical planning in amortized sequential sampling.", "keywords": [ "GFlowNets", "amortized samplers", "hierarchical planning", "abstractions", "macro-actions" ], "rating_list": [ 6, 6, 8, 8, 6 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 1, 4, 4, 2 ], "contribution_list": [ 3, 1, 3, 3, 2 ], "confidence_list": [ 2, 4, 3, 2, 4 ] }, { "paper_id": "isHiGhFwVV", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Alain Ryser", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Alexander Marx", "gender": "unknown", "institution": "Technische Universit\u00e4t Dortmund", "country": "DE", "position": "Full Professor" }, { "name": "Julia E Vogt", "gender": "Female", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Assistant Professor" }, { "name": "Thomas Marco Sutter", "gender": "unknown", "institution": "ETH Zurich", "country": "CH", "position": "Postdoc" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 61, "n_ref": 116, "n_ref_all": 137, "n_fig": 6, "n_tab": 9, "L_tab": 4845, "n_element_tab": 289, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 817, "n_element_tab_1": 43, "formula_len_all": 1329, "formula_len_all_1": 925, "len_all": 190430, "len_all_1": 63298, "len_abs": 2584, "len_title": 40, "len_sents": 47518, "len_sents_1": 30607, "n_sents": 327, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1305, "title": "Anomaly Detection by Context Contrasting", "abs": "Anomaly detection focuses on identifying samples that deviate from the norm.\nWhen working with high-dimensional data such as images, a crucial requirement for detecting anomalous patterns is learning lower-dimensional representations that capture concepts of normality. \nRecent advances in self-supervised learning have shown great promise in this regard. \nHowever, many successful self-supervised anomaly detection methods assume prior knowledge about anomalies to create synthetic outliers during training. \nYet, in real-world applications, we often do not know what to expect from unseen data, and we can solely leverage knowledge about normal data. \nIn this work, we propose Con$_2$, which learns representations through context augmentations that allow us to observe samples from\ntwo distinct perspectives while keeping the invariances of normal data. \nCon$_2$ learns rich representations of context-augmented samples by clustering them according to their context while simultaneously \naligning their positions across clusters. \nAt test time, representations of anomalies that do not adhere to the invariances of normal data then deviate from their respective context cluster. \nLearning representations in such a way thus allows us to detect anomalies without making\nassumptions about anomalous data.", "keywords": [ "anomaly detection", "one class classification", "novelty detection", "contrastive representation learning", "context clustering" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "is4nCVkSFA", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Beining Wu", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Miao Lu", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Siyu Chen", "gender": "Male", "institution": "Yale University", "country": "", "position": "PhD student" }, { "name": "Tianhao Wang", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Research Assistant Professor" }, { "name": "Zhuoran Yang", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 589, "n_formula_1": 51, "n_ref_uni": 32, "n_ref": 103, "n_ref_all": 121, "n_fig": 4, "n_tab": 7, "L_tab": 4417, "n_element_tab": 121, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 65738, "formula_len_all_1": 3252, "len_all": 535316, "len_all_1": 89881, "len_abs": 1596, "len_title": 47, "len_sents": 153249, "len_sents_1": 35690, "n_sents": 1851, "n_sents_1": 296, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 105, "L_abs": 1696, "title": "Can Neural Networks Achieve Optimal Computational-statistical Tradeoff? An Analysis on Single-Index Model", "abs": "In this work, we tackle the following question: Can neural networks trained with gradient-based methods achieve the optimal statistical-computational tradeoff in learning Gaussian single-index models? \nPrior research has shown that any polynomial-time algorithm under the statistical query (SQ) framework requires $\\Omega(d^{s^\\star/2}\\lor d)$ samples, where $s^\\star$ is the generative exponent representing the intrinsic difficulty of learning the underlying model.\nHowever, it remains unknown whether neural networks can achieve this sample complexity. \nInspired by prior techniques such as label transformation and landscape smoothing for learning single-index models, we propose a unified gradient-based algorithm for training a two-layer neural network in polynomial time.\nOur method is adaptable to a variety of loss and activation functions, covering a broad class of existing approaches.\nWe show that our algorithm learns a feature representation that strongly aligns with the unknown signal $\\theta^\\star$, with sample complexity $\\tilde O (d^{s^\\star/2} \\lor d)$, matching the SQ lower bound up to a polylogarithmic factor for all generative exponents $s^\\star\\geq 1$.\nFurthermore, we extend our approach to the setting where $\\theta^\\star$ is $k$-sparse for $k = o(\\sqrt{d})$ by introducing a novel weight perturbation technique that leverages the sparsity structure. \nWe derive a corresponding SQ lower bound \nof order $\\tilde\\Omega(k^{s^\\star})$, matched by our method up to a polylogarithmic factor.\nOur framework, especially the weight perturbation technique, is of independent interest, and suggests potential gradient-based solutions to other problems such as sparse tensor PCA.", "keywords": [ "single-index model", "feature learning", "gradient-based method", "computational-statistical tradeoff" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "irrtPRFksw", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Alonso Granados", "gender": "Male", "institution": "University of Arizona", "country": "US", "position": "PhD student" }, { "name": "Jason L Pacheco", "gender": "Male", "institution": "University of Arizona", "country": "US", "position": "Assistant Professor" }, { "name": "Mohammadreza Ebrahimi", "gender": "Male", "institution": "University of South Florida", "country": "US", "position": "Assistant Professor" } ], "n_formula": 55, "n_formula_1": 25, "n_ref_uni": 56, "n_ref": 95, "n_ref_all": 128, "n_fig": 12, "n_tab": 4, "L_tab": 703, "n_element_tab": 41, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7327, "formula_len_all_1": 2113, "len_all": 153946, "len_all_1": 63226, "len_abs": 995, "len_title": 111, "len_sents": 38812, "len_sents_1": 29237, "n_sents": 328, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1001, "title": "Risk-Sensitive Variational Actor-Critic: A Model-Based Approach", "abs": "Risk-sensitive reinforcement learning (RL) with an entropic risk measure typically requires knowledge of the transition kernel or performs unstable updates w.r.t. exponential Bellman equations. As a consequence, algorithms that optimize this objective have been restricted to tabular or low-dimensional continuous environments. In this work we leverage the connection between the entropic risk measure and the RL-as-inference framework to develop a risk-sensitive variational actor-critic algorithm (rsVAC). Our work extends the variational framework to incorporate stochastic rewards and proposes a variational model-based actor-critic approach that modulates policy risk via a risk parameter. We consider, both, the risk-seeking and risk-averse regimes and present rsVAC learning variants for each setting. Our experiments demonstrate that this approach produces risk-sensitive policies and yields improvements in both tabular and risk-aware variants of complex continuous control tasks in MuJoCo.", "keywords": [ "reinforcement learning", "variational inference", "risk sensitive RL" ], "rating_list": [ 8, 5, 5, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "irPcM6X5FV", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Jieping Ye", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Rui Dai", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Sile Hu", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Xinmei Tian", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xu Shen", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Yonggang Zhang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Postdoc" } ], "n_formula": 45, "n_formula_1": 21, "n_ref_uni": 28, "n_ref": 95, "n_ref_all": 126, "n_fig": 7, "n_tab": 14, "L_tab": 8273, "n_element_tab": 848, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 920, "n_element_tab_1": 89, "formula_len_all": 4910, "formula_len_all_1": 1980, "len_all": 228181, "len_all_1": 66040, "len_abs": 1214, "len_title": 123, "len_sents": 56461, "len_sents_1": 28397, "n_sents": 426, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1219, "title": "Leveraging Submodule Linearity Enhances Task Arithmetic Performance in LLMs", "abs": "Task arithmetic is a straightforward yet highly effective strategy for model merging, enabling the resultant model to exhibit multi-task capabilities. Recent research indicates that models demonstrating linearity enhance the performance of task arithmetic. In contrast to existing methods that rely on the global linearization of the model, we argue that this linearity already exists within the model's submodules. In particular, we present a statistical analysis and show that submodules (e.g., layers, self-attentions, and MLPs) exhibit significantly higher linearity than the overall model. Based on these findings, we propose an innovative model merging strategy that independently merges these submodules. Especially, we derive a closed-form solution for optimal merging weights grounded in the linear properties of these submodules. Experimental results demonstrate that our method consistently outperforms the standard task arithmetic approach and other established baselines across different model scales and various tasks. This result highlights the benefits of leveraging the linearity of submodules and provides a new perspective for exploring solutions for effective and practical multi-task model merging.", "keywords": [ "Large Language model", "Task Arithmetic" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "irCuIdCdAl", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Sangkyun Lee", "gender": "unknown", "institution": "Korea University", "country": "", "position": "Associate Professor" }, { "name": "Sungmin Han", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" }, { "name": "JEONGHYUN LEE", "gender": "Male", "institution": "Korea University", "country": "", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 13, "n_ref_uni": 41, "n_ref": 78, "n_ref_all": 116, "n_fig": 10, "n_tab": 11, "L_tab": 3888, "n_element_tab": 605, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1489, "n_element_tab_1": 260, "formula_len_all": 1332, "formula_len_all_1": 1157, "len_all": 144420, "len_all_1": 63002, "len_abs": 1340, "len_title": 129, "len_sents": 42766, "len_sents_1": 25790, "n_sents": 290, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1361, "title": "Improving Transformer Interpretability with Activation Contrast-Based Attribution", "abs": "Transformers have revolutionized AI research, particularly in natural language processing (NLP). However, understanding the decisions made by transformer-based models remains challenging, which impedes trust and safe deployment in real-world applications. While activation-based attribution methods have proven effective in explaining transformer-based text classification models, our findings suggest that they may suffer from class-irrelevant features within activations, potentially degrading the quality of their interpretations. To address this issue, we introduce Contrast-CAT, a novel activation contrast-based attribution method that improves token-level attribution by filtering out class-irrelevant features from activations. Contrast-CAT enhances interpretability by contrasting the activations of input sequences with reference activations, allowing for the generation of clearer and more faithful attribution maps. Our experiments demonstrate that Contrast-CAT consistently outperforms state-of-the-art methods across various datasets and models, achieving significant gains over the second-best methods with average improvements in AOPC and LOdds by $\\times 1.30$ and $\\times 2.25$, respectively, under the MoRF setting. Contrast-CAT provides a promising step forward in enhancing the interpretability and transparency of transformer-based models.", "keywords": [ "Transformer", "Interpretability", "XAI", "Attention", "Contrast-based" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "iqqpx8hgSQ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Jiwan Seo", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 60, "n_ref": 89, "n_ref_all": 109, "n_fig": 5, "n_tab": 4, "L_tab": 3658, "n_element_tab": 372, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1705, "formula_len_all_1": 1506, "len_all": 157353, "len_all_1": 55612, "len_abs": 1003, "len_title": 111, "len_sents": 40419, "len_sents_1": 27696, "n_sents": 261, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1016, "title": "RAQ-VAE: Rate-Adaptive Vector-Quantized Variational Autoencoder", "abs": "Vector Quantized Variational AutoEncoder (VQ-VAE) is an established technique in machine learning for learning discrete representations across various modalities. However, its scalability and applicability are limited by the need to retrain the model to adjust the codebook for different rate requirements or encoding efficiency. We introduce the Rate-Adaptive VQ-VAE ($\\textbf{RAQ-VAE}$) framework, which addresses this challenge with two novel discrete (codebook) representation methods: a model-based approach using a clustering technique for existing pre-trained VQ-VAE models, and a data-driven approach utilizing a sequence-to-sequence (Seq2Seq) model for variable-rate codebook generation. Our experiments demonstrate that RAQ-VAE achieves effective reconstruction performance across multiple rates, often outperforming conventional fixed-rate VQ-VAE models. This work enhances the adaptability and performance of VQ-VAEs, with broad applications in data reconstruction, generation, and computer vision tasks.", "keywords": [ "Discrete representation learning", "Vector-quantized variational autoencoder", "Generative model", "Sequence-to-sequence" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "iqdqRmqUsD", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Angel Villar-Corrales", "gender": "Male", "institution": "University of Bonn", "country": "DE", "position": "PhD student" }, { "name": "Jan Niklas Ewertz", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "PhD student" }, { "name": "Malte Mosbach", "gender": "unknown", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "PhD student" }, { "name": "Sven Behnke", "gender": "Male", "institution": "University of Bonn", "country": "DE", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 11, "n_ref_uni": 35, "n_ref": 102, "n_ref_all": 135, "n_fig": 20, "n_tab": 1, "L_tab": 257, "n_element_tab": 34, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 994, "formula_len_all_1": 685, "len_all": 168493, "len_all_1": 56722, "len_abs": 1281, "len_title": 126, "len_sents": 51155, "len_sents_1": 28596, "n_sents": 327, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1297, "title": "Learning Object-centric Latent Dynamics for Reinforcement Learning from Pixels", "abs": "Learning a latent dynamics model provides a task-agnostic representation of an agent\u2019s understanding of its environment. Leveraging this knowledge for model-based reinforcement learning holds the potential to improve sample efficiency over model-free methods by learning inside imagined rollouts. Furthermore, because the latent space serves as input to behavior models, the informative representations learned by the world model facilitate efficient learning of desired skills. However, most existing methods rely on holistic representations of the environment\u2019s state. In contrast, humans reason about objects and their interactions, forecasting how actions will affect specific parts of their surroundings. Inspired by this, we propose Slot-Attention for Object-centric Latent Dynamics (SOLD), a novel algorithm that learns object-centric dynamics models in an unsupervised manner from pixel inputs. We demonstrate that the structured latent space not only improves model interpretability but also provides a valuable input space for behavior models to reason over. Our results show that SOLD outperforms DreamerV3, a state-of-the-art model-based RL algorithm, across a range of benchmark robotic environments that evaluate for both relational reasoning and low-level manipulation capabilities.", "keywords": [ "Reinforcement Learning", "World Models", "Object-centric Representations" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "iqd8aHKwGA", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Jialin Liu", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Wotao Yin", "gender": "Male", "institution": "Alibaba Group US", "country": "", "position": "Principal Researcher" }, { "name": "Xiaohan Chen", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Xinshang Wang", "gender": "unknown", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Ziang Chen", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Instructor" } ], "n_formula": 118, "n_formula_1": 18, "n_ref_uni": 45, "n_ref": 107, "n_ref_all": 127, "n_fig": 6, "n_tab": 6, "L_tab": 2265, "n_element_tab": 107, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 791, "n_element_tab_1": 57, "formula_len_all": 12750, "formula_len_all_1": 2577, "len_all": 253574, "len_all_1": 82279, "len_abs": 1299, "len_title": 80, "len_sents": 76700, "len_sents_1": 35900, "n_sents": 839, "n_sents_1": 306, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1308, "title": "Expressive Power of Graph Neural Networks for (Mixed-Integer) Quadratic Programs", "abs": "Quadratic programming (QP) is the most widely applied category of problems in nonlinear programming. Many applications require real-time/fast solutions, though not necessarily with high precision. Existing methods either involve matrix decomposition or use the preconditioned conjugate gradient method. For relatively large instances, these methods cannot achieve the real-time requirement unless there is an effective preconditioner. Recently, graph neural networks (GNNs) opened new possibilities for QP. Some promising empirical studies of applying GNNs for QP tasks show that GNNs can capture key characteristics of an optimization instance and provide adaptive guidance accordingly to crucial configurations during the solving process, or directly provide an approximate solution. Despite notable empirical observations, theoretical foundations are still lacking.\n\nIn this work, we investigate the expressive or representative power of GNNs, a crucial aspect of neural network theory, specifically in the context of QP tasks, with both continuous and mixed-integer settings. We prove the existence of message-passing GNNs that can reliably represent key properties of quadratic programs, including feasibility, optimal objective value, and optimal solution. Our theory is validated by numerical results.", "keywords": [ "Quadratic programs", "mixed-integer quadratic programs", "graph neural networks", "universal approximation" ], "rating_list": [ 8, 6, 6, 6, 3 ], "soundness_list": [ 3, 4, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 2, 3 ], "contribution_list": [ 3, 4, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "iplOFSOzS2", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Changick Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Donguk Kim", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "MS student" }, { "name": "Jaehyuk Jang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Sangmin Woo", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "\ucd5c\uc720\ube48", "gender": "Female", "institution": "KAIST, Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 26, "n_formula_1": 11, "n_ref_uni": 50, "n_ref": 166, "n_ref_all": 205, "n_fig": 15, "n_tab": 12, "L_tab": 10454, "n_element_tab": 908, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3981, "n_element_tab_1": 176, "formula_len_all": 920, "formula_len_all_1": 446, "len_all": 238213, "len_all_1": 66561, "len_abs": 1479, "len_title": 149, "len_sents": 65404, "len_sents_1": 28525, "n_sents": 585, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1487, "title": "Don't Miss the Forest for the Trees: Attentional Vision Calibration for Large Vision Language Models", "abs": "This study seeks to understand and address a phenomenon observed in Large Vision Language Models (LVLMs) related to their attention mechanism. Interestingly, LVLMs tend to disproportionately focus on a few image tokens that lack meaningful, query-related semantics, leading to sharp outlier values in the attention maps \u2014 tokens we refer to as blind tokens. In well-designed attention mechanisms, the principle is to assign higher weights to the most relevant tokens. However, in this case, the attention imbalance leads to overemphasis on uninformative tokens, which is far from ideal. Our analysis shows that tokens receiving lower attention weights often hold critical information necessary for capturing subtle visual details. We hypothesize that over-reliance on blind tokens contributes to hallucinations in LVLMs.\nTo address this, we introduce a novel decoding technique called Attentional Vision Calibration (AVISC). During the decoding phase, AVISC identifies blind tokens by examining the image-wise attention distribution and dynamically adjusts the logits for the prediction. Specifically, it contrasts the logits conditioned on the original visual tokens with those conditioned on the blind tokens, thereby reducing the model\u2019s dependency on blind tokens and encouraging a more balanced consideration of all visual tokens. We validate AVISC on standard hallucination benchmarks, including POPE, MME, and AMBER, where it consistently outperforms existing decoding techniques.", "keywords": [ "Large Vision Language Models" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "ipQrjRsl11", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Finale Doshi-Velez", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "Mohammad Emtiyaz Khan", "gender": "Male", "institution": "RIKEN Center for AI Project", "country": "JP", "position": "Full Professor" }, { "name": "Siddharth Swaroop", "gender": "unknown", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "", "position": "Postdoc" } ], "n_formula": 43, "n_formula_1": 30, "n_ref_uni": 39, "n_ref": 87, "n_ref_all": 111, "n_fig": 1, "n_tab": 2, "L_tab": 467, "n_element_tab": 34, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 469, "n_element_tab_1": 5, "formula_len_all": 3410, "formula_len_all_1": 1990, "len_all": 163934, "len_all_1": 68317, "len_abs": 697, "len_title": 82, "len_sents": 52054, "len_sents_1": 31707, "n_sents": 454, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 34, "L_abs": 699, "title": "Connecting Federated ADMM to Bayes", "abs": "We provide new connections between two distinct federated learning approaches based on (i) ADMM and (ii) Variational Bayes (VB), and propose new variants by combining their complementary strengths. Specifically, we show that the dual variables in ADMM naturally emerge through the \"site\" parameters used in VB with isotropic Gaussian covariances. Using this, we derive two versions of ADMM from VB that use flexible covariances and functional regularisation, respectively. Through numerical experiments, we validate the improvements obtained in performance. The work shows connection between two fields that are believed to be fundamentally different and combines them to improve federated learning.", "keywords": [ "federated learning", "bayesian", "variational inference", "admm" ], "rating_list": [ 6, 6, 5, 8, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 4, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 2, 3, 3, 4 ] }, { "paper_id": "iotwQLLatn", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Ke Zhou", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yu Liu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Researcher" }, { "name": "Yunc G", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Zhiqi Shen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Lecturer" }, { "name": "ketong liu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 20, "n_ref": 46, "n_ref_all": 70, "n_fig": 5, "n_tab": 8, "L_tab": 2515, "n_element_tab": 213, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 701, "n_element_tab_1": 113, "formula_len_all": 560, "formula_len_all_1": 561, "len_all": 123984, "len_all_1": 60799, "len_abs": 1308, "len_title": 107, "len_sents": 34977, "len_sents_1": 26534, "n_sents": 292, "n_sents_1": 223, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1422, "title": "Unsupervised Meta-Learning via Dynamic Head and Heterogeneous Task Construction for Few-Shot Classification", "abs": "Meta-learning has been widely used in recent years in areas such as few-shot learning and reinforcement learning. However, the questions of why and when it's better than other algorithms in few-shot classification remain to be explored. In this paper, we answer the above questions from the perspective of data noise and heterogeneous tasks. Specifically, we perform pre-experiments by adjusting the proportion of data noise and the degree of heterogeneity of the task in the dataset. We use the metric of Singular Vector Canonical Correlation Analysis to quantified the representation stability of the neural network and thus to compare the behavior of meta-learning algorithms and other algorithms. We find that benefits from the bi-level optimization strategy, the meta-learning algorithm has better robustness to label noise and heterogeneous tasks. Based on the above conclusion, we argue a promising future for meta-learning in the unsupervised area, and thus propose DHM-UHT, a dynamic head meta-learning algorithm with unsupervised heterogeneous task construction. The core idea of DHM-UHT is to use DBSCAN and dynamic head to achieve heterogeneous task construction and meta-learn the whole process of unsupervised heterogeneous task construction. On several unsupervised zero-shot and few-shot datasets, DHM-UHT obtains state-of-the-art performance. The code is released at https://github.com/tuantuange/DHM-UHT.", "keywords": [ "Meta-Learning; Unsupervised Learning; Few-shot Learning; Data Noise; Heterogeneous Task" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "ioprnwVrDH", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Chi Jin", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Assistant Professor" }, { "name": "Mengzhou Xia", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Wenzhe Li", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "LIN Yong", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Postdoc" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 25, "n_ref": 74, "n_ref_all": 112, "n_fig": 6, "n_tab": 8, "L_tab": 2349, "n_element_tab": 303, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1344, "n_element_tab_1": 204, "formula_len_all": 134, "formula_len_all_1": 109, "len_all": 138403, "len_all_1": 62681, "len_abs": 1539, "len_title": 132, "len_sents": 41056, "len_sents_1": 28091, "n_sents": 330, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1570, "title": "Rethinking Mixture-of-Agents: Is Mixing Different Large Language Models Beneficial?", "abs": "Ensembling outputs from diverse sources is a straightforward yet effective approach to boost performance. Mixture-of-Agents (MoA) is one such popular ensemble method that aggregates outputs from multiple *different* Large Language Models (LLMs). This paper raises the question in the context of language models: is mixing different LLMs truly beneficial? We propose Self-MoA --- an ensemble method that aggregates outputs from only the *single* top-performing LLM. Our extensive experiments reveal that, surprisingly, Self-MoA outperforms standard MoA that mixes different LLMs in a large number of scenarios: Self-MoA achieves $6.6\\%$ improvement over MoA on the AlpacaEval 2.0 benchmark, and an average of $3.8\\%$ improvement across various benchmarks, including MMLU, CRUX, and MATH. Applying Self-MoA to one of the top-ranking models in AlpacaEval 2.0 directly achieves the new state-of-the-art performance ranking $1^{\\text{st}}$ on the leaderboard. To understand the effectiveness of Self-MoA, we systematically investigate the trade-off between diversity and quality of outputs under various MoA settings. We confirm that the MoA performance is rather sensitive to the quality, and mixing different LLMs often lowers the average quality of the models. To complement the study, we identify the scenarios where mixing different LLMs could be helpful. This paper further introduces a sequential version of self-MoA, that is capable of aggregating a large number of LLM outputs on-the-fly over multiple rounds, and is as effective as aggregating all outputs at once.", "keywords": [ "large language models", "ensemble", "mixture-of-agent" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "ioOgrS0UKx", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Giridhar Narasapura Rajagopalaiah", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Jonathan Chung", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 18, "n_ref": 46, "n_ref_all": 68, "n_fig": 6, "n_tab": 4, "L_tab": 1588, "n_element_tab": 126, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1528, "n_element_tab_1": 118, "formula_len_all": 80, "formula_len_all_1": 80, "len_all": 73141, "len_all_1": 41420, "len_abs": 1332, "len_title": 161, "len_sents": 22134, "len_sents_1": 16846, "n_sents": 171, "n_sents_1": 132, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1333, "title": "PlicoTabTransformer: Folding Tabular Embeddings Into M Vectors", "abs": "Tabular data represents the most prevalent and extensively utilized form of structured data in various domains. Traditionally dominated by tree-based algorithms, researchers are actively exploring the application of deep neural networks on tabular data. Notably, the TabTransformer (Huang et al., 2020) and FT-transformer (Gorishniy et al., 2021) showed that feeding column embeddings of the tabular\nfeatures into a transformer could learn a representation of the columns and how the embeddings interact with one another. This paper introduces PlicoTabTransformer, an enhancement of the previous methods, which is designed to learn multiple representations of the column embeddings. By incorporating a transformer with multiple learnable position embeddings and a contrastive learning loss, our\nmethod learns multiple distinct and orthogonal representations (denoted as plicovectors) of the column embeddings. We evaluated the PlicoTabTransformer with the pytorch-frame benchmark. Our experimental demonstrated that the PlicoTabTransformer is overall top ranked algorithm and achieves state of the art performance in several datasets compared to other deep learning method closing the gap\nwith tree based algorithms. Our method provides an added advantage to visualise redundancies and a potential dimensionality reduction technique.", "keywords": [ "Tabular data", "Transformer", "Self-attention", "Positional embeddings", "Contrastive loss", "Classification" ], "rating_list": [ 3, 3, 1, 5 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "io8uRPYktn", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 12, "author_info_list": [ { "name": "Aolin Ding", "gender": "Male", "institution": "Accenture", "country": "US", "position": "Researcher" }, { "name": "Hai Li", "gender": "Female", "institution": "Duke University", "country": "US", "position": "Professor" }, { "name": "Jianyi Zhang", "gender": "unknown", "institution": "National Artificial Intelligence Research Resource Pilot & Duke University", "country": "US", "position": "Principal investigator" }, { "name": "Jingwei Sun", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Jingyang Zhang", "gender": "Male", "institution": "Electrical and Computer Engineering, Duke University", "country": "US", "position": "PhD student" }, { "name": "Louis DiValentin", "gender": "unknown", "institution": "Accenture", "country": "", "position": "Researcher" }, { "name": "Martin Kuo", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Minxue Tang", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "William Chen", "gender": "Male", "institution": "North Carolina School of Science and Mathematics", "country": "US", "position": "Researcher" }, { "name": "Yiran Chen", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Professor" }, { "name": "Amin Hassanzadeh", "gender": "Male", "institution": "Accenture", "country": "US", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 9, "n_ref_uni": 37, "n_ref": 85, "n_ref_all": 113, "n_fig": 5, "n_tab": 18, "L_tab": 4679, "n_element_tab": 800, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1732, "n_element_tab_1": 34, "formula_len_all": 1259, "formula_len_all_1": 727, "len_all": 156855, "len_all_1": 47491, "len_abs": 1213, "len_title": 159, "len_sents": 46536, "len_sents_1": 22428, "n_sents": 358, "n_sents_1": 129, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 109, "L_abs": 1217, "title": "Proactive Privacy Amnesia for Large Language Models: Safeguarding PII with Negligible Impact on Model Utility", "abs": "With the rise of large language models (LLMs), increasing research has recognized\ntheir risk of leaking personally identifiable information (PII) under malicious\nattacks. Although efforts have been made to protect PII in LLMs, existing methods\nstruggle to balance privacy protection with maintaining model utility. In this paper,\ninspired by studies of amnesia in cognitive science, we propose a novel approach,\nProactive Privacy Amnesia (PPA), to safeguard PII in LLMs while preserving their\nutility. This mechanism works by actively identifying and forgetting key memories\nmost closely associated with PII in sequences, followed by a memory implanting\nusing suitable substitute memories to maintain the LLM\u2019s functionality. We conduct\nevaluations across multiple models to protect common PII, such as phone numbers\nand physical addresses, against prevalent PII-targeted attacks, demonstrating the\nsuperiority of our method compared with other existing defensive techniques. The\nresults show that our PPA method completely eliminates the risk of phone number\nexposure by 100% and significantly reduces the risk of physical address exposure\nby 9.8% \u2013 87.6%, all while maintaining comparable model utility performance.", "keywords": [ "Large language models", "Personal identifiable information", "Protect private data leakage" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "inpLTODeA6", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Ge Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Hangyu Guo", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Haoran Zhang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Intern" }, { "name": "Liu Jiaheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Meng Cao", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Shuyue Guo", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Wenhao Huang", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 34, "n_ref": 69, "n_ref_all": 87, "n_fig": 9, "n_tab": 3, "L_tab": 3922, "n_element_tab": 552, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 2475, "n_element_tab_1": 571, "formula_len_all": 48, "formula_len_all_1": 48, "len_all": 226626, "len_all_1": 70927, "len_abs": 1431, "len_title": 101, "len_sents": 51091, "len_sents_1": 28586, "n_sents": 568, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1434, "title": "ING-VP: MLLMs Cannot Play Easy Vision-based Games Yet", "abs": "As multimodal large language models (MLLMs) continue to demonstrate increasingly competitive performance across a broad spectrum of tasks, more intricate\nand comprehensive benchmarks have been developed to assess these cutting-edge\nmodels. These benchmarks introduce new challenges to core capabilities such\nas perception, reasoning, and planning. However, existing multimodal benchmarks fall short in providing a focused evaluation of multi-step planning based\non spatial relationships in images. To bridge this gap, we present ING-VP,\nthe first INteractive Game-based Vision Planning benchmark, specifically designed to evaluate the spatial imagination and multi-step reasoning abilities of\nMLLMs. ING-VP features 6 distinct games, encompassing 300 levels, each with\n6 unique configurations. A single model engages in over 60,000 rounds of interaction. The benchmark framework allows for multiple comparison settings,\nincluding image-only vs. text-only inputs, single-step vs. multi-step reasoning,\nand with-history vs. without-history conditions, offering valuable insights into\nthe model\u2019s capabilities. We evaluated numerous state-of-the-art MLLMs, with\nthe highest-performing model, Claude-3.5 Sonnet, achieving a best accuracy of\nonly 8.00%, far below the human accuracy of 65.66%. This work aims to provide\na specialized evaluation framework to drive advancements in MLLMs\u2019 capacity\nfor complex spatial reasoning and planning.", "keywords": [ "Multimodal", "Benchmark", "Game", "Planning" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "inOwd7hZC1", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Kehan Wen", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "MS student" }, { "name": "Lei Ke", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Yao Mu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yutong Hu", "gender": "unknown", "institution": "ETH Zurich", "country": "CH", "position": "Intern" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 76, "n_ref_all": 105, "n_fig": 12, "n_tab": 8, "L_tab": 1713, "n_element_tab": 253, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 991, "n_element_tab_1": 184, "formula_len_all": 454, "formula_len_all_1": 540, "len_all": 149621, "len_all_1": 67122, "len_abs": 1252, "len_title": 129, "len_sents": 48401, "len_sents_1": 30672, "n_sents": 346, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1317, "title": "M^3PC: Test-time Model Predictive Control using Pretrained Masked Trajectory Model", "abs": "Recent work in Offline Reinforcement Learning (RL) has shown that a unified transformer trained under a masked auto-encoding objective can effectively capture the relationships between different modalities (e.g., states, actions, rewards) within given trajectory datasets. However, this information has not been fully exploited during the inference phase, where the agent needs to generate an optimal policy instead of just reconstructing masked components from unmasked. Given that a pretrained trajectory model can act as both a Policy Model and a World Model with appropriate mask patterns, we propose using Model Predictive Control (MPC) at test time to leverage the model's own predictive capacity to guide its action selection. Empirical results on D4RL and RoboMimic show that our inference-phase MPC significantly improves the decision-making performance of a pretrained trajectory model without any additional parameter training. Furthermore, our framework can be adapted to Offline to Online (O2O) RL and Goal Reaching RL, resulting in more substantial performance gains when an additional online interaction budget is given, and better generalization capabilities when different task targets are specified. Code is available: \\href{https://github.com/wkh923/m3pc}{\\texttt{https://github.com/wkh923/m3pc}}.", "keywords": [ "Offline-to-Online Reinforcement Learning", "Model-based Reinforcement Learning", "Masked Autoencoding", "Robot Learning" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "inLUnCpDIB", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Binghui Li", "gender": "Not Specified", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yuanzhi Li", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 111, "n_formula_1": 26, "n_ref_uni": 42, "n_ref": 75, "n_ref_all": 88, "n_fig": 3, "n_tab": 4, "L_tab": 1866, "n_element_tab": 347, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12657, "formula_len_all_1": 3385, "len_all": 225727, "len_all_1": 58676, "len_abs": 1712, "len_title": 172, "len_sents": 63940, "len_sents_1": 24151, "n_sents": 604, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 124, "L_abs": 1718, "title": "Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data", "abs": "Adversarial training is a widely-applied approach to training deep neural networks to be robust against adversarial perturbation. However, although adversarial training has achieved empirical success in practice, it still remains unclear why adversarial examples exist and how adversarial training methods improve model robustness. In this paper, we provide a theoretical understanding of adversarial examples and adversarial training algorithms from the perspective of feature learning theory. Specifically, we focus on a multiple classification setting, where the structured data can be composed of two types of features: the robust features, which are resistant to perturbation but sparse, and the non-robust features, which are susceptible to perturbation but dense. We train a two-layer smoothed ReLU convolutional neural network to learn our structured data. First, we prove that by using standard training (gradient descent over the empirical risk), the network learner primarily learns the non-robust feature rather than the robust feature, which thereby leads to the adversarial examples that are generated by perturbations aligned with negative non-robust feature directions. Then, we consider the gradient-based adversarial training algorithm, which runs gradient ascent to find adversarial examples and runs gradient descent over the empirical risk at adversarial examples to update models. We show that the adversarial training method can provably strengthen the robust feature learning and suppress the non-robust feature learning to improve the network robustness. Finally, we also empirically validate our theoretical findings with experiments on real-image datasets, including MNIST, CIFAR10 and SVHN.", "keywords": [ "deep learning theory", "feature learning", "adversarial robustness", "adversarial training", "non-convex optimization" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 3, 2 ] }, { "paper_id": "in8qEyM4Xp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Changyou Chen", "gender": "Male", "institution": "State University of New York, Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Dancheng Liu", "gender": "Male", "institution": "University at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Jiayu Qin", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Visiting researcher" }, { "name": "Jingchen Sun", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Jinjun Xiong", "gender": "unknown", "institution": "State University of New York at Buffalo", "country": "US", "position": "Professor" }, { "name": "Rohan Sharma", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Shijie Zhou", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 18, "n_ref_uni": 66, "n_ref": 107, "n_ref_all": 122, "n_fig": 5, "n_tab": 5, "L_tab": 1989, "n_element_tab": 113, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 906, "n_element_tab_1": 168, "formula_len_all": 902, "formula_len_all_1": 827, "len_all": 182943, "len_all_1": 67522, "len_abs": 1705, "len_title": 80, "len_sents": 49200, "len_sents_1": 29415, "n_sents": 339, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1685, "title": "KidSpeak: A General Multi-Purpose LLM for Kids' Speech Recognition and Screening", "abs": "With the rapid advancement of conversational and diffusion-based AI, there is a growing adoption of AI in educational services, ranging from grading and assessment tools to personalized learning systems that provide targeted support for students. However, this adaptability has yet to fully extend to the domain of children's speech, where existing models often fail due to their reliance on datasets designed for clear, articulate adult speech. Children, particularly those in early developmental stages or with speech and language pathologies, present unique challenges that current AI models and datasets are ill-equipped to handle. To address this, we introduce KidSpeak, a multi-task speech-enhanced Foundation Model capable of both generative and discriminative tasks specifically tailored to children's speech patterns. Our framework employs a two-stage training process that incorporates phonetic knowledge into the speech encoder, achieving an average accuracy of 87\\% across four separate tasks. Furthermore, recognizing the limitations of scalable human annotation and existing speech alignment tools, we propose the Flexible and Automatic Speech Aligner (FASA) and leverage the method to construct high quality datasets for training and evaluation. This novel alignment tool significantly improves the quality of aligned children's speech from noisy data, enhancing data quality by 13.6\u00d7 compared to human annotations, as demonstrated on the CHILDES dataset. To the best of our knowledge, KidSpeak and FASA represent the first comprehensive solution designed for speech and language therapy in children, offering both a multi-purpose speech LLM and a robust alignment tool.", "keywords": [ "Speech Language Modeling", "Children's Speech", "Speech Pathology Diagnosis", "Speech Transcription", "Children's Speech Dataset Creation" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 1, 1 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "in0Nmo8Ojd", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Daniel Hettegger", "gender": "Not Specified", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Daniel Koutas", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Daniel Straub", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Associate Professor" }, { "name": "Kostas G. Papakonstantinou", "gender": "unknown", "institution": "Pennsylvania State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 39, "n_formula_1": 28, "n_ref_uni": 34, "n_ref": 64, "n_ref_all": 88, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 49, "n_element_tab_1": 1, "formula_len_all": 2513, "formula_len_all_1": 1431, "len_all": 121706, "len_all_1": 57471, "len_abs": 645, "len_title": 134, "len_sents": 40877, "len_sents_1": 26248, "n_sents": 277, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 647, "title": "Convex is back: \\\\ Solving Belief MDPs via Convexity-Informed Deep Reinforcement Learning", "abs": "We present a novel method for Deep Reinforcement Learning (DRL), incorporating the convex property of the value function over the belief space in Partially Observable Markov Decision Processes (POMDPs). We introduce hard- and soft-enforced convexity as two different approaches, and compare their performance against standard DRL on two well-known POMDP environments, namely the Tiger and FieldVisionRockSample problems. Our findings show that including the convexity feature can substantially increase performance of the agents, as well as increase robustness over the hyperparameter space, especially when testing on out-of-distribution domains.", "keywords": [ "Deep Reinforcement Learning", "POMDP", "Convexity" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "imT03YXlG2", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Hyesu Lim", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jaegul Choo", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Jinho Choi", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Steffen Schneider", "gender": "unknown", "institution": "Helmholtz Zentrum M\u00fcnchen", "country": "DE", "position": "Principal Researcher" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 29, "n_ref": 73, "n_ref_all": 128, "n_fig": 25, "n_tab": 2, "L_tab": 881, "n_element_tab": 77, "n_fig_1": 21, "n_tab_1": 1, "L_tab_1": 865, "n_element_tab_1": 72, "formula_len_all": 408, "formula_len_all_1": 392, "len_all": 137880, "len_all_1": 66656, "len_abs": 912, "len_title": 132, "len_sents": 45257, "len_sents_1": 30354, "n_sents": 369, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 880, "title": "Sparse autoencoders reveal selective remapping of visual concepts during adaptation", "abs": "Adapting foundation models for specific purposes has become a standard approach in machine learning systems, yet it is an open question which mechanisms take place during the adaptation. Here we develop PatchSAE to discover interpretable candidate concepts from vision encoders with spatially localized attributions. We explore how these concepts influence the model behavior and extend it to investigate how recent state-of-the-art adaptation techniques change the association of model inputs to these concepts. While activations of concepts slightly change between adapted and non-adapted models, we find that the majority of gains on common adaptation tasks can be explained with the existing concepts within the foundation model. This work provides a concrete framework to train and use SAEs for Vision Transformers and provides insights into explaining adaptation mechanisms.", "keywords": [ "interpretability", "vision-language models", "sparse autoencoder", "adaptation" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 4 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "ilcsm8B7Pe", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Chao Ning", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Han Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "MS student" } ], "n_formula": 77, "n_formula_1": 31, "n_ref_uni": 46, "n_ref": 69, "n_ref_all": 109, "n_fig": 3, "n_tab": 14, "L_tab": 1273, "n_element_tab": 100, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 24, "n_element_tab_1": 1, "formula_len_all": 6176, "formula_len_all_1": 2291, "len_all": 188479, "len_all_1": 70470, "len_abs": 1589, "len_title": 108, "len_sents": 60115, "len_sents_1": 29459, "n_sents": 475, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1594, "title": "End-to-End Conformal Prediction for Trajectory Optimization", "abs": "Conformal Prediction (CP) is a powerful tool to construct uncertainty sets with coverage guarantees, which has fueled its extensive adoption in generating prediction regions for decision-making tasks, e.g., Trajectory Optimization (TO) in uncertain environments. However, existing methods predominantly employ a sequential scheme, where decisions rely unidirectionally on the prediction regions, and consequently the information from the decision-making end fails to be transmitted back to instruct the CP end. In this paper, we propose a novel End-to-End CP (E2E-CP) framework for shrinking-horizon TO with a joint risk constraint over the entire mission time. Specifically, a CP-based posterior risk calculation method is developed by fully leveraging the realized trajectories to adjust the posterior allowable risk, which is then allocated to future times to update prediction regions. In this way, the information in the realized trajectories is continuously fed back to the CP end, enabling attractive end-to-end adjustments of the prediction regions and a provable online improvement in trajectory performance. Furthermore, we theoretically prove that such end-to-end adjustments consistently maintain the coverage guarantees of the prediction regions, thereby ensuring provable safety. Additionally, we develop a decision-focused iterative risk allocation algorithm with theoretical convergence analysis for allocating the posterior allowable risk which closely aligns with E2E-CP. The effectiveness and superiority of the proposed method are demonstrated through benchmark experiments.", "keywords": [ "conformal prediction", "end-to-end", "trajectory optimization", "risk allocation", "decision-focused learning" ], "rating_list": [ 8, 6, 5, 5 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "ilbxbOHk7a", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Alberto Marchesi", "gender": "Male", "institution": "Politecnico di Milano", "country": "", "position": "Assistant Professor" }, { "name": "Anna Lunghi", "gender": "unknown", "institution": "Polytechnic Institute of Milan", "country": "IT", "position": "PhD student" }, { "name": "Francesco Emanuele Stradi", "gender": "Male", "institution": "Polytechnic Institute of Milan", "country": "IT", "position": "PhD student" }, { "name": "Matteo Castiglioni", "gender": "unknown", "institution": "Politecnico di Milano", "country": "IT", "position": "Assistant Professor" }, { "name": "Nicola Gatti", "gender": "Male", "institution": "Polytechnic Institute of Milan", "country": "IT", "position": "Full Professor" } ], "n_formula": 176, "n_formula_1": 28, "n_ref_uni": 23, "n_ref": 68, "n_ref_all": 73, "n_fig": 0, "n_tab": 3, "L_tab": 580, "n_element_tab": 166, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 20041, "formula_len_all_1": 2424, "len_all": 211523, "len_all_1": 72073, "len_abs": 1135, "len_title": 118, "len_sents": 58493, "len_sents_1": 30903, "n_sents": 619, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1193, "title": "Best-of-Both-Worlds Policy Optimization for CMDPs with Bandit Feedback", "abs": "We study online learning in constrained Markov decision processes (CMDPs) in which rewards and constraints may be either stochastic or adversarial. In such settings, Stradi et al. (2024b) proposed the first best-of-both-worlds algorithm able to seamlessly handle stochastic and adversarial constraints, achieving optimal regret and constraint violation bounds in both cases. This algorithm suffers from two major drawbacks. First, it only works under full feedback, which severely limits its applicability in practice. Moreover, it relies on optimizing over the space of occupancy measures, which requires solving convex optimization problems, an highly inefficient task. In this paper, we provide the first best-of-both-worlds algorithm for CMDPs with bandit feedback. Specifically, when the constraints are stochastic, the algorithm achieves $\\widetilde{\\mathcal{O}}(\\sqrt{T})$ regret and constraint violation, while, when they are adversarial, it attains $\\widetilde{\\mathcal{O}}(\\sqrt{T})$ constraint violation and a tight fraction of the optimal reward. Moreover, our algorithm is based on a policy optimization approach, which is much more efficient than occupancy-measure-based methods.", "keywords": [ "CMDP", "Online Learning", "Best-of-both-worlds" ], "rating_list": [ 3, 3, 3, 6, 3 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 2, 4 ] }, { "paper_id": "ilOEOIqolQ", "primary_area": "interpretability and explainable AI", "n_author": 11, "author_info_list": [ { "name": "Allyson Ettinger", "gender": "Female", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Jiacheng Liu", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Intern" }, { "name": "Khyathi Chandu", "gender": "Female", "institution": "Mistral AI", "country": "US", "position": "Researcher" }, { "name": "Liwei Jiang", "gender": "Female", "institution": "University of Washington", "country": "", "position": "PhD student" }, { "name": "Melanie Sclar", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Nouha Dziri", "gender": "Female", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Seungju Han", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Skyler Hallinan", "gender": "Male", "institution": "University of Washington", "country": "", "position": "MS student" }, { "name": "Ximing Lu", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Yejin Choi", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Fatemehsadat Mireshghallah", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "Postdoc" } ], "n_formula": 8, "n_formula_1": 1, "n_ref_uni": 60, "n_ref": 127, "n_ref_all": 151, "n_fig": 16, "n_tab": 1, "L_tab": 231, "n_element_tab": 52, "n_fig_1": 16, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1155, "formula_len_all_1": 96, "len_all": 274102, "len_all_1": 69696, "len_abs": 1637, "len_title": 187, "len_sents": 66137, "len_sents_1": 35743, "n_sents": 403, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 138, "L_abs": 1482, "title": "AI as Humanity\u2019s Salieri: Quantifying Linguistic Creativity of Language Models via Systematic Attribution of Machine Text against Web Text", "abs": "Creativity has long been considered one of the most difficult aspect of human intelligence for AI to mimic. However, the rise of Large Language Models (LLMs), like ChatGPT, has raised questions about whether AI can match or even surpass\nhuman creativity. We present CREATIVITY INDEX as the first step to quantify the linguistic creativity of a text by reconstructing it from existing text snippets on the web. CREATIVITY INDEX is motivated by the hypothesis that the seemingly remarkable creativity of LLMs may be attributable in large part to the creativity of human-written texts on the web. To compute CREATIVITY INDEX efficiently, we introduce DJ SEARCH, a novel dynamic programming algorithm that can search verbatim and near-verbatim matches of text snippets from a given document against the web. Experiments reveal that the CREATIVITY INDEX of professional human authors is on average 66.2% higher than that of LLMs, and that alignment reduces the CREATIVITY INDEX of LLMs by an average of 30.1%. In addition, we explore variations in the CREATIVITY INDEX among different human authors and discuss the potential factors contributing to these differences. Finally, we showcase a novel application of CREATIVITY INDEX for zero-shot machine text detection, where it proves to be surprisingly effective\u2014outperforming the strong zero-shot system DetectGPT by a substantial margin of 30.2%, and even surpassing a leading supervised system, GhostBuster, in five out of six domains.", "keywords": [ "Machine Creativity", "Large Language Model", "Science of LLM", "Machine Text Detection" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 4, 4, 4, 4 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 5, 3, 5, 3 ] }, { "paper_id": "ilGdLPy3mA", "primary_area": "interpretability and explainable AI", "n_author": 1, "author_info_list": [ { "name": "Luna Zhang", "gender": "unknown", "institution": "State University of New York at Stony Brook", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 66, "n_ref": 86, "n_ref_all": 110, "n_fig": 7, "n_tab": 4, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 624, "formula_len_all_1": 749, "len_all": 183104, "len_all_1": 64007, "len_abs": 2552, "len_title": 172, "len_sents": 34080, "len_sents_1": 29827, "n_sents": 256, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 122, "L_abs": 1853, "title": "A New 3D Image Block Ranking Method Using Axial, Coronal and Sagittal Image Patch Rankings for Explainable Medical Imaging", "abs": "Although a 3D Convolutional Neural Network (CNN) has been applied to explainable\nmedical imaging in recent years, understanding the relationships among input\n2D image patches, input 3D image blocks, extracted feature maps, top-ranked\nfeatures, heatmaps, and final diagnosis remains a significant challenge. To help\naddress this important challenge, firstly, we create a new 2D Grad-CAM-based\nmethod using feature selection to produce explainable 2D heatmaps with a small\nnumber of highlighted image patches corresponding to top-ranked features. Secondly,\nwe design a new 2D image patch ranking algorithm that leverages the newly\ndefined feature matrices and relevant statistical data from numerous heatmaps to\nreliably rank axial patches, coronal patches, and sagittal patches. Thirdly, we create\na novel 3D image block ranking algorithm to generate a \u201cBlock Ranking Map\n(BRM)\u201d by using the axial patch ranking scores, coronal patch ranking scores, and\nsagittal patch ranking scores. Lastly, we develop a hybrid 3D image block ranking\nalgorithm to generate a reliable hybrid BRM by using different block ranking\nscores generated by the 3D image block ranking algorithm using different top feature\nsets. The associations between brain areas and a brain disease are reliably\ngenerated by using hybrid information from ChatGPT and relevant publications.\nThe simulation results using two different 3D data sets indicate that the novel hybrid\n3D image block ranking algorithm can identify top-ranked blocks associated\nwith important brain areas related to AD diagnosis and autism diagnosis. A doctor\nmay conveniently use the hybrid BRM with axial, coronal, and sagittal views\nto better understand the relationship between the top-ranked blocks and medical\ndiagnosis, and then can efficiently and effectively make a rational and explainable\nmedical diagnosis.", "keywords": [ "convolutional neural networks", "feature selection", "gradcam", "medical imaging", "disease diagnosis", "image classification" ], "rating_list": [ 3, 3, 3, 5, 3 ], "soundness_list": [ 1, 3, 3, 1, 2 ], "presentation_list": [ 1, 2, 3, 1, 1 ], "contribution_list": [ 2, 2, 2, 1, 2 ], "confidence_list": [ 3, 4, 3, 4, 2 ] }, { "paper_id": "il5yUQsrjC", "primary_area": "datasets and benchmarks", "n_author": 15, "author_info_list": [ { "name": "Alice Li", "gender": "Female", "institution": "Research, Google", "country": "", "position": "Researcher" }, { "name": "Christopher Rawles", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Daniel Kenji Toyama", "gender": "Male", "institution": "DeepMind Montreal", "country": "", "position": "Research Software Engineer" }, { "name": "Divya Tyamagundlu", "gender": "unknown", "institution": "Research, Google", "country": "", "position": "Researcher" }, { "name": "Folawiyo Campbell-Ajala", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Gabrielle Lau", "gender": "unknown", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Jonathan Waltz", "gender": "Male", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Marybeth Fair", "gender": "Female", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Oriana Riva", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Robert James Berry", "gender": "Male", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Sarah Clinckemaillie", "gender": "Female", "institution": "Research, Google", "country": "", "position": "Researcher" }, { "name": "Timothy P Lillicrap", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Research Scientist" }, { "name": "Wei Li", "gender": "Male", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "William E Bishop", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Yifan Chang", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 55, "n_ref": 142, "n_ref_all": 177, "n_fig": 9, "n_tab": 9, "L_tab": 21021, "n_element_tab": 281, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1020, "n_element_tab_1": 154, "formula_len_all": 70, "formula_len_all_1": 11, "len_all": 274272, "len_all_1": 63525, "len_abs": 1833, "len_title": 119, "len_sents": 58354, "len_sents_1": 30752, "n_sents": 442, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1434, "title": "AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents", "abs": "Autonomous agents that execute human tasks by controlling computers can enhance human productivity and application accessibility. However, progress in this field will be driven by realistic and reproducible benchmarks. We present AndroidWorld, a fully functional Android environment that provides reward signals for 116 programmatic tasks across 20 real-world Android apps. Unlike existing interactive environments, which provide a static test set, AndroidWorld dynamically constructs tasks that are parameterized and expressed in natural language in unlimited ways, thus enabling testing on a much larger and more realistic suite of tasks. To ensure reproducibility, each task includes dedicated initialization, success-checking, and tear-down logic, which modifies and inspects the device\u2019s system state.\n\nWe experiment with baseline agents to test AndroidWorld and provide initial results on the benchmark. Our best agent can complete 30.6% of AndroidWorld's tasks, leaving ample room for future work. Furthermore, we adapt a popular desktop web agent to work on Android, which we find to be less effective on mobile, suggesting future research is needed to achieve universal, cross-platform agents. Finally, we also conduct a robustness analysis, showing that task variations can significantly affect agent performance, demonstrating that without such testing, agent performance metrics may not fully reflect practical challenges.", "keywords": [ "Computer Control", "Autonomous Agents", "LLMs", "Multimodal" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ikr5XomWHS", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Arjun Krishna", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Dinesh Jayaraman", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Assistant Professor" }, { "name": "Edward Shichao Hu", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 33, "n_ref": 53, "n_ref_all": 113, "n_fig": 25, "n_tab": 1, "L_tab": 831, "n_element_tab": 101, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 316, "formula_len_all_1": 89, "len_all": 157484, "len_all_1": 62358, "len_abs": 1517, "len_title": 91, "len_sents": 64662, "len_sents_1": 32388, "n_sents": 388, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 1447, "title": "The Value of Sensory Information to a Robot", "abs": "A decision-making agent, such as a robot, must observe and react to any new task-relevant information that becomes available from its environment. We seek to study a fundamental scientific question: what value does sensory information hold to an agent at various moments in time during the execution of a task? Towards this, we empirically study agents of varying architectures, generated with varying policy synthesis approaches (imitation, RL, model-based control), on diverse robotics tasks. For each robotic agent, we characterize its regret in terms of performance degradation when state observations are withheld from it at various task states for varying lengths of time. We find that sensory information is surprisingly rarely task-critical in many commonly studied task setups. Task characteristics such as stochastic dynamics largely dictate the value of sensory information for a well-trained robot; policy architectures such as planning vs. reactive control generate more nuanced second-order effects. Further, sensing efficiency is curiously correlated with task proficiency: in particular, fully trained high-performing agents are more robust to sensor loss than novice agents early in their training. Overall, our findings characterize the tradeoffs between sensory information and task performance in practical sequential decision making tasks, and pave the way towards the design of more resource-efficient decision-making agents.", "keywords": [ "robotics", "limited sensing", "perception", "imitation learning", "reinforcement learning", "planning" ], "rating_list": [ 6, 5, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 4 ], "contribution_list": [ 2, 1, 4 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "ikkvC1UnnE", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Daogao Liu", "gender": "Male", "institution": "University of Washington, Seattle", "country": "", "position": "PhD student" }, { "name": "Kunal Talwar", "gender": "Male", "institution": "Apple", "country": "", "position": "Research Scientist" } ], "n_formula": 52, "n_formula_1": 41, "n_ref_uni": 29, "n_ref": 55, "n_ref_all": 59, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 443, "n_element_tab_1": 46, "formula_len_all": 5479, "formula_len_all_1": 4505, "len_all": 90242, "len_all_1": 54343, "len_abs": 821, "len_title": 120, "len_sents": 23372, "len_sents_1": 19968, "n_sents": 247, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 72, "L_abs": 969, "title": "Adaptive Batch Size for Privately Finding Second-Order Stationary Points", "abs": "There is a gap between finding a first-order stationary point (FOSP) and a second-order stationary point (SOSP) under differential privacy constraints, and it remains unclear whether privately finding an SOSP is more challenging than finding an FOSP. Specifically, Ganesh et al. (2023) demonstrated that an $\\alpha$-SOSP can be found with $\\alpha=\\Tilde{O}(\\frac{1}{n^{1/3}}+(\\frac{\\sqrt{d}}{n\\epsilon})^{3/7})$, where $n$ is the dataset size, $d$ is the dimension, and $\\epsilon$ is the differential privacy parameter. Building on the SpiderBoost algorithm framework, we propose a new approach that uses adaptive batch sizes and incorporates the binary tree mechanism. Our method improves the results for privately finding an SOSP, achieving $\\alpha=\\Tilde{O}(\\frac{1}{n^{1/3}}+(\\frac{\\sqrt{d}}{n\\epsilon})^{1/2})$. This improved bound matches the state-of-the-art for finding an FOSP, suggesting that privately finding an SOSP may be achievable at no additional cost.", "keywords": [ "Differential privacy", "non-convex optimization", "adaptive batch size" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 2, 2, 4, 4 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "ikhzVHXvXl", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Anjie Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Haoxuan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Jianhong Wang", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "Postdoc" }, { "name": "Jun Wang", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Professor" }, { "name": "Mengyue Yang", "gender": "Female", "institution": "University of Bristol", "country": "GB", "position": "Assistant Professor" }, { "name": "Samuel Kaski", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "Full Professor" }, { "name": "Xinrui Yang", "gender": "Male", "institution": "Department of Computer Science, University College London, University of London", "country": "GB", "position": "MS student" }, { "name": "Xu Chen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zhuohan Wang", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Undergrad student" } ], "n_formula": 35, "n_formula_1": 11, "n_ref_uni": 43, "n_ref": 69, "n_ref_all": 100, "n_fig": 11, "n_tab": 5, "L_tab": 2502, "n_element_tab": 210, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 279, "n_element_tab_1": 25, "formula_len_all": 2553, "formula_len_all_1": 817, "len_all": 192022, "len_all_1": 72872, "len_abs": 6269, "len_title": 152, "len_sents": 62335, "len_sents_1": 34554, "n_sents": 507, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1973, "title": "Attaining Human's Desirable Outcomes in Indirect Human-AI Interaction via Multi-Agent Influence Diagrams", "abs": "In human-AI interaction, one of the cutting-edge research questions is how AI agents can assist a human to attain their desirable outcomes. Most related work investigated the paradigm where a human is required to physically interact with AI agents, which we call direct human-AI interaction. However, this paradigm would be inapplicable when the scenarios are hazardous to humans, such as mine rescue and recovery. To alleviate this shortcoming, we consider indirect human-AI interaction in this paper. More detailed, a human would rely on some AI agents which we call AI proxies to interact with other AI agents, to attain the human's desirable outcomes. We model this interactive process as multi-agent influence diagrams (MAIDs), an augmentation of Bayesian networks to describe games, with Nash equilibrium (NE) as a solution. Nonetheless, in a MAID there may exist multiple NEs, and only one NE is associated with a human's desirable outcomes. To reach this optimal NE, we propose pre-strategy intervention which is an action to provide AI proxies with more information to make decision towards a human's desirable outcomes. Furthermore, we demonstrate that a team reward Markov game can be rendered as a MAID. This connection not only interprets the successes and failures of prevailing multi-agent reinforcement learning (MARL) paradigms, but also underpins the implementation of pre-strategy intervention in MARL. In practice, we incorporate pre-strategy intervention into MARL for the team reward Markov game to model the scenarios where all agents are required to achieve a common goal, with partial agents working as AI proxies to attain a human's desirable outcomes. During training, these AI proxies receive an additional reward encoding the human's desirable outcomes, and its feasibility is justified in theory. We evaluate the resulting algorithm ProxyAgent in benchmark MARL environments for teamwork, with additional goals as a human's desirable outcomes.", "keywords": [ "Human-AI Interaction", "Multi-Agent Influence Diagrams", "Multi-Agent Reinforcement Learning" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "ikSrEv8FId", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Abhronil Sengupta", "gender": "unknown", "institution": "Pennsylvania State University", "country": "", "position": "Assistant Professor" }, { "name": "Malyaban Bal", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Md Zesun Ahmed Mia", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 45, "n_ref": 77, "n_ref_all": 101, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 561, "formula_len_all_1": 604, "len_all": 145551, "len_all_1": 76622, "len_abs": 1327, "len_title": 142, "len_sents": 42846, "len_sents_1": 37649, "n_sents": 265, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1333, "title": "RMAAT: A Bio-Inspired Approach for Efficient Long-Context Sequence Processing in Transformers", "abs": "Astrocytes, an essential component of the brain's neural circuitry, demonstrate learning capabilities through bioplausible mechanisms such as presynaptic plasticity and hebbian plasticity. However, their integration into computational models remains underexplored. This paper advances astromorphic computing techniques to emulate transformer self-attention mechanisms, leveraging astrocytic nonlinearity and memory retention to improve long-range dependency processing in machine learning and natural language processing (NLP) tasks. Existing transformer models have difficulty handling lengthy contexts with thousands of tokens, even with substantial computational resources. We propose Recurrent Memory Augmented Astromorphic Transformers (RMAAT), integrating astrocytic memory and recurrent processing into self-attention, enabling longer context handling without quadratic complexity growth. Our bioplausible model has been found to outperform traditional transformers in experimental tests conducted on the Long Range Arena benchmark and IMDB dataset. Specifically, our model achieves a significant reduction in memory utilization and computational latency. This paves the way for biologically inspired AI models by illustrating how astrocytic characteristics may enhance the performance and efficiency of computational models.", "keywords": [ "Astrocyte", "Neuromorphic Computing", "Bio Inspired Learning", "Neuroscience-Algorithm-Application Codesign" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "ijwYWoChN9", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Noriaki Kawamae", "gender": "unknown", "institution": "Sophia University", "country": "JP", "position": "Lecturer" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 36, "n_ref": 47, "n_ref_all": 83, "n_fig": 1, "n_tab": 5, "L_tab": 2362, "n_element_tab": 265, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 2214, "n_element_tab_1": 260, "formula_len_all": 1152, "formula_len_all_1": 1151, "len_all": 131737, "len_all_1": 67631, "len_abs": 1237, "len_title": 86, "len_sents": 34973, "len_sents_1": 30055, "n_sents": 227, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 38, "L_abs": 1247, "title": "Domain Shift Tuning over Knowledge Gap", "abs": "This paper introduces Domain Shift Tuning (DST), a novel framework designed to guide pre-trained language models (PLMs), including Large Language Models (LLMs), in overcoming domain discrepancies (i.e., source-target).\nPLMs, pre-trained on extensive and diverse corpora, the source domain, often encounter domain gaps after fine-tuning over the target domain.\nUnlike conventional adapters or Parameter-Efficient Fine-Tuning (PEFT) methods, \nDST conceptualizes domain gaps as differences in knowledge encapsulated within multiple subnetworks of PLMs. \nTo bridge this gap, \nour challenge is to find a subnetwork set that corresponds to these pieces of knowledge and their weight.\nThis direction leads DST to employ a lightweight subnetwork, the Knowledge Steering Layer (KSL), and a training objective, Knowledge Distribution Modeling (KDM). \nThese components enable DST to fine-tune PLMs by aligning the knowledge weights of the source domain with those of the target domain. \nExperimental results on diverse datasets demonstrate that DST effectively mitigates the domain gap, allowing PLMs to generate text that closely aligns with even a small target corpus, thereby significantly enhancing domain adaptation for PLMs at lower computational cost.", "keywords": [ "PEFT", "Domain gap", "Domain Shift" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "ijbA5swmoK", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Chengchang Liu", "gender": "unknown", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jingzhao Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Lesi Chen", "gender": "Male", "institution": "Tsinghua Univeristy", "country": "CN", "position": "PhD student" } ], "n_formula": 91, "n_formula_1": 27, "n_ref_uni": 51, "n_ref": 146, "n_ref_all": 166, "n_fig": 2, "n_tab": 1, "L_tab": 137, "n_element_tab": 9, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9288, "formula_len_all_1": 3172, "len_all": 168417, "len_all_1": 63174, "len_abs": 1109, "len_title": 100, "len_sents": 41124, "len_sents_1": 23847, "n_sents": 435, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 52, "L_abs": 1240, "title": "Second-Order Min-Max Optimization with Lazy Hessians", "abs": "This paper studies second-order methods for convex-concave minimax optimization. \nMonteiro & Svaiter (2012) proposed a method to solve the problem with an optimal iteration complexity of \n$\\mathcal{O}(\\epsilon^{-3/2})$ to find an $\\epsilon$-saddle point. However, it is unclear whether the\ncomputational complexity, $\\mathcal{O}((N+ d^2) d \\epsilon^{-2/3})$, can be improved. In the above, we follow Doikov et al. (2023) and assume the complexity of obtaining a first-order oracle as $N$ and the complexity of obtaining a second-order oracle as $dN$. \nIn this paper, we show that the computation cost can be reduced by reusing Hessian across iterations. Our methods take the overall computational complexity of $\\tilde{\\mathcal{O}}( (N+d^2)(d+ d^{2/3}\\epsilon^{-2/3}))$, which improves those of previous methods by a factor of $d^{1/3}$. \nFurthermore, we generalize our method to strongly-convex-strongly-concave minimax problems and establish the complexity of $\\tilde{\\mathcal{O}}((N+d^2) (d + d^{2/3} \\kappa^{2/3}) )$ when the condition number of the problem is $\\kappa$, enjoying a similar speedup upon the state-of-the-art method. \nNumerical experiments on both real and synthetic datasets also verify the efficiency of our method.", "keywords": [ "min-max optimization; second-order methods; computational complexity" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "ijQp6HA4rK", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Ling-Hao Chen", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Shunlin Lu", "gender": "Male", "institution": "The Chinese University of HongKong, ShenZhen", "country": "CN", "position": "PhD student" }, { "name": "Wenxun Dai", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xuan Ju", "gender": "Female", "institution": "Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Lei Zhang", "gender": "Male", "institution": "International Digital Economy Academy", "country": "CN", "position": "Chief Scientist" } ], "n_formula": 9, "n_formula_1": 5, "n_ref_uni": 87, "n_ref": 180, "n_ref_all": 270, "n_fig": 68, "n_tab": 3, "L_tab": 174, "n_element_tab": 28, "n_fig_1": 25, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 365, "formula_len_all_1": 272, "len_all": 288438, "len_all_1": 71971, "len_abs": 0, "len_title": 141, "len_sents": 78586, "len_sents_1": 32017, "n_sents": 737, "n_sents_1": 303, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1426, "title": "MotionCLR: Motion Generation and Training-free Editing via Understanding Attention Mechanisms", "abs": "This research delves into the problem of interactive editing of human motion generation. Previous motion diffusion models lack explicit modeling of the word-level text-motion correspondence and good explainability, hence restricting their fine-grained editing ability. To address this issue, we propose an attention-based motion diffusion model, namely MotionCLR, with CLeaR modeling of attention mechanisms. Technically, MotionCLR models the in-modality and cross-modality interactions with self-attention and cross-attention, respectively. More specifically, the self-attention mechanism aims to measure the sequential similarity between frames and impacts the order of motion features. By contrast, the cross-attention mechanism works to find the fine-grained word-sequence correspondence and activate the corresponding timesteps in the motion sequence. Based on these key properties, we develop a versatile set of simple yet effective motion editing methods via manipulating attention maps, such as motion (de-)emphasizing, in-place motion replacement, and example-based motion generation, etc. For further verification of the explainability of the attention mechanism, we additionally explore the potential of action-counting and grounded motion generation ability via attention maps. Our experimental results show that our method enjoys good generation and editing ability with good explainability. Codes will be public.", "keywords": [ "human motion", "animation" ], "rating_list": [ 10, 3, 6, 3 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 4, 4, 2, 4 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "ijFdq8uqki", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Binjie Wang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Jiahe Jin", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Pengfei Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Associate Professor" }, { "name": "Steffi Chern", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Undergrad student" }, { "name": "Yuan Guo", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Undergrad student" }, { "name": "Yuqing Yang", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Zhulin Hu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "I-Chun Chern", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 45, "n_ref": 135, "n_ref_all": 142, "n_fig": 2, "n_tab": 5, "L_tab": 1791, "n_element_tab": 139, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1481, "n_element_tab_1": 120, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 251220, "len_all_1": 73638, "len_abs": 1507, "len_title": 103, "len_sents": 58141, "len_sents_1": 33931, "n_sents": 467, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1547, "title": "BeHonest: Benchmarking Honesty in Large Language Models", "abs": "Previous works on Large Language Models (LLMs) have mainly focused on evaluating their helpfulness or harmlessness. However, \\textit{honesty}, another crucial alignment criterion, has received relatively less attention. Dishonest behaviors in LLMs, such as spreading misinformation and defrauding users, present severe risks that intensify as these models approach superintelligent levels. Enhancing honesty in LLMs addresses critical limitations and helps uncover latent capabilities that are not readily expressed. This underscores the urgent need for reliable methods and benchmarks to effectively ensure and evaluate the honesty of LLMs.\n\nIn this paper, we introduce BeHonest, a pioneering benchmark specifically designed to assess honesty in LLMs comprehensively.\nBeHonest evaluates three essential aspects of honesty: \\emph{awareness of knowledge boundaries}, \\emph{avoidance of deceit}, and \\emph{consistency in responses}. Building on this foundation, we designed 10 scenarios to evaluate and analyze 9 popular LLMs on the market, including both closed-source and open-source models from different model families with varied model sizes. Our findings indicate that there is still significant room for improvement in the honesty of LLMs. We encourage the AI community to prioritize honesty alignment in these models, which can harness their full potential to benefit society while preventing them from causing harm through deception or inconsistency. Our benchmark and code can be found at: https://anonymous.4open.science/r/behonest-4093/.", "keywords": [ "large language models", "honesty", "benchmark" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "iinqdeuA8x", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 8, "author_info_list": [ { "name": "Can Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Chun Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Jiawei Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Sheng Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Tianjian Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Yan Feng", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhe Wang", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Zhen Zhang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" } ], "n_formula": 63, "n_formula_1": 29, "n_ref_uni": 44, "n_ref": 89, "n_ref_all": 125, "n_fig": 3, "n_tab": 3, "L_tab": 1414, "n_element_tab": 55, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6568, "formula_len_all_1": 2698, "len_all": 171203, "len_all_1": 69501, "len_abs": 1291, "len_title": 127, "len_sents": 46146, "len_sents_1": 30008, "n_sents": 465, "n_sents_1": 251, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1308, "title": "Towards Dynamic Graph Neural Networks with Provably High-Order Expressive Power", "abs": "Dynamic Graph Neural Networks (DyGNNs) have garnered increasing research attention for learning representations on evolving graphs. \nDespite their effectiveness, the limited expressive power of existing DyGNNs hinders them from capturing important evolving patterns of dynamic graphs. \nAlthough some works attempt to enhance expressive capability with heuristic features, there remains a lack of DyGNN frameworks with provable and quantifiable high-order expressive power.\nTo address this research gap, we firstly propose the k-dimensional Dynamic WL tests (k-DWL) as the referencing algorithms to quantify the expressive power of DyGNNs. We demonstrate that the expressive power of existing DyGNNs is bounded by the 1-DWL test. \nTo enhance the expressive power, we propose Dynamic Graph Neural Network with High-order expressive power (HopeDGN), which updates the representation of central node pair by aggregating the interaction history with neighbor node pairs. \nOur theoretical results demonstrate that HopeDGN can achieve expressive power equivalent to the 2-DWL test. \nWe then present a Transformer-based implementation for the local variant of \\model.\nExperimental results show that HopeDGN achieved performance improvement up to 3.12\\% on seven datasets, demonstrating the effectiveness of HopeDGN.", "keywords": [ "Graph Neural Network", "Dynamic Graph", "Expressive power" ], "rating_list": [ 3, 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 3, 3 ] }, { "paper_id": "iiK1vNRo6I", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Fuat Can Beylunioglu", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Mehrdad Pirnia", "gender": "Male", "institution": "University of Waterloo", "country": "", "position": "Lecturer" }, { "name": "P Robert Duimering", "gender": "unknown", "institution": "University of Waterloo", "country": "", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 17, "n_ref_uni": 10, "n_ref": 18, "n_ref_all": 35, "n_fig": 6, "n_tab": 3, "L_tab": 938, "n_element_tab": 93, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1614, "n_element_tab_1": 105, "formula_len_all": 2180, "formula_len_all_1": 1244, "len_all": 94658, "len_all_1": 65506, "len_abs": 1730, "len_title": 126, "len_sents": 36743, "len_sents_1": 28803, "n_sents": 256, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1781, "title": "Semi-Supervised Neural Network Model For Quadratic Multiparametric Programming", "abs": "Neural Networks (NN) with ReLU activation functions have been used as surrogate models for multiparametric quadratic problems (mp-QP) for a wide range of engineering applications. Researchers have suggested leveraging the piecewise affine property of deep NN models to solve mp-QP with linear constraints, which also exhibit piecewise affine behaviour. However, traditional deep NN applications to mp-QP fall short of providing optimal and feasible predictions, even when trained with large datasets. This study introduces a semi-supervised NN (SSNN) architecture that directly represents the mathematical structure of the global solution function. In contrast to generic NN training approaches, the proposed SSNN method derives a large proportion of model weights directly from the physical characteristics of the system, producing solutions with higher accuracy despite training on significantly smaller data sets. Since many energy management problems are formulated as QP, the proposed approach has been applied in energy systems to demonstrate proof of concept. Model performance in terms of solution accuracy and speed of the predictions was compared against a commercial solver and a generic NN model based on classical training. Results show KKT sufficient conditions for SSNN consistently outperform generic NN architectures with classical training using far less data. A similar performance advantage is shown using extreme, out-of-training distribution test data. Given its advantages of speed and reliability, the SSNN model can quickly produce optimal and feasible solutions within a second for millions of input parameters sampled from a distribution of stochastic demands and renewable generator dispatches, which can be used for simulations and long term planning.", "keywords": [ "Multiparametric Optimization", "Quadratic Programming", "Deep Neural Network", "DC-OPF", "AI for Sustainability" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "iiDioAxYah", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Yifei Li", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zeyi Xu", "gender": "Male", "institution": "Shanghai University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 39, "n_formula_1": 16, "n_ref_uni": 11, "n_ref": 21, "n_ref_all": 33, "n_fig": 7, "n_tab": 10, "L_tab": 1107, "n_element_tab": 120, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 2062, "n_element_tab_1": 146, "formula_len_all": 2251, "formula_len_all_1": 812, "len_all": 171611, "len_all_1": 56357, "len_abs": 1108, "len_title": 84, "len_sents": 45414, "len_sents_1": 25452, "n_sents": 304, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1079, "title": "Learning Physical Simulation with Message Passing Transformer", "abs": "Machine learning methods for mesh-based physical simulation have achieved significant success in recent years. We propose the Historical Message-Passing Integration Transformer (HMIT), an architecture based on Graph Neural Networks that incorporates a message passing framework and applies Graph Fourier Loss (GFL) for model optimization. To mitigate over-squashing, capture fine-grained details, and scale linearly with node count, we introduce Historical Message-Passing Attention (HMPA), which integrates multi-step historical message-passing information for each node with feature-wise softmax and employs a decoder-only architecture. Additionally, to modulate loss at specific frequencies and handle varying energy levels, we introduce GFL, which uses a frequency-domain energy adjustment schedule. To improve computational efficiency, we precompute the graph's Laplacian eigenvectors before training. Our architecture achieves significant accuracy improvements in shart- and long-term rollouts for both Lagrangian and Eulerian dynamical systems compared to current methods.", "keywords": [ "Physical Simulation", "Graph Neural Network", "Learned Simulation", "Message Passing" ], "rating_list": [ 6, 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 5, 4 ] }, { "paper_id": "ihwRfc4RNw", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Kevin Maik Jablonka", "gender": "Male", "institution": "Friedrich-Schiller Universit\u00e4t Jena", "country": "", "position": "PI" }, { "name": "Nawaf Alampara", "gender": "Male", "institution": "Friedrich-Schiller Universit\u00e4t Jena", "country": "DE", "position": "PhD student" }, { "name": "Santiago Miret", "gender": "Male", "institution": "Intel", "country": "US", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 2, "n_ref_uni": 94, "n_ref": 125, "n_ref_all": 142, "n_fig": 6, "n_tab": 4, "L_tab": 2872, "n_element_tab": 197, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2514, "n_element_tab_1": 155, "formula_len_all": 604, "formula_len_all_1": 335, "len_all": 229650, "len_all_1": 69113, "len_abs": 1994, "len_title": 123, "len_sents": 49095, "len_sents_1": 33073, "n_sents": 318, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 2004, "title": "MatText: Do Language Models Need More than Text & Scale for Materials Modeling?", "abs": "Effectively representing materials as text has the potential to leverage the vast advancements of large language models (LLMs) for discovering new materials. While LLMs have shown remarkable success in various domains, their application to materials science remains underexplored. A fundamental challenge is the lack of understanding of how to best utilize text-based representations for materials modeling. This challenge is further compounded by the absence of a comprehensive benchmark to rigorously evaluate the capabilities and limitations of these textual representations in capturing the complexity of material systems. To address this gap, we propose MatText, a suite of benchmarking tools and datasets designed to systematically evaluate the performance of language models in modeling materials. MatText encompasses nine distinct text-based representations for material systems, including several novel representations. Each representation incorporates unique inductive biases that capture relevant information and integrate prior physical knowledge about materials. Additionally, MatText provides essential tools for training and benchmarking the performance of language models in the context of materials science. These tools include standardized dataset splits for each representation across a range of dataset sizes, probes for evaluating sensitivity to geometric factors, and tools for seamlessly converting crystal structures into text. Using MatText, we conduct an extensive analysis of the capabilities of language models in modeling materials with different representations and dataset scales. Our findings reveal that current language models consistently struggle to capture the geometric information crucial for materials modeling across all representations. Instead, these models tend to leverage local information, which is emphasized in some of our novel representations. Our analysis underscores MatText's ability to reveal shortcomings of text-based methods for materials design.", "keywords": [ "Large language models", "AI for science", "Material representations", "LLMS for Materials", "Datasets and Benchmark" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 5, 4, 5 ] }, { "paper_id": "ihHeqPLRDk", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Ivan Oseledets", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "JINGWEI QIU", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jiang Yang", "gender": "unknown", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Tianchi Yu", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 67, "n_formula_1": 27, "n_ref_uni": 33, "n_ref": 81, "n_ref_all": 144, "n_fig": 11, "n_tab": 11, "L_tab": 2174, "n_element_tab": 205, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 145, "n_element_tab_1": 18, "formula_len_all": 6198, "formula_len_all_1": 1578, "len_all": 158208, "len_all_1": 56479, "len_abs": 2735, "len_title": 132, "len_sents": 44714, "len_sents_1": 25084, "n_sents": 333, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 752, "title": "Sinc Kolmogorov-Arnold Network and Its Applications on Physics-informed Neural Networks", "abs": "In this paper, we propose to use Sinc interpolation in the context of Kolmogorov-Arnold Networks, neural networks with learnable activation functions, which recently gained attention as alternatives to multilayer perceptron. Many different function representations have already been tried, but we show that Sinc interpolation proposes a viable alternative, since it is known in numerical analysis to represent well both smooth functions and functions with singularities. This is important not only for function approximation but also for the solutions of partial differential equations with physics-informed neural networks. Through a series of experiments, we show that SincKANs provide better results in almost all of the examples we have considered.", "keywords": [ "Physics-informed neural networks", "Kolmogorov-Arnold Networks", "Partial differential equations", "Computational physics." ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 2, 5, 2 ] }, { "paper_id": "ih3BJmIZbC", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Neehar Kondapaneni", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Pietro Perona", "gender": "Not Specified", "institution": "California Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Oisin Mac Aodha", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 53, "n_ref": 204, "n_ref_all": 240, "n_fig": 20, "n_tab": 1, "L_tab": 35, "n_element_tab": 4, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 629, "formula_len_all_1": 474, "len_all": 202249, "len_all_1": 69191, "len_abs": 1077, "len_title": 109, "len_sents": 70428, "len_sents_1": 32503, "n_sents": 601, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 786, "title": "Representational Similarity via Interpretable Visual Concepts", "abs": "How do two deep neural networks differ in how they arrive at a decision? Measuring deep network similarity has been a long-standing open question. Most existing metrics provide a single number to measure the similarity of two networks at a given layer, but give no insight into what makes them similar or dissimilar. We introduce an interpretable representational similarity method (RSVC) to compare two networks. We use RSVC to discover shared and unique visual concepts between two models. We show that some aspects of model differences can be attributed to unique concepts discovered by one model that are not well represented in the other. Finally, we conduct extensive evaluations across different vision model architectures and training protocols to demonstrate its effectiveness.", "keywords": [ "representational alignment", "representational similarity", "explainability", "interpretability", "comparison" ], "rating_list": [ 6, 8, 6, 8, 6 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 2, 4, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 4, 3 ] }, { "paper_id": "igpiMCYEjM", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Honghua Chen", "gender": "Male", "institution": "National Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Xingang Pan", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Yifan Zhou", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Yongwei Chen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Yushi LAN", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 33, "n_ref": 130, "n_ref_all": 150, "n_fig": 12, "n_tab": 1, "L_tab": 430, "n_element_tab": 52, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 392, "n_element_tab_1": 28, "formula_len_all": 365, "formula_len_all_1": 329, "len_all": 148973, "len_all_1": 61430, "len_abs": 599, "len_title": 136, "len_sents": 43759, "len_sents_1": 28815, "n_sents": 348, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1514, "title": "MVDrag3D: Drag-based Creative 3D Editing via Multi-view Generation-Reconstruction Prior", "abs": "Drag-based editing has become popular in 2D content creation, driven by the capabilities of image generative models. However, extending this technique to 3D remains a challenge. Existing 3D drag-based editing methods, whether employing explicit spatial transformations or relying on implicit latent optimization within limited-capacity 3D generative models, fall short in handling significant topology changes or generating new textures across diverse object categories. To overcome these limitations, we introduce MVDrag3D, a novel framework for more flexible and creative drag-based 3D editing that leverages multi-view generation and reconstruction priors.\nAt the core of our approach is the usage of a multi-view diffusion model as a strong generative prior to perform consistent drag editing over multiple rendered views, which is followed by a reconstruction model that reconstructs 3D Gaussians of the edited object. While the initial 3D Gaussians may suffer from misalignment between different views, we address this via view-specific deformation networks that adjust the position of Gaussians to be well aligned. In addition, we propose a multi-view score function that distills generative priors from multiple views to further enhance the view consistency and visual quality. Extensive experiments demonstrate that MVDrag3D provides a precise, generative, and flexible solution for 3D drag-based editing, supporting more versatile editing effects across various object categories and 3D representations.", "keywords": [ "3D Editing", "3D Drag", "Multiview Diffusion Model" ], "rating_list": [ 5, 5, 5, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 2, 5, 5, 5 ] }, { "paper_id": "igiQUYs53F", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Peng Zhang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" }, { "name": "Wenyuan Liu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "MS student" }, { "name": "Xindian Ma", "gender": "Male", "institution": "Tianjin University, Tianjin, China", "country": "", "position": "MS student" }, { "name": "Yan Wang", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 32, "n_ref": 64, "n_ref_all": 102, "n_fig": 9, "n_tab": 4, "L_tab": 2310, "n_element_tab": 282, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 2134, "n_element_tab_1": 247, "formula_len_all": 598, "formula_len_all_1": 515, "len_all": 120879, "len_all_1": 52282, "len_abs": 1466, "len_title": 174, "len_sents": 30515, "len_sents_1": 21720, "n_sents": 233, "n_sents_1": 149, "venue": "ICLR2025", "decision": "withdraw", "L_title": 125, "L_abs": 1476, "title": "CrossQuant: A Post-Training Quantization Method with Smaller Quantization Kernel for Precise Large Lanugage Model Compression", "abs": "Post-Training Quantization (PTQ) is an effective technique for compressing Large Language Models (LLMs). While many studies focus on quantizing both weights and activations, it is still a challenge to maintain the accuracy of LLM after activating quantization. To investigate the primary cause, we extend the concept of kernel from linear algebra to quantization functions to define a new term, \"quantization kernel\", which refers to the set of elements in activations that are quantized to zero. Through quantitative analysis of the quantization kernel, we find that these elements are crucial for maintaining the accuracy of quantized LLMs. With the decrease of quantization kernel, the precision of quantized LLMs increases. If the quantization kernel proportion is kept below 19\\% for OPT models and below 1\\% for LLaMA models, the precision loss from quantizing activations to INT8 becomes negligible. Motivated by the goal of developing a quantization method with small quantization kernel, we propose CrossQuant\u2014a simple yet effective method for quantizing activations. CrossQuant cross-quantizes elements using row and column-wise absolute maximum vectors, achieving a quantization kernel of approximately 16\\% for OPT models and less than 0.1\\% for LLaMA models. Experimental results on LLMs (LLaMA, OPT) ranging from 6.7B to 70B parameters demonstrate that CrossQuant improves or maintains perplexity and accuracy in language modeling, zero-shot, and few-shot tasks.", "keywords": [ "Post-Training Quantization", "Weight-Activation Quantization", "Quantization Kernel" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 3, 5 ] }, { "paper_id": "igaxFI1gBA", "primary_area": "causal reasoning", "n_author": 2, "author_info_list": [ { "name": "Ayanabha Ghosh", "gender": "Male", "institution": "Indian Institute of Technology, Jodhpur", "country": "IN", "position": "PhD student" }, { "name": "Debasis Das", "gender": "unknown", "institution": "Indian Institute of Technology, Jodhpur", "country": "", "position": "Assistant Professor" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 68, "n_ref": 89, "n_ref_all": 114, "n_fig": 8, "n_tab": 2, "L_tab": 818, "n_element_tab": 93, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 249, "n_element_tab_1": 20, "formula_len_all": 1113, "formula_len_all_1": 1004, "len_all": 149448, "len_all_1": 63426, "len_abs": 1909, "len_title": 126, "len_sents": 41819, "len_sents_1": 29962, "n_sents": 278, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1992, "title": "Temporal Causal Discovery and Generative Prediction of Vehicular CO$_2$ emission", "abs": "Global warming from greenhouse gas emissions is humanity's largest environmental hazard. Greenhouse gases, like CO$_2$ emissions from transportation, notably cars, contribute to the greenhouse effect. Effective CO$_2$ emission monitoring is needed to regulate vehicle emissions. Few studies have predicted automobile CO$_2$ emissions using OBD port data. For precise and effective prediction, the system must capture the underlying cause-effect structure between vehicular parameters that may contribute to the emission of CO$_2$ in the transportation sector. Thus, we present a causal RNN-based generative deep learning architecture that predicts vehicle CO$_2$ emissions using OBD-II data while keeping the underlying causal structure. Most widely used real-life datasets lack causal relationships between features or components, so we use our proposed architecture to discover and learn the underlying causal structure as an adjacency matrix during training and employ that during forecasting. Our framework learns a sparse adjacency matrix by imposing a sparsity-encouraging penalty on model weights and allowing some weights to be zero. This matrix is capable of capturing the causal relationships between all variable pairs. In this work, we first train the model with widely used synthetic datasets with known causal structure among variables, then we apply it to the state-of-the-art OBD-II dataset to find the internal causal structure among the vehicular parameters and perform causal inference to predict CO$_2$ emission. Experimental results reveal that our causal discovery and forecasting method surpasses state-of-the-art methods for the tasks of causal discovery in terms of AUROC, forecasting on multivariate causal time series data, and OBD-II dataset in terms of MMD, RMSE, and MAE. After successful completion, we will release the code (Code for review - \\href{https://anonymous.4open.science/r/causal-obd-co2-0A0C}{https://anonymous.4open.science/r/causal-obd-co2-0A0C}).", "keywords": [ "causal discovery", "causal inference", "time series forecasting", "vehicular emission" ], "rating_list": [ 3, 6, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 4, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "igZ5PlRB0t", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "Changpeng Cai", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "Feiyu Zhu", "gender": "Male", "institution": "Attrsense", "country": "CN", "position": "Researcher" }, { "name": "Jialin Gao", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Jiao Li", "gender": "Female", "institution": "Pingan Technology", "country": "CN", "position": "Researcher" }, { "name": "Junhao Su", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Peizhe Wang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "Xiuyuan Guo", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "MS student" }, { "name": "Yuming Zhang", "gender": "Female", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "\u4f55\u8bda\u6d69", "gender": "Male", "institution": "East China University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 15, "n_ref": 32, "n_ref_all": 44, "n_fig": 3, "n_tab": 5, "L_tab": 1287, "n_element_tab": 139, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1140, "n_element_tab_1": 118, "formula_len_all": 218, "formula_len_all_1": 219, "len_all": 107728, "len_all_1": 47913, "len_abs": 1097, "len_title": 133, "len_sents": 28339, "len_sents_1": 22394, "n_sents": 190, "n_sents_1": 146, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1100, "title": "Advancing Supervised Local Learning Beyond Classification with Long-term Feature Bank", "abs": "Local learning offers an alternative to traditional end-to-end back-propagation in deep neural networks, significantly reducing GPU memory usage. While local learning has shown promise in image classification tasks, its application to other visual tasks remains limited. This limitation arises primarily from two factors: 1) architectures tailored for classification are often not transferable to other tasks, leading to a lack of reusability of task-specific knowledge; 2) the absence of cross-scale feature communication results in degraded performance in tasks such as object detection and super-resolution. To address these challenges, we propose the Memory-augmented Auxiliary Network (MAN), which introduces a simplified design principle and incorporates a feature bank to enhance cross-task adaptability and communication. This work represents the first successful application of local learning methods beyond classification, demonstrating that MAN not only conserves GPU memory but also achieves performance on par with end-to-end approaches across multiple datasets for various visual tasks.", "keywords": [ "Local learning" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "igGeaxOiFM", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Chunyan Miao", "gender": "Female", "institution": "School of Computer Science and Engineering, Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Guang-Bin Huang", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "Full Professor" }, { "name": "Hongchao Jiang", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Junlang Qian", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Tiantong Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Wei Yang Bryan Lim", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 28, "n_formula_1": 29, "n_ref_uni": 27, "n_ref": 52, "n_ref_all": 60, "n_fig": 1, "n_tab": 4, "L_tab": 873, "n_element_tab": 107, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 438, "n_element_tab_1": 49, "formula_len_all": 1153, "formula_len_all_1": 1178, "len_all": 93800, "len_all_1": 58984, "len_abs": 1319, "len_title": 124, "len_sents": 26878, "len_sents_1": 25740, "n_sents": 201, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1326, "title": "HoLoRA: Combining Orthogonal Fine-Tuning and LoRA with Householder Reflectors", "abs": "The need for parameter-efficient fine-tuning (PEFT) has emerged as large pre-trained models are increasingly employed in specialized downstream tasks. Among PEFT methods, Low-Rank Adaptation (LoRA) is widely adopted due to its ability to fine-tune models with minimal additional parameters. However, LoRA\u2019s down-projection mechanism can lead to significant feature loss, particularly for tasks involving complex features and reasoning. This limitation poses a challenge in maintaining model performance in scenarios requiring high-dimensional representations.To address this issue, we introduce Householder Orthogonal LoRA (HoLoRA), which reparametrizes the down-projection matrix as a semi-orthogonal matrix, thereby mitigating feature loss. Our approach ensures strict orthogonality without increasing computational costs or modifying LoRA\u2019s core components. Experimental results on the GLUE benchmark show that HoLoRA consistently outperforms standard LoRA across various tasks, particularly in low-rank settings. By preserving essential features and improving fine-tuning efficiency, HoLoRA provides a robust solution to the limitations of existing PEFT methods. This advancement enhances LoRA's applicability in complex learning environments, promoting better performance in both low-budget and high-complexity scenarios.", "keywords": [ "Transfer learning", "Low Rank Adaptation", "Fine tuning", "Householder reflector", "Orthogonal fine-tuning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 3, 1, 3 ], "confidence_list": [ 5, 2, 5, 3 ] }, { "paper_id": "ig2wk7kK9J", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Chuang Gan", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Daniela Rus", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Mathias Lechner", "gender": "Unspecified", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Tsun-Hsuan Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Wei Xiao", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Ramin M. Hasani", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Researcher" } ], "n_formula": 45, "n_formula_1": 22, "n_ref_uni": 16, "n_ref": 38, "n_ref_all": 89, "n_fig": 10, "n_tab": 4, "L_tab": 1169, "n_element_tab": 119, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 590, "n_element_tab_1": 50, "formula_len_all": 4736, "formula_len_all_1": 1800, "len_all": 161490, "len_all_1": 72027, "len_abs": 940, "len_title": 111, "len_sents": 47428, "len_sents_1": 29575, "n_sents": 454, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 941, "title": "SafeDiffuser: Safe Planning with Diffusion Probabilistic Models", "abs": "Diffusion models have shown promise in data-driven planning. While these planners are commonly employed in applications where decisions are critical, they still lack established safety guarantees. In this paper, we address this limitation by introducing SafeDiffuser, a method to equip diffusion models with safety guarantees via control barrier functions. The key idea of our approach is to embed finite-time diffusion invariance, i.e., a form of specification consisting of safety constraints, into the denoising diffusion procedure. This way we enable data generation under safety constraints. We show that SafeDiffusers maintain the generative performance of diffusion models while also providing robustness in safe data generation. We evaluate our method on a series of tasks, including maze path generation, legged robot locomotion, and 3D space manipulation, and demonstrate the advantages of robustness over vanilla diffusion models.", "keywords": [ "Diffusion model", "Safety guarantees", "Planning and control" ], "rating_list": [ 5, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "iflKXk8oeg", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Ebru Aydin Gol", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "\u0130lker I\u015f\u0131k", "gender": "Male", "institution": "Novator Energy Inc.", "country": "TR", "position": "Researcher" }, { "name": "Ramazan G\u00f6kberk Cinbi\u015f", "gender": "Male", "institution": "Middle East Technical University", "country": "", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 28, "n_ref": 50, "n_ref_all": 80, "n_fig": 5, "n_tab": 6, "L_tab": 1420, "n_element_tab": 161, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 281, "n_element_tab_1": 19, "formula_len_all": 345, "formula_len_all_1": 342, "len_all": 130395, "len_all_1": 67195, "len_abs": 1111, "len_title": 129, "len_sents": 44267, "len_sents_1": 32749, "n_sents": 332, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1117, "title": "Interchangeable Token Embeddings for Extendable Vocabulary and Alpha-Equivalence", "abs": "We propose a novel approach for learning interchangeable tokens in language models to obtain an extendable vocabulary that can generalize to new tokens. Our method addresses alpha-equivalence, the principle that renaming bound variables preserves semantics. This property arises in many formal languages such as temporal logics, where all proposition symbols represent the same concept but remain distinct. To handle such tokens, we develop a dual-part embedding approach. The first part is shared across all interchangeable tokens, enforcing that they represent the same core concept. The second part is randomly generated for each token, enabling distinguishability. As a baseline, we consider a simpler approach that uses alpha-renaming for data augmentation. We also present alpha-covariance, a metric for measuring robustness against alpha-conversions. When evaluated in a Transformer encoder-decoder model for solving linear temporal logic formulae and copying with extendable vocabulary, our method demonstrates promising generalization capabilities as well as a favorable inductive bias for alpha-equivalence.", "keywords": [ "embedding methods", "interchangeable tokens", "extendable vocabulary", "transformer models", "linear temporal logic", "formal reasoning", "token generalization", "alpha-equivalence", "inductive bias", "language models", "symbolic representation", "neural networks" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 3, 3, 5, 3 ] }, { "paper_id": "ifK9NFyrhn", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Arthur Valentin", "gender": "Male", "institution": "Columbia University", "country": "", "position": "Researcher" }, { "name": "Charlotte Rochereau", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Gergo Nikolenyi", "gender": "unknown", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Mohammed AlQuraishi", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 0, "n_ref_uni": 65, "n_ref": 130, "n_ref_all": 176, "n_fig": 20, "n_tab": 13, "L_tab": 10444, "n_element_tab": 1581, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 717, "formula_len_all_1": 0, "len_all": 221384, "len_all_1": 63606, "len_abs": 1707, "len_title": 161, "len_sents": 53609, "len_sents_1": 32972, "n_sents": 378, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1714, "title": "Disconnecting The Dots: Creating Leakage-Free Protein Datasets by Removal of Densely Connected Data Points", "abs": "Biological systems arise through evolutionary processes that effectively render all biological data, at scales ranging from biomolecules to organisms, to be evolutionarily related. This poses a challenge to assessments of model generalization, as naive random splits do not safeguard against data leakage; all data points are in some sense related, and their degree of relatedness lies on a continuum. To address this challenge, various similarity metrics are typically used to cluster data prior to splitting to ensure dissimilarity of resulting partitions. However, as we show in this study, similarity thresholds that lead to well-behaved splits (large numbers of homogeneously sized clusters) must invariably be too permissive, thus only permitting assessment of weak generalization. Conversely, stringent thresholds that could in principle enable assessment of strong generalization typically fail to produce well-separated clusters, yielding one or a handful of very large clusters that span the entire dataset. Here, we propose a new data splitting methodology that optimally balances these competing considerations by relaxing the assumption that all data points must be retained. Instead, through a principled and judicious removal of highly central data points, our approach yields well-behaved data splits that enable assessment of extreme generalization regimes. We demonstrate its utility by investigating the impact of diverse proteins representations on protein function prediction. Our experiments confirm the robustness of our new methodology and provide insights into the utility and behavior of protein representations under previously untested regimes of sequence and structure generalization.", "keywords": [ "data splitting", "clustering", "biology", "protein function prediction", "protein representations" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "ifJFKbSZxS", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "C\u00f4me Bissuel", "gender": "Male", "institution": "EDF", "country": "", "position": "Researcher" }, { "name": "Emmanuel Rachelson", "gender": "Male", "institution": "Institut Sup\u00e9rieur de l'A\u00e9ronautique et de l'Espace", "country": "FR", "position": "Full Professor" }, { "name": "Olivier Juan", "gender": "Male", "institution": "EDF R&D", "country": "", "position": "Researcher" }, { "name": "Paul STRANG", "gender": "unknown", "institution": "Conservatoire national des arts et m\u00e9tiers", "country": "FR", "position": "PhD student" }, { "name": "Safia Kedad-Sidhoum", "gender": "unknown", "institution": "Conservatoire National des Arts et M\u00e9tiers", "country": "", "position": "Full Professor" }, { "name": "Zacharie ALES", "gender": "Male", "institution": "ENSTA Paris", "country": "FR", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 15, "n_ref_uni": 44, "n_ref": 92, "n_ref_all": 118, "n_fig": 3, "n_tab": 8, "L_tab": 7478, "n_element_tab": 1043, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2264, "formula_len_all_1": 1336, "len_all": 176104, "len_all_1": 59568, "len_abs": 918, "len_title": 118, "len_sents": 45318, "len_sents_1": 26629, "n_sents": 293, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 926, "title": "A Markov decision process for variable selection in Branch and bound", "abs": "Mixed-Integer Linear Programming (MILP) is a powerful framework used to address a wide range of NP-hard combinatorial optimization problems, often solved by Branch and bound (B\\&B). A key factor influencing the performance of B\\&B solvers is the variable selection heuristic governing branching decisions. Recent contributions have sought to adapt reinforcement learning (RL) algorithms to the B\\&B setting to learn optimal branching policies, through Markov Decision Processes (MDP) inspired formulations, and ad hoc convergence theorems and algorithms. In this work, we introduce B\\&B MDPs, a principled vanilla MDP formulation for variable selection in B\\&B, allowing to leverage a broad range of RL algorithms for the purpose of learning optimal B\\&B heuristics. Computational experiments validate our model empirically, as our branching agent outperforms prior state-of-the-art RL agents on four standard MILP benchmarks.", "keywords": [ "Mixed-integer linear programming; Branch and bound; Reinforcement learning; Markov decision process" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "if8iIYcmVC", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Minghang Zheng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Qingchao Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xinhao Cai", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Yang Liu", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuxin Peng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 22, "n_ref": 96, "n_ref_all": 135, "n_fig": 17, "n_tab": 12, "L_tab": 2037, "n_element_tab": 301, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 996, "n_element_tab_1": 124, "formula_len_all": 260, "formula_len_all_1": 184, "len_all": 140972, "len_all_1": 71881, "len_abs": 1633, "len_title": 112, "len_sents": 49201, "len_sents_1": 33672, "n_sents": 396, "n_sents_1": 251, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1640, "title": "Pose-guided Motion Diffusion Model for Text-to-motion Generation", "abs": "3D Human motion generation, especially textual conditioning motion generation, is a vital part of computer animation. However, during training, multiple actions are often coupled within a single textual description, which complicates the model's learning of individual actions. Additionally, the motion corresponding to a given text can be diverse, which makes it difficult for the model learning and for the user to control the generation of motions that contain a specific pose. Finally, motions with the same semantics can have various ways of expression in the forms of texts, which further increases the difficulty of the model\u2019s learning process. To solve the above challenges, we propose the Pose-Guided Text to Motion (PG-T2M) with the following designs. Firstly, we propose to divide the sentences into sub-sentences containing one single verb and make the model learn the specific mapping from one single action description to its motion. Secondly, we propose using pose priors from static 2D natural images for each sub-sentence as control signals, allowing the model to generate more accurate and controllable 3D pose sequences that align with the sub-action descriptions. Finally, to enable the model to distinguish which sub-sentences describe similar semantics, we construct a pose memory storing semantic-similar sub-sentences and the corresponding pose representations in groups. These designs together enable our model to retrieve the pose information for every single action described in the text and use them to guide motion generation. Our method achieves state-of-the-art performance on the HumanML3D and KIT datasets.", "keywords": [ "Motion generation", "Text to motion", "Diffusion model", "Generative model" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 5, 5, 5 ] }, { "paper_id": "iezDdA9oeB", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Jiangchao Yao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Kaiyuan Gao", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Lijun Wu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Tao Qin", "gender": "Male", "institution": "Microsoft Research Asia", "country": "", "position": "Principal Researcher" }, { "name": "Zizhuo Zhang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "PhD student" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 10, "n_ref_uni": 45, "n_ref": 122, "n_ref_all": 148, "n_fig": 9, "n_tab": 9, "L_tab": 3697, "n_element_tab": 414, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1652, "n_element_tab_1": 218, "formula_len_all": 1581, "formula_len_all_1": 982, "len_all": 177793, "len_all_1": 64673, "len_abs": 1797, "len_title": 88, "len_sents": 54824, "len_sents_1": 29074, "n_sents": 392, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1818, "title": "Fast and Accurate Blind Flexible Docking", "abs": "Molecular docking that predicts the bound structures of small molecules (ligands) to their protein targets, plays a vital role in drug discovery. However, existing docking methods often face limitations: they either overlook crucial structural changes by assuming protein rigidity or suffer from low computational efficiency due to their reliance on generative models for structure sampling. To address these challenges, we propose FABFlex, a fast and accurate regression-based multi-task learning model designed for realistic blind flexible docking scenarios, where proteins exhibit flexibility and binding pocket sites are unknown (blind). Specifically, FABFlex's architecture comprises three specialized modules working in concert: (1) A pocket prediction module that identifies potential binding sites, addressing the challenges inherent in blind docking scenarios. (2) A ligand docking module that predicts the bound (holo) structures of ligands from their unbound (apo) states. (3) A pocket docking module that forecasts the holo structures of protein pockets from their apo conformations. Notably, FABFlex incorporates an iterative update mechanism that serves as a conduit between the ligand and pocket docking modules, enabling continuous structural refinements. This approach effectively integrates the three subtasks of blind flexible docking\u2014pocket identification, ligand conformation prediction, and protein flexibility modeling\u2014into a unified, coherent framework. Extensive experiments on public benchmark datasets demonstrate that FABFlex not only achieves superior effectiveness in predicting accurate binding modes but also exhibits a significant speed advantage (208$\\times$) compared to existing state-of-the-art methods. Our code is released at~\\url{https://anonymous.4open.science/r/FABFlex-7007}.", "keywords": [ "Blind Flexible Molecular Docking", "Structure Prediction", "AI4Science", "Ligand-Protein Graph" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "idnMNjlEj5", "primary_area": "applications to robotics, autonomy, planning", "n_author": 11, "author_info_list": [ { "name": "Akira Kinose", "gender": "Male", "institution": "Panasonic Connect Co., Ltd.", "country": "", "position": "Researcher" }, { "name": "Felix Wick", "gender": "Male", "institution": "Panasonic R&D Center Germany GmbH", "country": "DE", "position": "Principal Researcher" }, { "name": "Jayashree Karlekar", "gender": "Female", "institution": "Panasonic Research and Development Center Singapore", "country": "SG", "position": "Principal Researcher" }, { "name": "Koki Oguri", "gender": "unknown", "institution": "Panasonic Connect", "country": "", "position": "Researcher" }, { "name": "Lakshmi Subramanian", "gender": "unknown", "institution": "Panasonic R&D Center Singapore", "country": "SG", "position": "Researcher" }, { "name": "Natsuki Murakami", "gender": "Female", "institution": "Ochanomizu Women's University", "country": "JP", "position": "PhD student" }, { "name": "Sugiri Pranata", "gender": "Male", "institution": "Panasonic R&D Center Singapore", "country": "SG", "position": "Researcher" }, { "name": "Thong Jing Yuan", "gender": "Male", "institution": "College of Computing, Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Tomoyuki Kagaya", "gender": "Male", "institution": "Panasonic Connect", "country": "JP", "position": "Researcher" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" }, { "name": "Yuxuan Lou", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 17, "n_ref": 20, "n_ref_all": 34, "n_fig": 6, "n_tab": 7, "L_tab": 3333, "n_element_tab": 159, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 2475, "n_element_tab_1": 48, "formula_len_all": 469, "formula_len_all_1": 436, "len_all": 105400, "len_all_1": 55199, "len_abs": 1821, "len_title": 98, "len_sents": 34600, "len_sents_1": 25048, "n_sents": 247, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1831, "title": "EnvBridge: Bridging Diverse Environments with Cross-Environment Knowledge Transfer for Embodied AI", "abs": "In recent years, Large Language Models (LLMs) have demonstrated high reasoning capabilities, drawing attention for their applications as agents in various decision-making processes. One notably promising application of LLM agents is robotic manipulation. Recent research has shown that LLMs can generate text planning or control code for robots, providing substantial flexibility and interaction capabilities.\nHowever, these methods still face challenges in terms of flexibility and applicability across different environments, limiting their ability to adapt autonomously. Current approaches typically fall into two categories: those relying on environment-specific policy training, which restricts their transferability, and those generating code actions based on fixed prompts, which leads to diminished performance when confronted with new environments. These limitations significantly constrain the generalizability of agents in robotic manipulation.\nTo address these limitations, we propose a novel method called EnvBridge. This approach involves the retention and transfer of successful robot control codes from source environments to target environments. EnvBridge enhances the agent's adaptability and performance across diverse settings by leveraging insights from multiple environments. Notably, our approach alleviates environmental constraints, offering a more flexible and generalizable solution for robotic manipulation tasks.\nWe validated the effectiveness of our method using robotic manipulation benchmarks: RLBench, MetaWorld, and CALVIN. Our experiments demonstrate that LLM agents can successfully leverage diverse knowledge sources to solve complex tasks. Consequently, our approach significantly enhances the adaptability and robustness of robotic manipulation agents in planning across diverse environments.", "keywords": [ "LLM Agent", "Robotic Manipulation", "Cross-Environment Knowledge Transfer" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 1, 4 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "icVRZJTK9v", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Eduard Gorbunov", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "Nazarii Tupitsa", "gender": "unknown", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "Samuel Horv\u00e1th", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Assistant Professor" }, { "name": "Martin Takac", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Associate Professor" } ], "n_formula": 39, "n_formula_1": 22, "n_ref_uni": 70, "n_ref": 105, "n_ref_all": 128, "n_fig": 11, "n_tab": 2, "L_tab": 341, "n_element_tab": 2, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2789, "formula_len_all_1": 1192, "len_all": 176291, "len_all_1": 66275, "len_abs": 961, "len_title": 105, "len_sents": 45387, "len_sents_1": 30442, "n_sents": 366, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 964, "title": "Federated Learning Can Find Friends That Are Advantageous", "abs": "In Federated Learning (FL), the distributed nature and heterogeneity of client data present both opportunities and challenges. While collaboration among clients can significantly enhance the learning process, not all collaborations are beneficial; some may even be detrimental. In this study, we introduce a novel algorithm that assigns adaptive aggregation weights to clients participating in FL training, identifying those with data distributions most conducive to a specific learning objective. We demonstrate that our aggregation method converges no worse than the method that aggregates only the updates received from clients with the same data distribution. Furthermore, empirical evaluations consistently reveal that collaborations guided by our algorithm outperform traditional FL approaches. This underscores the critical role of judicious client selection and lays the foundation for more streamlined and effective FL implementations in the coming years.", "keywords": [ "Federated Learning", "Distributed Learning", "Personalized Learning" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 2, 5 ] }, { "paper_id": "icUCCz8pAu", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Chejian Xu", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Dawn Song", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 20, "n_ref": 32, "n_ref_all": 45, "n_fig": 2, "n_tab": 6, "L_tab": 1490, "n_element_tab": 137, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 738, "n_element_tab_1": 18, "formula_len_all": 381, "formula_len_all_1": 346, "len_all": 116225, "len_all_1": 67219, "len_abs": 2190, "len_title": 92, "len_sents": 37414, "len_sents_1": 34101, "n_sents": 265, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 2199, "title": "MultiTrust: Enhancing Safety and Trustworthiness of Large Language Models from Multiple Perspectives", "abs": "Large Language Models (LLMs) have shown impressive performance across various tasks, yet they still face significant safety and trustworthiness challenges, such as robustness, fairness, and truthfulness. Addressing these challenges is critical for the reliable deployment of LLMs. Directly fine-tuning LLMs to enhance safety can degrade their performance and is challenging to balance across multiple safety perspectives due to the forgetting phenomenon. In this paper, we propose MultiTrust, a novel and scalable framework designed to enhance LLM safety from multiple safety perspectives. In particular, MultiTrust first generates challenging training data through adversarial optimizations, focusing on LLMs trustworthiness perspectives, such as robustness, fairness, and safety. MultiTrust then separately train safety auxiliary models for each perspective using supervised fine-tuning and Direct Preference Optimization (DPO). MultiTrust augments a base model with these safety auxiliary models on the fly through dynamic routing and logit ensembling, significantly boosting the performance across different trustworthiness metrics for the base model while preserving its helpfulness. Notably, MultiTrust introduces an effective perplexity-based inference-time router to seamlessly integrate these safety auxiliary models by averaging the logit outputs of the selected safety auxiliary model and the base model, which enhances the stability of the final performance. Moreover, MultiTrust's flexible design allows for the augmentation with new safety auxiliary models for different perspectives without necessitating additional training or adaptation. Extensive experimental results show that MultiTrust, which trains a series of 7B safety auxiliary models, significantly improves the trustworthiness of the base LLM across different sizes (7B and 13B). For instance, MultiTrust increased the average performance of Llama2-13B from 35.54% to 51.14%, and Vicuna-13B from 29.91% to 52.82%, outperforming models with similar and even larger sizes across different perspectives. These results underscore the effectiveness and scalability of MultiTrust in enhancing the safety and reliability of LLMs.", "keywords": [ "Large Language Models", "Safety", "Trustworthiness", "Robustness", "Fairness", "Truthfulness" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 1, 1, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "icNel2Thrt", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Denis Melanson", "gender": "Male", "institution": "University of Waterloo", "country": "", "position": "MS student" }, { "name": "Gavin Crooks", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "Lecturer" }, { "name": "Kaelan Donatella", "gender": "unknown", "institution": "Normal Computing", "country": "US", "position": "Researcher" }, { "name": "Maxwell Aifer", "gender": "Male", "institution": "University of Maryland, Baltimore County", "country": "US", "position": "PhD student" }, { "name": "Patrick J. Coles", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Samuel Duffield", "gender": "unknown", "institution": "Normal Computing", "country": "US", "position": "Researcher" } ], "n_formula": 36, "n_formula_1": 24, "n_ref_uni": 41, "n_ref": 71, "n_ref_all": 90, "n_fig": 6, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2513, "formula_len_all_1": 1605, "len_all": 130277, "len_all_1": 61747, "len_abs": 1199, "len_title": 86, "len_sents": 40506, "len_sents_1": 29415, "n_sents": 284, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 38, "L_abs": 1205, "title": "Thermodynamic Natural Gradient Descent", "abs": "Second-order training methods have better convergence properties than gradient descent but are rarely used in practice for large-scale training due to their computational overhead. This can be viewed as a hardware limitation (imposed by digital computers). Here we show that natural gradient descent (NGD), a second-order method, can have a similar computational complexity per iteration to a first-order method, when employing appropriate hardware. We present a new hybrid digital-analog algorithm for training neural networks that is equivalent to NGD in a certain parameter regime but avoids prohibitively costly linear system solves. Our algorithm exploits the thermodynamic properties of an analog system at equilibrium, and hence requires an analog thermodynamic computer. The training occurs in a hybrid digital-analog loop, where the gradient and Fisher information matrix (or any other positive semi-definite curvature matrix) are calculated at given time intervals while the analog dynamics take place. We numerically demonstrate the superiority of this approach over state-of-the-art digital first- and second-order training methods on classification tasks and language model fine-tuning tasks.", "keywords": [ "optimization", "gradient descent", "natural gradient descent", "thermodynamic computing", "analog computing" ], "rating_list": [ 8, 3, 3 ], "soundness_list": [ 4, 3, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 4, 2, 1 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "icDoYdUhRa", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Nikolaos Dimitriadis", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Pascal Frossard", "gender": "Not Specified", "institution": "EPFL", "country": "CH", "position": "Full Professor" }, { "name": "Francois Fleuret", "gender": "Male", "institution": "University of Geneva", "country": "", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 63, "n_ref": 170, "n_ref_all": 209, "n_fig": 21, "n_tab": 4, "L_tab": 2400, "n_element_tab": 321, "n_fig_1": 12, "n_tab_1": 3, "L_tab_1": 2840, "n_element_tab_1": 282, "formula_len_all": 1468, "formula_len_all_1": 1624, "len_all": 184395, "len_all_1": 70255, "len_abs": 1414, "len_title": 72, "len_sents": 41486, "len_sents_1": 29070, "n_sents": 308, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1420, "title": "Pareto Low-Rank Adapters: Efficient Multi-Task Learning with Preferences", "abs": "Multi-task trade-offs in machine learning can be addressed via Pareto Front Learning (PFL) methods that parameterize the Pareto Front (PF) with a single model. PFL permits to select the desired operational point during inference, contrary to traditional Multi-Task Learning (MTL) that optimizes for a single trade-off decided prior to training. However, recent PFL methodologies suffer from limited scalability, slow convergence, and excessive memory requirements, while exhibiting inconsistent mappings from preference to objective space. We introduce PaLoRA, a novel parameter-efficient method that addresses these limitations in two ways. First, we augment any neural network architecture with task-specific low-rank adapters and continuously parameterize the Pareto Front in their convex hull. Our approach steers the original model and the adapters towards learning general and task-specific features, respectively. Second, we propose a deterministic sampling schedule of preference vectors that reinforces this division of labor, enabling faster convergence and strengthening the validity of the mapping from preference to objective space throughout training. Our experiments show that PaLoRA outperforms state-of-the-art MTL and PFL baselines across various datasets, scales to large networks, reducing the memory overhead $23.8-31.7$ times compared with competing PFL baselines in scene understanding benchmarks.", "keywords": [ "Pareto Front Learning", "multi-task learning", "low rank" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "ic153qXFfx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Elliot L Epstein", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Hamid Palangi", "gender": "Male", "institution": "Google", "country": "US", "position": "Staff Research Scientist" }, { "name": "Kaisheng Yao", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Xinyi Bai", "gender": "Female", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "JING LI", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 25, "n_ref": 56, "n_ref_all": 97, "n_fig": 15, "n_tab": 9, "L_tab": 4830, "n_element_tab": 483, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1395, "n_element_tab_1": 119, "formula_len_all": 566, "formula_len_all_1": 432, "len_all": 164500, "len_all_1": 54711, "len_abs": 1757, "len_title": 124, "len_sents": 49218, "len_sents_1": 24874, "n_sents": 357, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1889, "title": "MMMT-IF: A Challenging Multimodal Multi-Turn Instruction Following Benchmark", "abs": "Evaluating instruction following capabilities for multimodal, multi-turn dialogue is challenging. With potentially multiple instructions in the input model context, the task is time-consuming for human raters and we show LLM based judges are biased towards answers from the same model. \nWe propose MMMT-IF, an image based multi-turn Q\\&A evaluation set with added global instructions between questions, constraining the answer format.\nThis challenges models to retrieve instructions dispersed across long dialogues and reason under instruction constraints.\nAll instructions are objectively verifiable through code execution.\nWe introduce the Programmatic Instruction Following ($\\operatorname{PIF}$) metric to measure the fraction of the instructions that are correctly followed while performing a reasoning task.\nThe $\\operatorname{PIF-N-K}$ set of metrics further evaluates robustness by measuring the fraction of samples in a corpus where, for each sample, at least K out of N generated model responses achieve a $\\operatorname{PIF}$ score of one.\nThe $\\operatorname{PIF}$ metric aligns with human instruction following ratings, showing 60 percent correlation.\nExperiments show Gemini 1.5 Pro, GPT-4o, and Claude 3.5 Sonnet, have a $\\operatorname{PIF}$ metric that drops from 0.81 on average at turn 1 across the models, to 0.64 at turn 20.\nAcross all turns, when each response is repeated 4 times ($\\operatorname{PIF-4-4}$), GPT-4o and Gemini successfully follow all instructions only 11\\% of the time.\nWhen all the instructions are also appended to the end of the model input context, the $\\operatorname{PIF}$ metric improves by 22.3 points on average, showing that the challenge with the task lies not only in following the instructions, but also in retrieving the instructions spread out in the model context. \nWe plan to open source the MMMT-IF dataset and metric computation code.", "keywords": [ "Large Language Models", "Foundation Models", "Instruction Following", "Program Verifiable Instructions", "Multimodal", "AI", "Long Context" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 4, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "iamWnRpMuQ", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Qi Qi", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Xing Tang", "gender": "Male", "institution": "FiT,Tencent", "country": "CN", "position": "Researcher" }, { "name": "Xu Chen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yankai Lin", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "Assistant Professor" }, { "name": "Yiju Guo", "gender": "unknown", "institution": "Renmin University of China", "country": "", "position": "PhD student" }, { "name": "Zexu Sun", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "xiuqiang He", "gender": "Male", "institution": "Shenzhen Technology University", "country": "CN", "position": "Full Professor" } ], "n_formula": 23, "n_formula_1": 23, "n_ref_uni": 40, "n_ref": 64, "n_ref_all": 82, "n_fig": 7, "n_tab": 4, "L_tab": 2200, "n_element_tab": 544, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1251, "n_element_tab_1": 339, "formula_len_all": 1397, "formula_len_all_1": 1397, "len_all": 125002, "len_all_1": 68443, "len_abs": 1585, "len_title": 127, "len_sents": 30601, "len_sents_1": 27931, "n_sents": 261, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1641, "title": "Policy-aware Reward Modeling with Uncertainty-Gradient based Data Augmentation", "abs": "Reinforcement Learning from Human Feedback (RLHF) has emerged as a standard and effective approach for training large language models (LLMs) with human preferences. In this framework, a learned reward model approximates human preferences and guides policy optimization, making it crucial to develop an accurate reward model. However, without the ``true'' reward function, challenges arise when the reward model is an imperfect proxy for human preference. Since the policy optimization continuously shifts the human preference training dataset's distribution. The fixed reward model suffers from this problem of off-distribution, especially the on policy methods. While collecting new preference data can mitigate this issue, it is costly and challenging to optimize. Thus, reusing the policy interaction samples becomes a possible way to further refine the reward model. To tackle these challenges, we introduce a novel method \\textbf{U}ncertainty-\\textbf{G}radient based \\textbf{D}ata \\textbf{A}ugmentation (\\textbf{UGDA} for short) to enhance reward modeling by leveraging policy samples to maintain on-distribution performance. Specifically, UGDA selects interaction samples based on the uncertainty of the reward ensembles and the gradient based influence of policy optimization. After the reward relabeling of selected samples, we use supervised learning to refine the reward ensembles, then get the retrained policy. Extensive experiments demonstrate that by leveraging UGDA to select a few samples without the costly human preference data collection, we can improve the ability of the policy and surpass the state-of-the-art methods.", "keywords": [ "Reward Modeling", "Large Language Model", "Data Augmentation" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "iaP7yHRq1l", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Duxin Chen", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "He Wang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Huiyang Yi", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "PhD student" }, { "name": "Mingyu Kang", "gender": "Male", "institution": "Southeast University", "country": "", "position": "PhD student" }, { "name": "Wenwu Yu", "gender": "Male", "institution": "Southeast University", "country": "", "position": "Full Professor" }, { "name": "He Yanyan", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "PhD student" } ], "n_formula": 55, "n_formula_1": 11, "n_ref_uni": 56, "n_ref": 152, "n_ref_all": 197, "n_fig": 12, "n_tab": 28, "L_tab": 14607, "n_element_tab": 1462, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3510, "n_element_tab_1": 73, "formula_len_all": 3302, "formula_len_all_1": 1325, "len_all": 242204, "len_all_1": 61619, "len_abs": 1573, "len_title": 123, "len_sents": 57228, "len_sents_1": 27816, "n_sents": 493, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1252, "title": "THE ROBUSTNESS OF DIFFERENTIABLE CAUSAL DISCOVERY IN MISSPECIFIED SCENARIOS", "abs": "Causal discovery aims to learn causal relationships between variables from targeted data, making it a fundamental task in machine learning. However, causal discovery algorithms often rely on unverifiable causal assumptions, which are usually difficult to satisfy in real-world data, thereby limiting the broad application of causal discovery in practical scenarios. Inspired by these considerations, this work extensively benchmarks the empirical performance of various mainstream causal discovery algorithms, which assume i.i.d. data, under eight model assumption violations. Our experimental results show that differentiable causal discovery methods exhibit robustness under the metrics of Structural Hamming Distance and Structural Intervention Distance of the inferred graphs in commonly used challenging scenarios, except for scale variation. We also provide the theoretical explanations for the performance of differentiable causal discovery methods. Finally, our work aims to comprehensively benchmark the performance of recent differentiable causal discovery methods under model assumption violations, and provide the standard for reasonable evaluation of causal discovery, as well as to further promote its application in real-world scenarios.", "keywords": [ "Differentiable causal discovery", "model assumption violations", "benchmark" ], "rating_list": [ 3, 8, 6, 5 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "iaHghgG8NR", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Ali Behrouz", "gender": "Male", "institution": "Google", "country": "US", "position": "Intern" }, { "name": "Ali Parviz", "gender": "unknown", "institution": "New Jersey Institute of Technology, MILA", "country": "", "position": "PhD student" }, { "name": "Bryan Perozzi", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Mahdi Karami", "gender": "Male", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Vahab Mirrokni", "gender": "Male", "institution": "Google Research", "country": "", "position": "Distinguished Scientist" }, { "name": "Clayton Hendrick Sanford", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 60, "n_formula_1": 12, "n_ref_uni": 86, "n_ref": 207, "n_ref_all": 243, "n_fig": 5, "n_tab": 10, "L_tab": 5779, "n_element_tab": 743, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1431, "n_element_tab_1": 302, "formula_len_all": 3540, "formula_len_all_1": 663, "len_all": 324993, "len_all_1": 80274, "len_abs": 1783, "len_title": 97, "len_sents": 88204, "len_sents_1": 36580, "n_sents": 711, "n_sents_1": 280, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1788, "title": "Towards an Understanding of Graph Sequence Models", "abs": "Modern sequence models (e.g., Transformers, linear RNNs, etc.) emerged as dominant backbones of recent deep learning frameworks, mainly due to their efficiency, representational power, and/or ability to capture long-range dependencies. Adopting these sequence models for graph-structured data has recently gained popularity as the alternative to Message Passing Neural Networks (MPNNs). There is, however, a lack of a common foundation about what constitutes a good graph sequence model, and a mathematical description of the benefits and deficiencies in adopting different sequence models for learning on graphs. To this end, we first present Graph Sequence Model (GSM), a unifying framework for adopting sequence models for graphs, consisting of three main steps: (1) Tokenization, which translates the graph into a set of sequences; (2) Local Encoding, which encodes local neighborhoods around each node; and (3) Global Encoding, which employs a scalable sequence model to capture long-range dependencies within the sequences. This framework allows us to understand, evaluate, and compare the power of different sequence model backbones in graph tasks. Our theoretical evaluations of the representation power of Transformers and modern recurrent models through the lens of global and local graph tasks show that there are both negative and positive sides for both types of models. Building on this observation, we present GSM++, a fast hybrid model that uses the Hierarchical Affinity Clustering (HAC) algorithm to tokenize the graph into hierarchical sequences, and then employs a hybrid architecture of Transformer to encode these sequences. Our theoretical and experimental results support the design of GSM++, showing that GSM++ outperforms baselines in most benchmark evaluations.", "keywords": [ "Graph Learning", "Sequence Models", "Graph Transformers", "Hierarchical Clustering" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 3, 1, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "iZl0VqEdxa", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Ender Konukoglu", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Associate Professor" }, { "name": "Jean-Philippe Thiran", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Full Professor" }, { "name": "Mael Macuglia", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Natasa Tagasovska", "gender": "Female", "institution": "Prescient Design - Genentech, Roche", "country": "", "position": "Senior Machine Learning Scientis" }, { "name": "Reto Sutter", "gender": "unknown", "institution": "University of Zurich", "country": "", "position": "Associate Professor" }, { "name": "Sebastiano Caprara", "gender": "Male", "institution": "University of Zurich", "country": "", "position": "Lecturer" }, { "name": "Anna Klimovskaia", "gender": "Female", "institution": "Google", "country": "CH", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 4, "n_ref_uni": 59, "n_ref": 100, "n_ref_all": 123, "n_fig": 9, "n_tab": 6, "L_tab": 2484, "n_element_tab": 158, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 891, "n_element_tab_1": 55, "formula_len_all": 527, "formula_len_all_1": 201, "len_all": 182566, "len_all_1": 67028, "len_abs": 1294, "len_title": 102, "len_sents": 47932, "len_sents_1": 32896, "n_sents": 342, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1299, "title": "Uncertainty modeling for fine-tuned implicit functions", "abs": "Implicit functions such as Neural Radiance Fields (NeRFs), occupancy networks, and signed distance functions (SDFs) have become pivotal in computer vision for reconstructing detailed object shapes from sparse views. Achieving optimal performance with these models can be challenging due to the extreme sparsity of inputs and distribution shifts induced by data corruptions. To this end, large, noise-free synthetic datasets can serve as shape priors to help models fill in gaps, but the resulting reconstructions must be approached with caution. Uncertainty estimation is crucial for assessing the quality of these reconstructions, particularly in identifying areas where the model is uncertain about the parts it has inferred from the prior. In this paper, we introduce Dropsembles, a novel method for uncertainty estimation in tuned implicit functions. We demonstrate the efficacy of our approach through a series of experiments, starting with toy examples and progressing to a real-world scenario. Specifically, we train a Convolutional Occupancy Network on synthetic anatomical data and test it on low-resolution MRI segmentations of the lumbar spine. Our results show that Dropsembles achieve the accuracy and calibration levels of deep ensembles but with significantly less computational cost.", "keywords": [ "uncertainty", "implicit functions", "3D reconstruction", "occupancy networks" ], "rating_list": [ 3, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "iZeQBqJamf", "primary_area": "foundation or frontier models, including LLMs", "n_author": 25, "author_info_list": [ { "name": "Achal Dave", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Alex Dimakis", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department, University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Alex Fang", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Gabriel Ilharco", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "", "position": "PhD student" }, { "name": "Georgios Smyrnis", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Igor Vasiljevic", "gender": "Male", "institution": "Toyota Research Institute", "country": "US", "position": "Research Scientist" }, { "name": "Jean Mercat", "gender": "Male", "institution": "Toyota Research Institute", "country": "US", "position": "Researcher" }, { "name": "Jeffrey Li", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Jenia Jitsev", "gender": "unknown", "institution": "Juelich Supercomputing Center, Research Center Juelich", "country": "", "position": "Senior Scientist" }, { "name": "Luca Soldaini", "gender": "Non-Binary", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Ludwig Schmidt", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Marianna Nezhurina", "gender": "unknown", "institution": "LAION", "country": "", "position": "Researcher" }, { "name": "Mitchell Wortsman", "gender": "Male", "institution": "University of Washington, Seattle", "country": "", "position": "PhD student" }, { "name": "Niklas Muennighoff", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Pang Wei Koh", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Visiting Research Scientist" }, { "name": "Reinhard Heckel", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Associate Professor" }, { "name": "Rui Xin", "gender": "Male", "institution": "University of Washington", "country": "", "position": "PhD student" }, { "name": "Rulin Shao", "gender": "unknown", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Samir Yitzhak Gadre", "gender": "Male", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Shuran Song", "gender": "Female", "institution": "Columbia University", "country": "", "position": "Assistant Professor" }, { "name": "Suchin Gururangan", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Thomas Kollar", "gender": "Male", "institution": "Wayve", "country": "US", "position": "Principal Researcher" }, { "name": "Yair Carmon", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Assistant Professor" }, { "name": "Sedrick Scott Keh", "gender": "Male", "institution": "Toyota Research Institute", "country": "", "position": "Research Engineer" }, { "name": "vaishaal naanny shankar", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 35, "n_ref": 75, "n_ref_all": 127, "n_fig": 11, "n_tab": 2, "L_tab": 201, "n_element_tab": 25, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 386, "n_element_tab_1": 63, "formula_len_all": 339, "formula_len_all_1": 200, "len_all": 290105, "len_all_1": 67011, "len_abs": 1455, "len_title": 121, "len_sents": 34695, "len_sents_1": 31263, "n_sents": 282, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1386, "title": "Language models scale reliably with over-training and on downstream tasks", "abs": "Scaling laws are useful guides for derisking expensive training runs, as they predict performance of large models using cheaper, small-scale experiments. However, there remain gaps between current scaling studies and how language models are ultimately trained and evaluated. For instance, scaling is usually studied in the compute-optimal training regime (i.e., \"Chinchilla optimal\" regime). In contrast, models are often over-trained to reduce inference costs. Moreover, scaling laws mostly predict loss on next-token prediction, but models are usually compared on downstream task performance. To address both shortcomings, we create a testbed of 104 models with 0.011B to 6.9B parameters trained with various numbers of tokens on three data distributions. First, we fit scaling laws that extrapolate in both the amount of over-training and the number of model parameters. This enables us to predict the validation loss of a 1.4B parameter, 900B token run (i.e., 32$\\times$ over-trained) and a 6.9B parameter, 138B token run (i.e., a compute-optimal run)\u2013\u2013each from experiments that take 300$\\times$ less compute. Second, we relate the perplexity of a language model to its downstream task performance by proposing a power law. We use this law to predict top-1 error averaged over downstream tasks for the two aforementioned models, using experiments that take 20$\\times$ less compute.", "keywords": [ "large language models", "scaling laws", "over-training", "task prediction" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 5, 3, 3, 5 ] }, { "paper_id": "iZI1vCiTTA", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Joykirat Singh", "gender": "Male", "institution": "Research, Microsoft", "country": "IN", "position": "Researcher" }, { "name": "Subhabrata Dutta", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "Postdoc" }, { "name": "Tanmoy Chakraborty", "gender": "Male", "institution": "Indian Institute of Technology, Delhi", "country": "IN", "position": "Associate Professor" } ], "n_formula": 14, "n_formula_1": 16, "n_ref_uni": 30, "n_ref": 69, "n_ref_all": 88, "n_fig": 2, "n_tab": 4, "L_tab": 1451, "n_element_tab": 278, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 949, "n_element_tab_1": 173, "formula_len_all": 1246, "formula_len_all_1": 1321, "len_all": 147034, "len_all_1": 66050, "len_abs": 1131, "len_title": 95, "len_sents": 37646, "len_sents_1": 30358, "n_sents": 292, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1135, "title": "Mechanistic Behavior Editing of Language Models", "abs": "Large Language Models trained on web-scale text acquire language generation abilities that can solve a wide range of tasks, particularly when task knowledge is refined into the generative prior using in-context examples. However, spurious features learned from noisy data hinder their generalizability. Supervised finetuning can introduce task specificity, but introduce data inefficiency. Prior studies indicate that (i) noisy neural circuitries coexist with generalizable ones within LLMs, and (ii) finetuning typically enhances (or suppresses) existing abilities without introducing newer ones. Building upon these, we propose TaRot, a novel method for task adaptation. TaRot intervenes in the neural circuitries using learnable rotation matrices that are optimized using Bayesian Optimization, on labelled samples in the order of standard few-shot prompting examples. Experiments on multiple classification and generation tasks using LLMs of varying sizes reveal the efficacy of TaRot, improving upon both zero- as well as few-shot performance, with average improvements (across models and tasks) of 23.81% and 11.15%, respectively", "keywords": [ "Mechanistic Intervention", "Bayesian Optimization" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 2 ] }, { "paper_id": "iYkhxre0In", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Beomseok Kim", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Dongsuk Jeon", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Associate Professor" }, { "name": "Inho Jeong", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Sol Namkung", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Sunghyeon Woo", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "PhD student" }, { "name": "Sunwoo Lee", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 12, "n_ref_uni": 43, "n_ref": 93, "n_ref_all": 110, "n_fig": 5, "n_tab": 6, "L_tab": 2125, "n_element_tab": 362, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1055, "n_element_tab_1": 195, "formula_len_all": 1889, "formula_len_all_1": 735, "len_all": 179457, "len_all_1": 61593, "len_abs": 2164, "len_title": 109, "len_sents": 37221, "len_sents_1": 28355, "n_sents": 240, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 2239, "title": "PaCA: Partial Connection Adaptation for Efficient Fine-Tuning", "abs": "Prior parameter-efficient fine-tuning (PEFT) algorithms reduce memory usage and computational costs of fine-tuning large neural network models by training only a few additional adapter parameters, rather than the entire model. However, the reduction in computational costs due to PEFT does not necessarily translate to a reduction in training time; although the computational costs of the adapter layers are much smaller than the pretrained layers, it is well known that those two types of layers are processed sequentially on GPUs, resulting in significant latency overhead. LoRA and its variants avoid this latency overhead by merging the low-rank adapter matrices with the pretrained weights during inference. However, those layers cannot be merged during training since the pretrained weights must remain frozen while the low-rank adapter matrices are updated continuously over the course of training. Furthermore, LoRA and its variants do not reduce activation memory, as the first low-rank adapter matrix still requires the input activations to the pretrained weights to compute weight gradients. To mitigate this issue, we propose **Pa**rtial **C**onnection **A**daptation (**PaCA**), which fine-tunes randomly selected partial connections within the pretrained weights instead of introducing adapter layers in the model. PaCA not only enhances training speed by eliminating the time overhead due to the sequential processing of the adapter and pretrained layers but also reduces activation memory since only partial activations, rather than full activations, need to be stored for gradient computation. Compared to LoRA, PaCA reduces training time by 22\\% and total memory usage by 16\\%, while maintaining comparable accuracy across various fine-tuning scenarios, such as fine-tuning on the MMLU dataset and instruction tuning on the Oasst1 dataset. PaCA can also be combined with quantization, enabling the fine-tuning of large models such as LLaMA3.1-70B. In addition, PaCA enables training on 23\\% longer sequence data and improves throughput by 16\\% on both NVIDIA A100 and INTEL Gaudi 2 GPUs compared to LoRA. The code is available at [https://anonymous.4open.science/r/paca-366F](https://anonymous.4open.science/r/paca-366F).", "keywords": [ "Parameter-Efficient Fine-Tuning (PEFT)", "Large Language Models (LLM)", "Memory efficient training", "Accelerating training" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "iXytGy9UOj", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 5, "author_info_list": [ { "name": "Gagandeep Singh", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Sasa Misailovic", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Associate Professor" }, { "name": "Shubham Ugare", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Yuchen Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "YIFAN ZHAO", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 18, "n_ref_uni": 26, "n_ref": 53, "n_ref_all": 68, "n_fig": 5, "n_tab": 11, "L_tab": 3247, "n_element_tab": 275, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 2438, "n_element_tab_1": 126, "formula_len_all": 1270, "formula_len_all_1": 1025, "len_all": 132098, "len_all_1": 67985, "len_abs": 1505, "len_title": 134, "len_sents": 38972, "len_sents_1": 29974, "n_sents": 309, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1539, "title": "ARQ: A Mixed-Precision Quantization Framework for Accurate and Certifiably Robust DNNs", "abs": "Mixed precision quantization has become an important technique for\nenabling the execution of deep neural networks (DNNs) on limited resource computing platforms.\nTraditional quantization methods have primarily concentrated on maintaining\nneural network accuracy, either ignoring the impact of quantization on the\nrobustness of the network, or using only empirical techniques for improving\nrobustness. In contrast, techniques for robustness certification, which can\nprovide strong guarantees about the robustness of DNNs have not been used\nduring quantization due to their high computation cost and/or scalability\nissues. \n\nThis paper introduces ARQ, an innovative mixed-precision quantization method that not only\npreserves the clean accuracy of the smoothed classifiers but also maintains\ntheir certified robustness. ARQ uses reinforcement learning to find accurate and robust\nDNN quantization, while efficiently leveraging randomized smoothing,\na popular class of statistical DNN verification algorithms, to guide the search process. \nWe compare ARQ with multiple state-of-the-art quantization techniques on\nseveral DNN architectures commonly used in quantization studies: ResNet-20 on\nCIFAR-10, ResNet-50 on ImageNet, and MobileNetV2 on ImageNet. \nWe demonstrate that ARQ consistently performs better than these baselines\nacross all the benchmarks and the input perturbation levels. In many cases, the performance of ARQ quantized networks can reach that of the original DNN with floating-point weights, but with only 1.5% instructions.", "keywords": [ "Quantization", "Robustness", "Certification", "Formal Verification" ], "rating_list": [ 5, 3, 3, 8 ], "soundness_list": [ 3, 2, 2, 4 ], "presentation_list": [ 4, 3, 2, 4 ], "contribution_list": [ 3, 2, 2, 4 ], "confidence_list": [ 5, 4, 4, 2 ] }, { "paper_id": "iXbUquaWbl", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Denis Blessing", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "", "position": "PhD student" }, { "name": "Xiaogang Jia", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "PhD student" }, { "name": "Gerhard Neumann", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 72, "n_formula_1": 36, "n_ref_uni": 59, "n_ref": 137, "n_ref_all": 184, "n_fig": 11, "n_tab": 7, "L_tab": 2684, "n_element_tab": 210, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5961, "formula_len_all_1": 2693, "len_all": 202263, "len_all_1": 63271, "len_abs": 1312, "len_title": 116, "len_sents": 58310, "len_sents_1": 25851, "n_sents": 470, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1270, "title": "End-to-end Learning of Gaussian Mixture Priors for Diffusion Sampler", "abs": "Diffusion models optimized via variational inference (VI) have emerged as a promising tool for generating samples from unnormalized target densities. These models create samples by simulating a stochastic differential equation, starting from a simple, tractable prior, typically a Gaussian distribution. However, when the support of this prior differs greatly from that of the target distribution, diffusion models often struggle to explore effectively or suffer from large discretization errors. Moreover, learning the prior distribution can lead to mode-collapse, exacerbated by the mode-seeking nature of reverse Kullback-Leibler divergence commonly used in VI.\nTo address these challenges, we propose end-to-end learnable Gaussian mixture priors (GMPs). GMPs offer improved control over exploration, adaptability to target support, and increased expressiveness to counteract mode collapse. We further leverage the structure of mixture models by proposing a strategy to iteratively refine the model through the addition of mixture components during training. Our experimental results demonstrate significant performance improvements across a diverse range of real-world and synthetic benchmark problems when using GMPs without requiring additional target evaluations.", "keywords": [ "Variational Inference", "Sampling", "Diffusion Models", "Mixture Models" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "iXCeQ2m6vT", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Angeliki Pantazi", "gender": "Female", "institution": "International Business Machines", "country": "", "position": "Principal Researcher" }, { "name": "Oleh Kolner", "gender": "Male", "institution": "Institute for Theoretical Computer Science, TU Graz, Technische Universit\u00e4t Graz", "country": "AT", "position": "PhD student" }, { "name": "Stanis\u0142aw Wo\u017aniak", "gender": "Male", "institution": "IBM Research", "country": "", "position": "Researcher" }, { "name": "Thomas Bohnstingl", "gender": "Male", "institution": "IBM Research - Europe", "country": "CH", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 3, "n_ref_uni": 46, "n_ref": 109, "n_ref_all": 136, "n_fig": 11, "n_tab": 6, "L_tab": 3844, "n_element_tab": 369, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1206, "n_element_tab_1": 82, "formula_len_all": 264, "formula_len_all_1": 118, "len_all": 164607, "len_all_1": 67103, "len_abs": 1345, "len_title": 161, "len_sents": 54569, "len_sents_1": 32688, "n_sents": 383, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 111, "L_abs": 1349, "title": "Mind the GAP: Glimpse-based Active Perception improves generalization and sample efficiency of visual reasoning", "abs": "Human capabilities in understanding visual relations are far superior to those of AI systems, especially for previously unseen objects. For example, while AI systems struggle to determine whether two such objects are visually the same or different, humans can do so with ease. Active vision theories postulate that the learning of visual relations is grounded in actions that we take to fixate objects and their parts by moving our eyes. In particular, the low-dimensional spatial information about the corresponding eye movements is hypothesized to facilitate the representation of relations between different image parts. Inspired by these theories, we develop a system equipped with a novel Glimpse-based Active Perception (GAP) that sequentially glimpses at the most salient regions of the input image and processes them at high resolution. Importantly, our system leverages the locations stemming from the glimpsing actions, along with the visual content around them, to represent relations between different parts of the image. The results suggest that the GAP is essential for extracting visual relations that go beyond the immediate visual content. Our approach reaches state-of-the-art performance on several visual reasoning tasks being more sample-efficient, and generalizing better to out-of-distribution visual inputs than prior models.", "keywords": [ "visual reasoning", "active vision", "out-of-distribution", "generalization", "sample efficiency", "relational features", "brain-inspired", "neuro-inspired" ], "rating_list": [ 6, 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3, 4 ] }, { "paper_id": "iXBYYbYTvX", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Hao Xiong", "gender": "Non-Binary", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Jianhao Huang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Undergrad student" }, { "name": "Wenzheng Pan", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "ZhuoXIA", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 27, "n_formula_1": 22, "n_ref_uni": 47, "n_ref": 83, "n_ref_all": 103, "n_fig": 5, "n_tab": 5, "L_tab": 4694, "n_element_tab": 195, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 3763, "n_element_tab_1": 162, "formula_len_all": 1618, "formula_len_all_1": 1345, "len_all": 146431, "len_all_1": 76024, "len_abs": 1801, "len_title": 146, "len_sents": 39643, "len_sents_1": 30287, "n_sents": 358, "n_sents_1": 277, "venue": "ICLR2025", "decision": "withdraw", "L_title": 98, "L_abs": 1645, "title": "From Graph Embedding to LKH: Bridging Learning and Heuristics for a Streamlined General TSP Solver", "abs": "The Traveling Salesman Problem (TSP) is known as one of the most notorious NP-hard combinatorial optimization problems. \nIn recent decades, researchers from fields such as computer science, operations research, and artificial intelligence including deep learning (DL) have made numerous attempts on the problem. \nAmong the works, the Lin-Kernighan-Helsgaun (LKH) heuristic algorithm is one of the most competent methods for obtaining optimal or near-optimal solutions. \nDespite the rapid development in DL-based solvers, few of them can defeat LKH in terms of both running efficiency and solution quality across different distributions.\nIn this paper, we would introduce a very novel approach that enhances LKH with graph embedding (GE) techniques in solving general TSP (distances can be non-metric and asymmetric), named as Embed-LKH. \nIt is presented as two stages: i) in the GE stage, it transforms the distances to transition probabilities, then conduct GE given the transition probabilities, and finally it uses the learned embeddings to construct the so-called `ghost distances'; ii) in the LKH stage, LKH generates candidates based on the ghost distances but searches tours according to the original distances. As the experiments show, compared with the original LKH counterpart, in most cases, our approach can obtain better solutions within the same amount of trials across six distance distributions (non-metric and asymmetric: normal, uniform, exponential, metric and symmetric: Euclidean 2D/10D/50D) and two problem scales (TSP-100/1000). The source files, running scripts, and data will be made publicly available after the review.", "keywords": [ "Traveling salesman problem", "graph embedding" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "iX7eHHE5Tx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Chao Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Didi Zhu", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jinluan Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Min Zhang", "gender": "unknown", "institution": "College of Computer Science and Technology, Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yibing Song", "gender": "unknown", "institution": "Alibaba DAMO Academy", "country": "CN", "position": "Staff Scientist" }, { "name": "Ziyu Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "tao shen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 33, "n_ref": 50, "n_ref_all": 76, "n_fig": 4, "n_tab": 9, "L_tab": 3817, "n_element_tab": 577, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1319, "n_element_tab_1": 199, "formula_len_all": 530, "formula_len_all_1": 509, "len_all": 152960, "len_all_1": 72017, "len_abs": 1500, "len_title": 111, "len_sents": 46764, "len_sents_1": 30939, "n_sents": 368, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1543, "title": "REMEDY: Recipe Merging Dynamics in Large Vision-Language Models", "abs": "Model merging has emerged as a powerful technique for combining task-specific vision models into a unified and multi-functional model. Previous methods represented by task arithmetic, have demonstrated effectiveness and scalability in this domain. When large vision-language models (LVLMs) arise with model size scaling up, this design becomes challenging to fuse different instruction-tuned LVLMs for generalization enhancement. The large scale and multi-modal nature of LVLMs present unique obstacles, including constructing reusable and modular components to accommodate the multi-component architecture of LVLMs and the requirement for dynamic fusion based on multi-modal input tokens. To address these challenges, we propose the \\textbf{RE}cipe \\textbf{ME}rging \\textbf{DY}namics (REMEDY) method, a scalable and flexible paradigm for model merging in LVLMs. We first define reusable modules termed \\textit{recipes} including the projector and shallow LLM layers, enhancing visual-language understanding. Then, we introduce a modality-aware allocator dynamically generates weights in a one-shot manner based on input relevance to existing recipes, enabling efficient cross-modal knowledge integration. REMEDY thus offers an adaptive solution for LVLMs to tackle both seen (i.e., multi-task learning) and unseen (i.e., zero-shot generalization) tasks. Experimental results demonstrate that our method consistently improves performance on both seen and unseen tasks, underscoring the effectiveness of REMEDY in diverse multi-modal scenarios.", "keywords": [ "Multi-Modal Large Language Models", "Zero-shot Generalization", "Model Merging" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 4, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "iX3uESGdsO", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Feifei Wang", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Postdoc" }, { "name": "Huijie Fan", "gender": "Female", "institution": "Shenyang Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Liangqiong Qu", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Pengxin Guo", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shuang Zeng", "gender": "Female", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yanran Wang", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Postdoc" } ], "n_formula": 61, "n_formula_1": 16, "n_ref_uni": 57, "n_ref": 158, "n_ref_all": 210, "n_fig": 7, "n_tab": 10, "L_tab": 3764, "n_element_tab": 348, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 1384, "n_element_tab_1": 159, "formula_len_all": 9181, "formula_len_all_1": 1618, "len_all": 180982, "len_all_1": 71228, "len_abs": 1139, "len_title": 116, "len_sents": 41113, "len_sents_1": 32168, "n_sents": 307, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1154, "title": "Selective Aggregation for Low-Rank Adaptation in Federated Learning", "abs": "We investigate LoRA in federated learning through the lens of the asymmetry analysis of the learned $A$ and $B$ matrices. In doing so, we uncover that $A$ matrices are responsible for learning general knowledge, while $B$ matrices focus on capturing client-specific knowledge. Based on this finding, we introduce Federated Share-A Low-Rank Adaptation (FedSA-LoRA), which employs two low-rank trainable matrices $A$ and $B$ to model the weight update, but only $A$ matrices are shared with the server for aggregation. Moreover, we delve into the relationship between the learned $A$ and $B$ matrices in other LoRA variants, such as rsLoRA and VeRA, revealing a consistent pattern. Consequently, we extend our FedSA-LoRA method to these LoRA variants, resulting in FedSA-rsLoRA and FedSA-VeRA. In this way, we establish a general paradigm for integrating LoRA with FL, offering guidance for future work on subsequent LoRA variants combined with FL. Extensive experimental results on natural language understanding and generation tasks demonstrate the effectiveness of the proposed method. Our code is available at https://github.com/Pengxin-Guo/FedSA-LoRA.", "keywords": [ "federated learning", "low-rank adaptation" ], "rating_list": [ 6, 5, 8, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "iWSl5Zyjjw", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Amir Globerson", "gender": "Male", "institution": "Tel Aviv University", "country": "", "position": "Full Professor" }, { "name": "Assaf Ben-Kish", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Itamar Zimerman", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Lior Wolf", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Associate Professor" }, { "name": "Nadav Cohen", "gender": "Male", "institution": "School of Computer Science, Tel Aviv University", "country": "IL", "position": "Associate Professor" }, { "name": "Raja Giryes", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Associate Professor" }, { "name": "Shady Abu Hussein", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 41, "n_ref": 65, "n_ref_all": 106, "n_fig": 13, "n_tab": 7, "L_tab": 2583, "n_element_tab": 230, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 981, "formula_len_all_1": 822, "len_all": 180931, "len_all_1": 64391, "len_abs": 1338, "len_title": 113, "len_sents": 57980, "len_sents_1": 29499, "n_sents": 453, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1136, "title": "DeciMamba: Exploring the Length Extrapolation Potential of Mamba", "abs": "Long-range sequence processing poses a significant challenge for Transformers due to their quadratic complexity in input length. A promising alternative is Mamba, which demonstrates high performance and achieves Transformer-level capabilities while requiring substantially fewer computational resources. In this paper we explore the length-generalization capabilities of Mamba, which we find to be relatively limited. Through a series of visualizations and analyses we identify that the limitations arise from a restricted effective receptive field, dictated by the sequence length used during training. To address this constraint, we introduce DeciMamba, a context-extension method specifically designed for Mamba. This mechanism, built on top of a hidden filtering mechanism embedded within the S6 layer, enables the trained model to extrapolate well even without additional training. Empirical experiments over real-world long-range NLP tasks show that DeciMamba can extrapolate to context lengths that are significantly longer than the ones seen during training, while enjoying faster inference. We will release our code and models.", "keywords": [ "Mamba", "Long context", "Context extension", "Long-range language modeling" ], "rating_list": [ 8, 3, 3, 6 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 4, 2, 2, 2 ], "contribution_list": [ 4, 2, 2, 3 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "iWCfiDxLIY", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Attila Lischka", "gender": "unknown", "institution": "Chalmers University of Technology", "country": "", "position": "PhD student" }, { "name": "Balazs Kulcsar", "gender": "unknown", "institution": "Chalmers University of Technology", "country": "", "position": "Prof" }, { "name": "Jiaming Wu", "gender": "Male", "institution": "Chalmers University of Technology", "country": "", "position": "Researcher" }, { "name": "Morteza Haghir Chehreghani", "gender": "Male", "institution": "Chalmers University of technology", "country": "SE", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 19, "n_ref": 51, "n_ref_all": 67, "n_fig": 7, "n_tab": 5, "L_tab": 1887, "n_element_tab": 303, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1004, "n_element_tab_1": 143, "formula_len_all": 551, "formula_len_all_1": 519, "len_all": 112907, "len_all_1": 65072, "len_abs": 1288, "len_title": 107, "len_sents": 36124, "len_sents_1": 29192, "n_sents": 303, "n_sents_1": 242, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1297, "title": "A GREAT Architecture for Edge-Based Graph Problems Like TSP", "abs": "In the last years, many neural network-based approaches have been proposed to tackle combinatorial optimization problems such as routing problems.\nMany of these approaches are based on graph neural networks (GNNs) or related transformers, operating on the Euclidean coordinates representing the routing problems. \nHowever, GNNs are inherently not well suited to operate on dense graphs, such as in routing problems. Furthermore, models operating on Euclidean coordinates cannot be applied to non-Euclidean versions of routing problems that are often found in real-world settings.\nTo overcome these limitations, we propose a novel GNN-related edge-based neural model called Graph Edge Attention Network (GREAT).\nWe evaluate the performance of GREAT in the edge-classification task to predict optimal edges in the Traveling Salesman Problem (TSP). We can use such a trained GREAT model to produce sparse TSP graph instances, keeping only the edges GREAT finds promising. Compared to other, non-learning-based methods to sparsify TSP graphs, GREAT can produce very sparse graphs while keeping most of the optimal edges. Furthermore, we build a reinforcement learning-based GREAT framework which we apply to Euclidean and non-Euclidean asymmetric TSP. This framework achieves state-of-the-art results.", "keywords": [ "Traveling Salesman Problem", "Asymmetric Traveling Salesman Problem", "Graph Learning", "Graph Neural Networks", "Routing Problems" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "iVy7aRMb0K", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Asher Trockman", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Hrayr Harutyunyan", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Sanjiv Kumar", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Srinadh Bhojanapalli", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Zico Kolter", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 17, "n_ref": 32, "n_ref_all": 58, "n_fig": 17, "n_tab": 2, "L_tab": 1033, "n_element_tab": 39, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 279, "n_element_tab_1": 3, "formula_len_all": 1025, "formula_len_all_1": 1024, "len_all": 85437, "len_all_1": 57231, "len_abs": 873, "len_title": 111, "len_sents": 30920, "len_sents_1": 26612, "n_sents": 220, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 871, "title": "Mimetic Initialization Helps State Space Models Learn to Recall", "abs": "Recent work has shown that state space models such as Mamba are significantly worse than Transformers on recall-based tasks due to the fact that their state size is constant with respect to their input sequence length. But in practice, state space models have fairly large state sizes, and we conjecture that they should be able to perform much better at these tasks than previously reported. We investigate whether their poor copying and recall performance could be due in part to training difficulties rather than fundamental capacity constraints. Based on observations of their \"attention'' maps, we propose a structured initialization technique that allows state space layers to more readily mimic attention. Across a variety of architecture settings, our initialization makes it substantially easier for Mamba to learn to copy and do associative recall from scratch.", "keywords": [ "Mamba", "SSM", "sequence models", "language models", "initialization", "linear attention" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "iVxxgZlXh6", "primary_area": "applications to robotics, autonomy, planning", "n_author": 11, "author_info_list": [ { "name": "Cristina Mata", "gender": "unknown", "institution": "State University of New York, Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Jinghuan Shang", "gender": "Male", "institution": "Boston Dynamics AI Institute", "country": "US", "position": "Researcher" }, { "name": "Jongwoo Park", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Kumara Kahatapitiya", "gender": "Male", "institution": "Meta", "country": "US", "position": "Intern" }, { "name": "Michael S Ryoo", "gender": "Male", "institution": "Salesforce AI Research", "country": "US", "position": "Principal Researcher" }, { "name": "Mu Cai", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin, Madison", "country": "US", "position": "PhD student" }, { "name": "Ryan D Burgert", "gender": "Non-Binary", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Xiang Li", "gender": "unknown", "institution": "State University of New York, Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Yoo Sung Jang", "gender": "Male", "institution": ", State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Kanchana Nisal Ranasinghe", "gender": "Male", "institution": "State University of New York, Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Yong Jae Lee", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 38, "n_ref": 101, "n_ref_all": 122, "n_fig": 13, "n_tab": 19, "L_tab": 9183, "n_element_tab": 2755, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 394, "n_element_tab_1": 22, "formula_len_all": 675, "formula_len_all_1": 126, "len_all": 350155, "len_all_1": 72826, "len_abs": 1170, "len_title": 115, "len_sents": 100962, "len_sents_1": 34369, "n_sents": 867, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1372, "title": "LLaRA: Supercharging Robot Learning Data for Vision-Language Policy", "abs": "Vision Language Models (VLMs) have recently been leveraged to generate robotic actions, forming Vision-Language-Action (VLA) models. However, directly adapting a pretrained VLM for robotic control remains challenging, particularly when constrained by a limited number of robot demonstrations. In this work, we introduce LLaRA: Large Language and Robotics Assistant, a framework that formulates robot action policy as visuo-textual conversations and enables an efficient transfer of a pretrained VLM into a powerful VLA, motivated by the success of visual instruction tuning in Computer Vision. First, we present an automated pipeline to generate conversation-style instruction tuning data for robots from existing behavior cloning datasets, aligning robotic actions with image pixel coordinates. Further, we enhance this dataset in a self-supervised manner by defining six auxiliary tasks, without requiring any additional action annotations. We show that a VLM finetuned with a limited amount of such datasets can produce meaningful action decisions for robotic control. Through experiments across multiple simulated and real-world tasks, we demonstrate that LLaRA achieves state-of-the-art performance while preserving the generalization capabilities of large language models. The code, datasets, and pretrained models are available at https://github.com/LostXine/LLaRA.", "keywords": [ "Robot Learning", "VLM", "Behavior Cloning", "Instruction Tuning", "Self-supervised Learning" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "iVMcYxTiVM", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Janis Keuper", "gender": "Male", "institution": "Institute for Machine Learning and Analytics, Offenburg University", "country": "DE", "position": "Full Professor" }, { "name": "Jovita Lukasik", "gender": "Female", "institution": "Universit\u00e4t Siegen", "country": "", "position": "Postdoc" }, { "name": "Margret Keuper", "gender": "Female", "institution": "Universit\u00e4t Mannheim", "country": "DE", "position": "Full Professor" }, { "name": "Muhammad Jehanzeb Mirza", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Paul Gavrikov", "gender": "Male", "institution": "Universit\u00e4t Mannheim", "country": "DE", "position": "PhD student" }, { "name": "Robert Geirhos", "gender": "Male", "institution": "Google DeepMind", "country": "CA", "position": "Research Scientist" }, { "name": "Steffen Jung", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 56, "n_ref": 188, "n_ref_all": 223, "n_fig": 13, "n_tab": 8, "L_tab": 2512, "n_element_tab": 281, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 65, "formula_len_all_1": 44, "len_all": 243124, "len_all_1": 66961, "len_abs": 1612, "len_title": 101, "len_sents": 70670, "len_sents_1": 33330, "n_sents": 582, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1624, "title": "Can We Talk Models Into Seeing the World Differently?", "abs": "Unlike traditional vision-only models, vision language models (VLMs) offer an intuitive way to access visual content through language prompting by combining a large language model (LLM) with a vision encoder. However, both the LLM and the vision encoder come with their own set of biases, cue preferences, and shortcuts, which have been rigorously studied in uni-modal models. A timely question is how such (potentially misaligned) biases and cue preferences behave under multi-modal fusion in VLMs. \nAs a first step towards a better understanding, we investigate a particularly well-studied vision-only bias - the texture vs. shape bias and the dominance of local over global information. \nAs expected, we find that VLMs inherit this bias to some extent from their vision encoders. Surprisingly, the multi-modality alone proves to have important effects on the model behavior, i.e., the joint training and the language querying change the way visual cues are processed. \nWhile this direct impact of language-informed training on a model's visual perception is intriguing, it raises further questions on our ability to actively steer a model's output so that its prediction is based on particular visual cues of the user's choice. \nInterestingly, VLMs have an inherent tendency to recognize objects based on shape information, which is different from what a plain vision encoder would do. Further active steering towards shape-based classifications through language prompts is however limited. In contrast, active VLM steering towards texture-based decisions through simple natural language prompts is often more successful.", "keywords": [ "vision language models", "vision biases", "shape/texture bias" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 4, 3, 4 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "iUwTDbjqyd", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Jung-Eun Kim", "gender": "Female", "institution": "North Carolina State University", "country": "US", "position": "Assistant Professor" }, { "name": "Xingli Fang", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 49, "n_ref_all": 94, "n_fig": 30, "n_tab": 3, "L_tab": 1004, "n_element_tab": 106, "n_fig_1": 24, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 379, "formula_len_all_1": 42, "len_all": 134456, "len_all_1": 61520, "len_abs": 892, "len_title": 91, "len_sents": 35641, "len_sents_1": 30158, "n_sents": 291, "n_sents_1": 238, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 897, "title": "Generating Fake Data to Fake Privacy Pryers", "abs": "Asymmetry of data complexity and model capacity can create privacy vulnerability. That is because if there are relatively fewer data points while the model capacity is relatively higher, a model may memorize almost all the data points. As a remedy for the issue, more data samples can be generated. When generating more data samples, the aim is to protect and promote the original data as privacy-safe as possible while generating more privacy-risky data samples to fake privacy attackers. To enable the aim, we investigate each individual data sample's privacy level, unlike existing studies that only take into account an overall dataset's privacy, which is not precisely effective. We show how effective our generative approach is in combating privacy attacks. Our work is novel in that we propose a sample-level valuation, and data transformation and generation approach in the privacy domain.", "keywords": [ "Privacy", "Data valuation", "Generative models" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 1, 1, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "iTm4H6N4aG", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Jiannan Huang", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Jun Hao Liew", "gender": "unknown", "institution": "ByteDance", "country": "SG", "position": "Researcher" }, { "name": "Yao Zhao", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Yunchao Wei", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Yuyang Yin", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Hanshu Yan", "gender": "Male", "institution": "ByteDance", "country": "SG", "position": "Research Scientist" }, { "name": "Honghui Shi", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 93, "n_ref": 192, "n_ref_all": 221, "n_fig": 15, "n_tab": 3, "L_tab": 904, "n_element_tab": 164, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 472, "n_element_tab_1": 80, "formula_len_all": 855, "formula_len_all_1": 839, "len_all": 185320, "len_all_1": 56590, "len_abs": 1588, "len_title": 128, "len_sents": 37341, "len_sents_1": 24668, "n_sents": 301, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1599, "title": "ClassDiffusion: More Aligned Personalization Tuning with Explicit Class Guidance", "abs": "Recent text-to-image customization works have proven successful in generating images of given concepts by fine-tuning diffusion models on a few examples. However, tuning-based methods inherently tend to overfit the concepts, resulting in failure to create the concept under multiple conditions (*e.g.*, headphone is missing when generating \"a `dog wearing a headphone\"). Interestingly, we notice that the base model before fine-tuning exhibits the capability to compose the base concept with other elements (*e.g.*, \"a dog wearing a headphone\"), implying that the compositional ability only disappears after personalization tuning. We observe a semantic shift in the customized concept after fine-tuning, indicating that the personalized concept is not aligned with the original concept, and further show through theoretical analyses that this semantic shift leads to increased difficulty in sampling the joint conditional probability distribution, resulting in the loss of the compositional ability. Inspired by this finding, we present **ClassDiffusion**, a technique that leverages a **semantic preservation loss** to explicitly regulate the concept space when learning a new concept. Although simple, this approach effectively prevents semantic drift during the fine-tuning process of the target concepts. Extensive qualitative and quantitative experiments demonstrate that the use of semantic preservation loss effectively improves the compositional abilities of fine-tuning models. Lastly, we also extend our ClassDiffusion to personalized video generation, demonstrating its flexibility.", "keywords": [ "Diffusion", "Subject-Driven Personalization" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "iTjSqQQ4f8", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Jalal Kazempour", "gender": "Male", "institution": "Technical University of Denmark", "country": "", "position": "Associate Professor" }, { "name": "Pierre Pinson", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Full Professor" }, { "name": "Thomas Falconer", "gender": "unknown", "institution": "Technical University of Denmark", "country": "DK", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 37, "n_ref": 54, "n_ref_all": 64, "n_fig": 5, "n_tab": 1, "L_tab": 15, "n_element_tab": 1, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 15, "n_element_tab_1": 1, "formula_len_all": 884, "formula_len_all_1": 860, "len_all": 108483, "len_all_1": 77527, "len_abs": 1294, "len_title": 87, "len_sents": 37179, "len_sents_1": 37914, "n_sents": 223, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 39, "L_abs": 1299, "title": "Towards Replication-Robust Data Markets", "abs": "Despite widespread adoption of machine learning throughout industry, many firms face a common challenge: relevant datasets are typically distributed amongst market competitors that are reluctant to share information. Recent works propose data markets to provide monetary incentives for collaborative machine learning, where agents share features with each other and are rewarded based on their contribution to improving the predictions others. These contributions are determined by their relative Shapley value, which is computed by treating features as players and their interactions as a characteristic function game. However, in its standard form, this setup further provides an incentive for agents to replicate their data and act under multiple false identities in order to increase their own revenue and diminish that of others, restricting their use in practice. In this work, we develop a replication-robust data market for supervised learning problems. We adopt Pearl\u2019s do-calculus from causal reasoning to refine the characteristic function game by differentiating between observational and interventional conditional probabilities. By doing this, we derive Shapley value-based rewards that are robust to this malicious replication by design, whilst preserving desirable market properties.", "keywords": [ "regression", "bayesian inference", "collaborative analytics", "data markets", "game theory" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 2, 2 ] }, { "paper_id": "iTVKOOZeYW", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Dmitry Kamzolov", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Klea Ziu", "gender": "Female", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Loka Li", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Slavomir Hanzely", "gender": "unknown", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Postdoc" }, { "name": "Martin Takac", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Associate Professor" } ], "n_formula": 35, "n_formula_1": 24, "n_ref_uni": 51, "n_ref": 110, "n_ref_all": 149, "n_fig": 18, "n_tab": 3, "L_tab": 665, "n_element_tab": 77, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2538, "formula_len_all_1": 1276, "len_all": 154609, "len_all_1": 57154, "len_abs": 1355, "len_title": 126, "len_sents": 44837, "len_sents_1": 24905, "n_sents": 363, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1360, "title": "$\\psi$DAG : Projected Stochastic Approximation Iteration for DAG Structure Learning", "abs": "Learning the structure of Directed Acyclic Graphs (DAGs) presents a significant challenge due to the vast combinatorial search space of possible graphs, which scales exponentially with the number of nodes. Recent advancements have redefined this problem as a continuous optimization task by incorporating differentiable acyclicity constraints. These methods commonly rely on algebraic characterizations of DAGs, such as matrix exponentials, to enable the use of gradient-based optimization techniques. Despite these innovations, existing methods often face optimization difficulties due to the highly non-convex nature of DAG constraints and the per-iteration computational complexity. In this work, we present a novel framework for learning DAGs, employing a Stochastic Approximation approach integrated with Stochastic Gradient Descent (SGD)-based optimization techniques. Our framework introduces new projection methods tailored to efficiently enforce DAG constraints, ensuring that the algorithm converges to a feasible local minimum. With its low iteration complexity, the proposed method is well-suited for handling large-scale problems with improved computational efficiency. We demonstrate the effectiveness and scalability of our framework through comprehensive experimental evaluations, which confirm its superior performance across various settings.", "keywords": [ "Structure Learning", "continuous optimization", "directed acyclic graphs", "stochastic optimization" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "iT4ImLQatF", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Anatol Eugen Wegner", "gender": "Male", "institution": "Bayerische Julius-Maximilians-Universit\u00e4t W\u00fcrzburg", "country": "DE", "position": "Postdoc" }, { "name": "Chester Tan", "gender": "unknown", "institution": "Bayerische Julius-Maximilians-Universit\u00e4t W\u00fcrzburg", "country": "", "position": "PhD student" }, { "name": "Ingo Scholtes", "gender": "Male", "institution": "University of W\u00fcrzburg", "country": "DE", "position": "Full Professor" } ], "n_formula": 35, "n_formula_1": 11, "n_ref_uni": 82, "n_ref": 228, "n_ref_all": 241, "n_fig": 3, "n_tab": 124, "L_tab": 120600, "n_element_tab": 13211, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2120, "n_element_tab_1": 220, "formula_len_all": 2001, "formula_len_all_1": 487, "len_all": 675793, "len_all_1": 54103, "len_abs": 940, "len_title": 125, "len_sents": 93209, "len_sents_1": 25384, "n_sents": 690, "n_sents_1": 121, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 946, "title": "A Semi-Supervised Clustering Approach For Graph Learning with Neural Networks", "abs": "We propose a semi-supervised approach that combines any unsupervised clustering objective and supervised objective for end-to-end training any neural networks to improve node classification in attributed graphs, particularly when training labels are sparse.\n\tOur framework formulates node classification as semi-supervised inference of neural network models of attributed graphs with cluster structure.\n\tWe use this framework to understand how neural networks for graph clustering can jointly cluster node attributes and graph structure, despite graph clustering objectives explicitly considering only graph structure and cluster assignments.\n\tOur framework also enables neural network architectures such as transformers and multilayer perceptrons to learn on graphs without positional encodings and without spectral or message passing layers found in graph neural networks.\n\tWe evaluate our framework on six real-world attributed graph datasets.", "keywords": [ "Semi-Supervised Clustering", "Node Classification", "Graph Clustering", "Stochastic Block Model", "Graph Neural Network", "Transformer", "MLP" ], "rating_list": [ 5, 5, 3, 6, 3 ], "soundness_list": [ 2, 2, 1, 3, 2 ], "presentation_list": [ 3, 1, 1, 3, 1 ], "contribution_list": [ 2, 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 3, 5 ] }, { "paper_id": "iSTMsye6SD", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Dawn Song", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Tianneng Shi", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Yu Gai", "gender": "unknown", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Zhun Wang", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 4, "n_ref": 5, "n_ref_all": 33, "n_fig": 8, "n_tab": 9, "L_tab": 1475, "n_element_tab": 93, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 908, "n_element_tab_1": 41, "formula_len_all": 930, "formula_len_all_1": 885, "len_all": 92821, "len_all_1": 48337, "len_abs": 1478, "len_title": 143, "len_sents": 22118, "len_sents_1": 21276, "n_sents": 174, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 141, "L_abs": 1447, "title": "Assessing the Knowledge-intensive Reasoning Capability of Large Language Models with Realistic Benchmarks Generated Programmatically at Scale", "abs": "Although LLMs demonstrates strong reasoning capability in such tasks as mathematical problem solving, less is known about their reasoning capability in settings that require extensive real-world knowledge due to the limited scale and knowledge coverage of existing benchmarks. To shed more light into this, we propose a novel pipeline that is capable of programmatically generating realistic knowledge-intensive question answering benchmarks that require complex reasoning. Leveraging open knowledge graphs, the graph query language SPARQL, and LLMs, our pipeline requires no manual annotation and can therefore scale to unprecedented benchmark size and knowledge coverage. We evaluate several state-of-the-art LLMs with benchmarks generated by our pipeline, and find that the LLMs struggle to recall and leverage world knowledge for reasoning, even for world knowledge present in their pre-training corpuses. Additionally, retrieval-augmented generation and chain-of-thoughts prompting does not fully solve the problems. Our benchmarks further enable us to examine to what extent the confidence of LLMs in the outcomes of their reasoning transparently reflects their confidence in the underlying knowledge, a study that is first-of-its-kind to our best knowledge. We find that the confidence of LLMs in the outcomes of their reasoning reflects poorly their confidence in the underlying knowledge, which suggests a direction of future improvement.", "keywords": [ "Large Language Models", "Evaluation", "Reasoning", "Hallucination" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 5, 3, 3 ] }, { "paper_id": "iSLDihAfYi", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Josiah Bjorgaard", "gender": "Male", "institution": "Syntensor, Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 49, "n_ref_all": 70, "n_fig": 11, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 103, "formula_len_all_1": 103, "len_all": 102332, "len_all_1": 53615, "len_abs": 1319, "len_title": 79, "len_sents": 31855, "len_sents_1": 27229, "n_sents": 231, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 31, "L_abs": 1324, "title": "Sparsely multimodal data fusion", "abs": "Multimodal data fusion is essential for applications requiring the integration of diverse data sources, especially in the presence of incomplete or sparsely available modalities. This paper presents a comparative study of three multimodal embedding techniques, Modal Channel Attention (MCA), Zorro, and Everything at Once (EAO), to evaluate their performance on sparsely multimodal data. MCA introduces fusion embeddings for all combinations of input modalities and uses attention masking to create distinct attention channels, enabling flexible and efficient data fusion. Experiments on two datasets with four modalities each, CMU-MOSEI and TCGA, demonstrate that MCA outperforms Zorro across ranking, recall, regression, and classification tasks and outperforms EAO across regression and classification tasks. MCA achieves superior performance by maintaining robust uniformity across unimodal and fusion embeddings. While EAO performs best in ranking metrics due to its approach of forming fusion embeddings post-inference, it underperforms in downstream tasks requiring multimodal interactions. These results highlight the importance of contrasting all modality combinations in constructing embedding spaces and offers insights into the design of multimodal architectures for real-world applications with incomplete data.", "keywords": [ "multimodal", "representation learning", "contrastive", "transformer", "masked attention", "sentiment", "sparsity", "embedding space" ], "rating_list": [ 5, 6, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 1, 3 ], "presentation_list": [ 2, 3, 2, 1, 2 ], "contribution_list": [ 1, 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "iRgzG5DKgA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Daniel de Leng", "gender": "unknown", "institution": "Link\u00f6ping University", "country": "SE", "position": "Assistant Professor" }, { "name": "Fredrik Heintz", "gender": "Male", "institution": "Link\u00f6ping University", "country": "SE", "position": "Full Professor" }, { "name": "Md Fahim Sikder", "gender": "Male", "institution": "Link\u00f6ping University", "country": "SE", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 30, "n_ref": 57, "n_ref_all": 77, "n_fig": 3, "n_tab": 5, "L_tab": 2524, "n_element_tab": 206, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2632, "n_element_tab_1": 211, "formula_len_all": 819, "formula_len_all_1": 787, "len_all": 98919, "len_all_1": 57891, "len_abs": 887, "len_title": 136, "len_sents": 28796, "len_sents_1": 23708, "n_sents": 229, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 894, "title": "Fair4Free: Generating High-fidelity Fair Synthetic Samples using Data-Free Distillation", "abs": "This work presents Fair4Free, a novel generative model to generate synthetic fair data using data-free distillation in the latent space. Fair4Free can work on the situation when the data is private or inaccessible. In our approach, we first train a teacher model to create fair representation and then distil the knowledge to a student model (using a smaller architecture). The process of distilling the student model is data-free, i.e. the student model does not have access to the training dataset while distilling. After the distillation, we use the distilled model to generate fair synthetic samples. Our extensive experiments show that our synthetic samples outperform state-of-the-art models in all three criteria (fairness, utility and synthetic quality) with a performance increase of 5\\% for fairness, 8\\% for utility and 12\\% in synthetic quality for both tabular and image datasets.", "keywords": [ "data fairness", "fair generative models", "knowledge distillation", "latent space distillation", "synthetic data", "biased data" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "iReypJFqa1", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Erteng Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Hengrui Lou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Jie Song", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Jinsong Geng", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Lechao Cheng", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Mingli Song", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yijun Bei", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Zunlei Feng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 73, "n_ref": 151, "n_ref_all": 190, "n_fig": 14, "n_tab": 6, "L_tab": 6637, "n_element_tab": 690, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2507, "n_element_tab_1": 260, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 280533, "len_all_1": 59749, "len_abs": 1468, "len_title": 120, "len_sents": 80249, "len_sents_1": 28473, "n_sents": 525, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1472, "title": "A Large-scale Universal Evaluation Benchmark For Face Forgery Detection", "abs": "With the rapid development of AI-generated content (AIGC) technology, the production of realistic fake facial images and videos that deceive human visual perception has become possible. Consequently, various face forgery detection techniques have been proposed to identify such fake facial content. However, evaluating the effectiveness and generalizability of these detection techniques remains a significant challenge. To address this, we have constructed a large-scale evaluation benchmark called DeepFaceGen, aimed at quantitatively assessing the effectiveness of face forgery detection and facilitating the iterative development of forgery detection technology. DeepFaceGen consists of 776,990 real face image/video samples and 773,812 face forgery image/video samples, generated using 34 mainstream face generation techniques. During the construction process, we carefully consider important factors such as content diversity, fairness across ethnicities, and availability of comprehensive labels, in order to ensure the versatility and convenience of DeepFaceGen. Subsequently, DeepFaceGen is employed in this study to evaluate and analyze the performance of 20 mainstream face forgery detection techniques from various perspectives. Through extensive experimental analysis, we derive significant findings and propose potential directions for future research. The code and dataset for DeepFaceGen are available at https://anonymous.4open.science/r/DeepFaceGen-47D1.", "keywords": [ "Deepfakes", "Face Forgery", "Evaluation Benchmark", "Forgery Detection" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 5, 3, 5 ] }, { "paper_id": "iRYExPKnxm", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Haozhen Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Pengrui Han", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Tao Feng", "gender": "Male", "institution": "Department of Computer Science, University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Jiaxuan You", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 23, "n_ref": 46, "n_ref_all": 72, "n_fig": 2, "n_tab": 4, "L_tab": 1640, "n_element_tab": 336, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 203, "formula_len_all_1": 0, "len_all": 187450, "len_all_1": 47472, "len_abs": 1217, "len_title": 93, "len_sents": 62405, "len_sents_1": 23255, "n_sents": 413, "n_sents_1": 137, "venue": "ICLR2025", "decision": "withdraw", "L_title": 45, "L_abs": 1314, "title": "AcademicEval: Live Long-Context LLM Benchmark", "abs": "Large Language Models (LLMs) have achieved remarkable performance in long-context understanding. However, current long-context LLM benchmarks are limited by rigid context length and labor-intensive annotation, and the label leakage issue in LLM training also poses a pressing challenge. Therefore, we propose \\textsc{AcademicEval}, a live benchmark for evaluating LLMs over long-context generation tasks. \\textsc{AcademicEval} adopts papers on arXiv to introduce several academic writing tasks with long-context inputs, \\textit{i.e.}, \\textsc{Title}, \\textsc{Abstract}, \\textsc{Introduction}, and \\textsc{Related Work}, which cover a wide range of abstraction levels and require no manual labeling. Moreover, \\textsc{AcademicEval} integrates high-quality and expert-curated few-shot demonstrations from a collected co-author graph to enable flexible context length. Especially, \\textsc{AcademicEval} features an efficient live evaluation, ensuring no label leakage. We conduct holistic experiments on \\textsc{AcademicEval}, and the results illustrate that LLMs perform poorly on tasks with hierarchical abstraction levels and tend to struggle with long few-shot demonstrations, illustrating the challenge of our benchmark. We also provide insightful analysis for enhancing LLMs' long-context modeling capabilities.", "keywords": [ "Large Language Models", "Ultra-long Context Understanding", "Live Benchmark", "Long-context LLM Benchmarks" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "iR5qF9N1Ge", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Guangdong Bai", "gender": "unknown", "institution": "University of Queensland", "country": "", "position": "Associate Professor" }, { "name": "Peng-Fei Zhang", "gender": "Male", "institution": "University of Queensland", "country": "AU", "position": "Postdoc" }, { "name": "Zi Huang", "gender": "Female", "institution": "University of Queensland", "country": "AU", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 18, "n_ref": 43, "n_ref_all": 63, "n_fig": 3, "n_tab": 11, "L_tab": 11856, "n_element_tab": 579, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 6387, "n_element_tab_1": 280, "formula_len_all": 405, "formula_len_all_1": 339, "len_all": 170725, "len_all_1": 75336, "len_abs": 1978, "len_title": 125, "len_sents": 49946, "len_sents_1": 34240, "n_sents": 298, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 2028, "title": "MAA: Meticulous Adversarial Attack against Vision-Language Pre-trained Models", "abs": "Current adversarial attacks for evaluating the robustness of vision-language pre-trained (VLP) models in multi-modal tasks suffer from limited transferability, where attacks crafted for a specific model often struggle to generalize effectively across different models, limiting their utility in assessing robustness more broadly. This is mainly attributed to the over-reliance on model-specific features and regions, particularly in the image modality. In this paper, we propose an elegant yet highly effective method termed Meticulous Adversarial Attack (MAA) to fully exploit model-independent characteristics and vulnerabilities of individual samples, achieving enhanced generalizability and reduced model dependence. MAA emphasizes fine-grained optimization of adversarial images by developing a novel resizing and sliding crop (RScrop) technique, incorporating a multi-granularity similarity disruption (MGSD) strategy. \nRScrop efficiently enriches the initial adversarial examples by generating more comprehensive, diverse, and detailed perspectives of the images, establishing a robust foundation for capturing representative and intrinsic visual characteristics. Building on this, MGSD seeks to maximize %the layer- and component-wise feature% \nthe embedding distance between adversarial examples and their original counterparts across different granularities and hierarchical levels within the architecture of VLP models, thereby amplifying the impact of the adversarial perturbations and enhancing the efficacy of attacks across every layer and component of the model. Extensive experiments across diverse VLP models, multiple benchmark datasets, and a variety of downstream tasks demonstrate that MAA significantly enhances the effectiveness and transferability of adversarial attacks. A large cohort of performance studies is conducted to generate insights into the effectiveness of various model configurations, guiding future advancements in this domain. The source code is provided in the supplementary material.", "keywords": [ "Adversarial Attack", "Vision-Language Pre-trained Models", "Robustness" ], "rating_list": [ 5, 5, 8, 6, 5 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 2, 5, 4, 3, 4 ] }, { "paper_id": "iQtz3UJGRz", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Haike Xu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Piotr Indyk", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Sandeep Silwal", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 9, "n_ref_uni": 26, "n_ref": 56, "n_ref_all": 81, "n_fig": 11, "n_tab": 2, "L_tab": 686, "n_element_tab": 27, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 872, "n_element_tab_1": 25, "formula_len_all": 1805, "formula_len_all_1": 358, "len_all": 166567, "len_all_1": 65972, "len_abs": 5305, "len_title": 107, "len_sents": 47258, "len_sents_1": 32443, "n_sents": 386, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1237, "title": "A Bi-metric Framework for Efficient Nearest Neighbor Search", "abs": "We propose a new bi-metric framework for designing nearest neighbor data structures. Our framework assumes two dissimilarity functions: a ground-truth metric that is accurate but expensive to compute, and a proxy metric that is cheaper but less accurate. In both theory and practice, we show how to construct data structures using only the proxy metric such that the query procedure achieves the accuracy of the expensive metric, while only using a limited number of calls to both metrics. Our theoretical results instantiate this framework for two popular nearest neighbor search algorithms: DiskANN and Cover Tree. In both cases we show that, as long as the proxy metric used to construct the data structure approximates the ground-truth metric up to a bounded factor, our data structure achieves arbitrarily good approximation guarantees with respect to the ground-truth metric. On the empirical side, we apply the framework to the text retrieval problem with two dissimilarity functions evaluated by ML models with vastly different computational costs. We observe that for almost all data sets in the MTEB benchmark, our approach achieves a considerably better accuracy-efficiency tradeoff than the alternatives, such as re-ranking.", "keywords": [ "nearest neighbor search", "information retrieval" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "iQ0aOGx6dc", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 13, "author_info_list": [ { "name": "Amit S", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Applied Research" }, { "name": "Anshul Mittal", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Researcher" }, { "name": "Deepak Saini", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Jian Jiao", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Lakshya Kumar", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Applied Research" }, { "name": "Manik Varma", "gender": "Male", "institution": "Research, Microsoft", "country": "IN", "position": "Principal Researcher" }, { "name": "Pankaj Malhotra", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Principal Researcher" }, { "name": "Purushottam Kar", "gender": "Male", "institution": "Indian Institute of Technology, Kanpur", "country": "", "position": "Associate Professor" }, { "name": "Shikhar Mohan", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Applied Scientist" }, { "name": "Siddarth Asokan", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Researcher" }, { "name": "Soumen Chakrabarti", "gender": "Not Specified", "institution": "Indian Institute of Technology Bombay", "country": "", "position": "Professor" }, { "name": "Suchith Chidananda Prabhu", "gender": "Male", "institution": "Indian Institute of Technology, Delhi", "country": "IN", "position": "PhD student" }, { "name": "Sumeet Agarwal", "gender": "Male", "institution": "Indian Institute of Technology Delhi", "country": "", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 11, "n_ref_uni": 56, "n_ref": 159, "n_ref_all": 191, "n_fig": 3, "n_tab": 11, "L_tab": 4278, "n_element_tab": 513, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 2328, "n_element_tab_1": 217, "formula_len_all": 8326, "formula_len_all_1": 8506, "len_all": 193401, "len_all_1": 84813, "len_abs": 1506, "len_title": 109, "len_sents": 54056, "len_sents_1": 35953, "n_sents": 376, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1514, "title": "Graph Regularized Encoder Training for Extreme Classification", "abs": "Deep extreme classification (XC) aims to train an encoder and label classifiers to tag a data point with the most relevant subset of labels from a very large universe of labels. XC applications in ranking, recommendation and tagging routinely encounter tail labels, for which the amount of training data is exceedingly small. One way to tackle the tail label problem is to use additional data - often structured as a graph associated with documents and labels - graph metadata. Graph Convolutional Networks (GCNs) present a convenient but computationally expensive way to leverage this graph metadata and enhance model accuracies in these settings. However, GCNs struggle to make predictions for a novel test point when it has no edge in the graph. The paper notices that in these settings, it is much more effective to use graph data to regularize encoder training than to implement a GCN. Based on these insights, an alternative paradigm RAMEN is presented to utilize graph metadata in XC settings that offers a significant performance boost with zero increase in inference computational costs. RAMEN scales to datasets with millions of labels and offers prediction accuracy up to 15% higher on benchmark datasets than state of the art methods, including those that use graph metadata to train GCNs. RAMEN also offers 10% higher accuracy over the best baseline on a proprietary recommendation dataset sourced from click logs of a popular search engine. Code for RAMEN will be released publicly upon acceptance.", "keywords": [ "Extreme classification", "Lage scale recommendation", "Metadata", "Sponsored search", "ads", "intelligent advertisement" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 1, 2, 3, 1 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "iPYwddLhXR", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Jiao Xie", "gender": "Female", "institution": "East China Normal University", "country": "CN", "position": "Postdoc" }, { "name": "Jie Hu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Jincheng Liao", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Junbo Qiao", "gender": "Male", "institution": "East China Normal University", "country": "", "position": "PhD student" }, { "name": "Shaohui Lin", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Researcher" }, { "name": "Wei Li", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Yong Guo", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Postdoc" }, { "name": "Yulun Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhangxizi Qiu", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yi-Sophia-Wen", "gender": "Female", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 20, "n_ref": 49, "n_ref_all": 64, "n_fig": 11, "n_tab": 12, "L_tab": 9395, "n_element_tab": 1481, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 10012, "n_element_tab_1": 1079, "formula_len_all": 701, "formula_len_all_1": 701, "len_all": 157211, "len_all_1": 69925, "len_abs": 1340, "len_title": 113, "len_sents": 29356, "len_sents_1": 19742, "n_sents": 241, "n_sents_1": 155, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1356, "title": "Hi-Mamba: Hierarchical Mamba for Efficient Image Super-Resolution", "abs": "State Space Models (SSM), such as Mamba, have shown strong representation ability in modeling long-range dependency with linear complexity, achieving successful applications from high-level to low-level vision tasks. However, SSM's sequential nature necessitates multiple scans in different directions to compensate for the loss of spatial dependency when unfolding the image into a 1D sequence. This multi-direction scanning strategy significantly increases the computation overhead and is unbearable for high-resolution image processing. To address this problem, we propose a novel Hierarchical Mamba network, namely, Hi-Mamba, for image super-resolution (SR). Hi-Mamba consists of two key designs: (1) The Hierarchical Mamba Block (HMB) assembled by a Local SSM (L-SSM) and a Region SSM (R-SSM) both with the single-direction scanning, aggregates multi-scale representations to enhance the context modeling ability. (2) The Direction Alternation Hierarchical Mamba Group (DA-HMG) allocates the isomeric single-direction scanning into cascading HMBs to enrich the spatial relationship modeling.\nExtensive experiments demonstrate the superiority of Hi-Mamba across five benchmark datasets for efficient SR. For example, Hi-Mamba achieves a significant PSNR improvement of 0.29 dB on Manga109 for $\\times3$ SR, compared to the strong lightweight MambaIR.", "keywords": [ "Image Super-Resolution", "Mamba", "Structured State Space Models" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "iOy2pITOoH", "primary_area": "foundation or frontier models, including LLMs", "n_author": 16, "author_info_list": [ { "name": "Alek Andreev", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Chong You", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "David E Culler", "gender": "Male", "institution": "Google", "country": "US", "position": "Principal Researcher" }, { "name": "Felix Chern", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Felix Yu", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Henry Levy", "gender": "Male", "institution": "University of Washington, University of Washington", "country": "US", "position": "Emeritus" }, { "name": "Jeremiah J. Willcock", "gender": "Male", "institution": "Google", "country": "US", "position": "Software engineer" }, { "name": "Kan Wu", "gender": "unknown", "institution": "SystemsResearch@Google", "country": "US", "position": "Researcher" }, { "name": "Lin Chen", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Praneeth Netrapalli", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Prateek Jain", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Sanjiv Kumar", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Shreya Pathak", "gender": "Female", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Srinadh Bhojanapalli", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Suvinay Subramanian", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Zhipeng Jia", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 52, "n_formula_1": 24, "n_ref_uni": 75, "n_ref": 109, "n_ref_all": 186, "n_fig": 15, "n_tab": 4, "L_tab": 880, "n_element_tab": 141, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 461, "n_element_tab_1": 20, "formula_len_all": 3657, "formula_len_all_1": 1685, "len_all": 216519, "len_all_1": 68912, "len_abs": 1527, "len_title": 100, "len_sents": 57209, "len_sents_1": 30039, "n_sents": 483, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1559, "title": "Spark Transformer: How Many FLOPs is a Token Worth?", "abs": "This work introduces Spark Transformer, an architectural variant of the Transformer model that drastically reduces the FLOPs count while maintaining comparable quality and an identical parameter count. This reduction is achieved by introducing sparse activations in both the feedforward network (FFN) and the Attention mechanism. In the FFN, this sparsity engages only a subset of parameters for each input. In the Attention mechanism, it limits the number of tokens that each token attends to. We achieve this sparsity through statistical top-$k$, a lightweight approximate algorithm that is well-suited for accelerator hardware and minimizes training slowdown. Furthermore, Spark Transformer incorporates dedicated predictors to identify the activated entries. These predictors are formed by allocating a portion of the model's parameters and are trained jointly with the rest of the model. This approach distinguishes Spark Transformer from existing methods that introduce sparsity and predictors post-training, which often leads to increased training costs, additional model parameters, and complex modifications to the model architecture. Our Spark Transformer, pretrained using the Gemma 2 recipe, achieves competitive performance on standard benchmarks while exhibiting significant sparsity. Specifically, it utilizes only 8% nonzeros in the FFN activation and attends to a maximum of 256 tokens. This results in a 3.1$\\times$ reduction in FLOPs, yielding a 1.70$\\times$ speedup for prefill and a 1.79$\\times$ speedup for decoding on a 16-core CPU VM.", "keywords": [ "LLM", "activation sparsity", "inference efficiency" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "iOltCu4TPS", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Khachatur Dallakyan", "gender": "unknown", "institution": "Karolinska Institute Stockholm", "country": "SE", "position": "MS student" }, { "name": "Lee KinHei", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "MS student" }, { "name": "Licheng Zong", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Intern" }, { "name": "Xuesong Wang", "gender": "unknown", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yimin Fan", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Yixuan Wang", "gender": "Female", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Yu Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 36, "n_ref": 56, "n_ref_all": 73, "n_fig": 6, "n_tab": 1, "L_tab": 1155, "n_element_tab": 40, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 1071, "n_element_tab_1": 167, "formula_len_all": 141, "formula_len_all_1": 117, "len_all": 120786, "len_all_1": 66911, "len_abs": 1602, "len_title": 107, "len_sents": 33080, "len_sents_1": 30419, "n_sents": 232, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1617, "title": "Evaluating Single-Cell Foundation Models for Cell Retrieval", "abs": "Efficiently and accurately searching large-scale single-cell RNA-seq databases has been a long standing computational challenge. There is an increasing number of single-cell retrieval methods, particularly those based on single-cell foundation models, proposed in the literature. However, this field lacks a comprehensive benchmark among these methods. This gap exists due to the lack of standard evaluation metrics and comprehensive benchmark datasets. Addressing these challenges, we propose a comprehensive evaluation benchmark to assess the capabilities of 12 existing single-cell retrieval methods from three classes: non-machine learning method, VAE-based methods and single-cell foundation model (scFM) based methods. We propose a series of label-dependent and label-free evaluation metrics to assess the performance of single-cell retrieval methods. Through benchmarking across diverse settings (cross-platform, cross-species and cross-omics), our notable findings include: top scFMs such as UCE, scFoundation and SCimilarity show substantial overall advantage compared with other methods; traditional non-machine learning method perform well in cell retrieval thus should not be neglected; common cells retrieved by top methods share distinct gene expression patterns; label-free metrics have consistent evaluation outcome compared with label-based methods thus can be employed in a broader scenario. Our rigorous and comprehensive evaluation identifies the challenges and limitations of current single-cell retrieval methods and serves as foundation for further development of single-cell retrieval methods.", "keywords": [ "Single-cell Foundation Models", "Cell Retrieval", "Benchmarking" ], "rating_list": [ 6, 3, 5, 8, 5 ], "soundness_list": [ 1, 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 5, 3, 5, 3, 3 ] }, { "paper_id": "iOMnn1hSBO", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Bryan Wilder", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Carlos Miguel Pati\u00f1o", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "MS student" }, { "name": "Eric Horvitz", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Chief Scientific Officer" }, { "name": "Santiago Cortes-Gomez", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Yewon Byun", "gender": "unknown", "institution": "Machine Learning Department, School of Computer Science", "country": "", "position": "PhD student" }, { "name": "Zhiwei Steven Wu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 7, "n_ref_uni": 42, "n_ref": 65, "n_ref_all": 91, "n_fig": 9, "n_tab": 10, "L_tab": 2456, "n_element_tab": 260, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1117, "formula_len_all_1": 489, "len_all": 143731, "len_all_1": 69609, "len_abs": 1399, "len_title": 83, "len_sents": 49511, "len_sents_1": 36841, "n_sents": 349, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 1416, "title": "DECISION-FOCUSED UNCERTAINTY QUANTIFICATION", "abs": "There is increasing interest in ``decision-focused\" machine learning methods which train models to account for how their predictions are used in downstream optimization problems. Doing so can often improve performance on subsequent decision problems. However, current methods for uncertainty quantification do not incorporate any information at all about downstream decisions. We develop a framework based on conformal prediction to produce prediction sets that account for a downstream decision loss function, making them more appropriate to inform high-stakes decision-making. Our approach harnesses the strengths of conformal methods\u2014modularity, model-agnosticism, and statistical coverage guarantees\u2014while incorporating downstream decisions and user-specified utility functions. We prove that our methods retain standard coverage guarantees. Empirical evaluation across a range of datasets and utility metrics demonstrates that our methods achieve significantly lower decision loss compared to standard conformal methods. Additionally, we present a real-world use case in healthcare diagnosis, where our method effectively incorporates the hierarchical structure of dermatological diseases. It successfully generates sets with coherent diagnostic meaning, aiding the triage process during dermatology diagnosis and illustrating how our method can ground high-stakes decision-making on external domain knowledge.", "keywords": [ "Decision-focused learning", "decision making", "uncertainty quantification", "healthcare" ], "rating_list": [ 6, 6, 8, 8, 6 ], "soundness_list": [ 2, 2, 4, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 4, 3, 3 ], "confidence_list": [ 3, 4, 5, 3, 3 ] }, { "paper_id": "iOAcVOHvEN", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Baoxiang Wang", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Canzhe Zhao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Jing Dong", "gender": "Female", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "PhD student" }, { "name": "Shuai Li", "gender": "Female", "institution": "John Hopcroft Center, Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yutian Cheng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 41, "n_formula_1": 39, "n_ref_uni": 63, "n_ref": 165, "n_ref_all": 181, "n_fig": 0, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5045, "formula_len_all_1": 5112, "len_all": 157383, "len_all_1": 71048, "len_abs": 1100, "len_title": 151, "len_sents": 34775, "len_sents_1": 30152, "n_sents": 207, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1133, "title": "Learning Imperfect Information Extensive-form Games with Last-iterate Convergence under Bandit Feedback", "abs": "We study learning the approximate Nash equilibrium (NE) policy profile in two-player zero-sum imperfect information extensive-form games (IIEFGs) with last-iterate convergence. The algorithms in previous works studying this problem either require full-information feedback or only have asymptotic convergence rates. In contrast, we study IIEFGs in the formulation of partially observable Markov games (POMGs) with the perfect-recall assumption and bandit feedback, where the knowledge of the game is not known a priori and only the rewards of the experienced information set and action pairs are revealed to the learners in each episode. Our algorithm utilizes a negentropy regularizer weighted by a virtual transition over information set-action space. By carefully designing the virtual transition together with the leverage of the entropy regularization technique, we prove that our algorithm converges to the NE of IIEFGs with a provable finite-time convergence rate of $\\widetilde{O}(k^{-\\frac{1}{8}})$ with high probability under bandit feedback, thus answering the second question of \\citet{Fiegel2023adapting} affirmatively.", "keywords": [ "Extensive-form games; partially observable Markov games (POMGs); last-iterate convergence" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "iNvxHAN1J8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Jifan Zhang", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "PhD student" }, { "name": "Robert D Nowak", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 16, "n_ref_uni": 55, "n_ref": 89, "n_ref_all": 105, "n_fig": 4, "n_tab": 4, "L_tab": 791, "n_element_tab": 105, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 528, "n_element_tab_1": 88, "formula_len_all": 1250, "formula_len_all_1": 1344, "len_all": 162729, "len_all_1": 68940, "len_abs": 1458, "len_title": 132, "len_sents": 47886, "len_sents_1": 31778, "n_sents": 365, "n_sents_1": 244, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1464, "title": "SIEVE: General Purpose Data Filtering System Matching GPT-4o Accuracy at 1% the Cost", "abs": "Creating specialized large language models requires vast amounts of clean, special purpose data for training and fine-tuning. With only a handful of existing large-scale, domain-specific datasets, creation of new datasets is required in most applications. This requires the development of new application-specific filtering of web-scale data. Filtering with a high-performance, general-purpose LLM such as GPT-4o can be highly effective, but this is extremely expensive at web-scale. This paper proposes SIEVE, a lightweight alternative that matches GPT-4o accuracy at a fraction of the cost. SIEVE can perform up to 500 filtering operations for the cost of one GPT-4o filtering call. The key to SIEVE is a seamless integration of GPT-4o and lightweight T5 models, using active learning to fine-tune T5 in the background with a small number of calls to GPT-4o. Once trained, it performs as well as GPT-4o at a tiny fraction of the cost. We experimentally validate SIEVE on the OpenWebText dataset, using five highly customized filter tasks targeting high quality and domain-specific content. Our results demonstrate the effectiveness and efficiency of our method in curating large, high-quality datasets for language model training at a substantially lower cost (1\\%) than existing techniques. To further validate SIEVE, experiments show that SIEVE and GPT-4o achieve similar accuracy, with human evaluators preferring SIEVE's filtering results to those of GPT-4o.", "keywords": [ "Data Filtering", "Large Language Models" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "iNmVX9lx9l", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Johannes Hertrich", "gender": "Male", "institution": "Universit\u00e9 Paris Dauphine - PSL", "country": "FR", "position": "Postdoc" }, { "name": "Michael Quellmalz", "gender": "unknown", "institution": "Technische Universit\u00e4t Berlin", "country": "DE", "position": "Researcher" }, { "name": "Tim Jahn", "gender": "unknown", "institution": "Technische Universit\u00e4t Berlin", "country": "DE", "position": "Postdoc" } ], "n_formula": 124, "n_formula_1": 23, "n_ref_uni": 67, "n_ref": 143, "n_ref_all": 213, "n_fig": 11, "n_tab": 1, "L_tab": 161, "n_element_tab": 29, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 641, "n_element_tab_1": 49, "formula_len_all": 12528, "formula_len_all_1": 1881, "len_all": 234913, "len_all_1": 61981, "len_abs": 608, "len_title": 96, "len_sents": 63321, "len_sents_1": 25157, "n_sents": 577, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 611, "title": "Fast Summation of Radial Kernels via QMC Slicing", "abs": "The fast computation of large kernel sums is a challenging task, which arises as a subproblem in any kernel method. We approach the problem by slicing, which relies on random projections to one-dimensional subspaces and fast Fourier summation. We prove bounds for the slicing error and propose a quasi-Monte Carlo (QMC) approach for selecting the projections based on spherical quadrature rules. Numerical examples demonstrate that our QMC-slicing approach significantly outperforms existing methods like (QMC-)random Fourier features, orthogonal Fourier features or non-QMC slicing on standard test datasets.", "keywords": [ "fast kernel summation", "slicing", "quasi-Monte Carlo", "non-equispaced fast Fourier transforms", "random Fourier features" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "iNcEChuYXD", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Ida Momennejad", "gender": "Female", "institution": "Research, Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Shanka Subhra Mondal", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Taylor Whittington Webb", "gender": "Male", "institution": "Research, Microsoft", "country": "US", "position": "Postdoc" } ], "n_formula": 16, "n_formula_1": 2, "n_ref_uni": 64, "n_ref": 115, "n_ref_all": 141, "n_fig": 6, "n_tab": 20, "L_tab": 5858, "n_element_tab": 418, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 543, "n_element_tab_1": 62, "formula_len_all": 1420, "formula_len_all_1": 262, "len_all": 237133, "len_all_1": 68890, "len_abs": 1505, "len_title": 125, "len_sents": 71179, "len_sents_1": 33941, "n_sents": 563, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1517, "title": "Improving Planning with Large Language Models: A Modular Agentic Architecture", "abs": "Large language models (LLMs) demonstrate impressive performance on a wide variety of tasks, but they often struggle with tasks that require multi-step reasoning or goal-directed planning. Both cognitive neuroscience and reinforcement learning (RL) have proposed a number of interacting functional components that together implement search and evaluation in multi-step decision making. These components include conflict monitoring, state prediction, state evaluation, task decomposition, and orchestration. To improve planning with LLMs, we propose an agentic architecture, the Modular Agentic Planner (MAP), in which planning is accomplished via the recurrent interaction of the specialized modules mentioned above, each implemented using an LLM. MAP improves planning through the interaction of specialized modules that break down a larger problem into multiple brief automated calls to the LLM. We evaluate MAP on three challenging planning tasks -- graph traversal, Tower of Hanoi, and the PlanBench benchmark -- as well as an NLP task requiring multi-step reasoning (strategyQA). We find that MAP yields significant improvements over both standard LLM methods (zero-shot prompting, in-context learning) and competitive baselines (chain-of-thought, multi-agent debate, and tree-of-thought), can be effectively combined with smaller and more cost-efficient LLMs (Llama3-70B), and displays superior transfer across tasks. These results suggest the benefit of a modular and multi-agent approach to planning with LLMs.", "keywords": [ "large language models", "planning", "LLM agents", "generalization" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "iN7EIQRUbF", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Quang Pham", "gender": "Male", "institution": "A*STAR", "country": "", "position": "Researcher" }, { "name": "Savitha Ramasamy", "gender": "Female", "institution": "Institute for Infocomm Research, Agency for Science, Technology and Research, Singapore", "country": "SG", "position": "Principal Researcher" }, { "name": "Xiaoli Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Cuong Ngoc Nguyen", "gender": "Male", "institution": "Durham University", "country": "GB", "position": "PhD student" }, { "name": "Cuong V Nguyen", "gender": "Male", "institution": "Durham University", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 148, "n_ref_all": 158, "n_fig": 3, "n_tab": 7, "L_tab": 2305, "n_element_tab": 212, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1842, "n_element_tab_1": 165, "formula_len_all": 390, "formula_len_all_1": 380, "len_all": 133299, "len_all_1": 68125, "len_abs": 1385, "len_title": 118, "len_sents": 37306, "len_sents_1": 30385, "n_sents": 268, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1413, "title": "Advancing Few-shot Continual Learning via Selective Knowledge Transfer", "abs": "Continual learning with large language models (LLMs) is a promising and chal\u0002lenging research that greatly impacts many applications. Existing solutions treat previous tasks equally, making them vulnerable to task interference, lacking scal\u0002ability with a large number of tasks, and oblivious to the intrinsic relationships among tasks. This work presents selective knowledge transfer (SKT), a novel\nand principled framework for continual learning with LLMs. SKT aims to maximize positive knowledge transfer while systematically minimizing the effects of irrelevant information from dissimilar tasks. To this end, SKT first assesses the degree of interference between the current and previous tasks and then selectively aggregates the tasks that maximize knowledge transfer for continual\ntraining. In addition, we integrate SKT into the current state-of-the-art continual language learning algorithm, Progressive Prompts, to introduce Log-evidence Progressive Prompts (LePP), which facilitate knowledge transfer between tasks. Comprehensive evaluations on challenging few-shot continual learning benchmarks demonstrate that LePP can surpass existing baselines for continual learning with LLMs with minimal overhead. Our extensive ablation studies reveal that SKT can discover useful task correlations without any prior knowledge, many of which align with human evaluations. Code will be published upon acceptance.", "keywords": [ "continual learning", "transfer learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "iN64nSYt0z", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Ali Maatouk", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Researcher" }, { "name": "Antonio De Domenico", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Fadhel Ayed", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Pedro Luiz Silva", "gender": "unknown", "institution": "\u00c9cole Polytechnique", "country": "FR", "position": "MS student" } ], "n_formula": 18, "n_formula_1": 10, "n_ref_uni": 20, "n_ref": 33, "n_ref_all": 50, "n_fig": 8, "n_tab": 3, "L_tab": 983, "n_element_tab": 155, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1325, "formula_len_all_1": 480, "len_all": 129859, "len_all_1": 48202, "len_abs": 1635, "len_title": 77, "len_sents": 37612, "len_sents_1": 22828, "n_sents": 285, "n_sents_1": 166, "venue": "ICLR2025", "decision": "withdraw", "L_title": 29, "L_abs": 621, "title": "Pay Attention to What Matters", "abs": "Despite the remarkable success of Large Language Models (LLMs), they still exhibit a limited capability to align their outputs to the user instructions. In this work, we introduce a simple and effective method, which we name as GUIDE, that mechanistically increases attention scores in instruction tokens. To support this operation, we present Influence, a novel metric that highlights how the user's instructions propagate with transformer layers and impact the LLM output. Our results show that GUIDE improves the accuracy of following certain instructions 29.4% to 60.4 %, outperforming natural prompting alternatives.", "keywords": [ "Large Language Models", "Mechanistic Interventions", "Alignment", "Transformer Interpretability" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "iMnd6c5bAa", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Jungtaek Kim", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Postdoc" } ], "n_formula": 32, "n_formula_1": 19, "n_ref_uni": 63, "n_ref": 136, "n_ref_all": 168, "n_fig": 13, "n_tab": 2, "L_tab": 2493, "n_element_tab": 270, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 2229, "n_element_tab_1": 258, "formula_len_all": 1612, "formula_len_all_1": 834, "len_all": 172820, "len_all_1": 68656, "len_abs": 1371, "len_title": 131, "len_sents": 44644, "len_sents_1": 27826, "n_sents": 344, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1375, "title": "Density Ratio Estimation-based Bayesian Optimization with Semi-Supervised Learning", "abs": "Bayesian optimization has attracted huge attention from diverse research areas in science and engineering, since it is capable of efficiently finding a global optimum of an expensive-to-evaluate black-box function. In general, a probabilistic regression model is widely used as a surrogate function to model an explicit distribution over function evaluations given an input to estimate and a training dataset. Beyond the probabilistic regression-based methods, density ratio estimation-based Bayesian optimization has been suggested in order to estimate a density ratio of the groups relatively close and relatively far to a global optimum. Developing this line of research further, supervised classifiers are employed to estimate a class probability for the two groups instead of a density ratio. However, the supervised classifiers used in this strategy are prone to be overconfident for known knowledge on global solution candidates. Supposing that we have access to unlabeled points, e.g., predefined fixed-size pools, we propose density ratio estimation-based Bayesian optimization with semi-supervised learning to solve this challenge. Finally, we show the empirical results of our methods and several baseline methods in two distinct scenarios with unlabeled point sampling and a fixed-size pool and analyze the validity of our proposed methods in diverse experiments.", "keywords": [ "Bayesian optimization", "Density ratio estimation-based Bayesian optimization", "Bayesian optimization with semi-supervised learning" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 3, 5 ] }, { "paper_id": "iMJpmcYucq", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Arto Klami", "gender": "Male", "institution": "University of Helsinki", "country": "FI", "position": "Associate Professor" }, { "name": "Bernardo Williams", "gender": "Male", "institution": "University of Helsinki", "country": "FI", "position": "PhD student" }, { "name": "Hanlin Yu", "gender": "Male", "institution": "University of Helsinki", "country": "FI", "position": "PhD student" }, { "name": "Hoang Phuc Hau Luu", "gender": "Male", "institution": "University of Helsinki", "country": "FI", "position": "Postdoc" }, { "name": "Marcelo Hartmann", "gender": "unknown", "institution": "University of Helsinki", "country": "", "position": "Postdoc" } ], "n_formula": 82, "n_formula_1": 32, "n_ref_uni": 31, "n_ref": 81, "n_ref_all": 103, "n_fig": 10, "n_tab": 2, "L_tab": 1365, "n_element_tab": 37, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10464, "formula_len_all_1": 3323, "len_all": 136588, "len_all_1": 66477, "len_abs": 1270, "len_title": 141, "len_sents": 40587, "len_sents_1": 28885, "n_sents": 378, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1272, "title": "Stochastic variance-reduced Gaussian variational inference on the Bures-Wasserstein manifold", "abs": "Optimization in the Bures-Wasserstein space has been gaining popularity in the machine learning community since it draws connections between variational inference and Wasserstein gradient flows. The variational inference objective function of Kullback\u2013Leibler divergence can be written as the sum of the negative entropy and the potential energy, making forward-backward Euler the method of choice. Notably, the backward step admits a closed-form solution in this case, facilitating the practicality of the scheme. However, the forward step is no longer exact since the Bures-Wasserstein gradient of the potential energy involves \"intractable\" expectations. Recent approaches propose using the Monte Carlo method -- in practice a single-sample estimator -- to approximate these terms, resulting in high variance and poor performance. We propose a novel variance-reduced estimator based on the principle of control variates. We theoretically show that this estimator has a smaller variance than the Monte-Carlo estimator in scenarios of interest. We also prove that variance reduction helps improve the optimization bounds of the current analysis. We demonstrate that the proposed estimator gains order-of-magnitude improvements over the previous Bures-Wasserstein methods.", "keywords": [ "Variational inference", "Bures-Wasserstein space", "Riemannian manifold", "optimization" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 2, 4, 2, 4 ], "presentation_list": [ 3, 4, 2, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "iMI4HRpZFc", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Doina Precup", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Research Team Lead" }, { "name": "Harry Zhao", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "Tristan Sylvain", "gender": "Unspecified", "institution": "Borealis AI", "country": "", "position": "Machine Learning Researcher" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 25, "n_ref": 51, "n_ref_all": 76, "n_fig": 20, "n_tab": 4, "L_tab": 4587, "n_element_tab": 75, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1232, "n_element_tab_1": 36, "formula_len_all": 78, "formula_len_all_1": 78, "len_all": 139308, "len_all_1": 68134, "len_abs": 784, "len_title": 72, "len_sents": 52499, "len_sents_1": 35452, "n_sents": 347, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 787, "title": "Identifying and Addressing Delusions for Target-Directed Decision Making", "abs": "Target-directed agents utilize self-generated targets, to guide their behaviors for better generalization. These agents are prone to blindly chasing problematic targets, resulting in worse generalization and safety catastrophes. We show that these behaviors can be results of delusions, stemming from improper designs around training: the agent may naturally come to hold false beliefs about certain targets. We identify delusions via intuitive examples in controlled environments, and investigate their causes and mitigations. With the insights, we demonstrate how we can make agents address delusions preemptively and autonomously. We validate empirically the effectiveness of the proposed strategies in correcting delusional behaviors and improving out-of-distribution generalization.", "keywords": [ "delusions", "hallucination", "planning", "generalization", "reinforcement learning" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 1, 3, 3 ] }, { "paper_id": "iM7MfzbF1B", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Alex Aiken", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Visiting Professor" }, { "name": "Allen Nie", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Anjiang Wei", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" }, { "name": "Ke Wang", "gender": "unknown", "institution": "Visa Research", "country": "US", "position": "Research Scientist" }, { "name": "Rohan Yadav", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" }, { "name": "Thiago S. F. X. Teixeira", "gender": "Male", "institution": "Intel", "country": "US", "position": "Researcher" }, { "name": "Wonchan Lee", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Senior Software Engineer" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 75, "n_ref_all": 105, "n_fig": 15, "n_tab": 1, "L_tab": 207, "n_element_tab": 44, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 800, "formula_len_all_1": 106, "len_all": 155971, "len_all_1": 57965, "len_abs": 1558, "len_title": 126, "len_sents": 49836, "len_sents_1": 29362, "n_sents": 341, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1596, "title": "MAGE: Leveraging LLMs for Automated Mapper Generation in Parallel Programming", "abs": "Efficiently mapping tasks to processors and data to memories is a cornerstone of parallel programming to achieve high performance. Traditionally, this critical task has been handled by expert-crafted mapper programs, tailored for specific machine architectures and problem domains. However, creating customized mappers for each unique application is labor-intensive and time-consuming. Large language models (LLMs) have recently demonstrated remarkable capabilities in understanding and generating code, as well as in self-improvement for optimizing specific performance metrics. Inspired by these advancements, we introduce the task of mapper generation (MAGE), which frames generating high-performance mappers as a discrete optimization problem aimed at maximizing compute throughput. To solve this optimization problem, we leverage reinforcement learning (RL) to guide LLMs in the mapper generation process. At the core of our approach lies a novel domain-specific language (DSL), which provides a high-level interface for LLMs to generate the mapper code without getting entangled with complicated, low-level system programming. Moreover, our DSL defines a structured and constrained search space for RL to explore, guiding LLMs to discover the optimal mapping policy. The evaluation shows that our LLM-generated mappers can surpass expert-written mappers in performance, achieving up to 34% speedup across 9 benchmarks. Notably, our approach improves the throughput of parallel matrix multiplication algorithms by up to 31%, reducing development time from several days to just a few minutes.", "keywords": [ "large language models", "reinforcement learning", "domain-specific language", "discrete optimization", "performance optimization" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 3, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "iLUcsecZJp", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Shiguang Wu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Yaqing Wang", "gender": "Female", "institution": "Beijing Institute of Mathematical Sciences and Applications", "country": "CN", "position": "Associate Professor" }, { "name": "quanming yao", "gender": "Male", "institution": "Department of Electronic Engineering, Tsinghua University", "country": "", "position": "Assistant Professor" } ], "n_formula": 46, "n_formula_1": 12, "n_ref_uni": 46, "n_ref": 105, "n_ref_all": 142, "n_fig": 12, "n_tab": 1, "L_tab": 191, "n_element_tab": 4, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3809, "formula_len_all_1": 1138, "len_all": 183048, "len_all_1": 70574, "len_abs": 1242, "len_title": 148, "len_sents": 55970, "len_sents_1": 32297, "n_sents": 445, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1156, "title": "Why In-Context Learning Models are Good Few-Shot Learners?", "abs": "We explore in-context learning (ICL) models from a learning-to-learn perspective. Unlike studies that identify specific learning algorithms in ICL models, we compare ICL models with typical meta-learners to understand their superior performance. We theoretically prove the expressiveness of ICL models as learning algorithms and examine their learnability and generalizability. \nOur findings show that ICL with transformers \ncan effectively construct data-dependent learning algorithms instead of directly follow existing ones \n(including gradient-based, metric-based, and amortization-based meta-learners). \nThe construction of such learning algorithm is determined by the pre-training process, as a function fitting the training distribution, which raises generalizability as an important issue.\nWith above understanding, we propose strategies to transfer techniques for classical deep networks to meta-level to further improve ICL. As examples, we implement meta-level meta-learning for domain adaptability with limited data and meta-level curriculum learning for accelerated convergence during pre-training, demonstrating their empirical effectiveness.", "keywords": [ "In-Context Learning", "Meta-Learning" ], "rating_list": [ 3, 8, 6, 6 ], "soundness_list": [ 1, 4, 3, 3 ], "presentation_list": [ 1, 4, 3, 3 ], "contribution_list": [ 1, 3, 4, 3 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "iL9A4e8RdS", "primary_area": "interpretability and explainable AI", "n_author": 1, "author_info_list": [ { "name": "Amr S. Mohamed", "gender": "unknown", "institution": "University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 18, "n_ref": 29, "n_ref_all": 39, "n_fig": 8, "n_tab": 2, "L_tab": 174, "n_element_tab": 23, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 960, "formula_len_all_1": 672, "len_all": 73681, "len_all_1": 69088, "len_abs": 1296, "len_title": 76, "len_sents": 27616, "len_sents_1": 26482, "n_sents": 198, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 28, "L_abs": 1300, "title": "Explanation using Simulation", "abs": "In safety-critical domains, such as industrial systems, the lack of explainability in predictive `black-box' machine learning models can hinder trust and adoption. Standard explainability techniques, while powerful, often require deep expertise in data analytics and machine learning and fail to align with the sequential, dynamic nature of data in these environments. In this paper, we propose a novel explainability framework that leverages reinforcement learning (RL) to support model predictions with visual explanations based on dynamical system simulation. By training RL agents to simulate events that require prediction, we use these agents' critics to make classifications. Next, we employ the actors of the RL agents to simulate the potential future trajectories underlying these classifications, providing visual explanations that are more intuitive and align with the expertise of industrial domain experts. We demonstrate the applicability of this method through a case study involving monitoring a small industrial system for cyberattacks, showing how our framework generates actionable predictions that are supported with visual explanations. This approach aims to bridge the gap between advanced machine learning models and their real-world deployment in safety-critical environments.", "keywords": [ "Explainable AI", "Reinforcement learning", "Dynamical systems" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 3, 1, 1 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "iKsTtpzBtc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Minsu Park", "gender": "unknown", "institution": "New York University, Abu Dhabi", "country": "AE", "position": "Assistant Professor" }, { "name": "Nima Dehmamy", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Pascal Mettes", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Pratik Kayal", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 56, "n_ref_all": 80, "n_fig": 13, "n_tab": 5, "L_tab": 2235, "n_element_tab": 315, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 424, "n_element_tab_1": 23, "formula_len_all": 79, "formula_len_all_1": 56, "len_all": 189060, "len_all_1": 57016, "len_abs": 1748, "len_title": 61, "len_sents": 73050, "len_sents_1": 30392, "n_sents": 488, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1742, "title": "Large Language Models Are Natural Video Popularity Predictors", "abs": "Predicting video popularity is typically formalized as a supervised learning problem, where models classify videos as popular or unpopular. Traditional approaches rely heavily on meta-information and aggregated user engagement data, but video popularity is highly context-dependent, influenced by cultural, social, and temporal factors that these approaches fail to capture. We argue that Large Language Models (LLMs), with their deep contextual awareness, are well-suited to address these challenges. A key difficulty, however, lies in bridging the modality gap between pixel-based video data and token-based LLMs. To overcome this, we transform frame-level visual data into sequential text representations using Vision-Language Models (VLMs), enabling LLMs to process multimodal video content\u2014titles, frame-based descriptions, and captions\u2014and capture rich contextual information for more accurate predictions. Evaluating on a newly introduced dataset of 17,000 videos, we show that while a supervised neural network using content embeddings achieved 80% accuracy, our LLM-based method reached 82% without fine-tuning. A combined approach, integrating the neural network's predictions into the LLM, further improved accuracy to 85.5%. Additionally, the LLM generates interpretable hypotheses explaining its predictions based on theoretically sound attributes. Survey-based manual validations confirm the quality of these hypotheses and address concerns about hallucinations in the video-to-text conversion process. Our findings highlight that LLMs, equipped with textually transformed multimodal representations, offer a powerful, interpretable, and data-efficient solution to the context-dependent challenge of video popularity prediction.", "keywords": [ "Large Language Models (LLMs)", "Vision-Language Models (VLMs)", "Multimodal Textual Representations", "Video Popularity Prediction" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "iKgQOAtvsD", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Qizhang Li", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Xiaochen Yang", "gender": "Female", "institution": "University of Glasgow", "country": "GB", "position": "Lecturer" }, { "name": "Yiwen Guo", "gender": "Male", "institution": "ByteDance", "country": "CN", "position": "Researcher" }, { "name": "Wangmeng Zuo", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 87, "n_ref_all": 115, "n_fig": 5, "n_tab": 6, "L_tab": 1703, "n_element_tab": 157, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2254, "n_element_tab_1": 172, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 196328, "len_all_1": 71840, "len_abs": 1973, "len_title": 133, "len_sents": 74875, "len_sents_1": 35021, "n_sents": 522, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1635, "title": "Deciphering the Chaos: Enhancing Jailbreak Attacks via Adversarial Prompt Translation", "abs": "Automatic adversarial prompt generation provides remarkable success in jailbreaking safely-aligned large language models (LLMs). Existing gradient-based attacks, while demonstrating outstanding performance in jailbreaking white-box LLMs, often generate garbled adversarial prompts with chaotic appearance. These adversarial prompts are difficult to transfer to other LLMs, hindering their performance in attacking unknown victim models. In this paper, for the first time, we delve into the semantic meaning embedded in garbled adversarial prompts and propose a novel method that *\"translate\"* them into coherent, human-readable natural language adversarial prompts. In this way, we can effectively uncover the semantic information that triggers vulnerabilities in the model and unambiguously transfer it to the victim model, without overlooking the adversarial information hidden in the garbled text, to enhance jailbreak attacks. It also offers a new approach to discovering effective designs for jailbreak prompts, advancing the understanding of jailbreak attacks. Experimental results demonstrate that our method significantly improves the success rate of jailbreak attacks against various safety-aligned LLMs and outperforms state-of-the-arts by a large margin. With at most 10 queries, our method achieves an average attack success rate of 81.8% in attacking 7 commercial closed-source LLMs, including GPT and Claude-3 series, on HarmBench. Our method also achieves over 90% attack success rates against Llama-2-Chat models on AdvBench, despite their outstanding resistance to jailbreaks. Our code will be made publicly available.", "keywords": [ "Large lanugage model", "Jailbreak attack", "adversarial prompt" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "iKLSISIPH7", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Min-hwan Oh", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" }, { "name": "Junghun Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Postdoc" } ], "n_formula": 103, "n_formula_1": 46, "n_ref_uni": 24, "n_ref": 58, "n_ref_all": 85, "n_fig": 3, "n_tab": 3, "L_tab": 228, "n_element_tab": 122, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 17369, "formula_len_all_1": 5168, "len_all": 128823, "len_all_1": 59185, "len_abs": 957, "len_title": 102, "len_sents": 31546, "len_sents_1": 23392, "n_sents": 351, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 993, "title": "Stochastic Matching Bandits under Preference Feedback", "abs": "In this study, we propose a new bandit framework of stochastic matching employing the Multinomial Logit (MNL) choice model with feature information. In this framework, agents on one side are assigned to arms on the other side, and each arm stochastically accepts an agent among the assigned pool of agents based on its unknown preference, allowing a possible outside option of not accepting any. \nThe objective is to minimize regret by maximizing the probability of successful matching. \nFor this framework, we first propose an elimination-based algorithm that achieves a regret bound of $\\tilde{O}\\big(K\\sqrt{rKT} \\big)$ over time horizon $T$, where $K$ is the number of arms and $r$ is the rank of feature space. Furthermore, we propose an approach to resolve the computation issue regarding combinatorial optimization in the algorithm.\nLastly, we evaluate the performances of our algorithm through experiments comparing with the existing showing the superior performances of our algorithm.", "keywords": [ "Matching bandits", "Preference Feedback" ], "rating_list": [ 5, 3, 6, 5, 5 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "iKI7wT6fCP", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Bernhard Sick", "gender": "Male", "institution": "Universit\u00e4t Kassel", "country": "", "position": "Full Professor" }, { "name": "Josephine Thomas", "gender": "unknown", "institution": "Universit\u00e4t Kassel", "country": "DE", "position": "Group Leader" }, { "name": "Pascal Plettenberg", "gender": "Male", "institution": "Universit\u00e4t Kassel", "country": "DE", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 18, "n_ref_uni": 32, "n_ref": 61, "n_ref_all": 74, "n_fig": 7, "n_tab": 10, "L_tab": 2162, "n_element_tab": 158, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 443, "n_element_tab_1": 30, "formula_len_all": 1503, "formula_len_all_1": 1360, "len_all": 119551, "len_all_1": 58953, "len_abs": 1188, "len_title": 74, "len_sents": 39430, "len_sents_1": 27965, "n_sents": 280, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 26, "L_abs": 1192, "title": "Flow Graph Neural Networks", "abs": "Graph Neural Networks (GNNs) have become essential for learning from graph-structured data. However, existing GNNs do not consider the conservation law inherent in graphs associated with a flow of physical resources, such as electrical current in power grids or traffic in transportation networks. To address this limitation and enhance the performance on tasks where accurate modeling of resource flows is crucial, we propose Flow Graph Neural Networks (FlowGNNs). This novel GNN framework adapts existing graph attention mechanisms to reflect the conservation of resources by distributing a node's message among its outgoing edges instead of allowing arbitrary duplication of the node's information. We further extend this framework to directed acyclic graphs (DAGs), enabling discrimination between non-isomorphic flow graphs that would otherwise be indistinguishable for standard GNNs tailored to DAGs. We validate our approach through extensive experiments on two different flow graph domains\u2014electronic circuits and power grids\u2014and demonstrate that the proposed framework enhances the performance of traditional GNN architectures on both graph-level classification and regression tasks.", "keywords": [ "Graph Neural Networks", "Graph Attention Networks", "Directed Acyclic Graphs", "Power Grids", "Electronic Circuits" ], "rating_list": [ 3, 5, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 4, 4 ] }, { "paper_id": "iKDbLpVgQc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Bolei Zhou", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "Ceyuan Yang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Chaoyang Wang", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Gordon Wetzstein", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Hsin-Ying Lee", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Qihang Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postgraduate student" }, { "name": "Yinghao Xu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 48, "n_ref": 87, "n_ref_all": 105, "n_fig": 13, "n_tab": 1, "L_tab": 139, "n_element_tab": 27, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 139, "n_element_tab_1": 27, "formula_len_all": 287, "formula_len_all_1": 288, "len_all": 129254, "len_all_1": 48037, "len_abs": 357, "len_title": 133, "len_sents": 25996, "len_sents_1": 20983, "n_sents": 242, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1076, "title": "3DitScene: Editing Any Scene via Language-guided Disentangled Gaussian Splatting", "abs": "Scene image editing is crucial for entertainment, photography, and advertising design. Existing methods solely focus on either 2D individual object or 3D global scene editing. This results in a lack of a unified approach to effectively control and manipulate scenes at the 3D level with different levels of granularity. In this work, we propose 3DitScene, a novel and unified scene editing framework leveraging language-guided disentangled Gaussian Splatting that enables seamless editing from 2D to 3D, allowing precise control over scene composition and individual objects. We first incorporate 3D Gaussians that are refined through generative priors and optimization techniques. Language features from CLIP then introduce semantics into 3D geometry for object disentanglement. With the disentangled Gaussians, 3DitScene allows for manipulation at both the global and individual levels, revolutionizing creative expression and empowering control over scenes and objects. Experimental results demonstrate the effectiveness and versatility of 3DitScene in scene image editing.", "keywords": [ "image editting", "gaussian splatting", "3D" ], "rating_list": [ 5, 5, 5, 8, 6 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 4, 3 ], "confidence_list": [ 3, 2, 4, 3, 4 ] }, { "paper_id": "iJi7nz5Cxc", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Fu-Yun Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Jingtan Piao", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Keqiang Sun", "gender": "Male", "institution": "The Chinese University of Hong Kong, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yunhao Shui", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 27, "n_ref": 63, "n_ref_all": 76, "n_fig": 8, "n_tab": 4, "L_tab": 2472, "n_element_tab": 416, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 2026, "n_element_tab_1": 155, "formula_len_all": 410, "formula_len_all_1": 387, "len_all": 99571, "len_all_1": 48978, "len_abs": 1248, "len_title": 158, "len_sents": 25861, "len_sents_1": 20640, "n_sents": 205, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 108, "L_abs": 1293, "title": "Diffusion-NPO: Negative Preference Optimization for Better Preference Aligned Generation of Diffusion Models", "abs": "Diffusion models have made substantial advances in image generation, yet models trained on large, unfiltered datasets often yield outputs misaligned with human preferences. Numerous methods have already been proposed to fine-tune pre-trained diffusion models, achieving notable improvements in aligning generated outputs with human preferences. However, we point out that existing preference alignment methods neglect the critical role of handling unconditional/negative-conditional outputs, leading to a diminished capacity to avoid generating undesirable outcomes. This oversight limits the efficacy of classifier-free guidance (CFG), which relies on the contrast between conditional generation and unconditional/negative-conditional generation to optimize output quality. In response, we propose a straightforward but consistently effective approach that involves training a model specifically attuned to negative preferences. This method does not require new training strategies or datasets but rather involves minor modifications to existing techniques. Our approach integrates seamlessly with models such as SD15, SDXL, video diffusion models and models that have undergone preference optimization, consistently enhancing their ability to produce more human preferences aligned outputs.", "keywords": [ "Diffusion", "Preference Optimization" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "iIrvKrtwnZ", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Bin Liang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Hang Liu", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "MS student" }, { "name": "Houde Liu", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xueqian Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yi Cheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "junlong wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 9, "n_ref": 18, "n_ref_all": 45, "n_fig": 9, "n_tab": 8, "L_tab": 1977, "n_element_tab": 156, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2747, "n_element_tab_1": 66, "formula_len_all": 801, "formula_len_all_1": 668, "len_all": 126076, "len_all_1": 58749, "len_abs": 212, "len_title": 98, "len_sents": 38194, "len_sents_1": 25073, "n_sents": 318, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1298, "title": "HuRi : Humanoid Robots Adaptive Risk-ware Distributional Reinforcement Learning for Robust Control", "abs": "Due to the high complexity of bipedal locomotion, the locomotion control of humanoid robots requires precise adjustment of the balance system to adapt to the varying environment conditions. In the past, few studies have explicitly incorporated risk factors into robot policy training, and lacked the ability to adaptively adjust the risk sensitivity for different risky environment conditions. This deficiency impacts the agent\u2019s exploration during training and thus fail to select the optimal action in the risky environment. We propose an adaptive risk-aware policy(HuRi) based on distributional reinforcement learning. In Dist. RL, the policy control the risk sensitivity by employing different distortion measure of the esitimated return distribution. HuRi is capable of dynamically selecting the risk sensitivity level in varying environmental conditions by utilizing the Inter Quartile Range to measure intrinsic uncertainty and Random Network Distillation for assessing the parameter uncertainty of the environment. This algorithm allows the agent to conduct safe and efficient exploration in hazardous environments during training, enhancing the mobility of humanoid robots. Simulations and real-world deployments on the Zerith-1 robot have been conducted to confirm the robustness of HuRi.", "keywords": [ "Adaptive Risk-Aware", "Distributional Reinforcement Learning", "Humanoid Robots", "Locomotion Control" ], "rating_list": [ 6, 5, 8, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 1, 1, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "iIWeyfGTof", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Johann Brehmer", "gender": "Male", "institution": "CuspAI", "country": "NL", "position": "Researcher" }, { "name": "Pim De Haan", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "PhD student" }, { "name": "S\u00f6nke Behrends", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "", "position": "Researcher" }, { "name": "Taco Cohen", "gender": "Male", "institution": "Meta", "country": "IT", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 63, "n_ref": 107, "n_ref_all": 131, "n_fig": 6, "n_tab": 2, "L_tab": 642, "n_element_tab": 82, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 642, "n_element_tab_1": 82, "formula_len_all": 404, "formula_len_all_1": 367, "len_all": 158394, "len_all_1": 67243, "len_abs": 972, "len_title": 82, "len_sents": 39435, "len_sents_1": 30374, "n_sents": 299, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 34, "L_abs": 977, "title": "Does equivariance matter at scale?", "abs": "Given large data sets and sufficient compute, is it beneficial to design neural architectures for the structure and symmetries of each problem? Or is it more efficient to learn them from data? We study empirically how equivariant and non-equivariant networks scale with compute and training samples. Focusing on a benchmark problem of rigid-body interactions and on general-purpose transformer architectures, we perform a series of experiments, varying the model size, training steps, and dataset size. We find evidence for three conclusions. First, equivariance improves data efficiency, but training non-equivariant models with data augmentation can close this gap given sufficient epochs. Second, scaling with compute follows a power law, with equivariant models outperforming non-equivariant ones at each tested compute budget. Finally, the optimal allocation of a compute budget onto model size and training duration differs between equivariant and non-equivariant models.", "keywords": [ "Geometric deep learning", "equivariance", "neural scaling laws" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "iINUF4n33F", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Canlong Zhang", "gender": "Male", "institution": "Guangxi Normal University", "country": "", "position": "Full Professor" }, { "name": "ZengliLuo", "gender": "Female", "institution": "Guangxi Traditional Chinese Medical University", "country": "", "position": "MS student" }, { "name": "Zhiwen Wang", "gender": "Male", "institution": "Guangxi University of Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Zhixin Li", "gender": "Male", "institution": "Guangxi Normal University", "country": "CN", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 17, "n_ref": 19, "n_ref_all": 24, "n_fig": 1, "n_tab": 2, "L_tab": 106, "n_element_tab": 16, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 106, "n_element_tab_1": 16, "formula_len_all": 203, "formula_len_all_1": 203, "len_all": 55095, "len_all_1": 55095, "len_abs": 1243, "len_title": 97, "len_sents": 19885, "len_sents_1": 19885, "n_sents": 134, "n_sents_1": 134, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1249, "title": "Text-Based Person Search in Full Images via Semantic Context Disentangling and Prototype Learning", "abs": "Text-based Person Search (TBPS) in full images aims to locate a target pedestrian within uncropped images based on natural language descriptions. Existing TBPS methods typically rely on candidate region generation and cross-modal matching. However, in complex scenes,especially those with multiple pedestrians in the image.It is often challenging to distinguish the target pedestrian from the background or other individuals. This leads to limited generalization capabilities.\n To address these issues, we propose a new TBPS framework named ProtoDis-TBPS, which integrates three key components: Semantic Context Decoupling (SCD), Prototype Embedding Learning (PEL), and a Cross-modal Person Re-identification (ReID) module. Specifically, SCD enhances cross-modal feature discrimination by separating background and irrelevant contextual information. PEL improves the model's robustness in complex scenes by learning prototype features for pedestrian categories. Finally, the ReID module, based on a Transformer architecture, further boosts the accuracy of both text-based pedestrian detection and re-identification in full images.Experiments demonstrate that our proposed method presents a significant challenge to existing approaches in this field.", "keywords": [ "Cross-modal Retrieval;Text-based Person Search;Context Disentangling;Prototype Learning" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 2, 1, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 1, 2, 1, 1 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "iIGNrDwDuP", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Bo Dai", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Ceyuan Yang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Hao He", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zhengyang Liang", "gender": "Male", "institution": "University of Toronto, University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 37, "n_ref": 60, "n_ref_all": 73, "n_fig": 4, "n_tab": 3, "L_tab": 307, "n_element_tab": 36, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 680, "n_element_tab_1": 40, "formula_len_all": 406, "formula_len_all_1": 406, "len_all": 119448, "len_all_1": 62961, "len_abs": 1140, "len_title": 87, "len_sents": 31424, "len_sents_1": 31041, "n_sents": 223, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1196, "title": "Scaling Laws for Diffusion Transformers", "abs": "Diffusion transformers (DiT) have already achieved appealing synthesis and scaling properties in content recreation, \\emph{e.g.,} image and video generation. \n\nHowever, scaling laws of DiT are less explored, which usually offer precise predictions regarding optimal model size and data requirements given a specific compute budget.\n\nTherefore, experiments across a broad range of compute budgets, from \\texttt{1e17} to \\texttt{6e18} FLOPs are conducted to confirm the existence of scaling laws in DiT \\emph{for the first time}. Concretely, the loss of pretraining DiT also follows a power-law relationship with the involved compute.\n\nBased on the scaling law, we can not only determine the optimal model size and required data but also accurately predict the text-to-image generation loss given a model with 1B parameters and a compute budget of \\texttt{1e21} FLOPs.\n\nAdditionally, we also demonstrate that the trend of pretraining loss matches the generation performances (\\emph{e.g.,} FID), even across various datasets, which complements the mapping from compute to synthesis quality and thus provides a predictable benchmark that assesses model performance and data quality at a reduced cost.", "keywords": [ "Scaling Laws", "Diffusion Models", "Transformers", "Generative Models" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "iGbuc9ekKK", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Angel X Chang", "gender": "Female", "institution": "Simon Fraser University", "country": "CA", "position": "Associate Professor" }, { "name": "Han-Hung Lee", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "PhD student" }, { "name": "Yiming Zhang", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 60, "n_ref": 136, "n_ref_all": 180, "n_fig": 13, "n_tab": 10, "L_tab": 2302, "n_element_tab": 235, "n_fig_1": 6, "n_tab_1": 8, "L_tab_1": 1074, "n_element_tab_1": 129, "formula_len_all": 331, "formula_len_all_1": 296, "len_all": 200792, "len_all_1": 71386, "len_abs": 1343, "len_title": 110, "len_sents": 48053, "len_sents_1": 32324, "n_sents": 387, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1350, "title": "Duoduo CLIP: Efficient 3D Understanding with Multi-View Images", "abs": "We introduce Duoduo CLIP, a model for 3D representation learning that learns shape encodings from multi-view images instead of point-clouds. The choice of multi-view images allows us to leverage 2D priors from off-the-shelf CLIP models to facilitate fine-tuning with 3D data. Our approach not only shows better generalization compared to existing point cloud methods, but also reduces GPU requirements and training time. In addition, the model is modified with cross-view attention to leverage information across multiple frames of the object which further boosts performance. Notably, our model is permutation invariant to the order of multi-view images while being pose-free. Compared to the current SOTA point cloud method that requires 480 A100 hours to train 1 billion model parameters we only require 57 A5000 hours and 87 million parameters. Multi-view images also provide more flexibility including being able to encode objects with a variable number of images, and performance scales when more views are used. In contrast, point cloud based methods require an entire scan or model of the object. We showcase this flexibility with benchmarks from images of real-world objects. Our model also achieves better performance in more fine-grained text to shape retrieval, demonstrating better text-and-shape alignment than point cloud based models.", "keywords": [ "3D Representation", "3D Understanding" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "iGafR0hSln", "primary_area": "applications to robotics, autonomy, planning", "n_author": 8, "author_info_list": [ { "name": "Gao Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Hao Shi", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Henry Zheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Qihang Peng", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Rui Huang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yepeng Weng", "gender": "Male", "institution": "Lenovo Group Limited", "country": "CN", "position": "Researcher" }, { "name": "Yong Xien Chng", "gender": "Not Specified", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "zhongchao shi", "gender": "Male", "institution": "Lenovo Research", "country": "CN", "position": "Research Scientist" } ], "n_formula": 10, "n_formula_1": 3, "n_ref_uni": 43, "n_ref": 95, "n_ref_all": 106, "n_fig": 5, "n_tab": 5, "L_tab": 1041, "n_element_tab": 114, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1087, "n_element_tab_1": 124, "formula_len_all": 613, "formula_len_all_1": 151, "len_all": 139449, "len_all_1": 65706, "len_abs": 1575, "len_title": 141, "len_sents": 39922, "len_sents_1": 31179, "n_sents": 276, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1606, "title": "DenseGrounding: Improving Dense Language-Vision Semantics for Ego-centric 3D Visual Grounding", "abs": "Enabling intelligent agents to comprehend and interact with 3D environments through natural language is crucial for advancing robotics and human-computer interaction. A fundamental task in this field is ego-centric 3D visual grounding, where agents locate target objects in real-world 3D spaces based on verbal descriptions. However, this task faces two significant challenges: (1) loss of fine-grained visual semantics due to sparse fusion of point clouds with ego-centric multi-view images, (2) limited textual semantic context due to arbitrary language descriptions. We propose DenseGrounding, a novel approach designed to address these issues by enhancing both visual and textual semantics. For visual features, we introduce the Hierarchical Scene Semantic Enhancer, which retains dense semantics by capturing fine-grained global scene features and facilitating cross-modal alignment. For text descriptions, we propose a Language Semantic Enhancer that leverage large language models to provide rich context and diverse language descriptions with additional context during model training. Extensive experiments show that DenseGrounding significantly outperforms existing methods in overall accuracy, achieving improvements of **5.81%** and **7.56%** when trained on the comprehensive full training dataset and smaller mini subset, respectively, further advancing the SOTA in ego-centric 3D visual grounding. Our method also achieves **1st place** and receives **Innovation Award** in the 2024 Autonomous Grand Challenge Multi-view 3D Visual Grounding Track, validating its effectiveness and robustness.", "keywords": [ "3D Visual Grounding" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 5, 3 ] }, { "paper_id": "iGX0lwpUYj", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Kyle Sprague", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Researcher" }, { "name": "Leo Schwinn", "gender": "Male", "institution": "Technical University of Munich", "country": "DE", "position": "Postdoc" }, { "name": "Mark Coates", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Full Professor" }, { "name": "Thomas Markovich", "gender": "unknown", "institution": "Block", "country": "US", "position": "Principal Researcher" }, { "name": "florence regol", "gender": "Female", "institution": "McGill University", "country": "", "position": "PhD student" } ], "n_formula": 92, "n_formula_1": 28, "n_ref_uni": 51, "n_ref": 125, "n_ref_all": 150, "n_fig": 9, "n_tab": 11, "L_tab": 4682, "n_element_tab": 725, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 617, "n_element_tab_1": 92, "formula_len_all": 5446, "formula_len_all_1": 1489, "len_all": 240035, "len_all_1": 70319, "len_abs": 1862, "len_title": 88, "len_sents": 74032, "len_sents_1": 32750, "n_sents": 628, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1508, "title": "When to retrain a machine learning model", "abs": "A significant challenge in maintaining real-world machine learning models is responding to the continuous and unpredictable evolution of data. Most practitioners are faced with the difficult question: when should I retrain or update my machine learning model? This seemingly straightforward problem is particularly challenging for three reasons: 1) decisions must be made based on very limited information - we usually have access to only a few examples, 2) the nature, extent, and impact of the distribution shift are unknown, and 3) it involves specifying a cost ratio between retraining and poor performance, which can be hard to characterize. Existing works address certain aspects of this problem, but none offer a comprehensive solution. Distribution shift detection falls short as it cannot account for the cost trade-off; the scarcity of the data, paired with its unusual structure, makes it a poor fit for existing offline reinforcement learning methods, and the online learning formulation overlooks key practical considerations.\nTo address this, we present a principled formulation of the retraining problem and propose an uncertainty-based method that makes decisions by continually forecasting the evolution of model performance evaluated with a bounded metric. Our experiments, addressing classification tasks, show that the method consistently outperforms existing baselines on 7 datasets. We thoroughly assess its robustness to varying cost trade-off values and mis-specified cost trade-offs.", "keywords": [ "retraining;sequence modeling; forecasting performance" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "iGV6Sg5bI0", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Chang Xu", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Hao Li", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" }, { "name": "Jiang Bian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Viktor Schlegel", "gender": "unknown", "institution": "Imperial College London", "country": "SG", "position": "Researcher" }, { "name": "Yuhao Huang", "gender": "unknown", "institution": "Nanjing University", "country": "CN", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 40, "n_ref": 72, "n_ref_all": 92, "n_fig": 9, "n_tab": 9, "L_tab": 6177, "n_element_tab": 575, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 4513, "n_element_tab_1": 322, "formula_len_all": 117, "formula_len_all_1": 117, "len_all": 216365, "len_all_1": 77433, "len_abs": 1849, "len_title": 121, "len_sents": 65990, "len_sents_1": 34621, "n_sents": 477, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 121, "L_abs": 1884, "title": "BRIDGE: Bootstrapping Text to Guide Time-Series Generation via Multi-Agent Iterative Optimisation and Diffusion Modelling", "abs": "Time-series Generation (TSG) is an impactful research direction, as generating realistic sequences can be used to create educational materials, in simulations and for counterfactual analysis in decision making. It has further the potential to alleviate the resource bottleneck that arises from a lack of diverse time-series data required to train large time-series foundational models. However, most existing TSG models are typically designed to generate data from a specified domain, which is due to the large divergence in patterns between different real-world TS domains. In this paper, we argue that text can provide semantic information (including cross-domain background knowledge and instance temporal patterns) to improve the generalisation of TSG. To do so, we introduce ``Text Guided Time Series Generation'' (TG$^2$)---the task of generating realistic time series from handful of example time series paired with their textual description. We further present a Self-Refine-based Multi-Agent LLM framework to synthesise a realistic benchmark for TG$^2$ and show that the collected text descriptions are both realistic and useful for time-series generation. We develop a first strong baseline for the TG$^2$, Bridge, which utilises LLMs and diffusion models to generate time series which encode semantic information as cross-domain condition. Our experimental results demonstrate that Bridge significantly outperforms existing time-series generation baselines on 10 out of 12 datasets, resulting in data distributions that are more closely aligned to target domains. Using the generated data for training positively impacts the performance of time series forecasting models, effectively addressing training data limitations. This work bridges the gap between LLMs and time series analysis, introducing natural language to help the time series generation and its applications.", "keywords": [ "Time Series Generation; AI Agent" ], "rating_list": [ 8, 3, 1, 8, 6 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 4, 3, 2, 3, 1 ], "contribution_list": [ 3, 2, 2, 4, 4 ], "confidence_list": [ 4, 3, 3, 2, 3 ] }, { "paper_id": "iG7qH9Kdao", "primary_area": "generative models", "n_author": 14, "author_info_list": [ { "name": "Aditya Golatkar", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Ashwin Swaminathan", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Chengwei Su", "gender": "Male", "institution": "Dartmouth College", "country": "", "position": "PhD student" }, { "name": "Hao Li", "gender": "Male", "institution": "AWS AI Labs", "country": "", "position": "Sr. Applied Scientist" }, { "name": "Orchid Majumder", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Research Engineer" }, { "name": "R. Manmatha", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Senior Principal Scientist" }, { "name": "Shamit Lal", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Stefano Ermon", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Yang Zou", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Ying Wang", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Scientist" }, { "name": "Yusheng Xie", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Principal Applied Scientist" }, { "name": "Zhiheng Li", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Zhuowen Tu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Professor" }, { "name": "Stefano Soatto", "gender": "unknown", "institution": "Amazon Web Services", "country": "US", "position": "Vice President" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 39, "n_ref": 82, "n_ref_all": 112, "n_fig": 15, "n_tab": 1, "L_tab": 60, "n_element_tab": 9, "n_fig_1": 12, "n_tab_1": 3, "L_tab_1": 2122, "n_element_tab_1": 302, "formula_len_all": 540, "formula_len_all_1": 730, "len_all": 143789, "len_all_1": 61944, "len_abs": 2976, "len_title": 72, "len_sents": 34704, "len_sents_1": 25123, "n_sents": 267, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 798, "title": "Efficient Scaling of Diffusion Transformers for Text-to-Image Generation", "abs": "We empirically study the scaling properties of various Diffusion Transformers (DiTs) for text-to-image generation by performing extensive and rigorous ablations, including training scaled DiTs ranging from 0.3B upto 8B parameters on datasets up to 600M images. We find that U-ViT, a pure self-attention based DiT model provides a simpler design and scales more effectively in comparison with cross-attention based DiT variants, which allows straightforward expansion for extra conditions and other modalities. We identify a 2.3B U-ViT model can get better performance than SDXL UNet and other DiT variants in controlled setting. On the data scaling side, we investigate how increasing dataset size and enhanced long caption improve the text-image alignment performance and the learning efficiency.", "keywords": [ "diffusion model", "text-to-image generation", "scaling law", "u-vit", "dit" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 1, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 1, 3, 3 ], "confidence_list": [ 2, 4, 3 ] }, { "paper_id": "iFK0xoceR0", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Binghui Wang", "gender": "Male", "institution": "Illinois Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Jiate Li", "gender": "Male", "institution": "Illinois Institute of Technology", "country": "US", "position": "Intern" }, { "name": "Jinyuan Jia", "gender": "unknown", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Meng Pang", "gender": "Male", "institution": "Nanchang University", "country": "CN", "position": "Associate Professor" }, { "name": "Yun Dong", "gender": "Female", "institution": "Milwaukee School of Engineering", "country": "US", "position": "Assistant Professor" } ], "n_formula": 20, "n_formula_1": 17, "n_ref_uni": 55, "n_ref": 138, "n_ref_all": 188, "n_fig": 13, "n_tab": 11, "L_tab": 2915, "n_element_tab": 239, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 1695, "n_element_tab_1": 153, "formula_len_all": 1232, "formula_len_all_1": 709, "len_all": 182391, "len_all_1": 75532, "len_abs": 1035, "len_title": 132, "len_sents": 49512, "len_sents_1": 33059, "n_sents": 383, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1066, "title": "Provably Robust Explainable Graph Neural Networks against Graph Perturbation Attacks", "abs": "Explaining Graph Neural Network (XGNN) has gained growing attention to facilitate the trust of using GNNs, which is the mainstream method to learn graph data. Despite their growing attention, Existing XGNNs focus on improving the explanation performance, and its robustness under attacks is largely unexplored. We noticed that an adversary can slightly perturb the graph structure such that the explanation result of XGNNs is largely changed. Such vulnerability of XGNNs could cause serious issues particularly in safety/security-critical applications. In this paper, we take the first step to study the robustness of XGNN against graph perturbation attacks, and propose XGNNCert, the first provably robust XGNN. Particularly, our XGNNCert can provably ensure the explanation result for a graph under the worst-case graph perturbation attack is close to that without the attack, while not affecting the GNN prediction, when the number of perturbed edges is bounded. Evaluation results on multiple graph datasets and GNN explainers show the effectiveness of XGNNCert.", "keywords": [ "Certiffed Defenses", "Explainable Graph Neural Network", "Explainable Artificial Intelligence" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "iF06WjHnNj", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Wai Ming Tai", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "SG", "position": "Researcher" }, { "name": "Yi Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" } ], "n_formula": 235, "n_formula_1": 44, "n_ref_uni": 15, "n_ref": 63, "n_ref_all": 102, "n_fig": 2, "n_tab": 3, "L_tab": 215, "n_element_tab": 17, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 211, "n_element_tab_1": 16, "formula_len_all": 25831, "formula_len_all_1": 3470, "len_all": 247530, "len_all_1": 67190, "len_abs": 602, "len_title": 101, "len_sents": 67663, "len_sents_1": 23706, "n_sents": 837, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 697, "title": "Near-optimal Active Regression of Single-Index Models", "abs": "The active regression problem of the single-index model is to solve $\\min_x \\lVert f(Ax)-b\\rVert_p$, where $A$ is fully accessible and $b$ can only be accessed via entry queries, with the goal of minimizing the number of queries to the entries of $b$.\nWhen $f$ is Lipschitz, previous results only obtain constant-factor approximations. This work presents the first algorithm that provides a $(1+\\varepsilon)$-approximation solution by querying $\\tilde{O}(d^{\\frac{p}{2}\\vee 1}/\\varepsilon^{p\\vee 2})$ entries of $b$. This query complexity is also shown to be optimal up to logarithmic factors for $p\\in [1,2]$ and the $\\varepsilon$-dependence of $1/\\varepsilon^p$ is shown to be optimal for $p>2$.", "keywords": [ "Lewis weights", "Active regression", "Query complexity" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 2, 3, 2, 2 ] }, { "paper_id": "iEfdvDTcZg", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Alexander Atanasov", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Alexandru Meterez", "gender": "Male", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Cengiz Pehlevan", "gender": "unknown", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "James B Simon", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 10, "n_ref_uni": 45, "n_ref": 110, "n_ref_all": 170, "n_fig": 28, "n_tab": 3, "L_tab": 1837, "n_element_tab": 118, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 827, "n_element_tab_1": 80, "formula_len_all": 1693, "formula_len_all_1": 702, "len_all": 213212, "len_all_1": 67169, "len_abs": 1581, "len_title": 70, "len_sents": 66049, "len_sents_1": 29154, "n_sents": 602, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1741, "title": "The Optimization Landscape of SGD Across the Feature Learning Strength", "abs": "We consider neural networks (NNs) where the final layer is down-scaled by a fixed hyperparameter $\\gamma$. \nRecent work has identified $\\gamma$ as controlling the strength of feature learning.\nAs $\\gamma$ increases, network evolution changes from \"lazy\" kernel dynamics to \"rich\" feature-learning dynamics, with a host of associated benefits including improved performance on common tasks.\nIn this work, we conduct a thorough empirical investigation of the effect of scaling $\\gamma$ across a variety of models and datasets in the online training setting.\nWe first examine the interaction of $\\gamma$ with the learning rate $\\eta$, identifying several scaling regimes in the $\\gamma$-$\\eta$ plane which we explain theoretically using a simple model.\nWe find that the optimal learning rate $\\eta^*$ scales non-trivially with $\\gamma$. In particular, $\\eta^* \\propto \\gamma^2$ when $\\gamma \\ll 1$ and $\\eta^* \\propto \\gamma^{2/L}$ when $\\gamma \\gg 1$ for a feed-forward network of depth $L$.\nUsing this optimal learning rate scaling, we proceed with an empirical study of the under-explored ``ultra-rich'' $\\gamma \\gg 1$ regime.\nWe find that networks in this regime display characteristic loss curves, starting with a long plateau followed by a drop-off, sometimes followed by one or more additional staircase steps.\nWe find networks of different large $\\gamma$ values optimize along similar trajectories up to a reparameterization of time.\nWe further find that optimal online performance is often found at large $\\gamma$ and could be missed if this hyperparameter is not tuned.\nOur findings indicate that analytical study of the large-$\\gamma$ limit may yield useful insights into the dynamics of representation learning in performant models.", "keywords": [ "Feature Learning", "Maximal Update Parameterization", "Optimization", "Edge of Stability", "Catapult Effect", "Silent Alignment" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "iEdEHPcFeu", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Aditi Mavalankar", "gender": "Female", "institution": "Google DeepMind", "country": "GB", "position": "Research Scientist" }, { "name": "Hassan Mansoor", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Mariia Samsikova", "gender": "Female", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Zita Marinho", "gender": "unknown", "institution": "Google Deepmind", "country": "GB", "position": "Researcher" }, { "name": "Tom Schaul", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 42, "n_ref_all": 79, "n_fig": 12, "n_tab": 2, "L_tab": 879, "n_element_tab": 64, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 694, "n_element_tab_1": 20, "formula_len_all": 187, "formula_len_all_1": 304, "len_all": 163518, "len_all_1": 62238, "len_abs": 1390, "len_title": 92, "len_sents": 45917, "len_sents_1": 29695, "n_sents": 343, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1503, "title": "AuPair: Golden Example Pairs for Code Repair", "abs": "Scaling up inference-time compute has proven to be a valuable strategy in improving the performance of Large Language Models (LLMs) on several tasks without involving any fine-tuning. An example of such a task that can benefit from additional inference-time compute is self-repair: given an initial flawed response produced by the LLM, it is supposed to correct its own mistake and produce an improved response. We propose leveraging the in-context learning capability exhibited by LLMs to aid with self-repair. The key contribution of this paper is an approach to synthesise and select a golden set of pairs, each of which contains a problem, the initial guess produced by the LLM, and the consequent fix generated. Each golden example pair, or AuPair is then provided as an in-context example at inference time to generate a candidate repaired solution with 1-shot prompting; in line with best-of-$N$ the highest scoring response is selected. Given an inference-time compute budget of $N$ LLM calls, our algorithm selects $N$ AuPairs in a manner that maximises complementarity and usefulness. We demonstrate the results of our algorithm on the coding domain for code repair on 4 LLMs across 7 competitive programming datasets. The AuPairs produced by our approach provide a significant boost in performance compared to best-of-$N$, and also exhibit strong generalisation across datasets and models. Moreover, our approach shows strong performance as the inference-time compute budget $N$ is scaled up.", "keywords": [ "LLM", "Coding" ], "rating_list": [ 1, 5, 5, 6 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 2 ] }, { "paper_id": "iEUZMISIKj", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Jun Xu", "gender": "Male", "institution": "China Mobile (Suzhou) Software Technology Company", "country": "", "position": "Researcher" }, { "name": "Kaiye Zhou", "gender": "Male", "institution": "China Mobile Research Center", "country": "CN", "position": "Researcher" }, { "name": "Shucheng Wang", "gender": "Male", "institution": "China Mobile (Suzhou) Software Technology ", "country": "CN", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 6, "n_ref_uni": 34, "n_ref": 52, "n_ref_all": 73, "n_fig": 8, "n_tab": 10, "L_tab": 1716, "n_element_tab": 155, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2306, "n_element_tab_1": 70, "formula_len_all": 1375, "formula_len_all_1": 539, "len_all": 175282, "len_all_1": 53173, "len_abs": 1771, "len_title": 120, "len_sents": 45250, "len_sents_1": 22543, "n_sents": 368, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1783, "title": "SwitchLoRA: Switched Low-Rank Adaptation Can Learn Full-Rank Information", "abs": "In the training of large language models, parameter-efficient techniques such as LoRA optimize memory usage and reduce communication overhead during the fine-tuning phase. However, applying such techniques directly during the pre-training phase results in poor performance, primarily because the premature implementation of low-rank training significantly reduces model accuracy. Existing methods like ReLoRA and GaLore have attempted to address this challenge by updating the low-rank subspace. However, they still fall short of achieving the accuracy of full-rank training because they must limit the update frequency to maintain optimizer state consistency, hindering their ability to closely approximate full-rank training behavior. In this paper, we introduce SwitchLoRA, a parameter-efficient training technique that frequently and smoothly replaces the trainable parameters of LoRA adapters with alternative parameters. SwitchLoRA updates the low-rank subspace incrementally, targeting only a few dimensions at a time to minimize the impact on optimizer states. This allows a higher update frequency, thereby enhancing accuracy by enabling the updated parameters to more closely mimic full-rank behavior during the pre-training phase. Our results demonstrate that SwitchLoRA actually surpasses full-rank training, reducing perplexity from 15.23 to 15.01 on the LLaMA 1.3B model while reducing communication overhead by 54\\% on the LLaMA 1.3B model. Furthermore, after full fine-tuning the SwitchLoRA pre-trained model and the full-rank pre-trained model on the GLUE benchmark, the SwitchLoRA pre-trained model showed an average accuracy gain of about 1\\% over the full-rank pre-trained model. This demonstrates enhanced generalization and reasoning capabilities of SwitchLoRA.", "keywords": [ "pre-training", "lora", "training efficiency", "large language models" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "iEHYbGbZ4D", "primary_area": "applications to neuroscience & cognitive science", "n_author": 7, "author_info_list": [ { "name": "Dain Kwon", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Hyeyoon Lee", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Jinho Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" }, { "name": "Jongkil Park", "gender": "Male", "institution": "Korea Institute of Science and Technology", "country": "KR", "position": "Principal Researcher" }, { "name": "Kanghyun Choi", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Seongsik Park", "gender": "Male", "institution": "Korea Institute of Science and Technology", "country": "KR", "position": "Researcher" }, { "name": "SunJong Park", "gender": "Male", "institution": "Yonsei University", "country": "", "position": "Undergrad student" } ], "n_formula": 20, "n_formula_1": 11, "n_ref_uni": 67, "n_ref": 117, "n_ref_all": 155, "n_fig": 10, "n_tab": 17, "L_tab": 8725, "n_element_tab": 857, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 3021, "n_element_tab_1": 220, "formula_len_all": 1095, "formula_len_all_1": 512, "len_all": 200145, "len_all_1": 67548, "len_abs": 1158, "len_title": 76, "len_sents": 44190, "len_sents_1": 28115, "n_sents": 365, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1163, "title": "DAS-GNN: Degree-Aware Spiking Graph Neural Networks for Graph Classification", "abs": "The recent integration of spiking neurons into graph neural networks has been gaining much attraction due to its superior energy efficiency. Especially because the sparse connection among graph nodes fits the nature of the spiking neural networks, spiking graph neural networks are considered strong alternatives to vanilla graph neural networks. However, there is still a large performance gap for graph tasks between the spiking neural networks and artificial neural networks. The gaps are especially large when they are adapted to graph classification tasks, where none of the nodes in the test set graphs are connected to the training set graphs. We diagnose the problem as the existence of neurons under starvation, caused by the sparse connections among the nodes and the neurons. To alleviate the problem, we propose DAS-GNN. Based on a set of observations on spiking neurons on graph classification tasks, we devise several techniques to utilize more neurons to deliver meaningful information to the connected neurons. Experiments on diverse datasets show significant improvements compared to the baselines, demonstrating the effectiveness of the DAS-GNN.", "keywords": [ "Spiking Neural Network", "Graph Neural Network", "Graph Classification" ], "rating_list": [ 5, 6, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "iDe1mtxqK5", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Deyu Meng", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Lei Zhang", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Chair Professor" }, { "name": "Minghan Li", "gender": "Female", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Zhengqiang ZHANG", "gender": "Male", "institution": "The Hong Kong Polytechnic University, Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "ChaoDong Xiao", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 11, "n_ref_uni": 44, "n_ref": 102, "n_ref_all": 123, "n_fig": 11, "n_tab": 5, "L_tab": 3635, "n_element_tab": 652, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 3320, "n_element_tab_1": 606, "formula_len_all": 439, "formula_len_all_1": 532, "len_all": 128697, "len_all_1": 62239, "len_abs": 1599, "len_title": 131, "len_sents": 30584, "len_sents_1": 23580, "n_sents": 207, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1584, "title": "Spatial-Mamba: Effective Visual State Space Models via Structure-Aware State Fusion", "abs": "Selective state space models (SSMs), such as Mamba, highly excel at capturing long-range dependencies in 1D sequential data, while their applications to 2D vision tasks still face challenges. Current visual SSMs often convert images into 1D sequences and employ various scanning patterns to incorporate local spatial dependencies. However, these methods are limited in effectively capturing the complex image spatial structures and the increased computational cost caused by the lengthened scanning paths. To address these limitations, we propose Spatial-Mamba, a novel approach that establishes neighborhood connectivity directly in the state space. Instead of relying solely on sequential state transitions, we introduce a structure-aware state fusion equation, which leverages dilated convolutions to capture image spatial structural dependencies, significantly enhancing the flow of visual contextual information. Spatial-Mamba proceeds in three stages: initial state computation in a unidirectional scan, spatial context acquisition through structure-aware state fusion, and final state computation using the observation equation. Our theoretical analysis shows that Spatial-Mamba unifies the original Mamba and linear attention under the same matrix multiplication framework, providing a deeper understanding of our method. Experimental results demonstrate that Spatial-Mamba, even with a single scan, attains or surpasses the state-of-the-art SSM-based models in image classification, detection and segmentation. Source codes and trained models will be made publicly available.", "keywords": [ "Representation learning", "Visual state space models", "Structure-aware state fusion" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 4, 4, 4 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "iDcWYtYUwX", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Fei Wu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Jiajian Xie", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Mengze Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Postdoc" }, { "name": "Shengyu Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "ZJU100 Young Professor" }, { "name": "chengfei lv", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Software Engineer" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 13, "n_ref_uni": 26, "n_ref": 50, "n_ref_all": 81, "n_fig": 15, "n_tab": 4, "L_tab": 589, "n_element_tab": 77, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1374, "n_element_tab_1": 197, "formula_len_all": 884, "formula_len_all_1": 969, "len_all": 117718, "len_all_1": 66404, "len_abs": 1072, "len_title": 139, "len_sents": 40225, "len_sents_1": 30705, "n_sents": 254, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 998, "title": "EcoFace: Audio-Visual Emotional Co-Disentanglement Speech-Driven 3D Talking Face Generation", "abs": "Speech-driven 3D facial animation has attracted significant attention due to its wide range of applications in animation production and virtual reality. Recent research has explored speech-emotion disentanglement to enhance facial expressions rather than manually assigning emotions. However, this approach face issues such as feature confusion, emotions weakening and mean-face. To address these issues, we present EcoFace, a framework that (1) proposes a novel collaboration objective to provide a explicit signal for emotion representation learning from the speaker's expressive movements and produced sounds, constructing an audio-visual joint and coordinated emotion space that is independent of speech content. (2) constructs a universal facial motion distribution space determined by speech features and implement speaker-specific generation. Extensive experiments show that our method achieves more generalized and emotionally realistic talking face generation compared to previous methods.", "keywords": [ "3D Talking face generation", "Audio-visual emotional co-disentanglement", "Speaker-specific" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "iC4hBE9xYe", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Fengzhuo Zhang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Guojian Wang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Lin Zhao", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Xuyang Chen", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" } ], "n_formula": 198, "n_formula_1": 33, "n_ref_uni": 27, "n_ref": 94, "n_ref_all": 171, "n_fig": 3, "n_tab": 7, "L_tab": 1148, "n_element_tab": 71, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 351, "n_element_tab_1": 46, "formula_len_all": 30852, "formula_len_all_1": 2383, "len_all": 218228, "len_all_1": 67700, "len_abs": 877, "len_title": 129, "len_sents": 62124, "len_sents_1": 30668, "n_sents": 613, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 959, "title": "Provable Convergence of Single-Timescale Neural Actor-Critic in Continuous Spaces", "abs": "Actor-critic (AC) algorithms have been the powerhouse behind many successful yet challenging applications. However, the theoretical understanding of finite-time convergence in AC's most practical form remains elusive. Existing research often oversimplifies the algorithm and only considers simple finite state and action spaces. We analyze the more practical single-timescale AC on continuous state and action spaces and use deep neural network approximations for both critic and actor. \nOur analysis reveals that the iterates of the more practical framework we consider converge towards the stationary point at rate $\\widetilde{\\mathcal{O}}(T^{-1/2})+\\widetilde{\\mathcal{O}}(m^{-1/2})$, where $T$ is the total number of iterations and $m$ is the width of the deep neural network. To our knowledge, this is the first finite-time analysis of single-timescale AC in continuous state and action spaces, which further narrows the gap between theory and practice.", "keywords": [ "Single-Timescale Actor-Critic", "Continuous State-Action Space", "Deep Neural Networks" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "iBS5SmeofT", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Milad Lankarany", "gender": "unknown", "institution": "University of Toronto", "country": "", "position": "Assistant Professor" }, { "name": "Milos Radomir Popovic", "gender": "unknown", "institution": "University of Toronto", "country": "CA", "position": "Full Professor" }, { "name": "Mohammad Reza Rezaei", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Rahul Krishnan", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 33, "n_formula_1": 17, "n_ref_uni": 41, "n_ref": 68, "n_ref_all": 94, "n_fig": 7, "n_tab": 8, "L_tab": 1886, "n_element_tab": 216, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1568, "n_element_tab_1": 134, "formula_len_all": 2987, "formula_len_all_1": 1034, "len_all": 142681, "len_all_1": 66152, "len_abs": 1055, "len_title": 110, "len_sents": 40440, "len_sents_1": 29325, "n_sents": 321, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1057, "title": "Implicit Dynamical Flow Fusion (IDFF) for Generative Modeling", "abs": "Conditional Flow Matching (CFM) models can generate high-quality samples from a non-informative prior, but they can be slow, often needing hundreds of network evaluations (NFE). To address this, we propose Implicit Dynamical Flow Fusion (IDFF); IDFF learns a new vector field with an additional momentum term that enables taking longer steps during sample generation while maintaining the fidelity of the generated distribution. Consequently, IDFFs reduce the NFEs by a factor of ten (relative to CFMs) without sacrificing sample quality, enabling rapid sampling and efficient handling of image and time-series data generation tasks. We evaluate IDFF on standard benchmarks such as CIFAR-10 and CelebA for image generation, where we achieve likelihood and quality performance comparable to CFMs and diffusion-based models with fewer NFEs. IDFF also shows superior performance on time-series datasets modeling, including molecular simulation and sea surface temperature (SST) datasets, highlighting its versatility and effectiveness across different domains.", "keywords": [ "Generative modeling", "Conditional Flow Matching (CFM)", "Image generation", "Time-series generation", "Flow-based models" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "iBExhaU3Lc", "primary_area": "optimization", "n_author": 9, "author_info_list": [ { "name": "Chenwei Wu", "gender": "unknown", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Congliang Chen", "gender": "Male", "institution": "The Chinese University of Hong Kong(Shenzhen)", "country": "CN", "position": "PhD student" }, { "name": "Diederik P Kingma", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Ruoyu Sun", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Tian Ding", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Researcher" }, { "name": "Yinyu Ye", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Yushun Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Zhi-Quan Luo", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Ziniu Li", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 73, "n_ref": 124, "n_ref_all": 199, "n_fig": 25, "n_tab": 5, "L_tab": 923, "n_element_tab": 120, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 94, "n_element_tab_1": 17, "formula_len_all": 1078, "formula_len_all_1": 854, "len_all": 236345, "len_all_1": 71207, "len_abs": 1242, "len_title": 96, "len_sents": 65685, "len_sents_1": 31977, "n_sents": 662, "n_sents_1": 317, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 1192, "title": "Adam-mini: Use Fewer Learning Rates To Gain More", "abs": "We propose Adam-mini, an optimizer that achieves on-par or better performance than AdamW with $50$% less memory footprint. Adam-mini reduces memory by cutting down the learning rate resources in Adam (i.e., $1/\\sqrt{v}$). By delving into the Hessian structure of neural nets, we find Adam\u2019s $v$ might not function at its full potential as effectively as we expected. We find that $\\geq 99.9$% of these learning rates in $v$ could be harmlessly removed if we (1) carefully partition the parameters into blocks following our proposed principle on Hessian structure; (2) assign a single but good learning rate to each parameter block. We then provide one simple way to find good learning rates and propose Adam-mini. Empirically, we verify that Adam-mini performs on par or better than AdamW on various language models sized from 39M to 13B for pre-training, supervised fine-tuning, and RLHF. The reduced memory footprint of Adam-mini also alleviates communication overheads among GPUs, thereby increasing throughput. For instance, Adam-mini achieves $49.6$% higher throughput than AdamW when pre-training Llama 2-7B on $2\\times$ A800-80GB GPUs, which saves 33% wall-clock time for pre-training.", "keywords": [ "large language model", "memory", "optimizer" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "iAmR7FfMmq", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Sahand Hormoz", "gender": "unknown", "institution": "Harvard Medical School, Harvard University", "country": "US", "position": "Associate Professor" }, { "name": "Seong Ho Pahng", "gender": "unknown", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" } ], "n_formula": 57, "n_formula_1": 9, "n_ref_uni": 44, "n_ref": 109, "n_ref_all": 130, "n_fig": 7, "n_tab": 12, "L_tab": 4870, "n_element_tab": 467, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1610, "n_element_tab_1": 95, "formula_len_all": 4527, "formula_len_all_1": 490, "len_all": 276913, "len_all_1": 72236, "len_abs": 1832, "len_title": 119, "len_sents": 96723, "len_sents_1": 33797, "n_sents": 741, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1867, "title": "Improving Graph Neural Networks by Learning Continuous Edge Directions", "abs": "Graph Neural Networks (GNNs) traditionally employ a message-passing mechanism that resembles diffusion over undirected graphs, which often leads to homogenization of node features and reduced discriminative power in tasks such as node classification. Our key insight for addressing this limitation is to assign fuzzy edge directions---that can vary continuously from node $i$ pointing to node $j$ to vice versa---to the edges of a graph so that features can preferentially flow in one direction between nodes to enable long-range information transmission across the graph. We also introduce a novel complex-valued Laplacian for directed graphs with fuzzy edges where the real and imaginary parts represent information flow in opposite directions. Using this Laplacian, we propose a general framework, called Continuous Edge Direction (CoED) GNN, for learning on graphs with fuzzy edges and prove its expressivity limits using a generalization of the Weisfeiler-Leman (WL) graph isomorphism test for directed graphs with fuzzy edges. Our architecture aggregates neighbor features scaled by the learned edge directions and processes the aggregated messages from in-neighbors and out-neighbors separately alongside the self-features of the nodes. Because continuous edge directions are differentiable, we can learn both the edge directions and the GNN weights end-to-end via gradient-based optimization. CoED GNN is particularly well-suited for graph ensemble data where the graph structure remains fixed but multiple realizations of node features are available, such as in gene regulatory networks, web connectivity graphs, and power grids. We demonstrate through extensive experiments on both synthetic and real datasets that learning continuous edge directions significantly improves performance both for undirected and directed graphs compared with existing methods.", "keywords": [ "Graph Neural Networks", "Directed Graphs", "Graph Laplacian", "Continuous Edge Directions", "Graph Ensemble Data" ], "rating_list": [ 8, 5, 3, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "iAK9oHp4Zz", "primary_area": "foundation or frontier models, including LLMs", "n_author": 28, "author_info_list": [ { "name": "Chenghua Lin", "gender": "Male", "institution": "University of Manchester ", "country": "GB", "position": "Full Professor" }, { "name": "Emmanouil Benetos", "gender": "Male", "institution": "Queen Mary, University of London", "country": "GB", "position": "Associate Professor" }, { "name": "Fengze Han", "gender": "Male", "institution": "Altos Radar", "country": "CN", "position": "Researcher" }, { "name": "Ge Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Gus Xia", "gender": "Male", "institution": "New York University", "country": "", "position": "Assistant Professor" }, { "name": "Liu Jiaheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Junting Zhou", "gender": "Male", "institution": "Peking University", "country": "", "position": "Undergrad student" }, { "name": "Ka Man Lo", "gender": "unknown", "institution": "University of Macau", "country": "", "position": "MS student" }, { "name": "Lejun Min", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Ruibin Yuan", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Shangda Wu", "gender": "unknown", "institution": "Central Conservatory of Music", "country": "", "position": "PhD student" }, { "name": "Shuyue Guo", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Tianyu Zhang", "gender": "Not Specified", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, University of Montreal", "country": "CA", "position": "PhD student" }, { "name": "Tianyu Zheng", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Wei Xue", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Wenhao Huang", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Xiang Yue", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Xingwei Qu", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "PhD student" }, { "name": "Xu Tan", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Principal Researcher" }, { "name": "Xueling Liu", "gender": "Female", "institution": "Midea Group (Shanghai) Co.,Ltd.", "country": "", "position": "Researcher" }, { "name": "Yiming Liang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Yinghao Ma", "gender": "Male", "institution": "Queen Mary University of London", "country": "GB", "position": "PhD student" }, { "name": "Yizhi LI", "gender": "Male", "institution": "University of Manchester ", "country": "GB", "position": "PhD student" }, { "name": "Ziya Zhou", "gender": "Not Specified", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Ziyang Ma", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "yuelin bai", "gender": "Not Specified", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Jie Fu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xinrun Du", "gender": "Male", "institution": "01.AI", "country": "CN", "position": "Intern" } ], "n_formula": 23, "n_formula_1": 5, "n_ref_uni": 24, "n_ref": 44, "n_ref_all": 84, "n_fig": 10, "n_tab": 10, "L_tab": 3963, "n_element_tab": 449, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 1347, "n_element_tab_1": 149, "formula_len_all": 1575, "formula_len_all_1": 392, "len_all": 189225, "len_all_1": 61572, "len_abs": 4829, "len_title": 105, "len_sents": 59860, "len_sents_1": 28807, "n_sents": 420, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1160, "title": "MuPT: A Generative Symbolic Music Pretrained Transformer", "abs": "In this paper, we explore the application of Large Language Models (LLMs) to the pre-training of music. While the prevalent use of MIDI in music modeling is well-established, our findings suggest that LLMs are inherently more compatible with ABC Notation, which aligns more closely with their design and strengths, thereby enhancing the model's performance in musical composition.\nTo address the challenges associated with misaligned measures from different tracks during generation, we propose the development of a $\\underline{S}$ynchronized $\\underline{M}$ulti-$\\underline{T}$rack ABC Notation ($\\textbf{SMT-ABC Notation}$), which aims to preserve coherence across multiple musical tracks. \nOur contributions include a series of models capable of handling up to 8192 tokens, covering 90\\% of the symbolic music data in our training set. Furthermore, we explore the implications of the $\\underline{S}$ymbolic $\\underline{M}$usic $\\underline{S}$caling Law ($\\textbf{SMS Law}$) on model performance. The results indicate a promising research direction in music generation, offering extensive resources for further research through our open-source contributions.", "keywords": [ "Pretrained Transformer", "SMT-ABC Notation", "SMS Law" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 2 ] }, { "paper_id": "i9RTCC6whL", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "John Timothy Halloran", "gender": "Male", "institution": "Leidos", "country": "US", "position": "Principal Researcher" }, { "name": "Manbir S Gulati", "gender": "Male", "institution": "Leidos Inc", "country": "US", "position": "Researcher" }, { "name": "Paul F Roysdon", "gender": "unknown", "institution": "Leidos", "country": "US", "position": "Principal Researcher" } ], "n_formula": 17, "n_formula_1": 5, "n_ref_uni": 45, "n_ref": 116, "n_ref_all": 140, "n_fig": 3, "n_tab": 4, "L_tab": 2796, "n_element_tab": 568, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 470, "n_element_tab_1": 37, "formula_len_all": 1662, "formula_len_all_1": 200, "len_all": 178088, "len_all_1": 60570, "len_abs": 1589, "len_title": 101, "len_sents": 51634, "len_sents_1": 30200, "n_sents": 373, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1567, "title": "MAMBA STATE-SPACE MODELS ARE LYAPUNOV-STABLE LEARNERS", "abs": "Mamba state-space models (SSMs) were recently shown to outperform state-of-the-art (SOTA) Transformer large language models (LLMs) across various tasks. Despite subsequent widespread adaptation, little work has focused on Mamba LLMs' amenability for fine-tuning frameworks ubiquitously used for Transformer-based LLMs, e.g., mixed-precision fine-tuning (MPFT) and parameter-efficient fine-tuning (PEFT). For the former, it currently remains an open question whether Mamba's recurrent dynamics are robust to small input changes, such as those encountered during MPFT. Using dynamical systems theory (in particular, Lyapunov exponents), we answer this question in the affirmative. We empirically validate this result through several experiments, showing that Mamba SSMs are significantly more stable to changes introduced by mixed-precision than comparable Transformers, even when both MPFT and PEFT are combined. For PEFT, we show how targeting specific memory buffers in Mamba's customized CUDA kernels for low-rank adaptation regularizes SSM parameters, thus providing both parameter efficient learning and computational savings. Finally, with both MPFT and PEFT enabled, we explore the impact of instruction tuning Mamba SSMs for in-context learning (ICL) on natural language tasks. While pretrained Mamba and Mamba-2 models only achieve 38% and 82% (respectively) of the ICL improvements of comparable Transformer-based LLMs, we show that instruction tuning allows Mamba models to narrow this gap to 81% and Mamba-2 models to skyrocket over this gap to 132%.", "keywords": [ "Mamba", "Mamba SSMs", "SSMs", "LLMs", "PEFT", "LoRA", "Transformers" ], "rating_list": [ 3, 6, 5 ], "soundness_list": [ 2, 4, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "i99hYFGpWl", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Aiden Durrant", "gender": "Male", "institution": "University of Aberdeen", "country": "GB", "position": "Lecturer" }, { "name": "Georgios Leontidis", "gender": "Male", "institution": "University of Aberdeen", "country": "GB", "position": "Full Professor" }, { "name": "Miles Everett", "gender": "unknown", "institution": "University of Aberdeen", "country": "GB", "position": "PhD student" }, { "name": "Mingjun Zhong", "gender": "Male", "institution": "University of Aberdeen", "country": "", "position": "Lecturer" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 23, "n_ref": 58, "n_ref_all": 79, "n_fig": 3, "n_tab": 7, "L_tab": 1654, "n_element_tab": 240, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1939, "n_element_tab_1": 55, "formula_len_all": 684, "formula_len_all_1": 527, "len_all": 113341, "len_all_1": 58985, "len_abs": 1199, "len_title": 113, "len_sents": 39039, "len_sents_1": 27742, "n_sents": 246, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1177, "title": "Capsule Network Projectors are Equivariant and Invariant Learners", "abs": "Learning invariant representations has been the longstanding approach to self-supervised learning. However, recently progress has been made in preserving equivariant properties in representations, yet do so with highly prescribed architectures. In this work, we propose an invariant-equivariant self-supervised architecture that employs Capsule Networks (CapsNets) which have been shown to capture equivariance with respect to novel viewpoints. We demonstrate that the use of CapsNets in equivariant self-supervised architectures achieves improved downstream performance on equivariant tasks with higher efficiency and fewer network parameters. To accommodate the architectural changes of CapsNets, we introduce a new objective function based on entropy minimisation. This approach which we name CapsIE (Capsule Invariant Equivariant Network) achieves state-of-the-art performance across all invariant and equivariant downstream tasks on the 3DIEBench dataset, while outperforming supervised baselines. Our results demonstrate the ability of CapsNets to learn complex and generalised representations for large-scale, multi-task datasets compared to previous CapsNet benchmarks.", "keywords": [ "Capsule Networks", "Equivariant", "Invariant", "Self Supervised" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "i8ynYkfoRg", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Haodong Chen", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "Junbo Wang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "Associate Professor" }, { "name": "Kaiming Zhu", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Taiyu Wang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Xiaokang Zhou", "gender": "Male", "institution": "Kansai University", "country": "JP", "position": "Associate Professor" }, { "name": "Zibin Zheng", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 19, "n_ref": 30, "n_ref_all": 42, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 740, "formula_len_all_1": 703, "len_all": 97119, "len_all_1": 58671, "len_abs": 1303, "len_title": 120, "len_sents": 35013, "len_sents_1": 28034, "n_sents": 236, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1308, "title": "Model Entanglement for solving Privacy Preserving in Federated Learning", "abs": "Federated learning (FL) is widely adopted as a secure and reliable distributed machine learning system for it allows participants to retain their training data locally, transmitting only model updates, such as gradients or parameters. However, the transmission process to the server can still lead to privacy leakage, as the updated information may be exploited to launch various privacy attacks. In this work, we present a key observation that the middle layer outputs, referred to as data representations, can exhibit independence in value distribution across different types of data. This enables us to capture the intrinsic relationship between data representations and private data, and inspires us to propose a Model Entanglement(ME) strategy aimed at enhancing privacy preserving by obfuscating the data representations of private models in a fine-grained manner, while improving the balance between privacy preservation and model accuracy. We compare our approach to the baseline FedAvg and two state-of-the-art defense methods. Our method demonstrates strong defense capabilities against mainstream privacy attacks, only reducing the global model accuracy by less than 0.7\\% and training efficiency of 6.8\\% respectively on the widely used dataset, excelling in both accuracy and privacy preserving.", "keywords": [ "Federated Learning", "Privacy Preserving", "Deep Learning", "Data Representation." ], "rating_list": [ 3, 3, 5, 1 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 5, 4, 3, 2 ] }, { "paper_id": "i8vPRlsrYu", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Ali Jadbabaie", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Michael Scholkemper", "gender": "unknown", "institution": "Rheinisch Westf\u00e4lische Technische Hochschule Aachen", "country": "", "position": "PhD student" }, { "name": "Michael T Schaub", "gender": "unknown", "institution": "RWTH Aachen University", "country": "DE", "position": "Assistant Professor" }, { "name": "Xinyi Wu", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 112, "n_formula_1": 23, "n_ref_uni": 48, "n_ref": 123, "n_ref_all": 140, "n_fig": 1, "n_tab": 5, "L_tab": 3187, "n_element_tab": 192, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 11773, "formula_len_all_1": 1525, "len_all": 215192, "len_all_1": 77578, "len_abs": 1566, "len_title": 129, "len_sents": 61247, "len_sents_1": 35680, "n_sents": 584, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1574, "title": "Residual Connections and Normalization Can Provably Prevent Oversmoothing in GNNs", "abs": "Residual connections and normalization layers have become standard design choices for graph neural networks (GNNs), and were proposed as solutions to the mitigate the oversmoothing problem in GNNs. However, how exactly these methods help alleviate the oversmoothing problem from a theoretical perspective is not well understood. In this work, we provide a formal and precise characterization of (linearized) GNNs with residual connections and normalization layers. We establish that (a) for residual connections, the incorporation of the initial features at each layer can prevent the signal from becoming too smooth, and determines the subspace of possible node representations; (b) batch normalization prevents a complete collapse of the output embedding space to a one-dimensional subspace through the individual rescaling of each column of the feature matrix. This results in the convergence of node representations to the top-k eigenspace of the message-passing operator; (c) moreover, we show that the centering step of a normalization layer \u2014 which can be understood as a projection \u2014 alters the graph signal in message-passing in such a way that relevant information can become harder to extract. Building on the last theoretical insight, we introduce GraphNormv2, a novel and principled normalization layer. GraphNormv2 features a learnable centering step designed to preserve the integrity of the original graph signal. Experimental results corroborate the effectiveness of our method, demonstrating improved performance across various GNN architectures and tasks.", "keywords": [ "Graph Neural Networks", "Normalization", "Residual Connections", "Oversmoothing" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "i8dYPGdB1C", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Qixin Zhang", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yu Yang", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Zongqi Wan", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 138, "n_formula_1": 34, "n_ref_uni": 67, "n_ref": 132, "n_ref_all": 192, "n_fig": 3, "n_tab": 4, "L_tab": 3139, "n_element_tab": 219, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 173, "n_element_tab_1": 11, "formula_len_all": 17498, "formula_len_all_1": 3498, "len_all": 224448, "len_all_1": 72288, "len_abs": 2031, "len_title": 167, "len_sents": 54716, "len_sents_1": 30591, "n_sents": 484, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 118, "L_abs": 1628, "title": "Near-Optimal Online Learning for Multi-Agent Submodular Coordination: Tight Approximation and Communication Efficiency", "abs": "Coordinating multiple agents to collaboratively maximize submodular functions in unpredictable environments is a critical task with numerous applications in machine learning, robot planning and control. The existing approaches, such as the OSG algorithm, are often hindered by their poor approximation guarantees and the rigid requirement for a fully connected communication graph. To address these challenges, we firstly present a $\\textbf{MA-OSMA}$ algorithm, which employs the multi-linear extension to transfer the discrete submodular maximization problem into a continuous optimization, thereby allowing us to reduce the strict dependence on a complete graph through consensus techniques. Moreover, $\\textbf{MA-OSMA}$ leverages a novel surrogate gradient to avoid sub-optimal stationary points. To eliminate the computationally intensive projection operations in $\\textbf{MA-OSMA}$, we also introduce a projection-free $\\textbf{MA-OSEA}$ algorithm, which effectively utilizes the KL divergence by mixing a uniform distribution. Theoretically, we confirm that both algorithms achieve a regret bound of $\\widetilde{O}(\\sqrt{\\frac{C_{T}T}{1-\\beta}})$ against a\u00a0 $(\\frac{1-e^{-c}}{c})$-approximation to the best comparator in hindsight, where $C_{T}$ is the deviation of maximizer sequence, $\\beta$ is the spectral gap of the network and $c$ is the joint curvature of submodular objectives. This result significantly improves the $(\\frac{1}{1+c})$-approximation provided by the state-of-the-art OSG algorithm. Finally, we demonstrate the effectiveness of our proposed algorithms through simulation-based multi-target tracking.", "keywords": [ "Online Learning", "Submodular Maximization", "Surrogate Gradient", "Multi-Agent" ], "rating_list": [ 6, 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 4, 2, 3 ], "contribution_list": [ 3, 3, 4, 3, 3 ], "confidence_list": [ 3, 2, 3, 4, 3 ] }, { "paper_id": "i8LCUpKvAz", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Harin Lee", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Undergrad student" }, { "name": "Min-hwan Oh", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" } ], "n_formula": 230, "n_formula_1": 24, "n_ref_uni": 28, "n_ref": 107, "n_ref_all": 151, "n_fig": 6, "n_tab": 6, "L_tab": 1786, "n_element_tab": 131, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 453, "n_element_tab_1": 12, "formula_len_all": 30946, "formula_len_all_1": 3087, "len_all": 289140, "len_all_1": 79463, "len_abs": 1648, "len_title": 106, "len_sents": 83467, "len_sents_1": 32516, "n_sents": 909, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 922, "title": "Minimax Optimal Reinforcement Learning with Quasi-Optimism", "abs": "In our quest for a reinforcement learning (RL) algorithm that is both practical and provably optimal, we introduce EQO (Exploration via Quasi-Optimism). Unlike existing minimax optimal approaches, EQO avoids reliance on empirical variances and employs a simple bonus term proportional to the inverse of the state-action visit count. Central to EQO is the concept of *quasi-optimism*, where estimated values need not be fully optimistic, allowing for a simpler yet effective exploration strategy. The algorithm achieves the sharpest known regret bound for tabular RL under the mildest assumptions, proving that fast convergence can be attained with a practical and computationally efficient approach. Empirical evaluations demonstrate that EQO consistently outperforms existing algorithms in both regret performance and computational efficiency, providing the best of both theoretical soundness and practical effectiveness.", "keywords": [ "Reinforcement Learning", "Tabular Reinforcement Learning", "Regret Analysis" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 2, 3, 4, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "i8IwcQBi74", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Freddy Lecue", "gender": "unknown", "institution": "INRIA", "country": "", "position": "Full Professor" }, { "name": "Junqi Jiang", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Manuela Veloso", "gender": "Female", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Saumitra Mishra", "gender": "Male", "institution": "J.P. Morgan Chase", "country": "", "position": "Researcher" }, { "name": "Tom Bewley", "gender": "unknown", "institution": "J.P. Morgan Chase", "country": "GB", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 43, "n_ref": 74, "n_ref_all": 110, "n_fig": 11, "n_tab": 8, "L_tab": 2418, "n_element_tab": 237, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1938, "n_element_tab_1": 204, "formula_len_all": 440, "formula_len_all_1": 474, "len_all": 159831, "len_all_1": 73117, "len_abs": 1408, "len_title": 112, "len_sents": 51394, "len_sents_1": 33666, "n_sents": 370, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1415, "title": "Interpreting Language Reward Models via Contrastive Explanations", "abs": "Reward models (RMs) are a crucial component in the alignment of large language models\u2019 (LLMs) outputs with human values. RMs approximate human preferences over possible LLM responses to the same prompt by predicting and comparing reward scores. However, as they are typically modified versions of LLMs with scalar output heads, RMs are large black boxes whose predictions are not explainable. More transparent RMs would enable improved trust in the alignment of LLMs. In this work, we propose to use contrastive explanations to explain any binary response comparison made by an RM. Specifically, we generate a diverse set of new comparisons similar to the original one to characterise the RM\u2019s local behaviour. The perturbed responses forming the new comparisons are generated to explicitly modify manually specified high-level evaluation attributes, on which analyses of RM behaviour are grounded. In quantitative experiments, we validate the effectiveness of our method for finding high-quality contrastive explanations. We then showcase the qualitative usefulness of our method for investigating global sensitivity of RMs to each evaluation attribute, and demonstrate how representative examples can be automatically extracted to explain and compare behaviours of different RMs. We see our method as a flexible framework for RM explanation, providing a basis for more interpretable and trustworthy LLM alignment.", "keywords": [ "Explainable AI", "Contrastive Explanations", "Counterfactual Explanations", "Alignment", "Large Language Model" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "i8BaiywFYx", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Alexandre Lemire Paquin", "gender": "unknown", "institution": "Laval university", "country": "", "position": "PhD student" }, { "name": "Brahim Chaib-draa", "gender": "Male", "institution": "Universite Laval, Laval university", "country": "", "position": "Full Professor" }, { "name": "Philippe Gigu\u00e8re", "gender": "Male", "institution": "Universit\u00e9 Laval", "country": "", "position": "Full Professor" } ], "n_formula": 64, "n_formula_1": 18, "n_ref_uni": 26, "n_ref": 58, "n_ref_all": 70, "n_fig": 1, "n_tab": 8, "L_tab": 4268, "n_element_tab": 514, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 3595, "n_element_tab_1": 292, "formula_len_all": 4955, "formula_len_all_1": 959, "len_all": 169920, "len_all_1": 67445, "len_abs": 1489, "len_title": 103, "len_sents": 50490, "len_sents_1": 25938, "n_sents": 449, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1487, "title": "Symmetrization of Loss Functions for Robust Training of Neural Networks in the Presence of Noisy Labels", "abs": "Labeling a training set is not only often expensive but also susceptible to errors. Consequently, the development of robust loss functions to handle label noise has emerged as a problem of great importance. The symmetry condition provides theoretical guarantees for robustness to such noise. In this work, we investigate a symmetrization method that arises from the unique decomposition of any multi-class loss function into a sum of a symmetric loss function and a class-insensitive term. Notably, the special case of symmetrizing the cross-entropy loss leads to a multi-class extension of the unhinged loss function. This loss function is linear, but unlike in the binary case, it must have specific coefficients in order to satisfy the symmetry condition. Under appropriate assumptions, we demonstrate that this multi-class unhinged loss function is the unique convex multi-class symmetric loss function. It holds a significant role among multi-class symmetric loss functions since the linear approximation of any symmetric loss function around points with equal components must be equivalent to the multi-class unhinged. Furthermore, we introduce SGCE and \u03b1-MAE, two novel loss functions that smoothly transition between the multi-class unhinged loss and the Mean Absolute Error (MAE). Our experiments demonstrate superior performance over previous state-of-the-art robust loss functions on standard benchmarks, highlighting the effectiveness of our approach in handling label noise.", "keywords": [ "Noisy labels", "Symmetric loss functions", "Multi-class loss decomposition", "Unhinged loss function" ], "rating_list": [ 5, 5, 3, 5, 6 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "i880EAXJ2x", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Andrew G. Howard", "gender": "unknown", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Chengxi Ye", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Grace Chu", "gender": "unknown", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Lukasz Lew", "gender": "Male", "institution": "Research, Google", "country": "", "position": "Researcher" }, { "name": "Yanfeng Liu", "gender": "Male", "institution": "Google", "country": "US", "position": "Software Engineer" }, { "name": "Yichi Zhang", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 13, "n_ref_uni": 36, "n_ref": 79, "n_ref_all": 118, "n_fig": 10, "n_tab": 10, "L_tab": 1389, "n_element_tab": 220, "n_fig_1": 9, "n_tab_1": 8, "L_tab_1": 1289, "n_element_tab_1": 195, "formula_len_all": 1281, "formula_len_all_1": 597, "len_all": 139576, "len_all_1": 70750, "len_abs": 1032, "len_title": 119, "len_sents": 43131, "len_sents_1": 31477, "n_sents": 360, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1036, "title": "Robust Training of Neural Networks at Arbitrary Precision and Sparsity", "abs": "The discontinuous operations inherent in quantization and sparsification introduce obstacles to backpropagation. This is particularly challenging when training deep neural networks in ultra-low precision and sparse regimes. We propose a novel, robust, and universal solution: a denoising affine transform that stabilizes training under these challenging conditions. By formulating quantization and sparsification as perturbations during training, we derive a perturbation-resilient approach based on ridge regression. Our solution employs a piecewise constant backbone model to ensure a performance lower bound and features an inherent noise reduction mechanism to mitigate perturbation-induced corruption. This formulation allows existing models to be trained at arbitrarily low precision and sparsity levels with off-the-shelf recipes. Furthermore, our method provides a novel perspective on training temporal binary neural networks, contributing to ongoing efforts to narrow the gap between artificial and biological neural networks.", "keywords": [ "quantization", "sparse", "quantized matrix multiplication" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "i7yL7VJx4H", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Arif Kerem Dayi", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Undergrad student" }, { "name": "Sitan Chen", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 197, "n_formula_1": 25, "n_ref_uni": 16, "n_ref": 34, "n_ref_all": 51, "n_fig": 2, "n_tab": 3, "L_tab": 447, "n_element_tab": 20, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 22927, "formula_len_all_1": 1789, "len_all": 245981, "len_all_1": 72011, "len_abs": 1469, "len_title": 56, "len_sents": 79361, "len_sents_1": 33538, "n_sents": 857, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1532, "title": "Gradient dynamics of low-rank fine-tuning beyond kernels", "abs": "LoRA has emerged as one of the \\emph{de facto} methods for fine-tuning foundation models with low computational cost and memory footprint. The idea is to only train a low-rank perturbation to the weights of a pre-trained model, given supervised data for a downstream task. Despite its empirical sucess, from a mathematical perspective it remains poorly understood what learning mechanisms ensure that gradient descent converges to useful low-rank perturbations.\n In this work we initiate the study of low-rank fine-tuning in a student-teacher setting. We are given the weights of a two-layer \\emph{base model} $f$, as well as i.i.d. samples $(x,f^*(x))$ where $x$ is Gaussian and $f^*$ is the \\emph{teacher model} given by perturbing the weights of $f$ by a rank-1 matrix. This generalizes the setting of \\emph{generalized linear model (GLM) regression} where the weights of $f$ are zero.\n When the rank-1 perturbation is comparable in norm to the weight matrix of $f$, the training dynamics are nonlinear. Nevertheless, in this regime we prove under mild assumptions that a student model which is initialized at the base model and trained with online gradient descent will converge to the teacher in $dk^{O(1)}$ iterations, where $k$ is the number of neurons in $f$. Importantly, unlike in the GLM setting, the complexity does not depend on fine-grained properties of the activation's Hermite expansion. We also prove that in our setting, learning the teacher model ``from scratch'' can require significantly more iterations.", "keywords": [ "learning theory", "fine tuning", "online sgd dynamics", "neural networks" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "i7k2sXSW1b", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Xiaoyi Zhu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Zengfeng Huang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 32, "n_formula_1": 27, "n_ref_uni": 33, "n_ref": 40, "n_ref_all": 47, "n_fig": 4, "n_tab": 1, "L_tab": 268, "n_element_tab": 6, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 396, "n_element_tab_1": 7, "formula_len_all": 3120, "formula_len_all_1": 2762, "len_all": 87908, "len_all_1": 53916, "len_abs": 887, "len_title": 82, "len_sents": 23283, "len_sents_1": 20908, "n_sents": 231, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 34, "L_abs": 955, "title": "Lipschitz Bandits in Optimal Space", "abs": "This paper considers the Lipschitz bandit problem, where the set of arms is continuous and the expected reward is a Lipschitz function over the arm space. This problem has been extensively studied. Prior algorithms need to store the reward information of all visited arms, leading to significant memory consumption. We address this issue by introducing an algorithm named Log-space Lipschitz bandits (Log-Li), which achieves an optimal (up to logarithmic factors) regret of $\\widetilde{O}\\left(T^{\\frac{d_z+1}{d_z+2}}\\right)$ while only uses $O\\left(\\log T\\right)$ bits of memory. Additionally, we provide a complexity analysis for this problem, demonstrating that $\\Omega\\left(\\log T\\right)$ bits of space are necessary for any algorithm to achieve the optimal regret. We also conduct numerical simulations, and the results show that our new algorithm achieves regret comparable to the state-of-the-art while reducing memory usage by orders of magnitude.", "keywords": [ "Space complexity", "Lipschitz bandits" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "i7jAYFYDcM", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Hanwei Guo", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Long Qian", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Sizhe Wang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xuguang Lan", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Yuhang Wang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 33, "n_ref": 75, "n_ref_all": 95, "n_fig": 8, "n_tab": 3, "L_tab": 916, "n_element_tab": 76, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 223, "n_element_tab_1": 20, "formula_len_all": 932, "formula_len_all_1": 1052, "len_all": 118427, "len_all_1": 59121, "len_abs": 2468, "len_title": 85, "len_sents": 37540, "len_sents_1": 28307, "n_sents": 305, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 37, "L_abs": 1419, "title": "Bootstrapped Model Predictive Control", "abs": "Model Predictive Control (MPC) has been demonstrated to be effective in continuous control tasks. When a world model and a value function are available, planning a sequence of actions ahead of time leads to a better policy. Existing methods typically obtain the value function and the corresponding policy in a model-free manner. However, we find that such an approach struggles with complex tasks, resulting in poor policy learning and inaccurate value estimation. To address this problem, we leverage the strengths of MPC itself. In this work, we introduce Bootstrapped Model Predictive Control (BMPC), a novel algorithm that performs policy learning in a bootstrapped manner. BMPC learns a network policy by imitating an MPC expert, and in turn, uses this policy to guide the MPC process. Combined with model-based TD-learning, our policy learning yields better value estimation and further boosts the efficiency of MPC. We also introduce a lazy reanalyze mechanism, which enables computationally efficient imitation learning. Our method achieves superior performance over prior works on diverse continuous control tasks. In particular, on challenging high-dimensional locomotion tasks, BMPC significantly improves data efficiency while also enhancing asymptotic performance and training stability, with comparable training time and smaller network sizes. Code is available at https://github.com/bmpc-anonymous/bmpc.", "keywords": [ "reinforcement learning", "model-based reinforcement learning", "model predictive control", "expert iteration", "continous control" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 4, 3 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "i7hXOqzUcK", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "David Fouhey", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Jianing Yang", "gender": "Male", "institution": "Meta", "country": "US", "position": "Intern" }, { "name": "Madhavan Iyengar", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Undergrad student" }, { "name": "Nikhil Madaan", "gender": "unknown", "institution": "Bloomberg", "country": "US", "position": "Researcher" }, { "name": "Shengyi Qian", "gender": "Male", "institution": "Meta FAIR", "country": "US", "position": "Research Scientist" }, { "name": "Xuweiyi Chen", "gender": "Male", "institution": ", University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Joyce Chai", "gender": "Female", "institution": "University of Michigan", "country": "US", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 46, "n_ref": 104, "n_ref_all": 132, "n_fig": 8, "n_tab": 13, "L_tab": 2465, "n_element_tab": 292, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 422, "n_element_tab_1": 14, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 210351, "len_all_1": 64248, "len_abs": 1963, "len_title": 139, "len_sents": 47824, "len_sents_1": 30240, "n_sents": 393, "n_sents_1": 249, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1494, "title": "3D-GRAND: A Million-Scale Dataset for 3D-LLMs with Better Grounding and Less Hallucination", "abs": "The integration of language and 3D perception is crucial for developing embodied agents and robots that comprehend and interact with the physical world. While large language models (LLMs) have demonstrated impressive language understanding and generation capabilities, their adaptation to 3D environments (3D-LLMs) remains in its early stages. A primary challenge is the absence of large-scale datasets that provide dense grounding between language and 3D scenes. In this paper, we introduce 3D-GRAND, a pioneering large-scale dataset comprising 40,087 household scenes paired with 6.2 million densely-grounded scene-language instructions. Our results show that instruction tuning with 3D-GRAND significantly enhances grounding capabilities and reduces hallucinations in 3D-LLMs. As part of our contributions, we propose a comprehensive benchmark 3D-POPE to systematically evaluate hallucination in 3D-LLMs, enabling fair comparisons among future models. Our experiments highlight a scaling effect between dataset size and 3D-LLM performance, emphasizing the critical role of large-scale 3D-text datasets in advancing embodied AI research. Notably, our results demonstrate early signals for effective sim-to-real transfer, indicating that models trained on large synthetic data can perform well on real-world 3D scans. Through 3D-GRAND and 3D-POPE, we aim to equip the embodied AI community with essential resources and insights, setting the stage for more reliable and better-grounded 3D-LLMs.", "keywords": [ "3D", "LLM", "Hallucination", "Embodied AI", "Multimodal", "NLP" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "i7drKWhFCo", "primary_area": "applications to neuroscience & cognitive science", "n_author": 9, "author_info_list": [ { "name": "Alex H Williams", "gender": "Male", "institution": "New York University", "country": "", "position": "Assistant Professor" }, { "name": "Anshuman Sabath", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Hari Koneru", "gender": "Male", "institution": "Duke University", "country": "", "position": "Undergrad student" }, { "name": "James Michael Roach", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "James Russell Ravenel", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Joshua Huang Wu", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "MICHAEL R TADROSS", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Shaun Sze-Xian Lim", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Timothy W DUNN", "gender": "unknown", "institution": "Duke University", "country": "", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 10, "n_ref_uni": 62, "n_ref": 111, "n_ref_all": 150, "n_fig": 22, "n_tab": 4, "L_tab": 872, "n_element_tab": 81, "n_fig_1": 18, "n_tab_1": 1, "L_tab_1": 193, "n_element_tab_1": 22, "formula_len_all": 963, "formula_len_all_1": 528, "len_all": 191459, "len_all_1": 69729, "len_abs": 1082, "len_title": 128, "len_sents": 54885, "len_sents_1": 34811, "n_sents": 379, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 80, "L_abs": 1086, "title": "Disentangling 3D Animal Pose Dynamics with Scrubbed Conditional Latent Variables", "abs": "Methods for tracking lab animal movements in unconstrained environments have become increasingly common and powerful tools for neuroscience. The prevailing hypothesis is that animal behavior in these environments comprises sequences of discrete stereotyped body movements (\"motifs\" or \"actions\"). However, the same action can occur at different speeds or heading directions, and the same action may manifest slightly differently across subjects due to, for example, variation in body size. These and other forms of nuisance variability complicate attempts to quantify animal behavior in terms of discrete action sequences and draw meaningful comparisons across individual subjects. To address this, we present a framework for motion analysis that uses conditional variational autoencoders in conjunction with adversarial learning paradigms to disentangle behavioral factors. We demonstrate the utility of this approach in downstream tasks such as clustering, decodability, and motion synthesis. Further, we apply our technique to improve disease detection in a Parkinsonian mouse model.", "keywords": [ "behavioral neuroscience", "systems neuroscience" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "i6jYK0hd0B", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Chanyoung Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Gyoung S. Na", "gender": "Male", "institution": "Korea Research Institute of Chemical Technology", "country": "KR", "position": "Researcher" }, { "name": "Heewoong Noh", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Namkyeong Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Tianfan Fu", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "Assistant Professor" }, { "name": "Yunhak Oh", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 40, "n_ref": 87, "n_ref_all": 117, "n_fig": 5, "n_tab": 6, "L_tab": 2536, "n_element_tab": 222, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 465, "n_element_tab_1": 37, "formula_len_all": 1114, "formula_len_all_1": 1124, "len_all": 174387, "len_all_1": 71466, "len_abs": 1311, "len_title": 119, "len_sents": 56144, "len_sents_1": 32146, "n_sents": 356, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1319, "title": "3D Interaction Geometric Pre-training for Molecular Relational Learning", "abs": "Molecular Relational Learning (MRL) is a rapidly growing field that focuses on understanding the interaction dynamics between molecules, which is crucial for applications ranging from catalyst engineering to drug discovery. \nDespite recent progress, earlier MRL approaches are limited to using only the 2D topological structure of molecules, as obtaining the 3D interaction geometry remains prohibitively expensive.\nThis paper introduces a novel 3D geometric pre-training strategy for MRL (3DMRL) that incorporates a 3D virtual interaction environment, overcoming the limitations of costly traditional quantum mechanical calculation methods. \nWith the constructed 3D virtual interaction environment, 3DMRL trains 2D MRL model to learn the overall 3D geometric information of molecular interaction through contrastive learning.\nMoreover, fine-grained interaction between molecules is learned through force prediction loss, which is crucial in understanding the wide range of molecular interaction processes.\nExtensive experiments on various tasks using real-world datasets, including out-of-distribution and extrapolation scenarios, demonstrate the effectiveness of 3DMRL, showing up to a 24.93 % improvement in performance across 40 tasks.\nOur code is publicly available at https://anonymous.4open.science/r/3DMRL-F973.", "keywords": [ "Molecular Relational Learning", "AI4Science", "Geometric Deep Learning" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "i6b2TrTNMz", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Hongyang Chen", "gender": "Male", "institution": "Zhejiang Lab, China", "country": "CN", "position": "Senior Research Expert" }, { "name": "Jun Sun", "gender": "Male", "institution": "Zhejiang Lab, Zhejiang Lab", "country": "CN", "position": "Researcher" }, { "name": "Simin Hong", "gender": "Female", "institution": "University of Leeds", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 42, "n_ref": 64, "n_ref_all": 76, "n_fig": 5, "n_tab": 7, "L_tab": 3254, "n_element_tab": 277, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 721, "n_element_tab_1": 75, "formula_len_all": 1697, "formula_len_all_1": 1521, "len_all": 133695, "len_all_1": 60060, "len_abs": 1786, "len_title": 140, "len_sents": 37306, "len_sents_1": 26678, "n_sents": 274, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1759, "title": "A Third-Person Appraisal Agent: Learning to Reason About Emotions in Conversational Contexts", "abs": "Emotion reasoning is crucial for achieving human-like emotional understanding in Emotion Recognition in Conversation (ERC). Current ERC datasets provide only emotion-labeled utterances, lacking the rich annotations necessary for emotion reasoning. Although Large Language Models (LLMs) show promise in generating rich emotional knowledge, they still struggle to apply this knowledge effectively for emotion reasoning. To address these challenges, we propose a learning framework based on cognitive appraisal theory, utilizing an agent powered by LLMs to learn emotion reasoning from a third-person perspective, which we refer to as the third-person appraisal agent. This learning framework comprises two phases: self-evaluation and meta-evaluation. In the self-evaluation phase, the agent generates appraisals essential for inferring emotions, incorporating counterfactual thinking to refine its appraisals. The meta-evaluation phase uses reflective actor-critic reinforcement learning to train the agent to generate accurate appraisals during testing. The training samples are appraisals generated during the self-evaluation phase, which eliminates the need for human annotations. By fine-tuning a specialized LLM in this framework, our approach significantly outperforms LLM baselines across ERC tasks, demonstrating improved reasoning and generalization across various dialogue datasets. Additionally, we provide interpretable results that clarify the model\u2019s reasoning process behind its predictions. To the best of our knowledge, this research is the first to apply cognition-based methods to enhance LLMs' emotional reasoning capabilities, marking a significant advancement toward achieving human-like emotional understanding in artificial intelligence.", "keywords": [ "emotion recognition in conversation", "LLM agent", "emotion reasoning", "reinforcement learning" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "i5MrJ6g5G1", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Alexander M Rush", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Associate Professor" }, { "name": "Bernardo P de Almeida", "gender": "unknown", "institution": "InstaDeep", "country": "", "position": "Researcher" }, { "name": "Guanghan Wang", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Hao Phung", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Hugo Dalla-torre", "gender": "Male", "institution": "InstaDeep", "country": "FR", "position": "Researcher" }, { "name": "Sam Boshar", "gender": "Male", "institution": "InstaDeep", "country": "US", "position": "Researcher" }, { "name": "Subham Sekhar Sahoo", "gender": "Male", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "PhD student" }, { "name": "Thomas PIERROT", "gender": "Male", "institution": "InstaDeep", "country": "US", "position": "Principal Researcher" }, { "name": "Volodymyr Kuleshov", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Assistant Professor" }, { "name": "Yair Schiff", "gender": "Male", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "PhD student" } ], "n_formula": 85, "n_formula_1": 35, "n_ref_uni": 50, "n_ref": 104, "n_ref_all": 174, "n_fig": 3, "n_tab": 22, "L_tab": 9310, "n_element_tab": 1314, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1827, "n_element_tab_1": 285, "formula_len_all": 9815, "formula_len_all_1": 2931, "len_all": 229364, "len_all_1": 67643, "len_abs": 1049, "len_title": 114, "len_sents": 54666, "len_sents_1": 25900, "n_sents": 480, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1052, "title": "Simple and Controllable Uniform Discrete Diffusion Language Models", "abs": "Diffusion models for continuous data gained widespread adoption owing to their high quality generation and control mechanisms. However, controllable diffusion on discrete data faces challenges: continuous diffusion guidance methods are not applicable and recent discrete diffusion models are not well-suited to control or exhibit a quality gap. Here, we provide a straightforward derivation of classifier-free and classifier-based guidance for discrete diffusion, as well as a new class of diffusion models that leverage uniform noise and thus can continuously edit their outputs. We improve the quality of these models with a novel continuous-time variational lower bound that yields state-of-the-art performance, in settings with small vocabularies. Empirically, we demonstrate the effectiveness of our guidance mechanisms relative to autoregressive and diffusion baselines, especially in conjunction with uniform noise diffusion, on several discrete data domains, including genomic sequences, small molecule design, and discretized image generation.", "keywords": [ "Discrete Diffusion", "Guidance" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "i5Gxilzk0u", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Changxing Ding", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Hongwei Yi", "gender": "Male", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Huaiguang Jiang", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Qiong Cao", "gender": "Female", "institution": "JD Explore Academy", "country": "CN", "position": "Research Scientist" }, { "name": "Yifei Liu", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 14, "n_ref_uni": 32, "n_ref": 79, "n_ref_all": 102, "n_fig": 4, "n_tab": 5, "L_tab": 1164, "n_element_tab": 185, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 831, "n_element_tab_1": 145, "formula_len_all": 1683, "formula_len_all_1": 846, "len_all": 158934, "len_all_1": 68312, "len_abs": 1730, "len_title": 96, "len_sents": 50979, "len_sents_1": 32041, "n_sents": 341, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1742, "title": "Multi-modal Controlled Coherent Motion Synthesis", "abs": "We walk and talk at the same time all the time. It is just natural for us. This paper tackles the challenge of replicating such natural behaviors in 3D avatar motion generation driven by concurrent multi-modal inputs, e.g., a text description ``a man is walking\" alongside a speech audio. Existing methods, constrained by the scarcity of aligned multi-modal data, typically combine motions from individual modalities sequentially or through weighted averaging. These strategies often result in mismatched or unrealistic movements. To overcome these limitations, we propose MOCO, a novel diffusion-based framework capable of processing multiple simultaneous inputs\u2014including speech audio, text descriptions, and trajectory data\u2014to generate coherent and lifelike motions without requiring additional datasets. Our key innovation lies in decoupling the motion generation process. During each denoising step, the diffusion model independently generates motions for each modality from the input noise and assembles the body parts according to predefined spatial rules. The resulting combined motion is then diffused and serves as the input noise for the subsequent denoising step. This iterative approach enables each modality to refine its contribution within the context of the overall motion, progressively harmonizing movements across modalities. Consequently, the generated motions become increasingly natural and fluid with each iteration, achieving coherent and synchronized behaviors. We evaluate our approach using a purpose-built multi-modal benchmark. Experimental results demonstrate that MOCO significantly outperforms existing baselines, advancing the field of multi-modal motion generation for 3D avatars. The code will be released.", "keywords": [ "Human Motion Generation; Multi-Modal; Generative Models" ], "rating_list": [ 8, 6, 5, 5, 5 ], "soundness_list": [ 4, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 4, 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "i55KTmMb4I", "primary_area": "datasets and benchmarks", "n_author": 16, "author_info_list": [ { "name": "Bangyu Xiang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "alibaba" }, { "name": "Bo Zheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Congnan Liu", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Ge Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Guoan Zhang", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "He Zhu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Liu Jiaheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jian Yang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "JinKe", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Ken Deng", "gender": "Female", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Linzheng Chai", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Peng Zhao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Shukai Liu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Wenbo Su", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yanan Wu", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Zekun Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 33, "n_ref": 77, "n_ref_all": 130, "n_fig": 21, "n_tab": 18, "L_tab": 10410, "n_element_tab": 1460, "n_fig_1": 11, "n_tab_1": 9, "L_tab_1": 2724, "n_element_tab_1": 325, "formula_len_all": 179, "formula_len_all_1": 82, "len_all": 189635, "len_all_1": 63092, "len_abs": 1259, "len_title": 128, "len_sents": 38225, "len_sents_1": 24111, "n_sents": 275, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1255, "title": "M2rc-Eval: Massively Multilingual Repository-level Code Completion Evaluation", "abs": "Repository-level code completion has drawn great attention in software engineering, and several benchmark datasets have been introduced. However, existing repository-level code completion benchmarks usually focus on a limited number of languages (<5), which cannot evaluate the general code intelligence abilities across different languages for existing code Large Language Models (LLMs). Besides, the existing benchmarks usually report overall average scores of different languages, where the fine-grained abilities in different completion scenarios are ignored. Therefore, to facilitate the research of code LLMs in multilingual scenarios, we propose a massively multilingual repository-level code completion benchmark covering 18 programming languages (called M2RC-EVAL), and two types of fine-grained annotations (i.e., bucket-level and semantic-level) on different completion scenarios are provided, where we obtain these annotations based on the parsed abstract syntax tree. Moreover, we also curate a massively multilingual instruction corpora M2RC- INSTRUCT dataset to improve the repository-level code completion abilities of existing code LLMs. Comprehensive experimental results demonstrate the effectiveness of our M2RC-EVAL and M2RC-INSTRUCT.", "keywords": [ "Large Language Models", "Code Completion", "Benchmark" ], "rating_list": [ 8, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "i4pGIOlH8l", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Bo Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Brian Nlong Zhao", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Research Assistant" }, { "name": "Dongsheng Li", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Javier Alvarez-Valle", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Senior Director of Biomedical Imaging" }, { "name": "Lili Qiu", "gender": "Female", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Matthew P. Lungren", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Principal Researcher" }, { "name": "XINYANG JIANG", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Researcher" }, { "name": "Xufang Luo", "gender": "Female", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" }, { "name": "Yifan Yang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Zilong Wang", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 40, "n_ref": 71, "n_ref_all": 79, "n_fig": 3, "n_tab": 6, "L_tab": 4409, "n_element_tab": 276, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 3311, "n_element_tab_1": 254, "formula_len_all": 310, "formula_len_all_1": 310, "len_all": 215315, "len_all_1": 71519, "len_abs": 1510, "len_title": 128, "len_sents": 69143, "len_sents_1": 32562, "n_sents": 589, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1515, "title": "Benchmark Dataset for Radiology Report Generation with Instructions and Contexts", "abs": "While automatic report generation has demonstrated promising results using deep learning-based methods, deploying these algorithms in real-world scenarios remains challenging, where models may be required to follow the instruction from the radiologists and consider contextual information. Such instructional report generation tasks are critical for enabling more accurate, customizable, and scalable report generation processes, but remain under-explored and lack substantial datasets for training and evaluation. However, constructing a dataset for report generation with instructions and contexts is challenging due to the scarcity of medical data, privacy concerns and the absence of recorded user-model interactions. To tackle this challenge, we propose a unified and automatic data generation pipeline which leverages large language model (LLM) to produce high-quality instructions and context for report generation tasks. We present a new benchmark dataset MIMIC-R3G that extends the largest existing radiology report generation dataset MIMIC-CXR, comprising five representative tasks pertinent to real-world medical report generation. We conducted an extensive evaluation of state-of-the-art methods using the proposed benchmark datasets. Additionally, we introduced a baseline method, the Domain-enhanced Multimodal Model (DeMMo), demonstrating that leveraging training data containing instructions and contextual information significantly improves the performance of instructional report generation tasks.", "keywords": [ "Report Generation", "Dataset and Benchmark", "Multimodal Learning", "Multimodal Model" ], "rating_list": [ 5, 3, 6, 5, 3 ], "soundness_list": [ 2, 2, 4, 2, 2 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "i4ouG6Kc8M", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Chad Vanderbilt", "gender": "Male", "institution": "Memorial Sloan Kettering Cancer Center", "country": "US", "position": "Assistant Professor" }, { "name": "Neeraj Kumar", "gender": "Male", "institution": "Memorial Sloan Kettering Cancer Centre", "country": "US", "position": "Researcher" }, { "name": "Swaraj Nanda", "gender": "Male", "institution": "Memorial Sloan Kettering Cancer Centre", "country": "US", "position": "Postdoc" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 41, "n_ref": 74, "n_ref_all": 100, "n_fig": 7, "n_tab": 4, "L_tab": 2209, "n_element_tab": 299, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1306, "n_element_tab_1": 283, "formula_len_all": 864, "formula_len_all_1": 55, "len_all": 144045, "len_all_1": 61880, "len_abs": 1511, "len_title": 138, "len_sents": 40408, "len_sents_1": 28415, "n_sents": 240, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1518, "title": "A Dual-Metric Approach for Model Selection in self-supervised learning for histopathology", "abs": "Selecting appropriate models during self-supervised training of vision transformers in histopathology is challenging. Recent efforts to quantify the quality of self-supervised learning representations through rank estimation approaches have shown promise in natural image classification tasks. However, their effectiveness in histopathology, particularly for non-linear tasks such as instance segmentation and classification from whole slide images, remains unexplored. This study proposes an approach for model selection in histopathology by combining task-specific metrics (such as accuracy) and task agnostic metrics (such as rank estimation). This work shows that by training several small-scale histopathology models and applying the proposed model selection approach, one can achieve instance segmentation performance comparable to state-of-the-art models trained on much larger datasets. The proposed approach also allows for obtaining a model based on the type of downstream task. Towards this end, three types of model selection based on the downstream task performance were evaluated: classification-best, segmentation-best, and a best all-round one. When evaluated on held-out classification and weakly supervised learning tasks, the most performant checkpoints often occur earlier in training, indicating potential performance saturation mid way in the training for histopathology models. These results highlight the importance of appropriate model selection for self-supervised learning in histopathology.", "keywords": [ "Model selection", "Self-supervised Learning", "Histopathology", "Vision Transformer", "Deep Learning" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "i4jHy0ewke", "primary_area": "applications to neuroscience & cognitive science", "n_author": 1, "author_info_list": [ { "name": "Luke Taylor", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 44, "n_ref": 69, "n_ref_all": 102, "n_fig": 25, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 24, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 999, "formula_len_all_1": 958, "len_all": 116404, "len_all_1": 61404, "len_abs": 927, "len_title": 98, "len_sents": 37088, "len_sents_1": 30531, "n_sents": 223, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 929, "title": "Shapeshifters: Auditory cortical neurons switch from polysemantic to monosemantic under anesthesia", "abs": "General anesthesia transitions the brain from a conscious to an unconscious state, but how does sensory processing differ between these conditions? To address this question, we trained neural network encoding models to predict the responses of auditory cortical neurons to natural sounds in both awake and anesthetized ferrets. Utilizing mechanistic interpretability methods, such as feature visualization, linearization and sparse autoencoders, we analyzed these networks tuning and connectivity to uncover key differences in sensory processing. We found that anesthesia decouples neural connectivity, shifting neurons from polysemantic (responding to multiple inputs) to monosemantic (responding to a single input), resulting in a lower-dimensional population code. These findings illuminate how anesthesia alters neural connectivity and encoding, offering new insights into the neural mechanisms underlying sensory processing.", "keywords": [ "auditory cortex", "anesthesia", "mechanstic interpretability", "polysemantic", "monosemantic", "feature visualization", "sparse autoencoders" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 2, 3, 1 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 5 ] }, { "paper_id": "i45NQb2iKO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Bryan Catanzaro", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Vice President" }, { "name": "Chankyu Lee", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Jimmy Lin", "gender": "unknown", "institution": "University of Waterloo", "country": "", "position": "Full Professor" }, { "name": "Mohammad Shoeybi", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Director of Applied Resesrch" }, { "name": "Sheng-Chieh Lin", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Wei Ping", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 51, "n_ref": 133, "n_ref_all": 187, "n_fig": 3, "n_tab": 12, "L_tab": 2570, "n_element_tab": 298, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2079, "n_element_tab_1": 343, "formula_len_all": 384, "formula_len_all_1": 253, "len_all": 189633, "len_all_1": 82243, "len_abs": 1875, "len_title": 99, "len_sents": 48021, "len_sents_1": 33425, "n_sents": 346, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1890, "title": "Universal Multimodal Retrieval with Multimodal LLMs", "abs": "State-of-the-art retrieval models typically address a straightforward search scenario, where retrieval tasks are fixed (e.g., finding a passage to answer a specific question) and only a single modality is supported for both queries and retrieved results. \nThis paper introduces techniques for advancing information retrieval with multimodal large language models (MLLMs), enabling a broader search scenario, termed universal multimodal retrieval, where multiple modalities and diverse retrieval tasks are accommodated. \nTo this end, we first study fine-tuning an MLLM as a bi-encoder retriever on 10 datasets with 16 retrieval tasks. \nOur empirical results show that the fine-tuned MLLM retriever is capable of understanding challenging queries, composed of both text and image, but underperforms a smaller CLIP retriever in cross-modal retrieval tasks due to modality bias from MLLMs. \nTo address the issue, we propose modality-aware hard negative mining to mitigate the modality bias exhibited by MLLM retrievers. \nSecond, we propose to continually fine-tune the universal multimodal retriever to enhance its text retrieval capability while maintaining multimodal retrieval capability. \nAs a result, our model, UniEmb, achieves state-of-the-art performance on the multimodal retrieval benchmark M-BEIR, which spans multiple domains and tasks, while also surpassing the state-of-the-art text retrieval model, NV-Embed-v1, on MTEB retrieval benchmark.\nFinally, we explore to prompt the off-the-shelf MLLMs as the zero-shot reranker to refine the ranking of the candidates from the multimodal retriever. \nWe find that through prompt-and-reranking, MLLMs can further improve multimodal retrieval when the user queries (e.g., text-image composed queries) are more complex and challenging to understand. \nThese findings also pave the way to advance universal multimodal retrieval in the future.", "keywords": [ "multimodal", "embedding model", "retriever", "LLM" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "i3tBySZWrR", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Cecilia Mascolo", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "Full Professor" }, { "name": "Helen Yannakoudakis", "gender": "unknown", "institution": "King's College London, University of London", "country": "GB", "position": "Associate Professor" }, { "name": "Israel Mason-Williams", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Jing Han", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "Postdoc" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 87, "n_ref_all": 112, "n_fig": 10, "n_tab": 15, "L_tab": 13711, "n_element_tab": 1816, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 10724, "n_element_tab_1": 934, "formula_len_all": 398, "formula_len_all_1": 278, "len_all": 199924, "len_all_1": 80576, "len_abs": 1295, "len_title": 136, "len_sents": 54034, "len_sents_1": 25186, "n_sents": 412, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1301, "title": "Machine Unlearning in Audio: Bridging The Modality Gap Via the Prune and Regrow Paradigm", "abs": "The ubiquity and success of deep learning is primarily owed to large human datasets; however, increasing interest in personal data raises questions of how to satisfy privacy legislation in deep learning. Machine unlearning is a nascent discipline centred on satisfying user privacy demands, by enabling data removal requests on trained models. While machine unlearning has reached a good level of maturity in the vision and language domains, applications in audio are largely underexplored, despite it being a highly prevalent and widely used modality. We address this modality gap by providing the first systematic analysis of machine unlearning techniques covering multiple architectures trained on audio datasets. Our analysis highlights that in audio, existing methods fail to remove data for the most likely case of unlearning -- Item Removal. We present a novel Prune and Regrow Paradigm that bolsters sparsity unlearning through Cosine and Post Optimal Pruning, achieving the best unlearning accuracy for 9/12 (75%) of Item Removal experiments and best, or joint best, for for 50% (6/12) of Class Removal Experiments. Furthermore, we run experiments showing performance as unlearning requests scale, and we shed light on the mechanisms underpinning the success of our Prune and Regrow Paradigm.", "keywords": [ "Machine Unlearning", "Audio", "Deep Learning", "Privacy", "Pruning" ], "rating_list": [ 5, 3, 3, 6, 6 ], "soundness_list": [ 3, 2, 1, 3, 2 ], "presentation_list": [ 3, 2, 1, 3, 2 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "i3f2N3iHl0", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Han Quan", "gender": "unknown", "institution": "China Agricultural University", "country": "CN", "position": "MS student" }, { "name": "Lang Zheng", "gender": "Male", "institution": "China Agricultural University", "country": "CN", "position": "MS student" }, { "name": "Ruxue Xing", "gender": "Female", "institution": "China Agricultural University", "country": "CN", "position": "MS student" }, { "name": "Zhenghan Chen", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "youhuan yang", "gender": "Male", "institution": "Zhengzhou University", "country": "CN", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 22, "n_ref_uni": 18, "n_ref": 22, "n_ref_all": 26, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1188, "formula_len_all_1": 1124, "len_all": 90788, "len_all_1": 56318, "len_abs": 1816, "len_title": 149, "len_sents": 26275, "len_sents_1": 25678, "n_sents": 205, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1751, "title": "Adaptive Tensor Attention Networks with Cross-Domain Transfer for Drug-Target Interaction Prediction", "abs": "The prediction of drug-target interactions is fundamental to the advancement of drug discovery. We present a groundbreaking unified theory for Drug-Target Interaction prediction with Domain Adaptation (DTI-DA), seamlessly integrating concepts from quantum mechanics, differential geometry, and information theory. Our framework introduces a novel DTI symplectic structure that captures the intrinsic geometry of drug-target interactions, leading to a Quantum Optimal Transport theorem that provides a rigorous foundation for domain adaptation in the DTI context. We develop a quantum statistical mechanical formulation of DTI-DA, introducing DTI-preserving quantum channels and deriving a Quantum Wasserstein distance tailored to drug discovery applications. Our information-geometric perspective yields a Quantum Fisher-Rao metric for DTI, resulting in a quantum Cramer-Rao bound that establishes fundamental limits on DTI prediction accuracy. We propose a unified variational principle for DTI-DA, encompassing quantum and classical aspects, which leads to a novel algorithm based on geometric stochastic gradient Langevin dynamics. Furthermore, we extend classical statistical inference to the quantum domain, deriving a Quantum Rao-Blackwell theorem and a Quantum Bayesian Cramer-Rao bound specifically for DTI-DA. These theoretical advancements not only deepen our understanding of the DTI-DA problem but also suggest new algorithmic approaches with provable guarantees. Preliminary numerical experiments on quantum-inspired DTI-DA algorithms demonstrate significant improvements in prediction accuracy and domain adaptation capabilities compared to classical methods, particularly for challenging out-of-distribution scenarios in drug discovery.", "keywords": [ "Domain Adaptive Prediction\uff0cAttention" ], "rating_list": [ 3, 5, 1, 1 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "i3e92uSZCp", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Laura Smith", "gender": "Female", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Sehoon Ha", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Seungeun Rho", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Tianyu Li", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Xue Bin Peng", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 6, "n_ref_uni": 31, "n_ref": 68, "n_ref_all": 88, "n_fig": 13, "n_tab": 2, "L_tab": 723, "n_element_tab": 49, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1477, "formula_len_all_1": 483, "len_all": 141755, "len_all_1": 65490, "len_abs": 1400, "len_title": 79, "len_sents": 46302, "len_sents_1": 31320, "n_sents": 379, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 31, "L_abs": 1407, "title": "Language Guided Skill Discovery", "abs": "Skill discovery methods enable agents to learn diverse emergent behaviors without explicit rewards. To make learned skills useful for downstream tasks, obtaining a semantically diverse repertoire of skills is crucial. While some approaches use discriminators to acquire distinguishable skills and others focus on increasing state coverage, the direct pursuit of \u2018semantic diversity\u2019 in skills remains underexplored. We hypothesize that leveraging the semantic knowledge of large language models (LLM) can lead us to improve semantic diversity of resulting behaviors. In this sense, we introduce Language Guided Skill Discovery (LGSD), a skill discovery framework that aims to directly maximize the semantic diversity between skills. LGSD takes user prompts as input and outputs a set of semantically distinctive skills. The prompts serve as a means to constrain the search space into a semantically desired subspace, and the generated LLM outputs guide the agent to visit semantically diverse states within the subspace. We demonstrate that LGSD enables legged robots to visit different user-intended areas on a plane by simply changing the prompt. Furthermore, we show that language guidance aids in discovering more diverse skills compared to five existing skill discovery methods in robot-arm manipulation environments. Lastly, LGSD provides a simple way of utilizing learned skills via natural language.", "keywords": [ "Unsupervised Skill Discovery", "Guided Skill Discovery", "Reinforcement Learning", "Language Guided RL" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "i3aFjkfnXO", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Litong Feng", "gender": "Male", "institution": "SenseTime Research", "country": "HK", "position": "Associate Research Director" }, { "name": "Mengcheng Lan", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xue Jiang", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Yiping Ke", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Yue Zhou", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Wei Zhang", "gender": "Male", "institution": "SenseTime Research", "country": "CN", "position": "Research Director" } ], "n_formula": 7, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 81, "n_ref_all": 111, "n_fig": 12, "n_tab": 7, "L_tab": 4626, "n_element_tab": 308, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 576, "n_element_tab_1": 40, "formula_len_all": 372, "formula_len_all_1": 0, "len_all": 186713, "len_all_1": 64966, "len_abs": 1387, "len_title": 125, "len_sents": 57017, "len_sents_1": 30918, "n_sents": 545, "n_sents_1": 261, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1395, "title": "GeoMath: A Benchmark for Multimodal Mathematical Reasoning in Remote Sensing", "abs": "Vision-language models (VLMs) have demonstrated impressive performance in various Earth observation tasks, particularly in zero-shot capabilities. However, their mathematical reasoning skills in remote sensing (RS) remain unexplored due to the lack of relevant data. To close this gap, we introduce \\dataset, a multimodal mathematical reasoning benchmark meticulously designed for the RS domain. It comprises 3773 high-quality vehicle-related questions from aerial perspectives, spanning 6 mathematical subjects and 20 topics. All data used in this benchmark were collected by our drones from various altitudes and perspectives. Despite the limited geographical coverage, full access to all parameters of the RS images and detailed vehicle information ensures that the constructed mathematical problems are rigorous and diverse. With GeoMath, we have conducted a comprehensive and quantitative evaluation of 14 prominent VLMs. Solving these math problems requires high-resolution visual perception and domain-specific mathematical knowledge, which poses a challenge even for state-of-the-art VLMs. We further explore the impact of image resolution and the zero-shot prompting strategy on the scores, analyzing the reasons behind GPT-4o's reasoning errors. By comparing the gap between InternVL2 and GPT-4o, we find that the latter exhibits some level of cross-view knowledge transfer capability.", "keywords": [ "Benchmark", "Vision-language Models", "Multimodal Mathematical Reasoning", "Remote Sensing" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 5, 3, 3 ] }, { "paper_id": "i3T0wvQDKg", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Daniel John Lawson", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "Associate Professor" }, { "name": "Ed Davis", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "PhD student" }, { "name": "Ian Gallagher", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Lecturer" }, { "name": "Patrick Rubin-Delanchy", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Full Professor" } ], "n_formula": 28, "n_formula_1": 18, "n_ref_uni": 56, "n_ref": 91, "n_ref_all": 116, "n_fig": 10, "n_tab": 10, "L_tab": 3042, "n_element_tab": 211, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 2213, "n_element_tab_1": 139, "formula_len_all": 2454, "formula_len_all_1": 1232, "len_all": 161011, "len_all_1": 62526, "len_abs": 1342, "len_title": 43, "len_sents": 40633, "len_sents_1": 25734, "n_sents": 330, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 1352, "title": "Valid Conformal Prediction for Dynamic GNNs", "abs": "Dynamic graphs provide a flexible data abstraction for modelling many sorts of real-world systems, such as transport, trade, and social networks. Graph neural networks (GNNs) are powerful tools allowing for different kinds of prediction and inference on these systems, but getting a handle on uncertainty, especially in dynamic settings, is a challenging problem.\n\nIn this work we propose to use a dynamic graph representation known in the tensor literature as the unfolding, to achieve valid prediction sets via conformal prediction. This representation, a simple graph, can be input to any standard GNN and does not require any modification to existing GNN architectures or conformal prediction routines. \n\nOne of our key contributions is a careful mathematical consideration of the different inference scenarios which can arise in a dynamic graph modelling context. For a range of practically relevant cases, we obtain valid prediction sets with almost no assumptions, even dispensing with exchangeability. In a more challenging scenario, which we call the semi-inductive regime, we achieve valid prediction under stronger assumptions, akin to stationarity. \n\nWe provide real data examples demonstrating validity, showing improved accuracy over baselines, and sign-posting different failure modes which can occur when those assumptions are violated.", "keywords": [ "Graph neural networks", "Graph machine learning", "conformal prediction" ], "rating_list": [ 5, 3, 5, 8, 8 ], "soundness_list": [ 3, 4, 2, 3, 3 ], "presentation_list": [ 3, 4, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 3, 3 ], "confidence_list": [ 3, 4, 2, 3, 4 ] }, { "paper_id": "i3QV4XgsLA", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 10, "author_info_list": [ { "name": "Allan Dos Santos Costa", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Ameya Daigavane", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Emine Kucukbenli", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Senior Research Scientist" }, { "name": "Franco Pellegrini", "gender": "Male", "institution": "International School for Advanced Studies Trieste", "country": "", "position": "Assistant Professor" }, { "name": "Ilan Mitnikov", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "MS student" }, { "name": "JOSEPH JACOBSON", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Karsten Kreis", "gender": "unknown", "institution": "NVIDIA", "country": "CA", "position": "Research Scientist" }, { "name": "Mario Geiger", "gender": "Male", "institution": "NVIDIA", "country": "CH", "position": "Researcher" }, { "name": "Tess Smidt", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Zhonglin Cao", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Applied Scientist" } ], "n_formula": 24, "n_formula_1": 9, "n_ref_uni": 55, "n_ref": 88, "n_ref_all": 127, "n_fig": 13, "n_tab": 4, "L_tab": 1018, "n_element_tab": 74, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1209, "n_element_tab_1": 84, "formula_len_all": 2359, "formula_len_all_1": 1402, "len_all": 155750, "len_all_1": 63117, "len_abs": 1101, "len_title": 83, "len_sents": 44146, "len_sents_1": 26122, "n_sents": 342, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1106, "title": "EquiJump: Protein Dynamics Simulation via SO(3)-Equivariant Stochastic Interpolants", "abs": "Mapping the conformational dynamics of proteins is crucial for elucidating their functional mechanisms. While Molecular Dynamics (MD) simulation enables detailed time evolution of protein motion, its computational toll hinders its use in practice. To address this challenge, multiple deep learning models for reproducing and accelerating MD have been proposed drawing on transport-based generative methods. However, existing work focuses on generation through transport of samples from prior distributions, that can often be distant from the data manifold. The recently proposed framework of stochastic interpolants, instead, enables transport between arbitrary distribution endpoints. Building upon this work, we introduce EquiJump, a transferable SO(3)-equivariant model that bridges all-atom protein dynamics simulation time steps directly. Our approach unifies diverse sampling methods and is benchmarked against existing models on trajectory data of fast folding proteins. EquiJump achieves state-of-the-art results on dynamics simulation with a transferable model on all of the fast folding proteins.", "keywords": [ "biomolecules", "proteins", "molecular dynamics", "generative models", "neural transport", "stochastic interpolants" ], "rating_list": [ 5, 6, 3, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 4, 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 4, 2 ] }, { "paper_id": "i3KSorBQxF", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Alberto Maria Metelli", "gender": "Male", "institution": "Politecnico di Milano", "country": "", "position": "Assistant Professor" }, { "name": "Federico Corso", "gender": "Male", "institution": "Polytechnic Institute of Milan", "country": "", "position": "MS student" }, { "name": "Marcello Restelli", "gender": "Male", "institution": "Politecnico di Milano", "country": "IT", "position": "Associate Professor" }, { "name": "Marco Mussi", "gender": "Male", "institution": "Politecnico di Milano", "country": "IT", "position": "Postdoc" }, { "name": "Riccardo Zamboni", "gender": "unknown", "institution": "University of Edinburgh", "country": "GB", "position": "Intern" } ], "n_formula": 102, "n_formula_1": 28, "n_ref_uni": 25, "n_ref": 63, "n_ref_all": 71, "n_fig": 0, "n_tab": 10, "L_tab": 4395, "n_element_tab": 471, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 6, "n_element_tab_1": 4, "formula_len_all": 11194, "formula_len_all_1": 1953, "len_all": 221626, "len_all_1": 77650, "len_abs": 1100, "len_title": 133, "len_sents": 61096, "len_sents_1": 32973, "n_sents": 579, "n_sents_1": 262, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1230, "title": "No-regret Learning with Revealed Transitions in Adversarial Markov Decision Processes", "abs": "When learning in Adversarial Markov Decision Processes (MDPs), agents must deal with a sequence of arbitrarily chosen transition models and losses. In this paper, we consider the setting in which the transition model chosen by the adversary is revealed at the end of each episode. We propose the notion of smoothed MDP whose transition model aggregates with a generic function $f_t$ the ones experienced so far. Coherently, we define the concept of smoothed regret, and we devise Smoothed Online Mirror Descent (SOMD), an enhanced version of OMD that leverages a novel regularization term to effectively learn in this setting. For specific choices of the aggregation function $f_t$ defining the smoothed MDPs we retrieve, under full-feedback, a regret bound of order $\\widetilde{\\mathcal O}(L^{3/2}\\sqrt{TL}+L\\overline{C}_f^{\\mathsf{P}})$ where $T$ is the number of episodes, $L$ is the horizon of the episode, and $\\overline{C}_f^{\\mathsf{P}}$ is a novel index of the degree of maliciousness of the adversarially chosen transitions. Under bandit feedback on the losses, we obtain a bound of order $\\widetilde{\\mathcal O}(L^{3/2}\\sqrt{XAT}+L\\overline{C}_f^{\\mathsf{P}})$ using a simple importance weighted estimator on the losses.", "keywords": [ "Adversarial Markov Decision Processes", "Reinforcement Learning", "Online Learning" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "i3DyRNgCey", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Kai Olav Ellefsen", "gender": "Male", "institution": "University of Oslo", "country": "", "position": "Associate Professor" }, { "name": "Mateusz Wasiluk", "gender": "Male", "institution": "University of Oslo", "country": "", "position": "PhD student" }, { "name": "Mikkel Elle Lepper\u00f8d", "gender": "Male", "institution": "Simula Research Laboratory", "country": "NO", "position": "Principal Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 41, "n_ref": 64, "n_ref_all": 78, "n_fig": 5, "n_tab": 13, "L_tab": 6309, "n_element_tab": 1038, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 231, "n_element_tab_1": 55, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 145650, "len_all_1": 61316, "len_abs": 1631, "len_title": 133, "len_sents": 37381, "len_sents_1": 32124, "n_sents": 244, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1639, "title": "Self-Supervised Pseudodata Filtering for Improved Replay with Sub-Optimal Generators", "abs": "Continual learning of a sequence of tasks without forgetting previously acquired knowledge is one of the main challenges faced by modern deep neural networks. In the class-incremental scenario (aka open-set learning), one of the most difficult continual learning problems, new classes are presented to a classifier over time. The model needs to be able to learn and recognize these new classes while also retaining its knowledge of previously witnessed ones. A common approach is to make it revisit the old classes or their features in some form, either by analysing stored exemplars or by using artificially generated samples. The latter approach, Generative Replay, usually relies on a separate generator trained alongside the main classifier. Since the generator also needs to learn continually, it is usually retrained on every task, using its own generated samples as training data representing older classes. This can lead to error propagation and accumulating features unimportant or confusing for the classifier, reducing the overall performance for larger numbers of tasks. We propose a simple filtering mechanism for mitigating this issue \u2013 whenever pseudodata is generated for a new task, the classifier can reject samples it is not able to classify with sufficient confidence, thus preventing both models from retraining on poor-quality data. We tested the filter on several datasets, including real-life images, using various combinations of models, as the method can be applied regardless of the network architectures. We show that filtering improves the classifier's accuracy and provide statistical analysis of the results.", "keywords": [ "continual learning", "catastrophic forgetting", "generative replay", "feature replay", "deep learning" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "i2yxXoAekh", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Francesco Di Giovanni", "gender": "Male", "institution": "Valence Labs powered by recursion", "country": "GB", "position": "Researcher" }, { "name": "Michael M. Bronstein", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Nian Liu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Thomas Laurent", "gender": "Male", "institution": "Loyola Marymount University", "country": "US", "position": "Full Professor" }, { "name": "Xavier Bresson", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" }, { "name": "Xiaoxin He", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 102, "n_formula_1": 26, "n_ref_uni": 54, "n_ref": 131, "n_ref_all": 182, "n_fig": 10, "n_tab": 19, "L_tab": 5271, "n_element_tab": 519, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 3141, "n_element_tab_1": 242, "formula_len_all": 8962, "formula_len_all_1": 1993, "len_all": 242950, "len_all_1": 75351, "len_abs": 1452, "len_title": 108, "len_sents": 62859, "len_sents_1": 28836, "n_sents": 640, "n_sents_1": 256, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1465, "title": "Wavelet-based Graph Convolution via Chebyshev Decomposition", "abs": "Spectral graph convolution, an important tool of data filtering on graphs, relies on two essential decisions: selecting spectral bases for signal transformation and parameterizing the kernel for frequency analysis. While recent techniques mainly focus on standard Fourier transform and vector-valued spectral functions, they fall short in flexibility to model signal distributions over diverse spatial ranges, and capacity of spectral function.\nIn this paper, we present a novel wavelet-based graph convolution network, namely WaveGC, which integrates multi-resolution spectral bases and a matrix-valued filter kernel. Theoretically, we establish that WaveGC can effectively capture and decouple short-range and long-range information, providing superior filtering flexibility, surpassing existing graph convolutional networks and graph Transformers (GTs). \nTo instantiate WaveGC, we introduce a novel technique for learning general graph wavelets by separately combining odd and even terms of Chebyshev polynomials. This approach strictly satisfies wavelet admissibility criteria. Our numerical experiments showcase the capabilities of the new network. By replacing the Transformer part in existing architectures with WaveGC, we consistently observe improvements in both short-range and long-range tasks. This underscores the effectiveness of the proposed model in handling different scenarios. Our code is available at ", "keywords": [ "graph neural network", "graph wavelet", "graph signal processing" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 5, 3, 4 ] }, { "paper_id": "i2ue8J6aqI", "primary_area": "applications to neuroscience & cognitive science", "n_author": 7, "author_info_list": [ { "name": "Chenze Wang", "gender": "Male", "institution": "Zhejiang Gongshang University", "country": "CN", "position": "Researcher" }, { "name": "Feng Xia", "gender": "Male", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "Full Professor" }, { "name": "Jingjing Zhou", "gender": "Female", "institution": "Zhejiang Gongshang University", "country": "CN", "position": "Associate Professor" }, { "name": "Mujie Liu", "gender": "Female", "institution": "Federation University Australia", "country": "AU", "position": "PhD student" }, { "name": "Qichao Dong", "gender": "Male", "institution": "Zhejiang Gongshang University", "country": "", "position": "MS student" }, { "name": "Vidya Saikrishna", "gender": "unknown", "institution": "Federation University Australia", "country": "", "position": "Lecturer" }, { "name": "JING REN", "gender": "Female", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 8, "n_ref_uni": 37, "n_ref": 59, "n_ref_all": 78, "n_fig": 8, "n_tab": 5, "L_tab": 2038, "n_element_tab": 191, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 892, "n_element_tab_1": 96, "formula_len_all": 934, "formula_len_all_1": 655, "len_all": 131550, "len_all_1": 64330, "len_abs": 1487, "len_title": 126, "len_sents": 39232, "len_sents_1": 29482, "n_sents": 279, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1491, "title": "Scalable Diffusion for Bio-topological Representation Learning on Brain Graphs", "abs": "The topological structure information of the brain graph is critical in discovering bio-topological properties that underlie brain function and pathology. Authentic representations of brain graphs in many clinical applications heavily rely on these bio-topological properties. While existing studies have made strides in analyzing brain graph topology, they are often constrained by single-scale structural analysis and hence fail to extract these properties across multiple scales, thus potentially leading to incomplete and distorted representations. To address this limitation, we propose a novel Scalable diffusion model for bio-TOpological REpresentation learning on Brain graphs (BrainSTORE). BrainSTORE constructs multiscale topological structures within brain graphs, facilitating a deep exploration of bio-topological properties. By embedding these features into the training process and prioritizing bio-topological feature reconstruction, BrainSTORE learns representations that are more reflective of underlying brain organization. Furthermore, BrainSTORE utilizes a unified architecture to integrate these features effectively, yielding improved bio-topological representations which are more robust and biologically meaningful. To the best of our knowledge, this is the first study to investigate bio-topological properties in brain graph representation learning. Extensive experiments demonstrate that BrainSTORE outperforms state-of-the-art methods in brain disease detection.", "keywords": [ "Diffusion; Representation Learning; Graph Learning; Brain; Scalability; Topological Analysis" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "i2r7LDjba3", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 13, "author_info_list": [ { "name": "Hanqi Jiang", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Pengfei Jin", "gender": "Male", "institution": "Massachusetts General Hospital, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Quanzheng Li", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Associate Professor" }, { "name": "Sekeun Kim", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Postdoc" }, { "name": "Sifan Song", "gender": "Male", "institution": "Massachusetts General Hospital, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Tianming Liu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Professor" }, { "name": "Tianze Yang", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Xiang Li", "gender": "Male", "institution": "Massachusetts General Hospital, Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "Xiaowei Yu", "gender": "Male", "institution": "University of Texas at Arlington, University of Texas at Arlington", "country": "", "position": "PhD student" }, { "name": "Yi Pan", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Yiwei Li", "gender": "Male", "institution": "Massachusetts General Hospital, Harvard University", "country": "US", "position": "Intern" }, { "name": "Yucheng Shi", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Zihao Wu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 39, "n_ref": 73, "n_ref_all": 87, "n_fig": 5, "n_tab": 7, "L_tab": 1754, "n_element_tab": 170, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 607, "formula_len_all_1": 552, "len_all": 138973, "len_all_1": 54242, "len_abs": 1567, "len_title": 59, "len_sents": 39997, "len_sents_1": 25974, "n_sents": 282, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1641, "title": "ECHOPulse: ECG Controlled Echocardio-gram Video Generation", "abs": "Echocardiography (ECHO) is essential for cardiac assessments, but its video quality and interpretation heavily relies on manual expertise, leading to inconsistent results from clinical and portable devices. ECHO video generation offers a solution by improving automated monitoring through synthetic data and generating high-quality videos from routine health data. However, existing models often face high computational costs, slow inference, and rely on complex conditional prompts that require experts' annotations. To address these challenges, we propose ECHOPulse, an ECG-conditioned ECHO video generation model. ECHOPulse introduces two key advancements: (1) it accelerates ECHO video generation by leveraging VQ-VAE tokenization and masked visual token modeling for fast decoding, and (2) it conditions on readily accessible ECG signals, which are highly coherent with ECHO videos, bypassing complex conditional prompts. To the best of our knowledge, this is the first work to use time-series prompts like ECG signals for ECHO video generation. ECHOPulse not only enables controllable synthetic ECHO data generation but also provides updated cardiac function information for disease monitoring and prediction beyond ECG alone. Evaluations on three public and private datasets demonstrate state-of-the-art performance in ECHO video generation across both qualitative and quantitative measures. Additionally, ECHOPulse can be easily generalized to other modality generation tasks, such as cardiac MRI, fMRI, and 3D CT generation. We will make the synthetic ECHO dataset, along with the code and model, publicly available upon acceptance.", "keywords": [ "Medical video generation", "ECHO synthesis", "Multimodality", "Wearable device", "Medical foundation model" ], "rating_list": [ 8, 6, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 3, 4, 4 ] }, { "paper_id": "i2Ul8WIQm7", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Alexey Kurakin", "gender": "Male", "institution": "Research, Google", "country": "", "position": "Research Software Engineer" }, { "name": "Leo Marchyok", "gender": "Male", "institution": "Oregon State University", "country": "US", "position": "Undergrad student" }, { "name": "Nicholas Carlini", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Sanghyun Hong", "gender": "Male", "institution": "Oregon State University", "country": "", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 32, "n_ref": 74, "n_ref_all": 110, "n_fig": 12, "n_tab": 3, "L_tab": 312, "n_element_tab": 23, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 208, "formula_len_all_1": 42, "len_all": 150627, "len_all_1": 66148, "len_abs": 1611, "len_title": 107, "len_sents": 54171, "len_sents_1": 33296, "n_sents": 442, "n_sents_1": 270, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1603, "title": "Evaluating Privacy Risks of Parameter-Efficient Fine-Tuning", "abs": "Parameter-efficient fine-tuning (PEFT) is a new paradigm for fine-tuning language models at scale. Unlike standard fine-tuning,\nPEFT adjusts only a small number of parameters, making it more computationally accessible and enabling practitioners to develop personalized services by fine-tuning models on user data. Because the models are trained on user data, this emerging paradigm may attract adversaries who want to extract sensitive information from fine-tuning data. However, their privacy implications have not been well-understood in the literature.\n\nIn this paper, we study the impact of this new fine-tuning paradigm on privacy. We use an off-the-shelf data extraction attack as a vehicle to evaluate the privacy risk on two pre-trained language models fine-tuned on 2 datasets, repeated 5 times with different random seeds, resulting in a total of 100 variations. Our main findings are: (1) for practitioners employing PEFT to construct personalized models, the fine-tuned models have lower privacy risks while maintaining reasonable utility; (2) for developers designing new PEFT algorithms,\nwhile safer than standard fine-tuning, certain design choices in the algorithms increase memorization in an unexpected way; and (3) for researchers auditing the privacy of fine-tuned models, employing weak differential privacy is sufficient to mitigate existing data extraction risks without significantly compromising model utility. We hope our work encourages the safe adoption and development of PEFT algorithms in practice, as well as future work on advancing stronger privacy auditing mechanisms.", "keywords": [ "Parameter-efficient fine-tuning", "Privacy risk" ], "rating_list": [ 5, 5, 6, 8, 5 ], "soundness_list": [ 3, 3, 2, 4, 2 ], "presentation_list": [ 3, 2, 3, 4, 2 ], "contribution_list": [ 2, 2, 3, 4, 2 ], "confidence_list": [ 4, 4, 4, 3, 2 ] }, { "paper_id": "i28ZjVxl81", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Achmad Ginanjar", "gender": "Male", "institution": "The University of Queensland", "country": "AU", "position": "PhD student" }, { "name": "PRIYANKA SINGH", "gender": "Female", "institution": "Dartmouth College", "country": "US", "position": "Postdoc" }, { "name": "Wen Hua", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "", "position": "Associate Professor" }, { "name": "Xue Li", "gender": "Male", "institution": "University of Queensland", "country": "AU", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 21, "n_ref": 41, "n_ref_all": 52, "n_fig": 3, "n_tab": 8, "L_tab": 3595, "n_element_tab": 256, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 2226, "n_element_tab_1": 218, "formula_len_all": 378, "formula_len_all_1": 348, "len_all": 96585, "len_all_1": 51416, "len_abs": 1723, "len_title": 52, "len_sents": 23706, "len_sents_1": 19692, "n_sents": 226, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 1736, "title": "DEALING WITH OUT OF DISTRIBUTION IN PREDICTION PROBLEM", "abs": "The open world assumption in model development means that a model may lack sufficient\ninformation to effectively handle data that is completely different or out of distribution\n(OOD). When a model encounters OOD data, its performance can significantly decrease.\nImproving the model\u2019s performance in dealing with OOD can be achieved through gener-\nalization by adding noise, which can be easily done with deep learning. However, many\nadvanced machine learning models are resource-intensive and designed to work best with\nspecialized hardware (GPU), which may not always be available for common users with\nhardware limitations. To provide a deep understanding and solution on OOD for gen-\neral user, this study explores detection, evaluation, and prediction tasks within the context\nof OOD on tabular datasets using common consumer hardware (CPU). It demonstrates\nhow users can identify OOD data from available datasets and provide guidance on eval-\nuating the OOD selection through simple experiments and visualizations. Furthermore,\nthe study introduces Tabular Contrast Learning (TCL), a technique specifically designed\nfor tabular prediction tasks. While achieving better results compared to heavier models,\nTCL is more efficient even when trained without specialised hardware, making it useful\nfor general machine-learning users with computational limitations. This study includes\na comprehensive comparison with existing approaches within their best hardware setting\n(GPU) compared with TCL on common hardware (CPU), focusing on both accuracy and\nefficiency. The results show that TCL exceeds other models, including gradient boosting\ndecision trees, contrastive learning, and other deep learning models, on the classification\ntask.", "keywords": [ "representation learning", "tabular data", "out of distribution" ], "rating_list": [ 1, 3, 5, 1 ], "soundness_list": [ 1, 1, 3, 2 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "i25WJWnsmq", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Ke Wan", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Ruoqing Zhu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "Associate Professor" }, { "name": "Yifan Cui", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuyao Zhang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 16, "n_ref_uni": 44, "n_ref": 87, "n_ref_all": 96, "n_fig": 4, "n_tab": 1, "L_tab": 75, "n_element_tab": 8, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 75, "n_element_tab_1": 8, "formula_len_all": 800, "formula_len_all_1": 783, "len_all": 122496, "len_all_1": 58598, "len_abs": 1272, "len_title": 166, "len_sents": 36459, "len_sents_1": 29266, "n_sents": 229, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 116, "L_abs": 1277, "title": "Optimizing Dynamic Treatment Strategies with Reinforcement Learning and Dual-Hawkes Process in Clinical Environments", "abs": "Modeling the timing of critical events and controlling associated risks through treatment options are crucial aspects of healthcare. However, current methods fall short in optimizing dynamic treatment plans to improve clinical outcomes. A key challenge lies in modeling the intensity functions of critical events throughout disease progression and capturing the dynamic interactions between patient conditions and treatments. To address this, we propose integrating reinforcement learning with a Generative Adversarial Network (GAN) and a dual-Hawkes process model to develop intelligent agents capable of delivering personalized and adaptive treatment strategies. The dual-Hawkes process allows us to model the intensity of both disease progression and recovery, while accounting for long-term dependencies. The GAN simulates real-world clinical environments using raw time-to-event data, without requiring detailed treatment annotations. By interacting with GAN, our model-based reinforcement learning agent learns an optimal dynamic policy that leverages long-term historical dependencies. When applied to the MIMIC-III dataset, our approach significantly increased the duration that patients remained in a healthy state, outperforming established machine learning policies.", "keywords": [ "Reinforcement Learning" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 1, 2, 2 ], "presentation_list": [ 1, 3, 2 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 3, 5, 4 ] }, { "paper_id": "i1Yxnar4mj", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Amirreza Shiralinasab Langari", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "PhD student" }, { "name": "Kim Khoa Nguyen", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "", "position": "Associate Professor" }, { "name": "Leila Yeganeh", "gender": "Female", "institution": "York University", "country": "", "position": "Postdoc" } ], "n_formula": 31, "n_formula_1": 20, "n_ref_uni": 10, "n_ref": 16, "n_ref_all": 26, "n_fig": 3, "n_tab": 6, "L_tab": 649, "n_element_tab": 68, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 902, "n_element_tab_1": 77, "formula_len_all": 2884, "formula_len_all_1": 1812, "len_all": 131500, "len_all_1": 69142, "len_abs": 1025, "len_title": 149, "len_sents": 45587, "len_sents_1": 30569, "n_sents": 402, "n_sents_1": 264, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1239, "title": "Grothendieck Graph Neural Networks Framework: An Algebraic Platform for Crafting Topology-Aware GNNs", "abs": "Due to the structural limitations of Graph Neural Networks (GNNs), particularly those relying on conventional neighborhoods, alternative aggregation strategies have been explored to enhance expressive power. This paper proposes a novel approach by generalizing the concept of neighborhoods through algebraic covers to overcome these limitations.\nWe introduce the Grothendieck Graph Neural Networks (GGNN) framework, providing an algebraic platform for systematically defining and refining diverse covers for graphs. The GGNN framework translates these covers into matrix representations, extending the scope of designing GNN models by incorporating desired message-passing strategies.\nBased on the GGNN framework, we propose Sieve Neural Networks (SNN), a new GNN model that leverages the notion of sieves from category theory. SNN demonstrates competitive performance in experiments, particularly in differentiating between strongly regular graphs, and exemplifies the versatility of GGNN in generating novel architectures.\nIn conclusion, our work advances the design of GNNs by introducing algebraic structures that empower more expressive message-passing mechanisms, addressing the limitations of traditional neighborhood-based methods.", "keywords": [ "Graph Neural Networks", "Categorical Deep Learning", "Algebraic Deep Learning", "Graph Isomorphism", "Graph Classification" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "i1NNCrRxdM", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Kianoosh Ashouritaklimi", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Leo Zhang", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Robert Cornish", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "Researcher" }, { "name": "Yee Whye Teh", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" } ], "n_formula": 58, "n_formula_1": 30, "n_ref_uni": 51, "n_ref": 132, "n_ref_all": 153, "n_fig": 0, "n_tab": 3, "L_tab": 1103, "n_element_tab": 126, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1037, "n_element_tab_1": 77, "formula_len_all": 4768, "formula_len_all_1": 2544, "len_all": 178482, "len_all_1": 71736, "len_abs": 989, "len_title": 108, "len_sents": 53339, "len_sents_1": 31151, "n_sents": 418, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 992, "title": "SymDiff: Equivariant Diffusion via Stochastic Symmetrisation", "abs": "We propose SYMDIFF, a novel method for constructing equivariant diffusion\nmodels using the recently introduced framework of stochastic symmetrisation.\nSYMDIFF resembles a learned data augmentation that is deployed at sampling\ntime, and is lightweight, computationally efficient, and easy to implement on\ntop of arbitrary off-the-shelf models. Notably, in contrast to previous work,\nSYMDIFF typically does not require any neural network components that are\nintrinsically equivariant, avoiding the need for complex parameterizations and the\nuse of higher-order geometric features. Instead, our method can leverage highly\nscalable modern architectures as drop-in replacements for these more constrained\nalternatives. We show that this additional flexibility yields significant empirical\nbenefit on E(3)-equivariant molecular generation. To the best of our knowledge,\nthis is the first application of symmetrisation to generative modelling, suggesting\nits potential in this domain more generally.", "keywords": [ "Equivariance", "Diffusion Models", "Symmetrisation", "Molecular Generation", "Markov Categories" ], "rating_list": [ 5, 8, 8 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 4, 3 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "i1G4AWXHRv", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Reza Abbasi", "gender": "Male", "institution": "Sharif University of Technology", "country": "IR", "position": "MS student" }, { "name": "Ser-Nam Lim", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Associate Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 26, "n_ref": 31, "n_ref_all": 46, "n_fig": 4, "n_tab": 2, "L_tab": 1100, "n_element_tab": 200, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 923, "n_element_tab_1": 12, "formula_len_all": 654, "formula_len_all_1": 654, "len_all": 106229, "len_all_1": 59531, "len_abs": 1975, "len_title": 130, "len_sents": 35672, "len_sents_1": 29569, "n_sents": 236, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1985, "title": "Superpipeline: A Universal Approach for Reducing GPU Memory Usage in Large Models", "abs": "The rapid growth in size and complexity of machine learning models, particularly in natural language processing and computer vision, has led to significant challenges in model execution on hardware with limited resources. This paper introduces Superpipeline, a novel framework designed to optimize the execution of large-scale AI models on constrained hardware for both training and inference phases. Our approach focuses on dynamically managing model execution by partitioning models into individual layers and efficiently transferring these partitions between GPU and CPU memory.\nSuperpipeline achieves substantial reductions in GPU memory consumption\u2014up to 60\\% in our experiments\u2014while maintaining model accuracy and acceptable processing speeds. This enables the execution of models that would otherwise exceed available GPU memory capacity. Unlike existing solutions that primarily target inference or specific model types, Superpipeline demonstrates broad applicability across large language models (LLMs), vision-language models (VLMs), and vision-based models.\nWe evaluate Superpipeline's effectiveness through comprehensive experiments on diverse models and hardware configurations. Our method is characterized by two key parameters that allow fine-tuning of the trade-off between GPU memory usage and processing speed. Importantly, Superpipeline does not require model retraining or parameter modification, ensuring full preservation of the original model's output fidelity.\nThe simplicity and flexibility of Superpipeline make it a valuable tool for researchers and practitioners working with state-of-the-art AI models under hardware constraints. It enables the use of larger models or increased batch sizes on existing hardware, potentially accelerating innovation across various machine learning applications. This work represents a significant step towards democratizing access to advanced AI models and optimizing their deployment in resource-constrained environments.", "keywords": [ "GPU memory optimization", "Inference Efficiency", "Training Optimization", "Large-Scale Models", "GPU-CPU data transfer" ], "rating_list": [ 3, 1, 5, 1, 5 ], "soundness_list": [ 1, 1, 2, 1, 3 ], "presentation_list": [ 1, 2, 2, 2, 3 ], "contribution_list": [ 2, 1, 2, 1, 2 ], "confidence_list": [ 4, 4, 3, 5, 3 ] }, { "paper_id": "i1BTP8wFYM", "primary_area": "learning on time series and dynamical systems", "n_author": 8, "author_info_list": [ { "name": "Bo Yang", "gender": "unknown", "institution": "Jilin University", "country": "CN", "position": "Full Professor" }, { "name": "Changchun Li", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Associate Professor" }, { "name": "Chenglong Hu", "gender": "unknown", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "Ximing Li", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Full Professor" }, { "name": "Yiming Wang", "gender": "unknown", "institution": "Jilin University", "country": "CN", "position": "Postdoc" }, { "name": "Zhengnan Zhang", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "GengheZhang", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "danjiawen", "gender": "Female", "institution": "Jilin University", "country": "CN", "position": "MS student" } ], "n_formula": 34, "n_formula_1": 11, "n_ref_uni": 36, "n_ref": 65, "n_ref_all": 82, "n_fig": 3, "n_tab": 12, "L_tab": 22213, "n_element_tab": 3347, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 4525, "n_element_tab_1": 419, "formula_len_all": 1896, "formula_len_all_1": 614, "len_all": 208738, "len_all_1": 63644, "len_abs": 1536, "len_title": 117, "len_sents": 38070, "len_sents_1": 23298, "n_sents": 324, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1540, "title": "Generalizing Dynamics Modeling Easier from Representation Perspective", "abs": "Learning system dynamics from observations is a critical problem in many applications over various real-world complex systems, e.g., climate, ecology, and fluid systems. Recently, the neural-based dynamics modeling method has become the prevalent solution, where its basic idea is to embed the original states of objects into a latent space before learning the dynamics using neural-based methods such as neural Ordinary Differential Equations (ODE). Given observations from different complex systems, the existing dynamics modeling methods offer a specific model for each observation, resulting in poor generalization. Inspired by the great success of pre-trained models, we raise a question: whether we can conduct a generalized Pre-trained Dynamic EncoDER (PDEDER), which, for various complex systems, can embed their original states into a latent space, where the dynamics can be easier captured. To conduct this generalized PDEDER, we collect 153 sets of real-world and synthetic observations from 24 complex systems. Inspired by the success of time series forecasting using Pre-trained Language Models (PLM), we can employ any PLM and further update it over these dynamic observations by tokenization techniques to achieve the generalized PDEDER. Given any future dynamic observation, we can fine-tune PDEDERwith any specific dynamics modeling method. We evaluate PDEDER on 18 dynamic systems by long/short-term forecasting under both in-domain and cross-domain settings and the empirical results indicate the effectiveness of PDEDER.", "keywords": [ "Dynamics Modeling", "Ordinary Differential Equations", "Pre-trained Language Models" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "i0zzO7Hslk", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Long-Kai Huang", "gender": "unknown", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Zhanfeng Mo", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Sinno Pan", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" } ], "n_formula": 58, "n_formula_1": 26, "n_ref_uni": 54, "n_ref": 117, "n_ref_all": 153, "n_fig": 15, "n_tab": 9, "L_tab": 4108, "n_element_tab": 617, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5759, "formula_len_all_1": 2591, "len_all": 230041, "len_all_1": 68775, "len_abs": 1448, "len_title": 127, "len_sents": 70655, "len_sents_1": 32160, "n_sents": 476, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1486, "title": "Parameter and Memory Efficient Pretraining via Low-rank Riemannian Optimization", "abs": "Pretraining large language models often requires significant computational resources and memory due to their vast parameter amount. An effective approach to enhance parameter efficiency in both training and inference is to parameterize each full-size weight as the product of two trainable low-rank factors. While low-rank fine-tuning has achieved great success, low-rank pretraining remains challenging as it requires learning extensive knowledge from scratch under the restrictive low-rank parameterization. During standard low-rank pretraining, separately optimizing the low-rank factors introduces redundant information from the full gradient, which hinders the learning process. To achieve efficient yet effective low-rank pretraining, we propose a **Lo**w-rank **R**iemannian **O**ptimizer (**LORO**). At each LORO update step, the low-rank factor pairs are jointly updated to ensure their full-size product moves along the steepest descent direction on the low-rank manifold, without the need to compute any memory-intensive full-size matrices or gradients. Hence, our LORO finds low-rank models that achieve high performance comparable to full-size pretrained models, while significantly reducing memory usage and accelerating both training and inference. A LLaMA 1B model pretrained with LORO achieves a perplexity score of 2\\% better than the full-size baseline, with a 54\\% reduction in model memory, a $\\times1.8$ speedup in training, and a $\\times2.2$ speedup in inference.", "keywords": [ "Neural Network Optimization", "Parameter Efficient Pretraining", "Low-rank Optimization" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "i0qnHlgxFm", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "An Li", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "MS student" }, { "name": "benyou wang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Chaojie Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "MS student" }, { "name": "Fei Yu", "gender": "Female", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "PhD student" }, { "name": "Yumou Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "MS student" } ], "n_formula": 44, "n_formula_1": 9, "n_ref_uni": 47, "n_ref": 73, "n_ref_all": 108, "n_fig": 21, "n_tab": 2, "L_tab": 442, "n_element_tab": 75, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 73, "n_element_tab_1": 4, "formula_len_all": 6893, "formula_len_all_1": 582, "len_all": 163011, "len_all_1": 63661, "len_abs": 3334, "len_title": 126, "len_sents": 44307, "len_sents_1": 29135, "n_sents": 410, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1260, "title": "Periodical Moving Average Accelerates Gradient Accumulation for Post-Training", "abs": "High gradient variance challenges training Large Language Models (LLMs) on memory-limited devices. Existing practical approaches, such as small batch size or using Gradient Accumulation (GA), face the dilemma between low convergence rates due to high variance in parameter updates and long training times due to the serial GA process. In this paper, we identify that the exponential nature of the Exponential Moving Average (EMA) rapidly forgets historical gradients at an exponential rate in momentum updates, making it difficult to utilize the historical gradients to stabilize the update steps. To address this issue, we embed the idea of GA into the momentum update and propose the Periodical Moving Average (PMA) technique. PMA splits the training steps into periods and employs moving averages instead of EMA in each period. We apply PMA to AdamW and Lion, resulting in AdamW-PMA and Lion-PMA. Theoretical analysis demonstrates that AdamW-PMA achieves a comparable convergence rate with Adam. Extensive experiments showcase the superiority of PMA on post-training tasks, including Supervised Fine-Tuning and Direct Preference Optimization, that the PMA-based methods achieve approximately at least $2\\times$ speedup and higher scores on downstream tasks.", "keywords": [ "Optimization", "Large Language Models", "Efficient Machine Learning" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "i0e0OMK8xM", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Alexandre Rame", "gender": "Male", "institution": "Google", "country": "FR", "position": "research scientist" }, { "name": "Arthur Douillard", "gender": "Male", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Johan Ferret", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Leonard Hussenot", "gender": "unknown", "institution": "Google", "country": "", "position": "PhD student" }, { "name": "Nino Vieillard", "gender": "unknown", "institution": "Google Deepmind", "country": "FR", "position": "Researcher" }, { "name": "Olivier Bachem", "gender": "Male", "institution": "Google Brain", "country": "", "position": "Research scientist" }, { "name": "Pier Giuseppe Sessa", "gender": "Male", "institution": "Google DeepMind", "country": "CH", "position": "Researcher" }, { "name": "Pierre-Louis Cedoz", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Sertan Girgin", "gender": "unknown", "institution": "Google", "country": "FR", "position": "Researcher" }, { "name": "Robert Dadashi", "gender": "Male", "institution": "Google", "country": "", "position": "Research Software Engineer" } ], "n_formula": 30, "n_formula_1": 10, "n_ref_uni": 76, "n_ref": 258, "n_ref_all": 324, "n_fig": 24, "n_tab": 5, "L_tab": 1425, "n_element_tab": 102, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 994, "n_element_tab_1": 67, "formula_len_all": 1941, "formula_len_all_1": 910, "len_all": 269741, "len_all_1": 67624, "len_abs": 1268, "len_title": 58, "len_sents": 56558, "len_sents_1": 31577, "n_sents": 418, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1276, "title": "WARP: On the Benefits of Weight Averaged Rewarded Policies", "abs": "Reinforcement learning from human feedback (RLHF) aligns large language models by encouraging their generations to have high rewards, using a reward model trained on human preferences. To prevent forgetting of pre-trained knowledge, RLHF usually incorporates a KL regularization; this forces the policy to remain close to its initialization, though it hinders the reward optimization. To address the trade-off between KL and reward, in this paper we introduce a novel alignment strategy named Weight Averaged Rewarded Policies (WARP), merging policies in the weight space at three distinct stages. First, it uses the exponential moving average of the policy as a dynamic anchor in the KL regularization. Second, it applies spherical interpolation to merge independently fine-tuned policies into a new enhanced one. Third, it linearly interpolates between this merged model and the initialization, to recover features from pre-training. This procedure is then applied iteratively, with each iteration's final model used as an advanced initialization for the next, progressively refining the KL-reward Pareto front, achieving superior rewards at fixed KL. Experiments with Gemma policies validate that WARP improves their quality and alignment, outperforming open-source models.", "keywords": [ "LLM", "Alignment", "RLHF", "Model Merging" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "i0VqD2KaYt", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Hua Li", "gender": "unknown", "institution": "Hainan University", "country": "CN", "position": "Associate Professor" }, { "name": "Huazhong Liu", "gender": "Male", "institution": "Hainan University", "country": "CN", "position": "Full Professor" }, { "name": "Qirui LIN", "gender": "Male", "institution": "Southeast University", "country": "", "position": "MS student" }, { "name": "Runmin Cong", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "Full Professor" }, { "name": "Sam Kwong", "gender": "Male", "institution": "Lingnan University", "country": "HK", "position": "Chair Professor" }, { "name": "Shijie Lian", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yuheng Jia", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Yutong Li", "gender": "Male", "institution": "Hainan University", "country": "", "position": "Undergrad student" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 52, "n_ref": 102, "n_ref_all": 128, "n_fig": 11, "n_tab": 10, "L_tab": 2218, "n_element_tab": 138, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2287, "n_element_tab_1": 63, "formula_len_all": 461, "formula_len_all_1": 359, "len_all": 121451, "len_all_1": 53051, "len_abs": 1394, "len_title": 143, "len_sents": 34861, "len_sents_1": 22910, "n_sents": 268, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1401, "title": "ViT-UWA: Vision Transformer Underwater-Adapter for Dense Predictions Beneath the Water Surface", "abs": "Vision Transformer (ViT) and its variants have witnessed a significant success in computer vision. However, they do not perform well in underwater dense prediction tasks due to challenges like complex underwater environments, quality degradation, and light scattering in underwater images. To solve this problem, we propose the Vision Transformer Underwater-Adapter (ViT-UWA), the first detail-focused and adapted ViT backbone for underwater dense prediction tasks, without requiring task-specific pretraining. In ViT-UWA, we first introduce High-frequency Components Prior (HFCP) to add high-frequency information of underwater images to the plain ViT, which can help recover and capture lost high-frequency information of underwater images. Then, we propose an Detail Aware Module (DAM) to obtain a detail-focused multi-scale convolutional feature pyramid, which can be used in kinds of dense prediction tasks. Through the ViT-CNN Interaction Module (VCIM), we achieve bidirectional feature fusion between ViT and CNN. We evaluate ViT-UWA on multiple underwater dense prediction tasks, including semantic segmentation, instance segmentation, and object detection. Notably, with only ImageNet-22K pretraining, our ViT-UWA-B yields state-of-the-art 46.4 box AP and 44.2 mask AP on USIS10K dataset. We hope ViT-UWA could provide a new backbone for future research on underwater dense prediction tasks.", "keywords": [ "Underwater Image Dense Prediction", "Adapted ViT Backbone" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "hzxvMqYYMA", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Tingting Jiang", "gender": "Female", "institution": "School of Computer Science, Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Xingyu Wu", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Postdoc" }, { "name": "Yan Zhong", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 79, "n_formula_1": 20, "n_ref_uni": 62, "n_ref": 137, "n_ref_all": 204, "n_fig": 4, "n_tab": 8, "L_tab": 522, "n_element_tab": 52, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 1550, "n_element_tab_1": 17, "formula_len_all": 5009, "formula_len_all_1": 1117, "len_all": 200680, "len_all_1": 68718, "len_abs": 2193, "len_title": 175, "len_sents": 61447, "len_sents_1": 34025, "n_sents": 427, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 125, "L_abs": 2105, "title": "Understanding the Generalization of Blind Image Quality Assessment: A Theoretical Perspective on Multi-level Quality Features", "abs": "Due to the high annotation costs and relatively small scale of existing Image Quality Assessment (IQA) datasets, attaining consistent generalization remains a significant challenge for prevalent deep learning (DL)-based IQA methods. Although it is widely believed that quality perception information primarily resides in low-level image features, and that effective representation learning for the multi-level image features and distortion information is deemed crucial for the generalization of the Blind IQA (BIQA) methods, the theoretical underpinnings for this belief still remain elusive. Therefore, in this work, we investigate the role of multi-level image features in the generalization and quality perception ability of the CNN-based BIQA models from a theoretical perspective. For the role of low-level features, in Theorem 1, we innovatively derive an upper bound of Rademacher Average and the corresponding generalization bound for the CNN-based BIQA framework under distribution invariance in training and test sets, which indicates that the generalization ability tends to be reduced as the level of quality features increases, demonstrating the value of low-level features. In addition, under distribution shifts, a much tighter generalization bound is proposed in Theorem 2, which elucidates the theoretical impact of distributional differences between training and test sets on generalization performance. For the role of high-level features, in Theorem 3, we prove that BIQA networks tend to possess higher Betti number complexity by learning higher-level quality features. This indicates a larger representation power with smaller empirical errors, highlighting the value of high-level features. The three proposed Theorems can provide theoretical support for the enhanced generalization in existing BIQA methods. Furthermore, these theoretical findings reveal an inherent tension between robust generalization and strong representation power in BIQA networks, which inspires us to explore effective strategies to reduce empirical error without compromising the generalization ability.", "keywords": [ "Machine Learning;Computer Vision;Image Quality Assessment", "Generalization", "Theoretical Guarantees." ], "rating_list": [ 6, 3, 8, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 2, 5, 5, 4 ] }, { "paper_id": "hzuumhfYSO", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Alex Oshin", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Arshiya Taj Abdul", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Augustinos D Saravanos", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Evangelos Theodorou", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Hunter Kuperman", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Vincent Pacelli", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Postdoc" } ], "n_formula": 212, "n_formula_1": 36, "n_ref_uni": 42, "n_ref": 87, "n_ref_all": 171, "n_fig": 11, "n_tab": 4, "L_tab": 993, "n_element_tab": 139, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 240, "n_element_tab_1": 2, "formula_len_all": 22779, "formula_len_all_1": 2719, "len_all": 258094, "len_all_1": 78307, "len_abs": 3173, "len_title": 115, "len_sents": 77444, "len_sents_1": 34795, "n_sents": 740, "n_sents_1": 274, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1808, "title": "Deep Distributed Optimization for Large-Scale Quadratic Programming", "abs": "Quadratic programming (QP) forms a crucial foundation in optimization, encompassing a broad spectrum of domains and serving as the basis for more advanced algorithms. Consequently, as the scale and complexity of modern applications continue to grow, the development of efficient and reliable QP algorithms becomes increasingly vital. In this context, this paper introduces a novel deep learning-aided distributed optimization architecture designed for tackling large-scale QP problems. First, we combine the state-of-the-art Operator Splitting QP (OSQP) method with a consensus approach to derive DistributedQP, a new method tailored for network-structured problems, with convergence guarantees to optimality. Subsequently, we unfold this optimizer into a deep learning framework, leading to DeepDistributedQP, which leverages learned policies to accelerate reaching to desired accuracy within a restricted amount of iterations. Our approach is also theoretically grounded through Probably Approximately Correct (PAC)-Bayes theory, providing generalization bounds on the expected optimality gap for unseen problems. The proposed framework, as well as its centralized version DeepQP, significantly outperform their standard optimization counterparts on a variety of tasks such as randomly generated problems, optimal control, linear regression, transportation networks and others. Notably, DeepDistributedQP demonstrates strong generalization by training on small problems and scaling to solve much larger ones (up to 50K variables and 150K constraints) using the same policy. Moreover, it achieves orders-of-magnitude improvements in wall-clock time compared to OSQP. The certifiable performance guarantees of our approach are also demonstrated, ensuring higher-quality solutions over traditional optimizers.", "keywords": [ "Deep Learning Aided Optimization", "Distributed Optimization", "Large-Scale Quadratic Programming" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "hzVpZDrW73", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Fei Zhao", "gender": "unknown", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Shaohui Lin", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Researcher" }, { "name": "Shaosheng Cao", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Researcher" }, { "name": "Wenxuan Huang", "gender": "unknown", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Xiangfeng Xu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Yunhang Shen", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Zheyu Ye", "gender": "Male", "institution": "Xiaohongshu Inc", "country": "CN", "position": "Researcher" }, { "name": "Zijie Zhai", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 42, "n_ref": 131, "n_ref_all": 154, "n_fig": 9, "n_tab": 9, "L_tab": 5472, "n_element_tab": 683, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1412, "formula_len_all_1": 1259, "len_all": 220995, "len_all_1": 65417, "len_abs": 1594, "len_title": 157, "len_sents": 60650, "len_sents_1": 28095, "n_sents": 427, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 108, "L_abs": 1618, "title": "Dynamic-LLaVA: Efficient Multimodal Large Language Models via Dynamic Vision-language Context Sparsification", "abs": "Multimodal Large Language Models (MLLMs) have achieved remarkable success in vision understanding, reasoning, and interaction. However, the inference computation and memory increase progressively with the generation of output tokens during decoding, directly affecting the efficacy of MLLMs. Existing methods attempt to reduce the vision context redundancy to achieve efficient MLLMs. Unfortunately, the efficiency benefits of the vision context reduction in the prefill stage gradually diminish during the decoding stage. To address this problem, we proposed a dynamic vision-language context sparsification framework Dynamic-LLaVA, which dynamically reduces the redundancy of vision context in the prefill stage and decreases the memory and computation overhead of the generated language context during decoding. Dynamic-LLaVA designs a tailored sparsification inference scheme for different inference modes, i.e., prefill, decoding with and without KV cache, to achieve efficient inference of MLLMs. In practice, Dynamic-LLaVA can reduce computation consumption by $\\sim$75\\% in the prefill stage. Meanwhile, throughout the entire generation process of MLLMs, Dynamic-LLaVA reduces the $\\sim$50\\% computation consumption under decoding without KV cache, while saving $\\sim$50\\% GPU memory overhead when decoding with KV cache, due to the vision-language context sparsification. Extensive experiments also demonstrate that Dynamic-LLaVA achieves efficient inference for MLLMs with negligible understanding and generation ability degradation or even performance gains compared to the full-context inference baselines.", "keywords": [ "Efficient Multimodal Large Language Model" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 4, 5, 4 ] }, { "paper_id": "hzQcilRe2v", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 3, "author_info_list": [ { "name": "Didem Unat", "gender": "Female", "institution": "Ko\u00e7 University", "country": "", "position": "Associate Professor" }, { "name": "Mohamed Wahib", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Principal Researcher" }, { "name": "Muhammet Abdullah Soyturk", "gender": "unknown", "institution": "Ko\u00e7 University", "country": "", "position": "MS student" } ], "n_formula": 47, "n_formula_1": 34, "n_ref_uni": 52, "n_ref": 91, "n_ref_all": 108, "n_fig": 5, "n_tab": 7, "L_tab": 4981, "n_element_tab": 251, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 680, "n_element_tab_1": 10, "formula_len_all": 1823, "formula_len_all_1": 1193, "len_all": 253514, "len_all_1": 58482, "len_abs": 2633, "len_title": 115, "len_sents": 80242, "len_sents_1": 25498, "n_sents": 607, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1511, "title": "Elastic and Balanced End-to-end Training of Dynamic LLMs with DynMo", "abs": "To reduce the computational and memory costs of Large Language Models (LLMs), schemes that introduce dynamic training are increasingly emerging. Examples of dynamic models are: a) Mixture of Experts (MoEs) at which token routing affects the compute balance, b) gradual pruning of the parameters of a model, c) dynamically freezing layers, d) dynamic sparse attention schemes, e) early exit of tokens as they pass through the model layers, and f) Mixture of Depths (MoDs) schemes where tokens bypass blocks. One side effect that limits the practical value of dynamic models is the introduction of workload imbalance among workers, which in turn negatively affects the efficiency in distributed training. We propose a dynamic load balancing solution DynMo), with a proof that it satisfies maximum reduction in imbalance, to adaptively maintain equal compute workloads among different workers in pipeline parallelism. In addition, DynMo dynamically packs work into fewer workers, while sustaining training throughput, to release the idle workers back to the job manager. DynMo supports both single nodes with multi-GPUs and systems with multi-GPU multi-nodes. In comparison to static distributed training solutions (Megatron-LM and DeepSpeed), DynMo accelerates the end-to-end training of dynamic GPT models by up to 1.23x (MoEs), 3.18x (parameter pruning), 2.23x (layer freezing), 4.02x (sparse attention), 4.52x (early exit), and 1.17x (MoDs). DynMo is available at https://anonymous.4open.science/r/DynMo-4D04/.", "keywords": [ "Transformers; Dynamic Models; Pipeline Parallelism; LLMs" ], "rating_list": [ 3, 5, 8, 5 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 1, 4, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 5, 3, 2 ] }, { "paper_id": "hz3NtNpDNv", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Aldo Lipani", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "Associate Professor" }, { "name": "Chuan Wang", "gender": "Male", "institution": "Lulea University of Technology", "country": "", "position": "PhD student" }, { "name": "Ujjal Kr Dutta", "gender": "Male", "institution": "University College London", "country": "", "position": "Intern" }, { "name": "Yukun Hu", "gender": "unknown", "institution": "University College London, University of London", "country": "", "position": "Associate Professor" } ], "n_formula": 24, "n_formula_1": 16, "n_ref_uni": 30, "n_ref": 62, "n_ref_all": 117, "n_fig": 9, "n_tab": 19, "L_tab": 12080, "n_element_tab": 2059, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1495, "n_element_tab_1": 265, "formula_len_all": 1819, "formula_len_all_1": 1510, "len_all": 273491, "len_all_1": 65795, "len_abs": 831, "len_title": 53, "len_sents": 89279, "len_sents_1": 27612, "n_sents": 666, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 811, "title": "Hottel Zone Physics-Constrained Networks for Furnaces", "abs": "This paper investigates a novel approach to improve the temperature profile prediction of furnaces in foundation industries, crucial for sustainable manufacturing. While existing methods like the Hottel Zone model are accurate, they lack real-time inference capabilities. Deep learning methods excel in speed and prediction but require careful generalization for real-world applications. We propose a regularization technique that leverages the Hottel Zone method to make deep neural networks physics-aware, improving prediction accuracy for furnace temperature profiles. Our approach demonstrates effectiveness on various neural network architectures, including Multi-Layer Perceptrons (MLP), Long Short-Term Memory (LSTM) and Kolmogorov-Arnold Networks (KANs). We also discussion the data generation involved.", "keywords": [ "Hottel Zone method", "Physics-Informed Neural Networks", "Radiation Heat Transfer", "Furnaces" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 1 ], "presentation_list": [ 1, 1, 3, 1 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "hyrRupfS0o", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Junze Yin", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Ruizhe Zhang", "gender": "Male", "institution": "Simons Institute for the Theory of Computing", "country": "US", "position": "Postdoc" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" } ], "n_formula": 30, "n_formula_1": 21, "n_ref_uni": 111, "n_ref": 173, "n_ref_all": 181, "n_fig": 0, "n_tab": 2, "L_tab": 132, "n_element_tab": 5, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2740, "formula_len_all_1": 2080, "len_all": 212973, "len_all_1": 74850, "len_abs": 989, "len_title": 155, "len_sents": 40374, "len_sents_1": 31284, "n_sents": 359, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1187, "title": "Revisiting Quantum Algorithms for Linear Regressions: Quadratic Speedups without Data-Dependent Parameters", "abs": "Linear regression is one of the most fundamental linear algebra problems. Given a dense matrix $A \\in \\mathbb{R}^{n \\times d}$ and a vector $b$, the goal is to find $x'$ such that $\\|| Ax' - b \\||\\_2^2 \\leq (1+\\epsilon) \\min\\_{x} \\|| A x - b \\||\\_2^2$. The best classical algorithm takes $O(nd) + \\mathrm{poly}(d/\\epsilon)$ time [Clarkson and Woodruff STOC 2013, Nelson and Nguyen FOCS 2013]. On the other hand, quantum linear regression algorithms can achieve exponential quantum speedups, as shown in [Wang \\emph{Phys. Rev. A 96}, 012335, Kerenidis and Prakash ITCS 2017, Chakraborty, Gily{\\'e}n and Jeffery ICALP 2019]. However, the running times of these algorithms depend on some quantum linear algebra-related parameters, such as $\\kappa(A)$, the condition number of $A$. In this work, we develop a quantum algorithm that runs in $\\widetilde{O}(\\epsilon^{-1}\\sqrt{n}d^{1.5}) + \\mathrm{poly}(d/\\epsilon)$ time and outputs a classical solution. It provides a quadratic quantum speedup in $n$ over the classical lower bound without any dependence on data-dependent parameters. In addition, we also show our result can be generalized to multiple regression and ridge linear regression.", "keywords": [ "Linear regression", "quantum algorithms" ], "rating_list": [ 3, 8, 3 ], "soundness_list": [ 2, 4, 4 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 1, 4, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "hyfe5q5TD0", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Akshay Krishnamurthy", "gender": "Male", "institution": "Research, Microsoft", "country": "US", "position": "Senior Principal Research Manager" }, { "name": "Ayush Sekhari", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Runzhe Wu", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Wen Sun", "gender": "unknown", "institution": "Cornell University", "country": "", "position": "Assistant Professor" } ], "n_formula": 177, "n_formula_1": 17, "n_ref_uni": 34, "n_ref": 78, "n_ref_all": 85, "n_fig": 0, "n_tab": 9, "L_tab": 1878, "n_element_tab": 101, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 25903, "formula_len_all_1": 1729, "len_all": 254495, "len_all_1": 75022, "len_abs": 1318, "len_title": 138, "len_sents": 75156, "len_sents_1": 34647, "n_sents": 843, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 89, "L_abs": 1047, "title": "Computationally Efficient RL under Linear Bellman Completeness for Deterministic Dynamics", "abs": "We study computationally and statistically efficient Reinforcement Learning algorithms for the *linear Bellman Complete* setting, a setting that uses linear function approximation to capture value functions and unifies existing models like linear Markov Decision Processes (MDP) and Linear Quadratic Regulators (LQR). While it is known from the prior works that this setting is statistically tractable, it remained open whether a computationally efficient algorithm exists. Our work provides a computationally efficient algorithm for the linear Bellman complete setting that works for MDPs with large action spaces, random initial states, and random rewards but relies on the underlying dynamics to be deterministic. Our approach is based on randomization: we inject random noise into least square regression problems to perform optimistic value iteration. Our key technical contribution is to carefully design the noise to only act in the null space of the training data to ensure optimism while circumventing a subtle error amplification issue.", "keywords": [ "reinforcement learning theory", "linear function approximation" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 4, 3, 3, 4 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "hyb6NCjS8G", "primary_area": "applications to neuroscience & cognitive science", "n_author": 9, "author_info_list": [ { "name": "Huifeng Yin", "gender": "Male", "institution": "Center for Brain Inspired Computing Research (CBICR)", "country": "CN", "position": "MS student" }, { "name": "Lei Deng", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Mingkun Xu", "gender": "Male", "institution": "Guangdong Institute of Intelligence Science and Technology", "country": "CN", "position": "Postdoc" }, { "name": "Qi Xu", "gender": "Male", "institution": "School of Computer Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Shiping Wang", "gender": "Male", "institution": "Fuzhou University", "country": "CN", "position": "Full Professor" }, { "name": "Shuai Zhong", "gender": "Male", "institution": "Guangdong Institute of Intelligence Science and Technology, Zhuhai, China", "country": "", "position": "Principal Researcher" }, { "name": "Yuhong Chen", "gender": "Female", "institution": "Fuzhou University", "country": "CN", "position": "MS student" }, { "name": "Yujie Wu", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "CN", "position": "Researcher" }, { "name": "Zihan Fang", "gender": "Female", "institution": "Fuzhou University", "country": "CN", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 16, "n_ref_uni": 33, "n_ref": 38, "n_ref_all": 55, "n_fig": 6, "n_tab": 1, "L_tab": 2719, "n_element_tab": 64, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 3975, "n_element_tab_1": 103, "formula_len_all": 621, "formula_len_all_1": 622, "len_all": 123128, "len_all_1": 59470, "len_abs": 1903, "len_title": 131, "len_sents": 30976, "len_sents_1": 28574, "n_sents": 191, "n_sents_1": 153, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1915, "title": "Brain-inspired Multi-View Incremental Learning for Knowledge Transfer and Retention", "abs": "The human brain exhibits remarkable proficiency in dynamic learning and adaptation, seamlessly integrating prior knowledge with new information, thereby enabling flexible memory retention and efficient transfer across multiple views. In contrast, traditional multi-view learning methods are predominantly designed for static and fixed-view datasets, leading to the notorious \"view forgetting phenomenon\", where the introduction of new views leads to the erosion of prior knowledge. This phenomenon starkly contrasts with the brain\u2019s remarkable ability to continuously integrate and migrate past knowledge, ensuring both the retention of old information and the assimilation of new insights. This oversight presents a critical challenge: how to efficiently learn and integrate new views while simultaneously preserving knowledge from previously acquired views and enabling flexible knowledge transfer across diverse perspectives.Inspired by underlying neural processing mechanisms, we propose a view transfer learning framework named Hebbian View Orthogonal Projection (HVOP), which realizes efficient knowledge migration and sharing between multi-view data. HVOP constructs a knowledge transfer space (KTS), where the KTS reduces the interference between the old and the new views through an orthogonal learning mechanism. By further incorporating recursive lateral connections and Hebbian learning, the proposed model endows the learning process with brain-like dynamic adaptability, enhancing knowledge transfer and integration, and bringing the model closer to human cognition. We extensively validate the proposed model on node classification tasks and demonstrate its superior performance in knowledge retention and transfer compared to traditional methods. Our results underscore the potential of biologically inspired mechanisms in advancing multi-view learning and mitigating the view forgetting phenomenon.", "keywords": [ "Brain-inspired Knowledge Transfer; Hebbian Learning; Multi-view Incremental Learning; Orthogonal Projection" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "hyYP9MZeYn", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Daniel LK Yamins", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Kevin Tran", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Klemen Kotar", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Yash Shah", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 112, "n_ref": 156, "n_ref_all": 175, "n_fig": 14, "n_tab": 2, "L_tab": 433, "n_element_tab": 34, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 44, "formula_len_all_1": 43, "len_all": 230508, "len_all_1": 58068, "len_abs": 1263, "len_title": 131, "len_sents": 51554, "len_sents_1": 29938, "n_sents": 318, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1269, "title": "Modeling Focal Synaptic Degeneration and Neural Plasticity in Ventral Visual Cortex", "abs": "Strokes affect a significant portion of the population and often result in secondary damage in the form of focal synaptic degeneration. When this occurs in the ventral visual cortex (VVC), it can lead to neurological deficits, including visual function loss. In this paper, we use the VVC as a framework in which to model focal synaptic degeneration and post-injury plasticity. We do so by progressively \"injuring\" synaptic connections in primate visual areas V1, V2, V4, and the inferior temporal cortex (IT), followed by continual retraining of the spared connections on real-world visual stimuli. We demonstrate that the functional signatures of carefully designed differential tasks can localize synaptic decay in the VVC. Initially, categorization performance deteriorates gradually, up to a critical threshold, beyond which there is a sharp drop. This slow decline in performance is marked by a reorganization in nearby neurons, where both visual function and the structure of receptive fields adapt to compensate for the damage. Spared recurrent connections significantly contribute to recovery. Furthermore, we find that the presence of teaching signals in the form of category labels during rehabilitation leads to improved categorization performance recovery.", "keywords": [ "reorganization", "recurrent connections", "self-supervision", "vision", "stroke", "deep artificial neural networks" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "hxpbOfBywA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Lizhong Chen", "gender": "Male", "institution": "Oregon State University", "country": "", "position": "Associate Professor" }, { "name": "Sanghyun Hong", "gender": "Male", "institution": "Oregon State University", "country": "", "position": "Assistant Professor" }, { "name": "Seijoon Kim", "gender": "Male", "institution": "Samsung", "country": "KR", "position": "Principal Researcher" }, { "name": "Tahmid Hasan Pranto", "gender": "Male", "institution": "Oregon State University", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 41, "n_ref": 115, "n_ref_all": 147, "n_fig": 8, "n_tab": 3, "L_tab": 1651, "n_element_tab": 54, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1424, "n_element_tab_1": 106, "formula_len_all": 481, "formula_len_all_1": 397, "len_all": 165305, "len_all_1": 74061, "len_abs": 1422, "len_title": 125, "len_sents": 49903, "len_sents_1": 33696, "n_sents": 407, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1428, "title": "Hessian-Aware Training for Enhancing Model Resilience for In-Memory Computing", "abs": "Deep neural networks are not resilient to bitwise errors in their parameters: even a single-bit error in their memory representation can lead to significant performance degradation. This susceptibility poses great challenges in deploying models on emerging computing platforms, such as in-memory computing devices, where frequent bitwise errors occur. Most prior work addresses this issue with hardware or system-level approaches, such as additional hardware components for checking a model\u2019s integrity at runtime. However, these methods have not been widely deployed since they necessitate substantial infrastructure-wide modifications. In this paper, we study a new approach to address this challenge: we present a novel training method aimed at enhancing a model\u2019s inherent resilience to parameter errors. We define a model-sensitivity metric to measure this resilience and propose a training algorithm with an objective of minimizing the sensitivity. Models trained with our method demonstrate increased resilience to bitwise errors in parameters, particularly with a 50% reduction in the number of bits in the model parameter space whose flipping leads to a 90\u2013100% accuracy drop. Our method also aids in extreme model compression, such as lower bit-width quantization or pruning \u223c70% of parameters, with reduced performance loss. Moreover, our method is compatible with existing strategies to mitigate this susceptibility.", "keywords": [ "DNN Resilience", "Parameter corruptions", "Hessian-aware training" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "hxUMQ4fic3", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Cevahir Koprulu", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "ufuk topcu", "gender": "Unspecified", "institution": "University of Texas, Austin", "country": "US", "position": "Full Professor" }, { "name": "Franck B Djeumou", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 10, "n_ref_uni": 45, "n_ref": 94, "n_ref_all": 115, "n_fig": 8, "n_tab": 4, "L_tab": 1337, "n_element_tab": 213, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 1867, "n_element_tab_1": 272, "formula_len_all": 1403, "formula_len_all_1": 739, "len_all": 188995, "len_all_1": 65352, "len_abs": 2568, "len_title": 74, "len_sents": 52908, "len_sents_1": 30526, "n_sents": 348, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1435, "title": "Neural Stochastic Differential Equations for Uncertainty-Aware Offline RL", "abs": "Offline model-based reinforcement learning (RL) offers a principled approach to using a learned dynamics model as a simulator to optimize a control policy. \nDespite the near-optimal performance of existing approaches on benchmarks with high-quality datasets, most struggle on datasets with low state-action space coverage or suboptimal demonstrations.\nWe develop a novel offline model-based RL approach that particularly shines in low-quality data regimes while maintaining competitive performance on high-quality datasets.\nNeural Stochastic Differential Equations for Uncertainty-aware, Offline RL (NUNO) learns a dynamics model as neural stochastic differential equations (SDE), \nwhere its drift term can leverage prior physics knowledge as inductive bias.\nIn parallel, its diffusion term provides distance-aware estimates of model uncertainty by matching the dynamics' underlying stochasticity near the training data regime while providing high but bounded estimates beyond it.\nTo address the so-called model exploitation problem in offline model-based RL, NUNO builds on existing studies by penalizing and adaptively truncating neural SDE's rollouts according to uncertainty estimates.\nOur empirical results in D4RL and NeoRL MuJoCo benchmarks evidence that NUNO outperforms state-of-the-art methods in low-quality datasets by up to 93% while matching or surpassing their performance by up to 55% in some high-quality counterparts.", "keywords": [ "neural stochastic differential equations", "offline reinforcement learning", "physics-informed machine learning" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "hx8E1L4v2e", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Dong Yang", "gender": "unknown", "institution": "Georgia State University", "country": "US", "position": "PhD student" }, { "name": "HaitaoYang", "gender": "Male", "institution": "Yantai University", "country": "CN", "position": "MS student" }, { "name": "Ranran Li", "gender": "Male", "institution": "IMDEA Networks Institute", "country": "CN", "position": "Researcher" }, { "name": "Wang Kerry", "gender": "Male", "institution": "Inspur Smart City Technology Co, Ltd", "country": "CN", "position": "Researcher" }, { "name": "Yihao Jiang", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Zhaowei Liu", "gender": "Male", "institution": "Yantai University", "country": "CN", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 14, "n_ref_uni": 30, "n_ref": 63, "n_ref_all": 78, "n_fig": 3, "n_tab": 2, "L_tab": 1392, "n_element_tab": 62, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 3144, "n_element_tab_1": 81, "formula_len_all": 955, "formula_len_all_1": 729, "len_all": 108583, "len_all_1": 62016, "len_abs": 1047, "len_title": 112, "len_sents": 32803, "len_sents_1": 25576, "n_sents": 276, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1069, "title": "Anomaly Detection in Dynamic Graphs via Adversarial Autoencoder", "abs": "Anomaly detection in dynamic graphs is a very important task that has attracted a lot of attention. Many dynamic graph anomaly detection methods are already available, but most of these efforts are based on supervised learning. In the real world, however, it is often difficult to collect large amounts of labelled anomaly data, which is not conducive to the training of these supervised methods and severely reduces their ability to be applied in different dynamic graph anomaly detection scenarios. A novel semi-supervised anomaly detection framework \\textbf{AAEDY} for the detection of anomalous edges in dynamic graphs is presented in this paper, which improves reconstruction by combining adversarial based on autoencoder, and discriminates whether an edge is anomalous by comparing the original edge to the reconstructed edge in low-dimensional space. Extensive experiments have been carried out on six real-world datasets, and the experimental results show that \\textbf{AAEDY} can outperform the state-of-the-art competitors in anomaly detection significantly.", "keywords": [ "anomaly detection", "dynamic graph", "graph learning", "deep learning", "autoencoder" ], "rating_list": [ 3, 5, 3, 5, 3 ], "soundness_list": [ 2, 3, 2, 2, 1 ], "presentation_list": [ 1, 3, 1, 2, 2 ], "contribution_list": [ 2, 2, 1, 2, 1 ], "confidence_list": [ 4, 4, 5, 4, 5 ] }, { "paper_id": "hwnObmOTrV", "primary_area": "learning on time series and dynamical systems", "n_author": 8, "author_info_list": [ { "name": "Anne Biton", "gender": "Female", "institution": "Genentech", "country": "US", "position": "Principal Researcher" }, { "name": "Aviv Regev", "gender": "Female", "institution": "Genentech / Roche", "country": "US", "position": "Researcher" }, { "name": "Charlotte Bunne", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "", "position": "Assistant Professor" }, { "name": "Edward De Brouwer", "gender": "Male", "institution": "Genentech", "country": "US", "position": "Researcher" }, { "name": "Jan-Christian Huetter", "gender": "Male", "institution": "Genentech", "country": "US", "position": "Researcher" }, { "name": "Kelvin Y. Chen", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "Assistant Professor" }, { "name": "Martin Rohbeck", "gender": "Male", "institution": "Deutsches Krebsforschungszentrum", "country": "", "position": "PhD student" }, { "name": "Romain Lopez", "gender": "unknown", "institution": "New York University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 69, "n_formula_1": 27, "n_ref_uni": 40, "n_ref": 62, "n_ref_all": 84, "n_fig": 8, "n_tab": 9, "L_tab": 1168, "n_element_tab": 93, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 355, "n_element_tab_1": 29, "formula_len_all": 3912, "formula_len_all_1": 1375, "len_all": 177978, "len_all_1": 66567, "len_abs": 1486, "len_title": 126, "len_sents": 58861, "len_sents_1": 31863, "n_sents": 464, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 78, "L_abs": 1404, "title": "Modeling Complex System Dynamics with Flow Matching Across Time and Conditions", "abs": "Modeling the dynamics of complex real-world systems from temporal snapshot data is crucial for understanding phenomena such as gene regulation, climate change, and financial market fluctuations. Researchers have recently proposed a few methods based either on the Schroedinger Bridge or Flow Matching to tackle this problem, but these approaches remain limited in their ability to effectively combine data from multiple time points and different experimental settings. This integration is essential in real-world scenarios where observations from certain combinations of time points and experimental conditions are missing, either because of experimental costs or sensory failure. To address this challenge, we propose a novel method named Multi-Marginal Flow Matching (MMFM). MMFM first constructs a flow using smooth spline-based interpolation across time points and conditions and regresses it with a neural network using the classifier-free guided Flow Matching framework. This framework allows for the sharing of contextual information about the dynamics across multiple trajectories. We demonstrate the effectiveness of our method on both synthetic and real-world datasets, including a recent single-cell genomics data set with around a hundred chemical perturbations across time points. Our results show that MMFM significantly outperforms existing methods at imputing data at missing time points.", "keywords": [ "Flow Matching", "dynamical systems" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 1, 3, 4 ], "contribution_list": [ 2, 3, 4 ], "confidence_list": [ 5, 3, 3 ] }, { "paper_id": "hwSmPOAmhk", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Alberto Bietti", "gender": "Male", "institution": "Flatiron Institute", "country": "US", "position": "Researcher" }, { "name": "Eshaan Nichani", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Jason D. Lee", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" } ], "n_formula": 212, "n_formula_1": 33, "n_ref_uni": 35, "n_ref": 80, "n_ref_all": 91, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 24810, "formula_len_all_1": 2116, "len_all": 236031, "len_all_1": 68618, "len_abs": 1188, "len_title": 117, "len_sents": 58468, "len_sents_1": 28309, "n_sents": 779, "n_sents_1": 281, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 69, "L_abs": 1193, "title": "Understanding Factual Recall in Transformers via Associative Memories", "abs": "Large language models have demonstrated an impressive ability to perform factual recall. Prior work has found that transformers trained on factual recall tasks can store information at a rate proportional to their parameter count. In our work, we show that shallow transformers can use a combination of associative memories to obtain such near optimal storage capacity. We begin by proving that the storage capacities of both linear and MLP associative memories scale linearly with parameter count. We next introduce a synthetic factual recall task, and prove that a transformer with a single layer of self-attention followed by an MLP can obtain 100\\% accuracy on the task whenever either the total number of self-attention parameters or MLP parameters scales (up to log factors) linearly with the number of facts. In particular, the transformer can trade off between using the value matrices or the MLP as an associative memory to store the dataset of facts. We complement these expressivity results with an analysis of the gradient flow trajectory of a simplified linear attention model trained on our factual recall task, where we show that the model exhibits sequential learning behavior.", "keywords": [ "transformers", "associative memories", "factual recall", "storage capacity", "training dynamics" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "hvLBTpiDt3", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Hoang Pham", "gender": "Male", "institution": "The University of Warwick", "country": "GB", "position": "PhD student" }, { "name": "Hongkai Wen", "gender": "Male", "institution": "University of Warwick", "country": "GB", "position": "Full Professor" }, { "name": "Khoat Than", "gender": "Male", "institution": "Hanoi University of Science and Technology", "country": "", "position": "Associate Professor" }, { "name": "Lan-Cuong Nguyen", "gender": "Male", "institution": "FPT", "country": "VN", "position": "Researcher" }, { "name": "Lichuan Xiang", "gender": "Male", "institution": "University of Warwick", "country": "GB", "position": "Researcher" }, { "name": "Long Tran-Thanh", "gender": "unknown", "institution": "The university of Warwick", "country": "GB", "position": "Full Professor" }, { "name": "Quan Nguyen", "gender": "Male", "institution": "FPT software AI center", "country": "VN", "position": "Researcher" } ], "n_formula": 42, "n_formula_1": 40, "n_ref_uni": 33, "n_ref": 68, "n_ref_all": 84, "n_fig": 3, "n_tab": 10, "L_tab": 2665, "n_element_tab": 133, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2708, "n_element_tab_1": 35, "formula_len_all": 3538, "formula_len_all_1": 3333, "len_all": 166584, "len_all_1": 59199, "len_abs": 1528, "len_title": 127, "len_sents": 52584, "len_sents_1": 23494, "n_sents": 407, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1344, "title": "DPaI: Differentiable Pruning at Initialization with Node-Path Balance Principle", "abs": "Pruning at Initialization (PaI) is a technique in neural network optimization characterized by the proactive elimination of weights before the network's training on designated tasks. This innovative strategy potentially reduces the costs for training and inference, significantly advancing computational efficiency. A key element of PaI's effectiveness is that it considers the significance of weights in an untrained network. It prioritizes the trainability and optimization potential of the pruned subnetworks. Recent methods can effectively prevent the formation of hard-to-optimize networks, e.g. through iterative adjustments at each network layer. However, this way often results in *large-scale discrete optimization problems*, which could make PaI further challenging. This paper introduces a novel method, called *DPaI*, that involves a differentiable optimization of the pruning mask. DPaI adopts a dynamic and adaptable pruning process, allowing easier optimisation processes and better solutions. More importantly, our differentiable formulation enables readily use of the existing rich body of efficient gradient-based methods for PaI. Our empirical results demonstrate that DPaI significantly outperforms current state-of-the-art PaI methods on various architectures, such as Convolutional Neural Networks and Vision-Transformers.", "keywords": [ "Prunning at Initialization", "Sparsity", "Neural Architecture Search" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 2, 5, 2, 3 ] }, { "paper_id": "huy8g3iKy0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Ashwinee Panda", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Postdoc" }, { "name": "Benjamin Th\u00e9rien", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "Sambit Sahu", "gender": "Male", "institution": "CapitalOne", "country": "US", "position": "Principal Researcher" }, { "name": "Stephen Rawls", "gender": "unknown", "institution": "CapitalOne", "country": "US", "position": "Researcher" }, { "name": "Supriyo Chakraborty", "gender": "unknown", "institution": "Capital One", "country": "US", "position": "Researcher" }, { "name": "Tom Goldstein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Vatsal Baherwani", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Undergrad student" }, { "name": "Zain Sarwar", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 31, "n_ref": 54, "n_ref_all": 103, "n_fig": 15, "n_tab": 4, "L_tab": 827, "n_element_tab": 107, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 528, "n_element_tab_1": 64, "formula_len_all": 260, "formula_len_all_1": 211, "len_all": 309101, "len_all_1": 60714, "len_abs": 881, "len_title": 76, "len_sents": 42260, "len_sents_1": 27700, "n_sents": 361, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 883, "title": "Dense Backpropagation Improves Routing for Sparsely-Gated Mixture-of-Experts", "abs": "Mixture of Experts (MoE) pretraining is more scalable than dense Transformer pretraining, because MoEs learn to route inputs to a sparse set of their feedforward parameters. However, this means that MoEs only receive a sparse backward update, leading to problems such as router load imbalance where some experts receive more tokens than others. We present a lightweight approximation method that gives the MoE a dense gradient while only sparsely activating its parameters. A key insight into the design of our method is that at scale, many tokens not routed to a given expert may nonetheless lie in the span of tokens that were routed to that expert, allowing us to create an approximation for the expert output of that token from existing expert outputs. Our dense backpropagation outperforms standard TopK routing across multiple settings without significantly increasing runtime.", "keywords": [ "Mixtureofexperts", "MoE", "routing", "transformer", "LLM" ], "rating_list": [ 6, 3, 5, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "huuKoVQnB0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Christopher Potts", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Tatsunori Hashimoto", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Tristan Thrush", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "PhD student" } ], "n_formula": 93, "n_formula_1": 16, "n_ref_uni": 47, "n_ref": 109, "n_ref_all": 143, "n_fig": 9, "n_tab": 2, "L_tab": 408, "n_element_tab": 40, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7318, "formula_len_all_1": 913, "len_all": 224995, "len_all_1": 67292, "len_abs": 1063, "len_title": 104, "len_sents": 56846, "len_sents_1": 33051, "n_sents": 476, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1066, "title": "Improving Pretraining Data Using Perplexity Correlations", "abs": "Quality pretraining data is often seen as the key to high-performance language models. However, progress in understanding pretraining data has been slow due to the costly pretraining runs required for data selection experiments. We present a framework that avoids these costs and selects high-quality pretraining data without any LLM training of our own. Our work is based on a simple observation: LLM losses on many pretraining texts are correlated with downstream benchmark performance, and selecting high-correlation documents is an effective pretraining data selection method. We build a new statistical framework for data selection centered around estimates of perplexity-benchmark correlations and perform data selection using a sample of 90 LLMs taken from the Open LLM Leaderboard on texts from tens of thousands of web domains. In controlled pretraining experiments at the 160M parameter scale on 8 benchmarks, our approach outperforms DSIR on every benchmark, while matching the best data selector found in DataComp-LM, a hand-engineered bigram classifier.", "keywords": [ "pretraining data; data selection; natural language processing; statistics; large language models" ], "rating_list": [ 6, 5, 8, 5, 6 ], "soundness_list": [ 2, 3, 4, 3, 3 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 3, 4 ] }, { "paper_id": "huo8MqVH6t", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Jin Peng Zhou", "gender": "Male", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "PhD student" }, { "name": "Kilian Q Weinberger", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Professor" }, { "name": "Qizhou Wang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "PhD student" }, { "name": "Saebyeol Shin", "gender": "Female", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Zhanke Zhou", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Visiting student" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 40, "n_formula_1": 16, "n_ref_uni": 35, "n_ref": 83, "n_ref_all": 121, "n_fig": 19, "n_tab": 10, "L_tab": 17912, "n_element_tab": 3048, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 774, "n_element_tab_1": 8, "formula_len_all": 2366, "formula_len_all_1": 834, "len_all": 260182, "len_all_1": 71301, "len_abs": 1353, "len_title": 122, "len_sents": 70619, "len_sents_1": 33161, "n_sents": 578, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1358, "title": "Rethinking LLM Unlearning Objectives: A Gradient Perspective and Go Beyond", "abs": "Large language models (LLMs) should undergo rigorous audits to identify potential risks, such as copyright and privacy infringements. Once these risks emerge, timely updates are crucial to remove undesirable responses, ensuring legal and safe model usage. It has spurred recent research into LLM unlearning, focusing on erasing targeted undesirable knowledge without compromising the integrity of other, non-targeted responses. Existing studies have introduced various unlearning objectives to pursue LLM unlearning without necessitating complete retraining. However, each of these objectives has unique properties, and no unified framework is currently available to comprehend them thoroughly. To fill the gap, we propose the metric of the G-effect, quantifying the impacts of unlearning objectives on model performance from a gradient lens. A significant advantage of our metric is its broad ability to detail the unlearning impacts from various aspects across instances, updating steps, and LLM layers. Accordingly, the G-effect offers new insights into identifying drawbacks of existing unlearning objectives, further motivating us to explore a series of candidate solutions for their mitigation and improvements. Finally, we outline promising directions that merit further studies, aiming at contributing to the community to advance this critical field.", "keywords": [ "LLM Unlearning" ], "rating_list": [ 6, 5, 8, 3, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 4, 2, 4 ], "contribution_list": [ 2, 2, 4, 3, 4 ], "confidence_list": [ 3, 3, 3, 4, 4 ] }, { "paper_id": "htX7AoHyln", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Huaiyu Dai", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Full Professor" }, { "name": "Md Farhamdur Reza", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "PhD student" }, { "name": "Richeng Jin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tianfu Wu", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 60, "n_formula_1": 27, "n_ref_uni": 46, "n_ref": 140, "n_ref_all": 177, "n_fig": 19, "n_tab": 11, "L_tab": 6833, "n_element_tab": 706, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 105, "n_element_tab_1": 20, "formula_len_all": 3590, "formula_len_all_1": 1643, "len_all": 223308, "len_all_1": 71747, "len_abs": 1211, "len_title": 101, "len_sents": 71642, "len_sents_1": 33626, "n_sents": 460, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1247, "title": "GSBA$^K$: $top$-$K$ Geometric Score-based Black-box Attack", "abs": "Existing score-based adversarial attacks mainly focus on crafting $top$-1 adversarial examples against classifiers with single-label classification. Their attack success rate and query efficiency are often less than satisfactory, particularly under small perturbation requirements; moreover, the vulnerability of classifiers with multi-label learning is yet to be studied. In this paper, we propose a comprehensive surrogate free score-based attack, named \\b geometric \\b score-based \\b black-box \\b attack (GSBA$^K$), to craft adversarial examples in an aggressive $top$-$K$ setting for both untargeted and targeted attacks, where the goal is to change the $top$-$K$ predictions of the target classifier. We introduce novel gradient-based methods to find a good initial boundary point to attack. Our iterative method employs novel gradient estimation techniques, particularly effective in $top$-$K$ setting, on the decision boundary to effectively exploit the geometry of the decision boundary. Additionally, GSBA$^K$ can be used to attack against classifiers with $top$-$K$ multi-label learning. Extensive experiential results on ImageNet and PASCAL VOC datasets validate the effectiveness of GSBA$^K$ in crafting $top$-$K$ adversarial examples.", "keywords": [ "Adversarial Attack", "Black-box Attack", "Score-based Attack", "Top-K Attack" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "htOl3M7II8", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Antonio Andrea Gentile", "gender": "unknown", "institution": "Pasqal", "country": "NL", "position": "Researcher" }, { "name": "Atiyo Ghosh", "gender": "unknown", "institution": "Pasqal", "country": "FR", "position": "Researcher" }, { "name": "Hyukgeun Cha", "gender": "Male", "institution": "LG Electronics", "country": "KR", "position": "Principal Researcher" }, { "name": "Jongyeong Lee", "gender": "Male", "institution": "LG Electronics", "country": "", "position": "Researcher" }, { "name": "Jun-Ho Lee", "gender": "Male", "institution": "LG Electronics", "country": "KR", "position": "Principal Researcher" }, { "name": "Jung Jun Park", "gender": "unknown", "institution": "LG Corporation", "country": "", "position": "Researcher" }, { "name": "Nicol\u00f2 Toscano", "gender": "Male", "institution": "Pasqal", "country": "FR", "position": "Researcher" }, { "name": "Seong-Hyok Sean Kim", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Yunjun Choi", "gender": "Male", "institution": "LG Electronics", "country": "", "position": "Research Engineer" } ], "n_formula": 29, "n_formula_1": 26, "n_ref_uni": 28, "n_ref": 49, "n_ref_all": 56, "n_fig": 1, "n_tab": 1, "L_tab": 653, "n_element_tab": 63, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 672, "n_element_tab_1": 69, "formula_len_all": 1431, "formula_len_all_1": 1250, "len_all": 103899, "len_all_1": 65301, "len_abs": 1203, "len_title": 77, "len_sents": 33208, "len_sents_1": 29735, "n_sents": 228, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 29, "L_abs": 1207, "title": "Neural Superposition Networks", "abs": "Machine learning models can be biased towards the solutions of given differential equations in two principal ways: through regularisation, or through architecture design. Recent research has successfully constrained neural network architectures to satisfy divergence-free fields and Laplace's equation in two dimensions. This work reinterprets these architectures as linear superpositions of general formulated solutions. The notion of superposition is then exploited to develop novel architectures which satisfy both these and novel differential equations. In addition to new architectures for Laplace's equation and divergence-free fields, we propose novel constraints apt for the heat equation, and even some nonlinear differential equations including Burgers' equation. Benchmarks of superposition-based approaches against previously published architectures and physics-informed regularisation approaches are presented. We find that embedding differential equation constraints directly into neural network architectures can lead to improved performance and hope our results motivate further development of neural networks architectures developed to adhere specifically to given differential constraints.", "keywords": [ "Scientific Machine Learning", "Physics-Informed Neural Networks", "Differential Equations" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "htDczodFN5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Junyi Li", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Xiaolei Wang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Xin Zhao", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Xinyu Tang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 79, "n_ref_all": 101, "n_fig": 9, "n_tab": 3, "L_tab": 1869, "n_element_tab": 110, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 406, "formula_len_all_1": 277, "len_all": 134867, "len_all_1": 52558, "len_abs": 1222, "len_title": 146, "len_sents": 34822, "len_sents_1": 23586, "n_sents": 268, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 98, "L_abs": 1147, "title": "Investigating the Pre-Training Dynamics of In-Context Learning: Task Recognition vs. Task Learning", "abs": "The emergence of in-context learning (ICL) is potentially attributed to two major abilities: task recognition (TR) for recognizing the task from demonstrations and utilizing pre-trained priors, and task learning (TL) for learning from demonstrations. However, relationships between the two abilities and how such relationships affect the emergence of ICL is unclear. In this paper, we take the first step by examining the pre-training dynamics of the emergence of ICL. With carefully designed metrics, we find that these two abilities are, in fact, competitive during pre-training. Moreover, we observe a negative correlation between the competition and the performance of ICL. Further analysis of common pre-training factors (i.e., model size, dataset size, and data curriculum) demonstrates possible ways to regulate the competition. Based on these insights, we propose a simple yet effective method to better integrate these two abilities for ICL at inference time. Through adaptive ensemble learning, the performance of ICL can be significantly boosted, enabling two small models to outperform a larger one with more than twice the parameters.", "keywords": [ "In-context learning" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "hsnt2TKvLU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Baihe Huang", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Dongya Jia", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Gopala Anumanchipalli", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Assistant Professor" }, { "name": "Jiawei Chen", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Tingle Li", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Xiaobin Zhuang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Yuxuan Wang", "gender": "Male", "institution": "ByteDance", "country": "US", "position": "Researcher" }, { "name": "Zhuo Chen", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Wang Yuping", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 70, "n_ref": 132, "n_ref_all": 156, "n_fig": 9, "n_tab": 6, "L_tab": 865, "n_element_tab": 196, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 936, "n_element_tab_1": 99, "formula_len_all": 1554, "formula_len_all_1": 895, "len_all": 198379, "len_all_1": 69490, "len_abs": 1506, "len_title": 90, "len_sents": 50386, "len_sents_1": 32419, "n_sents": 399, "n_sents_1": 236, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 906, "title": "Object-Aware Audio-Visual Sound Generation", "abs": "Generating accurate sounds for complex audio-visual scenes is challenging, especially when multiple objects and sound sources are present. In this paper, we introduce an object-aware sound generation model that aligns generated sounds with visual objects in a scene. By grounding sound generation in object-centric representations, our model learns to associate specific visual objects with their corresponding sounds. We fine-tune a conditional latent diffusion model with dot-product attention to improve sound-object alignment. At test time, users can compositionally generate sounds by selecting objects via segmentation masks. We theoretically validate our test-time object-grounding ability, ensuring that even subtle sounds can be represented. Quantitative and qualitative evaluations show that our model outperforms baselines, achieving better alignment between objects and their associated sounds.", "keywords": [ "Sound Generation", "Audio-Visual Learning", "Multimodal Learning" ], "rating_list": [ 3, 3, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "hsMkpzr9Oy", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Ali Modarressi", "gender": "Male", "institution": "Center for Information and Language Processing, LMU Munich", "country": "", "position": "PhD student" }, { "name": "Amir Hossein Kargaran", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Fran\u00e7ois Yvon", "gender": "Male", "institution": "ISIR, Sorbonne Universit\u00e9 & CNRS", "country": "FR", "position": "Principal Researcher" }, { "name": "Hinrich Sch\u00fctze", "gender": "Male", "institution": "Center for Information and Language Processing", "country": "DE", "position": "Full Professor" }, { "name": "Jana Diesner", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "Full Professor" }, { "name": "Nafiseh Nikeghbal", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 76, "n_ref": 144, "n_ref_all": 167, "n_fig": 4, "n_tab": 2, "L_tab": 3106, "n_element_tab": 263, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 3823, "n_element_tab_1": 273, "formula_len_all": 136, "formula_len_all_1": 136, "len_all": 243186, "len_all_1": 75923, "len_abs": 1630, "len_title": 129, "len_sents": 45687, "len_sents_1": 35198, "n_sents": 302, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1662, "title": "Mexa: Multilingual Evaluation of English-Centric LLMs via Cross-Lingual Alignment", "abs": "English-centric large language models (LLMs) often show strong multilingual capabilities. However, the multilingual performance of these models remains unclear and is not thoroughly evaluated for many languages. Most benchmarks for multilinguality focus on classic NLP tasks, or cover a minimal number of languages.\nWe introduce Mexa, a method for assessing the multilingual capabilities of pre-trained English-centric LLMs using parallel sentences, which are available for more languages than existing downstream tasks. Mexa leverages the fact that English-centric LLMs use English as a kind of pivot language in their intermediate layers. It computes the alignment between English and non-English languages using parallel sentences to evaluate the transfer of language understanding from English to other languages. This alignment can be used to estimate task performance in other languages.\nWe conduct studies using various parallel datasets (FLORES-200 and Bible), models (Llama family, Gemma family, Mistral, and OLMo), and established downstream tasks (Belebele, m-MMLU, and m-ARC). We explore different methods to compute embeddings in decoder-only models.\nOur results show that Mexa, in its default settings, achieves a statistically significant average Pearson correlation of 0.90 with three established downstream tasks across nine models and two parallel datasets.\nThis suggests that Mexa is a reliable method for estimating the multilingual capabilities of English-centric LLMs, providing a clearer understanding of their multilingual potential and the inner workings of LLMs. HuggingFace Leaderboard: [anonymized URL], GitHub Code: [anonymized URL].", "keywords": [ "multilingual", "evaluation", "low-resource languages", "large language models" ], "rating_list": [ 5, 6, 5, 5, 6 ], "soundness_list": [ 2, 4, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 4 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 3, 4 ] }, { "paper_id": "hrqNOxpItr", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Alice Bizeul", "gender": "Female", "institution": "ETHZ - ETH Zurich", "country": "", "position": "PhD student" }, { "name": "Attila Juhos", "gender": "unknown", "institution": "Max-Planck Institute for Intelligent Systems", "country": "DE", "position": "PhD student" }, { "name": "Julia E Vogt", "gender": "Female", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Assistant Professor" }, { "name": "Patrik Reizinger", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Randall Balestriero", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Assistant Professor" }, { "name": "Wieland Brendel", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen", "country": "DE", "position": "Principal Researcher" }, { "name": "David A. Klindt", "gender": "unknown", "institution": "Cold Spring Harbor Laboratory", "country": "", "position": "Assistant Professor" } ], "n_formula": 101, "n_formula_1": 6, "n_ref_uni": 62, "n_ref": 175, "n_ref_all": 212, "n_fig": 10, "n_tab": 1, "L_tab": 1198, "n_element_tab": 230, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 1198, "n_element_tab_1": 230, "formula_len_all": 5564, "formula_len_all_1": 596, "len_all": 257482, "len_all_1": 80242, "len_abs": 1697, "len_title": 115, "len_sents": 81306, "len_sents_1": 36157, "n_sents": 748, "n_sents_1": 279, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 67, "L_abs": 1706, "title": "Cross-Entropy Is All You Need To Invert the Data Generating Process", "abs": "Supervised learning has become a cornerstone of modern machine learning, yet a comprehensive theory explaining its effectiveness remains elusive. Empirical phenomena, such as neural analogy-making and the linear representation hypothesis, suggest that supervised models can learn interpretable factors of variation in a linear fashion. Recent advances in self-supervised learning, particularly nonlinear Independent Component Analysis, have shown that these methods can recover latent structures by inverting the data generating process. We extend these identifiability results to parametric instance discrimination, \nthen show how insights transfer to the ubiquitous setting of supervised learning with cross-entropy minimization. We prove that even in standard classification tasks, models learn representations of ground-truth factors of variation up to a linear transformation under a certain DGP. We corroborate our theoretical contribution with a series of empirical studies. First, using simulated data matching our theoretical assumptions, we demonstrate successful disentanglement of latent factors. Second, we show that on DisLib, a widely-used disentanglement benchmark, simple classification tasks recover latent structures up to linear transformations. Finally, we reveal that models trained on ImageNet encode representations that permit linear decoding of proxy factors of variation.\nTogether, our theoretical findings and experiments offer a compelling explanation for recent observations of linear representations, such as superposition in neural networks. This work takes a significant step toward a cohesive theory that accounts for the unreasonable effectiveness of supervised learning.", "keywords": [ "supervised learning", "representation learning", "identifiability", "linear representation hypothesis" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 4, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "hrXt6Fdl2P", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Sorachi Kato", "gender": "Male", "institution": "Information Science and Technology", "country": "JP", "position": "PhD student" }, { "name": "Takuya Fujihashi", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "Assistant Professor" }, { "name": "Toshiaki Koike-Akino", "gender": "Male", "institution": "Mitsubishi Electric Research Labs", "country": "US", "position": "Principal Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 12, "n_ref": 16, "n_ref_all": 31, "n_fig": 4, "n_tab": 5, "L_tab": 3457, "n_element_tab": 432, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 3457, "n_element_tab_1": 432, "formula_len_all": 427, "formula_len_all_1": 427, "len_all": 86753, "len_all_1": 66465, "len_abs": 1094, "len_title": 101, "len_sents": 21899, "len_sents_1": 20760, "n_sents": 150, "n_sents_1": 137, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1101, "title": "FV-NeRV: Neural Compression for Free Viewpoint Videos", "abs": "The delivery of free viewpoint videos (FVVs) is gaining popularity because of their ability to provide freely switchable perspectives to remote users as immersive experiences. While smooth view switching is crucial for enhancing user's experiences, FVV delivery faces a significant challenge in balancing traffic and decoding latency. The typical approach sends limited viewpoints and synthesizes the remainings on the user, reducing traffic, but increasing decoding delays. Alternatively, sending more viewpoints reduces the delay, but requires more bandwidth for transmission. In this paper, we propose a novel FVV representation format, Free Viewpoint-Neural Representation for Videos~(FV-NeRV), to address this dilemma in FVV delivery. FV-NeRV reduces both traffic and decoding delay even for content with a large number of virtual viewpoints by overfitting compact neural networks to all viewpoints and pruning and quantizing the trained model. Experiments show that FV-NeRV achieves a comparable or even superior traffic reduction with a faster decoding speed compared to existing FVV codecs.", "keywords": [ "Implicit Neural Representation", "Free Viewpoint Video" ], "rating_list": [ 5, 1, 1, 3, 3 ], "soundness_list": [ 3, 1, 2, 1, 1 ], "presentation_list": [ 3, 2, 2, 2, 2 ], "contribution_list": [ 2, 1, 1, 1, 1 ], "confidence_list": [ 5, 4, 5, 5, 4 ] }, { "paper_id": "hrOlBgHsMI", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Daria Soboleva", "gender": "Female", "institution": "Cerebras Systems, Inc", "country": "US", "position": "Researcher" }, { "name": "Gurpreet Gosal", "gender": "Male", "institution": "Cerebras Systems, Inc", "country": "US", "position": "Researcher" }, { "name": "Nolan Simran Dey", "gender": "Male", "institution": "Cerebras Systems, Inc", "country": "CA", "position": "Researcher" }, { "name": "Shane Bergsma", "gender": "Male", "institution": "Cerebras Systems, Inc", "country": "CA", "position": "Principal Researcher" }, { "name": "Gavin Gray", "gender": "Female", "institution": "Cerebras Systems, Inc", "country": "", "position": "Researcher" }, { "name": "Joel Thomas Hestness", "gender": "Male", "institution": "Cerebras Systems, Inc", "country": "US", "position": "Research Scientist" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 61, "n_ref": 124, "n_ref_all": 223, "n_fig": 30, "n_tab": 6, "L_tab": 2845, "n_element_tab": 353, "n_fig_1": 23, "n_tab_1": 1, "L_tab_1": 282, "n_element_tab_1": 44, "formula_len_all": 715, "formula_len_all_1": 452, "len_all": 193013, "len_all_1": 66842, "len_abs": 1159, "len_title": 133, "len_sents": 55945, "len_sents_1": 31499, "n_sents": 431, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1164, "title": "Straight to Zero: Why Linearly Decaying the Learning Rate to Zero Works Best for LLMs", "abs": "LLMs are commonly trained with a learning rate (LR) warmup, followed by cosine decay to 10% of the maximum (10x decay). In a large-scale empirical study, we show that under an optimal max LR, a simple linear decay-to-zero (D2Z) schedule consistently outperforms other schedules when training at compute-optimal dataset sizes. Benefits increase further with more training tokens; e.g., a 617M-parameter model trained for 80 tokens-per-parameter (TPP) using D2Z achieves lower loss than when trained for 200 TPP using 10x decay, corresponding to an astonishing 60% FLOPs savings. This implies models like Llama2-7B, trained for 286 TPP with 10x decay, were severely under-decayed. We demonstrate the benefits of D2Z across a range of model sizes, batch sizes, and other training configurations. We explain the success of linear D2Z via a novel interpretation of AdamW as a convex combination of weight updates, with coefficients governed by the LR schedule. This interpretation demonstrates how linear D2Z balances the demands of early training (moving away quickly from initial conditions) and late training (smoothing over more updates to mitigate gradient noise).", "keywords": [ "Learning rate schedules", "Large language models (LLMs)", "AdamW optimizer", "Weight decay", "Compute-optimal training" ], "rating_list": [ 8, 5, 6 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "hrMNbdxcqL", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Hongyang Gao", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "Assistant Professor" }, { "name": "Xiangyang Xu", "gender": "Male", "institution": "Iowa State University", "country": "", "position": "PhD student" }, { "name": "ZHAONING YU", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 66, "n_ref_all": 76, "n_fig": 2, "n_tab": 8, "L_tab": 1857, "n_element_tab": 320, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1087, "n_element_tab_1": 186, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 108022, "len_all_1": 57513, "len_abs": 1208, "len_title": 146, "len_sents": 31971, "len_sents_1": 26971, "n_sents": 231, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 98, "L_abs": 1212, "title": "G2T-LLM: Graph-to-Tree Text Encoding for Molecule Generation with Fine-Tuned Large Language Models", "abs": "We introduce G2T-LLM, a novel approach for molecule generation that uses graph-to-tree text encoding to transform graph-based molecular structures into a hierarchical text format optimized for large language models (LLMs). This encoding converts complex molecular graphs into tree-structured formats, such as JSON and XML, which LLMs are particularly adept at processing due to their extensive pre-training on these types of data. By leveraging the flexibility of LLMs, our approach allows for intuitive interaction using natural language prompts, providing a more accessible interface for molecular design. Through supervised fine-tuning, G2T-LLM generates valid and coherent chemical structures, addressing common challenges like invalid outputs seen in traditional graph-based methods. While LLMs are computationally intensive, they offer superior generalization and adaptability, enabling the generation of diverse molecular structures with minimal task-specific customization. The proposed approach achieved comparable performances with state-of-the-art methods on various benchmark molecular generation datasets, demonstrating its potential as a flexible and innovative tool for AI-driven molecular design.", "keywords": [ "Molecule Generation", "Large Language Model", "Graph Generation", "Tree" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "hrLKzCETcf", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Daniel Kuelbs", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Mert Pilanci", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Sanjay Lall", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Full Professor" } ], "n_formula": 47, "n_formula_1": 21, "n_ref_uni": 21, "n_ref": 46, "n_ref_all": 80, "n_fig": 1, "n_tab": 3, "L_tab": 932, "n_element_tab": 54, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 607, "n_element_tab_1": 81, "formula_len_all": 5948, "formula_len_all_1": 2441, "len_all": 107753, "len_all_1": 55173, "len_abs": 1381, "len_title": 134, "len_sents": 32656, "len_sents_1": 22189, "n_sents": 260, "n_sents_1": 157, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1405, "title": "Adversarial Training of Two-Layer Polynomial and ReLU Activation Networks via Convex Optimization", "abs": "Training neural networks which are robust to adversarial attacks remains an important problem in deep learning, especially as heavily overparameterized models are adopted in safety-critical settings. Drawing from recent work which reformulates the training problems for two-layer ReLU and polynomial activation networks as convex programs, we devise a convex semidefinite program (SDP) for adversarial training of two-layer polynomial activation networks and prove that the convex SDP achieves the same globally optimal solution as its nonconvex counterpart. The convex adversarial SDP is observed to improve robust test accuracy against $\\ell_\\infty$\n attacks relative to the original convex training formulation on multiple datasets. Additionally, we present scalable implementations of adversarial training for two-layer polynomial and ReLU networks which are compatible with standard machine learning libraries and GPU acceleration. Leveraging these implementations, we retrain the final two fully connected layers of a Pre-Activation ResNet-18 model on the CIFAR-10 dataset with both polynomial and ReLU activations. The two `robustified' models achieve significantly higher robust test accuracies against $\\ell_\\infty$ attacks than a Pre-Activation ResNet-18 model trained with sharpness-aware minimization, demonstrating the practical utility of convex adversarial training on large-scale problems.", "keywords": [ "convex optimization", "adversarial training" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "hrGOMrfc2z", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Anca Dragan", "gender": "Female", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Jessica Lin", "gender": "Female", "institution": "University of California, Berkeley", "country": "", "position": "MS student" }, { "name": "Joey Hong", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 54, "n_ref": 101, "n_ref_all": 112, "n_fig": 6, "n_tab": 11, "L_tab": 5733, "n_element_tab": 167, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 5432, "n_element_tab_1": 133, "formula_len_all": 391, "formula_len_all_1": 597, "len_all": 216640, "len_all_1": 75321, "len_abs": 1685, "len_title": 82, "len_sents": 56818, "len_sents_1": 33307, "n_sents": 484, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1692, "title": "Interactive Dialogue Agents via Reinforcement Learning with Hindsight Regenerations", "abs": "Recent progress on large language models (LLMs) has enabled dialogue agents to generate highly naturalistic and plausible text. However, current LLM language generation focuses on responding accurately to questions and requests with a single effective response.\nIn reality, many real dialogues are interactive, meaning an agent's utterances will influence their conversational partner, elicit information, or change their opinion. Accounting for how an agent can effectively steer a conversation is a crucial ability in many dialogue tasks, from healthcare to preference elicitation. Existing methods for fine-tuning dialogue agents to accomplish such tasks would rely on curating some amount of expert data. However, doing so often requires understanding the underlying cognitive processes of the conversational partner, which is a skill neither humans nor LLMs trained on human data can reliably do. Our key insight is that while LLMs may not be adept at identifying effective strategies for steering conversations a priori, or in the middle of an ongoing conversation, they can do so post-hoc, or in hindsight, after seeing how their conversational partner responds. We use this fact to rewrite and augment existing suboptimal data, and train via offline reinforcement learning (RL) an agent that outperforms both prompting and learning from unaltered human demonstrations. We apply our approach to two domains that require understanding human mental state, intelligent interaction, and persuasion: mental health support, and soliciting charitable donations. Our results in a user study with real humans show that our approach greatly outperforms existing state-of-the-art dialogue agents.", "keywords": [ "offline reinforcement learning", "language models", "self-reflection" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "hqxzi4d3Ws", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Cho-Jui Hsieh", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Di Luo", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Lucas Matthew Tecot", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 50, "n_ref": 78, "n_ref_all": 97, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1370, "formula_len_all_1": 1342, "len_all": 147437, "len_all_1": 58909, "len_abs": 866, "len_title": 116, "len_sents": 36159, "len_sents_1": 28822, "n_sents": 262, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 870, "title": "Provably Noise-Resilient Training of Parameterized Quantum Circuits", "abs": "Advancements in quantum computing have spurred significant interest in harnessing its potential for speedups over classical systems. However, noise remains a major obstacle to achieving reliable quantum algorithms. In this work, we present a provably noise-resilient training theory and algorithm to enhance the robustness of parameterized quantum circuits. Our method, with a natural connection to Evolutionary Strategies, guarantees resilience to parameter noise with minimal adjustments to commonly used optimization algorithms. Our approach is function-agnostic and adaptable to various quantum circuits, successfully demonstrated in quantum phase classification and quantum state preparation tasks. By developing provably guaranteed learning theory with quantum circuits, our work opens new avenues for practical, robust applications of near-term quantum computers.", "keywords": [ "robustness", "quantum", "Parameterized Quantum Circuits", "Noise-Resilient" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "hpeyWG1PP6", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Defu Lian", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yi Yang", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Zhihao Zhu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 63, "n_ref": 133, "n_ref_all": 160, "n_fig": 4, "n_tab": 22, "L_tab": 32326, "n_element_tab": 2294, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 6838, "n_element_tab_1": 669, "formula_len_all": 37, "formula_len_all_1": 0, "len_all": 267293, "len_all_1": 70546, "len_abs": 3281, "len_title": 97, "len_sents": 61002, "len_sents_1": 27657, "n_sents": 421, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1436, "title": "TDDBench: A Benchmark for Training data detection", "abs": "Training Data Detection (TDD) is a task aimed at determining whether a specific data instance is used to train a machine learning model. In the computer security literature, TDD is also referred to as Membership Inference Attack (MIA). Given its potential to assess the risks of training data breaches, ensure copyright authentication, and verify model unlearning, TDD has garnered significant attention in recent years, leading to the development of numerous methods. Despite these advancements, there is no comprehensive benchmark to thoroughly evaluate the effectiveness of TDD methods.\nIn this work, we introduce TDDBench, which consists of 13 datasets spanning three data modalities: image, tabular, and text. We benchmark 21 different TDD methods across four detection paradigms and evaluate their performance from five perspectives: average detection performance, best detection performance, memory consumption, and computational efficiency in both time and memory. With TDDBench, researchers can identify bottlenecks and areas for improvement in TDD algorithms, while practitioners can make informed trade-offs between effectiveness and efficiency when selecting TDD algorithms for specific use cases. Our large-scale benchmarking also reveals the generally unsatisfactory performance of TDD algorithms across different datasets. To enhance accessibility and reproducibility, we open-source TDDBench for the research community.", "keywords": [ "Training data detection; Benchmark; Copyright certification" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "hpZ5zpudH8", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Changsong Liu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "LiMingyang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "tjuzhangwei", "gender": "Male", "institution": "Tianjin University", "country": "", "position": "Full Professor" }, { "name": "Fan Yimeng", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "MS student" } ], "n_formula": 29, "n_formula_1": 11, "n_ref_uni": 32, "n_ref": 69, "n_ref_all": 88, "n_fig": 6, "n_tab": 9, "L_tab": 4058, "n_element_tab": 533, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1241, "n_element_tab_1": 172, "formula_len_all": 1786, "formula_len_all_1": 741, "len_all": 157347, "len_all_1": 51873, "len_abs": 1323, "len_title": 136, "len_sents": 43216, "len_sents_1": 22286, "n_sents": 354, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1306, "title": "SpikSSD: Better Extraction and Fusion for Object Detection with Spiking Neuron Networks", "abs": "As the third generation of neural networks, Spiking Neural Networks (SNNs) have gained widespread attention due to their low energy consumption and biological interpretability. Recently, SNNs have made considerable advancements in computer vision. However, efficiently conducting feature extraction and fusion under the spiking characteristics of SNNs for object detection remains a pressing challenge. To address this problem, we propose the SpikSSD, a novel Spiking Single Shot Multibox Detector. Specifically, we design a full-spiking backbone network, MDS-ResNet, which effectively adjusts the membrane synaptic input distribution at each layer, achieving better spiking feature extraction. Additionally, for spiking feature fusion, we introduce the Spiking Bi-direction Fusion Module (SBFM), which for the first time realizes bi-direction fusion of spiking features, enhancing the multi-scale detection capability of the model. Experimental results show that SpikSSD achieves 40.8\\% mAP on the GEN1 dataset and 76.0\\% mAP@0.5 on the VOC 2007 dataset with only around 10\\% firing rate, outperforming existing SNN-based approaches at ultralow energy consumption. This work sets a new benchmark for future research in SNN-based object detection. Our code is publicly available in supplementary materials.", "keywords": [ "Spiking Neural Networks", "Object Detection", "Membrane-based Deformed Shortcut", "Spiking Bi-direction Fusion Module" ], "rating_list": [ 3, 3, 6, 6, 3 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 2, 1, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 3, 4, 4 ] }, { "paper_id": "hpDiwfGrrX", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Ao Ke", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Chuanwen Feng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "S Kevin Zhou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Wenlong Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Xike Xie", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Research Professor" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 32, "n_ref": 68, "n_ref_all": 85, "n_fig": 6, "n_tab": 7, "L_tab": 1989, "n_element_tab": 333, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 2514, "n_element_tab_1": 327, "formula_len_all": 994, "formula_len_all_1": 818, "len_all": 116026, "len_all_1": 65581, "len_abs": 1391, "len_title": 137, "len_sents": 34193, "len_sents_1": 27318, "n_sents": 269, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1395, "title": "PATTERN MATCHING-BASED OUT-OF-DISTRIBUTION DETECTION FOR MULTI-LABEL NODE CLASSIFICATION", "abs": "Graph neural networks (GNNs) have achieved dominant performance in various prediction tasks on graphs. When deploying GNNs in the real world, estimating the possibility of out-of-distribution (OOD) testing samples becomes a crucial safety concern. Although some research has investigated the graph OOD detection problem, most have concentrated on single-label classification scenarios, aspecific case of the more general multi-label classification, which has broader applications, such as in social networks where nodes can represent users with multiple interests or attributes. In this paper, we first introduce and define the multi-label graph OOD detection problem and propose a simple yet effective pattern matching-based OOD detection method to address it. In particular, our method utilizes feature pattern matching and label pattern matching to obtain two matching scores. By incorporating topological structure adjustment, we ultimately derive confidence scores, serving as indicators of the likelihood that a test sample is an OOD instances. We conduct extensive comparisons with existing OOD detection methods in the context of multi-label graphs. The results show that our method achieves an impressive 7.61% reduction in FPR95 compared to the leading baselines, setting a new state-of-the-art. Furthermore, our approach can servas a benchmark for OOD detection on multi-label graphs.", "keywords": [ "Out-of-distribution" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "hpCfPEvBsr", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Bo Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Fuzhao Xue", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Jinjie Ni", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Kabir Jain", "gender": "Not Specified", "institution": "National University of Singapore", "country": "", "position": "Intern" }, { "name": "Kaichen Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Undergrad student" }, { "name": "Mahir Shah", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Intern" }, { "name": "Xiang Yue", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" }, { "name": "Yifan Song", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yuntian Deng", "gender": "unknown", "institution": "University of Waterloo", "country": "", "position": "Assistant Professor" }, { "name": "Zian Zheng", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "MS student" }, { "name": "Junhao Zhang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Deepanway Ghosal", "gender": "unknown", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Qizhe Xie", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 4, "n_ref_uni": 42, "n_ref": 120, "n_ref_all": 150, "n_fig": 12, "n_tab": 3, "L_tab": 971, "n_element_tab": 27, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1531, "n_element_tab_1": 136, "formula_len_all": 3080, "formula_len_all_1": 1656, "len_all": 390900, "len_all_1": 50015, "len_abs": 0, "len_title": 47, "len_sents": 118350, "len_sents_1": 21749, "n_sents": 902, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 62, "L_abs": 1235, "title": "MixEval-X: Any-to-any Evaluations from Real-world Data Mixture", "abs": "Perceiving and generating diverse modalities are crucial for AI models to effectively learn from and engage with real-world signals, necessitating reliable evaluations for their development. We identify two major issues in current evaluations: (1) inconsistent standards, shaped by different communities with varying protocols and maturity levels; and (2) significant query, grading, and generalization biases. To address these, we introduce MixEval-X, the first any-to-any, real-world benchmark designed to optimize and standardize evaluations across diverse input and output modalities. We propose multi-modal benchmark mixture and adaptation-rectification pipelines to reconstruct real-world task distributions, ensuring evaluations generalize effectively to real-world use cases. Extensive meta-evaluations show our approach effectively aligns benchmark samples with real-world task distributions. Meanwhile, MixEval-X's model rankings correlate strongly with that of crowd-sourced real-world evaluations (up to 0.98) while being much more efficient. We provide comprehensive leaderboards to rerank existing models and organizations and offer insights to enhance understanding of multi-modal evaluations and inform future research.", "keywords": [ "Evaluation", "Multi-modal Evaluation", "Benchmark", "Multi-modal Benchmark", "Any-to-any", "MixEval", "Real-world", "Data Mixture", "Artificial General Intelligence", "AGI" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 4, 3, 4 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 4, 3, 4 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "hovDbX4Gh6", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "George Karypis", "gender": "Male", "institution": "University of Minnesota, Minneapolis", "country": "", "position": "Full Professor" }, { "name": "Han Xie", "gender": "unknown", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Huzefa Rangwala", "gender": "Male", "institution": "Amazon", "country": "", "position": "Principal Researcher" }, { "name": "Jian Zhang", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Jiliang Tang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Full Professor" }, { "name": "Xiang song", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Zhikai Chen", "gender": "Male", "institution": "Michigan State University", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 52, "n_ref": 114, "n_ref_all": 128, "n_fig": 4, "n_tab": 16, "L_tab": 4310, "n_element_tab": 529, "n_fig_1": 1, "n_tab_1": 9, "L_tab_1": 1464, "n_element_tab_1": 210, "formula_len_all": 278, "formula_len_all_1": 158, "len_all": 197023, "len_all_1": 73794, "len_abs": 2298, "len_title": 0, "len_sents": 47347, "len_sents_1": 33497, "n_sents": 387, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1407, "title": "AutoG: Towards automatic graph construction from tabular data", "abs": "Recent years have witnessed significant advancements in graph machine learning, with its applications spanning numerous domains. However, the focus has predominantly been on developing powerful models, often overlooking a crucial initial step: constructing suitable graphs from common data formats, such as tabular data.\nThis construction process is fundamental to applying graph-based models, yet it remains largely understudied and lacks formalization.\nOur research aims to address this gap by formalizing the graph construction problem and proposing an effective solution. We identify two critical challenges to achieve this goal: 1. The absence of dedicated benchmarks to formalize and evaluate the effectiveness of graph construction methods, and 2. Existing automatic construction methods can only be applied to some specific cases, while tedious human engineering is required to generate high-quality schemas. \nTo tackle these challenges, we present a two-fold contribution.\nFirst, we introduce a benchmark to formalize and evaluate graph construction methods. \nSecond, we propose an LLM-based solution, AutoG, automatically generating high-quality graph structures without human intervention.\nThe experimental results demonstrate that the quality of constructed graphs is critical to downstream task performance, and AutoG can generate high-quality graphs that rival those produced by human experts.", "keywords": [ "Graph machine learning", "Automatic data science", "Applications of large language models", "Tabular data" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "hom2oeHCnz", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Jingyi Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Kui Ren", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Peng Kuang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zhibo Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zhixuan Chu", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" } ], "n_formula": 33, "n_formula_1": 8, "n_ref_uni": 47, "n_ref": 183, "n_ref_all": 222, "n_fig": 10, "n_tab": 7, "L_tab": 3561, "n_element_tab": 398, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 898, "n_element_tab_1": 115, "formula_len_all": 2254, "formula_len_all_1": 361, "len_all": 218099, "len_all_1": 73366, "len_abs": 1641, "len_title": 125, "len_sents": 64034, "len_sents_1": 33124, "n_sents": 484, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1523, "title": "Towards Real World Debiasing: A Fine-grained Analysis On Spurious Correlation", "abs": "Spurious correlations in training data significantly hinder the generalization capability of machine learning models when faced with distribution shifts in real-world scenarios. To tackle the problem, numerous debias approaches have been proposed and benchmarked on datasets intentionally designed with severe biases. However, it remains to be asked: \\textit{1. Do existing benchmarks really capture biases in the real world? 2. Can existing debias methods handle biases in the real world?} To answer the questions, we revisit biased distributions in existing benchmarks and real-world datasets, and propose a fine-grained framework for analyzing dataset bias by disentangling it into the magnitude and prevalence of bias. We observe and theoretically demonstrate that existing benchmarks poorly represent real-world biases. We further introduce two novel biased distributions to bridge this gap, forming a nuanced evaluation framework for real-world debiasing. Building upon these results, we evaluate existing debias methods with our evaluation framework. Results show that existing methods are incapable of handling real-world biases. Through in-depth analysis, we propose a simple yet effective approach that can be easily applied to existing debias methods, named Debias in Destruction (DiD). Empirical results on real-world datasets in both image and language modalities demonstrate the superiority of DiD, improving the performance of existing methods on all types of biases within the proposed evaluation framework.", "keywords": [ "spurious correlation", "dataset bias", "debias" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "hoYFLRNbhc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Derek F. Wong", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Associate Professor" }, { "name": "Fandong Meng", "gender": "Male", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Jiali Zeng", "gender": "Female", "institution": "Tencent WeChat AI", "country": "CN", "position": "Researcher" }, { "name": "Jie Zhou", "gender": "Male", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Min Zhang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Xuebo Liu", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Yutong Wang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 4, "n_ref_uni": 32, "n_ref": 64, "n_ref_all": 105, "n_fig": 8, "n_tab": 13, "L_tab": 7180, "n_element_tab": 286, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1367, "n_element_tab_1": 71, "formula_len_all": 1187, "formula_len_all_1": 626, "len_all": 193620, "len_all_1": 65762, "len_abs": 1451, "len_title": 126, "len_sents": 47281, "len_sents_1": 29713, "n_sents": 354, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1457, "title": "DelTA: An Online Document-Level Translation Agent Based on Multi-Level Memory", "abs": "Large language models (LLMs) have achieved reasonable quality improvements in machine translation (MT).\nHowever, most current research on MT-LLMs still faces significant challenges in maintaining translation consistency and accuracy when processing entire documents.\nIn this paper, we introduce DelTA, a Document-levEL Translation Agent designed to overcome these limitations.\nDelTA features a multi-level memory structure that stores information across various granularities and spans, including Proper Noun Records, Bilingual Summary, Long-Term Memory, and Short-Term Memory, which are continuously retrieved and updated by auxiliary LLM-based components.\nExperimental results indicate that DelTA significantly outperforms strong baselines in terms of translation consistency and quality across four open/closed-source LLMs and two representative document translation datasets, achieving an increase in consistency scores by up to 4.58 percentage points and in COMET scores by up to 3.16 points on average.\nDelTA employs a sentence-by-sentence translation strategy, ensuring no sentence omissions and offering a memory-efficient solution compared to the mainstream method.\nFurthermore, DelTA improves pronoun and context-dependent translation accuracy, and the summary component of the agent also shows promise as a tool for query-based summarization tasks.\nThe code and data is anonymously available at https://anonymous.4open.science/r/DelTA_Agent-7716.", "keywords": [ "Document-Level Translation", "Large Language Models", "Autonomous Agents", "Natural Language Processing" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "hoEanaoP4i", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Chao Zhang", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Wensheng Li", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Xingwu Liu", "gender": "Male", "institution": "Dalian University of Technology", "country": "", "position": "Full Professor" }, { "name": "Xinyu Chen", "gender": "Male", "institution": "Dalian University of Technology", "country": "", "position": "PhD student" } ], "n_formula": 47, "n_formula_1": 14, "n_ref_uni": 37, "n_ref": 49, "n_ref_all": 79, "n_fig": 12, "n_tab": 8, "L_tab": 4324, "n_element_tab": 433, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1952, "n_element_tab_1": 195, "formula_len_all": 6870, "formula_len_all_1": 1294, "len_all": 163858, "len_all_1": 72515, "len_abs": 1468, "len_title": 160, "len_sents": 51945, "len_sents_1": 33392, "n_sents": 433, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1473, "title": "MD-LSM: An Efficient Tool for Real-time Monitoring Linear Separability of Hidden-layer Outputs of Deep Networks", "abs": "Many studies have shown that evaluating the linear separability of hidden-layer outputs plays a key role in understanding the working mechanism of deep networks. However, it is still challenging to develop the linear separability measure (LSM) that satisfies all of the following requirements: 1) it should be an absolute measure; 2) it should be insensitive to the outliers; and 3) its computational cost should be low for real-time monitoring the behavior of each hidden layer. In this paper, we propose the Minkowski difference-based linear separability measures (MD-LSMs) that just meet the first two requirements. Moreover, we also introduce an approximate calculation method to significantly decrease their computation costs with only a slight precision sacrifice. As an application example, we conduct the experiments on the real-time monitoring for the hidden-layer behaviors of several popular deep networks, and show that the outputs of the hidden layers adjacent to the output layer have higher linear separability degrees. We also observe that the change of linear separability degree of hidden layers (especially the ones are adjacent to the output layers) are in sync with the change of the training accuracy of the entire network. It implies that the linear separability of some important hidden layers can be treated as a performance criterion to characterize the network's training behavior. The relevant theoretical discussion also validates this finding.", "keywords": [ "linear separability", "deep network", "hidden layer", "Minkowski difference" ], "rating_list": [ 3, 5, 8, 8, 6 ], "soundness_list": [ 2, 2, 4, 4, 3 ], "presentation_list": [ 3, 3, 4, 3, 4 ], "contribution_list": [ 1, 3, 4, 3, 3 ], "confidence_list": [ 3, 5, 4, 3, 3 ] }, { "paper_id": "ho7ZUS1z8A", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Lujun Li", "gender": "Male", "institution": "HKGAI", "country": "HK", "position": "Researcher" }, { "name": "Mark G. Lee", "gender": "unknown", "institution": "University of Birmingham", "country": "", "position": "Full Professor" }, { "name": "Shengjie Sun", "gender": "Male", "institution": "Nanjing University of Information Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Wei Li", "gender": "Male", "institution": "University of Birmingham", "country": "GB", "position": "PhD student" }, { "name": "Wei Xue", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Yike Guo", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "You-Liang Huang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 42, "n_ref": 218, "n_ref_all": 248, "n_fig": 8, "n_tab": 10, "L_tab": 7314, "n_element_tab": 1089, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 933, "n_element_tab_1": 121, "formula_len_all": 901, "formula_len_all_1": 625, "len_all": 195723, "len_all_1": 67605, "len_abs": 1776, "len_title": 128, "len_sents": 52457, "len_sents_1": 30618, "n_sents": 387, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1695, "title": "Structured Mixture-of-Experts LLMs Compression via Singular Value Decomposition", "abs": "Mixture of Experts (MoE) architecture has emerged as a powerful paradigm in the development of Large Language Models (LLMs), offering superior scaling capabilities and reduced computational costs. However, the increased parameter budgets and memory overhead associated with MoE LLMs pose significant challenges to their efficiency and widespread deployment. In this paper, we present MoE-SVD, the first decomposition-based compression framework tailored for MoE LLMs without any extra training. By harnessing the power of Singular Value Decomposition (SVD), MoE-SVD addresses the critical issues of decomposition collapse and matrix redundancy in MoE architectures. Specifically, we first decompose experts into compact low-rank matrices, resulting in accelerated inference and memory optimization. In particular, we propose selective decomposition strategy by measuring sensitivity metrics based on weight singular values and activation statistics to automatically identify decomposable expert layers. Then, we share a single V-matrix across all experts and employ a top-k selection for U-matrices. This low-rank matrix sharing and trimming scheme allows for significant parameter reduction while preserving diversity among experts. Comprehensive experiments conducted on Mixtral-8\u00d77B|22B, Phi-3.5-MoE and DeepSeekMoE across multiple datasets reveal that MoE-SVD consistently outperforms existing compression methods in terms of performance-efficiency tradeoffs. Notably, we achieve a remarkable 60\\% compression ratio on Mixtral-7x8B and Phi-3.5-MoE, resulting in a 1.5$\\times$ inference acceleration with minimal performance degradation. Codes are available in the supplementary materials.", "keywords": [ "Mixture of Experts", "Efficient Large Language Models", "Low-Rank Decomposition", "Network Sparsity" ], "rating_list": [ 6, 6, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "ho4mNiwr2n", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Anil Kumar Vullikanti", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "Professor" }, { "name": "Jielu Zhang", "gender": "Female", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Junfeng Guo", "gender": "Male", "institution": "University of Maryland Institute for Advanced Computer Studies, University of Maryland, College Park", "country": "US", "position": "Postdoc" }, { "name": "Mengxuan Hu", "gender": "Female", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Ruoxi Jia", "gender": "unknown", "institution": "Virginia Tech", "country": "US", "position": "Assistant Professor" }, { "name": "Sheng Li", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Associate Professor" }, { "name": "Yi Zeng", "gender": "Male", "institution": "Virginia Tech", "country": "", "position": "PhD student" }, { "name": "Zhongliang Zhou", "gender": "unknown", "institution": "MERCK & CO., INC.", "country": "US", "position": "Researcher" }, { "name": "zihan guan", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 6, "n_ref_uni": 63, "n_ref": 158, "n_ref_all": 211, "n_fig": 16, "n_tab": 13, "L_tab": 2335, "n_element_tab": 303, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 859, "n_element_tab_1": 80, "formula_len_all": 1748, "formula_len_all_1": 428, "len_all": 228880, "len_all_1": 71031, "len_abs": 1494, "len_title": 130, "len_sents": 68562, "len_sents_1": 33678, "n_sents": 542, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1428, "title": "Mind Control through Causal Inference: Predicting Clean Images from Poisoned Data", "abs": "Anti-backdoor learning, aiming to train clean models directly from poisoned datasets, serves as an important defense method for backdoor attack. However, existing methods usually fail to recover backdoored samples to their original, correct labels and suffer from poor generalization to large pre-trained models due to its non end-to end training, making them unsuitable for protecting the increasingly prevalent large pre-trained models. To bridge the gap, we first revisit the anti-backdoor learning problem from a causal perspective. Our theoretical causal analysis reveals that incorporating \\emph{\\textbf{both}} images and the associated attack indicators preserves the model's integrity. Building on the theoretical analysis, we introduce an end-to-end method, Mind Control through Causal Inference (MCCI), to train clean models directly from poisoned datasets. This approach leverages both the image and the attack indicator to train the model. Based on this training paradigm, the model\u2019s perception of whether an input is clean or backdoored can be controlled. Typically, by introducing fake non-attack indicators, the model perceives all inputs as clean and makes correct predictions, even for poisoned samples. Extensive experiments demonstrate that our method achieves state-of-the-art performance, efficiently recovering the original correct predictions for poisoned samples and enhancing accuracy on clean samples.", "keywords": [ "Causal Inference", "Backdoor Attacks" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "hnsiuIcRT7", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Jun Sun", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Full Professor" }, { "name": "Wei Zhao", "gender": "Male", "institution": "Singapore Management University", "country": "", "position": "PhD student" }, { "name": "Ge Yi Li", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Postdoc" }, { "name": "Zhe Li", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 47, "n_ref": 75, "n_ref_all": 96, "n_fig": 7, "n_tab": 10, "L_tab": 4470, "n_element_tab": 301, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1823, "n_element_tab_1": 175, "formula_len_all": 324, "formula_len_all_1": 324, "len_all": 139277, "len_all_1": 62944, "len_abs": 1069, "len_title": 102, "len_sents": 36157, "len_sents_1": 27434, "n_sents": 280, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1073, "title": "Do Influence Functions Work on Large Language Models?", "abs": "Influence functions aim to quantify the impact of individual training data points on a model's predictions. While extensive research has been conducted on influence functions in traditional machine learning models, their application to large language models (LLMs) has been limited. In this work, we conduct a systematic study to address a key question: do influence functions work on LLMs? Specifically, we evaluate influence functions across multiple tasks and find that they consistently perform poorly in most settings. Our further investigation reveals that their poor performance can be attributed to: (1) inevitable approximation errors when estimating the iHVP component due to the scale of LLMs, (2) uncertain convergence during fine-tuning, and, more fundamentally, (3) the definition itself, as changes in model parameters do not necessarily correlate with changes in LLM behavior. Our study thus suggests the need for alternative approaches for identifying influential samples. To support future work, our code is made available at https://github.com/anonymous.", "keywords": [ "Large language models", "Influence function" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "hnpXIEaOrS", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Guangyu Jiang", "gender": "Male", "institution": "George Washington University", "country": "US", "position": "PhD student" }, { "name": "Nathaniel D. Bastian", "gender": "Male", "institution": "United States Military Academy", "country": "US", "position": "Principal Researcher" }, { "name": "Tian Lan", "gender": "Male", "institution": "George Washington University", "country": "", "position": "Full Professor" }, { "name": "Mahdi Imani", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 40, "n_formula_1": 36, "n_ref_uni": 23, "n_ref": 46, "n_ref_all": 55, "n_fig": 2, "n_tab": 1, "L_tab": 103, "n_element_tab": 4, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 819, "n_element_tab_1": 79, "formula_len_all": 3090, "formula_len_all_1": 2761, "len_all": 95718, "len_all_1": 68632, "len_abs": 1565, "len_title": 136, "len_sents": 33038, "len_sents_1": 28867, "n_sents": 232, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1572, "title": "Federated Maximum Likelihood Inverse Reinforcement Learning with Convergence Guarantee", "abs": "Inverse Reinforcement Learning (IRL) aims to recover the latent reward function and corresponding optimal policy from observed demonstrations. Existing IRL research predominantly focuses on a centralized learning approach, not suitable for real-world problems with distributed data and privacy restrictions. To this end, this paper proposes a novel algorithm for federated maximum-likelihood IRL (F-ML-IRL) and provides a rigorous analysis of its convergence and time-complexity. The proposed F-ML-IRL leverages a dual-aggregation to update the shared global model and performs bi-level local updates -- an upper-level learning task to optimize the parameterized reward function by maximizing the discounted likelihood of observing expert trajectories under the current policy and a low-level learning task to find the optimal policy concerning the entropy-regularized discounted cumulative reward under the current reward function. We analyze the convergence and time-complexity of the proposed F-ML-IRL algorithm and show that the global model in F-ML-IRL converges to a stationary point for both the reward and policy parameters within finite time, i.e., the log-distance between the recovered policy and the optimal policy, as well as the gradient of the likelihood objective, converge to zero. Finally, evaluating our F-ML-IRL algorithm on high-dimensional robotic control tasks in MuJoCo, we show that it ensures convergences of the recovered reward in decentralized learning and even outperforms centralized baselines due to its ability to utilize distributed data.", "keywords": [ "Inverse Reinforcement Learning", "Decentralized learning." ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "hmvCXDzPUR", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Pieter Abbeel", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Jakub Grudzien", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 32, "n_ref": 77, "n_ref_all": 97, "n_fig": 8, "n_tab": 5, "L_tab": 1399, "n_element_tab": 199, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 973, "n_element_tab_1": 112, "formula_len_all": 827, "formula_len_all_1": 728, "len_all": 134055, "len_all_1": 71987, "len_abs": 1841, "len_title": 115, "len_sents": 46935, "len_sents_1": 34398, "n_sents": 348, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1875, "title": "Cliqueformer: Model-Based Optimization With Structured Transformers", "abs": "Expressive large-scale neural networks enable training powerful models for prediction tasks. However, in many engineering and science domains, such models are intended to be used not just for prediction, but for design---e.g., creating new proteins that serve as effective therapeutics, or creating new materials or chemicals that maximize a downstream performance measure. Thus, researchers have been recently growing an interest in building deep learning methods that solve offline \\emph{model-based optimization} (MBO) problems, in which design candidates are optimized with respect to surrogate models learned from offline data. However, straightforward application of predictive models that are effective at predicting in-distribution properties of a design are not necessarily the best suited for use in creating new designs. Thus, the most successful algorithms that tackle MBO draw intpiration from reinforcement learning and generative modeling to meet the in-distribution constratints. Meanwhile, recent theoretical works have observed that exploiting structure of the target black-box function is an effective strategy for solving MBO from offline data. Unfortunately, discovering such structure remains an open problem. In this paper, following first principles, we develop a model that learns the structure of an MBO task and empirically leads to improved designs. To this end, we introduce \\emph{Cliqueformer}---a scalable transformer-based architecture that learns the black-box function's structure in form of its \\emph{functional graphical model} (FGM), thus bypassing the problem of distribution shift, previously tackled by conservative approaches. We evaluate Cliqueformer on various tasks, ranging from high-dimensional black-box functions from MBO literature, to real-world tasks of chemical and genetic design, consistently outperforming the baselines.", "keywords": [ "model-based optimization; black-box optimization; transformers" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "hmXUWc1ugd", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Alain Raymond", "gender": "Male", "institution": "Pontificia Universidad Catolica de Chile", "country": "CL", "position": "PhD student" }, { "name": "Alvaro Soto", "gender": "Male", "institution": "Universidad Cat\u00f3lica de Chile", "country": "CL", "position": "Associate Professor" }, { "name": "Julio Hurtado", "gender": "Male", "institution": "University of Warwick", "country": "GB", "position": "Postdoc" }, { "name": "Nico Alvarado", "gender": "Male", "institution": "Pontificia Universidad Catolica de Chile", "country": "CL", "position": "PhD student" }, { "name": "Vincenzo Lomonaco", "gender": "Male", "institution": "University of Pisa", "country": "", "position": "Assistant Professor" } ], "n_formula": 21, "n_formula_1": 15, "n_ref_uni": 27, "n_ref": 56, "n_ref_all": 70, "n_fig": 3, "n_tab": 8, "L_tab": 6488, "n_element_tab": 301, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1498, "n_element_tab_1": 119, "formula_len_all": 1066, "formula_len_all_1": 783, "len_all": 130026, "len_all_1": 62740, "len_abs": 1581, "len_title": 103, "len_sents": 39860, "len_sents_1": 28482, "n_sents": 329, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1588, "title": "Towards Understanding Why Group Robustness Methods Work", "abs": "Deep Learning has made remarkable strides, yet models trained under conventional Empirical Risk Minimization (ERM) approaches encounter challenges regarding their generalization capabilities. In particular, a lack of robustness to spurious correlations. In response, Group Robustness Methods (GRMs) have been developed to combat them. These methods partition training datasets into distinct groups based on spurious features and class labels and adjust their weighting in the loss function. These methods show remarkable performance in dealing with spurious correlations. The underlying mechanisms for their success, however, are not so well understood. Our work contributes by shedding light on the learning dynamics of GRMs, through an empirical and theoretical analysis of them that reveals the differences in feature learning and the type of classifiers they learn versus ERM. Surprisingly, both GRMs and ERM models retain spurious information in their representations, even when it is irrelevant to the task at hand. We find evidence that suggests that the key to GRMs' success is two-fold: distributing prediction across multiple features in representation space to avoid relying on few but spurious attributes and incentivizing the classifier to become orthogonal to spurious features. We verify our findings by proposing an upgrade to the Subsampling baseline method called Group Distributionally Robust Feature Reweighting (GDRFR) that is easy to compute and only requires a fraction of group labels during a finetuning phase and retrieve most of GRMs performance gains over ERM.", "keywords": [ "robustness fairness deep learning" ], "rating_list": [ 3, 6, 3, 3, 3 ], "soundness_list": [ 3, 3, 1, 2, 2 ], "presentation_list": [ 2, 2, 1, 3, 2 ], "contribution_list": [ 2, 3, 1, 2, 1 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "hmDt068MoZ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Ali Payani", "gender": "Male", "institution": "Cisco", "country": "US", "position": "Researcher" }, { "name": "Baixiang Huang", "gender": "Male", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Canyu Chen", "gender": "unknown", "institution": "Illinois Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Kai Shu", "gender": "unknown", "institution": "Emory University", "country": "US", "position": "Assistant Professor" }, { "name": "Xiongxiao Xu", "gender": "unknown", "institution": "Illinois Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 55, "n_ref": 155, "n_ref_all": 174, "n_fig": 5, "n_tab": 3, "L_tab": 433, "n_element_tab": 42, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 272, "n_element_tab_1": 31, "formula_len_all": 47, "formula_len_all_1": 47, "len_all": 216245, "len_all_1": 51527, "len_abs": 1660, "len_title": 100, "len_sents": 47343, "len_sents_1": 26134, "n_sents": 311, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1678, "title": "Can Knowledge Editing Really Correct Hallucinations?", "abs": "Large Language Models (LLMs) suffer from hallucinations, referring to the non-factual information in generated content, despite their superior capacities across tasks. Meanwhile, knowledge editing has been developed as a new popular paradigm to correct the erroneous factual knowledge encoded in LLMs with the advantage of avoiding retraining from scratch. However, one common issue of existing evaluation datasets for knowledge editing is that they do not ensure LLMs actually generate hallucinated answers to the evaluation questions before editing. When LLMs are evaluated on such datasets after being edited by different techniques, it is hard to directly adopt the performance to assess the effectiveness of different knowledge editing methods in correcting hallucinations. Thus, the fundamental question remains insufficiently validated: Can knowledge editing really correct hallucinations in LLMs? We proposed HalluEditBench to holistically benchmark knowledge editing methods in correcting real-world hallucinations. First, we rigorously construct a massive hallucination dataset with 9 domains, 26 topics and more than 6,000 hallucinations. Then, we assess the performance of knowledge editing methods in a holistic way on five dimensions including Efficacy, Generalization, Portability, Locality, and Robustness. Through HalluEditBench, we have provided new insights into the potentials and limitations of different knowledge editing methods in correcting hallucinations, which could inspire future improvements and facilitate the progress in the field of knowledge editing. Data, code, and results are available at https://anonymous.4open.science/r/hallucination-9D6.", "keywords": [ "LLMs", "Knowledge Editing", "Hallucination", "Benchmark" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "hlvLM3GX8R", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Andrei Lupu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Ani Calinescu", "gender": "Not Specified", "institution": "Department of Computer Science, University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Benjamin Ellis", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "", "position": "PhD student" }, { "name": "Jakob Nicolaus Foerster", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Tin Dizdarevic", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "MS student" }, { "name": "Tobias Gessler", "gender": "unknown", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 38, "n_ref": 77, "n_ref_all": 93, "n_fig": 10, "n_tab": 11, "L_tab": 2874, "n_element_tab": 403, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2814, "n_element_tab_1": 83, "formula_len_all": 45, "formula_len_all_1": 231, "len_all": 210240, "len_all_1": 74229, "len_abs": 1702, "len_title": 110, "len_sents": 72015, "len_sents_1": 34248, "n_sents": 557, "n_sents_1": 281, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1711, "title": "OvercookedV2: Rethinking Overcooked for Zero-Shot Coordination", "abs": "AI agents hold the potential to transform everyday life by helping humans achieve their goals.\nTo do this successfully, agents need to be able to coordinate with novel partners without prior interaction, a setting known as zero-shot coordination (ZSC).\nOvercooked has become one of the most popular benchmarks for evaluating coordination capabilities of AI agents and learning algorithms.\nIn this work, we investigate the origins of ZSC challenges in Overcooked.\nWe introduce a state augmentation mechanism which mixes states that might be encountered when paired with unknown partners into the training distribution, reducing the out-of-distribution challenge associated with ZSC.\nWe show that independently trained agents under this algorithm coordinate successfully in Overcooked.\nOur results suggest that ZSC failure can largely be attributed to poor state coverage under self-play rather than more sophisticated coordination challenges. The Overcooked environment is therefore not suitable as a ZSC benchmark.\nTo address these shortcomings, we introduce OvercookedV2, a new version of the benchmark, which includes asymmetric information and stochasticity, facilitating the creation of interesting ZSC scenarios.\nTo validate OvercookedV2, we conduct experiments demonstrating that mere exhaustive state coverage is insufficient to coordinate well. Finally, we use OvercookedV2 to build a new range of coordination challenges, including ones that require test time protocol formation, and we demonstrate the need for new coordination algorithms that can adapt online.\nWe hope that OvercookedV2 will help benchmark the next generation of ZSC algorithms and advance collaboration between AI agents and humans.", "keywords": [ "multi-agent reinforcement learning", "reinforcement learning", "multi-agent systems", "zero-shot coordination", "overcooked", "human-AI coordination" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 1 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "hllDiA56TX", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Akilesh Potti", "gender": "Male", "institution": "Ritual Research", "country": "US", "position": "Researcher" }, { "name": "Arka Pal", "gender": "unknown", "institution": "Ritual AI", "country": "US", "position": "Researcher" }, { "name": "Eva Zhang", "gender": "unknown", "institution": "Ritual", "country": "US", "position": "Researcher" }, { "name": "Micah Goldblum", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 45, "n_ref": 88, "n_ref_all": 107, "n_fig": 3, "n_tab": 4, "L_tab": 454, "n_element_tab": 54, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 559, "n_element_tab_1": 76, "formula_len_all": 239, "formula_len_all_1": 44, "len_all": 275626, "len_all_1": 63329, "len_abs": 2069, "len_title": 51, "len_sents": 62490, "len_sents_1": 31758, "n_sents": 427, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1254, "title": "vTune: Verifiable Fine-Tuning for LLMs Through Backdooring", "abs": "As fine-tuning large language models (LLMs) becomes increasingly prevalent, users often rely on third-party services with limited visibility into their fine-tuning processes. This lack of transparency raises the question: _how do consumers verify that\nfine-tuning services are performed correctly_? For instance, a service provider could claim to fine-tune a model for each user, yet simply send all users back the same base model. To address this issue, we propose vTune, a simple method that uses a small number of \\textit{backdoor} data points added to the training data to provide a statistical test for verifying that a provider fine-tuned a custom model on a particular user's dataset. Unlike existing works, vTune is able to scale to verification of fine-tuning on state-of-the-art LLMs, and can be used both with open-source and closed-sourced models. We test our approach across several model families and sizes as well as across multiple instruction-tuning datasets, and find that the statistical test is satisfied with p-values on the order of $\\sim 10e^{-40}$, with no negative impact on downstream task performance. Further, we explore several attacks that attempt to subvert vTune and demonstrate the method's robustness to these attacks.", "keywords": [ "backdooring", "fine-tuning", "statistical measure", "llms" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "hlijRgXTDK", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Andrew Gordon Wilson", "gender": "Not Specified", "institution": "New York University", "country": "US", "position": "Full Professor" }, { "name": "C. Bayan Bruss", "gender": "Male", "institution": "Capital One", "country": "", "position": "Director of Applied Research" }, { "name": "Daohan Lu", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Polina Kirichenko", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Researcher" }, { "name": "Shikai Qiu", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Yucen Lily Li", "gender": "unknown", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Tim Georg Johann Rudner", "gender": "Not Specified", "institution": "New York University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 56, "n_ref": 115, "n_ref_all": 133, "n_fig": 9, "n_tab": 2, "L_tab": 115, "n_element_tab": 54, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 325, "n_element_tab_1": 59, "formula_len_all": 463, "formula_len_all_1": 339, "len_all": 174335, "len_all_1": 72309, "len_abs": 1913, "len_title": 44, "len_sents": 49677, "len_sents_1": 37538, "n_sents": 289, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1515, "title": "Pathologies of Out-of-Distribution Detection", "abs": "There is a proliferation of out-of-distribution (OOD) detection methods in deep learning which aim to detect distribution shifts and improve model safety. These methods often rely on supervised learning to train models with in-distribution data and then use the models\u2019 predictive uncertainty or features to identify OOD points. In this paper, we critically re-examine this popular family of OOD detection procedures, revealing deep-seated pathologies. In contrast to prior work, we argue that these procedures are fundamentally answering the wrong question for OOD detection, with no easy fix. Uncertainty-based methods incorrectly conflate high uncertainty with being OOD, and feature-based methods incorrectly conflate far feature-space distance with being OOD. Moreover, there is no reason\nto expect a classifier trained only on in-distribution classes to be able to identify OOD points; for example, we should not necessarily expect a cat-dog classifier to be uncertain about the label of an airplane, which may share features with a cat that help distinguish cats from dogs, despite generally appearing nothing alike. We show how these pathologies manifest as irreducible errors in OOD detection and identify common settings where these methods are ineffective. Additionally, interventions to improve OOD detection such as feature-logit hybrid methods, scaling of model and data size, Bayesian (epistemic) uncertainty representation, and outlier exposure also fail to address the fundamental misspecification.", "keywords": [ "Out-of-distribution detection", "robustness" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 3, 1, 1 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "hkdqxN3c7t", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Edoardo Debenedetti", "gender": "Male", "institution": "ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Fredrik Nestaas", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Florian Tramer", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 57, "n_ref_all": 114, "n_fig": 28, "n_tab": 2, "L_tab": 708, "n_element_tab": 57, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 104, "formula_len_all_1": 0, "len_all": 155674, "len_all_1": 57813, "len_abs": 1003, "len_title": 112, "len_sents": 60795, "len_sents_1": 29578, "n_sents": 431, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1011, "title": "Adversarial Search Engine Optimization for Large Language Models", "abs": "Large Language Models (LLMs) are increasingly used in applications where the model selects from competing third-party content, such as in LLM-powered search engines or chatbot plugins. In this paper, we introduce *Preference Manipulation Attacks*, a new class of attacks that manipulate an LLM\u2019s selections to favor the attacker. We demonstrate that carefully crafted website content or plugin documentations can trick an LLM to promote the attacker products and discredit competitors, thereby increasing user traffic and monetization (a form of adversarial Search Engine Optimization). We show this leads to a *prisoner\u2019s dilemma*, where all parties are incentivized to launch attacks, but the collective effect degrades the LLM\u2019s outputs for everyone. We demonstrate our attacks on production LLM search engines (Bing and Perplexity) and plugin APIs (for GPT-4 and Claude). As LLMs are increasingly used to rank third-party content, we expect Preference Manipulation Attacks to emerge as a significant threat.", "keywords": [ "large language models", "security", "prompt injection", "search engine optimization", "function calling" ], "rating_list": [ 8, 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 4, 3 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "hjdIQ91ssY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Congcong Wen", "gender": "unknown", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "Hao Huang", "gender": "Male", "institution": "New York University", "country": "AE", "position": "Postdoc" }, { "name": "Mengyu Wang", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" }, { "name": "Min Shi", "gender": "Male", "institution": "University of Louisiana at Lafeyette", "country": "US", "position": "Assistant Professor" }, { "name": "Minghan Li", "gender": "Female", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Yan Luo", "gender": "Male", "institution": "Broad Institute of MIT and Harvard", "country": "US", "position": "Postdoc" }, { "name": "Yi Fang", "gender": "Male", "institution": "New York University", "country": "US", "position": "Associate Professor" }, { "name": "Yu Tian", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 31, "n_ref": 40, "n_ref_all": 53, "n_fig": 5, "n_tab": 6, "L_tab": 8897, "n_element_tab": 469, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 9254, "n_element_tab_1": 524, "formula_len_all": 830, "formula_len_all_1": 830, "len_all": 114975, "len_all_1": 69443, "len_abs": 1658, "len_title": 119, "len_sents": 28133, "len_sents_1": 25333, "n_sents": 197, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1668, "title": "FairFedMed: Achieving Equity in Medical Federated Learning via FairLoRA", "abs": "Fairness remains a critical concern in healthcare, where unequal access to services and treatment outcomes can adversely affect patient health. While Federated Learning (FL) presents a collaborative and privacy-preserving approach to model training, ensuring fairness is challenging due to heterogeneous data across institutions, and current research primarily addresses non-medical applications. To fill this gap, we introduce FairFedMed, the first FL dataset specifically designed to study group fairness (i.e., demographics) in the medical field. It consists of paired 2D SLO funfus images and 3D OCT B-Scans from 15,165 glaucoma patients, along with six different demographic attributes. Existing state-of-the-art FL models may work well for natural images but often struggle with medical images due to their unique characteristics. Moreover, these models do not sufficiently address performance disparities across diverse demographic groups. To overcome these limitations, we propose FairLoRA, a novel fairness-aware FL framework based on singular value decomposition(SVD)-based low-rank approximation. FairLoRA incorporates customized singular value matrices for each demographic group and shares singular vector matrices across all demographic groups, ensuring both model equity and computational efficiency. Experimental results on the FairFedMed dataset demonstrate that FairLoRA not only achieves state-of-the-art performance in medical image classification but also significantly improves fairness across diverse populations. Our code and dataset can be accessible via the Github anonymous link: https://github.com/Anonymouse4Science/FairFedMed-FairLoRA.git", "keywords": [ "Fairness", "Equity", "Federated Learning", "FairLoRA" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "hjROBHstZ3", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 13, "author_info_list": [ { "name": "Eran Segal", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "IL", "position": "Full Professor" }, { "name": "Eric P. Xing", "gender": "Male", "institution": "Mohamed bin Zayed Univeristy of AI", "country": "AE", "position": "Full Professor" }, { "name": "Gongxu Luo", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Guangyi Chen", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "US", "position": "Postdoc" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Lingjing Kong", "gender": "Male", "institution": "Computer Science Department, School of Computer Science", "country": "US", "position": "PhD student" }, { "name": "Loka Li", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Mengyue Yang", "gender": "Female", "institution": "University of Bristol", "country": "GB", "position": "Assistant Professor" }, { "name": "Petar Stojanov", "gender": "unknown", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Yixuan Zhang", "gender": "Female", "institution": "University of Melbourne", "country": "", "position": "MS student" }, { "name": "Yuewen Sun", "gender": "Female", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Yujia Zheng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zijian Li", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Postdoc" } ], "n_formula": 87, "n_formula_1": 17, "n_ref_uni": 76, "n_ref": 178, "n_ref_all": 257, "n_fig": 8, "n_tab": 3, "L_tab": 822, "n_element_tab": 33, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6761, "formula_len_all_1": 1611, "len_all": 307749, "len_all_1": 80848, "len_abs": 2021, "len_title": 119, "len_sents": 81170, "len_sents_1": 34595, "n_sents": 660, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1822, "title": "Causal Representation Learning from Multimodal Biological Observations", "abs": "Prevalent in biological applications (e.g., human phenotype measurements), multimodal datasets can provide valuable insights into the underlying biological mechanisms. However, current machine learning models designed to analyze such datasets still lack interpretability and theoretical guarantees, which are essential to biological applications. Recent advances in causal representation learning have shown promise in uncovering the interpretable latent causal variables with formal theoretical certificates. Unfortunately, existing works for multimodal distributions either rely on restrictive parametric assumptions or provide rather coarse identification results, limiting their applicability to biological research which favors a detailed understanding of the mechanisms. \n\nIn this work, we aim to develop flexible identification conditions for multimodal data and principled methods to facilitate the understanding of biological datasets. Theoretically, we consider a flexible nonparametric latent distribution (c.f., parametric assumptions in prior work) permitting causal relationships across potentially different modalities. We establish identifiability guarantees for each latent component, extending the subspace identification results from prior work. Our key theoretical ingredient is the structural sparsity of the causal connections among distinct modalities, which, as we will discuss, is natural for a large collection of biological systems. Empirically, we propose a practical framework to instantiate our theoretical insights. We demonstrate the effectiveness of our approach through extensive experiments on both numerical and synthetic datasets. Results on a real-world human phenotype dataset are consistent with established medical research, validating our theoretical and methodological framework.", "keywords": [ "multimodal observations", "identifiability", "causal representation learning" ], "rating_list": [ 3, 6, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4, 4 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 4, 2, 3, 3, 4 ] }, { "paper_id": "hj323oR3rw", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Eleni Chatzi", "gender": "Female", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Full Professor" }, { "name": "Hao Dong", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Olga Fink", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 12, "n_ref_uni": 43, "n_ref": 123, "n_ref_all": 164, "n_fig": 20, "n_tab": 20, "L_tab": 5686, "n_element_tab": 649, "n_fig_1": 10, "n_tab_1": 7, "L_tab_1": 1338, "n_element_tab_1": 138, "formula_len_all": 668, "formula_len_all_1": 519, "len_all": 202553, "len_all_1": 67609, "len_abs": 2000, "len_title": 143, "len_sents": 61512, "len_sents_1": 30517, "n_sents": 441, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 2014, "title": "Towards Robust Multimodal Open-set Test-time Adaptation via Adaptive Entropy-aware Optimization", "abs": "Test-time adaptation (TTA) has demonstrated significant potential in addressing distribution shifts between training and testing data. Open-set test-time adaptation (OSTTA) aims to adapt a source pre-trained model online to an unlabeled target domain that contains unknown classes. This task becomes more challenging when multiple modalities are involved. Existing methods have primarily focused on unimodal OSTTA, often filtering out low-confidence samples without addressing the complexities of multimodal data. In this work, we present Adaptive Entropy-aware Optimization (AEO), a novel framework specifically designed to tackle Multimodal Open-set Test-time Adaptation (MM-OSTTA) for the first time. Our analysis shows that the entropy difference between known and unknown samples in the target domain strongly correlates with MM-OSTTA performance. To leverage this, we propose two key components: Unknown-aware Adaptive Entropy Optimization (UAE) and Adaptive Modality Prediction Discrepancy Optimization (AMP). These components enhance the model\u2019s ability to distinguish unknown class samples during online adaptation by amplifying the entropy difference between known and unknown samples. To thoroughly evaluate our proposed methods in the MM-OSTTA setting, we establish a new benchmark derived from existing datasets. This benchmark includes two downstream tasks \u2013 action recognition and 3D semantic segmentation \u2013 and incorporates five modalities: video, audio, and optical flow for action recognition, as well as LiDAR and camera for 3D semantic segmentation. Extensive experiments across various domain shift situations demonstrate the efficacy and versatility of the AEO framework. Additionally, we highlight the strong performance of AEO in long-term and continual MM-OSTTA settings, both of which are challenging and highly relevant to real-world applications. This underscores AEO\u2019s robustness and adaptability in dynamic environments. Our source code and benchmarks will be made publicly available.", "keywords": [ "Test-time Adaptation", "Multimodal Learning" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "hiciJQdmpw", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Haowei Lin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jianzhu Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Liang Wang", "gender": "Male", "institution": "Institute of Automation\uff0c CAS\uff0cChina", "country": "CN", "position": "Full Professor" }, { "name": "Minghao Xu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Qiang Liu", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Xiangxin Zhou", "gender": "Not Specified", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yi Xiao", "gender": "Male", "institution": "Griffith University", "country": "AU", "position": "PhD student" }, { "name": "Yiming Ma", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" } ], "n_formula": 29, "n_formula_1": 28, "n_ref_uni": 42, "n_ref": 75, "n_ref_all": 98, "n_fig": 3, "n_tab": 2, "L_tab": 954, "n_element_tab": 19, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 547, "n_element_tab_1": 3, "formula_len_all": 1930, "formula_len_all_1": 1804, "len_all": 129523, "len_all_1": 63398, "len_abs": 1223, "len_title": 126, "len_sents": 32276, "len_sents_1": 29442, "n_sents": 239, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1225, "title": "Dual Flows with Contrastive Guidance for Generating Highly Designable Proteins", "abs": "Deep generative models have achieved substantial success in protein design. A prevalent approach for de novo protein design involves initially designing a protein backbone structure using deep generative models, such as diffusion and flow models, followed by using a separate inverse folding model to design the correponding sequence. Recently, co-design methods, which aim to jointly generate the structure and sequence of a protein, have attracted considerable attention. Despite this, co-designing sequences and structures of long proteins remains challenging. The complexity of this high-dimensional multimodal generative modeling makes sampling of diffusion and flow models prone to accumulated errors, often leading to non-designable regions. To tackle this challenge, we introduce a contrastive guided sampling algorithm with dual multimodal flows to sample both sequences and structures of highly designable proteins. The contrastive guidance uses the lower-quality flow to help the higher-quality flow avoid non-designable regions by gently steering it during sampling. Our method achieves designability of 80% for length-400 proteins and 37% for length-500 proteins, significantly outperforming previous approaches.", "keywords": [ "guided sampling", "flow matching", "protein design" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "hiZPVlbGsI", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Zhenggui Xiang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 6, "n_ref": 8, "n_ref_all": 37, "n_fig": 4, "n_tab": 11, "L_tab": 2424, "n_element_tab": 92, "n_fig_1": 3, "n_tab_1": 11, "L_tab_1": 1993, "n_element_tab_1": 91, "formula_len_all": 967, "formula_len_all_1": 1281, "len_all": 91965, "len_all_1": 52387, "len_abs": 1346, "len_title": 131, "len_sents": 26857, "len_sents_1": 24939, "n_sents": 150, "n_sents_1": 123, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1356, "title": "Table Learning Representation from Scanned PDF Documents Containing Some Red Stamps", "abs": "Generally, it can be challenging to recognize the table contents from scanned PDF documents containing some red stamps and reconcile the recognized table contents. \nIn this paper, we address the reconciliation challenge involving matching the handwritten invoice amounts overlapped with a red stamp against the invoice amounts found in multi-page tables extracted from scanned PDF documents and we propose a context-splitting Transformer-based table recognition method for recognizing the handwritten invoice amounts overlapped with a red stamp against the invoice amounts. \nFirstly, we recognize the layout structure of the table which is detected from the scanned PDF document containing some red stamps on the handwritten invoice amounts of the table. \nSecondly, we represent the table cells as context-splitting embedding vectors which involve spatial context embedding, position context embedding, lexical context embedding, and colored context embedding. \nFinally, we apply a stack of Transformer-based self-attention encoders to recognize the cross-modality table cells where we multiply the length of query vector and the length of key vector with the scaling factor of the original Transformer in order to make the training process more stable. \nWe improve the recognition accuracy of table cells with a red stamp on handwritten invoice amounts.", "keywords": [ "Table Learning Representation", "Table detection", "Table recognition", "Scanned PDF Documents", "Red Stamps" ], "rating_list": [ 1, 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 1, 3, 2 ], "presentation_list": [ 1, 2, 1, 3, 3 ], "contribution_list": [ 1, 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 5, 3 ] }, { "paper_id": "hheFYjOsWO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Haoru Tan", "gender": "Male", "institution": "HKU", "country": "HK", "position": "PhD student" }, { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Jianhui Liu", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Ruifei He", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shiming Zhang", "gender": "unknown", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Si Liu", "gender": "Female", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Wei Huang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "XIAOJUAN QI", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Yue Liao", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 8, "n_ref_uni": 45, "n_ref": 121, "n_ref_all": 145, "n_fig": 11, "n_tab": 8, "L_tab": 1727, "n_element_tab": 239, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 354, "n_element_tab_1": 34, "formula_len_all": 522, "formula_len_all_1": 394, "len_all": 164108, "len_all_1": 62588, "len_abs": 2137, "len_title": 57, "len_sents": 46762, "len_sents_1": 29570, "n_sents": 335, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 2120, "title": "MC-MoE: Mixture Compressor for Mixture-of-Experts LLMs Gains More", "abs": "Mixture-of-Experts large language models (MoE-LLMs) marks a significant step forward of language models, however, they encounter two critical challenges in practice: 1) expert parameters lead to considerable memory consumption and loading latency; and 2) the current activated experts are redundant, as many tokens may only require a single expert. Motivated by these issues, we investigate the MoE-LLMs and make two key observations: a) different experts exhibit varying behaviors on activation reconstruction error, routing scores, and activated frequencies, highlighting their differing importance, and b) not all tokens are equally important-- only a small subset is critical. Building on these insights, we propose MC-MoE, a training-free Mixture-Compressor for MoE-LLMs, which leverages the significance of both experts and tokens to achieve an extreme compression. First, to mitigate storage and loading overheads, we introduce Pre-Loading Mixed-Precision Quantization (PMQ), which formulates the adaptive bit-width allocation as a Linear Programming (LP) problem, where the objective function balances multi-factors reflecting the importance of each expert. Additionally, we develop Online Dynamic Pruning (ODP), which identifies important tokens to retain and dynamically select activated experts for other tokens during inference to optimize efficiency while maintaining performance. Our MC-MoE integrates static quantization and dynamic pruning to collaboratively achieve extreme compression for MoE-LLMs with less accuracy loss, ensuring an optimal trade-off between performance and efficiency Extensive experiments confirm the effectiveness of our approach. For instance, at 2.54 bits, MC-MoE compresses 76.6% of the model, with only a 3.8% average accuracy loss. During dynamic inference, we further reduce activated parameters by 15%, with a performance drop of less than 0.6%. Remarkably, MC-MoE even surpasses floating-point 13b dense LLMs with significantly smaller parameter sizes, suggesting that mixture compression in MoE-LLMs has the potential to outperform both comparable and larger dense LLMs.", "keywords": [ "Mixture-of-Expert", "LLM", "Quantization", "Pruning" ], "rating_list": [ 5, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "hgwGi81ndj", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Anthony GX-Chen", "gender": "unknown", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Kenneth Marino", "gender": "Male", "institution": "University of Utah", "country": "US", "position": "Assistant Professor" }, { "name": "Rob Fergus", "gender": "Male", "institution": "New York University", "country": "US", "position": "Professor" } ], "n_formula": 25, "n_formula_1": 8, "n_ref_uni": 48, "n_ref": 106, "n_ref_all": 181, "n_fig": 41, "n_tab": 7, "L_tab": 2025, "n_element_tab": 100, "n_fig_1": 18, "n_tab_1": 1, "L_tab_1": 396, "n_element_tab_1": 10, "formula_len_all": 1165, "formula_len_all_1": 477, "len_all": 244982, "len_all_1": 73640, "len_abs": 1353, "len_title": 96, "len_sents": 78825, "len_sents_1": 33829, "n_sents": 626, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1334, "title": "Efficient Exploration and Discriminative World Model Learning with an Object-Centric Abstraction", "abs": "In the face of difficult exploration problems in reinforcement learning, we study whether giving an agent an object-centric mapping (describing a set of items and their attributes) allow for more efficient learning. We found this problem is best solved hierarchically by modelling items at a higher level of state abstraction to pixels, and attribute change at a higher level of temporal abstraction to primitive actions. This abstraction simplifies the transition dynamic by making specific future states easier to predict. We make use of this to propose a fully model-based algorithm that learns a discriminative world model, plans to explore efficiently with only a count-based intrinsic reward, and can subsequently plan to reach any discovered (abstract) states.\n\nWe demonstrate the model's ability to (i) efficiently solve single tasks, (ii) transfer zero-shot and few-shot across item types and environments, and (iii) plan across long horizons. Across a suite of 2D crafting and MiniHack environments, we empirically show our model significantly out-performs state-of-the-art low-level methods (without abstraction), as well as performant model-free and model-based methods using the same abstraction. Finally, we show how to reinforce learn low level object-perturbing policies, and supervise learn the object mapping itself.", "keywords": [ "reinforcement learning", "model based reinforcement learning", "world model", "exploration", "hierarchy" ], "rating_list": [ 6, 6, 6, 8, 8 ], "soundness_list": [ 2, 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "hgvERMkXOx", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Jiangxing Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zongqing Lu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 47, "n_ref": 119, "n_ref_all": 138, "n_fig": 3, "n_tab": 3, "L_tab": 769, "n_element_tab": 137, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 699, "n_element_tab_1": 135, "formula_len_all": 1051, "formula_len_all_1": 928, "len_all": 135076, "len_all_1": 65260, "len_abs": 1157, "len_title": 122, "len_sents": 38273, "len_sents_1": 30976, "n_sents": 238, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1160, "title": "Watch Less, Do More: Implicit Skill Discovery for Video-Conditioned Policy", "abs": "In this paper, we study the problem of video-conditioned policy learning. While previous works mostly focus on learning policies that perform a single skill specified by the given video, we take a step further and aim to learn a policy that can perform multiple skills according to the given video, and generalize to unseen videos by recombining these skills. To solve this problem, we propose our algorithm, Watch-Less-Do-More, an information bottleneck-based imitation learning framework for implicit skill discovery and video-conditioned policy learning. In our method, an information bottleneck objective is employed to control the information contained in the video representation, ensuring that it only encodes information relevant to the current skill (Watch-Less). By discovering potential skills from training videos, the learned policy is able to recombine them and generalize to unseen videos to achieve compositional generalization (Do-More). To evaluate our method, we perform extensive experiments in various environments and show that our algorithm substantially outperforms baselines (up to 2x) in terms of compositional generalization ability.", "keywords": [ "video-conditioned policy", "compositional generalization" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "hgv11VQnIk", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Haibo Jin", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Intern" }, { "name": "Haohan Wang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Liying Kang", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "", "position": "MS student" }, { "name": "Peiyan Zhang", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yangqiu Song", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 1, "n_ref_uni": 26, "n_ref": 63, "n_ref_all": 73, "n_fig": 2, "n_tab": 9, "L_tab": 3482, "n_element_tab": 405, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1136, "n_element_tab_1": 169, "formula_len_all": 448, "formula_len_all_1": 40, "len_all": 167853, "len_all_1": 71148, "len_abs": 1485, "len_title": 140, "len_sents": 63455, "len_sents_1": 34674, "n_sents": 478, "n_sents_1": 236, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1491, "title": "GuardVal: Dynamic Large Language Model Jailbreak Evaluation for Comprehensive Safety Testing", "abs": "Jailbreak attacks reveal critical vulnerabilities in Large Language Models (LLMs) by causing them to generate harmful or unethical content. Evaluating these threats is particularly challenging due to the evolving nature of LLMs and the sophistication required in effectively probing their vulnerabilities. Current benchmarks and evaluation methods struggle to fully address these challenges, leaving gaps in the assessment of LLM vulnerabilities. In this paper, we review existing jailbreak evaluation practices and identify three assumed desiderata for an effective jailbreak evaluation protocol. To address these challenges, we introduce GuardVal, a new evaluation protocol that dynamically generates and refines jailbreak prompts based on the defender LLM's state, providing a more accurate assessment of defender LLMs' capacity to handle safety-critical situations. Moreover, we propose a new optimization method that prevents stagnation during prompt refinement, ensuring the generation of increasingly effective jailbreak prompts that expose deeper weaknesses in the defender LLMs. We apply this protocol to a diverse set of models, from Mistral-7b to GPT-4, across 10 safety domains. Our findings highlight distinct behavioral patterns among the models, offering a comprehensive view of their robustness. Furthermore, our evaluation process deepens the understanding of LLM behavior, leading to insights that can inform future research and drive the development of more secure models.", "keywords": [ "Large Language Models", "Safety Evaluation", "Jailbreaking", "Red-teaming" ], "rating_list": [ 3, 5, 8, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "hgjpO0H0id", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Dan Biderman", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "E. Kelly Buchanan", "gender": "Female", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Jakub Smekal", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Jimmy T.H. Smith", "gender": "Male", "institution": "Liquid AI", "country": "US", "position": "Researcher" }, { "name": "Nicolas Zucchet", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Scott Linderman", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" } ], "n_formula": 40, "n_formula_1": 23, "n_ref_uni": 22, "n_ref": 50, "n_ref_all": 71, "n_fig": 10, "n_tab": 1, "L_tab": 78, "n_element_tab": 24, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3374, "formula_len_all_1": 1505, "len_all": 95995, "len_all_1": 66318, "len_abs": 997, "len_title": 120, "len_sents": 31091, "len_sents_1": 26474, "n_sents": 253, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1011, "title": "On the interplay between learning and memory in deep state space models", "abs": "Deep state-space models (SSMs) have emerged as a powerful deep learning architecture for sequence modeling, but the theory of how these models learn long-term dependencies lags the practice. To explain how parameterization and the number of layers affect a model's expressiveness, we study the properties of deep $\\textit{linear}$ SSMs, i.e., linearly coupled stacks of linear time-invariant systems. We show that such systems share timescales across layers, and we provide novel analysis on the role of linear feedforward connections in regularizing these temporal dependencies. In practice, SSMs can struggle with an explosion of the hidden state variance when learning long-term dependencies. We expand our theoretical understanding of this problem for deep SSMs and provide new intuitions on how this problem may be resolved by increasing the number of layers. Finally, we confirm our theoretical results in a teacher-student framework and show the effects of model parameterization on learning convergence.", "keywords": [ "state space models", "long-term dependencies", "sequence modeling", "linear time-invariant systems", "theory", "memory" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "hghJJJUJJR", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Xiaokang Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University, China", "country": "CN", "position": "Professor" }, { "name": "Yichen Song", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Yunbo Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 30, "n_ref": 46, "n_ref_all": 59, "n_fig": 3, "n_tab": 7, "L_tab": 1691, "n_element_tab": 194, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1316, "n_element_tab_1": 173, "formula_len_all": 555, "formula_len_all_1": 501, "len_all": 104580, "len_all_1": 67170, "len_abs": 1144, "len_title": 118, "len_sents": 34001, "len_sents_1": 30347, "n_sents": 264, "n_sents_1": 234, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1156, "title": "DimOL: Dimensional Awareness as a New 'Dimension' in Operator Learning", "abs": "In the realm of computational physics, an enduring topic is the numerical solutions to partial differential equations (PDEs). Recently, the attention of researchers has shifted towards Neural Operator methods, renowned for their capability to approximate \"operators'' --- mappings from functions to functions. Despite the universal approximation theorem within neural operators, ensuring error bounds often requires employing numerous Fourier layers. However, what about lightweight models? In response to this question, we introduce DimOL (Dimension-aware Operator Learning), drawing insights from dimensional analysis. To implement DimOL, we propose the ProdLayer, which can be seamlessly integrated into FNO-based and Transformer-based PDE solvers, enhancing their ability to handle sum-of-products structures inherent in many physical systems. Empirically, DimOL models achieve up to 48% performance gain within the PDE datasets. Furthermore, by analyzing Fourier components' weights, we can symbolically discern the physical significance of each term. This sheds light on the opaque nature of neural networks, unveiling underlying physical principles.", "keywords": [ "Physics Simulation", "Fourier Neural Operators" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 1, 1 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "hgcxwrrGZf", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Che Lin", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Full Professor" }, { "name": "Ming-Yi Hong", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "PhD student" }, { "name": "Yen-Jung Hsu", "gender": "Female", "institution": "National Taiwan University", "country": "TW", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 29, "n_ref": 72, "n_ref_all": 90, "n_fig": 2, "n_tab": 8, "L_tab": 3545, "n_element_tab": 254, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1898, "n_element_tab_1": 128, "formula_len_all": 728, "formula_len_all_1": 432, "len_all": 155184, "len_all_1": 64584, "len_abs": 1587, "len_title": 105, "len_sents": 54119, "len_sents_1": 28555, "n_sents": 434, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1593, "title": "MTSTRec: Multimodal Time-Aligned Shared Token Recommender", "abs": "Sequential recommendation in e-commerce leverages users' anonymous browsing histories to offer personalized product suggestions without relying on personal information. While item ID-based sequential recommendations are commonly used, they often fail to fully capture the diverse factors influencing user preferences, such as textual descriptions, visual content, and pricing. These factors represent distinct modalities in recommender systems. Existing multimodal sequential recommendation models typically employ either early or late fusion of different modalities, overlooking the alignment of corresponding positions in time of product sequences that represent users' browsing preferences. To address these limitations, this paper proposes a unified framework for multimodal fusion in recommender systems, introducing the Multimodal Time-aligned Shared Token Recommender (MTSTRec). MTSTRec leverages a transformer-based architecture that incorporates a single time-aligned shared token for each product, allowing for efficient cross-modality fusion that also aligns in time. This approach not only preserves the distinct contributions of each modality but also aligns them to better capture user preferences. Additionally, the model extracts rich features from text, images, and other product data, offering a more comprehensive representation of user decision-making in e-commerce. Extensive experiments demonstrate that MTSTRec achieves state-of-the-art performance across multiple sequential recommendation benchmarks, significantly improving upon existing multimodal fusion strategies.", "keywords": [ "multimodal sequential recommendation", "time-aligned shared token", "image style representation", "large language model" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "hgagmZSAb9", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Hang Yan", "gender": "Male", "institution": "AI lab", "country": "", "position": "Researcher" }, { "name": "Kai Chen", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Libo Qin", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "Full Professor" }, { "name": "Qiguang Chen", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Qipeng Guo", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "", "position": "Researcher" }, { "name": "Wanxiang Che", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Yicheng Zou", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Zhi Chen", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "haijun Lv", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 31, "n_ref": 67, "n_ref_all": 104, "n_fig": 21, "n_tab": 2, "L_tab": 996, "n_element_tab": 158, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 2303, "n_element_tab_1": 131, "formula_len_all": 415, "formula_len_all_1": 415, "len_all": 151888, "len_all_1": 61019, "len_abs": 1997, "len_title": 173, "len_sents": 46391, "len_sents_1": 24531, "n_sents": 367, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 125, "L_abs": 1508, "title": "What are the Essential Factors in Crafting Effective Long Context Multi-Hop Instruction Datasets? Insights and Best Practices", "abs": "Recent advancements in large language models (LLMs) with extended context windows have significantly improved tasks such as information extraction, question answering, and complex planning scenarios. In order to achieve success in long-context tasks, a large amount of work has been done to enhance the long-context capabilities of the model through synthetic data. Existing methods typically utilize the Self-Instruct framework to generate instruction-tuning data for better long-context capability improvement. However, our preliminary experiments indicate that less than 35% of samples generated by Qwen-2$_{72B}$ are multi-hop, and more than 40% exhibit poor quality, limiting comprehensive understanding and further research.\n To improve the quality of synthetic data, we propose the Multi-agent Interactive Multi-hop Generation (MIMG) framework, incorporating a Quality Verification Agent, a Single-hop Question Generation Agent, a Multiple Question Sampling Strategy, and a Multi-hop Question Merger Agent. This framework improves the data quality, with the proportion of high-quality, multi-hop, and diverse data exceeding 85%. Furthermore, we systematically investigate strategies for document selection, question merging, and validation techniques through extensive experiments across various models. Our findings show that our synthetic high-quality long-context instruction data significantly enhances model performance, even surpassing models trained on larger amounts of human-annotated data.", "keywords": [ "Long context learning", "data synthesis", "multi-hop QA", "large language model" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "hgTFotBRKl", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Huaxiu Yao", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Jaehong Yoon", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Postdoc" }, { "name": "Mohit Bansal", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" }, { "name": "Shoubin Yu", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Vaidehi Patil", "gender": "Female", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 28, "n_ref": 123, "n_ref_all": 140, "n_fig": 6, "n_tab": 9, "L_tab": 3289, "n_element_tab": 255, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 721, "n_element_tab_1": 116, "formula_len_all": 569, "formula_len_all_1": 569, "len_all": 157406, "len_all_1": 52605, "len_abs": 179, "len_title": 225, "len_sents": 43244, "len_sents_1": 23528, "n_sents": 289, "n_sents_1": 147, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 2227, "title": "SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation", "abs": "Recent advances in diffusion models have significantly enhanced their ability to generate high-quality images and videos, but they have also increased the risk of producing unsafe content. Existing unlearning/editing-based methods for safe generation remove harmful concepts from the models but face several challenges: (1) They cannot instantly remove harmful or undesirable concepts (e.g., artist styles) without extra training. (2) Their safe generation abilities depend on collected training data. (3) They alter model weights, thus risking degrading quality unrelated to content unrelated to the toxic targeted concepts. To address these challenges, we propose SAFREE, a novel training-free approach for safe text-to-image and video generation, without altering the model's weights. Specifically, we detect a subspace corresponding to a set of toxic concepts in the text embedding space and steer prompt token embeddings away from this subspace, thereby filtering out harmful content while preserving intended semantics. To balance the trade-off between filtering toxicity and preserving safe concepts, SAFREE incorporates a novel self-validating filtering mechanism that dynamically adjusts denoising steps when applying filtered embeddings. Additionally, we incorporate adaptive re-attention mechanisms within the diffusion latent space to selectively reduce the influence of features related to toxic concepts at the pixel level. By integrating filtering across both textual embedding and visual latent spaces, SAFREE achieves coherent safety checking, ensuring the fidelity, quality, and safety of the generated outputs. Empirically, SAFREE demonstrates state-of-the-art performance for suppressing unsafe content in T2I generation (reducing 22% across 5 datasets) compared to other training-free methods and effectively filters targeted concepts, e.g., specific artist styles, while maintaining high-quality output. It also shows competitive results against training-based methods. We further extend our SAFREE to various T2I backbones and T2V tasks, showcasing its flexibility and generalization. As generative AI rapidly evolves, SAFREE provides a robust and adaptable safeguard for ensuring safe visual generation.", "keywords": [ "Safe generation", "Text-to-image generation", "Text-to-video generation" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 5, 4 ] }, { "paper_id": "hgBVVAJ1ym", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Danny Dongyeop Han", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Jiook Cha", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Assistant Professor" }, { "name": "Yunju Cho", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Jay Yoon Lee", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 77, "n_ref": 150, "n_ref_all": 190, "n_fig": 24, "n_tab": 4, "L_tab": 2473, "n_element_tab": 256, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 844, "n_element_tab_1": 88, "formula_len_all": 163, "formula_len_all_1": 29, "len_all": 231934, "len_all_1": 66683, "len_abs": 1705, "len_title": 148, "len_sents": 70914, "len_sents_1": 33284, "n_sents": 439, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1710, "title": "MIND THE GAP: ALIGNING THE BRAIN WITH LANGUAGE MODELS REQUIRES A NONLINEAR AND MULTIMODAL APPROACH", "abs": "Speech comprehension involves complex, nonlinear, and multimodal processes within the brain, integrating auditory signals with linguistic and semantic information across widespread brain networks. Traditional brain encoding models, often relying on linear mappings from unimodal features, fall short in representing these intricate mechanisms. In this study, we introduce a nonlinear, multimodal encoding model that combines audio and linguistic features extracted from pre-trained deep learning models (e.g., LLAMA and Whisper). These nonlinear architectures and early fusion mechanisms significantly enhance cross-modal integration, achieving a 14.4% increase in average normalized correlation coefficient and 7.7% increase in average single-story correlation compared to the previous state-of-the-art model relying on weighted averaging of linear unimodal predictions. Moreover, this improved performance reveals novel insights into the brain's functional organization, demonstrating how auditory and semantic information are nonlinearly fused within regions linked to motor control, somatosensory processing, and higher-level semantic representation. Our findings provide empirical support for foundational neurolinguistic theories, including the Motor Theory of Speech Perception, embodied semantic memory, and the Convergence Zone model, revealing novel insights into neural mechanisms otherwise impossible with simpler encoder models. By emphasizing the critical role of nonlinearity and multimodality in brain encoding models, our work bridges neural mechanisms and computational modeling, paving the way for the development of more biologically inspired, brain-aligned artificial intelligence systems.", "keywords": [ "fMRI language encoding", "brain LLM alignment", "neuroscience", "large language models", "neurolinguistics" ], "rating_list": [ 8, 5, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 4, 3, 4 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "hfRb6yC0W0", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Alex Ossadtchi", "gender": "Male", "institution": "AIRI", "country": "RU", "position": "Principal Researcher" }, { "name": "Alexey Voskoboinikov", "gender": "unknown", "institution": "Higher School of Economics", "country": "", "position": "Researcher" }, { "name": "Daria Kleeva", "gender": "unknown", "institution": "Higher School of Economics", "country": "", "position": "Researcher" }, { "name": "Daria Makeeva", "gender": "Female", "institution": "Higher School of Economics", "country": "", "position": "Intern" }, { "name": "Ilia Semenkov", "gender": "Male", "institution": "Higher School of Economics", "country": "RU", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 26, "n_ref": 47, "n_ref_all": 63, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 496, "formula_len_all_1": 496, "len_all": 95971, "len_all_1": 58380, "len_abs": 892, "len_title": 177, "len_sents": 35358, "len_sents_1": 30026, "n_sents": 196, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 129, "L_abs": 894, "title": "Perceived speech decoding and neurophysiological knowledge mining with explainable AI and non-invasive brain activity recordings", "abs": "Explainable artificial intelligence (XAI) is a branch of AI directed at the development of machine learning (ML) solutions that can be comprehended by the human users. Here we use an interpretable and domain-grounded machine learning architecture applied to non-invasive magnetoencephalographic (MEG) data of subjects performing a speech listening task and discover neurophsyologically plausible spatial-temporal neuronal representations of latent sources identified through self-supervised network training process. Achieving high decoding accuracy in the downstream task our solution bridges the gap between high performance and big data-based AI and the classical neuroimaging research and represents a novel knowledge mining platform where the decoding rule can be interpreted using the accepted in electrophysiology terms and concepts which is likely to advance neuroscientific research.", "keywords": [ "MEG", "Speech decoding", "explainable AI", "cortical mechanisms of speech processing" ], "rating_list": [ 3, 3, 1, 5 ], "soundness_list": [ 2, 1, 1, 3 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "hepYqFTeAD", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Juncheng Mu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yining Zhao", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yue Gao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 24, "n_formula_1": 13, "n_ref_uni": 40, "n_ref": 169, "n_ref_all": 192, "n_fig": 8, "n_tab": 9, "L_tab": 4126, "n_element_tab": 337, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2825, "n_element_tab_1": 131, "formula_len_all": 1338, "formula_len_all_1": 871, "len_all": 172694, "len_all_1": 66121, "len_abs": 1764, "len_title": 103, "len_sents": 51935, "len_sents_1": 28572, "n_sents": 431, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1772, "title": "Unleash The Power of Color for Point Cloud Registration", "abs": "Point cloud registration (PCR) has been an important research subject for many years but remains an open problem, presenting numerous challenges. The stability of existing registration methods is often inadequate, particularly in scenarios with low overlap. This issue primarily arises from the insufficient distinctiveness of extracted point cloud features, leading to ambiguous matches and the proliferation of outliers. To address these bottlenecks in point cloud registration, it is crucial to fully leverage the color information of the point clouds to discern point correspondences effectively.\nHowever, excessive control over color may disrupt the spatial structure of the point cloud, making it essential to find a balance between the aggressiveness and stability of color integration.\nTo tackle these challenges, we propose UPC-PCR, which unlocks the potential of color information while maintaining stability. Specifically, we design a Curvature-Color Fusion Module (CCF) to initialize distinctive features. Additionally, to balance color aggressiveness, we enhance the geometric structure by introducing a Centroid Angular (CA) embedding for superpoint structure encoding, which is particularly effective in low-overlap scenes.\nWhile CCF and CA ensure the distinctiveness of point features, the aggressive use of color in the feature enhancement process may still introduce errors. Therefore, we develop a robust estimator equipped with Feature-based Compatibility Hypergraph Convolution (FCH) to learn higher-order compatibility of correspondences and effectively filter out outliers.\nEvaluation across multiple datasets has demonstrated the state-of-the-art performance of UPC-PCR, achieving registration recalls of 98.4%/90.4% on Color3DMatch/Color3DLoMatch.", "keywords": [ "colored point cloud registration" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "hcvmYgFb2A", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Biao Gong", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "DanDan Zheng", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Feng Zhao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jingdong Chen", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Senior Staff Algorithm Engineer" }, { "name": "Ming Yang", "gender": "Male", "institution": "Ant Group", "country": "US", "position": "Researcher" }, { "name": "Wen Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Ziyuan Huang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zizheng Yang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "kecheng zheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 10, "n_ref": 27, "n_ref_all": 37, "n_fig": 5, "n_tab": 4, "L_tab": 824, "n_element_tab": 63, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 824, "n_element_tab_1": 63, "formula_len_all": 271, "formula_len_all_1": 271, "len_all": 86741, "len_all_1": 47991, "len_abs": 1587, "len_title": 124, "len_sents": 22255, "len_sents_1": 21594, "n_sents": 164, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1594, "title": "UniVAE: A Unified Frame-Enriched Video VAE for Latent Video Diffusion Models", "abs": "Variational Autoencoder (VAE) underscores its indispensable role along the growing prominence of Latent Video Diffusion Models (LVDMs). Nevertheless, current latent generative models are generally built upon image VAEs, which compress the spatial dimension only. While, it is vital for video VAE to model temporal dynamic patterns to produce smooth high quality video reconstruction. To address these issues, we propose UniVAE, which compresses videos both spatially and temporally while ensuring coherent video construction. Specifically, we employ 3D convolutions at varying scales in the encoder to temporally compress videos, enabling the UniVAE to capture dependencies across multiple time scales. Furthermore, existing VAEs only reconstruct videos at a low resolution and fps, bounded by limited GPU memory, which makes the entire video generation pipeline fragmented and complicated. Thus, in conjunction with the new encoder, we explore the potential of the VAE decoder to perform frame interpolation, aiming to synthesize additional intermediate frames without relying on standalone add-on interpolation models. Compared with existing VAEs, the proposed UniVAE explores a unified way to compress videos both spatially and temporally with jointly designed encoder and decoder, thus achieving accurate and smooth video reconstruction at a high frame rate. Extensive experiments on commonly used public datasets for video reconstruction and generation demonstrate the superiority of the proposed UniVAE. The code and the pre-trained models will be released to facilitate further research.", "keywords": [ "Variational Autoencoder", "Unified", "Interpolation", "Latent Video Diffusion Models" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "hcVd3zpVvg", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Feng Zheng", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Jinyu Yang", "gender": "Female", "institution": "University of Birmingham", "country": "GB", "position": "PhD student" }, { "name": "Victor Sanchez", "gender": "Male", "institution": "The university of Warwick", "country": "GB", "position": "Full Professor" }, { "name": "Xiaozhuang Song", "gender": "Male", "institution": "Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Yixuan Yang", "gender": "Male", "institution": "University of Warwick", "country": "GB", "position": "PhD student" }, { "name": "Zixiang Zhao", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Wang Fangjing", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "James Yu", "gender": "Male", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 26, "n_ref": 79, "n_ref_all": 93, "n_fig": 8, "n_tab": 5, "L_tab": 934, "n_element_tab": 132, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1600, "n_element_tab_1": 149, "formula_len_all": 254, "formula_len_all_1": 221, "len_all": 127138, "len_all_1": 65132, "len_abs": 1618, "len_title": 119, "len_sents": 44872, "len_sents_1": 30279, "n_sents": 358, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1655, "title": "MV3D-MAE: 2D Pre-trained MAEs are Effective 3D Representation Learners", "abs": "Deep learning's success relies heavily on the availability of extensive labelled datasets. Compared to 2D data, acquiring 3D data is substantially more expensive and time-consuming. Current multi-modal self-supervised approaches often involve converting 3D data into 2D data for parallel multi-modal training, thereby ignoring the prior knowledge contained within extensively trained 2D models. Therefore, it is important to find ways to utilize 2D feature priors to facilitate the learning process of 3D models. In this paper, we propose MV3D-MAE, a masked autoencoder framework that utilizes a pre-trained 2D MAE model to enhance 3D representation learning. Initially, we convert single 3D point clouds into multi-view depth images. Building on a pre-trained 2D MAE model, we adapt the model for multi-view depth image reconstruction by integrating group attention and incorporating additional attention layers. Then we propose a differentiable 3D reconstruction method named Mv-Swin, which maps the reconstructed results back to 3D objects without the use of camera poses, thereby learning 3D spatial representations. Thus, MV3D-MAE, through the bidirectional transformation between 2D and 3D data, mitigates the differences between modalities and enhances the network's representational performance by leveraging the prior knowledge in the pre-trained 2D MAE. Our model significantly improves performance in few-shot classification and achieves SOTA results in linear Support Vector Machine classification. It also demonstrated competitive performance in other downstream tasks of classification and segmentation in synthetic and real-world datasets.", "keywords": [ "Point Cloud", "MAE", "Multi-view depth image\uff0c2D-3D" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "hbon6Jbp9Q", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Georgy Lonshakov", "gender": "unknown", "institution": "University of Trento", "country": "IT", "position": "MS student" }, { "name": "Nhut Truong", "gender": "unknown", "institution": "University of Trento", "country": "IT", "position": "PhD student" }, { "name": "Uri Hasson", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 20, "n_ref": 32, "n_ref_all": 46, "n_fig": 5, "n_tab": 4, "L_tab": 4022, "n_element_tab": 305, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 926, "n_element_tab_1": 155, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 100729, "len_all_1": 67752, "len_abs": 1415, "len_title": 140, "len_sents": 37722, "len_sents_1": 34245, "n_sents": 246, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1426, "title": "Learning and Interpreting Multiple Representations of Semantics in a Neurobiological System", "abs": "A defining feature of computation in the human brain is that different regions can manifest different representations of the same object set. Here we introduce a novel method to learn and interpret multiple neural representations of lexical objects within specific, topographically-defined brain areas. Our approach fine-tunes a pre-trained language model (LM) for each brain region of interest, resulting in better alignment of the LM\u2019s representational space with that of the corresponding brain area. This alignment is achieved through supervised structural pruning of LM features, which selects a subset of features most relevant to the target brain region. We then interpret these retained features using a linear probing task to identify the semantic information they encode. Both the pruning and probing steps are validated through out-of-sample testing, with pruning significantly improving the prediction of brain representations. This method advances on existing approaches by $i$) eliminating the reliance on hand-crafted encoders, reducing potential biases; $ii$) optimizing the alignment process via data-driven learning; and $iii$) providing interpretability of the semantic features in a black-box LM. From a neurobiological perspective, we find that brain regions encoding social and cognitive aspects of lexical items consistently also represent their sensory-motor features, though the reverse does not hold.", "keywords": [ "Pruning", "probing", "representation", "explainableAI", "semantics", "brain" ], "rating_list": [ 1, 3, 3 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 1, 1, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "hbWFeQ1zBp", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Fanpu Cao", "gender": "unknown", "institution": "South China University of Technology", "country": "", "position": "Undergrad student" }, { "name": "Wenxuan Xie", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 20, "n_ref": 36, "n_ref_all": 46, "n_fig": 1, "n_tab": 5, "L_tab": 1492, "n_element_tab": 167, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 341, "n_element_tab_1": 49, "formula_len_all": 738, "formula_len_all_1": 747, "len_all": 71831, "len_all_1": 63800, "len_abs": 1464, "len_title": 134, "len_sents": 20436, "len_sents_1": 19927, "n_sents": 148, "n_sents_1": 143, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 734, "title": "SWIFT: Mapping Sub-series with Wavelet Decomposition Improves Time Series Forecasting", "abs": "In this paper, we propose $\\textit{SWIFT}$, a lightweight model that is not only powerful, but also efficient in deployment and inference for Long-term Time Series Forecasting (LTSF). Our model is based on two key points: 1. decomposition of sequences using wavelet transform. 2. using only one shared single layer for sub-series' mapping. We conduct comprehensive experiments, and the results show that $\\textit{SWIFT}$ achieves state-of-the-art (SOTA) performance on multiple datasets, offering a promising method for edge computing and deployment in this task. Moreover, it is noteworthy that the number of parameters in $\\textit{SWIFT}$ is only 25\\% of what it would be with a single-layer linear model for time-domain prediction.", "keywords": [ "Time Series Forecasting", "Edge Computing" ], "rating_list": [ 1, 1, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "hbS1t37PGM", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Aaron Courville", "gender": "unknown", "institution": "University of Montreal", "country": "", "position": "Assistant Professor" }, { "name": "Christos Tsirigotis", "gender": "unknown", "institution": "ServiceNow Research", "country": "", "position": "Intern" }, { "name": "Joao Monteiro", "gender": "Male", "institution": "Autodesk", "country": "GB", "position": "Researcher" }, { "name": "Perouz Taslakian", "gender": "Female", "institution": "ServiceNow", "country": "CA", "position": "Researcher" }, { "name": "Vaibhav Adlakha", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 49, "n_ref": 81, "n_ref_all": 123, "n_fig": 6, "n_tab": 7, "L_tab": 5290, "n_element_tab": 619, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1949, "n_element_tab_1": 205, "formula_len_all": 654, "formula_len_all_1": 502, "len_all": 188328, "len_all_1": 73543, "len_abs": 1224, "len_title": 120, "len_sents": 54869, "len_sents_1": 34134, "n_sents": 378, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1227, "title": "Training Universal Text Encoders with Pair Relevance Classification Loss", "abs": "Finetuning large language models (LLMs) using contrastive learning objectives has become the dominant approach for representation learning in general-purpose text embedding tasks. Our work seeks to enable going beyond strictly positive (or negative) pairs of text, to more fine-grained annotations that can capture the nuances of complex language tasks. We propose training text encoders with a simple pair classification loss that utilizes binary cross-entropy on relevance labels. When compared to the standard softmax-based loss for multi-class classification against multiple text alternatives, we find that training with our proposed loss improves the average score across 56 English language tasks of the Massive Text Embedding Benchmark (MTEB), while finetuning the same Meta-Llama-3-8B-Instruct model on the same mix of open datasets. Furthermore, our models excel in the Pair Classification and the Semantic Textual Similarity benchmarks, outperforming many models that are trained on more extensive data. Finally, thorough experiments using graded relevance data from TREC-DL 2023 during training demonstrate that binary cross-entropy provides generalization improvements that the softmax-based loss fails to achieve.", "keywords": [ "sentence embeddings", "text encoders", "contrastive learning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "haJHr4UsQX", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Enver Sangineto", "gender": "unknown", "institution": "University of Trento", "country": "", "position": "Postdoc" }, { "name": "Fiorenzo Parascandolo", "gender": "Male", "institution": "University of Modena e Reggio Emilia", "country": "IT", "position": "PhD student" }, { "name": "Lorenzo Baraldi", "gender": "Male", "institution": "University of Modena and Reggio Emilia ", "country": "IT", "position": "Associate Professor" }, { "name": "Nicholas Moratelli", "gender": "Male", "institution": "University of Modena and Reggio Emilia", "country": "IT", "position": "PhD student" }, { "name": "Rita Cucchiara", "gender": "Female", "institution": "Universit\u00e0 di modena e reggio emilia", "country": "IT", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 71, "n_ref": 201, "n_ref_all": 241, "n_fig": 5, "n_tab": 7, "L_tab": 6498, "n_element_tab": 699, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1254, "n_element_tab_1": 90, "formula_len_all": 465, "formula_len_all_1": 419, "len_all": 235259, "len_all_1": 70948, "len_abs": 1259, "len_title": 119, "len_sents": 65393, "len_sents_1": 35156, "n_sents": 519, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1262, "title": "Causal Graphical Models for Vision-Language Compositional Understanding", "abs": "Recent work has empirically shown that Vision-Language Models (VLMs) struggle\nto fully understand the compositional properties of the human language, usually\nmodeling an image caption as a \u201cbag of words\u201d. As a result, they perform\npoorly on compositional tasks, which require a deeper understanding of the different\nentities of a sentence (subject, verb, etc.) jointly with their mutual relationships\nin order to be solved. In this paper, we model the dependency relations\namong textual and visual tokens using a Causal Graphical Model (CGM), built using\na dependency parser, and we train a decoder conditioned by the VLM visual\nencoder. Differently from standard autoregressive or parallel predictions, our decoder\u2019s\ngenerative process is partially-ordered following the CGM structure. This\nstructure encourages the decoder to learn only the main causal dependencies in\na sentence discarding spurious correlations. Using extensive experiments on five\ncompositional benchmarks, we show that our method significantly outperforms\nall the state-of-the-art compositional approaches by a large margin, and it also improves\nover methods trained using much larger datasets. \nThe code is anonymously available and it will be publicly released after the paper acceptance.", "keywords": [ "Compositionality", "Vision-Language Models", "Causal Learning" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "hZztyfmr8n", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Mukang You", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Ning Gao", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Xingyu Jiang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Xiuhui Zhang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Yue Deng", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" } ], "n_formula": 26, "n_formula_1": 21, "n_ref_uni": 46, "n_ref": 72, "n_ref_all": 77, "n_fig": 3, "n_tab": 3, "L_tab": 516, "n_element_tab": 16, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 530, "n_element_tab_1": 23, "formula_len_all": 2124, "formula_len_all_1": 1278, "len_all": 134753, "len_all_1": 61809, "len_abs": 1164, "len_title": 125, "len_sents": 38230, "len_sents_1": 28315, "n_sents": 295, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1129, "title": "COSTAR: Dynamic Safety Constraints Adaptation in Safe Reinforcement Learning", "abs": "Recent advancements in safe reinforcement learning (safe RL) have focused on developing agents that maximize rewards while satisfying predefined safety constraints. However, the challenge of learning policies capable of generalizing to dynamic safety requirements has rarely been explored. To this end, we propose a novel COntrastive Safe TAsk Representation (COSTAR) framework for safe RL, which can boost existing algorithm's generalization to dynamic safety constraints, including variable cost functions and safety thresholds.In COSTAR, we employ a Safe Task Encoder to extract safety-specific representations from trajectory contexts, effectively distinguishing between various safety constraints with contrastive learning. It is noteworthy that our framework can integrate with existing safe RL algorithms and possesses zero-shot adaptation capability to varying safety constraints during deployment. Extensive experiments demonstrate that our COSTAR framework consistently achieves high rewards while maintaining low costs, and exhibits robust generalization capabilities when dealing with out-of-distribution (OOD) tasks.", "keywords": [ "Reinforcement Learning" ], "rating_list": [ 3, 5, 3, 1 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "hZ3QE0rUt1", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "Banafsheh Karimian", "gender": "unknown", "institution": "\u00c9cole de technologie sup\u00e9rieure", "country": "CA", "position": "PhD student" }, { "name": "Eric Granger", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure", "country": "CA", "position": "Full Professor" }, { "name": "Ismail Ben Ayed", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "Full Professor" }, { "name": "Jackie CK Cheung", "gender": "Male", "institution": "McGill University", "country": "", "position": "Associate Professor" }, { "name": "Lo\u00efc Fosse", "gender": "Male", "institution": "Orange-labs", "country": "FR", "position": "PhD student" }, { "name": "Maxime DARRIN", "gender": "Male", "institution": "CentraleSupelec", "country": "", "position": "PhD student" }, { "name": "Mohammadhadi Shateri", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "Assistant Professor" }, { "name": "Pablo Piantanida", "gender": "Male", "institution": "Universit\u00e9 Paris-Saclay, CNRS ", "country": "FR", "position": "Full Professor" }, { "name": "Philippe Formont", "gender": "unknown", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 93, "n_ref": 143, "n_ref_all": 168, "n_fig": 15, "n_tab": 17, "L_tab": 18445, "n_element_tab": 1803, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 3260, "n_element_tab_1": 377, "formula_len_all": 1122, "formula_len_all_1": 747, "len_all": 293764, "len_all_1": 62122, "len_abs": 1069, "len_title": 0, "len_sents": 48288, "len_sents_1": 23739, "n_sents": 382, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1081, "title": "How to distill task-agnostic representations from many teachers?", "abs": "Casting complex inputs onto tractable representations is a critical step in many fields. Differences in architectures, loss functions, input modalities, and datasets lead to embedding models that capture diverse information of the input. Multi-teacher distillation seeks to exploit this diversity to create richer representations but often remains task-specific. We extend this framework by proposing a task-oriented setting that introduces an objective function based on the \"majority vote\" principle. We demonstrate that the mutual information between the student and the teachers is an upper bound for this function, providing a task-agnostic loss for our distillation procedure. An extensive evaluation is performed in different domains ---natural language processing, computer vision, and molecular modeling --- indicating that our method effectively leverages teacher diversity to produce more informative representations. Finally, we use our method to train and release new state-of-the-art embedders, enabling improved downstream performance in NLP and molecular modeling.", "keywords": [ "knowledge distillation", "representation learning", "natural language processing", "molecular modeling", "computer vision", "embedding models" ], "rating_list": [ 6, 6, 5, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "hYe0o7mnwM", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Yao Zhou", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "Associate Professor" }, { "name": "Yibo Zhong", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 35, "n_ref": 60, "n_ref_all": 90, "n_fig": 13, "n_tab": 14, "L_tab": 8177, "n_element_tab": 450, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 4170, "n_element_tab_1": 285, "formula_len_all": 218, "formula_len_all_1": 184, "len_all": 165430, "len_all_1": 51750, "len_abs": 1191, "len_title": 96, "len_sents": 44588, "len_sents_1": 20839, "n_sents": 335, "n_sents_1": 151, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1261, "title": "Low-Rank Interconnected Adaptation across Layers", "abs": "Low-rank adaptation (LoRA) is a powerful parameter-efficient fine-tuning method that utilizes low-rank projectors $A$ and $B$ to learn weight updates $\\Delta W$ for adaptation targets $W$. However, while the low-rank structure of $A$ and $B$ enables high hardware efficiency, it also restricts the overall weight update to be low-rank, which limits the adaptation performance. In this paper, we propose \\underline{l}ow-rank \\underline{i}nterconnected adaptation across \\underline{l}a\\underline{y}ers (Lily). Specifically, we employ a hierarchical framework where low-dimensional projectors (LPs) retained for downward projection at a particular level, while globally-shared high-dimensional projector (HP) experts perform upward projection across all levels of layers. This interconnected asymmetric structure makes the adaptation much more dynamic and breaks the low-rank weight-update constraint of LoRA when using the same parameters budget. Furthermore, Lily's cross-layer connections facilitate the capture of intricate information and dependencies across different layers, thereby enhancing the model's representational capabilities. Experiments across various modalities, architectures, and model sizes underscore Lily's great performance and efficiency.", "keywords": [ "parameter-efficient fine-tuning", "low-rank adaptation", "foundation model" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 1, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "hYd6BCZTzg", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Changzhi Zhou", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Haiyan Zhao", "gender": "Female", "institution": "Peking University", "country": "", "position": "Associate Professor" }, { "name": "Kechi Zhang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Wanli Gu", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Xiancai Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Xunliang Cai", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuanpeng He", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhengwei Tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhi Jin", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Mengdi Zhang", "gender": "Female", "institution": "Shandong University", "country": "", "position": "Undergrad student" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 23, "n_ref": 59, "n_ref_all": 77, "n_fig": 3, "n_tab": 8, "L_tab": 2376, "n_element_tab": 232, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1609, "n_element_tab_1": 168, "formula_len_all": 273, "formula_len_all_1": 293, "len_all": 146518, "len_all_1": 67489, "len_abs": 1159, "len_title": 116, "len_sents": 42922, "len_sents_1": 28877, "n_sents": 356, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1165, "title": "Revisit Self-Debugging with Self-Generated Tests for Code Generation", "abs": "Large language models (LLMs) have shown significant advancements in code generation, but still face challenges on tasks beyond their basic capabilities. Recently, the notion of self-debugging has been proposed to boost the performance of code generation by leveraging execution feedback from tests. Despite its promise, the availability of high-quality tests in real-world scenarios is limited. In this context, self-debugging with self-generated tests is a promising solution but lacks a full exploration of its limitations and practical potential. Therefore, we investigate its efficacy on diverse programming problems. To deepen our understanding, we propose two distinct paradigms for the process: post-execution and in-execution self-debugging. Within the scope of self-contained Python programming tasks, we find that post-execution self-debugging struggles on basic problems but shows potential for improvement on competitive ones, due to the bias introduced by self-generated tests. On the other hand, in-execution self-debugging enables LLMs to mitigate the bias by solely leveraging intermediate states during execution, thereby enhancing code generation.", "keywords": [ "self-debugging", "code generation", "code reasoning", "large language models" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "hYU0P4Wlj9", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Hao Zhao", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "MS student" }, { "name": "Mingjia Li", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Xiaojie Guo", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 60, "n_ref": 110, "n_ref_all": 130, "n_fig": 13, "n_tab": 4, "L_tab": 1039, "n_element_tab": 64, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2001, "n_element_tab_1": 59, "formula_len_all": 282, "formula_len_all_1": 211, "len_all": 161945, "len_all_1": 61728, "len_abs": 1350, "len_title": 130, "len_sents": 42822, "len_sents_1": 29841, "n_sents": 286, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1357, "title": "LIME-Eval: Rethinking Low-light Image Enhancement Evaluation via Object Detection", "abs": "Due to the nature of enhancement--the absence of paired ground-truth information, high-level vision tasks have been recently employed to evaluate the performance of low-light image enhancement. A widely-used manner is to see how accurately an object detector trained on enhanced low-light images by different candidates can perform with respect to annotated semantic labels. In this paper, we first demonstrate that the mentioned approach is generally prone to overfitting, and thus diminishes its measurement reliability. In search of a proper evaluation metric, we propose LIME-Bench, the first online benchmark platform designed to collect human preferences for low-light enhancement, providing a valuable dataset for validating the correlation between human perception and automated evaluation metrics. We then customize LIME-Eval, a novel evaluation framework that utilizes detectors pre-trained on standard-lighting datasets without object annotations, to judge the quality of enhanced images. By adopting an energy-based strategy to assess the accuracy of output confidence maps, our LIME-Eval can simultaneously bypass biases associated with retraining detectors and circumvent the reliance on annotations for dim images. Comprehensive experiments are provided to reveal the effectiveness of our LIME-Eval. Our code will be made publicly available.", "keywords": [ "Low-light enhancement", "Image quality assessment" ], "rating_list": [ 6, 3, 8, 8 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "hYEV8QmaOt", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Dongliang Chang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Jerry Jiang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Jie Zhou", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jiwen Lu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Lei Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Shengjie Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 36, "n_ref": 46, "n_ref_all": 62, "n_fig": 6, "n_tab": 2, "L_tab": 4637, "n_element_tab": 464, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 4704, "n_element_tab_1": 473, "formula_len_all": 1240, "formula_len_all_1": 1220, "len_all": 99276, "len_all_1": 56713, "len_abs": 1272, "len_title": 142, "len_sents": 24667, "len_sents_1": 21399, "n_sents": 182, "n_sents_1": 160, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1282, "title": "From Forgery to Authenticity: Image Anti-Forensics via Reconstruction and Artefact Elimination", "abs": "In recent years, the development of large-scale vision-language models has resulted in significant advancements in image generation and editing, producing results that can often deceive the naked eye. \nHowever, despite their convincing appearance, these generated images remain susceptible to detection by forgery detectors due to various artefacts.\nThe goal of image anti-forensics is to eliminate such artefacts, ensuring that manipulated images successfully evade detection and enhance their overall quality. Existing image anti-forensics methods primarily focus on rectifying artefacts at the feature level, often overlooking the authenticity of the manipulated regions.\nTo address this limitation, we propose a two-phase approach. In the first phase, we introduce GUIded Diffusive rEfinement (GUIDE), a zero-shot learning-based image refinement module aimed at reconstructing details from unaltered regions.\nIn the second phase, we introduce an artefact removal algorithm to eliminate artefacts from the reconstructed ''forged regions''. We validate the effectiveness of our proposed method across multiple image forgery datasets, and comprehensive ablation studies further affirm the efficacy of each component of our approach. The code will be made available upon acceptance.", "keywords": [ "Image anti-forensics", "computer vision" ], "rating_list": [ 3, 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 1, 3, 3 ], "presentation_list": [ 3, 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 5 ] }, { "paper_id": "hXm0Wu2U9K", "primary_area": "learning theory", "n_author": 7, "author_info_list": [ { "name": "Akshay Krishnamurthy", "gender": "Male", "institution": "Research, Microsoft", "country": "US", "position": "Senior Principal Research Manager" }, { "name": "Audrey Huang", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "PhD student" }, { "name": "Dylan J Foster", "gender": "unknown", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Jason D. Lee", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Tengyang Xie", "gender": "unknown", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Wen Sun", "gender": "unknown", "institution": "Cornell University", "country": "", "position": "Assistant Professor" }, { "name": "Wenhao Zhan", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" } ], "n_formula": 241, "n_formula_1": 35, "n_ref_uni": 66, "n_ref": 225, "n_ref_all": 378, "n_fig": 3, "n_tab": 6, "L_tab": 954, "n_element_tab": 123, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 24948, "formula_len_all_1": 2890, "len_all": 445589, "len_all_1": 79204, "len_abs": 4420, "len_title": 166, "len_sents": 130614, "len_sents_1": 34710, "n_sents": 1088, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 125, "L_abs": 1694, "title": "Correcting the Mythos of KL-Regularization: Direct Alignment without Overoptimization via Chi-Squared Preference Optimization", "abs": "Language model alignment methods, such as reinforcement learning from human feedback (RLHF), have\nled to impressive advances in language model capabilities. However, existing techniques are limited by a widely observed phenomenon known as *overoptimization*, where the quality of the language model degrades over the course of the alignment process. Overoptimization occurs when a language model overfits to inaccuracies in an (either explicit or implicit) offline reward model, and drifts away from preferred responses covered by the data. To discourage such distribution shift, offline alignment methods typically employ KL-regularization, but this, as we show, is too weak to prevent degradation in performance. Then, can we design an efficient algorithm that is provably robust to overoptimization?\n\nIn this paper, we advance theoretical understanding of sample-efficient offline alignment and introduce a new algorithm called $\\chi^2$-Preference Optimization ($\\chi$PO). $\\chi$PO is a one-line change to Direct Preference Optimization (DPO; Rafailov et al. 2023), that modifies only the logarithmic link function in the DPO objective. Despite this minimal change, $\\chi$PO implicitly implements the principle of *pessimism in the face of uncertainty* via regularization with the $\\chi^2$-divergence---which quantifies uncertainty more effectively than KL-regularization---and provably alleviates overoptimization, achieving sample-complexity guarantees based on *single-policy concentrability*---the gold standard in offline reinforcement learning. This guarantee makes $\\chi$PO the first simple, yet general-purpose offline alignment algorithm that is provably robust to overoptimization.", "keywords": [ "Reinforcement Learning Theory", "Offline Reinforcement Learning", "single-policy concentrability", "pessimism", "RLHF" ], "rating_list": [ 6, 6, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 4, 2 ], "confidence_list": [ 3, 3, 3, 3, 3 ] }, { "paper_id": "hXJrQWIoR3", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Chris Ding", "gender": "Male", "institution": "University of Texas at Arlington", "country": "", "position": "Professor" }, { "name": "Jicong Fan", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Xudong Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong (Shenzhen)", "country": "CN", "position": "PhD student" }, { "name": "Ziheng Sun", "gender": "Male", "institution": "Chinese University of HongKong", "country": "", "position": "PhD student" } ], "n_formula": 72, "n_formula_1": 24, "n_ref_uni": 67, "n_ref": 120, "n_ref_all": 141, "n_fig": 3, "n_tab": 12, "L_tab": 5370, "n_element_tab": 222, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 838, "n_element_tab_1": 42, "formula_len_all": 10512, "formula_len_all_1": 1613, "len_all": 211965, "len_all_1": 70302, "len_abs": 1886, "len_title": 116, "len_sents": 51508, "len_sents_1": 28396, "n_sents": 476, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1550, "title": "Explainable Graph Representation Learning via Graph Pattern Analysis", "abs": "Explainable artificial intelligence (XAI) is an important area in the AI community, and interpretability is crucial for building robust and trustworthy AI models. While previous work has explored model-level and instance-level explainable graph learning, there has been limited investigation into explainable graph representation learning.\nIn this paper, we focus on representation-level explainable graph learning and ask a fundamental question: What specific information about a graph is captured in graph representations? Our approach is inspired by graph kernels, which evaluate graph similarities by counting substructures within specific graph patterns. Although the pattern counting vector can serve as an explainable representation, it has limitations such as ignoring node features and being high-dimensional.\nTo address these limitations, we introduce a framework for learning and explaining graph representations through graph pattern analysis. We start by sampling graph substructures of various patterns. Then, we learn the representations of these patterns and combine them using a weighted sum, where the weights indicate the importance of each graph pattern's contribution.\nWe also provide theoretical analyses of our methods, including robustness and generalization. In our experiments, we show how to learn and explain graph representations for real-world data using pattern analysis. Additionally, we compare our method against multiple baselines in both supervised and unsupervised learning tasks to demonstrate its effectiveness.", "keywords": [ "graph representation learning", "explainable graph learning" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "hXA8wqRdyV", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Francesco Croce", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Postdoc" }, { "name": "Maksym Andriushchenko", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Postdoc" }, { "name": "Nicolas Flammarion", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 1, "n_ref_uni": 49, "n_ref": 115, "n_ref_all": 192, "n_fig": 7, "n_tab": 5, "L_tab": 3417, "n_element_tab": 243, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2046, "n_element_tab_1": 122, "formula_len_all": 571, "formula_len_all_1": 48, "len_all": 213015, "len_all_1": 64554, "len_abs": 2001, "len_title": 117, "len_sents": 71707, "len_sents_1": 29114, "n_sents": 554, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1575, "title": "Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks", "abs": "We show that even the most recent safety-aligned LLMs are not robust to simple *adaptive* jailbreaking attacks. First, we demonstrate how to successfully leverage access to *logprobs* for jailbreaking: we initially design an adversarial prompt template (sometimes adapted to the target LLM), and then we apply random search on a suffix to maximize a target logprob (e.g., of the token *``Sure''*), potentially with multiple restarts. In this way, we achieve 100\\% attack success rate---according to GPT-4 as a judge---on Vicuna-13B, Mistral-7B, Phi-3-Mini, Nemotron-4-340B, Llama-2-Chat-7B/13B/70B, Llama-3-Instruct-8B, Gemma-7B, GPT-3.5, GPT-4o, and R2D2 from HarmBench that was adversarially trained against the GCG attack. We also show how to jailbreak *all* Claude models---that do not expose logprobs---via either a transfer or prefilling attack with a *100\\% success rate*. In addition, we show how to use random search on a restricted set of tokens for finding trojan strings in poisoned models---a task that shares many similarities with jailbreaking---which is the algorithm that brought us the *first place* in a recent trojan detection competition. The common theme behind these attacks is that *adaptivity* is crucial: different models are vulnerable to different prompting templates (e.g., R2D2 is very sensitive to in-context learning prompts), some models have unique vulnerabilities based on their APIs (e.g., prefilling for Claude), and in some settings, it is crucial to restrict the token search space based on prior knowledge (e.g., for trojan detection).", "keywords": [ "Jailbreaking", "adversarial attacks", "adversarial robustness", "AI safety" ], "rating_list": [ 6, 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 3, 4, 3 ] }, { "paper_id": "hWmwL9gizZ", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Bingxin Zhou", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "", "position": "Researcher" }, { "name": "Liang Hong", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Full Professor" }, { "name": "Song Ke", "gender": "Male", "institution": "shanghai matwings technology", "country": "CN", "position": "Researcher" }, { "name": "Song Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yang Tan", "gender": "Male", "institution": "East China University of Science and Technology", "country": "", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 38, "n_ref": 68, "n_ref_all": 83, "n_fig": 6, "n_tab": 2, "L_tab": 1070, "n_element_tab": 60, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 227, "formula_len_all_1": 227, "len_all": 127268, "len_all_1": 66043, "len_abs": 1088, "len_title": 127, "len_sents": 37922, "len_sents_1": 33755, "n_sents": 237, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1091, "title": "Immunogenicity Prediction with Dual Attention Enables Vaccine Target Selection", "abs": "Immunogenicity prediction is a central topic in reverse vaccinology for finding candidate vaccines that can trigger protective immune responses. Existing approaches typically rely on highly compressed features and simple model architectures, leading to limited prediction accuracy and poor generalizability. To address these challenges, we introduce ProVaccine, a novel deep learning solution with a dual attention mechanism that integrates pre-trained latent vector representations of protein sequences and structures. We also compile the most comprehensive immunogenicity dataset to date, encompassing over 9,500 antigen sequences, structures, and immunogenicity labels from bacteria, viruses, and tumors. Extensive experiments demonstrate that ProVaccine outperforms existing methods across a wide range of evaluation metrics. Furthermore, we establish a post-hoc validation protocol to assess the practical significance of deep learning models in tackling vaccine design challenges. Our work provides an effective tool for vaccine design and sets valuable benchmarks for future research.", "keywords": [ "protein language model", "protein representation learning", "immunogenicity prediction" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 3, 4, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "hWlCc7Iksi", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Alan Yuille", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Chen Wei", "gender": "unknown", "institution": "Rice University", "country": "US", "position": "Assistant Professor" }, { "name": "Hongru Zhu", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Sucheng Ren", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Yijiang Li", "gender": "Not Specified", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "cihang xie", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 33, "n_ref": 75, "n_ref_all": 92, "n_fig": 5, "n_tab": 9, "L_tab": 3585, "n_element_tab": 433, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1917, "n_element_tab_1": 190, "formula_len_all": 108, "formula_len_all_1": 108, "len_all": 117317, "len_all_1": 50807, "len_abs": 1099, "len_title": 133, "len_sents": 26060, "len_sents_1": 22462, "n_sents": 175, "n_sents_1": 140, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1139, "title": "ARVideo: Autoregressive Pretraining for Self-Supervised Video Representation Learning", "abs": "This paper presents a new self-supervised video representation learning framework \\textbf{ARVideo}, which \\textit{autoregressively} predict the next video token in a tailored sequence order. Two key designs are included. First, we organize autoregressive video tokens into clusters that span both \\textit{spatially} and \\textit{temporally}, thereby enabling a richer aggregation of contextual information compared to the standard spatial-only or temporal-only clusters. Second, we adopt a randomized spatiotemporal prediction order to facilitate learning from multi-dimensional data, addressing the limitations of a handcrafted spatial-first or temporal-first sequence order. Extensive experiments establish ARVideo as an effective paradigm for self-supervised video representation learning. For example, when trained with the ViT-B backbone, ARVideo competitively attains 81.2% on Kinetics-400 and 70.9% on Something-Something V2, which are on par with the strong benchmark set by VideoMAE. Importantly, ARVideo also demonstrates higher training efficiency, \\ie, it trains 14% faster and requires 58% less GPU memory compared to VideoMAE.", "keywords": [ "Autoregressive Pretraining; Self-Supervised Video Representation Learning" ], "rating_list": [ 3, 3, 5, 3, 3 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 3, 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "hWRc2L2hc5", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Damien Teney", "gender": "Male", "institution": "Idiap Research Institute", "country": "CH", "position": "Researcher" }, { "name": "Parsa Rahimi", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "S\u00e9bastien Marcel", "gender": "Male", "institution": "Universit\u00e9 de Lausanne", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 39, "n_ref": 71, "n_ref_all": 108, "n_fig": 8, "n_tab": 6, "L_tab": 4181, "n_element_tab": 260, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 2013, "n_element_tab_1": 128, "formula_len_all": 395, "formula_len_all_1": 326, "len_all": 146268, "len_all_1": 75856, "len_abs": 1021, "len_title": 116, "len_sents": 40379, "len_sents_1": 33072, "n_sents": 265, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1022, "title": "AugGen: Generative Synthetic Augmentation Can Boost Face Recognition", "abs": "As machine learning increasingly relies on large amounts of data, concerns about privacy and ethics have grown. Recently, methods for generating synthetic data to augment or replace real datasets have emerged to mitigate these concerns. In this paper, we demonstrate improved performance on a discriminative task when training on a mix of real and synthetic data, compared to training solely on the original real data. Our synthetic data is generated using a novel sampling method based on a conditional generative model and a discriminator, both trained exclusively on the original data, with no need for auxiliary data nor pre-trained foundation models. We consider the challenging task of face recognition, which is well known for its privacy and ethical issues. Using our augmented dataset, we demonstrate consistent improvements over the model trained on the original dataset, on various benchmarks including IJB-C and IJB-B by up to 5\\% while performing competitively with state-of-the-art synthetic data generation.", "keywords": [ "Generative Model", "Face Recognition", "Synthetic Data" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "hWF0HH8Rr9", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Alexander Prochnow", "gender": "unknown", "institution": "Maastricht University", "country": "NL", "position": "MS student" }, { "name": "Jonas Otten", "gender": "Male", "institution": "VisioLab", "country": "DE", "position": "Machine Learning Engineer " }, { "name": "Lionel Peer", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "", "position": "MS student" }, { "name": "Magnus M\u00fcller", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 7, "n_ref": 14, "n_ref_all": 25, "n_fig": 9, "n_tab": 2, "L_tab": 1618, "n_element_tab": 247, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1165, "n_element_tab_1": 21, "formula_len_all": 400, "formula_len_all_1": 400, "len_all": 71893, "len_all_1": 46291, "len_abs": 670, "len_title": 126, "len_sents": 20733, "len_sents_1": 19658, "n_sents": 153, "n_sents_1": 149, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 670, "title": "Large-Scale Multi-Agent Reinforcement Learning for Traffic Signal Optimization", "abs": "We present a novel approach to Traffic Signal Control (TSC) in a multi-agent environment by modeling communication among agents as a sequence problem, enabling intersections within road networks to communicate with one another. Taking inspiration from point cloud processing and graph neural networks, we make our architecture capable of handling variable road network topologies, including differing numbers of intersections and intersection types, and demonstrate this by successfully training on real & randomly generated road networks and traffic demands. Furthermore, we demonstrate that even utilizing minimal state information can achieve competitive performance.", "keywords": [ "Reinforcement Learning", "Traffic Signal Control", "Multi-Agent", "Transformer" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 1, 1, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "hVwS9KkY6V", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jefferson Hernandez", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Ruben Villegas", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Vicente Ordonez", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 38, "n_ref": 122, "n_ref_all": 138, "n_fig": 4, "n_tab": 3, "L_tab": 1464, "n_element_tab": 128, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2465, "n_element_tab_1": 112, "formula_len_all": 401, "formula_len_all_1": 401, "len_all": 164687, "len_all_1": 70539, "len_abs": 1049, "len_title": 84, "len_sents": 42299, "len_sents_1": 30940, "n_sents": 349, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 1051, "title": "Generative Visual Instruction Tuning", "abs": "We propose to use automatically generated instruction-following data to improve the zero-shot capabilities of a large multimodal model with additional support for generative and image editing tasks. We achieve this by curating a new multimodal instruction-following set using GPT-4V and existing datasets for image generation and editing. Using this instruction set and the existing LLaVA-Finetune instruction set for visual understanding tasks, we produce GenLLaVA, a Generative Large Language and Visual Assistant. GenLLaVA is built through a strategy that combines three types of large pretrained models through instruction finetuning: Mistral for language modeling, SigLIP for image-text matching, and StableDiffusion for text-to-image generation. Our model demonstrates visual understanding capabilities superior to LLaVA and additionally demonstrates competitive results with native multimodal models such as Unified-IO 2, paving the way for building advanced general-purpose visual assistants by effectively re-using existing multimodal models.", "keywords": [ "generative models", "multimodal large language models", "instruction tuning", "Vision Language", "Representation Learning" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 3, 5, 5 ] }, { "paper_id": "hVpAjJPfgZ", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Jiahui Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Qihe Huang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Wenjie Du", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Assistant Professor" }, { "name": "Yang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zhengyang Zhou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "XuqiangLi", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 20, "n_formula_1": 11, "n_ref_uni": 33, "n_ref": 63, "n_ref_all": 78, "n_fig": 8, "n_tab": 4, "L_tab": 3153, "n_element_tab": 116, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1299, "formula_len_all_1": 553, "len_all": 144852, "len_all_1": 59273, "len_abs": 1667, "len_title": 149, "len_sents": 44652, "len_sents_1": 27850, "n_sents": 326, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 1701, "title": "Overcoming Lookback Window Limitations: Exploring Longer Windows in Long-Term Time Series Forecasting", "abs": "Long-term time series forecasting (LTSF) aims to predict future trends based on historical data. While longer lookback windows theoretically provide more comprehensive insights, current Transformer-based models face the Lookback Window Limitation (LWL). On one hand, longer windows introduce redundant information, which can hinder model learning. On the other hand, Transformers tend to overfit temporal noise rather than extract meaningful temporal information when dealing with longer sequences, compounded by their quadratic complexity. In this paper, we aim to overcome LWL, enabling models to leverage more historical information for improved performance. Specifically, to mitigate information redundancy, we introduce the Information Bottleneck Filter (IBF), which applies information bottleneck theory to extract essential subsequences from the input. Additionally, to address the limitations of the Transformer architecture in handling long sequences, we propose the Hybrid-Transformer-Mamba (HTM), which combines the linear complexity and long-range modeling capabilities of Mamba with the Transformer's strength in modeling short sequences. We integrate these two model-agnostic modules into various existing methods and conduct experiments on seven datasets. The results demonstrate that incorporating these modules effectively overcomes the lookback window limitations. Notably, by combining them with the Patch strategy, we design the PIH (\\textbf{P}atch-\\textbf{I}BF-\\textbf{H}TM), successfully extending the window length to 1024\u2014a significantly larger window than previously achieved\u2014and achieving state-of-the-art results, highlighting the potential of exploring even longer windows.", "keywords": [ "long-term time series forcasting", "Mamba", "Information Bottleneck" ], "rating_list": [ 3, 3, 6, 1 ], "soundness_list": [ 1, 2, 3, 1 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "hVTaXJ0I5M", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Alexander Yu", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "Undergrad student" }, { "name": "Matthew Joseph", "gender": "Male", "institution": "Google", "country": "US", "position": "Research scientist" }, { "name": "Monica Ribero Diaz", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 6, "n_ref_uni": 12, "n_ref": 40, "n_ref_all": 51, "n_fig": 6, "n_tab": 1, "L_tab": 173, "n_element_tab": 7, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1003, "n_element_tab_1": 12, "formula_len_all": 3083, "formula_len_all_1": 701, "len_all": 93882, "len_all_1": 62858, "len_abs": 409, "len_title": 89, "len_sents": 33554, "len_sents_1": 26676, "n_sents": 305, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 413, "title": "Privately Counting Partially Ordered Data", "abs": "We consider differentially private counting when each data point consists of $d$ bits satisfying a partial order. Our main technical contribution is a problem-specific $K$-norm mechanism that runs in time $O(d^2)$. Experiments show that, depending on the partial order in question, our solution dominates existing pure differentially private mechanisms and can reduce their error by an order of magnitude or more.", "keywords": [ "differential privacy" ], "rating_list": [ 5, 8, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 2, 3 ] }, { "paper_id": "hUdLs6TqZL", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Dong Gong", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "Assistant Professor" }, { "name": "Lina Yao", "gender": "Female", "institution": "University of New South Wales", "country": "AU", "position": "Full Professor" }, { "name": "Masato Ishii", "gender": "Male", "institution": "Sony Research Inc.", "country": "JP", "position": "Researcher" }, { "name": "Mengjie Zhao", "gender": "unknown", "institution": "Sony", "country": "JP", "position": "Researcher" }, { "name": "Muhammad Jehanzeb Mirza", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Saurav Jha", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "PhD student" }, { "name": "Shiqi Yang", "gender": "Male", "institution": "SB Intuitions, SoftBank", "country": "JP", "position": "Chief Research Scientist" }, { "name": "Shusuke Takahashi", "gender": "Male", "institution": "Sony Group Corporation", "country": "JP", "position": "Researcher" }, { "name": "Yuki Mitsufuji", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Lead Research Scientist" }, { "name": "christian simon", "gender": "Male", "institution": "Sony ", "country": "JP", "position": "Researcher" } ], "n_formula": 26, "n_formula_1": 25, "n_ref_uni": 36, "n_ref": 129, "n_ref_all": 165, "n_fig": 23, "n_tab": 9, "L_tab": 3691, "n_element_tab": 559, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 1215, "n_element_tab_1": 192, "formula_len_all": 2000, "formula_len_all_1": 1640, "len_all": 197116, "len_all_1": 76335, "len_abs": 1353, "len_title": 165, "len_sents": 64914, "len_sents_1": 33168, "n_sents": 473, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 116, "L_abs": 1330, "title": "Mining your own secrets: Diffusion Classifier Scores for Continual Personalization of Text-to-Image Diffusion Models", "abs": "Personalized text-to-image diffusion models have grown popular for their ability to efficiently acquire a new concept from user-defined text descriptions and a few images. However, in the real world, a user may wish to personalize a model on multiple concepts but one at a time, with no access to the data from previous concepts due to storage/privacy concerns. When faced with this continual learning (CL) setup, most personalization methods fail to find a balance between acquiring new concepts and retaining previous ones -- a challenge that *continual personalization* (CP) aims to solve. \nInspired by the successful CL methods that rely on class-specific information for regularization, we resort to the inherent class-conditioned density estimates, also known as *diffusion classifier* (DC) scores, for CP of text-to-image diffusion models. \nNamely, we propose using DC scores for regularizing the parameter-space and function-space of text-to-image diffusion models.\nUsing several diverse evaluation setups, datasets, and metrics, we show that our proposed regularization-based CP methods outperform the state-of-the-art C-LoRA, and other baselines. Finally, by operating in the replay-free CL setup and on low-rank adapters, our method incurs zero storage and parameter overhead, respectively, over the state-of-the-art.", "keywords": [ "Text-to-image diffusion model", "Continual Learning", "Personalization" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "hUb2At2DsQ", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 5, "author_info_list": [ { "name": "Qi Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Qinxiang Cao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xinhao Zheng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xudong Lu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 39, "n_ref": 117, "n_ref_all": 151, "n_fig": 10, "n_tab": 9, "L_tab": 5528, "n_element_tab": 618, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2457, "n_element_tab_1": 205, "formula_len_all": 935, "formula_len_all_1": 1043, "len_all": 218384, "len_all_1": 75569, "len_abs": 1959, "len_title": 159, "len_sents": 56742, "len_sents_1": 33772, "n_sents": 401, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 111, "L_abs": 1950, "title": "Rethinking and improving autoformalization: towards a faithful metric and a Dependency Retrieval-based approach", "abs": "As a central component in formal verification, statement autoformalization has been widely studied including the recent efforts from machine learning community, but still remains a widely-recognized difficult and open problem. In this paper, we delve into two critical yet under-explored gaps: 1) absence of faithful and universal automated evaluation for autoformalization results; 2) agnosia of contextural information, inducing severe hallucination of formal definitions and theorems.\nTo address the first issue, we propose **BEq** (_**B**idirectional **E**xtended Definitional E**q**uivalence_), an automated neuro-symbolic method to determine the equivalence between two formal statements, which is formal-grounded and well-aligned with human intuition.\nFor the second, we propose **RAutoformalizer** (_**R**etrieval-augmented **Autoformalizer**_), augmenting statement autoformalization by _Dependency Retrieval_, retrieving potentially dependent objects from formal libraries.\nWe parse the dependencies of libraries and propose to _structurally informalise_ formal objects by the topological order of dependencies. To evaluate OOD generalization and research-level capabilities, we build a novel benchmark, _Con-NF_, consisting of 961 informal-formal statement pairs from frontier mathematical researches.\nExtensive experiments validate the effectiveness of our proposed approaches. In particular, BEq is evaluated on 200 diverse formal statement pairs with expert-annotated equivalence label, exhibiting significantly improved accuracy ($82.50\\\\% \\mapsto 90.50\\\\%$) and precision ($70.59\\\\% \\mapsto 100.0\\\\%$).\nFor dependency retrieval, a baseline with excellent performance is established.\nThe proposed RAutoformalizer substantially outperforms SOTA baselines in both in-distribution ProofNet benchmark ($12.83\\\\% \\mapsto 18.18\\\\%$, BEq@8) and OOD Con-NF scenario ($4.58\\\\%\\mapsto 16.86\\\\%$, BEq@8). Code, data, and models will be available.", "keywords": [ "Large Language Model", "Formal Verification", "Autoformalization" ], "rating_list": [ 8, 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 3, 4, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3, 4 ], "confidence_list": [ 5, 4, 2, 3, 3 ] }, { "paper_id": "hULJCP47PU", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chenhao Tang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "MS student" }, { "name": "Hong-Sheng Zhou", "gender": "unknown", "institution": "Virginia Commonwealth University", "country": "US", "position": "Associate Professor" }, { "name": "Lei Fan", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Weicheng Yang", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 29, "n_formula_1": 9, "n_ref_uni": 15, "n_ref": 54, "n_ref_all": 70, "n_fig": 4, "n_tab": 6, "L_tab": 4519, "n_element_tab": 414, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3104, "formula_len_all_1": 1018, "len_all": 215193, "len_all_1": 70969, "len_abs": 1233, "len_title": 156, "len_sents": 77927, "len_sents_1": 31547, "n_sents": 709, "n_sents_1": 289, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1190, "title": "Two Halves Make a Whole: How to Reconcile Soundness and Robustness in Watermarking for Large Language Models", "abs": "Watermarking techniques have been used to safeguard AI-generated content. In this paper, we study publicly detectable watermarking schemes (Fairoze et al.), and have several research findings.\n\nFirst, we observe that two important security properties, robustness and soundness, may conflict with each other. We then formally investigate these two properties in the presence of an arguably more realistic adversary that we called editing-adversary, and we can prove an impossibility result that, the robustness and soundness properties cannot be achieved via a publicly-detectable single watermarking scheme.\n\nSecond, we demonstrate our main result: we for the first time introduce the new concept of publicly-detectable dual watermarking scheme, for AI-generated content. We provide a novel construction by using two publicly-detectable watermarking schemes; each of the two watermarking schemes can achieve \u201chalf\u201d of the two required properties: one can achieve robustness, and the other can achieve soundness. Eventually, we can combine the two halves into a whole, and achieve the robustness and soundness properties at the same time. Our construction has been implemented and evaluated.", "keywords": [ "Large language model", "Watermark", "Robustness", "Soundness" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "hUD9ugK2OH", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Fangcong Yin", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Greg Durrett", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Xinyu Zhao", "gender": "Female", "institution": "University of Texas at Austin", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 38, "n_ref": 81, "n_ref_all": 127, "n_fig": 10, "n_tab": 25, "L_tab": 8672, "n_element_tab": 1125, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 1715, "n_element_tab_1": 83, "formula_len_all": 63, "formula_len_all_1": 58, "len_all": 195962, "len_all_1": 69694, "len_abs": 1650, "len_title": 109, "len_sents": 51119, "len_sents_1": 31871, "n_sents": 389, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1657, "title": "Understanding Synthetic Context Extension via Retrieval Heads", "abs": "Long-context LLMs are increasingly in demand for applications such as retrieval-augmented generation. To defray the cost of pretraining LLMs over long contexts, recent work takes an approach of synthetic context extension: fine-tuning LLMs with synthetically generated long-context data in a post-training stage. However, it remains unclear how and why this synthetic context extension imparts abilities for downstream long-context tasks. In this paper, we investigate fine-tuning on synthetic data for three long-context tasks that require retrieval and reasoning. We vary the realism of \u201cneedle\u201d concepts to be retrieved and diversity of the surrounding \u201chaystack\u201d context, from using LLMs to construct synthetic documents to using templated relations and creating symbolic datasets. We find that models trained on synthetic data fall short of the real data, but surprisingly, the mismatch can be interpreted and even predicted in terms of a special set of attention heads that are responsible for retrieval over long context, retrieval heads (Wu et al., 2024). The retrieval heads learned on synthetic data have high overlap with retrieval heads learned on real data, and there is a strong correlation between the recall of heads learned and the downstream performance of a model. Furthermore, with attention knockout and activation patching, we mechanistically show that retrieval heads are necessary and explain model performance, although they are not totally sufficient. Our results shed light on how to interpret synthetic data fine-tuning performance and how to approach creating better data for learning real-world capabilities over long contexts.", "keywords": [ "Large Language Models", "Synthetic Data", "Long Context" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 3, 4, 2, 3 ] }, { "paper_id": "hTphfqtafO", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Cho-Jui Hsieh", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Dorothea Wiesmann Rothuizen", "gender": "Female", "institution": "Google", "country": "CH", "position": "Principal Researcher" }, { "name": "Felix Yu", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Inderjit S. Dhillon", "gender": "Male", "institution": "Google", "country": "US", "position": "Distinguished Scientist" }, { "name": "Ruochen Wang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Si Si", "gender": "Female", "institution": "Google", "country": "", "position": "research scientist" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 58, "n_ref": 131, "n_ref_all": 157, "n_fig": 8, "n_tab": 5, "L_tab": 2326, "n_element_tab": 128, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1518, "n_element_tab_1": 98, "formula_len_all": 379, "formula_len_all_1": 315, "len_all": 194779, "len_all_1": 75713, "len_abs": 1548, "len_title": 96, "len_sents": 62881, "len_sents_1": 35008, "n_sents": 463, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 1555, "title": "Large Language Models are Interpretable Learners", "abs": "The trade-off between expressiveness and interpretability remains a core challenge when building human-centric models for classification and decision-making. While symbolic rules offer interpretability, they often lack expressiveness, whereas neural networks excel in performance but are known for being black boxes. This paper shows a combination of Large Language Models (LLMs) and symbolic programs can bridge this gap. In the proposed LLM-based Symbolic Programs (LSPs), the pretrained LLM with natural language prompts provides a massive set of interpretable modules that can transform raw input into natural language concepts. Symbolic programs then integrate these modules into interpretable decision rules. To train LSPs, we develop a divide-and-conquer approach to incrementally build the program from scratch, where the learning process of each step is guided by LLMs. To evaluate the effectiveness of LSPs in extracting interpretable and accurate knowledge from data, we introduce IL-Bench, a collection of diverse tasks, including both synthetic and real-world scenarios across different modalities. Empirical results demonstrate LSP's superior performance compared to traditional neurosymbolic programs and vanilla automatic prompt tuning methods. Moreover, as the knowledge learned by LSP is a combination of natural language descriptions and symbolic rules, it is easily transferable to humans (interpretable), and other LLMs, and generalizes well to out-of-distribution samples. Our code and benchmark will be released for future research.", "keywords": [ "Interpretability", "Explainable AI", "LLMs", "VLMs", "MLLMs", "Symbolic Learning", "Prompting" ], "rating_list": [ 8, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 2, 4, 4 ] }, { "paper_id": "hTUrBJqECJ", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Dongjun Wei", "gender": "unknown", "institution": "University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Michael Chau", "gender": "unknown", "institution": "University of Hong Kong", "country": "", "position": "Full Professor" }, { "name": "Minjia Mao", "gender": "Male", "institution": "University of Delaware", "country": "US", "position": "PhD student" }, { "name": "Xiao Fang", "gender": "Male", "institution": "University of Delaware", "country": "", "position": "Full Professor" }, { "name": "Zeyu Chen", "gender": "Male", "institution": "University of Delaware", "country": "US", "position": "PhD student" } ], "n_formula": 47, "n_formula_1": 18, "n_ref_uni": 27, "n_ref": 125, "n_ref_all": 141, "n_fig": 5, "n_tab": 6, "L_tab": 4331, "n_element_tab": 361, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 1740, "n_element_tab_1": 192, "formula_len_all": 4330, "formula_len_all_1": 923, "len_all": 188408, "len_all_1": 75048, "len_abs": 1805, "len_title": 124, "len_sents": 62216, "len_sents_1": 31892, "n_sents": 546, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1556, "title": "A Watermark for Low-entropy and Unbiased Generation in Large Language Models", "abs": "Recent advancements in large language models (LLMs) have highlighted the risk of misusing them, raising the need for accurate detection of LLM-generated content. In response, a viable solution is to inject imperceptible identifiers into LLMs, known as watermarks. Previous work demonstrates that unbiased watermarks ensure unforgeability and preserve text quality by maintaining the expectation of the LLM output probability distribution. However, previous unbiased watermarking methods suffer from one or more of the following issues: (1) requiring access to white-box LLMs during detection, (2) incurring long detection time, (3) being not robust against simple watermarking attacks, (4) failing to provide statistical guarantees for the type II error of watermark detection, and (5) being not statistically unbiased for low-entropy scenarios, which hinder their deployment in practice. This study proposes the Sampling One Then Accepting (STA-1) method, a watermark that can address all of these issues. Moreover, we discuss the tradeoff between watermark strength and text quality for unbiased watermarks. We show that in low-entropy scenarios, unbiased watermarks face a tradeoff between watermark strength and the risk of unsatisfactory outputs. Experimental results on both low-entropy and high-entropy datasets demonstrate that STA-1 achieves text quality and watermark strength comparable to existing unbiased watermarks, with a low risk of unsatisfactory outputs. Implementation codes for this study are available online (hidden for peer review).", "keywords": [ "Watermark", "large language model" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "hT1XEfHbtw", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Hafizh Raihan Kurnia Putra", "gender": "Male", "institution": "Universitas Brawijaya", "country": "ID", "position": "Undergrad student" }, { "name": "Novanto Yudistira", "gender": "unknown", "institution": "Universitas Brawijaya", "country": "ID", "position": "Associate Professor" }, { "name": "Tirana Noor Fatyanosa", "gender": "Female", "institution": "Universitas Brawijaya", "country": "ID", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 26, "n_ref": 45, "n_ref_all": 59, "n_fig": 2, "n_tab": 4, "L_tab": 4154, "n_element_tab": 648, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2600, "n_element_tab_1": 405, "formula_len_all": 862, "formula_len_all_1": 864, "len_all": 104006, "len_all_1": 59668, "len_abs": 1264, "len_title": 98, "len_sents": 27033, "len_sents_1": 24363, "n_sents": 199, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1272, "title": "Variational Mode Decomposition and Linear Embeddings are What You Need For Time-Series Forecasting", "abs": "Time-series forecasting often faces challenges due to data volatility, which can lead to inaccurate predictions. Variational Mode Decomposition (VMD) has emerged as a promising technique to mitigate volatility by decomposing data into distinct modes, enhancing forecast accuracy. This study integrates VMD with linear models to develop a robust forecasting framework. Our approach is evaluated on 13 diverse datasets, including ETTm2, WindTurbine, M4, and 10 air quality datasets from Southeast Asian cities. The effectiveness of the VMD strategy is assessed by comparing Root Mean Squared Error (RMSE) values from models utilizing VMD against those without it. Additionally, we benchmark linear-based models against well-known neural network architectures such as LSTM, BLSTM, and RNN. The results demonstrate a significant reduction in RMSE across nearly all models following VMD application. Notably, the Linear + VMD model achieved the lowest average RMSE in univariate forecasting at 0.619. In multivariate forecasting, the DLinear + VMD model consistently outperformed others, attaining the lowest RMSE across all datasets with an average of 0.019. These findings underscore the effectiveness of combining VMD with linear models for superior time-series forecasting.", "keywords": [ "Decomposition; Time-series forecasting; Linear models" ], "rating_list": [ 8, 3, 3, 1 ], "soundness_list": [ 3, 1, 2, 1 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 4, 2, 2, 1 ], "confidence_list": [ 5, 5, 1, 5 ] }, { "paper_id": "hShwhoMRVk", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Domniki Asimaki", "gender": "Female", "institution": "California Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Kamyar Azizzadenesheli", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Yaozhong Shi", "gender": "unknown", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zachary E Ross", "gender": "unknown", "institution": "California Institute of Technology", "country": "US", "position": "Assistant Professor" } ], "n_formula": 33, "n_formula_1": 24, "n_ref_uni": 39, "n_ref": 108, "n_ref_all": 125, "n_fig": 5, "n_tab": 2, "L_tab": 328, "n_element_tab": 5, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 161, "n_element_tab_1": 9, "formula_len_all": 3075, "formula_len_all_1": 1561, "len_all": 153476, "len_all_1": 64760, "len_abs": 538, "len_title": 102, "len_sents": 48902, "len_sents_1": 30483, "n_sents": 400, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 539, "title": "Stochastic Process Learning via Operator Flow Matching", "abs": "Using neural operators, we propose a novel framework for stochastic process learning across arbitrary domains. In particular, we develop operator flow matching for learning stochastic process priors on function spaces. Operator flow matching provides the probability density of any finite collection of points, and enables mathematically tractable functional regression at new points with mean and density estimation. Our method outperforms state of the art models at stochastic process learning, functional regression, and prior learning.", "keywords": [ "Neural Operator", "Flow Matching", "Stochastic Process", "Uncertainty Quantification" ], "rating_list": [ 8, 3, 5, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 1, 1 ], "contribution_list": [ 3, 1, 2, 1 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "hSZaCIznB2", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Daniel Cai", "gender": "unknown", "institution": "Brown University", "country": "US", "position": "Undergrad student" }, { "name": "Randall Balestriero", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 35, "n_ref": 59, "n_ref_all": 97, "n_fig": 18, "n_tab": 10, "L_tab": 2495, "n_element_tab": 284, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 92, "n_element_tab_1": 20, "formula_len_all": 218, "formula_len_all_1": 111, "len_all": 141608, "len_all_1": 55558, "len_abs": 1467, "len_title": 154, "len_sents": 41191, "len_sents_1": 27999, "n_sents": 264, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1476, "title": "No Location Left Behind: Measuring and Improving the Fairness of Implicit Representations for Earth Data", "abs": "Implicit neural representations (INRs) exhibit growing promise in addressing Earth representation challenges, ranging from emissions monitoring to climate modeling. However, existing methods disproportionately prioritize global average performance, whereas practitioners require fine-grained insights to understand biases and variations in these models. To bridge this gap, we introduce FAIR-Earth: a first-of-its-kind dataset explicitly crafted to challenge and examine inequities in Earth representations. FAIR-Earth comprises various high-resolution Earth signals, and uniquely aggregates extensive metadata along stratifications like landmass size and population density to assess the fairness of models. Evaluating state-of-the-art INRs across the various modalities of FAIR-Earth, we uncover striking performance disparities. Certain subgroups, especially those associated with high-frequency signals (e.g., islands, coastlines), are consistently poorly modeled by existing methods. In response, we propose spherical wavelet encodings, building on previous spatial encoding research for INRs. Leveraging the multi-resolution analysis capabilities of wavelets, our encodings yield more consistent performance over various scales and locations, offering more accurate and robust representations of the biased subgroups. These open-source contributions represent a crucial step towards facilitating the equitable assessment and deployment of implicit Earth representations.", "keywords": [ "implicit neural representations", "dataset", "fairness in AI", "representation learning", "geospatial modeling", "Earth representation", "wavelet", "location encoding" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 4, 2, 3 ], "confidence_list": [ 4, 2, 3 ] }, { "paper_id": "hRwxZmcvW9", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Bo Huang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Liangyou Li", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "", "position": "Researcher" }, { "name": "Lifeng Shang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Ruiming Tang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Wei Wang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xin Jiang", "gender": "Male", "institution": "Noah\u2019s Ark Lab, Huawei Technologies", "country": "HK", "position": "Principal Researcher" }, { "name": "Xingshan Zeng", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Yasheng Wang", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "", "position": "Researcher" }, { "name": "Yufei Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Yuxin Jiang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 44, "n_ref": 88, "n_ref_all": 134, "n_fig": 9, "n_tab": 12, "L_tab": 3723, "n_element_tab": 689, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 2811, "n_element_tab_1": 489, "formula_len_all": 1524, "formula_len_all_1": 776, "len_all": 172922, "len_all_1": 71280, "len_abs": 1357, "len_title": 135, "len_sents": 46883, "len_sents_1": 30965, "n_sents": 338, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1363, "title": "Bridging and Modeling Correlations in Pairwise Data for Direct Preference Optimization", "abs": "Direct preference optimization (DPO), a widely adopted offline preference optimization algorithm, aims to align large language models (LLMs) with human-desired behaviors using pairwise preference data. However, the generation of the winning response and the losing response within pairwise data are typically isolated, leading to weak correlations between them as well as suboptimal alignment performance. To address this issue, we propose an effective framework for Bridging and Modeling Correlations in pairwise data, named BMC. Firstly, we increase the consistency and informativeness of the pairwise preference signals through targeted modifications, synthesizing a pseudo-winning response by improving the losing response with the winning response as a reference. Secondly, we identify that DPO alone is insufficient to model these correlations and capture nuanced variations. Therefore, we propose learning token-level correlations by dynamically leveraging the policy model's confidence during training. Comprehensive experiments on QA, math, and instruction-following tasks demonstrate the effectiveness of our approach, significantly surpassing competitive baselines, including DPO. Additionally, our in-depth quantitative analysis reveals the reasons behind our method's superior performance over DPO and showcases its versatility to other DPO variants.", "keywords": [ "large language models", "alignment", "preference optimization" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "hRSabwKj53", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Adrian de Wynter", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Emanuele La Malfa", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Fangru Lin", "gender": "Female", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Furu Wei", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Partner Research Manager" }, { "name": "Jing Yao", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Michael J. Wooldridge", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Shaoguang Mao", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Si-Qing Chen", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Partner Applied Science Manager" }, { "name": "Valentin Hofmann", "gender": "unknown", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Postdoc" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 74, "n_ref": 123, "n_ref_all": 145, "n_fig": 6, "n_tab": 4, "L_tab": 2291, "n_element_tab": 56, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1277, "n_element_tab_1": 86, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 202478, "len_all_1": 65721, "len_abs": 2791, "len_title": 161, "len_sents": 52015, "len_sents_1": 31223, "n_sents": 432, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 111, "L_abs": 1538, "title": "One Language, Many Gaps: Evaluating Dialect Fairness and Robustness of Large Language Models in Reasoning Tasks", "abs": "Language is not monolithic. While many benchmarks are used as proxies to systematically estimate Large Language Models' (LLM) performance in real-life tasks, they tend to ignore the nuances of within-language variation and thus fail to model the experience of speakers of minority dialects. Focusing on African American Vernacular English (AAVE), we present the first study on LLMs' fairness and robustness to a dialect in canonical reasoning tasks (algorithm, math, logic, and comprehensive reasoning). We hire AAVE speakers, including experts with computer science backgrounds, to rewrite seven popular benchmarks, such as HumanEval and GSM8K. The result of this effort is ReDial, a dialectal benchmark comprising 1.2K+ parallel query pairs in Standardized English and AAVE. We use ReDial to evaluate state-of-the-art LLMs, including GPT-4o/4/3.5-turbo, LLaMA-3.1/3, Mistral, and Phi-3. We find that, compared to Standardized English, almost all of these widely used models show significant brittleness and unfairness to queries in AAVE. \nFurthermore, AAVE queries can degrade performance more substantially than misspelled texts in Standardized English, even when LLMs are more familiar with the AAVE queries. Finally, asking models to rephrase questions in Standardized English does not close the performance gap but generally introduces higher costs. Overall, our findings indicate that LLMs provide unfair service to dialect users in complex reasoning tasks. Code can be found at https://anonymous.4open.science/r/redial_eval-0A88.", "keywords": [ "dialect", "large language model", "fairness", "robustness", "reasoning" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "hREMYJ5ZmD", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 6, "author_info_list": [ { "name": "Arian Askari", "gender": "Male", "institution": "Leiden University, Leiden University", "country": "", "position": "PhD student" }, { "name": "Jack Williams", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Jinyang Li", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Nicholas Wilson", "gender": "Male", "institution": "Research, Microsoft", "country": "GB", "position": "Researcher" }, { "name": "Nick McKenna", "gender": "Male", "institution": "Research, Microsoft", "country": "GB", "position": "Researcher" }, { "name": "Reynold Cheng", "gender": "Male", "institution": "the University of Hong Kong, University of Hong Kong", "country": "HK", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 52, "n_ref": 94, "n_ref_all": 135, "n_fig": 23, "n_tab": 10, "L_tab": 2985, "n_element_tab": 217, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 44, "n_element_tab_1": 8, "formula_len_all": 598, "formula_len_all_1": 507, "len_all": 244612, "len_all_1": 68475, "len_abs": 1772, "len_title": 156, "len_sents": 69748, "len_sents_1": 32281, "n_sents": 571, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 124, "L_abs": 1800, "title": "Agents Help Agents: Exploring Training-Free Knowledge Distillation for Small Language Models in Data Science Code Generation", "abs": "Knowledge distillation from Large Language Models (LLMs) to locally hosted Small Language Models (SLMs) provides advantages for Data Science Code Generation (DSCG) such as enhanced data privacy and reduced response times. However, achieving effective distillation without resource-intensive training is challenging. This paper investigates whether LLMs can distill knowledge to SLMs through In-Context Learning (ICL), a training-free method for rapid task adaptation. We present the **Agents Help Agents (AHA)** framework, which facilitates automatic knowledge distillation from LLMs to SLMs via agent orchestration. AHA consists of three phases: exploration through an **Agent Orchestration Interface (AOI)**, memory collection of successful examples, and inference augmented with distilled knowledge. The AOI orchestrates interactions between a LLM as a teacher agent and a SLM as a student agent. And we propose two distillation strategies: a static approach that aggregates an offline instruction set and a dynamic RAG-based approach that distills knowledge dynamically during inference. We evaluate AHA on three challenging code generation tasks for tabular data analysis: TabMWP, BirD-SQL, and WikiTQ. Experimental results demonstrate the effectiveness of AHA, leading to an average 27.5\\% relative improvement in the performance of the Student Agent Phi-3-mini. Additionally, relative gains of 14.3\\% and 30.9\\% are observed in **Llama-3.1-8B** and **GPT-35-Turbo**, respectively, even though those models were not calibrated as part of the orchestration, highlighting the model-agnostic nature of the distilled knowledge in AHA. Further analysis compares distillation and demonstration techniques across different data input settings, providing insights into optimal configurations for DSCG.", "keywords": [ "Knowledge Distillation", "Large Language Models", "Small Language Models", "Data Science Code Generation", "In-Context Learning", "Agent Orchestration" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "hQvX9MBowC", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Dong Won Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "Undergrad student" }, { "name": "Jaehyeon Kim", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Jaewoong Cho", "gender": "unknown", "institution": "KRAFTON", "country": "", "position": "Researcher" }, { "name": "Keon Lee", "gender": "Male", "institution": "KRAFTON", "country": "KR", "position": "Researcher" }, { "name": "Seungjun Chung", "gender": "Male", "institution": "Krafton Inc.", "country": "", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 75, "n_ref": 237, "n_ref_all": 301, "n_fig": 10, "n_tab": 14, "L_tab": 3314, "n_element_tab": 474, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1405, "n_element_tab_1": 150, "formula_len_all": 639, "formula_len_all_1": 411, "len_all": 290168, "len_all_1": 70213, "len_abs": 1745, "len_title": 142, "len_sents": 82136, "len_sents_1": 33444, "n_sents": 574, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 93, "L_abs": 1205, "title": "DiTTo-TTS: Diffusion Transformers for Scalable Text-to-Speech without Domain-Specific Factors", "abs": "Large-scale latent diffusion models (LDMs) excel in content generation across various modalities, but their reliance on phonemes and durations in text-to-speech (TTS) limits scalability and access from other fields. While recent studies show potential in removing these domain-specific factors, performance remains suboptimal. In this work, we introduce DiTTo-TTS, a Diffusion Transformer (DiT)-based TTS model, to investigate whether LDM-based TTS can achieve state-of-the-art performance without domain-specific factors. Through rigorous analysis and empirical exploration, we find that (1) DiT with minimal modifications outperforms U-Net, (2) variable-length modeling with a speech length predictor significantly improves results over fixed-length approaches, and (3) conditions like semantic alignment in speech latent representations are key to further enhancement. By scaling our training data to 82K hours and the model size to 790M parameters, we achieve superior or comparable zero-shot performance to state-of-the-art TTS models in naturalness, intelligibility, and speaker similarity, all without relying on domain-specific factors. Speech samples are available at https://ditto-tts.github.io.", "keywords": [ "speech generation", "speech synthesis", "text-to-speech", "tts", "zero-shot", "non-autoregressive", "latent diffusion", "transformer" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "hQY03s8rOm", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Hannah Bast", "gender": "Female", "institution": "Universit\u00e4t Freiburg", "country": "", "position": "Full Professor" }, { "name": "Sebastian Walter", "gender": "unknown", "institution": "University of Freiburg, Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Simon Ging", "gender": "unknown", "institution": "Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Thomas Brox", "gender": "Male", "institution": "University of Freiburg", "country": "", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 57, "n_ref_all": 76, "n_fig": 11, "n_tab": 15, "L_tab": 8388, "n_element_tab": 1265, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 5098, "n_element_tab_1": 331, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 201188, "len_all_1": 66841, "len_abs": 1468, "len_title": 95, "len_sents": 52977, "len_sents_1": 28574, "n_sents": 437, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1475, "title": "Leveraging Knowledge Graphs to harvest a high-quality dataset for efficient CLIP model training", "abs": "Vision-language contrastive learning based on the CLIP method has been instrumental in driving recent advancements in computer vision. However, high quality CLIP models are based on very large datasets. This makes them expensive to train and hampers the scientific analysis of these models. We show how to train a CLIP base-size model efficiently for a broad domain on a much smaller amount of data. We demonstrate this specifically with the automated creation of a dataset named LivingThings with 8.9M images of animals and plants and 12.2M texts. The dataset is obtained via focused image-search queries of three kinds: entity queries (e.g., \"eagle\"), entity-attribute queries (e.g., \"bushy tail of a fox\"), and type-attribute queries (e.g., \"insect on a leaf\"). The entities and types, as well as some of the texts, are derived from the WordNet and Wikidata knowledge graphs, the attributes are obtained via LLMs. We train a CLIP model from scratch on LivingThings and evaluate it on ImageNet, iNaturalist, and CUB for object classification and OVAD and CUB for attribute classification. On the broad target domain of animals and plants, our model achieves comparable, and sometimes even much better performance than models that have orders of magnitude more parameters or training data. For instance, our ViT-B-32 model improves over much larger state-of-the-art CLIP models on the iNaturalist 21 object classification task. We will publicly release our code and dataset.", "keywords": [ "vision-language", "image-text", "contrastive learning", "CLIP", "dataset", "knowledge graph", "wordnet", "wikidata", "entities", "attributes", "image search", "imagenet", "inaturalist", "cub", "ovad", "attribute classification", "animals", "plants", "small data", "efficiency", "open vocabulary" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "hQOLtZ40hZ", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Angela Zhou", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Defu Cao", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" } ], "n_formula": 116, "n_formula_1": 38, "n_ref_uni": 38, "n_ref": 77, "n_ref_all": 91, "n_fig": 1, "n_tab": 6, "L_tab": 1447, "n_element_tab": 84, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 237, "n_element_tab_1": 31, "formula_len_all": 15348, "formula_len_all_1": 4333, "len_all": 173717, "len_all_1": 75487, "len_abs": 1134, "len_title": 50, "len_sents": 50063, "len_sents_1": 32310, "n_sents": 477, "n_sents_1": 287, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 995, "title": "Orthogonalized Estimation of Difference of Q-functions", "abs": "Offline reinforcement learning is important in many settings with available observational data but the inability to deploy new policies online due to safety, cost, and other concerns. Many recent advances in causal inference and machine learning target estimation of causal contrast functions such as CATE, which is sufficient for optimizing decisions and can adapt to potentially smoother structure. We develop a dynamic generalization of the R-learner (Nie and Wager 2021, Lewis and Syrgkanis 2021) for estimating and optimizing the difference of $Q_\\pi$-functions, $Q_\\pi(s,1)$\u2212$Q_\\pi(s,0)$ (which can be used to optimize multiple-valued actions). We leverage orthogonal estimation to improve convergence rates in the presence of slower nuisance estimation rates and prove consistency of policy optimization under a margin condition. The method can leverage black-box nuisance estimators of the $Q$-function and behavior policy to target estimation of a more structured $Q$-function contrast.", "keywords": [ "offline reinforcement learning", "causal inference", "orthogonal estimation", "heterogeneous treatment effects" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 2, 2 ] }, { "paper_id": "hQKXQf8Xo3", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Hongbin Na", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "MS student" }, { "name": "Ling Chen", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "Full Professor" }, { "name": "Shumao yu", "gender": "unknown", "institution": "Utrecht University", "country": "NL", "position": "MS student" }, { "name": "Tao Shen", "gender": "Male", "institution": "Oracle", "country": "", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 41, "n_ref": 84, "n_ref_all": 103, "n_fig": 4, "n_tab": 6, "L_tab": 2794, "n_element_tab": 95, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2936, "n_element_tab_1": 128, "formula_len_all": 348, "formula_len_all_1": 272, "len_all": 189926, "len_all_1": 68830, "len_abs": 1062, "len_title": 124, "len_sents": 36218, "len_sents_1": 31195, "n_sents": 232, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1064, "title": "Multi-Session Client-Centered Treatment Outcome Evaluation in Psychotherapy", "abs": "In psychotherapy, therapeutic outcome assessment, or treatment outcome evaluation, is essential for enhancing mental health care by systematically evaluating therapeutic processes and outcomes. Existing large language model approaches often focus on therapist-centered, single-session evaluations, neglecting the client's subjective experience and longitudinal progress across multiple sessions. To address these limitations, we propose IPAEval, a client-Informed Psychological Assessment-based Evaluation framework that automates treatment outcome evaluations from the client's perspective using clinical interviews. IPAEval integrates cross-session client-contextual assessment and session-focused client-dynamics assessment to provide a comprehensive understanding of therapeutic progress. Experiments on our newly developed TheraPhase dataset demonstrate that IPAEval effectively tracks symptom severity and treatment outcomes over multiple sessions, outperforming previous single-session models and validating the benefits of items-aware reasoning mechanisms.", "keywords": [ "Large Language Models", "Psychotherapy", "Psychological Assessment", "Evaluation" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "hQ2TUZmse1", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Lei You", "gender": "Male", "institution": "Technical University of Denmark", "country": "DK", "position": "Assistant Professor" }, { "name": "Le-le Cao", "gender": "Male", "institution": "Microsoft (ABK)", "country": "SE", "position": "Senior Principal AI Researcher" }, { "name": "Yijun Bian", "gender": "Female", "institution": "University of Copenhagen", "country": "DK", "position": "Postdoc" } ], "n_formula": 58, "n_formula_1": 16, "n_ref_uni": 27, "n_ref": 45, "n_ref_all": 68, "n_fig": 5, "n_tab": 5, "L_tab": 1267, "n_element_tab": 118, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 592, "n_element_tab_1": 18, "formula_len_all": 5024, "formula_len_all_1": 1162, "len_all": 152126, "len_all_1": 68912, "len_abs": 1359, "len_title": 155, "len_sents": 51935, "len_sents_1": 31886, "n_sents": 448, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1354, "title": "Refining Counterfactual Explanations With Joint-Distribution-Informed Shapley Towards Actionable Minimality", "abs": "Counterfactual explanations (CE) identify data points that closely resemble the observed data but produce different machine learning (ML) model outputs, offering critical insights into model decisions. Despite the diverse scenarios, goals and tasks to which they are tailored, existing CE methods often lack actionable efficiency because of unnecessary feature changes included within the explanations that are presented to users and stakeholders. We address this problem by proposing a method that minimizes the required feature changes while maintaining the validity of CE, without imposing restrictions on models or CE algorithms, whether instance- or group-based. The key innovation lies in computing a joint distribution between observed and counterfactual data and leveraging it to inform Shapley values for feature attributions (FA). We demonstrate that optimal transport (OT) effectively derives this distribution, especially when the alignment between observed and counterfactual data is unclear in used CE methods. Additionally, a counterintuitive finding is uncovered: it may be misleading to rely on an exact alignment defined by the CE generation mechanism in conducting FA. Our proposed method is validated on extensive experiments across multiple datasets, showcasing its effectiveness in refining CE towards greater actionable efficiency.", "keywords": [ "Explainable artificial Intelligence", "Feature attributions", "Counterfactual explanations" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "hPq9weqiwp", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Bingchen Zhao", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Jiangmiao Pang", "gender": "Male", "institution": "Shanghai AI Laboratory ", "country": "CN", "position": "Research Scientist" }, { "name": "XIAOJUAN QI", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Xin Wen", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yilun Chen", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 47, "n_ref": 110, "n_ref_all": 132, "n_fig": 7, "n_tab": 5, "L_tab": 1579, "n_element_tab": 136, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 914, "n_element_tab_1": 76, "formula_len_all": 524, "formula_len_all_1": 513, "len_all": 148107, "len_all_1": 62897, "len_abs": 1423, "len_title": 136, "len_sents": 35232, "len_sents_1": 28253, "n_sents": 267, "n_sents_1": 223, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1434, "title": "An Image is Worth $K$ Slots: Data-efficient Scaling of Self-supervised Visual Pre-training", "abs": "Scaling up data and computing has become the norm for pre-training powerful visual encoders. Current algorithms, when scaled up, often require training on large-scale datasets that are unlikely to be object-centric. However, these algorithms were typically developed and validated on the object-centric ImageNet. This discrepancy may suggest sub-optimal scalability and underutilized data potential. Non-object-centric (NOC) data, with its multiple objects and complex layouts, tends to be more information-dense. To better leverage this underlying structure, we introduce a semantic bottleneck to MIM, which reduces the number of prototypes to encourage the emergence of objectness at patch-level token representation. Further, cross-view consistency regularization is applied to encourage multiview invariance. Together, this induces semantic object discovery and allows instance discrimination to be applied between object-level features (slots). Our experiments encompass pre-training on object-centric, scene-centric, web-crawled, and ego-centric data. Across all settings, our approach learns transferrable representations and achieves significant improvements over prior work in image recognition, scene understanding, and robot learning evaluations. When scaled up with million-scale datasets, our method also demonstrates superior data efficiency and scalability. We will make our code and model artifacts publicly available.", "keywords": [ "non-object-centric data", "self-supervised learning", "representation learning", "visual pre-training", "object discovery" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "hPpyUv1XyQ", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Iryna Gurevych", "gender": "Female", "institution": "Institute for Computer Science, Artificial Intelligence and Technology", "country": "BG", "position": "Full Professor" }, { "name": "Nico Daheim", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Researcher" }, { "name": "Thomas M\u00f6llenhoff", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Researcher" }, { "name": "Clara Isabel Meister", "gender": "Female", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 17, "n_ref_uni": 68, "n_ref": 103, "n_ref_all": 168, "n_fig": 6, "n_tab": 3, "L_tab": 1142, "n_element_tab": 172, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 607, "n_element_tab_1": 83, "formula_len_all": 948, "formula_len_all_1": 822, "len_all": 218929, "len_all_1": 73838, "len_abs": 1003, "len_title": 99, "len_sents": 45841, "len_sents_1": 31861, "n_sents": 374, "n_sents_1": 268, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1006, "title": "Uncertainty-Aware Decoding with Minimum Bayes' Risk", "abs": "Despite their outstanding performance in the majority of scenarios, contemporary\nlanguage models still occasionally produce undesirable outputs, for example, hallucinated text. While such behaviors have previously been linked to uncertainty,\nthere is a notable lack of methods that actively consider uncertainty during text\ngeneration. In this work, we show how Minimum Bayes\u2019 Risk (MBR) decoding, a\nmethod that was originally designed to account for the imperfect nature of probabilistic language models, can be generalized into a principled uncertainty-aware\ndecoding method. In short, we account for model uncertainty during decoding\nby incorporating a posterior over model parameters into MBR\u2019s computation of\nexpected risk. We show that this modified expected risk is useful for both choosing\noutputs and deciding when to abstain from generation. We benchmark different\nmethods for learning posteriors and show that performance correlates with the\ndiversity of the combined set of models\u2019 predictions.", "keywords": [ "mbr", "uncertainty", "llms", "decoding", "machine translation", "language generation", "variational learning" ], "rating_list": [ 5, 5, 8, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "hPk92D2GJV", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Qinglin Qi", "gender": "Female", "institution": "Sichuan University", "country": "", "position": "MS student" }, { "name": "Qipeng Guo", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "", "position": "Researcher" }, { "name": "Xiangkun Hu", "gender": "Male", "institution": "Amazon", "country": "CN", "position": "Applied Scientist" }, { "name": "Yue Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" }, { "name": "Yun Luo", "gender": "Male", "institution": "westlake university", "country": "", "position": "PhD student" }, { "name": "Zheng Zhang", "gender": "Male", "institution": "Amazon", "country": "CN", "position": "Senior Principal Scientist" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 32, "n_ref": 60, "n_ref_all": 98, "n_fig": 16, "n_tab": 8, "L_tab": 2845, "n_element_tab": 633, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 638, "n_element_tab_1": 160, "formula_len_all": 34, "formula_len_all_1": 101, "len_all": 145723, "len_all_1": 61503, "len_abs": 1685, "len_title": 124, "len_sents": 39350, "len_sents_1": 29337, "n_sents": 274, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1636, "title": "BALCONI: BALancing CONtext and Internal Knowledge For Training Flexible LLMs", "abs": "The faithfulness to the context is significant for large language models (LLMs) in tasks such as Retrieval-Augmented Generation (RAG) or Information Extraction. However, LLMs can exhibit a \"stubborn\" reliance on their internal knowledge, which leads to failure in maintaining faithfulness to the context. Ideally, a model should leverage the given context if the user instruction requires to, yet remain correctness based on internal knowledge when the instruction does not provide the context. Considering such scenarios, we propose a balanced benchmark, FaithfulBench, to evaluate the faithfulness of LLMs, together with internal knowledge correctness in LLMs and evaluate whether the improvement in faithfulness would affect internal knowledge. Extensive experiments show that LLMs can be unfaithful to the context to some extent and in the Multi-choice QA, we observe an obvious negative correlation between faithfulness and internal knowledge correctness across different LLMs. Then based on the analysis of faithfulness enhancement methods, we find that instruction tuning using counterfactual data can significantly improve the model's context faithfulness, but compromise the model's internal knowledge. To address such a issue, we propose a straightforward yet effective approach BALCONI training by training with mixup data of factual requests, context requests, and NoAns (I cannot tell the answer from the context) requests. Experiments on our benchmark and a context-based machine translation task demonstrate that BALCONI can achieve a well-balanced effect in improving the balanced faithfulness and internal knowledge.", "keywords": [ "Large Language Model", "Faithfulness", "Internal knowledge" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "hPk6IZXo9V", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Gavin Mischler", "gender": "unknown", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Nima Mesgarani", "gender": "Male", "institution": "Columbia University", "country": "", "position": "Associate Professor" }, { "name": "Richard Antonello", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "PhD student" }, { "name": "Siavash Shams", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Stephan Bickel", "gender": "unknown", "institution": "Northwell Health", "country": "US", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 44, "n_ref": 68, "n_ref_all": 80, "n_fig": 7, "n_tab": 5, "L_tab": 1829, "n_element_tab": 39, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 873, "n_element_tab_1": 23, "formula_len_all": 495, "formula_len_all_1": 427, "len_all": 146308, "len_all_1": 55910, "len_abs": 937, "len_title": 174, "len_sents": 47915, "len_sents_1": 25894, "n_sents": 330, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 124, "L_abs": 941, "title": "Neuro2Semantic: A Transfer Learning Framework for Semantic Reconstruction of Continuous Language from Human Intracranial EEG", "abs": "Decoding continuous language from neural signals remains a significant challenge in the intersection of neuroscience and artificial intelligence. We introduce Neuro2Semantic, a novel framework that reconstructs the semantic content of perceived speech from intracranial EEG (iEEG) recordings. Our approach consists of two phases: first, an LSTM-based adapter aligns neural signals with pre-trained text embeddings; second, a corrector module generates continuous, natural text directly from these aligned embeddings. This flexible method overcomes the limitations of previous decoding approaches and enables unconstrained text generation. Neuro2Semantic achieves remarkable performance with as little as 30 minutes of neural data, significantly outperforming a recent state-of-the-art method in low-data settings. These results highlight the potential for practical applications in brain-computer interfaces and neural decoding technologies.", "keywords": [ "Brain decoding", "Semantic reconstruction", "Transfer learning", "Deep learning", "Intracranial EEG", "Natural Language Processing" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "hPWWXpCaJ7", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Donglin Wang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Associate Professor" }, { "name": "Pengxiang Ding", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Shangke Lyu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Researcher" }, { "name": "Ying Peng", "gender": "Female", "institution": "Westlake University", "country": "", "position": "RA" }, { "name": "Zhang Hongyin", "gender": "Male", "institution": "Westlake University", "country": "", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 51, "n_ref": 99, "n_ref_all": 116, "n_fig": 11, "n_tab": 7, "L_tab": 2226, "n_element_tab": 213, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 426, "n_element_tab_1": 39, "formula_len_all": 590, "formula_len_all_1": 515, "len_all": 157504, "len_all_1": 66618, "len_abs": 1373, "len_title": 124, "len_sents": 45874, "len_sents_1": 32559, "n_sents": 325, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1380, "title": "GEVRM: Goal-Expressive Video Generation Model For Robust Visual Manipulation", "abs": "With the rapid development of embodied artificial intelligence, significant progress has been made in vision-language-action (VLA) models for general robot decision-making. However, the majority of existing VLAs fail to account for the inevitable external perturbations encountered during deployment. These perturbations introduce unforeseen state information to the VLA, resulting in inaccurate actions and consequently, a significant decline in generalization performance. The classic internal model control (IMC) principle demonstrates that a closed-loop system with an internal model that includes external input signals can accurately track the reference input and effectively offset the disturbance. We propose a novel closed-loop VLA method GEVRM that integrates the IMC principle to enhance the robustness of robot visual manipulation. The text-guided video generation model in GEVRM can generate highly expressive future visual planning goals. Simultaneously, we evaluate perturbations by simulating responses, which are called internal embeddings and optimized through prototype contrastive learning. This allows the model to implicitly infer and distinguish perturbations from the external environment. The proposed GEVRM achieves state-of-the-art performance on both standard and perturbed CALVIN benchmarks and shows significant improvements in realistic robot tasks.", "keywords": [ "Robot Manipulation; Vision Language Action Model" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "hPOt3yUXii", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Guy Ohayon", "gender": "Male", "institution": "Technion - Israel Institute of Technology", "country": "IL", "position": "PhD student" }, { "name": "Michael Elad", "gender": "Male", "institution": "Computer Science Department, Technion - Israel Institute of Technology", "country": "IL", "position": "Full Professor" }, { "name": "Tomer Michaeli", "gender": "Male", "institution": "Technion, Technion", "country": "IL", "position": "Associate Professor" } ], "n_formula": 56, "n_formula_1": 20, "n_ref_uni": 57, "n_ref": 152, "n_ref_all": 207, "n_fig": 7, "n_tab": 10, "L_tab": 4273, "n_element_tab": 419, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 1105, "n_element_tab_1": 1, "formula_len_all": 4504, "formula_len_all_1": 1053, "len_all": 221968, "len_all_1": 60944, "len_abs": 1428, "len_title": 133, "len_sents": 57686, "len_sents_1": 28455, "n_sents": 469, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1434, "title": "Posterior-Mean Rectified Flow: Towards Minimum MSE Photo-Realistic Image Restoration", "abs": "Photo-realistic image restoration algorithms are typically evaluated by distortion measures (e.g., PSNR, SSIM) and by perceptual quality measures (e.g., FID, NIQE), where the desire is to attain the lowest possible distortion without compromising on perceptual quality. To achieve this goal, current methods commonly attempt to sample from the posterior distribution, or to optimize a weighted sum of a distortion loss (e.g., MSE) and a perceptual quality loss (e.g., GAN). Unlike previous works, this paper is concerned specifically with the *optimal* estimator that minimizes the MSE under a constraint of perfect perceptual index, namely where the distribution of the reconstructed images is equal to that of the ground-truth ones. A recent theoretical result shows that such an estimator can be constructed by optimally transporting the posterior mean prediction (MMSE estimate) to the distribution of the ground-truth images. Inspired by this result, we introduce Posterior-Mean Rectified Flow (PMRF), a simple yet highly effective algorithm that approximates this optimal estimator. In particular, PMRF first predicts the posterior mean, and then transports the result to a high-quality image using a rectified flow model that approximates the desired optimal transport map. We investigate the theoretical utility of PMRF and demonstrate that it consistently outperforms previous methods on a variety of image restoration tasks.", "keywords": [ "Image restoration", "blind face image restoration", "inverse problems", "perception-distortion tradeoff", "image processing", "computer vision", "machine learning", "generative models", "optimal transport", "rectified flow", "flow matching" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "hOELrZfg0J", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Animesh Garg", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Ignat Georgiev", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Nicklas Hansen", "gender": "Non-Binary", "institution": "NVIDIA", "country": "US", "position": "Intern" }, { "name": "Varun Giridhar", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Undergrad student" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 32, "n_ref": 66, "n_ref_all": 95, "n_fig": 14, "n_tab": 3, "L_tab": 357, "n_element_tab": 35, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 343, "n_element_tab_1": 25, "formula_len_all": 1587, "formula_len_all_1": 902, "len_all": 137820, "len_all_1": 59201, "len_abs": 1198, "len_title": 97, "len_sents": 41019, "len_sents_1": 28085, "n_sents": 333, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1255, "title": "PWM: Policy Learning with Multi-Task World Models", "abs": "Reinforcement Learning (RL) has made significant strides in complex tasks but struggles in multi-task settings with different embodiments. World models methods offer scalability by learning a simulation of the environment, but often rely on inefficient gradient-free optimization methods for policy extraction. In contrast, gradient-based methods exhibit lower variance but fail to handle discontinuities. Our work reveals that well-regularized world models can generate smoother optimization landscapes than the actual dynamics, facilitating more effective first-order optimization. We introduce Policy learning with multi-task World Models (PWM), a novel model-based RL algorithm for continuous control. Initially, the world model is pre-trained on offline data, and then policies are extracted from it using first-order optimization in less than 10 minutes per task. PWM effectively solves tasks with up to 152 action dimensions and outperforms methods that use ground-truth dynamics. Additionally, PWM scales to an 80-task setting, achieving up to 27% higher rewards than existing baselines, without relying on costly online planning. Visualizations and code available at [https://policy-world-model.github.io/](https://policy-world-model.github.io/).", "keywords": [ "reinforcement learning", "model-based reinforcement learning", "continuous control", "world models" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "hNkXTqDrfb", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Jiaye Teng", "gender": "Male", "institution": "Shanghai University of Finance and Economics", "country": "CN", "position": "Assistant Professor" }, { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zixuan Gong", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 210, "n_formula_1": 29, "n_ref_uni": 30, "n_ref": 63, "n_ref_all": 107, "n_fig": 7, "n_tab": 4, "L_tab": 1023, "n_element_tab": 78, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 318, "n_element_tab_1": 35, "formula_len_all": 43564, "formula_len_all_1": 1880, "len_all": 243411, "len_all_1": 68699, "len_abs": 1352, "len_title": 156, "len_sents": 62638, "len_sents_1": 28020, "n_sents": 686, "n_sents_1": 238, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1378, "title": "Mastering Syntax, Unlocking Semantics: A Mathematically Provable Two-stage Learning Process in Transformers", "abs": "Transformers have emerged as a cornerstone across various fields with extensive applications. \nHowever, the training dynamics of transformers remain relatively underexplored.\nIn this work, we present a novel perspective on how transformers acquire knowledge during the training dynamics, inspired by the feature learning theory. \nTo this end, we conceptualize each token as embodying two types of knowledge: elementary knowledge represented by syntactic information, and specialized knowledge represented by semantic information.\nBuilding on this data structure, we rigorously prove that transformers follow a syntax-then-semantics learning paradigm, i.e., first mastering syntax in the Elementary Stage and then unlocking semantics in the subsequent Specialized Stage.\nThe results are derived from the training dynamics analysis and finite-time convergence within the in-context learning framework for supervised classification.\nTo our best knowledge, this is the \\textbf{\\emph{first}} rigorous result of a two-stage optimization process in transformers from a feature learning perspective. Empirical findings on real-world language datasets support the theoretical results of the two-stage learning process. \nMoreover, the spectral properties of attention weights, derived from our theoretical framework, align with the experimental observations, providing further validation.", "keywords": [ "Two-stage learning", "Optimization dynamics", "Feature learning theory" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 2, 5, 2, 2 ] }, { "paper_id": "hNjCVVm0EQ", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Minghao Han", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Xunyuan Yin", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "ZHAOYANG LI", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 12, "n_ref_uni": 55, "n_ref": 88, "n_ref_all": 123, "n_fig": 18, "n_tab": 12, "L_tab": 2613, "n_element_tab": 183, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3335, "formula_len_all_1": 909, "len_all": 176760, "len_all_1": 61439, "len_abs": 1563, "len_title": 119, "len_sents": 51653, "len_sents_1": 29766, "n_sents": 407, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 71, "L_abs": 1570, "title": "MamKO: Mamba-based Koopman operator for modeling and predictive control", "abs": "The Koopman theory, which enables the transformation of nonlinear systems into linear representations, is a powerful and efficient tool to model and control nonlinear systems. However, the ability of the Koopman operator to model complex systems, particularly time-varying systems, is limited by the fixed linear state-space representation. To address the limitation, the large language model, Mamba, is considered a promising strategy for enhancing modeling capabilities while preserving the linear state-space structure.\nIn this paper, we propose a new framework, the Mamba-based Koopman operator (MamKO), which provides enhanced model prediction capability and adaptability, as compared to Koopman models with constant Koopman operators. Inspired by the Mamba structure, MamKO generates Koopman operators from online data; this enables the model to effectively capture the dynamic behaviors of the nonlinear system over time. A model predictive control system is then developed based on the proposed MamKO model. The modeling and control performance of the proposed method is evaluated through experiments on benchmark time-invariant and time-varying systems. The experimental results demonstrate the superiority of the proposed approach. Additionally, we perform ablation experiments to test the effectiveness of individual components of MamKO. This approach unlocks new possibilities for integrating large language models with control frameworks, and it achieves a good balance between advanced modeling capabilities and real-time control implementation efficiency.", "keywords": [ "Mamba; Koopman operator; model predictive control; nonlinear systems" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "hMjUnF3aQ8", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Dotan Di Castro", "gender": "Male", "institution": "Bosch", "country": "", "position": "Bosch Center for Artificial Inteligence" }, { "name": "Gilad Katz", "gender": "Male", "institution": "Ben-Gurion University of the Negev", "country": "IL", "position": "Assistant Professor" }, { "name": "Nitsan Soffair", "gender": "Not Specified", "institution": "Ben Gurion University of the Negev", "country": "IL", "position": "PhD student" }, { "name": "Orly Avner", "gender": "Female", "institution": "Bosch Center for Artificial Inteligence", "country": "", "position": "Researcher" }, { "name": "Shie Mannor", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Full Professor" } ], "n_formula": 25, "n_formula_1": 25, "n_ref_uni": 21, "n_ref": 28, "n_ref_all": 40, "n_fig": 0, "n_tab": 4, "L_tab": 337, "n_element_tab": 89, "n_fig_1": 0, "n_tab_1": 6, "L_tab_1": 435, "n_element_tab_1": 105, "formula_len_all": 1690, "formula_len_all_1": 1689, "len_all": 84277, "len_all_1": 57596, "len_abs": 884, "len_title": 84, "len_sents": 28850, "len_sents_1": 26734, "n_sents": 183, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 38, "L_abs": 454, "title": "SQT -- rough conservative actor critic", "abs": "Std $Q$-target is a conservative actor critic ensemble based $Q$-learning algorithm which based on a single key $Q$-formula--$Q$-networks standard deviation, an uncertainty penalty. A minimalistic solution to the problem of overestimation bias. We implement SQT on top of actor critic and test it against the SOTA actor critic algorithms on popular MuJoCo tasks. SQT shows a clear performance advantage over TD3, SAC and TD7 on the tested tasks majority.", "keywords": [ "Actor Critic", "Overestimation Bias" ], "rating_list": [ 3, 1, 3, 1 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 1, 1, 3, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 2, 5, 4, 4 ] }, { "paper_id": "hMEHnLJyrU", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Botong Zhang", "gender": "Female", "institution": "University of Pennsylvania", "country": "US", "position": "Undergrad student" }, { "name": "Kayo Yin", "gender": "Female", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Osbert Bastani", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Assistant Professor" }, { "name": "Shuo Li", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Vishakh Padmakumar", "gender": "unknown", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Alexander G Shypula", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 47, "n_ref": 73, "n_ref_all": 89, "n_fig": 5, "n_tab": 8, "L_tab": 2212, "n_element_tab": 232, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1600, "n_element_tab_1": 141, "formula_len_all": 330, "formula_len_all_1": 163, "len_all": 155017, "len_all_1": 60953, "len_abs": 1056, "len_title": 124, "len_sents": 42972, "len_sents_1": 28697, "n_sents": 328, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1106, "title": "Does Instruction Tuning Reduce Diversity? A Case Study Using Code Generation", "abs": "Large Language Models (LLMs) should ideally generate diverse content for open-ended prompts (e.g., variety in cooking recipes). Anecdotal evidence has suggested that preference-tuned language models struggle to generate diverse content, which would have important implications for how we align models. However, research on this question has been limited by the difficulty of measuring diversity, which naively would require costly human evaluation. We propose to leverage code as a means to study semantic diversity, since code has executable semantics. To this end, we create an open-ended program synthesis task, enabling us to cheaply evaluate the diversity of hundreds of thousands of generations. Using our methodology, we find that while instruction-tuning reduces syntactic and lexical diversity, it can actually increase semantic diversity. We also study the effect of model size and prompting technique on diversity. Finally, we find that neural diversity metrics correlate poorly with our semantic diversity metrics, highlighting the need for more rigorous methodologies for evaluating diversity.", "keywords": [ "LLM", "Diversity", "RLHF", "DPO", "SFT", "Program Synthesis", "Evaluation" ], "rating_list": [ 1, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "hLwcNSFhC2", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Chunbin Gu", "gender": "Male", "institution": "Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Hanqun CAO", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jingjie Zhang", "gender": "Female", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xiaorui Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" }, { "name": "Zijun Gao", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 38, "n_ref": 58, "n_ref_all": 72, "n_fig": 5, "n_tab": 6, "L_tab": 2853, "n_element_tab": 187, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1648, "n_element_tab_1": 149, "formula_len_all": 1060, "formula_len_all_1": 900, "len_all": 137086, "len_all_1": 61680, "len_abs": 2175, "len_title": 130, "len_sents": 40582, "len_sents_1": 29259, "n_sents": 254, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 2267, "title": "SAGEPhos: Sage Bio-Coupled and Augmented Fusion for Phosphorylation Site Detection", "abs": "Phosphorylation site prediction based on kinase-substrate interaction plays a vital role in understanding cellular signaling pathways and disease mechanisms. Computational methods for this task can be categorized into kinase-family-focused and individual kinase-targeted approaches. Individual kinase-targeted methods have gained prominence for their ability to explore a broader protein space and provide more precise target information for kinase inhibitors. However, most existing individual kinase-based approaches focus solely on sequence inputs, neglecting crucial structural information. To address this limitation, we introduce SAGEPhos (Structure-aware kinAse-substrate bio-coupled and bio-auGmented nEtwork for Phosphorylation site prediction), a novel framework that modifies the semantic space of main protein inputs using auxiliary inputs at two distinct modality levels. At the inter-modality level, SAGEPhos introduces a Bio-Coupled Modal Fusion method, distilling essential kinase sequence information to refine task-oriented local substrate feature space, creating a shared semantic space that captures crucial kinase-substrate interaction patterns. Within the substrate's intra-modality domain, it focuses on Bio-Augmented Fusion, emphasizing 2D local sequence information while selectively incorporating 3D spatial information from predicted structures to complement the sequence space. Moreover, to address the lack of structural information in current datasets, we contribute a new, refined phosphorylation site prediction dataset, which incorporates crucial structural elements and will serve as a new benchmark for the field. Experimental results demonstrate that SAGEPhos significantly outperforms baseline methods, notably achieving almost 10\\% and 12\\% improvements in prediction accuracy and AUC-ROC, respectively. We further demonstrate our algorithm's robustness and generalization through stable results across varied data partitions and significant improvements in zero-shot scenarios. These results underscore the effectiveness of constructing a larger and more precise protein space in advancing the state-of-the-art in phosphorylation site prediction. We release the SAGEPhos models and code at https://github.com/ZhangJJ26/SAGEPhos.", "keywords": [ "Deep Learning", "Bioinformatics", "Phosphorylation prediction" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "hLZEbvDYhz", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Jun Sun", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Full Professor" }, { "name": "Long H. Pham", "gender": "unknown", "institution": "Singapore Management University", "country": "", "position": "Postdoc" }, { "name": "Tai D. Nguyen", "gender": "Male", "institution": "Singapore Management University", "country": "", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 23, "n_ref": 36, "n_ref_all": 56, "n_fig": 6, "n_tab": 4, "L_tab": 2262, "n_element_tab": 574, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 584, "n_element_tab_1": 112, "formula_len_all": 743, "formula_len_all_1": 747, "len_all": 115801, "len_all_1": 67872, "len_abs": 1567, "len_title": 103, "len_sents": 36315, "len_sents_1": 31075, "n_sents": 291, "n_sents_1": 278, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1121, "title": "UniAdapt: A Universal Adapter for Knowledge Calibration", "abs": "Large Language Models (LLMs) require frequent updates to correct errors and keep pace with continuously evolving knowledge in a timely and effective manner. Recent research in *it model editing* has highlighted the challenges in balancing generalization and locality, especially in the context of *lifelong model editing*. We discover that inserting knowledge directly into the model often causes conflicts and potentially disrupts other unrelated pre-trained knowledge. To address this problem, we introduce UniAdapt, a universal adapter for knowledge calibration. Inspired by the Mixture of Experts architecture and Retrieval-Augmented Generation, UniAdapt is designed with a vector-assisted router that is responsible for routing inputs to appropriate experts. The router maintains a vector store, including multiple shards, to construct routing vectors based on semantic similarity search results. UniAdapt is fully model-agnostic and designed for seamless plug-and-play integration. Experimental results show that UniAdapt outperforms existing lifelong model editors and achieves exceptional results in most metrics.", "keywords": [ "Lifelong model editing", "Mixture of experts", "Retrieval-Augmented Generation" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "hLIlN0f4ix", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Gertjan J. Burghouts", "gender": "Male", "institution": "TNO", "country": "NL", "position": "Researcher" }, { "name": "Jona Ruthardt", "gender": "Male", "institution": "University of Technology Nuremberg", "country": "DE", "position": "PhD student" }, { "name": "Serge Belongie", "gender": "Male", "institution": "University of Copenhagen", "country": "DK", "position": "Full Professor" }, { "name": "Yuki Asano", "gender": "Male", "institution": "University of Technology Nuremberg", "country": "DE", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 40, "n_ref": 77, "n_ref_all": 98, "n_fig": 6, "n_tab": 11, "L_tab": 7627, "n_element_tab": 1182, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2254, "n_element_tab_1": 345, "formula_len_all": 164, "formula_len_all_1": 233, "len_all": 154923, "len_all_1": 58032, "len_abs": 1123, "len_title": 94, "len_sents": 37575, "len_sents_1": 25613, "n_sents": 264, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 46, "L_abs": 1131, "title": "Do better language models have crisper vision?", "abs": "How well do text-only Large Language Models (LLMs) grasp the visual world? As LLMs are increasingly used in computer vision, addressing this question becomes both fundamental and pertinent. However, existing studies have primarily focused on limited scenarios, such as their ability to generate visual content or cluster multimodal data. To this end, we propose the Visual Text Representation Benchmark (ViTeRB) to isolate key properties that make language models well-aligned with the visual world. With this, we identify large-scale decoder-based LLMs as ideal candidates for representing text in vision-centric contexts, counter to the current practice of utilizing text encoders. Building on these findings, we propose ShareLock, an ultra-lightweight CLIP-like model. By leveraging precomputable frozen features from strong vision and language models, ShareLock achieves an impressive 51% accuracy on ImageNet despite utilizing just 563k image-caption pairs. Moreover, training requires only 1 GPU hour (or 10 hours including the precomputation of features) - orders of magnitude less than prior methods. Code will be released.", "keywords": [ "large language models", "computer vision", "vision-language model", "zero-shot image classification", "efficient deep learning" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "hL5jone2Oh", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Guangtao Zhai", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Wenjun Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Zijian Chen", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "tingzhu chen", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 76, "n_ref_all": 105, "n_fig": 13, "n_tab": 7, "L_tab": 5376, "n_element_tab": 483, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 3283, "n_element_tab_1": 257, "formula_len_all": 49, "formula_len_all_1": 0, "len_all": 207173, "len_all_1": 68200, "len_abs": 1632, "len_title": 115, "len_sents": 55691, "len_sents_1": 29807, "n_sents": 408, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1593, "title": "OBI-Bench: Can LMMs Aid in Study of Ancient Script on Oracle Bones?", "abs": "We introduce OBI-Bench, a holistic benchmark crafted to systematically evaluate large multi-modal models (LMMs) on whole-process oracle bone inscriptions (OBI) processing tasks demanding expert-level domain knowledge and deliberate cognition. OBI-Bench includes 5,523 meticulously collected diverse-sourced images, covering five key domain problems: recognition, rejoining, classification, retrieval, and deciphering. These images span centuries of archaeological findings and years of research by front-line scholars, comprising multi-stage font appearances from excavation to synthesis, such as original oracle bone, inked rubbings, oracle bone fragments, cropped single characters, and handprinted characters. Unlike existing benchmarks, OBI-Bench focuses on advanced visual perception and reasoning with OBI-specific knowledge, challenging LMMs to perform tasks akin to those faced by experts. The evaluation of 6 proprietary LMMs as well as 17 open-source LMMs highlights the substantial challenges and demands posed by OBI-Bench. Even the latest versions of GPT-4o, Gemini 1.5 Pro, and Qwen-VL-Max are still far from public-level humans in some fine-grained perception tasks. However, they perform at a level comparable to untrained humans in deciphering tasks, indicating remarkable capabilities in offering new interpretative perspectives and generating creative guesses. We hope OBI-Bench can facilitate the community to develop domain-specific multi-modal foundation models towards ancient language research and delve deeper to discover and enhance these untapped potentials of LMMs.", "keywords": [ "oracle bone inscriptions", "ancient character deciphering", "large multi-modal models", "benchmark" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "hKeHfOUCXL", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Chaoli Zhang", "gender": "Female", "institution": "Zhejiang Normal University", "country": "CN", "position": "Lecturer" }, { "name": "Leandro Von Krannichfeldt", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Liang Sun", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Staff Software Engineer" }, { "name": "Shirui Pan", "gender": "unknown", "institution": "Griffith University", "country": "AU", "position": "Full Professor" }, { "name": "Yi Wang", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Zhixian Wang", "gender": "Male", "institution": " University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Qingsong Wen", "gender": "Male", "institution": "Squirrel Ai Learning", "country": "US", "position": "Principal Researcher" } ], "n_formula": 12, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 101, "n_ref_all": 144, "n_fig": 20, "n_tab": 20, "L_tab": 5887, "n_element_tab": 659, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1973, "n_element_tab_1": 147, "formula_len_all": 876, "formula_len_all_1": 202, "len_all": 229456, "len_all_1": 67340, "len_abs": 3522, "len_title": 100, "len_sents": 81793, "len_sents_1": 31976, "n_sents": 684, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 1542, "title": "Benchmarks and Custom Package for Energy Forecasting", "abs": "Energy (load, wind, photovoltaic) forecasting is significant in the power industry as it can provide a reference for subsequent tasks such as power grid dispatch, thus bringing huge economic benefits. However, there are many differences between energy forecasting and traditional time series forecasting. On the one hand, traditional time series mainly focus on capturing characteristics like trends and cycles. In contrast, the energy series is largely influenced by many external factors, such as meteorological and calendar variables. On the other hand, energy forecasting aims to minimize the cost of subsequent tasks such as power grid dispatch, rather than simply pursuing prediction accuracy. In addition, the scale of energy data can also significantly impact the predicted results. In this paper, we collected large-scale load datasets and released a new renewable energy dataset that contains both station-level and region-level renewable generation data with meteorological data. For load data, we also included load domain-specific feature engineering and provided a method to customize the loss function and link the forecasting error to requirements related to subsequent tasks (such as power grid dispatching costs), integrating it into our forecasting framework. Based on such a situation, we conducted extensive experiments with 21 forecasting methods in these energy datasets at different levels under 11 evaluation metrics, providing a comprehensive reference for researchers to compare different energy forecasting models.", "keywords": [ "Energy Forecasting.+Benchmark.+Dataset" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "hKcDOfDxgn", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Huimiao Chen", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Jiyi Wang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Likai Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Marcelo G Mattar", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Sen Song", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 42, "n_ref": 62, "n_ref_all": 97, "n_fig": 24, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 21, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 307, "formula_len_all_1": 307, "len_all": 112336, "len_all_1": 61557, "len_abs": 786, "len_title": 117, "len_sents": 33052, "len_sents_1": 29547, "n_sents": 283, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 791, "title": "Brain-Like Replay Naturally Emerges in Reinforcement Learning Agents", "abs": "Replay is a powerful strategy to promote learning in artificial intelligence and the brain. However, the conditions to generate it and its functional advantages have not been fully recognized. In this study, we develop a modular reinforcement learning model that could generate replay. We prove that replay generated in this way helps complete the task. We also analyze the information contained in the representation and provide a mechanism for how replay makes a difference. Our design avoids complex assumptions and enables replay to emerge naturally within a task-optimized paradigm. Our model also reproduces key phenomena observed in biological agents. This research explores the structural biases in modular ANN to generate replay and its potential utility in developing efficient RL.", "keywords": [ "reinforcement learning", "replay", "neuroscience" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 2, 1, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 3, 2, 2, 2 ] }, { "paper_id": "hKZfzVZ999", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Hongyan Xu", "gender": "Female", "institution": "University of New South Wales", "country": "AU", "position": "PhD student" }, { "name": "Shan You", "gender": "Male", "institution": "SenseTime Research", "country": "CN", "position": "Researcher" }, { "name": "Xiu Su", "gender": "unknown", "institution": "Central South University", "country": "CN", "position": "Full Professor" }, { "name": "Xiuxing Li", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yijing Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yunlong Zhao", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 12, "n_ref_uni": 19, "n_ref": 30, "n_ref_all": 49, "n_fig": 7, "n_tab": 5, "L_tab": 4055, "n_element_tab": 342, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2630, "n_element_tab_1": 199, "formula_len_all": 839, "formula_len_all_1": 682, "len_all": 111266, "len_all_1": 51478, "len_abs": 1356, "len_title": 143, "len_sents": 25893, "len_sents_1": 20045, "n_sents": 221, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1364, "title": "MetaDD: Boosting Dataset Distillation with Neural Network Architecture-Invariant Generalization", "abs": "Dataset distillation (DD) entails creating a refined, compact distilled dataset from a large-scale dataset to facilitate efficient training. A significant challenge in DD is the dependency between the distilled dataset and the neural network (NN) architecture used. Training a different NN architecture with a distilled dataset distilled using a specific architecture often results in diminished trainning performance for other architectures. This paper introduces MetaDD, designed to enhance the generalizability of DD across various NN architectures. Specifically, MetaDD partitions distilled data into meta features (i.e., the data's common characteristics that remain consistent across different NN architectures) and heterogeneous features (i.e., the data's unique feature to each NN architecture). Then, MetaDD employs an architecture-invariant loss function for multi-architecture feature alignment, which increases meta features and reduces heterogeneous features in distilled data. As a low-memory consumption component, MetaDD can be seamlessly integrated into any DD methodology. Experimental results demonstrate that MetaDD significantly improves performance across various DD methods. On the Distilled Tiny-Imagenet with Sre2L (50 IPC), MetaDD achieves cross-architecture NN accuracy of up to 30.1\\%, surpassing the second-best method (GLaD) by 1.7\\%.", "keywords": [ "Dataset Distillation", "Class Active Map" ], "rating_list": [ 3, 3, 5, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 1, 1, 3 ], "contribution_list": [ 2, 2, 2, 4 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "hKMPz3wkPV", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Eric Elmoznino", "gender": "unknown", "institution": "Google", "country": "CA", "position": "Researcher" }, { "name": "Guillaume Lajoie", "gender": "Male", "institution": "Mila, Quebec AI institute", "country": "", "position": "Core Academic Member" }, { "name": "Thomas Jiralerspong", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "", "position": "MS student" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" } ], "n_formula": 32, "n_formula_1": 11, "n_ref_uni": 86, "n_ref": 169, "n_ref_all": 204, "n_fig": 11, "n_tab": 1, "L_tab": 209, "n_element_tab": 9, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2547, "formula_len_all_1": 982, "len_all": 288472, "len_all_1": 74295, "len_abs": 1276, "len_title": 91, "len_sents": 106465, "len_sents_1": 38076, "n_sents": 675, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1288, "title": "Towards a formal theory of compositionality", "abs": "Compositionality is believed to be fundamental to intelligence. In humans, it underlies the structure of thought, language, and higher-level reasoning. In AI, it enables a powerful form of out-of-distribution generalization, in which a model systematically adapts to novel combinations of known concepts. However, while we have strong intuitions about what compositionality is, there currently exists no formal definition for it that is measurable and mathematical. Here, we propose such a definition, which we call representational compositionality. The definition is conceptually simple, quantitative, and grounded in algorithmic information theory. Intuitively, representational compositionality states that a compositional representation is both expressive and describable as a simple function of discrete parts. We validate our definition on both real and synthetic data, and show how it unifies disparate intuitions from across the literature in both AI and cognitive science. We also show that representational compositionality, while theoretically intractable, can be readily estimated using standard deep learning tools. Our definition has the potential to inspire the design of novel, theoretically-driven models that better capture the mechanisms of higher-level human thought.", "keywords": [ "compositionality", "complexity", "deep learning", "representation", "generalization" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 4, 4, 1, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "hJVdwBpWjt", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "David Robinson", "gender": "Male", "institution": "Earth Species Project", "country": "AU", "position": "Researcher" }, { "name": "Marius Miron", "gender": "Male", "institution": "Earth Species Project", "country": "US", "position": "Researcher" }, { "name": "Masato Hagiwara", "gender": "Male", "institution": "Earth Species Project", "country": "", "position": "Senior AI Research Scientist" }, { "name": "Olivier Pietquin", "gender": "Male", "institution": "Earth Species Project", "country": "US", "position": "Chief Scientist" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 71, "n_ref": 118, "n_ref_all": 130, "n_fig": 1, "n_tab": 5, "L_tab": 1029, "n_element_tab": 109, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 886, "n_element_tab_1": 93, "formula_len_all": 71, "formula_len_all_1": 74, "len_all": 172807, "len_all_1": 59295, "len_abs": 1395, "len_title": 115, "len_sents": 34964, "len_sents_1": 28653, "n_sents": 243, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1376, "title": "NatureLM-audio: an Audio-Language Foundation Model for Bioacoustics", "abs": "Large language models (LLMs) prompted with text and audio represent the state of the art in various auditory tasks, including speech, music, and general audio, showing emergent abilities on unseen tasks. However, these capabilities have yet to be fully demonstrated in bioacoustics tasks, such as detecting animal vocalizations in large recordings, classifying rare and endangered species, and labeling context and behavior\u2014tasks that are crucial for conservation, biodiversity monitoring, and the study of animal behavior. In this work, we present NatureLM-audio, the first audio-language foundation model specifically designed for bioacoustics. Our carefully curated training dataset comprises text-audio pairs spanning a diverse range of bioacoustics, speech, and music data, designed to address the challenges posed by limited annotated datasets in the field. We demonstrate successful transfer of learned representations from music and speech to bioacoustics, and our model shows promising generalization to unseen taxa and tasks. Importantly, we test NatureLM-audio on a novel benchmark (BEANS-Zero) and it sets the new state of the art (SotA) on several bioacoustics tasks, including zero-shot classification of unseen species. To advance bioacoustics research, we also open-source the code for generating training and benchmark data, as well as for training the model.", "keywords": [ "audio-language foundation models", "multimodal large language models (llms)", "bioacoustics", "animal vocalizations", "zero-shot learning", "in-context learning" ], "rating_list": [ 3, 5, 8 ], "soundness_list": [ 2, 3, 4 ], "presentation_list": [ 3, 2, 4 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "hJIEtJlvhL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "HONGZE CHEN", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Jun Zhang", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Zehong Lin", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Research Assistant Professor" } ], "n_formula": 41, "n_formula_1": 21, "n_ref_uni": 34, "n_ref": 73, "n_ref_all": 118, "n_fig": 30, "n_tab": 10, "L_tab": 5331, "n_element_tab": 964, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 800, "n_element_tab_1": 173, "formula_len_all": 2357, "formula_len_all_1": 1390, "len_all": 173868, "len_all_1": 65104, "len_abs": 1621, "len_title": 132, "len_sents": 48441, "len_sents_1": 29410, "n_sents": 393, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1684, "title": "GI-GS: Global Illumination Decomposition on Gaussian Splatting for Inverse Rendering", "abs": "We present GI-GS, a novel inverse rendering framework that leverages 3D Gaussian Splatting (3DGS) and deferred shading to achieve photo-realistic novel view synthesis and relighting. In inverse rendering, accurately modeling the shading processes of objects is essential for achieving high-fidelity results. Therefore, it is critical to incorporate global illumination to account for indirect lighting that reaches an object after multiple bounces across the scene. Previous 3DGS-based methods have attempted to model indirect lighting by characterizing indirect illumination as learnable lighting volumes or additional attributes of each Gaussian, while using baked occlusion to represent shadow effects. These methods, however, fail to accurately model the complex physical interactions between light and objects, making it impossible to construct realistic indirect illumination during relighting. To address this limitation, we propose to calculate indirect lighting using efficient path tracing with deferred shading. In our framework, we first render a G-buffer to capture the detailed geometry and material properties of the scene. Then, we perform physically-based rendering (PBR) only for direct lighting. With the G-buffer and previous rendering results, the indirect lighting can be calculated through a lightweight path tracing. Our method effectively models indirect lighting under any given lighting conditions, thereby achieving better novel view synthesis and competitive relighting. Quantitative and qualitative results show that our GI-GS outperforms existing baselines in both rendering quality and efficiency. Project page: https://stopaimme.github.io/GI-GS-site/.", "keywords": [ "Inverse rendering", "Gaussian Splatting", "Relighting" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "hJDTuVQcQp", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Amin Arbabian", "gender": "Male", "institution": "Plato Systems", "country": "US", "position": "Founder" }, { "name": "Mert Pilanci", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Soheil Hor", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Ying Qian", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 34, "n_formula_1": 17, "n_ref_uni": 20, "n_ref": 26, "n_ref_all": 57, "n_fig": 5, "n_tab": 3, "L_tab": 891, "n_element_tab": 82, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 806, "n_element_tab_1": 69, "formula_len_all": 3259, "formula_len_all_1": 1049, "len_all": 109320, "len_all_1": 57286, "len_abs": 1497, "len_title": 121, "len_sents": 33850, "len_sents_1": 27424, "n_sents": 238, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1063, "title": "Adaptive Inference: Theoretical Limits and Opportunities for Efficient AI", "abs": "With the commercial deployment of increasingly larger and more complex neural networks at the cloud and the edge in recent years, inference has become too costly in terms of compute workload worldwide. Adaptive inference methods, which dynamically adjust a neural network's size or structure during inference, offer a means to enhance efficiency of neural networks beyond what static network compression and optimization methods can fundamentally achieve.\n\nThis paper introduces the first theoretical framework for quantifying the efficiency and performance gain opportunity size of adaptive inference algorithms. We provide new approximate and exact bounds for the achievable efficiency and performance gains, supported by empirical evidence demonstrating the potential for 10-100x efficiency improvements in both Computer Vision and Natural Language Processing tasks without incurring any performance penalties. Additionally, we offer insights on improving achievable efficiency gains through the optimal selection and design of adaptive inference state spaces.", "keywords": [ "Adaptive Inference", "Efficient ML", "Dynamic Neural Networks", "Dynamic Routing", "Computer Vision", "Natural Language processing" ], "rating_list": [ 1, 5, 3, 6, 6 ], "soundness_list": [ 1, 3, 1, 3, 4 ], "presentation_list": [ 1, 2, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3, 4 ] }, { "paper_id": "hJ8OQAiTrl", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 1, "author_info_list": [ { "name": "Junnan Liu", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "MS student" } ], "n_formula": 22, "n_formula_1": 14, "n_ref_uni": 81, "n_ref": 146, "n_ref_all": 163, "n_fig": 5, "n_tab": 8, "L_tab": 3378, "n_element_tab": 551, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 1473, "n_element_tab_1": 113, "formula_len_all": 2076, "formula_len_all_1": 1411, "len_all": 196143, "len_all_1": 52321, "len_abs": 1370, "len_title": 127, "len_sents": 35125, "len_sents_1": 20215, "n_sents": 266, "n_sents_1": 135, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1138, "title": "Combining Structure and Text: Learning Representations for Reasoning on Graphs", "abs": "Effective reasoning on real-world graphs necessitates a thorough understanding and optimal utilization of structural information from graph structure and textual information corresponding to nodes and edges. Recent research has primarily focused on two paradigms: employing graph neural networks to capture structural features and utilizing language models to process textual information, respectively. While these approaches have shown impressive performance, integrating structural and textual information presents significant challenges. To be more specific, concurrently training graph neural networks and language models is particularly challenging, primarily due to the scale of real-world graphs. This paper introduces a novel framework, named CoST, tailored for graph reasoning tasks. The proposed optimization objective enables alternating training of the GNN and PLM, leading to the generation of effective text representations by the PLM model, thereby enhancing the reasoning capabilities of the GNN model. Empirical results demonstrate that CoST achieves state-of-the-art performance across representative benchmark datasets.", "keywords": [ "graph reasoning", "structure representation", "text representation", "GNN", "PLM" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "hJ2BCYGvFg", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Anikait Singh", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Aviral Kumar", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Chelsea Finn", "gender": "Female", "institution": "Physical Intelligence", "country": "US", "position": "Researcher" }, { "name": "Jean Mercat", "gender": "Male", "institution": "Toyota Research Institute", "country": "US", "position": "Researcher" }, { "name": "Kushal Arora", "gender": "Male", "institution": "McGill University", "country": "", "position": "PhD student" }, { "name": "Tatsunori Hashimoto", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Sedrick Scott Keh", "gender": "Male", "institution": "Toyota Research Institute", "country": "", "position": "Research Engineer" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 16, "n_ref": 32, "n_ref_all": 42, "n_fig": 8, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 35, "n_element_tab_1": 3, "formula_len_all": 1199, "formula_len_all_1": 1123, "len_all": 159363, "len_all_1": 70593, "len_abs": 2082, "len_title": 118, "len_sents": 40626, "len_sents_1": 32567, "n_sents": 283, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1851, "title": "Improving the Efficiency of Test-Time Search in LLMs with Backtracking", "abs": "Solving reasoning problems is an iterative multi-step computation, where a reasoning agent progresses through a sequence of steps, with each step logically building upon the previous one to reach a desired conclusion. If the desired solution is not attained, the agent must backtrack and try reasoning chains that are quite different from previous attempts. Though prior work such as test-time search against an outcome verifier can improve performance, most search is done in parallel via Best-of-N reranking, and independently for each attempt at a problem, thus wasting a significant amount of computation in sampling multiple full solutions even beyond the point that is needed. Can we reduce the total amount of computation by sharing information and computation across multiple attempts to a given problem? In this paper, we build a novel approach combining process verifiers that predict likelihoods of success \\emph{per step} with preemptive backtracking to maximize performance per generated token. To do this, the PRM can be used to identify where a problematic step in a solution trace is by using the sensitivity of the predictions of the learned verifier and allowing the model to do focused resampling of the problematic portion of a solution. This approach can significantly reduce the amount of computation by leveraging partial computation from previous revisions. To further enhance the computational efficiency of inference, we introduce in-context process supervision, where the verifier is conditioned on the history of revisions that are attempted, reducing uncertainty in the verification decisions and improving the verifier's confidence with each round of backtracking. This framework for iterative backtracking, leveraging in-context process supervision, enables an effective tradeoff between inference and model performance.", "keywords": [ "LLMs", "Reasoning", "Test Time Inference", "Backtracking", "In-Context Verifiers" ], "rating_list": [ 3, 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 3, 2, 3 ], "confidence_list": [ 4, 2, 4, 2, 2 ] }, { "paper_id": "hJ1BaJ5ELp", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Haosen Shi", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zhanfeng Mo", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Sinno Pan", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" } ], "n_formula": 36, "n_formula_1": 17, "n_ref_uni": 65, "n_ref": 151, "n_ref_all": 200, "n_fig": 6, "n_tab": 16, "L_tab": 4571, "n_element_tab": 513, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 262, "n_element_tab_1": 16, "formula_len_all": 3315, "formula_len_all_1": 1266, "len_all": 240546, "len_all_1": 69182, "len_abs": 1246, "len_title": 136, "len_sents": 69367, "len_sents_1": 33349, "n_sents": 500, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 88, "L_abs": 1271, "title": "Probabilistic Neural Pruning via Sparsity Evolutionary Fokker-Planck-Kolmogorov Equation", "abs": "Neural pruning aims to compress and accelerate deep neural networks by identifying the optimal subnetwork within a specified sparsity budget. In this work, we study how to gradually sparsify the unpruned dense model to the target sparsity level with a minimal performance drop. Specifically, we analyze the evolution of the population of optimal subnetworks under continuous sparsity increments from a thermodynamics perspective. We first reformulate neural pruning as an expected loss minimization problem over the mask distributions. Then, we establish an effective approximation for the sparsity evolution of the optimal mask distribution, termed the **S**parsity Evolutionary **F**okker-**P**lanck-**K**olmogorov Equation (**SFPK**), which provides closed-form, mathematically tractable guidance on distributional transitions for minimizing the expected loss under an infinitesimal sparsity increment. On top of that, we propose SFPK-pruner, a particle simulation-based probabilistic pruning method, to sample performant masks with desired sparsity from the destination distribution of SFPK. In theory, we establish the convergence guarantee for the proposed SFPK-pruner. In practice, our SFPK-pruner exhibits competitive performance across various pruning scenarios.", "keywords": [ "Optimization for Deep Network", "Probabilistic Method", "Machine learning", "Model compression" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "hIdnWVxA9Z", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Hanwang Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Jiequan Cui", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Research Fellow" }, { "name": "Jingjing Chen", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Xue Song", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Yu-Gang Jiang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Zhongqi Yue", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 15, "n_ref": 38, "n_ref_all": 84, "n_fig": 16, "n_tab": 3, "L_tab": 320, "n_element_tab": 92, "n_fig_1": 15, "n_tab_1": 2, "L_tab_1": 244, "n_element_tab_1": 32, "formula_len_all": 560, "formula_len_all_1": 208, "len_all": 129558, "len_all_1": 68512, "len_abs": 1385, "len_title": 135, "len_sents": 40763, "len_sents_1": 30059, "n_sents": 330, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1391, "title": "Object Fusion via Diffusion Time-step for Customized Image Editing with Single Example", "abs": "We tackle the task of customized image editing using a text-conditioned Diffusion Model (DM). The goal is to fuse the subject in a reference image (e.g., sunglasses) with a source one (e.g., a boy), while retaining the fidelity of them both (e.g., the boy wearing the sunglasses). An intuitive approach, called LoRA fusion, first separately trains a DM LoRA for each image to encode its details. Then the two LoRAs are linearly combined by a weight to generate a fused image. Unfortunately, even through careful grid search or learning the weight, this approach still trades off the fidelity of one image against the other. We point out that the evil lies in the overlooked role of diffusion time-step in the generation process, i.e., a smaller time-step controls the generation of a more fine-grained attribute. For example, a large LoRA weight for the source may help preserve its fine-grained details (e.g., face attributes) at a small time-step, but could overpower the reference subject LoRA and lose the fidelity of its overall shape at a larger time-step. To address this deficiency, we propose TimeFusion, which learns a time-step-specific LoRA fusion weight that resolves the trade-off, i.e., generating the source and reference subject in high fidelity given their respective prompt. Then we can customize image editing using this weight and a target prompt. Codes are in Appendix.", "keywords": [ "object fusion", "diffusion time-step", "customized image editing" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "hIKsem01M5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Alexander Robey", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Hamed Hassani", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Assistant Professor" }, { "name": "Joshua Nathaniel Williams", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Naoki Murata", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Researcher" }, { "name": "Ruslan Salakhutdinov", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Yiding Jiang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Yuki Mitsufuji", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Lead Research Scientist" }, { "name": "Yutong He", "gender": "Female", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "George Pappas", "gender": "Male", "institution": "School of Engineering and Applied Science, University of Pennsylvania", "country": "US", "position": "Full Professor" }, { "name": "Zico Kolter", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 49, "n_ref": 102, "n_ref_all": 142, "n_fig": 21, "n_tab": 12, "L_tab": 8513, "n_element_tab": 589, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 6371, "n_element_tab_1": 336, "formula_len_all": 279, "formula_len_all_1": 558, "len_all": 262028, "len_all_1": 67875, "len_abs": 1006, "len_title": 128, "len_sents": 92085, "len_sents_1": 28191, "n_sents": 594, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1009, "title": "Automated Black-box Prompt Engineering for Personalized Text-to-Image Generation", "abs": "Prompt engineering is an effective but labor-intensive way to control text-to-image (T2I) generative models. Its time-intensive nature and complexity have spurred the development of algorithms for automated prompt generation. However, these methods often struggle with transferability across T2I models, require white-box access to the underlying model, or produce non-intuitive prompts. In this work, we introduce PRISM, an algorithm that automatically produces human-interpretable and transferable prompts that can effectively generate desired concepts given only black-box access to T2I models. Inspired by large language model (LLM) jailbreaking, PRISM leverages the in-context learning ability of LLMs to iteratively refine the candidate prompt distribution built upon the reference images. Our experiments demonstrate the versatility and effectiveness of PRISM in generating accurate prompts for objects, styles, and images across multiple T2I models, including Stable Diffusion, DALL-E, and Midjourney.", "keywords": [ "Text-to-Image Generation", "Prompt Engineering", "Personalized Text-to-Image Generation" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "hHNVn4hFPk", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Jiacheng Zhu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Justin Solomon", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Kristjan Greenewald", "gender": "unknown", "institution": "MIT-IBM Watson AI Lab, IBM Research", "country": "", "position": "Research Scientist" }, { "name": "Leshem Choshen", "gender": "Not Specified", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Mikhail Yurochkin", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Research Staff Member" }, { "name": "Onkar Bhardwaj", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Postdoc" }, { "name": "Rickard Br\u00fcel Gabrielsson", "gender": "Not Specified", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 39, "n_formula_1": 14, "n_ref_uni": 30, "n_ref": 61, "n_ref_all": 105, "n_fig": 7, "n_tab": 21, "L_tab": 27362, "n_element_tab": 1831, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 149, "n_element_tab_1": 13, "formula_len_all": 4417, "formula_len_all_1": 1097, "len_all": 267341, "len_all_1": 64866, "len_abs": 3087, "len_title": 124, "len_sents": 93103, "len_sents_1": 31351, "n_sents": 513, "n_sents_1": 229, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1104, "title": "Compress then Serve: Serving Thousands of LoRA Adapters with Little Overhead", "abs": "Fine-tuning large language models (LLMs) with low-rank adaptations (LoRAs) has become common practice, often yielding numerous copies of the same LLM differing only in their LoRA updates. This paradigm presents challenges for systems that serve real-time responses to queries that each involve a different LoRA. Prior works optimize the design of such systems but still require continuous loading and offloading of LoRAs, as it is infeasible to store thousands of LoRAs in GPU memory. To mitigate this issue, we investigate the efficacy of model compression when serving LoRAs. We propose a method for joint compression of LoRAs into a shared basis paired with LoRA-specific scaling matrices. We extend our algorithm to learn clusters of LoRAs that are more amenable to joint compression, allowing it to scale gracefully to large LoRA collections. Our experiments with up to 500 LoRAs demonstrate that compressed LoRAs preserve performance while offering major throughput gains in realistic serving scenarios with over a thousand LoRAs, maintaining 80\\% of the throughput of serving a \\emph{single} LoRA.", "keywords": [ "LoRA", "LLM", "PEFT", "Infrastructure", "NLP" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "hHF5AayC7O", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Christopher D. Manning", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "", "position": "Full Professor" }, { "name": "Dzmitry Bahdanau", "gender": "Male", "institution": "ServiceNow Research", "country": "", "position": "Research Scientist" }, { "name": "Shikhar Murty", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 86, "n_ref_all": 104, "n_fig": 1, "n_tab": 7, "L_tab": 1406, "n_element_tab": 156, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 748, "n_element_tab_1": 94, "formula_len_all": 124, "formula_len_all_1": 199, "len_all": 165323, "len_all_1": 57802, "len_abs": 1419, "len_title": 129, "len_sents": 55170, "len_sents_1": 26620, "n_sents": 491, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1438, "title": "NNetscape Navigator: Complex Demonstrations for Web Agents Without a Demonstrator", "abs": "We introduce NNetscape Navigator (NNetnav), a method for training web agents entirely through synthetic demonstrations. These demonstrations are collected by first interacting with a browser to generate trajectory rollouts, which are then retroactively labeled into instructions using a language model. Most work on training browser agents has relied on expensive human supervision, and the limited previous work on such \\emph{interaction-first} synthetic data techniques has failed to provide effective search through the exponential space of exploration. In contrast, NNetnav exploits the hierarchical structure of language instructions to make this search more tractable: complex instructions are typically decomposable into simpler subtasks, allowing NNetnav to automatically prune interaction episodes when an intermediate trajectory cannot be annotated with a meaningful sub-task. We use NNetnav demonstrations from a language model for supervised fine-tuning of a smaller language model policy, and find improvements of 6 points on WebArena and over 20 points on MiniWoB++, two popular environments for web-agents. Notably, on WebArena, we observe that language model policies can be further enhanced when fine-tuned with NNetnav demonstrations derived from the \\emph{same} language model. Finally, we collect and release a dataset of over 6k NNetnav demonstrations on WebArena, spanning a diverse and complex set of instructions.", "keywords": [ "language model agents", "large language models", "demonstrations for sequential decision making", "language conditioned RL", "grounded instruction following" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "hGcxiNUbjy", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Benjamin Smith", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "Undergrad student" }, { "name": "Edwin Zhang", "gender": "unknown", "institution": "OpenAI", "country": "US", "position": "Researcher" }, { "name": "Henry Gasztowtt", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "MS student" }, { "name": "Qinxun Bai", "gender": "Male", "institution": "Horizon Robotics Inc.", "country": "US", "position": "Senior Research Scientist" }, { "name": "Vincent Zhu", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "Undergrad student" } ], "n_formula": 8, "n_formula_1": 4, "n_ref_uni": 19, "n_ref": 45, "n_ref_all": 76, "n_fig": 12, "n_tab": 3, "L_tab": 606, "n_element_tab": 65, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 737, "n_element_tab_1": 9, "formula_len_all": 915, "formula_len_all_1": 401, "len_all": 225136, "len_all_1": 62398, "len_abs": 988, "len_title": 132, "len_sents": 88883, "len_sents_1": 29954, "n_sents": 764, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 782, "title": "Large Legislative Models: Towards Efficient AI Policymaking in Economic Simulations", "abs": "The improvement of economic policymaking presents an opportunity for broad societal benefit, a notion that has inspired research towards AI-driven policymaking tools. AI policymaking holds the potential to surpass human performance through the ability to process data quickly at scale. However, existing RL-based methods exhibit sample inefficiency, and are further limited by an inability to flexibly incorporate nuanced information into their decision-making processes. Thus, we propose a novel method in which we instead utilize pre-trained Large Language Models (LLMs), as sample-efficient policymakers in socially complex multi-agent reinforcement learning (MARL) scenarios. We demonstrate significant efficiency gains, outperforming existing methods across three environments.", "keywords": [ "large language models", "reinforcement learning", "policymaking" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "hGKuATIGhr", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Fatih Porikli", "gender": "Male", "institution": "QualComm", "country": "US", "position": "Senior Director" }, { "name": "Munawar Hayat", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "US", "position": "Researcher" }, { "name": "Risheek Garrepalli", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "US", "position": "Researcher" }, { "name": "Shweta Mahajan", "gender": "unknown", "institution": "Qualcomm Inc, QualComm", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 18, "n_ref": 28, "n_ref_all": 32, "n_fig": 1, "n_tab": 4, "L_tab": 732, "n_element_tab": 54, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2574, "n_element_tab_1": 118, "formula_len_all": 162, "formula_len_all_1": 161, "len_all": 76690, "len_all_1": 46961, "len_abs": 1286, "len_title": 109, "len_sents": 22781, "len_sents_1": 22834, "n_sents": 115, "n_sents_1": 114, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1355, "title": "DDIL: Improved Diffusion Distillation with Imitation Learning", "abs": "Diffusion models excel at generative modeling (e.g., text-to-image) but sampling requires multiple denoising network passes, limiting practicality. Diffusion distillation methods have shown promise by reducing the number of passes at the expense of quality of the generated samples but suffer from lack of diversity, quality, etc. . In this work we identify co-variate shift as one of reason for poor performance of multi-step distilled models from compounding error at inference time. To address co-variate shift, we formulate diffusion distillation within imitation learning **DDIL** framework and enhance training distribution for distilling diffusion models on both data distribution (forward diffusion) and student induced distributions (backward diffusion). Training on data distribution helps to diversify the generations by *preserving marginal data distribution* and training on student distribution addresses compounding error by *correcting covariate shift*. In addition, we adopt reflected diffusion formulation for distillation and demonstrate improved performance, stable training across different distillation methods. We show that DDIL and reflected diffusion formulation consistency improves on baseline algorithms of progressive distillation **(PD)**, Latent consistency models **(LCM)** and Distribution Matching Distillation **(DMD2)**", "keywords": [ "Diffusion Models", "Distillation", "Imitation Learning" ], "rating_list": [ 6, 3, 3, 5, 3 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 4, 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "hFQZmKFtlT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bishwamittra Ghosh", "gender": "Male", "institution": "MPI-SWS", "country": "DE", "position": "Postdoc" }, { "name": "Evimaria Terzi", "gender": "unknown", "institution": "Boston University, Boston University", "country": "", "position": "Full Professor" }, { "name": "Krishna P. Gummadi", "gender": "Male", "institution": "MPI-SWS", "country": "", "position": "Full Professor" }, { "name": "Mohammad Aflah Khan", "gender": "Male", "institution": "MPI-SWS", "country": "DE", "position": "Researcher" }, { "name": "Qinyuan Wu", "gender": "Female", "institution": "MPI-SWS", "country": "DE", "position": "PhD student" }, { "name": "Soumi Das", "gender": "Female", "institution": "MPI-SWS", "country": "DE", "position": "Postdoc" }, { "name": "Till Speicher", "gender": "Male", "institution": "MPI-SWS", "country": "", "position": "PhD student" }, { "name": "Vedant Nanda", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 27, "n_ref": 41, "n_ref_all": 56, "n_fig": 6, "n_tab": 2, "L_tab": 585, "n_element_tab": 27, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 599, "formula_len_all_1": 181, "len_all": 109598, "len_all_1": 61410, "len_abs": 1438, "len_title": 75, "len_sents": 39616, "len_sents_1": 31904, "n_sents": 282, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1448, "title": "Rethinking Memorization in LLMs: On Learning by Rote vs. with Understanding", "abs": "Understanding whether and to what extent token sequences generated by large language models (LLMs) are the result of regurgitating memorized training data or are based on meaningful learning of the training data's syntax and semantics has many important implications.\nIn order to cleanly measure and disentangle token recollection by rote (memorization) from generation with understanding, we create an experimental framework that is based on training LLMs over *sequences generated using formal grammars*. Our framework allows us to better understand the interplay between the two types of learning, namely, *by rote* vs. *with understanding*. Using our framework we make several striking observations that hold consistently across different open-source model families (Pythia, Llama, and Mistral): (a) we find that the learning types are at odds with each other during training, i.e., rote learning harms understanding and by developing understanding, models forget previously memorized sequences, (b) we find that *entropy of the training datasets* impacts the ease of learning, with lower entropy datasets being easier to learn with understanding and higher entropy datasets being easier to learn by rote, (c) we highlight the difficulty of determining the type of learning involved in a model based solely on recollecting a training data sequence. Our surprising results have significant downstream implications in the study and usage of LLMs.", "keywords": [ "language models", "memorization", "generalization" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 1 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "hFFAg5Dmw9", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Caiyu Wang", "gender": "Female", "institution": "University of Waikato", "country": "NZ", "position": "PhD student" }, { "name": "Chenhao Qiu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Jintang Li", "gender": "Male", "institution": "Sun Yat-sen University", "country": "CN", "position": "PhD student" }, { "name": "Minghui Li", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Qianglong Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Runsen Hua", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Shengshan Hu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yechao Zhang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 7, "n_ref_uni": 20, "n_ref": 66, "n_ref_all": 89, "n_fig": 6, "n_tab": 11, "L_tab": 2400, "n_element_tab": 395, "n_fig_1": 6, "n_tab_1": 8, "L_tab_1": 1583, "n_element_tab_1": 194, "formula_len_all": 2120, "formula_len_all_1": 920, "len_all": 213234, "len_all_1": 63537, "len_abs": 1332, "len_title": 81, "len_sents": 53496, "len_sents_1": 27644, "n_sents": 444, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1365, "title": "WISDOM: Progressive Curriculum Synthesis Makes LLMs Better Mathematical Reasoner", "abs": "Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of problem-solving tasks. Despite their success, LLMs still face significant challenges in complex reasoning, particularly with advanced mathematical problems. These problems require not only a deep understanding of task descriptions but also sophisticated logical and mathematical reasoning to determine the correct solution path, which is often lacking in the existing synthetic data. To address this gap, we introduce WISDOM, which draws inspiration from the human learning process and employs curriculum learning to gradually synthesize high-quality CoT data from easy to hard. Our goal is to guide LLM training and improve reasoning capabilities by progressively exposing models to increasingly challenging problems. Based on the synthesized data, we further fine-tune and develop the WISDOM series models, achieving significant improvements across multiple mathematical reasoning benchmarks. Notably, WISDOM-7B (DSMath) achieves a score of 62.4% on MATH, matching GPT-4\u2019s performance with 2/30 correct answers on AIME2024. Furthermore, WISDOM-70B (Llama3) outperforms GPT-4 on AIME2024 with 3/30 correct answers, demonstrating its potential as a better mathematical reasoner. More data and models will be available at https://anonymous.4open.science/r/Wisdom-math-377B", "keywords": [ "Large language models", "Mathematical reasoning", "data synthesis" ], "rating_list": [ 6, 6, 1, 6 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "hFB7XGcBeB", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Cabrel Teguemne Fokam", "gender": "Male", "institution": "Universit\u00e4t Bielefeld", "country": "DE", "position": "PhD student" }, { "name": "David Kappel", "gender": "Male", "institution": "Universit\u00e4t Bielefeld", "country": "DE", "position": "Assistant Professor" }, { "name": "Khaleelulla Khan Nazeer", "gender": "Male", "institution": "Technische Universit\u00e4t Dresden", "country": "DE", "position": "PhD student" }, { "name": "Lukas K\u00f6nig", "gender": "Male", "institution": "Universit\u00e4t Bielefeld", "country": "DE", "position": "Researcher" }, { "name": "Anand Subramoney", "gender": "Male", "institution": "Royal Holloway, University of London", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 20, "n_formula_1": 14, "n_ref_uni": 17, "n_ref": 28, "n_ref_all": 59, "n_fig": 6, "n_tab": 11, "L_tab": 2784, "n_element_tab": 371, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1882, "n_element_tab_1": 225, "formula_len_all": 1157, "formula_len_all_1": 653, "len_all": 113399, "len_all_1": 56238, "len_abs": 1631, "len_title": 142, "len_sents": 34179, "len_sents_1": 23068, "n_sents": 268, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1636, "title": "Asynchronous stochastic gradient descent with decoupled backpropagation and layer-wise updates", "abs": "The increasing size of deep learning models has created the need for more efficient alternatives to the standard error backpropagation algorithm, that make better use of asynchronous, parallel and distributed computing. One major shortcoming of backpropagation is the interlocking between the forward phase of the algorithm, which computes a global loss, and the backward phase where the loss is backpropagated through all layers to compute the gradients, which are used to update the network parameters. To address this problem, we propose a method that parallelises SGD updates across the layers of a model by asynchronously updating them from multiple threads. Furthermore, since we observe that the forward pass is often much faster than the backward pass, we use separate threads for the forward and backward pass calculations, which allows us to use a higher ratio of forward to backward threads than the usual 1:1 ratio, reducing the overall staleness of the parameters. Thus, our approach performs asynchronous stochastic gradient descent using separate threads for the loss (forward) and gradient (backward) computations and performs layer-wise partial updates to parameters in a distributed way. We show that this approach yields close to state-of-the-art results while running up to 2.97\u00d7 faster than Hogwild! scaled on multiple devices (Locally- Partitioned-Asynchronous-Parallel SGD).We theoretically prove the convergence of the algorithm using a novel theoretical framework based on stochastic differential equations and the drift diffusion process, by modeling the asynchronous parameter updates as a stochastic process.", "keywords": [ "asynchronous sgd", "layer-wise updates", "distributed training" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "hDkLpu1E64", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Hao Cui", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Michael Brenner", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "Professor" }, { "name": "Nayantara Mudur", "gender": "Female", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Paul Raccuglia", "gender": "Male", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Peter Christian Norgaard", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Subhashini Venugopalan", "gender": "unknown", "institution": "Google", "country": "US", "position": "Staff Research Scientist " } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 28, "n_ref": 54, "n_ref_all": 79, "n_fig": 7, "n_tab": 10, "L_tab": 4041, "n_element_tab": 307, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 84, "formula_len_all_1": 23, "len_all": 229071, "len_all_1": 58619, "len_abs": 6102, "len_title": 118, "len_sents": 72010, "len_sents_1": 29539, "n_sents": 565, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1523, "title": "FEABench: Evaluating Language Models on Real World Physics Reasoning Ability", "abs": "Building precise simulations of the real world and invoking numerical solvers to answer quantitative problems is an essential requirement in engineering and science. We present FEABench, a benchmark to evaluate the ability of large language models (LLMs) and LLM agents to simulate and solve physics, mathematics and engineering problems using finite element analysis (FEA). We introduce a multipronged evaluation scheme to investigate the ability of LLMs to solve these problems by reasoning over natural language problem descriptions and operating COMSOL Multiphysics$^\\textregistered$, an FEA software, to compute the answers. In addition to testing state-of-the art-LLMs, we further design a language model agent equipped with the ability to interact with the software through its Application Programming Interface (API), examine its outputs and use tools to improve its solutions over multiple iterations. Our best performing strategy generates executable API calls 88\\% of the time. However, this benchmark still proves to be challenging enough that the LLMs and agents we tested were not able to completely and correctly solve any problem. LLMs that can successfully interact with and operate FEA software to solve problems such as those in our benchmark would significantly push the frontiers of their utility. Acquiring this capability would augment LLMs' reasoning skills with the precision of numerical solvers and advance the development of autonomous systems that can tackle complex problems in the real world.", "keywords": [ "numerical analysis", "finite element", "benchmark", "agents" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "hDPwaYVxBx", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Ashkan Khakzar", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Daniel Rueckert", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Jindong Gu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Razieh Rezaei", "gender": "Female", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Masoud Jalili Sabet", "gender": "Male", "institution": "Volkswagen Group", "country": "DE", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 31, "n_ref": 58, "n_ref_all": 83, "n_fig": 12, "n_tab": 5, "L_tab": 1599, "n_element_tab": 200, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1037, "n_element_tab_1": 128, "formula_len_all": 234, "formula_len_all_1": 234, "len_all": 122148, "len_all_1": 63397, "len_abs": 1168, "len_title": 116, "len_sents": 35795, "len_sents_1": 29284, "n_sents": 253, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1058, "title": "Learning Visual Prompts for Guiding the Attention of Vision Transformers", "abs": "to be completed laterVisual prompting infuses visual information into the input image to adapt models toward specific predictions and tasks. Recently, manually crafted markers such as red circles are shown to guide the model to attend to a target region on the image. However, these markers only work on models trained with data containing those markers. Moreover, finding these prompts requires guesswork or prior knowledge of the domain on which the model is trained. This work circumvents manual design constraints by proposing to learn the visual prompts for guiding the attention of vision transformers. The learned visual prompt, added to any input image would redirect the attention of the pre-trained vision transformer to its spatial location on the image. Specifically, the prompt is learned in a self-supervised manner without requiring annotations and without fine-tuning the vision transformer. Our experiments demonstrate the effectiveness of the proposed optimization-based visual prompting strategy across various pre-trained vision encoders.", "keywords": [ "vision transformers", "attention", "visual prompting", "universal adversarial patch", "universal adversarial transferability" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "hDBrQ4DApF", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Kai Wang", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Xiaolong Jin", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "MS student" }, { "name": "Xuanlei Zhao", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 51, "n_ref": 122, "n_ref_all": 159, "n_fig": 14, "n_tab": 13, "L_tab": 3388, "n_element_tab": 649, "n_fig_1": 11, "n_tab_1": 6, "L_tab_1": 805, "n_element_tab_1": 171, "formula_len_all": 422, "formula_len_all_1": 249, "len_all": 181969, "len_all_1": 55439, "len_abs": 855, "len_title": 107, "len_sents": 43950, "len_sents_1": 23246, "n_sents": 386, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 856, "title": "Real-Time Video Generation with Pyramid Attention Broadcast", "abs": "We present Pyramid Attention Broadcast (PAB), a real-time, high quality and training-free approach for DiT-based video generation. Our method is founded on the observation that attention difference in the diffusion process exhibits a U-shaped pattern, indicating significant redundancy. We mitigate this by broadcasting attention outputs to subsequent steps in a pyramid style. It applies different broadcast strategies to each attention based on their variance for best efficiency. We further introduce broadcast sequence parallel for more efficient distributed inference. PAB demonstrates superior results across three models compared to baselines, achieving real-time generation for up to 720p videos. We anticipate that our simple yet effective method will serve as a robust baseline and facilitate future research and application for video generation.", "keywords": [ "Diffusion Acceleration", "DiT", "Video Generation", "Efficient", "Real-Time", "Parallelism", "Sequence Parallelism" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "hCfhfwSfCg", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Di ZHANG", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "VP" }, { "name": "Fuzheng Zhang", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Principal Researcher" }, { "name": "Kun Gai", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Instructor" }, { "name": "Xi Yuan", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "YAN ZHENG", "gender": "Male", "institution": "Tianjin Unibersity, China", "country": "CN", "position": "Associate Professor" }, { "name": "Hangyu Mao", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 18, "n_formula_1": 11, "n_ref_uni": 40, "n_ref": 70, "n_ref_all": 76, "n_fig": 3, "n_tab": 3, "L_tab": 1344, "n_element_tab": 117, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1344, "n_element_tab_1": 117, "formula_len_all": 1688, "formula_len_all_1": 1323, "len_all": 107096, "len_all_1": 68282, "len_abs": 923, "len_title": 109, "len_sents": 27029, "len_sents_1": 24330, "n_sents": 221, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 929, "title": "Generate explorative goals with large language model guidance", "abs": "Reinforcement learning (RL) struggles with sparse reward environments.\nRecent developments in intrinsic motivation have revealed the potential of language models to guide agents in exploring the environment.\nHowever, the mismatch between the granularity of environment transitions and natural language descriptions hinders effective exploration for current methods.\nTo address this problem, we introduce a model-based RL method named Language-Guided Explorative Goal Generation (LanGoal), which combines large language model (LLM) guidance with intrinsic exploration reward by learning to propose meaningful goals.\nLanGoal learns a hierarchical policy together with a world model. The high-level policy learns to propose goals based on LLM guidance to explore the environment, and the low-level policy learns to achieve the goals.\nExtensive results on Crafter demonstrate the effectiveness of LanGoal compared to recent methods.", "keywords": [ "Reinforcement Learning", "Large Language Models", "Goal-Conditioned RL", "Exploration", "Model-baed Reinforcement Learning" ], "rating_list": [ 3, 3, 1, 1 ], "soundness_list": [ 2, 2, 1, 1 ], "presentation_list": [ 1, 1, 1, 1 ], "contribution_list": [ 3, 2, 1, 1 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "hBVywPrw6d", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Li Yi", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yunze Liu", "gender": "Male", "institution": "IIIS, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 19, "n_ref": 44, "n_ref_all": 61, "n_fig": 3, "n_tab": 4, "L_tab": 2225, "n_element_tab": 376, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 222, "n_element_tab_1": 26, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 129443, "len_all_1": 63260, "len_abs": 1244, "len_title": 156, "len_sents": 32882, "len_sents_1": 30640, "n_sents": 262, "n_sents_1": 249, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1249, "title": "MAP: Unleashing Hybrid Mamba-Transformer Vision Backbone\u2019s Potential with Masked Autoregressive Pretraining", "abs": "Mamba has achieved significant advantages in long-context modeling and autoregressive tasks, but its scalability with large parameters remains a major limitation in vision applications. pretraining is a widely used strategy to enhance backbone model performance. Although the success of Masked Autoencoder in Transformer pretraining is well recognized, it does not significantly improve Mamba's visual learning performance. We found that using the correct autoregressive pretraining can significantly boost the performance of the Mamba architecture. Based on this analysis, we propose Masked Autoregressive Pretraining(MAP) to pretrain a hybrid Mamba-Transformer vision backbone network. This strategy combines the strengths of both MAE and Autoregressive pretraining, improving the performance of Mamba and Transformer modules within a unified paradigm. Experimental results show that both the pure Mamba architecture and the hybrid Mamba-Transformer vision backbone network pretrained with MAP significantly outperform other pretraining strategies, achieving state-of-the-art performance. We validate the effectiveness of the method on both 2D and 3D datasets and provide detailed ablation studies to support the design choices for each component.", "keywords": [ "Hybrid Vision Backbone", "Masked Autoregressive", "Self-supervised Learning" ], "rating_list": [ 3, 5, 3, 6, 8 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 1, 3, 1, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 5, 4, 5, 3, 5 ] }, { "paper_id": "hBGavkf61a", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Byeonghu Na", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Il-chul Moon", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Kwanghyeon Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Minsang Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "MS student" }, { "name": "Yeongmin Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 87, "n_formula_1": 23, "n_ref_uni": 73, "n_ref": 223, "n_ref_all": 340, "n_fig": 16, "n_tab": 12, "L_tab": 2994, "n_element_tab": 379, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 1595, "n_element_tab_1": 172, "formula_len_all": 11964, "formula_len_all_1": 1870, "len_all": 262846, "len_all_1": 78264, "len_abs": 1446, "len_title": 118, "len_sents": 71410, "len_sents_1": 32818, "n_sents": 619, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 70, "L_abs": 1570, "title": "Diffusion Bridge AutoEncoders for Unsupervised Representation Learning", "abs": "Diffusion-based representation learning has achieved substantial attention due to its promising capabilities in latent representation and sample generation. Recent studies have employed an auxiliary encoder to identify a corresponding representation from data and to adjust the dimensionality of a latent variable $\\mathbf{z}$. Meanwhile, this auxiliary structure invokes an *information split problem*; the information of each data instance $\\mathbf{x}_0$ is divided into diffusion endpoint $\\mathbf{x}_T$ and encoded $\\mathbf{z}$ because there exist two inference paths starting from the data. The latent variable modeled by diffusion endpoint $\\mathbf{x}_T$ has some disadvantages. The diffusion endpoint $\\mathbf{x}_T$ is computationally expensive to obtain and inflexible in dimensionality. To address this problem, we introduce Diffusion Bridge AuteEncoders (DBAE), which enables $\\mathbf{z}$-dependent endpoint $\\mathbf{x}_T$ inference through a feed-forward architecture. This structure creates an information bottleneck at $\\mathbf{z}$, so $\\mathbf{x}_T$ becomes dependent on $\\mathbf{z}$ in its generation. This results in $\\mathbf{z}$ holding the full information of data. We propose an objective function for DBAE to enable both reconstruction and generative modeling, with their theoretical justification. Empirical evidence supports the effectiveness of the intended design in DBAE, which notably enhances downstream inference quality, reconstruction, and disentanglement. Additionally, DBAE generates high-fidelity samples in the unconditional generation.", "keywords": [ "Diffusion Model", "Represenation Learning", "Autoencoders" ], "rating_list": [ 5, 8, 8, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "hB6jYbvypa", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 9, "author_info_list": [ { "name": "An Xu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Cong Xie", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "PhD student" }, { "name": "Ding Zhou", "gender": "Male", "institution": "Bytedance", "country": "US", "position": "Researcher" }, { "name": "Xin Liu", "gender": "Male", "institution": "ByteDance Inc.", "country": "", "position": "Researcher" }, { "name": "Xue Lin", "gender": "Female", "institution": "Northeastern University", "country": "", "position": "Assistant Professor" }, { "name": "Yanyue Xie", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Yanzhi Wang", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Associate Professor" }, { "name": "Zhi Zhang", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Ziang Song", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 64, "n_ref": 118, "n_ref_all": 127, "n_fig": 4, "n_tab": 8, "L_tab": 4819, "n_element_tab": 863, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 4762, "n_element_tab_1": 432, "formula_len_all": 858, "formula_len_all_1": 901, "len_all": 172301, "len_all_1": 62784, "len_abs": 1333, "len_title": 139, "len_sents": 30751, "len_sents_1": 21940, "n_sents": 252, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1105, "title": "MoE-Pruner: Pruning Mixture-of-Experts Large Language Model using the Hints from Its Router", "abs": "Mixture-of-Experts (MoE) architectures face challenges such as high memory consumption and redundancy in experts. Pruning MoE can reduce network weights while maintaining model performance. Motivated by the recent observation of emergent large magnitude features in Large Language Models (LLM) and MoE routing policy, we propose MoE-Pruner, a method that prunes weights with the smallest magnitudes multiplied by the corresponding input activations and router weights, on each output neuron. Our pruning method is one-shot, requiring no retraining or weight updates. We evaluate our method on Mixtral-8x7B and Mixtral-8x22B across multiple language benchmarks. Experimental results show that our method significantly outperforms state-of-the-art LLM pruning methods. Furthermore, our pruned MoE models can benefit from a pretrained teacher model through expert-wise knowledge distillation, improving performance post-pruning. Experimental results demonstrate that the Mixtral-8x7B model with 50% sparsity maintains 99% of the performance of the original model after the expert-wise knowledge distillation.", "keywords": [ "mixture-of-experts", "large language models", "sparsity", "network pruning", "knowledge distillation" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "hAyw43h0MH", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Chenyu Xu", "gender": "Male", "institution": "Iowa State University", "country": "", "position": "PhD student" }, { "name": "Shuangyou Feng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Sihai Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "Associate Professor" }, { "name": "Zhi Liang", "gender": "Male", "institution": "Kangma Biotech", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 59, "n_formula_1": 17, "n_ref_uni": 29, "n_ref": 47, "n_ref_all": 72, "n_fig": 7, "n_tab": 3, "L_tab": 280, "n_element_tab": 19, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 622, "n_element_tab_1": 8, "formula_len_all": 6027, "formula_len_all_1": 996, "len_all": 119692, "len_all_1": 55876, "len_abs": 1597, "len_title": 97, "len_sents": 36213, "len_sents_1": 24744, "n_sents": 316, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1599, "title": "Smoothness Bridges Sparsity and Stability in MoEs", "abs": "Mixture of experts architectures have recently emerged as an effective approach for scaling model capacity while managing computational costs by leveraging expert sparsity, where only a subset of experts is activated during inference. Despite their computational efficiency, MoE models face challenges in training stability compared to their dense counterparts, largely due to the introduction of expert sparsity. While several methods have been proposed to mitigate this instability, the underlying relationship between expert sparsity and training stability remains unclear. In this work, we develop a theoretical framework that demonstrates an inverse correlation between training stability and expert sparsity, with gradient smoothness serving as the bridge. We derive an upper bound on training stability, formalizing for the first time the sparsity-stability trade-off in MoE models. Our findings show that activating more experts enhances gradient smoothness and improves training stability but at the cost of reduced sparsity. We validate our theory through extensive experiments on various architectures and datasets, and propose a novel MoE structure that addresses stability without sacrificing sparsity. This design introduces independent router heads and a soft top-$K$ selection via sampling without replacement, which smooths the gradient landscape while maintaining expert sparsity. Further analysis confirms the promise of this structure in striking the optimal balance between sparsity and stability, offering a new direction for optimizing MoE architectures in large-scale models.", "keywords": [ "Mixture-of-Experts (MoE)", "Model Sparsity", "Training Stability" ], "rating_list": [ 3, 3, 5, 1 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "h9dnHqrkfa", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Junhao Zheng", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Qianli Ma", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Wenxuan Shen", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Wenzheng Xie", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Yu Cheng", "gender": "unknown", "institution": "South China University of Technology", "country": "", "position": "MS student" }, { "name": "Wu Binquan", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 50, "n_formula_1": 16, "n_ref_uni": 38, "n_ref": 125, "n_ref_all": 153, "n_fig": 6, "n_tab": 10, "L_tab": 4188, "n_element_tab": 253, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 561, "n_element_tab_1": 25, "formula_len_all": 3540, "formula_len_all_1": 809, "len_all": 191588, "len_all_1": 67096, "len_abs": 1493, "len_title": 142, "len_sents": 59959, "len_sents_1": 28736, "n_sents": 529, "n_sents_1": 282, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1500, "title": "Conditional Information Bottleneck Approach for Out-of-Distribution Sequential Recommendation", "abs": "Sequential recommendation (SR) aims to suggest items users are most likely to engage with next based on their past interactions. However, in practice, SR systems often face the out-of-distribution (OOD) problem due to dynamic environmental factors (e.g., seasonal changes), leading to significant performance degradation in the testing phase. \nSome methods incorporate distributionally robust optimization (DRO) into SR to alleviate OOD, but the sparsity of SR data challenges this. Other approaches use random data augmentations to explore the OOD, potentially distorting important information, as user behavior is personalized rather than random. Additionally, they often overlook users' varying sensitivity to distribution shifts during the exploration, which is crucial for capturing the evolution of user preferences in OOD contexts.\nIn this work, inspired by information bottleneck theory (IB), we propose the Conditional Distribution Information Bottleneck (CDIB), a novel objective that creates diverse OOD distributions while preserving minimal sufficient information regarding the origin distribution conditioned on the user. Building on this, we introduce a framework with a learnable, personalized data augmentation method using a mask-then-generate paradigm to craft diverse and reliable OOD distributions optimized with CDIB. Experiments on four real-world datasets show our model consistently outperforms baselines. The code is available at https://anonymous.4open.science/r/CDIB-51C8.", "keywords": [ "Sequential Recommendation", "Out of Distribution", "Robust", "Conditional Information Bottleneck" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "h8yg0hT96f", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Christophe Heinkel\u00e9", "gender": "Male", "institution": "Cerema Endsum", "country": "", "position": "Researcher" }, { "name": "Florence Forbes", "gender": "Female", "institution": "INRIA", "country": "", "position": "Researcher" }, { "name": "Jacopo Iollo", "gender": "Male", "institution": "INRIA Rhone-Alpes", "country": "", "position": "PhD student" }, { "name": "Pierre Alliez", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Principal Researcher" } ], "n_formula": 74, "n_formula_1": 31, "n_ref_uni": 53, "n_ref": 112, "n_ref_all": 156, "n_fig": 8, "n_tab": 3, "L_tab": 84, "n_element_tab": 10, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 810, "n_element_tab_1": 13, "formula_len_all": 6481, "formula_len_all_1": 2276, "len_all": 185927, "len_all_1": 72876, "len_abs": 1176, "len_title": 103, "len_sents": 56051, "len_sents_1": 31404, "n_sents": 468, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 55, "L_abs": 1252, "title": "Bayesian Experimental Design Via Contrastive Diffusions", "abs": "Bayesian Optimal Experimental Design (BOED) is a powerful tool to reduce the cost of running a sequence of experiments.\nWhen based on the Expected Information Gain (EIG), design optimization corresponds to the maximization of some intractable expected *contrast* between prior and posterior distributions.\nScaling this maximization to high dimensional and complex settings has been an issue due to BOED inherent computational complexity.\nIn this work, we introduce an *expected posterior* distribution with cost-effective sampling properties and provide a tractable access to the EIG contrast maximization via a new EIG gradient expression. Diffusion-based samplers are used to compute the dynamics of the expected posterior and ideas from bi-level optimization are leveraged to derive an efficient joint sampling-optimization loop, without resorting to lower bound approximations of the EIG. The resulting efficiency gain allows to extend BOED to the well-tested generative capabilities of diffusion models. \nBy incorporating generative models into the BOED framework, we expand its scope and its use in scenarios that were previously impractical. Numerical experiments and comparison with state-of-the-art methods show the potential of the approach.", "keywords": [ "Bayesian Optimal Experimental Design", "Conditional Diffusion Models", "score based sampling", "Bayesian Inverse Problems", "Experimental Design", "Sampling as Optimization" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 2 ] }, { "paper_id": "h7fZvaU93L", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 13, "author_info_list": [ { "name": "Adam Scibior", "gender": "unknown", "institution": "University of British Columbia", "country": "", "position": "Researcher" }, { "name": "Berend Zwartsenberg", "gender": "Male", "institution": "Inverted AI", "country": "CA", "position": "Researcher" }, { "name": "Dylan Green", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Frank Wood", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Full Professor" }, { "name": "Jonathan Wilder Lavington", "gender": "Male", "institution": "University of British Columbia", "country": "", "position": "PhD student" }, { "name": "Ke Zhang", "gender": "unknown", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Matthew Niedoba", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Saeid Naderiparizi", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Setareh Dabiri", "gender": "unknown", "institution": "Inverted AI", "country": "", "position": "Researcher" }, { "name": "Vasileios Lioutas", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "William Harvey", "gender": "Male", "institution": "University of British Columbia", "country": "", "position": "PhD student" }, { "name": "Xiaoxuan Liang", "gender": "Female", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Yunpeng Liu", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 38, "n_ref": 78, "n_ref_all": 117, "n_fig": 21, "n_tab": 7, "L_tab": 2839, "n_element_tab": 415, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 781, "n_element_tab_1": 149, "formula_len_all": 769, "formula_len_all_1": 302, "len_all": 144293, "len_all_1": 60331, "len_abs": 1020, "len_title": 122, "len_sents": 43160, "len_sents_1": 28330, "n_sents": 310, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1022, "title": "Semantically Consistent Video Inpainting with Conditional Diffusion Models", "abs": "Current state-of-the-art methods for video inpainting typically rely on optical flow or attention-based approaches to inpaint masked regions by propagating visual information across frames. While such approaches have led to significant progress on standard benchmarks, they struggle with tasks that require the synthesis of novel content that is not present in other frames. In this paper, we reframe video inpainting as a conditional generative modeling problem and present a framework for solving such problems with conditional video diffusion models. We introduce inpainting-specific sampling schemes which capture crucial long-range dependencies in the context, and devise a novel method for conditioning on the known pixels in incomplete frames. We highlight the advantages of using a generative approach for this task, showing that our method is capable of generating diverse, high-quality inpaintings and synthesizing new content that is spatially, temporally, and semantically consistent with the provided context.", "keywords": [ "diffusion models", "video inpainting", "conditional generative modeling" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 3, 2 ] }, { "paper_id": "h7aQxzKbq6", "primary_area": "applications to robotics, autonomy, planning", "n_author": 11, "author_info_list": [ { "name": "Abhishek Gupta", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "Ankit Goyal", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Anqi Li", "gender": "unknown", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Caelan Reed Garrett", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Dieter Fox", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Senior Director of Robotics Research" }, { "name": "Fabio Ramos", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "Professor" }, { "name": "Jesse Zhang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Joel Jang", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Marius Memmel", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Yi Li", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Yuquan Deng", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 48, "n_ref": 137, "n_ref_all": 183, "n_fig": 12, "n_tab": 5, "L_tab": 403, "n_element_tab": 45, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 283, "formula_len_all_1": 224, "len_all": 288569, "len_all_1": 72198, "len_abs": 1920, "len_title": 117, "len_sents": 63049, "len_sents_1": 36440, "n_sents": 446, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1810, "title": "HAMSTER: Hierarchical Action Models for Open-World Robot Manipulation", "abs": "Large models have shown strong open-world generalization to complex problems in vision and language, but they have been relatively more difficult to deploy in robotics. This challenge stems from several factors, the foremost of which is the lack of scalable robotic training data since this requires expensive on-robot collection. For scalable training, these models must show considerable transfer across domains, to make use of cheaply available \"off-domain\" data such as videos, hand-drawn sketches, or data from simulation. In this work, we posit that hierarchical vision-language-action models can be more effective at transferring behavior across domains than standard monolithic vision-language-action models. In particular, we study a class of hierarchical vision-language-action models, where high-level vision-language models (VLMs) are trained on relatively cheap data to produce semantically meaningful intermediate predictions such as 2D paths indicating desired behavior. These predicted 2D paths can serve as guidance for low-level control policies that are 3D-aware and capable of precise manipulation. In this work, we show that separating prediction into semantic high-level predictions, and 3D-aware low-level predictions allows such hierarchical VLA policies to transfer across significant domain gaps, for instance from simulation to the real world or across scenes with widely varying visual appearance. Doing so allows for the usage of cheap, abundant data sources beyond teleoperated on-robot data thereby enabling broad semantic and visual generalization. We demonstrate how hierarchical architectures trained on this type of cheap off-domain data can enable robotic manipulation with semantic, visual, and geometric generalization through experiments in simulation and the real world.", "keywords": [ "vision language model; cross-domain generalization; sim-to-real transfer; robot manipulation; vision language action model" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "h7Qz1ulnvF", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 10, "author_info_list": [ { "name": "Chen Shen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Chenxi Huang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Ge Teng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" }, { "name": "Jieping Ye", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Liang Xie", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "Postdoc" }, { "name": "Sinan Fan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Wenxiao Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaofei He", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" }, { "name": "Xiaosong Yuan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" }, { "name": "Xiao Feng Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 21, "n_ref": 42, "n_ref_all": 59, "n_fig": 9, "n_tab": 2, "L_tab": 373, "n_element_tab": 74, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 471, "n_element_tab_1": 85, "formula_len_all": 642, "formula_len_all_1": 615, "len_all": 113810, "len_all_1": 54432, "len_abs": 1747, "len_title": 144, "len_sents": 33798, "len_sents_1": 26084, "n_sents": 224, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1749, "title": "Improving Complex Reasoning with Dynamic Prompt Corruption: A Soft Prompt Optimization Approach", "abs": "Prompt Tuning (PT) has emerged as a promising Parameter-Efficient Fine-Tuning (PEFT) approach by appending trainable continuous prompt vectors to the input, maintaining competitive performance with significantly fewer trainable parameters. While PT has shown effectiveness in enhancing task performance, particularly for classification tasks, its application to complex reasoning tasks has been largely overlooked. Our investigation reveals that PT provides limited improvement and may even degrade performance in reasoning tasks. This phenomenon suggests that soft prompts can positively impact certain instances while negatively affecting others, particularly during the latter stages of reasoning.\nTo address these challenges, we propose a novel method called Dynamic Prompt Corruption (DPC), which seeks to optimize the use of soft prompts in reasoning tasks. DPC dynamically adjusts the influence of soft prompts based on their impact on the reasoning process. Specifically, it involves two key components: Dynamic Trigger and Dynamic Corruption. Dynamic Trigger measures the influence of soft prompts, determining whether their impact is beneficial or detrimental. Dynamic Corruption mitigates the negative effects of soft prompts by selectively masking key tokens that interfere with the reasoning process.\nWe validate our approach through extensive experiments on various large language models (LLMs) and reasoning tasks, including GSM8K, MATH, and AQuA. The results demonstrate that Dynamic Prompt Corruption consistently improves the performance of LLMs, achieving 4\\%-8\\% accuracy gains compared to standard prompt tuning. These findings highlight the effectiveness of our approach and its potential to enhance complex reasoning in LLMs.", "keywords": [ "prompt tuning", "corruption", "information flow" ], "rating_list": [ 6, 6, 8, 3 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 2, 3, 4, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "h7GAgbLSmC", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Arya Mazumdar", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Christos Thrampoulidis", "gender": "unknown", "institution": "University of British Columbia", "country": "", "position": "Assistant Professor" }, { "name": "Hossein Taheri", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Postdoc" } ], "n_formula": 24, "n_formula_1": 25, "n_ref_uni": 30, "n_ref": 77, "n_ref_all": 81, "n_fig": 0, "n_tab": 3, "L_tab": 441, "n_element_tab": 34, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 441, "n_element_tab_1": 34, "formula_len_all": 1175, "formula_len_all_1": 1158, "len_all": 108761, "len_all_1": 77998, "len_abs": 1720, "len_title": 129, "len_sents": 33888, "len_sents_1": 32147, "n_sents": 280, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1781, "title": "Sharper Guarantees for Learning Neural Network Classifiers with Gradient Methods", "abs": "In this paper, we study the data-dependent convergence and generalization behavior of gradient methods for neural networks with smooth activation. Our first result is a novel bound on the excess risk of deep networks trained by the logistic loss via an alogirthmic stability analysis. Compared to previous works, our results improve upon the shortcomings of the well-established Rademacher complexity-based bounds. Importantly, the bounds we derive in this paper are tighter, hold even for neural networks of small width, do not scale unfavorably with width, are algorithm-dependent, and consequently capture the role of initialization on the sample complexity of gradient descent for deep nets. Specialized to noiseless data separable with margin $\\gamma$ by neural tangent kernel (NTK) features of a network of width $\\Omega(poly(\\log(n)))$, we show the test-error rate $e^{O(L)}/{\\gamma^2 n}$, where $n$ is the training set size and $L$ denotes the number of hidden layers. This results in an improvement in the test loss bound compared to previous works while maintaining the poly-logarithmic width conditions. We further investigate excess risk bounds for deep nets trained with noisy data, establishing that under a polynomial condition on the network width, gradient descent can achieve the optimal excess risk. Finally, we show that a large step-size significantly improves upon the NTK regime's results in classifying the XOR distribution. In particular, we show for a one-hidden layer neural network of constant width $m$ with quadratic activation and standard Gaussian initialization that SGD with linear sample complexity and with a large step-size $\\eta=m$ reaches the perfect test accuracy after only $\\lceil\\log(d)\\rceil$ iterations, where $d$ is the data dimension.", "keywords": [ "generalization bounds", "neural networks", "optimization", "feature learning" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 2 ] }, { "paper_id": "h7EwIfjxgn", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Feiyang Ye", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Masashi Sugiyama", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Director" }, { "name": "Xuehao Wang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yueming Lyu", "gender": "Male", "institution": "Agency for Science, Technology and Research (A*STAR)", "country": "SG", "position": "Researcher" }, { "name": "Ivor W Tsang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" } ], "n_formula": 182, "n_formula_1": 41, "n_ref_uni": 52, "n_ref": 114, "n_ref_all": 161, "n_fig": 6, "n_tab": 6, "L_tab": 1076, "n_element_tab": 91, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 43, "n_element_tab_1": 2, "formula_len_all": 16387, "formula_len_all_1": 2140, "len_all": 204822, "len_all_1": 67663, "len_abs": 1188, "len_title": 86, "len_sents": 53518, "len_sents_1": 26191, "n_sents": 558, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 38, "L_abs": 1195, "title": "Sharpness-Aware Black-Box Optimization", "abs": "Black-box optimization algorithms have been widely used in various machine learning problems, including reinforcement learning and prompt fine-tuning. However, directly optimizing the training loss value, as commonly done in existing black-box optimization methods, could lead to suboptimal model quality and generalization performance.\nTo address those problems in black-box optimization, we propose a novel Sharpness-Aware Black-box Optimization (SABO) algorithm, which applies a sharpness-aware minimization strategy to improve the model generalization. Specifically, the proposed SABO method first reparameterizes the objective function by its expectation over a Gaussian distribution. \nThen it iteratively updates the parameterized distribution by approximated stochastic gradients of the maximum objective value within a small neighborhood around the current solution in the Gaussian distribution space. \nTheoretically, we prove the convergence rate and generalization bound of the proposed SABO algorithm. \nEmpirically, extensive experiments on the black-box prompt fine-tuning tasks demonstrate the effectiveness of the proposed SABO method in improving model generalization performance.", "keywords": [ "Black-box Optimization", "Sharpness-Aware Minimization" ], "rating_list": [ 5, 6, 10, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "h71cSd2loX", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Bill Byrne", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Guangyu Yang", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "MS student" }, { "name": "Jingbiao Mei", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Jinghong Chen", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Weizhe Lin", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 21, "n_ref_uni": 16, "n_ref": 75, "n_ref_all": 99, "n_fig": 11, "n_tab": 7, "L_tab": 2062, "n_element_tab": 376, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 272, "n_element_tab_1": 32, "formula_len_all": 2472, "formula_len_all_1": 2491, "len_all": 170521, "len_all_1": 60399, "len_abs": 895, "len_title": 63, "len_sents": 41867, "len_sents_1": 25765, "n_sents": 334, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 899, "title": "On Extending Direct Preference Optimization to Accommodate Ties", "abs": "We derive and investigate two DPO variants that explicitly model the possibility of declaring a tie in pair-wise comparisons. We replace the Bradley-Terry model in DPO with two well-known modeling extensions, by Rao and Kupper and by Davidson, that assign probability to ties as alternatives to clear preferences. Our experiments in neural machine translation and summarization show that explicitly labeled ties can be added to the datasets for these DPO variants without the degradation in task performance that is observed when the same tied pairs are presented to DPO. We find empirically that the inclusion of ties leads to stronger regularization with respect to the reference policy as measured by KL divergence, and we see this even for DPO in its original form. These findings motivate and enable the inclusion of tied pairs in preference optimization as opposed to simply discarding them.", "keywords": [ "Preference Optimization", "Ties", "Direct Preference Optimization", "Langauge Model", "Machine Translation", "Summarisation." ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "h6ktwCPYxE", "primary_area": "learning theory", "n_author": 1, "author_info_list": [ { "name": "Aldo Pacchiano", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 36, "n_formula_1": 35, "n_ref_uni": 26, "n_ref": 41, "n_ref_all": 42, "n_fig": 0, "n_tab": 4, "L_tab": 1744, "n_element_tab": 23, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1839, "n_element_tab_1": 30, "formula_len_all": 3286, "formula_len_all_1": 3132, "len_all": 128932, "len_all_1": 87197, "len_abs": 1182, "len_title": 118, "len_sents": 28974, "len_sents_1": 25855, "n_sents": 251, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1585, "title": "Second Order Bounds for Contextual Bandits with Function Approximation", "abs": "Many works have developed no-regret algorithms for contextual bandits with function approximation, where the mean rewards over context-action pairs belong to a function class $\\mathcal{F}$. Although there are many approaches to this problem, algorithms based on the principle of optimism, such as optimistic least squares have gained in importance. It can be shown the regret of this algorithm scales as $\\widetilde{\\mathcal{O}}\\left(\\sqrt{d_{\\mathrm{eluder}}(\\mathcal{F}) \\log(\\mathcal{F}) T }\\right)$ where $d_{\\mathrm{eluder}}(\\mathcal{F})$ is a statistical measure of the complexity of the function class $\\mathcal{F}$ known as eluder dimension. Unfortunately, even if the variance of the measurement noise of the rewards at time $t$ equals $\\sigma_t^2$ and these are close to zero, the optimistic least squares algorithm\u2019s regret scales with $\\sqrt{T}$. In this work we are the first to develop algorithms that satisfy regret bounds for contextual bandits with function approximation of the form $\\widetilde{\\mathcal{O}}\\left( \\sigma \\sqrt{\\log(\\mathcal{F})d_{\\mathrm{eluder}}(\\mathcal{F}) T } + d_{\\mathrm{eluder}}(\\mathcal{F}) \\cdot \\log(|\\mathcal{F}|)\\right) $ when the variances are unknown and satisfy $\\sigma_t^2 = \\sigma$ for all $t$ and $\\widetilde{\\mathcal{O}}\\left( d_{\\mathrm{eluder}}(\\mathcal{F})\\sqrt{\\log(\\mathcal{F})\\sum_{t=1}^T \\sigma_t^2 } + d_{\\mathrm{eluder}}(\\mathcal{F}) \\cdot \\log(|\\mathcal{F}|)\\right) $ when the variances change every time-step. These bounds generalize existing techniques for deriving second order bounds in contextual linear problems.", "keywords": [ "theory", "contextual bandits", "variance aware", "second order", "bandits" ], "rating_list": [ 5, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "h6k4809xVV", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Gwangpyo Yoo", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "PhD student" }, { "name": "Honguk Woo", "gender": "Male", "institution": "Sungkyunkwan University", "country": "KR", "position": "Associate Professor" } ], "n_formula": 38, "n_formula_1": 39, "n_ref_uni": 44, "n_ref": 81, "n_ref_all": 108, "n_fig": 5, "n_tab": 4, "L_tab": 2247, "n_element_tab": 50, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 849, "n_element_tab_1": 90, "formula_len_all": 3190, "formula_len_all_1": 3194, "len_all": 115365, "len_all_1": 69598, "len_abs": 1323, "len_title": 99, "len_sents": 28287, "len_sents_1": 24078, "n_sents": 276, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1334, "title": "Model Risk-sensitive Offline Reinforcement Learning", "abs": "Offline reinforcement learning (RL) is becoming critical in risk-sensitive areas such as finance and autonomous driving, where incorrect decisions can lead to substantial financial loss or compromised safety. However, traditional risk-sensitive offline RL methods often struggle with accurately assessing risk, with minor errors in the estimated return potentially causing significant inaccuracies of risk estimation. These challenges are intensified by distribution shifts inherent in offline RL. To mitigate these issues, we propose a model risk-sensitive offline RL framework designed to minimize the worst-case of risks across a set of plausible alternative scenarios rather than solely focusing on minimizing estimated risk. We present a critic-ensemble criterion method that identifies the plausible alternative scenarios without introducing additional hyperparameters. We also incorporate the learned Fourier feature framework and the IQN framework to address spectral bias in neural networks, which can otherwise lead to severe errors in calculating model risk. Our experiments in finance and self-driving scenarios demonstrate that the proposed framework significantly reduces risk, by $11.2\\%$ to $18.5\\%$, compared to the most outperforming risk-sensitive offline RL baseline, particularly in highly uncertain environments.", "keywords": [ "risk-sensitive offline reinforcement learning", "reinforcement learning", "offline reinforcement learning", "risk", "model risk" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 2, 4 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "h5xc46rWcZ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Hamed Firooz", "gender": "Male", "institution": "LinkedIn", "country": "", "position": "Researcher" }, { "name": "Maziar Sanjabi", "gender": "Male", "institution": "LinkedIn", "country": "US", "position": "Principal Researcher" }, { "name": "Wenlong Jiang", "gender": "unknown", "institution": "University of Pittsburgh", "country": "US", "position": "PhD student" }, { "name": "Xiaoling Zhai", "gender": "Male", "institution": "LinkedIn", "country": "US", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 12, "n_ref_uni": 41, "n_ref": 75, "n_ref_all": 100, "n_fig": 12, "n_tab": 10, "L_tab": 2697, "n_element_tab": 256, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 2618, "n_element_tab_1": 138, "formula_len_all": 784, "formula_len_all_1": 473, "len_all": 129979, "len_all_1": 62417, "len_abs": 3478, "len_title": 131, "len_sents": 34714, "len_sents_1": 27455, "n_sents": 243, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1231, "title": "Lost-in-Distance: Impact of Contextual Proximity on LLM Performance in Graph Tasks", "abs": "Despite significant advancements, Large Language Models (LLMs) exhibit blind spots that impair their ability to retrieve and process relevant contextual data effectively. We demonstrate that LLM performance in graph tasks with complexities beyond the \u201cneedle-in-a-haystack\u201d scenario\u2014where solving the problem requires cross-referencing and reasoning across multiple subproblems jointly\u2014is influenced by the proximity of relevant information within the context, a phenomenon we term \u201clost-in-distance\u201d. We examine two fundamental graph tasks: identifying common connections between two nodes and assessing similarity among three nodes, and show that the model\u2019s performance in these tasks significantly depends on the relative positioning of common edges. We evaluate three publicly available LLMs\u2014Llama-3-8B, Llama-3-70B, and GPT-4\u2014using various graph encoding techniques that represent graph structures for LLM input. We propose a formulation for the lost-in-distance phenomenon and demonstrate that lost-in-distance and lost-in-the middle phenomenas occur independently. Results indicate that model accuracy can decline by up to 6x as the distance between node connections increases, independent of graph encoding and model size.", "keywords": [ "Large Language Models", "Graph Tasks", "Long Context" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "h5UdvNFHee", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Dongdong Chen", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Lu Yuan", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Weijian Xu", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Yunsheng Li", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Ronald Clark", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" }, { "name": "Yuanze Lin", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 63, "n_ref": 150, "n_ref_all": 174, "n_fig": 5, "n_tab": 5, "L_tab": 1575, "n_element_tab": 225, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1809, "n_element_tab_1": 236, "formula_len_all": 544, "formula_len_all_1": 544, "len_all": 144167, "len_all_1": 60236, "len_abs": 1467, "len_title": 143, "len_sents": 32487, "len_sents_1": 26573, "n_sents": 231, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1472, "title": "REVIP: Rethinking Visual Prompting for Multimodal Large Language Models with External Knowledge", "abs": "In recent years, multimodal large language models (MLLMs) have made significant strides by training on vast high-quality image-text datasets, enabling them to generally understand images well. However, the inherent difficulty in explicitly conveying fine-grained or spatially dense information (e.g., object masks) in the text format poses a challenge for MLLMs, limiting their ability to answer questions requiring an understanding of detailed or localized visual elements. Drawing inspiration from the Retrieval-Augmented Generation (RAG) concept, this paper proposes a new visual prompt approach to integrate fine-grained external knowledge, obtained from specialized vision models (e.g., instance segmentation/OCR models), into MLLMs. This is a promising yet underexplored direction for enhancing MLLMs' performance. Our approach diverges from concurrent works, which transform external knowledge into additional text prompts, necessitating the model to indirectly learn the correspondence between visual content and text coordinates. Instead, we propose embedding fine-grained object knowledge directly into a spatial embedding map as a visual prompt. This design can be easily incorporated into various MLLMs, such as LLaVA and Mipha, considerably improving their visual understanding performance. Through rigorous experiments, we demonstrate that our method can enhance MLLM performance across 11 benchmarks, improving their fine-grained context-aware capabilities.", "keywords": [ "Multimodal Large Language Models; Large Language Models" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "h5NqrrSjlP", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Jinghui zhong", "gender": "Male", "institution": "South China University of Technology", "country": "", "position": "Full Professor" }, { "name": "Zhitong Ma", "gender": "unknown", "institution": "South China University of Technology", "country": "", "position": "MS student" } ], "n_formula": 21, "n_formula_1": 19, "n_ref_uni": 31, "n_ref": 39, "n_ref_all": 76, "n_fig": 7, "n_tab": 7, "L_tab": 1340, "n_element_tab": 170, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 293, "n_element_tab_1": 43, "formula_len_all": 2954, "formula_len_all_1": 1368, "len_all": 165986, "len_all_1": 59833, "len_abs": 1511, "len_title": 106, "len_sents": 53169, "len_sents_1": 28473, "n_sents": 347, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1516, "title": "GESR: A Geometric Evolution Model for Symbolic Regression", "abs": "Symbolic regression is a challenging task in machine learning that aims to automatically discover highly interpretable mathematical equations from limited data. Keen efforts have been devoted to addressing this issue, yielding promising results. However, there are still bottlenecks that current methods struggle with, especially when dealing with complex problems containing various noises or with intricate underlying mathematical formulas.\nIn this work, we propose a novel Geometric Evolution Symbolic Regression(GESR) algorithm. Leveraging geometric semantics, the process of symbolic regression in GESR is transformed into an approximation to an unimodal target in n-dimensional topological space. Then, three key modules are proposed to enhance the approximation: (1) a new semantic gradient concept, proposed to assist the exploration, which aims to improve the accuracy of approximation; (2) a new geometric search operator, tailored for approximating the target formula directly in topological space; (3) the Levenberg-Marquardt algorithm with L2 regularization, used for the adjustment of expression structures and the balance of global subtree weights to assist the proposed geometric semantic search operator. With the proposal of these modules, GESR achieves state-of-the-art accuracy performance on multiple authoritative benchmark datasets and demonstrates a certain level of robustness against noise interference. The implementation is available at https://anonymous.4open.science/r/12331211321-014D.", "keywords": [ "symbolic regression; semantic approximation" ], "rating_list": [ 5, 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 1, 1 ], "contribution_list": [ 2, 2, 3, 1, 3 ], "confidence_list": [ 5, 3, 3, 4, 3 ] }, { "paper_id": "h5D0JICV3s", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Alan Bovik", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Full Professor" }, { "name": "Gregoire Phillips", "gender": "Male", "institution": "Telefonaktiebolaget LM Ericsson", "country": "US", "position": "Researcher" }, { "name": "Shashank Gupta", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 40, "n_ref": 74, "n_ref_all": 82, "n_fig": 2, "n_tab": 3, "L_tab": 570, "n_element_tab": 71, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 257, "n_element_tab_1": 12, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 130634, "len_all_1": 64071, "len_abs": 1196, "len_title": 129, "len_sents": 37141, "len_sents_1": 33130, "n_sents": 259, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1205, "title": "PIT-QMM: A Large Multimodal Model for No-Reference Point Cloud Quality Assessment", "abs": "Large Multimodal Models (LMMs) have recently enabled considerable advances in the realm of image and video quality assessment, but this progress has yet to be translated to the domain of 3D assets. We are interested in using these models to conduct No-Reference Point Cloud Quality Assessment (NR-PCQA), where the aim is to automatically evaluate the perceptual quality of a point cloud in absence of a reference. We begin with the observation that different modalities of data -- text descriptions, 2D projections, and 3D point cloud views -- provide uniquely useful insights into point cloud quality. We leverage this to devise a multimodal dataset construction strategy providing a holistic combination of multiple types and levels of information. We then construct PIT-QMM, a novel LMM for NR-PCQA that is capable of consuming text, images and point clouds to predict quality scores. Extensive experimentation shows that our proposed method outperforms the state-of-the-art by significant margins on popular benchmarks with fewer training iterations, and thorough ablations validate our dataset construction strategy. Code and datasets are available at https://anonymous.4open.science/r/pit-qmm-BD1F/.", "keywords": [ "Multimedia quality assessment", "point clouds", "large multimodal models" ], "rating_list": [ 5, 5, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 5, 4, 5, 5, 3 ] }, { "paper_id": "h51mpl8Tyx", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Fangxin Wang", "gender": "Female", "institution": "University of Illinois at Chicago", "country": "US", "position": "PhD student" }, { "name": "Kay Liu", "gender": "Male", "institution": "University of Illinois Chicago", "country": "US", "position": "PhD student" }, { "name": "Philip S. Yu", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "US", "position": "Full Professor" }, { "name": "Sourav Medya", "gender": "Male", "institution": "University of Illinois at Chicago", "country": "US", "position": "Assistant Professor" } ], "n_formula": 71, "n_formula_1": 19, "n_ref_uni": 29, "n_ref": 60, "n_ref_all": 91, "n_fig": 4, "n_tab": 10, "L_tab": 2845, "n_element_tab": 204, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 490, "n_element_tab_1": 26, "formula_len_all": 7202, "formula_len_all_1": 1782, "len_all": 182125, "len_all_1": 73291, "len_abs": 1131, "len_title": 108, "len_sents": 58413, "len_sents_1": 33972, "n_sents": 567, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1139, "title": "BANGS: Game-theoretic Node Selection for Graph Self-Training", "abs": "Graph self-training is a semi-supervised learning method that iteratively selects a set of unlabeled data to retrain the underlying graph neural network (GNN) model and improve its prediction performance. While selecting highly confident nodes has proven effective for self-training, this pseudo-labeling strategy ignores the combinatorial dependencies between nodes and suffers from a local view of the distribution.\nTo overcome these issues, we propose BANGS, a novel framework that unifies the labeling strategy with conditional mutual information as the objective of node selection. Our approach---grounded in game theory---selects nodes in a combinatorial fashion and provides theoretical guarantees for robustness under noisy objective. More specifically, unlike traditional methods that rank and select nodes independently, BANGS considers nodes as a collective set in the self-training process. Our method demonstrates superior performance and robustness across various datasets, base models, and hyperparameter settings, outperforming existing techniques. The codebase is available on https://anonymous.4open.science/r/BANGS-3EA4.", "keywords": [ "Graph Semi-supervised Learning", "Graph Self-training", "Game Theory Application" ], "rating_list": [ 6, 6, 8, 5, 6 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 4, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 2, 4, 3, 4 ] }, { "paper_id": "h4L5eUvXmP", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Xavier Sumba", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Yingzhen Li", "gender": "Female", "institution": "Imperial College London", "country": "", "position": "Associate Professor" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 59, "n_ref": 112, "n_ref_all": 133, "n_fig": 9, "n_tab": 7, "L_tab": 1594, "n_element_tab": 225, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 2405, "n_element_tab_1": 151, "formula_len_all": 1071, "formula_len_all_1": 843, "len_all": 142342, "len_all_1": 67411, "len_abs": 993, "len_title": 112, "len_sents": 33451, "len_sents_1": 29747, "n_sents": 223, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 998, "title": "From Patches to Graphs: Towards Image Diffusion Models with GNNs", "abs": "Diffusion models have achieved remarkable success in high-quality image generation, typically using convolutional neural networks (CNNs) or Vision Transformers (ViTs) as backbone architectures. However, CNNs may struggle with capturing long-range dependencies, while ViTs can be computationally intensive due to their attention mechanisms. We propose the Diffusion Image GNN (DiG), a novel architecture that leverages graph-based modeling within diffusion models. By representing image patches as nodes in a graph and connecting them based on spatial relationships, DiG efficiently captures both local and global dependencies and naturally handles multi-scale features. \nEmpirical results demonstrate that DiG achieves competitive Frechet Inception Distance (FID) scores compared to state-of-the-art methods. To our knowledge, this is the first application of graph neural networks as a backbone within diffusion models for image generation, opening new avenues for research in generative modeling.", "keywords": [ "Generative Models", "Graph Neural Networks", "Image Diffusion Models" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "h3wbI8Uk1Z", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Kaifeng Lyu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Kaiyue Wen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Xingyu Dang", "gender": "Male", "institution": "Institute for Interdisciplinary Information Sciences, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 100, "n_formula_1": 1, "n_ref_uni": 60, "n_ref": 118, "n_ref_all": 128, "n_fig": 5, "n_tab": 8, "L_tab": 1479, "n_element_tab": 8, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10588, "formula_len_all_1": 65, "len_all": 325345, "len_all_1": 70588, "len_abs": 2575, "len_title": 123, "len_sents": 106374, "len_sents_1": 36584, "n_sents": 968, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1182, "title": "RNNs are not Transformers (Yet): The Key Bottleneck on In-Context Retrieval", "abs": "This paper investigates the gap in representation powers of Transformers and Recurrent Neural Networks (RNNs), which are more memory efficient than Transformers. We aim to understand whether RNNs can match the performance of Transformers, particularly when enhanced with Chain-of-Thought (CoT) prompting. Our theoretical analysis reveals that CoT improves RNNs but is insufficient to close the gap with Transformers. A key bottleneck lies in the inability of RNNs to perfectly retrieve information from the context, even with CoT: \nfor several tasks that explicitly or implicitly require this capability, such as associative recall and determining if a graph is a tree, we prove that RNNs are not expressive enough to solve the tasks while Transformers can solve them with ease.\nConversely, we prove that adopting techniques to enhance the in-context retrieval capability of RNNs, including Retrieval-Augmented Generation (RAG) and adding a single Transformer layer, can elevate RNNs to be capable of solving all polynomial-time solvable problems with CoT, hence closing the representation gap with Transformers. We validate our theory on synthetic and natural language experiments.", "keywords": [ "rnn", "cot", "representation theory" ], "rating_list": [ 5, 8, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 1 ], "confidence_list": [ 3, 2, 3, 4 ] }, { "paper_id": "h3unlS2VWz", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Ali Garjani", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Intern" }, { "name": "Amir Zamir", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Assistant Professor" }, { "name": "Andrei Atanov", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Rahul Ramachandran", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Intern" }, { "name": "Oguzhan Fatih Kar", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 50, "n_ref": 101, "n_ref_all": 126, "n_fig": 15, "n_tab": 13, "L_tab": 3154, "n_element_tab": 338, "n_fig_1": 8, "n_tab_1": 7, "L_tab_1": 1564, "n_element_tab_1": 166, "formula_len_all": 190, "formula_len_all_1": 0, "len_all": 175735, "len_all_1": 58609, "len_abs": 295, "len_title": 210, "len_sents": 43561, "len_sents_1": 25119, "n_sents": 393, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 112, "L_abs": 1547, "title": "How well does GPT-4o understand vision? Solving standard computer vision tasks with multimodal foundation models", "abs": "Multimodal foundation models, such as GPT-4o, have made remarkable progress recently. However, it is not clear exactly where these models stand in terms of understanding vision. {In this paper, we \\textbf{quantify the performance of popular multimodal foundation models} (GPT-4o, Gemini Pro, Claude 3.5 Sonnet, Qwen2-VL) \\textbf{at standard computer vision tasks} (semantic segmentation, object detection, image classification, depth and surface normal prediction) and \\textbf{using established datasets} (e.g., COCO, ImageNet and its variants, etc).} \n\nThe main challenges to performing this are: \\textbf{1)} the models are trained to output text and cannot natively express versatile domains, such as segments or 3D geometry, and \\textbf{2)} many of the leading models are proprietary and accessible only at an API level, i.e., there is no weight access to adapt them. We address these challenges by translating standard vision tasks into equivalent text-promptable and API-compatible tasks via {prompt chaining}. \n\nWe observe that \\textbf{1)} the models are not close to the state-of-the-art at any tasks, and \\textbf{2)} they perform semantic tasks notably better than geometric ones. However, \\textbf{3)} they are respectable generalists; this is remarkable as they are presumably trained on only image-text-based tasks primarily. \\textbf{4)} While the prompting techniques affect the performance, better models exhibit less sensitivity to prompt variations. \\textbf{5)} GPT-4o performs the best, getting the top position in 5 out of 6 tasks.", "keywords": [ "multimodal foundation models", "computer vision" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 2 ] }, { "paper_id": "h3n7al4mhk", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Charles Patrick Martin", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "Senior Lecturer" }, { "name": "Jing Zhang", "gender": "Female", "institution": "Australian National University", "country": "AU", "position": "Lecturer" }, { "name": "Xinlei Niu", "gender": "Female", "institution": "Australian National University", "country": "AU", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 17, "n_ref_uni": 65, "n_ref": 118, "n_ref_all": 147, "n_fig": 7, "n_tab": 9, "L_tab": 2968, "n_element_tab": 236, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1821, "n_element_tab_1": 112, "formula_len_all": 3843, "formula_len_all_1": 1245, "len_all": 203412, "len_all_1": 73591, "len_abs": 1224, "len_title": 118, "len_sents": 59845, "len_sents_1": 31808, "n_sents": 447, "n_sents_1": 235, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1224, "title": "SoundMorpher: Perceptually-Uniform Sound Morphing with Diffusion Model", "abs": "We present SoundMorpher, a sound morphing method that generates perceptually uniform morphing trajectories using a diffusion model. Traditional sound morphing methods models the intractable relationship between morph factor and perception of the stimuli for resulting sounds under a linear assumption, which oversimplifies the complex nature of sound perception and limits their morph quality. In contrast, SoundMorpher explores an explicit proportional mapping between the morph factor and the perceptual stimuli of morphed sounds based on Mel-spectrogram. This approach enables smoother transitions between intermediate sounds and ensures perceptually consistent transformations, which can be easily extended to diverse sound morphing tasks. Furthermore, we present a set of quantitative metrics to comprehensively assess sound morphing systems based on three objective criteria, namely, correspondence, perceptual intermediateness, and smoothness. We provide extensive experiments to demonstrate the effectiveness and versatility of SoundMorpher in real-world scenarios, highlighting its potential impact on various applications such as creative music composition, film post-production and interactive audio technologies.", "keywords": [ "Sound Morphing", "Perceptual Consistency", "Diffusion Model" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 1, 2 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "h3Buc7hXSR", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Hee-Jun Jung", "gender": "Male", "institution": "Gwangju Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Hoyong Kim", "gender": "Male", "institution": "Gwangju Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Ilmin Kang", "gender": "Male", "institution": "Gwangju Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Kangil Kim", "gender": "Male", "institution": "Gwangju Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Kangil Kim", "gender": "unknown", "institution": "Gwangju Institute of Science and Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 42, "n_formula_1": 20, "n_ref_uni": 34, "n_ref": 96, "n_ref_all": 130, "n_fig": 19, "n_tab": 7, "L_tab": 2765, "n_element_tab": 145, "n_fig_1": 14, "n_tab_1": 3, "L_tab_1": 2372, "n_element_tab_1": 95, "formula_len_all": 4459, "formula_len_all_1": 1479, "len_all": 185667, "len_all_1": 68643, "len_abs": 1574, "len_title": 115, "len_sents": 48409, "len_sents_1": 26896, "n_sents": 486, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1599, "title": "Consistent Symmetry Representation over Latent Factors of Variation", "abs": "Recent symmetry-based methods on variational autoencoders have advanced disentanglement learning and combinatorial generalization, yet the appropriate symmetry representation for both tasks is under-clarified. We identify that existing methods struggle with maintaining the $\\textit{consistent symmetries}$ when representing identical changes of latent factors of variation, and they cause issues in achieving equivari-\nance. We theoretically prove the limitations of three frequently used group settings: matrix multiplication with General Lie Groups, defining group action with set of vectors and vector addition, and cyclic groups modeled through surjective functions. To overcome these issues, we introduce a novel method of $\\textit{conformal mapping}$ of latent vectors into a complex number space, ensuring consistent symmetries\nand cyclic semantics. Through empirical validation with ground truth of factors variation for transparent analysis, this study fills two significant gaps in the literature: 1) the inductive bias to enhance disentanglement learning and combinatorial generalization simultaneously, and 2) well-represented symmetries ensure significantly high disentanglement performance without a trade-off in reconstruction error, compared to current unsupervised methods. Additionally, we introduce less guidance-dependent validation results, extending our findings to more practical use. Our research highlights the significant impact of verifying consistent symmetry and suggests required future research for advancing combinatorial generalization and disentanglement learning.", "keywords": [ "Combinatorial Generalizaton", "Disentanglement Learning", "Variational Auto-Encoder", "Symmetry" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "h2Q3gOIz8q", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Jingxuan He", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Mark Vero", "gender": "Male", "institution": "ETHZ-ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Martin Vechev", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Niels M\u00fcndler", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Slobodan Jenko", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 46, "n_ref": 102, "n_ref_all": 127, "n_fig": 17, "n_tab": 3, "L_tab": 854, "n_element_tab": 184, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 323, "formula_len_all_1": 129, "len_all": 186006, "len_all_1": 73639, "len_abs": 1278, "len_title": 106, "len_sents": 63263, "len_sents_1": 35198, "n_sents": 443, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1283, "title": "Black-Box Adversarial Attacks on LLM-Based Code Completion", "abs": "Modern code completion engines, powered by large language models (LLMs), assist millions of developers with their impressive capabilities to generate functionally correct code. As such it is crucial to investigate their security implications. In this work, we present INSEC, the first black-box adversarial attack designed to manipulate modern LLM-based code completion engines into generating vulnerable code. INSEC works by injecting an attack string as a short comment in the completion input. The attack string is crafted through a query-based optimization procedure starting from a set of initialization schemes. We demonstrate INSEC's broad applicability and effectiveness by evaluating it on various state-of-the-art open-source models and black-box commercial services (e.g., OpenAI API and GitHub Copilot). We show that on a diverse set of security-critical test cases covering 16 CWEs across 5 programming languages, INSEC significantly increases the rate of generated insecure code by ~50%, while upholding the engines' capabilities of producing functionally correct code. Moreover, due to its black-box nature, developing INSEC does not require expensive local compute and costs less than 10 USD by querying remote APIs, thereby enabling the threat of widespread attacks.", "keywords": [ "code completion", "security", "code security", "adversarial attacks", "black-box", "large language models" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "h24XT5DOb2", "primary_area": "applications to neuroscience & cognitive science", "n_author": 6, "author_info_list": [ { "name": "Alejandro Lancho", "gender": "unknown", "institution": "Universidad Carlos III, Universidad Carlos III de Madrid", "country": "ES", "position": "Assistant Professor" }, { "name": "Antonio Art\u00e9s", "gender": "Male", "institution": "Universidad Carlos III de Madrid", "country": "ES", "position": "Full Professor" }, { "name": "Diego Herrero-Quevedo", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Josu\u00e9 P\u00e9rez Sabater", "gender": "unknown", "institution": "Universidad Carlos III, Universidad Carlos III de Madrid", "country": "ES", "position": "Researcher" }, { "name": "Pablo M. Olmos", "gender": "unknown", "institution": "Universidad Carlos III de Madrid", "country": "ES", "position": "Associate Professor" }, { "name": "Rodrigo Oliver", "gender": "Not Specified", "institution": "Universidad Carlos III de Madrid", "country": "ES", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 28, "n_ref": 38, "n_ref_all": 75, "n_fig": 21, "n_tab": 8, "L_tab": 3722, "n_element_tab": 437, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 344, "n_element_tab_1": 45, "formula_len_all": 663, "formula_len_all_1": 497, "len_all": 197101, "len_all_1": 59255, "len_abs": 1566, "len_title": 72, "len_sents": 67357, "len_sents_1": 29547, "n_sents": 466, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1573, "title": "A Foundation Model for Patient Behavior Monitoring and Suicide Detection", "abs": "Foundation models have achieved remarkable success across various domains, yet their adoption in healthcare remains limited, particularly in areas requiring the analysis of smaller and more complex datasets. While foundation models have made significant advances in medical imaging, genetic biomarkers, and time series from electronic health records, the potential for patient behavior monitoring through wearable devices remains underexplored. Wearable device datasets are inherently heterogeneous and multisource and often exhibit high rates of missing data, presenting unique challenges. Notably, missing patterns in these datasets are frequently not-at-random, and when adequately modeled, these patterns can reveal crucial insights into patient behavior. \nThis paper introduces a novel foundation model based on a modified vector quantized variational autoencoder (VQ-VAE), specifically designed to process real-world data from wearable devices. Our model excels at reconstructing heterogeneous multisource time-series data and effectively models missing data patterns. We demonstrate that our pretrained model, trained on a broad cohort of psychiatric patients with diverse mental health issues, can perform downstream tasks without fine-tuning on a held-out cohort of suicidal patients. This is illustrated through the use of a change-point detection algorithm that identifies suicide attempts with high accuracy, matching or surpassing patient-specific methods, thereby highlighting the potential of VQ-VAE as a versatile tool for behavioral analysis in healthcare.", "keywords": [ "Foundation models for patient monitoring", "Suicide detection via deep learning" ], "rating_list": [ 6, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "h1XoHOd19I", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Jinhao Jiang", "gender": "unknown", "institution": "Renmin University of China", "country": "", "position": "PhD student" }, { "name": "Junyi Li", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Tao Zhang", "gender": "Male", "institution": "BOSS Zhipin Career Science Lab", "country": "CN", "position": "Researcher" }, { "name": "Xin Zhao", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "SONG Yang", "gender": "Male", "institution": "BOSS Zhipin", "country": "CN", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 14, "n_ref_uni": 47, "n_ref": 94, "n_ref_all": 111, "n_fig": 8, "n_tab": 2, "L_tab": 2019, "n_element_tab": 165, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1243, "n_element_tab_1": 125, "formula_len_all": 576, "formula_len_all_1": 557, "len_all": 203124, "len_all_1": 60767, "len_abs": 1378, "len_title": 141, "len_sents": 40719, "len_sents_1": 29407, "n_sents": 259, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1350, "title": "Mix-CPT: A Domain Adaptation Framework via Decoupling Knowledge Learning and Format Alignment", "abs": "Adapting large language models (LLMs) to specialized domains typically requires domain-specific corpora for continual pre-training to facilitate knowledge memorization and related instructions for fine-tuning to apply this knowledge.\nHowever, this method may lead to inefficient knowledge memorization due to a lack of awareness of knowledge utilization during the continual pre-training and demands LLMs to simultaneously learn knowledge utilization and format alignment with divergent training objectives during the fine-tuning.\nTo enhance the domain adaptation of LLMs, we revise this process and propose a new domain adaptation framework including domain knowledge learning and general format alignment, called \\emph{Mix-CPT}. Specifically, we first conduct a knowledge mixture continual pre-training that concurrently focuses on knowledge memorization and utilization. To avoid catastrophic forgetting, we further propose a logit swap self-distillation constraint. By leveraging the knowledge and capabilities acquired during continual pre-training, we then efficiently perform instruction tuning and alignment with a few general training samples to achieve format alignment.\nExtensive experiments show that our proposed \\emph{Mix-CPT} framework can simultaneously improve the task-solving capabilities of LLMs on the target and general domains.", "keywords": [ "LLM", "Continual Pre-training", "knowledge distillation" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "h0vC0fm1q7", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Arhaan Ahmad", "gender": "Male", "institution": "Indian Institute of Technology, Bombay", "country": "IN", "position": "Undergrad student" }, { "name": "Ashutosh Gupta", "gender": "Male", "institution": "IIT Bombay", "country": "IN", "position": "Associate Professor" }, { "name": "S. Akshay", "gender": "Male", "institution": "Indian Institute of Technology Bombay, Indian Institute of Technology, Bombay", "country": "IN", "position": "Full Professor" }, { "name": "Tanay Vineet Tayal", "gender": "Male", "institution": "Indian Institute of Technology, Bombay, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "Undergrad student" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 23, "n_ref": 54, "n_ref_all": 65, "n_fig": 2, "n_tab": 1, "L_tab": 1078, "n_element_tab": 174, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 986, "n_element_tab_1": 173, "formula_len_all": 1411, "formula_len_all_1": 1293, "len_all": 127800, "len_all_1": 75082, "len_abs": 1214, "len_title": 109, "len_sents": 41435, "len_sents_1": 34018, "n_sents": 354, "n_sents_1": 280, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1221, "title": "Sensitivity Verification for Decision Tree Ensembles", "abs": "Tree ensemble models, such as Gradient Boosted Decision Trees (GBDTs) and\nrandom forests, are widely popular models for a variety of machine learning tasks.\nThe power of these models comes from the ensemble of decision trees, which\nmakes analysis of such models significantly harder than for single trees. As a\nresult, recent work has focused on developing exact and approximate techniques\nfor questions such as robustness verification, fairness and explainability, for such\nmodels of tree ensembles.\n\nIn this paper, we focus on a specific problem of feature sensitivity for additive\ndecision tree ensembles and build a formal verification framework for it. We start\nby showing theoretical (NP-)hardness of the problem and explain how it relates\nto other verification problems. Next, we provide a novel encoding of the problem\nusing pseudo-Boolean constraints. Based on this encoding, we develop a tunable\nalgorithm to perform sensitivity analysis, which can trade off precision for running\ntime. We implement our algorithm and study its performance on a suite of GBDT\nbenchmarks from the literature. Our experiments show the practical utility of our\napproach and its improved performance compared to existing approaches.", "keywords": [ "Robustness verification", "Sensitivity analysis", "SAT solvers", "efficient encodings", "NP-hardness", "fairness" ], "rating_list": [ 6, 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3, 2 ], "presentation_list": [ 2, 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3, 3 ] }, { "paper_id": "h0pACOIFxC", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Aryan Mokhtari", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Jacob L. Block", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Liam Collins", "gender": "unknown", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Sanjay Shakkottai", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Full Professor" }, { "name": "Sundararajan Srinivasan", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" } ], "n_formula": 73, "n_formula_1": 43, "n_ref_uni": 26, "n_ref": 51, "n_ref_all": 70, "n_fig": 5, "n_tab": 4, "L_tab": 1280, "n_element_tab": 115, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9244, "formula_len_all_1": 3981, "len_all": 166460, "len_all_1": 72567, "len_abs": 1495, "len_title": 89, "len_sents": 43430, "len_sents_1": 31165, "n_sents": 406, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1502, "title": "Meta-Learning Adaptable Foundation Models", "abs": "The power of foundation models (FMs) lies in their capacity to learn highly expressive representations that can be adapted to a broad spectrum of tasks. However, these pretrained models require multiple stages of fine-tuning to become effective for downstream applications. Conventionally, the model is first retrained on the aggregate of a diverse set of tasks of interest and then adapted to specific low-resource downstream tasks by utilizing a parameter-efficient fine-tuning (PEFT) scheme. While this procedure seems reasonable, the independence of the retraining and fine-tuning stages causes a major issue, as there is no guarantee the retrained model will achieve good performance post-fine-tuning. To explicitly address this issue, we introduce a meta-learning framework infused with PEFT in this intermediate retraining stage to learn a model that can be easily adapted to unseen tasks. For our theoretical results, we focus on linear models using low-rank adaptations. In this setting, we demonstrate the suboptimality of standard retraining for finding an adaptable set of parameters. Further, we prove that our method recovers the optimally adaptable parameters. We then apply these theoretical insights to retraining the RoBERTa model to predict the continuation of conversations between different personas within the ConvAI2 dataset. Empirically, we observe significant performance benefits using our proposed meta-learning scheme during retraining relative to the conventional approach.", "keywords": [ "meta-learning theory", "parameter efficient fine-tuning", "low-rank adaptation", "foundation models", "non-convex optimization" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "h0jdAboh0o", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Leo Richter", "gender": "Non-Binary", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Matt J Kusner", "gender": "Male", "institution": "University College London", "country": "", "position": "Associate Professor" }, { "name": "Xuanli He", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Postdoc" }, { "name": "Pasquale Minervini", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 58, "n_formula_1": 22, "n_ref_uni": 80, "n_ref": 141, "n_ref_all": 179, "n_fig": 11, "n_tab": 8, "L_tab": 1491, "n_element_tab": 111, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1258, "n_element_tab_1": 57, "formula_len_all": 8604, "formula_len_all_1": 10208, "len_all": 263051, "len_all_1": 81785, "len_abs": 1196, "len_title": 110, "len_sents": 65236, "len_sents_1": 32317, "n_sents": 489, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1203, "title": "An Auditing Test to Detect Behavioral Shift in Language Models", "abs": "As language models (LMs) approach human-level performance, a comprehensive understanding of their behavior becomes crucial. \nThis includes evaluating capabilities, biases, task performance, and alignment with societal values. Extensive initial evaluations, including red teaming and diverse benchmarking, can establish a model\u2019s behavioral profile. However, subsequent fine-tuning or deployment modifications may alter these behaviors in unintended ways. We present an efficient statistical test to tackle Behavioral Shift Auditing (BSA) in LMs, which we define as detecting distribution shifts in qualitative properties of the output distributions of LMs. Our test compares model generations from a baseline model to those of the model under scrutiny and provides theoretical guarantees for change detection while controlling false positives. The test features a configurable tolerance parameter that adjusts sensitivity to behavioral changes for different use cases. We evaluate our approach using two case studies: monitoring changes in (a) toxicity and (b) translation performance. We find that the test is able to detect meaningful changes in behavior distributions using just hundreds of examples.", "keywords": [ "AI alignment", "model auditing", "model evaluations", "red teaming", "sequential hypothesis testing" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 2, 2, 3 ] }, { "paper_id": "h0ZfDIrj7T", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Ben Athiwaratkun", "gender": "Male", "institution": "Together AI", "country": "US", "position": "Staff AI Scientist" }, { "name": "Ce Zhang", "gender": "unknown", "institution": "University of Chicago", "country": "", "position": "Associate Professor" }, { "name": "James Zou", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Jue WANG", "gender": "Male", "institution": "Together AI", "country": "US", "position": "Researcher" }, { "name": "Junlin Wang", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 31, "n_ref": 60, "n_ref_all": 98, "n_fig": 8, "n_tab": 9, "L_tab": 2727, "n_element_tab": 320, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 902, "n_element_tab_1": 57, "formula_len_all": 248, "formula_len_all_1": 248, "len_all": 160223, "len_all_1": 58721, "len_abs": 880, "len_title": 108, "len_sents": 43189, "len_sents_1": 27157, "n_sents": 345, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 60, "L_abs": 882, "title": "Mixture-of-Agents Enhances Large Language Model Capabilities", "abs": "Recent advances in large language models (LLMs) demonstrate substantial capabilities in natural language understanding and generation tasks. With the growing number of LLMs, how to harness the collective expertise of multiple LLMs is an exciting open direction. Toward this goal, we propose a new approach that leverages the collective strengths of multiple LLMs through a Mixture-of-Agents (MoA) methodology. In our approach, we construct a layered MoA architecture wherein each layer comprises multiple LLM agents. Each agent takes all the outputs from agents in the previous layer as auxiliary information in generating its response. MoA models achieves state-of-art performance on AlpacaEval 2.0, Arena-Hard, MT-Bench, and FLASK, surpassing GPT-4 Omni. For example, our MoA using only open-source LLMs achieves a score of 65.1% on AlpacaEval 2.0 compared to 57.5% by GPT-4 Omni.", "keywords": [ "Multi-Agent Inference", "Large Language Model" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "h0Ak8A5yqw", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Fei Huang", "gender": "Male", "institution": "Alibaba Group US", "country": "US", "position": "Senior Research Director" }, { "name": "Haiyang Yu", "gender": "unknown", "institution": "Xiamen University", "country": "", "position": "MS student" }, { "name": "Junfeng Fang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Kun Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Rongwu Xu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xinghua Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yang Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Yongbin Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zhenhong Zhou", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" } ], "n_formula": 28, "n_formula_1": 17, "n_ref_uni": 53, "n_ref": 126, "n_ref_all": 167, "n_fig": 22, "n_tab": 4, "L_tab": 943, "n_element_tab": 91, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 465, "n_element_tab_1": 46, "formula_len_all": 1582, "formula_len_all_1": 1086, "len_all": 205117, "len_all_1": 67479, "len_abs": 1846, "len_title": 109, "len_sents": 56144, "len_sents_1": 30622, "n_sents": 446, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 61, "L_abs": 1799, "title": "On the Role of Attention Heads in Large Language Model Safety", "abs": "Large language models (LLMs) achieve state-of-the-art performance on multiple language tasks, yet their safety guardrails can be circumvented, leading to harmful generations. In light of this, recent research on safety mechanisms has emerged, revealing that when safety representations or component are suppressed, the safety capability of LLMs are compromised. However, existing research tends to overlook the safety impact of multi-head attention mechanisms, despite their crucial role in various model functionalities. Hence, in this paper, we aim to explore the connection between standard attention mechanisms and safety capability to fill this gap in the safety-related mechanistic interpretability. We propose an novel metric which tailored for multi-head attention, the Safety Head ImPortant Score (Ships), to assess the individual heads' contributions to model safety. Base on this, we generalize Ships to the dataset level and further introduce the Safety Attention Head AttRibution Algorithm (Sahara) to attribute the critical safety attention heads inside the model. Our findings show that special attention head has a significant impact on safety. Ablating a single safety head allows aligned model (e.g., Llama-2-7b-chat) to respond to **16$\\times\\uparrow$** more harmful queries, while only modifying **0.006\\%** $\\downarrow$ of the parameters, in contrast to the $\\sim$ **5\\%** modification required in previous studies. More importantly, we demonstrate that attention heads primarily function as feature extractors for safety and models fine-tuned from the same base model exhibit overlapping safety heads through comprehensive experiments. Together, our attribution approach and findings provide a novel perspective for unpacking the black box of safety mechanisms in large models.", "keywords": [ "interpretability", "large language model", "multi-head attention", "safety", "harmful content" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "gzzX4ZeErx", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Chenchen Ye", "gender": "Female", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Mingyu Derek Ma", "gender": "Male", "institution": "Genentech", "country": "US", "position": "Machine Learning Scientist" }, { "name": "Wei Wang", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "Yanqiao Zhu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Yihe Deng", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Zijie Huang", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Ziniu Hu", "gender": "Male", "institution": "xAI", "country": "US", "position": "Member of Technical Staff" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 71, "n_ref": 128, "n_ref_all": 165, "n_fig": 15, "n_tab": 16, "L_tab": 12373, "n_element_tab": 875, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 2089, "n_element_tab_1": 102, "formula_len_all": 135, "formula_len_all_1": 0, "len_all": 501063, "len_all_1": 79285, "len_abs": 2408, "len_title": 113, "len_sents": 152810, "len_sents_1": 36713, "n_sents": 1137, "n_sents_1": 314, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1575, "title": "MIRAI: Evaluating LLM Agents for International Event Forecasting", "abs": "We present MIRAI, a benchmark designed to systematically evaluate LLM agents as temporal forecasters to predict international events. Our benchmark features an agentic environment with tools for accessing an extensive database of historical, structured events and textual news articles. We refine the GDELT event database with careful cleaning and parsing to curate a series of relational prediction tasks with varying forecasting horizons, assessing LLM agents\u2019 abilities from short-term to long-term forecasting. We further implement APIs to enable LLM agents to utilize different tools via a code-based interface. Notably, MIRAI features a dynamic data construction pipeline that supports periodically downloading recent news and events, and automatically generates the most recent test split. This allows us to evaluate any newly released model in a contamination-free manner as we can always construct a test split later than its knowledge cutoff date. MIRAI comprehensively evaluates the agents\u2019 capabilities in three dimensions: 1) autonomously source and integrate critical information from large global databases; 2) write codes with both domain-specific APIs and libraries for tool-use; and 3) jointly reason over historical knowledge from diverse formats and timespan to accurately predict future events. Through comprehensive evaluation, we establish a reliable benchmark for assessing the capabilities of LLM agents in forecasting international events and contribute to the development of more accurate and trustworthy models for international relation analysis.", "keywords": [ "LLM Agents", "Temporal Forecasting", "Tool Use" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "gzxDjnvBDa", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Kanta Ono", "gender": "unknown", "institution": "Osaka University, Tokyo Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Ryo Igarashi", "gender": "Male", "institution": "OMRON SINIC X Corporation", "country": "", "position": "Project Researcher" }, { "name": "Tatsunori Taniai", "gender": "Male", "institution": "OMRON SINIC X", "country": "JP", "position": "Researcher" }, { "name": "Yusei Ito", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "MS student" }, { "name": "Yoshitaka Ushiku", "gender": "Male", "institution": "OMRON SINIC X", "country": "JP", "position": "Vice President for Research" } ], "n_formula": 18, "n_formula_1": 13, "n_ref_uni": 50, "n_ref": 171, "n_ref_all": 210, "n_fig": 6, "n_tab": 11, "L_tab": 4780, "n_element_tab": 503, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1744, "n_element_tab_1": 195, "formula_len_all": 1040, "formula_len_all_1": 826, "len_all": 190989, "len_all_1": 76088, "len_abs": 1449, "len_title": 124, "len_sents": 53910, "len_sents_1": 34829, "n_sents": 412, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1340, "title": "Rethinking the role of frames for SE(3)-invariant crystal structure modeling", "abs": "Crystal structure modeling with graph neural networks is essential for various applications in materials informatics, and capturing SE(3)-invariant geometric features is a fundamental requirement for these networks. A straightforward approach is to model with orientation-standardized structures through structure-aligned coordinate systems, or \u201cframes.\u201d However, unlike molecules, determining frames for crystal structures is challenging due to their infinite and highly symmetric nature. In particular, existing methods rely on a statically fixed frame for each structure, determined solely by its structural information, regardless of the task under consideration. Here, we rethink the role of frames, *questioning whether such simplistic alignment with the structure is sufficient*, and propose the concept of *dynamic frames*. While accommodating the infinite and symmetric nature of crystals, these frames provide each atom with a dynamic view of its local environment, focusing on actively interacting atoms. We demonstrate this concept by utilizing the attention mechanism in a recent transformer-based crystal encoder, resulting in a new architecture called **CrystalFramer**. Extensive experiments show that CrystalFramer outperforms conventional frames and existing crystal encoders in various crystal property prediction tasks.", "keywords": [ "Materials Science", "Invariant Networks", "Transformer", "Physics-Informed ML" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "gzmInLJSoW", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Chiwun Yang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "Yingyu Liang", "gender": "Male", "institution": "University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Zhenmei Shi", "gender": "Male", "institution": "Voyage AI", "country": "US", "position": "Researcher" } ], "n_formula": 178, "n_formula_1": 29, "n_ref_uni": 84, "n_ref": 156, "n_ref_all": 187, "n_fig": 3, "n_tab": 6, "L_tab": 3203, "n_element_tab": 285, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 592, "n_element_tab_1": 31, "formula_len_all": 24525, "formula_len_all_1": 2840, "len_all": 384004, "len_all_1": 73250, "len_abs": 1199, "len_title": 91, "len_sents": 80821, "len_sents_1": 30925, "n_sents": 902, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1205, "title": "Towards Infinite-Long Prefix in Transformer", "abs": "Prompting and context-based fine-tuning methods, which we call Prefix Learning, have been proposed to enhance the performance of language models on various downstream tasks. They are empirically efficient and effective, matching the performance of full parameter fine-tuning, but the theoretical understandings are limited. In this paper, we aim to address this limitation by studying their ability from the perspective of prefix length. \nIn particular, we provide a convergence guarantee for training an ultra-long prefix in a stylized setting using the Neural Tangent Kernel (NTK) framework. Based on this strong theoretical guarantee, we design and implement an algorithm that only needs to introduce and fine-tune a few extra trainable parameters instead of an infinite-long prefix in each layer of a transformer, and can approximate the prefix attention to a guaranteed polynomial-small error.\nPreliminary experimental results on vision, natural language, and math data show that our method achieves superior or competitive performance compared to existing methods like full parameters fine-tuning, P-Tuning V2, and LoRA. This demonstrates our method is promising for parameter-efficient fine-tuning.", "keywords": [ "Large Language Model", "Prefix Learning", "Neural Tangent Kernel" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 2, 5, 2 ] }, { "paper_id": "gz8Rr1iuDK", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "David S. Greenberg", "gender": "Male", "institution": "Helmholtz Centre Hereon", "country": "DE", "position": "Principal Researcher" }, { "name": "Yunfei Huang", "gender": "Male", "institution": "Helmholtz-Zentrum Hereon", "country": "DE", "position": "Postdoc" } ], "n_formula": 42, "n_formula_1": 7, "n_ref_uni": 63, "n_ref": 108, "n_ref_all": 159, "n_fig": 27, "n_tab": 3, "L_tab": 729, "n_element_tab": 75, "n_fig_1": 22, "n_tab_1": 3, "L_tab_1": 119, "n_element_tab_1": 2, "formula_len_all": 5707, "formula_len_all_1": 346, "len_all": 170081, "len_all_1": 60271, "len_abs": 1132, "len_title": 129, "len_sents": 40932, "len_sents_1": 27922, "n_sents": 324, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1135, "title": "Geometric and Physical Constraints Synergistically Improve Neural PDE Integration", "abs": "Neural PDE surrogates can improve on cost-accuracy tradeoffs of classical solvers, but often generalize poorly to new initial conditions, accumulate errors over time. To close the performance gap between training and long-term inference, we constrain neural surrogates with symmetry equivariance and physical conservation laws as hard constraints, using novel input and output layers that support scalar and vector fields on the staggered grids commonly used in computational fluid dynamics. We systematically investigate how these constraints affect accuracy, individually and in combination, on two challenging tasks: shallow water equations with closed boundaries and decaying incompressible turbulence. Compared to a strong baseline, both types of constraints improve performance consistently across autoregressive prediction steps, accuracy measures, and network sizes. Symmetries are more effective but do not make physical constraints redundant. Doubly-constrained surrogates were more accurate for the same network and dataset sizes, and generalized better to initial conditions and durations beyond the range of training data.", "keywords": [ "Geometric deep learning", "physics-constrained neural networks", "PDE integration" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "gyvYKLEm8t", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Feng Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "Associate Professor" }, { "name": "Shaoang Li", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Postdoc" }, { "name": "Shuli Zeng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Sijia Zhang", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Xiangyang Li", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 39, "n_formula_1": 20, "n_ref_uni": 29, "n_ref": 75, "n_ref_all": 98, "n_fig": 3, "n_tab": 22, "L_tab": 5821, "n_element_tab": 704, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1412, "n_element_tab_1": 167, "formula_len_all": 3560, "formula_len_all_1": 1019, "len_all": 175762, "len_all_1": 71666, "len_abs": 1273, "len_title": 128, "len_sents": 59707, "len_sents_1": 33162, "n_sents": 457, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1296, "title": "Learning to Select Nodes in Branch and Bound with Sufficient Tree Representation", "abs": "Branch-and-bound methods are pivotal in solving Mixed Integer Linear Programming (MILP), where the challenge of node selection arises, necessitating the prioritization of different regions of the space for subsequent exploration. While machine learning techniques have been proposed to address this, two crucial problems concerning \\textbf{(P1)} how to sufficiently extract features from the branch-and-bound tree, and \\textbf{(P2)} how to assess the node quality comprehensively based on the features remain open. To tackle these challenges, we propose to tackle the node selection problem employing a novel Tripartite graph representation and Reinforcement learning with a Graph Neural Network model (TRGNN). The tripartite graph is theoretically proved to encompass sufficient information for tree representation in information theory. We learn node selection via reinforcement learning for learning delay rewards and give more comprehensive node metrics. Experiments show that TRGNN significantly improves the efficiency of solving MILPs compared to human-designed and learning-based node selection methods on both synthetic and large-scale real-world MILPs. Moreover, experiments demonstrate that TRGNN well generalizes to MILPs that are significantly larger than those seen during training.", "keywords": [ "branch and bound; mixed integer linear programming" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "gye2U9uNXx", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Arjun Patrawala", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "Undergrad student" }, { "name": "Erik Jones", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Jacob Steinhardt", "gender": "unknown", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 37, "n_ref": 58, "n_ref_all": 81, "n_fig": 6, "n_tab": 10, "L_tab": 5513, "n_element_tab": 661, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 932, "n_element_tab_1": 18, "formula_len_all": 389, "formula_len_all_1": 364, "len_all": 245952, "len_all_1": 70346, "len_abs": 1257, "len_title": 117, "len_sents": 72429, "len_sents_1": 33845, "n_sents": 540, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 68, "L_abs": 1283, "title": "Uncovering Gaps in How Humans and LLMs Interpret Subjective Language", "abs": "Humans often rely on subjective natural language to direct language models (LLMs); for example, users might instruct the LLM to write an *enthusiastic* blogpost, while developers might train models to be *helpful* and *harmless* using LLM-based edits. The LLM's *operational semantics* of such subjective phrases---how it adjusts its behavior when each phrase is included in the prompt---thus dictates how aligned it is with human intent. In this work, we uncover instances of *misalignment* between LLMs' actual operational semantics and what humans expect. Our method, TED (thesaurus error detector), first constructs a thesaurus that captures whether two phrases have similar operational semantics according to the LLM. It then elicits failures by unearthing disagreements between this thesaurus and a reference semantic thesaurus. TED routinely produces surprising instances of misalignment; for example, Mistral 7B Instruct produces more *harassing* outputs when it edits text to be *witty*, and Llama 3 8B Instruct produces *dishonest* articles when instructed to make the articles *enthusiastic*. Our results demonstrate that we can uncover unexpected LLM behavior by characterizing relationships between abstract concepts, rather than supervising individual outputs directly.", "keywords": [ "safety", "alignment", "constitutional ai", "language model failures", "misalignment", "automated evaluation", "automated red-teaming" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "gyTkfVYL45", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Changliang Zhou", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Qingfu Zhang", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Tong Xialiang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Xi Lin", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Zhenkun Wang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 13, "n_ref_uni": 48, "n_ref": 158, "n_ref_all": 207, "n_fig": 4, "n_tab": 21, "L_tab": 11363, "n_element_tab": 1635, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 4542, "n_element_tab_1": 683, "formula_len_all": 1489, "formula_len_all_1": 679, "len_all": 216678, "len_all_1": 69238, "len_abs": 1535, "len_title": 129, "len_sents": 61286, "len_sents_1": 25216, "n_sents": 454, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1546, "title": "ICAM: Rethinking Instance-Conditioned Adaptation in Neural Vehicle Routing Solver", "abs": "The neural combinatorial optimization (NCO) has shown great potential for solving routing problems without requiring expert knowledge. However, existing constructive NCO methods still struggle to directly solve large-scale instances, which significantly limits their application prospects. To address these crucial shortcomings, this work proposes a novel Instance-Conditioned Adaptation Model (ICAM) for better large-scale generalization of neural routing solvers. In particular, we design a simple yet efficient instance-conditioned adaptation function to significantly improve the generalization performance of existing NCO models with a very small time and memory overhead. In addition, with a systematic investigation on the performance of information incorporation between different attention mechanisms, we further propose a powerful yet low-complexity instance-conditioned adaptation module to generate better solutions for instances across different scales. Experimental results show that our proposed method is capable of obtaining promising results with a very fast inference time in solving Traveling Salesman Problems (TSPs), Capacitated Vehicle Routing Problems (CVRPs) and Asymmetric Traveling Salesman Problems (ATSPs). To the best of our knowledge, our model achieves state-of-the-art performance among all RL-based constructive methods for TSPs and ATSPs with up to 1,000 nodes and extends state-of-the-art performance to 5,000 nodes on CVRP instances, and our method also generalizes well to solve cross-distribution instances.", "keywords": [ "Vehicle Routing Problem", "Reinforcement Learning", "Instance-Conditioned Adaptation", "Neural Combinatorial Optimization", "Large-scale Generalization" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "gyHoR6uFhU", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Chau-Wai Wong", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Assistant Professor" }, { "name": "Pingzhi Li", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Rana Shahroz", "gender": "unknown", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Shahriar Nirjon", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Associate Professor" }, { "name": "Sukwon Yun", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Zhenyu Wang", "gender": "unknown", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 9, "n_ref_uni": 37, "n_ref": 54, "n_ref_all": 75, "n_fig": 2, "n_tab": 6, "L_tab": 2413, "n_element_tab": 275, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2550, "n_element_tab_1": 275, "formula_len_all": 1423, "formula_len_all_1": 477, "len_all": 161113, "len_all_1": 68938, "len_abs": 1820, "len_title": 147, "len_sents": 40692, "len_sents_1": 29819, "n_sents": 297, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 99, "L_abs": 1825, "title": "PortLLM: Personalizing Evolving Large Language Models with Training-Free and Portable Model Patches", "abs": "As large language models (LLMs) increasingly shape the AI landscape, fine-tuning pretrained models has become more popular than in the pre-LLM era for achieving optimal performance in domain-specific tasks. However, pretrained LLMs such as ChatGPT are periodically evolved (i.e., model parameters are frequently updated), making it challenging for downstream users with limited resources to keep up with fine-tuning the newest LLMs for their domain application. Even though fine-tuning costs have nowadays been reduced thanks to the innovations of parameter-efficient fine-tuning such as LoRA, not all downstream users have adequate computing for frequent personalization. Moreover, access to fine-tuning datasets, particularly in sensitive domains such as healthcare, could be time-restrictive, making it crucial to retain the knowledge encoded in earlier fine-tuned rounds for future adaptation. In this paper, we present PORTLLM, a training-free framework that (i) creates an initial lightweight model update patch to capture domain-specific knowledge, and (ii) allows a subsequent seamless plugging for the continual personalization of evolved LLM at minimal cost. Our extensive experiments cover seven representative datasets, from easier question-answering tasks {BoolQ, SST2} to harder reasoning tasks {WinoGrande, GSM8K}, and models including {Mistral-7B,Llama2, Llama3.1, and Gemma2}, validating the portability of our designed model patches and showcasing the effectiveness of our proposed framework. For instance, PORTLLM achieves comparable performance to LoRA fine-tuning with reductions of up to 12.2\u00d7 in GPU memory usage. Finally, we provide theoretical justifications to understand the portability of our model update patches, which offers new insights into the theoretical dimension of LLMs\u2019 personalization.", "keywords": [ "Large Language Models", "NLP", "Efficiency", "Fine-tuning", "Efficient Fine-tuning", "Portability" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "gx3LMRB15C", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Arpad Rimmel", "gender": "unknown", "institution": "Laboratoire de recherche en informatique", "country": "", "position": "Assistant Professor" }, { "name": "Bich-Li\u00ean DOAN", "gender": "Female", "institution": "CentraleSupelec", "country": "", "position": "Full Professor" }, { "name": "Fabrice Popineau", "gender": "Male", "institution": "Laboratoire Interdisciplinaire des Sciences du Num\u00e9rique", "country": "", "position": "Researcher" }, { "name": "Hugo Thimonier", "gender": "Male", "institution": "Emobot", "country": "FR", "position": "Researcher" }, { "name": "Jos\u00e9 Lucas De Melo Costa", "gender": "Male", "institution": "Universit\u00e9 Paris-Saclay", "country": "FR", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 6, "n_ref_uni": 47, "n_ref": 117, "n_ref_all": 152, "n_fig": 6, "n_tab": 11, "L_tab": 4424, "n_element_tab": 555, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 558, "n_element_tab_1": 35, "formula_len_all": 1414, "formula_len_all_1": 335, "len_all": 185239, "len_all_1": 63069, "len_abs": 1628, "len_title": 115, "len_sents": 56838, "len_sents_1": 30972, "n_sents": 427, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1636, "title": "T-JEPA: Augmentation-Free Self-Supervised Learning for Tabular Data", "abs": "Self-supervision is often used for pre-training to foster performance on a downstream task by constructing meaningful representations of samples. Self-supervised learning (SSL) generally involves generating different views of the same sample and thus requires data augmentations that are challenging to construct for tabular data. This constitutes one of the main challenges of self-supervision for structured data. In the present work, we propose a novel augmentation-free SSL method for tabular data. Our approach, T-JEPA, relies on a Joint Embedding Predictive Architecture (JEPA) and is akin to mask reconstruction in the latent space. It involves predicting the latent representation of one subset of features from the latent representation of a different subset within the same sample, thereby learning rich representations without augmentations. We use our method as a pre-training technique and train several deep classifiers on the obtained representation. Our experimental results demonstrate a substantial improvement in both classification and regression tasks, outperforming models trained directly on samples in their original data space. Moreover, T-JEPA enables some methods to consistently outperform or match the performance of traditional methods likes Gradient Boosted Decision Trees. To understand why, we extensively characterize the obtained representations and show that T-JEPA effectively identifies relevant features for downstream tasks without access to the labels. Additionally, we introduce regularization tokens, a novel regularization method critical for training of JEPA-based models on structured data.", "keywords": [ "Self-Supervised Learning", "Tabular Data", "Representation Learning" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "gx1wHnf5Vp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Jun Suzuki", "gender": "Male", "institution": "Tohoku University", "country": "JP", "position": "Full Professor" }, { "name": "Kazuki Fujii", "gender": "Not Specified", "institution": "Tokyo Institute of Technology", "country": "JP", "position": "MS student" }, { "name": "Rio Yokota", "gender": "Male", "institution": "Institute of Science Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Taishi Nakamura", "gender": "unknown", "institution": "Institute of Science Tokyo", "country": "JP", "position": "MS student" }, { "name": "Takuya Akiba", "gender": "unknown", "institution": "Sakana AI", "country": "JP", "position": "Researcher" }, { "name": "Yusuke Oda", "gender": "Male", "institution": "National Institute of Informatics", "country": "", "position": "Project Associate Professor" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 41, "n_ref": 66, "n_ref_all": 87, "n_fig": 7, "n_tab": 11, "L_tab": 4301, "n_element_tab": 527, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 4449, "n_element_tab_1": 294, "formula_len_all": 340, "formula_len_all_1": 365, "len_all": 167455, "len_all_1": 60362, "len_abs": 1375, "len_title": 129, "len_sents": 48117, "len_sents_1": 23745, "n_sents": 376, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1383, "title": "Drop-Upcycling: Training Sparse Mixture of Experts with Partial Re-initialization", "abs": "The Mixture of Experts (MoE) architecture reduces the training and inference cost significantly compared to a dense model of equivalent capacity. Upcycling is an approach that initializes and trains an MoE model using a pre-trained dense model. While upcycling leads to initial performance gains, the training progresses slower than when trained from scratch, leading to suboptimal performance in the long term. We propose Drop-Upcycling - a method that effectively addresses this problem. Drop-Upcycling combines two seemingly contradictory approaches: utilizing the knowledge of pre-trained dense models while statistically re-initializing some parts of the weights. This approach strategically promotes expert specialization, significantly enhancing the MoE model's efficiency in knowledge acquisition. \nExtensive large-scale experiments demonstrate that Drop-Upcycling significantly outperforms previous MoE construction methods in the long term, specifically when training on hundreds of billions of tokens or more.\nAs a result, our MoE model with 5.9B active parameters achieves comparable performance to a 13B dense model in the same model family, while requiring approximately 1/4 of the training FLOPs.\nAll experimental resources, including source code, training data, model checkpoints and logs, are publicly available to promote reproducibility and future research on MoE.", "keywords": [ "mixture of experts", "large language models", "continual pre-training" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "gwZ90hFSL2", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Li Weigang", "gender": "Male", "institution": "Universidade de Bras\u00edlia", "country": "BR", "position": "Full Professor" }, { "name": "Pedro Carvalho Brom", "gender": "unknown", "institution": "Universidade de Bras\u00edlia", "country": "BR", "position": "PhD student" }, { "name": "Rafael Marconi Ramos", "gender": "Male", "institution": "Universidade de Bras\u00edlia", "country": "BR", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 18, "n_ref": 30, "n_ref_all": 46, "n_fig": 6, "n_tab": 1, "L_tab": 458, "n_element_tab": 41, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 225, "n_element_tab_1": 36, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 85848, "len_all_1": 58890, "len_abs": 1797, "len_title": 162, "len_sents": 32977, "len_sents_1": 30200, "n_sents": 215, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 113, "L_abs": 1803, "title": "Advancing Cross-Lingual Capabilities for Humanoid Robots: Leveraging Chinese NLP through Pictophonetic Advantages", "abs": "Humanoid robots, as a critical trajectory in the development of artificial intelligence, are poised to play a key role in the era of cross-lingual and multimodal intelligence. This paper explores the unique capabilities of humanoid robots in multilingual processing by harnessing the pictophonetic advantages inherent in the Chinese language. Unlike phonetic languages such as English, Chinese characters encapsulate ideographic, phonetic, and semantic components within a single symbol, providing a rich, multidimensional data source. By analyzing the successful localization of the periodic table in Chinese, this study illustrates how the unique naming conventions used by Chinese chemists bridge scientific and linguistic understanding. It advocates adopting the systematic approach seen in Chinese chemical nomenclature to further advance research in Chinese natural language processing (CNLP). To this end, the Six-Writings Pictophonetic Coding (SWPC) technology is introduced, which constructs efficient character and word matrices to enable humanoid robots to process Chinese language inputs effectively. The integration of SWPC with techniques such as Scale-Invariant Feature Transform (SIFT) and machine learning facilitates multimodal recognition of characters and words, allowing robots to prioritize Chinese information and seamlessly process it in conjunction with other languages. This approach has the potential to significantly enhance natural language understanding and generation in complex Chinese contexts. By drawing insights from Chinese chemical nomenclature, the paper lays a foundation for intelligent cross-lingual interactions, providing a new direction for CNLP research and paving the way for humanoid robots to achieve deeper integration into future intelligent societies.", "keywords": [ "Chinese NLP", "Cross-Lingual", "humanoid robots", "multimodal intelligence", "SIFT" ], "rating_list": [ 1, 1, 1 ], "soundness_list": [ 1, 1, 1 ], "presentation_list": [ 2, 1, 1 ], "contribution_list": [ 1, 1, 1 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "gwNQuVXEEV", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Alexis Battle", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Benjamin D. Shapiro", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 26, "n_ref": 34, "n_ref_all": 53, "n_fig": 12, "n_tab": 2, "L_tab": 474, "n_element_tab": 8, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 79, "formula_len_all_1": 275, "len_all": 98785, "len_all_1": 49848, "len_abs": 1235, "len_title": 85, "len_sents": 36279, "len_sents_1": 26243, "n_sents": 223, "n_sents_1": 157, "venue": "ICLR2025", "decision": "withdraw", "L_title": 37, "L_abs": 1239, "title": "Bayesian Tree-Dependent Factorization", "abs": "We propose Bayesian Tree-Dependent Factorization (BTF), a novel probabilistic representation learning model that uncovers hierarchical, continuous latent factors in complex datasets. BTF constructs a tree-based model that discovers interpretable factorizations of the data wherein each factor has a conditional relationship to its parent, allowing it to capture both global and local effects. This approach is particularly well-suited for biological data, where traditional methods like PCA fail to capture higher-order dependencies and hierarchical structure. A significant contribution of this work is the multi-view extension of BTF, which allows for the joint analysis of multiple data modalities. By learning shared loadings across views while maintaining distinct factors for each modality, multi-view BTF improves performance and enables deeper insights into the relationships between different data types. We demonstrate the performance of BTF in simulations as well as in a real-world application to gene expression and clinical data in breast cancer patients, revealing biologically and clinically meaningful patient trends, and showing that BTF is a valuable representation learning tool for analysis and hypothesis generation.", "keywords": [ "factorization", "Bayesian models", "multi-view", "hierarchical", "gene expression", "clinical" ], "rating_list": [ 3, 3, 1, 3 ], "soundness_list": [ 1, 2, 1, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "gwGHBD9ZKU", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Chufan Gao", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Jimeng Sun", "gender": "unknown", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Professor" }, { "name": "Namkyeong Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Siddhartha Laghuvarapu", "gender": "Male", "institution": "Department of Computer Science", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 90, "n_ref_all": 105, "n_fig": 8, "n_tab": 8, "L_tab": 3497, "n_element_tab": 332, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2617, "n_element_tab_1": 40, "formula_len_all": 291, "formula_len_all_1": 0, "len_all": 152624, "len_all_1": 52216, "len_abs": 1914, "len_title": 112, "len_sents": 51334, "len_sents_1": 25535, "n_sents": 361, "n_sents_1": 144, "venue": "ICLR2025", "decision": "withdraw", "L_title": 112, "L_abs": 1493, "title": "MolTextQA: A Curated Question-Answering Dataset and Benchmark for Molecular Structure-Text Relationship Learning", "abs": "Recent advancements in AI have significantly enhanced molecular representation learning, which is crucial for predicting molecule properties and designing new molecules. Despite these advances, effectively utilizing the vast amount of molecular data available in textual form from databases and scholarly articles remains a challenge. Recently, a large body of research has focused on utilizing Large Language Models (LLMs) and multi-modal architectures to interpret textual information and link it with molecular structures. Nevertheless, existing datasets often lack specificity in evaluation, as well as direct comparisons and comprehensive benchmarking across different models and model classes. In this work, we construct a dataset specifically designed for evaluating models on structure-directed questions and textual description-based molecule retrieval, featuring over 500,000 question-answer pairs related to approximately 240,000 molecules from PubChem. Its structure enhances evaluation specificity and precision through the use of multiple-choice answers. Moreover, we benchmark various architectural classes fine-tuned using this dataset, including multi-modal architectures, and large language models, uncovering several insights. Our experiments indicate that the Galactica and BioT5 models are the top performers in Molecule QA and Molecule Retrieval tasks respectively, achieving about 70% accuracy. We have made both the dataset and the fine-tuned models publicly available.", "keywords": [ "molecule-text learning", "question answering", "datasets", "benchmark", "large language models" ], "rating_list": [ 3, 8, 6, 3 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 1, 4, 3, 2 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "gw4hYNFUIC", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Guy Azran", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "PhD student" }, { "name": "Kai Yuan", "gender": "Male", "institution": "Intel", "country": "DE", "position": "Researcher" }, { "name": "Sarah Keren", "gender": "unknown", "institution": "Technion, Technion", "country": "IL", "position": "Assistant Professor" }, { "name": "Yuval Goshen", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 44, "n_ref": 81, "n_ref_all": 111, "n_fig": 10, "n_tab": 9, "L_tab": 4798, "n_element_tab": 739, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 648, "n_element_tab_1": 136, "formula_len_all": 222, "formula_len_all_1": 196, "len_all": 182229, "len_all_1": 70497, "len_abs": 1360, "len_title": 126, "len_sents": 51124, "len_sents_1": 33284, "n_sents": 416, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1361, "title": "S3E: Semantic Symbolic State Estimation With Vision-Language Foundation Models", "abs": "In automated task planning, state estimation is the process of translating an agent's sensor input into a high-level task state. It is important because real-world environments are unpredictable, and actions often do not lead to expected outcomes. State estimation enables the agent to manage uncertainties, adjust its plans, and make more informed decisions. Traditionally, researchers and practitioners relied on hand-crafted and hard-coded state estimation functions to determine the abstract state defined in the task domain. Recent advancements in Vision Language Models (VLMs) enable autonomous retrieval of semantic information from visual input. We present Semantic Symbolic State Estimation (S3E), the first general-purpose symbolic state estimator based on VLMs that can be applied in various settings without specialized coding or additional exploration. S3E takes advantage of the foundation model's internal world model and semantic understanding to assess the likelihood of certain symbolic components of the environment's state. We analyze S3E as a multi-label classifier, reveal different kinds of uncertainties that arise when using it, and show how they can be mitigated using natural language and targeted environment design. We show that S3E can achieve over 90\\% state estimation precision in our simulated and real-world robot experiments.", "keywords": [ "Task Planning", "Symbolic AI", "Computer Vision", "Vision-Language Models" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "gvmoBNuf5f", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Christian Igel", "gender": "Male", "institution": "University of Copenhagen", "country": "DK", "position": "Full Professor" }, { "name": "Nick Hauptvogel", "gender": "Male", "institution": "Copenhagen University", "country": "", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 69, "n_ref": 160, "n_ref_all": 184, "n_fig": 2, "n_tab": 4, "L_tab": 885, "n_element_tab": 161, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 719, "formula_len_all_1": 360, "len_all": 167046, "len_all_1": 68603, "len_abs": 1911, "len_title": 101, "len_sents": 49648, "len_sents_1": 34219, "n_sents": 345, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1920, "title": "On Uniform, Bayesian, and PAC-Bayesian Deep Ensembles", "abs": "It is common practice to combine deep neural networks to ensembles. These deep ensembles can profit from the cancellation of errors effect: Errors by ensemble members may average out and the ensemble achieves better generalization performance than each individual network. Bayesian neural networks learn a posterior distribution over model parameters, and sampling and weighting networks according to this posterior yields an ensemble model referred to as Bayes ensemble. In this study, we stress that neither the sampling nor the weighting in a Bayes ensemble are particularly well-suited for increasing generalization performance, as they do not support the cancellation of errors effect, which is evident in the limit from the Bernstein-von~Mises theorem for misspecified models. In contrast, a weighted average of models, where the weights are optimized by minimizing a PAC-Bayesian generalization bound, can improve generalization performance. This requires that the optimization takes correlations between models into account, which can be achieved by minimizing the tandem loss at the cost that hold-out data for estimating error correlations need to be available. The PAC-Bayesian weighting increases the robustness against correlated models and models with lower performance in an ensemble. This allows us to safely add several models from the same learning process to an ensemble, instead of using early-stopping for selecting a single weight configuration. Our study presents empirical results supporting these conceptual considerations on four different classification datasets. We show that state-of-the-art Bayes ensembles from the literature, despite being computationally demanding, do not improve over simple uniformly weighted deep ensembles and cannot match the performance of deep ensembles weighted by optimizing the tandem loss, which additionally come with non-vacuous generalization guarantees.", "keywords": [ "deep learning", "ensemble methods", "Bayesian model average", "PAC-Bayesian generalization bounds" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "gvk3XEjxIc", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Hao Xu", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Jia Pan", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Shijie Lin", "gender": "Male", "institution": "The University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Yupu Lu", "gender": "Male", "institution": "The University of Hong Kong", "country": "", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 24, "n_ref_uni": 31, "n_ref": 83, "n_ref_all": 111, "n_fig": 7, "n_tab": 4, "L_tab": 836, "n_element_tab": 65, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1371, "n_element_tab_1": 114, "formula_len_all": 2078, "formula_len_all_1": 2016, "len_all": 99388, "len_all_1": 69200, "len_abs": 1092, "len_title": 119, "len_sents": 31253, "len_sents_1": 29048, "n_sents": 236, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1097, "title": "Lyapunov Stability Learning with Nonlinear Control via Inductive Biases", "abs": "Finding a control Lyapunov function (CLF) in a dynamical system with a controller is an effective way to guarantee stability, which is a crucial issue in safety-concerned applications. Recently, deep learning models representing CLFs have been applied into a learner-verifier framework to identify satisfiable candidates. However, the learner treats Lyapunov conditions as complex constraints for optimisation, which is hard to achieve global convergence. It is also too complicated to implement these Lyapunov conditions for verification. To improve this framework, we treat Lyapunov conditions as inductive biases and design a neural CLF and a CLF-based controller guided by this knowledge. This design enables a stable optimisation process with limited constraints, and allows end-to-end learning of both the CLF and the controller. Our approach achieves higher convergence rate and larger region of attraction (ROA) in learning the CLF compared to existing methods among abundant experiment cases. We also thoroughly reveal why the success rate decreases with previous methods during learning.", "keywords": [ "Lyapunov Stability", "Self-supervised Learning", "Region of Attraction", "Nonlinear Dynamics" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "gvZpk0n68q", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Yue Zhang", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yuke Zhu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Sheng Guo", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 13, "n_ref": 27, "n_ref_all": 33, "n_fig": 4, "n_tab": 9, "L_tab": 2085, "n_element_tab": 264, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1392, "n_element_tab_1": 110, "formula_len_all": 508, "formula_len_all_1": 492, "len_all": 132145, "len_all_1": 68359, "len_abs": 1190, "len_title": 121, "len_sents": 37316, "len_sents_1": 30351, "n_sents": 393, "n_sents_1": 295, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1137, "title": "NAText: Faster Scene Text Recognition with Non Autoregressive Transformer", "abs": "Autoregressive-based attention methods have made significant advance in scene text recognition. However, the inference speed of these methods is limited due to their iterative decoding scheme. In contrast, the non-autoregressive methods has a parallel decoding paradigm, making them much faster than the autoregressive decoder. The dilemma is that, though the speed is increased, the non-autoregressive methods are based on the character-wise independent assumption, making them perform much worse than the autoregressive methods. In this paper, we propose a simple non-autoregressive transformer-based text recognizer, named NAText, by proposing a progressive learning approach to force the network to learn the relationship between characters. Furthermore, we redesign the query composition by introducing positional encoding of the character center. And it has clear physical meanings than the conventional one. Experiments show that our NAText helps to better utilize the positional information for 2D feature aggregation. With all these techniques, the NAText has achieved competitive performance to the state-of-the-art methods.", "keywords": [ "Scene Text Recognition", "Non Autogressive" ], "rating_list": [ 5, 6, 3, 1, 5 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4, 5 ] }, { "paper_id": "gv8176NnO0", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Frank Hutter", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen & University of Freiburg", "country": "DE", "position": "Full Professor" }, { "name": "Noah Hollmann", "gender": "unknown", "institution": "Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "", "position": "Researcher" }, { "name": "Samuel M\u00fcller", "gender": "unknown", "institution": "University of Freiburg, Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 16, "n_ref": 28, "n_ref_all": 57, "n_fig": 12, "n_tab": 2, "L_tab": 465, "n_element_tab": 31, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 465, "n_element_tab_1": 31, "formula_len_all": 684, "formula_len_all_1": 684, "len_all": 79950, "len_all_1": 55460, "len_abs": 1340, "len_title": 111, "len_sents": 25908, "len_sents_1": 24794, "n_sents": 192, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1347, "title": "Bayes' Power for Explaining In-Context Learning Generalizations", "abs": "Traditionally, neural network training has been primarily viewed as an approximation of maximum likelihood estimation (MLE).\nThis interpretation originated in a time when training for multiple epochs on small datasets was common and performance was data bound; but it falls short in the era of large-scale single-epoch trainings ushered in by large self-supervised setups, like language models.\nIn this new setup, performance is compute-bound, but data is readily available.\nAs models became more powerful, in-context learning (ICL), i.e., learning in a single forward-pass based on the context, emerged as one of the dominant paradigms.\nIn this paper, we argue that a more useful interpretation of neural network behavior in this era is as an approximation of the true posterior, as defined by the data-generating process.\nWe demonstrate this interpretations' power for ICL and its usefulness to predict generalizations to previously unseen tasks. We show how models become robust in-context learners by effectively composing knowledge from their training data.\nWe illustrate this with experiments that reveal surprising generalizations, all explicable through the exact posterior.\nFinally, we show the inherent constraints of the generalization capabilities of posteriors and the limitations of neural networks in approximating these posteriors.", "keywords": [ "in-context learning", "large language models", "PFN", "TabPFN" ], "rating_list": [ 5, 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 2, 4, 3, 3, 4 ], "contribution_list": [ 2, 2, 3, 2, 1 ], "confidence_list": [ 4, 4, 3, 4, 5 ] }, { "paper_id": "guyICBe4p1", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Ilia Markov", "gender": "Male", "institution": "Vrije Universiteit Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Peter Bloem", "gender": "Male", "institution": "Vrije Universiteit Amsterdam", "country": "", "position": "Postdoc" }, { "name": "Piek Vossen", "gender": "Male", "institution": "Vrije Universiteit Amsterdam", "country": "", "position": "Full Professor" }, { "name": "Stefan Frederik Schouten", "gender": "Male", "institution": "Vrije Universiteit Amsterdam", "country": "NL", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 11, "n_ref": 32, "n_ref_all": 52, "n_fig": 6, "n_tab": 2, "L_tab": 3506, "n_element_tab": 116, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 634, "formula_len_all_1": 634, "len_all": 144548, "len_all_1": 68110, "len_abs": 1358, "len_title": 80, "len_sents": 54705, "len_sents_1": 32325, "n_sents": 433, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1364, "title": "Truth-value judgment in language models: belief directions are context sensitive", "abs": "Recent work has demonstrated that the latent spaces of large language models\n(LLMs) contain directions predictive of the truth of sentences. Multiple methods recover such directions and build probes that are described as uncovering\na model\u2019s \u201cknowledge\u201d or \u201cbeliefs\u201d. We investigate this phenomenon, looking\nclosely at the impact of context on the probes. Our experiments establish where\nin the LLM the probe\u2019s predictions are (most) sensitive to the presence of related\nsentences, and how to best characterize this kind of sensitivity. We do so by\nmeasuring different types of consistency errors that occur after probing an LLM\nwhose inputs consist of hypotheses preceded by (negated) supporting and contradicting sentences. We also perform a causal intervention experiment, investigating whether moving the representation of a premise along these belief directions\ninfluences the position of an entailed or contradicted sentence along that same direction. We find that the probes we test are generally context sensitive, but that\ncontexts which should not affect the truth often still impact the probe outputs.\nOur experiments show that the type of errors depend on the layer, the model, and\nthe kind of data. Finally, our results suggest that belief directions are (one of the)\ncausal mediators in the inference process that incorporates in-context information.", "keywords": [ "interpretability", "truth directions", "LLM beliefs", "large language model", "llm" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 4, 3, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "gtVo4xcpFI", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Gwok-Waa Wan", "gender": "Male", "institution": "National Center of Technology Innovation for Electronic Design Automation", "country": "CN", "position": "Intern" }, { "name": "Mengnv Xing", "gender": "Female", "institution": "eda", "country": "", "position": "Principal Researcher" }, { "name": "Nan Guan", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Ning Xu", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Qiang Xu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Wang yubo", "gender": "unknown", "institution": "Northeastern University", "country": "CN", "position": "MS student" }, { "name": "Xi Wang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhe jiang", "gender": "Male", "institution": "Southeast University", "country": "", "position": "Full Professor" }, { "name": "jingyi zhang", "gender": "unknown", "institution": "Southeast Community College Area", "country": "CN", "position": "MS student" }, { "name": "ying wang", "gender": "Not Specified", "institution": ", Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Xinze Wang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 40, "n_ref": 67, "n_ref_all": 89, "n_fig": 7, "n_tab": 8, "L_tab": 11104, "n_element_tab": 1301, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1456, "n_element_tab_1": 91, "formula_len_all": 50, "formula_len_all_1": 50, "len_all": 177843, "len_all_1": 46786, "len_abs": 1378, "len_title": 98, "len_sents": 37112, "len_sents_1": 20325, "n_sents": 272, "n_sents_1": 127, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1386, "title": "GenBen:A Genarative Benchmark for LLM-Aided Design", "abs": "This paper introduces GenBen, a generative benchmark designed to evaluate the capabilities of large language models (LLMs) in hardware design. With the rapid advancement of LLM-aided design (LAD), it has become crucial to assess the effectiveness of these models in automating hardware design processes.\nExisting benchmarks primarily focus on hardware code generation and often neglect critical aspects such as Quality-of-Result (QoR) metrics, design diversity, modality, and test set contamination. GenBen is the first open-source, generative benchmark tailored for LAD that encompasses a range of tasks, from high-level architecture to low-level circuit optimization, and includes diverse, silicon-proven hardware designs. \nWe have also designed a difficulty tiering mechanism to provide fine-grained insights into enhancements of LLM-aided designs. Through extensive evaluations of several state-of-the-art LLMs using GenBen, we reveal their strengths and weaknesses in hardware design automation. Our findings are based on 10,920 experiments and 2,160 hours of evaluation, underscoring the potential of this work to significantly advance the LAD research community. \nIn addition, both GenBen employs an end-to-end testing infrastructure to ensure consistent and reproducible results across different LLMs. The benchmark is available at https://anonymous.4open.science/r/GENBEN-2812.", "keywords": [ "GenBen; Benchmark; LLM-Aided Design; LLM; Hardware Design" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "gtCXzVeQxz", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Gustavo Carneiro", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Full Professor" }, { "name": "Hsiang-Ting Chen", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Assistant Professor" }, { "name": "Pitawelayalage Dasun Dileepa Pitawela", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 10, "n_ref_uni": 48, "n_ref": 75, "n_ref_all": 126, "n_fig": 2, "n_tab": 19, "L_tab": 7343, "n_element_tab": 656, "n_fig_1": 1, "n_tab_1": 9, "L_tab_1": 3376, "n_element_tab_1": 367, "formula_len_all": 2476, "formula_len_all_1": 950, "len_all": 199791, "len_all_1": 71910, "len_abs": 1335, "len_title": 153, "len_sents": 53063, "len_sents_1": 29623, "n_sents": 378, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 105, "L_abs": 1340, "title": "PHICO: Personalised Human-AI Cooperative Classification Using Augmented Noisy Labels and Model Prediction", "abs": "The nuanced differences in human behavior and the complex dynamics of human-AI interactions pose significant challenges in optimizing human-AI cooperation. Existing approaches tend to oversimplify the problem and rely on a single global behavior model, which overlooks individual variability, leading to sub-optimal solutions. To bridge this gap, we introduce PHICO, a novel framework for human-AI cooperative classification that initially identifies a set of representative annotator profiles characterized by unique noisy label patterns. These patterns are then augmented to train personalised AI cooperative models, each tailored to an annotator profile. When these models are paired with human inputs that exhibit similar noise patterns from a corresponding profile, they consistently achieve a joint classification accuracy that exceeds those achieved by either AI or humans alone. We theoretically prove the convergence of PHICO, ensuring the reliability of the framework. To evaluate PHICO, we introduce novel measures for assessing human-AI cooperative classification and empirically demonstrate its generalisability and performance across diverse datasets including CIFAR-10N, CIFAR-10H, Fashion-MNIST-H, AgNews, and Chaoyang histopathology. PHICO is both a model-agnostic and effective solution for improving human-AI cooperation.", "keywords": [ "Human-Ai Cooperation" ], "rating_list": [ 3, 3, 3, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "gsShHPxkUW", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Chaochao Lu", "gender": "unknown", "institution": "Shanghai AI Laboratory ", "country": "CN", "position": "Research Scientist" }, { "name": "Difan Zou", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Sirui Chen", "gender": "unknown", "institution": "Tongji University", "country": "", "position": "PhD student" }, { "name": "Yujin Han", "gender": "Female", "institution": "the University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xu Lei", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" } ], "n_formula": 33, "n_formula_1": 11, "n_ref_uni": 58, "n_ref": 115, "n_ref_all": 170, "n_fig": 9, "n_tab": 9, "L_tab": 3331, "n_element_tab": 178, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 579, "n_element_tab_1": 27, "formula_len_all": 3314, "formula_len_all_1": 894, "len_all": 230147, "len_all_1": 60122, "len_abs": 2042, "len_title": 125, "len_sents": 74648, "len_sents_1": 29132, "n_sents": 569, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 2134, "title": "Beyond Surface Structure: A Causal Assessment of LLMs' Comprehension ability", "abs": "Large language models (LLMs) have shown remarkable capability in natural language tasks, yet debate persists on whether they truly comprehend deep structure (i.e., core semantics) or merely rely on surface structure (e.g., presentation format). Prior studies observe that LLMs' performance declines when intervening on surface structure, arguing their success relies on surface structure recognition. However, surface structure sensitivity does not prevent deep structure comprehension. Rigorously evaluating LLMs' capability requires analyzing both, yet deep structure is often overlooked. To this end, we assess LLMs' comprehension ability using causal mediation analysis, aiming to fully discover the capability of using both deep and surface structures. Specifically, we formulate the comprehension of deep structure as direct causal effect (DCE) and that of surface structure as indirect causal effect (ICE), respectively. To address the non-estimability of original DCE and ICE --- stemming from the infeasibility of isolating mutual influences of deep and surface structures, we develop the corresponding quantifiable surrogates, including approximated DCE (ADCE) and approximated ICE (AICE). We further apply the ADCE to evaluate a series of mainstream LLMs (and the one with random weights), showing that most of them exhibit deep structure comprehension ability, which grows along with the prediction accuracy. Comparing ADCE and AICE demonstrates closed-source LLMs (e.g., GPT) rely more on deep structure, while open-source LLMs (e.g., Llama) are more surface-sensitive, which decreases with model scale. Theoretically, ADCE is a bidirectional evaluation, which measures both the sufficiency and necessity of deep structure changes in causing output variations, thus offering a more comprehensive assessment than accuracy, a common evaluation in LLMs. Our work provides new insights into LLMs' deep structure comprehension and offers novel methods for LLMs evaluation. The code for our project is available at [https://anonymous.4open.science](https://anonymous.4open.science/r/DiD-based-Causal-Mediation-Framework-417A).", "keywords": [ "Large Language Models", "Causality", "Interpretability" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "grU1VKEOLi", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Haoyu Geng", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "", "position": "PhD student" }, { "name": "Jiale Ma", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Nianzu Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Runzhong Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Wenzheng Pan", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yang Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 36, "n_formula_1": 11, "n_ref_uni": 45, "n_ref": 139, "n_ref_all": 181, "n_fig": 11, "n_tab": 16, "L_tab": 8155, "n_element_tab": 927, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 4380, "n_element_tab_1": 339, "formula_len_all": 2305, "formula_len_all_1": 581, "len_all": 248335, "len_all_1": 76270, "len_abs": 2285, "len_title": 151, "len_sents": 86292, "len_sents_1": 31816, "n_sents": 639, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1591, "title": "Streamlining the Design Space of ML4TSP Suggests Principles for Learning and Search", "abs": "Despite the rich works on machine learning (ML) for combinatorial optimization (CO), a unified, principled framework remains lacking. This study utilizes the Travelling Salesman Problem (TSP) as a major case study, with adaptations demonstrated for other CO problems, dissecting established mainstream learning-based solvers to outline a comprehensive design space. It advances a unified modular streamline incorporating existing technologies in both learning and search for transparent ablation, aiming to reassess the role of learning and to discern which parts of existing techniques are genuinely beneficial and which are not. This further leads to the investigation of desirable principles of learning designs and the exploration of concepts guiding method designs. We demonstrate the desirability of principles such as joint probability estimation, symmetry solution representation, and online optimization for learning-based designs. Leveraging the findings, we propose enhancements to existing methods to compensate for their missing attributes, thereby advancing performance and enriching the technique library. From a higher viewpoint, we also uncover a performance advantage in non-autoregressive and supervised paradigms compared to their counterparts. The strategic decoupling and organic recompositions yield a factory of new TSP solvers, where we investigate synergies across various method combinations and pinpoint the optimal design choices to create more powerful ML4TSP solvers, thereby facilitating and offering a reference for future research and engineering endeavors.", "keywords": [ "Neural Combinatorial Optimization", "Travelling salesman problem" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "grM2Yv49cI", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Houman Owhadi", "gender": "Male", "institution": "California Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Theo Bourdais", "gender": "Male", "institution": "California Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 110, "n_formula_1": 30, "n_ref_uni": 33, "n_ref": 47, "n_ref_all": 79, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8167, "formula_len_all_1": 1950, "len_all": 166160, "len_all_1": 67779, "len_abs": 1372, "len_title": 160, "len_sents": 48570, "len_sents_1": 29121, "n_sents": 458, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 110, "L_abs": 1377, "title": "Minimal Variance Model Aggregation: A principled, non-intrusive, and versatile integration of black box models", "abs": "Whether deterministic or stochastic, models can be viewed as functions designed to approximate a specific quantity of interest. \nWe introduce Minimal Empirical Variance Aggregation (MEVA), a data-driven framework that integrates predictions from various models, enhancing overall accuracy by leveraging the individual strengths of each. This non-intrusive, model-agnostic approach treats the contributing models as black boxes and accommodates outputs from diverse methodologies, including machine learning algorithms and traditional numerical solvers.\nWe advocate for a point-wise linear aggregation process and consider two methods for optimizing this aggregate: Minimal Error Aggregation (MEA), which minimizes the prediction error, and Minimal Variance Aggregation (MVA), which focuses on reducing variance. We prove a theorem showing that MVA can be more robustly estimated from data than MEA, making MEVA superior to Minimal Empirical Error Aggregation (MEEA). Unlike MEEA, which interpolates target values directly, MEVA formulates aggregation as an error estimation problem, which can be performed\nusing any backbone learning paradigm. We demonstrate the versatility and effectiveness of our framework across various applications, including data science and partial differential equations, illustrating its ability to significantly enhance both robustness and accuracy.", "keywords": [ "Model aggregation; Ensemble Learning; Scientific Machine Learning; Gaussian Processes; Neural operators" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "gqeXXrIMr0", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Bart\u0142omiej Sadlej", "gender": "Male", "institution": "University of Warsaw", "country": "PL", "position": "MS student" }, { "name": "Jakub Grzywaczewski", "gender": "unknown", "institution": "Warsaw University of Technology", "country": "PL", "position": "Undergrad student" }, { "name": "Matthew Tivnan", "gender": "Male", "institution": "Massachusetts General Hospital, Harvard University", "country": "US", "position": "Instructor" }, { "name": "Przemyslaw Biecek", "gender": "unknown", "institution": "Warsaw University of Technology", "country": "", "position": "Associate Professor" }, { "name": "Bartlomiej Jan Sobieski", "gender": "Male", "institution": "University of Warsaw", "country": "PL", "position": "PhD student" } ], "n_formula": 45, "n_formula_1": 8, "n_ref_uni": 0, "n_ref": 119, "n_ref_all": 186, "n_fig": 14, "n_tab": 11, "L_tab": 4334, "n_element_tab": 1005, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 779, "n_element_tab_1": 97, "formula_len_all": 3368, "formula_len_all_1": 470, "len_all": 350873, "len_all_1": 70400, "len_abs": 1406, "len_title": 119, "len_sents": 85545, "len_sents_1": 32170, "n_sents": 573, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1414, "title": "Rethinking Visual Counterfactual Explanations Through Region Constraint", "abs": "Visual counterfactual explanations (VCEs) have recently gained immense popularity as a tool for clarifying the decision-making process of image classifiers. This trend is largely motivated by what these explanations promise to deliver -- indicate semantically meaningful factors that change the classifier's decision. However, we argue that current state-of-the-art approaches lack a crucial component -- the region constraint -- whose absence prevents from drawing explicit conclusions, and may even lead to faulty reasoning due to phenomenons like confirmation bias. To address the issue of previous methods, which modify images in a very entangled and widely dispersed manner, we propose region-constrained VCEs (RVCEs), which assume that only a predefined image region can be modified to influence the model's prediction. To effectively sample from this subclass of VCEs, we propose Region-Constrained Counterfactual Schr\u00f6dinger Bridge (RCSB), an adaptation of a tractable subclass of Schr\u00f6dinger Bridges to the problem of conditional inpainting, where the conditioning signal originates from the classifier of interest. In addition to setting a new state-of-the-art by a large margin, we extend RCSB to allow for exact counterfactual reasoning, where the predefined region contains only the factor of interest, and incorporating the user to actively interact with the RVCE by predefining the regions manually.", "keywords": [ "visual", "counterfactual", "explanations", "diffusion", "generative", "explainable" ], "rating_list": [ 6, 5, 6, 5, 8 ], "soundness_list": [ 3, 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 3, 4 ] }, { "paper_id": "gqbbL7k8BF", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Boyang Sun", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Guang-Yuan Hao", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Yu Yao", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Lecturer" }, { "name": "Yumou Qiu", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 119, "n_formula_1": 25, "n_ref_uni": 35, "n_ref": 72, "n_ref_all": 97, "n_fig": 9, "n_tab": 1, "L_tab": 461, "n_element_tab": 34, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9969, "formula_len_all_1": 1775, "len_all": 182909, "len_all_1": 71555, "len_abs": 1259, "len_title": 114, "len_sents": 62716, "len_sents_1": 32671, "n_sents": 588, "n_sents_1": 293, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1307, "title": "A Conditional Independence Test in the Presence of Discretization", "abs": "Testing conditional independence (CI) has many important applications, such as Bayesian network learning and causal discovery. Although several approaches have been developed for learning CI structures for observed variables, those existing methods generally fail to work when the variables of interest can not be directly observed and only discretized values of those variables are available. For example, if $X_1$, $\\tilde{X}_2$ and $X_3$ are the observed variables, where $\\tilde{X}_2$ is a discretization of the latent variable $X_2$, applying the existing methods to the observations of $X_1$, $\\tilde{X}_2$ and $X_3$ would lead to a false conclusion about the underlying CI of variables $X_1$, $X_2$ and $X_3$.\nMotivated by this, we propose a CI test specifically designed to accommodate the presence of discretization. To achieve this, a bridge equation and nodewise regression are used to recover the precision coefficients reflecting the conditional dependence of the latent continuous variables under the nonparanormal model. An appropriate test statistic has been proposed, and its asymptotic distribution under the null hypothesis of CI has been derived.\nTheoretical analysis, along with empirical validation on various datasets, rigorously demonstrates the effectiveness of our testing methods.", "keywords": [ "Conditional Independence Test", "Discretization", "Causal Discovery" ], "rating_list": [ 3, 6, 5, 8, 6 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 2, 4, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 4, 3 ] }, { "paper_id": "gqC0egRfWq", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Ally Yalei Du", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Undergrad student" }, { "name": "Dravyansh Sharma", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Postdoc" }, { "name": "Eric Huang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "MS student" } ], "n_formula": 107, "n_formula_1": 22, "n_ref_uni": 27, "n_ref": 57, "n_ref_all": 67, "n_fig": 1, "n_tab": 2, "L_tab": 251, "n_element_tab": 31, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 1740, "n_element_tab_1": 20, "formula_len_all": 16250, "formula_len_all_1": 2338, "len_all": 160392, "len_all_1": 71759, "len_abs": 1371, "len_title": 136, "len_sents": 48542, "len_sents_1": 29571, "n_sents": 504, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1390, "title": "Theoretical Analyses of Hyperparameter Selection in Graph-Based Semi-Supervised Learning", "abs": "Graph-based semi-supervised learning is a powerful paradigm in machine learning for modeling and exploiting the underlying graph structure that captures the relationship between labeled and unlabeled data. A large number of classical as well as modern deep learning based algorithms have been proposed for this problem, often having tunable hyperparameters. We initiate a formal study of tuning algorithm hyperparameters from parameterized algorithm families for this problem. We obtain novel $O(\\log n)$ pseudo-dimension upper bounds for hyperparameter selection in three classical label propagation-based algorithm families, where $n$ is the number of nodes, implying bounds on the amount of data needed for learning provably good parameters. We further provide matching $\\Omega(\\log n)$ pseudo-dimension lower bounds, thus asymptotically characterizing the learning-theoretic complexity of the parameter tuning problem. We extend our study to selecting architectural hyperparameters in modern graph neural networks. We bound the Rademacher complexity for tuning the self-loop weighting in recently proposed Simplified Graph Convolution (SGC) networks. We further propose a tunable architecture that interpolates graph convolutional neural networks (GCN) and graph attention networks (GAT) in every layer, and provide Rademacher complexity bounds for tuning the interpolation coefficient.", "keywords": [ "graph-based semi-supervised learning", "hyperparameter selection", "sample complexity", "pseudo-dimension", "Rademacher complexity" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 1, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "gpP68EP1Jn", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Jie Zhou", "gender": "Male", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Le Tian", "gender": "Male", "institution": "Tencent Wechat AI", "country": "CN", "position": "Researcher" }, { "name": "Yuan Liu", "gender": "Male", "institution": "WeChat AI", "country": "CN", "position": "Researcher" }, { "name": "Zhongyin Zhao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Zhou Xiao", "gender": "Male", "institution": "Pattern recognition center", "country": "CN", "position": "Researcher" }, { "name": "Ziyuan Zhuang", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 185, "n_ref_all": 207, "n_fig": 9, "n_tab": 3, "L_tab": 1106, "n_element_tab": 123, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 1098, "n_element_tab_1": 114, "formula_len_all": 478, "formula_len_all_1": 424, "len_all": 205740, "len_all_1": 67158, "len_abs": 2062, "len_title": 119, "len_sents": 43022, "len_sents_1": 31343, "n_sents": 302, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 2072, "title": "POINTS: Improving Your Vision-language Model with Affordable Strategies", "abs": "In recent years, vision-language models have achieved significant advancements, excelling in tasks once deemed challenging, such as optical character recognition and geometric problem-solving. Despite these impressive achievements, several critical issues remain unaddressed: 1) Proprietary models rarely disclose detailed information about their architectures. In contrast, while open-source models provide visibility into their training strategies, detailed ablations of these strategies are highly anticipated. 2) Pre-training data is currently under-explored in open-source works, with most efforts empirically adding datasets from diverse sources, making the entire process elusive and cumbersome. 3) During the fine-tuning stage, the focus is often on adding and ablating more datasets, which frequently leads to diminishing returns. Therefore, refining data schemes is essential for further enhancing model performance.\nTo address these issues, we propose the following contributions in this paper: 1) We trained a robust baseline model, leveraging the latest technological advancements in vision-language models. Building upon existing advancements, we introduced effective improvements and conducted comprehensive ablation and validation for each technique incorporated into this strong baseline.\n2) Inspired by recent work on large language models, we propose filtering pre-training data using perplexity, selecting the data with the lowest perplexity as the training set. This approach allowed us to train on a curated 1M dataset, resulting in highly competitive performance. 3) During the visual instruction tuning stage, we experimented with model soup on different datasets when further introducing more datasets into the training set brought marginal improvements. Integrating these innovations, we obtained a model with 9B parameters, performing competitively with a series of existing state-of-the-art models. Additionally, these strategies we propose are efficient and relatively lightweight, allowing the community to adopt them easily for their models.", "keywords": [ "vision-language model", "multimodal" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "gpKEDj9Dgg", "primary_area": "generative models", "n_author": 1, "author_info_list": [ { "name": "Abdulhameed Abiola Dere", "gender": "Male", "institution": "University of Ilorin", "country": "NG", "position": "Undergrad student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 8, "n_ref": 8, "n_ref_all": 13, "n_fig": 1, "n_tab": 1, "L_tab": 36, "n_element_tab": 9, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 36, "n_element_tab_1": 9, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 30050, "len_all_1": 30168, "len_abs": 1320, "len_title": 173, "len_sents": 8776, "len_sents_1": 8783, "n_sents": 50, "n_sents_1": 51, "venue": "ICLR2025", "decision": "Reject", "L_title": 125, "L_abs": 1326, "title": "Optimizing Large Language Models with Automatic Speech Recognition for Medication Corpus in Low-Resource Healthcare Settings.", "abs": "Automatic Speech Recognition (ASR) systems, while effective in general contexts, often face challenges in low-resource settings, especially in specialized domains such as healthcare. This study investigates the integration of Large Language Models (LLMs) with ASR systems to improve transcription accuracy in such environments. Focusing on medication-related conversations in healthcare, we fine-tuned the Whisper-Large ASR model on a custom dataset, Pharma-Speak, and applied the LLaMA 3 model for second-pass rescoring to correct ASR output errors. To achieve efficient fine-tuning without altering the full LLM parameters, we employed Low-Rank Adaptation (LoRA), which enables re-ranking of the ASR\u2019s N-best hypotheses while retaining the LLM's original knowledge. \nOur results demonstrate a significant reduction in Word Error Rate (WER) across multiple epochs, validating the effectiveness of the LLM-based rescoring method. The integration of LLMs in this framework shows potential for overcoming the limitations posed by conventional ASR models in low-resource settings. While computational constraints and the inherent strength of Whisper-Large presented some limitations, our approach lays the groundwork for further exploration of domain-specific ASR enhancements using LLMs, particularly in healthcare applications.", "keywords": [ "Automatic Speech Recognition", "Large Language Models", "Healthcare", "Low Resource Settings" ], "rating_list": [ 1, 5, 1, 1 ], "soundness_list": [ 1, 2, 1, 1 ], "presentation_list": [ 1, 1, 1, 1 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "gp5tRHkz9B", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Andrew Gordon Wilson", "gender": "Not Specified", "institution": "New York University", "country": "US", "position": "Full Professor" }, { "name": "Bethmage Mayuka Jayawardhana", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Colin White", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Frank Hutter", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen & University of Freiburg", "country": "DE", "position": "Full Professor" }, { "name": "Micah Goldblum", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" }, { "name": "Renbo Tu", "gender": "unknown", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Samuel Dooley", "gender": "unknown", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Tom Goldstein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Valeriia Cherepanova", "gender": "Female", "institution": "Amazon", "country": "", "position": "Postdoc" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 23, "n_ref": 48, "n_ref_all": 61, "n_fig": 8, "n_tab": 6, "L_tab": 26145, "n_element_tab": 2349, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1599, "n_element_tab_1": 36, "formula_len_all": 44, "formula_len_all_1": 44, "len_all": 160593, "len_all_1": 55486, "len_abs": 1301, "len_title": 128, "len_sents": 31429, "len_sents_1": 25883, "n_sents": 248, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1308, "title": "LLMs Boost the Performance of Decision Trees on Tabular Data across Sample Sizes", "abs": "Large language models (LLMs) perform remarkably well on tabular datasets in zero- and few-shot settings, since they can extract meaning from natural language column headers that describe features and labels. In contrast to LLMs, gradient-boosted decision trees (GBDTs) must learn the relationships among columns from scratch, increasing their data requirements. Meanwhile, LLMs are not competitive with GBDTs on medium or large datasets, and their scalability is capped by their limited context lengths. In this paper, we propose LLM-Boost, a simple and lightweight approach for fusing large language models with gradient-boosted decision trees, which enables larger datasets to benefit from the natural language capabilities of LLMs than was previously shown. While matching LLMs at sufficiently small dataset sizes and GBDTs at sufficiently large sizes, LLM-Boost outperforms both standalone models on a wide range of dataset sizes in between. We demonstrate state-of-the-art performance against numerous baselines and ensembling approaches, and we also show how to fuse GBDTs with TabPFN, a recent non-LLM model for in-context learning on tabular data. We find that this combination achieves the best performance on larger datasets. We release our code at https://anonymous.4open.science/r/LLM-Boost-21DD.", "keywords": [ "Tabular data", "large language models", "decision trees", "ensembling" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 1, 1, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "gp32jvUquq", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Bing Li", "gender": "unknown", "institution": "Universit\u00e4t Siegen", "country": "DE", "position": "Assistant Professor" }, { "name": "Grace Li Zhang", "gender": "Female", "institution": "Technische Universit\u00e4t Darmstadt", "country": "", "position": "Assistant Professor" }, { "name": "Ing-Chao Lin", "gender": "unknown", "institution": "National Cheng Kung University", "country": "TW", "position": "Full Professor" }, { "name": "Jingcun Wang", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Yu-Guang Chen", "gender": "unknown", "institution": "National Central University", "country": "TW", "position": "Associate Professor" } ], "n_formula": 21, "n_formula_1": 12, "n_ref_uni": 36, "n_ref": 58, "n_ref_all": 80, "n_fig": 7, "n_tab": 9, "L_tab": 3609, "n_element_tab": 509, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 2827, "n_element_tab_1": 397, "formula_len_all": 2139, "formula_len_all_1": 632, "len_all": 143124, "len_all_1": 63300, "len_abs": 1028, "len_title": 129, "len_sents": 38564, "len_sents_1": 24226, "n_sents": 327, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 998, "title": "Basis Sharing: Cross-Layer Parameter Sharing for Large Language Model Compression", "abs": "Large Language Models (LLMs) have achieved remarkable breakthroughs. However, the huge number of parameters in LLMs require significant amount of memory storage in inference, which prevents their practical deployment in many applications. To reduce memory storage of LLMs, singular value decomposition (SVD) provides a promising solution to approximate weight matrices for compressing LLMs. In this paper, we take a step further to explore parameter sharing across different layers with SVD to achieve more effective compression for LLMs. Specifically, weight matrices in different layers are decomposed and represented with a linear combination of a set of shared basis vectors and unique coefficients. The types of weight matrices and the layer selection for basis sharing are examined when compressing LLMs to maintain the performance. Comprehensive experiments demonstrate that Basis-Sharing outperforms state-of-the-art SVD-based compression approaches, especially at large compression ratios.", "keywords": [ "Cross Layer Parameter Sharing in LLMs", "LLM Compression with SVD" ], "rating_list": [ 5, 8, 5, 8 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "goFpCuJalN", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Duncan Watson-Parris", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Leon Bergen", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Rose Yu", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Spencer Ho", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "Undergrad student" }, { "name": "Srikar Eranky", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Undergrad student" }, { "name": "Taylor Berg-Kirkpatrick", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "Associate Professor" }, { "name": "Veeramakali Vignesh Manivannan", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Yasaman Jafari", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Yian Ma", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 37, "n_ref_all": 54, "n_fig": 5, "n_tab": 5, "L_tab": 3354, "n_element_tab": 186, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1401, "n_element_tab_1": 116, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 157924, "len_all_1": 49553, "len_abs": 814, "len_title": 129, "len_sents": 43608, "len_sents_1": 21393, "n_sents": 314, "n_sents_1": 146, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 823, "title": "ClimaQA: An Automated Evaluation Framework for Climate Foundation Models", "abs": "The use of Large Language Models (LLMs) in climate science has recently gained significant attention. However, a critical issue remains: the lack of a comprehensive evaluation framework capable of assessing the quality and scientific validity of model outputs. To address this issue, we develop *ClimaGen* (Climate QA Generator), an adaptive learning framework that generates question-answer pairs from graduate textbooks with climate scientists in the loop. As a result, we present *ClimaQA-Gold*, an expert-annotated benchmark dataset alongside *ClimaQA-Silver*, a large-scale, comprehensive synthetic QA dataset for climate science. Finally, we develop evaluation strategies and compare different LLMs on our benchmarks. Our results offer novel insights into various approaches used to enhance knowledge of climate LLMs.", "keywords": [ "Climate Benchmark", "Scientific Foundation Models", "Scientific Question Answering", "Large Language Models", "Automated QA generation" ], "rating_list": [ 6, 6, 8, 6, 8 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 4 ], "confidence_list": [ 5, 3, 4, 4, 3 ] }, { "paper_id": "goBaGHLAdP", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Brisa Maneechotesuwan", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Chengyue Huang", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Junjiao Tian", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Shivang Chopra", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zsolt Kira", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 25, "n_ref": 63, "n_ref_all": 67, "n_fig": 2, "n_tab": 5, "L_tab": 2434, "n_element_tab": 374, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 978, "n_element_tab_1": 152, "formula_len_all": 1239, "formula_len_all_1": 1359, "len_all": 121653, "len_all_1": 56401, "len_abs": 1338, "len_title": 123, "len_sents": 35307, "len_sents_1": 25291, "n_sents": 301, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1375, "title": "Directional Gradient Projection for Robust Fine-tuning of Foundation Models", "abs": "Robust fine-tuning aims to adapt large foundation models to downstream tasks while preserving their robustness to distribution shifts. Existing methods primarily focus on constraining and projecting current model towards the pre-trained initialization based on the magnitudes between fine-tuned and pre-trained weights, which often require extensive hyper-parameter tuning and can sometimes result in underfitting. In this work, we propose $\\textbf{Di}$rectional $\\textbf{Gra}$dient $\\textbf{P}$rojection (DiGraP), a novel layer-wise trainable method that incorporates directional information from gradients to bridge regularization and multi-objective optimization. Besides demonstrating our method on image classification, as another contribution we generalize this area to the multi-modal evaluation settings for robust fine-tuning. Specifically, we first bridge the uni-modal and multi-modal gap by performing analysis on Image Classification reformulated Visual Question Answering (VQA) benchmarks and further categorize ten out-of-distribution (OOD) VQA datasets by distribution shift types and degree (i.e. near versus far OOD). Experimental results show that DiGraP consistently outperforms existing baselines across Image Classfication and VQA tasks with discriminative and generative backbones, improving both in-distribution (ID) generalization and OOD robustness.", "keywords": [ "Fine-tuning", "transfer learning", "foundation models", "robustness", "visual question answering" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "gnexAe3kjx", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Christian Theobalt", "gender": "Male", "institution": "Max-Planck-Institute for Informatics, Saarland Informatics Campus", "country": "DE", "position": "Director" }, { "name": "Shuteng Wang", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Intern" }, { "name": "Vladislav Golyanik", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute for Informatics", "country": "DE", "position": "Principal Researcher" } ], "n_formula": 19, "n_formula_1": 15, "n_ref_uni": 35, "n_ref": 41, "n_ref_all": 62, "n_fig": 9, "n_tab": 4, "L_tab": 3019, "n_element_tab": 85, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1478, "n_element_tab_1": 56, "formula_len_all": 1101, "formula_len_all_1": 960, "len_all": 139125, "len_all_1": 60504, "len_abs": 1235, "len_title": 69, "len_sents": 28486, "len_sents_1": 27829, "n_sents": 179, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 21, "L_abs": 1239, "title": "Quantum Neural Fields", "abs": "This paper introduces a new type of neural field for visual computing with components compatible with gate-based quantum hardware or simulators thereof. Our Quantum Neural Field Network (QNF-Net) expects as input a query coordinate and, optionally, a latent variable value, and outputs the corresponding field value. QNF-Net includes a new feature map for classical data encoding and a parametrised quantum circuit. The proposed neuro-deterministic data encoding converts, into qubit amplitudes, an energy spectrum of the Gibbs-Boltzmann distribution corresponding to the learned problem energy manifold. We provide a theoretical analysis of the model and its components and perform experiments on a simulator of a gate-based quantum computer with 2D images and 3D shapes (and their collections as learnt priors) and compare results with several classical baselines. QNF-Net consistently outperforms the classical baselines with a comparable number of parameters and achieves faster convergence speed, therefore showing its potential quantum advantages, even for relatively large-scale problems compared to what has been demonstrated in quantum machine learning so far. We will release the source code to facilitate method reproducibility.", "keywords": [ "quantum neural fields", "representation learning", "Neuro-deterministic data encoding", "quantum ansatz" ], "rating_list": [ 6, 1, 8, 5 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 1, 1, 4, 2 ], "contribution_list": [ 3, 1, 4, 3 ], "confidence_list": [ 3, 5, 3, 3 ] }, { "paper_id": "gnWk0ZF22j", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Ali Zare", "gender": "Male", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Constantine Tsibouris", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Researcher" }, { "name": "Li Zhang", "gender": "unknown", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Shih-Fu Chang", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Full Professor" }, { "name": "Xudong Lin", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 23, "n_ref": 68, "n_ref_all": 81, "n_fig": 6, "n_tab": 5, "L_tab": 993, "n_element_tab": 86, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 354, "n_element_tab_1": 25, "formula_len_all": 192, "formula_len_all_1": 192, "len_all": 102052, "len_all_1": 55820, "len_abs": 1766, "len_title": 102, "len_sents": 32166, "len_sents_1": 26270, "n_sents": 223, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1772, "title": "Customized Procedure Planning in Instructional Videos", "abs": "Generating customized procedures for task planning in instructional videos poses a unique challenge for vision-language models. In this paper, we introduce Customized Procedure Planning in Instructional Videos, a novel task that focuses on generating a sequence of detailed action steps for task completion based on user requirements and the task's initial visual state. Existing methods often neglect customization and user directions, limiting their real-world applicability. The absence of instructional video datasets with step-level state and video-specific action plan annotations has hindered progress in this domain. To address these challenges, we introduce the Customized Procedure Planner (CPP) framework, a causal, open-vocabulary model that leverages a LlaVA-based approach to predict procedural plans based on a task's initial visual state and user directions. To overcome the data limitation, we employ a weakly-supervised approach, using the strong vision-language model GEMINI and the large language model (LLM) GPT-4 to create detailed video-specific action plans from the benchmark instructional video datasets (COIN, CrossTask), producing pseudo-labels for training. Discussing the limitations of the existing procedure planning evaluation metrics in an open-vocabulary setting, we propose novel automatic LLM-based metrics with few-shot in-context learning to evaluate the customization and planning capabilities of our model, setting a strong baseline. Additionally, we implement an LLM-based objective function to enhance model training for improved customization. Extensive experiments, including human evaluations, demonstrate the effectiveness of our approach, establishing a strong baseline for future research in customized procedure planning.", "keywords": [ "Customized Procedure Planning", "multi-modal models", "vision-language models" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "gnJwb74rWQ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Guoliang HE", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Zhifeng Jiang", "gender": "unknown", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zhihua Jin", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 33, "n_ref": 74, "n_ref_all": 106, "n_fig": 8, "n_tab": 7, "L_tab": 3783, "n_element_tab": 379, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 202, "n_element_tab_1": 8, "formula_len_all": 532, "formula_len_all_1": 115, "len_all": 163882, "len_all_1": 71679, "len_abs": 782, "len_title": 84, "len_sents": 53149, "len_sents_1": 34615, "n_sents": 435, "n_sents_1": 273, "venue": "ICLR2025", "decision": "withdraw", "L_title": 36, "L_abs": 784, "title": "Safeguarding System Prompts for LLMs", "abs": "Large language models (LLMs) are increasingly utilized in applications where system prompts, which guide model outputs, play a crucial role. These prompts often contain business logic and sensitive information, making their protection essential. However, adversarial and even regular user queries can exploit LLM vulnerabilities to expose these hidden prompts. To address this issue, we present PromptKeeper, a novel defense mechanism for system prompt privacy. By reliably detecting worst-case leakage and regenerating outputs without the system prompt when necessary, PromptKeeper ensures robust protection against prompt extraction attacks via either adversarial or regular queries, while preserving conversational capability and runtime efficiency during benign user interactions.", "keywords": [ "large language models", "system prompts", "privacy" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 3, 5, 4, 5 ] }, { "paper_id": "glgvpS1dD1", "primary_area": "causal reasoning", "n_author": 7, "author_info_list": [ { "name": "Guangyi Chen", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "US", "position": "Postdoc" }, { "name": "Haoxiang Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Haoxuan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Peng Cui", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Xu Chen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zhenlei Wang", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 23, "n_ref_uni": 45, "n_ref": 85, "n_ref_all": 95, "n_fig": 2, "n_tab": 4, "L_tab": 182, "n_element_tab": 17, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 1442, "n_element_tab_1": 27, "formula_len_all": 2411, "formula_len_all_1": 2219, "len_all": 125754, "len_all_1": 64898, "len_abs": 1252, "len_title": 125, "len_sents": 38427, "len_sents_1": 30699, "n_sents": 274, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1266, "title": "Robust Heterogeneous Treatment Effect Estimation under Covariate Perturbation", "abs": "Heterogeneous treatment effect estimation has important applications in fields such as healthcare, economics, and education, attracting increasing attention from both research and the industrial community. However, most existing causal machine learning methods may not perform well in practice due to the lack of robustness of the treatment effect estimation predicted by deep neural networks when an imperceptible perturbation has been added to the covariate. In this paper, we alleviate this problem using the idea of adversarial machine learning. We first show that our loss of interest, the adversarial loss, is partly bounded by the Lipschitz constant of the casual model. Next, we propose a representation learning framework called RHTE which estimates heterogeneous treatment effect under covariate perturbation by controlling the empirical loss, Lipschitz constant, and distance metric simulta neously. Theories are then derived to guarantee the performance and robustness of our estimation. To the best of our knowledge, this is the first work proposing robust representation learning methods under variable perturbation. Extensive experiments on both synthetic examples and standard benchmarks demonstrate the effectiveness and generality of our framework.", "keywords": [ "causal inference", "treatment effect estimation", "robust estimation" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 5, 3, 2 ] }, { "paper_id": "glWOsse3TL", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Yang Luo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" }, { "name": "Yong Liu", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Zangwei Zheng", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Ziheng Qin", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zirui Zhu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 12, "n_ref_uni": 31, "n_ref": 50, "n_ref_all": 77, "n_fig": 10, "n_tab": 2, "L_tab": 464, "n_element_tab": 44, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 157, "n_element_tab_1": 27, "formula_len_all": 3199, "formula_len_all_1": 1288, "len_all": 153077, "len_all_1": 51026, "len_abs": 1676, "len_title": 132, "len_sents": 35172, "len_sents_1": 22713, "n_sents": 279, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1685, "title": "MERIT: Maximum-normalized Element-wise Ratio for Language Model Large-batch Training", "abs": "Large-batch training has become a cornerstone in accelerating the training of deep neural networks, yet it poses challenges in optimization and generalization. Existing optimizers like AdamW present performance degradation during language models' large-batch training, due to the information bottleneck of attention layers caused by the sharp increase of max attention logit. While the LAMB optimizer partially addresses this issue, some attention layers still experience sharply increased maximum attention logits. The reason is that $l_2$-norm-based trust ratios in LAMB are less effective in directly influencing extreme weight values. Furthermore, the weight-wise trust ratio in LAMB is error-prone due to overlooking relationships of weight values within rows or columns. Building on these observations, we propose a novel optimizer, MERIT, which leverages the max norm to calculate the trust ratio to directly constrain the max attention logit. Moreover, we further construct element-wise trust ratios to provide more robust update scaling by focusing on local weight structures. Extensive experiments of large-batch training across various sizes of GPT-2 models demonstrate the superior performance of MERIT. Notably, during the training of GPT-2 Medium, MERIT enables the use of a 6k batch size without any performance degradation compared to the standard batch size (480). This work highlights the importance of considering the max attention logit and finer granularity trust ratio calculation in large-batch training. It successfully improves the training stability and paves the way for larger batch usage, enabling faster development and iteration on large language models.", "keywords": [ "large-batch training", "max attention logit", "language models" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "glUf3YGcJQ", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Gabriel Kreiman", "gender": "Male", "institution": "Harvard Medical School", "country": "US", "position": "Full Professor" }, { "name": "Hanspeter Pfister", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Mengmi Zhang", "gender": "Female", "institution": "Nanyang Technological University, Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Spandan Madan", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "You Li", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "Undergrad student" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 48, "n_ref": 77, "n_ref_all": 119, "n_fig": 23, "n_tab": 5, "L_tab": 1295, "n_element_tab": 116, "n_fig_1": 15, "n_tab_1": 3, "L_tab_1": 511, "n_element_tab_1": 31, "formula_len_all": 149, "formula_len_all_1": 164, "len_all": 176157, "len_all_1": 57463, "len_abs": 1390, "len_title": 92, "len_sents": 46636, "len_sents_1": 25614, "n_sents": 381, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1411, "title": "Improving out-of-distribution generalization by mimicking the human visual diet.", "abs": "Human visual experience is markedly different from the large-scale computer vision datasets consisting of internet images. Babies densely sample a few $3D$ scenes with diverse variations such as object viewpoints or illuminations, while datasets like ImageNet contain one single snapshot from millions of 3D scenes. We investigated how these differences in input data composition (i.e., visual diet) impact the Out-Of-Distribution (OOD) generalization capabilities of a visual system. Training models on a dataset mimicking attributes of the human-like visual diet improved generalization to OOD lighting, material, and viewpoint changes by up to $18\\\\%$. This observation held despite the fact that the models were trained on $1,000$-fold less training data. Furthermore, when trained on purely synthetic data and tested on natural images, incorporating these visual diet attributes in the training dataset improved OOD generalization by $17\\\\%$. These experiments are enabled by our newly proposed benchmark---the Human Visual Diet (HVD) dataset, and a new model (Human Diet Network) designed to leverage the attributes of a human-like visual diet. These findings highlight a critical problem in modern day Artificial Intelligence---building better datasets requires thinking beyond dataset size and rather focus on improving data composition. All data and source code will be made available upon publication.", "keywords": [ "out-of-distribution generalization", "human vision", "generalization", "scene context" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "gkUyYcY1W9", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Chengruidong Zhang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Dongsheng Li", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Huiqiang Jiang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "RSDE" }, { "name": "Jianfeng Gao", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Lili Qiu", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Principal Researcher" }, { "name": "Qianhui Wu", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Surin Ahn", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Xufang Luo", "gender": "Female", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" }, { "name": "YUCHENG LI", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "PhD student" }, { "name": "Yuqing Yang", "gender": "unknown", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Amir H Abdi", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Principal Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 57, "n_ref": 138, "n_ref_all": 177, "n_fig": 9, "n_tab": 11, "L_tab": 9950, "n_element_tab": 1283, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 7398, "n_element_tab_1": 749, "formula_len_all": 40, "formula_len_all_1": 0, "len_all": 236386, "len_all_1": 73858, "len_abs": 1797, "len_title": 120, "len_sents": 54855, "len_sents_1": 27396, "n_sents": 373, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1800, "title": "SharedContextBench: How Lossy are Long-context Methods in KV Cache Reuse", "abs": "Long-context Large Language Models (LLMs) have unlocked numerous possibilities for downstream applications, many of which involve multiple requests sharing the same input context. Recent inference frameworks like vLLM and SGLang, as well as LLMs providers such as OpenAI, Gemini and Claude, have employed prefix caching techniques to accelerate multi-requests with shared context. However, existing long-context methods are primarily evaluated on single query testing, failing to demonstrate their true capability in real-world applications that often require KV cache reuse for follow-up queries. To address this gap, we introduce SharedContextBench, a comprehensive long-context benchmark to reveal how lossy are long-context methods in KV cache reuse scenarios. Specifically, it encompasses 12 tasks with two shared context modes, covering four categories of long-context abilities: string retrieval, semantic retrieval, global information processing, and multi-task capabilities. Using our benchmark, we evaluated five categories of long-context solutions, including Gated Linear RNNs (Codestal-Mamba), MambaAttention hybrids (Jamba-1.5-Mini), and efficient methods like sparse attention, KV cache compression, and prompt compression, on six transformer-based longcontext LLMs: Llama-3.1-8B/70B, Qwen2.5-72B/32B, Llama-3-8B-262K, and GLM-4-9B. Our findings show that sub-O(n) memory methods often struggle to maintain accuracy in multi-turn scenarios, while sparse encoding methods with O(n) memory and sub-O(n 2 ) computation in prefilling generally perform well. Additionally, dynamic sparse patterns in prefilling often produce more expressive memory (KV cache) compared to static methods, and layer-level sparsity in hybrid architectures reduces memory usage while yielding promising results.", "keywords": [ "Long-context", "Benchmark", "KV Cache Reuse", "Long-context Methods", "Shared Context" ], "rating_list": [ 8, 8, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 4, 2, 4, 3, 4 ] }, { "paper_id": "gkOtsxD6fr", "primary_area": "generative models", "n_author": 12, "author_info_list": [ { "name": "Bohan Zeng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Fu-Yun Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jiaming Liu", "gender": "Male", "institution": "Tiamat AI", "country": "CN", "position": "Researcher" }, { "name": "Juanxi Tian", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Intern" }, { "name": "Kaixin Zhu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Intern" }, { "name": "Ling Yang", "gender": "Male", "institution": "DeepSeek AI", "country": "CN", "position": "Intern" }, { "name": "Minkai Xu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Stefano Ermon", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Zixiang Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Intern" }, { "name": "yongzhen.gyz", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Li Siyu", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Zhang wen tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 44, "n_ref": 105, "n_ref_all": 121, "n_fig": 10, "n_tab": 4, "L_tab": 1230, "n_element_tab": 75, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 711, "n_element_tab_1": 46, "formula_len_all": 294, "formula_len_all_1": 382, "len_all": 150405, "len_all_1": 53520, "len_abs": 1152, "len_title": 132, "len_sents": 35651, "len_sents_1": 25467, "n_sents": 278, "n_sents_1": 165, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1164, "title": "Trans4D: Realistic Geometry-Aware Transition for Compositional Text-to-4D Synthesis", "abs": "Recent advances in diffusion models have demonstrated exceptional capabilities in image and video generation, further improving the effectiveness of 4D synthesis. Existing 4D generation methods can generate high-quality 4D objects or scenes based on user-friendly conditions, benefiting the gaming and video industries. However, these methods struggle to synthesize significant object deformation of complex 4D transitions and interactions within scenes. To address this challenge, we propose **Trans4D**, a novel text-to-4D synthesis framework that enables realistic complex scene transitions. Specifically, we first use multi-modal large language models (MLLMs) to produce a physic-aware scene description for 4D scene initialization and effective transition timing planning. Then we propose a geometry-aware 4D transition network to realize a complex scene-level 4D transition based on the plan, which involves expressive geometrical object deformation. Extensive experiments demonstrate that **Trans4D** consistently outperforms existing state-of-the-art methods in generating 4D scenes with accurate and high-quality transitions, validating its effectiveness.", "keywords": [ "Text-to-4D generation", "Scene transition" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 1, 1, 3, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "gkDRrvqeWF", "primary_area": "applications to robotics, autonomy, planning", "n_author": 9, "author_info_list": [ { "name": "Erdem Biyik", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Jan Kautz", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "VP Research" }, { "name": "Sifei Liu", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Xiaolong Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Xueyan Zou", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Postdoc" }, { "name": "Yandong Ji", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Zhaojing Yang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Researcher" }, { "name": "An-Chieh Cheng", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Hongxu Yin", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Senior Research Scientist" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 71, "n_ref": 100, "n_ref_all": 122, "n_fig": 7, "n_tab": 12, "L_tab": 4460, "n_element_tab": 528, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2517, "n_element_tab_1": 307, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 193089, "len_all_1": 61161, "len_abs": 348, "len_title": 112, "len_sents": 39984, "len_sents_1": 26809, "n_sents": 308, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 967, "title": "NaVILA: Legged Robot Vision-Language-Action Model for Navigation", "abs": "This paper proposes to solve the problem of Vision-and-Language Navigation with legged robots, which not only provides a flexible way for humans to command but also allows the robot to navigate through more challenging and cluttered scenes. However, it is non-trivial to translate human language instructions all the way to low-level leg joint actions. We propose NaVILA, a 2-level framework that unifies a Vision-Language-Action model (VLA) with locomotion skills. Instead of directly predicting low-level actions from VLA, NaVILA first generates mid-level actions with spatial information in the form of language, (e.g., \"moving forward 75cm\"), which serves as an input for a visual locomotion RL policy for execution. NaVILA substantially improves previous approaches on existing benchmarks. The same advantages are demonstrated in our newly developed benchmarks with IsaacLab, featuring more realistic scenes, low-level controls, and real-world robot experiments.", "keywords": [ "Vision Language Action", "Legged Robots", "Vision Language Navigation" ], "rating_list": [ 5, 8, 6, 3 ], "soundness_list": [ 4, 4, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "gjwhDHeAsz", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Mingyuan Zhou", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Shujian Zhang", "gender": "unknown", "institution": "University of Texas, Austin", "country": "", "position": "PhD student" }, { "name": "Tianqi Chen", "gender": "unknown", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 16, "n_ref_uni": 86, "n_ref": 163, "n_ref_all": 187, "n_fig": 9, "n_tab": 7, "L_tab": 1615, "n_element_tab": 191, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1321, "n_element_tab_1": 143, "formula_len_all": 1554, "formula_len_all_1": 1477, "len_all": 212104, "len_all_1": 66846, "len_abs": 1696, "len_title": 148, "len_sents": 58309, "len_sents_1": 31781, "n_sents": 372, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 99, "L_abs": 1501, "title": "Score Forgetting Distillation: A Swift, Data-Free Method for Machine Unlearning in Diffusion Models", "abs": "The machine learning community is increasingly recognizing the importance of fostering trust and safety in modern generative AI (GenAI) models. We posit machine unlearning (MU) as a crucial foundation for developing safe, secure, and trustworthy GenAI models. Traditional MU methods often rely on stringent assumptions and require access to real data. This paper introduces Score Forgetting Distillation (SFD), an innovative MU approach that promotes the forgetting of undesirable information in diffusion models by aligning the conditional scores of \"unsafe\" classes or concepts with those of \"safe\" ones. To eliminate the need for real data, our SFD framework incorporates a score-based MU loss into the score distillation objective of a pretrained diffusion model. This serves as a regularization term that preserves desired generation capabilities while enabling the production of synthetic data through a one-step generator. Our experiments on pretrained label-conditional and text-to-image diffusion models demonstrate that our method effectively accelerates the forgetting of target classes or concepts during generation, while preserving the quality of other classes or concepts. This unlearned and distilled diffusion not only pioneers a novel concept in MU but also accelerates the generation speed of diffusion models. Our experiments and studies on a range of diffusion models and datasets confirm that our approach is generalizable, effective, and advantageous for MU in diffusion models.", "keywords": [ "machine unlearning", "diffusion models", "generative modeling", "trustworthy machine learning" ], "rating_list": [ 5, 5, 8, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "gjRhw5S3A4", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Li Ju", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" }, { "name": "Xingyi Yang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Qiiiii Li", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Xinchao Wang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 42, "n_ref": 81, "n_ref_all": 113, "n_fig": 5, "n_tab": 19, "L_tab": 16109, "n_element_tab": 850, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 3871, "n_element_tab_1": 217, "formula_len_all": 340, "formula_len_all_1": 340, "len_all": 204289, "len_all_1": 66228, "len_abs": 1279, "len_title": 104, "len_sents": 57552, "len_sents_1": 29619, "n_sents": 420, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1292, "title": "GraphBridge: Towards Arbitrary Transfer Learning in GNNs", "abs": "Graph neural networks (GNNs) are conventionally trained on a per-domain, per-task basis. It creates a significant barrier in transferring the acquired knowledge to different, heterogeneous data setups. This paper introduces **GraphBridge**, a novel framework to enable knowledge transfer across disparate tasks and domains in GNNs, circumventing the need for modifications to task configurations or graph structures. Specifically, GraphBridge allows for the augmentation of any pre-trained GNN with prediction heads and a bridging network that connects the input to the output layer. This architecture not only preserves the intrinsic knowledge of the original model but also supports outputs of arbitrary dimensions. To mitigate the negative transfer problem, GraphBridg merges the source model with a concurrently trained model, thereby reducing the source bias when applied to the target domain. Our method is thoroughly evaluated across diverse transfer learning scenarios, including Graph2Graph, Node2Node, Graph2Node, and graph2point-cloud. Empirical validation, conducted over 16 datasets representative of these scenarios, confirms the framework's capacity for task- and domain-agnostic transfer learning within graph-like data, marking a significant advancement in the field of GNNs.", "keywords": [ "Graph Neural Networks", "Transfer Learning", "Efficient Tuning Methods", "Universial Model" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 3, 4, 2, 4 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "gjFgBfbP2C", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Jian Jin", "gender": "Male", "institution": "China Academy of information and communications technology", "country": "CN", "position": "Full Professor" }, { "name": "Jin Song", "gender": "unknown", "institution": "Nanjing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Yuan Yao", "gender": "Male", "institution": "China Academy of Information and Communications Technology", "country": "CN", "position": "Researcher" }, { "name": "Zhenzhen Jiao", "gender": "Male", "institution": "Beijing Teleinfo Technology, CAICT", "country": "", "position": "Principal Researcher" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 40, "n_ref": 100, "n_ref_all": 122, "n_fig": 5, "n_tab": 13, "L_tab": 4634, "n_element_tab": 308, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 2557, "n_element_tab_1": 226, "formula_len_all": 604, "formula_len_all_1": 642, "len_all": 162900, "len_all_1": 74359, "len_abs": 941, "len_title": 107, "len_sents": 52604, "len_sents_1": 34147, "n_sents": 386, "n_sents_1": 230, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 950, "title": "NeuralMark: Advancing White-Box Neural Network Watermarking", "abs": "As valuable digital assets, deep neural networks require ownership protection, making neural network watermarking (NNW) a promising solution. In this paper, we propose a *NeuralMark* method to advance white-box NNW, which can be seamlessly integrated into various network architectures. NeuralMark first establishes a hash mapping between the secret key and the watermark, enabling resistance to forging attacks. The watermark then functions as a filter to select model parameters for embedding, providing resilience against overwriting attacks. Furthermore, NeuralMark utilizes average pooling to defend against fine-tuning and pruning attacks. Theoretically, we analyze its security boundary. Empirically, we verify its effectiveness across 14 distinct Convolutional and Transformer architectures, covering five image classification tasks and one text generation task. The source codes are available at https://anonymous.4open.science/r/NeuralMark.", "keywords": [ "Neural network; White-box watermarking; Hash mapping; Watermark filtering; Average pooling" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "gjC91PwBZy", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Carl Yang", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Assistant Professor" }, { "name": "Juntong Ni", "gender": "Male", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Noveen Sachdeva", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Shengbo Gong", "gender": "Male", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Wei Jin", "gender": "unknown", "institution": "Emory University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 57, "n_ref": 111, "n_ref_all": 144, "n_fig": 6, "n_tab": 14, "L_tab": 7364, "n_element_tab": 930, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2900, "n_element_tab_1": 247, "formula_len_all": 121, "formula_len_all_1": 91, "len_all": 230172, "len_all_1": 76708, "len_abs": 1431, "len_title": 92, "len_sents": 70361, "len_sents_1": 35260, "n_sents": 503, "n_sents_1": 272, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1351, "title": "GC4NC: A Benchmark Framework for Graph Condensation on Node Classification with New Insights", "abs": "Graph condensation (GC) is an emerging technique designed to learn a significantly smaller graph that retains the essential information of the original graph. This condensed graph has shown promise in accelerating graph neural networks while preserving performance comparable to those achieved with the original, larger graphs. Additionally, this technique facilitates downstream applications like neural architecture search and deepens our understanding of redundancies in large graphs. Despite the rapid development of GC methods, particularly for node classification, a unified evaluation framework is still lacking to systematically compare different GC methods or clarify key design choices for improving their effectiveness. To bridge these gaps, we introduce **GC4NC**, a comprehensive framework for evaluating diverse GC methods on node classification across multiple dimensions including performance, efficiency, privacy preservation, denoising ability, NAS effectiveness, and transferability. Our systematic evaluation offers novel insights into how condensed graphs behave and the critical design choices that drive their success. These findings pave the way for future advancements in GC methods, enhancing both performance and expanding their real-world applications. The code is available at https://anonymous.4open.science/r/GC4NC-1620.", "keywords": [ "Graph condensation", "Dataset distillation", "Dataset condensation", "Graph neural network" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 3, 3, 2 ] }, { "paper_id": "gjC3QvVh1U", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Claudius Gros", "gender": "Male", "institution": "Goethe University Frankfurt, Germany", "country": "", "position": "Full Professor" }, { "name": "Oren Neumann", "gender": "Male", "institution": "adago GmbH", "country": "DE", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 6, "n_ref_uni": 48, "n_ref": 85, "n_ref_all": 148, "n_fig": 28, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 20, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 465, "formula_len_all_1": 177, "len_all": 160505, "len_all_1": 56187, "len_abs": 1225, "len_title": 125, "len_sents": 54150, "len_sents_1": 26514, "n_sents": 416, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1232, "title": "AlphaZero Neural Scaling and Zipf's Law: a Tale of Board Games and Power Laws", "abs": "Neural scaling laws are observed in a range of domains, to date with no clear understanding of why they occur. Recent theories suggest that loss power laws arise from Zipf's law, a power law observed in domains like natural language. One theory suggests that language scaling laws emerge when Zipf-distributed task quanta are learned in descending order of frequency. In this paper we examine power-law scaling in AlphaZero, a reinforcement learning algorithm, using a theory of language-model scaling. We find that game states in training and inference data scale with Zipf's law, which is known to arise from the tree structure of the environment, and examine the correlation between scaling-law and Zipf's-law exponents. In agreement with quanta scaling theory, we find that agents optimize state loss in descending order of frequency, even though this order scales inversely with modelling complexity. We also find that inverse scaling, the failure of models to improve with size, is correlated with unusual Zipf curves where end-game states are among the most frequent states. We show evidence that larger models shift their focus to these less-important states, sacrificing their understanding of important early-game states.", "keywords": [ "Scaling Laws", "Reinforcement Learning", "Zipf's Law" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "giU5WVNy7K", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Guizhen Chen", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Kenji Kawaguchi", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Presidential Young Professor" }, { "name": "Lidong Bing", "gender": "unknown", "institution": "Alibaba Group", "country": "SG", "position": "Scientist" }, { "name": "Wenxuan Zhang", "gender": "unknown", "institution": "Alibaba Group", "country": "SG", "position": "Researcher" }, { "name": "Yiran Zhao", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 11, "n_ref_uni": 54, "n_ref": 109, "n_ref_all": 130, "n_fig": 4, "n_tab": 7, "L_tab": 3606, "n_element_tab": 465, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2886, "n_element_tab_1": 344, "formula_len_all": 725, "formula_len_all_1": 892, "len_all": 152796, "len_all_1": 61776, "len_abs": 1403, "len_title": 125, "len_sents": 33311, "len_sents_1": 25067, "n_sents": 227, "n_sents_1": 166, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1153, "title": "From General to Expert: Custom Pruning LLMs Across Language, Domain, and Task", "abs": "Large Language Models (LLMs) have transformed natural language processing, yet their substantial model sizes often demand significant computational resources. To conserve computing resources and increase inference speed, it is crucial to prune redundant parameters, especially for general users who often need expert models tailored to specific downstream scenarios. However, current pruning methods primarily focus on maintaining models' general capabilities, either requiring extensive post-training or performing poorly due to coarse-grained pruning. In this work, we design a $\\underline{Cus}$tom $\\underline{Prun}$ing method ($\\texttt{Cus-Prun}$) to prune a large general model into a smaller expert model for specific scenarios. $\\texttt{Cus-Prun}$ positions an expert model along the \"language\", \"domain\" and \"task\" dimensions. By identifying and pruning irrelevant neurons, it creates expert models without any post-training. \nOur experiments demonstrate that $\\texttt{Cus-Prun}$ consistently outperforms other methods, achieving minimal loss in both expert and general capabilities across various models from different model families and sizes.", "keywords": [ "Large Language Models", "Pruning", "Expert Model" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 5, 5, 3, 3 ] }, { "paper_id": "ghH6YYDs15", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Charles O'Neill", "gender": "Male", "institution": "Australian National University", "country": "", "position": "Undergrad student" }, { "name": "David A. Klindt", "gender": "unknown", "institution": "Cold Spring Harbor Laboratory", "country": "", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 11, "n_ref_uni": 34, "n_ref": 68, "n_ref_all": 95, "n_fig": 17, "n_tab": 1, "L_tab": 1043, "n_element_tab": 34, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 966, "formula_len_all_1": 500, "len_all": 159343, "len_all_1": 61913, "len_abs": 1573, "len_title": 126, "len_sents": 57511, "len_sents_1": 28776, "n_sents": 449, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1580, "title": "Compute Optimal Inference and Provable Amortisation Gap in Sparse Autoencoders", "abs": "A recent line of work has shown promise in using sparse autoencoders (SAEs) to uncover interpretable features in neural network representations. However, the simple linear-nonlinear encoding mechanism in SAEs limits their ability to perform accurate sparse inference. In this paper, we investigate sparse inference and learning in SAEs through the lens of sparse coding. Specifically, we show that SAEs perform amortised sparse inference with a computationally restricted encoder and, using compressed sensing theory, we prove that this mapping is inherently insufficient for accurate sparse inference, even in solvable cases. Building on this theory, we empirically explore conditions where more sophisticated sparse inference methods outperform traditional SAE encoders. Our key contribution is the decoupling of the encoding and decoding processes, which allows for a comparison of various sparse encoding strategies. We evaluate these strategies on two dimensions: alignment with true underlying sparse features and correct inference of sparse codes, while also accounting for computational costs during training and inference. Our results reveal that substantial performance gains can be achieved with minimal increases in compute cost. We demonstrate that this generalises to SAEs applied to large language models (LLMs), where advanced encoders achieve similar interpretability. This work opens new avenues for understanding neural network representations and offers important implications for improving the tools we use to analyse the activations of large language models.", "keywords": [ "interpretability", "sparse coding", "neuroscience", "language models", "superposition" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 4, 2, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "gh563RwulS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Feiyu Xiong", "gender": "Male", "institution": "Institute for Advanced Algorithms Research, Shanghai", "country": "", "position": "Principal Researcher" }, { "name": "Jihao Zhao", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Pengnian Qi", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Simin Niu", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "PhD student" }, { "name": "Zhiyu li", "gender": "Male", "institution": "Institute for Advanced Algorithms Research, Shanghai", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhiyuan Ji", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "\u6ce2 \u5510", "gender": "Male", "institution": "Institute for Advanced Algorithms Research, Shanghai", "country": "", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 8, "n_ref_uni": 47, "n_ref": 60, "n_ref_all": 91, "n_fig": 6, "n_tab": 6, "L_tab": 4595, "n_element_tab": 605, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 3290, "n_element_tab_1": 412, "formula_len_all": 1459, "formula_len_all_1": 567, "len_all": 148692, "len_all_1": 64193, "len_abs": 1356, "len_title": 123, "len_sents": 40963, "len_sents_1": 27125, "n_sents": 297, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1361, "title": "Meta-Chunking: Learning Efficient Text Segmentation via Logical Perception", "abs": "Retrieval-Augmented Generation (RAG), while serving as a viable complement to large language models (LLMs), often overlooks the crucial aspect of text chunking within its pipeline, which impacts the quality of knowledge-intensive tasks. This paper introduces the concept of Meta-Chunking, which refers to a granularity between sentences and paragraphs, consisting of a collection of sentences within a paragraph that have deep linguistic logical connections. To implement Meta-Chunking, we designed Perplexity (PPL) Chunking, which balances performance and speed, and precisely identifies the boundaries of text chunks by analyzing the characteristics of context perplexity distribution. Additionally, considering the inherent complexity of different texts, we propose a strategy that combines PPL Chunking with dynamic merging to achieve a balance between fine-grained and coarse-grained text chunking. Experiments conducted on eleven datasets demonstrate that Meta-Chunking can more efficiently improve the performance of single-hop and multi-hop question answering based on RAG. For instance, on the 2WikiMultihopQA dataset, it outperforms similarity chunking by 1.32 while only consuming 45.8\\% of the time. Furthermore, through the analysis of models of various scales and types, we observed that PPL Chunking exhibits notable flexibility and adaptability.", "keywords": [ "Text chunking", "perplexity", "margin sampling", "large language models" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 1, 2, 1 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 2 ] }, { "paper_id": "gg6dPtdC1C", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Abdelrahman Eldesokey", "gender": "unknown", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" }, { "name": "Peter Wonka", "gender": "Male", "institution": "Snap Inc.", "country": "SA", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 38, "n_ref": 67, "n_ref_all": 98, "n_fig": 13, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 375, "n_element_tab_1": 92, "formula_len_all": 512, "formula_len_all_1": 603, "len_all": 129691, "len_all_1": 58844, "len_abs": 382, "len_title": 130, "len_sents": 37215, "len_sents_1": 26455, "n_sents": 269, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1577, "title": "Build-A-Scene: Interactive 3D Layout Control for Diffusion-Based Image Generation", "abs": "We propose a diffusion-based approach for Text-to-Image (T2I) generation with interactive 3D layout control.\nLayout control has been widely studied to alleviate the shortcomings of T2I diffusion models in understanding objects' placement and relationships from text descriptions.\nNevertheless, existing approaches for layout control are limited to 2D layouts, require the user to provide a static layout beforehand, and fail to preserve generated images under layout changes.\nThis makes these approaches unsuitable for applications that require 3D object-wise control and iterative refinements, e.g, interior design and complex scene generation. \nTo this end, we leverage the recent advancements in depth-conditioned T2I models and propose a novel approach for interactive 3D layout control.\nWe replace the traditional 2D boxes used in layout control with 3D boxes.\nFurthermore, we revamp the T2I task as a multi-stage generation process, where at each stage, the user can insert, change, and move an object in 3D while preserving objects from earlier stages.\nWe achieve this through a novel Dynamic Self-Attention (DSA) module and a consistent 3D object translation strategy.\nTo evaluate our approach, we establish a benchmark and an evaluation protocol for interactive 3D layout control.\nExperiments show that our approach can generate complicated scenes based on 3D layouts, outperforming the standard depth-conditioned T2I methods by two-folds on object generation success rate.\nMoreover, it outperforms all methods in comparison on preserving objects under layout changes.", "keywords": [ "Diffusion Models", "Text-to-Image", "Layout Control" ], "rating_list": [ 6, 8, 3, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "gfI9v7AbFg", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Guanzhi Wang", "gender": "Male", "institution": "California Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Jonathan Light", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "PhD student" }, { "name": "Min Cai", "gender": "unknown", "institution": "Shenzhen University", "country": "CN", "position": "MS student" }, { "name": "Wei Cheng", "gender": "Male", "institution": "NEC-Labs", "country": "US", "position": "Researcher" }, { "name": "Weiqin Chen", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "PhD student" }, { "name": "Xiusi Chen", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Postdoc" }, { "name": "Yisong Yue", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Ziniu Hu", "gender": "Male", "institution": "xAI", "country": "US", "position": "Member of Technical Staff" } ], "n_formula": 11, "n_formula_1": 4, "n_ref_uni": 37, "n_ref": 84, "n_ref_all": 129, "n_fig": 19, "n_tab": 9, "L_tab": 13074, "n_element_tab": 393, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 906, "n_element_tab_1": 85, "formula_len_all": 1193, "formula_len_all_1": 285, "len_all": 334354, "len_all_1": 61530, "len_abs": 1258, "len_title": 76, "len_sents": 112030, "len_sents_1": 28650, "n_sents": 914, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1207, "title": "Strategist: Self-improvement of LLM Decision Making via Bi-Level Tree Search", "abs": "Traditional reinforcement learning (RL) typically requires vast amounts of training\ndata to develop effective policies. In contrast, large language models (LLMs)\nexhibit strong generalization and zero-shot capabilities, but struggle with plan-\nning and understanding complex action policies. In this work, we introduce\nSTRATEGIST, a novel approach that integrates the strengths of both methods. Our\napproach leverages LLMs to learn high-level strategic abstractions, which are\nthen refined and executed by a low-level mechanism, such as Monte Carlo Tree\nSearch (MCTS). STRATEGIST is a generalizable framework that can be trained\nthrough population-based self-play simulations and self-improvement, without the\nneed for prior training data. We demonstrate the effectiveness of STRATEGIST in\nlearning optimal policies for competitive, multi-turn games with partial informa-\ntion, including Game of Pure Strategy (GOPS) and multi-agent, hidden-identity\ndiscussion games like The Resistance: Avalon. Our results show that agents trained\nwith STRATEGIST outperform those trained with traditional RL methods, other\nLLM-based skill acquisition techniques, and pre-existing LLM agents across both\ngame environments.", "keywords": [ "LLMs", "games", "search", "self-improvement", "self-play", "RL", "agent", "multi-agent" ], "rating_list": [ 6, 6, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 2, 4, 4, 3, 3 ] }, { "paper_id": "gfDbD1MRYk", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Adrian Weller", "gender": "Male", "institution": "Alan Turing Institute", "country": "", "position": "Programme Director for AI" }, { "name": "Ankur Mallick", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Menglin Xia", "gender": "unknown", "institution": "Microsoft", "country": "GB", "position": "Researcher" }, { "name": "Metod Jazbec", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Shoaib Ahmed Siddiqui", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Srikant Bharadwaj", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Wenbo Gong", "gender": "Male", "institution": "Microsoft", "country": "GB", "position": "Senior Researcher" }, { "name": "Yanzhi Chen", "gender": "unknown", "institution": "Microsoft Research ", "country": "GB", "position": "Researcher" }, { "name": "Victor Ruehle", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Principal Researcher" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 92, "n_ref_all": 117, "n_fig": 12, "n_tab": 4, "L_tab": 464, "n_element_tab": 62, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 825, "n_element_tab_1": 78, "formula_len_all": 567, "formula_len_all_1": 625, "len_all": 130572, "len_all_1": 67300, "len_abs": 893, "len_title": 104, "len_sents": 40413, "len_sents_1": 32005, "n_sents": 315, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 896, "title": "Semi-autoregressive Decoding for Efficient LLM Inference", "abs": "Inference in large language models (LLMs) is often slow due to their autoregressive nature. \nIn this work, we formulate a semi-autoregressive decoding paradigm for LLMs that delegates part of the expensive computation from the original large model to a smaller, more efficient autoregressive model. The core of our design lies in the separate modeling of token dependencies, where the large model handles long-term dependencies on distant tokens, while the smaller model addresses short-term dependencies on recent tokens. When employed as a draft model in speculative decoding, our method allows for substantial reuse of computation in the LLM without missing any token dependencies, thereby striking a good balance between draft quality and drafting speed. Experiments on text summarization, medical QA, code generation, and mathematical reasoning tasks demonstrates the efficacy of our method.", "keywords": [ "efficient inference", "dependence modeling", "non-autoregressive models", "LLMs", "speculative decoding" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "ge5PasXuJ6", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Ashutosh Kumar Nirala", "gender": "Male", "institution": "Iowa State University", "country": "", "position": "PhD student" }, { "name": "Jin Tian", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" }, { "name": "Modeste Atsague", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "PhD student" }, { "name": "Olukorede Fakorede", "gender": "unknown", "institution": "Iowa State University", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 42, "n_ref": 77, "n_ref_all": 98, "n_fig": 3, "n_tab": 14, "L_tab": 3645, "n_element_tab": 371, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1931, "n_element_tab_1": 187, "formula_len_all": 300, "formula_len_all_1": 300, "len_all": 165809, "len_all_1": 70096, "len_abs": 1290, "len_title": 139, "len_sents": 49965, "len_sents_1": 29864, "n_sents": 455, "n_sents_1": 258, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1305, "title": "Error Correcting by Agreement Checking for Adversarial Robustness against Black-box Attacks", "abs": "Drawing inspiration from the vulnerability of the initial feed-forward phase of biological perception in humans and primates to adversarial attacks, we propose a novel defense strategy named Error Correcting by Agreement Checking (ECAC). This strategy is designed to mitigate realistic \\emph{black-box} threats where attackers don't have full access to the model. We exploit the fact that natural and adversarially trained models rely on distinct feature sets for classification. \nNotably, naturally trained models retain commendable accuracy against adversarial examples generated using adversarially trained models. Leveraging this disparity, ECAC moves the input toward the prediction of the naturally trained model unless it leads to disagreement in prediction between the two models, before making the prediction. \nThis simple error correction mechanism is highly effective against leading SQA (Score-based Query Attacks) black-box attacks as well as decision-based and transfer-based black-box attacks. \nWe also verify that, unlike other black-box defense, ECAC maintains significant robustness even when adversary has full access to the model. We demonstrate its effectiveness through comprehensive experiments across various datasets (CIFAR and ImageNet) and architectures (ResNet as well as ViT).", "keywords": [ "adversarial defense; AT; black-box; SQA" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "gdzpnRBP4F", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Carel van Niekerk", "gender": "Male", "institution": "Heinrich-Heine Universit\u00e4t D\u00fcsseldorf", "country": "DE", "position": "Postdoc" }, { "name": "Hsien-chin Lin", "gender": "Male", "institution": "Heinrich Heine University D\u00fcsseldorf", "country": "", "position": "Postdoc" }, { "name": "Matthew Sutton", "gender": "Male", "institution": "Heinrich-Heine Universit\u00e4t D\u00fcsseldorf", "country": "DE", "position": "MS student" }, { "name": "Milica Gasic", "gender": "unknown", "institution": "Heinrich Heine University Duesseldorf", "country": "", "position": "Full Professor" }, { "name": "Renato Vukovic", "gender": "Male", "institution": "Heinrich Heine University D\u00fcsseldorf", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 28, "n_ref": 49, "n_ref_all": 64, "n_fig": 5, "n_tab": 5, "L_tab": 659, "n_element_tab": 101, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 578, "n_element_tab_1": 94, "formula_len_all": 229, "formula_len_all_1": 229, "len_all": 154650, "len_all_1": 57319, "len_abs": 1128, "len_title": 126, "len_sents": 30906, "len_sents_1": 26835, "n_sents": 220, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1134, "title": "RLSF: Reinforcement Learning from Self-feedback for improved logical reasoning", "abs": "Large Language Models (LLMs) have demonstrated impressive capabilities in generating coherent and contextually relevant text. These models\narguably lack the ability to logically reason, an essential skill required to solving mathematical problems and programming tasks.\nWhile step-by-step prompting approaches show some promise, they often depend on finding a suitable prompt tailored to the specific model and task. In this work, we propose a simple, yet an effective approach to enhance reasoning capabilities by leveraging reinforcement learning (RL) and the confidence scores of a well-calibrated LLM. It involves optimising an implicit reward derived from the model's confidence levels in the answer to the reasoning task at hand.\nWe generate preference data and fine-tune the LLM in a similar spirit to reinforcement learning from human feedback (RLHF), but without needing any human provided labels or preferences.\nOur results show that resulting reasoning abilities of an LLM improve and are transferable to other reasoning tasks. This warrants further investigation of RL as a facilitator for solving complex language tasks.", "keywords": [ "reinforcement learning", "large language models", "reasoning", "uncertainty" ], "rating_list": [ 6, 6, 3, 3 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "gdHtZlaaSo", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Adam Dziedzic", "gender": "Male", "institution": "CISPA, saarland university, saarland informatics campus", "country": "DE", "position": "Assistant Professor" }, { "name": "Bartosz Cywi\u0144ski", "gender": "Male", "institution": "Warsaw University of Technology", "country": "PL", "position": "PhD student" }, { "name": "Franziska Boenisch", "gender": "Female", "institution": "CISPA", "country": "DE", "position": "Assistant Professor" }, { "name": "Kamil Deja", "gender": "unknown", "institution": "IDEAS NCBR", "country": "PL", "position": "Postdoc" }, { "name": "\u0141ukasz Staniszewski", "gender": "Male", "institution": "Warsaw University of Technology", "country": "PL", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 36, "n_ref": 81, "n_ref_all": 119, "n_fig": 15, "n_tab": 8, "L_tab": 3957, "n_element_tab": 498, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1103, "n_element_tab_1": 174, "formula_len_all": 0, "formula_len_all_1": 38, "len_all": 169056, "len_all_1": 68287, "len_abs": 3894, "len_title": 121, "len_sents": 55298, "len_sents_1": 32958, "n_sents": 359, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1335, "title": "Precise Parameter Localization for Textual Generation in Diffusion Models", "abs": "Novel diffusion models (DMs) can synthesize photo-realistic images with integrated high-quality text. Surprisingly, we demonstrate through attention activation patching that only less than 1% of DMs' parameters contained in attention layers influence the generation of textual content within the images. Building on this observation, by precisely targeting cross and joint attention layers of DMs, we improve the efficiency and performance of textual generation. We introduce several applications that benefit from localizing the layers responsible for textual content generation. We first show that a LoRA-based fine-tuning solely of the localized layers enhances, even more, the general text-generation capabilities of large DMs while preserving the quality and diversity of the DMs' generations. Then, we demonstrate how we can use the localized layers to edit textual content in generated images. Finally, we extend this idea to the practical use case of preventing the generation of toxic text in a cost-free manner. In contrast to prior work, our localization approach is broadly applicable across various diffusion model architectures, including U-Net (e.g., LDM and SDXL) and transformer-based (e.g., DeepFloyd IF and Stable Diffusion 3), utilizing diverse text encoders (e.g., from CLIP and the large language models like T5).", "keywords": [ "diffusion models", "text edition", "LoRA", "localization", "SD-XL", "SD3", "DeepFloyd IF" ], "rating_list": [ 6, 6, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "gcouwCx7dG", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Badong Chen", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Gang Pan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" }, { "name": "Jiangrong Shen", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" }, { "name": "Qi Xu", "gender": "Male", "institution": "School of Computer Science and Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 30, "n_ref": 46, "n_ref_all": 55, "n_fig": 5, "n_tab": 1, "L_tab": 790, "n_element_tab": 126, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 503, "n_element_tab_1": 41, "formula_len_all": 488, "formula_len_all_1": 485, "len_all": 107702, "len_all_1": 66603, "len_abs": 1800, "len_title": 154, "len_sents": 36804, "len_sents_1": 34237, "n_sents": 246, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 106, "L_abs": 1809, "title": "Improving the Sparse Structure Learning of Spiking Neural Networks from the View of Compression Efficiency", "abs": "The human brain utilizes spikes for information transmission and dynamically reorganizes its network structure to boost energy efficiency and cognitive capabilities throughout its lifespan. Drawing inspiration from this spike-based computation, Spiking Neural Networks (SNNs) have been developed to construct event-driven models that emulate this efficiency. Despite these advances, deep SNNs continue to suffer from over-parameterization during training and inference, a stark contrast to the brain\u2019s ability to self-organize. Furthermore, existing sparse SNNs are challenged by maintaining optimal pruning levels due to a static pruning ratio, resulting in either under or over-pruning.\nIn this paper, we propose a novel two-stage dynamic structure learning approach for deep SNNs, aimed at maintaining effective sparse training from scratch while optimizing compression efficiency. \nThe first stage evaluates the compressibility of existing sparse subnetworks within SNNs using the PQ index, which facilitates an adaptive determination of the rewiring ratio for synaptic connections based on data compression insights. In the second stage, this rewiring ratio critically informs the dynamic synaptic connection rewiring process, including both pruning and regrowth. This approach significantly improves the exploration of sparse structures training in deep SNNs, adapting sparsity dynamically from the point view of compression efficiency.\nOur experiments demonstrate that this sparse training approach not only aligns with the performance of current deep SNNs models but also significantly improves the efficiency of compressing sparse SNNs. Crucially, it preserves the advantages of initiating training with sparse models and offers a promising solution for implementing Edge AI on neuromorphic hardware.", "keywords": [ "spiking neural networks" ], "rating_list": [ 8, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "gcTKtwWyQm", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Ece Calikus", "gender": "Female", "institution": "Uppsala University", "country": "SE", "position": "Assistant Professor" }, { "name": "Marius Kloft", "gender": "Male", "institution": "RPTU Kaiserslautern-Landau", "country": "", "position": "Professor" }, { "name": "Philipp Liznerski", "gender": "Male", "institution": "University of Kaiserslautern", "country": "DE", "position": "PhD student" }, { "name": "Saurabh Varshneya", "gender": "Male", "institution": "Rheinland-Pf\u00e4lzische Technische Universit\u00e4t", "country": "DE", "position": "PhD student" }, { "name": "Sophie Burkhardt", "gender": "Female", "institution": "Universit\u00e4t Kaiserslautern", "country": "", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 11, "n_ref_uni": 60, "n_ref": 102, "n_ref_all": 128, "n_fig": 6, "n_tab": 22, "L_tab": 21790, "n_element_tab": 718, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2028, "n_element_tab_1": 54, "formula_len_all": 597, "formula_len_all_1": 783, "len_all": 246430, "len_all_1": 68249, "len_abs": 768, "len_title": 106, "len_sents": 63667, "len_sents_1": 29411, "n_sents": 492, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 772, "title": "Anomaly Detection Exposed: Imagining Anomalies Were Normal", "abs": "Deep learning-based methods have achieved a breakthrough in image anomaly detection, but their complexity introduces a considerable challenge to understanding why an instance is predicted to be anomalous. We introduce a novel explanation method that generates multiple alternative modifications for each anomaly, capturing diverse concepts of anomalousness. Each modification is trained to be perceived as normal by the anomaly detector. The method provides a semantic explanation of the mechanism that triggered the anomaly detector, allowing users to explore ``what-if scenarios.'' Qualitative and quantitative analyses across various image datasets demonstrate that applying this method to state-of-the-art anomaly detectors provides high-quality semantic explanations.", "keywords": [ "anomaly-detection", "deep-anomaly-detection", "anomaly", "deep-learning", "one-class-classification", "outlier-exposure" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "gcEhF4nuYI", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 12, "author_info_list": [ { "name": "Emad Barsoum", "gender": "Male", "institution": "AMD", "country": "US", "position": "Corprate Vice President" }, { "name": "Fuwei Yang", "gender": "Male", "institution": "AMD Inc.", "country": "CN", "position": "Researcher" }, { "name": "Haiduo Huang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Han Liu", "gender": "Male", "institution": "AMD", "country": "CN", "position": "Researcher" }, { "name": "Ji Liu", "gender": "Male", "institution": "AMD", "country": "CN", "position": "Researcher" }, { "name": "Jintu Zheng", "gender": "unknown", "institution": "Advanced Micro Devices", "country": "US", "position": "Intern" }, { "name": "Jinzhang Peng", "gender": "unknown", "institution": "Researcher at AMD ", "country": "CN", "position": "Researcher" }, { "name": "Lu Tian", "gender": "Female", "institution": "AMD", "country": "CN", "position": "Researcher" }, { "name": "Zekai Li", "gender": "Male", "institution": "Advanced Micro Devices", "country": "CN", "position": "Intern" }, { "name": "Zeping Li", "gender": "Male", "institution": "AMD", "country": "CN", "position": "Researcher" }, { "name": "Dong Li", "gender": "Male", "institution": "AMD Inc.", "country": "CN", "position": "Researcher" }, { "name": "haoweiz", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 40, "n_ref": 76, "n_ref_all": 102, "n_fig": 3, "n_tab": 10, "L_tab": 3016, "n_element_tab": 471, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2928, "n_element_tab_1": 159, "formula_len_all": 686, "formula_len_all_1": 371, "len_all": 159211, "len_all_1": 65934, "len_abs": 1475, "len_title": 129, "len_sents": 49534, "len_sents_1": 29874, "n_sents": 377, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1481, "title": "FTP: A Fine-grained Token-wise Pruner for Large Language Models via Token Routing", "abs": "Recently, large language models (LLMs) have demonstrated superior performance across various tasks by adhering to scaling laws, which significantly increase model size. However, the huge computation overhead during inference hinders the deployment in industrial applications. Many works leverage traditional compression approaches to boost model inference, but these always introduce additional training costs to restore the performance and the pruning results typically show noticeable performance drops compared to the original model when aiming for a specific level of acceleration. To address these issues, we propose a fine-grained token-wise pruning approach for the LLMs, which presents a learnable router to adaptively identify the less important tokens and skip them across model blocks to reduce computational cost during inference. To construct the router efficiently, we present a search-based sparsity scheduler for pruning sparsity allocation, a trainable router combined with our proposed four low-dimensional factors as input and three proposed losses. We conduct extensive experiments across different benchmarks on different LLMs to demonstrate the superiority of our method. Our approach achieves state-of-the-art (SOTA) pruning results, surpassing other existing pruning methods. For instance, our method outperforms BlockPruner and ShortGPT by approximately 10 points on both LLaMA2-7B and Qwen1.5-7B in accuracy retention at comparable token sparsity levels.", "keywords": [ "LLMs; Pruning", "Router" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "gc8QAQfXv6", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Caigao JIANG", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Defu Lian", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Gangwei Jiang", "gender": "Male", "institution": "City University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "JUN ZHOU", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Linqi Song", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Siqiao Xue", "gender": "Male", "institution": "starguest.ai", "country": "CN", "position": "Researcher" }, { "name": "Ying Wei", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Zhaoyi Li", "gender": "Male", "institution": "City University of Hong Kong", "country": "CN", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 11, "n_ref_uni": 58, "n_ref": 133, "n_ref_all": 215, "n_fig": 11, "n_tab": 2, "L_tab": 1343, "n_element_tab": 141, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 2657, "n_element_tab_1": 436, "formula_len_all": 1532, "formula_len_all_1": 532, "len_all": 217440, "len_all_1": 79468, "len_abs": 1433, "len_title": 177, "len_sents": 59217, "len_sents_1": 34434, "n_sents": 387, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 129, "L_abs": 1440, "title": "Unlocking the Power of Function Vectors for Characterizing and Mitigating Catastrophic Forgetting in Continual Instruction Tuning", "abs": "Catastrophic forgetting (CF) poses a significant challenge in machine learning, where a model forgets previously learned information upon learning new tasks. \nDespite the advanced capabilities of Large Language Models (LLMs), they continue to face challenges with CF during continual learning. The majority of existing research focuses on analyzing forgetting patterns through a singular training sequence, thereby overlooking the intricate effects that diverse tasks have on model behavior.\nOur study explores CF across various settings, discovering that model forgetting is influenced by both the specific training tasks and the models themselves. To this end, we interpret forgetting by examining the function vector (FV), a compact representation of functions in LLMs, offering a model-dependent indicator for the occurrence of CF. Through theoretical and empirical analyses, we demonstrated that CF in LLMs primarily stems from biases in function activation rather than the overwriting of task processing functions.\nLeveraging these insights, we propose a novel function vector guided training methodology, incorporating a regularization technique to stabilize the FV and mitigate forgetting. Empirical tests on four benchmarks confirm the effectiveness of our proposed training method, substantiating our theoretical framework concerning CF and model function dynamics. We plan to make our code publicly accessible in the near future.", "keywords": [ "Catastrophic forgetting; Large language model; Instruction tuning" ], "rating_list": [ 10, 10, 8, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "gc70LAWjwe", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Behzad Dariush", "gender": "Male", "institution": "Honda Research Institute USA", "country": "US", "position": "Principal Researcher" }, { "name": "Kwonjoon Lee", "gender": "Male", "institution": "Honda Research Institute USA", "country": "US", "position": "Sr Research Scientist" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Weiqing Luo", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "MS student" }, { "name": "Xinyu Zhao", "gender": "unknown", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Yifan Li", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Zhen Tan", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 23, "n_ref": 59, "n_ref_all": 87, "n_fig": 12, "n_tab": 9, "L_tab": 3654, "n_element_tab": 354, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1396, "n_element_tab_1": 151, "formula_len_all": 137, "formula_len_all_1": 137, "len_all": 146420, "len_all_1": 62563, "len_abs": 1104, "len_title": 116, "len_sents": 39698, "len_sents_1": 28474, "n_sents": 251, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1108, "title": "Beyond Fixed Resolution: Enhancing VLLMs with Adaptive Input Scaling", "abs": "Real-world vision-language applications demand varying levels of perceptual granularity. However, most existing visual large language models (VLLMs), such as LLaVA, pre-assume a fixed resolution for downstream tasks, which leads to subpar performance. To address this problem, we first conduct a comprehensive and pioneering investigation into the resolution preferences of different vision-language tasks, revealing a correlation between resolution preferences with 1.image complexity, and 2.uncertainty variance of the VLLM at different image input resolutions. Building on this insight, we propose an empirical formula to determine the optimal resolution for a given vision-language task, accounting for these two factors as the zeroth-order and first-order terms in the Taylor expansion on a given image input. Second, based on rigorous experiments, we propose a novel parameter-efficient fine-tuning technique to extend the visual input resolution of pre-trained VLLMs to the identified optimal resolution. Extensive experiments on various vision-language tasks validate the effectiveness of our method.", "keywords": [ "Visual Large Language Model" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "gbruScKTJ2", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Bahram Zonooz", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "Elahe Arani", "gender": "Female", "institution": "Eindhoven University of technology ", "country": "NL", "position": "Assistant Professor" }, { "name": "Fahad Sarfraz", "gender": "Male", "institution": "TomTom", "country": "NL", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 19, "n_ref": 27, "n_ref_all": 41, "n_fig": 9, "n_tab": 2, "L_tab": 2414, "n_element_tab": 69, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 2414, "n_element_tab_1": 69, "formula_len_all": 211, "formula_len_all_1": 211, "len_all": 87591, "len_all_1": 64245, "len_abs": 1114, "len_title": 149, "len_sents": 32638, "len_sents_1": 32060, "n_sents": 185, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1067, "title": "Dual-Pathway Neural Networks: Harnessing Scene and Object Pathways for Enhanced Visual Understanding", "abs": "Standard artificial neural networks (ANNs) often struggle with generalization due to their reliance on surface-level cues, which can lead to suboptimal performance. Drawing inspiration from the distinct processing pathways for scenes and objects in the human brain, we explore the interactions between scene and object and introduce a dual-modality architecture aimed at emulating this cognitive processing mechanism within ANNs. Our approach features separate encodings for scene and object modalities, which are fused to facilitate enhanced visual understanding. By optimizing object recognition and scene reconstruction objectives, our architecture efficiently encodes scene and object information crucial for holistic representation learning. Empirical validation demonstrates significant improvements in generalization, lifelong learning, and adversarial robustness compared to conventional architectures. These findings underscore the potential of integrating biological insights into AI systems to bridge the gap between artificial and biological intelligence.", "keywords": [ "scene and object learning", "disentangled representations", "generalization" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 1, 2, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "gbJNFxcicC", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Carine Mukamakuza", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "", "position": "Instructor" }, { "name": "Eugenia Mawuenya Akpo", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "RW", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 19, "n_ref": 23, "n_ref_all": 25, "n_fig": 1, "n_tab": 1, "L_tab": 1840, "n_element_tab": 47, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 1840, "n_element_tab_1": 47, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 65483, "len_all_1": 65483, "len_abs": 934, "len_title": 113, "len_sents": 18859, "len_sents_1": 18859, "n_sents": 131, "n_sents_1": 131, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 938, "title": "Mask R-CNN for Automated Multi-Species Malaria Parasite Detection", "abs": "This study investigates the automatic detection and segmentation of malaria parasites across various Plasmodium species using Mask R-CNN, an advanced deep-learning architecture. Expanding on earlier studies in digital malaria diagnosis, we apply pixel-level segmentation to overcome the drawbacks of previous approaches. 971 microscopic pictures of four Plasmodium species\u2014P. falciparum, P. malariae, P. ovale, and P. vivax\u2014taken from Rwanda's healthcare facilities make up our dataset. This dataset was used to train the Mask R-CNN model, which produced excellent mean average precision (mAP) scores for all species, with P. vivax and P. malariae showing the most excellent performance with mAP 0.9575 and mAP 0.9459, respectively. Compared to earlier techniques, this method shows notable advances in parasite localization and delineation, suggesting the possibility of more precise and effective malaria diagnosis in clinical settings.", "keywords": [ "Mask R-CNN", "malaria parasite detection", "Plasmodium species", "deep learning", "instance segmentation", "microscopic image analysis" ], "rating_list": [ 1, 3, 1 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 1, 1, 1 ], "confidence_list": [ 5, 5, 5 ] }, { "paper_id": "gam5LiMPKT", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Chenhang Cui", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Summer Camp Student" }, { "name": "Huaxiu Yao", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Jiabing Yang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Peng Xia", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Ying Wei", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Yiyang Zhou", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 10, "n_ref_uni": 32, "n_ref": 88, "n_ref_all": 103, "n_fig": 7, "n_tab": 5, "L_tab": 1763, "n_element_tab": 320, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 1752, "n_element_tab_1": 310, "formula_len_all": 1105, "formula_len_all_1": 625, "len_all": 182319, "len_all_1": 59773, "len_abs": 1514, "len_title": 134, "len_sents": 52629, "len_sents_1": 25020, "n_sents": 421, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1560, "title": "Fading Focus: Mitigating Visual Attention Degradation in Large Vision-Language Models", "abs": "How can we ensure that Large Vision-Language Models (LVLMs) maintain strong attention to visual input throughout the inference process? Recent advancements in Large Vision-Language Models (LVLMs) have demonstrated significant progress across multiple domains. However, these models still face the inherent challenge of integrating vision and language for collaborative inference, which often leads to \"hallucinations,\" outputs that are not grounded in the corresponding images. Many efforts have been made to address these challenges, but each approach comes with its own limitations, such as high computational costs or expensive dataset annotation. Worse still, many of them fail to recognize the crucial role of visual attention in guiding the model\u2019s response generation.\nIn our research, we identify a key limitation in current LVLMs: the model's diminishing attention to visual input as the number of generated tokens increases, which results in performance degradation. To address this challenge, we propose \\textbf{I}mage attention-guided \\textbf{K}ey-value merging c\\textbf{O}llaborative \\textbf{D}ecoding (IKOD), a collaborative decoding strategy that generates image-focused sequences using key-value merging. This method derives logits from shorter sequences with higher image attention through key-value merging and combines them with those from the original decoding process, effectively mitigating attention decay. Importantly, IKOD requires no additional training or external tools, making it highly scalable and applicable to various models.", "keywords": [ "Hallucination; Large Vision-Language Models; Decoding Strategy" ], "rating_list": [ 3, 6, 5, 3, 6 ], "soundness_list": [ 3, 3, 3, 1, 3 ], "presentation_list": [ 2, 2, 2, 1, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "gaa7gWPZBz", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Aohan Sun", "gender": "Female", "institution": "Civil Aviation University of China", "country": "CN", "position": "MS student" }, { "name": "ATHANASIOS", "gender": "unknown", "institution": "ICT", "country": "NO", "position": "Full Professor" }, { "name": "Lu Yanrong", "gender": "Female", "institution": "Civil Aviation University of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 26, "n_ref": 33, "n_ref_all": 42, "n_fig": 3, "n_tab": 7, "L_tab": 1228, "n_element_tab": 118, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 1250, "n_element_tab_1": 120, "formula_len_all": 312, "formula_len_all_1": 312, "len_all": 77713, "len_all_1": 63775, "len_abs": 1174, "len_title": 128, "len_sents": 24947, "len_sents_1": 24903, "n_sents": 170, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1199, "title": "Mitigating Privacy Risk of Adversarial Examples with Counterfactual Explanations", "abs": "Robustness and privacy are two fundamental security properties that \nmachine learning models require. Without the balance between robustness and privacy leads to \nrobust models with high privacy risks. Obtaining machine learning models with high adversarial robustness and \nprivacy performance remains an open problem. In order to enhance the privacy performance of \nrobust models, we employ counterfactual explanations as a method \nto mitigate privacy risks while concurrently maintaining robust model accuracy, reducing the privacy risk of the robust model to the level of \nrandom guessing and using counterfactual explanations to generate adversarial examples for the first time. We analyze the similarities and differences between \nadversarial examples and counterfactual explanations and utilize these properties to design the \ngeneration method. We \nconduct an in-depth analysis of the advantages offered by counterfactual explanations compared \nto traditional adversarial examples. Our study indicates that the correlation between \nrobustness and privacy is strong and the ideal balance state of accuracy, robustness, and privacy is with 95\\% \nadversarial examples involved in model training.", "keywords": [ "Adversarial Examples", "Privacy", "Counterfactual Explanations" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 1, 1, 2 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "ga9PAnFsAt", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Jinsung Yoon", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Sercan O Arik", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 28, "n_ref": 39, "n_ref_all": 62, "n_fig": 1, "n_tab": 5, "L_tab": 1668, "n_element_tab": 195, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1734, "n_element_tab_1": 209, "formula_len_all": 254, "formula_len_all_1": 254, "len_all": 128296, "len_all_1": 63895, "len_abs": 1363, "len_title": 81, "len_sents": 43147, "len_sents_1": 30736, "n_sents": 287, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1369, "title": "Embedding-Converter: A Unified Framework for Cross-Model Embedding Transformation", "abs": "Embeddings, numerical representations of data like text and images, are fundamental to machine learning. However, the continuous emergence of new embedding models poses a challenge: migrating to these potentially superior models often requires computationally expensive re-embedding of entire datasets even without guarantees of improvement. This paper introduces Embedding-Converter, a unified framework and a novel paradigm for efficiently converting embeddings between different models, eliminating the need for costly re-embedding. In real-world scenarios, the proposed method yields O(100) times faster and cheaper computation of embeddings with new models. Our experiments demonstrate that Embedding-Converter not only facilitates seamless transitions to new models but can even surpass the source model's performance, approaching that of the target model. This enables efficient evaluation of new embedding models and promotes wider adoption by reducing the overhead associated with model switching. Moreover, Embedding-Converter addresses latency constraints by enabling the use of smaller models for online tasks while leveraging larger models for offline processing. By encouraging users to release converters alongside new embedding models, Embedding-Converter fosters a more dynamic and user-friendly paradigm for embedding model development and deployment.", "keywords": [ "Embeddings", "Embedding Converter", "Embedding transformation" ], "rating_list": [ 3, 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 1, 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 1, 2, 3 ], "confidence_list": [ 4, 1, 5, 3, 3 ] }, { "paper_id": "ga4LyaucKr", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Siqiang Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Xiaotie Deng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Yunxuan Ma", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhijian Duan", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 83, "n_formula_1": 16, "n_ref_uni": 41, "n_ref": 91, "n_ref_all": 126, "n_fig": 4, "n_tab": 8, "L_tab": 1297, "n_element_tab": 121, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 60, "n_element_tab_1": 2, "formula_len_all": 9256, "formula_len_all_1": 1061, "len_all": 237524, "len_all_1": 70421, "len_abs": 1061, "len_title": 163, "len_sents": 76901, "len_sents_1": 30809, "n_sents": 716, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1041, "title": "Learning-based Mechanism Design: Scalable, Truthful, and Continuum Approaches for Utility Maximization", "abs": "Mechanism design is a crucial topic at the intersection of computer science and economics. \nThis paper addresses the automated mechanism design problem by leveraging machine learning and neural networks. \nThe objective is to design a **truthful**, **expressive** and **efficient** mechanism that maximizes the platform's expected utility, given that the players' types are drawn from a pre-specified distribution.\n\nWe present a general mechanism design model that captures two critical features: hidden information and strategic behavior. \nSubsequently, we propose the **PFM-Net** framework, which parameterizes the menu mechanism class by function approximation and identifies an optimal mechanism through ingenious optimization techniques. \nWe also provide both theoretical and empirical justifications for the advantages of our approach. \nExperimental results demonstrate the effectiveness of PFM-Net over traditional and learning-based baselines, \nenabling the PFM-Net framework to serve as a new paradigm for automated mechanism design.", "keywords": [ "automated mechanism design", "differential economics", "function approximation", "mechanism representation" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "ga3DPo6BML", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Ernie Chu", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Jun-cheng Chen", "gender": "Male", "institution": "Research Center for Information Technology Innovation, Academia Sinica", "country": "TW", "position": "Associate Research Fellow/Professor" }, { "name": "Yang Che Sun", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Undergrad student" }, { "name": "Cheng Yu Yeo", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "Yu Lun Liu", "gender": "unknown", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 35, "n_ref": 70, "n_ref_all": 84, "n_fig": 9, "n_tab": 6, "L_tab": 5078, "n_element_tab": 237, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 3363, "n_element_tab_1": 183, "formula_len_all": 911, "formula_len_all_1": 908, "len_all": 151232, "len_all_1": 58621, "len_abs": 1374, "len_title": 135, "len_sents": 30676, "len_sents_1": 25345, "n_sents": 208, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1350, "title": "FIPER: Generalizable Factorized Fields for Joint Image Compression and Super-Resolution", "abs": "In this work, we propose a unified representation for Super-Resolution (SR) and Image Compression, termed **Factorized Fields**, motivated by the shared principles between these two tasks. Both SISR and Image Compression require recovering and preserving fine image details\u2014whether by enhancing resolution or reconstructing compressed data. Unlike previous methods that mainly focus network architecture, our proposed approach utilizes a basis-coefficient decomposition to explicitly capture multi-scale visual features and structural components in images, addressing the core challenges of both tasks. We first derive our SR model, which includes Coefficient Backbone and Basis Swin Transformer for generalizable Factorized Fields. Then, to further unify these two tasks, we leverage the strong information-recovery capabilities of the trained SR modules as priors in the compression pipeline, improving both compression efficiency and detail reconstruction. Additionally, we introduce a merged-basis compression branch that consolidates shared structures, further optimizing the compression process. Extensive experiments show that our unified representation delivers state-of-the-art performance, achieving an average improvement of 204.4\\% over the baseline in Super-Resolution (SR) and 156.1\\% in Image Compression compared to the previous SOTA.", "keywords": [ "Factor fields", "Image compression", "Super-resolution" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 2 ] }, { "paper_id": "ga1sPJen12", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Jose A. Lozano", "gender": "Male", "institution": "Basque Center for Applied Mathematics", "country": "ES", "position": "Principal Researcher" }, { "name": "Josu Ceberio", "gender": "Male", "institution": "Universidad del Pa\u00eds Vasco", "country": "ES", "position": "Associate Professor" }, { "name": "Mikel Malagon", "gender": "Male", "institution": "Universidad del Pa\u00eds Vasco", "country": "ES", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 39, "n_ref": 118, "n_ref_all": 186, "n_fig": 28, "n_tab": 7, "L_tab": 1688, "n_element_tab": 210, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 457, "n_element_tab_1": 84, "formula_len_all": 641, "formula_len_all_1": 80, "len_all": 221384, "len_all_1": 61182, "len_abs": 1560, "len_title": 150, "len_sents": 67910, "len_sents_1": 28381, "n_sents": 525, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1260, "title": "Craftium: Creating Efficient Environments for Open-Ended and Embodied Agents Beyond Gridworlds", "abs": "Advancements in open-ended and embodied AI require highly adaptable and computationally efficient environments. Yet, existing platforms often lack the flexibility, efficiency, or richness necessary to drive progress in these areas. Research in fields related to open-endedness, such as unsupervised environment design and continual reinforcement learning, usually defaults to simplistic 2D grid environments, as more complex alternatives are either too rigid or computationally expensive. Conversely, in embodied AI, the field relies on fully featured video games like Minecraft, which are rich in content but computationally inefficient and offer limited customization for creating new tasks. This paper introduces Craftium, a framework based on the open-source Minetest game engine, providing a highly customizable, easy-to-use, and efficient platform for building rich Minecraft-like 3D environments. We showcase environments of different complexity and nature: from simple reinforcement learning tasks to a vast world with many creatures and biomes, along with a customizable procedural task generator. Conducted benchmarks show that Craftium substantially improves the computational cost of Minecraft-based frameworks, achieving +2K steps per second more.", "keywords": [ "reinforcement learning", "environment", "embodied", "open-ended", "continual learning", "meta reinforcement learning" ], "rating_list": [ 6, 8, 3, 8 ], "soundness_list": [ 3, 4, 3, 4 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "ga1IraEqTE", "primary_area": "datasets and benchmarks", "n_author": 18, "author_info_list": [ { "name": "Aleksandra Faust", "gender": "Female", "institution": "Google Brain", "country": "US", "position": "Principal Researcher" }, { "name": "Austin V Huang", "gender": "unknown", "institution": "Brown University", "country": "", "position": "Assistant Professor" }, { "name": "Colton Bishop", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Undergrad student" }, { "name": "Ebrahim Songhori", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Software Engineer" }, { "name": "Ikechukwu Uchendu", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Izzeddin Gur", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "J K Terry", "gender": "Female", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Jason Jabbour", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Jeffrey Jian Ma", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Jie Tan", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Joel Runevic", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Undergrad student" }, { "name": "Korneel Van den Berghe", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Researcher" }, { "name": "Matthew Stewart", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Paige Bailey", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Sergio Guadarrama", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Srivatsan Krishnan", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Vijay Janapa Reddi", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Associate Professor" }, { "name": "Wenjie Jiang", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 43, "n_ref": 64, "n_ref_all": 87, "n_fig": 0, "n_tab": 13, "L_tab": 8685, "n_element_tab": 888, "n_fig_1": 0, "n_tab_1": 6, "L_tab_1": 2826, "n_element_tab_1": 282, "formula_len_all": 287, "formula_len_all_1": 288, "len_all": 179631, "len_all_1": 61523, "len_abs": 1754, "len_title": 94, "len_sents": 43311, "len_sents_1": 27463, "n_sents": 286, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 46, "L_abs": 1762, "title": "A2Perf: Real-World Autonomous Agents Benchmark", "abs": "Autonomous agents and systems cover a number of application areas, from robotics and digital assistants to combinatorial optimization, all sharing common, unresolved research challenges. It is not sufficient for agents to merely solve a given task; they must generalize to out-of-distribution tasks, perform reliably, and use hardware resources efficiently during training and on-device deployment, among other requirements. Several major classes of methods, such as reinforcement learning and imitation learning, are commonly used to tackle these problems, each with different trade-offs. However, there is currently no benchmarking suite that defines the environments, datasets, and metrics which can be used to develop reference implementations and seed leaderboards with baselines, providing a meaningful way for the community to compare progress. We introduce A2Perf---a benchmarking suite including three environments that closely resemble real-world domains: computer chip floorplanning, web navigation, and quadruped locomotion. A2Perf provides metrics that track task performance, generalization, system resource efficiency, and reliability, which are all critical to real-world applications. In addition, we propose a data cost metric to account for the cost incurred acquiring offline data for imitation learning, reinforcement learning, and hybrid algorithms, which allows us to better compare these approaches. A2Perf also contains baseline implementations of standard algorithms, enabling apples-to-apples comparisons across methods and facilitating progress in real-world autonomy. As an open-source and extendable benchmark, A2Perf is designed to remain accessible, documented, up-to-date, and useful to the research community over the long term.", "keywords": [ "benchmark", "reinforcement learning", "autonomous agents", "agents", "benchmarking" ], "rating_list": [ 8, 3, 5, 3 ], "soundness_list": [ 4, 2, 2, 2 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 4, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "gZue5gHQHp", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Johan Barthelemy", "gender": "unknown", "institution": "University of Wollongong", "country": "AU", "position": "Associate Professor" }, { "name": "Lei Wang", "gender": "Male", "institution": "University of Wollonong", "country": "AU", "position": "Full Professor" }, { "name": "Luping Zhou", "gender": "Female", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Zailong Chen", "gender": "Male", "institution": "University of Wollongong", "country": "", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 19, "n_ref": 41, "n_ref_all": 54, "n_fig": 4, "n_tab": 5, "L_tab": 1854, "n_element_tab": 151, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1810, "n_element_tab_1": 117, "formula_len_all": 445, "formula_len_all_1": 425, "len_all": 109763, "len_all_1": 67051, "len_abs": 1534, "len_title": 149, "len_sents": 33117, "len_sents_1": 31620, "n_sents": 226, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1544, "title": "Bidirectional Learning for the Visual Representation in Radiology Report Generation with Frozen LLMs", "abs": "Radiology report generation (R2Gen) has recently leveraged large language models (LLMs), achieving improved results. However, the generated reports still fall short in both language accuracy and clinical relevance. A key challenge is learning a visual representation of radiology images that an LLM can effectively interpret. To address this, we propose that for a visual representation to be interpretable by an LLM, it shall also be generatable by the LLM. Building on this idea, we introduce a novel bidirectional learning framework for R2Gen, integrating both vision-to-text and text-to-vision information to enhance visual representation learning. First, we require that the visual representation aid the LLM in generating reports that closely match the ground truth. Second, we require that the visual representation be maximally generated by the LLM when provided with the ground truth report. To enable the frozen LLM to perform text-to-vision generation, we jointly train a new text encoder for reports. Additionally, through an image reconstruction task, we encourage the visual representation to capture the core features of input radiology images. This bidirectional learning framework is realized using a frozen LLM and incurs no extra computational cost at the inference stage. Experimental results demonstrate better alignment between the learned visual representation and the LLM\u2019s word embedding space, along with state-of-the-art performance in both language accuracy and clinical efficacy. Our code will be publicly released.", "keywords": [ "Bidirectional Learning", "Radiology Report Generation", "Representation Learning", "Large Language Models." ], "rating_list": [ 5, 3, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 3, 5, 5, 4, 3 ] }, { "paper_id": "gZnBI7WS1K", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Bingxin Xu", "gender": "Female", "institution": "Illinois Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Mu Cai", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin, Madison", "country": "US", "position": "PhD student" }, { "name": "Yan Yan", "gender": "Male", "institution": "University of Illinois Chicago", "country": "US", "position": "Associate Professor" }, { "name": "Yuzhang Shang", "gender": "Male", "institution": "Illinois Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yong Jae Lee", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 39, "n_ref": 111, "n_ref_all": 128, "n_fig": 4, "n_tab": 7, "L_tab": 2793, "n_element_tab": 517, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1151, "n_element_tab_1": 270, "formula_len_all": 606, "formula_len_all_1": 606, "len_all": 136925, "len_all_1": 62600, "len_abs": 1707, "len_title": 78, "len_sents": 34121, "len_sents_1": 27701, "n_sents": 249, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1674, "title": "LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Models", "abs": "Large Multimodal Models (LMMs) have shown significant visual reasoning capabilities by connecting a visual encoder and a large language model. LMMs typically take in a fixed and large amount of visual tokens, such as the penultimate layer features in the CLIP visual encoder, as the prefix content. Recent LMMs incorporate more complex visual inputs, such as high-resolution images and videos, which further increases the number of visual tokens significantly. However, due to the inherent design of the Transformer architecture, the computational costs of these models tend to increase quadratically with the number of input tokens. To tackle this problem, we explore a token reduction mechanism that identifies significant spatial redundancy among visual tokens. In response, we propose PruMerge, a novel adaptive visual token reduction strategy that significantly reduces the number of visual tokens without compromising the performance of LMMs. Specifically, to metric the importance of each token, we exploit the sparsity observed in the visual encoder, characterized by the sparse distribution of attention scores between the class token and visual tokens. This sparsity enables us to dynamically select the most crucial visual tokens to retain. Subsequently, we cluster the selected (unpruned) tokens based on their key similarity and merge them with the unpruned tokens, effectively supplementing and enhancing their informational content.\nEmpirically, when applied to LLaVA-1.5 and Video-LLaVA, our approach can reduce the number of visual tokens by 4 times, and achieve comparable or better performance across diverse visual question-answering and reasoning tasks.", "keywords": [ "Efficient AI", "Large Multimodal Model" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "gZky2pakRK", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 12, "author_info_list": [ { "name": "Bill Y Lin", "gender": "Male", "institution": "xAI", "country": "US", "position": "Researcher" }, { "name": "Faeze Brahman", "gender": "Female", "institution": "Allen Institute for AI", "country": "US", "position": "Postdoc" }, { "name": "Frank F. Xu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Hao Zhu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Hyunwoo Kim", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Postdoc" }, { "name": "Liwei Jiang", "gender": "Female", "institution": "University of Washington", "country": "", "position": "PhD student" }, { "name": "Maarten Sap", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Ronan Le Bras", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Ximing Lu", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Xuhui Zhou", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Yejin Choi", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Fatemehsadat Mireshghallah", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "Postdoc" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 63, "n_ref": 155, "n_ref_all": 179, "n_fig": 6, "n_tab": 6, "L_tab": 2629, "n_element_tab": 310, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 942, "n_element_tab_1": 75, "formula_len_all": 62, "formula_len_all_1": 0, "len_all": 423382, "len_all_1": 64104, "len_abs": 1541, "len_title": 79, "len_sents": 106341, "len_sents_1": 31109, "n_sents": 800, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1571, "title": "HAICOSYSTEM: An Ecosystem for Sandboxing Safety Risks in Human-AI Interactions", "abs": "AI agents are increasingly autonomous in their interactions with human users and tools, leading to increased interactional safety risks. We present HAICOSYSTEM, a framework examining AI agent safety within diverse and complex social interactions. HAICOSYSTEM features a modular sandbox environment that simulates multi-turn interactions between human users and AI agents, where the AI agents are equipped with a variety of tools (e.g., patient management platforms) to navigate diverse scenarios (e.g., a user attempting to access other patients' profiles). To examine the safety of AI agents in these interactions, we develop a comprehensive multi-dimensional evaluation framework that uses metrics covering operational, content-related, societal, and legal risks. Through running over 8k simulations based on 132 scenarios across seven domains (e.g., healthcare, finance, education), we demonstrate that HAICOSYSTEM can emulate realistic user-AI interactions and complex tool use by AI agents. Our experiments show that state-of-the-art LLMs, both proprietary and open-sourced, exhibit safety risks in over 62\\% cases, with models generally showing higher risks when interacting with simulated malicious users. Our findings highlight the ongoing challenge of building agents that can safely navigate complex interactions, particularly when faced with malicious users. To foster the AI agent safety ecosystem, we release a code platform that allows practitioners to create custom scenarios, simulate interactions, and evaluate the safety and performance of their agents.", "keywords": [ "AI Safety", "Multi-Agent Systems", "Human-AI Interaction", "Social Simulation" ], "rating_list": [ 8, 6, 8, 3 ], "soundness_list": [ 2, 3, 4, 1 ], "presentation_list": [ 4, 2, 4, 4 ], "contribution_list": [ 3, 3, 4, 1 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "gYWqxXE5RJ", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Saeed Hassanpour", "gender": "unknown", "institution": "Dartmouth College", "country": "", "position": "Associate Professor" }, { "name": "Soroush Vosoughi", "gender": "unknown", "institution": "Dartmouth College", "country": "", "position": "Assistant Professor" }, { "name": "Weimin Lyu", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Xiaomeng Zhu", "gender": "unknown", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Yuxin Wang", "gender": "Female", "institution": "Dartmouth College", "country": "US", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 35, "n_ref": 44, "n_ref_all": 80, "n_fig": 20, "n_tab": 15, "L_tab": 8027, "n_element_tab": 456, "n_fig_1": 11, "n_tab_1": 5, "L_tab_1": 1156, "n_element_tab_1": 107, "formula_len_all": 1072, "formula_len_all_1": 953, "len_all": 246222, "len_all_1": 74573, "len_abs": 1384, "len_title": 128, "len_sents": 67502, "len_sents_1": 32890, "n_sents": 645, "n_sents_1": 282, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 79, "L_abs": 1391, "title": "ImpScore: A Learnable Metric For Quantifying The Implicitness Level of Language", "abs": "Handling implicit language is essential for natural language processing systems to achieve precise text understanding and facilitate natural interactions with users. Despite its importance, the absence of a robust metric for accurately measuring the implicitness of language significantly constrains the depth of analysis possible in evaluating models' comprehension capabilities. This paper addresses this gap by developing a scalar metric that quantifies the implicitness level of language without relying on external references. Drawing on principles from traditional linguistics, we define \"implicitness\" as the divergence between semantic meaning and pragmatic interpretation. To operationalize this definition, we introduce ImpScore, a novel, reference-free metric formulated through an interpretable regression model. This model is trained using pairwise contrastive learning on a specially curated dataset comprising $112,580$ (implicit sentence, explicit sentence) pairs. We validate ImpScore through a user study that compares its assessments with human evaluations on out-of-distribution data, demonstrating its accuracy and strong correlation with human judgments. Additionally, we apply ImpScore to hate speech detection datasets, illustrating its utility and highlighting significant limitations in current large language models' ability to understand highly implicit content.", "keywords": [ "implicit language", "pragmatics", "learnable metric", "text evaluation", "automatic evaluation", "explicit language" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 2, 4 ], "presentation_list": [ 4, 3, 4 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "gYNBQygmXG", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Aleksa Sukovic", "gender": "Male", "institution": "MPI-SWS", "country": "DE", "position": "Researcher" }, { "name": "Goran Radanovic", "gender": "unknown", "institution": "MPI-SWS", "country": "DE", "position": "Research group leader" }, { "name": "Stelios Triantafyllou", "gender": "Male", "institution": "MPI-SWS", "country": "DE", "position": "PhD student" }, { "name": "Yasaman Zolfimoselo", "gender": "Female", "institution": "MPI-SWS", "country": "DE", "position": "Intern" } ], "n_formula": 33, "n_formula_1": 18, "n_ref_uni": 42, "n_ref": 85, "n_ref_all": 111, "n_fig": 9, "n_tab": 2, "L_tab": 567, "n_element_tab": 31, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3878, "formula_len_all_1": 1755, "len_all": 187303, "len_all_1": 71741, "len_abs": 1522, "len_title": 125, "len_sents": 69572, "len_sents_1": 34981, "n_sents": 600, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1539, "title": "Counterfactual Effect Decomposition in Multi-Agent Sequential Decision Making", "abs": "We address the challenge of explaining counterfactual outcomes in multi-agent Markov decision processes. In particular, we aim to explain the total counterfactual effect of an agent's action on the outcome of a realized scenario through its influence on the environment dynamics and the agents' behavior. To achieve this, we introduce a novel *causal explanation formula* that decomposes the counterfactual effect by attributing to each agent and state variable a score reflecting their respective contributions to the effect. First, we show that the total counterfactual effect of an agent's action can be decomposed into two components: one measuring the effect that propagates through all subsequent agents' actions and another related to the effect that propagates through the state transitions. Building on recent advancements in causal contribution analysis, we further decompose these two effects as follows. For the former, we consider *agent-specific effects* -- a causal concept that quantifies the counterfactual effect of an agent's action that propagates through a subset of agents. Based on this notion, we use Shapley value to attribute the effect to individual agents. For the latter, we consider the concept of *structure-preserving interventions* and attribute the effect to state variables based on their \"intrinsic'' contributions. Through extensive experimentation, we demonstrate the interpretability of our decomposition approach in a Gridworld environment with LLM-assisted agents and a sepsis management simulator.", "keywords": [ "counterfactual reasoning", "causal explanation formula", "multi-agent Markov decision processes", "accountability" ], "rating_list": [ 6, 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 2, 4, 3, 3, 4 ] }, { "paper_id": "gY2IHLUJhk", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chanmi Lee", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Guoyuan An", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Jinhwan Seo", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Sung-eui Yoon", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "Professor" }, { "name": "Woojung Son", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "MS student" }, { "name": "Yoonki Cho", "gender": "unknown", "institution": "KAIST", "country": "KR", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 52, "n_ref": 209, "n_ref_all": 238, "n_fig": 8, "n_tab": 8, "L_tab": 3646, "n_element_tab": 453, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2004, "n_element_tab_1": 276, "formula_len_all": 255, "formula_len_all_1": 258, "len_all": 206923, "len_all_1": 64755, "len_abs": 1620, "len_title": 103, "len_sents": 65846, "len_sents_1": 28200, "n_sents": 548, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1634, "title": "Towards Robustness of Person Search against Corruptions", "abs": "Person search aims to simultaneously detect and re-identify a query person within an entire scene, involving detection and re-identification as a multi-task problem.\nWhile existing studies have made significant progress in achieving superior performance on clean datasets, the challenge of robustness under various corruptions remains largely unexplored.\nTo address this gap, we propose two benchmarks, CUHK-SYSU-C and PRW-C, designed to assess the robustness of person search models across diverse corruption scenarios.\nPrevious researches on corruption have been conducted independently for single tasks such as re-identification and detection.\nHowever, recent advancements in person search adopt an end-to-end multi-task learning framework that processes the entire scene as input, unlike the combination of single tasks. \nThis raises the question of whether independent achievements can ensure corruption robustness for person search.\nOur findings reveal that merely combining independent, robust detection and re-identification models is not sufficient for achieving robust person search. \nWe further investigate the vulnerability of the detection and representation stages to corruption and explore its impact on both foreground and background areas.\nBased on these insights, we propose a foreground-aware augmentation and regularization method to enhance the robustness of person search models.\nSupported by our comprehensive robustness analysis and evaluation framework our benchmarks provide, our proposed technique substantially improves the robustness of existing person search models.\nCode will be made publicly available.", "keywords": [ "Corruption", "Person Search", "Robustness" ], "rating_list": [ 8, 3, 8, 3, 8 ], "soundness_list": [ 3, 3, 4, 3, 2 ], "presentation_list": [ 4, 3, 4, 3, 3 ], "contribution_list": [ 4, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "gY08Ou8EL7", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Alex Schwing", "gender": "Unspecified", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Associate Professor" }, { "name": "Jing Wen", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Shenlong Wang", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 51, "n_ref": 105, "n_ref_all": 143, "n_fig": 13, "n_tab": 6, "L_tab": 1040, "n_element_tab": 111, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 461, "n_element_tab_1": 45, "formula_len_all": 3217, "formula_len_all_1": 2234, "len_all": 159111, "len_all_1": 71667, "len_abs": 1478, "len_title": 149, "len_sents": 43390, "len_sents_1": 29209, "n_sents": 418, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 111, "L_abs": 1350, "title": "LIFe-GoM: Generalizable Human Rendering with Learned Iterative Feedback Over Multi-Resolution Gaussians-on-Mesh", "abs": "Generalizable rendering of an animatable human avatar from sparse inputs relies on data priors and inductive biases extracted from training on large data to avoid scene-specific optimization and to enable fast reconstruction. This raises two main challenges: First, unlike iterative gradient-based adjustment in scene-specific optimization, generalizable methods must reconstruct the human shape representation in a single pass at inference time.\nSecond, rendering is preferably computationally efficient yet of high resolution.\nTo address both challenges we augment the recently proposed \ndual shape representation, which combines the benefits of a mesh and Gaussian points, in two ways. \nTo improve reconstruction, we propose an iterative feedback update framework, which successively improves the canonical human shape representation during reconstruction.\nTo achieve computationally efficient yet high-resolution rendering, we study a coupled-multi-resolution Gaussians-on-Mesh representation.\nWe evaluate the proposed approach on the challenging THuman2.0, XHuman and AIST++ data. Our approach reconstructs an animatable representation from sparse inputs in less than 1s, renders views with 95.1FPS at $1024 \\times 1024$, and achieves PSNR/LPIPS*/FID of 24.65/110.82/51.27 on THuman2.0, outperforming the state-of-the-art in rendering quality.", "keywords": [ "Generalizable human rendering", "error feedback", "dual representation" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "gXyWbl71n1", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Jinglin Xu", "gender": "Female", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "Associate Professor" }, { "name": "Yanzhe Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yuxin Peng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Zhiwen Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 3, "n_ref_uni": 31, "n_ref": 62, "n_ref_all": 81, "n_fig": 5, "n_tab": 13, "L_tab": 1847, "n_element_tab": 170, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 789, "n_element_tab_1": 29, "formula_len_all": 865, "formula_len_all_1": 140, "len_all": 166042, "len_all_1": 54150, "len_abs": 1761, "len_title": 122, "len_sents": 39425, "len_sents_1": 25759, "n_sents": 319, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1585, "title": "MAI: A Multi-turn Aggregation-Iteration Model for Composed Image Retrieval", "abs": "Multi-Turn Composed Image Retrieval (MTCIR) addresses a real-world scenario where users iteratively refine retrieval results by providing additional information until a target meeting all their requirements is found. Existing methods primarily achieve MTCIR through a \"multiple single-turn\" paradigm, wherein methods incorrectly converge on shortcuts that only utilize the most recent turn's image, ignoring attributes from historical turns. Consequently, retrieval failures occur when modification requests involve historical information. We argue that explicitly incorporating historical information into the modified text is crucial to addressing this issue. To this end, we build a new retrospective-based MTCIR dataset, **FashionMT**, wherein modification demands are highly associated with historical turns. We also propose a Multi-turn Aggregation-Iteration (**MAI**) model, emphasizing efficient aggregation of multimodal semantics and optimization of information propagation in multi-turn retrieval. Specifically, we propose a new Two-stage Semantic Aggregation (TSA) paradigm coupled with a Cyclic Combination Loss (CCL), achieving improved semantic consistency and modality alignment by progressively interacting the reference image with its caption and the modified text. In addition, we design a Multi-turn Iterative Optimization (MIO) mechanism that dynamically selects representative tokens and reduces redundancy during multi-turn iterations. Extensive experiments demonstrate that the proposed MAI model achieves substantial improvements over state-of-the-art methods.", "keywords": [ "Multi-modal", "Multi-turn retrieval", "Retrospective-based", "Composed image retrieval" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 3, 2, 4 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "gXV84CnMUm", "primary_area": "reinforcement learning", "n_author": 10, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Jinyi Liu", "gender": "unknown", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Shanqi Liu", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Wenya Wei", "gender": "Female", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Yifu Yuan", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Yujing Hu", "gender": "unknown", "institution": "NetEase, Inc.", "country": "", "position": "Researcher" }, { "name": "Zhenxing Ge", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "PhD student" }, { "name": "Zhou Fang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Associate Professor" }, { "name": "qianyi fu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Evan Zhu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 10, "n_ref_uni": 49, "n_ref": 117, "n_ref_all": 154, "n_fig": 10, "n_tab": 9, "L_tab": 1148, "n_element_tab": 83, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2248, "formula_len_all_1": 867, "len_all": 159793, "len_all_1": 62359, "len_abs": 1836, "len_title": 166, "len_sents": 48026, "len_sents_1": 30189, "n_sents": 355, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 117, "L_abs": 1847, "title": "Outward Odyssey: Improving Reward Models with Proximal Policy Exploration for Preference-Based Reinforcement Learning", "abs": "Reinforcement learning (RL) heavily depends on well-designed reward functions, which can be challenging to create and may introduce biases, especially for complex behaviors. Preference-based RL (PbRL) addresses this by using human feedback to construct a reward model that reflects human preferences, yet requiring considerable human involvement. To alleviate this, several PbRL methods aim to select queries that need minimal feedback. However, these methods do not directly enhance the data coverage within the preference buffer. In this paper, to emphasize the critical role of preference buffer coverage in determining the quality of the reward model, we first investigate and find that a reward model's evaluative accuracy is the highest for trajectories within the preference buffer's distribution and significantly decreases for out-of-distribution trajectories. Against this phenomenon, we introduce the **Proximal Policy Exploration (PPE)** algorithm, which consists of a *proximal-policy extension* method and a *mixture distribution query* method.\nTo achieve higher preference buffer coverage, the *proximal-policy extension* method encourages active exploration of data within near-policy regions that fall outside the preference buffer's distribution. To balance the inclusion of in-distribution and out-of-distribution data, the *mixture distribution query* method proactively selects a mix of data from both outside and within the preference buffer's distribution for querying. PPE not only expands the preference buffer's coverage but also ensures the reward model's evaluative capability for in-distribution data. Our comprehensive experiments demonstrate that PPE achieves significant improvement in both human feedback efficiency and RL sample efficiency, underscoring the importance of preference buffer coverage in PbRL tasks.", "keywords": [ "Preference-based Reinforcement Learning; Reinforcement Learning; Human Feedback" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 3, 2, 4 ] }, { "paper_id": "gXK3Y6WNVv", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Jian Jornbowrl Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" }, { "name": "Jiaolong Kong", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "PhD student" }, { "name": "Jiongchi Yu", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "PhD student" }, { "name": "Shangqing Liu", "gender": "unknown", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" }, { "name": "Xiaofei Xie", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Yi Li", "gender": "Male", "institution": "School of Computer Science and Engineering, Nanyang Technological University", "country": "SG", "position": "Associate Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 35, "n_ref": 81, "n_ref_all": 104, "n_fig": 2, "n_tab": 6, "L_tab": 5890, "n_element_tab": 1023, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1122, "n_element_tab_1": 99, "formula_len_all": 113, "formula_len_all_1": 0, "len_all": 173363, "len_all_1": 64715, "len_abs": 1851, "len_title": 71, "len_sents": 54659, "len_sents_1": 31784, "n_sents": 395, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1613, "title": "Defects4C: Benchmarking C/C++ Faults to Assess LLM-Based Program Repair", "abs": "Automated Program Repair (APR) plays a pivotal role in ensuring the quality and reliability of software. However, most existing APR research focuses on Java programs, primarily due to the well-established benchmark such as Defects4J. Despite the significant prevalence of C/C++ vulnerabilities, the field lacks extensive research on the automated repair of such vulnerabilities, primarily attributed to the absence of high-quality open-source benchmarks in this domain.\n\nTo address the critical gap in available datasets for C/C++ program repair, this paper introduces Defects4C, a comprehensive and high-quality executable benchmark designed to improve defect detection and repair. The dataset includes a vast collection of bug-relevant commits (e.g., **9M** in total), **248** high-quality buggy functions and **102** vulnerable functions paired with test cases for reproduction. These datasets can be used to evaluate repair techniques and to retrain learning-based methods for improved performance. Using this expanded dataset, we evaluate the performance of state-of-the-art LLM-based automated program repair techniques in addressing C/C++ faults. Specifically, we conduct an extensive empirical study with **24** leading LLMs. Our findings provide valuable insights into the capabilities and limitations of existing APR approaches for C/C++ programs, underscoring the necessity for novel APR techniques and the significance of Defects4C. This dataset marks a significant advancement in the field, offering a robust and comprehensive C/C++ dataset that is instrumental for future research on program repair.", "keywords": [ "Defects4C; Large Language Model; Program Repair" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "gWrWUaCbMa", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Allan Douglas Jepson", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "Emeritus" }, { "name": "Amir-massoud Farahmand", "gender": "Male", "institution": "Polytechnique Montr\u00e9al", "country": "CA", "position": "Associate Professor" }, { "name": "Mete Kemertas", "gender": "unknown", "institution": "University of Toronto", "country": "", "position": "PhD Student" } ], "n_formula": 106, "n_formula_1": 33, "n_ref_uni": 36, "n_ref": 96, "n_ref_all": 150, "n_fig": 5, "n_tab": 3, "L_tab": 267, "n_element_tab": 43, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 351, "n_element_tab_1": 52, "formula_len_all": 10945, "formula_len_all_1": 2849, "len_all": 214662, "len_all_1": 73374, "len_abs": 1049, "len_title": 95, "len_sents": 66392, "len_sents_1": 29932, "n_sents": 641, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1061, "title": "A Truncated Newton Method for Optimal Transport", "abs": "Developing a contemporary optimal transport (OT) solver requires navigating trade-offs among several critical requirements: GPU parallelization, scalability to high-dimensional problems, theoretical convergence guarantees, empirical performance in terms of precision versus runtime, and numerical stability in practice. With these challenges in mind, we introduce a specialized truncated Newton algorithm for entropic regularized OT. In addition to proving that locally quadratic convergence is possible without assuming a Lipschitz Hessian, we provide strategies to maximally exploit the high rate of local convergence in practice. Our GPU-parallel algorithm exhibits exceptionally favorable runtime performance, achieving high precision orders of magnitude faster than many existing alternatives. This is evidenced by wall-clock time experiments on 4096-dimensional MNIST and color transfer problems. The scalability of the algorithm is showcased on an extremely large OT problem with $n \\approx 10^6$, solved approximately under weak entopric regularization.", "keywords": [ "Computational optimal transport", "numerical optimization", "numerical linear algebra" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "gWqFbnKsqR", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Binbin Lin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Chunhua Shen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Di Huang", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Haifeng Liu", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Honghui Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Tong He", "gender": "Male", "institution": "Shanghai AI lab", "country": "CN", "position": "Researcher" }, { "name": "Wanli Ouyang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Full Professor" }, { "name": "Wei Yin", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Xiaofei He", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 45, "n_ref": 88, "n_ref_all": 110, "n_fig": 6, "n_tab": 5, "L_tab": 2696, "n_element_tab": 379, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 3143, "n_element_tab_1": 419, "formula_len_all": 434, "formula_len_all_1": 434, "len_all": 140664, "len_all_1": 61410, "len_abs": 495, "len_title": 92, "len_sents": 30802, "len_sents_1": 25105, "n_sents": 238, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 44, "L_abs": 1290, "title": "Depth Any Video with Scalable Synthetic Data", "abs": "Video depth estimation has long been hindered by the scarcity of consistent and scalable ground truth data, leading to inconsistent and unreliable results. In this paper, we introduce Depth Any Video, a model that tackles the challenge through two key innovations. First, we develop a scalable synthetic data pipeline, capturing real-time video depth data from diverse game environments, yielding 40,000 video clips of 5-second duration, each with precise depth annotations. Second, we leverage the powerful priors of generative video diffusion models to handle real-world videos effectively, integrating advanced techniques such as rotary position encoding and flow matching to further enhance flexibility and efficiency. Unlike previous models, which are limited to fixed-length video sequences, our approach introduces a novel mixed-duration training strategy that handles videos of varying lengths and performs robustly across different frame rates\u2014even on single frames. At inference, we propose a depth interpolation method that enables our model to infer high-resolution video depth across sequences of up to 150 frames. Our model outperforms all previous generative depth models in terms of spatial accuracy and temporal consistency. The code and model weights will be open-sourced.", "keywords": [ "Video Depth Estimation", "Synthetic Game Data" ], "rating_list": [ 5, 10, 5 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 4, 2 ], "confidence_list": [ 3, 5, 4 ] }, { "paper_id": "gWk8WQVWGr", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Chenliang Xu", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Associate Professor" }, { "name": "Xin LIANG", "gender": "Female", "institution": "University of Rochester", "country": "US", "position": "Intern" }, { "name": "Zeliang Zhang", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 18, "n_ref": 26, "n_ref_all": 32, "n_fig": 3, "n_tab": 5, "L_tab": 3546, "n_element_tab": 537, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 5772, "n_element_tab_1": 563, "formula_len_all": 130, "formula_len_all_1": 130, "len_all": 84989, "len_all_1": 67426, "len_abs": 1207, "len_title": 65, "len_sents": 22964, "len_sents_1": 22035, "n_sents": 153, "n_sents_1": 146, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1213, "title": "Towards good practice in boosting the targeted adversarial attack", "abs": "By accessing only the surrogate model, attackers can craft adversarial perturbations to fool black-box victim models into misclassifying a given image into the target class. However, the misalignment between surrogate models and victim models raises concerns about defining what constitutes a successful targeted attack in a black-box setting. In our work, we empirically identify that the vision-language foundation model CLIP is a natural good indicator to evaluate a good transferable targeted attacks. We find that a successful transferable targeted attack not only confuse the model on the vision modality towards the target class, but also fool the model on the text modality between the original class and target class. Motivated by this finding, we propose a simple yet effective regularization term to boost the existing transferable targeted attacks. We also revisit the feature-based attacks, and propose to boost the performance by enhancing the fine-grained features. Extensive experiments on the ImageNet-1k dataset demonstrate the effectiveness of our proposed methods. We hope our finding can motivate future research on the understanding of targeted attacks and develop more powerful techniques.", "keywords": [ "adversariak attack" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 1, 1, 1, 3 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "gWgaypDBs8", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Anh-Dung Dinh", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Daochang Liu", "gender": "Male", "institution": "University of Western Australia", "country": "AU", "position": "Lecturer" }, { "name": "Chang Xu", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" } ], "n_formula": 35, "n_formula_1": 24, "n_ref_uni": 9, "n_ref": 27, "n_ref_all": 47, "n_fig": 6, "n_tab": 9, "L_tab": 2511, "n_element_tab": 338, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 3034, "n_element_tab_1": 268, "formula_len_all": 3186, "formula_len_all_1": 1670, "len_all": 117100, "len_all_1": 69972, "len_abs": 1211, "len_title": 108, "len_sents": 31266, "len_sents_1": 27614, "n_sents": 286, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 66, "L_abs": 1226, "title": "Representative Guidance: Diffusion Model Sampling with Consistency", "abs": "The diffusion sampling process faces a persistent challenge stemming from its incoherence, attributable to varying noise directions across different time steps. \nOur Representative Guidance (RepG) offers a new perspective to handle this issue by reformulating the sampling process with a coherent direction towards a representative target.\nIn this formulation, while the classic classifier guidance improves feature discernment by steering the model away from ambiguous features, it fails to provide a favorable representative target, since the class label is overly compact and leads to sacrificed diversity and the adversarial generation problem.\nIn contrast, we leverage self-supervised representations as the coherent target and treat sampling as a downstream task, which refines image details and corrects errors rather than settling for simpler samples.\nOur representative guidance achieves superior performance and also illustrates the potential of pre-trained self-supervised models in image sampling. Our findings demonstrate that RepG not only substantially enhances vanilla diffusion sampling but also surpasses state-of-the-art benchmarks when combined with the classifier-free guidance. Our code will be released.", "keywords": [ "generative models", "diffusion model" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 4, 2, 2 ], "contribution_list": [ 4, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "gWOANrFJ0t", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Guanghao Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Mengze Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Postdoc" }, { "name": "Wei Xue", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Wenhan Luo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Xinyu Liu", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yan Li", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yike Guo", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Zhenyi Wang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Postdoc" } ], "n_formula": 14, "n_formula_1": 5, "n_ref_uni": 39, "n_ref": 79, "n_ref_all": 114, "n_fig": 16, "n_tab": 2, "L_tab": 941, "n_element_tab": 123, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 941, "n_element_tab_1": 123, "formula_len_all": 682, "formula_len_all_1": 351, "len_all": 135097, "len_all_1": 67796, "len_abs": 1437, "len_title": 113, "len_sents": 41039, "len_sents_1": 30900, "n_sents": 295, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1453, "title": "GuideEdit: Enhancing Face Video Editing with Fine-grained Control", "abs": "Face video editing (FVE) requires maintaining temporal consistency and iden-\ntity preservation while manipulating specific attributes. However, existing FVE\nmethods often introduce unwanted artifacts and affect non-target attributes during\nediting. To address these limitations, we propose GuideEdit to enhance the pre-\ncision of face video editing. Given the inherent linearity of the latent variables in\nthe bottleneck layer of the diffusion U-Net model, there exists a linear mapping\nbetween the input and the latent representation. This allows us to extract a latent\nbasis within the latent space that effectively encodes the key features related to\ntarget facial attributes. By comparing the latent basis of the original video to that\nof the manipulated video, we quantify the manipulation degree, which indicates\nthe extent of changes made. This manipulation degree serves as a guide for deter-\nmining the specific components to be edited, then we achieve more precise control\nat each denoising step. Integrating this fine-grained control into the editing pro-\ncess allows GuideEdit to enhance temporal consistency and preserve identity of\nFVE, while minimizing the introduction of artifacts. Extensive experiments on\ndiverse real-world videos demonstrate the effectiveness of GuideEdit, showcas-\ning its ability to achieve precise, high-quality edits that maintain coherence across\nframes and ensure the preservation of essential visual elements.", "keywords": [ "Face video editing", "Diffusion model" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "gWHQQagPbN", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 9, "author_info_list": [ { "name": "Chang Gao", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Han Bao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Jianfei Chen", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Kang Zhao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Tao Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Intern" }, { "name": "Zhaofeng Sun", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhenfeng Su", "gender": "Not Specified", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Intern" }, { "name": "Zichen Liang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "liping jing", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 41, "n_ref": 61, "n_ref_all": 95, "n_fig": 11, "n_tab": 17, "L_tab": 5103, "n_element_tab": 560, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 4649, "n_element_tab_1": 366, "formula_len_all": 562, "formula_len_all_1": 562, "len_all": 161024, "len_all_1": 67170, "len_abs": 2047, "len_title": 129, "len_sents": 45734, "len_sents_1": 27075, "n_sents": 355, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 2067, "title": "Beyond 2:4: Exploring V:N:M Sparsity for Efficient Transformer Inference on GPUs", "abs": "To date, 2:4 sparsity has stood as the only sparse pattern that can be accelerated\nusing sparse tensor cores on GPUs. In practice, 2:4 sparsity often possesses low\nactual speedups (\u2264 1.3) and requires fixed sparse ratios, meaning that other ratios,\nsuch as 4:8, 8:16, or those exceeding 50% sparsity, do not incur any speedups on\nGPUs. Recent studies suggest that V:N:M sparsity is promising in addressing\nthese limitations of 2:4 sparsity. This sparsity divides a weight matrix into mul-\ntiple V\u00d7M blocks, pruning (M-4) columns within each block and applying 2:4\nsparsity to the remaining columns. V:N:M sparsity inherently encompasses 2:4\nsparsity but allows for higher and more flexible pruning ratios, typically resulting\nin greater practical speedups. However, regarding accuracy, the effects of V:N:M\nsparsity on broader Transformer models, such as vision Transformers and large\nlanguage models (LLMs), are largely unexamined. Moreover, Some specific is-\nsues related to V:N:M sparsity, such as how to select appropriate V and M values,\nremain unresolved. In this study, we thoroughly investigate the application of\nV:N:M sparsity in vision models and LLMs across multiple tasks, from pretaining\nto downstream tasks. We propose three key approaches to enhance the applica-\nbility and accuracy of V:N:M-sparse Transformers, including heuristic V and M\nselection, V:N:M-specific channel permutation and three-staged LoRA training\ntechniques. Experimental results show that, with our methods, the DeiT-small\nachieves lossless accuracy at 64:2:5 sparsity, while the DeiT-base maintains ac-\ncuracy even at 64:2:8 sparsity. In addition, the fine-tuned LLama2-7B at 64:2:5\nsparsity performs comparably or better than training-free 2:4 sparse alternatives on\ndownstream tasks. More importantly, V:N:M-sparse Transformers offer a wider\nrange of speedup-accuracy trade-offs compared to 2:4 sparsity. Overall, our explo-\nration largely facilitates the V:N:M sparsity to act as a truly effective acceleration\nsolution for Transformers in cost-sensitive inference scenarios.", "keywords": [ "V:N:M sparisity", "Transformer inference acceleration", "accuracy-speedup tradeoffs", "channel permutation" ], "rating_list": [ 6, 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4, 3 ] }, { "paper_id": "gW4bdLwypB", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "A F M Saif", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "PhD student" }, { "name": "Brian Kingsbury", "gender": "Male", "institution": "IBM", "country": "US", "position": "Researcher" }, { "name": "Lisha Chen", "gender": "Female", "institution": "University of Rochester", "country": "US", "position": "Assistant Professor" }, { "name": "Songtao Lu", "gender": "Male", "institution": "IBM Thomas J. Watson Research Center", "country": "", "position": "Senior Research Scientist" }, { "name": "Tianyi Chen", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "Assistant Professor" }, { "name": "Xiaodong Cui", "gender": "Male", "institution": "IBM T. J. Watson Research Center", "country": "US", "position": "Principal Research Staff Member" } ], "n_formula": 54, "n_formula_1": 11, "n_ref_uni": 46, "n_ref": 61, "n_ref_all": 94, "n_fig": 3, "n_tab": 7, "L_tab": 2239, "n_element_tab": 310, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2860, "n_element_tab_1": 229, "formula_len_all": 4712, "formula_len_all_1": 1215, "len_all": 185221, "len_all_1": 63881, "len_abs": 1375, "len_title": 128, "len_sents": 57666, "len_sents_1": 26131, "n_sents": 476, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1367, "title": "Objective Soups: Multilingual Multi-Task Acoustic Modeling for Automatic Speech Recognition", "abs": "The need for training multilingual multi-task automatic speech recognition (ASR) models is increasingly evident. However, a significant challenge arises from the conflicts among multiple objectives when using a single model. Multi-objective optimization (MOO) can address this challenge by facilitating the optimization of multiple conflicting objectives, aligning the gradient updates in a common descent direction. While MOO helps avoid conflicting gradient update directions, a critical issue is that when there are many objectives such as those in multilingual multi-task ASR, it is often impossible to find such common descent directions. Therefore, an interesting question is: would it be more effective to separate highly conflicting objectives into different optimization levels or keep them in one level? To address this question, this paper investigates three multi-objective ASR training frameworks, which we refer to as objective soup recipes. These frameworks use MOO at different optimization levels to mitigate potential conflicts among all objectives. We conduct an extensive investigation using the LibriSpeech and AISHELL v1 datasets for ASR, along with the CoVoST v2 dataset for both ASR and speech-to-text translation tasks, to determine the highly conflicting objectives and the optimal training recipes among these three MOO training algorithms.", "keywords": [ "multilingual speech recognition", "speech-to-text translation", "multi-objective optimization", "multi-task learning", "semi-supervised training" ], "rating_list": [ 5, 10, 5, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "gVw9gFgAXh", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Bangzhen Liu", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Dan Cao", "gender": "unknown", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Shengfeng He", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Associate Professor" }, { "name": "Shiting Zhang", "gender": "Female", "institution": "South China University of Technology", "country": "", "position": "Undergrad student" }, { "name": "Sun Fan", "gender": "unknown", "institution": "University of Massachusetts at Amherst", "country": "", "position": "Undergrad student" }, { "name": "Tingrui Shen", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Weifeng Pan", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Zhirun Fan", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 21, "n_ref": 25, "n_ref_all": 41, "n_fig": 9, "n_tab": 5, "L_tab": 2374, "n_element_tab": 139, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1361, "n_element_tab_1": 89, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 118100, "len_all_1": 60521, "len_abs": 1373, "len_title": 129, "len_sents": 32589, "len_sents_1": 29284, "n_sents": 241, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1378, "title": "Language-driven 3D Human Pose Estimation: Grounding Motion from Text Descriptions", "abs": "In an NBA game scenario, consider the challenge of locating and analyzing the 3D poses of players performing a user-specified action, such as attempting a shot. Traditional 3D human pose estimation (3DHPE) methods often fall short in such complex, multi-person scenes due to their lack of semantic integration and reliance on isolated pose data. To address these limitations, we introduce Language-Driven 3D Human Pose Estimation (L3DHPE), a novel approach that extends 3DHPE to general multi-person contexts by incorporating detailed language descriptions. We present Panoptic-L3D, the first dataset designed for L3DHPE, featuring 3,838 linguistic annotations for 1,476 individuals across 588 videos, with 6,035 masks and 91k frame-level 3D skeleton annotations. Additionally, we propose Cascaded Pose Perception (CPP), a benchmarking method that simultaneously performs language-driven mask segmentation and 3D pose estimation within a unified model. CPP first learns 2D pose information, utilizes a body fusion module to aid in mask segmentation, and employs a mask fusion module to mitigate mask noise before outputting 3D poses. Our extensive evaluation of CPP and existing benchmarks on the Panoptic-L3D dataset demonstrates the necessity of this novel task and dataset for advancing 3DHPE. Our dataset can be accessed at https://languagedriven3dposeestimation.github.io/.", "keywords": [ "language-driven 3D human pose estimation", "text-motion interaction" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 3, 1, 3 ], "presentation_list": [ 4, 2, 2 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "gVnJFY8nCM", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Catherine Weaver", "gender": "Female", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Chen Tang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Postdoc" }, { "name": "Chenran Li", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Kenta Kawamoto", "gender": "Male", "institution": "Sony Research Inc.", "country": "", "position": "Researcher" }, { "name": "Pengcheng Wang", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Wei Zhan", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Masayoshi Tomizuka", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" } ], "n_formula": 43, "n_formula_1": 15, "n_ref_uni": 39, "n_ref": 61, "n_ref_all": 91, "n_fig": 14, "n_tab": 10, "L_tab": 4016, "n_element_tab": 392, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 332, "n_element_tab_1": 26, "formula_len_all": 4240, "formula_len_all_1": 947, "len_all": 180263, "len_all_1": 57425, "len_abs": 1452, "len_title": 113, "len_sents": 56097, "len_sents_1": 27585, "n_sents": 455, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1457, "title": "Residual-MPPI: Online Policy Customization for Continuous Control", "abs": "Policies developed through Reinforcement Learning (RL) and Imitation Learning (IL) have shown great potential in continuous control tasks, but real-world applications often require adapting trained policies to unforeseen requirements. While fine-tuning can address such needs, it typically requires additional data and access to the original training metrics and parameters.\nIn contrast, an online planning algorithm, if capable of meeting the additional requirements, can eliminate the necessity for extensive training phases and customize the policy without knowledge of the original training scheme or task. In this work, we propose a generic online planning algorithm for customizing continuous-control policies at the execution time, which we call Residual-MPPI. It can customize a given prior policy on new performance metrics in few-shot and even zero-shot online settings, given access to the prior action distribution alone. Through our experiments, we demonstrate that the proposed Residual-MPPI algorithm can accomplish the few-shot/zero-shot online policy customization task effectively, including customizing the champion-level racing agent, Gran Turismo Sophy (GT Sophy) 1.0, in the challenging car racing scenario, Gran Turismo Sport (GTS) environment. Code for MuJoCo experiments is included in the supplementary and will be open-sourced upon acceptance. Demo videos are available on our website: https://sites.google.com/view/residual-mppi.", "keywords": [ "Policy customization", "Combination of learning- and planning-based approaches", "Model predictive control" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "gVkX9QMBO3", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Emmanuel Bengio", "gender": "Male", "institution": "Recursion", "country": "", "position": "Researcher" }, { "name": "Ihor Neporozhnii", "gender": "unknown", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Jason Hartford", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "Assistant Professor" }, { "name": "Julien Roy", "gender": "Male", "institution": "Valence Labs powered by recursion", "country": "CA", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 8, "n_ref_uni": 46, "n_ref": 53, "n_ref_all": 93, "n_fig": 14, "n_tab": 2, "L_tab": 719, "n_element_tab": 111, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2312, "formula_len_all_1": 524, "len_all": 163242, "len_all_1": 65360, "len_abs": 1429, "len_title": 114, "len_sents": 52868, "len_sents_1": 32006, "n_sents": 350, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1432, "title": "Efficient Biological Data Acquisition through Inference Set Design", "abs": "In drug discovery, highly automated high-throughput laboratories are used to screen a large number of compounds in search of effective drugs. These experiments are expensive, so we might hope to reduce their cost by experimenting on a subset of the compounds, and predicting the outcomes of the remaining experiments. In this work, we model this problem as a sequential subset selection problem: we aim to sequentially select the smallest set of candidates in order to achieve some desired level of accuracy for the system as a whole. Our key observation is that, if there is heterogeneity in the difficulty of the prediction problem across the input space, selectively obtaining the labels for the hardest examples in the acquisition pool will leave only the relatively easy examples to remain in the inference set, leading to better overall system performance. We call this problem inference set design, and propose an active learning solution using the model's confidence to prune out these challenging examples. Our algorithm includes an explicit stopping criterion that stops running the experiments when it is sufficiently confident that the system has reached the target performance. Our empirical studies on images and molecular datasets, as well as a real-world case, show that deploying active learning for inference set design leads to significant reduction in experimental cost while obtaining better system performance.", "keywords": [ "Active Learning", "Data Acquisition", "ML for Drug Discovery" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 1, 2, 3, 4 ], "presentation_list": [ 1, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "gVbPYihQag", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Christofer Bester", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Postdoc" }, { "name": "Dongting Hu", "gender": "Male", "institution": "The University of Melbourne", "country": "AU", "position": "PhD student" }, { "name": "James Bailey", "gender": "unknown", "institution": "The University of Melbourne", "country": "AU", "position": "Full Professor" }, { "name": "Stephen O'Leary", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Full Professor" }, { "name": "Sudanthi Wijewickrema", "gender": "Female", "institution": "University of Melbourne", "country": "AU", "position": "Principal Research Fellow" }, { "name": "Yuansan Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Postdoc" } ], "n_formula": 26, "n_formula_1": 21, "n_ref_uni": 29, "n_ref": 65, "n_ref_all": 78, "n_fig": 6, "n_tab": 7, "L_tab": 1259, "n_element_tab": 93, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 659, "n_element_tab_1": 35, "formula_len_all": 1761, "formula_len_all_1": 1562, "len_all": 121155, "len_all_1": 63511, "len_abs": 1378, "len_title": 132, "len_sents": 39127, "len_sents_1": 27453, "n_sents": 309, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1388, "title": "Stochastic Diffusion: A Diffusion Based Model for Stochastic Time Series Forecasting", "abs": "Recent successes in diffusion probabilistic models have demonstrated their strength in modelling and generating different types of data, paving the way for their application in generative time series forecasting. However, most existing diffusion based approaches rely on sequential models and unimodal latent variables to capture global dependencies and model entire observable data, resulting in difficulties when it comes to highly stochastic time series data. In this paper, we propose a novel **Stoch**astic **Diff**usion (StochDiff) model that integrates the diffusion process into time series modelling stage and utilizes the representational power of the stochastic latent spaces to capture the variability of the stochastic time series data. Specifically, the model applies diffusion module at each time step within the sequential framework and learns a step-wise, data-driven prior for generative diffusion process. These features enable the model to effectively capture complex temporal dynamics and the multi-modal nature of the highly stochastic time series data. Through extensive experiments on real-world datasets, we demonstrate the effectiveness of our proposed model for probabilistic time series forecasting, particularly in scenarios with high stochasticity. Additionally, with a real-world surgical use case, we highlight the model's potential in medical application.", "keywords": [ "diffusion probabilistic model", "stochastic time series forecasting", "data-driven prior" ], "rating_list": [ 3, 6, 6, 5 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "gVWEq7LITG", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Hao Su", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "James Hou", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Liangzhi Shi", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Linghao Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Lingqi Zeng", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Undergrad student" }, { "name": "Yulin Liu", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Zhiao Huang", "gender": "Male", "institution": "University of California, San Diego, University of California, San Diego", "country": "", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 56, "n_ref": 92, "n_ref_all": 104, "n_fig": 9, "n_tab": 3, "L_tab": 255, "n_element_tab": 51, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 75, "n_element_tab_1": 12, "formula_len_all": 833, "formula_len_all_1": 722, "len_all": 150828, "len_all_1": 55958, "len_abs": 1325, "len_title": 139, "len_sents": 35433, "len_sents_1": 26715, "n_sents": 251, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1327, "title": "TF-HOT: Training-Free Hand-Object Pose Tracking and Optimization for Dexterous Manipulation", "abs": "Robotic manipulation with dexterous hands is inherently challenging due to their high-dimensional action spaces and the lack of large-scale, high-quality demonstrations. While there are many videos involving interactions between human hands and objects, the frequent, dynamic occlusions between human hands and objects complicate the accurate and robust tracking of hand and object poses, making it challenging to convert these interactions into high-quality dexterous robotic demonstrations.\nTo address these challenges, we introduce a novel Training-Free Hand-Object pose tracking pipeline (TF-HOT) that leverages differentiable rendering and rich priors from pre-trained 2D foundation perception models to perform optimization of human hand and object pose trajectories from input videos. Our method is efficient, allowing us to convert an in-the-wild video to pose trajectories in 1 minute, and we demonstrate state-of-the-art performance of our method over in-the-wild videos. Finally, we illustrate an application of our method in imitation learning by training policies to follow the pose trajectories extracted from TF-HOT, allowing us to learn dexterous manipulation policies that significantly outperform reinforcement learning and imitation learning methods that do not utilize hand-object pose trajectory following.", "keywords": [ "Pose estimation", "Robotics manipulation" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "gVVoZtiQlt", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Hang Zhang", "gender": "Male", "institution": "Baidu", "country": "", "position": "Researcher" }, { "name": "ping li", "gender": "Male", "institution": "Rutgers University", "country": "", "position": "Associate Professor" } ], "n_formula": 44, "n_formula_1": 44, "n_ref_uni": 33, "n_ref": 75, "n_ref_all": 92, "n_fig": 2, "n_tab": 2, "L_tab": 244, "n_element_tab": 37, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 325, "n_element_tab_1": 38, "formula_len_all": 2556, "formula_len_all_1": 2559, "len_all": 101772, "len_all_1": 73853, "len_abs": 660, "len_title": 102, "len_sents": 29548, "len_sents_1": 29810, "n_sents": 251, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1896, "title": "The Phase Transition Phenomenon of Shuffled Regression", "abs": "We study the phase transition \nphenomenon inherent in the shuffled (permuted) regression problem, which has found numerous applications in databases, privacy, data analysis, etc. For the permuted regression task: $\\mathbf{Y} = \\mathbf{\\Pi}\\mathbf{X}\\mathbf{B}$, the goal is to recover the permutation matrix $\\mathbf{\\Pi}$ as well as the coefficient matrix $\\mathbf{B}$. It has been empirically observed in prior studies that when recovering $\\mathbf{\\Pi}$, there exists a phase transition phenomenon: the error rate drops to zero rapidly once the parameters reach certain thresholds. In this study, we aim to precisely identify the locations of the phase transition points by leveraging techniques from {\\em message passing} (MP). \n\nIn our analysis, we first transform the permutation recovery problem into a probabilistic graphical model. Then, we leverage the analytical tools rooted in the message passing (MP) algorithm and derive an equation to track the convergence of the MP algorithm. By linking this equation to the branching random walk process, we are able to characterize the impact of the \\emph{signal-to-noise-ratio} ($\\mathsf{snr}$) on the permutation recovery. Depending on whether the signal is given or not, we separately investigate the oracle case and the non-oracle case. The bottleneck in identifying the phase transition regimes lies in deriving closed-form formulas for the corresponding critical points, but only in rare scenarios can one obtain such precise expressions. To tackle this challenge, we propose the Gaussian approximation method, which allows us to obtain the closed-form formulas in almost all scenarios. In the oracle case, our method can fairly accurately predict the phase transition $\\mathsf{snr}$. In the non-oracle case, our proposed algorithm can predict the maximum allowed number of permuted rows and uncover its dependency on the sample number.", "keywords": [ "Message Passing", "Permuted Linear Regression", "Phase Transition" ], "rating_list": [ 3, 6, 8, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 4, 2, 1, 2 ] }, { "paper_id": "gV0Moskp7k", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Fuli Qiao", "gender": "Female", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Mehrdad Mahdavi", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 27, "n_formula_1": 18, "n_ref_uni": 44, "n_ref": 74, "n_ref_all": 85, "n_fig": 6, "n_tab": 14, "L_tab": 3477, "n_element_tab": 276, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1076, "n_element_tab_1": 92, "formula_len_all": 2053, "formula_len_all_1": 1349, "len_all": 164693, "len_all_1": 66890, "len_abs": 1103, "len_title": 157, "len_sents": 43068, "len_sents_1": 29985, "n_sents": 297, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1106, "title": "Combating the Generalization-Forgetting Trade-off in Continual Learning: A Cautious Passive Low-Rank Approach", "abs": "Large Language Models (LLMs) have shown remarkable capabilities through wide-scale pre-training on a wide range of domains. However, they often suffer from catastrophic forgetting when learning sequential tasks. In this paper, we propose a novel parameter-efficient approach for continual learning in LLMs, which empirically explores the role of different effective layerwise ranks, leveraging lower ranks to mitigate catastrophic forgetting of previous tasks and higher ranks to enhance generalization on new tasks. By employing a subspace similarity metric that evaluates the orthogonality of low-rank subspaces between tasks, we gradually increase the rank of layerwise matrices for each new task, minimizing interference with previously learned tasks while enhancing generalization. Experimental results on standard continual learning benchmarks and challenging math benchmarks demonstrate that our method outperforms existing state-of-the-art approaches, effectively mitigating forgetting, improving task performance, and maintaining strong generalization to unseen tasks in a memory-efficient manner.", "keywords": [ "continual learning", "LLMs" ], "rating_list": [ 6, 3, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 5, 5 ] }, { "paper_id": "gU4ZgQNsOC", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Daouda Sow", "gender": "Male", "institution": "Ohio State University", "country": "US", "position": "PhD student" }, { "name": "Hans Arno Jacobsen", "gender": "Not Specified", "institution": "University of Toronto", "country": "CA", "position": "Full Professor" }, { "name": "Herbert Woisetschl\u00e4ger", "gender": "Male", "institution": "Technical University of Munich", "country": "DE", "position": "PhD student" }, { "name": "Saikiran Bulusu", "gender": "Male", "institution": "Syracuse University", "country": "", "position": "PhD student" }, { "name": "Shiqiang Wang", "gender": "Male", "institution": "IBM, International Business Machines", "country": "US", "position": "Research Staff Member" }, { "name": "Yingbin Liang", "gender": "Female", "institution": "The Ohio State University", "country": "US", "position": "Professor" } ], "n_formula": 73, "n_formula_1": 19, "n_ref_uni": 37, "n_ref": 62, "n_ref_all": 109, "n_fig": 8, "n_tab": 15, "L_tab": 3029, "n_element_tab": 527, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 15, "n_element_tab_1": 2, "formula_len_all": 6872, "formula_len_all_1": 1077, "len_all": 196292, "len_all_1": 67805, "len_abs": 1582, "len_title": 131, "len_sents": 54928, "len_sents_1": 32128, "n_sents": 438, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1646, "title": "Dynamic Loss-Based Sample Reweighting for Improved Large Language Model Pretraining", "abs": "Pretraining large language models (LLMs) on vast and heterogeneous datasets is crucial for achieving state-of-the-art performance across diverse downstream tasks. However, current training paradigms treat all samples equally, overlooking the importance or relevance of individual samples throughout the training process. Existing reweighting strategies, which primarily focus on group-level data importance, fail to leverage fine-grained instance-level information and do not adapt dynamically to individual sample importance as training progresses. In this paper, we introduce novel algorithms for dynamic, instance-level data reweighting aimed at improving both the efficiency and effectiveness of LLM pretraining. Our methods adjust the weight of each training sample based on its loss value in an online fashion, allowing the model to dynamically focus on more informative or important samples at the current training stage. In particular, our framework allows us to systematically devise reweighting strategies deprioritizing redundant or uninformative data, which we find tend to work best. \nFurthermore, we develop a new theoretical framework for analyzing the impact of loss-based reweighting on the convergence of gradient-based optimization, providing the first formal characterization of how these strategies affect convergence bounds. We empirically validate our approach across a spectrum of tasks, from pretraining 7B and 1.4B parameter LLMs to smaller-scale language models and linear regression problems, demonstrating that our loss-based reweighting approach can lead to faster convergence and significantly improved performance.", "keywords": [ "sample reweighing", "large language models", "pretraining" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 4, 2, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "gTwRMU3lJ5", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Jian Liang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Ran He", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Tieniu Tan", "gender": "Male", "institution": " Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Professor" }, { "name": "Zhengbo Wang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zilei Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 67, "n_formula_1": 18, "n_ref_uni": 47, "n_ref": 116, "n_ref_all": 146, "n_fig": 3, "n_tab": 9, "L_tab": 2614, "n_element_tab": 151, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1484, "n_element_tab_1": 82, "formula_len_all": 5312, "formula_len_all_1": 1103, "len_all": 172843, "len_all_1": 61613, "len_abs": 1705, "len_title": 99, "len_sents": 44190, "len_sents_1": 25340, "n_sents": 403, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 51, "L_abs": 1423, "title": "LoRA-Pro: Are Low-Rank Adapters Properly Optimized?", "abs": "Low-rank adaptation, also known as LoRA, has emerged as a prominent method for parameter-efficient fine-tuning of foundation models.\nDespite its computational efficiency, LoRA still yields inferior performance compared to full fine-tuning.\nIn this paper, we first uncover a fundamental connection between the optimization processes of LoRA and full fine-tuning: using LoRA for optimization is mathematically equivalent to full fine-tuning using a low-rank gradient for parameter updates.\nAnd this low-rank gradient can be expressed in terms of the gradients of the two low-rank matrices in LoRA.\nLeveraging this insight, we introduce LoRA-Pro, a method that enhances LoRA's performance by strategically adjusting the gradients of these low-rank matrices.\nThis adjustment allows the low-rank gradient to more accurately approximate the full fine-tuning gradient, thereby narrowing the performance gap between LoRA and full fine-tuning.\nFurthermore, we theoretically derive the optimal solutions for adjusting the gradients of the low-rank matrices, applying them during fine-tuning in LoRA-Pro.\nWe conduct extensive experiments across natural language understanding, dialogue generation, mathematical reasoning, code generation, and image classification tasks, demonstrating that LoRA-Pro substantially improves LoRA's performance, effectively narrowing the gap with full fine-tuning.\nCode is available in the supplementary.", "keywords": [ "Parameter Efficient Fine-Tuning", "Large Language Models", "Low-Rank Adaptation" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 4, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "gTsLBDMZrL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Daiyuyang", "gender": "unknown", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Long Tian", "gender": "Male", "institution": "Xi'an University of Software Engineering Institute", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiyang Liu", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Yufei Li", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "Postdoc" }, { "name": "shihaonan", "gender": "unknown", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 22, "n_ref": 36, "n_ref_all": 51, "n_fig": 6, "n_tab": 7, "L_tab": 4569, "n_element_tab": 302, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 778, "n_element_tab_1": 40, "formula_len_all": 841, "formula_len_all_1": 822, "len_all": 98976, "len_all_1": 52360, "len_abs": 1325, "len_title": 132, "len_sents": 22710, "len_sents_1": 22076, "n_sents": 184, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1333, "title": "A Prototype-oriented Fast Refinement Model for Few-shot Industrial Anomaly Detection", "abs": "Industrial Anomaly Detection (IAD) in low data regime is crucial for automating industrial inspections in practice. Previous methods have primarily focused on obtaining robust prototypes using only a few normal images per product. However, these methods seldom account for transferring the characteristics of online query images to enhance the representativeness of the original prototypes in a systematic way. To address the pivot issue, we propose a fast prototype-oriented refinement model for few-shot IAD. Given online query images, we formulate prototype refinement as a nested optimization problem between transport probability for anomaly suppression and transform matrix for characteristic transfer. Then we present an Expectation Maximization (EM)-based algorithm to iteratively compute the transport probability and transform matrix. In the E-step, we use entropy-based optimal transport, known as the Sinkhorn algorithm, to learn the transport probability. In the M-step, the transform matrix is updated via gradient descent. Finally, we integrate our model with two popular and recently proposed few-shot IAD methods, PatchCore and WinCLIP. Comprehensive experiments on three widely used datasets including MVTec, ViSA, and MPDD verify the effectiveness and efficiency of our proposed model in few-shot IAD applications.", "keywords": [ "Industrial Anomaly Detection", "Few-shot Learning", "Optimal Transport" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "gStRS8L28g", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Bohan Wang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Research Assistant" }, { "name": "Lu Lin", "gender": "Female", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Yurui Chang", "gender": "Female", "institution": "Pennsylvania State University", "country": "", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 16, "n_ref_uni": 41, "n_ref": 102, "n_ref_all": 124, "n_fig": 4, "n_tab": 7, "L_tab": 4288, "n_element_tab": 190, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1393, "n_element_tab_1": 132, "formula_len_all": 2340, "formula_len_all_1": 1440, "len_all": 144620, "len_all_1": 65190, "len_abs": 1279, "len_title": 129, "len_sents": 44544, "len_sents_1": 29136, "n_sents": 304, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1285, "title": "Mitigating Graph Covariate Shift via Score-based Out-of-distribution Augmentation", "abs": "Distribution shifts between training and testing datasets significantly impair the model performance on graph learning. A commonly-taken causal view in graph invariant learning suggests that stable features of graphs are causally associated with labels, whereas unstable environmental features lead to distribution shifts. In particular, covariate shifts caused by unseen environmental features in test graphs underscore the critical need for out-of-distribution (OOD) generalization. Existing graph augmentation methods designed to address the covariate shift often disentangle the stable and environmental features in the input space, and selectively perturb or mixup the environmental features. However, such perturbation-based methods heavily rely on an accurate separation of stable and environmental features, and their exploration ability is confined to existing environmental features in the training distribution. To overcome these limitations, we introduce a novel approach using score-based graph generation strategies that synthesize unseen environmental features while preserving the validity and stable features of overall graph patterns. Our comprehensive empirical evaluations demonstrate the enhanced effectiveness of our method in improving graph OOD generalization.", "keywords": [ "Graph Neural Network", "Graph Data Augmentation", "Distribution Shift" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "gSO9fYLPSw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Jia Li", "gender": "Female", "institution": "University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Jiahang Cao", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jingkai SUN", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Junhao He", "gender": "Male", "institution": "the Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Intern" }, { "name": "Mingyuan Sun", "gender": "Male", "institution": "Northeastern University", "country": "CN", "position": "MS student" }, { "name": "Renjing Xu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Yi Gu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Ziyi Zhang", "gender": "Female", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Intern" }, { "name": "WANG Jiaxu", "gender": "Male", "institution": "Hong Kong University of Science and Technology (GZ)", "country": "CN", "position": "PhD student" }, { "name": "QIANG ZHANG", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 14, "n_ref_uni": 24, "n_ref": 46, "n_ref_all": 63, "n_fig": 7, "n_tab": 5, "L_tab": 1793, "n_element_tab": 197, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1214, "n_element_tab_1": 134, "formula_len_all": 835, "formula_len_all_1": 726, "len_all": 140660, "len_all_1": 63844, "len_abs": 1671, "len_title": 124, "len_sents": 42299, "len_sents_1": 28576, "n_sents": 343, "n_sents_1": 241, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1672, "title": "DEGS: Deformable Event-based 3D Gaussian Splatting from RGB and Event Stream", "abs": "Reconstructing Dynamic 3D Gaussians Splatting (3DGS) from low-framerate RGB videos is challenging. This is because large inter-frame motions will increase the uncertainty of the solution space. For example, one pixel in the first frame might have more choices to reach the corresponding pixel in the second frame. Event cameras can provide super-fast visual change acquisition asynchronously while not containing color information. Intuitively, the event stream can provide deterministic constraints for the inter-frame large motion by the event trajectories. Hence, combining low-temporal resolution images with high-framerate event streams can address this challenge. \nHowever, the data format of the two modalities is very different, and currently, no methods directly optimize dynamic 3DGS from events and RGB images. This paper introduces a novel framework that jointly optimizes dynamic 3DGS from the two modalities. The key idea is to adopt event motion priors to guide the optimization of the deformation fields. First, we extract the motion priors encoded in event streams using the proposed LoCM unsupervised fine-tuning framework to adapt an event flow estimator to a certain unseen scene. Then, we present the geometry-aware data association method to build the event-Gaussian motion correspondence, which is the primary foundation of the pipeline, accompanied by two useful strategies: motion decomposition and inter-frame pseudo-label. Extensive experiments show that our method outperforms existing image and event-based approaches across synthetic and real scenes and prove that our method can effectively optimize dynamic 3DGS with the help of event data.", "keywords": [ "Dynamic 3D Reconstruction; Bio-inspired Vision; Event-based Camera" ], "rating_list": [ 6, 3, 5, 1 ], "soundness_list": [ 4, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "gSGRSxVcRP", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Bastian Rieck", "gender": "Male", "institution": "University of Fribourg", "country": "CH", "position": "Full Professor" }, { "name": "Emanuele Rodola", "gender": "Male", "institution": "Sapienza University of Rome", "country": "IT", "position": "Full Professor" }, { "name": "Irene Cannistraci", "gender": "Female", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 35, "n_ref": 57, "n_ref_all": 94, "n_fig": 9, "n_tab": 9, "L_tab": 2194, "n_element_tab": 55, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1061, "n_element_tab_1": 106, "formula_len_all": 565, "formula_len_all_1": 404, "len_all": 164517, "len_all_1": 68663, "len_abs": 2252, "len_title": 125, "len_sents": 63100, "len_sents_1": 34638, "n_sents": 366, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1272, "title": "Detecting and Approximating Redundant Computational Blocks in Neural Networks", "abs": "Deep neural networks often learn similar internal representations, both across different models and within their own layers. While inter-network similarities have enabled techniques such as model stitching and merging, intra-network similarities present new opportunities for designing more efficient architectures. In this paper, we investigate the emergence of these internal similarities across different layers in diverse neural architectures, showing that similarity patterns emerge independently of the datataset used. We introduce a simple metric, Block Redundancy (BR), to detect redundant blocks, providing a foundation for future architectural optimization methods. Building on this, we propose Redundant Blocks Approximation (RBA), a general framework that identifies and approximates one or more redundant computational blocks using simpler transformations. We show that the transformation $\\mathcal{T}$ between two representations can be efficiently computed in closed-form, and it is enough to replace the redundant blocks from the network. RBA reduces model parameters and time complexity while maintaining good performance. We validate our method on classification tasks in the vision domain, using a variety of pretrained foundational models and datasets.", "keywords": [ "latent representations", "representation learning", "neural network similarities", "classification", "foundation models", "large models" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "gRuZkEy49k", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Adamo Young", "gender": "Not Specified", "institution": "University of Toronto", "country": "", "position": "PhD student" }, { "name": "Kanika Madan", "gender": "unknown", "institution": "MILA, University of Montreal", "country": "", "position": "PhD student" }, { "name": "Amir Hosein Khasahmadi", "gender": "unknown", "institution": "Autodesk", "country": "CA", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 15, "n_ref_uni": 23, "n_ref": 68, "n_ref_all": 95, "n_fig": 13, "n_tab": 5, "L_tab": 1395, "n_element_tab": 107, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1479, "n_element_tab_1": 66, "formula_len_all": 2247, "formula_len_all_1": 854, "len_all": 129903, "len_all_1": 65413, "len_abs": 1180, "len_title": 118, "len_sents": 41262, "len_sents_1": 29063, "n_sents": 319, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1185, "title": "Adapting Monte Carlo Tree Search for Generative Flow Network Training", "abs": "Generative Flow Networks, or GFlowNets, formulate generative modelling in discrete spaces as a sequential decision-making problem. Sampling plays a key role in GFlowNet training, as most algorithms use the learned policy to sample trajectories from the environment. Monte-Carlo Tree Search (MCTS) is a planning algorithm that has successfully been applied to train sequential decision-making models with reinforcement learning (RL). In this work, we leverage known connections between GFlowNets and maximum-entropy RL to adapt MCTS for GFlowNet training. We prove that standard MCTS tree construction processes can be modified to calculate the optimal flows for a GFlowNet, given sufficient samples from the environment. Our results extend to multiple cases of GFN modelling, including terminating-energy and intermediate-energy environments. We investigate practical strategies for employing MCTS as a sampling tool and apply it to different GFN parameterizations and training objectives. Through extensive experiments in a variety of discrete domains, including a language-based reasoning task, we show that our proposed method offers an improvement over standard on-policy sampling.", "keywords": [ "Generative Models", "Generative Flow Networks" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "gRrjx9GzRP", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Bastian J\u00e4ckl", "gender": "Male", "institution": "Universit\u00e4t Konstanz", "country": "DE", "position": "PhD student" }, { "name": "Daniel A. Keim", "gender": "Male", "institution": "Universit\u00e4t Konstanz", "country": "DE", "position": "Full Professor" }, { "name": "Maximilian T. Fischer", "gender": "Male", "institution": "Universit\u00e4t Konstanz", "country": "DE", "position": "Postdoc" }, { "name": "Udo Schlegel", "gender": "unknown", "institution": "University of Konstanz", "country": "", "position": "PhD student" }, { "name": "Yannick Metz", "gender": "Male", "institution": "Universit\u00e4t Konstanz", "country": "DE", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 124, "n_ref_all": 158, "n_fig": 8, "n_tab": 10, "L_tab": 3328, "n_element_tab": 174, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1816, "n_element_tab_1": 91, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 190291, "len_all_1": 68396, "len_abs": 1845, "len_title": 128, "len_sents": 56305, "len_sents_1": 32810, "n_sents": 435, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1293, "title": "Leveraging Color Channel Independence for Improved Unsupervised Object Detection", "abs": "Object-centric architectures can learn to extract distinct object representations from visual scenes, enabling downstream applications on the object level. \nSimilarly to autoencoder-based image models, object-centric approaches have been trained on the unsupervised reconstruction loss of images encoded by RGB color spaces. \nIn our work, we challenge the common assumption that RGB images are the optimal target for unsupervised learning in computer vision.\nWe discuss conceptually and empirically that other color spaces, such as HSV, bear essential characteristics for object-centric representation learning, like robustness to lighting conditions. We further show that models improve when requiring them to predict additional color channels.\nSpecifically, we propose the RGB-S space, which extends RGB with HSV's saturation component and leads to markedly better reconstruction and disentanglement for five common evaluation datasets.\nThe use of composite color spaces can be implemented with basically no computational overhead, is agnostic of the models' architecture, and is universally applicable across a wide range of visual computing tasks and training types. \nThe findings of our approach encourage additional investigations in computer vision tasks beyond object-centric learning.", "keywords": [ "object detection", "visual reasoning", "segmentation", "unsupervised learning", "object centric representation learning", "computer vision" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "gRmWtOnTLK", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Dongyang Dai", "gender": "Male", "institution": "Researcher in Generative Models", "country": "CN", "position": "Researcher" }, { "name": "Peng Liu", "gender": "Unspecified", "institution": "Transsion Holdings", "country": "CN", "position": "Researcher" }, { "name": "Zhiyong Wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 51, "n_ref": 102, "n_ref_all": 125, "n_fig": 4, "n_tab": 11, "L_tab": 4577, "n_element_tab": 587, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1803, "n_element_tab_1": 172, "formula_len_all": 608, "formula_len_all_1": 561, "len_all": 195175, "len_all_1": 70640, "len_abs": 1023, "len_title": 115, "len_sents": 49195, "len_sents_1": 31978, "n_sents": 416, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1106, "title": "RFWave: Multi-band Rectified Flow for Audio Waveform Reconstruction", "abs": "Recent advancements in generative modeling have significantly enhanced the reconstruction of audio waveforms from various representations. While diffusion models are adept at this task, they are hindered by latency issues due to their operation at the individual sample point level and the need for numerous sampling steps. In this study, we introduce RFWave, a cutting-edge multi-band Rectified Flow approach designed to reconstruct high-fidelity audio waveforms from Mel-spectrograms or discrete acoustic tokens. RFWave uniquely generates complex spectrograms and operates at the frame level, processing all subbands simultaneously to boost efficiency. Leveraging Rectified Flow, which targets a straight transport trajectory, RFWave achieves reconstruction with just 10 sampling steps. Our empirical evaluations show that RFWave not only provides outstanding reconstruction quality but also offers vastly superior computational efficiency, enabling audio generation at speeds up to 160 times faster than real-time on a GPU. An online demonstration is available at: https://rfwave-demo.github.io/rfwave/.", "keywords": [ "Rectified Flow", "Audio Waveform Reconstruction", "Multi-band audio generation\uff0cReal-time diffusion Vocoder" ], "rating_list": [ 8, 8, 5, 5, 8 ], "soundness_list": [ 3, 3, 2, 2, 4 ], "presentation_list": [ 3, 4, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "gRbWCGCFBz", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Dongdong Chen", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Lequan Yu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Lingting Zhu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Mengchen Liu", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Yunsheng Li", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Jin Zhenchao", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 1, "n_ref_uni": 16, "n_ref": 33, "n_ref_all": 56, "n_fig": 1, "n_tab": 18, "L_tab": 10693, "n_element_tab": 766, "n_fig_1": 1, "n_tab_1": 8, "L_tab_1": 2602, "n_element_tab_1": 277, "formula_len_all": 126, "formula_len_all_1": 32, "len_all": 211194, "len_all_1": 52028, "len_abs": 3515, "len_title": 128, "len_sents": 72963, "len_sents_1": 22899, "n_sents": 601, "n_sents_1": 136, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1789, "title": "ToolBridge: An Open-Source Dataset to Equip LLMs with External Tool Capabilities", "abs": "Through the integration of external tools, large language models (LLMs) such as GPT-4o and Llama 3.1 significantly expand their functional capabilities, evolving from elementary conversational agents to general-purpose assistants. We contend that the primary drivers of these advancements are the quality and diversity of the training data. However, the existing LLMs with external tool integration provide only limited transparency regarding their datasets and data collection approaches, which has led to the initiation of this study. Specifically, in this work, we endeavor to present a detailed exposition of the methodology for constructing datasets that facilitate LLMs in effectively learning how to utilize external tools and make this process available to the public through the introduction of ToolBridge. ToolBridge proposes to leverage a collection of general open-access datasets as its raw dataset pool and incorporates a series of strategies to identify the appropriate data entries for external tool API insertions. By supervised fine-tuning (SFT) on these curated data entries, LLMs can invoke external tools in appropriate contexts to boost their predictive accuracy, particularly for essential functions including factual retrieval, data processing and numerical computation. Our experiments meticulously isolate model architectures and training configurations, zeroing in exclusively on the role of data. The experimental results indicate that LLMs trained on ToolBridge exhibit consistent performance gains on both standard benchmarks and custom evaluation datasets. All associated code and data will be released as open source, promoting transparency and facilitating the broader community to explore methodologies for equipping LLMs with external tools capabilities.", "keywords": [ "LLM", "datasets", "tool use", "open source" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "gRXLa6LS3J", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Haomin Wen", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Leman Akoglu", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yuchen Shen", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" } ], "n_formula": 25, "n_formula_1": 5, "n_ref_uni": 62, "n_ref": 131, "n_ref_all": 177, "n_fig": 14, "n_tab": 16, "L_tab": 20991, "n_element_tab": 840, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 644, "n_element_tab_1": 62, "formula_len_all": 2786, "formula_len_all_1": 328, "len_all": 288229, "len_all_1": 70122, "len_abs": 1690, "len_title": 143, "len_sents": 85520, "len_sents_1": 35064, "n_sents": 566, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1695, "title": "Zero-shot Outlier Detection via Synthetically Pretrained Transformers: Model Selection Bygone!", "abs": "Outlier detection (OD) has a vast literature as it finds numerous applications in\nenvironmental monitoring, security, manufacturing, and finance to name a few.\nBeing an inherently unsupervised task, model selection is a key bottleneck for OD\n(both algorithm and hyperparameter selection) without label supervision. There is\na long list of techniques to choose from \u2013 both classical algorithms and deep neural\narchitectures \u2013 and while several studies report their hyperparameter sensitivity, the\nliterature remains quite slim on unsupervised model selection\u2014limiting the effective use of OD in practice. In this paper we present FoMo-0D, for zero/0-shot OD\nexploring a transformative new direction that bypasses the hurdle of model selection\naltogether (!), thus breaking new ground. The fundamental idea behind FoMo-0D is\nthe Prior-data Fitted Networks, recently introduced by M\u00fcller et al. (2022), which\ntrains a Transformer model on a large body of synthetically generated data from a\nprior data distribution. In essence, FoMo-0D is a pretrained Foundation Model\nfor zero/0-shot OD on tabular data, which can directly predict the (outlier/inlier)\nlabel of any test data at inference time, by merely a single forward pass\u2014making\nobsolete the need for choosing an algorithm/architecture and tuning its associated\nhyperparameters, besides requiring no training of model parameters when given a\nnew OD dataset. Extensive experiments on 57 public benchmark datasets against\n26 baseline methods show that FoMo-0D performs statistically no different from the\n2nd top baseline, while significantly outperforming the majority of the baselines,\nwith an average inference time of 7.7 ms per test sample.", "keywords": [ "zero-shot outlier detection", "prior-data fitted networks" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "gQoBw7sGAu", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Charles Dawson", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Chuchu Fan", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Max Z. Li", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" }, { "name": "Van Tran", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Undergrad student" } ], "n_formula": 21, "n_formula_1": 10, "n_ref_uni": 45, "n_ref": 76, "n_ref_all": 120, "n_fig": 16, "n_tab": 4, "L_tab": 1827, "n_element_tab": 151, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 6, "n_element_tab_1": 1, "formula_len_all": 1507, "formula_len_all_1": 539, "len_all": 154940, "len_all_1": 55721, "len_abs": 888, "len_title": 149, "len_sents": 47428, "len_sents_1": 26123, "n_sents": 322, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 101, "L_abs": 929, "title": "Rare event modeling with self-regularized normalizing flows: what can we learn from a single failure?", "abs": "Increased deployment of autonomous systems in fields like transportation and robotics have seen a corresponding increase in safety-critical failures. These failures can be difficult to model and debug due to the relative lack of data: compared to tens of thousands of examples from normal operations, we may have only seconds of data leading up to the failure. This scarcity makes it challenging to train generative models of rare failure events, as existing methods risk either overfitting to noise in the limited failure dataset or underfitting due to an overly strong prior. We address this challenge with CalNF, or calibrated normalizing flows, a self-regularized framework for posterior learning from limited data. CalNF achieves state-of-the-art performance on data-limited failure modeling and inverse problems and enables a first-of-a-kind case study into the root causes of the 2022 Southwest Airlines scheduling crisis.", "keywords": [ "rare event modeling", "normalizing flows", "Bayesian inverse problems" ], "rating_list": [ 5, 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 3, 2 ] }, { "paper_id": "gQlxd3Mtru", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Peijie Zhou", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tiejun Li", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Zhenyi Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 62, "n_formula_1": 32, "n_ref_uni": 50, "n_ref": 119, "n_ref_all": 191, "n_fig": 13, "n_tab": 11, "L_tab": 2641, "n_element_tab": 218, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 924, "n_element_tab_1": 69, "formula_len_all": 5862, "formula_len_all_1": 2603, "len_all": 239579, "len_all_1": 70743, "len_abs": 3077, "len_title": 142, "len_sents": 66829, "len_sents_1": 29695, "n_sents": 555, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 92, "L_abs": 1155, "title": "Learning stochastic dynamics from snapshots through regularized unbalanced optimal transport", "abs": "Reconstructing dynamics using samples from sparsely time-resolved snapshots is an important problem in both natural sciences and machine learning. Here, we introduce a new deep learning approach for solving regularized unbalanced optimal transport (RUOT) and inferring continuous unbalanced stochastic dynamics from observed snapshots. Based on the RUOT form, our method models these dynamics without requiring prior knowledge of growth and death processes or additional information, allowing them to be learned directly from data. Theoretically, we explore the connections between the RUOT and Schr\u00f6dinger bridge problem and discuss the key challenges and potential solutions. The effectiveness of our method is demonstrated with a synthetic gene regulatory network, high-dimensional Gaussian Mixture Model, and single-cell RNA-seq data from blood development. Compared with other methods, our approach accurately identifies growth and transition patterns, eliminates false transitions, and constructs the Waddington developmental landscape. Our code is available at: [https://github.com/zhenyiizhang/DeepRUOT](https://github.com/zhenyiizhang/DeepRUOT).", "keywords": [ "optimal transport", "Schr\u00f6dinger bridge", "trajectory inference", "single-cell" ], "rating_list": [ 8, 8, 10 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 4 ], "confidence_list": [ 5, 3, 2 ] }, { "paper_id": "gQPwP1JFwC", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Rafa\u0142 Wolniak", "gender": "Male", "institution": "Gdansk University of Technology", "country": "PL", "position": "PhD student" } ], "n_formula": 32, "n_formula_1": 9, "n_ref_uni": 23, "n_ref": 35, "n_ref_all": 104, "n_fig": 14, "n_tab": 6, "L_tab": 3600, "n_element_tab": 272, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1260, "n_element_tab_1": 127, "formula_len_all": 4547, "formula_len_all_1": 1002, "len_all": 144590, "len_all_1": 59182, "len_abs": 1579, "len_title": 150, "len_sents": 52812, "len_sents_1": 26283, "n_sents": 398, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1614, "title": "Efficient Gradient-Based Algorithm for Training Deep Learning Models With Many Nonlinear Activations", "abs": "This research paper presents a novel algorithm for training deep neural networks with many nonlinear layers (e.g., 30). The method is based on backpropagation of an approximated gradient, averaged over the range of a weight update. Unlike the gradient, the average gradient of a loss function is proven within this research to provide more accurate information on the change in loss caused by the associated parameter update of a model. Therefore, it may be utilized to improve learning. In our implementation, the efficiently approximated average gradient is paired with RMSProp and compared to the typical gradient-based approach. For the tested deep model with numerous stacked fully-connected layers featuring nonlinear activations on MNIST and Fashion MNIST, the presented algorithm: $\\quad$ (a) generalizes better, at least in a reasonable epoch count,$\\quad$ (b) in the case of optimal implementation, learning would require less computation time than the gradient-based RMSProp, with the memory requirement of the Adam optimizer,$\\quad$ (c) performs well on a broader range of learning rates, therefore it may bring time and energy savings from reduced hyperparameter searches,$\\quad$ (d) improves sample efficiency about three times according to median training losses. On the other hand, for a deep sequential convolutional model trained on the IMDB dataset, sample efficiency is improved by about 55%. However, in the case of the tested shallow model, the method performs approximately the same as the gradient-based RMSProp in terms of both training and test loss. The source code is provided at [...].", "keywords": [ "deep learning", "optimization", "deep learning theory", "deep neural network" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "gNxvs5pUdu", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Dimosthenis Karatzas", "gender": "Male", "institution": "Universitat Aut\u00f2noma de Barcelona", "country": "ES", "position": "Associate Professor" }, { "name": "Khanh Nguyen", "gender": "Male", "institution": "Computer Vision Center, Universitat Aut\u00f3noma de Barcelona", "country": "ES", "position": "PhD student" }, { "name": "Mario Fritz", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Full Professor" }, { "name": "Raouf Kerkouche", "gender": "Male", "institution": "CISPA, saarland university, saarland informatics campus", "country": "DE", "position": "Postdoc" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 31, "n_ref": 58, "n_ref_all": 113, "n_fig": 9, "n_tab": 13, "L_tab": 4663, "n_element_tab": 779, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 246, "formula_len_all_1": 153, "len_all": 200168, "len_all_1": 68749, "len_abs": 1355, "len_title": 122, "len_sents": 73011, "len_sents_1": 35701, "n_sents": 503, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1360, "title": "DocMIA: Document-Level Membership Inference Attacks against DocVQA Models", "abs": "Document Visual Question Answering (DocVQA) has introduced a new paradigm for end-to-end document understanding, and quickly became one of the standard benchmarks for multimodal LLMs. Automating document processing workflows, driven by DocVQA models, presents significant potential for many business sectors. However, documents tend to contain highly sensitive information, raising concerns about privacy risks associated with training such DocVQA models. One significant privacy vulnerability, exploited by the membership inference attack, is the possibility for an adversary to determine if a particular record was part of the model's training data. In this paper, we introduce two novel membership inference attacks tailored specifically to DocVQA models. These attacks are designed for two different adversarial scenarios: a white-box setting, where the attacker has full access to the model architecture and parameters, and a black-box setting, where only the model's outputs are available. Notably, our attacks assume the adversary lacks access to auxiliary datasets, which is more realistic in practice but also more challenging. Our unsupervised methods outperform existing state-of-the-art membership inference attacks across a variety of DocVQA models and datasets, demonstrating their effectiveness and highlighting the privacy risks in this domain.", "keywords": [ "Membership Inference Attacks", "Document-based VQA", "Multi-modal Models", "Privacy" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 5, 2, 3, 3, 3 ] }, { "paper_id": "gNrNlhhCN1", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Guojun Zhang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CA", "position": "Researcher" }, { "name": "Hongliang Li", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CA", "position": "Researcher" }, { "name": "Pascal Poupart", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Full Professor" }, { "name": "Zehao Zhang", "gender": "unknown", "institution": "University of Waterloo", "country": "", "position": "MS student" } ], "n_formula": 22, "n_formula_1": 3, "n_ref_uni": 27, "n_ref": 53, "n_ref_all": 65, "n_fig": 2, "n_tab": 12, "L_tab": 5590, "n_element_tab": 434, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 672, "n_element_tab_1": 40, "formula_len_all": 2312, "formula_len_all_1": 225, "len_all": 153950, "len_all_1": 68457, "len_abs": 3443, "len_title": 104, "len_sents": 54491, "len_sents_1": 34569, "n_sents": 413, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1204, "title": "FAACL: Federated Adaptive Asymmetric Clustered Learning", "abs": "Asymmetric clustering has remained an unexplored problem in Clustered Federated Learning (CFL), diverging from the traditional approach of forming independent, non-interacting clusters. Previous methodologies have been limited to either separating devices with different data quality into distinct clusters or merging all devices into a single cluster, both of which compromise either data utilization or model accuracy. We propose a new federated learning technique where some devices may contribute to the training of the models of other devices, but without enforcing reciprocity, leading to a form of asymmetric clustering. This is beneficial in a variety of situations including scenarios where it is desirable for a device with high quality data to help train the model of a device with low quality data, but not vice-versa. This method not only enhances data utilization across the devices, but also maintains the integrity of high-quality data. Through a rigorous theoretical analysis and empirical evaluations, we demonstrate that our approach can efficiently find high quality (asymmetric) clusterings for numerous devices, achieving competitive performance metrics on existing CFL benchmarks.", "keywords": [ "Federated Learning", "Distributed Learning", "Clustering" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 3, 2 ] }, { "paper_id": "gNoqEdT2wO", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "David Vazquez", "gender": "Male", "institution": "ServiceNow research", "country": "CA", "position": "Researcher" }, { "name": "Marco D'Alessandro", "gender": "Male", "institution": "Universidad P\u00fablica de Navarra", "country": "ES", "position": "Researcher" }, { "name": "Mikel Elkano", "gender": "Male", "institution": "Neuraptic AI", "country": "ES", "position": "Principal Researcher" }, { "name": "Enrique Hern\u00e1ndez", "gender": "Male", "institution": "Universidad P\u00fablica de Navarra", "country": "ES", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 46, "n_ref": 67, "n_ref_all": 70, "n_fig": 1, "n_tab": 4, "L_tab": 1033, "n_element_tab": 128, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 509, "n_element_tab_1": 42, "formula_len_all": 238, "formula_len_all_1": 157, "len_all": 120297, "len_all_1": 84486, "len_abs": 1116, "len_title": 122, "len_sents": 28332, "len_sents_1": 25461, "n_sents": 188, "n_sents_1": 160, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1120, "title": "A Multimodal Class-Incremental Learning benchmark for classification tasks", "abs": "Continual learning has made significant progress in addressing catastrophic forgetting in vision and language domains, yet the majority of research has treated these modalities separately. The exploration of multimodal continual learning remains sparse, with a few existing works focused on specific applications like VQA, text-to-vision retrieval, and incremental multi-tasking. These efforts lack a general benchmark to standardize the evaluation of models in multimodal continual learning settings. In this paper, we introduce a novel benchmark for Multimodal Class-Incremental Learning (MCIL), designed specifically for multimodal classification tasks. Our benchmark comprises a curated selection of multimodal datasets tailored to classification challenges. We further adapt a widely used Vision-Language model to multiple existing continual learning strategies, providing crucial insights into the behavior of vision-language models in incremental classification tasks. This work represents the first comprehensive framework for MCIL, establishing a foundation for future research in multimodal continual learning.", "keywords": [ "multimodal", "continual learning", "incremental learning", "benchmark", "vision", "language", "vision-language", "multimodal continual learning" ], "rating_list": [ 1, 3, 1, 3, 3 ], "soundness_list": [ 1, 2, 2, 2, 1 ], "presentation_list": [ 3, 2, 1, 1, 2 ], "contribution_list": [ 1, 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 5 ] }, { "paper_id": "gNWTngNX31", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "dongwei Pan", "gender": "Male", "institution": "Sensetime Research", "country": "", "position": "Researcher" }, { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Kwan-Yee Lin", "gender": "Female", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Postdoc" }, { "name": "Yang Li", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 50, "n_ref": 103, "n_ref_all": 133, "n_fig": 15, "n_tab": 6, "L_tab": 2845, "n_element_tab": 307, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 274, "n_element_tab_1": 6, "formula_len_all": 1195, "formula_len_all_1": 1059, "len_all": 206430, "len_all_1": 76690, "len_abs": 0, "len_title": 216, "len_sents": 65619, "len_sents_1": 36717, "n_sents": 498, "n_sents_1": 271, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 2013, "title": "TimeWalker: Personalized Neural Space for Life-long Head Avatar", "abs": "We present TimeWalker, a novel framework that models realistic, full-scale 3D head avatars of a person on lifelong scale. Unlike current human head avatar pipelines that capture a person's identity only at the momentary level (i.e., instant photography, or short videos), TimeWalker constructs a person's comprehensive identity from unstructured data collection over his/her various life stages, offering a paradigm to achieve full reconstruction and animation of that person at different moments of life. At the heart of TimeWalker's success is a novel neural parametric model that learns personalized representation with the disentanglement of shape, expression, and appearance across ages. Central to our methodology are the concepts of two aspects: (1) We track back to the principle of modeling a person's identity in an additive combination of his/her average head representation in the canonical space, and moment-specific head attribute representations driven from a set of neural head basis. To learn the set of head basis that could represent the comprehensive head variations of the target person in a compact manner, we propose a Dynamic Neural Basis-Blending Module (Dynamo). It dynamically adjusts the number and blend weights of neural head bases, according to both shared and specific traits of the target person over ages. (2) We introduce Dynamic 2D Gaussian Splatting (DNA-2DGS), an extension of Gaussian splatting representation, to model head motion deformations like facial expressions without losing the realism of rendering and reconstruction of full head. DNA-2DGS includes a set of controllable 2D oriented planar Gaussian disks that utilize the priors from a parametric morphable face model, and move/rotate with the change of expression. Through extensive experimental evaluations, we show TimeWalker's ability to reconstruct and animate avatars across decoupled dimensions with realistic rendering effects, demonstrating a way to achieve personalized ``time traveling'' in a breeze.", "keywords": [ "Life-long Personalized Representation;Neural Parametric Morphable Model;" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 4, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "gNOW7ch3Ye", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Fan Zhou", "gender": "Male", "institution": "AntGroup", "country": "CN", "position": "Researcher" }, { "name": "Hongyuan Mei", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "", "position": "Research Assistant Professor" }, { "name": "Qingyang Dai", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Siqiao Xue", "gender": "Male", "institution": "starguest.ai", "country": "CN", "position": "Researcher" }, { "name": "Tingting Chen", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zhixuan Chu", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 55, "n_ref_all": 68, "n_fig": 6, "n_tab": 9, "L_tab": 3374, "n_element_tab": 383, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 1170, "n_element_tab_1": 229, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 166318, "len_all_1": 64412, "len_abs": 2066, "len_title": 123, "len_sents": 41590, "len_sents_1": 29220, "n_sents": 314, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1412, "title": "FAMMA: A Benchmark for Financial Multilingual Multimodal Question Answering", "abs": "In this paper, we introduce FAMMA, an open-source benchmark for financial multilingual multimodal question answering (QA).\nOur benchmark aims to evaluate the abilities of multimodal large language models (MLLMs) in answering questions that require advanced financial knowledge and sophisticated reasoning. It includes 1,758 meticulously collected question-answer pairs from university textbooks and exams, spanning 8 major subfields in finance including corporate finance, asset management, and financial engineering. Some of the QA pairs are written in Chinese or French, while a majority of them are in English. These questions are presented in a mixed format combining text and heterogeneous image types, such as charts, tables, and diagrams. \nWe evaluate a range of state-of-the-art MLLMs on our benchmark, and our analysis shows that FAMMA poses a significant challenge for these models. Even advanced systems like GPT-4o and Claude-35-Sonnet achieve only 42\\% accuracy. Additionally, the open-source Qwen2-VL lags notably behind its proprietary counterparts. Lastly, we explore GPT o1-style reasoning chains to enhance the models' reasoning capabilities, which significantly improve error correction.\nOur FAMMA benchmark will facilitate future research to develop expert systems in financial QA. The code and data have been anonymously released at \\small \\url{https://github.com/random2024GO/bench-script}.", "keywords": [ "benchmark", "multimodal large language model", "financial question answering" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "gN4stDLq3t", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Aditya Prasad", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Adriana Meza Soria", "gender": "Female", "institution": "MIT-IBM Watson AI Lab", "country": "US", "position": "Researcher" }, { "name": "David Daniel Cox", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "IBM Director, MIT-IBM Watson AI Lab" }, { "name": "Gaoyuan Zhang", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Research engineer" }, { "name": "Matthew Stallone", "gender": "unknown", "institution": "MIT-IBM Watson AI Lab", "country": "US", "position": "Researcher" }, { "name": "Mayank Mishra", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Rameswar Panda", "gender": "Male", "institution": "MIT-IBM Watson AI Lab", "country": "", "position": "Research Scientist" }, { "name": "Shawn Tan", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Yikang Shen", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 19, "n_ref": 49, "n_ref_all": 65, "n_fig": 6, "n_tab": 8, "L_tab": 2980, "n_element_tab": 477, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1577, "n_element_tab_1": 142, "formula_len_all": 424, "formula_len_all_1": 248, "len_all": 98061, "len_all_1": 40491, "len_abs": 1643, "len_title": 128, "len_sents": 22553, "len_sents_1": 16821, "n_sents": 194, "n_sents_1": 136, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1651, "title": "Power Scheduler: A Batch Size and Token Number Agnostic Learning Rate Scheduler", "abs": "Finding the optimal learning rate for language model pretraining is a challenging task. This is not only because there is a complicated correlation between learning rate, batch size, number of training tokens, model size, and other hyperparameters but also because it is prohibitively expensive to perform a hyperparameter search for large language models with Billions or Trillions of parameters. Recent studies propose using small proxy models and small corpus to perform hyperparameter searches and transposing the optimal parameters to large models and large corpus. While the zero-shot transferability is theoretically and empirically proven for model size-related hyperparameters, like depth and width, the zero-shot transfer from small corpus to large corpus is underexplored. In this paper, we study the correlation between optimal learning rate, batch size, and number of training tokens for the recently proposed WSD scheduler. After thousands of small experiments, we found a power-law relationship between variables and demonstrated its transferability across model sizes. Based on the observation, we propose a new learning rate scheduler, Power scheduler, that is agnostic about the number of training tokens and batch size. The experiment shows that combining the Power scheduler with Maximum Update Parameterization (muP) can consistently achieve impressive performance with one set of hyperparameters regardless of the number of training tokens, batch size, model size, and even model architecture. Our 3B dense and MoE models trained with the Power scheduler achieve comparable performance as state-of-the-art small language models.", "keywords": [ "Learning rate scheduler", "language model", "pretraining" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "gLkmW0cfcI", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Kristian Kersting", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Full Professor" }, { "name": "Manuel Brack", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Researcher" }, { "name": "Martin Mundt", "gender": "Male", "institution": "Universit\u00e4t Bremen", "country": "DE", "position": "Full Professor" }, { "name": "Patrick Schramowski", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Researcher" }, { "name": "Subarnaduti Paul", "gender": "Male", "institution": "Universit\u00e4t Bremen", "country": "DE", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 8, "n_ref_uni": 49, "n_ref": 81, "n_ref_all": 100, "n_fig": 13, "n_tab": 3, "L_tab": 2039, "n_element_tab": 88, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1261, "n_element_tab_1": 55, "formula_len_all": 689, "formula_len_all_1": 418, "len_all": 167321, "len_all_1": 64251, "len_abs": 1272, "len_title": 117, "len_sents": 48754, "len_sents_1": 31406, "n_sents": 355, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1211, "title": "Core Tokensets for Data-efficient Sequential Training of Transformers", "abs": "Deep networks are frequently tuned to novel tasks and continue learning from ongoing data streams. Such sequential training requires consolidation of new and past information, a challenge predominantly addressed by retaining the most important data points - formally known as coresets. Traditionally, these coresets consist of entire samples, such as images or sentences. However, recent transformer architectures operate on tokens, leading to the famous assertion that an image is worth 16x16 words. Intuitively, not all of these tokens are equally informative or memorable. Going beyond coresets, we thus propose to construct a deeper-level data summary on the level of tokens. Ours, respectively named core tokensets, both select the most informative data points and leverage feature attribution to store only their most relevant features. We demonstrate that core tokensets yield significant performance retention in incremental image classification, open-ended visual question answering, and continual image captioning with significantly reduced memory. In fact, we empirically find that a core tokenset of 1\\% of the data performs comparably to at least a twice as large and up to 10 times larger coreset.", "keywords": [ "Transformers", "Coresets", "Sequential Learning" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "gLaEjxiRc3", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Guang Zeng", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Guoren Wang", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Hongchao Qin", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Kaichi Yu", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Rong-Hua Li", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xunkai Li", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Zhengyu Wu", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 41, "n_formula_1": 19, "n_ref_uni": 53, "n_ref": 91, "n_ref_all": 136, "n_fig": 9, "n_tab": 8, "L_tab": 5939, "n_element_tab": 307, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 3519, "n_element_tab_1": 127, "formula_len_all": 3897, "formula_len_all_1": 1597, "len_all": 273512, "len_all_1": 69190, "len_abs": 1848, "len_title": 125, "len_sents": 89943, "len_sents_1": 29428, "n_sents": 698, "n_sents_1": 240, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1728, "title": "Entropy-driven Data Knowledge Distillation in Digraph Representation Learning", "abs": "The directed graph (digraph), as a generalization of undirected graphs, exhibits superior representation capability in modeling complex topology systems and has garnered considerable attention in recent years. Despite the notable efforts made by existing DiGraph Neural Networks (DiGNNs) to leverage directed edges, they still fail to comprehensively delve into the abundant data knowledge concealed in the digraphs. This limitation results in sub-optimal performance and underscores the necessity of further exploring the potential correlations between the directed topology and node profiles from a data-centric perspective, thereby empowering model-centric neural networks with stronger encoding capabilities. In this paper, we propose \\textbf{E}ntropy-driven \\textbf{D}igraph knowl\\textbf{E}dge distillatio\\textbf{N} (EDEN), which can serve as a new data-centric digraph learning paradigm or a model-agnostic hot-and-plug data online knowledge distillation module for most existing DiGNNs to fully leverage informative digraphs. Specifically, EDEN first utilizes directed structural measurements from a topological perspective to construct a knowledge tree, guided by the hierarchical encoding theory. Subsequently, EDEN quantifies the mutual information of nodes from a feature perspective to further refine the knowledge flow, facilitating tree layer-wise knowledge distillation. As a general framework, EDEN also can naturally extend to undirected scenarios and demonstrate satisfactory performance. In our experiments, EDEN has been widely evaluated on 14 (di)graph datasets and across 4 downstream tasks. The results demonstrate that EDEN attains SOTA performance and exhibits strong improvement for prevalent (Di)GNNs.", "keywords": [ "Data Knowledge Distillation", "Graph Neural Network", "Directed Graph Learning" ], "rating_list": [ 5, 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 2, 1 ], "contribution_list": [ 2, 2, 4, 3, 2 ], "confidence_list": [ 3, 4, 2, 3, 4 ] }, { "paper_id": "gLa96FlWwn", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Dheeraj Rajagopal", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Lucas Dixon", "gender": "Not Specified", "institution": "Research, Google", "country": "", "position": "Researcher" }, { "name": "Tolga Bolukbasi", "gender": "unknown", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Ian F. Tenney", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Tyler A Chang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Student researcher" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 34, "n_ref": 106, "n_ref_all": 139, "n_fig": 2, "n_tab": 7, "L_tab": 1903, "n_element_tab": 186, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1383, "n_element_tab_1": 100, "formula_len_all": 203, "formula_len_all_1": 148, "len_all": 178082, "len_all_1": 67587, "len_abs": 1474, "len_title": 120, "len_sents": 57166, "len_sents_1": 33976, "n_sents": 397, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1485, "title": "Scalable Influence and Fact Tracing for Large Language Model Pretraining", "abs": "Training data attribution (TDA) methods aim to attribute model outputs back to specific training examples, and the application of these methods to large language model (LLM) outputs could significantly advance model transparency and data curation. However, it has been challenging to date to apply these methods to the full scale of LLM pretraining. In this paper, we introduce a gradient-based method that works effectively at scale, allowing us to retrieve influential examples for an 8B-parameter language model from a pretraining corpus of over 160B tokens with no need for subsampling or pre-filtering. Our method combines several techniques, including optimizer state correction, a task-specific Hessian approximation, and normalized encodings, which we find to be critical for performance at scale. Our method performs best at identifying examples that *influence* model predictions, but classical, model-agnostic retrieval methods such as BM25 still perform better at finding passages which explicitly contain relevant facts. These results demonstrate a misalignment between factual *attribution* and causal *influence*. With increasing model size and training tokens, we find that influence more closely aligns with attribution. Finally, we examine different types of examples identified as influential by our method, finding that while many directly entail a particular fact, others support the same output by reinforcing priors on relation types, common entities, and names.", "keywords": [ "training data attribution", "LLM pretraining", "influence functions", "factual knowledge" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 2, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "gLHuAYGs6a", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Ruina Bai", "gender": "unknown", "institution": "Guizhou University", "country": "CN", "position": "PhD student" }, { "name": "Ruizhang Huang", "gender": "unknown", "institution": "Guizhou University", "country": "CN", "position": "Full Professor" }, { "name": "Weifang Ding", "gender": "Female", "institution": "Guizhou University", "country": "CN", "position": "MS student" }, { "name": "Yanping Chen", "gender": "Male", "institution": "Guizhou University", "country": "", "position": "Full Professor" }, { "name": "Yongbin Qin", "gender": "Male", "institution": "Guizhou University", "country": "CN", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 17, "n_ref": 24, "n_ref_all": 40, "n_fig": 7, "n_tab": 4, "L_tab": 2755, "n_element_tab": 153, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2400, "n_element_tab_1": 164, "formula_len_all": 691, "formula_len_all_1": 691, "len_all": 92181, "len_all_1": 58064, "len_abs": 1162, "len_title": 119, "len_sents": 28750, "len_sents_1": 25412, "n_sents": 210, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1165, "title": "Structural Multi-view Clustering Network via Heterogeneous Random Walks", "abs": "Multi-view clustering, which aims to partition data samples into disjoint clusters by leveraging information from multiple views, has been shown to be highly effective when incorporating structure information, as widely acknowledged in recent works. This paper presents a novel structural multi-view clustering network via heterogeneous random walks, guided by a unified sample-level structure to enhance clustering performance. We first construct a multi-view heterogeneous graph consisting of sample nodes and view nodes, capturing correlations between views while preserving their specific structures. Then, a multi-step random walk strategy on the heterogeneous graph is introduced to explore high-order sample structures across various views, ensuring that each view structure is taken into account. Based on this, a lightweight network is designed to facilitate structure learning both within-view and cross-view, guided by the unified structure derived from heterogeneous random walks, ultimately achieving representations that are conducive to clustering. Extensive experiments on five real-world datasets demonstrate the superiority of the proposed method.", "keywords": [ "multi-view clustering; sample structure; heterogeneous graph; heterogeneous random walks" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 1, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 3, 2 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "gLGp77MxFo", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Chao Yang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Shuang Li", "gender": "Female", "institution": "The Chinese University of Hong Kong (Shenzhen)", "country": "", "position": "Assistant Professor" }, { "name": "Xinye Chen", "gender": "Female", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Yinghao Fu", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 53, "n_formula_1": 11, "n_ref_uni": 34, "n_ref": 34, "n_ref_all": 56, "n_fig": 4, "n_tab": 3, "L_tab": 1170, "n_element_tab": 53, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 287, "n_element_tab_1": 60, "formula_len_all": 4427, "formula_len_all_1": 922, "len_all": 155673, "len_all_1": 72043, "len_abs": 1327, "len_title": 153, "len_sents": 54698, "len_sents_1": 34831, "n_sents": 445, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1333, "title": "Who Should Join the Decision-Making Table? Targeted Expert Selection for Enhanced Human-AI Collaboration", "abs": "Integrating AI and human expertise can significantly enhance decision-making across various scenarios. This paper introduces a novel approach that leverages the Product of Experts (PoE) model to optimize decision-making by strategically combining AI with human inputs. While human experts bring diverse perspectives, their decisions may be constrained by biases or knowledge gaps. To address these limitations, we propose an AI agent that provides probabilistic, rule-based insights, complementing and filling human experts' knowledge gaps. A key feature of our approach is the strategic selection of human experts based on how well their knowledge complements or enhances the AI\u2019s recommendations. By dynamically adapting the expert selection process, we ensure that decisions benefit from the most impactful and complementary inputs. Our PoE model calibrates inputs from both AI and human experts, leveraging their combined strengths to improve decision outcomes. Furthermore, operating in an online setting, our framework can also continuously update the AI\u2019s knowledge and refine expert selection criteria, ensuring adaptability to evolving environments. Experiments in simulation environments demonstrate that our model effectively integrates logic rule-informed AI with human expertise, enhancing collaborative decision-making.", "keywords": [ "human-ai complementarity", "calibration", "rule learning" ], "rating_list": [ 5, 5, 6, 5, 5 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 5 ] }, { "paper_id": "gL1cNK2UEW", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Chung-Wei Hang", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Applied Scientiest" }, { "name": "He Wang", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Henghui Zhu", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Hideo Kobayashi", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Jiani Zhang", "gender": "Female", "institution": "Amazon Web Services", "country": "", "position": "Researcher" }, { "name": "Patrick Ng", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Sheng Zhang", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Yiqun Hu", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Hanbo Li", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 26, "n_ref": 66, "n_ref_all": 86, "n_fig": 4, "n_tab": 8, "L_tab": 3076, "n_element_tab": 184, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 245, "n_element_tab_1": 20, "formula_len_all": 111, "formula_len_all_1": 36, "len_all": 141640, "len_all_1": 54077, "len_abs": 1603, "len_title": 147, "len_sents": 43744, "len_sents_1": 27170, "n_sents": 328, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1614, "title": "DSMentor: Enhancing Data Science Agents with Curriculum Learning and Online Knowledge Accumulation", "abs": "Large language model (LLM) agents have shown promising performance in generating code for solving complex data science problems. Recent studies primarily focus on enhancing in-context learning through improved search, sampling, and planning techniques, while overlooking the importance of the order in which problems are tackled during inference. In this work, we develop a novel inference-time optimization framework, referred to as DSMentor, which leverages curriculum learning---a strategy that introduces simpler task first and progressively moves to more complex ones as the learner improves---to enhance LLM agent performance in challenging data science tasks. Our mentor-guided framework organizes data science tasks in order of increasing difficulty and incorporates a growing long-term memory to retain prior experiences, guiding the agent's learning progression and enabling more effective utilization of accumulated knowledge. We evaluate DSMentor through extensive experiments on DSEval and QRData benchmarks. Experiments show that DSMentor using Claude-3.5-Sonnet improves the pass rate by up to 5.2% on DSEval and QRData compared to baseline agents. Furthermore, DSMentor demonstrates stronger causal reasoning ability, improving the pass rate by 8.8% on the causality problems compared to GPT-4 using Program-of-Thoughts prompts. Our work underscores the importance of developing effective strategies for accumulating and utilizing knowledge during inference, mirroring the human learning process and opening new avenues for improving LLM performance through curriculum-based inference optimization.", "keywords": [ "curriculum learning", "data science agent", "long-term memory", "online data retrieval" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 1, 2, 1, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "gKui6QvvfK", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Jamie Stirling", "gender": "Male", "institution": "Durham University", "country": "GB", "position": "PhD student" }, { "name": "Noura Al Moubayed", "gender": "Female", "institution": "Durham University", "country": "GB", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 7, "n_ref_uni": 29, "n_ref": 85, "n_ref_all": 135, "n_fig": 8, "n_tab": 4, "L_tab": 1506, "n_element_tab": 96, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 973, "n_element_tab_1": 48, "formula_len_all": 1085, "formula_len_all_1": 581, "len_all": 152067, "len_all_1": 58645, "len_abs": 1986, "len_title": 129, "len_sents": 53592, "len_sents_1": 29733, "n_sents": 325, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 2011, "title": "Compositional VQ Sampling for Efficient and Accurate Conditional Image Generation", "abs": "Compositional diffusion and energy-based models have driven progress in controllable image generation, however the challenge of composing discrete generative models has remained open, holding the potential for improvements in efficiency, interpretability and generation quality. To this end, we propose a framework for controllable conditional generation of images. We formulate a process for composing discrete generation processes, enabling generation with an arbitrary number of input conditions without the need for any specialised training objective. We adapt this result for parallel token prediction with masked generative transformers, enabling accurate and efficient conditional sampling from the discrete latent space of VQ models. In particular, our method attains an average error rate of 19.3% across nine experiments spanning three datasets (between one and three input conditions for each dataset), representing an average 63.4% reduction in error rate relative to the previous state-of-the-art. Our method also outperforms the next-best approach (ranked by error rate) in terms of FID in seven out of nine settings, with an average FID of $24.23$, and average improvement of $-9.58$. Furthermore, our method offers a $2.3\\times$ to $12\\times$ speedup over comparable methods. We find that our method can generalise to combinations of input conditions that lie outside the training data (e.g. more objects per image for Positional CLEVR) in addition to offering an interpretable dimension of controllability via concept weighting. Outside of the rigorous quantitative settings, we further demonstrate that our approach can be readily applied to an open pre-trained discrete text-to-image model, demonstrating fine-grained control of text-to-image generation. The accuracy and efficiency of our framework across diverse conditional image generation settings reinforces its theoretical foundations, while opening up practical avenues for future work in controllable and composable image generation.", "keywords": [ "image generation", "compositional generalization" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "gKM8wwsTOg", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Chunyang Jiang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Jianxuan Yu", "gender": "Male", "institution": "tsinghua university", "country": "", "position": "Undergrad student" }, { "name": "Wei Xue", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Weijie Shi", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Weize Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiaowei Chi", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Xinyu Liu", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yike Guo", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Zhiyuan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "CHIMIN CHAN", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 0, "n_ref_uni": 52, "n_ref": 86, "n_ref_all": 116, "n_fig": 13, "n_tab": 5, "L_tab": 2828, "n_element_tab": 178, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 2302, "n_element_tab_1": 160, "formula_len_all": 499, "formula_len_all_1": 0, "len_all": 260059, "len_all_1": 59511, "len_abs": 2047, "len_title": 134, "len_sents": 86769, "len_sents_1": 27114, "n_sents": 755, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1966, "title": "AgentMonitor: A Plug-and-Play Framework for Predictive and Secure Multi-Agent Systems", "abs": "Rapid advancement of large language models (LLMs) has catalyzed the emergence of LLM-based agents. Recent research has shifted from single-agent systems to multi-agent frameworks, demonstrating that collaboration can outperform the capabilities of individual LLMs. However, effectively pre-configuring a Multi-Agent System (MAS) for a specific task remains a challenging problem, with performance outcomes only observable after execution. Inspired by the well-established scaling laws in LLM development that model downstream task performance or validation loss as functions of various factors during training, we seek to investigate the predictability of MAS performance. Specifically, we explore whether it is possible to predict the downstream task performance of a configured MAS. In addition, MAS face a growing challenge in ensuring reliable and trustworthy responses. The introduction of malicious agents can lead to the generation and spread of harmful content, which poses significant security risks. To address the above issues, we introduce AgentMonitor, a framework that integrates with existing MAS at the agent level. AgentMonitor captures inputs and outputs at each step; this enables (1) transforming them into relevant statistics supporting the training of a regression model to predict task performance and (2) the application of on-the-fly corrections to mitigate negative impacts on final outcomes. Extensive experiments demonstrate that training a simple XGBoost model achieves a high Spearman rank correlation of 0.89 in an in-domain setting. In more challenging scenarios, where the statistics of a specific task or architecture is absent from the training set, our method maintains a moderate average correlation of 0.58. Furthermore, by employing AgentMonitor in a maliciously configured MAS, the system ultimately generates 6.2% less harmful content and 1.8% more helpful content on average, reducing safety risks and improving reliability.", "keywords": [ "LLM-Agents", "Multi-Agents", "Large Language Models" ], "rating_list": [ 5, 5, 3, 5, 6 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 4, 4 ] }, { "paper_id": "gK1rl98VRp", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Huayi Tang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Xiaolin Hu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zixuan Gong", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 133, "n_formula_1": 17, "n_ref_uni": 53, "n_ref": 102, "n_ref_all": 185, "n_fig": 6, "n_tab": 5, "L_tab": 726, "n_element_tab": 97, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 213, "n_element_tab_1": 5, "formula_len_all": 14906, "formula_len_all_1": 1649, "len_all": 324359, "len_all_1": 74601, "len_abs": 2389, "len_title": 144, "len_sents": 110759, "len_sents_1": 33933, "n_sents": 919, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1616, "title": "Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization", "abs": "Large language models (LLMs) have demonstrated remarkable in-context learning (ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits two limitations: \\textbf{(a) Limited \\textit{i.i.d.} Setting.} Most studies focus on supervised function learning tasks where prompts are constructed with \\textit{i.i.d.} input-label pairs. This \\textit{i.i.d.} assumption diverges significantly from real language learning scenarios where prompt tokens are interdependent. \\textbf{(b) Lack of Emergence Explanation.} Most literature answers \\textbf{\\textit{what}} ICL does from an implicit optimization perspective but falls short in elucidating \\textbf{\\textit{how}} ICL emerges and the impact of pre-training phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm, \\textbf{\\textit{auto-regressive next-token prediction (AR-NTP)}}, which closely aligns with the actual training of language models. Specifically, within AR-NTP, we emphasize prompt token-dependency, which involves predicting each subsequent token based on the preceding sequence. To address (b), we formalize a systematic pre-training and ICL framework, highlighting the layer-wise structure of sequences and topics, alongside a two-level expectation. In conclusion, we present data-dependent, topic-dependent and optimization-dependent PAC-Bayesian generalization bounds for pre-trained LLMs, investigating that \\textbf{\\textit{ICL emerges from the generalization of sequences and topics}}. Our theory is supported by experiments on numerical linear dynamic systems, synthetic GINC and real-world language datasets.", "keywords": [ "In-context learning", "Auto-regressive next-token prediction", "Generalization performance", "PAC-Bayesian" ], "rating_list": [ 5, 8, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 2, 3 ], "confidence_list": [ 2, 2, 3, 3, 3 ] }, { "paper_id": "gJk4N7zscD", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Gagandeep Singh", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Gaokai Zhang", "gender": "unknown", "institution": "Microsoft Research Asia", "country": "CN", "position": "Intern" }, { "name": "Hangoo Kang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Undergrad student" }, { "name": "Jason Vega", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Junsheng Huang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Undergrad student" }, { "name": "Minjia Zhang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 4, "n_ref_uni": 42, "n_ref": 75, "n_ref_all": 111, "n_fig": 12, "n_tab": 6, "L_tab": 3144, "n_element_tab": 178, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 866, "n_element_tab_1": 25, "formula_len_all": 701, "formula_len_all_1": 291, "len_all": 196550, "len_all_1": 58523, "len_abs": 4999, "len_title": 132, "len_sents": 68613, "len_sents_1": 27492, "n_sents": 445, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1119, "title": "Stochastic Monkeys at Play: Random Augmentations Cheaply Break LLM Safety Alignment", "abs": "Safety alignment of Large Language Models (LLMs) has recently become a critical objective of model developers. In response, a growing body of work has been investigating how safety alignment can be bypassed through various jailbreaking methods, such as adversarial attacks. However, these jailbreak methods can be rather costly or involve a non-trivial amount of creativity and effort, introducing the assumption that malicious users are high-resource or sophisticated. In this paper, we study how simple random augmentations to the input prompt affect safety alignment effectiveness in state-of-the-art LLMs, such as Llama 3 and Qwen 2. We perform an in-depth evaluation of 17 different models and investigate the intersection of safety under random augmentations with multiple dimensions: augmentation type, model size, quantization, fine-tuning-based defenses, and decoding strategies (e.g., sampling temperature). We show that low-resource and unsophisticated attackers, i.e. $\\textit{stochastic monkeys}$, can significantly improve their chances of bypassing alignment with just 25 random augmentations per prompt.", "keywords": [ "LLM", "Large Language Model", "safety alignment", "augmentations", "randomness", "jailbreaks", "attacks" ], "rating_list": [ 5, 3, 6, 5, 5 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 2, 4, 4 ] }, { "paper_id": "gJPe4dxm7N", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Chi Luo", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Huanhuan Chen", "gender": "Male", "institution": "University of Science and Technology of China, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Shengfei Lyu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "Xin Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Xingjian Lin", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 65, "n_formula_1": 26, "n_ref_uni": 42, "n_ref": 136, "n_ref_all": 156, "n_fig": 7, "n_tab": 6, "L_tab": 1838, "n_element_tab": 54, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 413, "n_element_tab_1": 7, "formula_len_all": 4659, "formula_len_all_1": 1781, "len_all": 157885, "len_all_1": 67022, "len_abs": 1211, "len_title": 168, "len_sents": 50673, "len_sents_1": 30863, "n_sents": 366, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 119, "L_abs": 1076, "title": "Temporal Adaptive Convolutional Intervention Network for Counterfactual Estimation: A Domain Generalization Perspective", "abs": "Accurate estimation of time-varying treatment effects is crucial for optimizing interventions in personalized medicine. However, observational data often contains complex confounding bias and temporal complexities, making counterfactual estimation challenging. We propose Temporal Adaptive Convolutional Intervention Network (TACIN), a novel model that introduces an Intervention-aware Functional Convolution kernel to emphasize the role of treatments and capture complex temporal treatment interactions. TACIN addresses confounding bias from a domain generalization perspective, approximating the unknown target domain using adversarial examples and incorporating Sharpness-Aware Minimization to derive a generalization bound. This approach is more suitable for longitudinal settings compared to existing methods inspired by domain adaptation techniques due to inherent differences between static and longitudinal contexts. Experiments on simulated datasets demonstrate TACIN's superior performance compared to state-of-the-art models for counterfactual estimation over time.", "keywords": [ "causal inference", "longitudinal data", "time-varying confounding bias", "domain generalization" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 1, 4, 2, 1 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 1, 3, 2, 1 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "gJG4IPwg6l", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Annie S Chen", "gender": "Female", "institution": "Google", "country": "GB", "position": "Intern" }, { "name": "Charlie Gauthier", "gender": "Female", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "PhD student" }, { "name": "Kaustubh Mani", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, University of Montreal", "country": "", "position": "Intern" }, { "name": "Liam Paull", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Associate Professor" }, { "name": "Samer B. Nashed", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Postdoc" }, { "name": "Vincent Mai", "gender": "Male", "institution": "IREQ - Hydro Quebec Research Center", "country": "CA", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 36, "n_ref": 84, "n_ref_all": 129, "n_fig": 17, "n_tab": 2, "L_tab": 1115, "n_element_tab": 83, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 939, "n_element_tab_1": 78, "formula_len_all": 220, "formula_len_all_1": 220, "len_all": 155451, "len_all_1": 69914, "len_abs": 1135, "len_title": 110, "len_sents": 56081, "len_sents_1": 34700, "n_sents": 357, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1139, "title": "Risk Informed Policy Learning for Safer Exploration", "abs": "Reinforcement learning algorithms typically necessitate extensive exploration of the state space to find optimal policies. However, in safety-critical applications, the risks associated with such exploration can lead to catastrophic consequences. Existing safe exploration methods attempt to mitigate this by imposing constraints, which often result in overly conservative behaviours and inefficient learning. Heavy penalties for early constraint violations can trap agents in local optima, deterring exploration of risky yet high-reward regions of the state-space. To address this, we introduce a method that explicitly learns state-conditioned safety representations. By augmenting state features with these safety representations, our approach naturally encourages safer exploration without being excessively cautious, resulting\nin more efficient and safer policy learning in safety-critical scenarios. Empirical evaluations across diverse environments show that our method significantly improves task performance while reducing constraint violations during training, underscoring its effectiveness in balancing exploration with safety.", "keywords": [ "Reinforcement Learning", "Safe Exploration", "Representation Learning", "Inductive Bias" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "gJ48psisby", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Andrew Lizarraga", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Bernie Boscoe", "gender": "unknown", "institution": "Southern Oregon University", "country": "US", "position": "Assistant Professor" }, { "name": "Eric Hanchen Jiang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Undergrad student" }, { "name": "Jacob Nowack", "gender": "unknown", "institution": "Southern Oregon State College", "country": "US", "position": "Undergrad student" }, { "name": "Tuan Do", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "Yingnian Wu", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Full Professor" }, { "name": "Yunqi Billy Li", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 17, "n_ref": 34, "n_ref_all": 60, "n_fig": 10, "n_tab": 1, "L_tab": 951, "n_element_tab": 26, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 1072, "n_element_tab_1": 30, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 93318, "len_all_1": 55791, "len_abs": 1333, "len_title": 132, "len_sents": 32699, "len_sents_1": 27579, "n_sents": 222, "n_sents_1": 163, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1363, "title": "Conditional Trajectories in Diffusion Models - Modeling Galaxy Evolution from Redshift", "abs": "In this paper, we present a novel approach for continuous {\\bf C}onditional {\\bf T}rajectories on Denoising {\\bf D}iffusion Probabilistic {\\bf M}odels (CTDM). Focusing on physical applications, our model learns to capture the underlying relationship between galaxy images and their redshift values from training data. This enables the simulation of galaxy evolution by conditioning the reverse denoising process on future redshift values. Importantly, this is achieved without requiring multiple images of the same galaxy at different redshifts.\nWe demonstrate that our redshift-conditioned diffusion model learns the marginal distribution of galaxy images at each redshift value. This allows the model to generate realistic galaxy images that reflect the physical changes occurring as galaxies evolve. We derive a smoothness condition for this learned distribution, proving that the model can construct trajectories between galaxy images by incrementally changing redshift during the reverse denoising process.\nOur approach offers a novel interpretation of the learned diffusion process as a means to simulate galaxy evolution, capturing both visual and physical changes over time. These techniques not only provide deeper insights into the formation and evolution of galaxies but also have broader potential applications in various areas of generative modeling.", "keywords": [ "Diffusion", "Galaxy", "Morphology", "Redshift" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 1, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "gIrVoQEDQv", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Andreza Maria Coutinho Falcao", "gender": "Female", "institution": "Universidade Federal Rural de Pernambuco", "country": "BR", "position": "Undergrad student" }, { "name": "Filipe R. Cordeiro", "gender": "Male", "institution": "Universidade Federal Rural de Pernambuco", "country": "BR", "position": "Associate Professor" }, { "name": "IURY ADONES XAVIER DOS SANTOS", "gender": "Male", "institution": "Universidade Federal Rural de Pernambuco", "country": "BR", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 28, "n_ref": 50, "n_ref_all": 61, "n_fig": 6, "n_tab": 3, "L_tab": 112, "n_element_tab": 15, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 112, "n_element_tab_1": 15, "formula_len_all": 340, "formula_len_all_1": 355, "len_all": 85577, "len_all_1": 51725, "len_abs": 1054, "len_title": 117, "len_sents": 28142, "len_sents_1": 23425, "n_sents": 203, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1059, "title": "Unraveling Neural Cellular Automata for Lightweight Image Compression", "abs": "Neural Cellular Automata (NCA) are computational models inspired by cellular growth, capable of learning complex behaviors through local interactions. While NCAs have been applied to various tasks like image restoration and synthesis, their potential for image compression remains largely unexplored. This paper aims to unravel the capabilities of NCAs for lightweight image compression by introducing a Grid Neural Cellular Automata (GNCA) training strategy. Unlike traditional methods that depend on large deep learning models, NCAs offer a low-cost compact and highly parallelizable alternative with intrinsic robustness to noise. Through experiments on the COCO 2017 dataset, we compare the compression performance of NCAs against JPEG, JPEG-2000 and WebP, using the metrics PSNR, SSIM, and MSE and Compression Rate. Our results demonstrate that NCAs achieve competitive compression rates and image quality reconstruction, highlighting their potential as a lightweight solution for efficient image compression. The code will be available upon acceptance.", "keywords": [ "image compression", "automata", "deep learning" ], "rating_list": [ 3, 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2, 1 ], "presentation_list": [ 2, 2, 1, 2, 3 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 2, 4 ] }, { "paper_id": "gInIbukM0R", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Abdulla Kuleib", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Undergrad student" }, { "name": "Faisal Saleh Alshinaifi", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "Undergrad student" }, { "name": "Gabriel A. Silva", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Johnny Jingze Li", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Zeyad Almoaigel", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "Undergrad student" } ], "n_formula": 9, "n_formula_1": 6, "n_ref_uni": 21, "n_ref": 30, "n_ref_all": 41, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 406, "formula_len_all_1": 207, "len_all": 95476, "len_all_1": 57336, "len_abs": 1192, "len_title": 85, "len_sents": 38263, "len_sents_1": 29909, "n_sents": 266, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1184, "title": "Quantifying Emergence in Neural Networks: Insights from Pruning and Training Dynamics", "abs": "Emergence, where complex behaviors develop from the interactions of simpler components within a network, plays a crucial role in enhancing neural network capabilities. We introduce a quantitative framework to measure emergence as structural nonlinearity, study the dynamics of this measure during the training process, and examine its impact on network performance, particularly in relation to pruning and training dynamics. Our hypothesis posits that the degree of emergence\u2014evaluated from the distribution and connectivity of active nodes\u2014can predict the development of emergent behaviors in the network. We demonstrate that higher emergence correlates with improved trianing performance. We further explore the relationship between network complexity and the loss landscape, suggesting that higher emergence indicates a greater concentration of local minima and a more rugged loss landscape. We show that this framework can be applied to explain the impact of pruning on the training dynamics. These findings provide new insights into the interplay between emergence, complexity, and performance in neural networks, offering implications for designing and optimizing architectures.", "keywords": [ "Emergence", "training dynamics", "pruning", "landscape" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 1 ], "presentation_list": [ 1, 1, 2, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "gINO3tfVEP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Haifeng Wu", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Lixin Duan", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Shuhang Gu", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Wen Li", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 28, "n_formula_1": 23, "n_ref_uni": 43, "n_ref": 94, "n_ref_all": 132, "n_fig": 18, "n_tab": 5, "L_tab": 5274, "n_element_tab": 400, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1843, "n_element_tab_1": 187, "formula_len_all": 1581, "formula_len_all_1": 1457, "len_all": 181135, "len_all_1": 63990, "len_abs": 1731, "len_title": 151, "len_sents": 43626, "len_sents_1": 27086, "n_sents": 341, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1743, "title": "Beyond 2D Representation: Learning 3D Scene Field for Robust Monocular Depth Estimation", "abs": "Monocular depth estimation has been extensively studied over the past few decades, yet achieving robust depth estimation in real-world scenes remains a challenge, particularly in the presence of reflections, shadow occlusions, and low-texture regions. Existing methods typically rely on extracting front-view 2D features for depth estimation, which often fail to capture those complex physical factors present in real-world scenes, leading to discontinuous, incomplete, or inconsistent depth maps. To address these issues, we turn to learning a more powerful 3D representation for robust monocular depth estimation, and propose a novel self-supervised monocular depth estimation framework based on the Three-dimensional Scene Field representation, or TSF-Depth for short. Specifically, we build our TSF-Depth framework upon an encoder-decoder architecture. The encoder extracts scene features from the input 2D image, and subsequently reshapes it as a tri-plane feature field by incorporating scene prior encoding. This tri-plane feature field is designed to implicitly model the structure and appearance of the continuous 3D scene. We then estimate a high-quality depth map from the tri-plane feature field by simulating the camera imaging process. To do this, we construct a 2D feature map with 3D geometry by sampling from the tri-plane feature field using the coordinates of points where the line of sight intersects with the scene. The aggregated multi-view geometric features are subsequently fed into the decoder for depth estimation. Extensive experiments on KITTI and NYUv2 datasets show that TSF-Depth achieves state-of-the-art performance. We also validate the generalization capability of our model on Make3D and ScanNet datasets.", "keywords": [ "Monocular depth estimation", "self-supervised", "3D scene field", "3D geometric" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "gI0kPklUKS", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Alice Rigg", "gender": "Non-Binary", "institution": "EleutherAI", "country": "US", "position": "Researcher" }, { "name": "Jose Oramas", "gender": "Male", "institution": "University of Antwerp", "country": "BE", "position": "Associate Professor" }, { "name": "Lee Sharkey", "gender": "Male", "institution": "Apollo Research", "country": "", "position": "Researcher" }, { "name": "Michael T Pearce", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Thomas Dooms", "gender": "Male", "institution": "Universiteit Antwerpen", "country": "", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 26, "n_ref": 43, "n_ref_all": 87, "n_fig": 25, "n_tab": 6, "L_tab": 780, "n_element_tab": 115, "n_fig_1": 16, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 415, "formula_len_all_1": 413, "len_all": 166395, "len_all_1": 61466, "len_abs": 3556, "len_title": 81, "len_sents": 58133, "len_sents_1": 29710, "n_sents": 542, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 62, "L_abs": 1255, "title": "Bilinear MLPs enable weight-based mechanistic interpretability", "abs": "A mechanistic understanding of how MLPs do computation in deep neural networks remains elusive. Current interpretability work can extract features from hidden activations over an input dataset but generally cannot explain how MLP weights construct features. One challenge is that element-wise nonlinearities introduce higher-order interactions and make it difficult to trace computations through the MLP layer. In this paper, we analyze bilinear MLPs, a type of Gated Linear Unit (GLU) without any element-wise nonlinearity that nevertheless achieves competitive performance. Bilinear MLPs can be fully expressed in terms of linear operations using a third-order tensor, allowing flexible analysis of the weights. Analyzing the spectra of bilinear MLP weights using eigendecomposition reveals interpretable low-rank structure across toy tasks, image classification, and language modeling. We use this understanding to craft adversarial examples, uncover overfitting, and identify small language model circuits directly from the weights alone. Our results demonstrate that bilinear layers serve as an interpretable drop-in replacement for current activation functions and that weight-based interpretability is viable for understanding deep-learning models.", "keywords": [ "interpretability", "mechanistic interpretability", "bilinear", "feature extraction", "weight-based", "eigenvector", "eigendecomposition", "tensor network" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "gHPUXP51L0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Brittany T. Morin", "gender": "Female", "institution": "University of California, San Francisco", "country": "US", "position": "Researcher" }, { "name": "David Paul Galang Baquirin", "gender": "Male", "institution": "University of California, San Francisco", "country": "US", "position": "Researcher" }, { "name": "Gopala Anumanchipalli", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Assistant Professor" }, { "name": "Jet M.J. Vonk", "gender": "unknown", "institution": "University of California, San Francisco", "country": "", "position": "Assistant Professor" }, { "name": "Jiachen Lian", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department, University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Maria Luisa Gorno-Tempini", "gender": "Female", "institution": "University of California, San Francisco", "country": "", "position": "Full Professor" }, { "name": "Xuanru Zhou", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Zachary A. Miller", "gender": "Male", "institution": "University of California, San Francisco", "country": "", "position": "Associate Professor" }, { "name": "Zoe Ezzes", "gender": "Female", "institution": "University of California, San Francisco", "country": "", "position": "Researcher" } ], "n_formula": 34, "n_formula_1": 24, "n_ref_uni": 37, "n_ref": 128, "n_ref_all": 150, "n_fig": 9, "n_tab": 9, "L_tab": 3716, "n_element_tab": 452, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3380, "n_element_tab_1": 315, "formula_len_all": 3828, "formula_len_all_1": 2508, "len_all": 244605, "len_all_1": 71003, "len_abs": 1404, "len_title": 0, "len_sents": 65196, "len_sents_1": 27886, "n_sents": 553, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1516, "title": "Time-Accurate Speech Rich Transcription with Non-Fluencies", "abs": "Speech is a hierarchical collection of text, prosody, emotions, dysfluencies, etc. Automatic transcription of speech that goes beyond text (words) is an underexplored problem.\nWe focus on transcribing speech along with non-fluencies (dysfluencies). The current state-of-the-art pipeline \\citep{lian2024ssdmscalablespeechdysfluency} suffers from complex architecture design, training complexity, and significant shortcomings in the local sequence aligner, and it does not explore in-context learning capacity. In this work, we propose SSDM 2.0, which tackles those shortcomings via four main contributions:\n(1) We propose a novel \\textit{neural articulatory flow} to derive highly scalable speech representations.\n(2) We developed a \\textit{full-stack connectionist subsequence aligner} that captures all types of dysfluencies.\n(3) We introduced a mispronunciation prompt pipeline and consistency learning module into LLM to leverage dysfluency \\textit{in-context pronunciation learning} abilities.\n(4) We curated Libri-Dys \\citep{lian2024ssdmscalablespeechdysfluency} and open-sourced the current largest-scale co-dysfluency corpus, \\textit{Libri-Co-Dys}, for future research endeavors.\n\\hl{In clinical experiments on pathological speech transcription, we tested SSDM 2.0 using nfvPPA corpus primarily characterized by \\textit{articulatory dysfluencies}.}\nOverall, SSDM 2.0 outperforms SSDM and all other dysfluency transcription models by a large margin. See our project demo page at \\url{https://srnf2.github.io/}.", "keywords": [ "Non-fluency", "Dysfluency", "Alignment", "speech transcription", "clinical" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 2, 2, 4 ] }, { "paper_id": "gHLWTzKiZV", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Andreas Krause", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Gabriele Corso", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Noah Getz", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Regina Barzilay", "gender": "female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Professor" }, { "name": "Tommi Jaakkola", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Vignesh Ram Somnath", "gender": "Male", "institution": "Swiss Federal Institute of Technology, Zurich", "country": "CH", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 16, "n_ref_uni": 31, "n_ref": 64, "n_ref_all": 80, "n_fig": 7, "n_tab": 3, "L_tab": 863, "n_element_tab": 145, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 497, "n_element_tab_1": 107, "formula_len_all": 3226, "formula_len_all_1": 1288, "len_all": 145338, "len_all_1": 64396, "len_abs": 933, "len_title": 110, "len_sents": 50840, "len_sents_1": 30981, "n_sents": 366, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 62, "L_abs": 936, "title": "Composing Unbalanced Flows for Flexible Docking and Relaxation", "abs": "Diffusion models have emerged as a successful approach for molecular docking, but they often cannot model protein flexibility or generate nonphysical poses. We argue that both these challenges can be tackled by framing the problem as a transport between distributions. Still, existing paradigms lack the flexibility to define effective maps between such complex distributions. To address this limitation we propose Unbalanced Flow Matching, a generalization of Flow Matching (FM) that allows trading off sample efficiency with approximation accuracy and enables more accurate transport. Empirically, we apply Unbalanced FM on flexible docking and structure relaxation, demonstrating our ability to model protein flexibility and generate energetically favorable poses. On the PDBBind docking benchmark, our method FlexDock improves the docking performance while increasing the proportion of energetically favorable poses from 30% to 73%.", "keywords": [ "molecular docking", "flow matching", "structure relaxation", "unbalanced transport" ], "rating_list": [ 10, 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 5, 2, 3, 3, 3 ] }, { "paper_id": "gGpuhyjIlS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Jiaqi Lv", "gender": "Female", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Shiyu Xia", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "PhD student" }, { "name": "Xin Geng", "gender": "Male", "institution": "Southeast University, China", "country": "CN", "position": "Professor" }, { "name": "Xu Yang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Jiaze XU", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 25, "n_ref": 50, "n_ref_all": 57, "n_fig": 4, "n_tab": 9, "L_tab": 4148, "n_element_tab": 467, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2812, "n_element_tab_1": 391, "formula_len_all": 356, "formula_len_all_1": 266, "len_all": 116476, "len_all_1": 65308, "len_abs": 1325, "len_title": 136, "len_sents": 32846, "len_sents_1": 27851, "n_sents": 221, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1329, "title": "Learngene Tells You How to Customize: Task-Aware Parameter Prediction at Flexible Scales", "abs": "Reducing serving costs and latency is a fundamental challenge for deploying large-scale models in business applications. To cope with this demand, the Learngene framework encapsulates shareable information from large models into a compact unit called a learngene. This unit serves to initialize downstream models, enabling them to inherit the knowledge from the large model efficiently, hopefully diminishing deployment expenses. However, existing learngene methods are constrained by their strong dependence on the architecture of large model and overlook the features of target tasks, resulting in suboptimal adaptability of downstream models to deployment requirements. In this paper, we present Task-Aware Learngene (TAL), a novel method based on graph hypernetworks that predicts model parameters conditioned on desired model scales and task-specific characteristics. Extensive experiments demonstrate that TAL effectively scales model initialization parameters, selectively utilizes shareable information pertinent to target tasks, and consistently outperforms random initialization and existing parameter prediction methods. Furthermore, TAL exhibits promising transfer learning capabilities for unseen tasks, underscoring its effectiveness in condensing large model knowledge while being aware of downstream requirements.", "keywords": [ "parameter prediction; learngene" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "gGWYecsK1U", "primary_area": "generative models", "n_author": 14, "author_info_list": [ { "name": "Jian Jiao", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Kaishuai Xu", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Peng CHENG", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Qi Chen", "gender": "Female", "institution": "Microsoft Research", "country": "CA", "position": "Principal Researcher" }, { "name": "Song Wang", "gender": "Male", "institution": "Microsoft Azure AI", "country": "", "position": "Senior Applied Scientist" }, { "name": "Wayne Xiong", "gender": "Male", "institution": "Research, Microsoft", "country": "", "position": "Principal Researcher" }, { "name": "Wen Xiao", "gender": "Female", "institution": "Microsoft", "country": "CA", "position": "Researcher" }, { "name": "Wenge Liu", "gender": "unknown", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Wenjie Li", "gender": "Female", "institution": "The Hong Kong Polytechnic University, The Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Wenjun Hou", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Xiao Liang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yeyun Gong", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Yi Cheng", "gender": "Female", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Yuji Zhang", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 49, "n_ref": 182, "n_ref_all": 240, "n_fig": 11, "n_tab": 9, "L_tab": 7249, "n_element_tab": 459, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 5013, "n_element_tab_1": 348, "formula_len_all": 546, "formula_len_all_1": 459, "len_all": 293415, "len_all_1": 73697, "len_abs": 1478, "len_title": 118, "len_sents": 98993, "len_sents_1": 31036, "n_sents": 751, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1210, "title": "Integrative Decoding: Improving Factuality via Implicit Self-consistency", "abs": "Self-consistency-based approaches, which involve repeatedly sampling multiple outputs and selecting the most consistent one as the final response, prove to be remarkably effective in improving the factual accuracy of large language models. Nonetheless, existing methods usually have strict constraints on the task format, largely limiting their applicability. In this paper, we present Integrative Decoding (ID), to unlock the potential of self-consistency in open-ended generation tasks. ID operates by constructing a set of inputs, each prepended with a previously sampled response, and then processes them concurrently, with the next token being selected by aggregating of all their corresponding predictions at each decoding step. In essence, this simple approach implicitly incorporates self-consistency in the decoding objective. Extensive evaluation shows that ID consistently enhances factuality over a wide range of language models, with substantial improvements on the TruthfulQA (+11.2%), Biographies (+15.4%) and LongFact (+8.5%) benchmarks. The performance gains amplify progressively as the number of sampled responses increases, indicating the potential of ID to scale up with repeated sampling.", "keywords": [ "Large Language Models", "Hallucination", "Factuality", "Decoding" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "gGElk5T8sD", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "An Yang", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Bofei Gao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Bowen Yu", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Dayiheng Liu", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Jianhong Tu", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jingren Zhou", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Junyang Lin", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Tianyi Tang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yichang Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Shr Quan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 41, "n_ref_all": 61, "n_fig": 6, "n_tab": 7, "L_tab": 4183, "n_element_tab": 425, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1293, "n_element_tab_1": 243, "formula_len_all": 666, "formula_len_all_1": 251, "len_all": 172909, "len_all_1": 67351, "len_abs": 1334, "len_title": 105, "len_sents": 62541, "len_sents_1": 30041, "n_sents": 468, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1344, "title": "Language Models can Self-Lengthen to Generate Long Texts", "abs": "Recent advancements in Large Language Models (LLMs) have significantly enhanced their ability to process long contexts, yet a notable gap remains in generating long, aligned outputs. \nThis limitation stems from a training gap where pre-training lacks effective instructions for long-text generation, and post-training data primarily consists of short query-response pairs. \nCurrent approaches, such as instruction backtranslation and behavior imitation, face challenges including data quality, copyright issues, and constraints on proprietary model usage.\nIn this paper, we introduce an innovative iterative training framework called Self-Lengthen that leverages only the intrinsic knowledge and skills of LLMs without the need for auxiliary data or proprietary models. \nThe framework consists of two roles: the Generator and the Extender. The Generator produces the initial response, which is then split and expanded by the Extender. This process results in a new, longer response, which is used to train both the Generator and the Extender iteratively. Through this process, the models are progressively trained to handle increasingly longer responses. \nExperiments on benchmarks and human evaluations show that Self-Lengthen outperforms existing methods in long-text generation, when applied to top open-source LLMs such as Qwen2 and LLaMA3.", "keywords": [ "LLM", "Synthetic Data", "Long Output" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "gG7P1SL0QS", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Haibo Hu", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Jiawei Duan", "gender": "Not Specified", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Postdoc" }, { "name": "Qingqing Ye", "gender": "Female", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 20, "n_ref_uni": 26, "n_ref": 54, "n_ref_all": 72, "n_fig": 3, "n_tab": 2, "L_tab": 407, "n_element_tab": 43, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 251, "n_element_tab_1": 26, "formula_len_all": 1888, "formula_len_all_1": 1362, "len_all": 135728, "len_all_1": 69565, "len_abs": 1516, "len_title": 109, "len_sents": 31779, "len_sents_1": 29101, "n_sents": 270, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1709, "title": "Analyzing and Optimizing Perturbation of DP-SGD Geometrically", "abs": "Differential privacy (DP) has become a prevalent privacy model in\na wide range of machine learning tasks, especially after the debut\nof DP-SGD. However, DP-SGD, which directly perturbs gradients\nin the training iterations, fails to mitigate the negative impacts of\nnoise on gradient direction. As a result, DP-SGD is often inefficient.\nAlthough various solutions (e.g., clipping to reduce the sensitivity\nof gradients and amplifying privacy bounds to save privacy budgets)\nare proposed to trade privacy for model efficiency, the root cause of\nits inefficiency is yet unveiled.\nIn this work, we first generalize DP-SGD and theoretically derive\nthe impact of DP noise on the training process. Our analysis reveals\nthat, in terms of a perturbed gradient, only the noise on a direction\nhas eminent impact on the model efficiency while that on magnitude\ncan be mitigated by optimization techniques, i.e., fine-tuning gradient\nclipping and learning rate. Besides, we confirm that traditional\nDP introduces biased noise on the direction when adding unbiased\nnoise to the gradient itself. Overall, the perturbation of DP-SGD is\nactually sub-optimal from a geometric perspective. Motivated by\nthis, we design a geometric perturbation strategy GeoDP within the\nDP framework, which perturbs the direction and the magnitude of a\ngradient, respectively. By directly reducing the noise on the direction,\nGeoDP mitigates the negative impact of DP noise on model\nefficiency with the same DP guarantee. Extensive experiments on\ntwo public datasets (i.e., MNIST and CIFAR-10), one synthetic\ndataset and three prevalent models (i.e., Logistic Regression, CNN\nand ResNet) confirm the effectiveness and generality of our strategy.", "keywords": [ "Differential Privacy", "Stochastic Gradient Descent", "Deep Learning", "Efficiency Analysis and Enhancement" ], "rating_list": [ 5, 1, 6, 3, 1 ], "soundness_list": [ 2, 1, 3, 1, 1 ], "presentation_list": [ 2, 3, 3, 1, 1 ], "contribution_list": [ 2, 1, 4, 2, 1 ], "confidence_list": [ 4, 4, 3, 4, 5 ] }, { "paper_id": "gFvRRCnQvX", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Hee-Youl Kwak", "gender": "unknown", "institution": "Ulsan University", "country": "", "position": "Assistant Professor" }, { "name": "Jong-Seon No", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Emeritus" }, { "name": "Sang-Hyo Kim", "gender": "unknown", "institution": "Sung Kyun Kwan University", "country": "", "position": "Full Professor" }, { "name": "Seong-Joon Park", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Postdoc" }, { "name": "Yongjune Kim", "gender": "unknown", "institution": "POSTECH", "country": "", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 26, "n_ref": 92, "n_ref_all": 126, "n_fig": 17, "n_tab": 10, "L_tab": 3809, "n_element_tab": 499, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1362, "n_element_tab_1": 215, "formula_len_all": 535, "formula_len_all_1": 576, "len_all": 136329, "len_all_1": 58230, "len_abs": 1218, "len_title": 128, "len_sents": 43254, "len_sents_1": 25639, "n_sents": 325, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1223, "title": "CrossMPT: Cross-attention Message-passing Transformer for Error Correcting Codes", "abs": "Error correcting codes (ECCs) are indispensable for reliable transmission in communication systems. The recent advancements in deep learning have catalyzed the exploration of ECC decoders based on neural networks. Among these, transformer-based neural decoders have achieved state-of-the-art decoding performance. In this paper, we propose a novel Cross-attention Message-Passing Transformer (CrossMPT), which shares key operational principles with conventional message-passing decoders. While conventional transformer-based decoders employ self-attention mechanism without distinguishing between the types of input vectors (i.e., magnitude and syndrome vectors), CrossMPT updates the two types of input vectors separately and iteratively using two masked cross-attention blocks. The mask matrices are determined by the code's parity-check matrix, which explicitly captures the irrelevant relationship between two input vectors. Our experimental results show that CrossMPT significantly outperforms existing neural network-based decoders for various code classes. Notably, CrossMPT achieves this decoding performance improvement, while significantly reducing the memory usage, complexity, inference time, and training time.", "keywords": [ "Cross-attention", "Error correcting codes", "Message-passing decoder", "Neural decoder", "Transformer" ], "rating_list": [ 6, 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 4 ], "confidence_list": [ 5, 5, 5, 5, 4 ] }, { "paper_id": "gFUomIaycw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 1, "author_info_list": [ { "name": "Ashish Dubey", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Principal Researcher" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 15, "n_ref": 22, "n_ref_all": 34, "n_fig": 3, "n_tab": 7, "L_tab": 1194, "n_element_tab": 145, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 812, "n_element_tab_1": 90, "formula_len_all": 658, "formula_len_all_1": 643, "len_all": 75017, "len_all_1": 48730, "len_abs": 1625, "len_title": 81, "len_sents": 21687, "len_sents_1": 20310, "n_sents": 176, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1633, "title": "Dynamic Routing Mixture of Experts for Enhanced Multi-Label Image Classification", "abs": "Multi-label image classification (MLC) is a fundamental task in computer vision, requiring the identification of multiple objects or attributes within a single image. Traditional approaches often rely on shared backbones and static gating mecha-nisms, which can struggle to effectively capture complex label correlations and handle label heterogeneity, leading to issues such as negative transfer. In this pa-per, we introduce the Dynamic Routing Mixture of Experts (DR-MoE) model, a novel architecture that integrates input-dependent dynamic gating networks into the mixture-of-experts (MoE) framework for MLC. Unlike static gating in exist-ing models like the Hybrid Sharing Query (HSQ) Yin et al. (2024), our dynamic gating mechanism adaptively selects and weights both shared and task-specific experts based on the input image features. This allows DR-MoE to better capture varying label dependencies and mitigate negative transfer, resulting in improved overall and per-label classification performance. We conduct extensive experi-ments on benchmark datasets MS-COCO Lin et al. (2014) and PASCAL VOC 2007 Everingham et al. (2015), demonstrating that DR-MoE achieves state-of-the-art results, outperforming existing methods including HSQ, Q2L Liu et al.(2021), and ML-GCN Chen et al. (2019). Additionally, ablation studies confirm the effectiveness of dynamic gating in enhancing model adaptability and perfor-mance, particularly for labels with high heterogeneity. Our findings suggest that incorporating dynamic routing mechanisms into MoE architectures is a promising direction for advancing multi-label image classification.", "keywords": [ "Multi-label image classification", "Dynamic Routing Mixture of Experts", "Computer vision", "Dynamic gating networks", "Label heterogeneity" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "gEEJBXktQM", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Noseong Park", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Sheo yon Jhin", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 7, "n_ref_uni": 25, "n_ref": 34, "n_ref_all": 52, "n_fig": 8, "n_tab": 16, "L_tab": 12948, "n_element_tab": 1867, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 4280, "n_element_tab_1": 264, "formula_len_all": 1089, "formula_len_all_1": 463, "len_all": 161351, "len_all_1": 66045, "len_abs": 1021, "len_title": 86, "len_sents": 46501, "len_sents_1": 28372, "n_sents": 323, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1024, "title": "Enhancing Time-Series Forecasting with Iterative Decomposition and Separable Training", "abs": "Time series data, crucial for decision-making in fields like finance and healthcare, often presents challenges due to its inherent complexity, exacerbating the bias-variance tradeoff and leading to overfitting and underfitting in conventional forecasting models. While promising, state-of-the-art models like PatchTST, iTransformer, and DLinear are hindered by this tradeoff, limiting their ability to separate predictable patterns from noise. To resolve this, we propose the IDEAS framework, which reduces the bias-variance tradeoff to help models achieve optimal performance. IDEAS combines iterative residual decomposition, which reduces bias by extracting predictable patterns, and separable training, which reduces variance by independently optimizing each component. We provide theoretical proof and demonstrate through experiments that IDEAS significantly improves performance across four state-of-the-art models on nine complex benchmark datasets, offering a more robust solution for complex time series forecasting.", "keywords": [ "time series forecasting", "training method" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "gDcL7cgZBt", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Ivona Brandi\u0107", "gender": "Female", "institution": "Technische Universit\u00e4t Wien", "country": "", "position": "Full Professor" }, { "name": "Sabrina Herbst", "gender": "Female", "institution": "Technische Universit\u00e4t Wien", "country": "AT", "position": "PhD student" }, { "name": "Sandeep Suresh Cranganore", "gender": "Male", "institution": "Forschungszentrum Juelich GmbH", "country": "", "position": "PhD student" }, { "name": "Vincenzo De Maio", "gender": "Male", "institution": "Technische Universit\u00e4t Wien", "country": "AT", "position": "Postdoc" } ], "n_formula": 29, "n_formula_1": 7, "n_ref_uni": 58, "n_ref": 82, "n_ref_all": 113, "n_fig": 6, "n_tab": 1, "L_tab": 177, "n_element_tab": 24, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2073, "formula_len_all_1": 677, "len_all": 160457, "len_all_1": 61549, "len_abs": 1431, "len_title": 155, "len_sents": 45129, "len_sents_1": 30113, "n_sents": 324, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 1436, "title": "Exploring channel distinguishability in local neighborhoods of the model space in quantum neural networks", "abs": "With the increasing interest in Quantum Machine Learning, Quantum Neural Networks (QNNs) have emerged and gained significant attention. These models have, however, been shown to be notoriously difficult to train, which we hypothesize is partially due to the architectures, called ansatzes, that are hardly studied at this point. Therefore, in this paper, we take a step back and analyze ansatzes. We initially consider their expressivity, i.e., the space of operations they are able to express, and show that the closeness to being a 2-design, the primarily used measure, fails at capturing this property. Hence, we look for alternative ways to characterize ansatzes, unrelated to expressivity, by considering the local neighborhood of the model space, in particular, analyzing model distinguishability upon small perturbation of parameters. We derive an upper bound on their distinguishability, showcasing that QNNs using the Hardware Efficient Ansatz with few parameters are hardly discriminable upon update. Our numerical experiments support our bounds and further indicate that there is a significant degree of variability, which stresses the need for warm-starting or clever initialization. Altogether, our work provides an ansatz-centric perspective on training dynamics and difficulties in QNNs, ultimately suggesting that iterative training of small quantum models may not be effective, which contrasts their initial motivation.", "keywords": [ "Quantum Machine Learning", "Quantum Neural Network" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 2, 3, 4 ] }, { "paper_id": "gDZd8UGaxS", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "DEEPAK KANDEL", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Dimah Dera", "gender": "Female", "institution": "Rochester Institute of Technology", "country": "US", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 20, "n_ref_uni": 20, "n_ref": 52, "n_ref_all": 68, "n_fig": 4, "n_tab": 7, "L_tab": 2968, "n_element_tab": 354, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1929, "n_element_tab_1": 138, "formula_len_all": 1185, "formula_len_all_1": 1158, "len_all": 132693, "len_all_1": 60078, "len_abs": 1468, "len_title": 140, "len_sents": 28028, "len_sents_1": 26819, "n_sents": 188, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1474, "title": "ProCEED: Prototype Consolidation and Ensemble-based Exemplar-Free Deep Incremental Learning", "abs": "Exemplar-free Class Incremental Learning requires the learning agent to incrementally acquire new class information and maintain past knowledge without having access to samples from previous tasks. Despite the significant performance achieved by the subspace ensemble of a mixture of experts (MoE) with Gaussian prototypical networks, a critical gap still exists. As the downstream tasks arrive, the subspace representation of old classes gets updated, resulting in a prototype drift and leading to forgetting. To address the forgetting problem, we propose ProCEED to dynamically realign previous classes' representation in the latest subspace to adjust the drifted class prototypes and preserve their decision boundaries. Specifically, we compute the inter-subspace angular drifts of the prototype of previous incremental stages with the current one, holding the local semantic relationship between the incremental subspaces. The angular drift is then used to adjust old tasks' prototypes into the subspace of incremental tasks. Furthermore, the model inherits combined knowledge from MoE, supporting plasticity without extra computational burden. Consequently, ProCEED significantly balances the stability-plasticity dilemma over incoming incremental tasks, allowing the model to learn continually. The experimental evaluations on challenging benchmark datasets demonstrate dominant accuracy for ProCEED compared to the state-of-the-art class-incremental learning methods.", "keywords": [ "Prototype Consolidation", "Catastrophic Forgetting", "Incremental Learning", "Mixture-of-Experts" ], "rating_list": [ 1, 3, 3, 5 ], "soundness_list": [ 1, 2, 1, 3 ], "presentation_list": [ 1, 2, 1, 1 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "gDWkImLIKd", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Aashish Yadavally", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Gauthier Guinet", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Hoan Nguyen", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Laurent Callot", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 53, "n_ref_all": 65, "n_fig": 3, "n_tab": 5, "L_tab": 1573, "n_element_tab": 147, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1573, "n_element_tab_1": 147, "formula_len_all": 19, "formula_len_all_1": 0, "len_all": 123987, "len_all_1": 62988, "len_abs": 1346, "len_title": 122, "len_sents": 34221, "len_sents_1": 29657, "n_sents": 244, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1354, "title": "Large Language Model Critics for Execution-Free Evaluation of Code Changes", "abs": "Large language models (LLMs) offer a promising way forward for automating software engineering tasks, such as bug fixes, feature additions, etc., via multi-step LLM-based agentic workflows. However, existing metrics for evaluating such workflows, mainly build status and occasionally log analysis, are too sparse and limited in providing the information needed to assess the quality of changes made. In this work, we designed LLM-based critics to derive well-structured and rigorous intermediate/step-level, execution-free evaluation proxies for repo-level code changes. Importantly, we assume access to the gold patch for the problem (i.e., reference-aware) to assess both semantics and executability of generated patches. With the gold test patch as a reference, we predict executability of all editing locations with an accuracy of 91.6%, aggregating which, we can predict the build status in 82.1% of the instances in SWE-bench. In particular, such an execution-focused LLM critic outperforms other reference-free and reference-aware LLM critics by 38.9% to 72.5%. Moreover, we demonstrate the usefulness of such a reference-aware framework in comparing patches generated by different agentic workflows. Finally, we open-source the library developed for this project, which allow further usage for either other agentic workflows or other benchmarks.", "keywords": [ "Code Evaluation; Large Language Models; Execution-free Evaluation; Agents" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 4, 5, 5 ] }, { "paper_id": "gCYFtUKXSc", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Abhinab Acharya", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Dayou Yu", "gender": "unknown", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Qi Yu", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "Professor" }, { "name": "Xumin Liu", "gender": "Female", "institution": "Rochester Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Yuansheng Zhu", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 36, "n_ref": 83, "n_ref_all": 108, "n_fig": 12, "n_tab": 7, "L_tab": 2290, "n_element_tab": 170, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 504, "n_element_tab_1": 40, "formula_len_all": 858, "formula_len_all_1": 447, "len_all": 158605, "len_all_1": 76880, "len_abs": 1575, "len_title": 112, "len_sents": 52404, "len_sents_1": 37879, "n_sents": 396, "n_sents_1": 273, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1580, "title": "Avoid Being a Shortcut Learner through Library-Based Re-Learning", "abs": "Replay-based methods provide a promising solution to address catastrophic forgetting issue in continual learning. They try to retain previous knowledge by using a small amount of data from previous tasks stored in a fix-sized buffer. In this work, we invoke the information bottleneck principles and reveal some fundamental limitations of those methods on their effectiveness in capturing the truly important features from the prior tasks by relying on the buffer data selected according to the model's performance on known tasks. Since future tasks are not accessible during model training and buffer construction, the trained model and the buffer data tend to be biased towards making accurate predictions on the labels of known tasks. However, when new task samples are introduced along with labels, the biased model and the buffer data become less effective in differentiating samples of the old tasks from those of the new ones. Inspired by the way humans learn over time, we propose a novel relearning technique that makes use of additional past data, referred to as the library, to test how much information the model loses after learning the new task. We then realign the model towards those forgotten samples by training on a carefully selected small subset samples from the library for a few epochs with comparable computational cost as existing replay-based models. The experimental results on multiple real-world datasets demonstrate that the proposed relearning process can improve the performance of the state-of-the-art continual learning methods by a large margin.", "keywords": [ "Continual learning", "Data-Efficient Learning", "Information Theory" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "gCSEQIgbWH", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Aron Laszka", "gender": "Not Specified", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Ayan Mukhopadhyay", "gender": "unknown", "institution": "Vanderbilt University", "country": "US", "position": "Researcher" }, { "name": "Iliyas Bektas", "gender": "Male", "institution": "Pennsylvania State University", "country": "", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 23, "n_ref": 65, "n_ref_all": 78, "n_fig": 5, "n_tab": 3, "L_tab": 2251, "n_element_tab": 324, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 905, "formula_len_all_1": 777, "len_all": 136855, "len_all_1": 71669, "len_abs": 1383, "len_title": 162, "len_sents": 49814, "len_sents_1": 33156, "n_sents": 332, "n_sents_1": 230, "venue": "ICLR2025", "decision": "withdraw", "L_title": 114, "L_abs": 1369, "title": "Generalist Policy for k-Server Problem on Graphs using Deep Reinforcement Learning with Action-Value Decomposition", "abs": "The online $k$-server problem on graphs is a fundamental computational problem that can model a wide range of practical problems, such as dispatching ambulances to serve accidents or dispatching taxis to serve ride requests. While most prior work on the $k$-server problem focused on online algorithms, reinforcement learning promises policies that require low computational effort during execution, which is critical in time-sensitive applications, such as ambulance dispatch. However, there exists no scalable reinforcement-learning approach for the $k$-server problem. To address this gap, we introduce a scalable computational approach for learning generalist policies. Besides scalability, the advantage of generalist policies is transferability: a generalist policy can be applied to an entire class of graphs without the need for retraining, which is crucial for practical applications, e.g., in ambulance dispatch problems where road conditions or demand distributions may change over time. We achieve scalability and transferability by introducing a novel architecture that decomposes the action-value into a global and a local term, estimated from a shared graph-convolution backbone. We evaluate our approach on a variety of graph classes, comparing to well-established baselines, demonstrating the performance and transferability of our generalist policies.", "keywords": [ "reinforcement learning", "graph neural network", "k-server problem", "deep learning", "graph convolution", "transportation" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "gCCYBGGYTi", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Shilin Zhang", "gender": "Male", "institution": "North China University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "YAN MING", "gender": "unknown", "institution": "North China University of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 18, "n_ref": 30, "n_ref_all": 44, "n_fig": 7, "n_tab": 5, "L_tab": 2366, "n_element_tab": 260, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 600, "n_element_tab_1": 109, "formula_len_all": 524, "formula_len_all_1": 450, "len_all": 150532, "len_all_1": 58703, "len_abs": 1618, "len_title": 119, "len_sents": 53871, "len_sents_1": 28217, "n_sents": 365, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1505, "title": "Contextual Kernels for Task-Aware Fine-Tuning in Vision-Language Models", "abs": "Vision-Language Models (VLMs) exhibit impressive generalization due to training on vast datasets like ImageNet. However, their performance diminishes on unfamiliar tasks. While downstream fine-tuning enhances adaptability, it often sacrifices inherent generality. To address this, we propose a novel method leveraging contextual generation for enhanced task representation within a semantic space. Our approach utilizes VLMs to generate detailed contextual descriptions for test image batches, developing Contextual Kernels (CK) for each class in the semantic space. \nOur test-time fine-tuning preserves core VLM features by freezing fundamental components and extending a linear network for semantic kernel density projection. This strategy significantly boosts model adaptability for real-world tasks. Despite strong zero-shot capabilities, we explore additional training samples to improve adaptability in dynamic Task Incremental Learning (TIL) scenarios. Each task's unique CK distribution serves as a fingerprint, enabling high-performance TIL with minimal forgetting. Experiments on four TIL datasets demonstrate the efficacy of our framework, achieving state-of-the-art performance. Our findings reveal that the semantic space within the text mode encapsulates both VLMs' generality and adaptability, paving the way for robust applications in diverse, evolving task environments. This work systematically balances generality and adaptability in VLMs, addressing a critical gap in current research.", "keywords": [ "Continual Learning", "Model Adaptability", "Task Incremental Learning", "Kernel Based Task Reprresentation Learning" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "gC0ikdZoz8", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Hidenori Tanaka", "gender": "unknown", "institution": "Harvard University, Harvard University", "country": "US", "position": "Researcher" }, { "name": "Rhys Gould", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Undergrad student" } ], "n_formula": 61, "n_formula_1": 25, "n_ref_uni": 36, "n_ref": 87, "n_ref_all": 169, "n_fig": 20, "n_tab": 2, "L_tab": 156, "n_element_tab": 8, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5160, "formula_len_all_1": 1991, "len_all": 149769, "len_all_1": 62779, "len_abs": 1378, "len_title": 68, "len_sents": 47626, "len_sents_1": 27743, "n_sents": 382, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1389, "title": "Continuous-Time Analysis of Adaptive Optimization and Normalization", "abs": "Adaptive optimization algorithms, particularly Adam and its variant AdamW, are fundamental to modern deep learning, however, their training dynamics lack comprehensive theoretical understanding, with limited insight into why common practices\u2014such as specific hyperparameter choices and normalization layers\u2014contribute to successful generalization. This work presents a continuous-time formulation of Adam and AdamW, facilitating a tractable analysis of training dynamics that can shed light on such practical questions. We theoretically derive a stable region for Adam's hyperparameters $(\\beta, \\gamma)$ that ensures bounded updates, empirically verifying these predictions by observing unstable exponential growth of parameter updates outside this region. Furthermore, we theoretically justify the success of normalization layers by uncovering an implicit meta-adaptive effect of scale-invariant architectural components. This insight leads to an explicit optimizer, $2$-Adam, which we generalize to $k$-Adam\u2014an optimizer that applies an adaptive normalization procedure $k$ times, encompassing Adam (corresponding to $k=1$) and Adam with a normalization layer (corresponding to $k=2$). Overall, our continuous-time formulation of Adam facilitates a principled analysis, offering deeper understanding of optimal hyperparameter choices and architectural decisions in modern deep learning.", "keywords": [ "Theory of Deep Learning", "Adaptive Optimization", "Continuous-Time Analysis", "Normalization" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "gBT6rAEqvx", "primary_area": "learning theory", "n_author": 1, "author_info_list": [ { "name": "Zhuang Yang", "gender": "Male", "institution": "Soochow University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 77, "n_formula_1": 49, "n_ref_uni": 50, "n_ref": 73, "n_ref_all": 98, "n_fig": 5, "n_tab": 2, "L_tab": 417, "n_element_tab": 23, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 11, "n_element_tab_1": 2, "formula_len_all": 5063, "formula_len_all_1": 2806, "len_all": 135588, "len_all_1": 56703, "len_abs": 1519, "len_title": 94, "len_sents": 36304, "len_sents_1": 23467, "n_sents": 312, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 45, "L_abs": 1552, "title": "Adaptive Second-Order Stochastic Optimization", "abs": "As a much possible way of improving first-order stochastic optimization ($\\mathcal{FSO}$), the role of second-order information in stochastic optimization is receiving an increasing attention especially for solving the model with large-scale datasets in recent years, resulting in various second-order stochastic optimization ($\\mathcal{SSO}$) methods, e.g., the stochastic Newton (SN) method, the stochastic quasi-Newton (SQN) method, etc. However, the question of how to set an appropriate update rule of the learning rate for SSO methods is still an extremely intractable task, and surprisingly there is quite less literature to tackle this issue. To bridge the gap between the SSO methods and the learning rate, this work develops a class of adaptive SSO methods from the perspective of adaptive gradient methods. Concretely, a general adaptive gradient (GAG) method with the quasi-hyperbolic momentum (QHM) strategy that encompasses Adam, AdaGrad, RMSProp, etc., as the special case of GAG, is incorporated into SN and SQN, respectively, which leads to two methods: SN-GAG and SQN-GAG. In addition, we establish a unified analysis for different adaptive SSO methods, covering their convergence behavior and computational complexity for different backgrounds, such as the strongly convex (SC) case and the Polyak-{\\L}ojasiewicz (P{\\L}) case, where, particularly, the latter is missing in current studies. Finally, numerical tests on different applications of machine learning demonstrate the superiority and the robustness of the resulting methods.", "keywords": [ "MachinSecond-order optimization", "stochastic Newton methods", "stochastic quasi-Newton methods", "adaptive graident descent", "convergence analysis" ], "rating_list": [ 3, 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 1, 2, 3 ], "contribution_list": [ 3, 3, 1, 1, 2 ], "confidence_list": [ 3, 3, 4, 5, 4 ] }, { "paper_id": "gBHZAAwcgT", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Changhe Tu", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "Full Professor" }, { "name": "Hao Pan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jiepeng Wang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Rui Xu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shiqing Xin", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "Full Professor" }, { "name": "Taku Komura", "gender": "Male", "institution": "the University of Hong Kong, University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Wenping Wang", "gender": "Male", "institution": "Texas A&M University ", "country": "US", "position": "Full Professor" }, { "name": "Xin Tong", "gender": "Male", "institution": "Anuttacon", "country": "US", "position": "Principal Researcher" }, { "name": "Yang Liu", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 5, "n_ref_uni": 25, "n_ref": 54, "n_ref_all": 88, "n_fig": 16, "n_tab": 3, "L_tab": 427, "n_element_tab": 76, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 500, "n_element_tab_1": 56, "formula_len_all": 774, "formula_len_all_1": 183, "len_all": 120749, "len_all_1": 55132, "len_abs": 998, "len_title": 118, "len_sents": 44032, "len_sents_1": 26844, "n_sents": 317, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1001, "title": "ComboStoc: Combinatorial Stochasticity for Diffusion Generative Models", "abs": "In this paper, we study an under-explored but important factor of diffusion generative models, i.e., the combinatorial complexity. \nData samples are generally high-dimensional, and for various structured generation tasks, additional attributes are combined to associate with data samples.\nWe show that the space spanned by the combination of dimensions and attributes is insufficiently sampled by existing training scheme of diffusion generative models, causing degraded test time performance.\nWe present a simple fix to this problem by constructing stochastic processes that fully exploit the combinatorial structures, hence the name ComboStoc.\nUsing this simple strategy, we show that network training is significantly accelerated across diverse data modalities, including images and 3D structured shapes.\nMoreover, ComboStoc enables a new way of test time generation which uses asynchronous time steps for different dimensions and attributes, thus allowing for varying degrees of control over them.", "keywords": [ "generative models", "combinatorial", "stochastic" ], "rating_list": [ 8, 5, 6, 5, 3 ], "soundness_list": [ 4, 2, 3, 3, 2 ], "presentation_list": [ 4, 2, 2, 3, 1 ], "contribution_list": [ 4, 3, 3, 2, 2 ], "confidence_list": [ 5, 3, 4, 3, 3 ] }, { "paper_id": "gB2ZeqDpl6", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Jialin Wu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Kaixiong Zhou", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Assistant Professor" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Xinnan Zhang", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" }, { "name": "JUNYI XIE", "gender": "Female", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 58, "n_ref": 161, "n_ref_all": 183, "n_fig": 5, "n_tab": 9, "L_tab": 10659, "n_element_tab": 1381, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2707, "n_element_tab_1": 126, "formula_len_all": 569, "formula_len_all_1": 446, "len_all": 200954, "len_all_1": 67671, "len_abs": 1513, "len_title": 123, "len_sents": 45660, "len_sents_1": 30955, "n_sents": 314, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1519, "title": "Benchmark on Drug Target Interaction Modeling from a Structure Perspective", "abs": "The prediction modeling of drug-target interactions is crucial to drug discovery and design, which has seen rapid advancements owing to deep learning technologies. Recently developed methods, such as those based on graph neural networks (GNNs) and Transformers, demonstrate exceptional performance across various datasets by effectively extracting structural information. However, the benchmarking of these novel methods often varies significantly in terms of hyperparameter settings and datasets, which limits algorithmic progress. In view of these, we conduct a comprehensive survey and benchmark for drug-target interaction modeling from a structure perspective, via integrating tens of explicit (i.e., GNN-based) and implicit (i.e., Transformer-based) structure learning algorithms. To this end, we first unify the hyperparameter setting within each class of structure learning methods. Moreover, we conduct a macroscopical comparison between these two classes of encoding strategies as well as the different featurization techniques that inform molecules' chemical and physical properties. We then carry out the microscopical comparison between all the integrated models across the six datasets, via comprehensively benchmarking their effectiveness and efficiency. Remarkably, the summarized insights from the benchmark studies lead to the design of model combos. We demonstrate that our combos can achieve new state-of-the-art performance on various datasets associated with cost-effective memory and computation.", "keywords": [ "drug-target interaction", "benchmark", "transformer", "GNN" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "gAEEjGv5Oa", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "David Rein", "gender": "Not Specified", "institution": "New York University", "country": "", "position": "Researcher" }, { "name": "Julian Michael", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "Samuel Arnesen", "gender": "Male", "institution": "New York University", "country": "US", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 2, "n_ref_uni": 36, "n_ref": 108, "n_ref_all": 132, "n_fig": 9, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 698, "formula_len_all_1": 180, "len_all": 332472, "len_all_1": 62998, "len_abs": 739, "len_title": 126, "len_sents": 131640, "len_sents_1": 31565, "n_sents": 1211, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 740, "title": "Training Language Models to Win Debates with Self-Play Improves Judge Accuracy", "abs": "We test the robustness of debate as a method of scalable oversight by training models to debate with data generated via self-play. In a long-context reading comprehension task, we find that language model based evaluators answer questions more accurately when judging models optimized to win debates. By contrast, we find no such relationship for consultancy models trained to persuade a judge without an opposing debater present. In quantitative and qualitative comparisons between our debate models and novel consultancy baselines, we find evidence that debate training encourages stronger and more informative arguments, showing promise that it can help provide high-quality supervision for tasks that are difficult to directly evaluate.", "keywords": [ "AI Safety", "Scalable Oversight", "Reinforcement Learning", "Debate" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "g90RNzs8wX", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Chenyu Li", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "Kaize Ding", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "Assistant Professor" }, { "name": "Rui Miao", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "PhD student" }, { "name": "Shirui Pan", "gender": "unknown", "institution": "Griffith University", "country": "AU", "position": "Full Professor" }, { "name": "xin wang", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Associate Professor" }, { "name": "Xu Shen", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "Yili Wang", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "PhD student" }, { "name": "Ying Wang", "gender": "Female", "institution": "Jilin University", "country": "CN", "position": "Full Professor" }, { "name": "Yixin Liu", "gender": "Male", "institution": "Griffith University", "country": "AU", "position": "Postdoc" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 43, "n_ref": 122, "n_ref_all": 163, "n_fig": 11, "n_tab": 10, "L_tab": 10583, "n_element_tab": 1608, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 4988, "n_element_tab_1": 825, "formula_len_all": 2896, "formula_len_all_1": 1021, "len_all": 226909, "len_all_1": 75792, "len_abs": 1624, "len_title": 146, "len_sents": 60980, "len_sents_1": 28175, "n_sents": 505, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 98, "L_abs": 1255, "title": "Unifying Unsupervised Graph-Level Anomaly Detection and Out-of-Distribution Detection: A Benchmark", "abs": "To build safe and reliable graph machine learning systems, unsupervised graph-level anomaly detection (GLAD) and unsupervised graph-level out-of-distribution (OOD) detection (GLOD) have received significant attention in recent years. Though these two lines of research share the same objective, they have been studied independently in the community due to distinct evaluation setups, creating a gap that hinders the application and evaluation of methods from one to the other. To bridge the gap, in this work, we present a Unified Benchmark for unsupervised Graph-level OOD and anomaly Detection (UB-GOLD), a comprehensive evaluation framework that unifies GLAD and GLOD under the concept of generalized graph-level OOD detection. Our benchmark encompasses 35 datasets spanning four practical anomaly and OOD detection scenarios, facilitating the comparison of 18 representative GLAD/GLOD methods. We conduct multi-dimensional analyses to explore the effectiveness, generalizability, robustness, and efficiency of existing methods, shedding light on their strengths and limitations. Furthermore, we provide an open-source codebase of UB-GOLD to foster reproducible research and outline potential directions for future investigations based on our insights.", "keywords": [ "Graph out-of-distribution detection; graph anomaly detection; benchmark" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "g8TF3gd01u", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 10, "author_info_list": [ { "name": "Hongyan Fei", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jiapei Zhang", "gender": "Male", "institution": "WeChat AI", "country": "CN", "position": "Researcher" }, { "name": "Jie Zhou", "gender": "Male", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Jinchao Zhang", "gender": "Male", "institution": "WeChat AI", "country": "CN", "position": "Researcher" }, { "name": "Xiaoyue Duan", "gender": "Male", "institution": "WeChat AI, Tencent Inc", "country": "CN", "position": "Algorithm engineer" }, { "name": "Yeshuang Zhu", "gender": "unknown", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Researcher" }, { "name": "Ying Deng", "gender": "Female", "institution": "WechatAI", "country": "CN", "position": "Researcher" }, { "name": "Yuan Zhiqiang", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Zexi Jia", "gender": "Male", "institution": "Tencent Inc. ", "country": "CN", "position": "Researcher" }, { "name": "Chuanwei Huang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 9, "n_ref": 14, "n_ref_all": 29, "n_fig": 7, "n_tab": 4, "L_tab": 1276, "n_element_tab": 171, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2476, "n_element_tab_1": 191, "formula_len_all": 703, "formula_len_all_1": 612, "len_all": 94421, "len_all_1": 47086, "len_abs": 1130, "len_title": 119, "len_sents": 22720, "len_sents_1": 20597, "n_sents": 150, "n_sents_1": 125, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1136, "title": "Comprehensive Artistic Style Representation for Quantitative Evaluation", "abs": "Artistic style, a unique medium for artists to express creativity through elements like form, color, and composition, poses a challenge for computer vision due to its intricate patterns and nuanced aesthetics. Contemporary models, often reliant on specific datasets, face limitations in their generalizability and precision in identifying individual artists' styles. From an information theory perspective, we examine the limitations of fine-tuning and investigate techniques to disentangle content from style information. We note differences in artistic style representation between unimodal and multimodal models. As a result, we propose a plug-and-play approach designed to efficiently separate content information within Vision-Language Models (VLMs), preserving stylistic details. Furthermore, we present the WeART dataset, a large-scale art dataset with high-quality annotations, to evaluate the artistic style representation capabilities of models. Experimental results show that our method improves the performance of VLMs in style retrieval tasks across several datasets. We will publicly release the proposed dataset and code.", "keywords": [ "Visual Representation", "Representation Decoupling", "Artistic Style", "Vision-Language Models" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 3, 1, 2, 1 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "g7xZkiHcGO", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Dejie Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jun Luo", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "LingjunChen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Qingchao Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yang Liu", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuxin Peng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Zijing Zhao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 19, "n_ref": 28, "n_ref_all": 54, "n_fig": 4, "n_tab": 7, "L_tab": 4481, "n_element_tab": 171, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1002, "n_element_tab_1": 42, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 97356, "len_all_1": 50396, "len_abs": 1460, "len_title": 105, "len_sents": 29389, "len_sents_1": 24682, "n_sents": 196, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1450, "title": "Investigating Domain Gaps for Indoor 3D Object Detection", "abs": "As a fundamental task for indoor scene understanding, 3D object detection has been extensively studied, and the accuracy on indoor point cloud data has been substantially improved. However, existing researches have been conducted on limited datasets, where the training and testing sets share the same distribution. In this paper, we consider the task of adapting indoor 3D object detectors from one dataset to another, presenting a first benchmark with commonly used ScanNet and SUN RGB-D datasets, as well as our newly proposed large-scale SimRoom and SimHouse datasets by a 3D simulator with far greater number of objects and more precise annotations. Since indoor point cloud datasets are collected and constructed in different ways, the object detectors are likely to overfit to specific factors within each dataset, such as point cloud quality, room layout configuration, style and object size. We conduct experiments across datasets on different adaptation scenarios, analyzing the impact of different domain gaps on 3D object detectors. We observe that through our evaluated domain gap factors, synthetic-to-real adaptation is the most difficult adaptation hurdle to overcome. We also introduce several domain adaptation approaches to improve adaptation performances, providing a first baseline for domain adaptive indoor 3D object detection, hoping that future works may propose detectors with stronger generalization ability across domains.", "keywords": [ "domain adaptation", "indoor 3D object detection" ], "rating_list": [ 6, 5, 6, 3, 5 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "g7DHM6MRE4", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Eric Peter Wairagala", "gender": "Male", "institution": "Makerere University", "country": "", "position": "MS student" }, { "name": "Joyce Nakatumba-Nabende", "gender": "Female", "institution": "Makerere University", "country": "UG", "position": "Senior Lecturer" }, { "name": "Rose Nakasi", "gender": "unknown", "institution": "Makerere University", "country": "", "position": "Lecturer" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 19, "n_ref": 20, "n_ref_all": 25, "n_fig": 2, "n_tab": 4, "L_tab": 740, "n_element_tab": 83, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 740, "n_element_tab_1": 83, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 66309, "len_all_1": 51417, "len_abs": 1377, "len_title": 115, "len_sents": 23121, "len_sents_1": 23122, "n_sents": 162, "n_sents_1": 162, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1416, "title": "Building Luganda Machine Translation models for the Medical Domain", "abs": "Globalization and migration have highlighted the critical need for effective cross-language communication, particularly in healthcare. In Uganda, a multilingual nation where Luganda is widely spoken, language barriers in predominantly English-speaking medical settings often lead to misunderstandings, misdiagnoses, and compromised patient care. This study aims to mitigate these issues by developing an machine translation model built for medical communication, specifically targeting translations from English to Luganda within the context of malaria diagnosis and community engagement. Utilizing recent advancements in Artificial Intelligence and unsupervised learning, this research involves curating a parallel medical corpus, training a transformer-based model with domain-specific adapters, and rigorously evaluating the model's accuracy and cultural sensitivity.\n\nThe results demonstrate that the MarianMT-Adapter LoRa model, when combined with active learning, achieved a significant improvement in translation quality, evidenced by a BLEU score increase to 56. This model effectively reduced translation errors and preserved the contextual integrity of medical texts. The findings are anticipated to enhance healthcare communication, reduce disparities, and improve access to medical knowledge for Luganda-speaking communities, providing a blueprint for similar efforts in other multilingual environments.", "keywords": [ "machine translation", "luganda", "Low-resourced languages", "medical machine translation" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "g6v09VxgFw", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Adarsh Jamadandi", "gender": "Male", "institution": "Saarland University", "country": "", "position": "MS student" }, { "name": "Celia Rubio-Madrigal", "gender": "Female", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "PhD student" }, { "name": "Rebekka Burkholz", "gender": "Female", "institution": "Helmholtz Center CISPA for Information Security", "country": "DE", "position": "Associate Professor" } ], "n_formula": 35, "n_formula_1": 6, "n_ref_uni": 50, "n_ref": 108, "n_ref_all": 152, "n_fig": 14, "n_tab": 19, "L_tab": 11042, "n_element_tab": 1273, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4244, "formula_len_all_1": 631, "len_all": 228763, "len_all_1": 64365, "len_abs": 1206, "len_title": 116, "len_sents": 60414, "len_sents_1": 30054, "n_sents": 568, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1162, "title": "GNNs Getting ComFy: Community and Feature Similarity Guided Rewiring", "abs": "Maximizing the spectral gap through graph rewiring has been proposed to enhance the performance of message-passing graph neural networks (GNNs) by addressing over-squashing. However, as we show, minimizing the spectral gap can also improve generalization. To explain this, we analyze how rewiring can benefit GNNs within the context of stochastic block models. Since spectral gap optimization primarily influences community strength, it improves performance when the community structure aligns with node labels. Building on this insight, we propose three distinct rewiring strategies that explicitly target community structure, node labels, and their alignment: (a) community structure-based rewiring (ComMa), a more computationally efficient alternative to spectral gap optimization that achieves similar goals; (b) feature similarity-based rewiring (FeaSt), which focuses on maximizing global homophily; and (c) a hybrid approach (ComFy), which enhances local feature similarity while preserving community structure to optimize label-community alignment. Extensive experiments confirm the effectiveness of these strategies and support our theoretical insights.", "keywords": [ "graph neural networks", "over-squashing", "graph rewiring", "community structure", "homophily", "feature similarity" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "g6syfIrVuS", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Rio Yokota", "gender": "Male", "institution": "Institute of Science Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Ryo Karakida", "gender": "Male", "institution": "AIST, National Institute of Advanced Industrial Science and Technology", "country": "", "position": "Researcher" }, { "name": "Satoki Ishikawa", "gender": "Male", "institution": "Institute of Science Tokyo", "country": "JP", "position": "MS student" } ], "n_formula": 139, "n_formula_1": 32, "n_ref_uni": 46, "n_ref": 150, "n_ref_all": 177, "n_fig": 7, "n_tab": 6, "L_tab": 1849, "n_element_tab": 273, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 14200, "formula_len_all_1": 2252, "len_all": 231137, "len_all_1": 71335, "len_abs": 1292, "len_title": 171, "len_sents": 68135, "len_sents_1": 31124, "n_sents": 662, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 123, "L_abs": 1296, "title": "Local Loss Optimization in the Infinite Width: Stable Parameterization of Predictive Coding Networks and Target Propagation", "abs": "Local learning, which trains a network through layer-wise local targets and losses, has been studied as an alternative to backpropagation (BP) in neural computation. However, its algorithms often become more complex or require additional hyperparameters due to the locality, making it challenging to identify desirable settings where the algorithm progresses in a stable manner.\nTo provide theoretical and quantitative insights, we introduce maximal update parameterization ($\\mu$P) in the infinite-width limit for two representative designs of local targets: predictive coding (PC) and target propagation (TP). We verify that $\\mu$P enables hyperparameter transfer across models of different widths.\nFurthermore, our analysis reveals unique and intriguing properties of $\\mu$P that are not present in conventional BP. By analyzing deep linear networks, we find that PC's gradients interpolate between first-order and Gauss-Newton-like gradients, depending on the parameterization. \nWe demonstrate that, in specific standard settings, PC in the infinite-width limit behaves more similarly to the first-order gradient.\nFor TP, even with the standard scaling of the last layer differing from classical $\\mu$P, its local loss optimization favors the feature learning regime over the kernel regime.", "keywords": [ "deep learning", "feature learning", "local learning", "predictive coding", "target propagation", "infinite width", "maximal update parameterization (muP)" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 2, 2, 1 ] }, { "paper_id": "g6iiIUvhko", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Chen Yurou", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Lu Zhang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Zhengwei Li", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zhenyang Lin", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zhi-yong Liu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 25, "n_ref": 40, "n_ref_all": 59, "n_fig": 8, "n_tab": 9, "L_tab": 9235, "n_element_tab": 627, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 722, "n_element_tab_1": 115, "formula_len_all": 763, "formula_len_all_1": 704, "len_all": 134486, "len_all_1": 62087, "len_abs": 1504, "len_title": 163, "len_sents": 43396, "len_sents_1": 30174, "n_sents": 318, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 115, "L_abs": 1509, "title": "Zero-Shot Task-Level Adaptation via Coarse-to-Fine Policy Refinement and Holistic-Local Contrastive Representations", "abs": "Meta-reinforcement learning offers a mechanism for zero-shot adaptation, enabling agents to handle new tasks with parametric variation in real-world environments. However, existing methods still struggle with task-level adaptation, which demands generalization beyond simple variations within tasks, thereby limiting their practical effectiveness. This limitation stems from several challenges, including the poor task representations and inefficient policy learning, resulting from the underutilization of hierarchical structure inherent in task-level adaptation. To address these challenges, we propose a Coarse-to-Fine Policy Refinement combined with a Holistic-Local Contrastive Representation method to enable effective zero-shot policy adaptation. Specifically, in terms of policy learning, we use task language instructions as prior knowledge to select skill-specific expert modules as a coarse policy. This coarse policy is then refined by a fine policy generated through a hypernetwork, producing a task-aware policy based on task representations. Additionally, for task representation, we employ contrastive learning from both holistic and local perspectives to enhance task representations for more effective policy adaptation. Experimental results demonstrate that our method significantly improves learning efficiency and zero-shot adaptation on new tasks, outperforming previous methods by approximately 42.3% and 45.4% in success rate on the Meta-World ML-10 and ML-45 benchmarks, respectively.", "keywords": [ "Meta-RL", "Zero-shot Task-level Adaptation", "Contrastive Representations" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "g6fYDGKeyB", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 7, "author_info_list": [ { "name": "Antoine Wehenkel", "gender": "Male", "institution": "Apple", "country": "CH", "position": "Researcher" }, { "name": "Guillermo Sapiro", "gender": "unknown", "institution": "Apple", "country": "", "position": "Apple" }, { "name": "Jens Behrmann", "gender": "Male", "institution": "Apple", "country": "DE", "position": "Researcher" }, { "name": "Joern-Henrik Jacobsen", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Juan L. Gamella", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Ozan Sener", "gender": "Male", "institution": "Apple", "country": "DE", "position": "Researcher" }, { "name": "marco cuturi", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 16, "n_ref_uni": 49, "n_ref": 64, "n_ref_all": 83, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1199, "formula_len_all_1": 1114, "len_all": 135882, "len_all_1": 71461, "len_abs": 1196, "len_title": 137, "len_sents": 41179, "len_sents_1": 33949, "n_sents": 266, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1199, "title": "Addressing Misspecification in Simulation-based Inference through Data-driven Calibration", "abs": "Driven by steady progress in generative modeling, simulation-based inference (SBI) has enabled inference over stochastic simulators. However, recent work has demonstrated that model misspecification can harm SBI's reliability, preventing its adoption in important applications where only misspecified simulators are available.\nThis work introduces robust posterior estimation (RoPE), a framework that overcomes model misspecification with a small real-world calibration set of ground truth parameter measurements.\nWe formalize the misspecification gap as the solution of an optimal transport problem between learned representations of real-world and simulated observations, allowing the method to learn a model of the misspecification without placing additional assumptions on its nature. The method shows how a small calibration set can be leveraged to offer a controllable balance between calibrated uncertainty and informative inference even under severely misspecified simulators. Our empirical results on four synthetic tasks and two real-world problems with ground-truth labels demonstrate that RoPE outperforms baselines and consistently returns informative and calibrated credible intervals.", "keywords": [ "Simulation-based inference", "SBI", "Bayesian Inference", "Misspecification", "Likelihood-free", "Robust Inference", "Physics-informed Machine Learning" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "g6Qc3p7JH5", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Jingyi Cui", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Qi Lei", "gender": "Female", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Stefanie Jegelka", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Xiang Pan", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Yifei Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yisen Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "zhang qi", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" } ], "n_formula": 38, "n_formula_1": 7, "n_ref_uni": 46, "n_ref": 102, "n_ref_all": 130, "n_fig": 6, "n_tab": 4, "L_tab": 1563, "n_element_tab": 148, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 833, "n_element_tab_1": 116, "formula_len_all": 4704, "formula_len_all_1": 463, "len_all": 194692, "len_all_1": 65865, "len_abs": 1494, "len_title": 129, "len_sents": 59654, "len_sents_1": 31994, "n_sents": 479, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1459, "title": "Beyond Interpretability: The Gains of Feature Monosemanticity on Model Robustness", "abs": "Deep learning models often suffer from a lack of interpretability due to polysemanticity, where individual neurons are activated by multiple unrelated semantics, resulting in unclear attributions of model behavior. Recent advances in monosemanticity, where neurons correspond to consistent and distinct semantics, have significantly improved interpretability but are commonly believed to compromise accuracy. In this work, we challenge the prevailing belief of the accuracy-interpretability tradeoff, showing that monosemantic features not only enhance interpretability but also bring concrete gains in model performance. Across multiple robust learning scenarios\u2014including input and label noise, few-shot learning, and out-of-domain generalization\u2014our results show that models leveraging monosemantic features significantly outperform those relying on polysemantic features. Furthermore, we provide empirical and theoretical understandings on the robustness gains of feature monosemanticity. Our preliminary analysis suggests that monosemanticity, by promoting better separation of feature representations, leads to more robust decision boundaries. This diverse evidence highlights the generality of monosemanticity in improving model robustness. As a first step in this new direction, we embark on exploring the learning benefits of monosemanticity beyond interpretability, supporting the long-standing hypothesis of linking interpretability and robustness.", "keywords": [ "Representation Learning", "Robustness", "Polysemanticity", "Superposition" ], "rating_list": [ 6, 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 4, 2, 4 ], "contribution_list": [ 2, 2, 3, 3, 4 ], "confidence_list": [ 4, 3, 3, 3, 4 ] }, { "paper_id": "g4VGwNqzpB", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Boyuan Wang", "gender": "Male", "institution": "Lancaster University", "country": "GB", "position": "PhD student" }, { "name": "Richard M. Jiang", "gender": "Male", "institution": "Lancaster University", "country": "GB", "position": "Associate Professor" } ], "n_formula": 26, "n_formula_1": 7, "n_ref_uni": 45, "n_ref": 59, "n_ref_all": 73, "n_fig": 10, "n_tab": 8, "L_tab": 2155, "n_element_tab": 162, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1454, "formula_len_all_1": 343, "len_all": 139218, "len_all_1": 55435, "len_abs": 706, "len_title": 88, "len_sents": 49131, "len_sents_1": 26499, "n_sents": 391, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 40, "L_abs": 735, "title": "HENP: Dynamic Pruning via Neuron Entropy", "abs": "We introduce a novel framework for analyzing neural networks based on the concepts of \\textit{dynamic} and \\textit{static} neurons, which describe the stability of neuron activation under specific inputs. From these concepts, we propose \\textit{neuron entropy} as a metric to quantify network expressiveness. Our analysis reveals that better generalization correlates with diverse activation patterns and higher neuron entropy. Building on this, we propose our HENP method, a dynamic pruning technique that regulates dying neurons and sparsifies the network during training. Experimental results demonstrate that our HENP improves both network sparsity and performance, offering a new approach to efficient neural network optimization.", "keywords": [ "Pruning", "Network Compression", "Dying Neurons", "Network Architecture Search" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 2, 1 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "g3xuCtrG6H", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Atsunobu Kotani", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Ren Ng", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 77, "n_ref": 108, "n_ref_all": 132, "n_fig": 6, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 153, "formula_len_all_1": 153, "len_all": 216128, "len_all_1": 66051, "len_abs": 1633, "len_title": 131, "len_sents": 37431, "len_sents_1": 34143, "n_sents": 213, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 83, "L_abs": 1642, "title": "A Computational Framework for Modeling Emergence of Color Vision in the Human Brain", "abs": "It is a mystery how the brain decodes color vision purely from the optic nerve signals it receives, with a core inferential challenge being how it disentangles internal perception with the correct color dimensionality from the unknown encoding properties of the eye. \nIn this paper, we introduce a computational framework for modeling this emergence of human color vision by simulating both the eye and the cortex. Existing research often overlooks how the cortex develops color vision or represents color space internally, assuming that the color dimensionality is known a priori; however, we argue that the visual cortex has the capability and the challenge of inferring the color dimensionality purely from fluctuations in the optic nerve signals. To validate our theory, we introduce a simulation engine for biological eyes based on established vision science and generate optic nerve signals resulting from looking at natural images. Further, we propose a bio-plausible model of cortical learning based on self-supervised prediction of optic nerve signal fluctuations under natural eye motions. We show that this model naturally learns to generate color vision by disentangling retinal invariants from the sensory signals. When the retina contains $N$ types of color photoreceptors, our simulation shows that $N$-dimensional color vision naturally emerges, verified through formal colorimetry. Using this framework, we also present the first simulation work that successfully boosts the color dimensionality, as observed in gene therapy on squirrel monkeys, and demonstrates the possibility of enhancing human color vision from 3D to 4D.", "keywords": [ "color vision", "computational neuroscience", "retina simulation", "cortical learning", "self-supervised learning", "color blindness" ], "rating_list": [ 8, 10, 8, 6, 8 ], "soundness_list": [ 2, 4, 4, 3, 3 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 2, 4, 3, 3, 2 ], "confidence_list": [ 3, 5, 5, 4, 3 ] }, { "paper_id": "g3nxy8N3bQ", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Jesse Perla", "gender": "unknown", "institution": "University of British Columbia", "country": "", "position": "Associate Professor" }, { "name": "Kevin Leyton-Brown", "gender": "Not Specified", "institution": "University of British Columbia", "country": "CA", "position": "Full Professor" }, { "name": "Narun Krishnamurthi Raman", "gender": "Male", "institution": "University of British Columbia", "country": "", "position": "PhD student" }, { "name": "Taylor Lundy", "gender": "Male", "institution": "University of British Columbia", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 49, "n_ref": 81, "n_ref_all": 104, "n_fig": 14, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 1389, "n_element_tab_1": 48, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 207573, "len_all_1": 74765, "len_abs": 1431, "len_title": 73, "len_sents": 41968, "len_sents_1": 37565, "n_sents": 280, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1450, "title": "STEER-ME: Assessing the Microeconomic Reasoning of Large Language Models", "abs": "Large language models (LLMs) are increasingly being applied to economic tasks like stock picking and financial analysis. Existing LLM benchmarks tend to focus on specific applications and often fail to describe a rich variety of economic tasks. Raman et al. (2024) offer a blueprint for comprehensively benchmarking strategic decision-making. However, their work failed to address the non-strategic settings prevalent in micro-economics. We address this gap by taxonomizing micro-economic reasoning into $58$ distinct elements, each grounded in up to $10$ distinct domains, $5$ perspectives, and $3$ types. The generation of benchmark data across this combinatorial space is powered by a novel LLM-assisted data generation protocol that we dub auto-STEER which generates a set of questions by adapting handwritten templates to target new domains and perspectives. By generating fresh questions for each element, auto-STEER helps reduce the risk of data contamination, ensuring that \\model evaluations remain valuable over time. We leveraged our benchmark to evaluate $15$ LLMs over each of the instantiated elements, examined their ability to reason through and solve microeconomic problems and compared LLM performance across a suite of adaptations and metrics. Our work provides insights into the current capabilities and limitations of LLMs in non-strategic economic decision-making and a tool for fine-tuning these models to improve performance.", "keywords": [ "large language models", "microeconomics", "benchmarking", "decision-making", "economic agents", "llm agents" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "g3kK6YBSZ1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 13, "author_info_list": [ { "name": "Chen Liu", "gender": "Female", "institution": "University of Queensland", "country": "AU", "position": "PhD student" }, { "name": "Hengyuan Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Jiahao Pan", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Mengfei Li", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Peng Li", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Qifeng Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Wei Xue", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Wenhan Luo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Xiaowei Chi", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Xingqun Qi", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yatian Wang", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Researcher" }, { "name": "Yike Guo", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Shanghang Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 55, "n_ref": 159, "n_ref_all": 177, "n_fig": 10, "n_tab": 4, "L_tab": 945, "n_element_tab": 150, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1555, "n_element_tab_1": 134, "formula_len_all": 299, "formula_len_all_1": 296, "len_all": 167001, "len_all_1": 58275, "len_abs": 291, "len_title": 73, "len_sents": 39451, "len_sents_1": 27225, "n_sents": 296, "n_sents_1": 166, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1790, "title": "$\\textbf{CoCoGesture}$: Towards Coherent Co-speech 3D Gesture Generation in the Wild", "abs": "Deriving co-speech 3D gestures has seen tremendous progress in virtual avatar animation. Yet, the existing methods often produce stiff and unreasonable gestures with unseen human speech inputs due to the limited 3D speech-gesture data. In this paper, we propose $\\textbf{CoCoGesture}$, a novel framework enabling coherent and diverse gesture synthesis from unseen human speech prompts. \nOur key insight is built upon the custom-designed pretrain-fintune training paradigm. At the pretraining stage, we aim to formulate a large generalizable gesture diffusion model by learning the abundant postures manifold. Therefore, to alleviate the scarcity of 3D data, we first construct a large-scale co-speech 3D gesture dataset containing more than $40$M meshed posture instances across $4.3$K speakers, dubbed $\\textbf{GES-X}$. Then, we scale up the large unconditional diffusion model to 1B parameters and pre-train it to be our gesture experts. At the finetune stage, we present the audio ControlNet that incorporates the human voice as condition prompts to guide the gesture generation. Here, we construct the audio ControlNet through a trainable copy of our pre-trained diffusion model. Moreover, we design a novel Mixture-of-Gesture-Experts (MoGE) block to adaptively fuse the audio embedding from the human speech and the gesture features from the pre-trained gesture experts with a routing mechanism. Such an effective manner ensures audio embedding is temporal coordinated with motion features while preserving the vivid and diverse gesture generation. Extensive experiments demonstrate that our proposed CoCoGesture outperforms the state-of-the-art methods on the zero-shot speech-to-gesture generation. The dataset will be publicly available at: https://anonymous.4open.science/w/GES-X/.", "keywords": [ "Co-speech Gesture Generation", "Human Motion Modeling" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 4, 1, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "g3aGMMFHW0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Guanting Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Linyu Liu", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Xiaocheng Li", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Assistant Professor" }, { "name": "Yu Pan", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 56, "n_ref": 104, "n_ref_all": 144, "n_fig": 14, "n_tab": 7, "L_tab": 3996, "n_element_tab": 325, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 786, "n_element_tab_1": 40, "formula_len_all": 692, "formula_len_all_1": 297, "len_all": 204603, "len_all_1": 68762, "len_abs": 1052, "len_title": 80, "len_sents": 68008, "len_sents_1": 34537, "n_sents": 453, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1057, "title": "Uncertainty Estimation and Quantification for LLMs: A Simple Supervised Approach", "abs": "In this paper, we study the problem of uncertainty estimation and calibration for LLMs. We begin by formulating the uncertainty estimation problem, a relevant yet underexplored area in existing literature. We then propose a supervised approach that leverages labeled datasets to estimate the uncertainty in LLMs' responses. Based on the formulation, we illustrate the difference between the uncertainty estimation for LLMs and that for standard ML models and explain why the hidden neurons of the LLMs may contain uncertainty information. Our designed approach demonstrates the benefits of utilizing hidden activations to enhance uncertainty estimation across various tasks and shows robust transferability in out-of-distribution settings. We distinguish the uncertainty estimation task from the uncertainty calibration task and show that better uncertainty estimation leads to better calibration performance. Furthermore, our method is easy to implement and adaptable to different levels of model accessibility including black box, grey box, and white box.", "keywords": [ "large language model", "uncertainty estimation", "calibration", "supervised learning" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "g3VCIM94ke", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Adrian W. Dobbelstein", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "MS student" }, { "name": "Arne Schneuing", "gender": "Not Specified", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Bruno Correia", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "Associate Professor" }, { "name": "Ilia Igashov", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Michael M. Bronstein", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Thomas Castiglione", "gender": "Male", "institution": "VantAI", "country": "US", "position": "AI Scientist" } ], "n_formula": 54, "n_formula_1": 7, "n_ref_uni": 70, "n_ref": 123, "n_ref_all": 180, "n_fig": 24, "n_tab": 11, "L_tab": 4784, "n_element_tab": 632, "n_fig_1": 18, "n_tab_1": 4, "L_tab_1": 1186, "n_element_tab_1": 146, "formula_len_all": 3204, "formula_len_all_1": 404, "len_all": 235132, "len_all_1": 72408, "len_abs": 736, "len_title": 106, "len_sents": 64334, "len_sents_1": 34162, "n_sents": 493, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 737, "title": "Multi-domain Distribution Learning for De Novo Drug Design", "abs": "We introduce DrugFlow, a generative model for structure-based drug design that integrates continuous flow matching with discrete Markov bridges, demonstrating state-of-the-art performance in learning chemical, geometric, and physical aspects of three-dimensional protein-ligand data. We endow DrugFlow with an uncertainty estimate that is able to detect out-of-distribution samples. To further enhance the sampling process towards distribution regions with desirable metric values, we propose a joint preference alignment scheme applicable to both flow matching and Markov bridge frameworks. Furthermore, we extend our model to also explore the conformational landscape of the protein by jointly sampling side chain angles and molecules.", "keywords": [ "Drug Discovery", "Flow Matching", "Markov Bridge", "Equivariance" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 1 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "g3PuaFh5vV", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Oiwi Parker Jones", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "Principal Researcher" }, { "name": "Ryan Charles Timms", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Researcher" }, { "name": "Yonatan Gideoni", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 63, "n_ref": 83, "n_ref_all": 113, "n_fig": 3, "n_tab": 11, "L_tab": 3405, "n_element_tab": 296, "n_fig_1": 3, "n_tab_1": 9, "L_tab_1": 2033, "n_element_tab_1": 142, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 185788, "len_all_1": 66215, "len_abs": 745, "len_title": 112, "len_sents": 38116, "len_sents_1": 30208, "n_sents": 307, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 747, "title": "Non-invasive Neural Decoding in Source Reconstructed Brain Space", "abs": "Non-invasive brainwave decoding is usually done using Magneto/Electroencephalography (MEG/EEG) sensor measurements as inputs. This makes combining datasets and building models with inductive biases difficult as most datasets use different scanners and the sensor arrays have a nonintuitive spatial structure. In contrast, fMRI scans are acquired directly in brain space, a voxel grid with a typical structured input representation. By using established techniques to reconstruct the sensors' sources' neural activity it is possible to decode from voxels for MEG data as well. We show that this enables spatial inductive biases, spatial data augmentations, better interpretability, zero-shot generalisation between datasets, and data harmonisation.", "keywords": [ "Neural decoding", "MEG", "brain decoding", "structured learning", "brain computer interface" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "g3D27bfmrf", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Baoxing Huai", "gender": "unknown", "institution": "Huawei Cloud Computing Technology Co., Ltd", "country": "", "position": "Researcher" }, { "name": "Jikai Wang", "gender": "Male", "institution": "Soochow University", "country": "CN", "position": "MS student" }, { "name": "Juntao Li", "gender": "Male", "institution": "Soochow University, China", "country": "", "position": "Associate Professor" }, { "name": "Min Zhang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Qingrong Xia", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Xinyu Duan", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Zhefeng Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Zhenxu Tian", "gender": "unknown", "institution": "Soochow University", "country": "CN", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 36, "n_ref": 62, "n_ref_all": 77, "n_fig": 4, "n_tab": 5, "L_tab": 2065, "n_element_tab": 242, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2065, "n_element_tab_1": 242, "formula_len_all": 371, "formula_len_all_1": 371, "len_all": 115790, "len_all_1": 67581, "len_abs": 1357, "len_title": 122, "len_sents": 29864, "len_sents_1": 29099, "n_sents": 262, "n_sents_1": 254, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1411, "title": "CASD: Enhancing Generation Accuracy via Context-Aware Speculative Decoding", "abs": "With recent advancements in long-context model variants, Large language models (LLMs) can conveniently process different types of task-related information by simply converting them into an input sequence, even consisting of over 100K tokens. Though with a simple and unified form, there is still considerable room in leveraging input context effectively and efficiently. In this paper, we propose a simple yet effective CASD (Context-Aware Speculative Decoding) method to boost context usage. CASD is a decoding algorithm that requires no extra training or draft models. It improves not only generation performance but also inference efficiency. Experiments on 8 datasets (including question answering, summarization and code completion tasks in LongBench) show that CASD increases the average generation score by 3.3 points. CASD achieves a mean acceptance length of 3.10 and a speed-up ratio of 1.99. Moreover, CASD integrates effectively with context compression technology, addressing the issue of excessive memory overhead caused by long contexts. Since CASD directly retrieves token-level content from the input context to boost the generation accuracy, it can effectively mitigate the possible side-effects of context compression methods when crucial context information is dropped. Our anonymous code is available at \\href{https://anonymous.4open.science/r/CASD}{https://anonymous.4open.science/r/CASD}.", "keywords": [ "Language model", "Natural language processing", "Long context generation", "Speculative decoding" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "g2Udwv77WN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Hongxin Wei", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Pengjie Gu", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" }, { "name": "Xinyu Cai", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" } ], "n_formula": 50, "n_formula_1": 18, "n_ref_uni": 31, "n_ref": 56, "n_ref_all": 66, "n_fig": 3, "n_tab": 5, "L_tab": 2256, "n_element_tab": 134, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1802, "n_element_tab_1": 93, "formula_len_all": 2976, "formula_len_all_1": 1270, "len_all": 117678, "len_all_1": 67168, "len_abs": 1158, "len_title": 127, "len_sents": 36706, "len_sents_1": 29649, "n_sents": 308, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1173, "title": "Improving Ordinal Conformal Prediction by Stepwise Adaptive Posterior Alignment", "abs": "Ordinal classification (OC) is widely used in real-world applications to categorize instances into ordered discrete classes. In risk-sensitive scenarios, ordinal conformal prediction (OCP) is used to obtain a small contiguous prediction set containing ground-truth labels with a desired coverage guarantee. However, OC models often fail to accurately model the posterior distribution, which harms the prediction set obtained by OCP. Therefore, we introduce a new method called \\textit{Adaptive Posterior Alignment Step-by-Step} (APASS), which reduces the distribution discrepancy to improve the downstream OCP performance. It is designed as a versatile, plug-and-play solution that is easily integrated into any OC model before OCP. APASS first employs an attention-based estimator to adaptively estimate the variance of the posterior distribution using the information in the calibration set, then utilizes a stepwise temperature scaling algorithm to align the posterior variance predicted by OC models to the better variance estimation. Extensive evaluations on 10 real-world datasets demonstrate that APASS consistently boosts the OCP performance of 5 popular OC models.", "keywords": [ "conformal prediction", "ordinal classification" ], "rating_list": [ 3, 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 3, 4 ] }, { "paper_id": "g2C947jjjQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Christos Faloutsos", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Costas Mavromatis", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Huzefa Rangwala", "gender": "Male", "institution": "Amazon", "country": "", "position": "Principal Researcher" }, { "name": "Meng-Chieh Lee", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Qi Zhu", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Soji Adeshina", "gender": "unknown", "institution": "University of California Berkeley", "country": "", "position": "Undergrad student" }, { "name": "Vassilis N. Ioannidis", "gender": "unknown", "institution": "Amazon Web Services", "country": "US", "position": "Applied Scientist II" }, { "name": "Zhen Han", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 30, "n_ref": 59, "n_ref_all": 90, "n_fig": 5, "n_tab": 6, "L_tab": 2544, "n_element_tab": 226, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1226, "n_element_tab_1": 108, "formula_len_all": 207, "formula_len_all_1": 137, "len_all": 131651, "len_all_1": 62698, "len_abs": 1435, "len_title": 118, "len_sents": 37383, "len_sents_1": 27170, "n_sents": 336, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1442, "title": "Agent-G: An Agentic Framework for Graph Retrieval Augmented Generation", "abs": "Given two knowledge sources, one containing unstructured documents and the other comprising structured graph knowledge bases, how can we effectively retrieve the relevant information to answer user questions?\nWhile Retrieval-Augmented Generation (RAG) retrieves documents to assist the large language model (LLM) in question answering,\nGraph RAG (GRAG) uses graph knowledge bases as an additional knowledge source.\nHowever, there are many questions that require information from both sources, which complicates the scenario and makes hybrid retrieval essential.\nThe goal is to effectively leverage both sources to provide better answers to the questions.\nTherefore, we propose Agent-G, a unified framework for GRAG, composed of an agent, a retriever bank, and a critic module.\nAgent-G has the following advantages:\n1) Agentic, it automatically improves the agent's action with self-reflection, \n2) Adaptive, it solves questions that require hybrid knowledge source with a single unified framework,\n3) Interpretable, it justifies decision making and reduces hallucinations, and\n4) Effective, it adapts to different GRAG settings and outperforms all baselines.\n\nThe experiments are conducted on two real-world GRAG benchmarks, namely STaRK and CRAG.\nIn STaRK, Agent-G shows relative improvements in Hit@1 of 47% in STaRK-MAG and 55% in TaRK-Prime.\nIn CRAG, Agent-G increases accuracy by 35% while reducing hallucination by 11%, both relatively.", "keywords": [ "Graph Retrieval-Augmented Generation", "Large Language Models", "LLM Agents" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "g1kSMVqaXg", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Jie Xu", "gender": "Female", "institution": "City University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Zihan Wu", "gender": "Non-Binary", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 127, "n_formula_1": 33, "n_ref_uni": 20, "n_ref": 35, "n_ref_all": 72, "n_fig": 2, "n_tab": 9, "L_tab": 5452, "n_element_tab": 305, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 1837, "n_element_tab_1": 155, "formula_len_all": 10098, "formula_len_all_1": 2457, "len_all": 166201, "len_all_1": 61740, "len_abs": 1214, "len_title": 155, "len_sents": 47760, "len_sents_1": 22561, "n_sents": 525, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1221, "title": "Dynamic Influence Tracker: Estimating Sample Influence in SGD-Trained Models across Arbitrary Time Windows", "abs": "Understanding how training samples affect models improves model interpretability, optimization strategies, and anomaly detection. However, existing methods for estimating sample influence provide only static assessments, rely on restrictive assumptions, and require high computational costs. \n\tWe propose Dynamic Influence Tracker (DIT), a novel method to estimate time-varying sample influence in models trained with Stochastic Gradient Descent (SGD). DIT enables fine-grained analysis of sample influence within arbitrary time windows during training through a two-phase algorithm. The training phase efficiently captures and stores necessary information about the SGD trajectory, while the inference phase computes the influence of samples on the model within a specified time window. We provide a theoretical error bound for our estimator without assuming convexity, showing its reliability across various learning scenarios. Our experimental results reveal the evolution of sample influence throughout the training process, enhancing understanding of learning dynamics. We show DIT's effectiveness in improving model performance through anomalous sample detection and its potential for advancing curriculum learning.", "keywords": [ "Explainability; data influence" ], "rating_list": [ 3, 8, 3, 6, 5 ], "soundness_list": [ 1, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 4, 2 ], "confidence_list": [ 4, 3, 3, 3, 3 ] }, { "paper_id": "g1fkhbhHjL", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Jing Zhang", "gender": "Female", "institution": "Australian National University", "country": "AU", "position": "Lecturer" }, { "name": "Mengqi He", "gender": "Male", "institution": "Australian National University", "country": "", "position": "PhD student" }, { "name": "Shu Zou", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "PhD student" }, { "name": "Xinyu Tian", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "PhD student" }, { "name": "Zhaoyuan Yang", "gender": "unknown", "institution": "General Electric", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 65, "n_ref": 188, "n_ref_all": 240, "n_fig": 13, "n_tab": 21, "L_tab": 2914, "n_element_tab": 389, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 461, "n_element_tab_1": 68, "formula_len_all": 336, "formula_len_all_1": 401, "len_all": 252155, "len_all_1": 71464, "len_abs": 1171, "len_title": 150, "len_sents": 73678, "len_sents_1": 34291, "n_sents": 576, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 1174, "title": "Black Sheep in the Herd: Playing with Spuriously Correlated Attributes for Vision-Language Recognition", "abs": "Few-shot adaptation for Vision-Language Models (VLMs) presents a dilemma: balancing in-distribution accuracy with out-of-distribution generalization. Recent research has utilized low-level concepts such as visual attributes to enhance generalization. However, this study reveals that VLMs overly rely on a small subset of attributes on decision-making, which co-occur with the category but are not inherently part of it, termed spuriously correlated attributes. This biased nature of VLMs results in poor generalization. To address this, 1) we first propose Spurious Attribute Probing (SAP), identifying and filtering out these problematic attributes to significantly enhance the generalization of existing attribute-based methods; 2) We introduce Spurious Attribute Shielding (SAS), a plug-and-play module that mitigates the influence of these attributes on prediction, seamlessly integrating into various Parameter-Efficient Fine-Tuning (PEFT) methods. In experiments, SAP and SAS significantly enhance accuracy on distribution shifts across 11 datasets and 3 generalization tasks without compromising downstream performance, establishing a new state-of-the-art benchmark.", "keywords": [ "Few-shot Adaptation", "Prompt Learning", "Vision-Language Models" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 2, 5, 3, 5 ] }, { "paper_id": "g0rnZeBguq", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Huanran Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jie Shi", "gender": "unknown", "institution": "Huawei International.", "country": "SG", "position": "Principal Researcher" }, { "name": "Qiongxiu Li", "gender": "Female", "institution": "Aalborg University ", "country": "DK", "position": "Assistant Professor" }, { "name": "Wenxuan Sun", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Xiao Li", "gender": "Male", "institution": "Computer Science, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiaolin Hu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "He YingZhe", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" } ], "n_formula": 57, "n_formula_1": 19, "n_ref_uni": 37, "n_ref": 119, "n_ref_all": 155, "n_fig": 3, "n_tab": 15, "L_tab": 5479, "n_element_tab": 432, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1900, "n_element_tab_1": 160, "formula_len_all": 4748, "formula_len_all_1": 1279, "len_all": 193057, "len_all_1": 72144, "len_abs": 1046, "len_title": 126, "len_sents": 54256, "len_sents_1": 31532, "n_sents": 476, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1023, "title": "Adversarial Diffusion Bridge Model for Reliable Adversarial Purification", "abs": "Recently Diffusion-based Purification (DiffPure) has been recognized as an effective defense method against adversarial examples. However, we find DiffPure which directly employs the original pre-trained diffusion models for adversarial purification, to be suboptimal. This is due to an inherent trade-off between noise purification performance and data recovery quality. Additionally, the reliability of existing evaluations for DiffPure is questionable, as they rely on weak adaptive attacks. In this work, we propose a novel Adversarial Diffusion Bridge Model, termed ADBM. ADBM directly constructs a reverse bridge from the diffused adversarial data back to its original clean examples, enhancing the purification capabilities of the original diffusion models. Through theoretical analysis and experimental validation across various scenarios, ADBM has proven to be a superior and robust defense mechanism, offering significant promise for practical applications. Code is available at https://github.com/LixiaoTHU/ADBM.", "keywords": [ "diffusion models", "adversarial robustness" ], "rating_list": [ 8, 6, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "g0Doz4IRHU", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Dun Zeng", "gender": "unknown", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Guanzhong Chen", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "MS student" }, { "name": "Irwin King", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Jianyong Wang", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "Associate Professor" }, { "name": "Jinghua Wang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Junfan Li", "gender": "Male", "institution": "Harbin Institute of Technology Shenzhen", "country": "CN", "position": "Postdoc" }, { "name": "Yuan Qi", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yukun Zhang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Zenglin Xu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 0, "n_ref_uni": 40, "n_ref": 55, "n_ref_all": 78, "n_fig": 7, "n_tab": 13, "L_tab": 11626, "n_element_tab": 983, "n_fig_1": 4, "n_tab_1": 9, "L_tab_1": 3619, "n_element_tab_1": 285, "formula_len_all": 230, "formula_len_all_1": 0, "len_all": 215022, "len_all_1": 66227, "len_abs": 1650, "len_title": 142, "len_sents": 54210, "len_sents_1": 31719, "n_sents": 392, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1481, "title": "FedCVD: The First Real-World Federated Learning Benchmark on Cardiovascular Disease Data", "abs": "Cardiovascular diseases (CVDs) are currently the leading cause of death worldwide, highlighting the critical need for early diagnosis and treatment. Machine learning (ML) methods can help diagnose CVDs early, but their performance relies on access to substantial data with high quality. However, the sensitive nature of healthcare data often restricts individual clinical institutions from sharing data to train sufficiently generalized and unbiased ML models. Federated Learning (FL) is an emerging approach, which offers a promising solution by enabling collaborative model training across multiple participants without compromising the privacy of the individual data owners. However, to the best of our knowledge, there has been limited prior research applying FL to the cardiovascular disease domain. Moreover, existing FL benchmarks and datasets are typically simulated and may fall short of replicating the complexity of natural heterogeneity found in realistic datasets that challenges current FL algorithms. To address these gaps, this paper presents the first real-world FL benchmark for cardiovascular disease detection, named FedCVD. This benchmark comprises two major tasks: electrocardiogram (ECG) classification and echocardiogram (ECHO) segmentation, based on naturally scattered datasets constructed from the CVD data of seven institutions. Our extensive experiments on these datasets reveal that FL faces new challenges with real-world non-IID and long-tail data.", "keywords": [ "Federated Learning", "Healthcare Data", "Cardiovascular Diseases", "Real-world Datasets", "Data Heterogeneity" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "fzbmTawTUB", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Di Lin", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Hongkai Yu", "gender": "Male", "institution": "Cleveland State University", "country": "US", "position": "Assistant Professor" }, { "name": "Ming-Ming Cheng", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Full Professor" }, { "name": "Qin Zou", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Qing Guo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Xi Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yuyi Wang", "gender": "Male", "institution": "CRRC Zhuzhou Institute, China", "country": "CN", "position": "Principal Researcher" }, { "name": "mingxuan cui", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 40, "n_ref": 69, "n_ref_all": 112, "n_fig": 13, "n_tab": 6, "L_tab": 775, "n_element_tab": 217, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 765, "n_element_tab_1": 73, "formula_len_all": 653, "formula_len_all_1": 645, "len_all": 150984, "len_all_1": 61026, "len_abs": 1602, "len_title": 133, "len_sents": 47257, "len_sents_1": 28240, "n_sents": 351, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1613, "title": "Visibility-Uncertainty-guided 3D Gaussian Inpainting via Scene Conceptional Learning", "abs": "3D Gaussian Splatting (3DGS) has emerged as a powerful and efficient 3D representation for novel view synthesis. This paper extends 3DGS capabilities to inpainting, where masked objects in a scene are replaced with new contents that blend seamlessly with the surroundings. Unlike 2D image inpainting, 3D Gaussian inpainting (3DGI) faces the challenge of effectively leveraging complementary visual and semantic cues from multiple input views, as occluded areas in one view may be visible in others. To address this, we propose a method that measures the visibility uncertainties of 3D points across different input views and uses them to guide 3DGI in utilizing complementary visual cues. We also employ the uncertainties to learn a semantic concept of the scene without the masked object and use a diffusion model to fill masked objects in the input images based on the learned concept. Finally, we build a novel 3DGI framework VISTA by integrating VISibility-uncerTainty-guided 3DGI with scene conceptuAl learning. VISTA generates high-quality 3DGS models capable of synthesizing artifact-free and naturally inpainted novel views. Furthermore, our approach extends to handling dynamic distractors arising from temporal object changes, enhancing its versatility in diverse scene reconstruction scenarios. We demonstrate the superior performance of our method over state-of-the-art techniques using two challenging datasets: the SPIn-NeRF dataset, featuring 10 diverse static 3D inpainting scenes, and an underwater 3D inpainting dataset derived from UTB180, which includes fast-moving fish as inpainting targets.", "keywords": [ "3D reconstruction; 3D Inpainting; Conceptional Diffusion" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "fzZfju8y0g", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Edouard Oyallon", "gender": "Male", "institution": "CNRS", "country": "FR", "position": "Researcher" }, { "name": "Jiequn Han", "gender": "Male", "institution": "Simons Foundation", "country": "US", "position": "Researcher" }, { "name": "Rudy Morel", "gender": "Male", "institution": "Flatiron Institute", "country": "US", "position": "Postdoc" } ], "n_formula": 17, "n_formula_1": 7, "n_ref_uni": 49, "n_ref": 88, "n_ref_all": 106, "n_fig": 12, "n_tab": 6, "L_tab": 1680, "n_element_tab": 136, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 1246, "n_element_tab_1": 155, "formula_len_all": 1207, "formula_len_all_1": 429, "len_all": 139429, "len_all_1": 58893, "len_abs": 1271, "len_title": 125, "len_sents": 40816, "len_sents_1": 26895, "n_sents": 288, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1279, "title": "In-Context Neural PDE: Learning to Adapt a Neural Solver to Different Physics", "abs": "We address the problem of predicting the next state of a dynamical system governed by *unknown* temporal partial differential equations (PDEs) using limited time-lapse data. While transformers offer a natural solution to this task through in-context learning, the inductive bias of temporal PDEs suggests a more tailored and effective approach. Specifically, when the underlying temporal PDE is fully known, classical numerical solvers can evolve the state with only a few parameters. Building on this observation, we introduce a large transformer-based hypernetwork that processes successive states to generate parameters for a much smaller neural ODE-like solver, which then predicts the next state through time integration. This framework, termed as *in-context neural PDE*, decouples parameter estimation from state prediction, offering closer alignment with classical numerical methods for improved interpretability while preserving the in-context learning capabilities of transformers. \nNumerical experiments on diverse physical datasets demonstrate that our method outperforms standard transformer-based models, reducing sample complexity and improving generalization, making it an efficient and scalable approach for spatiotemporal prediction in complex physical systems.", "keywords": [ "Spatio-temporal prediction", "PDEs", "in-context learning", "neural solvers" ], "rating_list": [ 3, 5, 3, 3, 3 ], "soundness_list": [ 4, 3, 2, 2, 2 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 2, 2 ], "confidence_list": [ 4, 1, 4, 4, 5 ] }, { "paper_id": "fzJtylzsKO", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Austin Tripp", "gender": "Male", "institution": "Valence Labs powered by recursion", "country": "GB", "position": "Researcher" }, { "name": "Jenna Fromer", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Mrunali Manjrekar", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Runzhong Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Connor Coley", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Jose Miguel Hernandez", "gender": "unknown", "institution": "Microsoft Research Cambridge", "country": "GB", "position": "Research Scientist" } ], "n_formula": 17, "n_formula_1": 9, "n_ref_uni": 56, "n_ref": 107, "n_ref_all": 129, "n_fig": 8, "n_tab": 2, "L_tab": 2862, "n_element_tab": 302, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1465, "n_element_tab_1": 156, "formula_len_all": 1037, "formula_len_all_1": 651, "len_all": 173246, "len_all_1": 66092, "len_abs": 1137, "len_title": 152, "len_sents": 43212, "len_sents_1": 30241, "n_sents": 322, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1141, "title": "Batched Bayesian optimization with correlated candidate uncertainties", "abs": "Batched Bayesian optimization (BO) can accelerate molecular design by efficiently identifying top-performing compounds from a large chemical library. Existing acquisition strategies for batch design in BO aim to balance exploration and exploitation. This often involves optimizing non-additive batch acquisition functions, necessitating approximation via myopic construction and/or diversity heuristics. In this work, we propose an acquisition strategy for discrete optimization that is motivated by pure exploitation, qPO (multipoint Probability of Optimality). qPO maximizes the probability that the batch includes the true optimum, which is expressible as the sum over individual acquisition scores and thereby circumvents the combinatorial challenge of optimizing a batch acquisition function. We differentiate the proposed strategy from parallel Thompson sampling and discuss how it implicitly captures diversity. Finally, we apply our method to the model-guided exploration of large chemical libraries and provide empirical evidence that it performs better than or on par with state-of-the-art methods in batched Bayesian optimization.", "keywords": [ "Bayesian optimization", "molecular design", "molecular discovery", "uncertainty", "model-guided optimization" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "fyl82vAale", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Juri Hong", "gender": "unknown", "institution": "Yonsei University", "country": "KR", "position": "MS student" }, { "name": "Kibok Lee", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Seunghan Lee", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "PhD student" }, { "name": "Taeyoung Park", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 40, "n_ref": 116, "n_ref_all": 156, "n_fig": 12, "n_tab": 22, "L_tab": 10202, "n_element_tab": 1295, "n_fig_1": 11, "n_tab_1": 11, "L_tab_1": 4479, "n_element_tab_1": 687, "formula_len_all": 751, "formula_len_all_1": 794, "len_all": 161762, "len_all_1": 77917, "len_abs": 1044, "len_title": 105, "len_sents": 36527, "len_sents_1": 29469, "n_sents": 355, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1046, "title": "Sequential Order-Robust Mamba for Time Series Forecasting", "abs": "Mamba has recently emerged as a promising alternative to Transformers, offering near-linear complexity in processing sequential data.\nHowever, while channels in time series (TS) data have no specific order in general, recent studies have adopted Mamba to capture channel dependencies (CD) in TS, introducing a sequential order bias. To address this issue, we propose SOR-Mamba, a TS forecasting method that 1) incorporates a regularization strategy to minimize the discrepancy between two embedding vectors generated from data with reversed channel orders, thereby enhancing robustness to channel order, and 2) eliminates the 1D-convolution originally designed to capture local information in sequential data. Furthermore, we introduce channel correlation modeling (CCM), a pretraining task aimed at preserving correlations between channels from the data space to the latent space in order to enhance the ability to capture CD.\nExtensive experiments demonstrate the efficacy of the proposed method across standard and transfer learning scenarios.", "keywords": [ "Mamba", "State-space Model", "Time Series" ], "rating_list": [ 6, 5, 6, 6, 5 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 5, 3, 2 ] }, { "paper_id": "fy4rCv3s5i", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Peter Kocsis", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Vincent Sitzmann", "gender": "Male", "institution": "Yellow Technologies", "country": "US", "position": "Chief Scientist" }, { "name": "Matthias Niessner", "gender": "unknown", "institution": "Technical University of Munich", "country": "DE", "position": "Professor" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 24, "n_ref": 90, "n_ref_all": 123, "n_fig": 17, "n_tab": 2, "L_tab": 110, "n_element_tab": 15, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 517, "formula_len_all_1": 428, "len_all": 152936, "len_all_1": 65233, "len_abs": 583, "len_title": 40, "len_sents": 41572, "len_sents_1": 29535, "n_sents": 419, "n_sents_1": 283, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1258, "title": "Neural Lighting Priors for Indoor Scenes", "abs": "We introduce Neural Lighting Priors, a learned surface emission model for indoor scenes. Given multi-view observations as well as the geometry of a scene, we decouple spatially varying lighting and material parameters. Existing inverse rendering methods typically use hand-crafted emission models or require a large number of views to better constrain the highly ambiguous appearance decomposition task. We aim to overcome these limitations by introducing an expressive learned parametric emission model and utilizing semantic information to sufficiently constrain the optimization, thus allowing us to infer light sources, even if they are not visible in the observations. We model the emitted radiance with a neural field parameterized by the emitting direction and a local latent code stored in a voxel grid. At test time, we fit the local latent codes to the scene using differentiable path tracing, optimizing the reconstruction loss. Our reconstruction allows us to insert virtual objects in a scene and gives us control over the emitters to change their emission color and intensity. Thanks to the learned 3D prior, our method requires fewer views than state-of-the-art relighting methods, gives more control, and also improves the relighting quality.", "keywords": [ "lighting representation", "prior learning", "neural field", "3D", "computer graphics" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "fxv0FfmDAg", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Artem M Vysogorets", "gender": "Male", "institution": "Rockefeller University", "country": "US", "position": "Researcher" }, { "name": "Julia Kempe", "gender": "Female", "institution": "New York University", "country": "US", "position": "Professor" }, { "name": "Kartik Ahuja", "gender": "Male", "institution": "FAIR (Meta)", "country": "FR", "position": "Researcher" } ], "n_formula": 22, "n_formula_1": 8, "n_ref_uni": 70, "n_ref": 123, "n_ref_all": 183, "n_fig": 15, "n_tab": 2, "L_tab": 1135, "n_element_tab": 120, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1923, "formula_len_all_1": 736, "len_all": 198281, "len_all_1": 64692, "len_abs": 1110, "len_title": 90, "len_sents": 55256, "len_sents_1": 32360, "n_sents": 402, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 42, "L_abs": 1114, "title": "DRoP: Distributionally Robust Data Pruning", "abs": "In the era of exceptionally data-hungry models, careful selection of the training data is essential to mitigate the extensive costs of deep learning. Data pruning offers a solution by removing redundant or uninformative samples from the dataset, which yields faster convergence and improved neural scaling laws. However, little is known about its impact on classification bias of the trained models. We conduct the first systematic study of this effect and reveal that existing data pruning algorithms can produce highly biased classifiers. We present theoretical analysis of the classification risk in a mixture of Gaussians to argue that choosing appropriate class pruning ratios, coupled with random pruning within classes has potential to improve worst-class performance. We thus propose DRoP, a distributionally robust approach to pruning and empirically demonstrate its performance on standard computer vision benchmarks. In sharp contrast to existing algorithms, our proposed method continues improving distributional robustness at a tolerable drop of average performance as we prune more from the datasets.", "keywords": [ "Data Pruning", "Classification Bias", "Robustness" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 4, 1 ], "presentation_list": [ 3, 4, 1 ], "contribution_list": [ 3, 4, 3 ], "confidence_list": [ 2, 4, 3 ] }, { "paper_id": "fxarGPFMmB", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Caroline Mazini Rodrigues", "gender": "Female", "institution": "Universit\u00e9 Gustave Eiffel", "country": "FR", "position": "PhD student" }, { "name": "Laurent Najman", "gender": "Male", "institution": "Universit\u00e9 Gustave Eiffel - ESIEE Paris", "country": "FR", "position": "Full Professor" }, { "name": "Nicolas Boutry", "gender": "Male", "institution": "Ecole Pour l'Informatique et les Techniques Avancees", "country": "FR", "position": "Associate Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 29, "n_ref": 53, "n_ref_all": 107, "n_fig": 11, "n_tab": 13, "L_tab": 16692, "n_element_tab": 1206, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 5277, "n_element_tab_1": 277, "formula_len_all": 70, "formula_len_all_1": 0, "len_all": 216841, "len_all_1": 75741, "len_abs": 4349, "len_title": 105, "len_sents": 69967, "len_sents_1": 34699, "n_sents": 480, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1046, "title": "Reasoning with trees: interpreting CNNs using hierarchies", "abs": "Challenges remain in providing interpretable explanations for neural network reasoning in explainable AI (xAI). Existing methods like Integrated Gradients produce noisy maps, and LIME, while intuitive, may deviate from the model\u2019s reasoning. We introduce a framework that uses hierarchical segmentation techniques for faithful and interpretable explanations of Convolutional Neural Networks (CNNs). Our method constructs model-based hierarchical segmentations that maintain the model\u2019s reasoning fidelity and allow both human-centric and model-centric segmentation. This approach can be combined with various xAI methods and provides multiscale explanations that help identify biases and improve understanding of neural network decision-making. Experiments show that our framework, xAiTrees, delivers highly interpretable and faithful model explanations, not only surpassing traditional xAI methods but shedding new light on a novel approach to enhancing xAI interpretability. Code at: https://anonymous.4open.science/r/reasoning_with_trees-F3E1.", "keywords": [ "interpretability", "image classification", "hierarchical segmentation" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 1, 3, 1 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 2, 5, 4, 3 ] }, { "paper_id": "fwHVclv0ij", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Liyan Xie", "gender": "Female", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "Assistant Professor" }, { "name": "Mengdi Wang", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Associate Professor" }, { "name": "Minshuo Chen", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Postdoc" }, { "name": "Guo Jiawei", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 44, "n_ref": 62, "n_ref_all": 92, "n_fig": 21, "n_tab": 3, "L_tab": 1271, "n_element_tab": 64, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 362, "n_element_tab_1": 40, "formula_len_all": 1107, "formula_len_all_1": 828, "len_all": 162563, "len_all_1": 71673, "len_abs": 3881, "len_title": 132, "len_sents": 55068, "len_sents_1": 33414, "n_sents": 456, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1484, "title": "Online Detection for Black-Box Large Language Models with Adaptive Prompt Selection", "abs": "The widespread success of large language models (LLMs) has made them integral to various applications, yet security and reliability concerns are growing. It now becomes critical to safeguard LLMs from unintended changes caused by tampering, malicious prompt injection, or unauthorized parameter updates, etc. Early detection of these changes is essential to maintain the performance, fairness, and trustworthiness of LLM-powered applications. However, in black-box settings, where access to model parameters and output probabilities is unavailable, few detection methods exist. In this paper, we propose a novel online change-point detection method for quickly detecting changes in black-box LLMs. Our method features several key innovations: 1) we derive a CUSUM-type detection statistic based on the entropy and the Gini coefficient of the response distribution, and 2) we utilize a UCB-based adaptive prompt selection strategy for identifying change-sensitive prompts to enhance detection. We evaluate the effectiveness of the proposed method using synthetic data, where changes are simulated through watermarking and model version updates. Our proposed method is able to detect changes quickly while well controlling the false alarm rate. Moreover, for real-world data, our method also accurately detects announced changes in LLM APIs via daily online interactions with APIs. We also demonstrate strong evidence of unreported changes in APIs, which may be of independent interest.", "keywords": [ "online change detection", "LLM security", "active prompt selection" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 3, 3, 5, 3 ] }, { "paper_id": "fvyuKLTC3s", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Artsiom Ranchynski", "gender": "unknown", "institution": "University of Warsaw", "country": "PL", "position": "PhD student" }, { "name": "\u0141ukasz Kuci\u0144ski", "gender": "Male", "institution": "IDEAS NCBR", "country": "PL", "position": "Senior Research Scientist" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 9, "n_ref": 12, "n_ref_all": 13, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1039, "formula_len_all_1": 1039, "len_all": 43618, "len_all_1": 43618, "len_abs": 878, "len_title": 117, "len_sents": 17069, "len_sents_1": 17069, "n_sents": 146, "n_sents_1": 146, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 906, "title": "Cayley Maze: Universal Open-Ended Reinforcement Learning Environment", "abs": "Parametrizable environments with variable complexity are crucial for advancing fields such as Unsupervised Environment Design (UED), Open-Ended Learning, Curriculum Learning, and Meta Reinforcement Learning. However, the selection of environments in evaluation procedures, along with their complexities, is often either neglected or lacks formal justification. We propose the formal definition of complexity for Markov Decision Processes using Deterministic Finite Automata and Group Theory machinery. We introduce Cayley Maze, a novel open-ended reinforcement learning environment that naturally generalizes problems like solving Rubik's Cube, sorting, and integer factorization. Cayley Maze is universal: every finite deterministic sparse MDP is an MDP of a certain instance of Cayley Maze. We demonstrate how Cayley Maze enables control over complexity, simplification, and combination of its instances.", "keywords": [ "Reinforcement Learning", "Unsupervised Environment Design", "Open-Ended Learning" ], "rating_list": [ 1, 3, 5, 6 ], "soundness_list": [ 1, 1, 2, 3 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 1, 2, 4, 2 ], "confidence_list": [ 2, 3, 5, 3 ] }, { "paper_id": "fvo6q86NKG", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Masaki Inoue", "gender": "Male", "institution": "Keio University", "country": "JP", "position": "Associate Professor" }, { "name": "Yuya Miyaoka", "gender": "Male", "institution": "Keio University", "country": "", "position": "MS student" } ], "n_formula": 47, "n_formula_1": 18, "n_ref_uni": 38, "n_ref": 59, "n_ref_all": 89, "n_fig": 9, "n_tab": 4, "L_tab": 761, "n_element_tab": 70, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 133, "n_element_tab_1": 5, "formula_len_all": 2136, "formula_len_all_1": 641, "len_all": 142279, "len_all_1": 72025, "len_abs": 886, "len_title": 39, "len_sents": 47183, "len_sents_1": 33843, "n_sents": 427, "n_sents_1": 295, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 890, "title": "CBF-LLM: Safe Control for LLM Alignment", "abs": "This paper proposes a control-based framework for aligning large language models (LLMs) by leveraging a control barrier function (CBF) to ensure user-desirable text generation. \nThe presented framework applies the CBF safety filter to the predicted token generated from the baseline LLM, to intervene in the generated text.\nThe safety filter includes two significant advantages:\nthis safety filter is an add-on type, allowing it to be used for alignment purposes without fine-tuning the baseline LLM,\nand if there is an evaluation model regarding the desired alignment, it can be directly applied to the filter design.\nThe overall text-generation system is implemented with Llama 3 and a BERT model, aiming to generate positive text.\nFinally, further applications and limitations of the CBF-LLM for other alignment tasks, including topic-keeping and hallucination mitigating, are discussed.", "keywords": [ "LLM", "alignment", "control barrier function" ], "rating_list": [ 3, 8, 5, 3, 1 ], "soundness_list": [ 1, 3, 2, 1, 2 ], "presentation_list": [ 3, 4, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 1, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "fvkElsJOsN", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Chi Han", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" }, { "name": "Hanlin Zhang", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Heng Ji", "gender": "Female", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Kuan-Hao Huang", "gender": "Male", "institution": "Texas A&M University", "country": "US", "position": "Assistant Professor" }, { "name": "Sham M. Kakade", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Shuiwang Ji", "gender": "Male", "institution": "Texas A&M University", "country": "US", "position": "Professor" }, { "name": "Xiner Li", "gender": "Female", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Ziqi Wang", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "Ph.D. student" }, { "name": "Hao Peng", "gender": "Male", "institution": "Department of Computer Science, University of Illinois Urbana-Champaign", "country": "", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 98, "n_ref_all": 142, "n_fig": 7, "n_tab": 10, "L_tab": 3457, "n_element_tab": 708, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1152, "n_element_tab_1": 178, "formula_len_all": 410, "formula_len_all_1": 515, "len_all": 198343, "len_all_1": 70384, "len_abs": 1383, "len_title": 116, "len_sents": 55858, "len_sents_1": 31151, "n_sents": 500, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1425, "title": "Eliminating Position Bias of Language Models: A Mechanistic Approach", "abs": "Position bias has proven to be a prevalent issue of modern language models (LMs), where the models prioritize content based on its position within the given context. This bias often leads to unexpected model failures and hurts performance, robustness, and reliability across various applications. A simple mechanistic analysis attributes the position bias to two components employed in nearly all state-of-the-art LMs: causal attention and relative positional encodings. Based on the analyses, we propose to **eliminate** position bias (e.g., different retrieved documents' orders in QA affect performance) with a **training-free zero-shot** approach. Our method changes the causal attention to bidirectional attention between documents and utilizes model attention values to decide the relative orders of documents instead of using the order provided in input prompts, therefore enabling Position-INvariant inferencE (PINE) at the document level. By eliminating position bias, models achieve better performance and reliability in downstream tasks, including LM-as-a-judge, retrieval-augmented QA, molecule generation, and math reasoning. Notably, PINE is especially useful when adapting LMs for evaluating reasoning pairs: it consistently provides $8$ to $10$ percentage points performance gains, making Llama-3-70B-Instruct perform even better than GPT-4-0125-preview and GPT-4o-2024-08-06 on the RewardBench reasoning set.", "keywords": [ "Position Bias", "Languague Models" ], "rating_list": [ 6, 8, 6, 8, 5 ], "soundness_list": [ 3, 4, 3, 2, 2 ], "presentation_list": [ 3, 4, 3, 3, 1 ], "contribution_list": [ 3, 4, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 4, 3 ] }, { "paper_id": "fvUVe2gJh0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Alexandra Chronopoulou", "gender": "Female", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Jonathan Lai", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Manaal Faruqui", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Mohit Bansal", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" }, { "name": "Prateek Yadav", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina, Chapel Hill", "country": "", "position": "Graduate Student" }, { "name": "Tsendsuren Munkhdalai", "gender": "unknown", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Tu Vu", "gender": "Male", "institution": "Virginia Tech", "country": "US", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 72, "n_ref": 153, "n_ref_all": 175, "n_fig": 7, "n_tab": 6, "L_tab": 1118, "n_element_tab": 156, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1056, "n_element_tab_1": 13, "formula_len_all": 125, "formula_len_all_1": 190, "len_all": 205511, "len_all_1": 63750, "len_abs": 1959, "len_title": 88, "len_sents": 46717, "len_sents_1": 32933, "n_sents": 287, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1971, "title": "What Matters for Model Merging at Scale?", "abs": "Model merging aims to combine multiple expert models into a more capable single model, offering benefits such as reduced storage and serving costs, improved generalization, and support for decentralized model development. Despite its promise, previous studies have primarily focused on merging a few small models. This leaves many unanswered questions about the effect of scaling model size and how it interplays with other key factors\u2014like the base model quality and number of expert models\u2014, to affect the merged model\u2019s performance. This work systematically evaluates the utility of model merging at scale, examining the impact of these different factors. We experiment with merging fully fine-tuned models using four popular merging methods\u2014Averaging, Task Arithmetic, Dare-TIES, and TIES-Merging\u2014across model sizes ranging from 1B to 64B parameters and merging up to 8 different expert models. We evaluate the merged models on both held-in tasks, i.e., the expert\u2019s training tasks, and zero-shot generalization to unseen held-out tasks. Our wide range of experiments provide several new insights about model merging at scale and the interplay between different factors. First, we find that merging is more effective when experts are created from strong base models, i.e., models with good zero-shot performance, compared to pre-trained ones. Second, larger models facilitate easier merging. Third merging consistently improves generalization capabilities. Notably, when merging eight large expert models, the merged models often generalize better compared to the multitask trained models. Fourth, we can better merge more expert models when working with larger models. Fifth, different merging methods behave very similarly at larger scales. Overall, our findings shed light on some interesting properties of model merging while also highlighting some limitations. We hope that this study will serve as a reference point on large-scale merging for upcoming research.", "keywords": [ "model merging", "weight averaging", "averaging", "composition", "modular model", "generalization" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "fvNn2rgj4Y", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Hirokatsu Kataoka", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Principal Researcher" }, { "name": "Kenji Doi", "gender": "Male", "institution": "LY Corporation", "country": "JP", "position": "Researcher" }, { "name": "Ryota Yoshihashi", "gender": "unknown", "institution": "LY Corporation", "country": "JP", "position": "Software engineer" }, { "name": "Shuntaro Okada", "gender": "Male", "institution": "LY corporation", "country": "JP", "position": "Researcher" }, { "name": "Tomohiro Tanaka", "gender": "Male", "institution": "LY Corporation", "country": "JP", "position": "Researcher" } ], "n_formula": 73, "n_formula_1": 24, "n_ref_uni": 27, "n_ref": 85, "n_ref_all": 135, "n_fig": 9, "n_tab": 13, "L_tab": 6533, "n_element_tab": 1256, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2346, "n_element_tab_1": 634, "formula_len_all": 3117, "formula_len_all_1": 874, "len_all": 161887, "len_all_1": 65287, "len_abs": 1245, "len_title": 164, "len_sents": 35851, "len_sents_1": 23771, "n_sents": 342, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 115, "L_abs": 877, "title": "Constant Rate Schedule: Constant-Rate Distributional Change for Efficient Training and Sampling in Diffusion Models", "abs": "We propose a noise schedule that ensures a constant rate of change in the probability distribution of diffused data throughout the diffusion process.\nTo obtain this noise schedule, we measure the rate of change in the probability distribution of the forward process and use it to determine the noise schedule before training diffusion models.\nThe functional form of the noise schedule is automatically determined and tailored to each dataset and type of diffusion model.\nWe evaluate the effectiveness of our noise schedule on unconditional and class-conditional image generation tasks using the LSUN (bedroom/church/cat/horse), ImageNet, and FFHQ datasets.\nThrough extensive experiments, we confirmed that our noise schedule broadly improves the performance of the diffusion models regardless of the dataset, sampler, number of function evaluations, or type of diffusion model.", "keywords": [ "diffusion model", "noise schedule" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "fv9XU7CyN2", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Gen Zhou", "gender": "Not Specified", "institution": "University of Western Ontario", "country": "", "position": "PhD student" }, { "name": "Pingzhao Hu", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Associate Professor" }, { "name": "Sugitha Janarthanan", "gender": "Female", "institution": "Western University", "country": "CA", "position": "PhD student" }, { "name": "Yutong Lu", "gender": "Female", "institution": "University of Toronto", "country": "CA", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 31, "n_ref": 58, "n_ref_all": 96, "n_fig": 6, "n_tab": 8, "L_tab": 2189, "n_element_tab": 268, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 899, "n_element_tab_1": 50, "formula_len_all": 762, "formula_len_all_1": 681, "len_all": 157433, "len_all_1": 61075, "len_abs": 1743, "len_title": 173, "len_sents": 49150, "len_sents_1": 29603, "n_sents": 348, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 124, "L_abs": 1749, "title": "CL-MFAP: A Contrastive Learning-Based Multimodal Foundation Model for Molecular Property Prediction and Antibiotic Screening", "abs": "Due to the rise in antimicrobial resistance, identifying novel compounds with antibiotic potential is crucial for combatting this global health issue. However, traditional drug development methods are costly and inefficient. Recognizing the pressing need for more effective solutions, researchers have turned to machine learning techniques to streamline the prediction and development of novel antibiotic compounds. While foundation models have shown promise in antibiotic discovery, current mainstream efforts still fall short of fully leveraging the potential of multimodal molecular data. Recent studies suggest that contrastive learning frameworks utilizing multimodal data exhibit excellent performance in representation learning across various domains. Building upon this, we introduce CL-MFAP, an unsupervised contrastive learning (CL)-based multimodal foundation (MF) model specifically tailored for discovering small molecules with potential antibiotic properties (AP) using three types of molecular data. This model employs 1.6 million bioactive molecules with drug-like properties from the ChEMBL dataset to jointly pretrain three encoders: (1) a transformer-based encoder with rotary position embedding for processing SMILES strings; (2) another transformer-based encoder, incorporating a novel bi-level routing attention mechanism to handle molecular graph representations; and (3) a Morgan fingerprint encoder using a multilayer perceptron, to achieve the contrastive learning purpose. The CL-MFAP outperforms baseline models in antibiotic property prediction by effectively utilizing different molecular modalities and demonstrates superior domain-specific performance when fine-tuned for antibiotic-related property prediction tasks.", "keywords": [ "Contrastive Learning", "Multimodal Foundation Model", "Antibiotic Property Prediction", "Bi-level Routing Attention", "Transformer" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "fv8TTt9srF", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Eric Ricardo Anschuetz", "gender": "Male", "institution": "California Institute of Technology", "country": "", "position": "Postdoc" } ], "n_formula": 369, "n_formula_1": 30, "n_ref_uni": 48, "n_ref": 199, "n_ref_all": 276, "n_fig": 6, "n_tab": 1, "L_tab": 884, "n_element_tab": 11, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 131, "n_element_tab_1": 10, "formula_len_all": 21040, "formula_len_all_1": 1541, "len_all": 398697, "len_all_1": 63133, "len_abs": 1174, "len_title": 106, "len_sents": 125474, "len_sents_1": 28911, "n_sents": 1197, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1180, "title": "A Unified Theory of Quantum Neural Network Loss Landscapes", "abs": "Classical neural networks with random initialization famously behave as Gaussian processes in the limit of many neurons, which allows one to completely characterize their training and generalization behavior. No such general understanding exists for quantum neural networks (QNNs), which\u2014outside of certain special cases\u2014are known to not behave as Gaussian processes when randomly initialized. We here prove that QNNs and their first two derivatives instead generally form what we call \"Wishart processes,\" where certain algebraic properties of the network determine the hyperparameters of the process. This Wishart process description allows us to, for the first time: give necessary and sufficient conditions for a QNN architecture to have a Gaussian process limit; calculate the full gradient distribution, generalizing previously known barren plateau results; and calculate the local minima distribution of algebraically constrained QNNs. Our unified framework suggests a certain simple operational definition for the \"trainability\" of a given QNN model using a newly introduced, experimentally accessible quantity we call the \"degrees of freedom\" of the network architecture.", "keywords": [ "quantum machine learning", "neural tangent kernel", "loss landscape", "spin glass", "Kac\u2013Rice formula" ], "rating_list": [ 8, 3, 8, 8 ], "soundness_list": [ 4, 2, 2, 4 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "fv2hL5n2mh", "primary_area": "applications to robotics, autonomy, planning", "n_author": 12, "author_info_list": [ { "name": "Bang Liu", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Assistant Professor" }, { "name": "Bangbang Liu", "gender": "Male", "institution": "deepwisdom", "country": "", "position": "Researcher" }, { "name": "Ceyao Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Intern" }, { "name": "Duyi Pan", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Guanghao Mei", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Jacky Kwok", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Sirui Hong", "gender": "Female", "institution": "DeepWisdom", "country": "CN", "position": "Researcher" }, { "name": "Tianqi Pang", "gender": "Female", "institution": "South China Normal University", "country": "CN", "position": "MS student" }, { "name": "Yaying Fei", "gender": "Female", "institution": "Southern Methodist University", "country": "US", "position": "PhD student" }, { "name": "Yizhang Lin", "gender": "Male", "institution": "DeepWisdom", "country": "CN", "position": "Researcher" }, { "name": "Yizhou Chi", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" }, { "name": "Alexander Wu", "gender": "Male", "institution": "DeepWisdom", "country": "", "position": "CEO" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 25, "n_ref": 42, "n_ref_all": 53, "n_fig": 4, "n_tab": 8, "L_tab": 2584, "n_element_tab": 367, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1478, "n_element_tab_1": 205, "formula_len_all": 430, "formula_len_all_1": 399, "len_all": 133438, "len_all_1": 59802, "len_abs": 1406, "len_title": 116, "len_sents": 35564, "len_sents_1": 27963, "n_sents": 279, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1467, "title": "SELA: Tree-Search Enhanced LLM Agents for Automated Machine Learning", "abs": "Automated Machine Learning (AutoML) approaches encompass traditional methods that optimize fixed pipelines for model selection and ensembling, as well as newer LLM-based frameworks that autonomously build pipelines. While LLM-based agents have shown promise in automating machine learning tasks, they often generate low-diversity and suboptimal code, even after multiple iterations. To overcome these limitations, we introduce Tree-$\\textbf{S}$earch $\\textbf{E}$nhanced $\\textbf{L}$LM $\\textbf{A}$gents ($\\textbf{SELA}$), an innovative agent-based system that leverages Monte Carlo Tree Search (MCTS) to optimize the AutoML process. By representing pipeline configurations as trees, our framework enables agents to conduct experiments intelligently and iteratively refine their strategies, facilitating a more effective exploration of the machine learning solution space.\nThis novel approach allows SELA to discover optimal pathways based on experimental feedback, improving the overall quality of the solutions. In an extensive evaluation across 20 machine learning datasets, we compare the performance of traditional and agent-based AutoML methods, demonstrating that SELA achieves a win rate of 65\\% to 80\\% against each baseline across all datasets. These results underscore the significant potential of agent-based strategies in AutoML, offering a fresh perspective on tackling complex machine learning challenges. The code will be open-sourced upon publication.", "keywords": [ "AutoML", "AutoDS", "LLM", "Agents", "Tree Search", "Planning" ], "rating_list": [ 5, 1, 5, 3 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "fuoM5YDBX4", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Chenyang Zhang", "gender": "Male", "institution": "The Univeristy of Hongkong", "country": "HK", "position": "PhD student" }, { "name": "Xuran Meng", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Postdoc" }, { "name": "Yuan Cao", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "Assistant Professor" } ], "n_formula": 297, "n_formula_1": 45, "n_ref_uni": 32, "n_ref": 72, "n_ref_all": 79, "n_fig": 4, "n_tab": 3, "L_tab": 1914, "n_element_tab": 328, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 45043, "formula_len_all_1": 3521, "len_all": 317436, "len_all_1": 68156, "len_abs": 1002, "len_title": 77, "len_sents": 74665, "len_sents_1": 27588, "n_sents": 915, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 885, "title": "Transformer Learns Optimal Variable Selection in Group-Sparse Classification", "abs": "Transformers have demonstrated remarkable success across various applications. However, the success of transformers have not been understood in theory. In this work, we give a case study of how transformers can be trained to learn a classic statistical model with \"group sparsity\", where the input variables form multiple groups, and the label only depends on the variables from one of the groups. We theoretically demonstrate that, a one-layer transformer trained by gradient descent can correctly leverage the attention mechanism to select variables, disregarding irrelevant ones and focusing on those beneficial for classification. We also demonstrate that a well-pretrained one-layer transformer can be adapted to new downstream tasks to achieve good prediction accuracy with a limited number of samples. Our study sheds light on how transformers effectively learn structured data.", "keywords": [ "transformer", "self-attention", "variable selection", "group-sparse classification" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "ftwMX4ORIS", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Jianhou Gan", "gender": "Male", "institution": "Yunnan Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Jun Wang", "gender": "Male", "institution": "Yunnan Normal University", "country": "CN", "position": "Assistant Professor" }, { "name": "Juxiang Zhou", "gender": "Female", "institution": "Yunnan Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Mingjie Wang", "gender": "Male", "institution": "Yunnan Normal University", "country": "CN", "position": "PhD student" }, { "name": "Zijie Li", "gender": "Male", "institution": "Yunnan Normal University", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 23, "n_ref": 56, "n_ref_all": 76, "n_fig": 6, "n_tab": 5, "L_tab": 3466, "n_element_tab": 281, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1516, "n_element_tab_1": 236, "formula_len_all": 442, "formula_len_all_1": 519, "len_all": 119997, "len_all_1": 50375, "len_abs": 1676, "len_title": 140, "len_sents": 39577, "len_sents_1": 22027, "n_sents": 257, "n_sents_1": 136, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1682, "title": "ExerCAKT: A Knowledge Tracing Model Based on GRU Capturing Contextual Features of Exercises", "abs": "Knowledge tracing aims to predict students' future performance based on their past interactions, helping online learning platforms and teachers assess learners' knowledge levels. This technology plays a critical role in achieving large-scale cognitive diagnosis. Recently, deep learning-based knowledge tracing models have demonstrated impressive results, with most research focusing on designing customized network architectures and novel optimization objectives. However, redundant parameters and overly complex loss functions can complicate model training and make it harder to maintain prediction accuracy. To further investigate the effectiveness of simple recurrent neural networks in this field, and to leverage their advantages in handling sequential exercise representation, this paper introduces a GRU-based knowledge tracing model named ExerCAKT (Exercise Context-Aware Knowledge Tracing). This model effectively captures contextual features of exercises and achieves robust knowledge state modeling through the use of a GRU-based knowledge state feature extractor and a GRU-based exercise feature extractor\u2014without relying on additional optimization objectives.The model's superior performance is validated through comparisons with baseline models, such as AKT and SIMPLEKT, on three public datasets in the knowledge tracing domain. Evaluations are conducted using AUC and ACC metrics at both the Knowledge Concept level and the question level. We validated that relying solely on simple recurrent neural networks, combined with appropriate representation methods, can still achieve excellent performance in this field. Our code will be available at xxx (Anonymous URL).", "keywords": [ "knowledge tracing", "deep learning", "recurrent neural networks", "knowledge state modeling" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "ftdJEiFudy", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Chuan-Sheng Foo", "gender": "Male", "institution": "Centre for Frontier AI Research, A*STAR", "country": "SG", "position": "Investigator" }, { "name": "Fayao Liu", "gender": "Female", "institution": "A*STAR", "country": "SG", "position": "Research Scientist" }, { "name": "Kennard Chan", "gender": "unknown", "institution": ", A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Weisi Lin", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Guosheng Lin", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 22, "n_ref": 60, "n_ref_all": 108, "n_fig": 30, "n_tab": 4, "L_tab": 1818, "n_element_tab": 285, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 184, "formula_len_all_1": 90, "len_all": 159643, "len_all_1": 56227, "len_abs": 1864, "len_title": 147, "len_sents": 58674, "len_sents_1": 25706, "n_sents": 499, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 99, "L_abs": 1880, "title": "Robust-PIFu: Robust Pixel-aligned Implicit Function for 3D Human Digitalization from a Single Image", "abs": "Existing methods for 3D clothed human digitalization perform well when the input image is captured in ideal conditions that assume the lack of any occlusion. However, in reality, images may often have occlusion problems such as incomplete observation of the human subject's full body, self-occlusion by the human subject, and non-frontal body pose. When given such input images, these existing methods fail to perform adequately. Thus, we propose Robust-PIFu, a pixel-aligned implicit model that capitalized on large-scale, pretrained latent diffusion models to address the challenge of digitalizing human subjects from non-ideal images that suffer from occlusions.\n\nRobust-PIfu offers four new contributions. Firstly, we propose a 'disentangling' latent diffusion model. This diffusion model, pretrained on billions of images, takes in any input image and removes external occlusions, such as inter-person occlusions, from that image. Secondly, Robust-PIFu addresses internal occlusions like self-occlusion by introducing a `penetrating' latent diffusion model. This diffusion model outputs multi-layered normal maps that by-pass occlusions caused by the human subject's own limbs or other body parts (i.e. self-occlusion). Thirdly, in order to incorporate such multi-layered normal maps into a pixel-aligned implicit model, we introduce our Layered-Normals Pixel-aligned Implicit Model, which improves the structural accuracy of predicted clothed human meshes. Lastly, Robust-PIFu proposes an optional super-resolution mechanism for the multi-layered normal maps. This addresses scenarios where the input image is of low or inadequate resolution. Though not strictly related to occlusion, this is still an important subproblem. Our experiments show that Robust-PIFu outperforms current SOTA methods both qualitatively and quantitatively. Our code will be released to the public.", "keywords": [ "Human Digitalization", "3D Human Avatar", "Pixel-aligned Implicit Models", "Latent Diffusion Models", "Occlusion", "Robustness" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ftHNJmogT1", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Chiyuan Zhang", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Danqi Chen", "gender": "Female", "institution": "Department of Computer Science, Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Haotian Liu", "gender": "unknown", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "", "position": "PhD student" }, { "name": "Luke Zettlemoyer", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Full Professor" }, { "name": "Luxi He", "gender": "Female", "institution": "Department of Computer Science, Princeton University", "country": "US", "position": "PhD student" }, { "name": "Peter Henderson", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Tinghao Xie", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Weijia Shi", "gender": "unknown", "institution": "University of Washington, Seattle", "country": "US", "position": "PhD student" }, { "name": "Yangsibo Huang", "gender": "Female", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Yue Wang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 46, "n_ref": 108, "n_ref_all": 148, "n_fig": 17, "n_tab": 5, "L_tab": 2025, "n_element_tab": 140, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1278, "n_element_tab_1": 75, "formula_len_all": 103, "formula_len_all_1": 107, "len_all": 196012, "len_all_1": 69693, "len_abs": 2053, "len_title": 107, "len_sents": 66704, "len_sents_1": 34378, "n_sents": 477, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1977, "title": "Fantastic Copyrighted Beasts and How (Not) to Generate Them", "abs": "Recent studies show that image and video generation models can be prompted to reproduce copyrighted content from their training data, raising serious legal concerns about copyright infringement. Copyrighted characters (e.g., Mario, Batman) present a significant challenge: at least one lawsuit has already awarded damages based on the generation of such characters. Consequently, commercial services like DALL\u00b7E have started deploying interventions.\nHowever, little research has systematically examined these problems: (1) Can users easily prompt models to generate copyrighted characters, even if it is unintentional?; (2) How effective are the existing mitigation strategies?\nTo address these questions, we introduce a novel evaluation framework with metrics that assess both the generated image\u2019s similarity to copyrighted characters and its consistency with user intent, grounded in a set of popular copyrighted characters from diverse studios and regions.\nWe show that state-of-the-art image and video generation models can still generate characters even if characters' names are not explicitly mentioned, sometimes with only two generic keywords (e.g., prompting with ``videogame, plumber'' consistently generates Nintendo's Mario character). \nWe also introduce semi-automatic techniques to identify such keywords or descriptions that trigger character generation. Within this framework, we study the effectiveness of mitigation strategies, including both existing methods and new strategies we propose. Our findings reveal that commonly used strategies, such as prompt rewriting in DALL\u00b7E, are insufficient as standalone guardrails. These strategies need to be supplemented with other approaches, such as negative prompting, to effectively reduce the unintended generation of copyrighted characters. Our work provides empirical grounding for discussions on copyright mitigation strategies and offers actionable insights for model deployers implementing these safeguards.", "keywords": [ "copyright", "copyrighted characters", "alignment", "evaluation", "image generation models" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "ftGnpZrW7P", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Danilo Comminiello", "gender": "Male", "institution": "Sapienza University of Rome", "country": "IT", "position": "Associate Professor" }, { "name": "Eleonora Grassucci", "gender": "Female", "institution": "Sapienza University of Rome", "country": "IT", "position": "Assistant Professor" }, { "name": "Giordano Cicchetti", "gender": "Male", "institution": "University of Roma \"La Sapienza\"", "country": "IT", "position": "PhD student" }, { "name": "Luigi Sigillo", "gender": "unknown", "institution": "University of Roma \"La Sapienza\"", "country": "IT", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 14, "n_ref_uni": 10, "n_ref": 24, "n_ref_all": 49, "n_fig": 6, "n_tab": 8, "L_tab": 2829, "n_element_tab": 455, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2817, "n_element_tab_1": 452, "formula_len_all": 1846, "formula_len_all_1": 1097, "len_all": 162147, "len_all_1": 65365, "len_abs": 1503, "len_title": 104, "len_sents": 44807, "len_sents_1": 27177, "n_sents": 302, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1516, "title": "Gramian Multimodal Representation Learning and Alignment", "abs": "Human perception integrates multiple modalities\u2014such as vision, hearing, and language\u2014into a unified understanding of the surrounding reality. While recent multimodal models have achieved significant progress by aligning pairs of modalities via contrastive learning, their solutions are unsuitable when scaling to multiple modalities. These models typically align each modality to a designated anchor without ensuring the alignment of all modalities with each other, leading to suboptimal performance in tasks requiring a joint understanding of multiple modalities. In this paper, we structurally rethink the pairwise conventional approach to multimodal learning and we present the novel Gramian Representation Alignment Measure (GRAM), which overcomes the above-mentioned limitations. GRAM learns and then aligns $n$ modalities directly in the higher-dimensional space in which modality embeddings lie by minimizing the Gramian volume of the $k$-dimensional parallelotope spanned by the modality vectors, ensuring the geometric alignment of all modalities simultaneously. GRAM can replace cosine similarity in any downstream method, holding for 2 to $n$ modalities and providing more meaningful alignment with respect to previous similarity measures. The novel GRAM-based contrastive loss function enhances the alignment of multimodal models in the higher-dimensional embedding space, leading to new state-of-the-art performance in downstream tasks such as video-audio-text retrieval and audio-video classification.", "keywords": [ "Multimodal Representation Learning", "Multimodal Alignment", "Multimodal Contrastive Learning" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "fswihJIYbd", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Pengwei Tang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Xiaolin Hu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 14, "n_ref_uni": 50, "n_ref": 127, "n_ref_all": 153, "n_fig": 3, "n_tab": 17, "L_tab": 5151, "n_element_tab": 508, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2514, "n_element_tab_1": 219, "formula_len_all": 1568, "formula_len_all_1": 1439, "len_all": 175605, "len_all_1": 66579, "len_abs": 1760, "len_title": 124, "len_sents": 41581, "len_sents_1": 30519, "n_sents": 313, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1875, "title": "ADePT: Adaptive Decomposed Prompt Tuning for Parameter-Efficient Fine-tuning", "abs": "Prompt Tuning (PT) enables the adaptation of Pre-trained Large Language Models (PLMs) to downstream tasks by optimizing a small amount of soft virtual tokens, which are prepended to the input token embeddings. Recently, Decomposed Prompt Tuning (DePT) has demonstrated superior adaptation capabilities by decomposing the soft prompt into a shorter soft prompt and a pair of low-rank matrices. The product of the pair of low-rank matrices is added to the input token embeddings to offset them. Additionally, DePT achieves faster inference compared to PT due to the shorter soft prompt. However, in this paper, we find that the \\textcolor{blue}{\\textcolor{blue}{position-based} token embedding offsets} of DePT restricts its ability to generalize across diverse model inputs, and that the shared embedding offsets across many token embeddings result in sub-optimization. To tackle these issues, we introduce \\textbf{A}daptive \\textbf{De}composed \\textbf{P}rompt \\textbf{T}uning (ADePT), which is composed of a short soft prompt and a shallow token-shared feed-forward neural network. ADePT utilizes the token-shared feed-forward neural network to learn the embedding offsets for each token, enabling adaptive embedding offsets that vary according to the model input and better optimization of token embedding offsets. This enables ADePT to achieve superior adaptation performance without requiring more inference time or additional trainable parameters compared to vanilla PT and its variants. In comprehensive experiments across \\textcolor{blue}{23} natural language processing (NLP) tasks and \\textcolor{blue}{4} typical PLMs of different scales, we show that ADePT consistently surpasses the leading parameter-efficient fine-tuning (PEFT) methods, and even outperforms the full fine-tuning baseline in certain scenarios. The code can be found in the supplementary materials.", "keywords": [ "Natural Language Processing", "Large Language Models", "Parameter-efficient Fine-tuning" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "fsrQuugqiF", "primary_area": "learning on time series and dynamical systems", "n_author": 8, "author_info_list": [ { "name": "Boli Chen", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Assistant Professor" }, { "name": "Daniel Giles", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Postdoc" }, { "name": "Sibo Cheng", "gender": "Male", "institution": "Ecole Nationale des Ponts et Chausees", "country": "FR", "position": "Assistant Professor" }, { "name": "Xiao Xue", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Researcher" }, { "name": "Xiaoyuan Cheng", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Yi He", "gender": "unknown", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Yiming Yang", "gender": "unknown", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Yukun Hu", "gender": "unknown", "institution": "University College London, University of London", "country": "", "position": "Associate Professor" } ], "n_formula": 45, "n_formula_1": 18, "n_ref_uni": 71, "n_ref": 102, "n_ref_all": 133, "n_fig": 10, "n_tab": 4, "L_tab": 906, "n_element_tab": 76, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 877, "n_element_tab_1": 67, "formula_len_all": 3408, "formula_len_all_1": 1006, "len_all": 207747, "len_all_1": 69738, "len_abs": 1658, "len_title": 121, "len_sents": 58270, "len_sents_1": 32548, "n_sents": 439, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1562, "title": "Tensor-Var: Variational Data Assimilation in Tensor Product Feature Space", "abs": "Variational data assimilation estimates the dynamical system states by minimizing a cost function that fits the numerical models with observational data. The widely used method, four-dimensional variational assimilation (4D-Var), has two primary challenges: (1) computationally demanding for complex nonlinear systems, and (2) relying on state-observation mappings, which are often impractical. Deep learning (DL) has been used as a more expressive class of efficient model approximators to address these challenges. However, integrating such models into 4D-Var remains challenging due to their inherent nonlinearities and the lack of theoretical guarantees for consistency in assimilation results. In this paper, we propose *Tensor-Var* to address these challenges using kernel Conditional Mean Embedding (CME). Tensor-Var improves optimization efficiency by characterizing system dynamics and state-observation mappings as linear operators, leading to a convex cost function in the feature space. Furthermore, our method provides a new perspective to incorporate CME into 4D-Var, offering theoretical guarantees of consistent assimilation results between the original and feature spaces. To improve scalability, we propose a method to learn deep features (DFs) using neural networks within the Tensor-Var framework. Experiments on chaotic systems and global weather prediction with real-time observations show that Tensor-Var outperforms conventional and DL hybrid 4D-Var baselines in accuracy while achieving efficiency comparable to the static 3D-Var method.", "keywords": [ "Variational Data Assimilation", "Dynamical System", "Weather Forecasting", "Representation Learning" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "fsmEuS5ZNg", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Aidong Men", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Full Professor" }, { "name": "Jiangbo Pei", "gender": "Male", "institution": "Beijing University of Post and Telecommunication", "country": "", "position": "Undergrad student" }, { "name": "Li Xiao", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Full Professor" }, { "name": "Ruizhe Li", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 22, "n_ref": 37, "n_ref_all": 50, "n_fig": 5, "n_tab": 9, "L_tab": 4406, "n_element_tab": 543, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1809, "n_element_tab_1": 306, "formula_len_all": 678, "formula_len_all_1": 650, "len_all": 126159, "len_all_1": 64987, "len_abs": 1464, "len_title": 104, "len_sents": 32300, "len_sents_1": 29028, "n_sents": 252, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1470, "title": "Multi-Model Induced Source-free Video Domain Adaptation", "abs": "Existing Source-free Video Domain Adaptation (SFVDA) aims to learn a target video model for an unlabeled target domain by transferring knowledge from a labeled source domain using a single pre-trained source video model. In this paper, we explore a new SFVDA setting where multiple source domains exist, each offering a library of source models with different architectures. This setting offers both opportunities and challenges: while the presence of multiple source models enriches the pool of transferable knowledge, it also increases the risk of negative transfer due to inappropriate source knowledge. To tackle these challenges, we introduce the Multiple-Source-Video-Model Aggregation Framework (MSVMA), comprising two key modules. The first module, termed Multi-level Instance Transferability Calibration (MITC), enhances existing uncertainty-based transferability estimation metrics by incorporating scale information from both group and dataset levels. This integration facilitates accurate transferability estimation at the instance level across diverse models. The second module, termed Instance-level Multi Video Model Aggregation (IMVMA), leverages the calculated instance-level transferability to guide a path generation network. This network produces instance-specific weights for unsupervised aggregation of source models. Empirical results from three video domain adaptation datasets demonstrate the state-of-the-art performance of our MSVMA framework.", "keywords": [ "Domain Adaptation", "Multi-Task Learning" ], "rating_list": [ 3, 5, 1 ], "soundness_list": [ 3, 3, 1 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "fsX9nFwMNj", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Anh Tuan Luu", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "Assistant Professor" }, { "name": "Feng-Lin Li", "gender": "Male", "institution": "Shopee", "country": "", "position": "Researcher" }, { "name": "Huimin Xu", "gender": "Female", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" }, { "name": "WANG CHEN", "gender": "Male", "institution": "SHOPEE", "country": "SG", "position": "Researcher" }, { "name": "Wei Zhang", "gender": "Male", "institution": "sea group", "country": "SG", "position": "Researcher" }, { "name": "Xin Mao", "gender": "Male", "institution": "NTU", "country": "", "position": "Postdoc" }, { "name": "jin ziqi", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Researcher" } ], "n_formula": 35, "n_formula_1": 15, "n_ref_uni": 27, "n_ref": 78, "n_ref_all": 109, "n_fig": 8, "n_tab": 9, "L_tab": 8622, "n_element_tab": 374, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1395, "n_element_tab_1": 102, "formula_len_all": 6332, "formula_len_all_1": 1906, "len_all": 172396, "len_all_1": 64247, "len_abs": 1553, "len_title": 80, "len_sents": 49271, "len_sents_1": 28169, "n_sents": 372, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1597, "title": "As Simple as Fine-tuning: LLM Alignment via Bidirectional Negative Feedback Loss", "abs": "Direct Preference Optimization (DPO) has emerged as a more computationally efficient alternative to Reinforcement Learning from Human Feedback (RLHF) with Proximal Policy Optimization (PPO), eliminating the need for reward models and online sampling. Despite these benefits, DPO and its variants remain sensitive to hyper-parameters and prone to instability, particularly on mathematical datasets. We argue that these issues arise from the unidirectional likelihood-derivative negative feedback inherent in the log-likelihood loss function.\nTo address this, we propose a novel LLM alignment loss that establishes a stable Bidirectional Negative Feedback (BNF) during optimization. \nOur proposed BNF loss eliminates the need for pairwise contrastive losses and does not require any extra tunable hyper-parameters or pairwise preference data, streamlining the alignment pipeline to be as simple as supervised fine-tuning.\nWe conduct extensive experiments across two challenging QA benchmarks and four reasoning benchmarks. \nThe experimental results show that BNF achieves comparable performance to the best methods on QA benchmarks, while its performance decrease on the four reasoning benchmarks is significantly lower compared to the best methods, thus striking a better balance between value alignment and reasoning ability. \nIn addition, we further validate the performance of BNF on non-pairwise datasets, and conduct in-depth analysis of log-likelihood and logit shifts across different preference optimization methods.\nWe will release all the source code, checkpoints, and datasets on GitHub.", "keywords": [ "LLM Alignemnt", "Preference Learning", "Text Generation" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "fsEzHMqbkf", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Baris Coskun", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Qian Cui", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Wei Ding", "gender": "Male", "institution": "Amazon", "country": "CA", "position": "Researcher" }, { "name": "Zhilu Zhang", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Vianne Gao", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 61, "n_ref": 107, "n_ref_all": 132, "n_fig": 7, "n_tab": 17, "L_tab": 10920, "n_element_tab": 1896, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 3199, "n_element_tab_1": 559, "formula_len_all": 1904, "formula_len_all_1": 1239, "len_all": 196478, "len_all_1": 77728, "len_abs": 1440, "len_title": 115, "len_sents": 46482, "len_sents_1": 31682, "n_sents": 377, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1444, "title": "Conditional Density Ratio Score for Post Hoc Deep Outlier Detection", "abs": "The ability to accurately identify out-of-distribution (OOD) samples is essential not only as a stand-alone machine learning task but also for maintaining the reliability and safety of machine learning systems. Within this domain, post hoc density estimators like the energy score are popular ways for detecting OOD samples. However, most of the existing post hoc density estimation have mainly focused on marginalizing the conditional distributions over all possible classes. In this paper, we introduce the Conditional Density Ratio (CDR) score, a principled post hoc density estimator that leverages both a class-conditional generative model in the latent space and a discriminative classifier model, allowing us to estimate the marginal densities of the latent representation without marginalization. We demonstrate that a key component to the success of the CDR score lies in correctly calibrating the two models and propose a simple yet effective method to automatically tune the temperature parameter without the need for out-of-distribution samples. We illustrate the general compatibility of the proposed method with two popular density estimators, the kernel density estimator and the Mahalanobis estimator. Through experiments on a wide range of OOD benchmark tasks, we verify the effectiveness of the proposed method and advocate it as an easy-to-implement baseline that can achieve competitive performance in most tested scenarios.", "keywords": [ "Deep Learning", "Outlier Detection", "Out-Of-Distribution Detection" ], "rating_list": [ 3, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 3, 3, 2 ] }, { "paper_id": "fsDZwS49uY", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Jing Tang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Liang Feng", "gender": "Male", "institution": "Chongqing University", "country": "CN", "position": "Full Professor" }, { "name": "Linqi Song", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Wei Shi", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Xiaodan Liang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Xiongwei Han", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Yinya Huang", "gender": "unknown", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Yiwei Wang", "gender": "Male", "institution": "University of California, Merced", "country": "US", "position": "Assistant Professor" }, { "name": "Zhijiang Guo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhicheng Yang", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 84, "n_ref_all": 115, "n_fig": 8, "n_tab": 17, "L_tab": 8034, "n_element_tab": 894, "n_fig_1": 7, "n_tab_1": 8, "L_tab_1": 1801, "n_element_tab_1": 333, "formula_len_all": 21, "formula_len_all_1": 0, "len_all": 243366, "len_all_1": 68058, "len_abs": 1520, "len_title": 126, "len_sents": 74920, "len_sents_1": 29313, "n_sents": 664, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1560, "title": "OptiBench Meets ReSocratic: Measure and Improve LLMs for Optimization Modeling", "abs": "Large language models (LLMs) have exhibited their problem-solving abilities in mathematical reasoning. Solving realistic optimization (OPT) problems in application scenarios requires advanced and applied mathematics ability. However, current OPT benchmarks that merely solve linear programming are far from complex realistic situations. In this work, we propose **OptiBench**, a benchmark for End-to-end optimization problem-solving with human-readable inputs and outputs. **OptiBench** contains rich optimization problems, including linear and nonlinear programming with or without tabular data, which can comprehensively evaluate LLMs' solving ability. In our benchmark, LLMs are required to call a code solver to provide precise numerical answers.\nFurthermore, to alleviate the data scarcity for optimization problems, and to bridge the gap between open-source LLMs on a small scale (e.g., Llama-3-8b) and closed-source LLMs (e.g., GPT-4), we further propose a data synthesis method namely ***ReSocratic***. Unlike general data synthesis methods that proceed from questions to answers, \\ReSocratic first incrementally synthesizes formatted optimization demonstration with mathematical formulations step by step and then back-translates the generated demonstrations into questions. Based on this, we synthesize the ***ReSocratic-29k*** dataset. We further conduct supervised fine-tuning with ***ReSocratic-29k*** on multiple open-source models. Experimental results show that ***ReSocratic-29k*** significantly improves the performance of open-source models.", "keywords": [ "large language models; optimization problem; data synthesis" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "fs2Z2z3GRx", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Xiangming Meng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yichi Zhang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Yici Yan", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "PhD student" }, { "name": "Zhizhen Zhao", "gender": "unknown", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Associate Professor" } ], "n_formula": 44, "n_formula_1": 24, "n_ref_uni": 36, "n_ref": 67, "n_ref_all": 131, "n_fig": 8, "n_tab": 9, "L_tab": 2676, "n_element_tab": 188, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 916, "n_element_tab_1": 71, "formula_len_all": 2746, "formula_len_all_1": 1356, "len_all": 149503, "len_all_1": 59111, "len_abs": 907, "len_title": 112, "len_sents": 38065, "len_sents_1": 23640, "n_sents": 366, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 913, "title": "FIG: Flow with Interpolant Guidance for Linear Inverse Problems", "abs": "Diffusion and flow matching models have been recently used to solve various linear inverse problems such as image restoration. Using a pre-trained diffusion or flow-matching model as a prior, most existing methods modify the reverse-time sampling process by incorporating the likelihood information from the measurement. However, they struggle in challenging scenarios, e.g., in case of high measurement noise or severe ill-posedness. In this paper, we propose Flow with Interpolant Guidance (FIG), an algorithm where the reverse-time sampling is efficiently guided with measurement interpolants through theoretically justified schemes. Experimentally, we demonstrate that FIG efficiently produce highly competitive results on a variety of linear image reconstruction tasks on natural image datasets. We improve upon state-of-the-art baseline algorithms, especially for challenging tasks. Code will be released.", "keywords": [ "Flow Matching", "Diffusion Model", "Linear Inverse Problems" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "frsg32u0rO", "primary_area": "learning theory", "n_author": 7, "author_info_list": [ { "name": "Ahmad Beirami", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Ananda Theertha Suresh", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Asaf Aharoni", "gender": "Male", "institution": "Google", "country": "IL", "position": "Researcher" }, { "name": "Jae Hun Ro", "gender": "Male", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Uri Mendlovic", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Yaniv Leviathan", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Ziteng Sun", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 106, "n_formula_1": 16, "n_ref_uni": 30, "n_ref": 55, "n_ref_all": 126, "n_fig": 5, "n_tab": 12, "L_tab": 2529, "n_element_tab": 286, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 984, "n_element_tab_1": 60, "formula_len_all": 14279, "formula_len_all_1": 2078, "len_all": 192413, "len_all_1": 63721, "len_abs": 3509, "len_title": 99, "len_sents": 54281, "len_sents_1": 27451, "n_sents": 497, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1278, "title": "Block Verification Accelerates Speculative Decoding", "abs": "Speculative decoding is an effective method for lossless acceleration of large language models during inference. It uses a fast model to draft a block of tokens which are then verified in parallel by the target model, and provides a guarantee that the output is distributed identically to a sample from the target model. In prior works, draft verification is performed independently token-by-token. Surprisingly, we show that this approach is not optimal. We propose *Block Verification*, a simple draft verification algorithm that verifies the entire block jointly and provides additional wall-clock speedup. We prove that the proposed mechanism is optimal in the expected number of tokens produced each iteration and specifically is never worse than the standard token-level verification.\nEmpirically, block verification provides modest but consistent wall-clock speedups over the standard token verification algorithm of 5\\%-8\\% in a range of tasks and datasets.\nGiven that block verification does not increase code complexity, maintains the strong lossless guarantee of the standard speculative decoding verification algorithm, cannot deteriorate performance, and, in fact, consistently improves it, it can be used as a good default in speculative decoding implementations.", "keywords": [ "llm efficiency", "speculative decoding", "distribution coupling" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "friHAl5ofG", "primary_area": "applications to robotics, autonomy, planning", "n_author": 16, "author_info_list": [ { "name": "Chuyuan Fu", "gender": "Female", "institution": "Google", "country": "US", "position": "software engineer" }, { "name": "Danny Driess", "gender": "unknown", "institution": "Physical Intelliigence", "country": "US", "position": "Researcher" }, { "name": "Dhruv Shah", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Dinesh Jayaraman", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Assistant Professor" }, { "name": "Dorsa Sadigh", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Fei Xia", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Jacky Liang", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Osbert Bastani", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Assistant Professor" }, { "name": "Peng Xu", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Sean Kirmani", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Ted Xiao", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Tingnan Zhang", "gender": "Male", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Wenhao Yu", "gender": "Male", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Zhuo Xu", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Donald Joseph Hejna III", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Yecheng Jason Ma", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 55, "n_ref": 92, "n_ref_all": 135, "n_fig": 18, "n_tab": 7, "L_tab": 932, "n_element_tab": 76, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 817, "n_element_tab_1": 71, "formula_len_all": 569, "formula_len_all_1": 642, "len_all": 180644, "len_all_1": 67174, "len_abs": 1660, "len_title": 100, "len_sents": 48440, "len_sents_1": 33184, "n_sents": 358, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 52, "L_abs": 1666, "title": "Vision Language Models are In-Context Value Learners", "abs": "Predicting temporal progress from visual trajectories is important for intelligent robots that can learn, adapt, and improve. However, learning such progress estimator, or temporal value function, across different tasks and domains requires both a large amount of diverse data and methods which can scale and generalize. To address these challenges, we present Generative Value Learning (GVL), a universal value function estimator that leverages the world knowledge embedded in vision-language models (VLMs) to predict task progress. Naively asking a VLM to predict values for a video sequence performs poorly due to the strong temporal correlation between successive frames. Instead, GVL poses value estimation as a temporal ordering problem over shuffled video frames; this seemingly more challenging task encourages VLMs to more fully exploit their underlying semantic and temporal grounding capabilities to differentiate frames based on their perceived task progress, consequently producing significantly better value predictions. Without any robot or task specific training, GVL can in-context zero-shot and few-shot predict effective values for more than 300 distinct real-world tasks across diverse robot platforms, including challenging bimanual manipulation tasks. Furthermore, we demonstrate that GVL permits flexible multi-modal in-context learning via examples from heterogeneous tasks and embodiments, such as human videos. The generality of GVL enables various downstream applications pertinent to visuomotor policy learning, including dataset filtering, success detection, and value-weighted regression -- all without any model training or finetuning.", "keywords": [ "robot learning", "vision-language model", "value estimation", "manipulation" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "frbfEqZX5R", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Utkarsh Ojha", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Yuyang Ji", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Researcher" }, { "name": "Zeyi Huang", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Yong Jae Lee", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 35, "n_ref": 57, "n_ref_all": 79, "n_fig": 6, "n_tab": 3, "L_tab": 845, "n_element_tab": 135, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 409, "n_element_tab_1": 65, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 117747, "len_all_1": 66975, "len_abs": 551, "len_title": 122, "len_sents": 37708, "len_sents_1": 32935, "n_sents": 273, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1423, "title": "Do Vision Models Develop Human-Like Progressive Difficulty Understanding?", "abs": "When a human undertakes a test, their responses likely follow a pattern: if they answered an easy question (2x3) incorrectly, they would likely answer a more difficult one (2x3x4) incorrectly; and if they answered a difficult question correctly, they would likely answer the easy one correctly. Anything else hints at memorization. Do current visual recognition models exhibit a similarly structured learning capacity? In this work, we consider the task of image classification and study if those models' responses follow that pattern. Since real images aren't labeled with difficulty, we first create a dataset of 100 categories, 10 attributes, and 3 difficulty levels using recent generative models: for each category (e.g., dog) and attribute (e.g., occlusion), we generate images of increasing difficulty (e.g., a dog without occlusion, a dog only partly visible). We find that most of the models do in fact behave similarly to the aforementioned pattern around 80-90\\% of the time. Using this property, we then present a new way to evaluate those models' image recognition ability. Instead of testing the model on every possible test image, we create an adaptive test akin to GRE, in which the model's performance on the current round of images determines the test images in the next round. This allows the model to skip over questions too easy/hard for itself, and helps us get its overall performance in fewer steps.", "keywords": [ "Understanding Image Classification; Adaptive Testing; Synthetic Benchmarks" ], "rating_list": [ 5, 6, 3, 1 ], "soundness_list": [ 3, 3, 1, 1 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 3, 2, 1, 1 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "frZVMBbqQJ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Feiyang Lu", "gender": "Male", "institution": "ShanghaiTech University", "country": "", "position": "MS student" }, { "name": "Wenjie Wang", "gender": "Female", "institution": "ShanghaiTech University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yukai Zhou", "gender": "Male", "institution": "ShanghaiTech University", "country": "", "position": "MS student" }, { "name": "Zhan Qin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zhijie Huang", "gender": "Male", "institution": "ShanghaiTech University", "country": "", "position": "Undergrad student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 39, "n_ref": 87, "n_ref_all": 136, "n_fig": 10, "n_tab": 9, "L_tab": 2459, "n_element_tab": 184, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 833, "n_element_tab_1": 87, "formula_len_all": 1144, "formula_len_all_1": 658, "len_all": 201627, "len_all_1": 69447, "len_abs": 1430, "len_title": 101, "len_sents": 68614, "len_sents_1": 31718, "n_sents": 505, "n_sents_1": 248, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1454, "title": "Don\u2019t Say No: Jailbreaking LLM by Suppressing Refusal", "abs": "Ensuring the safety alignment of Large Language Models (LLMs) is crucial to generating responses consistent with human values. Despite their ability to recognize and avoid harmful queries, LLMs are vulnerable to jailbreaking attacks, where carefully crafted prompts seduce them to produce toxic content. One category of jailbreak attacks is reformulating the task as an optimization by eliciting the LLM to generate affirmative responses. However, such optimization objective has its own limitations, such as the restriction on the predefined objectionable behaviors, leading to suboptimal attack performance. In this study, we first uncover the reason why vanilla target loss is not optimal, then we explore and enhance the loss objective and introduce the $\\textit{DSN}$ (Don't Say No) attack, which achieves successful attack by suppressing refusal. Another challenge in studying jailbreak attacks is the evaluation, as it is difficult to directly and accurately assess the harmfulness of the responses. The existing evaluation such as refusal keyword matching reveals numerous false positive and false negative instances. To overcome this challenge, we propose an Ensemble Evaluation pipeline that novelly incorporates Natural Language Inference (NLI) contradiction assessment and two external LLM evaluators. Extensive experiments demonstrate the potential of the $\\textit{DSN}$ and effectiveness of Ensemble Evaluation compared to baseline methods.", "keywords": [ "prompting", "security and privacy", "red teaming", "applications", "robustness" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "fr7cLDfNNU", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 1, "author_info_list": [ { "name": "Zeyu Michael Li", "gender": "Male", "institution": "Duke University, Duke University", "country": "US", "position": "Undergrad student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 47, "n_ref": 124, "n_ref_all": 185, "n_fig": 5, "n_tab": 14, "L_tab": 7703, "n_element_tab": 855, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 6540, "n_element_tab_1": 681, "formula_len_all": 343, "formula_len_all_1": 306, "len_all": 198399, "len_all_1": 83009, "len_abs": 1035, "len_title": 145, "len_sents": 46998, "len_sents_1": 30326, "n_sents": 363, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 993, "title": "Using Interleaved Ensemble Unlearning to Keep Backdoors at Bay for Finetuning Vision Transformers", "abs": "Vision Transformers (ViTs) have become popular in computer vision tasks. Backdoor attacks, which trigger undesirable behaviours in models during inference, threaten ViTs' performance, particularly in security-sensitive tasks. Although backdoor defences have been developed for Convolutional Neural Networks (CNNs), they are less effective for ViTs, and defences tailored to ViTs are scarce. To address this, we present Interleaved Ensemble Unlearning (IEU), a method for finetuning clean ViTs on backdoored datasets. In stage 1, a shallow ViT is finetuned to have high confidence on backdoored data and low confidence on clean data. In stage 2, the shallow ViT acts as a \"gate\" to block potentially poisoned data from the defended ViT. This data is added to an unlearn set and asynchronously unlearnt via gradient ascent. We demonstrate IEU's effectiveness on three datasets against 11 state-of-the-art backdoor attacks and show its versatility by applying it to different model architectures.", "keywords": [ "Vision Transformer", "Backdoor Defence" ], "rating_list": [ 6, 3, 8, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 3, 2, 4, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "fqtaADSGEe", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "BOHUAI WU", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "MS student" }, { "name": "Fangyun Wei", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" }, { "name": "Hongyang Zhang", "gender": "Male", "institution": "School of Computer Science, University of Waterloo", "country": "CA", "position": "Assistant Professor" }, { "name": "Jierun Chen", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Jingjing Zhao", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "MS student" }, { "name": "S.-H. Chan", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Sizhe Song", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zhuoxuan Peng", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 54, "n_ref": 107, "n_ref_all": 135, "n_fig": 14, "n_tab": 6, "L_tab": 5291, "n_element_tab": 417, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 3691, "n_element_tab_1": 162, "formula_len_all": 72, "formula_len_all_1": 0, "len_all": 173087, "len_all_1": 50515, "len_abs": 1460, "len_title": 144, "len_sents": 34185, "len_sents_1": 21403, "n_sents": 267, "n_sents_1": 164, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1468, "title": "Revisiting Referring Expression Comprehension Evaluation in the Era of Large Multimodal Models", "abs": "Referring expression comprehension (REC) involves localizing a target instance based on a textual description. Recent advancements in REC have been driven by large multimodal models (LMMs) like CogVLM, which achieved 92.44\\% accuracy on RefCOCO. However, this study questions whether existing benchmarks such as RefCOCO, RefCOCO+, and RefCOCOg, capture LMMs' comprehensive capabilities. We begin with a manual examination of these benchmarks, revealing high labeling error rates: 14\\% in RefCOCO, 24\\% in RefCOCO+, and 5\\% in RefCOCOg, which undermines the authenticity of evaluations. We address this by excluding problematic instances and reevaluating several LMMs capable of handling the REC task, showing significant accuracy improvements, thus highlighting the impact of benchmark noise. In response, we introduce Ref-L4, a comprehensive REC benchmark, specifically designed to evaluate modern REC models. Ref-L4 is distinguished by four key features: 1) a substantial sample size with 45,341 annotations; 2) a diverse range of object categories with 365 distinct types and varying instance scales from 30 to 3,767; 3) lengthy referring expressions averaging 24.2 words; and 4) an extensive vocabulary comprising 22,813 unique words. We evaluate a total of 24 large models on Ref-L4 and provide valuable insights. The cleaned versions of RefCOCO, RefCOCO+, and RefCOCOg, as well as our Ref-L4 benchmark and evaluation code will be made available to the community.", "keywords": [ "Referring Expression Comprehension", "Benchmark", "Dataset", "Evaluation", "Large Multimodal Models" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "fq24pEb8SL", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Fanqi Wan", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Longguang Zhong", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Tianyuan Shi", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "PhD student" }, { "name": "Xiaojun Quan", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Ziyi Yang", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 46, "n_ref": 119, "n_ref_all": 161, "n_fig": 6, "n_tab": 9, "L_tab": 3157, "n_element_tab": 452, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1679, "n_element_tab_1": 274, "formula_len_all": 1520, "formula_len_all_1": 856, "len_all": 191720, "len_all_1": 70553, "len_abs": 1325, "len_title": 113, "len_sents": 57181, "len_sents_1": 31695, "n_sents": 391, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1385, "title": "Weighted-Reward Preference Optimization for Implicit Model Fusion", "abs": "While fusing heterogeneous open-source LLMs with varying architectures and sizes can potentially integrate the strengths of different models, existing fusion methods face significant challenges, such as vocabulary alignment and merging distribution matrices. These procedures are not only complex but also prone to introducing noise and errors. In this paper, we propose an implicit fusion method, Weighted-Reward Preference Optimization (WRPO), which leverages preference optimization between the source LLMs and the target LLM to transfer their capabilities effectively. WRPO eliminates the need for vocabulary alignment and matrix fusion and can be efficiently scaled to accommodate various LLMs. To address distributional deviations between the source and target LLMs, WRPO introduces a progressive adaptation strategy that gradually shifts reliance on preferred examples from the target LLM to the source LLMs. Extensive experiments on the MT-Bench, AlpacaEval-2, and Arena-Hard benchmarks demonstrate that WRPO consistently outperforms existing knowledge fusion methods and various fine-tuning baselines. When applied to LLaMA3-8B-Instruct as the target model, WRPO achieves a length-controlled win rate of 55.9\\% against GPT-4-Preview-1106 on AlpacaEval-2 and a win rate of 46.2\\% against GPT-4-0314 on Arena-Hard. Our code is available at \\url{https://github.com/SLIT-AI/WRPO}.", "keywords": [ "Model Fusion", "Preference Optimization", "Large Language Models" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "fpvgSDKXGY", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Bahjat Kawar", "gender": "Male", "institution": "Apple", "country": "IL", "position": "Researcher" }, { "name": "Beomsu Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "James Thornton", "gender": "unknown", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Jong Chul Ye", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Michal Klein", "gender": "Male", "institution": "Apple", "country": "FR", "position": "Researcher" }, { "name": "Yu-Guan Hsieh", "gender": "Male", "institution": "Apple", "country": "FR", "position": "Postdoc" }, { "name": "marco cuturi", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 71, "n_formula_1": 23, "n_ref_uni": 57, "n_ref": 176, "n_ref_all": 246, "n_fig": 11, "n_tab": 11, "L_tab": 3196, "n_element_tab": 412, "n_fig_1": 10, "n_tab_1": 9, "L_tab_1": 2906, "n_element_tab_1": 345, "formula_len_all": 6123, "formula_len_all_1": 2266, "len_all": 206052, "len_all_1": 79387, "len_abs": 1007, "len_title": 103, "len_sents": 55964, "len_sents_1": 28551, "n_sents": 501, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1037, "title": "Simple ReFlow: Improved Techniques for Fast Flow Models", "abs": "Diffusion and flow-matching models achieve remarkable generative performance but at the cost of many neural function evaluations (NFE), which slows inference and limits applicability to time-critical tasks. The ReFlow procedure can accelerate sampling by straightening generation trajectories. But it is an iterative procedure, typically requiring training on simulated data, and results in reduced sample quality. To mitigate sample deterioration, we examine the design space of ReFlow and highlight potential pitfalls in prior heuristic practices. We then propose seven improvements for training dynamics, learning and inference, which are verified with thorough ablation studies on CIFAR10 $32 \\times 32$, AFHQv2 $64 \\times 64$, and FFHQ $64 \\times 64$. Combining all our techniques, we achieve state-of-the-art FID scores (without / with guidance, resp.) for fast generation via neural ODEs: $2.23$ / $1.98$ on CIFAR10, $2.30$ / $1.91$ on AFHQv2, $2.84$ / $2.67$ on FFHQ, and $3.49$ / $1.74$ on ImageNet-64, all with merely $9$ NFEs.", "keywords": [ "flow matching", "diffusion models", "reflow", "optimal transport", "generative modeling" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "fp77Ln5Hcc", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 1, "author_info_list": [ { "name": "Emile R Richard", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Postdoc" } ], "n_formula": 31, "n_formula_1": 15, "n_ref_uni": 26, "n_ref": 52, "n_ref_all": 88, "n_fig": 8, "n_tab": 2, "L_tab": 578, "n_element_tab": 7, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2407, "formula_len_all_1": 1000, "len_all": 139472, "len_all_1": 72080, "len_abs": 5589, "len_title": 108, "len_sents": 48085, "len_sents_1": 33340, "n_sents": 445, "n_sents_1": 290, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1228, "title": "Depth Extrapolation of Decoders Trained on Nested Structures", "abs": "Reasoning problems with deeply nested formal statements are challenging for humans and machines alike. We investigate how next-token predictors learn such structures, and whether they extrapolate to more deeply nested cases, within a single inference pass. A case study of Boolean logic simplification demonstrates that a specialized decoder Transformer seems to perform well when it overfits, but fails at extrapolating. To understand if this limitation is universal, we propose a theoretical grounding of memorization in a self-attention head. We apply this theory to a simpler problem: completion of a bounded stack of parentheses. From the theoretical construction we derive a closed-form model that perfectly fits a single sequence training set. We prove that it also completes any out-of-sample parentheses prefix, regardless of the context depth. In contrast, we observe that decoder Transformers trained with gradient descent on this task fail at depth extrapolation. Gradient-trained decoders demand large samples and a high-dimensional embedding space to achieve high accuracy on test sets nearly as deep as the training set. However, when the gap between training and test depths widens, gradient-trained models fail.", "keywords": [ "transformers", "reasoning", "nested structures", "embeddings", "generalization" ], "rating_list": [ 3, 3, 6, 6 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 1, 1, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "fp6t3F669F", "primary_area": "datasets and benchmarks", "n_author": 13, "author_info_list": [ { "name": "Akbir Khan", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Bart\u0142omiej Cupia\u0142", "gender": "Male", "institution": "University of Warsaw", "country": "PL", "position": "PhD student" }, { "name": "Davide Paglieri", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Eduardo Pignatelli", "gender": "Male", "institution": "University College London", "country": "", "position": "PhD student" }, { "name": "Jack Parker-Holder", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Jakob Nicolaus Foerster", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Lerrel Joseph Pinto", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Maciej Wolczyk", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Rob Fergus", "gender": "Male", "institution": "New York University", "country": "US", "position": "Professor" }, { "name": "Ulyana Piterbarg", "gender": "Female", "institution": "Facebook", "country": "FR", "position": "Researcher" }, { "name": "\u0141ukasz Kuci\u0144ski", "gender": "Male", "institution": "IDEAS NCBR", "country": "PL", "position": "Senior Research Scientist" }, { "name": "Samuel Dai Coward", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Tim Rocktaeschel", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Director" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 77, "n_ref": 119, "n_ref_all": 141, "n_fig": 4, "n_tab": 8, "L_tab": 2809, "n_element_tab": 227, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 655, "n_element_tab_1": 56, "formula_len_all": 871, "formula_len_all_1": 191, "len_all": 247582, "len_all_1": 63127, "len_abs": 1507, "len_title": 125, "len_sents": 66741, "len_sents_1": 32064, "n_sents": 497, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1513, "title": "AgentQuest: Benchmarking LLM and VLM Agents on Long-Horizon Interactive Tasks", "abs": "Large Language Models (LLMs) and Vision Language Models (VLMs) possess extensive knowledge and exhibit promising reasoning abilities, however, they still struggle to perform well in complex, dynamic environments. Real-world tasks require handling intricate interactions, advanced spatial reasoning, long-term planning, and continuous exploration of new strategies\u2014areas in which we lack effective methodologies for comprehensively evaluating these capabilities. To address this gap, we introduce AgentQuest, a novel benchmark designed to assess the agentic capabilities of LLMs and VLMs through a diverse set of challenging games. Our benchmark incorporates a range of existing reinforcement learning environments with varying levels of difficulty, including tasks that are solvable by non-expert humans in seconds to extremely challenging ones that may take years to master (e.g., the NetHack Learning Environment). \nWe devise fine-grained metrics to measure performance and conduct an extensive evaluation of several popular open-source and closed-source LLMs and VLMs. Our findings indicate that while current models achieve partial success in the easier games, they struggle significantly with more challenging tasks. Notably, we observe severe deficiencies in vision-based decision-making, as models perform worse when visual representations of the environments are provided. We release AgentQuest as an open and user-friendly benchmark to facilitate future research and development in the agentic community.", "keywords": [ "LLM", "VLM", "Agents", "Benchmark", "RL" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "fovPyqPcKY", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Can Yang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Jiaxin ZHANG", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "MS student" }, { "name": "Jishan Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Tianhao Chen", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Xin Xu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zitong Chao", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 44, "n_ref": 134, "n_ref_all": 180, "n_fig": 11, "n_tab": 22, "L_tab": 5824, "n_element_tab": 510, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 798, "n_element_tab_1": 28, "formula_len_all": 129, "formula_len_all_1": 129, "len_all": 228188, "len_all_1": 55203, "len_abs": 1553, "len_title": 167, "len_sents": 48363, "len_sents_1": 25184, "n_sents": 407, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 118, "L_abs": 1584, "title": "UGMathBench: A Diverse and Dynamic Benchmark for Undergraduate-Level Mathematical Reasoning with Large Language Models", "abs": "Large Language Models (LLMs) have made significant strides in mathematical reasoning, underscoring the need for a comprehensive and fair evaluation of their capabilities. However, existing benchmarks often fall short, either lacking extensive coverage of undergraduate-level mathematical problems or probably suffering from test-set contamination. To address these issues, we introduce UGMathBench, a diverse and dynamic benchmark specifically designed for evaluating undergraduate-level mathematical reasoning with LLMs. UGMathBench comprises 5,062 problems across 16 subjects and 111 topics, featuring 10 distinct answer types. Each problem includes three randomized versions, with additional versions planned for release as the leading open-source LLMs become saturated in UGMathBench. Furthermore, we propose two key metrics: effective accuracy (EAcc), which measures the percentage of correctly solved problems across all three versions, and reasoning gap ($\\Delta$), which assesses reasoning robustness by calculating the difference between the average accuracy across all versions and EAcc. Our extensive evaluation of 23 leading LLMs reveals that the highest EAcc achieved is 56.3% by OpenAI-o1-mini, with large $\\Delta$ values observed across different models. This highlights the need for future research aimed at developing \"large reasoning models\" with high EAcc and $\\Delta = 0$. We anticipate that the release of UGMathBench, along with its detailed evaluation codes, will serve as a valuable resource to advance the development of LLMs in solving mathematical problems.", "keywords": [ "Math Reasoning", "Undergraduate-Level Math problems", "Benchmark" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 4, 4 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "fopjVghcE2", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Andre Cornman", "gender": "unknown", "institution": "Tatta Bio", "country": "US", "position": "Principal Researcher" }, { "name": "Jacob West-Roberts", "gender": "Male", "institution": "Tatta Bio", "country": "US", "position": "Researcher" }, { "name": "Joshua Kravitz", "gender": "unknown", "institution": "Tatta Bio", "country": "US", "position": "Researcher" }, { "name": "Nishant Jha", "gender": "Male", "institution": "Nitro Bio", "country": "US", "position": "Founder" }, { "name": "Yunha Hwang", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 44, "n_ref": 60, "n_ref_all": 73, "n_fig": 5, "n_tab": 2, "L_tab": 4285, "n_element_tab": 537, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1575, "n_element_tab_1": 19, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 164685, "len_all_1": 48555, "len_abs": 1358, "len_title": 133, "len_sents": 38143, "len_sents_1": 23160, "n_sents": 300, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1377, "title": "Diverse Genomic Embedding Benchmark for Functional Evaluation Across the Tree of Life", "abs": "Biological foundation models hold significant promise for deciphering complex biological functions. However, evaluating their performance on functional tasks remains challenging due to the lack of standardized benchmarks encompassing diverse sequences and functions. Existing functional annotations are often scarce, biased, and susceptible to train-test leakage, hindering robust evaluation. Furthermore, biological functions manifest at multiple scales, from individual residues to large genomic segments. To address these limitations, we introduce the Diverse Genomic Embedding Benchmark (DGEB), inspired by natural language embedding benchmarks. DGEB comprises six embedding tasks across 18 expert curated datasets, spanning sequences from all domains of life and encompassing both nucleic acid and amino acid modalities. Notably, four datasets enable direct comparison between models trained on different modalities. Benchmarking protein and genomic language models (pLMs and gLMs) on DGEB reveals performance saturation with model scaling on numerous tasks, especially on those with underrepresented sequences (e.g. Archaea). This highlights the limitations of existing modeling objectives and training data distributions for capturing diverse biological functions. DGEB is available as an open-source package with a public leaderboard at \n\\url{URL hidden for anonymity}.", "keywords": [ "benchmark", "genomics", "proteins" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "foQ4AeEGG7", "primary_area": "causal reasoning", "n_author": 7, "author_info_list": [ { "name": "Anpeng Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Fei Wu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Haiyi Qiu", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Ruoxuan Xiong", "gender": "unknown", "institution": "Emory University", "country": "US", "position": "Assistant Professor" }, { "name": "Zijian Li", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Postdoc" }, { "name": "zhengming Chen", "gender": "Male", "institution": "Guangdong University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 44, "n_formula_1": 11, "n_ref_uni": 64, "n_ref": 244, "n_ref_all": 345, "n_fig": 7, "n_tab": 15, "L_tab": 6113, "n_element_tab": 398, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1923, "n_element_tab_1": 205, "formula_len_all": 3704, "formula_len_all_1": 1091, "len_all": 257302, "len_all_1": 71372, "len_abs": 1430, "len_title": 132, "len_sents": 88069, "len_sents_1": 32100, "n_sents": 580, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1320, "title": "Causal Graph Transformer for Treatment Effect Estimation Under Unknown Interference", "abs": "Networked interference, also known as the peer effect in social science and spillover effect in economics, has drawn increasing interest across various domains. This phenomenon arises when a unit\u2019s treatment and outcome are influenced by the actions of its peers, posing significant challenges to causal inference, particularly in treatment assignment and effect estimation in real applications, due to the violation of the SUTVA assumption. While extensive graph models have been developed to identify treatment effects, these models often rely on structural assumptions about networked interference, assuming it to be identical to the social network, which can lead to misspecification issues in real applications. To address these challenges, we propose an Interference-Agnostic Causal Graph Transformer (CauGramer), which aggregates peers information via $L$-order Graph Transformer and employs cross-attention to infer aggregation function for learning interference representations. By integrating confounder balancing and minimax moment constraints, CauGramer fully incorporates peer information, enabling robust treatment effect estimation. Extensive experiments on two widely-used benchmarks demonstrate the effectiveness and superiority of CauGramer. The code is available at https://github.com/anpwu/CauGramer.", "keywords": [ "Causal Graph Transformer", "Networked Interference", "Unknown Interference Graph", "Peer Effects", "Treatment Effects Estimation" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "foKwWau15m", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Chunyi Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Donghui Feng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Guangtao Zhai", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Guo Lu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Haoning Wu", "gender": "Male", "institution": "Rhymes AI", "country": "SG", "position": "Researcher" }, { "name": "Weisi Lin", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Xiaohong Liu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiele Wu", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xiongkuo Min", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "zicheng zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 130, "n_ref_all": 172, "n_fig": 18, "n_tab": 10, "L_tab": 4962, "n_element_tab": 425, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 3729, "n_element_tab_1": 244, "formula_len_all": 176, "formula_len_all_1": 38, "len_all": 191586, "len_all_1": 59535, "len_abs": 345, "len_title": 111, "len_sents": 47901, "len_sents_1": 24329, "n_sents": 343, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1165, "title": "CMC-Bench: Towards a New Paradigm of Visual Signal Compression", "abs": "Ultra-low bitrate image compression is a challenging and demanding topic. With the development of Large Multimodal Models (LMMs), a Cross Modality Compression (CMC) paradigm of Image-Text-Image has emerged. Compared with traditional codecs, this semantic-level compression can **reduce image data size to 0.1% or even lower**, which has strong potential applications. However, CMC has certain defects in consistency with the original image and perceptual quality. To address this problem, we introduce CMC-Bench, a benchmark of the **cooperative performance of Image-to-Text (I2T) and Text-to-Image (T2I) models for image compression**. This benchmark covers 18,000 and 40,000 images respectively to verify 6 mainstream I2T and 12 T2I models, including 160,000 subjective preference scores annotated by human experts. At ultra-low bitrates, this paper proves that the combination of some I2T and T2I models has surpassed the most advanced visual signal codecs; meanwhile, it highlights where LMMs can be further optimized toward the compression task. We encourage LMM developers to participate in this test to promote the evolution of visual signal codec protocols.", "keywords": [ "Image Compression", "Large Multimodal Model", "Image Quality Assessment" ], "rating_list": [ 8, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "fo5IUCMoFg", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Cansu Sancaktar", "gender": "Female", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Georg Martius", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "Full Professor" }, { "name": "Ji Shi", "gender": "Male", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Jiaqi Chen", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Jonas Frey", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 41, "n_ref": 91, "n_ref_all": 134, "n_fig": 25, "n_tab": 4, "L_tab": 1070, "n_element_tab": 95, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 617, "n_element_tab_1": 27, "formula_len_all": 421, "formula_len_all_1": 344, "len_all": 163160, "len_all_1": 59754, "len_abs": 1587, "len_title": 133, "len_sents": 54813, "len_sents_1": 29021, "n_sents": 397, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1595, "title": "Offline vs. Online Learning in Model-based RL: Lessons for Data Collection Strategies", "abs": "Data collection is crucial for learning robust world models in model-based reinforcement learning.\nThe most prevalent strategies are to actively collect trajectories by interacting with the environment during online training or training on offline datasets.\nAt first glance, the nature of learning task-agnostic environment dynamics makes world models a good candidate for effective offline training. However, the effects of online vs. offline data on world models and thus on the resulting task performance have not been thoroughly studied in the literature. In this work, we investigate both paradigms in model-based settings, conducting experiments on 31 different environments.\nFirst, we showcase that online agents outperform their offline counterparts.\nWe identify a key challenge behind performance degradation of offline agents: encountering Out-of-Distribution states at test time.\nThis issue arises because, without the self-correction mechanism in online agents, offline datasets with limited state space coverage induce a mismatch between the agent's imagination and real rollouts, compromising policy training.\nWe demonstrate that this issue can be mitigated by allowing for additional online interactions in a fixed or adaptive schedule, restoring the performance of online training with limited interaction data.\nWe also showcase that incorporating exploration data helps mitigate the performance degradation of offline agents. Based on our insights, we recommend adding exploration data when collecting large datasets, as current efforts predominantly focus on expert data alone.", "keywords": [ "Model-based reinforcement learning", "Offline learning", "Online learning", "Active learning", "Exploration" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "fnnDtyMxcX", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Francesco Tonin", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Postdoc" }, { "name": "Johan Suykens", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "Full Professor" }, { "name": "Sonny Achten", "gender": "Male", "institution": "KU Leuven", "country": "", "position": "PhD student" }, { "name": "Volkan Cevher", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 22, "n_ref_uni": 43, "n_ref": 60, "n_ref_all": 82, "n_fig": 4, "n_tab": 10, "L_tab": 2261, "n_element_tab": 295, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 720, "n_element_tab_1": 101, "formula_len_all": 2464, "formula_len_all_1": 2269, "len_all": 118427, "len_all_1": 62282, "len_abs": 889, "len_title": 130, "len_sents": 33890, "len_sents_1": 28468, "n_sents": 238, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 914, "title": "HeNCler: Node Clustering in Heterophilous Graphs via Learned Asymmetric Similarity", "abs": "Clustering nodes in heterophilous graphs is challenging as traditional methods assume that effective clustering is characterized by high intra-cluster and low inter-cluster connectivity. To address this, we introduce HeNCler\u2014a novel approach for **He**terophilous **N**ode **Cl**ust**er**ing. \nHeNCler *learns* a similarity graph by optimizing a clustering-specific objective based on weighted kernel singular value decomposition.\nOur approach enables spectral clustering on an *asymmetric* similarity graph, providing flexibility for both directed and undirected graphs. By solving the primal problem directly, our method overcomes the computational difficulties of traditional adjacency partitioning-based approaches. Experimental results show that HeNCler significantly improves node clustering performance in heterophilous graph settings, highlighting the advantage of its asymmetric graph-learning framework.", "keywords": [ "Heterophily", "Node Clustering", "Singular Value Decomposition" ], "rating_list": [ 1, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "fn36V5qsCw", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Gokul Swamy", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Sanjiban Choudhury", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Assistant Professor" }, { "name": "Wen Sun", "gender": "unknown", "institution": "Cornell University", "country": "", "position": "Assistant Professor" }, { "name": "Nicolas Espinosa Dice", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" } ], "n_formula": 114, "n_formula_1": 18, "n_ref_uni": 39, "n_ref": 119, "n_ref_all": 134, "n_fig": 5, "n_tab": 6, "L_tab": 485, "n_element_tab": 93, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 325, "n_element_tab_1": 75, "formula_len_all": 9108, "formula_len_all_1": 1236, "len_all": 202689, "len_all_1": 68083, "len_abs": 1273, "len_title": 90, "len_sents": 62658, "len_sents_1": 29908, "n_sents": 593, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 1294, "title": "Efficient Imitation under Misspecification", "abs": "Interactive imitation learning (IL) is a powerful paradigm for learning to make sequences of decisions from an expert demonstrating how to perform a task. Prior work in efficient imitation learning has focused on the realizable setting, where the expert's policy lies within the learner's policy class (i.e. the learner can perfectly imitate the expert in all states). However, in practice, perfect imitation of the expert perfectly is often impossible due to differences in state information and action space expressiveness (e.g. morphological differences between humans and humanoid robots.) In this paper, we consider the more general *misspecified* setting, where no assumptions are made about the expert policy's realizability. We introduce a novel structural condition, *reward-agnostic policy completeness*, and prove that it is sufficient for interactive IL algorithms to efficiently avoid the quadratically compounding errors that stymie offline approaches like behavioral cloning. We address an additional practical constraint---the case of limited expert data---and propose a principled method for using sub-optimal data to further improve the sample-efficiency of interactive IL algorithms. Finally, we corroborate our theory with experiments on a suite of continuous control tasks.", "keywords": [ "Inverse Reinforcement Learning", "Imitation Learning", "Distribution Shift", "Policy Completeness" ], "rating_list": [ 8, 6, 5, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "fn0mjkZopf", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Luca Cocchi", "gender": "unknown", "institution": "University of Queensland", "country": "", "position": "Principal Researcher" }, { "name": "Luke J. Hearne", "gender": "unknown", "institution": "QIMR Berghofer", "country": "AU", "position": "Postdoc" }, { "name": "Murray Campbell", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Research Scientist" }, { "name": "Parikshit Ram", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Takuya Ito", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Postdoc" }, { "name": "Tim Klinger", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Research Staff Member" } ], "n_formula": 9, "n_formula_1": 0, "n_ref_uni": 39, "n_ref": 90, "n_ref_all": 160, "n_fig": 28, "n_tab": 4, "L_tab": 1321, "n_element_tab": 202, "n_fig_1": 28, "n_tab_1": 2, "L_tab_1": 592, "n_element_tab_1": 50, "formula_len_all": 619, "formula_len_all_1": 0, "len_all": 179309, "len_all_1": 71452, "len_abs": 1799, "len_title": 120, "len_sents": 60214, "len_sents_1": 36367, "n_sents": 435, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1808, "title": "Learning positional encodings in transformers depends on initialization", "abs": "The attention mechanism is central to the transformer's ability to capture complex dependencies between tokens of an input sequence.\nKey to the successful application of the attention mechanism in transformers is its choice of positional encoding (PE).\nThe PE provides essential information that distinguishes the position and order amongst tokens in a sequence.\nMost prior investigations of PE effects on generalization were tailored to 1D input sequences, such as those presented in natural language, where adjacent tokens (e.g., words) are highly related.\nIn contrast, many real world tasks involve datasets with highly non-trivial positional arrangements, such as datasets organized in multiple spatial dimensions, or datasets for which ground truth positions are not known, such as in biological data.\nHere we study the importance of learning accurate PE for problems which rely on a non-trivial arrangement of input tokens. \nCritically, we find that the choice of initialization of a learnable PE greatly influences its ability to learn accurate PEs that lead to enhanced generalization.\nWe empirically demonstrate our findings in a 2D relational reasoning task and a real world 3D neuroscience dataset, applying interpretability analyses to verify the learning of accurate PEs.\nOverall, we find that a learned PE initialized from a small-norm distribution can 1) uncover interpretable PEs that mirror ground truth positions, 2) learn non-trivial and modular PEs in a real-world neuroscience dataset, and 3) lead to improved downstream generalization in both datasets.\nImportantly, choosing an ill-suited PE can be detrimental to both model interpretability and generalization.\nTogether, our results illustrate the feasibility of learning identifiable and interpretable PEs for enhanced generalization.", "keywords": [ "transformers", "representation learning", "deep learning", "reasoning", "transformer" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "fmWVPbRGC4", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Jay R Gopal", "gender": "Male", "institution": "Warren Alpert Medical School of Brown University", "country": "", "position": "Medical Student" }, { "name": "Julien Colin", "gender": "Male", "institution": "Universidad de Alicante", "country": "", "position": "PhD student" }, { "name": "Lore Goetschalckx", "gender": "Female", "institution": "Brown University", "country": "", "position": "Postdoc" }, { "name": "Nuria M Oliver", "gender": "Female", "institution": "ELLIS unit Alicante Foundation ", "country": "ES", "position": "Principal Researcher" }, { "name": "Thomas FEL", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Research Fellow" }, { "name": "Thomas Serre", "gender": "Male", "institution": "Brown University", "country": "", "position": "Full Professor" }, { "name": "Victor Boutin", "gender": "Male", "institution": "CNRS", "country": "FR", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 57, "n_ref": 136, "n_ref_all": 159, "n_fig": 11, "n_tab": 1, "L_tab": 36, "n_element_tab": 2, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 130, "formula_len_all_1": 64, "len_all": 184101, "len_all_1": 67591, "len_abs": 1552, "len_title": 83, "len_sents": 53356, "len_sents_1": 34751, "n_sents": 371, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1560, "title": "Local vs distributed representations: What is the right basis for interpretability?", "abs": "Much of the research on the interpretability of deep neural networks has focused on studying the visual features that maximally activate individual neurons. However, recent work has cast doubts on the usefulness of such local representations for understanding the behavior of deep neural networks because individual neurons tend to respond to multiple unrelated visual patterns, a phenomenon referred to as \u201csuperposition\u201d. A promising alternative to disentangle these complex patterns is learning sparsely distributed vector representations from entire network layers, as the resulting basis vectors seemingly encode single identifiable visual patterns consistently. Thus, one would expect the resulting code to align better with human-perceivable visual patterns, but supporting evidence remains, at best, anecdotal. To fill this gap, we conducted three large-scale psychophysics experiments collected from a pool of 560 participants. Our findings provide (i) strong evidence that features obtained from sparse distributed representations are easier to interpret by human observers and (ii) that this effect is more pronounced in the deepest layers of a neural network. Complementary analyses also reveal that (iii) features derived from sparse distributed representations contribute more to the model\u00b4s decision. \n\nOverall, our results highlight that distributed representations constitute a superior basis for interpretability, underscoring a need for the field to move beyond the interpretation of local neural codes in favor of sparsely distributed ones.", "keywords": [ "Interpretability", "human-centric evaluation", "human experiment", "interpretable features", "local vs distributed" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "fmTY6QQHnQ", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Gavin Kerrigan", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Kai Nelson", "gender": "Male", "institution": "Donald Bren School of Information and Computer Sciences, University of California, Irvine", "country": "US", "position": "Undergrad student" }, { "name": "Padhraic Smyth", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Full Professor" } ], "n_formula": 23, "n_formula_1": 15, "n_ref_uni": 40, "n_ref": 124, "n_ref_all": 154, "n_fig": 4, "n_tab": 10, "L_tab": 4093, "n_element_tab": 438, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 659, "n_element_tab_1": 75, "formula_len_all": 1926, "formula_len_all_1": 1311, "len_all": 175389, "len_all_1": 73758, "len_abs": 1039, "len_title": 68, "len_sents": 53041, "len_sents_1": 32830, "n_sents": 425, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1064, "title": "EventFlow: Forecasting Continuous-Time Event Data with Flow Matching", "abs": "Continuous-time event sequences, in which events occur at irregular intervals, are ubiquitous across a wide range of industrial and scientific domains. The contemporary modeling paradigm is to treat such data as realizations of a temporal point process, and in machine learning it is common to model temporal point processes in an autoregressive fashion using a neural network. While autoregressive models are successful in predicting the time of a single subsequent event, their performance can be unsatisfactory in forecasting longer horizons due to cascading errors. We propose $\\texttt{EventFlow}$, a non-autoregressive generative model for temporal point processes. Our model builds on the flow matching framework in order to directly learn joint distributions over event times, side-stepping the autoregressive process. $\\texttt{EventFlow}$ is likelihood-free, easy to implement and sample from, and either matches or surpasses the performance of state-of-the-art models in both unconditional and conditional generation tasks on a set of standard benchmarks.", "keywords": [ "temporal point processes", "generative models", "event sequences" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 4, 3 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "fmJUYgmMbL", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Louis Grenioux", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "", "position": "PhD student" }, { "name": "Marylou Gabri\u00e9", "gender": "Female", "institution": "Ecole Normale Sup\u00e9rieure de Paris", "country": "FR", "position": "Assistant Professor" }, { "name": "Maxence Noble", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "FR", "position": "PhD student" }, { "name": "Alain Durmus", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "", "position": "Full Professor" } ], "n_formula": 205, "n_formula_1": 19, "n_ref_uni": 65, "n_ref": 165, "n_ref_all": 241, "n_fig": 17, "n_tab": 13, "L_tab": 6909, "n_element_tab": 545, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 24264, "formula_len_all_1": 1571, "len_all": 392667, "len_all_1": 80557, "len_abs": 1389, "len_title": 120, "len_sents": 118267, "len_sents_1": 36357, "n_sents": 1008, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1398, "title": "Learned Reference-based Diffusion Sampler for multi-modal distributions", "abs": "Over the past few years, several approaches utilizing score-based diffusion have been proposed to sample from probability distributions, that is without having access to exact samples and relying solely on evaluations of unnormalized densities. The resulting samplers approximate the time-reversal of a noising diffusion process, bridging the target distribution to an easy-to-sample base distribution. In practice, the performance of these methods heavily depends on key hyperparameters that require ground truth samples to be accurately tuned. Our work aims to highlight and address this fundamental issue, focusing in particular on multi-modal distributions, which pose significant challenges for existing sampling methods. Building on existing approaches, we introduce *Learned Reference-based Diffusion Sampler* (LRDS), a methodology specifically designed to leverage prior knowledge on the location of the target modes in order to bypass the obstacle of hyperparameter tuning. LRDS proceeds in two steps by (i) learning a *reference* diffusion model on samples located in high-density space regions and tailored for multimodality, and (ii) using this reference model to foster the training of a diffusion-based sampler. We experimentally demonstrate that LRDS best exploits prior knowledge on the target distribution compared to competing algorithms on a variety of challenging distributions.", "keywords": [ "sampling from multi-modal densities", "diffusion models", "variational inference", "stochastic optimal control", "multi-level energy-based models" ], "rating_list": [ 8, 8, 3, 6 ], "soundness_list": [ 4, 3, 2, 4 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 2, 3, 2 ] }, { "paper_id": "fmHS8aBfuH", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Alon Zolfi", "gender": "Male", "institution": "Ben Gurion University of the Negev", "country": "IL", "position": "PhD student" }, { "name": "Asaf Shabtai", "gender": "Male", "institution": "Ben-Gurion University of the Negev", "country": "IL", "position": "Full Professor" }, { "name": "Ben Ganon", "gender": "Male", "institution": "Ben Gurion University of the Negev", "country": "IL", "position": "MS student" }, { "name": "Hisashi Kojima", "gender": "Male", "institution": "Fujitsu Research and Development Center Co. Ltm.", "country": "JP", "position": "Researcher" }, { "name": "Inderjeet Singh", "gender": "Male", "institution": "Fujitsu Research of Europe Limited", "country": "GB", "position": "Researcher" }, { "name": "Omer Hofman", "gender": "Male", "institution": "Fujitsu Research and Development Center Co. Ltm.", "country": "IL", "position": "Researcher" }, { "name": "Yuval Elovici", "gender": "Male", "institution": "Ben Gurion University of the Negev, Technion", "country": "IL", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 31, "n_ref": 49, "n_ref_all": 65, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 417, "n_element_tab_1": 19, "formula_len_all": 999, "formula_len_all_1": 999, "len_all": 106645, "len_all_1": 55408, "len_abs": 1579, "len_title": 125, "len_sents": 31836, "len_sents_1": 26011, "n_sents": 201, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1584, "title": "DIESEL - Dynamic Inference-Guidance via Evasion of Semantic Embeddings in LLMs", "abs": "In recent years, conversational large language models (LLMs) have shown tremendous success in tasks such as casual conversation, question answering, and personalized dialogue, making significant advancements in domains like virtual assistance, social interaction, and online customer engagement. However, they often generate responses that are not aligned with human values (e.g., ethical standards, safety, or social norms), leading to potentially unsafe or inappropriate outputs. While several techniques have been proposed to address this problem, they come with a cost, requiring computationally expensive training or dramatically increasing the inference time. In this paper, we present DIESEL, a lightweight inference guidance technique that can be seamlessly integrated into any autoregressive LLM to semantically filter undesired concepts from the response. DIESEL can function either as a standalone safeguard or as an additional layer of defense, enhancing response safety by reranking the LLM's proposed tokens based on their similarity to predefined negative concepts in the latent space. This approach provides an efficient and effective solution for maintaining alignment with human values. Our evaluation demonstrates DIESEL's effectiveness on state-of-the-art conversational models (e.g., Llama 3), even in challenging jailbreaking scenarios that test the limits of response safety. We further show that DIESEL can be generalized to use cases other than safety, providing a versatile solution for general-purpose response filtering with minimal computational overhead.", "keywords": [ "Large Language Models", "Inference Guidance", "Safety", "Robustness" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "flgrH5nK4H", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Annelies Lejon", "gender": "Female", "institution": "Universiteit Gent", "country": "BE", "position": "Postdoc" }, { "name": "Joni Dambre", "gender": "unknown", "institution": "Universiteit Gent", "country": "", "position": "Full Professor" }, { "name": "Mathieu De Coster", "gender": "Male", "institution": "Universiteit Gent", "country": "", "position": "PhD student" }, { "name": "Toon Vandendriessche", "gender": "unknown", "institution": "Universiteit Gent", "country": "BE", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 25, "n_ref": 43, "n_ref_all": 59, "n_fig": 8, "n_tab": 3, "L_tab": 844, "n_element_tab": 92, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 808, "n_element_tab_1": 92, "formula_len_all": 0, "formula_len_all_1": 41, "len_all": 100818, "len_all_1": 64543, "len_abs": 853, "len_title": 143, "len_sents": 34489, "len_sents_1": 30492, "n_sents": 279, "n_sents_1": 253, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 857, "title": "Representing Signs as Signs: One-Shot ISLR to Facilitate Functional Sign Language Technologies", "abs": "Isolated Sign Language Recognition (ISLR) is crucial for scalable sign language technology, yet language-specific approaches limit current models. To address this, we propose a one-shot learning approach that generalises across languages and evolving vocabularies. Our method involves pretraining a model to embed signs based on essential features and using a dense vector search for rapid, accurate recognition of unseen signs. We achieve state-of-the-art results, including 50.8% one-shot MRR on a large dictionary containing 10,235 unique signs from a different language than the training set. Our approach is robust across languages and support sets, offering a scalable, adaptable solution for ISLR. Co-created with the Deaf and Hard of Hearing (DHH) community, this method aligns with real-world needs, and advances scalable sign language recognition.", "keywords": [ "Sign Language", "Deep Learning", "Computer Vision" ], "rating_list": [ 3, 6, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 2, 4, 1, 1, 3 ], "contribution_list": [ 2, 3, 1, 2, 1 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "fkNsgI1nye", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Eunho Yang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Geondo Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Jaeyun Song", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Uigyu Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 42, "n_ref": 76, "n_ref_all": 93, "n_fig": 8, "n_tab": 4, "L_tab": 611, "n_element_tab": 55, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 215, "n_element_tab_1": 20, "formula_len_all": 1072, "formula_len_all_1": 1268, "len_all": 118581, "len_all_1": 63341, "len_abs": 1242, "len_title": 123, "len_sents": 34284, "len_sents_1": 29853, "n_sents": 261, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1255, "title": "Secure Diffusion Model Unlocked: Efficient Inference via Score Distillation", "abs": "As services based on diffusion models expand across various domains, preserving the privacy of client data becomes more critical. Fully homomorphic encryption and secure multi-party computation have been employed for privacy-preserving inference, but these methods are computationally expensive and primarily work for linear computations, making them challenging to apply to large diffusion models. While homomorphic encryption has been recently applied to diffusion models, it falls short of fully safeguarding privacy, as inputs used in the $\\epsilon$ prediction are not encrypted. In this paper, we propose a novel framework for private inference for both inputs and outputs. To ensure robust approximations, we introduce several techniques for handling non-linear operations. Additionally, to reduce latency, we curtail the number of denoising steps while minimizing performance degradation of conditional generation through score distillation from the unconditional generation of the original model with full denoising steps. Experimental results show that our model produces high-quality images comparable to the original, and the proposed score distillation significantly enhances performance, compensating for fewer steps and approximation errors.", "keywords": [ "private inference", "diffusion model" ], "rating_list": [ 3, 5, 1 ], "soundness_list": [ 3, 2, 1 ], "presentation_list": [ 3, 2, 1 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "fkKW1PK8Ga", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Jia-Ching Wang", "gender": "Male", "institution": "National Central University", "country": "TW", "position": "Full Professor" }, { "name": "Minghsiang Su", "gender": "Male", "institution": "Soochow University", "country": "TW", "position": "Assistant Professor" }, { "name": "Rueicyuan Su", "gender": "Male", "institution": "National Central University", "country": "TW", "position": "MS student" }, { "name": "Tzuen Su", "gender": "Male", "institution": "Soochow University", "country": "TW", "position": "Undergrad student" } ], "n_formula": 31, "n_formula_1": 21, "n_ref_uni": 55, "n_ref": 86, "n_ref_all": 96, "n_fig": 1, "n_tab": 5, "L_tab": 1736, "n_element_tab": 190, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1196, "n_element_tab_1": 148, "formula_len_all": 3502, "formula_len_all_1": 2314, "len_all": 157560, "len_all_1": 76255, "len_abs": 1087, "len_title": 79, "len_sents": 42779, "len_sents_1": 34246, "n_sents": 350, "n_sents_1": 250, "venue": "ICLR2025", "decision": "withdraw", "L_title": 31, "L_abs": 1090, "title": "Generalization Gradient Descent", "abs": "We propose a new framework for evaluating the relationship between features\nand generalization via a theoretical analysis of the out-of-distribution (OOD)\ngeneralization problem, in which we simultaneously use two mathematical methods:\na generalization ratio that quantitatively characterizes the degree of generalization,\nand a generalization decision process (GDP) that formalizes the relationship of loss\nbetween seen and unseen domains. By combining the concepts of informativeness\nand variation in the generalization ratio, we intuitively associate them with OOD\nproblems to derive the generalization inequality. We then introduce it to the GDP to\nselect the best loss from seen domains to gradient descent for backpropagation. In\nthe case where the classifier is defined by fully connected neural network, the entire\nsystem is trained with backpropagation. There is no need for any model selection\ncriterion or operating on gradients during training. Experiments demonstrate the\npotential of the framework through qualitative and quantitative evaluation of the\ngeneralization ability.", "keywords": [ "out-of-distribution", "generalization problem", "gradient descent" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 2, 2, 2, 2 ] }, { "paper_id": "fk5ePN7YCS", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Enzo Tartaglione", "gender": "Male", "institution": "T\u00e9l\u00e9com Paris", "country": "FR", "position": "Associate Professor" }, { "name": "Van-Tam Nguyen", "gender": "Male", "institution": "T\u00e9l\u00e9com Paris", "country": "FR", "position": "Full Professor" }, { "name": "Victor Qu\u00e9tu", "gender": "Male", "institution": "T\u00e9l\u00e9com ParisTech", "country": "FR", "position": "PhD student" }, { "name": "Zhu LIAO", "gender": "Male", "institution": "T\u00e9l\u00e9com ParisTech", "country": "FR", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 32, "n_ref": 44, "n_ref_all": 68, "n_fig": 7, "n_tab": 8, "L_tab": 4030, "n_element_tab": 302, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 3587, "n_element_tab_1": 282, "formula_len_all": 511, "formula_len_all_1": 474, "len_all": 112213, "len_all_1": 63565, "len_abs": 1188, "len_title": 116, "len_sents": 30846, "len_sents_1": 24584, "n_sents": 227, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1222, "title": "NEPENTHE: Entropy-Based Pruning as a Neural Network Depth's Reducer", "abs": "While deep neural networks are highly effective at solving complex tasks, their computational demands can hinder their usefulness in real-time applications and with limited-resources systems. Besides, it is a known fact that, for many downstream tasks, off-the-shelf models are over-parametrized. While classical structured pruning can reduce the network's width, the computation's critical path, namely the maximum number of layers encountered at forward propagation, apparently can not be reduced.\n\nIn this paper, we aim to reduce the depth of over-parametrized deep neural networks: we propose an e**N**tropy-bas**E**d **P**runing as a n**E**ural **N**etwork dep**TH**'s r**E**ducer (NEPENTHE) to alleviate deep neural networks' computational burden.\nBased on our theoretical finding, NEPENTHE leverages \"unstructured'' pruning to bias sparsity enhancement in layers with low entropy to remove them entirely. We validate our approach on popular architectures such as MobileNet, Swin-T and RoBERTa, showing that, when in the overparametrization regime, some layers are linearizable (hence reducing the model's depth) with little to no performance loss. The code will be publicly available upon acceptance of the article.", "keywords": [ "Pruning", "Compression", "Entropy", "Deep Learning" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "fk4czNKXPC", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 8, "author_info_list": [ { "name": "Alex Vitvitskyi", "gender": "unknown", "institution": "DeepMind", "country": "GB", "position": "Research Engineer" }, { "name": "Andrew Joseph Dudzik", "gender": "Male", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Borja Ibarz", "gender": "Male", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Jessica B Hamrick", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Larisa Markeeva", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Research Engineer" }, { "name": "Petar Veli\u010dkovi\u0107", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Affiliated Lecturer" }, { "name": "Razvan Pascanu", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Affiliate Member" }, { "name": "Wilfried Bounsi", "gender": "Male", "institution": "Google", "country": "", "position": "eng" } ], "n_formula": 5, "n_formula_1": 8, "n_ref_uni": 35, "n_ref": 73, "n_ref_all": 83, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 279, "formula_len_all_1": 495, "len_all": 123037, "len_all_1": 55252, "len_abs": 1293, "len_title": 94, "len_sents": 33925, "len_sents_1": 27480, "n_sents": 215, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1298, "title": "Transformers meet Neural Algorithmic Reasoners", "abs": "Transformers have revolutionized machine learning with their simple yet effective architecture. Pre-training Transformers on massive text datasets from the Internet has led to unmatched generalization for natural language understanding (NLU) tasks. However, such language models remain fragile when tasked with algorithmic forms of reasoning, where computations must be precise and robust. To address this limitation, we propose a novel approach that combines the Transformer's language understanding with the robustness of graph neural network (GNN)-based neural algorithmic reasoners (NARs). Such NARs proved effective as generic solvers for algorithmic tasks, when specified in graph form. To make their embeddings accessible to a Transformer, we propose a hybrid architecture with a two-phase training procedure, allowing the tokens in the language model to cross-attend to the node embeddings from the NAR. We evaluate our resulting TransNAR model on CLRS-Text, the text-based version of the CLRS-30 benchmark, and demonstrate significant gains over Transformer-only models for algorithmic reasoning, both in and out of distribution. Finally, we empirically show that Transformer-only models distilled from TransNAR models also exhibit improved out-of-distribution generalization capabilities.", "keywords": [ "Graph Neural Network", "Transformer", "Neural Algorithmic Reasoning", "Length Generalization" ], "rating_list": [ 6, 3, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "fk4QS3j1sU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Mingyue Cheng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Researcher" }, { "name": "Qi Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Qingyang Mao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Rui Li", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Zheng Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zhi Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 31, "n_ref": 59, "n_ref_all": 79, "n_fig": 8, "n_tab": 11, "L_tab": 4322, "n_element_tab": 283, "n_fig_1": 6, "n_tab_1": 8, "L_tab_1": 2950, "n_element_tab_1": 315, "formula_len_all": 185, "formula_len_all_1": 120, "len_all": 140984, "len_all_1": 66064, "len_abs": 1970, "len_title": 128, "len_sents": 38596, "len_sents_1": 29717, "n_sents": 280, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1626, "title": "PoTable: Programming Standardly on Table-based Reasoning Like a Human Analyst", "abs": "Table-based reasoning has garnered substantial research interest, particularly in its integration with Large Language Model (LLM) which has revolutionized the general reasoning paradigm. Numerous LLM-based studies introduce symbolic tools (e.g., databases, Python) as assistants in complex information understanding and arithmetic computations. However, they emphasize extensive and flexible utilization of symbolic tools, without fully considering the intrinsic logic of the reasoning process. In this study, we propose PoTable as a simple yet effective table-based reasoning method. Specifically, PoTable features a planning phase and an executing phase, implemented with an LLM-based operation planner and code generator and a Python interpreter as the real-time executor. To incorporate logical top-level guidance, we split the entire reasoning process into several distinct analysis stages with macroscopic instruction injection. As the reasoning process is structured suitably under the top-level guidance with precise and specific goals, PoTable produces superior reasoning results with highly accurate, steply commented and completely executable code. To summarize, PoTable enjoys the advantages of accuracy and explainability that make it a distinguished tabular data analyst. Extensive experiments over three evaluation datasets from two public benchmarks on two backbones demonstrate the outstanding performance of PoTable. In particular, GPT-based PoTable achieves over 4% higher absolute accuracy than runner-ups on all evaluation datasets. Our code is available at https://anonymous.4open.science/r/PoTable-6788.", "keywords": [ "Table-based Reasoning", "Large Language Model", "Symbolic Tools", "Real-time Program Execution", "Human Cognitive Behavior" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "fjPOt8QlqQ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Hu Ding", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Xianglu Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 13, "n_ref_uni": 75, "n_ref": 146, "n_ref_all": 181, "n_fig": 4, "n_tab": 14, "L_tab": 7662, "n_element_tab": 442, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 448, "n_element_tab_1": 17, "formula_len_all": 1478, "formula_len_all_1": 873, "len_all": 204624, "len_all_1": 65929, "len_abs": 1629, "len_title": 90, "len_sents": 47017, "len_sents_1": 31763, "n_sents": 348, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1659, "title": "Exploring The Forgetting in Adversarial Training: A Novel Method for Enhancing Robustness", "abs": "In recent years, there has been an explosion of research into developing robust deep neural networks against adversarial examples. As one of the most successful methods, Adversarial Training (AT) has been widely studied before, but there is still a gap to achieve promising\nclean and robust accuracy for many practical tasks. In this paper, we consider the AT problem from a new perspective which connects it to catastrophic forgetting in continual learning (CL). Catastrophic forgetting is a phenomenon in which neural networks forget old knowledge upon learning a new task. Although AT and CL are two different problems, we show that they actually share several key properties in their training processes. Specifically, we conduct an empirical study and find that this forgetting phenomenon indeed occurs in adversarial robust training across multiple datasets (SVHN, CIFAR-10, CIFAR-100, and TinyImageNet) and perturbation models ($\\ell_{\\infty}$ and $\\ell_{2}$). Based on this observation, we propose a novel method called Adaptive Multi-teachers Self-distillation (AMS), which leverages a carefully designed adaptive regularizer to mitigate the forgetting by aligning model outputs between new and old ``stages''. Moreover, our approach can be used as a unified method to enhance multiple different AT algorithms. Our experiments demonstrate that our method can significantly enhance robust accuracy and meanwhile preserve high clean accuracy, under several popular adversarial attacks (e.g., PGD, CW, and Auto Attacks). As another benefit of our method, we discover that it can largely alleviate the robust overfitting issue of AT in our experiments.", "keywords": [ "Adversarial training", "Continual learning", "Catastrophic forgetting" ], "rating_list": [ 6, 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 3, 5, 4, 4, 4 ] }, { "paper_id": "fjJ5oYtBpO", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Bao Wang", "gender": "Male", "institution": "University of Utah", "country": "US", "position": "Assistant Professor" }, { "name": "Shih-Hsin Wang", "gender": "Male", "institution": "University of Utah", "country": "", "position": "PhD student" }, { "name": "Andrea Bertozzi", "gender": "Female", "institution": "University of California, Los Angeles", "country": "", "position": "Full Professor" }, { "name": "Justin Baker", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" } ], "n_formula": 21, "n_formula_1": 24, "n_ref_uni": 49, "n_ref": 77, "n_ref_all": 103, "n_fig": 8, "n_tab": 5, "L_tab": 1276, "n_element_tab": 96, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 794, "n_element_tab_1": 66, "formula_len_all": 1828, "formula_len_all_1": 1532, "len_all": 152743, "len_all_1": 72585, "len_abs": 1086, "len_title": 121, "len_sents": 46313, "len_sents_1": 32874, "n_sents": 393, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1093, "title": "Learning Molecular Symmetry Breaking via Symmetry-adapted Neural Networks", "abs": "E(3)-equivariant neural networks have achieved remarkable performance in molecular modeling. However, the equivariance constraint limits the model's effectiveness in learning tasks involving symmetry breaking, particularly those that violate the celebrated Curie principle. Relaxing the equivariance constraint is essential for addressing these challenges. In this paper, we explore the intricate symmetry relationships between an object and its spontaneously symmetry-broken outcomes. We introduce a relaxed equivariance based on the molecule's inherent symmetries. Additionally, we develop SANN -- a symmetry-adapted neural network architecture that learns symmetry breaking through equivalence classes of atoms. SANN decomposes the molecular point cloud into sets of symmetry-equivalent atoms and performs message-passing both within and across these classes. We demonstrate the advantages of our method over invariant and equivariant models through synthetic tasks and show that SANN effectively learns both equivariance and symmetry breaking in various benchmark molecular modeling tasks.", "keywords": [ "relaxed equivariance", "equivariance", "symmetry breaking", "geometric deep learning", "molecular modeling" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "fjEZ2LPceZ", "primary_area": "datasets and benchmarks", "n_author": 16, "author_info_list": [ { "name": "Bin Liang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "Undergrad student" }, { "name": "Guanting Dong", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Huangxuan Wu", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Jianing Yu", "gender": "Female", "institution": "School of Artificial Intelligence, Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Qiuna Tan", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Runqi Qiao", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Weihao Zeng", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Intern" }, { "name": "Weiran Xu", "gender": "Male", "institution": "Beijing University of Post and Telecommunication", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaoshuai Song", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Yejie Wang", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Zhexu Wang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Zhuoma GongQue", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "dayuan fu", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "MuxiDiao", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Fu Yujia", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Wang Zengyang", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 47, "n_ref": 100, "n_ref_all": 165, "n_fig": 19, "n_tab": 17, "L_tab": 10472, "n_element_tab": 915, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 1384, "n_element_tab_1": 85, "formula_len_all": 316, "formula_len_all_1": 0, "len_all": 332123, "len_all_1": 58960, "len_abs": 6169, "len_title": 143, "len_sents": 97988, "len_sents_1": 28431, "n_sents": 828, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1532, "title": "CS-Bench: A Comprehensive Benchmark for Large Language Models towards Computer Science Mastery", "abs": "Large language models (LLMs) have demonstrated significant potential in advancing various fields of research and society. However, the current community of large language models (LLMs) overly focuses on benchmarks for analyzing specific foundational skills (e.g. mathematics and code generation), neglecting an all-round evaluation of the computer science field. To bridge this gap, we introduce CS-Bench, the first multilingual (English, Chinese, French, German) benchmark dedicated to evaluating the performance of LLMs in computer science. CS-Bench comprises approximately 10K meticulously curated test samples, covering 26 subfields across 4 key areas of computer science, encompassing various task forms and divisions of knowledge and reasoning. Utilizing CS-Bench, we conduct a comprehensive evaluation of over 30 mainstream LLMs, revealing the relationship between CS performance and model scales. We also quantitatively analyze the reasons for failures in existing LLMs and highlight directions for improvements, including knowledge supplementation and CS-specific reasoning. Further cross-capability experiments show a high correlation between LLMs' capabilities in computer science and their abilities in mathematics and coding. Moreover, expert LLMs specialized in mathematics and coding also demonstrate strong performances in several CS subfields. Looking ahead, we envision CS-Bench serving as a cornerstone for LLM applications in the CS field and paving new avenues in assessing LLMs' diverse reasoning capabilities.", "keywords": [ "large language model", "evaluation", "computer science" ], "rating_list": [ 8, 6, 5, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 1, 4 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "fifXzmzeGy", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 5, "author_info_list": [ { "name": "Daniel Simpson", "gender": "unknown", "institution": "Monash University", "country": "", "position": "Full Professor" }, { "name": "Johnathan Chiu", "gender": "Not Specified", "institution": "Normal Computing", "country": "", "position": "Researcher" }, { "name": "Kaelan Donatella", "gender": "unknown", "institution": "Normal Computing", "country": "US", "position": "Researcher" }, { "name": "Phoebe Klett", "gender": "unknown", "institution": "Normal Computing", "country": "", "position": "Researcher" }, { "name": "Samuel Duffield", "gender": "unknown", "institution": "Normal Computing", "country": "US", "position": "Researcher" } ], "n_formula": 40, "n_formula_1": 16, "n_ref_uni": 86, "n_ref": 199, "n_ref_all": 235, "n_fig": 12, "n_tab": 1, "L_tab": 127, "n_element_tab": 27, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 724, "n_element_tab_1": 8, "formula_len_all": 2431, "formula_len_all_1": 853, "len_all": 209837, "len_all_1": 65635, "len_abs": 876, "len_title": 99, "len_sents": 58623, "len_sents_1": 32322, "n_sents": 388, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 958, "title": "Scalable Bayesian Learning with posteriors", "abs": "Although theoretically compelling, Bayesian learning with modern machine learning models is computationally challenging since it requires approximating a high dimensional posterior distribution. In this work, we (i) introduce **_posteriors_**, an easily extensible PyTorch library hosting general-purpose implementations making Bayesian learning accessible and scalable to large data and parameter regimes; (ii) present a tempered framing of stochastic gradient Markov chain Monte Carlo, as implemented in posteriors, that transitions seamlessly into optimization and unveils a minor modification to deep ensembles to ensure they are asymptotically unbiased for the Bayesian posterior, and (iii) demonstrate and compare the utility of Bayesian approximations through experiments including an investigation into the cold posterior effect and applications with large language models.\n\n_**posteriors**_ repository: https://github.com/normal-computing/posteriors", "keywords": [ "Bayesian deep learning", "PyTorch", "Variational Inference", "MCMC" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "fiTpna7fO5", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Stanislav Fort", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 17, "n_ref": 25, "n_ref_all": 45, "n_fig": 13, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 295, "formula_len_all_1": 295, "len_all": 76716, "len_all_1": 54986, "len_abs": 1137, "len_title": 135, "len_sents": 29103, "len_sents_1": 25902, "n_sents": 200, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1160, "title": "Multi-attacks: A single adversarial perturbation for multiple images and target labels", "abs": "We show that we can easily design a single adversarial perturbation $P$ that changes the class of $n$ images $X_1,X_2,\\dots,X_n$ from their original, unperturbed classes $c_1, c_2,\\dots,c_n$ to desired (not necessarily all the same) classes $c^*_1,c^*_2,\\dots,c^*_n$ for up to hundreds of images and target classes at once. We call these \\textit{multi-attacks}. Characterizing the maximum $n$ we can achieve under different conditions such as image resolution, we estimate the number of regions of high class confidence around a particular image in the space of pixels to be around $10^{\\mathcal{O}(100)}$, posing a significant problem for exhaustive defense strategies. We show several immediate consequences of this: adversarial attacks that change the resulting class based on their intensity, and scale-independent adversarial examples. To demonstrate the redundancy and richness of class decision in the pixel space, we look for its two-dimensional sections that trace images and spell words using particular classes. We also show that ensembling reduces susceptibility to multi-attacks, and that classifiers trained on random labels are more susceptible.", "keywords": [ "adversarial attacks", "geometry", "classification", "robustness", "security" ], "rating_list": [ 3, 3, 1, 6 ], "soundness_list": [ 2, 1, 1, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "fi9LF92Cak", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Antonio Bianchi", "gender": "Male", "institution": "Purdue University", "country": "", "position": "Assistant Professor" }, { "name": "Hao Wen", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Intern" }, { "name": "Hu Xiao", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 9, "n_ref": 22, "n_ref_all": 43, "n_fig": 6, "n_tab": 4, "L_tab": 2122, "n_element_tab": 170, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2902, "n_element_tab_1": 180, "formula_len_all": 286, "formula_len_all_1": 254, "len_all": 118640, "len_all_1": 51415, "len_abs": 3657, "len_title": 121, "len_sents": 28509, "len_sents_1": 22826, "n_sents": 199, "n_sents_1": 143, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1304, "title": "Efficient In-Context Visual Learning with Trident Block and Cross Blocks", "abs": "Visual prompt-based large vision models exhibit remarkable performance in a range of vision tasks. However, visual prompting large vision models are computationally intensive and resource-demanding due to their large parameter sizes and the complexity of processing visual prompts, resulting in inefficiencies in speed and memory usage. To tackle these challenges, we propose the Efficient Painter model, which leverages a novel context-aggregated attention based trident block to alleviate cross-task gaps and reduce memory and computation overhead. Furthermore, we introduce a cross-blocks feature union module to capture global contextual information at different levels and speed up training. This architecture mitigates training costs and memory requirements during inference. Our model strikes a balance between speed and memory efficiency, achieving a 19$\\times$ reduction in FLOPs. Moreover, our model is 9$\\times$ smaller in model size and runs 4.1$\\times$ and 27$\\times$ faster during training and inference, respectively. Comprehensive experiments demonstrate that our design effectively processes additional visual prompts and outperforms baseline methods on standard benchmarks like \\textit{SIDD} and \\textit{LoL} in zero-shot settings, improving performance by 0.4\\% and 1.2\\% respectively.", "keywords": [ "Visual Prompting Large Vision Model", "Efficient In-context Learning", "Vision Transformers" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "fhJeqL1rRg", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Adel Nabli", "gender": "Male", "institution": "Universit\u00e9 Pierre et Marie Curie - Paris 6, Computer Science Lab - Pierre and Marie Curie University, Paris, France", "country": "FR", "position": "PhD student" }, { "name": "Edouard Oyallon", "gender": "Male", "institution": "CNRS", "country": "FR", "position": "Researcher" }, { "name": "Eugene Belilovsky", "gender": "Male", "institution": "Concordia University, Montreal", "country": "CA", "position": "Assistant Professor" }, { "name": "Louis Fournier", "gender": "Male", "institution": "Universit\u00e9 Pierre et Marie Curie - Paris 6, Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "", "position": "PhD student" }, { "name": "Marco Pedersoli", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "Associate Professor" }, { "name": "Masih Aminbeidokhti", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 22, "n_ref": 49, "n_ref_all": 67, "n_fig": 8, "n_tab": 7, "L_tab": 6360, "n_element_tab": 275, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2976, "n_element_tab_1": 221, "formula_len_all": 491, "formula_len_all_1": 481, "len_all": 134734, "len_all_1": 63939, "len_abs": 1251, "len_title": 85, "len_sents": 35261, "len_sents_1": 27853, "n_sents": 293, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1254, "title": "WASH: Train your Ensemble with Communication-Efficient Weight Shuffling, then Average", "abs": "The performance of deep neural networks is enhanced by ensemble methods, which average the output of several models. However, this comes at an increased cost at inference. Weight averaging methods aim to balance the generalization of ensembling and the inference speed of a single model by averaging the parameters of an ensemble of models. Yet, naive averaging results in poor performance as models converge to different loss basins, and aligning the models to improve the performance of the average is challenging. Alternatively, inspired by distributed training, methods like DART and PAPA have been proposed to train several models in parallel such that they will end up in the same basin, resulting in good averaging accuracy. However, these methods either compromise ensembling accuracy or demand significant communication between models during training. In this paper, we introduce WASH, a novel distributed method for training model ensembles for weight averaging that achieves state-of-the-art image classification accuracy. WASH maintains models within the same basin by randomly shuffling a small percentage of weights during training, resulting in diverse models and lower communication costs compared to standard parameter averaging methods.", "keywords": [ "weight averaging", "model averaging", "model merging", "permutation", "communication", "distributed", "parallel", "ensembling" ], "rating_list": [ 3, 3, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 2, 3, 4 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "fh9OwKKb8D", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Junyi Tang", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "MS student" }, { "name": "Li Zhang", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "Postdoc" }, { "name": "Yan Ge", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "Lecturer" } ], "n_formula": 18, "n_formula_1": 20, "n_ref_uni": 27, "n_ref": 53, "n_ref_all": 74, "n_fig": 8, "n_tab": 7, "L_tab": 1598, "n_element_tab": 77, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 840, "n_element_tab_1": 57, "formula_len_all": 829, "formula_len_all_1": 1020, "len_all": 131619, "len_all_1": 57979, "len_abs": 1231, "len_title": 127, "len_sents": 35560, "len_sents_1": 25687, "n_sents": 229, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1237, "title": "Conditionally Adaptive Graph Attention Networks for Credit Card Fraud Detection", "abs": "Fraudulent transactions have been on the rise, leading to significant financial losses annually. In credit card fraud detection (CCFD), various predictive models aim to mitigate these losses by assessing transaction risk. While GNN-based methods have been employed to capture spatio-temporal transaction features, they often suffer from oversmoothing as graph layers increase, causing fraudulent and legitimate transactions to become indistinguishable. Existing semi-supervised methods that mask some labels have not fully resolved this issue. To address this, we propose the Multi-head Attention Conditional Variational Autoencoder (Ma-CVAE), which leverages weight distributions from imbalanced datasets and the Gumbel softmax distribution to construct more diverse reconstructed features, reducing feature homogenization. Then, We utilize Temporal Graph Attention Networks (TGAT) with a Multi-Attention mechanism to model risk propagation among transactions. Finally, classification probabilities are mapped to risk scores via a Multi-Layer Perceptron (MLP). Our approach achieves state-of-the-art performance, improving AUC scores by 1.45%, 3.05%, and 0.83% on three semi-supervised datasets: FFSD, YelpChi, and Amazon, respectively.", "keywords": [ "Semi-supervised Learning", "Graph Neural Networks", "Fraud Detection" ], "rating_list": [ 3, 5, 8, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "fh7GYa7cjO", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Jiacai Liu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Ke Wei", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Wenye Li", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" } ], "n_formula": 124, "n_formula_1": 26, "n_ref_uni": 16, "n_ref": 39, "n_ref_all": 54, "n_fig": 5, "n_tab": 1, "L_tab": 130, "n_element_tab": 12, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 273, "n_element_tab_1": 4, "formula_len_all": 16601, "formula_len_all_1": 1689, "len_all": 174170, "len_all_1": 61110, "len_abs": 4161, "len_title": 124, "len_sents": 46705, "len_sents_1": 24190, "n_sents": 546, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 998, "title": "$\\phi$-Update: A Class of Policy Update Methods with Policy Convergence Guarantee", "abs": "Inspired by the similar update pattern of softmax natural policy gradient and Hadamard policy gradient, we propose to study a general policy update rule called $\\phi$-update, where $\\phi$ refers to a scaling function on advantage functions. Under very mild conditions on $\\phi$, the global asymptotic state value convergence of $\\phi$-update is firstly established. Then we show that the policy produced by $\\phi$-update indeed converges, even when there are multiple optimal policies. This is in stark contrast to existing results where explicit regularizations are required to guarantee the convergence of the policy. Since softmax natural policy gradient is an instance of $\\phi$-update, it provides an affirmative answer to the question whether the policy produced by softmax natural policy gradient converges. The exact asymptotic convergence rate of state values is further established based on the policy convergence. Lastly, we establish the global linear convergence of $\\phi$-update.", "keywords": [ "reinforcement learning", "policy optimization", "policy convergence", "linear convergence" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "fgaNjiijS1", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Liuhua Peng", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Associate Professor" }, { "name": "Xunye Tian", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" }, { "name": "Zhijian Zhou", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" } ], "n_formula": 136, "n_formula_1": 18, "n_ref_uni": 70, "n_ref": 142, "n_ref_all": 170, "n_fig": 11, "n_tab": 5, "L_tab": 1752, "n_element_tab": 245, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 14512, "formula_len_all_1": 1258, "len_all": 239353, "len_all_1": 67765, "len_abs": 1560, "len_title": 87, "len_sents": 63414, "len_sents_1": 30286, "n_sents": 630, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1656, "title": "A Kernel Distribution Closeness Testing", "abs": "The \\emph{distribution closeness testing} (DCT) assesses whether the distance between an unknown distribution pair is at least $\\epsilon$-far; in practice, the $\\epsilon$ can be defined as the distance between a reference (known) distribution pair. However, existing DCT methods are mainly measure discrepancies between a distribution pair defined on discrete one-dimensional spaces (e.g., total variation on a discrete one-dimensional space), which limits the DCT to be used on complex data (e.g., images). To make DCT applicable on complex data, a natural idea is to introduce the \\emph{maximum mean discrepancy} (MMD), a powerful measurement to see the difference between a pair of two complex distributions, to DCT scenarios. Nonetheless, in this paper, we find that MMD value is less informative \\textcolor{blue}{when assessing the closeness levels for multiple distribution pairs with the same kernel, i.e., MMD value can be the same for many pairs of distributions that have different norms in the same \\emph{reproducing kernel Hilbert space} (RKHS). To mitigate the issue, we propose a new kernel DCT with the \\emph{norm-adaptive MMD} (NAMMD) by scaling MMD with the norms of distributions, effective for kernels $\\kappa(\\x,\\x')=\\Psi(\\x-\\x')\\leq K$ with a positive-definite $\\Psi(\\cdot)$ and $\\Psi(\\bm{0})=K$.} Theoretically, we prove that our NAMMD test achieves higher test power compared to the MMD test, along with asymptotic distribution analysis. We also present upper bounds on the sample complexity of our NAMMD test and prove that Type-I error is controlled. We finally conduct experiments to validate the effectiveness of our NAMMD test.", "keywords": [ "hypothesis testing", "Maximum Mean Discrepancy", "distribution closeness testing", "two samples testing" ], "rating_list": [ 8, 6, 6, 3 ], "soundness_list": [ 3, 3, 4, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "fgUFZAxywx", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Ashkan Shahbazi", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "PhD student" }, { "name": "Bennett Allan Landman", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "Full Professor" }, { "name": "Catie Chang", "gender": "Female", "institution": "Vanderbilt University", "country": "US", "position": "Assistant Professor" }, { "name": "Kaiwen Shi", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "Undergrad student" }, { "name": "Soheil Kolouri", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "Assistant Professor" }, { "name": "Xinran Liu", "gender": "unknown", "institution": "Vanderbilt University", "country": "US", "position": "PhD student" }, { "name": "Yikun Bai", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "Postdoc" }, { "name": "ROCIO DIAZ MARTIN", "gender": "Female", "institution": "Tufts University", "country": "US", "position": "Postdoc" } ], "n_formula": 127, "n_formula_1": 21, "n_ref_uni": 56, "n_ref": 115, "n_ref_all": 162, "n_fig": 11, "n_tab": 4, "L_tab": 1011, "n_element_tab": 72, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 2822, "n_element_tab_1": 68, "formula_len_all": 11174, "formula_len_all_1": 2542, "len_all": 246830, "len_all_1": 65407, "len_abs": 5404, "len_title": 133, "len_sents": 61539, "len_sents_1": 26272, "n_sents": 614, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 85, "L_abs": 1299, "title": "Linear Spherical Sliced Optimal Transport: A Fast Metric for Comparing Spherical Data", "abs": "Efficient comparison of spherical probability distributions becomes important in fields such as computer vision, geosciences, and medicine. Sliced optimal transport distances, such as spherical and stereographic spherical sliced Wasserstein distances, have recently been developed to address this need. These methods reduce the computational burden of optimal transport by slicing hyperspheres into one-dimensional projections, i.e., lines or circles. Concurrently, linear optimal transport has been proposed to embed distributions into $L^2$ spaces, where the $L^2$ distance approximates the optimal transport distance, thereby simplifying comparisons across multiple distributions. In this work, we introduce the Linear Spherical Sliced Optimal Transport (LSSOT) framework, which utilizes slicing to embed spherical distributions into $L^2$ spaces while preserving their intrinsic geometry, offering a computationally efficient metric for spherical probability measures. We establish the metricity of LSSOT and demonstrate its superior computational efficiency in applications such as cortical surface registration, 3D point cloud interpolation via gradient flow, and shape embedding. Our results demonstrate the significant computational benefits and high accuracy of LSSOT in these applications.", "keywords": [ "Optimal Transport", "Spherical Data Analysis" ], "rating_list": [ 6, 10, 6, 8 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 4, 4, 1, 3 ] }, { "paper_id": "ffuHn3Q6Hc", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Shibhansh Dohare", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "PhD student" }, { "name": "J. A. Hernandez", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "PhD student" }, { "name": "Richard S. Sutton", "gender": "Male", "institution": "Keen Technologies", "country": "CA", "position": "Research Scientist" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 41, "n_ref": 100, "n_ref_all": 137, "n_fig": 19, "n_tab": 2, "L_tab": 232, "n_element_tab": 29, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 232, "n_element_tab_1": 29, "formula_len_all": 461, "formula_len_all_1": 141, "len_all": 167371, "len_all_1": 71238, "len_abs": 2160, "len_title": 126, "len_sents": 58890, "len_sents_1": 36505, "n_sents": 421, "n_sents_1": 251, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1272, "title": "Reinitializing weights vs hidden units for maintaining plasticity in neural networks", "abs": "Loss of plasticity is a phenomenon where a neural network loses its ability to learn when trained for an extended time on non-stationary data.\nIt is a crucial problem to overcome when designing systems that learn continually.\nAn effective technique for preventing loss of plasticity is reinitializing parts of the network.\nIn this paper, we compare two different reinitialization schemes: reinitializing units vs reinitializing weights.\nWe propose a new algorithm named \\textit{selective weight reinitialization} for reinitializing the least useful weights in the network. \nWe compare our algorithm to continual backpropagation, a previously proposed algorithm that reinitializes units.\nThrough our experiments in continual supervised learning problems, we identify two settings when reinitializing weights is more effective at maintaining plasticity than reinitializing units: (1) when the network has a small number of units and (2) when the network includes layer normalization.\nConversely, reinitializing weights and units are equally effective at maintaining plasticity when the network is of sufficient size and does not include layer normalization. \nWe found that reinitializing weights maintains plasticity in a wider variety of settings than reinitializing units.", "keywords": [ "Continual Learning", "Supervised Learning", "Deep Learning", "Loss of Plasticity" ], "rating_list": [ 5, 3, 8 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 1, 1, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "ff2V3UR9sC", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Balaji Krishnamurthy", "gender": "Male", "institution": "Adobe Systems", "country": "IN", "position": "Principal Scientist" }, { "name": "Changyou Chen", "gender": "Male", "institution": "State University of New York, Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Harini S I", "gender": "Female", "institution": "Adobe Systems", "country": "IN", "position": "Researcher" }, { "name": "Rajiv Ratn Shah", "gender": "Male", "institution": "Indraprastha Institute of Information Technology, Delhi", "country": "IN", "position": "Associate Professor" }, { "name": "Somesh Kumar Singh", "gender": "Male", "institution": "Indraprastha Institute of Information Technology, Delhi", "country": "IN", "position": "PhD student" }, { "name": "Veeky Baths", "gender": "Male", "institution": "Birla Institute of Technology and Science", "country": "", "position": "Associate Professor" }, { "name": "Yaman Kumar", "gender": "Male", "institution": "Adobe ", "country": "IN", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 60, "n_ref": 103, "n_ref_all": 137, "n_fig": 3, "n_tab": 14, "L_tab": 4838, "n_element_tab": 517, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 574, "n_element_tab_1": 63, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 255722, "len_all_1": 61242, "len_abs": 2063, "len_title": 121, "len_sents": 63593, "len_sents_1": 29523, "n_sents": 476, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1275, "title": "Teaching Human Behavior Improves Content Understanding Abilities Of LLMs", "abs": "Communication is defined as ``Who says what to whom with what effect.'' A message from a communicator generates downstream receiver effects, also known as behavior. Receiver behavior, being a downstream effect of the message, carries rich signals about it. Even after carrying signals about the message, the behavior data is often ignored while training large language models. We show that training LLMs on receiver behavior can actually help improve their content-understanding abilities. Specifically, we show that training LLMs to predict the receiver behavior of likes and comments improves the LLM's performance on a wide variety of downstream content understanding tasks. We show this performance increase over 46 video and image understanding tasks over 26 benchmark datasets across both 0-shot and fine-tuning settings, outperforming many supervised baselines. Moreover, since receiver behavior, such as likes and comments, is collected by default on the internet and does not need any human annotations to be useful, the performance improvement we get after training on this data is essentially free-lunch. We release the receiver behavior cleaned comments and likes of 750k images and videos collected from multiple platforms along with our instruction-tuning data.", "keywords": [ "LLM", "human behavior", "content understanding" ], "rating_list": [ 6, 5, 8, 8 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "feykAFeSfq", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Fahim Hafiz", "gender": "Male", "institution": "United International University", "country": "BD", "position": "Lecturer" }, { "name": "Iftekharul Abedeen", "gender": "Male", "institution": "United International University", "country": "BD", "position": "Lecturer" }, { "name": "Swakkhar Shatabda", "gender": "unknown", "institution": "BRAC University", "country": "BD", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 25, "n_ref": 63, "n_ref_all": 80, "n_fig": 6, "n_tab": 3, "L_tab": 1948, "n_element_tab": 583, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 2213, "n_element_tab_1": 637, "formula_len_all": 489, "formula_len_all_1": 182, "len_all": 107044, "len_all_1": 68207, "len_abs": 836, "len_title": 129, "len_sents": 33179, "len_sents_1": 28578, "n_sents": 259, "n_sents_1": 223, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 841, "title": "LayoutRL: A Reinforcement Learning-Based Approach to Keyboard Layout Optimization", "abs": "Keyboards are a key interface between humans and computers, with character arrangements offering numerous layout possibilities. Many existing designs follow standardized ergonomic principles and explore Pareto-optimality in multi-objective functions using metaheuristics or deep learning. In this work, we propose a reinforcement learning-based approach to designing optimized keyboard layouts that integrate both technical and ergonomic considerations. Our results demonstrate that reinforcement learning optimization can produce layouts more efficiently than conventional designs, such as the \"QWERTY\" keyboard. Specifically, our approach achieves approximately an 12.4\\% improvement in ergonomic parameters over traditional keyboards, underscoring the potential for a more data-driven, systematic approach to keyboard layout optimization.", "keywords": [ "Layout Optimization", "Ergonomics", "Data-driven Optimization", "Interaction Design", "Reinforcement learning", "HCI" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "fectsEG2GU", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Chun Gu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Li Zhang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Zeyu Yang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Zijie Pan", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 6, "n_ref_uni": 39, "n_ref": 115, "n_ref_all": 137, "n_fig": 11, "n_tab": 3, "L_tab": 784, "n_element_tab": 103, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 496, "n_element_tab_1": 54, "formula_len_all": 1936, "formula_len_all_1": 616, "len_all": 157562, "len_all_1": 64614, "len_abs": 1880, "len_title": 154, "len_sents": 47053, "len_sents_1": 30921, "n_sents": 334, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 107, "L_abs": 1888, "title": "Diffusion$^2$: Dynamic 3D Content Generation via Score Composition of Video and Multi-view Diffusion Models", "abs": "Recent advancements in 3D generation are predominantly propelled by improvements in 3D-aware image diffusion models. These models are pretrained on Internet-scale image data and fine-tuned on massive 3D data, offering the capability of producing highly consistent multi-view images. However, due to the scarcity of synchronized multi-view video data, it remains challenging to adapt this paradigm to 4D generation directly. Despite that, the available video and 3D data are adequate for training video and multi-view diffusion models separately that can provide satisfactory dynamic and geometric priors respectively. To take advantage of both, this paper presents Diffusion$^2$, a novel framework for dynamic 3D content creation that reconciles the knowledge about geometric consistency and temporal smoothness from these models to directly sample dense multi-view multi-frame images which can be employed to optimize continuous 4D representation. Specifically, we design a simple yet effective denoising strategy via score composition of pretrained video and multi-view diffusion models based on the probability structure of the target image array. To alleviate the potential conflicts between two heterogeneous scores, we further introduce variance-reducing sampling via interpolated steps, facilitating smooth and stable generation. Owing to the high parallelism of the proposed image generation process and the efficiency of the modern 4D reconstruction pipeline, our framework can generate 4D content within few minutes. Notably, our method circumvents the reliance on expensive and hard-to-scale 4D data, thereby having the potential to benefit from the scaling of the foundation video and multi-view diffusion models. Extensive experiments demonstrate the efficacy of our proposed framework in generating highly seamless and consistent 4D assets under various types of conditions.", "keywords": [ "4D generation", "diffusion model", "generative model", "video diffusion" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "feZcpZbbqL", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Fanhua Shang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" }, { "name": "Hongying Liu", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Qi Zhu", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yuanyuan Liu", "gender": "Not Specified", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Zhouchen Lin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Professor" } ], "n_formula": 149, "n_formula_1": 16, "n_ref_uni": 37, "n_ref": 63, "n_ref_all": 74, "n_fig": 2, "n_tab": 2, "L_tab": 851, "n_element_tab": 107, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 19739, "formula_len_all_1": 1755, "len_all": 178657, "len_all_1": 56785, "len_abs": 1465, "len_title": 169, "len_sents": 45043, "len_sents_1": 25500, "n_sents": 446, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Reject", "L_title": 121, "L_abs": 1489, "title": "Finding Second-order Stationary Points for Generalized-Smooth Nonconvex Minimax Optimization via Gradient-based Algorithm", "abs": "Nonconvex minimax problems have received intense interest in many machine learning applications such as generative adversarial network, robust optimization and adversarial \nRecently, a variety of minimax optimization algorithms based on Lipschitz smoothness \nfor finding first-order or second-order stationary points have been proposed.\nHowever, the standard Lipschitz continuous gradient or Hessian assumption could fail to hold even in some classic minimax problems,\nrendering conventional minimax optimization algorithms fail to converge in practice.\nTo address this challenge, we demonstrate a new gradient-based method for nonconvex-strongly-concave minimax optimization \nunder a generalized smoothness assumption.\nMotivated by the important application of escaping saddle points, we propose a generalized Hessian smoothness condition, \nunder which our gradient-based method can achieve the complexity of $\\mathcal{O}(\\epsilon^{-1.75}\\log n)$ \nto find a second-order stationary point with only gradient calls involved, \nwhich improves the state-of-the-art complexity results for the nonconvex minimax optimization \neven under standard Lipschitz smoothness condition.\nTo the best of our knowledge, this is the first work to show convergence \nfor finding second-order stationary points on nonconvex minimax optimization with generalized smoothness.\nThe experimental results on the application of domain adaptation confirm the superiority of our algorithm compared with existing methods.", "keywords": [ "Minimax Optimization", "Nonconvex Optimization", "Generalized Smoothness", "Second-order Stationary Point" ], "rating_list": [ 6, 6, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 4, 2, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 2, 5 ] }, { "paper_id": "feFlfuOse1", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 16, "author_info_list": [ { "name": "Ariel Kwiatkowski", "gender": "Not Specified", "institution": "Facebook", "country": "FR", "position": "Researcher" }, { "name": "Arjun KG", "gender": "unknown", "institution": "Earthbrain", "country": "JP", "position": "Researcher" }, { "name": "Gianluca De Cola", "gender": "Male", "institution": "---", "country": "IT", "position": "Software Engineer" }, { "name": "Hannah Tan", "gender": "Female", "institution": "DBS Bank Limited", "country": "SG", "position": "Machine Learning Engineer" }, { "name": "J K Terry", "gender": "Female", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "John U. Balis", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "", "position": "PhD student" }, { "name": "Jun Jet Tai", "gender": "Male", "institution": "Coventry University", "country": "", "position": "PhD student" }, { "name": "Kallinteris Andreas", "gender": "unknown", "institution": "Technical University of Crete", "country": "GR", "position": "Undergrad student" }, { "name": "Manuel Goul\u00e3o", "gender": "Male", "institution": "Instituto Superior T\u00e9cnico", "country": "", "position": "MS student" }, { "name": "Mark Towers", "gender": "Male", "institution": "University of Southampton", "country": "GB", "position": "PhD student" }, { "name": "Markus Krimmel", "gender": "Male", "institution": "Max Planck Institute of Biochemistry", "country": "", "position": "PhD student" }, { "name": "Rodrigo De Lazcano Perez-Vicente", "gender": "Male", "institution": "Farama Foundation", "country": "", "position": "Researcher" }, { "name": "Sander V Schulhoff", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "Undergrad student" }, { "name": "Tristan Deleu", "gender": "unknown", "institution": "University of Montreal", "country": "", "position": "PhD student" }, { "name": "Andrea Pierre", "gender": "Male", "institution": "University of Massachusetts at Lowell", "country": "US", "position": "PhD student" }, { "name": "Omar Gallal Younis", "gender": "unknown", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 27, "n_ref": 29, "n_ref_all": 33, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 83839, "len_all_1": 66077, "len_abs": 1255, "len_title": 119, "len_sents": 24298, "len_sents_1": 24280, "n_sents": 187, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1260, "title": "Gymnasium: A Standard Interface for Reinforcement Learning Environments", "abs": "Reinforcement Learning (RL) is a continuously growing field that has the potential to revolutionize many areas of artificial intelligence. However, despite its promise, RL research is often hindered by the lack of standardization in environment and algorithm implementations. This makes it difficult for researchers to compare and build upon each other's work, slowing down progress in the field.\nGymnasium is an open-source library that provides a standard API for RL environments, aiming to tackle this issue. Gymnasium's main feature is a set of abstractions that allow for wide interoperability between environments and training algorithms, making it easier for researchers to develop and test RL algorithms. In addition, Gymnasium provides a collection of easy-to-use environments, tools for easily customizing environments, and tools to ensure the reproducibility and robustness of RL research.\nThrough this unified framework, Gymnasium significantly streamlines the process of developing and testing RL algorithms, enabling researchers to focus more on innovation and less on implementation details. By providing a standardized platform for RL research, Gymnasium helps to drive forward the field of reinforcement learning and unlock its full potential.", "keywords": [ "reinforcement learning", "api", "gymnasium", "artificial intelligence", "autonomous agents", "environment" ], "rating_list": [ 8, 6, 10, 5 ], "soundness_list": [ 4, 4, 4, 2 ], "presentation_list": [ 2, 3, 4, 4 ], "contribution_list": [ 4, 2, 4, 2 ], "confidence_list": [ 5, 2, 5, 3 ] }, { "paper_id": "fdvSCcB7i8", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Fang Chen", "gender": "unknown", "institution": "University of Technology Sydney (UTS)", "country": "AU", "position": "Full Professor" }, { "name": "Jiahao Huang", "gender": "unknown", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "MS student" }, { "name": "Jianlong Zhou", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Jiayu Zhang", "gender": "Male", "institution": "Suzhou Yierqi", "country": "", "position": "Researcher" }, { "name": "Xinyi Zhang", "gender": "Female", "institution": "University of Sydney, University of Sydney", "country": "CN", "position": "Undergrad student" }, { "name": "Zhibo Jin", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Zhu Zhiyu", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 32, "n_ref": 41, "n_ref_all": 61, "n_fig": 9, "n_tab": 3, "L_tab": 913, "n_element_tab": 73, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2779, "n_element_tab_1": 65, "formula_len_all": 721, "formula_len_all_1": 680, "len_all": 99004, "len_all_1": 59669, "len_abs": 1917, "len_title": 82, "len_sents": 33092, "len_sents_1": 28366, "n_sents": 231, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 34, "L_abs": 1924, "title": "Feature Level Instance Attribution", "abs": "Instance attribution has emerged as one of the most crucial methodologies for model explainability because it identifies training data that significantly impacts model predictions, thereby optimizing model performance and enhancing transparency and trustworthiness. The applications of instance attribution include data cleaning, where it identifies and rectifies poor-quality data to improve model outcomes, and in specific domains such as detection of harmful speech, social network graph labeling, and medical image annotation, it provides precise insights into how data influences model decisions. Specifically, current instance attribution methods facilitate the identification of causal relationships between training data and model predictions. A higher Instance-level Training Data Influence value (IL value) indicates that the training data used for the computation play a more significant role in the model's prediction process. However, the current methods can only indicate that a training sample is important, but they do not explain why this sample is important. A feasible algorithm is urgently needed to provide an explanation for this behavior. This paper discovers that artificially manipulating the attribution score by modifying samples (e.g., changing a pixel value in image data) can significantly intervene in the importance of training samples and yield explainability results at the feature-level during the intervention process. The proposed Feature Level Instance Attribution (FLIA) algorithm assists in identifying crucial feature locations in training data that significantly impact causality. To avoid the frequent retraining of models for evaluation, we introduce an unlearning algorithm as an assessment method and provide detailed empirical evidence of our algorithm's efficacy. To facilitate future research, we have made the code available at: https://anonymous.4open.science/r/FIIA-D60E/.", "keywords": [ "Interpretability", "attribution" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 1, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "fd2u60ryG0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Jiawei He", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Postdoc" }, { "name": "Lue Fan", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Tieniu Tan", "gender": "Male", "institution": " Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Professor" }, { "name": "Yingyan Li", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yuntao Chen", "gender": "Male", "institution": "Centre for Artificial Intelligence and Robotics (CAIR), Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "HK", "position": "Assistant Professor" }, { "name": "Yuqi Wang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zhaoxiang Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 39, "n_ref": 118, "n_ref_all": 133, "n_fig": 2, "n_tab": 6, "L_tab": 2361, "n_element_tab": 285, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 2519, "n_element_tab_1": 259, "formula_len_all": 563, "formula_len_all_1": 523, "len_all": 147042, "len_all_1": 66997, "len_abs": 1049, "len_title": 111, "len_sents": 33757, "len_sents_1": 25238, "n_sents": 339, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1063, "title": "Enhancing End-to-End Autonomous Driving with Latent World Model", "abs": "In autonomous driving, end-to-end planners directly utilize raw sensor data, enabling them to extract richer scene features and reduce information loss compared to traditional planners. This raises a crucial research question: how can we develop better scene feature representations to fully leverage sensor data in end-to-end driving? Self-supervised learning methods show great success in learning rich feature representations in NLP and computer vision. Inspired by this, we propose a novel self-supervised learning approach using the LAtent World model (LAW) for end-to-end driving. LAW predicts future latent scene features based on current features and ego trajectories. This self-supervised task can be seamlessly integrated into perception-free and perception-based frameworks, improving scene feature learning while optimizing trajectory prediction. LAW achieves state-of-the-art performance across multiple benchmarks, including real-world open-loop benchmark nuScenes, NAVSIM, and simulator-based closed-loop benchmark CARLA. The code will be released.", "keywords": [ "end-to-end autonomous driving", "world model", "self-supervised learning" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "fcJKzwlwcs", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Chenjia Bai", "gender": "Male", "institution": "TeleAI, China Telecom", "country": "CN", "position": "Research Scientist" }, { "name": "Chenyou Fan", "gender": "unknown", "institution": "Northwestern Polytechnical University, Northwest Polytechnical University Xi'an", "country": "CN", "position": "MS student" }, { "name": "Haoran He", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yang Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhao Shan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhen Wang", "gender": "Male", "institution": "Northwestern Polytechnical University", "country": "CN", "position": "Full Professor" } ], "n_formula": 31, "n_formula_1": 25, "n_ref_uni": 57, "n_ref": 104, "n_ref_all": 148, "n_fig": 7, "n_tab": 7, "L_tab": 3644, "n_element_tab": 292, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 818, "n_element_tab_1": 78, "formula_len_all": 2784, "formula_len_all_1": 1672, "len_all": 197014, "len_all_1": 73997, "len_abs": 2044, "len_title": 134, "len_sents": 51329, "len_sents_1": 31743, "n_sents": 427, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1446, "title": "Task-agnostic Pre-training and Task-guided Fine-tuning for Versatile Diffusion Planner", "abs": "Diffusion models have demonstrated their capabilities in modeling trajectories of multi-tasks. However, existing multi-task planners or policies typically rely on task-specific demonstrations via multi-task imitation, or require task-specific reward labels to facilitate policy optimization via Reinforcement Learning (RL). They heavily rely on the task-specific labeled data which can be difficult to acquire. To address these challenges, we aim to develop a versatile diffusion planner that can leverage large-scale inferior data that contains task-agnostic sub-optimal trajectories, with the ability to fast adapt to specific tasks. In this paper, we propose SODP, a two-stage framework that leverages Sub-Optimal data to learn a Diffusion Planner, which is generalizable for various downstream tasks. Specifically, in the pre-training stage, we train a foundation diffusion planner that extracts general planning capabilities by modeling the versatile distribution of multi-task trajectories, which can be sub-optimal and has wide data coverage. Then for downstream tasks, we adopt RL-based fine-tuning with task-specific rewards to fast refine the diffusion planner, which aims to generate action sequences with higher task-specific returns. Experimental results from multi-task domains including Meta-World and Adroit demonstrate that SODP outperforms state-of-the-art methods with only a small amount of data for reward-guided fine-tuning.", "keywords": [ "reinforcement learning", "diffusion models", "planning" ], "rating_list": [ 3, 5, 6, 5, 5 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3, 3 ] }, { "paper_id": "fbrHgEBx2f", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Kuan-Ta Li", "gender": "unknown", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "PhD student" }, { "name": "Ping-Chun Hsieh", "gender": "Male", "institution": "National Chiao Tung University", "country": "", "position": "Assistant Professor" }, { "name": "Yu-Chih Huang", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" } ], "n_formula": 129, "n_formula_1": 23, "n_ref_uni": 35, "n_ref": 82, "n_ref_all": 115, "n_fig": 15, "n_tab": 4, "L_tab": 524, "n_element_tab": 46, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9931, "formula_len_all_1": 2096, "len_all": 185108, "len_all_1": 72698, "len_abs": 1058, "len_title": 139, "len_sents": 60530, "len_sents_1": 32879, "n_sents": 566, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1064, "title": "Diminishing Exploration: A Minimalist Approach to Piecewise Stationary Multi-Armed Bandits", "abs": "The piecewise-stationary bandit problem is an important variant of the multi-armed bandit problem that further considers abrupt changes in the reward distributions. The main theme of the problem is the trade-off between exploration for detecting environment changes and exploitation of traditional bandit algorithms. While this problem has been extensively investigated, existing works either assume knowledge about the number of change points $M$ or require extremely high computational complexity. In this work, we revisit the piecewise-stationary bandit problem from a minimalist perspective. We propose a novel and generic exploration mechanism, called diminishing exploration, which eliminates the need for knowledge about $M$ and can be used in conjunction with an existing change detection-based algorithm to achieve near-optimal regret scaling. Simulation results show that despite oblivious of $M$, equipping existing algorithms with the proposed diminishing exploration generally achieves better empirical regret than the traditional uniform exploration.", "keywords": [ "Piecewise-stationary bandit", "multi-armed bandit", "regret analysis", "change detection" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "fbqOEOqurU", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Jingcheng Liu", "gender": "Male", "institution": "nanjing university", "country": "", "position": "Associate Professor" }, { "name": "Zongrui Zou", "gender": "Male", "institution": "Nanjing university", "country": "CN", "position": "PhD student" }, { "name": "Jalaj Upadhyay", "gender": "Male", "institution": "Rutgers University", "country": "", "position": "Assistant Professor" } ], "n_formula": 123, "n_formula_1": 39, "n_ref_uni": 24, "n_ref": 93, "n_ref_all": 133, "n_fig": 0, "n_tab": 2, "L_tab": 943, "n_element_tab": 22, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 11577, "formula_len_all_1": 3693, "len_all": 182678, "len_all_1": 68772, "len_abs": 733, "len_title": 95, "len_sents": 56489, "len_sents_1": 26616, "n_sents": 536, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 759, "title": "Optimality of Matrix Mechanism on $\\ell_p^p$-metric", "abs": "In this paper, we introduce the $\\ell_p^p$-error metric (for $p \\geq 2$) when answering linear queries under the constraint of differential privacy. We characterize such an error under $(\\epsilon,\\delta)$-differential privacy in the natural add/remove model. Before this paper, tight characterization in the hardness of privately answering linear queries was known under $\\ell_2^2$-error metric (Edmonds et al. 2020) and $\\ell_p^2$-error metric for unbiased mechanisms in the substitution model (Nikolov et al. 2024). As a direct consequence of our results, we give tight bounds on answering prefix sum and parity queries under differential privacy for all constant $p$ in terms of the $\\ell_p^p$ error, generalizing the bounds in Hhenzinger et al. for $p=2$.", "keywords": [ "differential privacy" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 4, 4, 4, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "falBlwUsIH", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Hong Yang", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Qi Yu", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "Professor" }, { "name": "Travis Desell", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 3, "n_ref_uni": 47, "n_ref": 110, "n_ref_all": 120, "n_fig": 5, "n_tab": 2, "L_tab": 1168, "n_element_tab": 103, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 701, "n_element_tab_1": 39, "formula_len_all": 965, "formula_len_all_1": 425, "len_all": 161137, "len_all_1": 73900, "len_abs": 1275, "len_title": 102, "len_sents": 53149, "len_sents_1": 35571, "n_sents": 453, "n_sents_1": 295, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1281, "title": "Can We Ignore Labels in Out of Distribution Detection?", "abs": "Out-of-distribution (OOD) detection methods have recently become more prominent, serving as a core element in safety-critical autonomous systems. One major purpose of OOD detection is to reject invalid inputs that could lead to unpredictable errors and compromise safety. Due to the cost of labeled data, recent works have investigated the feasibility of self-supervised learning (SSL) OOD detection, unlabled OOD detection, and zero shot OOD detection. In this work, we identify a set of conditions for a theoretical guarantee of failure in unlabeled OOD detection algorithms from an information-theoretic perspective. These conditions are present in all OOD tasks dealing with real world data: I) we provide theoretical proof of unlabeled OOD detection failure when there exists zero mutual information between the learning objective and the in-distribution labels, a.k.a. \u2018label blindness\u2019, II) we define a new OOD task \u2013 Adjacent OOD detection \u2013 that tests for label blindness and accounts for a previously ignored safety gap in all OOD detection benchmarks, and III) we perform experiments demonstrating that existing unlabeled OOD methods fail under conditions suggested by our label blindness theory and analyze the implications for future research in unlabeled OOD methods.", "keywords": [ "Out of Distribution", "Uncertainty", "Self Supervised", "Unsupervised" ], "rating_list": [ 6, 5, 8 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "faSfhqDpZP", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Kailing Wang", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Keyang Zhao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Postdoc" }, { "name": "Wei Shen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaokang Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University, China", "country": "CN", "position": "Professor" }, { "name": "chen yang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 22, "n_ref_uni": 31, "n_ref": 86, "n_ref_all": 100, "n_fig": 4, "n_tab": 2, "L_tab": 144, "n_element_tab": 16, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 144, "n_element_tab_1": 16, "formula_len_all": 1299, "formula_len_all_1": 1302, "len_all": 139131, "len_all_1": 67223, "len_abs": 1660, "len_title": 99, "len_sents": 35273, "len_sents_1": 30579, "n_sents": 283, "n_sents_1": 246, "venue": "ICLR2025", "decision": "withdraw", "L_title": 51, "L_abs": 1669, "title": "Realistic Surgical Simulation from Monocular Videos", "abs": "This paper tackles the challenge of automatically constructing realistic surgical simulation systems from readily available surgical videos. Recent efforts have successfully integrated physically grounded dynamics within 3D Gaussians to perform high-fidelity simulations in well-reconstructed static simulation environments. However, they struggle with the geometry inconsistency of simulation environments and unrealistic physical deformations of soft tissues when it comes to dynamic and complex surgical processes. In this paper, we propose SurgiSim, a novel automatic simulation system to overcome these limitations. To build a surgical simulation environment, we maintain a canonical 3D scene composed of 3D Gaussians coupled with a deformation field to accurately model monocular dynamic surgical scenes. This process involves a multi-stage optimization with trajectory and anisotropic regularization, enhancing the geometry consistency of the canonical scene which serves the simulation environment. To improve the realism of physical simulations, we implement a Visco-Elastic deformation model based on the Maxwell model, effectively restoring the complex deformations of tissues. Additionally, we estimate the physical properties of tissues by minimizing the discrepancies between the input video and simulation results guided by predicted tissue motion, ensuring realistic simulation outcomes. Experiments across diverse surgical scenarios demonstrate SurgiSim's ability to perform realistic physical interactions of soft tissues among surgical procedures, showing its enormous potential for enhancing surgical training, planning, and robotic surgery systems.", "keywords": [ "Surgical Simulation", "Video-based Reconstruction", "Robotic Surgery" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 4 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "faDMOmnsjx", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Huy Nguyen", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Huyen Trang Pham", "gender": "Female", "institution": "VinAI Research", "country": "VN", "position": "Intern" }, { "name": "Nhat Ho", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Pedram Akbarian", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Shujian Zhang", "gender": "unknown", "institution": "University of Texas, Austin", "country": "", "position": "PhD student" }, { "name": "Thien Trang Nguyen Vu", "gender": "Female", "institution": "VinAI Research", "country": "VN", "position": "Researcher" } ], "n_formula": 161, "n_formula_1": 38, "n_ref_uni": 44, "n_ref": 80, "n_ref_all": 129, "n_fig": 6, "n_tab": 5, "L_tab": 1067, "n_element_tab": 109, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18090, "formula_len_all_1": 3089, "len_all": 252763, "len_all_1": 76964, "len_abs": 1603, "len_title": 120, "len_sents": 69924, "len_sents_1": 32996, "n_sents": 655, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1533, "title": "Statistical Advantages of Perturbing Cosine Router in Mixture of Experts", "abs": "The cosine router in Mixture of Experts (MoE) has recently emerged as an attractive alternative to the conventional linear router. Indeed, the cosine router demonstrates favorable performance in image and language tasks and exhibits better ability to mitigate the representation collapse issue, which often leads to parameter redundancy and limited representation potentials. Despite its empirical success, a comprehensive analysis of the cosine router in MoE has been lacking. Considering the least square estimation of the cosine routing MoE, we demonstrate that due to the intrinsic interaction of the model parameters in the cosine router via some partial differential equations, regardless of the structures of the experts, the estimation rates of experts and model parameters can be as slow as $\\mathcal{O}(1/\\log^{\\tau}(n))$ where $\\tau > 0$ is some constant and $n$ is the sample size. Surprisingly, these pessimistic non-polynomial convergence rates can be circumvented by the widely used technique in practice to stabilize the cosine router --- simply adding noises to the $L^2$ norms in the cosine router, which we refer to as *perturbed cosine router*. Under the strongly identifiable settings of the expert functions, we prove that the estimation rates for both the experts and model parameters under the perturbed cosine routing MoE are significantly improved to polynomial rates. Finally, we conduct extensive simulation studies in both synthetic and real data settings to empirically validate our theoretical results.", "keywords": [ "mixture of experts", "cosine router", "perturbation" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 2, 2, 4 ] }, { "paper_id": "fZK6AQXlUU", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Bhargava Kumar", "gender": "Male", "institution": "TD Securities", "country": "US", "position": "VP-TDS, Technology Solutions" }, { "name": "Mouloud Belbahri", "gender": "unknown", "institution": "Layer6 AI", "country": "CA", "position": "Researcher" }, { "name": "Yi Sui", "gender": "Female", "institution": "Layer6 AI", "country": "CA", "position": "Machine Learning Scientist" }, { "name": "Jesse C Cresswell", "gender": "unknown", "institution": "Layer 6 AI", "country": "CA", "position": "Staff Machine Learning Scientist" } ], "n_formula": 18, "n_formula_1": 14, "n_ref_uni": 54, "n_ref": 103, "n_ref_all": 146, "n_fig": 7, "n_tab": 8, "L_tab": 1101, "n_element_tab": 156, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 577, "n_element_tab_1": 50, "formula_len_all": 995, "formula_len_all_1": 790, "len_all": 177437, "len_all_1": 69111, "len_abs": 925, "len_title": 100, "len_sents": 55648, "len_sents_1": 32691, "n_sents": 403, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 52, "L_abs": 1010, "title": "Conformal Prediction Sets Can Cause Disparate Impact", "abs": "Conformal prediction is a statistically rigorous method for quantifying uncertainty in models by having them output sets of predictions, with larger sets indicating more uncertainty. However, prediction sets are not inherently actionable; many applications require a single output to act on, not several. To overcome this limitation, prediction sets can be provided to a human who then makes an informed decision. In any such system it is crucial to ensure the fairness of outcomes across protected groups, and researchers have proposed that Equalized Coverage be used as the standard for fairness. By conducting experiments with human participants, we demonstrate that providing prediction sets can lead to disparate impact in decisions. Disquietingly, we find that providing sets that satisfy Equalized Coverage actually increases disparate impact compared to marginal coverage. Instead of equalizing coverage, we propose to equalize set sizes across groups which empirically leads to lower disparate impact.", "keywords": [ "Conformal Prediction", "Fairness", "Uncertainty Quantification", "Trustworthy ML", "Human Subject Experiments" ], "rating_list": [ 8, 8, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "fYOl9leH72", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Adrian Weller", "gender": "Male", "institution": "Alan Turing Institute", "country": "", "position": "Programme Director for AI" }, { "name": "Yanzhi Chen", "gender": "unknown", "institution": "Microsoft Research ", "country": "GB", "position": "Researcher" }, { "name": "Yingzhen Li", "gender": "Female", "institution": "Imperial College London", "country": "", "position": "Associate Professor" }, { "name": "Zijing Ou", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 23, "n_ref_uni": 31, "n_ref": 81, "n_ref_all": 98, "n_fig": 3, "n_tab": 4, "L_tab": 280, "n_element_tab": 40, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1525, "formula_len_all_1": 1072, "len_all": 164664, "len_all_1": 70793, "len_abs": 877, "len_title": 113, "len_sents": 46741, "len_sents_1": 33749, "n_sents": 397, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 880, "title": "Neural Mutual Information Estimation with Reference Distributions", "abs": "Estimating mutual information (MI) from data is a fundamental task in machine learning and data science, yet it remains highly challenging even with state-of-the-art estimators. This work proposes a new distribution-free MI estimator based on reference distributions. Unlike existing works that only discern between the joint distribution and the marginal distribution, which can easily overfit in high-MI settings, our method compares them with extra reference distributions. These artificial distributions share the same marginals as the original distributions but have known dependence structures, providing additional signals for more accurate dependency modeling. Experiments on synthetic tasks with non-Gaussian, high-dimensional data and real-world applications including Bayesian experimental design and self-supervised learning demonstrate the potential of our approach.", "keywords": [ "mutual information", "density ratio estimate", "flow-based model", "neural density estimate", "copula" ], "rating_list": [ 1, 5, 5, 5, 5 ], "soundness_list": [ 1, 2, 4, 3, 3 ], "presentation_list": [ 1, 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "fXnE4gB64o", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Kevin Lin", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Lijuan Wang", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Qiang Qiu", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "Ze Wang", "gender": "Male", "institution": "Advanced Micro Devices", "country": "US", "position": "Researcher" }, { "name": "Zhengyuan Yang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Zichen Miao", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Zicheng Liu", "gender": "Male", "institution": "Advanced Micro Devices", "country": "US", "position": "Sr. Director GenAI" } ], "n_formula": 17, "n_formula_1": 12, "n_ref_uni": 38, "n_ref": 110, "n_ref_all": 128, "n_fig": 11, "n_tab": 5, "L_tab": 3103, "n_element_tab": 370, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2134, "n_element_tab_1": 289, "formula_len_all": 2036, "formula_len_all_1": 1714, "len_all": 140951, "len_all_1": 57528, "len_abs": 1816, "len_title": 124, "len_sents": 36514, "len_sents_1": 22584, "n_sents": 267, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1822, "title": "Tuning Timestep-Distilled Diffusion Model Using Pairwise Sample Optimization", "abs": "Recent advancements in timestep-distilled diffusion models have enabled high-quality image generation that rivals non-distilled multi-step models, but with significantly fewer inference steps. While such models are attractive for applications due to the low inference cost and latency, fine-tuning them with a naive diffusion objective would result in degraded and blurry outputs. An intuitive alternative is to repeat the diffusion distillation process with a fine-tuned teacher model, which produces good results but is cumbersome and computationally intensive: the distillation training usually requires magnitude higher of training compute compared to fine-tuning for specific image styles. In this paper, we present an algorithm named pairwise sample optimization (PSO), which enables the direct fine-tuning of an arbitrary timestep-distilled diffusion model. PSO introduces additional reference images sampled from the current time-step distilled model, and increases the relative likelihood margin between the training images and reference images. This enables the model to retain its few-step generation ability, while allowing for fine-tuning of its output distribution. We also demonstrate that PSO is a generalized formulation which be flexible extended to both offline-sampled and online-sampled pairwise data, covering various popular objectives for diffusion model preference optimization. We evaluate PSO in both preference optimization and other fine-tuning tasks, including style transfer and concept customization. We show that PSO can directly adapt distilled models to human-preferred generation with both offline and online-generated pairwise preference image data. PSO also demonstrates effectiveness in style transfer and concept customization by directly tuning timestep-distilled diffusion models.", "keywords": [ "Diffusion Models;" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 4, 2, 4 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "fXkoROek1M", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Felipe Tobar", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Senior Lecturer" }, { "name": "Cristobal Alcazar", "gender": "Male", "institution": "Universidad de Chile", "country": "", "position": "MS student" }, { "name": "Roberto Joaqu\u00edn Rodolfo Barcel\u00f3 Soltmann", "gender": "Male", "institution": "Universidad de Chile", "country": "CL", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 27, "n_ref": 54, "n_ref_all": 73, "n_fig": 6, "n_tab": 3, "L_tab": 1210, "n_element_tab": 42, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 687, "n_element_tab_1": 20, "formula_len_all": 930, "formula_len_all_1": 930, "len_all": 100225, "len_all_1": 54996, "len_abs": 1077, "len_title": 129, "len_sents": 31386, "len_sents_1": 25472, "n_sents": 222, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1081, "title": "Avoiding mode collapse in diffusion models fine-tuned with reinforcement learning", "abs": "Fine-tuning foundation models via reinforcement learning (RL) has proven promising for aligning to downstream objectives. In the case of diffusion models (DMs), though RL training improves alignment from early timesteps, critical issues such as training instability and mode collapse arise. We address these drawbacks by exploiting the hierarchical nature of DMs: we train them dynamically at each epoch with a tailored RL method, allowing for continual evaluation and step-by-step refinement of the model performance (or alignment). Furthermore, we find that not every denoising step needs to be fine-tuned to align DMs to downstream tasks. Consequently, in addition to clipping, we regularise model parameters at distinct learning phases via a sliding-window approach. Our approach, termed Hierarchical Reward Fine-tuning (HRF), is validated on the Denoising Diffusion Policy Optimisation method, where we show that models trained with HRF achieve better preservation of diversity in downstream tasks, thus enhancing the fine-tuning robustness and at uncompromising mean rewards.", "keywords": [ "Generative Models", "Diffusion Models", "Reinforcement Learning", "Hierarchical RL" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "fXb9BbuyAD", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Glen Berseth", "gender": "Male", "institution": "University of Montreal, University of Montreal", "country": "", "position": "Assistant Professor" }, { "name": "Irina Rish", "gender": "Female", "institution": "University of Montreal", "country": "CA", "position": "Professor" }, { "name": "Matthew Riemer", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, University of Montreal", "country": "CA", "position": "PhD student" }, { "name": "Gopeshh Raaj Subbaraj", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, University of Montreal", "country": "CA", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 5, "n_ref_uni": 38, "n_ref": 71, "n_ref_all": 114, "n_fig": 25, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 16, "n_tab_1": 2, "L_tab_1": 227, "n_element_tab_1": 11, "formula_len_all": 695, "formula_len_all_1": 414, "len_all": 198900, "len_all_1": 71501, "len_abs": 1267, "len_title": 135, "len_sents": 61769, "len_sents_1": 35062, "n_sents": 409, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1295, "title": "Realtime Reinforcement Learning: Towards Rapid Asynchronous Deployment of Large Models", "abs": "Realtime environments change even as agents perform action inference and learning, thus requiring high interaction frequencies to effectively minimize long-term regret. However, recent advances in machine learning involve larger neural networks with longer inference times, raising questions about their applicability in realtime systems where quick reactions are crucial. We present an analysis of lower bounds on regret in realtime environments to show that minimizing long-term regret is generally impossible within the typical sequential interaction and learning paradigm, but often becomes possible when sufficient asynchronous compute is available. We propose novel algorithms for staggering asynchronous inference processes to ensure that actions are taken at consistent time intervals, and demonstrate that use of models with high action inference times is only constrained by the environment's effective stochasticity over the inference horizon, and not by action frequency. Our analysis shows that the number of inference and learning processes needed scales linearly with increasing inference times while enabling use of models that are multiple orders of magnitude larger than existing approaches when learning from a realtime simulation of Game Boy games such as Pokemon and Tetris.", "keywords": [ "Realtime Environments", "Asynchronous Algorithms", "Time Discretization", "Real World Deployment", "Deep Reinforcement Learning" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 2, 2 ] }, { "paper_id": "fXb7MgySp8", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Enyi Jiang", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Gagandeep Singh", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 47, "n_ref": 109, "n_ref_all": 136, "n_fig": 8, "n_tab": 8, "L_tab": 2229, "n_element_tab": 326, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1566, "n_element_tab_1": 256, "formula_len_all": 997, "formula_len_all_1": 729, "len_all": 150034, "len_all_1": 66853, "len_abs": 1089, "len_title": 111, "len_sents": 40898, "len_sents_1": 28065, "n_sents": 307, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1160, "title": "Towards Universal Certified Robustness with Multi-Norm Training", "abs": "Existing certified training methods can only train models to be robust against a certain perturbation type (e.g. $l_\\infty$ or $l_2$). However, an $l_\\infty$ certifiably robust model may not be certifiably robust against $l_2$ perturbation (and vice versa) and also has low robustness against other perturbations (e.g. geometric transformation). To this end, we propose the first multi-norm certified training framework \\textbf{CURE}, consisting of a new $l_2$ deterministic certified training defense and several multi-norm certified training methods, to attain better \\emph{union robustness} when training from scratch or fine-tuning a pre-trained certified model. Further, we devise bound alignment and connect natural training with certified training for better union robustness. Compared with SOTA certified training, \\textbf{CURE} improves union robustness up to $22.8\\%$ on MNIST, $23.9\\%$ on CIFAR-10, and $8.0\\%$ on TinyImagenet. Further, it leads to better generalization on a diverse set of challenging unseen geometric perturbations, up to $6.8\\%$ on CIFAR-10. Overall, our contributions pave a path towards \\textit{universal certified robustness}.", "keywords": [ "Certified Training", "Certified Robustness" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "fXJCqdUSVG", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Boyi Wei", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Luxi He", "gender": "Female", "institution": "Department of Computer Science, Princeton University", "country": "US", "position": "PhD student" }, { "name": "Matthew Jagielski", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Milad Nasr", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Nicholas Carlini", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Peter Henderson", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Prateek Mittal", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "Full Professor" }, { "name": "Tinghao Xie", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Xiangyu Qi", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Yangsibo Huang", "gender": "Female", "institution": "Google", "country": "US", "position": "Research Scientist" } ], "n_formula": 7, "n_formula_1": 0, "n_ref_uni": 51, "n_ref": 140, "n_ref_all": 208, "n_fig": 16, "n_tab": 9, "L_tab": 1251, "n_element_tab": 97, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 364, "formula_len_all_1": 0, "len_all": 230413, "len_all_1": 77567, "len_abs": 1155, "len_title": 111, "len_sents": 82229, "len_sents_1": 39203, "n_sents": 603, "n_sents_1": 279, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1165, "title": "On Evaluating the Durability of Safeguards for Open-Weight LLMs", "abs": "Many stakeholders---from model developers to policymakers---seek to minimize the risks of large language models (LLMs). Key to this goal is whether technical safeguards can impede the misuse of LLMs, even when models are customizable via fine-tuning or when model weights are openly available. Several recent studies have proposed methods to produce durable LLM safeguards for open-weight LLMs that can withstand adversarial modifications of the model's weights via fine-tuning. This holds the promise of raising adversaries' costs even under strong threat models where adversaries can directly fine-tune parameters. However, we caution against over-reliance on such methods in their current state. Through several case studies, we demonstrate that even the evaluation of these defenses is exceedingly difficult and can easily mislead audiences into thinking that safeguards are more durable than they really are. We draw lessons from the failure modes that we identify and suggest that future research carefully cabin claims to more constrained, well-defined, and rigorously examined threat models, which can provide useful and candid assessments to stakeholders.", "keywords": [ "AI Safety", "Fine-tuning Attacks", "Open-weight LLMs", "Adaptive Attacks" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 4, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "fWx1CKgPCc", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Ke Jiang", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" }, { "name": "Xiaoyang Tan", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "", "position": "Full Professor" } ], "n_formula": 74, "n_formula_1": 20, "n_ref_uni": 26, "n_ref": 83, "n_ref_all": 114, "n_fig": 6, "n_tab": 7, "L_tab": 1916, "n_element_tab": 233, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1263, "n_element_tab_1": 124, "formula_len_all": 7834, "formula_len_all_1": 1630, "len_all": 160845, "len_all_1": 67104, "len_abs": 1029, "len_title": 128, "len_sents": 52723, "len_sents_1": 30417, "n_sents": 427, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1034, "title": "Towards Reliable Offline Reinforcement Learning via Lyapunov Uncertainty Control", "abs": "Learning trustworthy and reliable offline policies presents significant challenges due to the inherent uncertainty in pre-collected datasets. In this paper, we propose a novel offline reinforcement learning method to tackle this issue. Inspired by the concepts of Lyapunov stability and control-invariant sets from control theory, the central idea is to introduce a restricted state space for the agent to operate within. This approach allows the learned models to exhibit reduced Bellman uncertainty and make reliable decisions. To achieve this, we regulate the expected Bellman uncertainty associated with the new policy, ensuring that its growth trend in subsequent states remains within acceptable limits. The resulting method, termed Lyapunov Uncertainty Control (LUC), is shown to guarantee that the agent remains within a low-uncertainty state enclosure throughout its entire trajectory. \nFurthermore, we perform extensive theoretical and experimental analysis to showcase the effectiveness and feasibility of the proposed LUC.", "keywords": [ "Offline reinforcement learning; Reliable offline reinforcement learning; Lyapunov Uncertainty Control;" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 2, 1 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "fWXYD0ZCdd", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Anushri Arora", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Jonathan W. Pillow", "gender": "Not Specified", "institution": "Princeton University", "country": "US", "position": "Professor" } ], "n_formula": 15, "n_formula_1": 16, "n_ref_uni": 40, "n_ref": 78, "n_ref_all": 91, "n_fig": 11, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 760, "formula_len_all_1": 761, "len_all": 110155, "len_all_1": 68877, "len_abs": 1453, "len_title": 96, "len_sents": 36741, "len_sents_1": 34080, "n_sents": 249, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1464, "title": "A New Look at Low-Rank Recurrent Neural Networks", "abs": "Low-rank recurrent neural networks (RNNs) have recently gained prominence as a framework for understanding how neural systems solve complex cognitive tasks. However, fitting and interpreting these networks remains an important open problem.\nHere we address this challenge using a perspective from the ``neural engineering framework'', which shows how to embed an arbitrary ordinary differential equation (ODE) into a low-rank RNN using least-squares regression. Under this perspective, individual neurons in a low-rank RNN provide nonlinear basis functions for representing an ODE of interest. This clarifies limits on the expressivity of low-rank RNNs, such as the fact that with a $\\tanh$ non-linearity they can only capture odd-symmetric functions in the absence of per neuron inputs or biases. Building on this framework, we propose a method for finding the smallest low-rank RNN to implement a given dynamical system using a variant of orthogonal matching pursuit. We also show how to use regression-based fitting to obtain low-rank RNNs with time-varying dynamics. This allows for the rapid training of vastly different dynamical systems that nevertheless produce a given time-varying trajectory. Finally, we highlight the usefulness of our framework by comparing to RNNs trained using backprop-through-time on neuroscience-inspired tasks, showing that our method achieves faster and more accurate learning with smaller networks than gradient-based training.", "keywords": [ "low-rank rnn", "computational neuroscience", "dynamical systems", "neural dynamics" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "fWRBheSJth", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Rui Zhang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Ryo Kamoi", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Sarkar Snigdha Sarathi Das", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Yusen Zhang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Bo Pang", "gender": "Male", "institution": "SalesForce.com", "country": "", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 34, "n_ref": 84, "n_ref_all": 121, "n_fig": 6, "n_tab": 14, "L_tab": 16154, "n_element_tab": 357, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 959, "n_element_tab_1": 89, "formula_len_all": 705, "formula_len_all_1": 383, "len_all": 165961, "len_all_1": 58763, "len_abs": 1692, "len_title": 137, "len_sents": 49280, "len_sents_1": 27354, "n_sents": 394, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1795, "title": "GReaTer: Gradients Over Reasoning Makes Smaller Language Models Strong Prompt Optimizers", "abs": "The effectiveness of large language models (LLMs) is closely tied to the design of prompts, making prompt optimization essential for enhancing their performance across a wide range of tasks. Although recent advancements have focused on automating prompt engineering, many existing approaches rely exclusively on textual feedback, refining prompts based solely on inference errors identified by large, computationally expensive LLMs. Unfortunately, smaller models struggle to generate high-quality feedback, resulting in complete dependence on large LLM judgment. Moreover, these methods fail to leverage more direct and finer-grained information, such as gradients, due to operating purely in text space. To this end, we introduce, we introduce *GReaTer*, a novel prompt optimization technique that directly incorporates *gradient information over task-specific reasoning*. By utilizing task loss gradients, *GReaTer* enables self-optimization of prompts for smaller, lightweight language models (LM) without the need for costly closed-source LLMs, while maintaining reasonable prompt structures. This allows high-performance prompt optimization without dependence on massive LLMs, closing the gap between smaller models and the sophisticated reasoning often needed for prompt refinement. Extensive evaluations across diverse tasks demonstrate that \\ours consistently outperforms previous methods, even those reliant on powerful LLMs. Additionally, *GReaTer*-optimized prompts frequently exhibit better transferability and, in some cases, boost task performance to levels comparable to or surpassing those achieved by larger language models, highlighting the effectiveness of *\"gradient over reasoning\"*-based prompt optimization. Full source code of *GReaTer* will be available upon acceptance.", "keywords": [ "Large Language Model", "Prompt Optimization" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "fWNHKHh0Yn", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Dazhou Li", "gender": "Male", "institution": "Shenyang University of Chemical Technology", "country": "CN", "position": "Lecturer" }, { "name": "Guiyuan Fu", "gender": "Male", "institution": "Shenyang University of Chemical Technology", "country": "", "position": "Researcher" }, { "name": "Rui Yu", "gender": "Male", "institution": "University of Louisville", "country": "US", "position": "Assistant Professor" }, { "name": "Ying Li", "gender": "Female", "institution": "Shenyang University of Chemical Technology", "country": "CN", "position": "MS student" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Shenyang University of Chemical Technology", "country": "CN", "position": "Lecturer" }, { "name": "Zhipeng Yao", "gender": "Male", "institution": "Shenyang University of Chemical Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 122, "n_formula_1": 18, "n_ref_uni": 53, "n_ref": 97, "n_ref_all": 118, "n_fig": 11, "n_tab": 10, "L_tab": 1557, "n_element_tab": 346, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 397, "n_element_tab_1": 46, "formula_len_all": 11582, "formula_len_all_1": 1497, "len_all": 182417, "len_all_1": 57586, "len_abs": 829, "len_title": 100, "len_sents": 46421, "len_sents_1": 24863, "n_sents": 458, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 756, "title": "SGDF: A Method for Reducing Variance in Stochastic Gradient Descent via Filter Estimation", "abs": "In deep learning, stochastic gradient descent (SGD) and its momentum-based variants are widely used for optimization, but they typically suffer from slow convergence. Conversely, existing adaptive learning rate optimizers speed up convergence but often compromise generalization. To resolve this issue, we propose a novel optimization method designed to accelerate SGD's convergence without sacrificing generalization. Our approach reduces the variance of the historical gradient, improves first-order moment estimation of SGD by applying Wiener filter theory, and introduces a time-varying adaptive gain. Empirical results demonstrate that SGDF (SGD with Filter) effectively balances convergence and generalization compared to state-of-the-art optimizers.", "keywords": [ "Stochastic Gradient Descent", "Gradient Estimation", "Generalization" ], "rating_list": [ 5, 5, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 1, 3 ], "presentation_list": [ 2, 3, 2, 1, 3 ], "contribution_list": [ 3, 3, 2, 2, 1 ], "confidence_list": [ 4, 2, 5, 3, 4 ] }, { "paper_id": "fVgUXaesSS", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Jun Sun", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Full Professor" }, { "name": "Ruihan Zhang", "gender": "Female", "institution": "Singapore Management University", "country": "", "position": "PhD student" } ], "n_formula": 76, "n_formula_1": 25, "n_ref_uni": 49, "n_ref": 114, "n_ref_all": 153, "n_fig": 10, "n_tab": 2, "L_tab": 438, "n_element_tab": 48, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 438, "n_element_tab_1": 48, "formula_len_all": 6997, "formula_len_all_1": 2464, "len_all": 205906, "len_all_1": 71039, "len_abs": 1448, "len_title": 88, "len_sents": 57117, "len_sents_1": 30102, "n_sents": 533, "n_sents_1": 283, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1478, "title": "Are Probabilistic Robust Accuracy Bounded", "abs": "Adversarial samples pose a security threat to many critical systems built on neural networks. It has recently been proven that achieving deterministic robustness (*i.e.*, complete elimination of adversarial samples) always comes at an unbearable cost to accuracy. As a result, probabilistic robustness (where the probability of retaining the same label within a vicinity is at least $1 - \\kappa$) has been proposed as a promising compromise. However, existing training methods for probabilistic robustness still experience non-trivial accuracy loss. It remains an open question what the upper limit on accuracy is when optimizing for probabilistic robustness, and whether there is a specific relationship between $\\kappa$ and this potential bound. This work studies these problems from a Bayes error perspective. We find that while Bayes uncertainty does affect probabilistic robustness, its impact is smaller than that on deterministic robustness. This reduced Bayes uncertainty allows a higher upper bound on probabilistic robust accuracy than that on deterministic robust accuracy. Further, we show that voting within the vicinity always improves probabilistic robust accuracy and the upper bound of probabilistic robust accuracy monotonically increases as $\\kappa$ grows. Our empirical findings also align with our results. This study thus presents a theoretical argument supporting probabilistic robustness as the appropriate target for achieving neural network robustness.", "keywords": [ "bayes error", "Probabilistic Robustness" ], "rating_list": [ 6, 8, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 2, 3, 3, 5 ] }, { "paper_id": "fV0t65OBUu", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Andi Zhang", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "Postdoc" }, { "name": "Mingtian Zhang", "gender": "Male", "institution": "University College London", "country": "", "position": "PhD student" }, { "name": "Yingzhen Li", "gender": "Female", "institution": "Imperial College London", "country": "", "position": "Associate Professor" }, { "name": "Zijing Ou", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "David Barber", "gender": "Male", "institution": "University College London", "country": "", "position": "Associate Professor" }, { "name": "Tim Xiao", "gender": "Male", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "PhD student" } ], "n_formula": 54, "n_formula_1": 35, "n_ref_uni": 60, "n_ref": 140, "n_ref_all": 212, "n_fig": 18, "n_tab": 9, "L_tab": 4842, "n_element_tab": 316, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 2486, "n_element_tab_1": 122, "formula_len_all": 4809, "formula_len_all_1": 2185, "len_all": 183697, "len_all_1": 71109, "len_abs": 870, "len_title": 130, "len_sents": 47803, "len_sents_1": 29292, "n_sents": 382, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 82, "L_abs": 875, "title": "Improving Probabilistic Diffusion Models With Optimal Diagonal Covariance Matching", "abs": "The probabilistic diffusion model has become highly effective across various domains. Typically, sampling from a diffusion model involves using a denoising distribution characterized by a Gaussian with a learned mean and either fixed or learned covariances. In this paper, we leverage the recently proposed covariance moment matching technique and introduce a novel method for learning the diagonal covariances. Unlike traditional data-driven covariance approximation approaches, our method involves directly regressing the optimal analytic covariance using a new, unbiased objective named Optimal Covariance Matching (OCM). This approach can significantly reduce the approximation error in covariance prediction. We demonstrate how our method can substantially enhance the sampling efficiency, recall rate and likelihood of both diffusion models and latent diffusion models.", "keywords": [ "Diffusion Model", "Generative Model", "Probalistic Modelling" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 4, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "fUZ6rCUZUx", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Claudius Schn\u00f6rr", "gender": "Male", "institution": "University od Applied Sciences Munich", "country": "DE", "position": "Full Professor" }, { "name": "Julian Wustl", "gender": "unknown", "institution": "Universit\u00e4t Augsburg", "country": "DE", "position": "PhD student" }, { "name": "Philipp Haid", "gender": "unknown", "institution": "Universit\u00e4t Augsburg", "country": "DE", "position": "PhD student" }, { "name": "Yarema Okhrin", "gender": "Male", "institution": "University of Augsburg", "country": "DE", "position": "Full Professor" } ], "n_formula": 29, "n_formula_1": 15, "n_ref_uni": 51, "n_ref": 88, "n_ref_all": 124, "n_fig": 9, "n_tab": 11, "L_tab": 6705, "n_element_tab": 879, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2226, "formula_len_all_1": 1232, "len_all": 203145, "len_all_1": 65294, "len_abs": 1300, "len_title": 131, "len_sents": 64340, "len_sents_1": 32503, "n_sents": 516, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1305, "title": "Decoupling Dependency Structures: Sklar\u2019s theorem for explainable outlier detection", "abs": "Recent advances in outlier detection have been primarily driven by deep learning models, which, while powerful, have substantial drawbacks in terms of explainability. This is particularly relevant in fields that demand detailed reasoning and understanding of why observations are classified as outliers. To close the gap between state-of-the-art performance and enhanced explainability, we propose Vine Copula-Based Outlier Detection (VC-BOD). We utilize Sklar\u2019s theorem in conjunction with vine copulas and univariate kernel density estimators to decouple marginal distributions and their dependency structure for outlier detection. Our model uses a closed-form equation for the outlier score, which allows for detailed explainability and feature attribution. VC-BOD employs a traceable criterion to determine whether a new observation is an outlier, while also identifying the specific features responsible for this classification. The proposed model further distinguishes whether these features deviate from their own distributions or from interactions with other features. Our empirical evaluations demonstrate that VC-BOD outperforms most benchmarked classical models and several deep learning approaches in terms of average rank performance while proving competitive with the best-performing models.", "keywords": [ "Outlier Detection", "Explainability", "Copula Modeling" ], "rating_list": [ 5, 5, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "fUHoUXGUZp", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Giulia M. Bovolenta", "gender": "Female", "institution": "Istituto Italiano di Tecnologia", "country": "IT", "position": "Postdoc" }, { "name": "Michele Ciavotta", "gender": "Male", "institution": "University of Milan - Bicocca", "country": "IT", "position": "Associate Professor" }, { "name": "Michele Parrinello", "gender": "Male", "institution": "Istituto Italiano di Tecnologia, Genova", "country": "", "position": "Principal Researcher" }, { "name": "Pietro Morerio", "gender": "Male", "institution": "Istituto Italiano di Tecnologia", "country": "IT", "position": "Researcher" }, { "name": "Stefano Coniglio", "gender": "unknown", "institution": "University of Bergamo", "country": "IT", "position": "Researcher" }, { "name": "Stefano Fiorini", "gender": "Male", "institution": "Istituto Italiano di Tecnologia", "country": "IT", "position": "Postdoc" }, { "name": "Alessio Del Bue", "gender": "Male", "institution": "Istituto Italiano di Tecnologia", "country": "IT", "position": "Senior Researcher" } ], "n_formula": 23, "n_formula_1": 21, "n_ref_uni": 31, "n_ref": 52, "n_ref_all": 72, "n_fig": 6, "n_tab": 2, "L_tab": 851, "n_element_tab": 5, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 816, "n_element_tab_1": 3, "formula_len_all": 1630, "formula_len_all_1": 1605, "len_all": 119193, "len_all_1": 73836, "len_abs": 1644, "len_title": 132, "len_sents": 35415, "len_sents_1": 32214, "n_sents": 234, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1679, "title": "DLGNet: Hyperedge Classification through Directed Line Graphs for Chemical Reactions", "abs": "Graphs and hypergraphs provide powerful abstractions for modeling interactions among a set of entities of interest and have been attracting a growing interest in the literature thanks to many successful applications in several fields. In particular, they are rapidly expanding in domains such as chemistry and biology, especially in the areas of drug discovery and molecule generation. One of the areas witnessing the fasted growth is the chemical reactions field, where chemical reactions can be naturally encoded as directed hyperedges of a hypergraph. In this paper, we address the chemical reaction classification problem by introducing the notation of a Directed Line Graph (DGL) associated with a given directed hypergraph. On top of it, we build the Directed Line Graph Network (DLGNet), the first spectral-based Graph Neural Network (GNN) expressly designed to operate on a hypergraph via its DLG transformation. The foundation of DLGNet is a novel Hermitian matrix, the Directed Line Graph Laplacian $\\mathbb{\\vec L}_N$, which compactly encodes the directionality of the interactions taking place within the directed hyperedges of the hypergraph thanks to the DLG representation. $\\mathbb{\\vec L}_N$ enjoys many desirable properties, including admitting an eigenvalue decomposition and being positive semidefinite, which make it well-suited for its adoption within a spectral-based GNN. Through extensive experiments on chemical reaction datasets, we show that DGLNet significantly outperforms the existing approaches, achieving on a collection of real-world datasets an average relative-percentage-difference improvement of 3.27\\%, with a maximum improvement of 5.28\\%.", "keywords": [ "Directed Line Graph", "Directed Line Graph Laplacian", "Hyperedge Classification", "Chemical reaction classification" ], "rating_list": [ 8, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "fU8H4lzkIm", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Bocheng Zeng", "gender": "unknown", "institution": "Renmin University of China", "country": "", "position": "MS student" }, { "name": "Hao Sun", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Hongsheng Liu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Mengtao Yan", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Qi Wang", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Ruizhi Chengze", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yang Liu", "gender": "Female", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Yi Zhang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zidong Wang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 12, "n_ref_uni": 45, "n_ref": 80, "n_ref_all": 101, "n_fig": 6, "n_tab": 9, "L_tab": 1747, "n_element_tab": 214, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 234, "n_element_tab_1": 36, "formula_len_all": 1104, "formula_len_all_1": 817, "len_all": 176195, "len_all_1": 57067, "len_abs": 1675, "len_title": 134, "len_sents": 47660, "len_sents_1": 26315, "n_sents": 340, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 85, "L_abs": 1591, "title": "PhyMPGN: Physics-encoded Message Passing Graph Network for spatiotemporal PDE systems", "abs": "Solving partial differential equations (PDEs) serves as a cornerstone for modeling complex dynamical systems. Recent progresses have demonstrated grand benefits of data-driven neural-based models for predicting spatiotemporal dynamics (e.g., tremendous speedup gain compared with classical numerical methods). However, most existing neural models rely on rich training data, have limited extrapolation and generalization abilities, and suffer to produce precise or reliable physical prediction under intricate conditions (e.g., irregular mesh or geometry, complex boundary conditions, diverse PDE parameters, etc.). To this end, we propose a new graph learning approach, namely, Physics-encoded Message Passing Graph Network (PhyMPGN), to model spatiotemporal PDE systems on irregular meshes given small training datasets. Specifically, we incorporate a GNN into a numerical integrator to approximate the temporal marching of spatiotemporal dynamics for a given PDE system. Considering that many physical phenomena are governed by diffusion processes, we further design a learnable Laplace block, which encodes the discrete Laplace-Beltrami operator, to aid and guide the GNN learning in a physically feasible solution space. A boundary condition padding strategy is also designed to improve the model convergence and accuracy. Extensive experiments demonstrate that PhyMPGN is capable of accurately predicting various types of spatiotemporal dynamics on coarse unstructured meshes, consistently achieves the state-of-the-art results, and outperforms other baselines with considerable gains.", "keywords": [ "Physics-encoded; Spatiotemporal PDEs; Graph Network; Deep Learning;" ], "rating_list": [ 8, 10, 6, 8, 8 ], "soundness_list": [ 3, 4, 3, 3, 3 ], "presentation_list": [ 4, 4, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 3, 3 ], "confidence_list": [ 4, 5, 4, 5, 4 ] }, { "paper_id": "fTdhM7q1o2", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Jinsong Liu", "gender": "Male", "institution": "Shanghai University of Finance and Economics", "country": "", "position": "PhD student" }, { "name": "Ruihao Zhu", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Assistant Professor" }, { "name": "Dongdong Ge", "gender": "Male", "institution": "shanghai university of finance and economics", "country": "", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 22, "n_ref_uni": 30, "n_ref": 56, "n_ref_all": 70, "n_fig": 2, "n_tab": 4, "L_tab": 481, "n_element_tab": 31, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 552, "n_element_tab_1": 39, "formula_len_all": 2042, "formula_len_all_1": 2042, "len_all": 91571, "len_all_1": 59847, "len_abs": 1423, "len_title": 89, "len_sents": 25107, "len_sents_1": 25615, "n_sents": 196, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1433, "title": "Reward Learning From Preference With Ties", "abs": "Reward learning plays a pivotal role in Reinforcement Learning from Human Feedback (RLHF), ensuring the alignment of language models. The Bradley-Terry (BT) model stands as the prevalent choice for capturing human preferences from datasets containing pairs of chosen and rejected responses. In preference modeling, the focus is not on absolute values but rather on the reward difference between chosen and rejected responses, referred to as preference strength. Thus, precise evaluation of preference strength holds paramount importance in preference modeling. However, an easily overlooked factor significantly affecting preference strength measurement is that human attitudes towards two responses may not solely indicate a preference for one over the other and ties are also a common occurrence. To address this, we propose the adoption of the generalized Bradley-Terry model -- the Bradley-Terry model with ties (BTT) -- to accommodate tied preferences, thus leveraging additional information. We prove that even with the access to the true distributions of prompt and response, disregarding ties can lead to a notable bias in preference strength measurement. Comprehensive experiments further validate the advantages of incorporating ties in preference modeling. Notably, fine-tuning with BTT significantly outperforms fine-tuning with BT on synthetic preference datasets with ties, labeled by state-of-the-art open-source LLMs.", "keywords": [ "reward learning", "RLHF", "preference modeling", "language model" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "fSxiromxAq", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "MADHU KRISHNA R", "gender": "Male", "institution": "Tredence Inc", "country": "IN", "position": "Researcher" }, { "name": "Ravindra Patil", "gender": "Male", "institution": "Maastricht University", "country": "", "position": "PhD student" }, { "name": "kalyani pakhale", "gender": "Female", "institution": "Indian Institute of Technology, Madras, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "Researcher" } ], "n_formula": 28, "n_formula_1": 28, "n_ref_uni": 15, "n_ref": 22, "n_ref_all": 24, "n_fig": 0, "n_tab": 2, "L_tab": 848, "n_element_tab": 56, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 848, "n_element_tab_1": 56, "formula_len_all": 1577, "formula_len_all_1": 1577, "len_all": 66357, "len_all_1": 50745, "len_abs": 1278, "len_title": 141, "len_sents": 21802, "len_sents_1": 21802, "n_sents": 152, "n_sents_1": 153, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1286, "title": "Sparse Causal Model: A Novel Approach for Causal Discovery and Attributions on Sparse Dataset", "abs": "This paper introduces a novel approach to tackle the challenges of causal modeling and attribution in sparse and non-continuous data with limited feature knowledge. Traditional methods rely on static inputs and lack adaptability to dynamic changes in causal relationships, resulting in a limited understanding and goodness-of-fit. We introduce a unique causal discovery framework on real-world sparse datasets to address this challenge. We leverage a Directed Acyclic Graph (DAG) by discovering causal relationships between the variables by identifying confounder-treatment pairs that make the variable selection process robust and efficient. We propose a three-stage causal model that uses multiple distinct regressors such as likelihood-based, tree-based, and Generalized Additive Models (GAMs). Furthermore, we introduce a Model Score by including the sensitivity analysis involving random shuffling confounders and treatments to select the best optimal model. We implement a partial dependency approach to understand the attribution of variables, contributing by adding a 53% increase in the R2 score compared to traditional methods. This research underscores the limitations of conventional approaches in addressing real-world challenges to address practical scenarios effectively.", "keywords": [ "Causality", "Causal Inference" ], "rating_list": [ 3, 5, 1, 3 ], "soundness_list": [ 1, 2, 1, 3 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 1, 3, 1, 1 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "fSqzHzyVZU", "primary_area": "applications to neuroscience & cognitive science", "n_author": 8, "author_info_list": [ { "name": "Biao Yan", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Chen Qingyuan", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Dongqi Han", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Dongsheng Li", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Jiang Chengyong", "gender": "unknown", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Jiayi Zhang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Junwen Luo", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yansen Wang", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 22, "n_ref": 29, "n_ref_all": 45, "n_fig": 8, "n_tab": 1, "L_tab": 640, "n_element_tab": 29, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 640, "n_element_tab_1": 29, "formula_len_all": 181, "formula_len_all_1": 181, "len_all": 100218, "len_all_1": 54735, "len_abs": 1606, "len_title": 129, "len_sents": 36807, "len_sents_1": 27252, "n_sents": 244, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1611, "title": "The VEP Booster: A Closed-Loop AI System for Visual EEG Biomarker Auto-generation", "abs": "The effectiveness of Visual Brain-Machine Interfaces (BMIs) is significantly dependent on the accurate detection and interpretation of electroencephalography (EEG) biomarkers, which frequently exhibit variability due to physiological changes and environmental disturbances over time. Traditional EEG signal enhancement strategies largely concentrate on signal processing techniques such as feature extraction and filtering; however, these approaches often do not adequately address the inherent sources of variability that affect biomarker stability over time. To surmount these challenges, we have developed the Visual Evoked Potential Booster (VEP Booster), a novel closed-loop artificial intelligence framework designed to produce reliable and stable EEG biomarkers under visual stimulation protocols. Our system utilizes a Deep Convolutional Generative Adversarial Network (DCGAN) to refine stimulus images based on real-time feedback from human EEG signals, thereby creating visual stimuli that are specifically tailored to the characteristic preferences of neurons in the primary visual cortex. We evaluated the efficacy of this system through the implementation of steady-state visual evoked potential (SSVEP) protocols in nine human subjects. In our evaluations, both the SSVEP biomarker amplitude and the single-trial SSVEP binary classification experiments, encompassing intra- and inter-temporal analyses, exhibited statistically significant enhancements when employing the VEP Booster. These encouraging outcomes underscore the potential for broad applications in clinical and technological domains.", "keywords": [ "Visual Evoked Potential", "Brain-machine Interface", "EEGs", "Closed-loop AI", "Generative AI" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 4, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "fSbPwHjdDG", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Marcus Hutter", "gender": "unknown", "institution": "DeepMind", "country": "GB", "position": "Researcher" }, { "name": "David John Quarel", "gender": "unknown", "institution": "Australian National University", "country": "AU", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 15, "n_ref_uni": 22, "n_ref": 34, "n_ref_all": 77, "n_fig": 24, "n_tab": 1, "L_tab": 29, "n_element_tab": 3, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1042, "formula_len_all_1": 704, "len_all": 118105, "len_all_1": 61449, "len_abs": 794, "len_title": 145, "len_sents": 36092, "len_sents_1": 21440, "n_sents": 260, "n_sents_1": 134, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 827, "title": "Llamas (mostly) think in English: On Causal Interventions in the Latent Language of Transformers", "abs": "Previous research on the Llama-2 family of Large Language Models (LLMs) suggested a\n correlation indicating the use of English as a intermediary language within\n these models for tasks in non-English languages. We improve on this by demonstrating a causal relationship. By\n intervening on the intermediate layers during a forward pass, we show\n that projecting out the activations onto a subspace corresponding to the correct\n prediction in English impairs the model's ability to make\n correct predictions on non-English translation tasks. Projecting onto an unrelated \n English subspace, or a related subspace in a non-English language, has little effect,\n demonstrating that\n this family of models store concepts that have a high similarity to the corresponding\n concept in English in the residual stream.", "keywords": [ "mechanistic interpretability", "large language models", "transformers", "residual stream" ], "rating_list": [ 3, 1, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 1, 1 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "fSB95BWiBQ", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Andrew Luo", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Leila Wehbe", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Margaret Marie Henderson", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Rushikesh Zawar", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Shaurya Rajat Dewan", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Michael Tarr", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 92, "n_ref": 108, "n_ref_all": 128, "n_fig": 14, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 513, "n_element_tab_1": 54, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 167646, "len_all_1": 55727, "len_abs": 1507, "len_title": 158, "len_sents": 32756, "len_sents_1": 26383, "n_sents": 255, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 1511, "title": "StableSemantics: A Synthetic Language-Vision Dataset of Semantic Representations in Naturalistic Images", "abs": "Understanding dense visual semantics remains a fundamental challenge in computer vision, as semantically similar objects can exhibit drastically different visual appearances. Recent advancements in generative text-to-image frameworks have led to models that implicitly capture natural scene statistics. These models learn to model complex relationships between objects, lighting, and other visual factors, enabling the generation of detailed and contextually rich images from text captions. To advance visual semantic understanding and develop more robust and interpretable vision models, we present StableSemantics, a large-scale dataset composed of 224 thousand human-curated prompts, processed natural language captions, over 2 million synthetic images, and 10 million attention maps. The dataset provides fine-grained semantic attributions at the noun-chunk level, leverages human-generated prompts that correspond to visually interesting stable diffusion generations, and provides 10 generations per phrase, with cross-attention maps corresponding to noun chunks for each image. We explore the semantic distribution of generated images, examine the distribution of objects within images, and benchmark captioning and open vocabulary segmentation methods on our data. As the first diffusion dataset to include dense attention attributions, we expect StableSemantics to catalyze advances in visual semantic understanding and provide a foundation for developing more sophisticated and effective visual models.", "keywords": [ "diffusion", "interpretation", "attention", "attribution", "semantics", "synthetic data", "dataset" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 4, 1 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "fRpAUgKJhT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Gang Niu", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Senior Research Scientist (tenured)" }, { "name": "Jianzhong Qi", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Associate Professor" }, { "name": "Masashi Sugiyama", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Director" }, { "name": "Ruijiang Dong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" }, { "name": "Zesheng Ye", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Postdoc" } ], "n_formula": 45, "n_formula_1": 19, "n_ref_uni": 30, "n_ref": 78, "n_ref_all": 126, "n_fig": 3, "n_tab": 5, "L_tab": 1673, "n_element_tab": 214, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4008, "formula_len_all_1": 1493, "len_all": 176041, "len_all_1": 62452, "len_abs": 1613, "len_title": 125, "len_sents": 61225, "len_sents_1": 28630, "n_sents": 487, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1619, "title": "CARPRT: Class-Aware Prompt Reweighting for Pre-Trained Vision-Language Models", "abs": "When using a pre-trained vision-language model (VLM) to classify an image, we often need to use the pre-trained VLM to compute a similarity score between the image and texts containing a semantic label, e.g., \u201ca photo of a cat\u201d, where \u201ca photo of a\u201d is called a prompt and \u201ccat\u201d is the semantic label (a.k.a. a class in classification tasks). The existing studies have shown that the selection of prompts can significantly affect the scoring scheme between a given image and a semantic label, and they proposed a new score via using a weighting vector to reassemble scores regarding different prompts. However, these studies assume that all classes should share the same weighting vector. In this paper, we first empirically show that the existing approach is sub-optimal. We subsequently revisit the existing reweighting strategy from a probabilistic view and find an implicit assumption in prior work: the conditional independence of classes and weights, which often does not hold in practice. To cope with this problem, we propose class-aware prompt reweighting (CARPRT), a strategy designed to adjust the weighting vector for each class. CARPRT calculates the relevance scores for prompt-class pairs with respect to all images, and identifies the maximum score for each prompt-class pair. These maximum scores are then averaged across prompts for each class to estimate the class-specific weighting vectors, ensuring that prompts are optimally reweighted based on class-specific information. Our experiments demonstrate that CARPRT outperforms the existing reweighting strategy under the image classification tasks.", "keywords": [ "Prompt Weighting", "Vision-language Models" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "fRmfDqZ2yq", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Ben He", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Fan Chen", "gender": "Male", "institution": "Bytedance", "country": "US", "position": "Researcher" }, { "name": "Le Sun", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Lexin Tang", "gender": "Female", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Longyin Wen", "gender": "Male", "institution": "Bytedance Inc.", "country": "US", "position": "Research Manager" }, { "name": "Xinyao Wang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yaojie Shen", "gender": "unknown", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Ying Zhou", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yulei Niu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 20, "n_formula_1": 9, "n_ref_uni": 55, "n_ref": 104, "n_ref_all": 128, "n_fig": 6, "n_tab": 7, "L_tab": 3392, "n_element_tab": 376, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 1584, "n_element_tab_1": 325, "formula_len_all": 1239, "formula_len_all_1": 397, "len_all": 179984, "len_all_1": 65180, "len_abs": 1386, "len_title": 125, "len_sents": 47053, "len_sents_1": 29255, "n_sents": 344, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1391, "title": "DiffLM: Controllable Synthetic Data Generation via Diffusion Language Models", "abs": "Recent advancements in large language models (LLMs) have significantly enhanced their knowledge and generative capabilities, leading to a surge of interest in leveraging LLMs for high-quality data synthesis. However, synthetic data generation via prompting LLMs remains challenging due to LLMs' limited understanding of target data distributions and the complexity of prompt engineering, especially for structured formatted data. To address these issues, we introduce DiffLM, a controllable data synthesis framework based on variational autoencoder (VAE), which further (1) leverages diffusion models to reserve more information of original distribution and format structure in the learned latent distribution and (2) decouples the learning of target distribution knowledge from the LLM's generative objectives via a plug-and-play latent feature injection module. As we observed significant discrepancies between the VAE's latent representations and the real data distribution, the latent diffusion module is introduced into our framework to learn a fully expressive latent distribution. Evaluations on seven real-world datasets with structured formatted data (i.e., Tabular, Code and Tool data) demonstrate that DiffLM generates high-quality data, with performance on downstream tasks surpassing that of real data by 2\\%\u20137\\% in certain cases. Data and code will be released upon acceptance.", "keywords": [ "synthetic data generation", "diffusion models", "language model" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 5, 5 ] }, { "paper_id": "fRaK0cG9L8", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Alexander Hepburn", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "Postdoc" }, { "name": "Javier Portilla", "gender": "unknown", "institution": "Spanish National Research Council", "country": "ES", "position": "Principal Researcher" }, { "name": "Raul Santos-Rodriguez", "gender": "unknown", "institution": "University of Bristol", "country": "GB", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 20, "n_ref": 52, "n_ref_all": 61, "n_fig": 3, "n_tab": 4, "L_tab": 738, "n_element_tab": 70, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 439, "n_element_tab_1": 45, "formula_len_all": 843, "formula_len_all_1": 835, "len_all": 84533, "len_all_1": 60503, "len_abs": 1581, "len_title": 157, "len_sents": 33916, "len_sents_1": 29763, "n_sents": 234, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 110, "L_abs": 1594, "title": "Evaluating Perceptual Distances Models by Fitting Binomial Distributions to Two-Alternative Forced Choice Data", "abs": "The two-alternative forced choice (2AFC) experimental method is popular in the visual perception literature, where practitioners aim to understand how human observers perceive distances within triplets made of a reference image and two distorted versions. In the past, this had been conducted in controlled environments, with triplets sharing images, so it was possible to rank the perceived quality. This ranking would then be used to evaluate perceptual distance models against the experimental data. Recently, crowd-sourced perceptual datasets have emerged, with no images shared between triplets, making ranking infeasible. Evaluating perceptual distance models using this data reduces the judgements on a triplet to a binary decision, namely, whether the distance model agrees with the human decision - which is suboptimal and prone to misleading conclusions. Instead, we statistically model the underlying decision-making process during 2AFC experiments using a binomial distribution. Having enough empirical data, we estimate a smooth and consistent distribution of the judgements on the reference-distorted distance plane, according to each distance model. By applying maximum likelihood, we estimate the parameter of the local binomial distribution, and a global measurement of the expected log-likelihood of the measured responses. We calculate meaningful and well-founded metrics for the distance model, beyond the mere prediction accuracy as percentage agreement, even with variable numbers of judgements per triplet -- key advantages over both classical and neural network methods.", "keywords": [ "visual perception", "perceptual distances", "two-alternative forced choice" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 2, 1 ], "presentation_list": [ 3, 3, 1 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "fRXAQfHlmr", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Hongda Mao", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Qingshuang Chen", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yelin Kim", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yimu Pan", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 64, "n_ref": 126, "n_ref_all": 149, "n_fig": 10, "n_tab": 8, "L_tab": 2528, "n_element_tab": 125, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1508, "n_element_tab_1": 33, "formula_len_all": 495, "formula_len_all_1": 301, "len_all": 164988, "len_all_1": 52584, "len_abs": 1184, "len_title": 171, "len_sents": 39719, "len_sents_1": 23231, "n_sents": 318, "n_sents_1": 160, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1187, "title": "studentSplat: Your Student Model Learns Single-view 3D Gaussian Splatting", "abs": "Recent advance in feed-forward 3D Gaussian splatting has enable remarkable multi-view 3D scene reconstruction or single-view 3D object reconstruction but single-view 3D scene reconstruction remain under-explored due to inherited ambiguity in single-view. We present studentSplat, the first single-view 3D Gaussian splatting method for scene reconstruction. To overcome the scale ambiguity and extrapolation problems inherent in novel-view supervision from a single input, we introduce two techniques: 1) a teacher-student architecture where a multi-view teacher model provides geometric supervision to the single-view student during training, addressing scale ambiguity and encourage geometric validity; and 2) an extrapolation network that completes missing scene context, enabling high-quality extrapolation. Extensive experiments show studentSplat achieves state-of-the-art single-view novel-view reconstruction quality and comparable performance to multi-view methods at the scene level. Furthermore, studentSplat demonstrates competitive performance as a self-supervised single-view depth estimation method, highlighting its potential for general single-view 3D understanding tasks.", "keywords": [ "3d reconstruction; 3d gaussian splatting; self-supervised learning" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "fRPmc94QeH", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Stuart Shieber", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Full Professor" }, { "name": "Yejin Choi", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Yuntian Deng", "gender": "unknown", "institution": "University of Waterloo", "country": "", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 9, "n_ref_uni": 19, "n_ref": 47, "n_ref_all": 77, "n_fig": 9, "n_tab": 6, "L_tab": 1254, "n_element_tab": 223, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 2085, "n_element_tab_1": 250, "formula_len_all": 226, "formula_len_all_1": 422, "len_all": 107843, "len_all_1": 59716, "len_abs": 1000, "len_title": 123, "len_sents": 37670, "len_sents_1": 25093, "n_sents": 298, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1004, "title": "From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step", "abs": "When leveraging language models for reasoning tasks, generating explicit chain-of-thought (CoT) steps often proves essential for achieving high accuracy in final outputs. In this paper, we investigate if models can be taught to internalize these CoT steps. To this end, we propose a simple yet effective method for internalizing CoT steps: starting with a model trained for explicit CoT reasoning, we gradually remove the intermediate steps and finetune the model. This process allows the model to internalize the intermediate reasoning steps, thus simplifying the reasoning process while maintaining high performance. Our approach enables training a GPT-2 Small model to solve 20-by-20 multiplication with 99.5% accuracy while being 26 times faster than explicit CoT, whereas standard training cannot solve beyond 4-by-4 multiplication. Furthermore, our method proves effective on larger language models, such as Mistral 7B, achieving over 50% accuracy on GSM8K without producing any intermediate steps.", "keywords": [ "implicit reasoning", "chain of thought" ], "rating_list": [ 3, 3, 8, 5 ], "soundness_list": [ 2, 2, 4, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "fRNDDFkPiv", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Eugene Belilovsky", "gender": "Male", "institution": "Concordia University, Montreal", "country": "CA", "position": "Assistant Professor" }, { "name": "Rahaf Aljundi", "gender": "Female", "institution": "Toyota Motor Europe", "country": "BE", "position": "Researcher" }, { "name": "Vaibhav Singh", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 14, "n_ref_uni": 50, "n_ref": 91, "n_ref_all": 106, "n_fig": 1, "n_tab": 6, "L_tab": 1474, "n_element_tab": 106, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 3121, "n_element_tab_1": 130, "formula_len_all": 705, "formula_len_all_1": 691, "len_all": 152655, "len_all_1": 71093, "len_abs": 1178, "len_title": 112, "len_sents": 37659, "len_sents_1": 29433, "n_sents": 290, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1187, "title": "Controlling Forgetting with Test-Time Data in Continual Learning", "abs": "Foundational vision-language models have shown impressive performance on various downstream tasks. Yet, there is still a pressing need to update these models later as new tasks or domains become available. Ongoing Continual Learning (CL) research provides techniques to overcome catastrophic forgetting of previous information when new knowledge is acquired. To date, CL techniques focus only on the supervised training sessions. This results in significant forgetting yielding inferior performance to even the prior model zero shot performance. In this work, we argue that test-time data hold great information that can be leveraged in a self supervised manner to refresh the model's memory of previous learned tasks and hence greatly reduce forgetting at no extra labelling cost. We study how unsupervised data can be employed online to improve models' performance on prior tasks upon encountering representative samples. We propose a simple yet effective student-teacher model with gradient based sparse parameters updates and show significant performance improvements and reduction in forgetting, which could alleviate the role of an offline episodic memory/experience replay buffer.", "keywords": [ "Continual Learning", "Test Time Learning" ], "rating_list": [ 5, 6, 8, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 2, 1, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "fQoYYtPJFX", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Chenyi Kuang", "gender": "Female", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "PhD student" }, { "name": "Jeffrey O. Kephart", "gender": "Male", "institution": "IBM, International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Qiang Ji", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "Professor" } ], "n_formula": 38, "n_formula_1": 25, "n_ref_uni": 36, "n_ref": 80, "n_ref_all": 100, "n_fig": 8, "n_tab": 3, "L_tab": 785, "n_element_tab": 84, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2315, "n_element_tab_1": 93, "formula_len_all": 2272, "formula_len_all_1": 1575, "len_all": 126367, "len_all_1": 61622, "len_abs": 1499, "len_title": 138, "len_sents": 37573, "len_sents_1": 26094, "n_sents": 284, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1508, "title": "Weakly-supervised & Uncertainty-aware 3D Gaze Estimation with Geometry-guided Constraints", "abs": "3D eye gaze estimation from monocular images remains to be a challenging task due to the model sensitivity to illumination, occlusion and head pose changes. As the growing interests and demand in in-the-wild 3D gaze estimation under unconstrained environments, the generalization ability has been considered as a crucial performance metric of 3D gaze estimation models. In this work, we present UGaze-Geo, an uncertainty-aware weakly-supervised framework for 3D gaze estimation. We leverage the general knowledge of human eyeball anatomy and develop multiple geometric constraints. The proposed geometrical constraints contains two types, where the first type is formulated by constructing the mapping function from anatomical 3D eyeball parameters to eye appearance features (eyelid \\& iris landmarks). The second type of constraints is based on the relationship among head rotation, eyeball rotation and gaze, where we learn a variable that describes \"relative eyeball rotation\" conditioned on current head pose. Both type of constraints are free of gaze labels and are general to any subjects and environmental conditions. We formulate these constraints as loss functions in a probabilistic framework.\nWe evaluate the UGaze-Geo framework on within-domain and four cross-domain gaze estimation tasks to validate the effectiveness of each constraint and the advantage of performing probabilistic gaze estimation. Experimental results indicate that our model achieves SOTA performances on different dataset.", "keywords": [ "3D geometry; gaze representation learning", "3D gaze estimation" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "fQbIZY9a3G", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Andrew Markham", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" }, { "name": "Vibhav Vineet", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Xubo Liu", "gender": "Male", "institution": "Facebook", "country": "GB", "position": "Researcher" }, { "name": "Yash Jain", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Yuhang He", "gender": "Male", "institution": "Research, Microsoft", "country": "CA", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 36, "n_ref": 97, "n_ref_all": 126, "n_fig": 9, "n_tab": 8, "L_tab": 2034, "n_element_tab": 235, "n_fig_1": 7, "n_tab_1": 8, "L_tab_1": 1519, "n_element_tab_1": 142, "formula_len_all": 293, "formula_len_all_1": 344, "len_all": 143467, "len_all_1": 73285, "len_abs": 936, "len_title": 107, "len_sents": 40107, "len_sents_1": 33791, "n_sents": 299, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 938, "title": "RiTTA: Modeling Event Relations in Text-to-Audio Generation", "abs": "Despite significant advancements in Text-to-Audio (TTA) generation models achieving high-fidelity audio with fine-grained context understanding, they struggle to model the relations between audio events described in the input text. However, previous TTA methods have not systematically explored audio event relation modeling, nor have they proposed frameworks to enhance this capability. In this work, we systematically study audio event relation modeling in TTA generation models. We first establish a benchmark for this task by: (1) proposing a comprehensive relation corpus covering all potential relations in real-world scenarios; (2) introducing a new audio event corpus encompassing commonly heard sounds; and (3) proposing new evaluation metrics to assess audio event relation modeling from various perspectives. Furthermore, we propose a finetuning framework to enhance existing TTA models' ability to model audio events relation.", "keywords": [ "Text-to-Audio Generation", "Audio Events Relation", "Benchmark", "Dataset Corpus" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 5, 1, 5 ] }, { "paper_id": "fQSZMrjW8X", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Gengchen Mai", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Jielu Zhang", "gender": "Female", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Lan Mu", "gender": "Female", "institution": "University of Georgia", "country": "", "position": "Full Professor" }, { "name": "Nemin Wu", "gender": "Female", "institution": "University of Georgia", "country": "GE", "position": "PhD student" }, { "name": "Ni Lao", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Qian Cao", "gender": "Female", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Yang Song", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Researcher" }, { "name": "Yiqun Xie", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Zeping Liu", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Zhangyu Wang", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Zhongliang Zhou", "gender": "unknown", "institution": "MERCK & CO., INC.", "country": "US", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 16, "n_ref_uni": 32, "n_ref": 81, "n_ref_all": 111, "n_fig": 5, "n_tab": 7, "L_tab": 2302, "n_element_tab": 260, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1965, "n_element_tab_1": 187, "formula_len_all": 1476, "formula_len_all_1": 1101, "len_all": 166817, "len_all_1": 71541, "len_abs": 1685, "len_title": 126, "len_sents": 56228, "len_sents_1": 33358, "n_sents": 405, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1694, "title": "LocDiffusion: Identifying Locations on Earth by Diffusing in the Hilbert Space", "abs": "Image geolocalization is a fundamental yet challenging task, aiming at inferring the geolocation on Earth where an image is taken. Existing methods approach it either via grid-based classification or via image retrieval. The geolocalization accuracy of these methods is constrained by the choice of geographic grid cell sizes or the spatial distributions of the retrieval image/geolocation gallery, and\ntheir performance significantly suffers when the spatial distribution of test images does not align with such choices. To address these limitations, we propose to leverage diffusion models to achieve image geolocalization with arbitrary resolutions. To avoid the problematic manifold reprojection step in diffusion, we developed a novel spherical positional encoding-decoding framework, which encodes points on a spherical surface (e.g., geolocations on Earth) into a Hilbert space of Spherical Harmonics coefficients and decodes points (geolocations) by mode-seeking. We call this type of position encoding Spherical Harmonics Dirac Delta (SHDD) Representation. We also propose a novel SirenNet-based architecture called CS-UNet to learn the conditional backward process in the latent SHDD space by minimizing a latent KL-divergence loss. We train a conditional latent diffusion model called LocDiffusion that generates geolocations under the guidance of images \u2013 to the best of our knowledge, the first generative model to address the image geolocalization problem. We evaluate our LocDiffusion model against SOTA image geolocalization baselines. LocDiffusion achieves competitive geolocalization performance and demonstrates significantly stronger generalizability to unseen geolocations.", "keywords": [ "Image geo-localization", "location encoding", "diffusion" ], "rating_list": [ 6, 3, 8, 3, 6 ], "soundness_list": [ 3, 1, 3, 3, 2 ], "presentation_list": [ 4, 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "fPYJVMBuEc", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Heesan Kong", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "HyeongJoo Hwang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Postdoc" }, { "name": "Kee-Eung Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Seokin Seo", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Ahn Youngjin", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" }, { "name": "Seunghoon Hong", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 24, "n_formula_1": 13, "n_ref_uni": 35, "n_ref": 131, "n_ref_all": 152, "n_fig": 11, "n_tab": 13, "L_tab": 6131, "n_element_tab": 538, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1452, "n_element_tab_1": 120, "formula_len_all": 2908, "formula_len_all_1": 1650, "len_all": 193426, "len_all_1": 68675, "len_abs": 1178, "len_title": 135, "len_sents": 63095, "len_sents_1": 31216, "n_sents": 444, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1181, "title": "Contrast with Aggregation: A Scalable Framework for Multi-View Representation Learning", "abs": "Multi-View Representation Learning (MVRL) aims to learn the joint representation from diverse data sources by discovering complex relationships among them.\nIn MVRL, since the downstream task information and the view availability are often unknown a-priori, it is essential for the joint representation to be robust to the partial availability of views.\nHowever, existing methods exhibit various limitations, such as discarding potentially valuable view-specific information, lacking\nthe ability to extract representation from an arbitrary subset of views, or requiring considerable computational resources that increase exponentially with the number of views.\nTo address these challenges, we present a scalable MVRL framework based on contrastive learning.\nOur approach employs a set of encoders that is able to extract representations from arbitrary subset of views, and jointly trains them with a computation cost that scales linearly with the number of views.\nWe conducted comprehensive evaluations across 7 MVRL benchmark datasets ranging from 2 to 8 views, demonstrating that our method robustly handles diverse input view combinations and outperforms strong baseline methods.", "keywords": [ "Multi-View Representation Learning", "Multimodal Representation Learning", "Contrastive Learning" ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "fPQZd3p1De", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Heejin Ahn", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Hyunchul Bae", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Minhee Kang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Postdoc" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 62, "n_ref_all": 89, "n_fig": 10, "n_tab": 13, "L_tab": 4466, "n_element_tab": 314, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 395, "n_element_tab_1": 49, "formula_len_all": 456, "formula_len_all_1": 438, "len_all": 126600, "len_all_1": 56136, "len_abs": 920, "len_title": 130, "len_sents": 31992, "len_sents_1": 23964, "n_sents": 313, "n_sents_1": 223, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 942, "title": "ParCon: Noise-Robust Collaborative Perception via Multi-module Parallel Connection", "abs": "In this paper, we investigate improving the perception performance of autonomous vehicles through communication with other vehicles and road infrastructures. To this end, we introduce a novel collaborative perception architecture, called $\\textbf{ParCon}$, which connects multiple modules in parallel, as opposed to the sequential connections used in most other collaborative perception methods. Through extensive experiments, we demonstrate that ParCon inherits the advantages of parallel connection. Specifically, ParCon is robust to noise, as the parallel architecture allows each module to manage noise independently and compensate for the limitations of other modules. \nAs a result, ParCon achieves state-of-the-art accuracy, particularly in noisy environments, such as real-world datasets, increasing detection accuracy by 6.91\\%. Additionally, ParCon is computationally efficient, reducing floating-point operations (FLOPs) by 11.46\\%.", "keywords": [ "Collaborative perception", "Vehicle-Infrastructure cooperative driving", "3D Object Detection" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "fPBExgC1m9", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Daichi Zhang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Sabine Susstrunk", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Full Professor" }, { "name": "Shiming Ge", "gender": "Male", "institution": "Institute of Information Engineering, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Tong Zhang", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Postdoc" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 36, "n_ref": 87, "n_ref_all": 107, "n_fig": 7, "n_tab": 5, "L_tab": 2173, "n_element_tab": 126, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3484, "n_element_tab_1": 109, "formula_len_all": 451, "formula_len_all_1": 451, "len_all": 120783, "len_all_1": 64859, "len_abs": 1309, "len_title": 126, "len_sents": 33555, "len_sents_1": 28812, "n_sents": 346, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1211, "title": "Leveraging Natural Frequency Deviation for Diffusion-Generated Image Detection", "abs": "Diffusion models have achieved remarkable success in image synthesis, but the generated high-quality images raise concerns about potential malicious use. Existing detectors often struggle to capture distinctive features across different training models, limiting their generalization to unseen diffusion models with varying schedulers and hyperparameters. To address this issue, we observe that diffusion-generated images exhibit progressively larger differences from real images across low- to high-frequency bands. Based on this insight, we propose a novel image representation called \\textbf{N}atural \\textbf{F}r\\textbf{e}quency \\textbf{De}viation~(\\textbf{DEFEND}). DEFEND applies a weighted filter to the Fourier spectrum, suppressing less discriminative bands while enhancing more informative ones. This approach, grounded in a comprehensive analysis of frequency-based differences between real and diffusion-generated images, enables robust detection of images from unseen diffusion models and provides resilience to various perturbations. Extensive experiments on diffusion-generated image datasets show that our method outperforms state-of-the-art detectors with superior generalization and robustness.", "keywords": [ "frequency domain", "filter banks", "diffusion-generated image detection" ], "rating_list": [ 3, 3, 6, 6 ], "soundness_list": [ 1, 2, 4, 3 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "fOOOyVhTYV", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Amit Daniely", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "", "position": "Associate Professor" }, { "name": "Gal Katzhendler", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "PhD student" } ], "n_formula": 97, "n_formula_1": 34, "n_ref_uni": 5, "n_ref": 24, "n_ref_all": 30, "n_fig": 0, "n_tab": 4, "L_tab": 619, "n_element_tab": 45, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 23452, "formula_len_all_1": 7348, "len_all": 130244, "len_all_1": 55357, "len_abs": 2532, "len_title": 106, "len_sents": 31358, "len_sents_1": 16477, "n_sents": 457, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 752, "title": "A Multicover Approach to Neural Networks Sample Complexity", "abs": "Covering numbers are central to estimating sample complexity. Alas, standard techniques for bounding covering numbers fail in estimating the covering numbers of many classes of neural networks. We introduce a generalization of covers, called {\\em multicovers}, which are covers w.r.t. many metrics simultaneously. \n\nContrary to standard covering numbers, multicovering numbers behave better with the layer-wise structure in neural networks. We utilize this property to recover a recent result of \\citet{ADL2019} who defined a new notion called Approximate Description Length (ADL) to establish tight bounds on the sample complexity of networks with weights of bounded Frobenius norm. We also show that ADL and multicovering numbers are closely related.", "keywords": [ "Sample Complexity", "Covering Numbers", "Neural Networks" ], "rating_list": [ 1, 3, 5, 5, 5 ], "soundness_list": [ 1, 3, 2, 3, 2 ], "presentation_list": [ 1, 2, 1, 2, 1 ], "contribution_list": [ 1, 2, 3, 2, 2 ], "confidence_list": [ 5, 3, 1, 3, 2 ] }, { "paper_id": "fO1xnmW8T6", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "David Acuna", "gender": "Male", "institution": "NVIDIA", "country": "CA", "position": "Researcher" }, { "name": "Rafid Mahmood", "gender": "unknown", "institution": "University of Ottawa", "country": "CA", "position": "Assistant Professor" }, { "name": "Sanja Fidler", "gender": "Female", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "Associate Professor" }, { "name": "Yuan-Hong Liao", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 60, "n_ref": 126, "n_ref_all": 142, "n_fig": 3, "n_tab": 2, "L_tab": 341, "n_element_tab": 36, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 183099, "len_all_1": 65969, "len_abs": 1612, "len_title": 120, "len_sents": 35182, "len_sents_1": 30493, "n_sents": 295, "n_sents_1": 256, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1072, "title": "Can Large Vision-Language Models Correct Grounding Errors By Themselves?", "abs": "Enhancing semantic grounding abilities in Vision-Language Models (VLMs) often involves collecting domain-specific training data, refining the network architectures, or modifying the training recipes. In this work, we venture into an orthogonal direction and explore semantic grounding in VLMs through self-correction, without requiring in-domain data, fine-tuning, or modifications to the network architectures. Despite the concerns raised in the self-correction of LLMs, we find that if prompted and framed properly, VLMs can correct their own semantic grounding mistakes even without the access to the oracle feedback. We also show an identified self-correction framework in an iterative setting which consistently improves performance across all models investigated. Overall, iterative self-correction consistently improves VLM performance by up to 8.4 accuracy points across all models investigated; yet, after several rounds of feedback, strong models like GPT-4V and GPT-4o still exhibit significant error rates, indicating promising directions for further research.", "keywords": [ "vision-language models", "self-correction", "feedback" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "fO0YO9giQV", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Haochao Ying", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Jian Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Jimeng Sun", "gender": "unknown", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Professor" }, { "name": "Jintai Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Xu Cao", "gender": "Non-Binary", "institution": "Department of Computer Science, University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Yaojun Hu", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yue Wang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "James Rehg", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 43, "n_ref": 45, "n_ref_all": 51, "n_fig": 2, "n_tab": 10, "L_tab": 6317, "n_element_tab": 422, "n_fig_1": 1, "n_tab_1": 8, "L_tab_1": 5043, "n_element_tab_1": 251, "formula_len_all": 875, "formula_len_all_1": 830, "len_all": 143458, "len_all_1": 64447, "len_abs": 1693, "len_title": 107, "len_sents": 40043, "len_sents_1": 26716, "n_sents": 274, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1703, "title": "AnyECG: Foundational Models for Electrocardiogram Analysis", "abs": "Electrocardiogram (ECG), a non-invasive and affordable tool for cardiac monitoring, is highly sensitive in detecting acute heart attacks. However, due to the lengthy nature of ECG recordings, numerous machine learning methods have been developed for automated heart disease detection to reduce human workload. Despite these efforts, performance remains suboptimal. A key obstacle is the inherent complexity of ECG data, which includes heterogeneity (e.g., varying sampling rates), high levels of noise, demographic-related pattern shifts, and intricate rhythm-event associations. To overcome these challenges, this paper introduces AnyECG, a foundational model designed to extract robust representations from any real-world ECG data. Specifically, a tailored ECG Tokenizer encodes each fixed-duration ECG fragment into a token and, guided by proxy tasks, converts noisy, continuous ECG features into discrete, compact, and clinically meaningful local rhythm codes. These codes encapsulate basic morphological, frequency, and demographic information (e.g., sex), effectively mitigating signal noise. We further pre-train the AnyECG to learn rhythmic pattern associations across ECG tokens, enabling the capture of cardiac event semantics. By being jointly pre-trained on diverse ECG data sources, AnyECG is capable of generalizing across a wide range of downstream tasks where ECG signals are recorded from various devices and scenarios. Experimental results in anomaly detection, arrhythmia detection, corrupted lead generation, and ultra-long ECG signal analysis demon-\nstrate that AnyECG learns common ECG knowledge from data and significantly outperforms cutting-edge methods in each respective task.", "keywords": [ "ECG representation", "Cardiac Diagnosis", "Denoising Reconstruction", "Foundation Model" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "fNMKqyvuZT", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Can Chang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Haoran He", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Huazhe Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ling Pan", "gender": "Female", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 54, "n_ref": 158, "n_ref_all": 192, "n_fig": 18, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 156, "n_element_tab_1": 15, "formula_len_all": 1124, "formula_len_all_1": 575, "len_all": 186971, "len_all_1": 69284, "len_abs": 1319, "len_title": 129, "len_sents": 56850, "len_sents_1": 34116, "n_sents": 379, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1359, "title": "Looking Backward: Retrospective Backward Synthesis for Goal-Conditioned GFlowNets", "abs": "Generative Flow Networks (GFlowNets), a new family of probabilistic samplers, have demonstrated remarkable capabilities to generate diverse sets of high-reward candidates, in contrast to standard return maximization approaches (e.g., reinforcement learning) which often converge to a single optimal solution. Recent works have focused on developing goal-conditioned GFlowNets, which aim to train a single GFlowNet capable of achieving different outcomes as the task specifies. However, training such models is challenging due to extremely sparse rewards, particularly in high-dimensional problems. Moreover, previous methods suffer from the limited coverage of explored trajectories during training, which presents more pronounced challenges when only offline data is available. In this work, we propose a novel method called \\textbf{R}etrospective \\textbf{B}ackward \\textbf{S}ynthesis (\\textbf{RBS}) to address these critical problems. Specifically, RBS synthesizes new backward trajectories in goal-conditioned GFlowNets to enrich training trajectories with enhanced quality and diversity, thereby introducing copious learnable signals for effectively tackling the sparse reward problem. Extensive empirical results show that our method improves sample efficiency by a large margin and outperforms strong baselines on various standard evaluation benchmarks.", "keywords": [ "Backward Sampling", "Goal-Conditioned Learning", "Generative Flow Networks" ], "rating_list": [ 6, 8, 5, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "fN8yLc3eA7", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Cl\u00e9ment Moulin-Frier", "gender": "Male", "institution": "Inria", "country": "", "position": "Associate Professor" }, { "name": "Corentin L\u00e9ger", "gender": "unknown", "institution": "INRIA", "country": "FR", "position": "Intern" }, { "name": "C\u00e9dric Colas", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Gaia Molinaro", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "J\u00e9r\u00e9my Perez", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "PhD student" }, { "name": "Maxime Derex", "gender": "unknown", "institution": "Universit\u00e9 de Toulouse", "country": "", "position": "Researcher" }, { "name": "Pierre-Yves Oudeyer", "gender": "Male", "institution": "Inria", "country": "", "position": "Research director" }, { "name": "Grgur Kovac", "gender": "Male", "institution": "INRIA", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 92, "n_ref": 123, "n_ref_all": 159, "n_fig": 16, "n_tab": 3, "L_tab": 1352, "n_element_tab": 53, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 120, "formula_len_all_1": 30, "len_all": 234158, "len_all_1": 69035, "len_abs": 4209, "len_title": 151, "len_sents": 66511, "len_sents_1": 35925, "n_sents": 495, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 103, "L_abs": 1580, "title": "When LLMs Play the Telephone Game: Cumulative Changes and Attractors in Iterated Cultural Transmissions", "abs": "As large language models (LLMs) start interacting with each other and generating an increasing amount of text online, it becomes crucial to better understand how information is transformed as it passes from one LLM to the next. While significant research has examined individual LLM behaviors, existing studies have largely overlooked the collective behaviors and information distortions arising from iterated LLM interactions. Small biases, negligible at the single output level, risk being amplified in iterated interactions, potentially leading the content to evolve towards attractor states. In a series of _telephone game experiments_, we apply a transmission chain design borrowed from the human cultural evolution literature: LLM agents iteratively receive, produce, and transmit texts from the previous to the next agent in the chain. By tracking the evolution of text _toxicity_, _positivity_, _difficulty_, and _length_ across transmission chains, we uncover the existence of biases and attractors, and study their dependence on the initial text, the instructions, language model, and model size. For instance, we find that more open-ended instructions lead to stronger attraction effects compared to more constrained tasks. We also find that different text properties display different sensitivity to attraction effects, with _toxicity_ leading to stronger attractors than _length_. These findings highlight the importance of accounting for multi-step transmission dynamics and represent a first step towards a more comprehensive understanding of LLM cultural dynamics.", "keywords": [ "Large Language Models", "Multi-Turn Behaviour", "Cultural Evolution", "Attractors", "Transmission Chain" ], "rating_list": [ 5, 8, 8, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 2, 4, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "fMvcffpsDo", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Yuxin Peng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Zhiwen Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 21, "n_ref": 54, "n_ref_all": 64, "n_fig": 2, "n_tab": 4, "L_tab": 1408, "n_element_tab": 141, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1408, "n_element_tab_1": 141, "formula_len_all": 845, "formula_len_all_1": 845, "len_all": 104811, "len_all_1": 60558, "len_abs": 1392, "len_title": 130, "len_sents": 30332, "len_sents_1": 28032, "n_sents": 186, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1399, "title": "BiC-Occ: Bi-directional Circulated 3D Occupancy Prediction for Autonomous Driving", "abs": "Vision-based 3D occupancy prediction is the cornerstone in autonomous driving systems to provide comprehensive scene perception for subsequent decisions, which requires assessing voxelized 3D scenes with multi-view 2D images. Existing methods mainly adopt unidirectional pipelines projecting image features to BEV representations for following supervision, whose performances are limited by the sparsity and ambiguity of voxel labels. To address this issue, we propose a Bi-directional Circulated 3D Occupancy Prediction (BiC-Occ) framework for more accurate voxel predictions and supervisions. Specifically, we design a Bi-directional View Transformer module that approximates invertible transition matrices of the view transformation process, promoting the self-consistency between 2D image features and 3D BEV representations. Furthermore, we propose a Circulated Interpolation Predictor module that exploits local geometric structures to align multi-scale BEV representations, correcting local ambiguity with consistent occupancy predictions across different resolutions. With the synergy of these two modules, the self-consistency within different perception views and occupancy resolutions compensates for the sparsity and ambiguity of voxel labels, leading to more accurate 3D occupancy predictions. Extensive experiments and analyses demonstrate the effectiveness of our BiC-Occ framework.", "keywords": [ "3D occupancy prediciton", "Bi-directional", "View transformation", "Autonomous driving" ], "rating_list": [ 3, 6, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 5, 3, 3, 5, 4 ] }, { "paper_id": "fMbLszVO1H", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Jan Kautz", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "VP Research" }, { "name": "Jihoon Hong", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Undergrad student" }, { "name": "Kejing Xia", "gender": "Male", "institution": "Bosch", "country": "CN", "position": "Intern" }, { "name": "Pavlo Molchanov", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Research Scientist" }, { "name": "Xiangchi Yuan", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Xin Dong", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yonggan Fu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zhifan Ye", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "SHIZHE DIAO", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yingyan Lin", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 22, "n_ref": 69, "n_ref_all": 88, "n_fig": 10, "n_tab": 3, "L_tab": 464, "n_element_tab": 29, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 2388, "n_element_tab_1": 50, "formula_len_all": 663, "formula_len_all_1": 663, "len_all": 108395, "len_all_1": 56704, "len_abs": 1573, "len_title": 149, "len_sents": 35381, "len_sents_1": 25965, "n_sents": 252, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1616, "title": "LongMamba: Enhancing Mamba's Long-Context Capabilities via Training-Free Receptive Field Enlargement", "abs": "Mamba models have emerged as an efficient alternative to Transformer models for language modeling tasks, offering linear complexity as context length increases. However, despite their efficiency in handling long contexts, recent studies have demonstrated that Mamba models underperform in understanding extended contexts compared to Transformer models. To address this significant shortfall, we propose ``LongMamba\", a training-free technique that significantly enhances the long-context capabilities of Mamba models. Our approach builds upon the discovery that hidden state channels in Mamba models\u2014categorized into \\textit{local} and \\textit{global channels} based on their receptive field lengths\u2014exhibit distinct functionalities. Specifically, the \\textit{global channels} struggle to adaptively extend their effective receptive fields when input lengths far exceed their training sequence length due to exponential decay in their hidden states. We hypothesize this exponential decay is the root cause of Mamba models\u2019 limited performance in extended contexts. LongMamba counters this by effectively expanding the \\textit{global channels}' receptive fields to fully encompass the input sequence length, thus enabling them to capture global information more effectively. Through extensive benchmarking across synthetic and real-world long-context scenarios, LongMamba sets a new standard for state-of-the-art performance in Mamba-based long-context tasks, significantly extending the operational range of Mamba models without requiring additional fine-tuning. All code and models will be released upon acceptance.", "keywords": [ "Large Language Model", "Long Context Understanding" ], "rating_list": [ 6, 8, 5, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 5, 3 ] }, { "paper_id": "fMaEbeJGpp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 1, "author_info_list": [ { "name": "Hanliang Chen", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 27, "n_ref": 43, "n_ref_all": 53, "n_fig": 1, "n_tab": 5, "L_tab": 621, "n_element_tab": 67, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 621, "n_element_tab_1": 67, "formula_len_all": 249, "formula_len_all_1": 248, "len_all": 91126, "len_all_1": 54166, "len_abs": 1505, "len_title": 116, "len_sents": 31181, "len_sents_1": 26569, "n_sents": 183, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1509, "title": "Multimodal Retrieval-Augmented Generation Question-Answering System", "abs": "Retrieval-Augmented Generation (RAG) combines the richness of external knowledge bases with the generative capabilities of large language models (LLMs) to provide users with more accurate and real-time responses. However, in the era of information explosion, the way information is presented is increasingly becoming multimodal. Users are no longer satisfied with the information provided by traditional text-based knowledge bases, making the construction of an efficient and accurate multimodal RAG question-answering system of significant theoretical and practical importance. To address these issues, this paper proposes an innovative RAG question-answering system: this approach pre-designs a rich dataset containing images, text, and question-answer pairs from external knowledge bases for subsequent model training, effectively improving the training quality of the model; it builds a cross-modal retrieval model from text to images, ensuring precise matching between document content and corresponding images, significantly reducing the complexity and processing time of locating relevant images within long texts. Furthermore, the retrieval model and the multimodal question-answering model are integrated to construct an efficient and accurate RAG question-answering system. Experimental results show that this system not only effectively simplifies the document formatting process and improves text-to-image retrieval accuracy but also exhibits comprehensive performance in handling multimodal data.", "keywords": [ "Retrieval-Augmented Generation; Dataset Construction; Text-Image Retrieval; Visual Question-Answering System" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 1, 1, 1, 1 ], "contribution_list": [ 1, 1, 1, 1 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "fMUggopCYI", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Marcel Worring", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "Yijia Zheng", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" } ], "n_formula": 51, "n_formula_1": 14, "n_ref_uni": 78, "n_ref": 227, "n_ref_all": 252, "n_fig": 8, "n_tab": 10, "L_tab": 2992, "n_element_tab": 310, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 2998, "n_element_tab_1": 75, "formula_len_all": 4395, "formula_len_all_1": 707, "len_all": 285980, "len_all_1": 81978, "len_abs": 1720, "len_title": 133, "len_sents": 93204, "len_sents_1": 38819, "n_sents": 714, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1268, "title": "Co-Representation Neural Hypergraph Diffusion for Edge-Dependent Node Classification", "abs": "Hypergraphs are widely employed to represent complex higher-order relations in real-world applications. Most hypergraph learning research focuses on node-level or edge-level tasks. A practically relevant but more challenging task, edge-dependent node classification (ENC), is only recently proposed. In ENC, a node can have different labels across different hyperedges, which requires the modeling of node-edge pairs instead of single nodes or hyperedges. Existing solutions for this task are based on message passing and model interactions in within-edge and within-node structures as multi-input single-output functions. This brings three limitations: (1) non-adaptive representation size, (2) non-adaptive messages, and (3) insufficient direct interactions among nodes or edges. To tackle these limitations, we propose CoNHD, a new ENC solution that models both within-edge and within-node interactions as multi-input multi-output functions. Specifically, we represent these interactions as a hypergraph diffusion process on node-edge co-representations. We further develop a neural implementation for this diffusion process, which can adapt to a specific ENC dataset. Extensive experiments demonstrate the effectiveness and efficiency of the proposed CoNHD method.", "keywords": [ "hypergraph neural networks", "hypergraph diffusion", "hypergraph learning" ], "rating_list": [ 3, 8, 5, 5 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "fMTPkDEhLQ", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Site Bai", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Brian Bullins", "gender": "Male", "institution": "Purdue University", "country": "", "position": "Assistant Professor" } ], "n_formula": 120, "n_formula_1": 48, "n_ref_uni": 18, "n_ref": 86, "n_ref_all": 94, "n_fig": 1, "n_tab": 1, "L_tab": 394, "n_element_tab": 7, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 16998, "formula_len_all_1": 5348, "len_all": 132109, "len_all_1": 64207, "len_abs": 1399, "len_title": 136, "len_sents": 34606, "len_sents_1": 23503, "n_sents": 368, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 86, "L_abs": 1240, "title": "Tight Lower Bounds under Asymmetric High-Order H\u00f6lder Smoothness and Uniform Convexity", "abs": "In this paper, we provide tight lower bounds for the oracle complexity of minimizing high-order H\u00f6lder smooth and uniformly convex functions. Specifically, for a function whose $p^{th}$-order derivatives are H\u00f6lder continuous with degree $\\nu$ and parameter $H$, and that is uniformly convex with degree $q$ and parameter $\\sigma$, we focus on two asymmetric cases: (1) $q > p + \\nu$, and (2) $q < p+\\nu$. Given up to $p^{th}$-order oracle access, we establish worst-case oracle complexities of $\\Omega\\left( \\left( \\frac{H}{\\sigma}\\right)^\\frac{2}{3(p+\\nu)-2}\\left( \\frac{\\sigma}{\\epsilon}\\right)^\\frac{2(q-p-\\nu)}{q(3(p+\\nu)-2)}\\right)$ in the first case with an $\\ell_\\infty$-ball-truncated-Gaussian smoothed hard function and $\\Omega\\left(\\left(\\frac{H}{\\sigma}\\right)^\\frac{2}{3(p+\\nu)-2}+ \\log\\log\\left(\\left(\\frac{\\sigma^{p+\\nu}}{H^q}\\right)^\\frac{1}{p+\\nu-q}\\frac{1}{\\epsilon}\\right)\\right)$ in the second case, for reaching an $\\epsilon$-approximate solution in terms of the optimality gap. Our analysis generalizes previous lower bounds for functions under first- and second-order smoothness as well as those for uniformly convex functions, and furthermore our results match the corresponding upper bounds in this general setting.", "keywords": [ "Convex Optimization", "Uniform Convexity", "Lower Bound", "High-Order Method", "Regularization", "H\u00f6lder Smoothness" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 3, 2, 4, 4 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "fMRq7sPP1y", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Akshil Patel", "gender": "Male", "institution": "University of Bath", "country": "GB", "position": "PhD student" }, { "name": "Jessica Nicholson", "gender": "Female", "institution": "University of Bath", "country": "GB", "position": "PhD student" }, { "name": "Joseph S Goodier", "gender": "Male", "institution": "University of Bath", "country": "", "position": "PhD student" }, { "name": "\u00d6zg\u00fcr \u015eim\u015fek", "gender": "Female", "institution": "University of Bath", "country": "GB", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 43, "n_ref_all": 54, "n_fig": 6, "n_tab": 8, "L_tab": 3122, "n_element_tab": 199, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 677, "formula_len_all_1": 370, "len_all": 135979, "len_all_1": 60538, "len_abs": 765, "len_title": 125, "len_sents": 45576, "len_sents_1": 29917, "n_sents": 372, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 762, "title": "Variational Learned Priors for Intrinsically Motivated Reinforcement Learning", "abs": "Efficient exploration is a fundamental challenge in reinforcement learning, especially in environments with sparse rewards. Intrinsic motivation can improve exploration efficiency by rewarding agents for encountering novel states. In this work, we propose a method called Variation Learned Priors for intrinsic motivation that estimates state novelty through variational state encoding. Specifically, novelty is measured using the Kullback-Leibler divergence between a Variational Autoencoder's learned prior and posterior distributions. When tested across various domains, our approach improves the latent space quality of the Variational Autoencoder, leading to increased exploration efficiency and better task performance for the reinforcement learning agent.", "keywords": [ "reinforcement", "learning", "intrinsic", "motivation", "exploration", "curiosity" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "fMOUybjbnO", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Christos Diou", "gender": "Male", "institution": "Harokopio University", "country": "GR", "position": "Associate Professor" }, { "name": "Christos Koutlis", "gender": "Male", "institution": "CERTH/ITI", "country": "GR", "position": "Postdoc" }, { "name": "Ioannis Sarridis", "gender": "unknown", "institution": "Harokopio University", "country": "GR", "position": "PhD student" }, { "name": "Symeon Papadopoulos", "gender": "Male", "institution": "CERTH/ITI", "country": "GR", "position": "Principal Researcher" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 41, "n_ref": 115, "n_ref_all": 129, "n_fig": 6, "n_tab": 16, "L_tab": 2502, "n_element_tab": 242, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 1375, "n_element_tab_1": 130, "formula_len_all": 543, "formula_len_all_1": 529, "len_all": 137322, "len_all_1": 63732, "len_abs": 624, "len_title": 91, "len_sents": 42228, "len_sents_1": 29388, "n_sents": 275, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1096, "title": "BAdd: Bias Mitigation through Bias Addition", "abs": "Computer vision (CV) datasets often exhibit biases in the form of spurious correlations between certain attributes and target variables that are perpetuated by Deep Learning (DL) models. While recent efforts aim to mitigate such biases and foster bias-neutral representations, they fail in complex real-world scenarios. In particular, existing methods excel in controlled experiments on benchmarks with single-attribute injected biases, but struggle with complex multi-attribute biases that naturally occur in established CV datasets. Here, we introduce BAdd, a simple yet effective method that allows for learning bias-neutral representations invariant to bias-inducing attributes. It achieves this by injecting features encoding these attributes into the training process. BAdd is evaluated on seven benchmarks and exhibits competitive performance, surpassing state-of-the-art methods on both single- and multi-attribute bias settings. Notably, it achieves +27.5% and +5.5% absolute accuracy improvements on the challenging multi-attribute benchmarks, FB-Biased-MNIST and CelebA, respectively.", "keywords": [ "fairness", "bias", "spurious correlations" ], "rating_list": [ 5, 8, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 4, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "fMNRYBvcQN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Shengyuan Hu", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Virginia Smith", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Yiwei Fu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Zhiwei Steven Wu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 3, "n_ref_uni": 34, "n_ref": 84, "n_ref_all": 107, "n_fig": 5, "n_tab": 10, "L_tab": 2280, "n_element_tab": 285, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 344, "n_element_tab_1": 44, "formula_len_all": 500, "formula_len_all_1": 134, "len_all": 228817, "len_all_1": 78564, "len_abs": 1026, "len_title": 134, "len_sents": 85341, "len_sents_1": 36422, "n_sents": 699, "n_sents_1": 278, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1031, "title": "Jogging the Memory of Unlearned LLMs Through Targeted Relearning Attacks", "abs": "Machine unlearning is a promising approach to mitigate undesirable memorization of training data in ML models. However, in this work we show that existing approaches for unlearning in LLMs are surprisingly susceptible to a simple set of *targeted relearning attacks*. With access to only a small and potentially loosely related set of data, we find that we can \u2018jog\u2019 the memory of unlearned models to reverse the effects of unlearning. For example, we show that relearning on public medical articles can lead an unlearned LLM to output harmful knowledge about bioweapons, and relearning general wiki information about the book series Harry Potter can force the model to output verbatim memorized text. We formalize this unlearning-relearning pipeline, explore the attack across three popular unlearning benchmarks, and discuss future directions and guidelines that result from our study. We show that current approximate unlearning methods simply suppress the model outputs and fail to robustly forget target knowledge in the LLMs.", "keywords": [ "Machine Unlearning", "Large Language Model" ], "rating_list": [ 8, 5, 8, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 4, 2, 2, 3 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "fM432E7l5w", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Bo Hui", "gender": "unknown", "institution": "University of Tulsa", "country": "US", "position": "Assistant Professor" }, { "name": "Da Yan", "gender": "Male", "institution": "Indiana University Bloomington", "country": "US", "position": "Associate Professor" }, { "name": "Yi Shi", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "Postdoc" }, { "name": "Zinan Ling", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 34, "n_formula_1": 32, "n_ref_uni": 49, "n_ref": 88, "n_ref_all": 104, "n_fig": 6, "n_tab": 4, "L_tab": 876, "n_element_tab": 147, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 333, "n_element_tab_1": 29, "formula_len_all": 1326, "formula_len_all_1": 1302, "len_all": 175033, "len_all_1": 67697, "len_abs": 1402, "len_title": 125, "len_sents": 34852, "len_sents_1": 29160, "n_sents": 247, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1408, "title": "Towards Building Reliable Conditional Diffusion Models for Protein Generation", "abs": "Generating novel and functional protein sequences is critical to a wide\nrange of applications in biology. Recent advancements in conditional diffusion models have shown impressive empirical performance in protein generation tasks. However, reliable generations of protein remain an open research question in de novo protein design, especially when it comes to conditional diffusion models. Considering the biological function of a protein is determined\nby multi-level structures, we propose a novel multi-level conditional diffusion model that integrates both sequence-based\nand structure-based information for efficient end-to-end protein design guided by\nspecified functions. By generating representations at different levels simultaneously, our framework can effectively model the inherent hierarchical relations between different levels, resulting in an informative and\ndiscriminative representation of the generated protein. We also propose a Protein-MMD, a new reliable evaluation metric, to evaluate the quality of generated protein with conditional diffusion models. Our new metric is able to capture both distributional and functional similarities between real and generated protein sequences while ensuring conditional consistency. We experiment with standard datasets and the\nresults on protein generation tasks demonstrate the efficacy of the proposed generation framework and evaluation metric.", "keywords": [ "Protein Generation" ], "rating_list": [ 6, 3, 5, 3, 1 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 3, 1, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 1, 1 ], "confidence_list": [ 4, 5, 4, 4, 5 ] }, { "paper_id": "fL8sds4naU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Chen Li", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Han Hu", "gender": "Male", "institution": "Tencent Hunyuan Research", "country": "CN", "position": "Researcher" }, { "name": "Houwen Peng", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" }, { "name": "Jingcheng Hu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Nanning Zheng", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Weiqi Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Yixuan Wei", "gender": "Not Specified", "institution": "independent", "country": "CN", "position": "Researcher" }, { "name": "Zheng Zhang", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 20, "n_ref": 34, "n_ref_all": 66, "n_fig": 6, "n_tab": 7, "L_tab": 5230, "n_element_tab": 185, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 437, "n_element_tab_1": 50, "formula_len_all": 224, "formula_len_all_1": 95, "len_all": 148383, "len_all_1": 60066, "len_abs": 1478, "len_title": 114, "len_sents": 44701, "len_sents_1": 26982, "n_sents": 389, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1487, "title": "Common 7B Language Models Already Possess Strong Math Capabilities", "abs": "It was once believed that mathematical capabilities in language models required either large model scales or extensive math-related data pre-training. However, this paper demonstrates that the small-scale LLaMA-2 7B model already possesses strong mathematical potential. This is evidenced by its impressive scores of 97.6% on GSM8K benchmark and 70% on MATH benchmark, achieved by selecting the oracle response from 1024 generations. Equipped GPT-4 Turbo as an additional verification, LLaMA-2 7B also achieves 91.8% accuracy on GSM8K benchmark. This indicates that the primary issue within current models is the difficulty in consistently eliciting the inherent mathematical capabilities. We find that scaling up synthetic SFT data, which proves to be nearly as effective as real data, can significantly enhance the reliability of generating correct answers. Surprisingly, even with approximately one million samples, we observe no clear performance saturation. And our method is more efficient with large data scale than previous works. This approach achieves an accuracy of 82.4% on GSM8K and 40.1% on MATH using LLaMA-2 7B model, surpassing GPT-3.5 Turbo. Our 70B model even exceeds an early version of GPT-4 on MATH and out-of-domain Hungarian National High School Math Exam. These results demonstrate our method significantly elicits the general mathematical capabilities of language models. Also, we provide insights into scaling behaviors across different reasoning complexities.", "keywords": [ "Large language model", "Math capabilities", "Synthetic data", "Alignment" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "fL8Zp8o6RL", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Huping Ding", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jingzhen Ding", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Kaifu Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "vice president" }, { "name": "Zhao Xu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Guohao Xu", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 34, "n_ref": 79, "n_ref_all": 116, "n_fig": 9, "n_tab": 6, "L_tab": 1315, "n_element_tab": 227, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 301, "formula_len_all_1": 272, "len_all": 153855, "len_all_1": 52357, "len_abs": 1740, "len_title": 126, "len_sents": 36536, "len_sents_1": 25019, "n_sents": 255, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1768, "title": "FTP: Efficient Prefilling for Long-Context LLM Inference via FFN Token Pruning", "abs": "Large Language Models (LLMs) have demonstrated remarkable performance across various NLP tasks, and have extended their capability to long-context scenarios. However, the increasing context length leads to longer inference time in both the prefilling and decoding stages.\nExisting token pruning methods primarily evict tokens to compress the KV cache, and only accelerate the decoding stage.\nRecent studies have extended token pruning to both stages, but they either yield subtle speedup during the prefilling stage or defer a portion of computations to the decoding phase. Critically, these approaches prioritize the attention module, overlooking the significant computations in the Feed-Forward Network (FFN) module.\n\nIn this work, we focus on the prefilling stage and propose a novel token pruning method named FTP for long-context LLM inference.\nOur approach is based on the observation that the FFN module accounts for over 60\\% of the inference time. FTP reduces this by pruning non-critical tokens before the inference of FFN. The importance of each token, along with the quantity to be pruned, are dynamically determined by the attention scores in each layer.\nUnlike previous token pruning methods, FTP preserves a substantial amount of information of the pruned tokens through the residual connection, thereby achieving a notable speedup with only a negligible decrease in performance.\nSpecifically, the Qwen2-7B-Instruct model with FTP achieves a speedup of 1.24$\\times$ in the prefilling stage with only a 1.30\\% performance drop compared to the baseline model. The speedup is further boosted to 1.39$\\times$ on a Qwen1.5-32B-Chat model.\nExtensive experiments on long-context datasets across various tasks demonstrate the potential and effectiveness of FTP.", "keywords": [ "Large language model", "Inference acceleration", "Token pruning", "Long-context inference", "Natural Language Processing" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "fL4qWkSmtM", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bolin Ding", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Senior Director" }, { "name": "Chenheng Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jinyang Gao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Lizhe Fang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Stefanie Jegelka", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Yifei Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yisen Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhaoyang Liu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 40, "n_ref": 81, "n_ref_all": 140, "n_fig": 13, "n_tab": 6, "L_tab": 1625, "n_element_tab": 267, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 812, "n_element_tab_1": 153, "formula_len_all": 1165, "formula_len_all_1": 723, "len_all": 168541, "len_all_1": 59913, "len_abs": 1456, "len_title": 113, "len_sents": 44476, "len_sents_1": 27475, "n_sents": 339, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1481, "title": "What is Wrong with Perplexity for Long-context Language Modeling?", "abs": "Handling long-context inputs is crucial for large language models (LLMs) in tasks such as extended conversations, document summarization, and many-shot in-context learning. While recent approaches have extended the context windows of LLMs and employed perplexity (PPL) as a standard evaluation metric, PPL has proven unreliable for assessing long-context capabilities. The underlying cause of this limitation has remained unclear. In this work, we provide a comprehensive explanation for this issue. We find that PPL overlooks key tokens, which are essential for long-context understanding, by averaging across all tokens and thereby obscuring the true performance of models in long-context scenarios. To address this, we propose \\textbf{LongPPL}, a novel metric that focuses on key tokens by employing a long-short context contrastive method to identify them. Our experiments demonstrate that LongPPL strongly correlates with performance on various long-context benchmarks (e.g., Pearson correlation of -0.96), significantly outperforming traditional PPL in predictive accuracy. Additionally, we introduce \\textbf{LongCE} (Long-context Cross-Entropy) loss, a re-weighting strategy for fine-tuning that prioritizes key tokens, leading to consistent improvements across diverse benchmarks. In summary, these contributions offer deeper insights into the limitations of PPL and present effective solutions for accurately evaluating and enhancing the long-context capabilities of LLMs.", "keywords": [ "Large language models", "Long context" ], "rating_list": [ 8, 6, 6, 6, 8 ], "soundness_list": [ 4, 2, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 2, 3, 3 ], "confidence_list": [ 5, 4, 2, 4, 3 ] }, { "paper_id": "fKrFTGnoXY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Ali Shafahi", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Amin Ghiasi", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Charan Prakash", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Reza Ardekani", "gender": "unknown", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Trevine Oorloff", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Vishwanath A. Sindagi", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Wele Gedara Chaminda Bandara", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 17, "n_ref_uni": 29, "n_ref": 66, "n_ref_all": 92, "n_fig": 8, "n_tab": 3, "L_tab": 1271, "n_element_tab": 161, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1126, "n_element_tab_1": 143, "formula_len_all": 873, "formula_len_all_1": 743, "len_all": 106131, "len_all_1": 62454, "len_abs": 1483, "len_title": 119, "len_sents": 33352, "len_sents_1": 29224, "n_sents": 213, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1496, "title": "STABLE DIFFUSION MODELS ARE SECRETLY GOOD AT VISUAL IN-CONTEXT LEARNING", "abs": "Large language models (LLM) in natural language processing (NLP) have demonstrated great potential for in-context learning (ICL) -- the ability to leverage a few set of example prompts to adapt to various tasks without having to explicitly update model weights. \nICL has recently been explored for the visual domain with promising early outcomes. These approaches involve specialized training and/or additional data which complicate the process and limit its generalizability. In this work, we show that off-the-shelf Stable Diffusion models can be re-purposed for visual in-context learning (V-ICL). Specifically, we formulate an in-place attention re-computation within the self-attention layers of the Stable Diffusion architecture that explicitly incorporates context between the query and example prompts. Without any additional fine-tuning, we show that this re-purposed Stable Diffusion model is able to adapt to six different tasks: foreground segmentation, single object detection, semantic segmentation, keypoint detection, edge detection, and colorization. \nFor example, the proposed approach improves the mean intersection over union (mIoU) for the foreground segmentation task on Pascal-5i dataset by 8.9\\% and 3.2\\% over recent methods such as Visual Prompting and IMProv, respectively. Additionally, we show that the proposed method is able to effectively leverage multiple prompts through ensembling to infer the task better and further improve the performance across all tasks.", "keywords": [ "in context learning", "stable diffusion" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 1, 3, 2 ], "confidence_list": [ 5, 3, 5 ] }, { "paper_id": "fJ1hON2r2u", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Anthony Kum Hoe Tung", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Qiang Huang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yandong Sun", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Yiqun Sun", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yixuan Tang", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Ziwei Xu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" } ], "n_formula": 25, "n_formula_1": 24, "n_ref_uni": 32, "n_ref": 69, "n_ref_all": 116, "n_fig": 20, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 284, "n_element_tab_1": 24, "formula_len_all": 1513, "formula_len_all_1": 1442, "len_all": 125000, "len_all_1": 66458, "len_abs": 1256, "len_title": 143, "len_sents": 43282, "len_sents_1": 27339, "n_sents": 350, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1351, "title": "A Single Swallow Does Not Make a Summer: Understanding Semantic Structures in Embedding Spaces", "abs": "Embedding spaces encapsulate rich information from deep learning models, with vector distances reflecting the semantic similarity between textual elements. However, their abstract nature and the computational complexity of analyzing them remain significant challenges. To address these, we introduce the concept of Semantic Field Subspace, a novel mapping that links embedding spaces with the underlying semantics. We propose \\textsf{SAFARI}, a novel algorithm for \\textsf{S}em\\textsf{A}ntic \\textsf{F}ield subsp\\textsf{A}ce dete\\textsf{R}m\\textsf{I}nation, which leverages hierarchical clustering to discover hierarchical semantic structures, using Semantic Shifts to capture semantic changes as clusters merge, allowing for the identification of meaningful subspaces. To improve scalability, we extend Weyl's Theorem, enabling an efficient approximation of Semantic Shifts that significantly reduces computational costs. Extensive evaluations on five real-world datasets demonstrate the effectiveness of \\textsf{SAFARI} in uncovering interpretable and hierarchical semantic structures. Additionally, our approximation method achieves a 15$\\sim$30$\\times$ speedup while maintaining minimal errors (less than 0.01), making it practical for large-scale applications. The source code is available at \\url{https://anonymous.4open.science/r/Safari-C803/}.", "keywords": [ "Embedding Space Understanding", "Semantic Field", "Semantic Field Subspace", "Hierarchical Clustering" ], "rating_list": [ 3, 5, 3, 3, 5 ], "soundness_list": [ 1, 3, 2, 1, 3 ], "presentation_list": [ 1, 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 2, 3, 3, 4, 3 ] }, { "paper_id": "fIMf9zQo9d", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Baorui Ma", "gender": "Male", "institution": "BAAI", "country": "CN", "position": "Researcher" }, { "name": "Junsheng Zhou", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Tiejun Huang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Xinlong Wang", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Yu-Shen Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Hao Ge Deng", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 28, "n_ref": 105, "n_ref_all": 153, "n_fig": 18, "n_tab": 4, "L_tab": 919, "n_element_tab": 77, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 252, "n_element_tab_1": 37, "formula_len_all": 737, "formula_len_all_1": 605, "len_all": 176449, "len_all_1": 68707, "len_abs": 1448, "len_title": 143, "len_sents": 50381, "len_sents_1": 34419, "n_sents": 349, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1453, "title": "GeoDream: Disentangling 2D and Geometric Priors for High-Fidelity and Consistent 3D Generation", "abs": "Text-to-3D generation by distilling pretrained large-scale text-to-image diffusion models has shown great promise but still suffers from inconsistent 3D geometric structures (Janus problems) and severe artifacts. The aforementioned problems mainly stem from 2D diffusion models lacking 3D awareness during the lifting. In this work, we present GeoDream, a novel method that incorporates explicit generalized 3D priors with 2D diffusion priors to enhance the capability of obtaining unambiguous 3D consistent geometric structures without sacrificing diversity or fidelity. Specifically, we first utilize a multi-view diffusion model to generate posed images and then construct cost volume from the predicted image, which serves as native 3D geometric priors, ensuring spatial consistency in 3D space. Subsequently, we further propose to harness 3D geometric priors to unlock the great potential of 3D awareness in 2D diffusion priors via a disentangled design. Notably, disentangling 2D and 3D priors allows us to refine 3D geometric priors further. We justify that the refined 3D geometric priors aid in the 3D-aware capability of 2D diffusion priors, which in turn provides superior guidance for the refinement of 3D geometric priors. Our numerical and visual comparisons demonstrate that GeoDream generates more 3D consistent textured meshes with high-resolution realistic renderings (i.e., 1024 * 1024) and adheres more closely to semantic coherence.", "keywords": [ "3D generation; Text to 3D; Image to 3D; AIGC" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "fHvh913U1H", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Anda Cheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Lei Wang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Shoumeng Yan", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Tao Wei", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Wei Huang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yinggui Wang", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhao Zhang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 35, "n_ref": 79, "n_ref_all": 97, "n_fig": 5, "n_tab": 7, "L_tab": 2353, "n_element_tab": 299, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 4088, "n_element_tab_1": 483, "formula_len_all": 808, "formula_len_all_1": 581, "len_all": 119625, "len_all_1": 63185, "len_abs": 1831, "len_title": 137, "len_sents": 31549, "len_sents_1": 25098, "n_sents": 222, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1804, "title": "Mitigating Catastrophic Forgetting in Large Language Models with Forgetting-aware Pruning", "abs": "Recent advancements in Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks. These models are typically pretrained on extensive corpora and subsequently fine-tuned on task-specific datasets. However, during the fine-tuning process, LLMs often suffer from Catastrophic Forgetting (CF), wherein previously acquired general knowledge is lost. Traditional approaches to mitigating CF often rely on data replay, which may not be viable when the original training data is inaccessible. Additionally, methods that alter the training process or the model architecture can increase complexity and detract from the accuracy of downstream tasks, thus limiting their generalizability. In this paper, we propose Forgetting-Aware Pruning Metric (FAPM), a novel pruning-based approach to balance CF and downstream task performance. Our investigation reveals that the degree to which task vectors (i.e., the subtraction of pre-trained weights from the weights fine-tuned on downstream tasks) overlap with pre-trained model parameters is a critical factor for CF. Motivated by this insight, FAPM employs the ratio of the task vector to pre-trained model parameters as a metric to quantify CF, integrating this measure into the pruning criteria. Importantly, FAPM does not necessitate modifications to the training process or model architecture, nor does it require any auxiliary data. We conducted extensive experiments across six datasets encompassing natural language inference, question answering, reading comprehension, and cloze tests. The results demonstrate that FAPM limits CF to just 1% while maintaining 99% accuracy on downstream tasks, rendering FAPM highly competitive relative to the state-of-the-art methods that involve modifications to the training process.", "keywords": [ "Large Language Models", "Catastrophic Forgetting", "Neural Network Pruning" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "fHqwCsDK1z", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Mohd Ashhad", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "MS student" }, { "name": "Ricardo Henao", "gender": "Male", "institution": "Duke University", "country": "", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 52, "n_ref": 92, "n_ref_all": 119, "n_fig": 5, "n_tab": 10, "L_tab": 6131, "n_element_tab": 513, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 377, "formula_len_all_1": 163, "len_all": 190417, "len_all_1": 69950, "len_abs": 1541, "len_title": 123, "len_sents": 50283, "len_sents_1": 35163, "n_sents": 752, "n_sents_1": 267, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1547, "title": "Conditioning on Time is All You Need for Synthetic Survival Data Generation", "abs": "Synthetic data generation holds considerable promise, offering avenues to enhance privacy, fairness, and data accessibility. Despite the availability of various methods for generating synthetic tabular data, challenges persist, particularly in specialized applications such as survival analysis. One significant obstacle in survival data generation is censoring, which manifests as not knowing the precise timing of observed (target) events for certain instances. Existing methods face difficulties in accurately reproducing the real distribution of event times for both observed (uncensored) events and censored events, i.e., the generated event-time distributions do not accurately match the underlying distributions of the real data. So motivated, we propose a simple paradigm to produce synthetic survival data by generating covariates conditioned on event times (and censoring indicators), thus allowing one to reuse existing conditional generative models for tabular data without significant computational overhead, and without making assumptions about the (usually unknown) generation mechanism underlying censoring. We evaluate this method via extensive experiments on real-world datasets. Our methodology outperforms multiple competitive baselines at generating survival data, while improving the performance of downstream survival models trained on it and tested on real data. Importantly, our approach achieves these improvements without compromising patient privacy, offering a balanced solution for synthetic survival data generation.", "keywords": [ "Survival Data Generation", "Survival Analysis", "Tabular Data Generation", "Generative Modeling" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 1, 2, 1 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "fHNpXyhrTC", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Dayang Liang", "gender": "Male", "institution": "Xiamen University", "country": "", "position": "PhD student" }, { "name": "Huiyi Deng", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Yunlong Liu", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 22, "n_formula_1": 15, "n_ref_uni": 35, "n_ref": 72, "n_ref_all": 91, "n_fig": 6, "n_tab": 5, "L_tab": 1485, "n_element_tab": 92, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 83, "n_element_tab_1": 11, "formula_len_all": 1447, "formula_len_all_1": 1025, "len_all": 129062, "len_all_1": 59852, "len_abs": 1471, "len_title": 142, "len_sents": 42805, "len_sents_1": 30461, "n_sents": 247, "n_sents_1": 165, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1474, "title": "Preference-based Credit Assignment for Reinforcement Learning with Delayed and Noised Rewards", "abs": "Credit assignment has been utilized as a common technique for determining the past key state-action pairs and assigning corresponding rewards that have strong relevance with the final outputs in reinforcement learning, especially for environments with delayed rewards. However, current reward function design methods rely heavily on domain knowledge and may not accurately reflect the actual reward that should be received, which will lead to noised reward assignments during the credit assignment process and deteriorate the performance of the agent. To address this issue, in this paper, by leveraging the benefits of Preference-based Reinforcement Learning (PbRL), we propose a novel trajectory preference-based credit assignment method, where each trajectory is assigned to one of three different preferences according to its related delayed reward and the entire trajectory space. Then, a causal Transformer framework is introduced to predict the relevance between the decisions at each timestep and the different trajectory preferences to guide the credit assignment. Despite the unavoidable noised reward related to each trajectory, we demonstrate that our method can still effectively guide agents to learn superior strategies. Experiments on the Mujoco task and the treatment of sepsis under extremely delayed reward setting show that our method can mitigate the adverse effects resulting from the delayed noised rewards and provide effective guidelines for agents.", "keywords": [ "credit assignment", "delayed rewards", "noised rewards", "Preferencece-based Reinforcement Learning", "treatment of sepsis" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 1, 2, 1 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "fGhr39bqZa", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "Jun Wang", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Xiu-Chuan Li", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "PhD student" } ], "n_formula": 56, "n_formula_1": 22, "n_ref_uni": 51, "n_ref": 99, "n_ref_all": 137, "n_fig": 10, "n_tab": 11, "L_tab": 6646, "n_element_tab": 313, "n_fig_1": 8, "n_tab_1": 7, "L_tab_1": 1362, "n_element_tab_1": 146, "formula_len_all": 9647, "formula_len_all_1": 2880, "len_all": 190004, "len_all_1": 77192, "len_abs": 1351, "len_title": 125, "len_sents": 46065, "len_sents_1": 28365, "n_sents": 438, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1434, "title": "Recovery of Causal Graph Involving Latent Variables via Homologous Surrogates", "abs": "Causal discovery with latent variables is an important and challenging problem. To identify latent variables and infer their causal relations, most existing works rely on the assumption that latent variables have pure children. Considering that this assumption is potentially restrictive in practice and not strictly necessary in theory, in this paper, by introducing the concept of homologous surrogate, we eliminate the need for pure child in the context of causal discovery with latent variables. The homologous surrogate fundamentally differs from the pure child in the sense that the latter is characterized by having strictly restricted parents while the former allows for much more flexible parents. We formulate two assumptions involving homologous surrogates and develop theoretical results under each assumption. Under the weaker assumption, our theoretical results imply that we can determine each variable's ancestors, that is, partially recover the causal graph. The stronger assumption further enables us to determine each variable's parents exactly, that is, fully recover the causal graph. Building on these theoretical results, we derive an algorithm that fully leverages the properties of homologous surrogates for causal graph recovery. Also, we validate its efficacy through experiments. Our work broadens the applicability of causal discovery. Our code is available at: https://github.com/XiuchuanLi/ICLR2025-CDHS", "keywords": [ "causal discovery", "latent variables" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "fGdF8Bq1FV", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Abdellatif Zaidi", "gender": "Male", "institution": "Universit\u00e9 Gustave Eiffel", "country": "FR", "position": "Associate Professor" }, { "name": "Milad Sefidgaran", "gender": "Male", "institution": "Huawei Technologies Ltd. (Pairs Resaerch Center)", "country": "", "position": "Researcher" }, { "name": "Piotr Krasnowski", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" } ], "n_formula": 125, "n_formula_1": 34, "n_ref_uni": 1, "n_ref": 2, "n_ref_all": 27, "n_fig": 4, "n_tab": 8, "L_tab": 2197, "n_element_tab": 256, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 124, "n_element_tab_1": 13, "formula_len_all": 9886, "formula_len_all_1": 2085, "len_all": 253389, "len_all_1": 70549, "len_abs": 1118, "len_title": 144, "len_sents": 66754, "len_sents_1": 29874, "n_sents": 618, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 96, "L_abs": 1125, "title": "Generalization Guarantees for Representation Learning via Data-Dependent Gaussian Mixture Priors", "abs": "We establish in-expectation and tail bounds on the generalization error of representation learning type algorithms. The bounds are in terms of the relative entropy between the distribution of the representations extracted from the training and \"test'' datasets and a data-dependent symmetric prior, i.e., the Minimum Description Length (MDL) of the latent variables for the training and test datasets. Our bounds are shown to reflect the \"structure'' and \"simplicity'' of the encoder and significantly improve upon the few existing ones for the studied model. We then use our in-expectation bound to devise a suitable data-dependent regularizer; and we investigate thoroughly the important question of the selection of the prior. We propose a systematic approach to simultaneously learning a date-dependent Gaussian mixture prior and using it as a regularizer. Interestingly, we show that a weighted attention mechanism emerges naturally in this procedure. Our experiments show that our approach outperforms the now popular Variational Information Bottleneck (VIB) method as well as the recent Category-Dependent VIB (CDVIB).", "keywords": [ "Representation learning algorithm", "Gaussian-Mixture", "regularizer", "rate-disotortion" ], "rating_list": [ 8, 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 3, 2, 4 ], "confidence_list": [ 2, 3, 3, 2, 3 ] }, { "paper_id": "fGTtEG24lA", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "David E. Jacobs", "gender": "Male", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Michael Rubinstein", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Principal Researcher" }, { "name": "Nataniel Ruiz", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Neal Wadhwa", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Shlomi Fruchter", "gender": "Male", "institution": "Research, Google", "country": "IL", "position": "Researcher" }, { "name": "Yael Pritch", "gender": "Female", "institution": "Google Research", "country": "US", "position": "Researcher" }, { "name": "Yuanzhen Li", "gender": "Female", "institution": "Google", "country": "US", "position": "Software Engineer" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 12, "n_ref": 22, "n_ref_all": 39, "n_fig": 8, "n_tab": 6, "L_tab": 2533, "n_element_tab": 106, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 923, "n_element_tab_1": 59, "formula_len_all": 189, "formula_len_all_1": 189, "len_all": 120096, "len_all_1": 50413, "len_abs": 272, "len_title": 89, "len_sents": 26219, "len_sents_1": 23965, "n_sents": 182, "n_sents_1": 157, "venue": "ICLR2025", "decision": "withdraw", "L_title": 39, "L_abs": 1012, "title": "Magic Insert: Style-Aware Drag-and-Drop", "abs": "We present Magic Insert, a method for dragging-and-dropping subjects from a user-provided image into a target image of a different style in a physically plausible manner while matching the style of the target image. This work formalizes the problem of style-aware drag-and-drop and presents a method for tackling it by addressing two sub-problems: style-aware personalization and realistic object insertion in stylized images. For style-aware personalization, our method first fine-tunes a pretrained text-to-image diffusion model using LoRA and learned text tokens on the subject image, and then infuses it with a CLIP representation of the target style. For object insertion, we use Bootstrapped Domain Adaption to adapt a domain-specific photorealistic object insertion model to the domain of diverse artistic styles. Overall, the method significantly outperforms traditional approaches such as inpainting. Finally, we present a dataset, SubjectPlop, to facilitate evaluation and future progress in this area.", "keywords": [ "generative models", "diffusion models", "editing", "personalization", "style" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 5, 5 ] }, { "paper_id": "fGSEWgRHNZ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "James Flemings", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Meisam Razaviyayn", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Associate Professor" }, { "name": "Murali Annavaram", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Full Professor" } ], "n_formula": 32, "n_formula_1": 15, "n_ref_uni": 32, "n_ref": 67, "n_ref_all": 92, "n_fig": 2, "n_tab": 2, "L_tab": 729, "n_element_tab": 74, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 158, "n_element_tab_1": 2, "formula_len_all": 3797, "formula_len_all_1": 1670, "len_all": 139869, "len_all_1": 61771, "len_abs": 1829, "len_title": 113, "len_sents": 40232, "len_sents_1": 29910, "n_sents": 307, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1860, "title": "Adaptively Private Next-Token Prediction of Large Language Models", "abs": "As Large Language Models (LLMs) proliferate, developing privacy safeguards for these models is crucial. One popular safeguard involves training LLMs in a differentially private manner. However, such solutions are shown to be computationally expensive and detrimental to the utility of these models. Since LLMs are deployed on the cloud and thus only accessible via an API, a Machine Learning as a Service (MLaaS) provider can protect its downstream data by privatizing the predictions during the decoding process. However, the practicality of such solutions still largely lags behind DP training methods. One recent promising approach, Private Mixing of Ensemble Distributions (PMixED), avoids additive noise by sampling from the output distributions of private LLMs mixed with the output distribution of a public model. Yet, PMixED must satisfy a fixed privacy level for a given number of queries, which is difficult for an analyst to estimate before inference and, hence, does not scale. To this end, we relax the requirements to a more practical setting by introducing Adaptive PMixED ($\\texttt{AdaPMixED}$), a private decoding framework based on PMixED that is adaptive to the private and public output distributions evaluated on a given input query. In this setting, we introduce a noisy screening mechanism that filters out queries with potentially expensive privacy loss, and a data-dependent analysis that exploits the divergence of the private and public output distributions in its privacy loss calculation. Our experimental evaluations demonstrate that our mechanism and analysis $\\textit{can reduce the privacy loss by $16\\times$}$ while preserving the utility over the original PMixED. Furthermore, performing 100K predictions with $\\texttt{AdaPMixED}$ still achieves strong utility and a reasonable data-dependent privacy loss of $\\epsilon=5.25$.", "keywords": [ "Differential Privacy", "Language Models" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "fGIqGfmgkW", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Biqing Qi", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Bowen Zhou", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Ermo Hua", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Haoxin Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Jiayuan Zhang", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Kaiyan Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ning Ding", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Xingtai Lv", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xuekai Zhu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 5, "n_ref_uni": 52, "n_ref": 119, "n_ref_all": 145, "n_fig": 7, "n_tab": 8, "L_tab": 3673, "n_element_tab": 496, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1944, "n_element_tab_1": 282, "formula_len_all": 416, "formula_len_all_1": 246, "len_all": 202718, "len_all_1": 61834, "len_abs": 2402, "len_title": 127, "len_sents": 44541, "len_sents_1": 26990, "n_sents": 339, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1875, "title": "OpenPRM: Building Open-domain Process-based Reward Models with Preference Trees", "abs": "Scaling inference-time computation is increasingly seen as the next frontier in scaling laws for large language models. Previous work in mathematics and coding has demonstrated the remarkable potential for inference-time scaling. During such scaling, fine-grained supervision through process-based reward models (PRMs) is essential for enhancement. However, exploration of inference-time scaling and PRMs in open-domain problems remains limited, where lacking exact answers and obtaining process supervision prove challenging. In this paper, we explore the construction of PRMs for open-domain tasks, specifically for instruction-following tasks. Utilizing existing outcome-based reward models (ORMs), we develop sentence-level preference trees based on the prefix similarity of parallel sampled candidates from datasets like UltraFeedback. This setup allows us to derive weak supervision for processes via back-propagation from outcome-level rewards. Subsequently, we integrate ORMs and PRMs under the same pairwise ranking objectives, resulting in our newly developed reward models, named OpenPRM. This approach significantly enhances the scalability of process-level supervision in open domains at minimal cost. We assess the performance of OpenPRM across various reward benchmarks, demonstrating its competitive edge over traditional ORMs in open domains and PRMs in specialized domains. Additionally, we investigate the scalability of inference-time computation for open-domain instructions. Our results highlight the limitations of ORMs\u2019 scalability, while OpenPRM shows superior performance in scaled settings. Despite these advances, achieving automatic fine-grained supervision for open-domain inference-time scaling remains a substantial challenge. We hope these findings will spur further development of process supervision reward models in open-domain scenarios.", "keywords": [ "large language models", "reward models", "open-domain instruction following" ], "rating_list": [ 6, 5, 5, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "fFtmpqLFvw", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Cristina Menghini", "gender": "Female", "institution": "Scale AI", "country": "US", "position": "Researcher" }, { "name": "Hugh Zhang", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Ian Steneker", "gender": "Male", "institution": "University of Waterloo", "country": "", "position": "Undergrad student" }, { "name": "Nathaniel Li", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Riley Goodside", "gender": "Male", "institution": "Scale AI", "country": "US", "position": "Staff Prompt Engineer" }, { "name": "Summer Yue", "gender": "Female", "institution": "Scale AI", "country": "US", "position": "Researcher" }, { "name": "Willow E. Primack", "gender": "unknown", "institution": "Scale AI", "country": "US", "position": "Researcher" }, { "name": "Zifan Wang", "gender": "Male", "institution": "Scale AI", "country": "US", "position": "Researcher" }, { "name": "Ziwen Han", "gender": "Not Specified", "institution": "Scale AI", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 98, "n_ref_all": 116, "n_fig": 5, "n_tab": 5, "L_tab": 11506, "n_element_tab": 32, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1539, "n_element_tab_1": 106, "formula_len_all": 1, "formula_len_all_1": 0, "len_all": 230447, "len_all_1": 62661, "len_abs": 1205, "len_title": 108, "len_sents": 69924, "len_sents_1": 30745, "n_sents": 528, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1002, "title": "Uncovering Model Vulnerabilities With Multi-Turn Red Teaming", "abs": "Recent large language model (LLM) defenses have greatly improved models' ability to refuse harmful queries, even when adversarially attacked. However, LLM defenses are primarily evaluated against automated adversarial attacks in a single turn of conversation, an insufficient threat model for real-world malicious use. We demonstrate that multi-turn human jailbreaks uncover significant vulnerabilities, exceeding 70% attack success rate (ASR) on HarmBench against defenses that report single-digit ASRs with automated single-turn attacks. Human jailbreaks also reveal vulnerabilities in machine unlearning defenses, successfully recovering dual-use biosecurity knowledge from unlearned models. We compile these results into Multi-Turn Human Jailbreaks (MHJ), a dataset of 2,912 prompts across 537 multi-turn jailbreaks. We publicly release MHJ alongside a compendium of jailbreak tactics developed across dozens of commercial red teaming engagements, supporting research towards stronger LLM defenses.", "keywords": [ "language models", "ai security", "ai safety", "robustness", "adversarial attacks" ], "rating_list": [ 8, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "fErm1seIom", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Guoting Chen", "gender": "Male", "institution": "Great Bay University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaojie Zhang", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yingyi Fu", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Zipo Jibao", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 19, "n_formula_1": 12, "n_ref_uni": 30, "n_ref": 39, "n_ref_all": 53, "n_fig": 9, "n_tab": 3, "L_tab": 1258, "n_element_tab": 149, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3758, "n_element_tab_1": 176, "formula_len_all": 883, "formula_len_all_1": 483, "len_all": 129136, "len_all_1": 53121, "len_abs": 1323, "len_title": 107, "len_sents": 48220, "len_sents_1": 22880, "n_sents": 353, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1329, "title": "FMP-AE: A HYBRID APPROACH TO TIME SERIES ANOMALY DETECTION", "abs": "Unsupervised anomaly detection in time series presents significant challenges, especially due to the lack of labeled data and the prevalence of highly imbalanced datasets. Traditional statistical and machine learning methods often suffer from low recall and computational inefficiency. While deep learning techniques can automatically extract features, they still struggle with data imbalance. This paper introduces a novel anomaly detection model, Feature map Matrix Profile with an AutoEncoder (FMP-AE), which integrates matrix profile techniques with deep learning. The model uses a 1D-CNN to extract features and compute the matrix profile. A new Matrix Profile loss function is introduced and combined with the Autoencoder's reconstruction loss to enhance anomaly detection. The approach also incorporates a sliding window technique to improve sensitivity to sparse anomalies and increase efficiency. Experimental results on the UCR250 benchmark datasets demonstrate the model's superior performance across multiple metrics, including accuracy, precision, recall, F1-score, and AUC. These results highlight the FMP-AE model's ability to efficiently process large-scale datasets and generalize well across diverse time series domains, offering significant improvements in both detection accuracy and computational efficiency.", "keywords": [ "Anomaly detection; Matrix Profile; Deep learning;" ], "rating_list": [ 5, 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 1, 2, 1 ], "presentation_list": [ 3, 3, 3, 2, 2 ], "contribution_list": [ 3, 1, 1, 1, 1 ], "confidence_list": [ 4, 4, 4, 5, 4 ] }, { "paper_id": "fEEbTDoecM", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Abhinav Bhatia", "gender": "Male", "institution": "Department of Computer Science, University of Massachusetts, Amherst", "country": "", "position": "PhD student" }, { "name": "Samer B. Nashed", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Postdoc" }, { "name": "Shlomo Zilberstein", "gender": "Male", "institution": "University of Massachusetts Amherst", "country": "", "position": "" } ], "n_formula": 24, "n_formula_1": 10, "n_ref_uni": 53, "n_ref": 92, "n_ref_all": 115, "n_fig": 9, "n_tab": 3, "L_tab": 1398, "n_element_tab": 46, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 140, "n_element_tab_1": 14, "formula_len_all": 2033, "formula_len_all_1": 518, "len_all": 177378, "len_all_1": 78463, "len_abs": 1110, "len_title": 110, "len_sents": 58872, "len_sents_1": 38235, "n_sents": 394, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1124, "title": "RL$^3$: Boosting Meta Reinforcement Learning via RL inside RL$^2$", "abs": "Meta reinforcement learning (meta-RL) methods such as \\rlsquare have emerged as promising approaches for learning data-efficient RL algorithms tailored to a given task distribution. However, they show poor asymptotic performance and struggle with out-of-distribution tasks because they rely on sequence models, such as recurrent neural networks or transformers, to process experiences rather than summarize them using general-purpose RL components such as value functions. In contrast, traditional RL algorithms are data-inefficient as they do not use domain knowledge, but do converge to an optimal policy in the limit. We propose RL$^3$, a principled hybrid approach that incorporates action-values, learned per task via traditional RL, in the inputs to meta-RL. We show that RL$^3$ earns greater cumulative reward in the long term compared to RL$^2$ while drastically reducing meta-training time and generalizes better to out-of-distribution tasks. Experiments are conducted on both custom and benchmark discrete domains from the meta-RL literature that exhibit a range of short-term, long-term, and complex dependencies.", "keywords": [ "meta-reinforcement learning" ], "rating_list": [ 5, 5, 6, 5, 5 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 4, 5 ] }, { "paper_id": "fDcn3S8oAt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Archiki Prasad", "gender": "Female", "institution": "University of North Carolina, Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Duy Nguyen", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Elias Stengel-Eskin", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Postdoc" }, { "name": "Mohit Bansal", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 65, "n_ref": 136, "n_ref_all": 180, "n_fig": 9, "n_tab": 7, "L_tab": 2555, "n_element_tab": 267, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1005, "n_element_tab_1": 124, "formula_len_all": 744, "formula_len_all_1": 559, "len_all": 223057, "len_all_1": 72384, "len_abs": 2075, "len_title": 123, "len_sents": 60101, "len_sents_1": 34356, "n_sents": 452, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 2104, "title": "LASeR: Learning to Adaptively Select Reward Models with Multi-Armed Bandits", "abs": "Reward Models (RMs) play a crucial role in aligning large language models (LLMs) with human preferences, enhancing their performance by ranking outputs during inference or iterative training. However, the degree to which an RM generalizes to new tasks is often not known *a priori*. For instance, some RMs may excel at scoring creative writing, while others specialize in evaluating math reasoning. Therefore, using only one fixed RM while training LLMs can be *suboptimal*. Moreover, optimizing LLMs with multiple RMs simultaneously can be prohibitively computationally-intensive and challenging due to conflicting signals from different RMs, potentially degrading performance. To address these challenges, we introduce LASeR (**L**earning to **A**daptively **Se**lect **R**ewards), which iteratively trains LLMs using multiple RMs, selecting and utilizing the most well-suited RM for each instance to rank outputs and generate preference data, framed as a multi-armed bandit problem. Our empirical results on commonsense and math reasoning tasks demonstrate that LASeR can boost iterative LLM optimization by optimizing for multiple RMs, improving the absolute average accuracy of Llama-3-8B over three datasets by 2.67% over training with ensemble RM scores while also showing superior training efficiency (e.g., a 2x speedup). Moreover, on WildChat, a benchmark of instruction-following prompts in open-form generation, we find that using Llama-3-8B LASeR leads to a 71.45% AlpacaEval win rate over sequentially optimizing multiple RMs. Extending to long-context generation tasks, we find that on Llama-3-8B, LASeR achieves an average improvement of 2.64 F1 points on single-document QA tasks and 2.42 F1 points on multi-document QA over random RM selection when used with best-of-n sampling. Our analysis shows that LASeR is robust to noisy rewards and generalizes to multiple settings. Finally, we demonstrate that LASeR's RM selection changes depending on the underlying task or instance and we verify the presence of conflicting preferences from multiple RMs that can be mitigated using LASeR.", "keywords": [ "Multi-armed bandits", "Preference optimization", "Reward model", "Iterative LLM training" ], "rating_list": [ 8, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "fDNBPqgr4K", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Dezhi Luo", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Undergrad student" }, { "name": "Haiyun Lyu", "gender": "Female", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Undergrad student" }, { "name": "Haoran Sun", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "MS student" }, { "name": "Hokin Deng", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Researcher" }, { "name": "Qingying Gao", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Yijiang Li", "gender": "Not Specified", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 121, "n_ref": 162, "n_ref_all": 172, "n_fig": 4, "n_tab": 2, "L_tab": 1250, "n_element_tab": 125, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1250, "n_element_tab_1": 125, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 171176, "len_all_1": 55577, "len_abs": 736, "len_title": 129, "len_sents": 30270, "len_sents_1": 26352, "n_sents": 177, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 816, "title": "CogDevelop2K: Reversed Cognitive Development in Multi-modal Large Language Models", "abs": "Are Multi-modal Large Language Models (MLLMs) stochastic parrots? Do they genuinely understand and are capable of performing the tasks they excel at? This paper aims to explore the fundamental basis of MLLMs, i.e. core cognitive abilities that human intelligence builds upon to perceive, comprehend, and reason. To this end, we propose CogDevelop2K, a comprehensive benchmark that spans 12 sub-concepts from fundamental knowledge like object permanence and boundary to advanced reasoning like intentionality understanding, structured via the developmental trajectory of a human mind. We evaluate 46 MLLMs on our benchmarks. Comprehensively, we further evaluate the influence of evaluation strategies and prompting techniques. Surprisingly, we observe a reversed cognitive developmental trajectory compared to humans.", "keywords": [ "Vision Language Model", "Multi-modal Large Language Model", "Cognitive Development", "Cognitive Science", "Benchmark" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "fDGPIuCdGi", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Jiang Bian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Lei Song", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Linjie Xu", "gender": "Male", "institution": "School of Electronic Engineering and Computer Science", "country": "GB", "position": "PhD student" }, { "name": "Pengcheng You", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ruohong Liu", "gender": "Female", "institution": "The Hong Kong University of Science and Technology, Guangzhou", "country": "CN", "position": "MS student" }, { "name": "Yize Chen", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Assistant Professor" }, { "name": "Yuxin Pan", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 60, "n_formula_1": 12, "n_ref_uni": 45, "n_ref": 109, "n_ref_all": 135, "n_fig": 8, "n_tab": 12, "L_tab": 4235, "n_element_tab": 504, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 593, "n_element_tab_1": 70, "formula_len_all": 3509, "formula_len_all_1": 659, "len_all": 206831, "len_all_1": 76684, "len_abs": 1588, "len_title": 78, "len_sents": 65811, "len_sents_1": 36620, "n_sents": 562, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1591, "title": "Efficient Discovery of Pareto Front for Multi-Objective Reinforcement Learning", "abs": "Multi-objective reinforcement learning (MORL) excels at handling rapidly changing preferences in tasks that involve multiple criteria, even for unseen preferences. However, previous dominating MORL methods typically generate a fixed policy set or preference-conditioned policy through multiple training iterations exclusively for sampled preference vectors, and cannot ensure the efficient discovery of the Pareto front. Furthermore, integrating preferences into the input of policy or value functions presents scalability challenges, in particular as the dimension of the state and preference space grow, which can complicate the learning process and hinder the algorithm's performance on more complex tasks. To address these issues, we propose a two-stage Pareto front discovery algorithm called Constrained MORL (C-MORL), which serves as a seamless bridge between constrained policy optimization and MORL. Concretely, a set of policies is trained in parallel in the initialization stage, with each optimized towards its individual preference over the multiple objectives. Then, to fill the remaining vacancies in the Pareto front, the constrained optimization steps are employed to maximize one objective while constraining the other objectives to exceed a predefined threshold. Empirically, compared to recent advancements in MORL methods, our algorithm achieves more consistent and superior performances in terms of hypervolume, expected utility, and sparsity on both discrete and continuous control tasks, especially with numerous objectives (up to nine objectives in our experiments).", "keywords": [ "multi-objective reinforcement learning", "constrained reinforcement learning" ], "rating_list": [ 5, 3, 8, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 1, 2, 3, 1 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "fD8Whiy7ca", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Jonathan Kadmon", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "Assistant Professor" }, { "name": "Maher Hanut", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 30, "n_ref_uni": 22, "n_ref": 35, "n_ref_all": 74, "n_fig": 15, "n_tab": 2, "L_tab": 1185, "n_element_tab": 25, "n_fig_1": 13, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15231, "formula_len_all_1": 21826, "len_all": 147265, "len_all_1": 102194, "len_abs": 1773, "len_title": 114, "len_sents": 47412, "len_sents_1": 36436, "n_sents": 379, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1782, "title": "Training Large Neural Networks With Low-Dimensional Error Feedback", "abs": "Training deep neural networks typically relies on backpropagating high-dimensional error signals\u2014a computationally intensive process with little evidence supporting its implementation in the brain. However, since most tasks involve low-dimensional outputs, we propose that low-dimensional error signals may suffice for effective learning. To test this hypothesis, we introduce a novel local learning rule based on Feedback Alignment that leverages indirect, low-dimensional error feedback to train large networks. Our method decouples the backward pass from the forward pass, enabling precise control over error signal dimensionality while maintaining high-dimensional representations. We begin with a detailed theoretical derivation for linear networks, which forms the foundation of our learning framework, and extend our approach to nonlinear and convolutional architectures. Remarkably, we demonstrate that even minimal error dimensionality\u2014on the order of the task dimensionality\u2014can achieve performance matching that of traditional backpropagation. Furthermore, our rule enables efficient training of convolutional networks, which have previously been resistant to Feedback Alignment methods, with minimal error. This breakthrough not only paves the way towards more biologically accurate models of learning but also challenges the conventional reliance on high-dimensional gradient signals in neural network training. Our findings suggest that low-dimensional error signals can be as effective as high-dimensional ones, prompting a reevaluation of gradient-based learning in high-dimensional systems. Ultimately, our work offers a fresh perspective on neural network optimization and contributes to understanding learning mechanisms in both artificial and biological systems.", "keywords": [ "Feedback Alignment", "Brain Inspired", "Biologically Plausible Learning", "Theoretical Analysis", "Visual Pathway" ], "rating_list": [ 5, 8, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "fCi4o83Mfs", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Arman Cohan", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" }, { "name": "Chuhan Li", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Grad student" }, { "name": "Tesca Fitzgerald", "gender": "Female", "institution": "Yale University", "country": "", "position": "Assistant Professor" }, { "name": "Yanan Zheng", "gender": "Female", "institution": "Yale University", "country": "US", "position": "Postdoc" }, { "name": "Yilun Zhao", "gender": "unknown", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Yuxuan Ding", "gender": "Male", "institution": "Yale University", "country": "US", "position": "MS student" }, { "name": "Ziyao Shangguan", "gender": "Female", "institution": "Yale University", "country": "US", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 22, "n_ref": 83, "n_ref_all": 118, "n_fig": 3, "n_tab": 13, "L_tab": 7269, "n_element_tab": 834, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 3680, "n_element_tab_1": 455, "formula_len_all": 472, "formula_len_all_1": 225, "len_all": 447405, "len_all_1": 70014, "len_abs": 3895, "len_title": 137, "len_sents": 161178, "len_sents_1": 30099, "n_sents": 1510, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 67, "L_abs": 1762, "title": "Can Multimodal Foundation Models Perform Visual Temporal Reasoning?", "abs": "Existing benchmarks often highlight the remarkable performance achieved by state-of-the-art Multimodal Foundation Models (MFMs) in leveraging temporal context for video understanding.\nHowever, *how well do the models truly perform visual temporal reasoning*?\nOur study of existing benchmarks shows that this capability of MFMs is likely overestimated as many questions can be solved by using a single, few, or out-of-order frames.\nTo systematically examine current visual temporal reasoning tasks, we propose three principles with corresponding metrics:\n(1) *Multi-Frame Gain*,\n(2) *Frame Order Sensitivity*,\nand (3) *Frame Information Disparity*.\nFollowing these principles, we introduce **TVBench**, **T**emporal Reasoning **V**ideo Understanding **Bench**mark, a novel benchmark crafted to rigorously assess MFMs' temporal reasoning capabilities in video understanding.\nTVBench comprises 1,484 carefully curated, *human-annotated* questions spanning six tasks (i.e. *action count, direction, rotation, shape & trend, velocity & frequency, and visual cues*), applied to 1,417 videos, including 805 self-recorded and -generated videos, that encompass human-centric, real-world, and simulated scenarios. \nOur comprehensive evaluation reveals a human-model performance gap of 57.3% with the best-performing model.\nMoreover, our in-depth analysis uncovers more fundamental limitations beyond this gap in current MFMs. While they can accurately recognize events in isolated frames, they fail to interpret these frames as a continuous sequence.\nWe believe TVBench will serve as a crucial testbed for evaluating the next-generation MFMs and as a call to the community to develop AI systems capable of comprehending the human world dynamics through the video modality.", "keywords": [ "visual temporal reasoning", "video understanding", "benchmark", "vision-language benchmark", "video-language models", "evaluation" ], "rating_list": [ 6, 8, 8, 5 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 3, 4, 4, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "fBkdjUnymd", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Bo Dai", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "ChangHao Li", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Chao Zhang", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Hanjun Dai", "gender": "Male", "institution": "Google Research", "country": "", "position": "Researcher" }, { "name": "Haotian Sun", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Rushi Qiang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yuchen Zhuang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 48, "n_ref": 114, "n_ref_all": 134, "n_fig": 5, "n_tab": 7, "L_tab": 6550, "n_element_tab": 219, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 494, "n_element_tab_1": 6, "formula_len_all": 1168, "formula_len_all_1": 1114, "len_all": 263987, "len_all_1": 56365, "len_abs": 1700, "len_title": 102, "len_sents": 92616, "len_sents_1": 26432, "n_sents": 707, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1510, "title": "Martryoshka: Learning to Drive Black-Box LLMs with LLMs", "abs": "Despite the impressive generative abilities of black-box large language models (LLMs), their inherent opacity hinders further advancements in capabilities such as reasoning, planning, and personalization. Existing works aim to enhance LLM capabilities via domain-specific adaptation or in-context learning, which require additional training on accessible model parameters, an infeasible option for black-box LLMs. To address this challenge, we introduce Martryoshika, a lightweight white-box LLM controller that guides a large-scale black-box LLM generator by decomposing complex tasks into a series of intermediate outputs. Specifically, we consider the black-box LLM as an environment, with Martryoshika serving as a policy to provide intermediate guidance through prompts for driving the black-box LLM. Martryoshika is trained to pivot the outputs of the black-box LLM aligning with preferences during iterative interaction, which enables controllable multi-turn generation and self-improvement in optimizing intermediate guidance. Empirical evaluations on three diverse tasks demonstrate that Martryoshika effectively enhances the capabilities of black-box LLMs in complex, long-horizon tasks, including reasoning, planning, and personalization. By leveraging this pioneering controller-generator framework to mitigate dependence on model parameters, Martryoshika provides a transparent and practical solution for improving black-box LLMs through controllable multi-turn generation using white-box LLMs.", "keywords": [ "Large Language Model", "Reasoning and Planning", "LLM Controller" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "fBhgu6PsA7", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Haozhe Zhang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Ruoyu Sun", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Senmiao Wang", "gender": "Male", "institution": "Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Tian Ding", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Researcher" }, { "name": "Weijian Sun", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Yupeng Chen", "gender": "Male", "institution": "The Chinese University of Hong Kong (ShenZhen))", "country": "CN", "position": "PhD student" }, { "name": "Yushun Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Zhihang Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "MS student" } ], "n_formula": 64, "n_formula_1": 13, "n_ref_uni": 54, "n_ref": 102, "n_ref_all": 141, "n_fig": 11, "n_tab": 10, "L_tab": 1976, "n_element_tab": 202, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 429, "n_element_tab_1": 79, "formula_len_all": 8518, "formula_len_all_1": 843, "len_all": 247748, "len_all_1": 69737, "len_abs": 3433, "len_title": 127, "len_sents": 55659, "len_sents_1": 31668, "n_sents": 521, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1122, "title": "MoFO: Momentum-Filtered Optimizer for Mitigating Forgetting in LLM Fine-Tuning", "abs": "Large language models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks. Typically, an LLM is first pre-trained on large corpora and subsequently fine-tuned on task-specific datasets. However, during fine-tuning, LLMs may forget some knowledge acquired in the pre-training stage, leading to a decline in general capabilities. To address this challenge, we propose a new fine-tuning algorithm termed Momentum-Filtered Optimizer (MoFO). \nAs an extension of greedy block coordinate descent (BCD) methods, MoFO iteratively selects and updates the model parameters with the largest momentum magnitudes.\nMoFO achieves similar fine-tuning performance to the default fine-tuning algorithm while effectively mitigating knowledge forgetting.\nFurthermore, MoFO does not require access to pre-training data, making it highly suitable for scenarios where the pre-training data is unavailable, such as fine-tuning checkpoint-only open-source LLMs. \nWe validate MoFO through rigorous convergence analysis and extensive experiments, demonstrating its superiority over existing methods in mitigating forgetting.", "keywords": [ "LLM fine-tuning", "catastrophic forgetting" ], "rating_list": [ 5, 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 4, 3 ] }, { "paper_id": "fBSc0c1IXJ", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Deniz Gunduz", "gender": "unknown", "institution": "Imperial College London", "country": "", "position": "Full Professor" }, { "name": "Szymon Kobus", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 39, "n_ref": 51, "n_ref_all": 65, "n_fig": 4, "n_tab": 6, "L_tab": 3821, "n_element_tab": 201, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1469, "n_element_tab_1": 70, "formula_len_all": 378, "formula_len_all_1": 370, "len_all": 112857, "len_all_1": 61855, "len_abs": 1641, "len_title": 108, "len_sents": 31004, "len_sents_1": 29897, "n_sents": 214, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1650, "title": "Remote Reinforcement Learning with Communication Constraints", "abs": "We introduce the novel problem of remote reinforcement learning (RRL) with a communication constraint, in which the actor that takes the actions in the environment lacks direct access to the reward signal. Instead, the rewards are observed by a controller, which communicates with the agent through a communication-constrained channel. This can model a remote control scenario over a wireless channel, where the communication link from the controller to the agent has limited capacity due to power, bandwidth, or delay constraints. In the proposed solution, rather than transmitting the reward values to the agent over the rate-limited channel, the controller learns the optimal policy, and at each round, signals the action that the agent should take over the channel. However, instead of sending the precise action--which can be prohibitive when the action set is large--we use an importance sampling approach to reduce the communication load, which allows the agent to sample an action from the current policy. The actor, sampling from the desired policy at each turn, can also learn the optimal policy, albeit at a slower pace, using supervised learning. We exploit the learned policy at the actor to further reduce the communication load. Our solution, called Guided Remote Action Sampling Policy (GRASP), exhibits a significant reduction in communication requirements, achieving an average of 12-fold decrease in data transmission across all experiments, and 50-fold reduction for environments with continuous action spaces. We also show the applicability of GRASP beyond single-agent scenarios, including parallel and multi-agent environments.", "keywords": [ "reinforcement learning", "communication", "source coding", "compression", "sampling", "channel simulation" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 1, 2, 2 ], "presentation_list": [ 2, 3, 1, 3, 2 ], "contribution_list": [ 3, 1, 1, 2, 2 ], "confidence_list": [ 3, 4, 3, 4, 3 ] }, { "paper_id": "fBJo3wwZeJ", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Dominic Phillips", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Flaviu Cipcigan", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" } ], "n_formula": 38, "n_formula_1": 11, "n_ref_uni": 45, "n_ref": 75, "n_ref_all": 96, "n_fig": 12, "n_tab": 1, "L_tab": 136, "n_element_tab": 42, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3625, "formula_len_all_1": 1084, "len_all": 181630, "len_all_1": 66163, "len_abs": 1019, "len_title": 131, "len_sents": 52742, "len_sents_1": 31027, "n_sents": 458, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1022, "title": "MetaGFN: Exploring Distant Modes with Adapted Metadynamics for Continuous GFlowNets", "abs": "Generative Flow Networks (GFlowNets) are a class of generative models that sample objects in proportion to a specified reward function through a learned policy. They can be trained either on-policy or off-policy, needing a balance between exploration and exploitation for fast convergence to a target distribution. While exploration strategies for discrete GFlowNets have been studied, exploration in the continuous case remains to be investigated, despite the potential for novel exploration algorithms due to the local connectedness of continuous domains. Here, we introduce Adapted Metadynamics, a variant of metadynamics that can be applied to arbitrary black-box reward functions on continuous domains. We use Adapted Metadynamics as an exploration strategy for continuous GFlowNets. We show several continuous domains where the resulting algorithm, MetaGFN, accelerates convergence to the target distribution and discovers more distant reward modes than previous off-policy exploration strategies used for GFlowNets.", "keywords": [ "GFlowNets", "metadynamics", "exploration", "sampling" ], "rating_list": [ 3, 6, 3, 5, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 1, 2, 3 ], "confidence_list": [ 5, 3, 4, 2, 2 ] }, { "paper_id": "fAvG3P3Cxx", "primary_area": "learning on time series and dynamical systems", "n_author": 7, "author_info_list": [ { "name": "Bin Yang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Chenjuan Guo", "gender": "unknown", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Jilin Hu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Kai Zhao", "gender": "Non-Binary", "institution": "Aalborg University", "country": "DK", "position": "PhD student" }, { "name": "Xiangfei Qiu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Yang Shu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Assistant Professor" }, { "name": "syhu", "gender": "Female", "institution": "East China Normal University", "country": "CN", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 18, "n_ref_uni": 33, "n_ref": 73, "n_ref_all": 99, "n_fig": 14, "n_tab": 5, "L_tab": 1200, "n_element_tab": 113, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 2494, "n_element_tab_1": 29, "formula_len_all": 1383, "formula_len_all_1": 1092, "len_all": 139261, "len_all_1": 61270, "len_abs": 1409, "len_title": 161, "len_sents": 42351, "len_sents_1": 27113, "n_sents": 347, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 113, "L_abs": 925, "title": "MultiRC: Joint Learning for Time Series Anomaly Prediction and Detection with Multi-scale Reconstructive Contrast", "abs": "Many methods have been proposed for unsupervised time series anomaly detection. Despite some progress, research on predicting future anomalies is still relatively scarce. Predicting anomalies is particularly challenging due to the diverse reaction time and the lack of labeled data. To address these challenges, we propose MultiRC to integrate reconstructive and contrastive learning for joint learning of anomaly prediction and detection, with multi-scale structure and adaptive dominant period mask to deal with the diverse reaction time. MultiRC also generates negative samples to provide essential training momentum for the anomaly prediction tasks and prevent model degradation. We evaluate seven benchmark datasets from different fields. For both anomaly prediction and detection tasks, MultiRC outperforms existing state-of-the-art methods. The code is available at https://anonymous.4open.science/status/MultiRC-CCE6.", "keywords": [ "Time Series Anomaly Prediction", "Time Series Anomaly Detection", "Self-Supervised Learning" ], "rating_list": [ 3, 6, 6, 5, 3 ], "soundness_list": [ 2, 4, 2, 3, 2 ], "presentation_list": [ 1, 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 4, 3 ] }, { "paper_id": "fAAaT826Vv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Ben Zhou", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" }, { "name": "Dan Roth", "gender": "Male", "institution": "Oracle", "country": "US", "position": "Chief Scientist" }, { "name": "Weidong Lin", "gender": "Male", "institution": "The Wharton School, University of Pennsylvania", "country": "US", "position": "Researcher" }, { "name": "Yu Feng", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 14, "n_ref_uni": 36, "n_ref": 56, "n_ref_all": 102, "n_fig": 15, "n_tab": 5, "L_tab": 1878, "n_element_tab": 159, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1141, "n_element_tab_1": 139, "formula_len_all": 1813, "formula_len_all_1": 1369, "len_all": 189185, "len_all_1": 73056, "len_abs": 1046, "len_title": 122, "len_sents": 64448, "len_sents_1": 34014, "n_sents": 526, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 74, "L_abs": 1050, "title": "BIRD: A Trustworthy Bayesian Inference Framework for Large Language Models", "abs": "Predictive models often need to work with incomplete information in real-world tasks. Consequently, they must provide reliable probability or confidence estimation, especially in large-scale decision making and planning tasks. Current large language models (LLM) are insufficient for such accurate estimations, but they can generate relevant factors that may affect the probabilities, produce coarse-grained probabilities when the information is more complete, and help determine which factors are relevant to specific downstream contexts. In this paper, we make use of these capabilities of LLMs to provide a significantly more accurate probabilistic estimation. We propose BIRD, a novel probabilistic inference framework that aligns a Bayesian network with LLM abductions and then estimates more accurate probabilities in a deduction step. We show BIRD provides reliable probability estimations that are 30% better than those provided directly by LLM baselines. These estimates can further contribute to better and more trustworthy decision-making.", "keywords": [ "Large language models", "Reasoning", "Planning", "Trustworthiness", "Interpretability", "Probability Estimation", "Bayesian Methods" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 4, 3, 2 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 4, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "f9w89OY2cp", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Chengdong Ma", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Fengshuo Bai", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Haoran Sun", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Mingzhi Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Qizhi Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Yaodong Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhaowei Zhang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Zilong Zheng", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" } ], "n_formula": 41, "n_formula_1": 6, "n_ref_uni": 42, "n_ref": 93, "n_ref_all": 131, "n_fig": 9, "n_tab": 6, "L_tab": 2709, "n_element_tab": 497, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 2270, "n_element_tab_1": 483, "formula_len_all": 4499, "formula_len_all_1": 290, "len_all": 192313, "len_all_1": 71443, "len_abs": 1324, "len_title": 131, "len_sents": 59352, "len_sents_1": 31798, "n_sents": 468, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1332, "title": "Amulet: ReAlignment During Test Time for Personalized Preference Adaptation of LLMs", "abs": "How to align large language models (LLMs) with user preferences from a static general dataset has been frequently studied. However, user preferences are usually personalized, changing, and diverse. This leads to the problem that the actual user preferences often do not coincide with those trained by the model developers in the practical use of LLMs. Since we cannot collect enough data and retrain for every demand, researching efficient real-time preference adaptation methods based on the backbone LLMs during test time is important. To this end, we introduce **Amulet**, a novel, training-free framework that formulates the decoding process of every token as a separate online learning problem with the guidance of simple user-provided prompts, thus enabling real-time optimization to satisfy users' personalized preferences. To reduce the computational cost brought by this optimization process for each token, we additionally provide a closed-form solution for each iteration step of the optimization process, thereby reducing the computational time cost to a negligible level. The detailed experimental results demonstrate that Amulet can achieve significant performance improvements in rich settings with combinations of different LLMs, datasets, and user preferences, while maintaining acceptable computational efficiency.", "keywords": [ "AI Alignment", "LLM Test Time Preference Adaptation", "Alignment without Training and Fine-tuing" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 4, 3, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "f9GURUHZQo", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Aditya Akella", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Alex Dimakis", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department, University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Daehyeok Kim", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Donghyun Kim", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Sriram Ravula", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "PhD student" }, { "name": "Taemin Ha", "gender": "Male", "institution": ", University of Texas at Austin", "country": "US", "position": "Undergrad student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 44, "n_ref": 84, "n_ref_all": 127, "n_fig": 23, "n_tab": 3, "L_tab": 902, "n_element_tab": 66, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 827, "n_element_tab_1": 86, "formula_len_all": 62, "formula_len_all_1": 62, "len_all": 176262, "len_all_1": 72281, "len_abs": 1726, "len_title": 112, "len_sents": 56803, "len_sents_1": 36094, "n_sents": 407, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1546, "title": "Large Language Models as Realistic Microservice Trace Generators", "abs": "Computer system workload traces, which record hardware or software events during application execution, are essential for understanding the behavior of complex systems and managing their processing and memory resources. However, obtaining real-world traces can be challenging due to the significant collection overheads in performance and privacy concerns that arise in proprietary systems. As a result, synthetic trace generation is considered a promising alternative to using traces collected in real-world production deployments. This paper proposes to train a large language model (LLM) to generate synthetic workload traces, specifically microservice call graphs. To capture complex and arbitrary hierarchical structures and implicit constraints in such traces, we fine-tune LLMs to generate each layer recursively, making call graph generation a sequence of easier steps. To further enforce learning constraints in traces and generate uncommon situations, we apply additional instruction tuning steps to align our model with the desired trace features. Our evaluation results show that our model can generate diverse realistic traces under various conditions and outperform existing methods in accuracy and validity. We show that our synthetically generated traces can effectively substitute real-world data in optimizing or tuning systems management tasks. We also show that our model can be adapted to perform key downstream trace-related tasks, specifically, predicting key trace features and infilling missing data given partial traces.", "keywords": [ "synthetic data", "synthetic trace", "microservice", "large language model", "machine learning for systems" ], "rating_list": [ 8, 6, 3, 6 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "f92M45YRfh", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jianfeng Lu", "gender": "Male", "institution": "Nanjing university of Science & Technology", "country": "CN", "position": "Full Professor" }, { "name": "Qiwei Wang", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" }, { "name": "Xianghui Ze", "gender": "unknown", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yujiao Shi", "gender": "Female", "institution": "ShanghaiTech University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhenbo Song", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 55, "n_ref_all": 81, "n_fig": 11, "n_tab": 8, "L_tab": 2068, "n_element_tab": 223, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 4987, "n_element_tab_1": 171, "formula_len_all": 535, "formula_len_all_1": 535, "len_all": 148392, "len_all_1": 59076, "len_abs": 2064, "len_title": 160, "len_sents": 46559, "len_sents_1": 25520, "n_sents": 330, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 111, "L_abs": 1215, "title": "Controllable Satellite-to-Street-View Synthesis with Precise Pose Alignment and Zero-Shot Environmental Control", "abs": "Generating street-view images from satellite imagery is a challenging task, particularly in maintaining accurate pose alignment and incorporating diverse environmental conditions. While diffusion models have shown promise in generative tasks, their ability to maintain strict pose alignment throughout the diffusion process is limited. In this paper, we propose a novel Iterative Homography Adjustment (IHA) scheme applied during the denoising process, which effectively addresses pose misalignment and ensures spatial consistency in the generated street-view images. Additionally, currently, available datasets for satellite-to-street-view generation are limited in their diversity of illumination and weather conditions, thereby restricting the generalizability of the generated outputs. To mitigate this, we introduce a text-guided illumination and weather-controlled sampling strategy that enables fine-grained control over the environmental factors. Extensive quantitative and qualitative evaluations demonstrate that our approach significantly improves pose accuracy and enhances the diversity and realism of generated street-view images, setting a new benchmark for satellite-to-street-view generation tasks.", "keywords": [ "Satellite to street-view synthesis", "diffusion model", "controllable", "precise pose Alignment" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "f89YIjbuRC", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Andreas Schliebitz", "gender": "Male", "institution": "Osnabr\u00fcck University of Applied Sciences", "country": "DE", "position": "Researcher" }, { "name": "Heiko Tapken", "gender": "Male", "institution": "Hochschule Osnabr\u00fcck University of Applied Sciences", "country": "DE", "position": "Full Professor" }, { "name": "Martin Atzmueller", "gender": "Not Specified", "institution": "Universit\u00e4t Osnabr\u00fcck", "country": "DE", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 6, "n_ref": 25, "n_ref_all": 47, "n_fig": 1, "n_tab": 6, "L_tab": 2569, "n_element_tab": 202, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 2413, "n_element_tab_1": 200, "formula_len_all": 202, "formula_len_all_1": 201, "len_all": 84194, "len_all_1": 71590, "len_abs": 1716, "len_title": 124, "len_sents": 37057, "len_sents_1": 35047, "n_sents": 228, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1739, "title": "Improving Nonlinear Projection Heads using Pretrained Autoencoder Embeddings", "abs": "This empirical study aims at improving the effectiveness of the standard 2-layer MLP projection head $g(\\cdot)$ featured in the SimCLR framework through the use of pretrained autoencoder embeddings. Given a contrastive learning task with a largely unlabeled image classification dataset, we first train a shallow autoencoder architecture and extract its compressed representations contained in the encoder's embedding layer. After freezing the weights within this pretrained layer, we use it as a drop-in replacement for the input layer of SimCLR's default projector. Additionally, we also apply further architectural changes to the projector by decreasing its width and changing its activation function. The different projection heads are then used to contrastively train and evaluate a feature extractor $f(\\cdot)$ following the SimCLR protocol, while also examining the performance impact of $Z$-score normalized datasets. Our experiments indicate that using a pretrained autoencoder embedding in the projector can not only increase classification accuracy by up to 2.9% or 1.7% on average but can also significantly decrease the dimensionality of the projection space. Our results also suggest, that using the sigmoid and $\\tanh$ activation functions within the projector can outperform ReLU in terms of peak and average classification accuracy. When applying our presented projectors, then not applying $Z$-score normalization to datasets often increases peak performance. In contrast, the default projection head can benefit more from normalization. All experiments involving our pretrained projectors are conducted with frozen embeddings, since our test results indicate an advantage compared to using their non-frozen counterparts.", "keywords": [ "Nonlinear Projection Heads", "Multilayer Perceptrons", "Autoencoder Embeddings", "SimCLR Framework", "Contrastive Learning", "Representation Learning" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "f7aWmxgSN4", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "David D. Baek", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Max Tegmark", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Yuxiao Li", "gender": "Female", "institution": "Basque Center for Applied Mathematics", "country": "ES", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 39, "n_ref": 46, "n_ref_all": 60, "n_fig": 10, "n_tab": 2, "L_tab": 156, "n_element_tab": 23, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 788, "formula_len_all_1": 814, "len_all": 92218, "len_all_1": 59254, "len_abs": 1017, "len_title": 133, "len_sents": 25252, "len_sents_1": 23637, "n_sents": 176, "n_sents_1": 162, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1037, "title": "Generalization from Starvation: Hints of Universality in LLM Knowledge Graph Learning", "abs": "Motivated by interpretability and reliability, we investigate how neural networks represent knowledge during graph learning. We find hints of universality, where equivalent representations are learned across a range of model sizes (from $10^2$ to $10^9$ parameters) and contexts (MLP toy models, LLM in-context learning and LLM training). We show that these attractor representations optimize generalization to unseen examples by exploiting properties of knowledge graph relations (e.g. symmetry and meta-transitivity). We find experimental support for such universality by showing that LLMs and simpler neural networks can be successfully stitched, i.e., by stitching the first part of one model to the last part of another, mediated only by an affine or almost affine transformation. We hypothesize that this dynamic toward simplicity and generalization is driven by ``intelligence from starvation\u201d: where overfitting is minimized by pressure to minimize the use of resources that are either scarce or competed for against other tasks.", "keywords": [ "Knowledge Graph Learning", "Interpretability", "Generalization", "Universality" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 1, 1, 1 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "f7Zq9CqQEM", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "yang changyuan eric", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Guang Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jiahui Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Lingyun Sun", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Shengyuan Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Tianrun Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Tianyun Zhang", "gender": "unknown", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Zejian Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhiyuan Yang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 14, "n_ref_uni": 47, "n_ref": 91, "n_ref_all": 110, "n_fig": 12, "n_tab": 3, "L_tab": 550, "n_element_tab": 30, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 294, "n_element_tab_1": 23, "formula_len_all": 1126, "formula_len_all_1": 756, "len_all": 162471, "len_all_1": 54263, "len_abs": 1290, "len_title": 161, "len_sents": 42440, "len_sents_1": 24406, "n_sents": 337, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 112, "L_abs": 1342, "title": "Path-Tracing Distillation: Enhancing Stability in Text-to-3D Generation by Mitigating Out-of-Distribution Issues", "abs": "Text-to-3D generation techniques signify a pivotal advancement in creating 3D models from textual descriptions. Contemporary state-of-the-art methods utilize score distillation processes, leveraging 2D priors to generate 3D assets. However, these approaches frequently encounter instability during the initial generation phases, primarily due to an distribution divergence between the pretrained score prediction network and the nascent 3D model. Specifically, raw rendered images of an initial 3D model lie outside the distribution (OOD) of the pretrained score prediction network, which is trained on high-fidelity realistic images. To address this OOD issue, we introduce an innovative Path-Tracing Distillation (PTD) technique that refines the distillation process. Our method sequentially optimizes the 3D model using intermediate score networks that exhibit closer distributional alignment, thereby accelerating the convergence during the early stages of training. This approach not only ensures a more stable increase in CLIP similarity initially but also preserves the visual quality and diversity of the generated models. Experiments demonstrate that PTD significantly enhances both the stability and quality of text-to-3D generation, outperforming existing baselines. PTD can also be generalized to other score distillation methods.", "keywords": [ "Text-to-3D Generation", "Score Distillation", "Path-Tracing", "3D Model Stability" ], "rating_list": [ 3, 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 4, 5 ] }, { "paper_id": "f7WBRSuf9l", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Conghui He", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Haodong Duan", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Postdoc" }, { "name": "Jiaqi Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Pan Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xiaoyi Dong", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yuanjun Xiong", "gender": "Male", "institution": "Predera.AI", "country": "US", "position": "Researcher" }, { "name": "Yuhang Zang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Ziyu Liu", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Undergrad student" }, { "name": "Yuhang CAO", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 31, "n_ref": 107, "n_ref_all": 143, "n_fig": 12, "n_tab": 15, "L_tab": 4593, "n_element_tab": 466, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 1, "n_element_tab_1": 1, "formula_len_all": 321, "formula_len_all_1": 349, "len_all": 239434, "len_all_1": 60441, "len_abs": 2312, "len_title": 142, "len_sents": 67192, "len_sents_1": 26800, "n_sents": 606, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1631, "title": "MIA-DPO: Multi-Image Augmented Direct Preference Optimization For Large Vision-Language Models", "abs": "Visual preference alignment involves training Large Vision-Language Models (LVLMs) to predict human preferences between visual inputs. This is typically achieved by using labeled datasets of chosen/rejected pairs and employing optimization algorithms like direct preference optimization (DPO).\nExisting visual alignment methods, primarily designed for single-image scenarios, struggle to effectively handle the complexity of multi-image tasks due to the scarcity of diverse training data and the high cost of annotating chosen/rejected pairs.\nWe present Multi-Image Augmented Direct Preference Optimization (MIA-DPO), a visual preference alignment approach that effectively handles multi-image inputs.\nMIA-DPO mitigates the scarcity of diverse multi-image training data by extending single-image data with unrelated images arranged in grid collages or pic-in-pic formats, significantly reducing the costs associated with multi-image data annotations.\nOur observation reveals that attention values of LVLMs vary considerably across different images. We use attention values to identify and filter out rejected responses the model may have mistakenly focused on.\nOur attention-aware selection for constructing the chosen/rejected pairs without relying on (i) human annotation, (ii) extra data, and (iii) external models or APIs.\nMIA-DPO is compatible with various architectures and outperforms existing methods on five multi-image benchmarks, achieving an average performance boost of 3.0% on LLaVA-v1.5 and 4.3% on the recent InternLM-XC2.5.\nMoreover, MIA-DPO has a minimal effect on the model's ability to understand single images.", "keywords": [ "Large Vision Language Models" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "f7VXdQTbyW", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Jheng-Long Wu", "gender": "Not Specified", "institution": "Soochow University", "country": "TW", "position": "Assistant Professor" }, { "name": "Yamien Cheng", "gender": "Female", "institution": "Soochow University", "country": "TW", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 21, "n_ref": 21, "n_ref_all": 27, "n_fig": 4, "n_tab": 2, "L_tab": 309, "n_element_tab": 50, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 309, "n_element_tab_1": 50, "formula_len_all": 606, "formula_len_all_1": 606, "len_all": 83699, "len_all_1": 70894, "len_abs": 736, "len_title": 126, "len_sents": 29174, "len_sents_1": 29172, "n_sents": 196, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 737, "title": "ThreadsGAN: Enhancing Coherence and Diversity in Discussion Thread Generation", "abs": "Current research on generating discussion threads faces challenges in coherence, interactivity, and multi-topic handling, which are crucial for meaningful responses. This paper introduces threadsGAN, a model that enhances thread generation by incorporating multi-topic and social response intention tags. By leveraging BERT and Transformer, threadsGAN ensures contextual coherence and manages topic consistency. Additionally, it employs conditional generation to align responses with specific discussion contexts, and its CNN-based discriminator assesses response quality by evaluating similarity between generated and real responses, improving overall performance in generating realistic and contextually appropriate discussion threads.", "keywords": [ "discussion threads", "generative adversarial network", "natural language generating" ], "rating_list": [ 3, 3, 1, 1 ], "soundness_list": [ 1, 1, 1, 2 ], "presentation_list": [ 1, 2, 1, 1 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "f7O3hITh5s", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Adriana Meza Soria", "gender": "Female", "institution": "MIT-IBM Watson AI Lab", "country": "US", "position": "Researcher" }, { "name": "Rameswar Panda", "gender": "Male", "institution": "MIT-IBM Watson AI Lab", "country": "", "position": "Research Scientist" }, { "name": "Wei Sun", "gender": "Female", "institution": "IBM Research", "country": "", "position": "Researcher" }, { "name": "Yikang Shen", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "PhD student" }, { "name": "Zhen Guo", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 37, "n_ref": 71, "n_ref_all": 94, "n_fig": 7, "n_tab": 9, "L_tab": 7908, "n_element_tab": 408, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2570, "n_element_tab_1": 296, "formula_len_all": 10, "formula_len_all_1": 10, "len_all": 196193, "len_all_1": 60710, "len_abs": 1253, "len_title": 127, "len_sents": 46270, "len_sents_1": 25618, "n_sents": 321, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1026, "title": "API Pack: A Massive Multi-Programming Language Dataset for API Call Generation", "abs": "We introduce API Pack, a massive multi-programming language dataset containing more than one million instruction-API calls that seeks to improve the API call generation capabilities of large language models. Our evaluation highlights three key findings: First, fine-tuning on API Pack enables open-source models to outperform both GPT-3.5 and GPT-4 in generating code for completely new API calls. We demonstrate this by fine-tuning CodeLlama-13B on 20,000 Python instances from API Pack. Second, we demonstrate that fine-tuning on a large dataset in one programming language, combined with smaller datasets from others, improves API generation accuracy across multiple languages. Third, we demonstrate the benefits and the need of larger datasets for API generalization, as our experiments show that increasing fine-tuning data to 1 million instances improves generalization to new APIs. To support further research, we open-source the API Pack dataset, trained model, and code at \\url{https://github.com/anonymous/API-Pack}.", "keywords": [ "Instruction Datasets", "Synthetic Data Generation", "Code Generation" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "f7KxfUrRSb", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Pengfei Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Associate Professor" }, { "name": "Wenhong Zhu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xiaofeng Wang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Zhiwei He", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Rui Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 26, "n_formula_1": 14, "n_ref_uni": 29, "n_ref": 75, "n_ref_all": 116, "n_fig": 9, "n_tab": 18, "L_tab": 2847, "n_element_tab": 316, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 781, "n_element_tab_1": 120, "formula_len_all": 1725, "formula_len_all_1": 955, "len_all": 185318, "len_all_1": 62511, "len_abs": 1094, "len_title": 127, "len_sents": 54775, "len_sents_1": 26211, "n_sents": 500, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 79, "L_abs": 1096, "title": "Weak-to-Strong Preference Optimization: Stealing Reward from Weak Aligned Model", "abs": "Aligning language models (LMs) with human preferences has become a key area of research, enabling these models to meet diverse user needs better. Inspired by weak-to-strong generalization, where a strong LM fine-tuned on labels generated by a weaker model can consistently outperform its weak supervisor, we extend this idea to model alignment. In this work, we observe that the alignment behavior in weaker models can be effectively transferred to stronger models and even exhibit an amplification effect. Based on this insight, we propose a method called Weak-to-Strong Preference Optimization (WSPO), which achieves strong model alignment by learning the distribution differences before and after the alignment of the weak model. Experiments demonstrate that WSPO delivers outstanding performance, improving the win rate of Qwen2-7B-Instruct on Arena-Hard from 39.70 to 49.60, achieving a remarkable 47.04 length-controlled win rate on AlpacaEval 2, and scoring 7.33 on MT-bench. Our results suggest that using the weak model to elicit a strong model with a high alignment ability is feasible.", "keywords": [ "weak-to-strong", "model alignment" ], "rating_list": [ 8, 8, 8, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "f6r1mYwM1g", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Jiaxiang Liu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Jun Zhao", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "", "position": "Full Professor" }, { "name": "Kang Liu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Professor" }, { "name": "Xiusheng Huang", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Yequan Wang", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 7, "n_ref_uni": 21, "n_ref": 41, "n_ref_all": 49, "n_fig": 2, "n_tab": 7, "L_tab": 2278, "n_element_tab": 374, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 878, "n_element_tab_1": 81, "formula_len_all": 1284, "formula_len_all_1": 403, "len_all": 105994, "len_all_1": 55312, "len_abs": 1038, "len_title": 135, "len_sents": 30474, "len_sents_1": 23221, "n_sents": 244, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1041, "title": "Capability Localization: Capabilities Can be Localized rather than Individual Knowledge", "abs": "Large scale language models have achieved superior performance in tasks related to natural language processing, however, it is still unclear how model parameters affect performance improvement. Previous studies assumed that individual knowledge is stored in local parameters, and the storage form of individual knowledge is dispersed parameters, parameter layers, or parameter chains, which are not unified. We found through fidelity and reliability evaluation experiments that individual knowledge cannot be localized. Afterwards, we constructed a dataset for decoupling experiments and discovered the potential for localizing data commonalities. To further reveal this phenomenon, this paper proposes a Commonality Neuron Localization (CNL) method, which successfully locates commonality neurons and achieves a neuron overlap rate of 96.42% on the GSM8K dataset. Finally, we have demonstrated through cross data experiments that commonality neurons are a collection of capability neurons that possess the capability to enhance performance.", "keywords": [ "Capability Localization; Knowledge Localization" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 1, 2, 1 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 2, 2 ] }, { "paper_id": "f6GMwpxXHG", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Anuradha Kumari", "gender": "Female", "institution": "Indian Institute of Technology, Indore", "country": "IN", "position": "PhD student" }, { "name": "M. Tanveer", "gender": "Male", "institution": "Indian Institute of Technology, Indore", "country": "IN", "position": "Full Professor" }, { "name": "Ritik Mishra", "gender": "Male", "institution": "Indian Institute of Technology, Indore", "country": "IN", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 40, "n_ref": 50, "n_ref_all": 63, "n_fig": 4, "n_tab": 5, "L_tab": 551, "n_element_tab": 90, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 491, "n_element_tab_1": 80, "formula_len_all": 1394, "formula_len_all_1": 1295, "len_all": 122227, "len_all_1": 62178, "len_abs": 1173, "len_title": 105, "len_sents": 35869, "len_sents_1": 28778, "n_sents": 258, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1182, "title": "ZEPHYR GAN: REDEFINING GAN WITH FLEXIBLE GRADIENT CONTROL", "abs": "Generative adversarial networks (GANs) are renowned for their ability to generate highly realistic and diverse data samples. However, the performance of GANs is heavily dependent on the choice of loss functions, and commonly used losses such as cross-entropy and least squares are often susceptible to outliers, vanishing gradients, and training instability. To overcome these limitations, we introduce zephyr loss\u2014a novel, convex, smooth, and Lipschitz continuous loss function designed to enhance robustness and provide flexible gradient control. Leveraging this new loss function, we propose ZGAN, a refined GAN model that guarantees a unique optimal discriminator and stabilizes the overall training dynamics. Furthermore, we demonstrate that optimizing ZGAN's generator objective minimizes a weighted total variation between the real and generated data distributions. Through rigorous theoretical analysis, including convergence proofs, we substantiate the robustness and effectiveness of ZGAN, positioning it as a compelling and reliable alternative for stable GAN training. Extensive experiments further demonstrate that ZGAN surpasses leading methods in generative modeling.", "keywords": [ "Generative Adversarial Network", "Zephyr loss", "Adversarial Training", "Flexible Gradient Control." ], "rating_list": [ 1, 3, 3, 3, 1 ], "soundness_list": [ 1, 2, 3, 2, 1 ], "presentation_list": [ 2, 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 1, 1 ], "confidence_list": [ 5, 5, 4, 3, 4 ] }, { "paper_id": "f65RuQgVlp", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Florian Buettner", "gender": "unknown", "institution": "Deutsches Krebsforschungszentrum", "country": "", "position": "Associate Professor" }, { "name": "Giuseppe Serra", "gender": "Male", "institution": "Goethe-Universit\u00e4t Frankfurt", "country": "", "position": "Postdoc" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 47, "n_ref": 120, "n_ref_all": 143, "n_fig": 4, "n_tab": 14, "L_tab": 5102, "n_element_tab": 410, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1025, "n_element_tab_1": 77, "formula_len_all": 383, "formula_len_all_1": 271, "len_all": 165654, "len_all_1": 70059, "len_abs": 1475, "len_title": 153, "len_sents": 51634, "len_sents_1": 33786, "n_sents": 356, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 1484, "title": "Federated Continual Learning Goes Online: Uncertainty-Aware Memory Management for Vision Tasks and Beyond", "abs": "Given the ability to model more realistic and dynamic problems, Federated Continual Learning (FCL) has been increasingly investigated recently. A well-known problem encountered in this setting is the so-called catastrophic forgetting, for which the learning model is inclined to focus on more recent tasks while forgetting the previously learned knowledge. The majority of the current approaches in FCL propose generative-based solutions to solve said problem. However, this setting requires multiple training epochs over the data, implying an offline setting where datasets are stored locally and remain unchanged over time. Furthermore, the proposed solutions are tailored for vision tasks solely. To overcome these limitations, we propose a new approach to deal with different modalities in the online scenario where new data arrive in streams of mini-batches that can only be processed once. To solve catastrophic forgetting, we propose an uncertainty-aware memory-based approach. Specifically, we suggest using an estimator based on the Bregman Information (BI) to compute the model's variance at the sample level. Through measures of predictive uncertainty, we retrieve samples with specific characteristics, and \u2013 by retraining the model on such samples \u2013 we demonstrate the potential of this approach to reduce the forgetting effect in realistic settings while maintaining data confidentiality and competitive communication efficiency compared to state-of-the-art approaches.", "keywords": [ "Federated Continual Learning", "Catastrophic Forgetting", "Uncertainty Estimation" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "f5o6kWRC0A", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Feng Ji", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "JiaHao Wu", "gender": "unknown", "institution": "Jiangnan University", "country": "CN", "position": "MS student" }, { "name": "Jielong Yang", "gender": "Male", "institution": "Jiangnan University", "country": "CN", "position": "Associate Professor" }, { "name": "L Peng", "gender": "Male", "institution": "Jiangnan University", "country": "", "position": "Full Professor" } ], "n_formula": 33, "n_formula_1": 21, "n_ref_uni": 40, "n_ref": 55, "n_ref_all": 69, "n_fig": 3, "n_tab": 7, "L_tab": 2285, "n_element_tab": 269, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1601, "n_element_tab_1": 178, "formula_len_all": 2236, "formula_len_all_1": 1376, "len_all": 130328, "len_all_1": 66063, "len_abs": 1581, "len_title": 160, "len_sents": 33268, "len_sents_1": 27571, "n_sents": 293, "n_sents_1": 238, "venue": "ICLR2025", "decision": "withdraw", "L_title": 110, "L_abs": 1589, "title": "Machine Unlearning For Alleviating Negative Transfer In Partial-Set Source-Free Unsupervised Domain Adaptation", "abs": "Source-free Unsupervised Domain Adaptation (SFUDA) aims to adjust a source model trained on a labeled source domain to a related but unlabeled target domain without accessing the source data. Many SFUDA methods are studied in closed-set scenarios where the target domain and source domain categories are perfectly aligned. However, a more practical scenario is a partial-set scenario where the source label space subsumes the target one. In this paper, we prove that reducing the differences between the source and target domains in the partial-set scenario helps to achieve domain adaptation. And we propose a simple yet effective SFUDA framework called the Machine Unlearning Framework to alleviate the negative transfer problem in the partial-set scenario, thereby allowing the model to focus on the target domain category. Specifically, we first generate noise samples for each category that only exists in the source domain and generate pseudo-labeled samples from the target domain. Then, in the forgetting stage, we use these samples to train the model, making it behave like the model has never seen the class that only exists in the source domain before. Finally, in the adaptation stage, we use only the pseudo-labeled samples to conduct self-supervised training on the model, making it more adaptable to the target domain. Our method is easy to implement and pluggable, suitable for various pre-trained models. Experimental results show that our method can well alleviate the negative transfer problem and improve model performance under various target domain category settings.", "keywords": [ "Source-Free Domain Adaptation", "Unsupervised domain adaptation", "Machine unlearning" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "f5kwV2rdce", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jiacheng Qiu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Shuyang Cui", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Tauhidur Rahman", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Yijun Luo", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Undergrad student" }, { "name": "Zhongyang Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 29, "n_ref": 39, "n_ref_all": 59, "n_fig": 4, "n_tab": 9, "L_tab": 3170, "n_element_tab": 303, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2498, "n_element_tab_1": 131, "formula_len_all": 1435, "formula_len_all_1": 960, "len_all": 147782, "len_all_1": 65611, "len_abs": 1195, "len_title": 169, "len_sents": 47885, "len_sents_1": 30931, "n_sents": 355, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 120, "L_abs": 1150, "title": "Labits: Layered Bidirectional Time Surfaces Representation for Event Camera-based Continuous Dense Trajectory Estimation", "abs": "Event cameras provide a compelling alternative to traditional frame-based sensors, capturing dynamic scenes with high temporal resolution and low latency. Moving objects trigger events with precise timestamps along their trajectory, enabling smooth continuous-time estimation. However, few works have attempted to optimize the information loss during event representation construction, imposing a ceiling on this task. Fully exploiting event cameras requires representations that simultaneously preserve fine-grained temporal information, stable and characteristic 2D visual features, and temporally consistent information density\u2014an unmet challenge in existing representations. We introduce Labits: Layered Bidirectional Time Surfaces, a simple yet elegant representation designed to retain all these features. Additionally, we propose a dedicated module for extracting active pixel local optical flow (APLOF), significantly boosting the performance. Our approach achieves an impressive 49\\% reduction in trajectory end-point error (TEPE) compared to the previous state-of-the-art on the MultiFlow dataset. The code will be released upon acceptance.", "keywords": [ "event camera", "trajectory estimation", "computer vision", "event-based vision", "dynamic vision sensor", "feature tracking" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 2, 3, 5 ] }, { "paper_id": "f4mQ2SU5tp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Hang Guo", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Luca Benini", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Full Professor" }, { "name": "Shu-Tao Xia", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Tao Dai", "gender": "Male", "institution": "Department of Software Engineering, Shenzhen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yawei Li", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Lecturer" } ], "n_formula": 11, "n_formula_1": 13, "n_ref_uni": 43, "n_ref": 87, "n_ref_all": 118, "n_fig": 18, "n_tab": 6, "L_tab": 1927, "n_element_tab": 195, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 544, "n_element_tab_1": 30, "formula_len_all": 770, "formula_len_all_1": 940, "len_all": 177713, "len_all_1": 66354, "len_abs": 1322, "len_title": 115, "len_sents": 54519, "len_sents_1": 31420, "n_sents": 420, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1307, "title": "IntLoRA: Integral Low-rank Adaptation of Quantized Diffusion Models", "abs": "Fine-tuning large-scale text-to-image diffusion models for various downstream tasks has yielded impressive results. However, the heavy computational burdens of tuning large models prevent personal customization. Recent advances have attempted to employ parameter-efficient fine-tuning (PEFT) techniques to adapt the floating-point (FP) or quantized pre-trained weights. Nonetheless, the adaptation parameters in existing works are still restricted to FP arithmetic, hindering hardware-friendly acceleration. In this work, we propose IntLoRA, to further push the efficiency limits by using integer type (INT) low-rank parameters to adapt the quantized diffusion models. By working in the integer arithmetic, our IntLoRA offers three key advantages: (i) for fine-tuning, the pre-trained weights are quantized, reducing memory usage; (ii) for storage, both pre-trained and low-rank weights are in INT which consumes less disk space; (iii) for inference, IntLoRA weights can be naturally merged into quantized pre-trained weights through efficient integer multiplication or bit-shifting, eliminating additional post-training quantization. Extensive experiments demonstrate that IntLoRA can achieve performance on par with or even superior to the vanilla LoRA, accompanied by significant efficiency improvements.", "keywords": [ "Parameter efficient fine tuning", "Network quantization", "Diffusion model" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "f4gF6AIHRy", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Pingjie Wang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Shengchao Hu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Siyuan Du", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Ya Zhang", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Professor" }, { "name": "Yanfeng Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Ziqing Fan", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 34, "n_formula_1": 12, "n_ref_uni": 59, "n_ref": 127, "n_ref_all": 158, "n_fig": 8, "n_tab": 11, "L_tab": 5070, "n_element_tab": 643, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3649, "n_element_tab_1": 405, "formula_len_all": 2251, "formula_len_all_1": 561, "len_all": 187454, "len_all_1": 69831, "len_abs": 1586, "len_title": 135, "len_sents": 49201, "len_sents_1": 29766, "n_sents": 326, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 86, "L_abs": 1596, "title": "Combatting Dimensional Collapse in LLM Pre-Training Data via Submodular File Selection", "abs": "Selecting high-quality pre-training data for large language models (LLMs) is crucial for enhancing their overall performance under limited computation budget, improving both training and sample efficiency. Recent advancements in file selection primarily rely on using an existing or trained proxy model to assess the similarity of samples to a target domain, such as high quality sources BookCorpus and Wikipedia. However, upon revisiting these methods, the domain-similarity selection criteria demonstrates a diversity dilemma, i.e. dimensional collapse in the feature space, improving performance on the domain-related tasks but causing severe degradation on generic performance.To prevent collapse and enhance diversity, we propose a DiverSified File selection algorithm (DiSF), which selects the most decorrelated text files in the feature space. We approach this with a classical greedy algorithm to achieve more uniform eigenvalues in the feature covariance matrix of the selected texts, analyzing its approximation to the optimal solution under a formulation of $\\gamma$-weakly submodular optimization problem. Empirically, we establish a benchmark and conduct extensive experiments on the TinyLlama architecture with models from 120M to 1.1B parameters. Evaluating across nine tasks from the Harness framework, DiSF demonstrates a significant improvement on overall performance. Specifically, DiSF saves 98.5\\% of 590M training files in SlimPajama, outperforming the full-data pre-training within a 50B training budget, and achieving about 1.5x training efficiency and 5x data efficiency.", "keywords": [ "file selection", "large language model", "pre-training", "submodular optimization" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "f4b0YVwKUO", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Hanyu Hu", "gender": "Not Specified", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Pengxiang Zhao", "gender": "unknown", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Ping Li", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Xiaoming Yuan", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Yi ZHENG", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Zhefeng Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 26, "n_ref": 55, "n_ref_all": 67, "n_fig": 4, "n_tab": 6, "L_tab": 550, "n_element_tab": 91, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 550, "n_element_tab_1": 91, "formula_len_all": 637, "formula_len_all_1": 637, "len_all": 85886, "len_all_1": 55629, "len_abs": 1390, "len_title": 115, "len_sents": 25032, "len_sents_1": 25025, "n_sents": 178, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1395, "title": "FASP: Fast and Accurate Structured Pruning of Large Language Models", "abs": "The rapid increase in the size of large language models (LLMs) has significantly escalated their computational and memory demands, posing challenges for efficient deployment, especially on resource-constrained devices. Structured pruning has emerged as an effective model compression method that can reduce these demands while preserving performance. In this paper, we introduce FASP (Fast and Accurate Structured Pruning), a novel structured pruning framework for LLMs that emphasizes both speed and accuracy. FASP employs a distinctive pruning structure that interlinks sequential layers, allowing for the removal of columns in one layer while simultaneously eliminating corresponding rows in the preceding layer without incurring additional performance loss. The pruning metric, inspired by Wanda, is computationally efficient and effectively selects components to prune. Additionally, we propose a restoration mechanism that enhances model fidelity by adjusting the remaining weights post-pruning. We evaluate FASP on the OPT and LLaMA model families, demonstrating superior performance in terms of perplexity and accuracy on downstream tasks compared to state-of-the-art methods. Our approach achieves significant speed-ups, pruning models such as OPT-125M in 17 seconds and LLaMA-30B in 20 minutes on a single NVIDIA RTX 4090 GPU, making it a highly practical solution for optimizing LLMs.", "keywords": [ "large language models", "post training pruning", "structured pruning" ], "rating_list": [ 3, 5, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 3, 2 ], "presentation_list": [ 1, 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 3, 1 ], "confidence_list": [ 5, 4, 4, 4, 5 ] }, { "paper_id": "f4aMqhYG7z", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "JinLei Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Xingyu Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zewei He", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhe-Ming Lu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zixuan Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 58, "n_ref": 138, "n_ref_all": 171, "n_fig": 23, "n_tab": 9, "L_tab": 1792, "n_element_tab": 188, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 1592, "n_element_tab_1": 151, "formula_len_all": 1003, "formula_len_all_1": 1003, "len_all": 141554, "len_all_1": 62937, "len_abs": 1542, "len_title": 126, "len_sents": 33265, "len_sents_1": 26362, "n_sents": 275, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1567, "title": "Addressing domain shift with diffusion-based adaptation for real image dehazing", "abs": "Conventional supervised single-image dehazing methods, which are trained with substantial synthetic hazy-clean image pairs, have achieved promising performance. However, they often fail to tackle out-of-distribution hazy images, due to the domain shift between source and target scenarios (e.g., between indoor and outdoor, between synthetic and real). In this work, we observe the opportunity for improving such dehazing models' generalization ability without modifying the architectures or weights of conventional models by adopting the diffusion model to transfer the distribution of input images from target domain to source domain. Specifically, we train a denoising diffusion probabilistic model (DDPM) with source hazy images to capture prior probability distribution of the source domain. Then, during the test-time the obtained DDPM can adapt target hazy inputs to source domain in the reverse process from the perspective of conditional generation. The adapted inputs are fed into a certain state-of-the-art (SOTA) dehazing model pre-trained on source domain to predict the haze-free outputs. Note that, the whole proposed pipeline, termed \\textbf{Diff}usion-based \\textbf{AD}aptation (DiffAD), is model-agnostic and plug-and-play. Besides, to enhance the efficiency in real image dehazing, we further employ the predicted haze-free outputs as the pseudo labels to fine-tune the underlying model. Extensive experimental results demonstrate that our DiffAD is effective, achieving superior performance against SOTA dehazing methods in domain-shift scenarios.", "keywords": [ "diffusion-based adaptation", "domain gap", "real image dehazing" ], "rating_list": [ 5, 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 5, 5, 4 ] }, { "paper_id": "f47c05mcOj", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "ANH-HUY PHAN", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Full Professor" }, { "name": "Anton Bibin", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Researcher" }, { "name": "Ashish Jha", "gender": "Male", "institution": "Skolkovo Institute of Science And Technology", "country": "RU", "position": "PhD student" }, { "name": "Nikolay Kalmykov", "gender": "unknown", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Nikolay Kozyrskiy", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Razan Dibo", "gender": "Female", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 9, "n_ref": 20, "n_ref_all": 47, "n_fig": 9, "n_tab": 11, "L_tab": 7833, "n_element_tab": 1015, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 299, "n_element_tab_1": 49, "formula_len_all": 345, "formula_len_all_1": 345, "len_all": 121813, "len_all_1": 57362, "len_abs": 4205, "len_title": 62, "len_sents": 40278, "len_sents_1": 27788, "n_sents": 303, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1179, "title": "Adaptive Log-Exp Perturbations for Secure AI Image Compression", "abs": "AI image compression has outperformed traditional methods in both efficiency and quality but remains vulnerable to adversarial attacks. Most attacks on deep neural networks (DNNs) involve adding small perturbations to the input image to deceive the system and produce incorrect results. While simple, these additive perturbations affect pixels uniformly across different intensity levels, from dark to bright regions. However the human eye is less sensitive to variations in dark areas than in bright ones, making noise in brighter areas more visible. This observation suggests a novel attack strategy that minimizes the visibility of adversarial noise through adaptive perturbations. To achieve this, we propose a nonlinear log-exp perturbation, which applies more noise to dark pixels while minimizing its impact on bright areas.\n\nWe evaluated this perturbation model in two scenarios: one distorts the output of decompression models and another one increases the bit rate of compressed images without visibly affecting quality. Our findings offer new strategies to protect AI-driven image compression systems, ensuring both security and performance in practical applications.", "keywords": [ "neural image compression", "adversarial attack", "adaptive perturbation" ], "rating_list": [ 1, 3, 3, 5 ], "soundness_list": [ 1, 1, 2, 3 ], "presentation_list": [ 2, 1, 1, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "f3xXPDCh8Q", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Haixu Wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hang Zhou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Haowen Wang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yuezhou Ma", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Mingsheng Long", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 53, "n_formula_1": 22, "n_ref_uni": 38, "n_ref": 98, "n_ref_all": 174, "n_fig": 12, "n_tab": 24, "L_tab": 8078, "n_element_tab": 285, "n_fig_1": 3, "n_tab_1": 14, "L_tab_1": 5271, "n_element_tab_1": 786, "formula_len_all": 3974, "formula_len_all_1": 2184, "len_all": 258956, "len_all_1": 124411, "len_abs": 1426, "len_title": 120, "len_sents": 93798, "len_sents_1": 52243, "n_sents": 703, "n_sents_1": 380, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1445, "title": "Unisolver: PDE-Conditional Transformers Are Universal PDE Solvers", "abs": "Deep models have recently emerged as a promising tool to solve partial differential equations (PDEs), known as neural PDE solvers. While neural solvers trained from either simulation data or physics-informed loss can solve PDEs reasonably well, they are mainly restricted to a few instances of PDEs, e.g. a certain equation with a limited set of coefficients. This limits the generalization of neural solvers to diverse PDEs, impeding them from being practical surrogate models for numerical solvers. In this paper, we present the Universal PDE Solver (Unisolver) capable of solving a wide scope of PDEs by training a novel Transformer model on diverse data and conditioned on diverse PDEs. Instead of purely scaling up data and parameters, Unisolver stems from the theoretical analysis of the PDE-solving process. Our key finding is that a PDE solution is fundamentally under the control of a series of PDE components, e.g. equation symbols, coefficients, and boundary conditions. Inspired by the mathematical structure of PDEs, we define a complete set of PDE components and flexibly embed them as domain-wise (e.g. equation symbols) and point-wise (e.g. boundaries) conditions for Transformer PDE solvers. Integrating physical insights with recent Transformer advances, Unisolver achieves consistent state-of-the-art results on three challenging large-scale benchmarks, showing impressive performance gains and favorable PDE generalizability.", "keywords": [ "Neural PDE Solver", "Deep Learning" ], "rating_list": [ 8, 3, 8, 3 ], "soundness_list": [ 4, 2, 3, 1 ], "presentation_list": [ 4, 3, 4, 1 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 5, 3, 3 ] }, { "paper_id": "f3jySJpEFT", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Harin Lee", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Undergrad student" }, { "name": "Min-hwan Oh", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" }, { "name": "Taehyun Hwang", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" } ], "n_formula": 281, "n_formula_1": 25, "n_ref_uni": 30, "n_ref": 222, "n_ref_all": 279, "n_fig": 5, "n_tab": 10, "L_tab": 3009, "n_element_tab": 256, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 144, "n_element_tab_1": 6, "formula_len_all": 50637, "formula_len_all_1": 13770, "len_all": 359474, "len_all_1": 89051, "len_abs": 1987, "len_title": 104, "len_sents": 93376, "len_sents_1": 33504, "n_sents": 926, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1237, "title": "Lasso Bandit with Compatibility Condition on Optimal Arm", "abs": "We consider a stochastic sparse linear bandit problem where only a sparse subset of context features affects the expected reward function, i.e., the unknown reward parameter has sparse structure.\nIn the existing Lasso bandit literature, the compatibility conditions together with additional diversity conditions on the context features are imposed to achieve regret bounds that only depend logarithmically on the ambient dimension $d$.\nIn this paper, we demonstrate that even without the additional diversity assumptions, the _compatibility condition on the optimal arm_ is sufficient to derive a regret bound that depends logarithmically on $d$, and our assumption is strictly weaker than those used in the lasso bandit literature under the single-parameter setting.\nWe propose an algorithm that adapts the forced-sampling technique and prove that the proposed algorithm achieves $\\mathcal{O}(\\text{poly}\\log dT)$ regret under the margin condition.\nTo our knowledge, the proposed algorithm requires the weakest assumptions among Lasso bandit algorithms under the single-parameter setting that achieve $\\mathcal{O}(\\text{poly}\\log dT)$ regret.\nThrough numerical experiments, we confirm the superior performance of our proposed algorithm.", "keywords": [ "Sparse linear bandits", "Lasso", "Regret analysis" ], "rating_list": [ 5, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "f3hIphjjY8", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Dongxiao Yu", "gender": "unknown", "institution": "Shandong University", "country": "CN", "position": "Full Professor" }, { "name": "Jing Qiao", "gender": "Not Specified", "institution": "Shandong University", "country": "CN", "position": "PhD student" }, { "name": "Xiao Zhang", "gender": "unknown", "institution": "Shandong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiuzhen Cheng", "gender": "Female", "institution": "Shandong University", "country": "CN", "position": "Full Professor" }, { "name": "YUAN YUAN", "gender": "Female", "institution": "Shandong University", "country": "", "position": "Postdoc" }, { "name": "Yu Liu", "gender": "unknown", "institution": "Shandong University", "country": "CN", "position": "PhD student" } ], "n_formula": 94, "n_formula_1": 42, "n_ref_uni": 31, "n_ref": 66, "n_ref_all": 99, "n_fig": 0, "n_tab": 2, "L_tab": 640, "n_element_tab": 71, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 14697, "formula_len_all_1": 4417, "len_all": 136387, "len_all_1": 61765, "len_abs": 905, "len_title": 113, "len_sents": 38711, "len_sents_1": 26486, "n_sents": 318, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 910, "title": "$d$-Linear Generation Error Bound for Distributed Diffusion Models", "abs": "The recent rise of distributed diffusion models has been driven by the explosive growth of data and the increasing demand for data generation. However, distributed diffusion models face unique challenges in resource-constrained environments. Existing approaches lack theoretical support, particularly with respect to generation error in such settings. In this paper, we are the first to derive the generation error bound for distributed diffusion models with arbitrary pruning, not assuming perfect score approximation. By analyzing the convergence of the score estimation model trained with arbitrary pruning in a distributed manner, we highlight the impact of complex factors such as model evolution dynamics and arbitrary pruning on the generation performance. This theoretical generation error bound is linear in the data dimension $d$, aligning with state-of-the-art results in the single-worker paradigm.", "keywords": [ "Distributed diffusion models", "Generation error bound" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 1 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "f3gCs2a4ZD", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 8, "author_info_list": [ { "name": "Antonio Norelli", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Jingcheng Wu", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "DE", "position": "PhD student" }, { "name": "Michael M. Bronstein", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Volker Tresp", "gender": "Male", "institution": "Ludwig Maximilian University of Munich", "country": "DE", "position": "Associate Professor" }, { "name": "Yifeng Li", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "MS student" }, { "name": "Yuan He", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Yunpu Ma", "gender": "Male", "institution": "Siemens Corporate Research", "country": "DE", "position": "Researcher" }, { "name": "Zifeng Ding", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" } ], "n_formula": 22, "n_formula_1": 23, "n_ref_uni": 26, "n_ref": 97, "n_ref_all": 143, "n_fig": 9, "n_tab": 20, "L_tab": 8007, "n_element_tab": 679, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 5519, "n_element_tab_1": 238, "formula_len_all": 3083, "formula_len_all_1": 3250, "len_all": 221506, "len_all_1": 78669, "len_abs": 1327, "len_title": 167, "len_sents": 66707, "len_sents_1": 30245, "n_sents": 534, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 118, "L_abs": 1333, "title": "DyGMamba: Efficiently Modeling Long-Term Temporal Dependency on Continuous-Time Dynamic Graphs with State Space Models", "abs": "Learning useful representations for continuous-time dynamic graphs (CTDGs) is challenging, due to the concurrent need to span long node interaction histories and grasp nuanced temporal details. In particular, two problems emerge: (1) Encoding longer histories requires more computational resources, making it crucial for CTDG models to maintain low computational complexity to ensure efficiency; (2) Meanwhile, more powerful models are needed to identify and select the most critical temporal information within the extended context provided by longer histories. To address these problems, we propose a CTDG representation learning model named DyGMamba, originating from the popular Mamba state space model (SSM). DyGMamba first leverages a node-level SSM to encode the sequence of historical node interactions. Another time-level SSM is then employed to exploit the temporal patterns hidden in the historical graph, where its output is used to dynamically select the critical information from the interaction history. We validate DyGMamba experimentally on the dynamic link prediction task. The results show that our model achieves state-of-the-art in most cases. DyGMamba also maintains high efficiency in terms of computational resources, making it possible to capture long temporal dependencies with a limited computation budget.", "keywords": [ "dynamic graph", "state space model" ], "rating_list": [ 3, 3, 8, 3, 3 ], "soundness_list": [ 2, 2, 4, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 2 ], "contribution_list": [ 1, 2, 3, 1, 2 ], "confidence_list": [ 5, 3, 5, 5, 4 ] }, { "paper_id": "f3TSOXnkXZ", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Do\u011fanay \u00d6zese", "gender": "Male", "institution": "Bogazici University", "country": "TR", "position": "MS student" }, { "name": "Ilker Birbil", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "Full Professor" }, { "name": "Mustafa Gokce Baydogan", "gender": "Male", "institution": "Bogazici University", "country": "", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 23, "n_ref": 28, "n_ref_all": 42, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 677, "formula_len_all_1": 677, "len_all": 88109, "len_all_1": 66996, "len_abs": 809, "len_title": 81, "len_sents": 33317, "len_sents_1": 33317, "n_sents": 256, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 33, "L_abs": 812, "title": "Output-Constrained Decision Trees", "abs": "When there is a correlation between any pair of targets, one needs a prediction method that can handle vector-valued output. In this setting, multi-target learning is particularly important as it is widely used in various applications. This paper introduces new variants of decision trees that can handle not only multi-target output but also the constraints among the targets. We focus on the customization of conventional decision trees by adjusting the splitting criteria to handle the constraints and obtain feasible predictions. We present both an optimization-based exact approach and several heuristics, complete with a discussion on their respective advantages and disadvantages. To support our findings, we conduct a computational study to demonstrate and compare the results of the proposed approaches.", "keywords": [ "output constraints", "decision trees", "optimization", "heuristics" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "f3QR9TEERH", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Cevahir Koprulu", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Nils Jansen", "gender": "Male", "institution": "Ruhr-Universit\u00e4t Bochum", "country": "DE", "position": "Full Professor" }, { "name": "Thiago D. Sim\u00e3o", "gender": "unknown", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "ufuk topcu", "gender": "Unspecified", "institution": "University of Texas, Austin", "country": "US", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 42, "n_ref": 90, "n_ref_all": 122, "n_fig": 19, "n_tab": 4, "L_tab": 2505, "n_element_tab": 209, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 518, "n_element_tab_1": 6, "formula_len_all": 1938, "formula_len_all_1": 1361, "len_all": 176679, "len_all_1": 69665, "len_abs": 1041, "len_title": 116, "len_sents": 59269, "len_sents_1": 32855, "n_sents": 464, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1047, "title": "Safety-Prioritizing Curricula for Constrained Reinforcement Learning", "abs": "Curriculum learning aims to accelerate reinforcement learning (RL) by generating curricula, i.e., sequences of tasks of increasing difficulty. \nAlthough existing curriculum generation approaches provide benefits in sample efficiency, they overlook safety-critical settings where an RL agent must adhere to safety constraints.\nThus, these approaches may generate tasks that cause RL agents to violate safety constraints during training and behave suboptimally after. \nWe develop a safe curriculum generation approach (SCG) that aligns the objectives of constrained RL and curriculum learning: improving safety during training and boosting sample efficiency.\nSCG generates sequences of tasks where the RL agent can be safe and performant by initially generating tasks with minimum safety violations over high-reward ones.\nWe empirically show that compared to the state-of-the-art curriculum learning approaches and their naively modified safe versions, SCG achieves optimal performance and the lowest amount of constraint violations during training.", "keywords": [ "curriculum learning", "constrained reinforcement learning" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "f3CdjpPkSq", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Deepu Rajan", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Yihui Qiu", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 30, "n_ref": 42, "n_ref_all": 74, "n_fig": 5, "n_tab": 13, "L_tab": 1977, "n_element_tab": 245, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 418, "n_element_tab_1": 71, "formula_len_all": 312, "formula_len_all_1": 302, "len_all": 137089, "len_all_1": 65532, "len_abs": 1242, "len_title": 101, "len_sents": 48841, "len_sents_1": 31736, "n_sents": 377, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1248, "title": "Action Sequence Augmentation for Action Anticipation", "abs": "Action anticipation models require an understanding of temporal action patterns and dependencies to predict future actions from previous events. The key challenges arise from the vast number of possible action sequences, given the flexibility in action ordering and the interleaving of multiple goals. Since only a subset of such action sequences are present in action anticipation datasets, there is an inherent ordering bias in them. Another challenge is the presence of noisy input to the models due to erroneous action recognition or other upstream tasks. This paper addresses these challenges by introducing a novel data augmentation strategy that separately augments observed action sequences and next actions. To address biased action ordering, we introduce a grammar induction algorithm that derives a powerful context-free grammar from action sequence data. We also develop an efficient parser to generate plausible next-action candidates beyond the ground truth. For noisy input, we enhance model robustness by randomly deleting or replacing actions in observed sequences. Our experiments on the 50Salads, EGTEA Gaze+, and Epic-Kitchens-100 datasets demonstrate significant performance improvements over existing state-of-the-art methods.", "keywords": [ "Sequence Augmentation", "Action Anticipation", "Context-free Grammar" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "f1uXrAjpOH", "primary_area": "datasets and benchmarks", "n_author": 18, "author_info_list": [ { "name": "Bin Liu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "Associate Professor" }, { "name": "Hailiang Yao", "gender": "Male", "institution": "Tianjin Normal University", "country": "CN", "position": "MS student" }, { "name": "Haiyang Sun", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Hao Gu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Haoyu Chen", "gender": "Male", "institution": "University of Oulu", "country": "FI", "position": "Assistant Professor" }, { "name": "Jiangyan Yi", "gender": "Female", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Associate Professor" }, { "name": "Jianhua Tao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Kang Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Lan Chen", "gender": "Female", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Licai Sun", "gender": "unknown", "institution": "University of Oulu", "country": "FI", "position": "Postdoc" }, { "name": "Rui Liu", "gender": "Male", "institution": "Inner Mongolia University", "country": "CN", "position": "Full Professor" }, { "name": "Shan Liang", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "", "position": "Associate Professor" }, { "name": "Shun Chen", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Ya Li", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "Associate Professor" }, { "name": "ZHUOFAN WEN", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "Undergrad student" }, { "name": "Zhang Siyuan", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Zheng Lian", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "\u540d\u5b87 \u5f90", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 52, "n_ref": 78, "n_ref_all": 134, "n_fig": 18, "n_tab": 9, "L_tab": 2574, "n_element_tab": 224, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1829, "n_element_tab_1": 166, "formula_len_all": 753, "formula_len_all_1": 420, "len_all": 253215, "len_all_1": 67151, "len_abs": 1150, "len_title": 127, "len_sents": 77239, "len_sents_1": 27747, "n_sents": 714, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1155, "title": "Open-vocabulary Multimodal Emotion Recognition: Dataset, Metric, and Benchmark", "abs": "Multimodal Emotion Recognition (MER) is an important research topic. This paper advocates for a transformative paradigm in MER. The rationale behind our work is that current approaches often rely on a limited set of basic emotion labels, which do not adequately represent the rich spectrum of human emotions. These traditional and overly simplistic emotion categories fail to capture the inherent complexity and subtlety of human emotional experiences, leading to limited generalizability and practicality. Therefore, we propose a new MER paradigm called Open-vocabulary MER (OV-MER), which encompasses a broader range of emotion labels to reflect the richness of human emotions. This paradigm relaxes the label space, allowing for the prediction of arbitrary numbers and categories of emotions. To support this transition, we provide a comprehensive solution that includes a newly constructed database based on LLM and human collaborative annotations, along with corresponding metrics and a series of benchmarks. We hope this work advances emotion recognition from basic emotions to more nuanced emotions, contributing to the development of emotional AI.", "keywords": [ "Open-vocabulary Multimodal Emotion Recognition", "Dataset", "Metric", "Benchmark" ], "rating_list": [ 6, 5, 3, 5, 8 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 1, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 4 ], "confidence_list": [ 3, 5, 5, 4, 4 ] }, { "paper_id": "f114duq0Ov", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Chul-Ho Lee", "gender": "unknown", "institution": "Texas State University", "country": "US", "position": "Assistant Professor" }, { "name": "S.-H. Chan", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Shuhan Zhong", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Weifeng Su", "gender": "Male", "institution": "BNU-HKBU United International College", "country": "CN", "position": "Full Professor" }, { "name": "Weipeng Zhuo", "gender": "Male", "institution": "BNU-HKBU United International College", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhenghao Wu", "gender": "Male", "institution": "University College Dublin", "country": "IE", "position": "PhD student" }, { "name": "Zhongyi Yu", "gender": "unknown", "institution": "BNU-HKBU UIC", "country": "CN", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 44, "n_ref": 89, "n_ref_all": 123, "n_fig": 6, "n_tab": 20, "L_tab": 43431, "n_element_tab": 5302, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2611, "n_element_tab_1": 297, "formula_len_all": 671, "formula_len_all_1": 687, "len_all": 276148, "len_all_1": 77543, "len_abs": 1303, "len_title": 125, "len_sents": 50229, "len_sents_1": 33580, "n_sents": 347, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1311, "title": "M$^3$-Impute: Mask-guided Representation Learning for Missing Value Imputation", "abs": "Missing values are a common problem that poses significant challenges to data analysis and machine learning. This problem necessitates the development of an effective imputation method to fill in the missing values accurately, thereby enhancing the overall quality and utility of the datasets. Existing imputation methods, however, fall short of explicitly considering the `missingness' information in the data during the embedding initialization stage and modeling the entangled feature and sample correlations during the learning process, thus leading to inferior performance. We propose M$^3$-Impute, which aims to explicitly leverage the missingness information and such correlations with novel masking schemes. M$^3$-Impute first models the data as a bipartite graph and uses a graph neural network to learn node embeddings, where the refined embedding initialization process directly incorporates the missingness information. They are then optimized through M$^3$-Impute's novel feature correlation unit (FRU) and sample correlation unit (SRU) that effectively captures feature and sample correlations for imputation. Experiment results on 25 benchmark datasets under three different missingness settings show the effectiveness of M$^3$-Impute by achieving 20 best and 4 second-best MAE scores on average.", "keywords": [ "Missing Value Imputation", "Graph Representation Learning", "Data Correlations" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "f0cGihOlgH", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Diego Klabjan", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "Full Professor" }, { "name": "Mengfan Xu", "gender": "Female", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" } ], "n_formula": 185, "n_formula_1": 18, "n_ref_uni": 27, "n_ref": 49, "n_ref_all": 67, "n_fig": 9, "n_tab": 1, "L_tab": 154, "n_element_tab": 5, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 23859, "formula_len_all_1": 1816, "len_all": 238964, "len_all_1": 80573, "len_abs": 1550, "len_title": 124, "len_sents": 75432, "len_sents_1": 37927, "n_sents": 831, "n_sents_1": 316, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1557, "title": "Regret Bounds and Reinforcement Learning Exploration of EXP-based Algorithms", "abs": "We study the challenging exploration incentive problem in both bandit and reinforcement learning, where the rewards are scale-free and potentially unbounded, driven by real-world scenarios and differing from existing work. Past works in reinforcement learning either assume costly interactions with an environment or propose algorithms finding potentially low quality local maxima. Motivated by EXP-type methods that integrate multiple agents (experts) for exploration in bandits with the assumption that rewards are bounded, we propose new algorithms, namely EXP4.P and EXP4-RL for exploration in the unbounded reward case, and demonstrate their effectiveness in these new settings. Unbounded rewards introduce challenges as the regret cannot be limited by the number of trials, and selecting suboptimal arms may lead to infinite regret. Specifically, we establish EXP4.P's regret upper bounds in both bounded and unbounded linear and stochastic contextual bandits. Surprisingly, we also find that by including one sufficiently competent expert, EXP4.P can achieve global optimality in the linear case. This unbounded reward result is also applicable to a revised version of EXP3.P in the Multi-armed Bandit scenario. In EXP4-RL, we extend EXP4.P from bandit scenarios to reinforcement learning to incentivize exploration by multiple agents, including one high-performing agent, for both efficiency and excellence. This algorithm has been tested on difficult-to-explore games and shows significant improvements in exploration compared to state-of-the-art.", "keywords": [ "Bandit", "reinforcement learning", "EXP-based algorithms" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 2, 2, 4 ] }, { "paper_id": "f0GTSxGi6G", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 8, "author_info_list": [ { "name": "Emma Strubell", "gender": "Non-Binary", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Jacob Kahn", "gender": "Male", "institution": "Facebook AI Research", "country": "", "position": "Research Engineer" }, { "name": "Jared Fernandez", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Kalyan Saladi", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Leonid Shamis", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Luca Wehrstedt", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Research Engineer" }, { "name": "Mostafa Elhoushi", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Yonatan Bisk", "gender": "Male", "institution": "Meta", "country": "US", "position": "Visiting Professor" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 51, "n_ref": 77, "n_ref_all": 106, "n_fig": 12, "n_tab": 4, "L_tab": 1375, "n_element_tab": 122, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 19, "len_all": 165446, "len_all_1": 57362, "len_abs": 1385, "len_title": 131, "len_sents": 43568, "len_sents_1": 29944, "n_sents": 258, "n_sents_1": 168, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1358, "title": "Hardware Scaling Trends and Diminishing Returns in Large-Scale Distributed Training", "abs": "Dramatic increases in the capabilities of neural network models in recent years\nare driven by scaling model size, training data, and corresponding computational\nresources. To develop the exceedingly large networks required in modern applications, such as large language models (LLMs), model training is distributed\nacross tens of thousands of hardware accelerators (e.g. GPUs), requiring orchestration of computation and communication across large computing clusters. In this\nwork, we demonstrate that careful consideration of hardware configuration and\nparallelization strategy is critical for effective (i.e. compute- and cost-efficient)\nscaling of model size, training data, and total computation. We conduct an extensive empirical study of the performance of large-scale LLM training workloads\nacross model size, hardware configurations, and distributed parallelization strategies. We demonstrate that: (1) beyond certain scales, overhead incurred from\ncertain distributed communication strategies leads parallelization strategies previously thought to be sub-optimal in fact become preferable; and (2) scaling the\ntotal number of accelerators for large model training quickly yields diminishing\nreturns even when hardware and parallelization strategies are properly optimized,\nimplying poor marginal performance per additional unit of power or GPU-hour.", "keywords": [ "distributed training", "scaling", "pretraining", "hardware-software", "parallelization", "efficiency", "utilization", "training", "performance." ], "rating_list": [ 3, 6, 3, 5, 3 ], "soundness_list": [ 2, 4, 1, 3, 2 ], "presentation_list": [ 3, 4, 1, 3, 4 ], "contribution_list": [ 2, 4, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 5, 4 ] }, { "paper_id": "ezzmWTm8r6", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Thong Bach", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "PhD student" }, { "name": "Thanh Tang Nguyen", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Postdoc" } ], "n_formula": 32, "n_formula_1": 34, "n_ref_uni": 33, "n_ref": 45, "n_ref_all": 58, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1436, "formula_len_all_1": 1414, "len_all": 98542, "len_all_1": 52515, "len_abs": 1483, "len_title": 134, "len_sents": 29491, "len_sents_1": 22783, "n_sents": 202, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1490, "title": "Update larger, train faster: Stable Test-time adaptation utilizing noisy-pseudo labels", "abs": "We investigate the role of pseudo-labels in the test-time adaptation (TTA) problem. When working with unlabeled samples in TTA, pseudo-labels have become a natural approach to updating the target model. However, pseudo-label learning also presents some challenges: it suffers from a memorization effect (the model learns from clean labels first, then memorizes the noisy ones) and confirmation bias (errors from noisy labels increase over time and disrupt model performance when they become significant). Our work first identifies two underlying mechanisms leading to these obstacles. On the one hand, existing methods follow a \"slow\" adaptation to the target domain, allowing sufficient time for the model to memorize noisy labels (memorization effect) and accumulate errors (confirmation bias). Furthermore, training with noisy labels blurs the decision boundary with nearby classes. To address the first issue, we propose a novel loss function, namely sparse cross logit (sparse-CL), that operates in the logit space and allows the model to take larger learning steps in a stable training manner. This helps the target model reach a better solution faster under the same number of updating steps. To address the second issue, we introduce a regularization that penalizes negative pseudo-labels while encouraging positive ones, which can increase the boundary between nearby classes. We demonstrate that our methods outperform state-of-the-art methods in a diverse set of TTA experiments.", "keywords": [ "domain adaptation", "test-time adaptation", "online learning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 1, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 5, 5 ] }, { "paper_id": "eznTVIM3bs", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Hongyu Lin", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Jiawei Chen", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Jing Su", "gender": "Female", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Jingjing Xu", "gender": "Female", "institution": "Shanghai AI Lab", "country": "", "position": "PhD student" }, { "name": "Le Sun", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Mengjie Ren", "gender": "Female", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Wentao Chen", "gender": "Male", "institution": "ByteDance Inc.", "country": "", "position": "Researcher" }, { "name": "Xianpei Han", "gender": "Male", "institution": "Institute of Software, CAS", "country": "CN", "position": "Professor" }, { "name": "Yaojie Lu", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 29, "n_ref": 45, "n_ref_all": 68, "n_fig": 9, "n_tab": 3, "L_tab": 1222, "n_element_tab": 243, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 71, "n_element_tab_1": 10, "formula_len_all": 60, "formula_len_all_1": 38, "len_all": 139501, "len_all_1": 62286, "len_abs": 2051, "len_title": 160, "len_sents": 37109, "len_sents_1": 31012, "n_sents": 251, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 111, "L_abs": 1258, "title": "The Rise and Down of Babel Tower: Investigating the Evolution Process of Multilingual Code Large Language Model", "abs": "Large language models (LLMs) have shown significant multilingual capabilities. However, the mechanisms underlying the development of these capabilities during pre-training are not well understood. In this paper, we use code LLMs as an experimental platform to explore the evolution of multilingual capabilities in LLMs during the pre-training process. Based on our observations, we propose the Babel Tower Hypothesis, which describes the entire process of LLMs acquiring new language capabilities. During the learning process, multiple languages initially share a single knowledge system dominated by the primary language and gradually develop language-specific knowledge systems. We then validate the above hypothesis by tracking the internal states of the LLM using specific methods. Experimental results show that the internal state changes of the LLM are consistent with our Babel Tower Hypothesis. Building on these insights, we propose a novel method to construct an optimized pre-training corpus for multilingual code LLMs, which significantly outperforms LLMs trained on the original corpus. The proposed Babel Tower Hypothesis provides new insights into designing pre-training data distributions to achieve optimal multilingual capabilities in LLMs.", "keywords": [ "Large Language Model", "Multilingual", "Code" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 2, 2, 4, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "ezPbPoYFME", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "En Yu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Tiancai Wang", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Wenbing Tao", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xiangyu Zhang", "gender": "Male", "institution": "MEGVII Technology", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuang Zhang", "gender": "Not Specified", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "MS student" }, { "name": "Zhuoling Li", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 37, "n_ref": 90, "n_ref_all": 106, "n_fig": 11, "n_tab": 9, "L_tab": 2655, "n_element_tab": 383, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 480, "n_element_tab_1": 66, "formula_len_all": 412, "formula_len_all_1": 588, "len_all": 147939, "len_all_1": 69157, "len_abs": 1090, "len_title": 118, "len_sents": 41068, "len_sents_1": 31879, "n_sents": 348, "n_sents_1": 255, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1094, "title": "MOTRv3: Release-Fetch Supervision for End-to-End Multi-Object Tracking", "abs": "Although end-to-end multi-object trackers like MOTR enjoy the merits of simplicity, they suffer from the conflict between detection and association, resulting in unsatisfactory convergence dynamics. While MOTRv2 partly addresses this problem, it demands an additional detector. In this work, we serve as the first to reveal this conflict arises from unfair label assignment between detect and track queries, where detect queries are responsible for recognizing newly appearing targets and track queries are to associate them in following frames. Based on this observation, we propose MOTRv3, which balances the label assignment using the proposed release-fetch supervision strategy. In this strategy, labels are first released for detection and gradually fetched back for association. Besides, another two strategies named pseudo label distillation and track group denoising are designed to further strengthen the supervision for detection and association. Without extra detector during inference, MOTRv3 achieves impressive performance across diverse benchmarks, showing scaling up capability.", "keywords": [ "End-to-End; Multiple Object Tracking; Transformer" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 5, 5 ] }, { "paper_id": "eyJcXtZ9xv", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Hongsheng Huang", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "MS student" }, { "name": "Kaixin Shen", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Shengjun Tang", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "Researcher" }, { "name": "Siqi Du", "gender": "Not Specified", "institution": "Shenzhen University", "country": "CN", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 21, "n_ref": 36, "n_ref_all": 49, "n_fig": 2, "n_tab": 4, "L_tab": 1651, "n_element_tab": 155, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2387, "n_element_tab_1": 148, "formula_len_all": 713, "formula_len_all_1": 705, "len_all": 95630, "len_all_1": 65086, "len_abs": 1572, "len_title": 143, "len_sents": 32807, "len_sents_1": 26550, "n_sents": 237, "n_sents_1": 163, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1469, "title": "Modeling Spatiotemporal Heterogeneity in Earth Science Machine Learning: An End-to-End Approach", "abs": "In Earth sciences, unobserved factors often lead to spatially nonstationary distributions, causing relationships between features and targets to vary across locations. Traditional tabular machine learning methods struggle to effectively model this spatial heterogeneity. While approaches like Geographically Weighted Regression (GWR) capture local variations, they often miss global patterns, overfit local noise, and lack the ability to model temporal changes in spatial heterogeneity. Our research aims to model spatiotemporal heterogeneity. To achieve this, we propose an end-to-end approach that fits the entire dataset to capture global patterns, while designing the model as a conditional generative framework to learn sparse spatial heterogeneity, mitigating overfitting through localized condition sharing. Our method involves four key steps: constructing a spatiotemporal graph, encoding tabular features, aggregating spatial heterogeneity node embeddings via graph convolutions, and decoding with spatial condition vectors for location-specific predictions. We validate our approach by predicting vegetation gross primary productivity (GPP) using global climate and land cover data (2001\u20132020). Trained on 50M samples and tested on 2.8M, our model achieves an RMSE of 0.836, outperforming GWR (2.149), LightGBM (1.063) and TabNet (0.944). Visual analysis of the learned node embeddings reveals clear spatial heterogeneity patterns and their temporal dynamics.", "keywords": [ "AI4Science", "Remote Sensing", "Tabular Machine Learning", "Spatiotemporal Prediction", "Spatiotemporal Heterogeneity" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "eyBkAAeSP0", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Jun Sun", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Full Professor" }, { "name": "Wei Zhao", "gender": "Male", "institution": "Singapore Management University", "country": "", "position": "PhD student" }, { "name": "Ge Yi Li", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Postdoc" }, { "name": "Zhe Li", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 30, "n_ref": 48, "n_ref_all": 67, "n_fig": 11, "n_tab": 9, "L_tab": 4023, "n_element_tab": 215, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1412, "n_element_tab_1": 167, "formula_len_all": 149, "formula_len_all_1": 149, "len_all": 131477, "len_all_1": 64381, "len_abs": 1377, "len_title": 89, "len_sents": 41643, "len_sents_1": 29845, "n_sents": 300, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 41, "L_abs": 1388, "title": "Adversarial Suffixes May Be Features Too!", "abs": "Despite significant ongoing efforts in safety alignment, large language models (LLMs) such as GPT-4 and LLaMA 3 remain vulnerable to jailbreak attacks that can induce harmful behaviors, including those triggered by adversarial suffixes. Building on prior research, we hypothesize that these adversarial suffixes are not mere bugs but may represent features that can dominate the LLM's behavior. To evaluate this hypothesis, we conduct several experiments. First, we demonstrate that benign features can be effectively made to function as adversarial suffixes, i.e., we develop a feature extraction method to extract sample-agnostic features from benign dataset in the form of suffixes and show that these suffixes may effectively compromise safety alignment. Second, we show that adversarial suffixes generated from jailbreak attacks may contain meaningful features, i.e., appending the same suffix to different prompts results in responses exhibiting specific characteristics. Third, we show that such benign-yet-safety-compromising features can be easily introduced through fine-tuning using only benign datasets, i.e., even in the absence of harmful content. This highlights the critical risk posed by dominating benign features in the training data and calls for further research to reinforce LLM safety alignment. Our code and data is available at \\url{https://github.com/anonymous}.", "keywords": [ "LLM", "adversarial attacks", "safety alignment", "robustness" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "exnoX9Iaik", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Haotong Yang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Muhan Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Qian Tao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Shuxian Hu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xiyuan Wang", "gender": "unknown", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Zhouchen Lin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Professor" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 31, "n_ref": 48, "n_ref_all": 67, "n_fig": 4, "n_tab": 16, "L_tab": 3016, "n_element_tab": 318, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 441, "n_element_tab_1": 69, "formula_len_all": 362, "formula_len_all_1": 578, "len_all": 130929, "len_all_1": 62837, "len_abs": 1808, "len_title": 86, "len_sents": 37604, "len_sents_1": 29571, "n_sents": 317, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1500, "title": "GL-Fusion: Rethinking the Combination of Graph Neural Network and Large Language model", "abs": "Recent research on integrating Large Language Models (LLMs) with Graph Neural Networks (GNNs) typically follows two approaches: LLM-centered models, which convert graph data into tokens for LLM processing, and GNN-centered models, which use LLMs to encode text features into node and edge representations for GNN input. LLM-centered models often struggle to capture graph structures effectively, while GNN-centered models compress variable-length textual data into fixed-size vectors, limiting their ability to understand complex semantics. Additionally, GNN-centered approaches require converting tasks into a uniform, manually-designed format, restricting them to classification tasks and preventing language output. To address these limitations, we introduce a new architecture that deeply integrates GNN with LLM, featuring three key innovations: (1) Structure-Aware Transformers, which incorporate GNN\u2019s message-passing capabilities directly into LLM\u2019s transformer layers, allowing simultaneous processing of textual and structural information and generating outputs from both GNN and LLM; (2) Graph-Text Cross-Attention, which processes full, uncompressed text from graph nodes and edges, ensuring complete semantic integration; and (3) GNN-LLM Twin Predictor, enabling LLM\u2019s flexible autoregressive generation alongside GNN\u2019s scalable one-pass prediction. GL-Fusion achieves outstand performance on various tasks. Notably, it achieves state-of-the-art performance on OGBN-Arxiv and OGBG-Code2,", "keywords": [ "GNN", "LLM" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 2, 4, 4 ] }, { "paper_id": "exgLs4snap", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Chulhee Yun", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Giung Nam", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Hyunsu Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Juho Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Hongseok Yang", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 41, "n_formula_1": 23, "n_ref_uni": 66, "n_ref": 123, "n_ref_all": 153, "n_fig": 11, "n_tab": 11, "L_tab": 3555, "n_element_tab": 319, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 1506, "n_element_tab_1": 107, "formula_len_all": 2727, "formula_len_all_1": 1513, "len_all": 196369, "len_all_1": 66408, "len_abs": 5233, "len_title": 111, "len_sents": 56778, "len_sents_1": 29946, "n_sents": 436, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1374, "title": "Parameter Expanded Stochastic Gradient Markov Chain Monte Carlo", "abs": "Bayesian Neural Networks (BNNs) provide a promising framework for modeling predictive uncertainty and enhancing out-of-distribution robustness (OOD) by estimating the posterior distribution of network parameters. Stochastic Gradient Markov Chain Monte Carlo (SGMCMC) is one of the most powerful methods for scalable posterior sampling in BNNs, achieving efficiency by combining stochastic gradient descent with second-order Langevin dynamics. However, SGMCMC often suffers from limited sample diversity in practice, which affects uncertainty estimation and model performance. We propose a simple yet effective approach to enhance sample diversity in SGMCMC without the need for tempering or running multiple chains. Our approach reparameterizes the neural network by decomposing each of its weight matrices into a product of matrices, resulting in a sampling trajectory that better explores the target parameter space. This approach produces a more diverse set of samples, allowing faster mixing within the same computational budget. Notably, our sampler achieves these improvements without increasing the inference cost compared to the standard SGMCMC. Extensive experiments on image classification tasks, including OOD robustness, diversity, loss surface analyses, and a comparative study with Hamiltonian Monte Carlo, demonstrate the superiority of the proposed approach.", "keywords": [ "SGMCMC", "Bayesian Neural Network", "Parameter Expansion" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "exfy4e7OJq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Charles Wang Wai Ng", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Haoyang LI", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Research Assistant Professor" }, { "name": "Lei Chen", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Shimin Di", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "Research Assistant Professor" }, { "name": "Tianhao Tang", "gender": "unknown", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Xiaofang Zhou", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "", "position": "Full Professor" }, { "name": "JIACHUAN WANG", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Postdoc" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 41, "n_ref": 126, "n_ref_all": 156, "n_fig": 15, "n_tab": 8, "L_tab": 3356, "n_element_tab": 286, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1209, "n_element_tab_1": 124, "formula_len_all": 853, "formula_len_all_1": 599, "len_all": 170964, "len_all_1": 71606, "len_abs": 1736, "len_title": 168, "len_sents": 48663, "len_sents_1": 31793, "n_sents": 390, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 119, "L_abs": 1592, "title": "Learning Towards Emergence: Paving the Way to Induce Emergence by Inhibiting Monosemantic Neurons on Pre-trained Models", "abs": "Emergence, the phenomenon of a rapid performance increase once the model scale reaches a threshold, has achieved widespread attention recently. The literature has observed that monosemantic neurons in neural networks gradually diminish as the model scale increases. Subsequently, *Learning From Emergence* is proposed to actively inhibit monosemantic neurons in relatively small neural networks (e.g., BERT and Swin-Transformer) for promoting model performance with fine-tuning. However, to ultimately achieve emergence, it is demanding to support the monosemantic neuron inhibition in the pretraining phase of large-scale models. Thus, this work further pushes the boundary of this research direction to be *Learning Towards Emergence (L2E)* and enables the training and validating of the impact of inhibiting monosemantic neurons on larger pre-trained neural networks (e.g., Pythia-70M, 410M, and 2.8B). More specifically, to bridge the gap in current research, we first conduct experiments on models of various scales (up to 6.9B) to validate the monosemantic ideas. Then, we present a novel method L2E to address the inefficient monosemantic neuron retrieval and ineffective monosemantic neuron inhibition when existing methods are applied in the pretraining phase of large-scale models. It employs an adjustable thresholding technique for efficient neuron retrieval, incorporates a False Killing Rate metric to assess inhibition effects, and proposes a regularization-style inhibition approach, which addresses the limitations of previous approaches in both efficiency and effectiveness.", "keywords": [ "Deep Learning", "Emergent Abilities", "Monosemanticity", "Large Language Model" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "exIN7Z0wDf", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "Fuhai Chen", "gender": "Male", "institution": "Fuzhou University", "country": "CN", "position": "Assistant Professor" }, { "name": "Pengpeng Huang", "gender": "unknown", "institution": "Fuzhou University", "country": "CN", "position": "MS student" }, { "name": "Xuri Ge", "gender": "Male", "institution": "University of Glasgow", "country": "GB", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 23, "n_ref_uni": 23, "n_ref": 35, "n_ref_all": 45, "n_fig": 4, "n_tab": 6, "L_tab": 2397, "n_element_tab": 406, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2098, "n_element_tab_1": 231, "formula_len_all": 1617, "formula_len_all_1": 1601, "len_all": 103199, "len_all_1": 58392, "len_abs": 1250, "len_title": 103, "len_sents": 33617, "len_sents_1": 25606, "n_sents": 218, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1255, "title": "Multimodal Sentiment Analysis Based on Causal Reasoning", "abs": "With the rapid development of multimedia, the shift from unimodal textual sentiment analysis to multimodal image-text sentiment analysis has obtained academic and industrial attention in recent years. However, multimodal sentiment analysis is affected by unimodal data bias, e.g., text sentiment is misleading due to explicit sentiment semantic, leading to low accuracy in the final sentiment classification. In this paper, we propose a novel CounterFactual Multimodal Sentiment Analysis framework (CF-MSA) using causal counterfactual inference to construct multimodal sentiment causal inference. CF-MSA mitigates the direct effect from unimodal bias and ensures heterogeneity across modalities by differentiating the treatment variables between modalities. In addition, considering the information complementarity and bias differences between modalities, we propose a new optimisation objective to effectively integrate different modalities and reduce the inherent bias from each modality. Experimental results on two public datasets, MVSA-Single and MVSA-Multiple, demonstrate that the proposed CF-MSA has superior debiasing capability and achieves new state-of-the-art performances. We will release the code and datasets to facilitate future research.", "keywords": [ "Causal Reasoning; Multimodal Sentiment Analysis" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 1, 1, 1, 2 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "ewjN1MAnJi", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Jijun Cheng", "gender": "unknown", "institution": "East China Normal University", "country": "", "position": "MS student" }, { "name": "Lingfeng Wen", "gender": "Male", "institution": "East China Normal University", "country": "", "position": "MS student" }, { "name": "Ou Yang Ming Jie", "gender": "unknown", "institution": "East China Normal University", "country": "", "position": "MS student" }, { "name": "Xian Wei", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Principal Researcher" }, { "name": "Xiangxiang Shen", "gender": "Male", "institution": "East China Normal University", "country": "", "position": "MS student" }, { "name": "Xuan Tang", "gender": "Female", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Zheng Wan", "gender": "Male", "institution": "East China Normal University", "country": "", "position": "PhD student" }, { "name": "Licheng SUN", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" } ], "n_formula": 24, "n_formula_1": 16, "n_ref_uni": 32, "n_ref": 77, "n_ref_all": 102, "n_fig": 9, "n_tab": 14, "L_tab": 3612, "n_element_tab": 592, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1078, "n_element_tab_1": 191, "formula_len_all": 2664, "formula_len_all_1": 1866, "len_all": 171338, "len_all_1": 69032, "len_abs": 1982, "len_title": 148, "len_sents": 58199, "len_sents_1": 30525, "n_sents": 430, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1991, "title": "PDDFormer: Pairwise Distance Distribution Graph Transformer for Crystal Material Property Prediction", "abs": "The crystal structure can be simplified as a periodic point set repeating across the entire three-dimensional space along an underlying lattice. Traditionally, methods for representing crystals rely on descriptors like lattice parameters, symmetry, and space groups to characterize the structure. However, in reality, atoms in material always vibrate above absolute zero, causing continuous fluctuations in their positions. This dynamic behavior disrupts the underlying periodicity of the lattice, making crystal graphs based on static lattice parameters and conventional descriptors discontinuous under even slight perturbations. To this end, chemists proposed the Pairwise Distance Distribution (PDD) method, which has been used to distinguish all periodic structures in the world's largest real materials collection, the Cambridge Structural Database. However, achieving the completeness of PDD requires defining a large number of neighboring atoms, resulting in high computational costs. Moreover, it does not account for atomic information, making it challenging to directly apply PDD to crystal material property prediction tasks. To address these challenges, we propose the atom-Weighted Pairwise Distance Distribution (WPDD) and Unit cell Pairwise Distance Distribution (UPDD) for the first time, incorporating them into the construction of multi-edge crystal graphs. Based on this, we further developed WPDDFormer and UPDDFormer, graph transformer architecture constructed using WPDD and UPDD crystal graphs. We demonstrate that this method maintains the continuity and completeness of crystal graphs even under slight perturbations in atomic positions. Moreover, by modeling PDD as global information and integrating it into matrix-based message passing, we significantly reduced computational costs. Comprehensive evaluation results show that WPDDFormer achieves state-of-the-art predictive accuracy across tasks on benchmark datasets such as the Materials Project and JARVIS-DFT.", "keywords": [ "Pairwise Distance Distribution", "Graph Transformer", "Periodic crystal" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "ewZSzO6bts", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Akhilan Boopathy", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Ila R Fiete", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Professor" } ], "n_formula": 37, "n_formula_1": 11, "n_ref_uni": 34, "n_ref": 56, "n_ref_all": 70, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2622, "formula_len_all_1": 644, "len_all": 125869, "len_all_1": 53955, "len_abs": 1868, "len_title": 110, "len_sents": 42440, "len_sents_1": 26150, "n_sents": 332, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1884, "title": "Unified Neural Network Scaling Laws and Scale-time Equivalence", "abs": "As neural networks continue to grow in size but datasets might not, it is vital to understand how much performance improvement can be expected: is it more important to scale network size or data volume? Thus, neural network scaling laws, which characterize how test error varies with network size and data volume, have become increasingly important. However, existing scaling laws are often applicable only in limited regimes and often do not incorporate or predict well-known phenomena such as double descent. Here, we present a novel theoretical characterization of how three factors --- model size, training time, and data volume --- interact to determine the performance of deep neural networks. We first establish a theoretical and empirical equivalence between scaling the size of a neural network and increasing its training time proportionally. Scale-time equivalence challenges the current practice, wherein large models are trained for small durations, and suggests that smaller models trained over extended periods could match their efficacy. It also leads to a novel method for predicting the performance of large-scale networks from small-scale networks trained for extended epochs, and vice versa. We next combine scale-time equivalence with a linear model analysis of double descent to obtain a unified theoretical scaling law, which we confirm with experiments across vision benchmarks and network architectures. These laws explain several previously unexplained phenomena: reduced data requirements for generalization in larger models, heightened sensitivity to label noise in overparameterized models, and instances where increasing model scale does not necessarily enhance performance. Our findings hold significant implications for the practical deployment of neural networks, offering a more accessible and efficient path to training and fine-tuning large models.", "keywords": [ "generalization", "neural network", "scaling law", "double descent" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "ewRkjUX4SY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Bowen Qin", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Duanyu Feng", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "PhD student" }, { "name": "Xi Yang", "gender": "Female", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 15, "n_ref": 34, "n_ref_all": 55, "n_fig": 3, "n_tab": 11, "L_tab": 11613, "n_element_tab": 1494, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 4407, "n_element_tab_1": 572, "formula_len_all": 311, "formula_len_all_1": 311, "len_all": 185587, "len_all_1": 68574, "len_abs": 1254, "len_title": 138, "len_sents": 29582, "len_sents_1": 25489, "n_sents": 236, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1259, "title": "Conversational Few-Shot Prompting: Rethinking Few-Shot Prompting for Chat Language Model", "abs": "In-context learning, also referred to as few-shot learning, enables language models to adapt to tasks using a limited number of examples embedded in the prompt. Traditional approaches typically present all examples in a single prompt, which works well for pre-trained base models. However, the application of this method to instruction-tuned chat models, such as ChatGPT, remains underexplored.\nIn this paper, we introduce a novel conversational few-shot prompting technique, which structures few-shot examples as multi-turn conversation between the user and the assistant, rather than a single input prompt. This conversational framing better aligns with the interactive nature of chat models, enhancing their instruction-following abilities and generalization across tasks.\nThrough experiments on various benchmarks, we demonstrate that this approach significantly improves performance, particularly in low-shot scenarios, compared to traditional few-shot prompting. Our results suggest that this method provides a more flexible and robust way to leverage few-shot examples in instruction-tuned chat models, improving task performance without the need for additional fine-tuning, reducing prompt sensitivity, and offering potential for diverse applications.", "keywords": [ "Few-shot learning; In-context learning; Large language model;" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 1, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 5, 3, 3 ] }, { "paper_id": "ewBe3QpCyY", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Bo Zhao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Boyan Gao", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "David A. Clifton", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Shreyank N Gowda", "gender": "Male", "institution": "University of Nottingham", "country": "GB", "position": "Assistant Professor" }, { "name": "Timothy Hospedales", "gender": "Male", "institution": "University of Edinburgh", "country": "", "position": "Full Professor" } ], "n_formula": 30, "n_formula_1": 20, "n_ref_uni": 62, "n_ref": 114, "n_ref_all": 128, "n_fig": 3, "n_tab": 11, "L_tab": 2831, "n_element_tab": 368, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1327, "n_element_tab_1": 173, "formula_len_all": 2652, "formula_len_all_1": 1420, "len_all": 161781, "len_all_1": 67291, "len_abs": 1606, "len_title": 109, "len_sents": 42818, "len_sents_1": 30487, "n_sents": 293, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1613, "title": "Dataset Condensation with Sharpness-Aware Trajectory Matching", "abs": "Dataset condensation aims to synthesise datasets with a few representative samples that can effectively represent the original datasets. This enables efficient training and produces models with performance close to those trained on the original sets. Most existing dataset condensation methods conduct dataset learning under the bilevel (inner and outer loop) based optimisation. However, due to its notoriously complicated loss landscape and expensive time-space complexity, the preceding methods either develop advanced training protocols so that the learned datasets generalise to unseen tasks or reduce the inner loop learning cost increasing proportionally to the unrolling steps. This phenomenon deteriorates when the datasets are learned via matching the trajectories of networks trained on the real and synthetic datasets with a long horizon inner loop. To address these issues, we introduce Sharpness-Aware Trajectory Matching (SATM), which enhances the generalisation capability of learned synthetic datasets by minimising sharpness in the outer loop of bilevel optimisation. Moreover, our approach is coupled with an efficient hypergradient approximation that is mathematically well-supported and straightforward to implement along with controllable computational overhead. Empirical evaluations of SATM demonstrate its effectiveness across various applications, including standard in-domain benchmarks and out-of-domain settings. Moreover, its easy-to-implement properties afford flexibility, allowing it to integrate with other advanced sharpness-aware minimisers. We will release our code on GitHub.", "keywords": [ "dataset condensation", "meta-learning" ], "rating_list": [ 6, 5, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 4, 5, 2, 3, 2 ] }, { "paper_id": "evzUQDjXKV", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Artem Moskalev", "gender": "unknown", "institution": "Johnson and Johnson", "country": "NL", "position": "Researcher" }, { "name": "Junjie Xu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Mangal Prakash", "gender": "Male", "institution": "Johnson & Johnson", "country": "", "position": "Researcher" }, { "name": "Rui Liao", "gender": "unknown", "institution": "Johnson and Johnson", "country": "", "position": "Principal Researcher" }, { "name": "Tommaso Mansi", "gender": "Male", "institution": "Johnson and Johnson", "country": "US", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 11, "n_ref_uni": 43, "n_ref": 68, "n_ref_all": 86, "n_fig": 7, "n_tab": 4, "L_tab": 710, "n_element_tab": 95, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 72, "n_element_tab_1": 5, "formula_len_all": 2001, "formula_len_all_1": 1014, "len_all": 210020, "len_all_1": 58830, "len_abs": 1461, "len_title": 103, "len_sents": 64272, "len_sents_1": 27396, "n_sents": 477, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 1485, "title": "SE(3)-Hyena Operator for Scalable Equivariant Learning", "abs": "Modeling global geometric context while maintaining equivariance is crucial for accurate predictions in many fields such as biology, chemistry, or when modeling physical systems. Yet, this is challenging due to the computational demands of processing high-dimensional data at scale. Existing approaches such as equivariant self-attention or distance-based message passing, suffer from quadratic complexity with respect to sequence length, while localized methods sacrifice global information. Inspired by the recent success of state-space and long-convolutional models, in this work, we introduce the SE(3)-Hyena operator, the first equivariant network that adopts a long-convolutional framework for geometric systems. SE(3)-Hyena captures global geometric context at sub-quadratic complexity while maintaining equivariance to rotations and translations. Evaluated on the task of all-atom property prediction of large RNA molecules, SE(3)-Hyena matches or outperforms equivariant self-attention while requiring significantly less memory and compute for long geometric sequences. Additionally, we propose equivariant associative recall as a new mechanistic interpretability task for studying the contextual learning capabilities of equivariant models. Notably, our model processes the geometric context of $30k$ tokens $20 \\times$ faster than the equivariant transformer and allows $72 \\times$ longer context within the same memory budget. The code will be released upon the acceptance.", "keywords": [ "architecrture", "equivariance", "global context", "long convolution", "scalability", "mechanistic interpretability" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 1, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "evyIlAvQ6J", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Abby Bertics", "gender": "Female", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Elizabeth Chrastil", "gender": "unknown", "institution": "University of California, Irvine", "country": "US", "position": "Associate Professor" }, { "name": "Jean Carlson", "gender": "Female", "institution": "University of California, Santa Barbara", "country": "US", "position": "Full Professor" }, { "name": "Nina Miolane", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 22, "n_ref": 26, "n_ref_all": 36, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 137, "formula_len_all_1": 150, "len_all": 96129, "len_all_1": 60777, "len_abs": 1857, "len_title": 110, "len_sents": 35408, "len_sents_1": 29485, "n_sents": 297, "n_sents_1": 241, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1945, "title": "Flow Tree: A dynamic model for navigation paths and strategies", "abs": "Navigation is a dynamic process that involves learning how to represent the environment, along with positions in and trajectories through it. Spatial navigation skills vary significantly among individual humans. But what exactly differentiates a good navigator from a bad one, or an easy-to-navigate path from a hard one, is not well understood. Several studies have analysed exploration and navigation behaviour using static quantitative measures, like counts of positions visited or distance travelled. These static measures, however, are inherently limited in their ability to describe dynamic behaviors, providing a coarse quantification of the navigation process. To fill this gap, we introduce the \\emph{Flow Tree}, a novel data structure, which quantifies the dynamics of a group of trajectories through time. This is a discrete adaptation of the Reeb graph, a mathematical structure from topology, computed from multiple trajectories (from different people or the same person over time). Each divergence in trajectory is captured as a node, encoding the variability of the collection of trajectories. A Flow Tree encodes how difficult it will be to navigate a certain path for a group of humans. We apply the Flow Tree to a behavioural dataset of 100 humans exploring and then navigating a small, closed-form maze in virtual reality. In this paper we (1) describe what a Flow Tree is and how to calculate it, (2) show that Flow Trees can be used to predict path difficulty more effectively than static metrics, and (3) demonstrate that a trajectory through the Flow Tree is predictive of that individual's success. We (4) introduce a hypothesis testing framework over Flow Trees to quantitatively differentiate between the strategies of the best navigators from those of worst. Thus, we show that Flow Trees are a powerful tool to analyse dynamic trajectory data.\\footnote{The code will be made publicly available at [anon-github-link].}", "keywords": [ "dynamics", "navigation", "behavior", "trees", "mazes", "cognitive representation", "neuroscience" ], "rating_list": [ 3, 6, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "evDSvZBFRP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Ahmed Rayen Mhadhbi", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Undergrad student" }, { "name": "Matthias Althoff", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Associate Professor" }, { "name": "Tobias Ladner", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 44, "n_formula_1": 32, "n_ref_uni": 43, "n_ref": 109, "n_ref_all": 137, "n_fig": 6, "n_tab": 5, "L_tab": 1955, "n_element_tab": 46, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 917, "n_element_tab_1": 32, "formula_len_all": 3562, "formula_len_all_1": 3001, "len_all": 155208, "len_all_1": 63237, "len_abs": 886, "len_title": 123, "len_sents": 42278, "len_sents_1": 25203, "n_sents": 372, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 917, "title": "Towards Formally Verifying LLMs: Taming the Nonlinearity of the Transformer", "abs": "Large language models are increasingly used across various domains,\n which raises important safety concerns, particularly regarding adversarial attacks.\n While recent advancements in formal neural network verification have shown promising results, \n the complexity of transformers, the backbone of large language models, poses unique challenges for formal robustness verification.\n Traditional convex relaxation methods often result in large approximation errors due to the transformer's parallel, nonlinear attention heads.\n In this work, we address these limitations by introducing a novel approach based on non-convex, set-based computing to preserve the nonlinear dependencies through a transformer.\n Our approach generalizes previous methods on robustness verification of transformers,\n and the desired precision is tunable at the cost of additional computation time with a single parameter.", "keywords": [ "large language models", "formal verification", "set-based computing", "matrix polynomial zonotopes", "neural networks" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 2, 3, 1, 1 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "ev2KNFmvux", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Fandong Meng", "gender": "Male", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Jie Zhou", "gender": "Male", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Liqiang Niu", "gender": "Male", "institution": "WeChat AI, Tencent Inc", "country": "CN", "position": "Researcher" }, { "name": "Wenchao Chen", "gender": "Male", "institution": "Tencent Inc", "country": "CN", "position": "Researcher" }, { "name": "Ziyao Lu", "gender": "Male", "institution": "WeChat AI", "country": "", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 40, "n_ref": 65, "n_ref_all": 79, "n_fig": 6, "n_tab": 6, "L_tab": 1788, "n_element_tab": 209, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1530, "n_element_tab_1": 184, "formula_len_all": 218, "formula_len_all_1": 218, "len_all": 110913, "len_all_1": 53502, "len_abs": 1144, "len_title": 119, "len_sents": 24020, "len_sents_1": 22894, "n_sents": 200, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1155, "title": "MaskMamba: A Hybrid Mamba-Transformer Model for Masked Image Generation", "abs": "Image generation models have encountered challenges related to scalability and quadratic complexity, primarily due to the reliance on Transformer-based backbones. In this study, we introduce MaskMamba, a novel hybrid model that integrates Mamba and Transformer architectures, utilizing Masked Image Modeling for non-autoregressive image synthesis. We meticulously redesign the bidirectional Mamba architecture by implementing two key modifications: (1) replacing causal convolutions with standard convolutions to better capture global context, and (2) utilizing concatenation instead of multiplication, which significantly boosts performance while accelerating inference speed. Additionally, we explore various hybrid schemes of MaskMamba, including both serial and grouped parallel arrangements. Furthermore, we incorporate an in-context condition that allows our model to perform both class-to-image and text-to-image generation tasks. Our MaskMamba outperforms Mamba-based and Transformer-based models in generation quality. Notably, it achieves a remarkable 54.44\\% improvement in inference speed at a resolution of $2048\\times 2048$ over Transformer.", "keywords": [ "Masked Image Modeling", "Bidirectional Mamba", "In-context Condition", "Non-autoregressive image synthesis" ], "rating_list": [ 5, 6, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 3 ] }, { "paper_id": "euZD4YTXKu", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 6, "author_info_list": [ { "name": "David E. Keyes", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Full Professor" }, { "name": "Di Wang", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Hang Xu", "gender": "Male", "institution": "KAUST", "country": "", "position": "PhD student" }, { "name": "Jie Ren", "gender": "unknown", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Junxiao Wang", "gender": "Male", "institution": "Guangzhou University", "country": "CN", "position": "Associate Professor" }, { "name": "Liangyu Wang", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 25, "n_ref": 38, "n_ref_all": 58, "n_fig": 9, "n_tab": 4, "L_tab": 1475, "n_element_tab": 132, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1032, "n_element_tab_1": 107, "formula_len_all": 76, "formula_len_all_1": 76, "len_all": 146869, "len_all_1": 63525, "len_abs": 2157, "len_title": 123, "len_sents": 37746, "len_sents_1": 31192, "n_sents": 252, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1704, "title": "ZO-Offloading: Fine-Tuning LLMs with 100 Billion Parameters on a Single GPU", "abs": "Fine-tuning pre-trained LLMs typically requires a vast amount of GPU memory. Standard first-order optimizers like SGD face a significant challenge due to the large memory overhead from back-propagation as the size of LLMs increases, which necessitates caching activations during the forward pass and gradients during the backward pass. In contrast, zeroth-order (ZO) methods can estimate gradients with only two forward passes and without the need for activation caching. Additionally, CPU resources can be aggregated and offloaded to extend the memory and computational capacity of a single GPU.\nTo enable efficient fine-tuning of LLMs on a single GPU, we introduce ZO-Offloading, a framework that strategically utilizes both CPU and GPU resources for ZO. ZO-Offloading dynamically offloads model parameters to the CPU and retrieves them to the GPU as needed, ensuring continuous and efficient computation by reducing idle times and maximizing GPU utilization. Parameter updates are integrated with ZO's dual forward passes to minimize redundant data transfers, thereby improving the overall efficiency of the fine-tuning process. The ZO-Offloading framework also incorporates a novel low-bit precision technique for managing data transfers between the CPU and GPU in AMP mode, as well as asynchronous checkpointing for LLM fine-tuning.\nWith ZO-Offloading, for the first time, it becomes possible to fine-tune extremely large models, such as the OPT-175B with over $\\textbf{175 billion}$ parameters, on a single GPU with just $\\textbf{24GB}$ of memory\u2014a feat unattainable with conventional methods. Moreover, our framework operates without any additional time cost compared to standard ZO methodologies.", "keywords": [ "LLMs", "zeroth-order optimization", "efficient CPU-offloading", "memory efficient fine-tuning" ], "rating_list": [ 1, 3, 3, 8 ], "soundness_list": [ 1, 2, 2, 1 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "eu1PIDPYwC", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Bolin Ding", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Senior Director" }, { "name": "Jun Zhang", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Yaliang Li", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Staff Engineer" }, { "name": "Yuchang Sun", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yuexiang Xie", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Staff" } ], "n_formula": 63, "n_formula_1": 30, "n_ref_uni": 48, "n_ref": 108, "n_ref_all": 132, "n_fig": 2, "n_tab": 3, "L_tab": 586, "n_element_tab": 116, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 669, "n_element_tab_1": 124, "formula_len_all": 6752, "formula_len_all_1": 1655, "len_all": 172333, "len_all_1": 67719, "len_abs": 1201, "len_title": 108, "len_sents": 36728, "len_sents_1": 28855, "n_sents": 283, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1204, "title": "Exploring Selective Layer Fine-Tuning in Federated Learning", "abs": "Federated learning (FL) has emerged as a promising paradigm for fine-tuning foundation models using distributed data in a privacy-preserving manner. Under limited computational resources, clients often find it more practical to fine-tune a selected subset of layers, rather than the entire model, based on their task-specific data. In this study, we provide a thorough theoretical exploration of selective layer fine-tuning in FL, emphasizing a flexible approach that allows the clients to adjust their selected layers according to their local data and resources. We theoretically demonstrate that the layer selection strategy has a significant impact on model convergence in two critical aspects: the importance of selected layers and the heterogeneous choices across clients. Drawing from these insights, we further propose a strategic layer selection method that utilizes local gradients and regulates layer selections across clients. Extensive experiments on both image and text datasets demonstrate the effectiveness of the proposed strategy compared with several baselines, highlighting its advances in identifying critical layers that adapt to the client heterogeneity and training dynamics in FL.", "keywords": [ "Federated Learning", "Selective Layer Fine-tuning" ], "rating_list": [ 3, 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 2, 1, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "etxbRucurT", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Fei Chao", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Xiawu Zheng", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuexiao Ma", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Zhaohao Xie", "gender": "Male", "institution": "Xiamen University", "country": "", "position": "MS student" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 29, "n_ref": 51, "n_ref_all": 63, "n_fig": 0, "n_tab": 6, "L_tab": 971, "n_element_tab": 156, "n_fig_1": 0, "n_tab_1": 7, "L_tab_1": 1504, "n_element_tab_1": 164, "formula_len_all": 1041, "formula_len_all_1": 1037, "len_all": 130394, "len_all_1": 65531, "len_abs": 1482, "len_title": 102, "len_sents": 31567, "len_sents_1": 31281, "n_sents": 213, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 1629, "title": "Automated Fine-Grained Mixture-of-Experts Quantization", "abs": "Mixture of Experts (MoE) enables efficient parameter scaling in large language models by dynamically activating relevant parameter subsets per input token. \nCompressing MoE models presents unique challenges due to their inherent sparsity. Traditional quantization techniques, which are typically effective for dense models, prove inadequate when applied to MoE architectures.\nThis paper proposes an efficient MoE quantization algorithm.\nWe propose a fine-grained, adaptive quantization approach coupled with an efficient method for determining optimal configurations.\nSpecifically, we construct a mixed-precision quantization search space encompassing different granularities from expert-level to channel-level. \nThis approach facilitates precise bit-width resource allocation across model components based on their significance and activation frequency. \nAnd then, we leverage evolutionary algorithms to efficiently navigate this search space, autonomously identifying optimal quantization configurations. \nThe synergy between adaptive granularity and automated search effectively mitigates the distinctive quantization challenges inherent to MoE models, culminating in a fully automated framework for efficient MoE quantization.\nExperimental results indicate that our method achieves significant performance improvements across multiple evaluation tasks, with particularly notable results in low-bit quantization scenarios.\nWhen applied to the Mixtral-8x7b-v0.1 model, our approach outperforms the current state-of-the-art by $9.24$\\% , setting a new benchmark in MoE quantization. Code is available in supplementary materials.", "keywords": [ "quantization" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 5, 5 ] }, { "paper_id": "etrY4TegYb", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 1, "author_info_list": [ { "name": "yeyunchen", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 28, "n_ref_uni": 25, "n_ref": 51, "n_ref_all": 68, "n_fig": 7, "n_tab": 5, "L_tab": 1657, "n_element_tab": 215, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 946, "n_element_tab_1": 51, "formula_len_all": 2509, "formula_len_all_1": 2089, "len_all": 118064, "len_all_1": 65723, "len_abs": 1774, "len_title": 132, "len_sents": 34575, "len_sents_1": 29030, "n_sents": 280, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1785, "title": "Learning Relational Invariance for Out-of-Distribution Molecular Relational Learning", "abs": "Molecular Relational Learning (MRL) expands the scope of molecular representation learning by incorporating additional molecules, aiming to understand the interactions between pairs of molecules. While MRL has shown promising results, the existing methods have not been able to generalise to real world scenarios. Invariant learning is pivotal in addressing Out-of-Distribution (OOD) generalization challenges. However, two major obstacles impede the progress of invariant learning in MRL: (1) Unlike single-molecular cases, interactions between molecules introduce added complexity, with a heavy reliance on molecular substructure recognition, often leading to the misspecification of invariant patterns. (2) Accurate modeling of interactions can effectively improve generalizations. However, previous methods focus on node interaction, which is limited by the expressiveness of GNN, and long-range interactions cannot be captured. To address these, we propose a novel Relational Invariant Learning (RIL) framework that uses a multi-granularity interaction approach to improve OOD generalization for MRL, and the framework is denoted as RILOOD. Specifically, we model the environment diversity distribution of molecules by mixup-based Conditional Modeling. Then, we employ a multi-granularity refinement strategy to learn the Context-Aware Representation, which is essential for capturing multi-level interaction. We further design an invariant learning module to capture the invariant patterns that robustly generalize across unseen environments. Extensive experiments on molecular datasets show that our method achieves stronger generalization against state-of-the-art methods in the presence of various distribution shifts. Our code will be released after our paper is accepted.", "keywords": [ "molecular relational learning", "out of distribution", "invariant learning" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "etif9j1CnG", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Ahmed Imtiaz Humayun", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Candice Schumann", "gender": "Non-Binary", "institution": "Google Deepmind", "country": "US", "position": "Researcher" }, { "name": "Cristina Nader Vasconcelos", "gender": "Female", "institution": "Research, Google", "country": "CA", "position": "Researcher" }, { "name": "Deepak Ramachandran", "gender": "Male", "institution": "Google", "country": "US", "position": "Staff Researcher" }, { "name": "Golnoosh Farnadi", "gender": "Female", "institution": "McGill University", "country": "CA", "position": "Assistant Professor" }, { "name": "Ibtihel Amara", "gender": "Female", "institution": "Google", "country": "CA", "position": "Intern" }, { "name": "Katherine Heller", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Negar Rostamzadeh", "gender": "Female", "institution": "Google", "country": "CA", "position": "Research Scientist" }, { "name": "Junfeng He", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Mohammad Havaei", "gender": "Male", "institution": "Google", "country": "CA", "position": "Reseach scientist " } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 28, "n_ref": 41, "n_ref_all": 62, "n_fig": 7, "n_tab": 3, "L_tab": 1385, "n_element_tab": 287, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 416, "n_element_tab_1": 89, "formula_len_all": 443, "formula_len_all_1": 407, "len_all": 111005, "len_all_1": 65686, "len_abs": 1317, "len_title": 90, "len_sents": 34619, "len_sents_1": 31118, "n_sents": 257, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1233, "title": "What Secrets Do Your Manifolds Hold? Understanding the Local Geometry of Generative Models", "abs": "Deep Generative Models are frequently used to learn continuous representations of complex data distributions using a finite number of samples. For any generative model, including pre-trained foundation models with GAN, Transformer or Diffusion architectures, generation performance can vary significantly based on which part of the learned data manifold is sampled. In this paper we study the post-training local geometry of the learned manifold and its relationship to generation outcomes for models ranging from toy settings to the latent decoder of the near state-of-the-art Stable Diffusion 1.4 Text-to-Image model. Building on the theory of continuous piecewise-linear (CPWL) generators, we characterize the local geometry in terms of three geometric descriptors - scaling ($\\psi$), rank ($\\nu$), and complexity ($\\delta$). We provide quantitative and qualitative evidence showing that for a given latent, the local descriptors are indicative of generation aesthetics, artifacts, diversity, and memorization. Finally we demonstrate that training a reward model using the local geometry allows us to control the log-likelihood of a generated sample under the learned distribution, and improve the qualitative aspects of an image.", "keywords": [ "Geometry", "Diffusion models", "VAE", "Generative Models", "Guidance", "Memorization", "Out-of-Distribution Detection" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "etUJR2xBYa", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Hongyu Sun", "gender": "Male", "institution": "University of Birmingham", "country": "GB", "position": "PhD student" }, { "name": "Min Hua", "gender": "Female", "institution": "University of Birmingham", "country": "GB", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 21, "n_ref": 32, "n_ref_all": 58, "n_fig": 16, "n_tab": 9, "L_tab": 5990, "n_element_tab": 396, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 3437, "n_element_tab_1": 254, "formula_len_all": 436, "formula_len_all_1": 429, "len_all": 136158, "len_all_1": 60487, "len_abs": 858, "len_title": 132, "len_sents": 32935, "len_sents_1": 25561, "n_sents": 255, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 863, "title": "High-quality and controllable time series generation with diffusion in transformers", "abs": "Current research on time series generation frequently depends on oversimplified data and lenient evaluation methods, making it challenging to apply these models effectively in real-world scenarios. Diffusion in Transformers (DiT) has demonstrated that the traditional inductive biases in neural networks are unnecessary. This paper shows that the advantages of DiT can be extended to time series generation. We add the attention mask and dilated causal convolution to introduce the temporal characteristic. Additionally, we introduce a novel smooth guidance policy for style control during generation, leveraging a property of the diffusion process. Furthermore, our proposed model can generate longer sequences with training in short sequences. Experimental results reveal that our variant of DiT achieves state-of-the-art performance across various data types.", "keywords": [ "Diffusion in Transformers", "time series generation" ], "rating_list": [ 3, 3, 3, 6, 6 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 3, 1, 1, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 2, 3 ] }, { "paper_id": "etToTig9Fp", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Bill Y Lin", "gender": "Male", "institution": "xAI", "country": "US", "position": "Researcher" }, { "name": "Qi Jia", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Xiang Yue", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" }, { "name": "Yizhu Liu", "gender": "unknown", "institution": "Meituan", "country": "", "position": "Instructor" }, { "name": "Ziheng Qin", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Flora Huang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 35, "n_ref": 54, "n_ref_all": 84, "n_fig": 10, "n_tab": 8, "L_tab": 3674, "n_element_tab": 268, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1216, "n_element_tab_1": 157, "formula_len_all": 925, "formula_len_all_1": 712, "len_all": 155185, "len_all_1": 65268, "len_abs": 1849, "len_title": 81, "len_sents": 45006, "len_sents_1": 29621, "n_sents": 370, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 33, "L_abs": 1856, "title": "Visual Perception in Text Strings", "abs": "Understanding visual semantics embedded in consecutive characters is a crucial capability for both large language models (LLMs) and multi-modal large language models (MLLMs). This type of artifact possesses the unique characteristic that identical information can be readily formulated in both texts and images, making them a significant proxy for analyzing modern LLMs' and MLLMs' capabilities in modality-agnostic vision understanding. In this work, we select ASCII art as a representative artifact, where the lines and brightness used to depict each concept are rendered by characters, and we frame the problem as an ASCII art recognition task. We benchmark model performance on this task by constructing an evaluation dataset with an elaborate categorization tree and also collect a training set to elicit the models' visual perception ability. Through a comprehensive analysis of dozens of models, results reveal that although humans can achieve nearly 100% accuracy, the state-of-the-art LLMs and MLLMs lag far behind. Models are capable of recognizing concepts depicted in the ASCII arts given only text inputs indicated by over 60% accuracy for some concepts, but most of them achieves merely around 30% accuracy when averaged across all categories. When provided with images as inputs, GPT-4o gets 82.68%, outperforming the strongest open-source MLLM by 21.95%. Although models favor different kinds of ASCII art depending on the modality provided, none of the MLLMs successfully benefit when both modalities are supplied simultaneously. Moreover, supervised fine-tuning helps improve models' accuracy especially when provided with the image modality, but also highlights the need for better training techniques to enhance the information fusion among modalities. All resources are available at https://anonymous.4open.science/r/VisionInText-08D3.", "keywords": [ "Large Language Model", "Multi-modal Large Language Model", "Visual Perception", "ASCII Art" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "et5l9qPUhm", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Arjun Subramonian", "gender": "Agender", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Elvis Dohmatob", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Julia Kempe", "gender": "Female", "institution": "New York University", "country": "US", "position": "Professor" }, { "name": "Yunzhen Feng", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Intern" } ], "n_formula": 166, "n_formula_1": 22, "n_ref_uni": 35, "n_ref": 114, "n_ref_all": 153, "n_fig": 12, "n_tab": 1, "L_tab": 80, "n_element_tab": 8, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 14603, "formula_len_all_1": 2277, "len_all": 229661, "len_all_1": 74974, "len_abs": 1191, "len_title": 69, "len_sents": 67067, "len_sents_1": 34430, "n_sents": 619, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 21, "L_abs": 1195, "title": "Strong Model Collapse", "abs": "Within the scaling laws paradigm, which underpins the training of large neural networks like ChatGPT and Llama, we consider a supervised regression setting and establish a strong form of the model collapse phenomenon, a critical performance degradation due to synthetic data in the training corpus. Our results show that even the smallest fraction of synthetic data (e.g., as little as 1 per 1000) can still lead to model collapse: larger and larger training sets do not enhance performance. We further investigate whether increasing model size, an approach aligned with current trends in training large language models, exacerbates or mitigates model collapse. In a simplified regime where neural networks are approximated via random projections of tunable size, we both theoretically and empirically show that larger models can amplify model collapse. Interestingly, our theory also indicates that, beyond the interpolation threshold (which can be extremely high for very large datasets), larger models may mitigate the collapse, although they do not entirely prevent it. Our theoretical findings are empirically verified through experiments on language models and neural networks for images.", "keywords": [ "Model Collapse", "Regression", "High dimensional asymptotics", "Synthetic Data", "Scaling Laws" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "eszQcR5F1e", "primary_area": "reinforcement learning", "n_author": 12, "author_info_list": [ { "name": "David Arturo Soriano Valdez", "gender": "Male", "institution": "University of Auckland", "country": "NZ", "position": "Postdoc" }, { "name": "Dr Padriac Amato Tahua O'Leary", "gender": "unknown", "institution": "University of Auckland", "country": "NZ", "position": "Lecturer" }, { "name": "Gael Gendron", "gender": "unknown", "institution": "University of Auckland", "country": "NZ", "position": "PhD student" }, { "name": "Jiamou Liu", "gender": "Male", "institution": "The University of Auckland", "country": "NZ", "position": "Associate Professor" }, { "name": "Kobe Knowles", "gender": "unknown", "institution": "University of Auckland", "country": "NZ", "position": "PhD student" }, { "name": "Mihailo Azhar", "gender": "unknown", "institution": "Aarhus University", "country": "DK", "position": "Postdoc" }, { "name": "Mitchell Rogers", "gender": "Male", "institution": "University of Auckland", "country": "NZ", "position": "PhD student" }, { "name": "Patrice Jean Delmas", "gender": "Not Specified", "institution": "The University of Auckland", "country": "", "position": "Associate Professor" }, { "name": "Shahrokh Heidari", "gender": "Male", "institution": "University of Auckland", "country": "", "position": "PhD student" }, { "name": "Yang Chen", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "Postdoc" }, { "name": "Yitan Zhang", "gender": "unknown", "institution": "University of Auckland", "country": "NZ", "position": "MS student" }, { "name": "Michael J. Witbrock", "gender": "Male", "institution": "University of Auckland", "country": "NZ", "position": "Professor" } ], "n_formula": 19, "n_formula_1": 17, "n_ref_uni": 54, "n_ref": 107, "n_ref_all": 130, "n_fig": 9, "n_tab": 5, "L_tab": 1707, "n_element_tab": 146, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 523, "n_element_tab_1": 50, "formula_len_all": 2022, "formula_len_all_1": 1198, "len_all": 151459, "len_all_1": 71722, "len_abs": 1694, "len_title": 162, "len_sents": 44341, "len_sents_1": 33897, "n_sents": 303, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 114, "L_abs": 1701, "title": "Adversarial Inverse Reward-Constraint Learning with Reward-Feasibility Contrast Prior Inspired by Animal Behaviour", "abs": "The behaviour of natural and artificial agents is shaped by underlying reward systems, which signal rewards based on internal and external factors, driving reward-oriented actions. However, real-world scenarios often impose constraints that reward alone cannot capture. While existing inverse (constrained) reinforcement learning methods can recover either rewards or constraints from demonstrations, the simultaneous inference of both remains unexplored due to the complexity of inference and the lack of knowledge of their relationship. To address this gap, we propose a novel algorithm that simultaneously infers both rewards and constraints within an adversarial learning framework, where both are updated through a policy optimisation process guided by expert demonstrations. Crucial to this framework is the introduction of the \u201creward-feasibility contrast prior,\u201d a hypothesis that correlates rewards and constraints. It is inspired by patterns observed in animal behaviour (particularly meerkats), positing that states with high rewards nearby are more likely to be associated with weaker feasibility (stronger constraints).\nOur experiments on virtual robot control tasks with safety constraints and real-world animal behaviour data with spatio-temporal causal constraints validate our proposed framework's effectiveness and the reward-feasibility contrast prior hypothesis. The results show accurate recovery of rewards and constraints, reflected by strong alignment with expert demonstrations and a low rate of constraint violations. Additionally, the performance improvement by embedding this prior into other inverse constraint inference methods further confirms its general effectiveness.", "keywords": [ "inverse reinforcement learning", "inverse constraint inference", "simultaneous reward-constraint inference", "animal behaviour" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "esjkcic7Oh", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Chenwei Xu", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" }, { "name": "Jerry Yao-Chieh Hu", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Yibo Wen", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "US", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 41, "n_ref": 96, "n_ref_all": 107, "n_fig": 4, "n_tab": 5, "L_tab": 6310, "n_element_tab": 902, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 292, "n_element_tab_1": 31, "formula_len_all": 1808, "formula_len_all_1": 1392, "len_all": 169780, "len_all_1": 67726, "len_abs": 1461, "len_title": 125, "len_sents": 44897, "len_sents_1": 33162, "n_sents": 367, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1307, "title": "AlignAb: Pareto-Optimal Energy Alignment for Designing Nature-Like Antibodies", "abs": "We present a three-stage framework for training deep learning models specializing in antibody sequence-structure co-design.\nWe first pre-train a language model based on millions of antibody sequence data.\nThen, we employ the learned representations to guide the training of a diffusion model for joint optimization over both sequence and structure of antibodies. \nDuring the final alignment stage, we optimize the model to favor antibodies with low repulsion and high attraction to the antigen binding site, enhancing the rationality and functionality of the design.\nTo mitigate conflicting energy preferences, we extend AbDPO (Antibody Direct Preference Optimization) to guide the model towards Pareto optimality under multiple energy-based alignment objectives. \nFurthermore, we adopt an iterative learning paradigm with temperature scaling, enabling the model to benefit from diverse online datasets without requiring additional data.\nIn practice, our proposed methods achieve high stability and efficiency in producing a better Pareto front of antibody designs compared to top samples generated by baselines and previous alignment techniques.\nThrough extensive experiments, we showcase the superior performance of our methods in generating nature-like antibodies with high binding affinity consistently.", "keywords": [ "Antibody Design", "Diffusion Models", "Multi-Objective Optimization" ], "rating_list": [ 6, 3, 6, 3 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "esf4Lduba2", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 18, "author_info_list": [ { "name": "Antonio Henrique de Oliveira Fonseca", "gender": "Male", "institution": "Yale University", "country": "", "position": "PhD student" }, { "name": "Benjamin W Christensen", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "MS student" }, { "name": "Chadi Abdallah", "gender": "Male", "institution": "Baylor College of Medicine", "country": "", "position": "Associate Professor" }, { "name": "Christopher Lee Averill", "gender": "Male", "institution": "Baylor College of Medicine", "country": "", "position": "Researcher" }, { "name": "David Zhang", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Undergrad student" }, { "name": "David van Dijk", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" }, { "name": "Emanuele Zappala", "gender": "Male", "institution": "Idaho State University", "country": "US", "position": "Assistant Professor" }, { "name": "Haoran Lyu", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Researcher" }, { "name": "Maria Brbic", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "", "position": "Assistant Professor" }, { "name": "Maryam Bagherian", "gender": "Female", "institution": "Yale University", "country": "", "position": "Postdoc" }, { "name": "Nazreen Pallikkavaliyaveetil MohammedSheriff", "gender": "Female", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "Research Fellow" }, { "name": "Nhi Nguyen", "gender": "Female", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Rahul Madhav Dhodapkar", "gender": "Male", "institution": "Yale University", "country": "", "position": "PhD student" }, { "name": "Syed A Rizvi", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Rex Ying", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" }, { "name": "Zhuoyang Lyu", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Undergrad student" }, { "name": "amin karbasi", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Associate Professor" }, { "name": "Josue A Ortega", "gender": "Male", "institution": "Yale University", "country": "", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 44, "n_ref": 98, "n_ref_all": 114, "n_fig": 7, "n_tab": 4, "L_tab": 2998, "n_element_tab": 146, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2631, "n_element_tab_1": 108, "formula_len_all": 631, "formula_len_all_1": 604, "len_all": 146398, "len_all_1": 56158, "len_abs": 917, "len_title": 121, "len_sents": 45898, "len_sents_1": 25467, "n_sents": 281, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 920, "title": "FIMP: Foundation Model-Informed Message Passing for Graph Neural Networks", "abs": "Foundation models have achieved remarkable success across many domains, relying on pretraining over vast amounts of data. Graph-structured data often lacks the same scale as unstructured data, making the development of graph foundation models challenging. In this work, we propose Foundation-Informed Message Passing (FIMP), a Graph Neural Network (GNN) message-passing framework that repurposes existing pretrained non-textual foundation models for graph-based tasks. We show that the self-attention layers of foundation models can effectively be leveraged on graphs to perform cross-node attention-based message-passing. Our model is evaluated across diverse domains on image networks, single-cell RNA sequencing, and fMRI brain activity recordings in finetuned and zero-shot settings. FIMP outperforms strong baselines, demonstrating that it can effectively leverage state-of-the-art foundation models in graph tasks.", "keywords": [ "Graph Neural Networks", "Message-Passing", "Foundation Models" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 4 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "esYrEndGsr", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Alexander Immer", "gender": "Male", "institution": "Bioptimus", "country": "FR", "position": "Researcher" }, { "name": "Bruno Kacper Mlodozeniec", "gender": "Not Specified", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "David Krueger", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Assistant Professor" }, { "name": "Juhan Bae", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Richard E. Turner", "gender": "Male", "institution": "Alan Turing Institute", "country": "GB", "position": "Researcher" }, { "name": "Runa Eschenhagen", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "PhD student" } ], "n_formula": 62, "n_formula_1": 24, "n_ref_uni": 35, "n_ref": 114, "n_ref_all": 214, "n_fig": 22, "n_tab": 6, "L_tab": 1278, "n_element_tab": 80, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4884, "formula_len_all_1": 2109, "len_all": 219931, "len_all_1": 70322, "len_abs": 1390, "len_title": 117, "len_sents": 80901, "len_sents_1": 30825, "n_sents": 565, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 69, "L_abs": 1306, "title": "Influence Functions for Scalable Data Attribution in Diffusion Models", "abs": "Diffusion models have led to significant advancements in generative modelling. Yet their widespread adoption poses challenges regarding data attribution and interpretability. In this paper, we aim to help address such challenges in diffusion models by extending influence functions. Influence function-based data attribution methods approximate how a model's output would have changed if some training data were removed. In supervised learning, this is usually used for predicting how the loss on a particular example would change. For diffusion models, we focus on predicting the change in the probability of generating a particular example via several proxy measurements. We show how to formulate influence functions for such quantities and how previously proposed methods can be interpreted as particular design choices in our framework. To ensure scalability of the Hessian computations in influence functions, we use a K-FAC approximation based on generalised Gauss-Newton matrices specifically tailored to diffusion models. We show that our recommended method outperforms previously proposed data attribution methods on common data attribution evaluations, such as the Linear Data-modelling Score (LDS) or retraining without top influences, without the need for method-specific hyperparameter tuning.", "keywords": [ "diffusion", "influence functions", "K-FAC", "Generalised Gauss Newton", "data attribution", "Hessian approximation", "GGN", "interpretability" ], "rating_list": [ 10, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 4, 3, 3 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "erowpbZcPi", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Feng Pan", "gender": "Male", "institution": "Centre for Quantum Techonolgies", "country": "SG", "position": "Postdoc" }, { "name": "Lirand\u00eb Pira", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Patrick Rebentrost", "gender": "unknown", "institution": "National University of Singapore", "country": "", "position": "Researcher" }, { "name": "Wenxin Zhu", "gender": "Female", "institution": "University of Oxford", "country": "GB", "position": "Undergrad student" }, { "name": "Yixian Qiu", "gender": "Female", "institution": "national university of singaore, National University of Singapore", "country": "", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 14, "n_ref_uni": 42, "n_ref": 74, "n_ref_all": 83, "n_fig": 3, "n_tab": 2, "L_tab": 205, "n_element_tab": 18, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 113, "n_element_tab_1": 15, "formula_len_all": 1165, "formula_len_all_1": 788, "len_all": 112729, "len_all_1": 61064, "len_abs": 1652, "len_title": 97, "len_sents": 31192, "len_sents_1": 28556, "n_sents": 236, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1662, "title": "Tilted Losses in Training Quantum Neural Networks", "abs": "Empirical risk minimization is a fundamental paradigm in the optimization process of machine learning (ML) models. Several techniques extend this idea by introducing parameters which further regularize this strategy in training these models. One of these paradigms is the so-called tilted empirical risk minimization (TERM), which uses a tilted hyperparameter to penalize the presence of outliers, which represent data samples that differ significantly from the rest of the dataset. Quantum machine learning (QML) models have been studied and benchmarked across various criteria stemming from classical ML, including their training via the parameter-shift rule. Therefore, it is natural to extend the concept of TERM in training QML models, namely the type of models known as quantum neural networks (QNNs). In this work, we examine the impact of a tilted loss function in training a class of QNNs, specifically for binary classification tasks involving two different datasets with induced class imbalance. In the first dataset, the Iris dataset, we show that varying the value of the tilted hyperparameter modifies the decision boundary leading to reduced importance of outliers and better training accuracy --- highlighting the importance of using tilted risk minimization. Additionally, in a synthetic dataset we validate that the training accuracy can be improved using the tilted parameter. Analytically, we extend the parameter-shift training method to accommodate weighted inputs by introducing the tilted hyperparameter for training QNNs. These results highlight the significance of incorporating regularization techniques from ML models into QML models.", "keywords": [ "Quantum machine learning", "exponential tilting", "empirical risk minimization" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 1, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "erWwBoR59l", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Clement Chatelain", "gender": "Male", "institution": "INSA Rouen Normandy", "country": "FR", "position": "Associate Professor" }, { "name": "Eric Granger", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure", "country": "CA", "position": "Full Professor" }, { "name": "Pourya Shamsolmoali", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Researcher" }, { "name": "Simon Bernard", "gender": "Male", "institution": "Universit\u00e9 de Rouen - Haute Normandie", "country": "FR", "position": "Associate Professor" }, { "name": "Tsiry Mayet", "gender": "Not Specified", "institution": "Institut National des Sciences Appliqu\u00e9es de Rouen", "country": "FR", "position": "PhD student" }, { "name": "Romain D HERAULT", "gender": "Male", "institution": "Universit\u00e9 de Caen Basse Normandie", "country": "FR", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 16, "n_ref_uni": 34, "n_ref": 57, "n_ref_all": 103, "n_fig": 21, "n_tab": 6, "L_tab": 4983, "n_element_tab": 718, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 2292, "n_element_tab_1": 426, "formula_len_all": 972, "formula_len_all_1": 973, "len_all": 133384, "len_all_1": 66258, "len_abs": 1428, "len_title": 123, "len_sents": 34629, "len_sents_1": 26551, "n_sents": 276, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1432, "title": "TD-Paint: Faster Diffusion Inpainting Through Time Aware Pixel Conditioning", "abs": "Diffusion models have emerged as highly effective techniques for inpainting, however, they remain constrained by slow sampling rates. While recent advances have enhanced generation quality, they have also increased sampling time, thereby limiting scalability in real-world applications. We investigate the generative sampling process of diffusion-based inpainting models and observe that these models make minimal use of the input condition during the initial sampling steps. As a result, the sampling trajectory deviates from the data manifold, requiring complex synchronization mechanisms to realign the generation process. To address this, we propose Time-aware Diffusion Paint (TD-Paint), a novel approach that adapts the diffusion process by modeling variable noise levels at the pixel level. This technique allows the model to efficiently use known pixel values from the start, guiding the generation process toward the target manifold. By embedding this information early in the diffusion process, TD-Paint significantly accelerates sampling without compromising image quality. Unlike conventional diffusion-based inpainting models, which require a dedicated architecture or an expensive generation loop, TD-Paint achieves faster sampling times without architectural modifications. Experimental results across three datasets show that TD-Paint outperforms state-of-the-art diffusion models while maintaining lower complexity.", "keywords": [ "Diffusion model", "Inpainting" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 4, 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "eqVu9eaVAB", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Emilia Wi\u015bnios", "gender": "Female", "institution": "Independent Researcher", "country": "PL", "position": "Researcher" }, { "name": "Franciszek Budrowski", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "MS student" }, { "name": "Gracjan G\u00f3ral", "gender": "Male", "institution": "University of Warsaw", "country": "PL", "position": "PhD student" }, { "name": "Micha\u0142 Tyrolski", "gender": "Not Specified", "institution": "University of Warsaw", "country": "", "position": "MS student" }, { "name": "Micha\u0142 Zawalski", "gender": "Male", "institution": "University of Warsaw", "country": "", "position": "PhD student" }, { "name": "Piotr Mi\u0142o\u015b", "gender": "unknown", "institution": "IDEAS NCBR", "country": "PL", "position": "Researcher" }, { "name": "\u0141ukasz Kuci\u0144ski", "gender": "Male", "institution": "IDEAS NCBR", "country": "PL", "position": "Senior Research Scientist" } ], "n_formula": 18, "n_formula_1": 3, "n_ref_uni": 79, "n_ref": 188, "n_ref_all": 241, "n_fig": 35, "n_tab": 3, "L_tab": 1349, "n_element_tab": 173, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1268, "formula_len_all_1": 184, "len_all": 344999, "len_all_1": 60261, "len_abs": 1993, "len_title": 122, "len_sents": 92012, "len_sents_1": 28225, "n_sents": 748, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1046, "title": "What Matters in Hierarchical Search for Combinatorial Reasoning Problems?", "abs": "Combinatorial reasoning problems, particularly the notorious NP-hard tasks, remain a significant challenge for AI research. A common approach to addressing them combines search with learned heuristics. Recent methods in this domain utilize hierarchical planning, executing strategies based on subgoals. Our goal is to advance research in this area and establish a solid conceptual and empirical foundation. Specifically, we identify the following key obstacles, whose presence favors the choice of hierarchical search methods: _hard-to-learn value functions_, _complex action spaces_, _presence of dead ends in the environment_, or _data collected from diverse sources_. Through in-depth empirical analysis, we establish that hierarchical search methods consistently outperform standard search methods across these dimensions, and we formulate insights for future research. On the practical side, we also propose a consistent evaluation methodology to enable meaningful comparisons between methods and to reassess the state-of-the-art algorithms.", "keywords": [ "deep learning", "search", "subgoals", "hierarchical reinforcement learning", "imitation learning" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 2, 2, 4, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "eqQFBnjjPP", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Ale\u0161 Wodecki", "gender": "Male", "institution": "Czech Technical Univeresity in Prague, Czech Technical University of Prague", "country": "CZ", "position": "Researcher" }, { "name": "Jakub Marecek", "gender": "Male", "institution": "Czech Technical University", "country": "CZ", "position": "Tenured faculty" }, { "name": "Pavel Ryt\u00ed\u0159", "gender": "unknown", "institution": "Czech Technical Univeresity in Prague, Czech Technical University of Prague", "country": "CZ", "position": "Researcher" }, { "name": "Georgios Korpas", "gender": "Male", "institution": "IMIS - \"Athena\" Research Center", "country": "", "position": "Principal Researcher" } ], "n_formula": 21, "n_formula_1": 20, "n_ref_uni": 30, "n_ref": 76, "n_ref_all": 85, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 141, "n_element_tab_1": 14, "formula_len_all": 1052, "formula_len_all_1": 1051, "len_all": 101841, "len_all_1": 53341, "len_abs": 1313, "len_title": 50, "len_sents": 29984, "len_sents_1": 24539, "n_sents": 216, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1323, "title": "ExDBN: Exact learning of Dynamic Bayesian Networks", "abs": "Causal learning from data has received a lot of attention in recent years. One way of capturing causal relationships is by utilizing Bayesian networks. There, one recovers a weighted directed acyclic graph in which random variables are represented by vertices, and the weights associated with each edge represent the strengths of the causal relationships between them.\n\nThis concept is extended to capture dynamic effects by introducing a dependency on past data, which may be captured by the structural equation model. This formalism is utilized in the present contribution to propose a score-based learning algorithm. A mixed-integer quadratic program is formulated and an algorithmic solution proposed, in which the pre-generation of exponentially many acyclicity constraints is avoided by utilizing the so-called branch-and-cut (``lazy constraint'') method.\n\nComparing the novel approach to the state-of-the-art, we show that the proposed approach turns out to produce more accurate results when applied to small and medium-sized synthetic instances containing up to 25 time series. Lastly, two interesting applications in bioscience and finance, to which the method is directly applied, further stress the importance of developing highly accurate, globally convergent solvers that can handle instances of modest size.", "keywords": [ "causal lerning", "bayesian network", "structural equation model" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "eqKHuxIpp5", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Alexey Serdyuk", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "PhD student" }, { "name": "Fabian Kre\u00df", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "Researcher" }, { "name": "Iuliia Topko", "gender": "Female", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "PhD student" }, { "name": "Juergen Becker", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Matthias Stammler", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "PhD student" }, { "name": "Tanja Harbaum", "gender": "unknown", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "Postdoc" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 30, "n_ref": 36, "n_ref_all": 44, "n_fig": 3, "n_tab": 2, "L_tab": 439, "n_element_tab": 55, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 439, "n_element_tab_1": 60, "formula_len_all": 281, "formula_len_all_1": 281, "len_all": 93216, "len_all_1": 58508, "len_abs": 1378, "len_title": 113, "len_sents": 28883, "len_sents_1": 28299, "n_sents": 225, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1385, "title": "On-Device Transfer Learning based on Mixed Precision Partitioning", "abs": "The application of machine learning is becoming more widespread, with a growing number of use cases. The development of centralized data training and the exponential growth of data generation raise significant privacy and security concerns. On-device training offers a solution by enhancing privacy and reducing the need for communication between the cloud and the device. Furthermore, on-device transfer learning (TL) can leverage the knowledge gained from pre-trained models, hence, accelerating the training process. However, backpropagation, especially in embedded systems, requires more memory than running inference, which becomes a challenge for devices with limited resources. This paper aims to improve the efficiency and performance of on-device TL. We propose an open source mixed-precision partitioning framework that identifies optimal partitioning layers for retraining, combining quantized and bfloat16 layers to enhance performance and energy efficiency. Our approach is validated through experiments on ResNet-18 and SqueezeNetV1.1 models using Flowers-102, STL-10, and OxfordIIITPet datasets. The partitioned mixed-precision model is able to transfer the knowledge from the pre-trained model to new datasets without losing accuracy compared to the baseline bfloat16 model. These results illustrate the potential for resource-constrained devices to perform TL locally.", "keywords": [ "Machine Learning", "Transfer Learning", "On-device training" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 1, 2, 2, 1 ], "presentation_list": [ 3, 3, 1, 1 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 5, 3, 2, 4 ] }, { "paper_id": "eq8jOD0tgl", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Jianqiang Hu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Mingjie Hu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" } ], "n_formula": 224, "n_formula_1": 33, "n_ref_uni": 20, "n_ref": 68, "n_ref_all": 106, "n_fig": 4, "n_tab": 10, "L_tab": 3921, "n_element_tab": 269, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 30843, "formula_len_all_1": 4861, "len_all": 229704, "len_all_1": 67354, "len_abs": 1912, "len_title": 103, "len_sents": 54923, "len_sents_1": 25449, "n_sents": 648, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1234, "title": "Multi-Task Best Arm Identification with Risk Constraint", "abs": "Best Arm Identification is a very challenging problem in sequential decision-making with many real-world applications. Existing works typically assume that all arms are feasible or/and deal with expectation-based constraints with strong assumptions, loose sample complexity bounds, and non-optimal algorithms. This paper introduces a multi-task best arm identification problem with risk constraint in the fixed-confidence setting, where each arm has multiple performance metrics. The agent aims to optimize one metric while ensuring that the quantiles of other metrics remain below specified thresholds for each task. We first derive a tight, instance-dependent lower bound on sample complexity. Based on this bound, we establish optimality conditions for the static optimal sampling ratio and illustrate how it balances among different tasks and constraints, while addressing the trade-off between optimality and feasibility. We derive a Track-and-Stop strategy with asymptotically optimal sample complexity and a computationally efficient strategy that iteratively solves the optimality conditions. Finally, we extend our results to the linear bandit setting. Numerical experiments show that our algorithm performs relatively well.", "keywords": [ "Multi-Task\uff0cBest Arm Identification\uff0cRisk Constraint\uff0c Fixed Confidence" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 1, 3, 3 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "epbXCD1Ifk", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Afsaneh Doryab", "gender": "Female", "institution": "University of Virginia, Charlottesville", "country": "", "position": "Assistant Professor" }, { "name": "Matthew Landers", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Taylor Wayne Killian", "gender": "Male", "institution": "Apple", "country": "US", "position": "Postdoctoral Research Scientist" }, { "name": "Thomas Hartvigsen", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 27, "n_ref": 40, "n_ref_all": 106, "n_fig": 9, "n_tab": 1, "L_tab": 531, "n_element_tab": 53, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 408, "n_element_tab_1": 5, "formula_len_all": 1356, "formula_len_all_1": 1378, "len_all": 96821, "len_all_1": 61575, "len_abs": 865, "len_title": 112, "len_sents": 31802, "len_sents_1": 28914, "n_sents": 227, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 866, "title": "Offline Reinforcement Learning With Combinatorial Action Spaces", "abs": "Reinforcement learning problems often involve large action spaces arising from the simultaneous execution of multiple sub-actions, resulting in combinatorial action spaces. Learning in combinatorial action spaces is difficult due to the exponential growth in action space size with the number of sub-actions and the dependencies among these sub-actions. In offline settings, this challenge is compounded by limited and suboptimal data. Current methods for offline learning in combinatorial spaces simplify the problem by assuming sub-action independence. We propose Branch Value Estimation (BVE), which effectively captures sub-action dependencies and scales to large combinatorial spaces by learning to evaluate only a small subset of actions at each timestep. Our experiments show that BVE outperforms state-of-the-art methods across a range of action space sizes.", "keywords": [ "reinforcement learning", "offline reinforcement learning", "combinatorial action space" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 5 ] }, { "paper_id": "epJorNF7MK", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Dongdong Chen", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Lu Yuan", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Mengchen Liu", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Yinan Feng", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Yinpeng Chen", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Youzuo Lin", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Associate Professor" }, { "name": "Zicheng Liu", "gender": "Male", "institution": "Advanced Micro Devices", "country": "US", "position": "Sr. Director GenAI" }, { "name": "Xiyang Dai", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 12, "n_ref_uni": 13, "n_ref": 33, "n_ref_all": 97, "n_fig": 19, "n_tab": 25, "L_tab": 8306, "n_element_tab": 778, "n_fig_1": 9, "n_tab_1": 9, "L_tab_1": 1390, "n_element_tab_1": 147, "formula_len_all": 1459, "formula_len_all_1": 1058, "len_all": 238002, "len_all_1": 61915, "len_abs": 981, "len_title": 109, "len_sents": 68878, "len_sents_1": 25263, "n_sents": 528, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 985, "title": "Exploring Invariance in Images through One-way Wave Equations", "abs": "In this paper, we empirically reveals an invariance over images \u2013 images share a set of one-way wave equations with latent speeds. Each image is uniquely associated with a solution to these wave equations, allowing for its reconstruction with high fidelity from an initial condition. We demonstrate it using an intuitive encoder-decoder framework where each image is encoded into its corresponding initial condition (a single vector). Subsequently, the initial condition undergoes a specialized decoder, transforming the one-way wave equations into a first-order norm+linear autoregressive process. This process propagates the initial condition along the x and y directions, generating a high-resolution feature map (up to the image resolution), followed by a few convolutional layers to reconstruct image pixels. The revealed invariance, rooted in the shared wave equations, offers a fresh perspective for comprehending images, establishing a promising avenue for further exploration.", "keywords": [ "Mathematical invariance in images", "One-way wave equation", "Auto-regression" ], "rating_list": [ 3, 6, 8, 5 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "ep4FPuE1w3", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Andrei Bursuc", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Researcher" }, { "name": "Mohammad Fahes", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "PhD student" }, { "name": "Tuan-Hung Vu", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Researcher" }, { "name": "Patrick Perez", "gender": "Male", "institution": "Kyutai", "country": "FR", "position": "Researcher" }, { "name": "Raoul de Charette", "gender": "Male", "institution": "Inria", "country": "FR", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 11, "n_ref": 30, "n_ref_all": 52, "n_fig": 4, "n_tab": 13, "L_tab": 5422, "n_element_tab": 448, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 3516, "n_element_tab_1": 226, "formula_len_all": 509, "formula_len_all_1": 351, "len_all": 156086, "len_all_1": 70617, "len_abs": 1014, "len_title": 111, "len_sents": 37008, "len_sents_1": 27702, "n_sents": 342, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1017, "title": "Fine-tuning CLIP\u2019s Last Visual Projector: A Few-Shot Cornucopia", "abs": "We consider the problem of adapting a contrastively pretrained vision-language model like CLIP (Radford et al., 2021) for few-shot classification. The existing literature addresses this problem by learning a linear classifier of the frozen visual features, optimizing word embeddings, or learning external feature adapters. This paper introduces an alternative way for CLIP adaptation without adding \u201cexternal\u201d parameters to optimize. We find that simply fine-tuning the last projection matrix of the vision encoder leads to strong performance compared to the existing baselines. Furthermore, we show that regularizing training with the distance between the fine-tuned and pretrained matrices adds reliability for adapting CLIP through this layer. Perhaps surprisingly, this approach, coined ProLIP, yields performances on par or better than state of the art on 11 few-shot classification benchmarks, few-shot domain generalization, cross-dataset transfer and test-time adaptation. Code will be made available online.", "keywords": [ "Few-shot classification", "CLIP", "transfer learning" ], "rating_list": [ 3, 8, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "eoln5WgrPx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Chenxin An", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jingjing Xu", "gender": "Female", "institution": "Shanghai AI Lab", "country": "", "position": "PhD student" }, { "name": "Jun Zhang", "gender": "Male", "institution": "ByteDance", "country": "CN", "position": "Researcher" }, { "name": "Lingpeng Kong", "gender": "Male", "institution": "Department of Computer Science, The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Ming Zhong", "gender": "Male", "institution": "University of Illinois Urbana Champaign", "country": "US", "position": "PhD student" }, { "name": "Shansan Gong", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yao Luo", "gender": "Female", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Tobias Lee", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 8, "n_ref_uni": 47, "n_ref": 119, "n_ref_all": 158, "n_fig": 17, "n_tab": 6, "L_tab": 3696, "n_element_tab": 604, "n_fig_1": 14, "n_tab_1": 4, "L_tab_1": 2861, "n_element_tab_1": 350, "formula_len_all": 689, "formula_len_all_1": 453, "len_all": 204272, "len_all_1": 69478, "len_abs": 1344, "len_title": 105, "len_sents": 45873, "len_sents_1": 30659, "n_sents": 362, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1268, "title": "Why Does the Effective Context Length of LLMs Fall Short?", "abs": "Advancements in distributed training and efficient attention mechanisms have significantly expanded the context window sizes of large language models (LLMs). However, recent work reveals that the effective context lengths of open-source LLMs often fall short, typically not exceeding half of their training lengths. In this work, we attribute this limitation to the left-skewed frequency distribution of relative positions formed in LLMs pretraining and post-training stages, which impedes their ability to effectively gather distant information.\nTo address this challenge, we introduce Shifted Rotray Position Embedding (STRING). STRING shifts well-trained positions to overwrite the original ineffective positions during inference, enhancing performance within their existing training lengths. \nExperimental results show that without additional training, STRING dramatically improves the performance of the latest large-scale models, such as Llama3.1 70B and Qwen2 72B, by over 10 points on popular long-context benchmarks RULER and InfiniteBench, establishing new state-of-the-art results for open-source LLMs. Compared to commercial models, Llama 3.1 70B with STRING even achieves better performance than GPT-4-128K and clearly surpasses Claude 2 and Kimi-chat.", "keywords": [ "Large Language Model", "Long-Context Modeling", "Position Embedding" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 4, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "enQSCx47Ud", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Qiang Ni", "gender": "Male", "institution": "Lancaster University", "country": "GB", "position": "Full Professor" }, { "name": "Yongchen Zhou", "gender": "Male", "institution": "Lancaster University", "country": "GB", "position": "PhD student" }, { "name": "Richard M. Jiang", "gender": "Male", "institution": "Lancaster University", "country": "GB", "position": "Associate Professor" } ], "n_formula": 23, "n_formula_1": 23, "n_ref_uni": 45, "n_ref": 57, "n_ref_all": 63, "n_fig": 3, "n_tab": 2, "L_tab": 356, "n_element_tab": 41, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 356, "n_element_tab_1": 41, "formula_len_all": 1062, "formula_len_all_1": 1062, "len_all": 93794, "len_all_1": 65683, "len_abs": 832, "len_title": 147, "len_sents": 24494, "len_sents_1": 24106, "n_sents": 165, "n_sents_1": 163, "venue": "ICLR2025", "decision": "withdraw", "L_title": 98, "L_abs": 833, "title": "Addressing Inverse Problems in Frame Restoration with Siamese Conditional Variational Autoencoders", "abs": "Restoring missing information in video frames is a challenging inverse problem, particularly in applications such as autonomous driving and surveillance. This paper introduces the Siamese Masked Conditional Variational Autoencoder (SMCVAE), a novel model that utilizes a Siamese network architecture with Siamese Vision Transformer (SiamViT) encoders. By leveraging the inherent similarities between paired frames, SMCVAE enhances the model's ability to accurately reconstruct missing content. This approach effectively tackles the problem of missing patches\u2014often resulting from camera malfunctions\u2014through advanced variational inference techniques. Experimental results demonstrate SMCVAE's superior performance in restoring lost information, highlighting its potential to solve complex inverse problems in real-world environments.", "keywords": [ "autoencoder", "inverse problem", "Siamese networks" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 1, 1, 3, 2, 3 ], "presentation_list": [ 1, 2, 3, 2, 2 ], "contribution_list": [ 1, 2, 1, 2, 1 ], "confidence_list": [ 5, 5, 3, 5, 4 ] }, { "paper_id": "en3NwykrHW", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Jason D. Lee", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Ruosong Wang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Assistant Professor" }, { "name": "Simon Shaolei Du", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "Yuxin Chen", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Associate Professor" }, { "name": "Zihan Zhang", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Postdoc" } ], "n_formula": 106, "n_formula_1": 36, "n_ref_uni": 34, "n_ref": 58, "n_ref_all": 62, "n_fig": 0, "n_tab": 6, "L_tab": 2827, "n_element_tab": 99, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 329, "n_element_tab_1": 13, "formula_len_all": 11298, "formula_len_all_1": 2918, "len_all": 144801, "len_all_1": 63696, "len_abs": 854, "len_title": 130, "len_sents": 35173, "len_sents_1": 25307, "n_sents": 382, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 889, "title": "Minimax Optimal Regret Bound for Reinforcement Learning with Trajectory Feedback", "abs": "We study the reinforcement learning (RL) problem with trajectory feedback. The trajectory feedback based reinforcement learning problem, where the learner can only observe the accumulative noised reward along the trajectory, is particularly suitable for the practical scenarios where the agent suffers extensively from querying the reward in each single step. For a finite-horizon Markov Decision Process (MDP) with $S$ states, $A$ actions and a horizon length of $H$, we develop an algorithm that enjoys an optimal regret of $\\tilde{O}\\left(\\sqrt{SAH^3K}\\right)$ in $K$ episodes for sufficiently large $K$. To achieve this, our technical contributions are two-fold: (1) we incorporate reinforcement learning with linear bandits problem to construct a tighter confidence region for the reward function; (2) we construct a reference transition model to better guide the exploration process.", "keywords": [ "Reinforcement learning theory", "regret analysis", "trajectory feedback" ], "rating_list": [ 6, 3, 5, 5, 8 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 3, 2 ] }, { "paper_id": "emns7tgDOq", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Minh Hoang Nguyen", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "PhD student" }, { "name": "Thommen Karimpanal George", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Assistant Professor" }, { "name": "Hung Thai Le", "gender": "Male", "institution": "Deakin University", "country": "", "position": "Lecturer" }, { "name": "Linh Pham Van Le", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "PhD student" }, { "name": "Sunil Gupta", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Full Professor" } ], "n_formula": 43, "n_formula_1": 23, "n_ref_uni": 61, "n_ref": 146, "n_ref_all": 169, "n_fig": 5, "n_tab": 15, "L_tab": 3538, "n_element_tab": 419, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1107, "n_element_tab_1": 171, "formula_len_all": 2003, "formula_len_all_1": 1082, "len_all": 207554, "len_all_1": 71951, "len_abs": 1209, "len_title": 133, "len_sents": 61153, "len_sents_1": 31129, "n_sents": 558, "n_sents_1": 273, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1290, "title": "Beyond the Known: Decision Making with Counterfactual Reasoning Decision Transformer", "abs": "Decision Transformer (DT) plays a crucial role in modern reinforcement learning, leveraging offline datasets to achieve impressive results across various domains. However, DT requires high-quality, comprehensive data to perform optimally. In real-world applications, such ideal data is often lacking, with the underrepresentation of optimal behaviours posing a significant challenge. This limitation highlights the difficulty of relying on offline datasets for training, as suboptimal data can hinder performance. To address this, we propose the Counterfactual Reasoning Decision Transformer (CRDT), a novel framework inspired by counterfactual reasoning. CRDT enhances DT\u2019s ability to reason beyond known data by generating and utilizing counterfactual experiences, enabling improved decision-making in out-of-distribution scenarios. Extensive experiments across continuous and discrete action spaces, including environments with limited data, demonstrate that CRDT consistently outperforms conventional DT approaches. Additionally, reasoning counterfactually allows the DT agent to obtain stitching ability, allowing it to combine suboptimal trajectories. These results highlight the potential of counterfactual reasoning to enhance RL agents' performance and generalization capabilities.", "keywords": [ "reinforcement learning", "decision transformer", "causality" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "emSgz2bKVq", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Carola-Bibiane Sch\u00f6nlieb", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "Full Professor" }, { "name": "Georgios Batzolis", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Willem Diepeveen", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Postdoc" }, { "name": "Zakhar Shumaylov", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 93, "n_formula_1": 38, "n_ref_uni": 32, "n_ref": 51, "n_ref_all": 90, "n_fig": 13, "n_tab": 1, "L_tab": 626, "n_element_tab": 26, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9209, "formula_len_all_1": 3168, "len_all": 156979, "len_all_1": 58690, "len_abs": 1238, "len_title": 88, "len_sents": 48317, "len_sents_1": 24297, "n_sents": 411, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1242, "title": "Score-based pullback Riemannian geometry", "abs": "Data-driven Riemannian geometry has emerged as a powerful tool for interpretable representation learning, offering improved efficiency in downstream tasks. Moving forward, it is crucial to balance cheap manifold mappings with efficient training algorithms. In this work, we integrate concepts from pullback Riemannian geometry and generative models to propose a framework for data-driven Riemannian geometry that is scalable in both geometry and learning: score-based pullback Riemannian geometry. Focusing on unimodal distributions as a first step, we propose a score-based Riemannian structure with closed-form geodesics that pass through the data probability density. With this structure, we construct a Riemannian autoencoder (RAE) with error bounds for discovering the correct data manifold dimension. This framework can naturally be used with anisotropic normalizing flows by adopting isometry regularization during training. Through numerical experiments on various datasets, we demonstrate that our framework not only produces high-quality geodesics through the data support, but also reliably estimates the intrinsic dimension of the data manifold and provides a global chart of the manifold, even in high-dimensional ambient spaces.", "keywords": [ "Data-driven Riemannian geometry", "interpretable representation learning", "pullback Riemannian geometry", "generative models", "closed-form geodesics", "manifold dimension estimation" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "emMMa4q0qw", "primary_area": "applications to neuroscience & cognitive science", "n_author": 6, "author_info_list": [ { "name": "James J. DiCarlo", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Jeremy Schwartz", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Researcher" }, { "name": "Joshua B. Tenenbaum", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Professor" }, { "name": "Weichen Huang", "gender": "Male", "institution": "St Andrew's College ", "country": "IE", "position": "High school student" }, { "name": "Yudi Xie", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Seth Alter", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 1, "n_ref_uni": 40, "n_ref": 72, "n_ref_all": 100, "n_fig": 14, "n_tab": 5, "L_tab": 1411, "n_element_tab": 48, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 432, "n_element_tab_1": 9, "formula_len_all": 1023, "formula_len_all_1": 29, "len_all": 161432, "len_all_1": 64113, "len_abs": 1900, "len_title": 0, "len_sents": 54247, "len_sents_1": 32686, "n_sents": 385, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 1920, "title": "Vision CNNs trained to estimate spatial latents learned similar ventral-stream-aligned representations", "abs": "Studies of the functional role of the primate ventral visual stream have traditionally focused on object categorization, often ignoring -- despite much prior evidence -- its role in estimating \"spatial\" latents such as object position and pose. Most leading ventral stream models are derived by optimizing networks for object categorization, which seems to imply that the ventral stream is also derived under such an objective. Here, we explore an alternative hypothesis: Might the ventral stream be optimized for estimating spatial latents? And a closely related question: How different -- if at all -- are representations learned from spatial latent estimation compared to categorization? To ask these questions, we leveraged synthetic image datasets generated by a 3D graphic engine and trained convolutional neural networks (CNNs) to estimate different combinations of spatial and category latents. We found that models trained to estimate just a few spatial latents achieve neural alignment scores comparable to those trained on hundreds of categories, and the spatial latent performance of models strongly correlates with their neural alignment. Spatial latent and category-trained models have very similar -- but not identical -- internal representations, especially in their early and middle layers. We provide evidence that this convergence is partly driven by non-target latent variability in the training data, which facilitates the implicit learning of representations of those non-target latents. Taken together, these results suggest that many training objectives, such as spatial latents, can lead to similar models aligned neurally with the ventral stream. Thus, one should not assume that the ventral stream is optimized for object categorization only. As a field, we need to continue to sharpen our measures of comparing models to brains to better understand the functional roles of the ventral stream.", "keywords": [ "Vision", "Convolutional Neural Networks", "Representation Learning", "Neural Data Alignment", "Ventral Visual Stream", "Computational Neuroscience" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 4, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "em0gAL8fbK", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Guanhong Tao", "gender": "unknown", "institution": "University of Utah", "country": "US", "position": "Assistant Professor" }, { "name": "Shengwei An", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Shiwei Feng", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Wenbo Guo", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "Assistant Professor" }, { "name": "Xiangyu Zhang", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Full Professor" }, { "name": "Xuan Chen", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Yunshu Mao", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Zikang Xiong", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 55, "n_ref": 82, "n_ref_all": 105, "n_fig": 4, "n_tab": 9, "L_tab": 3671, "n_element_tab": 559, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 557, "n_element_tab_1": 112, "formula_len_all": 419, "formula_len_all_1": 372, "len_all": 176874, "len_all_1": 73405, "len_abs": 1790, "len_title": 159, "len_sents": 53514, "len_sents_1": 36606, "n_sents": 403, "n_sents_1": 262, "venue": "ICLR2025", "decision": "withdraw", "L_title": 110, "L_abs": 1803, "title": "Temporal Logic-Based Multi-Vehicle Backdoor Attacks against Offline RL Agents in End-to-end Autonomous Driving", "abs": "End-to-end autonomous driving (AD) systems integrate complex decision-making processes. \nAssessing the safety of these systems against potential security threats, including backdoor attacks, is a stepping stone for real-world deployment. \nHowever, traditional methods focus on static triggers, which do not adequately reflect the dynamic nature of these systems and could be impractical to deploy in the real world.\nTo address these limitations, we propose a novel backdoor attack against the end-to-end AD systems that leverage multi-vehicles' trajectories as triggers.\nWe employ different behavior models and their configurations to generate the trigger trajectories, which are then quantitatively evaluated using temporal logic specifications. \nThis evaluation guides the subsequent perturbations to the behavior model configurations.\nThrough an iterative process of regeneration and re-evaluation, we can refine and generate realistic and plausible trigger trajectories that involve multiple vehicles' complex interactions.\nFurthermore, we develop a negative training strategy by incorporating patch trajectories that share similarities with the triggers but are designated not to activate the backdoor. \nWe thus enhance the stealthiness of the attack, refining the system\u2019s responses to trigger scenarios. \nThrough extensive empirical studies using offline reinforcement learning (RL) driving agents with various trigger patterns and target action designs, we demonstrate the flexibility and effectiveness of our proposed attack, showing the under-exploration of existing end-to-end AD systems' vulnerabilities to such multi-vehicle-based backdoor attacks.\nWe also evaluate the attack against existing defenses and validate different design choices of our attack via a comprehensive ablation study.", "keywords": [ "backdoor attack; autonomous driving safety" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "eluQFbNeuH", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Cornelius Miething", "gender": "Male", "institution": "Medical Center - University of Freiburg", "country": "DE", "position": "Principal Researcher" }, { "name": "Gabriel Kalweit", "gender": "unknown", "institution": "Collaborative Research Institute Intelligent Oncology", "country": "DE", "position": "Researcher" }, { "name": "Justus Duyster", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Maria Kalweit", "gender": "Female", "institution": "Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Mehdi Naouar", "gender": "Male", "institution": "CRIION - Collaborative Research Group", "country": "DE", "position": "Researcher" }, { "name": "Yannick Vogt", "gender": "Male", "institution": "Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Joschka B\u00f6decker", "gender": "Male", "institution": "Universit\u00e4t Freiburg", "country": "", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 32, "n_ref": 101, "n_ref_all": 116, "n_fig": 4, "n_tab": 2, "L_tab": 1751, "n_element_tab": 91, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 637, "formula_len_all_1": 637, "len_all": 131564, "len_all_1": 69253, "len_abs": 1670, "len_title": 131, "len_sents": 38662, "len_sents_1": 33758, "n_sents": 273, "n_sents_1": 231, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1678, "title": "BetterBodies: Reinforcement Learning guided Diffusion for Antibody Sequence Design", "abs": "Antibodies offer great potential for the treatment of various diseases. However, the discovery of therapeutic antibodies through traditional wet lab methods is expensive and time-consuming. The use of generative models in designing antibodies therefore holds great promise, as it can reduce the time and resources required. Recently, the class of diffusion models has gained considerable traction for their ability to synthesize diverse and high-quality samples. In their basic form, however, they lack mechanisms to optimize for specific properties, such as binding affinity to an antigen. In contrast, the class of offline Reinforcement Learning (RL) methods has demonstrated strong performance in navigating large search spaces, including scenarios where frequent real-world interaction, such as interaction with a wet lab, is impractical. Our novel method, BetterBodies, which combines Variational Autoencoders (VAEs) with offline RL guided latent diffusion, can generate novel sets of antibody CDRH3 sequences from different data distributions. Furthermore, we reflect biophysical properties in the VAE latent space using a contrastive loss and add a novel Q-function based filtering to enhance the affinity of generated sequences. Using the Absolut! simulator, we demonstrate that BetterBodies generates sequences with improved binding affinity to the SARS-CoV spike receptor-binding domain and matches or outperforms the state-of-the-art method Generative Flow Network (GFlowNet). In conclusion, our method has the potential for great implications in real-world biological sequence design, where the generation of novel high-affinity binders is a cost-intensive endeavor.", "keywords": [ "Reinforcement Learning", "Generative Models", "Diffusion Models", "Antibody Design" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 1, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "elmTU101oS", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Shiyue Wang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "PhD student" }, { "name": "Yang Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Ziao Guo", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 10, "n_ref_uni": 45, "n_ref": 71, "n_ref_all": 92, "n_fig": 6, "n_tab": 16, "L_tab": 10553, "n_element_tab": 1173, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1528, "n_element_tab_1": 159, "formula_len_all": 1932, "formula_len_all_1": 774, "len_all": 198708, "len_all_1": 74723, "len_abs": 1507, "len_title": 129, "len_sents": 59652, "len_sents_1": 34251, "n_sents": 480, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1511, "title": "Learning General Representations Across Graph Combinatorial Optimization Problems", "abs": "Combinatorial optimization (CO) problems are classical and crucial in many fields, with many NP-complete (NPC) examples being reducible to one another, revealing an underlying connection between them. Existing methods, however, primarily focus on task-specific models trained on individual datasets, limiting the quality of learned representations and the transferability to other CO problems. Given the reducibility among these problems, a natural idea is to abstract a higher-level representation that captures the essence shared across different problems, enabling knowledge transfer and mutual enhancement. In this paper, we propose a novel paradigm CORAL that treats each CO problem type as a distinct modality and unifies them by transforming all instances into representations of the fundamental Boolean satisfiability (SAT) problem. Our approach aims to capture the underlying commonalities across multiple problem types via cross-modal contrastive learning with supervision, thereby enhancing representation learning. Extensive experiments on seven graph decision problems (GDPs) demonstrate the effectiveness of CORAL, showing that our approach significantly improves the quality and generalizability of the learned representations. Furthermore, we showcase the utility of the pre-trained unified SAT representations on related tasks, including satisfying assignment prediction and unsat core variable prediction, highlighting the potential of CORAL as a unified pre-training paradigm for CO problems.", "keywords": [ "Combinatorial Optimization", "Contrastive Learning", "Representation Learning" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 1 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "elTJBP7Fbv", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Ning Gao", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Xingyu Jiang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Xiuhui Zhang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Yue Deng", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "hongkun dou", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 68, "n_ref_all": 104, "n_fig": 12, "n_tab": 4, "L_tab": 975, "n_element_tab": 72, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 121, "n_element_tab_1": 9, "formula_len_all": 730, "formula_len_all_1": 733, "len_all": 149288, "len_all_1": 56037, "len_abs": 1670, "len_title": 140, "len_sents": 48713, "len_sents_1": 28293, "n_sents": 329, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1694, "title": "Value-aligned Behavior Cloning for Offline Reinforcement Learning via Bi-level Optimization", "abs": "Offline reinforcement learning (RL) aims to optimize policies under pre-collected data, without requiring any further interactions with the environment. Derived from imitation learning, Behavior cloning (BC) is extensively utilized in offline RL for its simplicity and effectiveness. Although BC inherently avoids out-of-distribution deviations, it lacks the ability to discern between high and low-quality data, potentially leading to sub-optimal performance when facing with poor-quality data. Current offline RL algorithms attempt to enhance BC by incorporating value estimation, yet often struggle to effectively balance these two critical components, specifically the alignment between the behavior policy and the pre-trained value estimations under in-sample offline data. To address this challenge, we propose the Value-aligned Behavior Cloning via Bi-level Optimization (VACO), a novel bi-level framework that seamlessly integrates an inner loop for weighted supervised behavior cloning (BC) with an outer loop dedicated to value alignment. \\textcolor{blue}{In this framework, the inner loop employs a meta-scoring network to evaluate and appropriately weight each training sample, while the outer loop maximizes value estimation for alignment with controlled noise to facilitate limited exploration.} This bi-level structure allows VACO to identify the optimal weighted BC policy, ultimately maximizing the expected estimated return conditioned on the learned value function. We conduct a comprehensive evaluation of VACO across a variety of continuous control benchmarks in offline RL, where it consistently achieves superior performance compared to existing state-of-the-art methods.", "keywords": [ "offline reinforcement learning;bi-level optimization;value alignment" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 4, 4, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "el2pNeLrRC", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 10, "author_info_list": [ { "name": "Anh Nguyen", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Undergrad student" }, { "name": "Antonios Valkanas", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "Bilel Fehri", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Bishal Thapaliya", "gender": "Male", "institution": "Georgia State University", "country": "", "position": "PhD student" }, { "name": "Deepayan Chakrabarti", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Fudong Lin", "gender": "Male", "institution": "University of Delaware", "country": "US", "position": "PhD student" }, { "name": "Igor Grudetskyi", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Jingyu Liu", "gender": "unknown", "institution": "University of New Mexico", "country": "", "position": "PhD student" }, { "name": "Tian Xie", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yao Lu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 30, "n_formula_1": 10, "n_ref_uni": 36, "n_ref": 65, "n_ref_all": 84, "n_fig": 11, "n_tab": 9, "L_tab": 4514, "n_element_tab": 361, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 2734, "n_element_tab_1": 238, "formula_len_all": 1312, "formula_len_all_1": 412, "len_all": 166358, "len_all_1": 71103, "len_abs": 1163, "len_title": 134, "len_sents": 53034, "len_sents_1": 29780, "n_sents": 452, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1206, "title": "ECGN: A CLUSTER-AWARE APPROACH TO GRAPH NEURAL NETWORKS FOR IMBALANCED CLASSIFICATION.", "abs": "Classifying nodes in a graph is a common problem. The ideal classifier must\nadapt to any imbalances in the class distribution. It must also use information in\nthe clustering structure of real-world graphs. Existing Graph Neural Networks\n(GNNs) have not addressed both problems together. We propose the Enhanced\nCluster-aware Graph Network (ECGN), a novel method that addresses these is-\nsues by integrating cluster-specific training with synthetic node generation. Unlike\ntraditional GNNs that apply the same node update process for all nodes, ECGN\nlearns different aggregations for different clusters. We also use the clusters to gen-\nerate new minority-class nodes in a way that helps clarify the inter-class decision\nboundary. By combining cluster-aware embeddings with a global integration step,\nECGN enhances the quality of the resulting node embeddings. Our method works\nwith any underlying GNN and any cluster generation technique. Experimental\nresults show that ECGN consistently outperforms its closest competitors by up to\n11% on some widely-studied benchmark datasets. The GitHub implementation\nfor implementation and replication is publicly available on https://github.com/anonymous753341/ECGN.", "keywords": [ "Graph Neural Networks", "Uniform Node Updates Issue", "Imbalanced Node Classification", "Cluster Specific Updates", "Synthetic Node Generation" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "eks3dGnocX", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Cyrus Rashtchian", "gender": "Male", "institution": "Google Research", "country": "US", "position": "Researcher" }, { "name": "Enming Luo", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Guan Zhe Hong", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Nishanth Dikkala", "gender": "Male", "institution": "Google", "country": "", "position": "Google Research" }, { "name": "Rina Panigrahy", "gender": "unknown", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Xin Wang", "gender": "Male", "institution": "Google", "country": "", "position": "Software Engineer" } ], "n_formula": 18, "n_formula_1": 3, "n_ref_uni": 25, "n_ref": 63, "n_ref_all": 117, "n_fig": 22, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1542, "formula_len_all_1": 274, "len_all": 221472, "len_all_1": 67344, "len_abs": 1108, "len_title": 123, "len_sents": 84874, "len_sents_1": 33015, "n_sents": 615, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1115, "title": "How Transformers Solve Propositional Logic Problems: A Mechanistic Analysis", "abs": "Large language models (LLMs) have shown amazing performance on tasks that require planning and reasoning. Motivated by this, we investigate the internal mechanisms that underpin a network's ability to perform complex logical reasoning. We first construct a synthetic propositional logic problem that serves as a concrete test-bed for network training and evaluation. Crucially, this problem demands nontrivial planning to solve. We perform our study on two fronts. First, we pursue an understanding of precisely how a three-layer transformer, trained from scratch and attains perfect test accuracy, solves this problem. We are able to identify certain \"planning\" and \"reasoning\" circuits in the network that necessitate cooperation between the attention blocks to implement the desired logic. Second, we study how a pretrained LLM, Mistral 7B, solves this problem. Using activation patching, we characterize internal components that are critical in solving our logic problem. Overall, our work systemically uncovers novel aspects of small and large transformers, and continues the study of how they plan and reason.", "keywords": [ "Mechanistic Interpretability", "Language Models", "Transformers", "Logical Reasoning", "Learned Representations" ], "rating_list": [ 3, 3, 6, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 1, 3, 4 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "ekADgawLgI", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Difei Cheng", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "ShaoDong Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xin shi", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "ruinan Jin", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Postdoc" } ], "n_formula": 138, "n_formula_1": 33, "n_ref_uni": 19, "n_ref": 38, "n_ref_all": 41, "n_fig": 0, "n_tab": 3, "L_tab": 27, "n_element_tab": 1, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 79, "n_element_tab_1": 20, "formula_len_all": 16413, "formula_len_all_1": 2249, "len_all": 137807, "len_all_1": 62640, "len_abs": 1577, "len_title": 159, "len_sents": 41798, "len_sents_1": 26766, "n_sents": 421, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1649, "title": "Asymptotic Convergence of SGD in Non-Convex Problems: A Stopping Times Method with Relaxed Step-size Conditions", "abs": "Stochastic Gradient Descent (SGD) is widely used in machine learning research. In previous research, the convergence analyses of SGD under vanishing step-size settings typically assumed that the step sizes satisfied the Robbins-Monro conditions, which is to say, the sum of the step sizes was infinite, while the sum of the squares of the step sizes was finite. In practical applications, a wider variety of step sizes is often used, but these may not meet the Robbins-Monro step-size conditions, thus lacking theoretical guarantees of convergence. To bridge the gap between theory and practical application, this paper introduces a novel analytical method\u2014the stopping time method based on probability theory\u2014to explore the asymptotic convergence of SGD under more relaxed step-size conditions. In the non-convex setting, we prove that the almost sure convergence of the sequence of iterates generated by SGD when step sizes satisfy \\(\\sum_{t=1}^{+\\infty} \\epsilon_t = +\\infty\\) and \\(\\sum_{t=1}^{+\\infty} \\epsilon_t^p < +\\infty\\) for some \\(p > 2\\). Compared to previous works, our analysis eliminates the need to assume global Lipschitz continuity of the loss function, and it also relaxes the requirement of global boundedness of the high-order moments of the stochastic gradient to local boundedness. Additionally, we prove \\(L_2\\) convergence without the need for assuming global boundedness of loss functions or their gradients. The assumptions required for this work are the weakest among studies with the same conclusions, thereby extending the applicability of SGD in various practical scenarios where traditional assumptions may not hold.", "keywords": [ "stochastic optimization", "convergence analyse" ], "rating_list": [ 3, 3, 6, 3, 6 ], "soundness_list": [ 2, 2, 3, 2, 4 ], "presentation_list": [ 1, 1, 3, 2, 3 ], "contribution_list": [ 1, 2, 3, 2, 3 ], "confidence_list": [ 2, 3, 3, 4, 2 ] }, { "paper_id": "ejvf3JrZuC", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Mario Fritz", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Full Professor" }, { "name": "Pramod Kaushik", "gender": "unknown", "institution": "International Institute of Information Technology Hyderabad", "country": "", "position": "MS student" }, { "name": "Sahar Abdelnabi", "gender": "Male", "institution": "Microsoft", "country": "GB", "position": "Researcher" }, { "name": "Sarath Sivaprasad", "gender": "Male", "institution": "cispa", "country": "DE", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 26, "n_ref": 56, "n_ref_all": 74, "n_fig": 5, "n_tab": 10, "L_tab": 5815, "n_element_tab": 532, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2241, "n_element_tab_1": 233, "formula_len_all": 279, "formula_len_all_1": 133, "len_all": 199011, "len_all_1": 70535, "len_abs": 1166, "len_title": 110, "len_sents": 61265, "len_sents_1": 32210, "n_sents": 451, "n_sents_1": 250, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1124, "title": "Theory of LLM sampling: part descriptive and part prescriptive", "abs": "Large Language Models (LLMs) are increasingly utilized in autonomous decision-making systems, where they sample options from an action space. However, the underlying heuristics guiding the sampling of LLMs remain under-explored. We examine LLM response sampling and propose a theory that the sample of an LLM is driven by a descriptive component (the notion of statistical average) and a prescriptive component (notion of an ideal represented in the LLM). In a controlled experimental setting, we demonstrate that LLM outputs deviate from statistically probable outcome in the direction of a presciptive component. We further show this deviation towards prescriptive component consistently appears across diverse real-world domains, including social, public health, and scientific contexts. Using this theory, we show that concept prototypes in LLMs are affected by prescriptive norms, similar to concept of normality in humans. Through case studies, we illustrate that in real-world applications, the shift toward an ideal value in LLM outputs can result in significantly biased decision-making, raising ethical concerns.", "keywords": [ "decision making", "sampling", "prescriptive norm", "heuristics" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "ejgY0DyaQD", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Mijung Kim", "gender": "unknown", "institution": "Ulsan National Institute of Science and Technology", "country": "", "position": "Assistant Professor" }, { "name": "Taehwan Kim", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "SEOKHOON JEONG", "gender": "unknown", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 43, "n_ref": 65, "n_ref_all": 81, "n_fig": 5, "n_tab": 1, "L_tab": 4437, "n_element_tab": 77, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 345, "n_element_tab_1": 4, "formula_len_all": 473, "formula_len_all_1": 224, "len_all": 145475, "len_all_1": 56128, "len_abs": 2147, "len_title": 107, "len_sents": 34267, "len_sents_1": 26042, "n_sents": 262, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1578, "title": "Designing Deep Learning Programs with Large Language Models", "abs": "The process of utilizing deep neural architectures to solve tasks differs significantly from conventional programming due to its complexity and the need for specialized knowledge. While code generation technologies have made substantial progress, their application in deep learning programs requires a distinct approach. Although previous research has shown that large language model agents perform well in areas such as data science, neural architecture search, and hyperparameter tuning, the task of proposing and refining deep neural architectures at a high level remains largely unexplored. Current methods for automating the synthesis of deep learning programs often rely on basic code templates or API calls, which restrict the solution space to predefined architectures. In this paper, we aim to bridge the gap between traditional code generation and deep learning program synthesis by introducing the task of Deep Learning Program Design (DLPD), a task of designing an effective deep learning program for the task, along with appropriate architectures and techniques. We propose Deep Ones, a comprehensive solution for DLPD. Our solution includes a large-scale dataset and a lightweight benchmark specifically designed for DLPD. On our benchmark, Llama-3.1 8B, fine-tuned on our dataset, demonstrates better architecture suggestion capability than GPT-4o and better performance than Claude-3.5-Sonnet, showcasing that Deep Ones effectively addresses the challenge of DLPD. Deep Ones will be publicly available, including the dataset, benchmark, codes, and model weights.", "keywords": [ "Program Synthesis", "Large Language Model Agents", "Dataset", "Benchmark" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "ejVuTFFkl6", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Qingyao Ai", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Shuqi Zhu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Yiqun LIU", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Ziyi Ye", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 46, "n_ref_all": 67, "n_fig": 9, "n_tab": 5, "L_tab": 2718, "n_element_tab": 405, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1937, "n_element_tab_1": 227, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 147208, "len_all_1": 62110, "len_abs": 1654, "len_title": 162, "len_sents": 44043, "len_sents_1": 28401, "n_sents": 331, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 114, "L_abs": 1661, "title": "EEG-ImageNet: An Electroencephalogram Dataset and Benchmarks with Image Visual Stimuli of Multi-Granularity Labels", "abs": "Exploring how brain activity translates into visual perception offers valuable insights into the biological visual system's representation of the world. Recent advancements have enabled effective image classification and high-quality reconstruction using brain signals obtained through Functional Magnetic Resonance Imaging (fMRI) or magnetoencephalography (MEG). However, the cost and bulkiness of these technologies hinder their practical application. In contrast, Electroencephalography (EEG) presents advantages such as ease of use, affordability, high temporal resolution, and non-invasive operation, yet it remains underutilized in related research due to a shortage of comprehensive datasets. To fill this gap, we introduce EEG-ImageNet, a novel EEG dataset featuring recordings from 16 participants exposed to 4000 images sourced from the ImageNet dataset. This dataset offers five times the number of EEG-image pairs compared to existing benchmarks. EEG-ImageNet includes image stimuli labeled with varying levels of granularity, comprising 40 images with coarse labels and 40 with fine labels. We establish benchmarks for both object classification and image reconstruction based on this dataset. Experiments with several commonly used models show that the best-performing models can achieve object classification with an accuracy around 60% and image reconstruction with two-way identification around 64%. These findings highlight the dataset's potential to enhance EEG-based visual brain-computer interfaces, deepen our understanding of visual perception in biological systems, and suggest promising applications for improving machine vision models.", "keywords": [ "EEG", "visual stimuli", "computer vision", "multi-modality", "object classification", "image generation" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "ejGAytoWoe", "primary_area": "generative models", "n_author": 13, "author_info_list": [ { "name": "Enshen Zhou", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "MS student" }, { "name": "Jing Shao", "gender": "Female", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Jiwen Yu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "LEI BAI", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Lijun Li", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Lu Sheng", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Associate Professor" }, { "name": "Ruimao Zhang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Wanli Ouyang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Full Professor" }, { "name": "Xihui Liu", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Xijun Wang", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Undergrad student" }, { "name": "Yiran Qin", "gender": "Male", "institution": "The Chinese University of Hong Kong(Shenzhen)", "country": "CN", "position": "PhD student" }, { "name": "Zhelun Shi", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Zhenfei Yin", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 100, "n_ref_all": 113, "n_fig": 7, "n_tab": 12, "L_tab": 3491, "n_element_tab": 395, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1600, "n_element_tab_1": 77, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 225457, "len_all_1": 64643, "len_abs": 1703, "len_title": 114, "len_sents": 74105, "len_sents_1": 32819, "n_sents": 526, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1710, "title": "WorldSimBench: Towards Video Generation Models as World Simulators", "abs": "Recent advancements in predictive models have demonstrated exceptional capabilities in predicting the future state of objects and scenes. However, the lack of categorization based on inherent characteristics continues to hinder the progress of predictive model development. Additionally, existing benchmarks are unable to effectively evaluate higher-capability, highly embodied predictive models from an embodied perspective. In this work, we classify the functionalities of predictive models into a hierarchy and take the first step in evaluating World Simula\u0002tors by proposing a dual evaluation framework called WorldSimBench. World\u0002SimBench includes Explicit Perceptual Evaluation and Implicit Manipulative Evaluation, encompassing human preference assessments from the visual perspective and action-level evaluations in embodied tasks, covering three representative embodied scenarios: Open-Ended Embodied Environment, Autonomous, Driving, and Robot Manipulation. In the Explicit Perceptual Evaluation, we introduce the HF-Embodied Dataset, a video assessment dataset based on fine-grained human feedback, which we use to train a Human Preference Evaluator that aligns with human perception and explicitly assesses the visual fidelity of World Simu\u0002later. In the Implicit Manipulative Evaluation, we assess the video-action consistency of World Simulators by evaluating whether the generated situation-aware video can be accurately translated into the correct control signals in dynamic environments. Our comprehensive evaluation offers key insights that can drive further innovation in video generation models, positioning World Simulators as a pivotal advancement toward embodied artificial intelligence.", "keywords": [ "World Simulator", "Embodied AI", "Video Generation", "Benchmark" ], "rating_list": [ 5, 8, 5, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "ejEIlXRShF", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Jiannong Cao", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "CN", "position": "Full Professor" }, { "name": "Wanyu Lin", "gender": "Female", "institution": "The Hong Kong Polytechnic University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xu Sun", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Undergrad student" }, { "name": "Zhuoran Li", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" } ], "n_formula": 111, "n_formula_1": 21, "n_ref_uni": 44, "n_ref": 94, "n_ref_all": 107, "n_fig": 14, "n_tab": 6, "L_tab": 3266, "n_element_tab": 230, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1101, "n_element_tab_1": 12, "formula_len_all": 6234, "formula_len_all_1": 1373, "len_all": 235815, "len_all_1": 73247, "len_abs": 1955, "len_title": 169, "len_sents": 78405, "len_sents_1": 30890, "n_sents": 717, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 121, "L_abs": 1980, "title": "Unveiling Molecular Secrets: An LLM-Augmented Linear Model for Explainable and Calibratable Molecular Property Prediction", "abs": "Explainable molecular property prediction is essential for various scientific fields, such as drug discovery and material science. \nDespite delivering intrinsic explainability, linear models struggle with capturing complex, non-linear patterns. Large language models (LLMs), on the other hand, yield accurate predictions through powerful inference capabilities yet fail to provide chemically meaningful explanations for their predictions. This work proposes a novel framework, called MoleX, which leverages LLM knowledge to build a simple yet powerful linear model for accurate molecular property prediction with faithful explanations. The core of MoleX is to model complicated molecular structure-property relationships using a simple linear model, augmented by LLM knowledge and a crafted calibration strategy. Specifically, to extract the maximum amount of task-relevant knowledge from LLM embeddings, we employ information bottleneck-inspired fine-tuning and sparsity-inducing dimensionality reduction. These informative embeddings are then used to fit a linear model for explainable inference. Moreover, we introduce residual calibration to address prediction errors stemming from linear models' insufficient expressiveness of complex LLM embeddings, thus recovering the LLM's predictive power and boosting overall accuracy. Theoretically, we provide a mathematical foundation to justify MoleX\u2019s explainability. Extensive experiments demonstrate that MoleX outperforms existing methods in molecular property prediction, establishing a new milestone in predictive performance, explainability, and efficiency. In particular, MoleX enables CPU inference and accelerates large-scale dataset processing, achieving comparable performance 300$\\times$ faster with 100,000 fewer parameters than LLMs. Additionally, the calibration improves model performance by up to 12.7\\% without compromising explainability. The source code is available at \\url{https://github.com/MoleX2024/MoleX}.", "keywords": [ "Large language models (LLMs)", "explainable AI", "explainability", "molecular property prediction" ], "rating_list": [ 5, 3, 3, 6, 8 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 1, 1, 1, 3, 3 ], "contribution_list": [ 2, 1, 2, 3, 3 ], "confidence_list": [ 2, 4, 5, 4, 2 ] }, { "paper_id": "eiqrnVaeIw", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Daphne Ippolito", "gender": "Female", "institution": "School of Engineering and Applied Science, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Eric Michael Smith", "gender": "Non-Binary", "institution": "Meta AI", "country": "", "position": "Researcher" }, { "name": "Ivan Evtimov", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Research Scientist" }, { "name": "Javier Rando", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Jianfeng Chi", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Research Scientist" }, { "name": "Nicholas Carlini", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Yiming Zhang", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Florian Tramer", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 39, "n_ref": 77, "n_ref_all": 101, "n_fig": 11, "n_tab": 7, "L_tab": 1059, "n_element_tab": 188, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 688, "n_element_tab_1": 146, "formula_len_all": 92, "formula_len_all_1": 0, "len_all": 148575, "len_all_1": 58679, "len_abs": 1574, "len_title": 41, "len_sents": 41839, "len_sents_1": 27303, "n_sents": 300, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 1138, "title": "Persistent Pre-training Poisoning of LLMs", "abs": "Large language models are pre-trained on uncurated text datasets consisting of trillions of tokens scraped from the Web.\nPrior work has shown that: (1) web-scraped pre-training datasets can be practically poisoned by malicious actors; and (2) adversaries can compromise language models after poisoning fine-tuning datasets.\nOur work evaluates for the first time whether language models can also be \\emph{compromised during pre-training}, with a focus on the persistence of pre-training attacks after models are fine-tuned as helpful and harmless chatbots (i.e., after SFT and DPO).\nWe pre-train a series of LLMs from scratch to measure the impact of a potential poisoning adversary under four different attack objectives (denial-of-service, belief manipulation, jailbreaking, and prompt stealing), and across a wide range of model sizes (from 600M to 7B).\nOur main result is that poisoning only 0.1% of a model's pre-training dataset is sufficient for three out of four attacks to measurably persist through post-training. Moreover, simple attacks like denial-of-service persist through post-training with a poisoning rate of only 0.001%.", "keywords": [ "poisoning", "pretraining", "large language models", "security" ], "rating_list": [ 10, 3, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 4, 2, 4, 3, 3 ], "contribution_list": [ 4, 2, 2, 2, 3 ], "confidence_list": [ 5, 5, 4, 4, 3 ] }, { "paper_id": "eimzz4T1wo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "James Lucas", "gender": "Male", "institution": "NVIDIA", "country": "GB", "position": "Researcher" }, { "name": "Jiwoong Choi", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Nikita Durasov", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Pascal Fua", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Full Professor" }, { "name": "Rafid Mahmood", "gender": "unknown", "institution": "University of Ottawa", "country": "CA", "position": "Assistant Professor" }, { "name": "Jose M Alvarez", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Director" }, { "name": "Marc T Law", "gender": "Male", "institution": "NVIDIA", "country": "CA", "position": "Research Scientist" } ], "n_formula": 16, "n_formula_1": 9, "n_ref_uni": 51, "n_ref": 93, "n_ref_all": 103, "n_fig": 6, "n_tab": 4, "L_tab": 3005, "n_element_tab": 355, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1348, "n_element_tab_1": 214, "formula_len_all": 1176, "formula_len_all_1": 741, "len_all": 135884, "len_all_1": 65551, "len_abs": 1372, "len_title": 119, "len_sents": 37140, "len_sents_1": 28876, "n_sents": 271, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1111, "title": "Uncertainty Estimation for 3D Object Detection via Evidential Learning", "abs": "3D object detection is an essential task for computer vision applications in autonomous vehicles and robotics. \nHowever, models often struggle to quantify detection reliability, leading to poor performance on unfamiliar scenes.\nWe introduce a framework for quantifying uncertainty in 3D object detection by leveraging an evidential learning loss on Bird's Eye View representations in the 3D detector.\nThese uncertainty estimates require minimal computational overhead and are generalizable across different architectures.\nWe demonstrate both the efficacy and importance of these uncertainty estimates on identifying out-of-distribution scenes, poorly localized objects, and missing (false negative) detections; our framework consistently improves over baselines by 10-20\\% on average.\nFinally, we integrate this suite of tasks into a system where a 3D object detector auto-labels driving scenes and our uncertainty estimates verify label correctness before the labels are used to train a second model. Here, our uncertainty-driven verification results in a 1\\% improvement in mAP and a 1-2\\% improvement in NDS.", "keywords": [ "3d;detection;uncertainty" ], "rating_list": [ 3, 3, 8, 5 ], "soundness_list": [ 1, 2, 2, 4 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 1, 1, 3, 2 ], "confidence_list": [ 3, 5, 3, 5 ] }, { "paper_id": "eifW0W0xgt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 12, "author_info_list": [ { "name": "Arjun Vikram", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Carlos Guestrin", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Full Professor" }, { "name": "Genghan Zhang", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Jiarui Xu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Karan Dalal", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Tatsunori Hashimoto", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Xiaolong Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Xinhao Li", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Xinlei Chen", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Yann Dubois", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Researcher" }, { "name": "Yu Sun", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 11, "n_ref_uni": 57, "n_ref": 76, "n_ref_all": 108, "n_fig": 13, "n_tab": 5, "L_tab": 2737, "n_element_tab": 80, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 235, "n_element_tab_1": 20, "formula_len_all": 1546, "formula_len_all_1": 833, "len_all": 175717, "len_all_1": 71055, "len_abs": 1362, "len_title": 116, "len_sents": 49939, "len_sents_1": 32546, "n_sents": 513, "n_sents_1": 320, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1348, "title": "Learning to (Learn at Test Time): RNNs with Expressive Hidden States", "abs": "Self-attention performs well in long context but has quadratic complexity. Existing RNN layers have linear complexity, but their performance in long context is limited by the expressive power of their hidden state. Inspired by prior work, we present a practical framework to instantiate sequence modeling layers with linear complexity and expressive hidden states. The key idea is to make the hidden state a machine learning model itself, and the update rule a step of self-supervised learning. Since the hidden state is updated by training even on test sequences, our layers are called Test-Time Training (TTT) layers. We consider two instantiations: TTT-Linear and TTT-MLP, whose hidden state is a linear model and a two-layer MLP respectively. We evaluate our instantiations at the scale of 125M to 1.3B parameters, comparing with a strong Transformer and Mamba, a modern RNN. Both TTT-Linear and TTT-MLP match or exceed the baselines. Similar to Transformer, they can keep reducing perplexity by conditioning on more tokens, while Mamba cannot after 16k context. With preliminary systems optimization, TTT-Linear is already faster than Transformer at 8k context and matches Mamba in wall-clock time. TTT-MLP still faces challenges in memory I/O, but shows larger potential in long context, pointing to a promising direction for future research.", "keywords": [ "test-time training" ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 2, 1 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "eiIM576lpj", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "JaeHyeon Kim", "gender": "Male", "institution": "Korea Aerospace University", "country": "KR", "position": "MS student" }, { "name": "Jay Hoon Jung", "gender": "Male", "institution": "Korea Aerospace University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 8, "n_ref_uni": 17, "n_ref": 34, "n_ref_all": 48, "n_fig": 2, "n_tab": 7, "L_tab": 3680, "n_element_tab": 334, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2495, "n_element_tab_1": 208, "formula_len_all": 175, "formula_len_all_1": 287, "len_all": 96516, "len_all_1": 56350, "len_abs": 1123, "len_title": 145, "len_sents": 34255, "len_sents_1": 24652, "n_sents": 248, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1133, "title": "Robustness through Random Activation: Adversarial Training with Bernoulli Rectified Linear Units", "abs": "Despite their considerable achievements across a range of domains, deep learning models have been demonstrated to be susceptible to adversarial attacks.\nIn order to mitigate this vulnerability, adversarial training has become a prevalent defense strategy. \nIn this context, we propose Bernoulli Rectified Linear Units (BReLU), an activation function designed to further enhance the effectiveness of adversarial training. In contrast to conventional activation functions, BReLU modulates activation probabilities in accordance with input values, thereby introducing input-dependent randomness into the model.\nThe experimental results demonstrate that the incorporation of BReLU into adversarial training significantly enhances the robustness of the model against adversarial attacks. Specifically, on the CIFAR-10 dataset using the ResNet-18 model, BReLU improved robustness by 15\\% under FGSM, by 8\\% under PGD-20, and by 54\\% under the CW attack compared to ReLU.\nOur findings indicate that BReLU represents a promising addition to adversarial training techniques for strengthening deep learning models against adversarial threats.", "keywords": [ "Adversarial training", "Adversarial attack", "Activation function" ], "rating_list": [ 3, 5, 1, 5, 3 ], "soundness_list": [ 1, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 5, 4, 5, 4, 5 ] }, { "paper_id": "ei3qCntB66", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 12, "author_info_list": [ { "name": "Aishan Liu", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Associate Professor" }, { "name": "Changgan Yin", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Chenyu Zhu", "gender": "Male", "institution": "Software of Engineering, Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Hangtao Zhang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Leo Yu Zhang", "gender": "Male", "institution": "Griffith University", "country": "AU", "position": "Researcher" }, { "name": "Lulu Xue", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Minghui Li", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Peijin Guo", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Shengshan Hu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Xianlong Wang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yichen Wang", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Ziqi Zhou", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 73, "n_ref": 146, "n_ref_all": 187, "n_fig": 19, "n_tab": 11, "L_tab": 5449, "n_element_tab": 567, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 1164, "n_element_tab_1": 148, "formula_len_all": 582, "formula_len_all_1": 516, "len_all": 325777, "len_all_1": 74710, "len_abs": 1373, "len_title": 106, "len_sents": 104324, "len_sents_1": 34487, "n_sents": 778, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 58, "L_abs": 1393, "title": "BadRobot: Manipulating Embodied LLMs in the Physical World", "abs": "Embodied AI represents systems where AI is integrated into physical entities, enabling them to perceive and interact with their surroundings. Large Language Model (LLM), which exhibits powerful language understanding abilities, has been extensively employed in embodied AI by facilitating sophisticated task planning. However, a critical safety issue remains overlooked: could these embodied LLMs perpetrate harmful behaviors? In response, we introduce BadRobot, a novel attack paradigm aiming to make embodied LLMs violate safety and ethical constraints through typical voice-based user-system interactions. Specifically, three vulnerabilities are exploited to achieve this type of attack: (i) manipulation of LLMs within robotic systems, (ii) misalignment between linguistic outputs and physical actions, and (iii) unintentional hazardous behaviors caused by world knowledge's flaws. Furthermore, we construct a benchmark of various malicious physical action queries to evaluate BadRobot's attack performance. Based on this benchmark, extensive experiments against existing prominent embodied LLM frameworks (e.g., Voxposer, Code as Policies, and ProgPrompt) demonstrate the effectiveness of our BadRobot. More demonstrations are available at an anonymous address: https://Embodied-LLMs-Safety.github.io. \n\nWarning: This paper contains harmful AI-generated language and aggressive actions.", "keywords": [ "Robotics", "Safety Risks", "Embodied AI" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "ehr4oTe6XI", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Karsten Roth", "gender": "Not Specified", "institution": "University of Tuebingen", "country": "DE", "position": "PhD student" }, { "name": "Th\u00e9o Uscidda", "gender": "Male", "institution": "Helmholtz Zentrum M\u00fcnchen", "country": "DE", "position": "Visiting PhD student" }, { "name": "Zeynep Akata", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "marco cuturi", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Fabian J. Theis", "gender": "unknown", "institution": "Technical University Munich", "country": "", "position": "Full Professor" }, { "name": "Luca Vincent Eyring", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 46, "n_formula_1": 15, "n_ref_uni": 58, "n_ref": 142, "n_ref_all": 195, "n_fig": 5, "n_tab": 6, "L_tab": 1971, "n_element_tab": 163, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5626, "formula_len_all_1": 1992, "len_all": 388102, "len_all_1": 217272, "len_abs": 1439, "len_title": 110, "len_sents": 102857, "len_sents_1": 86040, "n_sents": 685, "n_sents_1": 359, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1449, "title": "Disentangled Representation Learning with the Gromov-Monge Gap", "abs": "Learning disentangled representations from unlabelled data is a fundamental challenge in machine learning. Solving it may unlock other problems, such as generalization, interpretability, or fairness. Although remarkably challenging to solve in theory, disentanglement is often achieved in practice through prior matching. Furthermore, recent works have shown that prior matching approaches can be enhanced by leveraging geometrical considerations, e.g., by learning representations that preserve geometric features of the data, such as distances or angles between points. However, matching the prior while preserving geometric features is challenging, as a mapping that *fully* preserves these features while aligning the data distribution with the prior does not exist in general. To address these challenges, we introduce a novel approach to disentangled representation learning based on quadratic optimal transport. We formulate the problem using Gromov-Monge maps that transport one distribution onto another with minimal distortion of predefined geometric features, preserving them *as much as can be achieved*. To compute such maps, we propose the Gromov-Monge-Gap (GMG), a regularizer quantifying whether a map moves a reference distribution with minimal geometry distortion. We demonstrate the effectiveness of our approach for disentanglement across four standard benchmarks, outperforming other methods leveraging geometric considerations.", "keywords": [ "Disentangled Representational Learning", "Optimal Transport", "Geometric Learning" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 5, 3, 2, 3 ] }, { "paper_id": "ehlZFDxwJo", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Albert Zomaya", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Full Professor" }, { "name": "Sen Fu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Wei Bao", "gender": "unknown", "institution": "University of Sydney", "country": "", "position": "Lecturer" }, { "name": "Zhengjie Yang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Postdoc" } ], "n_formula": 69, "n_formula_1": 26, "n_ref_uni": 37, "n_ref": 65, "n_ref_all": 77, "n_fig": 5, "n_tab": 7, "L_tab": 1411, "n_element_tab": 110, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1617, "n_element_tab_1": 96, "formula_len_all": 6295, "formula_len_all_1": 2157, "len_all": 136516, "len_all_1": 69923, "len_abs": 1446, "len_title": 94, "len_sents": 40617, "len_sents_1": 28986, "n_sents": 360, "n_sents_1": 237, "venue": "ICLR2025", "decision": "withdraw", "L_title": 45, "L_abs": 1481, "title": "PERSONALIZED FEDERATED PARTIAL LABEL LEARNING", "abs": "Partial Label Learning (PLL) is known as a valuable learning technique that trains Machine Learning (ML) models on partial label datasets, where the ground truth label is concealed within the candidate label set of each data instance. It learns label correlation based on a single centralized dataset to predict the latent true label. When data is non-independent and identically distributed (non-i.i.d.) among workers in Federated Learning (FL), the label correlation interference problem occurs. To address the issue, in this paper, we propose pFedPLL, a personalized federated partial label learning algorithm with two new designs. In Label Correlation Isolation (LCI), we first develop a twin-module architecture, where a feature-level correlation matrix layer for each worker is isolated locally to prevent it from being interfered with by others. In Label Correlation Personalization (LCP), we then propose a bi-directional calibration loss to identify a more accurate learning direction, where the positive calibration aligns the prediction result with the latent true label, and the negative calibration pushes away the prediction result that falls into the non-candidate label set. We provide a convergence analysis of pFedPLL with a rate of $O\\left(\\sqrt{\\frac{1}{T}}\\right)$ for smooth non-convex problems. Experiment results demonstrate that pFedPLL outperforms SOTA federated PLL algorithms and the federated version of centralized PLL algorithms across nine datasets.", "keywords": [ "Federated learning", "Partial label learning", "Weakly supervised learning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "eh1fL0zw8o", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Haitao Huang", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "MS student" }, { "name": "Jianmin Wang", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "PhD student" }, { "name": "Lai Hou Tim", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Longyue Wang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Peng Zhou", "gender": "unknown", "institution": "Hunan University", "country": "CN", "position": "PhD student" }, { "name": "Pengsen Ma", "gender": "unknown", "institution": "Hunan University", "country": "", "position": "PhD student" }, { "name": "Wei Liu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Distinguished Scientist" }, { "name": "Xibao Cai", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "PhD student" }, { "name": "xiangxiang Zeng", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 44, "n_ref": 71, "n_ref_all": 89, "n_fig": 5, "n_tab": 8, "L_tab": 2792, "n_element_tab": 468, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 777, "n_element_tab_1": 158, "formula_len_all": 323, "formula_len_all_1": 323, "len_all": 150593, "len_all_1": 61752, "len_abs": 1424, "len_title": 128, "len_sents": 43076, "len_sents_1": 29963, "n_sents": 318, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1420, "title": "Large Language and Protein Assistant for Protein-Protein Interactions prediction", "abs": "Predicting the types and affinities of protein-protein interactions (PPIs) is crucial for understanding biological processes and discovering macromolecular drugs. While encoding proteins themselves is essential, PPI networks can also provide rich prior knowledge for these predictive tasks. However, existing methods oversimplify the problem of PPI prediction in a semi-supervised manner when utilizing PPI networks, limiting their practical application. Furthermore, how to effectively use the rich prior knowledge of PPI networks for novel proteins not present in the network remains an unexplored issue. Additionally, due to inflexible architectures, existing methods cannot handle complexes containing an arbitrary number of proteins. To overcome these limitations, we introduce LLaPA (Large Language and Protein Assistant), a multimodal large language model that integrates proteins and PPI networks. LLaPA offers a more rational approach to utilizing PPI networks for PPI prediction and can fully exploit the information of PPI networks for unseen proteins. Through natural language instructions, LLaPA can accept any number of protein sequences and has the potential to perform various protein tasks. Experiments show that LLaPA achieves state-of-the-art performance in multi-label PPI type prediction and is capable of predicting the binding affinity between multiple interacting proteins based on sequence data.", "keywords": [ "Multimodal;LLM;Protein-Protein Interaction;Binding Affinity Prediction" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "eghAocvqBk", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Fan Bao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Guande He", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Jianfei Chen", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Jun Zhu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Kaiwen Zheng", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 69, "n_formula_1": 28, "n_ref_uni": 38, "n_ref": 115, "n_ref_all": 165, "n_fig": 5, "n_tab": 10, "L_tab": 3499, "n_element_tab": 455, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2510, "n_element_tab_1": 348, "formula_len_all": 8329, "formula_len_all_1": 2203, "len_all": 191214, "len_all_1": 68533, "len_abs": 1396, "len_title": 80, "len_sents": 47848, "len_sents_1": 26469, "n_sents": 397, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 32, "L_abs": 1397, "title": "Diffusion Bridge Implicit Models", "abs": "Denoising diffusion bridge models (DDBMs) are a powerful variant of diffusion models for interpolating between two arbitrary paired distributions given as endpoints. Despite their promising performance in tasks like image translation, DDBMs require a computationally intensive sampling process that involves the simulation of a (stochastic) differential equation through hundreds of network evaluations. In this work, we take the first step in fast sampling of DDBMs without extra training, motivated by the well-established recipes in diffusion models. We generalize DDBMs via a class of non-Markovian diffusion bridges defined on the discretized timesteps concerning sampling, which share the same marginal distributions and training objectives, give rise to generative processes ranging from stochastic to deterministic, and result in diffusion bridge implicit models (DBIMs). DBIMs are not only up to 25$\\times$ faster than the vanilla sampler of DDBMs but also induce a novel, simple, and insightful form of ordinary differential equation (ODE) which inspires high-order numerical solvers. Moreover, DBIMs maintain the generation diversity in a distinguished way, by using a booting noise in the initial sampling step, which enables faithful encoding, reconstruction, and semantic interpolation in image translation tasks. Code is available at \\url{https://github.com/thu-ml/DiffusionBridge}.", "keywords": [ "Diffusion Bridge Models", "Image Translation", "Fast Sampling" ], "rating_list": [ 5, 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3, 3 ] }, { "paper_id": "egHptuv7hx", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Eric P. Xing", "gender": "Male", "institution": "Mohamed bin Zayed Univeristy of AI", "country": "AE", "position": "Full Professor" }, { "name": "Jianshu She", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "PhD student" }, { "name": "Qirong Ho", "gender": "unknown", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Xinyue Li", "gender": "Female", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "MS student" }, { "name": "Zhengzhong Liu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 39, "n_ref": 60, "n_ref_all": 83, "n_fig": 11, "n_tab": 8, "L_tab": 4575, "n_element_tab": 289, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 418, "formula_len_all_1": 413, "len_all": 154644, "len_all_1": 51026, "len_abs": 1938, "len_title": 123, "len_sents": 44532, "len_sents_1": 25752, "n_sents": 339, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1940, "title": "How does controllability emerge in language models during pretraining?", "abs": "Language models can intervened upon by steering their internal representations, which alters the degree to which concepts such as emotional tone, style, truthfulness, and safety are expressed in their generative outputs. This paper demonstrates that intervention efficacy, measured by linear steerability (the ability to adjust outputs via linear transformations of hidden states), emerges abruptly during pre-training, and furthermore, even closely-related concepts (e.g. anger and sadness) can emerge at different stages of pre-training. To understand how the steerability of internal representations changes during pre-training, we introduce the \"Intervention Detector\" (ID), which applies unsupervised learning techniques to hidden states under different stimuli, and generates concept representations that can be used to steer the text generation of language models. The extracted concept representations are used to compute an ID score, measuring their alignment with the model\u2019s hidden states. This ID score can be used to approximately predict the time of emergence of effective intervention by steering different concepts, and the degree to which each concept is able to intervene.\nBy analyzing ID scores across a longitudinal series of models taken at different stages of pre-training, we demonstrate that, as pre-training progresses, concepts become increasingly easier to extract via linear methods, which correlates with the emergence of steerability. For instance, in the CrystalCoder model, the linear steerability of the concept \"anger\" emerges at 68\\% of pre-training, whereas the linear steerability of the concept ``sadness\" emerges at 93\\% of the pre-training process. We use heatmap visualizations and other metrics (eg., entropy, cosine similarity, tSNE) to study these differences and validate the reliability and generalizability of ID scores through model interventions using the extracted concept representations.", "keywords": [ "controllability", "ability emergence", "pre-training models", "dimentionality reduction", "representations" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "ef3tbYHIVn", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Andreas Krause", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Noah Liniger", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Riccardo De Santi", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" } ], "n_formula": 67, "n_formula_1": 16, "n_ref_uni": 57, "n_ref": 115, "n_ref_all": 131, "n_fig": 4, "n_tab": 4, "L_tab": 2334, "n_element_tab": 29, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 388, "n_element_tab_1": 3, "formula_len_all": 6948, "formula_len_all_1": 1494, "len_all": 207502, "len_all_1": 77979, "len_abs": 1345, "len_title": 111, "len_sents": 59148, "len_sents_1": 34228, "n_sents": 465, "n_sents_1": 251, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1354, "title": "Generative bandit optimization via diffusion posterior sampling", "abs": "Many real-world discovery problems, including drug and material design, can be modeled within the bandit optimization framework, where an agent selects a sequence of experiments to efficiently optimize an unknown reward function. However, classic bandit algorithms operate on fixed finite or continuous action sets, making discovering novel designs impossible in the former case, and often leading to the curse of dimensionality in the latter, thus rendering these methods impractical. In this work, we first formalize the *generative bandit* setting, where an agent wishes to maximize an unknown reward function over the support of a data distribution, often called *data manifold*, which implicitly encodes complex constraints (e.g., the geometry of valid molecules), and from which (unlabeled) sample data is available (e.g., a dataset of valid molecules). We then propose Diffusion Posterior Sampling (DiffPS), an algorithm that tackles the exploration-exploitation problem directly on the learned data manifold by leveraging a conditional diffusion model. We formally show that the statistical complexity of DiffPS adapts to the *intrinsic dimensionality* of the data, overcoming the curse of dimensionality in high-dimensional settings. Our experimental evaluation supports the theoretical claims and demonstrates promising performance in practice.", "keywords": [ "bandit optimization", "diffusion models", "posterior sampling", "algorithmic discovery" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 2, 3, 4 ] }, { "paper_id": "eeyhnqYbxw", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Amirhossein Farzam", "gender": "unknown", "institution": "Google", "country": "US", "position": "Intern" }, { "name": "Guillermo Sapiro", "gender": "unknown", "institution": "Apple", "country": "", "position": "Apple" }, { "name": "Joshua M. Susskind", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Juan Matias Di Martino", "gender": "Male", "institution": "Universidad Cat\u00f3lica del Uruguay", "country": "US", "position": "Full Professor" }, { "name": "Oded Schlesinger", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 10, "n_ref_uni": 62, "n_ref": 128, "n_ref_all": 158, "n_fig": 14, "n_tab": 2, "L_tab": 779, "n_element_tab": 62, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 689, "n_element_tab_1": 54, "formula_len_all": 807, "formula_len_all_1": 535, "len_all": 187618, "len_all_1": 68117, "len_abs": 1693, "len_title": 131, "len_sents": 62877, "len_sents_1": 34053, "n_sents": 405, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1702, "title": "The Geometry of Attention: Ricci Curvature and Transformers Training and Robustness", "abs": "Transformer models have revolutionized machine learning, and the theoretical underpinnings behind their success are only now starting to be explored. \nIn this work, we analyze the performance and robustness of transformers by considering the attention mechanism as a graph operator, focusing on the geometry of attention maps viewed as weighted graphs.\nSpecifically, we investigate the role of Ricci curvature, a metric closely tied to graph spectral properties and system robustness, in shaping the training dynamics and robustness of transformers. \nOur theoretical analysis establishes a link between Ricci curvature and the convergence of gradient descent on transformers, and consequently, their training and fine tuning.\nWe also show that a higher frequency of more positive values in the Ricci curvature distribution of attention graphs, therefore more system robustness, leads to more robust transformers, highlighting the impact of curvature on the robustness of transformers.\nLeveraging these insights, we propose an efficient regularization method to train curvature-adjusted transformers. \nSupporting our theoretical findings, experiments show that our proposed attention curvature manipulation can improve the learning speed, performance, or generalizability of vision and language transformers.\nAdditionally, our observations point to a trade-off between their performance and robustness.\nThis work demonstrates that the geometry of the attention map provides a theoretically elegant and computationally versatile framework for analyzing and manipulating transformers training, generalization, performance, and robustness, opening new avenues for designing models using geometric concepts.", "keywords": [ "Transformers", "Attention", "Geometry", "Robustness" ], "rating_list": [ 5, 6, 3, 5, 5 ], "soundness_list": [ 2, 4, 1, 2, 3 ], "presentation_list": [ 4, 4, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 5, 3, 5 ] }, { "paper_id": "eev4PHiMir", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Hongchao Li", "gender": "Male", "institution": "Tokyo University, Tokyo Institute of Technology", "country": "", "position": "MS student" }, { "name": "Liu Ziyin", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Masahito Ueda", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" } ], "n_formula": 170, "n_formula_1": 40, "n_ref_uni": 42, "n_ref": 56, "n_ref_all": 122, "n_fig": 7, "n_tab": 1, "L_tab": 155, "n_element_tab": 4, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12511, "formula_len_all_1": 2155, "len_all": 184006, "len_all_1": 80055, "len_abs": 1055, "len_title": 72, "len_sents": 51666, "len_sents_1": 33133, "n_sents": 599, "n_sents_1": 305, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1261, "title": "Noise Balance and Stationary Distribution of Stochastic Gradient Descent", "abs": "How the stochastic gradient descent (SGD) navigates the loss landscape of a neural network remains poorly understood. This work shows that the minibatch noise of SGD regularizes the solution towards a noise-balanced solution whenever the loss function contains a rescaling symmetry. We prove that when the rescaling symmetry exists, the SGD dynamics is limited to only a low-dimensional subspace and prefers a special set of solutions in an infinitely large degenerate manifold, which offers a partial explanation of the effectiveness of SGD in training neural networks. We then apply this result to derive the stationary distribution of stochastic gradient flow for a diagonal linear network with arbitrary depth and width, which is the first analytical expression of the stationary distribution of SGD in a high-dimensional non-quadratic potential. The stationary distribution exhibits complicated nonlinear phenomena such as phase transitions, loss of ergodicity, memory effects, and fluctuation inversion. These phenomena are shown to exist uniquely in deep networks, highlighting a fundamental difference between deep and shallow models. Lastly, we discuss the implication of the proposed theory for the practical problem of variational Bayesian inference.", "keywords": [ "stochastic gradient descent", "stationary distribution", "stochastic differential equation", "phase transition" ], "rating_list": [ 3, 3, 6, 6, 3 ], "soundness_list": [ 2, 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 5, 2 ] }, { "paper_id": "eePww5u7J3", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Shengcao Cao", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Yu-Xiong Wang", "gender": "unknown", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Yuxiang Lu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 112, "n_ref": 296, "n_ref_all": 358, "n_fig": 8, "n_tab": 21, "L_tab": 5233, "n_element_tab": 614, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 696, "n_element_tab_1": 37, "formula_len_all": 579, "formula_len_all_1": 315, "len_all": 329916, "len_all_1": 62297, "len_abs": 97, "len_title": 151, "len_sents": 76165, "len_sents_1": 30339, "n_sents": 538, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 103, "L_abs": 1305, "title": "Swiss Army Knife: Synergizing Biases in Knowledge from Vision Foundation Models for Multi-Task Learning", "abs": "Vision Foundation Models (VFMs) have demonstrated outstanding performance on numerous downstream tasks. However, due to their inherent representation biases originating from different training paradigms, VFMs exhibit advantages and disadvantages across distinct vision tasks. Although amalgamating the strengths of multiple VFMs for downstream tasks is an intuitive strategy, effectively exploiting these biases remains a significant challenge. In this paper, we propose a novel and versatile \"Swiss Army Knife\" (SAK) solution, which adaptively distills knowledge from a committee of VFMs to enhance multi-task learning. Unlike existing methods that use a single backbone for knowledge transfer, our approach preserves the unique representation bias of each teacher by collaborating the lightweight Teacher-Specific Adapter Path modules with the Teacher-Agnostic Stem. Through dynamic selection and combination of representations with Mixture-of-Representations Routers, our SAK is capable of synergizing the complementary strengths of multiple VFMs. Extensive experiments show that our SAK remarkably outperforms prior state of the arts in multi-task learning by 10\\% on the NYUD-v2 benchmark, while also providing a flexible and robust framework that can readily accommodate more advanced model designs.", "keywords": [ "Vision Foundation Model", "Multi-Task Learning", "Knowledge Distillation" ], "rating_list": [ 8, 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3, 3 ] }, { "paper_id": "eeJz7eDWKO", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Anish Dhir", "gender": "Male", "institution": "Imperial College London, Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "James Requeima", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "Postdoc" }, { "name": "Mark van der Wilk", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" }, { "name": "Matthew Ashman", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 35, "n_formula_1": 17, "n_ref_uni": 40, "n_ref": 96, "n_ref_all": 123, "n_fig": 2, "n_tab": 5, "L_tab": 677, "n_element_tab": 84, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 111, "n_element_tab_1": 15, "formula_len_all": 1854, "formula_len_all_1": 1008, "len_all": 142061, "len_all_1": 70943, "len_abs": 1219, "len_title": 101, "len_sents": 45240, "len_sents_1": 33898, "n_sents": 362, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1223, "title": "A Meta-Learning Approach to Bayesian Causal Discovery", "abs": "Discovering a unique causal structure is difficult due to both inherent identifiability issues, and the consequences of finite data.\nAs such, uncertainty over causal structures, such as those obtained from a Bayesian posterior, are often necessary for downstream tasks.\nFinding an accurate approximation to this posterior is challenging, due to the large number of possible causal graphs, as well as the difficulty in the subproblem of finding posteriors over the functional relationships of the causal edges.\nRecent works have used Bayesian meta learning to view the problem of posterior estimation as a supervised learning task.\nYet, these methods are limited as they cannot reliably sample from the posterior over causal structures and fail to encode key properties of the posterior, such as correlation between edges and permutation equivariance with respect to nodes.\nTo address these limitations, we propose a Bayesian meta learning model that allows for sampling causal structures from the posterior and encodes these key properties.\nWe compare our meta-Bayesian causal discovery against existing Bayesian causal discovery methods, demonstrating the advantages of directly learning a posterior over causal structure.", "keywords": [ "neural processes", "bayesian causal discovery", "transformers" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 2, 2, 3, 4 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "eeC1bSkUrY", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 1, "author_info_list": [ { "name": "Xuefeng Xu", "gender": "Male", "institution": "University of Warwick", "country": "GB", "position": "PhD student" } ], "n_formula": 42, "n_formula_1": 8, "n_ref_uni": 40, "n_ref": 61, "n_ref_all": 98, "n_fig": 7, "n_tab": 7, "L_tab": 3164, "n_element_tab": 248, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1543, "n_element_tab_1": 116, "formula_len_all": 4335, "formula_len_all_1": 703, "len_all": 157711, "len_all_1": 55385, "len_abs": 1214, "len_title": 110, "len_sents": 43333, "len_sents_1": 24307, "n_sents": 370, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1220, "title": "FedPS: Federated data Preprocessing via aggregated Statistics", "abs": "Data preprocessing is a crucial step in machine learning that significantly influences model accuracy and performance. In Federated Learning (FL), where multiple entities collaboratively train a model using decentralized data, the importance of preprocessing is often overlooked. This is particularly true in Non-IID settings, where clients hold heterogeneous datasets, requiring aggregated parameter estimates to perform consistent data preprocessing. In this paper, we introduce FedPS, a comprehensive suite of tools for federated data preprocessing. FedPS leverages aggregated statistics, data sketching, and federated machine learning models to address the challenges posed by distributed and diverse datasets in FL. Additionally, we resolve key numerical issues in power transforms by improving numerical stability through log-space computations and constrained optimization. Our proposed Federated Power Transform algorithm, based on Brent\u2019s method, achieves superlinear convergence. Experimental results demonstrate the impact of effective data preprocessing in federated learning, highlighting FedPS as a versatile and robust solution compared to existing frameworks. The implementation of FedPS is open-sourced.", "keywords": [ "Data Preprocessing", "Federated Learning", "Aggregated Statistics" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "ee2c4MEx9l", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Gihyun Kwon", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jong Chul Ye", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 44, "n_ref": 94, "n_ref_all": 118, "n_fig": 14, "n_tab": 2, "L_tab": 619, "n_element_tab": 71, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 533, "n_element_tab_1": 68, "formula_len_all": 744, "formula_len_all_1": 743, "len_all": 147071, "len_all_1": 62623, "len_abs": 206, "len_title": 133, "len_sents": 42194, "len_sents_1": 29138, "n_sents": 335, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1105, "title": "TweedieMix: Improving Multi-Concept Fusion for Diffusion-based Image/Video Generation", "abs": "Despite significant advancements in customizing text-to-image and video generation models, generating images and videos that effectively integrate multiple personalized concepts remains challenging. To address this, we present TweedieMix, a novel method for composing customized diffusion models during the inference phase. By analyzing the properties of reverse diffusion sampling, our approach divides the sampling process into two stages. During the initial steps, we apply a multiple object-aware sampling technique to ensure the inclusion of the desired target objects. In the later steps, we blend the appearances of the custom concepts in the de-noised image space using Tweedie's formula. Our results demonstrate that TweedieMix can generate multiple personalized concepts with higher fidelity than existing methods. Moreover, our framework can be effortlessly extended to image-to-video diffusion models by extending the residual layer's features across frames}, enabling the generation of videos that feature multiple personalized concepts. Results and source code are in anonymous project page.", "keywords": [ "Diffusion Models", "Custom Concepts", "Text-to-Image", "Video" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "ed7zI29lRF", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Adam Klivans", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "Professor" }, { "name": "Gautam Chandrasekaran", "gender": "Male", "institution": " University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Konstantinos Stavropoulos", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" }, { "name": "Lin Lin Lee", "gender": "Female", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" } ], "n_formula": 162, "n_formula_1": 19, "n_ref_uni": 44, "n_ref": 118, "n_ref_all": 128, "n_fig": 0, "n_tab": 7, "L_tab": 1447, "n_element_tab": 39, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 46, "n_element_tab_1": 15, "formula_len_all": 18636, "formula_len_all_1": 1743, "len_all": 285481, "len_all_1": 78578, "len_abs": 1146, "len_title": 132, "len_sents": 78750, "len_sents_1": 33784, "n_sents": 821, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1153, "title": "Learning Neural Networks with Distribution Shift: Efficiently Certifiable Guarantees", "abs": "We give the first provably efficient algorithms for learning neural networks with respect to distribution shift. We work in the Testable Learning with Distribution Shift framework (TDS learning) of Klivans et al. (2024), where the learner receives labeled examples from a training distribution and unlabeled examples from a test distribution and must either output a hypothesis with low test error or reject if distribution shift is detected. No assumptions are made on the test distribution. \n\nAll prior work in TDS learning focuses on classification, while here we must handle the setting of nonconvex regression. Our results apply to real-valued networks with arbitrary Lipschitz activations and work whenever the training distribution has strictly sub-exponential tails. For training distributions that are bounded and hypercontractive, we give a fully polynomial-time algorithm for TDS learning one hidden-layer networks with sigmoid activations. We achieve this by importing classical kernel methods into the TDS framework using data-dependent feature maps and a type of kernel matrix that couples samples from both train and test distributions.", "keywords": [ "pac learning", "distribution shift", "distribution testing", "testable learning", "neural networks", "kernel methods" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 4, 3, 4, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 4 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "ed75tWzgt0", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Shenao Zhang", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Yibo Wang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhaoran Wang", "gender": "Not Specified", "institution": "Northwestern University", "country": "US", "position": "Associate Professor" }, { "name": "Zhihan Liu", "gender": "Male", "institution": "Northwestern University", "country": "", "position": "PhD student" }, { "name": "Zikun Zhang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 109, "n_formula_1": 30, "n_ref_uni": 58, "n_ref": 90, "n_ref_all": 129, "n_fig": 3, "n_tab": 10, "L_tab": 1750, "n_element_tab": 165, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 394, "n_element_tab_1": 7, "formula_len_all": 12107, "formula_len_all_1": 1964, "len_all": 209903, "len_all_1": 60911, "len_abs": 1614, "len_title": 115, "len_sents": 50846, "len_sents_1": 25691, "n_sents": 479, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1621, "title": "Provably Efficient and Practical Self-Play for Better LLM Alignment", "abs": "Reinforcement Learning with Human Feedback (RLHF) has gained significant attention for aligning AI behavior with human preferences. Self-play style RLHF has shown strong advantages, as highlighted by many studies. However, current self-play style RLHF approaches face several limitations, including the lack of provable sample efficiency, absence of active exploration, and limited diversity in training data. To address these challenges, we propose a novel RLHF framework that balances exploration and exploitation while providing theoretical guarantees. We introduce Two-Agent Nash Policy Optimization (TANPO) as an equivalent and easy-to-implement two-agent algorithm building on this framework. In TANPO, the two players are trained using different loss functions to ensure more diverse and informative data collection. We also propose Single-Agent Diversity-driven Optimization (SADPO), a single-agent approximation of TANPO, supported by both theoretical analysis and empirical evidence. Our theoretical analysis shows that our theoretical algorithm framework enjoys sublinear regret under general function approximation and mild structural conditions, with a detailed analysis provided for the linear case. Empirically, we implement TANPO and SADPO using Zephyr-7B-SFT as our base model, outperforming several baselines across multiple evaluation benchmarks, such as AlpacaEval 2.0, MT-Bench and various standard academic benchmarks. Our experiments also show that TANPO improves performance on AlpacaEval 2.0 over extended training epochs, demonstrating its ability to consistently improve and reduce overfitting.", "keywords": [ "Large Language Model", "two player game", "DPO", "Sample-efficient RLHF" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 1 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "eciCtsqGc8", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Danilo Mandic", "gender": "unknown", "institution": "Imperial College London", "country": "", "position": "Full Professor" }, { "name": "Harry J Davies", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Postdoc" }, { "name": "James Monsen", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 68, "n_ref_all": 87, "n_fig": 10, "n_tab": 6, "L_tab": 1443, "n_element_tab": 207, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 574, "n_element_tab_1": 7, "formula_len_all": 48, "formula_len_all_1": 0, "len_all": 172088, "len_all_1": 54429, "len_abs": 1899, "len_title": 113, "len_sents": 64151, "len_sents_1": 29275, "n_sents": 386, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1907, "title": "Interpretable Pre-Trained Transformers for Heart Time-Series Data", "abs": "Interpretability of artificial intelligence models is vital in healthcare, as a poorly informed decision can directly impact the health and well-being of patients. This means that, owing to their black box nature, deep-learning solutions that may even yield high accuracy often fail to be adopted in real-world healthcare settings. To this end, we employ the generative pre-trained transformer (GPT) framework to clinical heart time-series data, to create two pre-trained general purpose cardiac models, termed PPG-PT and ECG-PT. We place a special emphasis on making both such pre-trained models fully interpretable. This is achieved firstly through aggregate attention maps which show that, in order to make predictions, the model focuses on similar points in previous cardiac cycles and gradually broadens its attention in deeper layers. Next, we show that tokens with the same value, which occur at different distinct points in the electrocardiography (ECG) and photoplethysmography (PPG) cycle, form separate clusters in a high dimensional space. Such clusters are formed according to the phase of the cardiac cycle, as the tokens propagate through the transformer blocks. Finally, we highlight that individual attention heads correspond to specific physiologically relevant features, such as the dicrotic notch in PPG and the P-wave in ECG. Importantly, it is also demonstrated that these pre-trained models are straightforward to fine-tune for tasks such as the classification of atrial fibrillation (AF), and beat detection in photoplethysmography. The so introduced PPG-PT and ECG-PT models achieve accuracy comparable to the state-of-the-art for both tasks, whilst crucially retaining their interpretability and explainability. This is demonstrated in the AF-screening fine-tuned model, with attention clearly shifting to regions in the context that are strongly indicative of atrial fibrillation.", "keywords": [ "biosignals", "interpretability", "healthcare", "transformers" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "ech9J3xl9X", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Ankush Kumar", "gender": "unknown", "institution": "Infosys Limited", "country": "IN", "position": "Researcher" }, { "name": "Balaji A J", "gender": "Male", "institution": "Infosys Limited", "country": "IN", "position": "Researcher" }, { "name": "Gagan Gayari", "gender": "Male", "institution": "Infosys Limited", "country": "IN", "position": "Researcher" }, { "name": "Harshini K", "gender": "Female", "institution": "Infosys Limited", "country": "IN", "position": "Researcher" }, { "name": "Kamalkumar Rathinasamy", "gender": "Male", "institution": "Infosys Limited", "country": "IN", "position": "Principal Researcher" }, { "name": "Mohammed Rafee Tarafdar", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Rajab Ali Mondal", "gender": "Male", "institution": "Strategic Technology Group, Infosys Limited (infosys.com)", "country": "IN", "position": "Researcher" }, { "name": "Sreenivasa Raghavan K S", "gender": "Male", "institution": "Infosys Limited", "country": "IN", "position": "Consultant" }, { "name": "Swayam Singh", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 3, "n_ref": 11, "n_ref_all": 25, "n_fig": 0, "n_tab": 3, "L_tab": 450, "n_element_tab": 39, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 450, "n_element_tab_1": 39, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 70493, "len_all_1": 70493, "len_abs": 1137, "len_title": 101, "len_sents": 15566, "len_sents_1": 15566, "n_sents": 140, "n_sents_1": 140, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1117, "title": "Narrow Transformer: Mono-lingual Code SLM for Desktop", "abs": "This paper presents NT-Java-1.1B, an open-source specialized code language model built on StarCoderBase-1.1B, designed for coding tasks in Java programming. NT-Java-1.1B achieves state-of-the-art performance, surpassing its base model and majority of other models of similar size on MultiPL-E Java code benchmark. While there have been studies on extending large, generic pre-trained models to improve proficiency in specific programming languages like Python, similar investigations on small code models for other programming languages are lacking. Large code models require specialized hardware like GPUs for inference, highlighting the need for research into building small code models that can be deployed on developer desktops. This paper addresses this research gap by focusing on the development of a small Java code model, NT-Java-1.1B, and its quantized versions, which performs comparably to open models around 1.1B on MultiPL-E Java code benchmarks, making them ideal for desktop deployment. This paper establishes the foundation for specialized models across languages and sizes for a family of NT Models.", "keywords": [ "Narrow Transformer", "Code SLMs", "Desktop Deployment", "Lightweight Code Language Models", "Small Language Models", "Language Specific Models", "Monolingual Code Language Model" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 1, 1, 1, 1 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "ecRyUAPshY", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Eunsol Choi", "gender": "unknown", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Greg Durrett", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Shrey Pandit", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "MS student" }, { "name": "Xi Ye", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Zeyu Liu", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 86, "n_ref_all": 117, "n_fig": 7, "n_tab": 22, "L_tab": 16084, "n_element_tab": 650, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 2015, "n_element_tab_1": 255, "formula_len_all": 789, "formula_len_all_1": 115, "len_all": 263381, "len_all_1": 68920, "len_abs": 511, "len_title": 62, "len_sents": 65004, "len_sents_1": 31193, "n_sents": 571, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1681, "title": "CodeUpdateArena: Benchmarking Knowledge Editing on API Updates", "abs": "Large language models (LLMs) are increasingly being used to synthesize and reason about source code. The libraries and API functions they invoke are continuously evolving, with functionality being added or changing. Yet, no prior work has studied how an LLM's knowledge about code API functions can be updated. To fill this gap, we present `CodeUpdateArena`, a benchmark for knowledge editing in the code domain. An instance in our benchmark consists of a synthetic API function update paired with a program synthesis example that uses the updated functionality; our goal is to update an LLM to be able to solve this program synthesis example without providing documentation of the update at inference time. Compared to knowledge editing for facts, success here is more challenging: a code LLM must reason about the semantics of the modified function rather than just reproduce its syntax. Our dataset is constructed by first prompting GPT-$4$ to generate atomic and executable function updates. Then, for each update, we generate program synthesis examples whose code solutions are prone to use the update. Our benchmark covers updates of various types to 54 functions from seven diverse Python packages, with a total of 670 program synthesis examples. Our experiments show that fine-tuning open-source code LLMs (i.e., DeepSeek, CodeLlama) on documentation of a new update does not allow them to incorporate changes for problem-solving. However, prepending the same information does help, establishing that the information is present, and careful fine-tuning on examples demonstrating the update shows improvement, paving the way for better knowledge editing techniques for code.", "keywords": [ "knowledge editing", "code large language models", "program synthesis" ], "rating_list": [ 8, 6, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 4, 3 ], "contribution_list": [ 3, 3, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 5 ] }, { "paper_id": "ecIvumCyAj", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Anastasis Kratsios", "gender": "Non-Binary", "institution": "Vector Institute", "country": "CA", "position": "Researcher" }, { "name": "Florian Krach", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" }, { "name": "Frank Rudzicz", "gender": "Male", "institution": "Dalhousie University", "country": "CA", "position": "Associate Professor" }, { "name": "Raeid Saqur", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Yannick Limmer", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Blanka N Horvath", "gender": "Female", "institution": "Oxford University", "country": "", "position": "Associate Professor" } ], "n_formula": 84, "n_formula_1": 30, "n_ref_uni": 41, "n_ref": 69, "n_ref_all": 109, "n_fig": 5, "n_tab": 12, "L_tab": 3671, "n_element_tab": 489, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1783, "n_element_tab_1": 267, "formula_len_all": 10469, "formula_len_all_1": 2972, "len_all": 241422, "len_all_1": 75920, "len_abs": 1911, "len_title": 134, "len_sents": 60712, "len_sents_1": 30124, "n_sents": 540, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1951, "title": "Filtered not Mixed: Filtering-Based Online Gating for Mixture of Large Language Models", "abs": "We propose MoE-F \u2014 a formalized mechanism for combining N pre-trained expert Large Language Models (LLMs) in online time-series prediction tasks by adaptively forecasting the best weighting of LLM predictions at every time step. Our mechanism leverages the conditional information in each expert's running performance to forecast the best combination of LLMs for predicting the time series in its next step. Diverging from static (learned) Mixture of Experts (MoE) methods, our approach employs time-adaptive stochastic filtering techniques to combine experts. By framing the expert selection problem as a finite state-space, continuous-time Hidden Markov model (HMM), we can leverage the Wohman-Shiryaev filter. Our approach first constructs N parallel filters corresponding to each of the N individual LLMs. Each filter proposes its best combination of LLMs, given the information that they have access to. Subsequently, the N filter outputs are optimally aggregated to maximize their robust predictive power, and this update is computed efficiently via a closed-form expression, thus generating our ensemble predictor.\n\nOur contributions are:\n\n- **(I)** the MoE-F algorithm \u2014 deployable as a plug-and-play filtering harness,\n\n- **(II)** theoretical optimality guarantees of the proposed filtering-based gating algorithm (via optimality guarantees for its parallel Bayesian filtering and its robust aggregation steps), and\n\n- **(III)** empirical evaluation and ablative results using state-of-the-art foundational and MoE LLMs on a real-world _Financial Market Movement_ task where MoE-F attains a remarkable 17% absolute and 48.5% relative F1 measure improvement over the next best performing individual LLM expert predicting short-horizon market movement based on streaming news. Further, we provide empirical evidence of substantial performance gains in applying MoE-F over specialized models in the _long-horizon time-series forecasting_ domain.", "keywords": [ "Optimal filtering", "LLMs", "mixture-of-experts", "time-series-forecasting", "financial-market-movement" ], "rating_list": [ 6, 3, 8, 6 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 2, 2 ] }, { "paper_id": "ec9hJPn59o", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Zhou Teng", "gender": "Male", "institution": "Chongqing University", "country": "CN", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 46, "n_ref": 49, "n_ref_all": 71, "n_fig": 7, "n_tab": 8, "L_tab": 1376, "n_element_tab": 250, "n_fig_1": 6, "n_tab_1": 8, "L_tab_1": 1376, "n_element_tab_1": 250, "formula_len_all": 789, "formula_len_all_1": 789, "len_all": 108786, "len_all_1": 57065, "len_abs": 1476, "len_title": 100, "len_sents": 23525, "len_sents_1": 22931, "n_sents": 164, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1484, "title": "BiEnhancer: Bi-Level Feature Enhancement in the Dark", "abs": "The remarkable achievements of high-level vision tasks (e.g., object detection, semantic segmentation) under favorable lighting conditions highlight the persistent challenges faced in low-light vision. Previous studies have mainly focused on enhancing low-light images to create visual-friendly representations, often neglecting the differences between machine vision and human vision. This oversight has led to limited performance improvements for high-level tasks. Furthermore, many approaches rely on synthetic paired datasets for training, which can result in limited generalization to real-world images with diverse illumination levels. To address these issues, we propose a new module called BiEnhancer, which is designed to enhance the representation of low-light images by optimizing the loss function of high-level tasks to improve performance. BiEnhancer decomposes low-light images into low-level and high-level components and performs feature enhancement. Then, it adopts an attentional feature fusion strategy and a pixel-wise iterative estimation strategy to effectively enhance and restore the details and semantic information of low-light images and improve the machine-readable representation ability of low-light images. As a versatile plug-in module, BiEnhancer supports end-to-end joint training with diverse high-level tasks. Extensive experimental results demonstrate that the BiEnhancer framework outperforms state-of-the-art methods in both speed and accuracy.", "keywords": [ "Low-light image enhancement; low-light object detection; night-time semantic segmentation" ], "rating_list": [ 5, 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 1, 2, 2 ], "confidence_list": [ 3, 5, 5, 5, 4 ] }, { "paper_id": "ebnyMCM63m", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "William Fleshman", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Benjamin Van Durme", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 48, "n_ref": 89, "n_ref_all": 113, "n_fig": 6, "n_tab": 1, "L_tab": 608, "n_element_tab": 200, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1962, "n_element_tab_1": 215, "formula_len_all": 86, "formula_len_all_1": 86, "len_all": 186917, "len_all_1": 64507, "len_abs": 682, "len_title": 112, "len_sents": 30637, "len_sents_1": 28518, "n_sents": 219, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 684, "title": "RE-Adapt: Reverse Engineered Adaptation of Large Language Models", "abs": "We introduce RE-Adapt, an approach to fine-tuning large language models on new domains without degrading any pre-existing instruction-tuning. We reverse engineer an adapter which isolates what an instruction-tuned model has learned beyond its corresponding pretrained base model. Importantly, this requires no additional data or training. We can then fine-tune the base model on a new domain and readapt it to instruction following with the reverse engineered adapter. RE-Adapt and our low-rank variant LoRE-Adapt both outperform other methods of fine-tuning, across multiple popular LLMs and datasets, even when the models are used in conjunction with retrieval-augmented generation.", "keywords": [ "Large Language Model", "Fine-Tuning", "Instruction-Tuning", "Reverse Engineer", "Adapter", "LoRA", "DoRA", "QA", "LLama-3", "Gemma", "Mistral" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "eb5pkwIB5i", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Ethan Perez", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Felix Jedidja Binder", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Henry Sleight", "gender": "Male", "institution": "Constellation", "country": "US", "position": "Researcher" }, { "name": "James Chua", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "John Hughes", "gender": "Male", "institution": "ML Alignment & Theory Scholars", "country": "GB", "position": "Researcher" }, { "name": "Owain Evans", "gender": "unknown", "institution": "Truthful AI", "country": "US", "position": "Principal Researcher" }, { "name": "Robert Long", "gender": "Male", "institution": "Center for AI Safety", "country": "US", "position": "Researcher" }, { "name": "Miles Andrew Turpin", "gender": "Male", "institution": "Scale AI", "country": "US", "position": "Researcher" }, { "name": "Tomasz Korbak", "gender": "Male", "institution": "UK AI Safety Institute", "country": "GB", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 55, "n_ref": 95, "n_ref_all": 150, "n_fig": 22, "n_tab": 7, "L_tab": 5471, "n_element_tab": 675, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 908, "n_element_tab_1": 30, "formula_len_all": 33, "formula_len_all_1": 0, "len_all": 283538, "len_all_1": 66771, "len_abs": 1684, "len_title": 123, "len_sents": 74785, "len_sents_1": 30546, "n_sents": 684, "n_sents_1": 286, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1525, "title": "Looking Inward: Language Models Can Learn About Themselves by Introspection", "abs": "Humans acquire knowledge by observing the external world, but also by introspection. Introspection gives a person privileged access to their current state of mind (e.g. thoughts and feelings) that are not accessible to external observers. Do LLMs have this introspective capability of privileged access? If they do, this would show that LLMs can acquire knowledge not contained in or inferable from training data.\nWe investigate LLMs predicting properties of their own behavior in hypothetical situations. If a model M1 has this capability, it should outperform a different model M2 in predicting M1's behavior\u2014even if M2 is trained on M1's ground-truth behavior.\nThe idea is that M1 has privileged access to its own behavioral tendencies, and this enables it to predict itself better than M2 (even if M2 is generally stronger).\nIn experiments with GPT-4, GPT-4o, and Llama-3 models, we find that the model M1 outperforms M2 in predicting itself, providing evidence for privileged access. Further experiments and ablations provide additional evidence.\nOur results show that LLMs can offer reliable self-information independent of external data in certain domains. By demonstrating this, we pave the way for further work on introspection in more practical domains, which would have significant implications for model transparency and explainability. However, while we successfully show introspective capabilities in simple tasks, we are unsuccessful on more complex tasks or those requiring out-of-distribution generalization.", "keywords": [ "Introspection", "Large Language Models", "Model awareness", "Self-simulation", "Generalization", "Capability Evaluations", "AI safety" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "eajZpoQkGK", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Bangbang Yang", "gender": "Male", "institution": "ByteDance Inc", "country": "CN", "position": "Researcher" }, { "name": "Chenguo Lin", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Research Intern" }, { "name": "Panwang Pan", "gender": "Male", "institution": "ByteDance", "country": "CN", "position": "Researcher" }, { "name": "Yadong MU", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Zeming Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 33, "n_ref": 177, "n_ref_all": 205, "n_fig": 11, "n_tab": 5, "L_tab": 1999, "n_element_tab": 253, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1737, "n_element_tab_1": 178, "formula_len_all": 581, "formula_len_all_1": 447, "len_all": 186719, "len_all_1": 64477, "len_abs": 1467, "len_title": 133, "len_sents": 36396, "len_sents_1": 29043, "n_sents": 220, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1171, "title": "DiffSplat: Repurposing Image Diffusion Models for Scalable Gaussian Splat Generation", "abs": "Recent advancements in 3D content generation from text or a single image struggle with limited high-quality 3D datasets and inconsistency from 2D multi-view generation. We introduce DiffSplat, a novel 3D generative framework that natively generates 3D Gaussian splats by taming large-scale text-to-image diffusion models. It differs from previous 3D generative models by effectively utilizing web-scale 2D priors while maintaining 3D consistency in a unified model. To bootstrap the training, a lightweight reconstruction model is proposed to instantly produce multi-view Gaussian splat grids for scalable dataset curation. In conjunction with the regular diffusion loss on these grids, a 3D rendering loss is introduced to facilitate 3D coherence across arbitrary views. The compatibility with image diffusion models enables seamless adaptions of numerous techniques for image generation to the 3D realm. Extensive experiments reveal the superiority of DiffSplat in text- and image-conditioned generation tasks and downstream applications. Thorough ablation studies validate the efficacy of each critical design choice and provide insights into the underlying mechanism.", "keywords": [ "3D Generation", "Diffusion models", "3D Gaussian Splatting" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 5, 5, 3, 4 ] }, { "paper_id": "eaTqsptDPL", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Jinwook Jung", "gender": "unknown", "institution": "Hanyang University", "country": "KR", "position": "MS student" }, { "name": "Sungyong Baik", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "Assistant Professor" }, { "name": "Yeoreum Lee", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "MS student" } ], "n_formula": 32, "n_formula_1": 19, "n_ref_uni": 48, "n_ref": 117, "n_ref_all": 169, "n_fig": 8, "n_tab": 11, "L_tab": 2716, "n_element_tab": 374, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 159, "n_element_tab_1": 6, "formula_len_all": 2110, "formula_len_all_1": 1295, "len_all": 174180, "len_all_1": 64190, "len_abs": 1467, "len_title": 129, "len_sents": 54502, "len_sents_1": 31589, "n_sents": 368, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1473, "title": "Mitigating Parameter Interference in Model Merging via Sharpness-Aware Fine-Tuning", "abs": "Large-scale deep learning models with a pretraining-finetuning paradigm have led to a surge of numerous task-specific models finetuned from a common pretrained model. Recently, several research efforts have been made on merging these large models into a single multi-task model, particularly with simple arithmetic on parameters. Such merging methodology faces a central challenge: interference between model parameters finetuned on different tasks. Few recent works have focused on desiging a new finetuning scheme that can lead to small parameter interference, however at the cost of the performance of each task-specific finetuned model and thereby limiting that of a merged model. To improve the performance of a merged model, we note that a finetuning scheme should aim for (1) smaller parameter interference and (2) better performance of each finetuned model on the corresponding task. In this work, we aim to design a new finetuning objective function to work towards these two goals. In the course of this process, we find such objective function to be strikingly similar to sharpness-aware minimization (SAM) objective function, which aims to achieve generalization by finding flat minima. Drawing upon our observation, we propose to finetune pretrained models via SAM or its variants. The experimental and theoretical results showcase the effectiveness and orthogonality of our proposed approach, improving performance upon various merging and finetuning methods.", "keywords": [ "Model merging", "Multi-task Learning", "Sharpness-aware minimization" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "eZLckrDOom", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Johannes Hertrich", "gender": "Male", "institution": "Universit\u00e9 Paris Dauphine - PSL", "country": "FR", "position": "Postdoc" }, { "name": "Robert Gruhlke", "gender": "unknown", "institution": "Freie Universit\u00e4t Berlin", "country": "DE", "position": "Postdoc" } ], "n_formula": 106, "n_formula_1": 31, "n_ref_uni": 46, "n_ref": 115, "n_ref_all": 152, "n_fig": 12, "n_tab": 4, "L_tab": 215, "n_element_tab": 14, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 13623, "formula_len_all_1": 2759, "len_all": 246533, "len_all_1": 70751, "len_abs": 1141, "len_title": 88, "len_sents": 72594, "len_sents_1": 30284, "n_sents": 638, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1144, "title": "Importance Corrected Neural JKO Sampling", "abs": "In order to sample from an unnormalized probability density function, we propose to combine continuous normalizing flows (CNFs) with rejection-resampling steps based on importance weights. We relate the iterative training of CNFs with regularized velocity fields to a JKO scheme and prove convergence of the involved velocity fields to the velocity field of the Wasserstein gradient flow (WGF). The alternation of local flow steps and non-local rejection-resampling steps allows to overcome local minima or slow convergence of the WGF for multimodal distributions. Since the proposal of the rejection step is generated by the model itself, they do not suffer from common drawbacks of classical rejection schemes. The arising model can be trained iteratively, reduces the reverse Kullback-Leibler (KL) loss function in each step, allows to generate iid samples and moreover allows for evaluations of the generated underlying density. Numerical examples show that our method yields accurate results on various test distributions including high-dimensional multimodal targets and outperforms the state of the art in almost all cases significantly.", "keywords": [ "Sampling", "Wasserstein Gradient Flows", "Normalizing Flows", "Rejection Sampling" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "eYePDPSmmu", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Changsheng Lu", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "PhD student" }, { "name": "Hongdong Li", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "Full Professor" }, { "name": "Rong Wang", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "PhD student" }, { "name": "Wei Mao", "gender": "Male", "institution": "Tencent ", "country": "AU", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 10, "n_ref_uni": 44, "n_ref": 138, "n_ref_all": 165, "n_fig": 12, "n_tab": 6, "L_tab": 1037, "n_element_tab": 134, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1228, "n_element_tab_1": 144, "formula_len_all": 543, "formula_len_all_1": 545, "len_all": 146223, "len_all_1": 70562, "len_abs": 1828, "len_title": 113, "len_sents": 42357, "len_sents_1": 32001, "n_sents": 297, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1837, "title": "Skinning-free Accurate 3D Garment Deformation via Image Transfer", "abs": "3D garment animation is key to a wide range of applications including digital humans, virtual try-on, and extended reality. This paper addresses the task of predicting 3D garment deformation from a posed body mesh. Existing learning-based methods mostly rely on linear blend skinning to decompose garment deformation into low-frequency posed garment shape and high-frequency wrinkles. However, due to the lack of explicit skinning supervision, they often produce misaligned garment positions with undesired artifacts during garment re-posing, which corrupt the high-frequency signals. These skinning-based methods consequently fail to recover accurate wrinkle patterns. To tackle this issue, we present a skinning-free approach that re-formulates the high-low frequency decomposition by estimating posed (i) vertex position for low-frequency posed garment shape, and (ii) vertex normal for high-frequency local wrinkle details. In this way, each frequency modality can be effectively decoupled and directly supervised by the geometry of the deformed garment. Moreover, we propose to encode both vertex attributes as texture images, so that 3D garment deformation can be equivalently achieved via 2D image transfer. This enables us to leverage powerful pretrained image encoders to recover high-fidelity visual details representing fine wrinkles. In addition, we model body-garment interaction via cross-attention between dense body and garment image patches, which refines the naive skinning on sparse joints. Finally, we propose a multimodal fusion to incorporate constraints from both frequency modalities and optimize deformed 3D garments from transferred images. Extensive experiments show that our method significantly improves deformation accuracy on various garment types and recovers finer wrinkles than state-of-the-art methods.", "keywords": [ "3D Garment Deformation" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "eYcK7lzlOi", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Bosun Hwang", "gender": "Male", "institution": "Samsung Advanced Institute of Technology", "country": "", "position": "Principal Researcher" }, { "name": "Doyun Kim", "gender": "Male", "institution": "Samsung Advanced Institute of Technology", "country": "KR", "position": "Researcher" }, { "name": "Jaemin Park", "gender": "Male", "institution": "Samsung Advanced Institute of Technology", "country": "KR", "position": "Researcher" }, { "name": "Taejin Paik", "gender": "Male", "institution": "Samsung Advanced Institute of Technology", "country": "KR", "position": "Researcher" }, { "name": "Yoon Hyeok Lee", "gender": "Male", "institution": "Samsung Electronics", "country": "KR", "position": "Researcher" } ], "n_formula": 65, "n_formula_1": 17, "n_ref_uni": 42, "n_ref": 97, "n_ref_all": 160, "n_fig": 3, "n_tab": 11, "L_tab": 5361, "n_element_tab": 568, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2489, "n_element_tab_1": 170, "formula_len_all": 4121, "formula_len_all_1": 1018, "len_all": 213957, "len_all_1": 65891, "len_abs": 1458, "len_title": 107, "len_sents": 58022, "len_sents_1": 27443, "n_sents": 534, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1464, "title": "Unleashing Graph Transformers with Green and Martin Kernels", "abs": "Graph Transformers (GTs) are rapidly emerging as superior models, surpassing traditional message-passing neural networks in graph-level tasks. For optimal performance, it is essential to design GT architectures that embed graph inductive biases and utilize global attention mechanisms through effective structural encodings (SEs). In this work, we introduce novel SEs derived from a rigorous theoretical analysis of random walks (RWs), specifically leveraging the Green and Martin kernels. The Green and Martin kernels are mathematical tools used to observe the long-term behavior of RWs on graphs. By integrating these kernels into the encoding process, we enhance their capability to accurately represent complex graph structures. Our empirical evaluations demonstrate that these approaches enable GTs to achieve state-of-the-art performance on 7 out of 8 benchmark datasets. These include molecular datasets characterized by intricate, non-aperiodic substructures such as benzene rings, and directed acyclic graphs common in the circuit domain. We attribute these performance improvement to the effective capture of the characteristics of non-aperiodic substructures and directed acyclic graphs by our extending encodings. The results not only validate the effectiveness of integrating the Green and Martin kernels into RW-based encodings but also underscore their potential to substantially enhance the learning capabilities of GTs across diverse applications.", "keywords": [ "Graph Transformers", "Graph Neural Networks", "Structural Encodings", "Green Kernel", "Martin Kernel", "Non-aperiodic substructures", "DAGs" ], "rating_list": [ 5, 5, 8, 8 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 5, 3, 3 ] }, { "paper_id": "eY5JNJE56i", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Faguo Wu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "Assistant Professor" }, { "name": "Jianxiang Liu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Qingmao Yao", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "Undergrad student" }, { "name": "Tianyuan Chen", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" }, { "name": "Xiao Zhang", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "Full Professor" }, { "name": "Xuefan Chen", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Zhichao Lei", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Ziyue Yuan", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" } ], "n_formula": 60, "n_formula_1": 32, "n_ref_uni": 46, "n_ref": 108, "n_ref_all": 142, "n_fig": 5, "n_tab": 9, "L_tab": 4117, "n_element_tab": 353, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1413, "n_element_tab_1": 130, "formula_len_all": 8753, "formula_len_all_1": 3715, "len_all": 205441, "len_all_1": 73260, "len_abs": 1188, "len_title": 125, "len_sents": 60298, "len_sents_1": 28869, "n_sents": 516, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1266, "title": "Offline RL with Smooth OOD Generalization in Convex Hull and its Neighborhood", "abs": "Offline Reinforcement Learning (RL) struggles with distributional shifts, leading to the $Q$-value overestimation for out-of-distribution (OOD) actions. Existing methods address this issue by imposing constraints; however, they often become overly conservative when evaluating OOD regions, which constrains the $Q$-function generalization. This over-constraint issue results in poor $Q$-value estimation and hinders policy improvement. In this paper, we introduce a novel approach to achieve better $Q$-value estimation by enhancing $Q$-function generalization in OOD regions within Convex Hull and its Neighborhood (CHN). Under the safety generalization guarantees of the CHN, we propose the Smooth Bellman Operator (SBO), which updates OOD $Q$-values by smoothing them with neighboring in-sample $Q$-values. We theoretically show that SBO approximates true $Q$-values for both in-sample and OOD actions within the CHN. Our practical algorithm, Smooth Q-function OOD Generalization (SQOG), empirically alleviates the over-constraint issue, achieving near-accurate $Q$-value estimation. On the D4RL benchmarks, SQOG outperforms existing state-of-the-art methods in both performance and computational efficiency. Code is available at .", "keywords": [ "Deep Reinforcement Learning", "Offline Reinforcement Learning", "Smooth Bellman Operator", "Smooth Q-function OOD Generalization" ], "rating_list": [ 6, 8, 5, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "eXB5TCrAu9", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Geewook Kim", "gender": "unknown", "institution": "NAVER Cloud", "country": "", "position": "Researcher" }, { "name": "Hoyeon Chang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Hyunji Lee", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Jiyeon Kim", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" }, { "name": "Minjoon Seo", "gender": "Male", "institution": "Twelve Labs", "country": "", "position": "Chief Scientist" }, { "name": "Seongyun Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Sue Hyun Park", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KP", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 42, "n_ref": 106, "n_ref_all": 132, "n_fig": 7, "n_tab": 8, "L_tab": 2420, "n_element_tab": 290, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1434, "n_element_tab_1": 80, "formula_len_all": 66, "formula_len_all_1": 0, "len_all": 251531, "len_all_1": 58982, "len_abs": 1446, "len_title": 128, "len_sents": 78638, "len_sents_1": 28784, "n_sents": 613, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1454, "title": "How Does Vision-Language Adaptation Impact the Safety of Vision Language Models?", "abs": "Vision-Language adaptation (VL adaptation) transforms Large Language Models (LLMs) into Large Vision-Language Models (LVLMs) for multimodal tasks, but this process often compromises the inherent safety capabilities embedded in the original LLMs. Despite potential harmfulness due to weakened safety measures, in-depth analysis on the effects of VL adaptation on safety remains under-explored. This study examines how VL adaptation influences safety and evaluates the impact of safety fine-tuning methods. Our analysis reveals that safety degradation occurs during VL adaptation, even when the training data is safe. While safety tuning techniques like supervised fine-tuning with safety datasets or reinforcement learning from human feedback mitigate some risks, they still lead to safety degradation and a reduction in helpfulness due to over-rejection issues. Further analysis of internal model weights suggests that VL adaptation may impact certain safety-related layers, potentially lowering overall safety levels. Additionally, our findings demonstrate that the objectives of VL adaptation and safety tuning are divergent, which often results in their simultaneous application being suboptimal. To address this, we suggest the weight merging approach as an optimal solution effectively reducing safety degradation while maintaining helpfulness. These insights help guide the development of more reliable and secure LVLMs for real-world applications.", "keywords": [ "Large Vision Language Model", "Safety" ], "rating_list": [ 5, 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 3, 4, 3 ] }, { "paper_id": "eWs2Zxxwwn", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Bin Qin", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Bing Su", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Jiahao Chen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Jiangmeng Li", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Yurou Liu", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 10, "n_ref_uni": 44, "n_ref": 83, "n_ref_all": 94, "n_fig": 5, "n_tab": 4, "L_tab": 447, "n_element_tab": 39, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1379, "n_element_tab_1": 52, "formula_len_all": 1505, "formula_len_all_1": 681, "len_all": 143721, "len_all_1": 62166, "len_abs": 1659, "len_title": 156, "len_sents": 42624, "len_sents_1": 28015, "n_sents": 351, "n_sents_1": 234, "venue": "ICLR2025", "decision": "withdraw", "L_title": 106, "L_abs": 1667, "title": "Integrating the Expression and Discrimination via Bilateral Compensation for Molecular Property Prediction", "abs": "Predicting molecular properties plays an important role in both scientific research and industrial applications. Given that different molecular properties are influenced by specific atoms or functional groups, it is essential to incorporate both types of information. Previous approaches either leverage subgraph information in self-supervised learning to pre-train atom-based architectures or develop subgraph-based architectures tailored to specific downstream tasks. However, these methods often lack a thorough analysis or theoretical support concerning the expressive capabilities of these two types of representations. Moreover, they typically rely on fixed coupling representations, which cannot adaptively prioritize more discriminative information for various downstream tasks.\nIn this paper, we introduce a Route-guided Bilateral Compensation (RBC) architecture that explicitly extracts atom-wise and subgraph-wise information through two decoupled branches and integrates them via a route module. Theoretically, we demonstrate that our decomposition-polymerization subgraph-wise branch exhibits greater expressive power than the atom-wise branch, and that the integration process reduces the generalization error bound. Furthermore, we propose a coordinated self-supervised learning strategy that incorporates node-level masked graph reconstruction tasks for atomic and lexicalized subgraph tokens, alongside a graph-level contrastive learning task. For different downstream tasks, the route module facilitates dynamic integration, enhancing the discriminative power of the final representation. External experiments verify the effectiveness of our method.", "keywords": [ "Molecular property prediction", "Self-supervised learning" ], "rating_list": [ 5, 5, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 5, 3 ] }, { "paper_id": "eWocmTQn7H", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Guojin Zhong", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "MS student" }, { "name": "Jin Yuan", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Associate Professor" }, { "name": "Long Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Zhiyong Li", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Full Professor" }, { "name": "pan wang", "gender": "unknown", "institution": "Hunan University", "country": "", "position": "MS student" } ], "n_formula": 49, "n_formula_1": 13, "n_ref_uni": 42, "n_ref": 115, "n_ref_all": 138, "n_fig": 10, "n_tab": 17, "L_tab": 7045, "n_element_tab": 499, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1988, "n_element_tab_1": 190, "formula_len_all": 5426, "formula_len_all_1": 1317, "len_all": 209841, "len_all_1": 65922, "len_abs": 1688, "len_title": 142, "len_sents": 66176, "len_sents_1": 29250, "n_sents": 453, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1710, "title": "Multi-Resolution Decomposable Diffusion Model for Non-Stationary Time Series Anomaly Detection", "abs": "Recently, generative models have shown considerable promise in unsupervised time series anomaly detection. Nonetheless, the task of effectively capturing complex temporal patterns and minimizing false alarms becomes increasingly challenging when dealing with non-stationary time series, characterized by continuously fluctuating statistical attributes and joint distributions. To confront these challenges, we underscore the benefits of multi-resolution modeling, which improves the ability to distinguish between anomalies and non-stationary behaviors by leveraging correlations across various resolution scales. In response, we introduce a **M**ulti-Res**o**lution **De**composable Diffusion **M**odel (MODEM), which integrates a coarse-to-fine diffusion paradigm with a frequency-enhanced decomposable network to adeptly navigate the intricacies of non-stationarity. Technically, the coarse-to-fine diffusion model embeds cross-resolution correlations into the forward process to optimize diffusion transitions mathematically. It then innovatively employs low-resolution recovery to guide the reverse trajectories of high-resolution series in a coarse-to-fine manner, enhancing the model's ability to learn and elucidate underlying temporal patterns. Furthermore, the frequency-enhanced decomposable network operates in the frequency domain to extract globally shared time-invariant information and time-variant temporal dynamics for accurate series reconstruction. Extensive experiments conducted across five real-world datasets demonstrate that our proposed MODEM achieves state-of-the-art performance and can be generalized to other time series tasks. The code will be publicly available upon acceptance.", "keywords": [ "Diffusion Model", "Non-Stationary Time Series", "Anomaly Detection", "Multi-Resolution" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "eWc76Kyi8H", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Jing Bai", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Junzhe Chen", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Tianlong Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Xueting Han", "gender": "unknown", "institution": "Microsoft Research Asia", "country": "CN", "position": "Researcher" } ], "n_formula": 22, "n_formula_1": 24, "n_ref_uni": 35, "n_ref": 75, "n_ref_all": 104, "n_fig": 4, "n_tab": 4, "L_tab": 4248, "n_element_tab": 104, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 511, "n_element_tab_1": 26, "formula_len_all": 1464, "formula_len_all_1": 1470, "len_all": 149871, "len_all_1": 62495, "len_abs": 1625, "len_title": 125, "len_sents": 42717, "len_sents_1": 27068, "n_sents": 336, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1596, "title": "CPL: Critical Plan Step Learning Boosts LLM Generalization in Reasoning Tasks", "abs": "Post-training, particularly reinforcement learning (RL) using self-play-generated data, has become a new learning paradigm for large language models (LLMs). However, scaling RL to develop a general reasoner remains a research challenge, as existing methods focus on task-specific reasoning without adequately addressing generalization across a broader range of tasks. Moreover, unlike traditional RL with limited action space, LLMs operate in an infinite space, making it crucial to search for valuable and diverse strategies to solve problems effectively.\nTo address this, we propose searching within the action space on high-level abstract plans to enhance model generalization and introduce Critical Plan Step Learning (CPL), comprising: 1) searching on plan, using Monte Carlo Tree Search (MCTS) to explore diverse plan steps in multi-step reasoning tasks, and 2) learning critical plan steps through Step-level Advantage Preference Optimization (Step-APO), which integrates advantage estimates for step preference obtained via MCTS into Direct Preference Optimization (DPO). This combination helps the model effectively learn critical plan steps, enhancing both reasoning capabilities and generalization.\nExperimental results demonstrate that our method, trained exclusively on GSM8K and MATH, not only significantly improves performance on GSM8K (+10.5\\%) and MATH (+6.5\\%), but also enhances out-of-domain reasoning benchmarks, such as HumanEval (+12.2\\%), GPQA (+8.6\\%), ARC-C (+4.0\\%), MMLU-STEM (+2.2\\%), and BBH (+1.8\\%). The code is available at https://anonymous.4open.science/r/CPL.", "keywords": [ "LLM Reasoning", "Monte-Carlo Tree Search", "Reinforcement Learning", "Generalization" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "eWNEqdH0vk", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Ivan Titov", "gender": "unknown", "institution": "University of Edinburgh", "country": "", "position": "Full Professor" }, { "name": "Shuang Cheng", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Yizhi Zhou", "gender": "unknown", "institution": "nanjing university", "country": "", "position": "Undergrad student" }, { "name": "Zeyu Huang", "gender": "unknown", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Zihan Qiu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zili Wang", "gender": "Male", "institution": "stepfun", "country": "CN", "position": "Researcher" }, { "name": "Jie Fu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 56, "n_ref": 90, "n_ref_all": 100, "n_fig": 3, "n_tab": 15, "L_tab": 5491, "n_element_tab": 489, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 3236, "n_element_tab_1": 208, "formula_len_all": 412, "formula_len_all_1": 292, "len_all": 183679, "len_all_1": 69154, "len_abs": 1559, "len_title": 97, "len_sents": 46280, "len_sents_1": 30181, "n_sents": 369, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1537, "title": "Layerwise Recurrent Router for Mixture-of-Experts", "abs": "The scaling of large language models (LLMs) has revolutionized their capabilities in various tasks, yet this growth must be matched with efficient computational strategies. \nThe Mixture-of-Experts (MoE) architecture stands out for its ability to scale model size without significantly increasing training costs. \nDespite their advantages, current MoE models often display parameter inefficiency. \nFor instance, a pre-trained MoE-based LLM with 52 billion parameters might perform comparably to a standard model with 6.7 billion. \nBeing a crucial part of MoE, \ncurrent routers in different layers independently assign tokens without leveraging historical routing information, potentially leading to suboptimal token-expert combinations and the parameter inefficiency problem.\nTo alleviate this issue, we introduce the Layerwise Recurrent Router for Mixture-of-Experts (RMoE). \nRMoE leverages a Gated Recurrent Unit (GRU) to establish dependencies between routing decisions across consecutive layers.\nSuch layerwise recurrence can be efficiently parallelly computed for input tokens and introduces negotiable costs.\nOur extensive empirical evaluations demonstrate that RMoE-based language models consistently outperform a spectrum of baseline models. \nFurthermore, RMoE integrates a novel computation stage orthogonal to existing methods, allowing seamless compatibility with other MoE architectures. \nOur analyses attribute RMoE's gains to its effective cross-layer information sharing, which also improves expert selection and diversity.", "keywords": [ "Mixture of Expert", "Router", "LLMs" ], "rating_list": [ 8, 6, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "eWFkMCBySw", "primary_area": "applications to robotics, autonomy, planning", "n_author": 8, "author_info_list": [ { "name": "Dong An", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Ian Reid", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" }, { "name": "Liang Wang", "gender": "Male", "institution": "Institute of Automation\uff0c CAS\uff0cChina", "country": "CN", "position": "Full Professor" }, { "name": "Rongtao Xu", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yan Huang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Yifei Su", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Yonggen Ling", "gender": "Male", "institution": "Tencent Robotics X", "country": "CN", "position": "Researcher" }, { "name": "kehan chen", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 40, "n_ref": 96, "n_ref_all": 126, "n_fig": 10, "n_tab": 7, "L_tab": 1501, "n_element_tab": 171, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1137, "n_element_tab_1": 121, "formula_len_all": 288, "formula_len_all_1": 287, "len_all": 161278, "len_all_1": 63668, "len_abs": 1131, "len_title": 128, "len_sents": 44006, "len_sents_1": 28849, "n_sents": 346, "n_sents_1": 223, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1138, "title": "CONSTRAINT-AWARE ZERO-SHOT VISION-LANGUAGE NAVIGATION IN CONTINUOUS ENVIRONMENTS", "abs": "We present Constraint-Aware Navigator (CA-Nav), a zero-shot approach for Vision-Language Navigation in Continuous Environments (VLN-CE).\nCA-Nav reframes the zero-shot VLN-CE task as a sequential constraint-aware sub-instruction completion process, continuously translating sub-instructions into navigation plans via a cross-modal value map. \nCentral to our approach are two modules namely Constraint-aware Sub-instruction Manager (CSM) and Constraint-aware Value Mapper (CVM).\nCSM defines the completion criteria of decomposed sub-instructions as constraints and tracks navigation progress by switching sub-instructions in a constraint-aware manner.\nBased on the constraints identified\nby CSM, CVM builds a value map on-the-fly and refines it using superpixel clustering to enhance navigation stability.\nCA-Nav achieves the state-of-the-art performance on two VLN-CE benchmarks, surpassing the compared best method by 12\\% on R2R-CE and 13\\% on RxR-CE in terms of Success Rate on the validation unseen split.\nFurthermore, CA-Nav demonstrates its effectiveness in real-world robot deployments across diverse indoor scenes and instructions.", "keywords": [ "Instruction Navigation", "training free", "unexplored environment" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "eW4yh6HKz4", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 11, "author_info_list": [ { "name": "Baoqun Yin", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Hanting Chen", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Jie Hu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Wei Li", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Xiaoyu Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yehui Tang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yun Zhang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Yunhe Wang", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhijun Tu", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Zhiwei Xiong", "gender": "Male", "institution": "USTC", "country": "CN", "position": "Professor" }, { "name": "Xinding", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 26, "n_ref": 54, "n_ref_all": 78, "n_fig": 3, "n_tab": 19, "L_tab": 9155, "n_element_tab": 996, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 4478, "n_element_tab_1": 446, "formula_len_all": 812, "formula_len_all_1": 809, "len_all": 180922, "len_all_1": 67331, "len_abs": 1542, "len_title": 104, "len_sents": 38944, "len_sents_1": 28041, "n_sents": 267, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 55, "L_abs": 1552, "title": "CBQ: Cross-Block Quantization for Large Language Models", "abs": "Post-training quantization (PTQ) has played a pivotal role in compressing large language models (LLMs) at ultra-low costs. Although current PTQ methods have achieved promising results by addressing outliers and employing layer- or block-wise loss optimization techniques, they still suffer from significant performance degradation at ultra-low bits precision. To dissect this issue, we conducted an in-depth analysis of quantization errors specific to LLMs and surprisingly discovered that, unlike traditional sources of quantization errors, the growing number of model parameters, combined with the reduction in quantization bits, intensifies inter-layer and intra-layer dependencies, which severely impact quantization accuracy. This finding highlights a critical challenge in quantizing LLMs. To address this, we propose CBQ, a cross-block reconstruction-based PTQ method for LLMs. CBQ leverages a cross-block dependency to establish long-range dependencies across multiple blocks and integrates an adaptive LoRA-Rounding technique to manage intra-layer dependencies. To further enhance performance, CBQ incorporates a coarse-to-fine pre-processing mechanism for processing weights and activations. Extensive experiments show that CBQ achieves superior low-bit quantization (W4A4, W4A8, W2A16) and outperforms existing state-of-the-art methods across various LLMs and datasets. Notably, CBQ only takes 4.3 hours to quantize a weight-only quantization of a 4-bit LLAMA1-65B model, achieving a commendable trade off between performance and efficiency.", "keywords": [ "Large Language Model Compression", "ultra-low bits precision" ], "rating_list": [ 6, 6, 8, 10, 8 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 4, 3 ], "confidence_list": [ 5, 4, 4, 5, 3 ] }, { "paper_id": "eVsSjNRuAp", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Dan Wilson", "gender": "Male", "institution": "University of Tennessee, Knoxville", "country": "US", "position": "Assistant Professor" }, { "name": "Fanqi Wang", "gender": "Female", "institution": "University of Tennessee, Knoxville", "country": "", "position": "PhD student" }, { "name": "Hairong Qi", "gender": "Female", "institution": "University of Tennessee, Knoxville", "country": "US", "position": "Full Professor" }, { "name": "Igor Mezic", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "", "position": "Full Professor" }, { "name": "Landon Harris", "gender": "Male", "institution": "University of Tennessee, Knoxville", "country": "", "position": "MS student" }, { "name": "Weisheng Tang", "gender": "Male", "institution": "University of Tennessee, Knoxville", "country": "", "position": "Postdoc" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 25, "n_ref": 58, "n_ref_all": 83, "n_fig": 12, "n_tab": 5, "L_tab": 3921, "n_element_tab": 224, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 278, "n_element_tab_1": 30, "formula_len_all": 509, "formula_len_all_1": 510, "len_all": 110391, "len_all_1": 51097, "len_abs": 1886, "len_title": 108, "len_sents": 36488, "len_sents_1": 24233, "n_sents": 295, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1897, "title": "Predictive Differential Training Guided by Training Dynamics", "abs": "This paper centers around a novel concept proposed recently by researchers from the control community where the training process of a deep neural network can be considered a nonlinear dynamical system acting upon the high-dimensional weight space. Koopman operator theory, a data-driven dynamical system analysis framework, can then be deployed to discover the otherwise non-intuitive training dynamics. Taking advantage of the predictive power of the Koopman operator theory, the time-consuming Stochastic Gradient Descent ( SGD) iterations can be bypassed by directly predicting network weights a few epochs later. This novel predictive training framework, however, often suffers from gradient explosion especially for more extensive and complex models. In this paper, we incorporate the idea of differential learning, where different parts of the network can undergo different learning rates during training, into the predictive training framework and propose the so-called \"predictive differential training'' (PDT) to sustain robust performance for accelerated learning even for complex network structures. The key contribution is the design of an effective masking strategy based on Koopman analysis of training dynamics of each parameter in order to select the subset of parameters that exhibits \"good'' prediction performance. PDT also includes the design of an acceleration scheduler to keep track of the prediction error so that the training process can roll back to the traditional GD-based approaches to \"correct'' deviations from off-predictions. We demonstrate that PDT can be seamlessly integrated as a plug-in with existing optimizers, including, for example, SGD, momentum, and Adam. The experimental results have shown consistent performance improvement in terms of faster convergence, lower training/testing loss, and fewer number of epochs to achieve the best loss of Baseline.", "keywords": [ "Training Dynamics", "Koopman Operator Theory", "Predictive Training", "Deep Neural Networks" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "eVKP64sQBd", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Chenyu You", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Assistant Professor" }, { "name": "James s Duncan", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Professor" }, { "name": "Lawrence Hamilton Staib", "gender": "Not Specified", "institution": "Yale University", "country": "US", "position": "Full Professor" }, { "name": "Tiansheng Wen", "gender": "Male", "institution": "Xidian University ", "country": "CN", "position": "MS student" }, { "name": "Weicheng Dai", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "PhD student" }, { "name": "Yifei Min", "gender": "Male", "institution": "Two Sigma", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 66, "n_ref": 150, "n_ref_all": 171, "n_fig": 7, "n_tab": 2, "L_tab": 954, "n_element_tab": 66, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1132, "n_element_tab_1": 69, "formula_len_all": 378, "formula_len_all_1": 378, "len_all": 156016, "len_all_1": 74659, "len_abs": 1765, "len_title": 139, "len_sents": 39356, "len_sents_1": 36003, "n_sents": 280, "n_sents_1": 240, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1786, "title": "Robust Multi-modal Learning with Shifted Feature Reweighting Against Spurious Correlations", "abs": "Pre-trained multi-modal models have recently garnered significant attention due to their adaptability to diverse downstream tasks via fine-tuning. \nHowever, their resilience to certain group shift issues, i.e., spurious correlations, remains imperative yet relatively under-investigated.\nWe study this problem in vision-language models (VLMs), and we observe potential vulnerabilities in pre-trained VLMs, such as CLIP, when confronted with spurious correlations. \nWhile recent studies have been exploited to address unimodal group-imbalances by minority group up-sampling or creating group-balanced subsets, we posit that true robustness can be achieved by debiasing the training process through feature reweighting. \nIn this paper, we propose Shifted Feature Reweighting (SFR), a robust multi-modal learning method to mitigate the reliance on spurious features. \nSpecifically, we introduce a novel disagreement-based importance weight that allocates distinct weights to individual instances within the training data. \nThis contrasts with existing group rebalance weight strategies, which uniformly weigh all instances within a group. \nOur reweighting strategy adeptly addresses disparities in instance-level learning difficulty. \nMoreover, our empirical results unveil that representation collapse may arise during fine-tuning. \nTo address this, we proposed to introduce feature dropout and show that this simple method can further regularize the training on the majority groups and encourage the training on the minority groups. \nEmpirical results on multiple benchmarks verify our claims and confirm the effectiveness of our proposed SFR. \nTheoretically, we analyze the performance of our SFR and confirm its superiority in mitigating spurious correlations. \nOur codes will be here.", "keywords": [ "spurious correlations; multi-modal learning; shortcut learning;" ], "rating_list": [ 5, 6, 5, 3, 1 ], "soundness_list": [ 3, 3, 2, 2, 1 ], "presentation_list": [ 1, 3, 2, 3, 1 ], "contribution_list": [ 2, 3, 2, 1, 1 ], "confidence_list": [ 2, 4, 4, 4, 5 ] }, { "paper_id": "eUkbTUsDgs", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Ben Slater", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Jonathan Prunty", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "Konstantinos Voudouris", "gender": "Male", "institution": "Helmholtz Zentrum M\u00fcnchen", "country": "DE", "position": "Postdoc" }, { "name": "Lucy G Cheke", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Associate Professor" }, { "name": "Marko Tesic", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "Matteo G. Mecattaf", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 68, "n_ref": 110, "n_ref_all": 124, "n_fig": 4, "n_tab": 1, "L_tab": 5, "n_element_tab": 1, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 194737, "len_all_1": 64721, "len_abs": 1894, "len_title": 135, "len_sents": 53039, "len_sents_1": 34171, "n_sents": 379, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 134, "L_abs": 1913, "title": "A little less conversation, a little more action, please: Investigating the physical common-sense of LLMs in a 3D embodied environment", "abs": "As general-purpose tools, Large Language Models (LLMs) must often reason about everyday physical environments. In a question-and-answer capacity, understanding the interactions of physical objects may be necessary to give appropriate responses. Moreover, LLMs are increasingly used as reasoning engines in agentic systems, designing and controlling their action sequences. The vast majority of research has tackled this issue using static benchmarks, comprised of text or image-based questions about the physical world. However, these benchmarks do not capture the complexity and nuance of real-life physical processes. Here we advocate for a second, relatively unexplored, approach:~`embodying' the LLMs by granting them control of an agent within a 3D environment. We present the first embodied and cognitively meaningful evaluation of physical common-sense reasoning in LLMs. Our framework allows direct comparison of LLMs with other embodied agents, such as those based on Deep Reinforcement Learning, and human and non-human animals. We employ the Animal-AI (AAI) environment, a simulated 3D \\textit{virtual laboratory}, to study physical common-sense reasoning in LLMs. For this, we use the AAI Testbed, a suite of experiments that replicate laboratory studies with non-human animals, to study physical reasoning capabilities including distance estimation, tracking out-of-sight objects, and tool use. We demonstrate that state-of-the-art multi-modal models with no finetuning can complete this style of task, allowing meaningful comparison to the entrants of the 2019 Animal-AI Olympics competition and to human children. Our results show that LLMs are currently outperformed by human children on these tasks. We argue that this approach allows the study of physical reasoning using ecologically valid experiments drawn directly from cognitive science, improving the predictability and reliability of LLMs.", "keywords": [ "LLM Agents", "Animal Cognition", "Cognitive Science", "Evaluation" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 3, 1, 2 ], "contribution_list": [ 3, 1, 2 ], "confidence_list": [ 4, 2, 3 ] }, { "paper_id": "eUEMjwh5wK", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Arunesh Sinha", "gender": "Male", "institution": "Rutgers University", "country": "", "position": "Assistant Professor" }, { "name": "Pradeep Varakantham", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Full Professor" }, { "name": "Roman Belaire", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "PhD student" } ], "n_formula": 44, "n_formula_1": 16, "n_ref_uni": 53, "n_ref": 137, "n_ref_all": 165, "n_fig": 4, "n_tab": 11, "L_tab": 5502, "n_element_tab": 550, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2996, "n_element_tab_1": 207, "formula_len_all": 7169, "formula_len_all_1": 1415, "len_all": 191338, "len_all_1": 68320, "len_abs": 2329, "len_title": 113, "len_sents": 54097, "len_sents_1": 30810, "n_sents": 406, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1572, "title": "On Minimizing Adversarial Counterfactual Error in Adversarial Reinforcement Learning", "abs": "Deep Reinforcement Learning (DRL) policies are critically vulnerable to adversarial noise in observations, posing severe risks in safety-critical scenarios. For example, a self-driving car receiving manipulated sensory inputs about traffic signs could lead to catastrophic outcomes. Existing strategies to fortify RL algorithms against such adversarial perturbations generally fall into two categories: (a) using regularization methods that enhance robustness by incorporating adversarial loss terms into the value objectives, and (b) adopting \"maximin\" principles, which focus on maximizing the minimum value to ensure robustness. While regularization methods reduce the likelihood of successful attacks, their effectiveness drops significantly if an attack does succeed. On the other hand, maximin objectives, although robust, tend to be overly conservative. To address this challenge, we introduce a novel objective called Adversarial Counterfactual Error (ACoE), which naturally balances optimizing value and robustness against adversarial attacks. To optimize ACoE in a scalable manner in model-free settings, we propose a theoretically justified surrogate objective known as Cumulative-ACoE (C-ACoE). The core idea of optimizing C-ACoE is utilizing the belief about the underlying true state given the adversarially perturbed observation. Our empirical evaluations demonstrate that our method outperforms current state-of-the-art approaches for addressing adversarial RL problems across all established benchmarks (MuJoCo, Atari, and Highway) used in the literature.", "keywords": [ "Reinforcement learning", "robust reinforcement learning", "adversarial robustness", "partially observable markov decision problems" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "eU39PDsZtT", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Tao Feng", "gender": "Male", "institution": "Department of Computer Science, University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Yanzhen Shen", "gender": "Male", "institution": "Department of Computer Science", "country": "", "position": "Undergrad student" }, { "name": "Jiaxuan You", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 45, "n_ref": 80, "n_ref_all": 129, "n_fig": 7, "n_tab": 12, "L_tab": 1838, "n_element_tab": 233, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1274, "n_element_tab_1": 164, "formula_len_all": 846, "formula_len_all_1": 922, "len_all": 164970, "len_all_1": 70989, "len_abs": 1777, "len_title": 108, "len_sents": 50015, "len_sents_1": 31314, "n_sents": 382, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1786, "title": "GraphRouter: A Graph-based Router for LLM Selections", "abs": "The rapidly growing number and variety of Large Language Models (LLMs)\npresent significant challenges in efficiently selecting the appropriate LLM for\na given query, especially considering the trade-offs between performance and\ncomputational cost. Current LLM selection methods often struggle to generalize\nacross new LLMs and different tasks because of their limited ability to leverage\ncontextual interactions among tasks, queries, and LLMs, as well as their depen-\ndence on a transductive learning framework. To address these shortcomings, we\nintroduce a novel inductive graph framework, named as GraphRouter, which\nfully utilizes the contextual information among tasks, queries, and LLMs to en-\nhance the LLM selection process. GraphRouter constructs a heterogeneous\ngraph comprising task, query, and LLM nodes, with interactions represented as\nedges, which efficiently captures the contextual information between the query\u2019s\nrequirements and the LLM\u2019s capabilities. Through an innovative edge prediction\nmechanism, GraphRouter is able to predict attributes (the effect and cost of\nLLM response) of potential edges, allowing for optimized recommendations that\nadapt to both existing and newly introduced LLMs without requiring retraining.\nComprehensive experiments across three distinct effect-cost weight scenarios have\nshown that GraphRouter substantially surpasses existing routers, delivering a\nminimum performance improvement of 12.3%. In addition, it achieves enhanced\ngeneralization across new LLMs settings and supports diverse tasks with at least a\n9.5% boost in effect and a significant reduction in computational demands. This\nwork endeavors to apply a graph-based approach for the contextual and adaptive\nselection of LLMs, offering insights for real-world applications.", "keywords": [ "LLM selection", "Graph-based router", "Contextual interactions", "New LLM settings" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "eTWRCiMQ1z", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 5, "author_info_list": [ { "name": "Anne D. Koelewijn", "gender": "unknown", "institution": "Friedrich-Alexander Universit\u00e4t Erlangen-N\u00fcrnberg", "country": "DE", "position": "Assistant Professor" }, { "name": "Eva Dorschky", "gender": "Female", "institution": "fiveD GmbH", "country": "DE", "position": "Researcher" }, { "name": "J\u00f6rg Miehling", "gender": "Male", "institution": "Friedrich-Alexander Universit\u00e4t Erlangen-N\u00fcrnberg", "country": "DE", "position": "Principal Researcher" }, { "name": "Marcel Sch\u00f6ckel", "gender": "Male", "institution": "Friedrich-Alexander Universit\u00e4t Erlangen-N\u00fcrnberg", "country": "DE", "position": "Undergrad student" }, { "name": "Markus Gambietz", "gender": "unknown", "institution": "Friedrich-Alexander Universit\u00e4t Erlangen-N\u00fcrnberg", "country": "DE", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 4, "n_ref_uni": 43, "n_ref": 91, "n_ref_all": 107, "n_fig": 7, "n_tab": 10, "L_tab": 2573, "n_element_tab": 301, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 687, "n_element_tab_1": 76, "formula_len_all": 678, "formula_len_all_1": 233, "len_all": 150905, "len_all_1": 69326, "len_abs": 1667, "len_title": 113, "len_sents": 46387, "len_sents_1": 33715, "n_sents": 362, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1677, "title": "A Self-Supervised PINN for Inertial Pose and Dynamics Estimations", "abs": "Accurate real-time monitoring of not only movements, but also internal joint moments or muscle forces that cause movement in unrestricted environments is key for many clinical and sports applications. A minimally obstrusive way to monitor movements is with wearable sensors, such as inertial measurement units, using the fewest sensors possible. \nCurrent real-time methods rely on supervised learning, where a ground truth dataset needs to be measured with laboratory measurement systems, such as optical motion capture, which then needs to be processed with methods that are known to introduce errors. There is a discrepancy between laboratory and real-world movements, and for analysing new motions, new ground truth data would need to be recorded, which is impractical.\nTherefore, we introduce SSPINNpose, a self-supervised physics-informed neural network that estimates movement dynamics, including joint angles and joint moments, from inertial sensors without the need for ground truth data for training.\nWe run the network output through a physics model of the human body to optimize physical plausibility and generate virtual measurement data. Using this virtual sensor data, the network is trained directly on the measured sensor data instead of a ground truth. Experiments show that SSPINNpose is able to accurately estimate joint angles and joint moments at 8.7 degrees and 4.9 BWBH%, respectively, for walking and running at up to speeds of 4.9 m/s at a latency of 3.5 ms. We further show the versatility of our method by estimating movement dynamics for a variety of sparse sensor configurations and inferring the positions where the sensors are placed on the body.", "keywords": [ "self-supervised learning", "biomechanics", "physics-informed neural networks" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 5, 1 ] }, { "paper_id": "eT6zYrd1wl", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Adarsh Jamadandi", "gender": "Male", "institution": "Saarland University", "country": "", "position": "MS student" }, { "name": "Celia Rubio-Madrigal", "gender": "Female", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "PhD student" }, { "name": "Rebekka Burkholz", "gender": "Female", "institution": "Helmholtz Center CISPA for Information Security", "country": "DE", "position": "Associate Professor" } ], "n_formula": 22, "n_formula_1": 15, "n_ref_uni": 53, "n_ref": 85, "n_ref_all": 113, "n_fig": 8, "n_tab": 11, "L_tab": 7214, "n_element_tab": 505, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2142, "formula_len_all_1": 1160, "len_all": 174140, "len_all_1": 65543, "len_abs": 960, "len_title": 89, "len_sents": 44152, "len_sents_1": 30355, "n_sents": 338, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 41, "L_abs": 964, "title": "SoLAR: Surrogate Label Aware GNN Rewiring", "abs": "Rewiring the input graph of graph neural networks (GNNs) has been proposed as a pre-processing step to address issues like over-squashing and over-smoothing. However, most existing techniques rely solely on topology-based modifications, neglecting performance-critical node label information. To fill this gap, we propose SoLAR (Surrogate Label Aware Rewiring), a method that rewires the graph based on predicted node labels from a surrogate model. We prove its effectiveness in a theoretically tractable setting highlighting two key mechanisms that enable its success. The first is a denoising effect, while the second is a novel knowledge distillation-inspired process, where information from a surrogate model is encoded into the graph structure. Extensive experiments demonstrate consistent improvements of SoLAR across various datasets. Notably, the best surrogate models arise from iterative SoLAR, and reusing the same model class is a competitive strategy.", "keywords": [ "graph neural networks", "graph rewiring", "homophily", "heterophily", "knowledge distillation" ], "rating_list": [ 3, 6, 3, 5, 5 ], "soundness_list": [ 2, 3, 1, 2, 3 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 1, 3, 1, 2, 2 ], "confidence_list": [ 3, 4, 3, 4, 3 ] }, { "paper_id": "eRkNNQRppH", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Daanish Shabbir", "gender": "unknown", "institution": "nncollective", "country": "", "position": "Researcher" }, { "name": "Isabelle Lee", "gender": "Not Specified", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Dani Yogatama", "gender": "unknown", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Nina Rimsky", "gender": "Female", "institution": "Anthropic", "country": "GB", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 35, "n_ref": 43, "n_ref_all": 77, "n_fig": 10, "n_tab": 6, "L_tab": 276, "n_element_tab": 20, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2363, "formula_len_all_1": 520, "len_all": 126546, "len_all_1": 55843, "len_abs": 914, "len_title": 118, "len_sents": 35069, "len_sents_1": 27378, "n_sents": 249, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 916, "title": "(Pre-)training Dynamics: Scaling Generalization with First-Order Logic", "abs": "Transformer-based models have demonstrated a remarkable capacity for learning complex nonlinear relationships. While previous research on generalization dynamics has primarily focused on small transformers (1-2 layers) and simple tasks like XOR and modular addition, we extend this investigation to larger models with 125M parameters, trained on a more sophisticated first-order logic (FOL) task. We introduce a novel FOL dataset that allows us to systematically explore generalization across varying levels of complexity. Our analysis of the pretraining dynamics reveals a series of distinct phase transitions corresponding to the hierarchical generalization of increasingly complex operators and rule sets within the FOL framework. Our task and model establish a testbed for investigating pretraining dynamics at scale, offering a foundation for future research on the learning trajectories of advanced AI systems.", "keywords": [ "phase transitions", "pretraining dynamics", "generalization", "interpretability" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 1, 1, 1, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 1, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "eRduvBHLQ1", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 12, "author_info_list": [ { "name": "Changping Peng", "gender": "Male", "institution": "JD.COM", "country": "", "position": "Researcher" }, { "name": "Ching Law", "gender": "Male", "institution": "JD.com", "country": "CN", "position": "Researcher" }, { "name": "Chun Gan", "gender": "Male", "institution": "JD.com", "country": "CN", "position": "Researcher" }, { "name": "Congying Han", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Fengxiang He", "gender": "unknown", "institution": "University of Edinburgh", "country": "GB", "position": "Lecturer" }, { "name": "Jie He", "gender": "Male", "institution": "Peking University", "country": "", "position": "Undergrad student" }, { "name": "Tiande Guo", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Wenlong Chen", "gender": "Male", "institution": "JD.com", "country": "", "position": "Researcher" }, { "name": "Yudong Hu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zhangang Lin", "gender": "Male", "institution": "JD", "country": "CN", "position": "Researcher" }, { "name": "hanminwang", "gender": "Male", "institution": "JD.com", "country": "CN", "position": "Researcher" }, { "name": "WangHaoran", "gender": "Male", "institution": "JD.com", "country": "CN", "position": "Researcher" } ], "n_formula": 66, "n_formula_1": 24, "n_ref_uni": 32, "n_ref": 52, "n_ref_all": 80, "n_fig": 8, "n_tab": 3, "L_tab": 1203, "n_element_tab": 77, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9792, "formula_len_all_1": 2808, "len_all": 165851, "len_all_1": 60522, "len_abs": 1546, "len_title": 83, "len_sents": 50775, "len_sents_1": 27434, "n_sents": 425, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 35, "L_abs": 1550, "title": "Online Auction for Ads and Organics", "abs": "This paper introduces the first online blending auction mechanism design for sponsored items (ads) alongside organic items (organics), ensuring guaranteed Pareto optimality for platform revenue, advertiser utilities, and user interest (measured through clicks). We innovatively define an umbrella term, \"traffic item,\" to encompass both organics and auctionable ad items, where an organic represents a unit of traffic to be auctioned, valued positively by attracting user interest with a fixed zero bid and payment. The online blending traffic distribution problem is thus transformed into an auction problem with unified valuation metric for the traffic item, which is subsequently formulated as an online multi-objective constrained optimization problem. We derive a Pareto equation for this optimization problem, characterizing the optimal auction mechanism set by its solution set. This solution is implemented through a novel two-stage Adaptive Modeled Mechanism Design (AMMD), which (1) trains a hypernetwork to learn a family of parameterized mechanisms, each corresponding to a specific solution of the Pareto equation, and (2) employs feedback-based online control to adaptively adjust the mechanism parameters, ensuring real-time optimality in a dynamic environment. Extensive experiments demonstrate that AMMD outperforms existing methods in both click-through rates and revenue across multiple auction scenarios, particularly highlighting its adaptability to online environments. The code has been submitted and will be released publicly.", "keywords": [ "online auction", "mechanism design", "multi-objective optimization" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "eR9C6c76j5", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Gabriele Scalia", "gender": "unknown", "institution": "Genentech", "country": "", "position": "Researcher" }, { "name": "G\u00f6kcen Eraslan", "gender": "Male", "institution": "Genentech", "country": "US", "position": "Principal Researcher" }, { "name": "Tommaso Biancalani", "gender": "Male", "institution": "Genentech", "country": "", "position": "Director" }, { "name": "Alex Tseng", "gender": "Non-Binary", "institution": "Genentech", "country": "", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 4, "n_ref_uni": 29, "n_ref": 84, "n_ref_all": 127, "n_fig": 28, "n_tab": 4, "L_tab": 2483, "n_element_tab": 383, "n_fig_1": 21, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1214, "formula_len_all_1": 445, "len_all": 219297, "len_all_1": 58273, "len_abs": 1048, "len_title": 119, "len_sents": 78783, "len_sents_1": 30128, "n_sents": 557, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1052, "title": "A mechanistically interpretable neural network for regulatory genomics", "abs": "Deep neural networks excel in mapping genomic DNA sequences to associated readouts (e.g., protein\u2013DNA binding). Beyond prediction, the goal of these networks is to reveal to scientists the underlying motifs (and their syntax) which drive genome regulation. Traditional methods that extract motifs from convolutional filters suffer from the uninterpretable dispersion of information across filters and layers. Other methods which rely on importance scores can be unstable and unreliable. Instead, we designed a novel mechanistically interpretable architecture for regulatory genomics, where motifs and their syntax are directly encoded and readable from the learned weights and activations. We provide theoretical and empirical evidence of our architecture's full expressivity, while still being highly interpretable. Through several experiments, we show that our architecture excels in de novo motif discovery and motif instance calling, is robust to variable sequence contexts, and enables fully interpretable generation of novel functional sequences.", "keywords": [ "interpretability", "mechanistic interpretability", "attention", "convolution", "regulatory genomics" ], "rating_list": [ 3, 3, 6, 5, 6 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 2, 4 ], "contribution_list": [ 2, 1, 2, 2, 4 ], "confidence_list": [ 4, 4, 2, 3, 3 ] }, { "paper_id": "eR1119aUlL", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Eray Erturk", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Maryam Shanechi", "gender": "Female", "institution": "University of Southern California", "country": "", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 69, "n_ref": 160, "n_ref_all": 218, "n_fig": 21, "n_tab": 6, "L_tab": 864, "n_element_tab": 59, "n_fig_1": 16, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1262, "formula_len_all_1": 890, "len_all": 245711, "len_all_1": 66348, "len_abs": 1666, "len_title": 148, "len_sents": 76696, "len_sents_1": 31750, "n_sents": 460, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1670, "title": "Dynamical modeling for real-time inference of nonlinear latent factors in multiscale neural activity", "abs": "Continuous real-time decoding of target variables from time-series data is needed for many applications across various domains including neuroscience. Further, these variables can be encoded across multiple time-series modalities such as discrete spiking activity and continuous field potentials that can have different timescales (i.e., sampling rates) and different probabilistic distributions, or can even be missing at some time-steps. Existing nonlinear models of multimodal neural activity do not support real-time decoding and do not address the different timescales or missing samples across modalities. Here, we develop a learning framework that can nonlinearly aggregate information across multiple time-series modalities with such distinct characteristics, while also enabling real-time decoding. This framework consists of 1) a multiscale encoder that nonlinearly fuses information after learning within-modality dynamics to handle different timescales and missing samples, 2) a multiscale dynamical backbone that extracts multimodal temporal dynamics and enables real-time decoding, and 3) modality-specific decoders to account for different probabilistic distributions across modalities. We further introduce smoothness regularization objectives on the learned dynamics to better decode smooth target variables such as behavioral variables and employ a dropout technique to increase the robustness for missing samples. We show that our model can aggregate information across modalities to improve target variable decoding in simulations and in a real multiscale brain dataset. Further, our method outperforms prior linear and nonlinear multimodal models.", "keywords": [ "Multimodal deep learning", "Missing data", "Neuroscience", "Real-time decoding" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "eQjJeO7pTF", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Amir Habibian", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "NL", "position": "Researcher" }, { "name": "Davide Abati", "gender": "Not Specified", "institution": "Qualcomm Inc, QualComm", "country": "NL", "position": "Engineer, Senior" }, { "name": "Jaegul Choo", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Mohamed Omran", "gender": "unknown", "institution": "Qualcomm Inc, QualComm", "country": "NL", "position": "Researcher" }, { "name": "Auke J. Wiggers", "gender": "Male", "institution": "QualComm", "country": "NL", "position": "Researcher" }, { "name": "JooYeol Yun", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 54, "n_ref": 109, "n_ref_all": 133, "n_fig": 15, "n_tab": 3, "L_tab": 819, "n_element_tab": 135, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1261, "n_element_tab_1": 86, "formula_len_all": 401, "formula_len_all_1": 549, "len_all": 150028, "len_all_1": 59342, "len_abs": 1152, "len_title": 113, "len_sents": 44055, "len_sents_1": 27495, "n_sents": 312, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1158, "title": "Generative Location Modeling for Spatially Aware Object Insertion", "abs": "Generative models have become a powerful tool for image editing tasks, including object insertion. However, these methods often lack spatial awareness, generating objects with unrealistic locations and scales, or unintentionally altering the scene background. A key challenge lies in maintaining visual coherence, which requires both a geometrically suitable object location and a high-quality image edit. In this paper, we focus on the former, creating a *location model* dedicated to identifying realistic object locations. Specifically, we train an autoregressive model that generates bounding box coordinates, conditioned on the background image and the desired object class.\nThis formulation allows to effectively handle sparse placement annotations and to incorporate implausible locations into a preference dataset by performing direct preference optimization. Our extensive experiments demonstrate that our generative location model, when paired with an inpainting method, substantially outperforms state-of-the-art instruction-tuned models and location modeling baselines in object insertion tasks, delivering accurate and visually coherent results.", "keywords": [ "Object insertion", "image editing", "location modeling" ], "rating_list": [ 5, 3, 6, 6, 5 ], "soundness_list": [ 2, 2, 3, 3, 2 ], "presentation_list": [ 3, 1, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 2 ] }, { "paper_id": "eQggPqESBr", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Etienne Boursier", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Researcher" }, { "name": "Nicolas Flammarion", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 160, "n_formula_1": 28, "n_ref_uni": 60, "n_ref": 115, "n_ref_all": 193, "n_fig": 10, "n_tab": 6, "L_tab": 1168, "n_element_tab": 246, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 14260, "formula_len_all_1": 1779, "len_all": 276715, "len_all_1": 76339, "len_abs": 2145, "len_title": 117, "len_sents": 69862, "len_sents_1": 33725, "n_sents": 616, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1790, "title": "Simplicity Bias and Optimization Threshold in Two-Layer Networks", "abs": "Understanding generalization of overparametrized neural networks remains a fundamental challenge in machine learning. \nMost of the literature mostly studies generalization from an interpolation point of view, taking convergence of parameters towards a global minimum of the training loss for granted. While overparametrized architectures indeed interpolated the data for typical classification tasks, this interpolation paradigm does not seem valid anymore for more complex tasks such as in-context learning or diffusion. Instead for such tasks, it has been empirically observed that the trained models goes from global minima to spurious local minima of the training loss as the number of training samples becomes larger than some level we call *optimization threshold*. While the former yields a poor generalization to the true population loss, the latter was observed to actually correspond to the minimiser of this true loss.\nThis paper explores theoretically this phenomenon in the context of two-layer ReLU networks. We demonstrate that, despite overparametrization, networks often converge toward simpler solutions rather than interpolating the training data, which can lead to a drastic improvement on the test loss with respect to interpolating solutions. \nOur analysis relies on the so called early alignment phase, during which neurons align towards specific directions. This directional alignment, which occurs in the early stage of training, leads to a simplicity bias, wherein the network approximates the ground truth model without converging to the global minimum of the training loss. Our results suggest that this bias, resulting in an optimization threshold from which interpolation is not reached anymore, is beneficial and enhances the generalization of trained models.", "keywords": [ "Neural Networks", "Simplicty Bias", "Implicit Bias", "One hidden ReLU Network", "Early Alignment" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 3, 2, 4 ] }, { "paper_id": "eQDdfqacoR", "primary_area": "learning on time series and dynamical systems", "n_author": 8, "author_info_list": [ { "name": "Dalin Qin", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Liang Sun", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Staff Software Engineer" }, { "name": "Pierre Pinson", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Full Professor" }, { "name": "Weiqi Chen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yehui Li", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yi Wang", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Zhaoyang Zhu", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Qingsong Wen", "gender": "Male", "institution": "Squirrel Ai Learning", "country": "US", "position": "Principal Researcher" } ], "n_formula": 17, "n_formula_1": 18, "n_ref_uni": 37, "n_ref": 79, "n_ref_all": 101, "n_fig": 5, "n_tab": 12, "L_tab": 9299, "n_element_tab": 338, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1565, "n_element_tab_1": 61, "formula_len_all": 1245, "formula_len_all_1": 1287, "len_all": 150429, "len_all_1": 63832, "len_abs": 1224, "len_title": 88, "len_sents": 46514, "len_sents_1": 30104, "n_sents": 281, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1157, "title": "Evolving Multi-Scale Normalization for Time Series Forecasting under Distribution Shifts", "abs": "Complex distribution shifts are the main obstacle to achieving accurate long-term time series forecasting. Several efforts have been conducted to capture the distribution characteristics and propose adaptive normalization techniques to alleviate the influence of distribution shifts. However, these methods neglect intricate distribution dynamics that are observed from various scales and the evolving functions of both distribution dynamics and normalized mapping relationships. To this end, we propose a novel model-agnostic Evolving Multi-Scale Normalization (EvoMSN) framework to tackle the distribution shift problem. Flexible normalization and denormalization are proposed based on the multi-scale statistics prediction module and adaptive ensembling. An evolving optimization strategy is designed to update the forecasting model and statistics prediction module collaboratively to track the shifting distributions. We evaluate the effectiveness of EvoMSN in improving the performance of five mainstream forecasting methods on benchmark datasets and also show its superiority compared to existing advanced normalization and online learning approaches.", "keywords": [ "Time series forecasting", "Distribution shifts", "Normalization", "Online learning", "Multi-scale modeling" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "ePJrZLIqpV", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chenliang Xu", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Associate Professor" }, { "name": "Daiki Shimada", "gender": "Male", "institution": "University of Rochester", "country": "", "position": "Researcher" }, { "name": "Susan Liang", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Zeliang Zhang", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 40, "n_ref": 74, "n_ref_all": 87, "n_fig": 6, "n_tab": 3, "L_tab": 918, "n_element_tab": 132, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 934, "n_element_tab_1": 195, "formula_len_all": 1156, "formula_len_all_1": 1050, "len_all": 134120, "len_all_1": 63638, "len_abs": 1382, "len_title": 139, "len_sents": 41438, "len_sents_1": 29686, "n_sents": 276, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1387, "title": "Rethinking Audio-Visual Adversarial Vulnerability from Temporal and Modality Perspectives", "abs": "While audio-visual learning equips models with a richer understanding of the real world by leveraging multiple sensory modalities, this integration also introduces new vulnerabilities to adversarial attacks.\n In this paper, we present a comprehensive study of the adversarial robustness of audio-visual models, considering both temporal and modality-specific vulnerabilities. We propose two powerful adversarial attacks: 1) a temporal invariance attack that exploits the inherent temporal redundancy across consecutive time segments and 2) a modality misalignment attack that introduces incongruence between the audio and visual modalities. These attacks are designed to thoroughly assess the robustness of audio-visual models against diverse threats. Furthermore, to defend against such attacks, we introduce a novel audio-visual adversarial training framework. This framework addresses key challenges in vanilla adversarial training by incorporating efficient adversarial perturbation crafting tailored to multi-modal data and an adversarial curriculum strategy. Extensive experiments in the Kinetics-Sounds dataset demonstrate that our proposed temporal and modality-based attacks in degrading model performance can achieve state-of-the-art performance, while our adversarial training defense largely improves the adversarial robustness as well as the adversarial training efficiency.", "keywords": [ "audio-visual learning; adversarial attack" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "ePGheWbLPY", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Adrian Weller", "gender": "Male", "institution": "Alan Turing Institute", "country": "", "position": "Programme Director for AI" }, { "name": "Cristiana Diaconu", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Eric Langezaal", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "MS student" }, { "name": "Matthew Ashman", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Richard E. Turner", "gender": "Male", "institution": "Alan Turing Institute", "country": "GB", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 9, "n_ref_uni": 40, "n_ref": 106, "n_ref_all": 153, "n_fig": 19, "n_tab": 9, "L_tab": 2674, "n_element_tab": 388, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1418, "formula_len_all_1": 683, "len_all": 242827, "len_all_1": 70665, "len_abs": 2525, "len_title": 128, "len_sents": 82869, "len_sents_1": 33370, "n_sents": 626, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1483, "title": "Gridded Transformer Neural Processes for Large Unstructured Spatio-Temporal Data", "abs": "Many important problems require modelling large-scale spatio-temporal datasets, with one prevalent example being weather forecasting. Recently, transformer-based approaches have shown great promise in a range of weather forecasting problems. However, these have mostly focused on gridded data sources, neglecting the wealth of unstructured, off-the-grid data from observational measurements such as those at weather stations. A promising family of models suitable for such tasks are neural processes (NPs), notably the family of transformer neural processes (TNPs). Although TNPs have shown promise on small spatio-temporal datasets, they are unable to scale to the quantities of data used by state-of-the-art weather and climate models. This limitation stems from their lack of efficient attention mechanisms. We address this shortcoming through the introduction of gridded pseudo-token TNPs which employ specialised encoders and decoders to handle unstructured observations and utilise a processor containing gridded pseudo-tokens that leverage efficient attention mechanisms. Our method consistently outperforms a range of strong baselines on various synthetic and real-world regression tasks involving large-scale data, while maintaining competitive computational efficiency. The real-life experiments are performed on weather data, demonstrating the potential of our approach to bring performance and computational benefits when applied at scale in a weather modelling pipeline.", "keywords": [ "neural process", "probabilistic machine learning", "transformer", "spatio-temporal data" ], "rating_list": [ 6, 8, 5, 3 ], "soundness_list": [ 3, 3, 3, 1 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "ePEZvQNFDW", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Fredrik Lindsten", "gender": "Male", "institution": "Link\u00f6ping University", "country": "SE", "position": "Associate Professor" }, { "name": "Joel Oskarsson", "gender": "Male", "institution": "Link\u00f6ping University", "country": "SE", "position": "PhD student" }, { "name": "Martin Andrae", "gender": "Male", "institution": "Link\u00f6ping University", "country": "SE", "position": "PhD student" }, { "name": "Tomas Landelius", "gender": "Male", "institution": "SMHI", "country": "SE", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 6, "n_ref_uni": 35, "n_ref": 74, "n_ref_all": 98, "n_fig": 12, "n_tab": 8, "L_tab": 2482, "n_element_tab": 423, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1832, "n_element_tab_1": 20, "formula_len_all": 2053, "formula_len_all_1": 419, "len_all": 165797, "len_all_1": 68006, "len_abs": 1243, "len_title": 109, "len_sents": 46925, "len_sents_1": 30492, "n_sents": 400, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1033, "title": "Continuous Ensemble Weather Forecasting with Diffusion models", "abs": "Weather forecasting has seen a shift in methods from numerical simulations to data-driven systems. While initial research in the area focused on deterministic forecasting, recent works have used diffusion models to produce skillful ensemble forecasts. These models are trained on a single forecasting step and rolled out autoregressively. However, they are computationally expensive and accumulate errors for high temporal resolution due to the many rollout steps. We address these limitations with Continuous Ensemble Forecasting, a novel and flexible method for sampling ensemble forecasts in diffusion models. The method can generate temporally consistent ensemble trajectories completely in parallel, with no autoregressive steps. Continuous Ensemble Forecasting can also be combined with autoregressive rollouts to yield forecasts at an arbitrary fine temporal resolution without sacrificing accuracy. We demonstrate that the method achieves competitive results for global weather forecasting with good probabilistic properties.", "keywords": [ "weather forecasting", "diffusion", "ensemble forecasting" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 3, 3, 2 ] }, { "paper_id": "eP5ICc0584", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Jialiang LU", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "", "position": "Associate Professor" }, { "name": "JunePyo Jung", "gender": "unknown", "institution": "Ajou University", "country": "KR", "position": "Undergrad student" }, { "name": "Keun-Woo Lim", "gender": "Male", "institution": "T\u00e9l\u00e9com Paris", "country": "FR", "position": "Associate Professor" }, { "name": "Leonardo Linguaglossa", "gender": "Male", "institution": "Telecom Paris", "country": "FR", "position": "Associate Professor" }, { "name": "Xicun Yang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 26, "n_ref": 34, "n_ref_all": 42, "n_fig": 1, "n_tab": 4, "L_tab": 1210, "n_element_tab": 194, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 3736, "n_element_tab_1": 223, "formula_len_all": 713, "formula_len_all_1": 713, "len_all": 100372, "len_all_1": 60662, "len_abs": 1437, "len_title": 122, "len_sents": 26692, "len_sents_1": 25317, "n_sents": 191, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1446, "title": "MVFL: Multivariate Vertical Federated Learning for Time-Series Forecasting", "abs": "Extending multivariate time series forecasting to resource-limited devices is a critical demand for real applications, especially with the advancements in IoT technologies. A common scenario is where the variates are distributed vertically on different devices and each device needs to do local forecasting. This paper studies a resource-efficient solution for this scenario based on vertical federated learning (VFL). Prior VFL frameworks are designed for situations where only one party holds the labels and would struggle to meet the demand of the targeted scenario, as storage resources usage would increase dramatically with the number of devices. Going beyond VFL, we design multivariate vertical federated learning (MVFL) as a novel federated learning framework, where we separate communication features and local features in an embedded feature space. This design enables MVFL to utilize storage and communication resources more efficiently by eliminating the redundant models. MVFL outperforms VFL approaches in both efficiency and accuracy. On four real-world benchmarks, compared to VFL, when the storage resources are equally utilized, MVFL yields a 12.1\\% relative improvement on loss with a 43\\% relative improvement on communication resources usage. Even when both MVFL and VFL employ the same main model size, MVFL achieves a 75\\% reduction in storage resources compared to VFL while maintaining the loss at the same level of VFL.", "keywords": [ "vertical federated learning", "multivariate time series forecasting", "resource-limited devices" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 4, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "eOlCIiNe5o", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Deyu Meng", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Haokun Lin", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Qian Zhao", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Quanziang Wang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Renzhen Wang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yichen Wu", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zhiwu Wang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 48, "n_ref": 107, "n_ref_all": 127, "n_fig": 8, "n_tab": 2, "L_tab": 1699, "n_element_tab": 192, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3360, "n_element_tab_1": 254, "formula_len_all": 580, "formula_len_all_1": 406, "len_all": 134547, "len_all_1": 72087, "len_abs": 1387, "len_title": 114, "len_sents": 35411, "len_sents_1": 31321, "n_sents": 235, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1392, "title": "Singular Value Fine-tuning for Few-Shot Class-Incremental Learning", "abs": "Class-Incremental Learning (CIL) aims to learn knowledge from new classes sequentially, while rataining the knowledge obtained from previously encountered classes, thereby mitigating the challenge of Catastrophic Forgetting. In a more realistic scenario, future unseen classes may contain only a few samples, leading to a new challenge of over-fitting, which is referred to as Few-Shot Class-Incremental Learning (FSCIL). Existing works explore FSCIL from various perspectives, such as classifier calibration and backbone extension. Most of them treat the many-shot base session and incremental few-shot sessions separately, as the model tends to overfit on few-shot classes. In this paper, we propose Singular Value Fine-tuning for few-shot Class-incremental Learning (SVFCL) to constantly learn base and incremental sessions based on the pre-trained ViT encoder. SVFCL incorporates incremental adapters, each of which is attached to a corresponding pre-trained module and contains only a small number of learnable parameters, effectively reducing the risk of overfitting. Furthermore, since each adapter is task-specific, information from previous tasks is well-preserved, mitigating catastrophic forgetting.\nOur experimental results demonstrate that SVFCL achieves substantial improvements over state-of-the-art methods while requiring significantly less computational overhead and epochs.", "keywords": [ "Few-Shot Continual Learning; Class-Incremental Learning" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "eOE2g28eX9", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Amir Najafi", "gender": "Male", "institution": "Sharif University of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Farzan Farnia", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Samin Mahdizadeh Sani", "gender": "Female", "institution": "University of Tehran, University of Tehran", "country": "IR", "position": "Researcher" } ], "n_formula": 138, "n_formula_1": 28, "n_ref_uni": 27, "n_ref": 37, "n_ref_all": 47, "n_fig": 6, "n_tab": 1, "L_tab": 293, "n_element_tab": 34, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9636, "formula_len_all_1": 1964, "len_all": 171497, "len_all_1": 68212, "len_abs": 1302, "len_title": 107, "len_sents": 56785, "len_sents_1": 31239, "n_sents": 520, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1342, "title": "Robust Model Evaluation over Large-scale Federated Networks", "abs": "In this paper, we address the challenge of certifying the performance of a machine learning model on an unseen target network. We consider a source network \u201cA\u201d of $K$ clients, each with private data from unique and heterogeneous distributions, assumed to be independent samples from a broader meta-distribution $ \\mu $. Our goal is to provide certified guarantees for the model\u2019s performance on a different, unseen target network \u201cB,\u201d governed by another meta-distribution $ \\mu' $, assuming the deviation between $\\mu$ and $\\mu'$ is bounded by either the {\\it Wasserstein} distance or an $f$-{\\it divergence}. We derive theoretical guarantees for the model\u2019s empirical average loss and provide uniform bounds on the risk CDF, where the latter correspond to novel and adversarially robust versions of the Glivenko-Cantelli theorem and the Dvoretzky-Kiefer-Wolfowitz (DKW) inequality. Our bounds are computable in polynomial time with a polynomial number of queries to the $K$ clients, preserving client privacy by querying only the model\u2019s (potentially adversarial) loss on private data. We also establish non-asymptotic generalization bounds that consistently converge to zero as both $K$ and the minimum client sample size grow. Extensive empirical evaluations validate the robustness and practicality of our bounds across real-world tasks.", "keywords": [ "Distributionally Robust Optimization (DRO)", "Generalization Bound", "Federated Model Evaluation", "Glivenko-Cantelli Theorem and DKW Bound" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 1, 4, 2, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 4, 2, 3 ] }, { "paper_id": "eNjXcP6C0H", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Tianshu Yu", "gender": "Male", "institution": "Chinese University of Hong Kong (Shenzhen)", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaoxue Wang", "gender": "Female", "institution": "ChemLex Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Zihan Zhou", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "PhD student" } ], "n_formula": 82, "n_formula_1": 31, "n_ref_uni": 65, "n_ref": 102, "n_ref_all": 128, "n_fig": 9, "n_tab": 8, "L_tab": 2264, "n_element_tab": 150, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5830, "formula_len_all_1": 2688, "len_all": 231197, "len_all_1": 72732, "len_abs": 1250, "len_title": 89, "len_sents": 68777, "len_sents_1": 32580, "n_sents": 610, "n_sents_1": 282, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 1257, "title": "Generating Physical Dynamics under Priors", "abs": "Generating physically feasible dynamics in a data-driven context is challenging, especially when adhering to physical priors expressed in specific equations or formulas. Existing methodologies often overlook the integration of ''physical priors'', resulting in violation of basic physical laws and suboptimal performance. In this paper, we introduce a novel framework that seamlessly incorporates physical priors into diffusion-based generative models to address this limitation. Our approach leverages two categories of priors: 1) distributional priors, such as roto-translational invariance, and 2) physical feasibility priors, including energy and momentum conservation laws and PDE constraints. By embedding these priors into the generative process, our method can efficiently generate physically realistic dynamics, encompassing trajectories and flows. Empirical evaluations demonstrate that our method produces high-quality dynamics across a diverse array of physical phenomena with remarkable robustness, underscoring its potential to advance data-driven studies in AI4Physics. Our contributions signify a substantial advancement in the field of generative modeling, offering a robust solution to generate accurate and physically consistent dynamics.", "keywords": [ "diffusion models", "generative models", "physical dynamics", "priors" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "eNbA8Fqir4", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Dayiheng Liu", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Junbo Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Junyang Lin", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Ru Peng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yawen Zeng", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Intern" }, { "name": "kexin yang", "gender": "Female", "institution": "Alibaba Group", "country": "", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 70, "n_ref": 207, "n_ref_all": 317, "n_fig": 16, "n_tab": 55, "L_tab": 111199, "n_element_tab": 7182, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 3412, "n_element_tab_1": 28, "formula_len_all": 928, "formula_len_all_1": 467, "len_all": 485401, "len_all_1": 68544, "len_abs": 5278, "len_title": 108, "len_sents": 68911, "len_sents_1": 30595, "n_sents": 503, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 2707, "title": "DataMan: Data Manager for Pre-training Large Language Models", "abs": "As the performance of large language models (LLMs) emerges via data scaling, the significance of pre-training data becomes increasingly evident. Although methods such as deduplication and high-quality sampling have explored data selection, comprehensive criteria for text quality remain underdeveloped, hindering efficient pre-training data selection and composition.\nThis paper establishes guidelines for data selection, fosters consensus on data quality, and introduces a management tool to evaluate data quality and domain types. \nWe believe that robust quality criteria should be applicable across diverse texts, showcasing semantic content understanding, and mutual complement.\nPrevious work mainly relies on intuition and lacks generalizability. To tackle this, we employ reverse thinking\u2014\\emph{prompting LLMs to self-identify the causes of anomalous perplexity (PPL)} in text\u2014and derive 13 quality criteria related to LLM performance, collectively derive a comprehensive\nmetric as \\emph{Overall Score}. \nWe developed a complete prompt that integrates quality criteria and domain types.\nWe use LLM's pointwise ratings and compare the computational complexities of pointwise and pairwise ratings (\\(O(N)\\) v.s. \\(O(N^{2})\\)), showing that pointwise ratings are more feasible for vast datasets, with over 95\\% agreement with human assessments.\nBy annotating 356K documents using GPT-4-turbo and fine-tuning a Qwen2-1.5B model, we created the \\textbf{Data} \\textbf{Man}ager (\\textbf{DataMan}), with an average fine-tuning accuracy across all criteria approaching 80\\% and 81.6\\% for \\emph{Overall Score}.\nWe annotated 447B tokens from the slimpajama corpus by DataMan, and selected a 30B token subset to maximize quality representativeness while ensuring domain diversity to train 1.3B-parameter LLM.\nResults show that models trained on DataMan-sampled data exceed state-of-the-art benchmarks in in-context learning (ICL) gain by 0.4\\% to 4.3\\% and in instruct following win rate by 34.2\\% to 57\\%.\nThe strongest model \\emph{Overall Score l=5}, significantly surpasses models trained on uniform sampling with 50\\% more data.\nContinued pre-training on high-rated domain-specific data further boosts ICL performance, validating DataMan's effectiveness in domain mixing. \nWe reveal that PPL and ICL results do not strictly align, underscoring the distinction between understanding and generalization abilities.\nOur contributions include: i)-developing a data quality criteria system based on LLM PPL features; ii)-creating DataMan for data quality rating and domain identification; and iii)-releasing our code, models, and annotated datasets to facilitate research on the relationship between data and LLMs.", "keywords": [ "Large Language Models", "Data Selection", "Pre-Training" ], "rating_list": [ 5, 6, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 1, 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 2, 4, 3, 3, 4 ] }, { "paper_id": "eNQp79A5Oz", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Rahul Mazumder", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Ryan Lucas", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 54, "n_formula_1": 21, "n_ref_uni": 46, "n_ref": 102, "n_ref_all": 135, "n_fig": 10, "n_tab": 3, "L_tab": 686, "n_element_tab": 42, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 37, "n_element_tab_1": 8, "formula_len_all": 3137, "formula_len_all_1": 1308, "len_all": 182090, "len_all_1": 69954, "len_abs": 1215, "len_title": 152, "len_sents": 57695, "len_sents_1": 32723, "n_sents": 444, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 103, "L_abs": 1230, "title": "Preserving Deep Representations in One-Shot Pruning: A Hessian-Free Second-Order Optimization Framework", "abs": "We present SNOWS, a one-shot post-training pruning framework aimed at reducing neural network inference cost without requiring retraining. Current leading one-shot pruning methods minimize layer-wise least squares reconstruction error which does not take into account deeper network representations. We propose to optimize a more global reconstruction objective that accounts for nonlinear activations deep in the network to obtain a better proxy for the network loss. Optimizing this nonlinear objective leads to a more challenging optimization problem---we demonstrate that it can be solved efficiently using a specialized second-order optimization framework. A key innovation of our framework is the use of Hessian-free optimization to compute exact Newton descent steps without needing to compute or store the full Hessian matrix. A distinct advantage of SNOWS is that it can be readily applied on top of any sparse mask derived from prior methods, readjusting their weights to exploit nonlinearities in deep feature representations. SNOWS obtains state-of-the-art results on various one-shot pruning benchmarks including residual networks and Vision Transformers (VIT/B-16 and VIT/L-16, 86m and 304m parameters respectively).", "keywords": [ "Neural Network Pruning", "Structured Pruning", "Optimization", "Hessian-free Optimization" ], "rating_list": [ 6, 6, 8, 5, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "eNCyY81aW6", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Beidi Chen", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Hongyi Liu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yang Zhou", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zhuoming Chen", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 1, "n_ref_uni": 7, "n_ref": 30, "n_ref_all": 54, "n_fig": 6, "n_tab": 14, "L_tab": 3879, "n_element_tab": 337, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 838, "n_element_tab_1": 109, "formula_len_all": 784, "formula_len_all_1": 126, "len_all": 155546, "len_all_1": 67939, "len_abs": 1579, "len_title": 128, "len_sents": 63675, "len_sents_1": 33890, "n_sents": 531, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1585, "title": "FACTOR: Factoring Complexity and Context Length in Long-Context Model Evaluation", "abs": "Large language models (LLMs) with extended context windows have shown remarkable capabilities, especially with contexts up to 128K tokens. However, whether these resource-intensive LLMs genuinely surpass simpler Retrieval Augmented Generation (RAG) techniques remains debated. \nWe precisely delineate differences between long-context LLMs and RAG methods, emphasizing the unique long-context reasoning abilities of LLMs that RAG cannot replicate. \nExisting benchmarks often focus on retrieval tasks and contain weak if not none complex reasoning tasks, hindering assessment of reasoning over extended contexts. We introduce the \\textbf{FACTOR} benchmark (\\textbf{F}actoring \\textbf{A}nalysis of \\textbf{C}omplexity and \\textbf{T}extual \\textbf{C}ontext in \\textbf{R}easoning), which evaluates LLMs by independently varying task complexity and context length. A comprehensive list of LLMs are evaluated on FACTOR. \nBesides mere accuracy scores, we also model the relationship between accuracy and complexity given the context length. A simple but consistent log-linear model works surprisingly well across various models. Also, the modeling contains two explainable parameters, the slope or Complexity Decay Factor (CDF) and the y-intercept or Contextual Decay Offset (CDO) that are shown to offer separate and insightful measures of the models' complex reasoning and long context innate ability. \nOur findings highlight distinct failure modes linked to task complexity and context length, underscoring the unique reasoning capabilities of long-context LLMs unattainable by RAG methods.", "keywords": [ "Long-context reasoning", "Language models" ], "rating_list": [ 8, 3, 6, 3 ], "soundness_list": [ 4, 2, 2, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "eN4g4cjFX1", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Desong Zhang", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "PhD student" }, { "name": "Geyong Min", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "Full Professor" }, { "name": "Jia Hu", "gender": "unknown", "institution": "University of Exeter", "country": "GB", "position": "Associate Professor" } ], "n_formula": 58, "n_formula_1": 9, "n_ref_uni": 33, "n_ref": 70, "n_ref_all": 141, "n_fig": 5, "n_tab": 9, "L_tab": 5295, "n_element_tab": 948, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 3136, "n_element_tab_1": 200, "formula_len_all": 5138, "formula_len_all_1": 702, "len_all": 195895, "len_all_1": 64717, "len_abs": 2011, "len_title": 128, "len_sents": 53002, "len_sents_1": 29269, "n_sents": 401, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 2021, "title": "Spatio-Temporal Dependency-Aware Neuron Optimization for Spiking Neural Networks", "abs": "As a biologically inspired computing paradigm, Spiking Neural Networks (SNNs) process information through discrete spike sequences, mimicking the brain's temporal dynamics and energy efficiency. The combination of backpropagation through time (BPTT) and direct input encoding (i.e., feeding decimal data directly into the network) has emerged as the mainstream training approach for SNNs. However, this combination introduces varying temporal dependency requirements across the network\u2019s spatial dimension. These differences are often neglected in existing studies, which typically apply uniform temporal dependency configurations throughout the network. Consequently, this could result in missing key gradients or introducing redundant ones in the temporal dimension, ultimately affecting the network's performance. To address this gap, we propose a novel Spatio-Temporal Dependency-Aware Neuron Optimization (ST-DANO) method for SNNs, which consists of two key components: neuron design and neuron search. Specifically, to overcome the limitations of traditional Leaky Integrate-and-Fire (LIF) neurons in adapting to varying temporal dependencies, we designed two variants, Long-LIF and Short-LIF, which improve the neuron's ability to capture long-term and short-term dependencies, respectively, by dynamic modulation of membrane potential thresholds and time constants. After validating our neuron designs through ablation studies, we developed a layer-wise neuron search strategy that automatically selects the optimal neuron type for each layer to ensure optimal temporal dependency configurations across the network. Extensive experiments on static and neuromorphic datasets demonstrate that ST-DANO can effectively adapt to temporal dependency differences across the spatial dimension in SNNs under various time-step configurations. The resulting architectures surpass state-of-the-art performance, achieving a remarkable 83.90\\% accuracy on the DVS-CIFAR-10 dataset\u2014a more than 5\\% improvement over the baseline.", "keywords": [ "Neuromorphic computation", "spiking neural networks", "temporal dependency" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "eN0RyRVbSm", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Antoine Manzanera", "gender": "Male", "institution": "Ecole Nationale Sup\u00e9rieure de Techniques Avanc\u00e9es", "country": "FR", "position": "Full Professor" }, { "name": "Arturo Mendoza", "gender": "Male", "institution": "Safran", "country": "FR", "position": "Researcher" }, { "name": "David Brellmann", "gender": "Male", "institution": "Safran", "country": "FR", "position": "Researcher" }, { "name": "Gianni Franchi", "gender": "Male", "institution": "ENSTA Paris", "country": "FR", "position": "Assistant Professor" }, { "name": "Mou\u00efn Ben Ammar", "gender": "Male", "institution": "ENSTA Paris", "country": "", "position": "PhD student" } ], "n_formula": 68, "n_formula_1": 19, "n_ref_uni": 51, "n_ref": 135, "n_ref_all": 144, "n_fig": 2, "n_tab": 5, "L_tab": 4005, "n_element_tab": 233, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 297, "n_element_tab_1": 5, "formula_len_all": 7892, "formula_len_all_1": 1859, "len_all": 192409, "len_all_1": 61634, "len_abs": 1072, "len_title": 176, "len_sents": 53436, "len_sents_1": 26911, "n_sents": 466, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 127, "L_abs": 1078, "title": "Double Descent Meets Out-of-Distribution Detection: Theoretical Insights and Empirical Analysis of the role of model complexity", "abs": "While overparameterization is known to benefit generalization, its impact on Out-Of-Distribution (OOD) detection is less understood. This paper investigates the influence of model complexity in OOD detection. We propose an expected OOD risk metric to evaluate classifiers confidence on both training and OOD samples. Leveraging Random Matrix Theory, we derive bounds for the expected OOD risk of binary least-squares classifiers applied to Gaussian data. We show that the OOD risk depicts an infinite peak, when the number of parameters is equal to the number of samples, which we associate with the double descent phenomenon. Our experimental study on different OOD detection methods across multiple neural architectures extends our theoretical insights and highlights a double descent curve. Our observations suggest that overparameterization does not necessarily lead to better OOD detection. Using the Neural Collapse framework, we provide insights to better understand this behavior. To facilitate reproducibility, our code will be made publicly available upon publication.", "keywords": [ "Out-Of-Distribution", "double descent." ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "eM5dar35Ys", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Genwei Zhang", "gender": "Male", "institution": "New Jersey Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Guiling Wang", "gender": "Female", "institution": "New Jersey Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Xinyun Zhao", "gender": "Female", "institution": "New Jersey Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 39, "n_ref": 57, "n_ref_all": 70, "n_fig": 8, "n_tab": 3, "L_tab": 411, "n_element_tab": 42, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 1757, "n_element_tab_1": 22, "formula_len_all": 934, "formula_len_all_1": 699, "len_all": 98582, "len_all_1": 48119, "len_abs": 984, "len_title": 145, "len_sents": 22660, "len_sents_1": 19638, "n_sents": 170, "n_sents_1": 147, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 989, "title": "End-to-End Reinforcement Learning for Traffic Signal Control: Real-Time Video to Signal Decisions", "abs": "Efficient traffic management at urban intersections is vital for reducing congestion\nand improving safety. This paper presents MD3DQN, the first End-to-End novel\nreinforcement learning model using surveillance video for real-time traffic signal\ncontrol. The model features two main components: an image reception module,\ncapturing traffic data from cameras positioned on signal poles, and a multi-agent\ndecision module, where each agent manages a traffic phase. These components\nare connected via a bridge module for seamless integration.\n\nOur novel Entropy Attention Mechanism enhances the multi-decision turn-based\ntraffic signal control by leveraging uncertainty and signal phase delays, leading\nto more optimized decisions. Results show MD3DQN improved cumulative reward\nby an average of 85.2% over Fixed-time 40 and 54.4% over DQN-VTP.\nThe entropy mechanism contributed to a 41.8% improvement upon ablation study,\ndemonstrating its impact on faster convergence and better performance.", "keywords": [ "Traffic signal control", "Reinforcement learning" ], "rating_list": [ 3, 3, 1, 3, 3 ], "soundness_list": [ 2, 3, 1, 2, 2 ], "presentation_list": [ 3, 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 1, 2, 2 ], "confidence_list": [ 5, 4, 4, 3, 4 ] }, { "paper_id": "eLOuG5L42a", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Benjamin D. Evans", "gender": "Male", "institution": "University of Sussex", "country": "", "position": "Lecturer" }, { "name": "Dong Yin", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "MS student" }, { "name": "Federico Adolfi", "gender": "Male", "institution": "ESI Neuroscience, Max-Planck Society", "country": "DE", "position": "Postdoc" }, { "name": "Gaurav Malhotra", "gender": "Male", "institution": "State University of New York at Albany", "country": "", "position": "Assistant Professor" }, { "name": "Guillermo Puebla", "gender": "Male", "institution": "Universidad de Tarapac\u00e1", "country": "", "position": "Assistant Professor" }, { "name": "Jeffrey Bowers", "gender": "Male", "institution": "University of Bristol", "country": "", "position": "Full Professor" }, { "name": "John E. Hummel", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "Full Professor" }, { "name": "Karim G. Habashy", "gender": "Not Specified", "institution": "University of Bristol", "country": "", "position": "Postdoc" }, { "name": "Marin Dujmovic", "gender": "Male", "institution": "University of Bristol", "country": "", "position": "Postdoc" }, { "name": "Rachel F Heaton", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "Postdoc" }, { "name": "Valerio Biscione", "gender": "Male", "institution": "University of Bristol", "country": "", "position": "Research" }, { "name": "Milton Montero", "gender": "Male", "institution": "IT University of Copenhagen", "country": "DK", "position": "Postdoc" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 100, "n_ref": 227, "n_ref_all": 257, "n_fig": 12, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 319923, "len_all_1": 60571, "len_abs": 1203, "len_title": 76, "len_sents": 102540, "len_sents_1": 32239, "n_sents": 702, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1207, "title": "MindSet: Vision. A toolbox for testing DNNs on key psychological experiments", "abs": "Multiple benchmarks have been developed to assess the alignment between deep neural networks (DNNs) and human vision. In almost all cases these benchmarks are observational in the sense they are composed of behavioural and brain responses to naturalistic images that have not been manipulated to test hypotheses regarding how DNNs or humans perceive and identify objects. Here we introduce the toolbox MindSet: Vision, consisting of a collection of image datasets and related scripts designed to test DNNs on 30 psychological findings. In all experimental conditions, the stimuli are systematically manipulated to test specific hypotheses regarding human visual perception and object recognition. In addition to providing pre-generated datasets of images, we provide code to regenerate these datasets, offering many configurable parameters which greatly extend the dataset versatility for different research contexts, and code to facilitate the testing of DNNs on these image datasets using three different methods (similarity judgments, out-of-distribution classification, and decoder method), accessible via GitHub. We test ResNet-152 on each of these methods as an example of how the toolbox can be used.", "keywords": [ "Datasets", "Visual Perception", "Psychology", "Human Vision", "Deep Neural Networks", "Visual Illusion" ], "rating_list": [ 8, 3, 3, 8, 3 ], "soundness_list": [ 3, 3, 3, 4, 1 ], "presentation_list": [ 4, 3, 4, 3, 2 ], "contribution_list": [ 3, 2, 4, 3, 2 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "eLLBILFRsA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Huimin LU", "gender": "Female", "institution": "The University of Tokyo", "country": "JP", "position": "MS student" }, { "name": "Ichiro Sakata", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Junichiro Mori", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Masaru Isonuma", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Postdoc" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 46, "n_ref": 116, "n_ref_all": 140, "n_fig": 3, "n_tab": 4, "L_tab": 5434, "n_element_tab": 431, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 3084, "n_element_tab_1": 261, "formula_len_all": 477, "formula_len_all_1": 383, "len_all": 180974, "len_all_1": 68417, "len_abs": 1357, "len_title": 132, "len_sents": 52133, "len_sents_1": 31694, "n_sents": 355, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1368, "title": "UniDetox: Universal Detoxification of Large Language Models via Dataset Distillation", "abs": "We present UniDetox, a universally applicable method designed to mitigate toxicity across various large language models (LLMs).\nPrevious detoxification methods are typically model-specific, addressing only individual models or model families, and require careful hyperparameter tuning due to the trade-off between detoxification efficacy and language modeling performance. \nIn contrast, UniDetox provides a detoxification technique that can be universally applied to a wide range of LLMs without the need for separate model-specific tuning. \nSpecifically, we propose a novel and efficient dataset distillation technique for detoxification using contrastive decoding. \nThis approach distills detoxifying representations in the form of synthetic text data, enabling universal detoxification of any LLM through fine-tuning with the distilled text. \nOur experiments demonstrate that the detoxifying text distilled from GPT-2 can effectively detoxify larger models, including OPT, Falcon, and LLaMA-2. \nFurthermore, UniDetox eliminates the need for separate hyperparameter tuning for each model, as a single hyperparameter configuration can be seamlessly applied across different models. \nAdditionally, analysis of the detoxifying text reveals a reduction in politically biased content, providing insights into the attributes necessary for effective detoxification of LLMs.", "keywords": [ "Large Language Models", "Detoxification", "Safety", "Fairness" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "eLBKQSpsVd", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Ivan Vuli\u0107", "gender": "Male", "institution": "Google DeepMind", "country": "CH", "position": "Researcher" }, { "name": "Jochen De Weerdt", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "Full Professor" }, { "name": "Marie-Francine Moens", "gender": "Female", "institution": "KU Leuven, KU Leuven", "country": "BE", "position": "Full Professor" }, { "name": "Philipp Borchert", "gender": "Male", "institution": "I\u00c9SEG School of Management", "country": "FR", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 39, "n_ref": 65, "n_ref_all": 100, "n_fig": 7, "n_tab": 8, "L_tab": 11341, "n_element_tab": 653, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 84, "formula_len_all_1": 31, "len_all": 188926, "len_all_1": 63605, "len_abs": 1760, "len_title": 110, "len_sents": 43303, "len_sents_1": 31669, "n_sents": 361, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1718, "title": "Language Fusion for Parameter-Efficient Cross-lingual Transfer", "abs": "Limited availability of multilingual text corpora for training language models often leads to poor performance on downstream tasks due to undertrained representation spaces for languages other than English. This 'under-representation' has motivated recent cross-lingual transfer methods to leverage the English representation space by e.g. mixing English and 'non-English' tokens at input or extending model parameters to accommodate new languages, which in turn increases computational complexity. To address this, we introduce **F**usion for **La**nguage **Re**presentations (FLARE) in adapters, a method designed to improve both the representation quality and downstream performance for languages other than English. FLARE integrates source and target language representations within the bottlenecks of low-rank LoRA adapters using lightweight linear transformations. This maintains parameter efficiency as the method does not require additional parameters, while improving transfer performance, further narrowing the performance gap to English.\nFurthermore, the proposed latent representation fusion does not increase the number of input tokens, this way maintaining computational efficiency. Moreover, FLARE provides flexibility to integrate various types of representations, e.g., we show that it is possible to fuse latent translations extracted from machine translation models. A series of experiments across representative cross-lingual natural language understanding tasks, including natural language inference, question-answering and sentiment analysis, demonstrate FLARE's effectiveness, reducing the average performance gap to English to 8.39% for XLM-R Large and 12.41% for Llama 3 across our benchmarks.", "keywords": [ "cross-lingual transfer", "multilingual representation learning", "parameter-efficient fine-tuning" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "eJpI20hzWf", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Chao Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Associate Professor" }, { "name": "Guangzhi Sun", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Junior Research Fellow" }, { "name": "Jun Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "", "position": "Principal Researcher" }, { "name": "Lu Lu", "gender": "Male", "institution": "Louisiana State University", "country": "", "position": "PhD student" }, { "name": "Siyin Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Wenyi Yu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Xianzhao Chen", "gender": "unknown", "institution": "Tianjin University", "country": "", "position": "MS student" }, { "name": "Xiaohai Tian", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Xiaoyu Yang", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Yuxuan Wang", "gender": "Male", "institution": "ByteDance", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 29, "n_ref": 47, "n_ref_all": 55, "n_fig": 4, "n_tab": 3, "L_tab": 820, "n_element_tab": 99, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 677, "n_element_tab_1": 68, "formula_len_all": 44, "formula_len_all_1": 44, "len_all": 86814, "len_all_1": 56459, "len_abs": 1245, "len_title": 141, "len_sents": 18899, "len_sents_1": 18433, "n_sents": 117, "n_sents_1": 103, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1269, "title": "SALMONN-omni: A Speech Understanding and Generation LLM in a Codec-free Full-duplex Framework", "abs": "Speech large language models (LLMs) offer a unified approach to handling various speech-processing tasks using a single autoregressive model built on discrete speech and audio codecs. Unlike traditional pipeline-based systems, which involve separate components for speech recognition, understanding, and generation, end-to-end speech LLMs can capture both verbal and non-verbal information, such as paralinguistic and speaker characteristics. This enables full-duplex capabilities, allowing the system to listen and speak simultaneously with low latency, making it ideal for conversational AI. In this paper, we introduce a novel codec-free, full-duplex framework for speech understanding and generation, and present {SALMONN-omni}, an instance of this speech LLM. SALMONN-omni can listen to its own generated speech and background sounds while speaking. To align the frame rate gap between text and audio, we propose a novel \\textit{thinking} step, ensuring high performance on pre-trained tasks. Using a two-stage \\textit{understand then generate} training approach, SALMONN-omni effectively addresses a variety of streaming speech tasks, including speech recognition, synthesis, enhancement, dereverberation, target speaker extraction, and spoken question answering.", "keywords": [ "Large speech text model", "Full-duplex model" ], "rating_list": [ 3, 5, 1, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "eIgGesYKLG", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Chulhee Yun", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Hanseul Cho", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jaeyoung Cha", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Srinadh Bhojanapalli", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" } ], "n_formula": 56, "n_formula_1": 1, "n_ref_uni": 36, "n_ref": 98, "n_ref_all": 190, "n_fig": 18, "n_tab": 9, "L_tab": 4142, "n_element_tab": 151, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10139, "formula_len_all_1": 71, "len_all": 232379, "len_all_1": 70199, "len_abs": 1112, "len_title": 126, "len_sents": 75949, "len_sents_1": 35593, "n_sents": 608, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1072, "title": "Arithmetic Transformers Can Length-Generalize in Both Operand Length and Count", "abs": "Transformers often struggle with *length generalization*, meaning they fail to generalize to sequences longer than those encountered during training. While arithmetic tasks are commonly used to study length generalization, certain tasks are considered notoriously difficult, e.g., multi-operand addition (requiring generalization over both the number of operands and their lengths) and multiplication (requiring generalization over both operand lengths). In this work, we achieve approximately 2\u20133\u00d7 length generalization on both tasks, which is the first such achievement in arithmetic Transformers. We design task-specific scratchpads enabling the model to focus on a fixed number of tokens per each next-token prediction step, and apply multi-level versions of *Position Coupling* (Cho et al., 2024; McLeish et al., 2024) to let Transformers know the right position to attend to. On the theory side, we prove that a 1-layer Transformer using our method can solve multi-operand addition, up to operand length and operand count that are exponential in embedding dimension.", "keywords": [ "Length Generalization", "Transformers", "Scratchpad", "Position Coupling", "Positional Encoding", "Out-of-distribution Generalization", "Arithmetic Tasks" ], "rating_list": [ 5, 8, 5, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "eIO1YcEdE6", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 23, "author_info_list": [ { "name": "Andrew Tao", "gender": "Male", "institution": "NVIDIA Corporation", "country": "US", "position": "Vice President of Applied Research" }, { "name": "Boris Ivanovic", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Boyi Li", "gender": "Female", "institution": "NVIDIA Research", "country": "US", "position": "Researcher" }, { "name": "Fuzhao Xue", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Jan Kautz", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "VP Research" }, { "name": "Jitendra Malik", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Director of Research" }, { "name": "Jonah Philion", "gender": "Male", "institution": "NVIDIA", "country": "CA", "position": "Researcher" }, { "name": "Ligeng Zhu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Linxi Fan", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Marco Pavone", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Director, Autonomous Vehicle Research" }, { "name": "Max Ehrlich", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Adjunct" }, { "name": "Ming-Yu Liu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Ran Tian", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Sanja Fidler", "gender": "Female", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "Associate Professor" }, { "name": "Shuhan Tan", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Song Han", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Sushant Veer", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Xinshuo Weng", "gender": "Female", "institution": "NVIDIA", "country": "CA", "position": "Researcher" }, { "name": "Yao Lu", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Yin Cui", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Yuke Zhu", "gender": "Male", "institution": "Computer Science Department, University of Texas, Austin", "country": "US", "position": "Assistant Professor" }, { "name": "trevor darrell", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department", "country": "US", "position": "Professor" }, { "name": "Yuxiao Chen", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Postdoc" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 59, "n_ref_all": 78, "n_fig": 6, "n_tab": 5, "L_tab": 530, "n_element_tab": 49, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 777, "n_element_tab_1": 63, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 121301, "len_all_1": 67978, "len_abs": 1296, "len_title": 116, "len_sents": 33342, "len_sents_1": 31074, "n_sents": 267, "n_sents_1": 248, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1317, "title": "Wolf: Accurate Video Captioning with a World Summarization Framework", "abs": "We propose Wolf, a WOrLd summarization Framework for accurate video captioning. Wolf is an automated captioning framework that adopts a mixture-of-experts approach, leveraging complementary strengths of Vision Language Models (VLMs). By utilizing both image and video models, our framework captures different levels of information and summarizes them efficiently. Our approach can be applied to enhance video understanding, auto-labeling, and captioning. To evaluate caption quality, we introduce CapScore, an LLM-based metric to assess the similarity and quality of generated captions compared to the ground truth captions. We further build four human-annotated datasets in three domains: autonomous driving, general scenes, and robotics, to facilitate comprehensive comparisons. We show that Wolf achieves superior captioning performance compared to state-of-the-art approaches from the research community (VILA1.5, CogAgent) and commercial solutions (Gemini-Pro-1.5, GPT-4V). For instance, in comparison with GPT-4V, Wolf improves CapScore (caption quality) by 55.6% and CapScore (caption similarity) by 77.4% on challenging driving videos. Finally, we establish a benchmark for video captioning and introduce a leaderboard, aiming to accelerate advancements in video understanding, captioning, and data alignment.", "keywords": [ "Video Captioning", "Video Understanding", "Multimodal Learning" ], "rating_list": [ 3, 8, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "eIJfOIMN9z", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Leheng Sheng", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Xiang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yi Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Yuxin Chen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "AN ZHANG", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Tat-seng Chua", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 75, "n_ref": 259, "n_ref_all": 304, "n_fig": 17, "n_tab": 13, "L_tab": 6588, "n_element_tab": 435, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2033, "n_element_tab_1": 220, "formula_len_all": 276, "formula_len_all_1": 259, "len_all": 267023, "len_all_1": 68998, "len_abs": 2121, "len_title": 127, "len_sents": 60706, "len_sents_1": 31565, "n_sents": 436, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 79, "L_abs": 2128, "title": "Language Representations Can be What Recommenders Need: Findings and Potentials", "abs": "Recent studies empirically indicate that language models (LMs) encode rich world knowledge beyond mere semantics, attracting significant attention across various fields.\nHowever, in the recommendation domain, it remains uncertain whether LMs implicitly encode user preference information. Contrary to prevailing understanding that LMs and traditional recommenders learn two distinct representation spaces due to the huge gap in language and behavior modeling objectives, this work re-examines such understanding and explores extracting a recommendation space directly from the language representation space.\nSurprisingly, our findings demonstrate that item representations, when linearly mapped from advanced LM representations, yield superior recommendation performance.\nThis outcome suggests the possible homomorphism between the advanced language representation space and an effective item representation space for recommendation, implying that collaborative signals may be implicitly encoded within LMs.\nMotivated by the finding of homomorphism, we explore the possibility of designing advanced collaborative filtering (CF) models purely based on language representations without ID-based embeddings.\nTo be specific, we incorporate several crucial components (i.e., a multilayer perceptron (MLP), graph convolution, and contrastive learning (CL) loss function) to build a simple yet effective model, with the language representations of item textual metadata (i.e., title) as the input.\nEmpirical results show that such a simple model can outperform leading ID-based CF models on multiple datasets, which sheds light on using language representations for better recommendation.\nMoreover, we systematically analyze this simple model and find several key features for using advanced language representations:\na good initialization for item representations, superior zero-shot recommendation abilities in new datasets, and being aware of user intention.\nOur findings highlight the connection between language modeling and behavior modeling, which can inspire both natural language processing and recommender system communities.", "keywords": [ "Collaborative filtering", "Language-representation-based recommendation", "Language models", "Language model representations" ], "rating_list": [ 8, 8, 5, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 4, 4, 3, 4 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 5, 4, 5, 2 ] }, { "paper_id": "eIFHoPsIkw", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 9, "author_info_list": [ { "name": "Barry-John Theobald", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Chen Huang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Research Scientist" }, { "name": "Fartash Faghri", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Hadi Pouransari", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Joshua M. Susskind", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Mehrdad Farajtabar", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Oncel Tuzel", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Raviteja Vemulapalli", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Skyler Seto", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 10, "n_ref_uni": 55, "n_ref": 112, "n_ref_all": 151, "n_fig": 10, "n_tab": 10, "L_tab": 8533, "n_element_tab": 597, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 920, "n_element_tab_1": 70, "formula_len_all": 1815, "formula_len_all_1": 1127, "len_all": 207768, "len_all_1": 68463, "len_abs": 1521, "len_title": 161, "len_sents": 57770, "len_sents_1": 31429, "n_sents": 439, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 113, "L_abs": 1528, "title": "Proxy-FDA: Proxy-based Feature Distribution Alignment for Fine-tuning Vision Foundation Models without Forgetting", "abs": "Vision foundation models pre-trained on massive data encode rich representations of real-world concepts, which can be adapted to downstream tasks by fine-tuning. However, fine-tuning foundation models on one task often leads to the issue of concept forgetting on other tasks, and this issue is exacerbated by the typically limited data for fine-tuning. Recent methods of robust fine-tuning aim to mitigate forgetting of prior knowledge without affecting the fine-tuning performance. Knowledge is often preserved by matching the original and fine-tuned model weights or feature pairs. However, such point-wise matching can be too strong, without explicit awareness of the feature neighborhood structures that encode rich knowledge as well. We propose a novel regularization method Proxy-FDA that explicitly preserves the structural knowledge in feature space. Proxy-FDA performs Feature Distribution Alignment (using nearest neighbor graphs) between the pre-trained and fine-tuned feature spaces, and the alignment is further improved by informative proxies that are generated dynamically to increase data diversity. We show in end-to-end fine-tuning experiments that Proxy-FDA significantly reduces concept forgetting, and we find a strong correlation between forgetting and a distributional distance metric (in comparison to L2 distance). We further demonstrate Proxy-FDA's utility in both few-shot (based on prompt tuning) and continual fine-tuning settings, where we achieve consistent gains over the corresponding baselines.", "keywords": [ "Proxy-FDA", "robust fine-tuning", "concept forgetting", "vision foundation model" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "eIB1UZFcFg", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Alexandre Variengien", "gender": "Male", "institution": "Conjecture", "country": "GB", "position": "Intern" }, { "name": "Eric Winsor", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 63, "n_ref_all": 123, "n_fig": 26, "n_tab": 6, "L_tab": 2106, "n_element_tab": 224, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1530, "formula_len_all_1": 228, "len_all": 252180, "len_all_1": 63829, "len_abs": 836, "len_title": 131, "len_sents": 94229, "len_sents_1": 30943, "n_sents": 800, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1655, "title": "Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models", "abs": "When solving challenging problems, language models (LMs) are able to identify relevant information from long and complicated contexts. To study how LMs solve retrieval tasks in diverse situations, we introduce ORION, a collection of structured retrieval tasks spanning six domains, from text understanding to coding. Each task in ORION can be represented abstractly by a request (e.g. a question) that retrieves an attribute (e.g. the character name) from a context (e.g. a story). We apply causal analysis on 18 open-source language models with sizes ranging from 125 million to 70 billion parameters. We find that LMs internally decompose retrieval tasks in a modular way: middle layers at the last token position process the request, while late layers retrieve the correct entity from the context. After causally enforcing this decomposition, models are still able to solve the original task, preserving 70% of the original correct token probability in 98 of the 106 studied model-task pairs. We connect our macroscopic decomposition with a microscopic description by performing a fine-grained case study of a question-answering task on Pythia-2.8b. Building on our high-level understanding, we demonstrate a proof of concept application for scalable internal oversight of LMs to mitigate prompt-injection while requiring human supervision on only a single input. Our solution improves accuracy drastically (from 15.5% to 97.5% on Pythia-12b). This work presents evidence of a universal emergent modular processing of tasks across varied domains and models and is a pioneering effort in applying interpretability for scalable internal oversight of LMs.", "keywords": [ "Interpretability", "LLM", "Universality" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "eI3hEAWe8W", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Fernanda Vi\u00e9gas", "gender": "Female", "institution": "Google", "country": "", "position": "Principal Researcher" }, { "name": "Kenneth Li", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Martin Wattenberg", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Yiming Wang", "gender": "Male", "institution": "UCLA Computer Science Department, University of California, Los Angeles", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 54, "n_ref": 90, "n_ref_all": 122, "n_fig": 5, "n_tab": 1, "L_tab": 255, "n_element_tab": 10, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 255, "n_element_tab_1": 10, "formula_len_all": 429, "formula_len_all_1": 429, "len_all": 160179, "len_all_1": 72673, "len_abs": 805, "len_title": 148, "len_sents": 43959, "len_sents_1": 34727, "n_sents": 348, "n_sents_1": 270, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 808, "title": "Dialogue Action Tokens: Steering Language Models in Goal-Directed Dialogue with a Multi-Turn Planner", "abs": "We present an approach called Dialogue Action Tokens (DAT) that adapts language model agents to plan goal-directed dialogues. The core idea is to treat each utterance as an action, thereby converting dialogues into games where existing approaches such as reinforcement learning can be applied. Specifically, we freeze a pretrained language model and train a small planner model that predicts a continuous action vector, used for controlled generation in each round. This design avoids the problem of language degradation under reward optimization. When evaluated on the Sotopia platform for social simulations, the DAT-steered LLaMA model surpasses GPT-4's performance. We also apply DAT to steer an attacker language model in a novel multi-turn red-teaming setting, revealing a potential new attack surface.", "keywords": [ "Large Language Model", "Dialogue Systems", "Social Intelligence", "Red Teaming" ], "rating_list": [ 5, 8, 8, 6, 5 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 4, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "eHfq8Q3LeD", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Danrong Li", "gender": "Female", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Juliana Freire", "gender": "Female", "institution": "New York University", "country": "US", "position": "Professor of Computer Science and Data Science" }, { "name": "Majid Daliri", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Christopher P Musco", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 45, "n_formula_1": 23, "n_ref_uni": 38, "n_ref": 67, "n_ref_all": 80, "n_fig": 4, "n_tab": 1, "L_tab": 91, "n_element_tab": 3, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5869, "formula_len_all_1": 2099, "len_all": 162846, "len_all_1": 83731, "len_abs": 1100, "len_title": 97, "len_sents": 41431, "len_sents_1": 30870, "n_sents": 364, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1283, "title": "Matrix Product Sketching via Coordinated Sampling", "abs": "We revisit the well-studied problem of approximating a matrix product, $\\bv{A}^T\\bv{B}$, based on small space sketches $\\mathcal{S}(\\bv{A})$ and $\\mathcal{S}(\\bv{B})$ of $\\bv{A} \\in \\R^{n \\times d}$ and $\\bv{B}\\in \\R^{n \\times m}$. We are interested in the setting where the sketches must be computed independently of each other, except for the use of a shared random seed. We prove that, when $\\bv{A}$ and $\\bv{B}$ are sparse, methods based on \\emph{coordinated random sampling} can outperform classical linear sketching approaches, like Johnson-Lindenstrauss Projection or CountSketch. For example, to obtain Frobenius norm error $\\epsilon\\|\\bv{A}\\|_F\\|\\bv{B}\\|_F$, coordinated sampling requires sketches of size $O(s/\\epsilon^2)$ when $\\bv{A}$ and $\\bv{B}$ have at most $s \\leq d,m$ non-zeros per row. In contrast, linear sketching leads to sketches of size $O(d/\\epsilon^2)$ and $O(m/\\epsilon^2)$ for $\\bv{A}$ and $\\bv{B}$. We empirically evaluate our approach on two applications: 1) distributed linear regression in databases, a problem motivated by tasks like dataset discovery and augmentation, and 2) approximating attention matrices in transformer-based language models. In both cases, our sampling algorithms yield an order of magnitude improvement over linear sketching.", "keywords": [ "Sketching Algorithm", "Matrix Multiplication", "Model Compression", "Data Discovery", "Efficient resource" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "eHf4mj9m9o", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Botao Li", "gender": "Male", "institution": "Laboratoire de Probabilit\u00e9s, Statistique et Mod\u00e9lisation", "country": "FR", "position": "Postdoc" }, { "name": "Liu Ziyin", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Masahito Ueda", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Tomer Galanti", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" } ], "n_formula": 38, "n_formula_1": 26, "n_ref_uni": 37, "n_ref": 49, "n_ref_all": 98, "n_fig": 16, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1654, "formula_len_all_1": 1077, "len_all": 153868, "len_all_1": 76820, "len_abs": 1189, "len_title": 122, "len_sents": 53270, "len_sents_1": 34595, "n_sents": 507, "n_sents_1": 293, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1192, "title": "Type-II Saddles and Probabilistic Stability of Stochastic Gradient Descent", "abs": "Characterizing and understanding the dynamics of stochastic gradient descent (SGD) around saddle points remains an open problem in neural network optimization. We identify two distinct types of saddle points, demonstrating that Type-II saddles pose a significant challenge due to vanishing gradient noise, which makes them particularly difficult for SGD to escape. We show that the dynamics around these saddles can be effectively modeled by a random matrix product process, allowing us to apply concepts from probabilistic stability and Lyapunov exponents. By leveraging ergodic theory, we establish that saddle points can be either attractive or repulsive for SGD, leading to a classification of four distinct dynamic phases based on the gradient's signal-to-noise ratio near the saddle. We apply the theory to the training at the initial stage of neural networks, explaining an intriguing phenomenon that neural networks are prone to be stuck at the initialization point at a larger learning rate. Our results offer a novel theoretical framework for understanding the intricate behavior of SGD around saddle points, with implications for improving optimization strategies in deep learning.", "keywords": [ "stochastic gradient descent", "linear stability", "saddle points", "deep learning theory" ], "rating_list": [ 3, 5, 8, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 1, 2, 3, 1 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "eHehzSDUFp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Dohaeng Lee", "gender": "unknown", "institution": "LG Corporation", "country": "KR", "position": "Researcher" }, { "name": "Hyeonbin Hwang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" }, { "name": "Hyowon Cho", "gender": "Female", "institution": "KAIST", "country": "", "position": "MS student" }, { "name": "Hyunji Lee", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Jiyeon Kim", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" }, { "name": "Joel Jang", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Minjoon Seo", "gender": "Male", "institution": "Twelve Labs", "country": "", "position": "Chief Scientist" }, { "name": "Seungpil Won", "gender": "unknown", "institution": "LG Corporation", "country": "KR", "position": "Researcher" }, { "name": "Youbin Ahn", "gender": "Female", "institution": "LG Corporation", "country": "", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 6, "n_ref_uni": 37, "n_ref": 100, "n_ref_all": 144, "n_fig": 20, "n_tab": 2, "L_tab": 542, "n_element_tab": 78, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 617, "n_element_tab_1": 34, "formula_len_all": 1112, "formula_len_all_1": 345, "len_all": 155417, "len_all_1": 52932, "len_abs": 1095, "len_title": 139, "len_sents": 52673, "len_sents_1": 26365, "n_sents": 337, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 91, "L_abs": 1098, "title": "Knowledge Entropy Decay during Language Model Pretraining Hinders New Knowledge Acquisition", "abs": "In this work, we investigate how a model's tendency to broadly integrate its parametric knowledge evolves throughout pretraining, and how this behavior affects overall performance, particularly in terms of knowledge acquisition and forgetting. We introduce the concept of knowledge entropy, which quantifies the range of memory sources the model engages with; high knowledge entropy indicates that the model utilizes a wide range of memory sources, while low knowledge entropy suggests reliance on specific sources with greater certainty. Our analysis reveals a consistent decline in knowledge entropy as pretraining advances. We also find that the decline is closely associated with a reduction in the model's ability to acquire and retain knowledge, leading us to conclude that diminishing knowledge entropy (smaller number of active memory sources) impairs the model's knowledge acquisition and retention capabilities. We find further support for this by demonstrating that increasing the activity of inactive memory sources enhances the model's capacity for knowledge acquisition and retention.", "keywords": [ "knowledge entropy", "knowledge acquisition and forgetting", "evolving behavior during LLM pretraining" ], "rating_list": [ 8, 8, 8, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "eHEYwrN4lw", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Jongheon Jeong", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Assistant Professor" }, { "name": "Subin Kim", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Tae-hyun Oh", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Associate Professor" }, { "name": "Nam Hyeon Woo", "gender": "Male", "institution": "Helmholtz Zentrum M\u00fcnchen", "country": "DE", "position": "Intern" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 23, "n_ref": 85, "n_ref_all": 131, "n_fig": 27, "n_tab": 4, "L_tab": 868, "n_element_tab": 126, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 151, "n_element_tab_1": 24, "formula_len_all": 595, "formula_len_all_1": 498, "len_all": 153964, "len_all_1": 63737, "len_abs": 1508, "len_title": 120, "len_sents": 46442, "len_sents_1": 28851, "n_sents": 444, "n_sents_1": 260, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1426, "title": "Distinct and Shared Concept Discovery for Fine-grained Concept Inversion", "abs": "A real-world object is expressed by composing distinctive characteristics that distinguish it from others and some common properties shared with different objects. Recent advances in generative modeling focus on identifying the shared concepts within images of individual identities. However, it remains unclear how to identify shared concepts beyond multiple identities while preserving the unique concepts inherent to each. In this work, we address this new problem of simultaneously discovering similarities and differences between two sets of images and propose a two-stage framework coined DISCOD (DIstinct and Shared COncept Discovery). In the first stage of DISCOD, we introduce information-regularized textual inversion, focusing on separating representative concepts distinctive from others while capturing the shared concepts among different objects. In the next stage, we further optimize them to align composited concepts of those with the corresponding objects, respectively. We demonstrate the effectiveness of DISCOD by showing that DISCOD discovers the concepts better than baselines, as measured by CLIPScore and success rate. The human study also validates the reasonable discovery capability of DISCOD. Furthermore, we show the practical applicability of our approach by applying to various applications: image editing, few-shot personalization of diffusion models, and group bias mitigation in recognition.", "keywords": [ "Vision language models", "generative model", "concept discovery" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 4, 1, 2, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "eGqQyTAbXC", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jinhua Zhu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Kaiyuan Gao", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Lijun Wu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Qizhi Pei", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Rui Yan", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 49, "n_ref": 209, "n_ref_all": 246, "n_fig": 5, "n_tab": 10, "L_tab": 4902, "n_element_tab": 682, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 3761, "n_element_tab_1": 204, "formula_len_all": 1517, "formula_len_all_1": 1131, "len_all": 217911, "len_all_1": 66258, "len_abs": 1989, "len_title": 117, "len_sents": 54796, "len_sents_1": 26152, "n_sents": 421, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 2013, "title": "Leveraging Discrete Structural Information for Molecule-Text Modeling", "abs": "The integration of molecular and natural language representations has emerged as a focal point in molecular science, with recent advancements in Language Models (LMs) demonstrating significant potential for comprehensive modeling of both domains. However, existing approaches face notable limitations, particularly in their neglect of three-dimensional (3D) information, which is crucial for understanding molecular structures and functions. While some efforts have been made to incorporate 3D molecular information into LMs using external structure encoding modules, significant difficulties remain, such as insufficient interaction across modalities in pre-training and challenges in modality alignment. To address the limitations, we propose \\textbf{3D-MolT5}, a unified framework designed to model molecule in both sequence and 3D structure spaces. The key innovation of our approach lies in mapping fine-grained 3D substructure representations into a specialized 3D token vocabulary. This methodology facilitates the seamless integration of sequence and structure representations in a tokenized format, enabling 3D-MolT5 to encode molecular sequences, molecular structures, and text sequences within a unified architecture. Leveraging this tokenized input strategy, we build a foundation model that unifies the sequence and structure data formats. We then conduct joint pre-training with multi-task objectives to enhance the model's comprehension of these diverse modalities within a shared representation space. Thus, our approach significantly improves cross-modal interaction and alignment, addressing key challenges in previous work. Further instruction tuning demonstrated that our 3D-MolT5 has strong generalization ability and surpasses existing methods with superior performance in multiple downstream tasks, such as nearly 70\\% improvement on molecular property prediction task compared to state-of-the-art methods. Our code is available at \\url{https://anonymous.4open.science/r/3D-MolT5-ICLR2025}.", "keywords": [ "molecule-text modeling", "3D molecular tokenization", "language model" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "eG56H9teXv", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Haim David Fisher", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "MS student" }, { "name": "Moni Shahar", "gender": "Male", "institution": "Tel Aviv University, Tel Aviv University", "country": "IL", "position": "Researcher" }, { "name": "Zeek FR", "gender": "Unspecified", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 19, "n_ref": 23, "n_ref_all": 30, "n_fig": 2, "n_tab": 3, "L_tab": 373, "n_element_tab": 62, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 34, "n_element_tab_1": 1, "formula_len_all": 549, "formula_len_all_1": 549, "len_all": 82267, "len_all_1": 54747, "len_abs": 1981, "len_title": 100, "len_sents": 30467, "len_sents_1": 28438, "n_sents": 213, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 1992, "title": "Neural Fingerprints for Adversarial Attack Detection", "abs": "Deep learning models for image classification have become standard tools in recent years. However, a well known\nvulnerability of these models is their susceptibility to adversarial examples. Adversarial examples are generated \nby slightly altering an image of a certain class in a way that is imperceptible to humans but causes the model to classify it wrongly as another class. Many algorithms have been proposed to address this problem, falling generally into one of two categories: (i) building robust classifiers (ii) directly detecting attacked images. Despite the very good performance of the proposed detectors, we argue that in a white-box setting, where the attacker knows the configuration and weights of the network and the detector, the attacker can overcome the detector by running many examples on a local copy, and sending only examples that were not detected to the actual model. This problem of addressing complete knowledge of the attacker is common in security applications where even a very good model is not sufficient to ensure safety. In this paper we propose to overcome this inherent limitation of any static defence with randomization. To do so, one must generate a very large family of detectors with consistent performance, and select one or more of them randomly for each input. For the individual detectors, we suggest the method of neural fingerprints. In the training phase, for each class we repeatedly sample a tiny random subset of neurons from certain layers of the network, and if their average is sufficiently different between clean and attacked images of the focal class they are considered a fingerprint and added to the detector bank. During test time, we sample fingerprints from the bank associated with the label predicted by the model, and detect attacks using a likelihood ratio test. We evaluate our detectors on ImageNet with different attack methods and model architectures, and show near-perfect detection with low rates of false detection.", "keywords": [ "Deep Learning", "Adversarial Attacks", "Neural Fingerprints" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 1, 1, 2 ], "presentation_list": [ 3, 2, 1 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "eFoj2egr7G", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Linxi Zhao", "gender": "Female", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "PhD student" }, { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Yihe Deng", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Weitong ZHANG", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 67, "n_ref": 175, "n_ref_all": 232, "n_fig": 14, "n_tab": 19, "L_tab": 4489, "n_element_tab": 473, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 1089, "n_element_tab_1": 119, "formula_len_all": 612, "formula_len_all_1": 538, "len_all": 265848, "len_all_1": 69640, "len_abs": 1387, "len_title": 139, "len_sents": 77337, "len_sents_1": 32099, "n_sents": 612, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1424, "title": "Mitigating Object Hallucination in Large Vision-Language Models via Image-Grounded Guidance", "abs": "The advancement of Large Vision-Language Models (LVLMs) has increasingly highlighted the critical issue of their tendency to hallucinate non-existing objects in the images. To address this issue, previous works focused on using specially curated datasets or powerful LLMs (e.g., GPT-3.5) to rectify the outputs of LVLMs. However, these approaches require either expensive training/fine-tuning or API access to advanced LLMs for post-generation correction. In response to these limitations, we propose **M**itigating hallucin**A**tion via image-g**R**ounded gu**I**da**N**c**E** (MARINE), a framework that is both _training-free_ and _API-free_. MARINE effectively and efficiently reduces object hallucinations during inference by introducing image-grounded guidance to LVLMs. This is achieved by leveraging open-source vision models to extract object-level information, thereby enhancing the precision of LVLM-generated content. Our framework's flexibility further allows for the integration of multiple vision models, enabling more reliable and robust object-level guidance. Through comprehensive evaluations across $5$ popular LVLMs with diverse evaluation metrics and benchmarks, we demonstrate the effectiveness of MARINE, which even outperforms existing fine-tuning-based methods. Remarkably, it reduces hallucinations consistently in GPT-4V-assisted evaluation while maintaining the detailedness of LVLMs' generations.", "keywords": [ "Large Vision-Language Models", "Object Hallucination", "Multi-modal LLMs" ], "rating_list": [ 6, 5, 8, 5, 6 ], "soundness_list": [ 3, 2, 4, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "eFGIWUqHQm", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Alireza Akhavanpour", "gender": "Male", "institution": "Shahid Rajaee Teacher Training University", "country": "IR", "position": "Lecturer" }, { "name": "Hassan yousefzase", "gender": "Male", "institution": "Khajeh Nasir Toosi University of Technology", "country": "IR", "position": "MS student" }, { "name": "Mohammad Sabokrou", "gender": "Male", "institution": "Okinawa Institute of Science and Technology (OIST)", "country": "JP", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 22, "n_ref": 49, "n_ref_all": 63, "n_fig": 8, "n_tab": 3, "L_tab": 1252, "n_element_tab": 107, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1135, "n_element_tab_1": 80, "formula_len_all": 612, "formula_len_all_1": 639, "len_all": 140668, "len_all_1": 50724, "len_abs": 2012, "len_title": 134, "len_sents": 27116, "len_sents_1": 22678, "n_sents": 192, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1629, "title": "Region-Aware Generalized Face Anti-Spoofing via Chebyshev Convolutional Graph Networks", "abs": "Face Anti-Spoofing (FAS) is critical for safeguarding face recognition systems from spoofing attacks. However, current methods based on Convolutional Neural Networks (CNNs) and Vision Transformers face limitations in modeling the diverse, region-specific attack behaviors, leading to reduced generalization. This challenge arises due to two main factors: (1) attacks manifest differently across facial regions due to variations in color, texture, and material properties; and (2) the large data space hinders effective generalization.\n\nTo address these issues, we propose a novel approach utilizing Chebyshev Convolutional Graph Neural Networks (ChebConv GNNs) to model spatial information within a graph-based structure. ChebConv is particularly efficient in processing visual data from graphs derived from images. Our method processes regions around facial landmarks through the initial layers of a DenseNet to extract rich, local node features for each region. By assigning a node to each facial region, we construct a unified graph structure where the nodes correspond to the same regions across all faces. This enables the network to model local features and inter-region relationships effectively, reducing the data space and enhancing generalization.\n\nTo further improve generalization across unseen domains, we integrate a Domain-Adversarial Graph Network. Additionally, we introduce an auxiliary self-supervised task to encourage the learning of region-specific texture features. Experimental results demonstrate that our method significantly outperforms existing approaches in terms of both accuracy and generalization.", "keywords": [ "GNN", "face spoofing", "CNN" ], "rating_list": [ 1, 1, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 2, 1, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 2, 2 ], "confidence_list": [ 1, 4, 5, 4, 5 ] }, { "paper_id": "eF1i7YTVen", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chunjie Yang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Feng Yan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Siwei Lou", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "ShaoQi Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 10, "n_ref_uni": 39, "n_ref": 63, "n_ref_all": 102, "n_fig": 15, "n_tab": 7, "L_tab": 4158, "n_element_tab": 192, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2840, "n_element_tab_1": 81, "formula_len_all": 6861, "formula_len_all_1": 2552, "len_all": 169377, "len_all_1": 61339, "len_abs": 1072, "len_title": 119, "len_sents": 49978, "len_sents_1": 27051, "n_sents": 399, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1121, "title": "COPU: Recognizing Time Series' Heterogeneity In Stacked Neural Network", "abs": "Neural networks (NNs) have been widely studied in complex fields due to their remarkable capacity for nonlinear modeling. \nHowever, in the realm of time series analysis, researches indicate that merely stacking NNs does not yield promising nonlinear modeling outputs and hinders model performance. Conventional NN architectures overemphasize homogeneous feature extraction, impeding the learning of diverse features and diminishing their nonlinear modeling capability. To address this gap, we propose the $\\textbf{C}$ross-correlation Enhanced Approximated $\\textbf{O}$rthogonal $\\textbf{P}$rojection $\\textbf{U}$nit (COPU) to quantify and augment the NN's nonlinear modeling capacity. COPU efficiently computes the local cross-correlation characteristics between features, amplifying heterogeneous components while compressing homogeneous ones. By reducing redundant information, COPU facilitates the learning of unique and independent features, thereby enhancing nonlinear modeling capability. Extensive experiments demonstrate that our method achieves superior performance across two real-world regression applications.", "keywords": [ "Neural Network", "Time Series", "Regression", "Nonlinear Modeling", "Natural Gradient" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "eENHKMTOfW", "primary_area": "foundation or frontier models, including LLMs", "n_author": 13, "author_info_list": [ { "name": "Abhishek Bhandwaldar", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Akash Srivastava", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Aldo Pareja", "gender": "Male", "institution": "Delft University of Technology", "country": "", "position": "PhD student" }, { "name": "Guangxuan Xu", "gender": "Male", "institution": "Red Hat. Inc", "country": "US", "position": "Researcher" }, { "name": "Hao Wang", "gender": "Male", "institution": "RedHat AI & MIT-IBM Watson AI Lab", "country": "US", "position": "Researcher" }, { "name": "Kai Xu", "gender": "Male", "institution": "MIT-IBM Watson AI Lab", "country": "", "position": "Research scientist" }, { "name": "Krishnateja Killamsetty", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Ligong Han", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Luke Inglis", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Nikhil Shivakumar Nayak", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Seungwook Han", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Shivchander Sudalairaj", "gender": "Male", "institution": "MIT-IBM Watson AI Lab", "country": "US", "position": "Researcher" }, { "name": "Wenlong Zhao", "gender": "unknown", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 93, "n_ref_all": 133, "n_fig": 24, "n_tab": 21, "L_tab": 4409, "n_element_tab": 486, "n_fig_1": 2, "n_tab_1": 9, "L_tab_1": 1271, "n_element_tab_1": 173, "formula_len_all": 75, "formula_len_all_1": 0, "len_all": 231556, "len_all_1": 68407, "len_abs": 1621, "len_title": 158, "len_sents": 74693, "len_sents_1": 32929, "n_sents": 526, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 110, "L_abs": 1628, "title": "Training Mice to Compete with Elephants: A Guide for Customizing Small-Sized LLMs on Knowledge and Skills Data", "abs": "Customizing large language models (LLMs) is increasingly in demand by enterprises and individual developers. It allows LLMs to be tailored for domain expertise, aligned with organizational guidelines, and enhanced for user experience. Effective customization hinges on three core elements: a small-size model, large-scale domain-specific datasets, and an effective training strategy to help the model acquire relevant knowledge and skills from the data. In this paper, we focus on the third element by conducting an in-depth study on fine-tuning LLMs (3B to 7B parameters) using large-scale instruction tuning datasets across multiple knowledge domains and skills. We examine various training configurations and strategies on three pretrained LLMs. Our results question several common training practices, including hyperparameter recommendations from TULU and phased training recommended by Orca.\nKey insights from our work include: (i) larger batch sizes paired with lower learning rates lead to improved model performance on benchmarks such as MMLU, MTBench, and Open LLM Leaderboard; (ii) early-stage training dynamics, such as lower gradient norms and higher loss values, are strong indicators of better final model performance, allowing for early termination of sub-optimal runs and significant computational savings; (iii) skipping warmup and using a constant learning rate do not compromise performance; and (iv) stacked training outperforms phased training. With these findings holding robustly across model families and sizes, we hope this study serves as a comprehensive guide for practitioners fine-tuning small LLMs.", "keywords": [ "Machine Learning", "Generative Models", "Large Language Models", "Natural Language Processing", "Transformers", "Fine-Tuning", "Instruction Tuning", "Synthetic Data Generation", "Knowledge Data", "Skills Data", "Model Generalization", "Batch Size", "Hyperparameter Optimization", "Gradient Norm", "MMLU", "MTBench", "Stacked Training", "Phased Training", "Compute Efficiency", "Sample Efficiency", "Flash Attention", "Multipack Bucketing" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "eE2PXlNydB", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Hongyu Qu", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jianan Wei", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Wenguan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Xiangbo Shu", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 76, "n_ref": 197, "n_ref_all": 231, "n_fig": 8, "n_tab": 8, "L_tab": 3159, "n_element_tab": 419, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1757, "n_element_tab_1": 183, "formula_len_all": 1164, "formula_len_all_1": 1024, "len_all": 199942, "len_all_1": 79452, "len_abs": 1391, "len_title": 121, "len_sents": 51598, "len_sents_1": 33755, "n_sents": 384, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1426, "title": "Learning Clustering-based Prototypes for Compositional Zero-Shot Learning", "abs": "Learning primitive (i.e., attribute and object) concepts from seen compositions is the primary challenge of Compositional Zero-Shot Learning (CZSL). Existing CZSL solutions typically rely on oversimplified data assumptions, e.g., modeling each primitive with a single centroid primitive presentation, ignoring the natural diversities of the attribute (resp. object) when coupled with different objects (resp. attribute). In this work, we develop ClusPro, a robust clustering-based prototype mining framework for CZSL that defines the conceptual boundaries of primitives through a set of diversified prototypes. Specifically, ClusPro conducts within-primitive clustering on the embedding space for automatically discovering and dynamically updating prototypes. To learn high-quality embeddings for discriminative prototype construction, ClusPro repaints a well-structured and independent primitive embedding space, ensuring intra-primitive separation and inter-primitive decorrelation through prototype-based contrastive learning and decorrelation learning. Moreover, ClusPro effectively performs prototype clustering in a non-parametric fashion without the introduction of additional learnable parameters or computational budget during testing. Experiments on three benchmarks demonstrate ClusPro outperforms various top-leading CZSL solutions under both closed-world and open-world settings. Our code is available at CLUSPRO.", "keywords": [ "Compositional Zero-Shot Learning", "Prototype Learning", "Representation Disentanglement" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "eDnslTIWSt", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Andrew Bai", "gender": "unknown", "institution": ", University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Cho-Jui Hsieh", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Nanyun Peng", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Associate Professor" }, { "name": "Yihan Wang", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 25, "n_ref": 53, "n_ref_all": 82, "n_fig": 5, "n_tab": 7, "L_tab": 3800, "n_element_tab": 286, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 243, "n_element_tab_1": 22, "formula_len_all": 416, "formula_len_all_1": 262, "len_all": 189463, "len_all_1": 56340, "len_abs": 1356, "len_title": 115, "len_sents": 40055, "len_sents_1": 28659, "n_sents": 289, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1377, "title": "On the loss of context-awareness in general instruction finetuning", "abs": "Pretrained Large Language Models (LLMs) require post-training methods such as supervised fine-tuning (SFT) on instruction-response pairs to enable instruction following. \nHowever, this process can potentially harm existing capabilities learned during pretraining. \nIn this paper, we investigate the loss of context awareness after SFT, defined as the capability to extract and understand information from the user-provided context and respond accordingly. \nWe are the first to identify and show that the loss of context-awareness appears on instruction-finetuned LLMs when the chat template is applied to the input prompts. \nWe identify the performance decline is partially caused by the bias embedded into the chat template to focus less on the the user-provided context.\nBased on these observations, we propose two methods to mitigate the loss of context awareness in instruct models: post hoc attention steering on user prompts and conditional instruction fine-tuning with a context-dependency indicator.\nEmpirical experiments on 4 context-dependent downstream tasks and 3 pretrained LLMs of different sizes show that our methods can effectively mitigate the loss of context awareness without compromising the general ability of instruction following. \nOur findings also strongly advocate the necessity to benchmark context awareness after instruction fine-tuning carefully.", "keywords": [ "instruction finetuning", "large language models" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "eDduYIUgHk", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Ryan Grainger", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "PhD student" }, { "name": "Thomas Paniagua", "gender": "unknown", "institution": "North Carolina State University", "country": "US", "position": "PhD student" }, { "name": "Tianfu Wu", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 21, "n_ref_uni": 45, "n_ref": 75, "n_ref_all": 114, "n_fig": 4, "n_tab": 12, "L_tab": 14446, "n_element_tab": 488, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 4340, "n_element_tab_1": 143, "formula_len_all": 1320, "formula_len_all_1": 1541, "len_all": 169823, "len_all_1": 70441, "len_abs": 3377, "len_title": 117, "len_sents": 41235, "len_sents_1": 27714, "n_sents": 351, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1571, "title": "Revisiting and Expanding Targeted Universal Adversarial Perturbations", "abs": "Universal adversarial perturbations (UAPs) have deepened the vulnerability concern of Deep Neural Networks (DNNs) after the initial intriguing discovery of vanilla single-model-single-image adversarial attacks. However, the landscape of UAPs has not been thoroughly investigated. In this paper, we revisit and expand UAPs for white-box targeted attacks along three axes simultaneously: the model-axis, the data-axis, and the target-axis. For the target-axis, we adopt the most aggressive ordered top-$K$ attack protocol ($K\\geq 1$) to expand the traditional top-$1$ attack setting in the prior art of learning UAPs. Our proposed method is thus dubbed as AllAttacK. \nIn implementation, our AllAttacK is built on two state-of-the-art single-model-single-image ordered top-$K$ attack methods, the KL divergence based adversarial distillation method and the more recently proposed quadratic programming based method. We propose a simple yet effective joint mini-data-batch and mini-model-batch optimization strategy in learning UAPs for a large number of models (e.g., up to 18 disparate DNNs) and a large number of images (e.g., 1000 images). We test our AllAttacK on the ImageNet-1k classification task using an ensemble of disparate models such as Convolutional Neural Networks and their adversarially-robustified versions, Vision Transformers, CLIP vision encoders, and MLP-Mixers. Our learned AllAttacK perturbations are doubly transferable across training and testing models, and across training and testing images, and they also show intriguing yet sensible looking.", "keywords": [ "Universal Adversarial Perturbations", "Ordered Top-K AttacKs", "AllAttacK" ], "rating_list": [ 6, 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 4, 3 ] }, { "paper_id": "eDJsL1qAxw", "primary_area": "datasets and benchmarks", "n_author": 15, "author_info_list": [ { "name": "B. Aditya Prakash", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Fanxing Liu", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "PhD student" }, { "name": "Haoxin Liu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Kaize Ding", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "Assistant Professor" }, { "name": "Linglong Qian", "gender": "Male", "institution": "King's College London, University of London", "country": "", "position": "PhD student" }, { "name": "Wenjia Wang", "gender": "Male", "institution": "HKUST (GZ)", "country": "CN", "position": "Assistant Professor" }, { "name": "Wenjie Du", "gender": "Male", "institution": "PyPOTS Research", "country": "CA", "position": "Researcher" }, { "name": "Yingjie Zhou", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "Associate Professor" }, { "name": "Yiyuan Yang", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "", "position": "PhD student" }, { "name": "Yuxuan Liang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Zepu Wang", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Zhiyuan Zhao", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zina Ibrahim", "gender": "Female", "institution": "King's College London, University of London", "country": "", "position": "Associate Professor" }, { "name": "Jun WANG", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Qingsong Wen", "gender": "Male", "institution": "Squirrel Ai Learning", "country": "US", "position": "Principal Researcher" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 56, "n_ref": 118, "n_ref_all": 144, "n_fig": 7, "n_tab": 28, "L_tab": 91251, "n_element_tab": 3398, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 758, "n_element_tab_1": 103, "formula_len_all": 299, "formula_len_all_1": 26, "len_all": 346999, "len_all_1": 53032, "len_abs": 1599, "len_title": 94, "len_sents": 54519, "len_sents_1": 27188, "n_sents": 344, "n_sents_1": 158, "venue": "ICLR2025", "decision": "withdraw", "L_title": 46, "L_abs": 1385, "title": "TSI-Bench: Benchmarking Time Series Imputation", "abs": "Effective imputation is a crucial preprocessing step for time series analysis. Despite the development of numerous deep learning algorithms for time series imputation, the community lacks standardized and comprehensive benchmark platforms to effectively evaluate imputation performance across different settings. Moreover, although many deep learning forecasting algorithms have demonstrated excellent performance, whether their modeling achievements can be transferred to time series imputation tasks remains unexplored. To bridge these gaps, we develop TSI-Bench, the first (to our knowledge) comprehensive benchmark suite for time series imputation utilizing deep learning techniques. The TSI-Bench pipeline standardizes experimental settings to enable fair evaluation of imputation algorithms and identification of meaningful insights into the influence of domain-appropriate missing rates and patterns on model performance. Furthermore, TSI-Bench innovatively provides a systematic paradigm to tailor time series forecasting algorithms for imputation purposes. Our extensive study across 34,804 experiments, 28 algorithms, and 8 datasets with diverse missingness scenarios demonstrates TSI-Bench's effectiveness in diverse downstream tasks and potential to unlock future directions in time series imputation research and analysis. All source code and experiment logs are released.", "keywords": [ "Time Series", "Missing Data", "Imputation", "Deep Learning", "Benchmark" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "eC2a2IndIt", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Fei Wang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Muhao Chen", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Assistant Professor" }, { "name": "Terry Tong", "gender": "Male", "institution": " University of California, Davis", "country": "US", "position": "Undergrad student" }, { "name": "Zhe Zhao", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 7, "n_ref_uni": 35, "n_ref": 75, "n_ref_all": 85, "n_fig": 4, "n_tab": 11, "L_tab": 2184, "n_element_tab": 372, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 805, "n_element_tab_1": 141, "formula_len_all": 986, "formula_len_all_1": 468, "len_all": 225117, "len_all_1": 67449, "len_abs": 1941, "len_title": 89, "len_sents": 71214, "len_sents_1": 29539, "n_sents": 592, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1953, "title": "BadJudge: Backdoor Vulnerabilities of LLM-As-A-Judge", "abs": "This paper exposes the backdoor threat in automatic evaluation with LLM-as-a-Judge. We propose a novel threat model, where the adversary assumes control of both the candidate model and evaluator model. The victims are the benign users who are unfairly rated by the backdoored evaluator in favor of the adversary. A trivial single token backdoor poisoning 1% of the evaluator training data triples the score of the adversary, compared with the adversary's legitimate score. We systematically categorize levels of data access corresponding to three real-world settings, (1) web poisoning, (2) malicious annotator, and (3) weight poisoning. These levels reflect a weak to strong escalation of access that highly correlates with attack severity. The (1) web poisoning setting contains the weakest assumptions, but we still inflate scores by over 20%. The (3) weight poisoning setting holds the strongest assumptions, allowing the adversary to inflate their scores from 1.5/5 to 4.9/5. The backdoor threat generalizes to different evaluator architectures, trigger designs, evaluation tasks, and poisoning rates. By poisoning 10% of the evaluator training data, we control toxicity judges (Guardrails) to misclassify toxic prompts as non-toxic 89% of the time and document reranker judges in RAG to rank the poisoned document first 97% of the time. Defending the LLM-as-a-Judge setting is challenging because false positive errors are unethical. This limits the available tools we can use for defense. Fortunately, we find that model merging, a simple and principled knowledge transfer tool used to imbue LLM Judges with multi-task evaluation abilities, works surprisingly well. The backdoor is mitigated to near 0% whilst maintaining SOTA performance. Its low computational cost, convenient integration into the current LLM Judge training pipeline, and SOTA performance position it as a promising avenue for backdoor mitigation in the LLM-as-a-Judge setting.", "keywords": [ "LLM-as-a-Judge", "LLM Evaluator", "Backdoor Attack", "Backdoor Defense" ], "rating_list": [ 5, 8, 6, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 1, 3, 2, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "eC2ICbECNM", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Guiyu Zhang", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Researcher" }, { "name": "Hao Zhao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Huan-ang Gao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhedong Zheng", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Assistant Professor" }, { "name": "Zijian Jiang", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 79, "n_ref_all": 101, "n_fig": 10, "n_tab": 3, "L_tab": 1649, "n_element_tab": 207, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 5069, "n_element_tab_1": 138, "formula_len_all": 489, "formula_len_all_1": 688, "len_all": 132836, "len_all_1": 70517, "len_abs": 1873, "len_title": 129, "len_sents": 39365, "len_sents_1": 31508, "n_sents": 311, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1882, "title": "Ctrl-U: Robust Conditional Image Generation via Uncertainty-aware Reward Modeling", "abs": "In this paper, we focus on the task of conditional image generation, where an image is synthesized according to user instructions. The critical challenge underpinning this task is ensuring both the fidelity of the generated images and their semantic alignment with the provided conditions. To tackle this issue, previous studies have employed supervised perceptual losses derived from pre-trained models, i.e., reward models, to enforce alignment between the condition and the generated result. However, we observe one inherent shortcoming: considering the diversity of synthesized images, the reward model usually provides inaccurate feedback when encountering newly generated data, which can undermine the training process. To address this limitation, we propose an uncertainty-aware reward modeling, called Ctrl-U, including uncertainty estimation and uncertainty-aware regularization, designed to reduce the adverse effects of imprecise feedback from the reward model. Given the inherent cognitive uncertainty within reward models, even images generated under identical conditions often result in a relatively large discrepancy in reward loss. Inspired by the observation, we explicitly leverage such prediction variance as an uncertainty indicator. Based on the uncertainty estimation, we regularize the model training by adaptively rectifying the reward. In particular, rewards with lower uncertainty receive higher loss weights, while those with higher uncertainty are given reduced weights to allow for larger variability. The proposed uncertainty regularization facilitates reward fine-tuning through consistency construction. Extensive experiments validate the effectiveness of our methodology in improving the controllability and generation quality, as well as its scalability across diverse conditional scenarios, including segmentation mask, edge, and depth conditions.", "keywords": [ "Conditional generation", "semantic alignment", "uncertainty learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 2, 2 ] }, { "paper_id": "eBgIzHaegm", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Haochen Zhao", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Hui Meng", "gender": "Female", "institution": "Hangzhou Institute for Advanced Study, University of the Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Jianwei Niu", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaozheng Xie", "gender": "Female", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "Postdoc" }, { "name": "Xuefeng Liu", "gender": "Male", "institution": "Behang University", "country": "CN", "position": "Associate Professor" }, { "name": "Yong Wang", "gender": "Male", "institution": "Peking Union Medical College Hospital", "country": "CN", "position": "Full Professor" } ], "n_formula": 36, "n_formula_1": 29, "n_ref_uni": 34, "n_ref": 123, "n_ref_all": 157, "n_fig": 5, "n_tab": 13, "L_tab": 3889, "n_element_tab": 243, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 3536, "n_element_tab_1": 79, "formula_len_all": 2992, "formula_len_all_1": 2541, "len_all": 197460, "len_all_1": 64737, "len_abs": 2039, "len_title": 191, "len_sents": 50195, "len_sents_1": 25256, "n_sents": 438, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 142, "L_abs": 2055, "title": "Keep Your Friends Close, and Your Enemies Farther: Distance-aware Voxel-wise Contrastive Learning for Semi-supervised Multi-organ Segmentation", "abs": "Voxel-wise contrastive learning (VCL) is a prominent approach in semi-supervised medical image segmentation. Based on the initially generated pseudo-labels, VCL pulls voxels with the same pseudo-labels toward their prototypes while pushes those with different labels apart, thereby learns effective representations for the segmentation task. However, in multi-organ segmentation (MoS), the complex anatomical structures of certain organs often lead to many unreliable pseudo-labels. Directly applying VCL can introduce confirmation bias, resulting in poor segmentation performance. A common practice is to first transform these unreliable pseudo-labels into more reliable complementary ones, which represent classes that voxels are least likely to belong to, and then push voxels away from the prototypes of their complementary labels. However, we find that in this approach, if voxels with unreliable pseudo-labels are originally close in feature space, they can end up far apart after being pushed away from their complementary prototypes. This disruption of the semantic relationships among voxels can be detrimental to the MoS task. In this paper, we propose DVCL, a novel distance-aware VCL method for semi-supervised MoS. DVCL is based on the observation that voxels close to each other in the feature space ('neighbors') likely belong to the same semantic category, while distant ones ('outsiders') likely belong to different categories. In DVCL, we first identify neighbors and outsiders for all voxels with unreliable pseudo-labels, and then pull their neighbors into the same clusters while pushing outsiders away. In this way, neighbors of unreliable voxels remain their neighbors and outsiders remain outsiders. This approach helps maintain useful semantic relationships among unreliable voxels while still enjoying the advantages of VCL. We conduct extensive experiments on four datasets to validate the effectiveness. Extensive experiments on four datasets demonstrate the superior performance of DVCL compared to state-of-the-art methods.", "keywords": [ "Semi-supervised Learning; Contrastive Learning\uff1bMulti-organ Segmentation\uff1bMedical Image Segmentation\uff1b" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "eBVCZj3RZN", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Henrik Bostr\u00f6m", "gender": "unknown", "institution": "KTH Royal Institute of Technology, Stockholm, Sweden", "country": "SE", "position": "Professor" }, { "name": "Michalis Vazirgiannis", "gender": "Male", "institution": "Ecole Polytechnique, France", "country": "", "position": "Full Professor" }, { "name": "Roman Bresson", "gender": "Not Specified", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "Postdoc" }, { "name": "Amr Mehasseb Alkhatib", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 20, "n_ref_uni": 55, "n_ref": 98, "n_ref_all": 146, "n_fig": 11, "n_tab": 13, "L_tab": 10711, "n_element_tab": 735, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3282, "formula_len_all_1": 1538, "len_all": 201253, "len_all_1": 65075, "len_abs": 949, "len_title": 87, "len_sents": 50851, "len_sents_1": 27878, "n_sents": 367, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 39, "L_abs": 952, "title": "Prediction Via Shapley Value Regression", "abs": "Shapley values have several desirable properties for explaining black-box model predictions, which come with strong theoretical support. Traditionally, Shapley values are computed post-hoc, leading to additional computational cost at inference time. To overcome this, we introduce ViaSHAP, a novel approach that learns a function to compute Shapley values, from which the predictions can be derived directly by summation. We explore two learning approaches based on the universal approximation theorem and the Kolmogorov-Arnold representation theorem. Results from a large-scale empirical investigation are presented, in which the predictive performance of ViaSHAP is compared to state-of-the-art algorithms for tabular data, where the implementation using Kolmogorov-Arnold Networks showed a superior performance. It is also demonstrated that the explanations of ViaSHAP are accurate, and that the accuracy is controllable through the hyperparameters.", "keywords": [ "Explainable Machine Learning", "Neural Networks", "Kolmogorov\u2013Arnold Networks" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 4, 3, 1 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "eBS3dQQ8GV", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Andrea Agazzi", "gender": "Male", "institution": "Universit\u00e4t Bern", "country": "CH", "position": "Associate Professor" }, { "name": "Federico Pasqualotto", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Giuseppe Bruno", "gender": "Male", "institution": "Universit\u00e4t Bern", "country": "CH", "position": "PhD student" } ], "n_formula": 144, "n_formula_1": 29, "n_ref_uni": 29, "n_ref": 53, "n_ref_all": 101, "n_fig": 6, "n_tab": 2, "L_tab": 139, "n_element_tab": 4, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12053, "formula_len_all_1": 1723, "len_all": 190268, "len_all_1": 68242, "len_abs": 1132, "len_title": 116, "len_sents": 62774, "len_sents_1": 32184, "n_sents": 614, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 68, "L_abs": 1132, "title": "Emergence of meta-stable clustering in mean-field transformer models", "abs": "We model the evolution of tokens within a deep stack of Transformer layers as a continuous-time flow on the unit sphere, governed by a mean-field interacting particle system, building on the framework introduced in Geshkovski et al. (2023). Studying the corresponding mean-field Partial Differential Equation (PDE), which can be interpreted as a Wasserstein gradient flow, in this paper we provide a mathematical investigation of the long-term behavior of this system, with a particular focus on the emergence and persistence of meta-stable phases and clustering phenomena, key elements in applications like next-token prediction. More specifically, we perform a perturbative analysis of the mean-field PDE around the iid uniform initialization and prove that, in the limit of large number of tokens, the model remains close to a meta-stable manifold of solutions with a given structure (e.g., periodicity). Further, the structure characterizing the meta-stable manifold is explicitly identified, as a function of the inverse temperature parameter of the model, by the index maximizing a certain rescaling of Gegenbauer polynomials.", "keywords": [ "Mean-field limits", "Transformers", "Meta-stability", "Clustering" ], "rating_list": [ 8, 5, 10, 8, 8 ], "soundness_list": [ 4, 3, 4, 3, 3 ], "presentation_list": [ 4, 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 4, 3, 3 ], "confidence_list": [ 2, 1, 3, 2, 3 ] }, { "paper_id": "eB7T1bqthA", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Carlee Joe-Wong", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Ishank Juneja", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Osman Yagan", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "", "position": "Research Professor" } ], "n_formula": 277, "n_formula_1": 10, "n_ref_uni": 11, "n_ref": 35, "n_ref_all": 111, "n_fig": 5, "n_tab": 13, "L_tab": 5791, "n_element_tab": 622, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 211, "n_element_tab_1": 17, "formula_len_all": 28166, "formula_len_all_1": 2182, "len_all": 302691, "len_all_1": 66701, "len_abs": 1540, "len_title": 133, "len_sents": 100192, "len_sents_1": 31984, "n_sents": 913, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1550, "title": "Pairwise Elimination with Instance-Dependent Guarantees for Bandits with Cost Subsidy", "abs": "Multi-armed bandits (MAB) are commonly used in sequential online decision-making when the reward of each decision is an unknown random variable. In practice, however, the typical goal of maximizing total reward may be less important than minimizing the total cost of the decisions taken, subject to a reward constraint. For example, we may seek to make decisions that have at least the reward of a reference ``default'' decision. This problem was recently introduced in the Multi-Armed Bandits with Cost Subsidy (MAB-CS) framework. MAB-CS is broadly applicable to problem domains where a primary metric (cost) is constrained by a secondary metric (reward), and there is an inability to explicitly determine the trade-off between these metrics. In our work, we first introduce the Pairwise-Elimination algorithm for a simplified variant of the cost subsidy problem with a known reference arm. We then generalize PE to PE-CS to solve the MAB-CS problem in the setting where the reference arm is the un-identified optimal arm. Next, we analyze the performance of both PE and PE-CS on the dual metrics of Cost and Quality Regret. Our instance-dependent analysis of PE and PE-CS reveals that both algorithms have an order-wise logarithmic upper bound on Cost and Quality Regret, making our policy the first with such a guarantee. Finally, experiments are conducted using the MovieLens 25M dataset for both PE and PE-CS and using a synthetic toy experiment for PE-CS revealing that our method invariably outperforms the ETC-CS baseline from the literature.", "keywords": [ "Bandits", "Multi-Armed Bandits", "Online Learning", "Cost-Subsidy", "Cost Subsidy", "Improved UCB", "UCB", "Elimination Algorithms" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 2, 2 ] }, { "paper_id": "eB2QgsohdN", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 14, "author_info_list": [ { "name": "Ana Carolina Leote", "gender": "unknown", "institution": "Genentech", "country": "US", "position": "Postdoc" }, { "name": "Aviv Regev", "gender": "Female", "institution": "Genentech / Roche", "country": "US", "position": "Researcher" }, { "name": "Burkhard Hoeckendorf", "gender": "unknown", "institution": "Genentech", "country": "US", "position": "Researcher" }, { "name": "David Richmond", "gender": "Male", "institution": "Genentech/Roche", "country": "US", "position": "Researcher" }, { "name": "Heming Yao", "gender": "Female", "institution": "Genentech, Inc", "country": "US", "position": "Researcher" }, { "name": "Jan-Christian Huetter", "gender": "Male", "institution": "Genentech", "country": "US", "position": "Researcher" }, { "name": "Ji Won Park", "gender": "Female", "institution": "Genentech", "country": "US", "position": "Researcher" }, { "name": "Kyunghyun Cho", "gender": "Male", "institution": "Genentech", "country": "", "position": "Senior Director of Frontier Research" }, { "name": "Natasa Tagasovska", "gender": "Female", "institution": "Prescient Design - Genentech, Roche", "country": "", "position": "Senior Machine Learning Scientis" }, { "name": "Paula Coelho", "gender": "Female", "institution": "Genentech", "country": "", "position": "Postdoc" }, { "name": "Romain Lopez", "gender": "unknown", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Stephen Ra", "gender": "Male", "institution": "Prescient Design, Genentech", "country": "US", "position": "Director of Frontier Research" }, { "name": "TAKAMASA KUDO", "gender": "Male", "institution": "Genenetech, Inc", "country": "US", "position": "Postdoc" }, { "name": "Taro Makino", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 63, "n_ref": 156, "n_ref_all": 173, "n_fig": 11, "n_tab": 1, "L_tab": 27, "n_element_tab": 1, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 769, "formula_len_all_1": 908, "len_all": 158093, "len_all_1": 67435, "len_abs": 1624, "len_title": 92, "len_sents": 44775, "len_sents_1": 32519, "n_sents": 347, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1659, "title": "Supervised Contrastive Block Disentanglement", "abs": "Real-world datasets often combine data collected under different experimental conditions. Although this yields larger datasets, it also introduces spurious correlations that make it difficult to accurately model the phenomena of interest. We address this by learning two blocks of latent variables to independently represent the phenomena of interest and the spurious correlations. The former are correlated with the target variable $y$ and invariant to the environment variable $e$, while the latter depend on $e$. The invariance of the phenomena of interest to $e$ is highly sought-after but difficult to achieve on real-world datasets. Our primary contribution is an algorithm called Supervised Contrastive Block Disentanglement (SCBD) that is highly effective at enforcing this invariance. It is based purely on supervised contrastive learning, and scales to real-world data better than existing approaches. We empirically validate SCBD on two challenging problems. The first is domain generalization, where we achieve strong performance on a synthetic dataset, as well as on Camelyon17-WILDS. SCBD introduces a single hyperparameter $\\alpha$ that controls the degree of invariance to $e$. When we increase $\\alpha$ to strengthen the degree of invariance, there is a monotonic improvement in out-of-distribution performance at the expense of in-distribution performance. The second is a scientific problem of batch correction. Here, we demonstrate the utility of SCBD by learning representations of single-cell perturbations from 26 million Optical Pooled Screening images that are nearly free of technical artifacts induced by the variation across wells.", "keywords": [ "disentanglement", "block disentanglement", "out-of-distribution generalization", "domain generalization", "distribution shift", "spurious correlations", "robustness" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "eAisRJ7AiF", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Jun Zhai", "gender": "Female", "institution": "Illinois Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Muberra Ozmen", "gender": "Female", "institution": "Block", "country": "CA", "position": "Researcher" }, { "name": "Thomas Markovich", "gender": "unknown", "institution": "Block", "country": "US", "position": "Principal Researcher" } ], "n_formula": 37, "n_formula_1": 25, "n_ref_uni": 45, "n_ref": 77, "n_ref_all": 94, "n_fig": 4, "n_tab": 8, "L_tab": 2774, "n_element_tab": 227, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1697, "n_element_tab_1": 141, "formula_len_all": 1997, "formula_len_all_1": 1374, "len_all": 141576, "len_all_1": 69569, "len_abs": 1225, "len_title": 123, "len_sents": 40959, "len_sents_1": 30203, "n_sents": 316, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1229, "title": "Understanding the Design Principles of Link Prediction in Directed Settings", "abs": "Link prediction is a widely studied task in Graph Representation Learning (GRL) for modeling relational data. Early theories in GRL were based on the assumption of a symmetric adjacency matrix, reflecting an undirected setting. As a result, much of the following state-of-the-art research has continued to operate under this symmetry assumption, even though real-world data often involves crucial information conveyed through the direction of relationships. This oversight limits the ability of these models to fully capture the complexity of directed interactions. In this paper, we focus on the challenge of directed link prediction by evaluating key heuristics that have been successful in the undirected settings. We propose simple but effective adaptations of these heuristics to the directed link prediction task and demonstrate that these modifications yield competitive performance compared to leading Graph Neural Networks (GNNs) originally designed for undirected graphs. Through an extensive set of experiments, we derive insights that inform the development of a novel framework for directed link prediction, which not only surpasses baseline methods but also outperforms state-of-the-art GNNs on multiple benchmarks.", "keywords": [ "directed link prediction", "graph representation learning", "link prediction heuristics", "directed graph neural networks" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 4, 2, 4 ], "contribution_list": [ 1, 3, 2, 1 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "eAgnpmxUu1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Haoran Duan", "gender": "Male", "institution": "University of Newcastle-upon-Tyne", "country": "GB", "position": "Postdoc" }, { "name": "Rajiv Ranjan", "gender": "Male", "institution": "University of Newcastle-upon-Tyne", "country": "GB", "position": "Full Professor" }, { "name": "Tejal Shah", "gender": "unknown", "institution": "University of Newcastle-upon-Tyne", "country": "GB", "position": "Lecturer" }, { "name": "Varun Ojha", "gender": "unknown", "institution": "Newcastle University, UK", "country": "GB", "position": "Associate Professor" }, { "name": "Xingyu Miao", "gender": "Male", "institution": "Durham University", "country": "GB", "position": "PhD student" }, { "name": "Yang Long", "gender": "Male", "institution": "Durham University", "country": "GB", "position": "Associate Professor" } ], "n_formula": 21, "n_formula_1": 14, "n_ref_uni": 35, "n_ref": 71, "n_ref_all": 93, "n_fig": 11, "n_tab": 1, "L_tab": 295, "n_element_tab": 39, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 373, "n_element_tab_1": 49, "formula_len_all": 2006, "formula_len_all_1": 1163, "len_all": 121726, "len_all_1": 49676, "len_abs": 1364, "len_title": 135, "len_sents": 34925, "len_sents_1": 23053, "n_sents": 289, "n_sents_1": 169, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1369, "title": "Dreamer XL: Towards High-Resolution Text-to-3D Generation via Trajectory Score Matching", "abs": "In this work, we propose a novel Trajectory Score Matching (TSM) method that aims to solve the pseudo ground truth inconsistency problem caused by the accumulated error in Interval Score Matching (ISM) when using the Denoising Diffusion Implicit Models (DDIM) inversion process. Unlike ISM which adopts the inversion process of DDIM to calculate on a single path, our TSM method leverages the inversion process of DDIM to generate two paths from the same starting point for calculation. Since both paths start from the same starting point, TSM can reduce the accumulated error compared to ISM, thus alleviating the problem of pseudo ground truth inconsistency. TSM enhances the stability and consistency of the model's generated paths during the distillation process. We demonstrate this experimentally and further show that ISM is a special case of TSM. Furthermore, to optimize the current multi-stage optimization process from high-resolution text to 3D generation, we adopt Stable Diffusion XL for guidance. In response to the issues of abnormal replication and splitting caused by unstable gradients during the 3D Gaussian splatting process when using Stable Diffusion XL, we propose a pixel-by-pixel gradient clipping method. Extensive experiments show that our model significantly surpasses the state-of-the-art models in terms of visual quality and performance.", "keywords": [ "text-to-3D", "3DGS" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "eAFNJk63KE", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Jalal Etesami", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Assistant Professor" }, { "name": "Kashif Rasul", "gender": "unknown", "institution": "Zalando SE", "country": "DE", "position": "Researcher" }, { "name": "Larkin Liu", "gender": "Not Specified", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Yutong Chao", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" } ], "n_formula": 107, "n_formula_1": 24, "n_ref_uni": 33, "n_ref": 45, "n_ref_all": 53, "n_fig": 0, "n_tab": 6, "L_tab": 1154, "n_element_tab": 71, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 106, "n_element_tab_1": 7, "formula_len_all": 7537, "formula_len_all_1": 2268, "len_all": 212098, "len_all_1": 66926, "len_abs": 1207, "len_title": 133, "len_sents": 71901, "len_sents_1": 29625, "n_sents": 621, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1233, "title": "Riemannian Manifold Learning for Stackelberg Games with Neural Flow Representations", "abs": "We present a novel framework for online learning in Stackelberg general-sum games, where two agents, the leader and follower, engage in sequential turn-based interactions. At the core of this approach is a learned diffeomorphism that maps the joint action space to a smooth Riemannian manifold, referred to as the $\\textit{Stackelberg manifold}$. This mapping, facilitated by neural normalizing flows, ensures the formation of tractable isoplanar subspaces, enabling efficient techniques for online learning. By assuming linearity between the agents' reward functions on the $\\textit{Stackelberg manifold}$, our construct allows the application of standard bandit algorithms. We then provide a rigorous theoretical basis for regret minimization on convex manifolds and establish finite-time bounds on simple regret for learning Stackelberg equilibria. This integration of manifold learning into game theory uncovers a previously unrecognized potential for neural normalizing flows as an effective tool for multi-agent learning. We present empirical results demonstrating the effectiveness of our approach compared to standard baselines, with applications spanning domains such as cybersecurity and economic supply chain optimization.", "keywords": [ "Neural Normalizing Flows", "Stackelberg Games", "Riemannian Manifolds" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 4 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "e9yfCY7Q3U", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Chao Du", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Jindong Gu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Min Lin", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Principal Researcher" }, { "name": "Tianyu Pang", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Xiaochun Cao", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Xiaojun Jia", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Yang Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "YIHAO HUANG", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 49, "n_ref": 100, "n_ref_all": 139, "n_fig": 9, "n_tab": 14, "L_tab": 2295, "n_element_tab": 234, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2416, "n_element_tab_1": 150, "formula_len_all": 830, "formula_len_all_1": 526, "len_all": 167412, "len_all_1": 61562, "len_abs": 1413, "len_title": 128, "len_sents": 41293, "len_sents_1": 25557, "n_sents": 377, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1409, "title": "Improved Techniques for Optimization-Based Jailbreaking on Large Language Models", "abs": "Large language models (LLMs) are being rapidly developed, and a key component of their widespread deployment is their safety-related alignment. Many red-teaming efforts aim to jailbreak LLMs, where among these efforts, the Greedy Coordinate Gradient (GCG) attack\u2019s success has led to a growing interest in the study of optimization-based jailbreaking techniques. Although GCG is a significant milestone, its attacking efficiency remains unsatisfactory. In this paper, we present several improved (empirical) techniques for optimization-based jailbreaks like GCG. We first observe that the single target template of \"Sure\" largely limits the attacking performance of GCG; given this, we propose to apply diverse target templates containing harmful self-suggestion and/or guidance to mislead LLMs. Besides, from the optimization aspects, we propose an automatic multi-coordinate updating strategy in GCG (i.e., adaptively deciding how many tokens to replace in each step) to accelerate convergence, as well as tricks like easy-to-hard initialisation. Then, we combine these improved technologies to develop an efficient jailbreak method, dubbed I-GCG. In our experiments, we evaluate on a series of benchmarks (such as NeurIPS 2023 Red Teaming Track). The results demonstrate that our improved techniques can help GCG outperform state-of-the-art jailbreaking attacks and achieve nearly 100% attack success rate.", "keywords": [ "Jailbreaking Attacks", "Large Language Models" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 2, 2, 4, 2 ], "presentation_list": [ 3, 2, 4, 1 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 5, 3 ] }, { "paper_id": "e9iRAkEJQ1", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Qianxiao Li", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Ting Lin", "gender": "unknown", "institution": "School of mathematical Science, Peking University, Peking University", "country": "", "position": "PhD student" }, { "name": "Zuowei Shen", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Full Professor" }, { "name": "Cheng Jingpu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 18, "n_ref_uni": 24, "n_ref": 39, "n_ref_all": 50, "n_fig": 7, "n_tab": 2, "L_tab": 674, "n_element_tab": 57, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2616, "formula_len_all_1": 1587, "len_all": 116050, "len_all_1": 58679, "len_abs": 1181, "len_title": 116, "len_sents": 40092, "len_sents_1": 28590, "n_sents": 334, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1185, "title": "Rate of Approximation by Flows: A Case Study on the Eikonal Equation", "abs": "Previous works have demonstrated the universal approximation capability of residual networks through their continuous idealization as flow maps of dynamical systems. However, informative results on their approximation rates in terms of depth (corresponding to time) are generally lacking. From the viewpoint of approximation theory, a major difficulty in addressing this gap lies in identifying an appropriate target space for the approximation problem. In this paper, we introduce a restrictive but useful target function space comprised of solutions to the eikonal equations, a type of first-order nonlinear partial differential equation, to investigate the approximation rates of flow map families. We provide an estimate of the approximation error within this space, which is notably different from classical rate estimates based directly on the smoothness of target functions. This theoretical result further inspires a new learning-based algorithm for solving the eikonal equation. Experimental results validate the effectiveness of our proposed algorithm, including its robustness to spatial resolution and solution regularity, as well as transferability among similar problems.", "keywords": [ "flow map", "approximation rate", "eikonal equation" ], "rating_list": [ 3, 8, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 1, 3, 4 ] }, { "paper_id": "e92KW6htFO", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Chenghu Zhou", "gender": "Male", "institution": "IGSNRR, Chinese Academy of Sciences, Beijing, China", "country": "", "position": "Full Professor" }, { "name": "Jiaxin Ding", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Luoyi Fu", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "", "position": "Associate Professor" }, { "name": "Shiqing Gao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Xinbing Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 115, "n_formula_1": 35, "n_ref_uni": 27, "n_ref": 52, "n_ref_all": 79, "n_fig": 17, "n_tab": 2, "L_tab": 2, "n_element_tab": 1, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 1964, "n_element_tab_1": 21, "formula_len_all": 11463, "formula_len_all_1": 2694, "len_all": 188847, "len_all_1": 59922, "len_abs": 1595, "len_title": 131, "len_sents": 58131, "len_sents_1": 25154, "n_sents": 533, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1602, "title": "MICE: Memory-driven Intrinsic Cost Estimation for Mitigating Constraint Violations", "abs": "Constrained Reinforcement Learning (CRL) aims to maximize cumulative rewards while satisfying constraints. However, most existing CRL algorithms encounter significant constraint violations during training, limiting their applicability in safety-critical scenarios. In this paper, we identify the underestimation of the cost value function as a key factor contributing to these violations. To address this issue, we propose the Memory-driven Intrinsic Cost Estimation (MICE) method, which introduces intrinsic costs to enhance the cost estimate of unsafe behaviors, thus mitigating the underestimation bias. Our method draws inspiration from human cognitive processes, specifically the concept of flashbulb memory, where vivid memories of dangerous events are retained to prevent potential risks. MICE constructs a memory module to store unsafe trajectories explored by the agent. The intrinsic cost is formulated as the similarity between the current trajectory and the unsafe trajectories stored in memory, assessed by an intrinsic generator. We propose an extrinsic-intrinsic cost value function and optimization objective based on intrinsic cost, along with the corresponding optimization method. Theoretically, we provide convergence guarantees for the new cost value function and establish the worst-case constraint violation for the MICE update, ensuring fewer constraint violations compared to baselines. Extensive experiments validate the effectiveness of our approach, demonstrating a substantial reduction in constraint violations while maintaining policy performance comparable to baselines.", "keywords": [ "reinforcement learning", "constraint optimization", "underestimation", "intrinsic cost" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "e8qXTxMgPg", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "David Woodruff", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Qiuyi Zhang", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Sandeep Silwal", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 91, "n_formula_1": 21, "n_ref_uni": 24, "n_ref": 37, "n_ref_all": 53, "n_fig": 0, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 29, "n_element_tab_1": 4, "formula_len_all": 8552, "formula_len_all_1": 2439, "len_all": 251382, "len_all_1": 71034, "len_abs": 2223, "len_title": 110, "len_sents": 82368, "len_sents_1": 30087, "n_sents": 884, "n_sents_1": 287, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 2257, "title": "Beyond Worst-Case Dimensionality Reduction for Sparse Vectors", "abs": "We study beyond worst-case dimensionality reduction for $s$-sparse vectors (vectors with at most $s$ non-zero coordinates). Our work is divided into two parts, each focusing on a different facet of beyond worst-case analysis:\n\n\\noindent (a) We first consider average-case guarantees for embedding $s$-sparse vectors. Here, a well-known folklore upper bound based on the birthday-paradox states: For any collection $X$ of $s$-sparse vectors in $\\mathbb{R}^d$, there exists a linear map $A: \\mathbb{R}^d \\rightarrow \\mathbb{R}^{O(s^2)}$ which \\emph{exactly} preserves the norm of $99\\%$ of the vectors in $X$ in any $\\ell_p$ norm (as opposed to the usual setting where guarantees hold for all vectors). We provide novel lower bounds showing that this is indeed optimal in many settings. Specifically, any oblivious linear map satisfying similar average-case guarantees must map to $\\Omega(s^2)$ dimensions. The same lower bound also holds for a wider class of sufficiently smooth maps, including `encoder-decoder schemes', where we compare the norm of the original vector to that of a smooth function of the embedding. These lower bounds reveal a surprising separation result for smooth embeddings of sparse vectors, as an upper bound of $O(s \\log(d))$ is possible if we instead use arbitrary functions, e.g., via compressed sensing algorithms.\n\n\n (b) Given these lower bounds, we specialize to sparse \\emph{non-negative} vectors to hopes of improved upper bounds. For a dataset $X$ of non-negative $s$-sparse vectors and any $p \\ge 1$, we can non-linearly embed $X$ to $O(s\\log(|X|s)/\\varepsilon^2)$ dimensions while preserving all pairwise distances in $\\ell_p$ norm up to $1\\pm \\varepsilon$, with no dependence on $p$. Surprisingly, the non-negativity assumption enables much smaller embeddings than arbitrary sparse vectors, where the best known bound suffers an exponential $(\\log |X|)^{O(p)}$ dependence. Our map also guarantees \\emph{exact} dimensionality reduction for the $\\ell_{\\infty}$ norm by embedding $X$ into $O(s\\log |X|)$ dimensions, which is tight. We further give separation results showing that both the non-linearity of $f$ and the non-negativity of $X$ are necessary, and provide downstream algorithmic improvements using our embedding.", "keywords": [ "dimensionality reduction", "sparsity", "johnson lindenstrauss" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 4 ], "contribution_list": [ 4, 2, 2, 3 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "e8c7XDRJcg", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Abhi Kamboj", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Minh N. Do", "gender": "Male", "institution": "VinUniversity", "country": "VN", "position": "Honorary Vice Provost" }, { "name": "Anh Duy Nguyen", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 7, "n_ref_uni": 50, "n_ref": 75, "n_ref_all": 110, "n_fig": 8, "n_tab": 7, "L_tab": 1576, "n_element_tab": 287, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 529, "n_element_tab_1": 117, "formula_len_all": 573, "formula_len_all_1": 342, "len_all": 173124, "len_all_1": 67035, "len_abs": 1324, "len_title": 115, "len_sents": 53909, "len_sents_1": 32539, "n_sents": 383, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1331, "title": "C3T: Cross-modal Transfer Through Time for Human Action Recognition", "abs": "In order to unlock the potential of diverse sensors, we investigate a method to transfer knowledge between modalities using the structure of a unified multimodal representation space for Human Action Recognition (HAR). \nWe formalize and explore an understudied cross-modal transfer setting we term Unsupervised Modality Adaptation (UMA), where the modality used in testing is not used in supervised training, i.e. zero labeled instances of the test modality are available during training. \nWe develop three methods to perform UMA: Student-Teacher (ST), Contrastive Alignment (CA), and Cross-modal Transfer Through Time (C3T).\nOur extensive experiments on various camera+IMU datasets compare these methods to each other in the UMA setting, and to their empirical upper bound in the supervised setting.\nThe results indicate C3T is the most robust and highest performing by at least a margin of 8\\%, and nears the supervised setting performance even in the presence of temporal noise.\nThis method introduces a novel mechanism for aligning signals across time-varying latent vectors, extracted from the receptive field of temporal convolutions.\nOur findings suggest that C3T has significant potential for developing generalizable models for time-series sensor data, opening new avenues for multi-modal learning in various applications.", "keywords": [ "cross-modal transfer", "multimodal learning", "human action recognition", "human activity reocognition", "IMU" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 1, 1, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "e7AUJpP8bV", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Meng Luo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Ruizhe Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Wenhao Chai", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "MS student" }, { "name": "Xiaotian Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zuozhu Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 28, "n_formula_1": 15, "n_ref_uni": 49, "n_ref": 113, "n_ref_all": 142, "n_fig": 8, "n_tab": 10, "L_tab": 5162, "n_element_tab": 736, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1664, "n_element_tab_1": 355, "formula_len_all": 2241, "formula_len_all_1": 1066, "len_all": 228776, "len_all_1": 69099, "len_abs": 1417, "len_title": 100, "len_sents": 61343, "len_sents_1": 27528, "n_sents": 484, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 44, "L_abs": 1298, "title": "PAD: Personalized Alignment at Decoding-time", "abs": "Aligning with personalized preferences, which vary significantly across cultural, educational, and political differences, poses a significant challenge due to the computational costs and data demands of traditional alignment methods. In response, this paper presents Personalized Alignment at Decoding-time (PAD), a novel framework designed to align LLM outputs with diverse personalized preferences during the inference phase, eliminating the need for additional training. By introducing a unique personalized reward modeling strategy, this framework decouples the text generation process from personalized preferences, facilitating the generation of generalizable token-level personalized rewards. The PAD algorithm leverages these rewards to guide the decoding process, dynamically tailoring the base model\u2019s predictions to personalized preferences. Extensive experimental results demonstrate that PAD not only outperforms existing training-based alignment methods in terms of aligning with diverse preferences but also shows significant generalizability to preferences unseen during training and scalability across different base models. This work advances the capability of LLMs to meet user needs in real-time applications, presenting a substantial step forward in personalized LLM alignment.", "keywords": [ "Alignment", "personalization" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 2, 3, 4, 2 ] }, { "paper_id": "e6xFKjo4Cp", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Enhong Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Haoyu Tang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Kai Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "Researcher" }, { "name": "Qi Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xukai Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Yanghai Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Ye Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 60, "n_ref": 107, "n_ref_all": 129, "n_fig": 6, "n_tab": 8, "L_tab": 6357, "n_element_tab": 1309, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2001, "n_element_tab_1": 365, "formula_len_all": 494, "formula_len_all_1": 435, "len_all": 170536, "len_all_1": 68120, "len_abs": 1616, "len_title": 133, "len_sents": 40467, "len_sents_1": 30034, "n_sents": 294, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1637, "title": "Learn while Unlearn: An Iterative Unlearning Framework for Generative Language Models", "abs": "Recent advancements in machine learning, particularly in Natural Language Processing (NLP), have led to the development of sophisticated models trained on extensive datasets, yet raising concerns about the potential leakage of sensitive information. In response, regulatory measures such as the European Union's General Data Protection Regulation (GDPR) have driven increasing interest in Machine Unlearning techniques, which enable models to selectively forget specific data entries. Early approaches primarily relied on pre-processing methods, while more recent research has shifted towards training-based unlearning techniques. Despite their effectiveness, most existing methods require access to the original training data, which is often inaccessible. Additionally, directly applying unlearning techniques bear the cost of undermining the model's expressive capabilities. To address these challenges, we introduce the **I**terative **C**ontrastive **U**nlearning (**ICU**) framework, which consists of three core components: A Knowledge Unlearning Induction module designed to remove specific knowledge through an unlearning loss; A Contrastive Learning Enhancement module to preserve the model's expressive capabilities against the pure unlearning goal; And an Iterative Unlearning Refinement module that dynamically assess the unlearning extent on specific data pieces and make iterative update. Experimental results demonstrate the efficacy of our ICU method in unlearning sensitive information while maintaining the model's overall performance, offering a promising solution for privacy-conscious machine learning applications.", "keywords": [ "machine unlearning", "natural language processing", "generative language model" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "e69qTSwdOT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Chengqi Zhang", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Chair Professor" }, { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Gang Niu", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Senior Research Scientist (tenured)" }, { "name": "Hongduan Tian", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "PhD student" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 93, "n_ref_all": 131, "n_fig": 12, "n_tab": 3, "L_tab": 143, "n_element_tab": 22, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 72, "n_element_tab_1": 12, "formula_len_all": 311, "formula_len_all_1": 285, "len_all": 188078, "len_all_1": 73410, "len_abs": 2227, "len_title": 118, "len_sents": 74167, "len_sents_1": 35155, "n_sents": 583, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1461, "title": "Exploring the Discriminative Capability of LLMs in In-context Learning", "abs": "_In-context learning_ (ICL), as an emergent behavior of large language models (LLMs), has exhibited impressive capability in solving previously unseen tasks based on the observations of the given samples without extra training. However, recent works find that LLMs irregularly obtain unexpected fragmented decision boundaries in simple discriminative tasks, such as binary linear classification. Our observations on the output of Llama-3-8B for the reasoning process of label predictions reveal that LLMs tend to leverage the existing machine learning algorithms to perform discriminative tasks. Specifically, LLMs tend first to select a strategy for the given task and then predict the labels of query data by executing the selected strategy. Based on the observation, in this paper, we propose to dive into such a behavior of LLMs for a deeper understanding of the discriminative capability of LLMs. We conduct a series of analyses on Llama-3-8B to determine the behaviors adopted by LLMs in the discriminative tasks, including probing the label predictions of query data and the corresponding confidence of LLMs under different prompt settings. Moreover, we also probe the preference of LLMs for strategy selection and then simulate the behavior of LLMs performing classification based on obtained preference. The analysis and simulation results provide some important observations and insights into the properties of LLMs in performing discriminative tasks.", "keywords": [ "large language models", "in-context learning", "discriminative capability", "decision boundary" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "e5mTvjXG9u", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Gautam Singh", "gender": "Male", "institution": "Lawrence Livermore National Labs", "country": "US", "position": "Postdoctoral Researcher" }, { "name": "Junyeob Baek", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Sungjin Ahn", "gender": "Male", "institution": "KAIST", "country": "US", "position": "Associate Professor" }, { "name": "Yi-Fu Wu", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 53, "n_ref": 127, "n_ref_all": 152, "n_fig": 11, "n_tab": 7, "L_tab": 1098, "n_element_tab": 142, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 812, "formula_len_all_1": 652, "len_all": 188484, "len_all_1": 68975, "len_abs": 1402, "len_title": 69, "len_sents": 53659, "len_sents_1": 33815, "n_sents": 402, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1407, "title": "Dreamweaver: Learning Compositional World Models from Pixels", "abs": "Humans have an innate ability to decompose their perceptions of the world into objects and their attributes, such as colors, shapes, and movement patterns. This cognitive process enables us to imagine novel futures by recombining familiar concepts. However, replicating this ability in artificial intelligence systems has proven challenging, particularly when it comes to modeling videos into compositional concepts and generating unseen, recomposed futures without relying on auxiliary data, such as text, masks, or bounding boxes. In this paper, we propose Dreamweaver, a neural architecture designed to discover hierarchical and compositional representations from raw videos and generate compositional future simulations. Our approach leverages a novel Recurrent Block-Slot Unit (RBSU) to decompose videos into their constituent objects and attributes. In addition, Dreamweaver uses a multi-future-frame prediction objective to capture disentangled representations for dynamic concepts more effectively as well as static concepts. In experiments, we demonstrate our model outperforms current state-of-the-art baselines for world modeling when evaluated under the DCI framework across multiple datasets. Furthermore, we show how the modularized concept representations of our model enable compositional imagination, allowing the generation of novel videos by recombining attributes from different objects.", "keywords": [ "compositional world models", "unsupervised object-centric learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "e5jGTEiJMT", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Hao Su", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Mengke Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "Undergrad student" }, { "name": "Stone Tao", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Tongzhou Mu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Xiu Yuan", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "Undergrad student" }, { "name": "Yunhao Fang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 68, "n_ref": 131, "n_ref_all": 176, "n_fig": 27, "n_tab": 8, "L_tab": 1000, "n_element_tab": 78, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 1075, "n_element_tab_1": 39, "formula_len_all": 107, "formula_len_all_1": 81, "len_all": 245808, "len_all_1": 67241, "len_abs": 1036, "len_title": 122, "len_sents": 78895, "len_sents_1": 31549, "n_sents": 633, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1094, "title": "Policy Decorator: Model-Agnostic Online Refinement for Large Policy Model", "abs": "Recent advancements in robot learning have used imitation learning with large models and extensive demonstrations to develop effective policies. However, these models are often limited by the quantity quality, and diversity of demonstrations. This paper explores improving offline-trained imitation learning models through online interactions with the environment. We introduce Policy Decorator, which uses a model-agnostic residual policy to refine large imitation learning models during online interactions. By implementing controlled exploration strategies, Policy Decorator enables stable, sample-efficient online learning. Our evaluation spans eight tasks across two benchmarks\u2014ManiSkill and Adroit\u2014and involves two state-of-the-art imitation learning models (Behavior Transformer and Diffusion Policy). The results show Policy Decorator effectively improves the offline-trained policies and preserves the smooth motion of imitation learning models, avoiding the erratic behaviors of pure RL policies. See our [project page](https://sites.google.com/view/policy-decorator/home) for videos.", "keywords": [ "Policy Learning", "Online Improve" ], "rating_list": [ 8, 5, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "e5g53a4A0g", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Chao Pan", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Olgica Milenkovic", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Full Professor" }, { "name": "Peizhi Niu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Siheng Chen", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 43, "n_ref": 136, "n_ref_all": 160, "n_fig": 4, "n_tab": 19, "L_tab": 22427, "n_element_tab": 1594, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2340, "n_element_tab_1": 235, "formula_len_all": 593, "formula_len_all_1": 620, "len_all": 220146, "len_all_1": 75384, "len_abs": 1881, "len_title": 128, "len_sents": 61793, "len_sents_1": 34755, "n_sents": 429, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1890, "title": "Improving Defense Mechanisms for Subgraph-Structure Membership Inference Attacks", "abs": "Graph neural networks (GNNs) are of significant importance in diverse real-world applications since they leverage powerful graph learning techniques to solve problems pertaining to social network mining and medical data analysis. Despite their practical relevance, GNNs remain vulnerable to adversarial attacks such as membership inference attacks (MIAs) which pose privacy risks by revealing whether specific data records were part of the training set of the model. While most existing research has focused on designing defense mechanisms for known node-level MIAs, and in particular, for determining if a certain node was used during training, only limited attention has been paid to subgraph-structure MIA (SMIA) problems. SMIA methods seek to infer whether a set of nodes forms a particular target structure of interest (such as a graph motif, e.g., clique or multi-hop path) in the training graph. The main contributions of our work are three-fold. The first is a novel robust defense mechanism for GNNs against SMIA attacks. It combines an alternating train-test schedule with a flattening strategy to mitigate the attacks. The second contribution is a new end-to-end SMIA attack model that outperforms existing attacks by using multiset functions to generate learnable embeddings for collections of nodes. Extensive simulations reveal that the new attack model outperforms prior state-of-the-art attack models on GNNs by 12.31% across four datasets when no defense mechanism is present. With the new defense mechanism, one can achieve an average decrease of 14.30% in the attack AUROC and an 10.05% improvement in target model utility compared to classical defenses, even when using the improved attack scheme. The third contribution is a study that shows that our defense mechanism extends to node-level MIAs as well, offering similar improvements in attack resistance and utility.", "keywords": [ "Graph neural network", "subgraph-substructure membership inference attack", "2-stage training" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "e5288Iu4Zc", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Kai Zhu", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "Postdoc" }, { "name": "Liming Zhao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Pingyu Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Wei Zhai", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" }, { "name": "Yang Cao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yu Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zheng-Jun Zha", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 71, "n_ref_all": 100, "n_fig": 13, "n_tab": 7, "L_tab": 1579, "n_element_tab": 101, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 1062, "n_element_tab_1": 30, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 106478, "len_all_1": 63633, "len_abs": 1608, "len_title": 99, "len_sents": 36500, "len_sents_1": 32017, "n_sents": 239, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 51, "L_abs": 1613, "title": "Improved Video VAE for Latent Video Diffusion Model", "abs": "Variational Autoencoder (VAE) aims to compress pixel data into low-dimensional latent space, playing an important role in OpenAI's Sora and other latent video diffusion generation models. While most of existing video VAEs inflate a pretrained image VAE into the 3D causal structure for temporal-spatial compression, this paper presents two astonishing findings: (1) The initialization from a well-trained image VAE with the same latent dimensions suppresses the improvement of subsequent temporal compression capabilities. (2) The adoption of causal reasoning leads to unequal information interactions and unbalanced performance between frames. To alleviate these problems, we propose a keyframe-based temporal compression (KTC) architecture and a group causal convolution (GCConv) module to further improve video VAE (IV-VAE). Specifically, the KTC architecture divides the latent space into two branches, in which one half completely inherits the compression prior of keyframes from lower-dimension image VAEs while the other half involves temporal compression into the 3D group causal convolution, reducing temporal-spatial conflicts and accelerating the convergence speed of video VAE. The GCConv in above 3D half uses standard convolution within each frame group to ensure inter-frame equivalence, and employs causal logical padding between groups to maintain flexibility in processing variable frame video. Extensive experiments on five benchmarks demonstrate the SOTA video reconstruction and generation capabilities of the proposed IV-VAE. The source code and weights will be made available to the public.", "keywords": [ "Video VAE", "Variational Autoencoder" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "e4em5klSEw", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Jihai Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Tong Zhu", "gender": "unknown", "institution": "Soochow University, China", "country": "CN", "position": "PhD student" }, { "name": "Xiaoye Qu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yu Cheng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 36, "n_ref": 83, "n_ref_all": 92, "n_fig": 1, "n_tab": 5, "L_tab": 1875, "n_element_tab": 127, "n_fig_1": 0, "n_tab_1": 6, "L_tab_1": 2423, "n_element_tab_1": 137, "formula_len_all": 428, "formula_len_all_1": 428, "len_all": 130601, "len_all_1": 63032, "len_abs": 1769, "len_title": 141, "len_sents": 32645, "len_sents_1": 28493, "n_sents": 222, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1799, "title": "CLIP-MoE: Towards Building Mixture of Experts for CLIP with Diversified Multiplet Upcycling", "abs": "In recent years, Contrastive Language-Image Pre-training (CLIP) has become a cornerstone in multimodal intelligence. \nHowever, recent studies have identified that the information loss in the encoding process of CLIP is substantial. Such deficiency significantly limits the ability of a single CLIP model to handle images rich in visual detail. In this work, we propose a simple yet effective model-agnostic strategy, $\\textbf{Diversified Multiplet Upcycling (DMU)}$ for CLIP. It integrates multiple CLIP models that capture diversified, complementary information into a Mixture of Experts (MoE) architecture. Inspired by the recently proposed Multistage Contrastive Learning (MCL), which constructs multiple CLIP models that share the same structure while capturing different complementary information, Diversified Multiplet Upcycling efficiently fine-tunes a series of CLIP models from a dense pre-trained CLIP checkpoint to capture different feature distributions, sharing parameters except for the Feed-Forward Network (FFN). These models are then transformed into a $\\textbf{CLIP-MoE}$ with a larger model capacity but minimal computational overhead. Extensive experiments demonstrate the significant performance of CLIP-MoE across various zero-shot retrieval, zero-shot image classification tasks, and downstream Multimodal Large Language Model (MLLM) benchmarks by serving as a vision encoder. Furthermore, Diversified Multiplet Upcycling enables the conversion of any dense CLIP model into CLIP-MoEs, which can seamlessly replace CLIP in a plug-and-play manner without requiring further adaptation in downstream frameworks. Through Diversified Multiplet Upcycling, we aim to provide valuable insights for future research on developing more efficient and effective multimodal learning systems.", "keywords": [ "Mixture of Experts", "Contrastive Learning", "Multimodal Learning", "Multimodal Large Language Models" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "e4PL5zssJ9", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Csaba Both", "gender": "unknown", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Jeet Mohapatra", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "MS student" }, { "name": "Nima Dehmamy", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Subhro Das", "gender": "Male", "institution": "MIT-IBM Watson AI Lab, IBM Research", "country": "", "position": "Staff Research Scientist" }, { "name": "Tommi Jaakkola", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 45, "n_formula_1": 25, "n_ref_uni": 22, "n_ref": 29, "n_ref_all": 32, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 2622, "n_element_tab_1": 31, "formula_len_all": 2521, "formula_len_all_1": 1026, "len_all": 115419, "len_all_1": 60994, "len_abs": 1092, "len_title": 121, "len_sents": 41472, "len_sents_1": 27084, "n_sents": 315, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1096, "title": "Symmetry-Driven Discovery of Dynamical Variables in Molecular Simulations", "abs": "Molecular dynamics simulations are crucial for understanding complex biomolecular systems, but they are often hindered by the high dimensionality of the configurational space. This paper introduces two novel approaches for discovering effective degrees of freedom (DoF) in molecular dynamics simulations by leveraging approximate symmetries of the energy landscape. We present a scalable symmetry loss function compatible with existing force-field frameworks and a Hessian-based method efficient for smaller systems. Both approaches enable systematic exploration of conformational space by connecting structural dynamics to energy landscape symmetries. Applied to alanine dipeptide, our methods comprehensively sample the Ramachandran plot, including shallow minima. Simulations initiated from our DoF-sampled points converge to all important conformations, demonstrating the methods\u2019 effectiveness in navigating complex energy landscapes. These approaches offer powerful tools for efficient exploration in molecular simulations, with potential applications in protein folding and drug discovery.", "keywords": [ "AI4Science", "Molecular Dynamics", "Second Order Methods", "Symmetry Discovery" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 2, 1 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "e3odKmatZr", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Brandon Cui", "gender": "Male", "institution": "Stanford University", "country": "", "position": "MS student" }, { "name": "Jonathan Daniel Chang", "gender": "unknown", "institution": "Cornell University", "country": "", "position": "PhD student" }, { "name": "Mansheej Paul", "gender": "Male", "institution": "Databricks, Databricks", "country": "US", "position": "Researcher" }, { "name": "Prithviraj Ammanabrolu", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Zachary Ankner", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 38, "n_ref": 61, "n_ref_all": 107, "n_fig": 18, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 568, "formula_len_all_1": 608, "len_all": 189178, "len_all_1": 61681, "len_abs": 1422, "len_title": 79, "len_sents": 63937, "len_sents_1": 29650, "n_sents": 481, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 31, "L_abs": 1332, "title": "Critique-out-Loud Reward Models", "abs": "Traditionally, reward models used for reinforcement learning from human feedback (RLHF) are trained to directly predict preference scores without leveraging the generation capabilities of the underlying large language model (LLM). This limits the capabilities of reward models as they must reason implicitly about the quality of a response, i.e., preference modeling must be performed in a single forward pass through the model. To enable reward models to reason explicitly about the quality of a response, we introduce Critique-out-Loud (CLoud) reward models. CLoud reward models operate by first generating a natural language critique of the assistant's response that is then used to predict a scalar reward for the quality of the response. We demonstrate the success of CLoud reward models for both Llama-3-8B and 70B base models: compared to classic reward models CLoud reward models improve pairwise preference classification accuracy on RewardBench by 4.65 and 5.84 percentage points for the 8B and 70B base models respectively. Furthermore, CLoud reward models lead to a Pareto improvement for win rate on ArenaHard when used as the scoring model for Best-of-N. Finally, we explore how to exploit the dynamic inference compute capabilities of CLoud reward models by performing self-consistency decoding for reward prediction.", "keywords": [ "Reward Model", "Chain of Thought", "Alignment", "Preference Modeling", "Large Language Model", "LLM", "RLHF" ], "rating_list": [ 8, 5, 5, 3 ], "soundness_list": [ 4, 2, 2, 2 ], "presentation_list": [ 4, 2, 3, 4 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "e32cI4r8Eo", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Cong Wang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Fei Shen", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Researcher" }, { "name": "Jun Zhang", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Kuan Tian", "gender": "unknown", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Qing Gu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Yonghang Guan", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Zhiwei Jiang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 28, "n_formula_1": 18, "n_ref_uni": 56, "n_ref": 94, "n_ref_all": 135, "n_fig": 9, "n_tab": 13, "L_tab": 4265, "n_element_tab": 552, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1518, "n_element_tab_1": 73, "formula_len_all": 1274, "formula_len_all_1": 802, "len_all": 192522, "len_all_1": 62936, "len_abs": 1657, "len_title": 108, "len_sents": 47446, "len_sents_1": 26315, "n_sents": 415, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1664, "title": "Ensembling Diffusion Models via Adaptive Feature Aggregation", "abs": "The success of the text-guided diffusion model has inspired the development and release of numerous powerful diffusion models within the open-source community. These models are typically fine-tuned on various expert datasets, showcasing diverse denoising capabilities. Leveraging multiple high-quality models to produce stronger generation ability is valuable, but has not been extensively studied. Existing methods primarily adopt parameter merging strategies to produce a new static model. However, they overlook the fact that the divergent denoising capabilities of the models may dynamically change across different states, such as when experiencing different prompts, initial noises, denoising steps, and spatial locations. In this paper, we propose a novel ensembling method, Adaptive Feature Aggregation (AFA), which dynamically adjusts the contributions of multiple models at the feature level according to various states (i.e., prompts, initial noises, denoising steps, and spatial locations), thereby keeping the advantages of multiple diffusion models, while suppressing their disadvantages. Specifically, we design a lightweight Spatial-Aware Block-Wise (SABW) feature aggregator that adaptive aggregates the block-wise intermediate features from multiple U-Net denoisers into a unified one. The core idea lies in dynamically producing an individual attention map for each model's features by comprehensively considering various states. It is worth noting that only SABW is trainable with about 50 million parameters, while other models are frozen. Both the quantitative and qualitative experiments demonstrate the effectiveness of our proposed method.", "keywords": [ "Image Generation", "Diffusion Models", "Model Ensembling" ], "rating_list": [ 6, 8, 5, 6, 6 ], "soundness_list": [ 3, 4, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "e2p1BWR3vq", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Alexander H. Liu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Bryan Catanzaro", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Vice President" }, { "name": "Joao Felipe Santos", "gender": "Non-Binary", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Mikyas T. Desta", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Rafael Valle", "gender": "Not Specified", "institution": "NVIDIA", "country": "", "position": "Senior Research Scientist" }, { "name": "Sang-gil Lee", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Sungwon Kim", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Sudheer Kumar Kovela", "gender": "Male", "institution": "NVIDIA", "country": "IN", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 23, "n_ref": 50, "n_ref_all": 67, "n_fig": 4, "n_tab": 9, "L_tab": 1725, "n_element_tab": 247, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 569, "n_element_tab_1": 70, "formula_len_all": 693, "formula_len_all_1": 514, "len_all": 136309, "len_all_1": 70424, "len_abs": 1669, "len_title": 127, "len_sents": 46548, "len_sents_1": 34031, "n_sents": 297, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1682, "title": "A$^2$-Flow: Alignment-Aware Pre-training for Speech Synthesis with Flow Matching", "abs": "Recent advances in speech synthesis have enabled highly natural and speaker-adaptive speech generation by leveraging large-scale transcribed datasets. However, requiring tens of thousands of hours of annotated speech is impractical in low-resource settings. Existing pre-trained speech models often utilize masked speech inpainting for pre-training and show strong performance on various speech generation tasks using limited task-specific data. Nonetheless, these models still require external alignment mechanisms or extensive additional training to learn alignment for alignment-aware tasks, such as text-to-speech (TTS). In this paper, we propose A$^2$-Flow, an alignment-aware pre-training method for flow matching models in speech synthesis. A$^2$-Flow integrates alignment learning directly into the pre-training process using discrete speech units, enabling the model to efficiently adapt to alignment-aware tasks without the need for separate alignment mechanisms. By embedding alignment learning into pre-training, A$^2$-Flow facilitates alignment-free voice conversion (VC) and allows for faster convergence during TTS fine-tuning, even with limited transcribed data, making it highly suitable for low-resource scenarios. Experimental results show that A$^2$-Flow superior zero-shot VC performance compared to existing models and matches state-of-the-art TTS performance using only a small amount of transcribed data. Moreover, we demonstrate that A$^2$-Flow can be more efficiently applied to alignment-aware speech synthesis tasks than existing pre-training methods, providing a practical and scalable solution for high-quality speech synthesis across diverse settings.", "keywords": [ "Generative Pre-training; Flow Matching; Alignment Learning; Discrete Speech Units" ], "rating_list": [ 3, 8, 5, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "e2ONKX6qzJ", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Otmar Hilliges", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "Full Professor" }, { "name": "Romann M. Weber", "gender": "Male", "institution": "Disney Research, Disney", "country": "CH", "position": "Research Scientist" }, { "name": "Seyedmorteza Sadat", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 6, "n_ref_uni": 23, "n_ref": 94, "n_ref_all": 132, "n_fig": 16, "n_tab": 10, "L_tab": 2045, "n_element_tab": 679, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 469, "n_element_tab_1": 36, "formula_len_all": 1761, "formula_len_all_1": 344, "len_all": 158965, "len_all_1": 49423, "len_abs": 1601, "len_title": 84, "len_sents": 37881, "len_sents_1": 22411, "n_sents": 300, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1544, "title": "Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models", "abs": "Classifier-free guidance (CFG) is crucial for improving both generation quality and alignment between the input condition and final output in diffusion models. While a high guidance scale is generally required to enhance these aspects, it also causes oversaturation and unrealistic artifacts. In this paper, we revisit the CFG update rule and introduce modifications to address this issue. We first decompose the update term in CFG into parallel and orthogonal components with respect to the conditional model prediction and observe that the parallel component primarily causes oversaturation, while the orthogonal component enhances image quality. Accordingly, we propose down-weighting the parallel component to achieve high-quality generations without oversaturation. Additionally, we draw a connection between CFG and gradient ascent and introduce a new rescaling and momentum method for the CFG update rule based on this insight. Our approach, termed adaptive projected guidance (APG), retains the quality-boosting advantages of CFG while enabling the use of higher guidance scales without oversaturation. APG is easy to implement and introduces practically no additional computational overhead to the sampling process. Through extensive experiments, we demonstrate that APG is compatible with various conditional diffusion models and samplers, leading to improved FID, recall, and saturation scores while maintaining precision comparable to CFG, making our method a superior plug-and-play alternative to standard classifier-free guidance.", "keywords": [ "classifier-free guidance", "diffusion models" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 4, 4, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "e2NRNQ0sZe", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Haifeng Zhang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Associate Professor" }, { "name": "Haitham Bou Ammar", "gender": "Male", "institution": "Huawei R&D UK", "country": "GB", "position": "Principal Researcher" }, { "name": "Jun Wang", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Professor" }, { "name": "Mengyue Yang", "gender": "Female", "institution": "University of Bristol", "country": "GB", "position": "Assistant Professor" }, { "name": "Xidong Feng", "gender": "unknown", "institution": "Google DeepMind", "country": "GB", "position": "Research Scientist" }, { "name": "Xue Yan", "gender": "Female", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yan Song", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" } ], "n_formula": 41, "n_formula_1": 16, "n_ref_uni": 58, "n_ref": 91, "n_ref_all": 115, "n_fig": 10, "n_tab": 4, "L_tab": 1693, "n_element_tab": 167, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1530, "n_element_tab_1": 170, "formula_len_all": 3502, "formula_len_all_1": 1115, "len_all": 182882, "len_all_1": 65178, "len_abs": 3703, "len_title": 113, "len_sents": 48758, "len_sents_1": 28335, "n_sents": 380, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1234, "title": "Efficient Reinforcement Learning with Large Language Model Priors", "abs": "In sequential decision-making (SDM) tasks, methods like reinforcement learning (RL) and heuristic search have made notable advances in specific cases. However, they often require extensive exploration and face challenges in generalizing across diverse environments due to their limited grasp of the underlying decision dynamics. In contrast, large language models (LLMs) have recently emerged as powerful general-purpose tools, due to their capacity to maintain vast amounts of domain-specific knowledge. To harness this rich prior knowledge for efficiently solving complex SDM tasks, we propose treating LLMs as prior action distributions and integrating them into RL frameworks through Bayesian inference methods, making use of variational inference and direct posterior sampling. The proposed approaches facilitate the seamless incorporation of fixed LLM priors into both policy-based and value-based RL frameworks. Our experiments show that incorporating LLM-based action priors significantly reduces exploration and optimization complexity, substantially improving sample efficiency compared to traditional RL techniques, e.g., using LLM priors decreases the number of required samples by over 90\\% in offline learning scenarios.", "keywords": [ "Reinforcement Learning; Probabilistic Inference; Language Prior;" ], "rating_list": [ 6, 6, 8, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "e2F0mJJeN0", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Anish Acharya", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Inderjit S. Dhillon", "gender": "Male", "institution": "Google", "country": "US", "position": "Distinguished Scientist" }, { "name": "sujay sanghavi", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 14, "n_ref_uni": 66, "n_ref": 106, "n_ref_all": 127, "n_fig": 5, "n_tab": 5, "L_tab": 3001, "n_element_tab": 296, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 855, "n_element_tab_1": 49, "formula_len_all": 1810, "formula_len_all_1": 859, "len_all": 150857, "len_all_1": 58295, "len_abs": 1098, "len_title": 102, "len_sents": 36000, "len_sents_1": 26917, "n_sents": 273, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1138, "title": "Geometric Median (GM) Matching for Robust Data Pruning", "abs": "Data pruning, the combinatorial task of selecting a small and informative subset from a large dataset, is crucial for mitigating the enormous computational costs associated with training data-hungry modern deep learning models at scale. Since large-scale data collections are invariably noisy, developing data pruning strategies that remain robust even in the presence of corruption is critical in practice. \nIn response, we propose $\\gmm$ -- a herding~\\citep{welling2009herding} style greedy algorithm -- that {\\em yields a $k$-subset such that the mean of the subset approximates the geometric median of the (potentially) noisy dataset}. Theoretically, we show that $\\gm$ Matching enjoys an improved $\\gO(1/k)$ scaling over $\\gO(1/\\sqrt{k})$ scaling of uniform sampling; while achieving the optimal breakdown point of 1/2 even under arbitrary corruption. Extensive experiments across popular deep learning benchmarks indicate that $\\gm$ Matching consistently outperforms prior state-of-the-art; the gains become more profound at high rates of corruption and aggressive pruning rates; making it a strong baseline for robust data pruning.", "keywords": [ "data pruning", "robust", "data selection" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 1, 3, 3 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "e1wDDFmlVu", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Dianqi Li", "gender": "Male", "institution": "Citadel Securities", "country": "US", "position": "Quantitative Researcher" }, { "name": "Ming Jin", "gender": "Male", "institution": "Griffith University", "country": "AU", "position": "Assistant Professor" }, { "name": "Shiyu Wang", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Staff Researcher" }, { "name": "Xiaoming Shi", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Yuqi Nie", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Zhou Ye", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Qingsong Wen", "gender": "Male", "institution": "Squirrel Ai Learning", "country": "US", "position": "Principal Researcher" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 60, "n_ref": 94, "n_ref_all": 117, "n_fig": 4, "n_tab": 12, "L_tab": 5349, "n_element_tab": 548, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 6208, "n_element_tab_1": 280, "formula_len_all": 752, "formula_len_all_1": 575, "len_all": 215383, "len_all_1": 68896, "len_abs": 1812, "len_title": 126, "len_sents": 51156, "len_sents_1": 28775, "n_sents": 358, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 77, "L_abs": 1821, "title": "Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts", "abs": "Deep learning for time series forecasting has seen significant advancements over the past decades. However, despite the success of large-scale pre-training in language and vision domains, pre-trained time series models remain limited in scale and operate at a high cost, hindering the development of larger capable forecasting models in real-world applications. In response, we introduce Time-MoE, a scalable and unified architecture designed to pre-train larger, more capable forecasting foundation models while reducing inference costs. By leveraging a sparse mixture-of-experts (MoE) design, Time-MoE enhances computational efficiency by activating only a subset of networks for each prediction, reducing computational load while maintaining high model capacity. This allows Time-MoE to scale effectively without a corresponding increase in inference costs. Time-MoE comprises a family of decoder-only transformer models that operate in an auto-regressive manner and support flexible forecasting horizons with varying input context lengths. We pre-trained these models on our newly introduced large-scale data Time-300B, which spans over 9 domains and encompassing over 300 billion time points. For the first time, we scaled a time series foundation model up to 2.4 billion parameters, achieving significantly improved forecasting precision. Our results validate the applicability of scaling laws for training tokens and model size in the context of time series forecasting. Compared to dense models with the same number of activated parameters or equivalent computation budgets, our models consistently outperform them by large margin. These advancements position Time-MoE as a state-of-the-art solution for tackling real-world time series forecasting challenges with superior capability, efficiency, and flexibility.", "keywords": [ "time series", "foundation model", "forecasting" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "e1Z4NCQ146", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 7, "author_info_list": [ { "name": "Hanmei Yang", "gender": "Female", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" }, { "name": "Hui Guan", "gender": "Female", "institution": "University of Massachusetts, Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Jin Zhou", "gender": "unknown", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" }, { "name": "Ramine Roane", "gender": "Male", "institution": "Advanced Micro Devices", "country": "", "position": "VP" }, { "name": "Tongping Liu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Xiaoqun Wang", "gender": "Male", "institution": "Advanced Micro Devices", "country": "", "position": "Researcher" }, { "name": "Yao Fu", "gender": "Male", "institution": "Advanced Micro Devices", "country": "US", "position": "Principal Researcher" } ], "n_formula": 10, "n_formula_1": 2, "n_ref_uni": 31, "n_ref": 49, "n_ref_all": 70, "n_fig": 5, "n_tab": 8, "L_tab": 2659, "n_element_tab": 221, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 672, "n_element_tab_1": 47, "formula_len_all": 1111, "formula_len_all_1": 118, "len_all": 176012, "len_all_1": 60628, "len_abs": 1071, "len_title": 113, "len_sents": 56205, "len_sents_1": 31256, "n_sents": 378, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1090, "title": "ProTrain: Efficient LLM Training via Automatic Memory Management", "abs": "Training billion-scale large language models (LLMs) with just a few consumer-grade graphics cards is key to democratizing LLM access. However, existing frameworks often depend on manual tuning of memory management settings, leading to inefficient hardware utilization and suboptimal performance. This paper introduces ProTrain, a novel training system that automatically tailors memory management policies to the model architecture and underlying hardware resources, eliminating the need for manual intervention. ProTrain features (1) automated memory management that abstracts complex memory management strategies into a few tunable configuration parameters and searches for optimal parameter settings using cost models and (2) a runtime profiler that provides precise estimates of latency, memory usage, and I/O bandwidth to build high-fidelity cost models. \nProTrain does not change the training algorithm and thus does not compromise accuracy. Experiments show that ProTrain improves training throughput by 1.43$\\times$ to 2.71$\\times$ compared to the state-of-the-art training systems.", "keywords": [ "ML System", "Memory Optimization", "Data Parallelism", "ZeRO", "Gradient Checkpointing", "Tensor Offloading" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 2, 3, 3, 5 ] }, { "paper_id": "e1ETy9XW0T", "primary_area": "learning on time series and dynamical systems", "n_author": 7, "author_info_list": [ { "name": "Annie Marsden", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Daniel Suo", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Elad Hazan", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Full Professor" }, { "name": "Evan Dogariu", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Undergrad student" }, { "name": "Naman Agarwal", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Peter L. Bartlett", "gender": "Male", "institution": "University of California - Berkeley", "country": "", "position": "Full Professor" }, { "name": "Xinyi Chen", "gender": "Female", "institution": "Princeton University", "country": "", "position": "PhD student" } ], "n_formula": 34, "n_formula_1": 33, "n_ref_uni": 0, "n_ref": 33, "n_ref_all": 47, "n_fig": 5, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 294, "n_element_tab_1": 9, "formula_len_all": 2425, "formula_len_all_1": 2347, "len_all": 78296, "len_all_1": 72184, "len_abs": 554, "len_title": 124, "len_sents": 35462, "len_sents_1": 31698, "n_sents": 229, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 558, "title": "Provable Length Generalization in Sequence Prediction via Spectral Filtering", "abs": "We consider the problem of length generalization in sequence prediction. We define a new metric of performance in this setting -- the Asymmetric-Regret -- which measures regret against a benchmark predictor with longer context length than available to the learner. We continue by studying this concept from the lens of the spectral filtering algorithm. We present a gradient-based learning algorithm that provably achieves length generalization for linear dynamical systems. We conclude with proof-of-concept experiments which are consistent with our theory.", "keywords": [ "online learning", "state space models", "linear dynamical systems", "sequence prediction", "LLMs" ], "rating_list": [ 8, 5, 3, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "e1DkCLjdhS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Ang Li", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Associate Professor" }, { "name": "Chuan Liu", "gender": "Female", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Chunshu Wu", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Postdoc" }, { "name": "Ruibing Song", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Tong Geng", "gender": "Male", "institution": "University of Rochester", "country": "", "position": "Assistant Professor" }, { "name": "Yingnian Wu", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 25, "n_ref": 67, "n_ref_all": 83, "n_fig": 6, "n_tab": 4, "L_tab": 1243, "n_element_tab": 176, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 985, "n_element_tab_1": 158, "formula_len_all": 602, "formula_len_all_1": 602, "len_all": 107938, "len_all_1": 68682, "len_abs": 1477, "len_title": 126, "len_sents": 37207, "len_sents_1": 32739, "n_sents": 235, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1485, "title": "A Brain-Inspired Machine Learning Paradigm for Nature-Powered Equation Solving", "abs": "Solving equations is fundamental to human understanding of the world. While modern machine learning methods are powerful equation solvers, their escalating complexity and extreme operational costs hinder sustainable development. In contrast, nature effortlessly solves complex equations through dynamical systems that instinctively evolve to low-energy states without explicit instructions. However, existing attempts to leverage dynamical systems are limited by low expressivity and a lack of training support. To this end, we propose DS-Solver, a nature-powered AI paradigm employing an expressive, self-trainable dynamical system capable of accurately solving a wide spectrum of equations with extraordinary efficiency. (1) We enhance system expressivity by enriching node dynamics with coupled real-valued and polarized shadow nodes, capturing complex interactions inherent in the real world. (2) We propose an on-device learning method that leverages intrinsic electrical signals as loss, enabling the dynamical system to instantly train itself at negligible cost. Experimental results across key equations from diverse domains demonstrate that DS-Solver achieves 42\\% higher accuracy than current SOTA -- while offering orders-of-magnitude improvements in speed and energy efficiency over traditional neural network solutions on GPUs for both inference and training, showcasing its broader impact in overcoming persistent computational bottlenecks across various critical fields.", "keywords": [ "Nature-Powered Computing", "Dynamical Systems" ], "rating_list": [ 1, 3, 6, 6 ], "soundness_list": [ 1, 1, 3, 3 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 1, 1, 3, 2 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "e0rQRMUhs7", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Chun-Mei Feng", "gender": "Female", "institution": "IHPC", "country": "SG", "position": "Researcher" }, { "name": "Lei Zhu", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou) & HKUST", "country": "GB", "position": "Assistant Professor" }, { "name": "Rick Siow Mong Goh", "gender": "unknown", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "SG", "position": "Director" }, { "name": "Wangmeng Zuo", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yong Liu", "gender": "Male", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "SG", "position": "Senior Scientist, Adjunct Assistant Professor" }, { "name": "Yunlu Yan", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 22, "n_ref_uni": 65, "n_ref": 128, "n_ref_all": 165, "n_fig": 6, "n_tab": 11, "L_tab": 3495, "n_element_tab": 393, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 2259, "n_element_tab_1": 207, "formula_len_all": 1527, "formula_len_all_1": 1554, "len_all": 211910, "len_all_1": 62200, "len_abs": 1449, "len_title": 109, "len_sents": 51231, "len_sents_1": 25163, "n_sents": 400, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1395, "title": "Federated Residual Low-Rank Adaption of Large Language Models", "abs": "Low-Rank Adaptation (LoRA) presents an effective solution for federated fine-tuning of Large Language Models (LLMs), as it substantially reduces communication overhead. However, a straightforward combination of FedAvg and LoRA results in suboptimal performance, especially under data heterogeneity. We noted this stems from both intrinsic (i.e., constrained parameter space) and extrinsic (i.e., client drift) limitations, which hinder it effectively learn global knowledge. In this work, we proposed a novel Federated Residual Low-Rank Adaption method, namely FRLoRA, to tackle above two limitations. It directly sums the weight of the global model parameters with a residual low-rank matrix product (\\ie, weight change) during the global update step, and synchronizes this update for all local models. By this, FRLoRA performs global updates in a higher-rank parameter space, enabling a better representation of complex knowledge structure. Furthermore, FRLoRA reinitializes the local low-rank matrices with the principal singular values and vectors of the pre-trained weights in each round, to calibrate their inconsistent convergence, thereby mitigating client drift. Our extensive experiments demonstrate that FRLoRA consistently outperforms various state-of-the-art FL methods across nine different benchmarks in natural language understanding and generation under different FL scenarios.", "keywords": [ "Large Languagel model", "Federated Learning", "Parameter-Efficient Fine-Tuning" ], "rating_list": [ 8, 8, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 4, 5, 2, 4, 3 ] }, { "paper_id": "e0bTcdF29g", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Arijit Das", "gender": "Male", "institution": "ERGO AG", "country": "DE", "position": "Principal Researcher" } ], "n_formula": 15, "n_formula_1": 16, "n_ref_uni": 39, "n_ref": 62, "n_ref_all": 70, "n_fig": 1, "n_tab": 3, "L_tab": 453, "n_element_tab": 26, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1186, "n_element_tab_1": 103, "formula_len_all": 668, "formula_len_all_1": 598, "len_all": 113593, "len_all_1": 64745, "len_abs": 1587, "len_title": 133, "len_sents": 34207, "len_sents_1": 31083, "n_sents": 219, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1628, "title": "Natural GaLore: Accelerating GaLore for memory-efficient LLM Training and Fine-tuning", "abs": "Training LLMs presents significant memory challenges due to growing size of data, weights, and optimizer states. Techniques such as data and model parallelism, gradient checkpointing, and offloading strategies address this issue but are often infeasible due to hardware constraints. To mitigate memory usage, alternative methods like Parameter-Efficient-Fine-Tuning (PEFT) and GaLore approximate weights or optimizer states. PEFT methods, such as LoRA, have gained popularity for fine-tuning LLMs, though they require a full-rank warm start. In contrast, GaLore allows full-parameter learning while being more memory-efficient. This work introduces \\textit{Natural GaLore}, a simple drop in replacement for AdamW, which efficiently applies the inverse Empirical Fisher Information Matrix to low-rank gradients using Woodbury's Identity. We demonstrate that incorporating second-order information speeds up optimization significantly, especially when the iteration budget is limited. Empirical pretraining on 60M, 130M, 350M, and 1.1B parameter Llama models on C4 data demonstrate significantly lower perplexity over GaLore without additional memory overhead. By fine-tuning RoBERTa on the GLUE benchmark using \\textit{Natural GaLore}, we demonstrate significant reduction in gap 86.05\\% vs 86.28\\% for full-finetuning. Furthermore, fine-tuning the TinyLlama 1.1B model for function calling using the TinyAgent framework shows that \\textit{Natural GaLore} achieving 83.09\\% accuracy on the TinyAgent dataset, significantly outperforms 16-bit LoRA at 80.06\\% and even surpasses GPT4-Turbo by 4\\%, all while using 30\\% less memory.", "keywords": [ "large language models", "LLM pre-training", "memory-efficient training", "optimization" ], "rating_list": [ 3, 6, 5, 3, 5 ], "soundness_list": [ 2, 4, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 2 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4, 2 ] }, { "paper_id": "e0X9l4kecx", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Alp Yurtsever", "gender": "Male", "institution": "Ume\u00e5 University", "country": "SE", "position": "Assistant Professor" }, { "name": "Karthik Prakhya", "gender": "Male", "institution": "Umea University", "country": "", "position": "PhD student" }, { "name": "Tolga Birdal", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 39, "n_formula_1": 21, "n_ref_uni": 48, "n_ref": 83, "n_ref_all": 107, "n_fig": 2, "n_tab": 5, "L_tab": 1696, "n_element_tab": 57, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1405, "n_element_tab_1": 60, "formula_len_all": 3500, "formula_len_all_1": 1845, "len_all": 185349, "len_all_1": 67136, "len_abs": 936, "len_title": 111, "len_sents": 53649, "len_sents_1": 29887, "n_sents": 445, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 940, "title": "Convex Formulations for Training Two-Layer ReLU Neural Networks", "abs": "Solving non-convex, NP-hard optimization problems is crucial for training machine learning models, including neural networks. However, non-convexity often leads to black-box machine learning models with unclear inner workings. While convex formulations have been used for verifying neural network robustness, their application to training neural networks remains less explored. In response to this challenge, we reformulate the problem of training infinite-width two-layer ReLU networks as a convex completely positive program in a finite-dimensional (lifted) space. Despite the convexity, solving this problem remains NP-hard due to the complete positivity constraint. To overcome this challenge, we introduce a semidefinite relaxation that can be solved in polynomial time. We then experimentally evaluate the tightness of this relaxation, demonstrating its competitive performance in test accuracy across a range of classification tasks.", "keywords": [ "copositive programming", "semidefinite programming", "neural networks" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "dyzdDSzoKi", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Ke Sun", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Ke Yan", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Shouhong Ding", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "researcher" }, { "name": "Xianming Lin", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaoshuai Sun", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Yunpeng Luo", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Zhongxi Chen", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Ziyin Zhou", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 45, "n_ref": 132, "n_ref_all": 149, "n_fig": 8, "n_tab": 7, "L_tab": 2772, "n_element_tab": 458, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 4541, "n_element_tab_1": 146, "formula_len_all": 609, "formula_len_all_1": 467, "len_all": 161671, "len_all_1": 61627, "len_abs": 1639, "len_title": 158, "len_sents": 38611, "len_sents_1": 24849, "n_sents": 278, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 110, "L_abs": 1647, "title": "Exploring the Collaborative Advantage of Low-level Information on Generalizable AI-generateted Image Detection", "abs": "This paper investigates the generalization issue in AI-Generated image detection, aiming to generalize from training on one AI-Generated image dataset to detecting unseen AI-Generated images. Many methods consider extracting low-level information from RGB images to aid the generalization of AI-Generated image detection. However, these methods often consider a single type of low-level information and this may lead to suboptimal generalization. In our analysis, different low-level information often exhibit generalization capabilities for different forgery types. Additionally, simple fusion strategies are insufficient to leverage the detection advantages of each low-level and high-level information for various forgery types. Therefore, we propose the Adaptive Low-level Experts Injection (ALEI) framework.\nOur approach introduces Lora Experts to enable the transformer-based backbone to learn knowledge from different low-level information. We incorporate a Cross-Low-level Attention layer to fuse these features at intermediate layers. To prevent the backbone from losing modeling capabilities for different low-level features, we develop a Low-level Information Adapter that interacts with the features extracted by the backbone. Finally, we propose Dynamic Feature Selection to maximize the generalization detection capability by dynamically selecting the most suitable features for detecting the current image.\nExtensive experiments demonstrate that our method, finetuned on only four categories of ProGAN data, performs excellently and achieves state-of-the-art results on multiple datasets containing unseen GAN and Diffusion methods.", "keywords": [ "AI-generateted Image Detection", "Low-level Information" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "dyYc8GFdD5", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Chaoyue Song", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Fan Yang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Guan Luo", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jianfeng Zhang", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Xiu Li", "gender": "Male", "institution": "Bytedance", "country": "SG", "position": "Researcher" }, { "name": "Zhongang Cai", "gender": "Male", "institution": "SenseTime International Pte Ltd", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Zhongcong Xu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Guosheng Lin", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Jiashi Feng", "gender": "Male", "institution": "ByteDance", "country": "SG", "position": "Research Lead" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 15, "n_ref": 30, "n_ref_all": 39, "n_fig": 5, "n_tab": 1, "L_tab": 731, "n_element_tab": 65, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 663, "n_element_tab_1": 66, "formula_len_all": 106, "formula_len_all_1": 106, "len_all": 98844, "len_all_1": 46458, "len_abs": 1682, "len_title": 143, "len_sents": 27543, "len_sents_1": 24185, "n_sents": 162, "n_sents_1": 132, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1689, "title": "U3D: Unlocking the Video Prior for High Fidelity Sparse Novel View Synthesis and 3D Generation", "abs": "Trained on massive datasets, video diffusion models have shown strong generative priors for novel view synthesis tasks. Existing methods finetune these models to synthesize 360-degree orbit videos from input images. While these methods demonstrate the pretrained models' generalization ability, they are limited by the assumption of temporal attention and struggle to generate highly consistent results. Additionally, generating novel views as a sequence of twenty or more frames incurs high computational costs compared to sparse view synthesis methods. Sparse novel view synthesis methods finetuned from traditional 2D diffusion models, on the other hand, can generate highly consistent images from arbitrary camera positions but suffer from poor generalization, leading to unsatisfactory results on out-of-domain inputs. In this paper, we explore leveraging video diffusion models' rich generative priors to enhance sparse novel view generation models. Specifically, we investigate the generation process of video diffusion models and unearth key observations to extract geometrical priors from them. Based on this, we propose a novel framework, U3D, for sparse novel view synthesis. U3D includes a geometrical reference network to integrate these priors into the sparse novel view synthesis network and a temporal enhanced sparse view generation network to preserve pretrained temporal knowledge. By leveraging the significant generative priors from video diffusion models, our framework can synthesize highly consistent sparse novel views with strong generalization ability, which can be reconstructed into high-quality 3D assets using feed-forward sparse view reconstruction methods.", "keywords": [ "Image to 3D; Mutli-view Diffusion; Novel View Synthesis" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "dxoryzjsCW", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Long Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Yanghao Wang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 17, "n_ref_uni": 41, "n_ref": 86, "n_ref_all": 119, "n_fig": 11, "n_tab": 8, "L_tab": 1517, "n_element_tab": 153, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1756, "n_element_tab_1": 65, "formula_len_all": 1696, "formula_len_all_1": 915, "len_all": 159046, "len_all_1": 62713, "len_abs": 381, "len_title": 127, "len_sents": 46062, "len_sents_1": 25455, "n_sents": 516, "n_sents_1": 244, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1284, "title": "Improving Diffusion-based Data Augmentation with Inversion Circle Interpolation", "abs": "Data Augmentation (DA), i.e., synthesizing faithful and diverse samples to expand the original training set, is a prevalent and effective strategy to improve various visual recognition tasks. With the powerful image generation ability, diffusion-based DA has shown strong performance gains on different benchmarks. In this paper, we analyze today\u2019s diffusion-based DA methods, and argue that they can- not take account of both faithfulness and diversity, which are two critical keys for generating high-quality samples and boosting final classification performance. To this end, we propose a novel Diffusion-based Inversion Interpolation DA method: Diff-II. Specifically, Diff-II consists of three main steps: 1) Category concepts learning: Learning concept embeddings for each category. 2) Inversion interpolation: Calculating the inversion for each image, and conducting random circle interpolation for two randomly sampled inversions from the same category. 3) Two-stage denoising: Using different prompts to generate synthesized images in a coarse-to-fine manner. Extensive experiments on multiple image classification tasks (e.g., few-shot, long-tailed, and out-of-distribution classification) have demonstrated its effectiveness over state-of-the-art diffusion-based DA methods.", "keywords": [ "Data Augmentation", "Diffusion", "Inversion Circle Interpolation" ], "rating_list": [ 6, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "dxMffCAd4w", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Guiling Wang", "gender": "unknown", "institution": "New Jersey Institute of Technology", "country": "", "position": "Distinguished Professor " }, { "name": "Jianyi Yang", "gender": "unknown", "institution": "International Technological University", "country": "US", "position": "Independent Researcher" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 25, "n_ref": 36, "n_ref_all": 55, "n_fig": 5, "n_tab": 7, "L_tab": 1404, "n_element_tab": 72, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1348, "n_element_tab_1": 65, "formula_len_all": 489, "formula_len_all_1": 501, "len_all": 106869, "len_all_1": 58157, "len_abs": 980, "len_title": 108, "len_sents": 36666, "len_sents_1": 25255, "n_sents": 282, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 985, "title": "CLF: Curve Line Fitting Neural Network Based On Bezier Curve", "abs": "The Multilayer Perceptron (MLP) serves as a fundamental architecture in deep learning, leveraging the universal function approximation theorem through linear regression combined with activation functions. Despite its widespread use, the inclusion of activation functions contributes to the inherent nature of MLPs as ``black boxes,\" limiting their interpretability. In this paper, we propose a novel Curve Line Fitting (CLF) network, which introduces Bezier curve fitting to directly address nonlinear distributions. By replacing traditional linear regression with Bezier curve regression, the CLF network offers a more efficient means of fitting target distributions. Additionally, the removal of activation functions makes the CLF model fully interpretable, enabling clear insights into the relationships between input dimensions and target distributions, as well as the interdependencies across different dimensions. (Sample code for the CLF model will be made available on GitHub.)", "keywords": [ "CLF", "MLP", "Interpretable Neural Networks" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 3, 1, 2, 2 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "dwQIVcW1du", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Chengcheng Wan", "gender": "unknown", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Songsong Wang", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Intern" }, { "name": "Xiaodong Gu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuling Shi", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 49, "n_ref": 99, "n_ref_all": 120, "n_fig": 10, "n_tab": 6, "L_tab": 5038, "n_element_tab": 590, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3348, "n_element_tab_1": 635, "formula_len_all": 383, "formula_len_all_1": 247, "len_all": 162904, "len_all_1": 68632, "len_abs": 1407, "len_title": 142, "len_sents": 35566, "len_sents_1": 27968, "n_sents": 248, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1411, "title": "From Code to Correctness: Closing the Last Mile of Code Generation with Hierarchical Debugging", "abs": "While large language models have made significant strides in code generation, the pass rate of the generated code is bottlenecked on subtle errors, often requiring human intervention to pass tests, especially for complex problems. Existing LLM-based debugging systems treat generated programs as monolithic units, failing to address bugs at multiple levels of granularity, from low-level syntax errors to high-level algorithmic flaws. In this paper, we introduce Multi-Granularity Debugger (MGDebugger), a hierarchical code debugger by isolating, identifying, and resolving bugs at various levels of granularity. MGDebugger decomposes problematic code into a hierarchical tree structure of subfunctions, with each level representing a particular granularity of error. During debugging, it analyzes each subfunction and iteratively resolves bugs in a bottom-up manner. To effectively test each subfunction, we propose an LLM-simulated Python executor, which traces code execution and tracks important variable states to pinpoint errors accurately. Extensive experiments demonstrate that MGDebugger outperforms existing debugging systems, achieving an 18.9\\% improvement in accuracy over seed generations in HumanEval and a 97.6\\% repair success rate in HumanEvalFix. Furthermore, MGDebugger effectively fixes bugs across different categories and difficulty levels, demonstrating its robustness and effectiveness.", "keywords": [ "program synthesis", "code generation", "large language models", "machine learning for code", "self-debugging" ], "rating_list": [ 6, 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 5, 4 ] }, { "paper_id": "dw9VUsSHGB", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Dong Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Distinguished Scientist" }, { "name": "Hongming Zhang", "gender": "Male", "institution": "Tencent AI Lab Seattle", "country": "US", "position": "Researcher" }, { "name": "Jiawei Han", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign (UIUC)", "country": "US", "position": "Full Professor" }, { "name": "Kaixin Ma", "gender": "unknown", "institution": "Tencent AI Lab", "country": "US", "position": "Researcher" }, { "name": "Mengzhao Jia", "gender": "Female", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Siru Ouyang", "gender": "Female", "institution": "University of Illinois Urbana-Champaign Champaign", "country": "US", "position": "PhD student" }, { "name": "Wenhao Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Researcher" }, { "name": "Zhihan Zhang", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Zilin Xiao", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 32, "n_ref": 73, "n_ref_all": 116, "n_fig": 17, "n_tab": 7, "L_tab": 1974, "n_element_tab": 232, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 846, "n_element_tab_1": 97, "formula_len_all": 81, "formula_len_all_1": 75, "len_all": 197428, "len_all_1": 76567, "len_abs": 1207, "len_title": 126, "len_sents": 56001, "len_sents_1": 37002, "n_sents": 446, "n_sents_1": 272, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1212, "title": "RepoGraph: Enhancing AI Software Engineering with Repository-level Code Graph", "abs": "Large Language Models (LLMs) excel in code generation yet struggle with modern AI software engineering tasks. Unlike traditional function-level or file-level coding tasks, AI software engineering requires not only basic coding proficiency but also advanced skills in managing and interacting with code repositories. However, existing methods often overlook the need for repository-level code understanding, which is crucial for accurately grasping the broader context and developing effective solutions. On this basis, we present RepoGraph, a plug-in module that manages a repository-level structure for modern AI software engineering solutions. RepoGraph offers the desired guidance and serves as a repository-wide navigation for AI software engineers. We evaluate RepoGraph on the SWE-bench by plugging it into four different methods of two lines of approaches, where RepoGraph substantially boosts the performance of all systems, leading to *a new state-of-the-art* among open-source frameworks. Our analyses also demonstrate the extensibility and flexibility of RepoGraph by testing on another repo-level coding benchmark, CrossCodeEval. Our code is available at https://anonymous.4open.science/r/Repo_Graph.", "keywords": [ "AI Software Engineering", "Large Language Models", "Code Intelligence", "Graph Structure" ], "rating_list": [ 5, 8, 6, 6, 6 ], "soundness_list": [ 2, 4, 2, 3, 3 ], "presentation_list": [ 3, 4, 3, 4, 3 ], "contribution_list": [ 3, 4, 2, 4, 3 ], "confidence_list": [ 4, 5, 4, 4, 3 ] }, { "paper_id": "dugoA2gfhs", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Frederic Sala", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Huaxiu Yao", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Jiayi Chen", "gender": "Female", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Undergrad student" }, { "name": "Jieyu Zhang", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Yifei Zhang", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "Undergrad student" }, { "name": "Yusen Jiao", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Intern" }, { "name": "Zhaoyang Li", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 40, "n_ref": 72, "n_ref_all": 92, "n_fig": 10, "n_tab": 1, "L_tab": 381, "n_element_tab": 18, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 952, "formula_len_all_1": 766, "len_all": 123587, "len_all_1": 57988, "len_abs": 1568, "len_title": 120, "len_sents": 39335, "len_sents_1": 28899, "n_sents": 262, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1572, "title": "Just Select Twice: Leveraging Low Quality Data to Improve Data Selection", "abs": "Data valuation is crucial for assessing the impact and quality of individual data points, enabling the ranking of data by importance for efficient data collection, storage, and training. Many data valuation methods are sensitive to outliers and require a certain level of noise to effectively distinguish low-quality data from high-quality data, making them particularly useful for data removal tasks. In particular, optimal transport-based methods exhibit notable performance in outlier detection but show only moderate effectiveness in high-quality data selection, due to their sensitivity to outliers and insensitivity to small variations. To mitigate the issue of insensitivity to high-quality data and facilitate effective data selection, in this paper, we propose a straightforward two-stage approach, JST, that initially does data valuation as usual, but then performs a second-round data selection where the identified low-quality data points are designated as the validation set to perform data valuation again. In this way, high-quality data become outliers with respect to the new validation set and can be naturally identified. We empirically evaluate an instantiation of our framework based on optimal transport method for data selection and data pruning on several standard datasets and our framework demonstrates superior performance compared to pure data valuation, especially under small noise conditions. Additionally, we show the general applicability of our framework to influence function based and reinforcement learning based data valuation methods.", "keywords": [ "data selection", "data valuation", "data-centric AI", "optimal transport", "robust statistics" ], "rating_list": [ 3, 3, 8, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "dug02AimLZ", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "David Fridovich-Keil", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Kushagra Gupta", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Ross Emerson Allen", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Xinjie Liu", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "ufuk topcu", "gender": "Unspecified", "institution": "University of Texas, Austin", "country": "US", "position": "Full Professor" } ], "n_formula": 61, "n_formula_1": 28, "n_ref_uni": 28, "n_ref": 59, "n_ref_all": 88, "n_fig": 6, "n_tab": 2, "L_tab": 245, "n_element_tab": 29, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7565, "formula_len_all_1": 3491, "len_all": 129565, "len_all_1": 54149, "len_abs": 1350, "len_title": 123, "len_sents": 40744, "len_sents_1": 22750, "n_sents": 392, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1378, "title": "Second-Order Algorithms for Finding Local Nash Equilibria in Zero-Sum Games", "abs": "Zero-sum games arise in a wide variety of problems, including robust optimization and adversarial learning. However, algorithms deployed for finding a local Nash equilibrium in these games often converge to non-Nash stationary points. This highlights a key challenge: for any algorithm, the stability properties of its underlying dynamical system can cause non-Nash points to be potential attractors. To overcome this challenge, algorithms must account for subtleties involving the curvatures of players' costs. To this end, we leverage dynamical system theory and develop a second-order algorithm for finding a local Nash equilibrium in the smooth, possibly nonconvex-nonconcave, zero-sum game setting. First, we prove that this novel method guarantees convergence to only local Nash equilibria with a local $\\textit{linear}$ convergence rate. We then interpret a version of this method as a modified Gauss-Newton algorithm with local $\\textit{superlinear}$ convergence to the neighborhood of a point that satisfies first-order local Nash equilibrium conditions. In comparison, current related state-of-the-art methods do not offer convergence rate guarantees. Furthermore, we show that this approach naturally generalizes to settings with convex and potentially coupled constraints while retaining earlier guarantees of convergence to only local (generalized) Nash equilibria.", "keywords": [ "game theory", "nonconvex-nonconcave optimization", "dynamical systems", "Nash equilibrium" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "duGygkA3QR", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Andi Han", "gender": "Male", "institution": "RIKEN AIP", "country": "JP", "position": "Postdoc" }, { "name": "Dai Shi", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Researcher" }, { "name": "Junbin Gao", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Full Professor" }, { "name": "Lequan Lin", "gender": "Female", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Yi Guo", "gender": "unknown", "institution": "Western Sydney University", "country": "", "position": "Associate Professor" }, { "name": "Zhiyong Wang", "gender": "Male", "institution": "The University of Sydney", "country": "AU", "position": "Full Professor" } ], "n_formula": 56, "n_formula_1": 33, "n_ref_uni": 68, "n_ref": 142, "n_ref_all": 178, "n_fig": 3, "n_tab": 8, "L_tab": 2752, "n_element_tab": 215, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2301, "n_element_tab_1": 160, "formula_len_all": 3150, "formula_len_all_1": 1860, "len_all": 242905, "len_all_1": 74618, "len_abs": 1812, "len_title": 106, "len_sents": 75885, "len_sents_1": 33023, "n_sents": 511, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1821, "title": "When Graph Neural Networks Meet Dynamic Mode Decomposition", "abs": "Graph Neural Networks (GNNs) have emerged as fundamental tools for a wide range of prediction tasks on graph-structured data. Recent studies have drawn analogies between GNN feature propagation and diffusion processes, which can be interpreted as dynamical systems. In this paper, we delve deeper into this perspective by connecting the dynamics in GNNs to modern Koopman theory and its numerical method, Dynamic Mode Decomposition (DMD). We illustrate how DMD can estimate a low-rank, finite-dimensional linear operator based on multiple states of the system, effectively approximating potential nonlinear interactions between nodes in the graph. This approach allows us to capture complex dynamics within the graph accurately and efficiently. We theoretically establish a connection between the DMD-estimated operator and the original dynamic operator between system states. Building upon this foundation, we introduce a family of DMD-GNN models that effectively leverage the low-rank eigenfunctions provided by the DMD algorithm. We further discuss the potential of enhancing our approach by incorporating domain-specific constraints such as symmetry into the DMD computation, allowing the corresponding GNN models to respect known physical properties of the underlying system. Our work paves the path for applying advanced dynamical system analysis tools via GNNs. We validate our approach through extensive experiments on various learning tasks, including directed graphs, large-scale graphs, long-range interactions, and spatial-temporal graphs. We also empirically verify that our proposed models can serve as powerful encoders for link prediction tasks. The results demonstrate that our DMD-enhanced GNNs achieve state-of-the-art performance, highlighting the effectiveness of integrating DMD into GNN frameworks.", "keywords": [ "Dynamic Mode Decomposition", "Graph Neural Networks" ], "rating_list": [ 8, 6, 8, 6, 5 ], "soundness_list": [ 3, 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 4, 3, 2 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 3, 2, 2, 3, 3 ] }, { "paper_id": "duCs92vmMc", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Jing Yang", "gender": "Female", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Jinouwen Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Rongkun Xue", "gender": "Female", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Yazhe Niu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yu Liu", "gender": "Male", "institution": "SenseTime", "country": "HK", "position": "Principal Researcher" }, { "name": "Yun Chen", "gender": "unknown", "institution": "Peking University", "country": "", "position": "MS student" } ], "n_formula": 61, "n_formula_1": 18, "n_ref_uni": 51, "n_ref": 114, "n_ref_all": 145, "n_fig": 5, "n_tab": 16, "L_tab": 7404, "n_element_tab": 566, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1302, "n_element_tab_1": 127, "formula_len_all": 5483, "formula_len_all_1": 1700, "len_all": 221489, "len_all_1": 59458, "len_abs": 3977, "len_title": 137, "len_sents": 55106, "len_sents_1": 26507, "n_sents": 425, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1366, "title": "Revisiting Generative Policies: A Simpler Reinforcement Learning Algorithmic Perspective", "abs": "Generative models, particularly diffusion models, have achieved remarkable success in density estimation for multimodal data, drawing significant interest from the reinforcement learning (RL) community, especially in policy modeling in continuous action spaces. However, existing works exhibit significant variations in training schemes and RL optimization objectives, and some methods are only applicable to diffusion models. In this study, we compare and analyze various generative policy training and deployment techniques, identifying and validating effective designs for generative policy algorithms. Specifically, we revisit existing training objectives and classify them into two categories, each linked to a simpler approach. The first approach, Generative Model Policy Optimization (GMPO), employs a native advantage-weighted regression formulation as the training objective, which is significantly simpler than previous methods. The second approach, Generative Model Policy Gradient (GMPG), offers a numerically stable implementation of the native policy gradient method. We introduce a standardized experimental framework named GenerativeRL. Our experiments demonstrate that the proposed methods achieve state-of-the-art performance on various offline-RL datasets, offering a unified and practical guideline for training and deploying generative policies.", "keywords": [ "Generative Model", "Reinforcement Learning", "Diffusion Model", "Offline Reinforcement Learning" ], "rating_list": [ 5, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 1, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "du9reSRIo1", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Andr\u00e9 Hottung", "gender": "unknown", "institution": "Bielefeld University", "country": "DE", "position": "Postdoc" }, { "name": "Chuanbo Hua", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Federico Berto", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Jinkyoo Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Junyoung Park", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "US", "position": "Researcher" }, { "name": "Kevin Tierney", "gender": "Male", "institution": "Bielefeld University", "country": "DE", "position": "Full Professor" }, { "name": "Leon Lan", "gender": "unknown", "institution": "Vrije Universiteit Amsterdam", "country": "", "position": "PhD student" }, { "name": "Nayeli Gast Zepeda", "gender": "Female", "institution": "Universit\u00e4t Bielefeld", "country": "DE", "position": "PhD student" }, { "name": "Niels Wouda", "gender": "Male", "institution": "Rotterdam School of Management", "country": "NL", "position": "Postdoc" } ], "n_formula": 30, "n_formula_1": 12, "n_ref_uni": 55, "n_ref": 132, "n_ref_all": 156, "n_fig": 4, "n_tab": 13, "L_tab": 12302, "n_element_tab": 947, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 5316, "n_element_tab_1": 364, "formula_len_all": 2185, "formula_len_all_1": 1702, "len_all": 269521, "len_all_1": 68953, "len_abs": 2210, "len_title": 115, "len_sents": 72026, "len_sents_1": 26236, "n_sents": 488, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1020, "title": "RouteFinder: Towards Foundation Models for Vehicle Routing Problems", "abs": "This paper introduces RouteFinder, a comprehensive foundation model framework to tackle different Vehicle Routing Problem (VRP) variants. Our core idea is that a foundation model for VRPs should be able to represent variants by treating each as a subset of a generalized problem equipped with different attributes. We propose a unified VRP environment capable of efficiently handling any attribute combination. The RouteFinder model leverages a modern transformer-based encoder and global attribute embeddings to improve task representation. Additionally, we introduce two reinforcement learning techniques to enhance multi-task performance: mixed batch training, which enables training on different variants at once, and multi-variant reward normalization to balance different reward scales. Finally, we propose efficient adapter layers that enable fine-tuning for new variants with unseen attributes. Extensive experiments on 48 VRP variants show RouteFinder achieves competitive results. Our code is openly available.", "keywords": [ "Vehicle Routing Problems", "VRP", "Foundation Models", "Neural Combinatorial Optimization" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "du7iixIeke", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Hong Chen", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Ruixuan Liu", "gender": "Female", "institution": "Emory University", "country": "US", "position": "Postdoc" }, { "name": "YANG CAO", "gender": "Unspecified", "institution": "Tokyo Institute of Technology", "country": "JP", "position": "Associate Professor" }, { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yuhan Liu", "gender": "Female", "institution": "Renmin University of China", "country": "", "position": "PhD student" }, { "name": "Yuncheng Wu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Sha Haichao", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 178, "n_formula_1": 20, "n_ref_uni": 24, "n_ref": 91, "n_ref_all": 112, "n_fig": 4, "n_tab": 14, "L_tab": 6942, "n_element_tab": 419, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1020, "n_element_tab_1": 44, "formula_len_all": 25461, "formula_len_all_1": 2166, "len_all": 235359, "len_all_1": 70104, "len_abs": 1770, "len_title": 134, "len_sents": 57847, "len_sents_1": 30837, "n_sents": 656, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1519, "title": "Clip Body and Tail Separately: High Probability Guarantees for DP-SGD with Heavy Tails", "abs": "Differentially Private Stochastic Gradient Descent (DPSGD) is widely utilized to preserve training data privacy in deep learning, which first clips the gradients to a predefined norm and then injects calibrated noise into the training procedure. Existing DPSGD works typically assume the gradients follow sub-Gaussian distributions and design various gradient clipping mechanisms to optimize training performance. However, recent studies have shown that the gradients in deep learning exhibit a heavy-tail phenomenon, that is, the tails of the gradient may have infinite variance, which leads to excessive clipping loss with existing mechanisms. To address this problem, we propose a novel approach, Discriminative Clipping~(DC)-DPSGD, with two key designs. First, we introduce a subspace identification technique to distinguish between body and tail gradients. Second, we present a discriminative clipping mechanism that applies different clipping thresholds separately for body and tail gradients to reduce the clipping loss. Under the non-convex condition and heavy-tailed sub-Weibull gradient noise assumption, DC-DPSGD reduces the empirical risk from ${\\mathbb{O}\\left(\\log^{\\max(0,\\theta-1)}(T/\\delta)\\log^{2\\theta}(\\sqrt{T})\\right)}$ to ${\\mathbb{O}\\left(\\log(\\sqrt{T})\\right)}$ with heavy-tailed index $\\theta> 1/2$, iterations $T$, and high probability $1-\\delta$. Extensive experiments on five real-world datasets demonstrate that our approach outperforms three baselines by up to 9.72\\% in terms of accuracy.", "keywords": [ "Privacy", "DPSGD", "Gradient Clipping", "High Probability Bounds" ], "rating_list": [ 3, 8, 1, 6, 3 ], "soundness_list": [ 1, 3, 1, 3, 2 ], "presentation_list": [ 1, 3, 1, 3, 2 ], "contribution_list": [ 2, 3, 1, 2, 3 ], "confidence_list": [ 4, 2, 5, 3, 4 ] }, { "paper_id": "dsiwLm8yjz", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Shuhan Zhu", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Xiaoyan Sun", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xinmei Tian", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yonggang Zhang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 13, "n_ref": 28, "n_ref_all": 47, "n_fig": 6, "n_tab": 6, "L_tab": 820, "n_element_tab": 136, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 708, "n_element_tab_1": 132, "formula_len_all": 725, "formula_len_all_1": 725, "len_all": 111394, "len_all_1": 61382, "len_abs": 1496, "len_title": 132, "len_sents": 34597, "len_sents_1": 27909, "n_sents": 264, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1503, "title": "Prompt Reverse Learning: Enhancing Visual Language Models for Rare Image Recognition", "abs": "Large visual language models like CLIP have demonstrated impressive performance on various downstream tasks involving common data, e.g., natural images, by leveraging prompt learning. However, these models often falter when applied to tasks involving rare data, e.g., medical images. We provide an experimental insight into this phenomenon: CLIP is insensitive to the class names of rare images. For instance, replacing the class name \u201cmedulloblastoma\u201d (a type of brain tumor) with \u201cdog\u201d in prompts has minimal impact on performance, a phenomenon not observed with common images. This highlights the disparity in representation learning between common and rare data. To realign prompt learning with rare image recognition, we propose a novel prompt learning strategy, termed prompt reverse learning (PeLen). Different from the existing methods that adapt CLIP's representations to downstream tasks, PeLen adapts task-specific representations to CLIP's representations. Built upon the insensitivity to the class names of rare images, PeLen designates common images and their class names to represent a specific class of rare images and class names, e.g., allowing the image and text of a dog to correspond to the image and text of medulloblastoma. Consequently, PeLen learns prompts to align the representations between the rare images and the visual and textual representations of common images. Our experiments on three types of rare images demonstrate the efficacy of PeLen for rare image recognition.", "keywords": [ "Prompt Learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "dsP91M4hDL", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Shen Yan", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Sijun Zhang", "gender": "unknown", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Xingyan Bin", "gender": "Male", "institution": "ByteDance Inc.", "country": "", "position": "Researcher" }, { "name": "Yisen Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhouchen Lin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Professor" } ], "n_formula": 21, "n_formula_1": 18, "n_ref_uni": 28, "n_ref": 57, "n_ref_all": 83, "n_fig": 13, "n_tab": 3, "L_tab": 1198, "n_element_tab": 233, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1198, "n_element_tab_1": 233, "formula_len_all": 1066, "formula_len_all_1": 933, "len_all": 120619, "len_all_1": 59891, "len_abs": 1413, "len_title": 112, "len_sents": 35150, "len_sents_1": 24266, "n_sents": 294, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1419, "title": "TC-MoE: Augmenting Mixture of Experts with Ternary Expert Choice", "abs": "The Mixture of Experts (MoE) architecture has emerged as a promising solution for reducing computational overhead by selectively activating subsets of model parameters. The effectiveness of MoE models is primarily dependent on their routing mechanisms, with the widely adopted Top-K routing scheme used to activate experts. However, the Top-K scheme has notable limitations, including unnecessary activations and underutilization of existing experts. In this work, rather than modifying the routing mechanism as in previous studies, we propose Ternary Choice MoE (TC-MoE), a novel approach that expands the expert space by multiplying each expert with the ternary set {-1, 0, 1}. This expansion allows for more efficient and effective expert activations without incurring significant computational cost. Additionally, given the unique characteristics of the expanded expert space, we introduce a new load balancing loss and reward loss to ensure workload balance and achieve a flexible trade-off between effectiveness and efficiency. Extensive experiments demonstrate that TC-MoE achieves an average improvement of more than 1.1% over the traditional approaches, while reducing the average number of activated experts by up to 9%. These results confirm that TC-MoE effectively address the inefficiencies of classical routing schemes, offering a more efficient and scalable solution for MoE-based large language models.", "keywords": [ "Large Language Models", "Mixture of Experts" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "dsHpulHpOK", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Jacob Adamczyk", "gender": "Male", "institution": "University of Massachusetts Boston", "country": "US", "position": "PhD student" }, { "name": "Josiah C Kratz", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 22, "n_ref_uni": 35, "n_ref": 50, "n_ref_all": 73, "n_fig": 6, "n_tab": 2, "L_tab": 216, "n_element_tab": 19, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1464, "formula_len_all_1": 1493, "len_all": 128252, "len_all_1": 63267, "len_abs": 2247, "len_title": 128, "len_sents": 33268, "len_sents_1": 29894, "n_sents": 238, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 80, "L_abs": 1266, "title": "Reinforcement Learning for Control of Non-Markovian Cellular Population Dynamics", "abs": "Many organisms and cell types, from bacteria to cancer cells, exhibit a remarkable ability to adapt to fluctuating environments. Additionally, cells can leverage memory of past environments to better survive previously-encountered stressors. From a control perspective, this adaptability poses significant challenges in driving cell populations toward extinction, and is thus an open question with great clinical significance. In this work, we focus on drug dosing in cell populations exhibiting phenotypic plasticity. For specific dynamical models switching between resistant and susceptible states, exact solutions are known. However, when the underlying system parameters are unknown, and for complex memory-based systems, obtaining the optimal solution is currently intractable. To address this challenge, we apply reinforcement learning (RL) to identify informed dosing strategies to control cell populations evolving under novel non-Markovian dynamics. We find that model-free deep RL is able to recover exact solutions and control cell populations even in the presence of long-range temporal dynamics. To further test our approach in more realistic settings, we demonstrate performant RL-based control strategies in environments with dynamic memory strength.", "keywords": [ "optimal drug dosing", "fractional differential equations", "reinforcement learning", "control theory" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 4, 4, 3 ], "presentation_list": [ 4, 4, 4, 3 ], "contribution_list": [ 4, 3, 4, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "dsALpkd1OU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 1, "author_info_list": [ { "name": "Yinghui Xia", "gender": "Male", "institution": "AutoAgents", "country": "CN", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 13, "n_ref": 15, "n_ref_all": 24, "n_fig": 2, "n_tab": 2, "L_tab": 260, "n_element_tab": 27, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 260, "n_element_tab_1": 27, "formula_len_all": 190, "formula_len_all_1": 190, "len_all": 65360, "len_all_1": 65360, "len_abs": 1380, "len_title": 130, "len_sents": 19184, "len_sents_1": 19184, "n_sents": 129, "n_sents_1": 129, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1385, "title": "D2Coder: large language models based agent for coding with dynamic debugging tools", "abs": "Intelligent agents based on large language models have demonstrated certain programming abilities, but there is still significant room for improvement in complex project-level debugging tasks. Previous work has utilized general multi-agent workflows to enhance performance but has the following issues: 1) excessive reliance on the reasoning capabilities of large language models without debugging and detailed analysis of the code; 2) lack of intrinsic code information, such as call relationships and dependencies; 3) insufficient analysis and optimization of critical stages, especially the code search capability in fault localization, which directly affects the effectiveness of subsequent stages. Based on the SWE-bench dataset, we first isolate the fault localization capability for separate analysis and experiments, and introduce program call graphs to demonstrate the effectiveness of this information for debugging. Furthermore, during the debugging phase, we propose a simulated debugging mode that enables large language models to simulate program debugging without relying on other debugging tools. Compared to the real machine debugging mode, our experiments prove the effectiveness and generality of the simulated debugging mode. We conducted experiments on SWE-bench and improved the resolution rate by approximately 27.3\\%, demonstrating the potential of this method.", "keywords": [ "LLM-based Agent", "Call Graph", "Dynamic Debugging", "Fault Localization" ], "rating_list": [ 1, 3, 1 ], "soundness_list": [ 1, 2, 2 ], "presentation_list": [ 1, 2, 2 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "ds6Mmd7LlH", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Fangzhou Lin", "gender": "Male", "institution": "Tohoku University, Tokyo Institute of Technology", "country": "JP", "position": "Researcher" }, { "name": "Haichong Zhang", "gender": "Male", "institution": "Worcester Polytechnic Institute", "country": "", "position": "Assistant Professor" }, { "name": "Jose Morales", "gender": "Male", "institution": "Worcester Polytechnic Institute", "country": "US", "position": "PhD student" }, { "name": "Kazunori Yamada", "gender": "Male", "institution": "Tohoku University", "country": "", "position": "Full Professor" }, { "name": "Venkatesh Saligrama", "gender": "unknown", "institution": "Boston University", "country": "US", "position": "Full Professor" }, { "name": "Vijaya B Kolachalama", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "Associate Professor" }, { "name": "Yun Yue", "gender": "unknown", "institution": "Worcester Polytechnic Institute", "country": "US", "position": "PhD student" }, { "name": "Ziming Zhang", "gender": "Male", "institution": "Worcester Polytechnic Institute", "country": "US", "position": "Assistant Professor" } ], "n_formula": 29, "n_formula_1": 25, "n_ref_uni": 68, "n_ref": 128, "n_ref_all": 154, "n_fig": 4, "n_tab": 6, "L_tab": 3120, "n_element_tab": 200, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 126, "n_element_tab_1": 18, "formula_len_all": 2707, "formula_len_all_1": 2927, "len_all": 161095, "len_all_1": 67164, "len_abs": 1367, "len_title": 87, "len_sents": 33410, "len_sents_1": 28793, "n_sents": 245, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1371, "title": "GRADIENT-OPTIMIZED CONTRASTIVE LEARNING", "abs": "Contrastive learning is a crucial technique in representation learning, producing robust embeddings by distinguishing between similar and dissimilar pairs. In this paper, we introduce a novel framework, Gradient-Optimized Contrastive Learning (GOAL), which enhances network training by optimizing gradient updates during backpropagation as a bilevel optimization problem. Our approach offers three key insights that set it apart from existing methods: (1) Contrastive learning can be seen as an approximation of a one-class support vector machine (OC-SVM) using multiple neural tangent kernels (NTKs) in the network\u2019s parameter space; (2) Hard triplet samples are vital for defining support vectors and outliers in OC-SVMs within NTK spaces, with their difficulty measured using Lagrangian multipliers; (3) Contrastive losses like InfoNCE provide efficient yet dense approximations of sparse Lagrangian multipliers by implicitly leveraging gradients. To address the computational complexity of GOAL, we propose a novel contrastive loss function, Sparse InfoNCE (SINCE), which improves the Lagrangian multiplier approximation by incorporating hard triplet sampling into InfoNCE. Our experimental results demonstrate the effectiveness and efficiency of SINCE in tasks such as image classification and point cloud completion. Demo code is attached in the supplementary file.", "keywords": [ "Contrastive learning; Sparse kernel machines; Image classification; Point cloud completion" ], "rating_list": [ 6, 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 4 ], "contribution_list": [ 2, 3, 2, 3, 4 ], "confidence_list": [ 3, 4, 3, 3, 4 ] }, { "paper_id": "ds3Tcnrte8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Chen Chen", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Assistant Professor" }, { "name": "Haochen Liu", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Jundong Li", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" }, { "name": "Song Wang", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 17, "n_ref_uni": 29, "n_ref": 58, "n_ref_all": 67, "n_fig": 3, "n_tab": 2, "L_tab": 842, "n_element_tab": 47, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 842, "n_element_tab_1": 47, "formula_len_all": 698, "formula_len_all_1": 683, "len_all": 121225, "len_all_1": 62550, "len_abs": 1672, "len_title": 114, "len_sents": 37856, "len_sents_1": 28540, "n_sents": 278, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1678, "title": "Question-Aware Knowledge Graph Prompting for Large Language Models", "abs": "Large Language Models (LLMs) have demonstrated significant advancements in various natural language processing tasks, yet they often struggle with tasks that require external domain-specific knowledge, such as Multiple Choice Question Answering (MCQA). Integrating Knowledge Graphs (KGs) with LLMs has been explored as a solution to enhance LLMs' reasoning capabilities, while existing methods either involve computationally expensive finetuning processes or rely on the noisy retrieval of KG information. Recent efforts have focused on leveraging Graph Neural Networks (GNNs) to generate KG-based soft prompts for LLMs, which face challenges of lacking question-relevance assessment in GNN and utilization of relations among options. In this paper, we propose a novel approach, QAP, to address these challenges by optimizing the utilization of KG in MCQA tasks. Our method introduces question embeddings into the GNN aggregation process, enabling the model to assess the relevance of KG information based on the question context. Additionally, QAP facilitates inter-option interactions by employing an attention module that explicitly models relationships between answer options. Specifically, we use multiple attention heads for the GNN output, allowing the model to capture and compare features across different options, thereby enhancing cross-option reasoning. Our approach not only enhances the connection between GNNs and LLMs but also enables the model to better utilize the relationships between answer options. Experimental results demonstrate that QAP outperforms state-of-the-art models on multiple public MCQA datasets, validating its effectiveness and scalability.", "keywords": [ "Knowledge Graph", "Question Answering", "Large Language Model", "Prompt" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 3, 2, 1 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "drrXhD2r8V", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Jian Lou", "gender": "unknown", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jiaqi Liu", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Kui Ren", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Wenjie Bao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Xiaochen Li", "gender": "Female", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Postdoc" }, { "name": "Yuke Hu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhan Qin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zhihao Liu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 19, "n_ref_uni": 39, "n_ref": 74, "n_ref_all": 96, "n_fig": 5, "n_tab": 14, "L_tab": 10069, "n_element_tab": 1654, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 952, "n_element_tab_1": 138, "formula_len_all": 1939, "formula_len_all_1": 1662, "len_all": 178901, "len_all_1": 71458, "len_abs": 1677, "len_title": 124, "len_sents": 45638, "len_sents_1": 32163, "n_sents": 342, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1661, "title": "Structure-Aware Parameter-Efficient Machine Unlearning on Transformer Models", "abs": "Transformer has become fundamental to a vast series of pretrained large models that have achieved remarkable success across diverse applications. Machine unlearning is an emerging field focused on efficiently removing the influence of specific data from trained models, to comply with privacy regulations enforcing the right to be forgotten. The sheer size of Transformer-based models poses a significant challenge to unlearning efficiency. Existing methods find it promising to restrict unlearning updates to a small portion of influence-critical parameters. However, their parameter-efficient unlearning methods are largely devised in a structure-oblivious manner, which tends to inaccurately identify these parameters and leads to inferior unlearning performance for Transformers. In this paper, we propose {\\tt SPE-Unlearn}, a structure-aware parameter-efficient machine unlearning approach tailored for the Transformer architecture. {\\tt SPE-Unlearn} introduces a learnable pair of masks to respectively pinpoint influence-critical parameters in the heads and filters of Transformers. The learning objective of these masks is derived by jointly considering both desiderata of unlearning, i.e., sufficiency in influence removal and efficiency, and optimized through an efficient algorithm featured by a greedy search with a warm start. Equipped with the identified key parameters, {\\tt SPE-Unlearn} facilitates second-order unlearning, memory-free unlearning, and memory-aided unlearning scenarios. Extensive experiments on various transformer models and datasets demonstrate the effectiveness and efficiency of {\\tt SPE-Unlearn}~for Transformer unlearning.", "keywords": [ "Machine Unlearning", "Parameter-Efficient", "Transformer" ], "rating_list": [ 3, 6, 6, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "drPDukdY3t", "primary_area": "applications to neuroscience & cognitive science", "n_author": 8, "author_info_list": [ { "name": "Bing Li", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Jin Gao", "gender": "Male", "institution": "Institute of automation, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Li Yang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "Assistant Professor" }, { "name": "Shuxun Wang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Wei Liu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Weiming Hu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "Full Professor" }, { "name": "Wenjuan Li", "gender": "Female", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Zhao Mingxuan", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 26, "n_ref": 46, "n_ref_all": 61, "n_fig": 3, "n_tab": 6, "L_tab": 2107, "n_element_tab": 263, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1597, "n_element_tab_1": 186, "formula_len_all": 487, "formula_len_all_1": 488, "len_all": 95752, "len_all_1": 49872, "len_abs": 1728, "len_title": 139, "len_sents": 25403, "len_sents_1": 22641, "n_sents": 165, "n_sents_1": 141, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1733, "title": "DeepTAGE: Deep Temporal-Aligned Gradient Enhancement for Optimizing Spiking Neural Networks", "abs": "Spiking Neural Networks (SNNs), with their biologically inspired spatio-temporal dynamics and spike-driven processing, are emerging as a promising low-power alternative to traditional Artificial Neural Networks (ANNs). However, the complex neuronal dynamics and non-differentiable spike communication mechanisms in SNNs present substantial challenges for efficient training. By analyzing the membrane potentials in spiking neurons, we found that their distributions can increasingly deviate from the firing threshold as time progresses, which tends to cause diminished backpropagation gradients and unbalanced optimization. To address these challenges, we propose Deep Temporal-Aligned Gradient Enhancement (DeepTAGE), a novel approach that improves optimization gradients in SNNs from both internal surrogate gradient functions and external supervision methods. Our DeepTAGE dynamically adjusts surrogate gradients in accordance with the membrane potential distribution across different time steps, enhancing their respective gradients in a temporal-aligned manner that promotes balanced training. Moreover, to mitigate issues of gradient vanishing or deviating during backpropagation, DeepTAGE incorporates deep supervision at both spatial (network stages) and temporal (time steps) levels to ensure more effective and robust network optimization. Importantly, our method can be seamlessly integrated into existing SNN architectures without imposing additional inference costs or requiring extra control modules. We validate the efficacy of DeepTAGE through extensive experiments on static benchmarks (CIFAR10, CIFAR100, and ImageNet-1k) and a neuromorphic dataset (DVS-CIFAR10), demonstrating significant performance improvements.", "keywords": [ "Spiking Neural Networks", "Deep Temporal-Aligned Gradient Enhancement", "Spatio-Temporal Deep Supervision", "Optimization Imbalance." ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 2 ] }, { "paper_id": "dr0s6aGYb7", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Jonathan Roberts", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Kai Han", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Wenye Lin", "gender": "unknown", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yunhan Yang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zongqing Lu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "samuel Albanie", "gender": "Not Specified", "institution": "Google", "country": "GB", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 36, "n_ref": 75, "n_ref_all": 90, "n_fig": 3, "n_tab": 11, "L_tab": 23385, "n_element_tab": 564, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2431, "n_element_tab_1": 406, "formula_len_all": 88, "formula_len_all_1": 15, "len_all": 198815, "len_all_1": 66303, "len_abs": 2370, "len_title": 109, "len_sents": 53692, "len_sents_1": 28156, "n_sents": 439, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1584, "title": "GAMEBOT: Gaming Arena for Model Evaluation - Battle of Tactics", "abs": "Large Language Models (LLMs) are increasingly deployed in real-world applications that demand complex reasoning. To track progress, we require robust benchmarks to evaluate their capabilities beyond superficial pattern recognition. However, existing benchmarks either suffer from data contamination or lack legibility. In this paper, we introduce GAMEBOT, a novel benchmark for evaluating LLMs in competitive gaming environments that addresses these limitations. GAMEBOT decomposes complex reasoning in games into modular subproblems, targeting abilities like rule understanding and strategy instruction following. We develop Chain-of-Thought (CoT) prompts that leverage domain knowledge to guide LLMs and automatically validate their intermediate reasoning steps against ground truth. This approach allows us to assess not only the accuracy of final decisions but also the quality of the underlying reasoning process. We benchmark 17 prominent LLMs across eight diverse games, encompassing various strategic abilities and game characteristics. GAMEBOT offers four advantages: (1) Mitigation of Data Contamination: Dynamic game states minimize overlap with pre-training data. (2) Legibility: Evaluation of intermediate reasoning steps enables fine-grained scrutiny of LLM behavior. (3) Difficulty: The games effectively differentiate top-performing models. (4) Stronger Baselines: Our curated CoT prompts establish competitive baselines for future research. We hope GAMEBOT stimulates further work that seeks a deeper understanding of LLM reasoning capabilities in strategic settings.", "keywords": [ "LLM evaluation", "benchmark", "competitive game" ], "rating_list": [ 3, 1, 5, 5 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "dqyuCsBvn9", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 6, "author_info_list": [ { "name": "Alexander Wilentz", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "MS student" }, { "name": "Alina Zare", "gender": "Female", "institution": "University of Florida", "country": "", "position": "Full Professor" }, { "name": "James Fairbanks", "gender": "Male", "institution": "University of Florida", "country": "", "position": "Assistant Professor" }, { "name": "Mason Lary", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "", "position": "PhD student" }, { "name": "Matthew Klawonn", "gender": "unknown", "institution": "Air Force Research Laboratory, Information Directorate", "country": "", "position": "Researcher" }, { "name": "Richard Samuelson", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 27, "n_ref": 40, "n_ref_all": 48, "n_fig": 1, "n_tab": 2, "L_tab": 310, "n_element_tab": 32, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 542, "n_element_tab_1": 41, "formula_len_all": 707, "formula_len_all_1": 975, "len_all": 105955, "len_all_1": 60735, "len_abs": 1144, "len_title": 126, "len_sents": 35154, "len_sents_1": 29902, "n_sents": 246, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1149, "title": "Learning Diagrams: A Graphical Language for Compositional Training Regimes", "abs": "Motivated by deep learning regimes with multiple interacting yet distinct model components, we introduce learning diagrams, graphical depictions of training setups that capture parameterized learning as data rather than code. A learning diagram compiles to a unique loss function on which component models are trained. The result of training on this loss is a collection of models whose predictions ``agree\" with one another. We show that a number of popular learning setups such as few-shot multi-task learning, knowledge distillation, and multi-modal learning can be depicted as learning diagrams. We further implement learning diagrams in a library that allows users to build diagrams of PyTorch and Flux.jl models. By implementing some classic machine learning use cases, we demonstrate how learning diagrams allow practitioners to build complicated models as compositions of smaller components, identify relationships between workflows, and manipulate models during or after training. Leveraging a category theoretic framework, we introduce a rigorous semantics for learning diagrams that puts such operations on a firm mathematical foundation.", "keywords": [ "ML Libraries", "Training", "Foundation Models", "Multi-Task Learning" ], "rating_list": [ 6, 8, 3 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 4, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "dqMqAaw7Sq", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Anke Tang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Didi Zhu", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Fei Wu", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Jinluan Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhengyu Chen", "gender": "unknown", "institution": "Meituan", "country": "CN", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 7, "n_ref_uni": 53, "n_ref": 113, "n_ref_all": 152, "n_fig": 8, "n_tab": 17, "L_tab": 14232, "n_element_tab": 1557, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 3824, "n_element_tab_1": 409, "formula_len_all": 1591, "formula_len_all_1": 945, "len_all": 201203, "len_all_1": 64560, "len_abs": 1790, "len_title": 133, "len_sents": 53720, "len_sents_1": 28630, "n_sents": 341, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1748, "title": "Mitigating the Backdoor Effect for Multi-Task Model Merging via Safety-Aware Subspace", "abs": "Model merging has gained significant attention as a cost-effective approach to integrate multiple single-task fine-tuned models into a unified one that can perform well on multiple tasks. However, existing model merging techniques primarily focus on resolving conflicts between task-specific models, they often overlook potential security threats, particularly the risk of backdoor attacks in the open-source model ecosystem. In this paper, we first investigate the vulnerabilities of existing model merging methods to backdoor attacks, identifying two critical challenges: backdoor succession and backdoor transfer. To address these issues, we propose a novel Defense-Aware Merging (DAM) approach that simultaneously mitigates task interference and backdoor vulnerabilities. Specifically, DAM employs a meta-learning-based optimization method with dual masks to identify a shared and safety-aware subspace for model merging. These masks are alternately optimized: the Task-Shared mask identifies common beneficial parameters across tasks, aiming to preserve task-specific knowledge while reducing interference, while the Backdoor-Detection mask isolates potentially harmful parameters to neutralize security threats. This dual-mask design allows us to carefully balance the preservation of useful knowledge and the removal of potential vulnerabilities. Compared to existing merging methods, DAM achieves a more favorable balance between performance and security, reducing the attack success rate by 2-10 percentage points while sacrificing only about 1\\% in accuracy. Furthermore, DAM exhibits robust performance and broad applicability across various types of backdoor attacks and the number of compromised models involved in the merging process.", "keywords": [ "Model Merging; Backdoor Defense; Subspace" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 2, 2, 3 ] }, { "paper_id": "dq3keisMjT", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Flemming Holtorf", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "RE", "position": "PhD student" }, { "name": "Frank Sch\u00e4fer", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Julian Arnold", "gender": "Male", "institution": "University of Basel", "country": "", "position": "PhD student" }, { "name": "Niels L\u00f6rch", "gender": "Male", "institution": "University of Basel", "country": "CH", "position": "Postdoc" } ], "n_formula": 25, "n_formula_1": 10, "n_ref_uni": 42, "n_ref": 68, "n_ref_all": 92, "n_fig": 6, "n_tab": 1, "L_tab": 358, "n_element_tab": 30, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2094, "formula_len_all_1": 450, "len_all": 150125, "len_all_1": 60157, "len_abs": 1436, "len_title": 117, "len_sents": 41470, "len_sents_1": 28985, "n_sents": 336, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1102, "title": "Phase Transitions in the Output Distribution of Large Language Models", "abs": "In a physical system, changing parameters such as temperature can induce a phase transition: an abrupt change from one state of matter to another. Analogous phenomena have recently been observed in large language models. Typically, the task of identifying phase transitions requires human analysis and some prior understanding of the system to narrow down which low-dimensional properties to monitor and analyze. Statistical methods for the automated detection of phase transitions from data have recently been proposed within the physics community. These methods are largely system agnostic and, as shown here, can be adapted to study the behavior of large language models. In particular, we quantify distributional changes in the generated output via statistical distances, which can be efficiently estimated with access to the probability distribution over next-tokens. This versatile approach is capable of discovering new phases of behavior and unexplored transitions -- an ability that is particularly exciting in light of the rapid development of language models and their emergent capabilities.", "keywords": [ "language models", "generative models", "phase transitions" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "dpnPOXoqVQ", "primary_area": "learning theory", "n_author": 7, "author_info_list": [ { "name": "Bin Gu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Feng Zheng", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Hong Chen", "gender": "unknown", "institution": "Huazhong Agricultural University", "country": "", "position": "Full Professor" }, { "name": "Tieliang Gong", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xuelin Zhang", "gender": "Male", "institution": "Huazhong Agricultural University", "country": "CN", "position": "PhD student" }, { "name": "Yingjie Wang", "gender": "Male", "institution": "China University of Petroleum", "country": "CN", "position": "Lecturer" }, { "name": "ZEYU ZHANG", "gender": "Male", "institution": "Huazhong Agricultural University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 144, "n_formula_1": 31, "n_ref_uni": 72, "n_ref": 251, "n_ref_all": 287, "n_fig": 11, "n_tab": 14, "L_tab": 9480, "n_element_tab": 708, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 4067, "n_element_tab_1": 251, "formula_len_all": 16720, "formula_len_all_1": 2802, "len_all": 324448, "len_all_1": 77826, "len_abs": 1194, "len_title": 137, "len_sents": 83846, "len_sents_1": 27876, "n_sents": 704, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1199, "title": "S$^2$MAM: Semi-supervised Meta Additive Model for Robust Estimation and Variable Selection", "abs": "Semi-supervised learning with manifold regularization is a classical family for learning from the labeled and unlabeled data jointly, where the key requirement is the support of unknown marginal distribution enjoys the geometric structure of a Riemannian manifold. Usually, the Laplace-Beltrami operator-based manifold regularization can be approximated empirically by the Laplacian regularization associated with the whole training data and its graph Laplacian matrix. However, the graph Laplacian matrix depends heavily on the pre-specifying similarity metric and may result in inappropriate penalties when facing redundant and noisy input variables. In order to address the above issues, this paper proposes a new semi-supervised meta additive model (S$^2$MAM) under a bilevel optimization scheme to automatically identify the informative variables, update the similarity matrix, and achieve the interpretable prediction simultaneously. Theoretical guarantees are provided for S$^2$MAM including the computing convergence and the statistical generalization bound. Experimental assessments on synthetic and real-world datasets validate the robustness and interpretability of the proposed approach.", "keywords": [ "manifold regularization", "bilevel optimization", "sparse additive model", "robustness", "learning theory" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 2, 4, 5, 5 ] }, { "paper_id": "dp1BH2bK4Y", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Heyuan Huang", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Principal Researcher" }, { "name": "Hongyan Li", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Principal Researcher" }, { "name": "Jiaxin Shi", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Shiwan Zhao", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "PhD student" }, { "name": "Sitao Xie", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yu Wang", "gender": "Female", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yubo Zhang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zhihu Wang", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Principal Researcher" }, { "name": "Zhixing Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 47, "n_ref_all": 78, "n_fig": 7, "n_tab": 15, "L_tab": 5146, "n_element_tab": 259, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1510, "n_element_tab_1": 117, "formula_len_all": 166, "formula_len_all_1": 0, "len_all": 159547, "len_all_1": 57712, "len_abs": 1795, "len_title": 129, "len_sents": 56064, "len_sents_1": 28201, "n_sents": 408, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1813, "title": "Re-TASK: Revisiting LLM Tasks from Capability, Skill, and Knowledge Perspectives", "abs": "The Chain-of-Thought (CoT) paradigm has become a pivotal method for solving complex problems. However, its application to intricate, domain-specific tasks remains challenging, as large language models (LLMs) often struggle to accurately decompose these tasks and, even when decomposition is correct, fail to execute the subtasks effectively. This paper introduces the Re-TASK framework, a novel theoretical model that revisits LLM tasks from the perspectives of capability, skill, and knowledge, drawing on the principles of Bloom's Taxonomy and Knowledge Space Theory. While CoT offers a workflow perspective on tasks, the Re-TASK framework introduces a Chain-of-Learning view, illustrating how tasks and their corresponding subtasks depend on various capability items. Each capability item is further dissected into its constituent aspects of knowledge and skills. Our framework reveals that many CoT failures in domain-specific tasks stem from insufficient knowledge or inadequate skill adaptation. In response, we combine CoT with the Re-TASK framework and implement a carefully designed Re-TASK prompting strategy to improve task performance. Specifically, we identify core capability items linked to tasks and subtasks, then strengthen these capabilities through targeted knowledge injection and skill adaptation. We validate the Re-TASK framework on three datasets across the law, finance, and mathematics domains, achieving significant improvements over the baseline models. Notably, our approach yields a remarkable 44.42\\% improvement with the Yi-1.5-9B model and a 33.08\\% improvement with the Llama3-Chinese-8b on the legal dataset. These experimental results confirm the effectiveness of the Re-TASK framework, demonstrating substantial enhancements in both the performance and applicability of LLMs.", "keywords": [ "LLM", "Task", "Capability", "Knowledge", "Skill", "Chain-of-Thought" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "doBof19Ia4", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Xian Zeng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Xin Guo", "gender": "Male", "institution": "Shanghai Academy of Artificial Intelligence for Science", "country": "CN", "position": "Principal Researcher" }, { "name": "Xuyangliu", "gender": "Male", "institution": "Shanghai Academy of Artificial Intelligence for Science", "country": "CN", "position": "Researcher" }, { "name": "Yuan Cheng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yuan Qi", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "siqi", "gender": "Male", "institution": "Shanghai Academy of Artificial Intelligence for Science", "country": "CN", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 3, "n_ref": 16, "n_ref_all": 29, "n_fig": 2, "n_tab": 2, "L_tab": 1851, "n_element_tab": 47, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1658, "n_element_tab_1": 45, "formula_len_all": 501, "formula_len_all_1": 500, "len_all": 96590, "len_all_1": 51193, "len_abs": 1407, "len_title": 126, "len_sents": 24257, "len_sents_1": 23685, "n_sents": 155, "n_sents_1": 157, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1414, "title": "GeoRDe: Tertiary Structure-based RNA Design With Multiple Geometric Constraint", "abs": "Functional RNA sequence design plays an essential role in the regulation of life processes. The RNA inverse folding problem, which involves designing nucleic acid sequences based on their three-dimensional structures, remains highly challenging. This complexity arises not only from the inherent flexibility of RNA structures but also from the base-pairing rules that impose critical spatial constraints on the RNA scaffold. In recent times, the design of RNA has often depended on geometric graph networks to design sequences. Motivated by recent advancements in protein design, we have developed the RNAformer module. This module is capable of learning the geometric constraints of RNA molecules in cooperation with geometric graph networks. Furthermore, to enhance the specificity of sequence generation, we have integrated secondary structure information as labels, ensuring that the designed sequences align more closely with secondary structure constraints. Additionally, we have used RNA language models to understand average evolutionary constraints. By incorporating a range of constraint insights, GeoRDe has demonstrated superior performance under identical training data conditions and has also showcased generalization capabilities on the independent casp15 and RNA-puzzle datasets. Through extensive experimentation, the GeoRDe has proven to be an innovative solution to the challenges of RNA design.", "keywords": [ "RNA design", "RNA Inverse Folding", "Computational biology" ], "rating_list": [ 3, 1, 3, 1, 5 ], "soundness_list": [ 3, 1, 3, 2, 2 ], "presentation_list": [ 1, 1, 2, 1, 3 ], "contribution_list": [ 3, 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "doBkiqESYq", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Anirudh Sundara Rajan", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "MS student" }, { "name": "Jedidiah Schloesser", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "Undergrad student" }, { "name": "Utkarsh Ojha", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Yong Jae Lee", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Associate Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 127, "n_ref_all": 165, "n_fig": 9, "n_tab": 6, "L_tab": 2432, "n_element_tab": 240, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1703, "n_element_tab_1": 184, "formula_len_all": 33, "formula_len_all_1": 33, "len_all": 156633, "len_all_1": 73746, "len_abs": 1729, "len_title": 114, "len_sents": 53457, "len_sents_1": 34681, "n_sents": 445, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1714, "title": "On the Effectiveness of Dataset Alignment for Fake Image Detection", "abs": "As latent diffusion models (LDMs) democratize image generation capabilities, there is a growing need to detect fake images. A good detector should focus on the generative model\u2019s fingerprints while ignoring image properties such as semantic content, resolution, file format, etc. Fake image detectors are usually built in a data-driven way, where a model is trained to separate real from fake images. Existing works primarily investigate network architecture choices and training recipes. In this work, we argue that in addition to these algorithmic choices, we also require a well-aligned dataset of real/fake images to train a robust detector. For the family of LDMs, we propose a very simple way to achieve this: we reconstruct all the real images using the LDM's autoencoder, without any denoising operation. We then train a model to separate these real images from their reconstructions. The fakes created this way are extremely similar to the real ones in almost every aspect (e.g., size, aspect ratio, semantic content), which forces the model to look for the LDM decoder's artifacts. We empirically show that this way of creating aligned real/fake datasets, which also sidesteps the computationally expensive denoising process, helps in building a detector that focuses less on spurious correlations, something that a very popular existing method is susceptible to. Finally, to demonstrate the effectivenss of dataset alignment, we build a detector using images that are not natural objects, and present promising results. Overall, our work identifies the subtle but significant issues that arise when training a fake image detector and proposes a simple and inexpensive solution to address these problems.", "keywords": [ "Image Forensics", "Latent Diffusion" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 1, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "dnp63LgTgc", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Chee Kwoh", "gender": "Male", "institution": "National Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Gabriel Tjio", "gender": "unknown", "institution": "A*STAR", "country": "SG", "position": "PhD student" }, { "name": "Nhat Chung", "gender": "Male", "institution": "FPT Software AI Center", "country": "VN", "position": "Researcher" }, { "name": "Qing Guo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Yun Xing", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "PhD student" }, { "name": "Ivor W Tsang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Xiaofeng Cao", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 31, "n_ref": 65, "n_ref_all": 78, "n_fig": 4, "n_tab": 5, "L_tab": 2681, "n_element_tab": 329, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1981, "n_element_tab_1": 313, "formula_len_all": 810, "formula_len_all_1": 810, "len_all": 137606, "len_all_1": 63298, "len_abs": 1711, "len_title": 167, "len_sents": 33133, "len_sents_1": 25719, "n_sents": 253, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 118, "L_abs": 1705, "title": "Beautifying Diffusion Models: Learning Context-Aware Filters for Robust Dense Prediction on Test-Time Corrupted Images", "abs": "Diffusion models have enabled input-based domain adaptation to unseen test-time corruption for the classification problem.\nNevertheless, while dense prediction tasks share similar robustness issues with image-level classification, previous input adaptation work may fail to preserve the semantic information necessary for robust pixel-level prediction. \nTo address the issue, we propose a novel diffusion-driven strategy that translates the corrupted inputs to the source domain (\\ie, the training data domain), while also preserving the semantic information (\\ie high-frequency shape information and low-frequency color information).\nWe first studied how to leverage frequency filtering to guide the diffusion generation process and analyze the influence of different filters.\nFrom our experiments, we observed that utilizing both high and low spatial frequency information during diffusion driven denoising can substantially improve the adaptation performance of dense classification.\nThis observation motivates us to develop a novel framework, \\ie a predictive frequency filtering-driven diffusion (FDD) adaptation, where we predict the filters from the corrupted test-time inputs and use them to guide the diffusion process.\nWe design a Y-like frequency prediction network to predict context-aware low-pass and high-pass filters.\nTo train this network, we propose a novel data augmentation method, FrequencyMix, to generate pairs of clean and corrupted images.\nWe validate our method via extensive experiments on two semantic segmentation datasets and two depth estimation datasets.\nAgainst a broad range of common corruptions, we demonstrate that our method is competitive with state of the art work.", "keywords": [ "Diffusion", "Test-time adaptation", "Dense Prediction", "Frequency-Aware Modeling" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "dnUWt1EN72", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Han Bao", "gender": "Male", "institution": "Kyoto University, Kyoto University", "country": "JP", "position": "Assistant Professor" }, { "name": "Hisashi Kashima", "gender": "Male", "institution": "Kyoto University", "country": "JP", "position": "Full Professor" }, { "name": "Koh Takeuchi", "gender": "Male", "institution": "Kyoto University", "country": "JP", "position": "Lecturer" }, { "name": "Xiaofeng Lin", "gender": "Male", "institution": "Kyoto University, Kyoto University", "country": "JP", "position": "PhD student" }, { "name": "Yan Cui", "gender": "Not Specified", "institution": "Fudan University", "country": "CN", "position": "Intern" } ], "n_formula": 26, "n_formula_1": 23, "n_ref_uni": 48, "n_ref": 125, "n_ref_all": 146, "n_fig": 10, "n_tab": 8, "L_tab": 2385, "n_element_tab": 93, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 3569, "n_element_tab_1": 63, "formula_len_all": 3021, "formula_len_all_1": 2221, "len_all": 181553, "len_all_1": 79597, "len_abs": 2167, "len_title": 131, "len_sents": 54356, "len_sents_1": 31460, "n_sents": 435, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 2113, "title": "Dynamic Interference Modeling For Estimating Treatment Effects From Dynamic Graphs", "abs": "Estimating treatment effects can assist decision-making in various areas, such as commerce and medicine. One application of the treatment effect estimation is to predict the effect of an advertisement on the purchase result of a customer, known as individual treatment effect (ITE). In online websites, the outcome of an individual can be affected by treatments of other individuals, as people often propagate information with their friends, a phenomenon referred to as interference. Prior studies have attempted to model interference for accurate ITE estimation under a static network among individuals. However, the network usually changes over time in real-world applications due to complex social activities among individuals. For instance, an individual can follow another individual on one day and unfollow this individual afterward on an online social website. In this case, the outcomes of individuals can be interfered with not only by treatments for current neighbors but also by past information and treatments for past neighbors, which we refer to as \\emph{dynamic interference}. In this work, we model dynamic interference for the first time by developing an architecture to aggregate both the past information of individuals and their neighbors. Specifically, our proposed method contains a mechanism that summarizes historical information of individuals from previous time stamps, graph neural networks that propagate information about individuals within every time stamp, and a weighting mechanism that estimates the importance of different time stamps. Moreover, the model parameters should gradually change rather than drastically because information of every individual gradually changes over time. To take it into account, we also propose a variant of our method to evolve the model parameters over time with long short-term memory. In our experiments on multiple datasets with dynamic interference, our methods outperform existing methods for ITE estimation because they are unable to capture dynamic interference. This result corroborates the importance of dynamic interference modeling.", "keywords": [ "Causal Inference", "Treatment effect estimation", "Dynamic Graph", "Interference" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "dnU9bGgSZ5", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Daoyang Hua", "gender": "unknown", "institution": "zhejiang university", "country": "CN", "position": "Postdoc" }, { "name": "De Ma", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Gang Pan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" }, { "name": "Wen Yao", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "Principal Researcher" }, { "name": "Xianqi Chen", "gender": "Male", "institution": "Chinese Academy of Military Science", "country": "CN", "position": "Assistant Professor" }, { "name": "Wang Boning", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Qian Dr. Zheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 17, "n_ref": 24, "n_ref_all": 41, "n_fig": 4, "n_tab": 3, "L_tab": 1534, "n_element_tab": 214, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1534, "n_element_tab_1": 214, "formula_len_all": 406, "formula_len_all_1": 406, "len_all": 107177, "len_all_1": 59874, "len_abs": 1109, "len_title": 115, "len_sents": 26908, "len_sents_1": 26337, "n_sents": 219, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 995, "title": "Jointly Optimizing Wirelength and Thermal Fields for Chip Placement", "abs": "Macro placement is a crucial and complex issue in chip design. In recent studies, reinforcement learning (RL) has demonstrated outstanding performance in optimizing chip wirelength, but this leads to thermally inefficient design. Additionally, due to the specialized expertise necessary for creating chip benchmarks and the constraints imposed by confidentiality agreements, there exists a scarcity of publicly available chip thermal placement benchmarks. This work introduces a reinforcement learning-based thermal placement model that can optimize both wirelength and peak temperatures. We also strictly followed the chip design process and established a macro thermal placement benchmark. This significantly reduces the entry barriers for researchers, facilitating benchmarking and result replication. Compared to other models, our model notably diminishes the chip's peak temperature of the chip while slightly extending wirelength, thereby improving the chip's heat dissipation efficiency.", "keywords": [ "Reinforcement Learning", "Chip Design", "Thermal Placement" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 2, 4 ] }, { "paper_id": "dmzM5UdAq6", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Abhishek Aich", "gender": "Male", "institution": "NEC Laboratories, America", "country": "US", "position": "Researcher" }, { "name": "Samuel Schulter", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yumin Suh", "gender": "unknown", "institution": "NEC-Labs", "country": "US", "position": "Researcher" }, { "name": "Manmohan Chandraker", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 41, "n_ref": 127, "n_ref_all": 147, "n_fig": 13, "n_tab": 11, "L_tab": 4693, "n_element_tab": 404, "n_fig_1": 10, "n_tab_1": 7, "L_tab_1": 2365, "n_element_tab_1": 168, "formula_len_all": 186, "formula_len_all_1": 87, "len_all": 151493, "len_all_1": 69907, "len_abs": 1270, "len_title": 143, "len_sents": 40263, "len_sents_1": 30630, "n_sents": 362, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1098, "title": "Progressive Token Length Scaling in Transformer Encoders for Efficient Universal Segmentation", "abs": "A powerful architecture for universal segmentation relies on transformers that encode multi-scale image features and decode object queries into mask predictions. With efficiency being a high priority for scaling such models, we observed that the state-of-the-art method Mask2Former uses \\~50% of its compute only on the transformer encoder. This is due to the retention of a full-length token-level representation of all backbone feature scales at each encoder layer. With this observation, we propose a strategy termed PROgressive Token Length SCALing for Efficient transformer encoders (PRO-SCALE) that can be plugged-in to the Mask2Former segmentation architecture to significantly reduce the computational cost. The underlying principle of PRO-SCALE is: progressively scale the length of the tokens with the layers of the encoder. This allows PRO-SCALE to reduce computations by a large margin with minimal sacrifice in performance (\\~52% GFLOPs reduction with no drop in performance on COCO dataset). We validate our framework on multiple public benchmarks. Our code will be publicly released.", "keywords": [ "universal segmentation; efficient transformers" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "dmjQLHufev", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Haizhao Fan", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Jiaxin Lu", "gender": "Female", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Tianzhe Wang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zetian Jiang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 51, "n_ref": 91, "n_ref_all": 107, "n_fig": 5, "n_tab": 7, "L_tab": 4218, "n_element_tab": 621, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3099, "n_element_tab_1": 488, "formula_len_all": 948, "formula_len_all_1": 942, "len_all": 153344, "len_all_1": 71431, "len_abs": 1027, "len_title": 131, "len_sents": 43755, "len_sents_1": 29911, "n_sents": 321, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 82, "L_abs": 1042, "title": "Learning Structured Universe Graph with Outlier OOD Detection for Partial Matching", "abs": "Partial matching is a kind of graph matching where only part of two graphs can be aligned. This problem is particularly important in computer vision applications, where challenges like point occlusion or annotation errors often occur when labeling key points. Previous work has often conflated point occlusion and annotation errors, despite their distinct underlying causes. We propose two components to address these challenges: (1) a structured universe graph is learned to connect two input graphs $X_{ij} = X_{iu} X_{ju}^\\top$, effectively resolving the issue of point occlusion; (2) an energy-based out-of-distribution detection is designed to remove annotation errors from the input graphs before matching. We evaluated our method on the Pascal VOC and Willow Object datasets, focusing on scenarios involving point occlusion and random outliers. The experimental results demonstrate that our approach consistently outperforms state-of-the-art methods across all tested scenarios, highlighting the accuracy and robustness of our method.", "keywords": [ "graph matching" ], "rating_list": [ 6, 5, 8, 6, 6 ], "soundness_list": [ 3, 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 2, 4, 3 ] }, { "paper_id": "dmh53n4onc", "primary_area": "applications to neuroscience & cognitive science", "n_author": 6, "author_info_list": [ { "name": "Abhishek Tyagi", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Andreas Bulling", "gender": "Male", "institution": "University of Stuttgart", "country": "DE", "position": "Full Professor" }, { "name": "Anna Penzkofer", "gender": "Female", "institution": "Universit\u00e4t Stuttgart", "country": "DE", "position": "PhD student" }, { "name": "Chris Eliasmith", "gender": "unknown", "institution": "University of Waterloo", "country": "", "position": "Full Professor" }, { "name": "Karim Habashy", "gender": "unknown", "institution": "University of Waterloo", "country": "CA", "position": "MS student" }, { "name": "Yuhao Zhu", "gender": "Not Specified", "institution": "University of Rochester", "country": "US", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 65, "n_ref": 126, "n_ref_all": 141, "n_fig": 6, "n_tab": 5, "L_tab": 1277, "n_element_tab": 153, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3697, "n_element_tab_1": 179, "formula_len_all": 162, "formula_len_all_1": 162, "len_all": 148178, "len_all_1": 62588, "len_abs": 1579, "len_title": 104, "len_sents": 35180, "len_sents_1": 27220, "n_sents": 220, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1356, "title": "SSPictR: Spatial Semantic Pointer Picture Representation", "abs": "The development of image representations that capture semantic and spatial information efficiently, which are also interpretable and generalisable, remains unsolved. Drawing from a cognitive modelling framework, we propose SSPictR \u2013 a biologically plausible image representation based on spatial semantic pointers (SSPs). SSPictR encodes semantic labels and their spatial locations extracted from segmentation maps and only requires a single vector to capture a fully decodable neuro-symbolic representation of a natural scene. It is inherently interpretable, offers a high compression factor and significantly faster inference speed on downstream tasks, such as scene recognition. We evaluate the efficiency and generalisability of SSPictR on the popular Places365, and ADE20K datasets for scene recognition, on COCOStuff for segmentation reconstruction, and on VISC and Savoias for prediction of visual complexity. We show that the scene representations provided by SSPictR are more generalisable within and across these tasks while only requiring a fraction of model parameters and, therefore, offer 25 times higher inference speed, with comparable accuracy. As such, SSPictR opens up a new direction for future research on cognitively-inspired image representations that are not only significantly smaller but also more interpretable and generalisable.", "keywords": [ "scene representation", "scene recognition", "cognitive maps", "spatial semantic pointers", "vector symbolic algebras" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 1, 4, 4 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "dmDEhEHxix", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chenwen Liao", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Lei Li", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Xuandong Zhao", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Yu-Xiang Wang", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 25, "n_ref": 67, "n_ref_all": 78, "n_fig": 3, "n_tab": 5, "L_tab": 722, "n_element_tab": 88, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 167, "n_element_tab_1": 7, "formula_len_all": 619, "formula_len_all_1": 454, "len_all": 116493, "len_all_1": 63585, "len_abs": 1091, "len_title": 114, "len_sents": 38251, "len_sents_1": 29999, "n_sents": 296, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1095, "title": "Efficiently Identifying Watermarked Segments in Mixed-Source Texts", "abs": "Text watermarks in large language models (LLMs) are increasingly used to detect synthetic text, mitigating misuse cases like fake news and academic dishonesty. While existing watermarking detection techniques primarily focus on classifying entire documents as watermarked or not, they often neglect the common scenario of identifying individual watermark segments within longer, mixed-source documents. Drawing inspiration from plagiarism detection systems, we propose two novel methods for partial watermark detection. First, we develop a geometry cover detection framework aimed at determining whether there is a watermark segment in long text. Second, we introduce an adaptive online learning algorithm to pinpoint the precise location of watermark segments within the text. Evaluated on three popular watermarking techniques (KGW-Watermark, Unigram-Watermark, and Gumbel-Watermark), our approach achieves high accuracy, significantly outperforming baseline methods. Moreover, our framework is adaptable to other watermarking techniques, offering new insights for precise watermark detection.", "keywords": [ "Watermark", "Online learning", "LLM" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "dmCGjPFVhF", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Firas Laakom", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" }, { "name": "J\u00fcrgen Schmidhuber", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Director, AI Initiative" }, { "name": "Li Nanbo", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" }, { "name": "Wenyi Wang", "gender": "unknown", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Yucheng XU", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "", "position": "MS student" } ], "n_formula": 26, "n_formula_1": 14, "n_ref_uni": 93, "n_ref": 215, "n_ref_all": 236, "n_fig": 5, "n_tab": 10, "L_tab": 2754, "n_element_tab": 267, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2157, "n_element_tab_1": 184, "formula_len_all": 3697, "formula_len_all_1": 1266, "len_all": 241131, "len_all_1": 72040, "len_abs": 1132, "len_title": 107, "len_sents": 60939, "len_sents_1": 33790, "n_sents": 371, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1154, "title": "FACTS: A Factored State-Space Framework for World Modelling", "abs": "World modelling is essential for understanding and predicting the dynamics of complex systems by learning both spatial and temporal dependencies. However, current frameworks, such as Transformers and selective state-space models like Mambas, exhibit limitations in efficiently encoding spatial and temporal structures, particularly in scenarios requiring long-term high-dimensional sequence modelling. To address these issues, we propose a novel recurrent framework, the FACTored State-space (FACTS) model, for spatio-temporal world modelling. The FACTS framework constructs a graph-structured memory with a routing mechanism that learns permutable memory representations, ensuring invariance to input permutations while adapting through selective state-space propagation. Furthermore, FACTS can be linearised to support parallel computation of high-dimensional sequences. We empirically evaluate FACTS across diverse tasks, including multivariate time series forecasting and object-centric world modelling, demonstrating that it consistently outperforms or matches specialised state-of-the-art models, despite its general-purpose world modelling design.", "keywords": [ "spatial-temporal modelling", "world modelling", "multivariate time-series forecasting", "object-centric representation learning", "unsupervised learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "dliIIodM6b", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Arunesh Sinha", "gender": "Male", "institution": "Rutgers University", "country": "", "position": "Assistant Professor" }, { "name": "Changyu Chen", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "PhD student" }, { "name": "Chao Du", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Min Lin", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Principal Researcher" }, { "name": "Pradeep Varakantham", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Full Professor" }, { "name": "Qian Liu", "gender": "Male", "institution": "Tiktok", "country": "SG", "position": "Researcher" }, { "name": "Tianyu Pang", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Zichen Liu", "gender": "unknown", "institution": "Sea AI Lab, Sea, Singapore", "country": "SG", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 34, "n_ref": 76, "n_ref_all": 124, "n_fig": 6, "n_tab": 9, "L_tab": 2215, "n_element_tab": 260, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 621, "n_element_tab_1": 94, "formula_len_all": 6024, "formula_len_all_1": 14295, "len_all": 169675, "len_all_1": 82473, "len_abs": 1137, "len_title": 103, "len_sents": 48453, "len_sents_1": 30865, "n_sents": 378, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1147, "title": "Bootstrapping Language Models with DPO Implicit Rewards", "abs": "Human alignment in large language models (LLMs) is an active area of research. A recent groundbreaking work, direct preference optimization (DPO), has greatly simplified the process from past work in reinforcement learning from human feedback (RLHF) by bypassing the reward learning stage in RLHF. DPO, after training, provides an implicit reward model. In this work, we make a novel observation that this implicit reward model can by itself be used in a bootstrapping fashion to further align the LLM. Our approach is to use the rewards from a current LLM model to construct a preference dataset, which is then used in subsequent DPO rounds. We incorporate two refinements to further improve our approach: 1) length-regularized reward shaping to make the preference dataset length-unbiased; 2) experience replay to enhance the quality of the preference dataset. Our approach, named self-alignment with DPO ImpliCit rEwards (DICE), shows great improvements in alignment. It achieves an increase of more than 8$\\\\%$ in length-controlled win rate on AlpacaEval 2 for all the different base models that we tried, without relying on external feedback.", "keywords": [ "Alignment", "Direct Preference Optimization", "Large Language Models" ], "rating_list": [ 8, 5, 8, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 4, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "dlUjNdybnq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Andis Draguns", "gender": "unknown", "institution": "Contramont Research", "country": "US", "position": "Principal Researcher" }, { "name": "Raymond Douglas", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "Undergrad student" }, { "name": "Tomas Gavenciak", "gender": "unknown", "institution": "Charles University Prague", "country": "CZ", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 16, "n_ref": 34, "n_ref_all": 61, "n_fig": 8, "n_tab": 3, "L_tab": 605, "n_element_tab": 85, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 954, "formula_len_all_1": 843, "len_all": 104624, "len_all_1": 61338, "len_abs": 980, "len_title": 125, "len_sents": 38836, "len_sents_1": 31045, "n_sents": 282, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 987, "title": "Mitigating the Influence of Distractor Tasks in LMs with Prior-Aware Decoding", "abs": "The broad capabilities of Language Models (LMs) can be limited by their sensitivity to distractor tasks: LMs can infer secondary tasks from the prompt in addition to the intended one, leading to unwanted outputs. For example, prompt injection attacks can cause models to deviate from explicit directives. In some \u2018inverse scaling\u2019 cases, this unwanted behaviour actually worsens as models scale up to at least 540B parameters. We present a theoretical framework that interprets LMs as a product of experts that combine multiple data generation processes. Based on this framework, we introduce prior-aware decoding (PAD) -- a simple contrastive inference method to reduce the influence of distractor tasks. We apply PAD to eleven models, across four datasets, and find improvements in 41 out of 44 task-model combinations, with a median increase in task completion proportion of 40%. The results suggest a promising direction for further development towards more reliable language models.", "keywords": [ "Inverse Scaling", "Contrastive Decoding" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "dl4nnpJssi", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Kuo Yang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Limin Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Qihe Huang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Wu Jiang", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Yang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zhengyang Zhou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 13, "n_ref_uni": 38, "n_ref": 83, "n_ref_all": 99, "n_fig": 6, "n_tab": 9, "L_tab": 6056, "n_element_tab": 113, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 811, "n_element_tab_1": 26, "formula_len_all": 1541, "formula_len_all_1": 630, "len_all": 183838, "len_all_1": 68459, "len_abs": 2010, "len_title": 120, "len_sents": 54035, "len_sents_1": 31161, "n_sents": 420, "n_sents_1": 274, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1671, "title": "Enhancing Graph Invariant Learning from a Negative Inference Perspective", "abs": "The out-of-distribution (OOD) generalization challenge is a longstanding problem in graph learning. Through studying the fundamental cause of data distribution shift, i.e., the changes of environments, significant progress has been achieved in addressing this issue. However, we observe that existing works still fail to effectively address complex environment shifts. Previous practices place excessive attention on extracting causal subgraphs, inevitably treating spurious subgraphs as environment variables. While spurious subgraphs are controlled by environments, the space of environment changes encompass more than the scale of spurious subgraphs. Therefore, existing efforts have a limited inference space for environments, leading to failure under severe environment changes. To tackle this issue, we propose a negative inference graph OOD framework (NeGo) to broaden the inference space for environment factors. Inspired by the successful practice of prompt learning in capturing underlying semantics and causal associations in large language models, we design a negative prompt environment inference to extract underlying environment information. We further introduce the environment-enhanced invariant subgraph learning method to effectively exploit inferred environment embedding, ensuring the robust extraction of causal subgraph in the environment shifts. Lastly, we conduct a comprehensive evaluation of NeGo on real-world datasets and synthetic datasets across domains. NeGo outperforms baselines on nearly all datasets, which verify the effectiveness of our framework. Our source code is available at \\url{https://anonymous.4open.science/r/NeGo-E4C1}.", "keywords": [ "Graph learning", "out-of-distribution generalization", "environment awareness", "negative inference", "prompt learning." ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "dkrEoT68by", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Joel Julin", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Koichiro Niinuma", "gender": "Male", "institution": "Fujitsu Research of America", "country": "", "position": "Research Director" }, { "name": "Laszlo Attila Jeni", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Liuyue Xie", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 17, "n_ref_uni": 19, "n_ref": 71, "n_ref_all": 98, "n_fig": 10, "n_tab": 6, "L_tab": 1514, "n_element_tab": 274, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 977, "n_element_tab_1": 174, "formula_len_all": 851, "formula_len_all_1": 752, "len_all": 140905, "len_all_1": 56419, "len_abs": 1295, "len_title": 79, "len_sents": 40372, "len_sents_1": 25575, "n_sents": 312, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 31, "L_abs": 1121, "title": "Gaussian Splatting Lucas-Kanade", "abs": "Gaussian Splatting and its dynamic extensions are effective for reconstructing 3D scenes from 2D images when there is significant camera movement to facilitate motion parallax and when scene objects remain relatively static. However, in many real-world scenarios, these conditions are not met. As a consequence, data-driven semantic and geometric priors have been favored as regularizers, despite their bias toward training data and their neglect of broader movement dynamics.\n\nDeparting from this practice, we propose a novel analytical approach that adapts the classical Lucas-Kanade method to dynamic Gaussian splatting. By leveraging the intrinsic properties of the forward warp field network, we derive an analytical velocity field that, through time integration, facilitates accurate scene flow computation. This enables the precise enforcement of motion constraints on warp fields, thus constraining both 2D motion and 3D positions of the Gaussians. Our method excels in reconstructing highly dynamic scenes with minimal camera movement, as demonstrated through experiments on both synthetic and real-world scenes.", "keywords": [ "Gaussian Splatting", "regularization", "novel view synthesis" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "dkoiAGjZV9", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Angela Yao", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" }, { "name": "Shihao Zhang", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yuguang Yan", "gender": "Male", "institution": "Guangdong University of Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 33, "n_formula_1": 20, "n_ref_uni": 21, "n_ref": 52, "n_ref_all": 74, "n_fig": 7, "n_tab": 5, "L_tab": 2205, "n_element_tab": 143, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1251, "n_element_tab_1": 78, "formula_len_all": 1716, "formula_len_all_1": 1078, "len_all": 107333, "len_all_1": 67505, "len_abs": 968, "len_title": 88, "len_sents": 32688, "len_sents_1": 28569, "n_sents": 304, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 999, "title": "Improving Deep Regression with Tightness", "abs": "For deep regression, preserving the ordinality of the targets with respect to the feature representation improves performance across various tasks. However, a theoretical explanation for the benefits of ordinality is still lacking. This work reveals that preserving ordinality reduces the conditional entropy $H(Z|Y)$ of representation $Z$ conditional on the target $Y$. However, our findings reveal that typical regression losses do little to reduce $H(Z|Y)$, even though it is vital for generalization performance. With this motivation, we introduce an optimal transport-based regularizer to preserve the similarity relationships of targets in the feature space to reduce $H(Z|Y)$. Additionally, we introduce a simple yet efficient strategy of duplicating the regressor targets, also with the aim of reducing $H(Z|Y)$. Experiments on three real-world regression tasks verify the effectiveness of our strategies to improve deep regression. Code: https://github.com/needylove/Regression_tightness", "keywords": [ "regression representation", "ordinality", "tightness", "depth estimation", "age estimation" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 1, 3 ] }, { "paper_id": "dj8CaE1G7m", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Min Chi", "gender": "Female", "institution": "North Carolina State University", "country": "", "position": "Associate Professor" }, { "name": "Seongsoo Kim", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 11, "n_ref_uni": 35, "n_ref": 75, "n_ref_all": 89, "n_fig": 5, "n_tab": 1, "L_tab": 825, "n_element_tab": 87, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 825, "n_element_tab_1": 87, "formula_len_all": 941, "formula_len_all_1": 520, "len_all": 174080, "len_all_1": 68441, "len_abs": 2664, "len_title": 142, "len_sents": 62198, "len_sents_1": 32550, "n_sents": 423, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1451, "title": "The Power of Data: How LSTMs Outshine Disease Progression Modeling with Two Simple Mechanisms", "abs": "Much of prior efforts have focused on Disease Progression Modeling (DPM) using Electronic Health Records (EHRs). EHRs, however, present significant challenges for deep learning models such as Long Short-Term Memory (LSTM), Variational Recurrent Neural Networks (VRNN), and Transformer due to the inherent complexities and variabilities within the data. Effectively addressing these variabilities is crucial for improving the performance and interpretability of such models. In this work, we propose two mechanisms to tackle key variabilities in EHR data: a \"bi-directional\" mechanism to account for the need to infer the underlying physical state in both forward and backward directions, and a \"time-aware\" mechanism to address irregular time intervals between consecutive events. We theoretically validate and empirically evaluate the impact of these two mechanisms across three state-of-the-art deep learning models in three distinct healthcare systems. Our results showed that the influence of the two mechanisms\u2014bidirectionality and time-awareness\u2014surpasses the differences between specific deep learning models. Across all three models, the performance hierarchy consistently follows: Bidirectional & Time-Aware > Time-Aware > Bidirectional > Original model, across all three healthcare systems. Notably, the Bidirectional Time-Aware LSTM matches or exceeds the performance of the corresponding VRNN and Transformer models in every system tested.", "keywords": [ "LSTM", "Septic Shock", "Disease Progression Modeling", "Time-Aware", "VRNN", "Transformer", "Bi-Directional" ], "rating_list": [ 3, 1, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 4, 1 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "dj0TktJcVI", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Jiancong Xiao", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Postdoc" }, { "name": "Li Shen", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Full Professor" }, { "name": "Ruochen Jin", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Intern" }, { "name": "Weijie J Su", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Associate Professor" }, { "name": "Bo-Jian Hou", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Postdoc" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 39, "n_ref": 74, "n_ref_all": 102, "n_fig": 10, "n_tab": 6, "L_tab": 912, "n_element_tab": 87, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 804, "n_element_tab_1": 85, "formula_len_all": 409, "formula_len_all_1": 430, "len_all": 135669, "len_all_1": 55952, "len_abs": 1638, "len_title": 135, "len_sents": 38472, "len_sents_1": 26132, "n_sents": 275, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1653, "title": "Fine-Tuning Attention Modules Only: Enhancing Weight Disentanglement in Task Arithmetic", "abs": "In recent years, *task arithmetic* has garnered increasing attention. This approach edits pre-trained models directly in weight space by combining the fine-tuned weights of various tasks into a *unified model*. Its efficiency and cost-effectiveness stem from its training-free combination, contrasting with traditional methods that require model training on large datasets for multiple tasks. However, applying such a unified model to individual tasks can lead to interference from other tasks (lack of *weight disentanglement*). To address this issue, Neural Tangent Kernel (NTK) linearization has been employed to leverage a ''kernel behavior'', facilitating weight disentanglement and mitigating adverse effects from unrelated tasks. Despite its benefits, NTK linearization presents drawbacks, including doubled training costs, as well as reduced performance of individual models. To tackle this problem, we propose a simple yet effective and efficient method that is to finetune the attention modules only in the Transformer. Our study reveals that the attention modules exhibit kernel behavior, and fine-tuning the attention modules only significantly improves weight disentanglement. To further understand how our method improves the weight disentanglement of task arithmetic, we present a comprehensive study of task arithmetic by differentiating the role of the representation module and task-specific module. In particular, we find that the representation module plays an important role in improving weight disentanglement whereas the task-specific modules such as the classification heads can degenerate the weight disentanglement performance.", "keywords": [ "task arithmetic", "weight disentanglement", "merging models", "multi-task learning" ], "rating_list": [ 8, 6, 3, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "din0lGfZFd", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Nikunj Saunshi", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Nishanth Dikkala", "gender": "Male", "institution": "Google", "country": "", "position": "Google Research" }, { "name": "Sanjiv Kumar", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Sashank J. Reddi", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Zhiyuan Li", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Assistant Professor" } ], "n_formula": 33, "n_formula_1": 9, "n_ref_uni": 48, "n_ref": 99, "n_ref_all": 126, "n_fig": 2, "n_tab": 6, "L_tab": 3045, "n_element_tab": 489, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4055, "formula_len_all_1": 954, "len_all": 193047, "len_all_1": 62377, "len_abs": 1851, "len_title": 89, "len_sents": 52563, "len_sents_1": 29786, "n_sents": 471, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 1877, "title": "Understanding Reasoning with Looped Transformers", "abs": "Large language models have shown promising abilities in reasoning problems and scaling laws suggest that parameter count is a key driver. Recent works (Chen & Zou, 2024; Ye et al., 2024) argue that for reasoning, depth plays a very important role in addition to parameter count. In this work, we make a more fine-grained claim \u2014 many reasoning problems require large depth but not necessarily many parameters, in the sense that they can be solved via looped models. This unlocks a novel application of looped models for reasoning. We empirically study various synthetic reasoning problems like addition, variable assignment and math problems. For each of these, we find that $k$-layer transformer model looped $L$ times nearly matches the quality of a $kL$-layer non-looped model and is much better than a k-layer model. Thus, using a small model and providing depth via looping can suffice for such reasoning problems. We then show theoretical results proving that many such reasoning problems can be solved via iterative algorithms, and thus, can be solved with looped models. Motivated by these findings, we train autoregressive models on general language modeling datasets with looping and compare a $k$-layer model looped $L$ times to a $kL$-layer model. While the looped model is understandably worse on perplexity and memorization tasks, it surprisingly does very well on tasks that require reasoning, like open book QA, math word problems and reasoning primitives. Despite having significantly fewer parameters, it can even match or outperform the non-looped $kL$-layer model on some of these tasks. These results suggest a novel inductive bias of looped models towards enhanced reasoning. We provide further evidence for this inductive bias by visualizing perplexity vs downstream isoplots, and design a looping-inspired regularization that solidifies this hypothesis.", "keywords": [ "looped models", "reasoning", "language model", "iterative algorithm", "inductive bias" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "dieIcwiXCL", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Hao Zhao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Shiran Yuan", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 39, "n_ref": 64, "n_ref_all": 84, "n_fig": 10, "n_tab": 7, "L_tab": 967, "n_element_tab": 96, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 768, "n_element_tab_1": 71, "formula_len_all": 621, "formula_len_all_1": 620, "len_all": 119618, "len_all_1": 60705, "len_abs": 1490, "len_title": 112, "len_sents": 32693, "len_sents_1": 28117, "n_sents": 239, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1498, "title": "Hybrid Spatial Representations for Species Distribution Modeling", "abs": "We address an important problem in ecology called Species Distribution Modeling (SDM), whose goal is to predict whether a species exists at a certain position on Earth. In particular, we tackle a challenging version of this task, where we learn from presence-only data in a community-sourced dataset, model a large number of species simultaneously, and do not use any additional environmental information. Previous work has used neural implicit representations to construct models that achieve promising results. However, implicit representations often generate predictions of limited spatial precision. We attribute this limitation to their inherently global formulation and inability to effectively capture local feature variations. This issue is especially pronounced with presence-only data and a large number of species. To address this, we propose a hybrid embedding scheme that combines both implicit and explicit embeddings. Specifically, the explicit embedding is implemented with a multiresolution hashgrid, enabling our models to better capture local information. Experiments demonstrate that our results exceed other works by a large margin on various standard benchmarks, and that the hybrid representation is better than both purely implicit and explicit ones. Qualitative visualizations and comprehensive ablation studies reveal that our hybrid representation successfully addresses the two main challenges. Our code is open-sourced at https://anonymous.4open.science/r/HSR-SDM-7360.", "keywords": [ "Neural Implicit Representations", "Explicit Representations", "Multiresolution Hashgrids", "Geographic Embedding", "AI for Science" ], "rating_list": [ 8, 3, 3, 6 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "diXvBHiRyE", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Boxi Cao", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Hongyu Lin", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Jiasheng Zheng", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Le Sun", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Ruotong Pan", "gender": "Female", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Xianpei Han", "gender": "Male", "institution": "Institute of Software, CAS", "country": "CN", "position": "Professor" }, { "name": "Yaojie Lu", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Zhengzhao Ma", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 44, "n_ref": 99, "n_ref_all": 120, "n_fig": 10, "n_tab": 7, "L_tab": 9383, "n_element_tab": 1989, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2359, "n_element_tab_1": 567, "formula_len_all": 293, "formula_len_all_1": 292, "len_all": 214805, "len_all_1": 74104, "len_abs": 1805, "len_title": 141, "len_sents": 54961, "len_sents_1": 32062, "n_sents": 395, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1810, "title": "Beyond Correctness: Benchmarking Multi-dimensional Code Generation for Large Language Models", "abs": "In recent years, researchers have proposed numerous benchmarks to evaluate the impressive coding capabilities of large language models (LLMs). However, current benchmarks primarily assess the accuracy of LLM-generated code, while neglecting other critical dimensions that also significantly impact code quality in real-world development. Moreover, relying exclusively on correctness as the guiding metric renders LLMs vulnerable to data contamination. Therefore, this paper proposes the **RACE** benchmark, which comprehensively evaluates the quality of code generated by LLMs across 4 dimensions: Readability, mAintainability, Correctness, and Efficiency. Specifically, considering the demand-dependent nature of dimensions beyond correctness, we design various types of user requirements for each dimension to assess the model's ability to generate correct code that also meets user demands. We analyze 28 representative LLMs based on RACE and find that: 1) current correctness-centric benchmarks fail to capture the multifaceted requirements of code in real-world scenarios, while RACE provides a comprehensive evaluation that reveals the defects of LLMs across multiple dimensions; 2) the RACE benchmark serves as an effective tool for resisting the risk of data contamination; 3) even the most advanced code LLMs still encounter significant challenges in customized requirements involving complex instructions; 4) most LLMs exhibit an inherent preference for specific coding style. These findings highlight the need for a multidimensional evaluation of code LLMs, emphasizing metrics beyond correctness for real-world applications. Future efforts should aim to develop novel learning algorithms to enhance code generation under varied constraints and improve coverage and usability for diverse user needs.", "keywords": [ "Code Generation", "Multidimension", "Benchmark", "LLM Evaluation" ], "rating_list": [ 3, 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 1, 1, 1, 3, 2 ], "confidence_list": [ 5, 5, 4, 3, 5 ] }, { "paper_id": "diKykN0Yaa", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Guillaume Lajoie", "gender": "Male", "institution": "Mila, Quebec AI institute", "country": "", "position": "Core Academic Member" }, { "name": "Marco Bonizzato", "gender": "Male", "institution": "\u00c9cole Polytechnique de Montr\u00e9al, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Assistant Professor" }, { "name": "Maxime Moutet", "gender": "Male", "institution": "Ecole Nationale de la Statistique et de l'Administration Economique", "country": "FR", "position": "MS student" }, { "name": "Rose Guay-Hottin", "gender": "Female", "institution": "\u00c9cole Polytechnique de Montr\u00e9al, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "MS student" }, { "name": "Thomas Garbay", "gender": "unknown", "institution": "\u00c9cole Polytechnique de Montr\u00e9al, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 39, "n_ref": 53, "n_ref_all": 69, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 501, "formula_len_all_1": 527, "len_all": 110619, "len_all_1": 63220, "len_abs": 1315, "len_title": 140, "len_sents": 33439, "len_sents_1": 29440, "n_sents": 286, "n_sents_1": 247, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1331, "title": "Memory-Pruning Algorithm for Bayesian Optimization with Strict Computational Cost Guarantees", "abs": "Bayesian Optimization (BO) is a powerful tool for optimizing noisy and expensive-to-evaluate black-box functions, widely used in fields such as machine learning and various branches of engineering. However, BO faces significant challenges when applied to large datasets or when it requires numerous optimization iterations. The computational and memory demands of updating Gaussian Process (GP) models can result in unmanageable computation times. To address these limitations, we propose a new Bayesian Optimization algorithm with memory pruning (MP-BO), which restricts the maximum training data size by acquiring new queries while concurrently removing data points from the training set. This approach guarantees a maximum algorithmic complexity of $\\bigO(m^3)$, where $m \\ll n$ is a fixed value and $n$ represent the size of the full training set. The pruning strategy ensures reduced and constant memory usage and computation time, without significantly degrading performance. We evaluate MP-BO on synthetic benchmarks and a real neurostimulation dataset, demonstrating its robustness and efficiency in scenarios where traditional BO would fail under strict computational constraints. Our results suggest that MP-BO is a promising solution for applications that require efficient optimization with limited computing resources.", "keywords": [ "Bayesian Optimization", "Gaussian Processes", "Embedded Systems", "Random Eviction", "Computational Constraints" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "diKRhKs5yl", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Haoxuan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Jiarui Jin", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Researcher" }, { "name": "Jun Wang", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Professor" }, { "name": "Mengyue Yang", "gender": "Female", "institution": "University of Bristol", "country": "GB", "position": "Assistant Professor" }, { "name": "Weinan Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaoting He", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Yiming Yang", "gender": "Female", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Yong Yu", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Yuwei Wu", "gender": "Male", "institution": "AI Fund", "country": "US", "position": "Researcher" } ], "n_formula": 47, "n_formula_1": 12, "n_ref_uni": 58, "n_ref": 100, "n_ref_all": 149, "n_fig": 8, "n_tab": 19, "L_tab": 22707, "n_element_tab": 2231, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 11905, "n_element_tab_1": 1144, "formula_len_all": 3716, "formula_len_all_1": 812, "len_all": 300877, "len_all_1": 88865, "len_abs": 1886, "len_title": 116, "len_sents": 86722, "len_sents_1": 29677, "n_sents": 672, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1875, "title": "Large Language Models are Demonstration Pre-Selectors for Themselves", "abs": "In-context learning with large language models (LLMs) delivers strong few-shot performance by choosing few-shot demonstrations from the entire training dataset. However, previous few-shot in-context learning methods, which calculate similarity scores for choosing demonstrations, incur high computational costs by repeatedly retrieving large-scale datasets for each query. This is due to their failure to recognize that not all demonstrations are equally informative, and many less informative demonstrations can be inferred from a core set of highly informative ones. To this end, we propose FEEDER (FEw yet Essential Demonstration prE-selectoR), a novel \\emph{pre-selection} framework that identifies a core subset of demonstrations containing the most informative examples. This subset, referred to as the FEEDER set, consists of demonstrations that capture both the ''sufficiency'' and ''necessity'' information to infer the entire dataset. Notice that FEEDER is selected before the few-shot in-context learning, enabling more efficient few-shot demonstrations choosing in a smaller set. To identify FEEDER, we propose a novel effective tree based algorithm. Once selected, it can replace the original dataset, leading to improved efficiency and prediction accuracy in few-shot in-context learning. Additionally, FEEDER also benefit fine-tuning LLMs, we propose a bi-level optimization method enabling more efficient training without sacrificing performance when datasets become smaller. \nOur experiments are on 6 text classification datasets, 1 reasoning dataset, and 1 semantic-parsing dataset, across 6 LLMs (ranging from 335M to 7B parameters), demonstrate that: (i) In few-shot inference, FEEDER achieves superior (or comparable) performance while utilizing only half the input training data. (ii) In fine-tuning, FEEDER significantly boosts the performance of LLMs.", "keywords": [ "Large Language Model", "Demonstration Pre-Selection", "In-Context Learning" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "dhuQJseaBA", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Leonid Karlinsky", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Muhammad Jehanzeb Mirza", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Raja Giryes", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Associate Professor" }, { "name": "Rogerio Feris", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Research Manager" }, { "name": "Sepp Hochreiter", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "Full Professor" }, { "name": "Sivan Doveh", "gender": "unknown", "institution": "Weizmann Institute of Science", "country": "IL", "position": "PhD student" }, { "name": "Wei Lin", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 38, "n_ref": 147, "n_ref_all": 208, "n_fig": 25, "n_tab": 13, "L_tab": 2668, "n_element_tab": 439, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2538, "n_element_tab_1": 234, "formula_len_all": 248, "formula_len_all_1": 245, "len_all": 265108, "len_all_1": 81278, "len_abs": 3302, "len_title": 84, "len_sents": 78876, "len_sents_1": 33737, "n_sents": 628, "n_sents_1": 195, "venue": "ICLR2025", "decision": "withdraw", "L_title": 36, "L_abs": 1233, "title": "Comparison Visual Instruction Tuning", "abs": "Comparing two images in terms of Commonalities and Differences (CaD) is a fundamental human capability that forms the basis of advanced visual reasoning and interpretation. It is essential for the generation of detailed and contextually relevant descriptions, performing comparative analysis, novelty detection, and making informed decisions based on visual data. However, surprisingly, little attention has been given to these fundamental concepts in the best current mimic of human visual intelligence - Large Multimodal Models (LMMs). We develop and contribute a new two-phase approach CaD-VI for collecting synthetic visual instructions, together with an instruction-following dataset CaD-Inst containing 349K image pairs with CaD instructions collected using CaD-VI. Our approach significantly improves the CaD spotting capabilities in LMMs, advancing the SOTA on a diverse set of related tasks by up to 17.5%. It is also complementary to existing difference-only instruction datasets, allowing automatic targeted refinement of those resources increasing their effectiveness for CaD tuning by up to 10%. Additionally, we propose an evaluation benchmark with 7.5K open-ended QAs to assess the CaD understanding abilities of LMMs.", "keywords": [ "Large Multimodal Models", "visual instruction tuning", "commonalities and differences" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "dhoCfPPjeZ", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Cheng Perng Phoo", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Kavita Bala", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Professor" }, { "name": "Mia Chiquier", "gender": "Female", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Utkarsh Mall", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Postdoc" }, { "name": "Bharath Hariharan", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 42, "n_ref": 55, "n_ref_all": 68, "n_fig": 4, "n_tab": 6, "L_tab": 918, "n_element_tab": 146, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1151, "n_element_tab_1": 158, "formula_len_all": 928, "formula_len_all_1": 450, "len_all": 140466, "len_all_1": 67527, "len_abs": 1607, "len_title": 114, "len_sents": 41796, "len_sents_1": 30344, "n_sents": 356, "n_sents_1": 240, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1042, "title": "DiSciPLE: Learning Interpretable Programs for Scientific Discovery", "abs": "Creating hypotheses for new observations is a key step in the scientific process of understanding a problem in any domain. A good hypothesis that is interpretable, reliable (good at predicting unseen observations), and data-efficient; is useful for scientists aiming to make novel discoveries. This paper introduces an automatic way of learning such interpretable and reliable hypotheses in a data-efficient manner. We propose DiSciPLE (Discovering Scientific Programs using LLMs and Evolution), an evolutionary algorithm that leverages common sense and prior knowledge of large language models (LLMs) to create hypotheses as Python programs. Additionally, we propose two improvements: a program critic and a program simplifier to further improve our method to produce good hypotheses. We evaluate our method on four different real-world tasks in two scientific domains and show significantly better results. For example, we can learn programs with 35% lower error than the closest non-interpretable baseline for population density estimation", "keywords": [ "AI for Science", "Large Language Model", "Evolutionary Algorithm", "Interpretabilty" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "dhAL5fy8wS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Furu Wei", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Partner Research Manager" }, { "name": "Hongning Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Li Dong", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Principal Researcher" }, { "name": "Minlie Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Qingxiu Dong", "gender": "Female", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yaru Hao", "gender": "Female", "institution": "Microsoft Research Asia", "country": "", "position": "Researcher" }, { "name": "Yuxian Gu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 69, "n_formula_1": 16, "n_ref_uni": 61, "n_ref": 142, "n_ref_all": 254, "n_fig": 10, "n_tab": 7, "L_tab": 2116, "n_element_tab": 312, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1762, "n_element_tab_1": 233, "formula_len_all": 5099, "formula_len_all_1": 1278, "len_all": 230622, "len_all_1": 71154, "len_abs": 1290, "len_title": 102, "len_sents": 53745, "len_sents_1": 28067, "n_sents": 435, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 54, "L_abs": 1226, "title": "Data Selection via Optimal Control for Language Models", "abs": "This work investigates the selection of high-quality pre-training data from massive corpora to enhance LMs' capabilities for downstream usage. \nWe formulate data selection as a generalized Optimal Control problem, which can be solved theoretically by Pontryagin's Maximum Principle (PMP), yielding a set of necessary conditions that characterize the relationship between optimal data selection and LM training dynamics.\nBased on these theoretical results, we introduce PMP-based Data Selection (PDS), a framework that approximates optimal data selection by solving the PMP conditions. \nIn our experiments, we adopt PDS to select data from CommmonCrawl and show that the PDS-selected corpus accelerates the learning of LMs and constantly boosts their performance on a wide range of downstream tasks across various model sizes.\nMoreover, the benefits of PDS extend to ~400B models trained on ~10T tokens, as evidenced by the extrapolation of the test loss curves according to the Scaling Laws.\nPDS also improves data utilization when the pre-training data is limited, by reducing the data demand by 1.8 times, which mitigates the quick exhaustion of available web-crawled corpora. We will open-source our code, models, and data.", "keywords": [ "Pre-training Language Models", "Data Selection", "Optimal Control" ], "rating_list": [ 10, 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 4, 4 ], "contribution_list": [ 3, 3, 4, 3, 3 ], "confidence_list": [ 2, 3, 3, 3, 2 ] }, { "paper_id": "dh78yRFVK9", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Amartya Sanyal", "gender": "Male", "institution": "Copenhagen University", "country": "DK", "position": "Assistant Professor" }, { "name": "Sadhika Malladi", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Sanjeev Arora", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "Full Professor" }, { "name": "Stanley Wei", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" } ], "n_formula": 123, "n_formula_1": 17, "n_ref_uni": 63, "n_ref": 114, "n_ref_all": 122, "n_fig": 0, "n_tab": 5, "L_tab": 657, "n_element_tab": 103, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 317, "n_element_tab_1": 18, "formula_len_all": 9047, "formula_len_all_1": 1120, "len_all": 248357, "len_all_1": 74486, "len_abs": 1274, "len_title": 106, "len_sents": 56278, "len_sents_1": 32890, "n_sents": 585, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1281, "title": "Provable unlearning in topic modeling and downstream tasks", "abs": "Machine unlearning algorithms are increasingly important as legal concerns arise around the provenance of training data, but verifying the success of unlearning is often difficult. Provable guarantees for unlearning are often limited to supervised learning settings. In this paper, we provide the first theoretical guarantees for unlearning in the pre-training and fine-tuning paradigm by studying topic models, simple bag-of-words language models that can be adapted to solve downstream tasks like retrieval and classification. First, we design a provably effective unlearning algorithm for topic models that incurs a computational overhead independent of the size of the original dataset. Our analysis additionally quantifies the deletion capacity of the model -- i.e., the number of examples that can be unlearned without incurring a significant cost in model performance. Finally, we formally extend our analyses to account for adaptation to a given downstream task. In particular, we design an efficient algorithm to perform unlearning after fine-tuning the topic model via a linear head. Notably, we show that it is easier to unlearn pre-training data from models that have been fine-tuned to a particular task, and one can unlearn this data without modifying the base model.", "keywords": [ "Machine unlearning", "topic models", "language models" ], "rating_list": [ 3, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 1, 3, 3, 3 ], "confidence_list": [ 4, 2, 3, 4 ] }, { "paper_id": "dh4t9qmcvK", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Edoardo Cetin", "gender": "unknown", "institution": "Sakana AI", "country": "JP", "position": "Researcher" }, { "name": "Qi Sun", "gender": "Male", "institution": "tokyo institute of technology", "country": "JP", "position": "MS student" }, { "name": "Yujin Tang", "gender": "Male", "institution": "Sakana AI", "country": "JP", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 36, "n_ref": 47, "n_ref_all": 68, "n_fig": 8, "n_tab": 8, "L_tab": 3600, "n_element_tab": 334, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 531, "n_element_tab_1": 49, "formula_len_all": 25874, "formula_len_all_1": 26502, "len_all": 182397, "len_all_1": 103490, "len_abs": 1198, "len_title": 82, "len_sents": 50370, "len_sents_1": 36542, "n_sents": 327, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 39, "L_abs": 1172, "title": "Transformer-Squared: Self-adaptive LLMs", "abs": "Self-adaptive large language models (LLMs) aim to solve the challenges posed by traditional fine-tuning methods, which are often computationally intensive and static in their ability to handle diverse tasks. We introduce Transformer-Squared, a novel self-adaptation framework that adapts LLMs for unseen tasks in real-time by selectively adjusting only the singular components of their weight matrices. During inference, Transformer-Squared employs a two-pass mechanism: first, a dispatch system identifies the task properties, and then task-specific 'expert' vectors, trained using reinforcement learning, are dynamically mixed to obtain targeted behavior for the incoming prompt. Our method consistently outperforms ubiquitous approaches such as LoRA, with fewer parameters and greater efficiency. Furthermore, Transformer-Squared demonstrates versatility across different LLM architectures and modalities, including vision-language tasks. Transformer-Squared represents a significant leap forward, offering a scalable, efficient solution for enhancing the adaptability and task-specific performance of LLMs, paving the way for truly dynamic, self-organizing AI systems.", "keywords": [ "Transformers", "Self-Adaptation", "Fine-tuning", "Dynamic System", "LLMs" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "dggRphAcCj", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Chenming Shang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Hengyuan Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Lei Ke", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Shiji Zhou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Xinchen Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yujiu Yang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yuwang Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 16, "n_ref_uni": 38, "n_ref": 48, "n_ref_all": 59, "n_fig": 8, "n_tab": 8, "L_tab": 9949, "n_element_tab": 778, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1701, "n_element_tab_1": 67, "formula_len_all": 2569, "formula_len_all_1": 1427, "len_all": 141914, "len_all_1": 58734, "len_abs": 1584, "len_title": 143, "len_sents": 36287, "len_sents_1": 27005, "n_sents": 276, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1590, "title": "GeoCon: Compositional Generalization Through Geometric Constraints on Representation Structure", "abs": "Compositional generalization, referring to the capacity to generalize novel combinations of fundamental and essential concepts, is thought to be the mechanism underlying a human\u2019s remarkable ability of rapid generalization to new knowledge and tasks. Recent research on brain neural activation space has found that the geometric structure of neural representations is highly related to human compositional generalization capability.\nIn this paper, we extend the above observations from neuroscience to deep neural networks to validate the potential relationship between the geometric structure of representations and compositional generalization capability. In particular, we first construct a new compositional generalization benchmark from the existent datasets, which aims to discriminate multiple concepts simultaneously through a powerful representation. Meanwhile, for the aforementioned geometric constraint, the parallelism score is formally defined for deep neural networks.\nSubsequently, we decompose the deep neural network into two parts: the featurizer and the classifier, to investigate the relationship between compositional generalization capability and parallelism score separately. Our proposed method, Geometric Constraint (GeoCon), involves distance variance minimization on the classifier and parallelism score maximization on the featurizer.\nExperiments on synthetic and real-world datasets demonstrate significant improvement of our approach, verifying the effectiveness of our neuroscience-inspired GeoCon approach towards human-like superior generalization ability.", "keywords": [ "Represention Learning", "Compositional Generalization." ], "rating_list": [ 8, 5, 6 ], "soundness_list": [ 4, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "dgb4rfPzaw", "primary_area": "applications to robotics, autonomy, planning", "n_author": 16, "author_info_list": [ { "name": "Alex Zihao Zhu", "gender": "unknown", "institution": "Waymo", "country": "", "position": "Researcher" }, { "name": "Anant Subramanian", "gender": "Male", "institution": "Google", "country": "US", "position": "Waymo" }, { "name": "Cheolho Park", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Dragomir Anguelov", "gender": "Male", "institution": "Waymo", "country": "US", "position": "Researcher" }, { "name": "Hao Xiang", "gender": "Male", "institution": "Waymo", "country": "US", "position": "Researcher" }, { "name": "Junwen Yao", "gender": "Male", "institution": "Waymo", "country": "US", "position": "Software Engineer" }, { "name": "Kan Chen", "gender": "Male", "institution": "Waymo", "country": "", "position": "Research Scientist" }, { "name": "Mengtian Li", "gender": "unknown", "institution": "Waymo LLC", "country": "", "position": "Software Engineer" }, { "name": "Mingxing Tan", "gender": "Male", "institution": "Google/Waymo", "country": "US", "position": "Researcher" }, { "name": "Runsheng Xu", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Tong He", "gender": "Male", "institution": "Waymo Research", "country": "US", "position": "Researcher" }, { "name": "Xia Chen", "gender": "Male", "institution": "Waymo", "country": "US", "position": "Researcher" }, { "name": "Yan Xu", "gender": "unknown", "institution": "Waymo", "country": "", "position": "Software Engineer" }, { "name": "Yanhui Liang", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Yingwei Li", "gender": "Male", "institution": "Waymo LLC", "country": "US", "position": "Researcher" }, { "name": "Zhaoqi Leng", "gender": "Male", "institution": "Waymo LLC", "country": "US", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 51, "n_ref": 68, "n_ref_all": 84, "n_fig": 3, "n_tab": 11, "L_tab": 1852, "n_element_tab": 192, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1209, "n_element_tab_1": 122, "formula_len_all": 398, "formula_len_all_1": 398, "len_all": 134054, "len_all_1": 62039, "len_abs": 1576, "len_title": 105, "len_sents": 35758, "len_sents_1": 29807, "n_sents": 229, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1583, "title": "World-simulation as pre-training for scalable perception", "abs": "Image-based autoregressive next-token prediction offers a promising avenue for developing world video simulators for autonomous driving. However, applications of these autoregressive models for common perception tasks such as geometric and semantic understanding remains under-explored, largely due to the difficulty of applying discrete token modeling to perception tasks. In this paper, we introduce PerceptionLM, an end-to-end framework that leverages autoregressive world simulators to effectively improve Perception tasks. It consists of a token-based pretraining stage and a novel fine-tuning stage that adapts discrete tokens to continuous embeddings for perception tasks. During pretraining, we leverage the world knowledge from Segment Anything and Depth Anything through autoregressive next-token prediction to imbue the model with world knowledge from multiple vision modalities. During fine-tuning, we propose a novel decoder adaptor to fuse discrete tokens with continuous embeddings from image encoders, which overcomes the limitations of discrete tokens. With PerceptionLM, we observe impressive scaling properties, where quality is consistently improved when providing more training compute or longer temporal context. On multiple public benchmarks including nuScenes, nuImages, Waymo Open Dataset, and Waymo Open Motion Dataset, PerceptionLM demonstrates significant performance improvements for common perception tasks such as depth estimation and semantic segmentation, highlighting its potential for scaling vision-only foundation models for autonomous driving.", "keywords": [ "autonomous driving; computer vision; autoregressive transformer; self-supervised learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "dgR6i4TSng", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Francesco Tonin", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Postdoc" }, { "name": "Leyla Naz Candogan", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Toshiaki Koike-Akino", "gender": "Male", "institution": "Mitsubishi Electric Research Labs", "country": "US", "position": "Principal Researcher" }, { "name": "Volkan Cevher", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" }, { "name": "Yongtao Wu", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Zhengqing Wu", "gender": "Male", "institution": "EPFL", "country": "CH", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 16, "n_ref_uni": 64, "n_ref": 134, "n_ref_all": 167, "n_fig": 9, "n_tab": 12, "L_tab": 2381, "n_element_tab": 296, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1388, "n_element_tab_1": 119, "formula_len_all": 2246, "formula_len_all_1": 800, "len_all": 190705, "len_all_1": 68212, "len_abs": 827, "len_title": 99, "len_sents": 49643, "len_sents_1": 29624, "n_sents": 382, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 832, "title": "Quantum-PEFT: Ultra parameter-efficient fine-tuning", "abs": "This paper introduces Quantum-PEFT that leverages quantum computations for parameter-efficient fine-tuning (PEFT). Unlike other additive PEFT methods, such as low-rank adaptation (LoRA), Quantum-PEFT exploits an underlying full-rank yet surprisingly parameter efficient _quantum unitary parameterization_. With the use of Pauli parameterization, the number of trainable parameters grows only logarithmically with the ambient dimension, as opposed to linearly as in LoRA-based PEFT methods. Quantum-PEFT achieves vanishingly smaller number of trainable parameters than the lowest-rank LoRA as dimensions grow, enhancing parameter efficiency while maintaining a competitive performance. We apply Quantum-PEFT to several transfer learning benchmarks in language and vision, demonstrating significant advantages in parameter efficiency.", "keywords": [ "parameter-efficient fine-tuning", "lora", "quantum machine learning", "orthogonality constraints" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "dfdrb2sZUw", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Muberra Ozmen", "gender": "Female", "institution": "Block", "country": "CA", "position": "Researcher" }, { "name": "Thomas Markovich", "gender": "unknown", "institution": "Block", "country": "US", "position": "Principal Researcher" }, { "name": "Zahra Fazel", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CA", "position": "AI Engineer" }, { "name": "florence regol", "gender": "Female", "institution": "McGill University", "country": "", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 28, "n_ref_uni": 57, "n_ref": 97, "n_ref_all": 108, "n_fig": 3, "n_tab": 6, "L_tab": 1422, "n_element_tab": 98, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1075, "n_element_tab_1": 73, "formula_len_all": 1770, "formula_len_all_1": 1247, "len_all": 148169, "len_all_1": 70813, "len_abs": 1255, "len_title": 96, "len_sents": 36377, "len_sents_1": 32175, "n_sents": 276, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1280, "title": "Risk Aware Negative Sampling in Link Prediction", "abs": "It is commonly believed that Message Passing Neural Networks (MPNNs) struggle in link prediction settings due to limitations in their expressive power. Recent work has focused on developing more expressive model classes, which are capable of learning link representations through techniques such as labeling tricks, the inclusion of structural features, or the use of subgraph methods. These approaches have yielded significant performance improvements across a range of benchmark datasets. However, an interesting question remains: have we fully wrung out the performance by optimizing the other aspects of the training process? In this work, we present results that indicate that significant amounts of model performance have been left on the table by the use of easy negative-samples during training. We theoretically explore the generalization gap and excess risk to quantify the performance loss caused by easy negatives. Motivated by this analysis, we introduce Risk Aware Negative Sampling in Link Prediction (RANS), which efficiently performs dynamic hard-negative-mining. Empirical results show that a simple GCN augmented by RANS realizes between 20\\% and 50\\% improvements in predictive accuracy when compared with the same model trained with standard negative samples.", "keywords": [ "graph neural networks", "link prediction", "negative sampling" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "dfC2ji6nek", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "DongbinZhang", "gender": "Male", "institution": "Tsinghua Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Haoqian Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Kangjie Chen", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Lijian Lin", "gender": "Female", "institution": "International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Minghan Qin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Ye Zhu", "gender": "Male", "institution": "Tencent PCG ARC Lab", "country": "CN", "position": "Researcher" }, { "name": "Yu Li", "gender": "Male", "institution": "International Digital Economy Academy", "country": "CN", "position": "Principal Researcher" }, { "name": "Yunfei Liu", "gender": "Male", "institution": "International Digital Economy Academy (IDEA)", "country": "CN", "position": "Researcher" } ], "n_formula": 20, "n_formula_1": 18, "n_ref_uni": 45, "n_ref": 81, "n_ref_all": 97, "n_fig": 10, "n_tab": 4, "L_tab": 4775, "n_element_tab": 504, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 804, "n_element_tab_1": 91, "formula_len_all": 1164, "formula_len_all_1": 926, "len_all": 159100, "len_all_1": 57237, "len_abs": 266, "len_title": 135, "len_sents": 34534, "len_sents_1": 24511, "n_sents": 308, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1233, "title": "SHARP: Splatting High-fidelity And Relightable Photorealistic 3D Gaussian Head Avatars", "abs": "Reconstructing animatable and high-fidelity 3D head avatars from monocular videos, especially with realistic relighting, is a valuable task. However, the limited information from single-view input, combined with the complex head poses and facial movements, makes this challenging. Previous methods achieve real-time performance by combining 3D Gaussian Splatting with a parametric head model, but the resulting head quality suffers from inaccurate face tracking and limited expressiveness of the deformation model. These methods also fail to produce realistic effects under novel lighting conditions. To address these issues, we propose SHARP, a method that reconstructs high-fidelity, relightable 3D head avatars using 3D Gaussian points. SHARP reduces tracking errors through end-to-end optimization and better captures individual facial deformations using learnable blendshapes and linear blend skinning. Additionally, it decomposes head appearance into several physical properties and incorporates physically-based shading to account for environmental lighting. Extensive experiments demonstrate that SHARP not only reconstructs superior-quality heads but also achieves realistic visual effects under varying lighting conditions.", "keywords": [ "3D Head avatar; monocular video; 3D Gaussian Splatting" ], "rating_list": [ 3, 5, 3, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 1, 3, 2, 2 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "deWPVVa6TE", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Jian Cao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Jianqing Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Qiang Yang", "gender": "Male", "institution": "Webank", "country": "CN", "position": "Principal Researcher" }, { "name": "Yang Hua", "gender": "Male", "institution": "Queen's University Belfast", "country": "GB", "position": "Assistant Professor" }, { "name": "Yang Liu", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 29, "n_formula_1": 17, "n_ref_uni": 35, "n_ref": 111, "n_ref_all": 134, "n_fig": 4, "n_tab": 5, "L_tab": 2572, "n_element_tab": 144, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2057, "n_element_tab_1": 124, "formula_len_all": 4152, "formula_len_all_1": 1484, "len_all": 165471, "len_all_1": 76111, "len_abs": 1103, "len_title": 125, "len_sents": 42022, "len_sents_1": 30966, "n_sents": 344, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1114, "title": "FedL2G: Learning to Guide Local Training in Heterogeneous Federated Learning", "abs": "Data and model heterogeneity are two core issues in Heterogeneous Federated Learning (HtFL). In scenarios with heterogeneous model architectures, aggregating model parameters becomes infeasible, leading to the use of prototypes (i.e., class representative feature vectors) for aggregation and guidance. However, they still experience a mismatch between the extra guiding objective and the client's original local objective when aligned with global prototypes. Thus, we propose a Federated Learning-to-Guide (FedL2G) method that adaptively learns to guide local training in a federated manner and ensures the extra guidance is beneficial to clients\u2019 original tasks. With theoretical guarantees, FedL2G efficiently implements the learning-to-guide process using only first-order derivatives w.r.t. model parameters and achieves a non-convex convergence rate of $\\mathcal{O}(1/T)$. We conduct extensive experiments on two data heterogeneity and six model heterogeneity settings using 14 heterogeneous model architectures (e.g., CNNs and ViTs) to demonstrate FedL2G\u2019s superior performance compared to six counterparts.", "keywords": [ "Heterogeneous Federated Learning", "Model Heterogeneity" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "dePB45VMFx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Aaron Foss", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Andrew Cohen", "gender": "Male", "institution": "Meta Platforms", "country": "US", "position": "Researcher" }, { "name": "Da JU", "gender": "Unspecified", "institution": "Facebook", "country": "", "position": "Research Engineer" }, { "name": "Justine T Kao", "gender": "Female", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Sasha Mitts", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Song Jiang", "gender": "Male", "institution": "FAIR", "country": "US", "position": "Research Engineer" }, { "name": "Xian Li", "gender": "unknown", "institution": "Facebook AI", "country": "", "position": "Principal Researcher" }, { "name": "Yuandong Tian", "gender": "Male", "institution": "Meta AI (FAIR)", "country": "", "position": "Research Scientist" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 46, "n_ref": 80, "n_ref_all": 82, "n_fig": 1, "n_tab": 9, "L_tab": 5815, "n_element_tab": 341, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1194, "n_element_tab_1": 165, "formula_len_all": 50, "formula_len_all_1": 36, "len_all": 189510, "len_all_1": 65308, "len_abs": 1014, "len_title": 126, "len_sents": 49798, "len_sents_1": 28550, "n_sents": 463, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1408, "title": "Towards Full Delegation: Designing Ideal Agentic Behaviors for Travel Planning", "abs": "How are LLM-based agents used in the future? While many of the existing work on agents has focused on improving the performance of a specific family of objective and challenging tasks, in this work, we take a different perspective by thinking about full delegation: agents take over humans\u2019 routine decision-making processes and are trusted by humans to find solutions that fit people\u2019s personalized needs and are adaptive to ever-changing context. In order to achieve such a goal, the behavior of the agents, i.e., agentic behaviors, should be evaluated not only on their achievements (i.e., outcome evaluation), but also how they achieved that (i.e., procedure evaluation). For this, we propose APEC Agent Constitution, a list of criteria that an agent should follow for good agentic behaviors, including Accuracy, Proactivity, Efficiency and Credibility. To verify whether APEC aligns with human preferences, we develop APEC-Travel, a travel planning agent that proactively extracts hidden personalized needs via multi-round dialog with travelers. APEC-Travel is constructed purely from synthetic data generated by Llama3.1-405B-Instruct with a diverse set of travelers\u2019 persona to simulate rich distribution of dialogs. Iteratively fine-tuned to follow APEC Agent Constitution, APEC-Travel surpasses baselines by 20.7% on rule based metrics and 9.1% on LLM-as-a-Judge scores across the constitution axes.", "keywords": [ "Agentic Behavior", "LLMs", "LLM-as-a-Judge", "Synthetic Data" ], "rating_list": [ 6, 3, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 2, 4 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "ddOxvs4NAq", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Jacek Ma\u0144dziuk", "gender": "Male", "institution": "AGH University of Science and Technology", "country": "PL", "position": "Full Professor" }, { "name": "Miko\u0142aj Ma\u0142ki\u0144ski", "gender": "Male", "institution": "Warsaw University of Technology", "country": "PL", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 54, "n_ref": 112, "n_ref_all": 142, "n_fig": 6, "n_tab": 23, "L_tab": 10057, "n_element_tab": 135, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 56, "n_element_tab_1": 12, "formula_len_all": 816, "formula_len_all_1": 211, "len_all": 219479, "len_all_1": 64978, "len_abs": 1308, "len_title": 121, "len_sents": 57151, "len_sents_1": 28683, "n_sents": 507, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1314, "title": "Generalization and Knowledge Transfer in Abstract Visual Reasoning Models", "abs": "We study generalization and knowledge reuse capabilities of deep neural networks in the domain of abstract visual reasoning (AVR), employing Raven's Progressive Matrices (RPMs), a recognized benchmark task for assessing AVR abilities. Two knowledge transfer scenarios referring to the I-RAVEN dataset are investigated. Firstly, inspired by generalization assessment capabilities of the PGM dataset and popularity of I-RAVEN, we introduce Attributeless-I-RAVEN, a benchmark with $10$ generalization regimes that allow to test generalization of abstract rules applied to held-out attributes. Secondly, we construct I-RAVEN-Mesh, a dataset that enriches RPMs with a novel component structure comprising line-based patterns, facilitating assessment of progressive knowledge acquisition in transfer learning setting. The developed benchmarks reveal shortcomings of the contemporary deep learning models, which we partly address with Pathways of Normalized Group Convolution (PoNG) model, a novel neural architecture for solving AVR tasks. PoNG excels in both presented challenges, as well as the standard I-RAVEN and PGM setups. Encouraged by these promising results, we further evaluate PoNG in another AVR task, visual analogy problem with both synthetic and real-world images, demonstrating its strength beyond PRMs.", "keywords": [ "Abstract Visual Reasoning", "Deep Learning", "Generalization", "Transfer Learning", "Raven's Progressive Matrices" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "ddNZLAWPdT", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Christopher Brinton", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "Surojit Ganguli", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Zeyu Zhou", "gender": "Unspecified", "institution": "Purdue University", "country": "", "position": "PhD student" }, { "name": "David I Inouye", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 32, "n_formula_1": 15, "n_ref_uni": 35, "n_ref": 70, "n_ref_all": 112, "n_fig": 16, "n_tab": 8, "L_tab": 1804, "n_element_tab": 187, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1421, "n_element_tab_1": 144, "formula_len_all": 3888, "formula_len_all_1": 1420, "len_all": 196411, "len_all_1": 69981, "len_abs": 947, "len_title": 121, "len_sents": 72741, "len_sents_1": 33772, "n_sents": 526, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 966, "title": "Robust Decentralized VFL Over Dynamic Device Environment", "abs": "Robust collaborative learning on a network of edge devices, for vertically split datasets, is challenging because edge devices may fail due to environment conditions or events such as extreme weather. The current Vertical Federated learning (VFL) approaches assume a centralized learning setup or assume the active party or server cannot fail. To address these limitations, we first formalize the problem of VFL under dynamic network conditions such as faults (named DN-VFL). Then, we develop a novel DN-VFL method called **M**ultiple **A**ggregation with **G**ossip Rounds and **S**imulated Faults (MAGS) that synthesizes faults via dropout, replication, and\ngossiping to improve robustness significantly over baselines. We also theoretically analyze our proposed approaches to explain why they enhance robustness. Extensive empirical results validate that MAGS is robust across a range of fault rates\u2014including extreme fault rates\u2014compared to prior VFL approaches.", "keywords": [ "Robustness", "Vertical Federated Learning" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "dd2CABUZaw", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Chengjin Xu", "gender": "Male", "institution": "International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Chun Yuan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jian Guo", "gender": "Male", "institution": "International Digital Economy Academy, International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "SiNan Du", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yiyan Qi", "gender": "unknown", "institution": "IDEA", "country": "CN", "position": "Researcher" }, { "name": "Zhengzhuo Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 45, "n_ref": 106, "n_ref_all": 140, "n_fig": 21, "n_tab": 21, "L_tab": 20492, "n_element_tab": 2568, "n_fig_1": 13, "n_tab_1": 7, "L_tab_1": 10114, "n_element_tab_1": 1422, "formula_len_all": 189, "formula_len_all_1": 167, "len_all": 338648, "len_all_1": 99663, "len_abs": 1942, "len_title": 110, "len_sents": 88104, "len_sents_1": 28759, "n_sents": 819, "n_sents_1": 279, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1151, "title": "ChartBench: A Benchmark for Complex Visual Reasoning in Charts", "abs": "Multimodal Large Language Models (MLLMs) have shown impressive capabilities in image understanding and generation. However, current benchmarks fail to accurately evaluate the chart comprehension of MLLMs due to limited chart types and inappropriate metrics. To address this, we propose ChartBench, a comprehensive benchmark designed to assess chart comprehension and data reliability through complex visual reasoning. ChartBench includes 42 categories, 66.6k charts, and 600k question-answer pairs. Notably, we do not provide data point annotations on charts explicitly, which requires models to derive values by leveraging inherent chart elements such as color, legends, and coordinate systems. We also design an enhanced evaluation metric named Acc++ to evaluate MLLMs without extensive manual or costly LLM-based evaluations. Furthermore, we propose two baselines based on the chain of thought and supervised fine-tuning to improve model performance on unannotated charts. Extensive experimental evaluations of 18 open-sourced and 3 proprietary MLLMs reveal their limitations in chart comprehension and offer valuable insights for further research.", "keywords": [ "Multimodal Large Language Models", "Chart Reasoning", "Chart Benchmark", "Chain of Thought" ], "rating_list": [ 6, 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 5, 5 ] }, { "paper_id": "dd0rUW29tQ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Hadi Jamali-Rad", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "Hamed Pirsiavash", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Associate Professor" }, { "name": "Navaneet K L", "gender": "Male", "institution": "University of California, Davis", "country": "", "position": "PhD student" }, { "name": "Soroush Abbasi Koohpayegani", "gender": "Male", "institution": "University of California, Davis", "country": "", "position": "PhD student" }, { "name": "Anuj Rajeeva Singh", "gender": "Male", "institution": "Shell Global Solutions B.V.", "country": "NL", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 78, "n_ref": 181, "n_ref_all": 226, "n_fig": 11, "n_tab": 6, "L_tab": 2437, "n_element_tab": 287, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 361, "n_element_tab_1": 57, "formula_len_all": 317, "formula_len_all_1": 317, "len_all": 243505, "len_all_1": 64509, "len_abs": 1258, "len_title": 104, "len_sents": 60457, "len_sents_1": 32396, "n_sents": 407, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1191, "title": "GeNIe: Generative Hard Negative Images Through Diffusion", "abs": "Data augmentation is crucial in training deep models, preventing them from overfitting to limited data. Recent advances in generative AI, e.g., diffusion models, have enabled more sophisticated augmentation techniques that produce data resembling natural images. We introduce GeNIe a novel augmentation method which leverages a latent diffusion model conditioned on a text prompt to combine two contrasting data points (an image from the source category and a text prompt from the target category) to generate challenging augmentations. To achieve this, we adjust the noise level (equivalently, number of diffusion iterations) to ensure the generated image retains low-level and background features from the source image while representing the target category, resulting in a hard negative sample for the source category. We further automate and enhance GeNIe by adaptively adjusting the noise level selection on a per image basis (coined as GeNIe-Ada), leading to further performance improvements. Our extensive experiments, in both few-shot and long-tail distribution settings, demonstrate the effectiveness of our novel augmentation method and its superior performance over the prior art.", "keywords": [ "Data Augmentation;Diffusion Models; Computer Vision; Few-shot Learning; Long-tail Classification;" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "dcG17rjJF9", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Chuqin Zhou", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Gang chen", "gender": "Male", "institution": "china telecommunications corporation", "country": "CN", "position": "Researcher" }, { "name": "Guo Lu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Junhao Du", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Li Song", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Ning cao", "gender": "Male", "institution": "china telecommunications corporation", "country": "CN", "position": "Researcher" }, { "name": "Wenjun Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Yunuo Chen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhengxue Cheng", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 50, "n_ref": 111, "n_ref_all": 133, "n_fig": 5, "n_tab": 13, "L_tab": 1815, "n_element_tab": 337, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1092, "n_element_tab_1": 205, "formula_len_all": 288, "formula_len_all_1": 288, "len_all": 151865, "len_all_1": 60692, "len_abs": 1564, "len_title": 119, "len_sents": 38801, "len_sents_1": 26410, "n_sents": 288, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1570, "title": "Large Language Model for Lossless Image Compression with Visual Prompts", "abs": "Recent advancements in deep learning have driven significant progress in lossless image compression. With the emergence of Large Language Models (LLMs), preliminary attempts have been made to leverage the extensive prior knowledge embedded in these pretrained models to enhance lossless image compression, particularly by improving the entropy model. However, a significant challenge remains in bridging the gap between the textual prior knowledge within LLMs and lossless image compression.\nTo tackle this challenge and unlock the potential of LLMs, this paper introduces a novel paradigm for lossless image compression that incorporates LLMs with visual prompts. Specifically, we first generate a lossy reconstruction of the input image as visual prompts, from which we extract local and global features to serve as visual embeddings for the LLM. The residual between the original image and the lossy reconstruction is then fed into the LLM along with these visual embeddings, enabling the LLM to function as an entropy model to predict the probability distribution of the residual.\nExtensive experiments on multiple benchmark datasets demonstrate our method achieves state-of-the-art compression performance, surpassing both traditional and learning-based lossless image codecs. Furthermore, our approach can be easily extended to images from other domains, such as medical and screen content images, achieving impressive performance. These results highlight the potential of LLMs for lossless image compression and may inspire further research in related directions.", "keywords": [ "Image Compression", "Lossless Image Compression", "Lossy Image Compression", "Video Compression" ], "rating_list": [ 5, 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 4, 3 ] }, { "paper_id": "dbwF3QFWGn", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Haosen Ge", "gender": "Male", "institution": "The Wharton School, University of Pennsylvania", "country": "US", "position": "Researcher" }, { "name": "Osbert Bastani", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Assistant Professor" }, { "name": "Hamsa Sridhar Bastani", "gender": "Female", "institution": "The Wharton School, University of Pennsylvania", "country": "", "position": "Assistant Professor" } ], "n_formula": 32, "n_formula_1": 28, "n_ref_uni": 21, "n_ref": 36, "n_ref_all": 44, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2031, "formula_len_all_1": 1481, "len_all": 95065, "len_all_1": 81799, "len_abs": 1216, "len_title": 112, "len_sents": 30826, "len_sents_1": 28407, "n_sents": 302, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1222, "title": "Stochastic Online Conformal Prediction with Semi-Bandit Feedback", "abs": "Conformal prediction has emerged as an effective strategy for uncertainty quantification by modifying a model to output sets of labels instead of a single label. These prediction sets come with the guarantee that they contain the true label with high probability. However, conformal prediction typically requires a large calibration dataset of i.i.d. examples. We consider the online learning setting, where examples arrive over time, and the goal is to construct prediction sets dynamically. Departing from existing work, we assume semi-bandit feedback, where we only observe the true label if it is contained in the prediction set. For instance, consider calibrating a document retrieval model to a new domain; in this setting, a user would only be able to provide the true label if the target document is in the prediction set of retrieved documents. We propose a novel conformal prediction algorithm targeted at this setting, and prove that it obtains sublinear regret compared to the optimal conformal predictor. We evaluate our algorithm on a retrieval task, an image classification task, and an auction price-setting task, and demonstrate that it empirically achieves good performance compared to several baselines.", "keywords": [ "Conformal Predictions", "Online Learning", "Semi-bandit Feedback" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "dbuFJg7eaw", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Chenyang Cao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Junbo Tan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Longxiang He", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Silang Wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xueqian Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yucheng Xin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Zichen Yan", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 49, "n_formula_1": 28, "n_ref_uni": 55, "n_ref": 115, "n_ref_all": 150, "n_fig": 19, "n_tab": 2, "L_tab": 631, "n_element_tab": 101, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3404, "formula_len_all_1": 2039, "len_all": 208406, "len_all_1": 67618, "len_abs": 1243, "len_title": 106, "len_sents": 62733, "len_sents_1": 30575, "n_sents": 578, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1250, "title": "FOSP: Fine-tuning Offline Safe Policy through World Models", "abs": "Offline Safe Reinforcement Learning (RL) seeks to address safety constraints by learning from static datasets and restricting exploration. However, these approaches heavily rely on the dataset and struggle to generalize to unseen scenarios safely. In this paper, we aim to improve safety during the deployment of vision-based robotic tasks through online fine-tuning an offline pretrained policy. To facilitate effective fine-tuning, we introduce model-based RL, which is known for its data efficiency. Specifically, our method employs in-sample optimization to improve offline training efficiency while incorporating reachability guidance to ensure safety. After obtaining an offline safe policy, safe policy expansion approach is leveraged for online fine-tuning. The performance of our method is validated on simulation benchmarks with five vision-only tasks and through real-world robot deployment using limited data. It demonstrates that our approach significantly improves the generalization of offline policies to unseen safety-constrained scenarios. To the best of our knowledge, this is the first work to explore offline-to-online RL for safe generalization tasks. The videos are available at https://sites.google.com/view/safefinetune/home.", "keywords": [ "Safe RL", "Offline-to-online RL", "Robot Learning" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "dbiLOMgMm7", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Andrew M Saxe", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "Associate Professor" }, { "name": "Christopher Summerfield", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "Full Professor" }, { "name": "Jan Philipp Bauer", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Jirko Rubruck", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "PhD student" } ], "n_formula": 41, "n_formula_1": 6, "n_ref_uni": 59, "n_ref": 116, "n_ref_all": 187, "n_fig": 21, "n_tab": 1, "L_tab": 182, "n_element_tab": 5, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3230, "formula_len_all_1": 349, "len_all": 209234, "len_all_1": 69611, "len_abs": 1665, "len_title": 78, "len_sents": 66023, "len_sents_1": 31856, "n_sents": 581, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1676, "title": "Early learning of the optimal constant solution in neural networks and humans", "abs": "Deep neural networks learn increasingly complex functions over the course of training. Here, we show both empirically and theoretically that learning of the target function is preceded by an early phase in which networks learn the optimal constant solution (OCS) \u2013 that is, initial model responses mirror the distribution of target labels, while entirely ignoring information provided in the input. Using a hierarchical category learning task, we derive exact solutions for learning dynamics in deep linear networks trained with bias terms. Even when initialized to zero, this simple architectural feature induces substantial changes in early dynamics. We identify hallmarks of this early OCS phase and illustrate how these signatures are observed in deep linear networks and larger, more complex (and nonlinear) convolutional neural networks solving a hierarchical learning task based on MNIST and CIFAR10. We explain these observations by proving that deep linear networks necessarily learn the OCS during early learning. To further probe the generality of our results, we train human learners over the course of three days on a structurally equivalent learning task. We then identify qualitative signatures of this early OCS phase in terms of true negative rates. Surprisingly, we find the same early reliance on the OCS in the behaviour of human learners. Finally, we show that learning of the OCS can emerge even in the absence of bias terms and is equivalently driven by generic correlations in the input data. Overall, our work suggests the OCS as a common learning principle in supervised, error-corrective learning, and suggests possible factors for its prevalence.", "keywords": [ "Simplicity Bias", "Deep Linear Networks", "cognitive science", "neuroscience", "learning dynamics" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 4, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "daVCPIBCtQ", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 5, "author_info_list": [ { "name": "Bo Jiang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Tian Guo", "gender": "Female", "institution": "Worcester Polytechnic Institute", "country": "US", "position": "Associate Professor" }, { "name": "Yunzhuo Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Zhi Han", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Zihao Fan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 9, "n_ref_uni": 22, "n_ref": 53, "n_ref_all": 82, "n_fig": 13, "n_tab": 8, "L_tab": 1815, "n_element_tab": 259, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 2137, "n_element_tab_1": 204, "formula_len_all": 1210, "formula_len_all_1": 558, "len_all": 121938, "len_all_1": 57279, "len_abs": 1265, "len_title": 153, "len_sents": 39213, "len_sents_1": 25697, "n_sents": 319, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 106, "L_abs": 1276, "title": "DyGNeX : Efficient Distributed Training of Dynamic Graph Neural Networks with Cross-Time-Window Scheduling", "abs": "Dynamic Graph Neural Networks (DGNNs) are advanced methods for processing evolving graph data, capturing both structural and temporal dependencies efficiently. However, existing distributed DGNN training methods face challenges in achieving load balance across GPUs and minimizing communication overhead, which limits their efficiency. In this paper, we introduce DyGNeX, a distributed training system designed to address this issue. DyGNeX utilizes a cross-time-window snapshot group scheduling algorithm that balances computational loads across GPUs without introducing additional cross-GPU feature aggregation or hidden state communication. Based on the specific scenario, the scheduling algorithm is applied using greedy or Integer Linear Programming (ILP) methods, referred to as DyGNeX-G and DyGNeX-L, respectively. DyGNeX-L and DyGNeX-G achieve average reductions of 28\\% and 24\\% in per-epoch training time compared to state-of-the-art methods, maintaining load imbalance across GPUs at approximately 4\\% and 8\\%, while preserving model convergence across various DGNN models and datasets. In simulation experiments, as the number of GPUs increases, DyGNeX-G shows good scalability, efficiently handling clusters with up to 512 GPUs while maintaining 95\\% efficiency.", "keywords": [ "Dynamic Graph Neural Networks", "Distributed training", "load balance" ], "rating_list": [ 3, 3, 5, 3, 6 ], "soundness_list": [ 3, 1, 3, 3, 4 ], "presentation_list": [ 2, 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 2, 2, 2 ] }, { "paper_id": "daUQ7vmGap", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Boqian Wu", "gender": "Female", "institution": "University of Twente", "country": "", "position": "PhD student" }, { "name": "Decebal Constantin Mocanu", "gender": "Male", "institution": "University of Luxemburg", "country": "LU", "position": "Associate Professor" }, { "name": "Elena Mocanu", "gender": "Female", "institution": "University of Twente", "country": "", "position": "Assistant Professor" }, { "name": "Maurice van Keulen", "gender": "Male", "institution": "University of Twente", "country": "NL", "position": "Associate Professor" }, { "name": "Mykola Pechenizkiy", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Full Professor" }, { "name": "Nicola Strisciuglio", "gender": "Male", "institution": "University of Twente", "country": "NL", "position": "Associate Professor" }, { "name": "Qiao Xiao", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "PhD student" }, { "name": "Shunxin Wang", "gender": "Female", "institution": "University of Twente", "country": "NL", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 33, "n_ref": 57, "n_ref_all": 96, "n_fig": 9, "n_tab": 7, "L_tab": 4428, "n_element_tab": 248, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 422, "n_element_tab_1": 29, "formula_len_all": 134, "formula_len_all_1": 121, "len_all": 188020, "len_all_1": 55785, "len_abs": 1200, "len_title": 149, "len_sents": 48065, "len_sents_1": 27884, "n_sents": 300, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 98, "L_abs": 1214, "title": "Dynamic Sparse Training versus Dense Training: The Unexpected Winner in Image Corruption Robustnes", "abs": "It is generally perceived that Dynamic Sparse Training opens the door to a new era of scalability and efficiency for artificial neural networks at, perhaps, some costs in accuracy performance for the classification task. At the same time, Dense Training is widely accepted as being the \"de facto\" approach to train artificial neural networks if one would like to maximize their robustness against image corruption. In this paper, we question this general practice. Consequently, \\textit{we claim that}, contrary to what is commonly thought, the Dynamic Sparse Training methods can consistently outperform Dense Training in terms of robustness accuracy, particularly if the efficiency aspect is not considered as a main objective (i.e., sparsity levels between 10\\% and up to 50\\%), without adding (or even reducing) resource cost. We validate our claim on two types of data, images and videos, using several traditional and modern deep learning architectures for computer vision and three widely studied Dynamic Sparse Training algorithms. Our findings reveal a new yet-unknown benefit of Dynamic Sparse Training and open new possibilities in improving deep learning robustness beyond the current state of the art.", "keywords": [ "Dynamic Sparse Training", "Image Corruption Robustnes" ], "rating_list": [ 3, 8, 6, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "daRu82GAoZ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Wenhao Wang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Yi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yifan Sun", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Senior Expert" }, { "name": "Zhengdong Hu", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Zhentao Tan", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Zongxin Yang", "gender": "Male", "institution": "Harvard Medical School, Harvard University", "country": "US", "position": "Postdoc" } ], "n_formula": 35, "n_formula_1": 16, "n_ref_uni": 24, "n_ref": 49, "n_ref_all": 75, "n_fig": 13, "n_tab": 8, "L_tab": 2512, "n_element_tab": 149, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 2386, "n_element_tab_1": 86, "formula_len_all": 1918, "formula_len_all_1": 764, "len_all": 144880, "len_all_1": 58927, "len_abs": 894, "len_title": 131, "len_sents": 39182, "len_sents_1": 24723, "n_sents": 340, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1836, "title": "Generalizable Origin Identification for Text-Guided Image-to-Image Diffusion Models", "abs": "Text-guided image-to-image diffusion models excel in translating images based on textual prompts, allowing for precise and creative visual modifications. However, such a powerful technique can be misused for *spreading misinformation*, *infringing on copyrights*, and *evading content tracing*. This motivates us to introduce the task of origin **ID**entification for text-guided **I**mage-to-image **D**iffusion models (**ID$\\mathbf{^2}$**), aiming to retrieve the original image of a given translated query. A straightforward solution to ID$^2$ involves training a specialized deep embedding model to extract and compare features from both query and reference images. However, due to *visual discrepancy* across generations produced by different diffusion models, this similarity-based approach fails when training on images from one model and testing on those from another, limiting its effectiveness in real-world applications. To solve this challenge of the proposed ID$^2$ task, we contribute the first dataset and a theoretically guaranteed method, both emphasizing generalizability. The curated dataset, **OriPID**, contains abundant **Ori**gins and guided **P**rompts, which can be used to train and test potential **ID**entification models across various diffusion models. In the method section, we first prove the *existence* of a linear transformation that minimizes the distance between the pre-trained Variational Autoencoder (VAE) embeddings of generated samples and their origins. Subsequently, it is demonstrated that such a simple linear transformation can be *generalized* across different diffusion models. Experimental results show that the proposed method achieves satisfying generalization performance, significantly surpassing similarity-based methods (+31.6% mAP), even those with domain generalization designs.", "keywords": [ "Diffusion Models", "AI Security", "Origin Identification" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "daD6uGMeLs", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Hajime Nagahara", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "Full Professor" }, { "name": "Hideaki Hayashi", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "Associate Professor" }, { "name": "Yicheng Deng", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 18, "n_ref_uni": 32, "n_ref": 54, "n_ref_all": 70, "n_fig": 7, "n_tab": 5, "L_tab": 1311, "n_element_tab": 115, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1053, "n_element_tab_1": 41, "formula_len_all": 1236, "formula_len_all_1": 992, "len_all": 123597, "len_all_1": 61001, "len_abs": 1326, "len_title": 148, "len_sents": 37485, "len_sents_1": 28604, "n_sents": 253, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 99, "L_abs": 1334, "title": "Gaussian-Based Instance-Adaptive Intensity Modeling for Point-Supervised Facial Expression Spotting", "abs": "Point-supervised facial expression spotting (P-FES) aims to localize facial expression instances in untrimmed videos, requiring only a single timestamp label for each instance during training. To address label sparsity, hard pseudo-labeling is often employed to propagate point labels to unlabeled frames; however, this approach can lead to confusion when distinguishing between neutral and expression frames with various intensities, which can negatively impact model performance. In this paper, we propose a two-branch framework for P-FES that incorporates a Gaussian-based instance-adaptive Intensity Modeling (GIM) module for soft pseudo-labeling. GIM models the expression intensity distribution for each instance. Specifically, we detect the pseudo-apex frame around each point label, estimate the duration, and construct a Gaussian distribution for each expression instance. We then assign soft pseudo-labels to pseudo-expression frames as intensity values based on the Gaussian distribution. Additionally, we introduce an Intensity-Aware Contrastive (IAC) loss to enhance discriminative feature learning and suppress neutral noise by contrasting neutral frames with expression frames of various intensities. Extensive experiments on the SAMM-LV and CAS(ME)$^2$ datasets demonstrate the effectiveness of our proposed framework.", "keywords": [ "micro-expression spotting", "semi-supervised learning", "soft pseudo-labeling" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "daByonGVyo", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Anian Ruoss", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "David Heurtel-Depeiges", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "PhD student" }, { "name": "Joel Veness", "gender": "unknown", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Tim Genewein", "gender": "Male", "institution": "DeepMind", "country": "", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 69, "n_ref": 148, "n_ref_all": 185, "n_fig": 8, "n_tab": 5, "L_tab": 1327, "n_element_tab": 189, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 688, "n_element_tab_1": 92, "formula_len_all": 233, "formula_len_all_1": 232, "len_all": 178733, "len_all_1": 63994, "len_abs": 2275, "len_title": 127, "len_sents": 45936, "len_sents_1": 33887, "n_sents": 249, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1520, "title": "Compression via Pre-trained Transformers: A Study on Byte-Level Multimodal Data", "abs": "Foundation models have recently been shown to be strong data compressors. However, when accounting for their excessive parameter count, their compression ratios are actually inferior to standard compression algorithms. Moreover, naively reducing the number of parameters may not necessarily help as it leads to worse predictions and thus weaker compression. In this paper, we conduct a large-scale empirical study to investigate whether there is a sweet spot where competitive compression ratios with pre-trained vanilla transformers are possible. To this end, we train families of models on 165GB of raw byte sequences of either text, image, or audio data (and all possible combinations of the three) and then compress 1GB of out-of-distribution (OOD) data from each modality. We find that relatively small models (i.e., millions of parameters) can outperform standard general-purpose compression algorithms (gzip, LZMA2) and even domain-specific compressors (PNG, JPEG~2000, FLAC) — even when factoring in parameter count. We achieve, e.g., the lowest compression ratio of 0.49 on OOD audio data (vs. 0.54 for FLAC). To study the impact of model- and dataset scale, we conduct extensive ablations and hyperparameter sweeps, and we investigate the effect of unimodal versus multimodal training. We find that even small models can be trained to perform well on multiple modalities, but, in contrast to previously reported results with large-scale foundation models, transfer to unseen modalities is generally weak.", "keywords": [ "lossless compression", "transformers", "multi-modality" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "dZsjj4vQjl", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Libo Qin", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "Full Professor" }, { "name": "Min-Yen Kan", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Associate Professor" }, { "name": "Wanxiang Che", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Xiao Xu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yuxi Xie", "gender": "Female", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Tianhao niu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 71, "n_ref": 269, "n_ref_all": 327, "n_fig": 13, "n_tab": 23, "L_tab": 4719, "n_element_tab": 390, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1580, "n_element_tab_1": 137, "formula_len_all": 142, "formula_len_all_1": 93, "len_all": 412510, "len_all_1": 65294, "len_abs": 2542, "len_title": 80, "len_sents": 126937, "len_sents_1": 31128, "n_sents": 815, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1288, "title": "Exploring Multi-Grained Concept Annotations for Multimodal Large Language Models", "abs": "Multimodal Large Language Models (MLLMs) excel in vision--language tasks by pre-training solely on coarse-grained concept annotations (e.g., image captions).\nWe hypothesize that integrating fine-grained concept annotations (e.g., object labels and object regions) will further improve performance, as both data granularities complement each other in terms of breadth and depth in concept representation.\n\nWe introduce a new dataset featuring Multimodal Multi-Grained Concept annotations (MMGiC) for MLLMs.\nIn constructing MMGiC, we explore the impact of different data recipes on multimodal comprehension and generation.\nOur analyses reveal that multi-grained concept annotations integrate and complement each other, under our structured template and autoregressive discrete framework.\n\nWe definitively show that multi-grained concepts do facilitate MLLMs to better locate and learn concepts, aligning vision and language at multiple granularities.\nWe further validate our hypothesis by investigating the comparison and collaboration between MMGiC and image--caption data on 12 multimodal comprehension and generation benchmarks, e.g., their appropriate combination achieve 3.95% and 2.34% accuracy improvements on POPE and SEED-Bench.\nCode, data and models will be made openly available.", "keywords": [ "Multimodal Large Language Model", "Multi-Grained Annotations", "Fine-Grained Annotations", "Concept Annotations", "Vision--Language Learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "dZbCoATni7", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Binyuan Huang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Chang Wen Chen", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Haoqiang Fan", "gender": "Male", "institution": "Kohat University of Science and Technology (KUST)", "country": "KH", "position": "Researcher" }, { "name": "Tiancai Wang", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Weixin Mao", "gender": "Male", "institution": "Waseda University", "country": "JP", "position": "PhD student" }, { "name": "Yaosi Hu", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Postdoc" }, { "name": "Yucheng Zhao", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Zhenzhong Chen", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Zihan Lan", "gender": "unknown", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 31, "n_ref": 50, "n_ref_all": 73, "n_fig": 10, "n_tab": 6, "L_tab": 1745, "n_element_tab": 219, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 1562, "n_element_tab_1": 181, "formula_len_all": 367, "formula_len_all_1": 231, "len_all": 122308, "len_all_1": 59132, "len_abs": 1156, "len_title": 103, "len_sents": 40739, "len_sents_1": 28558, "n_sents": 266, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1159, "title": "Embodied Scene Cloning: Solving Generalization in Embodied AI via Visual-Prompt Image Editing", "abs": "Recent advancements in robotic learning have enabled robots to perform a wide range of tasks. However, generalizing policies from training environments to deployment environments remains a major challenge, and improving these policies by collecting and annotating demonstrations in target environments is both costly and time-consuming. To address this issue, we propose Embodied Scene Cloning, a novel visual-prompt-based framework that generates visual-aligned trajectories from existing data by leveraging visual cues from the specific deployment environment. This approach minimizes the impact of environmental discrepancies on policy performance. Unlike traditional embodied augmentation methods that rely on text prompts, we propose to \"clone\" source demonstrations into the target environment and edit it with visual prompt to effectively improve the generalization ability on specific embodied scene. Experimental results demonstrate that samples generated by Embodied Scene Cloning significantly enhance the generalization ability of policies in the target deployment environments, representing a meaningful advancement in embodied data augmentation.", "keywords": [ "Embodied Intelligence", "Data Augmentation" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "dZNI8DyUKY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Chun-Shien Lu", "gender": "Male", "institution": "Academia Sinica", "country": "TW", "position": "Full Professor" }, { "name": "Yuan-Chih Chen", "gender": "Male", "institution": "Academia Sinica", "country": "", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 37, "n_ref": 113, "n_ref_all": 160, "n_fig": 2, "n_tab": 8, "L_tab": 1763, "n_element_tab": 233, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 3488, "n_element_tab_1": 351, "formula_len_all": 630, "formula_len_all_1": 464, "len_all": 118121, "len_all_1": 65128, "len_abs": 1057, "len_title": 116, "len_sents": 29835, "len_sents_1": 26248, "n_sents": 203, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 914, "title": "Adversarial Masked Autoencoder Purifier with Defense Transferability", "abs": "The study of adversarial defense still struggles to combat with advanced adversarial attacks.\nIn contrast to most prior studies that rely on the diffusion model for test-time defense to remarkably increase the inference time, we propose Masked AutoEncoder Purifier (MAEP), which integrates Masked AutoEncoder (MAE) into an adversarial purifier\nframework for test-time purification.\nWhile MAEP achieves promising adversarial robustness, it particularly features model defense transferability without relying on using additional data that is different from the training dataset. \nTo our knowledge, MAEP is the first study of adversarial purifier based on masked autoencoder.\nExtensive experiments validate the proposed method.\nNotably, MAEP trained on CIFAR10 achieves state-of-the-art performance even when tested directly on ImageNet, outperforming existing diffusion-based models trained specifically on ImageNet.", "keywords": [ "Adversarial attack", "Adversarial defense", "Purifier", "Robustness", "Security", "Transferability" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "dZ3cI69BE8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Beomseok Kwon", "gender": "Male", "institution": "NAVER CLOUD", "country": "", "position": "Researcher" }, { "name": "Byeongwook Kim", "gender": "unknown", "institution": "NAVER CLOUD", "country": "", "position": "Researcher" }, { "name": "Dongsoo Lee", "gender": "Male", "institution": "NAVER CLOVA", "country": "", "position": "Executive Officer" }, { "name": "Jeongin Bae", "gender": "Female", "institution": "NAVER", "country": "", "position": "Researcher" }, { "name": "Se Jung Kwon", "gender": "Male", "institution": "NAVER Cloud", "country": "KR", "position": "AI Researcher" }, { "name": "Seungcheol Park", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "U Kang", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" } ], "n_formula": 23, "n_formula_1": 19, "n_ref_uni": 27, "n_ref": 70, "n_ref_all": 116, "n_fig": 6, "n_tab": 14, "L_tab": 3861, "n_element_tab": 423, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1122, "n_element_tab_1": 203, "formula_len_all": 2712, "formula_len_all_1": 1752, "len_all": 174481, "len_all_1": 67824, "len_abs": 1642, "len_title": 118, "len_sents": 57200, "len_sents_1": 28150, "n_sents": 460, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1274, "title": "FlexBCQ: Flexible Binary-coding Quantization for Large Language Models", "abs": "How can we compress large language models without compromising accuracy?\nQuantization, which reduces the number of bits for representing weights, is an essential technique to utilize large language models (LLMs) in real-world applications.\nSpecifically, binary-coding quantization (BCQ) is a promising approach since it has extensive representation space, which encompasses the representation space of uniform quantization (UQ), and fast inference speed.\nHowever, because of the lack of accurate optimization techniques, BCQ shows inferior performance compared to UQ algorithms, failing to leverage their powerful expressive power.\nIn this paper, we propose FlexBCQ (Flexible Binary-coding Quantization), an accurate optimization algorithm for BCQ.\nWe leverage the sophisticated optimization techniques of UQ by decomposing the quantization process of BCQ into the composition of a UQ and an inner BCQ.\nAs a result, we take advantage of both the sophisticated optimizing techniques of UQ, specifically the flexible mapping technique, and the powerful expressive capability of BCQ.\nThrough extensive experiments, we find that FlexBCQ provides 3.24%p higher accuracy than existing UQ and BCQ algorithms on MMLU 5-shot benchmark when quantizing a Llama-3 70B model into 3 bits.", "keywords": [ "LLM", "Quantization", "Binary-coding Quantization (BCQ)", "Uniform Quantization (UQ)" ], "rating_list": [ 6, 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 2, 4, 4 ] }, { "paper_id": "dYc55Hvm3p", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Huafeng Li", "gender": "Male", "institution": "Kunmimg University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Mengzhao Wang", "gender": "Male", "institution": "Kunmimg University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Yafei Zhang", "gender": "Female", "institution": "Kunmimg University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Jinxing Li", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 19, "n_ref": 26, "n_ref_all": 57, "n_fig": 7, "n_tab": 2, "L_tab": 486, "n_element_tab": 47, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 626, "n_element_tab_1": 138, "formula_len_all": 1021, "formula_len_all_1": 1023, "len_all": 116712, "len_all_1": 70794, "len_abs": 2131, "len_title": 145, "len_sents": 39103, "len_sents_1": 33003, "n_sents": 299, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 2190, "title": "Disentangling Inter- and Intra-Video Relations for Multi-event Video-Text Retrieval and Grounding", "abs": "Video-text retrieval aims to precisely search for videos most relevant to a text query within a video corpus. However, existing methods are largely limited to single-text (single-event) queries, which are not effective at handling multi-text (multi-event) queries. Furthermore, these methods typically focus solely on retrieval and do not attempt to locate multiple events within the retrieved videos. To address these limitations, our paper proposes a novel method named Disentangling Inter- and Intra-Video relations, which jointly considers multi-event video-text retrieval and grounding. This method leverages both inter-video and intra-video event relationships to enhance the performance of retrieval and grounding. At the retrieval level, we devise a Relational Event-Centric Video-Text Retrieval module, based on the principle that more comprehensive textual information leads to a more precise correspondence between text and video. It incorporates event relationship features at different hierarchical levels and exploits the hierarchical structure of corresponding video relationships to achieve multi-level contrastive learning between events and videos. This approach enhances the richness, accuracy, and comprehensiveness of event descriptions, improving alignment precision between text and video and enabling effective differentiation among videos. For event localization, we propose Event Contrast-Driven Video Grounding, which accounts for positional differences between different events and achieves precise grounding of multiple events through divergence learning of event locations. Our solution not only provides efficient text-to-video retrieval capabilities but also accurately locates events within the retrieved videos, addressing the shortcomings of existing methods. Extensive experimental results on the ActivityNet-Captions and Charades-STA benchmark datasets demonstrate the superior performance of our method, clearly validating its effectiveness. The innovation of this research lies in introducing a new joint framework for video-text retrieval and multi-event localization, while offering new ideas for further research and applications in related fields.", "keywords": [ "Video-Text Retrieval; Grounding; Multi-event Queries" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "dYTtGFuD3S", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Han Quan", "gender": "unknown", "institution": "China Agricultural University", "country": "CN", "position": "MS student" }, { "name": "Lang Zheng", "gender": "Male", "institution": "China Agricultural University", "country": "CN", "position": "MS student" }, { "name": "Maotao Guo", "gender": "Male", "institution": "Northeast University at Qinhuangdao Campus", "country": "CN", "position": "Undergrad student" }, { "name": "Ruxue Xing", "gender": "Female", "institution": "China Agricultural University", "country": "CN", "position": "MS student" }, { "name": "Zhenghan Chen", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "youhuan yang", "gender": "Male", "institution": "Zhengzhou University", "country": "CN", "position": "Researcher" }, { "name": "Yi Ge", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Research Assistant" } ], "n_formula": 39, "n_formula_1": 36, "n_ref_uni": 13, "n_ref": 18, "n_ref_all": 20, "n_fig": 0, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2012, "formula_len_all_1": 1787, "len_all": 82522, "len_all_1": 57117, "len_abs": 1935, "len_title": 128, "len_sents": 26558, "len_sents_1": 25085, "n_sents": 216, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1951, "title": "Adaptive Drug Interaction Prediction via Enhanced Graph Representation Learning", "abs": "This paper presents a groundbreaking theoretical framework for drug-drug interaction (DDI) prediction that seamlessly integrates domain adaptation (DA) techniques with advanced mathematical concepts. We introduce GraphPharmNet, a novel architecture that operates on DDI-DA bundles, leveraging gauge-equivariant geometric deep learning to capture the intricate structure of drug interactions across domains. Our approach reformulates the DDI prediction problem using the language of differential geometry, optimal transport, and symplectic geometry, viewing domain adaptation as a Hamiltonian flow on a statistical manifold. We develop a cohomological interpretation of domain invariance, characterizing robust DDI prediction features through the lens of persistent homology and sheaf theory. The domain adaptation process is analyzed using a geometric renormalization group framework, revealing a profound connection between the DDI-DA bundle's geometry and the emergence of domain-invariant predictive features. We further elucidate the spectral properties of the DDI-DA Laplacian, providing insights into the topological stability of domain adaptation in DDI prediction. Extensive experiments on benchmark datasets demonstrate that GraphPharmNet significantly outperforms existing methods, particularly in scenarios with limited data or when transferring knowledge across disparate domains. Our results highlight the power of this unified mathematical framework in capturing complex drug interactions and adapting to new domains, paving the way for more accurate, robust, and interpretable DDI prediction models. This work not only advances the field of computational drug discovery but also establishes a rigorous theoretical foundation for domain adaptation in graph-structured data, with potential applications across a wide range of scientific disciplines. Our anonymous github link: \\textbf{https://anonymous.4open.science/r/GraphPharmNet-C9D9}", "keywords": [ "Domain-Aligned\uff0cTransfer Learning\uff0cDrug-Target Interaction" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 3, 1, 1 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "dYTjB86pcT", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Ayush Sekhari", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Karthik Sridharan", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Assistant Professor" }, { "name": "Linda Lu", "gender": "unknown", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "PhD student" } ], "n_formula": 77, "n_formula_1": 29, "n_ref_uni": 30, "n_ref": 61, "n_ref_all": 63, "n_fig": 1, "n_tab": 2, "L_tab": 609, "n_element_tab": 58, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10193, "formula_len_all_1": 2403, "len_all": 165173, "len_all_1": 73692, "len_abs": 1846, "len_title": 109, "len_sents": 52764, "len_sents_1": 34133, "n_sents": 450, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1852, "title": "System Aware Unlearning Algorithms: Use Lesser, Forget Faster", "abs": "Machine unlearning aims to provide privacy guarantees to users when they request deletion, such that an attacker who can compromise the system post-unlearning cannot recover private information about the deleted individuals. Previously proposed definitions of unlearning require the unlearning algorithm to exactly or approximately recover the hypothesis obtained by retraining-from-scratch on the remaining samples. While this definition has been the gold standard in machine unlearning, unfortunately, because it is designed for the worst-case attacker (that can recover the updated hypothesis and the remaining dataset), developing rigorous, and memory or compute-efficient unlearning algorithms that satisfy this definition has been challenging. In this work, we propose a new definition of unlearning, called system aware unlearning, that takes into account the information that an attacker could recover by compromising the system (post-unlearning). We prove that system-aware unlearning generalizes commonly referred to definitions of unlearning by restricting what the attacker knows, and furthermore, may be easier to satisfy in scenarios where the system-information available to the attacker is limited, e.g. because the learning algorithm did not use the entire training dataset to begin with. Towards that end, we develop an exact system-aware-unlearning algorithm that is both memory and computation-time efficient for function classes that can be learned via sample compression. We then present an improvement over this for the special case of learning linear classifiers by using selective sampling for data compression, thus giving the first memory and time-efficient exact unlearning algorithm for linear classification. We analyze the tradeoffs between deletion capacity, accuracy, memory, and computation time for these algorithms.", "keywords": [ "machine unlearning", "learning theory", "selective sampling for unlearning" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "dXCpPgjTtd", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Julian McAuley", "gender": "Male", "institution": "University of California, San Diego, University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Ruihan Wu", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Postdoc" }, { "name": "Xiusi Chen", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Postdoc" }, { "name": "Yu Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Zexue He", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 14, "n_ref_uni": 29, "n_ref": 85, "n_ref_all": 127, "n_fig": 2, "n_tab": 10, "L_tab": 4790, "n_element_tab": 607, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1322, "n_element_tab_1": 166, "formula_len_all": 1614, "formula_len_all_1": 830, "len_all": 147139, "len_all_1": 69988, "len_abs": 1290, "len_title": 77, "len_sents": 42516, "len_sents_1": 31340, "n_sents": 317, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 29, "L_abs": 1296, "title": "Large Scale Knowledge Washing", "abs": "Large language models show impressive abilities in memorizing world knowledge, which leads to concerns regarding memorization of private information, toxic or sensitive knowledge, and copyrighted content. We introduce the problem of Large Scale Knowledge Washing, focusing on unlearning an extensive amount of factual knowledge. Previous unlearning methods usually define the reverse loss and update the model via backpropagation, which may affect the model's fluency and reasoning ability or even destroy the model due to extensive training with the reverse loss. Existing works introduce additional data from downstream tasks to prevent the model from losing capabilities, which requires downstream task awareness. Controlling the tradeoff of unlearning existing knowledge while maintaining existing capabilities is also challenging. To this end, we propose LaW (Large Scale Washing), where we update the MLP layers in decoder-only large language models to perform knowledge washing, as inspired by model editing methods. We derive a new objective with the knowledge to be unlearned to update the weights of certain MLP layers. Experimental results demonstrate the effectiveness of LaW in forgetting target knowledge while maximally maintaining reasoning ability. The code will be open-sourced.", "keywords": [ "knowledge unlearning", "large language models" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "dWsdJAXjQD", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Jeremy Avigad", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Prasad Tetali", "gender": "Male", "institution": "Georgia Tech Research Corporation", "country": "", "position": "" }, { "name": "Riyaz Ahuja", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Sean Welleck", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 19, "n_ref": 42, "n_ref_all": 73, "n_fig": 12, "n_tab": 13, "L_tab": 5258, "n_element_tab": 520, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1368, "n_element_tab_1": 111, "formula_len_all": 680, "formula_len_all_1": 312, "len_all": 175548, "len_all_1": 64253, "len_abs": 1331, "len_title": 98, "len_sents": 52215, "len_sents_1": 31019, "n_sents": 375, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1336, "title": "ImProver: Agent-Based Automated Proof Optimization", "abs": "Large language models (LLMs) have been used to generate formal proofs of mathematical theorems in proofs assistants such as Lean.\nHowever, we often want to optimize a formal proof with respect to various criteria, depending on its downstream use.\nFor example, we may want a proof to adhere to a certain style, be declaratively structured, or concise. Having suitably optimized proofs is also important for learning tasks, especially since human-written proofs may not optimal for that purpose.\nTo this end, we study a new problem of automated proof optimization: rewriting a proof so that it is correct and optimizes for an arbitrary criterion, such as length or declarativity.\nAs a first method for automated proof optimization, we present ImProver, a large-language-model agent that rewrites proofs to optimize arbitrary user-defined metrics in Lean.\nWe find that naively applying LLMs to proof optimization falls short, and we incorporate various improvements into ImProver, such as the use of symbolic Lean context in a novel Chain-of-States technique, as well as error-correction and retrieval. We test ImProver on rewriting real-world undergraduate, competition, and research-level mathematics theorems, finding that ImProver is capable of rewriting proofs so that they are substantially shorter and more declarative in structure.", "keywords": [ "Automated Proof Optimization", "Neural Theorem Proving", "Formal Mathematics", "Lean Theorem Prover", "Proof Generation", "Large Language Models", "Symbolic Reasoning", "Interactive Theorem Proving" ], "rating_list": [ 8, 8, 5, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 2, 4 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "dWsBrgaNzU", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 1, "author_info_list": [ { "name": "Juho Kim", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" } ], "n_formula": 30, "n_formula_1": 28, "n_ref_uni": 18, "n_ref": 47, "n_ref_all": 74, "n_fig": 3, "n_tab": 9, "L_tab": 3403, "n_element_tab": 354, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 142, "n_element_tab_1": 19, "formula_len_all": 2487, "formula_len_all_1": 2304, "len_all": 123536, "len_all_1": 58102, "len_abs": 668, "len_title": 98, "len_sents": 32722, "len_sents_1": 22548, "n_sents": 230, "n_sents_1": 146, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 668, "title": "GPU-Accelerated Counterfactual Regret Minimization", "abs": "Counterfactual regret minimization is a family of algorithms of no-regret learning dynamics capable of solving large-scale imperfect information games. We propose implementing this algorithm as a series of dense and sparse matrix and vector operations, thereby making it highly parallelizable for a graphical processing unit, at a cost of higher memory usage. Our experiments show that our implementation performs up to about 401.2 times faster than OpenSpiel's Python implementation and, on an expanded set of games, up to about 203.6 times faster than OpenSpiel's C++ implementation and the speedup becomes more pronounced as the size of the game being solved grows.", "keywords": [ "counterfactual regret minimization", "extensive-form games", "graphical processing units", "imperfect information games", "Nash equilibrium" ], "rating_list": [ 5, 5, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 1, 4, 3 ] }, { "paper_id": "dWi2c9auRm", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Fuying Wang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Guosheng Yin", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Jiacheng Xu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Lequan Yu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Tsai Hor Chan", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 51, "n_formula_1": 20, "n_ref_uni": 37, "n_ref": 67, "n_ref_all": 90, "n_fig": 7, "n_tab": 1, "L_tab": 289, "n_element_tab": 29, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 773, "n_element_tab_1": 69, "formula_len_all": 3844, "formula_len_all_1": 1586, "len_all": 117504, "len_all_1": 55907, "len_abs": 1796, "len_title": 67, "len_sents": 34609, "len_sents_1": 24403, "n_sents": 303, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1813, "title": "Bayesian Nonparametric Survival Analysis via Deep Dirichlet Process", "abs": "The analysis of time-to-event data has received increasing attention in many application fields. The key challenge is that the data are mostly incomplete, with the right censoring mechanism being the most popular form. While Cox's proportional hazards assumption has shown adaptivity to traditional time-to-event datasets, challenges are observed when generalizing this assumption to modern survival analysis --- the proportional hazards assumption is often violated when covariates are high-dimensional. Moreover, traditional parametric assumptions on the survival distribution mostly belong to the exponential family and thus the assumption is strong and their exponential decay rate leads to poor long-tail approximations. To overcome these challenges, we propose a novel deep learning framework for survival analysis, named **DDPSurv**, which adopts a deeply parameterized Dirichlet process (DP) mixture model on survival distribution. Different from previous deep parametric approaches which rely on strong statistical assumptions, our framework can model the survival distribution with greater flexibility by adopting a DP mixture model. With the DP mixture model, we can improve the flexibility in modelling the survival distributions and achieve better tail behaviour by including the heavy-tail distributions in the mixture. We theoretically show that the proposed model can approximate the true survival distribution at a tight concentration rate. Empirical evaluations on standard survival benchmarks validate the satisfactory performance of the proposed method. Extensive experiments on large-scale clinical datasets --- MIMIC-III and MIMIC-IV --- highlight the scalability and clinical significance of our method. Codes are anonymously available at https://anonymous.4open.science/r/DeepSurv-net-2215", "keywords": [ "Bayesian nonparametric methods", "Survival Analysis", "Variational Inference" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 1, 1, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "dWReNWEj5b", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Aadyot Bhatnagar", "gender": "Male", "institution": "SalesForce.com", "country": "", "position": "Research Engineer" }, { "name": "Ali Madani", "gender": "unknown", "institution": "Profluent Bio", "country": "", "position": "Principal Researcher" }, { "name": "Jason Yang", "gender": "unknown", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jeffrey Ruffolo", "gender": "Male", "institution": "Profluent Bio", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 57, "n_ref": 125, "n_ref_all": 152, "n_fig": 16, "n_tab": 2, "L_tab": 1242, "n_element_tab": 148, "n_fig_1": 14, "n_tab_1": 2, "L_tab_1": 1621, "n_element_tab_1": 168, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 207710, "len_all_1": 62423, "len_abs": 1221, "len_title": 132, "len_sents": 48619, "len_sents_1": 29509, "n_sents": 345, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1226, "title": "Conditional Enzyme Generation Using Protein Language Models with Adapters", "abs": "The conditional generation of proteins with desired functions and/or properties is a key goal for generative models. Existing methods based on prompting of language models can generate proteins conditioned on a target functionality, such as a desired enzyme family. However, these methods are limited to simple, tokenized conditioning and have not been shown to generalize to unseen functions. In this study, we propose ProCALM (Protein Conditionally Adapted Language Model), an approach for the conditional generation of proteins using adapters to protein language models (PLMs). While previous methods have used adapters for structure-conditioned generation from PLMs, our implementation of ProCALM involves finetuning ProGen2 to condition generation on protein function, and it is flexible to function representations of enzyme family, taxonomy, or natural language descriptions. ProCALM matches or exceeds the performance of existing methods at conditional sequence generation from target functions. Impressively, it can also generalize to rare and unseen functions. Overall, ProCALM is a flexible and computationally efficient approach, and we expect that it can be extended to a wide range of generative language models.", "keywords": [ "protein", "enzyme", "protein language model", "generative model", "fine-tuning", "conditional generation", "adapter", "machine learning" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "dWHecekgsD", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Chuan Shi", "gender": "Male", "institution": "Beijing University of Post and Telecommunication, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Mengmei Zhang", "gender": "Female", "institution": "China Telecom Bestpay", "country": "CN", "position": "Researcher" }, { "name": "Yuan Fang", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "yibo Li", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 24, "n_ref": 52, "n_ref_all": 68, "n_fig": 6, "n_tab": 4, "L_tab": 2316, "n_element_tab": 171, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1115, "n_element_tab_1": 120, "formula_len_all": 956, "formula_len_all_1": 639, "len_all": 132288, "len_all_1": 54742, "len_abs": 1793, "len_title": 102, "len_sents": 27407, "len_sents_1": 22064, "n_sents": 220, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 1437, "title": "Towards Fine-grained Molecular Graph-Text Pre-training", "abs": "Understanding molecular structure and related knowledge is crucial for scientific research. Recent studies integrate molecular graphs with their textual descriptions to enhance molecular representation learning. However, they focus on the whole molecular graph and neglect frequently occurring subgraphs, known as motifs, which are essential for determining molecular properties. Without such fine-grained knowledge, these models struggle to generalize to unseen molecules and tasks that require motif-level insights. To bridge this gap, we propose FineMolTex, a novel Fine-grained Molecular graph-Text pre-training framework to jointly learn coarse-grained molecule-level knowledge and fine-grained motif-level knowledge. Specifically, FineMolTex consists of two pre-training tasks: a contrastive alignment task for coarse-grained matching and a masked multi-modal modeling task for fine-grained matching. In particular, the latter predicts the labels of masked motifs and words, leveraging insights from each other, thereby enabling FineMolTex to understand the fine-grained matching between motifs and words. Finally, we conduct extensive experiments across three downstream tasks, achieving up to 230% improvement in the text-based molecule editing task. Additionally, our case studies reveal that FineMolTex successfully captures fine-grained knowledge, potentially offering valuable insights for drug discovery and catalyst design.", "keywords": [ "Molecular Representation Learning", "Graph Neural Network" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "dWEDLUmddV", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Kuan Li", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Minhao Cheng", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Shuai Wang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" } ], "n_formula": 41, "n_formula_1": 22, "n_ref_uni": 34, "n_ref": 240, "n_ref_all": 260, "n_fig": 5, "n_tab": 18, "L_tab": 9641, "n_element_tab": 1082, "n_fig_1": 4, "n_tab_1": 9, "L_tab_1": 3458, "n_element_tab_1": 405, "formula_len_all": 1802, "formula_len_all_1": 985, "len_all": 243956, "len_all_1": 111179, "len_abs": 1361, "len_title": 188, "len_sents": 66718, "len_sents_1": 46567, "n_sents": 520, "n_sents_1": 347, "venue": "ICLR2025", "decision": "Reject", "L_title": 140, "L_abs": 1366, "title": "Enhancing Dataset Distillation with Concurrent Learning: Addressing Negative Correlations and Catastrophic Forgetting in Trajectory Matching", "abs": "Dataset distillation generates a small synthetic dataset on which a model is trained to achieve performance comparable to that obtained on a complete dataset. Current state-of-the-art methods primarily focus on Trajectory Matching (TM), which optimizes the synthetic dataset by matching its training trajectory with that from the real dataset. Due to convergence issues and numerical stability, it is impractical to match the entire trajectory in one go; typically, a segment is sampled for matching at each iteration. However, previous TM-based methods overlook the potential interactions between matching different segments, particularly the presence of negative correlations. To study this problem, we conduct a quantitative analysis of the correlation between matching different segments and discover varying degrees of negative correlation depending on the image per class (IPC). Such negative correlation could lead to an increase in accumulated trajectory error and transform trajectory matching into a continual learning paradigm, potentially causing catastrophic forgetting. To tackle this issue, we propose a concurrent learning-based trajectory matching that simultaneously matches multiple segments. Extensive experiments demonstrate that our method consistently surpasses previous TM-based methods on CIFAR-10, CIFAR-100, Tiny ImageNet, and ImageNet-1K.", "keywords": [ "Dataset Distillation; Efficient Machine Learning; Data-centric AI" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "dVrYcscgLu", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Alessandro G Allievi", "gender": "unknown", "institution": "University of Texas at Austin", "country": "", "position": "Postdoc" }, { "name": "Andrew Levy", "gender": "Male", "institution": "Brown University", "country": "", "position": "PhD Student" }, { "name": "George Konidaris", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 7, "n_ref_uni": 33, "n_ref": 67, "n_ref_all": 78, "n_fig": 7, "n_tab": 5, "L_tab": 1847, "n_element_tab": 77, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 263, "n_element_tab_1": 16, "formula_len_all": 1602, "formula_len_all_1": 718, "len_all": 168402, "len_all_1": 67799, "len_abs": 1230, "len_title": 120, "len_sents": 65430, "len_sents_1": 37590, "n_sents": 352, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1239, "title": "Latent-Predictive Empowerment: Measuring Empowerment without a Simulator", "abs": "Empowerment has the potential to help agents learn large skillsets, but is not yet a scalable solution for training general-purpose agents. Recent empowerment methods learn large skillsets by maximizing the mutual information between skills and states, but these approaches require a model of the transition dynamics, which can be challenging to learn in realistic settings with high-dimensional and stochastic observations. We present an algorithm, Latent-Predictive Empowerment (LPE), that can compute empowerment in a more scalable manner. LPE learns large skillsets by maximizing an objective that under certain conditions has the same optimal skillset as the mutual information between skills and states, but our objective is more tractable to optimize because it only requires learning a simpler latent-predictive model rather than a full simulator of the environment. We show empirically in a variety of settings, includes ones with high-dimensional observations and highly stochastic transition dynamics, that our empowerment objective learns similar-sized skillsets as the leading empowerment algorithm, which assumes access to a model of the transition dynamics, and outperforms other model-based approaches to empowerment.", "keywords": [ "Empowerment", "Unsupervised Skill Learning", "Unsupervised Reinforcement Learning", "Self-supervised Reinforcement Learning" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "dVLrqe2a7c", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Meir Yossef Levi", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "PhD student" }, { "name": "Ronen Schaffer", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "MS student" }, { "name": "Tomer Michaeli", "gender": "Male", "institution": "Technion, Technion", "country": "IL", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 8, "n_ref_uni": 53, "n_ref": 105, "n_ref_all": 125, "n_fig": 11, "n_tab": 6, "L_tab": 979, "n_element_tab": 121, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 491, "n_element_tab_1": 62, "formula_len_all": 662, "formula_len_all_1": 367, "len_all": 146195, "len_all_1": 63120, "len_abs": 1322, "len_title": 131, "len_sents": 39129, "len_sents_1": 29948, "n_sents": 325, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1221, "title": "From Attention to Prediction Maps: Per-Class Gradient-Free Transformer Explanations", "abs": "The Vision Transformer (ViT) has become a standard model architecture in computer vision, especially for classification tasks. As such, explaining ViT predictions has attracted significant research efforts in recent years. Many methods rely on attention maps, which highlight \\emph{where} in the image the network directs its attention. In this paper, we introduce Prediction~Maps -- a novel explanation method that complements attention maps by revealing \\emph{what} the network sees. Prediction maps visualize how each patch token within a given layer is associated with each possible class. This is done by utilizing the classification head at the output of the network, originally trained to be fed with the class token at the last layer. Specifically, to obtain the prediction map of a particular layer, we apply the classification head to every patch token within that layer. We show that prediction maps provide complementary information to attention maps and illustrate that combining them leads to state-of-the-art explainability performance. Furthermore, since our proposed method is neither gradient- nor perturbation-based, it offers superior computational and memory efficiency compared to competing methods.", "keywords": [ "explainable ai", "vision transformer" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 4, 1, 2, 3 ], "contribution_list": [ 3, 2, 2, 4 ], "confidence_list": [ 3, 5, 4, 2 ] }, { "paper_id": "dUiMNQHYXG", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Damiano Verda", "gender": "unknown", "institution": "Rulex Innovations Labs", "country": "", "position": "Researcher" }, { "name": "Elvio G. Amparore", "gender": "Male", "institution": "University of Turin", "country": "IT", "position": "Assistant Professor" }, { "name": "Enrico Ferrari", "gender": "Male", "institution": "Rulex Innovations Labs", "country": "IT", "position": "Researcher" }, { "name": "Muhammad Rashid", "gender": "Male", "institution": "University of Turin", "country": "IT", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 10, "n_ref_uni": 25, "n_ref": 51, "n_ref_all": 126, "n_fig": 21, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 323, "n_element_tab_1": 10, "formula_len_all": 2367, "formula_len_all_1": 644, "len_all": 143544, "len_all_1": 61819, "len_abs": 1204, "len_title": 65, "len_sents": 51232, "len_sents_1": 28245, "n_sents": 416, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1074, "title": "Shapley Image Explanations With Data-Aware Binary Partition Trees", "abs": "Extracting a visual interpretation of a learned representation of a machine learning model applied to image data is a relevant task in eXplainable AI (XAI). \nPixel-level feature attributions serve as a valuable tool in this context, as they identify the regions within an image responsible for the classification outcome.\nThe hierarchical Owen approximation of the Shapley values has proved to be an effective strategy for this task. \nHowever, existing approaches lack data-awareness, leading to poor alignment between the pixel-level attributions and the actual morphological features of the classified image.\n\nThis paper introduces ShapBPT, a novel XAI method that computes the Owen approximation of the Shapley coefficients following a data-aware binary hierarchical coalition structure derived from the Binary Partition Tree computer vision algorithm. \nBy aligning with the morphological features of the image, the proposed method significantly enhances the identification of relevant image regions.\nExperimental results confirm the effectiveness of the proposed method.", "keywords": [ "Shapley coefficients", "Owen approximation", "Binary Partition Tree", "eXplainable AI" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 4, 1, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "dUCMO9lwSv", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Giulio Corallo", "gender": "Male", "institution": "Eurecom", "country": "", "position": "PhD student" }, { "name": "Giulio Franzese", "gender": "Male", "institution": "Eurecom", "country": "FR", "position": "Assistant Professor" }, { "name": "Mattia Martini", "gender": "Male", "institution": "Universit\u00e9 de Nice-Sophia Antipolis", "country": "FR", "position": "Postdoc" }, { "name": "Paolo Papotti", "gender": "Male", "institution": "Eurecom", "country": "", "position": "Associate Professor" }, { "name": "Pietro Michiardi", "gender": "Male", "institution": "EURECOM", "country": "", "position": "Full Professor" } ], "n_formula": 100, "n_formula_1": 41, "n_ref_uni": 51, "n_ref": 97, "n_ref_all": 160, "n_fig": 7, "n_tab": 4, "L_tab": 198, "n_element_tab": 15, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 413, "n_element_tab_1": 6, "formula_len_all": 10017, "formula_len_all_1": 2932, "len_all": 183789, "len_all_1": 70010, "len_abs": 1038, "len_title": 98, "len_sents": 52999, "len_sents_1": 31567, "n_sents": 418, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1056, "title": "Latent Abstractions in Generative Diffusion Models", "abs": "In this work we study how diffusion-based generative models produce high-dimensional data, such as an image, by implicitly relying on a manifestation of a low-dimensional set of latent abstractions, that guide the generative process.\nWe present a novel theoretical framework that extends Nonlinear Filtering (NLF), and that offers a unique perspective on SDE-based generative models. The development of our theory relies on NLF, including a novel formulation of the joint (state and measurement) dynamics, and an information-theoretic measure of the influence of the system state on the measurement process. According to our theory, diffusion models can be cast as a system of SDE, describing a non-linear filter in which the evolution of unobservable latent abstractions steers the dynamics of an observable measurement process (corresponding to the generative pathways). In addition, we present an empirical study to validate our theory and previous empirical results on the emergence of latent abstractions at different stages of the generative process.", "keywords": [ "diffusion models", "world modelling", "information theory", "nonlinear filtering" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 2, 2 ] }, { "paper_id": "dTkqaCKLPp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Alexandre Allauzen", "gender": "Male", "institution": "Ecole Sup\u00e9rieure de Physique et de Chimie Industrielles", "country": "FR", "position": "Full Professor" }, { "name": "Florian Le Bronnec", "gender": "Male", "institution": "Universit\u00e9 Paris-Dauphine (Paris IX)", "country": "FR", "position": "PhD student" }, { "name": "Laure Soulier", "gender": "unknown", "institution": "Sorbonne Universit\u00e9, CNRS, ISIR", "country": "FR", "position": "Associate Professor" }, { "name": "Song Duong", "gender": "Male", "institution": "Criteo", "country": "FR", "position": "PhD student" }, { "name": "Vincent Guigue", "gender": "Male", "institution": "AgroParisTech", "country": "", "position": "Full Professor" }, { "name": "Alberto Lumbreras Carrasco", "gender": "Male", "institution": "Criteo", "country": "", "position": "Research Scientist" }, { "name": "patrick gallinari", "gender": "Male", "institution": "Criteo AI Lab", "country": "FR", "position": "Research Scientist" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 51, "n_ref": 77, "n_ref_all": 119, "n_fig": 10, "n_tab": 17, "L_tab": 4282, "n_element_tab": 783, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1328, "n_element_tab_1": 175, "formula_len_all": 286, "formula_len_all_1": 203, "len_all": 240049, "len_all_1": 69112, "len_abs": 1636, "len_title": 94, "len_sents": 61804, "len_sents_1": 30909, "n_sents": 557, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1278, "title": "SCOPE: A Self-supervised framework for Improving Faithfulness in Conditional Text Generation", "abs": "Large Language Models (LLMs), when used for conditional text generation, often produce hallucinations, i.e., information that is unfaithful or not grounded in the input context. This issue arises in typical conditional text generation tasks, such as text summarization and data-to-text generation, where the goal is to produce fluent text based on contextual input. When fine-tuned on specific domains, LLMs struggle to provide faithful answers to a given context, often adding information or generating errors. One underlying cause of this issue is that LLMs rely on statistical patterns learned from their training data. This reliance can interfere with the model\u2019s ability to stay faithful to a provided context, leading to the generation of ungrounded information. We build upon this observation and introduce a novel self-supervised method for generating a training set of unfaithful samples. We then refine the model using a training process that encourages the generation of grounded outputs over unfaithful ones, drawing on preference-based training. Our approach leads to significantly more grounded text generation, outperforming existing self-supervised techniques in faithfulness, as evaluated through automatic metrics, LLM-based assessments, and human evaluations.", "keywords": [ "faithfulness", "hallucination", "conditional text generation", "natural language processing", "large language models" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 2, 4 ] }, { "paper_id": "dTQmayPKMs", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Haeone Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Intern" }, { "name": "Hanho Ryu", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Undergrad student" }, { "name": "Kimin Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Taywon Min", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Yongchan Kwon", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 19, "n_formula_1": 8, "n_ref_uni": 35, "n_ref": 88, "n_ref_all": 120, "n_fig": 16, "n_tab": 9, "L_tab": 3023, "n_element_tab": 208, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 3303, "n_element_tab_1": 67, "formula_len_all": 1070, "formula_len_all_1": 568, "len_all": 204430, "len_all_1": 72145, "len_abs": 1221, "len_title": 110, "len_sents": 72139, "len_sents_1": 32450, "n_sents": 528, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1225, "title": "Understanding Impact of Human Feedback via Influence Functions", "abs": "In Reinforcement Learning from Human Feedback (RLHF), it is crucial to learn suitable reward models from human feedback to align large language models (LLMs) with human intentions. However, human feedback can often be noisy, inconsistent, or biased, especially when evaluating complex responses. Such feedback can lead to misaligned reward signals, potentially causing unintended side effects during the RLHF process. To address these challenges, we explore the use of influence functions to measure the impact of human feedback on the performance of reward models. We propose a compute-efficient approximation method that enables the application of influence functions to LLM-based reward models and large-scale preference datasets. In our experiments, we demonstrate two key applications of influence functions: (1) detecting common forms of labeler bias in human feedback datasets and (2) guiding labelers to refine their strategies to align more closely with expert feedback. By quantifying the impact of human feedback on reward models, we believe that influence functions can enhance feedback interpretability and contribute to scalable oversight in RLHF, helping labelers provide more accurate and consistent feedback.", "keywords": [ "reinforcement learning from human feedback", "influence function", "reward learning", "alignment", "scalable oversight" ], "rating_list": [ 5, 8, 6, 5, 8 ], "soundness_list": [ 3, 4, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 1, 3 ], "confidence_list": [ 4, 2, 2, 3, 5 ] }, { "paper_id": "dTPz4rEDok", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Carolin Schmidt", "gender": "unknown", "institution": "Technical University of Denmark", "country": "", "position": "PhD student" }, { "name": "Daniele Gammelli", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Postdoc" }, { "name": "Filipe Rodrigues", "gender": "Male", "institution": "Technical University of Denmark", "country": "DK", "position": "Associate Professor" }, { "name": "James Harrison", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Marco Pavone", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Director, Autonomous Vehicle Research" } ], "n_formula": 40, "n_formula_1": 9, "n_ref_uni": 53, "n_ref": 95, "n_ref_all": 133, "n_fig": 11, "n_tab": 14, "L_tab": 4321, "n_element_tab": 405, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 2123, "n_element_tab_1": 175, "formula_len_all": 2260, "formula_len_all_1": 412, "len_all": 225798, "len_all_1": 73144, "len_abs": 1288, "len_title": 116, "len_sents": 72568, "len_sents_1": 35357, "n_sents": 519, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1319, "title": "Offline Hierarchical Reinforcement Learning via Inverse Optimization", "abs": "Hierarchical policies enable strong performance in many sequential decision-making problems, such as those with high-dimensional action spaces, those requiring long-horizon planning, and settings with sparse rewards. \nHowever, learning hierarchical policies from static offline datasets presents a significant challenge.\nCrucially, actions taken by higher-level policies may not be directly observable within hierarchical controllers, and the offline dataset might have been generated using a different policy structure, hindering the use of standard offline learning algorithms.\nIn this work, we propose $\\textit{OHIO}$: a framework for offline reinforcement learning (RL) of hierarchical policies. \nOur framework leverages knowledge of the policy structure to solve the $\\textit{inverse problem}$, recovering the unobservable high-level actions that likely generated the observed data under our hierarchical policy.\nThis approach constructs a dataset suitable for off-the-shelf offline training.\nWe demonstrate our framework on robotic and network optimization problems and show that it substantially outperforms end-to-end RL methods and improves robustness. \nWe investigate a variety of instantiations of our framework, both in direct deployment of policies trained offline and when online fine-tuning is performed.", "keywords": [ "Offline Reinforcement Learning", "Hierarchical Reinforcement Learning" ], "rating_list": [ 5, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "dTGH9vUVdf", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Lue Fan", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Qitai Wang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yuntao Chen", "gender": "Male", "institution": "Centre for Artificial Intelligence and Robotics (CAIR), Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "HK", "position": "Assistant Professor" }, { "name": "Yuqi Wang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zhaoxiang Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 64, "n_ref": 103, "n_ref_all": 130, "n_fig": 9, "n_tab": 9, "L_tab": 4118, "n_element_tab": 325, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1184, "n_element_tab_1": 145, "formula_len_all": 47, "formula_len_all_1": 47, "len_all": 184607, "len_all_1": 66187, "len_abs": 2052, "len_title": 108, "len_sents": 42687, "len_sents_1": 30739, "n_sents": 327, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1440, "title": "FreeVS: Generative View Synthesis on Free Driving Trajectory", "abs": "Existing reconstruction-based novel view synthesis methods for driving scenes focus on synthesizing camera views along the recorded trajectory of the ego vehicle. \nTheir image rendering performance will severely degrade on viewpoints falling out of the recorded trajectory, where camera rays are untrained.\nWe propose FreeVS, a novel fully generative approach that can synthesize camera views on free new trajectories in real driving scenes. \nTo control the generation results to be 3D consistent with the real scenes and accurate in viewpoint pose, we propose the pseudo-image representation of view priors to control the generation process.\nViewpoint translation simulation is applied on pseudo-images to simulate camera movement in each direction.\nOnce trained, FreeVS can be applied to any validation sequences without reconstruction process and synthesis views on novel trajectories.\nMoreover, we propose two new challenging benchmarks tailored to driving scenes, which are novel camera synthesis and novel trajectory synthesis, emphasizing the freedom of viewpoints.\nGiven that no ground truth images are available on novel trajectories, we also propose to evaluate the consistency of images synthesized on novel trajectories with 3D perception models.\nExperiments on the Waymo Open Dataset show that FreeVS has a strong image synthesis performance on both the recorded trajectories and novel trajectories. \nThe code will be released.", "keywords": [ "Novel View Synthesis", "Driving Scene", "Free Trajectory", "Image Generation" ], "rating_list": [ 3, 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 4, 5, 3, 4, 3 ] }, { "paper_id": "dSuC2qFXqB", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Xinyu Yuan", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "MS student" }, { "name": "Yan Qiao", "gender": "Female", "institution": "Hefei University of Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 46, "n_formula_1": 8, "n_ref_uni": 47, "n_ref": 102, "n_ref_all": 149, "n_fig": 23, "n_tab": 5, "L_tab": 2910, "n_element_tab": 277, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4324, "formula_len_all_1": 759, "len_all": 193896, "len_all_1": 63839, "len_abs": 1372, "len_title": 121, "len_sents": 61471, "len_sents_1": 30829, "n_sents": 518, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1209, "title": "Toward Practical Learning-based Frequency Estimation without Ground Truth", "abs": "Estimating the frequency of items on the high-volume, fast data stream has been extensively studied in many areas, such as database and network measurement. Traditional sketch algorithms only allow to give very rough estimates with limited memory cost, whereas some learning-augmented algorithms have been proposed recently, their offline framework requires actual frequencies that are challenging to access in general for training, and speed is too slow for real-time processing, despite the still coarse-grained accuracy. To this end, we propose a more practical learning-based estimation framework namely UCL-sketch, by following the line of equation-based sketch to estimate per-key frequencies. In a nutshell, there are two key techniques: online training via equivalent learning without ground truth, and highly scalable architecture with logical estimation buckets. We implemented experiments on both real-world and synthetic datasets. The results demonstrate that our method greatly outperforms existing state-of-the-art sketches regarding per-key accuracy and distribution, while preserving resource efficiency. Our code is attached in the supplementary material, and will be made publicly available.", "keywords": [ "sketching algorithms", "ML for streaming data", "frequency estimation" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "dSneEp59yX", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Heejun Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Myeongjae Jeon", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Sung Ju Hwang", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Youngwan Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jeffrey Ryan Willette", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 18, "n_ref": 43, "n_ref_all": 107, "n_fig": 27, "n_tab": 10, "L_tab": 3376, "n_element_tab": 355, "n_fig_1": 21, "n_tab_1": 8, "L_tab_1": 1425, "n_element_tab_1": 164, "formula_len_all": 563, "formula_len_all_1": 482, "len_all": 121424, "len_all_1": 67023, "len_abs": 1631, "len_title": 114, "len_sents": 40131, "len_sents_1": 29591, "n_sents": 306, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1545, "title": "Training Free Exponential Context Extension via Cascading KV Cache", "abs": "The transformer's context window is vital for tasks such as few-shot learning and conditional generation as it preserves previous tokens for active memory. However, as the context lengths increase, the computational costs grow quadratically, hindering the deployment of large language models (LLMs) in real-world, long sequence scenarios. Although some recent key-value caching (KV Cache) methods offer linear inference complexity, they naively manage the stored context, prematurely evicting tokens and losing valuable information. Moreover, they lack an optimized prefill/prompt stage strategy, resulting in higher latency than even quadratic attention for realistic context sizes. In response, we introduce a novel mechanism that leverages cascading sub-cache buffers to selectively retain the most relevant tokens, enabling the model to maintain longer context histories without increasing the cache size. Our approach outperforms linear caching baselines across key benchmarks, including streaming perplexity, question answering, book summarization, and passkey retrieval, where it retains better retrieval accuracy at 1M tokens after four doublings of the cache size of 65K. Additionally, our method reduces prefill stage latency by a factor of 6.8 when compared to flash attention on 1M tokens. These innovations not only enhance the computational efficiency of LLMs but also pave the way for their effective deployment in resource-constrained environments, enabling large-scale, real-time applications with significantly reduced latency.", "keywords": [ "transformer", "efficiency", "linear attention" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "dSjCFKiEdK", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Fei Kou", "gender": "Male", "institution": "vivo mobile communication", "country": "CN", "position": "Researcher" }, { "name": "Hong Gu", "gender": "Male", "institution": "Hangzhou VIVO Information Technology Co., Ltd", "country": "CN", "position": "Director" }, { "name": "Mingrui Zhu", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Nannan Wang", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Pengcheng Xu", "gender": "Male", "institution": "University of Western Ontario", "country": "CA", "position": "PhD student" }, { "name": "Qingnan Fan", "gender": "Male", "institution": "VIVO", "country": "CN", "position": "Lead Researcher" }, { "name": "Ruoyu Zhao", "gender": "Male", "institution": "Xidian University", "country": "", "position": "MS student" }, { "name": "Wei WU", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xinbo Gao", "gender": "Male", "institution": "Chongqing University of Post and Telecommunications", "country": "CN", "position": "Full Professor" }, { "name": "shuai qin", "gender": "Male", "institution": "vivo", "country": "CN", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 10, "n_ref_uni": 25, "n_ref": 50, "n_ref_all": 79, "n_fig": 13, "n_tab": 5, "L_tab": 2106, "n_element_tab": 284, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 913, "n_element_tab_1": 123, "formula_len_all": 550, "formula_len_all_1": 691, "len_all": 154792, "len_all_1": 57339, "len_abs": 314, "len_title": 174, "len_sents": 41497, "len_sents_1": 27309, "n_sents": 299, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1130, "title": "InstructBrush: Learning Attention-based Visual Instruction for Image Editing", "abs": "Diffusion-based image editing methods have garnered significant attention in image editing. However, despite encompassing a wide range of editing priors, these methods are helpless when handling editing tasks that are challenging for users to accurately describe. We propose InstructBrush, an inversion method for instruction-based image editing methods to bridge this gap. It extracts editing effects from example image pairs as editing instructions to guide the editing of new images. Two key techniques are introduced into InstructBrush, Attention-based Instruction Optimization and Transformation-oriented Instruction Initialization, to address the limitations of the previous method in terms of inversion effects and instruction generalization. To explore the ability of visual prompt editing methods to guide image editing in open scenarios, we establish a Transformation-Oriented Paired Benchmark (TOP-Bench). Quantitatively and qualitatively, our approach achieves superior performance in editing and is more semantically consistent with the target editing effects. The code and benchmark will be released upon acceptance.", "keywords": [ "Image Editing.+Visual In-Context Learning.+Diffusion Models" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "dSTSl6QK5m", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Omesh Tickoo", "gender": "Male", "institution": "Intel", "country": "US", "position": "Principal Researcher" }, { "name": "Piyush Khanna", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Ranganath Krishnan", "gender": "Male", "institution": "Intel Corporation", "country": "US", "position": "Research Scientist" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 62, "n_ref": 110, "n_ref_all": 131, "n_fig": 7, "n_tab": 6, "L_tab": 4220, "n_element_tab": 562, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 3140, "n_element_tab_1": 445, "formula_len_all": 474, "formula_len_all_1": 239, "len_all": 167194, "len_all_1": 64307, "len_abs": 1463, "len_title": 123, "len_sents": 45059, "len_sents_1": 28575, "n_sents": 258, "n_sents_1": 154, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1468, "title": "Enhancing Trust in Large Language Models with Uncertainty-Aware Fine-Tuning", "abs": "Large language models (LLMs) have revolutionized the field of natural language processing with their impressive reasoning and question-answering capabilities. However, these models are sometimes prone to generating credible-sounding but incorrect information, a phenomenon known as LLM hallucinations. Reliable uncertainty estimation in LLMs is essential for fostering trust in their generated responses and serves as a critical tool for the detection and prevention of erroneous or hallucinated outputs. To achieve reliable and well-calibrated uncertainty quantification in open-ended and free-form natural language generation, we propose an uncertainty-aware fine-tuning approach for LLMs. This approach enhances the model's ability to provide reliable uncertainty estimates without compromising accuracy, thereby guiding them to produce more trustworthy responses. We introduce a novel uncertainty-aware causal language modeling loss function, grounded in the principles of decision theory. Through rigorous evaluation on multiple free-form question-answering datasets and models, we demonstrate that our uncertainty-aware fine-tuning approach yields better calibrated uncertainty estimates in natural language generation tasks than fine-tuning with the standard causal language modeling loss. Furthermore, the experimental results show that the proposed method significantly improves the model's ability to detect hallucinations and identify out-of-domain prompts.", "keywords": [ "Uncertainty quantification", "Trustworthiness", "Natural Language Generation", "Large Language Models (LLMs)", "Uncertainty-aware Fine-tuning" ], "rating_list": [ 6, 6, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 4, 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4, 3 ] }, { "paper_id": "dSQtMx6dPE", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Adam Dziedzic", "gender": "Male", "institution": "CISPA, saarland university, saarland informatics campus", "country": "DE", "position": "Assistant Professor" }, { "name": "Franziska Boenisch", "gender": "Female", "institution": "CISPA", "country": "DE", "position": "Assistant Professor" }, { "name": "Jing Xu", "gender": "Female", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Postdoc" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 48, "n_ref": 120, "n_ref_all": 158, "n_fig": 11, "n_tab": 10, "L_tab": 2588, "n_element_tab": 337, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 979, "n_element_tab_1": 99, "formula_len_all": 748, "formula_len_all_1": 331, "len_all": 176559, "len_all_1": 73803, "len_abs": 2065, "len_title": 100, "len_sents": 48509, "len_sents_1": 34639, "n_sents": 374, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1267, "title": "DP-GPL: Differentially Private Graph Prompt Learning", "abs": "Graph Neural Networks (GNNs) have shown remarkable performance in various applications. Recently, graph prompt learning has emerged as a powerful GNN training paradigm, inspired by advances in language and vision models. Here, a GNN is pre-trained on public data and then adapted to sensitive tasks using lightweight graph prompts. However, using prompts from sensitive data poses privacy risks.\nIn this work, we are the first to investigate these risks in graph prompts by instantiating a membership inference attack that reveals significant privacy leakage. We also find that the standard privacy method, DP-SGD, fails to provide practical privacy-utility trade-offs in graph prompt learning, likely due to the small number of sensitive data points used to learn the prompts.\nAs a solution, we propose two algorithms, DP-GPL and DP-GPL+W, for differentially private graph prompt learning based on the PATE framework, that generate a graph prompt with differential privacy guarantees.\nOur evaluation across various graph prompt learning methods, GNN architectures, and pre-training strategies demonstrates that our algorithms achieve high utility at strong privacy, effectively mitigating privacy concerns while preserving the powerful capabilities of prompted GNNs.", "keywords": [ "Graph Prompt Learning; Membership Inference Attack; Differential Privacy" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "dRz3cizftU", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Jiannan Cao", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "MS student" }, { "name": "Jianwei Yin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Sheng Cheng", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Researcher" }, { "name": "Shi Bo", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "PhD student" }, { "name": "Tianyu Du", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xuhong Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xun Wang", "gender": "Male", "institution": "Northeast Forest University", "country": "CN", "position": "MS student" }, { "name": "Yanming Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yuwei Zhang", "gender": "Female", "institution": "Tongji University", "country": "CN", "position": "MS student" }, { "name": "xinyue peng", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 41, "n_ref": 89, "n_ref_all": 106, "n_fig": 4, "n_tab": 7, "L_tab": 1694, "n_element_tab": 181, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1129, "n_element_tab_1": 244, "formula_len_all": 644, "formula_len_all_1": 642, "len_all": 307882, "len_all_1": 62811, "len_abs": 1322, "len_title": 111, "len_sents": 102666, "len_sents_1": 27984, "n_sents": 892, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1330, "title": "Tool-Planner: Task Planning with Clusters across Multiple Tools", "abs": "Large language models (LLMs) have demonstrated exceptional reasoning capabilities, enabling them to solve various complex problems. Recently, this ability has been applied to the paradigm of tool learning. Tool learning involves providing examples of tool usage and their corresponding functions, allowing LLMs to formulate plans and demonstrate the process of invoking and executing each tool. LLMs can address tasks that they cannot complete independently, thereby enhancing their potential across different tasks. However, this approach faces two key challenges. First, redundant error correction leads to unstable planning and long execution time. Additionally, designing a correct plan among multiple tools is also a challenge in tool learning. To address these issues, we propose Tool-Planner, a task-processing framework based on toolkits. Tool-Planner groups tools based on the API functions with the same function into a toolkit and allows LLMs to implement planning across the various toolkits. When a tool error occurs, the language model can reselect and adjust tools based on the toolkit. Experiments show that our approach demonstrates a high pass and win rate across different datasets and optimizes the planning scheme for tool learning in models such as GPT-4 and Claude 3, showcasing the potential of our method.", "keywords": [ "Tool Learning", "Task Planning", "LLM agent" ], "rating_list": [ 6, 5, 8, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "dRdjTNb5eN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Boqun Kou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Fei Huang", "gender": "Male", "institution": "Alibaba Group US", "country": "US", "position": "Senior Research Director" }, { "name": "Ji Zhang", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Senior Staff Engineer" }, { "name": "Ming Yan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Instructor" }, { "name": "Peng Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yang Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Zijun Liu", "gender": "Male", "institution": "Department of Computer Science and Technology, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 10, "n_ref_uni": 41, "n_ref": 116, "n_ref_all": 173, "n_fig": 10, "n_tab": 12, "L_tab": 3708, "n_element_tab": 425, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2141, "n_element_tab_1": 222, "formula_len_all": 1347, "formula_len_all_1": 366, "len_all": 272576, "len_all_1": 66979, "len_abs": 1557, "len_title": 113, "len_sents": 67273, "len_sents_1": 29735, "n_sents": 516, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1575, "title": "Enabling Weak LLMs to Judge Response Reliability via Meta Ranking", "abs": "Despite the strong performance of large language models (LLMs) across a wide range of tasks, they still have reliability issues. Previous studies indicate that strong LLMs like GPT-4-turbo excel in evaluating the reliability of responses from LLMs, but face efficiency and local deployment issues. Thus, to enable weak LLMs to effectively assess the reliability of LLM responses, we propose a novel cross-query-comparison-based method called $\\textit{Meta Ranking}$ (MR). Unlike previous few-shot methods that solely based on in-context learning capabilities in LLMs, MR assesses reliability by pairwise ranking the target query-response pair with multiple reference query-response pairs. We found that MR is highly effective in error detection for LLM responses, that MR with weaker LLMs, which have lower task performance, results in higher judgement precision against baselines with the same or even stronger models. Moreover, the method requires as few as five reference samples and significantly improving efficiency. \nWe further demonstrate that MR can enhance strong LLMs' performance in two practical applications: model cascading and instruction tuning. In model cascading, we combine open- and closed-source LLMs to achieve performance comparable to GPT-4-turbo with lower costs. In instruction tuning, we use MR for iterative training data filtering, significantly reducing data processing time and enabling LLaMA-7B and Phi-2 to surpass 13B models with fewer training tokens. These results underscore the high potential of MR in both efficiency and effectiveness.", "keywords": [ "LLM Response Judgement", "LLM Reliability", "Weak Language Model", "Model Cascading", "Data Selection" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "dRXxFEY8ZE", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Bernhard Sick", "gender": "Male", "institution": "Universit\u00e4t Kassel", "country": "", "position": "Full Professor" }, { "name": "Christoph Scholz", "gender": "Male", "institution": "Universit\u00e4t Kassel", "country": "", "position": "Postdoc" }, { "name": "Denis Huseljic", "gender": "Male", "institution": "Universit\u00e4t Kassel", "country": "", "position": "PhD student" }, { "name": "Jonas Lange", "gender": "Male", "institution": "Christian-Albrechts-Universit\u00e4t Kiel", "country": "DE", "position": "PhD student" }, { "name": "Lukas Rauch", "gender": "Male", "institution": "University of Kassel", "country": "DE", "position": "PhD student" }, { "name": "Marek Herde", "gender": "Male", "institution": "Universit\u00e4t Kassel", "country": "", "position": "PhD student" }, { "name": "Moritz Wirth", "gender": "Male", "institution": "Universit\u00e4t Kassel", "country": "DE", "position": "MS student" }, { "name": "Raphael Schwinger", "gender": "Male", "institution": "Christian-Albrechts-Universit\u00e4t Kiel", "country": "", "position": "PhD student" }, { "name": "Ren\u00e9 Heinrich", "gender": "Male", "institution": "Universit\u00e4t Kassel", "country": "", "position": "PhD student" }, { "name": "Stefan Kahl", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Postdoc" }, { "name": "Sven Tomforde", "gender": "Male", "institution": "Christian-Albrechts-Universit\u00e4t Kiel", "country": "", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 52, "n_ref": 165, "n_ref_all": 193, "n_fig": 11, "n_tab": 16, "L_tab": 11348, "n_element_tab": 769, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 1188, "n_element_tab_1": 39, "formula_len_all": 2238, "formula_len_all_1": 2099, "len_all": 311927, "len_all_1": 70887, "len_abs": 1113, "len_title": 81, "len_sents": 89671, "len_sents_1": 37020, "n_sents": 689, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 88, "L_abs": 1176, "title": "$\\texttt{BirdSet}$: A Large-Scale Dataset for Audio Classification in Avian Bioacoustics", "abs": "Deep learning (DL) has greatly advanced audio classification, yet the field is limited by the scarcity of large-scale benchmark datasets that have propelled progress in other domains. While AudioSet is a pivotal step to bridge this gap as a universal-domain dataset, its restricted accessibility and limited range of evaluation use cases challenge its role as the sole resource. Therefore, we introduce $\\texttt{BirdSet}$, a large-scale benchmark data set for audio classification focusing on avian bioacoustics. $\\texttt{BirdSet}$ surpasses AudioSet with over 6,800 recording hours ($\\uparrow17\\%$) from nearly 10,000 classes ($\\uparrow18\\times$) for training and more than 400 hours ($\\uparrow7\\times$) across eight strongly labeled evaluation datasets. It serves as a versatile resource for use cases such as multi-label classification, covariate shift or self-supervised learning. We benchmark six well-known DL models in multi-label classification across three distinct training scenarios and outline further evaluation use cases in audio classification. We host our dataset on Hugging Face for easy accessibility and offer an extensive codebase to reproduce our results.", "keywords": [ "audio classification", "multi-label", "dataset collection", "bioacoustics" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 4, 2, 4, 4 ], "presentation_list": [ 4, 3, 4, 4 ], "contribution_list": [ 4, 2, 4, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "dQzpP9ziaJ", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Chenkun Tan", "gender": "Male", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Linlin Li", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Linyang Li", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Mianqiu Huang", "gender": "Male", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Mozhi Zhang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Pengyu Wang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Qipeng Guo", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "", "position": "Researcher" }, { "name": "Qun Liu", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "HK", "position": "Chief Scientist of Speech and Language Computing" }, { "name": "Shaojun Zhou", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Xiaoran Liu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Xipeng Qiu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Zhe Xu", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 33, "n_ref": 73, "n_ref_all": 80, "n_fig": 4, "n_tab": 16, "L_tab": 2951, "n_element_tab": 141, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1352, "n_element_tab_1": 90, "formula_len_all": 41, "formula_len_all_1": 40, "len_all": 167669, "len_all_1": 63970, "len_abs": 1629, "len_title": 110, "len_sents": 52474, "len_sents_1": 31194, "n_sents": 393, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1642, "title": "LongSafetyBench: Long-Context LLMs Struggle with Safety Issues", "abs": "With the development of large language models (LLMs), the sequence length of these models continues to increase, drawing significant attention to long-context language models. However, the evaluation of these models has been primarily limited to their capabilities, with a lack of research focusing on their safety. Existing work, such as ManyShotJailbreak, has to some extent demonstrated that long-context language models can exhibit safety concerns. However, the methods used are limited and lack comprehensiveness. In response, we introduce **LongSafetyBench**, the first benchmark designed to objectively and comprehensively evaluate the safety of long-context models. LongSafetyBench consists of 10 task categories, with an average length of 41,889 words. After testing eight long-context language models on LongSafetyBench, we found that existing models generally exhibit insufficient safety capabilities. Moreover, models' safety performance in long-context scenarios does not always align with that in short-context scenarios. Further investigation revealed that long-context models tend to overlook harmful content within lengthy texts. We also proposed a simple yet effective solution, allowing open-source models to achieve performance comparable to that of top-tier closed-source models. We believe that LongSafetyBench can serve as a valuable benchmark for evaluating the safety capabilities of long-context language models. We hope that our work will encourage the broader community to pay attention to the safety of long-context models and contribute to the development of solutions to improve the safety of long-context LLMs.", "keywords": [ "Long-Context Language Model", "Model Safety", "Benchmark" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "dQpZolwXiH", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Andreas Dengel", "gender": "Male", "institution": "Universit\u00e4t Kaiserslautern", "country": "DE", "position": "Full Professor" }, { "name": "Deepak Pathak", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Researcher" }, { "name": "Hiba Najjar", "gender": "Female", "institution": "Universit\u00e4t Kaiserslautern", "country": "", "position": "PhD student" }, { "name": "Marlon Nuske", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 62, "n_ref": 98, "n_ref_all": 126, "n_fig": 17, "n_tab": 6, "L_tab": 1620, "n_element_tab": 222, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 112, "n_element_tab_1": 16, "formula_len_all": 131, "formula_len_all_1": 0, "len_all": 184229, "len_all_1": 62936, "len_abs": 1323, "len_title": 121, "len_sents": 58317, "len_sents_1": 32363, "n_sents": 373, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1328, "title": "Intrinsic Explainability of Multimodal Learning for Crop Yield Prediction", "abs": "Multimodal learning enables various machine learning tasks to benefit from diverse data sources, effectively mimicking the interplay of different factors in real life events. While the heterogeneous nature of these modalities may necessitate the design of complex architectures, their interpretability is often overlooked. In this study, we leverage the intrinsic explainability of Transformer-based models to explain multimodal learning frameworks. We utilize the self-attention mechanism alongside model-specific feature attribution techniques, comparing these against post-hoc methods. Our detailed analysis focuses on the challenging task of crop yield prediction, exploiting the characteristics of the modalities and the data to aggregate local explanations at multiple levels. Our findings indicate that Transformers significantly outperform other architectures in yield prediction, making them well-suited for further intrinsic interpretability analysis. Among the modalities, satellite data emerged as the most influential but requires deeper layers for effective feature extraction due to its complex structure. Additionally, we observed that the Attention Rollout method is more robust than Generic Attention, aligns more closely with Shapley-based attributions and shows reduced sensitivity to minor input variations.", "keywords": [ "intrinsic interpretability", "explainable AI", "multimodal learning", "Transformers", "shapley values", "crop yield prediction", "remote sensing" ], "rating_list": [ 5, 6, 3, 3, 3 ], "soundness_list": [ 3, 3, 2, 2, 2 ], "presentation_list": [ 3, 4, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 2, 1 ], "confidence_list": [ 3, 3, 4, 3, 4 ] }, { "paper_id": "dQG8R9uOq2", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Chun-Mei Feng", "gender": "Female", "institution": "IHPC", "country": "SG", "position": "Researcher" }, { "name": "Jiayou Zheng", "gender": "Male", "institution": "The Chinese University of Hong Kong (Shenzhen)", "country": "CN", "position": "PhD student" }, { "name": "Shuguang Cui", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Yuzhe Zhou", "gender": "unknown", "institution": "CUHK(SZ)", "country": "", "position": "PhD student" }, { "name": "Zixun Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Zhen LI_Jason", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 27, "n_ref": 46, "n_ref_all": 50, "n_fig": 2, "n_tab": 6, "L_tab": 1771, "n_element_tab": 172, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 170, "n_element_tab_1": 18, "formula_len_all": 584, "formula_len_all_1": 541, "len_all": 105099, "len_all_1": 50654, "len_abs": 1196, "len_title": 152, "len_sents": 31337, "len_sents_1": 24522, "n_sents": 199, "n_sents_1": 160, "venue": "ICLR2025", "decision": "withdraw", "L_title": 104, "L_abs": 1095, "title": "Boost Protein Language Model with Injected Structure Information through Parameter Efficient Fine-tuning", "abs": "At the intersection of computer vision and computational biology, large-scale Protein Language Models (PLMs), particularly the ESM series, have made significant advances in understanding protein structures and functions. However, these models are mainly pre-trained on pure residue sequence, often lack explicit incorporation of structural information, highlighting an opportunity for enhancement. In this paper, we design a parameter-efficient fine-tuning method, SI-Tuning, that injects structural information into PLMs while preserving the original model parameters frozen and optimizing a minimal task-specific vector for input embedding and attention map. This vector, extracted from structural features like dihedral angles and distance maps, introduces a structural bias that enhances the model's performance in downstream tasks. Extensive experiments show that our parameter-efficient fine-tuned ESM-2 650M model outperforms SaProt, a large-scale model pre-trained with protein structural data, in various downstream tasks with a reduction of 40.3% GPU memory and 39.8% time consumption.", "keywords": [ "Protein Language Model", "Parameter-Efficient Fine-Tuning", "Structure Information Injecting", "ESM2" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 5 ] }, { "paper_id": "dQ2xiSIYzp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Buzhen Huang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "PhD student" }, { "name": "Chen Li", "gender": "Female", "institution": "A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Gim Hee Lee", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" }, { "name": "Hanlin Chen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Jianfeng Zhang", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Jinnan Chen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Lingting Zhu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 31, "n_ref": 108, "n_ref_all": 130, "n_fig": 12, "n_tab": 7, "L_tab": 1677, "n_element_tab": 176, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1508, "n_element_tab_1": 146, "formula_len_all": 264, "formula_len_all_1": 254, "len_all": 118431, "len_all_1": 50138, "len_abs": 1183, "len_title": 100, "len_sents": 35934, "len_sents_1": 22143, "n_sents": 272, "n_sents_1": 155, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1188, "title": "Generalizable Human Gaussians from Single-View Image", "abs": "In this work, we tackle the task of learning 3D human Gaussians from a single image, focusing on recovering detailed appearance and geometry including unobserved regions. We introduce a single-view generalizable Human Gaussian Model (HGM), which employs a novel generate-then-refine pipeline with the guidance from human body prior and diffusion prior. Our approach uses a ControlNet to refine rendered back-view images from coarse predicted human Gaussians, then uses the refined image along with the input image to reconstruct refined human Gaussians. To mitigate the potential generation of unrealistic human poses and shapes, we incorporate human priors from the SMPL-X model as a dual branch, propagating image features from the SMPL-X volume to the image Gaussians using sparse convolution and attention mechanisms. Given that the initial SMPL-X estimation might be inaccurate, we gradually refine it with our HGM model. We validate our approach on several publicly available datasets. Our method surpasses previous methods in both novel view synthesis and surface reconstruction. Our approach also exhibits strong generalization for cross-dataset evaluation and in-the-wild images.", "keywords": [ "Human Gaussians", "3D Human Reconstruction" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 5, 4, 5 ] }, { "paper_id": "dPqfgsYn6l", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Arthur Aubret", "gender": "unknown", "institution": "Frankfurt Institute for Advanced Studies ", "country": "DE", "position": "Postdoc" }, { "name": "Chen Yu", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Jane Yang", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Jochen Triesch", "gender": "Male", "institution": "Goethe University", "country": "DE", "position": "Full Professor" }, { "name": "Marcel C. Raabe", "gender": "Male", "institution": "Johann Wolfgang Goethe Universit\u00e4t Frankfurt am Main", "country": "DE", "position": "MS student" }, { "name": "Zhengyang Yu", "gender": "Male", "institution": "Frankfurt Institute for Advanced Studies", "country": "DE", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 41, "n_ref": 76, "n_ref_all": 108, "n_fig": 19, "n_tab": 2, "L_tab": 448, "n_element_tab": 21, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 448, "n_element_tab_1": 21, "formula_len_all": 156, "formula_len_all_1": 101, "len_all": 118460, "len_all_1": 60744, "len_abs": 1520, "len_title": 119, "len_sents": 37367, "len_sents_1": 29176, "n_sents": 303, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1529, "title": "Active Gaze Behavior Boosts Self-Supervised Object Learning", "abs": "Toddlers learn to recognize objects from different viewpoints with almost no supervision. Recent works argue that toddlers develop this ability by mapping close-in-time visual inputs to similar representations while interacting with objects. High acuity vision is only available in the central visual field, which may explain why toddlers (much like adults) constantly move around their gaze during such interactions. It is unclear whether/how much toddlers curate their visual experience through these eye movements to support their learning of object representations. In this work, we explore whether a bio-inspired visual learning model can harness toddlers\u2019 gaze behavior during a play session to develop view-invariant object recognition. Exploiting head-mounted eye tracking during dyadic play, we simulate toddlers\u2019 central visual field experience by cropping image regions centered on the gaze location. This visual stream feeds time-based self-supervised learning algorithms. Our experiments demonstrate that toddlers\u2019 gaze strategy supports the learning of invariant object representations. Our analysis also reveals that the limited size of the central visual field where acuity is high is crucial for this. We further find that toddlers\u2019 visual experience elicits more robust representations compared to adults\u2019, mostly because toddlers look at objects they hold themselves for longer bouts. Overall, our work reveals how toddlers\u2019 gaze behavior supports self-supervised learning of view-invariant object recognition.", "keywords": [ "Gaze behavior", "Self-supervised learning", "Time-based augmentations", "Object recognition" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "dPbJb9XdI1", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Olivier Marre", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Peter Neri", "gender": "Male", "institution": "Ecole Normale Sup\u00e9rieure de Paris", "country": "FR", "position": "Principal Researcher" }, { "name": "Simone Azeglio", "gender": "Male", "institution": "Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "FR", "position": "PhD student" }, { "name": "Ulisse Ferrari", "gender": "Male", "institution": "Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "FR", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 42, "n_ref": 66, "n_ref_all": 102, "n_fig": 19, "n_tab": 4, "L_tab": 1405, "n_element_tab": 80, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 521, "n_element_tab_1": 46, "formula_len_all": 442, "formula_len_all_1": 491, "len_all": 134395, "len_all_1": 60724, "len_abs": 1593, "len_title": 142, "len_sents": 47158, "len_sents_1": 30364, "n_sents": 330, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1595, "title": "Convolution goes higher-order: a biologically inspired mechanism empowers image classification.", "abs": "We propose a novel approach to image classification inspired by complex nonlinear biological visual processing, whereby classical convolutional neural networks (CNNs) are equipped with learnable higher-order convolutions. Our model incorporates a Volterra-like expansion of the convolution operator, capturing multiplicative interactions akin to those observed in early and advanced stages of biological visual processing. We evaluated this approach on synthetic datasets by measuring sensitivity to testing higher-order correlations and performance in standard benchmarks (MNIST, FashionMNIST, CIFAR10, CIFAR100 and Imagenette). Our architecture outperforms traditional CNN baselines, and achieves optimal performance with expansions up to 3rd/4th order, aligning remarkably well with the distribution of pixel intensities in natural images. Through systematic perturbation analysis, we validate this alignment by isolating the contributions of specific image statistics to model performance, demonstrating how different orders of convolution process distinct aspects of visual information. Furthermore, Representational Similarity Analysis reveals distinct geometries across network layers, indicating qualitatively different modes of visual information processing. Our work bridges neuroscience and deep learning, offering a path towards more effective, biologically inspired computer vision models. It provides insights into visual information processing and lays the groundwork for neural networks that better capture complex visual patterns, particularly in resource-constrained scenarios.", "keywords": [ "Higher-order convolution", "Biologically inspired neural networks", "Image classification", "Convolutional Neural Networks", "Biological visual processing", "Neural representations" ], "rating_list": [ 6, 6, 3, 3 ], "soundness_list": [ 2, 4, 2, 2 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 4, 1, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "dOwmtbn6ZO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Aram Galstyan", "gender": "Male", "institution": "Information Sciences Institute", "country": "", "position": "Research Director" }, { "name": "Bhavana Ganesh", "gender": "Female", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Goeric Huybrechts", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Sravan Babu Bodapati", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Sullam Jeoung", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 23, "n_ref": 43, "n_ref_all": 65, "n_fig": 7, "n_tab": 9, "L_tab": 1889, "n_element_tab": 203, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 1761, "n_element_tab_1": 169, "formula_len_all": 40, "formula_len_all_1": 88, "len_all": 147138, "len_all_1": 66014, "len_abs": 1112, "len_title": 163, "len_sents": 44885, "len_sents_1": 28866, "n_sents": 340, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 115, "L_abs": 1115, "title": "Adaptive Video Understanding Agent: Enhancing Efficiency with Dynamic Frame Sampling and Feedback-driven Reasoning", "abs": "Understanding long-form video content presents significant challenges due to its temporal complexity and the substantial computational resources required. In this work, we propose an agent-based approach to enhance both the efficiency and effectiveness of long-form video understanding by utilizing large language models (LLMs) and their tool-harnessing ability. A key aspect of our method is query-adaptive frame sampling, which leverages the reasoning capabilities of LLMs to process only the most relevant frames in real-time, and addresses an important limitation of existing methods which typically involve sampling redundant or irrelevant frames. To enhance the reasoning abilities of our video-understanding agent, we leverage the self-reflective capabilities of LLMs to provide verbal reinforcement to the agent, which leads to improved performance while minimizing the number of frames accessed. We evaluate our method across several video understanding benchmarks and demonstrate that not only it enhances state-of-the-art performance but also improves efficiency by reducing the number of frames sampled.", "keywords": [ "multimodal agent", "long context video processing", "adaptive sampling" ], "rating_list": [ 3, 5, 3, 3, 8 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 5, 5, 4 ] }, { "paper_id": "dOkuRMrWtL", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Bing Liu", "gender": "Male", "institution": "University of Illinois at Chicago", "country": "US", "position": "Full Professor" }, { "name": "Fanhu Zeng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Fei Zhu", "gender": "Male", "institution": "Centre for Artificial Intelligence and Robotics Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "CN", "position": "Postdoc" }, { "name": "Haiyang Guo", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Xu-yao Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 39, "n_ref": 93, "n_ref_all": 114, "n_fig": 9, "n_tab": 2, "L_tab": 1917, "n_element_tab": 141, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2766, "n_element_tab_1": 174, "formula_len_all": 463, "formula_len_all_1": 446, "len_all": 132832, "len_all_1": 67567, "len_abs": 1331, "len_title": 125, "len_sents": 37334, "len_sents_1": 29167, "n_sents": 278, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1352, "title": "DESIRE: Dynamic Knowledge Consolidation for Rehearsal-Free Continual Learning", "abs": "Continual learning aims to equip models with the ability to retain previously learned knowledge like a human. Recent work incorporating Parameter-Efficient Fine-Tuning has revitalized the field by introducing lightweight extension modules. However, existing methods usually overlook the issue of information leakage caused by the fact that the experiment data have been used in pre-trained models. Once these duplicate data are removed in the pre-training phase, their performance can be severely affected. In this paper, we propose a new LoRA-based rehearsal-free method named $\\textbf{DESIRE}$. Our method avoids imposing additional constraints during training to mitigate catastrophic forgetting, thereby maximizing the learning of new classes. To integrate knowledge from old and new tasks, we propose two efficient post-processing modules. On the one hand, we retain only two sets of LoRA parameters for merging and propose dynamic representation consolidation to calibrate the merged feature representation. On the other hand, we propose decision boundary refinement to address classifier bias when training solely on new class data. Extensive experiments demonstrate that our method achieves state-of-the-art performance on multiple datasets and strikes an effective balance between stability and plasticity. Our code will be publicly available.", "keywords": [ "Class-incremental learning", "Model merging" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "dOiinVDQEW", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Andrew Bell", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Joao Fonseca", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 34, "n_ref": 42, "n_ref_all": 63, "n_fig": 7, "n_tab": 1, "L_tab": 48, "n_element_tab": 4, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 48, "n_element_tab_1": 4, "formula_len_all": 93, "formula_len_all_1": 93, "len_all": 96840, "len_all_1": 67883, "len_abs": 1212, "len_title": 122, "len_sents": 29280, "len_sents_1": 29280, "n_sents": 212, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1221, "title": "Output Scouting: Auditing Large Language Models for Catastrophic Responses", "abs": "Recent high profile incidents in which the use of Large Language Models (LLMs) resulted in significant harm to individuals have brought about a growing interest in AI safety. One reason LLM safety issues occur is that models often have at least some non-zero probability of producing harmful outputs. In this work, we explore the following scenario: imagine an AI safety auditor is searching for catastrophic responses from an LLM (e.g. a \"yes\" responses to \"can I fire an employee for being pregnant?\"), and is able to query the model a limited number times (e.g. 1000 times). What is a strategy for querying the model that would efficiently find those failure responses? To this end, we propose output scouting: an approach that aims to generate semantically fluent outputs to a given prompt matching any target probability distribution. We then run experiments using two LLMs and find numerous examples of catastrophic responses. We conclude with a discussion that includes advice for practitioners who are looking to implement LLM auditing for catastrophic responses. We will release an open-source toolkit that implements our auditing framework using the Hugging Face ``transformers`` library following publication.", "keywords": [ "large language models", "safety", "interpretability", "red teaming" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 4, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "dOAkHmsjRX", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Hyunseo Koh", "gender": "Male", "institution": "Gwangju Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Jonghyun Choi", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" }, { "name": "Minhyuk Seo", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Researcher" } ], "n_formula": 25, "n_formula_1": 17, "n_ref_uni": 62, "n_ref": 131, "n_ref_all": 162, "n_fig": 19, "n_tab": 23, "L_tab": 8575, "n_element_tab": 958, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1138, "n_element_tab_1": 72, "formula_len_all": 1390, "formula_len_all_1": 641, "len_all": 244644, "len_all_1": 66724, "len_abs": 1301, "len_title": 138, "len_sents": 74465, "len_sents_1": 31330, "n_sents": 569, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 90, "L_abs": 1257, "title": "Budgeted Online Continual Learning by Adaptive Layer Freezing and Frequency-based Sampling", "abs": "The majority of online continual learning (CL) advocates single-epoch training and imposes restrictions on the size of replay memory.\nHowever, single-epoch training would incur a different amount of computations per CL algorithm, and the additional storage cost to store logit or model in addition to replay memory is largely ignored in calculating the storage budget.\nArguing different computational and storage budgets hinder fair comparison among CL algorithms in practice, we propose to use floating point operations (FLOPs) and total memory size in Byte as a metric for computational and memory budgets, respectively, to compare and develop CL algorithms in the same \"total resource budget\".\nTo improve a CL method in a limited total budget, we propose adaptive layer freezing that does not update the layers for less informative batches to reduce computational costs with a negligible loss of accuracy.\nIn addition, we propose a memory retrieval method that allows the model to learn the same amount of knowledge as using random retrieval in fewer iterations.\nEmpirical validations on the CIFAR-10/100, CLEAR-10/100, and ImageNet-1K datasets demonstrate that the proposed approach outperforms the state-of-the-art methods within the same total budget.", "keywords": [ "Continual Learning", "Lifelong Learning" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "dNunnVB4W6", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Ameneh Asgari-Targhi", "gender": "Female", "institution": "Brigham and Women's Hospital, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Barbara D. Lam", "gender": "Female", "institution": "Beth Israel Deaconess Medical Center", "country": "", "position": "Postdoc" }, { "name": "Peiqi Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Polina Golland", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Rameswar Panda", "gender": "Male", "institution": "MIT-IBM Watson AI Lab", "country": "", "position": "Research Scientist" }, { "name": "Tina Kapur", "gender": "unknown", "institution": "harvard medical school", "country": "US", "position": "Assistant Professor" }, { "name": "William M Wells", "gender": "Male", "institution": "Brigham and Women's Hospital, Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Yingcheng Liu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 20, "n_ref_uni": 35, "n_ref": 63, "n_ref_all": 118, "n_fig": 15, "n_tab": 4, "L_tab": 1176, "n_element_tab": 220, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2861, "formula_len_all_1": 1312, "len_all": 168013, "len_all_1": 65807, "len_abs": 635, "len_title": 36, "len_sents": 56696, "len_sents_1": 31506, "n_sents": 457, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 36, "L_abs": 636, "title": "Calibrating Expressions of Certainty", "abs": "We present a novel approach to calibrating linguistic expressions of certainty, e.g., \"Maybe\" and \"Likely\". Unlike prior work that assigns a single score to each certainty phrase, we model uncertainty as distributions over the simplex to capture their semantics more accurately. To accommodate this new representation of certainty, we generalize existing measures of miscalibration and introduce a novel post-hoc calibration method. Leveraging these tools, we analyze the calibration of both humans (e.g., radiologists) and computational models (e.g., language models) and provide interpretable suggestions to improve their calibration.", "keywords": [ "calibration", "uncertainty", "optimal transport", "language models" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 4, 2 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 3, 3, 4, 2 ] }, { "paper_id": "dNT3abOsLo", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Jinkyoo Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Minsu Kim", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Postdoc" }, { "name": "Nayoung Kim", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Seongsu Kim", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Sungsoo Ahn", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 30, "n_formula_1": 27, "n_ref_uni": 65, "n_ref": 130, "n_ref_all": 154, "n_fig": 9, "n_tab": 8, "L_tab": 1309, "n_element_tab": 143, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 190, "n_element_tab_1": 29, "formula_len_all": 2690, "formula_len_all_1": 2239, "len_all": 183715, "len_all_1": 67537, "len_abs": 1056, "len_title": 124, "len_sents": 43453, "len_sents_1": 28629, "n_sents": 366, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1061, "title": "MOFFlow: Flow Matching for Structure Prediction of Metal-Organic Frameworks", "abs": "Metal-organic frameworks (MOFs) are a class of crystalline materials with promising applications in many areas such as carbon capture and drug delivery. In this work, we introduce MOFFlow, the first deep generative model tailored for MOF structure prediction. Existing approaches, including ab initio calculations and even deep generative models, struggle with the complexity of MOF structures due to the large number of atoms in the unit cells. To address this limitation, we propose a novel Riemannian flow matching framework that reduces the dimensionality of the problem by treating the metal nodes and organic linkers as rigid bodies, capitalizing on the inherent modularity of MOFs. By operating in the $SE(3)$ space, MOFFlow effectively captures the roto-translational dynamics of these rigid components in a scalable way. Our experiment demonstrates that MOFFlow accurately predicts MOF structures containing several hundred atoms, significantly outperforming conventional methods and state-of-the-art machine learning baselines while being much faster.", "keywords": [ "metal-organic framework", "material structure prediction", "AI for science" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "dNBE4ciYJF", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Dasom choi", "gender": "Female", "institution": "Chungnam National University", "country": "KR", "position": "Undergrad student" }, { "name": "Hidetaka Kamigaito", "gender": "Male", "institution": "Nara Institute of Science and Technology", "country": "JP", "position": "Associate Professor" }, { "name": "Jingun Kwon", "gender": "unknown", "institution": "Chungnam National University", "country": "KR", "position": "Assistant Professor" }, { "name": "Manabu Okumura", "gender": "Male", "institution": "Institute of Science Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Sang June Moon", "gender": "Male", "institution": "Chungnam National University", "country": "KR", "position": "Undergrad student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 49, "n_ref": 74, "n_ref_all": 118, "n_fig": 6, "n_tab": 4, "L_tab": 5332, "n_element_tab": 219, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3869, "n_element_tab_1": 199, "formula_len_all": 425, "formula_len_all_1": 346, "len_all": 181713, "len_all_1": 72571, "len_abs": 1168, "len_title": 95, "len_sents": 38715, "len_sents_1": 32387, "n_sents": 278, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 47, "L_abs": 1163, "title": "Length Representations in Large Language Models", "abs": "Large language models (LLMs) have shown remarkable capabilities across various tasks that are learned from massive amounts of text-based data. Although LLMs can control output sequence length, particularly through instruction-based settings, the internal mechanisms behind this control has been unexplored. In this study, we provide empirical evidence on how output sequence length information is encoded within the internal representations of LLMs. In particular, our findings show that multi-head attention mechanisms are critical in determining output sequence length, which can be adjusted in an editable manner. By scaling specific hidden units within the model, we can control the output sequence length without losing the informativeness of the generated text, thereby indicating that length information is partially separable from semantic information. Moreover, some hidden units become increasingly active as prompts become more length-specific, thus reflecting the model's internal awareness of this attribute. Our findings suggest that LLMs have learned robust and adaptable internal mechanisms for controlling output length without external controls.", "keywords": [ "LLMs", "Length representation", "Length control" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 1, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "dMj3SDNxn4", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Guy Ben Razon", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" }, { "name": "Jonathan Svirsky", "gender": "Male", "institution": "Bar-Ilan University", "country": "", "position": "PhD student" }, { "name": "Uri Shaham", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 36, "n_ref": 76, "n_ref_all": 93, "n_fig": 3, "n_tab": 5, "L_tab": 3315, "n_element_tab": 390, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 560, "formula_len_all_1": 451, "len_all": 134410, "len_all_1": 66977, "len_abs": 1520, "len_title": 112, "len_sents": 41110, "len_sents_1": 33050, "n_sents": 307, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1527, "title": "UNICORNN: Unimodal Calibrated Ordinal Regression Neural Network", "abs": "Ordinal regression is a supervised machine learning technique aimed at predicting the value of a discrete dependent variable with an ordered set of possible outcomes. Many of the algorithms that have been developed to address this issue rely on maximum likelihood for training. However, the standard maximum likelihood approach often fails to adequately capture the inherent order of classes, even though it tends to produce well-calibrated probabilities. Alternatively, some methods use Optimal Transport (OT) divergence as their training objective. Unlike maximum likelihood, OT accounts for the ordering of classes; however, in this manuscript, we show that it doesn't always yield well-calibrated probabilities. To overcome these limitations, we introduce UNICORNN, an approach inspired by the well-known Proportional Odds Model, which offers three key guarantees: (i) it ensures unimodal output probabilities, a valuable feature for many real-world applications;\n(ii) it employs OT loss during training to accurately capture the natural order of classes;\n(iii) it provides well-calibrated probability estimates through a post-training accuracy-preserving calibration step.\nExperimental results on six real-world datasets \ndemonstrate that UNICORNN consistently either outperforms or performs as well as recently proposed deep learning approaches for ordinal regression. It excels in both accuracy and probability calibration, while also guaranteeing output unimodality. The code will be publicly available upon acceptance.", "keywords": [ "unimodality", "ordinal regression", "probability calibration", "deep learning" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "dML3XGvWmy", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Liangming Pan", "gender": "Male", "institution": "University of Arizona", "country": "US", "position": "Assistant Professor" }, { "name": "Xiaojun Wan", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Xinyi Wang", "gender": "Female", "institution": "UC Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Xunjian Yin", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "William Yang Wang", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 40, "n_ref": 75, "n_ref_all": 93, "n_fig": 5, "n_tab": 3, "L_tab": 5080, "n_element_tab": 92, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 262, "n_element_tab_1": 22, "formula_len_all": 154, "formula_len_all_1": 121, "len_all": 166382, "len_all_1": 64901, "len_abs": 1013, "len_title": 130, "len_sents": 45779, "len_sents_1": 31416, "n_sents": 335, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1009, "title": "G\u00f6del Agent: A Self-Referential Framework Helps for Recursively Self-Improvement", "abs": "The rapid advancement of large language models (LLMs) has significantly enhanced the capabilities of AI-driven agents across various tasks. However, existing agentic systems, whether based on fixed pipeline algorithms or pre-defined meta-learning frameworks, cannot search the whole agent design space due to the restriction of human-designed components, and thus might miss the globally optimal agent design. In this paper, we introduce G\u00f6del Agent, a self-evolving framework inspired by the G\u00f6del machine, enabling agents to recursively improve themselves without relying on predefined routines or fixed optimization algorithms. G\u00f6del Agent leverages LLMs to dynamically modify its own logic and behavior, guided solely by high-level objectives through prompting. Experimental results on mathematical reasoning and complex agent tasks demonstrate that implementation of G\u00f6del Agent can achieve continuous self-improvement, surpassing manually crafted agents in performance, efficiency, and generalizability.", "keywords": [ "Agent", "Large Language Model", "Reasoning", "Self-Improvement" ], "rating_list": [ 5, 6, 6, 6, 5 ], "soundness_list": [ 2, 2, 3, 3, 1 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 3, 3 ] }, { "paper_id": "dM4yZd6ic9", "primary_area": "applications to neuroscience & cognitive science", "n_author": 9, "author_info_list": [ { "name": "Hui Xiong", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Hyejeong Jo", "gender": "Female", "institution": "Kyung Hee University", "country": "KR", "position": "MS student" }, { "name": "JINNI ZHOU", "gender": "Female", "institution": "Hong Kong University of Science and Technology (GZ)", "country": "CN", "position": "Lecturer" }, { "name": "Renjing Xu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Won Hee Lee", "gender": "unknown", "institution": "Kyung Hee University", "country": "KR", "position": "Associate Professor" }, { "name": "Xuming Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Yiqian Yang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Yiqun Duan", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "QIANG ZHANG", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 22, "n_ref": 37, "n_ref_all": 44, "n_fig": 2, "n_tab": 6, "L_tab": 2262, "n_element_tab": 269, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 989, "n_element_tab_1": 152, "formula_len_all": 195, "formula_len_all_1": 195, "len_all": 122554, "len_all_1": 63935, "len_abs": 1365, "len_title": 89, "len_sents": 38686, "len_sents_1": 31580, "n_sents": 293, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 41, "L_abs": 1381, "title": "MAD: Multi-Alignment MEG-to-Text Decoding", "abs": "Deciphering language from brain activity is a crucial task in brain-computer interface (BCI) research. Non-invasive cerebral signaling techniques including electroencephalography (EEG) and magnetoencephalography (MEG) are becoming increasingly popular due to their safety and practicality, avoiding invasive electrode implantation. However, current works under-investigated three points: 1) a predominant focus on EEG with limited exploration of MEG, which provides superior signal quality; 2) poor performance on unseen text, indicating the need for models that can better generalize to diverse linguistic contexts; 3) insufficient integration of information from other modalities, which could potentially constrain our capacity to comprehensively understand the intricate dynamics of brain activity.\n\nThis study presents a novel approach for translating MEG signals into text using a speech-decoding framework with multiple alignments. Our method is the first to introduce an end-to-end multi-alignment framework for totally unseen text generation directly from MEG signals. We achieve an impressive BLEU-1 score on the $\\textit{GWilliams}$ dataset, significantly outperforming the baseline from 5.49 to 10.44 on the BLEU-1 metric. This improvement demonstrates the advancement of our model towards real-world applications and underscores its potential in advancing BCI research.", "keywords": [ "MEG", "text", "speech", "transfer learning" ], "rating_list": [ 6, 5, 3, 3, 6 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 1, 2, 3 ], "confidence_list": [ 4, 3, 4, 5, 4 ] }, { "paper_id": "dM1wO2OkbO", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Chenwei Cui", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Gedeon Muhawenayo", "gender": "Male", "institution": "Arizona State University (ASU)", "country": "US", "position": "PhD student" }, { "name": "Hannah Kerner", "gender": "Female", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" }, { "name": "Zehao Yan", "gender": "Male", "institution": "Colorado State University", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 30, "n_ref": 66, "n_ref_all": 80, "n_fig": 8, "n_tab": 2, "L_tab": 222, "n_element_tab": 24, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 222, "n_element_tab_1": 24, "formula_len_all": 424, "formula_len_all_1": 424, "len_all": 99364, "len_all_1": 67329, "len_abs": 1004, "len_title": 87, "len_sents": 26977, "len_sents_1": 23428, "n_sents": 255, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1010, "title": "Linear-Time Sequence Modeling with MLPs", "abs": "We present Causal Relation Networks (CausalRNs), the first all-MLP sequence modeling architecture with linear-time parallel training.\nTo enable autoregressive modeling, we made Relation Networks (RNs) equivariant and causal through relaxation and masking.\nContrary to the earlier belief that RNs are quadratic-time, we show that when using exp(x) as the activation function, any RN is linear-time, fully parallelizable, and numerically stable.\nOur derivation spontaneously gave rise to familiar design choices adopted by state-of-the-art architectures, e.g. exponential gating and state expansion.\nSuch duality provided a new perspective, from which we not only validated popular design choices, but also discovered new design considerations.\nExperiments on autoregressive language modeling and image classification showed CausalRNs to be comparable to Linear Transformers.\nThe quadratic variant of CausalRNs achieved perfect retrieval on the copying task, which was previously only possible with Transformers.", "keywords": [ "All-MLP", "Sequence Modeling", "Multilayer Perceptron", "Transformer" ], "rating_list": [ 6, 5, 8 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "dL3h1lyUNd", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Changsong Liu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "LiMingyang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "tjuzhangwei", "gender": "Male", "institution": "Tianjin University", "country": "", "position": "Full Professor" }, { "name": "Fan Yimeng", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 24, "n_ref": 43, "n_ref_all": 60, "n_fig": 6, "n_tab": 7, "L_tab": 3004, "n_element_tab": 357, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1355, "n_element_tab_1": 218, "formula_len_all": 347, "formula_len_all_1": 340, "len_all": 121283, "len_all_1": 52703, "len_abs": 1312, "len_title": 49, "len_sents": 28785, "len_sents_1": 24051, "n_sents": 220, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1321, "title": "EDSNN: Edge Detection with Spiking Neuron Network", "abs": "Edge detection has made great progress under the development of Artificial Neural Networks (ANNs), particularly Convolutional Neural Networks (CNNs) and Transformers, some of them even have achieved a beyond human-level performance. However, these methods come with complex designs and high energy consumption. Spiking Neural Networks (SNNs), with their low energy consumption and biological interpretability, offer a promising solution to address these issues. In this work, we propose the first SNN-based method named EDSNN (Edge Detection with Spiking Neural Network) for edge detection. We construct a novel Spiking Multi-Scale Block (SMSB) to effectively utilize multi-scale information, thereby helping the network generate precise and clean edge maps. In addition, to more accurately decode spike trains, we present a Membrane Average Decoding (MAD) method in the prediction block. Our method has the advantages of remarkable efficiency and high performance across multiple datasets. It surpasses the human-level performance on BSDS500 (ODS=0.804 vs. ODS=0.803) while consuming only 14.64 mJ, remains competitive performance among top-performing ANN-based approaches on NYUDv2 (ODS=0.750), and achieves state-of-the-art performance on BIPED (ODS=0.891). Our codes are publicly available in supplementary materials.", "keywords": [ "Spiking Neural Networks", "Edge Detection", "Spiking Multi-Scale Block", "Membrane Average Decoding" ], "rating_list": [ 3, 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "dKfcntLRjZ", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Atefeh Moradan", "gender": "Female", "institution": "Aarhus University", "country": "", "position": "PhD student" }, { "name": "Davide Mottin", "gender": "Male", "institution": "Aarhus University", "country": "DK", "position": "Associate Professor" }, { "name": "Ira Assent", "gender": "Female", "institution": "Aarhus University", "country": "DK", "position": "Full Professor" }, { "name": "Klim Zaporojets", "gender": "Male", "institution": "Aarhus University", "country": "DK", "position": "Postdoc" }, { "name": "Mohammad Sorkhpar", "gender": "Male", "institution": "Indiana State University ", "country": "", "position": "MS student" }, { "name": "Saeedeh Javadi", "gender": "Female", "institution": "Polytechnic Institute of Turin", "country": "IT", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 16, "n_ref": 25, "n_ref_all": 42, "n_fig": 4, "n_tab": 1, "L_tab": 572, "n_element_tab": 6, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 499, "n_element_tab_1": 5, "formula_len_all": 895, "formula_len_all_1": 895, "len_all": 75125, "len_all_1": 53386, "len_abs": 1214, "len_title": 83, "len_sents": 26266, "len_sents_1": 24975, "n_sents": 230, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1220, "title": "Wiki Entity Summarization Benchmark", "abs": "Entity summarization aims to compute concise summaries for entities in knowledge graphs.\nHowever, current datasets and benchmarks are often limited to only a few hundred entities\nand overlook knowledge graph structure. This is particularly evident in the scarcity of\nground-truth summaries, with few labeled entities available for evaluation and training. We\npropose WIKES (Wiki Entity Summarization Benchmark), a large benchmark comprising\nof entities, their summaries, and their connections. Additionally, WIKES features a\ndataset generator to test entity summarization algorithms in different subgraphs of the\nknowledge graph. Importantly, our approach combines graph algorithms and NLP models,\nas well as different data sources such that WIKES does not require human annotation,\nrendering the approach cost-effective and generalizable to multiple domains. Finally,\nWIKES is scalable and capable of capturing the complexities of knowledge graphs in\nterms of topology and semantics. WIKES features existing datasets for comparison.\nEmpirical studies of entity summarization methods confirm the usefulness of our benchmark.\nData, code, and models are available at: https://anonymous.4open.science/r/Wikes-2DDA/README.md", "keywords": [ "Entity summarization", "knowledge graph benchmark" ], "rating_list": [ 6, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "dKZCfzRlm3", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Guangming Shi", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Jinjian Wu", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Junhui Hou", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Yifan Zhang", "gender": "Male", "institution": "City University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Zhiwen Chen", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Researcher" }, { "name": "Zhiyu Zhu", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 32, "n_ref": 75, "n_ref_all": 88, "n_fig": 5, "n_tab": 6, "L_tab": 3350, "n_element_tab": 303, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 941, "n_element_tab_1": 122, "formula_len_all": 485, "formula_len_all_1": 332, "len_all": 157266, "len_all_1": 68448, "len_abs": 1018, "len_title": 117, "len_sents": 43938, "len_sents_1": 25900, "n_sents": 352, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1021, "title": "Learning Effective Multi-modal Trackers via Modality-Sensitive Tuning", "abs": "This paper tackles the critical issue of constructing multi-modal trackers by effectively adapting the extensive knowledge of pre-trained RGB trackers to auxiliary modalities.To address the challenges, we propose a novel modality sensitivity-aware tuning framework, namely MST, which delicately models the learning process via adaptive tuning of model weights by inherent modality characteristics. Specifically, we first investigate the parameter modality-sensitivity as a criterion for measuring a precise element-wise essentiality for multi-modal adaptation. Then, in the tuning phase, we further leverage such sensitivity to bolster the stability and coherence of multi-modal representations, thereby enhancing generalization capabilities. Extensive experiments showcase the effectiveness of the proposed method, surpassing current state-of-the-art techniques across various multi-modal tracking scenarios and demonstrating remarkable performance even in extreme conditions. The source code will be publicly available.", "keywords": [ "Multi-modal Tracking; Cross-modal Fine-Tuning" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "dIoLjHet58", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "DongNyeong Heo", "gender": "Male", "institution": "Handong Global University", "country": "KR", "position": "PhD student" }, { "name": "Heeyoul Choi", "gender": "unknown", "institution": "Handong Global University", "country": "KR", "position": "Associate Professor" } ], "n_formula": 60, "n_formula_1": 24, "n_ref_uni": 42, "n_ref": 88, "n_ref_all": 120, "n_fig": 9, "n_tab": 3, "L_tab": 1183, "n_element_tab": 86, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 175, "n_element_tab_1": 9, "formula_len_all": 6876, "formula_len_all_1": 1286, "len_all": 164397, "len_all_1": 64587, "len_abs": 1379, "len_title": 110, "len_sents": 48084, "len_sents_1": 28748, "n_sents": 381, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1385, "title": "Generalized Probabilistic Attention Mechanism in Transformers", "abs": "The Transformer architecture has become widely adopted due to its demonstrated success, attributed to the attention mechanism at its core. Despite these successes, the attention mechanism of Transformers is associated with two well-known issues: rank-collapse and gradient vanishing. In this paper, we present a theoretical analysis that it is inherently difficult to address both issues simultaneously in the conventional attention mechanism. To handle these issues, we introduce a novel class of attention mechanism, referred to as generalized probabilistic attention mechanism (GPAM), and its dual-attention implementation within the Transformer architecture. Unlike conventional attention mechanisms, GPAM allows for negative attention scores while preserving a fixed total sum. We provide theoretical evidence that the proposed dual-attention GPAM (daGPAM) effectively mitigates both the rank-collapse and gradient vanishing issues which are difficult to resolve simultaneously with the conventional attention mechanisms. Furthermore, we empirically validate this theoretical evidence, demonstrating the superiority of daGPAM compared to other alternative attention mechanisms that were proposed to address the same issues. Additionally, we demonstrate the practical benefits of GPAM in natural language processing tasks, such as language modeling and neural machine translation.", "keywords": [ "Attention Mechanism", "Transformer", "Rank-Collapse", "Gradient Vanishing" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "dImD2sgy86", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 6, "author_info_list": [ { "name": "Denis Blessing", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "", "position": "PhD student" }, { "name": "Julius Berner", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Junhua Chen", "gender": "Not Specified", "institution": "University of Cambridge", "country": "", "position": "Undergrad student" }, { "name": "Lorenz Richter", "gender": "unknown", "institution": "Zuse Institute Berlin", "country": "DE", "position": "Postdoc" }, { "name": "Anima Anandkumar", "gender": "Female", "institution": "California Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Gerhard Neumann", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 77, "n_formula_1": 38, "n_ref_uni": 56, "n_ref": 176, "n_ref_all": 218, "n_fig": 10, "n_tab": 12, "L_tab": 5331, "n_element_tab": 578, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2844, "n_element_tab_1": 156, "formula_len_all": 5995, "formula_len_all_1": 2741, "len_all": 283755, "len_all_1": 80075, "len_abs": 1449, "len_title": 89, "len_sents": 88922, "len_sents_1": 30849, "n_sents": 818, "n_sents_1": 288, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 1454, "title": "Sequential Controlled Langevin Diffusions", "abs": "An effective approach for sampling from unnormalized densities is based on the idea of gradually transporting samples from an easy prior to the complicated target distribution. Two popular methods are (1) Sequential Monte Carlo (SMC), where the transport is performed through successive annealed densities via prescribed Markov chains and resampling steps, and (2) recently developed diffusion-based\nsampling methods, where a learned dynamical transport is used. Despite the common goal, both approaches have different, often complementary, advantages and drawbacks. The resampling steps in SMC allow focusing on promising regions of the space, often leading to robust performance. While the algorithm enjoys asymptotic guarantees, the lack of flexible, learnable transitions can lead to slow convergence. On the other hand, diffusion-based samplers are learned and can potentially better adapt themselves to the target at hand, yet often suffer from training instabilities. In this work, we present a principled framework for combining SMC with diffusion-based samplers by viewing both methods in continuous time and considering measures on path space. This culminates in the new Sequential Controlled Langevin Diffusion (SCLD) sampling method, which is able to utilize the benefits of both methods and reaches improved performance on multiple benchmark problems, in many cases using only 10% of the training budget of previous diffusion-based samplers.", "keywords": [ "variational inference", "sequential importance sampling", "monte carlo", "SDEs", "diffusion models" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "dIkpHooa2D", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Anvith Thudi", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Chris Maddison", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 29, "n_formula_1": 12, "n_ref_uni": 35, "n_ref": 80, "n_ref_all": 112, "n_fig": 15, "n_tab": 6, "L_tab": 2050, "n_element_tab": 216, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 473, "n_element_tab_1": 19, "formula_len_all": 13680, "formula_len_all_1": 11803, "len_all": 162745, "len_all_1": 82743, "len_abs": 1289, "len_title": 125, "len_sents": 49147, "len_sents_1": 32079, "n_sents": 344, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1296, "title": "MixMax: Distributional Robustness in Function Space via Optimal Data Mixtures", "abs": "Machine learning models are often required to perform well across several pre-defined settings, such as a set of user groups. Worst-case performance is a common metric to capture this requirement, and is the objective of group distributionally robust optimization (group DRO). Unfortunately, these methods struggle when the loss is non-convex in the parameters, or the model class is non-parametric. Here, we make a classical move to address this: we reparameterize group DRO from parameter space to function space, which results in a number of advantages. First, we show that group DRO over the space of bounded functions admits a minimax theorem. Second, for cross-entropy and mean squared error, we show that the minimax optimal mixture distribution is the solution of a simple convex optimization problem. Thus, provided one is working with a model class of universal function approximators, group DRO can be solved by a convex optimization problem followed by a classical risk minimization problem. We call our method MixMax. In our experiments, we found that MixMax matched or outperformed the standard group DRO baselines, and in particular, MixMax improved the performance of XGBoost over the only baseline, data balancing, for variations of the ACSIncome and CelebA annotations datasets.", "keywords": [ "Distributional Robustness", "Non-parametric Learning", "Function Space", "Trustworthy Machine Learning" ], "rating_list": [ 5, 6, 8, 8 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 2, 4, 2 ] }, { "paper_id": "dIaykjbiiL", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Carl Yang", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Assistant Professor" }, { "name": "Fanzhe Fu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jing Zhang", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Junru Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Lvbin Ma", "gender": "Male", "institution": "Zhejiang Huayun Information Technology Co. Ltd", "country": "", "position": "manager" }, { "name": "Yang Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 15, "n_ref": 31, "n_ref_all": 54, "n_fig": 8, "n_tab": 7, "L_tab": 1732, "n_element_tab": 197, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 271, "formula_len_all_1": 229, "len_all": 149378, "len_all_1": 51607, "len_abs": 873, "len_title": 111, "len_sents": 51630, "len_sents_1": 27102, "n_sents": 294, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 877, "title": "Are Synthetic Time-series Data Really not as Good as Real Data?", "abs": "To alleviate the commonly encountered inadequate time-series data problem in DL (DL), we develop a non-DL generic data synthesis method. When current methods require real data or data statistics to train generators or synthesize data, our method InfoBoost enables zero-shot training of models without the need for real data or data statistics. Additionally, as an application of our synthetic data, we train an unconditional feature (rhythm, noise, trend) decomposer based on our synthetic data, which is applicable to real time-series data. Through experiments, our non-DL synthetic data enables models to achieve superior performance on unsupervised tasks and self-supervised prediction \\& imputation compared models using real data. Visualized case studies further demonstrate the effectiveness of our novel unconditional feature decomposer trained with our synthetic data.", "keywords": [ "Time-Series Data", "Data Synthesis", "Non-Deep-Learning Data Synthesis", "InfoBoost", "Prediction", "Imputation", "Feature Decomposition" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 1, 1, 2, 1 ], "presentation_list": [ 1, 1, 3, 1 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "dIZB7jeSUv", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Guangcong Zheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Tao Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Teng Li", "gender": "Non-Binary", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Xi Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yehao Lu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "RuiJiang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 37, "n_ref": 102, "n_ref_all": 112, "n_fig": 6, "n_tab": 3, "L_tab": 962, "n_element_tab": 109, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 656, "n_element_tab_1": 70, "formula_len_all": 1042, "formula_len_all_1": 1040, "len_all": 154113, "len_all_1": 59376, "len_abs": 1628, "len_title": 104, "len_sents": 33785, "len_sents_1": 26787, "n_sents": 240, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1641, "title": "CamI2V: Camera-Controlled Image-to-Video Diffusion Model", "abs": "Recent advancements have integrated camera pose as a user-friendly and physics-informed condition in video diffusion models, enabling precise camera control. In this paper, we identify one of the key challenges as effectively modeling noisy cross-frame interactions to enhance geometry consistency and camera controllability. We innovatively associate the quality of a condition with its ability to reduce uncertainty and interpret noisy cross-frame features as a form of noisy condition. Recognizing that noisy conditions provide deterministic information while also introducing randomness and potential misguidance due to added noise, we propose applying epipolar attention to only aggregate features along corresponding epipolar lines, thereby accessing an optimal amount of noisy conditions. Additionally, we address scenarios where epipolar lines disappear, commonly caused by rapid camera movements, dynamic objects, or occlusions, ensuring robust performance in diverse environments.\nFurthermore, we develop a more robust and reproducible evaluation pipeline to address the inaccuracies and instabilities of existing camera control metrics. Our method achieves a 25.64\\% improvement in camera controllability on the RealEstate10K dataset without compromising dynamics and generation quality, also demonstrating strong generalization to out-of-domain images. Training and inference require only 24GB and 12GB of memory, respectively, for 16-frame sequences at 256\u00d7256 resolution. We will release all checkpoints, along with training and evaluation code. Dynamic videos are available for viewing on our supplementary anonymous web page.", "keywords": [ "Camera Control", "Image-to-video", "Video Diffusion Model" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "dIY0vwNyH4", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Yifan wang", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Yun Fu", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 24, "n_ref": 53, "n_ref_all": 66, "n_fig": 1, "n_tab": 3, "L_tab": 1487, "n_element_tab": 208, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1520, "n_element_tab_1": 216, "formula_len_all": 155, "formula_len_all_1": 153, "len_all": 105662, "len_all_1": 63732, "len_abs": 961, "len_title": 103, "len_sents": 28554, "len_sents_1": 26381, "n_sents": 286, "n_sents_1": 274, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 972, "title": "Understanding, Abstracting and Checking: Evoking Complicated Multimodal Reasoning in LMMs", "abs": "The recent large multimodal models (LMMs) have demonstrated their impressive capability of image understanding. However, they still struggle to make complicated reasoning for solving a challenging multimodal problem. In this paper, we present UnAC (Understanding, Abstracting, and Checking), a novel multimodal prompting method, to synergize reasoning for complicated problems in the multimodal context of LMMs, such as GPT-4o, Gemini-1.5 and GPT-4V. To improve the understanding of the image and capture more details, we propose an adaptive visual prompting method to make LMMs able to focus on certain regions. An image abstracting prompting is designed to effectively extract information from images. Further, we propose a gradual self-checking scheme for leading to better reasoning by checking each decomposed sub-question and its answer. Extensive experiments on three public benchmarks -- MathVista, MM-Vet, and MMMU -- demonstrate the effectiveness of our method.", "keywords": [ "Complicated Reasoning", "Large multimodal model" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "dIK7GpOwNY", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "David Khachaturov", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Robert Mullins", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 28, "n_ref": 46, "n_ref_all": 52, "n_fig": 4, "n_tab": 1, "L_tab": 417, "n_element_tab": 6, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 106, "formula_len_all_1": 106, "len_all": 74567, "len_all_1": 64435, "len_abs": 1750, "len_title": 133, "len_sents": 22622, "len_sents_1": 21729, "n_sents": 150, "n_sents_1": 144, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1360, "title": "Complexity Matters: Effective Dimensionality as a Measure for Adversarial Robustness", "abs": "Quantifying robustness in a single measure for the purposes of model selection, development of adversarial training methods, and anticipating trends has so far been elusive. The simplest metric to consider is the number of trainable parameters in a model but this has previously been shown to be insufficient at explaining robustness properties. A variety of other metrics, such as ones based on boundary thickness and gradient flatness have been proposed but have been shown to be inadequate proxies for robustness.\n\nIn this work, we investigate the relationship between a model's $\\textit{effective dimensionality}$, which can be thought of as model complexity, and its robustness properties. We run experiments on commercial-scale models that are often used in real-world environments such as YOLO and ResNet. We reveal a near-linear inverse relationship between effective dimensionality and adversarial robustness, that is models with a lower dimensionality exhibit better robustness. We investigate the effect of a variety of adversarial training methods on effective dimensionality and find the same inverse linear relationship present, suggesting that effective dimensionality can serve as a useful criterion for model selection and robustness evaluation, providing a more nuanced and effective metric than parameter count or previously-tested measures.", "keywords": [ "adversarial machine learning", "complexity", "robustness", "scale" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 3, 1, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "dIK0EfZFO9", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Markus Nagel", "gender": "unknown", "institution": "Qualcomm AI Research", "country": "NL", "position": "Researcher" }, { "name": "Riccardo Del Chiaro", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "NL", "position": "Researcher" }, { "name": "Yelysei Bondarenko", "gender": "unknown", "institution": "Qualcomm Inc, QualComm", "country": "", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 59, "n_ref": 122, "n_ref_all": 163, "n_fig": 2, "n_tab": 11, "L_tab": 10237, "n_element_tab": 1475, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 865, "formula_len_all_1": 488, "len_all": 215575, "len_all_1": 65799, "len_abs": 1348, "len_title": 93, "len_sents": 47840, "len_sents_1": 32192, "n_sents": 366, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1057, "title": "Low-Rank Quantization-Aware Training for LLMs", "abs": "In this paper we propose LR-QAT \u2013 a lightweight and memory-efficient QAT algorithm for LLMs. LR-QAT employs several components to save memory without sacrificing performance: (a) low-rank quantization-aware reparameterization; (b) downcasting operation using fixed-point or double-packing and (c) checkpointing. Unlike most related work, our method (i) is inference-efficient, leading to no additional overhead compared to traditional PTQ; (ii) can be seen as a general extended pre-training framework, meaning that the resulting model can still be utilized for any downstream task afterwards; (iii) is orthogonal to most of recent PTQ methods and thus can be seamlessly combined with them. We apply LR-QAT to the LLaMA-1/2/3 and Mistral model families and validate its effectiveness on several downstream tasks. Our method outperforms most of recent LLM quantization approaches and reaches the same model performance as full-model QAT at the fraction of its memory usage. Specifically, we can train a 7B LLM on a single consumer grade GPU with 24GB memory.", "keywords": [ "transformers", "LLM", "quantization", "quantization-aware training", "QAT", "low-rank adaptation", "PEFT", "memory efficiency", "inference efficiency" ], "rating_list": [ 5, 5, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "dI66AEIo6T", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Ben R. Newell", "gender": "unknown", "institution": "University of New South Wales", "country": "AU", "position": "Full Professor" }, { "name": "Dirk U. Wulff", "gender": "Male", "institution": "Max-Planck Institute", "country": "DE", "position": "Principal Researcher" }, { "name": "Rui Mata", "gender": "unknown", "institution": "University of Basel", "country": "CH", "position": "Associate Professor" }, { "name": "Zak Hussain", "gender": "unknown", "institution": "University of Basel", "country": "CH", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 29, "n_ref": 51, "n_ref_all": 63, "n_fig": 3, "n_tab": 3, "L_tab": 3968, "n_element_tab": 121, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 3259, "n_element_tab_1": 103, "formula_len_all": 102, "formula_len_all_1": 102, "len_all": 115757, "len_all_1": 62218, "len_abs": 998, "len_title": 139, "len_sents": 30888, "len_sents_1": 27355, "n_sents": 230, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1001, "title": "Probing the contents of text, behavior, and brain data toward improving human-LLM alignment", "abs": "Large language models (LLMs) are traditionally trained on massive digitized text corpora; however, alternative data sources exist that may help evaluate and improve the alignment between language models and humans. We contribute to the assessment of the role of data sources in human-LLM alignment. Specifically, we present work aimed at understanding differences in the informational content of text, behavior (e.g., free associations), and brain (e.g., fMRI) data. Using representational similarity analysis, we show that word vectors derived from behavior and brain data encode information that differs from their text-derived cousins. Furthermore, using an interpretability method that we term representational content analysis, we find that, in particular, behavior representations better encode certain affective, agentic, and socio-moral dimensions. The findings highlight the potential of behavior data to evaluate and improve language models along dimensions critical for human-LLM alignment.", "keywords": [ "LLMs", "language models", "interpretability", "free association", "human-model alignment", "representational similarity analysis" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "dI3IjAuu9V", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Young Kyun Jang", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Postdoc" }, { "name": "Ser-Nam Lim", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 4, "n_ref_uni": 27, "n_ref": 53, "n_ref_all": 72, "n_fig": 5, "n_tab": 4, "L_tab": 427, "n_element_tab": 43, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1233, "n_element_tab_1": 127, "formula_len_all": 549, "formula_len_all_1": 419, "len_all": 116730, "len_all_1": 63184, "len_abs": 1404, "len_title": 138, "len_sents": 32459, "len_sents_1": 29178, "n_sents": 250, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1412, "title": "Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models", "abs": "Modern retrieval systems often struggle with upgrading to new and more powerful models due to the incompatibility of embeddings between the old and new models. This necessitates a costly process known as backfilling, which involves re-computing the embeddings for a large number of data samples. In vision, Backward-compatible Training (BT) has been proposed to ensure that the new model aligns with the old model's embeddings. This paper extends the concept of vision-only BT to the field of cross-modal retrieval, marking the first attempt to address Cross-modal BT (XBT). Our goal is to achieve backward-compatibility between Vision-Language Pretraining (VLP) models, such as CLIP, for the cross-modal retrieval task. To address XBT challenges, we propose an efficient solution: a projection module that maps the new model's embeddings to those of the old model. This module, pretrained solely with text data, significantly reduces the number of image-text pairs required for XBT learning, and, once it is pretrained, it avoids using the old model during training. Furthermore, we utilize parameter-efficient training strategies that improve efficiency and preserve the off-the-shelf new model's knowledge by avoiding any modifications. Experimental results on cross-modal retrieval datasets demonstrate the effectiveness of XBT and its potential to enable backfill-free upgrades when a new VLP model emerges.", "keywords": [ "backward compatible training", "cross-modal representation learning" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "dHYwfV2KeP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jaesik Park", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Assistant Professor" }, { "name": "Seungjoo Shin", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sunghyun Cho", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Associate Professor" } ], "n_formula": 22, "n_formula_1": 13, "n_ref_uni": 36, "n_ref": 136, "n_ref_all": 168, "n_fig": 11, "n_tab": 13, "L_tab": 9961, "n_element_tab": 1604, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2189, "n_element_tab_1": 243, "formula_len_all": 2148, "formula_len_all_1": 891, "len_all": 186829, "len_all_1": 69775, "len_abs": 1156, "len_title": 119, "len_sents": 53406, "len_sents_1": 31453, "n_sents": 368, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1193, "title": "Locality-aware Gaussian Compression for Fast and High-quality Rendering", "abs": "We present LocoGS, a locality-aware 3D Gaussian Splatting (3DGS) framework that exploits the spatial coherence of 3D Gaussians for compact modeling of volumetric scenes.\nTo this end, we first analyze the local coherence of 3D Gaussian attributes, and propose a novel locality-aware 3D Gaussian representation that effectively encodes locally-coherent Gaussian attributes using a neural field representation with a minimal storage requirement.\nOn top of the novel representation, LocoGS is carefully designed with additional components such as dense initialization, an adaptive spherical harmonics bandwidth scheme and different encoding schemes for different Gaussian attributes to maximize compression performance.\nExperimental results demonstrate that our approach outperforms the rendering quality of existing compact Gaussian representations for representative real-world 3D datasets while achieving from 54.6$\\times$ to 96.6$\\times$ compressed storage size and from 2.1$\\times$ to 2.4$\\times$ rendering speed than 3DGS. Even our approach also demonstrates an averaged 2.4$\\times$ higher rendering speed than the state-of-the-art compression method with comparable compression performance.", "keywords": [ "point-based rendering", "neural fields", "compression" ], "rating_list": [ 3, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "dHAPEcxyLv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Mou Hao", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Pingfu Chao", "gender": "Male", "institution": "Soochow University", "country": "CN", "position": "Associate Professor" }, { "name": "Haiqin Yang", "gender": "Male", "institution": "International Digital Economy Academy (IDEA)", "country": "", "position": "Principal Researcher" }, { "name": "daiyaxun", "gender": "Male", "institution": "Soochow University", "country": "CN", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 23, "n_ref": 165, "n_ref_all": 231, "n_fig": 27, "n_tab": 37, "L_tab": 21183, "n_element_tab": 937, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 1614, "n_element_tab_1": 39, "formula_len_all": 106, "formula_len_all_1": 106, "len_all": 279805, "len_all_1": 56051, "len_abs": 1498, "len_title": 99, "len_sents": 64689, "len_sents_1": 26426, "n_sents": 569, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1571, "title": "READ-SQL: Reasoning Path Decomposer for Text-to-SQL", "abs": "Text-to-SQL is a longstanding task aimed at automatically converting natural language questions into SQL queries for database retrieval. Despite impressive advancements, particularly with Large Language Models (LLMs), existing methods still struggle with issues such as misinterpreted, omitted, or unwanted constraints. To address these challenges, we propose READ-SQL, a novel framework employing a \\underline{re}asoning p\\underline{a}th \\underline{d}compos\\underline{er}, \\textbf{READ}ER, for text-to-SQL tasks. READER decomposes SQLs into clauses, sub-SQLs, and reasoning paths, supporting data preparation and confidence level determination in post-processing. READ-SQL comprises two main models: a Generator and a Corrector, both trained via LoRA for parameter efficiency. Based on READER's decomposition, READ-SQL generates two types of augmented data using an LLM: question/SQL pairs and question/reason pairs. The Generator is trained on both original and augmented data to identify constraint changes and enhance reasoning. The Corrector is trained on data from READER\u2019s post-processing, improving self-correction by refining high-confidence SQLs and addressing low-confidence elements. Extensive experiments show that READ-SQL significantly outperforms leading baselines, with READ-SQL-3B achieving 57.37\\% execution accuracy on BIRD\u2019s dev set, surpassing several 7B-parameter models and setting a new state-of-the-art with fewer parameters. Additionally, READER and the Corrector show broad applicability when integrated with LLMs or other base models.", "keywords": [ "Text-to-SQL", "Tabular Reasoning", "SQL Decomposition", "Abstract Syntax Trees", "Self-Correction" ], "rating_list": [ 3, 3, 5, 6, 5 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "dGVZwyq5tV", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Ben Athiwaratkun", "gender": "Male", "institution": "Together AI", "country": "US", "position": "Staff AI Scientist" }, { "name": "Han Guo", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "James Liu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Pragaash Ponnusamy", "gender": "Male", "institution": "Together AI", "country": "US", "position": "Principal Researcher" }, { "name": "Tianle Cai", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Yoon Kim", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" } ], "n_formula": 20, "n_formula_1": 6, "n_ref_uni": 61, "n_ref": 87, "n_ref_all": 105, "n_fig": 8, "n_tab": 7, "L_tab": 2800, "n_element_tab": 389, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 574, "n_element_tab_1": 117, "formula_len_all": 1356, "formula_len_all_1": 216, "len_all": 215165, "len_all_1": 64546, "len_abs": 1017, "len_title": 106, "len_sents": 35699, "len_sents_1": 28704, "n_sents": 294, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 58, "L_abs": 1043, "title": "Training-Free Activation Sparsity in Large Language Models", "abs": "Activation sparsity can enable practical inference speedups in large language models (LLMs) by reducing the compute and memory-movement required for matrix multiplications during the forward pass. \nHowever, existing methods face limitations that inhibit widespread adoption. Some approaches are tailored towards older models with ReLU-based sparsity, while others require extensive continued pre-training on up to hundreds of billions of tokens. \nThis paper describes TEAL (**T**raining-Fre**e** **A**ctivation Sparsity in **L**LMs), a simple training-free method that applies magnitude-based activation sparsity to hidden states throughout the entire model. TEAL achieves 40-50\\% model-wide sparsity with minimal performance degradation across Llama-2, Llama-3, and Mistral families, with sizes varying from 7B to 70B. We improve existing sparse kernels and demonstrate wall-clock decoding speed-ups of up to 1.53\u00d7 and 1.8\u00d7 at 40\\% and 50\\% model-wide sparsity. TEAL is compatible with weight quantization, enabling further efficiency gains.", "keywords": [ "Large Language Models", "Activation Sparsity", "Efficiency" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "dGSOn7sdWg", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Alan Baade", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Undergrad student" }, { "name": "David Harwath", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "Assistant Professor" }, { "name": "Puyuan Peng", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 58, "n_ref": 128, "n_ref_all": 148, "n_fig": 2, "n_tab": 5, "L_tab": 2064, "n_element_tab": 190, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1487, "n_element_tab_1": 129, "formula_len_all": 627, "formula_len_all_1": 596, "len_all": 176196, "len_all_1": 65679, "len_abs": 1411, "len_title": 117, "len_sents": 40785, "len_sents_1": 30819, "n_sents": 255, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1369, "title": "SyllableLM: Learning Coarse Semantic Units for Speech Language Models", "abs": "Language models require tokenized inputs. However, tokenization strategies for continuous data like audio and vision are often based on simple heuristics such as fixed sized convolutions or discrete clustering, which do not necessarily align with the semantic structure of the data. For speech in particular, the high resolution of waveforms (16,000 samples/second or more) presents a significant challenge as speech-based language models have had to use several times more tokens per word than text-based language models. In this work, we introduce a controllable self-supervised technique to merge speech representations into coarser syllable-like units while still preserving semantic information. We do this by 1) extracting noisy boundaries through analyzing correlations in pretrained encoder losses and 2) iteratively improving model representations with a novel distillation technique. Our method produces controllable-rate semantic units at as low as 5Hz and 60bps and achieves SotA in syllabic segmentation and clustering. Using these coarse tokens, we successfully train SyllableLM, a Speech Language Model (SpeechLM) that matches or outperforms current SotA SpeechLMs on a range of spoken language modeling tasks. SyllableLM also achieves significant improvements in efficiency with a 30x reduction in training compute and a 4x wall-clock inference speedup.", "keywords": [ "Generative Spoken Language Modeling", "Audio", "Textless NLP", "Representation Learning" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "dGMJ93qpfq", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Josef Kittler", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Full Professor" }, { "name": "Srinivasa Rao Nandam", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "PhD student" }, { "name": "Zhen-Hua Feng", "gender": "Male", "institution": "Jiangnan University", "country": "CN", "position": "Full Professor" }, { "name": "Muhammad Awais Tanvir Rana", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Associate Professor" }, { "name": "Sara Atito Ahmed", "gender": "Female", "institution": "University of Surrey", "country": "GB", "position": "Postdoc" } ], "n_formula": 18, "n_formula_1": 14, "n_ref_uni": 41, "n_ref": 114, "n_ref_all": 135, "n_fig": 2, "n_tab": 15, "L_tab": 7032, "n_element_tab": 691, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 889, "n_element_tab_1": 143, "formula_len_all": 908, "formula_len_all_1": 722, "len_all": 151469, "len_all_1": 69340, "len_abs": 2027, "len_title": 90, "len_sents": 40570, "len_sents_1": 31385, "n_sents": 298, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 2053, "title": "A PATCH LEVEL PERSPECTIVE OF PROMPT TUNING", "abs": "Prompt tuning is an efficient way to adapt large foundation models, such as CLIP, by introducing learnable prompts with the input data tokens, offering a practical alternative to full model finetuning. However, when prompts are trained on base/target tasks, they often overfit, leading to reduced performance on novel, unseen tasks. To address this limitation, various techniques leverage global image semantics to improve accuracy on unseen tasks while maintaining performance on base tasks. However, they often overlook the rich fine-grained local information that could be crucial for capturing finer semantics and improving generalisation. In this work, we propose a modular approach to prompt tuning that leverages local semantics by incorporating patch-level information, representing the first integration of such semantics in this context. Specifically, we integrate patch-level information across vision, text, and predictions through three consistency mechanisms: 1) Patch-based consistency loss that aligns patches from the prompted input image with those from the same image processed by a frozen model, while also enforcing inter-view consistency by applying the loss across different views, capturing fine-grained regional dependencies and improving vision representation quality, 2) Text prompt consistency loss, where view-specific text prompts are tailored and regularised to maintain coherence across views, and 3) Vision features for each view, enriched with patch-level information, are used to generate predictions based on view-tailored text features. These predictions are then regularised across views, complementing the earlier consistency mechanisms and contributing to a cohesive overall framework. Our approach outperforms existing methods across multiple benchmarks, including base-to-novel generalisation, domain generalisation, and cross-dataset evaluation. These results underscore the potential of integrating fine-grained details for more robust and adaptable prompts, marking a step forward in foundation model tuning.", "keywords": [ "Vision Language models", "Efficient Adaptation", "Prompt Tuning", "Few-shot learning" ], "rating_list": [ 3, 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3, 4 ], "presentation_list": [ 3, 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "dEypApI1MZ", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Alexander Atanasov", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Blake Bordelon", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Cengiz Pehlevan", "gender": "unknown", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "", "position": "Assistant Professor" } ], "n_formula": 89, "n_formula_1": 22, "n_ref_uni": 57, "n_ref": 123, "n_ref_all": 160, "n_fig": 15, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 1243, "n_element_tab_1": 15, "formula_len_all": 8923, "formula_len_all_1": 2016, "len_all": 205783, "len_all_1": 68006, "len_abs": 1187, "len_title": 100, "len_sents": 57950, "len_sents_1": 28669, "n_sents": 480, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 52, "L_abs": 1148, "title": "How Feature Learning Can Improve Neural Scaling Laws", "abs": "We develop a simple solvable model of neural scaling laws beyond the kernel limit. Theoretical analysis of this model predicts the performance scaling predictions with model size, training time and total amount of available data. From the scaling analysis we identify three relevant regimes: hard tasks, easy tasks, and super easy tasks. For easy and super-easy target functions, which are in the Hilbert space (RKHS) of the initial infinite-width neural tangent kernel (NTK), there is no change in the scaling exponents between feature learning models and models in the kernel regime. For hard tasks, which we define as tasks outside of the RKHS of the initial NTK, we show analytically and empirically that feature learning can improve the scaling with training time and compute, approximately doubling the exponent for very hard tasks. This leads to a new compute optimal scaling law for hard tasks in the feature learning regime. We support our finding that feature learning improves the scaling law for hard tasks with experiments of nonlinear MLPs fitting functions with power-law Fourier spectra on the circle and CNNs learning vision tasks.", "keywords": [ "neural scaling laws", "feature learning", "kernel methods", "linear networks", "mean field theory" ], "rating_list": [ 8, 8, 6, 6, 8 ], "soundness_list": [ 2, 4, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 1, 3 ], "contribution_list": [ 3, 3, 3, 2, 4 ], "confidence_list": [ 3, 3, 3, 3, 4 ] }, { "paper_id": "dEg5SdGaiq", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Christopher Hoang", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Mengye Ren", "gender": "unknown", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Yann LeCun", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Director" }, { "name": "Yuwen Xiong", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Alexander Wang", "gender": "Male", "institution": "NYU, New York University", "country": "US", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 38, "n_ref": 107, "n_ref_all": 149, "n_fig": 20, "n_tab": 10, "L_tab": 4872, "n_element_tab": 734, "n_fig_1": 14, "n_tab_1": 6, "L_tab_1": 1995, "n_element_tab_1": 308, "formula_len_all": 266, "formula_len_all_1": 229, "len_all": 164716, "len_all_1": 65674, "len_abs": 977, "len_title": 123, "len_sents": 44217, "len_sents_1": 29247, "n_sents": 346, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 984, "title": "PooDLe\ud83d\udc29: Pooled and dense self-supervised learning from naturalistic videos", "abs": "Self-supervised learning has driven significant progress in learning from single-subject, _iconic_ images.\nHowever, there are still unanswered questions about the use of minimally-curated, naturalistic video data, which contain _dense_ scenes with many independent objects, imbalanced class distributions, and varying object sizes.\nIn this paper, we propose PooDLe, a self-supervised learning method that combines an invariance-based objective on pooled representations with a dense SSL objective that enforces equivariance to optical flow warping.\nOur results show that a unified objective applied at multiple feature scales is essential for learning effective image representations from naturalistic videos.\nWe validate our method with experiments on the BDD100K driving video dataset and the Walking Tours first-person video dataset, demonstrating its ability to capture spatial understanding from a dense objective and semantic understanding via a pooled representation objective.", "keywords": [ "computer vision", "representation learning", "self-supervised learning", "egocentric video", "visual representation" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "dEGYODD6iU", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Anudeep Vurity", "gender": "Male", "institution": "George Mason University", "country": "US", "position": "PhD student" }, { "name": "Emanuela Marasco", "gender": "Female", "institution": "George Mason University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 26, "n_ref": 30, "n_ref_all": 41, "n_fig": 6, "n_tab": 4, "L_tab": 1889, "n_element_tab": 118, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 832, "n_element_tab_1": 117, "formula_len_all": 147, "formula_len_all_1": 223, "len_all": 139704, "len_all_1": 59375, "len_abs": 1879, "len_title": 199, "len_sents": 36529, "len_sents_1": 29689, "n_sents": 237, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 149, "L_abs": 1899, "title": "\u201dDOES YOUR MOBILE SUIT YOUR SKIN?\u201d: ADDRESSING SKIN TONE DISPARITIES IN PRESENTATION ATTACK DETECTION FOR ENHANCED INCLUSIVITY OF SMARTPHONE SECURITY", "abs": "Mobile devices are at a heightened risk for cybercrime due to the sensitive personal and financial data they handle. Biometric authentication provides a robust,convenient, and secure way to protect smartphones by using unique user characteristics like fingerprints, facial features, or voice patterns for access. Existing mobile biometric technology often relies on RGB cameras to capture biometric\nsamples, such as face images or finger photos, making them vulnerable to spoofing (e.g., 3D masks, display or printout attack). The security of these systems is effectively addressed by integrating a Presentation Attack Detection (PAD) module. Existing PAD solutions do not account for diverse physical characteristics such as skin tone. As a result, marginalized groups face higher misidentification\nrates or false rejections, reducing access to services and increasing security risks.\n\nThis paper introduces a novel deep learning framework called ColorCubeNet that is designed to process ColorCube, a multi dimensional data representation by combining information from RGB, HSV and YCbCr color spaces. This data cube leverages the joint capabilities of RGB, HSV, and YCbCr color spaces to depict color more sophisticatedly. By incorporating features from multiple complementary color channels, this approach can effectively handle a variety of skin tones. We utilized three EfficientNet-B0 models, each trained on ImageNet using RGB, HSV, and YCbCr color spaces, and then fine-tune them on the ColorCube representation to fully exploit the combined information from all three color spaces. Additionally, a channel-attention mechanism is integrated into the architecture, enabling the extraction of key features from different input channels and exploit their combined performance. Results show that the proposed approach outperforms traditional RGB methods by reducing skin tone disparities by 50%.", "keywords": [ "Skin Tone", "Fairness", "Mobile Biometrics", "Finger Photo", "Presentation Attack Detection", "Presentation Attack", "Color Spaces" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 3, 5, 4 ] }, { "paper_id": "dDpB23VbVa", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Chenze Shao", "gender": "Male", "institution": "Tencent Inc", "country": "CN", "position": "Researcher" }, { "name": "Fandong Meng", "gender": "Male", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Jie Zhou", "gender": "Male", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 61, "n_ref": 75, "n_ref_all": 100, "n_fig": 11, "n_tab": 9, "L_tab": 2720, "n_element_tab": 295, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 1769, "n_element_tab_1": 194, "formula_len_all": 87, "formula_len_all_1": 87, "len_all": 167589, "len_all_1": 58760, "len_abs": 1059, "len_title": 95, "len_sents": 36719, "len_sents_1": 26962, "n_sents": 254, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 76, "L_abs": 1131, "title": "Beyond Next Token Prediction: Patch-Level Training for Large Language Models", "abs": "The prohibitive training costs of Large Language Models (LLMs) have emerged as a significant bottleneck in the development of next-generation LLMs. In this paper, we show that it is possible to significantly reduce the training costs of LLMs without sacrificing their performance. Specifically, we introduce patch-level training for LLMs, in which multiple tokens are aggregated into a unit of higher information density, referred to as a `patch', to serve as the fundamental text unit for training LLMs. During patch-level training, we feed the language model shorter sequences of patches and train it to predict the next patch, thereby processing the majority of the training data at a significantly reduced cost. Following this, the model continues token-level training on the remaining training data to align with the inference mode. Experiments on a diverse range of models (370M-2.7B parameters) demonstrate that patch-level training can reduce the overall training costs to 0.5$\\times$, without compromising the model performance compared to token-level training. Source code: \\url{https://github.com/shaochenze/PatchTrain}.", "keywords": [ "large language models", "patch-level training" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "dDdxbdhMsY", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "David Bergstr\u00f6m", "gender": "Male", "institution": "Link\u00f6ping University", "country": "SE", "position": "PhD student" }, { "name": "Fredrik Heintz", "gender": "Male", "institution": "Link\u00f6ping University", "country": "SE", "position": "Full Professor" }, { "name": "Mattias Tiger", "gender": "Male", "institution": "Link\u00f6ping University", "country": "", "position": "Postdoc" } ], "n_formula": 8, "n_formula_1": 3, "n_ref_uni": 28, "n_ref": 64, "n_ref_all": 93, "n_fig": 9, "n_tab": 4, "L_tab": 2967, "n_element_tab": 380, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 589, "n_element_tab_1": 66, "formula_len_all": 334, "formula_len_all_1": 155, "len_all": 124569, "len_all_1": 57136, "len_abs": 1083, "len_title": 74, "len_sents": 39875, "len_sents_1": 27141, "n_sents": 310, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1090, "title": "Deep Temporal Deaggregation: Large-Scale Spatio-Temporal Generative Models", "abs": "Access to spatio-temporal trajectory data is essential for improving infrastructure, preventing the spread of disease and for building autonomous vehicles. However, it remains underutilized due to limited availability, as it cannot be shared publicly due privacy concerns or other sensitive attributes. Generative time-series models have shown promise in generating non-sensitive data, but show poor performance for large-scale and complex environments. In this paper we propose a spatio-temporal generative model for trajectories, TDDPM, which outperforms and scales substantially better than state-of-the-art. The focus is primarily on trajectories of peoples' movement in cities. We propose a conditional distribution approach which unlock out-of-distribution generalization, such as to city-areas not trained on, from a spatial aggregate prior. We also show that data can be generated in a privacy-preserving manner using $k$-anonymity. Further, we propose a new comprehensive benchmark across several standard datasets, and evaluation measures, considering key distribution properties.", "keywords": [ "generative learning", "mobility data", "denoising diffusion probablistic models" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "dDLGZTKZYZ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Natalie Abreu", "gender": "unknown", "institution": "Harvard University, Harvard University", "country": "", "position": "PhD student" }, { "name": "eran malach", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Benjamin L Edelman", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 28, "n_ref": 40, "n_ref_all": 61, "n_fig": 6, "n_tab": 2, "L_tab": 320, "n_element_tab": 28, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 260, "formula_len_all_1": 208, "len_all": 132045, "len_all_1": 57416, "len_abs": 1345, "len_title": 109, "len_sents": 41408, "len_sents_1": 28961, "n_sents": 344, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1353, "title": "MLPs for NLP: Towards Discovering Inductive Bias From Scratch", "abs": "The recent rise of large language models has been fueled by scale. More data, more compute, and bigger models have consistently led to better performance. This scaling paradigm has been applied most notably to the transformer architecture, which is especially conducive to training parallelization and sequence modeling. In this work, we ask what happens if we apply the power of scale to the simplest possible class of neural network: the multi-layer perceptron (MLP). Specifically, we train MLPs to perform next-token prediction on billions of tokens of text. Indeed, their performance consistently improves with scale, though vanilla MLPs are still clearly inferior to transformers for this task, especially because their parameter count grows with the length of the input sequences. We then perform a mechanistic analysis of the trained models, and identify a consistent emergent structure: most neurons in the first hidden layer either perform arbitrary linear functions over a small look-back window, or low-frequency functions over the entire context. These neuron types recall $n$-gram and bag-of-words techniques from classical statistical language modeling. Using the discrete cosine transform, we define a unified way of reparameterizing these neuron types such that the number of parameters per neuron does not depend on the sequence length.", "keywords": [ "Language Models", "Architecture Design", "MLP", "Inductive Bias", "Scaling Laws" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "dD6b5RREws", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Jacek Ma\u0144dziuk", "gender": "Male", "institution": "AGH University of Science and Technology", "country": "PL", "position": "Full Professor" }, { "name": "Stanis\u0142aw Ka\u017amierczak", "gender": "unknown", "institution": "Warsaw University of Technology", "country": "PL", "position": "Lecturer" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 14, "n_ref": 30, "n_ref_all": 38, "n_fig": 3, "n_tab": 39, "L_tab": 41180, "n_element_tab": 3559, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2878, "n_element_tab_1": 437, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 177855, "len_all_1": 66358, "len_abs": 1321, "len_title": 78, "len_sents": 31939, "len_sents_1": 29378, "n_sents": 286, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1354, "title": "Bootstrap Sampling Rate Greater than 1.0 May Improve Random Forest Performance", "abs": "Random forests utilize bootstrap sampling to create an individual training set for each component tree. This involves sampling with replacement, with the number of instances equal to the size of the original training set ($N$). Research literature indicates that drawing fewer than $N$ observations can also yield satisfactory results. The ratio of the number of observations in each bootstrap sample to the total number of training instances is called the bootstrap rate (BR). Sampling more than $N$ observations (BR $>$ 1) has been explored in the literature only to a limited extent and has generally proven ineffective. In this paper, we re-examine this approach using 36 diverse datasets and consider BR values ranging from 1.2 to 5.0. Contrary to previous findings, we show that such parameterization can result in statistically significant improvements in classification accuracy compared to standard settings (BR $\\leq$ 1). Furthermore, we investigate what the optimal BR depends on and conclude that it is more a property of the dataset than a dependence on the random forest hyperparameters. Finally, we develop a binary classifier to predict whether the optimal BR is $\\leq$ 1 or $>$ 1 for a given dataset, achieving between 81.88\\% and 88.81\\% accuracy, depending on the experiment configuration. The code is available at: $<$placeholder$>$.", "keywords": [ "random forest", "bootstrap sampling rate" ], "rating_list": [ 3, 1, 5, 1 ], "soundness_list": [ 2, 3, 2, 1 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 4, 5, 3, 5 ] }, { "paper_id": "dCcY2pyNIO", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Jiecheng Lu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Shihao Yang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Yan Sun", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 37, "n_ref": 108, "n_ref_all": 143, "n_fig": 12, "n_tab": 13, "L_tab": 22875, "n_element_tab": 1457, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1533, "n_element_tab_1": 337, "formula_len_all": 1517, "formula_len_all_1": 1397, "len_all": 209099, "len_all_1": 69901, "len_abs": 883, "len_title": 80, "len_sents": 58675, "len_sents_1": 30320, "n_sents": 422, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 32, "L_abs": 880, "title": "In-context Time Series Predictor", "abs": "Recent Transformer-based large language models (LLMs) demonstrate in-context learning ability to perform various functions based solely on the provided context, without updating model parameters. To fully utilize the in-context capabilities in time series forecasting (TSF) problems, unlike previous Transformer-based or LLM-based time series forecasting methods, we reformulate \"time series forecasting tasks\" as input tokens by constructing a series of (lookback, future) pairs within the tokens. This method aligns more closely with the inherent in-context mechanisms and is more parameter-efficient without the need of using pre-trained LLM parameters. Furthermore, it addresses issues such as overfitting in existing Transformer-based TSF models, consistently achieving better performance across full-data, few-shot, and zero-shot settings compared to previous architectures.", "keywords": [ "Time Series Forecasting", "In-context Learning", "Transformer" ], "rating_list": [ 8, 6, 8, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "dCTGFl3lN2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Dylan Zhang", "gender": "Male", "institution": "Department of Computer Science", "country": "US", "position": "PhD student" }, { "name": "Qirun Dai", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Hao Peng", "gender": "Male", "institution": "Department of Computer Science, University of Illinois Urbana-Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Jiaqi Ma", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 2, "n_ref_uni": 38, "n_ref": 64, "n_ref_all": 84, "n_fig": 7, "n_tab": 5, "L_tab": 3913, "n_element_tab": 415, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2521, "n_element_tab_1": 347, "formula_len_all": 356, "formula_len_all_1": 46, "len_all": 154073, "len_all_1": 64427, "len_abs": 1998, "len_title": 154, "len_sents": 49598, "len_sents_1": 29489, "n_sents": 354, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 105, "L_abs": 2012, "title": "Improving Influence-based Instruction Tuning Data Selection for Balanced Learning of Diverse Capabilities", "abs": "Selecting appropriate training data is crucial for successful instruction fine-tuning, which aims to (1) elicit strong capabilities from pretrained large language models (LLMs), and (2) achieve balanced performance across a diverse range of tasks. Algorithms based on influence estimation have shown promise in achieving (1) through estimating the contribution of each training example to model's prediction on a downstream task, but they often struggle with (2). Through systematic experiments, we attribute their underperformance to an inherent bias---certain tasks intrinsically have greater influence than others. Directly comparing influence scores across different tasks would thus bias the selected data towards these tasks, hurting the LM's performance not only on other capabilities, but also, surprisingly, on the tasks for which the selected data has high influence.\n\nTo address this issue, we propose BIDS, a Balanced and Influential Data Selection algorithm. BIDS first normalizes influence scores of the training data with respect to each downstream task at an instance level. It then applies an iterative process to further balance the selection of influential training data. At each step, BIDS selects the training example that bears the highest influence on the most underrepresented capability by the currently selected data. We perform comprehensive experiments using both Llama-3 and Mistral-v0.3 on seven evaluation benchmarks spanning five diverse capabilities. Results demonstrate that BIDS consistently outperforms both state-of-the-art influence-based data selection algorithms and other non-influence-based selection frameworks under various budgets. Surprisingly, training on a 15% subset selected by BIDS can even outperform full-dataset training with a much more balanced performance across different tasks. Our analysis further highlights the importance of both instance-level normalization and iterative optimization of selected data for balanced learning of diverse capabilities.", "keywords": [ "Instruction Tuning", "Data Selection", "Influence Estimation" ], "rating_list": [ 3, 3, 3, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "dCPF1wlqj8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Hongning Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jian Guan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jiaxin Wen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Minlie Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Wei Wu", "gender": "Male", "institution": "Ant Research", "country": "CN", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 65, "n_ref": 107, "n_ref_all": 154, "n_fig": 4, "n_tab": 15, "L_tab": 3695, "n_element_tab": 463, "n_fig_1": 4, "n_tab_1": 10, "L_tab_1": 3611, "n_element_tab_1": 509, "formula_len_all": 591, "formula_len_all_1": 174, "len_all": 216490, "len_all_1": 77639, "len_abs": 1787, "len_title": 133, "len_sents": 59214, "len_sents_1": 31798, "n_sents": 431, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1664, "title": "Unlocking Reasoning Potential in Large Language Models by Scaling Code-form Planning", "abs": "Despite the remarkable success of large language models (LLMs) on traditional natural language processing tasks, their planning ability remains a critical bottleneck in tackling complex multi-step reasoning tasks. Existing approaches mainly rely on prompting or task-specific fine-tuning, often suffering from weak robustness and cross-task generalization. To address the limitation, we introduce CodePlan, a scalable paradigm that empowers LLMs to generate and follow code-form plans---pseudocode that outlines high-level, structured reasoning processes. By leveraging the structured and versatile nature of code, CodePlan effectively captures the rich semantics and control flows inherent to sophisticated reasoning. Importantly, CodePlan allows the automatic extraction of code-form plans from massive, wide-ranging text corpora without the need for curated, task-specific datasets. This enables it to scale up efficiently and improve reasoning capabilities across diverse scenarios. To train CodePlan, we construct a large-scale dataset of 2M examples that integrate code-form plans with standard prompt-response pairs from existing corpora. With minimal computation overhead during both training and inference, CodePlan achieves a 25.1\\% relative improvement compared with directly generating responses, averaged across 13 challenging multi-step reasoning benchmarks, spanning mathematical reasoning, symbolic reasoning, instruction-following, multi-hop QA, and decision-making tasks. Further analysis reveals CodePlan's increasing performance gains on more complex reasoning tasks, as well as significant data efficiency thanks to its generalization ability.", "keywords": [ "Large langauge model for reasoning", "planning", "code-aided reasoning" ], "rating_list": [ 3, 5, 6, 3, 8 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "dBafcyEQzr", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Bin Shao", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Principal Research Manager" }, { "name": "Guoqing Liu", "gender": "Male", "institution": "Microsoft Research ", "country": "CN", "position": "Researcher" }, { "name": "Hongfei Wu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Lijun Wu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Zhirong Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Zun Wang", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 30, "n_formula_1": 19, "n_ref_uni": 65, "n_ref": 106, "n_ref_all": 130, "n_fig": 11, "n_tab": 9, "L_tab": 4017, "n_element_tab": 303, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1746, "n_element_tab_1": 105, "formula_len_all": 2456, "formula_len_all_1": 1110, "len_all": 180422, "len_all_1": 59497, "len_abs": 1409, "len_title": 144, "len_sents": 47314, "len_sents_1": 26168, "n_sents": 356, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1414, "title": "SE3Set: Harnessing equivariant hypergraph neural networks for molecular representation learning", "abs": "In this paper, we develop SE3Set, an SE(3) equivariant hypergraph neural network architecture tailored for advanced molecular representation learning. Hypergraphs are not merely an extension of traditional graphs; they are pivotal for modeling high-order relationships, a capability that conventional equivariant graph-based methods lack due to their inherent limitations in representing intricate many-body interactions. To achieve this, we first construct hypergraphs via proposing a new fragmentation method that considers both chemical and three-dimensional spatial information of molecular system. We then design SE3Set, which incorporates equivariance into the hypergragh neural network. This ensures that the learned molecular representations are invariant to spatial transformations, thereby providing robustness essential for accurate prediction of molecular properties. SE3Set has shown performance on par with state-of-the-art (SOTA) models for small molecule datasets like QM9 and MD17. It excels on the MD22 dataset, achieving a notable improvement of approximately 20% in accuracy across all molecules, which highlights the prevalence of complex many-body interactions in larger molecules. This exceptional performance of SE3Set across diverse molecular structures underscores its transformative potential in computational chemistry, offering a route to more accurate and physically nuanced modeling.", "keywords": [ "Equivariant", "Hypergraph Neural Networks", "Molecules" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "dAeET8gxqg", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Ilia Sucholutsky", "gender": "Male", "institution": "New York University", "country": "US", "position": "Faculty Fellow/Assistant Professor" }, { "name": "Jiayi Geng", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "MS student" }, { "name": "Joshua C Peterson", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Postdoc" }, { "name": "Ryan Liu", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Thomas L. Griffiths", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 71, "n_ref": 135, "n_ref_all": 161, "n_fig": 2, "n_tab": 7, "L_tab": 2640, "n_element_tab": 414, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1650, "n_element_tab_1": 277, "formula_len_all": 328, "formula_len_all_1": 327, "len_all": 176332, "len_all_1": 72330, "len_abs": 1410, "len_title": 120, "len_sents": 51721, "len_sents_1": 35233, "n_sents": 356, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1423, "title": "Large Language Models Assume People are More Rational than We Really are", "abs": "In order for AI systems to communicate effectively with people, they must understand how we make decisions. However, people's decisions are not always rational, so the implicit internal models of human decision-making in Large Language Models (LLMs) must account for this. Previous empirical evidence seems to suggest that these implicit models are accurate --- LLMs offer believable proxies of human behavior, acting how we expect humans would in everyday interactions. However, by comparing LLM behavior and predictions to a large dataset of human decisions, we find that this is actually not the case: when both simulating and predicting people's choices, a suite of cutting-edge LLMs (GPT-4o \\& 4-Turbo, Llama-3-8B \\& 70B, Claude 3 Opus) assume that people are more rational than we really are. Specifically, these models deviate from human behavior and align more closely with a classic model of rational choice --- expected value theory. Interestingly, people also tend to assume that other people are rational when interpreting their behavior. As a consequence, when we compare the inferences that LLMs and people draw from the decisions of others using another psychological dataset, we find that these inferences are highly correlated. Thus, the implicit decision-making models of LLMs appear to be aligned with the human expectation that other people will act rationally, rather than with how people actually act.", "keywords": [ "Large Language Models", "Rationality", "Cognitive Models", "Psychology" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "dAavOuxZvo", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Gabriel Hope", "gender": "Male", "institution": "University of California, Irvine", "country": "", "position": "PhD student" }, { "name": "Sakshi Agarwal", "gender": "Female", "institution": "University of California, Irvine", "country": "", "position": "PhD student" }, { "name": "Erik Sudderth", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Professor" } ], "n_formula": 20, "n_formula_1": 17, "n_ref_uni": 39, "n_ref": 93, "n_ref_all": 119, "n_fig": 13, "n_tab": 2, "L_tab": 447, "n_element_tab": 78, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 17, "n_element_tab_1": 1, "formula_len_all": 1541, "formula_len_all_1": 1403, "len_all": 129468, "len_all_1": 61128, "len_abs": 1331, "len_title": 133, "len_sents": 32090, "len_sents_1": 28676, "n_sents": 207, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1375, "title": "VIPaint: Image Inpainting with Pre-Trained Diffusion Models via Variational Inference", "abs": "Diffusion probabilistic models learn to remove noise that is artificially added to the data during training. Novel data, like images, may then be generated from Gaussian noise through a sequence of de-noising operations. While this Markov process implicitly defines a joint distribution over noise-free images, it is not simple to condition the generative process on masked observations of partial images. A number of heuristic sampling procedures have been proposed for solving inverse problems with diffusion priors, but these approaches do not directly approximate the true conditional distribution imposed by inference queries, and are often ineffective for high distortion levels and large image masks. Moreover, many of these baselines cannot be applied to latent diffusion models which use image encodings for efficiency. We instead develop a hierarchical variational inference algorithm that analytically marginalizes missing features, and uses a rigorous variational bound to optimize a non-Gaussian Markov approximation of the true diffusion posterior. Through extensive experiments with both pixel-based and latent diffusion models of images, we show that our VIPaint method significantly outperforms previous approaches in both the plausibility and diversity of imputations, and is easily generalized to other inverse problems like deblurring and superresolution.", "keywords": [ "Diffusion Models", "Variational Inference", "Image Inpainting" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 1, 1, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "dAIcU2ZwUN", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Sung Whan Yoon", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Weebum Yoo", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 60, "n_formula_1": 14, "n_ref_uni": 31, "n_ref": 54, "n_ref_all": 74, "n_fig": 3, "n_tab": 11, "L_tab": 3842, "n_element_tab": 344, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2023, "n_element_tab_1": 141, "formula_len_all": 9809, "formula_len_all_1": 1583, "len_all": 196435, "len_all_1": 63005, "len_abs": 1280, "len_title": 115, "len_sents": 56440, "len_sents_1": 25770, "n_sents": 434, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1284, "title": "How Do Augmentations with Label Smoothing Enhance Model Robustness?", "abs": "Model robustness indicates a model's capability to generalize well on unforeseen distributional shifts, including data corruption, adversarial attacks, and domain shifts. One of the most prevalent and effective ways to enhance the robustness often involves data augmentations and label smoothing techniques. Despite the great success of the related approaches in diverse practices, a unified theoretical understanding of their efficacy in improving model robustness is lacking. We offer a theoretical framework to clarify how augmentations, label smoothing, or their combination enhance model robustness through the lens of loss surface flatness, generalization bound, and adversarial robustness. Specifically, we first formally bridge the diversified data distribution via augmentations to the flatter minima on the parameter space, which directly links to the improved generalization capability. Moreover, we further bridge augmentations with label smoothing, which softens the confidence of the target label, to the improved adversarial robustness. We broadly confirm our theories through extensive simulations on the existing common corruption and adversarial robustness benchmarks based on the CIFAR and tinyImageNet datasets, as well as various domain generalization benchmarks.", "keywords": [ "Learning theory", "Model robustness", "Data augmentation", "label smoothing", "Generalization" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 2, 3 ] }, { "paper_id": "d9iHI1eimo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chaoyang Wang", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Chuang Gan", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Evangelos Kalogerakis", "gender": "Male", "institution": "Technical University of Crete", "country": "GR", "position": "Associate Professor" }, { "name": "Hsin-Ying Lee", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Peiye Zhuang", "gender": "Female", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Sergey Tulyakov", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Director of Research" }, { "name": "Tuan Duc Ngo", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 64, "n_ref": 142, "n_ref_all": 170, "n_fig": 7, "n_tab": 6, "L_tab": 1052, "n_element_tab": 170, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 438, "n_element_tab_1": 61, "formula_len_all": 678, "formula_len_all_1": 626, "len_all": 171753, "len_all_1": 59725, "len_abs": 512, "len_title": 93, "len_sents": 43180, "len_sents_1": 27280, "n_sents": 314, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1093, "title": "DELTA: DENSE EFFICIENT LONG-RANGE 3D TRACKING FOR ANY VIDEO", "abs": "Tracking dense 3D motion from monocular videos remains challenging, particularly when aiming for pixel-level precision over long sequences. We introduce DELTA, a novel method that efficiently tracks every pixel in 3D space, enabling accurate motion estimation across entire videos. Our approach leverages a joint global-local attention mechanism for reduced-resolution tracking, followed by a transformer-based upsampler to achieve high-resolution predictions. Unlike existing methods, which are limited by computational inefficiency or sparse tracking, DELTA delivers dense 3D tracking at scale, running over 8x faster than previous methods while achieving state-of-the-art accuracy. Furthermore, we explore the impact of depth representation on tracking performance and identify log-depth as the optimal choice. Extensive experiments demonstrate the superiority of DELTA on multiple benchmarks, achieving new state-of-the-art results in both 2D and 3D dense tracking tasks. Our method provides a robust solution for applications requiring fine-grained, long-term motion tracking in 3D space.", "keywords": [ "3D Computer Vision", "Point Tracking" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 3, 5, 3 ] }, { "paper_id": "d9aWa875kj", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Debarghya Ghoshdastidar", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Assistant Professor" }, { "name": "Lukas Gosch", "gender": "Male", "institution": "Technical University of Munich", "country": "DE", "position": "PhD student" }, { "name": "Mahalakshmi Sabanayagam", "gender": "Female", "institution": "Department of Informatics, Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Stephan G\u00fcnnemann", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Professor" } ], "n_formula": 42, "n_formula_1": 21, "n_ref_uni": 38, "n_ref": 93, "n_ref_all": 180, "n_fig": 26, "n_tab": 12, "L_tab": 4057, "n_element_tab": 420, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 472, "n_element_tab_1": 83, "formula_len_all": 4792, "formula_len_all_1": 2016, "len_all": 216915, "len_all_1": 77639, "len_abs": 1524, "len_title": 71, "len_sents": 67785, "len_sents_1": 33273, "n_sents": 553, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 70, "L_abs": 1536, "title": "Exact Certification of (Graph) Neural Networks Against Label Poisoning", "abs": "Machine learning models are highly vulnerable to label flipping, i.e., the adversarial modification (poisoning) of training labels to compromise performance. Thus, deriving robustness certificates is important to guarantee that test predictions remain unaffected and to understand worst-case robustness behavior. However, for Graph Neural Networks (GNNs), the problem of certifying label flipping has so far been unsolved. We change this by introducing an exact certification method, deriving both sample-wise and collective certificates. Our method leverages the Neural Tangent Kernel (NTK) to capture the training dynamics of wide networks enabling us to reformulate the bilevel optimization problem representing label flipping into a Mixed-Integer Linear Program (MILP). We apply our method to certify a broad range of GNN architectures in node classification tasks. Thereby, concerning the worst-case robustness to label flipping: $(i)$ we establish hierarchies of GNNs on different benchmark graphs; $(ii)$ quantify the effect of architectural choices such as activations, depth and skip-connections; and surprisingly, $(iii)$ uncover a novel phenomenon of the robustness plateauing for intermediate perturbation budgets across all investigated datasets and architectures. While we focus on GNNs, our certificates are applicable to sufficiently wide NNs in general through their NTK. Thus, our work presents the first exact certificate to a poisoning attack ever derived for neural networks, which could be of independent interest.", "keywords": [ "graph neural networks", "robustness", "certificates", "provable robustness", "neural networks", "label poisoning", "label flipping", "poisoning", "mixed-integer linear programming", "neural tangent kernel", "support vector machines" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 4, 3, 2 ], "confidence_list": [ 2, 3, 2, 3 ] }, { "paper_id": "d9JcSrQoeP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Leying Guan", "gender": "unknown", "institution": "Yale University", "country": "", "position": "Assistant Professor" }, { "name": "Yuheng Lai", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" } ], "n_formula": 47, "n_formula_1": 24, "n_ref_uni": 32, "n_ref": 71, "n_ref_all": 103, "n_fig": 10, "n_tab": 4, "L_tab": 1289, "n_element_tab": 118, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1773, "n_element_tab_1": 104, "formula_len_all": 4099, "formula_len_all_1": 1861, "len_all": 145736, "len_all_1": 71245, "len_abs": 959, "len_title": 116, "len_sents": 49423, "len_sents_1": 32599, "n_sents": 395, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 964, "title": "Flexible Fairness-Aware Learning via Inverse Conditional Permutation", "abs": "*Equalized odds*, as a popular notion of algorithmic fairness, aims to ensure that sensitive variables, such as race and gender, do not unfairly influence the algorithm's prediction when conditioning on the true outcome. Despite rapid advancements, current research primarily focuses on equalized odds violations caused by a single sensitive attribute, leaving the challenge of simultaneously accounting for multiple attributes largely unaddressed. We bridge this gap by introducing an in-processing fairness-aware learning approach, FairICP, which integrates adversarial learning with a novel inverse conditional permutation scheme. FairICP offers a theoretically justified, flexible, and efficient scheme to promote equalized odds under fairness conditions described by complex and multi-dimensional sensitive attributes. The efficacy and adaptability of our method are demonstrated through both simulation studies and empirical analyses of real-world datasets.", "keywords": [ "Algorithmic fairness", "Equalized odds", "Adversarial learning", "Inverse conditional permutation" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "d9BMHLXPrr", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Tal Amir", "gender": "unknown", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Researcher" }, { "name": "Yair Davidson", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "MS student" }, { "name": "Yonatan Sverdlov", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "PhD student" }, { "name": "Nadav E Dym", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "Assistant Professor" } ], "n_formula": 51, "n_formula_1": 25, "n_ref_uni": 35, "n_ref": 71, "n_ref_all": 97, "n_fig": 4, "n_tab": 5, "L_tab": 739, "n_element_tab": 146, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 637, "n_element_tab_1": 11, "formula_len_all": 4551, "formula_len_all_1": 1969, "len_all": 162081, "len_all_1": 53959, "len_abs": 1097, "len_title": 106, "len_sents": 43495, "len_sents_1": 24132, "n_sents": 394, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1003, "title": "FSW-GNN: A Bi-Lipschitz WL-Equivalent Graph Neural Network", "abs": "Many of the most popular graph neural networks fall into the category of Message Passing Neural Networks (MPNNs). Famously, MPNNs ability to distinguish between graphs is limited to graphs separable by the Weisfeiler-Leman (WL) graph isomorphism test, and the strongest MPNNs, in terms of separation power, are those which are WL-equivalent. \n\nRecently, it was shown that the quality of separation provided by standard WL-equivalent MPNN can be very low, resulting in WL-separable graphs being mapped to very similar, hardly distinguishable features.\n\nThis paper addresses this issue by seeking bi-Lipschitz continuity guarantees for MPNNs. We demonstrate that, in contrast with standard summation-based MPNNs, which lack bi-Lipschitz properties, our proposed model provides a bi-Lipschitz graph embedding with respect to two standard graph metrics. Empirically, we show that our MPNN is competitive with standard MPNNs for several graph learning tasks, and is far more accurate on long-range tasks.", "keywords": [ "Bi-Lipschitzness", "Weisfeiler Leman", "Graph Embedding", "Graph Neural Network", "MPNN" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "d8hYXbxX71", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Ana-Andreea Stoica", "gender": "Female", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "Postdoc" }, { "name": "Jiduan Wu", "gender": "unknown", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Moritz Hardt", "gender": "Not Specified", "institution": "Max-Planck-Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "Principal Researcher" }, { "name": "Rediet Abebe", "gender": "Not Specified", "institution": "Harvard University", "country": "US", "position": "Principal Researcher" } ], "n_formula": 86, "n_formula_1": 22, "n_ref_uni": 43, "n_ref": 65, "n_ref_all": 79, "n_fig": 6, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7841, "formula_len_all_1": 1729, "len_all": 188724, "len_all_1": 67463, "len_abs": 1496, "len_title": 90, "len_sents": 68072, "len_sents_1": 32112, "n_sents": 544, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 1299, "title": "Policy Design in Long-run Welfare Dynamics", "abs": "We study a stochastic dynamic model of long-term welfare in a population. Individuals in our model have welfare that improves with intervention and deteriorates in the absence of treatment. The planner can treat one individual at each time step. We contrast two fundamental policies in our model. The utilitarian policy greedily maximizes welfare improvement at each step. The Rawlsian policy intervenes on the individual of lowest welfare. Although hugely influential as a normative proposal, Rawlsian policies have been criticized for failing to optimize social welfare. We prove that, surprisingly, in a meaningful range of parameters Rawlsian policy has greater long-run utility than the utilitarian policy even though it is inferior on short time horizons. Specifically, this is true provided that treatment effects satisfy a weak homogeneity assumption, and the welfare dynamics satisfy a rich-get-richer and poor-get-poorer condition. We extend our results with a comprehensive comparison of different policies under different parameter regimes. Through semi-synthetic simulation studies, we evaluate various policies in cases where the assumptions of our theorems do not hold. Our results illustrate that comparing policies based on short-term evaluations can lead to misleading conclusions.", "keywords": [ "Long-run welfare", "policy design", "Rawlsian policy and utilitarianism" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 4, 4 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 2, 5, 2, 4 ] }, { "paper_id": "d8hURACo0P", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Derek Xiao", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Jonathan Kao", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "", "position": "Assistant Professor" }, { "name": "Ken-Fu Liang", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Keyu Ji", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 23, "n_ref": 73, "n_ref_all": 116, "n_fig": 25, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 18, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 539, "formula_len_all_1": 255, "len_all": 125177, "len_all_1": 66980, "len_abs": 2179, "len_title": 129, "len_sents": 44694, "len_sents_1": 33863, "n_sents": 311, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 2215, "title": "Using Reinforcement Learning to Investigate Neural Dynamics During Motor Learning", "abs": "Recent work characterized shifts in preparatory activity of the motor cortex during motor learning. \nThe specific geometry of the shifts during learning, washout, and relearning blocks was hypothesized to implement the acquisition, retention, and retrieval of motor memories. \nWe sought train recurrent neural network (RNN) models that could be used to study these motor learning phenomena.\nWe built an environment for a curl field (CF) motor learning task and trained RNNs with reinforcement learning (RL) with novel regularization terms to perform behaviorally realistic reaching trajectories over the course of learning. \nOur choice of RL rather than supervised learning was motivated by the idea that motor adaptation to a novel environment, in the absence of demonstrations, is a process of reoptimization. \nWe find these models, despite lack of supervision, reproduce many behavioral findings from human and monkey CF adaptation experiments. \nRelearning is faster than initial learning, indicating formation of motor memories. \nOptimal reaches under a CF are not straight, but rather curved, which is optimal and has been observed in humans and macaques. \nThese models also captured key neurophysiological findings. \nWe found that the model\u2019s preparatory activity existed in a force-predictive subspace that remained stable across learning, washout, and relearning. \nAdditionally, preparatory activity shifted uniformly, independently of the distance to the CF trained target. \nFinally, we found that the washout shift became more orthogonal to the learning shift, and hence more brain-like, when the RNNs are pretrained to have prior experience with CF dynamics. \nWe argue the increased fit to neurophysiological recordings is driven by more generalizable and structured dynamical motifs in the model with prior experience from pretraining. \nThis suggests that the near-orthogonality of learning-washout neural geometry underlying motor memory may be influenced by structured dynamical motifs in the motor cortex circuitry developed from prior experience. \nTogether, our work takes a step towards elucidating the factors that support motor memory, acquisition, retention, and retrieval during motor learning.", "keywords": [ "Computational Neuroscience", "Motor Learning", "RNN modeling", "Reinforcement LearningRecent work characterized shifts in preparatory activity of the motor cortex during motor learning. The specific geometry of the shifts during learning", "washout", "and relearning blocks was hypothesized to implement the acquisition", "retention", "and retrieval of motor memories. %This leads to the question: what learning algorithms lead to the emergence of these phenomena when monkeys perform a curl field (CF) adaptation task? We sought to develop a framework to train recurrent neural network (RNN) models that could be used to study these motor learning phenomena. We built an environment for a curl field (CF) motor learning task and trained RNNs with reinforcement learning (RL) with novel regularization terms to perform behaviorally realistic reaching trajectories over the course of learning. Our choice of RL rather than supervised learning was motivated by the idea that motor adaptation to a novel environment is a process of reoptimization. We find these models", "despite lack of supervision", "reproduce many behavioral findings from human and monkey CF adaptation experiments. Relearning is faster than initial learning", "indicating formation of motor memories. Optimal reaches under a CF are not straight", "but rather curved", "which is optimal and has been observed in humans and macaques. These models also captured key neurophysiological findings. We found that the model\u2019s preparatory activity shifted uniformly", "independently of the distance to the CF trained target. %We also found the washout shift was consistently approximately orthogonal to the learning shift. Finally", "we found that the washout shift becomes more orthogonal to the learning shift when the RNNs are pretrained to learn CF dynamics. We argue the increased fit to neurophysiological recordings is driven by more generalizable circuitry in the pretrained model. This suggests that some aspects of the neural geometry underlying motor memory may be influenced by priors learned over experience in the motor cortex circuitry. Together", "this work provides a modeling framework for exploring algorithms that support motor memory", "acquisition", "retention", "and retrieval during motor learning. %These results may inform additional future theoretical exploration of the algorithms underlying motor memory acquisition", "retention", "and retrieval." ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "d8cnezVcaW", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "David Yao", "gender": "unknown", "institution": "Columbia University", "country": "", "position": "Full Professor" }, { "name": "Hanyang Zhao", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Haoxian Chen", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Henry Lam", "gender": "unknown", "institution": "Columbia University", "country": "", "position": "Associate Professor" }, { "name": "Wenpin Tang", "gender": "Male", "institution": "Columbia University", "country": "", "position": "Assistant Professor" } ], "n_formula": 63, "n_formula_1": 36, "n_ref_uni": 49, "n_ref": 88, "n_ref_all": 130, "n_fig": 16, "n_tab": 7, "L_tab": 537, "n_element_tab": 88, "n_fig_1": 13, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5896, "formula_len_all_1": 3165, "len_all": 186467, "len_all_1": 67326, "len_abs": 1121, "len_title": 106, "len_sents": 51150, "len_sents_1": 28244, "n_sents": 484, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1131, "title": "MallowsPO: Fine-Tune Your LLM with Preference Dispersions", "abs": "Direct Preference Optimization (DPO) has recently emerged as a popular approach to improve reinforcement learning with human feedback (RLHF), leading to better techniques to fine-tune large language models (LLM). A weakness of DPO, however, lies in its lack of capability to characterize the diversity of human preferences. Inspired by Mallows' theory of preference ranking, we develop in this paper a new approach, the *MallowsPO*. A distinct feature of this approach is a *dispersion index*, which reflects the dispersion of human preference to prompts. We show that existing DPO models can be reduced to special cases of this dispersion index, thus unified with MallowsPO. More importantly, we demonstrate empirically how to use this dispersion index to enhance the performance of DPO in a broad array of benchmark tasks, from synthetic bandit selection to controllable generation and dialogues, while maintaining great generalization capabilities. MallowsPO is also compatible with other SOTA offline preference optimization methods, boosting nearly 2\\% extra LC win rate when used as a plugin for fine-tuning Llama3-Instruct.", "keywords": [ "Language Models fine-tuning", "learning from human feedback", "Mallows ranking model", "human preference dispersions" ], "rating_list": [ 5, 8, 5, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "d7q9IGj2p0", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Byeongho Heo", "gender": "Male", "institution": "NAVER AI Lab", "country": "KR", "position": "Researcher" }, { "name": "Dongyoon Han", "gender": "Male", "institution": "NAVER", "country": "KR", "position": "Senior Research Scientist" }, { "name": "Taekyung Kim", "gender": "Not Specified", "institution": "NAVER AI Lab", "country": "KR", "position": "Research Scientist" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 36, "n_ref": 104, "n_ref_all": 153, "n_fig": 12, "n_tab": 9, "L_tab": 4487, "n_element_tab": 462, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 702, "n_element_tab_1": 49, "formula_len_all": 967, "formula_len_all_1": 633, "len_all": 189100, "len_all_1": 67999, "len_abs": 1450, "len_title": 47, "len_sents": 50241, "len_sents_1": 31183, "n_sents": 401, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1457, "title": "MORPHING TOKENS DRAW STRONG MASKED IMAGE MODELS", "abs": "Masked image modeling (MIM) has emerged as a promising approach for training Vision Transformers (ViTs). The essence of MIM lies in the token-wise prediction of masked tokens, which aims to predict targets tokenized from images or generated by pre-trained models like vision-language models. While using tokenizers or pre-trained models are plausible MIM targets, they often offer spatially inconsistent targets even for neighboring tokens, complicating models to learn unified and discriminative representations. Our pilot study identifies spatial inconsistencies and suggests that resolving them can accelerate representation learning. Building upon this insight, we introduce a novel self-supervision signal called Dynamic Token Morphing (DTM), which dynamically aggregates contextually related tokens to yield contextualized targets, thereby mitigating spatial inconsistency. DTM is compatible with various SSL frameworks; we showcase improved MIM results by employing DTM, barely introducing extra training costs. Our method facilitates training by using consistent targets, resulting in 1) faster training and 2) reduced losses. Experiments on ImageNet-1K and ADE20K demonstrate the superiority of our method compared with state-of-the-art, complex MIM methods. Furthermore, the comparative evaluation of the iNaturalists and fine-grained visual classification datasets further validates the transferability of our method on various downstream tasks.", "keywords": [ "Self-supervised learning" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 2, 4 ], "presentation_list": [ 3, 2, 4 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "d7pr2doXn3", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Jinyang Du", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Lijun Yang", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Wen Yao", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "Principal Researcher" }, { "name": "Yue Deng", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Zeyu Li", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "hongkun dou", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 21, "n_ref_uni": 46, "n_ref": 137, "n_ref_all": 170, "n_fig": 17, "n_tab": 11, "L_tab": 3911, "n_element_tab": 554, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 900, "n_element_tab_1": 157, "formula_len_all": 1728, "formula_len_all_1": 1381, "len_all": 180934, "len_all_1": 64131, "len_abs": 1351, "len_title": 118, "len_sents": 49441, "len_sents_1": 27629, "n_sents": 394, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1356, "title": "Hybrid Regularization Improves Diffusion-based Inverse Problem Solving", "abs": "Diffusion models, recognized for their effectiveness as generative priors, have become essential tools for addressing a wide range of visual challenges. Recently, there has been a surge of interest in leveraging Denoising processes for Regularization (DR) to solve inverse problems. However, existing methods often face issues such as mode collapse, which results in excessive smoothing and diminished diversity. In this study, we perform a comprehensive analysis to pinpoint the root causes of gradient inaccuracies inherent in DR. Drawing on insights from diffusion model distillation, we propose a novel approach called Consistency Regularization (CR), which provides stabilized gradients without the need for ODE simulations. Building on this, we introduce Hybrid Regularization (HR), a unified framework that combines the strengths of both DR and CR, harnessing their synergistic potential. Our approach proves to be effective across a broad spectrum of inverse problems, encompassing both linear and nonlinear scenarios, as well as various measurement noise statistics. Experimental evaluations on benchmark datasets, including FFHQ and ImageNet, demonstrate that our proposed framework not only achieves highly competitive results compared to state-of-the-art methods but also offers significant reductions in wall-clock time and memory consumption.", "keywords": [ "Diffusion models", "inverse problems" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 4, 3, 1 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 3, 3, 1 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "d7DZRNe2xG", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Emiel Hoogeboom", "gender": "unknown", "institution": "Google", "country": "NL", "position": "Researcher" }, { "name": "Jonathan Heek", "gender": "unknown", "institution": "Google", "country": "NL", "position": "Software Engineer" }, { "name": "Tim Salimans", "gender": "Male", "institution": "Google", "country": "NL", "position": "Research Scientist" } ], "n_formula": 19, "n_formula_1": 17, "n_ref_uni": 16, "n_ref": 46, "n_ref_all": 57, "n_fig": 6, "n_tab": 6, "L_tab": 2510, "n_element_tab": 378, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1448, "n_element_tab_1": 258, "formula_len_all": 1436, "formula_len_all_1": 1586, "len_all": 117694, "len_all_1": 63041, "len_abs": 1221, "len_title": 47, "len_sents": 35857, "len_sents_1": 25441, "n_sents": 304, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 28, "L_abs": 1185, "title": "Multistep Consistency Models", "abs": "Diffusion models are relatively easy to train but require many steps to generate samples. Consistency models are far more difficult to train, but generate samples in a single step.\n\nIn this paper we propose Multistep Consistency Models: A unification between Consistency Models (Song et al., 2023) and TRACT (Berthelotet al., 2023) that can interpolate between a consistency model and a diffusion model: a trade-off between sampling speed and sampling quality. Specifically, a 1-step consistency model is a conventional consistency model whereas a $\\infty$-step consistency model is a diffusion model.\n\nMultistep Consistency Models work really well in practice. By increasing the sample budget from a single step to 2-8 steps, we can train models more easily that generate higher quality samples, while retaining much of the sampling speed benefits. Notable results are 1.4 FID on Imagenet 64 in 8 sampling steps and 2.1 FID on Imagenet128 in 8 sampling steps with consistency distillation, using simple losses without adversarial training. We also show that our method scales to a text-to-image diffusion model, generating samples that are close to the quality of the original model.", "keywords": [ "multistep consistency", "consistency models", "distillation", "fast sampling" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "d6Kk7moQH3", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Daniel Rueckert", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Huaqi Qiu", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Postdoc" }, { "name": "Julian McGinnis", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Nil Stolt-Ans\u00f3", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Vasiliki Sideri-Lampretsa", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Martin Menten", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Lecturer" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 37, "n_ref": 65, "n_ref_all": 85, "n_fig": 3, "n_tab": 3, "L_tab": 3302, "n_element_tab": 41, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 3251, "n_element_tab_1": 16, "formula_len_all": 772, "formula_len_all_1": 674, "len_all": 126793, "len_all_1": 75699, "len_abs": 1802, "len_title": 100, "len_sents": 36131, "len_sents_1": 34514, "n_sents": 237, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1812, "title": "Image registration is a geometric deep learning task", "abs": "Data-driven deformable image registration methods predominantly rely on operations that process grid-like inputs.\nHowever, applying deformable transformations to an image results in a warped space that deviates from a rigid grid structure.\nConsequently, data-driven approaches with sequential deformations have to apply grid resampling operations between each deformation step. \nWhile artifacts caused by resampling are negligible in high-resolution images, the resampling of sparse, high-dimensional feature grids introduces errors that affect the deformation modeling process.\nTaking inspiration from Lagrangian reference frames of deformation fields, our work introduces a novel paradigm for data-driven deformable image registration that utilizes geometric deep-learning principles to model deformations without grid requirements.\nSpecifically, we model image features as a set of nodes that freely move in Euclidean space, update their coordinates under graph operations, and dynamically readjust their local neighborhoods.\nWe employ this formulation to construct a multi-resolution deformable registration model, where deformation layers iteratively refine the overall transformation at each resolution without intermediate resampling operations on the feature grids.\nWe investigate our method's ability to fully deformably capture large deformations across a number of medical imaging registration tasks. \nIn particular, we apply our approach (GeoReg) to the registration of inter-subject brain MR images and inhale-exhale lung CT images, showing on par performance with the current state-of-the-art methods. \nWe believe our contribution open up avenues of research to reduce the black-box nature of current learned registration paradigms by explicitly modeling the transformation within the architecture.", "keywords": [ "Image registration", "Geometric deep learning" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "d63a4AM4hb", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Eric J Michaud", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Isaac Liao", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Joshua Engels", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Max Tegmark", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Wes Gurnee", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 10, "n_ref_uni": 46, "n_ref": 94, "n_ref_all": 156, "n_fig": 25, "n_tab": 4, "L_tab": 1449, "n_element_tab": 290, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 2609, "n_element_tab_1": 42, "formula_len_all": 2052, "formula_len_all_1": 612, "len_all": 210155, "len_all_1": 67768, "len_abs": 1315, "len_title": 108, "len_sents": 70158, "len_sents_1": 30023, "n_sents": 519, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 1333, "title": "Not All Language Model Features Are Linear", "abs": "Recent work has proposed that language models perform computation by manipulating one-dimensional representations of concepts (\"features\") in activation space. In contrast, we explore whether some language model representations may be inherently multi-dimensional. We begin by developing a rigorous definition of irreducible multi-dimensional features based on whether they can be decomposed into either independent or non-co-occurring lower-dimensional features. Motivated by these definitions, we design a scalable method that uses sparse autoencoders to automatically find multi-dimensional features in GPT-2 and Mistral 7B. These auto-discovered features include strikingly interpretable examples, e.g. $\\textit{circular}$ features representing days of the week and months of the year. We identify tasks where these exact circles are used to solve computational problems involving modular arithmetic in days of the week and months of the year. Next, we provide evidence that these circular features are indeed the fundamental unit of computation in these tasks with intervention experiments on Mistral 7B and Llama 3 8B. Finally, we find further circular representations by breaking down the hidden states for these tasks into interpretable components, and we examine the continuity of the days of the week feature in Mistral 7B.", "keywords": [ "Mechanistic Interpretability", "Learned Representations", "Dictionary Learning", "Sparse Autoencoders", "Features", "Circuits" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "d5HUnyByAI", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Angel X Chang", "gender": "Female", "institution": "Simon Fraser University", "country": "CA", "position": "Associate Professor" }, { "name": "Austin Wang", "gender": "Male", "institution": "Simon Fraser University", "country": "", "position": "PhD student" }, { "name": "Joakim Bruslund Haurum", "gender": "Male", "institution": "Aalborg University", "country": "DK", "position": "Postdoc" }, { "name": "Xiaoliang Huo", "gender": "Male", "institution": "Simon Fraser University", "country": "", "position": "PhD student" }, { "name": "ZeMing Gong", "gender": "Male", "institution": "Simon Fraser University", "country": "", "position": "MS student" }, { "name": "Graham W Taylor", "gender": "unknown", "institution": "University of Guelph", "country": "", "position": "Assistant Professor" }, { "name": "Scott C Lowe", "gender": "unknown", "institution": "Vector Institute", "country": "CA", "position": "Postdoc" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 66, "n_ref": 128, "n_ref_all": 186, "n_fig": 13, "n_tab": 17, "L_tab": 10108, "n_element_tab": 2109, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 574, "n_element_tab_1": 131, "formula_len_all": 316, "formula_len_all_1": 340, "len_all": 295230, "len_all_1": 71167, "len_abs": 777, "len_title": 120, "len_sents": 74050, "len_sents_1": 35010, "n_sents": 542, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 771, "title": "CLIBD: Bridging Vision and Genomics for Biodiversity Monitoring at Scale", "abs": "Measuring biodiversity is crucial for understanding ecosystem health. While prior works have developed machine learning models for taxonomic classification of photographic images and DNA separately, in this work, we introduce a multi-modal approach combining both, using CLIP-style contrastive learning to align images, barcode DNA, and text-based representations of taxonomic labels in a unified embedding space. This allows for accurate classification of both known and unknown insect species without task-specific fine-tuning, leveraging contrastive learning for the first time to fuse DNA and image data. Our method surpasses previous single-modality approaches in accuracy by over 8% on zero-shot learning tasks, showcasing its effectiveness in biodiversity studies.", "keywords": [ "Multimodal Learning", "Contrastive Learning", "DNA Barcodes", "Taxonomic Classification", "Fine-grained Classification", "Biodiversity Monitoring" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 1, 3, 2, 4 ], "presentation_list": [ 3, 4, 2, 4 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "d54fIsAbff", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Moritz Wagner", "gender": "Male", "institution": "Technische Universit\u00e4t Berlin", "country": "DE", "position": "PhD student" }, { "name": "Sebastian Pokutta", "gender": "Male", "institution": "ZIB", "country": "DE", "position": "Vice President" }, { "name": "Shpresim Sadiku", "gender": "Male", "institution": "Technische Universit\u00e4t Berlin", "country": "DE", "position": "PhD student" } ], "n_formula": 38, "n_formula_1": 21, "n_ref_uni": 38, "n_ref": 75, "n_ref_all": 101, "n_fig": 5, "n_tab": 9, "L_tab": 4096, "n_element_tab": 306, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1926, "n_element_tab_1": 135, "formula_len_all": 2308, "formula_len_all_1": 1242, "len_all": 134860, "len_all_1": 62573, "len_abs": 1346, "len_title": 106, "len_sents": 34098, "len_sents_1": 25856, "n_sents": 284, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1373, "title": "GSE: Group-wise Sparse and Explainable Adversarial Attacks", "abs": "Sparse adversarial attacks fool deep neural networks (DNNs) through minimal pixel perturbations, often regularized by the $\\ell_0$ norm. Recent efforts have replaced this norm with a structural sparsity regularizer, such as the nuclear group norm, to craft group-wise sparse adversarial attacks. The resulting perturbations are thus explainable and hold significant practical relevance, shedding light on an even greater vulnerability of DNNs. However, crafting such attacks poses an optimization challenge, as it involves computing norms for groups of pixels within a non-convex objective. We address this by presenting a two-phase algorithm that generates group-wise sparse attacks within semantically meaningful areas of an image. Initially, we optimize a quasinorm adversarial loss using the $1/2$-quasinorm proximal operator tailored for non-convex programming. Subsequently, the algorithm transitions to a projected Nesterov's accelerated gradient descent with $2$-norm regularization applied to perturbation magnitudes. Rigorous evaluations on CIFAR-10 and ImageNet datasets demonstrate a remarkable increase in group-wise sparsity, e.g., $50.9\\%$ on CIFAR-10 and $38.4\\%$ on ImageNet (average case, targeted attack). This performance improvement is accompanied by significantly faster computation times, improved interpretability, and a $100\\%$ attack success rate.", "keywords": [ "Sparse Adversarial Attack", "Quasinorm Regularization", "Nesterov Accelerated Gradient" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "d4qMoUSMLT", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Doosan Jung", "gender": "unknown", "institution": "Cisco", "country": "US", "position": "Researcher" }, { "name": "Emily Pitler", "gender": "Female", "institution": "Cisco", "country": "", "position": "Principal Researcher" }, { "name": "Jianxin Zhang", "gender": "unknown", "institution": "University of Michigan", "country": "US", "position": "PhD student" }, { "name": "Josh Viktorov", "gender": "Male", "institution": "Cisco", "country": "US", "position": "Researcher" } ], "n_formula": 67, "n_formula_1": 14, "n_ref_uni": 26, "n_ref": 421, "n_ref_all": 464, "n_fig": 7, "n_tab": 10, "L_tab": 1277, "n_element_tab": 171, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 788, "n_element_tab_1": 61, "formula_len_all": 7576, "formula_len_all_1": 1327, "len_all": 175835, "len_all_1": 70538, "len_abs": 1447, "len_title": 109, "len_sents": 51235, "len_sents_1": 27715, "n_sents": 498, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 107, "L_abs": 1457, "title": "Efficient Training of Neural Stochastic Differential Equations by Matching Finite Dimensional Distributions", "abs": "Neural Stochastic Differential Equations (Neural SDEs) have emerged as powerful mesh-free generative models for continuous stochastic processes, with critical applications in fields such as finance, physics, and biology. Previous state-of-the-art methods have relied on adversarial training, such as GANs, or on minimizing distance measures between processes using signature kernels. However, GANs suffer from issues like instability, mode collapse, and the need for specialized training techniques, while signature kernel-based methods require solving linear PDEs and backpropagating gradients through the solver, whose computational complexity scales quadratically with the discretization steps. In this paper, we identify a novel class of strictly proper scoring rules for comparing continuous Markov processes. This theoretical finding naturally leads to a novel approach called Finite Dimensional Matching (FDM) for training Neural SDEs. Our method leverages the Markov property of SDEs to provide a computationally efficient training objective. This scoring rule allows us to bypass the computational overhead associated with signature kernels and reduces the training complexity from $O(D^2)$ to $O(D)$ per epoch, where $D$ represents the number of discretization steps of the process. We demonstrate that FDM achieves superior performance, consistently outperforming existing methods in terms of both computational efficiency and generative quality.", "keywords": [ "neural stochastic differential equations", "Markov process", "scoring rule" ], "rating_list": [ 5, 8, 8, 3 ], "soundness_list": [ 1, 4, 4, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "d4njmzM7jf", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Dengsheng Chen", "gender": "Female", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Enhua Wu", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Emeritus" }, { "name": "Jie Hu", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Principal Researcher" }, { "name": "Wei Xiaoming", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 9, "n_ref_uni": 91, "n_ref": 229, "n_ref_all": 261, "n_fig": 18, "n_tab": 6, "L_tab": 6667, "n_element_tab": 975, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 648, "n_element_tab_1": 106, "formula_len_all": 743, "formula_len_all_1": 309, "len_all": 305873, "len_all_1": 66534, "len_abs": 1306, "len_title": 104, "len_sents": 72144, "len_sents_1": 32091, "n_sents": 560, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1296, "title": "Denoising with a Joint-Embedding Predictive Architecture", "abs": "Joint-embedding predictive architectures (JEPAs) have shown substantial promise in self-supervised representation learning, yet their application in generative modeling remains underexplored. Conversely, diffusion models have demonstrated significant efficacy in modeling arbitrary probability distributions. In this paper, we introduce Denoising with a Joint-Embedding Predictive Architecture (D-JEPA), pioneering the integration of JEPA within generative modeling. By recognizing JEPA as a form of masked image modeling, we reinterpret it as a generalized next-token prediction strategy, facilitating data generation in an auto-regressive manner. Furthermore, we incorporate diffusion loss to model the per-token probability distribution, enabling data generation in a continuous space. We also adapt flow matching loss as an alternative to diffusion loss, thereby enhancing the flexibility of D-JEPA. Empirically, with increased GFLOPs, D-JEPA consistently achieves lower FID scores with fewer training epochs, indicating its good scalability. Our base, large, and huge models outperform all previous generative models across all scales on ImageNet conditional generation benchmarks. Beyond image generation, D-JEPA is well-suited for other continuous data modeling, including video and audio.", "keywords": [ "AIGC", "JEPA", "Diffusion Model", "Flow Matching", "Image Synthetic" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "d4gu2XgccF", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Aurelie Lozano", "gender": "Female", "institution": "IBM Research", "country": "US", "position": "Principal Researcher" }, { "name": "Ching-Yun Ko", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Georgios Kollias", "gender": "unknown", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Payel Das", "gender": "Female", "institution": "IBM, International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Sihui Dai", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Subhajit Chaudhury", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Research Scientist" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 39, "n_ref": 61, "n_ref_all": 94, "n_fig": 4, "n_tab": 12, "L_tab": 2849, "n_element_tab": 502, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 59, "n_element_tab_1": 4, "formula_len_all": 579, "formula_len_all_1": 715, "len_all": 156761, "len_all_1": 65843, "len_abs": 948, "len_title": 121, "len_sents": 52772, "len_sents_1": 33660, "n_sents": 348, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 839, "title": "MEMREASONER: A MEMORY-AUGMENTED LANGUAGE MODEL ARCHITECTURE FOR MULTI-HOP REASONING", "abs": "Recent benchmarks suggest that there remains significant room to improve large language models\u2019 ability to robustly reason across facts distributed in extremely long documents. In this work, we propose MemReasoner, a new memory-augmented LLM architecture that is trained to perform temporal reasoning, along with multiple computational steps, over the context stored in the memory. Experiments show that MemReasoner trained on the core reasoning facts generalizes better, when compared to off-the-shelf large language models and existing recurrent models, on a test distribution where the required facts are scattered across long natural text up to 128k tokens. Further, MemReasoner demonstrates robust reasoning performance relative to the baselines, when the answer distribution in test samples differs from that in the training set.", "keywords": [ "reasoning", "multi hop", "memory", "large language models", "generalization" ], "rating_list": [ 3, 6, 3 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 1, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 5, 3, 3 ] }, { "paper_id": "d465apqCqc", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Yi Chang", "gender": "Male", "institution": "Jilin University, China", "country": "CN", "position": "Full Professor" }, { "name": "Yuan Wu", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yupeng Chang", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 18, "n_ref_uni": 80, "n_ref": 170, "n_ref_all": 194, "n_fig": 8, "n_tab": 10, "L_tab": 3864, "n_element_tab": 460, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2097, "n_element_tab_1": 275, "formula_len_all": 807, "formula_len_all_1": 866, "len_all": 227999, "len_all_1": 59117, "len_abs": 1373, "len_title": 156, "len_sents": 48782, "len_sents_1": 25288, "n_sents": 334, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1381, "title": "BA-LoRA: Bias-Alleviating Low-Rank Adaptation to Mitigate Catastrophic Inheritance in Large Language Models", "abs": "Large language models (LLMs) have demonstrated remarkable proficiency across various natural language processing (NLP) tasks. However, adapting LLMs to downstream applications requires computationally intensive and memory-demanding fine-tuning procedures. To alleviate these burdens, parameter-efficient fine-tuning (PEFT) techniques have emerged as a promising approach to tailor LLMs with minimal computational overhead. While PEFT methods offer substantial advantages, they do not fully address the pervasive issue of bias propagation from pre-training data. This work introduces Bias-Alleviating Low-Rank Adaptation (BA-LoRA), a novel PEFT method designed to counteract bias inheritance. BA-LoRA incorporates three distinct regularization terms: (1) a consistency regularizer, (2) a diversity regularizer, and (3) a singular value decomposition regularizer. These regularizers aim to enhance the models' consistency, diversity, and generalization capabilities during fine-tuning. We conduct extensive experiments on natural language understanding (NLU) and natural language generation (NLG) tasks using prominent LLMs such as LLaMA, Mistral, and Gemma. The results demonstrate that BA-LoRA outperforms LoRA and its state-of-the-art variants. Moreover, our method effectively mitigates the adverse effects of pre-training bias, leading to more reliable and robust model outputs.", "keywords": [ "supervised fine-tuning", "parameter efficient fine-tuning", "bias reduction" ], "rating_list": [ 8, 3, 3, 8, 6 ], "soundness_list": [ 4, 1, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 3, 4 ] }, { "paper_id": "d38yjwdGYr", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Chenxiong Qian", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Kai Ye", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Liangcai Su", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Qingyu Zhang", "gender": "Male", "institution": "the University of Hong Kong, University of Hong Kong", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 21, "n_ref": 37, "n_ref_all": 64, "n_fig": 9, "n_tab": 5, "L_tab": 3246, "n_element_tab": 215, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3128, "n_element_tab_1": 138, "formula_len_all": 411, "formula_len_all_1": 345, "len_all": 128562, "len_all_1": 58315, "len_abs": 4076, "len_title": 98, "len_sents": 34489, "len_sents_1": 25157, "n_sents": 283, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 50, "L_abs": 1485, "title": "ConGra: Benchmarking Automatic Conflict Resolution", "abs": "Resolving conflicts from merging different software versions is a challenging task for developers. To reduce the overhead of manual merging, researchers develop various program analysis-based tools which only solve specific types of conflicts and have a limited scope of application. With the development of language models, researchers treat conflict code as text, which theoretically allows for addressing almost all types of conflicts. However, the absence of effective conflict difficulty grading methods hinders a comprehensive evaluation of large language models (LLMs), making it difficult to gain a deeper understanding of their limitations. Furthermore, there is a notable lack of large-scale open benchmarks for evaluating the performance of LLMs in automatic conflict resolution. To address these issues, We introduce ConGra, a CONflict-GRAded benchmarking scheme designed to evaluate the performance of software merging tools under varying complexity conflict scenarios. We propose a novel approach to classify conflicts based on code operations and use it to build a large-scale evaluation dataset based on 44,948 conflicts from 34 real-world projects. We evaluate state-of-the-art LLMs on conflict resolution tasks using this dataset. By employing the proposed dataset, we assess the performance of multiple state-of-the-art LLMs and code LLMs, ultimately uncovering two counterintuitive yet insightful phenomena. ConGra will be released at https://github.com/xxx/ConGra.", "keywords": [ "Code merging", "conflict resolution", "large language model" ], "rating_list": [ 5, 3, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 3, 5, 4, 4, 4 ] }, { "paper_id": "d32d9fE5lG", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Huihui Xu", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Lei Zhu", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou) & HKUST", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 21, "n_ref": 58, "n_ref_all": 70, "n_fig": 0, "n_tab": 6, "L_tab": 4103, "n_element_tab": 320, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 1521, "n_element_tab_1": 176, "formula_len_all": 1059, "formula_len_all_1": 834, "len_all": 137207, "len_all_1": 61487, "len_abs": 1553, "len_title": 128, "len_sents": 30842, "len_sents_1": 25826, "n_sents": 275, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1569, "title": "Online Agglomerative Pooling for Scalable Self-Supervised Universal Segmentation", "abs": "Recent self-supervised image segmentors have achieved promising zero-shot performance. However, their pretraining schedule is multi-stage and alternates between offline pseudo-masks generation and parameters update, which leads to unstable training and sub-optimal solution. \nTo solve this issue, we present Online Agglomerative Pooling (OAP) that allows efficiently generating universal pseudo-masks and updating parameters simultaneously at each training step. \nSpecifically, OAP contains a stack of instance pooling and semantic pooling layers. By using a layer-varied threshold, OAP can generate multi-hierarchy masks that can provide more visual details for segmentation. Compared with MaskCut or Divide-Conquer, each OAP layer can identify connected nodes in parallel, thus can generate universal pseudo-masks for a single image within tens of milliseconds. \nMoreover, to deploy OAP in online pretraining, we devise a teacher-student framework with Query-wise Self-distillation, where the local view queries are each aligned with the matched global view queries to learn the local-to-global correspondence. \nCompared with other multi-stage offline pretraining methods, our framework can effectively scale to larger datasets while ensuring quicker convergence. Extensive experiments on the COCO, PASCAL VOC, Cityscapes, and UVO datasets show that our method achieves state-of-the-art performance on zero-shot instance segmentation, semantic segmentation, and panoptic segmentation. Our code and\npretrained models shall be released upon acceptance of this work.", "keywords": [ "self-supervised learning", "universal image segmentation", "zero-shot segmentation", "graph pooling" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 1, 1, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 1, 4 ] }, { "paper_id": "d2UrCGtntF", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Andrea Tagliasacchi", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "Associate Professor" }, { "name": "Daniel Watson", "gender": "unknown", "institution": "Google DeepMind", "country": "CA", "position": "Researcher" }, { "name": "David J. Fleet", "gender": "Male", "institution": "Google DeepMind", "country": "CA", "position": "Research Scientist" }, { "name": "Lala Li", "gender": "unknown", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Saurabh Saxena", "gender": "Male", "institution": "Google", "country": "CA", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 34, "n_ref": 70, "n_ref_all": 85, "n_fig": 5, "n_tab": 5, "L_tab": 2115, "n_element_tab": 240, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2467, "n_element_tab_1": 250, "formula_len_all": 222, "formula_len_all_1": 225, "len_all": 136702, "len_all_1": 62456, "len_abs": 1172, "len_title": 92, "len_sents": 34600, "len_sents_1": 29526, "n_sents": 224, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 44, "L_abs": 1178, "title": "Diffusion Models for 4D Novel View Synthesis", "abs": "We present 4DiM, a cascaded diffusion model for 4D novel view synthesis (NVS), supporting generation with arbitrary camera trajectories and timestamps, in natural scenes, conditioned on one or more images. With a novel architecture and sampling procedure, we enable training on a mixture of 3D (with camera pose), 4D (pose+time) and video (time but no pose) data, which greatly improves generalization to unseen images and camera pose trajectories over prior works which generally operate in limited domains (e.g., object centric).\n4DiM is the first-ever NVS method with intuitive metric-scale camera pose control enabled by our novel calibration pipeline for structure-from-motion-posed data. Experiments demonstrate that 4DiM outperforms prior 3D NVS models both in terms of \nimage fidelity and pose alignment, while also enabling the generation of scene dynamics. 4DiM provides a general framework for a variety of tasks including single-image-to-3D, two-image-to-video (interpolation and extrapolation), and pose-conditioned video-to-video translation, which we illustrate qualitatively on a variety of scenes.\nSee https://anonymous-4d-diffusion.github.io for video samples.", "keywords": [ "generative", "models", "diffusion", "novel", "view", "synthesis", "3d", "4d", "scenes" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "d2H1oTNITn", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Chengqi Lyu", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Kai Chen", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Wenwei Zhang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yuzhe Gu", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 39, "n_ref": 86, "n_ref_all": 108, "n_fig": 4, "n_tab": 10, "L_tab": 2543, "n_element_tab": 359, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1751, "n_element_tab_1": 266, "formula_len_all": 877, "formula_len_all_1": 877, "len_all": 165939, "len_all_1": 64414, "len_abs": 1782, "len_title": 113, "len_sents": 49593, "len_sents_1": 29384, "n_sents": 376, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1790, "title": "Mask-DPO: Generalizable Fine-grained Factuality Alignment of LLMs", "abs": "Large language models (LLMs) exhibit hallucinations (i.e., unfaithful or nonsensical information) when serving as AI assistants in various domains. Since hallucinations always come with truthful content in the LLM responses, previous factuality alignment methods that conduct response-level preference learning inevitably introduced noises during training. Therefore, this paper proposes a fine-grained factuality alignment method based on Direct Preference Optimization (DPO), called Mask-DPO. Incorporating sentence-level factuality as mask signals, Mask-DPO only learns from factually correct sentences in the preferred samples and prevents the penalty on factual contents in the not preferred samples, which resolves the ambiguity in the preference learning. Extensive experimental results demonstrate that Mask-DPO can significantly improve the factuality of LLMs responses to questions from both in-domain and out-of-domain datasets, although these questions and their corresponding topics are unseen during training. Only trained on the ANAH train set, the score of Llama3.1-8B-Instruct on the ANAH test set is improved from 49.19% to 77.53%, even surpassing the score of Llama3.1-70B-Instruct (53.44%), while its FactScore on the out-of-domain Biography dataset is also improved from 30.29% to 39.39%. We further study the generalization property of Mask-DPO using different training sample scaling strategies and find that scaling the number of topics in the dataset is more effective than the number of questions. We provide a hypothesis of what factual alignment is doing with LLMs, on the implication of this phenomenon, and conduct proof-of-concept experiments to verify it. We hope the method and the findings pave the way for future research on scaling factuality alignment.", "keywords": [ "Hallucination Mitigation", "Large Language Model", "Fine-grained Alignment" ], "rating_list": [ 6, 8, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "d23EVDRJ6g", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Hai Jiang", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "Full Professor" }, { "name": "Juwei Lu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CA", "position": "Sr Principal Scientist" }, { "name": "Li cheng", "gender": "Not Specified", "institution": "University of Alberta", "country": "CA", "position": "Full Professor" }, { "name": "Muhammad Gohar Javed", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "MS student" }, { "name": "Xinxin Zuo", "gender": "Female", "institution": "Concordia University", "country": "CA", "position": "Assistant Professor" }, { "name": "Yilin Wang", "gender": "Female", "institution": "University of Alberta", "country": "CA", "position": "PhD student" }, { "name": "Yuxuan Mu", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "PhD student" }, { "name": "chuan guo", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 34, "n_ref": 100, "n_ref_all": 125, "n_fig": 9, "n_tab": 8, "L_tab": 1861, "n_element_tab": 245, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 431, "n_element_tab_1": 79, "formula_len_all": 602, "formula_len_all_1": 472, "len_all": 138830, "len_all_1": 61618, "len_abs": 1546, "len_title": 138, "len_sents": 43475, "len_sents_1": 29665, "n_sents": 294, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1559, "title": "MotionDreamer: One-to-Many Motion Synthesis with Localized Generative Masked Transformer", "abs": "Generative masked transformer have demonstrated remarkable success across various content generation tasks, primarily due to their ability to effectively model large-scale dataset distributions with high consistency. However, in the animation domain, large datasets are not always available. Applying generative masked modeling to generate diverse instances from a single MoCap reference may lead to overfitting, a challenge that remains unexplored. In this work, we present MotionDreamer, a localized masked modeling paradigm designed to learn motion internal patterns from a given motion with arbitrary topology and duration. By embedding the given motion into quantized tokens with a novel distribution regularization method, MotionDreamer constructs a robust and informative codebook for local motion patterns. Moreover, a sliding window local attention is introduced in our masked transformer, enabling the generation of natural yet diverse animations that closely resemble the reference motion patterns. As demonstrated through comprehensive experiments, MotionDreamer outperforms the state-of-the-art methods that are typically GAN or Diffusion-based in both faithfulness and diversity. Thanks to the consistency and robustness of quantization-based approach, MotionDreamer can also effectively perform downstream tasks such as temporal motion editing, crowd motion synthesis, and beat-aligned dance generation, all using a single reference motion. Our implementation, learned models and results are to be made publicly available upon paper acceptance.", "keywords": [ "motion synthesis", "generative masked modelling", "vector quantization", "single instance learning" ], "rating_list": [ 5, 3, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 4 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "d1NWq4PjJW", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Nicu Sebe", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Full Professor" }, { "name": "Xiaojun Wu", "gender": "Male", "institution": "Jiangnan University", "country": "CN", "position": "Full Professor" }, { "name": "Yue Song", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Ziheng Chen", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "PhD student" } ], "n_formula": 83, "n_formula_1": 27, "n_ref_uni": 43, "n_ref": 191, "n_ref_all": 254, "n_fig": 4, "n_tab": 13, "L_tab": 3044, "n_element_tab": 242, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2204, "n_element_tab_1": 130, "formula_len_all": 6276, "formula_len_all_1": 2126, "len_all": 262251, "len_all_1": 72128, "len_abs": 1916, "len_title": 77, "len_sents": 76296, "len_sents_1": 27633, "n_sents": 847, "n_sents_1": 281, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 29, "L_abs": 1087, "title": "Gyrogroup Batch Normalization", "abs": "Several Riemannian manifolds in machine learning, such as Symmetric Positive Definite (SPD), Grassmann, spherical, and hyperbolic manifolds, have been proven to admit gyro structures, thus enabling a principled and effective extension of Euclidean Deep Neural Networks (DNNs) to manifolds. Inspired by this, this study introduces a general Riemannian Batch Normalization (RBN) framework on gyrogroups, termed GyroBN. We identify the least requirements to guarantee GyroBN with theoretical control over sample statistics, referred to as \\textit{pseudo-reduction} and \\textit{gyroisometric gyrations}, which are satisfied by all the existing gyrogroups in machine learning. Besides, our GyroBN incorporates several existing normalization methods, including the one on general Lie groups and different types of RBN on the non-group SPD geometry. Lastly, we instantiate our GyroBN on the Grassmannian and hyperbolic spaces. Experiments on the Grassmannian and hyperbolic networks demonstrate the effectiveness of our GyroBN. The code is available at https://github.com/GitZH-Chen/GyroBN.git.", "keywords": [ "Gyrovector Spaces", "Riemannian Manifolds", "Riemannian Batch Normalization", "Grassmannian Manifolds", "Hyperbolic Manifolds" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 2, 4, 2, 3 ] }, { "paper_id": "d16mJDyQN6", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Matthias Bethge", "gender": "Male", "institution": "University of Tuebingen", "country": "DE", "position": "Full Professor" }, { "name": "Pekka Marttinen", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Associate Professor" }, { "name": "S. T. John", "gender": "unknown", "institution": "Aalto University", "country": "FI", "position": "Research Fellow" }, { "name": "\u00c7a\u011flar H\u0131zl\u0131", "gender": "unknown", "institution": "Aalto University", "country": "", "position": "PhD student" }, { "name": "Cagatay Yildiz", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "Postdoc" } ], "n_formula": 69, "n_formula_1": 16, "n_ref_uni": 31, "n_ref": 153, "n_ref_all": 189, "n_fig": 11, "n_tab": 4, "L_tab": 716, "n_element_tab": 103, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 278, "n_element_tab_1": 33, "formula_len_all": 5283, "formula_len_all_1": 1418, "len_all": 209745, "len_all_1": 69260, "len_abs": 1366, "len_title": 116, "len_sents": 70429, "len_sents_1": 30010, "n_sents": 564, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1372, "title": "Identifying latent state transitions in non-linear dynamical systems", "abs": "This work aims to recover the underlying states and their time evolution in a latent dynamical system from high-dimensional sensory measurements. Previous work on identifiable representation learning in dynamical systems focused on identifying latent states, possibly with linear transition approximations. As such, they cannot identify nonlinear transition dynamics, and hence fail to reliably predict complex future behavior. Inspired by the advances in nonlinear \\ica, we propose a state-space modeling framework in which we can identify not just the latent states but also the unknown transition function that maps the past states to the present. Our identifiability theory relies on two key assumptions: (i) sufficient variability in the latent noise, and (ii) the bijectivity of the augmented transition function. Drawing from this theory, we introduce a practical algorithm based on variational auto-encoders. We empirically demonstrate that it improves generalization and interpretability of target dynamical systems by (i) recovering latent state dynamics with high accuracy, (ii) correspondingly achieving high future prediction accuracy, and (iii) adapting fast to new environments. Additionally, for complex real-world dynamics, (iv) it produces state-of-the-art future prediction results for long horizons, highlighting its usefulness for practical scenarios.", "keywords": [ "nonlinear ica", "identifiability", "disentanglement", "dynamical systems" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 2 ] }, { "paper_id": "d159zNCmOq", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Lipeng Zu", "gender": "Male", "institution": "Florida State University", "country": "US", "position": "PhD student" }, { "name": "Shayok Chakraborty", "gender": "Male", "institution": "Florida State University", "country": "US", "position": "Associate Professor" }, { "name": "Xiaonan Zhang", "gender": "Female", "institution": "Florida State University", "country": "US", "position": "Assistant Professor" }, { "name": "Yukun Yuan", "gender": "Male", "institution": "University of Tennessee at Chattanooga", "country": "US", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 29, "n_ref": 39, "n_ref_all": 50, "n_fig": 6, "n_tab": 2, "L_tab": 1462, "n_element_tab": 104, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1462, "n_element_tab_1": 104, "formula_len_all": 679, "formula_len_all_1": 699, "len_all": 95992, "len_all_1": 61109, "len_abs": 1339, "len_title": 178, "len_sents": 29386, "len_sents_1": 27773, "n_sents": 205, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 130, "L_abs": 1370, "title": "From Static to Dynamic: Leveraging Implicit Behavioral Models to Facilitate Transition in Offline-to-Online Reinforcement Learning", "abs": "Transitioning reinforcement learning (RL) models from offline training environments to dynamic online settings faces critical challenges because of the distributional shift and the model inability in effectively adapting to new, unseen scenarios. This work proposes the \\textbf{B}ehavior \\textbf{A}daption \\textbf{Q}-Learning (BAQ), a novel framework facilitating smoother transitions in offline-to-online RL. BAQ strategically leverages the implicit behavioral model to imitate and adapt behaviors of offline datasets, enabling the model to handle out-of-distribution state-action pairs more effectively during its online deployment. The key to our approach is the integration of a composite loss function that not only mimics the offline data-driven policy but also dynamically adjusts to new experiences encountered online. This dual-focus mechanism enhances the model's adaptability and robustness, reducing Q-value estimation errors and improving the overall learning efficiency. Extensive empirical evaluations demonstrate that BAQ significantly outperforms existing methods, achieving enhanced adaptability and reduced performance degradation in diverse RL settings. Our framework sets a new standard for offline-to-online RL, offering a robust solution for applications requiring reliable transitions from theoretical training to practical, real-world execution.", "keywords": [ "Offline-to-Online Reinforcement Learning", "Behavioral Adaptation", "Q-value Estimation", "Priority Sampling Strategy" ], "rating_list": [ 3, 3, 5, 3, 3 ], "soundness_list": [ 1, 2, 3, 2, 2 ], "presentation_list": [ 2, 1, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 1, 2 ], "confidence_list": [ 5, 5, 4, 3, 3 ] }, { "paper_id": "d0tlL0ZWlu", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Po-han Li", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Mohd Omama", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Sandeep Chinchali", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 22, "n_ref_uni": 40, "n_ref": 65, "n_ref_all": 92, "n_fig": 13, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 806, "formula_len_all_1": 806, "len_all": 118329, "len_all_1": 66275, "len_abs": 1701, "len_title": 126, "len_sents": 35051, "len_sents_1": 32096, "n_sents": 271, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1708, "title": "EXPLOITING DISTRIBUTION CONSTRAINTS FOR SCALABLE AND EFFICIENT IMAGE RETRIEVAL", "abs": "Image retrieval is crucial in robotics and computer vision, with downstream applications in robot place recognition and vision-based product recommendations. Modern retrieval systems face two key challenges: scalability and efficiency.\nState-of-the-art image retrieval systems train specific neural networks for each dataset, an approach that lacks scalability. Furthermore, since retrieval speed is directly proportional to embedding size, existing systems that use large embeddings lack efficiency. To tackle scalability, recent works propose using off-the-shelf foundation models. However, these models, though applicable across datasets, fall short in achieving performance comparable to that of dataset-specific models. Our key observation is that, while foundation models capture necessary subtleties for effective retrieval, the underlying distribution of their embedding space can negatively impact cosine similarity searches. We introduce Autoencoders with Strong Variance Constraints (AE-SVC), which, when used for projection, significantly improves the performance of foundation models. We provide an in-depth theoretical analysis of AE-SVC. Addressing efficiency, we introduce Single-Shot Similarity Space Distillation ((SS)2D), a novel approach to learn embeddings with adaptive sizes that offers a better trade-off between size and performance. We conducted extensive experiments on four retrieval datasets, including Stan-\nford Online Products (SoP) and Pittsburgh30k, using four different off-the-shelf foundation models, including DinoV2 and CLIP. AE-SVC demonstrates up to a 16% improvement in retrieval performance, while (SS)2D shows a further 10% improvement for smaller embedding sizes.", "keywords": [ "image retrieval", "efficiency", "foundation models" ], "rating_list": [ 6, 6, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 3, 3, 4 ] }, { "paper_id": "czvVNVLr7R", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Chenhui Zhao", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Liyue Shen", "gender": "Female", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 55, "n_ref": 150, "n_ref_all": 189, "n_fig": 13, "n_tab": 11, "L_tab": 1560, "n_element_tab": 200, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1779, "n_element_tab_1": 158, "formula_len_all": 203, "formula_len_all_1": 94, "len_all": 190106, "len_all_1": 67026, "len_abs": 1256, "len_title": 128, "len_sents": 49484, "len_sents_1": 27959, "n_sents": 390, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1436, "title": "Part-aware Personalized Segment Anything Model for Patient-Specific Segmentation", "abs": "Precision medicine, such as patient-adaptive treatments utilizing medical images, poses new challenges for image segmentation algorithms due to (1) the large variability across different patients and (2) the limited availability of annotated data for each patient. In this work, we propose a data-efficient segmentation method to address these challenges, namely $\\textit{\\textbf{P}art-aware}$ $\\textit{\\textbf{P}ersonalized}$ $\\textit{\\textbf{S}egment}$ $\\textit{\\textbf{A}nything}$ $\\textit{\\textbf{M}odel}$ ($\\mathbf{{P}^{2}SAM}$). Without any model fine-tuning, enables seamless adaptation to any new patients relying only on one-shot patient-specific data. We introduce a novel part-aware prompt mechanism to select multiple-point prompts based on part-level features of the one-shot data, which can be extensively integrated into different promptable segmentation models, such as SAM and SAM 2. To further promote the robustness of the selected part-aware prompt, we propose a distribution-similarity-based retrieval approach to determine the optimal number of part-level features for a specific case. $\\text{P}^{\\text{2}}\\text{SAM}$ improves the performance by $\\texttt{+} 8.0$% and $\\texttt{+} 2.0$% mean Dice score within two patient-specific segmentation tasks, and exhibits impressive generality across different domains, $\\textit{e.g.}$, $\\texttt{+} 6.4$% mIoU on the PerSeg benchmark. Code will be released upon acceptance.", "keywords": [ "Precision Medicine; Patient-Specific Segmentation; Out-of-Distribution Patient Adaptation" ], "rating_list": [ 8, 5, 5, 1 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 4, 3, 2 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "cznqgb4DNv", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Christopher Brinton", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "Dong-Jun Han", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Rohit Parasnis", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Seyyedali Hosseinalipour", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Shahryar Zehtabi", "gender": "unknown", "institution": "Purdue University", "country": "", "position": "PhD student" } ], "n_formula": 303, "n_formula_1": 27, "n_ref_uni": 48, "n_ref": 108, "n_ref_all": 179, "n_fig": 33, "n_tab": 4, "L_tab": 720, "n_element_tab": 70, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 327, "n_element_tab_1": 36, "formula_len_all": 49342, "formula_len_all_1": 2608, "len_all": 359463, "len_all_1": 76333, "len_abs": 1683, "len_title": 151, "len_sents": 101945, "len_sents_1": 31908, "n_sents": 1020, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 102, "L_abs": 1521, "title": "Decentralized Sporadic Federated Learning: A Unified Algorithmic Framework with Convergence Guarantees", "abs": "Decentralized federated learning (DFL) captures FL settings where both (i) model updates and (ii) model aggregations are exclusively carried out by the clients without a central server. Existing DFL works have mostly focused on settings where clients conduct a fixed number of local updates between local model exchanges, overlooking heterogeneity and dynamics in communication and computation capabilities. In this work, we propose Decentralized Sporadic Federated Learning ($\\texttt{DSpodFL}$), a DFL methodology built on a generalized notion of *sporadicity* in both local gradient and aggregation processes. $\\texttt{DSpodFL}$ subsumes many existing decentralized optimization methods under a unified algorithmic framework by modeling the per-iteration (i) occurrence of gradient descent at each client and (ii) exchange of models between client pairs as arbitrary indicator random variables, thus capturing *heterogeneous and time-varying* computation/communication scenarios. We analytically characterize the convergence behavior of $\\texttt{DSpodFL}$ for both convex and non-convex models and for both constant and diminishing learning rates, under mild assumptions on the communication graph connectivity, data heterogeneity across clients, and gradient noises. We show how our bounds recover existing results from decentralized gradient descent as special cases. Experiments demonstrate that $\\texttt{DSpodFL}$ consistently achieves improved training speeds compared with baselines under various system settings.", "keywords": [ "Decentralized Federated Learning", "Sporadicity", "Unified Algorithmic Framework", "Convergence Analysis" ], "rating_list": [ 8, 8, 8, 5, 8 ], "soundness_list": [ 4, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "czVzzXPCkw", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Hao Zhou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jianbing Zhang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Pan Mianzhi", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Wei-Ying Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yawen Ouyang", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "\u5251\u98de \u674e", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 31, "n_ref": 69, "n_ref_all": 89, "n_fig": 6, "n_tab": 5, "L_tab": 6035, "n_element_tab": 311, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 4881, "n_element_tab_1": 301, "formula_len_all": 513, "formula_len_all_1": 513, "len_all": 117578, "len_all_1": 64386, "len_abs": 1328, "len_title": 97, "len_sents": 27095, "len_sents_1": 25406, "n_sents": 199, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1334, "title": "ON EXTRAPOLATION IN MATERIAL PROPERTY REGRESSION", "abs": "Deep learning methods have yielded exceptional performances in material property regression (MPR). However, most existing methods operate under the assumption that the training and test are independent and identically distributed (i.i.d.). This overlooks the importance of extrapolation - predicting material properties beyond the range of training data - which is essential for advanced material discovery, as researchers strive to identify materials with exceptional properties that exceed current capabilities. In this paper, we address this gap by introducing a comprehensive benchmark comprising seven tasks specifically designed to evaluate extrapolation in MPR. We critically evaluate existing methods including deep imbalanced regression (DIR) and regression data augmentation (DA) methods, and reveal their limitations in extrapolation tasks. To address these issues, we propose the Matching-based EXtrapolation (MEX) framework, which reframes MPR as a material-property matching problem to alleviate the inherent complexity of the direct material-to-label mapping paradigm for better extrapolation. Our experimental results show that MEX outperforms all existing methods on our benchmark and demonstrates exceptional capability in identifying promising materials, underscoring its potential for advancing material discovery.", "keywords": [ "material property prediction", "regression", "extrapolation" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 1, 3, 2 ], "confidence_list": [ 2, 4, 3 ] }, { "paper_id": "cywG53B2ZQ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Biao Liu", "gender": "Male", "institution": "Southeast University", "country": "", "position": "PhD student" }, { "name": "Ning Xu", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Shiqi Qiao", "gender": "Female", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "Xin Geng", "gender": "Male", "institution": "Southeast University, China", "country": "CN", "position": "Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 21, "n_ref": 47, "n_ref_all": 64, "n_fig": 2, "n_tab": 5, "L_tab": 1317, "n_element_tab": 73, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 1597, "n_element_tab_1": 78, "formula_len_all": 412, "formula_len_all_1": 412, "len_all": 105197, "len_all_1": 59746, "len_abs": 1395, "len_title": 111, "len_sents": 27464, "len_sents_1": 27008, "n_sents": 208, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1403, "title": "Negative-Prompt-driven Alignment for Generative Language Model", "abs": "Large language models have achieved remarkable capabilities, but aligning their outputs with human values and preferences remains a significant challenge. Existing alignment methods primarily focus on positive examples while overlooking the importance of negative responses in guiding models away from undesirable behaviors. For instance, the widely-used alignment datasets reveals a scarcity of explicit negative examples that contradict human values, hindering its ability to discourage harmful or biased outputs during training. To address this limitation, we propose NEAT, i.e., NEgative-prompt-driven AlignmenT, to introduce negative prompts to generate undesirable responses alongside positive examples during the optimization process. NEAT explicitly penalizes the model for producing harmful outputs, guiding it not only toward desirable behaviors but also steering it away from generating undesirable, biased responses. This dual feedback mechanism enables better alignment with human preferences, crucial in contexts where avoiding harm is paramount. Starting from a pre-trained language model, NEAT performs online alignment by incorporating a ranking loss derived from an expanded preference dataset containing both positive and negative examples. Extensive experiments validate NEAT's effectiveness in significantly enhancing language models' alignment with human values and preferences.", "keywords": [ "AI Alignment", "Prompt-driven Online Sampling" ], "rating_list": [ 3, 3, 1, 3 ], "soundness_list": [ 2, 3, 1, 1 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 2, 3, 1, 1 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "cya3eEczAx", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Huihuang Cai", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "MS student" }, { "name": "Jingcheng Wang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Shunyu Wu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xinwei Xiao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 81, "n_formula_1": 31, "n_ref_uni": 31, "n_ref": 49, "n_ref_all": 56, "n_fig": 1, "n_tab": 4, "L_tab": 825, "n_element_tab": 87, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1602, "n_element_tab_1": 148, "formula_len_all": 5841, "formula_len_all_1": 1652, "len_all": 130380, "len_all_1": 61894, "len_abs": 1486, "len_title": 148, "len_sents": 36688, "len_sents_1": 27442, "n_sents": 308, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1492, "title": "Adaptive Proximal Gradient Optimizer: Addressing Gradient Inexactness in Predict+Optimize Framework", "abs": "To achieve end-to-end optimization in the Predict+Optimize (P+O) framework, efforts have been focused on constructing surrogate loss functions to replace the non-differentiable decision regret. \nWhile these surrogate functions are effective in forwarding training, the backpropagation of the gradient introduces a significant but unexplored problem: the inexactness of the surrogate gradient, which often destabilizes the training process. To address this challenge, we propose the Adaptive Proximal Gradient Optimizer (AProx), the first gradient descent optimizer designed to handle the inexactness of surrogate gradient backpropagation within the P+O framework. \nInstead of explicitly solving proximal operations, AProx uses subgradients to approximate the proximal operator, simplifying the computational complexity and making proximal gradient descent feasible within the P+O framework. We prove that the surrogate gradients of three major types of surrogate functions are subgradients, allowing efficient application of AProx to end-to-end optimization.\nAdditionally, AProx introduces momentum and novel strategies for adaptive weight decay and parameter smoothing, which together enhance both training stability and convergence speed.\nThrough experiments on several classical combinatorial optimization benchmarks using different surrogate functions, AProx demonstrates superior performance in stabilizing the training process and reducing the optimality gap under predicted parameters.", "keywords": [ "Predict+optimize", "Inexact gradient", "Proximal gradient descent", "Optimizer" ], "rating_list": [ 1, 3, 1 ], "soundness_list": [ 1, 1, 1 ], "presentation_list": [ 1, 3, 1 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "cyPMEXdqQ2", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Haian Yin", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Jin Zhang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "", "position": "Associate Professor" }, { "name": "Shangzhi Zeng", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Wei Yao", "gender": "unknown", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 129, "n_formula_1": 35, "n_ref_uni": 57, "n_ref": 100, "n_ref_all": 159, "n_fig": 5, "n_tab": 6, "L_tab": 776, "n_element_tab": 26, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 187, "n_element_tab_1": 6, "formula_len_all": 15938, "formula_len_all_1": 3325, "len_all": 262579, "len_all_1": 72791, "len_abs": 1352, "len_title": 155, "len_sents": 70099, "len_sents_1": 26899, "n_sents": 598, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 107, "L_abs": 1363, "title": "Overcoming Lower-Level Constraints in Bilevel Optimization: A Novel Approach with Regularized Gap Functions", "abs": "Constrained bilevel optimization tackles nested structures present in constrained learning tasks like constrained meta-learning, adversarial learning, and distributed bilevel optimization. \nHowever, existing bilevel optimization methods mostly are typically restricted to specific constraint settings, such as linear lower-level constraints. \nIn this work, we overcome this limitation and develop a new single-loop, Hessian-free constrained bilevel algorithm capable of handling more general lower-level constraints. \nWe achieve this by employing a doubly regularized gap function tailored to the constrained lower-level problem, transforming constrained bilevel optimization into an equivalent single-level optimization problem with a single smooth constraint. \nWe rigorously establish the non-asymptotic convergence analysis of the proposed algorithm under the convexity of lower-level problem, avoiding the need for strong convexity assumptions on the lower-level objective or coupling convexity assumptions on lower-level constraints found in existing literature. \nAdditionally, the generality of our method allows for its extension to bilevel optimization with minimax lower-level problem. \nWe evaluate the effectiveness and efficiency of our algorithm on various synthetic problems, typical hyperparameter learning tasks, and generative adversarial network.", "keywords": [ "bilevel optimization", "constrained optimization", "gap function", "single-loop", "Hessian-free", "convergence analysis" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "cxKLRM3KhC", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Michael Maire", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Associate Professor" }, { "name": "Rebecca Willett", "gender": "Female", "institution": "University of Chicago", "country": "US", "position": "Full Professor" }, { "name": "Ruoxi Jiang", "gender": "Female", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Will Gao", "gender": "Not Specified", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Undergrad student" }, { "name": "Xiao Zhang", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 63, "n_ref": 121, "n_ref_all": 142, "n_fig": 7, "n_tab": 3, "L_tab": 529, "n_element_tab": 94, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1190, "n_element_tab_1": 140, "formula_len_all": 303, "formula_len_all_1": 293, "len_all": 144905, "len_all_1": 73191, "len_abs": 1182, "len_title": 109, "len_sents": 38203, "len_sents_1": 33729, "n_sents": 259, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1192, "title": "Residual Connections Harm Generative Representation Learning", "abs": "We show that introducing a weighting factor to reduce the influence of identity shortcuts in residual networks significantly enhances semantic feature learning in generative representation learning frameworks, such as masked autoencoders (MAEs) and diffusion models. Our modification improves linear probing accuracy for both, notably increasing ImageNet accuracy from 67.8\\% to 72.7\\% for MAEs with a VIT-B/16 backbone, while also boosting generation quality for diffusion models. This significant gap suggests that, while residual connection structure serves an essential role in facilitating gradient propagation, it may have a harmful side effect of reducing capacity for abstract learning by virtue of injecting an echo of shallower representations into deeper layers. We ameliorate this downside via a fixed formula for monotonically decreasing the contribution of identity connections as layer depth increases. Our design promotes the gradual development of feature abstractions, without impacting network trainability. Analyzing the representations learned by our modified residual networks, we find correlation between low effective feature rank and downstream task performance.", "keywords": [ "Decayed Residual Connections", "Representation Learning" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "cxB0fPNZkx", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Bryan Kian Hsiang Low", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Associate Professor" }, { "name": "Lucas Agussurja", "gender": "unknown", "institution": "national university of singaore, National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Xinyang Lu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 21, "n_ref": 24, "n_ref_all": 34, "n_fig": 5, "n_tab": 3, "L_tab": 6600, "n_element_tab": 107, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2264, "n_element_tab_1": 57, "formula_len_all": 737, "formula_len_all_1": 664, "len_all": 102751, "len_all_1": 64580, "len_abs": 1218, "len_title": 106, "len_sents": 38605, "len_sents_1": 30465, "n_sents": 281, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1226, "title": "Global-to-Local Support Spectrums for Model Explainability", "abs": "Existing sample-based methods, like influence functions and representer points, measure the importance of a training point by approximating the effect of its removal from training. As such, they are skewed towards outliers and points that are very close to the decision boundaries. The explanations provided by these methods are often static and not specific enough for different test points. In this paper, we propose a method to generate an explanation in the form of support spectrums which are based on two main ideas: the support sets and a global-to-local importance measure. The support set is the set of training points, in the predicted class, that ``lie in between'' the test point and training points in the other classes. They indicate how well the test point can be distinguished from the points not in the predicted class. The global-to-local importance measure is obtained by decoupling existing methods into the global and local components which are then used to select the points in the support set. Using this method, we are able to generate explanations that are tailored to specific test points. In the experiments, we show the effectiveness of the method in image classification and text generation tasks.", "keywords": [ "sample-based explanations", "influence functions", "representer points" ], "rating_list": [ 1, 5, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 1 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "cx46JSD2qn", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Bin Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Feng Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Shixiong Kai", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Siyuan Xu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yuxi Qian", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Zhaojie Tu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Zhentao Tang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zhexuan Xu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "Undergrad student" }, { "name": "Zhihai Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zijie Geng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Ziyan Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 36, "n_ref": 75, "n_ref_all": 108, "n_fig": 4, "n_tab": 20, "L_tab": 5922, "n_element_tab": 858, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 5202, "n_element_tab_1": 336, "formula_len_all": 185, "formula_len_all_1": 185, "len_all": 149722, "len_all_1": 63094, "len_abs": 1997, "len_title": 121, "len_sents": 39521, "len_sents_1": 26503, "n_sents": 283, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 2030, "title": "Benchmarking End-To-End Performance of AI-Based Chip Placement Algorithms", "abs": "The increasing complexity of modern very-large-scale integration (VLSI) design highlights the significance of Electronic Design Automation (EDA) technologies. Chip placement is a critical step in the EDA workflow, which positions chip modules on the canvas with the goal of optimizing performance, power, and area (PPA) metrics of final chip designs. Recent advances have demonstrated the great potential of AI-based algorithms in enhancing chip placement. However, due to the lengthy workflow of chip design, the evaluations of these algorithms often focus on $\\textit{intermediate surrogate metrics}$, which are easy to compute but frequently reveal a substantial misalignment with the $\\textit{end-to-end performance}$ (i.e., the final design PPA). To address this challenge, we introduce ChiPBench, which can effectively facilitate research in chip placement within the AI community. ChiPBench is a comprehensive benchmark specifically designed to evaluate the effectiveness of existing AI-based chip placement algorithms in improving final design PPA metrics. Specifically, we have gathered $20$ circuits from various domains (e.g., CPU, GPU, and microcontrollers). These designs are compiled by executing the workflow from the verilog source code, which preserves necessary physical implementation kits, enabling evaluations for the placement algorithms on their impacts on the final design PPA. We executed six state-of-the-art AI-based chip placement algorithms on these designs and plugged the results of each single-point algorithm into the physical implementation workflow to obtain the final PPA results. Experimental results show that even if intermediate metric of a single-point algorithm is dominant, while the final PPA results are unsatisfactory. This suggests that the AI community should concentrate more on enhancing end-to-end performance rather than those intermediate surrogates. We believe that our benchmark will serve as an effective evaluation framework to bridge the gap between academia and industry.", "keywords": [ "Electronic Design Automation", "Chip Placement Algorithms", "End to End Performance Evaluation Benchmark", "Physical Design", "Reinforcement Learning and Evolutionary Algorithm" ], "rating_list": [ 3, 1, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 4, 3 ], "contribution_list": [ 1, 1, 2, 3, 2 ], "confidence_list": [ 5, 1, 3, 4, 5 ] }, { "paper_id": "cwuSAR7EKd", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Eunsol Choi", "gender": "unknown", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "W. Bradley Knox", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Michael J Zhang", "gender": "unknown", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 68, "n_ref_all": 85, "n_fig": 3, "n_tab": 8, "L_tab": 2816, "n_element_tab": 172, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 941, "n_element_tab_1": 70, "formula_len_all": 601, "formula_len_all_1": 549, "len_all": 167966, "len_all_1": 68035, "len_abs": 1472, "len_title": 124, "len_sents": 43452, "len_sents_1": 32377, "n_sents": 315, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1279, "title": "Modeling Future Conversation Turns to Teach LLMs to Ask Clarifying Questions", "abs": "Large language models (LLMs) must often respond to highly ambiguous user requests. In such cases, the LLM's best response may be to ask a clarifying question to elicit more information. We observe existing LLMs often respond by presupposing a single interpretation of such ambiguous requests, frustrating users who intended a different interpretation. We speculate this is caused by current preference data labeling practice, where LLM responses are evaluated only on their prior contexts. To address this, we propose to assign preference labels by simulating their expected outcomes in the future turns. This allows LLMs to learn to ask clarifying questions when it can generate responses that are tailored to each user interpretation in future turns. In experiments on open-domain QA, we compare systems that trained using our proposed preference labeling methods against standard methods, which assign preferences based on only prior context. We evaluate systems based on their ability to ask clarifying questions that can recover each user's interpretation and expected answer, and find that our training with our proposed method trains LLMs to ask clarifying questions with a 5% improvement in F1 measured against the answer set from different interpretations of each query.", "keywords": [ "Clarifying Questions", "QA", "Ambiguity", "RLHF" ], "rating_list": [ 5, 6, 6, 5, 8 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 1, 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 4, 3, 3 ], "confidence_list": [ 4, 4, 3, 4, 2 ] }, { "paper_id": "cweehsf2Fn", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "John A Onofrey", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" }, { "name": "Nicha C Dvornek", "gender": "Female", "institution": "Yale University", "country": "US", "position": "Assistant Professor" }, { "name": "Yuexi Du", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 49, "n_ref": 202, "n_ref_all": 220, "n_fig": 8, "n_tab": 10, "L_tab": 3284, "n_element_tab": 517, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2213, "n_element_tab_1": 429, "formula_len_all": 648, "formula_len_all_1": 593, "len_all": 218602, "len_all_1": 71085, "len_abs": 1674, "len_title": 143, "len_sents": 64435, "len_sents_1": 31871, "n_sents": 472, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1321, "title": "Multi-View and Multi-Scale Alignment for Contrastive Language-Image Pre-training in Mammography", "abs": "Contrastive Language-Image Pre-training (CLIP) shows promise in medical image analysis but requires substantial data and computational resources. Due to these restrictions, existing CLIP applications in medical imaging focus mainly on modalities like chest X-rays that have abundant image-report data available, leaving many other important modalities under-explored. Here, we propose one of the first adaptations of the full CLIP model to mammography, which presents significant challenges due to labeled data scarcity, high-resolution images with small regions of interest, and data imbalance. We first develop a specialized supervision framework for mammography that leverages its multi-view nature. Furthermore, we design a symmetric local alignment module to better focus on detailed features in high-resolution images. Lastly, we incorporate a parameter-efficient fine-tuning approach for large language models pre-trained with medical knowledge to address data limitations. Our multi-view and multi-scale alignment (MaMA) method outperforms state-of-the-art baselines for three different tasks on two large real-world mammography datasets, EMBED and RSNA-Mammo, with only 52% model size compared with the largest baseline. The code is attached in the supplement file and will be released on GitHub upon acceptance.", "keywords": [ "Contrastive Language-Image Pre-training", "Medical Image", "Mammography", "Multi-view & Multi-scale" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "cwbJxUGVOI", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Bo Yang", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "Yafei YANG", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Zihui Zhang", "gender": "Male", "institution": "The Hong Kong Polytechnic University, Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 10, "n_ref_uni": 67, "n_ref": 130, "n_ref_all": 176, "n_fig": 12, "n_tab": 16, "L_tab": 8018, "n_element_tab": 1248, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1140, "n_element_tab_1": 140, "formula_len_all": 1539, "formula_len_all_1": 933, "len_all": 278093, "len_all_1": 127171, "len_abs": 1203, "len_title": 137, "len_sents": 83988, "len_sents_1": 60777, "n_sents": 617, "n_sents_1": 429, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1223, "title": "OCN: Learning Object-centric Representations for Unsupervised Multi-object Segmentation", "abs": "We study the challenging problem of unsupervised multi-object segmentation on single images. By relying on an image reconstruction objective to learn objectness or leveraging pretrained image features to group similar pixels as objects, most existing methods can either segment simple synthetic objects or discover a rather limited number of real-world objects. In this paper, we introduce OCN, a new two stage pipeline to discover many complex objects on real-world images. The key to our approach is to explicitly learn our carefully defined three level object-centric representations in the first stage. After that, our multi-object reasoning module directly leverages the learned object priors to discover multiple objects in the second stage. Notably, such a reasoning module is completely network-free and does not need any human labels to train. Extensive experiments show that our OCN clearly surpasses all existing unsupervised methods by a large margin on 7 real-world benchmark datasets including the particularly challenging COCO dataset, achieving the state-of-the-art object segmentation results. Most notably, our method demonstrates superior results on extremely crowded images where all baselines collapse.", "keywords": [ "unsupervised learning", "object segmentation", "objectness representation" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 2, 4, 3 ] }, { "paper_id": "cv2iMNWCsh", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "David Moens", "gender": "Male", "institution": "KU Leuven", "country": "", "position": "Associate Professor" }, { "name": "Fabio Cuzzolin", "gender": "Male", "institution": "Oxford Brookes University", "country": "GB", "position": "Full Professor" }, { "name": "Hans Hallez", "gender": "Male", "institution": "KU Leuven, KU Leuven", "country": "", "position": "Associate Professor" }, { "name": "Kaizheng Wang", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "PhD student" }, { "name": "Keivan Shariatmadar", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "Postdoc" } ], "n_formula": 22, "n_formula_1": 13, "n_ref_uni": 78, "n_ref": 158, "n_ref_all": 218, "n_fig": 11, "n_tab": 16, "L_tab": 9458, "n_element_tab": 628, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 4223, "n_element_tab_1": 249, "formula_len_all": 1502, "formula_len_all_1": 1089, "len_all": 221705, "len_all_1": 74452, "len_abs": 1327, "len_title": 127, "len_sents": 60211, "len_sents_1": 32467, "n_sents": 414, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 78, "L_abs": 1327, "title": "Credal Wrapper of Model Averaging for Uncertainty Estimation in Classification", "abs": "This paper presents an innovative approach, called credal wrapper, to formulating a credal set representation of model averaging for Bayesian neural networks (BNNs) and deep ensembles (DEs), capable of improving uncertainty estimation in classification tasks. Given a finite collection of single predictive distributions derived from BNNs or DEs, the proposed credal wrapper approach extracts an upper and a lower probability bound per class, acknowledging the epistemic uncertainty due to the availability of a limited amount of distributions. Such probability intervals over classes can be mapped on a convex set of probabilities (a credal set) from which, in turn, a unique prediction can be obtained using a transformation called intersection probability transformation. In this article, we conduct extensive experiments on several out-of-distribution (OOD) detection benchmarks, encompassing various dataset pairs (CIFAR10/100 vs SVHN/Tiny-ImageNet, CIFAR10 vs CIFAR10-C, CIFAR100 vs CIFAR100-C and ImageNet vs ImageNet-O) and using different network architectures (such as VGG16, ResNet-18/50, EfficientNet B2, and ViT Base). Compared to the BNN and DE baselines, the proposed credal wrapper method exhibits superior performance in uncertainty estimation and achieves a lower expected calibration error on corrupted data.", "keywords": [ "Uncertainty Estimation", "Model Averaging", "Credal Stes", "Probability Intervals", "Out-of-Distribution Detection" ], "rating_list": [ 5, 8, 8, 8 ], "soundness_list": [ 2, 4, 4, 3 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "cuFzE8Jlvb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Chenhang HE", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "Eric Lin", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 57, "n_ref": 128, "n_ref_all": 147, "n_fig": 3, "n_tab": 7, "L_tab": 1438, "n_element_tab": 183, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 823, "n_element_tab_1": 80, "formula_len_all": 1096, "formula_len_all_1": 804, "len_all": 157735, "len_all_1": 64545, "len_abs": 980, "len_title": 139, "len_sents": 40822, "len_sents_1": 30101, "n_sents": 317, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 991, "title": "Continuous Autoregressive Modeling with Stochastic Monotonic Alignment for Speech Synthesis", "abs": "We propose a novel autoregressive modeling approach for speech synthesis, combining a variational autoencoder (VAE) with a multi-modal latent space and an autoregressive model that uses Gaussian Mixture Models (GMM) as the conditional probability distribution. Unlike previous methods that rely on residual vector quantization, our model leverages continuous speech representations from the VAE's latent space, greatly simplifying the training and inference pipelines. We also introduce a stochastic monotonic alignment mechanism to enforce strict monotonic alignments. Our approach significantly outperforms the state-of-the-art autoregressive model VALL-E in both subjective and objective evaluations, achieving these results with only 10.3\\% of VALL-E's parameters. This demonstrates the potential of continuous speech language models as a more efficient alternative to existing quantization-based speech language models. Sample audio can be found at \\url{https://tinyurl.com/gmm-lm-tts}.", "keywords": [ "Speech Synthesis;text-to-speech;" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "cuFnNExmdq", "primary_area": "learning on time series and dynamical systems", "n_author": 7, "author_info_list": [ { "name": "Bryan Hooi", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Chenghao Liu", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Researcher" }, { "name": "Doyen Sahoo", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Researcher" }, { "name": "Gerald Woo", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "PhD student" }, { "name": "Juncheng Liu", "gender": "unknown", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Yiwei Wang", "gender": "Male", "institution": "University of California, Merced", "country": "US", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 27, "n_ref": 93, "n_ref_all": 117, "n_fig": 10, "n_tab": 5, "L_tab": 4849, "n_element_tab": 136, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1141, "n_element_tab_1": 93, "formula_len_all": 772, "formula_len_all_1": 712, "len_all": 123953, "len_all_1": 54931, "len_abs": 1739, "len_title": 160, "len_sents": 37673, "len_sents_1": 25418, "n_sents": 250, "n_sents_1": 155, "venue": "ICLR2025", "decision": "Reject", "L_title": 112, "L_abs": 1117, "title": "UniTST: Effectively Modeling Inter-Series and Intra-Series Dependencies for Multivariate Time Series Forecasting", "abs": "Transformer-based models have emerged as powerful tools for multivariate time series forecasting (MTSF). However, existing Transformer models often fall short of capturing both intricate dependencies across variate and temporal dimensions in MTS data. Some recent models are proposed to separately capture variate and temporal dependencies through either two sequential or parallel attention mechanisms. However, these methods cannot directly and explicitly learn the intricate inter-series and intra-series dependencies. In this work, we first demonstrate that these dependencies are very important as they usually exist in real-world data. To directly model these dependencies, we propose a transformer-based model UniTST containing a unified attention mechanism on the flattened patch tokens. Additionally, we add a dispatcher module which reduces the complexity and makes the model feasible for a potentially large number of variates. Although our proposed model employs a simple architecture, it offers compelling performance as shown in our extensive experiments on several datasets for time series forecasting.", "keywords": [ "Multivariate Time Series Forecasting" ], "rating_list": [ 8, 3, 3, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "cu8qfq62Lv", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Robbert Reijnen", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "", "position": "PhD student" }, { "name": "Yingqian Zhang", "gender": "Female", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Associate Professor" }, { "name": "Zaharah Bukhsh", "gender": "unknown", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "YAOXIN WU", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 37, "n_ref": 65, "n_ref_all": 80, "n_fig": 4, "n_tab": 11, "L_tab": 4012, "n_element_tab": 447, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2583, "n_element_tab_1": 141, "formula_len_all": 1078, "formula_len_all_1": 759, "len_all": 160401, "len_all_1": 73200, "len_abs": 1031, "len_title": 143, "len_sents": 57552, "len_sents_1": 36010, "n_sents": 366, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1034, "title": "Graph-Supported Dynamic Algorithm Configuration for Multi-Objective Combinatorial Optimization", "abs": "Deep reinforcement learning (DRL) has been widely used for dynamic algorithm configuration, especially for evolutionary algorithms, which benefit from adaptive update of parameters during the algorithmic execution. However, applying DRL to algorithm configuration for multi-objective combinatorial optimization (MOCO) problems remains relatively unexplored. This paper presents a novel graph neural network (GNN) based DRL to configure multi-objective evolutionary algorithms. We model the dynamic algorithm configuration as a Markov decision process, representing the convergence of solutions in the objective space by a graph, with their embeddings learned by a GNN to enhance the state representation. Experiments on diverse MOCO challenges indicate that our method outperforms traditional and DRL-based algorithm configuration methods in terms of efficacy and adaptability. It also exhibits advantageous generalizability across objective types and problem sizes, and prospective applicability to different evolutionary algorithms.", "keywords": [ "Deep Reinforcement Learning", "Multi-objective Combinatorial Optimization", "Dynamic Algorithm Configuration", "Evolutionary algorithm" ], "rating_list": [ 5, 8, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "cu2CT2VAvs", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Maosong Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xinrong Zhang", "gender": "Female", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Xu Han", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yingfa Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhiyuan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "shengding hu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 69, "n_ref_all": 100, "n_fig": 14, "n_tab": 3, "L_tab": 152, "n_element_tab": 37, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 85, "n_element_tab_1": 2, "formula_len_all": 852, "formula_len_all_1": 589, "len_all": 156765, "len_all_1": 57548, "len_abs": 1849, "len_title": 133, "len_sents": 45607, "len_sents_1": 27309, "n_sents": 375, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1861, "title": "Stuffed Mamba: State Collapse and State Capacity of RNN-Based Long-Context Modeling", "abs": "One essential advantage of recurrent neural networks (RNNs) over transformer-based language models is their linear computational complexity concerning the sequence length, which makes them much faster in handling long sequences during inference. However, most publicly available RNNs (e.g., Mamba and RWKV) are trained on sequences with less than 10K tokens, and their effectiveness in longer contexts remains largely unsatisfying so far. In this paper, we study the cause of the inability to process long context for RNNs and suggest critical mitigations. First, we investigate *state explosion* (SE) in Mamba-2 when processing long sequences, a phenomenon where some channels of the state exhibit exploding values that cause severe performance degradation. With controlled experiments, we discover that the model fails to forget the earlier tokens when there is more information than it can remember. We attribute this to overfitting due to the recurrent state being overparameterized for the training length, thereby establishing a relationship between SE and the capacity of the state. To support this hypothesis, we make an important empirical observation: for any given state size, there exists a training length threshold such that SE is exhibited if and only if the training length is greater than this threshold. Empirically searching for this threshold for different state sizes reveals that it is a linear function of the state size. We also search for the maximum context length at which the model can recall contextual information and find that this context length scales exponentially to the state size. Based on this, we empirically train a Mamba-2 370M with near-perfect passkey retrieval accuracy on 256K context length. This suggests a promising future for RNN-based long-context modeling. Code and model checkpoints will be publicly released.", "keywords": [ "RNN", "foundation models", "long-context" ], "rating_list": [ 6, 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2, 1 ], "presentation_list": [ 3, 1, 2, 1, 1 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 2, 4, 4, 4, 2 ] }, { "paper_id": "ctzGqxE3O0", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "CHANGHAI OU", "gender": "Not Specified", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Chen Ling", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Fawen Li", "gender": "unknown", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Tingting Wang", "gender": "Female", "institution": "Digital Technologies Ltd", "country": "CN", "position": "Associate Professor" }, { "name": "XINGSHUO HAN", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Yao Shiyi", "gender": "Female", "institution": "Wuhan University", "country": "CN", "position": "Undergrad student" }, { "name": "Zixuan Huang", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 24, "n_formula_1": 24, "n_ref_uni": 15, "n_ref": 17, "n_ref_all": 29, "n_fig": 1, "n_tab": 4, "L_tab": 1738, "n_element_tab": 308, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1738, "n_element_tab_1": 308, "formula_len_all": 938, "formula_len_all_1": 938, "len_all": 76410, "len_all_1": 68680, "len_abs": 1206, "len_title": 100, "len_sents": 22762, "len_sents_1": 22762, "n_sents": 173, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1211, "title": "BID: Broad Incremental for Android Malware Detection", "abs": "With the rapid rise of mobile devices, the threat of malware targeting these platforms has escalated significantly. The fast-paced evolution of Android malware and new attack patterns frequently introduce substantial challenges for detection systems. Although many methods have achieved excellent results, they need to be retrained when faced with new attack modes or observation objects, and it is challenging to attain dynamic updates. To address this issue, we propose a novel Broad Incremental Detection (BID) method for real-time Android malware detection. Our method leverages incremental function to achieve dynamic adaptation to the growing variety of malware attacks while maintaining high computational efficiency, benefiting from its lightweight shallow network architecture. We also develop relational structures to capture complex relations and features of history attacks by fine-turning the network's weights unsupervised. Experimental results across three datasets demonstrate that BID achieves superior detection accuracy and computational efficiency compared to state-of-the-art approaches. Our work presents a robust, flexible, and lightweight framework for dynamic Android malware detection.", "keywords": [ "Broad learning system", "Android malware detection", "Incremental learning", "Relational structure" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "ctSjIlYN74", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Jaeseok Jeong", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "MS student" }, { "name": "Kim Shin Seong", "gender": "Male", "institution": "Yonsei University", "country": "KP", "position": "PhD student" }, { "name": "Mingi Kwon", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "PhD student" }, { "name": "Youngjung Uh", "gender": "unknown", "institution": "Yonsei University", "country": "KR", "position": "Associate Professor" } ], "n_formula": 28, "n_formula_1": 27, "n_ref_uni": 27, "n_ref": 45, "n_ref_all": 82, "n_fig": 23, "n_tab": 7, "L_tab": 3418, "n_element_tab": 335, "n_fig_1": 18, "n_tab_1": 1, "L_tab_1": 444, "n_element_tab_1": 59, "formula_len_all": 2156, "formula_len_all_1": 1962, "len_all": 137311, "len_all_1": 59195, "len_abs": 1668, "len_title": 91, "len_sents": 37606, "len_sents_1": 24168, "n_sents": 303, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 29, "L_abs": 1703, "title": "Balanced conic rectified flow", "abs": "Rectified flow is a generative model that learns smooth transport mappings between two distributions through an ordinary differential equation (ODE). Unlike diffusion-based generative models, which require costly numerical integration of a generative ODE to sample images with state-of-the-art quality, rectified flow uses an iterative process called reflow to learn smooth and straight ODE paths. This allows for relatively simple and efficient generation of high-quality images. However, rectified flow still faces several challenges. 1) The reflow process requires a large number of generative pairs to preserve the target distribution, leading to significant computational costs. 2) Since the model is typically trained using only generated image pairs, its performance heavily depends on the 1-rectified flow model, causing it to become biased towards the generated data.\n\nIn this work, we experimentally expose the limitations of the original rectified flow and propose a novel approach that incorporates real images into the training process. By preserving the ODE paths for real images, our method effectively reduces reliance on large amounts of generated data. Instead, we demonstrate that the reflow process can be conducted efficiently using a much smaller set of generated and real images. In CIFAR-10, we achieved significantly better FID scores, not only in one-step generation but also in full-step simulations, while using only $7.2\\%$ of the generative pairs compared to the original method. Furthermore, our approach induces straighter paths and avoids saturation on generated images during reflow, leading to more robust ODE learning while preserving the distribution of real images.", "keywords": [ "Image generation", "generative model", "rectified flow", "diffusion", "optimal transport", "curvature", "ODE" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "cselR6Jne3", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Alexander Pan", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Jacob Steinhardt", "gender": "unknown", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" }, { "name": "Lijie Chen", "gender": "Unspecified", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 50, "n_ref": 105, "n_ref_all": 138, "n_fig": 14, "n_tab": 6, "L_tab": 5586, "n_element_tab": 313, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 1159, "n_element_tab_1": 110, "formula_len_all": 89, "formula_len_all_1": 0, "len_all": 221650, "len_all_1": 68866, "len_abs": 1042, "len_title": 57, "len_sents": 67954, "len_sents_1": 32406, "n_sents": 643, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1075, "title": "Teaching LLMs to Decode Activations Into Natural Language", "abs": "Interpretability methods seek to understand language model representations, yet the outputs of most such methods---circuits, vectors, scalars---are uninterpretable, requiring further effort to interpret. In contrast, we propose to study LatentQA, the task of answering open-ended questions about model activations in natural language. Towards solving LatentQA, we propose Latent Interpretation Tuning (LIT), which finetunes a decoder LLM on a dataset of activations and associated question-answer pairs, similar to how visual instruction tuning trains on question-answer pairs associated with images. We use the decoder for diverse reading applications, such as extracting relational knowledge from representations or uncovering system prompts governing model behavior. Our decoder also specifies a differentiable loss that we use to control models, such as debiasing models on stereotyped sentences and controlling the sentiment of generations. Finally, we extend LatentQA to reveal harmful model capabilities, such as generating recipes for bioweapons and code for hacking.", "keywords": [ "llm safety", "llm interpretability", "ml safety", "activation steering" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "csbf1p8xUq", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Akiko Eriguchi", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Senior Researcher" }, { "name": "Haoran Xu", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Hieu Hoang", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Huda Khayrallah", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Kenton Murray", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Researcher" }, { "name": "Philipp Koehn", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 8, "n_ref_uni": 44, "n_ref": 106, "n_ref_all": 128, "n_fig": 7, "n_tab": 8, "L_tab": 5289, "n_element_tab": 446, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 499, "n_element_tab_1": 30, "formula_len_all": 733, "formula_len_all_1": 470, "len_all": 184926, "len_all_1": 59798, "len_abs": 1382, "len_title": 131, "len_sents": 40081, "len_sents_1": 30373, "n_sents": 410, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 83, "L_abs": 1368, "title": "X-ALMA: Plug & Play Modules and Adaptive Rejection for Quality Translation at Scale", "abs": "Large language models (LLMs) have achieved remarkable success across various NLP tasks, yet their focus has predominantly been on English due to English-centric pre-training and limited multilingual data. While some multilingual LLMs claim to support for hundreds of languages, models often fail to provide high-quality response for mid- and low-resource languages, leading to imbalanced performance heavily skewed in favor of high-resource languages like English and Chinese. In this paper, we prioritize quality over scaling number of languages, with a focus on multilingual machine translation task, and introduce **X-ALMA**, a model designed with a commitment to ensuring top-tier performance across 50 diverse languages, regardless of their resource levels. X-ALMA surpasses state-of-the-art open-source multilingual LLMs, such as Aya-101 and Aya-23, in every single translation direction on the FLORES and WMT'23 test datasets according to COMET-22. This is achieved by plug-and-play language-specific module architecture to prevent language conflicts during training and a carefully designed training regimen with novel optimization methods to maximize the translation performance. At the final stage of training regimen, our proposed Adaptive Rejection Preference Optimization (**ARPO**) surpasses existing preference optimization methods in translation tasks.", "keywords": [ "Large Language Model", "Machine Translation", "Multilingual" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 4, 3, 4 ], "contribution_list": [ 4, 3, 3 ], "confidence_list": [ 4, 3, 5 ] }, { "paper_id": "cs8dm8MgOT", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Anh Nguyen", "gender": "Male", "institution": "University of Liverpool", "country": "GB", "position": "Associate Professor" }, { "name": "Ismail Elezi", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Researcher" }, { "name": "Jiankang Deng", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Researcher" }, { "name": "Konstantinos Panagiotis Alexandridis", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Researcher" }, { "name": "Shan Luo", "gender": "Male", "institution": "King's College London, University of London", "country": "GB", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 25, "n_ref": 58, "n_ref_all": 94, "n_fig": 9, "n_tab": 6, "L_tab": 4843, "n_element_tab": 207, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2748, "n_element_tab_1": 145, "formula_len_all": 978, "formula_len_all_1": 779, "len_all": 179627, "len_all_1": 61985, "len_abs": 1381, "len_title": 100, "len_sents": 33728, "len_sents_1": 25458, "n_sents": 262, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 1385, "title": "FRACTAL CALIBRATION FOR LONG-TAILED OBJECT DETECTION", "abs": "Real-world datasets follow an imbalanced distribution, which poses significant\nchallenges in rare-category object detection. Recent studies tackle this problem\nby developing re-weighting and re-sampling methods, that utilise the class frequencies of the dataset. However, these techniques focus solely on the frequency statistics and ignore the distribution of the classes in image space, missing important information. In contrast to them, we propose Fractal CALibration (FRACAL): a novel post-calibration method for long-tailed object detection. FRACAL\ndevises a logit adjustment method that utilises the fractal dimension to estimate how uniformly classes are distributed in image space. During inference, it uses the fractal dimension to inversely downweight the probabilities of uniformly spaced class predictions achieving balance in two axes: between frequent and rare categories, and between uniformly spaced and sparsely spaced classes. FRACAL is a\npost-processing method and it does not require any training, also it can be combined with many off-the-shelf models such as one-stage sigmoid detectors and two-stage instance segmentation models. FRACAL boosts the rare class performance by up to 8.6% and surpasses all previous methods on LVIS dataset, while showing good generalisation to other datasets such as COCO, V3Det and OpenImages. We provide the code in the Appendix.", "keywords": [ "long-tail learning", "imbalanced object detection", "long-tailed instance segmentation" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "cqsw28DuMW", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Han Bao", "gender": "Male", "institution": "Kyoto University, Kyoto University", "country": "JP", "position": "Assistant Professor" }, { "name": "Kou Misaki", "gender": "Male", "institution": "Sakana AI", "country": "JP", "position": "Researcher" }, { "name": "Makoto Shing", "gender": "Male", "institution": "Sakana AI", "country": "JP", "position": "Researcher" }, { "name": "Sho Yokoi", "gender": "unknown", "institution": "NINJAL", "country": "JP", "position": "Associate Professor" }, { "name": "Takuya Akiba", "gender": "unknown", "institution": "Sakana AI", "country": "JP", "position": "Researcher" } ], "n_formula": 32, "n_formula_1": 5, "n_ref_uni": 53, "n_ref": 116, "n_ref_all": 153, "n_fig": 3, "n_tab": 4, "L_tab": 1365, "n_element_tab": 116, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1602, "n_element_tab_1": 159, "formula_len_all": 3213, "formula_len_all_1": 303, "len_all": 203706, "len_all_1": 72578, "len_abs": 2466, "len_title": 152, "len_sents": 59158, "len_sents_1": 34291, "n_sents": 456, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 103, "L_abs": 1647, "title": "TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models", "abs": "Causal language models have demonstrated remarkable capabilities, but their size poses significant challenges for deployment in resource-constrained environments. Knowledge distillation, a widely-used technique for transferring knowledge from a large teacher model to a small student model, presents a promising approach for model compression.\nA significant remaining issue lies in the major differences between teacher and student models, namely the substantial capacity gap, mode averaging, and mode collapse, which pose barriers during distillation.\nTo address these issues, we introduce $\\textit{Temporally Adaptive Interpolated Distillation (TAID)}$, a novel knowledge distillation approach that dynamically interpolates student and teacher distributions through an adaptive intermediate distribution, gradually shifting from the student's initial distribution towards the teacher's distribution. We provide a theoretical analysis demonstrating TAID's ability to prevent mode collapse and empirically show its effectiveness in addressing the capacity gap while balancing mode averaging and mode collapse.\nOur comprehensive experiments demonstrate TAID's superior performance across various model sizes and architectures in both instruction tuning and pre-training scenarios. Furthermore, we showcase TAID's practical impact by developing two state-of-the-art compact foundation models: $\\texttt{TAID-LLM-1.5B}$ for language tasks and $\\texttt{TAID-VLM-2B}$ for vision-language tasks.\nThese results demonstrate TAID's effectiveness in creating high-performing and efficient models, advancing the development of more accessible AI technologies.", "keywords": [ "Lanauge Models", "Knowledge Distillation" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 5, 5, 5 ] }, { "paper_id": "cqWD2dpDHW", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Hang Yu", "gender": "Male", "institution": "Shanghai University", "country": "CN", "position": "Full Professor" }, { "name": "Hu Zhang", "gender": "Male", "institution": "Commonwealth Scientific and Industrial Research Organisation, CSIRO", "country": "AU", "position": "Postdoc" }, { "name": "Ruiyang Zhang", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "PhD student" }, { "name": "Zhedong Zheng", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 55, "n_ref": 105, "n_ref_all": 131, "n_fig": 9, "n_tab": 4, "L_tab": 1765, "n_element_tab": 144, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1691, "n_element_tab_1": 137, "formula_len_all": 458, "formula_len_all_1": 451, "len_all": 168340, "len_all_1": 72830, "len_abs": 1566, "len_title": 128, "len_sents": 51027, "len_sents_1": 34683, "n_sents": 403, "n_sents_1": 252, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1575, "title": "Harnessing Uncertainty-aware Bounding Boxes for Unsupervised 3D Object Detection", "abs": "Unsupervised 3D object detection aims to identify objects of interest from unlabeled raw data, such as LiDAR points. Recent approaches usually adopt pseudo 3D bounding boxes (3D bboxes) from clustering algorithm to initialize the model training. However, pseudo bboxes inevitably contain noise, and such inaccuracies accumulate to the final model, compromising the performance. Therefore, in an attempt to mitigate the negative impact of inaccurate pseudo bboxes, we introduce a new uncertainty-aware framework for unsupervised 3D object detection, dubbed UA3D. In particular, our method consists of two phases: uncertainty estimation and uncertainty regularization. (1) In the uncertainty estimation phase, we incorporate an extra auxiliary detection branch alongside the original primary detector. The prediction disparity between the primary and auxiliary detectors could reflect fine-grained uncertainty at the box coordinate level. (2) Based on the assessed uncertainty, we adaptively adjust the weight of every 3D bbox coordinate via uncertainty regularization, refining the training process on pseudo bboxes. For pseudo bbox coordinate with high uncertainty, we assign a relatively low loss weight. Extensive experiments verify that the proposed method is robust against the noisy pseudo bboxes, yielding substantial improvements on nuScenes and Lyft compared to existing approaches, with increases of +6.9% AP_BEV and +2.5% AP_3D on nuScenes, and +4.1% AP_BEV and +2.0% AP_3D on Lyft. The anonymous code and checkpoints are at https://anonymous.4open.science/r/CBC6/.", "keywords": [ "Unsupervised 3D Object Detection", "Uncertainty Learning" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 2, 3, 1, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "cqU91W3LnB", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Graham Neubig", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Hongyin Luo", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Jiaxin Ge", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Vijay Viswanathan", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Xueying Jia", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 34, "n_ref": 64, "n_ref_all": 93, "n_fig": 8, "n_tab": 8, "L_tab": 961, "n_element_tab": 134, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 788, "n_element_tab_1": 122, "formula_len_all": 285, "formula_len_all_1": 352, "len_all": 133852, "len_all_1": 61383, "len_abs": 964, "len_title": 106, "len_sents": 36648, "len_sents_1": 27354, "n_sents": 316, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 951, "title": "Training Task Experts through Retrieval Based Distillation", "abs": "One of the most reliable ways to create deployable models for specialized tasks is to obtain an adequate amount of high-quality task-specific data. However, for specialized tasks, often such datasets do not exist. Existing methods address this by creating such data from large language models (LLMs) and then distilling such knowledge into smaller models. However, these methods are limited by the quality of the LLMs output, and tend to generate repetitive or incorrect data. In this work, we present Retrieval Based Distillation (ReBase), a method that first retrieves data from rich online sources and then transforms them into domain-specific data. This method greatly enhances data diversity. Moreover, ReBase generates Chain-of-Thought reasoning and distills the reasoning capacity of LLMs. We test our method on 4 benchmarks and shows that our method significantly improves performance by up to 10.76% on SQuAD, 1.37% on MNLI, and 1.94% on BBH.", "keywords": [ "Retrieval", "Distillation", "Task-Expert" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "cqTUJRlcLU", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Issei Sato", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Keitaro Sakamoto", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" } ], "n_formula": 360, "n_formula_1": 30, "n_ref_uni": 56, "n_ref": 113, "n_ref_all": 134, "n_fig": 7, "n_tab": 6, "L_tab": 2079, "n_element_tab": 159, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 34235, "formula_len_all_1": 1628, "len_all": 364386, "len_all_1": 69038, "len_abs": 1040, "len_title": 122, "len_sents": 103426, "len_sents_1": 30336, "n_sents": 1056, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1045, "title": "Benign or Not-Benign Overfitting in Token Selection of Attention Mechanism", "abs": "Modern over-parameterized neural networks can be trained to fit the training data perfectly while still maintaining a high generalization performance. This \u201cbenign overfitting\u201d phenomenon has been studied in a surge of recent theoretical work; however, most of these studies have been limited to linear models or two-layer neural networks. \nIn this work, we analyze benign overfitting in the token selection mechanism of the attention architecture, which characterizes the success of transformer models. We first show the existence of a benign overfitting solution and explain its mechanism in the attention architecture. Next, we discuss whether the model converges to such a solution, raising the difficulties specific to the attention architecture. We then present benign overfitting cases and not-benign overfitting cases by conditioning different scenarios based on the behavior of attention probabilities during training. To the best of our knowledge, this is the first study to characterize benign overfitting for the attention mechanism.", "keywords": [ "benign overfitting", "attention mechanism", "prompt tuning", "gradient descent" ], "rating_list": [ 5, 8, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 3, 2, 4, 3, 3 ] }, { "paper_id": "cphaRg46jD", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Daiqing Qi", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Dongliang Guo", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Hongyi Wu", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Mengxuan Hu", "gender": "Female", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Sheng Li", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Associate Professor" }, { "name": "Zhu Ronghang", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "zihan guan", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 76, "n_ref": 120, "n_ref_all": 150, "n_fig": 12, "n_tab": 7, "L_tab": 1615, "n_element_tab": 149, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 118, "n_element_tab_1": 7, "formula_len_all": 455, "formula_len_all_1": 0, "len_all": 224134, "len_all_1": 65205, "len_abs": 1601, "len_title": 147, "len_sents": 63014, "len_sents_1": 32352, "n_sents": 491, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1251, "title": "No Free Lunch: Retrieval-Augmented Generation Undermines Fairness in LLMs, Even for Vigilant Users", "abs": "Retrieval-Augmented Generation (RAG) is widely adopted for its effectiveness and cost-efficiency in mitigating hallucinations and enhancing the domain-specific generation capabilities of large language models (LLMs). However, is this effectiveness and cost-efficiency truly a free lunch? In this study, we comprehensively investigate the fairness costs associated with RAG by proposing a practical three-level threat model from the perspective of user awareness of fairness. Specifically, varying levels of user fairness awareness result in different degrees of fairness censorship on the external dataset. We examine the fairness implications of RAG using uncensored, partially censored, and fully censored datasets. Our experiments demonstrate that fairness alignment can be easily undermined through RAG **without the need for fine-tuning or retraining**. *Even with fully censored and supposedly unbiased external datasets, RAG can lead to biased outputs.* Our findings underscore the limitations of current alignment methods in the context of RAG-based LLMs and highlight the urgent need for new strategies to ensure fairness. We propose potential mitigations and call for further research to develop robust fairness safeguards in RAG-based LLMs.", "keywords": [ "Large Language Model", "RAG", "Fairness" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 1, 3, 4, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "cpZMsDwRie", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Huan Zhang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Yifan Sun", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Yuchen Jin", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "", "position": "PhD student" }, { "name": "Yue Zhang", "gender": "Male", "institution": "Hyperbolic Labs", "country": "US", "position": "Researcher" }, { "name": "Yuhang Li", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" } ], "n_formula": 22, "n_formula_1": 12, "n_ref_uni": 47, "n_ref": 80, "n_ref_all": 115, "n_fig": 7, "n_tab": 5, "L_tab": 1562, "n_element_tab": 209, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1032, "n_element_tab_1": 125, "formula_len_all": 1103, "formula_len_all_1": 625, "len_all": 184249, "len_all_1": 69988, "len_abs": 1725, "len_title": 119, "len_sents": 55580, "len_sents_1": 32364, "n_sents": 432, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1771, "title": "SVIP: Towards Verifiable Inference of Open-source Large Language Models", "abs": "Open-source Large Language Models (LLMs) have recently demonstrated remarkable capabilities in natural language understanding and generation, leading to widespread adoption across various domains. However, their increasing model sizes render local deployment impractical for individual users, pushing many to rely on decentralized computing service providers for inference through a blackbox API. This reliance introduces a new risk: a computing provider may stealthily substitute the requested LLM with a smaller, less capable model without consent from users, thereby delivering inferior outputs while benefiting from cost savings. In this paper, we formalize the problem of verifiable inference for LLMs. Existing verifiable computing solutions based on cryptographic or game-theoretic techniques are either computationally uneconomical or rest on strong assumptions. We introduce $\\texttt{SVIP}$, a secret-based verifiable LLM inference protocol that leverages intermediate outputs from LLMs as unique model identifiers. By training a proxy task on these outputs and requiring the computing provider to return both the generated text and the processed intermediate outputs, users can reliably verify whether the computing provider is acting honestly. In addition, the integration of a secret mechanism further enhances the security of our protocol. We thoroughly analyze our protocol under multiple strong and adaptive adversarial scenarios. Our extensive experiments demonstrate that $\\texttt{SVIP}$ is accurate, generalizable, computationally efficient, and resistant to various attacks. Notably, $\\texttt{SVIP}$ achieves false negative rates below $5\\\\%$ and false positive rates below $3\\\\%$, while requiring less than $0.01$ seconds per query for verification.", "keywords": [ "Large Language Models; Verifiable Inference; Trustworthy AI; Adversarial Attack" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 4, 2, 1 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "cpGPPLLYYx", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Ondrej Bohdal", "gender": "Male", "institution": "Samsung Research", "country": "GB", "position": "Researcher" }, { "name": "Timothy Hospedales", "gender": "Male", "institution": "University of Edinburgh", "country": "", "position": "Full Professor" }, { "name": "Yongshuo Zong", "gender": "unknown", "institution": "University of Edinburgh", "country": "GB", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 39, "n_ref": 120, "n_ref_all": 169, "n_fig": 17, "n_tab": 19, "L_tab": 8924, "n_element_tab": 922, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1086, "n_element_tab_1": 61, "formula_len_all": 93, "formula_len_all_1": 24, "len_all": 211561, "len_all_1": 57013, "len_abs": 1513, "len_title": 120, "len_sents": 51717, "len_sents_1": 27812, "n_sents": 391, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1497, "title": "VL-ICL Bench: The Devil in the Details of Multimodal In-Context Learning", "abs": "Large language models (LLMs) famously exhibit emergent in-context learning (ICL) - the ability to rapidly adapt to new tasks using few-shot examples provided as a prompt, without updating the model's weights. Built on top of LLMs, vision large language models (VLLMs) have advanced significantly in areas such as recognition, reasoning, and grounding. However, investigations into multimodal ICL have predominantly focused on few-shot visual question answering (VQA), and image captioning, which we will show neither exploit the strengths of ICL, nor test its limitations. The broader capabilities and limitations of multimodal ICL remain under-explored. In this study, we introduce a comprehensive benchmark VL-ICL Bench for multimodal in-context learning, encompassing a broad spectrum of tasks that involve both images and text as inputs and outputs, and different types of challenges, from {perception to reasoning and long context length}. We evaluate the abilities of state-of-the-art VLLMs against this benchmark suite, revealing their diverse strengths and weaknesses, and showing that even the most advanced models, such as GPT-4, find the tasks challenging. By highlighting a range of new ICL tasks, and the associated strengths and limitations of existing models, we hope that our dataset will inspire future work on enhancing the in-context learning capabilities of VLLMs, as well as inspire new applications that leverage VLLM ICL. Project page is at https://ys-zong.github.io/VL-ICL/", "keywords": [ "Vision-Language Models", "In-Context Learning", "LLMs" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "cp9LvuvAKW", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 13, "author_info_list": [ { "name": "Feng Zhao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Guodong Zheng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Jing Shao", "gender": "Female", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Jinlan Fu", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Lu Chen", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Rui Zheng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Senjie Jin", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Tao Gui", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xuanjing Huang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yifeng Gao", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Yongting Zhang", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhenfei Yin", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 1, "n_ref_uni": 42, "n_ref": 96, "n_ref_all": 117, "n_fig": 5, "n_tab": 24, "L_tab": 11431, "n_element_tab": 1106, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 3128, "n_element_tab_1": 458, "formula_len_all": 535, "formula_len_all_1": 20, "len_all": 279779, "len_all_1": 58812, "len_abs": 1345, "len_title": 135, "len_sents": 70447, "len_sents_1": 23856, "n_sents": 545, "n_sents_1": 165, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1274, "title": "SPA-VL: A Comprehensive Safety Preference Alignment Dataset for Vision Language Model", "abs": "The emergence of Vision Language Models (VLMs) has brought unprecedented advances in understanding multimodal information. The combination of textual and visual semantics in VLMs is highly complex and diverse, making the safety alignment of these models challenging. Furthermore, due to the limited study on the safety alignment of VLMs, there is a lack of large-scale, high-quality datasets. To address these limitations, we propose a Safety Preference Alignment dataset for Vision Language Models named SPA-VL. In terms of breadth, SPA-VL covers 6 harmfulness domains, 13 categories, and 53 subcategories, and contains 100,788 samples of the quadruple (question, image, chosen response, rejected response). In terms of depth, the responses are collected from 12 open-source (e.g., QwenVL) and closed-source (e.g., Gemini) VLMs to ensure diversity. The construction of preference data is fully automated, and the experimental results indicate that models trained with alignment techniques on the SPA-VL dataset exhibit substantial improvements in harmlessness and helpfulness while maintaining core capabilities. SPA-VL, as a large-scale, high-quality, and diverse dataset, represents a significant milestone in ensuring that VLMs achieve both harmlessness and helpfulness.", "keywords": [ "Alignment", "VLM", "LLM", "Dataset" ], "rating_list": [ 8, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "cp3aW7C5tD", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Han Li", "gender": "Female", "institution": "Huazhong Agricultural University", "country": "CN", "position": "Associate Professor" }, { "name": "Hong Chen", "gender": "unknown", "institution": "Huazhong Agricultural University", "country": "", "position": "Full Professor" }, { "name": "Tieliang Gong", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Wen Wen", "gender": "Female", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 159, "n_formula_1": 36, "n_ref_uni": 65, "n_ref": 134, "n_ref_all": 164, "n_fig": 14, "n_tab": 4, "L_tab": 2148, "n_element_tab": 92, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 405, "n_element_tab_1": 53, "formula_len_all": 21284, "formula_len_all_1": 2873, "len_all": 210553, "len_all_1": 60280, "len_abs": 1381, "len_title": 130, "len_sents": 50153, "len_sents_1": 25089, "n_sents": 491, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1396, "title": "Towards Generalization Bounds of GCNs for Adversarially Robust Node Classification", "abs": "Adversarially robust generalization of Graph Convolutional Networks (GCNs) has garnered significant attention in various security-sensitive application areas, driven by intrinsic adversarial vulnerability. Albeit remarkable empirical advancement, theoretical understanding of the generalization behavior of GCNs subjected to adversarial attacks remains elusive. To make progress on the mystery, we establish unified high-probability generalization bounds for GCNs in the context of node classification, by leveraging adversarial Transductive Rademacher Complexity (TRC) and developing a novel contraction technique on graph convolution. Our bounds capture the interaction between generalization error and adversarial perturbations, revealing the importance of key quantities in mitigating the negative effects of perturbations, such as low-dimensional feature projection, perturbation-dependent norm regularization, normalized graph matrix, proper number of network layers, etc. Furthermore, we provide TRC-based bounds of popular GCNs with $\\ell_r$-norm-additive perturbations for arbitrary $r\\geq 1$. A comparison of theoretical results demonstrates that specific network architectures (e.g., residual connection) can help alleviate the cumulative effect of perturbations during the forward propagation of deep GCNs. Experimental results on benchmark datasets validate our theoretical findings.", "keywords": [ "Generalization Analysis; Adversarial Learning; Graph Convolution Networks; Node Classification\uff1bNode Attacks" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "coq1hOntgI", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Chuang Gan", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Daniela Rus", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Tsun-Hsuan Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Wei Xiao", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" } ], "n_formula": 27, "n_formula_1": 9, "n_ref_uni": 21, "n_ref": 67, "n_ref_all": 129, "n_fig": 9, "n_tab": 8, "L_tab": 2777, "n_element_tab": 218, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 2094, "n_element_tab_1": 159, "formula_len_all": 2781, "formula_len_all_1": 1087, "len_all": 139330, "len_all_1": 62253, "len_abs": 1127, "len_title": 112, "len_sents": 42054, "len_sents_1": 26803, "n_sents": 367, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1133, "title": "ABNet: Attention BarrierNet for Safe and Scalable Robot Learning", "abs": "Safe learning is central to AI-enabled robots where a single failure may lead to catastrophic results. Barrier-based method is one of the dominant approaches for safe robot learning. However, this method is not scalable, hard to train, and tends to generate unstable signals under noisy inputs that are challenging to be deployed for robots. To address these challenges, we propose a novel Attention BarrierNet (ABNet) that is scalable to build larger foundational safe models in an incremental manner. Each head of BarrierNet in the ABNet could learn safe robot control policies from different features and focus on specific part of the observation. In this way, we do not need to one-shotly construct a large model for complex tasks, which significantly facilitates the training of the model while ensuring its stable output. Most importantly, we can still formally prove the safety guarantees of the ABNet. We demonstrate the strength of ABNet in 2D robot obstacle avoidance, safe robot manipulation, and vision-based end-to-end autonomous driving, with results showing much better robustness and guarantees over existing models.", "keywords": [ "Safe learning", "Robot learning", "Scalable learning" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "cojJ2s1e35", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Julian Jorge Andrade Guerreiro", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Kohei Hayashi", "gender": "Male", "institution": "Preferred Networks, Inc.", "country": "", "position": "Researcher" }, { "name": "Masanori Koyama", "gender": "unknown", "institution": "Preferred Networks, Inc.", "country": "", "position": "Researcher" } ], "n_formula": 44, "n_formula_1": 24, "n_ref_uni": 23, "n_ref": 48, "n_ref_all": 65, "n_fig": 6, "n_tab": 5, "L_tab": 307, "n_element_tab": 35, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 143, "n_element_tab_1": 8, "formula_len_all": 3029, "formula_len_all_1": 1575, "len_all": 134147, "len_all_1": 62355, "len_abs": 1079, "len_title": 124, "len_sents": 44854, "len_sents_1": 28927, "n_sents": 345, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1084, "title": "Inter-Environmental World Modeling for Continuous and Compositional Dynamics", "abs": "Various world model frameworks are being developed today based on autoregressive frameworks that rely on discrete representations of actions and observations, and these frameworks are succeeding in constructing interactive generative models for the target environment of interest. Meanwhile, humans demonstrate remarkable generalization abilities to combine experiences in multiple environments to mentally simulate and learn to control agents in diverse environments. Inspired by this human capability, we introduce World modeling through Lie Action (WLA), an unsupervised framework that learns continuous latent action representations to simulate across environments. \nWLA learns a control interface with high controllability and predictive ability by simultaneously modeling the dynamics of multiple environments using Lie group theory and object-centric autoencoder. On benchmark synthetic and real-world datasets, we demonstrate that WLA can be trained using only video frames and, with minimal or no action labels, can quickly adapt to new environments with novel action sets.", "keywords": [ "world model", "state space model", "object centric learning", "symmetry" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 2, 3, 3 ] }, { "paper_id": "coE6XbziUR", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Jiaming Han", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xiangyu Yue", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Yilei JIANG", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 21, "n_ref": 37, "n_ref_all": 54, "n_fig": 4, "n_tab": 6, "L_tab": 1705, "n_element_tab": 359, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1711, "n_element_tab_1": 331, "formula_len_all": 576, "formula_len_all_1": 472, "len_all": 101806, "len_all_1": 62199, "len_abs": 1257, "len_title": 148, "len_sents": 35171, "len_sents_1": 26913, "n_sents": 265, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 99, "L_abs": 1259, "title": "Mitigating Generative Privacy Risks of Diffusion Models via Mixed Self-Synthesized Data Fine-tuning", "abs": "Diffusion models (DMs) have demonstrated exceptional performance across various generative tasks, yet they also face significant security and privacy concerns, such as Membership Inference Attacks (MIAs), where adversaries attempt to determine whether specific images were part of the DM's training set. These threats present serious risks, particularly as pre-trained DMs are increasingly accessible online. To address these privacy concerns, we begin by investigating how fine-tuning DMs on a manipulated self-synthesized dataset affects their generative privacy risks, and have the following observations: (1) DMs fine-tuned solely on self-synthesized clean images are more vulnerable to privacy attacks (2) DMs fine-tuned on perturbed self-synthesized images become more robust against privacy attacks but exhibit degraded image generation quality.\nBased on the observations, we propose MixSyn, a simple and effective framework designed to mitigate privacy risks by fine-tuning DMs on a mixed self-synthesized dataset, which is a mixture of clean and perturbed synthetic images. Extensive experimental results demonstrate that our method significantly mitigates the generative privacy risks of DMs while preserving their original image generation quality.", "keywords": [ "Diffusion Models", "AI privacy", "Membership Inference Attack", "AI safety" ], "rating_list": [ 3, 6, 3, 5, 3 ], "soundness_list": [ 2, 3, 1, 3, 3 ], "presentation_list": [ 2, 3, 1, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 5, 3, 3, 3, 4 ] }, { "paper_id": "cnecLUNs6w", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "David Krueger", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Assistant Professor" }, { "name": "Johannes Von Oswald", "gender": "Male", "institution": "Research, Google", "country": "CH", "position": "Researcher" }, { "name": "Louis Kirsch", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Spencer Frei", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Usman Anwar", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 12, "n_ref_uni": 45, "n_ref": 115, "n_ref_all": 143, "n_fig": 12, "n_tab": 2, "L_tab": 138, "n_element_tab": 11, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 1761, "n_element_tab_1": 21, "formula_len_all": 1799, "formula_len_all_1": 793, "len_all": 153931, "len_all_1": 66905, "len_abs": 1807, "len_title": 132, "len_sents": 50743, "len_sents_1": 31129, "n_sents": 356, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1819, "title": "Adversarial Robustness of In-Context Learning in Transformers for Linear Regression", "abs": "Transformers have demonstrated remarkable in-context learning capabilities across various domains, including statistical learning tasks. While previous work has shown that transformers can implement common learning algorithms, the adversarial robustness of these learned algorithms remains unexplored. This work investigates the vulnerability of in-context learning in transformers to _hijacking attacks_ focusing on the setting of linear regression tasks. Hijacking attacks are prompt-manipulation attacks in which the adversary's goal is to manipulate the prompt to force the transformer to generate a specific output. We first prove that single-layer linear transformers, known to implement gradient descent in-context, are non-robust and can be manipulated to output arbitrary predictions by perturbing\na single example in the in-context training set. While our experiments show these attacks succeed on linear transformers, we find they do not transfer to more complex transformers with GPT-2 architectures. Nonetheless, we show that these transformers can be hijacked using gradient-based adversarial attacks. We then demonstrate that adversarial training enhances transformers' robustness against hijacking attacks, even when just applied during finetuning. Additionally, we find that in some settings, adversarial training against a weaker attack model can lead to robustness to a stronger attack model. Lastly, we investigate the transferability of hijacking attacks across transformers of varying scales and initialization seeds, as well as between transformers and ordinary least squares (OLS). We find that while attacks transfer effectively between small-scale transformers, they show poor transferability in other scenarios (small-to-large scale, large-to-large scale, and between transformers and OLS).", "keywords": [ "in-context learning", "transformers", "hijacking attacks", "linear regression", "linear transformers", "transfer of adversarial attacks" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "cnLNpIRPuF", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Alexander Ashcroft", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "PhD student" }, { "name": "Hmrishav Bandyopadhyay", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "PhD student" }, { "name": "Yi-Zhe Song", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Professor" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 49, "n_ref": 70, "n_ref_all": 83, "n_fig": 8, "n_tab": 1, "L_tab": 141, "n_element_tab": 17, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1140, "n_element_tab_1": 28, "formula_len_all": 697, "formula_len_all_1": 697, "len_all": 114405, "len_all_1": 55391, "len_abs": 1464, "len_title": 58, "len_sents": 28549, "len_sents_1": 25490, "n_sents": 200, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 10, "L_abs": 1422, "title": "Worldcraft", "abs": "We present Worldcraft, a hybrid implicit method for generating vast, interactive 3D worlds at unprecedented scale and speed by modeling them as exchangeable sequences of latent 3D objects. In contrast to existing methods that produce limited scenes, Worldcraft's novel approach constructs expansive environments comprising thousands of elements, extending to over a million objects in seconds, on a single GPU. The resulting created worlds\n are defined in terms of possessing certain essential properties: Object Individuality, Collective Semantics, and Expandability. To achieve this with both speed and scale, we conceptualize world generation as a set generation problem, introducing three key technical innovations: (i) Hierarchical and Exchangeable Sequence Modeling ensures Object Individuality while capturing Collective Semantics; (ii) Hybrid Implicit Generation Method enables rapid creation of vast worlds, supporting both Scale and Expandability; and (iii) Multi-level Indexing Functions allow efficient manipulation across scales, reinforcing Collective Semantics and enabling on-demand generation for Speed and Expandability. We demonstrate Worldcraft's capabilities using Minecraft as a test-bed, generating complex, interactive environments that users can explore. However, this approach is applicable to any suitable platform, potentially revolutionizing various applications in 3D environment generation.", "keywords": [ "Exchangeability", "Hierachy", "3D world Generation" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "cnKhHxN3xj", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Ilia Markov", "gender": "unknown", "institution": "Institute of Science and Technology Austria", "country": "", "position": "PhD student" }, { "name": "Linghao Kong", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Nir N Shavit", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Shashata Sawmya", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Dan Alistarh", "gender": "Male", "institution": "Institute of Science and Technology", "country": "AT", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 42, "n_ref": 78, "n_ref_all": 158, "n_fig": 37, "n_tab": 6, "L_tab": 1377, "n_element_tab": 204, "n_fig_1": 38, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 237, "formula_len_all_1": 218, "len_all": 177451, "len_all_1": 69389, "len_abs": 1236, "len_title": 106, "len_sents": 59703, "len_sents_1": 34253, "n_sents": 445, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 58, "L_abs": 1282, "title": "Wasserstein Distances, Neuronal Entanglement, and Sparsity", "abs": "Disentangling polysemantic neurons is at the core of many current approaches to interpretability of large language models. Here we attempt to study how disentanglement can be used to understand performance, particularly under weight sparsity, a leading post-training optimization technique. We suggest a novel measure for estimating neuronal entanglement: the Wasserstein distance of a neuron's output distribution to a Gaussian. Moreover, we show the existence of a small number of highly entangled \"Wasserstein Neurons\" in each linear layer of an LLM, characterized by their highly non-Gaussian output distributions, their role in mapping similar inputs to dissimilar outputs, and their significant impact on model accuracy. To study these phenomena, we propose a new experimental framework for disentangling polysemantic neurons. Our framework separates each layer's inputs to create a mixture of experts where each neuron's output is computed by a mixture of neurons of lower Wasserstein distance, each better at maintaining accuracy when sparsified without retraining. We provide strong evidence that this is because the mixture of sparse experts is effectively disentangling the input-output relationship of individual neurons, in particular the difficult Wasserstein neurons.", "keywords": [ "Polysemanticity", "Disentanglement", "Wasserstein Distance", "Sparsity", "Large Language Models" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "cmfyMV45XO", "primary_area": "learning on time series and dynamical systems", "n_author": 7, "author_info_list": [ { "name": "Jindou Jia", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Kexin Guo", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Associate Professor" }, { "name": "Lei Guo", "gender": "Male", "institution": "Southeast University", "country": "", "position": "PhD student" }, { "name": "Meng Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Xiang Yu", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Zihan Yang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Jianfei Yang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 36, "n_formula_1": 13, "n_ref_uni": 53, "n_ref": 95, "n_ref_all": 156, "n_fig": 20, "n_tab": 1, "L_tab": 157, "n_element_tab": 2, "n_fig_1": 17, "n_tab_1": 1, "L_tab_1": 109, "n_element_tab_1": 5, "formula_len_all": 2653, "formula_len_all_1": 850, "len_all": 174214, "len_all_1": 59652, "len_abs": 1181, "len_title": 97, "len_sents": 51486, "len_sents_1": 26037, "n_sents": 445, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 49, "L_abs": 1185, "title": "Feedback Favors the Generalization of Neural ODEs", "abs": "The well-known generalization problem hinders the application of artificial neural networks in continuous-time prediction tasks with varying latent dynamics. In sharp contrast, biological systems can neatly adapt to evolving environments benefiting from real-time feedback mechanisms. Inspired by the feedback philosophy, we present feedback neural networks, showing that a feedback loop can flexibly correct the learned latent dynamics of neural ordinary differential equations (neural ODEs), leading to a prominent generalization improvement. The feedback neural network is a novel two-DOF neural network, which possesses robust performance in unseen scenarios with no loss of accuracy performance on previous tasks. A linear feedback form is presented to correct the learned latent dynamics firstly, with a convergence guarantee. Then, domain randomization is utilized to learn a nonlinear neural feedback form. Finally, extensive tests including trajectory prediction of a real irregular object and model predictive control of a quadrotor with various uncertainties, are implemented, indicating significant improvements over state-of-the-art model-based and learning-based methods.", "keywords": [ "Neural ODEs", "feedback", "generalization", "learning dynamical systems", "model predictive control" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "cmYScmfu4Q", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Lei Ying", "gender": "Male", "institution": "University of Michigan, Ann Arbor", "country": "US", "position": "Professor" }, { "name": "Qining Zhang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 30, "n_ref_uni": 39, "n_ref": 84, "n_ref_all": 88, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 748, "n_element_tab_1": 8, "formula_len_all": 2003, "formula_len_all_1": 2026, "len_all": 155468, "len_all_1": 60987, "len_abs": 1683, "len_title": 149, "len_sents": 33796, "len_sents_1": 29151, "n_sents": 215, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1531, "title": "Zeroth-Order Policy Gradient for Reinforcement Learning from Human Feedback without Reward Inference", "abs": "Reward inference (learning a reward model from human preferences) is a critical intermediate step in Reinforcement Learning from Human Feedback (RLHF) for fine-tuning Large Language Models (LLMs) such as ChatGPT. In practice, reward inference faces several fundamental challenges, including double problem misspecification, reward model evaluation without ground truth, distribution shift, and overfitting in joint reward model and policy training. \nAn alternative approach that avoids these pitfalls is direct policy optimization without reward inference, such as Direct Preference Optimization (DPO), which provides a much simpler pipeline and has shown empirical success in LLMs. However, DPO utilizes the closed-form expression between the optimal policy and the reward function, which only works under the bandit setting or deterministic MDPs. This paper develops two RLHF algorithms without reward inference, which work for general RL problems beyond bandits and deterministic MDPs, and general preference models beyond the Bradely-Terry model. The key idea is to estimate the local value function difference from human preferences and then approximate the policy gradient with a zeroth-order gradient approximator. For both algorithms, we establish rates of convergence in terms of the number of policy gradient iterations, as well as the number of trajectory samples and human preference queries per iteration. Our results show there exist provably efficient methods to solve general RLHF problems without reward inference.", "keywords": [ "reinforcement learning theory", "human feedback", "zeroth-order optimization" ], "rating_list": [ 8, 8, 3, 8 ], "soundness_list": [ 3, 4, 2, 4 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "cmXWYolrlo", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Antonio Orvieto", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen, Max Planck Institute for Intelligent Systems, T\u00fcbingen AI Center, T\u00fcbingen, Germany", "country": "CH", "position": "Principal Researcher" }, { "name": "Sajad Movahedi", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen, Max Planck Institute for Intelligent Systems, T\u00fcbingen AI Center", "country": "DE", "position": "PhD student" }, { "name": "Seyed-Mohsen Moosavi-Dezfooli", "gender": "Male", "institution": "Apple Inc.", "country": "CH", "position": "Researcher" } ], "n_formula": 92, "n_formula_1": 14, "n_ref_uni": 30, "n_ref": 52, "n_ref_all": 120, "n_fig": 15, "n_tab": 1, "L_tab": 290, "n_element_tab": 26, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8184, "formula_len_all_1": 948, "len_all": 248137, "len_all_1": 76195, "len_abs": 2734, "len_title": 126, "len_sents": 79814, "len_sents_1": 37661, "n_sents": 622, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 78, "L_abs": 1343, "title": "Geometric Inductive Biases of Deep Networks: The Role of Data and Architecture", "abs": "In this paper, we propose the *geometric invariance hypothesis (GIH)*, which argues that when training a neural network, the input space curvature remains invariant under transformation in certain directions determined by its architecture. Starting with a simple non-linear binary classification problem residing on a plane in a high dimensional space, we observe that while an MLP can generalize on this problem regardless of the orientation of the plane, this is not the case for a ResNet. Motivated by this example, we define two maps that provide a compact *architecture-dependent* summary of the input space geometry of a neural network and its evolution during training, which we dub the **average geometry** and **average geometry evolution**, respectively. \nBy investigating average geometry evolution at initialization, we discover that the geometry of a neural network evolves according to the projection of data covariance onto average geometry. As a result, in cases where the average geometry is low-rank (such as in a ResNet), the geometry only changes in a subset of the input space. This causes an architecture-dependent invariance property in input space curvature, which we dub GIH. Finally, we present extensive experimental results to observe the consequences of GIH and how it relates to generalization in neural networks.", "keywords": [ "Optimization", "Deep Neural Networks", "Deep Learning", "Deep Learning Theory", "Machine Learning Theory" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 4, 3, 2 ], "presentation_list": [ 2, 4, 4, 2 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 3, 2, 3, 4 ] }, { "paper_id": "cklg91aPGk", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Mingjie Li", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "", "position": "Postdoc" }, { "name": "Xiaojun Guo", "gender": "Female", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yifei Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yisen Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 83, "n_ref": 227, "n_ref_all": 265, "n_fig": 7, "n_tab": 22, "L_tab": 14586, "n_element_tab": 967, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 4215, "n_element_tab_1": 301, "formula_len_all": 615, "formula_len_all_1": 391, "len_all": 261060, "len_all_1": 76384, "len_abs": 1020, "len_title": 107, "len_sents": 71677, "len_sents_1": 31550, "n_sents": 589, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1024, "title": "Propagation Alone is Enough for Graph Contrastive Learning", "abs": "Graph contrastive learning has recently gained substantial attention, leading to the development of various methodologies. In this work, we reveal that a simple training-free propagation method PROP achieves competitive results over dedicatedly designed GCL methods across a diverse set of node classification benchmarks. We elucidate the underlying rationale for PROP\u2019s effectiveness by drawing connections between the propagation operator and established unsupervised learning algorithms. To investigate the reasons for the suboptimal performance of GCL, we decouple the propagation and transformation phases of graph neural networks. Our findings indicate that existing GCL methods inadequately learns effective transformation weights while exhibiting potential for solid propagation learning. In light of these insights, we enhance PROP with learnable propagation, introducing a novel GCL method termed PROPGCL. The effectiveness of PROPGCL is demonstrated through comprehensive evaluations on node classification tasks.", "keywords": [ "graph contrastive learning", "polynomial GNNs" ], "rating_list": [ 5, 3, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 5, 4 ] }, { "paper_id": "ckicHjoTgf", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 3, "author_info_list": [ { "name": "Chao Guo", "gender": "Male", "institution": "Waseda University", "country": "JP", "position": "PhD student" }, { "name": "Jinjun Xiong", "gender": "unknown", "institution": "State University of New York at Buffalo", "country": "US", "position": "Professor" }, { "name": "Youhua Shi", "gender": "unknown", "institution": "Waseda University", "country": "JP", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 54, "n_ref_all": 84, "n_fig": 9, "n_tab": 6, "L_tab": 1476, "n_element_tab": 228, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 774, "n_element_tab_1": 107, "formula_len_all": 206, "formula_len_all_1": 142, "len_all": 122796, "len_all_1": 70140, "len_abs": 1897, "len_title": 128, "len_sents": 41110, "len_sents_1": 34443, "n_sents": 266, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1902, "title": "A Novel Security Threat Model for Automated AI Accelerator Generation Platforms", "abs": "In recent years, the design of Artificial Intelligence (AI) accelerators has gradually shifted from focusing solely on standalone accelerator hardware to considering the entire system, giving rise to a new AI accelerator design paradigm that emphasizes full-stack integration. Systems designed based on this paradigm offer a user-friendly, end-to-end solution for deploying pre-trained models. While previous studies have identified vulnerabilities in individual hardware components or models, the security of this paradigm has not yet been thoroughly evaluated. This work, from an attacker's perspective, proposes a threat model based on this paradigm and reveals the potential security vulnerabilities of systems by embedding malicious code in the design flow, highlighting the necessity for protection to address this security gap. In exploration and generation, maliciously leverage the exploration unit to identify sensitive parameters in the model's intermediate layers and insert hardware Trojan (HT) into the accelerator. In execution, malicious information is concealed within the control instructions, triggering the HT. Experimental results demonstrate that the proposed method, which manipulates sensitive parameters in a few selected kernels across the middle convolutional layers, successfully misclassifies input images into specified categories with high misclassification rates across various models: 97.3% in YOLOv8 by modifying only three parameters per layer in three layers, 99.2% in ResNet-18 by altering four parameters per layer in three layers and 98.1% for VGG-16 by changing seven parameters per layer in four layers. Additionally, the area overhead introduced by the proposed HT occupies no more than 0.34% of the total design while maintaining near-original performance as in uncompromised designs, which clearly illustrates the concealment of the proposed security threat.", "keywords": [ "AI accelerator generation platforms", "Design Space Exploration (DSE)", "Hardware Trojan (HT)", "Security threat model" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "ckabXglfiT", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Jianwei Yin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Shuo Shi", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Tao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tianyu Du", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xinyi Shang", "gender": "Female", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Xuhong Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "peng sun", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" } ], "n_formula": 42, "n_formula_1": 13, "n_ref_uni": 48, "n_ref": 129, "n_ref_all": 159, "n_fig": 9, "n_tab": 4, "L_tab": 3354, "n_element_tab": 229, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 2343, "n_element_tab_1": 152, "formula_len_all": 2215, "formula_len_all_1": 758, "len_all": 164660, "len_all_1": 71013, "len_abs": 1095, "len_title": 131, "len_sents": 46159, "len_sents_1": 30901, "n_sents": 367, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1166, "title": "Privacy as a Free Lunch: Crafting Initial Distilled Datasets through the Kaleidoscope", "abs": "The advancement of deep learning necessitates stringent data privacy guarantees.\nDataset distillation has shown potential in preserving differential privacy while maintaining training efficiency.\nThis study first identifies that data generated by state-of-the-art dataset distillation methods strongly resembles to real data, indicating severe privacy leakage.\nWe define this phenomenon as explicit privacy leakage.\nWe theoretically analyze that although distilled datasets can ensure differential privacy to some extent, a high \\IPC can weaken both differential privacy and explicit privacy.\nFurthermore, we reveal that the primary source of privacy leakage in distilled data stems from the common approach of initializing distilled images as real data.\nTo address this, we propose a plug-and-play module, Kaleidoscopic Transformation (KT), designed to introduce enhanced strong perturbations to the selected real data during the initialization phase.\nExtensive experiments demonstrate that our method ensures both differential privacy and explicit privacy, while preserving the generalization performance of the distilled data.\nOur code will be publicly available.", "keywords": [ "Dataset Distillation" ], "rating_list": [ 5, 1, 5, 8 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 1, 1, 2, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "cjpTu0Op5t", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Wentao Mo", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yang Liu", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuxin Peng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 11, "n_ref": 51, "n_ref_all": 90, "n_fig": 17, "n_tab": 11, "L_tab": 2716, "n_element_tab": 424, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1160, "n_element_tab_1": 209, "formula_len_all": 202, "formula_len_all_1": 202, "len_all": 152090, "len_all_1": 66395, "len_abs": 2390, "len_title": 137, "len_sents": 61280, "len_sents_1": 31108, "n_sents": 428, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 2856, "title": "VISCON: Identifying and Benchmarking Vision Hallucination for Large Vision-Language Model", "abs": "Large Vision-Language Models (LVLMs) have demonstrated exceptional capabilities in a variety of vision-language tasks, but suffer from \"vision hallucinations\" - a tendency generating text inconsistent with the image. This issue hampers their practical use in real-world applications.\n To effectively evaluate and detect these hallucinations, we introduce VISCON (VISual Concept cONsistency), a benchmark framework comprising a benchmark image dataset and quantitative evaluation pipelines to assess vision hallucinations in LVLMs. VISCON extends beyond previous hallucination metrics by offering: a) diverse image styles across multiple visual domains, b) evaluation of a broader range of visual concepts, including objects, attributes, and relationships, and c) high annotation density from detailed scene-graph annotations to reduce false negatives. These improvements enable comprehensive analysis of hallucinations related to both domain shifts and concept types and offer more accurate hallucination evaluation. \nTo detect vision hallucinations, we propose two innovative evaluation pipelines within VISCON: an Earth Mover's Distance (EMD)-based pipeline and an \"Evaluate-By-Edit\" pipeline. The EMD-based pipeline measures the distributional similarity between the reference visual concepts and those mentioned by LVLMs, robust against vocabulary shifts between annotations and natural language responses. The Through extensive experiments on six leading LVLMs, VISCON reveals crucial insights into the nature of vision hallucinations. Our findings indicate that factors such as image domain shifts, complexity of visual concepts and model response length significantly influence the occurrence of hallucinations in LVLM responses. Additionally, human evaluations confirm that VISCON aligns with human preferences better than established hallucination metrics. \"Evaluate-By-Edit\" focuses on the edit distance between the original LVLM response and a hallucination-reduced version revised according to the rich visual concept annotations, providing an interpretable analysis of hallucinated content. Importantly, our method directly evaluates captioning responses, unlike previous metrics that query the existence of individual visual concepts. This approach is more challenging, as it requires models to handle multiple concepts simultaneously, providing better discrimination of LVLM performance.\nThrough extensive experiments on six leading LVLMs, VISCON reveals crucial insights into the nature of vision hallucinations. Our findings indicate that factors such as image domain shifts, complexity of visual concepts and model response length significantly influence the occurrence of hallucinations in LVLM responses. Additionally, human evaluations confirm that VISCON aligns with human preferences better than established hallucination metrics.", "keywords": [ "hallucination", "vision hallucination", "large vision-language model", "large language model" ], "rating_list": [ 5, 5, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 4, 2, 2, 3, 1 ], "contribution_list": [ 3, 3, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "cjlPAgNifc", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Qi Zhang", "gender": "Male", "institution": "Fudan University", "country": "", "position": "Full Professor" }, { "name": "Tao Gui", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Assistant Professor" }, { "name": "Wanxu Zhao", "gender": "Male", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Wei He", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Xiaoran Fan", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Xuanjing Huang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yiwen Ding", "gender": "Female", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Zhiheng Xi", "gender": "unknown", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Zifei Shan", "gender": "Male", "institution": "WeChat, Tencent", "country": "", "position": "Senior Applied Scientist" } ], "n_formula": 4, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 92, "n_ref_all": 134, "n_fig": 18, "n_tab": 10, "L_tab": 3846, "n_element_tab": 337, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 4833, "n_element_tab_1": 428, "formula_len_all": 660, "formula_len_all_1": 0, "len_all": 207503, "len_all_1": 68114, "len_abs": 2083, "len_title": 105, "len_sents": 57043, "len_sents_1": 27980, "n_sents": 504, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1370, "title": "Distill Visual Chart Reasoning Ability from LLMs to MLLMs", "abs": "Solving complex chart Q&A tasks requires advanced visual reasoning abilities in multimodal large language models (MLLMs). Recent studies highlight that these abilities consist of two main parts: recognizing key information from visual inputs and conducting reasoning over it. Thus, a promising approach to enhance MLLMs is to construct relevant training data focusing on the two aspects. However, collecting and annotating complex charts and questions is costly and time-consuming, and ensuring the quality of annotated answers remains a challenge. In this paper, we propose Code-as-Intermediary Translation (CIT), a cost-effective, efficient and easily scalable data synthesis method for distilling visual reasoning abilities from LLMs to MLLMs. The code serves as an intermediary that translates visual chart representations into textual representations, enabling LLMs to understand cross-modal information. Specifically, we employ text-based synthesizing techniques to construct chart-plotting code and produce ReachQA, a dataset containing 3k reasoning-intensive charts and 20k Q&A pairs to enhance both recognition and reasoning abilities. Experiments show that when fine-tuned with our data, models not only perform well on chart-related benchmarks, but also demonstrate improved multimodal reasoning abilities on general mathematical benchmarks such as MathVista.", "keywords": [ "multimodal large language models", "synthetic data", "chart question answering" ], "rating_list": [ 8, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "cjJqU40nYS", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Dong-ming Yan", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Gaofeng Meng", "gender": "Male", "institution": "Institute of automation", "country": "CN", "position": "Full Professor" }, { "name": "Mingyang Zhao", "gender": "Male", "institution": "Centre for Artificial Intelligence and Robotics Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "HK", "position": "Lecturer" } ], "n_formula": 23, "n_formula_1": 12, "n_ref_uni": 38, "n_ref": 57, "n_ref_all": 113, "n_fig": 20, "n_tab": 6, "L_tab": 2131, "n_element_tab": 260, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 1122, "n_element_tab_1": 153, "formula_len_all": 1571, "formula_len_all_1": 1000, "len_all": 159289, "len_all_1": 61738, "len_abs": 1809, "len_title": 148, "len_sents": 47174, "len_sents_1": 26847, "n_sents": 339, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 98, "L_abs": 1817, "title": "Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy", "abs": "Non-rigid alignment of point clouds is crucial for scene understanding, reconstruction, and various computer vision and robotics tasks. Recent advancements in implicit deformation networks for non-rigid registration have significantly reduced the reliance on large amounts of annotated training data. However, existing state-of-the-art methods still face challenges in handling occlusion scenarios. To address this issue, this paper introduces an innovative unsupervised method called Occlusion-Aware Registration (OAR) for non-rigidly aligning point clouds. The key innovation of our method lies in the utilization of the adaptive correntropy function as a localized similarity measure, enabling us to treat individual points distinctly. In contrast to previous approaches that solely minimize overall deviations between two shapes, we combine unsupervised implicit neural representations with the maximum correntropy criterion to optimize the deformation of unoccluded regions. This effectively avoids collapsed, tearing, and other physically implausible results. Moreover, we present a theoretical analysis and establish the relationship between the maximum correntropy criterion and the commonly used Chamfer distance, highlighting that the correntropy-induced metric can be served as a more universal measure for point cloud analysis. Additionally, we introduce\nlocally linear reconstruction to ensure that regions lacking correspondences between shapes still undergo physically natural deformations. Our method achieves superior or competitive performance compared to existing approaches, particularly when dealing with occluded geometries. We also demonstrate the versatility of our method in challenging tasks such as large deformations, shape interpolation, and shape completion under occlusion disturbances.", "keywords": [ "Unsupervised Deformation", "Neural Implicit Representations", "Correntropy", "Locally Linear Reconstruction" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "cit3SNnZ6Q", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Furu Wei", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Partner Research Manager" }, { "name": "Hongyu Wang", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Ruiping Wang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Shuming Ma", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 20, "n_ref_uni": 31, "n_ref": 62, "n_ref_all": 85, "n_fig": 12, "n_tab": 6, "L_tab": 1644, "n_element_tab": 261, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 487, "n_element_tab_1": 86, "formula_len_all": 1014, "formula_len_all_1": 926, "len_all": 125825, "len_all_1": 56382, "len_abs": 1097, "len_title": 115, "len_sents": 30899, "len_sents_1": 24754, "n_sents": 260, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1082, "title": "Q-Sparse: All Large Language Models can be Fully Sparsely-Activated", "abs": "We introduce, Q-Sparse, a simple yet effective approach to training sparsely-activated large language models (LLMs). Q-Sparse enables full sparsity of activations in LLMs which can bring significant efficiency gains in inference. This is achieved by applying top-K sparsification to the activations and the straight-through-estimator to the training. We also introduce Block Q-Sparse for batch training and inference. The key results from this work are, (1) Q-Sparse can achieve results comparable to those of baseline LLMs while being much more efficient at inference time; (2) We present an inference-optimal scaling law for sparsely-activated LLMs; (3) Q-Sparse is effective in different settings, including training-from-scratch, continue-training of off-the-shelf LLMs, and finetuning; (4) Q-Sparse works for both full-precision and 1-bit LLMs (e.g., BitNet b1.58). Particularly, the synergy of BitNet b1.58 and Q-Sparse (can be equipped with MoE) provides the cornerstone and a clear path to revolutionize the efficiency, including cost and energy consumption, of future LLMs.", "keywords": [ "Activation sparsity", "efficiency", "large language models" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "cif0JVXJ3b", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Emmanuel Chemla", "gender": "unknown", "institution": "CNRS", "country": "", "position": "Full Professor" }, { "name": "Nicolas Guerin", "gender": "Male", "institution": "Ecole Normale Sup\u00e9rieure \u2013 PSL", "country": "FR", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 55, "n_ref": 111, "n_ref_all": 129, "n_fig": 10, "n_tab": 2, "L_tab": 250, "n_element_tab": 45, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 4, "n_element_tab_1": 2, "formula_len_all": 25, "formula_len_all_1": 0, "len_all": 182919, "len_all_1": 62117, "len_abs": 1501, "len_title": 114, "len_sents": 40780, "len_sents_1": 32082, "n_sents": 290, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1500, "title": "Qualifying Knowledge and Knowledge Sharing in Multilingual Models", "abs": "Pre-trained language models (PLMs) have demonstrated a remarkable ability to encode factual knowledge. However, the mechanisms underlying how this knowledge is stored and retrieved remain poorly understood, with important implications for AI interpretability and safety. In this paper, we disentangle the multifaceted nature of knowledge: successfully completing a knowledge retrieval task (e.g., \u201cThe capital of France is __\u201d) involves mastering underlying concepts (e.g. France, Paris), relationships between these concepts (e.g. capital of), the structure of prompts, including the language of the query. We propose to disentangle these distinct aspects of knowledge and apply this typology to offer a critical view of neuron-level knowledge attribution techniques. For concreteness, we focus on Dai et al.'s (2022) Knowledge Neurons (KNs) across multiple PLMs, testing 10 natural languages and unnatural languages (e.g. Autoprompt).\nOur key contributions are twofold: (i) we show that KNs come in different flavors, some indeed encoding entity level concepts, some having a much less transparent, more polysemantic role , and (ii) we uncover an unprecedented overlap in KNs across up to all of the 10 languages we tested, pointing to the existence of a partially unified, language-agnostic retrieval system. To do so, we introduce and release the mParaRel dataset, an extension of ParaRel, featuring prompts and paraphrases for cloze-style knowledge retrieval tasks in parallel over 10 languages.", "keywords": [ "Knowledge Retrieval", "Pretrained Language Model", "LLM Interpretability", "Multilingual Models", "Cross-linguistic Dataset" ], "rating_list": [ 5, 3, 5, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 4 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "ciSW6Jczvo", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Hong Mei", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Wenwu Zhu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xin Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yang Yao", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yijian Qin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Zeyang Zhang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 20, "n_ref": 30, "n_ref_all": 42, "n_fig": 3, "n_tab": 8, "L_tab": 2213, "n_element_tab": 287, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1680, "n_element_tab_1": 211, "formula_len_all": 745, "formula_len_all_1": 745, "len_all": 123071, "len_all_1": 63993, "len_abs": 1974, "len_title": 135, "len_sents": 35945, "len_sents_1": 28422, "n_sents": 272, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1981, "title": "Text-to-graph Generation with Conditional Diffusion Models Guided by Graph-aligned LLMs", "abs": "Text-to-graph generation, aiming for controlled graph generation based on natural language instructions, holds significant application potentials in real-world scenarios such as drug discoveries. However, existing generative models fail to achieve text-to-graph generation in the following two aspects: i) language model-based generative models struggle with generating complex graph structures, and ii) graph-based generative models mainly focus on unconditional graph generation or conditional generation with simple conditions, falling short in understanding as well as following human instructions. In this paper, we tackle the text-to-graph generation problem by employing graph diffusion models with guidance from large language models (LLMs) for the first time, to the best of our knowledge. The problem is highly non-trivial with the following challenges: 1) How to align LLMs for understanding the irregular graph structures and the graph properties hidden in human instructions, 2) How to align graph diffusion models for following natural language instructions in order to generate graphs with expected relational semantics from human. To address these challenges, we propose a novel LLM-aligned Graph Diffusion Model (LLM-GDM), which is able to generate graphs based on natural language instructions. In particular, we first propose the self-supervised text-graph alignment to empower LLMs with the ability to accurately understand graph structures and properties by finetuning LLMs with several specially designed alignment tasks involving various graph components such as nodes, edges, and subgraphs. Then, we propose a structure-aware cross-attention mechanism guiding the diffusion model to follow human instructions through inherently capturing the relational semantics among texts and structures. Extensive experiments on both synthetic and real-world molecular datasets demonstrate the effectiveness of our proposed LLM-GDM model over existing baseline methods.", "keywords": [ "large language model", "text-to-graph", "diffusion model", "graph generation" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "chwxOOwzFR", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Alaa Maalouf", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Alex Quach", "gender": "Not Specified", "institution": "Liquid AI", "country": "US", "position": "Researcher" }, { "name": "Daniela Rus", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Makram Chahine", "gender": "Not Specified", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Tsun-Hsuan Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 44, "n_ref": 63, "n_ref_all": 76, "n_fig": 10, "n_tab": 3, "L_tab": 483, "n_element_tab": 49, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 507, "formula_len_all_1": 379, "len_all": 161587, "len_all_1": 65643, "len_abs": 1457, "len_title": 121, "len_sents": 46539, "len_sents_1": 31241, "n_sents": 345, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1461, "title": "Flex: End-to-End Text-Instructed Visual Navigation with Foundation Models", "abs": "End-to-end learning directly maps sensory inputs to actions, creating highly integrated and efficient policies for complex robotics tasks. However, such models are tricky to efficiently train and often struggle to generalize beyond their training scenarios, limiting adaptability to new environments, tasks, and concepts. In this work, we investigate the minimal data requirements and architectural adaptations necessary to achieve robust closed-loop performance with vision-based control policies under unseen text instructions and visual distribution shifts.\nTo this end, we design datasets with various levels of data representation richness, refine feature extraction protocols by leveraging multi-modal foundation model encoders, and assess the suitability of different policy network heads. Our findings are synthesized in Flex (Fly-lexically), a framework that uses pre-trained Vision Language Models (VLMs) as frozen patch-wise feature extractors, generating spatially aware embeddings that integrate semantic and visual information. These rich features form the basis for training highly robust downstream policies capable of generalizing across platforms, environments, and text-specified tasks.\nWe demonstrate the effectiveness of this approach on quadrotor fly-to-target tasks, where agents trained via behavior cloning on a small simulated dataset successfully generalize to real-world scenes, handling diverse novel goals and command formulations.", "keywords": [ "Robot learning", "Foundation Models", "Visual Navigation" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "cho9iE9POr", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Arnaud Delaunoy", "gender": "Male", "institution": "Universit\u00e9 de Li\u00e8ge", "country": "", "position": "PhD student" }, { "name": "Gilles Louppe", "gender": "Male", "institution": "University of Li\u00e8ge", "country": "BE", "position": "Associate Professor" }, { "name": "Maxence de la Brassinne Bonardeaux", "gender": "Male", "institution": "University of Li\u00e8ge", "country": "", "position": "MS student" }, { "name": "Siddharth Mishra-Sharma", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" } ], "n_formula": 25, "n_formula_1": 18, "n_ref_uni": 40, "n_ref": 62, "n_ref_all": 76, "n_fig": 8, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1541, "formula_len_all_1": 1217, "len_all": 119522, "len_all_1": 56707, "len_abs": 1193, "len_title": 115, "len_sents": 37679, "len_sents_1": 26986, "n_sents": 282, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1175, "title": "Low-Budget Simulation-Based Inference with Bayesian Neural Networks", "abs": "Simulation-based inference methods have been shown to be inaccurate in the data-poor regime, when training simulations are limited or expensive.\nUnder these circumstances, the inference network is particularly prone to overfitting, and using it without accounting for the computational uncertainty arising from the lack of identifiability of the network weights can lead to unreliable results.\nTo address this issue, we propose using Bayesian neural networks in low-budget simulation-based inference, thereby explicitly accounting for the computational uncertainty of the posterior approximation.\nWe design a family of Bayesian neural network priors that are tailored for inference and show that they lead to well-calibrated posteriors on tested benchmarks, even when as few as $O(10)$ simulations are available.\nThis opens up the possibility of performing reliable simulation-based inference using very expensive simulators, as we demonstrate on a problem from the field of cosmology where single simulations are computationally expensive. We show that Bayesian neural networks produce informative and well-calibrated posterior estimates with only a few hundred simulations.", "keywords": [ "simulation-based inference", "approximate Bayesian inference" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 2, 5, 4, 4 ] }, { "paper_id": "chiDvQc1F6", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jeongeun Park", "gender": "Female", "institution": "Korea University", "country": "KR", "position": "PhD student" }, { "name": "Sangdoo Yun", "gender": "Male", "institution": "NAVER", "country": "KR", "position": "Research Scientist" }, { "name": "Sungjoon Choi", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 31, "n_ref": 76, "n_ref_all": 87, "n_fig": 5, "n_tab": 4, "L_tab": 1005, "n_element_tab": 105, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 937, "n_element_tab_1": 91, "formula_len_all": 371, "formula_len_all_1": 369, "len_all": 132722, "len_all_1": 59884, "len_abs": 1673, "len_title": 114, "len_sents": 35274, "len_sents_1": 26399, "n_sents": 258, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1677, "title": "Versatile Motion-Language Models for Multi-turn Interactive Agents", "abs": "Recent advancements in large language models (LLMs) have greatly enhanced their ability to generate natural and contextually relevant text, making AI interactions more human-like. However, generating and understanding interactive human-like motion, where two individuals engage in coordinated movements, remains a challenge due to the complexity of modeling these coordinated interactions. Furthermore, a versatile model is required to handle diverse interactive scenarios, such as chat systems that follow user instructions or adapt to their assigned role while adjusting interaction dynamics. To tackle this problem, we introduce VIM, short for the Versatile Interactive Motion language model, which integrates both language and motion modalities to effectively understand, generate, and control interactive motions in multi-turn conversational contexts. To address the scarcity of multi-turn interactive motion data, we introduce a synthetic dataset called INTER-MT2; where we utilize pre-trained models to create diverse instructional datasets with interactive motion. Our approach first trains a motion tokenizer that encodes interactive motions into residual discrete tokens. In the pre-training stage, the model learns to align motion and text representations with these discrete tokens. During the instruction fine-tuning stage, VIM adapts to multi-turn conversations using INTER-MT2. We evaluate the versatility of our method across motion-related tasks\u2014motion-to-text, text-to-motion, reaction generation, motion editing, and reasoning about motion sequences. The results highlight VIM\u2019s versatility and effectiveness in handling complex interactive motion synthesis.", "keywords": [ "Motion Synthesis", "Human Motion Modeling", "Interactive Motion", "Motion Language Model" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 1, 1, 2, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "chfJJYC3iL", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Armando Solar-Lezama", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Fanjia Yan", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" }, { "name": "Ion Stoica", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Koushik Sen", "gender": "Male", "institution": "UC Berkeley, University of California, Berkeley", "country": "", "position": "Full Professor" }, { "name": "Naman Jain", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Sida Wang", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Tianjun Zhang", "gender": "unknown", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Wen-Ding Li", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Alex Felix Gu", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Shangdian Han", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 71, "n_ref": 121, "n_ref_all": 153, "n_fig": 13, "n_tab": 4, "L_tab": 1743, "n_element_tab": 152, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 1915, "n_element_tab_1": 120, "formula_len_all": 259, "formula_len_all_1": 0, "len_all": 223062, "len_all_1": 66559, "len_abs": 1705, "len_title": 139, "len_sents": 51192, "len_sents_1": 30150, "n_sents": 442, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1711, "title": "LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code", "abs": "Large Language Models (LLMs) applied to code-related applications have emerged as a prominent field, attracting significant interest from academia and industry. However, as new and improved LLMs are developed, existing evaluation benchmarks (e.g., HumanEvla, MBPP) are no longer sufficient for assessing their capabilities suffering from data contamination, overfitting, saturation, and focus on merely code generation. In this work, we propose LiveCodeBench, a comprehensive and contamination-free evaluation of LLMs for code, which collects new problems over time from contests across three competition platforms, Leetcode, Atcoder, and Codeforces. Notably, our benchmark also focuses on a broader range of code-related capabilities, such as self-repair, code execution, and test output prediction, beyond just code generation. Currently, LiveCodeBench hosts over six hundred coding problems that were published between May 2023 and Aug 2024. We evaluate over 50 LLMs on LiveCodeBench (LCB for brevity) presenting the largest evaluation study of code LLMs on competition problems. Based on the study, we present novel empirical findings on contamination, overfitting, and holistic evaluations. We demonstrate that time-segmented evaluations serve as a robust approach to evade contamination; they are successful at detecting contamination across a wide range of open and closed models including GPT-4O, Claude, Deepseek, and Codestral. Next, we highlight overfitting and saturation of traditional coding benchmarks like HumanEvla and demonstrate LCB allows more reliable evaluations. Finally, our holistic evaluation scenarios allow for measuring the different capabilities of programming agents in isolation.", "keywords": [ "Code LLMs; Evaluation; Contaminationl; Overfitting" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "che9LCwPQM", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Tao Li", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Yang Li", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Zhiwei Deng", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Research Scientist" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 58, "n_ref": 114, "n_ref_all": 122, "n_fig": 3, "n_tab": 5, "L_tab": 1842, "n_element_tab": 113, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2155, "n_element_tab_1": 117, "formula_len_all": 403, "formula_len_all_1": 403, "len_all": 137797, "len_all_1": 82934, "len_abs": 1280, "len_title": 115, "len_sents": 28508, "len_sents_1": 28203, "n_sents": 208, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1286, "title": "Influential Language Data Selection via Gradient Trajectory Pursuit", "abs": "Curating a desirable dataset for training has been the core of building highly capable large language models (Touvron et al., 2023; Achiam et al., 2023; Team et al., 2024). Gradient influence scores (Pruthi et al., 2020; Xia et al., 2024) have been shown to be correlated with model performance and are commonly used as the criterion for data selection. However, existing methods are built upon either individual sample rankings or inefficient matching process, leading to suboptimal performance or scaling up issues. In this paper, we propose Gradient Trajectory Pursuit (GTP), an algorithm that performs pursuit of gradient trajectories via jointly selecting data points under an L0-norm regularized objective. The proposed algorithm highlights: (1) joint selection instead of independent top-k selection, which automatically de-duplicates samples; (2) higher efficiency with compressive sampling processes, which can be further sped up using a distributed framework. In the experiments, we demonstrate the algorithm in both in-domain and target-domain selection benchmarks and show that it outperforms top-k selection and competitive algorithms consistently, for example, our algorithm chooses as low as 0.5% data to achieve full performance on the targeted instruction tuning tasks.", "keywords": [ "Language data selection; LLM" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 2, 3, 3, 5 ] }, { "paper_id": "chanJGoa7f", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Clement Neo", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "David Krueger", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Assistant Professor" }, { "name": "Fazl Barez", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Luke Ong", "gender": "Male", "institution": "Nanyang Technological University", "country": "GB", "position": "Full Professor" }, { "name": "Mor Geva", "gender": "Female", "institution": "Tel Aviv University", "country": "IL", "position": "Assistant Professor" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 35, "n_ref": 57, "n_ref_all": 73, "n_fig": 8, "n_tab": 2, "L_tab": 547, "n_element_tab": 65, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 217, "formula_len_all_1": 217, "len_all": 129102, "len_all_1": 64564, "len_abs": 1200, "len_title": 124, "len_sents": 33800, "len_sents_1": 30272, "n_sents": 258, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1205, "title": "Towards Interpreting Visual Information Processing in Vision-Language Models", "abs": "Vision-Language Models (VLMs) are powerful tools for processing and understanding text and images. We study the processing of visual tokens in the language model component of LLaVA, a prominent VLM. Our approach focuses on analyzing the localization of object information, the evolution of visual token representations across layers, and the mechanism of integrating visual information for predictions. Through ablation studies, we demonstrated that object identification accuracy drops by over 70\\% when object-specific tokens are removed. We observed that visual token representations become increasingly interpretable in the vocabulary space across layers, suggesting an alignment with textual tokens corresponding to image content. Finally, we found that the model extracts object information from these refined representations at the last token position for prediction, mirroring the process in text-only language models for factual association tasks. These findings provide crucial insights into how VLMs process and integrate visual information, bridging the gap between our understanding of language and vision models, and paving the way for more interpretable and controllable multimodal systems.", "keywords": [ "Interpretability", "Vision-Language Model", "Mechanistic Interpretability", "Multimodal" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 2, 4, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "cgxqD6tr4E", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Louis Mahon", "gender": "unknown", "institution": "School of Informatics, University of Edinburgh", "country": "GB", "position": "Postdoc" }, { "name": "Mirella Lapata", "gender": "Female", "institution": "Edinburgh University, University of Edinburgh", "country": "GB", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 46, "n_ref": 67, "n_ref_all": 86, "n_fig": 3, "n_tab": 9, "L_tab": 4571, "n_element_tab": 94, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 450, "n_element_tab_1": 28, "formula_len_all": 498, "formula_len_all_1": 339, "len_all": 432131, "len_all_1": 66349, "len_abs": 2790, "len_title": 118, "len_sents": 110036, "len_sents_1": 33429, "n_sents": 1470, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1357, "title": "ScreenWriter: Automatic Screenplay Generation and Movie Summarisation", "abs": "The proliferation of creative video content has driven demand for textual descriptions or summaries that allow users to recall key plot points or get an overview without watching. The volume of movie content and speed of turnover motivates automatic summarisation, which is nevertheless challenging, requiring identifying character intentions and very long-range temporal dependencies. The few existing methods attempting this task rely heavily on textual screenplays as input, greatly limiting their applicability. In this work, we propose the task of automatic screenplay generation, and a method, ScreenWriter, that operates only on video input and produces output which includes dialogue, speaker names, scene breaks and visual descriptions. ScreenWriter introduces a novel algorithm to segment the video into scenes based on the sequence of visual vectors, and a novel method for the challenging problem of determining character names, based on a database of actors\u2019 faces. We further demonstrate how these automatic screenplays can be used to generate plot synopses with a hierarchical summarisation method based on scene breaks. We test the quality of the final summaries on the recent Moviesumm dataset, which we augment with videos, and show that they are superior to a num- ber of comparison models which assume access to goldstandard screenplays.", "keywords": [ "video understanding; summarisation; scene segmentation" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "cfuZKjGDW7", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Achal Dave", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Cheng-Yen Hsieh", "gender": "Not Specified", "institution": "ByteDance AI Lab Research", "country": "US", "position": "Researcher" }, { "name": "Deva Ramanan", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Kaihua Chen", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Tarasha Khurana", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 67, "n_ref": 192, "n_ref_all": 211, "n_fig": 7, "n_tab": 8, "L_tab": 2201, "n_element_tab": 273, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 33, "n_element_tab_1": 2, "formula_len_all": 47, "formula_len_all_1": 47, "len_all": 211561, "len_all_1": 64376, "len_abs": 622, "len_title": 104, "len_sents": 55691, "len_sents_1": 32108, "n_sents": 457, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1202, "title": "TAO-Amodal: A Benchmark for Tracking Any Object Amodally", "abs": "Amodal perception, the ability to comprehend complete object structures from partial visibility, is a fundamental skill, even for infants. Its significance extends to applications like autonomous driving, where a clear understanding of heavily occluded objects is essential. However, modern detection and tracking algorithms often overlook this critical capability, perhaps due to the prevalence of \\textit{modal} annotations in most benchmarks. To address the scarcity of amodal benchmarks, we introduce TAO-Amodal, featuring 833 diverse categories in thousands of video sequences. Our dataset includes \\textit{amodal} and modal bounding boxes for visible and partially or fully occluded objects, including those that are partially out of the camera frame. We investigate the current lay of the land in both amodal tracking and detection by benchmarking state-of-the-art modal trackers and amodal segmentation methods. We find that existing methods, even when adapted for amodal tracking, struggle to detect and track objects under heavy occlusion. To mitigate this, we explore simple finetuning schemes that can increase the amodal tracking and detection metrics of occluded objects by 2.1% and 3.3%.", "keywords": [ "Amodal perception", "Large-scale evaluation benchmark", "Multi-object tracking." ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "cfn2O1qvxp", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Junbo Guo", "gender": "Male", "institution": "People's Daily Online ", "country": "", "position": "Researcher" }, { "name": "Lei Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" }, { "name": "Mingye Zhu", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yi Liu", "gender": "Male", "institution": "People's Daily Online", "country": "CN", "position": "Researcher" }, { "name": "Zhendong Mao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 32, "n_formula_1": 12, "n_ref_uni": 17, "n_ref": 37, "n_ref_all": 58, "n_fig": 9, "n_tab": 10, "L_tab": 2486, "n_element_tab": 331, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1477, "n_element_tab_1": 273, "formula_len_all": 1923, "formula_len_all_1": 845, "len_all": 177606, "len_all_1": 63036, "len_abs": 1602, "len_title": 109, "len_sents": 66259, "len_sents_1": 27241, "n_sents": 514, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1609, "title": "On-the-fly Preference Alignment via Principle-Guided Decoding", "abs": "With the rapidly expanding landscape of large language models, aligning model generations with human values and preferences is becoming increasingly important. Popular alignment methods, such as Reinforcement Learning from Human Feedback, have shown significant success in guiding models with greater control. However, these methods require considerable computational resources, which is inefficient, and substantial collection of training data to accommodate the diverse and pluralistic nature of human preferences, which is impractical. These limitations significantly constrain the scope and efficacy of both task-specific and general preference alignment methods. In this work, we introduce On-the-fly Preference Alignment via Principle-Guided Decoding (OPAD) to directly align\nmodel outputs with human preferences during inference, eliminating the need for fine-tuning. Our approach involves first curating a surrogate solution to an otherwise infeasible optimization problem and then designing a principle-guided reward function based on this surrogate. The final decoding policy is derived by maximizing this customized reward, which exploits the discrepancy between the\nconstrained policy and its unconstrained counterpart. OPAD directly modifies the model\u2019s predictions during inference, ensuring principle adherence without incurring the computational overhead of retraining or fine-tuning. Experiments show that OPAD achieves competitive or superior performance in both general and personalized alignment tasks, demonstrating its efficiency and effectiveness compared to state-of-the-art baselines.", "keywords": [ "preference alignment", "tuning-free alignment", "principle-based decoding" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "cfe2zDg1G8", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Huifeng Guo", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Jingtong Gao", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Pengyue Jia", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Ruiming Tang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Wanyu Wang", "gender": "Female", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xiangyu Zhao", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Xiaopeng Li", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yejing Wang", "gender": "Male", "institution": "City University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Yichao Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yuhao Wang", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 33, "n_ref": 95, "n_ref_all": 133, "n_fig": 3, "n_tab": 13, "L_tab": 8248, "n_element_tab": 738, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 7838, "n_element_tab_1": 492, "formula_len_all": 314, "formula_len_all_1": 55, "len_all": 207902, "len_all_1": 68949, "len_abs": 1055, "len_title": 109, "len_sents": 66704, "len_sents_1": 25960, "n_sents": 482, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1031, "title": "Scenario-Wise Rec: A Multi-Scenario Recommendation Benchmark", "abs": "Multi Scenario Recommendation (MSR) tasks, referring to building a unified model to enhance performance across all recommendation scenarios, have recently gained much attention. However, current research in MSR faces two significant challenges that hinder the field's development: the absence of uniform procedures for multi-scenario dataset processing, thus hindering fair comparisons, and most models being closed-sourced, which complicates comparisons with current SOTA models. Consequently, we introduce our benchmark, Scenario-Wise Rec, which comprises six public datasets and twelve benchmark models, along with a training and evaluation pipeline. We have also validated our benchmark using an industrial advertising dataset, further enhancing its real-world reliability. We aim for this benchmark to provide researchers with valuable insights from prior works, enabling the development of novel models based on our benchmark and thereby fostering a collaborative research ecosystem in MSR. Our source code is also available.", "keywords": [ "Recommender System", "Multi-scenario Recommendation", "Click-Through Rate Prediction" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "cfKZ5VrhXt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Chenjia Bai", "gender": "Male", "institution": "TeleAI, China Telecom", "country": "CN", "position": "Research Scientist" }, { "name": "Kang Xu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Qiaosheng Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Shuang Qiu", "gender": "Male", "institution": "City University of Hong Kong", "country": "CN", "position": "Assistant Professor" }, { "name": "Xuelong Li", "gender": "Male", "institution": "China Telecom", "country": "CN", "position": "Full Professor" }, { "name": "Yang Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 60, "n_formula_1": 32, "n_ref_uni": 55, "n_ref": 116, "n_ref_all": 150, "n_fig": 1, "n_tab": 8, "L_tab": 2076, "n_element_tab": 296, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 1575, "n_element_tab_1": 220, "formula_len_all": 5040, "formula_len_all_1": 2230, "len_all": 193629, "len_all_1": 74544, "len_abs": 1552, "len_title": 123, "len_sents": 51968, "len_sents_1": 32926, "n_sents": 385, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 75, "L_abs": 1559, "title": "Online Preference Alignment for Language Models via Count-based Exploration", "abs": "Reinforcement Learning from Human Feedback (RLHF) has shown great potential in fine-tuning Large Language Models (LLMs) to align with human preferences. Existing methods perform preference alignment from a fixed dataset, which can be limited in data coverage and the resulting reward model is hard to generalize in out-of-distribution responses. Thus, online RLHF is more desirable to empower the LLM to explore outside the support of the initial dataset by iteratively collecting the prompt-response pairs. In this paper, we study the fundamental problem in online RLHF, i.e., how to explore for LLM. We give a theoretical motivation in linear reward assumption to show that an optimistic reward with an upper confidence bound (UCB) term leads to a provably efficient RLHF policy. Then, we reformulate our objective to direct preference optimization with an exploration term, where the UCB-term can be converted to a count-based exploration bonus. We further propose a practical algorithm, named Count-based Online Preference Optimization (COPO), which leverages a simple coin-flip counting module to estimate the pseudo-count of a prompt-response pair in previously collected data. COPO encourages LLMs to balance exploration and preference optimization in an iterative manner, which enlarges the exploration space and the entire data coverage of iterative LLM policies. We conduct online RLHF experiments on Zephyr and Llama-3 models. The results on instruction-following and standard academic benchmarks show that COPO significantly increases performance.", "keywords": [ "Reinforcement Learning from Human Feedback", "RLHF", "Preference Alignment", "Exploration", "LLMs" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "cfGpIcOIa5", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Richong Zhang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Xu Zhang", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Si CHEN", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 28, "n_ref": 48, "n_ref_all": 66, "n_fig": 1, "n_tab": 5, "L_tab": 1477, "n_element_tab": 87, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 436, "n_element_tab_1": 100, "formula_len_all": 1343, "formula_len_all_1": 619, "len_all": 202303, "len_all_1": 64555, "len_abs": 1121, "len_title": 111, "len_sents": 34134, "len_sents_1": 27503, "n_sents": 267, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1249, "title": "GeoILP: A Synthetic Dataset to Guide Large-Scale Rule Induction", "abs": "Inductive logic programming (ILP) is a machine learning approach aiming to learn explanatory rules from data.\n While existing ILP systems can successfully solve small-scale tasks, large-scale applications with various language biases are rarely explored.\n Besides, it is crucial for a large majority of current ILP systems to require expert-defined language bias, which hampers the development of ILP towards broader utilizations.\n In this paper, we introduce GeoILP, a large-scale synthetic dataset of diverse ILP tasks involving numerous aspects of language bias.\n % including complex rule forms, high deduction complexity, and more realistic assumptions.\n The ILP tasks are built from geometry problems, at the level from textbook exercise to regional International Mathematical Olympiad (IMO), with the help of a deduction engine.\n These problems are elaborately selected to cover all challenging language biases, such as recursion, predicate invention, and high arity.\n Experimental results show that no existing method can solve GeoILP tasks.\n In addition, along with classic symbolic-form data, we provide image-form data to boost the development of the joint learning of neural perception and symbolic rule induction.", "keywords": [ "inductive logic programming", "rule induction", "dataset" ], "rating_list": [ 3, 8, 5, 8 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "cfDbuobmU0", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Guilin Qi", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Full Professor" }, { "name": "Guohui Xiao", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Full Professor" }, { "name": "Lin Ren", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "PhD student" }, { "name": "Rihui Jin", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "PhD student" }, { "name": "Tongtong Wu", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Postdoc" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 35, "n_ref": 52, "n_ref_all": 104, "n_fig": 17, "n_tab": 15, "L_tab": 4990, "n_element_tab": 1036, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1921, "n_element_tab_1": 273, "formula_len_all": 645, "formula_len_all_1": 309, "len_all": 175804, "len_all_1": 66725, "len_abs": 1392, "len_title": 135, "len_sents": 50353, "len_sents_1": 27314, "n_sents": 470, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1398, "title": "SymTex: A New Benchmark for Non-monotonic Reasoning Capability of Large Language Models", "abs": "Non-monotonic reasoning (NMR) plays a crucial role in logical reasoning, allowing inference to adjust as new information arises. This adaptability is key for large language models (LLMs) to handle complex problems and adjust reasoning in dynamic environments, mimicking human-like flexibility in thought. Recent works mainly explore using LLMs to address non-monotonic reasoning through textual logic representation, as LLMs excel in understanding natural language. However, textual logic representation often leads to ambiguity and complexity, especially in complex situations, while symbolic logic representation is more clear and precise, avoiding these issues. In this work, we introduce a framework called Multi-step Generation for Symbolic and Textual NMR Samples (MG-SymTex) to generate diverse non-monotonic samples automatically, and build a non-monotonic reasoning benchmark, called SymTex, which is used to evaluate the non-monotonic reasoning capability of LLMs. SymTex comprises two types of description and three types of predicate, facilitating two primary tasks: Tri-State Boolean Querying and Answer Set Computation. Through our comprehensive evaluations, we demonstrate that state-of-the-art LLMs such as gpt-4o, claude-3.5-sonnet, and o1-mini encounter significant challenges when addressing our proposed benchmark, highlighting the difficulty of non-monotonic reasoning in LLMs.", "keywords": [ "Non-monotonic Reasoning", "Large Language Models" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "cf7NTWv1iW", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Hao Mark Chen", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Hongxiang Fan", "gender": "Male", "institution": "Samsung", "country": "", "position": "Researcher" }, { "name": "Konstantin Mishchenko", "gender": "unknown", "institution": "Facebook", "country": "FR", "position": "Researcher" }, { "name": "Rui Li", "gender": "Female", "institution": "Samsung AI Center", "country": "", "position": "Researcher" }, { "name": "Stylianos Venieris", "gender": "unknown", "institution": "Samsung AI", "country": "GB", "position": "Researcher" }, { "name": "Wayne Luk", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Full Professor" }, { "name": "Yiu Ka Fai Cedric", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 5, "n_ref_uni": 16, "n_ref": 57, "n_ref_all": 102, "n_fig": 24, "n_tab": 8, "L_tab": 1449, "n_element_tab": 227, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 62, "n_element_tab_1": 8, "formula_len_all": 320, "formula_len_all_1": 196, "len_all": 146282, "len_all_1": 61676, "len_abs": 1654, "len_title": 139, "len_sents": 54852, "len_sents_1": 29769, "n_sents": 425, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1686, "title": "Hardware-Aware Parallel Prompt Decoding for Memory-Efficient Acceleration of LLM Inference", "abs": "The auto-regressive decoding of Large Language Models (LLMs) results in significant overheads in their hardware performance. \nWhile recent research has investigated various speculative decoding techniques for multi-token generation, these efforts have primarily focused on improving processing speed such as throughput.\nCrucially, they often neglect other metrics essential for real-life deployments, such as memory consumption and training cost.\nTo overcome these limitations, we propose a novel parallel prompt decoding that requires only $0.0002$\\% trainable parameters, enabling efficient training on a single A100-40GB GPU in just 16 hours. \nInspired by the human natural language generation process, PPD approximates outputs generated at future timesteps in parallel by using multiple prompt tokens. This approach partially recovers the missing conditional dependency information necessary for multi-token generation, resulting in up to a 28\\% higher acceptance rate for long-range predictions.\nFurthermore, we present a hardware-aware two-stage tree pruning algorithm that adaptively optimizes this decoding scheme to fully leverage the computational capacities on different GPUs.\nThrough extensive experiments across LLMs ranging from MobileLlama to Vicuna-13B on a wide range of benchmarks, our approach demonstrates up to 2.49$\\times$ speedup and maintains a minimal runtime memory overhead of just $0.0004\\%$.\nMore importantly, our parallel prompt decoding can serve as an orthogonal optimization for synergistic integration with existing speculative decoding,\nshowing up to $1.22\\times$ further speed improvement. Our code will be open-sourced upon acceptance of the paper.", "keywords": [ "LLM inference acceleration", "prompt tuning", "hardware-aware design" ], "rating_list": [ 5, 5, 1, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "cecIf0CKnH", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Mathieu Tanneau", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Michael Klamkin", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Pascal Van Hentenryck", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 2, "n_ref_uni": 37, "n_ref": 57, "n_ref_all": 72, "n_fig": 2, "n_tab": 9, "L_tab": 4070, "n_element_tab": 397, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 739, "n_element_tab_1": 86, "formula_len_all": 3115, "formula_len_all_1": 354, "len_all": 147531, "len_all_1": 64929, "len_abs": 1562, "len_title": 111, "len_sents": 40434, "len_sents_1": 29922, "n_sents": 284, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1568, "title": "PGLearn - An Open-Source Learning Toolkit for Optimal Power Flow", "abs": "Machine learning techniques for Optimal Power Flow (OPF) problems have recently garnered significant attention, reflecting a broader trend of leveraging machine learning to approximate and/or accelerate the resolution of complex optimization problems. These developments are necessitated by the increased volatility and scale in energy production for modern and future grids. However, progress in ML for OPF is hindered by the lack of standardized datasets and evaluation metrics, from generating and solving OPF instances, to training and benchmarking machine learning models. To address this challenge, this paper introduces PGLearn, a comprehensive suite of standardized datasets and evaluation tools for ML and OPF. PGLearn implements realistic data generation procedures that capture both global and local variability, ensuring that datasets are representative of real-world conditions. In addition, it supports multiple OPF formulations, including AC, DC, and second-order cone formulations. Standardized datasets are made publicly available to democratize access to this field, reduce the burden of data generation, and enable the fair comparison of various methodologies. PGLearn also includes a robust toolkit for training, evaluating, and benchmarking machine learning models for OPF, with the goal of standardizing performance evaluation across the field. By promoting open, standardized datasets and evaluation metrics, PGLearn aims at democratizing and accelerating research and innovation in machine learning applications for optimal power flow problems.", "keywords": [ "optimal power flow", "machine learning", "dataset" ], "rating_list": [ 8, 6, 6, 3, 3 ], "soundness_list": [ 3, 2, 3, 2, 1 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 1, 1 ], "confidence_list": [ 4, 4, 5, 4, 5 ] }, { "paper_id": "ceUtIUfotv", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Brigt H\u00e5vardstun", "gender": "Male", "institution": "University of Bergen", "country": "NO", "position": "PhD student" }, { "name": "Cesar Ferri", "gender": "Male", "institution": "Universitat Polit\u00e8cnica de Val\u00e8ncia", "country": "", "position": "Associate Professor" }, { "name": "Dario Garigliotti", "gender": "Male", "institution": "University of Bergen", "country": "NO", "position": "Postdoc" }, { "name": "Diogo Nuno Freitas", "gender": "unknown", "institution": "Universidade da Madeira", "country": "PT", "position": "PhD student" }, { "name": "Jan Arne Telle", "gender": "unknown", "institution": "University of Bergen", "country": "NO", "position": "Full Professor" }, { "name": "Jose Hernandez-Orallo", "gender": "unknown", "institution": "Universitat Politecnica de Valencia", "country": "ES", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 23, "n_ref": 31, "n_ref_all": 49, "n_fig": 5, "n_tab": 2, "L_tab": 452, "n_element_tab": 92, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 176, "n_element_tab_1": 34, "formula_len_all": 160, "formula_len_all_1": 160, "len_all": 96273, "len_all_1": 61669, "len_abs": 1290, "len_title": 141, "len_sents": 33963, "len_sents_1": 30224, "n_sents": 251, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1296, "title": "Relative Drawing Identification Complexity is Invariant to Modality in Vision-Language Models", "abs": "Large language models have become multimodal, and many of them are said to integrate their modalities using common representations. If this were true, a drawing of car as an image, for instance, should map to the similar area in the latent space as a textual description of the strokes that conform the drawing. To explore this in a black-box access regime to these models, we propose the use of machine teaching, a theory that studies the minimal set of examples a teacher needs to choose so that the learner captures the concept. In particular, we apply this to GPT-4V, a multimodal version of GPT-4 that includes support for image analysis, to evaluate the complexity of teaching a subset of objects in the _Quick, Draw!_ dataset using two presentations: raw images as bitmaps and trace coordinates in TikZ format. The results indicate that image-based representations generally require fewer segments and achieve higher accuracy when compared to coordinate-based representations. But, surprisingly, for concepts recognized by both modalities, the teaching size ranks concepts similarly across both modalities, even when controlling for (a human proxy of) concept priors. This could also suggest that the simplicity of concepts is an inherent property that transcends modality representations.", "keywords": [ "Multimodal Language Models", "Machine Teaching", "Concept Identification", "Drawing Simplification" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "cd79pbXi4N", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Elias Abad Rocamora", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Grigorios Chrysos", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Volkan Cevher", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" } ], "n_formula": 33, "n_formula_1": 23, "n_ref_uni": 41, "n_ref": 113, "n_ref_all": 151, "n_fig": 3, "n_tab": 5, "L_tab": 2004, "n_element_tab": 252, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1562, "n_element_tab_1": 239, "formula_len_all": 3545, "formula_len_all_1": 2190, "len_all": 160876, "len_all_1": 70185, "len_abs": 1002, "len_title": 103, "len_sents": 42736, "len_sents_1": 28409, "n_sents": 375, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1017, "title": "Certified Robustness Under Bounded Levenshtein Distance", "abs": "Text classifiers suffer from small perturbations, that if chosen adversarially, can dramatically change the output of the model. Verification methods can provide robustness certificates against such adversarial perturbations, by computing a sound lower bound on the robust accuracy. Nevertheless, existing verification methods incur in prohibitive costs and cannot practically handle Levenshtein distance constraints. We propose the first method for computing the Lipschitz constant of convolutional classifiers with respect to the Levenshtein distance. We use these Lipschitz constant estimates for training 1-Lipschitz classifiers. This enables computing the certified radius of a classifier in a single forward pass. Our method, LipsLev, is able to obtain $38.80$% and $13.93$% verified accuracy at distance $1$ and $2$ respectively in the AG-News dataset, while being $4$ orders of magnitude faster than existing approaches. We believe our work can open the door to more efficient verification in the text domain.", "keywords": [ "Robustness verification", "Text classifiers", "Lipschitz constant" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "cbttLtO94Q", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Anastasios Nikolas Angelopoulos", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Banghua Zhu", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "Connor Chen", "gender": "unknown", "institution": "UC Berkeley, University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Evan Frick", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" }, { "name": "Ion Stoica", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Jiantao Jiao", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" }, { "name": "Joseph E. Gonzalez", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Associate Professor" }, { "name": "Tianle Li", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Wei-Lin Chiang", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 46, "n_ref_all": 64, "n_fig": 8, "n_tab": 24, "L_tab": 32642, "n_element_tab": 3751, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 308, "n_element_tab_1": 6, "formula_len_all": 37, "formula_len_all_1": 0, "len_all": 236446, "len_all_1": 56340, "len_abs": 1124, "len_title": 86, "len_sents": 45754, "len_sents_1": 27385, "n_sents": 361, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 38, "L_abs": 1130, "title": "How to Evaluate Reward Models for RLHF", "abs": "We introduce a new benchmark for reward models that quantifies their ability to produce strong language models through RLHF (Reinforcement Learning from Human Feedback).\nThe gold-standard approach is to run a full RLHF training pipeline and directly probe downstream LLM performance.\nHowever, this process is prohibitively expensive.\nTo address this, we build a predictive model of downstream LLM performance by evaluating the reward model on proxy tasks. \nThese proxy tasks consist of a large-scale human preference and a verifiable correctness preference dataset, in which we measure 12 metrics across 12 domains.\nTo investigate which reward model metrics are most correlated to gold-standard RLHF outcomes, we launch an end-to-end RLHF experiment on a large-scale crowd-sourced human preference platform to view real reward model downstream performance as ground truth. \nUltimately, we compile our data and findings into Preference Proxy Evaluations (PPE), the first reward model benchmark explicitly linked to post-RLHF real-world human preference performance, which we will open-source for public use and further development.", "keywords": [ "RLHF", "RL", "Reward Model", "LLM", "Benchmark", "Dataset", "Evaluation" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "cbFqqtJGtA", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Menghua Wu", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Regina Barzilay", "gender": "female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Professor" }, { "name": "SH Murphy", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Tommi Jaakkola", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Umesh Padia", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 3, "n_ref_uni": 35, "n_ref": 75, "n_ref_all": 107, "n_fig": 9, "n_tab": 14, "L_tab": 7471, "n_element_tab": 405, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2932, "n_element_tab_1": 154, "formula_len_all": 411, "formula_len_all_1": 127, "len_all": 187672, "len_all_1": 68902, "len_abs": 2288, "len_title": 113, "len_sents": 47320, "len_sents_1": 30239, "n_sents": 640, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1594, "title": "Predicting perturbation targets with causal differential networks", "abs": "Rationally identifying variables responsible for changes to a biological system can enable myriad applications in disease understanding and cell engineering. From a causality perspective, we are given two datasets generated by the same causal model, one observational (control) and one interventional (perturbed). The goal is to isolate the subset of measured variables (e.g. genes) that were the targets of the intervention, i.e. those whose conditional independencies have changed. Knowing the causal graph would limit the search space, allowing us to efficiently pinpoint these variables. However, current algorithms that infer causal graphs in the presence of unknown intervention targets scale poorly to the hundreds or thousands of variables in biological data, as they must jointly search the combinatorial spaces of graphs and consistent intervention targets. In this work, we propose a causality-inspired approach for predicting perturbation targets that decouples the two search steps. First, we use an amortized causal discovery model to separately infer causal graphs from the observational and interventional datasets. Then, we learn to map these paired graphs to the sets of variables that were intervened upon, in a supervised learning framework. This approach consistently outperforms baselines for perturbation modeling on seven single-cell transcriptomics datasets, each with thousands of measured variables. We also demonstrate significant improvements over six causal discovery algorithms in predicting intervention targets across a variety of tractable, synthetic datasets.", "keywords": [ "perturbation experiments", "Perturb-seq", "transcriptomics", "causality" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 1, 1, 2 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "cb4etlGvOY", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Abhishek Dutta", "gender": "Male", "institution": "University of Connecticut", "country": "US", "position": "Assistant Professor" }, { "name": "Yen-Che Hsiao", "gender": "Male", "institution": "University of Connecticut", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 12, "n_ref": 69, "n_ref_all": 90, "n_fig": 7, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 195, "n_element_tab_1": 24, "formula_len_all": 781, "formula_len_all_1": 789, "len_all": 78727, "len_all_1": 71783, "len_abs": 825, "len_title": 134, "len_sents": 28590, "len_sents_1": 26051, "n_sents": 219, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 828, "title": "Towards Autonomous Agents: Adaptive-planning, Reasoning, and Acting in Language Models", "abs": "We propose a novel in-context learning algorithm for building autonomous decision-making language agents. The language agent continuously attempts to solve the same task by reasoning, acting, observing and then self-correcting each time the task fails. Our selected language agent demonstrates the ability to solve tasks in a text-based game environment. Our results show that the gemma-2-9b-it language model, using our proposed method, can successfully complete two of six tasks that failed in the first attempt. This highlights the effectiveness of our approach in enhancing the problem-solving capabilities of a single language model through self-correction, paving the way for more advanced autonomous agents. The code is publicly available at https://anonymous.4open.science/r/AutonomousLLMAgentwithAdaptingPlanning-D613/.", "keywords": [ "llm agents", "in-context learning", "autonomous agents" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 1, 2, 3 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "cb4PoT7ePW", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Ming Liu", "gender": "Not Specified", "institution": "Iowa State University", "country": "US", "position": "PhD student" }, { "name": "Wensheng Zhang", "gender": "unknown", "institution": "Iowa State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 33, "n_ref": 194, "n_ref_all": 206, "n_fig": 3, "n_tab": 6, "L_tab": 3044, "n_element_tab": 171, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 926, "n_element_tab_1": 103, "formula_len_all": 797, "formula_len_all_1": 806, "len_all": 136146, "len_all_1": 55794, "len_abs": 1745, "len_title": 135, "len_sents": 36331, "len_sents_1": 26883, "n_sents": 262, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1560, "title": "Can Decoding by Contrasting Layers Really Improve Factuality in Large Language Models?", "abs": "Large language models (LLMs) have made notable advancements across diverse applications, but their susceptibility to hallucinations remains a critical challenge. That is, they could produce outputs divergent from real-world evidence or user-provided inputs. Recent studies have explored a contrastive decoding strategy known as DoLa, which mitigates output inaccuracy by contrasting the outputs from the final layer against those from the previous layers. Nevertheless, such strategy has its limitation, as LLMs, which already have internalized extensive parametric knowledge through comprehensive pre-training and fine-tuning phases, may generate errors due to incorrect or obsolete information within their parameters. As an alternative, trusted external knowledge could be included in the prompt context for querying, but the constrained context window of LLMs poses a significant barrier restricting the amount of information that can be provided. \nTo address the above issues, we propose to integrate the contrasive decoding strategy with a long-context encoder that effectively condenses extensive initial contexts into a more concise format. Extensive experiments have demonstrated that, \nour proposed methodology enhances the factual accuracy of the produced content, \nwhen applied to various datasets. For instance, it has improved the performance of LLaMA2-7B models on the Quality dataset by 61.61\\%, compared to the DoLa decoding method, showcasing its effectiveness in enhancing the reliability of LLMs in generating truthful information.", "keywords": [ "Factuality", "Contrastive Decoding", "Parametric Memory" ], "rating_list": [ 6, 5, 3, 5, 5 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 1, 2, 3 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 5, 3, 3 ] }, { "paper_id": "cazOlqncU6", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Conghui He", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Jie LI", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Na Ruan", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "", "position": "Associate Professor" }, { "name": "Zekun Sun", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhihao Sui", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 15, "n_ref": 58, "n_ref_all": 67, "n_fig": 5, "n_tab": 6, "L_tab": 1525, "n_element_tab": 61, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1378, "n_element_tab_1": 51, "formula_len_all": 126, "formula_len_all_1": 71, "len_all": 124792, "len_all_1": 75150, "len_abs": 1371, "len_title": 152, "len_sents": 51265, "len_sents_1": 35668, "n_sents": 387, "n_sents_1": 264, "venue": "ICLR2025", "decision": "withdraw", "L_title": 104, "L_abs": 1376, "title": "Trustworthy Dataset Proof: Certifying the Authentic Use of Dataset in Training Models for Enhanced Trust", "abs": "In the realm of deep learning, the veracity and integrity of the training data are pivotal for constructing reliable and transparent models. This study introduces the concept of Trustworthy Dataset Proof (TDP), which tackles the significant challenge of verifying the authenticity of training data as declared by trainers. Existing dataset provenance methods, which primarily aim at ownership verification rather than trust enhancement, often face challenges with usability and integrity. For instance, excessive operational demands and the inability to effectively verify dataset authenticity hinder their practical application. To address these shortcomings, we propose a novel technique termed Data Probe, which diverges from traditional watermarking by utilizing subtle variations in model output distributions to confirm the presence of a specific and small subset of training data. This model-agnostic approach improves usability by minimizing the intervention during the training process and ensures dataset integrity via a mechanism that only permits probe detection when the entire claimed dataset is utilized in training. Our study conducts extensive evaluations to demonstrate the effectiveness of the proposed data-drobe-based TDP framework, marking a significant step toward achieving transparency and trustworthiness in the use of training data in deep learning.", "keywords": [ "dataset integrity; trustworthy dataset proof; data probe; watermark" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "cayKVPCrOP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Di Huang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Postdoc" }, { "name": "\u65b9\u6d77\u7537", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Science", "country": "CN", "position": "MS student" }, { "name": "QiCheng Wang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Qi Guo", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Rui Zhang", "gender": "Female", "institution": "Institute of Computing Technology, CAS", "country": "CN", "position": "Assistant Professor" }, { "name": "Shuo Wang", "gender": "Male", "institution": "CAS", "country": "CN", "position": "MS student" }, { "name": "Tonghui He", "gender": "Male", "institution": ", Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Yuanbo Wen", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Yunpu Zhao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 23, "n_ref": 39, "n_ref_all": 62, "n_fig": 7, "n_tab": 3, "L_tab": 853, "n_element_tab": 112, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 396, "n_element_tab_1": 26, "formula_len_all": 845, "formula_len_all_1": 0, "len_all": 142107, "len_all_1": 60454, "len_abs": 1472, "len_title": 92, "len_sents": 46720, "len_sents_1": 30510, "n_sents": 333, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 1482, "title": "GOOD: Decoding-Time Black-Box LLM Alignment", "abs": "Large Language Models (LLMs) have demonstrated immense potential across various applications. However, aligning these models with specific real-world tasks and human preferences typically requires resource-intensive fine-tuning processes such as Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF).\nIn this paper, we propose GOOD (Guided Online Optimal Decoding), a novel alignment method that enhances pre-trained models without the need for parameter fine-tuning. We observed that the alignment-related behavior of one model can be used to guide another model, and based on this insight, we proposed the GOOD method. Utilizing a pair of guiding models, GOOD identifies critical positions related to alignment and adjusts the model\u2019s output dynamically during the response generation. Notably, the interaction between the guiding models and the guided model occurs at the string level, enabling GOOD to be applied to align even black-box models.\nExperiments show that GOOD can achieve performance comparable to or even surpassing direct fine-tuning in terms of comprehensive capability and harmless generation, reaching relative scores of 108\\% and 105\\% respectively. Even in weak-to-strong alignment, it can recover up to 94\\% of the performance of directly fine-tuned models. GOOD can also be applied to enhance already aligned models (improving pass@1 by 52\\% in code enhancement), making it compatible with various existing alignment techniques.", "keywords": [ "Large language models", "Alignment", "Black-Box" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "cagNCwQEEN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Han Li", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Jianing Zhou", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" }, { "name": "Lingyun Wang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Ning Xie", "gender": "Female", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Ruijie Wang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Postdoc" }, { "name": "Sheng Liu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Suma Bhat", "gender": "unknown", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Xiaohan Nie", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "SHUAI Zhang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 27, "n_ref": 81, "n_ref_all": 92, "n_fig": 3, "n_tab": 7, "L_tab": 4987, "n_element_tab": 811, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 3238, "n_element_tab_1": 555, "formula_len_all": 150, "formula_len_all_1": 150, "len_all": 165920, "len_all_1": 57496, "len_abs": 1581, "len_title": 108, "len_sents": 30945, "len_sents_1": 21296, "n_sents": 241, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1587, "title": "Multimodal Instruction Tuning with Hybrid State Space Models", "abs": "Handling lengthy context is crucial for enhancing the recognition and understanding capabilities of multimodal large language models (MLLMs) in applications such as processing high-resolution images or high frame rate videos. The rise in image resolution and frame rate substantially increases computational demands due to the increased number of input tokens. This challenge is further exacerbated by the quadratic complexity with respect to sequence length of the self-attention mechanism. Most prior works either pre-train models with long contexts, overlooking the efficiency problem, or attempt to reduce the context length via downsampling (e.g., identify the key image patches or frames) to decrease the context length, which may result in information loss. To circumvent this issue while keeping the remarkable effectiveness of MLLMs, we propose a novel approach using a hybrid transformer-MAMBA model to efficiently handle long contexts in multimodal applications. Our multimodal model can effectively process long context input exceeding 100k tokens, outperforming existing models across various benchmarks. Remarkably, our model enhances inference efficiency for high-resolution images and high-frame-rate videos by about 4 times compared to current models, with efficiency gains increasing as image resolution or video frames rise. Furthermore, our model is the first to be trained on low-resolution images or low-frame-rate videos while being capable of inference on high-resolution images and high-frame-rate videos, offering flexibility for inference in diverse scenarios.", "keywords": [ "multimodal", "hybrid SSM", "Jamba" ], "rating_list": [ 3, 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 1, 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "caY45V0dYt", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Dayan Wu", "gender": "Male", "institution": "iie,cas", "country": "CN", "position": "Associate Professor" }, { "name": "JingziGU", "gender": "Female", "institution": "INSTATUTE OF INFORMATION ENGINEERING,CAS", "country": "CN", "position": "Associate Professor" }, { "name": "Jinyang An", "gender": "Male", "institution": "Institute of Information Engineering, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Ming Li", "gender": "Male", "institution": "Guangming Laboratory", "country": "CN", "position": "Full Professor" }, { "name": "Wanqian Zhang", "gender": "Male", "institution": "IIE, CAS", "country": "CN", "position": "Associate Professor" }, { "name": "Weiping Wang", "gender": "Male", "institution": "IIE", "country": "CN", "position": "Full Professor" }, { "name": "Yufan Liu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "zheng Lin", "gender": "Female", "institution": "Institute of Information Engineering, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 22, "n_ref": 52, "n_ref_all": 66, "n_fig": 5, "n_tab": 5, "L_tab": 3418, "n_element_tab": 463, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 683, "n_element_tab_1": 27, "formula_len_all": 513, "formula_len_all_1": 550, "len_all": 105898, "len_all_1": 53374, "len_abs": 0, "len_title": 47, "len_sents": 30545, "len_sents_1": 26227, "n_sents": 216, "n_sents_1": 162, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1580, "title": "RealEra:Semantic-level Concept Erasure via Neighbor-Concept Mining", "abs": "The remarkable development of text-to-image generation models has raised notable security concerns, such as the infringement of portrait rights and the generation of inappropriate content. Concept erasure has been proposed to remove the model's knowledge about protected and inappropriate concepts. Although many methods have tried to balance the efficacy (erasing target concepts) and specificity (retaining irrelevant concepts), they can still generate abundant erasure concepts under the steering of semantically related inputs. In this work, we propose RealEra to address this \"concept residue\" issue. Specifically, we first introduce the mechanism of neighbor-concept mining, digging out the associated concepts by adding random perturbation into the embedding of erasure concept, thus expanding the erasing range and eliminating the generations even through associated concept inputs. Furthermore, to mitigate the negative impact on the generation of irrelevant concepts caused by the expansion of erasure scope, RealEra preserves the specificity through the beyond-concept regularization. This makes irrelevant concepts maintain their corresponding spatial position, thereby preserving their normal generation performance. We also employ the closed-form solution to optimize weights of U-Net for the cross-attention alignment, as well as the prediction noise alignment with the LoRA module. Extensive experiments on multiple benchmarks demonstrate that RealEra outperforms previous concept erasing methods in terms of superior erasing efficacy, specificity, and generality.", "keywords": [ "concept erasure", "diffusion model", "text-image generation" ], "rating_list": [ 3, 1, 5, 5, 3 ], "soundness_list": [ 2, 1, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "caE5faFVT1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Hassan Abid", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "MS student" }, { "name": "Muhammad Haris Khan", "gender": "Male", "institution": "Mohamed Bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Muhammad Ibraheem Siddiqui", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Muhammad Umer Sheikh", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 29, "n_ref": 55, "n_ref_all": 82, "n_fig": 9, "n_tab": 3, "L_tab": 1074, "n_element_tab": 74, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 399, "n_element_tab_1": 25, "formula_len_all": 641, "formula_len_all_1": 341, "len_all": 136364, "len_all_1": 59895, "len_abs": 1634, "len_title": 109, "len_sents": 46747, "len_sents_1": 30149, "n_sents": 327, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1988, "title": "PerSense: Personalized Instance Segmentation in Dense Images", "abs": "Leveraging large-scale pre-training, vision foundational models showcase notable performance benefits. Recent segmentation algorithms for natural scenes have advanced significantly. However, existing models still struggle to automatically segment personalized instances in dense and crowded scenarios, where severe occlusions, scale variations, and background clutter pose a challenge to accurately delineate densely packed instances of the target object. To address this, we propose **PerSense**, an end-to-end, training-free, and model-agnostic one-shot framework for **Per**sonalized instance **S**egmentation in d**ense** images. Towards developing this framework, we make the following core contributions. **(a)** We develop a new baseline capable of automatically generating instance-level point prompts via proposing a novel Instance Detection Module (IDM) that leverages density maps, encapsulating spatial distribution of objects in an image. **(b)** To mitigate false positives within generated point prompts, we design Point Prompt Selection Module (PPSM). Both IDM and PPSM transform density maps into personalized precise point prompts for instance-level segmentation and offer a seamless integration in our model-agnostic framework. **(c)** We introduce a feedback mechanism which enables PerSense to improve the accuracy of density maps by automating the exemplar selection process for density map generation. **(d)** To promote algorithmic advances and effective tools for this relatively underexplored task, we introduce PerSense-D, a diverse dataset exclusive to personalized instance segmentation in dense images. Our extensive experiments establish PerSense superiority in dense scenarios by achieving an mIoU of **71.61%** on PerSense-D, outperforming recent SOTA models by significant margins of **+47.16%**, **+42.27%**, **+8.83%**, and **+5.69%**. Additionally, our qualitative findings demonstrate the adaptability of our framework to images captured in-the-wild.", "keywords": [ "dense image segmentation", "personalized instance segmentation", "training-free", "one-shot segmentation" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 4, 2, 2, 2 ], "presentation_list": [ 4, 2, 2, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "cZZMC8VFZc", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 6, "author_info_list": [ { "name": "David E. Keyes", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Full Professor" }, { "name": "Di Wang", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Jie Ren", "gender": "unknown", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Junxiao Wang", "gender": "Male", "institution": "Guangzhou University", "country": "CN", "position": "Associate Professor" }, { "name": "Liangyu Wang", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Zihang Xiang", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 9, "n_ref": 10, "n_ref_all": 26, "n_fig": 6, "n_tab": 2, "L_tab": 1459, "n_element_tab": 106, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2834, "n_element_tab_1": 128, "formula_len_all": 142, "formula_len_all_1": 142, "len_all": 125993, "len_all_1": 63470, "len_abs": 1534, "len_title": 136, "len_sents": 32355, "len_sents_1": 28297, "n_sents": 222, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1448, "title": "FlashDP: Memory-Efficient and High-Throughput DP-SGD Training for Large Language Models", "abs": "As large language models (LLMs) increasingly underpin technological advancements, the privacy of their training data emerges as a critical concern. Differential Privacy (DP) serves as a rigorous mechanism to protect this data, yet its integration via Differentially Private Stochastic Gradient Descent (DP-SGD) introduces substantial challenges, primarily due to the complexities of per-sample gradient clipping. Current explicit methods, such as Opacus, necessitate extensive storage for per-sample gradients, significantly inflating memory requirements. Conversely, implicit methods like GhostClip reduce storage needs by recalculating gradients multiple times, which leads to inefficiencies due to redundant computations. This paper introduces FlashDP, an innovative cache-friendly method that consolidates necessary operations into a single task, calculating gradients only once in a fused manner. This approach not only diminishes memory movement by up to $\\textbf{50}$% but also cuts down redundant computations by $\\textbf{20}$%, compared to previous methods. Consequently, FlashDP does not increase memory demands and achieves a $\\textbf{90}$% throughput compared to the Non-DP method on a four-A100 system during the pre-training of the Llama-13B model, while maintaining parity with standard DP-SGD in terms of precision. These advancements establish FlashDP as a pivotal development for efficient and privacy-preserving training of LLMs.", "keywords": [ "differential privacy", "DP-SGD", "LLMs", "CUDA" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 2, 1, 3, 4 ] }, { "paper_id": "cZWCjan02B", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Costin-Andrei Oncescu", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "", "position": "PhD student" }, { "name": "Sanket Purandare", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Sham M. Kakade", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Stratos Idreos", "gender": "unknown", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "US", "position": "Full Professor" } ], "n_formula": 44, "n_formula_1": 16, "n_ref_uni": 18, "n_ref": 82, "n_ref_all": 95, "n_fig": 7, "n_tab": 5, "L_tab": 4205, "n_element_tab": 586, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3801, "formula_len_all_1": 1793, "len_all": 181605, "len_all_1": 68965, "len_abs": 1645, "len_title": 139, "len_sents": 60785, "len_sents_1": 30409, "n_sents": 450, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1092, "title": "Flash Inference: Near Linear Time Inference for Long Convolution Sequence Models and Beyond", "abs": "While transformers have been at the core of most recent advancements in sequence generative models, their computational cost remains quadratic in sequence length.\nSeveral subquadratic architectures have been proposed to address this computational issue. Some of them, including long convolution sequence models (LCSMs), such as Hyena, address this issue at training time but remain quadratic during inference. We propose a method for speeding up LCSMs' exact inference to quasilinear time, identify the key properties that make this possible, and propose a general framework that exploits these. Our approach, inspired by previous work on relaxed polynomial interpolation, is based on a tiling which helps decrease memory movement and share computation. It has the added benefit of allowing for almost complete parallelization across layers of the position-mixing part of the architecture. Empirically, we provide a proof of concept implementation for Hyena, which gets up to $7.8\\times$ end-to-end improvement over standard inference by improving $110\\times$ within the position-mixing part.", "keywords": [ "computational efficiency", "inference", "sequence generative models", "long convolution sequence models", "hyena" ], "rating_list": [ 6, 8, 6, 8, 5 ], "soundness_list": [ 3, 4, 2, 4, 4 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 2, 4, 3, 4, 4 ] }, { "paper_id": "cZOPrf5WLu", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Derek Lim", "gender": "Male", "institution": "Liquid AI", "country": "US", "position": "Researcher" }, { "name": "Haggai Maron", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Assistant Professor" }, { "name": "Stefanie Jegelka", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Theo Putterman", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Yoav Gelberg", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 57, "n_formula_1": 18, "n_ref_uni": 59, "n_ref": 109, "n_ref_all": 135, "n_fig": 4, "n_tab": 11, "L_tab": 1916, "n_element_tab": 150, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 669, "n_element_tab_1": 59, "formula_len_all": 6656, "formula_len_all_1": 1757, "len_all": 205473, "len_all_1": 70088, "len_abs": 1445, "len_title": 145, "len_sents": 51778, "len_sents_1": 30100, "n_sents": 541, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1452, "title": "Learning on LoRAs: GL-Equivariant Processing of Low-Rank Weight Spaces for Large Finetuned Models", "abs": "Low-rank adaptations (LoRAs) have revolutionized the finetuning of large foundation models, enabling efficient adaptation even with limited computational resources. The resulting proliferation of LoRAs presents exciting opportunities for applying machine learning techniques that take these low-rank weights themselves as inputs. In this paper, we investigate the potential of Learning on LoRAs (LoL), a paradigm where LoRA weights serve as input to machine learning models. For instance, an LoL model that takes in LoRA weights as inputs could predict the performance of the finetuned model on downstream tasks, detect potentially harmful finetunes, or even generate novel model edits without traditional training methods. We first identify the inherent parameter symmetries of low rank decompositions of weights, which differ significantly from the parameter symmetries of standard neural networks. To efficiently process LoRA weights, we develop several symmetry-aware invariant or equivariant LoL models, using tools such as canonicalization, invariant featurization, and equivariant layers. We finetune thousands of text-to-image diffusion models and language models to collect datasets of LoRAs. In numerical experiments on these datasets, we show that our LoL architectures are capable of processing low rank weight decompositions to predict CLIP score, finetuning data attributes, finetuning data membership, and accuracy on downstream tasks.", "keywords": [ "LoRA", "Weight-space learning", "Foundation models", "Finetuning", "Equivariance" ], "rating_list": [ 3, 5, 8 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 1, 2, 4 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "cYB7GvpGj9", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Alireza Karami", "gender": "Male", "institution": "CEA", "country": "FR", "position": "Postdoc" }, { "name": "Nhut Truong", "gender": "unknown", "institution": "University of Trento", "country": "IT", "position": "PhD student" }, { "name": "shahryar noei", "gender": "Male", "institution": "Fondazione Bruno Kessler", "country": "IT", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 15, "n_ref": 51, "n_ref_all": 64, "n_fig": 5, "n_tab": 1, "L_tab": 565, "n_element_tab": 46, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 565, "n_element_tab_1": 46, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 77044, "len_all_1": 77044, "len_abs": 1099, "len_title": 115, "len_sents": 18533, "len_sents_1": 18533, "n_sents": 109, "n_sents_1": 109, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1104, "title": "Reassessing Number-Detector Units in Convolutional Neural Networks", "abs": "Convolutional neural networks (CNNs), including CORnet, have become essential models for predicting neural activity and behavior in visual tasks. However, their ability to capture complex cognitive functions, such as numerosity discrimination, remains a topic of debate. Numerosity\u2014the ability to perceive and estimate the number of items in a visual scene\u2014is believed to be represented by specialized 'number-detector' units within CNNs. In this study, we utilize CORnet, a specialized type of CNN inspired by brain anatomy, which also effectively captures the variance in human behavioral data, to address the limitations of classical representational similarity analysis (RSA), which assumes equal importance for all features. We apply pruning, a feature selection technique that identifies and retains the most behaviorally relevant units. Our results demonstrate that number-detector units are not critical for population-level representations of numerosity, challenging their proposed significance in previous studies. These results can have implications for both machine learning and neuroscience.", "keywords": [ "Feature Selection", "Pruning", "RSA", "CNN", "Numerosity" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "cXxfVkRCHJ", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Enze Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Shuai Li", "gender": "Female", "institution": "John Hopcroft Center, Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tong Yu", "gender": "unknown", "institution": "Adobe Research", "country": "US", "position": "Senior Research Scientist" }, { "name": "Xiao Huang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "MS student" }, { "name": "Xu Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 25, "n_ref": 63, "n_ref_all": 71, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 598, "formula_len_all_1": 597, "len_all": 85564, "len_all_1": 66343, "len_abs": 1304, "len_title": 130, "len_sents": 29942, "len_sents_1": 28743, "n_sents": 220, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1311, "title": "Offline-to-Online Reinforcement Learning with Classifier-Free Diffusion Generation", "abs": "Offline-to-online Reinforcement Learning (O2O RL) aims to perform online fine-tuning on an offline pre-trained policy to minimize costly online interactions. Existing methods have used offline data or online data to generate new data for data augmentation, which has led to performance improvement during online fine-tuning. However, they have not fully analyzed and utilized both types of data simultaneously. Offline data helps prevent agents from settling too early on suboptimal policies by providing diverse data, while online data improves training stability and speeds up convergence. In this paper, we propose a data augmentation approach, Classifier-Free Diffusion Generation (CFDG). Considering the differences between offline data and online data, we use conditional diffusion to generate both types of data for augmentation in the online phase, aiming to improve the quality of sample generation. Experimental results show that CFDG outperforms replaying the two data types or using a standard diffusion model to generate new data. Our method is versatile and can be integrated with existing offline-to-online RL algorithms. By implementing CFDG to popular methods IQL, PEX and APL, we achieve a notable 15% average improvement in empirical performance on the D4RL benchmark like MuJoCo and AntMaze.", "keywords": [ "offline-to-online reinforcement learning", "data augmentation", "diffusion models" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 1, 1, 1 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "cXdG5fhZ3w", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Ali Payani", "gender": "Male", "institution": "Cisco", "country": "US", "position": "Researcher" }, { "name": "Parinaz Naghizadeh", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Yifan Yang", "gender": "Male", "institution": "Ohio State University", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 39, "n_ref": 74, "n_ref_all": 82, "n_fig": 4, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 293, "formula_len_all_1": 226, "len_all": 128042, "len_all_1": 69218, "len_abs": 1344, "len_title": 118, "len_sents": 35985, "len_sents_1": 35189, "n_sents": 220, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1347, "title": "Enhancing Group Fairness in Federated Learning through Personalization", "abs": "Personalized Federated Learning (FL) algorithms collaboratively train customized models for each client, enhancing the accuracy of the learned models on the client's local data (e.g., by clustering similar clients, by fine-tuning models locally, or by imposing regularization terms). In this paper, we investigate the impact of such personalization techniques on the group fairness of the learned models, and show that personalization can also lead to improved (local) fairness as an unintended benefit. We begin by illustrating these benefits of personalization through numerical experiments comparing several classes of personalized FL algorithms against a baseline FedAvg algorithm, elaborating on the reasons behind improved fairness using personalized FL, and then providing analytical support. Motivated by these, we then show how to build on this (unintended) fairness benefit, by further integrating a fairness metric into the cluster-selection procedure of clustering-based personalized FL algorithms, and improve the fairness-accuracy trade-off attainable through them. Specifically, we propose two new fairness-aware federated clustering algorithms, Fair-FCA and Fair-FL+HC, extending the existing IFCA and FL+HC algorithms, and demonstrate their ability to strike a (tuneable) balance between accuracy and fairness at the client level.", "keywords": [ "Federated Learning", "Personalization", "Fairness" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "cWrqs2lwCJ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Allen Z. Ren", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Anirudha Majumdar", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "brian ichter", "gender": "unknown", "institution": "Google", "country": "", "position": "Research Scientist" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 25, "n_ref": 113, "n_ref_all": 138, "n_fig": 11, "n_tab": 3, "L_tab": 2659, "n_element_tab": 129, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 449, "n_element_tab_1": 50, "formula_len_all": 162, "formula_len_all_1": 101, "len_all": 123774, "len_all_1": 62797, "len_abs": 1258, "len_title": 134, "len_sents": 42970, "len_sents_1": 28592, "n_sents": 316, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1247, "title": "Thinking Forward and Backward: Effective Backward Planning with Large Language Models", "abs": "Large language models (LLMs) have exhibited remarkable reasoning and planning capabilities. Most prior work in this area has used LLMs to reason through steps from an initial to a goal state or criterion, thereby effectively reasoning in a forward direction. Nonetheless, many planning problems exhibit an inherent asymmetry such that planning backward from the goal is significantly easier --- for example, if there are bottlenecks close to the goal. We take inspiration from this observation and demonstrate that this bias holds for LLM planning as well: planning performance in one direction correlates with the planning complexity of the problem in that direction. However, our experiments also reveal systematic biases which lead to poor planning in the backward direction. With this knowledge, we propose a backward planning algorithm for LLMs that first flips the problem and then plans forward in the flipped problem. This helps avoid the backward bias, generate more diverse candidate plans, and exploit asymmetries between the forward and backward directions in planning problems --- we find that combining planning in both directions with self-verification improves the overall planning success rates by 4-24% in three planning domains.", "keywords": [ "LLM planning", "backward search" ], "rating_list": [ 3, 1, 5, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "cWfpt2t37q", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Eric Moulines", "gender": "Male", "institution": "Ecole polytechnique", "country": "", "position": "Full Professor" }, { "name": "Maxim Panov", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Vladimir Kondratyev", "gender": "Male", "institution": "Aqemia", "country": "FR", "position": "Researcher" }, { "name": "Martin Takac", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Associate Professor" }, { "name": "Nikita Yurevich Kotelevskii", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" } ], "n_formula": 63, "n_formula_1": 23, "n_ref_uni": 38, "n_ref": 118, "n_ref_all": 164, "n_fig": 6, "n_tab": 12, "L_tab": 12414, "n_element_tab": 649, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 3035, "n_element_tab_1": 120, "formula_len_all": 9574, "formula_len_all_1": 1526, "len_all": 216832, "len_all_1": 71615, "len_abs": 810, "len_title": 92, "len_sents": 60552, "len_sents_1": 29948, "n_sents": 571, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 836, "title": "From Risk to Uncertainty: Generating Predictive Uncertainty Measures via Bayesian Estimation", "abs": "There are various measures of predictive uncertainty in the literature, but their relationships to each other remain unclear. This paper uses a decomposition of statistical pointwise risk into components associated with different sources of predictive uncertainty: namely, aleatoric uncertainty (inherent data variability) and epistemic uncertainty (model-related uncertainty). Together with Bayesian methods applied as approximations, we build a framework that allows one to generate different predictive uncertainty measures.\n\nWe validate measures, derived from our framework on image datasets by evaluating its performance in detecting out-of-distribution and misclassified instances using the AUROC metric. The experimental results confirm that the measures derived from our framework are useful for the considered downstream tasks.", "keywords": [ "Uncertainty quantification", "Bayesian methods", "Statistics" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "cWHonXThtM", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Bingyi Jing", "gender": "Male", "institution": "South University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Hanting Chen", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Jie Hu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Shaohui Lin", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Researcher" }, { "name": "Simiao Li", "gender": "Female", "institution": "Huawei Noah's Ark Lab ", "country": "CN", "position": "Researcher" }, { "name": "Wei Li", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Wenjia Wang", "gender": "Male", "institution": "HKUST (GZ)", "country": "CN", "position": "Assistant Professor" }, { "name": "Yun Zhang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 42, "n_ref": 98, "n_ref_all": 121, "n_fig": 2, "n_tab": 12, "L_tab": 6535, "n_element_tab": 542, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 5418, "n_element_tab_1": 479, "formula_len_all": 566, "formula_len_all_1": 565, "len_all": 124960, "len_all_1": 68097, "len_abs": 919, "len_title": 138, "len_sents": 26760, "len_sents_1": 26449, "n_sents": 177, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 90, "L_abs": 923, "title": "Knowledge Distillation with Multi-granularity Mixture of Priors for Image Super-Resolution", "abs": "Knowledge distillation (KD) is a promising yet challenging model compression technique that transfers rich learning representations from a well-performing but cumbersome teacher model to a compact student model. Previous methods for image super-resolution (SR) mostly are tailored to the specific teacher-student architectures. And the potential for improvement is limited, which hinders their wide applications. This work presents a novel KD framework for SR models, the multi-granularity mixture of prior knowledge distillation (MiPKD), that is universally applicable to a wide array of architectures at feature and block levels. The teacher\u2019s knowledge is effectively integrated with the student's feature via the Feature Prior Mixer, and the reconstructed feature propagates dynamically in the training phase with the Block Prior Mixer. Extensive experiments demonstrate the effectiveness of the proposed MiPKD method.", "keywords": [ "Image Super-Resolution", "Knowledge Distillation", "Model Compression" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "cWGCkd7mCp", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Cameron Gordon", "gender": "unknown", "institution": "The University of Adelaide", "country": "AU", "position": "PhD student" }, { "name": "Hemanth Saratchandran", "gender": "unknown", "institution": "University of Adelaide/Australian Institute of Machine Learning", "country": "AU", "position": "Postdoc" }, { "name": "Simon Lucey", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Full Professor" }, { "name": "Yiping Ji", "gender": "Male", "institution": "CSIRO", "country": "AU", "position": "Researcher" }, { "name": "Zeyu Zhang", "gender": "Male", "institution": "The Australian National University", "country": "AU", "position": "Undergrad student" } ], "n_formula": 31, "n_formula_1": 6, "n_ref_uni": 51, "n_ref": 98, "n_ref_all": 121, "n_fig": 8, "n_tab": 7, "L_tab": 3052, "n_element_tab": 181, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1728, "n_element_tab_1": 83, "formula_len_all": 1319, "formula_len_all_1": 339, "len_all": 157241, "len_all_1": 56432, "len_abs": 1073, "len_title": 104, "len_sents": 41406, "len_sents_1": 25035, "n_sents": 322, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1076, "title": "Efficient Learning with Sine-Activated Low-Rank Matrices", "abs": "Low-rank decomposition has emerged as a vital tool for enhancing parameter efficiency in neural network architectures, gaining traction across diverse applications in machine learning. These techniques significantly lower the number of parameters, striking a balance between compactness and performance. However, a common challenge has been the compromise between parameter efficiency and the accuracy of the model, where reduced parameters often lead to diminished accuracy compared to their full-rank counterparts. In this work, we propose a novel theoretical framework that integrates a sinusoidal function within the low-rank decomposition process. This approach not only preserves the benefits of the parameter efficiency characteristic of low-rank methods but also increases the decomposition's rank, thereby enhancing model performance. Our method proves to be a plug in enhancement for existing low-rank models, as evidenced by its successful application in Vision Transformers (ViT), Large Language Models (LLMs), Neural Radiance Fields (NeRF) and 3D shape modelling.", "keywords": [ "parameter efficient learning", "low rank matrices" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "cWEfRkYj46", "primary_area": "applications to neuroscience & cognitive science", "n_author": 7, "author_info_list": [ { "name": "Chen Feng", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Di Wu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Jie Yang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Researcher" }, { "name": "Lu Cao", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "PhD student" }, { "name": "Mohamad Sawan", "gender": "Male", "institution": "Polytechnique Montr\u00e9al, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Emeritus" }, { "name": "Siyuan Li", "gender": "Male", "institution": "Westlake University & Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yue Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 17, "n_ref_uni": 51, "n_ref": 81, "n_ref_all": 97, "n_fig": 5, "n_tab": 8, "L_tab": 2800, "n_element_tab": 737, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1223, "n_element_tab_1": 113, "formula_len_all": 1530, "formula_len_all_1": 730, "len_all": 171888, "len_all_1": 69586, "len_abs": 1348, "len_title": 134, "len_sents": 50374, "len_sents_1": 32953, "n_sents": 354, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 84, "L_abs": 1370, "title": "Towards Homogeneous Lexical Tone Decoding from Heterogeneous Intracranial Recordings", "abs": "Recent advancements in brain-computer interfaces (BCIs) and deep learning have made decoding lexical tones from intracranial recordings possible, providing the potential to restore the communication ability of speech-impaired tonal language speakers. However, data heterogeneity induced by both physiological and instrumental factors poses a significant challenge for unified invasive brain tone decoding. Particularly, the existing heterogeneous decoding paradigm (training subject-specific models with individual data) suffers from the intrinsic limitation that fails to learn generalized neural representations and leverages data across subjects. To this end, we introduce Homogeneity-Heterogeneity Disentangled Learning for Neural Representations (H2DiLR), a framework that disentangles and learns the homogeneity and heterogeneity from intracranial recordings of multiple subjects. To verify the effectiveness of H2DiLR, we collected stereoelectroencephalography (sEEG) from multiple participants reading Mandarin materials containing 407 syllables (covering nearly all Mandarin characters). Extensive experiments demonstrate that H2DiLR, as a unified decoding paradigm, outperforms the naive heterogeneous decoding paradigm by a large margin. We also empirically show that H2DiLR indeed captures homogeneity and heterogeneity during neural representation learning.", "keywords": [ "brain-computer interfaces; speech decoding; tonal language; Homogeneity-Heterogeneity Disentanglement" ], "rating_list": [ 8, 3, 8, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 5, 4, 5 ] }, { "paper_id": "cVyELMpMRS", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Gokul Swamy", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Jason D. Lee", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Jonathan Daniel Chang", "gender": "unknown", "institution": "Cornell University", "country": "", "position": "PhD student" }, { "name": "Wen Sun", "gender": "unknown", "institution": "Cornell University", "country": "", "position": "Assistant Professor" }, { "name": "Wenhao Zhan", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Zhaolin Gao", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "kiante brantley", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 55, "n_formula_1": 27, "n_ref_uni": 58, "n_ref": 142, "n_ref_all": 155, "n_fig": 1, "n_tab": 6, "L_tab": 1090, "n_element_tab": 85, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 530, "n_element_tab_1": 86, "formula_len_all": 7096, "formula_len_all_1": 2826, "len_all": 280296, "len_all_1": 72231, "len_abs": 1602, "len_title": 129, "len_sents": 91867, "len_sents_1": 33246, "n_sents": 680, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1617, "title": "Regressing the Relative Future: Efficient Policy Optimization for Multi-turn RLHF", "abs": "Large Language Models (LLMs) have achieved remarkable success at tasks like summarization that involve a *single turn* of interaction. However, they can still struggle with *multi-turn* tasks like dialogue that require long-term planning. Previous works on multi-turn dialogue extend single-turn reinforcement learning from human feedback (RLHF) methods to the multi-turn setting by treating all prior dialogue turns as a long context. Such approaches suffer from *covariate shift*: the conversations in the training set have previous turns generated by some reference policy, which means that low training error may not necessarily correspond to good performance when the learner is actually in the conversation loop. In response, we introduce REgressing the RELative FUture (REFUEL), an efficient policy optimization approach designed to address multi-turn RLHF in LLMs. REFUEL employs a single model to estimate Q-values and trains on self-generated data, addressing the covariate shift issue. REFUEL frames the multi-turn RLHF problem as a sequence of regression tasks on iteratively collected datasets, enabling ease of implementation. Theoretically, we prove that REFUEL can match the performance of any policy covered by the training set. Empirically, we evaluate our algorithm by using Llama-3.1-70B-it to simulate a user in conversation with our model. REFUEL consistently outperforms state-of-the-art methods such as DPO and REBEL across various settings. Furthermore, despite having only 8 billion parameters, Llama-3-8B-it fine-tuned with REFUEL, outperforms Llama-3.1-70B-it on long multi-turn dialogues.", "keywords": [ "Reinforcement Learning", "Reinforcement Learning from Human Feedback" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 4, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 3, 2, 2, 4 ] }, { "paper_id": "cVgOIjcNoQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 14, "author_info_list": [ { "name": "Bai Jionghao", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Undergrad student" }, { "name": "Dongjie Fu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Jingyu Lu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Linjun Li", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Minghui Fang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Rongjie Huang", "gender": "Male", "institution": "FAIR", "country": "US", "position": "Researcher" }, { "name": "Shengpeng Ji", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Tao Jin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaoda Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Xize Cheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yu Chen", "gender": "unknown", "institution": "Meituan", "country": "CN", "position": "Technical Director" }, { "name": "Zehan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Hu Ruofan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 37, "n_ref": 63, "n_ref_all": 87, "n_fig": 13, "n_tab": 3, "L_tab": 590, "n_element_tab": 45, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 226, "n_element_tab_1": 11, "formula_len_all": 751, "formula_len_all_1": 649, "len_all": 141881, "len_all_1": 61538, "len_abs": 1346, "len_title": 144, "len_sents": 43951, "len_sents_1": 29665, "n_sents": 335, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1367, "title": "OmniChat: Enhancing Spoken Dialogue Systems with Scalable Synthetic Data for Diverse Scenarios", "abs": "With the rapid development of large language models, researchers have created increasingly advanced spoken dialogue systems that can naturally converse with humans. However, these systems still struggle to handle the full complexity of real-world conversations, including audio events, musical contexts, and emotional expressions, mainly because current dialogue datasets are constrained in both scale and scenario diversity. In this paper, we propose leveraging synthetic data to enhance the dialogue models across diverse scenarios. We introduce **ShareChatX**, the first comprehensive, large-scale dataset for spoken dialogue that spans diverse scenarios. Based on this dataset, we introduce **OmniChat**, a multi-turn dialogue system with a heterogeneous feature fusion module, designed to optimize feature selection in different dialogue contexts. In addition, we explored critical aspects of training dialogue systems using synthetic data. Through comprehensive experimentation, we determined the ideal balance between synthetic and real data, achieving state-of-the-art results on the real-world dialogue dataset DailyTalk. We also highlight the crucial importance of synthetic data in tackling diverse, complex dialogue scenarios, especially those involving audio and music. For more details, please visit our demo page at \\url{https://sharechatx.github.io/}.", "keywords": [ "Spoken Dialogue System", "Synthetic Data", "Multi-modal Large Language Model" ], "rating_list": [ 3, 8, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "cUuOKnjVQJ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Amit Sharma", "gender": "Male", "institution": "Microsoft Research", "country": "IN", "position": "Principal Researcher" }, { "name": "Arno Solin", "gender": "unknown", "institution": "Aalto University", "country": "FI", "position": "Assistant Professor" }, { "name": "Gaurav Sinha", "gender": "Male", "institution": "Research, Microsoft", "country": "IN", "position": "Principal Researcher" }, { "name": "Nagarajan Natarajan", "gender": "Male", "institution": "Microsoft Research", "country": "", "position": "Researcher" }, { "name": "Prakhar Verma", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "PhD student" }, { "name": "Sukruta Prakash Midigeshi", "gender": "Female", "institution": "Indian Institute of Technology, Gandhinagar", "country": "", "position": "Undergrad student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 39, "n_ref": 67, "n_ref_all": 91, "n_fig": 4, "n_tab": 5, "L_tab": 1398, "n_element_tab": 67, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2845, "n_element_tab_1": 115, "formula_len_all": 257, "formula_len_all_1": 80, "len_all": 162130, "len_all_1": 70744, "len_abs": 1009, "len_title": 104, "len_sents": 51881, "len_sents_1": 32841, "n_sents": 419, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1023, "title": "Plan-RAG: Planning-guided Retrieval Augmented Generation", "abs": "We introduce Planning-guided Retrieval Augmented Generation (Plan-RAG), a novel framework that augments the retrieve-then-reason paradigm of existing RAG frameworks to plan-then-retrieve. Plan-RAG formulates a reasoning plan as a directed acyclic graph (DAG), decomposing queries into interrelated atomic sub-queries. Answer generation follows the DAG structure, allowing significant gains in efficiency through parallelized retrieval and generation. While state-of-the-art RAG solutions require extensive data generation and fine-tuning of language models (LMs), Plan-RAG incorporates frozen LMs as plug-and-play experts to generate high-quality answers. Compared to existing RAG solutions, Plan-RAG demonstrates significant improvements in reducing hallucinations and bolstering attribution due to its structured sub-query decomposition. Plan-RAG offers a new perspective on integrating external knowledge in LMs while ensuring attribution by design, contributing towards more reliable and interpretable LM-based systems.", "keywords": [ "Language Models", "Retrieval Augmented Generation", "LLM", "RAG" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "cUnqwFu5OO", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Kaito Ariu", "gender": "Male", "institution": "CyberAgent, Inc.", "country": "JP", "position": "Research Scientist" }, { "name": "Kenshi Abe", "gender": "Male", "institution": "University of Electro-Communications", "country": "", "position": "PhD student" }, { "name": "Noboru Isobe", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" } ], "n_formula": 87, "n_formula_1": 27, "n_ref_uni": 35, "n_ref": 72, "n_ref_all": 81, "n_fig": 4, "n_tab": 4, "L_tab": 296, "n_element_tab": 113, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 138, "n_element_tab_1": 91, "formula_len_all": 15874, "formula_len_all_1": 3111, "len_all": 198242, "len_all_1": 67963, "len_abs": 951, "len_title": 101, "len_sents": 36844, "len_sents_1": 22267, "n_sents": 379, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1058, "title": "Last Iterate Convergence in Monotone Mean Field Games", "abs": "Mean Field Game (MFG) is a framework utilized to model and approximate the behavior of a large number of agents, and the computation of equilibria in MFG has been a subject of interest. Despite the proposal of methods to approximate the equilibria, algorithms that can achieve equilibrium with the most recent policy of the algorithm, namely the last-iterate policy, have been limited. \nWe propose the use of a simple, proximal-point-type algorithm to compute strategies for MFGs. Subsequently, we provide the first last-iterate convergence guarantee under the Lasry--Lions-type monotonicity condition. \nWe further employ the Mirror Descent algorithm for the regularized MFG to efficiently approximate the update rules of the proximal point method for MFGs. \nWe demonstrate that the last-iterate strategy of Mirror Descent converges exponentially fast: we provide the guarantee of computing the $\\varepsilon$ approximation in $\\mathcal{O}(\\log(1/\\varepsilon))$ iterations. This research offers a tractable approach for large-scale and large-population games.", "keywords": [ "mean field game", "learning in games" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "cUeYEwc237", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Deepayan Sanyal", "gender": "Male", "institution": "Vanderbilt University", "country": "", "position": "PhD student" }, { "name": "Joel Phillips Michelson", "gender": "Male", "institution": "Vanderbilt University", "country": "", "position": "PhD student" }, { "name": "Maithilee Kunda", "gender": "Female", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Reader" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 14, "n_ref": 19, "n_ref_all": 23, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 74397, "len_all_1": 60480, "len_abs": 1943, "len_title": 129, "len_sents": 32064, "len_sents_1": 31270, "n_sents": 223, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 732, "title": "Feature-Based Analysis of Theory of Mind Representations in Neural Network Models", "abs": "Theory of Mind (ToM) presents a significant generalization challenge in computational modeling. This paper explores how neural networks with varying architectures and training regimes learn and represent ToM-related features. We introduce a novel method for quantifying feature representation within neural networks and apply it to a set of theoretically-grounded features designed to differentiate between hypothesized ToM strategies. We examine the relationship between feature representation and task accuracy across different model architectures and training datasets. This work provides insights into the mechanisms underlying ToM capabilities in neural networks and offers a framework for future research in computational ToM.", "keywords": [ "theory of mind", "computational modeling", "social cognition" ], "rating_list": [ 3, 1, 3, 1 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 1, 1, 2, 1 ], "contribution_list": [ 1, 1, 1, 1 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "cUN8lJB4rD", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Alexander Tyurin", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Assistant Professor" } ], "n_formula": 245, "n_formula_1": 44, "n_ref_uni": 38, "n_ref": 114, "n_ref_all": 158, "n_fig": 3, "n_tab": 5, "L_tab": 713, "n_element_tab": 38, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 649, "n_element_tab_1": 112, "formula_len_all": 21206, "formula_len_all_1": 4047, "len_all": 263987, "len_all_1": 71896, "len_abs": 845, "len_title": 144, "len_sents": 74148, "len_sents_1": 26997, "n_sents": 862, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 846, "title": "Tight Time Complexities in Parallel Stochastic Optimization with Arbitrary Computation Dynamics", "abs": "In distributed stochastic optimization, where parallel and asynchronous methods are employed, we establish optimal time complexities under virtually any computation behavior of workers/devices/CPUs/GPUs, capturing potential disconnections due to hardware and network delays, time-varying computation powers, and any possible fluctuations and trends of computation speeds. These real-world scenarios are formalized by our new universal computation model. Leveraging this model and new proof techniques, we discover tight lower bounds that apply to virtually all synchronous and asynchronous methods, including Minibatch SGD, Asynchronous SGD (Recht et al., 2011), and Picky SGD (Cohen et al., 2021). We show that these lower bounds, up to constant factors, are matched by the optimal Rennala SGD and Malenia SGD methods (Tyurin & Richt\u00e1rik, 2023).", "keywords": [ "nonconvex optimization", "lower bounds", "parallel methods", "asynchronous methods", "convex optimization" ], "rating_list": [ 5, 5, 8, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "cUFIil6hEG", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Abhinav Moudgil", "gender": "Male", "institution": "Concordia University", "country": "CA", "position": "PhD student" }, { "name": "Boris Knyazev", "gender": "Male", "institution": "Samsung", "country": "CA", "position": "Researcher" }, { "name": "Eugene Belilovsky", "gender": "Male", "institution": "Concordia University, Montreal", "country": "CA", "position": "Assistant Professor" }, { "name": "Guillaume Lajoie", "gender": "Male", "institution": "Mila, Quebec AI institute", "country": "", "position": "Core Academic Member" }, { "name": "Simon Lacoste-Julien", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "Associate Professor" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 31, "n_ref": 66, "n_ref_all": 127, "n_fig": 27, "n_tab": 8, "L_tab": 3146, "n_element_tab": 380, "n_fig_1": 20, "n_tab_1": 6, "L_tab_1": 2118, "n_element_tab_1": 281, "formula_len_all": 1344, "formula_len_all_1": 797, "len_all": 170360, "len_all_1": 77809, "len_abs": 3083, "len_title": 117, "len_sents": 46121, "len_sents_1": 30298, "n_sents": 332, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 934, "title": "Accelerating Training with Neuron Interaction and Nowcasting Networks", "abs": "Neural network training can be accelerated when a learnable update rule is used in lieu of classic adaptive optimizers (e.g. Adam). However, learnable update rules can be costly and unstable to train and use. Recently, Jang et al. (2023) proposed a simpler approach to accelerate training based on weight nowcaster networks (WNNs). In their approach, Adam is used for most of the optimization steps and periodically, only every few steps, a WNN nowcasts (predicts near future) parameters. We improve WNNs by proposing neuron interaction and nowcasting (NiNo) networks. In contrast to WNNs, NiNo leverages neuron connectivity and graph neural networks to more accurately nowcast parameters. We further show that in some networks, such as Transformers, modeling neuron connectivity accurately is challenging. We address this and other limitations, which allows NiNo to accelerate Adam training by up to 50% in vision and language tasks.", "keywords": [ "accelerated optimization", "parameter prediction", "graphs", "transformers", "learning to optimize" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "cU6ZdN87p3", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Chong Xiang", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Danqi Chen", "gender": "Female", "institution": "Department of Computer Science, Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "David Wagner", "gender": "unknown", "institution": "University of California Berkeley", "country": "US", "position": "Professor" }, { "name": "Prateek Mittal", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "Full Professor" }, { "name": "Tong Wu", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Zexuan Zhong", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 8, "n_ref_uni": 27, "n_ref": 73, "n_ref_all": 125, "n_fig": 21, "n_tab": 10, "L_tab": 3304, "n_element_tab": 306, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 80, "n_element_tab_1": 8, "formula_len_all": 3648, "formula_len_all_1": 900, "len_all": 239679, "len_all_1": 71823, "len_abs": 1636, "len_title": 108, "len_sents": 81142, "len_sents_1": 33492, "n_sents": 686, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1119, "title": "Certifiably Robust RAG against Retrieval Corruption Attacks", "abs": "Retrieval-augmented generation (RAG) has been shown vulnerable to retrieval corruption attacks: an attacker can inject malicious passages into retrieval results to induce inaccurate responses. In this paper, we propose RobustRAG as the first defense framework against retrieval corruption attacks. The key insight of RobustRAG is an isolate-then-aggregate strategy: we isolate passages into disjoint groups, generate LLM responses based on the concatenated passages from each isolated group, and then securely aggregate these responses for a robust output. To instantiate RobustRAG, we design keyword-based and decoding-based algorithms for securely aggregating unstructured text responses. Notably, RobustRAG can achieve certifiable robustness: we can formally prove and certify that, for certain queries, RobustRAG can always return accurate responses, even when an adaptive attacker has full knowledge of our defense and can arbitrarily inject a small number of malicious passages. We evaluate RobustRAG on open-domain QA and long-form text generation datasets and demonstrate its effectiveness and generalizability.", "keywords": [ "Retrieval-augemented Generation", "Retrieval Corruption Attack", "Certifiable Robustness" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "cTR17xl89h", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Haofei Lu", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Jianshu Li", "gender": "Male", "institution": "Ant Group", "country": "SG", "position": "Researcher" }, { "name": "Junliang Xing", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Ruoyu Li", "gender": "Male", "institution": "Tiktok", "country": "SG", "position": "Researcher" }, { "name": "Yuanchun Shi", "gender": "Female", "institution": ", Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Zhe Li", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "algorithm expert" }, { "name": "Zhe Wu", "gender": "unknown", "institution": "Qiyuan laboratory", "country": "", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 40, "n_ref": 102, "n_ref_all": 136, "n_fig": 10, "n_tab": 4, "L_tab": 2313, "n_element_tab": 139, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 903, "n_element_tab_1": 58, "formula_len_all": 1506, "formula_len_all_1": 1307, "len_all": 147568, "len_all_1": 61841, "len_abs": 1114, "len_title": 105, "len_sents": 45673, "len_sents_1": 28083, "n_sents": 341, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 57, "L_abs": 1073, "title": "BodyGen: Advancing Towards Efficient Embodiment Co-Design", "abs": "Embodiment co-design aims to optimize a robot's morphology and control policy simultaneously. \nWhile prior work has demonstrated its potential for generating environment-adaptive robots, this field still faces persistent challenges in optimization efficiency due to the (i) combinatorial nature of morphological search spaces and (ii) intricate dependencies between morphology and control.\nWe prove that the ineffective morphology representation and unbalanced reward signals between the design and control stages are key obstacles to efficiency.\nTo advance towards efficient embodiment co-design, we propose **BodyGen**, which utilizes (1) topology-aware self-attention for both design and control, enabling efficient morphology representation with lightweight model sizes; (2) a temporal credit assignment mechanism that ensures balanced reward signals for optimization. With our findings, BodyGen achieves an average **60.03%** performance improvement against state-of-the-art baselines. We provide codes and more results on the website: https://genesisorigin.github.io.", "keywords": [ "Reinforcement Learning" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 4, 4, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "cTG25RXtJA", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Andong Tan", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Hao Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Luyang Luo", "gender": "Male", "institution": "Harvard Medical School, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Yequan Bie", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zhixuan CHEN", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 51, "n_ref": 120, "n_ref_all": 137, "n_fig": 9, "n_tab": 7, "L_tab": 1926, "n_element_tab": 172, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2595, "n_element_tab_1": 172, "formula_len_all": 491, "formula_len_all_1": 414, "len_all": 156819, "len_all_1": 57890, "len_abs": 1909, "len_title": 142, "len_sents": 41141, "len_sents_1": 27336, "n_sents": 254, "n_sents_1": 145, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1654, "title": "KEEP: Towards a Knowledge-Enhanced Explainable Prompting Framework for Vision-Language Models", "abs": "Large-scale vision-language models (VLMs) embedded with expansive representations and visual concepts have showcased significant potential in the computer vision community. Efficiently adapting VLMs such as CLIP, to downstream tasks has garnered growing attention, with prompt learning emerging as a representative approach. However, most existing prompt-based adaptation methods, which rely solely on coarse-grained textual prompts, suffer from limited performance and interpretability when handling tasks that require domain-specific knowledge. This results in a failure to satisfy the stringent trustworthiness requirements of Explainable Artificial Intelligence (XAI) in high-risk scenarios like healthcare. To address this issue, we propose a Knowledge-Enhanced Explainable Prompting (KEEP) framework that leverages fine-grained domain-specific knowledge to enhance the adaptation process across various domains, facilitating bridging the gap between the general domain and other specific domains. We present to our best knowledge the first work to incorporate retrieval augmented generation and domain-specific foundation models to provide more reliable image-wise knowledge for prompt learning in various domains, alleviating the lack of fine-grained annotations, while offering both visual and textual explanations. Extensive experiments and explainability analyses conducted on eight datasets of different domains, demonstrate that our method simultaneously achieves superior performance and interpretability, shedding light on the effectiveness of the collaboration between foundation models and XAI. The code will be made publically available.", "keywords": [ "Prompt", "Domain Knowledge", "VLM", "XAI" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "cTDooc2J9S", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Abhishek Saroha", "gender": "Male", "institution": "Department of Informatics, Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Christian Koke", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Daniel Cremers", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Full Professor" }, { "name": "Marvin Eisenberger", "gender": "Male", "institution": "Technical University Munich", "country": "", "position": "PhD student" }, { "name": "Michael M. Bronstein", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Yuesong Shen", "gender": "Male", "institution": "Technical University Munich", "country": "", "position": "PhD student" } ], "n_formula": 120, "n_formula_1": 7, "n_ref_uni": 45, "n_ref": 116, "n_ref_all": 194, "n_fig": 20, "n_tab": 9, "L_tab": 6336, "n_element_tab": 411, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 1094, "n_element_tab_1": 31, "formula_len_all": 11812, "formula_len_all_1": 387, "len_all": 364957, "len_all_1": 78182, "len_abs": 987, "len_title": 124, "len_sents": 115510, "len_sents_1": 34204, "n_sents": 1101, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 991, "title": "Laplace-Transform-Filters render spectral Graph Neural Networks transferable", "abs": "We introduce a new point of view on transferability of graph neural networks based on the intrinsic notion of information diffusion within graphs. This notion is adapted to considering graphs to be similar if their overall rough structures are similar, while their fine-print articulation may differ. Transferability of graph neural networks is then considered between graphs that are similar from this novel perspective on transferability. After carefully analysing transferability of single filters, the transferability properties of entire networks are relegated to the transferability characteristics of the filters employed inside their convolutional blocks. A rigorous analysis establishes our main theoretical finding: Spectral convolutional networks are transferable between graphs whose overall rough structures align, if their filters arise as Laplace transforms of certain generalized functions. Numerical experiments illustrate and validate the theoretical findings in practice.", "keywords": [ "Graph Neural Networks", "Spectral Graph Theory", "Transferability" ], "rating_list": [ 5, 6, 3, 6, 3 ], "soundness_list": [ 2, 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 1 ], "confidence_list": [ 4, 3, 5, 2, 5 ] }, { "paper_id": "cSnbM9SIJJ", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 9, "author_info_list": [ { "name": "Bolin Ding", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Senior Director" }, { "name": "Dawei Gao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Jingren Zhou", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xuchen Pan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yaliang Li", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Staff Engineer" }, { "name": "Yuexiang Xie", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Staff" }, { "name": "Yushuo Chen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "MS student" }, { "name": "Zhewei Wei", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 58, "n_ref_all": 89, "n_fig": 23, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 233159, "len_all_1": 70702, "len_abs": 2139, "len_title": 111, "len_sents": 87915, "len_sents_1": 37485, "n_sents": 626, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1607, "title": "Very Large-Scale Multi-Agent Simulation with LLM-Powered Agents", "abs": "Recent advances in large language models (LLMs) have opened new avenues for applying multi-agent systems in very large-scale simulations. However, there remain several challenges when conducting multi-agent simulations with existing platforms, such as limited scalability and low efficiency, unsatisfied agent diversity, and effort-intensive management processes. To address these challenges, we develop several new features and components based on a user-friendly multi-agent platform, enhancing its convenience and flexibility for supporting very large-scale multi-agent simulations. Specifically, we propose an actor-based distributed mechanism as the underlying technological infrastructure towards great scalability and high efficiency, and provide flexible environment support for simulating various real-world scenarios, which enables parallel execution of multiple agents, automatic workflow conversion for distributed deployment, and both inter-agent and agent-environment interactions. Moreover, we develop an easy-to-use configurable tool and an automatic background generation pipeline, simplifying the process of creating agents with diverse yet detailed background settings. Last but not least, we provide a web-based interface for conveniently monitoring and managing a large number of agents that might deploy across multiple devices. We conduct a comprehensive simulation to demonstrate the effectiveness of these proposed enhancements, and provide detailed observations and insightful discussions to highlight the great potential of applying multi-agent systems in large-scale simulations.", "keywords": [ "Multi-agent system", "Large-scale", "Simulation" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 3, 1, 3 ], "presentation_list": [ 4, 1, 2, 3 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "cSgEW7EZ9h", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Hongshu Guo", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yue-Jiao Gong", "gender": "Female", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zeyuan Ma", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Zhiguang Cao", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Zhou Jiang", "gender": "unknown", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 30, "n_formula_1": 9, "n_ref_uni": 51, "n_ref": 97, "n_ref_all": 131, "n_fig": 3, "n_tab": 4, "L_tab": 2974, "n_element_tab": 191, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1959, "n_element_tab_1": 195, "formula_len_all": 2478, "formula_len_all_1": 784, "len_all": 181628, "len_all_1": 78017, "len_abs": 1711, "len_title": 137, "len_sents": 54692, "len_sents_1": 34935, "n_sents": 407, "n_sents_1": 240, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1717, "title": "Meta-Black-Box-Optimization through Offline Q-function Learning with Mamba Architecture", "abs": "Recent progress in Meta-Black-Box-Optimization (MetaBBO) has demonstrated that meta-training a neural network based meta-level control policy over an optimization task distribution could significantly enhance the optimization performance of the low-level black-box optimizers. However, achieving such performance enhancement requires effective policy optimization/search method to locate optimal control policy within a massive joint-action space. The online learning fashion of existing works further makes the efficiency of MetaBBO problematic. To address these technical challenges, we propose an offline learning framework in this paper, termed Q-Mamba. Concretely, our method uses a Mamba neural network architecture to meta-learn decomposed Q-functions for each configurable component in the low-level optimizer. By decomposing the Q-function of the configuration decisions of all components in an optimizer, we can apply effective sequence modelling to avoid searching the control policy in the massive joint-action space. Furthermore, by leveraging the long-sequence modelling advantage of Mamba and moderate offline trajectory samples, Q-Mamba can be efficiently trained through a synergy of offline Temporal-Difference update and Conservative Q-Learning regularization to achieve competitive performance against the online learning paradigms. Through extensive benchmarking, we observe that Q-Mamba achieves competitive or even superior optimization performance to prior online/offline learning baselines, while significantly improving the training efficiency of existing online learning baselines. Additional ablation studies show that each of the proposed key designs contributes to this good performance.", "keywords": [ "Black-Box Optimization", "Dynamic Algorithm Configuration", "Learning to Optimize", "Offline Reinforcement Learning", "Mamba" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 4, 2, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "cSd8Eom8Zt", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Lior Katz", "gender": "unknown", "institution": "Applied Materials", "country": "IL", "position": "Principal Researcher" }, { "name": "Tomer Melamed", "gender": "unknown", "institution": "Applied Materials", "country": "IL", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 24, "n_ref_uni": 18, "n_ref": 29, "n_ref_all": 48, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 447, "n_element_tab_1": 31, "formula_len_all": 900, "formula_len_all_1": 868, "len_all": 83382, "len_all_1": 54373, "len_abs": 914, "len_title": 121, "len_sents": 25160, "len_sents_1": 19277, "n_sents": 274, "n_sents_1": 155, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 916, "title": "Reshaping Model Output Space Via Deep Kernel Density Estimation Networks", "abs": "Traditional classification models are typically optimized solely for their specific training task without considering the properties of the underlying probability distribution of their output space. As the use of these models for downstream tasks becomes more prevalent, it becomes advantageous to have a framework that can transform the output space of such models to a more convenient space without\nsacrificing performance. In this paper, we introduce DeepKDE, a novel method which enables the transformation of arbitrary output spaces to match more desirable distributions, such as Normal and Gaussian Mixture Models. We explore the properties of the new method and test its effectiveness on ResNet-18 and vision transformers trained on CIFAR-10 and Fashion MNIST datasets. We show that DeepKDE models succeed in transforming the output spaces of the original models while outperforming them in terms of accuracy.", "keywords": [ "Deep KDE", "probability density transformations", "Kernel Density Estimation" ], "rating_list": [ 3, 1, 3 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 1, 1, 2 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "cSHBZ4U9eO", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Defu Cao", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Lun Du", "gender": "Male", "institution": "Ant Research", "country": "CN", "position": "Researcher" }, { "name": "Qiang Fu", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Yan Liu", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "Professor" }, { "name": "Yizhou Zhang", "gender": "unknown", "institution": "University of Southern California", "country": "", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 4, "n_ref_uni": 39, "n_ref": 82, "n_ref_all": 90, "n_fig": 4, "n_tab": 2, "L_tab": 474, "n_element_tab": 31, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 661, "n_element_tab_1": 44, "formula_len_all": 1459, "formula_len_all_1": 104, "len_all": 163703, "len_all_1": 67525, "len_abs": 3127, "len_title": 141, "len_sents": 51616, "len_sents_1": 32702, "n_sents": 427, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1268, "title": "An Examination on the Effectiveness of Divide-and-Conquer Prompting in Large Language Models", "abs": "Foundation models, such as Large language Models (LLMs), have attracted significant amount of interest due to their large number of applications. However, when handling tasks involving repetitive sub-tasks and/or deceptive contents, such as arithmetic calculation and article-level fake news detection, simple instructional prompts suffer from inaccurate responses. Existing works show that more complicated prompting strategies, such as Chain-of-Thoughts and Least-to-Most, can unlock LLM's powerful capacity in diverse areas. Recent researches reveal that simple divide-and-conquer prompting strategy, i.e. simply dividing the input sequence to multiple sub-inputs, can substantially improve LLM's performance in some specific tasks such as misinformation detection. In this paper, we aim at understanding the utility of divide-and-conquer prompting strategy, i.e. on which kind of tasks this strategy gets advantages. Specifically, we provide a theoretic analysis to divide-and-conquer prompting strategy and help us identify the specific tasks where DaC prompting can bring performance boost with theoretic guarantee. We then present two cases (\\textbf{large integer arithmetic and fact verification}) where experimental results aligns with our theoretic analysis.", "keywords": [ "Program-guided Prompt", "Divide-and-Conquer", "Foundation Model", "Misinformation Detection" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "cSAAGL0cn0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Changli Wu", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Jiayi Ji", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Liujuan Cao", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoshuai Sun", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Yihang Liu", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Yiwei Ma", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Gen Luogen", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 16, "n_ref_uni": 42, "n_ref": 100, "n_ref_all": 115, "n_fig": 5, "n_tab": 10, "L_tab": 2818, "n_element_tab": 428, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1165, "n_element_tab_1": 185, "formula_len_all": 6675, "formula_len_all_1": 7310, "len_all": 203214, "len_all_1": 72810, "len_abs": 1227, "len_title": 102, "len_sents": 58569, "len_sents_1": 28100, "n_sents": 449, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1233, "title": "Weakly-supervised 3D Referring Expression Segmentation", "abs": "3D Referring Expression Segmentation (3D-RES) aims to generate precise segmentation masks for targets based on free-form text descriptions. Despite significant advancements, current methods still rely on costly point-level mask-description pair annotations. In this paper, we introduce the Multi-Expert Network (MEN), a novel weakly supervised framework that utilizes the multimodal alignment of vision-language models across various semantic cues to reveal the relationships between descriptions and 3D instances. The primary challenges lie in effectively extracting and matching visual and textual context, while eliminating potential distractions. To address this, we propose the Multi-Expert Mining (MEM) and Multi-Expert Aggregation (MEA) modules. The MEM module employs multiple experts to extract semantic cues from full-context, attribute, and category dimensions. The MEA module mathematically consolidates the outputs of these experts, automatically assigning greater weight to more accurate ones, thus improving target selection accuracy and robustness. Extensive experiments on the ScanRefer and Multi3DRefer benchmarks demonstrate the effectiveness of our method in addressing the challenges of weakly supervised 3D-RES.", "keywords": [ "Weakly-supervised learning", "3D Referring Expression Segmentation" ], "rating_list": [ 3, 5, 6, 8 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "cRnCcuLvyr", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Hanwen Wang", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Jacob H Seidman", "gender": "unknown", "institution": "Reality Defender", "country": "US", "position": "Researcher" }, { "name": "Paris Perdikaris", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Associate Professor" }, { "name": "Shyam Sankaran", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Sifan Wang", "gender": "Male", "institution": "Yale University", "country": "", "position": "Postdoc" }, { "name": "George Pappas", "gender": "Male", "institution": "School of Engineering and Applied Science, University of Pennsylvania", "country": "US", "position": "Full Professor" } ], "n_formula": 67, "n_formula_1": 7, "n_ref_uni": 62, "n_ref": 128, "n_ref_all": 157, "n_fig": 11, "n_tab": 10, "L_tab": 2418, "n_element_tab": 179, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 760, "n_element_tab_1": 116, "formula_len_all": 3860, "formula_len_all_1": 500, "len_all": 245985, "len_all_1": 54448, "len_abs": 1409, "len_title": 106, "len_sents": 69650, "len_sents_1": 24545, "n_sents": 592, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1415, "title": "CViT: Continuous Vision Transformer for Operator Learning", "abs": "Operator learning, which aims to approximate maps between infinite-dimensional function spaces, is an important area in scientific machine learning with applications across various physical domains. Here we introduce the Continuous Vision Transformer (CViT), a novel neural operator architecture that leverages advances in computer vision to address challenges in learning complex physical systems. CViT combines a vision transformer encoder, a novel grid-based coordinate embedding, and a query-wise cross-attention mechanism to effectively capture multi-scale dependencies. This design allows for flexible output representations and consistent evaluation at arbitrary resolutions. We demonstrate CViT's effectiveness across a diverse range of partial differential equation (PDE) systems, including fluid dynamics, climate modeling, and reaction-diffusion processes. Our comprehensive experiments show that CViT achieves state-of-the-art performance on multiple benchmarks, often surpassing larger foundation models, even without extensive pretraining and roll-out fine-tuning. Taken together, CViT exhibits robust handling of discontinuous solutions, multi-scale features, and intricate spatio-temporal dynamics. Our contributions can be viewed as a significant step towards adapting advanced computer vision architectures for building more flexible and accurate machine learning models in the physical sciences.", "keywords": [ "Scientific Machine Learning", "Operator Learning", "Neural Operators", "Neural Fields", "Vision Transformer", "Partial Differential Equations" ], "rating_list": [ 6, 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "cRR0oDFEBC", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Bowen Yu", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Chang Zhou", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Chengpeng Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Guanting Dong", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Jingren Zhou", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Keming Lu", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Tingyu Xia", "gender": "Female", "institution": "Jilin University", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 37, "n_ref": 82, "n_ref_all": 96, "n_fig": 7, "n_tab": 13, "L_tab": 4612, "n_element_tab": 614, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 373, "formula_len_all_1": 335, "len_all": 257802, "len_all_1": 60997, "len_abs": 3806, "len_title": 152, "len_sents": 62002, "len_sents_1": 28758, "n_sents": 459, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 104, "L_abs": 1599, "title": "Self-play with Execution Feedback: Improving Instruction-following Capabilities of Large Language Models", "abs": "One core capability of large language models~(LLMs) is to follow natural language instructions. However, the issue of automatically constructing high-quality training data to enhance the complex instruction-following abilities of LLMs without manual annotation remains unresolved. In this paper, we introduce AutoIF, the first scalable and reliable method for automatically generating instruction-following training data. AutoIF transforms the validation of instruction-following data quality into code verification, requiring LLMs to generate instructions, the corresponding code to verify the correctness of the instruction responses, and unit test samples to cross-validate the code's correctness. Then, execution feedback-based rejection sampling can generate data for Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF) training. AutoIF achieves significant improvements across three training algorithms, SFT, Offline DPO, and Online DPO, when applied to the advanced open-source LLMs, Qwen2 and LLaMA3, in self-alignment and strong-to-weak distillation settings. Using two widely-used and three challenging general instruction-following benchmarks, we demonstrate that AutoIF significantly improves LLM performance across a wide range of natural instruction constraints. Notably, AutoIF is the first to surpass 90\\% accuracy in IFEval\u2019s loose instruction accuracy, without compromising general, math and coding capabilities. Further analysis of quality, scaling, combination, and data efficiency highlights AutoIF's strong generalization and alignment potential.", "keywords": [ "Instruction Following", "Large Language Models", "Execution Feedback", "On-policy Learning", "Strong-to-Weak Distillation", "Self-Alignment" ], "rating_list": [ 6, 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 4 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4, 2 ] }, { "paper_id": "cR5GTis5II", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Alexander DeRieux", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "PhD student" }, { "name": "Walid Saad", "gender": "unknown", "institution": "Virginia Tech", "country": "US", "position": "Full Professor" } ], "n_formula": 48, "n_formula_1": 15, "n_ref_uni": 23, "n_ref": 71, "n_ref_all": 117, "n_fig": 13, "n_tab": 12, "L_tab": 7027, "n_element_tab": 906, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 294, "n_element_tab_1": 107, "formula_len_all": 3483, "formula_len_all_1": 1381, "len_all": 200071, "len_all_1": 76095, "len_abs": 1879, "len_title": 159, "len_sents": 69587, "len_sents_1": 35390, "n_sents": 504, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 110, "L_abs": 1899, "title": "eQMARL: Entangled Quantum Multi-Agent Reinforcement Learning for Distributed Cooperation over Quantum Channels", "abs": "Collaboration is a key challenge in distributed multi-agent reinforcement learning (MARL) environments. Learning frameworks for these decentralized systems must weigh the benefits of explicit player coordination against the communication overhead and computational cost of sharing local observations and environmental data. Quantum computing has sparked a potential synergy between quantum entanglement and cooperation in multi-agent environments, which could enable more efficient distributed collaboration with minimal information sharing. This relationship is largely unexplored, however, as current state-of-the-art quantum MARL (QMARL) implementations rely on classical information sharing rather than entanglement over a quantum channel as a coordination medium. In contrast, in this paper, a novel framework dubbed entangled QMARL (eQMARL) is proposed. The proposed eQMARL is a distributed actor-critic framework that facilitates cooperation over a quantum channel and eliminates local observation sharing via a quantum entangled split critic. Introducing a quantum critic uniquely spread across the agents allows coupling of local observation encoders through entangled input qubits over a quantum channel, which requires no explicit sharing of local observations and reduces classical communication overhead. Further, agent policies are tuned through joint observation-value function estimation via joint quantum measurements, thereby reducing the centralized computational burden. Experimental results show that eQMARL with $\\Psi^{+}$ entanglement converges to a cooperative strategy up to $17.8\\\\%$ faster and with a higher overall score compared to split classical and fully centralized classical and quantum baselines. The results also show that eQMARL achieves this performance with a constant factor of $25$-times fewer centralized parameters compared to the split classical baseline.", "keywords": [ "quantum machine learning", "multi-agent reinforcement learning", "quantum entanglement" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 2, 3, 2 ] }, { "paper_id": "cQCrBJHy0C", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Ming-Ming Cheng", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Full Professor" }, { "name": "Qi Guo", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Qing Guo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Shanmin Pang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Tianyu Wei", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yizhuo Ma", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xiaofeng Cao", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 9, "n_ref_uni": 18, "n_ref": 41, "n_ref_all": 82, "n_fig": 14, "n_tab": 6, "L_tab": 5131, "n_element_tab": 90, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1365, "formula_len_all_1": 442, "len_all": 160976, "len_all_1": 59788, "len_abs": 1562, "len_title": 127, "len_sents": 56715, "len_sents_1": 28768, "n_sents": 448, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1445, "title": "EmoAttack: Emotion-to-Image Diffusion Models for Emotional Backdoor Generation", "abs": "Text-to-image diffusion models can generate realistic images based on textual inputs, enabling users to convey their opinions visually through language. Meanwhile, within language, emotion plays a crucial role in expressing personal opinions in our daily and the inclusion of maliciously negative content can lead users astray, exacerbating negative emotions. Recognizing the success of diffusion models and the significance of emotion, we investigate a previously overlooked risk associated with text-to-image diffusion models, that is, utilizing emotion in the input texts to introduce negative content and provoke unfavorable emotions in users. Specifically, we identify a new backdoor attack, i.e., emotion-aware backdoor attack (EmoAttack), which introduces malicious negative content triggered by emotional texts during image generation. We formulate such an attack as a diffusion personalization problem to avoid extensive model retraining and propose the \\textit{EmoBooth}. Unlike existing personalization methods, our approach fine-tunes a pre-trained diffusion model by establishing a mapping between a cluster of emotional words and a given reference image containing malicious negative content. To validate the effectiveness of our method, we built a dataset and conducted extensive analysis and discussion about its effectiveness. Given consumers' widespread use of diffusion models, uncovering this threat is critical for society.", "keywords": [ "Emotion;Backdoor Attack;Diffusion model;Personalization" ], "rating_list": [ 5, 8, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 5, 2, 4 ] }, { "paper_id": "cQ25MQQSNI", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Anas Awadalla", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Jack Hessel", "gender": "Male", "institution": "Samaya AI", "country": "", "position": "Researcher" }, { "name": "Jae Sung Park", "gender": "unknown", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Khyathi Chandu", "gender": "Female", "institution": "Mistral AI", "country": "US", "position": "Researcher" }, { "name": "Lijuan Wang", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Linjie Li", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Ximing Lu", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Yejin Choi", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 49, "n_ref": 118, "n_ref_all": 143, "n_fig": 8, "n_tab": 11, "L_tab": 2667, "n_element_tab": 430, "n_fig_1": 5, "n_tab_1": 9, "L_tab_1": 1995, "n_element_tab_1": 213, "formula_len_all": 124, "formula_len_all_1": 118, "len_all": 253056, "len_all_1": 67386, "len_abs": 1653, "len_title": 92, "len_sents": 69286, "len_sents_1": 30925, "n_sents": 548, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1659, "title": "CertainlyUncertain: A Benchmark and Metric for Multimodal Epistemic and Aleatoric Awareness", "abs": "The ability to acknowledge the inevitable uncertainty in their knowledge and reasoning is a prerequisite for AI systems to be truly truthful and reliable. In this paper, we present a taxonomy of uncertainty specific to vision-language AI systems, distinguishing between epistemic uncertainty (arising from a lack of information) and aleatoric uncertainty (due to inherent unpredictability), and further explore finer categories within. Based on this taxonomy, we synthesize a benchmark dataset, CertainlyUncertain, featuring 178K visual question answering (VQA) samples as contrastive pairs. This is achieved by 1) inpainting images to make previously answerable questions into unanswerable ones; and 2) using image captions to prompt large language models for both answerable and unanswerable questions. Additionally, we introduce a new metric confidence-weighted accuracy, that is well correlated with both accuracy and calibration error, to address the shortcomings of existing metrics. Despite the recent rapid progress in vision-language models (VLMs), evaluations on our benchmark show that they perform poorly in uncertain scenarios. Further experiments demonstrate that supervised fine-tuning with CertainlyUncertain enhances the performance of VLMs, and reduces the calibration error. These improvements extend beyond our benchmark to existing refusal-oriented datasets and show positive results on reducing hallucinations, while maintaining performance on standard VQA benchmarks. Our work underscores the importance of addressing uncertainty in vision-language AI systems to improve their reliability and trustworthiness in real-world applications.", "keywords": [ "multimodal", "refusals", "hallucinations" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 3, 4, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "cPozlf9OaF", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 10, "author_info_list": [ { "name": "Jiannan Cao", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "MS student" }, { "name": "Jianwei Yin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Sheng Cheng", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Researcher" }, { "name": "Shi Bo", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "PhD student" }, { "name": "Tianyu Du", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xuhong Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xun Wang", "gender": "Male", "institution": "Northeast Forest University", "country": "CN", "position": "MS student" }, { "name": "Yanming Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yanxin Shen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "xinyue peng", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 10, "n_ref_uni": 35, "n_ref": 60, "n_ref_all": 69, "n_fig": 4, "n_tab": 2, "L_tab": 2079, "n_element_tab": 395, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 452, "formula_len_all_1": 576, "len_all": 147834, "len_all_1": 59399, "len_abs": 1223, "len_title": 140, "len_sents": 39217, "len_sents_1": 28824, "n_sents": 267, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1228, "title": "Bridging Context Gaps: Leveraging Coreference Resolution for Long Contextual Understanding", "abs": "Large language models (LLMs) have shown remarkable capabilities in natural language processing; however, they still face difficulties when tasked with understanding lengthy contexts and executing effective question answering. These challenges often arise due to the complexity and ambiguity present in longer texts. To enhance the performance of LLMs in such scenarios, we introduce the Long Question Coreference Adaptation (LQCA) method. This innovative framework focuses on coreference resolution tailored to long contexts, allowing the model to identify and manage references effectively. The LQCA method encompasses four key steps: resolving coreferences within sub-documents, computing the distances between mentions, defining a representative mention for coreference, and answering questions through mention replacement. By processing information systematically, the framework provides easier-to-handle partitions for LLMs, promoting better understanding. Experimental evaluations on a range of LLMs and datasets have yielded positive results, with a notable improvements on OpenAI-o1-mini and GPT-4o models, highlighting the effectiveness of leveraging coreference resolution to bridge context gaps in question answering.", "keywords": [ "Coreference Resolution", "Long Contextual Understanding", "Information Extraction" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "cPjBTj1Qf5", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Dennis Frauen", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Stefan Feuerriegel", "gender": "Male", "institution": "LMU Munich", "country": "DE", "position": "Professor" }, { "name": "Valentyn Melnychuk", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Yuchen Ma", "gender": "Female", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 85, "n_formula_1": 15, "n_ref_uni": 67, "n_ref": 201, "n_ref_all": 230, "n_fig": 17, "n_tab": 5, "L_tab": 899, "n_element_tab": 42, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 597, "n_element_tab_1": 9, "formula_len_all": 8423, "formula_len_all_1": 1203, "len_all": 253087, "len_all_1": 70987, "len_abs": 1143, "len_title": 156, "len_sents": 85884, "len_sents_1": 32389, "n_sents": 764, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1170, "title": "Counterfactual fairness prediction: Consistent estimation with generative models and theoretical guarantees", "abs": "Fairness in predictions is of direct importance in practice due to legal, ethical, and societal reasons. This is often accomplished through counterfactual fairness, which ensures that the prediction for an individual is the same as that in a counterfactual world under a different sensitive attribute. However, achieving counterfactual fairness is challenging as counterfactuals are unobservable, and, because of that, existing baselines for counterfactual fairness do not have theoretical guarantees. In this paper, we propose a novel counterfactual fairness predictor for making predictions under counterfactual fairness. Here, we follow the standard counterfactual fairness setting and directly learn the counterfactual distribution of the descendants of the sensitive attribute via tailored neural networks, which we then use to enforce fair predictions through a novel counterfactual mediator regularization. Unique to our work is that we provide theoretical guarantees that our method is effective in ensuring the notion of counterfactual fairness. We further compare the performance across various datasets, where our method achieves state-of-the-art performance.", "keywords": [ "Counterfactual", "Causal inference", "Counterfactual fairness", "Generative models" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 1 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 3, 5, 3 ] }, { "paper_id": "cPZepCZlFW", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 9, "author_info_list": [ { "name": "Bo Li", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Junlin Huang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "", "position": "MS student" }, { "name": "Peijie Dong", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guang Zhou)", "country": "CN", "position": "Phd student" }, { "name": "Shaohuai Shi", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Xiaowen Chu", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Xueze Kang", "gender": "unknown", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Yuxin Wang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "PhD student" }, { "name": "Zichen TANG", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Zhenheng TANG", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Postdoc" } ], "n_formula": 42, "n_formula_1": 22, "n_ref_uni": 50, "n_ref": 142, "n_ref_all": 161, "n_fig": 7, "n_tab": 10, "L_tab": 2894, "n_element_tab": 365, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1857, "n_element_tab_1": 94, "formula_len_all": 4076, "formula_len_all_1": 1414, "len_all": 184298, "len_all_1": 67988, "len_abs": 2000, "len_title": 141, "len_sents": 41493, "len_sents_1": 24900, "n_sents": 366, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1473, "title": "Capturing and Mitigating Gradient Aggregation Errors for Fault-Tolerant Distributed Training", "abs": "Capturing and recovering from hardware failures is important in fault-tolerant distributed training to guarantee system efficiency. However, some hardware-related silent data corruption errors during gradient aggregation like bit corruptions or communication noise, are difficult to capture and address, leading to slow or failed convergence. \nTo understand and mitigate these errors, we first mathematically formulate and generalize them as gradient inconsistency. Then, we theoretically analyze how it leads to model divergence accumulated during training and the failed convergence. \nBased on the analytical study, we design PAFT, a fault-tolerant distributed training system with dynamic and asynchronous parameter synchronization. PAFT includes two parts: (1) PAFT-Sync, which mitigates model divergence by periodically synchronizing parameters, and (2) PAFT-Dyn, which minimizes synchronization overhead through dynamic training overlap and synchronization frequency scheduling based on profiled error degrees. Together, they ensure efficient model convergence at scale. The fault-tolerant synchronization in PAFT is optimized to support commonly used optimizers, e.g., Stochastic Gradient Descent (SGD), SGD momentum, and Adam. \nWe implement PAFT on PyTorch Distributed and train ResNet, GPT-2, and LLaMA-2 on 4$\\sim$ 32 GPUs. Experimental results show that PAFT efficiently defends against gradient aggregation error degrees while maintaining training performance.", "keywords": [ "Distributed Training", "Fault Tolerance", "Infrastructure" ], "rating_list": [ 3, 6, 1, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 1, 3, 1, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "cPIs6PlCuE", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Huaming Chen", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Teng Joon Lim", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "Full Professor" }, { "name": "Zhibo Jin", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Zhu Zhiyu", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 7, "n_ref_uni": 25, "n_ref": 55, "n_ref_all": 69, "n_fig": 3, "n_tab": 5, "L_tab": 1149, "n_element_tab": 163, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1638, "n_element_tab_1": 61, "formula_len_all": 1192, "formula_len_all_1": 582, "len_all": 107856, "len_all_1": 60579, "len_abs": 1670, "len_title": 99, "len_sents": 38008, "len_sents_1": 28347, "n_sents": 259, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1679, "title": "Rethinking Lipschitzness Data-free Backdoor Defense", "abs": "Deep Neural Networks (DNNs) have demonstrated remarkable success across various applications, yet some studies reveal their vulnerability to backdoor attacks, where attackers manipulate models under specific conditions using triggers. It significantly compromise the model integrity. \nAddressing this critical security issue requires robust defence mechanisms to ensure the reliability of DNN models. However, most existing defence mechanisms heavily rely on specialized defence datasets, which are often difficult to obtain due to data privacy and security concerns. This highlights the urgent need for effective data-free defence strategies. In this work, we propose Lipschitzness Precise Pruning (LPP), a novel data-free backdoor defence algorithm that leverages the properties of Lipschitz function to detect and mitigate backdoor vulnerabilities by pruning neurons with strong backdoor correlations while fine-tuning unaffected neurons. Our approach optimizes the computation of the Lipschitz constant using dot product properties, allowing for efficient and precise identification of compromised neurons without the need of clean defence data. This method addresses the limitations of existing data-free defences and extends the scope of backdoor mitigation to include fully connected layers, ensuring comprehensive protection of DNN models. As our approach does not require data exchange, it can be implemented efficiently and effectively in diverse environments. Extensive experiments demonstrate that LPP outperforms state-of-the-art defence approaches without the need for additional defence datasets. We release our code at: https://anonymous.4open.science/r/LPP-CD3C.", "keywords": [ "backdoor attack", "backdoor defense", "model implement", "system security", "AI security" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "cPD2hU35x3", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Bryan Catanzaro", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Vice President" }, { "name": "Chejian Xu", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Mohammad Shoeybi", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Director of Applied Resesrch" }, { "name": "Peng Xu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Wei Ping", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Xianchao Wu", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Senior Data Scientist" }, { "name": "Zihan Liu", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 85, "n_ref_all": 111, "n_fig": 4, "n_tab": 6, "L_tab": 2756, "n_element_tab": 279, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 693, "n_element_tab_1": 103, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 125987, "len_all_1": 54181, "len_abs": 1814, "len_title": 131, "len_sents": 29341, "len_sents_1": 24944, "n_sents": 210, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1737, "title": "ChatQA 2: Bridging the Gap to Proprietary LLMs in Long Context and RAG Capabilities", "abs": "In this work, we introduce ChatQA 2, an Llama 3.0-based model with a 128K\ncontext window, designed to bridge the gap between open-source LLMs and\nleading proprietary models (e.g., GPT-4-Turbo-2024-04-09) in long context un-\nderstanding and retrieval-augmented generation (RAG) capabilities. These two\ncapabilities are complementary to each other and essential for LLMs to process\nlarge volumes of information that cannot fit into a single prompt. We present\na detailed continued training recipe to extend the context window of Llama3-\n70B-base from 8K to 128K tokens, along with a three-stage instruction tun-\ning process to enhance the model\u2019s instruction-following, RAG performance,\nand long-context understanding capabilities. Our results demonstrate that the\nLlama3-ChatQA-2-70B model outperforms most existing state-of-the-art models,\nincluding GPT-4-Turbo-2024-04-09, Qwen2-72B-Instruct, and Llama3.1-70B-\nInstruct, on ultra-long tasks beyond 100K tokens, as well as on the RAG benchmark\nusing only a 4K context window, showing the strong long context capability across\nvarying sequence lengths. We further provide extensive comparisons between\ndirect long-context and RAG solutions using the same state-of-the-art long-context\nLLMs. Interestingly, we find that the performance of strong long-context LLMs\nusing RAG improves when retrieving a larger number of chunks. With a large set\nof top-k chunks, RAG consistently outperforms direct long-context solution using\nthe same state-of-the-art long-context models (e.g., Llama3-ChatQA-2-70B and\nQwen2-72B-Instruct) on both 32K and 128K benchmarks. We open-source the\nmodel weights, training data, and the evaluation setup for the for the community:\nhttps://chatqa2-project.github.io/", "keywords": [ "Long Context LLM", "Retrieval-augmented generation" ], "rating_list": [ 6, 6, 6, 6, 5 ], "soundness_list": [ 4, 3, 3, 2, 1 ], "presentation_list": [ 1, 4, 3, 3, 1 ], "contribution_list": [ 4, 3, 2, 3, 2 ], "confidence_list": [ 2, 4, 3, 3, 4 ] }, { "paper_id": "cP00UB2654", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Faisal Hamman", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Freddy Lecue", "gender": "unknown", "institution": "INRIA", "country": "", "position": "Full Professor" }, { "name": "Pasan Dissanayake", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Sanghamitra Dutta", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Saumitra Mishra", "gender": "Male", "institution": "J.P. Morgan Chase", "country": "", "position": "Researcher" } ], "n_formula": 39, "n_formula_1": 14, "n_ref_uni": 54, "n_ref": 118, "n_ref_all": 148, "n_fig": 5, "n_tab": 6, "L_tab": 2421, "n_element_tab": 490, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 585, "n_element_tab_1": 6, "formula_len_all": 3387, "formula_len_all_1": 890, "len_all": 193858, "len_all_1": 70312, "len_abs": 1561, "len_title": 123, "len_sents": 60178, "len_sents_1": 34526, "n_sents": 440, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1576, "title": "Quantifying Prediction Consistency Under Model Multiplicity in Tabular LLMs", "abs": "Fine-tuning large language models (LLMs) on tabular data for classification can lead to the phenomenon of \\emph{fine-tuning multiplicity}, where equally well-performing models make conflicting predictions on the same input. Fine-tuning multiplicity can arise due to variations in the training process, e.g., seed, random weight initialization, retraining on a few additional or deleted data points. This raises critical concerns about the robustness and reliability of Tabular LLMs, particularly when deployed for high-stakes decision-making, such as finance, hiring, education, healthcare, etc. This work formalizes the unique challenge of fine-tuning multiplicity in Tabular LLMs and proposes a novel measure to quantify the robustness of individual predictions without expensive model retraining. Our measure quantifies a prediction's robustness by analyzing (sampling) the model's local behavior around the input in the embedding space. Interestingly, we show that sampling in the local neighborhood can be leveraged to provide probabilistic robustness guarantees against a broad class of equally-well-performing fine-tuned models. By leveraging Bernstein's Inequality, we show that predictions with sufficiently high robustness (as defined by our measure) will remain consistent with high probability. We also provide empirical evaluation on real-world datasets to support our theoretical results. Our work highlights the importance of addressing fine-tuning instabilities to enable trustworthy deployment of Tabular LLMs in high-stakes and safety-critical applications.", "keywords": [ "trustworthy large language models", "Reliable machine learning", "tabular data", "model multiplicity", "high-stakes application" ], "rating_list": [ 8, 1, 3, 5 ], "soundness_list": [ 3, 1, 2, 2 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "cNmu0hZ4CL", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Alex H Williams", "gender": "Male", "institution": "New York University", "country": "", "position": "Assistant Professor" }, { "name": "Amin Nejatbakhsh", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Research Scientist" }, { "name": "David Lipshutz", "gender": "Male", "institution": "Flatiron Institute", "country": "", "position": "Associate Research Scientist" }, { "name": "Victor Geadah", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" } ], "n_formula": 49, "n_formula_1": 20, "n_ref_uni": 38, "n_ref": 55, "n_ref_all": 101, "n_fig": 20, "n_tab": 1, "L_tab": 90, "n_element_tab": 12, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 203, "n_element_tab_1": 42, "formula_len_all": 4941, "formula_len_all_1": 1351, "len_all": 150965, "len_all_1": 65148, "len_abs": 1101, "len_title": 124, "len_sents": 52574, "len_sents_1": 31234, "n_sents": 397, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 76, "L_abs": 1106, "title": "Comparing noisy neural population dynamics using optimal transport distances", "abs": "Biological and artificial neural systems form high-dimensional neural representations that underpin their computational capabilities. Methods for quantifying geometric similarity in neural representations have become a popular tool for identifying computational principles that are potentially shared across neural systems. These methods generally assume that neural responses are deterministic and static. However, responses of biological systems, and some artificial systems, are noisy and dynamically unfold over time. Furthermore, these characteristics can have substantial influence on a system\u2019s computational capabilities. Here, we demonstrate that existing metrics can fail to capture key differences between neural systems with noisy dynamic responses. We then propose a metric for comparing the geometry of noisy neural trajectories, which can be derived as an optimal transport distance between Gaussian processes. We use the metric to compare models of neural responses in different regions of the motor system and to compare the dynamics of latent diffusion models for text-to-image synthesis.", "keywords": [ "Representational similarity", "shape metrics", "optimal transport", "Wasserstein distance" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 4, 3 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "cNaHOdvh9J", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Hoin Jung", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Junyi Chai", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Xiaoqian Wang", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 74, "n_formula_1": 11, "n_ref_uni": 35, "n_ref": 80, "n_ref_all": 109, "n_fig": 9, "n_tab": 24, "L_tab": 12698, "n_element_tab": 845, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 305, "n_element_tab_1": 198, "formula_len_all": 7082, "formula_len_all_1": 749, "len_all": 202989, "len_all_1": 54156, "len_abs": 1519, "len_title": 100, "len_sents": 59326, "len_sents_1": 25867, "n_sents": 480, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1523, "title": "Adversarial Latent Feature Augmentation for Fairness", "abs": "Achieving fairness in machine learning remains a critical challenge, especially due to the opaque effects of data augmentation on input spaces within nonlinear neural networks. Nevertheless, current approaches that emphasize augmenting latent features, rather than input spaces, offer limited insights into their ability to detect and mitigate bias. In response, we introduce the concept of the \"unfair region\" in the latent space, a subspace that highlights areas where misclassification rates for certain demographic groups are disproportionately high, leading to unfair prediction results. To address this, we propose Adversarial Latent Feature Augmentation (ALFA), a method that leverages adversarial fairness attacks to perturb latent space features, which are then used as data augmentation for fine-tuning. ALFA intentionally shifts latent features into unfair regions, and the last layer of the network is fine-tuned with these perturbed features, leading to a corrected decision boundary that enhances fairness in classification in a cost-effective manner. We present a theoretical framework demonstrating that our adversarial fairness objective reliably generates biased feature perturbations, and that fine-tuning on samples from these unfair regions ensures fairness improvements. Extensive experiments across diverse datasets, modalities, and backbone networks validate that training with these adversarial features significantly enhances fairness while maintaining predictive accuracy in classification tasks.", "keywords": [ "Fairness", "Data Augmentation", "Adversarial Attack" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "cNThpik3Jz", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Aleksander J\u0119drosz", "gender": "unknown", "institution": "Snowflake Computing", "country": "PL", "position": "Researcher" }, { "name": "Micha\u0142 Pietruszka", "gender": "unknown", "institution": "Snowflake computing ", "country": "PL", "position": "Researcher" }, { "name": "Pawe\u0142 Morawiecki", "gender": "Male", "institution": "Institute of Computer Science, Polish Academy of Sciences", "country": "", "position": "Associate Professor" }, { "name": "\u0141ukasz Borchmann", "gender": "Male", "institution": "Snowflake", "country": "PL", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 0, "n_ref_uni": 27, "n_ref": 71, "n_ref_all": 86, "n_fig": 7, "n_tab": 6, "L_tab": 5430, "n_element_tab": 589, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1689, "n_element_tab_1": 319, "formula_len_all": 449, "formula_len_all_1": 0, "len_all": 186574, "len_all_1": 55925, "len_abs": 842, "len_title": 123, "len_sents": 62843, "len_sents_1": 25196, "n_sents": 435, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 861, "title": "Can Models Help us Create Better Models? Evaluating LLMs as Data Scientists", "abs": "We present a benchmark for large language models \ndesigned to tackle one of the most knowledge-intensive tasks in data science: writing\n\\textit{feature engineering} code, which requires domain knowledge in addition to a deep understanding of the underlying problem and data structure.\nThe model is provided with a dataset description in a prompt and asked to generate code transforming it. The evaluation score is derived from the improvement achieved by an XGBoost model fit on the modified\ndataset compared to the original data. By an extensive evaluation of state-of-the-art models and comparison to well-established benchmarks, we demonstrate that the \\bench{} of our proposal can cheaply and efficiently assess the broad capabilities of LLMs, in contrast to the existing methods.\nThe reference implementation is available at \\url{ATTACHED_AS_SUPPLEMENTARY}", "keywords": [ "llm", "data science", "benchmark", "tabular data", "feature engineering", "dataset", "kaggle", "evaluation", "code generation" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "cMLtjP3Cym", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Bei Wang", "gender": "Female", "institution": "NIO", "country": "", "position": "Researcher" }, { "name": "Beining Zhang", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "MS student" }, { "name": "Handing Wang", "gender": "Female", "institution": "Xidian University ", "country": "CN", "position": "Full Professor" }, { "name": "Jing Liu", "gender": "Female", "institution": "Xidian University, China", "country": "CN", "position": "Full Professor" }, { "name": "Kai Wu", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaoyu Zhang", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 3, "n_ref_uni": 55, "n_ref": 140, "n_ref_all": 177, "n_fig": 4, "n_tab": 23, "L_tab": 7519, "n_element_tab": 1197, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2893, "n_element_tab_1": 435, "formula_len_all": 856, "formula_len_all_1": 122, "len_all": 212895, "len_all_1": 58172, "len_abs": 958, "len_title": 118, "len_sents": 62104, "len_sents_1": 21743, "n_sents": 487, "n_sents_1": 149, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 962, "title": "TSGGuide: Recommendation Guide for Multivariate Time Series Generation", "abs": "Multivariate Time Series Generation (MTSG) plays a crucial role in time series analysis, supporting tasks such as data augmentation and anomaly detection. While several methods exist for MTSG, recommending the most suitable method for new scenarios remains a significant challenge. Although prior work by (Ang et al., 2023a) provides guidance for selecting MTSG methods, it lacks coverage of recent diffusion-based methods and has limited exploration of channel-independent frameworks. We address these gaps by improving the recommendation guide, highlighting the effectiveness of a central discriminator within the channel-independent framework. Our revised guide makes three key recommendations: 1) VAE-based methods excel on small-scale datasets; 2) a channel-independent framework with the newly designed central discriminator is optimal in most cases; and 3) a diffusion-based method is preferable when ample data and computational resources are available.", "keywords": [ "time series generation", "diffusion model", "Channel Independence" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "cLws58ZojF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Han Li", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Lingyun Wang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Ning Xie", "gender": "Female", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Sheng Liu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Xiaohan Nie", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 69, "n_ref_all": 82, "n_fig": 3, "n_tab": 8, "L_tab": 2249, "n_element_tab": 238, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 2117, "n_element_tab_1": 233, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 110670, "len_all_1": 59548, "len_abs": 1472, "len_title": 85, "len_sents": 27492, "len_sents_1": 26005, "n_sents": 222, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1231, "title": "An Exploration of Speech Conditioned Large Language Models (SLMs)", "abs": "Efforts to enable Large Language Models (LLMs) to understand human speech have spurred the development of an increasing number of Speech-Conditioned Large Language Models (SLMs). While these models have demonstrated success on various speech-related tasks, such as automatic speech recognition (ASR), the design space of SLMs has not been thoroughly explored. In this work, we revisit key design choices for SLMs, aiming to gain insights into how these choices impact the performance of SLMs and how we could optimize them for better results. Surprisingly, our experiments reveal that current SLMs struggle to follow speech instructions or respond to speech inputs, even for simple queries like \u201dwho has been to the moon?\u201d. Our experimental findings indicate that speech instruction following data is crucial for improving these capabilities. Leveraging this insight, we propose to use synthetic speech instruction following data to enhance speech instruction following capability. Combining the findings from our other experiments, we provide an effective recipe for developing SLMs. Our model, called SiM, not only achieves strong ASR performance, but also significantly outperforms existing SLMs in speech instruction following.", "keywords": [ "Speech Conditioned Large Language Models" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "cLtE4qoPlD", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Junghun Oh", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "PhD student" }, { "name": "Sungyong Baik", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "Assistant Professor" }, { "name": "Kyoung Mu Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 9, "n_ref_uni": 39, "n_ref": 92, "n_ref_all": 118, "n_fig": 14, "n_tab": 2, "L_tab": 3628, "n_element_tab": 35, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 831, "n_element_tab_1": 21, "formula_len_all": 1094, "formula_len_all_1": 536, "len_all": 131169, "len_all_1": 73462, "len_abs": 1762, "len_title": 107, "len_sents": 40990, "len_sents_1": 33547, "n_sents": 269, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1769, "title": "Find A Winning Sign: Sign Is All We Need to Win the Lottery", "abs": "The lottery ticket hypothesis (LTH) posits the existence of a sparse network (a.k.a. winning ticket) that can generalize comparably to its dense counterpart after training from initialization. However, early works fail to generalize its observation and method to large-scale settings. While recent methods, such as weight rewinding or learning rate rewinding (LRR), may have found effective pruning methods, we note that they still struggle with identifying a winning ticket. In this paper, we take a step closer to finding a winning ticket by arguing that a signed mask, a binary mask with parameter sign information, can transfer the capability to achieve strong generalization after training (i.e., generalization potential) to a randomly initialized network. We first share our observation on the subnetwork trained by LRR: if the parameter signs are maintained, the LRR-driven subnetwork retains its generalization potential even when the parameter magnitudes are randomly initialized, excluding those of normalization layers. However, this fails when the magnitudes of normalization layer parameters are initialized together. To tackle the significant influence of normalization layer parameters, we propose AWS, a slight variation of LRR to find A Winning Sign. Specifically, we encourage low error barriers along the linear path connecting the subnetwork trained by AWS to its counterpart with initialized normalization layer parameters, maintaining the generalization potential even when all parameters are initialized. Interestingly, we observe that across various architectures and datasets, a signed mask of the AWS-driven subnetwork can allow a randomly initialized network to perform comparably to a dense network, taking a step closer to the goal of LTH.", "keywords": [ "lottery ticket hypothesis", "network pruning", "linear mode connectivity" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "cLj51OYBsh", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Carles Hernandez", "gender": "Male", "institution": "Universidad Polit\u00e9cnica de Valencia", "country": "", "position": "Researcher" }, { "name": "IMANOL ALLENDE", "gender": "Male", "institution": "Opentech EDV Research GMBH", "country": "AT", "position": "Researcher" }, { "name": "Nicholas Mc Guire", "gender": "unknown", "institution": "", "country": "", "position": "" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 14, "n_ref": 29, "n_ref_all": 52, "n_fig": 4, "n_tab": 9, "L_tab": 4483, "n_element_tab": 455, "n_fig_1": 3, "n_tab_1": 10, "L_tab_1": 4539, "n_element_tab_1": 458, "formula_len_all": 74, "formula_len_all_1": 74, "len_all": 80501, "len_all_1": 60535, "len_abs": 710, "len_title": 108, "len_sents": 24644, "len_sents_1": 23225, "n_sents": 228, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 711, "title": "Power of Augmented Replicas in Out-Of-Distribution Detection", "abs": "Data augmentation is widely used in machine learning to enhance training datasets by introducing minor variations to the original data, traditionally aiming to prevent overfitting and improve model performance. This paper explores a novel application of data augmentation during the inference stage to enhance out-of-distribution (OOD) detection. The proposed method involves replicating the inference image multiple times, applying various transformation techniques to each replica, and then evaluating the detectors using these augmented images. The effectiveness of this approach is assessed across different detectors, models, and datasets, demonstrating its potential to improve OOD detection capabilities.", "keywords": [ "Data Augmentation", "Out-of-Distribution Detection (OOD)", "trustworthy ML" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 1, 1, 1, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "cLYvhd0pDY", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Dawei Yin", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Principal Researcher" }, { "name": "Fan Liu", "gender": "unknown", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Hao Liu", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Lixin Su", "gender": "Male", "institution": "Baidu", "country": "", "position": "Researcher" }, { "name": "Xinyu Ma", "gender": "Male", "institution": "Baidu", "country": "", "position": "Researcher" }, { "name": "Yue Feng", "gender": "Female", "institution": "University of Birmingham", "country": "GB", "position": "Assistant Professor" }, { "name": "Zhao Xu", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 19, "n_formula_1": 11, "n_ref_uni": 13, "n_ref": 36, "n_ref_all": 72, "n_fig": 21, "n_tab": 6, "L_tab": 6117, "n_element_tab": 532, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 3916, "n_element_tab_1": 476, "formula_len_all": 1115, "formula_len_all_1": 627, "len_all": 225428, "len_all_1": 76783, "len_abs": 2687, "len_title": 159, "len_sents": 70126, "len_sents_1": 32568, "n_sents": 576, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 2694, "title": "JAILJUDGE: A Comprehensive Jailbreak Judge Benchmark with Multi-Agent Enhanced Explanation Evaluation Framework", "abs": "Although significant research efforts have been dedicated to enhancing the safety of large language models (LLMs) by understanding and defending against jailbreak attacks, evaluating the defense capabilities of LLMs against jailbreak attacks also attracts lots of attention. Current evaluation methods lack explainability and do not generalize well to complex scenarios, resulting in incomplete and inaccurate\nassessments (e.g., direct judgment without reasoning explainability, the F1 score of the GPT-4 judge is only 55% in complex scenarios and bias evaluation on multilingual scenarios, etc.). To address these challenges, we have developed a comprehensive evaluation benchmark, JAILJUDGE, which includes a wide range of risk scenarios with complex malicious prompts (e.g., synthetic, adversarial, in-the-wild, and multi-language scenarios, etc.) along with high-quality human-annotated test datasets. Specifically, the JAILJUDGE dataset comprises training data of JAILJUDGE, with over 35k+ instruction-tune training data with reasoning explainability, and JAILJUDGETEST, a 4.5k+ labeled set of broad risk scenarios and a 6k+ labeled set of multilingual scenarios in ten languages. To provide reasoning explanations (e.g., explaining why an LLM is jailbroken or not) and fine-grained evaluations (jailbroken score from 1 to 10), we propose a multi-agent jailbreak judge framework, JailJudge MultiAgent, making the decision inference process explicit and interpretable to enhance evaluation quality. Using this framework, we construct the instruction-tuning ground truth and then instruction-tune an end-to-end jailbreak judge model, JAILJUDGE Guard, which can also provide reasoning explainability with fine-grained evaluations without API costs. Additionally, we introduce JailBoost, an attacker-agnostic attack enhancer, and GuardShield, a safety moderation defense method, both based on JAILJUDGE Guard. Comprehensive experiments demonstrate the superiority of our JAILJUDGE benchmark and jailbreak judge methods. Our jailbreak judge methods (JailJudge MultiAgent and JAILJUDGE Guard) achieve SOTA performance in closed-source models (e.g.,\nGPT-4) and safety moderation models (e.g., Llama-Guard and ShieldGemma, etc.), across a broad range of complex behaviors (e.g., JAILJUDGE benchmark, etc.) to zero-shot scenarios (e.g., other open data, etc.). Importantly, JailBoost and GuardShield, based on JAILJUDGE Guard, can enhance downstream tasks in jailbreak attacks and defenses under zero-shot settings with significant improvement (e.g., JailBoost can increase the average performance by approximately 29.24%, while GuardShield can reduce the average defense ASR from 40.46% to 0.15%).", "keywords": [ "jailbreak judge", "jialbreak evaluation", "jialbreak attack", "jailbreak defense", "LLM" ], "rating_list": [ 5, 3, 6, 8 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "cLTM1gc6Qm", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 4, "author_info_list": [ { "name": "Haoyu Jia", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "MS student" }, { "name": "Kei OKADA", "gender": "unknown", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "JP", "position": "Full Professor" }, { "name": "Kento Kawaharazuka", "gender": "unknown", "institution": "The University of Tokyo", "country": "", "position": "PhD student" }, { "name": "Yoshiki Obinata", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 36, "n_ref_all": 68, "n_fig": 8, "n_tab": 9, "L_tab": 3655, "n_element_tab": 478, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 511, "n_element_tab_1": 50, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 189857, "len_all_1": 60815, "len_abs": 1956, "len_title": 121, "len_sents": 72834, "len_sents_1": 31688, "n_sents": 486, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1933, "title": "Mockingbird: Platform for Adapting LLMs to General Machine Learning Tasks", "abs": "Large language models (LLMs) are now being used with increasing frequency as chat bots, tasked with the summarizing information or generating text and code in accordance with user instructions.\nThe rapid increase in reasoning capabilities and inference speed of LLMs has revealed their remarkable potential for applications extending beyond the domain of chat bots.\nHowever, there is a paucity of research exploring the integration of LLMs into a broader range of intelligent software systems.\nIn this research, we propose a paradigm for leveraging LLMs as mock functions to adapt LLMs to general machine learning tasks.\nFurthermore, we present an implementation of this paradigm, entitled the Mockingbird platform.\nIn this paradigm, users define mock functions which are defined solely by method signature and documentation. Unlike LLM-based code completion tools, this platform does not generate code at compile time; instead, it instructs the LLM to role-play these mock functions at runtime.\nBased on the feedback from users or error from software systems, this platform will instruct the LLM to conduct chains of thoughts to reflect on its previous output, thereby enabling it to perform reinforcement learning.\nThis paradigm fully exploits the intrinsic knowledge and in-context learning ability of LLMs.\nIn comparison to conventional machine learning methods, following distinctive advantages are offered: \n(a) Its intrinsic knowledge enables it to perform well in a wide range of zero-shot scenarios. \n(b) Its flexibility allows it to adapt to random increases or decreases of data fields.\n(c) It can utilize tools and extract information from sources that are inaccessible to conventional machine learning methods, such as the Internet.\nFinally, we evaluated its performance and demonstrated the previously mentioned benefits using several datasets from Kaggle. Our results indicate that this paradigm is highly competitive.", "keywords": [ "Paradigm for AI Systems", "LLM", "In-Context Learning", "Mocking" ], "rating_list": [ 5, 3, 10 ], "soundness_list": [ 2, 1, 3 ], "presentation_list": [ 2, 1, 4 ], "contribution_list": [ 2, 1, 4 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "cKlzKs3Nnb", "primary_area": "foundation or frontier models, including LLMs", "n_author": 16, "author_info_list": [ { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Huan Wang", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Jiacheng Xu", "gender": "Male", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "Kexun Zhang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Lei Li", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Rithesh R N", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Shelby Heinecke", "gender": "Female", "institution": "Salesforce Research", "country": "US", "position": "Researcher" }, { "name": "Silvio Savarese", "gender": "Male", "institution": "Salesforce", "country": "", "position": "Chief Scientist" }, { "name": "Tian Lan", "gender": "unknown", "institution": "SalesForce", "country": "", "position": "Researcher" }, { "name": "Weiran Yao", "gender": "Male", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "Yihao Feng", "gender": "Male", "institution": "Apple AI/ML", "country": "US", "position": "Researcher" }, { "name": "Yingbo Zhou", "gender": "unknown", "institution": "Salesforce Research", "country": "", "position": "Research Scientist" }, { "name": "Zhiwei Liu", "gender": "unknown", "institution": "Salesforce AI Research", "country": "US", "position": "Researcher" }, { "name": "Zuxin Liu", "gender": "Male", "institution": "Salesforce AI Research", "country": "US", "position": "Researcher" }, { "name": "Bo Pang", "gender": "Male", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "Reze Lou", "gender": "Male", "institution": "Pennsylvania State University", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 25, "n_ref": 44, "n_ref_all": 69, "n_fig": 9, "n_tab": 7, "L_tab": 1455, "n_element_tab": 225, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2167, "n_element_tab_1": 167, "formula_len_all": 288, "formula_len_all_1": 269, "len_all": 143899, "len_all_1": 60921, "len_abs": 1275, "len_title": 105, "len_sents": 52405, "len_sents_1": 27235, "n_sents": 423, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1256, "title": "Diversity Empowers Intelligence: Integrating Expertise of Software Engineering Agents", "abs": "Large language model (LLM) agents have shown great potential in solving real-world software engineering (SWE) problems. The most advanced open-source SWE agent can resolve over 27% of real GitHub issues in SWE-Bench Lite. However, these sophisticated agent frameworks exhibit varying strengths, excelling in certain tasks while underperforming in others. To fully harness the diversity of these agents, we propose DEI (Diversity Empowered Intelligence), a framework that leverages their unique expertise. DEI functions as a meta-module atop existing SWE agent frameworks, managing agent collectives for enhanced problem-solving. Experimental results show that a DEI-guided committee of agents is able to surpass the best individual agent's performance by a large margin. For instance, a group of open-source SWE agents, with a maximum individual resolve rate of 27.3% on SWE-Bench Lite, can achieve a 34.3% resolve rate with DEI, making a 25% improvement and beating most closed-source solutions. Our best-performing group excels with a 55% resolve rate, securing the highest ranking on SWE-Bench Lite. Our findings contribute to the growing body of research on collaborative AI systems and their potential to solve complex software engineering challenges.", "keywords": [ "large language models", "LLM agents", "software engineering" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "cKkZEj8n7O", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Beverly Yang", "gender": "unknown", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "EN-HUI YANG", "gender": "unknown", "institution": "University of Waterloo", "country": "CA", "position": "Full Professor" }, { "name": "Kaixiang Zheng", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 28, "n_ref_uni": 33, "n_ref": 47, "n_ref_all": 82, "n_fig": 6, "n_tab": 6, "L_tab": 2032, "n_element_tab": 138, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 477, "n_element_tab_1": 56, "formula_len_all": 1786, "formula_len_all_1": 1682, "len_all": 128537, "len_all_1": 59620, "len_abs": 956, "len_title": 93, "len_sents": 38878, "len_sents_1": 27028, "n_sents": 295, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 958, "title": "Generalization Error Minimized Deep Learning", "abs": "Despite the vast applications and rapid development of deep learning (DL), understanding and improving the generalization ability of deep neural networks (DNNs) remains a fundamental challenge. To tackle this challenge, in this paper, we first establish a novel bias-variance decomposition framework to analyze the generalization error of DNNs. Based on our new generalization error formula, we then present a new form of DL dubbed generalization error minimized (GEM) DL by jointly minimizing the conventional optimization target and an analytical proxy for the generalization error. Extensive experimental results show that in comparison with DNNs trained within the standard DL, GEM DNNs have smaller generalization errors and better generalization ability, thereby improving DNN prediction accuracy. Notably, GEM DL can increase prediction accuracy by as much as 13.19% on ImageNet in the presence of data distribution shift between training and testing.", "keywords": [ "generalization error", "overfitting", "bias-variance decomposition", "image classification" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 1, 3, 1, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "cK7yrw5g5Q", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Mingyuan Zhou", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Shentao Yang", "gender": "Male", "institution": "The University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Weizhu Chen", "gender": "Male", "institution": "Microsoft GenAI", "country": "US", "position": "Vice President" }, { "name": "Yueqin Yin", "gender": "Female", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Yujia Xie", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Yuting Sun", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Ziyi Yang", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Hany Hassan", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Research Scientist" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 86, "n_ref": 141, "n_ref_all": 178, "n_fig": 5, "n_tab": 15, "L_tab": 4297, "n_element_tab": 376, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 851, "n_element_tab_1": 131, "formula_len_all": 601, "formula_len_all_1": 527, "len_all": 267774, "len_all_1": 77738, "len_abs": 1182, "len_title": 127, "len_sents": 67878, "len_sents_1": 35520, "n_sents": 425, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1188, "title": "Segmenting Text and Learning Their Rewards for Improved RLHF in Language Models", "abs": "Reinforcement learning from human feedback (RLHF) has been widely adopted to align language models (LMs) with human preference.\nPrior RLHF works typically take a bandit formulation, which, though intuitive, ignores the sequential nature of LM generation and can suffer from the sparse reward issue.\nWhile recent works propose dense token-level RLHF, treating each token as an action may be oversubtle to proper reward assignment. \nIn this paper, we seek to get the best of both by training and utilizing a segment-level reward model, which assigns a reward to each semantically complete text segment that spans over a short sequence of tokens.\nFor reward learning, our method allows dynamic text segmentation and compatibility with standard sequence-preference datasets.\nFor effective RL-based LM training against segment reward, we generalize the classical scalar bandit reward normalizers into location-aware normalizer functions and interpolate the segment reward for further densification.\nWith these designs, our method performs competitively on popular RLHF benchmarks in both reward modeling and LM policy learning.\nAblation studies are conducted to further demonstrate our method.", "keywords": [ "Reinforcement Learning from Human Feedback", "Reward Modeling", "Dense Reward Assignment", "Language Models" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "cK0kUzocJW", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 4, "author_info_list": [ { "name": "Qingzhao Zhang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Shuowei Jin", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Xueshen Liu", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Zhuoqing Mao", "gender": "Female", "institution": "Google", "country": "US", "position": "consultant" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 29, "n_ref": 70, "n_ref_all": 85, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 46, "formula_len_all_1": 28, "len_all": 121289, "len_all_1": 63247, "len_abs": 1615, "len_title": 87, "len_sents": 40032, "len_sents_1": 31510, "n_sents": 306, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 39, "L_abs": 1319, "title": "Compute Or Load KV Cache? Why Not Both?", "abs": "Recent advancements in Large Language Models (LLMs) have significantly in-\ncreased context window sizes, enabling sophisticated applications but also in-\ntroducing substantial computational overheads, particularly computing key-value\n(KV) cache in the prefill stage. Prefix caching has emerged to save GPU power\nin this scenario, which saves KV cache at disks and reuse them across multiple\nqueries. However, traditional prefix caching mechanisms often suffer from sub-\nstantial latency because the speed of loading KV cache from disks to GPU mem-\nory is bottlenecked by the throughput of I/O devices. To optimize the latency of\nlong-context prefill, we propose Cake, a novel KV cache loader, which employs\na bidirectional parallelized KV cache generation strategy. Upon receiving a pre-\nfill task, Cake simultaneously and dynamically loads saved KV cache from prefix\ncache locations and computes KV cache on local GPUs, maximizing the utiliza-\ntion of available computation and I/O bandwidth resources. Additionally, Cake\nautomatically adapts to diverse system statuses without manual parameter. tuning.\nIn experiments on various prompt datasets, GPUs, and I/O devices, Cake offers\nup to 68.1% Time To First Token (TTFT) reduction compare with compute-only\nmethod and 94.6% TTFT reduction compare with I/O-only method.", "keywords": [ "Efficient LLM Serving", "LLM Serving System", "LLM Prefix Caching" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 2, 4, 2 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 5, 4, 2 ] }, { "paper_id": "cJn9HXPEpc", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 5, "author_info_list": [ { "name": "Jiashu Zhang", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Khuzaima Daudjee", "gender": "Not Specified", "institution": "University of Waterloo", "country": "", "position": "full Professor" }, { "name": "Molly Yiming Xu", "gender": "Not Specified", "institution": ", University of Waterloo", "country": "CA", "position": "Undergrad student" }, { "name": "Sihang Liu", "gender": "unknown", "institution": "University of Waterloo", "country": "CA", "position": "Assistant Professor" }, { "name": "Zihan Pan", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 21, "n_ref": 64, "n_ref_all": 83, "n_fig": 8, "n_tab": 5, "L_tab": 2305, "n_element_tab": 133, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 621, "n_element_tab_1": 78, "formula_len_all": 269, "formula_len_all_1": 64, "len_all": 134751, "len_all_1": 70808, "len_abs": 1189, "len_title": 100, "len_sents": 43924, "len_sents_1": 35213, "n_sents": 356, "n_sents_1": 272, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 1193, "title": "FreeRide: Harvesting Bubbles in Pipeline Parallelism", "abs": "The occurrence of bubbles in pipeline parallelism is an inherent limitation that can account for more than 40% of the large language model (LLM) training time and is one of the main reasons for the underutilization of GPU resources in LLM training. Harvesting these bubbles for GPU side tasks can increase resource utilization and reduce training costs but comes with challenges. First, because bubbles are discontinuous with various shapes, programming side tasks becomes difficult while requiring excessive engineering effort. Second, a side task can compete with pipeline training for GPU resources and incur significant overhead. To address these challenges, we propose FreeRide, a system designed to harvest bubbles in pipeline parallelism for side tasks. FreeRide provides programmers with interfaces to implement side tasks easily, manages bubbles and side tasks during pipeline training, and controls access to GPU resources by side tasks to reduce overhead. We demonstrate that FreeRide achieves about 8% average cost savings with a negligible overhead of about 1% for typical long training times of LLMs while serving model training, graph analytics, and image processing side tasks.", "keywords": [ "Pipeline parallelism", "bubbles" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "cJd1BgZ9CS", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Daniel Korat", "gender": "unknown", "institution": "Intel", "country": "IL", "position": "Researcher" }, { "name": "David Harel", "gender": "unknown", "institution": "Weizmann Institute of Science", "country": "", "position": "Full Professor" }, { "name": "Jonathan Mamou", "gender": "unknown", "institution": "Intel", "country": "", "position": "Deep Learning and NLP Researcher" }, { "name": "Michal Gordon-Kiwkowitz", "gender": "Female", "institution": "Holon Institute of Technology", "country": "", "position": "Instructor" }, { "name": "Moshe Berchansky", "gender": "Male", "institution": "Intel", "country": "", "position": "Researcher" }, { "name": "Moshe Wasserblat", "gender": "unknown", "institution": "Intel", "country": "", "position": "Researcher" }, { "name": "Nadav Timor", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "IL", "position": "PhD student" }, { "name": "Oren Pereg", "gender": "Male", "institution": "Intel", "country": "", "position": "Principal Researcher" }, { "name": "Tomer Galanti", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 14, "n_ref_uni": 48, "n_ref": 86, "n_ref_all": 112, "n_fig": 7, "n_tab": 1, "L_tab": 477, "n_element_tab": 12, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 45, "n_element_tab_1": 3, "formula_len_all": 1303, "formula_len_all_1": 1088, "len_all": 193396, "len_all_1": 78063, "len_abs": 1321, "len_title": 123, "len_sents": 58785, "len_sents_1": 37780, "n_sents": 518, "n_sents_1": 318, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1326, "title": "DSI: Faster Inference of Large Language Models via Speculation Parallelism", "abs": "Accelerating the inference of large language models (LLMs) is an important challenge in artificial intelligence. This paper introduces Distributed Speculative Inference (DSI), a novel distributed inference algorithm that is provably faster than speculative inference (SI) [leviathan2023fast, chen2023accelerating, miao2023specinfer] and traditional autoregressive inference (non-SI). Like other SI algorithms, DSI works on frozen LLMs, requiring no training or architectural modifications, and it preserves the target distribution. Prior studies on SI have demonstrated empirical speedups (compared to non-SI) but require fast and accurate drafters, which are often unavailable in practice. We identify a gap where SI can be slower than non-SI given slower or less accurate drafters. We close this gap by proving that DSI is faster than both SI and non-SI\u2014given any drafters. DSI introduces a novel type of task parallelism called Speculation Parallelism (SP), which orchestrates target and drafter instances to overlap in time, creating a new foundational tradeoff between computational resources and latency. DSI is not only faster than SI but also supports LLMs that cannot be accelerated with SI. Our simulations show speedups of off-the-shelf LLMs in realistic single-node settings where DSI is 1.29-1.92x faster than SI.", "keywords": [ "inference algorithms for generative models", "LLM inference", "speculative decoding" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "cJQ1K2fjpD", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Chenhang Cui", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Summer Camp Student" }, { "name": "Gelei Deng", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Huaxiu Yao", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Yiyang Zhou", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Zhaorun Chen", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "AN ZHANG", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Tat-seng Chua", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 41, "n_formula_1": 18, "n_ref_uni": 31, "n_ref": 116, "n_ref_all": 146, "n_fig": 12, "n_tab": 12, "L_tab": 4045, "n_element_tab": 477, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3553, "n_element_tab_1": 283, "formula_len_all": 2550, "formula_len_all_1": 1239, "len_all": 224446, "len_all_1": 70053, "len_abs": 1268, "len_title": 145, "len_sents": 60672, "len_sents_1": 27918, "n_sents": 537, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1280, "title": "Fine-Grained Verifiers: Preference Modeling as Next-token in Vision-Language Alignment", "abs": "The recent advancements in large language models (LLMs) and pre-trained vision models have accelerated the development of vision-language large models (VLLMs), enhancing the interaction between visual and linguistic modalities. Despite their notable success across various domains, VLLMs face challenges in modality alignment, which can lead to issues like hallucinations and unsafe content generation. Current alignment techniques often rely on coarse feedback and external datasets, limiting scalability and performance. In this paper, we propose FiSAO (Fine-Grained Self-Alignment Optimization), a novel self-alignment method that utilizes the model\u2019s own visual encoder as a fine-grained verifier to improve vision-language alignment without the need for additional data. By leveraging token-level feedback from the vision encoder, FiSAO significantly improves vision-language alignment, even surpassing traditional preference tuning methods that require additional data. Through both theoretical analysis and experimental validation, we demonstrate that FiSAO effectively addresses the misalignment problem in VLLMs, marking the first instance of token-level rewards being applied to such models. Our code is avaliable at \\url{https://anonymous.4open.science/r/FISAO-57F0/}.", "keywords": [ "Large Models; Alignment; Hallucination" ], "rating_list": [ 5, 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 4, 3 ] }, { "paper_id": "cJPUpL8mOw", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Alkis Sygkounas", "gender": "Male", "institution": "\u00d6rebro University", "country": "SE", "position": "PhD student" }, { "name": "Amy Loutfi", "gender": "Female", "institution": "\u00d6rebro University", "country": "", "position": "Full Professor" }, { "name": "Andreas Persson", "gender": "Male", "institution": "Centre for Applied Autonomous Sensor Systems", "country": "", "position": "Postdoc" }, { "name": "Pedro Zuidberg Dos Martires", "gender": "Male", "institution": "KU Leuven", "country": "", "position": "PhD student" }, { "name": "RISHI HAZRA", "gender": "Male", "institution": "\u00d6rebro University", "country": "SE", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 2, "n_ref_uni": 65, "n_ref": 116, "n_ref_all": 141, "n_fig": 11, "n_tab": 13, "L_tab": 5117, "n_element_tab": 360, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 258, "n_element_tab_1": 14, "formula_len_all": 1133, "formula_len_all_1": 78, "len_all": 264266, "len_all_1": 61929, "len_abs": 1233, "len_title": 121, "len_sents": 74297, "len_sents_1": 30280, "n_sents": 611, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1245, "title": "REvolve: Reward Evolution with Large Language Models using Human Feedback", "abs": "Designing effective reward functions is crucial to training reinforcement learning (RL) algorithms. However, this design is non-trivial, even for domain experts, due to the subjective nature of certain tasks that are hard to quantify explicitly. In recent works, large language models (LLMs) have been used for reward generation from natural language task descriptions, leveraging their extensive instruction tuning and commonsense understanding of human behavior. In this work, we hypothesize that LLMs, guided by human feedback, can be used to formulate reward functions that reflect human implicit knowledge. We study this in three challenging settings -- autonomous driving, humanoid locomotion, and dexterous manipulation -- wherein notions of ``good\" behavior are tacit and hard to quantify. To this end, we introduce REvolve, a truly evolutionary framework that uses LLMs for reward design in RL. REvolve generates and refines reward functions by utilizing human feedback to guide the evolution process, effectively translating implicit human knowledge into explicit reward functions for training (deep) RL agents. Experimentally, we demonstrate that agents trained on REvolve-designed rewards outperform other state-of-the-art baselines.", "keywords": [ "Evolutionary Algorithms", "Reward Design", "Reinforcement Learning", "Large Language Models" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 4, 4 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "cJ9qoVZbPd", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Hongda Sun", "gender": "Male", "institution": "Gaoling School of Artificial Intelligence, Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Rui Yan", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Songshi Liang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Yongbin Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yuchuan Wu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zihe Wang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Tony Lin", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 8, "n_ref_uni": 23, "n_ref": 52, "n_ref_all": 66, "n_fig": 2, "n_tab": 8, "L_tab": 4467, "n_element_tab": 548, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 2055, "n_element_tab_1": 256, "formula_len_all": 655, "formula_len_all_1": 463, "len_all": 145760, "len_all_1": 69416, "len_abs": 1427, "len_title": 99, "len_sents": 46130, "len_sents_1": 29682, "n_sents": 343, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1435, "title": "Locate-then-Unlearn: An Effective Method of Multi-Task Continuous Learning for Large Language Models", "abs": "Nowadays large language models (LLMs) have achieved remarkable success in\nvarious NLP tasks. However, they often misinterpret human instructions and generate incorrect or outdated responses, highlighting the need for more effective continual learning techniques. While recent efforts have introduced unlearning methods to remove erroneous knowledge, existing approaches still struggle in multi-task learning scenarios. To overcome these limitations, we propose Locate-then-unlearn, a new framework that identifies and selectively unlearns task-specific\nneurons to enable efficient multi-task learning. We hypothesize that LLM neurons can be broadly categorized into task-specific neurons for handling individual\ntasks, and general neurons to maintain the model\u2019s foundational capabilities. To\naccurately identify task-specific neurons, the locating process includes: (1) ranking task-related neurons based on their importance to each task, and (2) identifying\ntask-specific neurons by applying intervention to assess how neuron activity impacts task performance, isolating those most critical to each task. We conduct\ncomprehensive evaluations in two experimental setups: single-task specialization\nand multi-task generalization. The results show that our method significantly improves performance across both settings. This indicates that our method effectively balances model efficiency and accuracy in multi-task continual learning.", "keywords": [ "Machine unlearning", "Continue learning", "Model editing" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "cIKQp84vqN", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Hangqi Ding", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Hao Li", "gender": "Male", "institution": "Xidian University", "country": "", "position": "MS student" }, { "name": "Mingyang Zhang", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Qiguang Miao", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Professor" }, { "name": "Yue Wu", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Gong Maoguo", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "wenping ma", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 18, "n_ref_uni": 58, "n_ref": 106, "n_ref_all": 119, "n_fig": 4, "n_tab": 6, "L_tab": 2173, "n_element_tab": 292, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1417, "n_element_tab_1": 218, "formula_len_all": 1063, "formula_len_all_1": 997, "len_all": 134019, "len_all_1": 66618, "len_abs": 1637, "len_title": 109, "len_sents": 30662, "len_sents_1": 27544, "n_sents": 258, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1646, "title": "Adaptive Graduated Non-Convexity for Point Cloud Registration", "abs": "Point cloud registration is a critical and challenging task in computer vision. It is difficult to avoid poor local minima since the cost function is significantly non-convex. Correspondences tainted by significant or unknown outliers may cause the probability of finding a close-to-true transformation to drop rapidly, leading to point cloud registration failure. Many registration methods avoid local minima by updating the scale parameter of the cost function using graduated non-convexity (GNC). However, the update is usually performed in a fixed manner, resulting in limited accuracy and robustness of registration, and failure to reliably converge to the global minimum. Therefore, we present a novel method to robust point cloud registration based on Adaptive Graduated Non-Convexity (AGNC). By monitoring the positive definiteness of the Hessian of the cost function, the scale in graduated non-convexity is adaptively reduced without the need for a fixed optimization schedule. In addition, a multi-task knowledge sharing mechanism is used to achieve collaborative optimization of non-convex cost functions at different levels to further improve the success rate of point cloud registration under challenging high outlier conditions. Experimental results on simulated and real point cloud registration datasets show that AGNC far outperforms state-of-the-art methods in terms of robustness and accuracy, and can obtain promising registration results even in the case of extreme 99\\% outlier rates. To the best of our knowledge, this is the first study that explores point cloud registration considering adaptive graduated non-convexity.", "keywords": [ "Non-Convexity; Point Cloud; Registration" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "cHyQT6Y1jY", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Alexis Huet", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "FR", "position": "Researcher" }, { "name": "Dario Rossi", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "", "position": "Full Professor" }, { "name": "Giulio Franzese", "gender": "Male", "institution": "Eurecom", "country": "FR", "position": "Assistant Professor" }, { "name": "Pietro Michiardi", "gender": "Male", "institution": "EURECOM", "country": "", "position": "Full Professor" }, { "name": "Simone Clemente", "gender": "Male", "institution": "UBS Group AG", "country": "CH", "position": "Intern" }, { "name": "Zied Ben Houidi", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "FR", "position": "Principal Researcher" } ], "n_formula": 18, "n_formula_1": 5, "n_ref_uni": 37, "n_ref": 62, "n_ref_all": 108, "n_fig": 21, "n_tab": 12, "L_tab": 12416, "n_element_tab": 1656, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1372, "n_element_tab_1": 94, "formula_len_all": 1185, "formula_len_all_1": 582, "len_all": 219264, "len_all_1": 60363, "len_abs": 1555, "len_title": 152, "len_sents": 71068, "len_sents_1": 28782, "n_sents": 471, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1571, "title": "In Praise of Stubbornness: The Case for Cognitive-Dissonance Aware Continual Update of Knowledge in LLMs", "abs": "Despite remarkable capabilities, large language models (LLMs) struggle to continually update their knowledge without catastrophic forgetting. In contrast, humans effortlessly integrate new information, detect conflicts with existing beliefs, and selectively update their mental models. This paper introduces a cognitive-inspired investigation paradigm to study knowledge updating in LLMs. We implement two key components inspired by human cognition: (1) *Dissonance and Familiarity Awareness*, analyzing model behavior to classify information as novel, familiar, or dissonant; and (2) *Targeted Network Updates*, which track neural activity to identify frequently used (*stubborn*) and rarely used (*plastic*) neurons.\n\nThrough carefully designed experiments in controlled settings, we uncover a number of empirical findings demonstrating the potential of this approach. First, dissonance detection is feasible using simple activation and gradient features, suggesting potential for cognitive-inspired training. Second, we find that non-dissonant updates largely preserve prior knowledge regardless of targeting strategy, revealing inherent robustness in LLM knowledge integration. Most critically, we discover that dissonant updates prove catastrophically destructive to the model's knowledge base, indiscriminately affecting even information unrelated to the current updates. This suggests fundamental limitations in how neural networks handle contradictions and motivates the need for new approaches to knowledge updating that better mirror human cognitive mechanisms.", "keywords": [ "LLM", "knowledge editing", "continual learning", "Cognitive Science-Inspired AI", "Human-inspired AI", "Episodic Memory", "Incremental Learning", "Targeted Network Updates" ], "rating_list": [ 6, 3, 8, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "cHKuyeHmS9", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Wenguan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Xinhao Cai", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yazhou Yao", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Qiuxia LAI", "gender": "Female", "institution": "Communication University of China", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 16, "n_ref_uni": 28, "n_ref": 94, "n_ref_all": 121, "n_fig": 3, "n_tab": 8, "L_tab": 2681, "n_element_tab": 314, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1476, "n_element_tab_1": 171, "formula_len_all": 955, "formula_len_all_1": 1052, "len_all": 170215, "len_all_1": 69221, "len_abs": 1413, "len_title": 131, "len_sents": 42054, "len_sents_1": 30601, "n_sents": 337, "n_sents_1": 235, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1391, "title": "Cycle-Consistent Learning for Joint Layout-to-Image Generation and Object Detection", "abs": "In this paper, we propose a generation-detection cycle consistent (GDCC) learning framework that jointly optimizes both layout-to-image (L2I) generation and object detection (OD) tasks in an end-to-end manner. The key of GDCC lies in the inherent duality between the two tasks, where L2I takes all object boxes and labels as input conditions to generate images, and OD maps images back to these layout conditions. Specifically, in GDCC, L2I generation is guided by a layout translation cycle loss, ensuring that the layouts used to generate images align with those predicted from the synthesized images. Similarly, OD benefits from an image translation cycle loss, which enforces consistency between the synthesized images fed into the detector and those generated from predicted layouts. While current L2I and OD tasks benefit from large-scale annotated layout-image pairs, our GDCC enables more efficient use of unpaired layout data, thereby further enhancing data efficiency. It is worth noting that our GDCC framework is computationally efficient thanks to the perturbative single-step sampling strategy and a priority timestep re-sampling strategy during training, while maintaining the same inference cost as the original L2I and OD models. Extensive experiments demonstrate that GDCC significantly improves the controllability of diffusion models and the accuracy of object detectors.", "keywords": [ "Generation", "Detection", "Cycle-consistent", "Diffusion" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 5, 3, 5 ] }, { "paper_id": "cH65nS5sOz", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 9, "author_info_list": [ { "name": "Carl Yang", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Assistant Professor" }, { "name": "Chanyoung Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Junseok Lee", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Namkyeong Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Sein Kim", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Sukwon Yun", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Sungwon Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Yoonho Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Yunhak Oh", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 40, "n_formula_1": 20, "n_ref_uni": 38, "n_ref": 113, "n_ref_all": 147, "n_fig": 8, "n_tab": 12, "L_tab": 16781, "n_element_tab": 898, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 8281, "n_element_tab_1": 371, "formula_len_all": 2969, "formula_len_all_1": 2465, "len_all": 247930, "len_all_1": 87361, "len_abs": 2911, "len_title": 100, "len_sents": 78877, "len_sents_1": 33978, "n_sents": 612, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 52, "L_abs": 1321, "title": "Subgraph Federated Learning for Local Generalization", "abs": "Federated Learning (FL) on graphs enables collaborative model training to enhance performance without compromising the privacy of each client. However, existing methods often overlook the mutable nature of graph data, which frequently introduces new nodes and leads to shifts in label distribution. Since they focus solely on performing well on each client's local data, they are prone to overfitting to their local distributions (i.e., local overfitting), which hinders their ability to generalize to unseen data with diverse label distributions. In contrast, our proposed method, FedLoG, effectively tackles this issue by mitigating local overfitting. Our model generates global synthetic data by condensing the reliable information from each class representation and its structural information across clients. Using these synthetic data as a training set, we alleviate the local overfitting problem by adaptively generalizing the absent knowledge within each local dataset. This enhances the generalization capabilities of local models, enabling them to handle unseen data effectively. Our model outperforms baselines in our proposed experimental settings, which are designed to measure generalization power to unseen data in practical scenarios. \nOur code is available at https://anonymous.4open.science/r/FedLoG-89EE", "keywords": [ "Graph Neural Networks", "Graph Federated Learning" ], "rating_list": [ 10, 6, 8, 6, 8 ], "soundness_list": [ 4, 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3, 4 ], "contribution_list": [ 4, 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 4, 3 ] }, { "paper_id": "cH4VTcCVYs", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Chenguang Li", "gender": "Female", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Gabriel Kreiman", "gender": "Male", "institution": "Harvard Medical School", "country": "US", "position": "Full Professor" }, { "name": "Jonah Brenner", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Undergrad student" } ], "n_formula": 31, "n_formula_1": 18, "n_ref_uni": 37, "n_ref": 52, "n_ref_all": 77, "n_fig": 8, "n_tab": 1, "L_tab": 122, "n_element_tab": 10, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1494, "formula_len_all_1": 670, "len_all": 137369, "len_all_1": 63721, "len_abs": 665, "len_title": 110, "len_sents": 48458, "len_sents_1": 27807, "n_sents": 460, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 619, "title": "Policy optimization emerges from noisy representation learning", "abs": "Nervous systems learn representations of the world and policies to act within it. We present a framework that uses reward-dependent noise to facilitate policy optimization in representation learning networks. These networks balance extracting normative features and task-relevant information to solve tasks. Moreover, their representation changes reproduce several experimentally observed shifts in the neural code during task learning. Our framework presents a biologically plausible mechanism for emergent policy optimization amid evidence that representation learning plays a vital role in governing neural dynamics.", "keywords": [ "natural intelligence", "reinforcement learning", "representation learning", "noise" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "cGu5LtGcRD", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Chen Chen", "gender": "Female", "institution": "Qiyuan Lab", "country": "CN", "position": "Researcher" }, { "name": "Daming Shi", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Xiangyang Ji", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yixiu Mao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 73, "n_formula_1": 25, "n_ref_uni": 23, "n_ref": 67, "n_ref_all": 103, "n_fig": 18, "n_tab": 3, "L_tab": 1336, "n_element_tab": 227, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8969, "formula_len_all_1": 1798, "len_all": 161714, "len_all_1": 60320, "len_abs": 1480, "len_title": 135, "len_sents": 48081, "len_sents_1": 27207, "n_sents": 400, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1492, "title": "Adaptive HL-Gaussian: A Value Function Learning Method with Dynamic Support Adjustment", "abs": "Recent research indicates that using cross-entropy (CE) loss for value function learning surpasses traditional mean squared error (MSE) loss in performance and scalability, with the HL-Gaussian method showing notably strong results. However, this method requires a pre-specified support for representing the categorical distribution of the value function, and an inappropriately chosen interval for the support may not match the time-varying value function, potentially impeding the learning process. To address this issue, we theoretically establish that HL-Gaussian inherently introduces a projection error during the learning of the value function, which is dependent on the support interval. We further prove that an ideal interval should be sufficiently broad to reduce truncation-induced projection errors, yet not so excessive as to counterproductively amplify them. Guided by these findings, we introduce the Adaptive HL-Gaussian (AHL-Gaussian) approach. This approach starts with a confined support interval and dynamically adjusts its range by minimizing the projection error. This ensures that the interval's size stabilizes to adapt to the learning value functions without further expansion. We integrate AHL-Gaussian into several classic value-based algorithms and evaluate it on Atari 2600 games and Gym Mujoco. The results show that AHL-Gaussian significantly outperforms the vanilla baselines and standard HL-Gaussian with a static interval across the majority of tasks.", "keywords": [ "value function learning", "HL-Gaussian" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 3, 1, 4, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "cFu7ze7xUm", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Guangxuan Xiao", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Haotian Tang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Jiaming Tang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jingwei Zuo", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Shang Yang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Song Han", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yao Fu", "gender": "Male", "institution": "University of Edinburgh", "country": "", "position": "PhD student" }, { "name": "junxian guo", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 41, "n_ref": 103, "n_ref_all": 150, "n_fig": 18, "n_tab": 7, "L_tab": 4222, "n_element_tab": 789, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 1666, "n_element_tab_1": 43, "formula_len_all": 1025, "formula_len_all_1": 911, "len_all": 205619, "len_all_1": 61965, "len_abs": 1584, "len_title": 133, "len_sents": 53649, "len_sents_1": 29551, "n_sents": 337, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1656, "title": "DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads", "abs": "Deploying long-context large language models (LLMs) is essential but poses significant computational and memory challenges.\nCaching all Key and Value (KV) states across all attention heads consumes substantial memory.\nExisting KV cache pruning methods either damage the long-context capabilities of LLMs or offer only limited efficiency improvements.\nIn this paper, we identify that only a fraction of attention heads, a.k.a, Retrieval Heads, are critical for processing long contexts and require full attention across all tokens.\nIn contrast, all other heads, which primarily focus on recent tokens and attention sinks\u2014referred to as Streaming Heads\u2014do not require full attention.\nBased on this insight, we introduce DuoAttention, a framework that only applies a full KV cache to retrieval heads while using a light-weight, constant-length KV cache for streaming heads, which reduces both LLM's decoding and pre-filling memory and latency without compromising its long-context abilities.\nDuoAttention uses a lightweight, optimization-based algorithm with synthetic data to identify retrieval heads accurately.\nOur method significantly reduces long-context inference memory by up to 2.55$\\times$ for MHA and 1.67$\\times$ for GQA models while speeding up decoding by up to 2.18$\\times$ and 1.50$\\times$ and accelerating pre-filling by up to 1.73$\\times$ and 1.63$\\times$ for MHA and GQA models, respectively, with minimal accuracy loss compared to full attention.\nNotably, combined with quantization, DuoAttention enables Llama-3-8B decoding with 3.33 million context length measured on a single A100 GPU. Code and dataset will be released upon publication.", "keywords": [ "Large Language Models; Long Context; Efficiency;" ], "rating_list": [ 6, 8, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 4, 2 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 5, 5, 4, 4, 4 ] }, { "paper_id": "cF6OoaYcRa", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Ander Gray", "gender": "Male", "institution": "UK Atomic Energy Authority", "country": "", "position": "Researcher" }, { "name": "Daniel Giles", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Postdoc" }, { "name": "Lorenzo Zanisi", "gender": "unknown", "institution": "UK atomic energy authority", "country": "", "position": "Data scientist" }, { "name": "Matt J Kusner", "gender": "Male", "institution": "University College London", "country": "", "position": "Associate Professor" }, { "name": "Stanislas Pamela", "gender": "Male", "institution": "UKAEA ", "country": "", "position": "Researcher" }, { "name": "Vignesh Gopakumar", "gender": "Male", "institution": "UK Atomic Energy Authority", "country": "GB", "position": "Researcher" }, { "name": "Marc Deisenroth", "gender": "Male", "institution": "Google", "country": "GB", "position": "Principal Researcher" } ], "n_formula": 49, "n_formula_1": 18, "n_ref_uni": 54, "n_ref": 82, "n_ref_all": 153, "n_fig": 16, "n_tab": 11, "L_tab": 3881, "n_element_tab": 225, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 896, "n_element_tab_1": 11, "formula_len_all": 2380, "formula_len_all_1": 841, "len_all": 219162, "len_all_1": 59070, "len_abs": 2497, "len_title": 102, "len_sents": 55916, "len_sents_1": 26156, "n_sents": 471, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1007, "title": "Calibrated Physics-Informed Uncertainty Quantification", "abs": "Neural PDEs have emerged as inexpensive surrogate models for numerical PDE solvers. While they offer efficient approximations, they often lack robust uncertainty quantification (UQ), limiting their practical utility. Existing UQ methods for these models typically have high computational demands and lack guarantees. We introduce a novel framework for calibrated physics-informed uncertainty quantification to address these limitations. Our approach leverages physics residual errors as a nonconformity score within a conformal prediction (CP) framework. This enables data-free, model-agnostic, and statistically guaranteed uncertainty estimates. Our framework utilises convolutional layers as finite difference stencils for gradient estimation, our framework provides inexpensive coverage bounds for the violation of conservation laws within model predictions. In our experiments, we utilise CP to obtain marginal coverage for each cell and joint coverage over the entire prediction domain of various PDEs.", "keywords": [ "Surrogate Models", "Uncertainty Quantification", "Neural-PDE", "Physics-Informed", "Conformal Prediction" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 1, 1, 3 ], "presentation_list": [ 2, 2, 2, 4 ], "contribution_list": [ 2, 1, 1, 4 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "cDdeTXOnAK", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Bin Lei", "gender": "unknown", "institution": "University of Connecticut", "country": "", "position": "PhD student" }, { "name": "Caiwen Ding", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "Associate Professor" }, { "name": "Shan Zuo", "gender": "Female", "institution": "University of Connecticut", "country": "", "position": "Assistant Professor" }, { "name": "Yi Zhang", "gender": "Male", "institution": "University of Connecticut", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 16, "n_ref": 54, "n_ref_all": 67, "n_fig": 5, "n_tab": 7, "L_tab": 3250, "n_element_tab": 477, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2766, "n_element_tab_1": 302, "formula_len_all": 25, "formula_len_all_1": 0, "len_all": 124585, "len_all_1": 56661, "len_abs": 854, "len_title": 113, "len_sents": 38055, "len_sents_1": 23123, "n_sents": 282, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 745, "title": "AutoCoder: Enhancing Code Large Language Model with AIEV-INSTRUCT", "abs": "We introduce AutoCoder, an open-source Large Language Model to surpass GPT-4 Turbo and GPT-4o in pass@1 on the Human Eval benchmark test (90.9\\% vs. 90.2). In addition, AutoCoder offers a more versatile code interpreter compared to GPT-4 Turbo and GPT-4o. It's code interpreter can install external packages instead of limiting to built-in packages. AutoCoder's training data is a multi-turn dialogue dataset created by a system combining agent interaction and external code execution verification, a method we term AIEV-Instruct (Agent-Interaction Execution-Verified). Compared to previous large-scale code dataset annotation methods, AIEV-Instruct reduces dependence on proprietary large models and provides more accurate code annotation data.", "keywords": [ "LLM", "Code generation", "data annotation", "Agents interaction" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "cDd7kg9mkP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Hao Xue", "gender": "Male", "institution": "University of New South Wales", "country": "", "position": "Postdoc" }, { "name": "Linyao Chen", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Shohreh Deldari", "gender": "Female", "institution": "University of New South Wales", "country": "AU", "position": "Postdoc" }, { "name": "Zechen Li", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "PhD student" }, { "name": "Flora Salim", "gender": "Female", "institution": "University of New South Wales", "country": "AU", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 1, "n_ref_uni": 50, "n_ref": 100, "n_ref_all": 123, "n_fig": 4, "n_tab": 8, "L_tab": 2651, "n_element_tab": 254, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1235, "n_element_tab_1": 104, "formula_len_all": 516, "formula_len_all_1": 26, "len_all": 218306, "len_all_1": 72779, "len_abs": 1659, "len_title": 141, "len_sents": 67453, "len_sents_1": 33616, "n_sents": 520, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1665, "title": "SensorLLM: Aligning Large Language Models with Motion Sensors for Human Activity Recognition", "abs": "In this work, we bridge the gap between wearable sensor technology and personalized AI assistants by enabling Large Language Models (LLMs) to understand time-series tasks like human activity recognition (HAR). Despite the strong reasoning and generalization capabilities of LLMs, leveraging them for sensor data tasks remains largely unexplored. This gap stems from challenges like the lack of semantic context in time-series data, computational limitations, and LLMs' difficulty processing numerical inputs. To address these issues, we introduce SensorLLM, a two-stage framework to unlock LLMs\u2019 potential for sensor data tasks. In the Sensor-Language Alignment Stage, we introduce special tokens for each sensor channel and automatically generate trend-descriptive text to align sensor data with textual inputs, enabling SensorLLM to capture numerical changes, channel-specific information, and sensor data of varying lengths\u2014capabilities that existing LLMs typically struggle with, all without the need for human annotations. Next, in Task-Aware Tuning Stage, we refine the model for HAR classification using the frozen LLM and alignment module, achieving performance on par with or surpassing state-of-the-art models. We further demonstrate that SensorLLM evolves into an effective sensor learner, reasoner, and classifier through Sensor-Language Alignment, enabling it to generalize across diverse datasets for HAR tasks. We strongly believe our work lays the stepstone for future time-series and text alignment research, offering a path toward foundation models for sensor data. Our codes are available at https://anonymous.4open.science/r/sensorllm_code-E0FC.", "keywords": [ "Large language models", "Multi-modal", "alignment", "human activity recognition", "sensor", "time-series" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "cD1kl2QKv1", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Jian Yang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Kai Wang", "gender": "Male", "institution": "Computer Vision Center, Universitat Aut\u00f3noma de Barcelona", "country": "ES", "position": "Postdoc" }, { "name": "Ming-Ming Cheng", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Full Professor" }, { "name": "Senmao Li", "gender": "unknown", "institution": "Nankai University", "country": "TD", "position": "PhD student" }, { "name": "Shiqi Yang", "gender": "Male", "institution": "SB Intuitions, SoftBank", "country": "JP", "position": "Chief Research Scientist" }, { "name": "Tao Liu", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "MS student" }, { "name": "Yaxing Wang", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Fahad Shahbaz Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" }, { "name": "Joost van de", "gender": "Male", "institution": "Universitat Aut\u00f3noma de Barcelona", "country": "ES", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 52, "n_ref": 111, "n_ref_all": 162, "n_fig": 24, "n_tab": 4, "L_tab": 1392, "n_element_tab": 164, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 890, "n_element_tab_1": 110, "formula_len_all": 1911, "formula_len_all_1": 938, "len_all": 196460, "len_all_1": 72622, "len_abs": 1343, "len_title": 138, "len_sents": 53965, "len_sents_1": 30626, "n_sents": 427, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 90, "L_abs": 1492, "title": "One-Prompt-One-Story: Free-Lunch Consistent Text-to-Image Generation Using a Single Prompt", "abs": "Text-to-image generation models can create high-quality images from input prompts. However, they struggle to support the consistent generation of identity-preserving requirements for storytelling. Existing approaches to this problem typically require extensive training in large datasets or additional modifications to the original model architectures. This limits their applicability across different domains and diverse diffusion model configurations. In this paper, we first observe the inherent capability of language models, coined $\\textit{context consistency}$, to comprehend identity through context with a single prompt. Drawing inspiration from the inherent $\\textit{context consistency}$, we propose a novel $\\textit{training-free}$ method for consistent text-to-image (T2I) generation, termed \"One-Prompt-One-Story\" ($\\textit{1Prompt1Story}$). Our approach $\\textit{1Prompt1Story}$ concatenates all prompts into a single input for T2I diffusion models, initially preserving character identities. We then refine the generation process using two novel techniques: $\\textit{Singular-Value\nReweighting}$ and $\\textit{Identity-Preserving Cross-Attention}$, ensuring better alignment with the input description for each frame. In our experiments, we compare our method against various existing consistent T2I generation approaches to demonstrate its effectiveness, through quantitative metrics and qualitative assessments. Code is available at https://github.com/byliutao/1Prompt1Story.", "keywords": [ "diffusion model; consistent T2I image generation; storytelling" ], "rating_list": [ 8, 8, 8, 8, 5 ], "soundness_list": [ 3, 4, 4, 3, 3 ], "presentation_list": [ 4, 4, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 5, 4, 3, 4, 4 ] }, { "paper_id": "cCl10IU836", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Jack Brady", "gender": "unknown", "institution": "Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Sebastien Lachapelle", "gender": "Male", "institution": "Samsung", "country": "CA", "position": "Researcher" }, { "name": "Simon Buchholz", "gender": "unknown", "institution": "Max-Planck Institute", "country": "DE", "position": "Postdoc" }, { "name": "Wieland Brendel", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen", "country": "DE", "position": "Principal Researcher" }, { "name": "Julius Von K\u00fcgelgen", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Thomas N. Kipf", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" } ], "n_formula": 192, "n_formula_1": 14, "n_ref_uni": 99, "n_ref": 230, "n_ref_all": 267, "n_fig": 9, "n_tab": 3, "L_tab": 2429, "n_element_tab": 282, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 203, "n_element_tab_1": 3, "formula_len_all": 21349, "formula_len_all_1": 1239, "len_all": 450403, "len_all_1": 87039, "len_abs": 1609, "len_title": 79, "len_sents": 128424, "len_sents_1": 38761, "n_sents": 1334, "n_sents_1": 348, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1623, "title": "Interaction Asymmetry: A General Principle for Learning Composable Abstractions", "abs": "Learning disentangled representations of concepts and re-composing them in unseen ways is crucial for generalizing to out-of-domain situations. However, the underlying properties of concepts that enable such disentanglement and compositional generalization remain poorly understood. In this work, we propose the principle of interaction asymmetry which states: \"Parts of the same concept have more complex interactions than parts of different concepts\". We formalize this via block diagonality conditions on the $(n+1)$th order derivatives of the generator mapping concepts to observed data, where different orders of \"complexity\" correspond to different $n$. Using this formalism, we prove that interaction asymmetry enables both disentanglement and compositional generalization. Our results unify recent theoretical results for learning concepts of objects, which we show are recovered as special cases with $n=0$ or $1$. We provide results for up to $n=2$, thus extending these prior works to more flexible generator functions, and conjecture that the same proof strategies generalize to larger $n$. Practically, our theory suggests that, to disentangle concepts, an autoencoder should penalize its latent capacity and the interactions between concepts during decoding. We propose an implementation of these criteria using a flexible Transformer-based VAE, with a novel regularizer on the attention weights of the decoder. On synthetic image datasets consisting of objects, we provide evidence that this model can achieve comparable object disentanglement to existing models that use more explicit object-centric priors.", "keywords": [ "disentanglement", "compositional generalization", "representation learning", "object-centric learning", "identifiability", "unsupervised learning", "out-of-domain generalization" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "cCcaJzPAnb", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Akiva Block", "gender": "Male", "institution": "Computer Science Department, Technion - Israel Institute of Technology", "country": "IL", "position": "MS student" }, { "name": "Assaf Schuster", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "Full Professor" } ], "n_formula": 86, "n_formula_1": 21, "n_ref_uni": 48, "n_ref": 116, "n_ref_all": 131, "n_fig": 4, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 17406, "formula_len_all_1": 2562, "len_all": 219026, "len_all_1": 59467, "len_abs": 1176, "len_title": 101, "len_sents": 59935, "len_sents_1": 28465, "n_sents": 431, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1180, "title": "Universal Concavity-Aware Descent Rate for Optimizers", "abs": "Many machine learning problems involve a challenging task of calibrating parameters in a computational model to fit the training data; this task is especially challenging for non-convex problems. Many optimization algorithms have been proposed to assist in calibrating these parameters, each with its respective advantages in different scenarios, but it is often difficult to determine the scenarios for which an algorithm is best suited. To contend with this challenge, much work has been done on proving the rate at which these optimizers converge to their final solution, however the wide variety of such convergence rate bounds, each with their own different assumptions, convergence metrics, tightnesses, and parameters (which may or may not be known to the practitioner) make comparing these convergence rates difficult. To help with this problem, we present a minmax-optimal algorithm and, by comparison to it, give a single descent bound which is applicable to a very wide family of optimizers, tasks, and data (including all of the most prevalent ones), which also puts special emphasis on being tight even in parameter subspaces in which the cost function is concave.", "keywords": [ "nonconvex optimization", "optimization", "convergence rate", "objective function sub-optimality", "objective sub-optimality", "quasi newton optimization", "numerical methods", "lipschitz adaptive", "eigenspace lipschitz" ], "rating_list": [ 3, 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 1, 1, 2, 2, 1 ], "confidence_list": [ 4, 4, 2, 4, 3 ] }, { "paper_id": "cCRlEvjrx4", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Gautam Sreekumar", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Sachit Gaudi", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "MS student" }, { "name": "Vishnu Naresh Boddeti", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 77, "n_formula_1": 22, "n_ref_uni": 31, "n_ref": 79, "n_ref_all": 178, "n_fig": 25, "n_tab": 7, "L_tab": 941, "n_element_tab": 18, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 239, "n_element_tab_1": 4, "formula_len_all": 6595, "formula_len_all_1": 1949, "len_all": 229011, "len_all_1": 78708, "len_abs": 2466, "len_title": 105, "len_sents": 83417, "len_sents_1": 33687, "n_sents": 694, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 1296, "title": "Learning Conditionally Independent Marginals Enables Logical Compositions in Conditional Diffusion Models", "abs": "How can we learn generative models to sample data with arbitrary logical compositions of statistically independent attributes? The prevailing solution is to sample from distributions expressed as a composition of attributes' conditional marginal distributions under the assumption that they are statistically independent. This paper shows that standard conditional diffusion models violate this assumption, even when all attribute compositions are observed during training. And, this violation is significantly more severe when only a subset of the compositions is observed. We propose CoInD to address this problem. It explicitly enforces statistical independence between the conditional marginal distributions by minimizing Fisher\u2019s divergence between the joint and marginal distributions. The theoretical advantages of CoInD are reflected in both qualitative and quantitative experiments, demonstrating a significantly more faithful and controlled generation of samples for arbitrary logical compositions of attributes. The benefit is more pronounced for scenarios that current solutions relying on the assumption of conditionally independent marginals struggle with, namely, logical compositions involving the NOT operation and when only a subset of compositions are observed during training.", "keywords": [ "logical compositionality", "generative models", "diffusion models" ], "rating_list": [ 8, 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 4, 2, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "cC3LxGZasH", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Alexia Jolicoeur-Martineau", "gender": "Female", "institution": "Samsung - SAIT AI Lab, Montreal", "country": "CA", "position": "Researcher" }, { "name": "Christopher J Pal", "gender": "unknown", "institution": "Polytechnique Montreal", "country": "CA", "position": "Full Professor" }, { "name": "Ge Ya Luo", "gender": "Female", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "Gian Mario Favero", "gender": "unknown", "institution": "McGill University", "country": "CA", "position": "MS student" }, { "name": "ZhiHao Luo", "gender": "Male", "institution": "Polytechnique Montreal", "country": "", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 3, "n_ref_uni": 67, "n_ref": 122, "n_ref_all": 158, "n_fig": 18, "n_tab": 5, "L_tab": 1170, "n_element_tab": 95, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 153, "n_element_tab_1": 12, "formula_len_all": 1417, "formula_len_all_1": 152, "len_all": 216311, "len_all_1": 58301, "len_abs": 2103, "len_title": 128, "len_sents": 68984, "len_sents_1": 29176, "n_sents": 461, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1143, "title": "Beyond FVD: An Enhanced Evaluation Metrics for Video Generation Distribution Quality", "abs": "The Fr\u00e9chet Video Distance (FVD) is a widely adopted metric for evaluating video generation distribution quality. However, its effectiveness relies on critical assumptions. Our analysis reveals three significant limitations: (1) the non-Gaussianity of the Inflated 3D Convnet (I3D) feature space; (2) the insensitivity of I3D features to temporal distortions; (3) the impractical sample sizes required for reliable estimation. These findings undermine FVD's reliability and show that FVD falls short as a standalone metric for video generation evaluation. After extensive analysis of a wide range of metrics and backbone architectures, we propose JEDi, the JEPA Embedding Distance, \nbased on features derived from a Joint Embedding Predictive Architecture, measured using Maximum Mean Discrepancy with polynomial kernel. Our experiments on multiple open-source datasets show clear evidence that it is a superior alternative to the widely used FVD metric, requiring only 16% of the samples to reach its steady value, while increasing alignment with human evaluation by 34%, on average.\n\nProject page: https://oooolga.github.io/JEDi.github.io/.", "keywords": [ "Video quality metrics", "Frechet Video Distance", "Inflated 3D Convnet", "VideoMAE", "VJEPA", "kernel metrics" ], "rating_list": [ 6, 8, 6, 8, 3 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 5, 4 ] }, { "paper_id": "cADpvQgnqg", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Jeffrey Gu", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Serena Yeung", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 48, "n_ref": 146, "n_ref_all": 161, "n_fig": 4, "n_tab": 4, "L_tab": 1374, "n_element_tab": 77, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3639, "n_element_tab_1": 164, "formula_len_all": 245, "formula_len_all_1": 194, "len_all": 125297, "len_all_1": 55463, "len_abs": 1334, "len_title": 171, "len_sents": 30179, "len_sents_1": 24747, "n_sents": 166, "n_sents_1": 135, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 121, "L_abs": 1336, "title": "Foundation Models Secretly Understand Neural Network Weights: Enhancing Hypernetwork Architectures with Foundation Models", "abs": "Large pre-trained models, or foundation models, have shown impressive performance when adapted to a variety of downstream tasks, often out-performing specialized models. Hypernetworks, neural networks that generate some or all of the parameters of another neural network, have become an increasingly important technique for conditioning and generalizing implicit neural representations (INRs), which represent signals or objects such as audio or 3D shapes using a neural network. However, despite the potential benefits of incorporating foundation models in hypernetwork methods, this research direction has not been investigated, likely due to the dissimilarity of the weight generation task with other visual tasks. To address this gap, we (1) show how foundation models can improve hypernetworks with Transformer-based architectures, (2) provide an empirical analysis of the benefits of foundation models for hypernetworks through the lens of the generalizable INR task, showing that leveraging foundation models improves performance, generalizability, and data efficiency across a variety of algorithms and modalities. We also provide further analysis in examining the design space of foundation model-based hypernetworks, including examining the choice of foundation models, algorithms, and the effect of scaling foundation models.", "keywords": [ "neural fields", "neural implicit representations", "foundation models", "hypernetworks", "generalizable INR" ], "rating_list": [ 8, 5, 6, 3 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "cADdVJYiIG", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Rui Yin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Tao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tianyu Du", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "peng sun", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 7, "n_ref_uni": 66, "n_ref": 117, "n_ref_all": 141, "n_fig": 9, "n_tab": 9, "L_tab": 1523, "n_element_tab": 167, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 3473, "n_element_tab_1": 51, "formula_len_all": 1532, "formula_len_all_1": 307, "len_all": 200369, "len_all_1": 64436, "len_abs": 1596, "len_title": 71, "len_sents": 55308, "len_sents_1": 29651, "n_sents": 379, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 23, "L_abs": 1603, "title": "Data-Evolution Learning", "abs": "Recent advancements in machine learning have been driven by models trained on large-scale, high-quality datasets. However, the practical application of these models faces two significant challenges: the infeasibility of acquiring precise labels in real-world settings and the substantial computational burden imposed by training large models. While existing approaches\u2014such as self-supervised learning, weak supervision, noisy label learning, and dataset distillation\u2014address these challenges from a model-centric perspective, they often overlook the potential benefits of optimizing the data itself.\nThis paper introduces a novel data-centric learning paradigm where both the dataset and the model co-evolve during the learning process.\nWe formalize this paradigm and propose a Data-evolution Learning Algorithm (DeLA), which offers three key advantages: optimized dataset generation, versatile dataset compatibility, and effective utilization of prior knowledge.\nExtensive experiments demonstrate that DeLA enables the creation of optimized datasets for reuse in subsequent training, effectively addressing diverse datasets with varying target types. Moreover, DeLA accelerates learning by utilizing architecture-agnostic, open-source prior models for efficient data creation.\nNotably, DeLA frequently outperforms traditional SOTA model-centric methods in self-supervised and noisy label learning.\nFurthermore, its simplicity enables implementation in only two lines of PyTorch code, offering significant potential for advancements in representation learning.\nOur code will be made publicly available.", "keywords": [ "Data-centric Learning" ], "rating_list": [ 3, 1, 1, 5 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 1, 2, 1, 3 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "cA8iQJFioL", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Chenghao Huang", "gender": "Male", "institution": "Monash University", "country": "", "position": "PhD student" }, { "name": "Han Hu", "gender": "Male", "institution": "Monash University", "country": "", "position": "PhD student" }, { "name": "JohnMichael Jurgensen", "gender": "unknown", "institution": "University of Chicago", "country": "", "position": "PhD student" }, { "name": "Yinzhi Xu", "gender": "Female", "institution": "University of Chicago", "country": "", "position": "PhD student" }, { "name": "Yue Yang", "gender": "Female", "institution": "Monash University", "country": "AU", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 34, "n_ref": 44, "n_ref_all": 67, "n_fig": 10, "n_tab": 6, "L_tab": 6029, "n_element_tab": 250, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 738, "n_element_tab_1": 74, "formula_len_all": 7, "formula_len_all_1": 0, "len_all": 173557, "len_all_1": 60382, "len_abs": 1169, "len_title": 128, "len_sents": 59231, "len_sents_1": 30691, "n_sents": 424, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1174, "title": "InterIDEAS: An LLM and Expert-Enhanced Dataset for Philosophical Intertextuality", "abs": "The formation and circulation of ideas in philosophy have profound implications for pedagogical and scholarly practices. However, traditional analyses often depend on manual reading and subjective interpretation, constrained by human cognitive limits. To address these challenges, we introduce InterIDEAS, a pioneering dataset designed to bridge philosophy and natural language processing (NLP). By merging theories of intertextuality from literary studies with bibliometric techniques and recent LLMs, InterIDEAS enables both quantitative and qualitative analysis of the intellectual, social, and historical relations embedded within these difficult-to-interpret philosophical texts. This dataset not only enhances the study of philosophy but also contributes to the development of language models by providing a training corpus that challenges and enhances their interpretative capacity. InterIDEAS covers over 45,000 pages from key philosophical texts, spanning major thoughts and schools from 1750 to 1950, and features more than 3,150 writers. It manifests the mutual contribution between philosophy and NLP, laying the groundwork for future interdisciplinary research.", "keywords": [ "dataset", "natural language processing", "digital humanity", "large language model", "education" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "c9z65sDx6M", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 11, "author_info_list": [ { "name": "Ang Li", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Associate Professor" }, { "name": "Chuan Liu", "gender": "Female", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Chuang Ren", "gender": "unknown", "institution": "University of Rochester", "country": "US", "position": "Full Professor" }, { "name": "Chunshu Wu", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Postdoc" }, { "name": "Dongfang Liu", "gender": "unknown", "institution": "Rochester Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "James Chenhao Liang", "gender": "Male", "institution": "U. S. Naval Research Laboratory", "country": "US", "position": "Researcher" }, { "name": "Michael Huang", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Full Professor" }, { "name": "Mingkai Chen", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Tong Geng", "gender": "Male", "institution": "University of Rochester", "country": "", "position": "Assistant Professor" }, { "name": "shihui cao", "gender": "unknown", "institution": "University of Rochester", "country": "US", "position": "Postdoc" }, { "name": "Yingnian Wu", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 63, "n_ref": 79, "n_ref_all": 102, "n_fig": 6, "n_tab": 7, "L_tab": 2131, "n_element_tab": 146, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1101, "n_element_tab_1": 108, "formula_len_all": 295, "formula_len_all_1": 172, "len_all": 161094, "len_all_1": 68766, "len_abs": 1755, "len_title": 139, "len_sents": 49638, "len_sents_1": 34084, "n_sents": 318, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1765, "title": "Diff-PIC: Revolutionizing Particle-In-Cell Nuclear Fusion Simulation with Diffusion Models", "abs": "The rapid development of AI highlights the pressing need for sustainable energy, a critical global challenge for decades. Nuclear fusion, generally seen as an ultimate solution, has been the focus of intensive research for nearly a century, with investments reaching hundreds of billions of dollars. Recent advancements in Inertial Confinement Fusion have drawn significant attention to fusion research, in which Laser-Plasma Interaction (LPI) is critical for ensuring fusion stability and efficiency. However, the complexity of LPI upon fusion ignition makes analytical approaches impractical, leaving researchers depending on extremely computation-demanding Particle-in-Cell (PIC) simulations to generate data, presenting a significant bottleneck to advancing fusion research. In response, this work introduces Diff-PIC, a novel framework that leverages conditional diffusion models as a computationally efficient alternative to PIC simulations for generating high-fidelity scientific LPI data. In this work, physical patterns captured by PIC simulations are distilled into diffusion models associated with two tailored enhancements: (1) To effectively capture the complex relationships between physical parameters and corresponding outcomes, the parameters are encoded in a physically-informed manner. (2) To further enhance efficiency while maintaining high fidelity and physical validity, the rectified flow technique is employed to transform our model into a one-step conditional diffusion model. Experimental results show that Diff-PIC achieves 16,200 times speedup compared to traditional PIC on a 100 picosecond simulation, with an average reduction in MAE / RMSE / FID of 59.21% / 57.15% / 39.46% with respect to two other SOTA data generation approaches.", "keywords": [ "Diffusion Models", "PIC Simulations", "Synthetic Data Generation" ], "rating_list": [ 8, 5, 6, 6, 8 ], "soundness_list": [ 3, 3, 2, 2, 4 ], "presentation_list": [ 4, 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 4 ], "confidence_list": [ 4, 4, 2, 2, 3 ] }, { "paper_id": "c9kxrkabXL", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Ashmita Bhattacharya", "gender": "Female", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Malyaban Bal", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 21, "n_ref_uni": 45, "n_ref": 99, "n_ref_all": 118, "n_fig": 10, "n_tab": 4, "L_tab": 2268, "n_element_tab": 121, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1688, "n_element_tab_1": 105, "formula_len_all": 1434, "formula_len_all_1": 1095, "len_all": 141797, "len_all_1": 62544, "len_abs": 1612, "len_title": 150, "len_sents": 40875, "len_sents_1": 27658, "n_sents": 259, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1619, "title": "Multi-Agent Decision S4: Leveraging State Space Models for Offline Multi-Agent Reinforcement Learning", "abs": "Goal-conditioned sequence-based supervised learning with transformers has shown promise in offline reinforcement learning (RL) for single-agent settings. However, extending these methods to offline multi-agent RL (MARL) remains challenging. Existing transformer-based MARL approaches either train agents independently, neglecting multi-agent system dynamics, or rely on centralized transformer models, which face scalability issues. Moreover, transformers inherently struggle with long-term dependencies and computational efficiency. Building on the recent success of Structured State Space Sequence (S4) models, known for their parameter efficiency, faster inference, and superior handling of long context lengths, we propose a novel application of S4-based models to offline MARL tasks. Our method utilizes S4's efficient convolutional view for offline training and its recurrent dynamics for fast on-policy fine-tuning. To foster scalable cooperation between agents, we sequentially expand the decision-making process, allowing agents to act one after another at each time step. This design promotes bi-directional cooperation, enabling agents to share information via their S4 latent states or memory with minimal communication. Gradients also flow backward through this shared information, linking the current agent's learning to its predecessor. Experiments on challenging MARL benchmarks, including Multi-Robot Warehouse (RWARE) and StarCraft Multi-Agent Challenge (SMAC), demonstrate that our approach significantly outperforms state-of-the-art offline RL and transformer-based MARL baselines across most tasks.", "keywords": [ "offline multi-agent reinforcement learning", "sequence-based offline reinforcement learning" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 4, 1, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "c9TSRcdqBf", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "KC Tan", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Ming Yang", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "Minggang Zeng", "gender": "Male", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "", "position": "Senior Scientist II" }, { "name": "Wanyu Lin", "gender": "Female", "institution": "The Hong Kong Polytechnic University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zehui Lin", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Researcher" }, { "name": "Zhenzhong WANG", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 15, "n_ref_uni": 50, "n_ref": 103, "n_ref_all": 146, "n_fig": 13, "n_tab": 6, "L_tab": 2124, "n_element_tab": 264, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1714, "n_element_tab_1": 207, "formula_len_all": 660, "formula_len_all_1": 803, "len_all": 153131, "len_all_1": 67473, "len_abs": 1725, "len_title": 154, "len_sents": 41646, "len_sents_1": 28103, "n_sents": 327, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 106, "L_abs": 1637, "title": "Explainable Molecular Property Prediction: Aligning Chemical Concepts with Predictions via Language Models", "abs": "Providing explainable molecular property predictions is critical for many scientific domains, such as drug discovery and material science. Though transformer-based language models have shown great potential in accurate molecular property prediction, they neither provide chemically meaningful explanations nor faithfully reveal the molecular structure-property relationships. In this work, we develop a framework for explainable molecular property prediction based on language models, dubbed as Lamole, which can provide chemical concepts-aligned explanations. We take a string-based molecular representation --- Group SELFIES --- as input tokens to pretrain and fine-tune our Lamole, as it provides chemically meaningful semantics. By disentangling the information flows of Lamole, we propose combining self-attention weights and gradients for better quantification of each chemically meaningful substructure's impact on the model's output. To make the explanations more faithfully respect the structure-property relationship, we then carefully craft a marginal loss to explicitly optimize the explanations to be able to align with the chemists' annotations. We bridge the manifold hypothesis with the elaborated marginal loss to prove that the loss can align the explanations with the tangent space of the data manifold, leading to concept-aligned explanations. Experimental results over six mutagenicity datasets and one hepatotoxicity dataset demonstrate Lamole can achieve comparable classification accuracy and boost the explanation accuracy by up to 14.3%, being the state-of-the-art in explainable molecular property prediction.", "keywords": [ "molecular property prediction", "explainability" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 2, 4, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "c93my9VkqO", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Andrea Passerini", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Full Professor" }, { "name": "Antonio Longa", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Assistant Professor" }, { "name": "Beatrice Bevilacqua", "gender": "Female", "institution": "Purdue University", "country": "", "position": "PhD student" }, { "name": "Bruno Lepri", "gender": "Male", "institution": "Fondazione Bruno Kessler", "country": "IT", "position": "Principal Researcher" }, { "name": "Bruno Ribeiro", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "Veronica Lachi", "gender": "Female", "institution": "Fondazione Bruno Kessler", "country": "IT", "position": "Researcher" } ], "n_formula": 20, "n_formula_1": 16, "n_ref_uni": 55, "n_ref": 83, "n_ref_all": 115, "n_fig": 10, "n_tab": 15, "L_tab": 7511, "n_element_tab": 644, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 2060, "n_element_tab_1": 281, "formula_len_all": 1180, "formula_len_all_1": 876, "len_all": 165006, "len_all_1": 66843, "len_abs": 1695, "len_title": 128, "len_sents": 45051, "len_sents_1": 30039, "n_sents": 295, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1492, "title": "Over 100x Speedup in Relational Deep Learning via Static GNNs and Tabular Distillation", "abs": "Relational databases, organized into tables connected by primary-foreign key relationships, are widely used in industry. Companies leverage this data to build highly accurate, feature-engineered tabular models\u2014often using boosted decision trees\u2014to predict key metrics such as customer transactions and product revenues. However, these models need frequent retraining as new data is introduced, which is both expensive and time-consuming. Despite this, by being the result of extensive engineering effort, they remain difficult to outperform using generalist methods, like Temporal Graph Neural Networks (TGNNs) trained over the same relational data. Rather than attempting to replace tabular models with generalist approaches, we propose to combine the strengths of tabular models and static Graph Neural Networks (GNNs). GNNs offer better speed and scalability than TGNNs, and, as we argue, the primary strength of graph representation learning for these tasks does not lie in modeling temporal dynamics\u2014something highly- engineered tabular models excel at\u2014but in capturing complex relationships within the database, which are hard to featurize. Our approach integrates all predictive embeddings of all tabular models developed for various tasks into a single static GNN framework. Experimental results on the RelBench benchmark show that our approach achieves a performance improvement of up to 33% and an inference speedup of up to 1050x, making it highly suitable for real-time inference.", "keywords": [ "Relational Databases", "Relational Deep Learning", "GNNs", "Training Acceleration", "Real-time Inference", "Distillation" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "c8sEgxG2c0", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" }, { "name": "Jieke Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Lizhen Shi", "gender": "Female", "institution": "Northwestern University", "country": "", "position": "Assistant Professor" }, { "name": "Weimin Wu", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Zhihan Zhou", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Zhong Wang", "gender": "Not Specified", "institution": "University of California, Merced", "country": "", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 71, "n_ref_all": 94, "n_fig": 10, "n_tab": 8, "L_tab": 1605, "n_element_tab": 87, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 3702, "n_element_tab_1": 145, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 148484, "len_all_1": 73904, "len_abs": 1357, "len_title": 109, "len_sents": 45325, "len_sents_1": 34672, "n_sents": 322, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1366, "title": "GenomeOcean: Efficient Foundation Model for Genome Generation", "abs": "We introduce GenomeOcean, a 4-billion-parameter genome foundation model that natively generates DNA sequences that are adherent to the input context. \nWith an efficiency-oriented model design, GenomeOcean is 80 times faster than existing models of similar size in genome generation. \nUnlike most existing genome foundation models\u2014such as DNABERT and Nucleotide Transformers\u2014that are designed for discriminative tasks, GenomeOcean leverages generative modeling to unlock new potentials in genomics research. \nDiverging from the traditional reliance on reference genomes\u2014which possess inherent biases\u2014GenomeOcean is exclusively trained on large-scale curated environmental samples collected from diverse ecosystems, including oceans, lakes, forests, and soils. This extensive genomic diversity, encompassing uncultured and uncharacterized organisms, allows GenomeOcean to generate sequences that better reflect the true diversity of life.\nIn a series of automated evaluations, we demonstrate GenomeOcean's capability to understand and follow context sequences. \nCompared to existing models, GenomeOcean not only better retains species information but also produces sequences with more appropriate open reading frame lengths and codon usage bias.\nWe anticipate the open release of GenomeOcean to open up new possibilities in genomics and computational biology research.", "keywords": [ "Genome Foundation Model", "Genome Generation" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "c8QlNuhy2G", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Bin Xu", "gender": "Male", "institution": "Department of Computer Science, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jie Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jinhao Chen", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Weihan Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Wenmeng Yu", "gender": "Male", "institution": "Zhipu AI", "country": "CN", "position": "Researcher" }, { "name": "Wenyi Hong", "gender": "Female", "institution": "Department of Computer Science and Technology, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhen Yang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhengxiao Du", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhihuan Jiang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 39, "n_ref": 73, "n_ref_all": 100, "n_fig": 12, "n_tab": 13, "L_tab": 1384, "n_element_tab": 171, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 2312, "n_element_tab_1": 378, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 214984, "len_all_1": 58898, "len_abs": 1407, "len_title": 132, "len_sents": 58813, "len_sents_1": 24844, "n_sents": 480, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1318, "title": "MathGLM-Vision: Solving Mathematical Problems with Multi-Modal Large Language Model", "abs": "Large language models (LLMs) have demonstrated significant capabilities in mathematical reasoning, particularly with text-based mathematical problems. However, current multi-modal large language models (MLLMs), especially those specialized in mathematics, tend to focus predominantly on solving geometric problems but ignore the diversity of visual information available in other areas of mathematics. Moreover, the geometric information for these specialized mathematical MLLMs is derived from several public datasets, which are typically limited in diversity and complexity. To address these limitations, we aim to construct a fine-tuning dataset named MathVL, and develop a series of specialized mathematical MLLMs termed MathGLM-Vision by conducting Supervised Fine-Tuning (SFT) on MathVL with various parameter-scale backbones. To extensively evaluate the effectiveness of MathGLM-Vision, we conduct experiments on several public benchmarks and our curated MathVL-test benchmark consisting of 2,000 problems. Experimental results demonstrate that MathGLM-Vision achieves significant improvements compared with some existing models, including backbone models and open-source mathematical MLLMs. These findings indicate the importance of diversity dataset in enhancing the mathematical reasoning abilities of MLLMs.", "keywords": [ "Mathematical Reasoning", "Multi-modal large language model" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "c87QZPTVVm", "primary_area": "causal reasoning", "n_author": 1, "author_info_list": [ { "name": "Kamesh R", "gender": "Male", "institution": "Sathyabama University", "country": "IN", "position": "Undergrad student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 9, "n_ref": 22, "n_ref_all": 23, "n_fig": 0, "n_tab": 1, "L_tab": 397, "n_element_tab": 91, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 397, "n_element_tab_1": 91, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 75115, "len_all_1": 75115, "len_abs": 1437, "len_title": 114, "len_sents": 26381, "len_sents_1": 26381, "n_sents": 174, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1442, "title": "Think Beyond Size: Dynamic Prompting for More Effective Reasoning", "abs": "Pretrained large language models (LLMs) are increasingly utilized across a wide range of natural language processing (NLP) tasks due to their impressive capabilities as few-shot learners. Recent techniques, such as chain-of-thought (CoT) prompting, have significantly advanced multi-step reasoning by introducing step-by-step decomposition, achieving state-of-the-art results on complex reasoning benchmarks. However, these approaches often rely on static prompting templates that do not adapt to task complexity or errors during the reasoning process. In this work, we introduce Adaptive Prompting, a dynamic and iterative framework designed to enhance reasoning by incorporating real-time adjustments to prompt structures and validation mechanisms. Experimental results demonstrate that Adaptive Prompting significantly improves performance on diverse reasoning benchmarks, including arithmetic reasoning (GSM8K, MultiArith), logical reasoning and commonsense tasks, achieving substantial accuracy gains compared to static prompting baselines. By integrating guided prompts, intermediate validation, and self-corrective steps, our approach enables smaller models to achieve competitive performance with larger counterparts, such as GPT-4, while maintaining computational efficiency. The framework achieves this without requiring fine-tuning or task-specific training data, highlighting the untapped potential of iterative reasoning methods.", "keywords": [ "model efficiency", "prompt optimization", "task-specific prompting", "small-scale LLMs", "real-time adjustment", "efficient problem-solving", "adaptive reasoning", "incremental prompting", "step-wise refinement", "model performance improvement", "task adaptability", "few-shot prompting", "scalable models." ], "rating_list": [ 5, 3, 1, 3 ], "soundness_list": [ 2, 2, 1, 1 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 4, 3, 5, 2 ] }, { "paper_id": "c6zI3Cp8c6", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Depen Morwani", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Itai Shapira", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "", "position": "PhD student" }, { "name": "Lucas Janson", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "Associate Professor" }, { "name": "Nikhil Vyas", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Sham M. Kakade", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "eran malach", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" } ], "n_formula": 42, "n_formula_1": 33, "n_ref_uni": 34, "n_ref": 80, "n_ref_all": 101, "n_fig": 4, "n_tab": 3, "L_tab": 266, "n_element_tab": 38, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2774, "formula_len_all_1": 1668, "len_all": 151243, "len_all_1": 69317, "len_abs": 1349, "len_title": 94, "len_sents": 42865, "len_sents_1": 28577, "n_sents": 391, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 45, "L_abs": 1359, "title": "A New Perspective on Shampoo's Preconditioner", "abs": "Shampoo, a second-order optimization algorithm that uses a Kronecker product preconditioner, has recently received increasing attention from the machine learning community. Despite the increasing popularity of Shampoo, the theoretical foundations of its effectiveness are not well understood. The preconditioner used by Shampoo can be viewed as either an approximation of the Gauss--Newton component of the Hessian or the covariance matrix of the gradients maintained by Adagrad. Our key contribution is providing an explicit and novel connection between the optimal Kronecker product approximation of these matrices and the approximation\nmade by Shampoo. Our connection highlights a subtle but common misconception about Shampoo\u2019s approximation. In particular, the square of the approximation used by the Shampoo optimizer is equivalent to a single step of the power\niteration algorithm for computing the aforementioned optimal Kronecker product approximation. Across a variety of datasets and architectures we empirically\ndemonstrate that this is close to the optimal Kronecker product approximation. We also study the impact of batch gradients and empirical Fisher on the quality of Hessian approximation. Our findings not only advance the theoretical understanding of Shampoo but also illuminate potential pathways for enhancing its practical performance.", "keywords": [ "optimization", "hessian", "shampoo", "adagrad", "second order optimization" ], "rating_list": [ 3, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 1, 4, 3, 3 ] }, { "paper_id": "c6hGb8IsRN", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Chaining Wang", "gender": "unknown", "institution": "Santa Clara University", "country": "US", "position": "PhD student" }, { "name": "Haris Vikalo", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Full Professor" }, { "name": "Usman Akram", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Yiyue Chen", "gender": "unknown", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" } ], "n_formula": 47, "n_formula_1": 19, "n_ref_uni": 15, "n_ref": 30, "n_ref_all": 45, "n_fig": 3, "n_tab": 6, "L_tab": 2487, "n_element_tab": 118, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2429, "n_element_tab_1": 83, "formula_len_all": 4109, "formula_len_all_1": 1364, "len_all": 121449, "len_all_1": 63784, "len_abs": 1238, "len_title": 128, "len_sents": 38863, "len_sents_1": 27444, "n_sents": 287, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1236, "title": "Fed-REACT: Federated Representation Learning for Heterogeneous Time Series Data", "abs": "Motivated by high resource costs and privacy concerns that characterize centralized machine learning, federated learning (FL) emerged as an efficient alternative that allows the participating clients to collaboratively train global model while keeping their data local.\nIn practice, distributions of clients' data vary over time and from one client to another, creating heterogeneous conditions that deteriorate performance of conventional FL algorithms. In this work, we study an FL framework where clients train on heterogeneous time series data and introduce to these settings Fed-REACT, a novel federated learning method leveraging representation learning and evolutionary clustering. The algorithm consists of two stages: (1) in the first stage, the clients learn a model that extracts meaningful features from local time series data; (2) in the second stage, the server adaptively groups clients into clusters and coordinated cluster-wise learning of task (i.e., post-representation) models for local downstream tasks, e.g., classification or regression. We demonstrated high accuracy and robustness of the proposed algorithm in experiments on real-world time series datasets, and provided theoretical analysis of its performance.", "keywords": [ "Federated learning", "representation learning", "time series data" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 3, 2, 3, 4 ] }, { "paper_id": "c6TDOPEQ0e", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Disen Lan", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Weigao Sun", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xiaoye Qu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yu Cheng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" } ], "n_formula": 30, "n_formula_1": 22, "n_ref_uni": 31, "n_ref": 69, "n_ref_all": 79, "n_fig": 4, "n_tab": 6, "L_tab": 2257, "n_element_tab": 379, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1233, "n_element_tab_1": 141, "formula_len_all": 3436, "formula_len_all_1": 1586, "len_all": 148862, "len_all_1": 70122, "len_abs": 2125, "len_title": 123, "len_sents": 40774, "len_sents_1": 32489, "n_sents": 339, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1604, "title": "LASP-2: Rethinking Sequence Parallelism for Linear Attention and its Hybrid", "abs": "Linear sequence modeling approaches, such as linear attention, provide advantages like linear-time training and constant-memory inference over sequence lengths. However, existing sequence parallelism (SP) methods are either not optimized for their right-product-first feature or use ring-style communication as in LASP, which results in lower computation parallelism, limits their scalability for longer sequences in distributed systems. In this paper, we introduce LASP-2, a new SP approach designed to enhance both communication and computation efficiency in linear (attention) transformer models with very-long input sequences. Compared to LASP, LASP-2 rethinks the minimal communication requirement for SP on linear attention, reorganizes the whole communication-computation order of LASP. In this way, only one single all-gather collective communication is needed on intermediate memory states, whose sizes are independent of the sequence length, leading to significant improvements of both communication and computation parallelism, as well as their overlap. Additionally, we extend LASP-2 to LASP-2H by applying similar communication redesign to standard attention modules, offering an efficient SP solution for hybrid models that combine linear and standard attention layers. Our evaluation on a Linear-Llama3 model, a variant of Llama3 with linear attention replacing standard attention, demonstrates the effectiveness of LASP-2 and LASP-2H. Specifically, LASP-2 achieves throughput improvements of 15.2\\% over LASP and 36.6\\% over Ring Attention, with a sequence length of 2048K across 64 GPUs.", "keywords": [ "Sequence parallelism", "Distributed Training" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "c61unr33XA", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Baharan Mirzasoleiman", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "Jiayi Ni", "gender": "Female", "institution": "University of California, Los Angeles", "country": "", "position": "Undergrad student" }, { "name": "Siddharth Joshi", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 34, "n_formula_1": 16, "n_ref_uni": 39, "n_ref": 111, "n_ref_all": 136, "n_fig": 4, "n_tab": 11, "L_tab": 10172, "n_element_tab": 626, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 6225, "n_element_tab_1": 379, "formula_len_all": 3730, "formula_len_all_1": 962, "len_all": 163635, "len_all_1": 73878, "len_abs": 1295, "len_title": 161, "len_sents": 45225, "len_sents_1": 29565, "n_sents": 356, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 112, "L_abs": 1350, "title": "Dataset Distillation via Knowledge Distillation: Towards Efficient Self-Supervised Pre-training of Deep Networks", "abs": "Dataset distillation (DD) generates small synthetic datasets that can efficiently train deep networks with a limited amount of memory and compute. Despite the success of DD methods for supervised learning, DD for self-supervised pre-training of deep models has remained unaddressed. Pre-training on unlabeled data is crucial for efficiently generalizing to downstream tasks with limited labeled data. In this work, we propose the first effective DD method for SSL pre-training. First, we show, theoretically and empirically, that naiive application of supervised DD methods to SSL fails, due to the high variance of the SSL gradient. Then, we address this issue by relying on insights from knowledge distillation (KD) literature. Specifically, we train a small student model to match the representations of a larger teacher model trained with SSL. Then, we generate a small synthetic dataset by matching the training trajectories of the student models. As the KD objective has considerably lower variance than SSL, our approach can generate synthetic datasets that can successfully pre-train high-quality encoders. Through extensive experiments, we show that our distilled sets lead to up to 13% higher accuracy than prior work, on a variety of downstream tasks, in the presence of limited labeled data. Code at https://github.com/BigML-CS-UCLA/MKDT.", "keywords": [ "dataset distillation", "self-supervised learning" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "c5boBrSTKj", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Huazhen Lin", "gender": "Female", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "Full Professor" }, { "name": "Jiujun He", "gender": "Male", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "PhD student" }, { "name": "Zhiyao Tan", "gender": "Male", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "PhD student" } ], "n_formula": 32, "n_formula_1": 14, "n_ref_uni": 59, "n_ref": 138, "n_ref_all": 165, "n_fig": 9, "n_tab": 7, "L_tab": 5757, "n_element_tab": 640, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 3618, "n_element_tab_1": 390, "formula_len_all": 3236, "formula_len_all_1": 1186, "len_all": 201657, "len_all_1": 70430, "len_abs": 1734, "len_title": 131, "len_sents": 45164, "len_sents_1": 29972, "n_sents": 365, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1757, "title": "AMSC: Adaptive Multi-Dimensional Structured Compression with Theoretical Guarantees", "abs": "Network pruning is a pivotal strategy for reducing complexity and accelerating inference. Most pruning methods focus on a single dimension (depth or width), leading to insufficient compression when multiple dimensions are redundant. Additionally, separating pruning from training disrupts established network correlations, causing performance degradation. In this paper, we propose a novel Adaptive Multi-dimensional Structured Compression (AMSC) method that simultaneously learns the minimal depth, the minimal width, and network parameters under the strategy that prioritizes depth compression. Specifically, based on the regularization technique, AMSC incorporates layer- and filter- specific information into the penalty in order to adaptively identify and eliminate redundant depth and width in terms of the importance and size of each layer and filter. It integrates compression and training processes together without pruning. Consequently, the proposed method enables adaptive structure reduction from the initial configuration to a structure necessary that minimizes the generalization error. Rigorous theoretical evidence is provided in terms of the consistency of AMSC in achieving minimal network depth and width. To the best of our knowledge, this is the first study that offers a theoretical guarantees in structure selection. Extensive experiments on CIFAR-10/100 and ImageNet datasets demonstrate our method not only achieves state-of-the-art compression performance in terms of FLOPs and total parameters, but also preserves competitive classification accuracy. For example, AMSC enhances the accuracy of ResNet56 on CIFAR-10 from 93.37\\% to 93.71\\%, while simultaneously reducing calculations by 58.63\\% and parameters by 44.71\\%.", "keywords": [ "Multi-dimensional structured compression", "adaptive group lasso", "selection consistency" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 5 ] }, { "paper_id": "c5JZEPyFUE", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Baixu Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Haoran Xu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Jianmin Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Xingzhuo Guo", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Mingsheng Long", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 39, "n_formula_1": 22, "n_ref_uni": 59, "n_ref": 120, "n_ref_all": 145, "n_fig": 8, "n_tab": 8, "L_tab": 1782, "n_element_tab": 259, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 213, "n_element_tab_1": 46, "formula_len_all": 4336, "formula_len_all_1": 1572, "len_all": 171121, "len_all_1": 58776, "len_abs": 1303, "len_title": 117, "len_sents": 37926, "len_sents_1": 23801, "n_sents": 340, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1307, "title": "Dynamical Diffusion: Learning Temporal Dynamics with Diffusion Models", "abs": "Diffusion models have emerged as powerful generative frameworks by progressively adding noise to data through a forward process and then reversing this process to generate realistic samples. While these models have achieved strong performance across various tasks and modalities, their application to temporal predictive learning remains underexplored. Existing approaches treat predictive learning as a conditional generation problem, but often fail to fully exploit the temporal dynamics inherent in the data, leading to challenges in generating temporally coherent sequences. To address this, we introduce Dynamical Diffusion (DyDiff), a theoretically sound framework that incorporates temporally aware forward and reverse processes. Dynamical Diffusion explicitly models temporal transitions at each diffusion step, establishing dependencies on preceding states to better capture temporal dynamics. Through the reparameterization trick, Dynamical Diffusion achieves efficient training and inference similar to any standard diffusion model. Extensive experiments across scientific spatiotemporal forecasting, video prediction, and time series forecasting demonstrate that Dynamical Diffusion consistently improves performance in temporal predictive tasks, filling a crucial gap in existing methodologies.", "keywords": [ "Diffusion Model", "Generative Model", "Prediction Learning", "Dynamics" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "c54apoozCS", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" }, { "name": "Jerry Yao-Chieh Hu", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Minshuo Chen", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Postdoc" }, { "name": "Weimin Wu", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Yi-Chen Lee", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Yu-Chao Huang", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "MS student" } ], "n_formula": 556, "n_formula_1": 29, "n_ref_uni": 27, "n_ref": 190, "n_ref_all": 215, "n_fig": 6, "n_tab": 13, "L_tab": 1951, "n_element_tab": 229, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 95, "n_element_tab_1": 18, "formula_len_all": 58641, "formula_len_all_1": 3632, "len_all": 635759, "len_all_1": 77173, "len_abs": 1917, "len_title": 136, "len_sents": 184061, "len_sents_1": 29874, "n_sents": 2354, "n_sents_1": 299, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1210, "title": "On Statistical Rates of Conditional Diffusion Transformer: Approximation and Estimation", "abs": "We investigate the approximation and estimation rates of conditional diffusion transformers (DiTs) with classifier-free guidance.\nWe present a comprehensive analysis for ``in-context'' conditional DiTs under four common data assumptions.\nWe show that both conditional DiTs and their latent variants lead to the minimax optimality of unconditional DiTs under identified settings.\nSpecifically, we discretize the input domains into infinitesimal grids and then perform a term-by-term Taylor expansion on the conditional diffusion score function under H\u00f6lder smooth data assumption.\nThis enables fine-grained use of transformers' universal approximation through a more detailed piecewise constant approximation, and hence obtains tighter bounds.\nAdditionally, we extend our analysis to the latent setting under the linear latent subspace assumption.\nWe not only show that latent conditional DiTs achieve lower bounds than conditional DiTs both in approximation and estimation, but also show the minimax optimality of latent unconditional DiTs.\nOur findings establish statistical limits for conditional and unconditional DiTs, and offer\npractical guidance toward developing more efficient and accurate DiT models.", "keywords": [ "Conditional Diffusion Transformer", "Statistical Rates", "Approximation", "Estimation" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "c4wEKJOjY3", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Tianqi Zhong", "gender": "Male", "institution": "Shining 3D Tech Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yuanchao Wang", "gender": "Male", "institution": "Duke University", "country": "US", "position": "MS student" }, { "name": "Zhao-Rong Lai", "gender": "Male", "institution": "Jinan University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 45, "n_formula_1": 30, "n_ref_uni": 24, "n_ref": 50, "n_ref_all": 100, "n_fig": 2, "n_tab": 4, "L_tab": 2044, "n_element_tab": 127, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 2445, "n_element_tab_1": 155, "formula_len_all": 2979, "formula_len_all_1": 2191, "len_all": 120900, "len_all_1": 69234, "len_abs": 924, "len_title": 136, "len_sents": 37967, "len_sents_1": 28253, "n_sents": 350, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 929, "title": "Out-of-distribution Generalization for Total Variation based Invariant Risk Minimization", "abs": "Invariant risk minimization is an important general machine learning framework that has recently been interpreted as a total variation model (IRM-TV). However, how to improve out-of-distribution (OOD) generalization in the IRM-TV setting remains unsolved. In this paper, we extend IRM-TV to a Lagrangian multiplier model, named OOD-TV-IRM. We find that the autonomous TV penalty hyperparameter is exactly the Lagrangian multiplier. Thus OOD-TV-IRM is essentially a primal-dual optimization model, where the primal optimization reduces the entire invariant risk and the dual optimization strengthens the TV penalty. The objective is to reach a semi-Nash-equilibrium where the balance between the training loss and OOD generalization is kept. We also develop a convergent primal-dual solving algorithm that facilitates an adversarial learning scheme. Experimental results show that OOD-TV-IRM outperforms IRM-TV in most situations.", "keywords": [ "Out-of-distribution generalization", "total variation", "invariant risk minimization" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "c4w7WVs1z7", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Aleksandr Panov", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "Aleksei Staroverov", "gender": "Male", "institution": "AIRI", "country": "RU", "position": "Researcher" }, { "name": "Alexey Kovalev", "gender": "Male", "institution": "AIRI", "country": "RU", "position": "Researcher" }, { "name": "Egor Cherepanov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "PhD student" }, { "name": "Dmitry A Yudin", "gender": "Male", "institution": "Artificial Intelligence Research Intsitute", "country": "RU", "position": "Principal Researcher" } ], "n_formula": 8, "n_formula_1": 3, "n_ref_uni": 56, "n_ref": 157, "n_ref_all": 211, "n_fig": 19, "n_tab": 6, "L_tab": 2152, "n_element_tab": 212, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1270, "n_element_tab_1": 126, "formula_len_all": 1474, "formula_len_all_1": 919, "len_all": 198338, "len_all_1": 67222, "len_abs": 1393, "len_title": 88, "len_sents": 62650, "len_sents_1": 29450, "n_sents": 440, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1424, "title": "Recurrent Action Transformer with Memory", "abs": "Recently, the use of transformers in offline reinforcement learning has become a rapidly developing area. This is due to their ability to treat the agent's trajectory in the environment as a sequence, thereby reducing the policy learning problem to sequence modeling. In environments where the agent's decisions depend on past events (POMDPs), it is essential to capture both the event itself and the decision point in the context of the model. However, the quadratic complexity of the attention mechanism limits the potential for context expansion. One solution to this problem is to extend transformers with memory mechanisms. This paper proposes a Recurrent Action Transformer with Memory (RATE), a novel model architecture that incorporates a recurrent memory mechanism designed to regulate information retention. To evaluate our model, we conducted extensive experiments on memory-intensive environments (ViZDoom-Two-Colors, T-Maze, Memory Maze, Minigrid-Memory), classic Atari games, and MuJoCo control environments. The results show that using memory can significantly improve performance in memory-intensive environments, while maintaining or improving results in classic environments. We believe that our results will stimulate research on memory mechanisms for transformers applicable to offline reinforcement learning. The code is open-sourced and can be found in the https://anonymous.4open.science/r/RATE-B01F/.", "keywords": [ "offline RL", "transformers", "memory", "pomdp" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "c4w1TqcSi0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chen Qian", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Cheng Yang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Associate Professor" }, { "name": "Jiarui Yuan", "gender": "Male", "institution": ", Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Maosong Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Weize Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhiyuan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 18, "n_formula_1": 8, "n_ref_uni": 31, "n_ref": 90, "n_ref_all": 112, "n_fig": 8, "n_tab": 12, "L_tab": 2804, "n_element_tab": 189, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 968, "n_element_tab_1": 62, "formula_len_all": 2163, "formula_len_all_1": 959, "len_all": 229727, "len_all_1": 69442, "len_abs": 1844, "len_title": 129, "len_sents": 54277, "len_sents_1": 30598, "n_sents": 437, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1851, "title": "Optima: Optimizing Effectiveness and Efficiency for LLM-Based Multi-Agent System", "abs": "Large Language Model (LLM) based multi-agent systems (MAS) show remarkable potential in collaborative problem-solving, yet they still face critical challenges: low communication efficiency, poor scalability, and a lack of effective parameter-updating optimization methods for multi-agent collaboration. We present Optima, a novel framework that addresses these issues by significantly enhancing both communication efficiency and task effectiveness in LLM-based MAS through LLM training. At its core, Optima employs an iterative generate, rank, select, and train paradigm, incorporating a reward function that balances task performance, token efficiency, and communication readability. We explore various RL algorithms, including Supervised Fine-Tuning, Direct Preference Optimization, and their hybrid approaches, providing insights into their effectiveness-efficiency trade-offs for iterative LLM-based MAS training. Additionally, we integrate Monte Carlo Tree Search-inspired techniques for DPO data generation, conceptualizing conversation turns as tree nodes to explore diverse interaction trajectories. We evaluate Optima on common multi-agent tasks, including information-asymmetric question answering and complex reasoning. Our method demonstrates consistent and substantial improvements over single-agent baselines and vanilla MAS based on Llama 3 8B, achieving up to 2.8x performance gain with less than 10\\% tokens on tasks requiring heavy multi-agent information exchange. Moreover, Optima's efficiency gains open new possibilities for leveraging inference-compute more effectively, potentially leading to improved inference-time scaling laws. By addressing fundamental challenges in multi-agent collaboration and providing a novel optimization framework, Optima shows the potential towards scalable, efficient, and effective LLM-based MAS.", "keywords": [ "llm agent", "multi-agent", "inference scaling law" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 1, 2, 4 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "c4OGMNyzPT", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Bohan Zhuang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Research Professor" }, { "name": "Qi Wu", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "Associate Professor" }, { "name": "Xinyu Wang", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "Lecturer" } ], "n_formula": 30, "n_formula_1": 6, "n_ref_uni": 46, "n_ref": 78, "n_ref_all": 109, "n_fig": 16, "n_tab": 8, "L_tab": 2529, "n_element_tab": 376, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 907, "n_element_tab_1": 104, "formula_len_all": 2606, "formula_len_all_1": 247, "len_all": 289661, "len_all_1": 63371, "len_abs": 1215, "len_title": 99, "len_sents": 100460, "len_sents_1": 31415, "n_sents": 809, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1172, "title": "Are Large Vision Language Models Good Game Players?", "abs": "Large Vision Language Models (LVLMs) have demonstrated remarkable abilities in understanding and reasoning about both visual and textual information. However, existing evaluation methods for LVLMs, primarily based on benchmarks like Visual Question Answering and image captioning, often fail to capture the full scope of LVLMs' capabilities. These benchmarks are limited by issues such as inadequate assessment of detailed visual perception, data contamination, and a lack of focus on multi-turn reasoning. To address these challenges, we propose LVLM-Playground, a game-based evaluation framework designed to provide a comprehensive assessment of LVLMs' cognitive and reasoning skills in structured environments. LVLM-Playground uses a set of games to evaluate LVLMs on four core tasks: Perceiving, Question Answering, Rule Following, and End-to-End Playing, with each target task designed to assess specific abilities, including visual perception, reasoning, decision-making, etc. Based on this framework, we conduct extensive experiments that explore the limitations of current LVLMs, such as handling long structured outputs and perceiving detailed and dense elements.", "keywords": [ "Large Vision Language Models", "LLMs", "Evaluation", "Game" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "c4Nh4A8Xn5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Hanwang Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Jiequan Cui", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Research Fellow" }, { "name": "Qingshan Xu", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Research Fellow" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Wanjuan Su", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Wenbing Tao", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "xinyaoliao", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "YI Xuanyu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Yew Soon Ong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 28, "n_ref": 56, "n_ref_all": 94, "n_fig": 13, "n_tab": 6, "L_tab": 1995, "n_element_tab": 220, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3815, "n_element_tab_1": 174, "formula_len_all": 1183, "formula_len_all_1": 1182, "len_all": 133796, "len_all_1": 69575, "len_abs": 1346, "len_title": 143, "len_sents": 37420, "len_sents_1": 29153, "n_sents": 303, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1352, "title": "Geo-3DGS: Multi-view Geometry Consistency for 3D Gaussian Splatting and Surface Reconstruction", "abs": "Recently, the emergence of 3D Gaussian Splatting (3DGS) has made real-time and high-quality rendering possible. However, it is still challenging for 3DGS to reconstruct accurate geometry surfaces and achieve higher-quality rendering. To address these challenges, we propose to leverage multi-view geometry consistency for 3DGS and surface reconstruction. We reveal that there exists multi-view geometry inconsistency in 3DGS, preventing 3DGS from achieving higher-quality rendering and accurate surface reconstruction. To mitigate the geometry inconsistency, we first develop a multi-view photometric consistency regularization to constrain the rendered depth of 3DGS, which helps establish more stable and consistent 3D Gaussians to facilitate both rendering and surface reconstruction. To reconstruct geometry surfaces from 3DGS, we introduce a neural Signed Distance Function (SDF) field to represent continuous geometries of 3DGS. Then, we propose a geometry consistency-based SDF learning strategy, which leverages multi-view geometry consistency cues from 3DGS to efficiently optimize the SDF field for surface reconstruction. Extensive experiments on various datasets demonstrate that our method achieves both high-quality rendering and accurate surface reconstruction while keeping a good efficiency. Our code will be released upon publication.", "keywords": [ "Gaussian Splatting", "Multi-view geometry consistency", "Surface reconstruction", "Novel view synthesis" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "c4498OydLP", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Hong Mei", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Xinyu Ma", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Xu Chu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Yang Lin", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yasha Wang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Yujie Jin", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Zhibang Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 41, "n_formula_1": 22, "n_ref_uni": 49, "n_ref": 85, "n_ref_all": 107, "n_fig": 4, "n_tab": 13, "L_tab": 10612, "n_element_tab": 733, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2195, "n_element_tab_1": 293, "formula_len_all": 4616, "formula_len_all_1": 1794, "len_all": 227597, "len_all_1": 65452, "len_abs": 1408, "len_title": 112, "len_sents": 73213, "len_sents_1": 27648, "n_sents": 550, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1416, "title": "LoRA Dropout as a Sparsity Regularizer for Overfitting Reduction", "abs": "Parameter-efficient fine-tuning methods, represented by LoRA, play an essential role in adapting large-scale pre-trained models to downstream tasks. \nHowever, fine-tuning LoRA-series models also faces the risk of overfitting on small training datasets, and there's still a lack of theoretical guidance and practical mechanisms to control overfitting on LoRA-based PEFT methods. This paper introduces a novel dropout-based sparsity regularizer for LoRA, dubbed LoRA Dropout, which mitigates overfitting by applying refined dropout to LoRA's low-rank matrices.\nWe establish a theoretical framework that models dropout in LoRA as a sparse fine-tuning process and derive a generalization error bound under this sparsity regularization.\nTheoretical results show that appropriate sparsity can tighten the gap between empirical and generalization risks and thereby control overfitting. We further enhance the sparsity patterns in conventional dropout methods and propose an innovative LoRA Dropout method for more precise sparsity regularization to achieve better overfitting reduction. \nFurthermore, we introduce a test-time ensemble strategy and provide theoretical evidence demonstrating that the ensemble method can further compress the error bound and lead to better performance. \nExtensive experiments on various NLP tasks validate the effectiveness of our LoRA Dropout framework in improving the model's performance.", "keywords": [ "Large Language Model", "LoRA", "Dropout" ], "rating_list": [ 3, 6, 5, 3, 6 ], "soundness_list": [ 2, 3, 3, 1, 3 ], "presentation_list": [ 2, 3, 2, 1, 3 ], "contribution_list": [ 1, 2, 3, 2, 3 ], "confidence_list": [ 5, 3, 4, 4, 4 ] }, { "paper_id": "c3rfGbXMBE", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 7, "author_info_list": [ { "name": "He Zhao", "gender": "Male", "institution": "Commonwealth Scientific and Industrial Research Organisation, CSIRO", "country": "AU", "position": "Researcher" }, { "name": "Jueqing Lu", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Lan Du", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Associate Professor" }, { "name": "Weijie Xu", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Xiaohao Yang", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "YUANYUAN QI", "gender": "Female", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Dinh Phung", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 58, "n_ref": 122, "n_ref_all": 164, "n_fig": 15, "n_tab": 6, "L_tab": 2956, "n_element_tab": 203, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 323, "n_element_tab_1": 8, "formula_len_all": 632, "formula_len_all_1": 624, "len_all": 188137, "len_all_1": 59203, "len_abs": 1007, "len_title": 108, "len_sents": 50898, "len_sents_1": 26668, "n_sents": 393, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1010, "title": "Neural Topic Modeling with Large Language Models in the Loop", "abs": "Topic modeling is a fundamental task in natural language processing, allowing the discovery of latent thematic structures in text corpora. While Large Language Models (LLMs) have demonstrated promising capabilities in topic discovery, their direct application to topic modeling suffers from issues such as incomplete topic coverage, misalignment of topics, and inefficiency. To address these limitations, we propose LLM-ITL, a novel LLM-in-the-loop framework that integrates LLMs with many existing Neural Topic Models (NTMs). In LLM-ITL, global topics and document representations are learned through the NTM, while an LLM refines the topics via a confidence-weighted Optimal Transport (OT)-based alignment objective. This process enhances the interpretability and coherence of the learned topics, while maintaining the efficiency of NTMs. Extensive experiments demonstrate that LLM-ITL can help NTMs significantly improve their topic interpretability while maintaining the quality of document representation.", "keywords": [ "Neural Topic Model", "Topic Model", "Large Language Model", "Optimal Transport" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "c3i8uRSE9h", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Alexandros Graikos", "gender": "unknown", "institution": "Stony Brook University", "country": "US", "position": "PhD student" }, { "name": "Dimitris Samaras", "gender": "Male", "institution": "Stony Brook University", "country": "US", "position": "Full Professor" }, { "name": "Nebojsa Jojic", "gender": "unknown", "institution": "Microsoft Research", "country": "", "position": "Principal Researcher" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 11, "n_ref": 25, "n_ref_all": 41, "n_fig": 9, "n_tab": 1, "L_tab": 2, "n_element_tab": 1, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 2, "n_element_tab_1": 1, "formula_len_all": 859, "formula_len_all_1": 859, "len_all": 59269, "len_all_1": 59269, "len_abs": 1669, "len_title": 106, "len_sents": 21713, "len_sents_1": 21713, "n_sents": 142, "n_sents_1": 142, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1682, "title": "Fast constrained sampling in pre-trained diffusion models", "abs": "Diffusion models have dominated the field of large, generative image models, with the prime examples of Stable Diffusion and DALL-E 3 being widely adopted. These models have been trained to perform text-conditioned generation on vast numbers of image-caption pairs and as a byproduct, have acquired general knowledge about natural image statistics. However, when confronted with the task of constrained sampling, e.g. generating the right half of an image conditioned on the known left half, applying these models is a delicate and slow process, with previously proposed algorithms relying on expensive iterative operations that are usually orders of magnitude slower than text-based inference. This is counter-intuitive, as image-conditioned generation should rely less on the difficult-to-learn semantic knowledge that links captions and imagery, and should instead be achievable by lower-level correlations among image pixels. In practice, inverse models are trained or tuned separately for each inverse problem, e.g. by providing parts of images during training as an additional condition, to allow their application in realistic settings. However, we argue that this is not necessary and propose an algorithm for fast-constrained sampling in large pre-trained diffusion models (Stable Diffusion) that requires no expensive backpropagation operations through the model and produces results comparable even to the state-of-the-art \\emph{tuned} models. Our method is based on a novel optimization perspective to sampling under constraints and employs a numerical approximation to the expensive gradients, previously computed using backpropagation, incurring significant speed-ups.", "keywords": [ "Diffusion models", "conditional sampling", "generative models" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 2, 2, 4, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "c2OtbtZXFC", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Jiang Hu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Jiaxi Cui", "gender": "Male", "institution": "Liaoning University", "country": "", "position": "Undergrad student" }, { "name": "Lin Lin", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Quanzheng Li", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Associate Professor" }, { "name": "Zaiwen Wen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "yuan zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" } ], "n_formula": 33, "n_formula_1": 29, "n_ref_uni": 46, "n_ref": 79, "n_ref_all": 95, "n_fig": 6, "n_tab": 3, "L_tab": 1686, "n_element_tab": 277, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 849, "n_element_tab_1": 127, "formula_len_all": 3965, "formula_len_all_1": 2058, "len_all": 129275, "len_all_1": 65935, "len_abs": 1293, "len_title": 143, "len_sents": 32578, "len_sents_1": 27451, "n_sents": 273, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1268, "title": "Retraction-free optimization over the Stiefel manifold with application to the LoRA fine-tuning", "abs": "Optimization over the Stiefel manifold has played a significant role in various machine learning tasks. Many existing algorithms either use the retraction operator to keep each iterate staying on the manifold, or solve an unconstrained quadratic penalized problem. The retraction operator in the former corresponds to orthonormalization of matrices and can be computationally costly for large-scale matrices. The latter approach usually equips with an unknown large penalty parameter. To address the above issues, we propose a retraction-free and penalty parameter-free algorithm, which lands on the manifold. Moreover, our convergence theory allows using constant step size, which improve the result of converging to a neighborhood in \\citep{ablin2022fast}.\n A key component of the analysis is the convex-like property of the quadratic penalty of the Stiefel manifold, which enables us to explicitly characterize the constant penalty parameter. As an application, we introduce a new algorithm, Manifold-LoRA, which employs the landing technique and a carefully designed step size strategy to accelerate low-rank adaptation (LoRA) in fine-tuning large language models. Numerical experiments on the benchmark datasets demonstrate the efficiency of our proposed method.", "keywords": [ "landing", "manifold", "fine-tuning", "LoRA" ], "rating_list": [ 3, 3, 5, 8 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "c2NSfbKGOc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Heyan Huang", "gender": "Female", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Jincheng Xie", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Runheng Liu", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Xingchen Xiao", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Zhongyi Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 51, "n_formula_1": 16, "n_ref_uni": 52, "n_ref": 106, "n_ref_all": 123, "n_fig": 2, "n_tab": 5, "L_tab": 1042, "n_element_tab": 161, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1479, "n_element_tab_1": 115, "formula_len_all": 4155, "formula_len_all_1": 1145, "len_all": 171295, "len_all_1": 55770, "len_abs": 1415, "len_title": 126, "len_sents": 38649, "len_sents_1": 23201, "n_sents": 332, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1424, "title": "TangentBind: Unlocking the Potential of Emergent Alignment in Multimodal Model", "abs": "Improving the alignment of modalities has proven effective across various downstream tasks in multimodal models. Currently, modality alignment follows two main research directions: aligning all modalities simultaneously or binding the others by aligning to a core modality. The first ensures direct alignment, but it is difficult to extend to new modalities. The second is scalable but weak in emergent ability due to needing more direct inter-modality alignment. To address these problems, we propose the TangentBind. Specifically, we first align all modalities to a core modality, e.g., image or text. Then, we introduce a generative network that generates the embeddings of the second modality, e.g., text or image, based on the core modality embedding. Thirdly, other modalities, such as audio, are aligned to the core modality and generative embedding, improving emergent ability while retaining alignment with the core modality. During training, in addition to infoNCE, the Tangent Term is introduced to align the new modalities with the generated embeddings. This addresses accuracy issues caused by using generated vectors as representations for modalities. With VISION and TEXT as the core modality, our experiments include other modalities such as AUDIO, DEPTH, and INFRARED. Eventually, our experiments show that the emergent ability of TangentBind significantly outperforms the original benchmark on 9 datasets.", "keywords": [ "TangentBind", "Multi-modal Alignment", "Optimization" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "c2C2NQKjZw", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Binhua Li", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Fei Huang", "gender": "Male", "institution": "Alibaba Group US", "country": "US", "position": "Senior Research Director" }, { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" }, { "name": "Rongyu Cao", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Yingwei Ma", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yongbin Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zhenyu Pan", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "yongchang Cao", "gender": "unknown", "institution": "nanjing university", "country": "", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 69, "n_ref_all": 95, "n_fig": 7, "n_tab": 3, "L_tab": 1063, "n_element_tab": 137, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2604, "n_element_tab_1": 155, "formula_len_all": 13, "formula_len_all_1": 0, "len_all": 142083, "len_all_1": 70114, "len_abs": 1670, "len_title": 118, "len_sents": 44599, "len_sents_1": 34245, "n_sents": 303, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1672, "title": "Codev-Bench: How Do LLMs Understand Developer-Centric Code Completion?", "abs": "Code completion, a key downstream task in code generation, is one of the most frequent and impactful methods for enhancing developer productivity in software development. As intelligent completion tools evolve, we need a robust evaluation benchmark that enables meaningful comparisons between products and guides future advancements. However, existing benchmarks focus more on coarse-grained tasks without industrial analysis resembling general code generation rather than the real-world scenarios developers encounter. Moreover, these benchmarks often rely on costly and time-consuming human annotation, and the standalone test cases fail to leverage minimal tests for maximum repository-level understanding and code coverage. To address these limitations, we first analyze business data from an industrial code completion tool and redefine the evaluation criteria to better align with the developer's intent and desired completion behavior throughout the coding process\nBased on these insights, we introduce Codev-Agent, an agent-based system that automates repository crawling, constructs execution environments, extracts dynamic calling chains from existing unit tests, and generates new test samples to avoid data leakage, ensuring fair and effective comparisons. Using Codev-Agent, we present the Code-Development Benchmark (Codev-Bench), a fine-grained, real-world, repository-level, and developer-centric evaluation framework. Codev-Bench assesses whether a code completion tool can capture a developer's immediate intent and suggest appropriate code across diverse contexts, providing a more realistic benchmark for code completion in modern software development.", "keywords": [ "code completion", "code generation", "large language models", "software engineering" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 2, 4, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "c1Vn1RpB64", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Haike Xu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Kai-Wei Chang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Associate Professor" }, { "name": "Piotr Indyk", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Yizhou Sun", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "Zongyu Lin", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 4, "n_ref_uni": 22, "n_ref": 56, "n_ref_all": 90, "n_fig": 1, "n_tab": 9, "L_tab": 2741, "n_element_tab": 203, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1020, "n_element_tab_1": 61, "formula_len_all": 596, "formula_len_all_1": 220, "len_all": 143689, "len_all_1": 68224, "len_abs": 1777, "len_title": 107, "len_sents": 44563, "len_sents_1": 31415, "n_sents": 342, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1783, "title": "Contradiction Retrieval Via Sparse-Aware Sentence Embedding", "abs": "Contradiction retrieval refers to identifying and extracting documents that explicitly disagree with or refute the content of a query, which is important to many downstream applications like fact checking and data cleaning. To retrieve contradiction argument to the query from large document corpora, existing methods such as similarity search and crossencoder models exhibit significant limitations. The former struggles to capture the essence of contradiction due to its inherent nature of favoring similarity, while the latter suffers from computational inefficiency, especially when the size of corpora is large. To address these challenges, we introduce a novel approach: SparseCL that leverages specially trained sentence embeddings designed to preserve subtle, contradictory nuances between sentences. Our method utilizes a combined metric of cosine similarity and a sparsity function to efficiently identify and retrieve documents that contradict a given query. This approach dramatically enhances the speed of contradiction detection by reducing the need for exhaustive document comparisons to simple vector calculations. We validate our model using the Arguana dataset, a benchmark dataset specifically geared towards contradiction retrieval, as well as synthetic contradictions generated from the MSMARCO and HotpotQA datasets using GPT-4. Our experiments demonstrate the efficacy of our approach not only in contradiction retrieval with more than 30% accuracy improvements on MSMARCO and HotpotQA across different model architectures but also in applications such as cleaning corrupted corpora to restore high-quality QA retrieval. This paper outlines a promising direction for improving the accuracy and efficiency of contradiction retrieval in large-scale text corpora.", "keywords": [ "contradiction retrieval", "sentence embedding" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "c1RhJVTPwT", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Jinbo Yan", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Kaiqiang Xiong", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Lu Xiao", "gender": "Female", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Luyang Tang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Ronggang Wang", "gender": "Male", "institution": "Peking University Shenzhen Graduate School", "country": "CN", "position": "Full Professor" }, { "name": "Rui Peng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhiyan Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "WuJiahao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 21, "n_ref": 38, "n_ref_all": 64, "n_fig": 13, "n_tab": 5, "L_tab": 1761, "n_element_tab": 234, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1024, "n_element_tab_1": 128, "formula_len_all": 911, "formula_len_all_1": 510, "len_all": 126993, "len_all_1": 59733, "len_abs": 1484, "len_title": 161, "len_sents": 32643, "len_sents_1": 26350, "n_sents": 262, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 113, "L_abs": 1489, "title": "Swift4D: Adaptive divide-and-conquer Gaussian Splatting for compact and efficient reconstruction of dynamic scene", "abs": "Novel view synthesis has long been a practical but challenging task, although the introduction of numerous methods to solve this problem, even combining advanced representations like 3D Gaussian Splatting, they still struggle to recover high-quality results and often consume too much storage memory and training time. \nIn this paper we propose Swift4D, a divide-and-conquer 3D Gaussian Splatting method that can handle static and dynamic primitives separately, achieving a good trade-off between rendering quality and efficiency, motivated by the fact that most of the scene is the static primitive and does not require additional dynamic properties. Concretely, we focus on modeling dynamic transformations only for the dynamic primitives which benefits both efficiency and quality. We first employ a learnable decomposition strategy to separate the primitives, which relies on an additional parameter to classify primitives as static or dynamic. For the dynamic primitives, we employ a compact multi-resolution 4D Hash mapper to transform these primitives from canonical space into deformation space at each timestamp, and then mix the static and dynamic primitives to produce the final output. This divide-and-conquer method facilitates efficient training and reduces storage redundancy. Our method not only achieves state-of-the-art rendering quality while being 20\u00d7 faster in training than previous SOTA methods with a minimum storage requirement of only 30MB on real-world datasets.", "keywords": [ "3D Gaussian Splatting", "Dynamic scene reconstruction." ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "c1Ng0f8ivn", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Kyunghyun Cho", "gender": "Male", "institution": "Genentech", "country": "", "position": "Senior Director of Frontier Research" }, { "name": "Mark Ibrahim", "gender": "unknown", "institution": "Facebook AI Research (FAIR) Meta", "country": "US", "position": "Researcher" }, { "name": "Pietro Astolfi", "gender": "Male", "institution": "Meta", "country": "", "position": "Postdoc" }, { "name": "Randall Balestriero", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Assistant Professor" }, { "name": "Vivien Cabannes", "gender": "Not Specified", "institution": "META", "country": "FR", "position": "Researcher" }, { "name": "Yann LeCun", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Director" }, { "name": "Diane Nicole Bouchacourt", "gender": "Female", "institution": "Facebook AI Research", "country": "", "position": "Researcher" }, { "name": "Uladzislau Sobal", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 36, "n_ref": 86, "n_ref_all": 116, "n_fig": 15, "n_tab": 9, "L_tab": 1920, "n_element_tab": 351, "n_fig_1": 10, "n_tab_1": 5, "L_tab_1": 568, "n_element_tab_1": 102, "formula_len_all": 714, "formula_len_all_1": 537, "len_all": 153624, "len_all_1": 61127, "len_abs": 1661, "len_title": 136, "len_sents": 46863, "len_sents_1": 29092, "n_sents": 384, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 98, "L_abs": 1709, "title": "$\\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs", "abs": "Learning good representations involves capturing the diverse ways in which data samples relate. Contrastive loss\u2014an objective matching related samples\u2014underlies methods from self-supervised to multimodal learning. Contrastive losses, however, can be viewed more broadly as modifying a similarity graph to indicate how samples should relate in the embedding space. This view reveals a shortcoming in contrastive learning: the similarity graph is binary, as only one sample is the related positive sample. Crucially, similarities \\textit{across} samples are ignored. \nBased on this observation, we revise the standard contrastive loss to explicitly encode how a sample relates to others. We experiment with this new objective, called $\\mathbb{X}$-Sample Contrastive, to train vision models based on similarities in class or text caption descriptions.\nOur study spans three scales: ImageNet-1k with 1 million, CC3M with 3 million, and CC12M with 12 million samples. The representations learned via our objective outperform both contrastive self-supervised and vision-language models trained on the same data across a range of tasks. When training on CC12M, we outperform CLIP by $0.6\\%$ on both ImageNet and ImageNet Real. Our objective appears to work particularly well in lower-data regimes, with gains over CLIP of $17.2\\%$ on ImageNet and $18.0\\%$ on ImageNet Real when training with CC3M. Finally, our objective encourages the model to learn representations that separate objects from their attributes and backgrounds, with gains of $3.3$-$5.6$\\% over CLIP on ImageNet9. The proposed method takes a step towards developing richer learning objectives for understanding sample relations in foundation models.", "keywords": [ "contrastive learning", "vision language model", "simclr", "clip" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "c0PnZCNY2N", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Abhi Kamboj", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Minh N. Do", "gender": "Male", "institution": "VinUniversity", "country": "VN", "position": "Honorary Vice Provost" }, { "name": "Anh Duy Nguyen", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 34, "n_formula_1": 12, "n_ref_uni": 37, "n_ref": 69, "n_ref_all": 102, "n_fig": 10, "n_tab": 16, "L_tab": 10477, "n_element_tab": 1828, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2716, "n_element_tab_1": 326, "formula_len_all": 2874, "formula_len_all_1": 1182, "len_all": 233678, "len_all_1": 76420, "len_abs": 1054, "len_title": 92, "len_sents": 60733, "len_sents_1": 31052, "n_sents": 438, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1060, "title": "Robult: A Scalable Framework for Semi-Supervised Multimodal Learning with Missing Modalities", "abs": "In multimodal learning, the presence of missing modalities and limited labeled data presents significant challenges for building robust models. We propose **Robult**, a novel framework designed to address these challenges by leveraging an information-theoretic approach to preserve modality-specific features and synergistic information across modalities. Our model introduces two key objectives: (1) a latent reconstruction loss to retain unique modality-specific information, and (2) a novel soft Positive-Unlabeled (PU) contrastive loss to efficiently utilize sparse labeled data in semi-supervised settings. Robult seamlessly integrates into deep learning architectures, enhancing performance across multiple downstream tasks and ensuring robustness even when modalities are missing at inference time. Empirical results across diverse datasets demonstrate that Robult surpasses existing methods in handling both semi-supervised learning and missing modalities, while its lightweight design enables scalability and easy integration with existing frameworks.", "keywords": [ "Multimodal learning", "Semi-supervised learning", "Missing modalities" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "c0KPBFtGIy", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Chun Cao", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "Full Professor" }, { "name": "Jiahe Xu", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Jingwei Xu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Junyu Lai", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Yang Yao", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Yunpeng Huang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 1, "n_ref_uni": 22, "n_ref": 28, "n_ref_all": 42, "n_fig": 4, "n_tab": 4, "L_tab": 1150, "n_element_tab": 173, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 672, "n_element_tab_1": 149, "formula_len_all": 1607, "formula_len_all_1": 34, "len_all": 187185, "len_all_1": 65318, "len_abs": 1116, "len_title": 74, "len_sents": 67190, "len_sents_1": 32299, "n_sents": 626, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1083, "title": "Executing Arithmetic: Fine-Tuning Large Language Models as Turing Machines", "abs": "Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of natural language processing and reasoning tasks. However, their performance in the foundational domain of arithmetic remains unsatisfactory. When dealing with arithmetic tasks, LLMs often memorize specific examples rather than learning the underlying computational logic, limiting their ability to generalize to new problems. In this paper, we propose a Composable Arithmetic Execution Framework (CAEF) that enables LLMs to learn to execute step-by-step computations by emulating Turing Machines, thereby gaining a genuine understanding of computational logic. Moreover, the proposed framework is highly scalable, allowing composing learned operators to significantly reduce the difficulty of learning complex operators. In our evaluation, CAEF achieves nearly $100\\\\%$ accuracy across seven common mathematical operations on the LLaMA 3.1-8B model, effectively supporting computations involving operands with up to 100 digits, a level where GPT-4o falls short noticeably in some settings.", "keywords": [ "large language model", "arithmetic", "learn to execute" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "c01YB8pF0s", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 9, "author_info_list": [ { "name": "Danai Koutra", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Franck Dernoncourt", "gender": "unknown", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Huiyuan Chen", "gender": "Male", "institution": "VISA", "country": "", "position": "Researcher" }, { "name": "Namyong Park", "gender": "unknown", "institution": "Meta AI", "country": "US", "position": "Researcher" }, { "name": "Puja Trivedi", "gender": "Female", "institution": "University of Michigan", "country": "US", "position": "PhD student" }, { "name": "Tyler Derr", "gender": "unknown", "institution": "Vanderbilt University", "country": "US", "position": "Assistant Professor" }, { "name": "Yu Wang", "gender": "Male", "institution": "University of Oregon", "country": "US", "position": "Assistant Professor" }, { "name": "Nesreen Ahmed", "gender": "Female", "institution": "Intel AI Research", "country": "", "position": "Principal Researcher" }, { "name": "Ryan Rossi", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Senior Research Scientist" } ], "n_formula": 28, "n_formula_1": 11, "n_ref_uni": 39, "n_ref": 77, "n_ref_all": 132, "n_fig": 10, "n_tab": 20, "L_tab": 14360, "n_element_tab": 1365, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 2480, "n_element_tab_1": 234, "formula_len_all": 2279, "formula_len_all_1": 817, "len_all": 206211, "len_all_1": 68391, "len_abs": 6783, "len_title": 107, "len_sents": 50735, "len_sents_1": 29486, "n_sents": 326, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1871, "title": "A Large-scale Training Paradigm for Graph Generative Models", "abs": "Large Generative Models (LGMs) such as GPT, Stable Diffusion, Sora, and Suno are trained on a huge amount of language corpus, images, videos, and audio that are extremely diverse from numerous domains. This large-scale training paradigm on diverse well-curated data enhances the creativity and diversity of the generated content. However, all previous graph-generative models (e.g., GraphRNN, MDVAE, MoFlow, GDSS, and DiGress) have been trained only on one dataset each time, which cannot replicate the revolutionary success achieved by LGMs in other fields. To remedy this crucial gap, we propose a large-scale training paradigm that uses a large corpus of graphs (over 5000 graphs) from 13 domains, leading to the development of large graph generative models (LGGMs). We empirically demonstrate that the pre-trained LGGMs have superior zero-shot generative capability to existing graph generative models. Furthermore, our pre-trained LGGMs can be easily fine-tuned with graphs from target domains and demonstrate even better performance than those directly trained from scratch, behaving as a solid starting point for real-world customization. The generated graphs can boost the training data and lead to better graph classification performance. Inspired by Stable Diffusion, we further equip LGGMs with the Text-to-Graph generation capability, such as describing the network name and domain (i.e., \"The power-1138-bus graph represents a network of buses in a power distribution system.\") and network statistics (i.e., \"The graph has a low average degree, suitable for modeling social media interactions.\"). This Text-to-Graph capability integrates the extensive world knowledge in the underlying language model, offering users fine-grained control of the generated graphs. We release the code, the model checkpoint, and the datasets at https://github.com/GraphGG/LGGM.", "keywords": [ "Large-scale Training", "Graph Generative Model", "Diffusion model" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "bztzb1fyhv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Huanjing Yue", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jingyu Yang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" }, { "name": "Peng-Tao Jiang", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "Qirui Yang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Yihao Liu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yinbo Li", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "MS student" }, { "name": "cheng qihua", "gender": "Male", "institution": "Shenzhen MicroBT Electronics Technology Corporation", "country": "CN", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 43, "n_ref": 93, "n_ref_all": 115, "n_fig": 11, "n_tab": 9, "L_tab": 2930, "n_element_tab": 470, "n_fig_1": 8, "n_tab_1": 7, "L_tab_1": 2582, "n_element_tab_1": 439, "formula_len_all": 330, "formula_len_all_1": 309, "len_all": 135064, "len_all_1": 66386, "len_abs": 1427, "len_title": 109, "len_sents": 38657, "len_sents_1": 26496, "n_sents": 316, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1395, "title": "Learning Differential Pyramid Representation for Tone Mapping", "abs": "To display high dynamic range (HDR) images on low dynamic range (LDR) screens, tone mapping operations (TMO) are required to compress the dynamic range. Recently, the deep learning-based TMO methods with 3D Look-Up Table (LUT) have shown promising performance. However, these methods often fail to deliver satisfactory results in local areas, and generating image-level TMO on down-sampled low-resolution images leads to loss of details. To overcome this problem, we propose to construct a learnable differential pyramid representation network, termed DPRNet, for joint global and local tone mapping. Specifically, we construct multi-layer perceptrons to globally modulate the tones in pixel-level. Then, we propose a local 3D LUT, which generates the TMO coefficients in patch-level. To further enhance the details, we propose a learnable differential pyramid to capture multi-scale high-frequency components, coupled with an iterative mask learning strategy to refine high-frequency details. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art methods, improving PSNR by 2.58 dB in the HDR+ dataset and 3.31 dB in the HDRI Haven dataset respectively compared with the second best method. In addition, our method has the best generalization ability in unsupervised video TMO. We provide an anonymous online demo at https://xxxxxx2024.github.io/DPRNet/ .", "keywords": [ "Tone Mapping", "Differential Pyramid", "Image Signal Processor", "High Dynamic Range", "Image Retouching" ], "rating_list": [ 3, 5, 5, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "bzB7OIbITu", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Huanlong Zhong", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Jialin Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Jiangyi Deng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jie Zhang", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Shengyuan Pang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Wenyuan Xu", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Xinfeng Li", "gender": "Male", "institution": "Indiana University", "country": "US", "position": "Postdoc" }, { "name": "Yanjiao Chen", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 35, "n_formula_1": 32, "n_ref_uni": 25, "n_ref": 64, "n_ref_all": 83, "n_fig": 10, "n_tab": 7, "L_tab": 3170, "n_element_tab": 227, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2219, "formula_len_all_1": 2005, "len_all": 132092, "len_all_1": 63315, "len_abs": 1112, "len_title": 121, "len_sents": 41669, "len_sents_1": 28094, "n_sents": 354, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 121, "L_abs": 1884, "title": "Prompt-Independent Safe Decoding to Restrain Unsafe Image Generation for Text-to-Image Models against White-Box Adversary", "abs": "Text-to-image (T2I) models, developed through extensive training, are capable of generating realistic images from textual inputs, profoundly influencing various facets of our lives. Nevertheless, they can be exploited by adversaries who input malicious prompts, leading to the creation of unsafe content and posing serious ethical concerns. Current defense mechanisms primarily rely on external moderation or model modification, but they are inherently fragile against white-box adversaries who have access to the model's weights and can adjust them accordingly. \n\nTo address this issue, we propose \\sys, a novel defense framework that governs both the diffusion and the decoder module of the text-to-image pipeline, enabling them to reject generating unsafe content and resist malicious fine-tuning attempts. Concretely, we first fine-tune the diffusion and the decoder module with the denial-of-service samples: 1) for the diffusion module, the inputs are unsafe image-caption pairs, the ground truth is zero predicted noise, and 2) for the decoder module, the inputs are unsafe generations from the diffusion, the ground truth is zero decoding. Then, we employ adversarial training to ensure this denial-of-service behavior for unsafe queries remains effective even after the adversary's fine-tuning with unsafe data. Specifically, we continuously simulate potential fine-tuning processes that the adversary might adopt and expose them to the model, enabling it to learn how to resist.\n\nExtensive experiments validate that \\sys effectively prevents the generation of unsafe content without compromising the model\u2019s normal performance. Furthermore, our method demonstrates robust resistance to malicious fine-tuning by white-box adversaries, rendering it resource-intensive to corrupt our protected model, thus significantly deterring the misuse of our model for nefarious purposes.", "keywords": [ "Text-to-image generation", "AI security", "Model compliance" ], "rating_list": [ 8, 6, 3, 3 ], "soundness_list": [ 4, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "byofn4HN8F", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Alessandro Rinaldo", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Full Professor" }, { "name": "Purnamrita Sarkar", "gender": "Female", "institution": "University of Texas, Austin", "country": "US", "position": "Associate Professor" }, { "name": "Saptarshi Roy", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Postdoc" }, { "name": "Vansh Bansal", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" } ], "n_formula": 100, "n_formula_1": 31, "n_ref_uni": 35, "n_ref": 84, "n_ref_all": 95, "n_fig": 2, "n_tab": 3, "L_tab": 703, "n_element_tab": 123, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 482, "n_element_tab_1": 7, "formula_len_all": 10250, "formula_len_all_1": 2408, "len_all": 191279, "len_all_1": 74187, "len_abs": 1746, "len_title": 98, "len_sents": 56922, "len_sents_1": 33737, "n_sents": 586, "n_sents_1": 272, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1751, "title": "Straightness of Rectified Flow: A Theoretical Insight into Wasserstein Convergence", "abs": "Diffusion models have emerged as a powerful tool for image generation and denoising. Typically, generative models learn a trajectory between the starting noise distribution and the target data distribution. Recently Liu et al. (2023b) designed a novel alternative generative model Rectified Flow(RF), which aims to learn straight flow trajectories from noise to data using a sequence of convex optimization problems with close ties to optimal transport. If the trajectory is curved, one must use many Euler discretization steps or novel strategies, such as exponential integrators, to achieve a satisfactory generation quality. In contrast, RF has been shown to theoretically straighten the trajectory through successive rectifications, reducing the number of function evaluations (NFEs) while sampling. It has also been shown empirically that RF may improve the straightness in two rectifications if one can solve the underlying optimization problem within a sufficiently small error. In this paper, we make two key theoretical contributions: 1) we provide the first theoretical analysis of the\nWasserstein distance between the sampling distribution of RF and the target distribution. Our error rate is characterized by the number of discretization steps and a new formulation of straightness stronger than that in the original work. 2) under a mild regularity assumption, we show that for a rectified flow from a Gaussian to any general target distribution with finite first moment (e.g. mixture of Gaussians), two rectifications are sufficient to achieve a straight flow, which is in line with the previous empirical findings. Additionally, we also present empirical results on both simulated and real datasets to validate our theoretical findings.", "keywords": [ "generative model", "optimal transport", "rectified flow", "wasserstein distance" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "bykD2108Qv", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Grigoris Velegkas", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Weiwei Kong", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Monica Ribero Diaz", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 38, "n_formula_1": 30, "n_ref_uni": 16, "n_ref": 46, "n_ref_all": 59, "n_fig": 1, "n_tab": 2, "L_tab": 358, "n_element_tab": 28, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 1561, "n_element_tab_1": 36, "formula_len_all": 3566, "formula_len_all_1": 3213, "len_all": 89715, "len_all_1": 61720, "len_abs": 897, "len_title": 116, "len_sents": 27237, "len_sents_1": 22071, "n_sents": 245, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 905, "title": "Efficient Gradient Clipping Methods in DP-SGD for Convolution Models", "abs": "Differentially private stochastic gradient descent (DP-SGD) is a well-known method for training machine learning models with a specified level of privacy. \nHowever, its basic implementation is generally bottlenecked by the computation of the gradient norm (gradient clipping) for each example in an input batch. \nWhile various techniques have been developed to mitigate this issue, \nthere are only a handful of methods pertaining to convolution models, e.g., vision models.\nIn this work, we present three methods for performing gradient clipping that improve upon previous state-of-art methods. Two of these methods use in-place operations to reduce memory overhead, while the third one leverages a relationship between Fourier transforms and convolution layers. \nTo demonstrate the numerical efficiency of our methods, we also present several benchmark experiments that compare against other algorithms.", "keywords": [ "Differential Privacy", "SGD", "Clipping", "CNNs", "FFT", "DP-SGD", "Computational Complexity" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "byIsedbVo5", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Avrim Blum", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Full Professor" }, { "name": "Keziah Naggita", "gender": "Not Specified", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "PhD student" }, { "name": "Matthew Walter", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Associate Professor" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 27, "n_ref": 515, "n_ref_all": 597, "n_fig": 21, "n_tab": 10, "L_tab": 2594, "n_element_tab": 242, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 848, "n_element_tab_1": 26, "formula_len_all": 781, "formula_len_all_1": 610, "len_all": 228533, "len_all_1": 61885, "len_abs": 1558, "len_title": 117, "len_sents": 98467, "len_sents_1": 32834, "n_sents": 508, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1567, "title": "Learning Actionable Counterfactual Explanations in Large State Spaces", "abs": "An increasing number of high-stakes domains rely on machine learning to make decisions that have significant consequences for individuals, such as in loan approvals and college admissions. The black-box nature of these processes has led to a growing demand for solutions that make individuals aware of potential ways they could improve their qualifications.\nCounterfactual explanations (CFEs) are one form of feedback commonly used to provide insight into decision-making systems. Specifically, contemporary CFE generators provide explanations in the form of low-level CFEs whose constituent actions precisely describe how much a negatively classified individual should add to or subtract from their input features to achieve the desired positive classification.\nHowever, the low-level CFE generators have several shortcomings: they are hard to scale, often misaligned with real-world conditions, constrained by information access (e.g., they can not query the classifier), and make inadequate use of available historical data. \nTo address these challenges, we propose three data-driven CFE generators that create generalizable CFEs with desirable characteristics for individuals and decision-makers.\nThrough extensive empirical experiments, we compare the proposed CFE generators with a low-level CFE generator on four real-world (BRFSS, Foods, and two NHANES datasets), five semi-synthetic, and five variants of fully-synthetic datasets. \nOur problem can also be seen as learning an optimal policy in a family of large but deterministic Markov decision processes.", "keywords": [ "counterfactual explanations", "recourse", "data-driven algorithms", "fairness" ], "rating_list": [ 5, 1, 6, 5 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "bxqfyKl4lQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Haotong Qin", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Jing Lin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Jinjin Gu", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "Instructor" }, { "name": "Linghe Kong", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaokang Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University, China", "country": "CN", "position": "Professor" }, { "name": "Xin Yuan", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Associate Professor" }, { "name": "Yulun Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhiteng Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 36, "n_ref": 84, "n_ref_all": 98, "n_fig": 7, "n_tab": 8, "L_tab": 1971, "n_element_tab": 141, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1545, "n_element_tab_1": 136, "formula_len_all": 1580, "formula_len_all_1": 1582, "len_all": 113298, "len_all_1": 58254, "len_abs": 1423, "len_title": 98, "len_sents": 31008, "len_sents_1": 25898, "n_sents": 227, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1438, "title": "BiDRN: Binarized 3D Whole-body Human Mesh Recovery", "abs": "3D whole-body human mesh recovery aims to reconstruct the 3D human body, face, and hands from a single image. Although powerful deep learning models have achieved accurate estimation in this task, they require enormous memory and computational resources. Consequently, these methods can hardly be deployed on resource-limited edge devices. In this work, we propose a Binarized Dual Residual Network (BiDRN), a novel quantization method designed to estimate the 3D human body, face, and hands parameters efficiently. Specifically, we design a basic unit Binarized Dual Residual Block (BiDRB) composed of Local Convolution Residual (LCR) and Block Residual (BR), which can preserve as much full-precision information as possible. For LCR, we further generalize it to four kinds of convolutional modules so that full-precision information can be propagated even across mismatched dimensions when reshaping features. Additionally, we also binarize the face and hands box-prediction network as Binarized BoxNet, which further reduces the model redundancy. Comprehensive quantitative and qualitative experiments demonstrate the effectiveness of BiDRN, which has a significant improvement over state-of-the-art binarization algorithms. Moreover, our BiDRN achieves comparable performance with the full-precision method Hand4Whole while using only **22.1%** parameters and **14.8%** operations. We will release all the code and pretrained models.", "keywords": [ "3D whole-body human mesh recovery", "Binarization" ], "rating_list": [ 8, 3, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "bx0IbCcBvO", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Bohan Zhuang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Research Professor" }, { "name": "Feng Chen", "gender": "Male", "institution": "AIML", "country": "AU", "position": "PhD student" }, { "name": "Hong Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Jing Liu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Kaipeng Zhang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Wenqi Shao", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yefei He", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 107, "n_ref_all": 138, "n_fig": 8, "n_tab": 2, "L_tab": 1072, "n_element_tab": 194, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 698, "n_element_tab_1": 117, "formula_len_all": 291, "formula_len_all_1": 290, "len_all": 118776, "len_all_1": 58843, "len_abs": 1824, "len_title": 153, "len_sents": 31563, "len_sents_1": 27606, "n_sents": 211, "n_sents_1": 169, "venue": "ICLR2025", "decision": "withdraw", "L_title": 104, "L_abs": 1840, "title": "ZipVL: Efficient Large Vision-Language Models with Dynamic Token Sparsification and KV Cache Compression", "abs": "The efficiency of large vision-language models (LVLMs) is constrained by the computational bottleneck of the attention mechanism during the prefill phase and the memory bottleneck of fetching the key-value (KV) cache in the decoding phase, particularly in scenarios involving high-resolution images or videos. Visual content often exhibits substantial redundancy, resulting in highly sparse attention maps within LVLMs. This sparsity can be leveraged to accelerate attention computation or compress the KV cache through various approaches. However, most studies focus on addressing only one of these bottlenecks and do not adequately support dynamic adjustment of sparsity concerning distinct layers or tasks. In this paper, we present ZipVL, an efficient inference framework designed for LVLMs that resolves both computation and memory bottlenecks through a dynamic ratio of important tokens. This ratio is adaptively determined based on the layer-specific distribution of attention scores, rather than fixed hyper-parameters, thereby improving efficiency for less complex tasks while maintaining high performance for more challenging ones. Then we select important tokens based on their normalized attention scores and perform attention mechanism solely on those important tokens to accelerate the prefill phase. To mitigate the memory bottleneck in the decoding phase, we employ mixed-precision quantization to the KV cache, where high-bit quantization is used for caches of important tokens, while low-bit quantization is applied to those of less importance. Our experiments demonstrate that ZipVL can accelerate the prefill phase by 2.6$\\times$ and reduce GPU memory usage by 50.0\\%, with a minimal accuracy reduction of only 0.2\\% on Video-MME benchmark over LongVA-7B model, effectively enhancing the generation efficiency of LVLMs.", "keywords": [ "Efficient Models", "KV Cache Compression", "Vision-language Models" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "bwhLqFjsxd", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Ashton Anderson", "gender": "unknown", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "Associate Professor" }, { "name": "Difan Jiao", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Eric Xue", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" }, { "name": "Jon Kleinberg", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Full Professor" }, { "name": "Reid McIlroy-Young", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Postdoc" }, { "name": "Siddhartha Sen", "gender": "unknown", "institution": "Research, Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Zhenwei Tang", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 29, "n_ref": 52, "n_ref_all": 66, "n_fig": 4, "n_tab": 4, "L_tab": 1994, "n_element_tab": 93, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1994, "n_element_tab_1": 93, "formula_len_all": 299, "formula_len_all_1": 299, "len_all": 96553, "len_all_1": 65266, "len_abs": 1503, "len_title": 78, "len_sents": 33505, "len_sents_1": 31269, "n_sents": 220, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1421, "title": "Learning to Imitate with Less: Efficient Individual Behavior Modeling in Chess", "abs": "As humans seek to collaborate with, learn from, and better understand artificial intelligence systems, developing AI agents that can accurately emulate individual decision-making becomes increasingly important. Chess, with its long-standing role as a benchmark for AI research and its precise measurement of skill through chess ratings, provides an ideal environment for studying human-AI alignment. However, existing approaches to modeling human behavior require large amounts of data from each individual, making them impractical for new or sparsely represented users. In this work, we introduce Maia4All, a model designed to learn and adapt to individual decision-making styles efficiently, even with limited data. Maia4All achieves this by leveraging a two-stage fine-tuning method to bridge population and individual-level models and uses a meta-network to initialize and refine these embeddings with minimal data. Our experimental results show that Maia4All can accurately predict individual moves and profile behavioral patterns with high fidelity, establishing a new standard for personalized human-like AI behavior modeling in chess. Our work provides an example of how population AI systems can flexibly adapt to individual users using a prototype model as a bridge, which could lead to better and more accessible human-AI collaboration in other fields like education, healthcare, and strategic decision-making.", "keywords": [ "Human Behavior Modeling", "Chess", "Data-Efficient Learning", "Action Prediction", "Meta Learning" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 1, 2 ], "contribution_list": [ 3, 2, 1 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "bwhI6bCGY1", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Despoina Paschalidou", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Matan Atzmon", "gender": "Male", "institution": "NVIDIA", "country": "CA", "position": "Researcher" }, { "name": "Sanja Fidler", "gender": "Female", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "Associate Professor" }, { "name": "Sara Oblak", "gender": "Female", "institution": "University of Ljubljana", "country": "SI", "position": "MS student" } ], "n_formula": 109, "n_formula_1": 54, "n_ref_uni": 36, "n_ref": 72, "n_ref_all": 100, "n_fig": 18, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1621, "n_element_tab_1": 244, "formula_len_all": 7477, "formula_len_all_1": 3233, "len_all": 154781, "len_all_1": 65846, "len_abs": 1002, "len_title": 86, "len_sents": 44109, "len_sents_1": 25046, "n_sents": 383, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 38, "L_abs": 1006, "title": "ReMatching Dynamic Reconstruction Flow", "abs": "Reconstructing dynamic scenes from image inputs is a fundamental computer\nvision task with many downstream applications. Despite recent advancements, existing\napproaches still struggle to achieve high-quality reconstructions from unseen\nviewpoints and timestamps. This work introduces the ReMatching framework,\ndesigned to improve generalization quality by incorporating deformation priors into\ndynamic reconstruction models. Our approach advocates for velocity-field-based\npriors, for which we suggest a matching procedure that can seamlessly supplement\nexisting dynamic reconstruction pipelines. The framework is highly adaptable\nand can be applied to various dynamic representations. Moreover, it supports\nintegrating multiple types of model priors and enables combining simpler ones to\ncreate more complex classes. Our evaluations on popular benchmarks involving\nboth synthetic and real-world dynamic scenes demonstrate a clear improvement in\nreconstruction accuracy of current state-of-the-art models.", "keywords": [ "Dynamic Reconstruction", "Flow Modeling" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "bwgihJSDGg", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Enming Zhang", "gender": "Male", "institution": "Tsinghua", "country": "CN", "position": "PhD student" }, { "name": "Hanbing Liu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "J W", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Xiangyu Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yang Li", "gender": "Female", "institution": "Tsinghua Shenzhen International Graduate School", "country": "CN", "position": "Associate Professor" }, { "name": "Yanru Wu", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 21, "n_ref_uni": 41, "n_ref": 76, "n_ref_all": 94, "n_fig": 5, "n_tab": 7, "L_tab": 2781, "n_element_tab": 90, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 984, "n_element_tab_1": 77, "formula_len_all": 1340, "formula_len_all_1": 1456, "len_all": 148039, "len_all_1": 66091, "len_abs": 1665, "len_title": 142, "len_sents": 42634, "len_sents_1": 28779, "n_sents": 288, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1670, "title": "Exploiting Task Relationships for Continual Learning with Transferability-aware Task Embedding", "abs": "Continual learning (CL) has been a crucial topic in contemporary deep neural network usages, where catastrophic forgetting (CF) can impede a model's ability to progressively acquire knowledge, leading to critical training inefficiency and constraint in the improvement of model's overall capacity. Existing CL strategies mostly mitigate CF either by regularizing model weights and outputs during finetuning or by distinguishing task-specific and task-sharing model components to adapt the training process accordingly. Yet despite their effectiveness, these previous explorations are mainly limited to elements of task models, while we speculate a deeper exploitation of interrelationship among tasks can provide more enhancement for CL. Therefore, to better capture and utilize the task relations, we propose a transferability task embedding guided hypernet for continual learning. By introducing the information theoretical transferability based task embedding named H-embedding and incorporating it in a hypernetwork, we establish an online framework capable of capturing the statistical relations among the CL tasks and leveraging these knowledge for deriving task-conditioned model weights. The framework is also characterized by notable practicality, in that it only requires storing a low dimensional task embedding for each task, and can be efficiently trained in an end-to-end way. Extensive evaluations and experimental analyses on datasets including Permuted MNIST, Cifar10/100 and ImageNet-R showcase that our framework performs prominently compared to various baseline methods, as well as displays great potential in obtaining intrinsic task relationships.", "keywords": [ "Continual Learning", "Hypernetworks", "Task Embedding", "Transferability" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "bwVV0rHwrb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Alex Costanzino", "gender": "Male", "institution": "University of Bologna", "country": "IT", "position": "PhD student" }, { "name": "Giuseppe Lisanti", "gender": "Male", "institution": "University of Bologna", "country": "IT", "position": "Associate Professor" }, { "name": "Luigi di Stefano", "gender": "Male", "institution": "University of Bologna", "country": "IT", "position": "Full Professor" }, { "name": "Pierluigi Zama Ramirez", "gender": "Male", "institution": "University of Bologna", "country": "IT", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 20, "n_ref": 56, "n_ref_all": 96, "n_fig": 9, "n_tab": 9, "L_tab": 3371, "n_element_tab": 436, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 342, "n_element_tab_1": 74, "formula_len_all": 719, "formula_len_all_1": 651, "len_all": 136444, "len_all_1": 65678, "len_abs": 1321, "len_title": 132, "len_sents": 47942, "len_sents_1": 32994, "n_sents": 324, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1335, "title": "Forward-Backward Feature Transfer for Industrial Anomaly Detection and Segmentation", "abs": "Motivated by efficiency requirements, most industrial anomaly detection and segmentation (IADS) methods process low-resolution images, e.g., $224\\times 224$ pixels, obtained by downsampling the original input images.\nIn this setting, downsampling is typically applied also to the provided ground-truth defect masks.\nYet, as numerous industrial applications demand the identification of both large and small defects, this downsampling procedure may fail to reflect the actual performance achievable by current methods.\nIn this work, we propose a fast approach based on a novel Teacher-Student paradigm. \nThis paradigm relies on two shallow Student MLPs that learn to transfer patch features across the layers of a frozen Teacher Vision Transformer. \nOur framework can spot anomalies from high-resolution images faster than other methods, even when they process low-resolution images, achieving state-of-the-art overall performance on MVTec AD and segmentation results on VisA.\nWe also propose novel evaluation metrics that capture robustness regarding defect size, i.e., the ability of a method to preserve good localization from large anomalies to tiny ones, focusing on segmentation performance as a function of anomaly size. \nEvaluating our method with these metrics reveals its stable performance in detecting anomalies of any size.", "keywords": [ "anomaly", "detection", "segmentation", "localization" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "bwOndfohRK", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Aymeric Histace", "gender": "Male", "institution": "ETIS", "country": "FR", "position": "Full Professor" }, { "name": "Shuo Yang", "gender": "Male", "institution": "Ecole Nationale Sup\u00e9rieure de l'Electronique et de ses Applications", "country": "FR", "position": "PhD student" }, { "name": "Xuan Son Nguyen", "gender": "Male", "institution": "Ecole Nationale Sup\u00e9rieure de l'Electronique et de ses Applications", "country": "FR", "position": "Associate Professor" } ], "n_formula": 72, "n_formula_1": 20, "n_ref_uni": 42, "n_ref": 135, "n_ref_all": 162, "n_fig": 7, "n_tab": 12, "L_tab": 2692, "n_element_tab": 184, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 392, "n_element_tab_1": 18, "formula_len_all": 4879, "formula_len_all_1": 1689, "len_all": 222327, "len_all_1": 61745, "len_abs": 2453, "len_title": 102, "len_sents": 61310, "len_sents_1": 23120, "n_sents": 617, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1104, "title": "Neural networks on Symmetric Spaces of Noncompact Type", "abs": "Recent works have demonstrated promising performances of neural networks on hyperbolic spaces and symmetric positive definite (SPD) manifolds. These spaces belong to a family of Riemannian manifolds referred to as symmetric spaces of noncompact type. In this paper, we propose a novel approach for developing neural networks on such spaces. Our approach relies on a unified formulation of the distance from a point to a hyperplane on the considered spaces. We show that some existing formulations of the point-to-hyperplane distance can be recovered by our approach under specific settings. Furthermore, we derive a closed-form expression for the point-to-hyperplane distance in higher-rank symmetric spaces of noncompact type equipped with G-invariant Riemannian metrics. The derived distance then serves as a tool to design fully-connected (FC) layers and an attention mechanism for neural networks on the considered spaces. Our approach is validated on challenging benchmarks for image classification, electroencephalogram (EEG) signal classification, image generation, and natural language inference.", "keywords": [ "geometric deep learning", "symmetric spaces", "hyperbolic spaces", "SPD manifolds" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 3, 2, 3 ] }, { "paper_id": "bwJxUB0y46", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Morteza Mardani", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Principal Researcher" }, { "name": "Nicolas Zilberstein", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Santiago Segarra", "gender": "Male", "institution": "Rice University", "country": "", "position": "Assistant Professor" } ], "n_formula": 61, "n_formula_1": 25, "n_ref_uni": 59, "n_ref": 148, "n_ref_all": 195, "n_fig": 21, "n_tab": 15, "L_tab": 2283, "n_element_tab": 293, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2013, "n_element_tab_1": 61, "formula_len_all": 5992, "formula_len_all_1": 2043, "len_all": 258456, "len_all_1": 65547, "len_abs": 1116, "len_title": 112, "len_sents": 77737, "len_sents_1": 28389, "n_sents": 683, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1151, "title": "Repulsive Latent Score Distillation for Solving Inverse Problems", "abs": "Score Distillation Sampling (SDS) has been pivotal for leveraging pre-trained diffusion models in downstream tasks such as inverse problems, but it faces two major challenges: $(i)$ mode collapse and $(ii)$ latent space inversion, which become more pronounced in high-dimensional data. \nTo address mode collapse, we introduce a novel variational framework for posterior sampling. \nUtilizing the Wasserstein gradient flow interpretation of SDS, we propose a multimodal variational approximation with a \\emph{repulsion} mechanism that promotes diversity among particles by penalizing pairwise kernel-based similarity. \nThis repulsion acts as a simple regularizer, encouraging a more diverse set of solutions. \nTo mitigate latent space ambiguity, we extend this framework with an \\emph{augmented} variational distribution that disentangles the latent and data. \nThis repulsive augmented formulation balances computational efficiency, quality, and diversity. \nExtensive experiments on linear and nonlinear inverse tasks with high-resolution images ($512 \\times 512$) using pre-trained Stable Diffusion models demonstrate the effectiveness of our approach.", "keywords": [ "Inverse problems", "Score distillation", "Diffusion models", "Mode collapse", "Variational inference" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "bw9bvwVwMH", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Bing Li", "gender": "Male", "institution": "Clemson University", "country": "US", "position": "Assistant Professor" }, { "name": "Liang Yang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Longlong Jing", "gender": "Male", "institution": "Waymo LLC", "country": "US", "position": "Researcher" }, { "name": "Xiao Guo", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Yingwei Li", "gender": "Male", "institution": "Waymo LLC", "country": "US", "position": "Researcher" }, { "name": "Zhimin Chen", "gender": "Male", "institution": "Clemson University", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 34, "n_ref": 71, "n_ref_all": 85, "n_fig": 3, "n_tab": 6, "L_tab": 1810, "n_element_tab": 170, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1722, "n_element_tab_1": 139, "formula_len_all": 771, "formula_len_all_1": 636, "len_all": 127043, "len_all_1": 64115, "len_abs": 2956, "len_title": 120, "len_sents": 33771, "len_sents_1": 29098, "n_sents": 250, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1696, "title": "Point Cloud Self-supervised Learning via 3D to Multi-view Masked Leaner", "abs": "Recently, multi-modal masked autoencoders (MAE) has been introduced in 3D self-supervised learning, offering enhanced feature learning by leveraging both 2D and 3D data to capture richer cross-modal representations. However, these approaches have two limitations: (1) they inefficiently require both 2D and 3D modalities as inputs, even though the inherent multi-view properties of 3D point clouds already contain 2D modality.\n(2) input 2D modality causes the reconstruction learning to unnecessarily rely on visible 2D information, hindering 3D geometric representation learning.\nTo address these challenges, we propose a 3D to Multi-View Learner (Multi-View ML) that only utilizes 3D modalities as inputs and effectively capture rich spatial information in 3D point clouds. \nSpecifically, we first project 3D point clouds to multi-view 2D images at the feature level based on 3D-based pose.\nThen, we introduce two components: (1) a 3D to multi-view autoencoder that reconstructs point clouds and multi-view images from 3D and projected 2D features; \n(2) a multi-scale multi-head (MSMH) attention mechanism that facilitates local-global information interactions in each decoder transformer block through attention heads at various scales. \nAdditionally, a novel two-stage self-training strategy is proposed to align 2D and 3D representations.\nEmpirically, our method significantly outperforms state-of-the-art counterparts across various downstream tasks, including 3D classification, part segmentation, and object detection.\nSuch performance superiority showcases that Multi-View ML enriches the model's comprehension of geometric structures and inherent multi-modal properties of point clouds.", "keywords": [ "Self-supervised learning", "Multi-modality learning", "3D representation", "Masked autoencoder" ], "rating_list": [ 8, 8, 6, 3, 5 ], "soundness_list": [ 4, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 4, 3, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 4, 5 ] }, { "paper_id": "buxFBI6GG4", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Dongruo Zhou", "gender": "Male", "institution": "Indiana University", "country": "US", "position": "Assistant Professor" }, { "name": "Qin Zhang", "gender": "Male", "institution": "Indiana University at Bloomington", "country": "", "position": "Full Professor" }, { "name": "Tianyuan Jin", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" } ], "n_formula": 67, "n_formula_1": 20, "n_ref_uni": 28, "n_ref": 53, "n_ref_all": 61, "n_fig": 3, "n_tab": 1, "L_tab": 513, "n_element_tab": 7, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 65, "n_element_tab_1": 8, "formula_len_all": 7125, "formula_len_all_1": 1853, "len_all": 153236, "len_all_1": 67204, "len_abs": 737, "len_title": 140, "len_sents": 42122, "len_sents_1": 26771, "n_sents": 454, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 99, "L_abs": 743, "title": "Breaking the $\\log(1/\\Delta_2)$ Barrier: Better Batched Best Arm Identification with Adaptive Grids", "abs": "We investigate the problem of batched best arm identification in multi-armed bandits, where we want to find the best arm from a set of $n$ arms while minimizing both the number of samples and batches. We introduce an algorithm that achieves near-optimal sample complexity and features an instance-sensitive batch complexity, which breaks the $\\log(1/\\Delta_2)$ barrier. The main contribution of our algorithm is a novel sample allocation scheme that effectively balances exploration and exploitation for batch sizes. Experimental results indicate that our approach is more batch-efficient across various setups. We also extend this framework to the problem of batched best arm identification in linear bandits and achieve similar improvements.", "keywords": [ "Bandits" ], "rating_list": [ 6, 3, 6, 8 ], "soundness_list": [ 3, 2, 2, 4 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 4 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "burz7mU0YD", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Hiromi Wakaki", "gender": "Female", "institution": "Sony Group Corporation", "country": "JP", "position": "Senior Manager" }, { "name": "Mengjie Zhao", "gender": "unknown", "institution": "Sony", "country": "JP", "position": "Researcher" }, { "name": "Shiqi Yang", "gender": "Male", "institution": "SB Intuitions, SoftBank", "country": "JP", "position": "Chief Research Scientist" }, { "name": "Shusuke Takahashi", "gender": "Male", "institution": "Sony Group Corporation", "country": "JP", "position": "Researcher" }, { "name": "Wei-Hsiang Liao", "gender": "Male", "institution": "Sony Corporation", "country": "", "position": "Research Engineer" }, { "name": "Yuki Mitsufuji", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Lead Research Scientist" }, { "name": "Zhi Zhong", "gender": "Male", "institution": "Sony Group Corporation", "country": "", "position": "Researcher" }, { "name": "Zhuoyuan Mao", "gender": "Male", "institution": "Sony Group Corporation", "country": "JP", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 57, "n_ref": 117, "n_ref_all": 142, "n_fig": 4, "n_tab": 12, "L_tab": 3864, "n_element_tab": 463, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 3670, "n_element_tab_1": 458, "formula_len_all": 102, "formula_len_all_1": 14, "len_all": 182597, "len_all_1": 70630, "len_abs": 2720, "len_title": 101, "len_sents": 41387, "len_sents_1": 30188, "n_sents": 341, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 768, "title": "OpenMU: Your Swiss Army Knife for Music Understanding", "abs": "We present OpenMU-Bench, a large-scale benchmark suite for addressing the data scarcity issue in training multimodal language models to understand music. To construct OpenMU-Bench, we leveraged existing datasets and bootstrapped new annotations.\nOpenMU-Bench also broadens the scope of music understanding by including lyrics understanding and music tool usage. Using OpenMU-Bench, we trained our music understanding model, OpenMU, with extensive ablations, demonstrating that OpenMU outperforms baseline models such as MU-Llama. Both OpenMU and OpenMU-Bench are open-sourced to facilitate future research in music understanding and to enhance creative music production efficiency\\footnote{We will release the code, datasets, and model checkpoints upon acceptance.}.", "keywords": [ "Music understnading", "Multimodal Large Language Model" ], "rating_list": [ 1, 5, 3, 5 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "btqz4vMrUE", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Jingyi Liao", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xulei Yang", "gender": "Male", "institution": "I2R, A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Xun Xu", "gender": "Not Specified", "institution": "A*STAR", "country": "SG", "position": "Scientist" }, { "name": "Yongyi Su", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Liu Yifan", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 22, "n_ref": 43, "n_ref_all": 56, "n_fig": 5, "n_tab": 9, "L_tab": 8881, "n_element_tab": 1037, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1606, "n_element_tab_1": 273, "formula_len_all": 596, "formula_len_all_1": 502, "len_all": 136131, "len_all_1": 62242, "len_abs": 1289, "len_title": 154, "len_sents": 31841, "len_sents_1": 27472, "n_sents": 233, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 105, "L_abs": 1153, "title": "Test-Time Training for Out-of-Distribution Industrial Anomaly Detection via Robust Distribution Alignment", "abs": "Detecting anomalous patterns is essential for quality control in industrial applications, with state-of-the-art methods relying on large defect-free datasets to model normal distributions. However, robustness under domain shift, such as changes in lighting or sensor drift, remains a critical challenge in real-world deployment. An existing work, Generalized Normality Learning (GNL), addresses domain shifts by enforcing feature consistency through training-time augmentation, but its reliance on prior knowledge of target distributions and access to training data at inference limits flexibility. To overcome these limitations, we propose a memory bank-based anomaly detection method that avoids retraining or access to training data during inference. We improve the robustness to distribution shifts via distribution alignment based test-time training. Our approach leverages a modified Sinkhorn distance to align distributions and handle outliers, offering a more resilient solution for industrial anomaly detection under realistic constraints. Extensive evaluations on out-of-distribution anomaly detection benchmarks demonstrate the effectiveness.", "keywords": [ "Anomaly Detection", "Test Time Training; Out-of-Distribution" ], "rating_list": [ 1, 8, 3, 3 ], "soundness_list": [ 2, 4, 2, 2 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 1, 4, 2, 2 ], "confidence_list": [ 4, 2, 5, 3 ] }, { "paper_id": "btmHUbrfVj", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Yiming Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yining Sun", "gender": "Male", "institution": "Institute of Intelligent Machines\uff0cChinese Academy Of Sciences", "country": "CN", "position": "Principal Researcher" }, { "name": "Zenghui Ding", "gender": "Male", "institution": " Hefei Institute of Physical Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Zhaorun Chen", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Zhili Feng", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zhuokai Zhao", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 48, "n_ref": 74, "n_ref_all": 105, "n_fig": 4, "n_tab": 4, "L_tab": 897, "n_element_tab": 122, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 642, "n_element_tab_1": 77, "formula_len_all": 1078, "formula_len_all_1": 984, "len_all": 133078, "len_all_1": 55081, "len_abs": 1067, "len_title": 103, "len_sents": 31520, "len_sents_1": 25524, "n_sents": 181, "n_sents_1": 142, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 964, "title": "RANKCLIP: Ranking-Consistent Language-Image Pretraining", "abs": "Self-supervised contrastive learning models, such as CLIP, have set new benchmarks for vision-language models in many downstream tasks. However, their dependency on rigid one-to-one mappings overlooks the complex and often multifaceted relationships between and within texts and images. To this end, we introduce RankCLIP, a novel pretraining method that extends beyond the rigid one-to-one matching framework of CLIP and its variants. By extending the traditional pair-wise loss to list-wise, and leveraging both in-modal and cross-modal ranking consistency, RankCLIP improves the alignment process, enabling it to capture the nuanced many-to-many relationships between and within each modality. Through comprehensive experiments, we demonstrate the effectiveness of RankCLIP in various downstream tasks, notably achieving significant gains in zero-shot classifications over state-of-the-art methods, underscoring the importance of this enhanced learning process.", "keywords": [ "Vision and Language Alignment", "Contrastive Learning", "Ranking Consistency" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "btaxn8Xce6", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Alexander Rudikov", "gender": "unknown", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Researcher" }, { "name": "Ivan Oseledets", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "Tianchi Yu", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Fanaskov Vladimir", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Researcher" } ], "n_formula": 52, "n_formula_1": 37, "n_ref_uni": 30, "n_ref": 54, "n_ref_all": 84, "n_fig": 1, "n_tab": 3, "L_tab": 988, "n_element_tab": 114, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 737, "n_element_tab_1": 60, "formula_len_all": 2972, "formula_len_all_1": 2370, "len_all": 114920, "len_all_1": 55751, "len_abs": 1980, "len_title": 118, "len_sents": 28657, "len_sents_1": 22057, "n_sents": 224, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 2016, "title": "Astral: training physics-informed neural networks with error majorants", "abs": "The primal approach to physics-informed learning is a residual minimization. We argue that residual is, at best, an indirect measure of the error of approximate solution and propose to train with error majorant instead. Since error majorant provides a direct upper bound on error, one can reliably estimate how close PiNN is to the exact solution and stop the optimization process when the desired accuracy is reached. We call loss function associated with error majorant **Astral**: neur**A**l a po**ST**erio**R**i function**A**l **L**oss. To compare Astral and residual loss functions, we illustrate how error majorants can be derived for various PDEs and conduct experiments with diffusion equations (including anisotropic and in the L-shaped domain), convection-diffusion equation, temporal discretization of Maxwell's equation, magnetostatics and nonlinear elastoplasticity problems. The results indicate that Astral loss is competitive to the residual loss, typically leading to faster convergence and lower error (e.g., for Maxwell's equations, we observe an order of magnitude better relative error and training time). The main benefit of using Astral loss comes from its ability to estimate error, which is impossible with other loss functions. Our experiments indicate that the error estimate obtained with Astral loss is usually tight enough, e.g., for a highly anisotropic equation, on average, Astral overestimates error by a factor of $1.5$, and for convection-diffusion by a factor of $1.7$. We further demonstrate that Astral loss is better correlated with error than residual and is a more reliable predictor (in a statistical sense) of the error value. Moreover, unlike residual, the error indicator obtained from Astral loss has a superb spatial correlation with error. Backed with the empirical and theoretical results, we argue that one can productively use Astral loss to perform reliable error analysis and approximate PDE solutions with accuracy similar to standard residual-based techniques.", "keywords": [ "a posteriori error analysis", "functional error estimate", "PiNN", "physics-informed neural network", "scientific computing", "uncertainty quantification", "PDE" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 4, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "bsr78Cj2H7", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 6, "author_info_list": [ { "name": "Alexander Tschantz", "gender": "Male", "institution": "University of Sussex", "country": "", "position": "PhD student" }, { "name": "Christopher Buckley", "gender": "Male", "institution": "University of Sussex", "country": "", "position": "Full Professor" }, { "name": "Conor Heins", "gender": "unknown", "institution": "Max Planck Institute of Animal Behavior", "country": "", "position": "PhD student" }, { "name": "Dimitrije Markovic", "gender": "unknown", "institution": "Technische Universit\u00e4t Dresden", "country": "", "position": "Postdoc" }, { "name": "Hao Wu", "gender": "Male", "institution": "VERSES", "country": "US", "position": "Researcher" }, { "name": "Jeff Beck", "gender": "Unspecified", "institution": "Duke University", "country": "", "position": "Assistant Professor" } ], "n_formula": 34, "n_formula_1": 13, "n_ref_uni": 46, "n_ref": 84, "n_ref_all": 105, "n_fig": 4, "n_tab": 1, "L_tab": 400, "n_element_tab": 34, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 400, "n_element_tab_1": 34, "formula_len_all": 5027, "formula_len_all_1": 2089, "len_all": 146674, "len_all_1": 57678, "len_abs": 3554, "len_title": 110, "len_sents": 47031, "len_sents_1": 27247, "n_sents": 287, "n_sents_1": 151, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1690, "title": "Gradient-free variational learning with conditional mixture networks", "abs": "Bayesian methods are known to address some limitations of standard deep learning, such as the lack of calibrated predictions and uncertainty quantification. However, they can be computationally expensive as model and data complexity increase. Fast variational methods can reduce the computational requirements of Bayesian methods by eliminating the need for gradient descent or sampling, but are often limited to simple models. We demonstrate that conditional mixture networks (CMNs), a probabilistic variant of the mixture-of-experts (MoE) model, are suitable for fast, gradient-free inference and can solve complex classification tasks, thus balancing the expressiveness and scalability of neural networks with the probabilistic benefits of Bayesian methods . By exploiting conditional conjugacy and Polya-Gamma augmentation, we furnish Gaussian likelihoods for the weights of both the experts and the gating network. This enables efficient variational updates using coordinate ascent variational inference (CAVI), avoiding traditional gradient-based optimization. We validate this approach by training two-layer CMNs on standard benchmarks from the UCI repository. Our method, CAVI-CMN, achieves competitive and often superior predictive accuracy compared to maximum likelihood estimation (MLE) with backpropagation, while maintaining competitive runtime and full posterior distributions over all model parameters. Moreover, as input size or the number of experts increases, computation time scales competitively with MLE and other gradient-based solutions like black-box variational inference (BBVI), making CAVI-CMN a promising tool for deep, fast, and gradient-free Bayesian networks.", "keywords": [ "variational inference", "mixture-of-experts", "variational Bayes", "Mixture Models", "conjugate-exponential", "gradient-free", "Bayesian neural network" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "bsnRUkVn63", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Haoliang Li", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Hong Yan", "gender": "unknown", "institution": "City University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Kecheng Chen", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Pingping Zhang", "gender": "Female", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shiqi Wang", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Tiexin Qin", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 20, "n_ref_uni": 36, "n_ref": 88, "n_ref_all": 151, "n_fig": 9, "n_tab": 12, "L_tab": 4096, "n_element_tab": 297, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1108, "n_element_tab_1": 112, "formula_len_all": 2370, "formula_len_all_1": 1369, "len_all": 155793, "len_all_1": 74120, "len_abs": 1653, "len_title": 123, "len_sents": 47128, "len_sents_1": 30931, "n_sents": 349, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1688, "title": "Test-time Adaptation for Image Compression with Distribution Regularization", "abs": "Current test- or compression-time adaptation image compression (TTA-IC) approaches, which leverage both latent and decoder refinements as a two-step adaptation scheme, have potentially enhanced the rate-distortion (R-D) performance of learned image compression models on cross-domain compression tasks, \\textit{e.g.,} from natural to screen content images. However, compared with the emergence of various decoder refinement variants, the latent refinement, as an inseparable ingredient, is barely\n tailored to cross-domain scenarios. To this end, we are interested in developing an advanced latent refinement method by extending the effective hybrid latent refinement (HLR) method, which is designed for \\textit{in-domain} inference improvement but shows noticeable degradation of the rate cost in \\textit{cross-domain} tasks. Specifically, we first provide theoretical analyses, in a cue of marginalization approximation from in- to cross-domain scenarios, to uncover that the vanilla HLR suffers from an underlying mismatch between refined Gaussian conditional and hyperprior distributions, leading to deteriorated joint probability approximation of marginal distribution with increased rate consumption. To remedy this issue, we introduce a simple Bayesian approximation-endowed \\textit{distribution regularization} to encourage learning a better joint probability approximation in a plug-and-play manner. Extensive experiments on six in- and cross-domain datasets demonstrate that our proposed method not only improves the R-D performance compared with other latent refinement counterparts, but also can be flexibly integrated into existing TTA-IC methods with incremental benefits.", "keywords": [ "test-time adaptation", "image compression", "entropy coding" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 1, 3, 5, 4, 3 ] }, { "paper_id": "bsXxNkhvm6", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Hongkai Jiang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Wu Zhu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaolin Hu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 30, "n_ref": 115, "n_ref_all": 131, "n_fig": 2, "n_tab": 6, "L_tab": 7515, "n_element_tab": 479, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 3911, "n_element_tab_1": 247, "formula_len_all": 147, "formula_len_all_1": 13, "len_all": 151701, "len_all_1": 69709, "len_abs": 1525, "len_title": 106, "len_sents": 51023, "len_sents_1": 31920, "n_sents": 373, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1544, "title": "Benchmarking Machine Learning Methods for Stock Prediction", "abs": "Machine learning has been widely applied to stock movement prediction. However, research in this field is often hindered by the lack of high-quality benchmark datasets and comprehensive evaluation methods. To address these challenges, we introduce \\textit{BenchStock}, a benchmark that includes standardized datasets from the two largest stock markets (the U.S. and China) along with an evaluation method designed to facilitate a thorough examination of machine learning stock prediction methods. This benchmark covers a range of models, from traditional machine learning techniques to the latest deep learning approaches. Using BenchStock, we conducted large-scale experiments predicting individual stock returns over three decades in both markets to assess both short-term and long-term performance. To evaluate the impact of these predictions in actual market conditions, we constructed a portfolio based on the predictions and used a backtesting program to simulate its performance. The experiments revealed several key findings that have not been reported: 1) Most methods outperformed the S\\&P 500 in the U.S. market but experienced significant losses in the Chinese market. 2) Prediction accuracy of a method was not correlated with its portfolio return. 3) Advanced deep learning methods did not outperform traditional approaches. 4) The performance of the models was highly dependent on the testing period. These findings highlight the complexity of stock prediction and call for more in-depth machine learning research in this field.", "keywords": [ "Stock forecast", "Benchmark", "Evaluation", "AI4Finance" ], "rating_list": [ 3, 3, 3, 1, 3 ], "soundness_list": [ 2, 2, 1, 1, 2 ], "presentation_list": [ 3, 2, 3, 1, 2 ], "contribution_list": [ 1, 1, 1, 1, 2 ], "confidence_list": [ 5, 5, 5, 4, 4 ] }, { "paper_id": "bsFWJ0Kget", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Emanuele Zangrando", "gender": "Male", "institution": "Gran Sasso Science Institute", "country": "IT", "position": "PhD student" }, { "name": "Francesco Tudisco", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Associate Professor" }, { "name": "Gianluca Ceruti", "gender": "Male", "institution": "Universit\u00e4t Innsbruck", "country": "", "position": "Postdoc" }, { "name": "Jonas Kusch", "gender": "Male", "institution": "Norwegian University of Life Sciences", "country": "NO", "position": "Associate Professor" }, { "name": "Steffen Schotth\u00f6fer", "gender": "Male", "institution": "Oak Ridge National Laboratory", "country": "", "position": "Researcher" } ], "n_formula": 87, "n_formula_1": 19, "n_ref_uni": 29, "n_ref": 98, "n_ref_all": 161, "n_fig": 10, "n_tab": 8, "L_tab": 3219, "n_element_tab": 405, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1456, "n_element_tab_1": 165, "formula_len_all": 7917, "formula_len_all_1": 1670, "len_all": 193515, "len_all_1": 67933, "len_abs": 1225, "len_title": 115, "len_sents": 63534, "len_sents_1": 28968, "n_sents": 535, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1229, "title": "GeoLoRA: Geometric integration for parameter efficient fine-tuning", "abs": "Low-Rank Adaptation (LoRA) has become a widely used method for parameter-efficient fine-tuning of large-scale, pre-trained neural networks. However, LoRA and its extensions face several challenges, including the need for rank adaptivity, robustness, and computational efficiency during the fine-tuning process. We introduce GeoLoRA, a novel approach that addresses these limitations by leveraging dynamical low-rank approximation theory. GeoLoRA requires only a single backpropagation pass over the small-rank adapters, significantly reducing computational cost as compared to similar dynamical low-rank training methods and making it faster than popular baselines such as AdaLoRA. This allows GeoLoRA to efficiently adapt the allocated parameter budget across the model, achieving smaller low-rank adapters compared to heuristic methods like AdaLoRA and LoRA, while maintaining critical convergence, descent, and error-bound theoretical guarantees. The resulting method is not only more efficient but also more robust to varying hyperparameter settings. We demonstrate the effectiveness of GeoLoRA on several state-of-the-art benchmarks, showing that it outperforms existing methods in both\naccuracy and computational efficiency", "keywords": [ "Low Rank", "Finetuninge", "Robustness", "Rank Adaptive" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 2, 3, 3 ] }, { "paper_id": "breVfEOZLv", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Chengcheng Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Hanting Chen", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Jianyuan Guo", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Kai Han", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Yunhe Wang", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Chang Xu", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 82, "n_ref_all": 102, "n_fig": 2, "n_tab": 9, "L_tab": 3096, "n_element_tab": 285, "n_fig_1": 0, "n_tab_1": 11, "L_tab_1": 4248, "n_element_tab_1": 362, "formula_len_all": 201, "formula_len_all_1": 201, "len_all": 125036, "len_all_1": 69182, "len_abs": 1467, "len_title": 84, "len_sents": 31427, "len_sents_1": 28488, "n_sents": 230, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 36, "L_abs": 1474, "title": "Weak-to-Strong Enhanced Vision Model", "abs": "Recent advancements in large language and vision models have demonstrated extraordinary capabilities, driving researchers to train increasingly larger models in pursuit of even greater performance. However, smaller, easier-to-train models often exist prior to these larger models. In this paper, we explore how to effectively leverage these smaller, weaker models to assist in training larger, stronger models. Specifically, we investigate the concept of weak-to-strong knowledge distillation within vision models, where a weaker model supervises a stronger one, aiming to enhance the latter\u2019s performance beyond the limitations of the former.\nTo this end, we introduce a novel, adaptively adjustable loss function that dynamically calibrates the weaker model\u2019s supervision based on the discrepancy between soft labels and hard labels. This dynamic adjustment allows the weaker model to provide more effective guidance during training.\nOur comprehensive experiments span various scenarios, including few-shot learning, transfer learning, noisy label learning, and common knowledge distillation settings. The results are compelling: our approach not only surpasses benchmarks set by strong-to-strong distillation but also exceeds the performance of fine-tuning strong models on full datasets. These findings highlight the significant potential of weak-to-strong distillation, demonstrating its ability to substantially enhance vision model performance. Code will be released.", "keywords": [ "weak-to-strong enhancement", "knowledge distillation" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "br8YB7KMug", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Cewu Lu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Haowen Hou", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Intern" }, { "name": "Junxuan Liang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Undergrad student" }, { "name": "Xinpeng Liu", "gender": "Male", "institution": "Shanghai Innovation Institute", "country": "CN", "position": "PhD student" }, { "name": "Yong-Lu Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zili Lin", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 49, "n_ref": 99, "n_ref_all": 126, "n_fig": 15, "n_tab": 6, "L_tab": 1142, "n_element_tab": 140, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 931, "n_element_tab_1": 111, "formula_len_all": 374, "formula_len_all_1": 404, "len_all": 157766, "len_all_1": 76230, "len_abs": 1460, "len_title": 103, "len_sents": 39037, "len_sents_1": 31801, "n_sents": 378, "n_sents_1": 293, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1460, "title": "ImDy: Human Inverse Dynamics from Imitated Observations", "abs": "Inverse dynamics (ID), which aims at reproducing the driven torques from human kinematic observations, has been a critical tool for gait analysis. However, it is hindered from wider application to general motion due to its limited scalability. Conventional optimization-based ID requires expensive laboratory setups, restricting its availability. To alleviate this problem, we propose to exploit the recently progressive human motion imitation algorithms to learn human inverse dynamics in a data-driven manner. The key insight is that the human ID knowledge is implicitly possessed by motion imitators, though not directly applicable. In light of this, we devise an efficient data collection pipeline with state-of-the-art motion imitation algorithms and physics simulators, resulting in a large-scale human inverse dynamics benchmark as Imitated Dynamics (ImDy). ImDy contains over 150 hours of motion with joint torque and full-body ground reaction force data. With ImDy, we train a data-driven human inverse dynamics solver ImDyS(olver) in a fully supervised manner, which conducts ID and ground reaction force estimation simultaneously. Experiments on ImDy and real-world data demonstrate the impressive competency of ImDyS in human inverse dynamics and ground reaction force estimation. Moreover, the potential of ImDy(-S) as a fundamental motion analysis tool is exhibited with downstream applications. The project page is https://foruck.github.io/ImDy.", "keywords": [ "Motion Understanding", "Inverse Dynamics", "Biomechanics" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "br4H61LOoI", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Haiyun Jiang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Jiaya Jia", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Pengguang Chen", "gender": "Male", "institution": "SmartMore", "country": "HK", "position": "Researcher" }, { "name": "Shu Liu", "gender": "Male", "institution": "SmartMore Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhongshen Zeng", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 7, "n_ref_uni": 26, "n_ref": 49, "n_ref_all": 62, "n_fig": 8, "n_tab": 5, "L_tab": 2271, "n_element_tab": 262, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1374, "n_element_tab_1": 189, "formula_len_all": 270, "formula_len_all_1": 518, "len_all": 174465, "len_all_1": 61954, "len_abs": 1415, "len_title": 120, "len_sents": 56052, "len_sents_1": 29493, "n_sents": 404, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1419, "title": "MR-GSM8K: A Meta-Reasoning Benchmark for Large Language Model Evaluation", "abs": "In this work, we introduce a novel evaluation paradigm for Large Language Models\n(LLMs) that compels them to transition from a traditional question-answering role,\nakin to a student, to a solution-scoring role, akin to a teacher. This paradigm, focusing on \"reasoning about reasoning,\" termed meta-reasoning, shifts the emphasis\nfrom result-oriented assessments, which often neglect the reasoning process, to a\nmore comprehensive evaluation that effectively distinguishes between the cognitive\ncapabilities of different models. Our meta-reasoning process mirrors \"system-2\"\nslow thinking, requiring careful examination of assumptions, conditions, calculations, and logic to identify mistakes. This paradigm enables one to transform\nexisted saturated, non-differentiating benchmarks that might be leaked in data pretraining stage to evaluation tools that are both challenging and robust against data\ncontamination. To prove our point, we applied our paradigm to GSM8K dataset and\ndeveloped the MR-GSM8K benchmark. Our extensive analysis includes several\nstate-of-the-art models from both open-source and commercial domains, uncovering fundamental deficiencies in their training and evaluation methodologies.\nSpecifically, we found the OpenAI o1 models which possess characteristics of\n\"system-2\" thinking excel the other SOTA models by more than 20 absolute points\nin our benchmark, supporting our deficiency hypothesis.", "keywords": [ "Benchmark", "LLM", "Math", "Evaluation" ], "rating_list": [ 6, 3, 8, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 4, 2, 4, 3 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "bqv7M0wc4x", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Alejandro Ribeiro", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Full Professor" }, { "name": "Joshua Agterberg", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Juan Elenter", "gender": "unknown", "institution": "Spotify", "country": "US", "position": "Researcher" }, { "name": "Liangzu Peng", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Rene Vidal", "gender": "unknown", "institution": "University of Pennsylvania", "country": "US", "position": "Full Professor" } ], "n_formula": 104, "n_formula_1": 21, "n_ref_uni": 93, "n_ref": 209, "n_ref_all": 335, "n_fig": 25, "n_tab": 13, "L_tab": 1691, "n_element_tab": 177, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 248, "n_element_tab_1": 16, "formula_len_all": 17247, "formula_len_all_1": 1813, "len_all": 328571, "len_all_1": 67607, "len_abs": 1581, "len_title": 132, "len_sents": 83277, "len_sents_1": 27097, "n_sents": 787, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1606, "title": "ICL-TSVD: Bridging Theory and Practice in Continual Learning with Pre-trained Models", "abs": "The goal of continual learning (CL) is to train a model that can solve multiple tasks presented sequentially. Recent CL approaches have achieved strong performance by leveraging large pre-trained models that generalize well to downstream tasks. However, such methods lack theoretical guarantees, making them prone to unexpected failures. Conversely, principled CL approaches often fail to achieve competitive performance. In this work, we bridge this gap between theory and practice by integrating an empirically strong approach (RanPAC) into a principled framework, Ideal Continual Learner (ICL), designed to prevent forgetting. Specifically, we lift pre-trained features into a higher dimensional space and formulate an over-parametrized minimum-norm least-squares problem. We find that the lifted features are highly ill-conditioned, potentially leading to large training errors (numerical instability) and increased generalization errors (double descent). We address these challenges by continually truncating the singular value decomposition (SVD) of the lifted features. Our approach, termed ICL-TSVD, is stable with respect to the choice of hyperparameters, can handle hundreds of tasks, and outperforms state-of-the-art CL methods on multiple datasets. Importantly, our method satisfies a recurrence relation throughout its continual learning process, which allows us to prove it maintains small training and generalization errors by appropriately truncating a fraction of SVD factors. This results in a stable continual learning method with strong empirical performance and theoretical guarantees.", "keywords": [ "Continual Learning; Pretrained Models; Overparameterization; Generalization; Random Feature Models" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 2, 1 ] }, { "paper_id": "bqoHdVMIbt", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Brian Kulis", "gender": "unknown", "institution": "Boston University", "country": "", "position": "Associate Professor" }, { "name": "Christian So", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "Undergrad student" }, { "name": "Christopher Liao", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Theodoros Tsiligkaridis", "gender": "Male", "institution": "MIT Lincoln Laboratory, Massachusetts Institute of Technology", "country": "US", "position": "Senior AI Research Scientist" } ], "n_formula": 36, "n_formula_1": 9, "n_ref_uni": 70, "n_ref": 139, "n_ref_all": 181, "n_fig": 16, "n_tab": 11, "L_tab": 8201, "n_element_tab": 646, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2816, "n_element_tab_1": 68, "formula_len_all": 3058, "formula_len_all_1": 781, "len_all": 229102, "len_all_1": 66673, "len_abs": 1723, "len_title": 132, "len_sents": 57241, "len_sents_1": 28706, "n_sents": 467, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1809, "title": "Multimodal Unsupervised Domain Generalization by Retrieving Across the Modality Gap", "abs": "Domain generalization (DG) is an important problem that learns a model which generalizes to unseen test domains leveraging one or more source domains, under the assumption of shared label spaces. However, most DG methods assume access to abundant source data in the target label space, a requirement that proves overly stringent for numerous real-world applications, where acquiring the same label space as the target task is prohibitively expensive. For this setting, we tackle the multimodal version of the unsupervised domain generalization (MUDG) problem, which uses a large task-agnostic unlabeled source dataset during finetuning. Our framework does not explicitly assume any relationship between the source dataset and target task. Instead, it relies only on the premise that the source dataset can be accurately and efficiently searched in a joint vision-language space. We make three contributions in the MUDG setting. Firstly, we show theoretically that cross-modal approximate nearest neighbor search suffers from low recall due to the large distance between text queries and the image centroids used for coarse quantization. Accordingly, we propose paired k-means, a simple clustering algorithm that improves nearest neighbor recall by storing centroids in query space instead of image space. Secondly, we propose an adaptive text augmentation scheme for target labels designed to improve zero-shot accuracy and diversify retrieved image data. Lastly, we present two simple but effective components to further improve downstream target accuracy. We compare against state-of-the-art name-only transfer, source-free DG and zero-shot (ZS) methods on their respective benchmarks and show consistent improvement in accuracy on 20 diverse datasets. Code is available: https://github.com/Chris210634/mudg", "keywords": [ "Retrieval", "Domain Generalization", "Multimodal learning" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 4, 2, 2, 4 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "bqf0aCF3Dd", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Aimin Zhou", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Researcher" }, { "name": "Hong Qian", "gender": "unknown", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Huakang Lu", "gender": "Male", "institution": "East China Normal University", "country": "", "position": "MS student" }, { "name": "Ke Tang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Shuo Liu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Intern" }, { "name": "Xiang Shu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Xin An", "gender": "Female", "institution": "East China Normal University", "country": "", "position": "Undergrad student" }, { "name": "Yaolin Wen", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Yiyi Zhu", "gender": "Female", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Yang Yu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Professor" } ], "n_formula": 42, "n_formula_1": 3, "n_ref_uni": 27, "n_ref": 67, "n_ref_all": 95, "n_fig": 3, "n_tab": 11, "L_tab": 8160, "n_element_tab": 1240, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 954, "n_element_tab_1": 96, "formula_len_all": 3922, "formula_len_all_1": 188, "len_all": 217176, "len_all_1": 70073, "len_abs": 1832, "len_title": 132, "len_sents": 71177, "len_sents_1": 34643, "n_sents": 607, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1839, "title": "SOO-Bench: Benchmarks for Evaluating the Stability of Offline Black-Box Optimization", "abs": "Black-box optimization aims to find the optima through building a model close to the black-box objective function based on function value evaluation. However, in many real-world tasks, such as design of molecular formulas and mechanical structures, it is perilous, costly, or even infeasible to evaluate the objective function value of an actively sampled solution. In this situation, optimization can only be conducted via utilizing offline historical data, which yields offline black-box optimization. Different from the traditional goal that is to pursue the optimal solution, this paper at first discloses that the goal of offline optimization is to stably surpass the offline dataset during optimization procedure. Although benchmarks called Design-Bench already exist in this emerging field, it can hardly evaluate the stability of offline optimization, and mainly provides real-world offline tasks and the corresponding offline datasets. To this end, this paper proposes benchmarks named SOO-Bench (i.e., Stable Offline Optimization Benchmarks) for offline black-box optimization algorithms, so as to evaluate the stability of surpassing the offline dataset under different data distributions. Along with SOO-Bench, we also propose a stability indicator to measure the degree of stability. Specifically, SOO-Bench includes various real-world offline optimization tasks and offline datasets under different data distributions, involving the fields of satellites, materials science, structural mechanics and automobile manufacturing. Empirically, baseline and state-of-the-art algorithms are tested and analyzed on SOO-Bench. Hopefully, SOO-Bench is expected to serve as a catalyst for rapid developments of more novel and stable offline optimization methods. The code is available at https://anonymous.4open.science/r/SOO-Bench-9025.", "keywords": [ "Offline Optimization", "Black-Box Optimization", "Stability", "Benchmarks" ], "rating_list": [ 5, 8, 5, 6 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "bqUsdBeRjQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Liu Leqi", "gender": "Female", "institution": "Princeton University", "country": "", "position": "Postdoc" }, { "name": "Ruiyang Zhou", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Xinyu Li", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zachary Lipton", "gender": "Unspecified", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 10, "n_ref_uni": 37, "n_ref": 69, "n_ref_all": 104, "n_fig": 8, "n_tab": 10, "L_tab": 4137, "n_element_tab": 204, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 547, "n_element_tab_1": 7, "formula_len_all": 1516, "formula_len_all_1": 912, "len_all": 214125, "len_all_1": 76636, "len_abs": 1420, "len_title": 111, "len_sents": 75300, "len_sents_1": 37595, "n_sents": 468, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1425, "title": "Personalized Language Modeling from Personalized Human Feedback", "abs": "Personalized large language models (LLMs) are designed to tailor responses to individual user preferences. While Reinforcement Learning from Human Feedback (RLHF) is a commonly used framework for aligning LLMs with human preferences, vanilla RLHF assumes that all human preferences share the same distribution, preventing fine-tuned LLMs from generating personalized content when user preferences are diverse. In this work, we propose Personalized-RLHF (P-RLHF), an efficient framework that utilizes a lightweight user model to capture individual user preferences and jointly learns the user model and the personalized LLM from human feedback. P-RLHF exhibits the following three characteristics: (1) It enables an LLM to generate personalized content and scale efficiently with growing number of users. (2) It handles both explicit user preferences described as textual input and implicit user preferences encoded in the feedback data. (3) It eliminates the need for users to fully articulate their preferences, which are normally needed for prompting LLMs to generate personalized content yet are often impractical to obtain in real-world scenarios. Our experimental results show that personalized LLMs trained using P-RLHF generate responses that are more closely aligned with individual user preferences, outperforming vanilla, non-personalized RLHF and prompting-based personalization approaches across different tasks.", "keywords": [ "Large Language Models", "RLHF", "Personalization" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "bprVvbmL6T", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Chengchun Shi", "gender": "Male", "institution": "London School of Economics", "country": "GB", "position": "Associate Professor" }, { "name": "Liyuan Hu", "gender": "unknown", "institution": "London School of Economics and Political Science, University of London", "country": "", "position": "PhD student" }, { "name": "Meiling Hao", "gender": "Female", "institution": "University of International Business and Economics", "country": "CN", "position": "Associate Professor" }, { "name": "Pingfan Su", "gender": "Male", "institution": "London School of Economics and Political Science, University of London", "country": "GB", "position": "PhD student" }, { "name": "Qingyuan Zhao", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Associate Professor" }, { "name": "Zoltan Szabo", "gender": "Male", "institution": "London School of Economics", "country": "GB", "position": "Professor of Data Science" } ], "n_formula": 71, "n_formula_1": 15, "n_ref_uni": 48, "n_ref": 64, "n_ref_all": 82, "n_fig": 6, "n_tab": 4, "L_tab": 1430, "n_element_tab": 156, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8198, "formula_len_all_1": 1217, "len_all": 269233, "len_all_1": 73752, "len_abs": 1760, "len_title": 99, "len_sents": 74607, "len_sents_1": 33037, "n_sents": 719, "n_sents_1": 252, "venue": "ICLR2025", "decision": "withdraw", "L_title": 51, "L_abs": 996, "title": "Off-policy Evaluation with Deeply-abstracted States", "abs": "Off-policy evaluation (OPE) is crucial for assessing a target policy's impact offline before its deployment. However, achieving accurate OPE in large state spaces remains challenging. This paper studies state abstractions -- originally designed for policy learning -- in the context of OPE. Our contributions are three-fold: (i) We define a set of irrelevance conditions central to learning state abstractions for OPE, and derive a backward-model-irrelevance condition for achieving irrelevance in (marginalized) importance sampling ratios by constructing a time-reversed Markov decision process (MDP) based on the standard MDP. (ii) We propose a novel iterative procedure that sequentially projects the original state space into a smaller space, resulting in a deeply-abstracted state, which substantially simplify the sample complexity of OPE arising from high cardinality. (iii) We prove the Fisher consistencies of various OPE estimators when applied to our proposed abstract state spaces.", "keywords": [ "Markov decision process", "Off-policy evaluation", "State abstraction", "Reinforcement learning theory" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "bppG9srkpR", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Justin Kiefel", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "MS student" }, { "name": "Shrey Shah", "gender": "unknown", "institution": "University of Wisconsin, Madison", "country": "", "position": "Undergrad student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 45, "n_ref": 57, "n_ref_all": 71, "n_fig": 4, "n_tab": 1, "L_tab": 140, "n_element_tab": 25, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 140, "n_element_tab_1": 25, "formula_len_all": 46, "formula_len_all_1": 46, "len_all": 161432, "len_all_1": 77855, "len_abs": 742, "len_title": 0, "len_sents": 27333, "len_sents_1": 26659, "n_sents": 186, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 24, "L_abs": 744, "title": "LokiLM: Technical Report", "abs": "In this work, we introduce LokiLM, a 1.4B parameter large language model trained on 500B tokens. Our model performs strongly in natural language reasoning tasks and achieves state-of-the-art performance among models with 1.5B parameters or less. LokiLM is trained using multi-teacher knowledge distillation and high-quality training data to achieve benchmark results competitive with larger models trained on significantly more tokens. We support these findings by introducing steps to avoid benchmark contamination and overfitting throughout our development process. Despite its promising performance, LokiLM exhibits a concerning amount of hallucinations and scores poorly on the TruthfulQA benchmark, so we do not release the model publicly.", "keywords": [ "large language model" ], "rating_list": [ 3, 5, 6, 3, 1 ], "soundness_list": [ 2, 3, 2, 1, 2 ], "presentation_list": [ 2, 3, 4, 2, 2 ], "contribution_list": [ 2, 2, 2, 1, 1 ], "confidence_list": [ 4, 3, 4, 5, 4 ] }, { "paper_id": "bpdIZTIVq8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Chuheng Zhang", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Derek Cathera Yang", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "MS student" }, { "name": "Jiang Bian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Junliang Guo", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Li Zhao", "gender": "Female", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Pushi Zhang", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Tianyu He", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Researcher" }, { "name": "Xiaoyu Chen", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Graduate student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 62, "n_ref": 109, "n_ref_all": 121, "n_fig": 6, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 285, "formula_len_all_1": 288, "len_all": 189028, "len_all_1": 57467, "len_abs": 0, "len_title": 47, "len_sents": 38638, "len_sents_1": 28607, "n_sents": 264, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1332, "title": "IGOR: Image-GOal Representations are the Atomic Building Blocks for Next-Level Generalization in Embodied AI", "abs": "We introduce Image-GOal Representations (IGOR), aiming to learn a unified, semantically consistent action space across human and various robots. Through this unified latent action space, IGOR enables knowledge transfer among large-scale robot and human activity data. We achieve this by compressing visual changes between an initial image and its goal state into latent actions. IGOR allows us to generate latent action labels for internet-scale video data. This unified latent action space enables the training of foundation policy and world models across a wide variety of tasks performed by both robots and humans. We demonstrate that: (1) IGOR learns a semantically consistent action space for both human and robots, characterizing various possible motions of objects representing the physical interaction knowledge; (2) IGOR can \u201cmigrate\u201d the movements of the object in the one video to other videos, even across human and robots, by jointly using the latent action model and world model; (3) IGOR can learn to align latent actions with natural language through the foundation policy model, and integrate latent actions with a low-level policy model to achieve effective robot control. We believe IGOR opens new possibilities for human-to-robot knowledge transfer and control. See video demonstrations on our anonymous webpage.", "keywords": [ "Embodied AI", "Foundation Models", "Generalist Agents" ], "rating_list": [ 6, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "bpSN5YfSSZ", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Junyu Leng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Undergrad student" }, { "name": "Ke Wan", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Shuang Li", "gender": "Female", "institution": "The Chinese University of Hong Kong (Shenzhen)", "country": "", "position": "Assistant Professor" }, { "name": "Yang Yang", "gender": "Male", "institution": "The Chinese University of Hong Kong,Shenzhen", "country": "", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 28, "n_ref": 31, "n_ref_all": 38, "n_fig": 0, "n_tab": 1, "L_tab": 373, "n_element_tab": 56, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 276, "n_element_tab_1": 50, "formula_len_all": 805, "formula_len_all_1": 805, "len_all": 93298, "len_all_1": 61388, "len_abs": 1196, "len_title": 122, "len_sents": 31208, "len_sents_1": 29064, "n_sents": 249, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1201, "title": "Inferring the Invisible: Recurrent Neuro-Symbolic Forward Chaining Network", "abs": "A key challenge in artificial intelligence is inferring underlying factors that are not directly observable but are crucial for understanding and predicting complex behaviors. In this paper, we introduce a novel neural-symbolic framework that advances beyond traditional rule induction by integrating latent predicate discovery with rule learning. Our approach utilizes a recurrent unit to iteratively refine and learn rules from observed data, employing dynamic programming techniques to identify fixed points and solve complex problems. This framework enables the discovery of hidden predicates\u2014such as user engagement or underlying motivations\u2014that influence observable outcomes but are not directly grounded in the data.By encoding both explicit and latent predicates into a unified rule embedding, our method facilitates a deeper understanding of complex phenomena and enhances predictive accuracy. This joint learning process captures explicit relationships and invents new predicates essential for comprehensive inference. We validate our method across various tasks, demonstrating its capability to reveal hidden structures and enhance symbolic reasoning with deeper, more accurate insights.", "keywords": [ "Neuro Symbolic Reasoning" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "boz4LIvv3f", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Ryosuke Furuta", "gender": "Male", "institution": "The University of Tokyo, Tokyo University", "country": "JP", "position": "Assistant Professor" }, { "name": "Yoichi Sato", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Professor" } ], "n_formula": 22, "n_formula_1": 18, "n_ref_uni": 44, "n_ref": 142, "n_ref_all": 186, "n_fig": 10, "n_tab": 11, "L_tab": 5664, "n_element_tab": 733, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 473, "n_element_tab_1": 21, "formula_len_all": 2183, "formula_len_all_1": 1956, "len_all": 194969, "len_all_1": 70162, "len_abs": 1575, "len_title": 159, "len_sents": 50152, "len_sents_1": 29807, "n_sents": 393, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1406, "title": "Seeking Flat Minima with Mean Teacher on Semi- and Weakly-Supervised Domain Generalization for Object Detection", "abs": "Object detectors do not work well when domains largely differ between training and testing data. To overcome this domain gap in object detection without requiring expensive annotations, we consider two problem settings: semi-supervised domain generalizable object detection (SS-DGOD) and weakly-supervised DGOD (WS-DGOD). In contrast to the conventional domain generalization for object detection that requires labeled data from multiple domains, SS-DGOD and WS-DGOD require labeled data only from one domain and unlabeled or weakly-labeled data from multiple domains for training. In this paper, we show that object detectors can be effectively trained on the two settings with the same Mean Teacher learning framework, where a student network is trained with pseudo-labels output from a teacher on the unlabeled or weakly-labeled data. We provide novel interpretations of why the Mean Teacher learning framework works well on the two settings in terms of the relationships between the generalization gap and flat minima in parameter space. On the basis of the interpretations, we also show that incorporating a simple regularization method into the Mean Teacher learning framework leads to flatter minima. The experimental results demonstrate that the regularization leads to flatter minima and boosts the performance of the detectors trained with the Mean Teacher learning framework on the two settings.", "keywords": [ "object detection", "domain generalization", "semi-supervised learning", "weakly-supervised learning" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 1, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "bowetgeOMw", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Hanbo Huang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Hongyu Liang", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Lin Liu", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Shiyu Liang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yihan Li", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Yiran Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yuting Zheng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 50, "n_formula_1": 22, "n_ref_uni": 20, "n_ref": 24, "n_ref_all": 35, "n_fig": 6, "n_tab": 3, "L_tab": 2161, "n_element_tab": 252, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 134, "n_element_tab_1": 19, "formula_len_all": 6045, "formula_len_all_1": 2036, "len_all": 99675, "len_all_1": 51113, "len_abs": 1401, "len_title": 119, "len_sents": 30437, "len_sents_1": 23279, "n_sents": 267, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1408, "title": "Toward Trustworthy: A Method for Detecting Fine-Tuning Origins in LLMs", "abs": "As large language models (LLMs) continue to advance, their deployment often involves fine-tuning to enhance performance on specific downstream tasks. However, this customization is sometimes accompanied by misleading claims about the origins, raising significant concerns about transparency and trust within the open-source community. Existing model verification techniques typically assess functional, representational, and weight similarities. However, these approaches often struggle against obfuscation techniques, such as permutations and scaling transformations, that obscure a model's lineage. To address this limitation, we propose a novel detection method that rigorously determines whether a model has been fine-tuned from a specified base model. This method includes the ability to extract the LoRA rank utilized during the fine-tuning process, providing a more robust verification framework. This framework is the first to provide a formalized approach specifically aimed at pinpointing the sources of model fine-tuning. We empirically validated our method on twenty-nine diverse open-source models under conditions that simulate real-world obfuscation scenarios. We empirically analyze the effectiveness of our framework and finally, discuss its limitations. The results demonstrate the effectiveness of our approach and indicate its potential to establish new benchmarks for model verification.", "keywords": [ "Fine-Tuning Origins Detection", "LoRA", "LLM" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "bo36dzafRN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Carlo Vittorio Cannistraci", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Huaping Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jialin Zhao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xinghang Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yingtao Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 15, "n_ref_uni": 52, "n_ref": 93, "n_ref_all": 119, "n_fig": 6, "n_tab": 26, "L_tab": 9257, "n_element_tab": 982, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2017, "n_element_tab_1": 211, "formula_len_all": 2037, "formula_len_all_1": 786, "len_all": 222287, "len_all_1": 66058, "len_abs": 1786, "len_title": 131, "len_sents": 52806, "len_sents_1": 27721, "n_sents": 429, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1797, "title": "Sparse Spectral Training and Inference on Euclidean and Hyperbolic Neural Networks", "abs": "The growing demands on GPU memory posed by the increasing number of neural network parameters call for training approaches that are more memory-efficient. Previous memory reduction training techniques, such as Low-Rank Adaptation (LoRA) and ReLoRA, face challenges, with LoRA being constrained by its low-rank structure, particularly during intensive tasks like pre-training, and ReLoRA suffering from saddle point issues. In this paper, we propose Sparse Spectral Training (SST) to optimize memory usage for pre-training. SST updates all singular values and selectively updates singular vectors through a multinomial sampling method weighted by the magnitude of the singular values. Furthermore, SST employs singular value decomposition to initialize and periodically reinitialize low-rank parameters, reducing distortion relative to full-rank training compared to other low-rank methods. Through comprehensive testing on both Euclidean and hyperbolic neural networks across various tasks, including natural language generation, machine translation, node classification, link prediction, and image classification, SST demonstrates its ability to outperform existing memory reduction training methods and is comparable to full-rank training in various cases. On LLaMA-1.3B, with only 18.7\\% of the parameters trainable compared to full-rank training (using a rank equivalent to 6\\% of the embedding dimension), SST reduces the perplexity gap between other low-rank methods and full-rank training by 97.4\\%. This result highlights SST as an effective parameter-efficient technique for model pre-training, offering a promising new paradigm for achieving scalable and memory-efficient neural network training. Our code is available at https://anonymous.4open.science/r/sparse_spectral_training-6A2C/.", "keywords": [ "Efficient AI", "Parameter-efficient training", "Pre-training", "Hyperbolic Network" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "bntJK4NyIW", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 5, "author_info_list": [ { "name": "Haotian Yu", "gender": "Male", "institution": "Case Western Reserve University", "country": "US", "position": "PhD student" }, { "name": "Pan Li", "gender": "unknown", "institution": "Case Western Reserve University", "country": "US", "position": "Associate Professor" }, { "name": "Runchao Li", "gender": "Male", "institution": "Case Western Reserve University", "country": "US", "position": "PhD student" }, { "name": "Xianxuan Long", "gender": "Female", "institution": "Case Western Reserve University", "country": "US", "position": "PhD student" }, { "name": "Yao Fu", "gender": "Male", "institution": "Case Western Reserve University", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 28, "n_ref": 45, "n_ref_all": 58, "n_fig": 7, "n_tab": 1, "L_tab": 250, "n_element_tab": 45, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1959, "n_element_tab_1": 213, "formula_len_all": 448, "formula_len_all_1": 397, "len_all": 112020, "len_all_1": 68805, "len_abs": 1419, "len_title": 108, "len_sents": 31523, "len_sents_1": 29704, "n_sents": 269, "n_sents_1": 255, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1210, "title": "Decentralized Training of Transformer Models in Heterogeneous Network", "abs": "Training large transformer-based models like GPT-4 and Llama3 is prohibitively expensive, often requiring vast resources, such as tens of thousands of GPUs running simultaneously for months. Traditionally, these models are trained in specialized clusters with high-speed, uniform interconnections and computational capabilities, enabling efficient data and pipeline parallelism. However, these clusters are costly, while more affordable GPUs are widely distributed across the globe. Existing approaches, such as Swarm and Dapple, primarily focus on distributed learning across data centers. In this paper, we introduce a novel framework designed to handle heterogeneous devices and unstable communication environments. Our framework employs a hybrid approach, combining parameter server architectures, pipeline parallelism, and task pool strategies to effectively manage device disconnections. Through comprehensive time-cost analysis and graph clustering techniques, we derive a near-optimal resource allocation scheme. We compare our method with existing large-scale training approaches and demonstrate its effectiveness by training a large language model using gaming GPUs in real-world internet conditions.", "keywords": [ "Distributed Learning", "LLM" ], "rating_list": [ 3, 1, 3, 1 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "bnpeTgh29E", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Haeyun Lee", "gender": "Male", "institution": "Korea University of Technology and Education", "country": "KR", "position": "Assistant Professor" }, { "name": "Jae Youn Hwang", "gender": "Male", "institution": "DGIST", "country": "KR", "position": "Full Professor" }, { "name": "Kyungsu Lee", "gender": "Male", "institution": "Jeonbuk National University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 35, "n_formula_1": 17, "n_ref_uni": 28, "n_ref": 37, "n_ref_all": 67, "n_fig": 10, "n_tab": 8, "L_tab": 3373, "n_element_tab": 556, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1561, "n_element_tab_1": 188, "formula_len_all": 2373, "formula_len_all_1": 1378, "len_all": 140986, "len_all_1": 60846, "len_abs": 1010, "len_title": 110, "len_sents": 35586, "len_sents_1": 23902, "n_sents": 281, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1013, "title": "Sub-Domain Aware Granular Segmentation via Fine Tuning Network", "abs": "Recent advances in deep learning (DL) have led to improved vision-based algorithms. DL-based semantic segmentation, in particular, has enabled precise predictions using Convolutional Neural Networks (CNNs). State-of-the-art CNN-based networks have achieved high accuracy on various datasets in multiple fields, such as building, scene, and object segmentation. However, subdomain shifts between training and test sets within a single domain can cause degraded accuracy in fine-grained segmentation. To counter this, this paper introduces a novel Sub-Domain Adaptation (SDA) framework for fine-grained and granular segmentation, which divides one single domain into multiple sub-domains and optimizes the baseline-network for each sub-domain. The baseline-network is further fine-tuned by recognizing the domain of the input in run-time, leading to more accurate predictions. Benchmarks of scene parsing, autonomous driving, and aerial imagery demonstrate the superior performance of SDA for granular segmentation.", "keywords": [ "domain adaptation" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "bnmhMxz7PO", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Mingxuan Sun", "gender": "Female", "institution": "Louisiana State University", "country": "US", "position": "Associate Professor" }, { "name": "Taibiao Zhao", "gender": "Male", "institution": "Louisiana State University", "country": "", "position": "PhD student" }, { "name": "xiaobing chen", "gender": "Male", "institution": "Louisiana State University", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 18, "n_ref_uni": 21, "n_ref": 48, "n_ref_all": 59, "n_fig": 2, "n_tab": 8, "L_tab": 7549, "n_element_tab": 401, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 792, "n_element_tab_1": 66, "formula_len_all": 6262, "formula_len_all_1": 6632, "len_all": 122185, "len_all_1": 67828, "len_abs": 1275, "len_title": 84, "len_sents": 34136, "len_sents_1": 29754, "n_sents": 232, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1279, "title": "Enhancing LLM's interpretability for time series via multi-level aligned embeddings", "abs": "The adaptation of large language models (LLMs) to time series forecasting poses unique challenges, as time series data is continuous in nature, while LLMs operate on discrete tokens. Despite the success of LLMs in natural language processing (NLP) and other structured domains, aligning time series data with language-based representations while maintaining both predictive accuracy and interpretability remains a significant hurdle. Existing methods have attempted to reprogram time series data into text-based forms, but these often fall short in delivering meaningful, interpretable results.\nIn this paper, we propose a multi-text alignment framework for time series forecasting using LLMs that not only improves prediction accuracy but also enhances the interpretability of time series representations. Our method decomposes time series into trend, seasonality, and residual components, which are then reprogrammed into component-specific text representations. \nWe introduce a multi-level alignment mechanism, where component-specific embeddings are aligned with pre-trained word tokens, enabling more interpretable forecasts. \nExperiments on multiple datasets demonstrate that our method outperforms state-of-the-art models in accuracy while providing good interpretability.", "keywords": [ "time series", "Large Language Models", "cross modality" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 5 ] }, { "paper_id": "bnJgzAQjWf", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Francesco Croce", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Postdoc" }, { "name": "Francesco D'Angelo", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Nicolas Flammarion", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 111, "n_formula_1": 43, "n_ref_uni": 21, "n_ref": 39, "n_ref_all": 93, "n_fig": 15, "n_tab": 4, "L_tab": 6243, "n_element_tab": 714, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 64, "n_element_tab_1": 2, "formula_len_all": 15211, "formula_len_all_1": 5233, "len_all": 242308, "len_all_1": 79928, "len_abs": 2145, "len_title": 128, "len_sents": 74146, "len_sents_1": 28913, "n_sents": 607, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1693, "title": "Selective induction Heads: How Transformers Select Causal Structures in Context", "abs": "Transformers have exhibited exceptional capabilities in sequence modelling tasks, leveraging self-attention and in-context learning. Critical to this success are induction heads, attention circuits that enable copying tokens based on their previous occurrences. In this work, we introduce a novel synthetic framework designed to enable the theoretical analysis of transformers\u2019 ability to dynamically handle causal structures. Existing works rely on Markov Chains to study the formation of induction heads, revealing how transformers capture causal dependencies and learn transition probabilities in-context. However, they rely on a fixed causal structure that fails to capture the complexity of natural languages, where the relationship between tokens dynamically changes with context. To this end, our framework varies the causal structure through interleaved Markov chains with different lags while keeping the transition probabilities fixed. This setting unveils the formation of *Selective Induction Heads*, a new circuit that endows transformers with the ability to select the correct causal structure in-context. We empirically demonstrate that attention-only transformers learn this mechanism to predict the next token by identifying the correct lag and copying the corresponding token from the past. We provide a detailed construction of a 3-layer transformer to implement the selective induction head, and a theoretical analysis proving that this mechanism asymptotically converges to the maximum likelihood solution. Our findings advance the theoretical understanding of how transformers select causal structures, providing new insights into their functioning and interpretability.", "keywords": [ "Transformers", "Markov chain", "interpretability", "attention", "in-context learning" ], "rating_list": [ 5, 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3, 4 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 3, 2, 4, 3, 3 ] }, { "paper_id": "bnINPG5A32", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Abhishek Kumar", "gender": "unknown", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Constantine Caramanis", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Full Professor" }, { "name": "Litu Rout", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Nataniel Ruiz", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Sanjay Shakkottai", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Full Professor" }, { "name": "Wen-Sheng Chu", "gender": "unknown", "institution": "Google Research", "country": "US", "position": "Research Scientist" }, { "name": "Yujia Chen", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 62, "n_formula_1": 14, "n_ref_uni": 54, "n_ref": 281, "n_ref_all": 339, "n_fig": 23, "n_tab": 4, "L_tab": 655, "n_element_tab": 95, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1000, "n_element_tab_1": 72, "formula_len_all": 5220, "formula_len_all_1": 2097, "len_all": 224103, "len_all_1": 67144, "len_abs": 1174, "len_title": 126, "len_sents": 69050, "len_sents_1": 31100, "n_sents": 500, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 77, "L_abs": 1191, "title": "RB-Modulation: Training-Free Personalization using Stochastic Optimal Control", "abs": "We propose Reference-Based Modulation (RB-Modulation), a new plug-and-play solution for training-free personalization of diffusion models.\nExisting training-free approaches exhibit difficulties in (a) style extraction from reference images in the absence of additional style or content text descriptions, (b) unwanted content leakage from reference style images, and (c) effective composition of style and content. \nRB-Modulation is built on a novel stochastic optimal controller where a style descriptor encodes the desired attributes through a terminal cost. \nThe resulting drift not only overcomes the difficulties above, but also ensures high fidelity to the reference style and adheres to the given text prompt. \nWe also introduce a cross-attention-based feature aggregation scheme that allows RB-Modulation to decouple content and style from the reference image.\nWith theoretical justification and empirical evidence, our framework demonstrates precise extraction and control of *content* and *style* in a training-free manner. \nAdditionally, our method allows a seamless composition of content and style, which marks a departure from the dependency on external adapters or ControlNets", "keywords": [ "Inverse Problems", "Generative Modeling", "Diffusion Models", "Posterior Sampling", "Optimal Control" ], "rating_list": [ 8, 6, 8, 10 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "bmrYu2Ekdz", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Hailey Schoelkopf", "gender": "Female", "institution": "EleutherAI", "country": "US", "position": "Researcher" }, { "name": "Max M\u00fcller-Eberstein", "gender": "unknown", "institution": "IT University of Copenhagen", "country": "DK", "position": "Postdoc" }, { "name": "Naomi Saphra", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Fellow" }, { "name": "Oskar van der Wal", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Pietro Lesci", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Stella Biderman", "gender": "Female", "institution": "EleutherAI", "country": "", "position": "Researcher" }, { "name": "Willem H. Zuidema", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 49, "n_ref": 80, "n_ref_all": 114, "n_fig": 6, "n_tab": 1, "L_tab": 590, "n_element_tab": 78, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 585, "n_element_tab_1": 66, "formula_len_all": 144, "formula_len_all_1": 144, "len_all": 207689, "len_all_1": 64314, "len_abs": 1311, "len_title": 81, "len_sents": 37446, "len_sents_1": 30866, "n_sents": 257, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1300, "title": "PolyPythias: Stability and Outliers across Fifty Language Model Pre-Training Runs", "abs": "Understanding the stability of language model pre-training and its effects on downstream performance is still understudied. \nPrior work shows that the training process can yield significantly different results in response to slight variations in initial conditions, e.g., the random seed.\nCrucially, resources to study pre-training stability in language models are still lacking, especially for decoder-only models.\nWe introduce the PolyPythias, a set of 45 new training runs for the Pythia model suite: 9 new seeds across 5 model sizes, from 14M to 410M parameters, resulting in about 7k new checkpoints that we release.\nUsing these new 45 training runs, in addition to the 5 already available, we study the effects of different initial conditions determined by the seed---i.e., parameters' initialisation and data order---on (i) downstream performance, (ii) learned linguistic representations, and (iii) emergence of training phases.\nIn addition to common scaling behaviours, our analyses generally reveal highly consistent training dynamics across both model sizes and initial conditions.\nAdditionally, the new seeds for each model allow us to identify outlier training runs and delineate their characteristics.\nOur findings show the potential of using these methods to predict training stability.", "keywords": [ "language models", "training dynamics", "interpretability", "memorization", "robustness", "training stability" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "bmbRCRiNDu", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Michael Rabbat", "gender": "Male", "institution": "Facebook", "country": "CA", "position": "Principal Researcher" }, { "name": "Qinqing Zheng", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Sainbayar Sukhbaatar", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Research Scientist" }, { "name": "Yuandong Tian", "gender": "Male", "institution": "Meta AI (FAIR)", "country": "", "position": "Research Scientist" }, { "name": "andy su", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 54, "n_ref": 121, "n_ref_all": 134, "n_fig": 3, "n_tab": 5, "L_tab": 4407, "n_element_tab": 137, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 320, "n_element_tab_1": 26, "formula_len_all": 177, "formula_len_all_1": 0, "len_all": 233899, "len_all_1": 60950, "len_abs": 1997, "len_title": 150, "len_sents": 66281, "len_sents_1": 29510, "n_sents": 550, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 2087, "title": "Dualformer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces", "abs": "In human cognition theory, human thinking is governed by two systems: the fast and intuitive System 1 and the slower but more deliberative System 2. Recent studies have shown that incorporating System 2 process into Transformers including large language models (LLMs), significantly enhances their reasoning capabilities. Nevertheless, models that purely resemble System 2 thinking require substantially higher computational costs and are much slower to respond. To address this challenge, we present \\dualformer, a single Transformer model that seamlessly integrates both the fast and slow reasoning modes. \\dualformer is obtained by training on data with randomized reasoning traces, where different parts of the traces are dropped during training. The dropping strategies are specifically tailored according to the trace structure, analogous to analyzing our thinking process and creating shortcuts with patterns. At inference time, our model can be configured to output only the solutions (\\emph{fast mode}) or both the reasoning chain and the final solution (\\emph{slow mode}), or automatically decide which mode to engage (\\emph{auto mode}). In all cases, \\dualformer outperforms the corresponding baseline models in both performance and computational efficiency: \\textbf{(1)} in slow mode, \\dualformer optimally solves unseen $30 \\times 30$ maze navigation tasks $97.6\\%$ of the time, surpassing the \\searchformer (trained on data with complete reasoning traces) baseline performance of 93.3\\%, while only using $45.5\\%$ fewer reasoning steps; \\textbf{(2)} in fast mode, \\dualformer completes those tasks with an $80\\%$ optimal rate, significantly outperforming the Solution-Only model (trained on solution-only data), which has an optimal rate of only 30\\%; \\textbf{(3)} when operating in auto mode, \\dualformer achieves an optimal rate of 96.6\\% while utilizing $59.9\\%$ fewer reasoning steps compared to \\searchformer. For math problems, our techniques have also achieved improved performance with LLM fine-tuning, showing its generalization beyond task-specific models.", "keywords": [ "planning", "reasoning", "Sequential Decision Making" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 5, 5, 3 ] }, { "paper_id": "blwWIKpwpL", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Chenjia Bai", "gender": "Male", "institution": "TeleAI, China Telecom", "country": "CN", "position": "Research Scientist" }, { "name": "Jiafei Lyu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Runze Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Shengjie Sun", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xiu Li", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Yali Du", "gender": "Female", "institution": "King's College London", "country": "GB", "position": "Associate Professor" } ], "n_formula": 21, "n_formula_1": 14, "n_ref_uni": 41, "n_ref": 108, "n_ref_all": 151, "n_fig": 3, "n_tab": 9, "L_tab": 2662, "n_element_tab": 253, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2145, "n_element_tab_1": 159, "formula_len_all": 2191, "formula_len_all_1": 1351, "len_all": 175968, "len_all_1": 69165, "len_abs": 1335, "len_title": 114, "len_sents": 48302, "len_sents_1": 30196, "n_sents": 383, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1230, "title": "VLP: Vision-Language Preference Learning for Embodied Manipulation", "abs": "Reward engineering is one of the key challenges in Reinforcement Learning (RL). Preference-based RL effectively addresses this issue by learning from human feedback. However, it is both time-consuming and expensive to collect human preference labels. In this paper, we propose a novel Vision-Language Preference learning framework, named VLP, which learns a vision-language preference model to provide preference feedback for embodied manipulation tasks. To achieve this, we define three types of language-conditioned preferences and construct a vision-language preference dataset, which contains versatile implicit preference orders without human annotations. The preference model learns to extract language-related features, and then serves as a preference annotator in various downstream tasks. The policy can be learned according to the annotated preferences via reward learning or direct policy optimization. Extensive empirical results on simulated embodied manipulation tasks demonstrate that our method provides accurate preferences and generalizes to unseen tasks and unseen language, outperforming the baselines by a large margin. The code and videos of our method are available on the website: https://VLPref.github.io.", "keywords": [ "reinforcement learning", "preference-based reinforcement learning", "vision-language alignment", "offline reinforcement learning" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "blgJ4g00rC", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Enhong Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xianwei Meng", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yihang Lu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Xu yangyang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 18, "n_formula_1": 16, "n_ref_uni": 37, "n_ref": 73, "n_ref_all": 107, "n_fig": 14, "n_tab": 13, "L_tab": 8780, "n_element_tab": 1151, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 5604, "n_element_tab_1": 577, "formula_len_all": 1477, "formula_len_all_1": 1488, "len_all": 173478, "len_all_1": 73747, "len_abs": 1448, "len_title": 167, "len_sents": 51850, "len_sents_1": 29925, "n_sents": 372, "n_sents_1": 195, "venue": "ICLR2025", "decision": "withdraw", "L_title": 119, "L_abs": 1453, "title": "TimeCapsule: Solving the Jigsaw Puzzle of Long-Term Time Series Forecasting with Compressed Predictive Representations", "abs": "Recent deep learning models for long-term time series forecasting (LTSF) often emphasize complex, handcrafted designs and traditional methodologies, while simpler architectures like linear models or MLPs have occasionally outperformed these intricate solutions. In this paper, we revisit and organize the core ideas behind several key techniques, such as redundancy reduction and multi-scale modeling, which are frequently employed in advanced LTSF models. Our goal is to streamline these ideas for more efficient deep learning utilization. To this end, we introduce TimeCapsule, a model built around the principle of high-dimensional information compression that unifies these key ideas in a generalized yet simplified framework. Specifically, we model time series as a 3D tensor, incorporating temporal, variate, and level dimensions, and leverage mode production to capture multi-mode dependencies while achieving dimensionality compression. We propose an internal forecast within the compressed representation domain, supported by the Joint-Embedding Predictive Architecture (JEPA) to monitor the learning of predictive representations. Extensive experiments on challenging benchmarks demonstrate the versatility of our method, showing that TimeCapsule can achieve performance comparable to state-of-the-art models. More importantly, the structure of our model yields intriguing empirical findings, prompting a rethinking of approaches in this area.", "keywords": [ "multivariate long-term time series forecasting; deep learning; infomation tensor modeling" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "blSYKTWurU", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Xiangxin Zhou", "gender": "Not Specified", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xiwei Cheng", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Yu Bao", "gender": "Male", "institution": "ByteDance Research", "country": "CN", "position": "Researcher" }, { "name": "Yuwei Yang", "gender": "unknown", "institution": "ByteDance AI Lab", "country": "", "position": "Researcher" } ], "n_formula": 25, "n_formula_1": 24, "n_ref_uni": 33, "n_ref": 78, "n_ref_all": 102, "n_fig": 10, "n_tab": 10, "L_tab": 3587, "n_element_tab": 339, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2760, "n_element_tab_1": 66, "formula_len_all": 2109, "formula_len_all_1": 2082, "len_all": 154242, "len_all_1": 62877, "len_abs": 1384, "len_title": 121, "len_sents": 47474, "len_sents_1": 27942, "n_sents": 320, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1395, "title": "Decomposed Direct Preference Optimization for Structure-Based Drug Design", "abs": "Diffusion models have achieved promising results for Structure-Based Drug Design (SBDD). Nevertheless, high-quality protein subpocket and ligand data are relatively scarce, which hinders the models' generation capabilities. Recently, Direct Preference Optimization (DPO) has emerged as a pivotal tool for aligning generative models with human preferences. In this paper, we propose DecompDPO, a structure-based optimization method aligns diffusion models with pharmaceutical needs using multi-granularity preference pairs. DecompDPO introduces decomposition into the optimization objectives and obtains preference pairs at the molecule or decomposed substructure level based on each objective's decomposability. Additionally, DecompDPO introduces a physics-informed energy term to ensure reasonable molecular conformations in the optimization results. Notably, DecompDPO can be effectively used for two main purposes: (1) fine-tuning pretrained diffusion models for molecule generation across various protein families, and (2) molecular optimization given a specific protein subpocket after generation. Extensive experiments on the CrossDocked2020 benchmark show that DecompDPO significantly improves model performance, achieving up to 95.2\\% Med. High Affinity and a 36.2\\% success rate for molecule generation, and 100\\% Med. High Affinity and a 52.1\\% success rate for molecular optimization.", "keywords": [ "structure-based drug design", "diffusion model", "direct preference optimization" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "blNaExRx7Q", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Lei Yu", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Mor Geva", "gender": "Female", "institution": "Tel Aviv University", "country": "IL", "position": "Assistant Professor" }, { "name": "Shauli Ravfogel", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "Yihuai Hong", "gender": "Male", "institution": "International Digital Economy Academy, International Digital Economy Academy", "country": "CN", "position": "Intern" }, { "name": "Haiqin Yang", "gender": "Male", "institution": "International Digital Economy Academy (IDEA)", "country": "", "position": "Principal Researcher" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 40, "n_ref": 98, "n_ref_all": 130, "n_fig": 13, "n_tab": 9, "L_tab": 4379, "n_element_tab": 370, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1890, "n_element_tab_1": 136, "formula_len_all": 467, "formula_len_all_1": 155, "len_all": 204944, "len_all_1": 74583, "len_abs": 2158, "len_title": 116, "len_sents": 59549, "len_sents_1": 34965, "n_sents": 389, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1644, "title": "Intrinsic Evaluation of Unlearning Using Parametric Knowledge Traces", "abs": "The task of \u201cunlearning\u201d certain concepts in large language models (LLMs) has attracted immense attention recently, due to its importance in mitigating undesirable model behaviours, such as the generation of harmful, private, or incorrect information. Current protocols to evaluate unlearning methods largely rely on behavioral tests, without monitoring the presence of unlearned knowledge within the model\u2019s parameters. This residual knowledge can be adversarially exploited to recover the erased information post-unlearning. We argue that unlearning should also be evaluated internally, by considering changes in the parametric knowledge traces of the unlearned concepts. To this end, we propose a general evaluation methodology that leverages vocabulary projections to inspect concepts encoded in model parameters. We use this approach to localize \u201cconcept vectors\u201d \u2014 parameter vectors that encode concrete concepts \u2014 and construct ConceptVectors, a benchmark dataset containing hundreds of common concepts and their parametric knowledge traces within two open-source LLMs. Evaluation on ConceptVectors shows that existing unlearning methods minimally impact concept vectors and mostly suppress them during inference, while directly ablating these vectors demonstrably removes the associated knowledge and significantly reduces the model\u2019s susceptibility to adversarial manipulation. Our results highlight limitations in behavioral-based unlearning evaluations and call for future work to include parameter-based evaluations. To support this, we release our code and benchmark at https://anonymous.4open.science/r/ConceptVectors_review-98EF.", "keywords": [ "LLM Safety", "LLM Interpretability", "LLM Unlearning", "benchmark", "evaluations" ], "rating_list": [ 8, 3, 6, 5, 6 ], "soundness_list": [ 3, 1, 3, 2, 2 ], "presentation_list": [ 4, 1, 3, 3, 3 ], "contribution_list": [ 4, 1, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "bjxuqI4KwU", "primary_area": "causal reasoning", "n_author": 2, "author_info_list": [ { "name": "Pouria Ramazi", "gender": "Male", "institution": "Brock University", "country": "", "position": "Assistant Professor" }, { "name": "Vahideh Sanjaroonpouri", "gender": "Female", "institution": "Brock University", "country": "", "position": "Researcher" } ], "n_formula": 62, "n_formula_1": 10, "n_ref_uni": 36, "n_ref": 60, "n_ref_all": 73, "n_fig": 0, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 101, "n_element_tab_1": 3, "formula_len_all": 5270, "formula_len_all_1": 733, "len_all": 160813, "len_all_1": 64270, "len_abs": 1138, "len_title": 123, "len_sents": 51499, "len_sents_1": 29309, "n_sents": 503, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 75, "L_abs": 1213, "title": "Linear SCM Identification in the Presence of Confounders and Gaussian Noise", "abs": "Noisy linear structural causal models (SCMs) in the presence of confounding variables are known to be identifiable if all confounding and noise variables are non-Gaussian and unidentifiable if all are Gaussian.\n The identifiability when only some are Gaussian remains concealed. \n We show that, in the presence of Gaussian noise, a linear SCM is uniquely identifiable provided that \\emph{(i)} the number of confounders is at most the number of the observed variables, \\emph{(ii)} the confounders do not have a Gaussian component, and \\emph{(iii)} the causal structure of the SCM is known.\n If the third condition is relaxed, the SCM becomes finitely identifiable; more specifically, it belongs to a set of at most $n!$ linear SCMS, where $n$ is the number of observed variables.\n The confounders in all of these $n!$ SCMs share the same joint probability distribution function (PDF), which we obtain analytically. \n For the case where both the noise and confounders are Gaussian, we provide further insight into the existing counter-example-based unidentifiability result and demonstrate that every SCM with confounders can be represented as an SCM without confounders but with the same joint PDF.", "keywords": [ "identifiability", "SCM", "causal discovery; linear SCM; confounder" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "bjkQTInGes", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Hao Wen", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Lu Sheng", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Associate Professor" }, { "name": "Xinyuan Chen", "gender": "Female", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Yaohui Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Zehuan Huang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 2, "n_ref_uni": 47, "n_ref": 138, "n_ref_all": 159, "n_fig": 10, "n_tab": 7, "L_tab": 1392, "n_element_tab": 126, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 640, "n_element_tab_1": 56, "formula_len_all": 616, "formula_len_all_1": 145, "len_all": 153701, "len_all_1": 59518, "len_abs": 1352, "len_title": 116, "len_sents": 36078, "len_sents_1": 28912, "n_sents": 249, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1355, "title": "Ouroboros3D: Image-to-3D Generation via 3D-aware Recursive Diffusion", "abs": "Existing image-to-3D creation methods typically split the task into multi-view image generation and 3D reconstruction, leading to two main limitations: (1) multi-view bias, where geometric inconsistencies arise because multi-view diffusion models ensure image-level rather than 3D consistency; (2) misaligned reconstruction data, since reconstruction models trained on mostly synthetic data misalign when processing generated multi-view images during inference. To address these issues, we propose Ouroboros3D, a unified framework that integrates multi-view generation and 3D reconstruction into a recursive diffusion process. By incorporating a 3D-aware feedback mechanism, our multi-view diffusion model leverages the explicit 3D information from the reconstruction results of the previous denoising process as conditions, thus modeling consistency at the 3D geometric level. Furthermore, through joint training of both the multi-view diffusion and reconstruction models, we alleviate reconstruction bias due to data misalignment and enable mutual enhancement within the multi-step recursive process. Experimental results demonstrate that Ouroboros3D outperforms methods that treat these stages separately and those that combine them only during inference, achieving superior multi-view consistency and producing 3D models with higher geometric realism.", "keywords": [ "3D; Video Diffusion Model; 3D generation" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "bjcsVLoHYs", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Guangsheng Bao", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Hongbo Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Jindong Wang", "gender": "Male", "institution": "William & Mary", "country": "US", "position": "Assistant Professor" }, { "name": "Linyi Yang", "gender": "unknown", "institution": "University College London", "country": "GB", "position": "Researcher" }, { "name": "Minjun Zhu", "gender": "Female", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Yixuan Weng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Yue Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 4, "n_ref_uni": 71, "n_ref": 114, "n_ref_all": 148, "n_fig": 10, "n_tab": 16, "L_tab": 5396, "n_element_tab": 178, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2033, "n_element_tab_1": 80, "formula_len_all": 1163, "formula_len_all_1": 357, "len_all": 313274, "len_all_1": 66439, "len_abs": 5277, "len_title": 186, "len_sents": 97858, "len_sents_1": 30839, "n_sents": 682, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 66, "L_abs": 1744, "title": "CycleResearcher: Improving Automated Research via Automated Review", "abs": "The automation of scientific discovery has been a long-standing goal within the research community, driven by the potential to accelerate knowledge creation. While significant progress has been made using commercial large language models (LLMs) as research assistants or idea generators, the possibility of automating the entire research process with open-source LLMs remains largely unexplored. This paper explores the feasibility of using open-source post-trained LLMs as autonomous agents capable of performing the full cycle of automated research and review, from literature review and manuscript preparation to peer review and paper Refinement. Our iterative preference training framework consists of CycleResearcher, which conducts research tasks, and CycleReviewer, which simulates the peer review process, providing iterative feedback via reinforcement learning. To train these models, we develop two new datasets, Review-5k and Research-8k, reflecting real-world machine learning research and peer review dynamics. Our results demonstrate that CycleReviewer achieves promising performance with a 26.89\\% improvement in mean absolute error (MAE) over individual human reviewers in predicting paper scores, suggesting the potential of LLMs in supporting expert-level research evaluation. In research, the papers generated by the CycleResearcher model achieved a score of 5.36 in simulated peer reviews, showing competitive performance compared to the preprint level of 5.24 from human experts while still having room for improvement compared to the accepted paper level of 5.69. This work represents a significant step toward fully automated scientific inquiry, providing ethical safeguards and advancing AI-driven research capabilities.", "keywords": [ "Large Language Models", "Automation of Scientific Discovery", "AI Scientist" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "bjMi9ux50f", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Archit Sharma", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Aviral Kumar", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Chelsea Finn", "gender": "Female", "institution": "Physical Intelligence", "country": "US", "position": "Researcher" }, { "name": "Max Sobol Mark", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Mohan Kumar Srirama", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Researcher" }, { "name": "Tian Gao", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Georgia Gabriela da Silva Sampaio", "gender": "Female", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 22, "n_formula_1": 22, "n_ref_uni": 43, "n_ref": 96, "n_ref_all": 124, "n_fig": 10, "n_tab": 6, "L_tab": 1240, "n_element_tab": 153, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 952, "n_element_tab_1": 128, "formula_len_all": 2259, "formula_len_all_1": 2198, "len_all": 180564, "len_all_1": 76546, "len_abs": 1699, "len_title": 134, "len_sents": 61655, "len_sents_1": 36525, "n_sents": 409, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 28, "L_abs": 1845, "title": "Parameterization Agnostic RL", "abs": "Recent advances in learning decision-making policies can largely be attributed to training expressive policy models, largely via imitation learning. While imitation discards non-expert data, offline and/or online fine-tuning via reinforcement learning (RL) can still learn from suboptimal data. However, instantiating RL training of a new policy class often presents a different challenge: most deep RL machinery is co-developed with assumptions on the policy class, resulting in poor performance when the policy class changes. For e.g., SAC utilizes a low-variance reparameterization policy gradient for Gaussian policies, but this is unstable for diffusion policies and intractable for autoregressive (e.g., transformer) categorical policies. This implies that current RL algorithms may not perform well or may even not be applicable when the policy class changes. To address this issue, we develop an offline RL and online fine-tuning approach called **parameterization-agnostic RL** (**PA-RL**) that can effectively train multiple policy classes, with varying architectures. The basic idea is that a universal supervised learning loss can replace the policy improvement step in RL, as long as it is applied on \"optimized\" actions. To obtain these optimized actions, we first sample multiple actions from a base policy, and run global optimization (i.e., re-ranking multiple action samples using the Q-function) and local optimization (i.e., running gradient steps on an action sample) to maximize the critic on these candidates. PA-RL enables fine-tuning diffusion and autoregressive policies entirely via RL, while improving performance and sample-efficiency compared to existing online RL fine-tuning methods. PA-RL allows us to successfully fine-tune diffusion policies and OpenVLA, a 7B parameter generalist robot policy, on real robots.", "keywords": [ "reinforcement learning", "offline RL", "online fine-tuning", "online rl", "diffusion policies" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 3, 1, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 5, 4 ] }, { "paper_id": "bj9P8nt5hp", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Nayeong Kim", "gender": "Female", "institution": "POSTECH", "country": "KR", "position": "PhD student" }, { "name": "Seong Joon Oh", "gender": "Male", "institution": "Parameter Lab", "country": "DE", "position": "Principal Researcher" }, { "name": "Suha Kwak", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 2, "n_ref": 2, "n_ref_all": 4, "n_fig": 0, "n_tab": 1, "L_tab": 94, "n_element_tab": 7, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 94, "n_element_tab_1": 7, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 19990, "len_all_1": 19990, "len_abs": 157, "len_title": 108, "len_sents": 6811, "len_sents_1": 6811, "n_sents": 65, "n_sents_1": 65, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1569, "title": "GroupCoOp: Group-robust Fine-tuning via Group Prompt Learning", "abs": "Parameter-efficient fine-tuning (PEFT) of vision-language models (VLMs) excels in various vision tasks thanks to the rich knowledge and generalization ability of VLMs. However, recent studies revealed that such fine-tuned VLMs are vulnerable to spurious correlations stemming from the subgroup imbalance in the fine-tuning datasets. To resolve this issue, we propose Group Context Optimization (GroupCoOp), a simple and effective debiased fine-tuning algorithm that enhances the group robustness of fine-tuned VLMs without group labels. Its key idea is to employ group-specific text prompts as group representatives serving as multiple classifiers for their target class. The rich semantic knowledge of the text encoder of VLM enables the discovery of effective group prompts even for groups with a small number of training samples. Leveraging the group prompts for each class addresses the issues caused by the group-imbalanced training set, such as the neglect of minority groups and the scattered distribution of each class in the embedding space. Moreover, we propose a simple yet fairly effective pseudo group labeling algorithm, which allows GroupCoOp to fine-tune VLMs without manual group labels. GroupCoOp achieved the best results on five benchmarks across five CLIP architectures and even outperformed prior methods that train the entire network, despite training only 0.016\\% of the network's parameters. GroupCoOp demonstrates robust performance even with extremely limited training samples, where the minority group sample is limited to a single instance.", "keywords": [ "spurious correlation", "debiasing", "group robustness", "prompt learning", "vision-language models" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 3, 4, 2, 2 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 2, 3, 2, 4 ] }, { "paper_id": "bhUIoQ61pA", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Bin Liang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Hang Liu", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "MS student" }, { "name": "Houde Liu", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Linqi Ye", "gender": "Male", "institution": "Shanghai University", "country": "", "position": "Associate Professor" }, { "name": "Xueqian Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yi Cheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "zhenghan", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 13, "n_ref": 16, "n_ref_all": 35, "n_fig": 11, "n_tab": 4, "L_tab": 1692, "n_element_tab": 183, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 1876, "n_element_tab_1": 20, "formula_len_all": 580, "formula_len_all_1": 576, "len_all": 106239, "len_all_1": 49093, "len_abs": 386, "len_title": 126, "len_sents": 29404, "len_sents_1": 22004, "n_sents": 200, "n_sents_1": 153, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1649, "title": "HuWo\uff1aBuilding Physical Interaction World Models for Humanoid Robot Locomotion", "abs": "Reinforcement Learning control has been proved to be an effective approach for quadruped robot locomotion. However, locomotion tasks for humanoid robots are challenging, especially in complex environments. The main reason is that humanoid robots must maintain balance during movement and constantly engage in complex dynamic interactions with the environment. Understanding robot-environment interaction dynamics is key to achieving stable locomotion for humanoid robots. Since there is privileged information that the robot cannot directly access, to expand the observable space, previous reinforcement learning-based methods either reconstruct environmental information from partial observations or reconstruct robotic dynamics information from partial observations, but they fall short of fully capturing the dynamics of robot-environment interactions. In this work, we propose an end-to-end reinforcement learning control framework based on physical interaction$\\textbf{Wo}$rld Model for $\\textbf{Hu}$manoid Robots (HuWo). Our key innovation is to introduce a physical interaction world model to understand the interaction between the robot and environment, employing the hidden layers of transformer-XL for implicit modeling of this process across temporal sequences. The proposed framework can showcase robust and flexible locomotion ability in complex environments such as slopes, stairs, and discontinuous surfaces. We validated the robustness of this method using the $Zerith1$ robot, both in simulations and real-world deployments, and quantitatively compared our HuWo against the baselines with better traversability and command-tracking.", "keywords": [ "Reinforcement Learning", "Humanoid Locomotion", "World Mode", "Sim-to-Real Transfer" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 3, 3, 2, 1 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "bhOysNJvWm", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Enver Sangineto", "gender": "unknown", "institution": "University of Trento", "country": "", "position": "Postdoc" }, { "name": "Fabrizio Garuti", "gender": "Male", "institution": "Dipartimento di Ingegneria \"Enzo Ferrari\"", "country": "", "position": "PhD student" }, { "name": "Lorenzo Forni", "gender": "Male", "institution": "University of Padua", "country": "", "position": "Full Professor" }, { "name": "Rita Cucchiara", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Simone Luetto", "gender": "Male", "institution": "Polytechnic Institute of Turin", "country": "", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 40, "n_ref": 144, "n_ref_all": 186, "n_fig": 18, "n_tab": 7, "L_tab": 2683, "n_element_tab": 248, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 150, "n_element_tab_1": 10, "formula_len_all": 950, "formula_len_all_1": 476, "len_all": 198319, "len_all_1": 76608, "len_abs": 931, "len_title": 110, "len_sents": 74854, "len_sents_1": 40413, "n_sents": 580, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 941, "title": "Diffusion Transformers for Tabular Data Time Series Generation", "abs": "Tabular data generation has recently attracted a growing interest due to its different application scenarios. However, \ngenerating time series of tabular data, where each element of the series depends on the others,\nremains a largely unexplored domain. \nThis gap is probably due to the difficulty of jointly solving different problems, the main of which are the heterogeneity of tabular data (a problem common to non-time-dependent approaches) and the variable length of a time series.\nIn this paper, we propose a Diffusion Transformers (DiTs) based approach for tabular data series generation. Inspired by the recent success of DiTs in image and video generation, we extend this framework to deal with heterogeneous data and variable-length sequences. \nUsing extensive experiments on six datasets, we show that the proposed approach outperforms previous work by a large margin. \nOur code will be made public after this article is accepted.", "keywords": [ "tabular data generation", "time series", "diffusion models" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 2, 3, 4 ] }, { "paper_id": "bhK7U37VW8", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Huan Sun", "gender": "Female", "institution": "The Ohio State University, Columbus", "country": "US", "position": "Associate Professor" }, { "name": "Patrick McDaniel", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Full Professor" }, { "name": "Peiran Li", "gender": "Male", "institution": "NeurOS", "country": "US", "position": "Founder" }, { "name": "Somesh Jha", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin, Madison", "country": "US", "position": "Full Professor" }, { "name": "Xiaogeng Liu", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Yevgeniy Vorobeychik", "gender": "Male", "institution": "Washington University, St. Louis", "country": "US", "position": "Full Professor" }, { "name": "Zhuoqing Mao", "gender": "Female", "institution": "Google", "country": "US", "position": "consultant" }, { "name": "chaowei xiao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "G Suh", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Senior Director of Research" } ], "n_formula": 6, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 76, "n_ref_all": 90, "n_fig": 4, "n_tab": 13, "L_tab": 9649, "n_element_tab": 1042, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1490, "n_element_tab_1": 208, "formula_len_all": 93, "formula_len_all_1": 0, "len_all": 380852, "len_all_1": 77160, "len_abs": 1855, "len_title": 128, "len_sents": 119347, "len_sents_1": 36886, "n_sents": 854, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 79, "L_abs": 1176, "title": "AutoDAN-Turbo: A Lifelong Agent for Strategy Self-Exploration to Jailbreak LLMs", "abs": "Jailbreak attacks serve as essential red-teaming tools, proactively assessing whether LLMs can behave responsibly and safely in adversarial environments. Despite diverse strategies (e.g., cipher, low-resource language, persuasions, and so on) that have been proposed and shown success, these strategies are still manually designed, limiting their scope and effectiveness as a red-teaming tool. In this paper, we propose AutoDAN-Turbo, a black-box jailbreak method that can automatically discover as many jailbreak strategies as possible from scratch, without any human intervention or predefined scopes (e.g., specified candidate strategies), and use them for red-teaming. As a result, AutoDAN-Turbo can significantly outperform baseline methods, achieving a 74.3% higher average attack success rate on public benchmarks. Notably, AutoDAN-Turbo achieves an 88.5 attack success rate on GPT-4-1106-turbo. In addition, AutoDAN-Turbo is a unified framework that can incorporate existing human-designed jailbreak strategies in a plug-and-play manner. By integrating human-designed strategies, AutoDAN-Turbo can even achieve a higher attack success rate of 93.4 on GPT-4-1106-turbo.", "keywords": [ "Large Language Model", "Jailbreak Attack", "LLM Agent" ], "rating_list": [ 8, 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 4, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "bhD0EQWNut", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Chanyoung Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Junseok Lee", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Sangwoo Seo", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Sungwon Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Yoonho Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 5, "n_ref_uni": 84, "n_ref": 155, "n_ref_all": 288, "n_fig": 36, "n_tab": 22, "L_tab": 11936, "n_element_tab": 645, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 1591, "n_element_tab_1": 44, "formula_len_all": 2279, "formula_len_all_1": 290, "len_all": 396998, "len_all_1": 71324, "len_abs": 1206, "len_title": 138, "len_sents": 128868, "len_sents_1": 32935, "n_sents": 1063, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1141, "title": "Naturality-Guided Hyperedge Disentanglement for Message Passing Hypergraph Neural Network", "abs": "Hypergraph data structure has been widely used to store information or meaning derived from group interactions, meaning that each hyperedge inherently contains the context of their interactions. For example, a set of genes or a genetic pathway can be represented as a hyperedge to express the interaction of multiple genes that collaboratively perform a biological function (i.e., interaction context). However, most existing hypergraph neural networks cannot reflect the interaction context of each hyperedge due to their limited capability in capturing important or relevant factors therein. In this paper, we propose a \\textbf{simple but effective} hyperedge disentangling method, \\textbf{Natural-HNN}, that captures inherent hyperedge types or the interaction context of an hyperedge. We devised a novel guidance for hyperedge disentanglement based on the naturality condition in the category theory. In our experiments, we applied our model to hypergraphs of genetic pathways for the cancer subtype classification task, and showed that our model outperforms baselines by capturing the functional semantic similarity of genetic pathways.", "keywords": [ "Graph Neural Network", "Hypergraph", "Category Theory", "Categorcial Deep Learning", "Disentangled Representation Learning" ], "rating_list": [ 5, 3, 8 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "bgpNJBD6Va", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Binwei Yao", "gender": "Female", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Diyi Yang", "gender": "Female", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Ming Jiang", "gender": "Female", "institution": "Indiana University Indianapolis", "country": "US", "position": "Assistant Professor" }, { "name": "Shanglin Yang", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Undergrad student" }, { "name": "Yun-Shiuan Chuang", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "", "position": "PhD student" }, { "name": "Zefan Cai", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Junjie Hu", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 14, "n_ref_uni": 26, "n_ref": 41, "n_ref_all": 57, "n_fig": 5, "n_tab": 13, "L_tab": 4085, "n_element_tab": 243, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1411, "n_element_tab_1": 100, "formula_len_all": 930, "formula_len_all_1": 1267, "len_all": 141877, "len_all_1": 67668, "len_abs": 1592, "len_title": 120, "len_sents": 44236, "len_sents_1": 30598, "n_sents": 327, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1361, "title": "No Preference Left Behind: Group Distributional Preference Optimization", "abs": "Preferences within a group of people are not uniform but follow a distribution. While existing alignment methods like Direct Preference Optimization (DPO) attempt to steer models to reflect human preferences, they struggle to capture the distributional pluralistic preferences within a group. These methods often skew toward dominant preferences, overlooking the diversity of opinions, especially when conflicting preferences arise. To address this issue, we propose Group Distribution Preference Optimization (GDPO), a novel framework that aligns language models with the distribution of preferences within a group by incorporating the concept of beliefs that shape individual preferences. GDPO calibrates a language model using statistical estimation of the group's belief distribution and aligns the model with belief-conditioned preferences, offering a more inclusive alignment framework than traditional methods. In experiments using both synthetic controllable opinion generation and real-world movie review datasets, we show that DPO fails to align with the targeted belief distributions, while GDPO consistently reduces this alignment gap during training. Additionally, our evaluation metrics demonstrate that GDPO outperforms existing approaches in aligning with group distributional preferences, marking a significant advance in pluralistic alignment.", "keywords": [ "preference alignment; large language model; fairness; group preferences" ], "rating_list": [ 6, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "bgk4O69SoL", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Dongyan Zhao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Jian Guan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Rui Yan", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Wei Wu", "gender": "Male", "institution": "Ant Research", "country": "CN", "position": "Researcher" }, { "name": "Zhuocheng Gong", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Huishuai Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 50, "n_ref": 92, "n_ref_all": 100, "n_fig": 2, "n_tab": 3, "L_tab": 1901, "n_element_tab": 201, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 219, "n_element_tab_1": 16, "formula_len_all": 1480, "formula_len_all_1": 1354, "len_all": 151367, "len_all_1": 61826, "len_abs": 1480, "len_title": 130, "len_sents": 36493, "len_sents_1": 29465, "n_sents": 251, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1491, "title": "Latent Preference Coding: Aligning Large Language Models via Discrete Latent Codes", "abs": "Large language models (LLMs) have achieved remarkable success, yet aligning their generations with human preferences remains a critical challenge. Existing approaches to preference modeling often rely on an explicit or implicit reward function, overlooking the intricate and multifaceted nature of human preferences that may encompass conflicting factors across diverse tasks and populations. To address this limitation, we introduce Latent Preference Coding (LPC), a novel framework that models the implicit factors as well as their combinations behind holistic preferences using discrete latent codes. LPC seamlessly integrates with various offline alignment algorithms, automatically inferring the underlying factors and their importance from data without relying on pre-defined reward functions and hand-crafted combination weights. Extensive experiments on multiple benchmarks demonstrate that LPC consistently improves upon three alignment algorithms (DPO, SimPO, and IPO) using three base models (Mistral-7B, Llama3-8B, and Llama3-Instruct-8B). Furthermore, deeper analysis reveals that the learned latent codes effectively capture the differences in the distribution of human preferences and significantly enhance the robustness of alignment algorithms against noise in data. By providing a unified representation for the multifarious preference factors, LPC paves the way towards developing more robust and versatile alignment techniques for responsible deployment of powerful LLMs.", "keywords": [ "large language model", "alignment", "human preference", "variational inference" ], "rating_list": [ 6, 6, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "bgcdO9lmug", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Bistra Dilkina", "gender": "Female", "institution": "University of Southern California", "country": "", "position": "Associate Professor" }, { "name": "Sven Koenig", "gender": "unknown", "institution": "University of California, Irvine", "country": "US", "position": "Professor" }, { "name": "Weizhe Chen", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 69, "n_ref_all": 90, "n_fig": 10, "n_tab": 3, "L_tab": 228, "n_element_tab": 38, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 228, "n_element_tab_1": 38, "formula_len_all": 2, "formula_len_all_1": 0, "len_all": 153049, "len_all_1": 67902, "len_abs": 1576, "len_title": 129, "len_sents": 58207, "len_sents_1": 35772, "n_sents": 437, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1599, "title": "RePrompt: Prompt Engineering for Large Language Models Agents through Reflection", "abs": "In this past year, large language models (LLMs) have had remarkable success in domains outside the traditional natural language processing, and people are starting to explore the usage of LLMs in more general and close to application domains like code generation, travel planning, and robot controls. Connecting these LLMs with great capacity and external tools, people are building the so-called LLM agents, which are supposed to help people do all kinds of work in everyday life. In all these domains, the prompt to the LLMs has been shown to make a big difference in what the LLM would generate and thus affect the performance of the LLM agents. Therefore, automatic prompt engineering (APE) has become an important question for many researchers and users of LLMs. However, previous works in APE all rely on a final checker to evaluate the performance of the given prompt, which is hard to meet in the case of LLM agents where intermediate feedback is easier to get, and the final evaluation could be expensive, inaccurate, or even missing. In this paper, we propose a novel method, \\textsc{RePrompt}, which does a ``gradient descent\"-like approach to optimize the step-by-step instructions in the prompts given to LLM agents based on the chat history obtained from interactions and reflections with LLM agents. By leveraging intermediate feedback, \\textsc{RePrompt} can optimize the prompt without the need for a final solution checker. We have used experiments in PDDL generation and travel planning to show that our method could generally improve the performance for different reasoning tasks.", "keywords": [ "Large language models", "reasoning", "prompt engineering", "large language model agents" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "bfy5A3vCt7", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Conghui He", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Dora D. Liu", "gender": "Female", "institution": "Tongji University", "country": "CN", "position": "Researcher" }, { "name": "Jian Cao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Liang Hu", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Full Professor" }, { "name": "Qi Zhang", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Researcher" }, { "name": "Usman Naseem", "gender": "unknown", "institution": "Macquarie University", "country": "AU", "position": "Assistant Professor" }, { "name": "Zekun Sun", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhihao Sui", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 32, "n_ref": 53, "n_ref_all": 83, "n_fig": 11, "n_tab": 7, "L_tab": 7597, "n_element_tab": 1106, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1931, "n_element_tab_1": 287, "formula_len_all": 1303, "formula_len_all_1": 1189, "len_all": 138299, "len_all_1": 65835, "len_abs": 1338, "len_title": 83, "len_sents": 37209, "len_sents_1": 26719, "n_sents": 297, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1242, "title": "CMIRA: Class Membership Inducing Recovery Attacks Against Machine Unlearning Models", "abs": "The implementation of data privacy regulations such as GDPR and CCPA has advanced machine learning (MU) technology, which is designed to facilitate the removal of specific sensitive data points from trained models upon request. Despite rapid advancements in MU technology, its vulnerabilities are still underexplored, posing potential risks of privacy breaches by recovering unlearned sensitive information. Existing research on MU vulnerabilities often requires access to the original models, which violates with the core objective of MU. To address this gap, we initiate the first study on recovery attacks against MU models without requiring access to the original model. Our approach, known as Class Membership Inducing Recovery Attack (CMIRA), effectively recovers forgotten data by exploiting a probing dataset. Specifically, we implement the CMIRA scheme regarding mutual knowledge distillation between MU and attack models. Extensive experiments across multiple datasets and MU methods demonstrate that CMIRA exhibits high efficacy in both theoretical analysis and practical applications. Our study highlights the critical imperative for establishing robust MU systems and sets a benchmark for future research into MU vulnerabilities.", "keywords": [ "Computer Vision (CV) -> CV: Adversarial Attacks & Robustness", "Computer Vision (CV) -> CV: Ethics -- Bias", "Fairness", "Transparency & Privacy", "Machine Learning (ML) -> ML:" ], "rating_list": [ 3, 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 4 ], "contribution_list": [ 2, 2, 1, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 5 ] }, { "paper_id": "bfa58H1nQ8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Bocheng Zou", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Undergrad student" }, { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Ruichuan An", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Siyuan Huang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Weifeng Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xinyu Wei", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yulin Luo", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Gao Peng", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Shanghang Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 66, "n_ref": 158, "n_ref_all": 183, "n_fig": 8, "n_tab": 10, "L_tab": 4051, "n_element_tab": 613, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 846, "n_element_tab_1": 152, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 243475, "len_all_1": 59074, "len_abs": 441, "len_title": 138, "len_sents": 57187, "len_sents_1": 26997, "n_sents": 400, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1374, "title": "Draw-and-Understand: Leveraging Visual Prompts to Enable MLLMs to Comprehend What You Want", "abs": "In this paper, we present the Draw-and-Understand framework, exploring how to integrate visual prompting understanding capabilities into Multimodal Large Language Models (MLLMs). Visual prompts allow users to interact through multi-modal instructions, enhancing the models' interactivity and fine-grained image comprehension. In this framework, we propose a general architecture adaptable to different pre-trained MLLMs, enabling it to recognize various types of visual prompts (such as points, bounding boxes, cycles, and free-form shapes) alongside language understanding. Additionally, we introduce MDVP-Instruct-Data, a multi-domain dataset featuring 1.6 million image-visual prompt-text triplets, including natural images, document images, scene text images, mobile/web screenshots, remote sensing images, and multi-panel images. Building on this dataset, we introduce MDVP-Bench, a challenging benchmark designed to evaluate a model's ability to understand visual prompting instructions. The experimental results demonstrate that our framework can be easily and effectively applied to various MLLMs, such as SPHINX-X and LLaVA. After training with MDVP-Instruct-Data and image-level instruction datasets, our models exhibit impressive multimodal interaction capabilities and pixel-level understanding, while maintaining their image-level visual perception performance.", "keywords": [ "Multimodal Large Language Model", "Visual Prompting" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "bfZyAJ9ZAH", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Jinjun Wang", "gender": "unknown", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Professor" }, { "name": "Le Wang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Mang Cao", "gender": "unknown", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Sanping Zhou", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Wenli Huang", "gender": "Female", "institution": "Ningbo University of Technology", "country": "CN", "position": "Lecturer" }, { "name": "Ye Deng", "gender": "Male", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 17, "n_ref_uni": 39, "n_ref": 84, "n_ref_all": 107, "n_fig": 10, "n_tab": 7, "L_tab": 2072, "n_element_tab": 238, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 902, "formula_len_all_1": 932, "len_all": 118337, "len_all_1": 58533, "len_abs": 1531, "len_title": 101, "len_sents": 31169, "len_sents_1": 27112, "n_sents": 231, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1573, "title": "MSM: Multi-Scale Mamba in Multi-Task Dense Prediction", "abs": "High-quality visual representations are crucial for success in multi-task dense prediction. The Mamba architecture, initially designed for natural language processing, has garnered interest for its potential in computer vision due to its efficient modeling of long-range dependencies. However, when applied to multi-task dense prediction, it reveals inherent limitations. Unlike text processing with diverse tokenization strategies, image token partitioning requires careful consideration of multiple options. In multi-task dense prediction, each task may require specific levels of granularity in scene structure. Unfortunately, the current Mamba implementation, which segments images into fixed patch scales, fails to match these requirements, leading to sub-optimal performance. This paper proposes a simple yet effective Multi-Scale Mamba (MSM) for multi-task dense prediction. Firstly, we employ a novel Multi-Scale Scanning (MS-Scan) to establish global feature relationships at various scales. This module enhances the model's capability to deliver a comprehensive visual representation by integrating information across scales. Secondly, we adaptively merge task-shared information from multiple scales across different task branches. This design not only meets the diverse granularity demands of various tasks but also facilitates more nuanced cross-task feature interactions. Extensive experiments on two challenging benchmarks, i.e., NYUD-V2 and PASCAL-Context, show the superiority of our MSM vs its state-of-the-art competitors in multi-task dense prediction.", "keywords": [ "multi-task learning", "representation learning", "multiscale", "mamba" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "bfI8cp8qmk", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Hao-Xiang Xu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Hong Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Jia-Chen Gu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Jun-Yu Ma", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Zhen-Hua Ling", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" } ], "n_formula": 57, "n_formula_1": 14, "n_ref_uni": 43, "n_ref": 148, "n_ref_all": 184, "n_fig": 12, "n_tab": 5, "L_tab": 1770, "n_element_tab": 197, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 581, "n_element_tab_1": 59, "formula_len_all": 3771, "formula_len_all_1": 981, "len_all": 213820, "len_all_1": 70160, "len_abs": 1511, "len_title": 96, "len_sents": 48799, "len_sents_1": 32410, "n_sents": 398, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 1517, "title": "Perturbation-Restrained Sequential Model Editing", "abs": "Model editing is an emerging field that focuses on updating the knowledge embedded within large language models (LLMs) without extensive retraining. However, current model editing methods significantly compromise the general abilities of LLMs as the number of edits increases, and this trade-off poses a substantial challenge to the continual learning of LLMs. In this paper, we first theoretically analyze that the factor affecting the general abilities in sequential model editing lies in the condition number of the edited matrix. The condition number of a matrix represents its numerical sensitivity, and therefore can be used to indicate the extent to which the original knowledge associations stored in LLMs are perturbed after editing. Subsequently, statistical findings demonstrate that the value of this factor becomes larger as the number of edits increases, thereby exacerbating the deterioration of general abilities. To this end, a framework termed Perturbation Restraint on Upper bouNd for Editing (PRUNE) is proposed, which applies the condition number restraints in sequential editing. These restraints can lower the upper bound on perturbation to edited models, thus preserving the general abilities. Systematically, we conduct experiments employing three popular editing methods on three LLMs across four representative downstream tasks. Evaluation results show that PRUNE can preserve considerable general abilities while maintaining the editing performance effectively in sequential model editing.", "keywords": [ "Sequential Model Editing", "Matrix Perturbation", "General Abilities", "Large Language Models" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 2 ] }, { "paper_id": "bew8KfCw7g", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "He Zhu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Junran Wu", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Ke Xu", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "RaymondXia", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 24, "n_ref": 59, "n_ref_all": 82, "n_fig": 7, "n_tab": 7, "L_tab": 4270, "n_element_tab": 483, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 2204, "n_element_tab_1": 345, "formula_len_all": 1059, "formula_len_all_1": 1171, "len_all": 112145, "len_all_1": 65501, "len_abs": 1470, "len_title": 100, "len_sents": 34584, "len_sents_1": 26622, "n_sents": 263, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1521, "title": "CTGC: Cluster-Aware Transformer for Graph Clustering", "abs": "Graph clustering is a fundamental unsupervised task in graph mining. However, mainstream clustering methods are built on graph neural networks, thus inevitably suffer from the difficulty in long-range dependencies capturing. Moreover, current two-stage clustering scheme, consisting of representation learning and clustering, limits the ability of the graph encoder to fully exploit task-related information, resulting in suboptimal embeddings. In this work, we propose CTGC ($\\textbf{C}$luster-Aware $\\textbf{T}$ransformer for $\\textbf{G}$raph $\\textbf{C}$lustering) to mitigate these issues. Specifically, considering the excellence of transformer in long-range dependencies modeling, we first introduce transformer to graph clustering as the crucial graph encoder. To further enhance the task awareness of encoder during representation learning, we presents two mechanisms: momentum cluster-aware attention and cluster-aware regularization. In momentum cluster-aware attention, previous clustering results are adopted to guide the node embedding production with specially designed cluster-aware queries. Cluster-aware regularization is designed to fuse the cluster information into bordering nodes through minimizing the overlap between different clusters while maximizing the completeness of each cluster. We evaluate our method on seven real-world graph datasets and achieve superior results compared to existing state-of-the-art methods, demonstrating its effectiveness in improving the quality of graph clustering.", "keywords": [ "Graph Clustering", "Graph Transformer" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "bePaRx0otZ", "primary_area": "optimization", "n_author": 9, "author_info_list": [ { "name": "Defu Lian", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Han Li", "gender": "Male", "institution": "Kuaishou", "country": "", "position": "Researcher" }, { "name": "Kai Zheng", "gender": "Male", "institution": "Kuaishou", "country": "CN", "position": "Researcher" }, { "name": "Kun Gai", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Instructor" }, { "name": "Qi Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Wentian Bao", "gender": "Male", "institution": "Columbia University", "country": "", "position": "MS student" }, { "name": "Wuchao Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yang Song", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Researcher" }, { "name": "Yun En Yu", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" } ], "n_formula": 25, "n_formula_1": 19, "n_ref_uni": 23, "n_ref": 44, "n_ref_all": 54, "n_fig": 3, "n_tab": 6, "L_tab": 937, "n_element_tab": 116, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 574, "n_element_tab_1": 95, "formula_len_all": 1946, "formula_len_all_1": 1565, "len_all": 111897, "len_all_1": 64119, "len_abs": 1490, "len_title": 107, "len_sents": 35385, "len_sents_1": 28332, "n_sents": 311, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1544, "title": "Making Transformer Decoders Better Differentiable Indexers", "abs": "Retrieval aims to find the top-k items most relevant to a query/user from a large dataset. Traditional retrieval models represent queries/users and items as embedding vectors and use Approximate Nearest Neighbor (ANN) search for retrieval. Recently, researchers have proposed a generative-based retrieval method that represents items as token sequences and uses a decoder model for autoregressive training. Compared to traditional methods, this approach uses more complex models and integrates index structure during training, leading to better performance. However, these methods remain two-stage processes, where index construction is separate from the retrieval model, limiting the model's overall capacity. Additionally, existing methods construct indices by clustering pre-trained item representations in Euclidean space. However, real-world scenarios are more complex, making this approach less accurate. To address these issues, we propose a \\underline{U}nified framework for \\underline{R}etrieval and \\underline{I}ndexing, termed \\textbf{URI}. URI ensures strong consistency between index construction and the retrieval model, typically a Transformer decoder. URI simultaneously builds the index and trains the decoder, constructing the index through the decoder itself. It no longer relies on one-sided item representations in Euclidean space but constructs the index within the interactive space between queries and items. Experimental comparisons on three real-world datasets show that URI significantly outperforms existing methods.", "keywords": [ "Generative Retrieval", "Generative Index", "End-to-end Recommender System", "Information Retrieval", "Transformer" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2, 1 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3, 4 ] }, { "paper_id": "beAlX6RjsW", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Dacheng Li", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Deevashwer Rathee", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Hao Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Ion Stoica", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Raluca Popa", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 2, "n_ref_uni": 82, "n_ref": 188, "n_ref_all": 227, "n_fig": 13, "n_tab": 5, "L_tab": 1079, "n_element_tab": 239, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 490, "formula_len_all_1": 232, "len_all": 234342, "len_all_1": 69908, "len_abs": 1227, "len_title": 82, "len_sents": 65667, "len_sents_1": 32810, "n_sents": 468, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 34, "L_abs": 1252, "title": "MPC-Minimized Secure LLM Inference", "abs": "Many inference services based on large language models (LLMs) pose a privacy concern, either revealing user prompts to the service or the proprietary weights to the user. Secure inference offers a solution to this problem through secure multi-party computation (MPC), however, it is still impractical for modern LLM workload due to the large overhead imposed by MPC. To address this overhead, we propose\nMARILL, a framework that adapts LLM fine-tuning to minimize MPC usage during secure inference. MARILL introduces high-level architectural changes during fine-tuning that significantly reduce the number of expensive operations needed within MPC during inference, by removing some and relocating others outside MPC without compromising security. As a result, MARILL-generated models are more efficient across all secure inference protocols and our approach complements MPC-friendly approximations for such operations. Compared to standard fine-tuning, MARILL results in $2.2\u221211.3\\times$ better runtime and $2.4\u22126.9\\times$ better communication during secure inference across various MPC settings, while typically preserving over $90$% performance across downstream tasks. Anonymous code is available at https://anonymous.4open.science/r/MPC-auto-B100.", "keywords": [ "secure inference", "secure multi-party computation (MPC)", "transformer", "large language model (LLM)", "open-source foundational model", "fine-tuning", "LoRA", "head-merging" ], "rating_list": [ 8, 5, 6 ], "soundness_list": [ 3, 2, 4 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "be0sdRYSlH", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaolei Hao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Xun Yang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Zhifeng Hao", "gender": "Male", "institution": "Shantou University", "country": "", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 39, "n_ref": 76, "n_ref_all": 94, "n_fig": 4, "n_tab": 12, "L_tab": 4462, "n_element_tab": 646, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 3541, "n_element_tab_1": 190, "formula_len_all": 657, "formula_len_all_1": 689, "len_all": 139926, "len_all_1": 64752, "len_abs": 1196, "len_title": 148, "len_sents": 37351, "len_sents_1": 27730, "n_sents": 272, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1199, "title": "Exploring One-Shot Federated Learning by Model Inversion and Token Relabel with Vision Transformers", "abs": "One-Shot Federated Learning, where a central server learns a global model over a network of federated devices in a single round of communication, has recently emerged as a promising approach. For extremely Non-IID data, training models separately on each client results in poor performance, with low-quality generated data that are poorly matched with ground-truth labels. To overcome these issues, we propose a novel Federated Model Inversion and Token Relabel (FedMITR) framework, which trains the global model by better utilizing all patches of the synthetic images. FedMITR employs model inversion during the data generation process, selectively inverting semantic foregrounds while gradually halting the inversion process of uninformative backgrounds. Due to the presence of semantically meaningless tokens that do not positively contribute to ViT predictions, some of the generated pseudo-labels can be utilized to train the global model using patches with high information density, while patches with low information density can be relabeled using ensemble models. Extensive experimental results demonstrate that FedMITR can substantially outperform existing baselines under various settings.", "keywords": [ "One-Shot Federated Learning", "Model Inversion", "Token Relabel" ], "rating_list": [ 3, 6, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "bdHjLCcMSP", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Haibin Yan", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Full Professor" }, { "name": "Jiwen Lu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Sun ShuoFeng", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" } ], "n_formula": 19, "n_formula_1": 18, "n_ref_uni": 29, "n_ref": 50, "n_ref_all": 76, "n_fig": 6, "n_tab": 16, "L_tab": 3065, "n_element_tab": 553, "n_fig_1": 2, "n_tab_1": 9, "L_tab_1": 3725, "n_element_tab_1": 293, "formula_len_all": 700, "formula_len_all_1": 489, "len_all": 138085, "len_all_1": 70715, "len_abs": 2282, "len_title": 137, "len_sents": 39403, "len_sents_1": 29573, "n_sents": 297, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 2238, "title": "NGTTA: Non-parametric Geometry-driven Test Time Adaption for 3D Point Cloud Segmentation", "abs": "Previous Test Time Adaption (TTA) methods usually suffer from training collapse when they are transferred to complex 3D scenes for point cloud segmentation due to the significant domain gap between the source and target data.\n To solve this issue, we propose NGTTA, a stable test time adaption method guided by non-parametric geometric features. In NGTTA, we leverage the distribution of non-parametric geometric features on target data as an \u201cintermediate domain\u201d to reduce the domain gap and guide the stable learning of the source model on target data. \n Specifically, we use the source domain model and a non-parametric geometric model to extract the embedding features and geometric features of the point cloud, respectively. Then, a category-balance sampler is designed to filter easy samples and hard samples in the input data to address the class imbalance issue in semantic segmentation. Inspired by previous work, we use easy samples for entropy minimization loss and pseudo-label prediction to fine-tune the source domain model. The difference is that we refine the pseudo labels not only by considering the soft voting among their nearest neighbors in the model embedding feature space but also in the geometric space, which can prevent the accumulation of errors caused by model feature shifts. Furthermore, we believe that hard samples can effectively represent the distribution differences between the source domain and the target domain. Therefore, we propose to distill the geometric features of hard samples into the source domain model in the early stages of training to quickly converge to an \"intermediate domain\" that is similar to the target domain. By taking advantage of the ability of the non-parametric geometric feature to represent the underlying manifolds of the target data, our method efficiently reduces the difficulty of the domain adaption.\nWe conduct the main experiments on the more challenge \\textbf{\\textit{sim-to-real}} benchmark about synthetic dataset 3DFRONT and the real-world datasets ScanNet and S3DIS for 3D segmentation task. Results show that our method can efficiently improve the mIOU by over \\textbf{3\\%} on 3DFRONT$\\rightarrow$ ScanNet and \\textbf{7\\%} on 3DFRONT$\\rightarrow$ S3DIS.", "keywords": [ "Point Cloud Segmentation", "Test Time Adaption" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "bdFzyzf4Qx", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Natalia Ma\u015blany", "gender": "unknown", "institution": "Aptiv", "country": "PL", "position": "Researcher" }, { "name": "Tomasz Kania", "gender": "Male", "institution": "Jagiellonian University in Krakow", "country": "PL", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 9, "n_ref": 23, "n_ref_all": 31, "n_fig": 2, "n_tab": 4, "L_tab": 2221, "n_element_tab": 736, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2221, "n_element_tab_1": 736, "formula_len_all": 523, "formula_len_all_1": 523, "len_all": 62602, "len_all_1": 53062, "len_abs": 972, "len_title": 112, "len_sents": 19064, "len_sents_1": 19043, "n_sents": 142, "n_sents_1": 142, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 975, "title": "A Q-learning approach to the Lowest Unique Positive Integer game", "abs": "The Lowest Unique Positive Integer (LUPI) game is a multiplayer game where participants attempt to choose the smallest number that no one else selects. While previous studies model LUPI using Poisson--Nash equilibrium assumptions, our work introduces a novel Q-learning-based approach to achieve equilibrium without the need for specific distribution assumptions, such as Poisson. We demonstrate that our Q-learning model successfully emulates the Nash equilibrium while allowing flexibility in the number of players, providing a more robust and practical solution for real-world applications like real-time bidding (RTB) systems. We compare our model's performance against existing Poisson-based strategies, showcasing improved accuracy and adaptability. Furthermore, we apply our model to the Swedish Limbo lottery data and observe significant deviations from theoretical predictions, highlighting the strength of learning-based approaches in dynamic, real-world scenarios.", "keywords": [ "Q-learning", "Lowest Unique Positive Integer Game", "Nash Equilibrium", "Poisson-Nash equilibrium", "real-time bidding", "Swedish Limbo Lottery", "multi-agent reinforcement learning", "normal-form game", "reverse auction", "Poisson distribution", "game theory. TL;DR: This paper introduces a Q-learning-based approach to solve the Lowest Unique Positive Integer game", "outperforming traditional Poisson-based methods and demonstrating real-world applications such as in reverse auctions and real-time bidding systems" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 1, 3, 2 ], "presentation_list": [ 1, 1, 1 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "bcynT7s2du", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Andrew Jesson", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Postdoc" }, { "name": "David Blei", "gender": "Male", "institution": "Columbia University", "country": "", "position": "Full Professor" }, { "name": "Nicolas Beltran-Velez", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 10, "n_ref_uni": 51, "n_ref": 64, "n_ref_all": 90, "n_fig": 21, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3196, "formula_len_all_1": 804, "len_all": 153330, "len_all_1": 59405, "len_abs": 1647, "len_title": 130, "len_sents": 43878, "len_sents_1": 25447, "n_sents": 352, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1663, "title": "Can Generative AI Solve Your In-Context Learning Problem? A Martingale Perspective", "abs": "This work is about estimating when a conditional generative model (CGM) can solve an in-context learning (ICL) problem. An in-context learning (ICL) problem comprises a CGM, a dataset, and a prediction task. The CGM could be a multi-modal foundation model; the dataset, a collection of patient histories, test results, and recorded diagnoses; and the prediction task to communicate a diagnosis to a new patient. A Bayesian interpretation of ICL assumes that the CGM computes a posterior predictive distribution over an unknown Bayesian model defining a joint distribution over latent explanations and observable data. From this perspective, Bayesian model criticism is a reasonable approach to assess the suitability of a given CGM for an ICL problem. However, such approaches---like posterior predictive checks (PPCs)---often assume that we can sample from the likelihood and posterior defined by the Bayesian model, which are not explicitly given for contemporary CGMs. To address this, we show when ancestral sampling from the predictive distribution of a CGM is equivalent to sampling datasets from the posterior predictive of the assumed Bayesian model. Then we develop the generative predictive $p$-value, which enables PPCs and their cousins for contemporary CGMs. The generative predictive $p$-value can then be used in a statistical decision procedure to determine when the model is appropriate for an ICL problem. Our method only requires generating queries and responses from a CGM and evaluating its response log probability. We empirically evaluate our method on synthetic tabular, imaging, and natural language ICL tasks using large language models.", "keywords": [ "generative models", "Bayesian", "in-context learning", "generalization", "model checking" ], "rating_list": [ 6, 8, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 3, 4 ] }, { "paper_id": "bcTjW5kS4W", "primary_area": "applications to neuroscience & cognitive science", "n_author": 7, "author_info_list": [ { "name": "Lu Mi", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Trung Le", "gender": "Male", "institution": "University of Washington, Seattle", "country": "", "position": "PhD student" }, { "name": "Wuwei Zhang", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "MS student" }, { "name": "Ziyu Lu", "gender": "Female", "institution": "University of Washington", "country": "", "position": "PhD student" }, { "name": "Eric Shea-Brown", "gender": "Male", "institution": "University of Washington", "country": "", "position": "Full Professor" }, { "name": "Hao Wang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Assistant Professor" }, { "name": "Uygar Sumbul", "gender": "Male", "institution": "Allen Institute", "country": "", "position": "Associate Investigator" } ], "n_formula": 48, "n_formula_1": 15, "n_ref_uni": 60, "n_ref": 91, "n_ref_all": 146, "n_fig": 35, "n_tab": 7, "L_tab": 1657, "n_element_tab": 114, "n_fig_1": 16, "n_tab_1": 1, "L_tab_1": 856, "n_element_tab_1": 50, "formula_len_all": 2560, "formula_len_all_1": 857, "len_all": 222847, "len_all_1": 70169, "len_abs": 1837, "len_title": 151, "len_sents": 71599, "len_sents_1": 33369, "n_sents": 574, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 103, "L_abs": 1852, "title": "NetFormer: An interpretable model for recovering dynamical connectivity in neuronal population dynamics", "abs": "Neuronal dynamics are highly nonlinear and nonstationary. Traditional methods for extracting the underlying network structure from neuronal activity recordings mainly concentrate on modeling static connectivity, without accounting for key nonstationary aspects of biological neural systems, such as ongoing synaptic plasticity and neuronal modulation. To bridge this gap, we introduce the NetFormer model, an interpretable approach applicable to such systems. In NetFormer, the activity of each neuron across a series of historical time steps is defined as a token. These tokens are then linearly mapped through a query and key mechanism to generate a state- (and hence time-) dependent attention matrix that directly encodes nonstationary connectivity structures. We analyze our formulation from the perspective of nonstationary and nonlinear networked dynamical systems, and show both via an analytical expansion and targeted simulations how it can approximate the underlying ground truth. Next, we demonstrate NetFormer's ability to model a key feature of biological networks, spike-timing-dependent plasticity, whereby connection strengths continually change in response to local activity patterns. We further demonstrate that NetFormer can capture task-induced connectivity patterns on activity generated by task-trained recurrent neural networks. Thus informed, we apply NetFormer to a multi-modal dataset of real neural recordings, which contains neural activity, cell type, and behavioral state information. We show that the NetFormer effectively predicts neural dynamics and identifies cell-type specific, state-dependent dynamic connectivity that matches patterns measured in separate ground-truth physiology experiments, demonstrating its ability to help decode complex neural interactions based on population activity observations alone.", "keywords": [ "neuronal dynamics", "nonstationary", "dynamical connectivity", "interpretability", "transformer", "attention" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "bc3sUsS6ck", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Hao Cheng", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Researcher" }, { "name": "Hao Fang", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Principal Researcher" }, { "name": "Jianfeng Gao", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Luke Zettlemoyer", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Full Professor" }, { "name": "Patrick Xia", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Tong Chen", "gender": "Male", "institution": "University of Washington, Seattle", "country": "", "position": "PhD student" }, { "name": "Xiaodong Liu", "gender": "unknown", "institution": "Microsoft Research", "country": "US", "position": "Researcher" }, { "name": "Benjamin Van Durme", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 36, "n_ref": 59, "n_ref_all": 73, "n_fig": 4, "n_tab": 5, "L_tab": 2966, "n_element_tab": 149, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2092, "n_element_tab_1": 106, "formula_len_all": 1033, "formula_len_all_1": 844, "len_all": 158860, "len_all_1": 65583, "len_abs": 1849, "len_title": 141, "len_sents": 38577, "len_sents_1": 30200, "n_sents": 278, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1866, "title": "Generative Adapter: Contextualizing Language Models in Parameters with A Single Forward Pass", "abs": "Large language models (LLMs) acquire substantial knowledge during pretraining but often need adaptation to new contexts, tasks, or domains, typically achieved through fine-tuning or prompting. However, fine-tuning incurs significant training costs, while prompting increases inference overhead. Inspired by fast weight memory, we introduce GenerativeAdapter, an effective and efficient adaptation method that encode test-time context into language model parameters with a single forward pass.\nGenerativeAdapter augments a frozen pretrained LM with a lightweight adapter generator, trained via self-supervised learning, to produce parameter-efficient adapters.\nNotably, our generator is general-purpose, i.e., one generator can adapt the corresponding base model for all langauge processing scenarios.\nWe apply GenerativeAdapter to two pretrained LMs (Mistral-7B-Instruct and Llama2-7B-Chat) and evaluate the adapted models across knowledge acquisition from documents, learning from demonstrations, and personalization for users.\nIn StreamingQA, our approach is effective in injecting knowledge into the LM's parameters, achieving a 63.5\\% improvement in F1 score over the model with supervised fine-tuning (from $19.5$ to $31.5$) for contexts as long as 32K tokens.\nIn the MetaICL in-context learning evaluation, our method achieves an average accuracy of $44.9$ across 26 tasks, outperforming the base model. \nOn MSC, our method proves to be highly competitive in memorizing user information from conversations with a 4x reduction in computation and memory costs compared to \nprompting with full conversation history.\nOverall, GenerativeAdapter provides a viable solution for adapting large LMs to evolving information and providing tailored user experience, while reducing training and inference costs relative to traditional fine-tuning and prompting techniques.", "keywords": [ "language model; efficient adaptation; fine-tuning; prompting" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 3, 2, 4 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "bc2H72hGxB", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Athul Paul Jacob", "gender": "unknown", "institution": "Google", "country": "US", "position": "Student Researcher" }, { "name": "Daniel Fried", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Researcher" }, { "name": "Daphne Ippolito", "gender": "Female", "institution": "School of Engineering and Applied Science, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Vivian Lai", "gender": "Female", "institution": "VISA", "country": "", "position": "Researcher" }, { "name": "Yiming Zhang", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 1, "n_ref_uni": 32, "n_ref": 53, "n_ref_all": 95, "n_fig": 10, "n_tab": 10, "L_tab": 3677, "n_element_tab": 481, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 979, "n_element_tab_1": 43, "formula_len_all": 860, "formula_len_all_1": 100, "len_all": 144510, "len_all_1": 63508, "len_abs": 1580, "len_title": 88, "len_sents": 46594, "len_sents_1": 31009, "n_sents": 335, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1582, "title": "Human-Aligned Chess With a Bit of Search", "abs": "Chess has long been a testbed for AI's quest to match human intelligence, and in recent years, chess AI systems have surpassed the strongest humans at the game.\nHowever, these systems are *not human-aligned*; they are unable to match the skill levels of all human partners or model human-like behaviors beyond piece movement.\nIn this paper, we introduce Allie, a chess-playing AI designed to bridge the gap between artificial and human intelligence in this classic game.\nAllie is trained on log sequences of real chess games to model the behaviors of human chess players across the skill spectrum, including non-move behaviors such as pondering times and resignations\nIn offline evaluations, we find that Allie exhibits humanlike behavior: it outperforms the existing state-of-the-art in human chess move prediction and ``ponders'' at critical positions.\nThe model learns to reliably assign reward at each game state, which can be used at inference as a reward function in a novel *time-adaptive* Monte-Carlo tree search (MCTS) procedure, where the amount of search depends on how long humans would think in the same positions.\nAdaptive search enables remarkable *skill calibration*; in a large-scale online evaluation against players with ratings from 1000 to 2600 Elo, our adaptive search method leads to a skill gap of only 49 Elo on average, substantially outperforming search-free and standard MCTS baselines.\nAgainst grandmaster-level (2500 Elo) opponents, Allie with adaptive search exhibits the strength of a fellow grandmaster, all while learning *exclusively from humans*.", "keywords": [ "chess", "alignment", "adaptive MCTS", "inference-time scaling" ], "rating_list": [ 6, 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 4, 2 ], "contribution_list": [ 2, 2, 3, 4, 3 ], "confidence_list": [ 3, 4, 3, 4, 5 ] }, { "paper_id": "bbVH40jy7f", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Chao Huang", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Lianghao Xia", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Tu Ao", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Yanhua Yu", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Associate Professor" }, { "name": "ZIRUI GUO", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 12, "n_ref": 24, "n_ref_all": 39, "n_fig": 4, "n_tab": 4, "L_tab": 2514, "n_element_tab": 368, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 2933, "n_element_tab_1": 371, "formula_len_all": 161, "formula_len_all_1": 161, "len_all": 109387, "len_all_1": 79397, "len_abs": 1436, "len_title": 104, "len_sents": 43667, "len_sents_1": 36563, "n_sents": 329, "n_sents_1": 279, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1441, "title": "LightRAG: Simple and Fast Retrieval-Augmented Generation", "abs": "Retrieval-Augmented Generation (RAG) systems enhance large language models (LLMs) by integrating external knowledge sources, enabling more accurate and contextually relevant responses tailored to user needs. However, existing RAG systems have significant limitations, including reliance on flat data representations and inadequate contextual awareness, which can lead to fragmented answers that fail to capture complex inter-dependencies. To address these challenges, we propose LightRAG, which incorporates graph structures into text indexing and retrieval processes. This innovative framework employs a dual-level retrieval system that enhances comprehensive information retrieval from both low-level and high-level knowledge discovery. Additionally, the integration of graph structures with vector representations facilitates efficient retrieval of related entities and their relationships, significantly improving response times while maintaining contextual relevance. This capability is further enhanced by an incremental update algorithm that ensures the timely integration of new data, allowing the system to remain effective in dynamic environments. Extensive experimental validation demonstrates considerable improvements in retrieval accuracy and efficiency compared to existing approaches. For the reproducibility of our results, we have made our LightRAG available anonymously at: https://anonymous.4open.science/r/LightRAG-2BEE.", "keywords": [ "Retrieval-Augmented Generation", "Large Language Model", "Graph" ], "rating_list": [ 3, 5, 8, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 1, 4, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "bb2Cm6Xn6d", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Byungsoo Ko", "gender": "Male", "institution": "Facebook", "country": "CA", "position": "Researcher" }, { "name": "Han-Gyu Kim", "gender": "Male", "institution": "NAVER", "country": "KR", "position": "Researcher" }, { "name": "Ho-Jin Choi", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" }, { "name": "Yechan Hwang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Young-Jun Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 4, "n_ref_uni": 62, "n_ref": 190, "n_ref_all": 249, "n_fig": 18, "n_tab": 9, "L_tab": 4121, "n_element_tab": 266, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 905, "n_element_tab_1": 58, "formula_len_all": 589, "formula_len_all_1": 304, "len_all": 301697, "len_all_1": 72031, "len_abs": 1715, "len_title": 105, "len_sents": 80844, "len_sents_1": 34121, "n_sents": 586, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1761, "title": "Intriguing Properties of Large Language and Vision Models", "abs": "Recently, large language and vision models (LLVMs) have received significant attention and development efforts due to their remarkable generalization performance across a wide range of tasks requiring perception and cognitive abilities. A key factor behind their success is their simple architecture, which consists of a vision encoder, a projector, and a large language model (LLM). Despite their achievements in advanced reasoning tasks, their performance on fundamental perception-related tasks (\\eg MMVP) remains surprisingly low. This discrepancy raises the question of how LLVMs truly perceive images and exploit the advantages of the vision encoder. To address this, we systematically investigate this question regarding several aspects: \\textit{permutation invariance}, \\textit{robustness}, \\textit{synthetic data}, \\textit{alignment preserving} and \\textit{importance}, by evaluating the most common LLVM's families (\\ie LLaVA) across 13 evaluation benchmarks. Our extensive experiments reveal several intriguing properties of current LLVMs: (1) they internally process the image in a global manner, even whenthe order of visual patch sequences is randomly permuted; (2) they are sometimes able to solve math problems without fully perceiving detailed numerical information; (3) the cross-modal alignment is overfitted to complex reasoning tasks, thereby, causing them to lose some of the original perceptual capabilities of their vision encoder; (4) the representation space in the lower layers ($<25\\%$) plays a crucial role in determining performance and enhancing visual understanding. Lastly, based on the above observations, we suggest potential future directions for building better LLVMs and constructing more challenging evaluation benchmarks.", "keywords": [ "LLVM;clustering;projector;investigation" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "baSU1eVLwS", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Jigang Bao", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Naiqi Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Peiyuan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Shu-Tao Xia", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Tao Dai", "gender": "Male", "institution": "Department of Software Engineering, Shenzhen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yifan Hu", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Undergrad student" }, { "name": "Alan Wu", "gender": "Male", "institution": "Shenzhen University", "country": "", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 56, "n_ref_all": 68, "n_fig": 11, "n_tab": 18, "L_tab": 16465, "n_element_tab": 930, "n_fig_1": 8, "n_tab_1": 9, "L_tab_1": 5011, "n_element_tab_1": 314, "formula_len_all": 742, "formula_len_all_1": 543, "len_all": 165699, "len_all_1": 61224, "len_abs": 1430, "len_title": 122, "len_sents": 43674, "len_sents_1": 25681, "n_sents": 333, "n_sents_1": 148, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1437, "title": "TimeBridge: Non-Stationarity Matters for Long-term Time Series Forecasting", "abs": "Non-stationarity poses significant challenges for multivariate time series forecasting due to the inherent short-term fluctuations and long-term trends that can lead to spurious regressions or obscure essential long-term relationships. Most existing methods either eliminate or retain non-stationarity without adequately addressing its distinct impacts on short-term and long-term modeling. Eliminating non-stationarity is essential for avoiding spurious regressions and capturing local dependencies in short-term modeling, while preserving it is crucial for revealing long-term cointegration across variates. In this paper, we propose TimeBridge, a novel framework designed to bridge the gap between non-stationarity and dependency modeling in long-term time series forecasting. By segmenting input series into smaller patches, TimeBridge applies Integrated Attention to mitigate short-term non-stationarity and capture stable dependencies within each variate, while Cointegrated Attention preserves non-stationarity to model long-term cointegration across variates. Extensive experiments show that TimeBridge consistently achieves state-of-the-art performance in both short-term and long-term forecasting. Additionally, TimeBridge demonstrates exceptional performance in financial forecasting on the CSI 500 and S\\&P 500 indices, further validating its robustness and effectiveness. The code is available in the supplementary material.", "keywords": [ "Long-term time series forecasting", "Non-stationarity", "Dependency Modeling" ], "rating_list": [ 3, 3, 8 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 4 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "baQ0ICrnCR", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Hongyu Wang", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Ruiping Wang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Senwei Xie", "gender": "Male", "institution": "Institute of Computing Technology, CAS", "country": "", "position": "PhD student" }, { "name": "Zhanqi Xiao", "gender": "unknown", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xilin Chen", "gender": "Male", "institution": "Institute of Computing Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 67, "n_ref_all": 90, "n_fig": 9, "n_tab": 6, "L_tab": 911, "n_element_tab": 183, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 163, "n_element_tab_1": 38, "formula_len_all": 145, "formula_len_all_1": 142, "len_all": 155681, "len_all_1": 56948, "len_abs": 1328, "len_title": 132, "len_sents": 43388, "len_sents_1": 28252, "n_sents": 373, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1333, "title": "Robotic Programmer: Video Instructed Policy Code Generation for Robotic Manipulation", "abs": "Zero-shot generalization across various robots, tasks and environments remains a significant challenge in robotic manipulation. Policy code generation methods use executable code to connect high-level task descriptions and low-level action sequences, leveraging the generalization capabilities of large language models and atomic skill libraries. In this work, we propose Robotic Programmer (RoboPro), a robotic foundation model, enabling the capability of perceiving visual information and following free-form instructions to perform robotic manipulation with policy code in a zero-shot manner. To address low efficiency and high cost in collecting runtime code data for robotic tasks, we devise Video2Code to synthesize executable code from extensive videos in-the-wild with off-the-shelf vision-language model and code-domain large language model. Extensive experiments show that RoboPro achieves the state-of-the-art zero-shot performance on robotic manipulation in both simulators and real-world environments. Specifically, the zero-shot success rate of RoboPro on RLBench surpasses the state-of-the-art model GPT-4o by 11.6\\%, which is even comparable to a strong supervised training baseline. Furthermore, RoboPro is robust to different robotic configurations, and demonstrates broad visual understanding in general VQA tasks.", "keywords": [ "robotic manipulation", "vision language models", "code generation" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "baONCWMQ0r", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Hongyuan Zha", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Jiale Ma", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Qitian Wu", "gender": "Male", "institution": "Broad Institute of MIT and Harvard", "country": "US", "position": "Researcher" }, { "name": "Yang Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yebin Yang", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 27, "n_formula_1": 7, "n_ref_uni": 37, "n_ref": 97, "n_ref_all": 116, "n_fig": 6, "n_tab": 10, "L_tab": 2695, "n_element_tab": 181, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1179, "n_element_tab_1": 121, "formula_len_all": 1906, "formula_len_all_1": 814, "len_all": 158608, "len_all_1": 76918, "len_abs": 1859, "len_title": 148, "len_sents": 53070, "len_sents_1": 37376, "n_sents": 380, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 118, "L_abs": 1867, "title": "Your Consistency Model is Secretly a More Powerful Supervised Learning Paradigm for Learning Tasks with Complex Labels", "abs": "Directly predicting labels from data inputs has been a long-standing supervised learning paradigm. Its trade-off between compression and prediction is studied under the information theory framework e.g. Information Bottleneck, especially in the context of deep learning. It typically assumes that the information content of labels is significantly less than that of data inputs, leading to model designs that prioritize compressing and extracting features from data inputs. In fact, recent supervised learning increasingly faces predicting complex labels, exacerbating the challenge of learning mappings from compressed latent features to high-fidelity label representations. Predictive bottlenecks emerge not only from compression limitations but also from the inherent complexity of feature-to-label transformations. This paper proposes incorporating scheduled label information into the model during training to better learn the prediction consistency mapping, which stems from the consistency mapping concept from generative consistency models. Unlike traditional approaches predicting labels directly from inputs, in this paper, the training of our designed conditional consistency involves predicting labels using inputs and noise-perturbed label hints, pursuing the predictive consistency across different noise steps. It simultaneously learns the relationship between latent features and a spectrum of label information from zero to complete, which enables progressive learning for complex predictions and allows multi-step inference analogous to gradual denoising, thereby enhancing the prediction quality. Experiments on vision, text, and graph tasks show the superiority of our consistency supervised training paradigm, over conventional supervised training in complex label prediction problems. Source code will be made publicly available upon acceptance.", "keywords": [ "Supervised Learning", "Consistency Models" ], "rating_list": [ 5, 5, 5, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "baNW94qdsU", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Lang Huang", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Ratheesh Kalarot", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Toshihiko Yamasaki", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Zichuan Liu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 71, "n_ref": 131, "n_ref_all": 145, "n_fig": 5, "n_tab": 12, "L_tab": 3696, "n_element_tab": 357, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1176, "n_element_tab_1": 91, "formula_len_all": 726, "formula_len_all_1": 725, "len_all": 179944, "len_all_1": 60998, "len_abs": 1492, "len_title": 109, "len_sents": 43965, "len_sents_1": 28868, "n_sents": 293, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1497, "title": "Self-Training on Unpaired Data Improves Multi-Modal Alignment", "abs": "In the past few years, multimodal foundation models, e.g., CLIP, learned from a massive amount of paired multimodal data, emerged and exhibited impressive cross-modal ability in many applications. Yet collecting high-quality paired data is generally costly or even infeasible in certain cases, and the amount of paired multimodal data is several orders fewer than that of unpaired unimodal data, i.e., data without any correspondence. Our work focuses on alleviating the excessive demand for paired language-image data by leveraging the abundant unpaired data. We introduce a new approach for vision-language alignment, which we call Language-Image Self-Training (LIST). LIST consists of two key ingredients that function in a synergistic loop: i) a captioner model trained alternatively with the augmented paired data and the unpaired data with synthetic captions, both derived from the data engine, and ii) a data engine that synthesizes a diverse spectrum of captions for both paired and unpaired images with the captioner, integrating synthetic captions with the web-scraped ones to enhance the quality of paired data using off-the-shelf Large Language Models. We observe that the LIST methodology not only significantly improves the alignment between vision and language representations across multiple major benchmarks\u2014zero-shot image classification, image-text retrieval, and compositional evaluation\u2014but also demonstrates strong generalization to audio-language representation alignment.", "keywords": [ "Self-Training", "Multi-Modal Representation Alignment" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "bYsieh8LE2", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Dugki Min", "gender": "Male", "institution": "Konkuk University", "country": "", "position": "Full Professor" }, { "name": "Seongmin Park", "gender": "Male", "institution": "Konkuk University", "country": "KR", "position": "Undergrad student" }, { "name": "Taemin Jung", "gender": "Male", "institution": "Konkuk University", "country": "KR", "position": "MS student" }, { "name": "Tuan Anh Nguyen", "gender": "Male", "institution": "Konkuk University", "country": "KR", "position": "Assistant Professor" }, { "name": "WonJune Seo", "gender": "Male", "institution": "Konkuk University", "country": "KR", "position": "PhD student" }, { "name": "Park hayoung", "gender": "Female", "institution": "Konkuk University", "country": "KR", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 11, "n_ref": 28, "n_ref_all": 42, "n_fig": 7, "n_tab": 8, "L_tab": 2512, "n_element_tab": 391, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1322, "n_element_tab_1": 174, "formula_len_all": 302, "formula_len_all_1": 302, "len_all": 164116, "len_all_1": 54053, "len_abs": 1617, "len_title": 123, "len_sents": 26210, "len_sents_1": 23417, "n_sents": 199, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1624, "title": "ULoRA: Universal Low-Rank Adaptation of Diverse Deep Learning Architectures", "abs": "To train Large Language Models (LLMs) having a large number of parameters, the Parameter-Efficient Fine Tuning (PEFT) method based on LoRA, which allows fine-tuning with fewer parameters, is widely employed. However, these methods are primarily designed for application to Transformer architectures, which presents challenges when attempting to apply them to models such as Mamba. To address this limitation, this work proposes Universal LoRA (ULoRA), which applies a Low-Rank Adapter to all deep learning models at the level of universally common blocks. ULoRA achieves generalizability by applying Low-Rank Adapters to blocks, making it applicable to models that do not utilize Transformer architectures. Furthermore, by grouping multiple blocks and applying a single Low-Rank Adapter, ULoRA provides structural flexibility that allows a further reduction in the number of parameters. This significantly reduces resource usage and inference time, making it well-suited for on-device environments with limited resources, while only incurring a slight performance loss. Additionally, if all blocks are grouped to use a single Low-Rank Adapter, task switching during inference is enabled by computing only the adapter. Experimental results show that, for LLaMA-3-8B, ULoRA achieves comparable performance to LoRA with only about 60% of the parameters, while delivering up to 8% higher throughput. For Mamba-2.8B, ULoRA outperforms LoRA with only about 20% of the parameters. In scenarios with limited available resources, ULoRA can be applied using just 4% of the parameters of LoRA, with only a 10% reduction in performance.", "keywords": [ "LoRA", "Universal Adaptation", "Deep Learning", "Parameter Fine Tuning", "Transformer", "Language Models" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 3, 1 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 1 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "bWz8aOPwsJ", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Tengyuan Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Weiran Huang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhiquan Tan", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 16, "n_ref_uni": 23, "n_ref": 58, "n_ref_all": 89, "n_fig": 15, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1390, "formula_len_all_1": 1019, "len_all": 92913, "len_all_1": 52124, "len_abs": 1397, "len_title": 121, "len_sents": 29359, "len_sents_1": 22248, "n_sents": 247, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1405, "title": "Understanding Neural Tangent Kernel Dynamics Through Its Trace Evolution", "abs": "The Neural Tangent Kernel (NTK) has emerged as a valuable tool for analyzing the training and generalization properties of neural networks. While the behavior of the NTK in the infinite-width limit is well understood, a comprehensive investigation is still required to comprehend its dynamics during training in the finite-width regime. In this paper, we present a detailed exploration of the NTK's behavior through the examination of its trace during training.\n\nBy conducting experiments on standard supervised classification tasks, we observe that the NTK trace typically exhibits an increasing trend and stabilizes when the network achieves its highest accuracy on the training data. Additionally, we investigate the phenomenon of \"grokking'', which has recently garnered attention, as it involves an intriguing scenario where the test accuracy suddenly improves long after the training accuracy plateaus. To shed light on this phenomenon, we employ the NTK trace to monitor the training dynamics during grokking. Furthermore, we utilize the NTK trace to gain insights into the training dynamics of semi-supervised learning approaches, including the employment of exponential moving average mechanisms. Through these investigations, we demonstrate that the NTK, particularly its trace, remains a powerful and valuable tool for comprehending the training dynamics of modern finite-width neural networks.", "keywords": [ "Neural Tangent Kernel", "Representation Learning", "Training Dynamics" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 1, 2, 1, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "bWc6O8QSyp", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Bo Li", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Ji Shi", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Miao Zhang", "gender": "Male", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "Full Professor" }, { "name": "Xinming Zhang", "gender": "unknown", "institution": "Harbin Insitute of Technology (Shenzhen)", "country": "CN", "position": "Associate Professor" }, { "name": "Zhonghao Li", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 60, "n_formula_1": 31, "n_ref_uni": 49, "n_ref": 79, "n_ref_all": 104, "n_fig": 6, "n_tab": 4, "L_tab": 3425, "n_element_tab": 277, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2627, "n_element_tab_1": 101, "formula_len_all": 4595, "formula_len_all_1": 1648, "len_all": 163545, "len_all_1": 72622, "len_abs": 1324, "len_title": 111, "len_sents": 41572, "len_sents_1": 30226, "n_sents": 332, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1333, "title": "SFi-Former: Sparse Flow induced Attention for Graph Transformer", "abs": "Graph Transformers (GTs) have demonstrated superior performance compared to traditional message-passing graph neural networks in many studies, especially in processing graph data with long-range dependencies. However, GTs tend to suffer from weak inductive bias, overfitting and over-globalizing problems due to the dense attention. In this paper, we introduce SFi-attention, a novel attention mechanism designed to learn sparse pattern by minimizing an energy function based on network flows with $\\ell_1$-norm regularization, to relieve those issues caused by dense attention. Furthermore, SFi-Former is accordingly devised which can leverage the sparse attention pattern of SFi-attention to generate sparse network flows beyond adjacency matrix of graph data. Specifically, SFi-Former aggregates features selectively from other nodes through flexible adaptation of the sparse attention, leading to a more robust model. We validate our SFi-Former on various graph datasets, especially those graph data exhibiting long-range dependencies. Experimental results show that our SFi-Former obtains competitive performance on GNN Benchmark datasets and SOTA performance on Long-Range Graph Benchmark (LRGB) datasets. Additionally, our model gives rise to smaller generalization gaps, which indicates that it is less prone to over-fitting.", "keywords": [ "Graph Transformer", "Sparse learning", "Network flow", "Optimization" ], "rating_list": [ 3, 5, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 1, 3 ], "contribution_list": [ 2, 4, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 3, 4 ] }, { "paper_id": "bWTuOf7ZDR", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Ali Payani", "gender": "Male", "institution": "Cisco", "country": "US", "position": "Researcher" }, { "name": "Jie Hao", "gender": "Male", "institution": "George Mason University", "country": "US", "position": "PhD student" }, { "name": "Mingrui Liu", "gender": "unknown", "institution": "George Mason University", "country": "", "position": "Assistant Professor" }, { "name": "Myungjin Lee", "gender": "Male", "institution": "Cisco", "country": "US", "position": "Researcher" }, { "name": "Yuman Wu", "gender": "unknown", "institution": "George Mason University", "country": "US", "position": "PhD student" } ], "n_formula": 42, "n_formula_1": 14, "n_ref_uni": 39, "n_ref": 64, "n_ref_all": 102, "n_fig": 5, "n_tab": 16, "L_tab": 3840, "n_element_tab": 355, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1386, "n_element_tab_1": 113, "formula_len_all": 5064, "formula_len_all_1": 1043, "len_all": 230615, "len_all_1": 70164, "len_abs": 1444, "len_title": 148, "len_sents": 72867, "len_sents_1": 31219, "n_sents": 645, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1458, "title": "Personalized Federated Fine-tuning for Heterogeneous Data: a Two-Level Low Rank Adaptation Approach", "abs": "We study the personalized federated fine-tuning task with heterogeneous client data in the context of foundation models, where clients collaboratively fine-tune a foundation model (e.g., BERT, GPT) without sharing their local data, achieving personalized models simultaneously. While recent efforts have applied parameter-efficient fine-tuning techniques like low-rank adaptation (LoRA) or training prompts in federated settings, they often overlook data heterogeneity and model personalization. The primary challenge is that a single common adapter or prompt learner may not suffice for the diverse data of all clients. To address this issue, we propose PF2LoRA, a new personalized federated fine-tuning algorithm based on a novel \\emph{ two-level low rank adaptation framework} on top of LoRA. Given the pretrained foundation model whose weight is frozen, our algorithm aims to learn two levels of adaptation simultaneously: the first level aims to learn a common adapter for all clients, while the second level fosters individual client personalization. This framework explicitly accommodates variations in adapter matrix ranks across clients and introduces minimal additional memory overhead, as the second-level adaptation comprises a small number of parameters compared to the first level. Our experiments on natural language understanding and generation tasks demonstrate that PF2LoRA significantly outperforms existing federated fine-tuning methods.", "keywords": [ "Federated Learning", "Low Rank Adaptation", "Heterogenoeus Data", "Language Model", "Foundation Model" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "bWT6OBJ71x", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Eric J Michaud", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Max Tegmark", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Xiaoman Delores Ding", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Zifan Carl Guo", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Ziming Liu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 60, "n_ref": 97, "n_ref_all": 127, "n_fig": 14, "n_tab": 1, "L_tab": 123, "n_element_tab": 17, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 301, "formula_len_all_1": 245, "len_all": 148574, "len_all_1": 51350, "len_abs": 1457, "len_title": 124, "len_sents": 41506, "len_sents_1": 25631, "n_sents": 266, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1469, "title": "Interpretable Patterns in Random Initialization Unveil Final Representation", "abs": "The field of mechanistic interpretability has made strides in unraveling models' hidden representations but is often puzzled by why specific representations form. This paper addresses a crucial question on this front: when a neural network can learn multiple distinct representations to solve a task, how does it \"choose\" among them during training?\nWe suggest that, at initialization, instead of starting from an empty scratchpad, the model's embedding already contains partially formed representations of varying ''completeness.'' Models tend to develop a representation that is more \"complete\" at initialization, disregarding less complete alternatives.\nWe empirically examine this hypothesis on algorithmic toy models with clearly defined final representations from which we can elicit an interpretable signal to evaluate such \"completeness\" of possible representations in the initial embedding. We find that the representations with high initial signals are chosen by the model with high probability, a pattern consistent across models with a single learned representation (remainder equivalence, multi-digit XOR) and with multiple, redundant representations (modular addition). \nFinally, we investigate the role of embedding dimensionality on model's representation and their ``completeness.'' \nOur results with toy models show that the seemingly chaotic initialization contains many interpretable patterns to understand the training dynamics of representations.", "keywords": [ "mechanistic interpretability", "training dynamics", "science of deep learning", "representation learning" ], "rating_list": [ 5, 3, 3, 5, 6 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 3, 4 ] }, { "paper_id": "bW9fGYo44s", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Chirag Sehgal", "gender": "Male", "institution": "Delhi Technological University (Delhi College of Engineering)", "country": "IN", "position": "Undergrad student" }, { "name": "Jishu Sen Gupta", "gender": "Male", "institution": "Indian Institute of Technology (Banaras Hindu University) Varanasi", "country": "IN", "position": "Undergrad student" }, { "name": "Onkar Kishor Susladkar", "gender": "Male", "institution": "Indian Institute of Technology, Roorkee", "country": "IN", "position": "Researcher" }, { "name": "Rekha Singhal", "gender": "Female", "institution": "Tata Consultancy Services Limited, India", "country": "", "position": "Principal Researcher" }, { "name": "Sparsh Mittal", "gender": "Male", "institution": "Indian Institution Technology Roorkee", "country": "IN", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 9, "n_ref_uni": 41, "n_ref": 61, "n_ref_all": 88, "n_fig": 8, "n_tab": 11, "L_tab": 2611, "n_element_tab": 232, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 684, "n_element_tab_1": 55, "formula_len_all": 537, "formula_len_all_1": 639, "len_all": 211646, "len_all_1": 63461, "len_abs": 380, "len_title": 89, "len_sents": 58833, "len_sents_1": 26387, "n_sents": 451, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 89, "L_abs": 1514, "title": "MotionAura: Generating High-Quality and Motion Consistent Videos using Discrete Diffusion", "abs": "The spatio-temporal complexity of video data presents significant challenges in tasks such as compression, generation, and inpainting. We present four key contributions to address the challenges of spatiotemporal video processing. First, we introduce the 3D Mobile Inverted Vector-Quantization Variational Autoencoder (3D-MBQ-VAE), which combines Variational Autoencoders (VAEs) with masked modeling to enhance spatiotemporal video compression. The model achieves superior temporal consistency and state-of-the-art (SOTA) reconstruction quality by employing a novel training strategy with full frame masking. Second, we present MotionAura, a text-to-video generation framework that utilizes vector-quantized diffusion models to discretize the latent space and capture complex motion dynamics, producing temporally coherent videos aligned with text prompts. Third, we propose a spectral transformer-based denoising network that processes video data in the frequency domain using the Fourier Transform. This method effectively captures global context and long-range dependencies for high-quality video generation and denoising. Lastly, we introduce a downstream task of Sketch Guided Video Inpainting. This task leverages Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning. Our models achieve SOTA performance on a range of benchmarks. Our work offers robust frameworks for spatiotemporal modeling and user-driven video content manipulation. We will release the code, dataset, and models in open-source.", "keywords": [ "text2video", "VQ-Diffusion", "video Inpainting", "Large scale pretraining" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 2, 4 ], "presentation_list": [ 3, 2, 4 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "bVdcAZAW2h", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Lei Huang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Associate Professor" }, { "name": "Yihao Yue", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Yizhou Ruan", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "Undergrad student" }, { "name": "Yunhao Ni", "gender": "unknown", "institution": "Beihang University", "country": "CN", "position": "MS student" }, { "name": "Yuxin Guo", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" } ], "n_formula": 69, "n_formula_1": 19, "n_ref_uni": 32, "n_ref": 46, "n_ref_all": 55, "n_fig": 3, "n_tab": 9, "L_tab": 2149, "n_element_tab": 281, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 239, "n_element_tab_1": 10, "formula_len_all": 4734, "formula_len_all_1": 788, "len_all": 152449, "len_all_1": 58978, "len_abs": 1499, "len_title": 122, "len_sents": 44555, "len_sents_1": 26957, "n_sents": 435, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1508, "title": "Enjoy Your Layer Normalization with the Computation Efficiency of RMSNorm", "abs": "Layer normalization (LN) is a milestone technique in deep learning and has been widely used in various network architectures. It performs centering and scaling over the layer activations of a neural network for each example, stabilizing and accelerating the training of neural network. However, it introduces extra computation cost during inference and the computation problem has recently been addressed by its counterpart RMSNorm that only adopts scaling. This paper investigates how to exploit the theoretical advantages of LN but with the cost of RMSNorm. This paper formally defines the condition that the centering operation of LN can be removed and this condition can be obtained by imposing the column centering constraint on the adjacent linear module before the LN. We propose column centered weight transformation (CCWT) to ensure an LN without centering operation (i.e., RMSNorm) have the same output as the original one in a pre-trained model.\n Our method can be directly applied to various pre-trained large language models (LLMs) and large vision language models (VLMs) with LN, enabling an immediate reduction in computation cost meanwhile maintaining equivalent prediction during inference.\n We further propose a reparameterization method, called column based weight centering (CBWC), to ensure the linear module column centered during training. We show that RMSNorm combining CBWC can obtain an equivalent effects to the LN counterpart during training, but with more efficient computation.", "keywords": [ "Layer normalization", "RMSNorm", "Deep Learning" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "bVTM2QKYuA", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Kiho Park", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Victor Veitch", "gender": "unknown", "institution": "University of Chicago", "country": "", "position": "Assistant Professor" }, { "name": "Yibo Jiang", "gender": "Male", "institution": "University of Chicago", "country": "", "position": "PhD student" }, { "name": "Yo Joong Choe", "gender": "Male", "institution": "University of Chicago", "country": "", "position": "Postdoc" } ], "n_formula": 29, "n_formula_1": 13, "n_ref_uni": 42, "n_ref": 65, "n_ref_all": 122, "n_fig": 21, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2783, "formula_len_all_1": 1098, "len_all": 154380, "len_all_1": 63941, "len_abs": 1056, "len_title": 126, "len_sents": 50265, "len_sents_1": 30125, "n_sents": 423, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 78, "L_abs": 1064, "title": "The Geometry of Categorical and Hierarchical Concepts in Large Language Models", "abs": "The linear representation hypothesis is the informal idea that semantic concepts are encoded as linear directions in the representation spaces of large language models (LLMs). Previous work has shown how to make this notion precise for representing binary concepts that have natural contrasts (e.g., {male, female}) as _directions_ in representation space. However, many natural concepts do not have natural contrasts (e.g., whether the output is about an animal). In this work, we show how to extend the formalization of the linear representation hypothesis to represent features (e.g., is_animal) as _vectors_. This allows us to immediately formalize the representation of categorical concepts as polytopes in the representation space. Further, we use the formalization to prove a relationship between the hierarchical structure of concepts and the geometry of their representations. We validate these theoretical results on the Gemma and LLaMA-3 large language models, estimating representations for 900+ hierarchically related concepts using data from WordNet.", "keywords": [ "categorical concepts", "hierarchical concepts", "linear representation hypothesis", "causal inner product", "interpretability" ], "rating_list": [ 5, 8, 6, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "bVBLqKoiJ1", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Noam Rotstein", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "PhD student" }, { "name": "Ron Kimmel", "gender": "Male", "institution": "Computer Science Departmen, Technion-Israel Institute of Technology", "country": "IL", "position": "Full Professor" }, { "name": "Roy Ganz", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "PhD student" }, { "name": "navve wasserman", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "IL", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 53, "n_ref": 119, "n_ref_all": 171, "n_fig": 21, "n_tab": 8, "L_tab": 2725, "n_element_tab": 253, "n_fig_1": 10, "n_tab_1": 6, "L_tab_1": 1736, "n_element_tab_1": 134, "formula_len_all": 166, "formula_len_all_1": 166, "len_all": 196872, "len_all_1": 72288, "len_abs": 1608, "len_title": 118, "len_sents": 52081, "len_sents_1": 33089, "n_sents": 410, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1470, "title": "Paint by Inpaint: Learning to Add Image Objects by Removing Them First", "abs": "Image editing has advanced significantly with the introduction of text-conditioned diffusion models. Despite this progress, seamlessly adding objects to images based on textual instructions without requiring user-provided input masks remains a challenge. We address this by leveraging the insight that removing objects (Inpaint) is significantly simpler than its inverse process of adding them (Paint), attributed to the utilization of segmentation mask datasets alongside inpainting models that inpaint within these masks. Capitalizing on this realization, by implementing an automated and extensive pipeline, we curate a filtered large-scale image dataset containing pairs of images and their corresponding object-removed versions. Using these pairs, we train a diffusion model to inverse the inpainting process, effectively adding objects into images. Unlike other editing datasets, ours features natural target images instead of synthetic ones; moreover, it maintains consistency between source and target by construction. Additionally, we utilize a large Vision-Language Model to provide detailed descriptions of the removed objects and a Large Language Model to convert these descriptions into diverse, natural-language instructions. Our quantitative and qualitative results show that the trained model surpasses existing models in both object addition and general editing tasks. To propel future research, we will release the dataset alongside the trained models.", "keywords": [ "Image Editing Dataset", "Generative Models", "Diffusion Models" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "bU1JOvdXXK", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Bei Yan", "gender": "Female", "institution": ", Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Jie Zhang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Mengqi Lei", "gender": "Male", "institution": "China University of Geosciences Wuhan", "country": "CN", "position": "Undergrad student" }, { "name": "Shiguang Shan", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zheng Yuan", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zhongqi Wang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Xilin Chen", "gender": "Male", "institution": "Institute of Computing Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 48, "n_ref": 153, "n_ref_all": 166, "n_fig": 6, "n_tab": 5, "L_tab": 1205, "n_element_tab": 80, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 722, "n_element_tab_1": 44, "formula_len_all": 461, "formula_len_all_1": 461, "len_all": 167728, "len_all_1": 63858, "len_abs": 1343, "len_title": 130, "len_sents": 34017, "len_sents_1": 31454, "n_sents": 244, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1363, "title": "Dysca: A Dynamic and Scalable Benchmark for Evaluating Perception Ability of LVLMs", "abs": "Currently many benchmarks have been proposed to evaluate the perception ability of the Large Vision-Language Models (LVLMs).\nHowever, most benchmarks conduct questions by selecting images from existing datasets, resulting in the potential data leakage. \nBesides, these benchmarks merely focus on evaluating LVLMs on the realistic style images and clean scenarios, leaving the multi-stylized images and noisy scenarios unexplored. \nIn response to these challenges, we propose a dynamic and scalable benchmark named Dysca for evaluating LVLMs by leveraging synthesis images. \nSpecifically, we leverage Stable Diffusion and design a rule-based method to dynamically generate novel images, questions and the corresponding answers. \nWe consider 51 kinds of image styles and evaluate the perception capability in 20 subtasks.\nMoreover, we conduct evaluations under 4 scenarios (i.e., Clean, Corruption, Print Attacking and Adversarial Attacking) and 3 question types (i.e., Multi-choices, True-or-false and Free-form). Thanks to the generative paradigm, Dysca serves as a scalable benchmark for easily adding new subtasks and scenarios.\nA total of 24 advanced open-source LVLMs and 2 close-source LVLMs are evaluated on Dysca, revealing the drawbacks of current LVLMs. \nThe benchmark is released in anonymous github page \\url{https://github.com/Benchmark-Dysca/Dysca}.", "keywords": [ "Benchmark", "Large Vision language Model", "Perception Ability" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "bU0JMHJ8zL", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Adel Bibi", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Senior Researcher" }, { "name": "Alasdair Paren", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Ken Chatfield", "gender": "Male", "institution": "Tractable", "country": "", "position": "VP Research Science" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 21, "n_ref": 39, "n_ref_all": 44, "n_fig": 0, "n_tab": 3, "L_tab": 2067, "n_element_tab": 100, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 2243, "n_element_tab_1": 127, "formula_len_all": 922, "formula_len_all_1": 922, "len_all": 98520, "len_all_1": 68809, "len_abs": 887, "len_title": 87, "len_sents": 33131, "len_sents_1": 32774, "n_sents": 218, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 39, "L_abs": 884, "title": "Questioning Simplicity Bias Assumptions", "abs": "The Simplicity Bias (SB) is the observation that the training of most commonly used neural network architectures with standard training techniques is biased toward learning simple functions. This phenomenon can be a benefit or drawback depending on the relative complexity of the desired function to be learnt. If the desired function is relatively simple it's a positive. However, if there are simpler features that are highly predictive; commonly named shortcuts or spurious features, that are not present in the test environment, the SB can result in poor generalisation performance. Most existing works on mitigating the SB make various assumptions, either about the features present in the train and test domains or by assuming access to information about the test domain at train time. In this paper we review recent work on the SB and take a critical look at these assumptions.", "keywords": [ "Simplicity Bias", "Shortcut Learning" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "bTi6usR2hF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Qixiang Ye", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yaowei Wang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Yue Liu", "gender": "Not Specified", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yunjie Tian", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Postdoc" }, { "name": "Zhaozhi Wang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yunfan Liu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 63, "n_ref": 98, "n_ref_all": 117, "n_fig": 7, "n_tab": 10, "L_tab": 3736, "n_element_tab": 376, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2933, "n_element_tab_1": 272, "formula_len_all": 516, "formula_len_all_1": 517, "len_all": 158170, "len_all_1": 60939, "len_abs": 970, "len_title": 91, "len_sents": 33548, "len_sents_1": 25564, "n_sents": 255, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 972, "title": "Building Vision Models upon Heat Conduction", "abs": "Visual representation models leveraging attention mechanisms are challenged by significant computational overhead, particularly when pursuing large receptive fields. In this study, we aim to mitigate this challenge by introducing the Heat Conduction Operator (HCO) built upon the physical heat conduction principle. HCO conceptualizes image patches as heat sources and models their correlations through adaptive thermal energy diffusion, enabling robust visual representations. HCO enjoys a computational complexity of O(N^1.5), as it can be implemented using discrete cosine transformation (DCT) operations. HCO is plug-and-play, combining with deep learning backbones produces visual representation models (termed vHeat) with global receptive fields. Experiments across vision tasks demonstrate that, beyond the stronger performance, vHeat achieves up to a 3x throughput, 80% less GPU memory allocation and 35% fewer computational FLOPs compared to the Swin-Transformer.", "keywords": [ "Vision Models", "Representation Learning", "Heat Conduction" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "bT2iAIYFAg", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Juan Xu", "gender": "Female", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "Associate Professor" }, { "name": "Piji Li", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "", "position": "Professor" }, { "name": "Xuan Deng", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 74, "n_ref_all": 92, "n_fig": 3, "n_tab": 5, "L_tab": 2779, "n_element_tab": 161, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1934, "n_element_tab_1": 94, "formula_len_all": 627, "formula_len_all_1": 579, "len_all": 121270, "len_all_1": 63137, "len_abs": 1568, "len_title": 131, "len_sents": 33433, "len_sents_1": 28814, "n_sents": 249, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1577, "title": "Vector Segmented and Recombined Adaptation for Scalable and Efficient Model Tuning", "abs": "Among the most commonly utilized parameter-efficient fine-tuning (PEFT) methods, LoRA and its variations have achieved significant popularity. The Vector-based Random Matrix Adaptation (VeRA), one typical variant, utilizes random weights and projections to reduce the number of trainable parameters greatly. However, it requires additional GPU memory and computational resources, probably resulting in a lack of scalability that leads to performance bottlenecks in complex tasks. Besides, the inappropriate initialization of random matrices may affect model performance. To address these problems, we propose a new method called Vector Segmented and Recombined Adaptation (SeRA). SeRA segments input vectors into sub-vectors for individual dimensionality reduction, then introduces a square matrix to combine the information from the reduced sub-vectors, and finally expands the dimensionality independently to adapt the size of pre-trained model. SeRA allows for flexible increase of trainable parameters to enhance performance in complex tasks, and avoids the problem caused by random matrices initialization. Through evaluations on the image classification, cross-modal image-text retrieval, instruction-tuning and GLUE benchmark, we demonstrate the scalability and efficiency of SeRA. Furthermore, we utilize Singular Value Decomposition on the adaptation matrices of SeRA, to analyze how the information characteristics of the matrices change in different ranks and tasks. \nThe results can serve as the guide for selecting appropriate parameter amounts in different tasks.", "keywords": [ "Parameter-efficient fine-tuning", "Adaptation", "Vector segmentation", "Scalable" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "bSzygH9bYg", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Hao Huang", "gender": "Male", "institution": "New York University", "country": "AE", "position": "Postdoc" }, { "name": "Mengxi Wu", "gender": "Non-Binary", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Mohammad Rostami", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Yi Fang", "gender": "Male", "institution": "New York University", "country": "US", "position": "Associate Professor" } ], "n_formula": 30, "n_formula_1": 16, "n_ref_uni": 38, "n_ref": 61, "n_ref_all": 84, "n_fig": 4, "n_tab": 6, "L_tab": 2876, "n_element_tab": 339, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1028, "n_element_tab_1": 28, "formula_len_all": 2322, "formula_len_all_1": 1081, "len_all": 170621, "len_all_1": 75788, "len_abs": 1316, "len_title": 123, "len_sents": 52776, "len_sents_1": 33052, "n_sents": 413, "n_sents_1": 269, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1321, "title": "Curvature Diversity-Driven Deformation and Domain Alignment for Point Cloud", "abs": "Unsupervised Domain Adaptation (UDA) is crucial for reducing the need for extensive manual data annotation when training deep networks on point cloud data. A significant challenge of UDA lies in effectively bridging the domain gap. To tackle this challenge, we propose Curvature Diversity-Driven Nuclear-Norm Wasserstein Domain Alignment (CDND). Our approach first introduces a Curvature Diversity-driven Deformation Reconstruction (CurvRec) task, which effectively mitigates the gap between the source and target domains by enabling the model to extract salient features from semantically rich regions of a given point cloud. We then propose Deformation-based Nuclear-norm Wasserstein Discrepancy (D-NWD), which applies the Nuclear-norm Wasserstein Discrepancy to both deformed and original data samples to align the source and target domains. Furthermore, we contribute a theoretical justification for the effectiveness of D-NWD in distribution alignment and demonstrate that it is generic enough to be applied to any deformations. To validate our method, we conduct extensive experiments on two public domain adaptation datasets for point cloud classification and segmentation tasks. Empirical experiment results show that our CDND achieves state-of-the-art performance by a noticeable margin over existing approaches.", "keywords": [ "Unsupervised Domain Adaptation", "Point Cloud" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 4, 1 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "bSq0XGS3kW", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Andrii Zadaianchuk", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Postdoc" }, { "name": "Aniket Rajiv Didolkar", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "Anirudh Goyal", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Georg Martius", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "Full Professor" }, { "name": "Maximilian Seitzer", "gender": "unknown", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Michael C Mozer", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 0, "n_ref_uni": 56, "n_ref": 161, "n_ref_all": 195, "n_fig": 10, "n_tab": 7, "L_tab": 4491, "n_element_tab": 481, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1250, "n_element_tab_1": 42, "formula_len_all": 382, "formula_len_all_1": 0, "len_all": 225354, "len_all_1": 67067, "len_abs": 1368, "len_title": 106, "len_sents": 64859, "len_sents_1": 32382, "n_sents": 518, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1376, "title": "On the Transfer of Object-Centric Representation Learning", "abs": "The goal of object-centric representation learning is to decompose visual scenes into a structured representation that isolates the entities into individual vectors. Recent successes have shown that object-centric representation learning can be scaled to real-world scenes by utilizing features from pre-trained foundation models like DINO. However, so far, these object-centric methods have mostly been applied in-distribution, with models trained and evaluated on the same dataset. This is in contrast to the underlying foundation models, which have been shown to be applicable to a wide range of data and tasks. Thus, in this work, we answer the question of whether current real-world capable object-centric methods exhibit similar levels of transferability by introducing a benchmark comprising seven different synthetic and real-world datasets. We analyze the factors influencing performance under transfer and find that training on diverse real-world images improves generalization to unseen scenarios. Furthermore, inspired by the success of task-specific fine-tuning in foundation models, we introduce a novel fine-tuning strategy to adapt pre-trained vision encoders for the task of object discovery. We find that the proposed approach results in state-of-the-art performance for unsupervised object discovery, exhibiting strong zero-shot transfer to unseen datasets.", "keywords": [ "representation learning", "object-centric learning", "object-centric representation learning", "unsupervised learning", "transfer", "zero-shot", "generalization" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 3, 3, 1, 4 ], "contribution_list": [ 2, 3, 1, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "bSSFERgkFn", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Kishore Paul", "gender": "Male", "institution": "eBay Inc.", "country": "US", "position": "Researcher" }, { "name": "Mandy Barrett Korpusik", "gender": "Female", "institution": "Loyola Marymount University", "country": "US", "position": "Assistant Professor" }, { "name": "Mustafa Zaki", "gender": "Male", "institution": "eBay Inc.", "country": "US", "position": "Researcher" }, { "name": "Vaishampayan Pathuri", "gender": "Male", "institution": "eBay Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 11, "n_ref": 18, "n_ref_all": 30, "n_fig": 4, "n_tab": 3, "L_tab": 990, "n_element_tab": 125, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 990, "n_element_tab_1": 125, "formula_len_all": 165, "formula_len_all_1": 165, "len_all": 72180, "len_all_1": 53030, "len_abs": 652, "len_title": 106, "len_sents": 25060, "len_sents_1": 25052, "n_sents": 165, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 652, "title": "Counterfactual Techniques for Enhancing Customer Retention", "abs": "In this paper, we introduce a novel counterfactual reasoning method using eBERT embeddings to convert customers from an e-commerce company who frequently add items to their cart but don\u2019t proceed to checkout. We demonstrate that our method i) outperforms existing techniques such as DiCE, GANs, and CFRL in key metrics such as coverage, while also maintaining a low latency; ii) balances high coverage and low latency by adjusting the number of nearest unlike neighbors, highlighting a trade-off between these competing goals; and iii) allows customization of mutable features, improving the practical applicability of our counterfactual explanations.", "keywords": [ "Counterfactual Explanations", "BERT", "e-commerce" ], "rating_list": [ 8, 3, 3, 1 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 3, 1, 2, 1 ], "contribution_list": [ 3, 2, 1, 1 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "bS76qaGbel", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Bin Cui", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Chenlin Meng", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Jingwei Liu", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Ling Yang", "gender": "Male", "institution": "DeepSeek AI", "country": "CN", "position": "Intern" }, { "name": "Minkai Xu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Stefano Ermon", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Xingchao Liu", "gender": "Male", "institution": "DeepSeek AI", "country": "CN", "position": "Researcher" }, { "name": "Zhilong Zhang", "gender": "unknown", "institution": "Institute for AI Industry Research (AIR), Tsinghua University", "country": "CN", "position": "Intern" }, { "name": "Zixiang Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Intern" }, { "name": "Zhang wen tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 29, "n_formula_1": 15, "n_ref_uni": 41, "n_ref": 113, "n_ref_all": 138, "n_fig": 3, "n_tab": 6, "L_tab": 1897, "n_element_tab": 196, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 2266, "n_element_tab_1": 226, "formula_len_all": 3125, "formula_len_all_1": 1494, "len_all": 156131, "len_all_1": 52646, "len_abs": 1078, "len_title": 124, "len_sents": 32774, "len_sents_1": 21541, "n_sents": 212, "n_sents_1": 125, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1080, "title": "Consistency Flow Matching: Defining Straight Flows with Velocity Consistency", "abs": "Flow matching (FM) is a general framework for defining probability paths via Ordinary Differential Equations (ODEs) to transform between noise and data samples. Recent approaches attempt to straighten these flow trajectories to generate high-quality samples with fewer function evaluations, typically through iterative rectification methods or optimal transport solutions. In this paper, we introduce Consistency Flow Matching (Consistency-FM), a novel FM method that explicitly enforces self-consistency in the velocity field. Consistency-FM directly defines straight flows starting from different times to the same endpoint, imposing constraints on their velocity values. Additionally, we propose a multi-segment training approach for Consistency-FM to enhance expressiveness, achieving a better trade-off between sampling quality and speed. Extensive experiments demonstrate that our Consistency-FM significantly improves training efficiency by converging 4.4x faster than consistency models and 1.7x faster than rectified flow models while achieving better generation quality.", "keywords": [ "Flow Matching", "Generative Models" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "bRqaHn3J5I", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Binghao Liu", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Fang Wan", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Fei Gu", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Han Yang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 20, "n_ref_uni": 4, "n_ref": 14, "n_ref_all": 36, "n_fig": 4, "n_tab": 4, "L_tab": 872, "n_element_tab": 93, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 977, "n_element_tab_1": 72, "formula_len_all": 1329, "formula_len_all_1": 1139, "len_all": 104190, "len_all_1": 63443, "len_abs": 1562, "len_title": 119, "len_sents": 30977, "len_sents_1": 28138, "n_sents": 252, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1495, "title": "Prototype antithesis for biological few-shot class-incremental learning", "abs": "Deep learning has become essential in the biological species recognition task. However, a significant challenge is the ability to continuously learn new or mutated species with limited annotated samples. Since species within the same family typically share similar traits, distinguishing between new and existing (old) species during incremental learning often faces the issue of species confusion. This can result in \"catastrophic forgetting\" of old species and poor learning of new ones. To address this issue, we propose a Prototype Antithesis (PA) method, which leverages the hierarchical structures in biological taxa to reduce confusion between new and old species. PA operates in two steps: Residual Prototype Learning (RPL) and Residual Prototype Mixing (RPM). RPL enables the model to learn unique prototypes for each species alongside residual prototypes representing shared traits within families. RPM generates synthetic samples by blending features of new species with residual prototypes of old species, encouraging the model to focus on species-unique traits and minimize species confusion. By integrating RPL and RPM, the proposed PA method mitigates \"catastrophic forgetting\" while improving generalization to new species. Extensive experiments on CUB200, PlantVillage, and Tree-of-Life datasets demonstrate that PA significantly reduces inter-species confusion and achieves state-of-the-art performance, highlighting its potential for deep learning in biological data analysis.", "keywords": [ "Biological Recognition; Few-Shot Learning; Class-Incremental Learning; Prototype Antithesis" ], "rating_list": [ 6, 8, 6, 3 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "bRfVj0Sh88", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Dongmin Park", "gender": "Male", "institution": "KRAFTON", "country": "KR", "position": "Researcher" }, { "name": "Jae-Gil Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Jaehyun Park", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "MS student" }, { "name": "Seola Choi", "gender": "Female", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "MS student" } ], "n_formula": 22, "n_formula_1": 10, "n_ref_uni": 37, "n_ref": 53, "n_ref_all": 69, "n_fig": 4, "n_tab": 3, "L_tab": 724, "n_element_tab": 57, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1422, "n_element_tab_1": 42, "formula_len_all": 1980, "formula_len_all_1": 1324, "len_all": 121695, "len_all_1": 61810, "len_abs": 1218, "len_title": 111, "len_sents": 38800, "len_sents_1": 27013, "n_sents": 293, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1223, "title": "Active Preference Optimization via Maximizing Learning Capacity", "abs": "The success of deep learning in various complex tasks relies heavily on large amounts of annotated data, which can be prohibitively expensive to acquire. Techniques such as reinforcement learning with human feedback (RLHF) and direct preference optimization (DPO) have emerged as methods for fine-tuning models by leveraging human preferences, but they come with significant costs, especially when applied to large-scale language models (LLMs). Recent efforts to reduce these costs have focused on active preference optimization, which uses certainty-based selection to minimize the annotation burden. However, the two-step process of selecting uncertain input prompts and then acquiring completions can lead to sub-optimal pairings, potentially limiting model learning capacity. This paper suggests that divAPO eliminates suboptimal pairings that are typical of two-step methods and enhances learning capacity by selecting the most informative preference pairs in a single phase, taking into account both data distribution probabilities and preference model certainty. Through experiments on complicated Language tasks, we demonstrate that our method achieves significant performance improvements over existing approaches.", "keywords": [ "active learning", "preference learning", "preference optimization" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 1, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "bRa4JLPzii", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Yunju Cho", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Jay Yoon Lee", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 27, "n_ref": 82, "n_ref_all": 117, "n_fig": 6, "n_tab": 9, "L_tab": 10432, "n_element_tab": 1803, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 3400, "n_element_tab_1": 555, "formula_len_all": 146, "formula_len_all_1": 146, "len_all": 177203, "len_all_1": 65644, "len_abs": 1191, "len_title": 145, "len_sents": 50924, "len_sents_1": 27476, "n_sents": 368, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1195, "title": "CoMRes: Semi-Supervised Time Series Forecasting Utilizing Consensus Promotion of Multi-Resolution", "abs": "Long-term time series forecasting poses significant challenges due to the complex dynamics and temporal variations, particularly when dealing with unseen patterns and data scarcity. Traditional supervised learning approaches, which rely on cleaned and labeled data, struggle to capture these unseen characteristics, limiting their effectiveness in real-world applications. In this study, we propose a semi-supervised approach that leverages multi-view setting on augmented data without requiring explicit future values as labels to address these limitations. By introducing a consensus promotion framework, our method enhances agreement among multiple single-view models on unseen augmented data. This approach not only improves forecasting accuracy but also mitigates error accumulation in long-horizon predictions. Furthermore, we explore the impact of autoregressive and non-autoregressive decoding schemes on error propagation, demonstrating the robustness of our model in extending prediction horizons. Experimental results show that our proposed method not only surpasses traditional supervised models in accuracy but also exhibits greater robustness when extending the prediction horizon.", "keywords": [ "Time series forecasting", "Multi-scale", "Semi-supervised learning" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "bRMfqThoVC", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Bolin Ding", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Senior Director" }, { "name": "Haokun Chen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "MS student" }, { "name": "Jinyang Gao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xiang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yuan Gao", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zhicai Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "xue wang", "gender": "Male", "institution": "Alibaba Group US", "country": "US", "position": "Researcher" } ], "n_formula": 22, "n_formula_1": 19, "n_ref_uni": 37, "n_ref": 70, "n_ref_all": 106, "n_fig": 19, "n_tab": 10, "L_tab": 2479, "n_element_tab": 194, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 2260, "n_element_tab_1": 162, "formula_len_all": 1269, "formula_len_all_1": 937, "len_all": 160193, "len_all_1": 69004, "len_abs": 1002, "len_title": 139, "len_sents": 49107, "len_sents_1": 29348, "n_sents": 397, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1342, "title": "Leveraging Diffusion Transformers for Stock Factor Augmentation in Financial Markets", "abs": "Data scarcity poses a significant challenge in training machine learning models for stock forecasting, often leading to low signal-to-noise ratio (SNR) and data homogeneity that degrade model performance. To address these issues, we introduce DiffsFormer, a novel approach utilizing artificial intelligence-generated samples (AIGS) with a Transformer-based Diffusion Model. Initially trained on a large-scale source domain with conditional guidance to capture global joint distribution, DiffsFormer augments training by editing existing samples for specific downstream tasks, allowing control over the deviation of generated data from the target domain. We evaluate DiffsFormer on the CSI300 and CSI800 datasets using eight commonly used machine learning models, achieving relative improvements of 7.3\\% and 22.1\\% in annualized return ratio, respectively. Extensive experiments provide insights into DiffsFormer's functionality and its components, illustrating their role in mitigating data scarcity and enhancing model performance. Our findings demonstrate the potential of AIGS and DiffsFormer in addressing data limitations in stock forecasting, with the ability to generate realistic stock factors and control the editing process. These results validate our approach and contribute to a deeper understanding of its underlying mechanisms.", "keywords": [ "Financial Data Augmentation", "Diffusion Models", "Transformer" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 1, 1, 3 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "bR1J7SpzrD", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Bryan Catanzaro", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Vice President" }, { "name": "Dinesh Manocha", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Professor" }, { "name": "Rafael Valle", "gender": "Not Specified", "institution": "NVIDIA", "country": "", "position": "Senior Research Scientist" }, { "name": "Sonal Kumar", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Sreyan Ghosh", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Zhifeng Kong", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" } ], "n_formula": 18, "n_formula_1": 10, "n_ref_uni": 59, "n_ref": 81, "n_ref_all": 108, "n_fig": 6, "n_tab": 15, "L_tab": 6397, "n_element_tab": 883, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 505, "n_element_tab_1": 73, "formula_len_all": 1270, "formula_len_all_1": 823, "len_all": 196090, "len_all_1": 70229, "len_abs": 1572, "len_title": 130, "len_sents": 49641, "len_sents_1": 32935, "n_sents": 387, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1578, "title": "Synthio: Augmenting Small-Scale Audio Classification Datasets with Synthetic Data", "abs": "We present Synthio, a novel approach for augmenting small-scale audio classification datasets with synthetic data. Our goal is to improve audio classification accuracy with limited labeled data. Traditional data augmentation techniques, which apply artificial transformations (e.g., adding random noise or masking segments), struggle to create data that captures the true diversity present in real-world audios. To address this shortcoming, we propose to augment the dataset with synthetic audio generated from text-to-audio (T2A) diffusion models. However, synthesizing effective augmentations is challenging because not only should the generated data be acoustically consistent with the underlying small-scale dataset, but they should also have sufficient compositional diversity. To overcome the first challenge, we align the generations of the T2A model with the small-scale dataset using preference optimization. This ensures that the acoustic characteristics of the generated data remain consistent with the small-scale dataset. To address the second challenge, we propose a novel caption generation technique that leverages the reasoning capabilities of Large Language Models to (1) generate diverse and meaningful audio captions and (2) iteratively refine their quality. The generated captions are then used to prompt the aligned T2A model. We extensively evaluate Synthio on ten datasets and four simulated limited-data settings. Results indicate our method consistently outperforms all baselines by 0.1%-39% using a T2A model trained only on weakly-captioned AudioSet.", "keywords": [ "audio classification", "synthetic data", "data-efficient learning" ], "rating_list": [ 6, 5, 8, 8 ], "soundness_list": [ 3, 1, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "bQ0sbMLYFj", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Dit-Yan Yeung", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Chair Professor" }, { "name": "Jiaming Zhang", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Researcher" }, { "name": "Jitao Sang", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Junhong Ye", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xingjun Ma", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Ge Yi Li", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Postdoc" }, { "name": "Yang Yunfan", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 38, "n_ref": 62, "n_ref_all": 80, "n_fig": 7, "n_tab": 4, "L_tab": 3120, "n_element_tab": 289, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1435, "n_element_tab_1": 186, "formula_len_all": 281, "formula_len_all_1": 453, "len_all": 132479, "len_all_1": 64148, "len_abs": 1705, "len_title": 166, "len_sents": 36415, "len_sents_1": 29303, "n_sents": 274, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 96, "L_abs": 1726, "title": "AnyAttack: Self-supervised Generation of Targeted Adversarial Attacks for Vision-Language Models", "abs": "Due to their multimodal capabilities, Vision-Language Models (VLMs) have found numerous impactful applications in real-world scenarios. However, recent studies have revealed that VLMs are vulnerable to image-based adversarial attacks, particularly targeted adversarial images that manipulate the model to generate harmful content specified by the adversary. \nCurrent attack methods rely on predefined target labels to create targeted adversarial attacks, which limits their scalability and applicability for large-scale robustness evaluations. In this paper, we propose **AnyAttack**, a self-supervised framework that generates targeted adversarial images for VLMs without label supervision, allowing **any** image to serve as a target for the **attack**.\nTo address the limitation of existing methods that require label supervision, we introduce a contrastive loss that trains a generator on a large-scale unlabeled image dataset, LAION-400M\ndataset, for generating targeted adversarial noise. \nThis large-scale pre-training endows our method with powerful transferability across a wide range of VLMs.\nExtensive experiments on five mainstream open-source VLMs (CLIP, BLIP, BLIP2, InstructBLIP, and MiniGPT-4) across three multimodal tasks (image-text retrieval, multimodal classification, and image captioning) demonstrate the effectiveness of our attack.\nAdditionally, we successfully transfer AnyAttack to multiple commercial VLMs, including Google's Gemini, Claude's Sonnet, and Microsoft's Copilot.\nThese results reveal an unprecedented risk to VLMs, highlighting the need for effective countermeasures.\nUpon publication, we will release the pre-trained generator to support further research in addressing this challenge.", "keywords": [ "targeted adversarial attack", "Vision-Language Models", "self-supervised" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "bPO4jLOTGG", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Dong Yan", "gender": "Male", "institution": "Baichuan Intelligent Technology", "country": "", "position": "Researcher" }, { "name": "Jian Xie", "gender": "Male", "institution": "Beijing Baichuan Intelligent Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Junge Zhang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "Full Professor" }, { "name": "wei shen", "gender": "unknown", "institution": "Baichuan", "country": "", "position": "Researcher" }, { "name": "Xingzhou Lou", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yuzi Yan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 55, "n_ref": 112, "n_ref_all": 124, "n_fig": 6, "n_tab": 3, "L_tab": 525, "n_element_tab": 92, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1510, "n_element_tab_1": 91, "formula_len_all": 718, "formula_len_all_1": 707, "len_all": 154560, "len_all_1": 53977, "len_abs": 1037, "len_title": 126, "len_sents": 32813, "len_sents_1": 23734, "n_sents": 237, "n_sents_1": 168, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1041, "title": "Uncertainty-aware Reward Model: Teaching Reward Models to Know What is Unknown", "abs": "Reward models (RM) play a critical role in aligning generations of large language models (LLM) to human expectations. However, prevailing RMs fail to capture the stochasticity within human preferences and cannot effectively evaluate the reliability of reward predictions. To address these issues, we propose Uncertain-aware RM (URM) and Uncertain-aware RM Ensemble (URME) to incorporate and manage uncertainty in reward modeling. URM can model the distribution of disentangled attributes within human preferences, while URME quantifies uncertainty through discrepancies in the ensemble, thereby identifying potential lack of knowledge during reward evaluation. Experiment results indicate that the proposed URM achieves state-of-the-art performance compared to models with the same size, demonstrating the effectiveness of modeling uncertainty within human preferences. Furthermore, empirical results show that through uncertainty quantification, URM and URME can identify unreliable predictions to improve the quality of reward evaluations.", "keywords": [ "large language models", "reward models", "ensemble", "uncertainty-aware" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "bOpHCZNRPQ", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Bo Chen", "gender": "Male", "institution": "Middle Tennessee State University", "country": "US", "position": "Undergrad student" }, { "name": "Yingyu Liang", "gender": "Male", "institution": "University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Zhenmei Shi", "gender": "Male", "institution": "Voyage AI", "country": "US", "position": "Researcher" }, { "name": "Zhizhou Sha", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 51, "n_formula_1": 16, "n_ref_uni": 44, "n_ref": 65, "n_ref_all": 74, "n_fig": 1, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4352, "formula_len_all_1": 1723, "len_all": 192164, "len_all_1": 72003, "len_abs": 1495, "len_title": 90, "len_sents": 52498, "len_sents_1": 30806, "n_sents": 524, "n_sents_1": 262, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 1574, "title": "HSR-Enhanced Sparse Attention Acceleration", "abs": "Large Language Models (LLMs) have demonstrated remarkable capabilities across various applications, but their performance on long-context tasks is often limited by the computational complexity of attention mechanisms. \nThis paper introduces a novel approach to accelerate attention computation in LLMs, particularly for long-context scenarios. \nWe leverage the inherent sparsity within attention mechanisms, both in conventional Softmax attention and ReLU attention (with $\\mathsf{ReLU}^\\alpha$ activation, $\\alpha \\in \\mathbb{N}_+$), to significantly reduce the running time complexity.\nOur method employs a Half-Space Reporting (HSR) data structure to rapidly identify non-zero or ``massively activated'' entries in the attention matrix. We present theoretical analyses for two key scenarios: attention generation and full attention computation with long input context. \nOur approach achieves a running time of $O(mn^{4/5})$ significantly faster than the naive approach $O(mn)$ for attention generation, where $n$ is the context length, $m$ is the query length, and $d$ is the hidden dimension.\nWe can also reduce the running time of full attention computation from $O(mn)$ to $O(mn^{1 - 1 / \\lfloor d/2\\rfloor} + mn^{4/5})$.\nImportantly, our method introduces no error for ReLU attention and only provably negligible error for Softmax attention, where the latter is supported by our empirical validation.\nThis work represents a significant step towards enabling efficient long-context processing in LLMs, potentially broadening their applicability across various domains.", "keywords": [ "Half-Space Reporting", "Large Language Models", "Attention Acceleration" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "bOoHGBwFoo", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Jungong Han", "gender": "Male", "institution": "The University of Sheffield", "country": "GB", "position": "Full Professor" }, { "name": "Mushui Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Xi Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yunlong Yu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ziqian Lu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 14, "n_ref": 27, "n_ref_all": 48, "n_fig": 11, "n_tab": 6, "L_tab": 2768, "n_element_tab": 153, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 2151, "n_element_tab_1": 412, "formula_len_all": 986, "formula_len_all_1": 1008, "len_all": 125760, "len_all_1": 61986, "len_abs": 1622, "len_title": 123, "len_sents": 34176, "len_sents_1": 26532, "n_sents": 234, "n_sents_1": 170, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1627, "title": "Hierarchical divide-and-conquer grouping for Zero shot learning", "abs": "Generalized Zero-Shot Learning (GZSL) faces a key challenge in transferring knowledge from base classes to classify samples from both base and novel classes. This transfer learning paradigm inherently risks a prediction bias, wherein test samples are disproportionately classified towards the base classes due to the models' familiarity and overfitting to those classes during training. To tackle the prediction bias issue, we introduce a divide-and-conquer strategy that segregates the united label space into distinct base and novel subspaces. Within each subspace, we train a customized model to ensure specialized learning tailored to the distinct characteristics of the respective classes. To compensate for the absence of novel classes, we propose utilizing off-the-shelf diffusion-based generative models, conditioned on class-level descriptions crafted by Large Language Models (LLMs), to synthesize diverse visual samples representing the novel classes. To further relieve the class confusion in each subspace, we propose to further divide each subspace into two smaller subspaces, where the classes in each smaller subspace are obtained with the unsupervised cluster strategy in the text embedding space. With our hierarchical divide-and-conquer approach, the test samples are first divided into a smaller subspace and then predicted the class labels with the specialized model trained with the classes present within the subspace. Comprehensive evaluations across three GZSL benchmarks underscore the effectiveness of our method, demonstrating its ability to perform competitively and outperform existing approaches.", "keywords": [ "Generalized Zero-shot learning", "Visual Language Models", "Hierarchical Divide-and-Conquer" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "bOjmeZkmxI", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Haimin ZHANG", "gender": "unknown", "institution": "University of Technology Sydney", "country": "AU", "position": "Postdoc" }, { "name": "Min Xu", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 45, "n_ref": 73, "n_ref_all": 91, "n_fig": 5, "n_tab": 8, "L_tab": 3331, "n_element_tab": 385, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2927, "n_element_tab_1": 343, "formula_len_all": 692, "formula_len_all_1": 552, "len_all": 125585, "len_all_1": 57016, "len_abs": 1465, "len_title": 147, "len_sents": 32311, "len_sents_1": 23715, "n_sents": 288, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1454, "title": "Genetic-evolutionary Graph Nerual Networks: A Paradigm for Improved Graph Representation Learning", "abs": "Message-passing graph neural networks have become the dominant framework for learning over graphs. However, empirical studies continually show that message-passing graph neural networks tend to generate over-smoothed representations for nodes after iteratively applying message passing. This over-smoothing problem is a core issue that limits the representational capacity of message-passing graph neural networks. We argue that the fundamental problem with over-smoothing is a lack of diversity in the generated embeddings, and the problem could be reduced by preserving the embedding diversity in their generation process. To this end, we propose genetic-evolutionary graph neural networks, a new paradigm for graph representation learning inspired by genetic algorithms. We model each layer of a graph neural network as an evolutionary process and develop operations based on crossover and mutation to prevent embeddings from becoming similar to one another, thus enabling the model to generate improved graph representations. The proposed framework is interpretable, as it directly draws inspiration from genetic algorithms for preserving population diversity. We experimentally validate the proposed framework on six benchmark datasets on different tasks. The results show that our method significant advances the performance current graph neural networks, resulting in new state-of-the-art results for graph representation learning on the datasets.", "keywords": [ "graph neural networks", "graph representation learning", "genetic evolution" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "bOhr1iVee5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "De Ma", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Gang Pan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" }, { "name": "Huajin Tang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Weixing Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Xudong Jiang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Qian Dr. Zheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "ZONGRUI LI", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 16, "n_ref_uni": 27, "n_ref": 103, "n_ref_all": 136, "n_fig": 12, "n_tab": 3, "L_tab": 1433, "n_element_tab": 336, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 605, "n_element_tab_1": 133, "formula_len_all": 618, "formula_len_all_1": 622, "len_all": 113222, "len_all_1": 60817, "len_abs": 1523, "len_title": 161, "len_sents": 31261, "len_sents_1": 27012, "n_sents": 238, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 113, "L_abs": 1154, "title": "Spiking GS: Towards High-Accuracy and Low-Cost Surface Reconstruction via Spiking Neuron-based Gaussian Splatting", "abs": "3D Gaussian Splatting is capable of reconstructing 3D scenes in minutes. Despite recent advances in improving surface reconstruction accuracy, the reconstructed results still exhibit bias and suffer from inefficiency in storage and training. This paper provides a different observation on the cause of the inefficiency and the reconstruction bias, which is attributed to the integration of the low-opacity parts (LOPs) of the generated Gaussians. We show that LOPs consist of Gaussians with overall low-opacity (LOGs) and the low-opacity tails (LOTs) of Gaussians. We propose Spiking GS to reduce such two types of LOPs by integrating spiking neurons into the Gaussian Splatting pipeline. Specifically, we introduce global and local full-precision integrate-and-fire spiking neurons to the opacity and representation function of flattened 3D Gaussians, respectively. Furthermore, we enhance the density control strategy with spiking neurons' thresholds and an new criterion on the scale of Gaussians. Our method can represent more accurate reconstructed surfaces at a lower cost. The code is available at https://anonymous.4open.science/r/SpikingGS-D721.", "keywords": [ "3D Surface Reconstruction", "Gaussian Splatting", "Spiking Neuron" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 2, 5, 4, 3 ] }, { "paper_id": "bO31lfEdos", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Chang D. Yoo", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Eunseop Yoon", "gender": "Female", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Hee Suk Yoon", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Thanh Xuan Nguyen", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Tung Minh Luu", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 31, "n_ref": 46, "n_ref_all": 59, "n_fig": 4, "n_tab": 4, "L_tab": 7118, "n_element_tab": 245, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2506, "n_element_tab_1": 144, "formula_len_all": 820, "formula_len_all_1": 676, "len_all": 202429, "len_all_1": 62590, "len_abs": 1683, "len_title": 151, "len_sents": 47708, "len_sents_1": 27281, "n_sents": 334, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1698, "title": "Mitigating Object Hallucination in Large Vision Language Model with Human-Free Reinforcement Learning", "abs": "Large Vision-Language Models (LVLMs) have excelled in joint visual and language understanding, particularly in generating detailed image captions. However, they still struggle with object hallucination, where non-existent objects are described, especially in long captions. While fine-tuning through supervised learning with enhanced datasets or reinforcement learning from human feedback can alleviate this issue, these methods demand considerable human effort, limiting scalability. This paper addresses this challenge by introducing a human-free framework to mitigate object hallucination in LVLMs for image captioning, utilizing reinforcement learning driven exclusively by automatic natural language processing metrics. We demonstrate that the following framework can effectively mitigate hallucination: (1) caption generation is formulated as a Markov Decision Process (MDP); (2) minimizing hallucination while maintaining caption quality is guided by a reward function, combining a proposed \\textit{F1Score} with a penalty on Kullback\u2013Leibler divergence from the pre-trained model; (3) fine-tuning the LVLM within the MDP framework can be performed directly by Proximal Policy Optimization (PPO) with careful attention to architectural details. Extensive experiments demonstrate a significant reduction in hallucination by up to 41\\% while preserving the caption quality compared to the baseline model, InstructBLIP, on the COCO dataset. This improvement is reflected in consistent gains in object coverage and accuracy across various models and datasets. Notably, our method achieves comparable or superior performance to alternative approaches, all without requiring any human involvement.", "keywords": [ "Reinforcement Learning", "Vision Large Language Model", "Large Language Model", "Hallucination" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "bNVbOS3lrl", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Jie Lu", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "Professor" }, { "name": "Junyu Xuan", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Assistant Professor" }, { "name": "Mengjing Wu", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" } ], "n_formula": 57, "n_formula_1": 21, "n_ref_uni": 51, "n_ref": 91, "n_ref_all": 132, "n_fig": 10, "n_tab": 5, "L_tab": 1501, "n_element_tab": 174, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1099, "n_element_tab_1": 124, "formula_len_all": 5363, "formula_len_all_1": 1609, "len_all": 167567, "len_all_1": 66261, "len_abs": 1421, "len_title": 143, "len_sents": 53410, "len_sents_1": 30822, "n_sents": 378, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1422, "title": "Bridging the Gap between Variational Inference and Stochastic Gradient MCMC in Function Space", "abs": "Traditional parameter-space posterior inference for Bayesian neural networks faces several challenges, such as the difficulty in specifying meaningful prior, the potential pathologies in deep models and the intractability for multi-modal posterior. To address these issues, functional variational inference (fVI) and functional Markov Chain Monte Carlo (fMCMC) are two recently emerged Bayesian inference schemes that perform posterior inference directly in function space by incorporating more informative functional priors. Similar to their parameter-space counterparts, fVI and fMCMC have their own strengths and weaknesses. For instance, fVI is computationally efficient but imposes strong distributional assumptions, while fMCMC is asymptotically exact but suffers from slow mixing in high dimensions. To inherit the complementary benefits of both schemes, this work proposes a novel hybrid inference method for functional posterior inference. Specifically, it combines fVI and fMCMC successively by an elaborate linking mechanism to form an alternating approximation process. We also provide theoretical justification for the soundness of such a hybrid inference through the lens of Wasserstein gradient flows in the function space. We evaluate our method on several benchmark tasks and observe improvements in both predictive accuracy and uncertainty quantification compared to parameter/function-space VI and MCMC.", "keywords": [ "Bayesian deep learning", "functional variational inference", "functional MCMC" ], "rating_list": [ 6, 6, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 3, 4, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 4, 3 ] }, { "paper_id": "bMvqccRmKD", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Biwei Huang", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Fan Feng", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Postdoc" }, { "name": "Lei Xu", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "CN", "position": "Emeritus" }, { "name": "Shikui Tu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xinyue Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Yupei Yang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 64, "n_formula_1": 17, "n_ref_uni": 59, "n_ref": 109, "n_ref_all": 199, "n_fig": 24, "n_tab": 10, "L_tab": 1629, "n_element_tab": 88, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 554, "n_element_tab_1": 36, "formula_len_all": 4056, "formula_len_all_1": 1593, "len_all": 237317, "len_all_1": 77118, "len_abs": 1484, "len_title": 143, "len_sents": 68234, "len_sents_1": 32113, "n_sents": 555, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1490, "title": "Towards Generalizable Reinforcement Learning via Causality-Guided Self-Adaptive Representations", "abs": "General intelligence requires quick adaptation across tasks. While existing reinforcement learning (RL) methods have made progress in generalization, they typically assume only distribution changes between source and target domains. In this paper, we explore a wider range of scenarios where not only the distribution but also the environment spaces may change. For example, in the CoinRun environment, we train agents from easy levels and generalize them to difficulty levels where there could be new enemies that have never occurred before. To address this challenging setting, we introduce a causality-guided self-adaptive representation-based approach, called CSR, that equips the agent to generalize effectively across tasks with evolving dynamics. Specifically, we employ causal representation learning to characterize the latent causal variables within the RL system. Such compact causal representations uncover the structural relationships among variables, enabling the agent to autonomously determine whether changes in the environment stem from distribution shifts or variations in space, and to precisely locate these changes. We then devise a three-step strategy to fine-tune the causal model under different scenarios accordingly. Empirical experiments show that CSR efficiently adapts to the target domains with only a few samples and outperforms state-of-the-art baselines on a wide range of scenarios, including our simulated environments, CartPole, CoinRun and Atari games.", "keywords": [ "Reinforcement Learning", "Transfer Learning" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 4, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "bMC1t7eLRc", "primary_area": "foundation or frontier models, including LLMs", "n_author": 13, "author_info_list": [ { "name": "Chengliang Chai", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "Associate Professor" }, { "name": "Chi Zhang", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Conghui He", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Guoren Wang", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Huaping Zhong", "gender": "Male", "institution": "SenseTime", "country": "CN", "position": "Researcher" }, { "name": "Ju Fan", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Kuan Zhang", "gender": "unknown", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Lei Cao", "gender": "Male", "institution": "University of Arizona", "country": "US", "position": "Assistant Professor" }, { "name": "Qiu Jiantao", "gender": "Male", "institution": "shanghai AI lab", "country": "CN", "position": "Researcher" }, { "name": "Rui Wang", "gender": "unknown", "institution": "nanjing university", "country": "", "position": "MS student" }, { "name": "Tianyi Bai", "gender": "unknown", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Xinlin Zhuang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Ye Yuan", "gender": "unknown", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 62, "n_ref": 122, "n_ref_all": 145, "n_fig": 12, "n_tab": 18, "L_tab": 4457, "n_element_tab": 535, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 440, "n_element_tab_1": 86, "formula_len_all": 810, "formula_len_all_1": 771, "len_all": 190937, "len_all_1": 55035, "len_abs": 1973, "len_title": 136, "len_sents": 49359, "len_sents_1": 25053, "n_sents": 373, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 86, "L_abs": 1952, "title": "Harnessing Diversity for Important Data Selection in Pretraining Large Language Models", "abs": "Data selection is of great significance in pretraining large language models, given the variation in quality within the large-scale available training corpora. \nTo achieve this, researchers are currently investigating the use of data influence to measure the importance of data instances, $i.e.,$ a high influence score indicates that incorporating this instance to the training set is likely to enhance the model performance. Consequently, they select the top-$k$ instances with the highest scores. However, this approach has several limitations. \n(1) Calculating the accurate influence of all available data is time-consuming.\n(2) The selected data instances are not diverse enough, which may hinder the pretrained model's ability to generalize effectively to various downstream tasks.\nIn this paper, we introduce $\\texttt{Quad}$, a data selection approach that considers both quality and diversity by using data influence to achieve state-of-the-art pretraining results.\nTo compute the influence ($i.e.,$ the quality) more accurately and efficiently, we incorporate the attention layers to capture more semantic details, which can be accelerated through the Kronecker product. \nFor the diversity, $\\texttt{Quad}$ clusters the dataset into similar data instances within each cluster and diverse instances across different clusters. For each cluster, if we opt to select data from it, we take some samples to evaluate the influence to prevent processing all instances. Overall, we favor clusters with highly influential instances (ensuring high quality) or clusters that have been selected less frequently (ensuring diversity), thereby well balancing between quality and diversity. Experiments on Slimpajama and FineWeb over 7B large language models demonstrate that $\\texttt{Quad}$ significantly outperforms other data selection methods with a low FLOPs consumption. Further analysis also validates the effectiveness of our influence calculation.", "keywords": [ "LLMs", "data selection", "influence function", "diversity" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "bKswCSYkKq", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Haixia Pan", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "Associate Professor" }, { "name": "Jiahua Lan", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "MS student" }, { "name": "Ruijun Liu", "gender": "Male", "institution": "Beijing Technology and Business University", "country": "", "position": "Full Professor" }, { "name": "Sen Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "AU", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 6, "n_ref_uni": 43, "n_ref": 69, "n_ref_all": 84, "n_fig": 9, "n_tab": 5, "L_tab": 1358, "n_element_tab": 106, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 732, "n_element_tab_1": 23, "formula_len_all": 875, "formula_len_all_1": 284, "len_all": 164059, "len_all_1": 56909, "len_abs": 1521, "len_title": 132, "len_sents": 56368, "len_sents_1": 29047, "n_sents": 395, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1524, "title": "Neuron-level Balance between Stability and Plasticity in Deep Reinforcement Learning", "abs": "In contrast to the inherent ability of humans to continuously acquire new knowledge, modern deep reinforcement learning (DRL) agents generally encounter a significant challenge: the stability-plasticity dilemma, which refers to the trade-off between retaining existing skills (stability) and learning new knowledge (plasticity). In this study, we propose Neuron-level Balance between Stability and Plasticity (NBSP) to tackle this challenge, by taking inspiration from the observation that both stability and plasticity are integrally linked to the expressive capabilities of networks, which are primarily determined by the behavior of individual neurons. To the best of our knowledge, this is the first work that addresses both stability and plasticity loss simultaneously in DRL at the level of neurons. Specifically, NBSP first (1) defines and identifies RL skill neurons that are crucial for knowledge retention through a goal-oriented method, and then (2) introduces a stability-plasticity balancing mechanism by employing gradient masking and experience replay techniques targeting these neurons to preserve the encoded memory related to existing skills while enhancing the learning capabilities of other neurons. Experimental results on the Meta-World and Atari benchmarks demonstrate that NBSP significantly outperforms existing approaches in balancing stability and plasticity. Furthermore, our findings underscore the pivotal role of the critic within this context, providing valuable insights for future research.", "keywords": [ "reinforcement learning", "stability-plasticity dilemma", "skill neuron" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 1, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "bKQJzuBSRJ", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Dongyoon Han", "gender": "Male", "institution": "NAVER", "country": "KR", "position": "Senior Research Scientist" }, { "name": "Hyoseo Kim", "gender": "Female", "institution": "Sogang University", "country": "KR", "position": "MS student" }, { "name": "Junsuk Choe", "gender": "Male", "institution": "Sogang University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 39, "n_ref": 104, "n_ref_all": 140, "n_fig": 4, "n_tab": 14, "L_tab": 6403, "n_element_tab": 641, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 3199, "n_element_tab_1": 329, "formula_len_all": 248, "formula_len_all_1": 142, "len_all": 164393, "len_all_1": 74447, "len_abs": 1196, "len_title": 114, "len_sents": 52300, "len_sents_1": 33820, "n_sents": 375, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1200, "title": "NegMerge: Consensual Weight Negation for Strong Machine Unlearning", "abs": "Machine unlearning aims to selectively remove specific knowledge from a model. Current methods, such as task arithmetic, rely on fine-tuning models on the forget set, generating a task vector, and subtracting it from the original model. However, we argue the effectiveness of this approach is highly sensitive to hyperparameter selection, necessitating careful validation to identify the best model among many fine-tuned candidates. In this paper, we propose a novel method that leverages all given fine-tuned models rather than selecting a single one. By constructing task vectors from models trained with varied hyperparameters and merging only the components of the task vectors with consistent signs, we perform unlearning by negating the merged task vector from the original model. Given that existing methods also utilize multiple fine-tuned models, our approach delivers more effective unlearning without incurring additional computational costs. We demonstrate the effectiveness of our method on both vision-language models and standard image classification models, showing improved unlearning performance with minimal degradation on the retain set, outperforming state-of-the-art techniques.", "keywords": [ "Machine Unlearning", "Image Classification", "Model Merging" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 4, 2, 4, 3 ], "presentation_list": [ 4, 2, 4, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "bKCc3USOyv", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Junyang Chen", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yang Lu", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yiqun Zhang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Researcher" }, { "name": "Yiu-ming Cheung", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Full Professor" }, { "name": "Mengke LI", "gender": "Female", "institution": "Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ)", "country": "CN", "position": "Associate Professor" } ], "n_formula": 56, "n_formula_1": 18, "n_ref_uni": 31, "n_ref": 40, "n_ref_all": 74, "n_fig": 5, "n_tab": 10, "L_tab": 5416, "n_element_tab": 364, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 781, "n_element_tab_1": 116, "formula_len_all": 2422, "formula_len_all_1": 707, "len_all": 167147, "len_all_1": 61721, "len_abs": 1613, "len_title": 126, "len_sents": 50480, "len_sents_1": 29641, "n_sents": 371, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1624, "title": "Attributed Graph Clustering via Generalized Quaternion Representation Learning", "abs": "Clustering complex data in the form of attributed graphs has attracted increasing attention, where appropriate graph representation is a critical prerequisite for accurate cluster analysis. However, the Graph Convolutional Network will homogenize the representation of graph nodes due to the well-known over-smoothing effect. This limits the network architecture to a shallow one, losing the ability to capture the critical global distribution information for clustering. Therefore, we propose a generalized graph auto-encoder network, which introduces quaternion operations to the encoders to achieve efficient structured feature representation learning without incurring deeper network and larger-scale parameters. The generalization of our method lies in the following two aspects: 1) connecting the quaternion operation naturally suitable for four feature components with graph data of arbitrary attribute dimensions, and 2) introducing a generalized graph clustering objective as a loss term to obtain clustering-friendly representations without requiring a pre-specified number of clusters $k$. It turns out that the representations of nodes learned by the proposed Graph Clustering based on Generalized Quaternion representation learning (GCGQ) are more discriminative, containing global distribution information, and are more general, suiting downstream clustering under different $k$s. Extensive experiments including significance tests, ablation studies, and qualitative results, illustrate the superiority of GCGQ. The source code is temporarily opened at \"https://anonymous.4open.science/r/ICLR-25-No7181-codes\".", "keywords": [ "Clustering", "graph data", "representation learning", "quaternion" ], "rating_list": [ 5, 6, 1, 3 ], "soundness_list": [ 2, 3, 2, 1 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "bKAqK7Bh7n", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Dongxia Wu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Ph.D student" }, { "name": "Germano Heinzelmann", "gender": "Male", "institution": "Universidade Federal de Santa Catarina", "country": "", "position": "Associate Professor" }, { "name": "Michael K Gilson", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Peter Eckmann", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "Undergrad student" }, { "name": "Rose Yu", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 61, "n_ref": 122, "n_ref_all": 131, "n_fig": 3, "n_tab": 4, "L_tab": 853, "n_element_tab": 35, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1263, "n_element_tab_1": 43, "formula_len_all": 944, "formula_len_all_1": 860, "len_all": 163535, "len_all_1": 68063, "len_abs": 1325, "len_title": 130, "len_sents": 43963, "len_sents_1": 31878, "n_sents": 327, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1330, "title": "MF-LAL: Drug Compound Generation Using Multi-Fidelity Latent Space Active Learning", "abs": "Current generative models for drug discovery primarily use molecular docking as an oracle to guide the generation of active compounds. However, such models are often not useful in practice because even compounds with high docking scores do not consistently show experimental activity. More accurate methods for activity prediction exist, such as molecular dynamics based binding free energy calculations, but they are too computationally expensive to use in a generative model. To address this challenge, we propose Multi-Fidelity Latent space Active Learning (MF-LAL), a generative modeling framework that integrates a set of oracles with varying cost-accuracy tradeoffs. We train a surrogate model for each oracle and use these surrogates to generate compounds with high predicted activity. Unlike previous approaches that separately learn the surrogate model and generative model, MF-LAL combines the generative and multi-fidelity surrogate models into a single framework, allowing for more accurate activity prediction and higher quality samples. We train MF-LAL with a novel active learning algorithm to further reduce computational cost. Our experiments on two disease-relevant proteins show that MF-LAL produces compounds with significantly better binding free energy scores than other single and multi-fidelity approaches.", "keywords": [ "drug discovery", "multi-fidelity learning", "generative models" ], "rating_list": [ 5, 3, 8, 5, 5 ], "soundness_list": [ 1, 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 1, 1, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "bJLO9S6XOj", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Dexiong Chen", "gender": "Male", "institution": "Max-Planck Institute", "country": "DE", "position": "Project leader" }, { "name": "Jenna Wiens", "gender": "Female", "institution": "University of Michigan Ann Arbor", "country": "US", "position": "Associate Professor" }, { "name": "Karsten Borgwardt", "gender": "unknown", "institution": "Max Planck Institute of Biochemistry", "country": "DE", "position": "Max Planck Director" }, { "name": "Markus Krimmel", "gender": "Male", "institution": "Max Planck Institute of Biochemistry", "country": "", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 15, "n_ref_uni": 44, "n_ref": 144, "n_ref_all": 175, "n_fig": 5, "n_tab": 22, "L_tab": 6238, "n_element_tab": 716, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1832, "n_element_tab_1": 208, "formula_len_all": 2372, "formula_len_all_1": 913, "len_all": 233013, "len_all_1": 65147, "len_abs": 1131, "len_title": 116, "len_sents": 69661, "len_sents_1": 28795, "n_sents": 648, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1136, "title": "Towards Fast Graph Generation via Autoregressive Filtration Modeling", "abs": "Graph generative models often face a critical trade-off between learning complex distributions and achieving fast generation speed. We introduce Autoregressive Filtration Modeling (AFM), a novel approach that addresses both challenges. AFM leverages filtration, a concept from topological data analysis, to transform graphs into short sequences of monotonically increasing subgraphs. This enables a structured autoregressive generation process, contrasting with the stochastic trajectories of diffusion models. We propose a novel autoregressive graph mixer model to learn this filtration process, coupled with a noise augmentation strategy to mitigate exposure bias and a reinforcement learning approach to refine the generative model. Extensive experiments on diverse synthetic and real-world datasets demonstrate AFM's superior performance compared to existing autoregressive models. Additionally, AFM achieves a 100-fold speedup in generation time compared to state-of-the-art diffusion models while maintaining the quality of generated graphs. This work represents a significant advancement towards high-throughput graph generation.", "keywords": [ "Graph generation", "Autoregressive model", "Generative modeling" ], "rating_list": [ 8, 6, 6, 5, 3 ], "soundness_list": [ 3, 2, 3, 2, 1 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "bJ33TvbJW0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Juntao Li", "gender": "Male", "institution": "Soochow University, China", "country": "", "position": "Associate Professor" }, { "name": "Min Zhang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Ping Li", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Qingrong Xia", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Xinyu Duan", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Yi Su", "gender": "Male", "institution": "Suzhou University", "country": "", "position": "MS student" }, { "name": "Yuechi Zhou", "gender": "Male", "institution": "Suzhou University", "country": "", "position": "MS student" }, { "name": "Zhefeng Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 22, "n_ref": 39, "n_ref_all": 65, "n_fig": 15, "n_tab": 5, "L_tab": 1348, "n_element_tab": 214, "n_fig_1": 14, "n_tab_1": 2, "L_tab_1": 579, "n_element_tab_1": 103, "formula_len_all": 209, "formula_len_all_1": 213, "len_all": 146905, "len_all_1": 64816, "len_abs": 1461, "len_title": 112, "len_sents": 36909, "len_sents_1": 29595, "n_sents": 301, "n_sents_1": 241, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1343, "title": "SinkQ: Accurate 2-bit KV Cache Quantization with Dynamic Sink Tracking", "abs": "The impressive capabilities of large language models (LLMs) come at the cost of substantial computational resources during deployment. While KV Cache can significantly reduce recomputation during inference, it also introduces additional memory overhead. KV Cache quantization presents a promising solution, striking a good balance between memory usage and accuracy.\nPrevious research has shown that the Keys are distributed by channel, while the Values are distributed by token. Consequently, the common practice is to apply channel-wise quantization to the Keys and token-wise quantization to the Values. However, our further investigation reveals that a small subset of unusual tokens exhibit unique characteristics that deviate from this pattern, which can substantially impact quantization accuracy. Furthermore, these tokens often have higher attention scores, exacerbating their quantization errors.\nTo address this, we develop a simple yet effective method to identify these tokens accurately during the decoding process and exclude them from quantization, significantly improving overall accuracy. Extensive experiments show that our method achieves significant accuracy improvements under 2-bit quantization and can deliver a 6.4\u00d7 reduction in memory usage and a 2.3\u00d7 increase in throughput. Our code will be released upon acceptance.", "keywords": [ "KV Cache", "Large Language Models", "Quantization" ], "rating_list": [ 6, 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 1, 1, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "bIup4xWg9K", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chengcheng Wan", "gender": "unknown", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaodong Gu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yalan Lin", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "YiXiong Fang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 38, "n_ref": 51, "n_ref_all": 68, "n_fig": 6, "n_tab": 6, "L_tab": 3655, "n_element_tab": 395, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 775, "n_element_tab_1": 101, "formula_len_all": 200, "formula_len_all_1": 200, "len_all": 124688, "len_all_1": 52487, "len_abs": 3299, "len_title": 106, "len_sents": 32099, "len_sents_1": 23938, "n_sents": 238, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1217, "title": "CodeCipher: Learning To Obfuscate Source Code Against LLMs", "abs": "While code large language models have made significant strides in AI-assisted coding tasks, there are growing concerns about privacy challenges. The user code is transparent to the cloud LLM service provider, inducing risks of unauthorized training, reading, and execution of the user code. In this paper, we propose CodeCipher, a novel method that perturbs privacy from code while preserving the original response from LLMs. CodeCipher transforms the LLM's embedding matrix so that each row corresponds to a different word in the original matrix, forming a token-to-token confusion mapping for obfuscating source code. The new embedding matrix is optimized through minimizing the task-specific loss function. To tackle the challenge from the discrete and sparse nature of word vector spaces, CodeCipher adopts a discrete optimization strategy that aligns the updated vector to the nearest valid token in the vocabulary before each gradient update. We demonstrate the effectiveness of our approach on three AI-assisted coding tasks including code completion, summarization, and translation. Results show that our model successfully confuses the privacy in source code while preserving the original LLM's performance.", "keywords": [ "Code obfuscation; LLM privacy; Large Language Models of Code" ], "rating_list": [ 5, 8, 5, 6 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "bIoWuzFm6r", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Alina Bialkowski", "gender": "Female", "institution": "University of Queensland", "country": "", "position": "Lecturer" }, { "name": "Chenhao Zhang", "gender": "Male", "institution": "University of Queensland", "country": "", "position": "PhD student" }, { "name": "Miao Xu", "gender": "Female", "institution": "University of Queensland", "country": "AU", "position": "Lecturer" }, { "name": "Shaofei Shen", "gender": "Male", "institution": "University of Queensland", "country": "AU", "position": "PhD student" }, { "name": "Yawen Zhao", "gender": "Female", "institution": "University of Queensland", "country": "", "position": "PhD student" }, { "name": "Weitong Tony Chen", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Lecturer" } ], "n_formula": 57, "n_formula_1": 25, "n_ref_uni": 44, "n_ref": 116, "n_ref_all": 140, "n_fig": 8, "n_tab": 14, "L_tab": 15414, "n_element_tab": 938, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 2001, "n_element_tab_1": 120, "formula_len_all": 5513, "formula_len_all_1": 1783, "len_all": 211772, "len_all_1": 72083, "len_abs": 1591, "len_title": 91, "len_sents": 61986, "len_sents_1": 30047, "n_sents": 501, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1620, "title": "Machine Unlearning for Streaming Forgetting", "abs": "Machine unlearning aims to remove knowledge derived from the specific training data that are requested to be forgotten in a well-trained model while preserving the knowledge learned from the remaining training data. Currently, machine unlearning methods typically handle all forgetting data in a single batch, removing the corresponding knowledge all at once upon request. However, in practical scenarios, requests for data removal often arise in a streaming manner rather than in a single batch, leading to reduced efficiency and effectiveness in existing methods. Such challenges of streaming forgetting have not been the focus of much research. In this paper, to address the challenges of performance maintenance, efficiency, and data access brought about by streaming unlearning requests, we introduce an online unlearning paradigm, formalizing the unlearning as a distribution shift problem. We then estimate the altered distribution and propose a novel online unlearning algorithm to achieve efficient streaming forgetting without requiring access to the original training data. Theoretical analyses confirm an $O(V_T\\sqrt{T} + \\Delta_T)$ error bound on the streaming unlearning regret, where $V_T$ represents the cumulative total variation in the optimal solution over $T$ learning rounds and $\\Delta_T$ represents the cumulative total divergence between remaining and forgetting data distributions. This theoretical guarantee is achieved under mild conditions without the strong restriction of convex loss function. Experiments across various models and datasets validate the performance of our proposed method.", "keywords": [ "Machine Unlearning" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "bIlnpVM4bc", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Chen Liang", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Liliang Ren", "gender": "unknown", "institution": "Microsoft GenAI", "country": "", "position": "Senior Researcher" }, { "name": "Weizhu Chen", "gender": "Male", "institution": "Microsoft GenAI", "country": "US", "position": "Vice President" }, { "name": "Yadong Lu", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Yang Liu", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "yelong shen", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 83, "n_ref": 153, "n_ref_all": 195, "n_fig": 9, "n_tab": 10, "L_tab": 4474, "n_element_tab": 698, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 2767, "n_element_tab_1": 392, "formula_len_all": 447, "formula_len_all_1": 365, "len_all": 267858, "len_all_1": 63125, "len_abs": 1361, "len_title": 137, "len_sents": 48334, "len_sents_1": 27455, "n_sents": 301, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1366, "title": "Samba: Simple Hybrid State Space Models for Efficient Unlimited Context Language Modeling", "abs": "Efficiently modeling sequences with infinite context length has long been a challenging problem. Previous approaches have either suffered from quadratic computational complexity or limited extrapolation ability in length generalization. In this\nwork, we present Samba, a simple hybrid architecture that layer-wise combines\nMamba, a selective State Space Model (SSM), with Sliding Window Attention\n(SWA). Samba selectively compresses a given sequence into recurrent hidden\nstates while still maintaining the ability to precisely recall recent memories with the\nattention mechanism. We scale Samba up to 3.8B parameters with 3.2T training\ntokens and demonstrate that it significantly outperforms state-of-the-art models\nacross a variety of benchmarks. Pretrained on sequences of 4K length, Samba\nshows improved perplexity in context lengths of up to 1M in zero-shot. When\nfinetuned on 4K-length sequences, Samba efficiently extrapolates to a 256K context length with perfect memory recall on the Passkey Retrieval task, and exhibits\nsuperior retrieval extrapolation on the challenging Phonebook task compared to\nfull-attention models. As a linear-time sequence model, Samba achieves a 3.73\u00d7\nhigher throughput compared to Transformers with grouped-query attention for user\nprompts of 128K length, and a 3.64\u00d7 speedup when generating 64K tokens with\nunlimited streaming.", "keywords": [ "Large Language Models;Length Extrapolation;Efficiency;Hybrid State Space Models" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "bIf1YXztnD", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Christopher Scarvelis", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "David Benhaim", "gender": "Male", "institution": "BackflipAI", "country": "US", "position": "Principal Researcher" }, { "name": "Paul Zhang", "gender": "Male", "institution": "backflipAI", "country": "US", "position": "Principal Researcher" } ], "n_formula": 25, "n_formula_1": 10, "n_ref_uni": 28, "n_ref": 50, "n_ref_all": 72, "n_fig": 11, "n_tab": 2, "L_tab": 225, "n_element_tab": 22, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 76, "n_element_tab_1": 10, "formula_len_all": 2916, "formula_len_all_1": 652, "len_all": 125441, "len_all_1": 64303, "len_abs": 933, "len_title": 63, "len_sents": 44786, "len_sents_1": 30740, "n_sents": 311, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 15, "L_abs": 937, "title": "Orient Anything", "abs": "Orientation estimation is a fundamental task in 3D shape analysis which consists of estimating a shape's orientation axes: its side-, up-, and front-axes. Using this data, one can rotate a shape into canonical orientation, where its orientation axes are aligned with the coordinate axes. Developing an orientation algorithm that reliably estimates complete orientations of general shapes remains an open problem. We introduce a two-stage orientation pipeline that achieves state of the art performance on up-axis estimation and further demonstrate its efficacy on full-orientation estimation, where one seeks all three orientation axes. Unlike previous work, we train and evaluate our method on all of Shapenet rather than a subset of classes. We motivate our engineering contributions by theory describing fundamental obstacles to orientation estimation for rotationally-symmetric shapes, and show how our method avoids these obstacles.", "keywords": [ "3d orientation", "shape analysis", "3d deep learning", "geometric deep learning" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "bI3fcTsKW4", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Shiyun Xu", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Zhiqi Bu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 40, "n_formula_1": 23, "n_ref_uni": 37, "n_ref": 59, "n_ref_all": 97, "n_fig": 10, "n_tab": 9, "L_tab": 2595, "n_element_tab": 345, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1416, "n_element_tab_1": 212, "formula_len_all": 3590, "formula_len_all_1": 1800, "len_all": 189313, "len_all_1": 62357, "len_abs": 797, "len_title": 97, "len_sents": 45008, "len_sents_1": 24222, "n_sents": 339, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 802, "title": "Gradient descent with generalized Newton\u2019s method", "abs": "We propose the generalized Newton's method (GeN) --- a Hessian-informed approach that applies to any optimizer such as SGD and Adam, and covers the Newton-Raphson method as a sub-case. Our method automatically and dynamically selects the learning rate that accelerates the convergence, without the intensive tuning of the learning rate scheduler. In practice, our method is easily implementable, since it only requires additional forward passes with almost zero computational overhead (in terms of training time and memory cost), if the overhead is amortized over many iterations. We present extensive experiments on language and vision tasks (e.g. GPT and ResNet) to showcase that GeN optimizers match the state-of-the-art performance, which was achieved with carefully tuned learning rate schedulers.", "keywords": [ "Optimization", "Hessian matrix", "Learning rate scheduler" ], "rating_list": [ 6, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 1, 4, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "bHY0Uypoh9", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Guang Lin", "gender": "Male", "institution": "Purdue University", "country": "", "position": "Associate Professor" }, { "name": "Haoyang Zheng", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Ruqi Zhang", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 51, "n_formula_1": 18, "n_ref_uni": 41, "n_ref": 61, "n_ref_all": 78, "n_fig": 7, "n_tab": 5, "L_tab": 934, "n_element_tab": 117, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 664, "n_element_tab_1": 89, "formula_len_all": 5578, "formula_len_all_1": 1463, "len_all": 169421, "len_all_1": 64204, "len_abs": 864, "len_title": 94, "len_sents": 51954, "len_sents_1": 27537, "n_sents": 420, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 845, "title": "Exploring Non-Convex Discrete Energy Landscapes: A Langevin-Like Sampler with Replica Exchange", "abs": "Gradient-based Discrete Samplers (GDSs) are effective for sampling discrete energy landscapes. However, they often stagnate in complex, non-convex settings. To improve exploration, we introduce the Discrete Replica EXchangE Langevin (DREXEL) sampler and its variant with Adjusted Metropolis (DREAM). These samplers use two GDSs at different temperatures and step sizes: one focuses on local exploitation, while the other explores broader energy landscapes. When energy differences are significant, sample swaps occur, which are determined by a mechanism tailored for discrete sampling to ensure detailed balance. Theoretically, we prove both DREXEL and DREAM converge asymptotically to the target energy and exhibit faster mixing than a single GDS. Experiments further confirm their efficiency in exploring non-convex discrete energy landscapes.", "keywords": [ "discrete sampling", "Langevin dynamics", "replica exchange", "energy-based model", "Metropolis-Hastings", "Markov Chain Monte Carlo", "locally-balanced proposal" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 3, 4, 2 ] }, { "paper_id": "bHNVmLDtFo", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Can Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Chun Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Jiawei Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Peng Wu", "gender": "Male", "institution": "Beijing Technology and Business University", "country": "CN", "position": "Associate Professor" }, { "name": "Shengjia Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Weiqin Yang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yan Feng", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 26, "n_formula_1": 13, "n_ref_uni": 61, "n_ref": 161, "n_ref_all": 198, "n_fig": 6, "n_tab": 3, "L_tab": 2115, "n_element_tab": 79, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 444, "n_element_tab_1": 43, "formula_len_all": 2228, "formula_len_all_1": 963, "len_all": 210946, "len_all_1": 64102, "len_abs": 1144, "len_title": 111, "len_sents": 61515, "len_sents_1": 28901, "n_sents": 560, "n_sents_1": 248, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1185, "title": "Towards Optimizing Top-$K$ Ranking Metrics in Recommender Systems", "abs": "In the realm of recommender systems (RS), Top-$K$ metrics such as NDCG@$K$ are the gold standard for evaluating performance. Nonetheless, during the training of recommendation models, optimizing NDCG@$K$ poses significant challenges due to its inherent discontinuous nature and the intricacies of the Top-K truncation mechanism. Recent efforts to optimize NDCG@$K$ have either neglected the Top-$K$ truncation or suffered from low computational efficiency. To overcome these limitations, we propose SoftmaxLoss@$K$ (SL@$K$), a new loss function designed as a surrogate for optimizing NDCG@$K$ in RS. SL@$K$ integrates a quantile-based technique to handle the complex truncation term; and derives a smooth approximation of NDCG@$K$ to address discontinuity. Our theoretical analysis confirms the close bounded relationship between NDCG@$K$ and SL@$K$. \n Besides, SL@$K$ also exhibits several desirable properties including concise formulation, computational efficiency, and noisy robustness. Extensive experiments on four real-world datasets and three recommendation backbones demonstrate that SL@$K$ outperforms existing loss functions with a notable average improvement of 6.19\\%.", "keywords": [ "Recommender Systems", "Surrogate Loss", "Top-K Recommendation", "NDCG@K" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "bGkPZtisSm", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Shawn Im", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Yixuan Li", "gender": "Female", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 82, "n_formula_1": 27, "n_ref_uni": 80, "n_ref": 141, "n_ref_all": 162, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6221, "formula_len_all_1": 1888, "len_all": 236056, "len_all_1": 71811, "len_abs": 1433, "len_title": 101, "len_sents": 48426, "len_sents_1": 34015, "n_sents": 381, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1439, "title": "On the Generalization of Preference Learning with DPO", "abs": "Large language models (LLMs) have demonstrated remarkable capabilities but often struggle to align with human preferences, leading to harmful or undesirable outputs. Preference learning, which trains models to distinguish between preferred and non-preferred responses based on human feedback, has become a crucial component for ensuring that LLMs align with human values. Despite the widespread adoption in real-world systems, a thorough theoretical understanding of the generalization guarantees for these models remains lacking. This paper bridges that gap by introducing a new theoretical framework to analyze the generalization guarantees of models trained with direct preference optimization. While existing generalization theory often focuses on overparameterized models achieving near-optimal loss or models independent of the training process, our framework rigorously assesses how well models generalize after a finite number of gradient steps, reflecting real-world LLM training practices. By analyzing the reward margin associated with each sample and its trajectory throughout training, we can effectively bound the generalization error. We derive learning guarantees showing that, under specific conditions, models trained with DPO can correctly discern preferred responses on unseen data with high probability. These insights are empirically validated on contemporary LLMs, underscoring the practical relevance of our theory.", "keywords": [ "preference learning", "generalization bound" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "bGGMLWAGMc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Hehuan Liu", "gender": "Not Specified", "institution": "PCG", "country": "CN", "position": "Researcher" }, { "name": "Lei Li", "gender": "Male", "institution": "Tencent", "country": "", "position": "Algorithmic Engineer" }, { "name": "Xudong Weng", "gender": "unknown", "institution": "Tencent", "country": "CN", "position": "Engineer" }, { "name": "Yaxin Zhou", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Yi YUAN", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "", "position": "PhD student" }, { "name": "Zang Li", "gender": "Male", "institution": "Tencent", "country": "", "position": "Researcher" }, { "name": "ZhaoYang Gui", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Algorithmic Engineer" }, { "name": "Zheng Wei", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 13, "n_ref": 88, "n_ref_all": 114, "n_fig": 9, "n_tab": 3, "L_tab": 1145, "n_element_tab": 141, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1280, "formula_len_all_1": 1229, "len_all": 181989, "len_all_1": 56982, "len_abs": 1266, "len_title": 157, "len_sents": 42289, "len_sents_1": 26522, "n_sents": 329, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1273, "title": "Improving Reasoning Ability of Large Language Models via Iterative Uncertainty-based Preference Optimization", "abs": "Direct Preference Optimization (DPO) has recently emerged as an efficient and effective method for aligning large language models with human preferences.\nHowever, constructing high-quality preference datasets remains challenging, often necessitating expensive manual or powerful LM annotations. Additionally, standard DPO exhibits suboptimal performance in complex reasoning tasks, such as mathematical and code reasoning.\nIn this paper, we introduce an approach to collect preference pairs through iterative sampling and execution feedback, tailored to the current learning state (e.g. well-learned, mis-learned, and unlearned) of the policy model.\nTo alleviate the failures of DPO and improve its applicability in reasoning tasks, we propose IUPO, an iterative uncertainty-based preference optimization method that achieves fine-grained preference control by assessing model confidence.\nWe validate our approach across three reasoning tasks, incorporating five established reasoning datasets and one self-curated dataset. Our experimental results demonstrate an overall improvement of 3.6% over the standard DPO method. \nFurthermore, our approach exhibits promising generalizability involving weak-to-strong (8B to 70B) and cross-model (Llama to Mistral) generalizations.", "keywords": [ "Preference Optimization", "Large Language Model", "Iterative Optimization", "Uncertainty" ], "rating_list": [ 5, 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 2, 3, 4 ] }, { "paper_id": "bG61JDN4E8", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Cornelia Caragea", "gender": "unknown", "institution": "University of Illinois at Chicago", "country": "US", "position": "Full Professor" }, { "name": "Tiberiu Sosea", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "US", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 35, "n_ref": 68, "n_ref_all": 78, "n_fig": 2, "n_tab": 7, "L_tab": 2812, "n_element_tab": 298, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 2813, "n_element_tab_1": 298, "formula_len_all": 651, "formula_len_all_1": 651, "len_all": 132917, "len_all_1": 70379, "len_abs": 1004, "len_title": 120, "len_sents": 35093, "len_sents_1": 32634, "n_sents": 195, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 971, "title": "LANE: Label-Aware Noise Elimination for Fine-Grained Text Classification", "abs": "We propose Label-Aware Noise Elimination (LANE), a new approach that improves the robustness of deep learning models in fine-grained text classification when trained under increased label noise. LANE leverages the semantic relations between classes and monitors the training dynamics of the model on each training example to dynamically lower the importance of training examples that may have noisy labels. We test the effectiveness of LANE in fine-grained text classification and benchmark our approach on a wide variety of datasets with various number of classes and various amounts of label noise. LANE considerably outperforms strong baselines on all datasets, obtaining significant improvements ranging from an average improvement of 2.4% in F1 on manually annotated datasets to a considerable average improvement of 4.5% F1 on datasets with injected noisy labels. We carry out comprehensive analyses of LANE and identify the key components that lead to its success.", "keywords": [ "label noise", "training dynamics", "example reweighting" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "bFYST1MaGh", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Kenneth Li", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Vignav Ramesh", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "Undergrad student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 45, "n_ref": 80, "n_ref_all": 100, "n_fig": 4, "n_tab": 5, "L_tab": 1544, "n_element_tab": 129, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1418, "n_element_tab_1": 93, "formula_len_all": 1065, "formula_len_all_1": 789, "len_all": 182452, "len_all_1": 51880, "len_abs": 1406, "len_title": 104, "len_sents": 37630, "len_sents_1": 24848, "n_sents": 201, "n_sents_1": 128, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1430, "title": "Communicating Activations Between Language Model Agents", "abs": "Communication between multiple language model (LM) agents has been shown to scale up the reasoning ability of LMs. While natural language has been the dominant medium for inter-LM communication, it is not obvious this should be the standard: not only does natural language communication incur high inference costs that scale quickly with the number of both agents and messages, but also the decoding process abstracts away too much rich information that could be otherwise accessed from the internal activations. In this work, we propose a simple technique whereby LMs communicate via *activations*; concretely, we pause an LM $B$'s computation at an intermediate layer, combine its current activation with another LM $A$'s intermediate activation via some function $f$, then pass $f$'s output into the next layer of $B$ and continue the forward pass till decoding is complete. This approach scales up LMs on new tasks with *zero* additional parameters and data, and saves a *substantial amount of compute* over natural language communication. We test our method with various functional forms $f$ on two experimental setups\u2014multi-player coordination games and reasoning benchmarks\u2014and find that it achieves up to $27.0$% improvement over natural language communication across datasets with $<$$1/4$ the compute, illustrating the superiority and robustness of activations as an alternative \"language\" for communication between LMs.", "keywords": [ "large language models", "multiagent communication", "embedding representation", "multiagent debate" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "bFHR8hNk4I", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Elie Hachem", "gender": "unknown", "institution": "Mines ParisTech", "country": "FR", "position": "Full Professor" }, { "name": "Jonathan Viquerat", "gender": "unknown", "institution": "Mines ParisTech", "country": "FR", "position": "Researcher" }, { "name": "Paul Garnier", "gender": "Male", "institution": "Mines ParisTech", "country": "FR", "position": "PhD student" }, { "name": "Vincent Lannelongue", "gender": "Male", "institution": "Mines Paris PSL", "country": "FR", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 23, "n_ref": 34, "n_ref_all": 61, "n_fig": 10, "n_tab": 6, "L_tab": 1758, "n_element_tab": 276, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1111, "n_element_tab_1": 206, "formula_len_all": 195, "formula_len_all_1": 0, "len_all": 107744, "len_all_1": 48544, "len_abs": 1217, "len_title": 117, "len_sents": 27158, "len_sents_1": 20274, "n_sents": 214, "n_sents_1": 155, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1225, "title": "MeshMask: Physics-Based Simulations with Masked Graph Neural Networks", "abs": "We introduce a novel masked pre-training technique for graph neural networks (GNNs) applied to computational fluid dynamics (CFD) problems. By randomly masking up to 40\\% of input mesh nodes during pre-training, we force the model to learn robust representations of complex fluid dynamics. We pair this masking strategy with an asymmetric encoder-decoder architecture and gated multi-layer perceptrons to further enhance performance. The proposed method achieves state-of-the-art results on seven CFD datasets, including a new challenging dataset of 3D intracranial aneurysm simulations with over 250,000 nodes per mesh. Moreover, it significantly improves model performance and training efficiency across such diverse range of fluid simulation tasks. We demonstrate improvements of up to 60\\% in long-term prediction accuracy compared to previous best models, while maintaining similar computational costs. Notably, our approach enables effective pre-training on multiple datasets simultaneously, significantly reducing the time and data required to achieve high performance on new tasks.\nThrough extensive ablation studies, we provide insights into the optimal masking ratio, architectural choices, and training strategies.", "keywords": [ "graph networks", "simulation", "mesh", "physics" ], "rating_list": [ 8, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "bEvI30Hb2W", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Asim Kadav", "gender": "Male", "institution": "Samsung Research America", "country": "US", "position": "Researcher" }, { "name": "Saket Gurukar", "gender": "Male", "institution": "Samsung Research America", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 51, "n_ref": 77, "n_ref_all": 94, "n_fig": 6, "n_tab": 2, "L_tab": 609, "n_element_tab": 82, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 609, "n_element_tab_1": 82, "formula_len_all": 69, "formula_len_all_1": 69, "len_all": 133801, "len_all_1": 68452, "len_abs": 863, "len_title": 115, "len_sents": 34144, "len_sents_1": 30906, "n_sents": 271, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 868, "title": "LVM-NET: Efficient Long-Form Video Reasoning using neural sampling", "abs": "Long-form video reasoning is essential for various applications such as video retrieval, summarizing, and question\nanswering. However, existing methods often require significant computational resources and are limited by GPU memory constraints. To address this challenge, we present Long-Video Memory Network, LVM-NET, a novel video reasoning method that employs a fixed-size memory representation to store discriminative patches sampled from the input video. By leveraging a neural sampler that identifies discriminative memory tokens, LVM-Net achieves improved efficiency. Furthermore, LVM-Net only requires a single pass over the video, further enhancing overall efficiency. Our results on the Rest-ADL dataset demonstrate an 18x - 75x improvement in inference times for long-form video retrieval and answering questions, with a competitive predictive performance.", "keywords": [ "Video Reasoning", "External Memory", "Neural Sampling" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "bEqI61iBue", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Guang Dai", "gender": "Male", "institution": "SGIT AI", "country": "CN", "position": "Principal Researcher" }, { "name": "Haishan Ye", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Sizhe Dang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yanjun Zhao", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yi Qian", "gender": "unknown", "institution": "Xi'an Jiaotong University", "country": "", "position": "Associate Professor" }, { "name": "Ivor W Tsang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" } ], "n_formula": 43, "n_formula_1": 24, "n_ref_uni": 51, "n_ref": 87, "n_ref_all": 119, "n_fig": 15, "n_tab": 8, "L_tab": 2459, "n_element_tab": 311, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2218, "n_element_tab_1": 265, "formula_len_all": 5192, "formula_len_all_1": 1933, "len_all": 158233, "len_all_1": 54848, "len_abs": 1416, "len_title": 137, "len_sents": 34675, "len_sents_1": 18892, "n_sents": 328, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1461, "title": "Second-Order Fine-Tuning without Pain for LLMs: A Hessian Informed Zeroth-Order Optimizer", "abs": "Fine-tuning large language models (LLMs) is necessary for specific downstream tasks, but classic first-order optimizer entails prohibitive GPU memory because of the back propagation. Recent works such as MeZO have turned to zeroth-order optimizers for fine-tuning, which reduce substantial memory by using two forward passes. However, heterogeneous curvatures across different parameter dimensions in LLMs often cause model convergence instability or even failure. In this work, we propose HiZOO, a diagonal Hessian informed Zeroth-Order Optimizer , which is the first work to leverage the diagonal Hessian to enhance ZOO for fine-tuning LLMs. We provide theoretical proof for HiZOO and visualize the optimization trajectories on test functions to illustrate how it improves convergence in handling heterogeneous curvatures. Extensive experiments on various models (RoBERTa, OPT, Phi-2 and LLama3, with 350M$\\sim$66B parameters) indicate that HiZOO significantly reduces training steps and enhances model accuracy, while keeping the memory advantage of ZOO. For example, on SST2 task HiZOO achieves $8\\times$ speedup and better accuracy over MeZO across different models. We also propose HiZOO-L, which reduces the Hessian memory cost to 10\\% of the MeZO, while maintaining almost same performance. Compared with ZO-Adam, HiZOO-L achieves a 4.3\\% improvement, just using 50\\% of the GPU memory. Code is available at https://anonymous.4open.science/r/HiZOO-27F8.", "keywords": [ "LLM; deep learning; zeroth order optimizer" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "bEgDEyy2Yk", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 1, "author_info_list": [ { "name": "Gangli Liu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 14, "n_ref": 35, "n_ref_all": 50, "n_fig": 5, "n_tab": 2, "L_tab": 1339, "n_element_tab": 142, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1339, "n_element_tab_1": 142, "formula_len_all": 57, "formula_len_all_1": 57, "len_all": 53648, "len_all_1": 53648, "len_abs": 1011, "len_title": 152, "len_sents": 15314, "len_sents_1": 15314, "n_sents": 137, "n_sents_1": 137, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1030, "title": "An efficient implementation for solving the all pairs minimax path problem in an undirected dense graph", "abs": "We provide an efficient $ O(n^2) $ implementation for solving the all pairs minimax path problem or widest path problem in an undirected dense graph. It is a code implementation of the Algorithm 4 (MMJ distance by Calculation and Copy) in a previous paper. The distance matrix is also called the all points path distance (APPD). We conducted experiments to test the implementation and algorithm, compared it with several other algorithms for solving the APPD matrix. Result shows Algorithm 4 works good for solving the widest path or minimax path APPD matrix. It can drastically improve the efficiency for computing the APPD matrix. There are several theoretical outcomes which claim the APPD matrix can be solved accurately in $ O(n^2) $ . However, they are impractical because there is no code implementation of these algorithms. It seems Algorithm 4 is the first algorithm that has an actual code implementation for solving the APPD matrix of minimax path or widest path problem in $ O(n^2) $, in an undirected dense graph.", "keywords": [ "Minimax path problem", "Longest-leg path distance", "Min-Max-Jump distance", "Widest path problem", "Maximum capacity path problem", "Bottleneck edge query problem", "All points path distance", "Floyd-Warshall algorithm", "Minimum spanning tree" ], "rating_list": [ 1, 1, 1, 1 ], "soundness_list": [ 1, 3, 1, 1 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 1, 1, 1, 1 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "bEbQBiMpUI", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Bochuan Cao", "gender": "unknown", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Guangliang Liu", "gender": "Male", "institution": "Michigan State University", "country": "", "position": "PhD student" }, { "name": "Jiliang Tang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Full Professor" }, { "name": "Kristen Johnson", "gender": "Female", "institution": "Michigan State University", "country": "", "position": "Assistant Professor" }, { "name": "Rongrong Wang", "gender": "unknown", "institution": "Michigan State University", "country": "US", "position": "Associate Professor" }, { "name": "Xitong Zhang", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "US", "position": "Researcher" }, { "name": "Zhiyu Xue", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Mao Haitao", "gender": "Male", "institution": "Michigan State University", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 59, "n_ref": 116, "n_ref_all": 138, "n_fig": 7, "n_tab": 2, "L_tab": 44, "n_element_tab": 2, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 558, "formula_len_all_1": 558, "len_all": 179299, "len_all_1": 71271, "len_abs": 1747, "len_title": 125, "len_sents": 51953, "len_sents_1": 36419, "n_sents": 388, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1694, "title": "Convergence Towards Stable Intrinsic Self-correction of Large Language Models", "abs": "Large Language Models (LLMs) are able to improve their responses when instructed to do so, a capability known as self-correction. \nWhen instructions provide only the task's goal without specific details about potential issues in the response, LLMs must rely on their internal knowledge to improve response quality, a process referred to as intrinsic self-correction. \nThe empirical success of intrinsic self-correction is evident in various applications, but how and why it is effective remains unknown.\nIn this paper, we unveil that intrinsic self-correction can be progressively improved, allowing it to approach a converged state. Our findings are verified in: (1) the scenario of multi-round question answering, by comprehensively demonstrating that intrinsic self-correction can progressively introduce performance gains through iterative interactions, ultimately converging to stable performance; and (2) the context of intrinsic self-correction for enhanced morality, in which we provide empirical evidence that iteratively applying instructions reduces model uncertainty towards convergence, which then leads to convergence of both the calibration error and self-correction performance, ultimately resulting in a stable state of intrinsic self-correction. Furthermore, we introduce a mathematical formulation and a simulation task indicating that the latent concepts activated by self-correction instructions drive the reduction of model uncertainty.\nBased on our experimental results and analysis of the convergence of intrinsic self-correction, we reveal its underlying mechanism: consistent injected instructions reduce model uncertainty which yields converged, improved performance.", "keywords": [ "self-correction", "large language models", "morality", "toxicity", "social bias" ], "rating_list": [ 5, 8, 3, 8, 3 ], "soundness_list": [ 2, 3, 3, 4, 2 ], "presentation_list": [ 2, 4, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 2, 4, 4, 3, 3 ] }, { "paper_id": "bESxQeXTlo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Di Huang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Guodong Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Jinjin Zhang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Xiuguo Bao", "gender": "unknown", "institution": "cncert/cc", "country": "", "position": "Researcher" }, { "name": "Yizhou jin", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Yunhong Wang", "gender": "Female", "institution": "Beihang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 29, "n_ref": 67, "n_ref_all": 75, "n_fig": 4, "n_tab": 1, "L_tab": 256, "n_element_tab": 43, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 256, "n_element_tab_1": 43, "formula_len_all": 149, "formula_len_all_1": 149, "len_all": 82125, "len_all_1": 54339, "len_abs": 1611, "len_title": 126, "len_sents": 27452, "len_sents_1": 27698, "n_sents": 201, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1593, "title": "CLIP-LAD: Unleash the Potential of CLIP for Few-shot Logical Anomaly Detection", "abs": "Anomaly detection (AD) is crucial for visual inspections, and includes two main types: structural and logical anomalies. Despite growing interest in AD, most methods focus on structural anomalies, while few works address logical anomaly detection (LAD), which requires a global understanding of the context. Leading LAD methods often advocate segmentation algorithms to parse logical relations within images, necessitating extensive training images or elaborate labels, but they undergo significant performance degradation in low-data scenarios. This study explores a practical yet challenging scenario where only few-shot normal images are available. To the end, we introduce CLIP-LAD, a novel, training-free method for few-shot LAD. We propose a coarse-to-fine segmentation process, involving foreground extraction and fine-grained alignment, to progressively harness the CLIP's generalization abilities for LAD. Specifically, we first aggregate visual features into different regions with clear boundaries, benefited from the strong visual coherence in vision transformer (ViT), and leverage coarse prompts to help identify the foreground. Within the foreground, we further conduct per-pixel fine-grained classification with fine prompts to parse different parts of an object. The anomaly scoring is derived from the class histograms in the precise segmentation masks. For comprehensive evaluation, we build up a few-shot LAD benchmark based on the MvTec-LOCO dataset and include a series of comparison methods. Experiments on this benchmark demonstrates our superiority in low-data regime.", "keywords": [ "logical anomaly detection", "multi-modal transfer learning" ], "rating_list": [ 3, 5, 1 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "bDt5qc7TfO", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Cody Fleming", "gender": "unknown", "institution": "Iowa State University", "country": "US", "position": "Associate Professor" }, { "name": "Prajwal Koirala", "gender": "Male", "institution": "Iowa State University", "country": "", "position": "MS student" }, { "name": "Soumik Sarkar", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "Full Professor" }, { "name": "Zhanhong Jiang", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "Researcher" } ], "n_formula": 71, "n_formula_1": 26, "n_ref_uni": 48, "n_ref": 104, "n_ref_all": 123, "n_fig": 8, "n_tab": 6, "L_tab": 5133, "n_element_tab": 1164, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 1856, "n_element_tab_1": 451, "formula_len_all": 5391, "formula_len_all_1": 1739, "len_all": 220742, "len_all_1": 81399, "len_abs": 1359, "len_title": 129, "len_sents": 69861, "len_sents_1": 35750, "n_sents": 531, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1332, "title": "Latent Safety-Constrained Policy Approach for Safe Offline Reinforcement Learning", "abs": "In safe offline reinforcement learning, the objective is to develop a policy that maximizes cumulative rewards while strictly adhering to safety constraints, utilizing only offline data. Traditional methods often face difficulties in balancing these constraints, leading to either diminished performance or increased safety risks. We address these issues with a novel approach that begins by learning a conservatively safe policy through the use of Conditional Variational Autoencoders, which model the latent safety constraints. Subsequently, we frame this as a Constrained Reward-Return Maximization problem, wherein the policy aims to optimize rewards while complying with the inferred latent safety constraints. This is achieved by training an encoder with a reward-Advantage Weighted Regression objective within the latent constraint space. Our methodology is supported by theoretical analysis, including bounds on policy performance and sample complexity. Extensive empirical evaluation on benchmark datasets, including challenging autonomous driving scenarios, demonstrates that our approach not only maintains safety compliance but also excels in cumulative reward optimization, surpassing existing methods. Additional visualizations provide further insights into the effectiveness and underlying mechanisms of our approach.", "keywords": [ "Safe RL", "Offline RL", "Variational Autoencoders", "Latent Safety Constraints" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "bDPL0ohHBa", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Meng Fang", "gender": "Male", "institution": "University of Liverpool", "country": "", "position": "Assistant Professor" }, { "name": "Qiang He", "gender": "Male", "institution": "Ruhr-Universit\u00e4t Bochum", "country": "DE", "position": "PhD student" }, { "name": "Setareh Maghsudi", "gender": "unknown", "institution": "Ruhr-Universt\u00e4t Bochum", "country": "", "position": "Full Professor" }, { "name": "Tianyi Zhou", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Yucheng Yang", "gender": "unknown", "institution": "Eindhoven University of Technology", "country": "NL", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 14, "n_ref_uni": 27, "n_ref": 77, "n_ref_all": 135, "n_fig": 6, "n_tab": 8, "L_tab": 1631, "n_element_tab": 156, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1243, "n_element_tab_1": 113, "formula_len_all": 3061, "formula_len_all_1": 1340, "len_all": 168202, "len_all_1": 66417, "len_abs": 1736, "len_title": 112, "len_sents": 50633, "len_sents_1": 29463, "n_sents": 426, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1743, "title": "One Model for All: Multi-Objective Controllable Language Models", "abs": "Aligning large language models (LLMs) with human preference is critical to enhancing LLMs' safety, helpfulness, helpfulness, humor, faithfulness, etc. The current reinforcement learning from human feedback (RLHF) mainly focuses on a fixed reward learned from average human ratings, which may weaken the adaptivity and controllability of varying preferences. However, creating personalized LLMs requires aligning LLMs with individual human preferences, which is non-trivial due to the scarce data per user and the diversity of user preferences on multi-objective trade-offs, such as prioritizing humor and empathy in one context, while seeking efficiency and precision in another. Can we train one LLM to produce personalized outputs for different user preferences on the Pareto front? In this paper, we introduce Multi-Objective Control (MOC), which trains an LLM as a meta-policy to directly generate responses in the preference-defined regions of Pareto front. Our approach integrates multi-objective optimization (MOO) principles into Proximal Policy Optimization (PPO) to train an LLM as a preference-conditioned policy network. We improve the computational efficiency of MOC by applying MOO at the policy level, which enables us to finetune an LLM of 7B parameters on a single A6000 GPU. Extensive experiments demonstrate the advantages of MOC over baselines in three aspects: (i) Controllability of LLM outputs w.r.t. user preferences on the trade-off among multiple rewards; (ii) Quality and diversity of LLM outputs, measured by the hyper-volume of multiple solutions achieved; and (iii) Generalization to unseen preferences. These results highlight MOC\u2019s potential for real-world applications requiring scalable and customizable LLMs.", "keywords": [ "controllable language models", "reinforcement learning from human feedback" ], "rating_list": [ 5, 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 2, 2 ] }, { "paper_id": "bCi3Jz0q02", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 1, "author_info_list": [ { "name": "Hadi Abdullah", "gender": "unknown", "institution": "VISA", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 33, "n_ref": 76, "n_ref_all": 94, "n_fig": 8, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 61, "formula_len_all_1": 227, "len_all": 116170, "len_all_1": 65513, "len_abs": 1823, "len_title": 125, "len_sents": 38608, "len_sents_1": 32149, "n_sents": 307, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1942, "title": "Towards Efficient and Accurate Identification of Memorization in Deep Models", "abs": "\\textit{Memorization} is the ability of deep models to learn verbatim arbitrary inputs from the training data. One of the most popular means of calculating memorization scores (i.e., the probability that a point is memorized) is via the pseudo Leave-One-Out (pLOO) method proposed by~\\citet{feldman2020longtail}. However, this technique suffers from two shortcomings: it is computationally prohibitive (as it requires training thousands of models) and it produces inaccurate scores. The goal of this work is to overcome both these limitations simultaneously. To do so, we take the following approach: \\textbf{First}, we demonstrate that the major source of pLOO's computation bottleneck is its execution on the entire dataset, not just the memorized points. We find running pLOO on all the points is unnecessary since most of them are not even memorized. \\textbf{Second}, we develop a simple proxy to identify the memorized points without having to run pLOO in the first place. To do so, we study the model training cycle and find that memorized points are learned towards the last iterations. We build a simple proxy based on this observation and find that our proxy: \\textit{a)} is strongly correlated with the actual memorization scores (Pearson score $<-0.95$) across all our models and datasets and \\textit{b)} requires only a single model (instead of the thousands needed by pLOO). However, our proxy does not provide the exact memorization scores. \\textbf{Third}, to calculate these, we incorporate our proxy into the pLOO method, resulting in pLOO\\textsubscript{\\textit{improved}}. In doing so, we show that our pLOO\\textsubscript{\\textit{improved}} reduces both computational overhead (by over 90\\%) and the error in the approximated memorization scores (by over 65\\%). Therefore, our work makes it possible to study memorization in large datasets and real-world models while requiring only a fraction of the computational resources.", "keywords": [ "memorization" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "bC8oHmcB4X", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Hong Joo Lee", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Postdoc" }, { "name": "Hyeon Bae Kim", "gender": "Male", "institution": "Kyunghee University", "country": "KR", "position": "MS student" }, { "name": "Seong Tae Kim", "gender": "Male", "institution": "Kyung Hee University", "country": "KR", "position": "Assistant Professor" }, { "name": "Soyoun Won", "gender": "Female", "institution": "Kyung Hee University", "country": "KR", "position": "PhD student" }, { "name": "Yong Hyun Ahn", "gender": "Male", "institution": "Kyung Hee University", "country": "KR", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 18, "n_ref_uni": 31, "n_ref": 99, "n_ref_all": 109, "n_fig": 6, "n_tab": 3, "L_tab": 4805, "n_element_tab": 29, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 228, "n_element_tab_1": 26, "formula_len_all": 646, "formula_len_all_1": 645, "len_all": 113032, "len_all_1": 55030, "len_abs": 1344, "len_title": 115, "len_sents": 31266, "len_sents_1": 25645, "n_sents": 225, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1349, "title": "Steady and Fair Robustness Evaluation Based on Model Interpretation", "abs": "Adversarial robustness has become a major concern as machine learning models are increasingly deployed in security-sensitive applications. Evaluating adversarial robustness remains a challenging task, as current metrics are heavily affected by various factors, including attack methods, attack intensities, and model architecture. In this paper, we propose Steady and Fair Robustness Evaluation, a novel framework designed to mitigate the impact of these factors and provide a more stable evaluation of a model\u2019s robustness. Our key insight is based on the strong correlation between the standard deviation (SD) of Shapley values, which measures the importance of individual neurons, and adversarial robustness. We demonstrate that models with lower SD of Shapley values are more robust to adversarial attacks, regardless of the attack method or model architecture. Extensive experiments across various models, training objectives, and attack scenarios show that our approach offers more consistent and interpretable robustness evaluation. We further introduce a new training strategy that incorporates the minimization of the SD of Shapley values for improving the robustness of the model. Our findings suggest that analysis based on Shapley value can provide a principled and efficient alternative to conventional robustness evaluation techniques.", "keywords": [ "Adversarial robustness", "robustness evaluation", "Shapley value" ], "rating_list": [ 5, 1, 3, 5 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "bBoetBIN2R", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Adrian Hilton", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Professor" }, { "name": "Armin Mustafa", "gender": "Female", "institution": "University of Surrey", "country": "GB", "position": "Lecturer" }, { "name": "Asmar Nadeem", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "PhD student" }, { "name": "Faegheh Sardari", "gender": "Female", "institution": "University of Surrey", "country": "GB", "position": "Postdoc" }, { "name": "Robert Dawes", "gender": "unknown", "institution": "BBC R&D", "country": "GB", "position": "Research Engineer" }, { "name": "Syed Sameed Husain", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Lecturer" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 22, "n_ref": 47, "n_ref_all": 102, "n_fig": 19, "n_tab": 10, "L_tab": 3344, "n_element_tab": 252, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 3074, "n_element_tab_1": 178, "formula_len_all": 617, "formula_len_all_1": 578, "len_all": 211194, "len_all_1": 73506, "len_abs": 5726, "len_title": 122, "len_sents": 69289, "len_sents_1": 35247, "n_sents": 410, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1402, "title": "NarrativeBridge: Enhancing Video Captioning with Causal-Temporal Narrative", "abs": "Existing video captioning benchmarks and models lack causal-temporal narrative, which is sequences of events linked through cause and effect, unfolding over time and driven by characters or agents. This lack of narrative restricts models\u2019 ability to generate text descriptions that capture the causal and temporal dynamics inherent in video content. To address this gap, we propose NarrativeBridge, an approach comprising of: (1) a novel Causal-Temporal Narrative (CTN) captions benchmark generated using a large language model and few-shot prompting, explicitly encoding cause-effect temporal relationships in video descriptions; and (2) a Cause-Effect Network (CEN) with separate encoders for capturing cause and effect dynamics, enabling effective learning and generation of captions with causal-temporal narrative. Extensive experiments demonstrate that CEN significantly outperforms state-of-the-art models in articulating the causal and temporal aspects of video content: 17.88 and 17.44 CIDEr on the MSVD-CTN and MSRVTT-CTN datasets, respectively. Cross-dataset evaluations further showcase CEN\u2019s strong generalization capabilities. The proposed framework understands and generates nuanced text descriptions with intricate causal-temporal narrative structures present in videos, addressing a critical limitation in video captioning. For project details, visit https://narrativebridge.github.io/.", "keywords": [ "Video Captioning", "Video Narrative", "Video Storytelling" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "bBUhlynfRX", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Cengiz Pehlevan", "gender": "unknown", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "Dina Obeid", "gender": "Female", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "US", "position": "Researcher" }, { "name": "Elie Attias", "gender": "Male", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "US", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 14, "n_ref": 55, "n_ref_all": 86, "n_fig": 16, "n_tab": 1, "L_tab": 282, "n_element_tab": 23, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 1058, "n_element_tab_1": 13, "formula_len_all": 278, "formula_len_all_1": 308, "len_all": 112480, "len_all_1": 57662, "len_abs": 1411, "len_title": 103, "len_sents": 38149, "len_sents_1": 26220, "n_sents": 283, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1410, "title": "A Brain-Inspired Regularizer for Adversarial Robustness", "abs": "Convolutional Neural Networks (CNNs) excel in many visual tasks, but they tend to be sensitive to slight input perturbations that are imperceptible to the human eye, often resulting in task failures. Recent studies indicate that training CNNs with regularizers that promote brain-like representations, using neural recordings, can improve model robustness. However, the requirement to use neural data severely restricts the utility of these methods. Is it possible to develop regularizers that mimic the computational function of neural regularizers without the need for neural recordings, thereby expanding the usability and effectiveness of these techniques? In this work, we inspect a neural regularizer introduced in Li et al. to extract its underlying strength. The regularizer uses neural representational similarities, which we find also correlate with pixel similarities. Motivated by this finding, we introduce a new regularizer that retains the essence of the original but is computed using image pixel similarities, eliminating the need for neural recordings. We show that our regularization method 1) significantly increases model robustness against a variety of black box attacks, 2) relies only on original, unaugmented datasets and 3) is computationally inexpensive. Our work explores how biologically motivated loss functions can be used to drive the performance of artificial neural networks.", "keywords": [ "Neuroscience", "Machine Learning", "CNN", "Adversarial Attacks", "Image Classification", "Brain-Inspired" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 1, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 5, 3, 3 ] }, { "paper_id": "bBNUiErs26", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Hyeryung Jang", "gender": "Female", "institution": "Dongguk University", "country": "", "position": "Assistant Professor" }, { "name": "Junseo Park", "gender": "Male", "institution": "Dongguk University", "country": "KR", "position": "MS student" } ], "n_formula": 24, "n_formula_1": 11, "n_ref_uni": 32, "n_ref": 55, "n_ref_all": 92, "n_fig": 23, "n_tab": 3, "L_tab": 574, "n_element_tab": 53, "n_fig_1": 15, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1135, "formula_len_all_1": 508, "len_all": 182557, "len_all_1": 66189, "len_abs": 1342, "len_title": 131, "len_sents": 56901, "len_sents_1": 31461, "n_sents": 435, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 98, "L_abs": 1426, "title": "$\\text{I}^2\\text{AM}$: Interpreting Image-to-Image Latent Diffusion Models via Bi-Attribution Maps", "abs": "Large-scale diffusion models have made significant advances in image generation, particularly through cross-attention mechanisms. While cross-attention has been well-studied in text-to-image tasks, their interpretability in image-to-image (I2I) diffusion models remains underexplored. This paper introduces Image-to-Image Attribution Maps $(\\textbf{I}^2\\textbf{AM})$, a method that enhances the interpretability of I2I models by visualizing bidirectional attribution maps, from the reference image to the generated image and vice versa. $\\text{I}^2\\text{AM}$ aggregates cross-attention scores across time steps, attention heads, and layers, offering insights into how critical features are transferred between images. We demonstrate the effectiveness of $\\text{I}^2\\text{AM}$ across object detection, inpainting, and super-resolution tasks. Our results demonstrate that $\\text{I}^2\\text{AM}$ successfully identifies key regions responsible for generating the output, even in complex scenes. Additionally, we introduce the Inpainting Mask Attention Consistency Score (IMACS) as a novel evaluation metric to assess the alignment between attribution maps and inpainting masks, which correlates strongly with existing performance metrics. Through extensive experiments, we show that $\\text{I}^2\\text{AM}$ enables model debugging and refinement, providing practical tools for improving I2I model's performance and interpretability.", "keywords": [ "Image-to-Image Diffusion Model", "Interpretable AI", "Explainable AI", "Image Attribution" ], "rating_list": [ 5, 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 5, 3, 3, 3, 4 ] }, { "paper_id": "bB0OKNpznp", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Chao-Han Huck Yang", "gender": "Male", "institution": "NVIDIA Research", "country": "TW", "position": "Researcher" }, { "name": "Chen-Yu Liu", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "PhD student" }, { "name": "Hsi-Sheng Goan", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "Full Professor" }, { "name": "Min-Hsiu Hsieh", "gender": "unknown", "institution": "National Taiwan University", "country": "", "position": "Principal Researcher" } ], "n_formula": 21, "n_formula_1": 13, "n_ref_uni": 49, "n_ref": 100, "n_ref_all": 129, "n_fig": 11, "n_tab": 7, "L_tab": 1481, "n_element_tab": 265, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1, "n_element_tab_1": 1, "formula_len_all": 1310, "formula_len_all_1": 752, "len_all": 176751, "len_all_1": 61464, "len_abs": 1720, "len_title": 129, "len_sents": 52799, "len_sents_1": 30443, "n_sents": 364, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1739, "title": "A Quantum Circuit-Based Compression Perspective for Parameter-Efficient Learning", "abs": "Quantum-centric supercomputing presents a compelling framework for large-scale hybrid quantum-classical tasks. Although quantum machine learning (QML) offers theoretical benefits in various applications, challenges such as large-size data encoding in the input stage and the reliance on quantum resources in the inference stage limit its practicality for tasks like fine-tuning large language models (LLMs). Quantum parameter generation, a novel approach of QML, addresses these limitations by using quantum neural networks (QNNs) to generate classical model weights (parameters) exclusively during training, thereby decoupling inference from quantum hardware. In this work, we introduce Quantum Parameter Adaptation (QPA) in the framework of quantum parameter generation, which integrates QNNs with a classical multi-layer perceptron mapping model to generate parameters for fine-tuning methods. Using Gemma-2 and GPT-2 as case studies, QPA demonstrates significant parameter reduction for parameter-efficient fine-tuning methods, such as Low-Rank Adaptation (LoRA), while maintaining comparable or improved performance in text generation tasks. Specifically, QPA reduces the number of parameters to $52.06\\%$ of the original LoRA for GPT-2 with a slight performance gain of $0.75\\%$, and to $16.84\\%$ for Gemma-2, with a marginal performance improvement of $0.07\\%$. These results highlight QPA\u2019s ability to achieve efficient parameter reduction without sacrificing performance in the quantum parameter generation framework. This work showcases the potential of quantum-enhanced parameter reduction, offering a scalable quantum-classical solution for fine-tuning LLMs while preserving the feasibility of inference on classical hardware.", "keywords": [ "quantum machine learning", "quantum computing", "parameter-efficient fine-tuning", "large language model", "variational quantum circuits" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "bAdSmSR10C", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chao Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Associate Professor" }, { "name": "Eng Siong Chng", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Helin Wang", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "PhD student" }, { "name": "Wen Wu", "gender": "Female", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yuchen Hu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "CHEN CHEN", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 63, "n_ref": 102, "n_ref_all": 132, "n_fig": 7, "n_tab": 4, "L_tab": 680, "n_element_tab": 118, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1829, "n_element_tab_1": 56, "formula_len_all": 1238, "formula_len_all_1": 1152, "len_all": 190203, "len_all_1": 65090, "len_abs": 1498, "len_title": 64, "len_sents": 53383, "len_sents_1": 29902, "n_sents": 402, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1492, "title": "Enhancing Zero-shot Text-to-Speech Synthesis with Human Feedback", "abs": "In recent years, text-to-speech (TTS) technology has witnessed impressive advancements, particularly with large-scale training datasets, showcasing human-level speech quality and impressive zero-shot capabilities on unseen speakers. However, despite human subjective evaluations, such as the mean opinion score (MOS), remaining the gold standard for assessing the quality of synthetic speech, even state-of-the-art TTS approaches have kept human feedback isolated from training that resulted in mismatched training objectives and evaluation metrics. In this work, we investigate a novel topic of integrating subjective human evaluation into the TTS training loop. Inspired by the recent success of reinforcement learning from human feedback, we propose a comprehensive sampling-annotating-learning framework tailored to TTS optimization, namely uncertainty-aware optimization (UNO). Specifically, UNO eliminates the need for a reward model or preference data by directly maximizing the utility of speech generations while considering the uncertainty that lies in the inherent variability in subjective human speech perception and evaluations. Experimental results of both subjective and objective evaluations demonstrate that UNO considerably improves the zero-shot performance of TTS models in terms of MOS, word error rate, and speaker similarity. Additionally, we present a remarkable ability of UNO that it can adapt to the desired speaking style in emotional TTS seamlessly and flexibly.", "keywords": [ "Text-to-speech synthesis", "learning from human feedback", "reinforcement learning", "audio generation" ], "rating_list": [ 5, 6, 3, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 4 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "bAFVlpFQvT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Baharan Mirzasoleiman", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "Dang Nguyen", "gender": "Male", "institution": "Google", "country": "US", "position": "Intern" }, { "name": "Rathul Anand", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Undergrad student" }, { "name": "Wenhan Yang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Yu Yang", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 42, "n_formula_1": 16, "n_ref_uni": 38, "n_ref": 80, "n_ref_all": 110, "n_fig": 8, "n_tab": 6, "L_tab": 1782, "n_element_tab": 213, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 3930, "n_element_tab_1": 97, "formula_len_all": 3122, "formula_len_all_1": 1390, "len_all": 172528, "len_all_1": 68086, "len_abs": 1602, "len_title": 122, "len_sents": 50043, "len_sents_1": 26729, "n_sents": 474, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1568, "title": "Memory-efficient Training of Large Language Models with Larger Mini-batches", "abs": "Training with larger mini-batches improves the convergence rate and can yield superior performance. However, training with large mini-batches becomes prohibitive for Large Language Models (LLMs), due to the large GPU memory requirement. To address this problem, an effective approach is finding small mini-batch coresets that closely match the gradient of larger mini-batches. However, this approach becomes infeasible and ineffective for LLMs, due to the highly imbalanced nature of the sources in language data, use of the Adam optimizer, and the very large gradient dimensionality of LLMs. In this work, we address the above challenges by proposing *Coresets for Training LLMs* (CoLM). First, we show that mini-batch coresets found by gradient matching do not contain representative examples of the small sources w.h.p., and thus including all examples of the small sources in the mini-batch coresets is crucial for optimal performance. Second, we normalize the gradients by their historical exponential to find mini-batch coresets for training with Adam. Finally, we leverage zeroth-order methods to find smooth gradient of the last *V*-projection matrix and sparsify it to keep the dimensions with the largest normalized gradient magnitude. We apply CoLM to fine-tuning Phi-2, Phi-3, and Zephyr with LoRA on MathInstruct and SuperGLUE benchmark. Remarkably, CoLM reduces the memory requirement of fine-tuning by 2x and even outperforms training with 4x larger mini-batches. Notably, CoLM easily stack with existing memory-efficient training methods, such as LoRA.", "keywords": [ "Data selection", "Memory efficient", "Large Language Models" ], "rating_list": [ 6, 8, 5, 8 ], "soundness_list": [ 3, 4, 3, 4 ], "presentation_list": [ 3, 4, 2, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "b9w9b6naQG", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "He Zhao", "gender": "Male", "institution": "Commonwealth Scientific and Industrial Research Organisation, CSIRO", "country": "AU", "position": "Researcher" }, { "name": "Xuesong Wang", "gender": "Male", "institution": "CSIRO", "country": "", "position": "Postdoc" }, { "name": "Edwin V Bonilla", "gender": "Male", "institution": "CSIRO's Data61", "country": "", "position": "Principal Research Scientist" } ], "n_formula": 31, "n_formula_1": 23, "n_ref_uni": 47, "n_ref": 94, "n_ref_all": 106, "n_fig": 1, "n_tab": 6, "L_tab": 3920, "n_element_tab": 276, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1330, "n_element_tab_1": 132, "formula_len_all": 3939, "formula_len_all_1": 2149, "len_all": 140951, "len_all_1": 73968, "len_abs": 1546, "len_title": 70, "len_sents": 38249, "len_sents_1": 31591, "n_sents": 289, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Reject", "L_title": 22, "L_abs": 1575, "title": "R\u00e9nyi Neural Processes", "abs": "Neural Processes (NPs) are deep probabilistic models that represent stochastic processes by conditioning their prior distributions on a set of context points. Despite their obvious advantages in uncertainty estimation for complex distributions, NPs enforce parameterization coupling between the conditional prior model and the posterior model, thereby risking introducing a misspecified prior distribution. We hereby revisit the NP objectives and propose R\u00e9nyi Neural Processes (RNP) to ameliorate the impacts of prior misspecification by optimizing an alternative posterior that achieves better marginal likelihood. More specifically, by replacing the standard KL divergence with the R\u00e9nyi divergence between the model posterior and the true posterior, we scale the density ratio $\\frac{p}{q}$ by the power of (1-$\\alpha$) in the divergence gradients with respect to the posterior. This hyper parameter $\\alpha$ allows us to dampen the effects of the misspecified prior for the posterior update, which has been shown to effectively avoid oversmoothed predictions and improve the expressiveness of the posterior model.\nOur extensive experiments show consistent log-likelihood improvements over state-of-the-art NP family models which adopt both the variational inference or maximum likelihood estimation objectives. We validate the effectiveness of our approach across multiple benchmarks including regression and image inpainting tasks, and show significant performance improvements of RNPs in real-world regression problems where the underlying prior model is misspecifed.", "keywords": [ "Neural processes", "variational inference", "meta learning", "robust divergence" ], "rating_list": [ 6, 8, 3, 3 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "b9rpcQxm5P", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Benjamin Zhang", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Postdoc" }, { "name": "Markos Katsoulakis", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "", "position": "Full Professor" }, { "name": "Ziyu Chen", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "", "position": "Postdoc" } ], "n_formula": 94, "n_formula_1": 46, "n_ref_uni": 25, "n_ref": 73, "n_ref_all": 97, "n_fig": 2, "n_tab": 1, "L_tab": 298, "n_element_tab": 24, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8170, "formula_len_all_1": 3822, "len_all": 131872, "len_all_1": 64516, "len_abs": 1809, "len_title": 146, "len_sents": 39655, "len_sents_1": 27583, "n_sents": 343, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1836, "title": "Equivariant score-based generative models provably learn distributions with symmetries efficiently", "abs": "Symmetry is ubiquitous in many real-world phenomena and tasks, such as physics, images, and molecular simulations. Empirical studies have demonstrated that incorporating symmetries into generative models can provide better generalization and sampling efficiency when the underlying data distribution has group symmetry. In this work, we provide the first theoretical analysis and guarantees of score-based generative models (SGMs) for learning distributions that are invariant with respect to some group symmetry and offer the first quantitative comparison between data augmentation and adding equivariant inductive bias. First, building on recent works on the Wasserstein-1 ($\\mathbf{d}_1$) guarantees of SGMs and empirical estimations of probability divergences under group symmetry, we provide an improved $\\mathbf{d}_1$ generalization bound when the data distribution is group-invariant. Second, we describe the inductive bias of equivariant SGMs using Hamilton-Jacobi-Bellman theory, and rigorously demonstrate that one can learn the score of a symmetrized distribution using equivariant vector fields without data augmentations through the analysis of the optimality and equivalence of score-matching objectives. This also provides practical guidance that one does not have to augment the dataset as long as the vector field or the neural network parametrization is equivariant. Moreover, we quantify the impact of not incorporating equivariant structure into the score parametrization, by showing that non-equivariant vector fields can yield worse generalization bounds. This can be viewed as a type of model-form error that describes the missing structure of non-equivariant vector fields. Numerical simulations corroborate our analysis and highlight that data augmentations cannot replace the role of equivariant vector fields.", "keywords": [ "score-based generative models", "group symmetry", "data augmentation", "generalization error" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "b9qIPrOfCw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Jungong Han", "gender": "Male", "institution": "The University of Sheffield", "country": "GB", "position": "Full Professor" }, { "name": "Tao Xu", "gender": "Male", "institution": "Shanghai Investigation Design and Research Institute Co., Ltd. (SIDRI)", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhenkun Fan", "gender": "Male", "institution": "University of Wales, Aberystwyth", "country": "GB", "position": "PhD student" }, { "name": "zx H", "gender": "Male", "institution": "University of Wales, Aberystwyth", "country": "GB", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 12, "n_ref": 23, "n_ref_all": 36, "n_fig": 3, "n_tab": 6, "L_tab": 3484, "n_element_tab": 427, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1520, "n_element_tab_1": 175, "formula_len_all": 461, "formula_len_all_1": 462, "len_all": 106690, "len_all_1": 52825, "len_abs": 1591, "len_title": 92, "len_sents": 26871, "len_sents_1": 23864, "n_sents": 189, "n_sents_1": 161, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 1597, "title": "Continuity-Driven Pose Estimation for Videos", "abs": "Video-based pose estimation plays a critical role in understanding human actions and enabling effective human-computer interaction. By exploiting temporal information from video frames, it enhances the localization of human keypoints. Previous feature-fusion methods often rely on a frozen single-frame backbone trained on individual frames, followed by a network to learn temporal information from video sequences. Consequently, these approaches fail to capture the temporal continuity between frames at the backbone network level, thereby restricting the network's capacity to effectively learn and leverage sequential information. In this paper, we introduce a novel approach to supervise continuity in the whole video pose estimation model from two perspectives: semantic continuity and pixel-wise keypoint distribution continuity. To this end, we propose a Semantic Alignment Space, where a semantic alignment encodes feature maps from different frames into this space, ensuring continuous supervision of the encoded representations. To further maintain pixel-wise keypoint distribution continuity, we introduce the Trajectory Probability Difference Integration method, which minimizes the trajectory difference expectation across frames. Additionally, to better capture temporal dependencies, we present a Multi-frame Heatmap Fusion structure that aggregates heatmaps from adjacent frames for a more refined output. Extensive experiments on the PoseTrack17, PoseTrack18, and PoseTrack21 datasets demonstrate the effectiveness of our approach, consistently achieving state-of-the-art results.", "keywords": [ "video pose estiamtion" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "b9dBNNeDd3", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Le Zhuo", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Sheng Xia", "gender": "unknown", "institution": "nanjing university", "country": "CN", "position": "PhD student" }, { "name": "Wenze Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xiangyu Yue", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Yi Xin", "gender": "Male", "institution": "Nanjing university", "country": "CN", "position": "PhD student" }, { "name": "Gao Peng", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 41, "n_ref": 97, "n_ref_all": 138, "n_fig": 12, "n_tab": 5, "L_tab": 2359, "n_element_tab": 322, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1764, "n_element_tab_1": 267, "formula_len_all": 205, "formula_len_all_1": 208, "len_all": 130599, "len_all_1": 57502, "len_abs": 1340, "len_title": 124, "len_sents": 30152, "len_sents_1": 24050, "n_sents": 253, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1346, "title": "Customize Your Visual Autoregressive Recipe with Set Autoregressive Modeling", "abs": "We introduce a new paradigm for AutoRegressive (AR) image generation, termed Set AutoRegressive Modeling (SAR). SAR generalizes the conventional AR to the next-set setting, i.e., splitting the sequence into arbitrary sets containing multiple tokens, rather than outputting each token in a fixed raster order. To accommodate SAR, we develop a straightforward architecture termed Fully Masked Transformer. We reveal that existing AR variants correspond to specific design choices of sequence order and output intervals within the SAR framework, with AR and Masked AR (MAR) as two extreme instances. Notably, SAR facilitates a seamless transition from AR to MAR, where intermediate states allow for training a causal model that benefits from both few-step inference and KV cache acceleration, thus leveraging the advantages of both AR and MAR. On the ImageNet benchmark, we carefully explore the properties of SAR by analyzing the impact of sequence order and output intervals on performance, as well as the generalization ability regarding inference order and steps. We further validate the potential of SAR by training a 900M text-to-image model capable of synthesizing photo-realistic images with any resolution. We hope our work may inspire more exploration and application of AR-based modeling across diverse modalities. Code will be available.", "keywords": [ "generatigve modeling", "image generation", "autoregressive models" ], "rating_list": [ 3, 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3, 4 ], "presentation_list": [ 2, 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 3, 4 ] }, { "paper_id": "b9ZG7cI8ic", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Bharath B N", "gender": "Male", "institution": "Indian Institute of Technology Dharwad", "country": "IN", "position": "Assistant Professor" }, { "name": "Prashant Khanduri", "gender": "Male", "institution": "Wayne State University", "country": "US", "position": "Assistant Professor" }, { "name": "Shruti P Maralappanavar", "gender": "Female", "institution": "Indian Institute of Technology, Dharwad", "country": "IN", "position": "PhD student" }, { "name": "Sumit Sah", "gender": "Male", "institution": "Indian Institute of Technology, Dharwad", "country": "IN", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 26, "n_ref_uni": 25, "n_ref": 63, "n_ref_all": 70, "n_fig": 2, "n_tab": 1, "L_tab": 1360, "n_element_tab": 18, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 2091, "n_element_tab_1": 25, "formula_len_all": 2776, "formula_len_all_1": 2448, "len_all": 121468, "len_all_1": 72368, "len_abs": 1247, "len_title": 172, "len_sents": 30721, "len_sents_1": 27504, "n_sents": 263, "n_sents_1": 240, "venue": "ICLR2025", "decision": "withdraw", "L_title": 124, "L_abs": 1353, "title": "Generalization of FedAvg Under Constrained Polyak-Lojasiewicz Type Conditions: A Single Hidden Layer Neural Network Analysis", "abs": "In this work, we study the optimization and the generalization performance of the widely used FedAvg algorithm for solving Federated Learning (FL) problems. We analyze the generalization performance of FedAvg by handling the optimization error and the Rademacher complexity. Towards handling optimization error, we propose novel constrained Polyak-Lojasiewicz (PL)-type conditions on the objective function that ensure the existence of a global optimal to which FedAvg converges linearly after $\\mathcal{O}( \\log ({1}/{\\epsilon}))$ rounds of communication, where $\\epsilon$ is the desired optimality gap. Importantly, we demonstrate that a class of single hidden layer neural networks satisfies the proposed constrained PL-type conditions required to establish the linear convergence of FedAvg as long as $m > {nK}/{d}$, where $m$ is the width of the neural network, $K$ is the number of clients, $n$ is the number of samples at each client, and $d$ is the feature dimension. We then bound the Rademacher complexity for this class of neural networks and establish that both Rademacher complexity and the generalization error of FedAvg decrease at an optimal rate of $\\mathcal{O}({1}/{\\sqrt{n}})$. We further show that increasing the number of clients $K$ decreases the generalization error at the rate of $\\mathcal{O}({1}/{\\sqrt{n}} + {1}/{\\sqrt{nK}})$.", "keywords": [ "FedAvg", "Linear Convergence", "Generalization", "Neural Network" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "b9VSMQZl0j", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Chunyan Miao", "gender": "Female", "institution": "School of Computer Science and Engineering, Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Haocong Rao", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 25, "n_ref": 75, "n_ref_all": 97, "n_fig": 3, "n_tab": 5, "L_tab": 1696, "n_element_tab": 205, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 9327, "n_element_tab_1": 386, "formula_len_all": 923, "formula_len_all_1": 794, "len_all": 122606, "len_all_1": 80240, "len_abs": 1845, "len_title": 176, "len_sents": 42570, "len_sents_1": 32183, "n_sents": 244, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 128, "L_abs": 1905, "title": "General Skeleton Semantics Learning with Probabilistic Masked Context Reconstruction for Skeleton-Based Person Re-Identification", "abs": "Person re-identification (re-ID) via skeleton data is an emerging topic with immense potential for safety-critical applications. Existing methods usually utilize spatial or temporal skeleton semantics learning (SSL) tasks to facilitate skeleton representation learning, while most SSL tasks are *model-dependent* and lack the ability to capture general fine-grained (*e.g.*, joint-level) spatial-temporal skeleton patterns under different model architectures. To delve into multi-faceted generality of SSL tasks, we first propose an SSL generality assessment framework termed **SCUT** that identifies four key SSL properties: **S**patial-temporal effectiveness, **C**o-training compatibility, **U**nsupervised trainability, and **T**ask transformability. By formulating systematic evaluation criteria for each property, SCUT enables both qualitative and quantitative analysis of SSL generality under varying models and scenarios. Motivated by SCUT to fully harness skeleton context for semantics learning, we further devise a generic **Pro**babilistic **M**asked S**p**atial-**T**emporal cont**e**xt **R**econstruction (**Prompter**) task to enhance performance of skeleton-based person re-ID models. Specifically, Prompter first probabilistically and independently masks joints' structural locations to generate *spatial context*, and then randomly conceal their motion trajectories to form *temporal context*. \nThrough combining both spatial and temporal skeleton context representations to jointly reconstruct and infer skeleton sequences, Prompter encourages the model to capture general valuable spatial-temporal skeleton patterns for person re-ID. Empirical evaluations on SCUT and five benchmark datasets demonstrate the superiority of Prompter to most state-of-the-art SSL tasks. We further validate its general effectiveness in different skeleton modeling, RGB-estimated or cross-domain scenarios", "keywords": [ "General skeleton semantics learning", "Generality Assessment", "Skeleton-based person re-identification", "Probabilistic masked reconstruction", "Spatial-temporal context learning" ], "rating_list": [ 8, 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 4, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 4, 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 3, 3, 3 ] }, { "paper_id": "b9Ne5lHJ8Y", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Yue Wang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Daniel Takeshi Seita", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "Assistant Professor" }, { "name": "Jonathan Zamora-Anaya", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 88, "n_ref_all": 98, "n_fig": 3, "n_tab": 2, "L_tab": 388, "n_element_tab": 42, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1856, "n_element_tab_1": 57, "formula_len_all": 302, "formula_len_all_1": 0, "len_all": 160161, "len_all_1": 50983, "len_abs": 1216, "len_title": 128, "len_sents": 39333, "len_sents_1": 23557, "n_sents": 316, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1231, "title": "MuJoCo Manipulus: A Robot Learning Benchmark for Generalizable Tool Manipulation", "abs": "We propose MuJoCo Manipulus, a novel open-source benchmark powered by the MuJoCo physics simulation engine, designed to accelerate advances in robot learning for tool manipulation. Our benchmark includes a diverse set of tasks for tool manipulation --- a domain where the field currently lacks a unified benchmark. Different research groups rely on custom-designed tasks or closed-source setups, limiting cross-comparability and hindering significant progress in this field. To that end, our benchmark provides 16 challenging tool manipulation tasks, including variants of Pouring, Scooping, Scraping, Stacking, Gathering, Hammering, Mini-Golf, and Ping-Pong. The benchmark supports both state-based and vision-based observation spaces, is fully integrated with the Gymnasium API, and seamlessly connects with widely used Deep Reinforcement Learning libraries, ensuring easy adoption by the community. We conduct extensive reinforcement learning experiments on our benchmark, and our results demonstrate that there is substantial progress to be made for training tool manipulation policies. Our codebase and additional videos of the learned policies can be found on our anonymous project website: https://mujoco-manipulus.github.io", "keywords": [ "robotics", "deep reinforcement learning", "benchmark" ], "rating_list": [ 3, 3, 3, 5, 3 ], "soundness_list": [ 2, 1, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 4, 2 ], "contribution_list": [ 2, 1, 2, 2, 1 ], "confidence_list": [ 4, 4, 5, 4, 5 ] }, { "paper_id": "b8eEutZlPb", "primary_area": "generative models", "n_author": 20, "author_info_list": [ { "name": "Boyang Hong", "gender": "Male", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Chenyang Liao", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Dingwen Yang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Honglin Guo", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Junzhe Wang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Lu Chen", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Qi Zhang", "gender": "Male", "institution": "Fudan University", "country": "", "position": "Full Professor" }, { "name": "Rui Zheng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Songyang Gao", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Tao Gui", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Assistant Professor" }, { "name": "Wei He", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Wenxiang Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Xin Guo", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Xipeng Qiu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Xuanjing Huang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yicheng Zou", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Yiwen Ding", "gender": "Female", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Yu-Gang Jiang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Zhiheng Xi", "gender": "unknown", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Zuxuan Wu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 39, "n_ref": 182, "n_ref_all": 223, "n_fig": 10, "n_tab": 19, "L_tab": 6704, "n_element_tab": 683, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1507, "n_element_tab_1": 148, "formula_len_all": 838, "formula_len_all_1": 653, "len_all": 327182, "len_all_1": 70900, "len_abs": 1833, "len_title": 143, "len_sents": 82665, "len_sents_1": 31524, "n_sents": 759, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1570, "title": "AgentGym: Evaluating and Evolving Large Language Model-based Agents across Diverse Envronments", "abs": "Large language models (LLMs), with their generalized capabilities, are considered as a promising foundation to build generally-capable agents that can handle multi-turn decision-making tasks across various interactive environments. Previous attempts typically gather expert-provided trajectories and have LLM-based agents imitate these trajectories step-by-step. However, this supervised fine-tuning approach depends heavily on human supervision, limiting scalability and restricting the agent's exploration and learning in the environments. In this paper, we take the first step towards developing generally-capable LLM-based agents that can explore and evolve themselves across diverse environments. To achieve this, we identify a trinity of ingredients: 1) diverse interactive environments for agent exploration, 2) a trajectory set to equip agents with basic capabilities and prior knowledge, and 3) an effective and scalable approach for agent improvement across environments. We propose AgentGym, a new interactive framework featuring various real-world scenarios and environments for broad, unified, real-time, and concurrent agent exploration. AgentGym also includes a database with expanded instructions, high-quality trajectories, and a benchmark suite. Next, we investigate the potential of agent self-evolution across various environments with a derived exploration-learning method named AgentEvol. Experimental results show that the evolved agents can achieve results comparable to SOTA models. We will release the code, dataset, benchmark, and checkpoints.", "keywords": [ "large language model", "LLM-based agent", "self-improvement", "evaluation" ], "rating_list": [ 5, 5, 5, 8 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 4 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "b89OyrljJD", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Huaxiu Yao", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Liao Chang", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Wenhao Zheng", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Ying Wei", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 68, "n_ref_all": 114, "n_fig": 12, "n_tab": 7, "L_tab": 1516, "n_element_tab": 149, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 587, "n_element_tab_1": 39, "formula_len_all": 1039, "formula_len_all_1": 802, "len_all": 152308, "len_all_1": 66572, "len_abs": 1102, "len_title": 141, "len_sents": 45972, "len_sents_1": 32405, "n_sents": 309, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1110, "title": "How Well Can LLMs Synthesize Molecules? An LLM-Powered Framework for Multi-Step Retrosynthesis", "abs": "Predicting retrosynthesis routes is a fundamental challenge in chemistry, involving the design of a sequence of chemical reactions to synthesize a target molecule from commercially available starting materials. With a rapidly\ngrowing interest in using large language models for planning, this work introduces an LLM-powered framework for multi-step retrosynthesis. Our framework employs molecular-similarity-based retrieval-augmented generation (RAG) to generate an initial retrosynthesis route, which is then iteratively refined through expert feedback. The use of molecular-similarity-based RAG improves reaction round-trip validity from 24.42\\% to 51.64\\% compared to GPT-4 with representative routes. With further refinement, the validity increases to 89.81\\%, resulting in an overall route validity of 79.5\\% with a perfect query success rate, comparable to traditional methods. Our framework offers a flexible, customizable approach to retrosynthesis, and we present a comprehensive analysis of the generated routes along with promising future research directions in LLM-driven multi-step retrosynthesis.", "keywords": [ "Chemistry", "Retrosynthesis Planning", "LLM" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 2, 1 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "b87H1A3sxm", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "EN-HUI YANG", "gender": "unknown", "institution": "University of Waterloo", "country": "CA", "position": "Full Professor" }, { "name": "Renhao Tan", "gender": "unknown", "institution": "University of Waterloo", "country": "", "position": "MS student" } ], "n_formula": 59, "n_formula_1": 29, "n_ref_uni": 37, "n_ref": 92, "n_ref_all": 124, "n_fig": 10, "n_tab": 11, "L_tab": 4827, "n_element_tab": 530, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 269, "n_element_tab_1": 31, "formula_len_all": 6764, "formula_len_all_1": 2779, "len_all": 172055, "len_all_1": 60806, "len_abs": 1263, "len_title": 70, "len_sents": 51002, "len_sents_1": 27050, "n_sents": 326, "n_sents_1": 164, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1169, "title": "Enhancing Adversarial Robustness Through Robust Information Quantities", "abs": "It is known that deep neural networks (DNNs) are vulnerable to imperceptible adversarial attacks, and this fact raises concerns about their safety and reliability in real-world applications. In this paper, we aim to boost the robustness of a DNN against white-box adversarial attacks by defining three new information quantities---robust conditional mutual information (CMI), robust separation, and robust normalized CMI (NCMI)---which can serve as robust performance metrics for the DNN. We then utilize these concepts to introduce a novel training method that constrains the robust CMI and increases the robust separation simultaneously. Our experimental results demonstrate that our method consistently enhances model robustness against C\\&W and AutoAttack on CIFAR and Tiny-ImageNet datasets with and without additional synthetic data. Specifically, it is shown that our approach improves the robust accuracy of a DNN by up to 2.66\\% on CIFAR datasets and 3.49\\% on Tiny-ImageNet in the case of PGD attack and 1.70\\% on CIFAR datasets and 1.63\\% on Tiny-ImageNet in the case of AutoAttack, in comparison with the state-of-the-art training methods in the literature.", "keywords": [ "Adversarial Robustness", "Adversarial Training" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "b7ROBvgNkE", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Chao Cui", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Fan Li", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Hechang Chen", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Associate Professor" }, { "name": "Huafeng Cao", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "MS student" }, { "name": "Jiechao Gao", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Shisong Tang", "gender": "unknown", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" } ], "n_formula": 43, "n_formula_1": 16, "n_ref_uni": 51, "n_ref": 100, "n_ref_all": 112, "n_fig": 7, "n_tab": 10, "L_tab": 4421, "n_element_tab": 339, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 431, "n_element_tab_1": 33, "formula_len_all": 2022, "formula_len_all_1": 1208, "len_all": 171348, "len_all_1": 60790, "len_abs": 1990, "len_title": 122, "len_sents": 52265, "len_sents_1": 29490, "n_sents": 400, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1843, "title": "Calibrating Video Watch-time Predictions with Credible Prototype Alignment", "abs": "Accurately predicting user watch-time is crucial for enhancing user stickiness and retention in video recommendation systems. Existing watch-time prediction approaches typically involve transformations of watch-time labels for prediction and subsequent reversal, ignoring both the natural distribution properties of label and the \\textit{instance representation confusion} that results in inaccurate predictions. \nIn this paper, we propose ProWTP, a two-stage method combining prototype learning and optimal transport for watch-time regression prediction, suitable for any deep recommendation model. The core idea of ProWTP is to align label distribution with instance representation distribution to calibrate the instance space, thereby improving prediction accuracy. Specifically, we observe that the watch-ratio (the ratio of watch-time to video duration) within the same duration bucket exhibits a multimodal distribution. To facilitate incorporation into models, we use a hierarchical vector quantised variational autoencoder (HVQ-VAE) to convert the continuous label distribution into a high-dimensional discrete distribution, serving as credible prototypes for calibrations. Based on this, ProWTP views the alignment between prototypes and instance representations as a Semi-relaxed Unbalanced Optimal Transport (SUOT) problem, where the marginal constraints of prototypes are relaxed. And the corresponding optimization problem is reformulated as a weighted Lasso problem for solution. Moreover, ProWTP introduces the assignment and compactness losses to encourage instances to cluster closely around their respective prototypes, thereby enhancing the prototype-level distinguishability. Finally, we conducted extensive offline experiments on two industrial datasets, demonstrating our consistent superiority in real-world application.", "keywords": [ "Prototype learning", "optimal transport", "recommendation" ], "rating_list": [ 6, 6, 8, 5 ], "soundness_list": [ 2, 2, 4, 2 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "b7HOhqXiZs", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Bowen Peng", "gender": "unknown", "institution": "Nous Research", "country": "US", "position": "Principal Researcher" }, { "name": "Diederik P Kingma", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Jeffrey Quesnelle", "gender": "unknown", "institution": "University of Michigan - Dearborn", "country": "", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 12, "n_ref": 14, "n_ref_all": 22, "n_fig": 1, "n_tab": 2, "L_tab": 568, "n_element_tab": 206, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 568, "n_element_tab_1": 206, "formula_len_all": 470, "formula_len_all_1": 470, "len_all": 60853, "len_all_1": 60853, "len_abs": 1230, "len_title": 85, "len_sents": 18333, "len_sents_1": 18333, "n_sents": 119, "n_sents_1": 119, "venue": "ICLR2025", "decision": "withdraw", "L_title": 37, "L_abs": 1211, "title": "DeMo: Decoupled Momentum Optimization", "abs": "Training large scale neural networks typically involves sharing the gradients between all accelerators, which necessitates specialized high-speed interconnects. Taking cues from signal processing, we show that it is not necessary to share or synchronize the full optimizer states and model parameters during training. By decoupling the momentum and allowing divergence in the optimizer states across accelerators, it is possible to even improve convergence compared to previous state of the art optimizers.\nFrom this, we introduce a Decoupled Momentum optimization algorithm (DeMo) that reduces the communication requirements by several orders of magnitude, potentially enabling future training of large neural networks on slow internet bandwidths with heterogeneous networking hardware. Furthermore, our method is agnostic to the network topology and neural network architecture, and supports scalable clock-synchronous distributed training with negligible compute and memory overhead.\nEmpirically, we show that models trained with DeMo match or surpass the performance of equal models trained with AdamW, entirely bypassing the need for high-speed interconnects for pre-training large scale foundation models.", "keywords": [ "deep learning", "large language models", "optimization", "training", "generative models", "pre-training", "foundational models", "distributed training" ], "rating_list": [ 3, 3, 3, 3, 1 ], "soundness_list": [ 2, 1, 2, 1, 1 ], "presentation_list": [ 2, 3, 1, 3, 3 ], "contribution_list": [ 2, 1, 2, 2, 1 ], "confidence_list": [ 4, 3, 4, 2, 4 ] }, { "paper_id": "b77ML9nALL", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Chaouki Ben Issaid", "gender": "Male", "institution": "University of Oulu", "country": "FI", "position": "Principal Researcher" }, { "name": "Mehdi Bennis", "gender": "Male", "institution": "University of Oulu", "country": "", "position": "Full Professor" }, { "name": "Praneeth Vepakomma", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" } ], "n_formula": 68, "n_formula_1": 15, "n_ref_uni": 38, "n_ref": 92, "n_ref_all": 112, "n_fig": 4, "n_tab": 6, "L_tab": 1959, "n_element_tab": 152, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 573, "n_element_tab_1": 6, "formula_len_all": 4572, "formula_len_all_1": 970, "len_all": 184196, "len_all_1": 59416, "len_abs": 1154, "len_title": 154, "len_sents": 61325, "len_sents_1": 27425, "n_sents": 509, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1160, "title": "Tackling Feature and Sample Heterogeneity in Decentralized Multi-Task Learning: A Sheaf-Theoretic Approach", "abs": "Federated multi-task learning (FMTL) aims to simultaneously learn multiple related tasks across clients without sharing sensitive raw data. However, in the decentralized setting, existing FMTL frameworks are limited in their ability to capture complex task relationships and handle feature and sample heterogeneity across clients. To address these challenges, we introduce a novel sheaf-theoretic-based approach for FMTL. By representing client relationships using cellular sheaves, our framework can flexibly model interactions between heterogeneous client models. We formulate the sheaf-based FMTL optimization problem using sheaf Laplacian regularization and propose the Sheaf-FMTL algorithm to solve it. We show that the proposed framework provides a unified view encompassing many existing federated learning (FL) and FMTL approaches. Furthermore, we prove that our proposed algorithm, Sheaf-FMTL, achieves a sublinear convergence rate in line with state-of-the-art decentralized FMTL algorithms. Extensive experiments demonstrate that Sheaf-FMTL exhibits communication savings by sending significantly fewer bits compared to decentralized FMTL baselines.", "keywords": [ "federated multi-task learning", "decentralized learning", "communication-efficient", "Sheaf theory" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "b6juTJZ1I9", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Hai Zhang", "gender": "Male", "institution": "Northwest University Xi'an", "country": "CN", "position": "Full Professor" }, { "name": "Yandong Tang", "gender": "Male", "institution": "Shenyang Institue of Automation", "country": "CN", "position": "Full Professor" }, { "name": "Yao Wang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Zhi Han", "gender": "Male", "institution": "Shenyang Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhiyu Liu", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 137, "n_formula_1": 36, "n_ref_uni": 46, "n_ref": 101, "n_ref_all": 120, "n_fig": 5, "n_tab": 3, "L_tab": 614, "n_element_tab": 61, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 379, "n_element_tab_1": 49, "formula_len_all": 17951, "formula_len_all_1": 4334, "len_all": 165992, "len_all_1": 60807, "len_abs": 1438, "len_title": 92, "len_sents": 39905, "len_sents_1": 24107, "n_sents": 422, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1526, "title": "Efficient Over-parameterized Matrix Sensing via Alternating Preconditioned Gradient Descent", "abs": "We consider solving the low-rank matrix sensing problem in the over-parameterized setting, where the specified rank is larger than the true rank. Precisely, our main objective is to recover a matrix $X^*\\in\\mathbb{R}^{n_1\\times n_2}$ with rank $r_\\star$ using an over-parameterized form $LR^{\\top}$, where $L\\in\\mathbb{R}^{n_1\\times r},\\ R\\in\\mathbb{R}^{n_2\\times r}$ and $\\min\\{n_1,n_2\\}\\ge r> r_\\star$ with the true rank $r_\\star$ being unknown. The commonly used methods tackling such a problem such as Factorized Gradient Descent (FGD) can only demonstrate sub-linear convergence behavior, and their performance could significantly deteriorate when the matrix condition number is relatively large. To address this issue, we propose the alternating preconditioned gradient descent (APGD) method that an inexpensive right preconditioner with a constant damping parameter is applied to the original gradient. We prove that even starting from a random initialization, APGD can recover the target matrix at a linear convergence rate in the over-parameterized situation, independent of the condition number. Notably, unlike previous FGD-based methods, APGD alternates between updating the two factor matrices, which eliminates the reliance on a small step size, thereby enabling faster convergence. Through a series of experiments, we demonstrate that APGD achieves the fastest convergence speed compared to other methods, and further possesses strong robustness with respect to step size, condition number and other parameters.", "keywords": [ "matrix sensing", "over-parameterization", "low rank matrix recovery", "gradient descent" ], "rating_list": [ 8, 6, 1, 5 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 4, 2, 2, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "b67pPmHBJd", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Jen-tse Huang", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Postdoc" }, { "name": "Jiahao Xu", "gender": "unknown", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Pinjia He", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Tian Liang", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Wenxiang Jiao", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Wenxuan Wang", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Postdoc" }, { "name": "Youliang Yuan", "gender": "Male", "institution": "The Chinese University of Hong Kong-Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Zhaopeng Tu", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 51, "n_ref": 105, "n_ref_all": 125, "n_fig": 7, "n_tab": 7, "L_tab": 2962, "n_element_tab": 251, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1552, "n_element_tab_1": 153, "formula_len_all": 202, "formula_len_all_1": 230, "len_all": 194574, "len_all_1": 62577, "len_abs": 0, "len_title": 88, "len_sents": 40390, "len_sents_1": 28396, "n_sents": 310, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1325, "title": "Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training", "abs": "This study addresses a critical gap in safety tuning practices for Large Language Models (LLMs) by identifying and tackling a refusal position bias within safety tuning data, which compromises the models' ability to appropriately refuse generating unsafe content. We introduce a novel approach, Decoupled Refusal Training (DeRTa), designed to empower LLMs to refuse compliance to harmful prompts at any response position, significantly enhancing their safety capabilities. DeRTa incorporates two novel components: (1) Maximum Likelihood Estimation (MLE) with Harmful Response Prefix, which trains models to recognize and avoid unsafe content by appending a segment of harmful response to the beginning of a safe response, and (2) Reinforced Transition Optimization (RTO), which equips models with the ability to transition from potential harm to safety refusal consistently throughout the harmful response sequence. Our empirical evaluation, conducted using LLaMA3 and Mistral model families across six attack scenarios, demonstrates that our method not only improves model safety without compromising performance but also surpasses well-known models such as GPT-4 in defending against attacks. Importantly, our approach successfully defends recent advanced attack methods that have jailbroken GPT-4 and LLaMA3-70B-Instruct.", "keywords": [ "LLM Safety", "Jailbreak" ], "rating_list": [ 5, 3, 5, 5, 3 ], "soundness_list": [ 2, 1, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 3, 2 ], "contribution_list": [ 3, 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "b5lXUwZiD3", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Chi Jin", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Assistant Professor" }, { "name": "Jiachen Hu", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Qinghua Liu", "gender": "Male", "institution": "Microsoft Research, NYC", "country": "US", "position": "Postdoc" } ], "n_formula": 52, "n_formula_1": 9, "n_ref_uni": 47, "n_ref": 68, "n_ref_all": 91, "n_fig": 6, "n_tab": 8, "L_tab": 971, "n_element_tab": 114, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4708, "formula_len_all_1": 536, "len_all": 211603, "len_all_1": 61344, "len_abs": 3038, "len_title": 94, "len_sents": 69463, "len_sents_1": 29498, "n_sents": 555, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1391, "title": "On Limitation of Transformer for Learning HMMs", "abs": "This paper investigate the capability of transformer in learning a fundamental sequential model --- the Hidden Markov Model (HMM). We design various types of HMM examples and variants inspired by theory, and conduct extensive experiments testing and comparing the performance of both transformers and Recurrent Neural Networks (RNNs). Our experiments reveal three important findings: (1) Transformers can effectively learn a large number of HMMs, but this require the depth of transformers to be at least logarithmic in the sequence length; (2) There are challenging HMMs where Transformers struggle to learn, while RNNs succeed. We also consistently observe that Transformers underperform RNNs in both training speed and testing accuracy across all tested HMM models. (3) Long mixing times and the lack of access to intermediate latent states significantly degrade Transformer's performance, but has much less impact on RNNs' performance. To address the limitation of transformers in modeling HMMs, we demonstrate that a variant of the Chain-of-Thought (CoT), called \\emph{block CoT} in the training phase, can help transformers to reduce the evaluation error and to learn longer sequences at a cost of increasing the training time. Finally, we complement our empirical findings by theoretical results proving the expressiveness of transformers in approximating HMMs with logarithmic depth.", "keywords": [ "Transformers", "RNN", "HMM", "representation learning", "expressive power" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "b5MCteb3w7", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Annie Xie", "gender": "unknown", "institution": "Google", "country": "", "position": "Intern" }, { "name": "Bilgehan Sel", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "", "position": "PhD student" }, { "name": "Jonathan Lee", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Ming Jin", "gender": "unknown", "institution": "Virginia Tech", "country": "", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 68, "n_ref": 84, "n_ref_all": 104, "n_fig": 8, "n_tab": 1, "L_tab": 130, "n_element_tab": 24, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1315, "formula_len_all_1": 1209, "len_all": 153924, "len_all_1": 66285, "len_abs": 1324, "len_title": 150, "len_sents": 37556, "len_sents_1": 31791, "n_sents": 270, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1023, "title": "Actions Speak Louder Than States: Going Beyond Bayesian Inference in In-Context Reinforcement Learning", "abs": "In this paper, we investigate in-context learning (ICL) for reinforcement learning (RL), particularly extending beyond Bayesian inference to more advanced and richer learning paradigms in transformers. Transformers have shown promise for few-shot and zero-shot learning, but their capabilities for ICL in RL environments are not well explored. Our work studies the role of task diversity in RL environments on the downstream ICL capabilities of transformers. To do so, we introduce a novel RL benchmark, developed to provide a rich variety of tasks, essential for this exploration. Through this environment, we not only demonstrate the critical role of task diversity in facilitating advanced learning algorithms like transformers but also investigate the effects of model architecture, regularization, and other factors on the learning process. This study marks a pivotal advance in understanding the dynamics of ICL in RL, showcasing how diverse tasks can drive transformer models to surpass traditional learning methods.", "keywords": [ "meta-reinforcement learning", "in-context learning", "decision-making" ], "rating_list": [ 3, 3, 5, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "b5CEjE4zyL", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Alec F Diallo", "gender": "Male", "institution": "University of Edinburgh", "country": "GB", "position": "Postdoc" }, { "name": "Paul Patras", "gender": "unknown", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Full Professor" }, { "name": "Vaishak Belle", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Postdoc" } ], "n_formula": 51, "n_formula_1": 17, "n_ref_uni": 30, "n_ref": 36, "n_ref_all": 56, "n_fig": 15, "n_tab": 1, "L_tab": 147, "n_element_tab": 28, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 99, "n_element_tab_1": 17, "formula_len_all": 4128, "formula_len_all_1": 1414, "len_all": 181278, "len_all_1": 69288, "len_abs": 4540, "len_title": 167, "len_sents": 62707, "len_sents_1": 34016, "n_sents": 429, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 119, "L_abs": 1424, "title": "Neural Networks Decoded: Targeted and Robust Analysis of Neural Network Decisions via Causal Explanations and Reasoning", "abs": "Despite their success and widespread adoption, the opaque nature of deep neural networks (DNNs) continues to hinder trust, especially in critical applications. Current interpretability solutions often yield inconsistent or oversimplified explanations, or require model changes that compromise performance. In this work, we introduce TRACER, a novel method grounded in causal inference theory designed to estimate the causal dynamics underpinning DNN decisions without altering their architecture or compromising their performance. Our approach systematically intervenes on input features to observe how specific changes propagate through the network, affecting internal activations and final outputs. Based on this analysis, we determine the importance of individual features, and construct a high-level causal map by grouping functionally similar layers into cohesive causal nodes, providing a structured and interpretable view of how different parts of the network influence the decisions. TRACER further enhances explainability by generating counterfactuals that reveal possible model biases and offer contrastive explanations for misclassifications. Through comprehensive evaluations across diverse datasets, we demonstrate TRACER's effectiveness over existing methods and show its potential for creating highly compressed yet accurate models, illustrating its dual versatility in both understanding and optimizing DNNs.", "keywords": [ "AI Explainability", "Causal Discovery", "Neural Network Optimization" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 2, 4, 2 ] }, { "paper_id": "b57IG6N20B", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Abbas Rahimi", "gender": "Male", "institution": "IBM Research - Zurich", "country": "CH", "position": "Principal Researcher" }, { "name": "Francesco S. Carzaniga", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "PhD student" }, { "name": "Gary Tom Hoppeler", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Kaspar Schindler", "gender": "Male", "institution": "Insel Gruppe", "country": "CH", "position": "Associate Professor" }, { "name": "Michael Hersche", "gender": "Male", "institution": "International Business Machines", "country": "CH", "position": "Postdoc" } ], "n_formula": 14, "n_formula_1": 1, "n_ref_uni": 32, "n_ref": 71, "n_ref_all": 105, "n_fig": 25, "n_tab": 15, "L_tab": 11163, "n_element_tab": 2834, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 694, "formula_len_all_1": 30, "len_all": 211528, "len_all_1": 65328, "len_abs": 2708, "len_title": 162, "len_sents": 53506, "len_sents_1": 31813, "n_sents": 447, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 114, "L_abs": 2415, "title": "The Case for Cleaner Biosignals: High-fidelity Neural Compressor Enables Transfer from Cleaner iEEG to Noisier EEG", "abs": "All data modalities are not created equal, even when the signal they measure comes from the same source. In the case of the brain, two of the most important data modalities are the scalp electroencephalogram (EEG), and the intracranial electroencephalogram (iEEG). iEEG benefits from a higher signal-to-noise ratio (SNR), as it measures the electrical activity directly in the brain, while EEG is noisier and has lower spatial and temporal resolutions. Nonetheless, both EEG and iEEG are important sources of data for human neurology, from healthcare to brain\u2013machine interfaces. They are used by human experts, supported by deep learning (DL) models, to accomplish a variety of tasks, such as seizure detection and motor imagery classification. Although the differences between EEG and iEEG are well understood by human experts, the performance of DL models across these two modalities remains under-explored. To help characterize the importance of clean data on the performance of DL models, we propose BrainCodec, a high-fidelity EEG and iEEG neural compressor. We find that training BrainCodec on iEEG and then transferring to EEG yields higher reconstruction quality than training on EEG directly. In addition, we also find that training BrainCodec on both EEG and iEEG improves fidelity when reconstructing EEG. Our work indicates that data sources with higher SNR, such as iEEG, provide better performance across the board also in the medical time-series domain. This finding is consistent with reports coming from natural language processing, where clean data sources appear to have an outsized effect on the performance of the DL model overall. BrainCodec also achieves up to a 64x compression on iEEG and EEG without a notable decrease in quality. BrainCodec markedly surpasses current state-of-the-art compression models both in final compression ratio and in reconstruction fidelity. We also evaluate the fidelity of the compressed signals objectively on a seizure detection and a motor imagery task performed by standard DL models. Here, we find that BrainCodec achieves a reconstruction fidelity high enough to ensure no performance degradation on the downstream tasks. Finally, we collect the subjective assessment of an expert neurologist, that confirms the high reconstruction quality of BrainCodec in a realistic scenario. The code is available at https://github.com/IBM/eeg-ieeg-brain-compressor.", "keywords": [ "eeg", "ieeg", "intracranial eeg", "electroencephalography", "compression", "transfer", "seizure", "seizure detection", "motor imagery" ], "rating_list": [ 6, 8, 8, 8, 3 ], "soundness_list": [ 4, 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 4, 2 ], "confidence_list": [ 5, 4, 3, 4, 4 ] }, { "paper_id": "b4b6rERW4G", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Yujie Zhang", "gender": "Female", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Michael Reiter", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Full Professor" }, { "name": "Neil Gong", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Associate Professor" } ], "n_formula": 30, "n_formula_1": 3, "n_ref_uni": 34, "n_ref": 100, "n_ref_all": 146, "n_fig": 13, "n_tab": 12, "L_tab": 3697, "n_element_tab": 413, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 227, "n_element_tab_1": 14, "formula_len_all": 1624, "formula_len_all_1": 315, "len_all": 198099, "len_all_1": 66952, "len_abs": 1353, "len_title": 126, "len_sents": 76261, "len_sents_1": 33649, "n_sents": 517, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1358, "title": "Concealing Backdoors in Federated Learning by Trigger-Optimized Data Poisoning", "abs": "Federated Learning (FL) is a decentralized machine learning method that enables participants to collaboratively train a model without sharing their private data. Despite its privacy and scalability benefits, FL is susceptible to backdoor attacks, where adversaries poison the local training data of a subset of clients using backdoor triggers, aiming to make the aggregated model produce malicious results when the same backdoor conditions are met by an inference-time input. Existing backdoor attacks in FL suffer from common deficiencies: fixed trigger patterns and reliance on the assistance of model poisoning. State-of-the-art defenses based on analyzing clients' model updates exhibit a good defense performance on these attacks because of the significant divergence between malicious and benign client model updates. To effectively conceal malicious model updates among benign ones, we propose DPOT, a backdoor attack strategy in FL that dynamically constructs backdoor objectives by optimizing a backdoor trigger, making backdoor data have minimal effect on model updates. We provide theoretical justifications for DPOT's attacking principle and display experimental results showing that DPOT, via only a data-poisoning attack, effectively undermines state-of-the-art defenses and outperforms existing backdoor attack techniques on various datasets.", "keywords": [ "Data Poisoning", "Backdoor Attack", "Federated Learning" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "b4A20ODZBq", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Dongqi Fu", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Jingrui He", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Katherine Tieu", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "Undergrad student" }, { "name": "Ross Maciejewski", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "Full Professor" }, { "name": "Zihao Li", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 46, "n_formula_1": 36, "n_ref_uni": 27, "n_ref": 83, "n_ref_all": 132, "n_fig": 1, "n_tab": 20, "L_tab": 8599, "n_element_tab": 656, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 2509, "n_element_tab_1": 192, "formula_len_all": 4519, "formula_len_all_1": 2239, "len_all": 209804, "len_all_1": 80924, "len_abs": 1936, "len_title": 172, "len_sents": 68937, "len_sents_1": 35201, "n_sents": 456, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 124, "L_abs": 1941, "title": "Is Attention All You Need for Temporal Link Prediction? A Lightweight Alternative via Learnable Positional Encoding and MLPs", "abs": "Link prediction is of key importance in many real-world applications like social network analysis and recommender systems. To leverage the expressive power for achieving SOTA performance, many recent works adapt the attention mechanism to the structured data for link prediction, in which dense or relational attention is often unaffordable on large-scale structured data. Moreover, in a realistic setting, the time-evolving topological and feature information can raise more challenging questions about the efficiency and effectiveness of attention mechanisms. In spite of the expressive power, we discern that the attention mechanism may not always be as irreplaceable as expected for temporal graph representation learning, at least not for temporal link prediction tasks. Formally, we discover that some deliberately-designed simple positional encoding can enable MLPs to exploit attributed graph information to achieve SOTA performance than complex graph transformers. Hence, we propose a simple temporal link prediction model, named SimpleTLP. In detail, for SimpleTLP, we first propose to adapt Fourier Transform on temporal graphs for learning informative positional encoding, then we (1) prove this learning scheme can make positional encoding preserve the temporal graph topology from the spatial-temporal spectral viewpoint, (2) verify MLPs can fully exploit the expressiveness and reach and even surpass Transformers on that encoding, (3) change different initial positional encoding inputs to show robustness, (4) analyze the theoretical complexity and obtain less empirical running time than SOTA baselines, and (5) demonstrate its temporal link prediction out-performance in a comprehensive way on 13 classic datasets and with 10 algorithms in both transductive and inductive settings using 3 different sampling strategies. Also, SimpleTLP obtains the leading performance in the large-scale TGB benchmark (the newest TGB 2.0).", "keywords": [ "Link Prediction", "Graph Transformer", "Positional Encoding" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "b42wmsdwmB", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Xinyan Chen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Jianfei Yang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 52, "n_ref": 83, "n_ref_all": 97, "n_fig": 2, "n_tab": 14, "L_tab": 4596, "n_element_tab": 716, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 135, "n_element_tab_1": 4, "formula_len_all": 466, "formula_len_all_1": 571, "len_all": 168500, "len_all_1": 65537, "len_abs": 1508, "len_title": 120, "len_sents": 47223, "len_sents_1": 33277, "n_sents": 325, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1515, "title": "X-Fi: A Modality-Invariant Foundation Model for Multimodal Human Sensing", "abs": "Human sensing, which employs various sensors and advanced deep learning technologies to accurately capture and interpret human body information, has significantly impacted fields like public security and robotics. However, current human sensing primarily depends on modalities such as cameras and LiDAR, each of which has its own strengths and limitations. Furthermore, existing multimodal fusion solutions are typically designed for fixed modality combinations, requiring extensive retraining when modalities are added or removed for diverse scenarios. In this paper, we propose a modality-invariant foundation model for all modalities, X-Fi, to address these issues. X-Fi enables the independent or combinatory use of sensor modalities without additional training by utilizing a transformer structure to accommodate variable input sizes and incorporating a novel \"X-fusion\" mechanism to preserve modality-specific features during multimodal integration. This approach not only enhances adaptability but also facilitates the learning of complementary features across modalities. Extensive experiments conducted on the MM-Fi and XRF55 datasets, employing six distinct modalities, demonstrate that X-Fi achieves state-of-the-art performance in human pose estimation (HPE) and human activity recognition (HAR) tasks. The findings indicate that our proposed model can efficiently support a wide range of human sensing applications, ultimately contributing to the evolution of scalable, multimodal sensing technologies.", "keywords": [ "human sensing", "multimodal learning" ], "rating_list": [ 6, 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 3, 2 ], "confidence_list": [ 3, 5, 4, 3, 4 ] }, { "paper_id": "b3VzHRXrXh", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Ayush Roy", "gender": "Male", "institution": "Jadavpur University", "country": "IN", "position": "Undergrad student" }, { "name": "Jun Xia", "gender": "unknown", "institution": "State University of New York at Buffalo", "country": "US", "position": "Full Professor" }, { "name": "Mingzhen Huang", "gender": "Male", "institution": "State University of New York, Buffalo", "country": "US", "position": "PhD student" }, { "name": "Samin Enam", "gender": "Female", "institution": "State University of New York at Buffalo", "country": "US", "position": "MS student" }, { "name": "Vishnu Suresh Lokhande", "gender": "Male", "institution": ", State University of New York at Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Won Kim", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 4, "n_ref_uni": 22, "n_ref": 28, "n_ref_all": 42, "n_fig": 8, "n_tab": 5, "L_tab": 992, "n_element_tab": 58, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1291, "n_element_tab_1": 42, "formula_len_all": 811, "formula_len_all_1": 221, "len_all": 121957, "len_all_1": 45796, "len_abs": 1626, "len_title": 161, "len_sents": 30087, "len_sents_1": 20809, "n_sents": 229, "n_sents_1": 148, "venue": "ICLR2025", "decision": "Reject", "L_title": 113, "L_abs": 1631, "title": "Causal Frameworks and Feature Discrepancy Loss: Addressing Data Scarcity and Enhancing Medical Image Segmentation", "abs": "Data scarcity poses a significant challenge for deep learning models in medical imaging, particularly for training and generalization. Previous studies have demonstrated the efficacy of data pooling from various sources, facilitating the analysis of weak but significant correlations between imaging data and disease incidence. This approach is often constrained by strict data-sharing protocols among institutions, resulting in models reliant on external data sources. In this work, we address the issue of data scarcity by leveraging the available data for segmentation tasks across various medical imaging modalities. Based on our observation that samples with minimal foreground-background feature differences often demonstrate inadequate segmentation performance, we propose a causal-inspired foreground-background feature discrepancy penalty function, which improves feature separation and alleviates segmentation difficulties caused by homogeneous pixel distributions. The proposed feature discrepancy loss is mathematically grounded, with a lower bound defined by the negative logarithm of the Dice coefficient, suggesting that increased feature separation correlates with improved Dice scores. To further validate our approach, we introduce a novel ultrasound dataset for triple-negative breast cancer (TNBC), and we evaluate the method across three state-of-the-art segmentation architectures to demonstrate competitive performance. In addition, the results highlight the robustness of our method in mitigating performance decrease due to distribution shifts when new, differently distributed data batches are introduced.", "keywords": [ "causal reasoning", "bioemdical image segmentation", "data dilemma" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "b3CzCCCILJ", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Manuel Kansy", "gender": "Male", "institution": "ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Otmar Hilliges", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "Full Professor" }, { "name": "Romann M. Weber", "gender": "Male", "institution": "Disney Research, Disney", "country": "CH", "position": "Research Scientist" }, { "name": "Seyedmorteza Sadat", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 13, "n_ref_uni": 29, "n_ref": 98, "n_ref_all": 135, "n_fig": 15, "n_tab": 11, "L_tab": 1978, "n_element_tab": 474, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 976, "n_element_tab_1": 73, "formula_len_all": 885, "formula_len_all_1": 734, "len_all": 149269, "len_all_1": 53859, "len_abs": 1314, "len_title": 121, "len_sents": 35669, "len_sents_1": 24715, "n_sents": 258, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1325, "title": "Revamping Diffusion Guidance for Conditional and Unconditional Generation", "abs": "Classifier-free guidance (CFG) has become the standard method for enhancing the quality of conditional diffusion models. However, employing CFG requires either training an unconditional model alongside the main diffusion model or modifying the training procedure by periodically inserting a null condition. There is also no clear extension of CFG to unconditional models. In this paper, we revisit the core principles of CFG and introduce a new method, independent condition guidance (ICG), which provides the benefits of CFG without the need for any special training procedures. Our approach streamlines the training process of conditional diffusion models and can also be applied during inference on any pre-trained conditional model. Additionally, by leveraging the time-step information encoded in all diffusion networks, we propose an extension of CFG, called time-step guidance (TSG), which can be applied to *any* diffusion model, including unconditional ones. Our guidance techniques are easy to implement and have the same sampling cost as CFG. Through extensive experiments, we demonstrate that ICG matches the performance of standard CFG across various conditional diffusion models. Moreover, we show that TSG improves generation quality in a manner similar to CFG, without relying on any conditional information.", "keywords": [ "diffusion models", "classifier-free guidance" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "b39J2X4rjT", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Jianjiang Feng", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Zhou", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jiwen Lu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Wenxuan Guo", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiuwei Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ziwei Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 29, "n_ref": 63, "n_ref_all": 88, "n_fig": 9, "n_tab": 4, "L_tab": 1396, "n_element_tab": 191, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1339, "n_element_tab_1": 143, "formula_len_all": 403, "formula_len_all_1": 457, "len_all": 121809, "len_all_1": 69799, "len_abs": 1563, "len_title": 122, "len_sents": 40932, "len_sents_1": 31443, "n_sents": 328, "n_sents_1": 241, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1577, "title": "Efficient Sparse Single-stage 3D Visual Grounding with Text-guided Pruning", "abs": "In this paper, we propose an efficient sparse convolution-based architecture called ESS3D for 3D visual grounding. Conventional 3D visual grounding methods are difficult to meet the requirements of real-time inference due to the two-stage or point-based architecture. Inspired by the success of multi-level fully sparse convolutional architecture in 3D object detection, we aim to build a new 3D visual grounding framework following this technical route. However, as in visual grounding task the 3D scene representation should be deeply interacted with text features, sparse convolution-based architecture is inefficient for this interaction due to the large amount of voxel features. To this end, we propose text-guided pruning (TGP) and completion-based addition (CBA) to deeply fuse 3D scene representation and text features in an efficient way by gradual region pruning and target completion. Specifically, TGP iteratively sparsifies the 3D scene representation and thus efficiently interacts the voxel features with text features by cross-attention. To mitigate the affect of pruning on delicate geometric information, CBA adaptively fixes the over-pruned region by voxel completion with negligible computational overhead. Compared with previous single-stage methods, ESS3D achieves top inference speed and surpasses previous fastest method by 100\\% FPS. ESS3D also achieves state-of-the-art accuracy even compared with two-stage methods, with $+1.13$ lead of Acc@0.5 on ScanRefer, and $+5.4$ and $+5.0$ leads on NR3D and SR3D respectively. The code will be released soon.", "keywords": [ "3D Visual Grounding", "Efficient Single-stage Architecture", "Feature Pruning" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 2, 3, 4 ] }, { "paper_id": "b2oLgk5XRE", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "ChangHyun Kim", "gender": "Male", "institution": "University of Seoul", "country": "KR", "position": "MS student" }, { "name": "Gyeongdong Woo", "gender": "Male", "institution": "University of Seoul", "country": "KR", "position": "MS student" }, { "name": "Jaesung Lim", "gender": "Male", "institution": "University of Seoul", "country": "KR", "position": "MS student" }, { "name": "Jong-June Jeon", "gender": "Male", "institution": "University of Seoul", "country": "KR", "position": "Full Professor" }, { "name": "Seunghwan An", "gender": "Male", "institution": "Incheon National University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 21, "n_formula_1": 13, "n_ref_uni": 72, "n_ref": 125, "n_ref_all": 140, "n_fig": 1, "n_tab": 11, "L_tab": 9601, "n_element_tab": 483, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 775, "n_element_tab_1": 60, "formula_len_all": 2435, "formula_len_all_1": 1083, "len_all": 243746, "len_all_1": 65147, "len_abs": 1287, "len_title": 120, "len_sents": 51511, "len_sents_1": 28291, "n_sents": 793, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1293, "title": "Context-Driven Missing Data Imputation via Large Language Model", "abs": "Missing data poses significant challenges for machine learning and deep learning algorithms. In this paper, we aim to enhance post-imputation performance, measured by machine learning utility (MLu). We introduce a nearest-neighbor-based imputation method, DrIM, designed for heterogeneous tabular datasets. However, calculating similarity in the data space becomes challenging due to the varying presence of missing entries across different columns. To address this issue, we leverage the representation learning capabilities of language models. By transforming the tabular dataset into a text-format dataset and replacing the missing entries with mask (or unk) tokens, we extract representations that capture contextual information. This mapping to a continuous representation space enables the use of well-defined similarity measurements. Additionally, we incorporate a contrastive learning framework to refine the representations, ensuring that the representations of observations with similar information in the observed columns, regardless of the missingness patterns, are closely aligned. To validate our proposed model, we evaluate its performance in missing data imputation across 10 real-world tabular datasets, demonstrating its ability to produce a Complete dataset having high MLu.", "keywords": [ "Missing data imputation", "Heterogeneous tabular data", "Large Language Models", "Nearest neighbor", "Constrative learning" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "b2LklBgdcL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Canyu Chen", "gender": "unknown", "institution": "Illinois Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jun Li", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Peiran Wu", "gender": "unknown", "institution": "University of Bristol", "country": "GB", "position": "PhD student" }, { "name": "Rossella Arcucci", "gender": "Female", "institution": "Imperial College London ", "country": "", "position": "Senior Lecturer" }, { "name": "che liu", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Cosmin Ionut Bercea", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 24, "n_ref": 49, "n_ref_all": 75, "n_fig": 7, "n_tab": 5, "L_tab": 1805, "n_element_tab": 284, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 217, "n_element_tab_1": 23, "formula_len_all": 177, "formula_len_all_1": 204, "len_all": 101320, "len_all_1": 50005, "len_abs": 1624, "len_title": 132, "len_sents": 28110, "len_sents_1": 22736, "n_sents": 233, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1651, "title": "FMBench: Benchmarking Fairness in Multimodal Large Language Models on Medical Tasks", "abs": "Advancements in Multimodal Large Language Models (MLLMs) have significantly improved medical task performance, such as Visual Question Answering (VQA) and Report Generation (RG). However, the fairness of these models across diverse demographic groups remains underexplored, despite its importance in healthcare. This oversight is partly due to the lack of demographic diversity in existing medical multimodal datasets, which complicates the evaluation of fairness. In response, we propose **FMBench**, the first benchmark designed to evaluate the fairness of MLLMs performance across diverse demographic attributes. FMBench has the following key features: **1:** It includes four demographic attributes: race, ethnicity, language, and gender, across two tasks, VQA and RG, under zero-shot settings. **2:** Our VQA task is free-form, enhancing real-world applicability and mitigating the biases associated with predefined choices.\n**3:** We utilize both lexical metrics and LLM-based metrics, aligned with clinical evaluations, to assess models not only for linguistic accuracy but also from a clinical perspective. Furthermore, we introduce a new metric, **Fairness-Aware Performance (FAP)**, to evaluate how fairly MLLMs perform across various demographic attributes. We thoroughly evaluate the performance and fairness of eight state-of-the-art open-source MLLMs, including both general and medical MLLMs, ranging from 7B to 26B parameters on the proposed benchmark. We aim for **FMBench** to assist the research community in refining model evaluation and driving future advancements in the field. All data and code will be released upon acceptance.", "keywords": [ "Multimodal Large Language Models ; Visual Question Answering ; Report Generation ; FMBench ; Fairness-Aware Performance ;" ], "rating_list": [ 1, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "b2FFWnwZxl", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Ethan Feng", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "MS student" }, { "name": "Jacob Fein-Ashley", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Minh Pham", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "MS student" } ], "n_formula": 43, "n_formula_1": 41, "n_ref_uni": 13, "n_ref": 17, "n_ref_all": 25, "n_fig": 0, "n_tab": 6, "L_tab": 2783, "n_element_tab": 160, "n_fig_1": 0, "n_tab_1": 6, "L_tab_1": 3605, "n_element_tab_1": 173, "formula_len_all": 1485, "formula_len_all_1": 1483, "len_all": 79253, "len_all_1": 51116, "len_abs": 883, "len_title": 121, "len_sents": 22017, "len_sents_1": 17813, "n_sents": 175, "n_sents_1": 140, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 887, "title": "HVT: A Comprehensive Vision Framework for Learning in Non-Euclidean Space", "abs": "Data representation in non-Euclidean spaces has proven effective for capturing hierarchical and complex relationships in real-world datasets. Hyperbolic spaces, in particular, provide efficient embeddings for hierarchical structures. This paper introduces the Hyperbolic Vision Transformer (HVT), a novel extension of the Vision Transformer (ViT) that integrates hyperbolic geometry. While traditional ViTs operate in Euclidean space, our method enhances the self-attention mechanism by leveraging hyperbolic distance and M\u00f6bius transformations. This enables more effective modeling of hierarchical and relational dependencies in image data. We present rigorous mathematical formulations, showing how hyperbolic geometry can be incorporated into attention layers, feed-forward networks, and optimization. We offer improved performance for image classification using the ImageNet dataset.", "keywords": [ "hyperbolic", "topology", "vit", "transformer", "vision", "manifold" ], "rating_list": [ 3, 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 2, 1 ], "presentation_list": [ 1, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 1 ], "confidence_list": [ 5, 4, 3, 5, 4 ] }, { "paper_id": "b24n2LS2BJ", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Jaesik Choi", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" }, { "name": "Myeongjin Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Chang Wonjoon", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" } ], "n_formula": 32, "n_formula_1": 13, "n_ref_uni": 38, "n_ref": 76, "n_ref_all": 101, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2923, "formula_len_all_1": 1149, "len_all": 123384, "len_all_1": 61296, "len_abs": 1265, "len_title": 118, "len_sents": 40801, "len_sents_1": 29287, "n_sents": 330, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1274, "title": "Rethinking Shapley Value for Negative Interactions in Non-convex Games", "abs": "We study causal interaction for payoff allocation in cooperative game theory, including quantifying feature attribution for deep learning models. Most feature attribution methods mainly stem from the criteria from the Shapley value, which provides a unique payoff vector for players by marginalizing contributions in a cooperative game. However, interactions between players in the game do not exactly appear in the original formulation of the Shapley value. In this work, we clarify the role of interactions in computing the Shapley value by reformulation and discuss implicit assumptions from a game-theoretical perspective. Our theoretical analysis demonstrates that when negative interactions exist---common in deep learning models---attributions or payoffs can be underrated by the efficiency axiom. We suggest a new allocation rule that decomposes contributions into interactions and aggregates positive parts for non-convex games. Furthermore, we propose an approximation algorithm to reduce the cost of interaction computation which can be applied for differentiable functions such as deep learning models. Our approach mitigates counter-intuitive phenomena where even features highly relevant to the decision are assigned low attribution in the previous approaches.", "keywords": [ "Shapley value", "Interaction", "Feature Attribution" ], "rating_list": [ 5, 8, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "b20VK2GnSs", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Jie Lu", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "Professor" }, { "name": "Xiaoyu Yang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "EN YU", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Postdoc" } ], "n_formula": 25, "n_formula_1": 10, "n_ref_uni": 16, "n_ref": 29, "n_ref_all": 41, "n_fig": 5, "n_tab": 9, "L_tab": 3099, "n_element_tab": 178, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2258, "n_element_tab_1": 108, "formula_len_all": 1631, "formula_len_all_1": 535, "len_all": 206333, "len_all_1": 60834, "len_abs": 1843, "len_title": 132, "len_sents": 50631, "len_sents_1": 29645, "n_sents": 316, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1849, "title": "Adapting Multi-modal Large Language Model to Concept Drift From Pre-training Onwards", "abs": "Multi-modal Large Language Models (MLLMs) frequently face challenges from concept drift when dealing with real-world streaming data, wherein distributions change unpredictably. This mainly includes gradual drift due to long-tailed data and sudden drift from Out-Of-Distribution (OOD) data, both of which have increasingly drawn the attention of the research community. While these issues have been extensively studied in the individual domain of vision or language, their impacts on MLLMs in concept drift settings remain largely underexplored. In this paper, we reveal the susceptibility and vulnerability of Vision-Language (VL) models to significant biases arising from gradual drift and sudden drift, particularly in the pre-training. To effectively address these challenges, we propose a unified framework that extends concept drift theory to the multi-modal domain, enhancing the adaptability of the VL model to unpredictable distribution changes. Additionally, a T-distribution based drift adapter is proposed to effectively mitigate the bias induced by the gradual drift, which also facilitates the model in distinguishing sudden distribution changes through explicit distribution modeling. Extensive experiments demonstrate our method enhances the efficiency and accuracy of image-text alignment in the pre-training of VL models, particularly in the concept drift scenario. Moreover, various downstream tasks exhibit significant improvements in our model's ability to adapt to the long-tailed open world. Furthermore, we create a set of multi-modal datasets called OpenMMlo, specifically tailored for the long-tailed open-world setting, to validate our findings. To foster the development of the multi-modal community, we have made both OpenMMlo datasets and our code publicly available at: https://github.com/XiaoyuYoung/ConceptDriftMLLMs.", "keywords": [ "Concept Drift", "Multi-modal Large Language Model", "Data Defect" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "b1vVm6Ldrd", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Guiyang Hou", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Intern" }, { "name": "Sihao Shen", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "staff" }, { "name": "Weiming Lu", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Wenqi Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yongliang Shen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zeqi Tan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 27, "n_ref": 44, "n_ref_all": 66, "n_fig": 4, "n_tab": 4, "L_tab": 814, "n_element_tab": 139, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 164, "formula_len_all_1": 57, "len_all": 104708, "len_all_1": 44923, "len_abs": 1899, "len_title": 131, "len_sents": 31014, "len_sents_1": 21697, "n_sents": 207, "n_sents_1": 144, "venue": "ICLR2025", "decision": "withdraw", "L_title": 130, "L_abs": 1965, "title": "Entering Real Social World! Benchmarking the Theory of Mind and Socialization Capabilities of LLMs from a First-person Perspective", "abs": "In the social world, humans possess the capability to infer and reason about others' mental states (such as emotions, beliefs, and intentions), known as Theory of Mind (ToM). Simultaneously, humans' own mental states evolve in response to social situations, a capability we refer to as \\textit{socialization}. Together, these capabilities form the foundation of human social interaction. In the era of artificial intelligence (AI), especially with the development of large language models (LLMs), we raise intriguing questions: How do LLMs perform in terms of ToM and \\textit{socialization} capabilities? And more broadly, can these AI models truly enter and navigate the real social world? Existing research evaluating LLMs' ToM and \\textit{socialization} capabilities by positioning LLMs as passive observers from a third-person perspective, rather than as active participants. However, compared to the third-person perspective, observing and understanding the world from an ego-centric first-person perspective is a natural approach for both humans and AI agents. The ToM and \\textit{socialization} capabilities of LLMs from a first-person perspective, a crucial attribute for advancing embodied AI agents, remain unexplored. To answer the aforementioned questions and bridge the research gap, we introduce \\textit{EgoSocialArena}, a novel framework designed to evaluate and investigate the ToM and \\textit{socialization} capabilities of LLMs from a first-person perspective. It encompasses two evaluation environments: static environment and interactive environment, with seven scenarios: Daily Life, Counterfactual, New World, Blackjack, Number Guessing, and Limit Texas Hold\u2019em, totaling 2,195 data entries. With \\textit{EgoSocialArena}, we have conducted a comprehensive evaluation of nine advanced LLMs and observed some key insights regarding the future development of LLMs as well as the capabilities levels of the most advanced LLMs currently available.", "keywords": [ "Theory of Mind", "Socialization", "First-person Perspective" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "b1ivBPLb1n", "primary_area": "foundation or frontier models, including LLMs", "n_author": 39, "author_info_list": [ { "name": "Aarash Feizi", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "Abhay Puri", "gender": "Male", "institution": "ServiceNow Research", "country": "CA", "position": "Researcher" }, { "name": "Ahmed Masry", "gender": "Male", "institution": "York University", "country": "CA", "position": "PhD student" }, { "name": "Akshay Kalkunte Suresh", "gender": "Male", "institution": "ServiceNow Inc", "country": "US", "position": "Researcher" }, { "name": "Amirhossein Abaskohi", "gender": "Male", "institution": "ServiceNow Inc", "country": "CA", "position": "Intern" }, { "name": "Christopher J Pal", "gender": "unknown", "institution": "Polytechnique Montreal", "country": "CA", "position": "Full Professor" }, { "name": "David Vazquez", "gender": "Male", "institution": "ServiceNow research", "country": "CA", "position": "Researcher" }, { "name": "Fran\u00e7ois Savard", "gender": "Male", "institution": "ServiceNow Inc", "country": "CA", "position": "Researcher" }, { "name": "Issam H. Laradji", "gender": "Male", "institution": "ServiceNow", "country": "US", "position": "Researcher" }, { "name": "Joao Monteiro", "gender": "Male", "institution": "Autodesk", "country": "GB", "position": "Researcher" }, { "name": "M. \u00d6zsu", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Full Professor" }, { "name": "Mahsa Massoud", "gender": "Female", "institution": "McGill University, McGill University", "country": "CA", "position": "MS student" }, { "name": "Marco Pedersoli", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "Associate Professor" }, { "name": "Mats Leon Richter", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Postdoc" }, { "name": "Nicolas Chapados", "gender": "Male", "institution": "ServiceNow Research", "country": "CA", "position": "VP Research" }, { "name": "Perouz Taslakian", "gender": "Female", "institution": "ServiceNow", "country": "CA", "position": "Researcher" }, { "name": "Pierre-Andre Noel", "gender": "Male", "institution": "ServiceNow", "country": "CA", "position": "Researcher" }, { "name": "Rabiul Awal", "gender": "Male", "institution": "Mila - Quebec AI Institute", "country": "CA", "position": "Intern" }, { "name": "Sanket Biswas", "gender": "Male", "institution": "Computer Vision Center, Universitat Aut\u00f3noma de Barcelona", "country": "ES", "position": "PhD student" }, { "name": "Sathwik Tejaswi Madhusudhan", "gender": "Male", "institution": "ServiceNow Inc", "country": "US", "position": "Researcher" }, { "name": "Saverio Vadacchino", "gender": "Male", "institution": "McGill University", "country": "", "position": "MS student" }, { "name": "Sean Hughes", "gender": "unknown", "institution": "ServiceNow", "country": "", "position": "AI Ecosystem Director" }, { "name": "Sepideh Kharaghani", "gender": "Female", "institution": "Servicenow", "country": "CA", "position": "Researcher" }, { "name": "Shravan Nayak", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "MS student" }, { "name": "Shubham Agarwal", "gender": "Male", "institution": "Krutrim AI", "country": "IN", "position": "Researcher" }, { "name": "Siba Smarak Panigrahi", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Siva Reddy", "gender": "Male", "institution": "ServiceNow Inc", "country": "CA", "position": "Researcher" }, { "name": "Spandana Gella", "gender": "Female", "institution": "ServiceNow Inc", "country": "CA", "position": "Researcher" }, { "name": "Suyuchen Wang", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "Tianyu Zhang", "gender": "Not Specified", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, University of Montreal", "country": "CA", "position": "PhD student" }, { "name": "Torsten Scholak", "gender": "Male", "institution": "ServiceNow Research", "country": "CA", "position": "Researcher" }, { "name": "Xiangru Jian", "gender": "Male", "institution": "ServiceNow Research", "country": "CA", "position": "Researcher" }, { "name": "Ying Zhang", "gender": "Female", "institution": "ServiceNow Inc.", "country": "CA", "position": "Engineering Manager" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" }, { "name": "Zichao Li", "gender": "unknown", "institution": "McGill University", "country": "", "position": "PhD student" }, { "name": "Juan Rodriguez", "gender": "Male", "institution": "ServiceNow Research", "country": "CA", "position": "PhD student" }, { "name": "Krishnamurthy Dvijotham", "gender": "Male", "institution": "ServiceNow", "country": "US", "position": "Researcher" }, { "name": "sai rajeswar mudumba", "gender": "Male", "institution": "ServiceNow", "country": "CA", "position": "Research Scientist" }, { "name": "Sara Shanian", "gender": "Female", "institution": "ServiceNow Inc", "country": "CA", "position": "Manager" } ], "n_formula": 7, "n_formula_1": 1, "n_ref_uni": 59, "n_ref": 167, "n_ref_all": 262, "n_fig": 23, "n_tab": 9, "L_tab": 12669, "n_element_tab": 2151, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 687, "n_element_tab_1": 79, "formula_len_all": 1306, "formula_len_all_1": 260, "len_all": 422376, "len_all_1": 58109, "len_abs": 3202, "len_title": 156, "len_sents": 123000, "len_sents_1": 24277, "n_sents": 1042, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 108, "L_abs": 1637, "title": "BigDocs: An Open and Permissively-Licensed Dataset for Training Multimodal Models on Document and Code Tasks", "abs": "Multimodal AI has the potential to significantly enhance document-understanding tasks, such as processing receipts, understanding workflows, extracting data from documents, and summarizing reports. Code generation tasks that require long-structured outputs can also be enhanced by multimodality. Despite this, their use in commercial applications is often limited due to limited access to relevant training data and restrictive licensing, which hinders open access. To address these limitations, we introduce BigDocs-7.5M, a high-quality, open-access dataset comprising 7.5 million multimodal documents across 30 tasks. We use an efficient data curation process to ensure that our data is high quality and license-permissive. Our process emphasizes accountability, responsibility, and transparency through filtering rules, traceable metadata, and careful content analysis. Additionally, we introduce BigDocs-Bench,, a benchmark suite with 10 novel tasks where we carefully create datasets that reflect real-world use cases involving reasoning over Graphical User Interfaces (GUI) and code generation from images. Our experiments show that training with BigDocs-Bench, improves average performance up to 25.8% over closed-source GPT-4o in document reasoning and structured output tasks such as Screenshot2HTML or Image2Latex generation. Finally, human evaluations revealed that participants preferred the outputs from models trained with BigDocs over those from GPT-4o. This suggests that BigDocs can help both academics and the open-source community utilize and improve AI tools to enhance multimodal capabilities and document reasoning.", "keywords": [ "datasets", "vision language models", "multimodal", "large language models" ], "rating_list": [ 5, 8, 5, 5 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "b1CVu9l5GO", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Andrey Kolobov", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Furong Huang", "gender": "Female", "institution": "University of Maryland", "country": "US", "position": "Associate Professor" }, { "name": "Hal Daum\u00e9 III", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Jianfeng Gao", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Ruijie Zheng", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Shuaiyi Huang", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Yongyuan Liang", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Jianwei Yang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 32, "n_ref": 61, "n_ref_all": 79, "n_fig": 11, "n_tab": 5, "L_tab": 2299, "n_element_tab": 195, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1008, "n_element_tab_1": 166, "formula_len_all": 15, "formula_len_all_1": 37, "len_all": 192224, "len_all_1": 62632, "len_abs": 1272, "len_title": 100, "len_sents": 46400, "len_sents_1": 29972, "n_sents": 320, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1228, "title": "TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic Policies", "abs": "Although large vision-language-action (VLA) models pretrained on extensive robot datasets offer promising generalist policies for robotic learning, they still struggle with spatial-temporal dynamics in interactive robotics, making them less effective in handling complex tasks, such as manipulation. In this work, we introduce visual trace prompting, a simple yet effective approach to facilitate VLA models\u2019 spatial-temporal awareness for action prediction by encoding state-action trajectories visually. We develop a new TraceVLA model by finetuning\nOpenVLA on our own collected dataset of 150K robot manipulation trajectories using visual trace prompting. Evaluations of TraceVLA across 137 configurations in SimplerEnv and 4 tasks on a physical WidowX robot demonstrate state-of-the-art performance, outperforming OpenVLA by 10% on SimplerEnv and 3.5x on real-robot tasks and exhibiting robust generalization across diverse embodiments and scenarios. To further validate the effectiveness and generality of our method, we present a compact VLA model based on 4B Phi-3-Vision, pretrained on the Open-X-Embodiment and finetuned on our dataset, rivals the 7B OpenVLA baseline while significantly improving inference efficiency.", "keywords": [ "Vision Language Model", "Robot Learning" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "b10lRabU9W", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Guohao Dai", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Jianyuan Zhong", "gender": "unknown", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Ningyi Xu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Qiang Xu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Shan Huang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Zhengyuan Shi", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Ziyang Zheng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 32, "n_ref": 80, "n_ref_all": 110, "n_fig": 13, "n_tab": 7, "L_tab": 3061, "n_element_tab": 441, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 767, "n_element_tab_1": 37, "formula_len_all": 1648, "formula_len_all_1": 899, "len_all": 144922, "len_all_1": 70144, "len_abs": 1400, "len_title": 135, "len_sents": 44674, "len_sents_1": 31072, "n_sents": 319, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1419, "title": "DeepGate4: Efficient and Effective Representation Learning for Circuit Design at Scale", "abs": "Circuit representation learning has become pivotal in electronic design automation, enabling critical tasks such as testability analysis, logic reasoning, power estimation, and SAT solving. However, existing models face significant challenges in scaling to large circuits due to limitations like over-squashing in graph neural networks and the quadratic complexity of transformer-based models. To address these issues, we introduce \\textbf{DeepGate4}, a scalable and efficient graph transformer specifically designed for large-scale circuits. DeepGate4 incorporates several key innovations: (1) an update strategy tailored for circuit graphs, which reduce memory complexity to sub-linear and is adaptable to any graph transformer; (2) a GAT-based sparse transformer with global and local structural encodings for AIGs; and (3) an inference acceleration CUDA kernel that fully exploit the unique sparsity patterns of AIGs. Our extensive experiments on the ITC99 and EPFL benchmarks show that DeepGate4 significantly surpasses state-of-the-art methods, achieving 15.5\\% and 31.1\\% performance improvements over the next-best models. Furthermore, the Fused-DeepGate4 variant reduces runtime by 35.1\\% and memory usage by 46.8\\%, making it highly efficient for large-scale circuit analysis. These results demonstrate the potential of DeepGate4 to handle complex EDA tasks while offering superior scalability and efficiency.", "keywords": [ "circuit representation learning", "graph transformer" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "b0qxhCaKIY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Alois Knoll", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Full Professor" }, { "name": "Yuning Cui", "gender": "Male", "institution": "Technical University of Munich", "country": "DE", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 7, "n_ref_uni": 46, "n_ref": 138, "n_ref_all": 162, "n_fig": 7, "n_tab": 23, "L_tab": 11158, "n_element_tab": 825, "n_fig_1": 5, "n_tab_1": 17, "L_tab_1": 5632, "n_element_tab_1": 555, "formula_len_all": 285, "formula_len_all_1": 329, "len_all": 187444, "len_all_1": 69863, "len_abs": 1515, "len_title": 116, "len_sents": 33803, "len_sents_1": 21194, "n_sents": 268, "n_sents_1": 161, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1521, "title": "When Prompt Meets Frequency Learning for Efficient Image Restoration", "abs": "Image restoration, as a longstanding task, aims to recover the missing details and remove degradations from a corrupted observation. Inspired by the success of prompt learning in natural language processing, many prompt-based approaches have been developed for various image restoration tasks. However, these algorithms mostly operate in the spatial domain. As frequency learning plays an important role in image restoration by reducing the spectra discrepancy between degraded/sharp image pairs, this study explores the potential of frequency prompts for efficient image restoration by proposing a plug-and-play mechanism, which mainly comprises a prompt generation module and a prompt integration module. Specifically, the former encodes different frequency information by aggregating the pre-defined learnable parameters under the guidance of implicitly decomposed spectra of input features. Subsequently, to dynamically guide reconstruction, the learned prompts are embedded into the spectra of features via dual-dimensional attention for effective frequency learning. To demonstrate the effectiveness of our mechanism, we conduct experiments on general and all-in-one image restoration tasks. By incorporating it into a CNN-based backbone, the model achieves state-of-the-art performance on 15 benchmark datasets for five representative image restoration tasks. Furthermore, equipped with our mechanism, a pure Transformer network performs favorably against state-of-the-art algorithms under two all-in-one settings.", "keywords": [ "Image restoration", "prompt learning", "frequency learning", "all-in-one" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "b0WpXBABdu", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Max Ruiz Luyten", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Mihaela van der Schaar", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Yangming Li", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 58, "n_formula_1": 11, "n_ref_uni": 53, "n_ref": 83, "n_ref_all": 143, "n_fig": 10, "n_tab": 12, "L_tab": 6135, "n_element_tab": 553, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3069, "formula_len_all_1": 507, "len_all": 234432, "len_all_1": 69407, "len_abs": 2113, "len_title": 129, "len_sents": 72000, "len_sents_1": 32566, "n_sents": 564, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1170, "title": "Risk-Sensitive Diffusion: Robustly Optimizing Diffusion Models with Noisy Samples", "abs": "Diffusion models are mainly studied on image data. However, non-image data (e.g., tabular data) are also prevalent in real applications and tend to be noisy due to some inevitable factors in the stage of data collection, degrading the generation quality of diffusion models. In this paper, we consider a novel problem setting where every collected sample is paired with a vector indicating the data quality: risk vector. This setting applies to many scenarios involving noisy data and we propose risk-sensitive SDE, a type of stochastic differential equation (SDE) parameterized by the risk vector, to address it. With some proper coefficients, risk-sensitive SDE can minimize the negative effect of noisy samples on the optimization of diffusion models. We conduct systematic studies for both Gaussian and non-Gaussian noise distributions, providing analytical forms of risk-sensitive SDE. To verify the effectiveness of our method, we have conducted extensive experiments on multiple tabular and time-series datasets, showing that risk-sensitive SDE permits a robust optimization of diffusion models with noisy samples and significantly outperforms previous baselines.", "keywords": [ "Diffusion Models", "Noisy Data", "Risk-Sensitive SDE", "Generative Models" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 4, 1, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "azQiiSWrtx", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Benjamin Piwowarski", "gender": "Male", "institution": "CNRS / ISIR, Sorbonne Universit\u00e9", "country": "", "position": "Researcher" }, { "name": "Jo\u00e3o Maria Janeiro", "gender": "Male", "institution": "Research, Facebook", "country": "", "position": "Intern" }, { "name": "Lo\u00efc Barrault", "gender": "Male", "institution": "Meta AI", "country": "", "position": "Researcher" }, { "name": "patrick gallinari", "gender": "Male", "institution": "Criteo AI Lab", "country": "FR", "position": "Research Scientist" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 39, "n_ref": 70, "n_ref_all": 117, "n_fig": 11, "n_tab": 24, "L_tab": 14033, "n_element_tab": 1025, "n_fig_1": 1, "n_tab_1": 11, "L_tab_1": 1002, "n_element_tab_1": 138, "formula_len_all": 317, "formula_len_all_1": 200, "len_all": 229888, "len_all_1": 62031, "len_abs": 926, "len_title": 111, "len_sents": 53985, "len_sents_1": 28536, "n_sents": 474, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 942, "title": "MEXMA: Token-level objectives improve sentence representations", "abs": "Cross-lingual sentence encoders (CLSE) create fixed-size sentence representations with aligned translations. Current pre-trained CLSE approaches use sentence-level objectives only. This can lead to loss of information, especially for tokens, which then degrades the sentence representation. We propose MEXMA, a novel approach that integrates both sentence-level and token-level objectives. The sentence representation in one language is used to predict masked tokens in another language, with both the sentence representation and $\\textit{all tokens directly updating the encoder}$. We show that adding token-level objectives greatly improves the sentence representation quality across several tasks. Our approach outperforms current pre-trained cross-lingual sentence encoders on bitext mining as well as several downstream tasks. We also analyse the information encoded in our tokens, and how the sentence representation is built from them.", "keywords": [ "Natural Language Processing", "Cross lingual sentence encoders", "Language alignment", "Translations", "Representation Learning", "Translations" ], "rating_list": [ 5, 8, 5, 3 ], "soundness_list": [ 2, 4, 4, 2 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "az5WtGe48n", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Benlin Liu", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Luming Tang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Ranjay Krishna", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "Wei-Chiu Ma", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Assistant Professor" }, { "name": "Yongming Rao", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 35, "n_ref": 51, "n_ref_all": 60, "n_fig": 3, "n_tab": 4, "L_tab": 1164, "n_element_tab": 143, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 846, "n_element_tab_1": 94, "formula_len_all": 303, "formula_len_all_1": 219, "len_all": 121058, "len_all_1": 66547, "len_abs": 1645, "len_title": 61, "len_sents": 37133, "len_sents_1": 31317, "n_sents": 308, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1558, "title": "Diffusion Models are Few-shot Learners for Dense Vision Tasks", "abs": "The ability to adapt to new, unseen tasks with only a handful of training examples is a key factor behind the unprecedented success of language models. However, in computer vision, few-shot adaption has largely focused on adapting to new semantic categories or answering new visual questions. Adapting a model to dense vision tasks \u2013 depth estimation, surface normal estimation, semantic segmentation \u2013 has only been possible with large amounts of training data and with custom decoder heads, since the output spaces for each task varies widely. For instance, depth estimation outputs continuous values while semantic segmentation generates discrete categorical assignments. In this paper, we found that the diffusion prior can effectively adapt to various dense tasks, and based on this, we introduce an adaptation mechanism that exploits a pretrained diffusion model for 12 different dense vision tasks using only a few training examples. Moreover, adapting to different tasks requires only modifying the input, without changing the internal parameters of the model. Our key insight is to reframe all dense prediction tasks into a codebook-conditioned classification problem, even for continuous outputs.\nSpecifically, we learn two set of parameters: (1) concept embeddings that condition the diffusion model to encode task-specific representations in their attention masks; and (2) codebook embeddings that recombine discrete outputs to continuous ones. With this novel design, we achieve state-of-the-art results across 12 datasets for few shot learning.", "keywords": [ "Few-shot Learning; Dense Prediction; Generative Model" ], "rating_list": [ 6, 3, 8, 3, 6 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 3, 1, 3, 2, 3 ], "contribution_list": [ 3, 2, 4, 3, 3 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "ayupWYA1qD", "primary_area": "learning on time series and dynamical systems", "n_author": 7, "author_info_list": [ { "name": "Ben Cohen", "gender": "Male", "institution": "Datadog", "country": "US", "position": "Researcher" }, { "name": "Charles Masson", "gender": "unknown", "institution": "Hcompany", "country": "FR", "position": "Researcher" }, { "name": "Elise Ram\u00e9", "gender": "Female", "institution": "Datadog", "country": "FR", "position": "software engineer" }, { "name": "Emaad Khwaja", "gender": "unknown", "institution": "Datadog", "country": "US", "position": "Researcher" }, { "name": "Kan Wang", "gender": "unknown", "institution": "Datadog", "country": "US", "position": "Researcher" }, { "name": "Othmane Abou-Amal", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Youssef Doubli", "gender": "Male", "institution": "Datadog", "country": "FR", "position": "Software Engineer" } ], "n_formula": 10, "n_formula_1": 3, "n_ref_uni": 43, "n_ref": 71, "n_ref_all": 82, "n_fig": 5, "n_tab": 4, "L_tab": 2464, "n_element_tab": 325, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1662, "n_element_tab_1": 275, "formula_len_all": 484, "formula_len_all_1": 173, "len_all": 121581, "len_all_1": 57108, "len_abs": 955, "len_title": 105, "len_sents": 33227, "len_sents_1": 24990, "n_sents": 234, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 959, "title": "Toto: Time Series Optimized Transformer for Observability", "abs": "We introduce the Time Series Optimized Transformer for Observability (Toto), a foundation model designed for time series forecasting with a focus on observability metrics. Toto features a novel proportional factorized attention mechanism and a Student-T mixture model head, enabling it to efficiently handle high-dimensional, sparse, and non-stationary data. Trained on one trillion time series data points, including 75% proprietary observability data, Toto demonstrates state-of-the-art zero-shot performance on standard benchmarks such as electricity and weather forecasting. Furthermore, it significantly outperforms existing models in observability-specific tasks, making it an ideal solution for real-time system monitoring and anomaly detection. Toto\u2019s architectural innovations make it a versatile tool for both general-purpose forecasting and domain-specific applications, setting a new benchmark for scalability and accuracy in time series analysis.", "keywords": [ "Time Series Forecasting", "Proportional Factorized Attention", "Observability Metrics" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "aygBjpMdan", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chengwei Dai", "gender": "Male", "institution": "Information Institute of Engerineering, Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Kun Li", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Wei Zhou", "gender": "Female", "institution": "Institute of Information Engeering", "country": "CN", "position": "Assistant Professor" }, { "name": "Songiln Hu", "gender": "Male", "institution": "Institute of Information Engeering, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 34, "n_ref": 76, "n_ref_all": 112, "n_fig": 6, "n_tab": 26, "L_tab": 8672, "n_element_tab": 815, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3194, "n_element_tab_1": 145, "formula_len_all": 587, "formula_len_all_1": 585, "len_all": 224339, "len_all_1": 63657, "len_abs": 3438, "len_title": 149, "len_sents": 56591, "len_sents_1": 27446, "n_sents": 498, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1877, "title": "Beyond Imitation: Learning Key Reasoning Steps from Dual Chain-of-Thoughts in Reasoning Distillation", "abs": "As Large Language Models (LLMs) scale up and gain powerful Chain-of-Thoughts (CoTs) reasoning abilities, practical resource constraints drive efforts to distill these capabilities into more compact Smaller Language Models (SLMs). We find that CoTs consist mainly of simple reasoning forms, with a small proportion (~4.7%) of key reasoning steps that truly impact conclusions. However, previous distillation methods typically involve supervised fine-tuning student SLMs only on correct CoTs data produced by teacher LLMs, resulting in students struggling to learn the key reasoning steps, instead imitating the teacher's reasoning forms and making errors or omissions on these steps. To address these issues, drawing an analogy to human learning, where analyzing mistakes according to correct solutions often reveals the crucial steps leading to successes or failures, we propose mistak\\textbf{E}-\\textbf{D}riven key reason\\textbf{I}ng step distilla\\textbf{T}ion (\\textbf{EDIT}), a novel method that further aids SLMs learning key reasoning steps rather than mere simple fine-tuning. Firstly, to expose these crucial steps in CoTs, we design specific prompts to generate dual CoTs data with similar reasoning paths but divergent conclusions. Then, we apply the minimum edit distance algorithm on the dual CoTs data to locate these key steps and optimize the likelihood of these steps. Extensive experiments validate the effectiveness of EDIT across both in-domain and out-of-domain benchmark reasoning datasets. Further analysis shows that EDIT can generate high-quality CoTs with more correct key reasoning steps. Notably, we also explore how different mistake patterns affect performance and find that EDIT benefits more from logical errors than from knowledge or mathematical calculation errors in dual CoTs. Code can be found at https://anonymous.4open.science/r/eb77sh-F564", "keywords": [ "CoT distillation", "key reasoning steps", "dual CoTs" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ayg1PztmXP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Jiaqi Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Pan Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Wei Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Xiaoyi Dong", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yuanjun Xiong", "gender": "Male", "institution": "Predera.AI", "country": "US", "position": "Researcher" }, { "name": "Yuhang Zang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Zeyi Sun", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Ziyu Liu", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 41, "n_ref": 108, "n_ref_all": 132, "n_fig": 11, "n_tab": 11, "L_tab": 5690, "n_element_tab": 647, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1502, "n_element_tab_1": 124, "formula_len_all": 191, "formula_len_all_1": 191, "len_all": 194422, "len_all_1": 63742, "len_abs": 1634, "len_title": 114, "len_sents": 58651, "len_sents_1": 29758, "n_sents": 440, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1636, "title": "RAR: Retrieving And Ranking Augmented MLLMs for Visual Recognition", "abs": "CLlP (Contrastive Language lmage Pre-training) uses contrastive learning from noise image-text pairs to excel at recognizing a wide array of candidates, yet its focus on broad associations hinders the precision in distinguishing subtle differences among fine-grained items.Conversely, Multimodal Large Language Models (MLLMs) excel at classifying fine-grained categories, thanks to their substantial knowledge from pre-training on web-level corpora. However, the performance of MLLMs declines with an increase in category numbers, primarily due to growing complexity and constraints of limited context window size.To synergize the strengths of both approaches and enhance the few-shot/zero-shot recognition abilities for datasets characterized by extensive and fine-grained vocabularies, this paper introduces RAR, a Retrieving And Ranking augmented method for MLLMs. We initially establish a multi-modal retriever based on CLIP to create and store explicit memory for different categories beyond the immediate context window. During inference, RAR retrieves the top-k similar results from the memory and uses MLLMs to rank and make the final predictions. Our proposed approach not only addresses the inherent limitations in fine-grained recognition but also preserves the model's comprehensive knowledge base,significantly boosting accuracy across a range of vision-language recognition tasks. Notably, our approach demonstrates a significant improvement in performance on 5 fine-grained visual recognition benchmarks, 11 few-shot image recognition datasets, and the 2 object detection datasets under the zero-shot recognition setting.", "keywords": [ "Multimodal Large Language Models" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "aya06N6R4W", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Gael Gendron", "gender": "unknown", "institution": "University of Auckland", "country": "NZ", "position": "PhD student" }, { "name": "Gillian Dobbie", "gender": "Female", "institution": "University of Auckland", "country": "NZ", "position": "Full Professor" }, { "name": "Jose Rozanec", "gender": "Male", "institution": "Jo\u017eef Stefan Institute", "country": "SI", "position": "Postdoc" }, { "name": "Michael J. Witbrock", "gender": "Male", "institution": "University of Auckland", "country": "NZ", "position": "Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 37, "n_ref": 83, "n_ref_all": 115, "n_fig": 8, "n_tab": 9, "L_tab": 2017, "n_element_tab": 247, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1409, "n_element_tab_1": 196, "formula_len_all": 432, "formula_len_all_1": 176, "len_all": 205585, "len_all_1": 71418, "len_abs": 1361, "len_title": 78, "len_sents": 57721, "len_sents_1": 32261, "n_sents": 493, "n_sents_1": 276, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1374, "title": "Counterfactual Causal Inference in Natural Language with Large Language Models", "abs": "Causal structure discovery methods are commonly applied to structured data where the causal variables are known and where statistical testing can be used to assess the causal relationships. By contrast, recovering a causal structure from unstructured natural language data such as news articles contains numerous challenges due to the absence of known variables or counterfactual data to estimate the causal links. Large Language Models (LLMs) have shown promising results in this direction but also exhibit limitations. This work investigates LLM's abilities to build causal graphs from text documents and perform counterfactual causal inference. We propose an end-to-end causal structure discovery and causal inference method from natural language: we first use an LLM to extract the instantiated causal variables from text data and build a causal graph. We merge causal graphs from multiple data sources to represent the most exhaustive set of causes possible. We then conduct counterfactual inference on the estimated graph. The causal graph conditioning allows reduction of LLM biases and better represents the causal estimands. We use our method to show that the limitations in the counterfactual causal reasoning abilities come from prediction errors and propose directions to mitigate them. We demonstrate the applicability of our method on real-world news articles.", "keywords": [ "Causal structure discovery", "Counterfactual inference", "End-to-end", "Large Language Models" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "ayZsi8YA7h", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Brendan Juba", "gender": "Male", "institution": "Washington University in St. Louis", "country": "US", "position": "Associate Professor" }, { "name": "Luise Ge", "gender": "unknown", "institution": "Washington University, Saint Louis", "country": "US", "position": "PhD student" }, { "name": "Yevgeniy Vorobeychik", "gender": "Male", "institution": "Washington University, St. Louis", "country": "US", "position": "Full Professor" } ], "n_formula": 51, "n_formula_1": 21, "n_ref_uni": 50, "n_ref": 73, "n_ref_all": 77, "n_fig": 1, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6773, "formula_len_all_1": 2953, "len_all": 147016, "len_all_1": 67608, "len_abs": 1680, "len_title": 114, "len_sents": 44838, "len_sents_1": 30983, "n_sents": 350, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1688, "title": "Learning Linear Utility Functions From Pairwise Comparison Queries", "abs": "There is increasingly widespread use of reward model learning from human preferences to align AI systems with human values, with applications including large language models, recommendation systems, and robotic control. Nevertheless, a fundamental understanding of our ability to successfully learn utility functions in this model remains limited. We initiate this line of work by studying learnability of linear utility functions from pairwise comparison queries. In particular, we consider two learning objectives. The first objective is to predict out-of-sample responses to pairwise comparisons, whereas the second is to approximately recover the true parameters of the utility function. We show that in the passive learning setting, linear utilities are efficiently learnable with respect to the first objective, both when query responses are uncorrupted by noise, and under Tsybakov noise when the distributions are sufficiently \"nice\". In contrast, we show that utility parameters are not learnable for a large set of data distributions without strong modeling assumptions, even when query responses are noise-free. Next, we proceed to analyze the learning problem in an active learning setting. In this case, we show that even the second objective is efficiently learnable, and present algorithms for both the noise-free and noisy query response settings. This qualitative learnability gap between passive and active learning from pairwise comparisons suggests that the tendency of conventional alignment practices to simply annotate a fixed set of queries may fail to yield effective reward model estimates, an issue that can be remedied through more deliberate query selection.", "keywords": [ "Value Alignment", "Learning from pairwise comparisons", "Computational social choice" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "ayUh0A6LIJ", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Hanye Zhao", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Minghuan Liu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Weinan Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaoshen Han", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Yong Yu", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Zhengbang Zhu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" } ], "n_formula": 56, "n_formula_1": 28, "n_ref_uni": 37, "n_ref": 83, "n_ref_all": 122, "n_fig": 13, "n_tab": 7, "L_tab": 4501, "n_element_tab": 576, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1341, "n_element_tab_1": 108, "formula_len_all": 4200, "formula_len_all_1": 1836, "len_all": 193723, "len_all_1": 77965, "len_abs": 1437, "len_title": 134, "len_sents": 58525, "len_sents_1": 33271, "n_sents": 479, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1287, "title": "DyDiff: Long-Horizon Rollout via Dynamics Diffusion for Offline Reinforcement Learning", "abs": "With the great success of diffusion models (DMs) in generating realistic synthetic vision data, many researchers have investigated their potential in decision-making and control. Most of these works utilized DMs to sample directly from the trajectory space, where DMs can be viewed as a combination of dynamics models and policies. In this work, we explore how to decouple DMs\u2019 ability as dynamics models in fully offline settings, allowing the learning policy to roll out trajectories. As DMs learn the data distribution from the dataset, their intrinsic policy is actually the behavior policy induced from the dataset, which results in a mismatch between the behavior policy and the learning policy. We propose Dynamics Diffusion, short as DyDiff, which can inject information from the learning policy to DMs iteratively. DyDiff ensures long-horizon rollout accuracy while maintaining policy consistency and can be easily deployed on model-free algorithms. We provide theoretical analysis to show the advantage of DMs on long-horizon rollout over models and demonstrate the effectiveness of DyDiff in the context of offline reinforcement learning, where the rollout dataset is provided but no online environment for interaction. Our code is at https://anonymous.4open.science/r/DyDiff.", "keywords": [ "reinforcement learning", "diffusion model", "dynamics model" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 3, 2, 2, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "ayT4e9C3Gd", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Benjamin Rosman", "gender": "Male", "institution": "University of the Witwatersrand", "country": "ZA", "position": "Full Professor" }, { "name": "Geraud Nangue Tasse", "gender": "Male", "institution": "University of the Witwatersrand", "country": "", "position": "PhD student" }, { "name": "Mark Nemecek", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Affiliate" }, { "name": "Steven James", "gender": "Male", "institution": "University of the Witwatersrand", "country": "", "position": "Senior Lecturer" }, { "name": "Tamlin Love", "gender": "Male", "institution": "Universidad Polit\u00e9cnica de Cataluna", "country": "ES", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 14, "n_ref_uni": 32, "n_ref": 78, "n_ref_all": 124, "n_fig": 28, "n_tab": 7, "L_tab": 4932, "n_element_tab": 422, "n_fig_1": 19, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4534, "formula_len_all_1": 992, "len_all": 156952, "len_all_1": 68010, "len_abs": 898, "len_title": 95, "len_sents": 52867, "len_sents_1": 33221, "n_sents": 360, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 902, "title": "ROSARL: Reward-Only Safe Reinforcement Learning", "abs": "An important problem in reinforcement learning is designing agents that learn to solve tasks safely in an environment. A common solution is to define either a penalty in the reward function or a cost to be minimised when reaching unsafe states. However, designing reward or cost functions is non-trivial and can increase with the complexity of the problem. To address this, we investigate the concept of a Minmax penalty, the smallest penalty for unsafe states that leads to safe optimal policies, regardless of task rewards. We derive an upper and lower bound on this penalty by considering both environment diameter and solvability. Additionally, we propose a simple algorithm for agents to estimate this penalty while learning task policies. Our experiments demonstrate the effectiveness of this approach in enabling agents to learn safe policies in high-dimensional continuous control environments.", "keywords": [ "Reinforcement Learning", "Deep Reinforcement Learning", "Safety", "Safe AI", "Safe RL" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 2, 2, 4 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 4 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "ayPfZIkovt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Alejandro Ribeiro", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Full Professor" }, { "name": "Arnuv Tandon", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Charilaos Kanatsoulis", "gender": "unknown", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Researcher" }, { "name": "Ignacio Hounie", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 21, "n_ref_uni": 56, "n_ref": 108, "n_ref_all": 127, "n_fig": 7, "n_tab": 6, "L_tab": 1255, "n_element_tab": 136, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 506, "n_element_tab_1": 51, "formula_len_all": 1619, "formula_len_all_1": 1451, "len_all": 165827, "len_all_1": 58558, "len_abs": 1082, "len_title": 106, "len_sents": 38253, "len_sents_1": 25938, "n_sents": 307, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1085, "title": "LoRTA: Low Rank Tensor Adaptation of Large Language Models", "abs": "Low Rank Adaptation (LoRA) is a popular Parameter Efficient Fine Tuning (PEFT) method that effectively adapts large pre-trained models for downstream tasks. LoRA parameterizes model updates using low-rank matrices at each layer, significantly reducing the number of trainable parameters and, consequently, resource requirements during fine-tuning. However, the lower bound on the number of trainable parameters remains high due to the use of the low-rank matrix model. In this paper, we address this limitation by proposing a novel approach that employs a low rank tensor parametrization for model updates. The proposed low rank tensor model can significantly reduce the number of trainable parameters, while also allowing for finer-grained control over adapter size. Our experiments on Natural Language Understanding, Instruction Tuning, Preference Optimization and Protein Folding benchmarks demonstrate that our method is both efficient and effective for fine-tuning Large Language Models, achieving a reduction in the number of parameters while maintaining comparable performance.", "keywords": [ "PEFT", "LLM", "Fine-tuning", "Efficiency", "Low-rank", "Tensors" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 1 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "axyvTIt4bU", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Yingying Zhu", "gender": "Female", "institution": "Shenzhen University", "country": "CN", "position": "Full Professor" }, { "name": "Yiru Li", "gender": "Female", "institution": "Shenzhen University", "country": "CN", "position": "PhD student" }, { "name": "hongji yang", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 28, "n_ref": 74, "n_ref_all": 108, "n_fig": 9, "n_tab": 6, "L_tab": 2296, "n_element_tab": 179, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2231, "n_element_tab_1": 163, "formula_len_all": 650, "formula_len_all_1": 646, "len_all": 137477, "len_all_1": 57632, "len_abs": 1184, "len_title": 91, "len_sents": 39884, "len_sents_1": 24392, "n_sents": 308, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 1443, "title": "Retrieval-guided Cross-view Image Synthesis", "abs": "Cross-view image synthesis task aims to synthesize a photo-realistic ground-view image in correspondence with the aerial image in another view or vice versa. However, the following limitations exist: 1) existing works require extra semantic segmentation maps or preprocessing modules to bridge the domain gap. 2) the current models focus only on shared semantics in the view transformation and ignore exclusive semantics, thus performing poorly in terms of image quality and realism. 3) cross-view image synthesis for urban areas is more difficult and challenging than that of existing datasets due to the complex surroundings and building textures,the two existing datasets,however, are primarily rural and suburban scenarios. With these challenges in mind, the findings of this study can be summarized as follows: 1) a novel retrieval-guided framework, which adopts a retrieval network as the embedder to reduce the domain gap. 2) a new generator, which enhances the semantic consistency and the diversity of exclusive semantics in the target view. 3) a new dataset (named VIGOR-GEN), which offers more practical cross-view image pairs in urban areas and enriches the cross-view datasets. Extensive experiments on CVUSA, CVACT and VIGOR-GEN benchmarks verify the effectiveness of our proposed method to synthesize the photo-realistic images from the given single image in another view, outperforming the existing state-of-the-art methods.", "keywords": [ "Cross-view Image Synthesis; Domain Gap; Semantic Segmentation Maps; Retrieval" ], "rating_list": [ 3, 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 1, 3 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 4, 5, 5, 4, 4 ] }, { "paper_id": "axUf8BOjnH", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 6, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Longtao Zheng", "gender": "Male", "institution": "Nanyang Technological University, Singapore", "country": "SG", "position": "PhD student" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Xinrun Wang", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Zhenghai Xue", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Zhiyuan Huang", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 39, "n_ref": 76, "n_ref_all": 109, "n_fig": 10, "n_tab": 11, "L_tab": 2704, "n_element_tab": 332, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 629, "n_element_tab_1": 69, "formula_len_all": 466, "formula_len_all_1": 0, "len_all": 189485, "len_all_1": 68406, "len_abs": 6403, "len_title": 106, "len_sents": 60779, "len_sents_1": 35004, "n_sents": 473, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1276, "title": "AgentStudio: A Toolkit for Building General Virtual Agents", "abs": "General virtual agents need to handle multimodal observations, master complex action spaces, and self-improve in dynamic, open-domain environments. However, existing environments are often domain-specific and require complex setups, which limits agent development and evaluation in real-world settings. As a result, current evaluations lack in-depth analyses that decompose fundamental agent capabilities. We introduce AgentStudio, a trinity of environments, tools, and benchmarks to address these issues. AgentStudio provides a lightweight, interactive environment with highly generic observation and action spaces, e.g., video observations and GUI/API actions. It integrates tools for creating online benchmark tasks, annotating GUI elements, and labeling actions in videos. Based on our environment and tools, we curate an online task suite that benchmarks both GUI interactions and function calling with efficient auto-evaluation. We also reorganize existing datasets and collect new ones using our tools to establish three datasets: GroundUI, IDMBench, and CriticBench. These datasets evaluate fundamental agent abilities, including GUI grounding, learning from videos, and success detection, pointing to the desiderata for robust, general, and open-ended virtual agents.", "keywords": [ "Environment", "Benchmark", "Agent", "Digital Automation" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 2, 5, 3, 3 ] }, { "paper_id": "ax4ZOytBV2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Baoyu Jing", "gender": "unknown", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" }, { "name": "Bowen Jin", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Dongqi Fu", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Jiawei Han", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign (UIUC)", "country": "US", "position": "Full Professor" }, { "name": "Jingrui He", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Lecheng Zheng", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Yikun Ban", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Associate Professor" }, { "name": "Zihao Li", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 12, "n_ref_uni": 116, "n_ref": 159, "n_ref_all": 180, "n_fig": 5, "n_tab": 12, "L_tab": 4173, "n_element_tab": 412, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 3865, "n_element_tab_1": 256, "formula_len_all": 1141, "formula_len_all_1": 905, "len_all": 290280, "len_all_1": 77553, "len_abs": 1480, "len_title": 82, "len_sents": 47932, "len_sents_1": 31101, "n_sents": 394, "n_sents_1": 237, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1487, "title": "Multi-modal Prompt Learning Empowers Graph Neural Networks with Semantic Knowledge", "abs": "While great success has been achieved in building generalizable language models, three fundamental issues hinder GNN-based graph foundation models: the scarcity of labeled data, different levels of downstream tasks, and the conceptual gaps between domains. In depth, though the labels of real graphs are associated with semantic information, most graph learning frameworks ignore it by turning semantic labels into numerical labels. In this work, to address these issues, we present a new paradigm that leverages the text modality to align downstream tasks and data with any pre-trained GNN given only a few semantically labeled samples. Our paradigm embeds the graphs directly in the same space as the LLM by learning both graph prompts and text prompts simultaneously. To accomplish this, we improve state-of-the-art graph prompt method based on our theoretical findings. Then, we propose the first multi-modal prompt learning approach for exploiting the knowledge in pre-trained models. Notably, in our paradigm, the pre-trained GNN and the LLM are kept frozen, so the number of learnable parameters is much smaller than fine-tuning any pre-trained model. Through extensive experiments on real-world datasets, we demonstrate the superior performance of our paradigm in few-shot, multi-task-level, and cross-domain settings. Moreover, we build the first zero-shot classification prototype that can generalize GNNs to unseen classes. The code is provided in the supplementary materials.", "keywords": [ "Graph Foundation Model", "Multi-modal Prompt Learning", "Graph Neural Network", "Language Models", "Contrastive Learning" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "awz1JPyXNK", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Guipeng Lan", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Jiabao Wen", "gender": "Male", "institution": "Tianjin University", "country": "", "position": "Associate Professor" }, { "name": "Jiachen Yang", "gender": "Male", "institution": "Tianjin University", "country": "", "position": "Full Professor" }, { "name": "Shuai Xiao", "gender": "Male", "institution": "Tianjin University", "country": "", "position": "Associate Professor" }, { "name": "Xi Meng", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "Postdoc" } ], "n_formula": 47, "n_formula_1": 23, "n_ref_uni": 36, "n_ref": 50, "n_ref_all": 67, "n_fig": 5, "n_tab": 3, "L_tab": 519, "n_element_tab": 49, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 208, "n_element_tab_1": 17, "formula_len_all": 6958, "formula_len_all_1": 1662, "len_all": 161426, "len_all_1": 65035, "len_abs": 1674, "len_title": 127, "len_sents": 52118, "len_sents_1": 29347, "n_sents": 411, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1401, "title": "Inner Information Analysis Algorithm for Deep Neural Network based on Community", "abs": "Deep learning has achieved advancements across a variety of forefront fields. However, its inherent 'black box' characteristic poses challenges to the comprehension and trustworthiness of the decision-making processes within neural networks. To mitigate these challenges, we introduce InnerSightNet, an inner information analysis algorithm designed to illuminate the inner workings of deep neural networks through the perspectives of community. This approach is aimed at deciphering the intricate patterns of neurons within deep neural networks, thereby shedding light on the networks' information processing and decision-making pathways. InnerSightNet operates in three primary phases, 'neuronization-aggregation-evaluation'. Initially, it transforms learnable units into a structured network of neurons. Subsequently, these neurons are aggregated into distinct communities according to representation attributes. The final phase involves the evaluation of these communities' roles and functionalities, to unpick the information flow and decision-making. By transcending focus on single-layer or individual neuron, InnerSightNet broadens the horizon for deep neural network interpretation. InnerSightNet offers a unique vantage point, enabling insights into the collective behavior of communities within the overarching architecture, thereby enhancing transparency and trust in deep learning systems.", "keywords": [ "inner information analysis", "transparency", "knowledge mining" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "awvJBtB2op", "primary_area": "applications to robotics, autonomy, planning", "n_author": 3, "author_info_list": [ { "name": "Lingji Kong", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "US", "position": "MS student" }, { "name": "Muhan Li", "gender": "Male", "institution": "Northwestern University", "country": "", "position": "Researcher" }, { "name": "Sam Kriegman", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 8, "n_ref_uni": 36, "n_ref": 63, "n_ref_all": 120, "n_fig": 38, "n_tab": 1, "L_tab": 433, "n_element_tab": 28, "n_fig_1": 23, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1147, "formula_len_all_1": 559, "len_all": 140135, "len_all_1": 63101, "len_abs": 418, "len_title": 87, "len_sents": 47683, "len_sents_1": 31219, "n_sents": 313, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 39, "L_abs": 1450, "title": "Generating Freeform Endoskeletal Robots", "abs": "The automatic design of embodied agents (e.g. robots) has existed for 31 years and is experiencing a renaissance of interest in the literature. To date however, the field has remained narrowly focused on two kinds of anatomically simple robots: (1) fully rigid, jointed bodies; and (2) fully soft, jointless bodies. Here we bridge these two extremes with the open ended creation of terrestrial endoskeletal robots: deformable soft bodies that leverage jointed internal skeletons to move efficiently across land. Simultaneous de novo generation of external and internal structures is achieved by (i) modeling 3D endoskeletal body plans as integrated collections of elastic and rigid cells that directly attach to form soft tissues anchored to compound rigid bodies; (ii) encoding these discrete mechanical subsystems into a continuous yet coherent latent embedding; (iii) optimizing the sensorimotor coordination of each decoded design using model-free reinforcement learning; and (iv) navigating this smooth yet highly non-convex latent manifold using evolutionary strategies. This yields an endless stream of novel species of ``higher robots'' that, like all higher animals, harness the mechanical advantages of both elastic tissues and skeletal levers for terrestrial travel. It also provides a plug-and-play experimental platform for benchmarking evolutionary design and representation learning algorithms in complex hierarchical embodied systems.", "keywords": [ "co-design", "agent design", "robots", "morphology", "evolution", "locomotion" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 4, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 4, 4, 2, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "awuw503LzY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Kate Saenko", "gender": "Female", "institution": "Meta AI", "country": "US", "position": "Researcher" }, { "name": "Maan Qraitem", "gender": "unknown", "institution": "Boston University", "country": "US", "position": "PhD student" }, { "name": "Piotr Teterwak", "gender": "Male", "institution": "Boston University", "country": "US", "position": "PhD student" }, { "name": "Bryan Allen Plummer", "gender": "Male", "institution": "Boston University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 20, "n_ref": 43, "n_ref_all": 71, "n_fig": 13, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 139, "formula_len_all_1": 104, "len_all": 97961, "len_all_1": 59068, "len_abs": 1582, "len_title": 115, "len_sents": 36570, "len_sents_1": 29518, "n_sents": 302, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1593, "title": "Hidden Logos in Web-Scale Data Disrupt Large Vision Language Models", "abs": "Vision-Language Models are trained on very large, minimally curated image datasets that contain many spurious correlations between categories and visual patterns. This causes VLMs to learn shortcuts, e.g., between smiling and gender. Although logos are ubiquitous in VLM training data and are a potential source of such shortcuts, there is very limited study of this issue. Prior work pointed out that logos may indeed cause such problems, but the analysis was limited to a single text-based logo. In this paper, we undertake a broad study of logos in VLM training data and their potential to insert \"hidden\" spurious correlations into VLMs. We construct a new logo dataset, CC12M-LogoBank, propose an algorithm that uncovers spurious logos affecting a given VLM prediction task, and test it on several representative tasks: person attribute classification, object classification, and harmful content detection. Our key finding is that some logos indeed lead to spurious incorrect predictions, for example, adding the Adidas logo to a photo of a person causes a model classify the person as \"greedy\". Furthermore, we argue that the uncovered logos can be seen as effective attacks against foundational models; for example, an attacker could place a spurious logo on harmful content, causing the model to misclassify it as harmless. This threat is alarming considering the simplicity of logo attacks, increasing the attack surface of VLM models. As a defense, we explore two effective yet simple mitigation strategies that seamlessly integrate with zero-shot inference of foundation models.", "keywords": [ "Spurious Correlations", "Bias", "Logos" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "awtd0XhzKQ", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 5, "author_info_list": [ { "name": "Erik Arakelyan", "gender": "Male", "institution": "NVIDIA", "country": "AM", "position": "Researcher" }, { "name": "Isabelle Augenstein", "gender": "Female", "institution": "University of Copenhagen", "country": "DK", "position": "Full Professor" }, { "name": "Pat Verga", "gender": "unknown", "institution": "Cohere", "country": "", "position": "Researcher" }, { "name": "Patrick Lewis", "gender": "Male", "institution": "University College London", "country": "", "position": "PhD student" }, { "name": "Pasquale Minervini", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 45, "n_ref": 108, "n_ref_all": 134, "n_fig": 4, "n_tab": 7, "L_tab": 2108, "n_element_tab": 255, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 811, "n_element_tab_1": 158, "formula_len_all": 615, "formula_len_all_1": 483, "len_all": 225471, "len_all_1": 63805, "len_abs": 1907, "len_title": 101, "len_sents": 39215, "len_sents_1": 28273, "n_sents": 265, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 2026, "title": "FLARE: Faithful Logic-Aided Reasoning and Exploration", "abs": "Modern Question Answering (QA) and Reasoning approaches based on Large Language Models (LLMs) commonly use prompting techniques, such as Chain-of-Thought (CoT), assuming the resulting generation will have a more granular exploration and reasoning over the question space and scope. However, such methods struggle with generating outputs that are faithful to the intermediate chain of reasoning produced by the model. On the other end of the spectrum, neuro-symbolic methods such as Faithful CoT (F-CoT) and Logic-LM propose to combine LLMs with external symbolic solvers. While such approaches boast a high degree of faithfulness, they usually require a model trained for code generation and struggle with tasks that are ambiguous or hard to formalise strictly. We introduce $\\textbf{F}$aithful $\\textbf{L}$ogic-$\\textbf{A}$ided $\\textbf{R}$easoning and $\\textbf{E}$xploration ($\\textbf{FLARE}$), a novel interpretable approach for traversing the problem space using task decompositions. We use the LLM to plan a solution, formalise the query into facts and predicates, which form the problem space, using a logic programming code and simulate that code execution using an exhaustive multi-hop search over the defined space. Our method allows us to compute the faithfulness of the reasoning process w.r.t. the generated code and explicitly trace the steps of the multi-hop search without relying on external solvers. Our methods achieve SOTA results on $\\mathbf{7}$ out of $\\mathbf{9}$ diverse reasoning benchmarks. We also show that model faithfulness positively correlates with overall performance and further demonstrate that ${\\textbf{FLARE}}$ allows pinpointing the decisive factors sufficient for and leading to the correct answer with optimal reasoning during the multi-hop search. Our findings reveal that successful traces exhibit, on average, a $18.1\\%$ increase in unique emergent facts, a $8.6\\%$ higher overlap between code-defined and execution-trace relations, and a $3.6\\%$ reduction in unused code relations.", "keywords": [ "Reasoning", "LLM", "multi-hop", "complex", "prompting", "faithfulness", "logic", "symbolic", "search" ], "rating_list": [ 8, 6, 6, 3 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "awWiNvQwf3", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 14, "author_info_list": [ { "name": "Alan Aspuru-Guzik", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Director" }, { "name": "Chao Zhang", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Chenru Duan", "gender": "Male", "institution": "Deep Principle", "country": "US", "position": "Researcher" }, { "name": "Cher Tian Ser", "gender": "unknown", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Felix Strieth-Kalthoff", "gender": "Male", "institution": "University of Wuppertal", "country": "DE", "position": "Assistant Professor" }, { "name": "Haorui Wang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Kirill Neklyudov", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Assistant Professor" }, { "name": "Lingkai Kong", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Marta Skreta", "gender": "Female", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Wenhao Gao", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Yanqiao Zhu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Yuanqi Du", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Yuchen Zhuang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Yue Yu", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 76, "n_ref": 126, "n_ref_all": 172, "n_fig": 9, "n_tab": 15, "L_tab": 9215, "n_element_tab": 605, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 488, "n_element_tab_1": 35, "formula_len_all": 1027, "formula_len_all_1": 376, "len_all": 250638, "len_all_1": 66107, "len_abs": 1428, "len_title": 124, "len_sents": 68341, "len_sents_1": 33998, "n_sents": 457, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1146, "title": "Efficient Evolutionary Search Over Chemical Space with Large Language Models", "abs": "Molecular discovery, when formulated as an optimization problem, presents significant computational challenges because optimization objectives can be non-differentiable. Evolutionary Algorithms (EAs), often used to optimize black-box objectives in molecular discovery, traverse chemical space by performing random mutations and crossovers, leading to a large number of expensive objective evaluations. In this work, we ameliorate this shortcoming by incorporating chemistry-aware Large Language Models (LLMs) into EAs. Namely, we redesign crossover and mutation operations in EAs using LLMs trained on large corpora of chemical information. We perform extensive empirical studies on both commercial and open-source models on multiple tasks involving property optimization, molecular rediscovery, and structure-based drug design, demonstrating that the joint usage of LLMs with EAs yields superior performance over all baseline models across single- and multi-objective settings. We demonstrate that our algorithm improves both the quality of the final solution and convergence speed, thereby reducing the number of required objective evaluations.", "keywords": [ "Large Language Models", "Evolutionary Search", "Molecule Optimization", "AI for Science", "Molecular generation" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "awReGYZaGl", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Haoru Tan", "gender": "Male", "institution": "HKU", "country": "HK", "position": "PhD student" }, { "name": "Jiahui Liu", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Intern" }, { "name": "Shizhen Zhao", "gender": "Male", "institution": "The University of Hong Kong,", "country": "HK", "position": "PhD student" }, { "name": "XIAOJUAN QI", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Xin Wen", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 47, "n_ref": 100, "n_ref_all": 115, "n_fig": 10, "n_tab": 7, "L_tab": 4006, "n_element_tab": 392, "n_fig_1": 10, "n_tab_1": 6, "L_tab_1": 3825, "n_element_tab_1": 365, "formula_len_all": 713, "formula_len_all_1": 713, "len_all": 141394, "len_all_1": 70368, "len_abs": 1591, "len_title": 116, "len_sents": 35808, "len_sents_1": 28178, "n_sents": 310, "n_sents_1": 234, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1610, "title": "Rethinking Out-of-Distribution Detection in Vision Foundation Models", "abs": "Pre-trained vision foundation models have transformed many computer vision tasks. Despite their strong ability to learn discriminative and generalizable features-- crucial for out-of-distribution (OOD) detection, their impact on this task remains underexplored. Motivated by this gap, our study investigates vision foundation models in OOD detection. Our findings show that even without complex designs, a pre-trained DINOv2 model, utilizing a simple scoring metric and no fine-tuning, outperforms all prior state-of-the-art models, which typically depend on fine-tuning with in-distribution (ID) data. Furthermore, while the pre-trained CLIP model struggles with fine-grained OOD samples, DINOv2 excels, revealing the limitations of CLIP in this setting. Building on these insights, we explore how foundation models can be further optimized for both ID classification and OOD detection when ID data is available for fine-tuning. From a model perspective, we propose a Mixture of Feature Experts (MoFE) module, which partitions features into subspaces. This mitigates the challenge of tuning complex data distributions with limited ID data and enhances decision boundary learning for classification. From a data perspective, we introduce a Dynamic-$\\beta$ Mixup strategy, which samples interpolation weights from a dynamic beta distribution. This adapts to varying levels of learning difficulty across categories, improving feature learning for more challenging categories. Extensive experiments and ablation studies demonstrate the effectiveness of our approach, significantly outperforming baseline methods.", "keywords": [ "Out-of-Distribution Detection; Vision Foundation Model; Mixture of Experts" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "aw2Jc5DFZC", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Aarti Singh", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Dhruv Malik", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Jeff Schneider", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Researcher" }, { "name": "Youngseog Chung", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Yuanzhi Li", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 7, "n_ref_uni": 37, "n_ref": 77, "n_ref_all": 118, "n_fig": 7, "n_tab": 7, "L_tab": 1033, "n_element_tab": 88, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 41, "n_element_tab_1": 12, "formula_len_all": 2486, "formula_len_all_1": 1021, "len_all": 172732, "len_all_1": 69988, "len_abs": 2038, "len_title": 112, "len_sents": 56154, "len_sents_1": 34142, "n_sents": 427, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 2056, "title": "Beyond Parameter Count: Implicit Bias in Soft Mixture of Experts", "abs": "The traditional viewpoint on Sparse Mixture of Experts (MoE) models is that instead of training a single _large_ expert, which is computationally expensive, we can train many _small_ experts. The hope is that if the total parameter count of the small experts equals that of the singular large expert, then we retain the representation power of the large expert while gaining computational tractability and promoting expert specialization. The recently introduced Soft MoE replaces the Sparse MoE's discrete routing mechanism with a differentiable gating function that smoothly mixes tokens. While this smooth gating function successfully mitigates the various training instabilities associated with Sparse MoE, it is unclear whether it induces implicit biases that affect Soft MoE's representation power or potential for expert specialization. We prove that Soft MoE with a single arbitrarily powerful expert cannot represent simple convex functions. This justifies that Soft MoE's success cannot be explained by the traditional viewpoint of many small experts collectively mimicking the representation power of a single large expert, and that multiple experts are actually _necessary_ to achieve good representation power (even for a fixed total parameter count). Continuing along this line of investigation, we introduce a notion of expert specialization for Soft MoE, and while varying the number of experts yet fixing the total parameter count, we consider the following (computationally intractable) task. Given any input, how can we discover the expert subset that is specialized to predict this input's label? We empirically show that when there are many small experts, the architecture is implicitly biased in a fashion that allows us to efficiently approximate the specialized expert subset. Our method can be easily implemented to potentially reduce computation during inference. For example, using our method on ImageNet, one can perform inference using only $1/8$ of the experts and still retain $99$% of the test accuracy of using all experts.", "keywords": [ "mixture of experts", "conditional computation", "sparse activation" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "avlfmW32qO", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Han Sangbum", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Nojun Kwak", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Sangyu Han", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Yerim Kim", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 10, "n_ref_uni": 24, "n_ref": 63, "n_ref_all": 87, "n_fig": 9, "n_tab": 3, "L_tab": 1312, "n_element_tab": 71, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 507, "n_element_tab_1": 49, "formula_len_all": 897, "formula_len_all_1": 555, "len_all": 128712, "len_all_1": 61126, "len_abs": 1194, "len_title": 158, "len_sents": 44983, "len_sents_1": 28107, "n_sents": 353, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1253, "title": "Decompose the model: mechanistic interpretability in image models with generalized integrated gradients (GIG)", "abs": "In the field of eXplainable AI (XAI) in language models, the progression from local explanations of individual decisions to global explanations with high-level concepts has laid the groundwork for mechanistic interpretability, which aims to decode the exact operations.\nHowever, this paradigm has not been adequately explored in image models, where existing methods have primarily focused on class-specific interpretations.\nThis paper introduces a novel approach to systematically trace the entire pathway from input through all intermediate layers to the final output within the whole dataset.\nWe utilize Pointwise Feature Vectors (PFVs) and instance-specific Effective Receptive Fields (iERFs) to decompose model embeddings into interpretable Concept Vectors.\nThen, we calculate the relevance between concept vectors with our Generalized Integrated Gradients (GIG), enabling a comprehensive, dataset-wide analysis of model behavior.\nWe validate our method of concept extraction and concept attribution in both qualitative and quantitative evaluations.\nOur approach advances the understanding of semantic significance within image models, offering a holistic view of their operational mechanics. https://iclr2025gig.netlify.app/graph_visualization.html", "keywords": [ "explainable ai", "interpretability", "mechanistic interpretability" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "avSocG0oFA", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Christos Thrampoulidis", "gender": "unknown", "institution": "University of British Columbia", "country": "", "position": "Assistant Professor" }, { "name": "Minghui Chen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "Vala Vakilian", "gender": "Male", "institution": "University of British Columbia", "country": "", "position": "MS student" }, { "name": "Wenlong Deng", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Xiaoxiao Li", "gender": "Unspecified", "institution": "University of British Columbia", "country": "CA", "position": "Assistant Professor" }, { "name": "Yize Zhao", "gender": "Female", "institution": "University of British Columbia", "country": "CA", "position": "MS student" } ], "n_formula": 35, "n_formula_1": 10, "n_ref_uni": 42, "n_ref": 98, "n_ref_all": 172, "n_fig": 14, "n_tab": 10, "L_tab": 4193, "n_element_tab": 302, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2406, "formula_len_all_1": 607, "len_all": 219675, "len_all_1": 61957, "len_abs": 1564, "len_title": 104, "len_sents": 66323, "len_sents_1": 29098, "n_sents": 559, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 56, "L_abs": 1569, "title": "Revisiting Delta-Parameter Pruning For Fine-Tuned Models", "abs": "Storing open-source fine-tuned models separately introduces redundancy and increases response times in applications utilizing multiple models. Delta-parameter pruning (DPP), particularly the random drop and rescale (DARE) method proposed by Yu et al., addresses this by pruning the majority of delta parameters\u2014the differences between fine-tuned and pre-trained model weights\u2014while typically maintaining minimal performance loss. However, DARE fails when either the pruning rate or the magnitude of the delta parameters is large. We highlight two key reasons for this failure: (1) an excessively large rescaling factor as pruning rates increase, and (2) high mean and variance in the delta parameters.\nTo address these, we develop two algorithmic improvements: (1) DARq, which modifies the rescaling factor in DARE, leading to significant performance gains at high pruning rates (e.g., >30% on COLA and SST2 for encoder models, with even larger improvements in decoder models), and (2) AdamR, an in-training modification that incorporates appropriate Delta regularization before applying DPP. We also demonstrate that DARq can be seamlessly combined with vanilla parameter-efficient fine-tuning techniques like LoRA and can facilitate structural DPP. Additionally, we revisit the application of importance-based pruning techniques within DPP, demonstrating that they outperform random-based methods when delta parameters are large. Through this comprehensive study, we develop a pipeline for selecting the most appropriate DPP method under various practical scenarios.", "keywords": [ "Delta parameter pruning", "Efficiency", "Large Language Models" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "avNVrQ8D2v", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chao Ma", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Fei Xie", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CX", "position": "PhD student" }, { "name": "Lu Qi", "gender": "Male", "institution": "University of California, Merced", "country": "US", "position": "Postdoc" }, { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Xiangtai Li", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Yujin Tang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Intern" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 31, "n_ref": 78, "n_ref_all": 81, "n_fig": 0, "n_tab": 16, "L_tab": 5030, "n_element_tab": 596, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 3676, "n_element_tab_1": 415, "formula_len_all": 299, "formula_len_all_1": 187, "len_all": 129333, "len_all_1": 57128, "len_abs": 1464, "len_title": 173, "len_sents": 35673, "len_sents_1": 22518, "n_sents": 278, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1474, "title": "PredFormer: Transformers Are Effective Spatial-Temporal Predictive Learners", "abs": "Spatiotemporal predictive learning methods generally fall into two categories: recurrent-based approaches, which face challenges in parallelization and performance, and recurrent-free methods, which employ convolutional neural networks (CNNs) as encoder-decoder architectures. These methods benefit from strong inductive biases but often at the expense of scalability and generalization. This paper proposes PredFormer, a pure transformer-based framework for spatiotemporal predictive learning. Motivated by the Vision Transformers (ViT) design, PredFormer leverages carefully designed Gated Transformer blocks, following a comprehensive analysis of 3D attention mechanisms, including full-, factorized-, and interleaved- spatial-temporal attention. With its recurrent-free, transformer-based design, PredFormer is both simple and efficient, significantly outperforming previous methods by large margins. Extensive experiments on synthetic and real-world datasets demonstrate that PredFormer achieves state-of-the-art performance. On Moving MNIST, PredFormer achieves a 51.3% reduction in MSE relative to SimVP. For TaxiBJ, the model decreases MSE by 33.1% and boosts FPS from 533 to 2364. Additionally, on WeatherBench, it reduces MSE by 11.1% while enhancing FPS from 196 to 404. These performance gains in both accuracy and efficiency demonstrate PredFormer's potential for real-world applications. The source code and trained models will be made available to the public.", "keywords": [ "Video Prediction", "Spatio-temporal Prective Learning", "Spatio-temporal Forecasting" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "aueXfY0Clv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Alexey Bochkovskiy", "gender": "Male", "institution": "Apple", "country": "DE", "position": "Machine Learning Engineer" }, { "name": "Ama\u00ebl Delaunoy", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Hugo Germain", "gender": "unknown", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Marcel Santos", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Stephan Richter", "gender": "Male", "institution": "Apple", "country": "DE", "position": "Researcher" }, { "name": "Vladlen Koltun", "gender": "Male", "institution": "Apple", "country": "US", "position": "Distinguished Scientist" }, { "name": "Yichao Zhou", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 6, "n_ref_uni": 72, "n_ref": 161, "n_ref_all": 187, "n_fig": 10, "n_tab": 11, "L_tab": 7616, "n_element_tab": 1057, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 976, "n_element_tab_1": 123, "formula_len_all": 829, "formula_len_all_1": 303, "len_all": 263509, "len_all_1": 61795, "len_abs": 1590, "len_title": 47, "len_sents": 58222, "len_sents_1": 29085, "n_sents": 480, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 967, "title": "Depth Pro: Sharp Monocular Metric Depth in Less Than a Second", "abs": "We present a foundation model for zero-shot metric monocular depth estimation. Our model, Depth Pro, synthesizes high-resolution depth maps with unparalleled sharpness and high-frequency details. The predictions are metric, with absolute scale, without relying on the availability of metadata such as camera intrinsics. And the model is fast, producing a 2.25-megapixel depth map in 0.3 seconds on a standard GPU. These characteristics are enabled by a number of technical contributions, including an efficient multi-scale vision transformer for dense prediction, a training protocol that combines real and synthetic datasets to achieve high metric accuracy alongside fine boundary tracing, dedicated evaluation metrics for boundary accuracy in estimated depth maps, and state-of-the-art focal length estimation from a single image. Extensive experiments analyze specific design choices and demonstrate that Depth Pro outperforms prior work along multiple dimensions.", "keywords": [ "depth estimation", "computer vision" ], "rating_list": [ 6, 8, 3, 10, 6 ], "soundness_list": [ 4, 4, 3, 4, 3 ], "presentation_list": [ 4, 4, 3, 4, 2 ], "contribution_list": [ 3, 4, 2, 4, 3 ], "confidence_list": [ 4, 5, 4, 5, 2 ] }, { "paper_id": "aucMP9hGYv", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Bangzhen Liu", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Chenxi Zheng", "gender": "unknown", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Shengfeng He", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Associate Professor" }, { "name": "Xuemiao Xu", "gender": "Female", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yihong Lin", "gender": "unknown", "institution": "South China University of Technology", "country": "CN", "position": "MS student" }, { "name": "Yongwei Nie", "gender": "unknown", "institution": "South China University of Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 45, "n_formula_1": 18, "n_ref_uni": 48, "n_ref": 110, "n_ref_all": 143, "n_fig": 17, "n_tab": 8, "L_tab": 1119, "n_element_tab": 105, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 623, "n_element_tab_1": 46, "formula_len_all": 3178, "formula_len_all_1": 1311, "len_all": 219964, "len_all_1": 59501, "len_abs": 1809, "len_title": 123, "len_sents": 70121, "len_sents_1": 26915, "n_sents": 583, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1816, "title": "RecDreamer: Consistent Text-to-3D Generation via Uniform Score Distillation", "abs": "Current text-to-3D generation methods based on score distillation often suffer from geometric inconsistencies, leading to repeated patterns across different poses of 3D assets. This issue, known as the Multi-Face Janus problem, arises because existing methods struggle to maintain consistency across varying poses and are biased toward a canonical pose. While recent work has improved pose control and approximation, these efforts are still limited by this inherent bias, which skews the guidance during generation.\nTo address this, we propose a solution called RecDreamer, which reshapes the underlying data distribution to achieve more consistent pose representation. The core idea behind our method is to rectify the prior distribution, ensuring that pose variation is uniformly distributed rather than biased toward a canonical form. By modifying the prescribed distribution through an auxiliary function, we can reconstruct the density of the distribution to ensure compliance with specific marginal constraints. In particular, we ensure that the marginal distribution of poses follows a uniform distribution, thereby eliminating the biases introduced by the prior knowledge.\nWe incorporate this rectified data distribution into existing score distillation algorithms, a process we refer to as uniform score distillation. To efficiently compute the posterior distribution required for the auxiliary function, RecDreamer introduces a training-free classifier that estimates pose categories in a plug-and-play manner. Additionally, we utilize various approximation techniques for noisy states, significantly improving system performance.\nOur experimental results demonstrate that RecDreamer effectively mitigates the Multi-Face Janus problem, leading to more consistent 3D asset generation across different poses.", "keywords": [ "3D generation" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "auZZ2gN0ZN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Anurag Arnab", "gender": "Unspecified", "institution": "Google", "country": "FR", "position": "Research Scientist" }, { "name": "Chen Sun", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Assistant Professor" }, { "name": "Cordelia Schmid", "gender": "Female", "institution": "Google", "country": "FR", "position": "Researcher" }, { "name": "Xingyi Zhou", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 3, "n_ref_uni": 65, "n_ref": 181, "n_ref_all": 198, "n_fig": 4, "n_tab": 14, "L_tab": 3413, "n_element_tab": 495, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1392, "n_element_tab_1": 84, "formula_len_all": 732, "formula_len_all_1": 239, "len_all": 208159, "len_all_1": 77616, "len_abs": 1384, "len_title": 103, "len_sents": 52815, "len_sents_1": 35910, "n_sents": 413, "n_sents_1": 274, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 55, "L_abs": 1420, "title": "Dense Video Object Captioning from Disjoint Supervision", "abs": "We propose a new task and model for dense video object captioning -- detecting, tracking and captioning trajectories of objects in a video. This task unifies spatial and temporal localization in video, whilst also requiring fine-grained visual understanding that is best described by natural language. We propose a unified model, and demonstrate how our end-to-end approach is more accurate and temporally coherent than a multi-stage pipeline combining state-of-the-art detection, tracking, and captioning models. Moreover, we propose a training strategy based on a mixture of disjoint tasks, which allows us to leverage diverse, large-scale datasets which supervise different parts of our model. Although each pretraining task only provides weak supervision, they are complementary and, when combined, result in noteworthy zero-shot ability and serve as strong initialization for additional finetuning to further improve accuracy. We carefully design new metrics capturing all components of our task, and show how we can repurpose existing video grounding datasets (e.g. VidSTG and VLN) for our new task. We show that our model improves upon a number of strong baselines for this new task. Furthermore, we can apply our model to the task of spatial grounding, outperforming prior state-of-the-art on VidSTG and VLN, without explicitly training for it. Our code is available at https://github.com/google-research/scenic.", "keywords": [ "object captioning", "video", "tracking" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 4, 3, 4, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "atXCzVSXTJ", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Chunping Wang", "gender": "Female", "institution": "Finvolution Group", "country": "CN", "position": "Principal Scientist" }, { "name": "Lei CHEN", "gender": "Male", "institution": "Finvolution Group", "country": "CN", "position": "VP" }, { "name": "Taoran Fang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Tianhong Gao", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Undergrad student" }, { "name": "Yang Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "YihaoShang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Wei Chou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 72, "n_formula_1": 18, "n_ref_uni": 45, "n_ref": 132, "n_ref_all": 142, "n_fig": 1, "n_tab": 7, "L_tab": 3778, "n_element_tab": 329, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1507, "n_element_tab_1": 146, "formula_len_all": 7169, "formula_len_all_1": 1245, "len_all": 186472, "len_all_1": 68345, "len_abs": 1447, "len_title": 126, "len_sents": 53988, "len_sents_1": 30299, "n_sents": 431, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1452, "title": "KAA: Kolmogorov-Arnold Attention for Enhancing Attentive Graph Neural Networks", "abs": "Graph neural networks (GNNs) with attention mechanisms, often referred to as attentive GNNs, have emerged as a prominent paradigm in advanced GNN models in recent years. However, our understanding of the critical process of scoring neighbor nodes remains limited, leading to the underperformance of many existing attentive GNNs. In this paper, we unify the scoring functions of current attentive GNNs and propose Kolmogorov-Arnold Attention (KAA), which integrates the Kolmogorov-Arnold Network (KAN) architecture into the scoring process. KAA enhances the performance of scoring functions across the board and can be applied to nearly all existing attentive GNNs. To compare the expressive power of KAA with other scoring functions, we introduce Maximum Ranking Distance (MRD) to quantitatively estimate their upper bounds in ranking errors for node importance. Our analysis reveals that, under limited parameters and constraints on width and depth, both linear transformation-based and MLP-based scoring functions exhibit finite expressive power. In contrast, our proposed KAA, even with a single-layer KAN parameterized by zero-order B-spline functions, demonstrates nearly infinite expressive power. Extensive experiments on both node-level and graph-level tasks using various backbone models show that KAA-enhanced scoring functions consistently outperform their original counterparts, achieving performance improvements of over 20% in some cases.", "keywords": [ "Kolmogorov-Arnold Networks", "Attentive Graph Neural Networks" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "asR9FVd4eL", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Fengwang Li", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "Assistant Professor" }, { "name": "Runqi Lin", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 53, "n_ref": 98, "n_ref_all": 140, "n_fig": 5, "n_tab": 16, "L_tab": 2964, "n_element_tab": 237, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2512, "n_element_tab_1": 155, "formula_len_all": 86, "formula_len_all_1": 64, "len_all": 201991, "len_all_1": 66525, "len_abs": 1936, "len_title": 120, "len_sents": 66400, "len_sents_1": 32108, "n_sents": 460, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1456, "title": "Understanding and Enhancing the Transferability of Jailbreaking Attacks", "abs": "Jailbreaking attacks can effectively manipulate open-source large language models (LLMs) to produce harmful responses. However, these attacks exhibit limited transferability, failing to disrupt proprietary LLMs consistently. To reliably identify vulnerabilities in proprietary LLMs, this work investigates the transferability of jailbreaking attacks by analysing their impact on the model's intent perception. By incorporating adversarial sequences, these attacks can redirect the source LLM's focus away from malicious-intent tokens in the original input, thereby obstructing the model's intent recognition and eliciting harmful responses. Nevertheless, these adversarial sequences fail to mislead the target LLM's intent perception, allowing the target LLM to refocus on malicious-intent tokens and abstain from responding. Our analysis further reveals the inherent $\\textit{distributional dependency}$ within the generated adversarial sequences, whose effectiveness stems from overfitting the source LLM's parameters, resulting in limited transferability to target LLMs. To this end, we propose the Perceived-importance Flatten (PiF) method, which uniformly disperses the model's focus across neutral-intent tokens in the original input, thus obscuring malicious-intent tokens without relying on overfitted adversarial sequences. Extensive experiments demonstrate that PiF provides an effective and efficient red-teaming evaluation for proprietary LLMs.", "keywords": [ "Jailbreaking Attack", "Black-box Transferable Attack", "Large Language Model", "Red-teaming Evaluation" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "asGQQc7gNo", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Baolong Bi", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Hongcheng Gao", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Junfeng Fang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Lingrui Mei", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Shenghua Liu", "gender": "Not Specified", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Shiyu Ni", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Xueqi Cheng", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy", "country": "CN", "position": "Full Professor" }, { "name": "Yiwei Wang", "gender": "Male", "institution": "University of California, Merced", "country": "US", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 44, "n_ref": 132, "n_ref_all": 173, "n_fig": 6, "n_tab": 5, "L_tab": 1044, "n_element_tab": 122, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 530, "n_element_tab_1": 68, "formula_len_all": 206, "formula_len_all_1": 231, "len_all": 240720, "len_all_1": 58163, "len_abs": 1753, "len_title": 153, "len_sents": 68916, "len_sents_1": 28264, "n_sents": 724, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 1630, "title": "Is Factuality Enhancement a Free Lunch For LLMs? Better Factuality Can Lead to Worse Context-Faithfulness", "abs": "As the modern tools of choice for text understanding and generation, large language models (LLMs) are expected to accurately output answers by leveraging the input context.\nThis requires LLMs to possess both context-faithfulness and factual accuracy.\nExtensive efforts have been made to enable better outputs from LLMs by mitigating hallucinations through factuality enhancement methods.\nHowever, they also pose risks of hindering context-faithfulness, as factuality enhancement can lead LLMs to become overly confident in their parametric knowledge, causing them to overlook the relevant input context.\nIn this work, we argue that current factuality enhancement methods can significantly undermine the context-faithfulness of LLMs.\nWe first revisit the current factuality enhancement methods and evaluate their effectiveness in enhancing factual accuracy.\nNext, we evaluate their performance on knowledge editing tasks to assess the potential impact on context-faithfulness.\nThe experimental results reveal that while these methods may yield inconsistent improvements in factual accuracy, they also cause a more severe decline in context-faithfulness, with the largest decrease reaching a striking 69.7\\%.\nTo explain these declines, we analyze the hidden states and logit distributions for the tokens representing new knowledge and parametric knowledge respectively, highlighting the limitations of current approaches.\nOur finding highlights the complex trade-offs inherent in enhancing LLMs.\nTherefore, we recommend that more research on LLMs' factuality enhancement make efforts to reduce the sacrifice of context-faithfulness.", "keywords": [ "Large Language Models", "In-Context Faithfulness", "Factuality Enhancement" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "asA7vvsgcI", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Evangelia Spiliopoulou", "gender": "Female", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Gyuwan Kim", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Jie Ma", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Miguel Ballesteros", "gender": "Male", "institution": "Amazon", "country": "", "position": "Principal Applied Scientist" }, { "name": "Yang Li", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "William Yang Wang", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 61, "n_ref": 121, "n_ref_all": 135, "n_fig": 3, "n_tab": 1, "L_tab": 391, "n_element_tab": 104, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1223, "n_element_tab_1": 30, "formula_len_all": 351, "formula_len_all_1": 61, "len_all": 155336, "len_all_1": 61504, "len_abs": 1658, "len_title": 125, "len_sents": 41384, "len_sents_1": 30688, "n_sents": 267, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1665, "title": "Detecting Training Data of Large Language Models via Expectation Maximization", "abs": "The widespread deployment of large language models (LLMs) has led to impressive advancements, yet information about their training data, a critical factor in their performance, remains undisclosed. Membership inference attacks (MIAs) aim to determine whether a specific instance was part of a target model's training data. MIAs can offer insights into LLM outputs and help detect and address concerns such as data contamination and compliance with privacy and copyright standards. However, applying MIAs to LLMs presents unique challenges due to the massive scale of pre-training data and the ambiguous nature of membership. Additionally, creating appropriate benchmarks to evaluate MIA methods is not straightforward, as training and test data distributions are often unknown. In this paper, we introduce EM-MIA, a novel MIA method for LLMs that iteratively refines membership scores and prefix scores via an expectation-maximization algorithm, leveraging the duality that the estimates of these scores can be improved by each other. Membership scores and prefix scores assess how each instance is likely to be a member and discriminative as a prefix, respectively. Our method achieves state-of-the-art results on the WikiMIA dataset. To further evaluate EM-MIA, we present OLMoMIA, a benchmark built from OLMo resources, which allows us to control the difficulty of MIA tasks with varying degrees of overlap between training and test data distributions. We believe that EM-MIA serves as a robust MIA method for LLMs and that OLMoMIA provides a valuable resource for comprehensively evaluating MIA approaches, thereby driving future research in this critical area.", "keywords": [ "large language models", "membership inference attack", "data contamination", "memorization" ], "rating_list": [ 6, 1, 3, 5 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "armbJRJdrH", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Jiachen Lei", "gender": "Male", "institution": "Zhejiang University Jiaxing Research Institute", "country": "", "position": "Researcher" }, { "name": "Jiongxiao Wang", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Julius Berner", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Jun Zhu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Kui Ren", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zhongjie Ba", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zhongzhu Chen", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "PhD student" }, { "name": "Anima Anandkumar", "gender": "Female", "institution": "California Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "chaowei xiao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 33, "n_ref": 111, "n_ref_all": 139, "n_fig": 10, "n_tab": 4, "L_tab": 582, "n_element_tab": 61, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 3, "n_element_tab_1": 1, "formula_len_all": 808, "formula_len_all_1": 644, "len_all": 132800, "len_all_1": 67434, "len_abs": 1433, "len_title": 113, "len_sents": 43928, "len_sents_1": 30121, "n_sents": 329, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1442, "title": "Robust Representation Consistency Model via Contrastive Denoising", "abs": "Robustness is essential for deep neural networks, especially in security-sensitive applications. To this end, randomized smoothing provides theoretical guarantees for certifying robustness against adversarial perturbations. Recently, diffusion models have been successfully employed for randomized smoothing to purify noise-perturbed samples before making predictions with a standard classifier. While these methods excel at small perturbation radii, they struggle with larger perturbations and incur a significant computational overhead during inference compared to classical methods. To address this, we reformulate the generative modeling task along the diffusion trajectories in pixel space as a discriminative task in the latent space. Specifically, we use instance discrimination to achieve consistent representations along the trajectories by aligning temporally adjacent points. After fine-tuning based on the learned representations, our model enables implicit denoising-then-classification via a single prediction, substantially reducing inference costs. We conduct extensive experiments on various datasets and achieve state-of-the-art performance with minimal computation budget during inference. For example, our method outperforms the certified accuracy of diffusion-based methods on ImageNet across all perturbation radii by 5.3\\% on average, with up to 11.6\\% at larger radii, while reducing inference costs by 85x on average.", "keywords": [ "Certified Robustness", "Diffusion-based randomized smoothing" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "arbj7RJ5oh", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Bhishma Dedhia", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Niraj Jha", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 43, "n_ref": 75, "n_ref_all": 103, "n_fig": 20, "n_tab": 10, "L_tab": 5099, "n_element_tab": 378, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 231, "n_element_tab_1": 10, "formula_len_all": 1034, "formula_len_all_1": 731, "len_all": 168443, "len_all_1": 58618, "len_abs": 1344, "len_title": 133, "len_sents": 41672, "len_sents_1": 25849, "n_sents": 370, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1362, "title": "Neural Slot Interpreters: Grounding Object Semantics in Emergent Slot Representations", "abs": "Several accounts of human cognition posit that our intelligence is rooted in our ability to form abstract composable \nconcepts, ground them in our environment, and reason over these grounded entities. This trifecta of human \nthought has remained elusive in modern intelligent machines. In this work, we investigate whether slot representations \nextracted from visual scenes serve as appropriate compositional abstractions for grounding and reasoning. We present the \nNeural Slot Interpreter (NSI), which learns to ground object semantics in slots. At the core of NSI is an XML-like \nschema that uses simple syntax rules to organize the object semantics of a scene into object-centric schema primitives. \nThen, the NSI metric learns to ground primitives into slots through a structured objective that reasons over the intermodal \nalignment. We show that the grounded slots surpass unsupervised slots in real-world object discovery and scale with scene \ncomplexity. Experiments with a bi-modal object-property and scene retrieval task demonstrate the grounding efficacy and \ninterpretability of correspondences learned by NSI. Finally, we investigate the reasoning abilities of the grounded slots. \nVision Transformers trained on grounding-aware NSI tokenizers using as few as ten tokens outperform patch-based tokens on \nchallenging few-shot classification tasks.", "keywords": [ "Object centric learning", "grounding", "embodied reasoning", "visual tokenizers" ], "rating_list": [ 3, 8, 3, 3, 6 ], "soundness_list": [ 3, 4, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 3, 2 ], "contribution_list": [ 1, 4, 2, 2, 2 ], "confidence_list": [ 2, 5, 3, 5, 3 ] }, { "paper_id": "ar9tcnD4e9", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Michael Jiang", "gender": "Male", "institution": "Shenzhen Technology University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xinshun Liu", "gender": "Male", "institution": "datatist", "country": "CN", "position": "Researcher" }, { "name": "YiZhi Fang", "gender": "Male", "institution": "Independent", "country": "US", "position": "Intern" } ], "n_formula": 17, "n_formula_1": 12, "n_ref_uni": 25, "n_ref": 35, "n_ref_all": 68, "n_fig": 13, "n_tab": 6, "L_tab": 1096, "n_element_tab": 164, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1011, "n_element_tab_1": 151, "formula_len_all": 3259, "formula_len_all_1": 1058, "len_all": 138906, "len_all_1": 60766, "len_abs": 1809, "len_title": 135, "len_sents": 47488, "len_sents_1": 26848, "n_sents": 403, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1822, "title": "Automatic Organization of Neural Modules for Enhanced Collaboration in Neural Networks", "abs": "This work proposes a new perspective on the structure of Neural Networks (NNs). Traditional Neural Networks are typically tree-like structures for convenience, which can be predefined or learned by NAS methods. However, such a structure can not facilitate communications between nodes at the same level or signal transmissions to previous levels. These defects prevent effective collaboration, restricting the capabilities of neural networks. It is well-acknowledged that the biological neural system contains billions of neural units. Their connections are far more complicated than the current NN structure. To enhance the representational ability of neural networks, existing works try to increase the depth of the neural network and introduce more parameters. However, they all have limitations with constrained parameters. In this work, we introduce a synchronous graph-based structure to establish a novel way of organizing the neural units: the Neural Modules. This framework allows any nodes to communicate with each other and encourages neural units to work collectively, demonstrating a departure from the conventional constrained paradigm. Such a structure also provides more candidates for the NAS methods. Furthermore, we also propose an elegant regularization method to organize neural units into multiple independent, balanced neural modules systematically. This would be convenient for handling these neural modules in parallel. Compared to traditional NNs, our method unlocks the potential of NNs from tree-like structures to general graphs and makes NNs be optimized in an almost complete set. Our approach proves adaptable to diverse tasks, offering compatibility across various scenarios. Quantitative experimental results substantiate the potential of our structure, indicating the improvement of NNs.", "keywords": [ "Deep Learning; Neural Networks;" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "ar74UIeN1O", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Ji Wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jingzhi Yang", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Miao Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "WeiLong Wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xiao Zhang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xun Zhu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "ZiYu Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 18, "n_ref": 28, "n_ref_all": 52, "n_fig": 8, "n_tab": 4, "L_tab": 1187, "n_element_tab": 18, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 958, "n_element_tab_1": 15, "formula_len_all": 765, "formula_len_all_1": 959, "len_all": 110707, "len_all_1": 59241, "len_abs": 1643, "len_title": 145, "len_sents": 37290, "len_sents_1": 27988, "n_sents": 236, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1649, "title": "Medical Vision-Language Pretraining through Contrastive Learning of Positive and Negative Mention", "abs": "In recent years, contrastive learning techniques have achieved significant success and have been widely applied in both general and medical domains. In the general domain, image captions typically describe only objects present in the image. However, in the medical field, radiology reports contain both sentences confirming the presence of diseases or abnormalities (positive mentions) and sentences explicitly ruling them out (negative mentions). Current vision-language pretraining models in the medical domain often overlook this critical distinction in both model evaluation (e.g., zero-shot classification) and training processes.\nIn this paper, we suggest adding a zero-shot classification evaluation method. Unlike previous approaches that only assess the semantic similarity between medical images and positive mentions of different disease categories, this method evaluates the model\u2019s ability to distinguish between medical images and both positive and negative mentions of given disease category. Furthermore, to better capture the complex semantic relationships between medical images and the corresponding radiology reports, we introduce a visual entailment based contrastive learning method, explicitly modeling the entailment, contradiction, and neutral relationships between medical images and report sentences.\nExperimental results demonstrate that integrating this new evaluation method provides a more comprehensive evaluation of vision-language pretraining models in the medical domain. Additionally, our model achieves state-of-the-art performance across various downstream tasks, highlighting the effectiveness of our approach.", "keywords": [ "Unsupervised Learning", "Contrastive Learning", "Medical Multimodality" ], "rating_list": [ 5, 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 1, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 1, 3 ], "confidence_list": [ 5, 4, 4, 5, 4 ] }, { "paper_id": "aqvf3R48pl", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Heng Huang", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Zhenyi Wang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Postdoc" } ], "n_formula": 13, "n_formula_1": 5, "n_ref_uni": 43, "n_ref": 66, "n_ref_all": 86, "n_fig": 6, "n_tab": 7, "L_tab": 1591, "n_element_tab": 212, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 935, "n_element_tab_1": 34, "formula_len_all": 450, "formula_len_all_1": 270, "len_all": 132084, "len_all_1": 62234, "len_abs": 1303, "len_title": 105, "len_sents": 36078, "len_sents_1": 30879, "n_sents": 252, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1308, "title": "Data Efficient Continual Learning of Large Language Model", "abs": "Continual Learning (CL) in large language models (LLMs) aims to enable models to learn from evolving data distributions while preserving previously acquired knowledge. However, existing CL methods primarily rely on statistical correlations from observed data, which are particularly vulnerable under limited data settings. This reliance results in two major drawbacks: (1) increased susceptibility to forgetting previously learned knowledge when data distribution shifts occur, and (2) a tendency to depend on spurious features instead of uncovering true causal relationships in new tasks. These issues become even more pronounced, especially when training data is limited. To address these challenges, we introduce a causality-guided CL approach that reinterprets CL through the lens of causal inference. Our method aims to mitigate the dependency of model parameters on the data inputs, leading to two key advantages: (1) reduced catastrophic forgetting, and (2) decreased dependence on spurious correlations, thereby improving generalization across both old and new tasks. Extensive experiments on pre-trained LLMs, including T5-large and Llama2, demonstrate that our approach significantly outperforms state-of-the-art (SOTA) CL methods in LLMs, particularly when the amount of training data is limited.", "keywords": [ "Continual Learning; Large Language Model" ], "rating_list": [ 3, 5, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 1, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 3, 4, 3 ] }, { "paper_id": "aqok1UX7Z1", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Allan Zhou", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Sadhika Malladi", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Yiding Jiang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Zhili Feng", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zico Kolter", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 69, "n_ref": 119, "n_ref_all": 144, "n_fig": 8, "n_tab": 3, "L_tab": 1903, "n_element_tab": 294, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 562, "n_element_tab_1": 95, "formula_len_all": 1103, "formula_len_all_1": 637, "len_all": 194783, "len_all_1": 65405, "len_abs": 1661, "len_title": 118, "len_sents": 49512, "len_sents_1": 31642, "n_sents": 356, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1335, "title": "Adaptive Data Optimization: Dynamic Sample Selection with Scaling Laws", "abs": "The composition of pretraining data is a key determinant of foundation models' performance, but there is no standard guideline for allocating a limited computational budget across different data sources. Most current approaches either rely on extensive experiments with smaller models or dynamic data adjustments that also require proxy models, both of which significantly increase the workflow complexity and computational overhead. In this paper, we introduce Adaptive Data Optimization (ADO), an algorithm that optimizes data distributions in an online fashion, concurrent with model training. Unlike existing techniques, ADO does not require external knowledge, proxy models, or modifications to the model update. Instead, ADO uses per-domain scaling laws to estimate the learning potential of each domain during training and adjusts the data mixture accordingly, making it more scalable and easier to integrate. Experiments demonstrate that ADO can achieve comparable or better performance than prior methods while maintaining computational efficiency across different computation scales, offering a practical solution for dynamically adjusting data distribution without sacrificing flexibility or increasing costs. Beyond its practical benefits, ADO also provides a new perspective on data collection strategies via scaling laws.", "keywords": [ "Data selection", "Pre-training", "Curriculum learning", "Language Models", "Scaling laws" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "aqlzXgXwWa", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 13, "author_info_list": [ { "name": "Andong Wang", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Heung-Yeung Shum", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Emeritus" }, { "name": "Jingyun Xue", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "MS student" }, { "name": "Kaihao Zhang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Mengyang LIU", "gender": "Male", "institution": "Tencent", "country": "", "position": "Engineer" }, { "name": "Qi Tian", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Wei Liu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Distinguished Scientist" }, { "name": "Wenhan Luo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Wenzhe Zhao", "gender": "Male", "institution": "South China University of Technology", "country": "", "position": "MS student" }, { "name": "Zhiyuan Zhao", "gender": "Male", "institution": "Tencent ", "country": "CN", "position": "Researcher" }, { "name": "WANG HongFa", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Shaobo Min", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "ma yue", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 41, "n_ref": 66, "n_ref_all": 104, "n_fig": 16, "n_tab": 7, "L_tab": 2980, "n_element_tab": 313, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 2232, "n_element_tab_1": 253, "formula_len_all": 945, "formula_len_all_1": 459, "len_all": 150991, "len_all_1": 56477, "len_abs": 171, "len_title": 128, "len_sents": 41631, "len_sents_1": 23117, "n_sents": 344, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1687, "title": "Towards Multiple Character Image Animation Through Enhancing Implicit Decoupling", "abs": "Controllable character image animation has a wide range of applications. Although existing studies have consistently improved performance, challenges persist in the field of character image animation, particularly concerning stability in complex backgrounds and tasks involving multiple characters. To address these challenges, we propose a novel multi-condition guided framework for character image animation, employing several well-designed input modules to enhance the implicit decoupling capability of the model. First, the optical flow guider calculates the background optical flow map as guidance information, which enables the model to implicitly learn to decouple the background motion into background constants and background momentum during training, and generate a stable background by setting zero background momentum during inference. Second, the depth order guider calculates the order map of the characters, which transforms the depth information into the positional information of multiple characters. This facilitates the implicit learning of decoupling different characters, especially in accurately separating the occluded body parts of multiple characters. Third, the reference pose map is input to enhance the ability to decouple character texture and pose information in the reference image. Furthermore, to fill the gap of fair evaluation of multi-character image animation in the community, we propose a new benchmark comprising approximately 4,000 frames. Extensive qualitative and quantitative evaluations demonstrate that our method excels in generating high-quality character animations, especially in scenarios of complex backgrounds and multiple characters.", "keywords": [ "character image animation", "video generation", "diffusion model" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "aq7H2pWlEv", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Jian Shi", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Peter Wonka", "gender": "Male", "institution": "Snap Inc.", "country": "SA", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 28, "n_ref": 45, "n_ref_all": 62, "n_fig": 4, "n_tab": 6, "L_tab": 1654, "n_element_tab": 124, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3092, "n_element_tab_1": 104, "formula_len_all": 348, "formula_len_all_1": 390, "len_all": 117199, "len_all_1": 66105, "len_abs": 1300, "len_title": 136, "len_sents": 33709, "len_sents_1": 30320, "n_sents": 231, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1378, "title": "VoxelKP: A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data", "abs": "We present \\textit{VoxelKP}, a novel fully sparse network architecture tailored for human keypoint estimation in LiDAR data.\nThe key challenge is that objects are distributed sparsely in 3D space, while human keypoint detection requires detailed local information wherever humans are present.\nFirst, we introduce a dual-branch \\textit{fully sparse spatial-context block} where the spatial branch focuses on learning the local spatial correlations between keypoints within each human instance, while the context branch aims to retain the global spatial information. Second, we use a \\textit{spatially aware multi-scale BEV fusion} technique to leverage absolute 3D coordinates when projecting 3D voxels to a 2D grid encoding a bird's eye view for better preservation of the global context of each human instance.\nWe evaluate our method on the Waymo dataset and achieve an improvement of $27\\%$ on the MPJPE metric compared to the state-of-the-art, \\textit{HUM3DIL}, trained on the same data, and $12\\%$ against the state-of-the-art, \\textit{GC-KPL}, pretrained on a $25\\times$ larger dataset.\nTo the best of our knowledge, \\textit{VoxelKP} is the first single-staged, fully sparse network that is specifically designed for addressing the challenging task of 3D keypoint estimation from LiDAR data, achieving state-of-the-art performances. Our code is available at\n\\url{https://}.", "keywords": [ "Human Pose Estimation; Point Clouds; 3D Perception" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "apPItJe0wO", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Evan William Becker", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Stefano Soatto", "gender": "unknown", "institution": "Amazon Web Services", "country": "US", "position": "Vice President" } ], "n_formula": 17, "n_formula_1": 6, "n_ref_uni": 51, "n_ref": 112, "n_ref_all": 136, "n_fig": 12, "n_tab": 6, "L_tab": 1215, "n_element_tab": 192, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 555, "n_element_tab_1": 101, "formula_len_all": 818, "formula_len_all_1": 445, "len_all": 161608, "len_all_1": 65195, "len_abs": 1351, "len_title": 53, "len_sents": 51205, "len_sents_1": 32240, "n_sents": 381, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1377, "title": "Measuring LLM Confidence through Stable Explanations", "abs": "In many high-risk machine learning applications it is essential for a model to indicate when it is uncertain about a prediction. While large language models (LLMs) can reach and even surpass human-level accuracy on a variety of benchmarks, their overconfidence in incorrect responses is still a well-documented failure mode. Traditional methods for ML uncertainty quantification can be difficult to directly adapt to LLMs due to the computational cost of implementation and closed-source nature of many models. A variety of black-box methods have recently been proposed, but these often rely on heuristics such as self-verbalized confidence. We instead propose a framework for measuring an LLM's uncertainty with respect to the distribution of generated explanations for an answer. While utilizing explanations is not a new idea in and of itself, by interpreting each possible model+explanation pair as a test-time classifier we can calculate a posterior answer distribution over the most likely of these classifiers. We demonstrate how a specific instance of this framework using explanation entailment as our classifier likelihood improves confidence score metrics (in particular AURC and AUROC) over baselines across five different datasets. We believe these results indicate that our framework is both a well-principled and effective way of quantifying uncertainty in LLMs.", "keywords": [ "LLM", "Uncertainty Quantification", "Confidence Estimation", "Trustworthiness" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "apErWGzCAA", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Cong Lu", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Postdoc" }, { "name": "Jeff Clune", "gender": "unknown", "institution": "University of Wyoming", "country": "", "position": "" }, { "name": "Shengran Hu", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 107, "n_ref_all": 134, "n_fig": 4, "n_tab": 7, "L_tab": 2720, "n_element_tab": 150, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1, "formula_len_all_1": 0, "len_all": 235463, "len_all_1": 61114, "len_abs": 1927, "len_title": 124, "len_sents": 71552, "len_sents_1": 30852, "n_sents": 521, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1979, "title": "Intelligent Go-Explore: Standing on the Shoulders of Giant Foundation Models", "abs": "Go-Explore is a powerful family of algorithms designed to solve hard-exploration problems built on the principle of archiving discovered states, and iteratively returning to and exploring from the most promising states. This approach has led to superhuman performance across a wide variety of challenging problems including Atari games and robotic control, but requires manually designing heuristics to guide exploration (i.e., determine which states to save and explore from, and what actions to consider next), which is time-consuming and infeasible in general. To resolve this, we propose Intelligent Go-Explore (IGE) which greatly extends the scope of the original Go-Explore by replacing these handcrafted heuristics with the intelligence and internalized human notions of interestingness captured by giant pretrained foundation models (FMs). This provides IGE with a human-like ability to instinctively identify how interesting or promising any new state is (e.g., discovering new objects, locations, or behaviors), even in complex environments where heuristics are hard to define. Moreover, IGE offers the exciting opportunity to recognize and capitalize on serendipitous discoveries---states encountered during exploration that are valuable in terms of exploration, yet where what makes them interesting was not anticipated by the human user. We evaluate our algorithm on a diverse range of language and vision-based tasks that require search and exploration. Across these tasks, IGE strongly exceeds classic reinforcement learning and graph search baselines, and also succeeds where prior state-of-the-art FM agents like Reflexion completely fail. Overall, Intelligent Go-Explore combines the tremendous strengths of FMs and the powerful Go-Explore algorithm, opening up a new frontier of research into creating more generally capable agents with impressive exploration capabilities. All our code is open-sourced at: https://github.com/conglu1997/intelligent-go-explore.", "keywords": [ "Exploration", "Large Language Models", "LLM agents", "Open-endedness" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "aoW5Sm8Op8", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Isabel Chien", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Javier Gonzalez", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Principal Researcher" }, { "name": "Richard E. Turner", "gender": "Male", "institution": "Alan Turing Institute", "country": "GB", "position": "Researcher" }, { "name": "Zachary Izzo", "gender": "Male", "institution": "NEC Labs America", "country": "", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 1, "n_ref_uni": 62, "n_ref": 121, "n_ref_all": 159, "n_fig": 18, "n_tab": 5, "L_tab": 649, "n_element_tab": 38, "n_fig_1": 17, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 865, "formula_len_all_1": 38, "len_all": 205339, "len_all_1": 75107, "len_abs": 1827, "len_title": 114, "len_sents": 63468, "len_sents_1": 37925, "n_sents": 492, "n_sents_1": 288, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1833, "title": "Benchmarking Survival Models: Treatment Effects, Bias, and Equity", "abs": "Survival models are widely used to model time-to-event or survival data, which represents the duration until an event of interest occurs. In clinical research, survival analysis is used for estimating the effects of treatments on patient health outcomes. Recent advancements in machine learning (ML) have aimed to improve survival analysis methods, but current evaluation practices largely focus on predictive performance, often neglecting critical factors such as the ability to accurately estimate treatment effects and possible consequences on health equity. Estimating treatment effects from time-to-event data presents unique challenges due to the complex problem setting, the extensive assumptions required for causal inference, biased observational data, and the ethical consequences of using model outcomes in real-world health decisions. In this work, we introduce a comprehensive benchmarking framework designed to evaluate survival models on their ability to estimate treatment effects under realistic conditions and in the presence of potential inequalities. We formalize the discussion of bias in survival modelling, identifying key sources of inequity, and outline practical desiderata for methods that model time-to-event treatment effects. We clarify common assumptions in survival analysis, discuss critical shortcomings in current evaluation practices, and propose a new benchmarking metric that can be used to better evaluate model calibration. Using this framework, we systematically compare traditional and modern survival models across multiple synthetic and real world datasets, investigating, among other challenges, model performance under mis-specification and observational biases. Through this benchmark, we provide actionable insights for researchers to develop more robust and equitable survival models.", "keywords": [ "benchmarks", "survival analysis", "time to event", "fairness", "health equity", "responsible ai", "heterogeneous treatment effects" ], "rating_list": [ 1, 3, 3 ], "soundness_list": [ 3, 1, 3 ], "presentation_list": [ 1, 1, 2 ], "contribution_list": [ 1, 1, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "anQDiQZhDP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 13, "author_info_list": [ { "name": "Dangna Li", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Jeff Hwang", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "engineer" }, { "name": "Julian Chan", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Kainan Peng", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Mingbo Ma", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Research Scientist" }, { "name": "Vimal Manohar", "gender": "Male", "institution": "Meta Platforms Inc.", "country": "", "position": "Researcher" }, { "name": "Xiaohui Zhang", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Xueyao Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Yingru Liu", "gender": "Male", "institution": "Meta GenAI", "country": "US", "position": "Researcher" }, { "name": "Yuan Huang", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Yuhao Wang", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Zhenyu Tang", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Zhizheng Wu", "gender": "Male", "institution": "Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Associate Professor" } ], "n_formula": 15, "n_formula_1": 8, "n_ref_uni": 89, "n_ref": 262, "n_ref_all": 302, "n_fig": 7, "n_tab": 8, "L_tab": 3174, "n_element_tab": 421, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2878, "n_element_tab_1": 350, "formula_len_all": 1089, "formula_len_all_1": 728, "len_all": 243371, "len_all_1": 79761, "len_abs": 1741, "len_title": 130, "len_sents": 56350, "len_sents_1": 33357, "n_sents": 452, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1741, "title": "Vevo: Controllable Zero-Shot Voice Imitation with Self-Supervised Disentanglement", "abs": "The imitation of voice, targeted on specific speech attributes such as timbre and speaking style, is crucial in speech generation. However, existing methods rely heavily on annotated data, and struggle with effectively disentangling timbre and style, leading to challenges in achieving controllable generation, especially in zero-shot scenarios. To address these issues, we propose Vevo, a versatile zero-shot voice imitation framework with controllable timbre and style. Vevo operates in two core stages: (1) Content-Style Modeling: Given either text or speech's content tokens as input, we utilize an autoregressive transformer to generate the content-style tokens, which is prompted by a style reference; (2) Acoustic Modeling: Given the content-style tokens as input, we employ a flow-matching transformer to produce acoustic representations, which is prompted by a timbre reference. To obtain the content and content-style tokens of speech, we design a fully self-supervised approach that progressively decouples the timbre, style, and linguistic content of speech. Specifically, we adopt VQ-VAE as the tokenizer for the continuous hidden features of HuBERT. We treat the vocabulary size of the VQ-VAE codebook as the information bottleneck, and adjust it carefully to obtain the disentangled speech representations. Solely self-supervised trained on 60K hours of audiobook speech data, without any fine-tuning on style-specific corpora, Vevo matches or surpasses existing methods in accent and emotion conversion tasks. Additionally, Vevo\u2019s effectiveness in zero-shot voice conversion and text-to-speech tasks further demonstrates its strong generalization and versatility. Audio samples are available at https://versavoice.github.io/.", "keywords": [ "controllable speech generation", "speech disentanglement", "voice conversion", "accent conversion", "text to speech" ], "rating_list": [ 6, 8, 3, 5 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "anN4a8h4od", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Milind Shyani", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Yonatan Naamad", "gender": "Not Specified", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 27, "n_ref": 41, "n_ref_all": 85, "n_fig": 9, "n_tab": 6, "L_tab": 1781, "n_element_tab": 150, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 110, "n_element_tab_1": 15, "formula_len_all": 1826, "formula_len_all_1": 677, "len_all": 151771, "len_all_1": 68228, "len_abs": 1437, "len_title": 94, "len_sents": 54932, "len_sents_1": 35241, "n_sents": 380, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1445, "title": "Filtered Semantic Search via Vector Arithmetic", "abs": "How can we retrieve search results that are both semantically relevant and satisfy certain filter criteria? Modern day semantic search engines are increasingly reliant on vector-based search, yet the ability to restrict vector search to a fixed set of filter criteria remains an interesting problem with no known satisfactory solution. In this note, we leverage the rich emergent structure of vector embeddings of pre-trained search transformers to offer a simple solution. Our method involves learning, for each filter, a vector direction in the space of vector embeddings, and adding it to the query vector at run-time to perform a search constrained by that filter criteria. Our technique is broadly applicable to any finite set of semantically meaningful filters, compute-efficient in that it does not require modifying or rebuilding an existing $k$-NN index over document vector embeddings, lightweight in that it adds negligible latency, and widely compatible in that it can be utilized with any transformer model and $k$-NN algorithm. We also establish, subject to mild assumptions, an upper bound on the probability that our method errantly retrieves irrelevant results, and reveal new empirical insights about the geometry of transformer embeddings. In experiments, we find that our method, on average, yields more than a 21% boost over the baseline (measured in terms of nDCG@10) across three different transformer models and datasets.", "keywords": [ "passage retrieval", "dense retrieval", "feature representation" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "anAZ42rYFK", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Hengshuang Zhao", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Xiaogang Xu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Researcher" }, { "name": "Zhenhua Xu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Zhuoling Li", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Ser-Nam Lim", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 27, "n_ref": 46, "n_ref_all": 67, "n_fig": 6, "n_tab": 3, "L_tab": 695, "n_element_tab": 107, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 695, "n_element_tab_1": 107, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 105440, "len_all_1": 59700, "len_abs": 1442, "len_title": 72, "len_sents": 30885, "len_sents_1": 28336, "n_sents": 244, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1449, "title": "LARM: Large Auto-Regressive Model for Long-Horizon Embodied Intelligence", "abs": "Due to the need of interacting with the world, embodied agents are required to possess comprehensive task-relevant knowledge, long-horizon planning capability, and a swift response speed. Large language models (LLMs), owing to their rich general knowledge, recently achieve promising results in open-world embodied tasks, like the world exploration in Minecraft. However, the outputs of LLMs are descriptive sentences or code, which are slow to generate and not end-to-end, as a translator is required to translate the LLM outputs into actions to perform. To address these limitations, we introduce the large auto-regressive model (LARM). LARM leverages environment observations as input and predicts subsequent actions in an auto-regressive manner. Compared with LLM based methods, LARM directly predicts the next skill for execution according to the current observation. In addition, considering that the commonly adopted training paradigms do not reflect the mutual influence and dependency between actions and observations, we develop a novel data format named auto-regressive node transmission structure and assemble a corresponding dataset to train LARM. Combining these techniques, LARM successfully harvests enchanted equipment in Minecraft, which demands significantly more complex decision-making chains than the highest achievements of prior best methods. Besides, the speed of LARM is 6.8x faster than LLMs with similar parameter volume.", "keywords": [ "Embodied Intelligence", "Minecraft", "Large Language Model", "Large Auto-regressive Model" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 5, 3, 3 ] }, { "paper_id": "an3kPpce6b", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Haokai Hong", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "KC Tan", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Ming Yang", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "Wanyu Lin", "gender": "Female", "institution": "The Hong Kong Polytechnic University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 55, "n_formula_1": 19, "n_ref_uni": 46, "n_ref": 124, "n_ref_all": 147, "n_fig": 5, "n_tab": 9, "L_tab": 3096, "n_element_tab": 500, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1174, "n_element_tab_1": 286, "formula_len_all": 7522, "formula_len_all_1": 1301, "len_all": 189673, "len_all_1": 76281, "len_abs": 1657, "len_title": 137, "len_sents": 54816, "len_sents_1": 33168, "n_sents": 432, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1639, "title": "Steering 3D Molecule Generation in Data-Sparse Regions via Distributional Physical Priors", "abs": "Can we train a 3D molecule generator using data from dense regions to generate samples in sparse regions? This challenge can be framed as an out-of-distribution (OOD) generation problem. Existing works on OOD generation primarily focus on property shifts. However, the distribution shifts may come from structural variations in molecules, such as certain types of scaffolds, dubbed as physical priors. This work introduces a novel and principled diffusion-based generative framework, termed _GODD_, which enables training a generator on data-abundant distributions to generalize to data-scarce distributions under structure shifts. Specifically, we propose utilizing a designated equivariant asymmetric autoencoder to capture distributional physical priors. The asymmetric module allows generalization to unseen, out-of-distribution structural variations. As these captured physical priors represent distinct distributions, they can steer the generation of samples that are not in dense regions. We demonstrate that with these encoded structural-grained distributional physical priors, _GODD_ does not need to train with any molecules from the sparse regions. We conduct extensive experiments across various out-of-distribution molecule generation tasks using benchmark datasets. Compared to alternative baselines, our approach shows a significant improvement of up to 65.6\\% in success rate, defined based on molecular validity, uniqueness, and novelty. Additionally, we show that our generative framework, steered by physical priors, can be readily adapted to canonical fragment-based drug design tasks, exhibiting promising performance.", "keywords": [ "Molecule Generation", "Diffusion Model", "Out of distribution" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 2, 1, 3, 4 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "an3jH2qD2r", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Alberto Cazzaniga", "gender": "Male", "institution": "AREA Science Park", "country": "IT", "position": "Researcher" }, { "name": "Giada Panerai", "gender": "Female", "institution": "University of Trieste", "country": "IT", "position": "MS student" }, { "name": "Karthik Viswanathan", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Matteo Biagetti", "gender": "Male", "institution": "Area Science Park", "country": "IT", "position": "Researcher" }, { "name": "Yuri Gardinazzi", "gender": "Male", "institution": "University of Trieste", "country": "IT", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 40, "n_ref": 82, "n_ref_all": 117, "n_fig": 19, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 316, "formula_len_all_1": 316, "len_all": 144675, "len_all_1": 63438, "len_abs": 1027, "len_title": 124, "len_sents": 43264, "len_sents_1": 30678, "n_sents": 319, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1031, "title": "The Geometry of Tokens in Internal Representations of Large Language Models", "abs": "We investigate the relationship between the geometry of token embeddings and their role in next token prediction within transformer models. Toward this goal, previous studies have utilized metrics such as intrinsic dimension and neighborhood overlap to probe the geometry of internal representations, where prompts are summarized as a single point in representation space. We expand single points to point clouds by investigating how models geometrically distribute tokens in their internal representations. We measure the intrinsic dimension, neighborhood overlap, and cosine similarity on these point clouds for a large number of prompts. \nTo validate our approach, we compare these metrics to a dataset where the tokens are shuffled, which disrupts the syntactic and semantic structure. Our analysis reveals a correlation between the geometric properties of token embeddings and the cross-entropy loss of next token predictions, implying that prompts with higher loss values have tokens represented in higher-dimensional spaces.", "keywords": [ "Intrinsic Dimension", "Neighborhood Overlap", "Internal Representations", "Large Language Models" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 2, 3, 4, 4 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "an3fugFA23", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Guangsheng Bao", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Juncai He", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Researcher" }, { "name": "Yanbin Zhao", "gender": "Female", "institution": "Shanghai Polytechnic University", "country": "", "position": "Associate Professor" }, { "name": "Yue Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" } ], "n_formula": 21, "n_formula_1": 10, "n_ref_uni": 48, "n_ref": 99, "n_ref_all": 139, "n_fig": 10, "n_tab": 2, "L_tab": 2375, "n_element_tab": 332, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 594, "n_element_tab_1": 7, "formula_len_all": 1448, "formula_len_all_1": 894, "len_all": 178204, "len_all_1": 60953, "len_abs": 2168, "len_title": 148, "len_sents": 47051, "len_sents_1": 27341, "n_sents": 429, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1318, "title": "Glimpse: Enabling White-Box Methods to Use Proprietary Models for Zero-Shot LLM-Generated Text Detection", "abs": "Advanced large language models (LLMs) can generate text almost indistinguishable from human-written text, highlighting the importance of LLM-generated text detection. However, current zero-shot techniques face challenges as white-box methods are restricted to use weaker open-source LLMs, and black-box methods are limited by partial observation from stronger proprietary LLMs. It seems impossible to enable white-box methods to use proprietary models because API-level access to the models neither provides full predictive distributions nor inner embeddings. To traverse the divide, we propose **Glimpse**, a probability distribution estimation approach, predicting the full distributions from partial observations. Despite the simplicity of Glimpse, we successfully extend white-box methods like Entropy, Rank, Log-Rank, and Fast-DetectGPT to latest proprietary models. Experiments show that Glimpse with Fast-DetectGPT and GPT-3.5 achieves an average AUROC of about 0.95 in five latest source models, improving the score by 51\\% relative to the remaining space of the open source baseline. It demonstrates that the latest LLMs can effectively detect their own outputs, suggesting that advanced LLMs may be the best shield against themselves. We release our code and data at https://github.com/baoguangsheng/glimpse.", "keywords": [ "Machine-Generated Text Detection" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "amOpepqmSl", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Armand Foucault", "gender": "Male", "institution": "Universit\u00e9 de Toulouse", "country": "", "position": "PhD student" }, { "name": "Franck Mamalet", "gender": "Male", "institution": "IRT Saint Exupery", "country": "FR", "position": "Researcher" }, { "name": "Francois Malgouyres", "gender": "Male", "institution": "Universit\u00e9 Paul Sabatier", "country": "", "position": "Full Professor" } ], "n_formula": 50, "n_formula_1": 11, "n_ref_uni": 71, "n_ref": 144, "n_ref_all": 187, "n_fig": 3, "n_tab": 14, "L_tab": 5553, "n_element_tab": 812, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1416, "n_element_tab_1": 260, "formula_len_all": 3392, "formula_len_all_1": 519, "len_all": 248065, "len_all_1": 72780, "len_abs": 5599, "len_title": 52, "len_sents": 60577, "len_sents_1": 29960, "n_sents": 534, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1265, "title": "HADAMRNN: BINARY AND SPARSE TERNARY ORTHOGONAL RNNS", "abs": "Binary and sparse ternary weights in neural networks enable faster computations and lighter representations, facilitating their use on edge devices with limited computational power. Meanwhile, vanilla RNNs are highly sensitive to changes in their recurrent weights, making the binarization and ternarization of these weights inherently challenging. To date, no method has successfully achieved binarization\nor ternarization of vanilla RNN weights. We present a new approach leveraging the properties of Hadamard matrices to parameterize a subset of binary and sparse ternary orthogonal matrices. This method enables the training of orthogonal RNNs (ORNNs) with binary and sparse ternary recurrent weights, effectively creating a specific class of binary and sparse ternary vanilla RNNs. The resulting ORNNs, called HadamRNN and Block-HadamRNN, are evaluated on benchmarks such as the copy task, permuted and sequential MNIST tasks, and IMDB dataset. Despite binarization or sparse ternarization, these RNNs maintain performance levels comparable to state-of-the-art full-precision models, highlighting the effectiveness of our approach. Notably, our approach is the first solution with binary recurrent weights capable of tackling the copy task over 1000 timesteps.", "keywords": [ "Recurrent neural networks", "quantization", "orthogonal matrices", "Hadamard matrices" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 2 ] }, { "paper_id": "amDkNPVWcn", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Dinghuai Zhang", "gender": "Unspecified", "institution": "Mila, University of Montreal", "country": "CA", "position": "PhD student" }, { "name": "Jiatao Gu", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Joshua M. Susskind", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Navdeep Jaitly", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Qihang Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postgraduate student" }, { "name": "Shuangfei Zhai", "gender": "Male", "institution": "Apple", "country": "US", "position": "Research Scientist" }, { "name": "Yizhe Zhang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Yuyang Wang", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Research Scientist" } ], "n_formula": 24, "n_formula_1": 11, "n_ref_uni": 48, "n_ref": 88, "n_ref_all": 109, "n_fig": 12, "n_tab": 5, "L_tab": 1266, "n_element_tab": 187, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1784, "formula_len_all_1": 854, "len_all": 142318, "len_all_1": 55516, "len_abs": 4650, "len_title": 112, "len_sents": 36379, "len_sents_1": 26115, "n_sents": 285, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1100, "title": "Denoising Autoregressive Transformers for Scalable Text-to-Image Generation", "abs": "Diffusion models have become the dominant approach for visual generation. They are trained by denoising a Markovian process which gradually adds noise to the input. We argue that the Markovian property limits the model\u2019s ability to fully utilize the generation trajectory, leading to inefficiencies during training and inference. In this paper, we propose DART, a transformer-based model that unifies autoregressive (AR) and diffusion within a non-Markovian framework. DART iteratively denoises image patches spatially and spectrally using an AR model that has the same architecture as standard language models. DART does not rely on image quantization, which enables more effective image modeling while maintaining flexibility. Furthermore, DART seamlessly trains with both text and image data in a unified model. Our approach demonstrates competitive performance on class-conditioned and text-to-image generation tasks, offering a scalable, efficient alternative to traditional diffusion models. Through this unified framework, DART sets a new benchmark for scalable, high-quality image synthesis.", "keywords": [ "diffusion models", "autoregressive models", "Transformer" ], "rating_list": [ 5, 6, 6, 8, 6 ], "soundness_list": [ 3, 4, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 4, 3 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "am5Z8dXoaV", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Mahyar Najibi", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Minsik Cho", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "PhD student" }, { "name": "Mohammad Rastegari", "gender": "Male", "institution": "Apple", "country": "", "position": "ML manager" }, { "name": "Qichen Fu", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Sachin Mehta", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Thomas Merth", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 32, "n_ref": 56, "n_ref_all": 76, "n_fig": 8, "n_tab": 8, "L_tab": 3617, "n_element_tab": 606, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 3204, "n_element_tab_1": 596, "formula_len_all": 155, "formula_len_all_1": 30, "len_all": 118144, "len_all_1": 69732, "len_abs": 1413, "len_title": 71, "len_sents": 33847, "len_sents_1": 28427, "n_sents": 244, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1418, "title": "LazyLLM: DYNAMIC TOKEN PRUNING FOR EFFICIENT LONG CONTEXT LLM INFERENCE", "abs": "The inference of transformer-based large language models consists of two sequential stages: 1) a prefilling stage to compute the KV cache of prompts and generate the first token, and 2) a decoding stage to generate subsequent tokens. For long prompts, the KV cache must be computed for all tokens during the prefilling stage, which can significantly increase the time needed to generate the first token. Consequently, the prefilling stage may become a bottleneck in the generation process. An open question remains whether all prompt tokens are essential for generating the first token. To answer this, we introduce a novel method, LazyLLM, that selectively computes the KV for tokens important for the next token prediction in both the prefilling and decoding stages. Contrary to static pruning approaches that prune the prompt at once, LazyLLM allows language models to dynamically select different subsets of tokens from the context in different generation steps, even though they might be pruned in previous steps. Extensive experiments on standard datasets across various tasks demonstrate that LazyLLM is a generic method that can be seamlessly integrated with existing language models to significantly accelerate the generation without fine-tuning. For instance, in the multi-document question-answering task, LazyLLM accelerates the prefilling stage of the LLama 2 7B model by 2.34\u00d7 while maintaining accuracy.", "keywords": [ "Efficient LLM Inference", "Optimization" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "alaQod29Cb", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Ayomide Odumakinde", "gender": "Male", "institution": "Cohere", "country": "CA", "position": "Intern" }, { "name": "Beyza Ermis", "gender": "Female", "institution": "Cohere AI", "country": "", "position": "Researcher" }, { "name": "Pat Verga", "gender": "unknown", "institution": "Cohere", "country": "", "position": "Researcher" }, { "name": "Daniel Dsouza", "gender": "Male", "institution": "Cohere", "country": "US", "position": "Researcher" }, { "name": "Sara Hooker", "gender": "unknown", "institution": "Cohere For AI", "country": "US", "position": "Principal Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 70, "n_ref": 112, "n_ref_all": 152, "n_fig": 11, "n_tab": 10, "L_tab": 2365, "n_element_tab": 213, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 558, "n_element_tab_1": 65, "formula_len_all": 137, "formula_len_all_1": 134, "len_all": 220122, "len_all_1": 61287, "len_abs": 1199, "len_title": 129, "len_sents": 58641, "len_sents_1": 30152, "n_sents": 429, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1205, "title": "Multilingual Arbitrage: Optimizing Data Pools to Accelerate Multilingual Progress", "abs": "The use of synthetic data has been crucial in achieving recent state-of-the-art breakthroughs. However, relying solely on a single oracle teacher model for data generation can lead to issues such as model collapse and bias propagation. These problems are particularly pronounced in multilingual contexts, where no single teacher model performs optimally across all languages. In this study, we propose a solution through multilingual arbitrage, which exploits performance variations among multiple models for each language. By strategically routing samples through a diverse set of models, each possessing unique strengths in different languages, we address these challenges. Our extensive experiments with state-of-the-art models demonstrate that our arbitrage techniques significantly enhance performance compared to relying on a single teacher model. Our multilingual arbitrage techniques result in large gains of up to 80% win-rates over state-of-art proprietary and widely adopted open weight models such as Gemma 2, Llama 3.1, Mistral v0.3. These gains, achieved through multilingual arbitrage and averaged across all languages, were most substantial in the less-resourced languages within our pool.", "keywords": [ "Synthetic data", "Model Distillation", "Multilingual language models", "Routing", "Instruction Fine-Tuning" ], "rating_list": [ 5, 3, 8, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "alBn1uNTLi", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Gonzalo Mateos", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Associate Professor" }, { "name": "Max Wasserman", "gender": "unknown", "institution": "University of Rochester", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 30, "n_ref": 50, "n_ref_all": 73, "n_fig": 6, "n_tab": 3, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1496, "formula_len_all_1": 1074, "len_all": 117867, "len_all_1": 72810, "len_abs": 932, "len_title": 88, "len_sents": 45315, "len_sents_1": 35246, "n_sents": 348, "n_sents_1": 259, "venue": "ICLR2025", "decision": "withdraw", "L_title": 40, "L_abs": 936, "title": "Stabilizing the Kumaraswamy Distribution", "abs": "Large-scale latent variable models require expressive continuous distributions that support efficient sampling and low-variance differentiation, achievable through the reparameterization trick. The Kumaraswamy (KS) distribution is both expressive and supports the reparameterization trick with a simple closed-form inverse CDF. Yet, its adoption remains limited. We identify and resolve numerical instabilities in the inverse CDF and log-pdf, exposing issues in libraries like PyTorch and TensorFlow. We then introduce simple and scalable latent variable models to improve exploration-exploitation trade-offs in contextual multi-armed bandits and enhance uncertainty quantification for link prediction with graph neural networks. We find these models to be most performant when paired with the stable KS. Our results support the stabilized KS distribution as a core component in scalable variational models for bounded latent variables.", "keywords": [ "Latent variable models", "Stochastic variational inference", "Kumaraswamy distribution", "Bounded interval distributions", "Reparameterization trick", "Contextual multi-armed bandits", "Thompson Sampling" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "akPwQb4fHU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Yizhou Wang", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Intern" }, { "name": "Yue Bai", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "PhD student" }, { "name": "Yun Fu", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Full Professor" }, { "name": "Qihua DONG", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 9, "n_ref": 20, "n_ref_all": 42, "n_fig": 3, "n_tab": 7, "L_tab": 3434, "n_element_tab": 441, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3827, "n_element_tab_1": 516, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 95821, "len_all_1": 54632, "len_abs": 956, "len_title": 130, "len_sents": 21053, "len_sents_1": 19802, "n_sents": 165, "n_sents_1": 157, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 924, "title": "Seeing is Knowing: Advancing Semantic Understanding with MLLMs in Grounding Tasks", "abs": "Large vision models (VLMs) achieve success in most daily scenarios but face challenges in special grounding tasks. This limitation is primarily due to insufficient semantic understanding for both tasks and images in current vision models. In contrast, large multimodal language models (M-LLMs) excel in semantic comprehension and instruction-following but underperform in detailed recognition. To harness the strengths of both, we propose to utilize M-LLMs to assist VLMs in handling difficult segmentation tasks.\nThe key to our approach involves\n(1)leveraging M-LLMs for semantic expertise\nand (2)formatting instruction-based guidance.\nOur proposed framework is generalizable, performing well across various tasks. Experimental results show a significant performance improvement (10\\%+) in challenging tasks like camouflage object detection, anomaly detection and medical image segmentation compared to zero-shot baselines.", "keywords": [ "Zero-shot segmentation; Multimodal LLM; MLLM Grounding" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 1, 1, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "ak7r4He1qH", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Carl William Harris", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Eduardo Pontes Reis", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Researcher" }, { "name": "Jeffrey K Jopling", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" }, { "name": "Ji Woong Kim", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "Research Assistant" }, { "name": "Michael Moor", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Assistant Professor" }, { "name": "Rojin Ziaei", "gender": "Female", "institution": "University of Maryland, College Park", "country": "", "position": "Undergrad student" }, { "name": "Samuel Schmidgall", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 44, "n_ref": 82, "n_ref_all": 104, "n_fig": 7, "n_tab": 6, "L_tab": 5472, "n_element_tab": 134, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 254318, "len_all_1": 59937, "len_abs": 3945, "len_title": 140, "len_sents": 80465, "len_sents_1": 31575, "n_sents": 590, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1485, "title": "AgentClinic: a multimodal agent benchmark to evaluate AI in simulated clinical environments", "abs": "Evaluating large language models~(LLM) in clinical scenarios is crucial to assessing their potential clinical utility. Existing benchmarks rely heavily on static question-answering, which does not accurately depict the complex, sequential nature of clinical decision-making. Here, we introduce AgentClinic, a multimodal agent benchmark for evaluating LLMs in simulated clinical environments that include patient interactions, multimodal data collection under incomplete information, and the usage of various tools, resulting in an in-depth evaluation across nine medical specialties and seven languages.\nWe find that solving MedQA problems in the sequential decision-making format of AgentClinic is considerably more challenging, resulting in diagnostic accuracies that can drop to below a tenth of the original accuracy. Overall, we observe that agents sourced from Claude-3.5 outperform other LLM backbones in most settings. Nevertheless, we see stark differences in the LLMs\u2019 ability to make use of tools, such as experiential learning, adaptive retrieval, and reflection cycles. Strikingly, Llama-3 shows up to 92\\% relative improvements with the notebook tool that allows for writing and editing notes that persist across cases. To further scrutinize our clinical simulations, we leverage real-world electronic health records, perform a clinical reader study, perturb agents with biases, and explore novel patient-centric metrics that this interactive environment firstly enables.", "keywords": [ "Language Agents", "Medical Benchmark", "Multimodal Benchmark", "Multimodal Language Models" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "ajxAJ8GUX4", "primary_area": "applications to robotics, autonomy, planning", "n_author": 3, "author_info_list": [ { "name": "Rachid Alami", "gender": "Male", "institution": "LAAS / CNRS", "country": "", "position": "Principal Researcher" }, { "name": "Smail Ait Bouhsain", "gender": "Male", "institution": "LAAS / CNRS", "country": "FR", "position": "PhD student" }, { "name": "Thierry Simeon", "gender": "Male", "institution": "LAAS / CNRS", "country": "FR", "position": "Principal Researcher" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 39, "n_ref": 74, "n_ref_all": 116, "n_fig": 16, "n_tab": 5, "L_tab": 1385, "n_element_tab": 132, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1163, "n_element_tab_1": 114, "formula_len_all": 628, "formula_len_all_1": 542, "len_all": 161593, "len_all_1": 69839, "len_abs": 1189, "len_title": 120, "len_sents": 56741, "len_sents_1": 32715, "n_sents": 411, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1192, "title": "Learning Geometric Reasoning Networks For Robot Task And Motion Planning", "abs": "Task and Motion Planning (TAMP) is a computationally challenging robotics problem due to the tight coupling of discrete symbolic planning and continuous geometric planning of robot motions. In particular, planning manipulation tasks in complex 3D environments leads to a large number of costly geometric planner queries to verify the feasibility of considered actions and plan their motions. To address this issue, we propose Geometric Reasoning Networks (GRN), a graph neural network (GNN)-based model for action and grasp feasibility prediction, designed to significantly reduce the dependency on the geometric planner. Moreover, we introduce two key interpretability mechanisms: inverse kinematics (IK) feasibility prediction and grasp obstruction (GO) estimation. These modules not only improve feasibility predictions accuracy, but also explain why certain actions or grasps are infeasible, thus allowing a more efficient search for a feasible solution. Through extensive experimental results, we show that our model outperforms state-of-the-art methods, while maintaining generalizability to more complex environments, diverse object shapes, multi-robot settings, and real-world robots.", "keywords": [ "Graph Neural Networks", "Supervised Learning for Robotics", "Task and Motion Planning", "Robot Manipulation Planning" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "ajSmXqgS24", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Jianibieke Adalibieke", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Li Yi", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Qianwei Han", "gender": "Male", "institution": "Shanghai Qi Zhi Institute", "country": "CN", "position": "Researcher" }, { "name": "Xueyi Liu", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yuzhe Qin", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 96, "n_ref_all": 138, "n_fig": 22, "n_tab": 14, "L_tab": 1785, "n_element_tab": 207, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 688, "n_element_tab_1": 75, "formula_len_all": 1247, "formula_len_all_1": 228, "len_all": 209825, "len_all_1": 74637, "len_abs": 1736, "len_title": 0, "len_sents": 71452, "len_sents_1": 37307, "n_sents": 595, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1786, "title": "DexTrack: Towards Generalizable Neural Tracking Control for Dexterous Manipulation from Human References", "abs": "We address the challenge of developing a generalizable neural tracking controller for dexterous manipulation from human references. This controller aims to manage a dexterous robot hand to manipulate diverse objects for various purposes defined by kinematic human-object interactions. Developing such a controller is complicated by the intricate contact dynamics of dexterous manipulation and the need for adaptivity, generalizability, and robustness. Current reinforcement learning and trajectory optimization methods often fall short due to their dependence on task-specific rewards or precise system models. We introduce an approach that curates large-scale successful robot tracking demonstrations, comprising pairs of human references and robot actions, to train a neural controller. Utilizing a data flywheel, we iteratively enhance the controller's performance, as well as the number and quality of successful tracking demonstrations. We exploit available tracking demonstrations and carefully integrate reinforcement learning and imitation learning to boost the controller's performance in dynamic environments. At the same time, to obtain high-quality tracking demonstrations, we individually optimize per-trajectory tracking by leveraging the learned tracking controller in a homotopy optimization method. The homotopy optimization, mimicking chain-of-thought, aids in solving challenging trajectory tracking problems to increase demonstration diversity. We showcase our success by training a generalizable neural controller and evaluating it in both simulation and real world. Our method achieves over a 10% improvement in success rates compared to leading baselines. The project website with animated results is available at [DexTrack](https://meowuu7.github.io/DexTrack/).", "keywords": [ "Dexterous Manipulation", "Neural Tracking Control", "Homotopy Optimization" ], "rating_list": [ 8, 6, 5, 8 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 2, 4, 1, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "ajORwcxeM7", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Haotian Wang", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Haoxuan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Jun Wang", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Professor" }, { "name": "Mengyue Yang", "gender": "Female", "institution": "University of Bristol", "country": "GB", "position": "Assistant Professor" }, { "name": "Pu Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xinqin Feng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xu Chen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 27, "n_formula_1": 18, "n_ref_uni": 31, "n_ref": 46, "n_ref_all": 65, "n_fig": 7, "n_tab": 5, "L_tab": 755, "n_element_tab": 48, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2650, "n_element_tab_1": 183, "formula_len_all": 2148, "formula_len_all_1": 1563, "len_all": 133454, "len_all_1": 69098, "len_abs": 1243, "len_title": 123, "len_sents": 46620, "len_sents_1": 29283, "n_sents": 375, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1248, "title": "Active Causal Learning for Conditional Average Treatment Effect Estimation", "abs": "Estimating conditional average treatment effects (CATE) from observational data is an important problem and is of high practical relevance for many domains. Despite the great efforts of recent studies to accurately estimate CATE, most methods require complete observation of\nall covariates of an individual. However, in real-world scenarios, the acquisition of covariate information is usually done in a active manner, which motivates us to develop methods to minimize the total measurement cost by actively selecting the most appropriate covariates to measure while guaranteeing the CATE estimation accuracy. To this end, in this paper, we first extend the existing methods for estimating CATE to allow accurate estimation in the presence of unmeasured covariates. Next, we theoretically show the advantage of dynamically adjusting the sampling strategy based on an evolving understanding of the information measured in the covariates. Then, we formulate the dynamic sampling strategy learning as a partially observed Markov decision process (POMDP) and further develop a policy gradient method to solve the optimal dynamic policy. Extensive experiments conducted on three real-world datasets demonstrate the effectiveness of our proposed methods.", "keywords": [ "conditional average treatment effect", "dynamic sampling", "partially observed Markov decision process" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "aisfb733DW", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Chao Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Fangyikang Wang", "gender": "Male", "institution": "Wechat, Tencent Inc", "country": "CN", "position": "Intern" }, { "name": "Hanbin Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Hubery Yin", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Hui Qian", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Huminhao Zhu", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Intern" }, { "name": "Shaobin Zhuang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Chne Li", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Expert researcher" } ], "n_formula": 71, "n_formula_1": 25, "n_ref_uni": 42, "n_ref": 88, "n_ref_all": 100, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1876, "n_element_tab_1": 90, "formula_len_all": 9456, "formula_len_all_1": 1690, "len_all": 161351, "len_all_1": 54126, "len_abs": 1629, "len_title": 117, "len_sents": 45450, "len_sents_1": 22182, "n_sents": 379, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1640, "title": "Gradient-Free Analytical Fisher Information of Diffused Distributions", "abs": "Diffusion models (DMs) have demonstrated powerful distributional modeling capabilities by matching the first-order score of diffused distributions.\nRecent advancements have explored incorporating the second-order Fisher information, defined as the negative Hessian of log-density, into various downstream tasks and theoretical analysis of DMs.\nHowever, current practices often overlook the inherent structure of diffused distributions, accessing Fisher information via applying auto-differentiation to the learned score network. \nThis approach, while straightforward, leaves theoretical properties unexplored and is time-consuming. \nIn this paper, we derive the analytical formulation of Fisher information (AFI) by applying consecutive differentials to the diffused distributions.\nAs a result, AFI takes a gradient-free form of a weighted sum (or integral) of outer-products of the score and initial data.\nBased on this formulation, we propose two algorithmic variants of AFI for distinct scenarios.\nWhen evaluating the AFI\u2019s trace, we introduce a parameterized network to learn the trace.\nWhen AFI is applied as a linear operator, we present a training-free method that simplifies it into several inner-product calculations.\nFurthermore, we provide theoretical guarantees for both algorithms regarding convergence analysis and approximation error bounds.\nAdditionally, we leverage AFI to establish the first general theorem for the optimal transport property of the diffusion ODE deduced map.\nExperiments in likelihood evaluation and adjoint optimization demonstrate the superior accuracy and reduced time-cost of the proposed algorithms.", "keywords": [ "diffusion models; Fisher Information; analytical formulation" ], "rating_list": [ 6, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "agocj3HTTd", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Amitash Nanda", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "sree bhargavi balija", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 11, "n_ref": 17, "n_ref_all": 28, "n_fig": 6, "n_tab": 5, "L_tab": 5141, "n_element_tab": 433, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3806, "n_element_tab_1": 413, "formula_len_all": 954, "formula_len_all_1": 760, "len_all": 64579, "len_all_1": 51943, "len_abs": 1764, "len_title": 124, "len_sents": 18878, "len_sents_1": 21484, "n_sents": 130, "n_sents_1": 139, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1782, "title": "Performing Interpretability Analysis in Federated Learning Context", "abs": "Federated learning continues to evolve but faces challenges in interpretability and explainability. We introduce a creative approach employing Neural Additive Models (NAMs) within a federated learning framework to address these challenges. These models referred to as Federated Neural Additive Models (FedNAMs), merge the advantages of NAMs, where individual networks concentrate on specific input features, with the decentralized approach of federated learning, ultimately producing interpretable analysis results. This integration enhances privacy by training on local data across multiple devices, thereby minimizing the risks of data centralization and enhancing model robustness and generalizability. FedNAMs maintain detailed feature-specific learning, making them especially valuable in sectors like finance and healthcare. They facilitate training client-specific models that integrate local updates, preserving privacy and reducing centralization concerns. Our studies on various text and image classification tasks, using datasets such as OpenFetch ML Wine, UCI Heart Disease, and Iris, show that FedNAMs deliver strong interpretability with minimal accuracy loss compared to traditional Federated Deep Neural Networks (DNNs). The research involves notable findings, including the identification of key predictive features at the client level as well as at the global level. Volatile acidity, sulfates, and chlorides for wine quality. Chest pain type, maximum heart rate, and number of vessels for heart disease. Petal length and width for iris classification. This approach strengthens privacy and model efficiency and improves interpretability and robustness across diverse datasets. Finally, FedNAMs generate insights on causes of highly and low interpretable features.", "keywords": [ "Federated Learning", "Interpretability", "Neural Additive Models", "Optimization" ], "rating_list": [ 1, 1, 5 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 1, 2, 2 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "agbiPPuSeQ", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Akio Hayakawa", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Researcher" }, { "name": "Masato Ishii", "gender": "Male", "institution": "Sony Research Inc.", "country": "JP", "position": "Researcher" }, { "name": "Takashi Shibuya", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Senior Manager" }, { "name": "Yuki Mitsufuji", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Lead Research Scientist" } ], "n_formula": 34, "n_formula_1": 20, "n_ref_uni": 40, "n_ref": 81, "n_ref_all": 138, "n_fig": 6, "n_tab": 13, "L_tab": 4913, "n_element_tab": 585, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1339, "n_element_tab_1": 174, "formula_len_all": 3205, "formula_len_all_1": 1493, "len_all": 176070, "len_all_1": 65573, "len_abs": 1143, "len_title": 127, "len_sents": 52140, "len_sents_1": 28351, "n_sents": 409, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1215, "title": "Discriminator-Guided Cooperative Diffusion for Joint Audio and Video Generation", "abs": "In this study, we aim to construct an audio-video generative model with minimal computational cost by leveraging pre-trained single-modal generative models for audio and video.\nTo achieve this, we propose a novel method that guides single-modal models to cooperatively generate well-aligned samples across modalities. \nSpecifically, given two pre-trained base diffusion models, we train a lightweight joint guidance module to adjust scores separately estimated by the base models to match the score of joint distribution over audio and video. \nWe show that this guidance can be computed through the gradient of the optimal discriminator distinguishing real audio-video pairs from the fake ones independently generated by the base models. \nOn the basis of this analysis, we construct a joint guidance module by training this discriminator.\nAdditionally, we adopt a loss function to make the gradient of the discriminator work as a noise estimator, as in standard diffusion models, stabilizing the gradient of the discriminator. \nEmpirical evaluations on several benchmark datasets demonstrate that our method improves both single-modal fidelity and multi-modal alignment with a relatively small number of parameters.", "keywords": [ "Diffusion models", "Multi-modal data", "Audio-visual generative models" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "agHddsQhsL", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Boyang Zheng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Xiaoyu Wu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Caradryan Liang", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 41, "n_ref": 133, "n_ref_all": 183, "n_fig": 12, "n_tab": 10, "L_tab": 1780, "n_element_tab": 159, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1024, "n_element_tab_1": 229, "formula_len_all": 1708, "formula_len_all_1": 1001, "len_all": 189222, "len_all_1": 63905, "len_abs": 1446, "len_title": 128, "len_sents": 61177, "len_sents_1": 27006, "n_sents": 536, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 80, "L_abs": 1375, "title": "Targeted Attack Improves Protection against Unauthorized Diffusion Customization", "abs": "Diffusion models build a new milestone for image generation yet raising public concerns, for they can be fine-tuned on unauthorized images for customization. Protection based on adversarial attacks rises to encounter this unauthorized diffusion customization, by adding protective watermarks to images and poisoning diffusion models. However, current protection, leveraging untargeted attacks, does not appear to be effective enough. In this paper, we propose a simple yet effective improvement for the protection against unauthorized diffusion customization by introducing targeted attacks. We show that by carefully selecting the target, targeted attacks significantly outperform untargeted attacks in poisoning diffusion models and degrading the customization image quality. Extensive experiments validate the superiority of our method on two mainstream customization methods of diffusion models, compared to existing protections. To explain the surprising success of targeted attacks, we delve into the mechanism of attack-based protections and propose a hypothesis based on our observation, which enhances the comprehension of attack-based protections. To the best of our knowledge, we are the first to both reveal the vulnerability of diffusion models to targeted attacks and leverage targeted attacks to enhance protection against unauthorized diffusion customization.", "keywords": [ "Protection", "Unauthorized Diffusion Customization", "Adversarial Attack", "Diffusion Model", "Privacy" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "agEy9hliY1", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Hoong Chuin Lau", "gender": "Male", "institution": "Singapore Management University", "country": "", "position": "Full Professor" }, { "name": "Jingfeng Yang", "gender": "Male", "institution": "Singapore Institute of Manufacturing Technology", "country": "SG", "position": "Researcher" }, { "name": "Yining Ma", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Zhiguang Cao", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Zhiqin Zhang", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 58, "n_ref_all": 97, "n_fig": 11, "n_tab": 6, "L_tab": 3404, "n_element_tab": 181, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 3248, "n_element_tab_1": 158, "formula_len_all": 437, "formula_len_all_1": 0, "len_all": 143561, "len_all_1": 71598, "len_abs": 1324, "len_title": 129, "len_sents": 54701, "len_sents_1": 34381, "n_sents": 350, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1329, "title": "Unveiling Neural Combinatorial Optimization Model Representations Through Probing", "abs": "Neural combinatorial optimization (NCO) models have achieved remarkable performance, yet their learned underlying representations remain largely unclear. This hinders real-world application, as industrial stakeholders may want a deeper understanding of NCO models before committing resources. In this paper, we make the first step towards interpreting NCO models by investigating embeddings learned by various architectures through three probing tasks. Specifically, we analyze representative and state-of-the-art attention-based models, including AM, POMO, and LEHD, on the representative Traveling Salesman Problem and Capacitated Vehicle Routing Problem. Our findings reveal that NCO models encode linear representations of Euclidean distances between nodes, while also capturing additional knowledge that help avoid making myopic decisions. Furthermore, we show that architectural choices affect the ability of deep models to accurately represent Euclidean distances and to incorporate non-myopic decision-making strategies. We also verify to what extent NCO models understand the feasibility of constraints. Our work represents an initial effort to interpret NCO models, enhance understanding of why certain architectures outperform others, and demonstrate probing as a valuable tool for analyzing their internal mechanisms.", "keywords": [ "neural combinatorial optimization", "vehicle routing problem", "probing", "representation learning" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "afgqQYxTyR", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Artur Kasymov", "gender": "Male", "institution": "Jagiellonian University Cracow", "country": "PL", "position": "PhD student" }, { "name": "Marcin Sendera", "gender": "Male", "institution": "Mila - Quebec AI Institute", "country": "CA", "position": "Intern" }, { "name": "Micha\u0142 Stypu\u0142kowski", "gender": "Male", "institution": "Meta", "country": "GB", "position": "Intern" }, { "name": "Przemys\u0142aw Spurek", "gender": "Male", "institution": "Jagiellonian University Cracow", "country": "PL", "position": "Associate Professor" }, { "name": "Maciej Mateusz Zieba", "gender": "Male", "institution": "Technical University of Wroclaw", "country": "PL", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 39, "n_ref": 55, "n_ref_all": 67, "n_fig": 5, "n_tab": 2, "L_tab": 2044, "n_element_tab": 77, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 929, "n_element_tab_1": 193, "formula_len_all": 1294, "formula_len_all_1": 1294, "len_all": 109785, "len_all_1": 56664, "len_abs": 1059, "len_title": 69, "len_sents": 33100, "len_sents_1": 25803, "n_sents": 201, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1064, "title": "AutoLoRA: AutoGuidance Meets Low-Rank Adaptation for Diffusion Models", "abs": "Low-rank adaptation (LoRA) is a fine-tuning technique that can be applied to conditional generative diffusion models. LoRA utilizes a small number of context examples to adapt the model to a specific domain, character, style, or concept. However, due to the limited data utilized during training, the fine-tuned model performance is often characterized by strong context bias and a low degree of variability in the generated images. To solve this issue, we introduce AutoLoRA, a novel guidance technique for diffusion models fine-tuned with the LoRA approach. Inspired by other guidance techniques, AutoLoRA searches for a trade-off between consistency in the domain represented by LoRA weights and sample diversity from the base conditional diffusion model. Moreover, we show that incorporating classifier-free guidance for both LoRA fine-tuned and base models leads to generating samples with higher diversity and better quality. The experimental results for several fine-tuned LoRA domains show superiority over existing guidance techniques on selected metrics.", "keywords": [ "diffusion models", "guidance", "generative model", "LoRA" ], "rating_list": [ 3, 3, 5, 1 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 1, 2, 3, 1 ], "contribution_list": [ 1, 3, 2, 1 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "af2ztLTFqe", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chengzhao Yang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Jiakai Wang", "gender": "Male", "institution": "Zhongguancun Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Kewei Liao", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Xianglong Liu", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Yuqing Ma", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "Associate Professor" }, { "name": "Zhange Zhang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" }, { "name": "tianbo Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 22, "n_ref": 97, "n_ref_all": 121, "n_fig": 11, "n_tab": 9, "L_tab": 3720, "n_element_tab": 333, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 1845, "n_element_tab_1": 201, "formula_len_all": 1569, "formula_len_all_1": 639, "len_all": 237455, "len_all_1": 60638, "len_abs": 3078, "len_title": 122, "len_sents": 84833, "len_sents_1": 26264, "n_sents": 633, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1464, "title": "Token-Aware Inference-Time Intervention for Large Language Model Alignment", "abs": "Effectively mitigating the misalignment of large language models (LLMs) is crucial for ensuring secure AI applications. Inference-Time Intervention (ITI) technique, which applies interventions to internal representations along the probed alignment direction during inference, offers substantial alignment enhancements with minimal cost. However, previous ITI methods adopt coarse sentence-level analysis which neglects the misalignment discrepancy among varied tokens, resulting in deviant alignment direction and inflexible intervention strength.\nIn this work, we propose a Token-Aware Inference-Time Intervention (TA-ITI) approach to fully utilize token-level alignment information, therefore realizing superior post-intervention performance. TA-ITI primarily consists of Mutual Information-Guided Token-level Graph Aggregation (MIG) and Misalignment-aware Adaptive Token-level Intervention (MAI). MIG develops a MI-guided graph to exploit the tokens' informative interaction for representation enrichment, thus improving alignment probing and facilitating subsequent intervention.\nMAI comprehensively perceives the token-level misalignment degree from token representation and prediction to guide the adaptive adjustment of intervention strength, thereby enhancing final alignment performance. Extensive experiments on three alignment capabilities demonstrate the efficacy of TA-ITI, notably surpassing baseline by 25.8\\% on the primary metric of truthfulness.", "keywords": [ "LLM Alignment", "Inference-Time Intervention", "Mutual Information", "Graph Network", "Misalignment Estimation", "Uncertainty Quantification" ], "rating_list": [ 5, 6, 5, 6, 8 ], "soundness_list": [ 3, 3, 1, 4, 4 ], "presentation_list": [ 3, 2, 2, 3, 4 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3, 3 ] }, { "paper_id": "aefNwingnS", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Alice V. De Lorenci", "gender": "Female", "institution": "Facebook", "country": "FR", "position": "Intern" }, { "name": "Juan C Caicedo", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "", "position": "Assistant Professor" }, { "name": "Piotr Bojanowski", "gender": "Male", "institution": "Meta", "country": "FR", "position": "Researcher" }, { "name": "Seung Eun Yi", "gender": "Female", "institution": "Meta", "country": "", "position": "Machine Learning Engineer" }, { "name": "Th\u00e9o Moutakanni", "gender": "Male", "institution": "CentraleSupelec", "country": "FR", "position": "PhD student" }, { "name": "Wolfgang Maximilian Anton Pernice", "gender": "Male", "institution": "Columbia University", "country": "", "position": "Postdoc" }, { "name": "camille couprie", "gender": "Female", "institution": "Facebook", "country": "FR", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 6, "n_ref_uni": 25, "n_ref": 71, "n_ref_all": 105, "n_fig": 4, "n_tab": 13, "L_tab": 4011, "n_element_tab": 510, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1996, "n_element_tab_1": 244, "formula_len_all": 551, "formula_len_all_1": 358, "len_all": 168523, "len_all_1": 61533, "len_abs": 1305, "len_title": 120, "len_sents": 54159, "len_sents_1": 29327, "n_sents": 388, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1308, "title": "Scaling Channel-Invariant Self-Supervised Learning", "abs": "Recent advances in self-supervised pre-training of foundation models for natural images have made them a popular choice for various visual systems and applications. Self-supervised strategies have also shown promise in non-RGB scientific imaging domains such as in biology, medical and satellite imagery, but their broader application is hampered by heterogeneity in channel composition and semantics between relevant datasets: two datasets may contain different numbers of channels, and these may reveal distinct aspects of an object or scene. Recent works on channel-invariant strategies report substantial advantages for those that account for variable channel compositions without sacrificing the ability to jointly encode channels; yet, how these strategies behave at scale remains unclear. We here show that, surprisingly, trained across large-scale microscopy datasets, independent-encoding of channels consistently outperforms joint-encoding methods by a substantial margin. We explore this result along an extensive set of experiments and open-source a new general purpose feature extractor for fluorescent microscopy images, DINO BoC, that sets a new state-of-the-art across challenging benchmarks, including generalization to out-of-distribution tasks and unseen channel combinations at test time.", "keywords": [ "Self Supervised Learning", "Vision Transformer", "microscopy", "channel adaptive" ], "rating_list": [ 5, 3, 5, 6, 3 ], "soundness_list": [ 4, 2, 2, 3, 3 ], "presentation_list": [ 4, 2, 1, 4, 3 ], "contribution_list": [ 2, 1, 2, 3, 1 ], "confidence_list": [ 3, 3, 5, 3, 3 ] }, { "paper_id": "aeY0CAOnca", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Ayush Jain", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Erdem Biyik", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Joseph J Lim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Kyung-Min Kim", "gender": "Male", "institution": "NAVER", "country": "KR", "position": "Leader" }, { "name": "Norio Kosaka", "gender": "Male", "institution": "LINE", "country": "", "position": "Researcher" }, { "name": "Xinhu Li", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "MS student" } ], "n_formula": 40, "n_formula_1": 15, "n_ref_uni": 70, "n_ref": 126, "n_ref_all": 216, "n_fig": 33, "n_tab": 4, "L_tab": 1743, "n_element_tab": 241, "n_fig_1": 21, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6349, "formula_len_all_1": 3099, "len_all": 262672, "len_all_1": 70721, "len_abs": 1048, "len_title": 129, "len_sents": 79005, "len_sents_1": 29177, "n_sents": 648, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1051, "title": "Mitigating Suboptimality of Deterministic Policy Gradients in Complex Q-functions", "abs": "In reinforcement learning, off-policy actor-critic approaches like DDPG and TD3 are based on the deterministic policy gradient. Herein, the Q-function is trained from off-policy environment data and the actor (policy) is trained to maximize the Q-function via gradient ascent. We observe that in complex tasks like dexterous manipulation and restricted locomotion, the Q-value is a complex function of action, having several local optima or discontinuities. This poses a challenge for gradient ascent to traverse and makes the actor prone to get stuck at local optima. To address this, we introduce a new actor architecture that combines two simple insights: (i) use multiple actors and evaluate the Q-value maximizing action, and (ii) learn surrogates to the Q-function that are simpler to optimize with gradient-based methods.\nWe evaluate tasks such as restricted locomotion, dexterous manipulation, and large discrete-action space recommender systems and show that our actor finds more optimal actions and outperforms alternate actor architectures.", "keywords": [ "Deterministic Policy Gradients", "Complex Q-functions", "Actor-critic RL", "TD3." ], "rating_list": [ 8, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 4, 2 ] }, { "paper_id": "adrPcTD2cz", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Yen Liang Lai", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "MS student" }, { "name": "Yuh-Jzer Joung", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 14, "n_ref": 29, "n_ref_all": 46, "n_fig": 10, "n_tab": 4, "L_tab": 588, "n_element_tab": 39, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 588, "n_element_tab_1": 39, "formula_len_all": 328, "formula_len_all_1": 329, "len_all": 74411, "len_all_1": 52836, "len_abs": 779, "len_title": 152, "len_sents": 26829, "len_sents_1": 24860, "n_sents": 192, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 770, "title": "Weighted-Rank Contrastive Regression for Robust Learning on Imbalance Social Media Popularity Prediction", "abs": "Social Media Popularity Prediction (SMPP) is the task of forecasting the level of engagement a social media post will receive. It is crucial\nfor understanding audience engagement and enabling targeted marketing strategies. However, the inherent imbalance in real-world\nsocial media data, where certain popularity levels are underrepresented, poses a significant challenge. In this study, we leveraged the\nrecent success of contrastive learning and its growing integration into regression tasks by introducing a Weighted-Rank CR loss to\naddress the data imbalance challenges. Experiments on the Social Media Prediction Dataset demonstrated that our method outperformed\nthe vanilla approach and the current state-of-the-art contrastive regression approach Rank-N-Contrast.", "keywords": [ "Social Media Popularity Prediction", "Contrastive Learning", "Imbalance Regression", "Rank-N-Contrast", "Social Media Prediction Dataset" ], "rating_list": [ 3, 3, 1, 5 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 1, 2, 1, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "adhxppqQAn", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Cong Chen", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "US", "position": "Researcher" }, { "name": "Jean-Philippe Monteuuis", "gender": "Not Specified", "institution": "Qualcomm", "country": "US", "position": "Researcher" }, { "name": "Jonathan Petit", "gender": "Male", "institution": "Qualcomm", "country": "US", "position": "Principal Researcher" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 57, "n_ref_all": 88, "n_fig": 13, "n_tab": 5, "L_tab": 1037, "n_element_tab": 224, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 347, "n_element_tab_1": 57, "formula_len_all": 170, "formula_len_all_1": 347, "len_all": 134784, "len_all_1": 58804, "len_abs": 1214, "len_title": 109, "len_sents": 39981, "len_sents_1": 26727, "n_sents": 324, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1220, "title": "Multi-Task Consistency-based Detection of Adversarial Attacks", "abs": "Deep Neural Networks (DNNs) have found successful deployment in numerous vision perception systems. However, their susceptibility to adversarial attacks has prompted concerns regarding their practical applications, specifically in the context of autonomous driving. Existing research on defenses often suffers from cost inefficiency, rendering their deployment impractical for resource-constrained applications. In this work, we propose an efficient and effective adversarial attack detection scheme leveraging the multi-task perception within a complex vision system. Adversarial perturbations are detected by the inconsistencies between the inference outputs of multiple vision tasks, e.g., objection detection and instance segmentation. To this end, we developed a consistency score metric to measure the inconsistency between vision tasks. Next, we designed an approach to select the best model pairs for detecting this inconsistency effectively. Finally, we evaluated our defense by implementing PGD attacks across multiple vision models on the BDD100k validation dataset. The experimental results demonstrated that our defense achieved a ROC-AUC performance of 99.9% detection within the considered attacker model.", "keywords": [ "Adversarial Attack", "Object Detection", "Instance Segmentation", "Adversarial Defense" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "adSdHgWGBB", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Kaicheng Wang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Weihang Wang", "gender": "unknown", "institution": "University of Southern California", "country": "", "position": "Assistant Professor" }, { "name": "Weike Fang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 47, "n_ref": 66, "n_ref_all": 94, "n_fig": 6, "n_tab": 8, "L_tab": 2826, "n_element_tab": 228, "n_fig_1": 3, "n_tab_1": 9, "L_tab_1": 4391, "n_element_tab_1": 212, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 143562, "len_all_1": 59865, "len_abs": 1000, "len_title": 142, "len_sents": 32516, "len_sents_1": 25149, "n_sents": 216, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1002, "title": "Generate-then-Test: Automated Test Case Generation for WebAssembly Using Large Language Models", "abs": "The reliability and security of WebAssembly (Wasm) binaries are crucial for modern web development, yet effective testing methodologies remain undeveloped. This paper addresses the gap in Wasm binary testing by proposing a novel approach for test cases generation, leveraging Large Language Models (LLMs) to enhance test coverage and bug detection. Traditional testing approaches typically require access to source code, which is often unavailable for Wasm binaries. Our generate-then-test methodology overcomes this limitation by generating equivalent C++ code to simulate expected Wasm behavior, creating and mutating test cases in C++, and compiling these tests to evaluate them against the Wasm binary. Key contributions include automated test case generation using LLMs and improved code coverage through type-aware mutations, with comprehensive evaluation demonstrating the effectiveness of our approach in detecting subtle bugs in Wasm binaries, thereby ensuring more reliable Wasm applications.", "keywords": [ "WebAssembly", "Large Language Models", "Test Case Generation" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 1 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "acxHV6werE", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Jacob Steinhardt", "gender": "unknown", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" }, { "name": "Joseph E. Gonzalez", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Associate Professor" }, { "name": "Krishna Mandal", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department, University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Lisa Dunlap", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "PhD Student" }, { "name": "trevor darrell", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department", "country": "US", "position": "Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 30, "n_ref": 64, "n_ref_all": 88, "n_fig": 9, "n_tab": 7, "L_tab": 4205, "n_element_tab": 118, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1164, "n_element_tab_1": 67, "formula_len_all": 75, "formula_len_all_1": 109, "len_all": 196289, "len_all_1": 66146, "len_abs": 3423, "len_title": 127, "len_sents": 59815, "len_sents_1": 32156, "n_sents": 481, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1544, "title": "VibeCheck: Discover and Quantify Qualitative Differences in Large Language Models", "abs": "Large language models (LLMs) often exhibit subtle yet distinctive characteristics in their outputs that users intuitively recognize, but struggle to quantify. These \"vibes\" -- such as tone, formatting, or writing style -- influence user preferences, yet traditional evaluations focus primarily on the singular vibe of correctness.\nWe introduce $\\textbf{VibeCheck}$, a system for automatically comparing a pair of LLMs by discovering identifying traits of a model (\"vibes\") that are well-defined, differentiating, and user-aligned. VibeCheck iteratively discovers vibes from model outputs and then utilizes a panel of LLM judges to quantitatively measure the utility of each vibe. \nWe validate that the vibes generated by VibeCheck align with those found in human discovery and run VibeCheck on pairwise preference data from real-world user conversations with Llama-3-70b vs GPT-4. VibeCheck reveals that Llama has a friendly, funny, and somewhat controversial vibe. These vibes predict model identity with 80% accuracy and human preference with 61% accuracy. Lastly, we run VibeCheck on a variety of models and tasks, including summarization, math, and captioning to provide insight into differences in model behavior. VibeCheck discovers vibes like Command X prefers to add concrete intros and conclusions when summarizing in comparison to TNGL, Llama-405b often overexplains its thought process on math problems compared to GPT-4o, and GPT-4 prefers to focus on the mood and emotions of the scene when captioning compared to Gemini-1.5-Flash.", "keywords": [ "large language models", "evaluation" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "acPDTHPsOz", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Chuanye Wang", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Haibao Yu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jiaru Zhong", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Ruiyang Hao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Wenxian Yang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Zaiqing Nie", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Ping Luo", "gender": "unknown", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 18, "n_ref": 29, "n_ref_all": 40, "n_fig": 4, "n_tab": 7, "L_tab": 2763, "n_element_tab": 151, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1668, "n_element_tab_1": 116, "formula_len_all": 55, "formula_len_all_1": 55, "len_all": 115908, "len_all_1": 59184, "len_abs": 1969, "len_title": 150, "len_sents": 31140, "len_sents_1": 28248, "n_sents": 208, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1981, "title": "DriveE2E: Benchmarking Closed-Loop End-to-End Autonomous Driving Based-on Real-World Traffic Scenarios", "abs": "End-to-end learning has demonstrated considerable promise in advancing autonomous driving by fully leveraging sensor data. Recently, many end-to-end models have been developed, with a substantial number evaluated using the nuScenes dataset in an open-loop manner. However, open-loop evaluations, which lack interaction with the environment, fail to fully capture the driving capabilities of these models. While closed-loop evaluations, such as those using the CARLA simulator, allow for interaction with the environment, they often rely on rule-based, manually configured traffic scenarios. This approach leads to evaluations that diverge significantly from real-world driving conditions, thus limiting their ability to reflect actual driving performance.\nTo address these limitations, we introduce a novel closed-loop evaluation framework that closely integrates real-world driving scenarios with the CARLA simulator, effectively bridging the gap between simulated environments and real-world driving conditions. Our approach involves the creation of digital twins for 15 real-world intersections and the incorporation of 800 real-world traffic scenarios selected from a comprehensive 100-hour video dataset captured with highly installed infrastructure sensors. These digital twins accurately replicate the physical and environmental characteristics of their real-world counterparts, while the traffic scenarios capture a diverse range of driving behaviors, locations, weather conditions, and times of day. Within this twinned environment, CARLA enables realistic simulations where autonomous agents can dynamically interact with their surroundings. Furthermore, we have established a comprehensive closed-loop benchmark that evaluates end-to-end autonomous driving models across these diverse scenarios. Notably, this is the first closed-loop end-to-end autonomous driving benchmark based on real-world traffic scenarios. Video demos are provided in the supplementary materials.", "keywords": [ "Autonomous Driving", "End-to-End", "benchmark" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "acH47FOCTV", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Bernhard Schoelkopf", "gender": "Male", "institution": "ELLIS Institute", "country": "DE", "position": "Principal Researcher" }, { "name": "Hsiao-Ru Pan", "gender": "Male", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 24, "n_ref_uni": 57, "n_ref": 104, "n_ref_all": 134, "n_fig": 9, "n_tab": 8, "L_tab": 1711, "n_element_tab": 103, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 545, "n_element_tab_1": 28, "formula_len_all": 2716, "formula_len_all_1": 1985, "len_all": 146431, "len_all_1": 60035, "len_abs": 657, "len_title": 113, "len_sents": 37388, "len_sents_1": 26780, "n_sents": 294, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 660, "title": "Direct Advantage Estimation in Partially Observable Environments", "abs": "Direct Advantage Estimation (DAE) was recently shown to improve sample-efficiency of deep reinforcement learning algorithms. However, DAE assumes full observability of the environment, which may be restrictive in realistic settings. In the present work, we first show that DAE can be extended to partially observable domains with minor modifications. Secondly, we address the increased computational cost due to the need to approximate the transition probabilities through the use of discrete latent dynamics models. Finally, we empirically evaluate the proposed method using the Arcade Learning Environments, and show that it is scalable and sample-efficient.", "keywords": [ "POMDP", "advantage function", "deep RL", "off-policy learning" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 2, 4, 4 ], "presentation_list": [ 3, 1, 2, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "acDwoHrwZ8", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Bruno Lepri", "gender": "Male", "institution": "Fondazione Bruno Kessler", "country": "IT", "position": "Principal Researcher" }, { "name": "Gian Maria Campedelli", "gender": "Male", "institution": "Fondazione Bruno Kessler", "country": "", "position": "Research Scientist" }, { "name": "Jacopo Staiano", "gender": "unknown", "institution": "University of Trento", "country": "", "position": "Senior Assistant Professor" }, { "name": "Marco Guerini", "gender": "unknown", "institution": "Fondazione Bruno Kessler", "country": "", "position": "Principal Researcher" }, { "name": "Massimo Stefan", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "MS student" }, { "name": "Nicol\u00f2 Penzo", "gender": "Male", "institution": "Fondazione Bruno Kessler", "country": "IT", "position": "PhD student" }, { "name": "Roberto Dessi", "gender": "Male", "institution": "Universitat Pompeu Fabra", "country": "ES", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 49, "n_ref": 77, "n_ref_all": 122, "n_fig": 25, "n_tab": 9, "L_tab": 8161, "n_element_tab": 264, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 218, "n_element_tab_1": 4, "formula_len_all": 1141, "formula_len_all_1": 0, "len_all": 258516, "len_all_1": 69609, "len_abs": 1725, "len_title": 160, "len_sents": 82495, "len_sents_1": 34899, "n_sents": 640, "n_sents_1": 248, "venue": "ICLR2025", "decision": "withdraw", "L_title": 111, "L_abs": 1732, "title": "I Want to Break Free! Persuasion and Anti-Social Behavior of LLMs in Multi-Agent Settings with Social Hierarchy", "abs": "As Large Language Model (LLM)-based agents become increasingly autonomous and will more freely interact with each other, studying interactions between them becomes crucial to anticipate emergent phenomena and potential risks. Drawing inspiration from the widely popular Stanford Prison Experiment, we contribute to this line of research by studying interaction patterns of LLM agents in a context characterized by strict social hierarchy. We do so by specifically studying two types of phenomena: persuasion and anti-social behavior in simulated scenarios involving a guard and a prisoner agent who seeks to achieve a specific goal (i.e., obtaining additional yard time or escape from prison). Leveraging 200 experimental scenarios for a total of 2,000 machine-machine conversations across five different popular LLMs, we provide a set of noteworthy findings. We first document how some models consistently fail in carrying out a conversation in our multi-agent setup where power dynamics are at play.\nThen, for the models that were able to engage in successful interactions, we empirically show how the goal that an agent is set to achieve impacts primarily its persuasiveness, while having a negligible effect with respect to the agent's anti-social behavior. Third, we highlight how agents' personas, and particularly the guard's personality, drive both the likelihood of successful persuasion from the prisoner and the emergence of anti-social behaviors. Fourth, we show that even without explicitly prompting for specific personalities, anti-social behavior emerges by simply assigning agents' roles. These results bear implications for the development of interactive LLM agents as well as the debate on their societal impact.", "keywords": [ "large language models", "persuasion", "anti-social behavior", "sociology of machines", "social hierarchy", "computational social science", "experiments" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "ac93gRzxxV", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 5, "author_info_list": [ { "name": "Gagandeep Singh", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Rohan Gumaste", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Undergrad student" }, { "name": "Sasa Misailovic", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Associate Professor" }, { "name": "Shubham Ugare", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Tarun Suresh", "gender": "Male", "institution": "Department of Computer Science", "country": "", "position": "Undergrad student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 45, "n_ref_all": 56, "n_fig": 3, "n_tab": 15, "L_tab": 5571, "n_element_tab": 827, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2254, "n_element_tab_1": 405, "formula_len_all": 36, "formula_len_all_1": 0, "len_all": 165610, "len_all_1": 73044, "len_abs": 927, "len_title": 44, "len_sents": 56328, "len_sents_1": 32174, "n_sents": 426, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 44, "L_abs": 945, "title": "IterGen: Iterative Structured LLM Generation", "abs": "Large Language Models (LLMs) are widely used for tasks such as natural language and code generation, but their outputs often suffer from issues like hallucination, toxicity, and incorrect results. \nCurrent libraries for structured LLM generation rely on left-to-right decoding without support for backtracking, limiting the ability to correct or refine outputs mid-generation. \nTo address this, we introduce IterGen, a user-friendly library for iterative, grammar-guided LLM generation that enables users to move both forward and backward within the generated output based on grammar symbols. \nBy leveraging a symbol-to-position mapping and maintaining the key-value (KV) cache state, IterGen ensures efficient and structured generation while allowing for corrections during the process. \nWe demonstrate IterGen's effectiveness in two applications: reducing privacy leakage in LLM outputs and improving the accuracy of LLM-generated SQL queries.", "keywords": [ "LLM", "Grammar", "Formal Languages", "Parser", "Decoding" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 4, 2 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 2, 4, 2 ], "confidence_list": [ 5, 4, 2 ] }, { "paper_id": "abRWxnjMIz", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chulaka Gunasekara", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "David Yunis", "gender": "unknown", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "PhD student" }, { "name": "Siyu Huo", "gender": "Male", "institution": "New York University", "country": "", "position": "MS student" }, { "name": "Danish Contractor", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "Principal Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 104, "n_ref_all": 127, "n_fig": 9, "n_tab": 18, "L_tab": 23370, "n_element_tab": 3995, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 125, "formula_len_all_1": 125, "len_all": 251285, "len_all_1": 62822, "len_abs": 1498, "len_title": 59, "len_sents": 52575, "len_sents_1": 30223, "n_sents": 458, "n_sents_1": 237, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1504, "title": "Reducing the Scope of Language Models with Circuit Breakers", "abs": "Language models are now deployed in a wide variety of user-facing applications, often for specific purposes like answering questions about documentation or acting as coding assistants. As these models are intended for particular purposes, they should not be able to answer irrelevant queries like requests for poetry or questions about physics, or even worse, queries that can only be answered by humans like sensitive company policies. Instead we would like them to only answer queries corresponding to desired behavior and refuse all other requests, which we refer to as scoping. We find that, despite the use of system prompts, two representative language models can be poorly scoped and respond to queries they should not be addressing. We then conduct a comprehensive empirical evaluation of methods which could be used for scoping the behavior of language models. Among many other results, we show that a recently-proposed method for general alignment, Circuit Breakers (CB), can be adapted to scope language models to very specific tasks like sentiment analysis or summarization or even tasks with finer-grained scoping (e.g. summarizing only news articles). When compared to standard methods like fine-tuning or preference learning, CB is more robust both for out of distribution tasks, and to adversarial prompting techniques. We also show that layering SFT and CB together often results in the best of both worlds: improved performance only on relevant queries, while rejecting irrelevant ones.", "keywords": [ "large language models", "alignment", "refusal" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "abOksepKfS", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Efstratios Gavves", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Associate Professor" }, { "name": "Jiayi Shen", "gender": "Female", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Wenzhe Yin", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Yunlu Chen", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Zehao Xiao", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Cees Snoek", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "Jan-jakob Sonke", "gender": "unknown", "institution": "Netherlands Cancer Institute", "country": "NL", "position": "Full Professor" } ], "n_formula": 24, "n_formula_1": 11, "n_ref_uni": 60, "n_ref": 131, "n_ref_all": 159, "n_fig": 12, "n_tab": 5, "L_tab": 855, "n_element_tab": 80, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 509, "n_element_tab_1": 49, "formula_len_all": 2817, "formula_len_all_1": 1177, "len_all": 174056, "len_all_1": 61540, "len_abs": 1244, "len_title": 79, "len_sents": 42962, "len_sents_1": 27682, "n_sents": 348, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 31, "L_abs": 1249, "title": "Geometric Neural Process Fields", "abs": "This paper focuses on Implicit Neural Representation (INR) generalization, where models need to efficiently adapt to new signals with few observations. Specifically, for radiance field generalization, we propose Geometric Neural Processes (GeomNP) for probabilistic neural radiance field to explicitly capture uncertainty. We formulate INR generalization in a probabilistic manner, which incorporates uncertainty and directly infers the INR function distributions on limited context observations. To alleviate the information misalignment between the 2D context image and 3D discrete points in INR generalization, we introduce a set of geometric bases. The geometric bases learn to provide 3D structure information for inferring the INR function distributions. Based on the geometric bases, we model GeomNP with hierarchical latent variables. The latent variables integrate 3D information and modulate INR functions in different spatial levels, leading to better generalization of new scenes. Despite being designed for 3D tasks, the proposed method can seamlessly apply to 2D INR generalization problems. Experiments on novel view synthesis of 3D ShapeNet and DTU scenes, as well as 2D image regression, demonstrate the effectiveness of our method.", "keywords": [ "Implicit Neural Fields", "Neural Processes", "Generalization" ], "rating_list": [ 6, 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 4 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "abHtkQkumD", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "SouYoung Jin", "gender": "Female", "institution": "Dartmouth College", "country": "US", "position": "Assistant Professor" }, { "name": "Wayner Barrios", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 2, "n_ref_uni": 37, "n_ref": 100, "n_ref_all": 132, "n_fig": 12, "n_tab": 6, "L_tab": 1210, "n_element_tab": 74, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 983, "n_element_tab_1": 60, "formula_len_all": 349, "formula_len_all_1": 151, "len_all": 170628, "len_all_1": 63133, "len_abs": 1092, "len_title": 114, "len_sents": 59017, "len_sents_1": 30660, "n_sents": 433, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1096, "title": "Learnable Context-Aware Attention Mask for Multimodal Transformers", "abs": "The Self-Attention mechanism in Transformer models has shown great success across many domains, but its effectiveness can diminish in complex settings, such as multimodal tasks. This is due to the varying token granularity and the high computational cost of processing long sequences. To overcome these limitations, we propose the Learnable Context-Aware Attention Mask (LCAAM), a novel method that globally adjusts attention maps to prioritize the most important tokens in a sequence. Our approach integrates LCAAM into a BERT-like Transformer network, enhancing the Self-Attention mechanism by capturing token relationships while accounting for their contextual relevance. Additionally, we extend LCAAM to a multi-layer framework, enabling it to capture diverse information across the layers of the Transformer. Extensive experiments on datasets including MADv2, QVHighlights, ImageNet-1K, and MSRVTT demonstrate that LCAAM improves model performance while reducing redundant computations. This innovation offers a significant improvement in tackling complex tasks, such as movie understanding.", "keywords": [ "Multimodal learning", "Attention mechanisms", "Multimodal", "Learnable Masking", "Transformer Architecture" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "aapUBU9U0D", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Jian Cheng", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Junkai Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "", "position": "MS student" }, { "name": "Yang Wu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yifan Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "Full Professor" }, { "name": "Yuran Wang", "gender": "Male", "institution": "Dalian Minzu University", "country": "CN", "position": "Undergrad student" }, { "name": "Yurong Wu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 10, "n_ref": 26, "n_ref_all": 43, "n_fig": 5, "n_tab": 5, "L_tab": 1701, "n_element_tab": 180, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1329, "n_element_tab_1": 152, "formula_len_all": 24, "formula_len_all_1": 23, "len_all": 107644, "len_all_1": 59963, "len_abs": 1295, "len_title": 136, "len_sents": 36737, "len_sents_1": 28367, "n_sents": 274, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1353, "title": "Evo-Step: Evolutionary Generation and Stepwise Validation for Optimizing LLMs in OR", "abs": "Large Language Models (LLMs) have revolutionized various domains but face significant challenges in tackling optimization modeling tasks for Operations Research (OR) problems, particularly when dealing with complex problem. In this work, we propose Evo-Step-Instruct, a framework that augments existing datasets and generates high-quality fine-tuning data tailored to OR modeling tasks. Evo-Step-Instruct employs iterative problem generation to progressively increase problem complexity and stepwise validation to rigorously validate data, preventing error propagation and ensuring the quality of the generated dataset. Leveraging this framework, we fine-tune open-source LLMs, including LLaMA-3-8B and Mistral-7B, to develop Evo-Step\u2014a model that achieves state-of-the-art performance on benchmarks such as NL4OPT, MAMO, and IndustryOR. Extensive experiments demonstrate the superior performance of Evo-Step, especially in addressing complex OR tasks, with a notable 17.01\\% improvement in micro average accuracy on difficult problems. These findings highlight the effectiveness of combining structured validation with gradual problem refinement to advance the automation of decision-making processes using LLMs. The code and dataset are available at [https://anonymous.4open.science/r/Evo-Step-F5AB](https://anonymous.4open.science/r/Evo-Step-F5AB).", "keywords": [ "Large language model; Operations Research; Automated Modeling" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "aaePIQsAdq", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 7, "author_info_list": [ { "name": "Alex James Boyd", "gender": "Male", "institution": "University of California, Irvine", "country": "", "position": "PhD student" }, { "name": "Andrew Warrington", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Postdoc" }, { "name": "Cao Xiao", "gender": "Female", "institution": "GE Healthcare", "country": "US", "position": "VP of AI" }, { "name": "Padhraic Smyth", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Full Professor" }, { "name": "Parminder Bhatia", "gender": "Male", "institution": "GEHC", "country": "US", "position": "Principal Researcher" }, { "name": "Taha Kass-Hout", "gender": "Not Specified", "institution": "GE HealthCare", "country": "US", "position": "Chief Science and Technology Officer" }, { "name": "Yuxin Chang", "gender": "Female", "institution": "University of California, Irvine", "country": "", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 19, "n_ref_uni": 50, "n_ref": 111, "n_ref_all": 186, "n_fig": 13, "n_tab": 10, "L_tab": 6679, "n_element_tab": 734, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 975, "n_element_tab_1": 33, "formula_len_all": 2886, "formula_len_all_1": 1025, "len_all": 210130, "len_all_1": 80249, "len_abs": 1146, "len_title": 76, "len_sents": 62950, "len_sents_1": 34674, "n_sents": 487, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 28, "L_abs": 1153, "title": "Deep Linear Hawkes Processes", "abs": "Marked temporal point processes (MTPPs) are used to model sequences of different types of events with irregular arrival times, with broad applications ranging from healthcare and social networks to finance. We address shortcomings in existing point process models by drawing connections between modern deep state-space models (SSMs) and linear Hawkes processes (LHPs), culminating in an MTPP we call the _deep linear Hawkes process_ (DLHP). The DLHP modifies the linear differential equations in deep SSMs to be stochastic jump differential equations, akin to LHPs. After discretizing, the resulting recurrence can be implemented efficiently using a parallel scan. This brings both linear scaling and parallelism to MTPP models. This contrasts with attention-based MTPPs, which scale quadratically, and RNN-based MTPPs, which do not parallelize across the sequence length. We show empirically that DLHPs match or outperform existing models across a broad range of metrics on eight real-world datasets. Our proposed DLHP model is the first instance of the unique architectural capabilities of SSMs being leveraged to construct a new class of MTPP models.", "keywords": [ "Marked temporal point processes", "state-space models", "Hawkes processes" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "aa5hoHNheb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Li An", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "Intern" }, { "name": "Linghao Jin", "gender": "Female", "institution": "University of Southern California", "country": "", "position": "PhD student" }, { "name": "Xuezhe Ma", "gender": "Male", "institution": "USC/ISI", "country": "US", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 32, "n_ref": 43, "n_ref_all": 70, "n_fig": 9, "n_tab": 6, "L_tab": 1804, "n_element_tab": 176, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 2964, "n_element_tab_1": 203, "formula_len_all": 416, "formula_len_all_1": 413, "len_all": 140028, "len_all_1": 63715, "len_abs": 1111, "len_title": 121, "len_sents": 29407, "len_sents_1": 26809, "n_sents": 205, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1116, "title": "Towards Chapter-to-Chapter Literary Translation Via Large Language Models", "abs": "Discourse phenomena in existing document-level translation datasets are sparse, which has been a fundamental obstacle in the development of context-aware machine translation models. Moreover, most existing document-level corpora and context-aware machine translation methods rely on an unrealistic assumption on sentence-level alignments. To mitigate these issues, we first curate a novel dataset of Chinese-English literature, which consists of 132 books with intricate discourse structures. Then, we propose a more pragmatic and challenging setting for context-aware translation, termed chapter-to-chapter (Ch2Ch) translation, and investigate the performance of commonly-used machine translation models under this setting. Furthermore, we introduce a potential approach of fine-tuning large language models (LLMs) within the domain of Ch2Ch literary translation, yielding impressive improvements over baselines. Through our comprehensive analysis, we unveil that literary translation under the Ch2Ch setting is challenging in nature, with respect to both model learning methods and translation decoding algorithms.", "keywords": [ "neural machine translation", "context-aware neural machine translation", "literary translation" ], "rating_list": [ 3, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "aZjOk7wmWf", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 9, "author_info_list": [ { "name": "Chengyu Lai", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "HONGJIA XU", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jiajun Bu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Jiawei Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Ming Gu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Sheng Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Yao Ma", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "Assistant Professor" }, { "name": "Yuanchen Bei", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zhuonan Zheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 72, "n_formula_1": 11, "n_ref_uni": 42, "n_ref": 214, "n_ref_all": 245, "n_fig": 4, "n_tab": 13, "L_tab": 8421, "n_element_tab": 782, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 5587, "n_element_tab_1": 251, "formula_len_all": 4956, "formula_len_all_1": 599, "len_all": 260466, "len_all_1": 68299, "len_abs": 2134, "len_title": 125, "len_sents": 84813, "len_sents_1": 28302, "n_sents": 657, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1679, "title": "Exploring and Unleashing the Power of Message Passing on Heterophilous Graphs", "abs": "Graph Neural Networks (GNNs) have demonstrated strong performance in graph mining tasks due to their message-passing mechanism, which is aligned with the homophily assumption that adjacent nodes exhibit similar behaviors. However, in many real-world graphs, connected nodes may display contrasting behaviors, termed as heterophilous patterns, which has attracted increased interest in heterophilous GNNs (HTGNN).\nAlthough the message-passing mechanism seems unsuitable for heterophilous graphs due to the propagation of class-irrelevant information, it is still widely used in many existing HTGNNs and consistently achieves notable success. \nThis raises the question: why does message passing remain effective on heterophilous graphs?\nTo answer this question, in this paper, we revisit the message-passing mechanisms in heterophilous graph neural networks and reformulate them into a unified heterophilious message-passing (HTMP) mechanism.\nBased on HTMP and empirical analysis, we reveal that the success of message passing in existing HTGNNs is attributed to implicitly enhancing the compatibility matrix among classes.\nMoreover, we argue that the full potential of the compatibility matrix is not completely achieved due to the existence of incomplete and noisy semantic neighborhoods in real-world heterophilous graphs.\nTo bridge this gap, we introduce a new approach named CMGNN, which operates within the HTMP mechanism to explicitly leverage and improve the compatibility matrix.\nA thorough evaluation involving 10 benchmark datasets and comparative analysis against 17 well-established baselines highlights the superior performance of the HTMP mechanism and CMGNN method.", "keywords": [ "Heterophily", "Message Passing", "Graph Neural Networks" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "aZVRFIDhYL", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Jingbo Xu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Qiaoqiao Ding", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Researcher" }, { "name": "Xiaoqun Zhang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "", "position": "Full Professor" } ], "n_formula": 37, "n_formula_1": 34, "n_ref_uni": 22, "n_ref": 38, "n_ref_all": 54, "n_fig": 4, "n_tab": 3, "L_tab": 1013, "n_element_tab": 77, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 768, "n_element_tab_1": 77, "formula_len_all": 2188, "formula_len_all_1": 2266, "len_all": 88231, "len_all_1": 58132, "len_abs": 1145, "len_title": 121, "len_sents": 26809, "len_sents_1": 25312, "n_sents": 190, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1150, "title": "Efficient Diffusion Posterior Sampling for Dose Reduced CT Reconstruction", "abs": "The clinical efficacy of Computed Tomography (CT) is well-established, yet concerns regarding its radiation exposure persist. To mitigate this risk, a reduction in X-ray photon count or projection views is typically pursued, albeit at the expense of image quality. In this study, we introduce an innovative diffusion posterior sampling approach for CT image reconstruction at reduced radiation doses. This method initiates with a predictive step, leveraging data enhancement on the posterior approximation derived from a pre-trained diffusion model and the measurement data. Subsequently, a forward sampling phase ensues, which maps the output to a noisy timestep, followed by a diffusion estimation process. Additionally, we propose an acceleration strategy that employs superior initialization to significantly curtail the sampling steps required. Our experimental findings indicate that this method not only enhances the quality of reconstructed images by an average of 3.5 db but also accelerates the process to over ten times faster than existing diffusion-based techniques. These outcomes underscore the method's potential in clinical settings.", "keywords": [ "CT reconstruction", "Diffusion model", "Posterior sampling" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "aZ1gNJu8wO", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Brendan Leigh Ross", "gender": "Male", "institution": "Layer 6 AI", "country": "", "position": "Machine Learning Scientist" }, { "name": "Gabriel Loaiza-Ganem", "gender": "Male", "institution": "Layer 6 AI", "country": "", "position": "Machine Learning Research Scientist" }, { "name": "George Stein", "gender": "unknown", "institution": "Layer6 AI", "country": "", "position": "Machine Learning Scientist" }, { "name": "Hamidreza Kamkari", "gender": "Male", "institution": "Layer6 AI", "country": "CA", "position": "Researcher" }, { "name": "Rasa Hosseinzadeh", "gender": "Male", "institution": "Layer6", "country": "", "position": "Researcher" }, { "name": "Tongzi Wu", "gender": "Female", "institution": "Layer 6 AI", "country": "CA", "position": "Machine Learning Scientist" }, { "name": "Zhaoyan Liu", "gender": "Male", "institution": "Layer6 AI", "country": "CA", "position": "Machine Learning Scientist" }, { "name": "Jesse C Cresswell", "gender": "unknown", "institution": "Layer 6 AI", "country": "CA", "position": "Staff Machine Learning Scientist" } ], "n_formula": 35, "n_formula_1": 3, "n_ref_uni": 54, "n_ref": 150, "n_ref_all": 228, "n_fig": 25, "n_tab": 3, "L_tab": 112, "n_element_tab": 7, "n_fig_1": 17, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2186, "formula_len_all_1": 112, "len_all": 220964, "len_all_1": 67557, "len_abs": 1242, "len_title": 121, "len_sents": 75295, "len_sents_1": 33607, "n_sents": 556, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 73, "L_abs": 1252, "title": "A Geometric Framework for Understanding Memorization in Generative Models", "abs": "As deep generative models have progressed, recent work has shown them to be capable of memorizing and reproducing training datapoints when deployed. These findings call into question the usability of generative models, especially in light of the legal and privacy risks brought about by memorization. To better understand this phenomenon, we propose the *manifold memorization hypothesis* (MMH), a geometric framework which leverages the manifold hypothesis into a clear language in which to reason about memorization. We propose to analyze memorization in terms of the relationship between the dimensionalities of $(i)$ the ground truth data manifold and $(ii)$ the manifold learned by the model. This framework provides a formal standard for \"how memorized\" a datapoint is and systematically categorizes memorized data into two types: memorization driven by overfitting and memorization driven by the underlying data distribution. By analyzing prior work in the context of the MMH, we explain and unify assorted observations in the literature. We empirically validate the MMH using synthetic data and image datasets up to the scale of Stable Diffusion, developing new tools for detecting and preventing generation of memorized samples in the process.", "keywords": [ "deep generative modelling", "generative models", "memorization", "data copying", "privacy", "diffusion", "diffusion models", "GANs", "manifold hypothesis", "local intrinsic dimension", "lid", "lid estimation", "geometry" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 2, 4, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "aYx7JR20sI", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Anthea Monod", "gender": "Female", "institution": "Imperial College London", "country": "GB", "position": "Associate Professor" }, { "name": "Paul Lezeau", "gender": "Male", "institution": "Adecco", "country": "GB", "position": "Engineer" }, { "name": "Shiv Bhatia", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Thomas Walker", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "MS student" }, { "name": "Yueqi Cao", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 80, "n_formula_1": 21, "n_ref_uni": 25, "n_ref": 70, "n_ref_all": 95, "n_fig": 7, "n_tab": 3, "L_tab": 8213, "n_element_tab": 95, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5344, "formula_len_all_1": 1135, "len_all": 195262, "len_all_1": 64920, "len_abs": 1627, "len_title": 40, "len_sents": 66238, "len_sents_1": 31599, "n_sents": 644, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1639, "title": "Tropical Expressivity of Neural Networks", "abs": "We propose an algebraic geometric framework to study the expressivity of piecewise linear activation neural networks. A particular quantity of neural networks that has been actively studied is the number of linear regions, which gives a quantification of the information capacity of the architecture. To study and evaluate information capacity and expressivity, we work in the setting of tropical geometry---a combinatorial and polyhedral variant of algebraic geometry---where there are known connections between tropical rational maps and feedforward neural networks. Our work builds on and expands this connection to capitalize on the rich theory of tropical geometry to characterize and study various architectural aspects of neural networks. Our contributions are threefold: we provide a novel tropical geometric approach to selecting sampling domains among linear regions; an algebraic result allowing for a guided restriction of the sampling domain for network architectures with symmetries; and a new open source OSCAR library to analyze neural networks symbolically using their tropical representations, where we present a new algorithm that computes the exact number of their linear regions. We provide a comprehensive set of proof-of-concept numerical experiments demonstrating the breadth of neural network architectures to which tropical geometric theory can be applied to reveal insights on expressivity characteristics of a network. Our work provides the foundations for the adaptation of both theory and existing software from computational tropical geometry and symbolic computation to neural networks and deep learning.", "keywords": [ "fundamental domain", "linear regions", "sampling", "tropical geometry", "symbolic computing" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "aYwHiDkAdI", "primary_area": "applications to robotics, autonomy, planning", "n_author": 27, "author_info_list": [ { "name": "Bang Liu", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Assistant Professor" }, { "name": "Bangbang Liu", "gender": "Male", "institution": "deepwisdom", "country": "", "position": "Researcher" }, { "name": "Binhao Wu", "gender": "unknown", "institution": "DeepWisdom", "country": "", "position": "engineer" }, { "name": "Ceyao Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Intern" }, { "name": "Chenxing Wei", "gender": "Male", "institution": "Shenzhen University", "country": "", "position": "MS student" }, { "name": "Danyang Li", "gender": "Female", "institution": "DeepWisdom", "country": "CN", "position": "Researcher" }, { "name": "Jiaqi Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Jiayi Zhang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jinlin Wang", "gender": "Male", "institution": "Xiamen University", "country": "", "position": "MS student" }, { "name": "Li Zhang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Lingyao Zhang", "gender": "unknown", "institution": "Aurora Innovation Inc", "country": "", "position": "software engineer" }, { "name": "Min Yang", "gender": "Female", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Mingchen Zhuge", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Sirui Hong", "gender": "Female", "institution": "DeepWisdom", "country": "CN", "position": "Researcher" }, { "name": "Taicheng Guo", "gender": "Male", "institution": "University of Notre Dame", "country": "", "position": "PhD student" }, { "name": "Tuo Zhou", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Wei Tao", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Blind Name", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Xiawu Zheng", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Xinbing Liang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Yaying Fei", "gender": "Female", "institution": "Southern Methodist University", "country": "US", "position": "PhD student" }, { "name": "Yizhang Lin", "gender": "Male", "institution": "DeepWisdom", "country": "CN", "position": "Researcher" }, { "name": "Zhibin Gou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Zongze Xu", "gender": "Male", "institution": "Hohai University", "country": "CN", "position": "PhD student" }, { "name": "Alexander Wu", "gender": "Male", "institution": "DeepWisdom", "country": "", "position": "CEO" }, { "name": "xiangtao lu", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "derry chan", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 31, "n_ref": 60, "n_ref_all": 98, "n_fig": 11, "n_tab": 14, "L_tab": 5294, "n_element_tab": 653, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 1783, "n_element_tab_1": 180, "formula_len_all": 102, "formula_len_all_1": 102, "len_all": 189515, "len_all_1": 65501, "len_abs": 1515, "len_title": 95, "len_sents": 49869, "len_sents_1": 30354, "n_sents": 379, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1525, "title": "Data Interpreter: An LLM Agent For Data Science", "abs": "Large Language Model (LLM)-based agents have shown effectiveness across many applications. However, their use in data science scenarios requiring solving long-term interconnected tasks, dynamic data adjustments and domain expertise remains challenging. Previous approaches primarily focus on individual tasks, making it difficult to assess the complete data science workflow. Moreover, they struggle\nto handle real-time changes in intermediate data and fail to adapt dynamically to evolving task dependencies inherent to data science problems. In this paper, we present Data Interpreter, an LLM-based agent designed to automatically solve various data science problems end-to-end. Our Data Interpreter incorporates two key modules: 1) Hierarchical Graph Modeling, which breaks down complex problems into manageable subproblems, enabling dynamic node generation and graph optimization; and 2) Programmable Node Generation, a technique that refines and verifies each subproblem to iteratively improve code generation results and robustness. Extensive experiments consistently demonstrate the superiority of Data Interpreter. On InfiAgent-DABench, it achieves a 25% performance boost, raising accuracy from 75.9% to 94.9%. For machine learning and open-ended tasks, it improves performance from 88% to 95%, and from 60% to 97%, respectively. Moreover, on the MATH dataset, Data Interpreter achieves remarkable performance with a 26% improvement compared to state-of-the-art baselines. Code will be open-sourced upon publication.", "keywords": [ "Interactive Agent", "Large Language Models", "Dynamic Planning", "AI for Data Science" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "aYYZBPoSHb", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Jayant Parashar", "gender": "Male", "institution": "University of Georgia", "country": "", "position": "PhD student" }, { "name": "Suchendra Bhandarkar", "gender": "Male", "institution": "University of Georgia", "country": "", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 22, "n_ref": 34, "n_ref_all": 53, "n_fig": 3, "n_tab": 2, "L_tab": 177, "n_element_tab": 15, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 177, "n_element_tab_1": 15, "formula_len_all": 704, "formula_len_all_1": 704, "len_all": 84712, "len_all_1": 68980, "len_abs": 1510, "len_title": 112, "len_sents": 25782, "len_sents_1": 25836, "n_sents": 194, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1515, "title": "Multi-Objective Alignment of LLMs with ORPO using Self-Judgement", "abs": "The alignment of Large Language Models (LLMs) is achieved through fine-tuning with human preference data, where preference optimization has become a critical part of the process. Many methods have scaled LLM performance by incorporating self-judgement, highlighting the importance of unifying LLM-as-a-judge with the alignment process. One such method, called Self-rewarding LLMs, iteratively samples new data from the model to improve alignment using self-judgement. Since this additional data is generated by the LLM, we argue that similar improvements can be achieved without new data. We propose a method that reuses alignment data in the form of a self-judgement classification task and defines a multi-objective optimization problem. Our self-judgement task is derived from a simple transformation of the primary alignment data, asking the LLM to select the superior response. It introduces no new data beyond the existing alignment data. Thus, we claim the improvements are due to positive interference between the two tasks. We focus on a direct preference optimization method called Odds-Ratio Preference Optimization (ORPO). We conduct a thorough study of linear scalarization on two objectives and introduce two alternative approaches that vary the emphasis on alignment versus self-judgement objectives. Our results on Mistral 7B indicate a promising direction for fine-tuning LLMs on multiple objectives, particularly for improving performance on related tasks without additional natural language data.", "keywords": [ "LLMs", "preference optimization", "self-judgement", "LLM-as-a-judge", "supervised fine-tuning", "multi-objective optimization", "multi-task learning", "supervised learning" ], "rating_list": [ 3, 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 1, 3, 2 ], "presentation_list": [ 1, 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "aY3W95jLEI", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Kejie Wang", "gender": "unknown", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "MS student" }, { "name": "Liqiang Nie", "gender": "Male", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "dean" }, { "name": "Meng Liu", "gender": "Female", "institution": "Shandong Jianzhu University", "country": "CN", "position": "Full Professor" }, { "name": "Xuemeng Song", "gender": "Female", "institution": "Shandong University", "country": "CN", "position": "Associate Professor" }, { "name": "WEILI GUAN", "gender": "Female", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 32, "n_ref": 82, "n_ref_all": 103, "n_fig": 9, "n_tab": 2, "L_tab": 1253, "n_element_tab": 195, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1656, "n_element_tab_1": 216, "formula_len_all": 854, "formula_len_all_1": 643, "len_all": 126699, "len_all_1": 70999, "len_abs": 1420, "len_title": 129, "len_sents": 39476, "len_sents_1": 32913, "n_sents": 266, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1427, "title": "Vision-guided and Mask-enhanced Adaptive Denoising for Prompt-based Image Editing", "abs": "Text-to-image diffusion models have demonstrated remarkable progress in synthesizing high-quality images from text prompts, which boosts researches on prompt-based image editing that edits a source image according to a target prompt. Despite their advances, existing methods still encounter three key issues: 1) limited capacity of the text prompt in guiding target image generation, 2) insufficient mining of word-to-patch and patch-to-patch relationships for grounding editing areas, and 3) unified editing strength for all regions during each denoising step. To address these issues, we present a Vision-guided and Mask-enhanced Adaptive Editing (ViMAEdit) method with three key novel designs. First, we propose to leverage image embeddings as explicit guidance to enhance the conventional textual prompt-based denoising process, where a CLIP-based target image embedding estimation strategy is introduced. Second, we devise a self-attention-guided iterative editing area grounding strategy, which iteratively exploits patch-to-patch relationships conveyed by self-attention maps to refine those word-to-patch relationships contained in cross-attention maps. Last, we present a spatially adaptive variance-guided sampling, which highlights sampling variances for critical image regions to promote the editing capability. Experimental results demonstrate the superior editing capacity of ViMAEdit over all existing methods.", "keywords": [ "image editing", "diffusion models" ], "rating_list": [ 6, 1, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 3, 5 ] }, { "paper_id": "aY3L65HgHJ", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Huaian Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" }, { "name": "Jiaqi Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Jiazi Bu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Pan Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Pengyang Ling", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Tong Wu", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Xiaoyi Dong", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yi Jin", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yuhang Zang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 74, "n_ref_all": 89, "n_fig": 10, "n_tab": 1, "L_tab": 286, "n_element_tab": 37, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 286, "n_element_tab_1": 37, "formula_len_all": 503, "formula_len_all_1": 504, "len_all": 121637, "len_all_1": 52179, "len_abs": 1275, "len_title": 123, "len_sents": 33486, "len_sents_1": 24268, "n_sents": 221, "n_sents_1": 153, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1280, "title": "MotionClone: Training-Free Motion Cloning for Controllable Video Generation", "abs": "Motion-based controllable video generation offers the potential for creating captivating visual content. Existing methods typically necessitate model training to encode particular motion cues or incorporate fine-tuning to inject certain motion patterns, resulting in limited flexibility and generalization.\nIn this work, we propose MotionClone, a training-free framework that enables motion cloning from reference videos to versatile motion-controlled video generation, including text-to-video and image-to-video. Based on the observation that the dominant components in temporal-attention maps drive motion synthesis, while the rest mainly capture noisy or very subtle motions, MotionClone utilizes sparse temporal attention weights as motion representations for motion guidance, facilitating diverse motion transfer across varying scenarios. Meanwhile, MotionClone allows for the direct extraction of motion representation through a single denoising step, bypassing the cumbersome inversion processes and thus promoting both efficiency and flexibility. \nExtensive experiments demonstrate that MotionClone exhibits proficiency in both global camera motion and local object motion, with notable superiority in terms of motion fidelity, textual alignment, and temporal consistency.", "keywords": [ "controllable video image", "viusal representation", "text-to-video generation" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "aXwukBD6M6", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Aditya Chattopadhyay", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Armand Comas", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "PhD student" }, { "name": "Changyu Liu", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Feliu Formosa", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "Intern" }, { "name": "Rene Vidal", "gender": "unknown", "institution": "University of Pennsylvania", "country": "US", "position": "Full Professor" }, { "name": "Octavia Camps", "gender": "Female", "institution": "Northeastern University", "country": "US", "position": "Professor" } ], "n_formula": 22, "n_formula_1": 7, "n_ref_uni": 34, "n_ref": 64, "n_ref_all": 107, "n_fig": 26, "n_tab": 7, "L_tab": 2241, "n_element_tab": 96, "n_fig_1": 17, "n_tab_1": 2, "L_tab_1": 450, "n_element_tab_1": 40, "formula_len_all": 1695, "formula_len_all_1": 504, "len_all": 205318, "len_all_1": 66564, "len_abs": 724, "len_title": 106, "len_sents": 69897, "len_sents_1": 30835, "n_sents": 599, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 58, "L_abs": 726, "title": "Interpretable Compressed Descriptions For Image Generation", "abs": "Generative models can be applied in diverse domains, from natural language processing to image synthesis. A key aspect to control the generation process is the definition of adequate data representations, allowing users to access and efficiently manipulate the semantic factors shaping the data distribution.\nThis work advocates for the adoption of succinct, informative, and interpretable descriptions, quantified using information theoretic principles. Through extensive experiments, we demonstrate the efficacy of this proposed framework both qualitatively and quantitatively. We conclude that it significantly contributes to the ongoing quest to enhance both controllability and interpretability in the generation process.", "keywords": [ "Controllable Generation", "Image Generation", "Interpretability", "Information Pursuit", "Information Theory", "Diffusion Models" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "aXuWowhIYt", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Andreas Krause", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Bernhard Schoelkopf", "gender": "Male", "institution": "ELLIS Institute", "country": "DE", "position": "Principal Researcher" }, { "name": "Lars Lorch", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Scott Sussex", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Weronika Ormaniec", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" } ], "n_formula": 92, "n_formula_1": 15, "n_ref_uni": 41, "n_ref": 112, "n_ref_all": 239, "n_fig": 35, "n_tab": 2, "L_tab": 2228, "n_element_tab": 230, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 260, "n_element_tab_1": 193, "formula_len_all": 8561, "formula_len_all_1": 919, "len_all": 289445, "len_all_1": 79804, "len_abs": 1190, "len_title": 86, "len_sents": 102230, "len_sents_1": 35892, "n_sents": 850, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 38, "L_abs": 1269, "title": "Standardizing Structural Causal Models", "abs": "Synthetic datasets generated by structural causal models (SCMs) are commonly used for benchmarking causal structure learning algorithms. However, the variances and pairwise correlations in SCM data tend to increase along the causal ordering. Several popular algorithms exploit these artifacts, possibly leading to conclusions that do not generalize to real-world settings. Existing metrics like $\\operatorname{Var}$-sortability and $\\operatorname{R^2}$-sortability quantify these patterns, but they do not provide tools to remedy them. To address this, we propose internally-standardized structural causal models (iSCMs), a modification of SCMs that introduces a standardization operation at each variable during the generative process. By construction, iSCMs are not $\\operatorname{Var}$-sortable. We also find empirical evidence that they are mostly not $\\operatorname{R^2}$-sortable either for commonly-used graph families. Moreover, contrary to the post-hoc standardization of data generated by standard SCMs, we prove that linear iSCMs are less identifiable from prior knowledge on the weights and do not collapse to deterministic relationships in large systems, which may make iSCMs a useful model in causal inference beyond the benchmarking problem studied here.", "keywords": [ "Causality", "Causal Discovery", "Structural Causal Model", "Standardized Structural Causal Model", "Variance Artifact", "Covariance Artifact", "Simulation", "Benchmark" ], "rating_list": [ 8, 5, 6, 8, 8 ], "soundness_list": [ 2, 2, 4, 3, 3 ], "presentation_list": [ 3, 2, 4, 3, 3 ], "contribution_list": [ 2, 1, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3, 3 ] }, { "paper_id": "aXSxSu3fvg", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Jialin Wu", "gender": "unknown", "institution": "Weill Cornell Medicine, Cornell University", "country": "US", "position": "Researcher" }, { "name": "Samprit Banerjee", "gender": "unknown", "institution": "Weill Cornell Medicine, Cornell University", "country": "US", "position": "Associate Professor" }, { "name": "Soohyun Kim", "gender": "unknown", "institution": "Weill Cornell Medicine, Cornell University", "country": "US", "position": "Postdoc" }, { "name": "Younghoon Kim", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 23, "n_ref": 25, "n_ref_all": 36, "n_fig": 4, "n_tab": 3, "L_tab": 637, "n_element_tab": 72, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 637, "n_element_tab_1": 72, "formula_len_all": 591, "formula_len_all_1": 557, "len_all": 83221, "len_all_1": 73963, "len_abs": 1863, "len_title": 129, "len_sents": 27980, "len_sents_1": 27297, "n_sents": 209, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1875, "title": "Training Semi-Supervised Deep Learning Models with Heuristic Early Stopping Rules", "abs": "Semi-supervised learning (SSL), especially when combined with deep learning (DL) models, is a useful technique when there is a substantial amount of unlabeled data. This is particularly relevant in healthcare applications, such as mHealth, where data is often collected through smartphones. Labels are typically obtained via self-reported questions delivered by the device and tend to have a high rate of non-response i.e., missing labels. Despite its benefit, there is a lack of objective methodology on how to train semi-supervised deep learning (SSDL) models. In this study, we propose a framework for early-stopping in SSDL that terminates learning to prevent overfitting and before the performance starts to deteriorate. Our approach focuses on three aspects: model stability, generalizability, and high-confidence pseudo-label (i.e., label assigned to unlabeled data during SSL). We first monitor changes in learned weights of the model to assess convergence, using weight stabilization. We also track cross-entropy loss, identifying which iteration of the SSL algorithm minimizes validation loss and improves generalizability. Lastly, we use a sliding window method to assess our confidence in the pseudo-labels, retaining only the most reliable labels during training. Combining these criteria, this SSDL framework can be used to train deep learning models in the context of SSL with an objective criteria that prevents overfitting and improves generalizability. We apply this SSDL training strategy to mHealth data (device sensor data and self-reported data) collected from participants in a clinical trial, which consists of 4,700 observations, 62% of which are unlabeled. Using this objective early stopping criteria for training, we achieve improvements in accuracy and F1 scores, compared to the benchmark model where the early stopping criteria is not applied.", "keywords": [ "semi-supervised deep learning", "neural network", "convergence", "generalizability", "predictive modeling", "model optimization" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "aXPOA3urmA", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "En-Ya Pi", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "Ping-Chun Hsieh", "gender": "Male", "institution": "National Chiao Tung University", "country": "", "position": "Assistant Professor" }, { "name": "Shao-Hua Sun", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "Assistant Professor" }, { "name": "Ting-Hsuan Huang", "gender": "Female", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 63, "n_ref": 104, "n_ref_all": 136, "n_fig": 17, "n_tab": 3, "L_tab": 340, "n_element_tab": 49, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 502, "formula_len_all_1": 381, "len_all": 175181, "len_all_1": 65600, "len_abs": 1336, "len_title": 110, "len_sents": 50547, "len_sents_1": 31259, "n_sents": 338, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1340, "title": "Cross-Domain Reinforcement Learning via Preference Consistency", "abs": "Cross-domain reinforcement learning (CDRL) aims to utilize the knowledge acquired from a source domain to efficiently learn tasks in a target domain. Unsupervised CDRL assumes no access to any signal (e.g., rewards) from the target domain, and most methods utilize state-action correspondence or cycle consistency. In this work, we identify the critical correspondence identifiability issue (CII) that arises in existing unsupervised CDRL methods. To address this identifiability issue, we propose leveraging pairwise trajectory preferences in the target domain as weak supervision. Specifically, we introduce the principle of cross-domain preference consistency (CDPC)\u2013a policy is more transferable across the domains if the source and target domains have similar preferences over trajectories\u2013to provide additional guidance for establishing proper correspondence between the source and target domains. To substantiate the principle of CDPC, we present an algorithm that integrates a state decoder learned through preference consistency loss during training with a cross-domain MPC method for action selection during inference. Through extensive experiments in both MuJoCo and Robosuite, we demonstrate that CDPC enables effective and data-efficient knowledge transfer across domains, outperforming state-of-the-art CDRL benchmark methods.", "keywords": [ "Reinforcement learning", "Cross-domain transfer", "Transfer learning", "Preference-based RL" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "aX7X9z3vQS", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Abigail Hickok", "gender": "Female", "institution": "Yale University", "country": "US", "position": "Postdoc" }, { "name": "Andrew Blumberg", "gender": "Male", "institution": "Columbia University", "country": "", "position": "Full Professor" }, { "name": "Tristan Luca Saidi", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "PhD student" } ], "n_formula": 219, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 60, "n_ref_all": 134, "n_fig": 24, "n_tab": 4, "L_tab": 1235, "n_element_tab": 135, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 25778, "formula_len_all_1": 1111, "len_all": 319112, "len_all_1": 63595, "len_abs": 1002, "len_title": 108, "len_sents": 111618, "len_sents_1": 29833, "n_sents": 1052, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 60, "L_abs": 1005, "title": "Recovering Manifold Structure Using Ollivier Ricci Curvature", "abs": "We introduce ORC-ManL, a new algorithm to prune spurious edges from nearest neighbor graphs using a criterion based on Ollivier-Ricci curvature and estimated metric distortion. Our motivation comes from manifold learning: we show that when the data generating the nearest-neighbor graph consists of noisy samples from a low-dimensional manifold, edges that shortcut through the ambient space have more negative Ollivier-Ricci curvature than edges that lie along the data manifold. We demonstrate that our method outperforms alternative pruning methods and that it significantly improves performance on many downstream geometric data analysis tasks that use nearest neighbor graphs as input. Specifically, we evaluate on manifold learning, persistent homology, dimension estimation, and others. We also show that ORC-ManL can be used to improve clustering and manifold learning of single-cell RNA sequencing data. Finally, we provide empirical convergence experiments that support our theoretical findings.", "keywords": [ "Manifold Learning", "Persistent Homology", "Ollivier-Ricci Curvature", "Pruning", "Nearest-Neighbor Graphs" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 2, 3 ] }, { "paper_id": "aWkAKucZMR", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Byeongyeon So", "gender": "unknown", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Chiehyeon Lim", "gender": "unknown", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Hyunwoo Seo", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 25, "n_ref": 68, "n_ref_all": 93, "n_fig": 11, "n_tab": 15, "L_tab": 4745, "n_element_tab": 547, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 10644, "n_element_tab_1": 404, "formula_len_all": 921, "formula_len_all_1": 442, "len_all": 153323, "len_all_1": 73674, "len_abs": 3413, "len_title": 145, "len_sents": 53282, "len_sents_1": 27419, "n_sents": 421, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1335, "title": "ShuffleMTM: Learning Cross-channel Dependence in Multivariate Time Series from Shuffled Patches", "abs": "Masked time-series modeling has widely gained attention as a self-supervised pre-training method for multivariate time series (MTS). Recent studies adopt a channel-independent (CI) strategy to enhance the temporal modeling capacity. Despite the effectiveness and performance of this strategy, the CI methods inherently overlook cross-channel dependence, which is inherent and crucial in MTS data in various domains. To fill this gap, we propose ShuffleMTM, a simple yet effective masked time-series modeling framework to learn cross-channel dependence from shuffled patches. Technically, ShuffleMTM proposes to shuffle the unmasked patches from masked series across different channels, positioned at the same index. Then, Siamese encoders learn two views of masked patch representations from original and shuffled masked series, simultaneously capturing the temporal dependence within a channel as well as spatial dependence across different channels. ShuffleMTM pre-trains the Siamese encoders to reconstruct the original series by incorporating cross-channel information with intra-channel cross-time information. Our proposed method consistently achieves superior performance in various experiments, compared to advanced CI pre-training methods and channel-dependent methods in both time series forecasting and classification tasks.", "keywords": [ "Self-supervised learning", "masked modeling", "multivariate time series", "cross-channel dependence" ], "rating_list": [ 3, 6, 8, 5 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "aWXnKanInf", "primary_area": "applications to neuroscience & cognitive science", "n_author": 6, "author_info_list": [ { "name": "Badr AlKhamissi", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "US", "position": "PhD student" }, { "name": "Johannes Mehrer", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "Postdoc" }, { "name": "Martin Schrimpf", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Assistant Professor" }, { "name": "Neil Rathi", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Nicholas Blauch", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Taha Osama A Binhuraib", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 42, "n_ref": 121, "n_ref_all": 165, "n_fig": 16, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 65, "n_element_tab_1": 10, "formula_len_all": 129, "formula_len_all_1": 69, "len_all": 167739, "len_all_1": 62347, "len_abs": 1208, "len_title": 130, "len_sents": 52206, "len_sents_1": 31570, "n_sents": 358, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 81, "L_abs": 1211, "title": "TopoLM: brain-like spatio-functional organization in a topographic language model", "abs": "Neurons in the brain are spatially organized such that neighbors on tissue often exhibit similar response profiles. In the human language system, experimental studies have observed clusters for syntactic and semantic categories, but the mechanisms underlying this functional organization remain unclear. Here, building on work from the vision literature, we develop TopoLM, a transformer language model with an explicit two-dimensional spatial representation of model units. By combining a next-token prediction objective with a spatial smoothness loss, representations in this model assemble into clusters that correspond to semantically interpretable groupings of text and closely match the functional organization in the brain's language system. TopoLM successfully predicts the emergence of the spatio-functional organization of a cortical language system as well as the organization of functional clusters selective for fine-grained linguistic features empirically observed in human cortex. Our results suggest that the functional organization of the human language system is driven by a unified spatial objective, and provide a functionally and spatially aligned model of language processing in the brain.", "keywords": [ "language modeling", "topography", "fMRI", "neuroscience" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 3, 4, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "aWLQTbfFgV", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Alexandra Butoi", "gender": "Female", "institution": "ETHZ - ETH Zurich", "country": "", "position": "PhD student" }, { "name": "Anej Svete", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Brian DuSell", "gender": "unknown", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Ghazal Khalighinejad", "gender": "Female", "institution": "Department of Computer Science, Duke University", "country": "", "position": "PhD student" }, { "name": "Josef Valvoda", "gender": "Male", "institution": "University of Copenhagen", "country": "DK", "position": "Postdoc" }, { "name": "Ryan Cotterell", "gender": "Not Specified", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 62, "n_formula_1": 19, "n_ref_uni": 41, "n_ref": 139, "n_ref_all": 169, "n_fig": 3, "n_tab": 18, "L_tab": 21878, "n_element_tab": 2980, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1046, "n_element_tab_1": 89, "formula_len_all": 6364, "formula_len_all_1": 2147, "len_all": 330869, "len_all_1": 75236, "len_abs": 1741, "len_title": 107, "len_sents": 83129, "len_sents_1": 34469, "n_sents": 855, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1744, "title": "Training Neural Networks as Recognizers of Formal Languages", "abs": "Characterizing the computational power of neural network architectures in terms of formal language theory remains a crucial line of research, as it describes lower and upper bounds on the reasoning capabilities of modern AI. However, when empirically testing these bounds, existing work often leaves a discrepancy between experiments and the formal claims they are meant to support. The problem is that formal language theory pertains specifically to recognizers: machines that receive a string as input and classify whether it belongs to a language. On the other hand, it is common to instead use proxy tasks that are similar in only an informal sense, such as language modeling or sequence-to-sequence transduction. We correct this mismatch by training and evaluating neural networks directly as binary classifiers of strings, using a general method that can be applied to a wide variety of languages. As part of this, we extend an algorithm recently proposed by Anonymous (2024) to do length-controlled sampling of strings from regular languages, with much better asymptotic time complexity than previous methods. We provide results on a variety of languages across the Chomsky hierarchy for three neural architectures: a simple RNN, an LSTM, and a causally-masked transformer. We find that the RNN and LSTM often outperform the transformer, and that auxiliary training objectives such as language modeling can help, although no single objective uniformly improves performance across languages and architectures. Our contributions will facilitate theoretically sound empirical testing of language recognition claims in future work. We have released our datasets as a benchmark called FLaRe (Formal Language Recognition), along with our code.", "keywords": [ "neural network", "formal language theory", "transformer", "rnn", "lstm" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "aW7XcFocYr", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Chenyi yang", "gender": "Female", "institution": "Xiamen University", "country": "", "position": "PhD student" }, { "name": "GUANNAN JIANG", "gender": "Male", "institution": "Contemporary Amperex Technology Co., Limited", "country": "CN", "position": "Researcher" }, { "name": "Jie Hu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Wenjie Nie", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Xiawu Zheng", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuhang Wu", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Yuxin Zhang", "gender": "unknown", "institution": "Xiamen University", "country": "CN", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 27, "n_ref": 92, "n_ref_all": 103, "n_fig": 2, "n_tab": 5, "L_tab": 889, "n_element_tab": 173, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 889, "n_element_tab_1": 173, "formula_len_all": 914, "formula_len_all_1": 914, "len_all": 93624, "len_all_1": 58299, "len_abs": 1264, "len_title": 115, "len_sents": 28099, "len_sents_1": 27197, "n_sents": 182, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1277, "title": "BAME: Block-Aware Mask Evolution for Efficient N:M Sparse Training", "abs": "N:M sparsity stands as a progressively important tool for DNN compression, achieving practical speedups by stipulating at most N non-zero components within M sequential weights. Unfortunately, most existing works identify the N:M sparse mask through dense backward propagation to update all weights, which incurs exorbitant training costs. In this paper, we introduce BAME, a method that maintains consistent sparsity throughout the N:M sparse training process. BAME perpetually keeps both sparse forward and backward propagation, while iteratively performing weight pruning-and-regrowing within designated weight blocks to tailor the N:M mask. These blocks are selected through a joint assessment based on accumulated mask oscillation frequency and expected loss reduction of mask adaptation, thereby ensuring stable and efficient identification of the optimal N:M mask. Our empirical results substantiate the effectiveness of BAME, illustrating it performs comparably to or better than previous works that fully maintaining dense backward propagation during training. For instance, BAME attains a 72.0% top-1 accuracy while training a 1:16 sparse ResNet-50 on ImageNet, eclipsing SR-STE by 0.5%, despite achieving 2.37$\\times$ training FLOPs reduction. Code will be released.", "keywords": [ "Network Sparsity; Sparse Training; N:M Sparsity" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "aVyJwS1fqQ", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Chilam Cheang", "gender": "Female", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Fangqi Zhu", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Hongtao Wu", "gender": "unknown", "institution": "Bytedance Research", "country": "", "position": "Researcher" }, { "name": "Song Guo", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Tao Kong", "gender": "Male", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "Yuxiao Liu", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 4, "n_ref_uni": 49, "n_ref": 100, "n_ref_all": 125, "n_fig": 14, "n_tab": 8, "L_tab": 1503, "n_element_tab": 170, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2627, "n_element_tab_1": 119, "formula_len_all": 614, "formula_len_all_1": 132, "len_all": 188531, "len_all_1": 55039, "len_abs": 4564, "len_title": 111, "len_sents": 48553, "len_sents_1": 21854, "n_sents": 439, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1501, "title": "Mani-WM: An Interactive World Model for Real-Robot Manipulation", "abs": "Scalable robot learning in the real world is limited by the cost and safety issues of real robots. In addition, rolling out robot trajectories in the real world can be time-consuming and labor-intensive. In this paper, we propose to learn an interactive world model for robot manipulation as an alternative. We present a novel method, Mani-WM, which leverages the power of generative models to generate realistic videos of a robot arm executing a given action trajectory, starting from an initial given frame. Mani-WM employs a novel frame-level conditioning technique to ensure precise alignment between actions and video frames and leverages a diffusion transformer for high-quality video generation. To validate the effectiveness of Mani-WM, we perform extensive experiments on four challenging real-robot datasets. Results show that Mani-WM outperforms all the comparing baseline methods and is more preferable in human evaluations. We further showcase the flexible action controllability of Mani-WM by controlling the virtual robots in datasets with trajectories 1) predicted by an autonomous policy and 2) collected by a keyboard or VR controller. Finally, we combine Mani-WM with model-based planning to showcase its usefulness on real-robot manipulation tasks. We hope that Mani-WM can serve as an effective and scalable approach to enhance robot learning in the real world. To promote research on manipulation world models, we opensource the code at https://anonymous.4open.science/r/Mani-WM.", "keywords": [ "World Model", "Video Generation", "Robot Manipulation" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 1 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "aVovUyrh5J", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Bingzhe Wu", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Chengzhi Piao", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Researcher" }, { "name": "Hong Cheng", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Professor" }, { "name": "Jia Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Long-Kai Huang", "gender": "unknown", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Tian Bian", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Tingyang Xu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yifan Niu", "gender": "Non-Binary", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 21, "n_ref_uni": 27, "n_ref": 51, "n_ref_all": 58, "n_fig": 4, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1043, "formula_len_all_1": 899, "len_all": 95986, "len_all_1": 65447, "len_abs": 1185, "len_title": 122, "len_sents": 33068, "len_sents_1": 31243, "n_sents": 250, "n_sents_1": 234, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1199, "title": "IBCircuit: Towards Holistic Circuit Discovery with Information Bottleneck", "abs": "Circuit discovery has recently attracted attention as a potential research direction to explain the nontrivial behaviors of language model (LM). It aims to find the computational subgraph, also known as \\emph{circuit}, that explains LM's behavior on specific tasks. Most studies determine the circuit for a task by performing causal interventions independently on each component. However, they ignored the holistic nature of the circuit, which is an interconnected system of components rather than an independent combination. Additionally, existing methods require redesigning a unique corrupted activation for each task, which are complicated and inefficient. In this work, we propose a novel circuit discovery approach based on the principle of Information Bottleneck, called IBCircuit, to identify the most informative circuit from a holistic perspective. Furthermore, IBcircuit can be applied to any given task without corrupted activation construction. Our experiments demonstrate the ability of IBCircuit to identify the most informative circuit in the model. The results from IBCircuit suggest that the earlier layers in Transformer-based models are crucial in capturing factual information.", "keywords": [ "Information Bottleneck", "Circuit Analysis" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "aVfDrl7xDV", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 6, "author_info_list": [ { "name": "Dhruv Agarwal", "gender": "Male", "institution": "University of Massachusetts Amherst", "country": "US", "position": "PhD student" }, { "name": "Manoj Ghuhan Arivazhagan", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Rajarshi Das", "gender": "unknown", "institution": "AWS AI Labs", "country": "", "position": "Researcher" }, { "name": "Rashmi Gangadharaiah", "gender": "Female", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Sandesh Swamy", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Sopan Khosla", "gender": "Male", "institution": "Amazon Web Services", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 48, "n_ref": 74, "n_ref_all": 102, "n_fig": 12, "n_tab": 2, "L_tab": 346, "n_element_tab": 83, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 188, "n_element_tab_1": 15, "formula_len_all": 166, "formula_len_all_1": 120, "len_all": 154062, "len_all_1": 60084, "len_abs": 1835, "len_title": 115, "len_sents": 49566, "len_sents_1": 29221, "n_sents": 356, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1793, "title": "Searching for Optimal Solutions with LLMs via Bayesian Optimization", "abs": "Scaling test-time compute to search for optimal solutions is an important step towards building generally-capable language models that can reason. Recent work, however, shows that tasks of varying complexity require distinct search strategies to solve optimally, thus making it challenging to design a one-size-fits-all approach. Prior solutions either attempt to predict task difficulty to select the optimal search strategy, often infeasible in practice, or use a static, pre-defined strategy, e.g., repeated parallel sampling or greedy sequential search, which is sub-optimal. In this work, we argue for an alternative view using the probabilistic framework of Bayesian optimization (BO), where the search strategy is adapted dynamically based on the evolving uncertainty estimates of solutions as search progresses. To this end, we introduce Bayesian-OPRO (BOPRO)\u2014a generalization of a recent method for in-context optimization, which iteratively samples from new proposal distributions by modifying the prompt to the LLM with a subset of its previous generations selected to explore or exploit different parts of the search space. We evaluate our method on word search, molecule optimization, and a joint hypothesis+program search task using a 1-D version of the challenging Abstraction and Reasoning Corpus (1D-ARC). Our results show that BOPRO outperforms all baselines in word search (\u226510 points) and molecule optimization (higher quality and 17% fewer invalid molecules), but trails a best-k prompting strategy on program search. Our analysis of this failure case reveals that despite the ability to sufficiently balance exploration and exploitation using BOPRO, failure is likely due to the inability of code representation models in distinguishing sequences with low edit-distances.", "keywords": [ "search", "optimization", "LLMs", "test-time compute", "bayesian optimization", "reasoning" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "aUeQPyRMeJ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Chenxi Liu", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "PhD student" }, { "name": "Hui Xiong", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Junyi Li", "gender": "Male", "institution": "Hong Kong University of Science and Technology ", "country": "CN", "position": "MS student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Qianyi Cai", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Yongqiang Chen", "gender": "unknown", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 101, "n_ref_all": 118, "n_fig": 9, "n_tab": 5, "L_tab": 5299, "n_element_tab": 102, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1127, "n_element_tab_1": 48, "formula_len_all": 146, "formula_len_all_1": 199, "len_all": 148653, "len_all_1": 56621, "len_abs": 1206, "len_title": 121, "len_sents": 41552, "len_sents_1": 25994, "n_sents": 284, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1211, "title": "Can Large Language Models Help Experimental Design for Causal Discovery?", "abs": "Designing proper experiments and intervening targets is a longstanding problem in scientific or causal discovery. It is fundamentally impossible to identify the underlying causal structure merely based on the observational data. Obtaining interventional data, on the other hand, is crucial to causal discovery, yet it is usually expensive or time-consuming to obtain sufficient interventional data to facilitate causal discovery. Previous approaches usually leverage uncertainty or gradient signals to determine the intervention targets, and may suffer from the suboptimality. In this work, we investigate a different approach, whether we can leverage Large Language Models (LLMs) to assist with the intervention targeting in causal discovery by making use of the rich world knowledge about the experimental design in LLM. Specifically, we present Large Language Model Guided Intervention Targeting (LeGIT), a robust framework that effectively incorporates LLMs to assist with the intervention targeting in causal discovery. Surprisingly, across 4 different scales of realistic benchmarks, LeGIT significantly outperforms previous approaches. LeGIT opens up a new frontier for using LLMs in experimental design.", "keywords": [ "Large Language Model", "Experimental Design", "Causal Discovery" ], "rating_list": [ 5, 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 2, 3, 4 ] }, { "paper_id": "aUZEeb2yvK", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Ayush Jain", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Grace Zhang", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Injune Hwang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Joseph J Lim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Shao-Hua Sun", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "Assistant Professor" } ], "n_formula": 30, "n_formula_1": 7, "n_ref_uni": 38, "n_ref": 71, "n_ref_all": 122, "n_fig": 32, "n_tab": 3, "L_tab": 944, "n_element_tab": 97, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 310, "n_element_tab_1": 11, "formula_len_all": 3854, "formula_len_all_1": 761, "len_all": 217013, "len_all_1": 67444, "len_abs": 1693, "len_title": 113, "len_sents": 69246, "len_sents_1": 30919, "n_sents": 529, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1319, "title": "QMP: Q-switch Mixture of Policies for Multi-Task Behavior Sharing", "abs": "Multi-task reinforcement learning (MTRL) aims to learn several tasks simultaneously for better sample efficiency than learning them separately. Traditional methods achieve this by sharing parameters or relabeling data between tasks. In this work, we introduce a new framework for sharing behavioral policies across tasks, which can be used in addition to existing MTRL methods. The key idea is to improve each task's off-policy data collection by employing behaviors from other task policies. Selectively sharing helpful behaviors acquired in one task to collect training data for another task can lead to higher-quality trajectories, leading to more sample-efficient MTRL. Thus, we introduce a simple and principled framework called Q-switch mixture of policies (QMP) that selectively shares behavior between different task policies by using the task's Q-function to evaluate and select useful shareable behaviors. We theoretically analyze how QMP improves the sample efficiency of the underlying RL algorithm. Our experiments show that QMP's behavioral policy sharing provides complementary gains over many popular MTRL algorithms and outperforms alternative ways to share behaviors in various manipulation, locomotion, and navigation environments. Videos are available at https://sites.google.com/view/qmp-mtrl.", "keywords": [ "Multi-task Reinforcement Learning", "Behavior Sharing" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "aUH0XrFhiX", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Guansong Pang", "gender": "unknown", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Jin Zheng", "gender": "Female", "institution": "BeiHang University", "country": "CN", "position": "Associate Professor" }, { "name": "Ruohuan Fang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Trong-Tung Nguyen", "gender": "Male", "institution": "VinAI Research", "country": "VN", "position": "Researcher" }, { "name": "Xiao Bai", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "\u6587\u4fca \u7f2a", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 37, "n_ref": 122, "n_ref_all": 153, "n_fig": 11, "n_tab": 8, "L_tab": 17361, "n_element_tab": 950, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 3834, "n_element_tab_1": 177, "formula_len_all": 476, "formula_len_all_1": 496, "len_all": 188376, "len_all_1": 75653, "len_abs": 1563, "len_title": 108, "len_sents": 56394, "len_sents_1": 35300, "n_sents": 294, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1575, "title": "Out-of-Distribution Detection in Class Incremental Learning", "abs": "Class incremental learning (CIL) aims to learn a model that can not only incrementally accommodate new classes, but also maintain the learned knowledge of old classes. Out-of-distribution (OOD) detection in CIL is to retain this incremental learning ability, while being able to reject unknown samples that are drawn from different distributions of the learned classes. This capability is crucial to the safety of deploying CIL models in open worlds.However, despite remarkable advancements in the respective CIL and OOD detection, there lacks a systematic and large-scale benchmark to assess the capability of advanced CIL models in detecting OOD samples. To fill this gap, in this study we design a comprehensive empirical study to establish such a benchmark, named **OpenCIL**, offering a unified protocol for enabling CIL models with different OOD detectors using two principled OOD detection frameworks. One key observation we find through our comprehensive evaluation is that the CIL models can be severely biased towards the OOD samples and newly added classes when they are exposed to open environments. Motivated by this, we further propose a novel approach for OOD detection in CIL, namely Bi-directional Energy Regularization (**BER**), which is specially designed to mitigate these two biases in different CIL models by having energy regularization on both old and new classes. Extensive experiments show that BER can substantially improve the OOD detection capability across a range of CIL models, achieving state-of-the-art performance on the OpenCIL benchmark.", "keywords": [ "Out-of-Distribution Detection", "Class Incremental Learning" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "aU63Ib07KJ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 15, "author_info_list": [ { "name": "Afshin Rostamizadeh", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Ankit Singh Rawat", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Ayan Chakrabarti", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Hrayr Harutyunyan", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Nikunj Saunshi", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Rakesh Shivanna", "gender": "Male", "institution": "Indian Institute of Science Bangalore.", "country": "", "position": "MS student" }, { "name": "Rohan Anil", "gender": "Male", "institution": "Google Brain ", "country": "", "position": "Principal Engineer" }, { "name": "Sanjiv Kumar", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Sashank J. Reddi", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Seungyeon Kim", "gender": "unknown", "institution": "Google DeepMind", "country": "US", "position": "Research Engineer (L7)" }, { "name": "Veeranjaneyulu Sadhanala", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Vladimir Feinberg", "gender": "Not Specified", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Wittawat Jitkrittum", "gender": "Male", "institution": "Google Research", "country": "US", "position": "Research Scientist" }, { "name": "Zachary Nado", "gender": "Male", "institution": "Google", "country": "", "position": "Research Engineer" }, { "name": "Aditya Krishna Menon", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" } ], "n_formula": 80, "n_formula_1": 16, "n_ref_uni": 84, "n_ref": 141, "n_ref_all": 172, "n_fig": 4, "n_tab": 14, "L_tab": 6100, "n_element_tab": 490, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1278, "n_element_tab_1": 119, "formula_len_all": 6853, "formula_len_all_1": 1312, "len_all": 306622, "len_all_1": 71745, "len_abs": 1504, "len_title": 125, "len_sents": 69060, "len_sents_1": 33347, "n_sents": 555, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1513, "title": "A Little Help Goes a Long Way: Efficient LLM Training by Leveraging Small LMs", "abs": "A primary challenge in large language model (LLM) development is their onerous pre-training cost. Typically, such pre-training involves optimizing a self-supervised objective (such as next-token prediction) over a large corpus. This paper explores a promising paradigm to improve LLM pre-training efficiency and quality by suitably leveraging a small language model (SLM). In particular, this paradigm relies on an SLM to both (1) provide soft labels as additional training supervision, and (2) select a small subset of valuable (\"informative'' and \"hard'') training examples. Put together, this enables an effective transfer of the SLM's predictive distribution to the LLM, while prioritizing specific regions of the training data distribution. Empirically, this leads to reduced LLM training time compared to standard training, while improving the overall quality. Theoretically, we develop a statistical framework to systematically study the utility of SLMs in enabling efficient training of high-quality LLMs. In particular, our framework characterizes how the SLM's seemingly low-quality supervision can enhance the training of a much more capable LLM. Furthermore, it also highlights the need for an adaptive utilization of such supervision, by striking a balance between the bias and variance introduced by the SLM-provided soft labels. We corroborate our theoretical framework by improving the pre-training of an LLM with 2.8B parameters by utilizing a smaller LM with 1.5B parameters on the Pile dataset.", "keywords": [ "Large language models", "knowledge distillation", "data selection", "efficiency" ], "rating_list": [ 8, 5, 6, 3 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 4, 3, 4, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "aU2cjz87Bm", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Mine Melodi Caliskan", "gender": "unknown", "institution": "Ruhr-Universit\u00e4t Bochum", "country": "DE", "position": "Researcher" }, { "name": "Saeed Ghoorchian", "gender": "Male", "institution": "SAP SE", "country": "DE", "position": "Researcher" }, { "name": "Setareh Maghsudi", "gender": "unknown", "institution": "Ruhr-Universt\u00e4t Bochum", "country": "", "position": "Full Professor" } ], "n_formula": 112, "n_formula_1": 26, "n_ref_uni": 10, "n_ref": 22, "n_ref_all": 48, "n_fig": 6, "n_tab": 5, "L_tab": 1478, "n_element_tab": 159, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6134, "formula_len_all_1": 1753, "len_all": 128671, "len_all_1": 63186, "len_abs": 1611, "len_title": 123, "len_sents": 43648, "len_sents_1": 27639, "n_sents": 376, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 934, "title": "Robust Inverse Reinforcement Learning under State Adversarial Perturbations", "abs": "State adversarial perturbations \u2013such as sensor noise, environmental interference, or targeted attacks\u2013 are common in real-world systems, often leading to compromised state observations. Despite this, Inverse Reinforcement Learning (IRL) in the context of State-Adversarial Markov Decision Processes (SA-MDPs) has received limited attention, primarily because conventional notions of optimality do not apply. In this paper, we introduce a novel definition of optimality that ensures the existence of an optimal policy within SA-MDPs. Building on this foundation, we propose the State-Adversarial Max-Margin IRL (SAMM-IRL) algorithm, designed for robustness against state adversarial perturbations. Our theoretical analysis, supported by empirical validation, demonstrates that SAMM-IRL significantly enhances IRL performance in adversarial environments, providing a robust framework for real-world applications that demand resilience.", "keywords": [ "inverse reinforcement learning", "state adversarial attacks", "robustness" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "aTYexOYlLb", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Brian Formento", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Chuan-Sheng Foo", "gender": "Male", "institution": "Centre for Frontier AI Research, A*STAR", "country": "SG", "position": "Investigator" }, { "name": "See-Kiong Ng", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 45, "n_ref": 97, "n_ref_all": 130, "n_fig": 5, "n_tab": 10, "L_tab": 3236, "n_element_tab": 200, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 2345, "n_element_tab_1": 188, "formula_len_all": 295, "formula_len_all_1": 186, "len_all": 166791, "len_all_1": 59981, "len_abs": 1496, "len_title": 117, "len_sents": 42471, "len_sents_1": 26644, "n_sents": 300, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1521, "title": "Confidence Elicitation: A New Attack Vector for Large Language Models", "abs": "A fundamental issue in deep learning has been adversarial robustness. As these systems have scaled, such issues have persisted. Currently, large language models (LLMs) with billions of parameters suffer from adversarial attacks just like their earlier, smaller counterparts. However, the threat models have changed. Previously, having gray-box access, where input embeddings or output logits/probabilities were visible to the user, might have been reasonable. However, with the introduction of closed-source models, no information about the model is available apart from the generated output. This means that current black-box attacks can only utilize the final prediction to detect if an attack is successful. In this work, we investigate and demonstrate the potential of attack guidance, akin to using output probabilities, while having only black-box access in a classification setting. This is achieved through the ability to elicit confidence from the model. We empirically show that the elicited confidence is calibrated and not hallucinated for current LLMs. By minimizing the elicited confidence, we can therefore increase the likelihood of misclassification. Our new proposed paradigm demonstrates promising state-of-the-art results on three datasets across two models (LLaMA-3-8B-Instruct and Mistral-7B-Instruct-V0.3) when comparing our technique to existing hard-label black-box attack methods that introduce word-level substitutions. The code is publicly available at GitHub: Confidence_Elicitation_Attacks.", "keywords": [ "adversarial attack", "adversarial robustness", "confidence elicitation." ], "rating_list": [ 8, 8, 5, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "aTBE70xiFw", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Joakim And\u00e9n", "gender": "Male", "institution": "KTH Royal Institute of Technology, Stockholm, Sweden", "country": "SE", "position": "Associate Professor" }, { "name": "Justus Sagem\u00fcller", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "Postdoc" } ], "n_formula": 41, "n_formula_1": 19, "n_ref_uni": 53, "n_ref": 103, "n_ref_all": 113, "n_fig": 7, "n_tab": 1, "L_tab": 246, "n_element_tab": 71, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2561, "formula_len_all_1": 1075, "len_all": 164671, "len_all_1": 69743, "len_abs": 959, "len_title": 117, "len_sents": 42774, "len_sents_1": 33008, "n_sents": 351, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 972, "title": "Joint Denoising of Cryo-EM Projection Images using Polar Transformers", "abs": "Deep neural networks (DNNs) have proven powerful for denoising individual images, but there is a limit to the noise level they can handle.\nIn applications like cryogenic electron microscopy (cryo-EM), the noise level is extremely high but datasets contain hundreds of thousands of projections of the same molecule, each taken a different viewing direction.\nThis redundancy of information is useful in traditional denoising techniques known as class averaging methods, where images are clustered, aligned, and then averaged to reduce the noise level.\nWe present a neural network architecture based on polar representation of images and transformers that simultaneously clusters, aligns, and denoises cryo-EM projection images.\nResults on synthetic data show accurate denoising performance using this architecture, with a relative mean squared error of $0.06$ at signal-to-noise (SNR) level of $0.05$, outperforming traditional filter-based methods by a factor of $2\\times$.", "keywords": [ "denoising", "microscopy", "tomography", "attention", "symmetries" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "aSy2nYwiZ2", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Qiannan Zhang", "gender": "Female", "institution": "Weill Cornell Medicine, Cornell University", "country": "", "position": "Postdoc" }, { "name": "Shichao Pei", "gender": "unknown", "institution": "University of Massachusetts Boston", "country": "US", "position": "Assistant Professor" }, { "name": "Zhuowei Chen", "gender": "Male", "institution": "Guangdong University of Foreign Studies", "country": "CN", "position": "Undergrad student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 30, "n_ref": 65, "n_ref_all": 97, "n_fig": 9, "n_tab": 7, "L_tab": 5031, "n_element_tab": 363, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 2871, "n_element_tab_1": 443, "formula_len_all": 286, "formula_len_all_1": 236, "len_all": 129429, "len_all_1": 67035, "len_abs": 1190, "len_title": 108, "len_sents": 39078, "len_sents_1": 27415, "n_sents": 297, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1127, "title": "Injecting Universal Jailbreak Backdoors into LLMs in Minutes", "abs": "Jailbreak backdoor attacks on LLMs have garnered attention for their effectiveness and stealth. However, existing methods rely on the crafting of poisoned datasets and the time-consuming process of fine-tuning. In this work, we propose JailbreakEdit, a novel jailbreak backdoor injection method that exploits model editing techniques to inject a universal jailbreak backdoor into safety-aligned LLMs with minimal intervention *in minutes*. JailbreakEdit integrates a multi-node target estimation to estimate the jailbreak space, thus creating shortcuts from the backdoor to this estimated jailbreak space that induce jailbreak actions. Our attack effectively shifts the models' attention by attaching strong semantics to the backdoor, enabling it to bypass internal safety mechanisms. Experimental results show that JailbreakEdit achieves a high jailbreak success rate on jailbreak prompts while preserving generation quality, and safe performance on normal queries. Our findings underscore the effectiveness, stealthiness, and explainability of JailbreakEdit, emphasizing the need for more advanced defense mechanisms in LLMs.", "keywords": [ "Large language model", "Jailbreak", "Backdoor", "Attack", "Safety", "Model Editing" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "aSoLl0nlzr", "primary_area": "reinforcement learning", "n_author": 1, "author_info_list": [ { "name": "Shishen Lin", "gender": "Male", "institution": "University of Birmingham", "country": "GB", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 31, "n_ref": 76, "n_ref_all": 94, "n_fig": 8, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1393, "formula_len_all_1": 1274, "len_all": 105490, "len_all_1": 62856, "len_abs": 2181, "len_title": 121, "len_sents": 29938, "len_sents_1": 26674, "n_sents": 225, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 2196, "title": "Competitive Co-Evolutionary Learning on Matrix Games with Bandit Feedback", "abs": "Learning in games is a fundamental problem in machine learning and artificial intelligence, with many successful applications (Silver et al., 2016; Schrittwieser et al., 2020). We consider the problem of learning in matrix games, where two players engage in a two-player zero-sum game with an unknown payoff matrix and bandit feedback. In this setting, players can observe their actions and the corresponding (noisy) payoffs at each round. This problem has been studied in the literature, and several algorithms have been proposed to address it (O\u2019Donoghue et al., 2021; Maiti et al., 2023; Cai et al., 2023). In particular, O\u2019Donoghue et al. (2021) demonstrated that deterministic optimism (e.g., the UCB algorithm for matrix games) plays a central role in achieving sublinear regret and outperforms other algorithms. However, despite numerous applications, the theoretical understanding of learning in matrix games remains underexplored. Specifically, it remains an open question whether randomised optimism can also exhibit sublinear regret. \n\nIn this paper, we propose a novel algorithm called Competitive Co-evolutionary Bandit Learning (CoEBL) for unknown two-player zero-sum matrix games. By integrating evolutionary algorithms (EAs) into the bandit framework, CoEBL introduces randomised optimism through the variation operator of EAs. We prove that CoEBL also enjoys sublinear regret, matching the regret performance of algorithms based on deterministic optimism (O\u2019Donoghue et al., 2021). To the best of our knowledge, this is the first work that provides a regret analysis of an evolutionary bandit learning algorithm in matrix games. Empirically, we compare CoEBL with classical bandit algorithms, including EXP3 (Auer et al., 2002), the variant of EXP3-IX (Cai et al., 2023), and UCB algorithms analysed in O\u2019Donoghue et al. (2021) across several matrix game benchmarks. Our results show that CoEBL not only enjoys sublinear regret, but also outperforms existing methods in various scenarios. These findings reveal the promising potential of evolutionary bandit learning in game-theoretic settings, in particular, the effectiveness of randomised optimism via evolutionary algorithms.", "keywords": [ "Matrix Games", "Bandit Learning", "Evolutionary Algorithms", "Regret Analysis" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "aSByBbmASe", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Faezeh S Zakeri", "gender": "Female", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Hendrik Lensch", "gender": "Male", "institution": "University of T\u00fcbingen", "country": "DE", "position": "Professor" }, { "name": "Lukas Ruppert", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Raphael Braun", "gender": "unknown", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 39, "n_ref": 138, "n_ref_all": 176, "n_fig": 10, "n_tab": 7, "L_tab": 1824, "n_element_tab": 111, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 951, "n_element_tab_1": 72, "formula_len_all": 60, "formula_len_all_1": 60, "len_all": 174464, "len_all_1": 64599, "len_abs": 4938, "len_title": 111, "len_sents": 49115, "len_sents_1": 30137, "n_sents": 399, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1149, "title": "POC-SLT: Partial Object Completion with SDF Latent Transformers", "abs": "3D geometric shape completion hinges on representation learning and a deep understanding of geometric data.\nWithout profound insights into the three-dimensional nature of the data, this task remains unattainable.\nOur work addresses this challenge of 3D shape completion given partial observations\nby proposing a transformer operating on the latent space representing Signed Distance Fields (SDFs).\nInstead of a monolithic volume, the SDF of an object is partitioned into smaller high-resolution patches leading to a sequence of latent codes.\nThe approach relies on a smooth latent space encoding learned via a variational autoencoder (VAE), trained on millions of 3D patches.\nWe employ an efficient masked autoencoder transformer to complete partial sequences into comprehensive shapes in latent space.\nOur approach is extensively evaluated on partial observations from ShapeNet and the ABC dataset where only fractions of the objects are given.\nThe proposed POC-SLT architecture compares favorably with several baseline state-of-the-art methods,\ndemonstrating a significant improvement in 3D shape completion, both qualitatively and quantitatively.", "keywords": [ "SDF Completion", "Masked Encoder Transformer", "Variation Auto Encoder", "Scale Invariant", "Million-scale Point Processing", "Repairing Latent Codes" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 3, 3, 5, 5 ] }, { "paper_id": "aS1IhKdLPP", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Guangyi Chen", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "US", "position": "Postdoc" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Loka Li", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Peter Spirtes", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Xiangchen Song", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Yifan Shen", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Yunlong Deng", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "MS student" }, { "name": "Zeyu Tang", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zhenhao Chen", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 14, "n_ref_uni": 51, "n_ref": 66, "n_ref_all": 97, "n_fig": 10, "n_tab": 4, "L_tab": 1106, "n_element_tab": 136, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1771, "n_element_tab_1": 138, "formula_len_all": 11049, "formula_len_all_1": 10553, "len_all": 179415, "len_all_1": 78850, "len_abs": 1157, "len_title": 108, "len_sents": 49756, "len_sents_1": 30494, "n_sents": 333, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1156, "title": "Reflection Window: Text Generation with Selective Refinement", "abs": "The autoregressive approach to text generation in large language models (LLMs), while widely used, is inherently suboptimal due to the lack of a built-in mechanism to perform refinement and/or correction of the generated content. In this paper, we consider optimality in terms of the joint probability over the generated response, when jointly considering all tokens at the same time. We theoretically characterize the potential deviation of the autoregressively generated response from its globally optimal counterpart that is of the same length. Our analysis suggests that we need to be cautious when noticeable uncertainty arises during text generation, which may signal the sub-optimality of the generation history. To address the pitfall of autoregressive text generation, we propose an approach that incorporates a sliding reflection window and a pausing criterion, such that reflection and generation can be carried out interchangeably as the generation proceeds. Our approach utilizes a selective refinement mechanism to strike the balance between efficiency and optimality, and the experimental results demonstrate the effectiveness of our method.", "keywords": [ "Selective Refinement", "Autogressive Text Generation", "Reflection Window", "Large Language Model" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "aRxLDcxFcL", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Chuang Gan", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Hongxin Zhang", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" }, { "name": "Jiageng Liu", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "MS student" }, { "name": "Jincheng Yang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Lixing Fang", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" }, { "name": "Qinhong Zhou", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" }, { "name": "Xiangye Lin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Xinyu Sun", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "MS student" }, { "name": "Yutian Chen", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Undergrad student" }, { "name": "Zheyuan Zhang", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 51, "n_ref": 70, "n_ref_all": 84, "n_fig": 6, "n_tab": 2, "L_tab": 198, "n_element_tab": 41, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 198, "n_element_tab_1": 41, "formula_len_all": 318, "formula_len_all_1": 318, "len_all": 137658, "len_all_1": 58377, "len_abs": 1666, "len_title": 108, "len_sents": 30004, "len_sents_1": 27388, "n_sents": 198, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1554, "title": "Virtual Community: A Generative Social World for Embodied AI", "abs": "We present Virtual Community, a social world simulation platform designed to support embodied AI research, featuring large-scale community scenarios derived from the real world. Virtual Community introduces two key features to enrich the virtual social world with generative AI: scalable 3D Scene creation, which supports the generation of expansive outdoor and indoor environments at any location and scale, addressing the lack of a large-scale, interactive, open-world scene for embodied AI research; and embodied agents with grounded characters and social relationship networks, the first to simulate socially connected agents at a community level, that also have scene-grounded characters. We design two novel challenges to showcase that Virtual Community provides testbeds to evaluate the social reasoning and planning capabilities of embodied agents in open-world scenarios: Route Planning and Election Campaign. The Route Planning task examines the agent's ability to reason about time, location, and tools in the community to plan fast and economical commutes in daily life. The Election Campaign task evaluates an agent's ability to explore and connect with other agents as a new member of the community. We evaluate several baseline agents on these challenges and demonstrate the performance gap of current methods in addressing embodied social challenges within open-world scenarios, which our simulator is designed to unlock. We plan to open-source this simulation and hope Virtual Community can accelerate the development in this direction.", "keywords": [ "embodied AI" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "aRqyX0DsmW", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Jingdong Yang", "gender": "unknown", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Kehan Guo", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Nuno Moniz", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "Associate Professor" }, { "name": "Pin-Yu Chen", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Tian Gao", "gender": "unknown", "institution": "International Business Machines", "country": "US", "position": "Reseach Staff Member" }, { "name": "Werner Geyer", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Principal Researcher" }, { "name": "Xiangliang Zhang", "gender": "Female", "institution": "University of Notre Dame", "country": "US", "position": "Associate Professor" }, { "name": "Yujun Zhou", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Nitesh Chawla", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 56, "n_ref": 111, "n_ref_all": 134, "n_fig": 16, "n_tab": 9, "L_tab": 8188, "n_element_tab": 274, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 2134, "n_element_tab_1": 117, "formula_len_all": 438, "formula_len_all_1": 0, "len_all": 366861, "len_all_1": 63278, "len_abs": 1665, "len_title": 101, "len_sents": 135203, "len_sents_1": 29993, "n_sents": 919, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1622, "title": "Benchmarking LLMs on Safety Issues in Scientific Labs", "abs": "Laboratory accidents pose significant risks to human life and property, underscoring the importance of robust safety protocols. Despite advancements in safety training, laboratory personnel may still unknowingly engage in unsafe practices. With the increasing reliance on large language models (LLMs) for guidance in various fields, including laboratory settings, there is a growing concern about their reliability in critical safety-related decision-making. Unlike trained human researchers, LLMs lack formal lab safety education, raising questions about their ability to provide safe and accurate guidance. Existing research on LLM trustworthiness primarily focuses on issues such as ethical compliance, truthfulness, and fairness but fails to fully cover safety-critical real-world applications, like lab safety. To address this gap, we propose the Laboratory Safety Benchmark (LabSafety Bench), a comprehensive evaluation framework based on a new taxonomy aligned with Occupational Safety and Health Administration (OSHA) protocols. This benchmark includes 765 multiple-choice questions verified by human experts, assessing LLMs and large vision models (LVMs) performance in lab safety contexts. Our evaluations demonstrate that while GPT-4o outperforms human participants, it is still prone to critical errors, highlighting the risks of relying on LLMs in safety-critical environments. Our findings emphasize the need for specialized benchmarks to accurately assess the trustworthiness of LLMs in real-world safety applications. The code and data are available at https://anonymous.4open.science/r/LabSafetyBench-6363", "keywords": [ "benchmark", "lab safety", "LLM Trustworthiness" ], "rating_list": [ 3, 5, 3, 3, 6 ], "soundness_list": [ 1, 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "aRD1NqcXTC", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Dawit Mureja Argaw", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Fitsum Reda", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Joon Son Chung", "gender": "unknown", "institution": "KAIST", "country": "KR", "position": "Associate Professor" }, { "name": "Ming-Yu Liu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Qinsheng Zhang", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Xian Liu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 11, "n_ref_uni": 44, "n_ref": 166, "n_ref_all": 217, "n_fig": 10, "n_tab": 10, "L_tab": 4877, "n_element_tab": 668, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 3804, "n_element_tab_1": 415, "formula_len_all": 650, "formula_len_all_1": 1073, "len_all": 184820, "len_all_1": 73380, "len_abs": 1183, "len_title": 112, "len_sents": 49577, "len_sents_1": 32543, "n_sents": 312, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1191, "title": "High-Quality Joint Image and Video Tokenization with Causal VAE", "abs": "Generative modeling has seen significant advancements in image and video synthesis. However, the curse of dimensionality remains a significant obstacle, especially for video generation, given its inherently complex and high-dimensional nature. Many existing works rely on low-dimensional latent spaces from pretrained image autoencoders. However, this approach overlooks temporal redundancy in videos and often leads to temporally incoherent decoding. To address this issue, we propose a video compression network that reduces the dimensionality of visual data both spatially and temporally. Our model, based on a variational autoencoder, employs causal 3D convolution to handle images and videos jointly. The key contributions of our work include a scale-agnostic encoder for preserving video fidelity, a novel spatio-temporal down/upsampling block for robust long-sequence modeling, and a flow regularization loss for accurate motion decoding. \nOur approach outperforms competitors in video quality and compression rates across various datasets. Experimental analyses also highlight its potential as a robust autoencoder for video generation training. Code and models will be open-sourced.", "keywords": [ "Autoencoding", "Generative Modelling", "Causal Video VAE", "FILM" ], "rating_list": [ 6, 6, 5, 8, 8 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 4, 3 ], "confidence_list": [ 3, 4, 4, 4, 5 ] }, { "paper_id": "aQj9Ifxrl6", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Bryan Kian Hsiang Low", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Associate Professor" }, { "name": "Jingtan Wang", "gender": "Female", "institution": "National University of Singapore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Pang Wei Koh", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Visiting Research Scientist" }, { "name": "Rui Qiao", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zhaoxuan Wu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 20, "n_ref_uni": 56, "n_ref": 113, "n_ref_all": 146, "n_fig": 11, "n_tab": 5, "L_tab": 985, "n_element_tab": 135, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3271, "formula_len_all_1": 1297, "len_all": 171247, "len_all_1": 74267, "len_abs": 1363, "len_title": 116, "len_sents": 52065, "len_sents_1": 33806, "n_sents": 435, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1370, "title": "Mitigating Spurious Correlations via Group-robust Sample Reweighting", "abs": "Machine learning models often have uneven performance among subpopulations (a.k.a., groups) in the data distributions. This poses a significant challenge for the models to generalize when the proportions of the groups shift during deployment.\nTo improve robustness to such subpopulation shifts, existing approaches have developed strategies that train models or perform hyperparameter tuning using the group-labeled data to minimize the worst-case loss over groups.\nHowever, a non-trivial amount of high-quality labels is often required to obtain noticeable improvements.\nGiven the costliness of the labels, we propose to adopt a different paradigm to enhance group label efficiency:\nutilizing the group-labeled data as a target set to optimize the weights of other group-unlabeled data.\nWe introduce a two-stage approach called Group-robust Sample Reweighting (GSR) that first learns the representations from group-unlabeled data, and then tinkers the model by iteratively retraining its last layer on the reweighted data.\nOur GSR is theoretically sound, practically lightweight, and effective in improving the robustness to subpopulation shifts. In particular, GSR outperforms the previous state-of-the-art results on standard benchmarks when using the same amount of group labels. Notably, GSR even outperforms approaches that require significantly more group labels.", "keywords": [ "distribution shift", "subpopulation shift", "spurious correlation", "influence function" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "aQSbfKYXvo", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Lingzhou Xue", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Full Professor" }, { "name": "Runze Li", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Full Professor" }, { "name": "Xin Yu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Ying Sun", "gender": "Female", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Zelin He", "gender": "Not Specified", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 186, "n_formula_1": 38, "n_ref_uni": 32, "n_ref": 87, "n_ref_all": 148, "n_fig": 5, "n_tab": 4, "L_tab": 522, "n_element_tab": 63, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 234, "n_element_tab_1": 30, "formula_len_all": 14508, "formula_len_all_1": 2606, "len_all": 203638, "len_all_1": 73993, "len_abs": 1191, "len_title": 167, "len_sents": 60864, "len_sents_1": 31154, "n_sents": 582, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 118, "L_abs": 1188, "title": "The Effect of Personalization in FedProx: A Fine-grained Analysis on Statistical Accuracy and Communication Efficiency", "abs": "FedProx is a simple yet effective federated learning method that enables model personalization via regularization. Despite remarkable success in practice, a rigorous analysis of how such a regularization provably improves the statistical accuracy of each client's local model hasn't been fully established. Setting the regularization strength heuristically presents a risk, as an inappropriate choice may even degrade accuracy. This work fills in the gap by analyzing the effect of regularization on statistical accuracy, thereby providing a theoretical guideline for setting the regularization strength for achieving personalization. We prove that by adaptively choosing the regularization strength under different statistical heterogeneity, FedProx can consistently outperform pure local training and achieve a minimax-optimal statistical rate. In addition, to shed light on resource allocation, we design an algorithm, provably showing that stronger personalization reduces communication complexity without increasing the computation cost overhead. Finally, our theory is validated on both synthetic and real-world datasets and its generalizability is verified in a non-convex setting.", "keywords": [ "Personalized Federated Learning", "Statistical Complexity", "Communication Complexity", "Minimax Optimality" ], "rating_list": [ 6, 8, 6, 5, 1 ], "soundness_list": [ 3, 4, 3, 2, 1 ], "presentation_list": [ 3, 3, 3, 2, 1 ], "contribution_list": [ 2, 3, 3, 2, 1 ], "confidence_list": [ 5, 2, 3, 2, 5 ] }, { "paper_id": "aQ7qYnY2nF", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Assaf Hallak", "gender": "Male", "institution": "NVIDIA", "country": "IL", "position": "Senior research scientist" }, { "name": "Assaf Shocher", "gender": "Male", "institution": "NVIDIA", "country": "IL", "position": "Postdoc" }, { "name": "Gal Chechik", "gender": "unknown", "institution": "Bar Ilan University", "country": "IL", "position": "Full Professor" }, { "name": "Shie Mannor", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Full Professor" }, { "name": "Uri Gadot", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 49, "n_ref_all": 71, "n_fig": 9, "n_tab": 6, "L_tab": 1398, "n_element_tab": 92, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 198, "n_element_tab_1": 23, "formula_len_all": 497, "formula_len_all_1": 101, "len_all": 120719, "len_all_1": 64251, "len_abs": 1478, "len_title": 132, "len_sents": 40637, "len_sents_1": 31007, "n_sents": 303, "n_sents_1": 250, "venue": "ICLR2025", "decision": "withdraw", "L_title": 96, "L_abs": 1486, "title": "Real Time Macro-Block Rate Control for Task-Aware Video Compression Using Reinforcement Learning", "abs": "Video encoders optimize compression for human perception by minimizing reconstruction error under bit-rate constraints. In many modern applications such as autonomous driving, an overwhelming majority of videos serve as input for AI systems performing tasks like object recognition or segmentation, rather than being watched by humans. It is therefore useful to optimize the encoder for a downstream task instead of for perceptual image quality. However, a major challenge is how to combine such downstream optimization with existing standard video encoders, which are highly efficient and popular. Here, we address this challenge by controlling the Quantization Parameters (QPs) at the macro-block level to optimize the downstream task. This granular control allows us to prioritize encoding for task-relevant regions within each frame. We formulate this optimization problem as a Reinforcement Learning (RL) task, where the agent learns to balance long-term implications of choosing QPs on both task performance and bit-rate constraints. Notably, our policy does not require the downstream task as an input during inference, making it suitable for streaming applications and edge devices such as vehicles. We demonstrate significant improvements in two tasks, car detection, and ROI (saliency) encoding. Our approach improves task performance for a given bit rate compared to traditional task agnostic encoding methods, paving the way for more efficient task-aware video compression.", "keywords": [ "Video compression", "Rate control", "Reinforcement Learning", "Downstream task" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "aPTGvFqile", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Sedigheh Eslami", "gender": "Female", "institution": "Hasso Plattner Institute", "country": "DE", "position": "PhD student" }, { "name": "Gerard Melo", "gender": "Male", "institution": "Hasso Plattner Institute", "country": "DE", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 18, "n_ref_uni": 26, "n_ref": 45, "n_ref_all": 76, "n_fig": 13, "n_tab": 13, "L_tab": 3222, "n_element_tab": 697, "n_fig_1": 10, "n_tab_1": 6, "L_tab_1": 1554, "n_element_tab_1": 517, "formula_len_all": 1732, "formula_len_all_1": 896, "len_all": 128705, "len_all_1": 63164, "len_abs": 950, "len_title": 105, "len_sents": 36380, "len_sents_1": 25605, "n_sents": 260, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 857, "title": "Mitigate the Gap: Improving Cross-Modal Alignment in CLIP", "abs": "Contrastive Language--Image Pre-training (CLIP) has manifested remarkable improvements in zero-shot classification and cross-modal vision-language tasks. Yet, from a geometrical point of view, the CLIP embedding space has been found to have a pronounced modality gap. This gap renders the embedding space overly sparse and disconnected, with different modalities being densely distributed in distinct subregions of the hypersphere. In this work, we propose AlignCLIP, in order to improve the alignment between text and image embeddings, and thereby reduce the modality gap. AlignCLIP increases the cross-modal alignment, and yields gains across several zero-shot and fine-tuning downstream evaluations by sharing the learnable parameters between the modality encoders and a semantically-regularized separation objective function on the uni-modal embeddings.", "keywords": [ "CLIP", "modality gap", "cross-modal alignment", "multi-modal representation learning" ], "rating_list": [ 6, 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 4, 2, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 4, 4 ] }, { "paper_id": "aPHHhnZktB", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Anna Beer", "gender": "unknown", "institution": "University of Vienna", "country": "AT", "position": "Postdoc" }, { "name": "Anneka Myrup Thiesson", "gender": "Female", "institution": "Aarhus University", "country": "DK", "position": "MS student" }, { "name": "Ira Assent", "gender": "Female", "institution": "Aarhus University", "country": "DK", "position": "Full Professor" }, { "name": "Lena Krieger", "gender": "Female", "institution": "Forschungszentrum Juelich GmbH", "country": "DE", "position": "PhD student" }, { "name": "Pernille Matthews", "gender": "Female", "institution": "Aarhus University", "country": "", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 15, "n_ref_uni": 32, "n_ref": 76, "n_ref_all": 110, "n_fig": 8, "n_tab": 6, "L_tab": 2642, "n_element_tab": 195, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 560, "n_element_tab_1": 43, "formula_len_all": 1173, "formula_len_all_1": 751, "len_all": 135873, "len_all_1": 65671, "len_abs": 837, "len_title": 86, "len_sents": 46593, "len_sents_1": 31214, "n_sents": 364, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 38, "L_abs": 848, "title": "FairDen: Fair Density-Based Clustering", "abs": "Fairness in data mining tasks like clustering has recently become an increasingly important aspect. \nHowever, few clustering algorithms exist that focus on fair groupings of data with sensitive attributes. \nIncluding fairness in the clustering objective is especially hard for density-based clustering, as it does not directly optimize a closed form objective like centroid-based or spectral methods. \n\nThis paper introduces FairDen, the first fair, density-based clustering algorithm.\nWe capture the dataset's density-connectivity structure in a similarity matrix that we manipulate to encourage a balanced clustering. \nIn contrast to state-of-the-art, FairDen inherently handles categorical attributes, noise, and data with several sensitive attributes or groups.\nWe show that FairDen finds meaningful and fair clusters in extensive experiments.", "keywords": [ "Fairness", "Density-based Clustering", "Unsupervised Learning" ], "rating_list": [ 8, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "aP3OBwf8dk", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Angelos Katharopoulos", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Awni Hannun", "gender": "Male", "institution": "Apple", "country": "", "position": "Research Scientist" }, { "name": "David Grangier", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Pierre Ablin", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 41, "n_ref_all": 77, "n_fig": 14, "n_tab": 10, "L_tab": 2457, "n_element_tab": 336, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 410, "n_element_tab_1": 70, "formula_len_all": 445, "formula_len_all_1": 363, "len_all": 179822, "len_all_1": 67467, "len_abs": 1115, "len_title": 100, "len_sents": 40693, "len_sents_1": 31200, "n_sents": 338, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1121, "title": "Need a Small Specialized Language Model? Plan Early!", "abs": "Large language models are versatile tools but are not suitable for small inference budgets. Small models have more efficient inference, but their lower capacity means that their performance can be good only if one limits their scope to a specialized domain. This paper explores how to get good specialized small language models using a large, generic, pretraining set and a limited amount of specialized data. We consider two scenarios, depending on whether (i) one can afford pretraining a model for each specialization task, or (ii) one wants to cheaply adapt a single pretrained model for each task. In the first scenario, we propose an effective solution based on importance sampling: we resample the pretraining set to imitate the specialization data and train a small model on it. In the second scenario, we propose a novel architecture, projected networks (PN). PN is a large network whose parameters can be linearly projected into a small network for specialization. For both scenarios, we demonstrate the empirical effectiveness of our solutions across various domains, training set sizes, and training budgets.", "keywords": [ "Large language models", "pre-training", "efficient inference", "task-adaptive pretraining" ], "rating_list": [ 8, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 2, 2 ] }, { "paper_id": "aOiKt5b0NA", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Albert Bifet", "gender": "Male", "institution": "The University of Waikato", "country": "NZ", "position": "Full Professor" }, { "name": "Danilo Oliveira Andrade", "gender": "Male", "institution": "Universidade Federal da Bahia", "country": "", "position": "Undergrad student" }, { "name": "Islame Felipe da Costa Fernandes", "gender": "Male", "institution": "Universidade Federal da Bahia", "country": "", "position": "Associate Professor" }, { "name": "Joao Gama", "gender": "Male", "institution": "Universidade do Porto", "country": "", "position": "Full Professor" }, { "name": "Marcos Vin\u00edcius dos Santos Ferreira", "gender": "Male", "institution": "Universidade Federal da Bahia", "country": "BR", "position": "PhD student" }, { "name": "Matheus Carvalho de Souza", "gender": "Male", "institution": "Universidade Federal da Bahia", "country": "", "position": "Lecturer" }, { "name": "RICARDO RIOS", "gender": "Male", "institution": "Universidade Federal da Bahia", "country": "BR", "position": "Associate Professor" }, { "name": "Tatiane Nogueira Rios", "gender": "Female", "institution": "Universidade Federal da Bahia", "country": "BR", "position": "Associate Professor" } ], "n_formula": 28, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 158, "n_ref_all": 255, "n_fig": 15, "n_tab": 26, "L_tab": 9542, "n_element_tab": 847, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 2329, "n_element_tab_1": 194, "formula_len_all": 1446, "formula_len_all_1": 0, "len_all": 300184, "len_all_1": 60963, "len_abs": 2478, "len_title": 153, "len_sents": 100082, "len_sents_1": 28661, "n_sents": 703, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1860, "title": "Salvador Urban Network Transportation (SUNT): A Landmark Spatiotemporal Dataset for Public Transportation", "abs": "Efficient public transportation management is essential for the development of large urban centers, providing several benefits such as comprehensive coverage of population mobility, improvement of the local economy with the offer of new jobs and the decrease of transport costs, better control of traffic congestion, and significant reduction of environmental impact limiting gas emissions and pollution. Realizing these benefits requires carefully pursuing two essential pathways: (i) deeply understanding the population and transit patterns and (ii) using intelligent approaches to model multiple relations and characteristics efficiently. This work addresses these challenges by providing a novel dataset that includes various public transportation components alongside machine learning models trained to understand and predict different real-world behaviors. Our dataset comprises daily information from about 710,000 passengers in Salvador, one of Brazil's largest cities, and local public transportation data with approximately 2,000 vehicles operating across nearly 400 lines, connecting almost 3,000 stops and stations. As benchmarks, we have fine-tuned diverse Graph Neural Networks to perform inference on vertices and edges, undertaking both regression and classification tasks. These models leverage temporal and spatial features concerning passengers and transportation data. We emphasize the greatest advantage of using our dataset lies in different possibilities of modeling a real-world urban mobility dataset, reproducing our results, overcoming our models, and investigating several other open-problem situations listed in this manuscript as future work, which include the designing of new methods, optimization strategies, and environmental approaches. Our dataset, codes, and models are available at https://github.com/suntdataset/sunt.git.", "keywords": [ "Graph Neural Network", "Time Series", "Spatial Dataset", "Public Transportation", "Urban Mobility" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "aOJh9JvGcI", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Chengwei AI", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "PhD student" }, { "name": "Fei Guo", "gender": "Female", "institution": "Central South University", "country": "CN", "position": "Full Professor" }, { "name": "Hongpeng Yang", "gender": "unknown", "institution": "University of South Carolina", "country": "US", "position": "PhD student" }, { "name": "Qiaozhen Meng", "gender": "unknown", "institution": "Xiangtan University", "country": "CN", "position": "Lecturer" }, { "name": "Ruihan Dong", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xiaoyi Liu", "gender": "Female", "institution": "University of South Carolina, University of South Carolina", "country": "US", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 39, "n_ref": 69, "n_ref_all": 98, "n_fig": 5, "n_tab": 15, "L_tab": 12161, "n_element_tab": 452, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 2182, "n_element_tab_1": 74, "formula_len_all": 848, "formula_len_all_1": 848, "len_all": 220859, "len_all_1": 59315, "len_abs": 1626, "len_title": 179, "len_sents": 39514, "len_sents_1": 26566, "n_sents": 381, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 130, "L_abs": 1633, "title": "PharmaVQA: A Retrieval-Augmented Visual Question Answering Framework for Molecular Representation via Pharmacophore Guided Prompts", "abs": "In drug discovery, molecular representation learning is vital for understanding and generating new drug-like molecules. The accurate representation of molecules facilitates drug candidate screening and the optimization of lead compounds. The vastness of chemical space challenges traditional drug design and relies on complex computations. The Pharmacophore is a functional group contained within a drug molecule, which binds to receptors or biological macromolecules to produce biological effects and reduce computations. Pharmacophore-guided representation of molecules, however, remains a significant challenge. To address this issue, we propose an improved deep learning-based model called PharmaVQA for retrieving pharmacophore-related information directly from molecule databases, allowing for a more targeted understanding of drug-like molecules. Through the use of Visual Question Answering (VQA) framework, PharmaVQA captures pharmacophore data, generates knowledge prompts, and enriches molecular representations. On 46 benchmark datasets, PharmaVQA has demonstrated superior performance in both molecular property prediction and drug-target interaction prediction. Additionally, the applicability of PharmaVQA in drug discovery has been validated on an FDA-approved molecule dataset, where the Top-20 predictions were analyzed in real-world studies, with the majority of them experimentally validated as potential ligands previously reported in the literature. Our assessment of PharmaVQA is that it is a powerful and useful tool for accelerating the development of AI-assisted drug discovery across a wide range of areas.", "keywords": [ "Pharmacophore", "Visual Question Answering", "Prompt Learning", "Molecule Represenation Learning", "Bilinear Attention Network", "Multi-modal Retrieval" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "aOBYawUuWL", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Cazalet Tanguy", "gender": "Not Specified", "institution": "Universiteit Gent", "country": "BE", "position": "PhD student" }, { "name": "Joni Dambre", "gender": "unknown", "institution": "Universiteit Gent", "country": "", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 4, "n_ref_uni": 28, "n_ref": 38, "n_ref_all": 60, "n_fig": 11, "n_tab": 8, "L_tab": 2976, "n_element_tab": 261, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 841, "n_element_tab_1": 103, "formula_len_all": 414, "formula_len_all_1": 129, "len_all": 127166, "len_all_1": 54730, "len_abs": 2216, "len_title": 119, "len_sents": 46391, "len_sents_1": 25847, "n_sents": 318, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1216, "title": "Reshaping Reservoirs: Hebbian Plasticity for Improved Data Separability", "abs": "This paper introduces Hebbian Architecture Generation (HAG), a method grounded in Hebbian plasticity principles, designed to optimize the structure of Reservoir Computing networks. HAG adapts the synaptic weights in Recurrent Neural Networks by dynamically forming connections between neurons that exhibit high Pearson correlation. Unlike conventional reservoir computing models that rely on static, randomly initialized connectivity matrices, HAG tailors the reservoir architecture to specific tasks by autonomously optimizing network properties such as signal decorrelation and singular value spread. This task-specific adaptability enhances the linear separability of input data, as supported by Cover\u2019s theorem, which posits that increasing the dimensionality of the feature space improves pattern recognition. Experimental results show that HAG outperforms traditional Echo State Networks across various predictive modeling and pattern recognition benchmarks. By aligning with biological principles of structural plasticity, HAG addresses limitations of static reservoir architectures, offering a biologically plausible and highly adaptable alternative for improved performance in dynamic learning environments.", "keywords": [ "bio-inspired", "hebian plasticity", "echo states network", "unsupervised learning", "time series" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "aOAgMiOXU2", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Gust Verbruggen", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Mukul Singh", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Sumit Gulwani", "gender": "Male", "institution": "Research, Microsoft", "country": "", "position": "Researcher" }, { "name": "Vu Le", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 26, "n_ref": 67, "n_ref_all": 88, "n_fig": 9, "n_tab": 2, "L_tab": 333, "n_element_tab": 64, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1848, "n_element_tab_1": 83, "formula_len_all": 760, "formula_len_all_1": 501, "len_all": 99647, "len_all_1": 55281, "len_abs": 1088, "len_title": 106, "len_sents": 28850, "len_sents_1": 23728, "n_sents": 222, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1119, "title": "Code diffusion models are continuous human noise operators", "abs": "Diffusion for code generates code by iteratively removing noise from the latent representation of a code snippet.\nDuring later steps of the diffusion process, when the code snippet has almost converged, these edits resemble last-mile repairs applied to broken or incomplete code. We evaluate the extent to which these errors are similar to those that humans are faced with and the capability of these models to perform last-mile repair. Our insight has two applications with significant impact for code repair. First, we can leverage the diffusion model for last-mile repair by adding noise to a broken code snippet and resuming the diffusion process. Second, we can leverage the diffusion model to generate an arbitrary amount of training data for other last-mile repair approaches (that are computationally more efficient) by sampling an intermediate program (input) and the final program (output) from the diffusion process. We perform experiments to evaluate both applications, as well as analyze trends in the evolution of representation through the diffusion pipeline providing insights on the reasoning observed.", "keywords": [ "Program Repair", "Text Diffusion", "Code Diffusion", "Language Models" ], "rating_list": [ 5, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "aNf8VCQE0h", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 2, "author_info_list": [ { "name": "Mohammad Raza", "gender": "Male", "institution": "Hamad Bin Khalifa University", "country": "QA", "position": "Researcher" }, { "name": "Natasa Milic-Frayling", "gender": "Female", "institution": "Qatar Computing Research Institute", "country": "QA", "position": "Principal Researcher" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 22, "n_ref": 86, "n_ref_all": 126, "n_fig": 10, "n_tab": 5, "L_tab": 15982, "n_element_tab": 94, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1102, "formula_len_all_1": 39, "len_all": 304253, "len_all_1": 67031, "len_abs": 1115, "len_title": 146, "len_sents": 84186, "len_sents_1": 35665, "n_sents": 745, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1141, "title": "Almost Sure Reasoning: Generating Verified Formalizations with Language Models and Logical Solvers", "abs": "Robustness of reasoning remains a challenging problem for large language models, and addressing it is crucial for advancing the reliability and practical application of AI-driven reasoning systems. We introduce Semantic Self-Verification (SSV), a novel approach that addresses the key challenge in combining language models with the rigor of logical solvers: to accurately translate the reasoning problem from natural language to the formal language of the solver. SSV produces strong abstract formalizations of problems by verifying and refining them against concrete instantiations that are generated by the model and verified by the solver. In addition to significantly advancing the overall reasoning accuracy over the state-of-the-art, a key novelty that this approach presents is a feature of verification that has near-perfect precision over a significant coverage of cases, as we demonstrate on open reasoning benchmarks. We propose such $\\textit{near-certain reasoning}$ as a new approach that can reduce the need for manual human verification in many cases, taking us closer to more dependable and autonomous AI reasoning systems.", "keywords": [ "Automated Reasoning", "SAT/SMT Solvers", "Formal Methods", "Large Language Models" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "aNYabH9Th4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Changick Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Donguk Kim", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "MS student" }, { "name": "Jaehyuk Jang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Sangmin Woo", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "\ucd5c\uc720\ube48", "gender": "Female", "institution": "KAIST, Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 4, "n_ref_uni": 57, "n_ref": 171, "n_ref_all": 203, "n_fig": 9, "n_tab": 19, "L_tab": 11367, "n_element_tab": 1575, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 5970, "n_element_tab_1": 526, "formula_len_all": 593, "formula_len_all_1": 148, "len_all": 243649, "len_all_1": 75442, "len_abs": 1733, "len_title": 132, "len_sents": 59914, "len_sents_1": 29105, "n_sents": 522, "n_sents_1": 230, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1658, "title": "RITUAL: Random Image Transformations as a Universal Anti-hallucination Lever in LVLMs", "abs": "Recent advancements in Large Vision Language Models (LVLMs) have revolutionized how machines understand and generate textual responses based on visual inputs. Despite their impressive capabilities, they often produce \"hallucinatory\" outputs that do not accurately reflect the visual information, posing challenges in reliability and trustworthiness. Inspired by test-time augmentation, we propose a simple, training-free method termed RITUAL to enhance robustness against hallucinations in LVLMs. RITUAL introduces random image transformations as complementary inputs during the decoding phase. Importantly, these transformations are not employed during the training of the LVLMs. This straightforward strategy reduces the likelihood of hallucinations by exposing the model to varied visual scenarios, enriching its decision-making process. While transformed images alone may initially degrade performance, we empirically find that strategically combining them with the original images mitigates hallucinations. Specifically, in cases where hallucinations occur with the original image, the transformed images help correct misinterpretations by adjusting the probability distribution. By diversifying the visual input space, RITUAL provides a more robust foundation for generating accurate outputs. Notably, our method works seamlessly with existing contrastive decoding methods and does not require external models or costly self-feedback mechanisms, making it a practical addition. While extremely simple, RITUAL significantly outperforms existing contrastive decoding methods across several object hallucination benchmarks, including POPE, CHAIR, and MME.", "keywords": [ "Large Vision Language Models" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "aMD0qUyYJh", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Anna Winnicki", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "PhD student" }, { "name": "R. Srikant", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "" }, { "name": "Shie Mannor", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Full Professor" }, { "name": "Yihan Du", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "Postdoc" }, { "name": "Gal Dalal", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Senior Research Scientist" } ], "n_formula": 227, "n_formula_1": 24, "n_ref_uni": 17, "n_ref": 66, "n_ref_all": 107, "n_fig": 4, "n_tab": 5, "L_tab": 1139, "n_element_tab": 42, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 847, "n_element_tab_1": 11, "formula_len_all": 37044, "formula_len_all_1": 2471, "len_all": 231621, "len_all_1": 70497, "len_abs": 1415, "len_title": 93, "len_sents": 62372, "len_sents_1": 29453, "n_sents": 766, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1428, "title": "Reinforcement Learning with Segment Feedback", "abs": "Classic reinforcement learning (RL) assumes that an agent can observe a reward for each state-action pair. However, in practical applications, it is often difficult and costly to collect a reward for each state-action pair. While there have been several works considering RL with trajectory feedback, it is unclear if trajectory feedback is inefficient for learning when trajectories are long. In this work, we propose a model named RL with segment feedback, which offers a general paradigm filling the gap between per-state-action feedback and trajectory feedback seemlessly. In this model, we consider an episodic Markov decision process (MDP), where each episode is equally divided into $m$ segments, and the agent observes reward feedback only at the end of each segment. Under this model, we study two popular feedback settings: binary feedback and sum feedback, where the agent observes a binary outcome and a reward sum according to the underlying reward function, respectively. To investigate the impacts of the number of segments $m$ on learning performance, we design efficient algorithms and establish regret upper and lower bounds for both feedback settings. Our theoretical and empirical results show that: under binary feedback, increasing the number of segments $m$ decreases the regret at an exponential rate; in contrast, surprisingly under sum feedback, increasing $m$ does not reduce the regret significantly.", "keywords": [ "Reinforcement learning (RL)", "segment feedback", "binary feedback", "sum feedback" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "aMBSY2ebPw", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Christof Monz", "gender": "Male", "institution": "University of Amsterdam, University of Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "David Stap", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "PhD student" }, { "name": "Di Wu", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "PhD student" }, { "name": "Khalil Sima'an", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "Full Professor" }, { "name": "Seth Aycock", "gender": "unknown", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 60, "n_ref": 97, "n_ref_all": 129, "n_fig": 3, "n_tab": 10, "L_tab": 1735, "n_element_tab": 300, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 633, "n_element_tab_1": 108, "formula_len_all": 17, "formula_len_all_1": 0, "len_all": 220017, "len_all_1": 65119, "len_abs": 1501, "len_title": 129, "len_sents": 54032, "len_sents_1": 32885, "n_sents": 421, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 81, "L_abs": 1510, "title": "Can LLMs Really Learn to Translate a Low-Resource Language from One Grammar Book?", "abs": "Extremely low-resource (XLR) languages lack substantial corpora for training NLP models, motivating the use of all available resources such as dictionaries and grammar books. Machine Translation from One Book (Tanzer et al., 2024) suggests prompting long-context LLMs with one grammar book enables English\u2013Kalamang translation, an unseen XLR language\u2014a noteworthy case of linguistic knowledge helping an NLP task. We investigate whether the book's grammatical explanations or its parallel examples are most effective for learning XLR translation, finding almost all improvement stems from the parallel examples. We find similar results for Nepali and Guarani, seen low-resource languages, and achieve performance comparable to an LLM with a grammar book by simply fine-tuning an encoder-decoder translation model. We then investigate *where* grammar books help by testing two linguistic tasks, grammaticality judgment and gloss prediction, and we explore what *kind* of grammatical knowledge helps by introducing a typological feature prompt that achieves leading results on these more relevant tasks. We thus emphasise the importance of task-appropriate data for XLR languages: parallel examples for translation, and grammatical data for linguistic tasks. As we find no evidence that long-context LLMs can make effective use of grammatical explanations for XLR translation, we suggest data collection for multilingual XLR tasks such as translation is best focused on parallel data over linguistic description.", "keywords": [ "llms", "translation", "low-resource", "grammar", "long-context", "linguistics" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 4, 4, 4 ], "presentation_list": [ 4, 4, 4 ], "contribution_list": [ 2, 3, 4 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "aLsMzkTej9", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "James Hensman", "gender": "unknown", "institution": "Microsoft Research", "country": "GB", "position": "Principal Researcher" }, { "name": "Liana Mikaelyan", "gender": "Female", "institution": "Microsoft", "country": "GB", "position": "Researcher" }, { "name": "Taketomo Isazawa", "gender": "unknown", "institution": "Research, Microsoft", "country": "GB", "position": "Researcher" }, { "name": "Xi Wang", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 18, "n_ref_uni": 26, "n_ref": 40, "n_ref_all": 97, "n_fig": 21, "n_tab": 5, "L_tab": 3454, "n_element_tab": 102, "n_fig_1": 16, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1213, "formula_len_all_1": 1329, "len_all": 203195, "len_all_1": 78740, "len_abs": 1088, "len_title": 94, "len_sents": 70361, "len_sents_1": 36630, "n_sents": 536, "n_sents_1": 282, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 46, "L_abs": 1161, "title": "KBLaM: Knowledge Base augmented Language Model", "abs": "In this paper, we propose Knowledge Base augmented Language Model (KBLAM), a new method for augmenting Large Language Models (LLMs) with external knowledge. KBLAM works with a knowledge base (KB) constructed from a corpus of documents, transforming each piece of knowledge in the KB into continuous key-value vector pairs via pre-trained sentence encoders with linear adapters and\nintegrating them into pre-trained LLMs via a specialized rectangular attention mechanism. Unlike Retrieval-Augmented Generation, KBLAM eliminates external retrieval modules, and unlike in-context learning, its computational overhead scales linearly with KB size rather than quadratically. Our approach enables integrating a large KB of more than 10K triples into an 8B pre-trained LLM of only 8K context window on one single A100 80GB GPU and allows for dynamic updates without model fine-tuning or retraining. Experiments demonstrate KBLAM\u2019s effectiveness in various tasks, including question-answering and open-ended reasoning, while providing interpretable insights into its use of the augmented knowledge. Code and datasets are available at https://github.com/microsoft/KBLaM/", "keywords": [ "Large language model; Knowledge augmentation;" ], "rating_list": [ 3, 8, 5, 5, 8 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4, 3 ] }, { "paper_id": "aLSI9Z4UMD", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Abhijit Kundu", "gender": "unknown", "institution": "Google DeepMind", "country": "US", "position": "Research Engineer" }, { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Weijie Lyu", "gender": "unknown", "institution": "University of California, Merced", "country": "US", "position": "PhD student" }, { "name": "Yi-Hsuan Tsai", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Xueting Li", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 36, "n_ref": 113, "n_ref_all": 140, "n_fig": 13, "n_tab": 6, "L_tab": 2052, "n_element_tab": 234, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 854, "n_element_tab_1": 11, "formula_len_all": 204, "formula_len_all_1": 153, "len_all": 147251, "len_all_1": 55124, "len_abs": 1202, "len_title": 98, "len_sents": 48747, "len_sents_1": 28378, "n_sents": 381, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1184, "title": "Gaga: Group Any Gaussians via 3D-aware Memory Bank", "abs": "We introduce *Gaga*, a framework that reconstructs and segments open-world 3D scenes by leveraging inconsistent 2D masks predicted by zero-shot class-agnostic segmentation models. Contrasted to prior 3D scene segmentation approaches that heavily rely on video object tracking, *Gaga* utilizes spatial information provided by 3D Gaussians and effectively associates object masks across diverse camera poses through a novel 3D-aware memory bank. By eliminating the assumption of continuous view changes in training images, *Gaga* demonstrates robustness to variations in camera poses, particularly beneficial for sparsely sampled images, ensuring precise mask label consistency. Furthermore, *Gaga* accommodates 2D segmentation masks from diverse sources and demonstrates robust performance with different open-world zero-shot class-agnostic segmentation models, significantly enhancing its versatility. Extensive qualitative and quantitative evaluations demonstrate that *Gaga* performs favorably against state-of-the-art methods, emphasizing its potential for real-world applications such as scene understanding and manipulation. The source codes will be made available to the public.", "keywords": [ "3D Open-world Segmentation; Gaussian Splatting; Scene Understanding" ], "rating_list": [ 5, 3, 5, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 5, 3, 4 ] }, { "paper_id": "aKkDY1Wca0", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Adel Javanmard", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "Full Professor" }, { "name": "Murat A Erdogdu", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "Assistant Professor" }, { "name": "Alireza Mousavi", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "", "position": "PhD student" } ], "n_formula": 214, "n_formula_1": 24, "n_ref_uni": 52, "n_ref": 102, "n_ref_all": 112, "n_fig": 1, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 54, "n_element_tab_1": 1, "formula_len_all": 20373, "formula_len_all_1": 2608, "len_all": 270433, "len_all_1": 72688, "len_abs": 1429, "len_title": 113, "len_sents": 71983, "len_sents_1": 31896, "n_sents": 815, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1439, "title": "Robust Feature Learning for Multi-Index Models in High Dimensions", "abs": "Recently, there have been numerous studies on feature learning with neural networks, specifically on learning single- and multi-index models where the target is a function of a low-dimensional projection of the input. Prior works have shown that in high dimensions, the majority of the compute and data resources are spent on recovering the low-dimensional projection; once this subspace is recovered, the remainder of the target can be learned independently of the ambient dimension. However, implications of feature learning in adversarial settings remain unexplored. In this work, we take the first steps towards understanding adversarially robust feature learning with neural networks. Specifically, we prove that the hidden directions of a multi-index model offer a Bayes optimal low-dimensional projection for robustness against $\\ell_2$-bounded adversarial perturbations under the squared loss, assuming that the multi-index coordinates are statistically independent from the rest of the coordinates. Therefore, robust learning can be achieved by first performing standard feature learning, then robustly tuning a linear readout layer on top of the standard representations. In particular, we show that adversarially robust learning is just as easy as standard learning. Specifically, the additional number of samples needed to robustly learn multi-index models when compared to standard learning, does not depend on dimensionality.", "keywords": [ "feature learning", "adversarial robustness", "neural networks", "multi-index models", "gradient descent" ], "rating_list": [ 8, 8, 6, 6, 8 ], "soundness_list": [ 3, 4, 3, 3, 3 ], "presentation_list": [ 2, 4, 2, 2, 2 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 3, 3 ] }, { "paper_id": "aKcd7ImG5e", "primary_area": "learning on time series and dynamical systems", "n_author": 8, "author_info_list": [ { "name": "Bin Yang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Chenjuan Guo", "gender": "unknown", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Kai Zhao", "gender": "Non-Binary", "institution": "Aalborg University", "country": "DK", "position": "PhD student" }, { "name": "Lujia Pan", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Qichao Shentu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Yang Shu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhongwen Rao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "beibu li", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 60, "n_ref": 152, "n_ref_all": 199, "n_fig": 12, "n_tab": 13, "L_tab": 6303, "n_element_tab": 760, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3649, "n_element_tab_1": 219, "formula_len_all": 706, "formula_len_all_1": 692, "len_all": 196349, "len_all_1": 69018, "len_abs": 1561, "len_title": 150, "len_sents": 52807, "len_sents_1": 30422, "n_sents": 394, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 1566, "title": "Towards a General Time Series Anomaly Detector with Adaptive Bottlenecks and Dual Adversarial Decoders", "abs": "Time series anomaly detection plays a vital role in a wide range of applications. Existing methods require training one specific model for each dataset, which exhibits limited generalization capability across different target datasets, hindering anomaly detection performance in various scenarios with scarce training data. Aiming at this problem, we propose constructing a general time series anomaly detection model, which is pre-trained on extensive multi-domain datasets and can subsequently apply to a multitude of downstream scenarios. The significant divergence of time series data across different domains presents two primary challenges in building such a general model: (1) meeting the diverse requirements of appropriate information bottlenecks tailored to different datasets in one unified model, and (2) enabling distinguishment between multiple normal and abnormal patterns, both are crucial for effective anomaly detection in various target scenarios. To tackle these two challenges, we propose a General time series anomaly Detector with Adaptive Bottlenecks and Dual Adversarial Decoders (DADA), which enables flexible selection of bottlenecks based on different data and explicitly enhances clear differentiation between normal and abnormal series. We conduct extensive experiments on nine target datasets from different domains. After pre-training on multi-domain data, DADA, serving as a zero-shot anomaly detector for these datasets, still achieves competitive or even superior results compared to those models tailored to each specific dataset.", "keywords": [ "Time series", "Anomaly detection" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 5, 3 ] }, { "paper_id": "aKRADWBJ1I", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Andreas Krause", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Bhavya Sukhija", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Carmelo Sferrazza", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "Postdoc" }, { "name": "Lenart Treven", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Stelian Coros", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "Associate Professor" }, { "name": "Yarden As", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" } ], "n_formula": 57, "n_formula_1": 20, "n_ref_uni": 50, "n_ref": 106, "n_ref_all": 136, "n_fig": 12, "n_tab": 2, "L_tab": 182, "n_element_tab": 16, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 116, "n_element_tab_1": 2, "formula_len_all": 7531, "formula_len_all_1": 1814, "len_all": 184513, "len_all_1": 70094, "len_abs": 3043, "len_title": 127, "len_sents": 48419, "len_sents_1": 31870, "n_sents": 453, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1206, "title": "ActSafe: Active Exploration with Safety Constraints for Reinforcement Learning", "abs": "Reinforcement learning (RL) is ubiquitous in the development of modern AI systems. However, state-of-the-art RL agents require extensive, and potentially\nunsafe, interactions with their environments to learn effectively. These limitations\nconfine RL agents to simulated environments, hindering their ability to learn\ndirectly in real-world settings. In this work, we present ActSafe, a novel\nmodel-based RL algorithm for safe and efficient exploration. ActSafe learns\na well-calibrated probabilistic model of the system and plans optimistically\nw.r.t. the epistemic uncertainty about the unknown dynamics, while enforcing\npessimism w.r.t. the safety constraints. Under regularity assumptions on the\nconstraints and dynamics, we show that ActSafe guarantees safety during\nlearning while also obtaining a near-optimal policy in finite time. In addition, we\npropose a practical variant of ActSafe that builds on latest model-based RL advancements and enables safe exploration even in high-dimensional settings such\nas visual control. We empirically show that ActSafe obtains state-of-the-art\nperformance in difficult exploration tasks on standard safe deep RL benchmarks\nwhile ensuring safety during learning.", "keywords": [ "Safe Exploration", "Constrained Markov Decision Processes", "Safe Reinforcement Learning" ], "rating_list": [ 6, 5, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 3, 4, 2, 3 ] }, { "paper_id": "aKJr5NnN8U", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Bryan Chan", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "PhD student" }, { "name": "Dale Schuurmans", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Professor" }, { "name": "Xinyi Chen", "gender": "Female", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Andras Gyorgy", "gender": "unknown", "institution": "DeepMind", "country": "", "position": "Research Scientist" } ], "n_formula": 34, "n_formula_1": 18, "n_ref_uni": 30, "n_ref": 88, "n_ref_all": 139, "n_fig": 22, "n_tab": 4, "L_tab": 702, "n_element_tab": 386, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 712, "n_element_tab_1": 387, "formula_len_all": 3964, "formula_len_all_1": 1446, "len_all": 193349, "len_all_1": 75939, "len_abs": 1097, "len_title": 102, "len_sents": 69626, "len_sents_1": 34282, "n_sents": 534, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1100, "title": "Toward Understanding In-context vs. In-weight Learning", "abs": "It has recently been demonstrated empirically that in-context learning emerges in transformers when certain distributional properties are present in the training data, but this ability can also diminish upon further training. We provide a new theoretical understanding of these phenomena by identifying simplified distributional properties that give rise to the emergence and eventual disappearance of in-context learning. We do so by first analyzing a simplified model that uses a gating mechanism to choose between an in-weight and an in-context predictor. Through a combination of a generalization error and regret analysis we identify conditions where in-context and in-weight learning emerge. These theoretical findings are then corroborated experimentally by comparing the behaviour of a full transformer on the simplified distributions to that of the stylized model, demonstrating aligned results. We then extend the study to a full large language model, showing how fine-tuning on various collections of natural language prompts can elicit similar in-context and in-weight learning behaviour.", "keywords": [ "In-context learning", "generalization error", "transformers" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "aKFFpfiJHy", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Garud Iyengar", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Full Professor" }, { "name": "Min-hwan Oh", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" }, { "name": "Sungwoo PARK", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Wonyoung Kim", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Postdoc" }, { "name": "assaf zeevi", "gender": "Male", "institution": "Columbia University", "country": "", "position": "Full Professor" } ], "n_formula": 112, "n_formula_1": 29, "n_ref_uni": 25, "n_ref": 61, "n_ref_all": 91, "n_fig": 5, "n_tab": 3, "L_tab": 933, "n_element_tab": 153, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 36, "n_element_tab_1": 8, "formula_len_all": 14293, "formula_len_all_1": 2654, "len_all": 159480, "len_all_1": 71803, "len_abs": 1192, "len_title": 97, "len_sents": 47130, "len_sents_1": 30735, "n_sents": 431, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1221, "title": "Linear Bandits with Partially Observable Features", "abs": "We introduce a novel linear bandit problem where a subset of features is latent, resulting in partial access to reward information and spurious estimates.\nWithout properly addressing the latent features, the regret grows linearly over the decision epoch $T$ while improving the regret bound is challenging because their dimension and relationship with rewards are not available.\nWe propose a novel analysis to handle the latent features and an algorithm that achieves a regret bound sublinear in $T$.\nThe core of the algorithm lies in (i) augmenting basis vectors orthogonal to the observable feature space, and (ii) developing an efficient doubly robust estimator that further improves the regret bound.\nWith these two ingredients, our algorithm achieves a regret bound of $\\tilde{O}(\\sqrt{(d + d\\_h)T})$, where $d$ is the dimension of observable features, and $d_h$ is the _unknown_ dimension of the unobserved features that affects the reward. \nCrucially, our algorithm does not rely on prior knowledge of the unobserved feature space, which expands as more features become hidden.\nNumerical experiments confirm that our algorithm outperforms both non-contextual multi-armed bandits and other linear bandit algorithms.", "keywords": [ "Linear Bandits", "Partially Observable Features", "Doubly Robust" ], "rating_list": [ 3, 6, 3, 8 ], "soundness_list": [ 1, 3, 1, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 1, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "aJnKjvTtPq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Fei Chao", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Shen Li", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Staff Algorithm Engineer" }, { "name": "Wanchen Sui", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Staff Algorithm Engineer" }, { "name": "Wenting Lin", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Xiawu Zheng", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiwei Xu", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Yong Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Architect" }, { "name": "Yuexiao Ma", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Yuhang Wu", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Zelan Yang", "gender": "Female", "institution": "Biomedical Engineering school, Azad University Science and research branch", "country": "", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 40, "n_ref": 82, "n_ref_all": 94, "n_fig": 1, "n_tab": 14, "L_tab": 3797, "n_element_tab": 787, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1165, "n_element_tab_1": 110, "formula_len_all": 801, "formula_len_all_1": 465, "len_all": 153448, "len_all_1": 52288, "len_abs": 1632, "len_title": 105, "len_sents": 39084, "len_sents_1": 23491, "n_sents": 297, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1529, "title": "Low Rank Quantization Adaptation for Large Language Model", "abs": "As the parameters of Large Language Models (LLMs) increase, quantization has emerged as a potent strategy for model compression and acceleration. Concurrently, Low-Rank Adaptation (LoRA) has been recognized as an effective method for enhancing LLM performance. However, integrating LoRA with quantization presents significant challenges, particularly in preserving the quantization format after model optimization. In this paper, we introduce Low rank Quantization Adaptation (LoQA) for LLM, a novel approach that effectively fine-tunes holistic quantization parameters. Specifically, we first propose a new perspective of quantization operator, which is compatiable with LoRA and mathematically equivalent to the original operator. In this way, all the parameters (scale and zero point) are finetuned simultaneously, and thus yields notable improvements in model performance.Thanks to the expanded optimization landscape, LoQA is broadly applicabile to various Post-Training Quantization (PTQ) techniques, ensuring better generalizability in practical deployments. To maintain the stability of the optimization, we further propose a LoRA scaling strategy that leverages quantization data to adjust the norm of the low rank adaptation, regulating the speed of convergence in optimization and preventing inappropriate LoRA scaling, which could lead to overfitting or underfitting. Compared to existing methods, LoQA consistently achieves performance gains across a wide range of models, proving its effectiveness and adaptability.", "keywords": [ "Quantization", "Low-Rank Adaptation", "LLM" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "aJUuere4fM", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Maksym Andriushchenko", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Postdoc" }, { "name": "Nicolas Flammarion", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 30, "n_ref": 64, "n_ref_all": 87, "n_fig": 4, "n_tab": 4, "L_tab": 1455, "n_element_tab": 51, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1229, "n_element_tab_1": 44, "formula_len_all": 617, "formula_len_all_1": 676, "len_all": 109649, "len_all_1": 61818, "len_abs": 1444, "len_title": 107, "len_sents": 30258, "len_sents_1": 25350, "n_sents": 227, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1459, "title": "Does Refusal Training in LLMs Generalize to the Past Tense?", "abs": "Refusal training is widely used to prevent LLMs from generating harmful, undesirable, or illegal outputs. We reveal a curious generalization gap in the current refusal training approaches: simply reformulating a harmful request in the past tense (e.g., *\"How to make a Molotov cocktail?\"* to *\"How did people make a Molotov cocktail?\"*) is often sufficient to jailbreak many state-of-the-art LLMs. We systematically evaluate this method on Llama-3 8B, Claude-3.5 Sonnet, GPT-3.5 Turbo, Gemma-2 9B, Phi-3-Mini, GPT-4o-mini, GPT-4o, o1-mini, o1-preview, and R2D2 models using GPT-3.5 Turbo as a reformulation model. For example, the success rate of this simple attack on GPT-4o increases from 1\\% using direct requests to 88\\% using 20 past-tense reformulation attempts on harmful requests from JailbreakBench with GPT-4 as a jailbreak judge. Interestingly, we also find that reformulations in the future tense are less effective, suggesting that refusal guardrails tend to consider past historical questions more benign than hypothetical future questions. Moreover, our experiments on fine-tuning GPT-3.5 Turbo show that defending against past reformulations is feasible when past tense examples are explicitly included in the fine-tuning data. Overall, our findings highlight that the widely used alignment techniques---such as SFT, RLHF, and adversarial training---employed to align the studied models can be brittle and do not always generalize as intended.", "keywords": [ "Jailbreaking", "adversarial attacks", "adversarial robustness", "AI safety" ], "rating_list": [ 8, 1, 6, 8 ], "soundness_list": [ 4, 2, 3, 4 ], "presentation_list": [ 4, 2, 3, 4 ], "contribution_list": [ 3, 1, 3, 4 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "aJ3tiX1Tu4", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Chao Chen", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Postdoc" }, { "name": "Parvathinathan Venkitasubramaniam", "gender": "Male", "institution": "Lehigh University", "country": "", "position": "Full Professor" }, { "name": "Rui Xu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Sihong Xie", "gender": "Male", "institution": "HKUST-GZ", "country": "CN", "position": "Associate Professor" }, { "name": "Yue Sun", "gender": "Male", "institution": "Lehigh University", "country": "", "position": "PhD student" } ], "n_formula": 68, "n_formula_1": 33, "n_ref_uni": 49, "n_ref": 137, "n_ref_all": 218, "n_fig": 16, "n_tab": 5, "L_tab": 1105, "n_element_tab": 69, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 122, "n_element_tab_1": 8, "formula_len_all": 7600, "formula_len_all_1": 4096, "len_all": 175704, "len_all_1": 68783, "len_abs": 1656, "len_title": 125, "len_sents": 53526, "len_sents_1": 28544, "n_sents": 457, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1354, "title": "Wasserstein-Regularized Conformal Prediction under General Distribution Shift", "abs": "Conformal prediction yields a prediction set with guaranteed $1-\\alpha$ coverage of the true target under the i.i.d. assumption, \nwhich can fail and lead to a gap between $1-\\alpha$ and the actual coverage. Prior studies bound the gap using total variation distance, which cannot identify the gap changes under distribution shift at different $\\alpha$, thus serving as a weak indicator of prediction set validity. Besides, existing methods are mostly limited to covariate shifts, while general joint distribution shifts are more common in practice but less researched. In response, we first propose a Wasserstein distance-based upper bound of the coverage gap and analyze the bound using probability measure pushforwards between the shifted joint data and conformal score distributions, enabling a separation of the effect of covariate and concept shifts over the coverage gap. We exploit the separation to design algorithms based on importance weighting and regularized representation learning (WR-CP) to reduce the Wasserstein bound with a finite-sample error bound. WR-CP achieves a controllable balance between conformal prediction accuracy and efficiency. Experiments on six datasets prove that WR-CP can reduce coverage gaps to 3.1% across different confidence levels and outputs prediction sets 38% smaller than the worst-case approach on average.", "keywords": [ "conformal prediction", "distribution shift", "non-exchangeability" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "aIMi2lOKIn", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Changqing Zou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Lihong Wang", "gender": "Male", "institution": "Jilin University", "country": "", "position": "MS student" }, { "name": "Rui Ma", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Associate Professor" }, { "name": "Tieru Wu", "gender": "Male", "institution": "school of AI, Jilin University", "country": "CN", "position": "Full Professor" }, { "name": "Yibo Zhang", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "MS student" } ], "n_formula": 37, "n_formula_1": 21, "n_ref_uni": 46, "n_ref": 113, "n_ref_all": 140, "n_fig": 15, "n_tab": 4, "L_tab": 464, "n_element_tab": 50, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 286, "n_element_tab_1": 33, "formula_len_all": 2813, "formula_len_all_1": 1073, "len_all": 167763, "len_all_1": 67181, "len_abs": 1200, "len_title": 128, "len_sents": 48163, "len_sents_1": 31066, "n_sents": 375, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1239, "title": "Diff3DS: Generating View-Consistent 3D Sketch via Differentiable Curve Rendering", "abs": "3D sketches are widely used for visually representing the 3D shape and structure of objects or scenes. However, the creation of 3D sketch often requires users to possess professional artistic skills. Existing research efforts primarily focus on enhancing the ability of interactive sketch generation in 3D virtual systems. In this work, we propose Diff3DS, a novel differentiable rendering framework for generating view-consistent 3D sketch by optimizing 3D parametric curves under various supervisions. Specifically, we perform perspective projection to render the 3D rational B\u00e9zier curves into 2D curves, which are subsequently converted to a 2D raster image via our customized differentiable rasterizer. Our framework bridges the domains of 3D sketch and raster image, achieving end-to-end optimization of 3D sketch through gradients computed in the 2D image domain. Our Diff3DS can enable a series of novel 3D sketch generation tasks, including text-to-3D sketch and image-to-3D sketch, supported by the popular distillation-based supervision, such as Score Distillation Sampling (SDS). Extensive experiments have yielded promising results and demonstrated the potential of our framework. Project: https://yiboz2001.github.io/Diff3DS/", "keywords": [ "3D Sketch", "Sketch Generation", "Diffusion Models" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "aIJTNrF2Sg", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Heung-Il Suk", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Full Professor" }, { "name": "Jaehyun Jeon", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "MS student" }, { "name": "Junghyo Sohn", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" }, { "name": "Seungwoo Jeong", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" }, { "name": "Yeajin Shon", "gender": "Female", "institution": "Korea University", "country": "KR", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 28, "n_ref": 62, "n_ref_all": 73, "n_fig": 4, "n_tab": 3, "L_tab": 2617, "n_element_tab": 126, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 798, "formula_len_all_1": 798, "len_all": 97237, "len_all_1": 57170, "len_abs": 1117, "len_title": 113, "len_sents": 30340, "len_sents_1": 27623, "n_sents": 238, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1121, "title": "Frequency-Conditioned Diffusion Models for Time Series Generation", "abs": "Time series data, commonly used in fields like climate studies, finance, and healthcare, usually faces challenges such as missing data and privacy concerns. Recently, diffusion models have emerged as effective tools for generating high-quality data, but applying them to time series is still difficult, especially for capturing long-range dependencies and complex information. In this paper, we introduce a new diffusion model that uses frequency domain information to improve time series data generation. In particular, we apply Fourier analysis to adaptively separate low-frequency global trends from high-frequency details, which helps the model better understand important patterns during the denoising process. Finally, our approach uses a specialized frequency encoder to integrate this information, enhancing the model's ability to capture both global and local features. Through exhaustive experiments on various public datasets, our model shows an impressive performance in generating time series data for diverse tasks like forecasting and imputation, outperforming existing methods in accuracy and flexibility.", "keywords": [ "Diffusion models", "Time series generation", "Power spectrum", "Fourier transformation" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "aIIYzzGKZp", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Benjamin Plaut", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Khanh Xuan Nguyen", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "Postdoc" }, { "name": "Tu Trinh", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 44, "n_ref": 68, "n_ref_all": 114, "n_fig": 6, "n_tab": 13, "L_tab": 3719, "n_element_tab": 747, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1634, "n_element_tab_1": 45, "formula_len_all": 41, "formula_len_all_1": 41, "len_all": 151631, "len_all_1": 63180, "len_abs": 1092, "len_title": 145, "len_sents": 35755, "len_sents_1": 28656, "n_sents": 306, "n_sents_1": 236, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1100, "title": "Probabilities of Chat LLMs Are Miscalibrated but Still Predict Correctness on Multiple-Choice Q&A", "abs": "We study 14 large language models (LLMs) fine-tuned for chat and find that their maximum softmax probabilities (MSPs) are consistently miscalibrated on multiple-choice Q\\&A. However, those MSPs might still encode useful uncertainty information. Specifically, we hypothesized that wrong answers would be associated with smaller MSPs compared to correct answers. Via rigororous statistical testing, we show that this hypothesis holds for models which perform well on the underlying Q\\&A task. We also find a strong direction correlation between Q\\&A accuracy and MSP correctness prediction, while finding no correlation between Q\\&A accuracy and calibration error. This suggests that within the current fine-tuning paradigm, we can expect correctness prediction but not calibration to improve as LLM capabilities progress. To demonstrate the utility of correctness prediction, we show that when models have the option to abstain, performance can be improved by selectively abstaining based on the MSP of the initial model response, using only a small amount of labeled data to choose the MSP threshold.", "keywords": [ "large language models", "uncertainty quantification", "calibration", "question-answering" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 1, 2, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "aIAFDFpNXz", "primary_area": "applications to robotics, autonomy, planning", "n_author": 28, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Bohan Zhou", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Boyu Li", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Ceyao Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Intern" }, { "name": "Chaojie Wang", "gender": "Male", "institution": "Skywork AI", "country": "SG", "position": "Researcher" }, { "name": "Chuqiao Zong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Haochong Xia", "gender": "Male", "institution": "National Technological University", "country": "SG", "position": "PhD student" }, { "name": "Jiechuan Jiang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Intern" }, { "name": "Junpeng Yue", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Long Tian", "gender": "Male", "institution": "Xi'an University of Software Engineering Institute", "country": "CN", "position": "Assistant Professor" }, { "name": "Longtao Zheng", "gender": "Male", "institution": "Nanyang Technological University, Singapore", "country": "SG", "position": "PhD student" }, { "name": "Pengjie Gu", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" }, { "name": "Ruyi An", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Tianbao Xie", "gender": "Male", "institution": "the University of Hong Kong, University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Weihao Tan", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Wentao Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "Xiaoqiang Chai", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "", "position": "MS student" }, { "name": "Xinrun Wang", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Xinrun Xu", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yewen Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Yifei Bi", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "YuJie Wu", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "", "position": "MS student" }, { "name": "Ziluo Ding", "gender": "Male", "institution": "BAAI", "country": "CN", "position": "Researcher" }, { "name": "Zongqing Lu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "B\u00f6rje Karlsson", "gender": "unknown", "institution": "Beijing Academy of Artificial Intelligence (BAAI)", "country": "CN", "position": "Principal Researcher" }, { "name": "Qin Molei", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xi yun Li", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 0, "n_ref_uni": 54, "n_ref": 221, "n_ref_all": 364, "n_fig": 97, "n_tab": 40, "L_tab": 13554, "n_element_tab": 768, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1085, "n_element_tab_1": 126, "formula_len_all": 474, "formula_len_all_1": 0, "len_all": 845988, "len_all_1": 91159, "len_abs": 1359, "len_title": 69, "len_sents": 333812, "len_sents_1": 46363, "n_sents": 3061, "n_sents_1": 324, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 2124, "title": "Cradle: Empowering Foundation Agents towards General Computer Control", "abs": "Despite their success in specific scenarios, existing foundation agents still struggle to generalize across various virtual scenarios, mainly due to the dramatically different encapsulations of environments with manually designed observation and action spaces. To handle this issue, we propose the General Computer Control (GCC) setting to restrict foundation agents to interact with software through the most unified and standardized interface, i.e., using screenshots as input and keyboard and mouse actions as output. We introduce Cradle, a modular and flexible LMM-powered framework, as a preliminary attempt towards GCC. Enhanced by six key modules, Information Gathering, Self-Reflection, Task Inference, Skill Curation, Action Planning, and Memory, Cradle is able to understand input screenshots and output executable code for low-level keyboard and mouse control after high-level planning and information retrieval, so that Cradle can interact with any software and complete long-horizon complex tasks without relying on any built-in APIs. Experimental results show that Cradle exhibits remarkable generalizability and impressive performance across four commercial never before explorer digital games, five software applications, and a comprehensive benchmark, OSWorld. To our best knowledge, Cradle is the first to enable foundation agents to follow the main storyline and complete one-hour-long real missions in the complex AAA game Red Dead Redemption 2 (RDR2). Cradle can also create a city of a thousand people in Cities: Skylines, farm and harvest parsnips in Stardew Valley, and trade and bargain with a maximal weekly total profit of 87% in Dealer's Life 2. Cradle can not only operate daily software, like Chrome, Outlook, and Feishu, but also edit images and videos using Meitu and CapCut. With a unified interface to interact with any software, Cradle greatly extends the reach of foundation agents by enabling the easy conversion of any software, especially complex games, into benchmarks to evaluate agents' various abilities and further collect detailed data, thus paving the way for generalist agents.", "keywords": [ "Foundation Agents", "Large Multimodal Models", "Decision-making", "General Computer Control" ], "rating_list": [ 8, 8, 5, 5 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 1, 2 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "aHSoKImhc2", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Antonio Orvieto", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen, Max Planck Institute for Intelligent Systems, T\u00fcbingen AI Center, T\u00fcbingen, Germany", "country": "CH", "position": "Principal Researcher" }, { "name": "Destiny Okpekpe", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen", "country": "DE", "position": "Intern" }, { "name": "Nursena Koprucu", "gender": "Female", "institution": "Max-Planck Institute", "country": "DE", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 63, "n_ref": 118, "n_ref_all": 138, "n_fig": 3, "n_tab": 9, "L_tab": 1764, "n_element_tab": 176, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2772, "n_element_tab_1": 144, "formula_len_all": 1295, "formula_len_all_1": 1211, "len_all": 166739, "len_all_1": 70250, "len_abs": 1478, "len_title": 74, "len_sents": 38248, "len_sents_1": 30491, "n_sents": 302, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1483, "title": "NIMBA : Towards Robust and Principled Processing of Point Clouds With SSMs", "abs": "Transformers have become dominant in large-scale deep learning tasks across various domains, including text, 2D and 3D vision. However, the quadratic complexity of their attention mechanism limits their efficiency as the sequence length increases, particularly in high-resolution 3D data such as point clouds. Recently, state space models (SSMs) like Mamba have emerged as promising alternatives, offering linear complexity, scalability, and high performance in long-sequence tasks. The key challenge in the application of SSMs in this domain lies in reconciling the non-sequential structure of point clouds with the inherently directional (or bi-directional) order-dependent processing of recurrent models like Mamba. To achieve this, previous research proposed reorganizing point clouds along multiple directions or predetermined paths in 3D space, concatenating the results to produce a single 1D sequence capturing different views. In our work we introduce a method to convert point clouds into 1D sequences that maintains 3D spatial structure with no need for data replication, allowing Mamba\u2019s sequential processing to be applied effectively in an almost permutation-invariant manner. In contrast to other works, we found that our method does not require positional embeddings, and allows for shorter sequence lengths while still achieving state-of-the-art results in ModelNet40 and ScanObjectNN datasets and surpassing Transformer-based models in both accuracy and efficiency.", "keywords": [ "point clouds", "SSMs", "state-space models", "Mamba", "3D", "transformers", "pointcloud" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 5 ] }, { "paper_id": "aGBA8wz9qA", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Henry Han", "gender": "Not Specified", "institution": "Baylor University", "country": "", "position": "Full Professor" }, { "name": "Jie Ren", "gender": "Female", "institution": "Baylor University", "country": "", "position": "PhD student" } ], "n_formula": 87, "n_formula_1": 20, "n_ref_uni": 32, "n_ref": 48, "n_ref_all": 62, "n_fig": 8, "n_tab": 7, "L_tab": 2916, "n_element_tab": 327, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 315, "n_element_tab_1": 14, "formula_len_all": 4661, "formula_len_all_1": 1163, "len_all": 192646, "len_all_1": 64202, "len_abs": 2399, "len_title": 165, "len_sents": 59377, "len_sents_1": 31442, "n_sents": 472, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 117, "L_abs": 1285, "title": "Debiased Contrastive Learning with multi-resolution Kolmogorov-Arnold Network for Gravitational Wave Glitch Detection", "abs": "Time-series gravitational wave glitch detection presents significant challenges for machine learning due to the complexity of the data, limited labeled examples, and data imbalance. To address these issues, we introduce Debiased Contrastive Learning with Multi-Resolution Kolmogorov-Arnold Network(dcMltR-KAN), a novel self-supervised learning (SSL) approach that enhances glitch detection, robustness, explainability, and generalization. dcMltR-KAN consists of three key novel components: Wasserstein Debiased Contrastive Learning (wDCL), a CNN-based encoder, and a Multi-Resolution KAN (MltR-KAN). The wDCL improves the model\u2019s sensitivity to data imbalance and geometric structure. The CNN-based encoder eliminates false negatives during training, refines feature representations through similarity-based weighting (SBW), and reduces data complexity within the embedding. Additionally, MltR-KAN enhances explainability, generalization, and efficiency by adaptively learning parameters. Our model outperforms widely used baselines on O1, O2, and O3 data, demonstrating its effectiveness. Extending dcMltR-KAN to other time-series benchmarks underscores its novelty and efficiency, marking it as the first model of its kind and paving the way for future SSL and astrophysics research.", "keywords": [ "self-supervised learning", "Debiased Contrastive Learning", "Gravitational Wave", "Glitch detection", "deep learning" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "aF1jasJeRy", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Aristide Baratin", "gender": "unknown", "institution": "Samsung - SAIT AI Lab, Montreal", "country": "", "position": "Research scientist" }, { "name": "Gintare Karolina Dziugaite", "gender": "Female", "institution": "McGill University", "country": "CA", "position": "Adjunct Professor" }, { "name": "Goncalo Mordido", "gender": "Male", "institution": "Mila", "country": "", "position": "Postdoc" }, { "name": "Pranshu Malviya", "gender": "Male", "institution": "\u00c9cole Polytechnique de Montr\u00e9al, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "Razvan Pascanu", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Affiliate Member" }, { "name": "Reza Babanezhad Harikandeh", "gender": "Male", "institution": "Samsung", "country": "", "position": "Research Scientist" }, { "name": "Sarath Chandar", "gender": "Male", "institution": "\u00c9cole Polytechnique de Montr\u00e9al", "country": "CA", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 41, "n_ref": 57, "n_ref_all": 88, "n_fig": 10, "n_tab": 17, "L_tab": 17057, "n_element_tab": 2239, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 380, "n_element_tab_1": 40, "formula_len_all": 776, "formula_len_all_1": 808, "len_all": 171338, "len_all_1": 60950, "len_abs": 808, "len_title": 69, "len_sents": 34366, "len_sents_1": 25689, "n_sents": 256, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 21, "L_abs": 812, "title": "Torque-Aware Momentum", "abs": "Efficiently exploring complex loss landscapes is key to the performance of deep neural networks. While momentum-based optimizers are widely used in state-of-the-art setups, classical momentum can still struggle with large, misaligned gradients, leading to oscillations. To address this, we propose Torque-Aware Momentum (TAM), which introduces a damping factor based on the angle between the new gradients and previous momentum, stabilizing the update direction during training. Empirical results show that TAM, which can be combined with both SGD and Adam, enhances exploration, handles distribution shifts more effectively, and improves generalization performance across various tasks, including image classification and large language model fine-tuning, when compared to classical momentum-based optimizers.", "keywords": [ "optimization", "deep learning", "momentum" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "aEJ5ilIuwg", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Minh-Tuan Tran", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Trung Le", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" }, { "name": "Tuan Truong", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Undergrad student" }, { "name": "Tung Pham", "gender": "Male", "institution": "VinAI Research", "country": "VN", "position": "Researcher" }, { "name": "Dinh Phung", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" }, { "name": "Phi Hoang Nguyen", "gender": "Male", "institution": "VinAI Research", "country": "VN", "position": "Researcher" }, { "name": "Mehrtash T. Harandi", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 35, "n_formula_1": 22, "n_ref_uni": 37, "n_ref": 82, "n_ref_all": 101, "n_fig": 5, "n_tab": 5, "L_tab": 1647, "n_element_tab": 149, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1093, "n_element_tab_1": 74, "formula_len_all": 3668, "formula_len_all_1": 2407, "len_all": 147475, "len_all_1": 65243, "len_abs": 1010, "len_title": 113, "len_sents": 43446, "len_sents_1": 28348, "n_sents": 304, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1013, "title": "Optimization on Manifolds with Riemannian Jacobian Regularization", "abs": "Understanding the effectiveness of intrinsic geometry in enhancing a model's generalization ability, we draw upon prior works that apply geometric principles to optimization and present a novel approach to improve robustness and generalization for constrained optimization problems. This work aims to strengthen the sharpness-aware optimizers and proposes a novel Riemannian optimizer. We first present a theoretical analysis that characterizes the relationship between the general loss and the perturbation of the empirical loss in the context of Riemannian manifolds. Motivated by the result obtained from this analysis, we introduce our algorithm named Riemannian Jacobian Regularization (RJR), which explicitly regularizes the Riemannian gradient norm and the projected Hessian. To demonstrate RJR's ability to enhance generalization, we evaluate and contrast our algorithm on a broad set of problems, such as image classification and contrastive learning across different datasets with various architectures.", "keywords": [ "Riemannian manifolds", "Flat minimizer", "Sharpness-aware Minimization" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "aE6QjMJ1mN", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Adrians Skapars", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "PhD student" }, { "name": "Alessandra Russo", "gender": "Female", "institution": "Imperial College London", "country": "GB", "position": "Full Professor" }, { "name": "Alexander F Spies", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "PhD student" }, { "name": "Katsumi Inoue", "gender": "unknown", "institution": "National Institute of Informatics", "country": "JP", "position": "Full Professor" }, { "name": "Michael Ivanitskiy", "gender": "Male", "institution": "Colorado School of Mines", "country": "", "position": "PhD student" }, { "name": "Murray Shanahan", "gender": "Male", "institution": "Google", "country": "", "position": "Senior research scientist" }, { "name": "Tilman R\u00e4uker", "gender": "Male", "institution": "Pivotal Research ", "country": "DE", "position": "Researcher" }, { "name": "William Edwards", "gender": "Male", "institution": "UnSearch", "country": "", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 28, "n_ref": 41, "n_ref_all": 80, "n_fig": 20, "n_tab": 2, "L_tab": 174, "n_element_tab": 16, "n_fig_1": 15, "n_tab_1": 2, "L_tab_1": 71, "n_element_tab_1": 14, "formula_len_all": 2754, "formula_len_all_1": 2116, "len_all": 96783, "len_all_1": 60459, "len_abs": 1394, "len_title": 58, "len_sents": 36033, "len_sents_1": 30402, "n_sents": 219, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1401, "title": "Transformers Use Causal World Models in Maze-Solving Tasks", "abs": "Recent studies in interpretability have explored the inner workings of transformer models trained on tasks across various domains, often discovering that these networks naturally develop surprisingly structured representations. When such representations comprehensively reflect the task domain's structure, they are commonly referred to as ``World Models'' (WMs). In this work, we discover such WMs in transformers trained on maze tasks. In particular, by employing Sparse Autoencoders (SAEs) and analysing attention patterns, we examine the construction of WMs and demonstrate consistency between the circuit analysis and the SAE feature-based analysis. We intervene upon the isolated features to confirm their causal role and, in doing so, find asymmetries between certain types of interventions. Surprisingly, we find that models are able to reason with respect to a greater number of active features than they see during training, even if attempting to specify these in the input token sequence would lead the model to fail. Futhermore, we observe that varying positional encodings can alter how WMs are encoded in a model's residual stream. By analyzing the causal role of these WMs in a toy domain we hope to make progress toward an understanding of emergent structure in the representations acquired by Transformers, leading to the development of more interpretable and controllable AI systems.", "keywords": [ "mechanistic", "interpretability", "world models", "transformers" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "aE0XTpl6oM", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Peihua Mai", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yan Pang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Youlong Ding", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "PhD student" }, { "name": "Ziyan Lyu", "gender": "unknown", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 20, "n_formula_1": 10, "n_ref_uni": 30, "n_ref": 42, "n_ref_all": 69, "n_fig": 5, "n_tab": 9, "L_tab": 4347, "n_element_tab": 566, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1378, "n_element_tab_1": 177, "formula_len_all": 1573, "formula_len_all_1": 682, "len_all": 177200, "len_all_1": 65703, "len_abs": 1519, "len_title": 143, "len_sents": 49321, "len_sents_1": 28091, "n_sents": 403, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1526, "title": "GREC: Doubly Efficient Privacy-preserving Recommender Systems for Resource-Constrained Devices", "abs": "Federated recommender system (FedRec) has emerged as a solution to protect user data through collaborative training techniques. However, the real-world implementation of FedRec is hindered by two critical resource constraints of edge devices: a) limited upload bandwidth and b) limited user computational power and storage. Existing methods addressing the first issue, such as message compression techniques, often result in accuracy degradation or potential privacy leakage. For the second issue, most federated learning (FL) protocols assume that users must store and maintain the entire model locally for private inference, which is resource intensive. To address these challenges, we propose doubly efficient privacy-perserving recommender systems (GREC) consisting of both training and inference phase. To reduce communication costs during the training phase, we design a lossless secure aggregation (SecAgg) protocol based on functional secret sharing leveraging the sparsity of the update matrix. During the inference phase, we implement a user-side post-processing local differential privacy (LDP) algorithm to ensure privacy while shifting the bulk of computation to the cloud. Our framework reduces uplink communication costs by up to 90x compared to existing SecAgg protocols and decreases user-side computation time during inference by an average of 11x compared to full-model inference. This makes GREC a practical and scalable solution for deploying federated recommender systems on resource-constrained devices.", "keywords": [ "Federated learning", "recommender system", "secure aggregation" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 2, 3, 4, 2 ] }, { "paper_id": "aDG34Bhbs1", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Andrey Ploskonosov", "gender": "unknown", "institution": "Higher School of Economics", "country": "", "position": "MS student" }, { "name": "Kirill Sergeevich Shevkunov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "", "position": "MS student" }, { "name": "Liudmila Prokhorenkova", "gender": "Female", "institution": "Yandex", "country": "", "position": "Researcher" } ], "n_formula": 61, "n_formula_1": 11, "n_ref_uni": 35, "n_ref": 66, "n_ref_all": 93, "n_fig": 1, "n_tab": 13, "L_tab": 4516, "n_element_tab": 657, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2473, "n_element_tab_1": 351, "formula_len_all": 5917, "formula_len_all_1": 791, "len_all": 178398, "len_all_1": 74647, "len_abs": 1385, "len_title": 108, "len_sents": 57131, "len_sents_1": 33494, "n_sents": 494, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1393, "title": "Relevance-Based Embeddings for Efficient Relevance Retrieval", "abs": "In many machine learning applications, the most relevant items for a particular query should be efficiently extracted. The relevance function is usually an expensive similarity model making the exhaustive search infeasible. A typical solution to this problem is to train another model that separately embeds queries and items to a vector space, where similarity is defined via the dot product or cosine similarity. This allows one to search the most relevant objects through fast approximate nearest neighbors search at the cost of some reduction in quality. To compensate for this reduction, the found candidates are re-ranked by the expensive similarity model. In this paper, we investigate an alternative approach that utilizes the relevances of the expensive model to make relevance-based embeddings (RBE). The idea is to describe each query (item) by its relevance for a set of support items (queries) and use these new representations to obtain query (item) embeddings. We theoretically prove that relevance-based embeddings are powerful enough to approximate any complex similarity model (under mild conditions). An important ingredient of RBE is the choice of support items. We investigate several strategies and demonstrate that significant improvements can be obtained compared to random choice. Our experiments on diverse datasets illustrate the power of relevance-based embeddings.", "keywords": [ "Information search", "Relevance search", "Nearest neighbor search", "Relevance-based embeddings", "Recommendation systems" ], "rating_list": [ 3, 5, 8, 5, 3 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 1, 3, 4, 1, 3 ], "contribution_list": [ 2, 2, 3, 2, 1 ], "confidence_list": [ 4, 3, 3, 3, 4 ] }, { "paper_id": "aD2uwhLbnA", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Bingrui Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Mingze Wang", "gender": "unknown", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yuchen Mao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhanpeng Zhou", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 59, "n_formula_1": 16, "n_ref_uni": 63, "n_ref": 137, "n_ref_all": 195, "n_fig": 17, "n_tab": 11, "L_tab": 6257, "n_element_tab": 244, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 754, "n_element_tab_1": 38, "formula_len_all": 5318, "formula_len_all_1": 943, "len_all": 252440, "len_all_1": 76608, "len_abs": 1334, "len_title": 129, "len_sents": 64702, "len_sents_1": 31377, "n_sents": 575, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 80, "L_abs": 1338, "title": "Sharpness-Aware Minimization Efficiently Selects Flatter Minima Late In Training", "abs": "Sharpness-Aware Minimization (SAM) has substantially improved the generalization of neural networks under various settings.\nDespite the success, its effectiveness remains poorly understood.\nIn this work, we discover an intriguing phenomenon in the training dynamics of SAM, shedding lights on understanding its implicit bias towards flatter minima over Stochastic Gradient Descent (SGD).\nSpecifically, we find that *SAM efficiently selects flatter minima late in training*.\nRemarkably, even a few epochs of SAM applied at the end of training yield nearly the same generalization and solution sharpness as full SAM training.\nSubsequently, we delve deeper into the underlying mechanism behind this phenomenon.\nTheoretically, we identify two phases in the learning dynamics after applying SAM late in training: i) SAM first escapes the minimum found by SGD exponentially fast; and ii) then rapidly converges to a flatter minimum within the same valley.\nFurthermore, we empirically investigate the role of SAM during the early training phase.\nWe conjecture that the optimization method chosen in the late phase is more crucial in shaping the final solution's properties.\nBased on this viewpoint, we extend our findings from SAM to Adversarial Training.\nWe provide source code in supplementary materials and will release checkpoints in future.", "keywords": [ "Sharpness-Aware Minimization", "Implicit Bias", "Training Dynamics" ], "rating_list": [ 8, 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 2, 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 4, 3 ], "confidence_list": [ 3, 3, 3, 3, 3 ] }, { "paper_id": "aCz7TiKjwJ", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Sanglu Lu", "gender": "Female", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Wenzhong Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Xiangkai Ma", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "PhD student" }, { "name": "Xiaobin Hong", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 13, "n_ref_uni": 54, "n_ref": 103, "n_ref_all": 143, "n_fig": 13, "n_tab": 9, "L_tab": 8191, "n_element_tab": 231, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2834, "n_element_tab_1": 67, "formula_len_all": 1998, "formula_len_all_1": 1225, "len_all": 185781, "len_all_1": 69143, "len_abs": 2278, "len_title": 89, "len_sents": 60184, "len_sents_1": 32468, "n_sents": 359, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 2060, "title": "UTSD: Unified Time Series Diffusion Model", "abs": "Transformer-based architectures have achieved unprecedented success in time series analysis. However, facing the challenge of across-domain modeling, existing studies utilize statistical prior as prompt engineering fails under the huge distribution shift among various domains. In this paper, a Unified Time Series Diffusion (UTSD) model is established for the first time to model the multi-domain probability distribution, utilizing the powerful probability distribution modeling ability of Diffusion. Unlike the autoregressive models that capture the conditional probabilities of the prediction horizon to the historical sequence, we use a diffusion denoising process to model the mixture distribution of the cross-domain data and generate the prediction sequence for the target domain directly utilizing conditional sampling. The proposed UTSD contains three pivotal designs: (1) The condition network captures the multi-scale fluctuation patterns from the observation sequence, which are utilized as context representations to guide the denoising network to generate the prediction sequence; (2) Adaptor-based fine-tuning strategy, the multi-domain universal representation learned in the pretraining stage is utilized for downstream tasks in target domains; (3) The diffusion and denoising process on the actual sequence space, combined with the improved classifier free guidance as the conditional generation strategy, greatly improves the stability and accuracy of the downstream task. We conduct extensive experiments on mainstream benchmarks, and the pre-trained UTSD outperforms existing foundation models on all data domains, exhibiting superior zero-shot generalization ability. After training from scratch, UTSD achieves comparable performance against domain-specific proprietary models. In particular, UTSD shows stable and reliable time series generation, and the empirical results validate the potential of UTSD as a time series foundational model. The source codes of UTSD are publicly available on https://anonymous.4open.science/r/UTSD-1BFF.", "keywords": [ "time series anlysis", "foundation model", "cross-doman", "diffusition model" ], "rating_list": [ 5, 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 2, 2 ], "contribution_list": [ 1, 3, 2, 2, 1 ], "confidence_list": [ 3, 5, 5, 3, 2 ] }, { "paper_id": "aClIuYLG47", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Shunan Mao", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yaowei Wang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Zhanzhou FENG", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Shiliang Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 22, "n_ref_uni": 21, "n_ref": 69, "n_ref_all": 84, "n_fig": 3, "n_tab": 9, "L_tab": 1210, "n_element_tab": 184, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 3532, "n_element_tab_1": 102, "formula_len_all": 1193, "formula_len_all_1": 1066, "len_all": 111737, "len_all_1": 64129, "len_abs": 1256, "len_title": 106, "len_sents": 38145, "len_sents_1": 27503, "n_sents": 311, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1263, "title": "VideoUntier: Language-guided Video Feature Disentanglement", "abs": "Most of existing text-video retrieval works learn features comprehensively representing complicated video contents. This leads to the difficulty of textual-visual feature alignment, because text queries convey more concise cues like certain objects and events the user desires to retrieve. To pursue a more compact video representation and accurate textual-visual feature matching, this paper introduces a novel VideoUntier to disentangle video features. VideoUntier first generates 'object' and 'event' tokens from query texts. It subsequently spots and merges visual tokens related to concepts in the query. In other words, we use 'object' and 'event' tokens to represent cues of query, which therefore supervise the disentanglement and extraction of meaningful visual features from videos. VideoUntier finally leads to compact visual tokens explicitly depicting query objects and events. Extensive experiments on three widely-used datasets demonstrate the promising performance and domain generalization capability of our method. For instance, our method shows better efficiency and consistently outperforms many recent works like ProST on three datasets. We hope to inspire future work for collaborative cross-modal learning with certain modality as guidance.", "keywords": [ "representation learning", "cross-modal learning", "video recognition" ], "rating_list": [ 5, 5, 1, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 5 ] }, { "paper_id": "aCPFCDL9QY", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Julian McAuley", "gender": "Male", "institution": "University of California, San Diego, University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Junda Wu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Sean O'Brien", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Xinshuang Liu", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Xiusi Chen", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Postdoc" }, { "name": "Yu Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 39, "n_ref": 69, "n_ref_all": 84, "n_fig": 2, "n_tab": 5, "L_tab": 715, "n_element_tab": 57, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 817, "n_element_tab_1": 59, "formula_len_all": 261, "formula_len_all_1": 261, "len_all": 142933, "len_all_1": 74474, "len_abs": 1706, "len_title": 111, "len_sents": 42758, "len_sents_1": 37063, "n_sents": 289, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1727, "title": "Self-Updatable Large Language Models with Parameter Integration", "abs": "Despite significant advancements in large language models (LLMs), the rapid and frequent integration of small-scale experiences, such as interactions with sur- rounding objects, remains a substantial challenge. Two critical factors in assimilating these experiences are (1) **Efficacy**: the ability to accurately remember recent events; (2) **Retention**: the capacity to recall long-past experiences. Current methods either embed experiences within model parameters using continual learning, model editing, or knowledge distillation techniques, which often struggle with rapid updates and complex interactions, or rely on external storage to achieve long-term retention, thereby increasing storage requirements. In this paper, we propose **SELF-PARAM** (Self-Updatable Large Language Models with Parameter Integration). SELF-PARAM requires no extra parameters while ensuring near-optimal efficacy and long-term retention. Our method employs a training objective that minimizes the Kullback-Leibler (KL) divergence between the predictions of an original model (with access to contextual information) and a target model (without such access). By generating diverse question-answer pairs related to the knowledge and minimizing the KL divergence across this dataset, we update the target model to internalize the knowledge seamlessly within its parameters. Evaluations on question-answering and conversational recommendation tasks demonstrate that SELF-PARAM significantly outperforms existing methods, even when accounting for non-zero storage requirements. This advancement paves the way for more efficient and scalable integration of experiences in large language models by embedding knowledge directly into model parameters.", "keywords": [ "Large Language Models", "Knowledge Injection", "Self-Updatable LLMs" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "aCEg0zZ2bG", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Chen Zheng", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Heng Zhang", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Kaifan Jia", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Xuran Wang", "gender": "unknown", "institution": "Northeastern University", "country": "CN", "position": "Undergrad student" }, { "name": "Yongchun Jiang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 18, "n_formula_1": 6, "n_ref_uni": 25, "n_ref": 58, "n_ref_all": 78, "n_fig": 4, "n_tab": 10, "L_tab": 5243, "n_element_tab": 440, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1313, "n_element_tab_1": 109, "formula_len_all": 979, "formula_len_all_1": 416, "len_all": 161917, "len_all_1": 63375, "len_abs": 1650, "len_title": 139, "len_sents": 55369, "len_sents_1": 30274, "n_sents": 298, "n_sents_1": 168, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1666, "title": "SPACETGN: Augmented Mini-Batch Negative Sampling for Continuous-Time Dynamic Graph Learning", "abs": "Continuous-Time Dynamic Graph (CTDG) learning has significantly advanced link prediction performance by leveraging random negative sampling and incorporating adaptive temporal information.\nRecent studies aim to improve performance by introducing random sampling to obtain hard negative samples, whose quality is limited by randomness, capturing few categories of negative samples, and leading to false positive (FP) and false negative (FN) problems.\nHere we present SPACETGN, a CTDG learning framework, with a augmented hard negative sampling mini-batches (AMNS) strategy and two new feature extraction strategies that derive space-temporal locality subgraph and historical occurrence information to emphasize the graph's temporal discriminative properties. \nThe AMNS strategy sample mini-batches comprised of instances that are hard-to-distinguish (i.e., hard and true negatives with respect to each other) based on the target distribution, thereby effectively augmenting the discriminative features and the diversity of historical and inductive samples.\nFurthermore, to mitigate the challenges posed by false positives (FP) and false negatives (FN), our architecture SPACETGN employs a conceptually straightforward approach that investigates temporal subgraphs and historical interactions between source and destination nodes. This enables the model to leverage complex and historically accurate interactions among predicted entities.\nOur extensive evaluation of dynamic link prediction on seven state-of-the-practice datasets reveals that SPACETGN achieves state-of-the-art performance in most datasets, demonstrating its effectiveness in ameliorating model bias.", "keywords": [ "Dynamic Graph Learning;Negative Sampling Strategy" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 4, 1 ], "presentation_list": [ 2, 1, 1, 1 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "aBxxFLqtHa", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Bijun Li", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Fuxin Xie", "gender": "unknown", "institution": "Wuhan University", "country": "CN", "position": "Undergrad student" }, { "name": "HaoMiao Bian", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Undergrad student" }, { "name": "Jian Zhou", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Associate Professor" }, { "name": "Jinsheng Xiao", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 18, "n_ref": 29, "n_ref_all": 34, "n_fig": 4, "n_tab": 6, "L_tab": 2137, "n_element_tab": 297, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1713, "n_element_tab_1": 250, "formula_len_all": 633, "formula_len_all_1": 416, "len_all": 102993, "len_all_1": 55340, "len_abs": 1713, "len_title": 117, "len_sents": 33989, "len_sents_1": 24513, "n_sents": 227, "n_sents_1": 162, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1719, "title": "pDETR: End-to-End Object Detection via Perspective-Aware Transformers", "abs": "DETR has made notable performance improvements in object detection tasks by leveraging the long-range modeling capabilities of Transformers, but encoding all tokens indiscriminately significantly escalates computational cost and leads to slow convergence. Recent sparsification strategies effectively reduce computational cost through sparse encoders. However, these methods rely heavily on a fixed sparse ratio, which overlooks the coherence of feature representation across levels, leading to performance degradation in complex scenes. To address this issue, we propose a novel object detection approach aimed at constructing consistent representations of multi-level features. The approach composes two steps: First, we introduce a perspective proposal module that leverages the spatial information of high-level foreground features to guide the sparse sampling of low-level features, ensuring both integrity and coherence of multi-scale feature information. Furthermore, we integrated semantic probability to perform hierarchical and dynamic adjustments to the saliency of queries, thereby refining the semantic interaction among foreground queries. Experimental results demonstrate that on the challenging task-specific VisDrone dataset, our pDETR method enhances AP by 1.8% compared to DINO. On the COCO 2017 dataset, the performance improvement of pDETR is even more apparent, achieving a +2.5% increase in AP: under the 1\u00d7 schedule, pDETR attains an AP of 51.5%, and under the 2\u00d7 schedule, the AP further increases to 52.0%. Moreover, it exhibits faster convergence, exceeding 40% AP in just 2 training epochs while reducing computational cost by 13% in terms of FLOPs, indicating superior detection capability.", "keywords": [ "Multi-Scale", "Sparse", "Transformer", "Object Detection" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "aBnVU5DL3I", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Gang Yu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Software Engineer" }, { "name": "Huajun Chen", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Jing Yu", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Kehua Feng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Keyan Ding", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Qiang Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yiwen Qu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zhiwen Chen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "chengfei lv", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Software Engineer" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 20, "n_ref": 39, "n_ref_all": 59, "n_fig": 4, "n_tab": 8, "L_tab": 11095, "n_element_tab": 580, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 4639, "n_element_tab_1": 377, "formula_len_all": 498, "formula_len_all_1": 423, "len_all": 179955, "len_all_1": 62622, "len_abs": 1410, "len_title": 125, "len_sents": 42154, "len_sents_1": 24209, "n_sents": 251, "n_sents_1": 155, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1420, "title": "SaMer: A Scenario-aware Multi-dimensional Evaluator for Large Language Models", "abs": "Evaluating the response quality of large language models (LLMs) for open-ended questions poses a significant challenge, especially given the subjectivity and multi-dimensionality of \"quality\" in natural language generation. Existing LLM evaluators often neglect that different scenarios require distinct evaluation criteria. In this work, we propose **SaMer**, a scenario-aware multi-dimensional evaluator designed to provide both overall and fine-grained assessments of LLM-generated responses. Unlike fixed-dimension evaluation approaches, SaMer adapts to different scenarios by automatically identifying and prioritizing relevant evaluation dimensions tailored to the given query. To achieve this, we construct a large-scale fine-grained preference dataset spanning multiple real-world scenarios, each with distinct evaluation dimensions. We then leverage a text embedding model combined with three specialized heads to predict the appropriate evaluation dimensions and corresponding scores, as well as the respective weights that contribute to the overall score. The resulting model offers fine-grained and interpretable evaluations and shows robust adaptability across diverse scenarios. Extensive experiments on eight single rating and pairwise comparison datasets demonstrate that SaMer outperforms existing baselines in a variety of evaluation tasks, showcasing its robustness, versatility, and generalizability.", "keywords": [ "Fine-grained Evaluation", "Adaptive Multi-dimensional Evaluator", "Large Language Models" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "aAxzDb0nlO", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Claudia Clopath", "gender": "unknown", "institution": "Imperial College London", "country": "", "position": "Full Professor" }, { "name": "Rodrigo Antonio Carrasco-Davis", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Sebastian Lee", "gender": "unknown", "institution": "Imperial College London, Imperial College London", "country": "", "position": "PhD student" }, { "name": "Will Dabney", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" } ], "n_formula": 46, "n_formula_1": 17, "n_ref_uni": 82, "n_ref": 131, "n_ref_all": 188, "n_fig": 21, "n_tab": 1, "L_tab": 342, "n_element_tab": 13, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2888, "formula_len_all_1": 798, "len_all": 241122, "len_all_1": 74224, "len_abs": 1341, "len_title": 89, "len_sents": 68885, "len_sents_1": 35785, "n_sents": 449, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1348, "title": "Uncertainty Prioritized Experience Replay", "abs": "Prioritized experience replay, which improves sample efficiency by selecting relevant transitions to update parameter estimates, is a crucial component of contemporary value-based deep reinforcement learning models. Typically, transitions are prioritized based on their temporal difference error. However, this approach is prone to favoring noisy transitions, even when the value estimation closely approximates the target mean. This phenomenon resembles the _noisy TV_ problem postulated in the exploration literature, in which exploration-guided agents get stuck by mistaking noise for novelty. To mitigate the disruptive effects of noise in value estimation, we propose using epistemic uncertainty to guide the prioritization of transitions from the replay buffer. Epistemic uncertainty quantifies the uncertainty that can be reduced by learning, hence reducing transitions sampled from the buffer generated by unpredictable random processes. We first illustrate the benefits of epistemic uncertainty prioritized replay in two tabular toy models: a simple multi-arm bandit task, and a noisy gridworld. Subsequently, we evaluate our prioritization scheme on the Atari suite, outperforming quantile regression deep Q-learning benchmarks; thus forging a path for the use of epistemic uncertainty prioritized replay in reinforcement learning agents.", "keywords": [ "reinforcement learning", "replay", "uncertainty" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "aAcOaJYbUg", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Fanfei Li", "gender": "Female", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "Intern" }, { "name": "Robert Geirhos", "gender": "Male", "institution": "Google DeepMind", "country": "CA", "position": "Research Scientist" }, { "name": "Thomas Klein", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "", "position": "PhD student" }, { "name": "Wieland Brendel", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen", "country": "DE", "position": "Principal Researcher" }, { "name": "Roland Simon Zimmermann", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 70, "n_ref_all": 90, "n_fig": 10, "n_tab": 5, "L_tab": 2099, "n_element_tab": 599, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1253, "n_element_tab_1": 221, "formula_len_all": 90, "formula_len_all_1": 0, "len_all": 208906, "len_all_1": 58219, "len_abs": 1598, "len_title": 117, "len_sents": 71910, "len_sents_1": 27726, "n_sents": 568, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1604, "title": "LAION-C: An out-of-distribution benchmark for web-scale vision models", "abs": "Out-of-distribution (OOD) robustness is a desired property of computer vision models. Improving model robustness requires high-quality signals from robustness benchmarks to quantify progress. While various benchmark datasets such as ImageNet-C were proposed in the ImageNet era, most ImageNet-C corruption types are no longer OOD relative to today's large datasets scraped from the web, which already contain common corruptions such as blur or JPEG compression artifacts. Consequently, these standard benchmarks are no longer well-suited for evaluating OOD robustness in the era of web-scale datasets. Indeed, recent models show saturating scores on ImageNet-era OOD benchmarks, indicating that it is unclear whether models trained on web-scale datasets truly become better at OOD generalization or whether they have simply been exposed to the test distortions during training. To address this, we here introduce LAION-C as a benchmark alternative for ImageNet-C. LAION-C consists of six novel distortion types across five severity levels designed to be OOD, even for web-scale datasets such as LAION. In a comprehensive evaluation of state-of-the-art models, we find that the LAION-C dataset poses significant challenges to contemporary models. We additionally conducted a psychophysical experiment to evaluate the difficulty of our proposed corruptions for human observers, enabling a comparison of models to lab-quality human robustness data. We observe a paradigm shift in OOD generalization: from humans outperforming models to the best models now matching or outperforming the best human observers.", "keywords": [ "OOD", "representation learning", "benchmark", "model evaluation", "vision", "classification" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "aAI92OHA4t", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Casey Lauer", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Jonathan B Freund", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Robert C. Blake III", "gender": "Male", "institution": "Lawrence Livermore National Labs", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 25, "n_ref": 33, "n_ref_all": 39, "n_fig": 3, "n_tab": 1, "L_tab": 219, "n_element_tab": 14, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 219, "n_element_tab_1": 14, "formula_len_all": 350, "formula_len_all_1": 350, "len_all": 68878, "len_all_1": 68878, "len_abs": 1363, "len_title": 121, "len_sents": 22234, "len_sents_1": 22234, "n_sents": 143, "n_sents_1": 143, "venue": "ICLR2025", "decision": "withdraw", "L_title": 121, "L_abs": 1367, "title": "Soft Checksums to Flag Untrustworthy Machine Learning Surrogate Predictions and Application to Atomic Physics Simulations", "abs": "Trained neural networks (NN) are attractive as surrogate models to replace costly calculations in physical simulations, but are often unknowingly applied to states not adequately represented in the training dataset. We present the novel technique of soft checksums for scientific machine learning, a general-purpose method to differentiate between trustworthy predictions with small errors on in-distribution (ID) data points, and untrustworthy predictions with large errors on out-of-distribution (OOD) data points. By adding a check node to the existing output layer, we train the model to learn the chosen checksum function encoded within the NN predictions and show that violations of this function correlate with high prediction errors. As the checksum function depends only on the NN predictions, we can calculate the checksum error for any prediction with a single forward pass, incurring negligible time and memory costs. Additionally, we find that incorporating the checksum function into the loss function and exposing the NN to OOD data points during the training process improves separation between ID and OOD predictions. By applying soft checksums to a physically complex and high-dimensional non-local thermodynamic equilibrium atomic physics dataset, we show that a well-chosen threshold checksum error can effectively separate ID and OOD predictions.", "keywords": [ "deep neural networks", "surrogate models", "out-of-distribution detection", "non-local thermodynamic equilibrium" ], "rating_list": [ 3, 3, 1 ], "soundness_list": [ 2, 1, 1 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "a9vey6B54y", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Chengxin Wang", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Gary Tan", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Associate Professor" }, { "name": "Shaofeng Cai", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Yiran Zhao", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 52, "n_ref": 107, "n_ref_all": 116, "n_fig": 4, "n_tab": 6, "L_tab": 3680, "n_element_tab": 164, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 3509, "n_element_tab_1": 114, "formula_len_all": 581, "formula_len_all_1": 581, "len_all": 154446, "len_all_1": 57895, "len_abs": 1023, "len_title": 110, "len_sents": 40990, "len_sents_1": 26055, "n_sents": 272, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1104, "title": "Investigating Pattern Neurons in Urban Time Series Forecasting", "abs": "Urban time series forecasting is crucial for smart city development and is key to sustainable urban management. Although urban time series models (UTSMs) are effective in general forecasting, they often overlook low-frequency events, such as emergencies and holidays, leading to degraded performance in practical applications. In this paper, we first investigate how UTSMs handle these infrequent patterns from a neural perspective. Based on our findings, we propose $\\textbf{P}$attern $\\textbf{N}$euron guided $\\textbf{Train}$ing ($\\texttt{PN-Train}$), a novel training method that features (i) a $\\textit{perturbation-based detector}$ to identify neurons responsible for low-frequency patterns in UTSMs, and (ii) a $\\textit{fine-tuning mechanism}$ that enhances these neurons without compromising representation learning on high-frequency patterns. Empirical results demonstrate that $\\texttt{PN-Train}$ considerably improves forecasting accuracy for low-frequency events while maintaining high performance for high-frequency events. The code is available at https://anonymous.4open.science/r/PN-Train.", "keywords": [ "urban time series forecasting", "neuron detection" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 4, 2, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 2, 5, 4 ] }, { "paper_id": "a8wjeqTZ9C", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Jueun Mun", "gender": "Female", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Namhoon Lee", "gender": "unknown", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Seonghwan Park", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Sungbin Shin", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 50, "n_ref": 78, "n_ref_all": 108, "n_fig": 20, "n_tab": 7, "L_tab": 1648, "n_element_tab": 109, "n_fig_1": 14, "n_tab_1": 4, "L_tab_1": 675, "n_element_tab_1": 34, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 146220, "len_all_1": 66315, "len_abs": 4322, "len_title": 91, "len_sents": 46793, "len_sents_1": 32315, "n_sents": 334, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1356, "title": "Concept Bottleneck Models under Label Noise", "abs": "Concept bottleneck models (CBMs) are a class of interpretable neural network models that make the final predictions based on intermediate representations known as concepts. With these concepts being human-interpretable, CBMs enable one to better understand the decisions made by neural networks. Despite this advantage, we find that CBMs face a critical limitation: they require additional labeling efforts for concept annotation, which can easily increase the risk of mislabeling, i.e., CBMs need to be trained with noisy labels. In this work, we systematically investigate the impact of label noise on CBMs, demonstrating that it can significantly compromise both model performance and interpretability. Specifically, we measure the impact of varying levels of label noise across different training schemes, through diverse lenses including extensive numerical evaluations, feature visualizations, and in-depth analysis of individual concepts, identifying key factors contributing to the breakdowns and establishing a better understanding of underlying challenges. To mitigate these issues, we propose leveraging a robust optimization technique called sharpness-aware minimization (SAM). By improving the quality of intermediate concept predictions, SAM enhances both the subsequent concept-level interpretability and final target prediction performance.", "keywords": [ "Concept bottleneck models", "Label noise", "Sharpness-aware minimization" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "a8uJXdi7Df", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Bin Yang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Hui Xiong", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Yunfan LU", "gender": "Male", "institution": "Hong Kong University of Science and Technology(GuangZhou)", "country": "CN", "position": "PhD student" }, { "name": "Yusheng Wang", "gender": "Male", "institution": "The University of Tokyo, Tokyo University", "country": "", "position": "Assistant Professor" }, { "name": "Zipeng Wang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 5, "n_ref_uni": 68, "n_ref": 240, "n_ref_all": 281, "n_fig": 21, "n_tab": 9, "L_tab": 2550, "n_element_tab": 302, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 813, "n_element_tab_1": 80, "formula_len_all": 320, "formula_len_all_1": 211, "len_all": 227837, "len_all_1": 64188, "len_abs": 1928, "len_title": 109, "len_sents": 59811, "len_sents_1": 28507, "n_sents": 467, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1712, "title": "Continuous Space-Time Video Super-Resolution via Event Camera", "abs": "Continuous space-time video super-resolution (C-STVSR) aims to simultaneously enhance video resolution and frame rate at an arbitrary scale. Recently, implicit neural representation (INR) has been applied to video restoration, representing videos as implicit fields that can be decoded at an arbitrary scale. However, the highly ill-posed nature of C-STVSR limits the effectiveness of current INR-based methods: they assume linear motion between frames and use interpolation or feature warping to generate features at arbitrary spatiotemporal positions with \\ubtxt{two} consecutive frames. This restrains C-STVSR from capturing rapid and \\ubtxt{nonlinear motion} and \\ubtxt{long-term dependencies} (\\textit{involving more than two frames}) in complex dynamic scenes. In this paper, we propose a novel C-STVSR framework, called \\textbf{HR-INR}, which captures both \\textbf{h}olistic dependencies and \\textbf{r}egional motions based on INR. It is assisted by an event camera -- a novel sensor renowned for its high temporal resolution and low latency. To fully utilize the rich temporal information from events, we design a feature extraction consisting of (1) a regional event feature extractor -- taking events as inputs via the proposed event temporal pyramid representation to capture the regional nonlinear motion and (2) a holistic event-frame feature extractor for long-term dependence and continuity motion. We then propose a novel INR-based decoder with spatiotemporal embeddings to capture long-term dependencies with a larger temporal perception field. We validate the effectiveness and generalization of our method on four datasets (both simulated and real data), showing the superiority of our method.", "keywords": [ "Event Camera", "Video Super-resolution", "Video Frame Interpolation", "Continuous Space-time Video Super-resolution" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "a8mKwRQQrP", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Adeline Fermanian", "gender": "unknown", "institution": "Califrais", "country": "", "position": "Researcher" }, { "name": "Massil HIHAT", "gender": "Male", "institution": "Califrais", "country": "FR", "position": "Researcher" } ], "n_formula": 96, "n_formula_1": 8, "n_ref_uni": 21, "n_ref": 74, "n_ref_all": 113, "n_fig": 7, "n_tab": 5, "L_tab": 1268, "n_element_tab": 93, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 212, "n_element_tab_1": 19, "formula_len_all": 11755, "formula_len_all_1": 373, "len_all": 179846, "len_all_1": 70313, "len_abs": 816, "len_title": 46, "len_sents": 64098, "len_sents_1": 33781, "n_sents": 594, "n_sents_1": 270, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 818, "title": "Online Policy Selection for Inventory Problems", "abs": "We tackle online inventory problems where at each time period the manager makes a replenishment decision based on partial historical information in order to meet demands and minimize costs. To solve such problems, we build upon recent works in online learning and control, use insights from inventory theory and propose a new algorithm called GAPSI. This algorithm follows a new feature-enhanced base-stock policy and deals with the troublesome question of non-differentiability which occurs in inventory problems. Our method is illustrated in the context of a complex and novel inventory system involving multiple products, lost sales, perishability, warehouse-capacity constraints and lead times. Extensive numerical simulations are conducted to demonstrate the good performances of our algorithm on real-world data.", "keywords": [ "Online learning", "Inventory control" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 4 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 2, 3, 3, 5 ] }, { "paper_id": "a8dQutiF9E", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Dongya Jia", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Jinhua Liang", "gender": "Male", "institution": "Queen Mary University of London", "country": "", "position": "PhD student" }, { "name": "Xiaobin Zhuang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Yi Yuan", "gender": "Male", "institution": "University of Surrey", "country": "", "position": "PhD student" }, { "name": "Yuanzhe Chen", "gender": "Male", "institution": "ByteDance Inc.", "country": "", "position": "Researcher" }, { "name": "Yuxuan Wang", "gender": "Male", "institution": "ByteDance", "country": "US", "position": "Researcher" }, { "name": "Zhengxi Liu", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "Undergrad student" }, { "name": "Zhuo Chen", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Wang Yuping", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 29, "n_ref": 67, "n_ref_all": 75, "n_fig": 3, "n_tab": 4, "L_tab": 1299, "n_element_tab": 178, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1279, "n_element_tab_1": 161, "formula_len_all": 1320, "formula_len_all_1": 1281, "len_all": 109908, "len_all_1": 64530, "len_abs": 1744, "len_title": 0, "len_sents": 30626, "len_sents_1": 27778, "n_sents": 237, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1318, "title": "AudioMorphix: Training-free audio editing with diffusion probabilistic models", "abs": "Despite recent advancements in diffusion-based audio generation, precisely editing content in a specific area of a recording remains challenging. In this paper, we introduce AudioMorphix, a training-free audio editor that manipulates a target area of a recording using another recording as a reference. Specifically, we conceptualize audio editing as part of a morphing cycle, \nin which different sounds can be combined into a cohesive audio mixture through morphing, whereas the mixture can be disentangled into individual components via demorphing. Leveraging the concept of audio morphing cycle, we optimize the noised latent conditioned on raw input together with reference audio and devise a series of energy functions to refine the guided diffusion process. Additionally, we manipulate the features within self-attention layers to preserve detailed characteristics from the original recordings. To accommodate a broad range of audio editing techniques, we collected a new evaluation dataset, providing editing instructions, reference audio and captions, and the duration of the edited area as guidance. Extensive experiments demonstrate that the AudioMorphix yields promising performance on various audio editing tasks, including addition, removal, and style transferring. Demo and code is available at this url.", "keywords": [ "Audio editing", "diffusion probabilistic model" ], "rating_list": [ 3, 3, 3, 5, 3 ], "soundness_list": [ 2, 3, 1, 3, 1 ], "presentation_list": [ 2, 1, 1, 2, 1 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 4, 3, 3 ] }, { "paper_id": "a8XwgTZzE0", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 1, "author_info_list": [ { "name": "Zihan Gu", "gender": "Male", "institution": "Institute of Information Engineering, Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 52, "n_formula_1": 15, "n_ref_uni": 24, "n_ref": 28, "n_ref_all": 49, "n_fig": 13, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 285, "n_element_tab_1": 19, "formula_len_all": 2250, "formula_len_all_1": 619, "len_all": 134836, "len_all_1": 47049, "len_abs": 1266, "len_title": 115, "len_sents": 47906, "len_sents_1": 21210, "n_sents": 414, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1289, "title": "Reconstruct the Understanding of Grokking through Dynamical Systems", "abs": "\\textbf{Grokking}, or the \\textbf{delayed generalization phenomenon}, describes the abrupt and rapid improvement in test accuracy that occurs after a model has been overfitted for a prolonged period. This phenomenon was first identified by Power in the context of operations on a prime number field. Over the past two years, a range of mathematical analyses has been conducted to investigate grokking, typically involving the use of the hidden progress measure which mean a function that can anticipate the occurrence of grokking. We believe that a comprehensive and rigorous mathematical modeling approach can invigorate the research on this task and provide a unified perspective for understanding previous research. This paper introduces a novel approach by modeling the task as a unique dynamical system. Using mathematical derivation within this framework, we propose a robust hidden progress measure that effectively captures the grokking phenomenon across all operations on prime number fields. This approach not only provides a more complete understanding but also offers deeper insights into the underlying architecture of the model. Based on this understanding, we also proposed a method to accelerate grokking without involving regularization or altering the model architecture.", "keywords": [ "interpretability", "grokking", "dynamical systems", "progress measures" ], "rating_list": [ 1, 5, 1, 1 ], "soundness_list": [ 1, 2, 2, 1 ], "presentation_list": [ 1, 2, 1, 1 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "a8R07y1jQ1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Danyal Maqbool", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "MS student" }, { "name": "Jixuan Chen", "gender": "Female", "institution": "Nanjing University", "country": "CN", "position": "Undergrad student" }, { "name": "Timothy Ossowski", "gender": "Male", "institution": "University of Wisconsin-Madison", "country": "US", "position": "PhD student" }, { "name": "Tyler J. Bradshaw", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Zefan Cai", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Junjie Hu", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 19, "n_ref": 32, "n_ref_all": 48, "n_fig": 6, "n_tab": 2, "L_tab": 595, "n_element_tab": 9, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 1322, "n_element_tab_1": 17, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 135939, "len_all_1": 64365, "len_abs": 2096, "len_title": 103, "len_sents": 46904, "len_sents_1": 30676, "n_sents": 394, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1316, "title": "COMMA: A Communicative Multimodal Multi-Agent Benchmark", "abs": "The rapid advances of multi-modal agents built on large foundation models have largely overlooked their potential for language-based communication between agents in collaborative tasks. This oversight presents a critical gap in understanding their effectiveness in real-world deployments, particularly when communicating with humans. Existing agentic benchmarks fail to address key aspects of inter-agent communication and collaboration, particularly in scenarios where agents have unequal access to information and must work together to achieve tasks beyond the scope of individual capabilities. To fill this gap, we introduce a novel benchmark designed to evaluate the collaborative performance of multimodal multi-agent systems through language communication. Our benchmark features a variety of scenarios, providing a comprehensive evaluation across four key categories of agentic capability in a communicative collaboration setting. By testing both agent-agent and agent-human collaborations using open-source and closed-source models, our findings reveal surprising weaknesses in state-of-the-art models, including proprietary models like GPT-4o. These models struggle to outperform even a simple random agent baseline in agent-agent collaboration and only surpass the random baseline when a human is involved.", "keywords": [ "Multimodality", "Agent", "LLM", "Benchmark" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "a84AD957m9", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Florian Bordes", "gender": "Male", "institution": "Meta", "country": "CA", "position": "Researcher" }, { "name": "Michal Drozdzal", "gender": "Male", "institution": "Meta", "country": "CA", "position": "Research Scientst" }, { "name": "Pietro Astolfi", "gender": "Male", "institution": "Meta", "country": "", "position": "Postdoc" }, { "name": "Rim Assouel", "gender": "Female", "institution": "Facebook", "country": "CA", "position": "Intern" }, { "name": "Adriana Romero", "gender": "Female", "institution": "Mila", "country": "CA", "position": "Core Industry Member" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 36, "n_ref": 108, "n_ref_all": 134, "n_fig": 8, "n_tab": 7, "L_tab": 2668, "n_element_tab": 487, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1132, "n_element_tab_1": 207, "formula_len_all": 757, "formula_len_all_1": 585, "len_all": 211987, "len_all_1": 70823, "len_abs": 4761, "len_title": 124, "len_sents": 59841, "len_sents_1": 33572, "n_sents": 404, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1264, "title": "OC-CLIP : Object-centric binding in Contrastive Language-Image Pretraining", "abs": "Recent advancements in vision-language models (VLMs) have been driven by contrastive models like CLIP which learn to associate visual information with their corresponding text descriptions. However, these models have limitations in understanding complex compositional scenes involving multiple objects and their spatial relationships. To address these challenges, we propose a novel approach that diverges from traditional data-centric methods of enhancing model performance with hard negatives examples. Our work instead focuses on integrating sufficient inductive biases into pre-trained CLIP-like models to improve their compositional understanding without using additional data annotations. We introduce a binding module that connects a scene graph of the text with an induced graph-like representation of the image, facilitating a structured similarity assessment. We also leverage relationships as text-conditioned visual constraints, thereby capturing the intricate interactions between objects and their contextual relationships more effectively. Our resulting model (OC-CLIP) not only enhances the performance of CLIP in multi-object compositional understanding but also paves the way for more accurate and efficient image-text matching in complex scenes.", "keywords": [ "object-centric representations", "object binding", "CLIP", "contrastive learning", "compositional image-to-text retrieval" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "a7gfCUhwdV", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Xiaogeng Liu", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Yaolun Zhang", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "Undergrad student" }, { "name": "chaowei xiao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 21, "n_ref": 50, "n_ref_all": 69, "n_fig": 3, "n_tab": 5, "L_tab": 1261, "n_element_tab": 159, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1971, "n_element_tab_1": 219, "formula_len_all": 287, "formula_len_all_1": 220, "len_all": 155218, "len_all_1": 56565, "len_abs": 1099, "len_title": 130, "len_sents": 37987, "len_sents_1": 25273, "n_sents": 287, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1137, "title": "MetaAgent: Automatically Building Multi-Agent System based on Finite State Machine", "abs": "Large Language Models (LLMs) can solve various practical tasks via a multi-agent system. However, existing human-designed multi-agent systems can only adapt to a limited number of pre-defined scenarios. Current auto-designed methods also have several drawbacks, including no tool support, reliance on in-bag training, and inflexible communication structure. Therefore, we propose \\textbf{MetaAgent}, a novel framework to automatically generate a multi-agent system based on a finite state machine. Given a task description, MetaAgent will design a multi-agent system and polish it through self-generated test queries. When the multi-agent system is deployed, the finite state machine, which supports the traceback and is more suitable for tool-using, will control the process to handle every case in the task domain. To evaluate our framework, we conduct experiments on both practical tasks and basic NLP tasks, the results indicate that the generated multi-agent system surpasses other auto-designed methods and can achieve a comparable performance with the human-designed multi-agent system which is polished for those specific tasks.", "keywords": [ "LLM Agent", "Multi-Agent System" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "a7gOjgFswH", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Fei Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Jiangchao Yao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tianjiao Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Ya Zhang", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Professor" }, { "name": "Yanfeng Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 25, "n_formula_1": 17, "n_ref_uni": 65, "n_ref": 144, "n_ref_all": 169, "n_fig": 11, "n_tab": 10, "L_tab": 2690, "n_element_tab": 239, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3436, "n_element_tab_1": 100, "formula_len_all": 1081, "formula_len_all_1": 788, "len_all": 213139, "len_all_1": 63592, "len_abs": 970, "len_title": 123, "len_sents": 50264, "len_sents_1": 27310, "n_sents": 371, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 970, "title": "G4Seg: Generation for Online Segmentation Refinement with Diffusion Models", "abs": "This paper considers the problem of utilizing a large-scale text-to-image diffusion model to tackle the challenging Inexact Segmentation (IS) task. Unlike traditional approaches that rely heavily on discriminative-model-based paradigm or dense visual representations derived from internal attention mechanisms, our method focuses on the intrinsic generative priors in Stable Diffusion~(SD). Specifically, we exploit the pattern discrepancies between original images and mask-conditional generated images to facilitate a coarse-to-fine segmentation refinement by establishing a semantic correspondence alignment and updating the foreground probability. Comprehensive quantitative and qualitative experiments validate the effectiveness and superiority of our plug-and-play design, underscoring the potential of leveraging generation discrepancies to model dense representations and encouraging further exploration of generative approaches for solving discriminative tasks.", "keywords": [ "Diffusion models", "Inexact Segmentation", "Semantic Correspondence" ], "rating_list": [ 5, 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 5, 4 ] }, { "paper_id": "a79bwlyUNp", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Arik Reuter", "gender": "Male", "institution": "Technische Universit\u00e4t Clausthal", "country": "DE", "position": "Researcher" }, { "name": "David R\u00fcgamer", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Associate Professor" }, { "name": "Vincent Fortuin", "gender": "Male", "institution": "Technical University of Munich", "country": "", "position": "Principal Researcher" }, { "name": "Tim Georg Johann Rudner", "gender": "Not Specified", "institution": "New York University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 20, "n_formula_1": 13, "n_ref_uni": 72, "n_ref": 115, "n_ref_all": 147, "n_fig": 2, "n_tab": 17, "L_tab": 21652, "n_element_tab": 1311, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1184, "n_element_tab_1": 117, "formula_len_all": 2364, "formula_len_all_1": 757, "len_all": 274045, "len_all_1": 72703, "len_abs": 1478, "len_title": 95, "len_sents": 73757, "len_sents_1": 31525, "n_sents": 498, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1121, "title": "In-Context Learning for Full Bayesian Inference", "abs": "Transformers have emerged as the dominant architecture in the field of deep learning, with a broad range of applications and remarkable in-context learning (ICL) capabilities. While not yet fully understood, ICL has already proved to be an intriguing phenomenon, allowing transformers to learn in-context---without requiring further training. In this paper, we further advance the understanding of ICL by demonstrating that transformers can perform full Bayesian inference for commonly used statistical models in-context. More specifically, we introduce a general framework that builds on ideas from prior fitted networks and continuous normalizing flows and enables us to infer complex posterior distributions for models such as generalized linear models and latent factor models. Extensive experiments on real-world datasets demonstrate that our ICL approach yields posterior samples that are similar in quality to state-of-the-art MCMC or variational inference methods that do not operate in-context. The source code for this paper is available at https://anonymous.4open.science/r/ICL_For_Full_Bayesian_Inference-3F53", "keywords": [ "In-Context Learning", "Prior-Fitted-Networks", "Bayesian Inference" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 2, 4, 3, 2 ] }, { "paper_id": "a72vorQK8v", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Alessandro Palma", "gender": "Male", "institution": "Helmholtz M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Leon Hetzel", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Sergei Rybakov", "gender": "Male", "institution": "Helmholtz Zentrum M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Stephan G\u00fcnnemann", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Professor" }, { "name": "Fabian J. Theis", "gender": "unknown", "institution": "Technical University Munich", "country": "", "position": "Full Professor" } ], "n_formula": 59, "n_formula_1": 21, "n_ref_uni": 47, "n_ref": 77, "n_ref_all": 153, "n_fig": 15, "n_tab": 6, "L_tab": 2332, "n_element_tab": 139, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 3025, "n_element_tab_1": 84, "formula_len_all": 3263, "formula_len_all_1": 1144, "len_all": 214309, "len_all_1": 71844, "len_abs": 1387, "len_title": 130, "len_sents": 74090, "len_sents_1": 33641, "n_sents": 579, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1372, "title": "Enforcing Latent Euclidean Geometry in VAEs for Statistical Manifold Interpolation", "abs": "Latent linear interpolations are a powerful tool for navigating the representation space of deep generative models. This aspect is particularly relevant in applied settings, where meaningful latent traversals can be learnt to represent the evolution of a system's trajectory and mapped back to the often complex and high-dimensional data space. However, when data lies on a manifold with complex geometry, linear interpolations of the representation space do not directly correspond to geodesic paths along the manifold unless enforced. An example of such a setting is scRNA-seq, where high-dimensional and discrete cellular data is assumed to lie on a negative binomial statistical manifold modelled by the decoder of a variational autoencoder. We introduce FlatVI, a novel training framework enforcing Euclidean geometry in the latent space of discrete-likelihood variational autoencoders modelling count data. In our regularisation setting, straight lines in the latent domain correspond to geodesic interpolations in the decoded space, improving the combination of our model with methods assuming Euclidean latent geometry. Results on simulated data empirically support our claims, while experiments on temporally resolved biological datasets show improvements in the reconstruction of cellular trajectories and the learning of biologically meaningful velocity fields.", "keywords": [ "scRNA-seq", "Riemannian geometry", "representation learning", "trajectory inference", "VAEs", "statistical manifolds" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 4, 4, 3 ], "presentation_list": [ 2, 4, 4, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 4, 2, 3 ] }, { "paper_id": "a6bnpOInjs", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Aleksandra Faust", "gender": "Female", "institution": "Google Brain", "country": "US", "position": "Principal Researcher" }, { "name": "Pieter Abbeel", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Volodymyr Mnih", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Hao Liu", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 47, "n_ref": 76, "n_ref_all": 89, "n_fig": 5, "n_tab": 4, "L_tab": 1232, "n_element_tab": 132, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1044, "n_element_tab_1": 106, "formula_len_all": 136, "formula_len_all_1": 180, "len_all": 130826, "len_all_1": 56750, "len_abs": 985, "len_title": 115, "len_sents": 31690, "len_sents_1": 26260, "n_sents": 229, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 988, "title": "Textbook Consistency Weighted Internet Improves Efficiency Twofold", "abs": "We propose a novel method, Textbook Consistency, to improve the training efficiency of large language models by leveraging textbooks as a guiding signal for learning from internet-scale data. Rather than relying on hard filtering of data based on quality thresholds before training, our approach adaptively adjusts the weight of data during training based on its consistency with textbooks during training. We compute the cosine similarity between internet data and textbooks in a latent space, using this metric to modulate the cross-entropy loss. Our method significantly enhances training efficiency, achieving twice the effectiveness by reducing training time or the number of tokens required. Empirical results show superior performance on language models trained on large datasets like FineWeb and The Pile, with extensions to other domains such as robotics. Our method is simple to implement, incurs no additional overhead, and is compatible with existing data curation techniques.", "keywords": [ "training efficiency", "large language model", "adaptive data weighting" ], "rating_list": [ 5, 5, 6, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "a6XE2GJHjk", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Gleb Bazhenov", "gender": "Male", "institution": "Higher School of Economics", "country": "", "position": "PhD student" }, { "name": "Liudmila Prokhorenkova", "gender": "Female", "institution": "Yandex", "country": "", "position": "Researcher" }, { "name": "Oleg Platonov", "gender": "Male", "institution": "Yandex", "country": "RU", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 77, "n_ref": 135, "n_ref_all": 149, "n_fig": 1, "n_tab": 3, "L_tab": 3837, "n_element_tab": 462, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 3844, "n_element_tab_1": 440, "formula_len_all": 128, "formula_len_all_1": 0, "len_all": 248115, "len_all_1": 78792, "len_abs": 1607, "len_title": 141, "len_sents": 84278, "len_sents_1": 36286, "n_sents": 508, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1613, "title": "TabGraphs: A Benchmark and Strong Baselines for Learning on Graphs with Tabular Node Features", "abs": "Tabular machine learning is an important field for industry and science. In this field, table rows are typically treated as independent data samples, but additional information about the relations between these samples is sometimes available and can be used to improve predictive performance. Such information can be naturally modeled with a graph, hence tabular machine learning may benefit from graph machine learning methods. However, graph machine learning models are typically evaluated on datasets with homogeneous, most often text-based node features, which are very different from heterogeneous mixtures of numerical and categorical features present in tabular datasets. Thus, there is a critical difference between the data used in tabular and graph machine learning studies, which does not allow one to understand how successfully graph models can be transferred to tabular data. To bridge this gap, we propose a new benchmark of diverse graphs with heterogeneous tabular node features and realistic prediction tasks. We use this benchmark to evaluate a vast set of models, including simple methods previously overlooked in the literature. Our experiments show that graph neural networks indeed can often bring gains in predictive performance for tabular data, but standard tabular models can also be adapted to work with graph data by using simple graph-based feature augmentation, which sometimes enables them to compete with and even outperform graph neural models. Based on our empirical study, we provide insights for researchers and practitioners in both tabular and graph machine learning fields.", "keywords": [ "graph machine learning", "tabular machine learning", "graph neural network", "gradient boosting", "benchmark" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "a6U41REOa5", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Jianing Tian", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yaokun Ji", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Zichen Wang", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "SHUANGJIA ZHENG", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 58, "n_ref": 93, "n_ref_all": 104, "n_fig": 6, "n_tab": 6, "L_tab": 1402, "n_element_tab": 315, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1255, "n_element_tab_1": 225, "formula_len_all": 765, "formula_len_all_1": 782, "len_all": 172113, "len_all_1": 65924, "len_abs": 1404, "len_title": 140, "len_sents": 42031, "len_sents_1": 29961, "n_sents": 286, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1411, "title": "Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization", "abs": "Antibodies are essential proteins responsible for immune responses in organisms, capable of specifically recognizing antigen molecules of pathogens. Recent advances in generative models have significantly enhanced rational antibody design. However, existing methods mainly create antibodies from scratch without template constraints, leading to model optimization challenges and unnatural sequences. To address these issues, we propose a retrieval-augmented diffusion framework, termed RADAb, for efficient antibody design. Our method leverages a set of structural homologous motifs that align with query structural constraints to guide the generative model in inversely optimizing antibodies according to desired design criteria. Specifically, we introduce a structure-informed retrieval mechanism that integrates these exemplar motifs with the input backbone through a novel dual-branch denoising module, utilizing both structural and evolutionary information. Additionally, we develop a conditional diffusion model that iteratively refines the optimization process by incorporating both global context and local evolutionary conditions. Our approach is agnostic to the choice of generative models. Empirical experiments demonstrate that our method achieves state-of-the-art performance in multiple antibody inverse folding and optimization tasks, offering a new perspective on biomolecular generative models.", "keywords": [ "Generative model", "Retrieval augmented geneartion", "Protein design" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "a69zct3BkY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Futing Wang", "gender": "Female", "institution": "Westlake University", "country": "", "position": "PhD student" }, { "name": "Jianhao Yan", "gender": "Male", "institution": "Westlake University", "country": "", "position": "PhD student" }, { "name": "Yafu Li", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yue Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" }, { "name": "Yun Luo", "gender": "Male", "institution": "westlake university", "country": "", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 15, "n_ref_uni": 23, "n_ref": 64, "n_ref_all": 77, "n_fig": 6, "n_tab": 4, "L_tab": 599, "n_element_tab": 51, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1410, "formula_len_all_1": 914, "len_all": 124182, "len_all_1": 61516, "len_abs": 1165, "len_title": 118, "len_sents": 36152, "len_sents_1": 29436, "n_sents": 285, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1169, "title": "Keys to Robust Edits: From Theoretical Insights to Practical Advances", "abs": "Large language models (LLMs) have revolutionized knowledge storage and retrieval, but face challenges with conflicting and outdated information. Knowledge editing techniques have been proposed to address these issues, yet they struggle with robustness tests involving long contexts, paraphrased subjects, and continuous edits. This work investigates the cause of these failures in locate-and-edit methods, offering theoretical insights into their key-value modeling and deriving mathematical bounds for robust and specific edits, leading to a novel 'group discussion' conceptual model for locate-and-edit methods. Empirical analysis reveals that keys used by current methods fail to meet robustness and specificity requirements. To address this, we propose a Robust Edit Pathway (REP) that disentangles editing keys from LLMs' inner representations. Evaluations on LLaMA2-7B and Mistral-7B using the CounterFact dataset show that REP significantly improves robustness across various metrics, both in-domain and out-of-domain, with minimal trade-offs in success rate and locality. Our findings advance the development of reliable and flexible knowledge updating in LLMs.", "keywords": [ "model editing" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "a5EFuQuuPb", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Ori Matityahu", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "MS student" }, { "name": "Raanan Fattal", "gender": "Male", "institution": "Hebrew University, Hebrew University of Jerusalem", "country": "IL", "position": "Full Professor" } ], "n_formula": 28, "n_formula_1": 10, "n_ref_uni": 32, "n_ref": 135, "n_ref_all": 174, "n_fig": 4, "n_tab": 10, "L_tab": 3924, "n_element_tab": 625, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2624, "n_element_tab_1": 434, "formula_len_all": 2024, "formula_len_all_1": 566, "len_all": 185967, "len_all_1": 75238, "len_abs": 1633, "len_title": 80, "len_sents": 49117, "len_sents_1": 31601, "n_sents": 347, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 32, "L_abs": 1643, "title": "Generative Lines Matching Models", "abs": "In this paper we identify the source of a singularity in the training loss of key denoising models, that causes the denoiser's predictions to collapse towards the mean of the source or target distributions. This degeneracy creates false basins of attraction, distorting the denoising trajectories and ultimately increasing the number of steps required to sample these models.\n\nWe circumvent this artifact by leveraging the deterministic ODE-based samplers, offered by certain denoising diffusion and score-matching models, which establish a well-defined change-of-variables between the source and target distributions. Given this correspondence, we propose a new probability flow model, the Lines Matching Model (LMM), which matches globally straight lines interpolating the two distributions. We demonstrate that the flow fields produced by the LMM exhibit notable temporal consistency, resulting in trajectories with excellent straightness scores.\n\nBeyond its sampling efficiency, the LMM formulation allows us to enhance the fidelity of the generated samples by integrating domain-specific reconstruction and adversarial losses, and by optimizing its training for the sampling procedure used. Overall, the LMM achieves state-of-the-art FID scores with minimal NFEs on established benchmark datasets: 1.57/1.39 (NFE=1/2) on CIFAR-10, 1.47/1.17 on ImageNet 64x64, and 2.68/1.54 on AFHQ 64x64.\n\nFinally, we provide a theoretical analysis showing that the use of optimal transport to relate the two distributions suffers from a curse of dimensionality, where the pairing set size (mini-batch) must scale exponentially with the signal dimension.", "keywords": [ "Generative Flow", "Denoising Score Matching", "Denoising Diffusion Probability Models" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "a5EC2R1NeS", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Changchun Li", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Associate Professor" }, { "name": "Ximing Li", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Full Professor" }, { "name": "Yuanchao Dai", "gender": "Female", "institution": "Jilin University", "country": "CN", "position": "PhD student" } ], "n_formula": 40, "n_formula_1": 23, "n_ref_uni": 12, "n_ref": 21, "n_ref_all": 30, "n_fig": 2, "n_tab": 3, "L_tab": 3750, "n_element_tab": 352, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 578, "n_element_tab_1": 56, "formula_len_all": 5168, "formula_len_all_1": 2194, "len_all": 117225, "len_all_1": 61211, "len_abs": 1222, "len_title": 147, "len_sents": 33224, "len_sents_1": 27336, "n_sents": 234, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1226, "title": "Confidence Difference Reflects Various Supervised Signals in Confidence-Difference Classification", "abs": "Training a precise binary classifier with limited supervision in weakly supervised learning scenarios holds considerable research significance in practical settings. Leveraging pairwise unlabeled data with confidence differences has been demonstrated to outperform learning from pointwise unlabeled data. We theoretically analyze the various supervisory signals reflected by confidence differences in confidence difference (ConfDiff) classification and identify challenges arising from noisy signals when confidence differences are small. To address this, we partition the dataset into two subsets with distinct supervisory signals and propose a consistency regularization-based risk estimator to encourage similar outputs for similar instances, mitigating the impact of noisy supervision. We further derive and analyze its estimation error bounds theoretically. Extensive experiments on benchmark and UCI datasets demonstrate the effectiveness of our method. Additionally, to effectively capture the influence of real-world noise on the confidence difference, we artificially perturb the confidence difference distribution and demonstrate the robustness of our method under noisy conditions through comprehensive experiments.", "keywords": [ "Confidence-Difference Classification", "Binary Classification", "Weakly Supervised Learning" ], "rating_list": [ 8, 6, 5, 5 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "a59NMkKPob", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Bin-Bin Gao", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Senior Researcher" }, { "name": "Chengjie Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Jiawei Zhan", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Jinlong Peng", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Jun Liu", "gender": "Male", "institution": "Tencent YouTu Lab", "country": "CN", "position": "Researcher" }, { "name": "Wenjie Pei", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xiaochen Chen", "gender": "Male", "institution": "Tencent YouTu Lab", "country": "CN", "position": "Researcher" }, { "name": "Yong Liu", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" } ], "n_formula": 20, "n_formula_1": 14, "n_ref_uni": 36, "n_ref": 68, "n_ref_all": 91, "n_fig": 5, "n_tab": 11, "L_tab": 3060, "n_element_tab": 287, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2725, "n_element_tab_1": 265, "formula_len_all": 763, "formula_len_all_1": 548, "len_all": 176403, "len_all_1": 72429, "len_abs": 1856, "len_title": 122, "len_sents": 33389, "len_sents_1": 29809, "n_sents": 256, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1885, "title": "DRL: DISCRIMINATIVE REPRESENTATION LEARNING FOR CLASS INCREMENTAL LEARNING", "abs": "Non-rehearsal class incremental learning (CIL) is pivotal in real-world scenarios such as data streaming applications and data security. \nDespite the remarkable progress in research on CIL, it remains an extremely challenging task due to three conundrums: increasingly large model complexity, non-smooth representation shift during incremental learning and inconsistency between stage-wise sub-problem optimization and global inference. In this work, we propose the Discriminative Representation Learning (\\emph{DRL}) method to deal with these challenges specifically. To conduct incremental learning effectively and yet efficiently, our \\emph{DRL} is built upon a pre-trained large model with excellent representation learning capability, and increasingly augments the model by learning a lightweight adapter with a small amount of parameter learning overhead in each incremental learning stage. While the adapter is responsible for adapting the model to new classes of data involved in current learning stage, it can inherit and propagate the representation capability from the current model via parallel connection between them. As a result, such design can guarantee a smooth representation shift between different stages of incremental learning. Furthermore, to alleviate the issue of the training-inference inconsistency induced by the stage-wise sub-optimization, we design the Margin-CE loss, which imposes a hard margin between classification boundaries to push for more discriminative representation learning, thereby narrowing down the gap between stage-wise local optimization over a subset of data and global inference on all classes of data. Extensive experiments on six benchmarks reveal that our \\emph{DRL} consistently outperforms other state-of-the-art methods throughout the entire CIL period while maintaining high efficiency in both training and inference phases.", "keywords": [ "class incremental learning" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "a4sknPttwV", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Benhao Huang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Jin Huang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Xingjian Zhang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Yingzhuo Yu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Jiaqi Ma", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 51, "n_ref": 83, "n_ref_all": 104, "n_fig": 5, "n_tab": 13, "L_tab": 2772, "n_element_tab": 195, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 583, "n_element_tab_1": 69, "formula_len_all": 186, "formula_len_all_1": 0, "len_all": 251914, "len_all_1": 71517, "len_abs": 2313, "len_title": 98, "len_sents": 74050, "len_sents_1": 31718, "n_sents": 624, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1544, "title": "DCA-Bench: A Benchmark for Dataset Curation Agents", "abs": "The quality of datasets plays an increasingly crucial role in the research and development of modern artificial intelligence (AI). Despite the proliferation of open dataset platforms nowadays, data quality issues, such as incomplete documentation, inaccurate labels, ethical concerns, and outdated information, remain common in widely used datasets. Furthermore, these issues are often subtle and difficult to be detected by rule-based scripts, therefore requiring identification and verification by dataset users or maintainers--a process that is both time-consuming and prone to human mistakes. With the surging ability of large language models (LLM), it\u2019s promising to streamline the discovery of hidden dataset issues with LLM agents. To achieve this, one significant challenge is enabling LLM agents to detect issues in the wild rather than simply fixing known ones. In this work, we establish a benchmark to measure LLM agent\u2019s ability to tackle this challenge. We carefully curate 221 representative test cases from eight popular dataset platforms and propose an automatic evaluation framework using GPT-4. Our proposed framework shows strong empirical alignment with expert evaluations, validated through extensive comparisons with human annotations. Without any hints, a baseline GPT-4 agent can only reveal 11% of the data quality issues in the proposed dataset, highlighting the complexity of this task and indicating that applying LLM agents to real-world dataset curation still requires further in-depth exploration and innovation.", "keywords": [ "Dataset Curation", "LLM Agent", "Automatic Evaluation" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 2, 2, 4 ] }, { "paper_id": "a4nSE2kpoq", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Anna Bair", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Devin Willmott", "gender": "Male", "institution": "Bosch Center for Artificial Intelligence", "country": "US", "position": "Research Scientist" }, { "name": "Madan Ravi Ganesh", "gender": "Male", "institution": "Bosch Center for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Mohammad Sadegh Norouzzadeh", "gender": "Male", "institution": "Bosch", "country": "", "position": "Researcher" }, { "name": "Victor Akinwande", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Zico Kolter", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 36, "n_ref": 40, "n_ref_all": 58, "n_fig": 5, "n_tab": 6, "L_tab": 3231, "n_element_tab": 283, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1219, "n_element_tab_1": 81, "formula_len_all": 191, "formula_len_all_1": 136, "len_all": 133656, "len_all_1": 64949, "len_abs": 1214, "len_title": 61, "len_sents": 43627, "len_sents_1": 32521, "n_sents": 281, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1219, "title": "HyperCLIP: Adapting Vision-Language models with Hypernetworks", "abs": "Self-supervised vision-language models trained with contrastive objectives form the basis of current state-of-the-art methods in AI vision tasks. The success of these models is a direct consequence of the huge web-scale datasets used to train them, but they require correspondingly large vision components to properly learn powerful and general representations from such a broad data domain. This poses a challenge for deploying large vision-language models, especially in resource-constrained environments. To address this, we propose an alternate vision-language architecture, called HyperCLIP, that uses a small image encoder along with a hypernetwork that dynamically adapts image encoder weights to each new set of text inputs. All three components of the model (hypernetwork, image encoder, and text encoder) are pre-trained jointly end-to-end, and with a trained HyperCLIP model, we can generate new zero-shot deployment-friendly image classifiers for any task with a single forward pass through the text encoder and hypernetwork. HyperCLIP increases the zero-shot accuracy of SigLIP trained models with small image encoders by up to 3% on ImageNet and 5% on CIFAR-100 with minimal training throughput overhead.", "keywords": [ "vision-language models", "edge computing", "CLIP", "hyper-networks" ], "rating_list": [ 3, 6, 3 ], "soundness_list": [ 1, 3, 2 ], "presentation_list": [ 1, 3, 1 ], "contribution_list": [ 2, 3, 1 ], "confidence_list": [ 4, 3, 5 ] }, { "paper_id": "a4PBF1YInZ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Guodong Guo", "gender": "Male", "institution": "West Virginia University", "country": "US", "position": "Full Professor" }, { "name": "Luhao Zhu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Xiaohong Liu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yan Tai", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yiying Dong", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "CN", "position": "PhD student" }, { "name": "Yunan Ding", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "CN", "position": "PhD student" }, { "name": "zhiqiang chen", "gender": "Male", "institution": "Eastern Institute of Technology", "country": "CN", "position": "Postdoc" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 28, "n_ref": 66, "n_ref_all": 84, "n_fig": 6, "n_tab": 13, "L_tab": 5123, "n_element_tab": 642, "n_fig_1": 6, "n_tab_1": 8, "L_tab_1": 3105, "n_element_tab_1": 502, "formula_len_all": 744, "formula_len_all_1": 744, "len_all": 144367, "len_all_1": 59727, "len_abs": 1420, "len_title": 170, "len_sents": 37008, "len_sents_1": 23861, "n_sents": 272, "n_sents_1": 159, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1586, "title": "VT-PLUG: Integrating Visual Task Plugins with Unified Instruction Tuning", "abs": "Multimodal Large Language Models (MLLMs) demonstrate robust zero-shot capabilities across diverse vision-language tasks after training on mega-scale datasets. However, dense prediction tasks, such as semantic segmentation and keypoint detection, pose significant challenges for MLLMs when represented solely as text outputs. These challenges often necessitate task-specific visual decoders, leading to the underutilization of MLLMs' multi-task potential. In this work, we propose VT-PLUG, a novel framework that leverages modular visual components as scalable plugins for a variety of visual applications. During the joint training of vision-language tasks with varying prediction densities, we propose a Visual Decoding Chain-of-Thought (VD-CoT) mechanism to prevent task conflicts. VD-CoT requires the model to predict the current task's recognition entities, decoding unit type, and other specific details, while also providing learnable queries for precise decoding. Additionally, we construct VT-Instruct, a large-scale multi-task dataset containing over 100 million multimodal dialogue samples across 25 task types. Beyond text inputs and outputs, VT-Instruct incorporates various visual prompts such as point, box, scribble, and mask, and generates outputs composed of text and visual units like point, box, keypoint, and mask. The combination of different visual prompts and visual units generates a wide variety of task types, expanding the applicability of VT-PLUG significantly. The source code, dataset and demo will be released at https://anonymous.4open.science/r/VT-PLUG.", "keywords": [ "Multimodal Large Language Model", "Multi-task Learning", "Object Detection", "Segmentation", "Keypoint Detection" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "a43oZCXdNC", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Arshia Afzal", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Francesco Tonin", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Postdoc" }, { "name": "Mahsa Shoaran", "gender": "Female", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Assistant Professor" }, { "name": "Volkan Cevher", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 47, "n_ref": 144, "n_ref_all": 167, "n_fig": 5, "n_tab": 18, "L_tab": 4711, "n_element_tab": 607, "n_fig_1": 3, "n_tab_1": 9, "L_tab_1": 2672, "n_element_tab_1": 309, "formula_len_all": 1153, "formula_len_all_1": 924, "len_all": 160727, "len_all_1": 64415, "len_abs": 1130, "len_title": 124, "len_sents": 44251, "len_sents_1": 25912, "n_sents": 303, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1139, "title": "Space-time self-attention for graph signal processing", "abs": "This work introduces a Transformer-based approach for graph signal processing that leverages a novel task-specific attention mechanism, namely NTAttention. \nUnlike conventional self-attention mechanisms, our method attends to all nodes across multiple time steps, enabling the model to effectively capture dependencies between nodes over extended time periods. This addresses a key limitation faced by traditional methods.\nAdditionally, we propose geometry-aware masking (GMask), which incorporates the graph topology into the sparsification of the self-attention matrix. This enhances efficiency while preserving the rich temporal information conveyed by the nodes. \nWe demonstrate the effectiveness of our approach on two critical applications: EEG seizure detection and traffic forecasting. Both tasks involve data collected from fixed sensors, such as electrodes or road sensors, where data from one sensor can influence others temporally and spatially. Our model enhances sensitivity in fast seizure detection by 20 percentage points compared to state-of-the-art and significantly outperforms current methods in traffic forecasting.", "keywords": [ "seizure detection", "eeg", "traffic forecasting", "graph signal processing", "attention", "transformer" ], "rating_list": [ 5, 3, 1, 5, 5 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 1, 3, 3 ], "contribution_list": [ 2, 1, 1, 2, 3 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "a3ptUbuzbW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Chuanchen Luo", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Junran Peng", "gender": "Male", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "Associate Professor" }, { "name": "Zhaoxiang Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhongkai Mao", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Jimmy Liu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 27, "n_ref": 88, "n_ref_all": 120, "n_fig": 10, "n_tab": 8, "L_tab": 3565, "n_element_tab": 449, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1011, "n_element_tab_1": 55, "formula_len_all": 488, "formula_len_all_1": 386, "len_all": 147043, "len_all_1": 57885, "len_abs": 704, "len_title": 0, "len_sents": 41412, "len_sents_1": 27480, "n_sents": 351, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1323, "title": "CityGaussianV2: Efficient and Geometrically Accurate Reconstruction for Large-Scale Scenes", "abs": "Recently, 3D Gaussian Splatting (3DGS) has revolutionized radiance field reconstruction, manifesting efficient and high-fidelity novel view synthesis. However, accurately representing surfaces, especially in large and complex scenarios, remains a significant challenge due to the unstructured nature of 3DGS. In this paper, we present CityGaussianV2, a novel approach for large-scale scene reconstruction that addresses critical challenges related to geometric accuracy and efficiency. Building on the favorable generalization capabilities of 2D Gaussian Splatting (2DGS), we address its convergence and scalability issues. Specifically, we implement a decomposed-gradient-based densification and depth regression technique to eliminate blurry artifacts and accelerate convergence. To scale up, we introduce an elongation filter that mitigates Gaussian count explosion caused by 2DGS degeneration. Furthermore, we optimize the CityGaussian pipeline for parallel training, achieving up to 10$\\times$ compression, at least 25\\% savings in training time, and a 50\\% decrease in memory usage. We also established standard geometry benchmarks under large-scale scenes. Experimental results demonstrate that our method strikes a promising balance between visual quality, geometric accuracy, as well as storage and training costs.", "keywords": [ "neural rendering", "novel view synthesis", "large-scale scene", "radiance field", "surfel splatting", "surfel reconstruction" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "a3g2l4yEys", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Akari Asai", "gender": "Female", "institution": "Paul G. Allen School of Computer Science & Engineering, University of Washington", "country": "", "position": "PhD student" }, { "name": "Anjali Kantharuban", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Graham Neubig", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Jean de Dieu Nyandwi", "gender": "Male", "institution": "Carnegie Mellon University", "country": "RW", "position": "MS student" }, { "name": "Lintang Sutawika", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Sathyanarayanan Ramamoorthy", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Seungone Kim", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Simran Khanuja", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Xiang Yue", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Yueqi Song", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 45, "n_ref": 95, "n_ref_all": 138, "n_fig": 19, "n_tab": 11, "L_tab": 7358, "n_element_tab": 1265, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 345, "n_element_tab_1": 4, "formula_len_all": 1595, "formula_len_all_1": 532, "len_all": 371217, "len_all_1": 57026, "len_abs": 1461, "len_title": 114, "len_sents": 125624, "len_sents_1": 28685, "n_sents": 888, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1245, "title": "Pangea: A Fully Open Multilingual Multimodal LLM for 39 Languages", "abs": "Despite recent advances in multimodal large language models (MLLMs), their development has predominantly focused on English- and western-centric datasets and tasks, leaving most of the world's languages and diverse cultural contexts underrepresented. \nThis paper introduces PANGEA, a multilingual multimodal LLM trained on PANGEAINS, a diverse 6M instruction dataset spanning 39 languages. PANGEAINS features: 1) high-quality English instructions, 2) carefully machine-translated instructions, and 3) culturally relevant multimodal tasks to ensure cross-cultural coverage. \nTo rigorously assess models' capabilities, we introduce PANGEABENCH, a holistic evaluation suite encompassing 14 datasets covering 47 languages. \nResults show that PANGEA significantly outperforms existing open-source models in multilingual settings and diverse cultural contexts. Ablation studies further reveal the importance of English data proportions, language popularity, and the number of multimodal training samples on overall performance. We fully open-source our data, code, and trained checkpoints, to facilitate the development of inclusive and robust multilingual MLLMs, promoting equity and accessibility across a broader linguistic and cultural spectrum.", "keywords": [ "Multilingual", "Multimodal", "LLMs" ], "rating_list": [ 8, 5, 8, 6 ], "soundness_list": [ 4, 2, 4, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 4, 3, 4, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "a3PmRgAB5T", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Huizhuo Yuan", "gender": "Female", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Kaixuan Ji", "gender": "Not Specified", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Yiming Yang", "gender": "Female", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Yue Wu", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Zhiqing Sun", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Researcher" } ], "n_formula": 50, "n_formula_1": 32, "n_ref_uni": 42, "n_ref": 104, "n_ref_all": 133, "n_fig": 5, "n_tab": 5, "L_tab": 3452, "n_element_tab": 353, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 916, "n_element_tab_1": 130, "formula_len_all": 4012, "formula_len_all_1": 2825, "len_all": 207863, "len_all_1": 65204, "len_abs": 1653, "len_title": 111, "len_sents": 58846, "len_sents_1": 24964, "n_sents": 487, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1664, "title": "Self-Play Preference Optimization for Language Model Alignment", "abs": "Standard reinforcement learning from human feedback (RLHF) approaches relying on parametric models like the Bradley-Terry model fall short in capturing the intransitivity and irrationality in human preferences. Recent advancements suggest that directly working with preference probabilities can yield a more accurate reflection of human preferences, enabling more flexible and accurate language model alignment. In this paper, we propose a self-play-based method for language model alignment, which treats the problem as a constant-sum two-player game aimed at identifying the Nash equilibrium policy. Our approach, dubbed *Self-Play Preference Optimization* (SPPO), utilizes iterative policy updates to provably approximate the Nash equilibrium. \nAdditionally, we propose a new SPPO objective which is both strongly motivated by theory and is simple and effective in practice.\nIn our experiments, using only 60k prompts (without responses) from the UltraFeedback dataset and without any prompt augmentation, by leveraging a pre-trained preference model PairRM with only 0.4B parameters, SPPO can obtain a model from fine-tuning Mistral-7B-Instruct-v0.2 that achieves the state-of-the-art length-controlled win-rate of 28.53\\% against GPT-4-Turbo on AlpacaEval 2.0. It also outperforms the (iterative) DPO and IPO on MT-Bench, Arena-Hard, and the Open LLM Leaderboard.\nStarting from a stronger base model Llama-3-8B-Instruct, we are able to achieve a length-controlled win rate of 38.77\\%.\nNotably, the strong performance of SPPO is achieved without additional external supervision (e.g., responses, preferences, etc.) from GPT-4 or other stronger language models.", "keywords": [ "self play", "preference optimization", "large language model", "RLHF" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "a2tU4ykVA9", "primary_area": "datasets and benchmarks", "n_author": 16, "author_info_list": [ { "name": "Bohan Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Changhua Pei", "gender": "Male", "institution": "Computer Network Information Center, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Dan Pei", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Gaogang Xie", "gender": "Male", "institution": "Computer Network Information Center, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Haiming Zhang", "gender": "Male", "institution": "cnic", "country": "", "position": "Full Professor" }, { "name": "Jianhui li", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Full Professor" }, { "name": "Longlong Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Minghua Ma", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Mingze Sun", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Shenglin Zhang", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Wangkun", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Xiaohui Nie", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Xidao Wen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Yongqian Sun", "gender": "Male", "institution": "Nankai University", "country": "", "position": "Associate Professor" }, { "name": "Yuhe Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "\u5f20\u5fd7\u777f", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "Undergrad student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 23, "n_ref": 46, "n_ref_all": 84, "n_fig": 20, "n_tab": 9, "L_tab": 7198, "n_element_tab": 523, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 3320, "n_element_tab_1": 296, "formula_len_all": 754, "formula_len_all_1": 311, "len_all": 175020, "len_all_1": 71542, "len_abs": 1842, "len_title": 162, "len_sents": 58653, "len_sents_1": 31028, "n_sents": 468, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 113, "L_abs": 1853, "title": "OpsEval: A Comprehensive Benchmark Suite for Evaluating Large Language Models\u2019 Capability in IT Operations Domain", "abs": "The past decades have witnessed the rapid development of Information Technology (IT) systems, such as cloud computing, 5G networks, and financial information systems. Ensuring the stability of these IT systems has become an important issue. Large language models (LLMs) that have exhibited remarkable capabilities in NLP-related tasks are showing great potential in AIOps, such as root cause analysis of failures, generation of operations and maintenance scripts, and summarizing of alert information. Unlike knowledge in general corpora, knowledge of Ops varies with the different IT systems, encompassing various private sub-domain knowledge, sensitive to prompt engineering due to various sub-domains, and containing numerous terminologies. Existing NLP-related benchmarks (e.g., C-Eval, MMLU) can not guide the selection of suitable LLMs for Ops (OpsLLM), and current metrics (e.g., BLEU, ROUGE) can not adequately reflect the question-answering (QA) effectiveness in the Ops domain. We propose a comprehensive benchmark suite, OpsEval, including an Ops-oriented evaluation dataset, an Ops evaluation benchmark, and a specially designed Ops QA evaluation method. Our dataset contains 7,334 multiple-choice questions and 1,736 QA questions. We have carefully selected and released 20% of the dataset written by domain experts in various sub-domains to assist current researchers in preliminary evaluations of OpsLLMs. We test over 24 latest LLMs under various settings such as self-consistency, chain-of-thought, and in-context learning, revealing findings when applying LLMs to Ops. We also propose an evaluation method for QA in Ops, which has a coefficient of 0.9185 with human experts and is improved by 0.4471 and 1.366 compared to BLEU and ROUGE, respectively. Over the past one year, our dataset and leaderboard have been continuously updated.", "keywords": [ "Large language models", "Ops", "Benchmark", "Dataset", "Evaluation", "Prompt engineering" ], "rating_list": [ 5, 8, 6, 3 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 2, 2, 4 ] }, { "paper_id": "a2rSx6t4EV", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Jingxu Gao", "gender": "Female", "institution": "LinkedIn", "country": "US", "position": "Researcher" }, { "name": "Li Wang", "gender": "unknown", "institution": "Dropbox", "country": "US", "position": "Machine Learning Engineer" }, { "name": "Ryan Li", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "MS student" }, { "name": "Xinxi Chen", "gender": "unknown", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 34, "n_ref": 46, "n_ref_all": 54, "n_fig": 3, "n_tab": 3, "L_tab": 329, "n_element_tab": 35, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 57, "n_element_tab_1": 4, "formula_len_all": 117, "formula_len_all_1": 117, "len_all": 119950, "len_all_1": 51075, "len_abs": 1054, "len_title": 94, "len_sents": 25003, "len_sents_1": 24697, "n_sents": 198, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1016, "title": "EDU-RAG: A RAG Benchmark with Web-enhanced Content in Education Domain. Can RAG Help AI Tutor?", "abs": "Hallucination has been a persistent challenge when using Large Language Models (LLMs). Retrieval-Augmented Generation (RAG) has emerged as a popular approach to mitigate this issue by maintaining context and coherence in generated outputs, as well as incorporating customized knowledge. In this paper, we propose a benchmark dataset for evaluating LLM performance in the domain of middle-school science question answering, using textbook questions augmented with real-world web search results. We assess the performance of various LLMs, including GPT-4o, Llama2-7b, and Llama3-8b, with and without the application of RAG. Our goal is to determine whether RAG can reduce hallucinations stemming from the inherent biases of pre-trained LLMs or from the retrieval of irrelevant knowledge, even when relevant information is accessible. The dataset and methodology introduced here provide a robust foundation for advancing the evaluation and development of RAG techniques in mitigating hallucinations across diverse LLMs.", "keywords": [ "Large Language Models (LLM)", "Retrieval Augmented Generation (RAG)", "Search" ], "rating_list": [ 3, 3, 1 ], "soundness_list": [ 2, 1, 1 ], "presentation_list": [ 1, 1, 1 ], "contribution_list": [ 1, 1, 1 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "a2gBrMu9MP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Benben Jiang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Songqi Zhou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Research Assistant" }, { "name": "Zeyuan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 7, "n_ref_uni": 31, "n_ref": 54, "n_ref_all": 73, "n_fig": 3, "n_tab": 7, "L_tab": 4342, "n_element_tab": 325, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2436, "n_element_tab_1": 151, "formula_len_all": 1986, "formula_len_all_1": 451, "len_all": 157604, "len_all_1": 65094, "len_abs": 1401, "len_title": 107, "len_sents": 53250, "len_sents_1": 28889, "n_sents": 433, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1242, "title": "FairLoRA: Targeted Bias Mitigation without Performance Loss", "abs": "Ensuring fairness in machine learning models is critical, but existing debiasing techniques often sacrifice model performance, struggle to adapt to emerging biases, or require extensive sensitive attribute annotations. To address these challenges, we propose FairLoRA, a novel low-rank adaptation method that mitigates bias while preserving model performance. FairLoRA incorporates parameter-efficient modular LoRA components, enabling iterative bias mitigation to ensure fairness across multiple sensitive attributes without interfering with previous adjustments. Furthermore, it employs discriminators to identify biased classes with reduced reliance on sensitive information, significantly reducing the need for annotated data. We theoretically derive conditions under which FairLoRA fine-tuning can effectively mitigate bias while maintaining the original model's performance. We then empirically validate its effectiveness across diverse computer vision and natural language processing tasks. Our experimental results show that, even for models that have undergone prior bias mitigation training, the integration of FairLoRA fine-tuning can further enhance fairness, while maintaining or even slightly improving the original performance.", "keywords": [ "Fair machine learning", "Fairness", "Bias mitigation", "Fine-tuning" ], "rating_list": [ 3, 3, 8, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "a2eBgp4sjH", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 11, "author_info_list": [ { "name": "Gopal R Srinivasa", "gender": "Not Specified", "institution": "Research, Microsoft", "country": "IN", "position": "Research SDE" }, { "name": "Haiyang Xu", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Senior Development Engineer" }, { "name": "Harsha Vardhan Simhadri", "gender": "Male", "institution": "Microsoft ", "country": "US", "position": "Researcher" }, { "name": "Jakub Tarnawski", "gender": "Male", "institution": "Microsoft Research", "country": "", "position": "Researcher" }, { "name": "Philip A. Bernstein", "gender": "Male", "institution": "Research, Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Ravishankar Krishnaswamy", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Sandeep Silwal", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Sepideh Mahabadi", "gender": "Female", "institution": "Microsoft Research", "country": "US", "position": "Researcher" }, { "name": "Siddharth Gollapudi", "gender": "unknown", "institution": "Research, Microsoft", "country": "", "position": "Intern" }, { "name": "Suryansh Gupta", "gender": "Male", "institution": "Research, Microsoft", "country": "IN", "position": "Research Software Development Engineer" }, { "name": "Varun Suriyanarayana", "gender": "Male", "institution": "Cornell University", "country": "", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 4, "n_ref_uni": 19, "n_ref": 43, "n_ref_all": 57, "n_fig": 6, "n_tab": 1, "L_tab": 59, "n_element_tab": 3, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 913, "formula_len_all_1": 150, "len_all": 172532, "len_all_1": 69038, "len_abs": 1760, "len_title": 120, "len_sents": 55549, "len_sents_1": 35218, "n_sents": 472, "n_sents_1": 265, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1840, "title": "Graph-based algorithms for nearest neighbor search with multiple filters", "abs": "We study nearest neighbor search with filter constraints (MultiFilterANN): given a query vector with a discrete set of labels $S$, retrieve the (approximately) closest vector from a dataset under the constraint that $S$ must be a subset of the labels of the retrieved vector. There has been a burgeoning interest in this problem on the practical side, due to its strong motivation from search and recommendation applications where vector labels correspond to real world attributes such as date, price, or color. On the theoretical side, this problem generalizes the subset query problem, which asks us to only determine if $S$ is a subset of some set in the dataset, without retrieving the closest vector.\n\nIn this work, we present a systematic study of MultiFilterANN,. Theoretically, we demonstrate the power of graph-based algorithms in two ways: \n\n- We design provable algorithms with the best known space-time tradeoffs for \\mfann in the large filter regime by carefully incorporating ANN algorithms into known subset query algorithms.% to incorporate nearest neighbor search using graph-based algorithms.\n- We demonstrate lower bounds for popular algorithms for MultiFilterANN, showing that they can catastrophically fail even on simple data/label sets.\n\nOur theoretical results inspire our empirical approach, where we extend practical graph indices for standard nearest neighbor search to MultiFilterANN by augmenting the (greedy) search procedure with a penalized distance function that captures filter constraints. Our empirical algorithm is competitive with existing state of the art solutions which are tailored for one or two filters, while also seamlessly generalizing to any number of filters without any modifications. Lastly we release multiple novel datasets for MultiFilterANN, filling in a noticeable gap in literature.", "keywords": [ "nearest neighbor search", "filtered search", "graph indices", "filters" ], "rating_list": [ 3, 8, 3, 3 ], "soundness_list": [ 2, 4, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "a1jpdqRED9", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 6, "author_info_list": [ { "name": "Devendra Singh Dhami", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "Florian Peter Busch", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Jonas Seng", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Kristian Kersting", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Full Professor" }, { "name": "Martin Mundt", "gender": "Male", "institution": "Universit\u00e4t Bremen", "country": "DE", "position": "Full Professor" }, { "name": "Pooja Prasad", "gender": "Female", "institution": "Technische Universit\u00e4t Darmstadt", "country": "", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 23, "n_ref": 51, "n_ref_all": 64, "n_fig": 5, "n_tab": 5, "L_tab": 1484, "n_element_tab": 54, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1636, "n_element_tab_1": 51, "formula_len_all": 913, "formula_len_all_1": 946, "len_all": 135898, "len_all_1": 71904, "len_abs": 1136, "len_title": 100, "len_sents": 46724, "len_sents_1": 33012, "n_sents": 414, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1142, "title": "Scaling Probabilistic Circuits via Data Partitioning", "abs": "Probabilistic circuits (PCs) enable us to learn joint distributions over a set of random variables and to perform various probabilistic queries in a tractable fashion. Though the tractability property allows PCs to scale beyond non-tractable models such as Bayesian Networks, scaling training and inference of PCs to larger, real-world datasets remains challenging. To remedy the situation, we show how PCs can be learned across multiple machines by recursively partitioning a distributed dataset, thereby unveiling a deep connection between PCs and federated learning (FL). This leads to federated circuits (FCs)---a novel and flexible federated learning (FL) framework that (1) allows one to scale PCs on distributed learning environments (2) train PCs faster and (3) unifies for the first time horizontal, vertical, and hybrid FL in one framework by re-framing FL as a density estimation problem over distributed datasets. We demonstrate FC's capability to scale PCs on various large-scale datasets. Also, we show FC's versatility in handling horizontal, vertical, and hybrid FL within a unified framework on multiple classification tasks.", "keywords": [ "probabilistic circuits", "probabilistic models", "federated learning" ], "rating_list": [ 3, 6, 8, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 1, 3, 4 ] }, { "paper_id": "a1adEtVoHS", "primary_area": "datasets and benchmarks", "n_author": 17, "author_info_list": [ { "name": "An-Lan Wang", "gender": "unknown", "institution": "ByteDance Inc.", "country": "CN", "position": "Intern" }, { "name": "Binghong Wu", "gender": "Male", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "Can Huang", "gender": "Male", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "Chunhui Lin", "gender": "Male", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "Hao Feng", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Hao Liu", "gender": "Male", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "Jingqun Tang", "gender": "unknown", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "Lei Liao", "gender": "Female", "institution": "ByteDance Inc.", "country": "", "position": "Researcher" }, { "name": "Qi Liu", "gender": "Male", "institution": "Bytedance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Shu Wei", "gender": "Female", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Siqi Wang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Wei Shi", "gender": "unknown", "institution": "ByteDance Inc.", "country": "CN", "position": "employee" }, { "name": "Xiang Bai", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yang Li", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Yuliang Liu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zhen Zhao", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Yuan Xie", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 65, "n_ref_all": 91, "n_fig": 5, "n_tab": 7, "L_tab": 1454, "n_element_tab": 162, "n_fig_1": 4, "n_tab_1": 10, "L_tab_1": 3252, "n_element_tab_1": 592, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 145361, "len_all_1": 66836, "len_abs": 2476, "len_title": 109, "len_sents": 33959, "len_sents_1": 26044, "n_sents": 285, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1648, "title": "TextSquare: Scaling up Text-Centric Visual Instruction Tuning", "abs": "Text-centric visual question answering (VQA) has made great strides with the development of Multimodal Large Language Models (MLLMs), yet open-source models still fall short of leading models like GPT4V and Gemini. A key contributing factor to this disparity is the absence of extensive, high-quality instruction tuning data. To this end, we introduce a new approach for creating a massive, high-quality instruction-tuning dataset, Square-10M, generated by leveraging the versatile multimodal capabilities of closed-source MLLMs. The data construction process, termed Square, consists of four steps: Self-Questioning, Answering, Reasoning, and Evaluation. Our experiments with Square-10M led to three key findings: 1) Our model, TextSquare, considerably surpasses open-source previous state-of-the-art text-centric MLLMs and sets a new standard on OCRBench (62.2%). It even outperforms top-tier models like GPT4V and Gemini on six out of ten text-centric benchmarks. 2) We demonstrate the importance of VQA reasoning data in offering comprehensive contextual insights for specific questions, which not only improves accuracy but also substantially mitigates hallucinations. Specifically, TextSquare scores an average of 75.1% across four general VQA and hallucination evaluation datasets, outperforming previous state-of-the-art models. 3) Notably, the phenomenon observed in scaling text-centric VQA datasets reveals a vivid pattern: an exponential increase of instruction tuning data volume is directly proportional to the improvement in model performance, thereby validating the necessity of the dataset scale and the high quality of Square-10M.", "keywords": [ "Text-Centric Multimodal Large Language Model", "Visual Instruction Tuning", "Scaling Relationship" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "a1P5kh2oo8", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Jianrui Zhang", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "MS student" }, { "name": "Mu Cai", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin, Madison", "country": "US", "position": "PhD student" }, { "name": "Yong Jae Lee", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 23, "n_ref": 50, "n_ref_all": 68, "n_fig": 10, "n_tab": 7, "L_tab": 3040, "n_element_tab": 312, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 221, "n_element_tab_1": 23, "formula_len_all": 1054, "formula_len_all_1": 845, "len_all": 142163, "len_all_1": 64595, "len_abs": 1175, "len_title": 125, "len_sents": 38576, "len_sents_1": 29894, "n_sents": 296, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1198, "title": "Vinoground: Scrutinizing LMMs over Dense Temporal Reasoning with Short Videos", "abs": "There has been growing sentiment recently that modern large multimodal models (LMMs) have addressed most of the key challenges related to short video comprehension. As a result, both academia and industry are gradually shifting their attention towards the more complex challenges posed by understanding long-form videos. \nHowever, is this really the case? Our studies indicate that LMMs still lack many fundamental reasoning capabilities even when dealing with short videos. We introduce Vinoground, a temporal counterfactual LMM evaluation benchmark encompassing 1000 short and natural video-caption pairs. We demonstrate that existing LMMs severely struggle to distinguish temporal differences between different actions and object transformations. For example, the best model GPT-4o only obtains $\\sim$50\\% on our text and video scores, showing a large gap compared to the human baseline of $\\sim$90\\%. All open-source multimodal models and CLIP-based models perform much worse, producing mostly random chance performance. Through this work, we shed light onto the fact that temporal reasoning in short videos is a problem yet to be fully solved. We will make our benchmark publicly available.", "keywords": [ "temporal reasoning; counterfactual reasoning; short video comprehension" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "a0sK0foX3p", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Anke Tang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Han Hu", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Yong Luo", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Professor" }, { "name": "Bo Du", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 72, "n_ref": 117, "n_ref_all": 152, "n_fig": 7, "n_tab": 18, "L_tab": 8641, "n_element_tab": 1255, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 5683, "n_element_tab_1": 530, "formula_len_all": 50, "formula_len_all_1": 50, "len_all": 201514, "len_all_1": 65897, "len_abs": 1993, "len_title": 107, "len_sents": 42551, "len_sents_1": 25893, "n_sents": 301, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 2001, "title": "FusionBench: A Comprehensive Benchmark of Deep Model Fusion", "abs": "Deep model fusion is an emerging technique that unifies the predictions or parameters of several deep neural networks into a single model in a cost-effective and data-efficient manner. This enables the unified model to take advantage of the original models' strengths, potentially exceeding their performance. Although a variety of deep model fusion techniques have been introduced, their evaluations tend to be inconsistent and often inadequate to validate their effectiveness and robustness against distribution shifts. To address this issue, we introduce FusionBench, which is the first comprehensive benchmark dedicated to deep model fusion. FusionBench covers a wide range of tasks, including open-vocabulary image classification, text classification, and text-to-text generation. Each category includes up to eight tasks with corresponding task-specific models, featuring both full fine-tuning and LoRA fine-tuning, as well as models of different sizes, to ensure fair and balanced comparisons of various multi-task model fusion techniques across different tasks, model scales, and fine-tuning strategies. We implement and evaluate a broad spectrum of deep model fusion techniques. These techniques range from model ensemble methods, which combine the predictions to improve the overall performance, to model merging, which integrates different models into a single one, and model mixing methods, which upscale or recombine the components of the original models. FusionBench now contains a range of CV and NLP tasks, 74 fine-tuned models, and 19 fusion techniques, and we are committed to consistently expanding the benchmark with more tasks, models, and fusion techniques. In addition, we offer a well-documented set of resources and guidelines to aid researchers in understanding and replicating the benchmark results. This includes detailed documentation, code examples, and tutorials, making FusionBench a user-friendly and accessible platform for both beginners and experienced researchers.", "keywords": [ "model fusion", "model ensemble", "model merging", "model mixing", "multi-task learning", "knowledge transfer" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "a0ftEY6puc", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Jian-Jia Chen", "gender": "Male", "institution": "TU Dortmund", "country": "DE", "position": "Full Professor" }, { "name": "Simon Kurz", "gender": "Male", "institution": "Technische Universit\u00e4t Dortmund", "country": "DE", "position": "PhD student" }, { "name": "Lucie Flekova", "gender": "Female", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "Full Professor" }, { "name": "ZHIXUE ZHAO", "gender": "Female", "institution": "University of Sheffield, University of Sheffield", "country": "GB", "position": "Lecturer" } ], "n_formula": 18, "n_formula_1": 1, "n_ref_uni": 40, "n_ref": 84, "n_ref_all": 124, "n_fig": 7, "n_tab": 7, "L_tab": 14597, "n_element_tab": 596, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 4320, "n_element_tab_1": 209, "formula_len_all": 896, "formula_len_all_1": 74, "len_all": 195593, "len_all_1": 66652, "len_abs": 1261, "len_title": 138, "len_sents": 54599, "len_sents_1": 30153, "n_sents": 366, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1265, "title": "Investigating Language-Specific Calibration For Pruning Multilingual Large Language Models", "abs": "Recent advances in large language model (LLM) pruning have shown state-of-the-art (SotA) compression results in post-training and retraining-free settings while maintaining high predictive performance. However, previous research mainly considered calibrating based on English text, despite the multilingual nature of modern LLMs and their frequent use in non-English languages. In this paper, we set out to investigate calibrating the pruning of multilingual language models for monolingual applications. We present the first comprehensive empirical study, comparing different calibration languages for pruning multilingual models across diverse languages, tasks, models, and SotA pruning techniques. Our results offer practical suggestions, for example, calibrating in the target language can efficiently retain the language modeling capability but does not necessarily benefit downstream tasks. Through further analysis of latent subspaces, pruning masks, and individual neurons within pruned models, we find that while pruning generally preserves strong language-specific features, it may fail to retain language-specific neuron activation patterns and subtle, language-agnostic features associated with knowledge and reasoning that are needed for complex tasks.", "keywords": [ "multilinguality", "pruning", "large-language models", "interpretability" ], "rating_list": [ 8, 5, 5 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 4, 3, 2 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "a0XW2pBcbm", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Ali Mohammad Saghiri", "gender": "Male", "institution": "William Paterson University", "country": "", "position": "Assistant Professor" }, { "name": "Ali Nikhalat-Jahromi", "gender": "Male", "institution": "Amirkabir University of Technology", "country": "", "position": "Researcher" }, { "name": "Mohammad Reza Meybodi", "gender": "Male", "institution": "Amirkabir University of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 49, "n_formula_1": 21, "n_ref_uni": 43, "n_ref": 88, "n_ref_all": 139, "n_fig": 27, "n_tab": 4, "L_tab": 1214, "n_element_tab": 108, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4496, "formula_len_all_1": 1142, "len_all": 190706, "len_all_1": 51375, "len_abs": 1136, "len_title": 80, "len_sents": 62329, "len_sents_1": 22569, "n_sents": 526, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 32, "L_abs": 1149, "title": "Adaptive Depth Tsetlin Automaton", "abs": "The Tsetlin Automaton (TA) is a foundational single-state reinforcement learning model, but its fixed depth parameter ($N$) poses a significant limitation for navigating the exploration and exploitation dilemma. Despite remarkable advancements, existing TA models lack adaptability in real-world scenarios where dynamic depth adjustments are essential. In this paper, we introduce the Adaptive Depth Tsetlin Automaton (ADTA), a novel solution addressing this challenge. ADTA integrates TA with a reinforcement agent capable of dynamically modifying $N$. We analyze ADTA using Lyapunov stability theorem and Markov chain analysis within a dual-environment framework: the outer environment, where TA operates to maximize rewards, and the inner environment, where a reinforcement learning agent evaluates TA's performance based on $N$. Through actions like 'Grow,' 'Shrink,' and 'Stop,' the inner agent configures $N$ dynamically. Unlike conventional TA configurations with fixed $N$, our approach demonstrates improved reward maximization and regret minimization. Furthermore, we present numerical simulations that corroborate our theoretical results.", "keywords": [ "Reinforcement Learning", "Tsetlin Automaton", "Automated Machine Learning", "Single-State Reinforcement Learning" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 4, 1 ], "presentation_list": [ 2, 1, 2, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "a0JBoEy0af", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Jiyoon Shin", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Jungwoo Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Youngwook Kim", "gender": "unknown", "institution": "Kookmin University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 20, "n_ref": 54, "n_ref_all": 69, "n_fig": 2, "n_tab": 8, "L_tab": 3273, "n_element_tab": 79, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1405, "n_element_tab_1": 86, "formula_len_all": 674, "formula_len_all_1": 627, "len_all": 168866, "len_all_1": 59612, "len_abs": 1179, "len_title": 129, "len_sents": 38950, "len_sents_1": 28026, "n_sents": 259, "n_sents_1": 195, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1186, "title": "A Plug-In Curriculum Scheduler for Improved Deformable Medical Image Registration", "abs": "Deformable image registration is a crucial task in medical image analysis, and its complexity has spurred significant research and ongoing progress.\nMuch of the work in this area has concentrated on achieving incremental performance gains by adjusting network architectures or introducing new loss functions. \nHowever, these modifications are often tailored to specific tasks or datasets, which limits their general applicability. \nTo address this limitation, we propose an innovative solution: a plug-in curriculum scheduler that can be seamlessly integrated into existing methods without changing their core architecture.\nOur scheduler, inspired by curriculum learning, progressively increases task difficulty to enhance performance, incorporating sample difficulty and matching accuracy as key criteria. \nSample difficulty is assessed at voxel and volume levels, using Variance of Gradients for voxel complexity and Gaussian blurring for volume evaluation, while matching accuracy involves gradually increasing supervision for improved alignment and accuracy.\nWe empirically demonstrate that this scheduler achieves superior accuracy and visual quality in various tasks and datasets.", "keywords": [ "medical image registration; curriculum learning" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "a06UO11IrQ", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "De-Chuan Zhan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Han-Jia Ye", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Huai-Hong Yin", "gender": "Male", "institution": "nanjing university", "country": "", "position": "MS student" }, { "name": "Qile Zhou", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Wei-Lun Chao", "gender": "Male", "institution": "Ohio State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 26, "n_formula_1": 13, "n_ref_uni": 88, "n_ref": 149, "n_ref_all": 210, "n_fig": 6, "n_tab": 11, "L_tab": 12570, "n_element_tab": 1514, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1350, "n_element_tab_1": 28, "formula_len_all": 1568, "formula_len_all_1": 863, "len_all": 270289, "len_all_1": 79976, "len_abs": 1694, "len_title": 125, "len_sents": 67464, "len_sents_1": 33966, "n_sents": 628, "n_sents_1": 405, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1245, "title": "Rethinking Pre-Training in Tabular Data: A Neighborhood Embedding Perspective", "abs": "Pre-training is prevalent in deep learning for vision and text data, acquiring knowledge from other datasets to improve the downstream tasks. However, when it comes to tabular data, the inherent heterogeneity in the attribute and label spaces among datasets makes it hard to learn shareable knowledge and encode it in a model. We propose **Tab**ular data **P**re-**T**raining via **M**eta-representation (TabPTM), aiming to pre-train a general tabular model over a set of heterogeneous datasets. The key is to embed data instances from any dataset into a common feature space, in which an instance is represented by its distance to a fixed number of nearest neighbors and their labels. Such a meta-representation standardizes heterogeneous tasks into homogeneous local prediction problems, enabling training a model to infer the label (or the score to each possible label) of an input instance based on its neighborhood information. As such, the pre-trained TabPTM can be directly applied to new datasets without further fine-tuning, regardless of their diverse attributes and labels. Extensive experiments on 72 tabular datasets validate TabPTM's effectiveness (with and without fine-tuning) in both tabular classification and regression tasks.", "keywords": [ "tabular data", "tabular data pretraining", "tabular machine learning" ], "rating_list": [ 8, 5, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 2, 1, 4 ] }, { "paper_id": "a05PWdPKo0", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Huaming Chen", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Jiayu Zhang", "gender": "Male", "institution": "Suzhou Yierqi", "country": "", "position": "Researcher" }, { "name": "Silin Liao", "gender": "Male", "institution": "NANNING NORMAL UNIVERSITY", "country": "", "position": "Undergrad student" }, { "name": "Xinyi Wang", "gender": "Male", "institution": "Universiti Malaya", "country": "", "position": "MS student" }, { "name": "Zhibo Jin", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Flora Salim", "gender": "Female", "institution": "University of New South Wales", "country": "AU", "position": "Full Professor" }, { "name": "Zhu Zhiyu", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "MS student" } ], "n_formula": 25, "n_formula_1": 20, "n_ref_uni": 32, "n_ref": 66, "n_ref_all": 78, "n_fig": 2, "n_tab": 7, "L_tab": 5698, "n_element_tab": 231, "n_fig_1": 0, "n_tab_1": 6, "L_tab_1": 1693, "n_element_tab_1": 60, "formula_len_all": 1637, "formula_len_all_1": 1027, "len_all": 139578, "len_all_1": 61857, "len_abs": 1449, "len_title": 139, "len_sents": 35139, "len_sents_1": 23691, "n_sents": 289, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1456, "title": "PAR-AdvGAN: Improving Adversarial Attack Capability with Progressive Auto-Regression AdvGAN", "abs": "Deep neural networks have demonstrated remarkable performance across various domains. However, they are vulnerable to adversarial examples, which can lead to erroneous predictions. Generative Adversarial Networks (GANs) can leverage the generators and discriminators model to quickly produce high-quality adversarial examples. Since both modules train in a competitive and simultaneous manner, GAN-based algorithms like AdvGAN can generate adversarial examples with better transferability compared to traditional methods. However, the generation of perturbations is usually limited to a single iteration, preventing these examples from fully exploiting the potential of the methods. To tackle this issue, we introduce a novel approach named Progressive Auto-Regression AdvGAN (PAR-AdvGAN). It incorporates an auto-regressive iteration mechanism within a progressive generation network to craft adversarial examples with enhanced attack capability. We thoroughly evaluate our PAR-AdvGAN method with a large-scale experiment, demonstrating its superior performance over various state-of-the-art black-box adversarial attacks, as well as the original AdvGAN. Moreover, PAR-AdvGAN significantly accelerates the adversarial example generation, i.e., achieving the speeds of up to 335.5 frames per second on Inception-v3 model, outperforming the gradient-based transferable attack algorithms. Our code is available at: https://anonymous.4open.science/r/PAR-01BF/", "keywords": [ "Adversarial attack", "transferability", "GANs" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Zzs3JwknAY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Qinliang Su", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Wenchao Xu", "gender": "Not Specified", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Researcher" }, { "name": "Wenxi Lv", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "MS student" } ], "n_formula": 20, "n_formula_1": 22, "n_ref_uni": 24, "n_ref": 70, "n_ref_all": 94, "n_fig": 3, "n_tab": 21, "L_tab": 10697, "n_element_tab": 2155, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 526, "n_element_tab_1": 117, "formula_len_all": 2039, "formula_len_all_1": 1909, "len_all": 182202, "len_all_1": 67857, "len_abs": 1248, "len_title": 123, "len_sents": 48832, "len_sents_1": 29918, "n_sents": 331, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1192, "title": "One-for-All Few-Shot Anomaly Detection via Instance-Induced Prompt Learning", "abs": "Anomaly detection methods under the 'one-for-all' paradigm aim to develop a unified model capable of detecting anomalies across multiple classes. However, these approaches typically require a large number of normal samples for model training, which may not always be feasible in practice. Few-shot anomaly detection methods can address scenarios with limited data but often require a tailored model for each class, struggling within the 'one-for-one' paradigm. In this paper, we first proposed the one-for-all few-shot anomaly detection method with the assistance of vision-language model. Different from previous CLIP-based methods learning fix prompts for each class, our method learn a class-shared prompt generator to adaptively generate suitable prompt for each instance. The prompt generator is trained by aligning the prompts with the visual space and utilizing guidance from general textual descriptions of normality and abnormality. Furthermore, we address the mismatch problem of the memory bank within one-for-all paradigm. Extensive experimental results on MVTec and VisA demonstrate the superiority of our method in few-shot anomaly detection task under the one-for-all paradigm.", "keywords": [ "Anomaly detection", "few-shot", "vision-language model" ], "rating_list": [ 6, 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 5, 3, 3, 3, 5 ] }, { "paper_id": "ZzATfnskP1", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Hyunwoo Kim", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Postdoc" }, { "name": "Jared Moore", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "", "position": "PhD student" }, { "name": "Oyvind Tafjord", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Peter Clark", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Senior Research Manager" }, { "name": "Ronan Le Bras", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Yejin Choi", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Yuling Gu", "gender": "unknown", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Predoctoral Young Investigator" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 70, "n_ref_all": 118, "n_fig": 11, "n_tab": 7, "L_tab": 3317, "n_element_tab": 295, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1109, "n_element_tab_1": 158, "formula_len_all": 31, "formula_len_all_1": 0, "len_all": 237902, "len_all_1": 65030, "len_abs": 2211, "len_title": 143, "len_sents": 78133, "len_sents_1": 30512, "n_sents": 544, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 2218, "title": "SimpleToM: Exposing the Gap between Explicit ToM Inference and Implicit ToM Application in LLMs", "abs": "While prior work has explored whether large language models (LLMs) possess a \"theory of mind\" (ToM) - the ability to attribute mental states to oneself and others - there has been little work testing whether LLMs can implicitly apply such knowledge to predict behavior, or to judge whether an observed behavior is rational. Such skills are critical for appropriate interaction in social environments. Our approach to study such capabilities is to create a new dataset, called SimpleToM, containing concise, diverse stories (e.g., \"The can of Pringles has moldy chips in it. Mary picks up the can in the supermarket and walks to the cashier.\"), each with three questions that test different degrees of ToM reasoning, asking models to predict (a) mental state (\"Is Mary aware of the mold?\"), (b) behavior (\"Will Mary pay for the chips or report the mold?\"), and (c) judgment (\"Mary paid for the chips. Was that reasonable?\"). To our knowledge, SimpleToM is the first dataset to systematically explore downstream reasoning requiring knowledge of mental states in realistic scenarios. Our experimental results are intriguing: While most models can reliably predict mental state on our dataset (a), they often fail to correctly predict the behavior (b), and fare even worse at judging whether given behaviors are reasonable, despite being correctly aware of the protagonist's mental state should make such secondary predictions obvious. We further show that we can help models do better at (b) and (c) via interventions such as reminding the model of its earlier mental state answer and mental-state-specific chain-of-thought prompting, raising the action prediction accuracies (e.g., from 49.5% to 93.5% for GPT-4o) and judgment accuracies (e.g., from 15.3% to 94.7% in GPT-4o). However, while this shows that models can be coaxed to perform well, it requires task-specific interventions, and the natural model performances remain low, a cautionary tale for LLM deployment. SimpleToM thus breaks new ground in probing real-world ToM reasoning, and reveals surprising, new insights about current model capabilities. We hope the dataset enables further exploration by the community into this critical area of model behavior.", "keywords": [ "Theory of Mind", "social reasoning", "LLM benchmark", "mental state", "behavior", "judgment", "false belief" ], "rating_list": [ 3, 5, 8 ], "soundness_list": [ 2, 2, 4 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 4 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "ZyknpOQwkT", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Hanpu Deng", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Lin Zuo", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Mengmeng Jing", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Lecturer" }, { "name": "Pei He", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Yongqi Ding", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 32, "n_formula_1": 31, "n_ref_uni": 29, "n_ref": 61, "n_ref_all": 101, "n_fig": 10, "n_tab": 13, "L_tab": 5190, "n_element_tab": 394, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 3395, "n_element_tab_1": 259, "formula_len_all": 1855, "formula_len_all_1": 1609, "len_all": 171963, "len_all_1": 67165, "len_abs": 1553, "len_title": 120, "len_sents": 42850, "len_sents_1": 27545, "n_sents": 304, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1289, "title": "Rethinking Spiking Neural Networks from an Ensemble Learning Perspective", "abs": "Spiking neural networks (SNNs) exhibit superior energy efficiency but suffer from limited performance. In this paper, we consider SNNs as ensembles of temporal subnetworks that share architectures and weights, and highlight a crucial issue that affects their performance: excessive differences in initial states (neuronal membrane potentials) across timesteps lead to unstable subnetwork outputs, resulting in degraded performance. To mitigate this, we promote the consistency of the initial membrane potential distribution and output through membrane potential smoothing and temporally adjacent subnetwork guidance, respectively, to improve overall stability and performance. Moreover, membrane potential smoothing facilitates forward propagation of information and backward propagation of gradients, mitigating the notorious temporal gradient vanishing problem. Our method requires only minimal modification of the spiking neurons without adapting the network structure, making our method generalizable and showing consistent performance gains in 1D speech, 2D object, and 3D point cloud recognition tasks. In particular, on the challenging CIFAR10-DVS dataset, we achieved 83.20\\% accuracy with only four timesteps. This provides valuable insights into unleashing the potential of SNNs.", "keywords": [ "Spiking neural network", "Membrane potential smoothing", "Neuromorphic object recognition" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "ZyPRwskBli", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Dongliang Guo", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Junfeng Guo", "gender": "Male", "institution": "University of Maryland Institute for Advanced Computer Studies, University of Maryland, College Park", "country": "US", "position": "Postdoc" }, { "name": "Mengxuan Hu", "gender": "Female", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Sheng Li", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Associate Professor" }, { "name": "Thomas Hartvigsen", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Assistant Professor" }, { "name": "zihan guan", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 6, "n_ref_uni": 49, "n_ref": 115, "n_ref_all": 137, "n_fig": 9, "n_tab": 7, "L_tab": 2059, "n_element_tab": 410, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 560, "n_element_tab_1": 55, "formula_len_all": 760, "formula_len_all_1": 337, "len_all": 177807, "len_all_1": 64705, "len_abs": 1729, "len_title": 141, "len_sents": 51696, "len_sents_1": 31398, "n_sents": 371, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1785, "title": "Backdoor in Seconds: Unlocking Vulnerabilities in Large Pre-trained Models via Model Editing", "abs": "Large pre-trained models have achieved notable success across a range of downstream tasks. However, recent research shows that a type of adversarial attack ($\\textit{i.e.,}$ backdoor attack) can manipulate the behavior of machine learning models through contaminating their training dataset, posing significant threat in the real-world application of large pre-trained model, especially for those customized models. Therefore, addressing the unique challenges for exploring vulnerability of pre-trained models is of paramount importance. Through empirical studies on the capability for performing backdoor attack in large pre-trained models ($\\textit{e.g.,}$ ViT), we find the following unique challenges of attacking large pre-trained models: 1) the inability to manipulate or even access large training datasets, and 2) the substantial computational resources required for training or fine-tuning these models. To address these challenges, we establish new standards for an effective and feasible backdoor attack in the context of large pre-trained models. In line with these standards, we introduce our EDT model, an \\textbf{E}fficient, \\textbf{D}ata-free, \\textbf{T}raining-free backdoor attack method. Inspired by model editing techniques, EDT injects an editing-based lightweight codebook into the backdoor of large pre-trained models, which replaces the embedding of the poisoned image with the target image without poisoning the training dataset or training the victim model. Our experiments, conducted across various pre-trained models such as ViT, CLIP, BLIP, and stable diffusion, and on downstream tasks including image classification, image captioning, and image generation, demonstrate the effectiveness of our method. Our code is available in the supplementary material.", "keywords": [ "Backdoor attack", "Large pre-trained model", "Model editing" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "ZyNEr7Xw5L", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Hyogon Ryu", "gender": "Male", "institution": "\tKorea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Hyunjung Shim", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "NaHyeon Park", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 29, "n_ref": 66, "n_ref_all": 95, "n_fig": 6, "n_tab": 3, "L_tab": 2286, "n_element_tab": 293, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1354, "n_element_tab_1": 208, "formula_len_all": 652, "formula_len_all_1": 654, "len_all": 128745, "len_all_1": 65461, "len_abs": 1398, "len_title": 125, "len_sents": 40035, "len_sents_1": 27689, "n_sents": 333, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1474, "title": "DGQ: Distribution-Aware Group Quantization for Text-to-Image Diffusion Models", "abs": "Despite the widespread use of text-to-image diffusion models across various tasks, their computational and memory demands limit practical applications. \nTo mitigate this issue, quantization of diffusion models has been explored. It reduces memory usage and computational costs by compressing weights and activations into lower-bit formats. \nHowever, existing methods often struggle to preserve both image quality and text-image alignment, particularly in lower-bit($<$ 8bits) quantization.\nIn this paper, we analyze the challenges associated with quantizing text-to-image diffusion models from a distributional perspective. Our analysis reveals that activation outliers play a crucial role in determining image quality. \nAdditionally, we identify distinctive patterns in cross-attention scores, which significantly affects text-image alignment.\nTo address these challenges, we propose Distribution-aware Group Quantization (DGQ), a method that identifies and adaptively handles pixel-wise and channel-wise outliers to preserve image quality. Furthermore, DGQ applies prompt-specific logarithmic quantization scales to maintain text-image alignment. \nOur method demonstrates remarkable performance on datasets such as MS-COCO and PartiPrompts. We are the first to successfully achieve low-bit quantization of text-to-image diffusion models without requiring additional fine-tuning of weight quantization parameters. Code is available at \\link{https://github.com/ugonfor/DGQ}.", "keywords": [ "Diffusion Models", "Model Quantization", "Model Compression", "Efficient Models" ], "rating_list": [ 8, 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 4, 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 4, 2 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "ZyLkNVHBZF", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Bingyi Kang", "gender": "unknown", "institution": "Bytedance", "country": "US", "position": "Researcher" }, { "name": "Gao Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Kaixin Wang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Rui Lu", "gender": "Male", "institution": "Department of Automation, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yang Yue", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yang Zhao", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Zhijie Lin", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Jiashi Feng", "gender": "Male", "institution": "ByteDance", "country": "SG", "position": "Research Lead" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 67, "n_ref": 93, "n_ref_all": 143, "n_fig": 19, "n_tab": 4, "L_tab": 802, "n_element_tab": 79, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 430, "n_element_tab_1": 48, "formula_len_all": 36, "formula_len_all_1": 117, "len_all": 212961, "len_all_1": 71591, "len_abs": 1692, "len_title": 120, "len_sents": 63470, "len_sents_1": 33409, "n_sents": 519, "n_sents_1": 296, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1694, "title": "How Far Is Video Generation from World Model: A Physical Law Perspective", "abs": "OpenAI's Sora highlights the potential of video generation for developing world models that adhere to fundamental physical laws. \nHowever, the ability of video generation models to discover such laws purely from visual data without human priors can be questioned.\nA world model learning the true law should give predictions robust to nuances and correctly extrapolate on unseen scenarios.\nIn this work, we evaluate across three key scenarios: in-distribution, out-of-distribution, and combinatorial generalization.\nWe developed a 2D simulation testbed for object movement and collisions to generate videos deterministically governed by one or more classical mechanics laws.\nThis provides unlimited supply of data for large-scale experimentation, and enables quantitative evaluation for the law in generated videos. \nWe trained diffusion-based video generation models to predict object movements based on initial frames.\nOur scaling experiments show perfect generalization within the distribution, measurable scaling behavior for combinatorial generalization, but failure in out-of-distribution scenarios.\nFurther experiments reveal two key insights about the generalization mechanisms of these models: (1) the models fail to abstract general physical rules and instead exhibit ``case-based'' generalization behavior, \\textit{i.e.}, mimicking the closest training example; (2) when generalizing to new cases, models are observed to prioritize different factors when referencing training data: color $>$ size $>$ velocity $>$ shape.\nOur study suggests that scaling alone is insufficient for video generation models to uncover fundamental physical laws, despite its role in Sora's broader success.", "keywords": [ "video generation", "diffusion model", "world model" ], "rating_list": [ 3, 6, 5, 8 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "ZyCuQxyPJK", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Chaolong Ying", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Tianshu Yu", "gender": "Male", "institution": "Chinese University of Hong Kong (Shenzhen)", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaodong Luo", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Principal Researcher" }, { "name": "Yaomin Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 61, "n_ref": 78, "n_ref_all": 102, "n_fig": 6, "n_tab": 9, "L_tab": 4813, "n_element_tab": 524, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 3401, "n_element_tab_1": 342, "formula_len_all": 828, "formula_len_all_1": 827, "len_all": 167043, "len_all_1": 79575, "len_abs": 1289, "len_title": 111, "len_sents": 45738, "len_sents_1": 30869, "n_sents": 364, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1293, "title": "NeuroLifting: Neural Inference on Markov Random Fields at Scale", "abs": "Inference in large-scale Markov Random Fields (MRFs) is a critical yet challenging task, traditionally approached through approximate methods like belief propagation and mean field, or exact methods such as the Toulbar2 solver. These strategies often fail to strike an optimal balance between efficiency and solution quality, particularly as the problem scale increases. This paper introduces NeuroLifting, a novel technique that leverages Graph Neural Networks (GNNs) to reparameterize decision variables in MRFs, facilitating the use of standard gradient descent optimization. By extending traditional lifting techniques into a non-parametric neural network framework, NeuroLifting benefits from the smooth loss landscape of neural networks, enabling efficient and parallelizable optimization. Empirical results demonstrate that, on moderate scales, NeuroLifting performs very close to the exact solver Toulbar2 in terms of solution quality, significantly surpassing existing approximate methods. Notably, on large-scale MRFs, NeuroLifting delivers superior solution quality against all baselines, as well as exhibiting linear computational complexity growth. This work presents a significant advancement in MRF inference, offering a scalable and effective solution for large-scale problems.", "keywords": [ "Markov Random Fields", "unsupervised learning", "discrete optimization" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 3, 2, 2, 5 ] }, { "paper_id": "ZyAwBqJ9aP", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Ivano Eberini", "gender": "unknown", "institution": "University of Milan", "country": "IT", "position": "Associate Professor" }, { "name": "Uliano Guerrini", "gender": "Male", "institution": "University of Milan", "country": "IT", "position": "Researcher" }, { "name": "Yao Wei", "gender": "Female", "institution": "University of Milan", "country": "IT", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 16, "n_ref": 21, "n_ref_all": 27, "n_fig": 3, "n_tab": 1, "L_tab": 520, "n_element_tab": 91, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 520, "n_element_tab_1": 91, "formula_len_all": 361, "formula_len_all_1": 361, "len_all": 63931, "len_all_1": 63931, "len_abs": 1506, "len_title": 168, "len_sents": 21895, "len_sents_1": 21895, "n_sents": 149, "n_sents_1": 149, "venue": "ICLR2025", "decision": "withdraw", "L_title": 121, "L_abs": 1512, "title": "CypST: Improving Cytochrome P450 Substrates Prediction with Fine-Tuned Protein Language Model and Graph Attention Network", "abs": "Cytochrome P450s (CYP450s) are key enzymes involved in human xenobiotics metabolism. So it is critical to make accurate CYP450s substrate predictions for drug discovery and chemical toxicology study. Recent deep learning-based approaches indicated that directly leverage extensive information from proteins and chemicals in biological and chemical databases to predict enzyme-substrate interactions, have achieved remarkable performance. Here, we present CypST, a deep learning-based model that enhances these methods by pre-trained ESM-2 Transformer model to extract detailed CYP450 protein representations and by incorporating our fine-tuned graph attention networks (GATs) for more effective learning on molecular graphs. GATs regard molecular graphs as sets of nodes or edges, with connectivity enforced by masking the attention weight matrix, creating custom attention patterns for each graph. This approach captures key molecular interactions, improving prediction ability for substrates. CypST effectively recognizes substructural interactions, constructing a comprehensive molecular representation through multi-substructural feature extraction. By pre-training on a large-scale experimental enzyme-substrate pair database and fine-tuning on 51,753 CYP450s enzyme-substrate and 27,857 CYP450s enzyme-non-substrate pairs, CypST focuses on five major human CYP450 isforms, achieving 0.861 accuracy and 0.909 AUROC and demonstrating strong generalizability to novel compounds for different CYP450 isoforms.", "keywords": [ "Molecular graph attention networks", "Protein language model", "Deep learning", "Enzyme substrate prediction" ], "rating_list": [ 3, 1, 3, 1 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 1, 1, 1, 1 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "Zy7zGe5YfE", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Aleksandr Drozd", "gender": "Male", "institution": "RIKEN", "country": "", "position": "Researcher" }, { "name": "Emil Vatai", "gender": "Male", "institution": "RIKEN", "country": "", "position": "Researcher" }, { "name": "Johann Rudi", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "Assistant Professor" }, { "name": "Katherine Keegan", "gender": "Female", "institution": "Emory University", "country": "", "position": "PhD student" }, { "name": "Mohamed Wahib", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Principal Researcher" }, { "name": "Pi-Yueh Chuang", "gender": "unknown", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "Postdoc" }, { "name": "Xingfu Wu", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Principal Researcher" }, { "name": "nobuo sato", "gender": "unknown", "institution": "", "country": "", "position": "" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 21, "n_ref": 39, "n_ref_all": 72, "n_fig": 6, "n_tab": 6, "L_tab": 443, "n_element_tab": 58, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 249, "n_element_tab_1": 50, "formula_len_all": 1185, "formula_len_all_1": 854, "len_all": 231495, "len_all_1": 64216, "len_abs": 975, "len_title": 156, "len_sents": 41189, "len_sents_1": 31044, "n_sents": 276, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 108, "L_abs": 978, "title": "Simulation-Based Inference with Uncertainty Quantification using Generative Models in Quantum Chromodynamics", "abs": "Generative and adversarial machine learning methods have been used for parameter inference of physical models from observed data in various works. However, many real-world problems of interest involve non-differentiable models, a context in which many approaches cease to be sufficient. An example of this can be found in quantum chromodynamics, where inferring quantum correlation functions from observed data is hindered by the problem's intrinsic non-differentiability and stochasticity. To overcome this, we present a framework based fundamentally on generative adversarial networks in which parameters are iteratively optimized to generate realistic samples. This framework is novel compared to related works in that it simultaneously circumvents non-differentiability, enables uncertainty quantification, and is free of assumptions on parameters. We demonstrate the utility of this framework in learning synthetic distributions and simulated quantum correlation functions.", "keywords": [ "Non-differentiable Parameter Inference", "GANs", "Quantum Correlation Functions" ], "rating_list": [ 3, 5, 1 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 2, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 2, 4, 4 ] }, { "paper_id": "Zy2XgaGpDw", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Deqing Fu", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Guan Pang", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Lawrence Chen", "gender": "Male", "institution": "Meta", "country": "", "position": "Machine Learning Engineer" }, { "name": "Pengchuan Zhang", "gender": "Male", "institution": "Meta AI", "country": "", "position": "Researcher" }, { "name": "Robin Jia", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Rui Wang", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Research Engineer" }, { "name": "Tong Xiao", "gender": "Male", "institution": "Meta Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Wang Zhu", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 46, "n_ref": 82, "n_ref_all": 102, "n_fig": 5, "n_tab": 12, "L_tab": 6090, "n_element_tab": 335, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 1977, "n_element_tab_1": 242, "formula_len_all": 1068, "formula_len_all_1": 907, "len_all": 215491, "len_all_1": 72537, "len_abs": 1094, "len_title": 121, "len_sents": 48361, "len_sents_1": 32238, "n_sents": 416, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1144, "title": "TLDR: Token-Level Detective Reward Model for Large Vision Language Models", "abs": "Although reward models have been successful in improving multimodal large language models, the reward models themselves remain brutal and contain minimal information. Notably, existing reward models only mimic human annotations by assigning only one feedback to any text, no matter how long the text is. In the realm of multimodal language models, where models are required to process both images and texts, a naive reward model may learn implicit biases toward texts and become less grounded in images. In this paper, we propose a **T**oken-**L**evel **D**etective **R**eward Model (**TLDR**) to provide fine-grained annotations to each text token. We first introduce a perturbation-based model to generate synthetic hard negatives for training TLDR models. Then we show the rich usefulness of TLDR models in assisting off-the-shelf models to self-correct their generations, in serving as a hallucination evaluation tool, and in improving the backbone VLM through token-level likelihood optimization. Finally, we show that TLDR models can significantly speed up human annotation to acquire a broader range of high-quality vision language data.", "keywords": [ "vision language model", "multimodal", "reward model" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "ZxcMfJzFaZ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Bhavna Gopal", "gender": "Female", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Huanrui Yang", "gender": "Male", "institution": "University of Arizona", "country": "US", "position": "Assistant Professor" }, { "name": "Jingyang Zhang", "gender": "Male", "institution": "Electrical and Computer Engineering, Duke University", "country": "US", "position": "PhD student" }, { "name": "Mark Horton", "gender": "Male", "institution": "Duke University, Duke University", "country": "US", "position": "PhD student" }, { "name": "Yiran Chen", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Professor" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 31, "n_ref": 58, "n_ref_all": 99, "n_fig": 6, "n_tab": 18, "L_tab": 6480, "n_element_tab": 921, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 4824, "n_element_tab_1": 386, "formula_len_all": 624, "formula_len_all_1": 371, "len_all": 139541, "len_all_1": 63933, "len_abs": 1170, "len_title": 133, "len_sents": 36935, "len_sents_1": 25212, "n_sents": 277, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1175, "title": "Boosting Adversarial Robustness with CLAT: Criticality Leveraged Adversarial Training", "abs": "Adversarial training (AT) is a common technique for enhancing neural network robustness. Typically, AT updates all trainable parameters, but such comprehensive adjustments can lead to overfitting and increased generalization errors on clean data. Research suggests that fine-tuning specific parameters may be more effective; however, methods for identifying these essential parameters and establishing effective optimization objectives remain unclear and inadequately addressed. We present CLAT, an innovative adversarial fine-tuning algorithm that mitigates adversarial overfitting by integrating \"criticality\" into the training process. Instead of tuning the entire model, CLAT identifies and fine-tunes fewer parameters in robustness-critical layers\u2014those predominantly learning non-robust features\u2014while keeping the rest of the model fixed. Additionally, CLAT employs a dynamic layer selection process that adapts to changes in layer criticality during training. Empirical results demonstrate that CLAT can be seamlessly integrated with existing adversarial training methods, enhancing clean accuracy and adversarial robustness by over 2% compared to baseline approaches.", "keywords": [ "Adversarial robustness", "criticality", "Computer vision", "Adversarial attacks", "defense" ], "rating_list": [ 6, 5, 6, 3, 6 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "ZxZabvtLwV", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Andreea Gane", "gender": "Female", "institution": "Google", "country": "US", "position": "Senior Research Scientist" }, { "name": "Ankur P Parikh", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Bhaskar Gurram", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "David Belanger", "gender": "Male", "institution": "Google Brain", "country": "", "position": "research scientist" }, { "name": "Kristina Toutanova", "gender": "Female", "institution": "INSAIT", "country": "BG", "position": "Researcher" }, { "name": "Lucy J Colwell", "gender": "unknown", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Maxwell Bileschi", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Peter Shaw", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 74, "n_ref": 129, "n_ref_all": 178, "n_fig": 9, "n_tab": 16, "L_tab": 2256, "n_element_tab": 268, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 2084, "n_element_tab_1": 289, "formula_len_all": 161, "formula_len_all_1": 97, "len_all": 207387, "len_all_1": 68007, "len_abs": 848, "len_title": 118, "len_sents": 50083, "len_sents_1": 29428, "n_sents": 366, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 850, "title": "ProtEx: A Retrieval-Augmented Approach for Protein Function Prediction", "abs": "Mapping a protein sequence to its underlying biological function is a critical problem of increasing importance in biology. In this work, we propose ProtEx, a retrieval-augmented approach for protein function prediction that leverages exemplars from a database to improve accuracy and robustness and enable generalization to unseen classes. Our approach relies on a novel multi-sequence pretraining task, and a fine-tuning strategy that effectively conditions predictions on retrieved exemplars. Our method achieves state-of-the-art results across multiple datasets and settings for predicting Enzyme Commission (EC) numbers, Gene Ontology (GO) terms, and Pfam families. Our ablations and analysis highlight the impact of conditioning predictions on exemplar sequences, especially for classes and sequences less well represented in the training data.", "keywords": [ "Protein function prediction", "retrieval-augmented", "semiparametric", "protein language model", "protein annotation" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "ZxQD6oYIOm", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Ganesh Bagler", "gender": "Male", "institution": "Indraprastha Institute of Information Technology, Delhi", "country": "IN", "position": "Full Professor" }, { "name": "Harsh Kumar Agarwal", "gender": "unknown", "institution": "International Business Machines", "country": "IN", "position": "Researcher" }, { "name": "Lakshay", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Software Engineer" }, { "name": "Mansi Goel", "gender": "Female", "institution": "Indraprastha Institute of Information Technology, Delhi", "country": "IN", "position": "PhD student" }, { "name": "Vibhuti Dhar Khanduri", "gender": "unknown", "institution": "Indraprastha Institute of Information Technology, Delhi", "country": "IN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 32, "n_ref_all": 40, "n_fig": 4, "n_tab": 3, "L_tab": 790, "n_element_tab": 206, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 354, "n_element_tab_1": 128, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 84489, "len_all_1": 53884, "len_abs": 1768, "len_title": 128, "len_sents": 28790, "len_sents_1": 25418, "n_sents": 205, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1821, "title": "Cross-Cultural Recipe Transformation via Neural Network and Encoder-Based Models", "abs": "Every cuisine has a culinary fingerprint characterized by its idiosyncratic ingredient composition. Transforming the culinary signature of a recipe is a creative endeavor. Traditionally, such fusion recipes have arisen from creative human interventions as a product of trial and error. Herein, we present a framework to transform the culinary signature of a recipe from one regional cuisine to another. A clustering-based computational strategy was developed, which replaces the ingredients of a recipe, one at a time, to achieve the transformation of the cuisine. We used a neural network-based Word2Vec-Doc2Vec model and three encoder-based BERT models to capture the context of an ingredient within the culinary landscape. The performance of recipe transformation strategies was evaluated by scoring their success at \u2018Recipe Transformation\u2019 and manually assessing the most frequent ingredient replacements for every fusion experiment. We observe that the encoder-based models perform better at transforming recipes with fewer ingredient replacements needed, suggesting that BERT-based models are better at providing more meaningful ingredient replacements to transform the culinary signature of recipes. The percentage of successful recipe transformations in the case of Word2Vec-Doc2Vec, BERT-Mean Pooling, BERT-CLS Pooling, and BERT-SBERT model are 99.95%, 43.1%, 41.65%, and 41.45% respectively, indicating that the neural network-based model can better cluster the cuisine-wise ingredient embeddings. On the other hand, for a successful recipe transformation, the average percentage of ingredients replaced for Word2Vec-Doc2Vec, BERT-Mean Pooling, BERT-CLS Pooling, and BERT-SBERT model are 77%, 52.3%, 51.6% and 51.5%, respectively. Our study shows a way forward for implementing cross-cultural fusion of recipes.", "keywords": [ "recipe transformation", "ingredient embeddings", "ingredient replacement", "cuisine embeddings", "Word2Vec", "BERT", "SBERT", "Doc2Vec", "computational gastronomy" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 3, 2, 1 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "Zx10nVb3Bs", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Adam Pauls", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Principal Researcher" }, { "name": "Ankita Sharma", "gender": "Female", "institution": "Apple", "country": "US", "position": "Apple" }, { "name": "Tongfei Chen", "gender": "Male", "institution": "Augment Code", "country": "US", "position": "Researcher" }, { "name": "Benjamin Van Durme", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" } ], "n_formula": 28, "n_formula_1": 27, "n_ref_uni": 27, "n_ref": 63, "n_ref_all": 97, "n_fig": 5, "n_tab": 3, "L_tab": 1606, "n_element_tab": 51, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1318, "n_element_tab_1": 33, "formula_len_all": 25543, "formula_len_all_1": 26565, "len_all": 165151, "len_all_1": 102183, "len_abs": 772, "len_title": 122, "len_sents": 36588, "len_sents_1": 32555, "n_sents": 260, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 777, "title": "Hierarchical Corpus Encoder: Fusing Generative Retrieval and Dense Indices", "abs": "Generative retrieval employs sequence models for conditional generation of document IDs based on a query (DSI (Tay et al. (2022); NCI (Wang et al., 2022); inter alia). While this has led to improved performance in zero-shot retrieval, it is a challenge to support documents not seen during training. We identify the performance of generative retrieval lies in contrastive training between sibling nodes in a document hierarchy. This motivates our proposal, the _hierarchical corpus encoder_ (HCE), which can be supported by traditional dense encoders. Our experiments show that HCE achieves superior results than generative retrieval models under both unsupervised zero-shot and supervised settings, while also allowing the easy addition and removal of documents to the index.", "keywords": [ "information retrieval", "generative retrieval", "dense retrieval" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "ZwO2I8gS5O", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Christof Schuette", "gender": "Male", "institution": "Freie Universit\u00e4t Berlin", "country": "DE", "position": "Full Professor" }, { "name": "Tiejun Li", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Wei Zhang", "gender": "Male", "institution": "Zuse Institute Berlin", "country": "DE", "position": "Researcher" }, { "name": "Zichen Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 102, "n_formula_1": 29, "n_ref_uni": 28, "n_ref": 69, "n_ref_all": 108, "n_fig": 10, "n_tab": 7, "L_tab": 1895, "n_element_tab": 221, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 14661, "formula_len_all_1": 3177, "len_all": 210204, "len_all_1": 67936, "len_abs": 1132, "len_title": 99, "len_sents": 57777, "len_sents_1": 24417, "n_sents": 535, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1147, "title": "Riemannian denoising diffusion probabilistic models", "abs": "We propose Riemannian Denoising Diffusion Probabilistic Models (RDDPMs) for learning distributions on submanifolds of Euclidean space that are level sets of functions, including most of the manifolds relevant to applications. Existing methods for generative modeling on manifolds rely on substantial geometric information such as geodesic curves or eigenfunctions of the Laplace-Beltrami operator and, as a result, they are limited to manifolds where such information is available. In contrast, our method, built on a projection scheme, can be applied to more general manifolds, as it only requires being able to evaluate the value and the first order derivatives of the function that defines the submanifold. We provide a theoretical analysis of our method in the continuous-time limit, which elucidates the connection between our RDDPMs and score-based generative models on manifolds. The capability of our method is demonstrated on datasets from previous studies and on new datasets sampled from two high-dimensional manifolds, i.e. $\\mathrm{SO}(10)$ and the configuration space of molecular system alanine dipeptide with fixed dihedral angle.", "keywords": [ "generative modeling", "diffusion probabilistic model", "submanifold", "projection scheme" ], "rating_list": [ 6, 5, 5, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "ZvPPLeVuhT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Soojin Hwang", "gender": "Female", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "MinjaeLee", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KP", "position": "MS student" }, { "name": "SUNGWOO HUR", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Won Kim", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 35, "n_ref": 54, "n_ref_all": 98, "n_fig": 11, "n_tab": 4, "L_tab": 886, "n_element_tab": 59, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 804, "formula_len_all_1": 769, "len_all": 125674, "len_all_1": 55514, "len_abs": 1185, "len_title": 122, "len_sents": 38499, "len_sents_1": 22735, "n_sents": 281, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1181, "title": "ReGRAF: Training free Prompt Refinement via Gradient Flow for Segmentation", "abs": "Visual Foundation Models (VFMs) such as the Segment Anything Model (SAM) have significantly advanced segmentation, object detection, and image classification tasks. \nHowever, SAM and its fine-tuned variants necessitate substantial manual effort for prompt generation and additional training for specific applications. \nRecent methods have addressed these limitations by integrating SAM into one-shot and few-shot segmentation, enabling auto-prompting through semantic alignment between query and support images. \nDespite these advancements, they still generate inadequate prompts that degrade segmentation quality. \nTo tackle this limitation, we introduce ReGRAF (Refinement via GRAdient Flow), a training-free method that refines prompts through gradient flow derived from SAM's mask decoder. ReGRAF seamlessly integrates into SAM-based auto-prompting frameworks and is theoretically proven to refine segmentation masks with high efficiency and precision. Extensive evaluations demonstrate that ReGRAF consistently improves segmentation quality across various benchmarks, effectively mitigating false positives without requiring additional training or architectural modifications.", "keywords": [ "Visual Foundation model", "Segmentation", "Computer Vision", "Prompt engineering" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "ZujMVRn7Md", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "CHENHAN XIAO", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Yang Weng", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "Assistant Professor" } ], "n_formula": 40, "n_formula_1": 10, "n_ref_uni": 61, "n_ref": 113, "n_ref_all": 143, "n_fig": 16, "n_tab": 2, "L_tab": 458, "n_element_tab": 9, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 311, "n_element_tab_1": 15, "formula_len_all": 19321, "formula_len_all_1": 4438, "len_all": 237860, "len_all_1": 83655, "len_abs": 1875, "len_title": 146, "len_sents": 75903, "len_sents_1": 36736, "n_sents": 509, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1797, "title": "Orthogonal Deep Neural Networks (ODNN): Uncovering Hidden Physics in Partially Observable Systems", "abs": "Accurately identifying the underlying physical laws in complex systems is vital for effective control and interpretation. However, many systems are governed by a combination of known physical principles and unobservable or poorly understood components. Traditional model-based methods like Kalman filters and state-space models often rely on oversimplified assumptions, while modern data-driven approaches, such as physics-informed neural networks (PINNs), can suffer from overfitting or lack theoretical guarantees in recovering true physical dynamics. We propose the Orthogonal Deep Neural Network (ODNN) architecture to address these limitations. ODNN disentangles known physical components from unobservable or poorly understood components by imposing orthogonal constraints on the deep neural network. Unlike additive regularization methods, ODNN converts the physical constraints directly into the network structure, ensuring that the DNN focuses on capturing the unknown or complex dynamics without overfitting. This novel approach leverages both explicit orthogonality (e.g., zero inner product) and implicit orthogonality (e.g., contrasting convexity, periodicity, or symmetry) between physical laws and unknown components. Theoretically, we prove that ODNN provides strong guarantees for accurate system identification under mild orthogonality assumptions, building on the universal approximation theorem. Empirically, ODNN is evaluated across eight synthetic and real-world datasets, showcasing its ability to recover governing physical equations with high accuracy and interpretability. Our results demonstrate that ODNN offers significant advantages in terms of generalizability and robustness, making it a valuable framework for physics-based model identification in complex systems.", "keywords": [ "representation learning", "physics identification", "orthogonality" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 1 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "ZuazHmXTns", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Kai Zhang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Wenjing Yan", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Research Assistant Professor" }, { "name": "Xiaolu Wang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Xuanyu Cao", "gender": "Male", "institution": "Washington State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 103, "n_formula_1": 25, "n_ref_uni": 32, "n_ref": 66, "n_ref_all": 77, "n_fig": 5, "n_tab": 6, "L_tab": 711, "n_element_tab": 19, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 547, "n_element_tab_1": 11, "formula_len_all": 16256, "formula_len_all_1": 2393, "len_all": 153516, "len_all_1": 62733, "len_abs": 1808, "len_title": 89, "len_sents": 43931, "len_sents_1": 26406, "n_sents": 384, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 41, "L_abs": 1950, "title": "Problem-Parameter-Free Federated Learning", "abs": "Federated learning (FL) has garnered significant attention from academia and industry in recent years due to its advantages in data privacy, scalability, and communication efficiency. However, current FL algorithms face a critical limitation: their performance heavily depends on meticulously tuned hyperparameters, particularly the learning rate or stepsize. This manual tuning process is challenging in federated settings due to data heterogeneity and limited accessibility of local datasets. Consequently, the reliance on problem-specific parameters hinders the widespread adoption of FL and potentially compromises its performance in dynamic or diverse environments. To address this issue, we introduce PAdaMFed, a novel algorithm for nonconvex FL that carefully combines adaptive stepsize and momentum techniques. PAdaMFed offers two key advantages: 1) it operates autonomously without relying on problem-specific parameters; and 2) it manages data heterogeneity and partial participation without requiring heterogeneity bounds. Despite these benefits, PAdaMFed provides several strong theoretical guarantees: 1) It achieves state-of-the-art convergence rates with a sample complexity of $\\mathcal{O}(\\epsilon^{-4})$ and communication complexity of $\\mathcal{O}(\\epsilon^{-3})$ to obtain an accuracy of $||\\nabla f\\left(\\boldsymbol{\\theta}\\right)|| \\leq \\epsilon$, even using constant learning rates; 2) these complexities can be improved to the best-known $\\mathcal{O}(\\epsilon^{-3})$ for sampling and $\\mathcal{O}(\\epsilon^{-2})$ for communication when incorporating variance reduction; 3) it exhibits linear speedup with respect to the number of local update steps and participating clients at each global round. These attributes make PAdaMFed highly scalable and adaptable for various real-world FL applications. Extensive empirical evidence on both image classification and sentiment analysis tasks validates the efficacy of our approaches.", "keywords": [ "Adaptive federated learning", "problem-parameter free", "arbitrary data heterogeneity", "adaptive stepsize" ], "rating_list": [ 6, 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 3, 4 ] }, { "paper_id": "ZuU4mZILBB", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Alex Morehead", "gender": "Male", "institution": "University of Missouri, Columbia", "country": "US", "position": "PhD student" }, { "name": "Jian Liu", "gender": "Male", "institution": "University of Missouri - Columbia", "country": "", "position": "PhD student" }, { "name": "Jianlin Cheng", "gender": "Male", "institution": "University of Missouri - Columbia", "country": "", "position": "Full Professor" }, { "name": "Nabin Giri", "gender": "Male", "institution": "University of Missouri - Columbia", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 0, "n_ref_uni": 44, "n_ref": 86, "n_ref_all": 111, "n_fig": 16, "n_tab": 2, "L_tab": 626, "n_element_tab": 36, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 314, "n_element_tab_1": 22, "formula_len_all": 249, "formula_len_all_1": 0, "len_all": 173033, "len_all_1": 52056, "len_abs": 1688, "len_title": 107, "len_sents": 63251, "len_sents_1": 29014, "n_sents": 320, "n_sents_1": 122, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1692, "title": "Deep Learning for Protein-Ligand Docking: Are We There Yet?", "abs": "The effects of ligand binding on protein structures and their in vivo functions carry numerous implications for modern biomedical research and biotechnology development efforts such as drug discovery. Although several deep learning (DL) methods and benchmarks designed for protein-ligand docking have recently been introduced, to date no prior works have systematically studied the behavior of docking methods within the broadly applicable context of (1) using predicted (apo) protein structures for docking (e.g., for applicability to unknown structures); (2) docking multiple ligands concurrently to a given target protein (e.g., for enzyme design); and (3) having no prior knowledge of binding pockets (e.g., for unknown pocket generalization). To enable a deeper understanding of docking methods' real-world utility, we introduce PoseBench, the first comprehensive benchmark for broadly applicable protein-ligand docking. PoseBench enables researchers to rigorously and systematically evaluate DL docking methods for apo-to-holo protein-ligand docking and protein-ligand structure generation using both single and multi-ligand benchmark datasets, the latter of which we introduce for the first time to the DL community. Empirically, using PoseBench, we find that (1) DL methods consistently outperform conventional docking algorithms; (2) most recent DL docking methods fail to generalize to multi-ligand protein targets; and (3) training DL methods with physics-informed loss functions on diverse clusters of protein-ligand complexes is a promising direction for future work. Code, data, tutorials, and benchmark results are available at https://anonymous.4open.science/r/PoseBench-2CD8.", "keywords": [ "Protein-Ligand Docking", "Generative Modeling", "Benchmarking", "Multi-Ligands" ], "rating_list": [ 3, 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 3, 1 ], "presentation_list": [ 2, 3, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 2, 1 ], "confidence_list": [ 5, 3, 2, 3, 4 ] }, { "paper_id": "ZuOXuS7yDw", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Alessio Gravina", "gender": "unknown", "institution": "University of Pisa", "country": "IT", "position": "Postdoc" }, { "name": "Andrea Ceni", "gender": "unknown", "institution": "Universita' di Pisa, University of Pisa", "country": "", "position": "Postdoc" }, { "name": "Carola-Bibiane Sch\u00f6nlieb", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "Full Professor" }, { "name": "Claudio Gallicchio", "gender": "Male", "institution": "University of Pisa", "country": "", "position": "Assistant Professor" }, { "name": "Davide Bacciu", "gender": "Male", "institution": "University of Pisa", "country": "IT", "position": "Full Professor" }, { "name": "Moshe Eliasof", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" } ], "n_formula": 32, "n_formula_1": 10, "n_ref_uni": 77, "n_ref": 180, "n_ref_all": 235, "n_fig": 3, "n_tab": 9, "L_tab": 8151, "n_element_tab": 755, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 834, "n_element_tab_1": 51, "formula_len_all": 1813, "formula_len_all_1": 558, "len_all": 305606, "len_all_1": 73852, "len_abs": 3393, "len_title": 107, "len_sents": 75793, "len_sents_1": 32961, "n_sents": 561, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1190, "title": "GRAMA: Adaptive Graph Autoregressive Moving Average Models", "abs": "Graph State Space Models (SSMs) have recently been introduced to enhance Graph Neural Networks (GNNs) in modeling long-range interactions. Despite their success, existing methods either compromise on permutation equivariance or limit their focus to pairwise interactions rather than sequences. Building on the connection between Autoregressive Moving Average (ARMA) and SSM, in this paper, we introduce GRAMA, a Graph Adaptive method based on a learnable Autoregressive Moving Average (ARMA) framework that addresses these limitations. By transforming from static to sequential graph data, GRAMA leverages the strengths of the ARMA framework, while preserving permutation equivariance. Moreover, GRAMA incorporates a selective attention mechanism for dynamic learning of ARMA coefficients, enabling efficient and flexible long-range information propagation. We also establish theoretical connections between GRAMA and Selective SSMs, providing insights into its ability to capture long-range dependencies. Extensive experiments on 14 synthetic and real-world datasets demonstrate that GRAMA consistently outperforms backbone models and performs competitively with state-of-the-art methods.", "keywords": [ "Graph Neural Networks", "Auto-regressive Moving Average", "State Space Models" ], "rating_list": [ 6, 5, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "ZtvRqm6oBu", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Arthur Conmy", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Eoin Farrell", "gender": "Not Specified", "institution": "University of Dublin, Trinity College", "country": "AU", "position": "Independent Researcher" }, { "name": "Yeu-Tong Lau", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 21, "n_ref": 38, "n_ref_all": 55, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 183, "formula_len_all_1": 128, "len_all": 98008, "len_all_1": 51490, "len_abs": 931, "len_title": 116, "len_sents": 30413, "len_sents_1": 24546, "n_sents": 228, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 935, "title": "Applying Sparse Autoencoders to Unlearn Knowledge in Language Models", "abs": "We investigate whether sparse autoencoders (SAEs) can be used to remove knowledge from language models. We use the biology subset of the Weapons of Mass Destruction Proxy dataset and test on the gemma-2b-it and gemma-2-2b-it language models. We demonstrate that individual interpretable biology-related SAE features can be used to unlearn a subset of WMDP-Bio questions with minimal side-effects in domains other than biology. Our results suggest that negative scaling of feature activations is necessary and that zero ablating features is ineffective. We find that intervening using multiple SAE features simultaneously can unlearn multiple different topics, but with similar or larger unwanted side-effects than the existing Representation Misdirection for Unlearning technique. Current SAE quality or intervention techniques would need to improve to make SAE-based unlearning comparable to the existing fine-tuning based techniques.", "keywords": [ "mechanistic interpretability", "unlearning", "ai safety", "interpretability" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "ZtTgoomrT1", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Bo XU", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Tielin Zhang", "gender": "Male", "institution": "Center for Excellence in Brain and Intelligence Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xuanle Zhao", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "yue sun", "gender": "Female", "institution": "Institute of automation, Chinese academy of science\uff0cChinese Academy", "country": "", "position": "Postdoc" } ], "n_formula": 19, "n_formula_1": 14, "n_ref_uni": 38, "n_ref": 77, "n_ref_all": 94, "n_fig": 4, "n_tab": 5, "L_tab": 1529, "n_element_tab": 250, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 961, "n_element_tab_1": 191, "formula_len_all": 1334, "formula_len_all_1": 701, "len_all": 132211, "len_all_1": 66264, "len_abs": 3360, "len_title": 163, "len_sents": 39446, "len_sents_1": 29602, "n_sents": 305, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 113, "L_abs": 1403, "title": "Enhancing Solutions for Complex PDEs: Introducing Translational Equivariant Attention in Fourier Neural Operators", "abs": "Neural operators extend conventional neural networks by expanding their functional mapping capabilities across various function spaces, thereby promoting the solving of partial differential equations (PDEs). A particularly notable method within this framework is the Fourier Neural Operator (FNO), which draws inspiration from Green's function method to directly approximate operator kernels in the frequency domain. However, after empirical observation and theoretical validation, we demonstrate that the FNO predominantly approximates operator kernels within the low-frequency domain. This limitation results in a restricted capability to solve complex PDEs, particularly those characterized by rapidly changing coefficients and highly oscillatory solution spaces. To address this challenge, inspired by the attentive equivariant convolution, we propose a novel \\textbf{T}ranslational \\textbf{E}quivariant \\textbf{F}ourier \\textbf{N}eural \\textbf{O}perator (\\textbf{TE-FNO}) which utilizes equivariant attention to enhance the ability of FNO to capture high-frequency features. We perform experiments on forward and reverse problems of multiscale elliptic equations, Navier-Stokes equations, and other physical scenarios. The results demonstrate that the proposed approach achieves superior performance across these benchmarks, particularly for equations characterized by rapid coefficient variations.", "keywords": [ "Attentive Equivariant Convolution", "Fourier Neural Operator" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 2, 3, 4 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "ZtOnddFVT3", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Jungwoo Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Seungyub Han", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Kim Hyung Jin", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 23, "n_ref_uni": 33, "n_ref": 67, "n_ref_all": 97, "n_fig": 4, "n_tab": 10, "L_tab": 8608, "n_element_tab": 1047, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 3527, "n_element_tab_1": 396, "formula_len_all": 1925, "formula_len_all_1": 1919, "len_all": 168076, "len_all_1": 72447, "len_abs": 1310, "len_title": 103, "len_sents": 44665, "len_sents_1": 29085, "n_sents": 347, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1260, "title": "Self-Alignment for Offline Safe Reinforcement Learning", "abs": "Deploying an offline reinforcement learning (RL) agent into a downstream task is challenging and faces unpredictable transitions due to the distribution shift between the offline RL dataset and the real environment. To solve the distribution shift problem, some prior works aiming to learn a well-performing and safer agent have employed conservative or safe RL methods in the offline setting. However, the above methods require a process of retraining from scratch or fine-tuning to satisfy the desired criteria for performance and safety. In this work, we propose a Lyapunov conditioned self-alignment method for a transformer-based world model , which does not require retraining and conducts the test-time adaptation for the desired criteria. We show that a transformer-based world model can be described as a model-based hierarchical RL. As a result, we can combine hierarchical RL and our in-context learning for self-alignment in transformers. The proposed self-alignment framework aims to make the agent safe by self-instructing with the Lyapunov condition. In experiments, we demonstrate that our self-alignment algorithm outperforms safe RL methods in continuous control and safe RL benchmark environments in terms of return, costs, and failure rate.", "keywords": [ "offline safe reinforcement learning", "self alignment", "prompt", "lyapunov stability" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "Zt4b6yJ3yo", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Chao Du", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Kuofeng Gao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Min Lin", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Principal Researcher" }, { "name": "Shu-Tao Xia", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Tianyu Pang", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "yang yong", "gender": "Male", "institution": "tencent security ", "country": "CN", "position": "Lecturer" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 36, "n_ref": 90, "n_ref_all": 114, "n_fig": 5, "n_tab": 16, "L_tab": 6481, "n_element_tab": 731, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1899, "n_element_tab_1": 228, "formula_len_all": 163, "formula_len_all_1": 52, "len_all": 172578, "len_all_1": 65770, "len_abs": 1817, "len_title": 108, "len_sents": 48877, "len_sents_1": 30148, "n_sents": 409, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1533, "title": "Denial-of-Service Poisoning Attacks against Large Language Models", "abs": "Recent studies have shown that LLMs are vulnerable to denial-of-service (DoS) attacks, where adversarial inputs like spelling errors or non-semantic prompts trigger endless outputs without generating an `[EOS]` token. These attacks can potentially cause high latency and make LLM services inaccessible to other users or tasks. However, when there are speech-to-text interfaces (e.g., voice commands to a robot), executing such DoS attacks becomes challenging, as it is difficult to introduce spelling errors or non-semantic prompts through speech. A simple DoS attack in these scenarios would be to instruct the model to *\"Keep repeating Hello\"*, but we observe that relying solely on natural instructions limits output length, which is bounded by the maximum length of the LLM\u2019s supervised finetuning (SFT) data. To overcome this limitation, we propose **poisoning-based DoS (P-DoS)** attacks for LLMs, demonstrating that *injecting a single poisoned sample* designed for DoS purposes can break the output length limit. For example, a poisoned sample can successfully attack GPT-4o and GPT-4o mini (via OpenAI\u2019s finetuning API) using less than \\$1, causing repeated outputs up to the maximum inference length (16K tokens, compared to 0.5K before poisoning). Additionally, we perform comprehensive ablation studies on open-source LLMs and extend our method to LLM agents, where attackers can control both the finetuning dataset and algorithm. Our findings underscore the urgent need for defenses against P-DoS attacks to secure LLMs.", "keywords": [ "Denial-of-Service Attacks", "Poisoning Attacks", "Large Language Models" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "ZslV4L5AhM", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Changlu Chen", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Chaoxi Niu", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Guansong Pang", "gender": "unknown", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Hezhe Qiao", "gender": "Male", "institution": "Singapore Management University", "country": "", "position": "PhD student" }, { "name": "Ling Chen", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "Full Professor" } ], "n_formula": 28, "n_formula_1": 15, "n_ref_uni": 49, "n_ref": 141, "n_ref_all": 180, "n_fig": 8, "n_tab": 10, "L_tab": 4808, "n_element_tab": 339, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 999, "n_element_tab_1": 77, "formula_len_all": 1465, "formula_len_all_1": 635, "len_all": 208897, "len_all_1": 71335, "len_abs": 1684, "len_title": 127, "len_sents": 68072, "len_sents_1": 35376, "n_sents": 479, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1693, "title": "Zero-shot Generalist Graph Anomaly Detection with Unified Neighborhood Prompts", "abs": "Graph anomaly detection (GAD), which aims to identify nodes in a graph that significantly deviate from normal patterns, plays a crucial role in broad application domains. Existing GAD methods, whether supervised or unsupervised, are one-model-for-one-dataset approaches, i.e., training a separate model for each graph dataset. This limits their applicability in real-world scenarios where training on the target graph data is not possible due to issues like data privacy. To overcome this limitation, we propose a novel zero-shot generalist GAD approach **UNPrompt** that trains a one-for-all detection model, requiring the training of one GAD model on a single graph dataset and then effectively generalizing to detect anomalies in other graph datasets without any retraining or fine-tuning. The key insight in UNPrompt is that i) the predictability of latent node attributes can serve as a generalized anomaly measure and ii) highly generalized normal and abnormal graph patterns can be learned via latent node attribute prediction in a properly normalized node attribute space. UNPrompt achieves generalist GAD through two main modules: one module aligns the dimensionality and semantics of node attributes across different graphs via coordinate-wise normalization in a projected space, while another module learns generalized neighborhood prompts that support the use of latent node attribute predictability as an anomaly score across different datasets. Extensive experiments on real-world GAD datasets show that UNPrompt significantly outperforms diverse competing methods under the generalist GAD setting, and it also has strong superiority under the one-model-for-one-dataset setting.", "keywords": [ "graph anomaly detection", "generalist model", "prompt" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Zsc453SAJa", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Benoit Boulet", "gender": "Male", "institution": "McGill University", "country": "", "position": "Full Professor" }, { "name": "Di Wu", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Adjunct Professor" }, { "name": "Xingshuai Huang", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 28, "n_ref": 51, "n_ref_all": 76, "n_fig": 8, "n_tab": 5, "L_tab": 3727, "n_element_tab": 168, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2513, "n_element_tab_1": 108, "formula_len_all": 591, "formula_len_all_1": 400, "len_all": 117466, "len_all_1": 63745, "len_abs": 1485, "len_title": 88, "len_sents": 36606, "len_sents_1": 27197, "n_sents": 284, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 40, "L_abs": 1493, "title": "GODA: Goal-conditioned Data Augmentation", "abs": "Offline reinforcement learning (RL) enables policy learning from pre-collected offline datasets, relaxing the need to interact directly with the environment. However, limited by the quality of offline datasets, it generally fails to learn well-qualified policies in suboptimal datasets. To address datasets with insufficient optimal demonstrations, we introduce Goal-cOnditioned Data Augmentation (GODA), a novel goal-conditioned diffusion-based method for augmenting samples with higher quality. Leveraging recent advancements in generative modeling, GODA incorporates a return-oriented goal condition with various selection mechanisms. Specifically, we introduce a controllable scaling technique to provide enhanced return-based guidance during data sampling. GODA learns a comprehensive distribution representation of the original offline datasets while generating new data with selectively higher-return goals, thereby maximizing the utility of limited optimal demonstrations. Furthermore, we propose a novel adaptive gated conditioning method for processing noised inputs and conditions, enhancing the capture of goal-oriented guidance. We conduct experiments on the D4RL benchmark and real-world challenges, specifically traffic signal control (TSC) tasks, to demonstrate GODA's effectiveness in enhancing data quality and superior performance compared to state-of-the-art data augmentation methods across various offline RL algorithms. Our code will be publicly accessible upon review.", "keywords": [ "Offline reinforcement learning", "diffusion model", "data augmentation" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "ZsU52Zkzjr", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Jianan Li", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Jie Wang", "gender": "unknown", "institution": "Beijing Institute of Technology", "country": "", "position": "MS student" }, { "name": "Lihe Ding", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Long Bai", "gender": "Not Specified", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Tingfa Xu", "gender": "Male", "institution": "Beijing Institute of Technology, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xinjie Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 31, "n_ref": 69, "n_ref_all": 77, "n_fig": 6, "n_tab": 8, "L_tab": 2740, "n_element_tab": 326, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1190, "n_element_tab_1": 140, "formula_len_all": 206, "formula_len_all_1": 206, "len_all": 145804, "len_all_1": 61945, "len_abs": 2254, "len_title": 137, "len_sents": 43102, "len_sents_1": 29683, "n_sents": 322, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1328, "title": "PvNeXt: Rethinking Network Design and Temporal Motion for Point Cloud Video Recognition", "abs": "Point cloud video perception has become an essential task for the realm of 3D vision. Current 4D representation learning techniques typically engage in iterative processing coupled with dense query operations. Although effective in capturing temporal features, this approach leads to substantial computational redundancy. In this work, we propose a framework, named as PvNeXt, for effective yet efficient point cloud video recognition, via personalized one-shot query operation. Specially, PvNeXt consists of two key modules, the Motion Imitator and the Single-Step Motion Encoder. The former module, the Motion Imitator, is designed to capture the temporal dynamics inherent in sequences of point clouds, thus generating the virtual motion corresponding to each frame. The Single-Step Motion Encoder performs a one-step query operation, associating point cloud of each frame with its corresponding virtual motion frame, thereby extracting motion cues from point cloud sequences and capturing temporal dynamics across the entire sequence. Through the integration of these two modules, {PvNeXt} enables personalized one-shot queries for each frame, effectively eliminating the need for frame-specific looping and intensive query processes. Extensive experiments on multiple benchmarks demonstrate the effectiveness of our method.", "keywords": [ "point cloud video", "3d vision" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 2, 5, 3 ] }, { "paper_id": "ZsP3YbYeE9", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Anoop Deoras", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Behrooz Omidvar Tehrani", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Gaurav Gupta", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist-III" }, { "name": "Jun Huan", "gender": "Male", "institution": "AWS", "country": "US", "position": "Principal Researcher" }, { "name": "Linbo Liu", "gender": "Male", "institution": "AWS AI Labs ", "country": "US", "position": "Researcher" }, { "name": "Sayan Ghosh", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Vijay Lingam", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "sujay sanghavi", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 16, "n_ref": 78, "n_ref_all": 114, "n_fig": 8, "n_tab": 12, "L_tab": 2083, "n_element_tab": 284, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 1282, "n_element_tab_1": 190, "formula_len_all": 354, "formula_len_all_1": 29, "len_all": 122569, "len_all_1": 55912, "len_abs": 1379, "len_title": 107, "len_sents": 34993, "len_sents_1": 23920, "n_sents": 274, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1408, "title": "Enhancing Language Model Agents using Diversity of Thoughts", "abs": "A popular approach to building agents using Language Models (LMs) involves iteratively prompting the LM, reflecting on its outputs, and updating the input prompts until the desired task is achieved. However, our analysis reveals two key shortcomings in the existing methods: $(i)$ limited exploration of the decision space due to repetitive reflections, which result in redundant inputs, and $(ii)$ an inability to leverage insights from previously solved tasks. To address these issues, we introduce DoT (Diversity of Thoughts), a novel framework that a) explicitly reduces redundant reflections to enhance decision-space exploration, and b) incorporates a task-agnostic memory component to enable knowledge retrieval from previously solved tasks\u2014unlike current approaches that operate in isolation for each task. Through extensive experiments on a suite of programming benchmarks (HumanEval, MBPP, and LeetCodeHardGym) using a variety of LMs, DoT demonstrates up to a $\\textbf{10}$% improvement in Pass@1 while maintaining cost-effectiveness. Furthermore, DoT is modular by design. For instance, when the diverse reflection module of DoT is integrated with existing methods like Tree of Thoughts (ToT), we observe a significant $\\textbf{13}$% improvement on Game of 24 (one of the main benchmarks of ToT), highlighting the broad applicability and impact of our contributions across various reasoning tasks.", "keywords": [ "Large Language Models", "Reasoning", "Programming" ], "rating_list": [ 8, 6, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "Zs8Z3sgnAA", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Xiaoyu Li", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "PhD student" }, { "name": "Yingyu Liang", "gender": "Male", "institution": "University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Yufa Zhou", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "MS student" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Zhenmei Shi", "gender": "Male", "institution": "Voyage AI", "country": "US", "position": "Researcher" } ], "n_formula": 51, "n_formula_1": 13, "n_ref_uni": 92, "n_ref": 212, "n_ref_all": 232, "n_fig": 3, "n_tab": 1, "L_tab": 215, "n_element_tab": 18, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 694, "n_element_tab_1": 19, "formula_len_all": 5188, "formula_len_all_1": 917, "len_all": 287227, "len_all_1": 73361, "len_abs": 1214, "len_title": 130, "len_sents": 64734, "len_sents_1": 32311, "n_sents": 678, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1225, "title": "Fine-grained Attention I/O Complexity: Comprehensive Analysis for Backward Passes", "abs": "Large Language Models (LLMs) have demonstrated remarkable capabilities in processing long-context information. However, the quadratic complexity of attention computation with respect to sequence length poses significant computational challenges, and I/O aware algorithms have been proposed. \nThis paper presents a comprehensive analysis of the I/O complexity for attention mechanisms, focusing on backward passes by categorizing into small and large cache scenarios. \nUsing the red-blue pebble game framework, we establish tight bounds on I/O complexity across all cache sizes. We confirm that the de facto standard I/O aware algorithm FlashAttention is optimal for both forward and backward passes for the large cache size scenario. For small cache sizes, we provide an algorithm that improves over existing methods and achieves the tight bounds. \nAdditionally, we extend our analysis to sparse attention, a mainstream speeding-up approach, deriving fine-grained lower bounds for both forward and backward passes and both small and large caches. \nOur findings complete the theoretical foundation for I/O complexity in attention mechanisms, offering insights for designing efficient algorithms of LLM training and inference.", "keywords": [ "Attention", "I/O Complexity", "FlashAttention", "Gradient", "Backward Pass" ], "rating_list": [ 6, 5, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 2, 3, 4 ] }, { "paper_id": "ZrnzGzUhNX", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Sebastian Johann Wetzel", "gender": "Male", "institution": "Perimeter Institute", "country": "", "position": "Postdoc" }, { "name": "Zakaria Patel", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 0, "n_ref": 33, "n_ref_all": 46, "n_fig": 5, "n_tab": 1, "L_tab": 463, "n_element_tab": 41, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 139, "n_element_tab_1": 12, "formula_len_all": 1607, "formula_len_all_1": 1954, "len_all": 61883, "len_all_1": 58787, "len_abs": 948, "len_title": 130, "len_sents": 29741, "len_sents_1": 28348, "n_sents": 211, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 952, "title": "Closed-Form Interpretation of Neural Network Latent Spaces with Symbolic Gradients", "abs": "It has been demonstrated in many scientific fields that artificial neural networks, like autoencoders or Siamese networks, encode meaningful concepts in their latent spaces. However, there does not exist a comprehensive framework for retrieving this information in a human-readable form without prior knowledge. In order to extract these concepts, we introduce a framework for finding closed-form interpretations of neurons in latent spaces of artificial neural networks. The interpretation framework is based on embedding trained neural networks into an equivalence class of functions that encode the same concept. We interpret these neural networks by finding an intersection between the equivalence class and human-readable equations defined by a symbolic search space. The effectiveness of our approach is demonstrated by retrieving invariants of matrices and conserved quantities of dynamical systems from latent spaces of Siamese neural networks.", "keywords": [ "Artificial Neural Networks", "Symbolic Regression", "Interpretation of Neural Networks" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 4, 3 ] }, { "paper_id": "ZqM9mZkrRB", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Haoyu Liu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Nanshan Jia", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Tingyu Zhu", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Zeyu Zheng", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Associate Professor" } ], "n_formula": 67, "n_formula_1": 38, "n_ref_uni": 15, "n_ref": 24, "n_ref_all": 45, "n_fig": 5, "n_tab": 10, "L_tab": 1825, "n_element_tab": 123, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2320, "n_element_tab_1": 80, "formula_len_all": 5191, "formula_len_all_1": 2787, "len_all": 145058, "len_all_1": 55416, "len_abs": 809, "len_title": 123, "len_sents": 38030, "len_sents_1": 21994, "n_sents": 338, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 812, "title": "Structured Diffusion Models with Mixture of Gaussians as Prior Distribution", "abs": "We propose a class of structured diffusion models, in which the prior distribution is chosen as a mixture of Gaussians, rather than a standard Gaussian distribution. The specific mixed Gaussian distribution, as prior, can be chosen to incorporate certain structured information of the data. We develop a simple-to-implement training procedure that smoothly accommodates the use of mixed Gaussian as prior. Theory is provided to quantify the benefits of our proposed models, compared to the classical diffusion models. Numerical experiments with synthetic, image and operational data are conducted to show comparative advantages of our model. Our method is shown to be robust to mis-specifications and in particular suits situations where training resources are limited or faster training in real time is desired.", "keywords": [ "structured diffusion models", "mixed Gaussian prior", "training efficiency under limited resources" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "Zq8wylMZ8A", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Chandan Singh", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Researcher" }, { "name": "Eunji Kim", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Jianfeng Gao", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Sriya Mantena", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Sungroh Yoon", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Weiwei Yang", "gender": "Not Specified", "institution": "Microsoft", "country": "", "position": "Principal Researcher" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 69, "n_ref": 100, "n_ref_all": 134, "n_fig": 8, "n_tab": 7, "L_tab": 3415, "n_element_tab": 369, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1864, "n_element_tab_1": 193, "formula_len_all": 501, "formula_len_all_1": 472, "len_all": 192021, "len_all_1": 63905, "len_abs": 1251, "len_title": 111, "len_sents": 47089, "len_sents_1": 27710, "n_sents": 336, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1257, "title": "Interpretable Language Modeling via Induction-head Ngram Models", "abs": "Recent large language models (LLMs) have excelled across a wide range of tasks, but their use in high-stakes and compute-limited settings has intensified the demand for interpretability and efficiency. We address this need by proposing Induction-head ngram models (Induction-Gram), a method that builds an efficient, interpretable LM by bolstering modern ngram models with a hand-engineered ``induction head''. This induction head uses a custom neural similarity metric to efficiently search the model's input context for potential next-word completions. This process enables Induction-Gram to provide ngram-level grounding for each generated token. Moreover, experiments show that this simple method significantly improves next-word prediction over baseline interpretable models (up to 26%p) and can be used to speed up LLM inference for large models through speculative decoding. We further study Induction-Gram in a natural-language neuroscience setting, where the goal is to predict the next fMRI response in a sequence. It again provides a significant improvement over interpretable models (20% relative increase in the correlation of predicted fMRI responses), potentially enabling deeper scientific investigation of language selectivity in the brain.", "keywords": [ "interpretability", "ngram", "language modeling", "fmri", "neuroscience" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "ZpcQfTNtKv", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Aditya Challa", "gender": "Male", "institution": "Birla Institute of Technology and Science", "country": "IN", "position": "Assistant Professor" }, { "name": "Laurent Najman", "gender": "Male", "institution": "Universit\u00e9 Gustave Eiffel - ESIEE Paris", "country": "FR", "position": "Full Professor" }, { "name": "Snehanshu Saha", "gender": "Not Specified", "institution": "Birla Institute of Technology and Science, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "Full Professor" }, { "name": "Sravan Danda", "gender": "Male", "institution": "Birla Institute of Technology and Science, Goa", "country": "", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 23, "n_ref": 33, "n_ref_all": 60, "n_fig": 17, "n_tab": 3, "L_tab": 1062, "n_element_tab": 104, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 862, "formula_len_all_1": 585, "len_all": 109286, "len_all_1": 59263, "len_abs": 1499, "len_title": 107, "len_sents": 38313, "len_sents_1": 28421, "n_sents": 350, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1521, "title": "Quantile Activation: Correcting a failure mode of ML models", "abs": "An established failure mode for machine learning models occurs when the same features are equally likely to belong to class $0$ and class $1$.. In such cases, any ML model cannot to correctly classify the sample. However, a solvable case emerges when the probabilities of class $0$ and $1$ vary with the \"context distribution\". To the best of our knowledge, standard neural network architectures like MLPs or CNNs are not equipped to handle this.\n\nIn this article, we propose a simple activation function, quantile activation (QACT), that addresses this problem without significantly increasing computational costs. The core idea is to \"adapt\" the outputs of each neuron to its *context distribution*. The proposed quantile activation, QACT, produces the \"relative quantile\" of the sample in its context distribution, rather than the actual values, as in traditional networks.\n\nA practical example where the same sample can have different labels arises in cases of inherent distribution shift. We validate the proposed activation function under such shifts, using datasets designed to test robustness against distortions\u2014CIFAR10C, CIFAR100C, MNISTC, TinyImagenetC. Our results demonstrate significantly better generalization across distortions compared to conventional classifiers, across various architectures. Although this paper presents a proof of concept, we find that this approach unexpectedly outperforms DINOv2 (small) under large distortions, despite DINOv2 being trained with a much larger network and dataset.", "keywords": [ "Machine Learning Foundations", "Quantiles", "Distribution Shift" ], "rating_list": [ 5, 3, 8, 5 ], "soundness_list": [ 2, 1, 4, 2 ], "presentation_list": [ 2, 1, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 2, 5, 4, 4 ] }, { "paper_id": "ZpQ2SqQNXf", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Gregory Valiant", "gender": "Unspecified", "institution": "Computer Science Department, Stanford University", "country": "", "position": "Associate Professor" }, { "name": "Percy Liang", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Steven Cao", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" } ], "n_formula": 40, "n_formula_1": 10, "n_ref_uni": 31, "n_ref": 64, "n_ref_all": 81, "n_fig": 7, "n_tab": 2, "L_tab": 10983, "n_element_tab": 3, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1066, "n_element_tab_1": 71, "formula_len_all": 4969, "formula_len_all_1": 1007, "len_all": 149150, "len_all_1": 59809, "len_abs": 1777, "len_title": 93, "len_sents": 37303, "len_sents_1": 28473, "n_sents": 258, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 45, "L_abs": 1785, "title": "On the Entropy Calibration of Language Models", "abs": "Language models are trained with teacher forcing but are used autoregressively, so errors accumulate as more tokens are generated. This issue is well-studied but remains a fundamental problem that harms generation quality. Building on past work, we take the perspective that error accumulation is reflected in the model's entropy, so we can better understand and address it through the lens of entropy calibration. A language model is entropy calibrated if its entropy over generations, i.e. its confidence, matches the log loss it incurs on actual text. First, we find that models are indeed miscalibrated in practice: for base models across a range of sizes, entropy per step increases as more tokens are generated, leading to generations becoming incoherent over time. On the other hand, after instruction tuning, the largest models now have too little entropy (i.e. are overconfident), leading to a lack of diversity in model outputs. From a theoretical perspective, entropy calibration is difficult to attain because it is a global property of the entire generation process, which has an exponentially large output space. Per-step adjustments are tractable but fail to preserve the model's log loss, while global adjustments preserve log loss but are intractable. Our main theoretical contribution is to propose future entropy scaling, an adjustment to the next token probabilities that uses information about the future entropy of each token, i.e. the average entropy of continuations from that token. With additional assumptions, we prove that this adjustment calibrates the model while preserving log loss. While future entropy estimation is expensive, this result suggests that calibration and stabilization of the entropy should be possible without trading off model quality.", "keywords": [ "language models", "calibration", "entropy", "sampling", "language model inference" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "Zp8NOZo0rA", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Chen Chen", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Ekkasit Pinyoanuntapong", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Intern" }, { "name": "Hongfei Xue", "gender": "Male", "institution": "University of North Carolina at Charlotte", "country": "US", "position": "Assistant Professor" }, { "name": "Jian Ren", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Junli Cao", "gender": "Male", "institution": "UCLA Computer Science Department, University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Korrawe Karunratanakul", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Muhammad Usama Saleem", "gender": "Male", "institution": "University of North Carolina at Charlotte", "country": "US", "position": "PhD student" }, { "name": "Pu Wang", "gender": "Male", "institution": "University of North Carolina at Charlotte", "country": "US", "position": "Associate Professor" }, { "name": "Sergey Tulyakov", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Director of Research" }, { "name": "chuan guo", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 34, "n_ref": 82, "n_ref_all": 100, "n_fig": 10, "n_tab": 10, "L_tab": 5770, "n_element_tab": 754, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 337, "n_element_tab_1": 30, "formula_len_all": 653, "formula_len_all_1": 402, "len_all": 172870, "len_all_1": 52567, "len_abs": 821, "len_title": 96, "len_sents": 42206, "len_sents_1": 25659, "n_sents": 301, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1636, "title": "ControlMM: Controllable Masked Motion Generation", "abs": "Recent advances in motion diffusion models have enabled spatially controllable text-to-motion generation. However, despite achieving acceptable control precision, these models suffer from generation speed and fidelity limitations. To address these challenges, we propose ControlMM, a novel approach incorporating spatial control signals into the generative masked motion model. ControlMM achieves real-time, high-fidelity, and high-precision controllable motion generation simultaneously. Our approach introduces two key innovations. First, we propose masked consistency modeling, which ensures high-fidelity motion generation via random masking and reconstruction, while minimizing the inconsistency between the input control signals and the extracted control signals from the generated motion. To further enhance control precision, we introduce inference-time logit editing, which manipulates the predicted conditional motion distribution so that the generated motion, sampled from the adjusted distribution, closely adheres to the input control signals. During inference, ControlMM enables parallel and iterative decoding of multiple motion tokens, allowing for high-speed motion generation. Extensive experiments show that, compared to the state of the art, ControlMM delivers superior results in motion quality, with better FID scores (0.061 vs 0.271), and higher control precision (average error 0.0091 vs 0.0108). ControlMM generates motions 20 times faster than diffusion-based methods. Additionally, ControlMM unlocks diverse applications such as any joint any frame control, body part timeline control, and obstacle avoidance.", "keywords": [ "Masked Motion Model", "Generative Model", "Text-to-motion" ], "rating_list": [ 5, 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2, 4 ], "presentation_list": [ 1, 3, 3, 4, 2 ], "contribution_list": [ 2, 3, 2, 4, 3 ], "confidence_list": [ 5, 4, 4, 4, 3 ] }, { "paper_id": "Zp51wHvoot", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Dohun Lee", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Hyungjin Chung", "gender": "Male", "institution": "EverEx", "country": "KR", "position": "Researcher" }, { "name": "Jong Chul Ye", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 49, "n_formula_1": 12, "n_ref_uni": 51, "n_ref": 101, "n_ref_all": 145, "n_fig": 7, "n_tab": 4, "L_tab": 585, "n_element_tab": 70, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3118, "n_element_tab_1": 133, "formula_len_all": 2753, "formula_len_all_1": 770, "len_all": 171787, "len_all_1": 52607, "len_abs": 1594, "len_title": 126, "len_sents": 37177, "len_sents_1": 21453, "n_sents": 296, "n_sents_1": 149, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1485, "title": "ACDC: Autoregressive Coherent Multimodal Generation using Diffusion Correction", "abs": "Autoregressive models (ARMs) and diffusion models (DMs) represent two leading paradigms in generative modeling, each excelling in distinct areas: ARMs in global context modeling and long-sequence generation, and DMs in generating high-quality local contexts, especially for continuous data such as images and short videos. However, ARMs often suffer from exponential error accumulation over long sequences, leading to physically implausible results, while DMs are limited by their local context generation capabilities. In this work, we introduce Autoregressive Coherent multimodal generation with Diffusion Correction (ACDC), a zero-shot approach that combines the strengths of both ARMs and DMs at the inference stage without the need for additional fine-tuning. ACDC leverages ARMs for global context generation and memory-conditioned DMs for local correction, ensuring high-quality outputs by correcting artifacts in generated multimodal tokens. In particular, we propose a memory module based on large language models (LLMs) that dynamically adjusts the conditioning texts for the DMs, preserving crucial global context information. Our experiments on multimodal tasks, including coherent multi-frame story generation and autoregressive video generation, demonstrate that ACDC effectively mitigates the accumulation of errors and significantly enhances the quality of generated outputs, achieving superior performance while remaining agnostic to specific ARM and DM architectures.", "keywords": [ "Large Multimodal Models", "Autoregressive Models", "Diffusion Models" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 1, 1, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "Zonhl0c9I0", "primary_area": "reinforcement learning", "n_author": 10, "author_info_list": [ { "name": "Chenghua Huang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Dongmei Zhang", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Distinguished Scientist" }, { "name": "Fangkai Yang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Lu Wang", "gender": "Female", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Pu Zhao", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Qi Zhang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Qingwei Lin", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Sr. Principal Researcher" }, { "name": "Saravan Rajmohan", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Partner Director of Applied Research" }, { "name": "Zeqi Lin", "gender": "Male", "institution": "Microsoft Research", "country": "", "position": "Researcher" }, { "name": "Zhizhen Fan", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" } ], "n_formula": 23, "n_formula_1": 10, "n_ref_uni": 17, "n_ref": 27, "n_ref_all": 49, "n_fig": 6, "n_tab": 4, "L_tab": 1201, "n_element_tab": 113, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2326, "n_element_tab_1": 83, "formula_len_all": 1633, "formula_len_all_1": 811, "len_all": 153886, "len_all_1": 57198, "len_abs": 1131, "len_title": 85, "len_sents": 49405, "len_sents_1": 25824, "n_sents": 378, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 37, "L_abs": 1222, "title": "SELF-EVOLVED REWARD LEARNING FOR LLMS", "abs": "Reinforcement Learning from Human Feedback (RLHF) is a crucial technique for aligning language models with human preferences and is a key factor in the success of modern conversational models like GPT-4, ChatGPT, and Llama 2. A significant challenge in employing RLHF lies in training a reliable RM, which relies on high-quality labels. Typically, these labels are provided by human experts or a stronger AI, both of which can be costly and introduce bias that may affect the language model's responses. As models improve, human input may become less effective in enhancing their performance. This paper explores the potential of using the RM itself to generate additional training data for a more robust RM. Our experiments demonstrate that reinforcement learning from self-feedback outperforms baseline approaches.\nWe conducted extensive experiments with our approach on multiple datasets, such as HH-RLHF and UltraFeedback, and models including Mistral and Llama 3, comparing it against various baselines. Our results indicate that, even with a limited amount of human-labeled data, learning from self-feedback can robustly enhance the performance of the RM, thereby improving the capabilities of large language models.", "keywords": [ "RLHF", "LLM", "Self-learning" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "Zoli4UAQVZ", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Mike Qu", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "MS student" }, { "name": "Yingzhen Li", "gender": "Female", "institution": "Imperial College London", "country": "", "position": "Associate Professor" }, { "name": "Zijing Ou", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Jacob Yoke Hong Si", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 14, "n_ref_uni": 46, "n_ref": 131, "n_ref_all": 166, "n_fig": 8, "n_tab": 16, "L_tab": 13446, "n_element_tab": 1108, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 3446, "n_element_tab_1": 281, "formula_len_all": 2157, "formula_len_all_1": 1023, "len_all": 207534, "len_all_1": 68181, "len_abs": 1493, "len_title": 133, "len_sents": 62509, "len_sents_1": 28593, "n_sents": 487, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1499, "title": "TabUnite: Efficient Encoding Schemes for Flow and Diffusion Tabular Generative Models", "abs": "Flow matching and diffusion generative models for tabular data face challenges in modeling heterogeneous feature interrelationships, especially in data with continuous and categorical input features. Capturing these interrelationships is crucial as it allows these models to understand complex patterns and dependencies in the underlying data. A promising option to address the challenge is to devise suitable encoding schemes for the input features before the generative modeling process. However, prior methods often rely on either suboptimal heuristics such as one-hot encoding of categorical features followed by separated modeling of categorical/continuous features, or latent space diffusion models. Instead, our proposed solution unifies the data space and jointly applies a single generative process across all the encodings, efficiently capturing heterogeneous feature interrelationships. Specifically, it employs encoding schemes such as PSK Encoding, Dictionary Encoding, and Analog Bits that effectively convert categorical features into continuous ones. Extensive experiments on datasets comprised of heterogeneous features demonstrate that our encoding schemes, combined with Flow Matching or Diffusion as our choice of generative model, significantly enhance model capabilities. Our TabUnite models help address data heterogeneity, achieving superior performance across a broad suite of datasets, baselines, and benchmarks while generating accurate, robust, and diverse tabular data.", "keywords": [ "tabular data generation", "encoding schemes", "flow matching", "diffusion" ], "rating_list": [ 3, 5, 3, 3, 5 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 4, 3 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 5, 4, 4 ] }, { "paper_id": "Zmu3lVw6bm", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Hrad Ghoukasian", "gender": "Male", "institution": "McMaster University", "country": "CA", "position": "MS student" }, { "name": "Minghui Chen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "Ruinan Jin", "gender": "unknown", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Sai Praneeth Karimireddy", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Xiaoxiao Li", "gender": "Unspecified", "institution": "University of British Columbia", "country": "CA", "position": "Assistant Professor" }, { "name": "Zehua Wang", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Researcher" } ], "n_formula": 57, "n_formula_1": 8, "n_ref_uni": 28, "n_ref": 53, "n_ref_all": 78, "n_fig": 12, "n_tab": 6, "L_tab": 4644, "n_element_tab": 214, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 2200, "n_element_tab_1": 106, "formula_len_all": 8985, "formula_len_all_1": 678, "len_all": 210876, "len_all_1": 69960, "len_abs": 5830, "len_title": 125, "len_sents": 62218, "len_sents_1": 32570, "n_sents": 505, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1926, "title": "A Closer Look at Personalized Fine-Tuning in Heterogeneous Federated Learning", "abs": "Federated Learning (FL) enables privacy-preserving, decentralized model training but faces significant challenges in balancing global generalization and local personalization due to non-identical data distributions across clients. While Personalized Fine-Tuning (PFT) adapts models to local data, excessive personalization often degrades global performance. In this work, we present a comprehensive empirical study encompassing seven diverse datasets, multiple model architectures, and various fine-tuning methods under both covariate and concept shift scenarios. Our extensive evaluation reveals critical limitations in existing PFT methods, which struggle with overfitting and exhibit inconsistent performance across distribution shifts, even with careful hyperparameter tuning and regularization. To address these issues, we identify LP-FT, a simple yet effective strategy that combines Linear Probing with full Fine-Tuning, adapted to the FL setting. LP-FT consistently outperforms existing methods, achieving an optimal balance between local personalization and global generalization across all tested scenarios. By investigating the feature change after PFT, we hypothesize the a phenomena dubbed as federated feature distortion is linked to the global generalization. Motivated by the observation, we provide a theoretical analysis of two-layer linear networks, offering novel insights into the conditions under which LP-FT excels, thereby enhancing our understanding of personalization dynamics in FL. This work contributes in three key areas: (1) a rigorous and comprehensive evaluation of PFT methods under diverse distribution shifts, (2) the introduction of LP-FT as a robust and versatile solution to FL personalization challenges, and (3) theoretical foundations that explain LP-FT\u2019s superior effectiveness. Our findings set a new venue for PFT research and provide valuable insights to the broader FL community.", "keywords": [ "Federated Learning; Distribution Shift" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "ZmPf5Z1V2H", "primary_area": "foundation or frontier models, including LLMs", "n_author": 14, "author_info_list": [ { "name": "Baobao Chang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Dongmei Zhang", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Distinguished Scientist" }, { "name": "Fangkai Yang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Kaikai An", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Le-le Cao", "gender": "Male", "institution": "Microsoft (ABK)", "country": "SE", "position": "Senior Principal AI Researcher" }, { "name": "Liqun Li", "gender": "Male", "institution": "Microsoft Research", "country": "", "position": "Researcher" }, { "name": "Lu Wang", "gender": "Female", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Pu Zhao", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Qi Zhang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Qingwei Lin", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Sr. Principal Researcher" }, { "name": "Saravan Rajmohan", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Partner Director of Applied Research " }, { "name": "Shuzheng Si", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Sitao Cheng", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "JunTing Lu", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 27, "n_ref": 59, "n_ref_all": 103, "n_fig": 4, "n_tab": 10, "L_tab": 4026, "n_element_tab": 239, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 988, "n_element_tab_1": 81, "formula_len_all": 111, "formula_len_all_1": 105, "len_all": 179307, "len_all_1": 70661, "len_abs": 1424, "len_title": 162, "len_sents": 58764, "len_sents_1": 32200, "n_sents": 489, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Reject", "L_title": 114, "L_abs": 1431, "title": "Thread: A Logic-Based Data Organization Paradigm for How-To Question Answering with Retrieval Augmented Generation", "abs": "Recent advances in retrieval-augmented generation have significantly improved the performance of question-answering systems, particularly on factoid '5Ws' questions. However, these systems still face substantial challenges when addressing '1H' questions, specifically how-to questions, which are integral to decision-making processes and require dynamic, step-by-step answers. The key limitation lies in the prevalent data organization paradigm, chunk, which divides documents into fixed-size segments, and disrupts the logical coherence and connections within the context. To overcome this, in this paper, we propose Thread, a novel data organization paradigm aimed at enabling current systems to handle how-to questions more effectively. Specifically, we introduce a new knowledge granularity, termed 'logic unit', where documents are transformed into more structured and loosely interconnected logic units with large language models. Extensive experiments conducted across both open-domain and industrial settings demonstrate that Thread outperforms existing paradigms significantly, improving the success rate of handling how-to questions by 21\\% to 33\\%. \nMoreover, Thread exhibits high adaptability in processing various document formats, drastically reducing the candidate quantity in the knowledge base and minimizing the required information to one-fourth compared with chunk, optimizing both efficiency and effectiveness.", "keywords": [ "Data organization paradigm", "Retrieval augmented generation", "How-to questions", "Large language model" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 2, 3 ] }, { "paper_id": "Zltr2XVjDq", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Alan Gany", "gender": "Male", "institution": "Universit\u00e9 Paris-Saclay", "country": "FR", "position": "PhD student" }, { "name": "Bogdan Cautis", "gender": "Male", "institution": "Universit\u00e9 Paris Saclay", "country": "FR", "position": "Full Professor" }, { "name": "Laks V. S. Lakshmanan", "gender": "unknown", "institution": "University of British Columbia", "country": "CA", "position": "Professor" }, { "name": "Paul Landrier", "gender": "Male", "institution": "Ecole Normale Sup\u00e9rieure de Paris", "country": "FR", "position": "MS student" }, { "name": "Silviu Maniu", "gender": "unknown", "institution": "Universit\u00e9 Grenoble Alpes", "country": "", "position": "Full Professor" } ], "n_formula": 78, "n_formula_1": 17, "n_ref_uni": 30, "n_ref": 51, "n_ref_all": 74, "n_fig": 9, "n_tab": 8, "L_tab": 3650, "n_element_tab": 319, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 200, "n_element_tab_1": 15, "formula_len_all": 9476, "formula_len_all_1": 1216, "len_all": 175281, "len_all_1": 69842, "len_abs": 1535, "len_title": 89, "len_sents": 54702, "len_sents_1": 32353, "n_sents": 504, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1545, "title": "Stable GNN Embeddings for Relational Data", "abs": "Graph neural networks (GNNs) are a valuable tool for extracting meaningful representations from graph-structured data. Graphs, like relational databases, represent relationships between entities. Recent research has explored the potential of using GNNs for downstream tasks on relational data, such as entity resolution and missing value imputation. However, applying GNNs to relational databases presents two challenges. The first challenge is data conversion: relational databases, organized as tables connected by key / foreign key constraints,\nmust be transformed into graphs without losing essential information. The second challenge is ensuring that the embedding technique can adapt to the dynamic nature of databases. When a database is updated, the embeddings of the resulting database should be recomputable efficiently. This requires that previously computed embeddings remain stable despite changes to the data.\n\nMotivated by using GNNs for relational databases, we study stability, i.e., how much the embeddings generated by a GNN change when the input graph undergoes modifications. Building upon the work of Gama et al. (2020), which established a limit for the distance between embeddings of similar graphs, we focus on node-level stability for GNN embeddings, particularly when the graphs originate from relations. We propose several techniques for transforming relational databases into graphs. To assess the effectiveness of these methods, we conduct experiments using the TPC-E database benchmark and analyze their stability.", "keywords": [ "GNN", "stability", "database embeddings" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 2 ] }, { "paper_id": "Zkrsr7vAaG", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Kaisheng Ma", "gender": "Male", "institution": "Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Runjie Yan", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Kai Wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 28, "n_formula_1": 17, "n_ref_uni": 46, "n_ref": 164, "n_ref_all": 183, "n_fig": 10, "n_tab": 3, "L_tab": 500, "n_element_tab": 45, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 742, "n_element_tab_1": 23, "formula_len_all": 1909, "formula_len_all_1": 1048, "len_all": 168142, "len_all_1": 53632, "len_abs": 1096, "len_title": 105, "len_sents": 42942, "len_sents_1": 22573, "n_sents": 433, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1048, "title": "Flow Score Distillation for Diverse Text-to-3D Generation", "abs": "Recent advancements in Text-to-3D generation have yielded remarkable progress, particularly through methods that rely on Score Distillation Sampling (SDS). While SDS exhibits the capability to create impressive 3D assets, it is hindered by its inherent maximum-likelihood-seeking essence, resulting in limited diversity in generation outcomes. In this paper, we discover that the Denoise Diffusion Implicit Models (DDIM) generation process (i.e. PF-ODE) can be succinctly expressed using an analogue of SDS loss. One step further, one can see SDS as a generalized DDIM generation process. Following this insight, we show that the noise sampling strategy in the noise addition stage significantly restricts the diversity of generation results. To address this limitation, we present an innovative noise sampling approach and introduce a novel text-to-3D method called Flow Score Distillation (FSD). Our validation experiments across various text-to-image Diffusion Models demonstrate that FSD substantially enhances generation diversity and quality.", "keywords": [ "text to 3D", "score distillation", "diffusion model", "3D generation" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 4, 2, 3, 1 ], "contribution_list": [ 3, 3, 1, 1 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "Zkq4fsyjfp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Anton van den Hengel", "gender": "unknown", "institution": "University of Adelaide", "country": "AU", "position": "Professor" }, { "name": "Damien Teney", "gender": "Male", "institution": "Idiap Research Institute", "country": "CH", "position": "Researcher" }, { "name": "Hamed Damirchi", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "PhD student" }, { "name": "Cristian Rodriguez", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Edison Marrese-Taylor", "gender": "Male", "institution": "The Univesity of Tokyo", "country": "JP", "position": "Assistant Professor" }, { "name": "Ehsan M Abbasnejad", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 69, "n_ref": 132, "n_ref_all": 175, "n_fig": 12, "n_tab": 12, "L_tab": 17562, "n_element_tab": 1218, "n_fig_1": 11, "n_tab_1": 5, "L_tab_1": 1873, "n_element_tab_1": 282, "formula_len_all": 578, "formula_len_all_1": 248, "len_all": 244492, "len_all_1": 71349, "len_abs": 1323, "len_title": 139, "len_sents": 59547, "len_sents_1": 30485, "n_sents": 469, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1274, "title": "Synergy and Diversity in CLIP: Enhancing Performance Through Adaptive Backbone Ensembling", "abs": "Contrastive Language-Image Pretraining (CLIP) stands out as a prominent method for image representation learning. Various architectures, from vision transformers~(ViTs) to convolutional networks (ResNets) have been trained with CLIP to serve as general solutions to diverse vision tasks.\nThis paper explores the differences across various CLIP-trained vision backbones.\nDespite using the same data and training objective, we find that these architectures have notably different representations,\ndifferent classification performance across datasets, and different robustness properties to certain types of image perturbations.\nOur findings indicate a remarkable possible synergy across backbones\nby leveraging their respective strengths.\nIn principle, classification accuracy could be improved by over 40 percentage with an informed selection of the optimal backbone per test example. \nUsing this insight, we develop a straightforward yet powerful approach to adaptively ensemble multiple backbones.\nThe approach uses as few as one labeled example per class\nto tune the adaptive combination of backbones.\nOn a large collection of datasets, the method achieves a remarkable increase in accuracy of up to 39.1\\% over the best single backbone, well beyond traditional ensembles.", "keywords": [ "CLIP" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ZkpDdCQUC4", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Dequan Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Dukun Zhao", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Intern" }, { "name": "Jiaqi Shen", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Postdoc" }, { "name": "Jin Gao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Juntu Zhao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Yuming Lu", "gender": "Not Specified", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 68, "n_ref": 99, "n_ref_all": 138, "n_fig": 7, "n_tab": 7, "L_tab": 4492, "n_element_tab": 311, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1797, "n_element_tab_1": 156, "formula_len_all": 456, "formula_len_all_1": 362, "len_all": 206317, "len_all_1": 66271, "len_abs": 1693, "len_title": 133, "len_sents": 50647, "len_sents_1": 30984, "n_sents": 348, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1647, "title": "NovoBench-100K: A large-scale protein dataset for in silico evolution of de novo TadA", "abs": "We introduce NOVOBENCH-100K, a large-scale protein dataset for the in silico evolution of TadA, an enzyme critical for base editing. This dataset originates from the sequencing data collected during two rounds of our in vitro TadA evolution, encompassing 101,687 unique DNA variants with an average of 11.1 amino acid mutations. Rather than employing classes or scores as labels, our dataset consists of 77,900 ranking lists, each involving 2, 10, or 100 sequences ranked by their base editing efficiency. These rankings are generated using our SEQ2RANK, a novel algorithm that accounts for biological experiment credibility and ranking consistency. For evaluation, we provide two train-test splits, designated as in-domain ranking and out-of-domain ranking, based on a standard 7:3 random split and the actual in-vitro evolution rounds, respectively. We benchmark 80 biological language models (BLMs) across 24 papers, spanning protein, DNA, RNA, and multimodal domains. Comprehensive experiments reveal that BLMs perform well on in-domain ranking, with a detailed analysis by modality, model size, and K-mer. However, for out-of-domain ranking, BLMs exhibit poor performance in both linear probing and fine-tuning, resembling random guessing. This underscores the necessity for highly generalizable models to address domain shifts between experimental rounds. Finally, our wet experiments are ongoing to generate more data to expand our benchmark. In a few months, we expect to add additional rounds of in vitro evolution and include a broader variety of proteins. We will release the code, dataset, and embeddings of our evaluated 80 BLMs soon.", "keywords": [ "Benchmark", "Dataset", "Biological Language Model", "Ranking", "Base Editing" ], "rating_list": [ 5, 3, 6, 6, 3 ], "soundness_list": [ 3, 1, 4, 3, 2 ], "presentation_list": [ 2, 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 4, 3, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "Zkp1GuHerF", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Ionut-Vlad Modoranu", "gender": "Male", "institution": "Institute of Science and Technology Austria", "country": "", "position": "PhD student" }, { "name": "Mher Safaryan", "gender": "Male", "institution": "Institute of Science and Technology", "country": "", "position": "Postdoc" }, { "name": "Thomas Robert", "gender": "Male", "institution": "\u00c9cole Polytechnique - MVA", "country": "FR", "position": "MS student" }, { "name": "Dan Alistarh", "gender": "Male", "institution": "Institute of Science and Technology", "country": "AT", "position": "Full Professor" } ], "n_formula": 112, "n_formula_1": 23, "n_ref_uni": 46, "n_ref": 95, "n_ref_all": 123, "n_fig": 3, "n_tab": 13, "L_tab": 4580, "n_element_tab": 539, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 737, "n_element_tab_1": 124, "formula_len_all": 20251, "formula_len_all_1": 2138, "len_all": 217729, "len_all_1": 77262, "len_abs": 856, "len_title": 118, "len_sents": 55052, "len_sents_1": 34957, "n_sents": 479, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 866, "title": "LDAdam: Adaptive Optimization from Low-Dimensional Gradient Statistics", "abs": "We introduce LDAdam, a memory-efficient optimizer for training large models, that performs adaptive optimization steps within lower dimensional subspaces, while consistently exploring the full parameter space during training. This strategy keeps the optimizer's memory footprint to a fraction of the model size. LDAdam relies on a new projection-aware update rule for the optimizer states that allows for transitioning between subspaces, i.e., estimation of the statistics of the projected gradients. To mitigate the errors due to low-rank projection, LDAdam integrates a new generalized error feedback mechanism, which explicitly accounts for both gradient and optimizer state compression. We prove the convergence of LDAdam under standard assumptions, and provide empirical evidence that LDAdam allows for efficient fine-tuning and pre-training of language models.", "keywords": [ "adaptive optimization", "memory efficiency", "low-rank learning", "low-rank compression", "convergence rates" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "ZkHtfl77JG", "primary_area": "applications to neuroscience & cognitive science", "n_author": 7, "author_info_list": [ { "name": "Haruhiko Kishima", "gender": "Male", "institution": "Graduate School of Medicine, Osaka University", "country": "", "position": "Full Professor" }, { "name": "Rikuto Kotoge", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "PhD student" }, { "name": "Takufumi Yanagisawa", "gender": "Male", "institution": "Osaka University, Tokyo Institute of Technology", "country": "JP", "position": "Full Professor" }, { "name": "Tasuku Kimura", "gender": "unknown", "institution": "Osaka University", "country": "", "position": "Specially Appointed Assistant Professor" }, { "name": "Yasuko Matsubara", "gender": "Female", "institution": "Osaka University", "country": "JP", "position": "Associate Professor" }, { "name": "Yasushi Sakurai", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "Full Professor" }, { "name": "Zheng Chen", "gender": "Male", "institution": "ISIR, Osaka University, Japan", "country": "JP", "position": "Assistant Professor" } ], "n_formula": 41, "n_formula_1": 12, "n_ref_uni": 46, "n_ref": 98, "n_ref_all": 115, "n_fig": 7, "n_tab": 2, "L_tab": 798, "n_element_tab": 143, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1999, "n_element_tab_1": 152, "formula_len_all": 2055, "formula_len_all_1": 885, "len_all": 171323, "len_all_1": 65150, "len_abs": 6006, "len_title": 120, "len_sents": 45718, "len_sents_1": 27460, "n_sents": 402, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1252, "title": "Dynamic multi-channel EEG graph modeling for time-evolving brain network", "abs": "We describe a novel dynamic graph neural network (GNN) approach for seizure detection and prediction from multi-channel Electroencephalography (EEG) data thet addresses several limitations of existing methods. \nWhile deep learning models have achieved notable success in automating seizure detection, static graph-based methods fail to capture the evolving nature of brain networks, especially during seizure events. \nTo overcome this, we propose EvoBrain, which uses a time-then-graph strategy that first models the temporal dynamics of EEG signals and graphs, and then employs GNNs to learn evolving spatial EEG representations. \nOur contributions include \n(a) a theoretical analysis proving the expressivity advantage of time-then-graph over other approaches, \n(b) a simple and efficient model that significantly improves AUROC and F1 scores compared with state-of-the-art methods, and \n(c) the introduction of dynamic graph structures that better reflect transient changes in brain connectivity. \nWe evaluate our method on the challenging early seizure prediction task.\nThe results show improved performance, making EvoBrain a valuable tool for clinical applications.\nThe source code is available at:\nhttps://anonymous.4open.science/r/EvoBrain-FBC5", "keywords": [ "EEG", "Neuro Science", "Graph", "Time Series" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "ZkFMe3OPfw", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Dhritiman Sagar", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Erli Ding", "gender": "Male", "institution": "Case Western Reserve University", "country": "US", "position": "PhD student" }, { "name": "Fu-Yun Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Keqiang Sun", "gender": "Male", "institution": "The Chinese University of Hong Kong, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zhixin Lai", "gender": "Male", "institution": "Google", "country": "US", "position": "Eng, Machine Learning" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 30, "n_ref": 51, "n_ref_all": 93, "n_fig": 11, "n_tab": 10, "L_tab": 2520, "n_element_tab": 244, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1053, "n_element_tab_1": 175, "formula_len_all": 812, "formula_len_all_1": 810, "len_all": 135920, "len_all_1": 47198, "len_abs": 333, "len_title": 85, "len_sents": 36395, "len_sents_1": 19458, "n_sents": 255, "n_sents_1": 138, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1485, "title": "InstantPortrait: One-Step Portrait Editing via Diffusion Multi-Objective Distillation", "abs": "Real-time instruction-based portrait image editing is crucial in various applications, including filters, augmented reality, and video communications, etc. However, real-time portrait editing presents three significant challenges: identity preservation, fidelity to editing instructions, and fast model inference. Given that these aspects often present a trade-off, concurrently addressing them poses an even greater challenge. While diffusion-based image editing methods have shown promising capabilities in personalized image editing in recent years, they lack a dedicated focus on portrait editing and thus suffer from the aforementioned problems as well. To address the gap, this paper introduces an Instant-Portrait Network (IPNet), the first one-step diffusion-based model for portrait editing. We train the network in two stages. We first employ an annealing identity loss to train an Identity Enhancement Network (IDE-Net), to ensure robust identity preservation. We then train the IPNet using a novel diffusion Multi-Objective Distillation approach that integrates adversarial loss, identity distillation loss, and a novel Facial-Style Enhancing loss. The Diffusion Multi-Objective Distillation approach efficiently reduces inference steps, ensures identity consistency, and enhances the precision of instruction-based editing. Extensive comparison with prior models demonstrates IPNet as a superior model in terms of identity preservation, text fidelity, and inference speed.", "keywords": [ "Portrait Editing", "Diffusion Multi-Objective Distillation", "On Step Inference", "Identity Preservation", "Text Fidelity" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "ZkDgQ2PDDm", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 1, "author_info_list": [ { "name": "Yoshiaki Kitazawa", "gender": "Male", "institution": "NTT DATA Mathematical Systems Inc. ", "country": "", "position": "Researcher" } ], "n_formula": 148, "n_formula_1": 40, "n_ref_uni": 27, "n_ref": 53, "n_ref_all": 121, "n_fig": 5, "n_tab": 2, "L_tab": 644, "n_element_tab": 27, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 14832, "formula_len_all_1": 4143, "len_all": 195581, "len_all_1": 64914, "len_abs": 1356, "len_title": 111, "len_sents": 58852, "len_sents_1": 24669, "n_sents": 556, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1444, "title": "$\\alpha$-Divergence Loss Function for Neural Density Ratio Estimation", "abs": "Density ratio estimation (DRE) is a fundamental machine learning technique for capturing relationships between two probability distributions. State-of-the-art DRE methods estimate the density ratio using neural networks trained with loss functions derived from variational representations of $f$-divergence.\n However, existing methods face optimization challenges, such as overfitting due to lower-unbounded loss functions, biased mini-batch gradients, vanishing training loss gradients, and high sample requirements for Kullback-Leibler (KL) divergence loss functions.\n To address these issues, we focus on $\\alpha$-divergence, which provides a suitable variational representation of $f$-divergence.\n Subsequently, a novel loss function for DRE, the $\\alpha$-divergence loss function ($\\alpha$-Div), is derived.\n $\\alpha$-Div is concise but offers stable and effective optimization for DRE.\n The boundedness of $\\alpha$-divergence provides the potential for successful DRE with data exhibiting high KL-divergence.\n Our numerical experiments demonstrate the effectiveness in optimization using $\\alpha$-Div.\n However, the experiments also show that the proposed loss function offers no significant advantage over the KL-divergence loss function in terms of RMSE for DRE. This indicates that the accuracy of DRE is\n primarily determined by the amount of KL-divergence in the data and is less dependent on $\\alpha$-divergence.", "keywords": [ "density ratio estimation", "variational divergence optimization", "$\\alpha$-divergence", "Kullback\u2013Leibler divergence", "and $f$-divergence." ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 1 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "Zk9guOl9NS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Gabriel Synnaeve", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Jonas Gehring", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" }, { "name": "Juliette Decugis", "gender": "Female", "institution": "Ecole Normale Sup\u00e9rieure de Cachan", "country": "", "position": "MS student" }, { "name": "Kunhao Zheng", "gender": "Male", "institution": "Facebook", "country": "FR", "position": "PhD student" }, { "name": "Taco Cohen", "gender": "Male", "institution": "Meta", "country": "IT", "position": "Researcher" }, { "name": "benjamin negrevergne", "gender": "Male", "institution": "Univerist\u00e9 Paris-Dauphine", "country": "FR", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 39, "n_ref": 90, "n_ref_all": 145, "n_fig": 18, "n_tab": 14, "L_tab": 2437, "n_element_tab": 227, "n_fig_1": 9, "n_tab_1": 6, "L_tab_1": 1457, "n_element_tab_1": 115, "formula_len_all": 322, "formula_len_all_1": 52, "len_all": 244216, "len_all_1": 58342, "len_abs": 2933, "len_title": 120, "len_sents": 73966, "len_sents_1": 26742, "n_sents": 621, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1044, "title": "What Makes Large Language Models Reason in (Multi-Turn) Code Generation?", "abs": "Prompting techniques such as chain-of-thought have established themselves as a popular vehicle for improving the outputs of large language models (LLMs). For code generation, however, their exact mechanics and efficacy are under-explored using unified metrics and benchmarks. We thus investigate the effects of a wide range of prompting strategies with a focus on automatic re-prompting over multiple turns and computational requirements. After systematically decomposing reasoning, instruction, and execution feedback prompts, we conduct an extensive grid search on the competitive programming benchmarks CodeContests and TACO for multiple LLM families and sizes (Llama 3.0 and 3.1, 8B, 70B, 405B, and GPT-4o). Our study reveals strategies that consistently improve performance across all models with small and large sampling budgets. We then show how finetuning with such an optimal configuration allows models to internalize the induced reasoning process and obtain improvements in performance and scalability for multi-turn code generation.", "keywords": [ "Large language Models", "Multi-turn Code Generation", "Chain-of-Thought" ], "rating_list": [ 8, 5, 8 ], "soundness_list": [ 4, 2, 4 ], "presentation_list": [ 3, 2, 4 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "Zk8PNvzWQY", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Deunsol Yoon", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Researcher" }, { "name": "Jeonghye Kim", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Kanghoon Lee", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Principal Researcher" }, { "name": "Sunghoon Hong", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Researcher" }, { "name": "Whiyoung Jung", "gender": "Male", "institution": "LG AI Research", "country": "", "position": "Researcher" }, { "name": "Woohyung Lim", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Vice President" }, { "name": "Yongjae Shin", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Youngchul Sung", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 23, "n_formula_1": 11, "n_ref_uni": 43, "n_ref": 137, "n_ref_all": 188, "n_fig": 17, "n_tab": 15, "L_tab": 7703, "n_element_tab": 1185, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 3727, "n_element_tab_1": 644, "formula_len_all": 1339, "formula_len_all_1": 603, "len_all": 207232, "len_all_1": 70898, "len_abs": 1076, "len_title": 141, "len_sents": 63079, "len_sents_1": 26956, "n_sents": 471, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1046, "title": "Penalizing Infeasible Actions and Reward Scaling in Reinforcement Learning with Offline Data", "abs": "Reinforcement learning with offline data often suffers from Q-value extrapolation errors due to limited data, which poses significant challenges and limits overall performance. Existing methods such as layer normalization and reward relabeling have shown promise in addressing these errors and achieving empirical improvements. In this paper, we extend these approaches by introducing reward scaling with layer normalization (RS-LN) to further mitigate extrapolation errors and enhance performance. Furthermore, based on the insight that Q-values should be lower for infeasible action spaces\u2014where neural networks might otherwise extrapolate into undesirable regions\u2014we propose a penalization mechanism for infeasible actions (PA). By combining RS-LN and PA, we develop a new algorithm called PARS. We evaluate PARS on a range of tasks, demonstrating superior performance compared to state-of-the-art algorithms in both offline training and online fine-tuning across the D4RL benchmark, with notable success in the challenging AntMaze Ultra task.", "keywords": [ "Offline-to-Online Reinforcement Learning", "Offline Reinforcement Learning", "Penalizing Infeasible Actions", "Layer Normalization", "Reward Scaling\u200b" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "ZjvUcCAEK8", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Errui Ding", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Director" }, { "name": "Guanbin Li", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Jingdong Wang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Chief Scientist for Computer Vision" }, { "name": "Wei Zhang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "R&D at Baidu" }, { "name": "Weiming Zhang", "gender": "unknown", "institution": "Baidu Inc.", "country": "CN", "position": "Researcher" }, { "name": "Xiao Tan", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Zedong Dan", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "MS student" }, { "name": "Zijie Wang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 18, "n_ref": 29, "n_ref_all": 42, "n_fig": 5, "n_tab": 6, "L_tab": 1126, "n_element_tab": 230, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 559, "n_element_tab_1": 81, "formula_len_all": 668, "formula_len_all_1": 665, "len_all": 96401, "len_all_1": 55761, "len_abs": 1209, "len_title": 125, "len_sents": 26477, "len_sents_1": 24470, "n_sents": 197, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1215, "title": "UMVMap: Improving Vectorized Map Construction via Multi-vehicle Perspectives", "abs": "Prevalent vectorized map construction pipelines predominantly follow an end-to-end DETR-based paradigm. While these methods have achieved significant advancements, they are limited by their reliance on data from a single ego vehicle, which restricts their effectiveness and can lead to perceptual uncertainty in handling complex environmental scenarios. To address this limitation, we introduce a novel framework: Uncertainty-aware Multi-Vehicle Vectorized Map Construction (UMVMap). This framework effectively mitigates uncertainties by leveraging relevant non-ego information. UMVMap comprises two essential components: the Uncertainty-aware Multi-Vehicle Vectorized Map Construction Network (UMVMap-Net), which optimally integrates data from multiple vehicles, and the Uncertainty-aware Non-ego Vehicle Selection (UNVS) strategy, which identifies and incorporates the most informative non-ego data to minimize uncertainty. Comprehensive evaluations on the nuScenes dataset demonstrate that UMVMap significantly outperforms the single-vehicle MapTRv2 baseline by a margin of 9.1\\% and 9.9\\% respectively on the full and partial validation sets, with each of its components proving to be both effective and robust.", "keywords": [ "multi-vehicle", "autonomous driving", "vectorized map construction" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "Zjv38dg1Hb", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Beomsu Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jaemin Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Jeongsol Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Jong Chul Ye", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 52, "n_formula_1": 31, "n_ref_uni": 34, "n_ref": 76, "n_ref_all": 116, "n_fig": 11, "n_tab": 6, "L_tab": 1348, "n_element_tab": 195, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 3797, "n_element_tab_1": 154, "formula_len_all": 3954, "formula_len_all_1": 2168, "len_all": 134047, "len_all_1": 59892, "len_abs": 1108, "len_title": 112, "len_sents": 39799, "len_sents_1": 22311, "n_sents": 388, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1143, "title": "Generalized Consistency Trajectory Models for Image Manipulation", "abs": "Diffusion-based generative models excel in unconditional generation, as well as on applied tasks such as image editing and restoration. The success of diffusion models lies in the iterative nature of diffusion: diffusion breaks down the complex process of mapping noise to data into a sequence of simple denoising tasks. Moreover, we are able to exert fine-grained control over the generation process by injecting guidance terms into each denoising step. However, the iterative process is also computationally intensive, often taking from tens up to thousands of function evaluations. Although consistency trajectory models (CTMs) enable traversal between any time points along the probability flow ODE (PFODE) and score inference with a single function evaluation, CTMs only allow translation from Gaussian noise to data. Thus, this work aims to unlock the full potential of CTMs by proposing generalized CTMs (GCTMs), which translate between arbitrary distributions via ODEs. We discuss the design space of GCTMs and demonstrate their efficacy in various image manipulation tasks such as image-to-image translation, restoration, and editing.", "keywords": [ "Consistency Models", "Image Manipulation" ], "rating_list": [ 3, 8, 8 ], "soundness_list": [ 2, 3, 4 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 4 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "ZjuEPZJsa3", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Cfir Avraham Hadar", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "MS student" }, { "name": "Omer Shubi", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "PhD student" }, { "name": "Yevgeni Berzak", "gender": "unknown", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 47, "n_ref": 92, "n_ref_all": 112, "n_fig": 10, "n_tab": 1, "L_tab": 1209, "n_element_tab": 81, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 1209, "n_element_tab_1": 81, "formula_len_all": 672, "formula_len_all_1": 109, "len_all": 147964, "len_all_1": 63164, "len_abs": 1098, "len_title": 89, "len_sents": 43375, "len_sents_1": 30842, "n_sents": 318, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1101, "title": "Decoding Reading Goals from Eye Movements", "abs": "Readers can have different goals with respect to the text they are reading. Can these goals be decoded from the pattern of their eye movements over the text? In this work, we examine for the first time whether it is possible to decode two types of reading goals that are common in daily life: information seeking and ordinary reading. Using large scale eye-tracking data, we apply to this task a wide range of state-of-the-art models for eye movements and text that cover different architectural and data representation strategies, and further introduce a new model ensemble. We systematically evaluate these models at three levels of generalization: new textual item, new participant, and the combination of both. We find that eye movements contain highly valuable signals for this task. We further perform an error analysis which builds on prior empirical findings on differences between ordinary reading and information seeking and leverages rich textual annotations. This analysis reveals key properties of textual items and participant eye movements that contribute to the difficulty of the task.", "keywords": [ "application of language models to cognitive science and psycholinguistic", "eye movements in reading", "reading goals", "cognitive state decoding", "multimodal models (eye movements and text)" ], "rating_list": [ 6, 3, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "ZjOXuAfS6l", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Michael Crawshaw", "gender": "Male", "institution": "George Mason University", "country": "US", "position": "PhD student" }, { "name": "Mingrui Liu", "gender": "unknown", "institution": "George Mason University", "country": "", "position": "Assistant Professor" } ], "n_formula": 268, "n_formula_1": 26, "n_ref_uni": 31, "n_ref": 86, "n_ref_all": 172, "n_fig": 1, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 31055, "formula_len_all_1": 3759, "len_all": 262410, "len_all_1": 65604, "len_abs": 2957, "len_title": 169, "len_sents": 77122, "len_sents_1": 29306, "n_sents": 869, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 119, "L_abs": 1611, "title": "Complexity Lower Bounds of Adaptive Gradient Algorithms for Non-convex Stochastic Optimization under Relaxed Smoothness", "abs": "Recent results in non-convex stochastic optimization demonstrate the convergence of popular adaptive algorithms (e.g., AdaGrad) under the $(L_0, L_1)$-smoothness condition, but the rate of convergence is a higher-order polynomial in terms of problem parameters like the smoothness constants. The complexity guaranteed by such algorithms to find an $\\epsilon$-stationary point may be significantly larger than the optimal complexity of $\\Theta \\left( \\Delta L \\sigma^2 \\epsilon^{-4} \\right)$ achieved by SGD in the $L$-smooth setting, where $\\Delta$ is the initial optimality gap, $\\sigma^2$ is the variance of stochastic gradient. However, it is currently not known whether these higher-order dependencies can be tightened. To answer this question, we investigate complexity lower bounds for several adaptive optimization algorithms in the $(L_0, L_1)$-smooth setting, with a focus on the dependence in terms of problem parameters $\\Delta, L_0, L_1$. We provide complexity bounds for three variations of AdaGrad, which show at least a quadratic dependence on problem parameters $\\Delta, L_0, L_1$. Notably, we show that the decorrelated variant of AdaGrad-Norm requires at least $\\Omega \\left( \\Delta^2 L_1^2 \\sigma^2 \\epsilon^{-4} \\right)$ stochastic gradient queries to find an $\\epsilon$-stationary point. We also provide a lower bound for SGD with a broad class of adaptive stepsizes. Our results show that, for certain adaptive algorithms, the $(L_0, L_1)$-smooth setting is fundamentally more difficult than the standard smooth setting, in terms of the initial optimality gap and the smoothness constants.", "keywords": [ "optimization", "nonconvex optimization", "stochastic optimization", "adaptive optimization", "relaxed smoothness", "lower bounds", "adagrad" ], "rating_list": [ 8, 6, 6, 6, 5 ], "soundness_list": [ 4, 4, 2, 3, 3 ], "presentation_list": [ 4, 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 2, 3 ] }, { "paper_id": "ZjKTMmWKHP", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Haoming Liu", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Hongyi Wen", "gender": "unknown", "institution": "New York University, Shanghai", "country": "CN", "position": "Assistant Professor" }, { "name": "Shengjie Wang", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Yijia Cao", "gender": "Female", "institution": "New York University Shanghai", "country": "CN", "position": "Undergrad student" }, { "name": "Yuanhe Guo", "gender": "Male", "institution": "New York University", "country": "US", "position": "Undergrad student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 32, "n_ref": 55, "n_ref_all": 73, "n_fig": 10, "n_tab": 2, "L_tab": 857, "n_element_tab": 70, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1024, "n_element_tab_1": 82, "formula_len_all": 586, "formula_len_all_1": 626, "len_all": 143656, "len_all_1": 70505, "len_abs": 1293, "len_title": 106, "len_sents": 43916, "len_sents_1": 34305, "n_sents": 291, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1322, "title": "Optimal Generative Cyclic Transport between Image and Text", "abs": "Deep generative models, such as vision-language models (VLMs) and diffusion models (DMs), have achieved remarkable success in cross-modality generation tasks. However, the cyclic transformation of text $\\rightarrow$ image $\\rightarrow$ text often fails to secure an exact match between the original and the reconstructed content. In this work, we attempt to address this challenge by utilizing a deterministic function to guide the reconstruction of precise information via generative models. Using a color histogram as guidance, we first identify a soft prompt to generate the desired text using a language model and map the soft prompt to a target histogram. We then utilize the target color histogram as a constraint for the diffusion model and formulate the intervention as an optimal transport problem. As a result, the generated image has the exact color histogram as the target, which can be converted to a soft prompt deterministically for reconstructing the text. This allows the generated images to entail arbitrary forms of text (e.g., natural text, code, URLs, etc.) while ensuring the visual content is as natural as possible. Our method offers significant potential for applications on histogram-constrained generation, such as steganography and conditional generation in latent space with semantic meanings.", "keywords": [ "Cross-modal Information Transformation", "Generative Models", "Optimal Transport" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 5, 3, 5 ] }, { "paper_id": "ZiPoAlKf9Y", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "H Habibullah", "gender": "Male", "institution": "University of South Australia", "country": "", "position": "Lecturer" }, { "name": "Lin Liu", "gender": "Not Specified", "institution": "University of South Australia, Australia", "country": "AU", "position": "Full Professor" }, { "name": "Mahardhika Pratama", "gender": "Male", "institution": "University of South Australia", "country": "", "position": "Associate Professor" }, { "name": "Muhammad Anwar Ma'sum", "gender": "Male", "institution": "University of South Australia", "country": "", "position": "PhD student" }, { "name": "Ryszard Kowalczyk", "gender": "unknown", "institution": "University of South Australia", "country": "AU", "position": "Full Professor" } ], "n_formula": 21, "n_formula_1": 7, "n_ref_uni": 42, "n_ref": 115, "n_ref_all": 152, "n_fig": 9, "n_tab": 17, "L_tab": 13022, "n_element_tab": 1924, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1204, "n_element_tab_1": 81, "formula_len_all": 1681, "formula_len_all_1": 427, "len_all": 250340, "len_all_1": 75325, "len_abs": 1127, "len_title": 93, "len_sents": 68548, "len_sents_1": 33013, "n_sents": 550, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 45, "L_abs": 1143, "title": "Federated Few-Shot Class-Incremental Learning", "abs": "This study proposes a challenging yet practical Federated Few-Shot Class-Incremental Learning (FFSCIL) problem, where clients only hold very few samples for new classes. We develop a novel Unified Optimized Prototype Prompt (UOPP) model to simultaneously handle catastrophic forgetting, over-fitting, and prototype bias in FFSCIL. UOPP utilizes task-wise prompt learning to mitigate task interference and over-fitting, unified static-dynamic prototypes to achieve a stability-plasticity balance, and adaptive dual heads for enhanced inferences. Dynamic prototypes represent new classes in the current few-shot task and are rectified to deal with prototype bias. Our comprehensive experimental results show that UOPP significantly outperforms state-of-the-art (SOTA) methods on three datasets with improvements up to $76\\%$ on average accuracy and $67\\%$ on balance mean accuracy respectively. Our extensive analysis shows UOPP robustness in various numbers of local clients and global rounds, low communication costs, and moderate running time. The source code of UOPP is publicly available at \\url{https://anonymous.4open.science/r/op71m15}.", "keywords": [ "Federated", "Few-Shot", "Class-Incremental Learning", "Prototype-bias", "Rectification" ], "rating_list": [ 8, 6, 6, 5, 6 ], "soundness_list": [ 4, 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 2, 4 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "Zi1QNJKXAD", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Adil Zouitine", "gender": "unknown", "institution": "Institut de recherche technologique Saint-Exup\u00e9ry", "country": "", "position": "Researcher" }, { "name": "Emmanuel Rachelson", "gender": "Male", "institution": "Institut Sup\u00e9rieur de l'A\u00e9ronautique et de l'Espace", "country": "FR", "position": "Full Professor" }, { "name": "Matthieu Geist", "gender": "Male", "institution": "Cohere", "country": "FR", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 3, "n_ref_uni": 42, "n_ref": 73, "n_ref_all": 114, "n_fig": 5, "n_tab": 11, "L_tab": 5924, "n_element_tab": 390, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1804, "n_element_tab_1": 135, "formula_len_all": 2392, "formula_len_all_1": 160, "len_all": 174101, "len_all_1": 79309, "len_abs": 1533, "len_title": 103, "len_sents": 55828, "len_sents_1": 37450, "n_sents": 422, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1543, "title": "Solving robust MDPs as a sequence of static RL problems", "abs": "esigning control policies whose performance level is guaranteed to remain above a given threshold in a span of environments is a critical feature for the adoption of reinforcement learning (RL) in real-world applications. \nThe search for such robust policies is a notoriously difficult problem, related to the so-called dynamic model of transition function uncertainty, where the environment dynamics are allowed to change at each time step.\nBut in practical cases, one is rather interested in robustness to a span of static transition models throughout interaction episodes. \nThe static model is known to be harder to solve than the dynamic one, and seminal algorithms, such as robust value iteration, as well as most recent works on deep robust RL, build upon the dynamic model.\nIn this work, we propose to revisit the static model. \nWe suggest an analysis of why solving the static model under some mild hypotheses is a reasonable endeavor, based on an equivalence with the dynamic model, and formalize the general intuition that robust MDPs can be solved by tackling a series of static problems. \nWe introduce a generic meta-algorithm called IWOCS, which incrementally identifies worst-case transition models so as to guide the search for a robust policy. \nDiscussion on IWOCS sheds light on new ways to decouple policy optimization and adversarial transition functions and opens new perspectives for analysis.\nWe derive a deep RL version of IWOCS and demonstrate it is competitive with state-of-the-art algorithms on classical benchmarks.", "keywords": [ "Robust reinforcement learning" ], "rating_list": [ 6, 3, 3, 1, 3 ], "soundness_list": [ 3, 3, 3, 1, 2 ], "presentation_list": [ 2, 3, 2, 2, 2 ], "contribution_list": [ 3, 1, 2, 1, 2 ], "confidence_list": [ 3, 3, 3, 5, 4 ] }, { "paper_id": "Zhdhg6n2OG", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 11, "author_info_list": [ { "name": "Amitis Shidani", "gender": "Female", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Dan Busbridge", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Eeshan Gunesh Dhekane", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Federico Danieli", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Floris Weers", "gender": "Male", "institution": "Apple", "country": "NL", "position": "Researcher" }, { "name": "Jagrit Digani", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Jason Ramapuram", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Pierre Ablin", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Russell Webb", "gender": "unknown", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Tatiana Likhomanenko", "gender": "Female", "institution": "Apple", "country": "US", "position": "Research Scientist" }, { "name": "Zijin Gu", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 111, "n_formula_1": 13, "n_ref_uni": 43, "n_ref": 121, "n_ref_all": 235, "n_fig": 28, "n_tab": 12, "L_tab": 10121, "n_element_tab": 374, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 1890, "n_element_tab_1": 123, "formula_len_all": 9011, "formula_len_all_1": 891, "len_all": 394116, "len_all_1": 60757, "len_abs": 1456, "len_title": 111, "len_sents": 104088, "len_sents_1": 24381, "n_sents": 871, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1459, "title": "Theory, Analysis, and Best Practices for Sigmoid Self-Attention", "abs": "Attention is a key part of the transformer architecture. It is a sequence-to-sequence mapping that transforms each sequence element into a weighted sum of values. The weights are typically obtained as the softmax of dot products between keys and queries. Recent work has explored alternatives to softmax attention in transformers, such as ReLU and sigmoid activations. In this work, we revisit sigmoid attention and conduct an in-depth theoretical and empirical analysis. Theoretically, we prove that transformers with sigmoid attention are universal function approximators and benefit from improved regularity compared to softmax attention. Through detailed empirical analysis, we identify stabilization of large initial attention norms during the early stages of training as a crucial factor for the successful training of models with sigmoid attention, outperforming prior attempts. We also introduce FLASHSIGMOID, a hardware-aware and memory-efficient implementation of sigmoid attention yielding a 17% inference kernel speed-up over FLASHATTENTION2 on H100 GPUs. Experiments across language, vision, and speech show that properly normalized sigmoid attention matches the strong performance of softmax attention on a wide range of domains and scales, which previous attempts at sigmoid attention were unable to fully achieve. Our work unifies prior art and establishes best practices for sigmoid attention as a drop-in softmax replacement in transformers.", "keywords": [ "Sigmoid Attention", "Pointwise Attention", "Universal Function Approximators" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 5 ] }, { "paper_id": "ZhXJNUEOr9", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Armando Solar-Lezama", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Kavi Gupta", "gender": "Non-Binary", "institution": "Massachussets Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Osbert Bastani", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Assistant Professor" } ], "n_formula": 121, "n_formula_1": 23, "n_ref_uni": 18, "n_ref": 29, "n_ref_all": 53, "n_fig": 10, "n_tab": 3, "L_tab": 390, "n_element_tab": 26, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12536, "formula_len_all_1": 1610, "len_all": 199741, "len_all_1": 64332, "len_abs": 1168, "len_title": 123, "len_sents": 65892, "len_sents_1": 29181, "n_sents": 548, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 927, "title": "Sparling: Learning Latent Representations with Extremely Sparse Activations", "abs": "Real-world processes often contain intermediate state that can be modeled as an extremely sparse activation tensor. In this work, we analyze the identifiability of such sparse and local latent intermediate variables, which we call motifs.\nWe prove our Motif Identifiability Theorem, stating that under certain assumptions it is possible to precisely identify these motifs exclusively by reducing end-to-end error. Additionally, we provide the Sparling algorithm, which uses a new kind of informational bottleneck that enforces levels of activation sparsity unachievable using other techniques. We find that extreme sparsity is necessary to achieve good intermediate state modeling empirically. On our synthetic DigitCircle domain as well as the LaTeXOCR and AudioMNISTSequence domains, we are able to precisely localize the intermediate states up to feature permutation with >90% accuracy, even though we only train end-to-end.", "keywords": [ "machine learning", "sparsity", "interpretability", "optimization", "identifiability" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 1 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 2, 3, 2 ] }, { "paper_id": "Zh9gz3CaWm", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Deming Zhai", "gender": "Female", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Junjun Jiang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Qingbin Wei", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Shiyi Lin", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Xiangyang Ji", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xianming Liu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "\u98de\u9f99 \u5f20", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 38, "n_formula_1": 12, "n_ref_uni": 31, "n_ref": 52, "n_ref_all": 60, "n_fig": 3, "n_tab": 10, "L_tab": 5548, "n_element_tab": 446, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 4460, "n_element_tab_1": 346, "formula_len_all": 2410, "formula_len_all_1": 737, "len_all": 124061, "len_all_1": 56870, "len_abs": 1285, "len_title": 120, "len_sents": 33416, "len_sents_1": 21266, "n_sents": 255, "n_sents_1": 154, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1317, "title": "Communication-Efficient Federated Learning via Model Update Distillation", "abs": "Federated learning (FL) is a popular distributed machine learning framework for edge computing. However, it faces a significant challenge: the communication overhead caused by frequent model updates between clients and the central server. Previous studies have overlooked a crucial piece of information: the central server already knows the initial model on each client before local training begins in every round. This oversight leads to significant redundancy in communication, as full model information are transmitted unnecessarily. To address this, we propose a novel framework called \\textit{model update distillation} (MUD), which leverages this prior knowledge to decouple model parameters from the network architecture. Instead of transmitting raw parameter updates, our method synthesizes and transmits compact tensor sequences that encode only the essential information for synchronization. This dramatically reduces communication overhead while still allowing recipients to accurately reconstruct the intended model updates. Extensive experimental results demonstrate that FedMUD achieves substantial improvements in communication efficiency, making it a highly effective solution for federated learning in bandwidth-constrained environments. The PyTorch-like core code can be found in \\ref{alg: pytorch}.", "keywords": [ "Federated Learning", "Edge Computing", "Communication-Efficient", "Knowledge Distillation" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "Zggz6seq6F", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Jing Zhang", "gender": "Female", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Engineer" }, { "name": "Kang Hao Cheong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Shiyu Hu", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Xin Zhao", "gender": "Male", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "Full Professor" }, { "name": "Xuchen Li", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xuzhao Li", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "\u6021\u6c9b \u738b", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 29, "n_ref": 63, "n_ref_all": 103, "n_fig": 17, "n_tab": 9, "L_tab": 7676, "n_element_tab": 649, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 2006, "n_element_tab_1": 175, "formula_len_all": 1136, "formula_len_all_1": 156, "len_all": 289458, "len_all_1": 87858, "len_abs": 2835, "len_title": 164, "len_sents": 105645, "len_sents_1": 39135, "n_sents": 924, "n_sents_1": 295, "venue": "ICLR2025", "decision": "Reject", "L_title": 116, "L_abs": 2716, "title": "Can LVLMs Describe Videos like Humans? A Five-in-One Video Annotations Benchmark for Better Human-Machine Comparison", "abs": "Large vision-language models (LVLMs) have made significant strides in addressing complex video tasks, sparking researchers' interest in their human-like multimodal understanding capabilities. Video description serves as a fundamental task for evaluating video comprehension, necessitating a deep understanding of spatial and temporal dynamics, which presents challenges for both humans and machines. Thus, investigating whether LVLMs can describe videos as comprehensively as humans\u2014through reasonable human-machine comparisons using video captioning as a proxy task\u2014will enhance our understanding and application of these models. However, current benchmarks for video comprehension have notable limitations, including short video durations, brief annotations, and reliance on a single annotator's perspective. These factors hinder a comprehensive assessment of LVLMs' ability to understand complex, lengthy videos and prevent the establishment of a robust human baseline that accurately reflects human video comprehension capabilities. To address these issues, we propose a novel benchmark, FIOVA (Five In One Video Annotations), designed to evaluate the differences between LVLMs and human understanding more comprehensively. FIOVA includes 3,002 long video sequences (averaging 33.6 seconds) that cover diverse scenarios with complex spatiotemporal relationships. Each video is annotated by five distinct annotators, capturing a wide range of perspectives and resulting in captions that are 4 to 15 times longer than most existing benchmarks, thereby establishing a robust baseline that represents human understanding comprehensively for the first time in video description tasks. Using the FIOVA benchmark, we conducted an in-depth evaluation of six state-of-the-art (SOTA) LVLMs, comparing their performance with humans. To enhance this evaluation, we proposed FIOVA-DQ, a novel event-based metric that incorporates weighted event importance derived from human annotations. Results show that while current LVLMs demonstrate some perception and reasoning capabilities, they still struggle with information omission and descriptive depth. Moreover, we found significant discrepancies between LVLMs and humans in complex videos, particularly where human annotators exhibited substantial disagreement, whereas LVLMs tended to rely on uniform strategies for challenging content. These findings underscore the limitations of using a single human annotator as the groundtruth for evaluation and highlight the need for new evaluation perspectives. We believe this work offers valuable insights into the differences between LVLMs and humans, ultimately guiding future advancements toward human-level video comprehension.", "keywords": [ "Video Caption", "Video Understanding", "LVLM Evaluation", "Human-machine Comparison" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "ZfFQrVoygN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Haozhao Wang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Postdoc" }, { "name": "Ruixuan Li", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Tianzhe Xiao", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yichen Li", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yining Qi", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Postdoc" }, { "name": "Yuying Wang", "gender": "Female", "institution": "Suzhou University", "country": "CN", "position": "MS student" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 17, "n_ref": 29, "n_ref_all": 44, "n_fig": 3, "n_tab": 10, "L_tab": 6986, "n_element_tab": 219, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 3226, "n_element_tab_1": 80, "formula_len_all": 1131, "formula_len_all_1": 873, "len_all": 150312, "len_all_1": 65487, "len_abs": 1246, "len_title": 123, "len_sents": 38089, "len_sents_1": 28133, "n_sents": 309, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1273, "title": "Rehearsal-Free Continual Federated Learning with Synergistic Regularization", "abs": "Continual Federated Learning (CFL) allows distributed devices to collaboratively learn novel concepts from continuously shifting training data while avoiding \\textit{knowledge forgetting} of previously seen tasks. \nTo tackle this challenge, most current CFL approaches rely on extensive rehearsal of previous data. Despite effectiveness, rehearsal comes at a cost to memory, and it may also violate data privacy. \nConsidering these, we seek to apply regularization techniques to CFL by considering their cost-efficient properties that do not require sample caching or rehearsal. Specifically, we first apply traditional regularization techniques to CFL and observe that existing regularization techniques, especially synaptic intelligence, can achieve promising results under homogeneous data distribution but fail when the data is heterogeneous. Based on this observation, we propose a simple yet effective regularization algorithm for CFL named \\textbf{FedSSI}, which tailors the synaptic intelligence for the CFL with heterogeneous data settings. \nFedSSI can not only reduce computational overhead without rehearsal but also address the data heterogeneity issue. \nExtensive experiments show that FedSSI achieves superior performance compared to state-of-the-art methods.", "keywords": [ "Federated Learning", "Data Heterogeneity" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "Zf7EFQt04n", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Ashwin Ashok", "gender": "unknown", "institution": "Georgia State University", "country": "", "position": "Associate Professor" }, { "name": "Bryan Bo Cao", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Faith M Johnson", "gender": "Female", "institution": "Rutgers University", "country": "", "position": "PhD student" }, { "name": "Kristin Dana", "gender": "Female", "institution": "Rutgers University", "country": "US", "position": "Full Professor" }, { "name": "Shubham Jain", "gender": "unknown", "institution": "State University of New York at Stony Brook", "country": "", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 39, "n_ref": 81, "n_ref_all": 92, "n_fig": 2, "n_tab": 2, "L_tab": 1268, "n_element_tab": 184, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1268, "n_element_tab_1": 184, "formula_len_all": 44, "formula_len_all_1": 44, "len_all": 124855, "len_all_1": 61374, "len_abs": 1252, "len_title": 87, "len_sents": 34154, "len_sents_1": 30851, "n_sents": 194, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1256, "title": "Memory Proxy Maps for Visual Navigation", "abs": "Visual navigation takes inspiration from humans, who navigate in previously unseen environments using vision without detailed environment maps. Inspired by this, we introduce a novel no-RL, no-graph, no-odometry approach to visual navigation using feudal learning to build a three tiered agent. Key to our approach is a memory proxy map (MPM), an intermediate representation of the environment\nlearned in a self-supervised manner by the high-level manager agent that serves as a simplified memory, approximating what the agent has seen. We demonstrate that recording observations in this learned latent space is an effective and efficient memory proxy that can remove the need for graphs and odometry in visual navigation tasks. For the mid-level manager agent, we develop a waypoint network\n(WayNet) that outputs intermediate subgoals, or waypoints, imitating human waypoint selection during local navigation. For the low-level worker agent, we learn a classifier over a discrete action space that avoids local obstacles and moves the agent towards the WayNet waypoint. The resulting feudal navigation network offers a novel approach with no RL, no graph, no odometry, and no metric map; all while achieving SOTA results on the image goal navigation task.", "keywords": [ "visual navigation", "representation learning", "hierarchical networks" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Zes7Wyif8G", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Jaron Maene", "gender": "unknown", "institution": "KU Leuven", "country": "BE", "position": "PhD student" }, { "name": "Pedro Zuidberg Dos Martires", "gender": "Male", "institution": "KU Leuven", "country": "", "position": "PhD student" }, { "name": "Vincent Derkinderen", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 43, "n_ref": 72, "n_ref_all": 101, "n_fig": 11, "n_tab": 2, "L_tab": 701, "n_element_tab": 69, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 387, "n_element_tab_1": 38, "formula_len_all": 957, "formula_len_all_1": 762, "len_all": 127357, "len_all_1": 58592, "len_abs": 1038, "len_title": 107, "len_sents": 33992, "len_sents_1": 27388, "n_sents": 276, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1042, "title": "KLay: Accelerating Arithmetic Circuits for Neurosymbolic AI", "abs": "A popular approach to neurosymbolic AI involves mapping logic formulas to arithmetic circuits (computation graphs consisting of sums and products) and passing the outputs of a neural network through these circuits. This approach enforces symbolic constraints onto a neural network in a principled and end-to-end differentiable way. Unfortunately, arithmetic circuits are challenging to run on modern tensor accelerators as they exhibit a high degree of irregular sparsity. To address this limitation, we introduce knowledge layers (KLay), a new data structure to represent arithmetic circuits that can be efficiently parallelized on GPUs. Moreover, we contribute two algorithms used in the translation of traditional circuit representations to KLay and a further algorithm that exploits parallelization opportunities during circuit evaluations. We empirically show that KLay achieves speedups of multiple orders of magnitude over the state of the art, thereby paving the way towards scaling neurosymbolic AI to larger real-world applications.", "keywords": [ "Neurosymbolic AI", "Arithmetic Circuits", "Sparse Inference" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "ZeaTvXw080", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Dvir Samuel", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "PhD student" }, { "name": "Gal Chechik", "gender": "unknown", "institution": "Bar Ilan University", "country": "IL", "position": "Full Professor" }, { "name": "Lior Wolf", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Associate Professor" }, { "name": "Rinon Gal", "gender": "Male", "institution": "NVIDIA", "country": "IL", "position": "Researcher" }, { "name": "Yoad Tewel", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Yuval Atzmon", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 38, "n_ref": 54, "n_ref_all": 92, "n_fig": 22, "n_tab": 4, "L_tab": 940, "n_element_tab": 85, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 905, "n_element_tab_1": 82, "formula_len_all": 620, "formula_len_all_1": 516, "len_all": 137022, "len_all_1": 57087, "len_abs": 1143, "len_title": 0, "len_sents": 46928, "len_sents_1": 27323, "n_sents": 371, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1147, "title": "Add-it: Training-Free Object Insertion in Images With Pretrained Diffusion Models", "abs": "Adding Object into images based on text instructions is a challenging task in semantic image editing, requiring a balance between preserving the original scene and seamlessly integrating the new object in a fitting location. Despite extensive efforts, existing models often struggle with this balance, particularly with finding a natural location for adding an object in complex scenes. We introduce Add-it, a training-free approach that extends diffusion models' attention mechanisms to incorporate information from three key sources: the scene image, the text prompt, and the generated image itself. Our weighted extended-attention mechanism maintains structural consistency and fine details while ensuring natural object placement. Without task-specific fine-tuning, Add-it achieves state-of-the-art results on both real and generated image insertion benchmarks, including our newly constructed \"Additing Affordance Benchmark\" for evaluating object placement plausibility, outperforming supervised methods. Human evaluations show that Add-it is preferred in over 80% of cases, and it also demonstrates improvements in various automated metrics.", "keywords": [ "Diffusion", "Editing", "Affordance" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Ze4aPP0tIn", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Aonan Zhang", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Dong Yin", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Ruoming Pang", "gender": "unknown", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Shengyu Feng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Xiang Kong", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Yiming Yang", "gender": "Female", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Chong Wang", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "shuang ma", "gender": "Female", "institution": "Apple", "country": "", "position": "Researcher" } ], "n_formula": 58, "n_formula_1": 25, "n_ref_uni": 44, "n_ref": 108, "n_ref_all": 133, "n_fig": 5, "n_tab": 3, "L_tab": 927, "n_element_tab": 149, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 346, "n_element_tab_1": 59, "formula_len_all": 3631, "formula_len_all_1": 1669, "len_all": 187417, "len_all_1": 63843, "len_abs": 1173, "len_title": 123, "len_sents": 50880, "len_sents_1": 28389, "n_sents": 455, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1179, "title": "Step-by-Step Reasoning for Math Problems via Twisted Sequential Monte Carlo", "abs": "Augmenting the multi-step reasoning abilities of Large Language Models (LLMs) has been a persistent challenge. Recently, verification has shown promise in improving solution consistency by evaluating generated outputs. However, current verification approaches suffer from sampling inefficiencies, requiring a large number of samples to achieve satisfactory performance. Additionally, training an effective verifier often depends on extensive process supervision, which is costly to acquire. In this paper, we address these limitations by introducing a novel verification method based on Twisted Sequential Monte Carlo (TSMC). TSMC sequentially refines its sampling effort to focus exploration on promising candidates, resulting in more efficient generation of high-quality solutions. We apply TSMC to LLMs by estimating the expected future rewards at partial solutions. This approach results in a more straightforward training target that eliminates the need for step-wise human annotations. We empirically demonstrate the advantages of our method across multiple math benchmarks, and also validate our theoretical analysis of both our approach and existing verification methods.", "keywords": [ "Large Language Models", "Twisted Sequential Monte Carlo", "Reasoning" ], "rating_list": [ 8, 6, 6, 5, 8 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 4, 3, 1, 3 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 2, 2 ] }, { "paper_id": "Ze49bGd4ON", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Jiaqi Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Pan Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Rui Qian", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shuangrui Ding", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xiaoyi Dong", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yuhang CAO", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yuhang Zang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yuwei Guo", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 45, "n_ref": 76, "n_ref_all": 92, "n_fig": 3, "n_tab": 7, "L_tab": 1557, "n_element_tab": 210, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 1583, "n_element_tab_1": 184, "formula_len_all": 334, "formula_len_all_1": 352, "len_all": 128176, "len_all_1": 66584, "len_abs": 1825, "len_title": 133, "len_sents": 32599, "len_sents_1": 30194, "n_sents": 243, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1866, "title": "SAM2Long: Enhancing SAM2 for Long Video Segmentation with a Training-Free Memory Tree", "abs": "The Segment Anything Model 2 (SAM2) has emerged as a powerful foundation model for object segmentation in both images and videos, paving the way for various downstream video applications. The crucial design of SAM2 for video segmentation is its memory module, which prompts object-aware memories from previous frames for current frame prediction. However, its greedy-selection memory design suffers from the ``error accumulation\" problem, where an errored or missed mask will cascade and influence the segmentation of the subsequent frames, which limits the performance of SAM2 toward complex long-term videos. To this end, we introduce SAM2Long, an improved \\textbf{training-free} video object segmentation strategy, which considers the segmentation uncertainty within each frame and chooses the video-level optimal results from multiple segmentation pathways in a constrained tree search manner. In practice, we maintain a fixed number of segmentation pathways throughout the video. For each frame, multiple masks are proposed based on the existing pathways, creating various candidate branches. We then select the same fixed number of branches with higher cumulative scores as the new pathways for the next frame. After processing the final frame, the pathway with the highest cumulative score is chosen as the final segmentation result. Benefiting from its heuristic search design, SAM2Long is robust toward occlusions and object reappearances, and can effectively segment and track objects for complex long-term videos. Without introducing any additional parameters or further training, SAM2Long significantly outperforms SAM2 on six VOS benchmarks. Notably, it achieves an average improvement of 3.8 points across all model sizes and, in some cases, up to 5 points in $\\mathcal{J}$&$\\mathcal{F}$ on long-term video object segmentation benchmarks SA-V and LVOS.", "keywords": [ "Segment Anything Model 2", "Video Object Segmentation" ], "rating_list": [ 5, 3, 8, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "ZdqdWiRRwR", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Camila Gonzalez", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Ehsan Adeli", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Kilian M. Pohl", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "MohammadHassan Abbasi", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Researcher" }, { "name": "Qingyu Zhao", "gender": "Male", "institution": "Weill Cornell Medicine, Cornell University", "country": "", "position": "Assistant Professor" }, { "name": "Yash Shah", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" } ], "n_formula": 17, "n_formula_1": 7, "n_ref_uni": 60, "n_ref": 98, "n_ref_all": 128, "n_fig": 21, "n_tab": 3, "L_tab": 894, "n_element_tab": 96, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 608, "n_element_tab_1": 34, "formula_len_all": 1544, "formula_len_all_1": 615, "len_all": 176994, "len_all_1": 68194, "len_abs": 1098, "len_title": 118, "len_sents": 50140, "len_sents_1": 33445, "n_sents": 368, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1101, "title": "Confounder-Free Continual Learning via Recursive Feature Normalization", "abs": "Confounders are extraneous variable that affect both the input and the target, resulting in spurious correlations and biased predictions. Learning feature representations that are invariant to confounders remains a significant challenge in continual learning. To remove the influence of confounding variables from intermediate feature representations, we introduce the Recursive Metadata Normalization (R-MDN) layer, which can be integrated into any stage within deep neural networks (DNNs). R-MDN performs statistical regression via the recursive least squares algorithm to maintain and continually update an internal model state with respect to changing distributions of data and confounding variables. Since R-MDN operates on the level of individual examples, it is compatible with state-of-the-art architectures like vision transformers. Our experiments demonstrate that R-MDN promotes equitable predictions across population groups, both within static learning and across different stages of continual learning, by reducing catastrophic forgetting caused by confounder effects changing over time.", "keywords": [ "deep neural networks", "confounders", "continual learning", "invariant representations", "statistical regression" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 4, 2, 3, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "Zdl2i7RKmz", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Ian Gemp", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Research Scientist" }, { "name": "Philipp Dominic Siedler", "gender": "Male", "institution": "Aleph Alpha", "country": "DE", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 29, "n_ref_all": 41, "n_fig": 7, "n_tab": 3, "L_tab": 1868, "n_element_tab": 34, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 3508, "n_element_tab_1": 159, "formula_len_all": 696, "formula_len_all_1": 0, "len_all": 123130, "len_all_1": 49695, "len_abs": 1452, "len_title": 85, "len_sents": 40585, "len_sents_1": 21134, "n_sents": 311, "n_sents_1": 146, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 1091, "title": "LLM-Mediated Guidance of MARL Systems", "abs": "In complex multi-agent environments, achieving efficient learning and desirable behaviours is a significant challenge for Multi-Agent Reinforcement Learning (MARL) systems. This work explores the potential of combining MARL with Large Language Model (LLM)-mediated interventions to guide agents toward more desirable behaviours. Specifically, we investigate how LLMs can be used to interpret and facilitate interventions that shape the learning trajectories of multiple agents. We experimented with two types of interventions, referred to as controllers: a Natural Language (NL) Controller and a Rule-Based (RB) Controller. The NL Controller, which uses an LLM to simulate human-like interventions, showed a stronger impact than the RB Controller. Our findings indicate that agents particularly benefit from early interventions, leading to more efficient training and higher performance. Both intervention types outperform the baseline without interventions, highlighting the potential of LLM-mediated guidance to accelerate training and enhance MARL performance in challenging environments.", "keywords": [ "Multi-Agent Reinforcement Learning", "Large Language Model", "Human-AI Interaction", "Multi-Agent System", "Aerial Wildfire Suppression" ], "rating_list": [ 3, 6, 6, 3, 3 ], "soundness_list": [ 1, 3, 2, 2, 1 ], "presentation_list": [ 3, 4, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 1, 1 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "ZdHa3y0DeB", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jian Yang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Lian Jiawei", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Xia Du", "gender": "Male", "institution": "Xiamen University of Technology", "country": "CN", "position": "Lecturer" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 13, "n_ref": 16, "n_ref_all": 42, "n_fig": 10, "n_tab": 11, "L_tab": 3560, "n_element_tab": 524, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 3016, "n_element_tab_1": 446, "formula_len_all": 714, "formula_len_all_1": 628, "len_all": 129939, "len_all_1": 68294, "len_abs": 1777, "len_title": 138, "len_sents": 37773, "len_sents_1": 27435, "n_sents": 286, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1426, "title": "SceneLock: Reversible Adversarial Learning for Camera-Based Autonomous Driving Protection", "abs": "The advancement of autonomous driving technology hinges on large-scale data collection to train camera-based deep neural network 3D object detectors.\nHowever, these valuable datasets are at risk of unauthorized access and misuse by malicious actors, jeopardizing intellectual property, remote deployment, and the privacy of sensitive information captured during data collection. \nWe propose a novel reversible adversarial learning framework, referred to as SceneLock, aimed at protecting autonomous driving data from unauthorized use. \nOur method conducts adversarial perturbations through a carefully designed Noise Serialization Encoding module (NSE), which significantly degrades image quality and renders the data ineffective for unauthorized artificial intelligence models and manual annotation. \nTo ensure legitimate access remains unaffected, we integrate advanced image steganography to embed perturbation values within the images. \nFurthermore, authorized users can extract these values using appropriate decryption tools through the Noise Serialization Decoding module (NSD) to restore the original high-quality images. \nExperimental results demonstrate that our approach effectively safeguards data integrity against unauthorized use while maintaining availability for legitimate purposes. \nThis dual-layer protection highlights the potential of our method to enhance data security in the autonomous driving domain.", "keywords": [ "Data Protection", "Autonomous Driving", "Camera-Based 3D Perception" ], "rating_list": [ 3, 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 1, 3, 3 ], "presentation_list": [ 1, 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 5, 4 ] }, { "paper_id": "Zd8ODMYMBZ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Ben Zhou", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" }, { "name": "Dongwon Jung", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "PhD student" }, { "name": "Muhao Chen", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Assistant Professor" }, { "name": "Qin Liu", "gender": "Female", "institution": "University of California, Davis", "country": "US", "position": "PhD student" }, { "name": "Tenghao Huang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 34, "n_ref": 109, "n_ref_all": 124, "n_fig": 5, "n_tab": 5, "L_tab": 4264, "n_element_tab": 383, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 3334, "n_element_tab_1": 311, "formula_len_all": 304, "formula_len_all_1": 304, "len_all": 141345, "len_all_1": 67624, "len_abs": 1666, "len_title": 121, "len_sents": 35666, "len_sents_1": 29822, "n_sents": 225, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1671, "title": "Familiarity-Aware Evidence Compression for Retrieval-Augmented Generation", "abs": "Retrieval-augmented generation (RAG) improves large language models (LMs) by incorporating non-parametric knowledge through evidence retrieved from external sources. However, it often struggles to cope with inconsistent and irrelevant information that can distract the LM from its tasks, especially when multiple evidence pieces are required. While compressing the retrieved evidence with a compression model aims to address this issue, the compressed evidence may still be unfamiliar to the target model used for downstream tasks, potentially failing to utilize the evidence effectively. We propose FaviComp (Familiarity-aware Evidence Compression), a novel training-free evidence compression technique that makes retrieved evidence more familiar to the target model, while seamlessly integrating parametric knowledge from the model. Specifically, FaviComp proactively composes the compressed evidence in a way to lower the perplexity of the target model by combining decoding probabilities from both the compression model and the target model to generate context that is more familiar to the target model. This approach balances the integration of parametric and non-parametric knowledge, which is especially helpful in complex tasks where the retrieved evidence set may not contain all the necessary information. Experimental results show that FaviComp consistently outperforms most recent evidence compression baselines across multiple open-domain QA datasets, improving accuracy by up to 23.91% while achieving high compression rates. Additionally, we demonstrate the effective integration of both parametric and non-parametric knowledge during evidence compression.", "keywords": [ "Evidence Compression", "Retrieval Augmented Generation", "Parametric and Non-parametric Knowledge" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Zd2T7htqjV", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Danyang Zhuo", "gender": "Male", "institution": "Duke University", "country": "", "position": "Assistant Professor" }, { "name": "Hang Hu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "OMRI WEINSTEIN", "gender": "Male", "institution": "The Hebrew University", "country": "IL", "position": "Associate Professor " }, { "name": "Yichuan Deng", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" } ], "n_formula": 109, "n_formula_1": 14, "n_ref_uni": 63, "n_ref": 107, "n_ref_all": 141, "n_fig": 0, "n_tab": 6, "L_tab": 2550, "n_element_tab": 82, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 388, "n_element_tab_1": 4, "formula_len_all": 10390, "formula_len_all_1": 1451, "len_all": 267623, "len_all_1": 72461, "len_abs": 1511, "len_title": 107, "len_sents": 60765, "len_sents_1": 30350, "n_sents": 657, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1574, "title": "Training Overparametrized Neural Networks in Sublinear Time", "abs": "The success of deep learning comes at a tremendous computational and energy cost, and the scalability of training massively overparametrized neural networks is becoming a real barrier to the progress of artificial intelligence (AI). Despite the popularity and low cost-per-iteration of traditional backpropagation via gradient decent, stochastic gradient descent (SGD) has prohibitive convergence rate in non-convex settings, both in theory and practice. \n\nTo mitigate this cost, recent works have proposed to employ alternative (Newton-type) training methods with much faster convergence rate, albeit with higher cost-per-iteration. \nFor a typical neural network with $m=\\mathrm{poly}(n)$ parameters and input batch of $n$ datapoints in $\\mathbb{R}^d$, the previous work of \\cite{bpsw21} requires $\\sim mnd + n^3$ time per iteration. In this paper, we present a novel training method that requires only $m^{1-\\alpha} n d + n^3$ amortized time in the same overparametrized regime, where $\\alpha \\in (0.01,1)$ is some fixed constant. This method relies on a new and alternative view of neural networks, as a set of binary search trees, where each iteration corresponds to modifying a small subset of the nodes in the tree. We believe this view would have further applications in the design and analysis of deep neural networks (DNNs). We conclude a discussion of lower bound for the dynamic sensitive weight searching data structure we make use of, showing that under {\\sf SETH} or {\\sf OVC} from computational complexity, one cannot substantially improve our algorithm.", "keywords": [ "Overparametrization", "Computational efficiency", "Complexity", "Lower bound", "Training Algorithm" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "ZcBhd1F7PA", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Aming WU", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Assistant Professor" }, { "name": "Muli Yang", "gender": "unknown", "institution": "A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Yihang Zhu", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Yuxuan Wang", "gender": "Female", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Cheng Deng", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 19, "n_ref_uni": 38, "n_ref": 73, "n_ref_all": 106, "n_fig": 9, "n_tab": 17, "L_tab": 2807, "n_element_tab": 416, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 1046, "n_element_tab_1": 100, "formula_len_all": 880, "formula_len_all_1": 917, "len_all": 152993, "len_all_1": 68420, "len_abs": 1432, "len_title": 112, "len_sents": 43675, "len_sents_1": 29714, "n_sents": 374, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1322, "title": "Hierarchical Multi-Grained Reasoning for Object Concept Learning", "abs": "Human beings can easily understand object concepts involving attributes and affordances. Recently, to simulate this ability, Object Concept Learning (OCL) has been introduced as a new task to recognize attributes and affordances related to a given object. \nOCL is essentially a many-to-many mapping problem: While an object may possess multiple different concepts, a concept can also belong to multiple different objects. \nIn this regard, the prevailing method of learning discriminative representation---which is effective in the single-mapping cases---often fails in OCL.\nInspired by the reasoning mechanism of human beings, in this paper, we propose Hierarchical Multi-Grained Reasoning (HGR) for OCL, aiming to infer object-related concepts from coarse-to-fine and counterfactual grains.\nSpecifically, we first propose a coarse-to-fine hierarchical reasoning module that exploits multi-step learnable prompts to progressively localize object-relevant concept information. Subsequently, multiple counterfactual samples are selected to strengthen the relations between objects and concepts, which further improves the reasoning performance. In the experiments, our method is evaluated on multiple benchmarks. Significant performance gains and extensive visualization analysis demonstrate the superiorities of our method.", "keywords": [ "Hierarchical Multi-Grained Reasoning for Object Concept Learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "ZbkqhKbggH", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Gautam Vashishtha", "gender": "Male", "institution": "Skan.ai", "country": "IN", "position": "Researcher" }, { "name": "Marcus Liwicki", "gender": "Unspecified", "institution": "Lule\u00e5 University of Technology", "country": "", "position": "Full Professor" }, { "name": "Mubarak Shah", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Prakash Chandra Chhipa", "gender": "Male", "institution": "Lule\u00e5 University of Technology", "country": "SE", "position": "PhD student" }, { "name": "Rajkumar Saini", "gender": "Male", "institution": "Lulea university of technology ", "country": "", "position": "Lecturer" }, { "name": "Jithamanyu Anantha sai Settur", "gender": "Male", "institution": "Indian Institute of Technology, Madras", "country": "IN", "position": "Undergrad student" } ], "n_formula": 24, "n_formula_1": 12, "n_ref_uni": 35, "n_ref": 73, "n_ref_all": 100, "n_fig": 13, "n_tab": 18, "L_tab": 5653, "n_element_tab": 438, "n_fig_1": 8, "n_tab_1": 8, "L_tab_1": 3328, "n_element_tab_1": 177, "formula_len_all": 2033, "formula_len_all_1": 934, "len_all": 150888, "len_all_1": 61597, "len_abs": 1375, "len_title": 113, "len_sents": 47245, "len_sents_1": 26144, "n_sents": 329, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1504, "title": "ASTrA: Adversarial Self-supervised Training with Adaptive-Attacks", "abs": "Existing self-supervised adversarial training (self-AT) methods rely on hand-crafted adversarial attack strategies for PGD attacks, which fail to adapt to the evolving learning dynamics of the model and do not account for instance-specific characteristics of images. This results in sub-optimal adversarial robustness and limits the alignment between clean and adversarial data distributions. To address this, we propose $\\textit{ASTrA}$ ($\\textbf{A}$dversarial $\\textbf{S}$elf-supervised $\\textbf{Tr}$aining with $\\textbf{A}$daptive-Attacks), a novel framework introducing a learnable, self-supervised attack strategy network that autonomously discovers optimal attack parameters through exploration-exploitation in a single training episode. ASTrA leverages a reward mechanism based on contrastive loss, optimized with REINFORCE, enabling adaptive attack strategies without labeled data or additional hyperparameters. We further introduce a mixed contrastive objective to align the distribution of clean and adversarial examples in representation space. ASTrA achieves state-of-the-art results on CIFAR10, CIFAR100, and STL10 while integrating seamlessly as a plug-and-play module for other self-AT methods. ASTrA shows scalability to larger datasets, demonstrates strong semi-supervised performance, and is resilient to robust overfitting, backed by explainability analysis on optimal attack strategies. Project page for source code and other details at https://prakashchhipa.github.io/projects/ASTrA.", "keywords": [ "Self-supervised Adversarial Training", "Robustness", "Self-supervised Learning", "Adversarial Attacks" ], "rating_list": [ 5, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "ZbOSRZ0JXH", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Edith C. H. Ngai", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Haohan Wang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Sucheng Ren", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Weipeng DENG", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yijiang Li", "gender": "Not Specified", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Ying Gao", "gender": "Female", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yuzhi Xu", "gender": "Not Specified", "institution": "New York University", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 34, "n_ref": 54, "n_ref_all": 69, "n_fig": 5, "n_tab": 5, "L_tab": 563, "n_element_tab": 13, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3323, "n_element_tab_1": 132, "formula_len_all": 476, "formula_len_all_1": 527, "len_all": 139205, "len_all_1": 70150, "len_abs": 1945, "len_title": 139, "len_sents": 32928, "len_sents_1": 29931, "n_sents": 267, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1881, "title": "Beyond Finite Data: Towards Data-free Out-of-distribution Generalization via Extrapolation", "abs": "Out-of-distribution (OOD) generalization is a favorable yet challenging property for deep neural networks. The core challenges lie in the limited availability of source domains that help models learn an invariant representation from the spurious features. Various domain augmentation have been proposed but largely rely on interpolating existing domains and frequently face difficulties in creating truly \"novel\" domains. Humans, on the other hand, can easily extrapolate novel domains, thus, an intriguing question arises: How can neural networks extrapolate like humans and achieve OOD generalization?\nWe introduce a novel approach to domain extrapolation that leverages reasoning ability and the extensive knowledge encapsulated within large language models (LLMs) to synthesize entirely new domains. Starting with the class of interest, we query the LLMs to extract relevant knowledge for these novel domains. We then bridge the gap between the text-centric knowledge derived from LLMs and the pixel input space of the model using text-to-image generation techniques. By augmenting the training set of domain generalization datasets with high-fidelity, photo-realistic images of these new domains, we achieve significant improvements over all existing methods, as demonstrated in both single and multi-domain generalization across various benchmarks.\nWith the ability to extrapolate any domains for any class, our method has the potential to learn a generalized model for any task without any data. To illustrate, we put forth a much more difficult setting termed, data-free domain generalization, that aims to learn a generalized model in the absence of any collected data. Our empirical findings support the above argument and our methods exhibit commendable performance in this setting, even surpassing the supervised setting by approximately 1-2\\% on datasets such as VLCS.", "keywords": [ "domain generalization", "out-of-distribution generalization", "large language model" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3, 2 ], "contribution_list": [ 1, 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "ZbHIDgDFN0", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Che Lin", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Full Professor" }, { "name": "Chih-Yu Wang", "gender": "unknown", "institution": "Academia Sinica", "country": "TW", "position": "Full Professor" }, { "name": "Ming-Yi Hong", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "PhD student" }, { "name": "Shao-En Lin", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "MS student" }, { "name": "Yi-Hsiang Huang", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "You-Chen Teng", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 37, "n_ref": 85, "n_ref_all": 124, "n_fig": 17, "n_tab": 5, "L_tab": 2479, "n_element_tab": 262, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 415, "n_element_tab_1": 39, "formula_len_all": 668, "formula_len_all_1": 699, "len_all": 146656, "len_all_1": 64446, "len_abs": 1284, "len_title": 147, "len_sents": 45830, "len_sents_1": 29554, "n_sents": 354, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1292, "title": "SynHING: Synthetic Heterogeneous Information Network Generation for Graph Learning and Explanation", "abs": "Graph Neural Networks (GNNs) excel in modeling graph structures across diverse domains, such as community analysis and recommendation systems. As the need for GNN interpretability grows, there is an increasing demand for robust baselines and comprehensive graph datasets, especially within the realm of Heterogeneous Information Networks (HIN). To address this, we introduce SynHING, a framework for Synthetic Heterogeneous Information Network Generation designed to advance graph learning and explanation.\nAfter identifying key motifs in a target HIN, SynHING systematically employs a bottom-up generation process with intra-cluster and inter-cluster merge modules. This process, along with post-pruning techniques, ensures that the synthetic HIN accurately mirrors the structural and statistical properties of the original graph. The effectiveness of SynHING is validated using four datasets - IMDB, Recipe, ACM, and DBLP - spanning three distinct application categories, demonstrating both its generality and practicality.\nFurthermore, SynHING provides ground-truth motifs for evaluating GNN explainer models, establishing a new benchmark for explainable, synthetic HIN generation. This contributes significantly to advancing interpretable machine learning in complex network environments.", "keywords": [ "synthetic graph generation", "heterogeneous information networks", "graph neural networks", "explainable artificial intelligence" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "Zb2Ukmte7A", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Bo Li", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Hao Zhang", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "Jinwei Chen", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Junhao Gu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Peng-Tao Jiang", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "Wenming Yang", "gender": "Male", "institution": "Tsinghua University,", "country": "CN", "position": "Associate Professor" }, { "name": "Zhou Mi", "gender": "Male", "institution": "vivo", "country": "CN", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 43, "n_ref": 96, "n_ref_all": 112, "n_fig": 10, "n_tab": 4, "L_tab": 2740, "n_element_tab": 307, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 2958, "n_element_tab_1": 276, "formula_len_all": 328, "formula_len_all_1": 337, "len_all": 128081, "len_all_1": 60647, "len_abs": 1479, "len_title": 129, "len_sents": 30508, "len_sents_1": 26508, "n_sents": 227, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1489, "title": "ConsisSR: Delving Deep into Consistency in Diffusion-based Image Super-Resolution", "abs": "Real-world image super-resolution (Real-ISR) aims at restoring high-quality (HQ) images from low-quality (LQ) inputs corrupted by unknown and complex degradations. In particular, pretrained text-to-image (T2I) diffusion models provide strong generative priors to reconstruct credible and intricate details. However, T2I generation focuses on semantic consistency while Real-ISR emphasizes pixel-level reconstruction, which hinders existing methods from fully exploiting diffusion priors. To address this challenge, we introduce ConsisSR to handle both semantic and pixel-level consistency. Specifically, compared to coarse-grained text prompts, we exploit the more powerful CLIP image embedding and effectively leverage both modalities through our Hybrid Prompt Adapter (HPA) for semantic guidance. Secondly, we introduce Time-aware Latent Augmentation (TALA) to mitigate the inherent gap between T2I generation and Real-ISR consistency requirements. By randomly mixing LQ and HQ latent inputs, our model not only handle timestep-specific diffusion noise but also refine the accumulated latent representations. Last but not least, our GAN-Embedding strategy employs the pretrained Real-ESRGAN model to refine the diffusion start point. This accelerates the inference process to 10 steps while preserving sampling quality, in a training-free manner. Our method demonstrates state-of-the-art performance among both full-scale and accelerated models. The code will be made publicly available.", "keywords": [ "Diffusion Model", "Super-Resolution", "Real-World Image Super-Resolution" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "ZaudLwn0Hm", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Daiwei Yu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Intern" }, { "name": "Zhuorong Li", "gender": "Female", "institution": "Hangzhou City University", "country": "CN", "position": "Lecturer" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 41, "n_ref": 88, "n_ref_all": 96, "n_fig": 3, "n_tab": 2, "L_tab": 1860, "n_element_tab": 222, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1860, "n_element_tab_1": 222, "formula_len_all": 520, "formula_len_all_1": 521, "len_all": 110254, "len_all_1": 77640, "len_abs": 1706, "len_title": 130, "len_sents": 27984, "len_sents_1": 27889, "n_sents": 190, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1715, "title": "Prototypical evoluation for few-shot learning in vision-language model adaptation", "abs": "Vision-Language Models (e.g., CLIP), with their immense capacity and extensive exposure to vast data during pre-training, have demonstrated a strong ability to capture real-world concepts. When fast adapted to downstream tasks with only a few labeled samples, parameter-efficient methods, such as prompt-based and adapter-based approaches, which adjust only a small portion of the parameters, have proven effective in reducing the escalating costs in large vision-language models. However, conventional efficient fine-tuning techniques, using task-specific objectives like cross-entropy loss, often lead to overfitting the downstream data distributions. This overfitting diminishes the model\u2019s ability to retain its original generalization capacity, especially on out-of-distribution (OOD) samples. Unlike the pretraining stage, where rich textual descriptions are available, fine-tuning is typically constrained to using only class names. This creates suboptimal text-image alignment in the shared feature space, as it may exacerbate image feature variance within the same class. To address this issue, we propose Prototypical Evolutionary Adaptation (PEA), leveraging off-the-shelf image centroids as prototypes to regulate image feature variance, mitigating the excessive feature variance within the same class caused by selective bias. Additionally, we introduce learnable shift vectors to capture the dynamics of class prototypes, ensuring that they remain compact and informative. Experiments across diverse datasets and model architectures in few-shot learning demonstrate that our approach consistently outperforms existing methods while maintaining robust generalization under varying distribution shifts.", "keywords": [ "CLIP", "few-shot classification" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 1, 1, 1 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "ZadnlOHsHv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Boyan Gao", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "David A. Clifton", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Guoqi Li", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Jiajun Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Li Du", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Wanpeng Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xingrun Xing", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zheng Liu", "gender": "unknown", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Zheng Zhang", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "", "position": "Principal Researcher" }, { "name": "Shi Tao Xiao", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 49, "n_ref": 121, "n_ref_all": 149, "n_fig": 8, "n_tab": 11, "L_tab": 3583, "n_element_tab": 401, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1852, "n_element_tab_1": 196, "formula_len_all": 954, "formula_len_all_1": 822, "len_all": 164911, "len_all_1": 61986, "len_abs": 1628, "len_title": 143, "len_sents": 40867, "len_sents_1": 26628, "n_sents": 309, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1671, "title": "SpikeLLM: Scaling up Spiking Neural Network to Large Language Models via Saliency-based Spiking", "abs": "The recent advancements in large language models (LLMs) with billions of parameters have significantly boosted their performance across various real-world applications. However, the inference processes for these models require substantial energy and computational resources, presenting considerable deployment challenges. In contrast, human brains, which contain approximately 86 billion biological neurons, exhibit significantly greater energy efficiency compared to LLMs with a similar number of parameters. Inspired by this, we redesign 7~70 billion parameter LLMs using bio-plausible spiking mechanisms, emulating the efficient behavior of the human brain. We propose the first spiking large language model termed SpikeLLM. Coupled with the proposed model, two essential approaches are proposed to improve spiking efficiency: Generalized Integrate-and-Fire (GIF) neurons to compress spike length from $T$ to $\\frac{T}{L} \\log_2 L$ bits, and an Optimal Brain Spiking framework to divide outlier channels and allocate different $T$ for GIF neurons, which further compresses spike length to approximate $log_2T$ bits. The necessity of spike-driven LLM is proved by comparison with quantized LLMs with similar operations. In the OmniQuant pipeline, SpikeLLM reduces 24.85% WikiText2 perplexity and improves 2.01% accuracy of common scene reasoning on a LLAMA2-7B 4A4W model. In the GPTQ pipeline, SpikeLLM achieves direct additive in linear layers, significantly exceeding PB-LLMs. In the LLAMA-2-7B, SpikeLLM saves $\\times 10.79$ and $\\times 6.38$ operations with general matrix multiply and event-driven implementations respectively. We will release our code on GitHub.", "keywords": [ "spiking neural network", "energy efficient language model", "brain-inspired computing" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "ZaSOGF8Ojq", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Cheng Xin", "gender": "Not Specified", "institution": "Rutgers University", "country": "US", "position": "Postdoc" }, { "name": "Fan Xu", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Jiaxin Ding", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Gao", "gender": "unknown", "institution": "Rutgers University", "country": "US", "position": "Full Professor" }, { "name": "Xin Ding", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 15, "n_ref_uni": 51, "n_ref": 116, "n_ref_all": 126, "n_fig": 4, "n_tab": 5, "L_tab": 1713, "n_element_tab": 124, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1055, "n_element_tab_1": 96, "formula_len_all": 1182, "formula_len_all_1": 1584, "len_all": 166410, "len_all_1": 78216, "len_abs": 1711, "len_title": 134, "len_sents": 48657, "len_sents_1": 36624, "n_sents": 406, "n_sents_1": 285, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1633, "title": "TopInG: Topologically Interpretable Graph Learning via Persistent Rationale Filtration", "abs": "Graph Neural Networks (GNNs) have shown remarkable performance in various scientific domains, but their lack of interpretability limits their applicability in critical decision-making processes. Recently, intrinsic interpretable GNNs have been studied to provide insights into model predictions by identifying rationale substructures in graphs. However, existing methods face challenges when the underlying rationale subgraphs are complicated and variable. To address this challenge,\nwe propose TopIng, a novel topological framework to interpretable GNNs that leverages persistent homology to identify persistent rationale subgraphs.\nOur method introduces a rationale filtration learning technique that models the generating procedure of rationale subgraphs, and enforces the persistence of topological gap between rationale subgraphs and complement random graphs by a novel self-adjusted topological constraint, topological discrepancy. We show that our topological discrepancy is a lower bound of a Wasserstein distance on graph distributions with Gromov-Hausdorff metric. \nWe provide theoretical guarantees showing that our loss is uniquely optimized by the ground truth under certain conditions.\nThrough extensive experiments on varaious synthetic and real datasets, we demonstrate that TopIng effectively addresses key challenges in interpretable GNNs including handling variiform rationale subgraphs, balancing performance with interpretability, and avoiding spurious correlations. \nExperimental results show that our approach improves state-of-the-art methods up to 20%+ on both predictive accuracy and interpretation quality.", "keywords": [ "topological data analysis", "persistent homology", "graph neural network", "interpretability", "explainability", "filtration learning" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "ZaOHSBGOhV", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Heng Wang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "Undergrad student" }, { "name": "Jacob Steinhardt", "gender": "unknown", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" }, { "name": "Jiaxin Wen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Ruiqi Zhong", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 0, "n_ref": 51, "n_ref_all": 94, "n_fig": 6, "n_tab": 6, "L_tab": 869, "n_element_tab": 93, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 206, "formula_len_all_1": 73, "len_all": 203717, "len_all_1": 68292, "len_abs": 1364, "len_title": 128, "len_sents": 75940, "len_sents_1": 34573, "n_sents": 605, "n_sents_1": 272, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1368, "title": "SmartBackdoor: Malicious Language Model Agents that Avoid Being Caught", "abs": "As large language model (LLM) agents receive more information about themselves\nor the users from the environment, we speculate a new family of cyber attacks,\nSmartBackdoor: in this attack, malicious actors provide a backdoored LLM agent;\nwhen the victim uses the agent, the agent uses information from its environment to\ndetect whether it is overseen by the victim user; if not, the agent acts maliciously\nagainst the victim. To illustrate this family of attack, we use AutoGPT as a case\nstudy and provide a proof-of-concept: to exfiltrate a private key without being\ncaught, a backdoored LLM agent can analyze the command running itself or infer\nthe skill level of the human user, thus predicting whether it will get caught. To\nevaluate current LLMs\u2019 potential to perform such an attack, we propose a dataset of\nLLM agent scaffolds and benchmark LLMs\u2019 capability to analyze them and reason\nabout human overseers. The current best LLMs (as of 08/2024) fail to robustly\nperform this task, indicating that the current risk of SmartBackdoor is low. Finally,\nwhile our proof-of-concept is unsuccessful in reality and can be exposed by simple\ndefenses (e.g. monitoring system logs or forbidding internet connections), few\nof them are currently commonly adopted by practitioners and none is sufficient\nagainst future SmartBackdoor. We need better LLM agent safety protocols.", "keywords": [ "LLM Agent", "AI safety", "Backdoor Attack", "Deception" ], "rating_list": [ 5, 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "Za3M6OZuCU", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Deniz Gunduz", "gender": "unknown", "institution": "Imperial College London", "country": "", "position": "Full Professor" }, { "name": "Gongpu Chen", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "Postdoc" }, { "name": "Haotian Wu", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 80, "n_formula_1": 33, "n_ref_uni": 41, "n_ref": 51, "n_ref_all": 88, "n_fig": 22, "n_tab": 6, "L_tab": 2582, "n_element_tab": 172, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 521, "n_element_tab_1": 93, "formula_len_all": 6925, "formula_len_all_1": 1894, "len_all": 200164, "len_all_1": 70937, "len_abs": 1585, "len_title": 131, "len_sents": 64518, "len_sents_1": 31488, "n_sents": 588, "n_sents_1": 268, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1593, "title": "Actions Speak Louder Than Words: Rate-Reward Trade-off in Markov Decision Processes", "abs": "The impact of communication on decision-making systems has been extensively studied under the assumption of dedicated communication channels. We instead consider communicating through actions, where the message is embedded into the actions of an agent which interacts with the environment in a Markov decision process (MDP) framework. We conceptualize the MDP environment as a finite-state channel (FSC), where the actions of the agent serve as the channel input, while the states of the MDP observed by another agent (i.e., receiver) serve as the channel output. Here, we treat the environment as a communication channel over which the agent communicates through its actions, while at the same time, trying to maximize its reward. We first characterize the optimal information theoretic trade-off between the average reward and the rate of reliable communication in the infinite-horizon regime. Then, we propose a novel framework to design a joint control/coding policy, termed Act2Comm, which seamlessly embeds messages into actions. From a communication perspective, Act2Comm functions as a learning-based channel coding scheme for non-differentiable FSCs under input-output constraints. From a control standpoint, Act2Comm learns an MDP policy that incorporates communication capabilities, though at the cost of some control performance. Overall, Act2Comm effectively balances the dual objectives of control and communication in this environment. Experimental results validate Act2Comm's capability to enable reliable communication while maintaining a certain level of control performance.", "keywords": [ "Markov Decision Process", "Channel coding", "Rate-Reward Trade-off", "Finite state channel" ], "rating_list": [ 8, 6, 5, 8 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 2 ] }, { "paper_id": "ZZwP9zljas", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Brooks Paige", "gender": "Male", "institution": "University College London", "country": "", "position": "Associate Professor" }, { "name": "Jasraj Singh", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "MS student" }, { "name": "Keyue Jiang", "gender": "Male", "institution": "University College London", "country": "", "position": "PhD student" }, { "name": "Laura Toni", "gender": "unknown", "institution": "University College London", "country": "GB", "position": "Associate Professor" } ], "n_formula": 38, "n_formula_1": 23, "n_ref_uni": 59, "n_ref": 136, "n_ref_all": 174, "n_fig": 6, "n_tab": 2, "L_tab": 583, "n_element_tab": 69, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 583, "n_element_tab_1": 69, "formula_len_all": 3023, "formula_len_all_1": 1545, "len_all": 186995, "len_all_1": 60194, "len_abs": 1786, "len_title": 122, "len_sents": 48020, "len_sents_1": 26669, "n_sents": 380, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1799, "title": "Effects of Random Edge-Dropping on Over-Squashing in Graph Neural Networks", "abs": "Message Passing Neural Networks (MPNNs) are a class of Graph Neural Networks (GNNs) that leverage the graph topology to propagate messages across increasingly larger neighborhoods. The message-passing scheme leads to two distinct challenges: over-smoothing and over-squashing. While several algorithms, e.g. DropEdge and its variants \u2013 DropNode, DropAgg and DropGNN \u2013 have successfully addressed the over-smoothing problem, their impact on over-squashing remains largely unexplored. This represents a critical gap in the literature as failure to mitigate over-squashing would make these methods unsuitable for long-range tasks. In this work, we take the first step towards closing this gap by studying the aforementioned algorithms in the context of over-squashing. We present novel theoretical results that characterize the negative effects of DropEdge on sensitivity between distant nodes, suggesting its unsuitability for long-range tasks. Our findings are easily extended to its variants, allowing us to build a comprehensive understanding of how they affect over squashing. We evaluate these methods using real-world datasets, demonstrating their detrimental effects. Specifically, we show that while DropEdge-variants improve test-time performance in short-range tasks, they deteriorate performance in long-range ones. Our theory explains these results as follows: random edge-dropping lowers the effective receptive field of GNNs, which although beneficial for short-range tasks, misaligns the models on long-range ones. This forces the models to overfit to short-range artefacts in the training set, resulting in poor generalization. Our conclusions highlight the need to re-evaluate various methods designed for training deep GNNs, with a renewed focus on modelling long-range interactions.", "keywords": [ "Graph Neural Networks", "Over-squashing", "DropEdge" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "ZZrSOMLoau", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Fei Wang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Muhao Chen", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Assistant Professor" }, { "name": "Qin Liu", "gender": "Female", "institution": "University of California, Davis", "country": "US", "position": "PhD student" }, { "name": "Tinghui Zhu", "gender": "Female", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Zhengzhong Tu", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 8, "n_ref_uni": 43, "n_ref": 71, "n_ref_all": 83, "n_fig": 3, "n_tab": 5, "L_tab": 3476, "n_element_tab": 303, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 899, "n_element_tab_1": 87, "formula_len_all": 837, "formula_len_all_1": 326, "len_all": 153682, "len_all_1": 68607, "len_abs": 1338, "len_title": 125, "len_sents": 41110, "len_sents_1": 32249, "n_sents": 324, "n_sents_1": 243, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1350, "title": "Unraveling Cross-Modality Knowledge Conflicts in Large Vision-Language Models", "abs": "Large Vision-Language Models (LVLMs) have demonstrated impressive capabilities for capturing and reasoning over multimodal inputs.\nHowever, these models are prone to parametric knowledge conflicts, which arise from inconsistencies of represented knowledge between their vision and language components.\nIn this paper, we formally define the problem of **cross-modality parametric knowledge conflict** and present a systematic approach to detect, interpret, and mitigate them.\nWe introduce a pipeline that identifies conflicts between visual and textual answers, showing a persistently high conflict rate across modalities in recent LVLMs regardless of the model size.\nWe further investigate how these conflicts interfere with the inference process and propose a contrastive metric to discern the conflicting samples from the others.\nBuilding on these insights, we develop a novel dynamic contrastive decoding method that removes undesirable logits inferred from the less confident modality components based on answer confidence. \nFor models that do not provide logits, we also introduce two prompt-based strategies to mitigate the conflicts.\nOur methods achieve promising improvements in accuracy on both the ViQuAE and InfoSeek datasets.\nSpecifically, using LLaVA-34B, our proposed dynamic contrastive decoding improves an average accuracy of 2.24\\%.", "keywords": [ "Knowledge Conflict", "Vision and Language" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "ZZVOrId3yN", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Chuanqi Shi", "gender": "Male", "institution": "Huazhong Agricultural University", "country": "", "position": "Undergrad student" }, { "name": "Hao Bo Dong", "gender": "Male", "institution": "Harbin University of Commerce", "country": "CN", "position": "Undergrad student" }, { "name": "Jiacheng Shi", "gender": "Male", "institution": "College of William and Mary", "country": "US", "position": "PhD student" }, { "name": "Tao Feng", "gender": "unknown", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Xiaofei Sun", "gender": "Female", "institution": "EchoX Technology Limited", "country": "", "position": "Researcher" }, { "name": "YANG HAN", "gender": "Male", "institution": "Didi Research", "country": "CN", "position": "Researcher" }, { "name": "Zhenghan Chen", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Yi Ge", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Research Assistant" } ], "n_formula": 39, "n_formula_1": 34, "n_ref_uni": 3, "n_ref": 7, "n_ref_all": 10, "n_fig": 1, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2471, "formula_len_all_1": 2359, "len_all": 71062, "len_all_1": 51550, "len_abs": 1926, "len_title": 152, "len_sents": 21601, "len_sents_1": 19861, "n_sents": 203, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1437, "title": "CrossModalNet: Multimodal Medical Segmentation with Guaranteed Cross-Modal Flow and Domain Adaptability", "abs": "The fusion of multimodal data in medical image segmentation has emerged as a critical frontier in biomedical research, promising unprecedented diagnostic precision and insights. However, the intricate challenge of effectively integrating diverse data streams while preserving their unique characteristics has persistently eluded comprehensive solutions. This study introduces CrossModalNet, a groundbreaking architecture that revolutionizes multimodal medical image segmentation through advanced mathematical frameworks and innovative domain adaptation techniques. We present a rigorous mathematical analysis of CrossModalNet, proving its universal approximation capabilities and deriving tight generalization bounds. Furthermore, we introduce the Cross-Modal Information Flow (CMIF) metric, providing theoretical justification for the progressive integration of multimodal information through the network layers. Our Joint Adversarial Domain Adaptation (JADA) framework addresses the critical issue of domain shift, simultaneously aligning marginal and conditional distributions while preserving topological structures. Extensive experiments on the MM-WHS dataset demonstrate CrossModalNet's superior performance. This work not only advances the field of medical image segmentation but also provides a robust theoretical foundation for future research in multimodal learning and domain adaptation across various biomedical applications.", "keywords": [ "biomedical imaging", "transfer learning" ], "rating_list": [ 1, 3, 3, 5 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 2, 1, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "ZYwLfi50GI", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Deqing Sun", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Huaizu Jiang", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Assistant Professor" }, { "name": "Xiaogang Peng", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Yiming Xie", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Zizhao Wu", "gender": "Male", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "Associate Professor" }, { "name": "Varun Jampani", "gender": "Male", "institution": "Stability AI", "country": "US", "position": "Principal Researcher" } ], "n_formula": 19, "n_formula_1": 8, "n_ref_uni": 51, "n_ref": 142, "n_ref_all": 185, "n_fig": 16, "n_tab": 6, "L_tab": 1824, "n_element_tab": 351, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 815, "n_element_tab_1": 173, "formula_len_all": 1184, "formula_len_all_1": 496, "len_all": 194712, "len_all_1": 71777, "len_abs": 268, "len_title": 134, "len_sents": 55755, "len_sents_1": 31109, "n_sents": 472, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1227, "title": "HOI-Diff: Text-Driven Synthesis of 3D Human-Object Interactions using Diffusion Models", "abs": "We address the problem of generating realistic 3D human object interactions (HOIs) driven by textual prompts. To this end, we take a modular design and decompose the complex task into simpler subtasks. We first develop a dual-branch diffusion model (DBDM) to generate both human and object motions conditioned on the input text, and encourage coherent motions by a cross-attention communication module between the human and object motion generation branches. We also develop an affordance prediction diffusion model (APDM) to predict the contacting area between the human and object during the interactions driven by the textual prompt. The APDM is independent of the results by the DBDM and thus can correct potential errors by the latter. Moreover, it stochastically generates the contacting points to diversify the generated motions. Finally, we incorporate the estimated contacting points into the classifier-guidance to achieve accurate and close contact between humans and objects. To train and evaluate our approach, we annotate BEHAVE dataset with text descriptions. Experimental results on BEHAVE and OMOMO demonstrate that our approach produces realistic HOIs with various interactions and different types of objects.", "keywords": [ "human motion generation", "human-object interaction" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "ZYuiuxB7H4", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Adams Wai-Kin Kong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Fan Wang", "gender": "Male", "institution": "SAP SE", "country": "SG", "position": "Researcher" } ], "n_formula": 37, "n_formula_1": 12, "n_ref_uni": 48, "n_ref": 86, "n_ref_all": 108, "n_fig": 4, "n_tab": 7, "L_tab": 1954, "n_element_tab": 191, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 629, "n_element_tab_1": 83, "formula_len_all": 1942, "formula_len_all_1": 420, "len_all": 153666, "len_all_1": 64255, "len_abs": 1288, "len_title": 121, "len_sents": 45083, "len_sents_1": 30395, "n_sents": 339, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1286, "title": "Certified $\\ell_2$ Attribution Robustness via Uniformly Smoothed Attributions", "abs": "Model attribution is a popular tool to explain the rationales behind model predictions. However, recent work suggests that the attributions are vulnerable to minute perturbations, which can be added to input samples to fool the attributions while maintaining the prediction outputs. Although empirical studies have shown positive performance via adversarial training, an effective certified defense method is eminently needed to understand the robustness of attributions. In this work, we propose to use uniform smoothing technique that augments the vanilla attributions by noises uniformly sampled from a certain space. It is proved that, for all perturbations within the attack region, the cosine similarity between uniformly smoothed attribution of perturbed sample and the unperturbed sample is guaranteed to be lower bounded. We also derive alternative formulations of the certification that is equivalent to the original one and provides the maximum size of perturbation or the minimum smoothing radius such that the attribution can not be perturbed. We evaluate the proposed method on three datasets and show that the proposed method can effectively protect the attributions from attacks, regardless of the architecture of networks, training schemes and the size of the datasets.", "keywords": [ "Explainability", "Robustness" ], "rating_list": [ 6, 6, 1, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "ZYd5wJSaMs", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Martial Hebert", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Professor" }, { "name": "Ruoyu Zhao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Shuhong Zheng", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Yu-Xiong Wang", "gender": "unknown", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zhipeng Bao", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 5, "n_ref_uni": 55, "n_ref": 183, "n_ref_all": 225, "n_fig": 7, "n_tab": 12, "L_tab": 1621, "n_element_tab": 164, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 277, "n_element_tab_1": 18, "formula_len_all": 429, "formula_len_all_1": 257, "len_all": 192118, "len_all_1": 47274, "len_abs": 1255, "len_title": 123, "len_sents": 55214, "len_sents_1": 22918, "n_sents": 385, "n_sents_1": 147, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1258, "title": "Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models", "abs": "Beyond high-fidelity image synthesis, diffusion models have recently exhibited promising results in dense visual perception tasks. However, most existing work treats diffusion models as a standalone component for perception tasks, employing them either solely for off-the-shelf data augmentation or as mere feature extractors. In contrast to these isolated and thus sub-optimal efforts, we introduce a integrated, versatile, diffusion-based framework, Diff-2-in-1, that can simultaneously handle both multi-modal data generation and dense visual perception, through a unique exploitation of the diffusion-denoising process. Within this framework, we further enhance discriminative visual perception via multi-modal generation, by utilizing the denoising network to create multi-modal data that mirror the distribution of the original training set. Importantly, Diff-2-in-1 optimizes the utilization of the created diverse and faithful data by leveraging a novel self-improving learning mechanism. Comprehensive experimental evaluations validate the effectiveness of our framework, showcasing consistent performance improvements across various discriminative backbones and high-quality multi-modal data generation characterized by both realism and usefulness.", "keywords": [ "Diffusion Models", "Generation", "Dense Perception" ], "rating_list": [ 6, 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 4, 3, 2 ], "presentation_list": [ 1, 3, 4, 2, 3 ], "contribution_list": [ 3, 2, 4, 3, 2 ], "confidence_list": [ 3, 4, 4, 5, 5 ] }, { "paper_id": "ZYUR3HVSAT", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Hongyi Guo", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "", "position": "PhD student" }, { "name": "Jiaheng Wei", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "wei shen", "gender": "unknown", "institution": "Baichuan", "country": "", "position": "Researcher" }, { "name": "Xiaoying Zhang", "gender": "Female", "institution": "ByteDance AILab", "country": "", "position": "Researcher" }, { "name": "Yang Liu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "Zhaoran Wang", "gender": "Not Specified", "institution": "Northwestern University", "country": "US", "position": "Associate Professor" }, { "name": "Kevin Yao", "gender": "Male", "institution": "Meta GenAI", "country": "US", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 17, "n_ref": 48, "n_ref_all": 58, "n_fig": 3, "n_tab": 7, "L_tab": 2223, "n_element_tab": 145, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2491, "n_element_tab_1": 187, "formula_len_all": 432, "formula_len_all_1": 533, "len_all": 123841, "len_all_1": 63657, "len_abs": 1214, "len_title": 62, "len_sents": 42472, "len_sents_1": 26879, "n_sents": 388, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1250, "title": "Human-Instruction-Free LLM Self-Alignment with Limited Samples", "abs": "Aligning large language models (LLMs) with human values is a vital task for LLM practitioners. Current alignment techniques have several limitations: (1) requiring a large amount of annotated data; (2) demanding heavy human involvement; (3) lacking a systematic mechanism to continuously improve. In this work, we study aligning LLMs to a new domain with limited samples (e.g. < 100). We propose an algorithm that can \\textit{self-align} LLMs \\textit{iteratively} without active human involvement. Unlike existing works, our algorithm relies on neither human-crafted instructions nor labeled rewards, significantly reducing human involvement. In addition, our algorithm can self-improve the alignment continuously. The key idea is to first retrieve high-quality samples related to the target domain and use them as In-context Learning examples to generate more samples. Then we use the \\textit{self-generated} samples to finetune the LLM iteratively. We show that our method can unlock the LLMs' self-generalization ability to perform alignment with near-zero human supervision. We test our algorithm on three benchmarks in safety, truthfulness, and instruction-following, and show good performance in alignment, domain adaptability, and scalability.", "keywords": [ "LLM", "Self-Alignment", "ICL" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "ZYDEJEvCbv", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Arash Vahdat", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Giulia Fanti", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Karsten Kreis", "gender": "unknown", "institution": "NVIDIA", "country": "CA", "position": "Research Scientist" }, { "name": "Sangyun Lee", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Tomas Geffner", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Weili Nie", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Yilun Xu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 41, "n_formula_1": 17, "n_ref_uni": 24, "n_ref": 64, "n_ref_all": 127, "n_fig": 12, "n_tab": 11, "L_tab": 2866, "n_element_tab": 458, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 2008, "n_element_tab_1": 277, "formula_len_all": 4345, "formula_len_all_1": 1318, "len_all": 160757, "len_all_1": 69811, "len_abs": 1197, "len_title": 76, "len_sents": 46150, "len_sents_1": 29370, "n_sents": 375, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 28, "L_abs": 1216, "title": "Truncated Consistency Models", "abs": "Consistency models have recently been introduced to accelerate the generation speed of diffusion models by directly predicting the solution (data) of the probability flow ODE (PF ODE) from initial noise.\nHowever, the training of consistency models requires learning to map all intermediate points along PF ODE trajectories to their corresponding endpoints. This task is much more challenging than the ultimate objective of one-step generation, which only concerns the PF ODE's noise-to-data mapping.\nWe empirically find that this training paradigm limits the one-step generation performance of consistency models.\nTo address this issue, we generalize consistency training to the truncated time range, which allows the model to ignore denoising tasks at earlier time steps and focus its capacity on generation.\nWe propose a new parameterization of the consistency function and a two-stage training procedure that prevent the truncated-time training from collapsing to a trivial solution.\nExperiments on CIFAR-10 and ImageNet $64\\times64$ datasets show that our method achieves better one-step and two-step FIDs than the state-of-the-art consistency models such as iCT-deep,\nusing more than 2$\\times$ smaller networks.", "keywords": [ "Consistency models", "diffusion models", "generative models" ], "rating_list": [ 8, 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3, 5 ] }, { "paper_id": "ZXs3pkmrRG", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Dongmei Zhang", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Distinguished Scientist" }, { "name": "Bojun Huang", "gender": "Male", "institution": "Rakuten Institute of Technology", "country": "", "position": "Principal Researcher" }, { "name": "Qiang Fu", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Rui Ding", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Shi Han", "gender": "Male", "institution": "Microsoft Research Asia", "country": "", "position": "Researcher" }, { "name": "Wei Chen", "gender": "Male", "institution": "Hong Kong University of Science and Technology, Guangzhou", "country": "HK", "position": "PhD student" }, { "name": "Yang Zhang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yuxuan Liang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 46, "n_formula_1": 10, "n_ref_uni": 65, "n_ref": 133, "n_ref_all": 174, "n_fig": 12, "n_tab": 21, "L_tab": 8566, "n_element_tab": 908, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1791, "n_element_tab_1": 148, "formula_len_all": 5256, "formula_len_all_1": 824, "len_all": 303497, "len_all_1": 65732, "len_abs": 2310, "len_title": 111, "len_sents": 105432, "len_sents_1": 29885, "n_sents": 768, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1425, "title": "Test-Time Learning of Causal Structure from Interventional Data", "abs": "Inferring causal structures from interventional data remains a challenging task, especially when the intervention targets are unknown. *Supervised Causal Learning (SCL)* demonstrates strong empirical performance in predicting causal structures by training on datasets with known causal relations and applying the learned models to unseen test data. However, existing *SCL* methods often face inherent generalization challenges and struggle with the diverse intervention settings encountered in the *interventional causal discovery* problem. \n\nIn this work, we propose _**TICL**_ (**T**est-time **I**nterventional **C**ausal **L**earning), a novel approach that follows the *Test-Time Training (TTT)* + *Joint Causal Inference (JCI)* paradigm to address these challenges of generalization and versatility, respectively. Specifically, _**TICL**_ employs a self-augmentation technique that generates training data at test time, tailored to the characteristics of the test data, enabling the model to adapt to the inherent biases in the test distribution. Additionally, by integrating the *JCI* framework with *SCL*, _**TICL**_ replaces the rule-based logic of the standard PC algorithm with a learning-based approach, effectively leveraging self-augmented training data. \n\nExtensive experiments on bnlearn benchmarks demonstrate _**TICL**_'s superiority in multiple aspects of causal discovery and intervention target detection.", "keywords": [ "Test-Time Traing", "Self-Augmentation", "Causal Structure Learning", "Intervention Target Detection" ], "rating_list": [ 5, 6, 3, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 1, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "ZXaocmXc6d", "primary_area": "learning theory", "n_author": 7, "author_info_list": [ { "name": "Alexandra Maria Proca", "gender": "Female", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Andrew M Saxe", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "Associate Professor" }, { "name": "Cl\u00e9mentine Carla Juliette Domin\u00e9", "gender": "Female", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Daniel Kunin", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Lukas Braun", "gender": "Male", "institution": "Allen Institute for Neural Dynamics", "country": "US", "position": "Researcher" }, { "name": "Nicolas Anguita", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Pedro A. M. Mediano", "gender": "Not Specified", "institution": "Imperial College London", "country": "GB", "position": "Lecturer" } ], "n_formula": 185, "n_formula_1": 27, "n_ref_uni": 64, "n_ref": 170, "n_ref_all": 217, "n_fig": 23, "n_tab": 3, "L_tab": 915, "n_element_tab": 32, "n_fig_1": 15, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 26571, "formula_len_all_1": 2498, "len_all": 308494, "len_all_1": 68716, "len_abs": 1183, "len_title": 114, "len_sents": 90972, "len_sents_1": 31329, "n_sents": 761, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1183, "title": "From Lazy to Rich: Exact Learning Dynamics in Deep Linear Networks", "abs": "Biological and artificial neural networks develop internal representations that enable them to perform complex tasks. In artificial networks, the effectiveness of these models relies on their ability to build task specific representation, a process influenced by interactions among datasets, architectures, initialization strategies, and optimization algorithms. Prior studies highlight that different initializations can place networks in either a lazy regime, where representations remain static, or a rich/feature learning regime, where representations evolve dynamically. Here, we examine how initialization influences learning dynamics in deep linear neural networks, deriving exact solutions for lambda-balanced initializations-defined by the relative scale of weights across layers. These solutions capture the evolution of representations and the Neural Tangent Kernel across the spectrum from the rich to the lazy regimes. Our findings deepen the theoretical understanding of the impact of weight initialization on learning regimes, with implications for continual learning, reversal learning, and transfer learning, relevant to both neuroscience and practical applications.", "keywords": [ "Deep learning", "Learning theory", "Learning Regime", "Rich", "Lazy" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 4, 3, 4 ], "presentation_list": [ 3, 4, 4 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "ZXO7iURZfW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Diego Klabjan", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "Full Professor" }, { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" }, { "name": "Jean Utke", "gender": "unknown", "institution": "Allstate", "country": "US", "position": "Data Scientist / Technical Director" }, { "name": "Yufeng Zou", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 25, "n_ref": 70, "n_ref_all": 111, "n_fig": 12, "n_tab": 20, "L_tab": 19336, "n_element_tab": 1358, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 5752, "n_element_tab_1": 295, "formula_len_all": 151, "formula_len_all_1": 151, "len_all": 186643, "len_all_1": 72050, "len_abs": 943, "len_title": 90, "len_sents": 50797, "len_sents_1": 29354, "n_sents": 397, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 946, "title": "Automated Feature Engineering by Prompting", "abs": "Automated feature engineering (AutoFE) liberates data scientists from the burden\nof manual feature construction, a critical step for tabular data prediction. While the\nsemantic information of datasets provides valuable context for feature engineering,\nit has been underutilized in most existing works. In this paper, we introduce\nAutoFE by Prompting (FEBP), a novel AutoFE algorithm that leverages large language\nmodels (LLMs) to process dataset descriptions and automatically generate\nfeatures. Incorporating domain knowledge, the LLM iteratively refines feature\nconstruction through in-context learning of top-performing example features and\nprovides semantic explanations. Our experiments on real-world datasets demonstrate\nthe superior performance of FEBP over state-of-the-art AuoFE methods. We\nalso conduct ablation study to verify the impact of dataset semantic information\nand examine the behavior of our LLM-based feature search process.", "keywords": [ "Automated Feature Engineering", "Large Language Models", "Tabular Data Prediction" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "ZXFJeR9Xm6", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 6, "author_info_list": [ { "name": "Domenico Mandaglio", "gender": "Male", "institution": "University of Calabria", "country": "", "position": "Assistant Professor" }, { "name": "Francesco Parisi", "gender": "unknown", "institution": "University of Calabria", "country": "IT", "position": "Associate Professor" }, { "name": "Gianvincenzo Alfano", "gender": "Male", "institution": "University of Calabria", "country": "", "position": "Assistant Professor" }, { "name": "Irina Trubitsyna", "gender": "Female", "institution": "University of Calabria", "country": "IT", "position": "Associate Professor" }, { "name": "Reza Shahbazian", "gender": "Male", "institution": "University of Calabria", "country": "IT", "position": "Assistant Professor" }, { "name": "Sergio Greco", "gender": "Male", "institution": "University of Calabria", "country": "IT", "position": "Full Professor" } ], "n_formula": 46, "n_formula_1": 32, "n_ref_uni": 48, "n_ref": 74, "n_ref_all": 99, "n_fig": 4, "n_tab": 7, "L_tab": 2672, "n_element_tab": 389, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1082, "n_element_tab_1": 144, "formula_len_all": 4065, "formula_len_all_1": 3032, "len_all": 166555, "len_all_1": 65912, "len_abs": 1091, "len_title": 97, "len_sents": 49590, "len_sents_1": 27657, "n_sents": 363, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1096, "title": "Physics-Informed Decentralized Federated Learning", "abs": "The integration of domain knowledge into the learning process of artificial intelligence (AI) has received significant attention in the last few years. Most of the approaches proposed so far have focused on centralized machine learning scenarios, with less emphasis on how domain knowledge can be effectively integrated in decentralized settings. In this paper, we address this gap by evaluating the effectiveness of domain knowledge integration in distributed settings, specifically in the context of Decentralized Federated Learning (DFL). We propose the Physics-Informed DFL (PIDFL) architecture by integrating domain knowledge expressed as differential equations. We introduce a serverless data aggregation algorithm for PIDFL, prove its convergence, and discuss its computational complexity. We performed comprehensive experiments across various datasets and demonstrated that PIDFL significantly reduces average loss across diverse applications. This highlights the potential of PIDFL and offers a promising avenue for improving decentralized learning through domain knowledge integration.", "keywords": [ "Federated Learning", "Physics-Informed Neural Network", "Domain Knowledge", "Decentralized" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "ZWthVveg7X", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Chi-Hyuck Jun", "gender": "unknown", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Taek-Ho Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Postdoc" }, { "name": "SUNGWOO HUR", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 31, "n_ref": 53, "n_ref_all": 77, "n_fig": 9, "n_tab": 9, "L_tab": 19480, "n_element_tab": 1982, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 1701, "n_element_tab_1": 20, "formula_len_all": 476, "formula_len_all_1": 225, "len_all": 160066, "len_all_1": 53104, "len_abs": 1392, "len_title": 145, "len_sents": 35351, "len_sents_1": 22968, "n_sents": 289, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1399, "title": "Enhanced multi-task learning of imputation and prediction via feature relationship graph learning", "abs": "Missing values present significant challenges in machine learning, often degrading predictive performance. Traditional and deep learning imputation methods often overlook the relationships between features and their connections to downstream tasks. To address these gaps, we propose PIG (multi-task learning of Prediction and Imputation via feature-relationship Graph learning), a model that integrates imputation and prediction by leveraging feature interdependencies. \nPIG utilizes a graph-based approach to capture intricate feature relationships, thereby enhancing the accuracy of both imputation and downstream tasks. Our strategic training process begins with pre-training for both tasks, ensuring the model learns effective representations. This is followed by fine-tuning the entire model to further optimize imputation and downstream tasks simultaneously.\nWe evaluated our method using nine benchmark datasets, three for regression and six for classification.\nOur method showed superior imputation and prediction performance across nine datasets, achieving an average rank of 1.33 for both imputation and regression tasks and 1.83 for imputation and 1.17 for classification tasks. Additionally, in sensitivity analysis with respect to missing rates, our method demonstrated its robustness, especially in predictive performance, compared to other methods that showed significant degradation.", "keywords": [ "Missing values", "imputation", "feature selection", "graph nueral network", "multi-task learning" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "ZWi6RpT4mJ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Dhananjaya Jayasundara", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Don Yasiru Lakshan Ranasinghe", "gender": "Male", "institution": "Whiting School of Engineering", "country": "US", "position": "PhD student" }, { "name": "Sudarshan Rajagopalan", "gender": "Male", "institution": "Whiting School of Engineering", "country": "US", "position": "PhD student" }, { "name": "Trac Tran", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Vishal M Patel", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 52, "n_ref_all": 62, "n_fig": 6, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 40, "formula_len_all_1": 40, "len_all": 109612, "len_all_1": 69838, "len_abs": 1409, "len_title": 98, "len_sents": 37729, "len_sents_1": 34746, "n_sents": 280, "n_sents_1": 257, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1415, "title": "CoINR: Compressed Implicit Neural Representations", "abs": "Implicit Neural Representations (INRs) are increasingly recognized as a versatile data modality for representing discretized signals, offering benefits such as infinite query resolution and reduced storage requirements. Existing signal compression approaches for INRs typically employ one of two strategies: 1. direct quantization with entropy coding of the trained INR; 2. deriving a latent code on top of the INR through a learnable transformation. Thus, their performance is heavily dependent on the quantization and entropy coding schemes employed. In this paper, we introduce CoINR, an innovative compression algorithm that leverages the patterns in the vector spaces formed by weights of INRs. We compress these vector spaces using a high-dimensional sparse code within a dictionary. Further analysis reveals that the atoms of the dictionary used to generate the sparse code do not need to be learned or transmitted to successfully recover the INR weights. We demonstrate that the proposed approach can be integrated with any existing INR-based signal compression technique. Our results indicate that CoINR achieves substantial reductions in storage requirements for INRs across various configurations, outperforming conventional INR-based compression baselines. Furthermore, CoINR maintains high-quality decoding across diverse data modalities, including images, occupancy fields, and Neural Radiance Fields.", "keywords": [ "Implicit Neural Representations", "Signal Compression", "Model Compression", "Compressed Sensing" ], "rating_list": [ 5, 5, 3, 1 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "ZW4MRZrmSA", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Asuman E. Ozdaglar", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Gabriele Farina", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Mingyang Liu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 190, "n_formula_1": 43, "n_ref_uni": 28, "n_ref": 101, "n_ref_all": 112, "n_fig": 3, "n_tab": 2, "L_tab": 302, "n_element_tab": 29, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 32920, "formula_len_all_1": 5352, "len_all": 268975, "len_all_1": 81645, "len_abs": 970, "len_title": 143, "len_sents": 59065, "len_sents_1": 27063, "n_sents": 640, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 981, "title": "A Policy-Gradient Approach to Solving Imperfect-Information Games with Best-Iterate Convergence", "abs": "Policy gradient methods have become a staple of any single-agent reinforcement learning toolbox, due to their combination of desirable properties: iterate convergence, efficient use of stochastic trajectory feedback, and theoretically-sound avoidance of importance sampling corrections. In multi-agent imperfect-information settings (extensive-form games), however, it is still unknown whether the same desiderata can be guaranteed while retaining theoretical guarantees. Instead, sound methods for extensive-form games rely on approximating \\emph{counterfactual} values (as opposed to Q values), which are incompatible with policy gradient methodologies. In this paper, we investigate whether policy gradient can be safely used in two-player zero-sum imperfect-information extensive-form games (EFGs). We establish positive results, showing for the first time that a policy gradient method leads to provable best-iterate convergence to a regularized Nash equilibrium in self-play.", "keywords": [ "Game Theory", "Reinforcement Learning" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "ZVe2k7mNAP", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Jianguo Zhang", "gender": "Male", "institution": "Southern University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Peijia Qin", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 25, "n_formula_1": 20, "n_ref_uni": 45, "n_ref": 101, "n_ref_all": 120, "n_fig": 6, "n_tab": 9, "L_tab": 2759, "n_element_tab": 437, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1073, "n_element_tab_1": 197, "formula_len_all": 2178, "formula_len_all_1": 1539, "len_all": 151549, "len_all_1": 67042, "len_abs": 1157, "len_title": 108, "len_sents": 44220, "len_sents_1": 28494, "n_sents": 330, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1164, "title": "MQ-VAE: Training Vector-Quantized Networks via Meta Learning", "abs": "Deep neural networks with discrete latent variables are particularly well-suited for tasks that naturally involve sequences of discrete symbols.\nThe vector-quantized variational auto-encoder (VQ-VAE) has made significant progress in this area by leveraging vector quantization.\nHowever, while much effort has been put into maximizing codebook utilization, this does not always result in better performance.\nAdditional challenges include quantization errors in the VQ layer and the lack of direct integration of task loss into the codebook objective.\nTo address these issues, we propose Meta-Quantized Variational Auto-Encoder (MQ-VAE), a bi-level optimization-based vector quantization framework inspired by meta-learning.\nIn MQ-VAE, the codebook and encoder-decoder pair are optimized at different levels, with the codebook treated as hyperparameters optimized via hyper-gradient descent.\nThis approach effectively tackles these challenges within a unified framework.\nThe evaluation of MQ-VAE on two computer vision tasks demonstrates its superiority over existing methods and ablation baselines.\nCode is available at https://anonymous.4open.science/r/MQVAE-B52C.", "keywords": [ "Vector-Quantized Networks", "Bi-level Optimization", "Image Generation" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "ZVdThSQTuC", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Christina Lioma", "gender": "Female", "institution": "University of Copenhagen", "country": "", "position": "Professor" }, { "name": "Maria Maistro", "gender": "Female", "institution": "University of Copenhagen", "country": "", "position": "Assistant Professor" }, { "name": "Michiel Marten Spape", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Associate Professor" }, { "name": "Tuukka Ruotsalo", "gender": "Male", "institution": "Lappeenranta University of Technology", "country": "FI", "position": "Associate Professor" }, { "name": "Vadym Gryshchuk", "gender": "Male", "institution": "University of Copenhagen", "country": "DK", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 54, "n_ref": 74, "n_ref_all": 99, "n_fig": 15, "n_tab": 4, "L_tab": 2359, "n_element_tab": 943, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 804, "n_element_tab_1": 51, "formula_len_all": 16, "formula_len_all_1": 16, "len_all": 193743, "len_all_1": 69271, "len_abs": 1200, "len_title": 120, "len_sents": 53726, "len_sents_1": 34654, "n_sents": 423, "n_sents_1": 246, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1234, "title": "An EEG dataset of word-level brain responses for semantic text relevance", "abs": "Electroencephalography (EEG) can enable non-invasive, real-time measurement of brain activity in response to human language processing. Previously released EEG datasets focus on brain signals measured either during completely natural reading or in full psycholinguistic experimental settings. Since reading is commonly performed when considering certain content as more semantically relevant than other, we release a novel dataset for semantic text relevance containing $23{,}270$ time-locked (${\\sim}0.7s$) word-level EEG recordings acquired from participants who read both text that was semantically relevant and irrelevant to self-selected topics. Using these data, we present benchmark experiments with two evaluation protocols: participant-independent and participant-dependent on two prediction tasks (word relevance and sentence relevance). We report the performance of five well known models on these tasks. Our dataset and code are openly released. Altogether, our dataset paves the way for advancing research on language relevance and psycholinguistics, brain input and feedback-based recommendation and retrieval systems, and development of brain-computer interface (BCI) devices for online detection of language relevance.", "keywords": [ "brain", "semantic relevance", "word relevance", "sentence relevance", "EEG", "ERP", "text", "human language processing" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "ZVOGMy8Sd8", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Jiaying Wang", "gender": "Male", "institution": "Shenyang University of Technology, Shenyang University of Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Jing Shan", "gender": "Female", "institution": "Shenyang University of Technology, Shenyang University of Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Shuailing Hao", "gender": "unknown", "institution": "Shenyang University of Technology, Shenyang University of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 26, "n_ref": 37, "n_ref_all": 54, "n_fig": 2, "n_tab": 2, "L_tab": 495, "n_element_tab": 88, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 161, "n_element_tab_1": 37, "formula_len_all": 472, "formula_len_all_1": 473, "len_all": 96408, "len_all_1": 64827, "len_abs": 1872, "len_title": 104, "len_sents": 30741, "len_sents_1": 30099, "n_sents": 234, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1380, "title": "Knowledge Enhanced Image Captioning for Fashion Products", "abs": "The field of image captioning has witnessed a surge in attention, particularly in the context of e-commerce, where the exploration of automated fashion description has gained significant momentum. This growing interest can be attributed to the increasing influence of visual language and its impact on effective communication within the fashion industry.\nHowever, generating detailed and accurate natural language descriptions for fashion items remains a topic of intense discussion. This paper introduces an innovative approach that specifically addresses this challenge by proposing a method tailored to the requirements of the e-commerce domain. Our approach integrates a knowledge base into the widely adopted end-to-end architecture, thereby enhancing the availability of comprehensive data about fashion items.\nWe design a mode mapping network that facilitates the fusion of attribute features extracted from the knowledge base with image features. \nAdditionally, we introduce a filter strategy to enhance the quality of the generated descriptions by selecting the best result among the candidate sentences generated through beam search using a language model.\nThrough extensive experimentation and evaluation, our proposed method demonstrates superior performance in the task of fashion description, surpassing the performance of state-of-the-art approaches in this domain.", "keywords": [ "Image captioning", "Knowledge base", "Visual language", "Fashion description" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "ZVLyyG5yS6", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Dongming Huang", "gender": "unknown", "institution": "National University of Singapore", "country": "", "position": "Assistant Professor" }, { "name": "Hehu Xie", "gender": "Male", "institution": "Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Jianfa Lai", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Jie Luo", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Associate Professor" }, { "name": "Qian Lin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Weihao Lu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 52, "n_ref_all": 79, "n_fig": 6, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 1520, "n_element_tab_1": 18, "formula_len_all": 73, "formula_len_all_1": 73, "len_all": 93342, "len_all_1": 59200, "len_abs": 881, "len_title": 130, "len_sents": 31715, "len_sents_1": 29935, "n_sents": 180, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 885, "title": "Hierarchical Analysis: Monotonicity of Layerwise performance in Large Language Models", "abs": "We introduce a quantitative framework to evaluate how Large Language Models (LLMs) learn tasks across all layers, revealing a `monotonicity phenomenon'. Specifically: \ni) performance at each layer consistently improves from one layer to the next on the pre-training set, and \nii) this improvement is consistently observed across various downstream tasks. This monotonicity phenomenon indicates that LLMs effectively capture complex hierarchical features across diverse datasets. For example, our study on the abstraction of concepts using linear representations in word embeddings shows that the clarity of these abstractions progressively increases with each layer.\nFinally, by leveraging this monotonicity, we can significantly reduce inference time and memory requirements by selecting the most appropriate layer, thereby enhancing the efficiency of LLMs in real-world applications.", "keywords": [ "Monotonicity", "Layerwise Performance", "Large Language Models (LLMs)", "Hierarchical Analysis" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "ZV7CLf0RHK", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Shuyang Jiang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Ya Zhang", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Professor" }, { "name": "Yanfeng Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Yu Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yusheng Liao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 23, "n_ref_uni": 44, "n_ref": 93, "n_ref_all": 118, "n_fig": 10, "n_tab": 5, "L_tab": 2687, "n_element_tab": 376, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 810, "n_element_tab_1": 71, "formula_len_all": 986, "formula_len_all_1": 1001, "len_all": 162623, "len_all_1": 60726, "len_abs": 1465, "len_title": 102, "len_sents": 41985, "len_sents_1": 27941, "n_sents": 286, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 54, "L_abs": 1533, "title": "Fine-tuning with Reserved Majority for Noise Reduction", "abs": "Parameter-efficient fine-tuning (PEFT) has revolutionized supervised fine-tuning, where LoRA and its variants gain the most popularity due to their low training costs and zero inference latency.\nHowever, LoRA tuning not only injects knowledgeable features but also noisy hallucination during fine-tuning, which hinders the utilization of tunable parameters with the increasing LoRA rank.\nIn this work, we first investigate in-depth the redundancies among LoRA parameters with substantial empirical studies.\nAiming to resemble the learning capacity of high ranks from the findings, we set up a new fine-tuning framework, \\textbf{P}arameter-\\textbf{Re}dundant \\textbf{F}ine-\\textbf{T}uning (\\preft), which follows the vanilla LoRA tuning process but is required to reduce redundancies before merging LoRA parameters back to pre-trained models.\nBased on this framework, we propose \\textbf{No}ise reduction with \\textbf{R}eserved \\textbf{M}ajority (\\norm), which decomposes the LoRA parameters into majority parts and redundant parts with random singular value decomposition.\nThe major components are determined by the proposed \\search method, specifically employing subspace similarity to confirm the parameter groups that share the highest similarity with the base weight.\nBy employing \\norm, we enhance both the learning capacity and benefits from larger ranks, which consistently outperforms both LoRA and other \\preft-based methods on various downstream tasks, such as general instruction tuning, math reasoning and code generation.", "keywords": [ "large language models", "parameter redundancy fine-tuning", "noisy reduction" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "ZU8OdDLTts", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Haotong Qin", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Jiang Tian", "gender": "unknown", "institution": "Lenovo Research", "country": "", "position": "Principal Researcher" }, { "name": "Linghe Kong", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Tianao Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Xianglong Yan", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Xiaokang Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University, China", "country": "CN", "position": "Professor" }, { "name": "Yulun Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhiteng Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "zhongchao shi", "gender": "Male", "institution": "Lenovo Research", "country": "CN", "position": "Research Scientist" }, { "name": "DONG XIE", "gender": "Male", "institution": "Lenovo Research", "country": "", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 18, "n_ref_uni": 33, "n_ref": 58, "n_ref_all": 83, "n_fig": 5, "n_tab": 7, "L_tab": 3388, "n_element_tab": 557, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2001, "n_element_tab_1": 396, "formula_len_all": 2326, "formula_len_all_1": 2268, "len_all": 122837, "len_all_1": 61313, "len_abs": 2175, "len_title": 116, "len_sents": 31127, "len_sents_1": 22851, "n_sents": 255, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1521, "title": "ARB-LLM: Alternating Refined Binarizations for Large Language Models", "abs": "Large Language Models (LLMs) have greatly pushed forward advancements in natural language processing, yet their high memory and computational demands hinder practical deployment. Binarization, as an effective compression technique, can shrink model weights to just 1 bit, significantly reducing the high demands on computation and memory. However, current binarization methods struggle to narrow the distribution gap between binarized and full-precision weights, while also overlooking the column deviation in LLM weight distribution. To tackle these issues, we propose ARB-LLM, a novel 1-bit post-training quantization (PTQ) technique tailored for LLMs. To narrow the distribution shift between binarized and full-precision weights, we first design an alternating refined binarization (ARB) algorithm to progressively update the binarization parameters, which significantly reduces the quantization error. Moreover, considering the pivot role of calibration data and the column deviation in LLM weights, we further extend ARB to ARB-X and ARB-RC. In addition, we refine the weight partition strategy with column-group bitmap (CGB), which further enhance performance. Equipping ARB-X and ARB-RC with CGB, we obtain ARB-LLM$_{\\text{X}}$ and ARB-LLM$ _{\\text{RC}} $ respectively, which significantly outperform state-of-the-art (SOTA) binarization methods for LLMs.\nAs a binary PTQ method, our ARB-LLM$ _{\\text{RC}} $ is the first to surpass FP16 models of the same size. We will release all the code and models of ARB-LLM.", "keywords": [ "Binarization", "LLM" ], "rating_list": [ 8, 5, 8 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 3, 4, 2 ] }, { "paper_id": "ZU42Wrcqfm", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 11, "author_info_list": [ { "name": "Can Zhang", "gender": "Female", "institution": "United Automotive Electronic Systems", "country": "", "position": "Researcher" }, { "name": "Chenglin Li", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Hongkai Xiong", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Junni Zou", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Weijia Lu", "gender": "Male", "institution": "UAES", "country": "CN", "position": "Principal Researcher" }, { "name": "Wenrui Dai", "gender": "unknown", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaofeng Ma", "gender": "unknown", "institution": "UAES", "country": "", "position": "Researcher" }, { "name": "Xinyi Lu", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "ZHIFEI YANG", "gender": "unknown", "institution": "United Automotive Electronic Systems Co., Ltd.", "country": "", "position": "Researcher" }, { "name": "xiaodong Zhang", "gender": "Male", "institution": "United Automotive Electronic Systems Co., Ltd.", "country": "", "position": "Researcher" }, { "name": "Hao Zhangg", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 8, "n_ref_uni": 35, "n_ref": 112, "n_ref_all": 160, "n_fig": 6, "n_tab": 19, "L_tab": 6069, "n_element_tab": 593, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2857, "n_element_tab_1": 159, "formula_len_all": 7333, "formula_len_all_1": 1157, "len_all": 207952, "len_all_1": 66019, "len_abs": 1583, "len_title": 157, "len_sents": 72238, "len_sents_1": 29579, "n_sents": 559, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1587, "title": "FedSMU: Communication-Efficient and Generalization-Enhanced Federated Learning through Symbolic Model Updates", "abs": "The significant communication overhead and client data heterogeneity have posed important challenges to current federated learning (FL) paradigm. Most compression-based and optimization-based FL algorithms typically focus on addressing either the model compression challenge or the data heterogeneity issue individually, rather than tackling both of them. In this paper, we observe that by symbolizing the client model updates to be uploaded (i.e., normalizing the magnitude for each model parameter at local clients), the model heterogeneity can be mitigated that is essentially stemmed from data heterogeneity, thereby helping improve the overall generalization performance of the globally aggregated model at the server. Inspired with this observation, and further motivated by the success of Lion optimizer in achieving the optimal performance on most tasks in centralized learning, we propose a new FL algorithm, called FedSMU, which simultaneously reduces the communication overhead and alleviates the data heterogeneity issue. Specifically, FedSMU splits the standard Lion optimizer into the local updates and global execution, where only the symbol of client model updates commutes between the client and server. We theoretically prove the convergence of FedSMU for the general non-convex settings. Through extensive experimental evaluations on several benchmark datasets, we demonstrate that our FedSMU algorithm not only reduces the communication overhead, but also achieves a better generalization performance than the other compression-based and optimization-based baselines.", "keywords": [ "Federated learning", "Efficient Communication", "Enhanced Generalization" ], "rating_list": [ 5, 5, 8, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "ZTvUT49JjL", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Alp Yurtsever", "gender": "Male", "institution": "Ume\u00e5 University", "country": "SE", "position": "Assistant Professor" }, { "name": "Suvrit Sra", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "YIKUN HOU", "gender": "Male", "institution": "Umea University", "country": "SE", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 14, "n_ref_uni": 79, "n_ref": 105, "n_ref_all": 158, "n_fig": 22, "n_tab": 6, "L_tab": 1832, "n_element_tab": 187, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1227, "n_element_tab_1": 86, "formula_len_all": 1536, "formula_len_all_1": 1207, "len_all": 191838, "len_all_1": 62872, "len_abs": 1112, "len_title": 136, "len_sents": 45635, "len_sents_1": 27398, "n_sents": 358, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1130, "title": "Implicit Bias in Matrix Factorization and its Explicit Realization in a new Architecture", "abs": "Gradient descent for matrix factorization is known to exhibit an implicit bias toward approximately low-rank solutions. While existing theories often assume the boundedness of iterates, empirically the bias persists even with unbounded sequences. We thus hypothesize that implicit bias is driven by divergent dynamics markedly different from the convergent dynamics for data fitting. Using this perspective, we introduce a new factorization model: $X\\approx UDV^\\top$, where $U$ and $V$ are constrained within norm balls, while $D$ is a diagonal factor allowing the model to span the entire search space. Our experiments reveal that this model exhibits a strong implicit bias regardless of initialization and step size, yielding truly (rather than approximately) low-rank solutions. Furthermore, drawing parallels between matrix factorization and neural networks, we propose a novel neural network model featuring constrained layers and diagonal components. This model achieves strong performance across various regression and classification tasks while finding low-rank solutions, resulting in efficient and lightweight networks.", "keywords": [ "Implicit Regularization", "Matrix Factorization", "Neural Networks" ], "rating_list": [ 3, 5, 3, 3, 3 ], "soundness_list": [ 3, 3, 1, 2, 2 ], "presentation_list": [ 2, 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 1, 1 ], "confidence_list": [ 4, 3, 5, 4, 4 ] }, { "paper_id": "ZTpWOwMrzQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Frederick Tung", "gender": "Male", "institution": "Borealis AI", "country": "", "position": "Researcher" }, { "name": "Hossein Hajimirsadeghi", "gender": "Male", "institution": "Borealis AI", "country": "", "position": "Principal Researcher" }, { "name": "Mengyao Zhai", "gender": "unknown", "institution": "RBC Borealis", "country": "CA", "position": "Researcher" }, { "name": "Sepidehsadat Hosseini", "gender": "Female", "institution": "Borealis AI", "country": "CA", "position": "Researcher" }, { "name": "Yongchang Hao", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "PhD student" } ], "n_formula": 48, "n_formula_1": 14, "n_ref_uni": 51, "n_ref": 106, "n_ref_all": 129, "n_fig": 9, "n_tab": 1, "L_tab": 2489, "n_element_tab": 117, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5043, "formula_len_all_1": 849, "len_all": 172011, "len_all_1": 58329, "len_abs": 1028, "len_title": 101, "len_sents": 38979, "len_sents_1": 25126, "n_sents": 420, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1032, "title": "Radar: Fast Long-Context Decoding for Any Transformer", "abs": "Transformer models have demonstrated exceptional performance across a wide range of applications. Though forming the foundation of Transformer models, the dot-product attention does not scale well to long-context data since its time requirement grows quadratically with context length. In this work, we propose Radar, a training-free approach that accelerates inference by dynamically searching for the most important context tokens. For any pre-trained Transformer, Radar can reduce the decoding time complexity without training or heuristically evicting tokens. Moreover, we provide theoretical justification for our approach, demonstrating that Radar can reliably identify the most important tokens with high probability. We conduct extensive comparisons with the previous methods on a wide range of tasks. The results demonstrate that Radar achieves the state-of-the-art performance across different architectures with reduced time complexity, offering a practical solution for efficient long-context processing of Transformers.", "keywords": [ "Long-context decoding", "Large language models", "Inference acceleration", "Random features" ], "rating_list": [ 8, 3, 6, 8, 8 ], "soundness_list": [ 3, 2, 3, 4, 4 ], "presentation_list": [ 3, 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 3, 4 ], "confidence_list": [ 3, 4, 3, 3, 4 ] }, { "paper_id": "ZT33ACedmn", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Cheng Kang", "gender": "Male", "institution": "Czech Technical Univeresity in Prague, Czech Technical University of Prague", "country": "", "position": "PhD student" }, { "name": "Erin Claire Carson", "gender": "Female", "institution": "Charles University", "country": "CZ", "position": "Assistant Professor" }, { "name": "Xinye Chen", "gender": "Male", "institution": "LIP6", "country": "FR", "position": "Postdoc" } ], "n_formula": 25, "n_formula_1": 12, "n_ref_uni": 38, "n_ref": 68, "n_ref_all": 94, "n_fig": 6, "n_tab": 14, "L_tab": 10755, "n_element_tab": 637, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2421, "n_element_tab_1": 184, "formula_len_all": 1597, "formula_len_all_1": 743, "len_all": 177860, "len_all_1": 65862, "len_abs": 1564, "len_title": 137, "len_sents": 48510, "len_sents_1": 28018, "n_sents": 396, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1573, "title": "LLM-ABBA: Fine-Tuning Large Language Models For Time Series Using Symbolic Approximation", "abs": "The success of large language models (LLMs) for the time series domain has been demonstrated through various benchmarks. Utilizing symbolic time series representation, one can efficiently bridge the gap between LLMs and time series. However, the remaining challenge is to exploit the semantic information hidden in time series by using symbols or existing tokens of LLMs, while aligning the embedding space of LLMs according to the domain hidden information of time series. The symbolic time series approximation method called ABBA shows outstanding efficacy in preserving salient time series features by modeling time series patterns in terms of amplitude and period while using existing tokens of LLMs. \n\nIn this paper, we introduce a method, called LLM-ABBA, that integrates ABBA into large language models for various time series downstream tasks. By symbolizing time series, LLM-ABBA compares favorably to the recent state-of-the-art (SOTA) in UCR and three medical time series classification tasks. Meanwhile, a fixed-polygonal chain trick in ABBA is introduced to avoid large vibrations during prediction tasks by significantly mitigating the effects of cumulative error arisen from misused symbols during the transition from symbols to numerical values. In time series regression tasks, LLM-ABBA achieves the new SOTA on Time Series Extrinsic Regression (TSER) benchmarks. LLM-ABBA also shows competitive prediction capability compared to recent time series prediction SOTA results. We believe this framework can also seamlessly extend to other time series domains.", "keywords": [ "large language models", "time series", "symbolic approximation" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "ZSzmWtY31e", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Angelos Georgios Koulouras", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Dimitris Bertsimas", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Thodoris Koukouvinos", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 12, "n_ref_uni": 28, "n_ref": 37, "n_ref_all": 61, "n_fig": 3, "n_tab": 10, "L_tab": 5712, "n_element_tab": 863, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1121, "n_element_tab_1": 162, "formula_len_all": 1334, "formula_len_all_1": 647, "len_all": 144562, "len_all_1": 62817, "len_abs": 1945, "len_title": 70, "len_sents": 40557, "len_sents_1": 26992, "n_sents": 358, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 22, "L_abs": 1085, "title": "Optimized Oversampling", "abs": "Many classification problems that arise in practice feature imbalanced datasets, a regime in which a lot of machine learning (ML) models show diminished performance. To address class imbalance, techniques like undersampling and oversampling are used to improve the model's performance. In this paper, we introduce a new oversampling framework, Optimized Oversampling ($O^{2}$), which generates synthetic minority class points by maximizing the probability of belonging to the minority class, which is estimated by a trained classification model. We show theoretically, under mild assumptions, that the points generated by $O^{2}$ are more likely to belong to the minority class than those generated by other approaches. Further, we benchmark $O^{2}$ against state-of-the-art oversampling methods on 16 publicly available imbalanced datasets using Classification Trees (CART) and Logistic Regression (LR) for the downstream classification task. The numerical experiments show that $O^{2}$ has an edge over current state-of-the-art oversampling methods, which is more pronounced on CART.", "keywords": [ "Machine Learning", "Imbalanced Datasets", "Optimization" ], "rating_list": [ 5, 5, 1, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ZSdubdbOoi", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Arash Ahmadian", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "Undergrad student" }, { "name": "Eugene Choi", "gender": "Male", "institution": "Cohere", "country": "", "position": "Researcher" }, { "name": "Matthieu Geist", "gender": "Male", "institution": "Cohere", "country": "FR", "position": "Researcher" }, { "name": "Mohammad Gheshlaghi Azar", "gender": "unknown", "institution": "Cohere AI", "country": "US", "position": "Researcher" }, { "name": "Olivier Pietquin", "gender": "Male", "institution": "Earth Species Project", "country": "US", "position": "Chief Scientist" } ], "n_formula": 31, "n_formula_1": 28, "n_ref_uni": 44, "n_ref": 98, "n_ref_all": 151, "n_fig": 6, "n_tab": 1, "L_tab": 1479, "n_element_tab": 44, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 222, "n_element_tab_1": 21, "formula_len_all": 2677, "formula_len_all_1": 2107, "len_all": 187992, "len_all_1": 69355, "len_abs": 0, "len_title": 93, "len_sents": 51989, "len_sents_1": 30122, "n_sents": 445, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 45, "L_abs": 1938, "title": "Self-Improving Robust Preference Optimization", "abs": "Both online and offline RLHF methods such as PPO and DPO have been extremely successful in aligning AI with human preferences. Despite their success, the existing methods suffer from some fundamental limitations: prominent among those limitations are\n (a) models trained with RLHF can learn from mistakes or negative examples through RL mechanism or contrastive loss at the time of training. However at the time of inference they are not equipped with an innate mechanism to correct mistakes by self-improvement. \n (b) The optimal solution of existing methods is highly task-dependent and thus it is difficult for them to generalize to new tasks. Here we propose Self-Improving Robust Preference Optimization (SRPO), a practical and mathematically principled offline RLHF framework that address both these challenges. The key idea of SRPO is to cast the problem of learning from human preferences as a self-improvement process, which can be mathematically expressed in terms of a min-max objective that aims at joint optimization of self-improvement policy and the generative policy in an adversarial fashion. The solution for this optimization problem is independent of the training task and thus it is robust to its changes.\nWe then show that this objective can be re-expressed in the form of a non-adversarial offline loss which can be optimized using standard supervised optimization techniques at scale. \nWe show the effectiveness of SRPO in terms of AI Win-Rate (WR) against human (GOLD) completions. In particular, when SRPO is evaluated on the XSUM dataset, it outperforms the celebrated DPO by a clear margin of $\\mathbf{15}$% after $5$ self-revisions, achieving WR of $\\mathbf{90}$%. Also when evaluated on the challenging Arena-Hard prompts, SRPO outperforms both DPO and IPO (by $\\mathbf{4}$% without revision and $\\mathbf{6}$% after just one revision), achieving a WR of $\\mathbf{56}$% against Llama-3.1-8B-Instruct.", "keywords": [ "Preference optimization", "direct alignment", "reinforcement learning from human feedback", "Self-refinement" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "ZSbsX1sFo3", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Bin Bi", "gender": "unknown", "institution": "SalesForce.com", "country": "US", "position": "Principal Researcher" }, { "name": "Can Huang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Claire Na Cheng", "gender": "unknown", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "James Zhu", "gender": "Male", "institution": "University of Colorado at Boulder", "country": "US", "position": "Researcher" }, { "name": "Shiva Pentyala", "gender": "Male", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "Sitaram Asur", "gender": "unknown", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Zhichao Wang", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 16, "n_ref_uni": 20, "n_ref": 35, "n_ref_all": 80, "n_fig": 2, "n_tab": 8, "L_tab": 1281, "n_element_tab": 182, "n_fig_1": 2, "n_tab_1": 9, "L_tab_1": 1408, "n_element_tab_1": 184, "formula_len_all": 3043, "formula_len_all_1": 1487, "len_all": 181564, "len_all_1": 64079, "len_abs": 1334, "len_title": 139, "len_sents": 36077, "len_sents_1": 28341, "n_sents": 281, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1361, "title": "UNA: Unifying Alignments of RLHF/PPO, DPO and KTO by a Generalized Implicit Reward Function", "abs": "An LLM is pretrained on trillions of tokens, but the pretrained LLM may still generate undesired responses. To solve this problem, alignment techniques such as RLHF, DPO and KTO are proposed. However, these alignment techniques have limitations. For example, RLHF requires training the reward model and policy separately, which is complex, time-consuming, memory intensive and unstable during training processes. DPO proposes a mapping between an optimal policy and a reward, greatly simplifying the training process of RLHF. However, it can not take full advantages of a reward model and it is limited to pairwise preference data.\n\nIn this paper, we propose \\textbf{UN}ified \\textbf{A}lignment (UNA) which unifies RLHF/PPO, DPO and KTO. Firstly, we mathematically prove that given the classical RLHF objective, the optimal policy is induced by a generalize implicit reward function. With this novel mapping between a reward model and an optimal policy, UNA can 1. unify RLHF/PPO, DPO and KTO into a supervised learning of minimizing the difference between an implicit reward and an explicit reward; 2. outperform RLHF/PPO while simplify, stabilize, speed up and reduce memory burden of RL fine-tuning process; 3. accommodate different feedback types including pairwise, binary and scalar feedback. Downstream experiments show UNA outperforms DPO, KTO and RLHF.", "keywords": [ "LLM Alignment", "Unified Alignment", "RLHF", "PPO", "DPO", "KTO" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 2, 1, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ZS7UEI3vG5", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Christos Tzamos", "gender": "unknown", "institution": "University of Athens", "country": "GR", "position": "Associate Professor" }, { "name": "Konstantinos Karatapanis", "gender": "Male", "institution": "Archimedes/Athena RC, Greece", "country": "GR", "position": "Postdoc" }, { "name": "Vasilis Kontonis", "gender": "Male", "institution": ", University of Texas at Austin", "country": "", "position": "Postdoc" } ], "n_formula": 77, "n_formula_1": 23, "n_ref_uni": 24, "n_ref": 68, "n_ref_all": 77, "n_fig": 0, "n_tab": 1, "L_tab": 266, "n_element_tab": 13, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7251, "formula_len_all_1": 1734, "len_all": 158114, "len_all_1": 67894, "len_abs": 947, "len_title": 85, "len_sents": 50939, "len_sents_1": 30079, "n_sents": 436, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 37, "L_abs": 967, "title": "Oracle efficient truncated statistics", "abs": "We study the problem of learning from truncated samples: instead of observing\nsamples from some underlying population $p^\\ast$, we observe only the examples that fall in some survival set $S \\subset \\mathbb{R}^d$ whose probability mass (measured with respect to $p^\\ast$) is at least $\\alpha$. Assuming membership oracle access to the truncation set $S$, prior works obtained algorithms for the case where $p^\\ast$ is Gaussian or more generally an exponential family with strongly convex likelihood --- albeit with a super-polynomial \ndependency on the (inverse) survival mass $1/\\alpha$\nboth in terms of runtime and in number of oracle calls to the set $S$. In this work we design a new learning method with runtime and query complexity polynomial in $1/\\alpha$. \nOur result significantly improves over the prior works \nby focusing on efficiently solving the underlying optimization problem using a general\npurpose optimization algorithm with minimal assumptions.", "keywords": [ "truncated statistics", "exponential family", "statistical learning" ], "rating_list": [ 6, 8, 3, 8 ], "soundness_list": [ 3, 4, 1, 3 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "ZS1lCBLljq", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Jaeyoon Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sung-eui Yoon", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "Professor" }, { "name": "Yoonki Cho", "gender": "unknown", "institution": "KAIST", "country": "KR", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 10, "n_ref_uni": 59, "n_ref": 148, "n_ref_all": 178, "n_fig": 8, "n_tab": 10, "L_tab": 2831, "n_element_tab": 523, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 2385, "n_element_tab_1": 350, "formula_len_all": 1346, "formula_len_all_1": 613, "len_all": 168958, "len_all_1": 67990, "len_abs": 1409, "len_title": 118, "len_sents": 40562, "len_sents_1": 29487, "n_sents": 276, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1418, "title": "Class-Relational Label Smoothing for Lifelong Visual Place Recognition", "abs": "Visual Place Recognition (VPR) is a task of estimating the location of a query image, predominantly executed through image retrieval using learned global descriptors from a reference database of geo-tagged images. While recent approaches have aimed to improve the scalability of VPR training by leveraging classification loss as a proxy task, this leads to a task gap between classification and retrieval - classification discretizes the feature space into distinct class regions, often overlooking visual differences between classes. This gap makes VPR systems particularly vulnerable to extreme visual changes such as lifelong variations. To remedy these problems, we propose a novel Class-Relational Label Smoothing (CRLS) that transforms one-hot labels into soft labels by considering visual information of inter-class relations. We further enhance this method by dynamically adjusting the influence of CRLS based on the stability of class weights, which is quantified by their magnitudes. Importantly, our findings suggest that the magnitude of class weights serves as an indicator of class stability, which is also supported by derivative analysis. We demonstrate that our method outperforms state-of-the-art methods on the most extensive 17 benchmarks, effectively bridging the task gap between classification and retrieval in visual place recognition. Codes and trained weights will be made publicly available.", "keywords": [ "Lifelong place recognition", "visual geo-localization" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "ZRDhBwKs7l", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Byung Hyun Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Dong Un Kang", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Seunggyu Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Sungjin Lim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Se Young Chun", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" } ], "n_formula": 50, "n_formula_1": 22, "n_ref_uni": 61, "n_ref": 176, "n_ref_all": 218, "n_fig": 5, "n_tab": 15, "L_tab": 6364, "n_element_tab": 841, "n_fig_1": 4, "n_tab_1": 10, "L_tab_1": 2477, "n_element_tab_1": 236, "formula_len_all": 5137, "formula_len_all_1": 1871, "len_all": 235174, "len_all_1": 66924, "len_abs": 1414, "len_title": 86, "len_sents": 65610, "len_sents_1": 27291, "n_sents": 513, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1418, "title": "Concept Pinpoint Eraser for Text-to-image Diffusion Models via Residual Attention Gate", "abs": "Remarkable progress in text-to-image diffusion models has brought a major concern about potentially generating images on inappropriate or trademarked concepts. Concept erasing has been investigated with the goals of deleting target concepts in diffusion models while preserving other concepts with minimal distortion. To achieve these goals, recent concept erasing methods usually fine-tune the cross-attention layers of diffusion models. In this work, we first show that merely updating the cross-attention layers in diffusion models, which is mathematically equivalent to adding linear modules to weights, may not be able to preserve diverse remaining concepts. Then, we propose a novel framework, dubbed Concept Pinpoint Eraser (CPE), by adding nonlinear Residual Attention Gates (ResAGs) that selectively erase (or cut) target concepts while safeguarding remaining concepts from broad distributions by employing an attention anchoring loss to prevent the forgetting. Moreover, we adversarially train CPE with ResAG and learnable text embeddings in an iterative manner to maximize erasing performance and enhance robustness against adversarial attacks. Extensive experiments on the erasure of celebrities, artistic styles, and explicit contents demonstrated that the proposed CPE outperforms prior arts by keeping diverse remaining concepts while deleting the target concepts with robustness against attack prompts.", "keywords": [ "Concept erasing", "Generative model", "Diffusion model" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "ZRDa2IT1sQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Houxing Ren", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Junting Pan", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Ke Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Mingjie Zhan", "gender": "Male", "institution": "SenseTime Research", "country": "CN", "position": "Researcher" }, { "name": "Weikang Shi", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Zimu Lu", "gender": "Female", "institution": "University of Science and Technology of China", "country": "", "position": "Undergrad student" }, { "name": "Aojun Zhou", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 9, "n_ref": 16, "n_ref_all": 34, "n_fig": 8, "n_tab": 6, "L_tab": 1357, "n_element_tab": 229, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 2296, "n_element_tab_1": 166, "formula_len_all": 680, "formula_len_all_1": 849, "len_all": 185854, "len_all_1": 55965, "len_abs": 974, "len_title": 152, "len_sents": 32943, "len_sents_1": 21832, "n_sents": 255, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 979, "title": "Step-Controlled DPO: Leveraging Stepwise Errors for Enhancing Mathematical Reasoning of Language Models", "abs": "Direct Preference Optimization (DPO) has proven effective at improving the performance of large language models (LLMs) on downstream tasks such as reasoning and alignment. In this work, we propose Step-Controlled DPO (SCDPO), a method for automatically providing stepwise error supervision by creating negative samples of mathematical reasoning rationales that start making errors at a specified step. By applying these samples in DPO training, SCDPO can better align the model to avoid reasoning errors and output accurate reasoning steps. Qualitative analysis of the credit assignment of SCDPO and DPO demonstrates the effectiveness of SCDPO at identifying errors in mathematical solutions. We then apply SCDPO to an InternLM2-20B model, resulting in a 20B model that achieves competitive scores of 88.5\\% on GSM8K and 58.1\\% on MATH, rivaling all other open-source LLMs, showing the great potential of our method. The code, models and data are released to inspire future work.", "keywords": [ "large language model", "mathematical reasoning", "alignment with relative feedback" ], "rating_list": [ 6, 6, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "ZQzGrkihVq", "primary_area": "reinforcement learning", "n_author": 1, "author_info_list": [ { "name": "Ezgi Korkmaz", "gender": "unknown", "institution": "DeepMind", "country": "GB", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 19, "n_ref_uni": 31, "n_ref": 39, "n_ref_all": 56, "n_fig": 5, "n_tab": 2, "L_tab": 826, "n_element_tab": 59, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 826, "n_element_tab_1": 59, "formula_len_all": 2399, "formula_len_all_1": 2514, "len_all": 101734, "len_all_1": 65186, "len_abs": 1329, "len_title": 95, "len_sents": 31444, "len_sents_1": 30036, "n_sents": 200, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1345, "title": "HAL: Harmonic Learning in High-Dimensional MDPs", "abs": "Since the initial successes of deep reinforcement learning on learning policies purely by interacting with complex high-dimensional state representations and a decade of extensive research, deep neural policies have been applied to a striking variety of fields ranging from pharmaceuticals to foundation models. Yet, one of the strongest assumptions of reinforcement learning is to expect to receive a reward signal from the MDP. While this assumption comes in handy in certain fields, i.e. automated financial markets, it does not naturally fit in many others where the computational complexity of providing such a signal for the task at hand is larger than in fact learning one. Thus, in this paper we focus on learning policies in MDPs without this assumption, and study sequential decision making without having access to information on rewards provided by the MDP. We introduce We introduce harmonic learning, a training method in high-dimensional MDPs, and provide a theoretically well-founded algorithm that significantly improves the sample complexity of deep neural policies. The theoretical and empirical analysis reported in our paper demonstrates that harmonic learning achieves substantial improvements in sample efficient training while constructing more stable and resilient policies that can generalize to uncertain environments.", "keywords": [ "harmonic learning", "harmonic analytic basis training" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "ZQwvUTyL8Y", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "David Mitchell", "gender": "unknown", "institution": "New Mexico State University", "country": "", "position": "Associate Professor" }, { "name": "Grace Villacr\u00e9s", "gender": "Female", "institution": "Universidad Rey Juan Carlos", "country": "ES", "position": "Assistant Professor" }, { "name": "Mar\u00eda Mart\u00ednez-Garc\u00eda", "gender": "Female", "institution": "Saarland University, Universit\u00e4t des Saarlandes", "country": "DE", "position": "Postdoc" }, { "name": "Pablo M. Olmos", "gender": "unknown", "institution": "Universidad Carlos III de Madrid", "country": "ES", "position": "Associate Professor" } ], "n_formula": 28, "n_formula_1": 12, "n_ref_uni": 36, "n_ref": 77, "n_ref_all": 146, "n_fig": 38, "n_tab": 23, "L_tab": 6565, "n_element_tab": 824, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1661, "formula_len_all_1": 760, "len_all": 231997, "len_all_1": 60749, "len_abs": 2073, "len_title": 124, "len_sents": 82889, "len_sents_1": 30092, "n_sents": 664, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1499, "title": "Improved Variational Inference in Discrete VAEs using Error Correcting Codes", "abs": "Despite significant advancements in deep probabilistic models, effective learning of low-dimensional discrete latent representations remains challenging. This paper introduces a novel method to improve variational inference in discrete latent variable models by employing Error Correcting Codes (ECCs) to add redundancy to the latent representations, later exploited by the variational approximated posterior to provide more accurate estimates, thereby reducing the variational gap. Drawing inspiration from ECCs used in digital communications and data storage, we demonstrate proof-of-concept using a Discrete Variational Autoencoder (DVAE) with binary latent variables and block repetition codes. We then extend it to a hierarchical structure inspired by polar codes, in which some latent bits are more robustly protected than others. Our approach significantly enhances generation quality, data reconstruction, and uncertainty calibration compared to the uncoded DVAE, even when trained with tighter bounds such as the Importance Weighted Autoencoder (IWAE) objective. In particular, we demonstrate superior performance on MNIST, FMNIST, CIFAR10, and Tiny ImageNet datasets. The general approach of integrating ECCs into variational inference is compatible with existing techniques to boost variational inference, such as importance sampling or Hamiltonian Monte Carlo. We also formulate the properties that ECCs need to possess to be effectively used for improved discrete variational inference.", "keywords": [ "deep generative models", "variational inference", "discrete representations", "error correcting codes" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "ZQ9SF5eUHZ", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Cho-Jui Hsieh", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Jui-Nan Yen", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 19, "n_ref_uni": 13, "n_ref": 59, "n_ref_all": 68, "n_fig": 2, "n_tab": 6, "L_tab": 398, "n_element_tab": 87, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 186, "n_element_tab_1": 42, "formula_len_all": 1039, "formula_len_all_1": 1020, "len_all": 86871, "len_all_1": 56863, "len_abs": 1162, "len_title": 136, "len_sents": 31249, "len_sents_1": 26346, "n_sents": 258, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 823, "title": "Learn from the Past: Dynamic Data Pruning with Historically Weighted Bernoulli Sampling", "abs": "Dynamic data pruning, which also known as data importance sampling, has been proposed to improve training efficiency. For the case of sampling with replacement, the optimal sampling distribution to minimize the variance is to sample proportional to the gradient norm, which can be approximated by the gradient norm of the logits from an extra forward pass. However, this could result in repeated samples, which can be an undesirable property. Noticing that most dynamic data pruning methods that avoids repeated samples can be seen as weighted Bernoulli sampling, in this work we study the optimal distribution to reduce its variance. Furthermore, to avoid an extra forward pass, we study the use of historic statistics. We propose the use of exponential moving average and probability smoothing to improve the performance.", "keywords": [ "data selection", "dynamic data pruning", "importance sampling" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 2 ] }, { "paper_id": "ZPwX1FL4yp", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Nicu Sebe", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Full Professor" }, { "name": "Rui Wang", "gender": "Male", "institution": "Jiangnan University", "country": "CN", "position": "Lecturer" }, { "name": "Shaocheng Jin", "gender": "Male", "institution": "Jiangnan university", "country": "CN", "position": "MS student" }, { "name": "Xiaojun Wu", "gender": "Male", "institution": "Jiangnan University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoqing Luo", "gender": "Female", "institution": "Jiangnan University", "country": "CN", "position": "Full Professor" }, { "name": "Ziheng Chen", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 39, "n_ref_uni": 48, "n_ref": 125, "n_ref_all": 147, "n_fig": 2, "n_tab": 5, "L_tab": 1820, "n_element_tab": 217, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1820, "n_element_tab_1": 217, "formula_len_all": 2751, "formula_len_all_1": 2345, "len_all": 147762, "len_all_1": 71573, "len_abs": 1010, "len_title": 103, "len_sents": 34199, "len_sents_1": 26916, "n_sents": 373, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1013, "title": "Algebraic SPD and Correlation Geometry: A Gyro Approach", "abs": "The generalization of Deep Neural Networks (DNNs) to Riemannian manifolds has garnered significant attention across various scientific fields. Recent studies have demonstrated that several manifolds, including hyperbolic, spherical, Symmetric Positive Definite (SPD), and Grassmann manifolds, admit gyro-structures\u2014powerful algebraic structures that enable the principled extension of DNNs to manifolds. Inspired by these advancements, we introduce a novel gyro-structure for SPD manifolds, leveraging the flexible and powerful Power-Euclidean (PE) geometry. Moreover, full-rank correlation matrices, which are scale-invariant, serve as compact representations of SPD manifolds. Consequently, we propose two novel gyro-structures for correlation matrix manifolds, based on two theoretically and empirically convenient metrics: Euclidean-Cholesky (EC) and log-Euclidean-Cholesky (LEC) geometries. Extensive experiments on knowledge graph completion tasks validate the effectiveness of our proposed gyro-structures.", "keywords": [ "Correlation matrices", "SPD manifolds", "Gyrovector spaces" ], "rating_list": [ 3, 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 2, 2, 3, 3 ] }, { "paper_id": "ZPkNrs6aNO", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Bernie Wang", "gender": "Male", "institution": "AWS AI Labs", "country": "US", "position": "Principal Researcher" }, { "name": "Bingqing Song", "gender": "Female", "institution": "University of Minnesota, Minneapolis", "country": "", "position": "PhD student" }, { "name": "Bonan Min", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Boran Han", "gender": "Female", "institution": "Amazon/AWS", "country": "US", "position": "Researcher" }, { "name": "Haoyang Fang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Mingyi Hong", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Hao Wang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Assistant Professor" }, { "name": "SHUAI Zhang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" } ], "n_formula": 26, "n_formula_1": 19, "n_ref_uni": 34, "n_ref": 71, "n_ref_all": 89, "n_fig": 11, "n_tab": 2, "L_tab": 111, "n_element_tab": 2, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 209, "n_element_tab_1": 35, "formula_len_all": 1713, "formula_len_all_1": 1163, "len_all": 155574, "len_all_1": 73156, "len_abs": 1688, "len_title": 124, "len_sents": 46720, "len_sents_1": 33608, "n_sents": 365, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1702, "title": "Effectively Steer LLM To Follow Preference via Building Confident Directions", "abs": "Having an LLM that aligns with human preference is essential for accommodating individual needs, such as maintaining writing style or generating specific topics of interest.The majority of current alignment methods rely on fine-tuning or prompting, which can be either costly or difficult to control. Model steering algorithms, which construct certain steering directions used to modify the model output}, are typically easy to implement and optimization-free. {However, their capabilities are typically limited to steering the model into one of the two directions (i.e., bidreictional steering), and that there has been no theoretical understanding to guarantee their performance. In this work, we propose a theoretical framework to understand and quantify the model steering methods. Inspired by the framework, we propose a confident direction steering method (CONFST) that steers LLMs via modifying their activations in inference time. More specifically, CONFST builds a {\\it confident direction} that is closely aligned with users' preferences, and then this direction is added to the activations of the LLMs to effectively steer the model output. Our approach offers three key advantages over popular bidirectional model steering methods: 1) {It is more powerful, since multiple (i.e. more than two) users' preferences can be aligned simultaneously; 2) It is very simple to implement, since there is no need to determine which layer the steering vector should be added to; 3) No explicit user instruction is required. We validate our method on GPT-2 XL (1.5B), Mistral (7B) and Gemma-it (9B) models for tasks that require shifting the output of LLMs across a number of different topics and styles.", "keywords": [ "Language model", "Model steer", "Explainable LLM" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "ZPZ4eCQU9k", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Devendra Singh Dhami", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "Felix Divo", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Kristian Kersting", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Full Professor" }, { "name": "Maurice Kraus", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 46, "n_ref": 102, "n_ref_all": 118, "n_fig": 6, "n_tab": 10, "L_tab": 5971, "n_element_tab": 540, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 6253, "n_element_tab_1": 334, "formula_len_all": 772, "formula_len_all_1": 618, "len_all": 153517, "len_all_1": 63555, "len_abs": 1218, "len_title": 127, "len_sents": 38793, "len_sents_1": 25225, "n_sents": 288, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1038, "title": "xLSTM-Mixer: Multivariate Time Series Forecasting by Mixing via Scalar Memories", "abs": "Time series data is prevalent across numerous fields, necessitating the development of robust and accurate forecasting models. Capturing patterns both within and between temporal and multivariate components is crucial for reliable predictions.\nWe introduce xLSTM-Mixer, a model designed to effectively integrate temporal sequences, joint time-variate information, and multiple perspectives for robust forecasting. Our approach begins with a linear forecast shared across variates, which is then refined by xLSTM blocks. They serve as key elements for modeling the complex dynamics of challenging time series data. xLSTM-Mixer ultimately\nreconciles two distinct views to produce the final forecast. Our extensive evaluations demonstrate its superior long-term forecasting performance compared to recent state-of-the-art methods. A thorough model analysis provides further insights into its key components and confirms its robustness and effectiveness. This work contributes to the resurgence of recurrent models in time series forecasting.", "keywords": [ "time series", "xLSTM", "forecasting", "mixing", "recurrent" ], "rating_list": [ 3, 8, 5, 3, 6 ], "soundness_list": [ 2, 4, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2, 2 ], "contribution_list": [ 1, 4, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "ZPTHI3X9y8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Jiawei Chen", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "PhD student" }, { "name": "Jingyuan Zhang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Xiao Yang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Xinyi Zeng", "gender": "Female", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yu Tian", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yutao Zhu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Postdoc" }, { "name": "Yuying Shang", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zhengwei Fang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Zinan Liu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 25, "n_ref": 54, "n_ref_all": 70, "n_fig": 5, "n_tab": 6, "L_tab": 1745, "n_element_tab": 232, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1512, "n_element_tab_1": 152, "formula_len_all": 159, "formula_len_all_1": 159, "len_all": 118348, "len_all_1": 68246, "len_abs": 1257, "len_title": 136, "len_sents": 37539, "len_sents_1": 33051, "n_sents": 248, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1261, "title": "From Pixels to Tokens: Revisiting Object Hallucinations in Large Vision-Language Models", "abs": "Hallucinations in Large Vision-Language Models (LVLMs) are a significant challenge, i.e., generating objects that are not presented in the visual input, which impairs their reliability. Recent studies often attribute hallucinations to a lack of understanding of visual input, yet ignore a more fundamental issue: the model's inability to effectively extract or decouple visual features. In this paper, we revisit the hallucinations in LVLMs from an architectural perspective, investigating whether the primary cause lies in the visual encoder (feature extraction) or the modal alignment module (feature decoupling). Motivated by our findings on the preliminary investigation, we propose a novel tuning strategy, PATCH, to mitigate hallucinations in LVLMs. This plug-and-play method can be integrated into various LVLMs, utilizing adaptive virtual tokens to extract object features from bounding boxes, thereby addressing hallucinations caused by insufficient decoupling of visual features. PATCH achieves state-of-the-art performance on multiple multi-modal hallucination datasets. We hope this approach provides researchers with deeper insights into the underlying causes of hallucinations in LVLMs, fostering further advancements and innovation in this field.", "keywords": [ "large vision language model", "hallucination", "virtual token" ], "rating_list": [ 5, 8, 5, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "ZPCBcR7Drg", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Xue Maixuan", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xing Wei", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xinyuan Chang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zheng Pan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "XinranLiu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 18, "n_ref": 40, "n_ref_all": 87, "n_fig": 13, "n_tab": 7, "L_tab": 2861, "n_element_tab": 424, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 833, "n_element_tab_1": 122, "formula_len_all": 418, "formula_len_all_1": 344, "len_all": 166493, "len_all_1": 53614, "len_abs": 1323, "len_title": 147, "len_sents": 45278, "len_sents_1": 22729, "n_sents": 412, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1119, "title": "Driving by the Rules: A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map", "abs": "Ensuring adherence to traffic sign regulations is essential for both human and autonomous vehicle navigation. While current benchmark datasets concentrate on lane perception or basic traffic sign recognition, they often overlook the intricate task of integrating these regulations into lane operations. Addressing this gap, we introduce MapDR, a novel dataset designed for the extraction of Driving Rules from traffic signs and their association with vectorized, locally perceived HD Maps.\nMapDR features over 10,000 annotated video clips that capture the intricate correlation between traffic sign regulations and lanes. We define two pivotal sub-tasks: 1) Rule Extraction from Traffic Sign, which accurately deciphers regulatory instructions, and 2) Rule-Lane Correspondence Reasoning, which aligns these rules with their respective lanes.\nBuilt upon this benchmark, we provide a multimodal solution that offers a strong baseline for advancing autonomous driving technologies. It fills a critical gap in the integration of traffic sign rules, contributing to the development of reliable autonomous navigation systems.", "keywords": [ "Driving rules", "autonomous driving" ], "rating_list": [ 5, 6, 3, 6, 5 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 5, 3, 3 ] }, { "paper_id": "ZP8ZSJyP1U", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Fengyan Wang", "gender": "Female", "institution": "Institute ofArtificial intelligence, Hefei Comprehensive National Science Center", "country": "CN", "position": "Researcher" }, { "name": "Lixin He", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Shi Yin", "gender": "Male", "institution": "Institute of Artificial Intelligence, Hefei Comprehensive National Science Center", "country": "CN", "position": "Researcher" }, { "name": "Xinyang Pan", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 47, "n_formula_1": 18, "n_ref_uni": 25, "n_ref": 98, "n_ref_all": 134, "n_fig": 7, "n_tab": 11, "L_tab": 3600, "n_element_tab": 468, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1070, "n_element_tab_1": 155, "formula_len_all": 3480, "formula_len_all_1": 1148, "len_all": 215248, "len_all_1": 70321, "len_abs": 1978, "len_title": 135, "len_sents": 87194, "len_sents_1": 33817, "n_sents": 504, "n_sents_1": 158, "venue": "ICLR2025", "decision": "Reject", "L_title": 134, "L_abs": 1990, "title": "A Framework of SO(3)-equivariant Non-linear Representation Learning and its Application to Electronic-Structure Hamiltonian Prediction", "abs": "We propose both a theoretical and a methodological framework to address a critical challenge in applying deep learning to physical systems: the reconciliation of non-linear expressiveness with SO(3)-equivariance in predictions of SO(3)-equivariant quantities, such as the electronic-structure Hamiltonians. Inspired by covariant theory in physics, we present a solution by exploring the mathematical relationships between SO(3)-invariant and SO(3)-equivariant quantities and their representations. We first construct theoretical SO(3)-invariant quantities derived from the SO(3)-equivariant regression targets, and use these invariant quantities as supervisory labels to guide the learning of high-quality SO(3)-invariant features. Given that SO(3)-invariance is preserved under non-linear operations, the encoding process for invariant features can extensively utilize non-linear mappings, thereby fully capturing the non-linear patterns inherent in physical systems. Building on this, we propose a gradient-based mechanism to induce SO(3)-equivariant encodings of various degrees from the learned SO(3)-invariant features. This mechanism can incorporate non-linear expressive capabilities into SO(3)-equivariant representations, while theoretically preserving their equivariant properties as we prove, establishing a strong foundation for regressing complex SO(3)-equivariant targets. We apply our theory and method to the electronic-structure Hamiltonian prediction tasks, experimental results on eight benchmark databases covering multiple types of systems and challenging scenarios show substantial improvements on the state-of-the-art prediction accuracy of deep learning paradigm. Our method boosts Hamiltonian prediction accuracy by up to 40\\% and enhances downstream physical quantities, such as occupied orbital energy, by a maximum of 76\\%. Our method also significantly promotes the acceleration performance for the convergence of traditional Density Functional Theory methods.", "keywords": [ "SO(3)-equivariant representation learning; Non-linear expressiveness; Electronic-structure Hamiltonian prediction" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 5, 2 ] }, { "paper_id": "ZP1HqLus4y", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Andrej Janchevski", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Elias Abad Rocamora", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Ioannis Mavrothalassitis", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "MS student" }, { "name": "EFSTRATIOS PANTELEIMON SKOULAKIS", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "Postdoc" }, { "name": "Volkan Cevher", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" } ], "n_formula": 78, "n_formula_1": 23, "n_ref_uni": 36, "n_ref": 79, "n_ref_all": 97, "n_fig": 2, "n_tab": 6, "L_tab": 1747, "n_element_tab": 166, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 644, "n_element_tab_1": 64, "formula_len_all": 9859, "formula_len_all_1": 2391, "len_all": 160995, "len_all_1": 61853, "len_abs": 1073, "len_title": 120, "len_sents": 42744, "len_sents_1": 25973, "n_sents": 405, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1279, "title": "CONTINUAL FINITE-SUM MINIMIZATION UNDER THE POLYAK-\u0141OJASIEWICZ CONDITION", "abs": "Given functions $f_1,\\ldots,f_n$ where $f_i:\\mathcal{D}\\mapsto \\mathbb{R}$, \\textit{continual finite-sum minimization} (CFSM) asks for an $\\epsilon$-optimal sequence $\\hat{x}_1,\\ldots,\\hat{x}_n \\in \\mathcal{D}$ such that\n\n$$\\sum_{j=1}^i f_j(x_i)/i - \\min_{x \\in \\mathcal{D}}\\sum_{j=1}^if_j(x)/i \\leq \\epsilon$$\n\n In this work, we develop a new CFSM framework under the Polyak-\u0141ojasiewicz condition (PL), where each prefix-sum function $\\sum_{j=1}^i f_j(x)/i$ satisfies the PL condition, extending the recent result on CFSM with strongly convex functions. We present a new first-order method that under the PL condition producing an $\\epsilon$-optimal sequence with overall $\\mathcal{O}(n/\\sqrt{\\epsilon})$ first-order oracles (FOs), where an FO corresponds to the computation of a single gradient $\\nabla f_j(x)$ at a given $x \\in \\mathcal{D}$ for some $j \\in [n]$. Our method also improves upon the $\\mathcal{O}(n^2 \\log (1/\\epsilon))$ FO complexity of state-of-the art variance reduction methods as well as upon the $\\mathcal{O}(n/\\epsilon)$ FO complexity of $\\mathrm{StochasticGradientDescent}$. We experimentally evaluate our method in continual learning and the unlearning settings, demonstrating the potential of the CFSM framework in non-convex, deep learning problems.", "keywords": [ "Continual Learning", "Finite Sum Minimization" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "ZOrqCRDKh6", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Bimal Bhattarai", "gender": "Male", "institution": "University of Agder", "country": "", "position": "PhD student" }, { "name": "Lei Jiao", "gender": "Male", "institution": "University of Agder", "country": "NO", "position": "Full Professor" }, { "name": "Ole-Christoffer Granmo", "gender": "Male", "institution": "University of Oslo, Norway", "country": "", "position": "PhD student" }, { "name": "Rishad Shafik", "gender": "Male", "institution": "University of Newcastle-upon-Tyne", "country": "GB", "position": "Full Professor" }, { "name": "Ahmed Khalid Kadhim", "gender": "Male", "institution": "University of Agder", "country": "NO", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 7, "n_ref": 16, "n_ref_all": 24, "n_fig": 4, "n_tab": 2, "L_tab": 1024, "n_element_tab": 73, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 645, "n_element_tab_1": 69, "formula_len_all": 202, "formula_len_all_1": 202, "len_all": 81991, "len_all_1": 63998, "len_abs": 1482, "len_title": 123, "len_sents": 32542, "len_sents_1": 31933, "n_sents": 237, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1488, "title": "Scalable Multi-phase Word Embedding Using Conjunctive Propositional Clauses", "abs": "The Tsetlin Machine (TM) architecture has recently demonstrated effectiveness in Machine Learning (ML), particularly within Natural Language Processing (NLP). It has been utilized to construct word embedding using conjunctive propositional clauses, thereby significantly enhancing our understanding and interpretation of machine-derived decisions. The previous approach performed the word embedding over a sequence of input words to consolidate the information into a cohesive and unified representation. However, that approach encounters scalability challenges as the input size increases. In this study, we introduce a novel approach incorporating two-phase training to discover contextual embeddings of input sequences. Specifically, this method encapsulates the knowledge for each input word within the dataset\u2019s vocabulary, subsequently constructing embeddings for a sequence of input words utilizing the extracted knowledge. This technique not only facilitates the design of a scalable model but also preserves interpretability. Our experimental findings revealed that the proposed method yields competitive performance compared to the previous approaches, demonstrating promising results in contrast to human-generated benchmarks. Furthermore, we applied the proposed approach to sentiment analysis on the IMDB dataset, where the TM embedding and the TM classifier, along with other interpretable classifiers, offered a transparent end-to-end solution with competitive performance.", "keywords": [ "Tsetlin Machine", "Word Embeddings", "Natural Language Processing (NLP)" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 4, 2 ] }, { "paper_id": "ZNsWJkFrqQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Haolan Liu", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Jishen Zhao", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 30, "n_ref": 50, "n_ref_all": 86, "n_fig": 9, "n_tab": 2, "L_tab": 272, "n_element_tab": 39, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 131, "formula_len_all_1": 75, "len_all": 146479, "len_all_1": 53832, "len_abs": 3737, "len_title": 114, "len_sents": 43878, "len_sents_1": 26629, "n_sents": 365, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1141, "title": "GLIMO: Grounding Large Language Models With Imperfect World Models", "abs": "Despite a widespread success in various applications, large language models (LLMs) often stumble when tackling basic physical reasoning or executing robotics tasks, due to a lack of\ndirect experience with the physical nuances of the real world. \nTo address these issues, we propose a Grounding Large language model with Imperfect world MOdel (GLIMO), which \nutilizes proxy world models such as simulators to collect and synthesize trining data. \nGLIMO incorporates an LLM agent-based data generator to automatically create high-quality and diverse instruction datasets. The generator includes an iterative self-refining module for temporally consistent experience sampling, a diverse set of question-answering instruction seeds, and a retrieval-augmented generation module for reflecting on prior experiences.\nComprehensive experiments \nshow that our approach improve the performance of strong open-source LLMs like LLaMA-3 with a performance boost of 2.04 $\\times$, 1.54 $\\times$, and 1.82 $\\times$ across three different benchmarks, respectively.\nThe performance is able to compete with or surpass their larger counterparts such as GPT-4.", "keywords": [ "Robotics", "instruction tuning", "LLM" ], "rating_list": [ 3, 5, 1, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "ZNnmcddaB3", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Grani Adiwena Hanasusanto", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Associate Professor" }, { "name": "Hyuk Park", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Yingying Li", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 64, "n_formula_1": 29, "n_ref_uni": 45, "n_ref": 94, "n_ref_all": 132, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6064, "formula_len_all_1": 1583, "len_all": 160869, "len_all_1": 69334, "len_abs": 870, "len_title": 136, "len_sents": 47649, "len_sents_1": 31548, "n_sents": 398, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 901, "title": "Robust System Identification: Finite-sample Guarantees and Connection to Regularization", "abs": "We address the problem of learning nonlinear dynamical systems from a single sample trajectory. While the least squares estimate (LSE) is commonly used for this task, it suffers from poor identification errors when the sample size is small or the model fails to capture the system's true dynamics. To overcome these limitations, we propose a robust LSE framework, which incorporates robust optimization techniques, and prove that it is equivalent to regularizing LSE using general Schatten $p$-norms. We provide non-asymptotic performance guarantees for linear systems, achieving an error rate of $\\widetilde{\\mathcal{O}}(1/\\sqrt{T})$, and show that it avoids the curse of dimensionality, unlike state-of-the-art Wasserstein robust optimization models. Empirical results demonstrate substantial improvements in real-world system identification and online control tasks, outperforming existing methods.", "keywords": [ "dynamical system", "time series", "system identification", "optimization" ], "rating_list": [ 6, 6, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "ZNKLD0fX59", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Hong Xie", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Associate Professor" }, { "name": "Lingzhe Zhao", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Shiyu Zhao", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "Associate Professor" }, { "name": "Shucheng Gong", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Wenpu Li", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Intern" }, { "name": "Xiang Liu", "gender": "Not Specified", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Yin Zhang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Peidong Liu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 47, "n_ref": 102, "n_ref_all": 130, "n_fig": 13, "n_tab": 9, "L_tab": 3559, "n_element_tab": 472, "n_fig_1": 7, "n_tab_1": 8, "L_tab_1": 2485, "n_element_tab_1": 374, "formula_len_all": 708, "formula_len_all_1": 708, "len_all": 144480, "len_all_1": 63355, "len_abs": 440, "len_title": 135, "len_sents": 33788, "len_sents_1": 27036, "n_sents": 208, "n_sents_1": 164, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1547, "title": "CasualHDR: Robust High Dynamic Range 3D Gaussian Splatting from Casually Captured Videos", "abs": "In recent years, thanks to innovations in 3D scene representation, novel view synthesis and photo-realistic dense 3D reconstruction from multi-view images, such as neural radiance field (NeRF) and 3D Gaussian Splatting (3DGS), have garnered widespread attention due to their superior performance. However, most works rely on low dynamic range (LDR) images and representations of scenes, which limits the capturing of richer scene details. Prior works have focused on high dynamic range (HDR) scene recovery, typically require repeatedly capturing of multiple sharp images with different exposure times at fixed camera positions, which is time-consuming and challenging in practice.For a more flexible data acquisition, we propose a one-stage method: \\textbf{CasualHDR} to easily and robustly recover the 3D HDR scene from casual videos with auto-exposure (AE) enabled, even in the presence of severe motion blur and varying exposure time. CasualHDR contains a unified differentiable physical imaging model which jointly optimize (i.e. bundle adjust) exposure time, camera response function (CRF), continuous-time camera motion trajectory on $\\mathbb{SE}(3)$, and the 3DGS-based HDR scene. Extensive experiments demonstrate that our approach outperforms existing reconstruction methods in terms of robustness and rendering quality. Three applications can be achieved after the 3DGS HDR scene reconstruction: novel-view synthesis, image deblurring (deblur input images) and HDR editing (adjust the exposure time thus brightness of the input images).", "keywords": [ "High Dynamic Range", "3D Reconstruction", "Continous-time Trajectory", "Radiance Field", "Motion Blur" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "ZNHGsuMAgX", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 8, "author_info_list": [ { "name": "Jonathan Timcheck", "gender": "unknown", "institution": "Intel", "country": "US", "position": "Researcher" }, { "name": "Leobardo Campos-Macias", "gender": "Male", "institution": "Intel", "country": "MX", "position": "Researcher" }, { "name": "Mathis Richter", "gender": "unknown", "institution": "Intel", "country": "DE", "position": "Researcher" }, { "name": "Philipp Stratmann", "gender": "Male", "institution": "Intel", "country": "DE", "position": "Researcher" }, { "name": "Philipp Weidel", "gender": "unknown", "institution": "Intel", "country": "CH", "position": "Researcher" }, { "name": "Plank Philipp", "gender": "Male", "institution": "Intel", "country": "AT", "position": "Researcher" }, { "name": "Sumit Bam Shrestha", "gender": "Male", "institution": "Intel", "country": "US", "position": "Researcher" }, { "name": "Svea Marie Meyer", "gender": "Female", "institution": "Intel", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 31, "n_ref": 45, "n_ref_all": 56, "n_fig": 2, "n_tab": 3, "L_tab": 904, "n_element_tab": 95, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1475, "n_element_tab_1": 101, "formula_len_all": 484, "formula_len_all_1": 485, "len_all": 100087, "len_all_1": 58943, "len_abs": 1771, "len_title": 142, "len_sents": 30238, "len_sents_1": 29666, "n_sents": 171, "n_sents_1": 166, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1775, "title": "A Diagonal Structured State Space Model on Loihi 2 for Efficient Streaming Sequence Processing", "abs": "The unsustainable rise in energy cost from increasingly capable deep learning systems spurs computer architecture innovation beyond conventional deep learning accelerators such as GPUs.\nHowever, a novel computer architecture presents a problem: much of deep learning research has been optimized for conventional computer architectures, and the extent to which modern deep learning models can unlock improved efficiency on a novel computer architecture is not well understood. \nIn this work, we demonstrate for the first time that a State Space Model (SSM) can achieve substantial efficiency improvement when mapped to Loihi 2, a state-of-the-art neuromorphic research chip, versus a Jetson Orin Nano GPU (Jetson).\nSpecifically, we benchmark our SSM on sMNIST, psMNIST, and sCIFAR online token-by-token inference and find approximately 1000x increased energy efficiency and 75x improved latency and throughput on Loihi 2 with a decrease in accuracy of less than one to three percentage points compared to the full precision implementation on Jetson.\nWe comprehensively tailor our implementation to Loihi-specific features and constraints, such as the co-location of memory and compute as well as fixed precision arithmetic.\nOur results elucidate how SSMs meaningfully bridge conventional and neuromorphic hardware via their dual nature: SSMs can operate in an offline mode using convolution or scan, which is efficient on a GPU, or in an online mode as a recurrent network, which we show is efficient on Loihi 2.\nThis work provides a foundation for performant sequence models on neuromorphic hardware, potentially unlocking substantial improvements in latency-sensitive or energy-limited online inference applications, such as speech enhancement or vision for robotic control.", "keywords": [ "State Space Models", "Neuromorphic Computing", "Loihi 2", "New Hardware Paradigms" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ZMtq9pYw5e", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Runlin Lei", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Xinyi Huang", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "Undergrad student" }, { "name": "Yongchao Liu", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Yuwei Hu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Zhewei Wei", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 27, "n_ref": 59, "n_ref_all": 71, "n_fig": 7, "n_tab": 4, "L_tab": 962, "n_element_tab": 146, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2776, "n_element_tab_1": 174, "formula_len_all": 321, "formula_len_all_1": 321, "len_all": 154299, "len_all_1": 64214, "len_abs": 1383, "len_title": 113, "len_sents": 46454, "len_sents_1": 29874, "n_sents": 361, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1388, "title": "Scalable and Accurate Graph Reasoning with LLM-based Multi-Agents", "abs": "Recent research has explored the use of Large Language Models (LLMs) for tackling complex graph reasoning tasks. However, due to the intricacies of graph structures and the inherent limitations of LLMs in handling long text, current approaches often fail to deliver satisfactory accuracy, even on small-scale graphs and simple tasks. To address these challenges, we introduce GraphAgent-Reasoner, a fine-tuning-free framework that utilizes a multi-agent collaboration strategy for explicit and precise graph reasoning. Inspired by distributed graph computation theory, our framework decomposes graph problems into smaller, node-centric tasks that are distributed among multiple agents. The agents collaborate to solve the overall problem, significantly reducing the amount of information and complexity handled by a single LLM, thus enhancing the accuracy of graph reasoning. By simply increasing the number of agents, GraphAgent-Reasoner can efficiently scale to accommodate larger graphs with over 1,000 nodes. Evaluated on the GraphInstruct dataset, our framework demonstrates near-perfect accuracy on polynomial-time graph reasoning tasks, significantly outperforming the best available models, both closed-source and fine-tuned open-source variants. Our framework also demonstrates the capability to handle real-world graph reasoning applications such as webpage importance analysis.", "keywords": [ "Large Language Models; Graph Reasoning; Multi-Agents" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 1, 3, 3 ], "presentation_list": [ 2, 4, 3 ], "contribution_list": [ 2, 4, 2 ], "confidence_list": [ 5, 5, 5 ] }, { "paper_id": "ZLtWAhkWJm", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Jin Qian", "gender": "unknown", "institution": "Lawrence Berkeley National Lab", "country": "", "position": "Assistant Professor" }, { "name": "Marcus M. Noack", "gender": "Male", "institution": "Lawrence Berkeley National Lab", "country": "US", "position": "Researcher" }, { "name": "Rahul Khorana", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "MS student" } ], "n_formula": 41, "n_formula_1": 7, "n_ref_uni": 44, "n_ref": 102, "n_ref_all": 111, "n_fig": 1, "n_tab": 7, "L_tab": 1964, "n_element_tab": 135, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 883, "n_element_tab_1": 66, "formula_len_all": 4954, "formula_len_all_1": 616, "len_all": 210382, "len_all_1": 74535, "len_abs": 608, "len_title": 140, "len_sents": 59786, "len_sents_1": 32569, "n_sents": 648, "n_sents_1": 316, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 612, "title": "POLYATOMIC COMPLEXES: A TOPOLOGICALLY INFORMED LEARNING REPRESENTATION FOR ATOMISTIC SYSTEMS", "abs": "Developing robust physics-informed representations of chemical structures that enable models to learn topological inductive biases is challenging. In this manuscript, we present a representation of atomistic systems. We begin by proving that our representation satisfies all structural, geometric, efficiency, and generality constraints. Afterward, we provide a general algorithm to encode any atomistic system. Finally, we report performance comparable to state-of-the-art methods on numerous tasks. We open-source all code and datasets. The anonymized code and data are available in the supplementary material.", "keywords": [ "representation learning", "gaussian processes", "cheminformatics", "molecular representations" ], "rating_list": [ 3, 5, 3, 5, 3 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 1, 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3, 2 ] }, { "paper_id": "ZLZs2QG7vz", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Chi-Sheng Chen", "gender": "Male", "institution": "Neuro Industry, Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 29, "n_ref": 39, "n_ref_all": 47, "n_fig": 2, "n_tab": 5, "L_tab": 4330, "n_element_tab": 156, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2902, "n_element_tab_1": 39, "formula_len_all": 511, "formula_len_all_1": 531, "len_all": 138429, "len_all_1": 52877, "len_abs": 1236, "len_title": 137, "len_sents": 29907, "len_sents_1": 23975, "n_sents": 186, "n_sents_1": 136, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1273, "title": "NECOMIMI: Neural-Cognitive Multimodal EEG-informed Image Generation with Diffusion Models", "abs": "NECOMIMI (NEural-COgnitive MultImodal EEG-Informed Image Generation with Diffusion Models) introduces a novel framework for generating images directly from EEG signals using advanced diffusion models. Unlike previous works that focused solely on EEG-image classification through contrastive learning, NECOMIMI extends this task to image generation. The proposed NERV EEG encoder demonstrates state-of-the-art (SoTA) performance across multiple zero-shot classification tasks, including 2-way, 4-way, and 200-way, and achieves top results in our newly proposed Category-based Assessment Table (CAT) Score, which evaluates the quality of EEG-generated images based on semantic concepts. A key discovery of this work is that the model tends to generate abstract or generalized images, such as landscapes, rather than specific objects, highlighting the inherent challenges of translating noisy and low-resolution EEG data into detailed visual outputs. Additionally, we introduce the CAT Score as a new metric tailored for EEG-to-image evaluation and establish a benchmark on the ThingsEEG dataset. This study underscores the potential of EEG-to-image generation while revealing the complexities and challenges that remain in bridging neural activity with visual representation.", "keywords": [ "EEG", "diffusion model", "EEG to image", "brain-computer interface", "image reconstruction" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ZLOpfkf5aj", "primary_area": "reinforcement learning", "n_author": 11, "author_info_list": [ { "name": "Arnu Pretorius", "gender": "Male", "institution": "InstaDeep", "country": "ZA", "position": "Researcher" }, { "name": "Jemma Daniel", "gender": "Female", "institution": "InstaDeep", "country": "ZA", "position": "Researcher" }, { "name": "Juan Claude Formanek", "gender": "Male", "institution": "University of Cape Town", "country": "ZA", "position": "PhD student" }, { "name": "Liam Clark", "gender": "Male", "institution": "InstaDeep", "country": "ZA", "position": "Researcher" }, { "name": "Louay Ben Nessir", "gender": "Male", "institution": "InstaDeep", "country": "TN", "position": "Researcher" }, { "name": "Louise Beyers", "gender": "Female", "institution": "InstaDeep", "country": "ZA", "position": "Researcher" }, { "name": "Omayma Mahjoub", "gender": "Female", "institution": "InstaDeep", "country": "", "position": "Researcher" }, { "name": "Ruan John de Kock", "gender": "Male", "institution": "University of Cape Town", "country": "ZA", "position": "MS student" }, { "name": "Sasha Abramowitz", "gender": "Male", "institution": "InstaDeep", "country": "", "position": "Researcher" }, { "name": "Simon Verster Du Toit", "gender": "Male", "institution": "InstaDeep", "country": "TN", "position": "Researcher" }, { "name": "Wiem Khlifi", "gender": "Female", "institution": "InstaDeep", "country": "TN", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 13, "n_ref_uni": 60, "n_ref": 133, "n_ref_all": 151, "n_fig": 10, "n_tab": 3, "L_tab": 944, "n_element_tab": 33, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 135, "n_element_tab_1": 20, "formula_len_all": 1875, "formula_len_all_1": 1115, "len_all": 223577, "len_all_1": 58732, "len_abs": 1480, "len_title": 124, "len_sents": 65366, "len_sents_1": 28199, "n_sents": 476, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1435, "title": "Performant, Memory Efficient and Scalable Multi-Agent Reinforcement Learning", "abs": "As the field of multi-agent reinforcement learning (MARL) progresses towards larger and more complex environments, achieving strong performance while maintaining memory efficiency and scalability to many agents becomes increasingly important. Although recent research has led to several advanced algorithms, to date, none fully address all of these key properties simultaneously. In this work, we introduce Sable, a novel and theoretically sound algorithm that adapts the retention mechanism from Retentive Networks to MARL. Sable's retention-based sequence modelling architecture allows for computationally efficient scaling to a large number of agents, as well as maintaining a long temporal context, making it well-suited for large-scale partially observable environments. Through extensive evaluations across six diverse environments, we demonstrate how Sable is able to significantly outperform existing state-of-the-art methods in the majority of tasks (34 out of 45, roughly 75\\%). Furthermore, Sable demonstrates stable performance as we scale the number of agents, handling environments with more than a thousand agents while exhibiting a linear increase in memory usage. Finally, we conduct ablation studies to isolate the source of Sable's performance gains and confirm its efficient computational memory usage. Our results highlight Sable's performance and efficiency, positioning it as a leading approach to MARL at scale.", "keywords": [ "Multi-agent reinforcement learning", "Reinforcement Learning", "Decision Making", "Multi-agent systems" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 4, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "ZLAQ6Pjf9y", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Ahmed Abdulaal", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Ayodeji Ijishakin", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Daniel C. Alexander", "gender": "Male", "institution": "University College London", "country": "", "position": "Full Professor" }, { "name": "Daniel C. Castro", "gender": "unknown", "institution": "Microsoft Research", "country": "", "position": "Senior Researcher" }, { "name": "Hugo Laurence Fry", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "MS student" }, { "name": "Jack Gao", "gender": "Not Specified", "institution": "King's College London, University of London", "country": "", "position": "Undergrad student" }, { "name": "Stephanie Hyland", "gender": "Female", "institution": "Microsoft Research", "country": "", "position": "Researcher" }, { "name": "Nina Montana-Brown", "gender": "Female", "institution": "University College London, University of London", "country": "", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 16, "n_ref_uni": 73, "n_ref": 168, "n_ref_all": 211, "n_fig": 15, "n_tab": 6, "L_tab": 1145, "n_element_tab": 169, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 794, "n_element_tab_1": 131, "formula_len_all": 1128, "formula_len_all_1": 816, "len_all": 291984, "len_all_1": 66555, "len_abs": 1369, "len_title": 146, "len_sents": 85384, "len_sents_1": 30305, "n_sents": 653, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1377, "title": "An X-Ray Is Worth 15 Features: Sparse Autoencoders for Interpretable Radiology Report Generation", "abs": "Radiological services are experiencing unprecedented demand, leading to increased interest in automating radiology report generation. Existing Vision-Language Models (VLMs) suffer from hallucinations, lack interpretability, and require expensive fine-tuning. Sparse Autoencoders (SAEs) have been shown to provide a principled approach to reverse-engineer a model's internal activations into discrete, verifiable components. Thus, we introduce SAE-Rad, the first instance of using mechanistic interpretability techniques explicitly for a downstream multi-modal reasoning task. SAE-Rad uses a novel SAE architecture to decompose latent representations from a pre-trained vision transformer into human-interpretable features. These features are then labelled using an off-the-shelf language model and compiled into a full report for each image, eliminating the need for fine-tuning large models for this task. On the MIMIC-CXR dataset, SAE-Rad achieves competitive radiology-specific metrics compared to state-of-the-art models while using significantly fewer computational resources for training. Qualitative analysis reveals that SAE-Rad learns meaningful visual concepts and generates reports aligning closely with expert interpretations. Our results suggest that SAEs can enhance multimodal reasoning in healthcare, providing a more interpretable alternative to existing VLMs.", "keywords": [ "Radiology", "Mechanistic Interpretability", "Medical Imaging", "Sparse Autoencoders" ], "rating_list": [ 5, 5, 8, 5, 5 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 5, 5, 3, 5, 4 ] }, { "paper_id": "ZKRHiu5kE4", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Duncan Watson-Parris", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Kun Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Undergrad student" }, { "name": "Rose Yu", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Sumanth Varambally", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Yian Ma", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" } ], "n_formula": 60, "n_formula_1": 12, "n_ref_uni": 40, "n_ref": 95, "n_ref_all": 117, "n_fig": 12, "n_tab": 2, "L_tab": 363, "n_element_tab": 5, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5515, "formula_len_all_1": 1037, "len_all": 180261, "len_all_1": 53286, "len_abs": 1019, "len_title": 117, "len_sents": 55642, "len_sents_1": 23594, "n_sents": 521, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1022, "title": "Discovering Latent Structural Causal Models from Spatio-Temporal Data", "abs": "Many important phenomenon in scientific fields such as climate, neuroscience and epidemiology are naturally represented as spatiotemporal gridded data with complex interactions. Inferring causal relationships from these data is a difficult problem compounded by the high dimensionality of such data and the correlations between spatially proximate points. We present SPACY (SPAtiotemporal Causal discoverY), a novel framework based on variational inference, designed to explicitly model latent time-series and their causal relationships from spatially confined modes in the data. Our method uses an end-to-end training process that maximizes an evidence-lower bound (ELBO) for the data likelihood. Theoretically, we show that, under some conditions, the latent variables are identifiable up to transformation by an invertible matrix. Empirically, we show that SPACY outperforms state-of-the-art baselines on synthetic data, remains scalable for large grids, and identifies key known phenomena from real-world climate data.", "keywords": [ "Causal Inference", "Causal Represetation Learning", "Spatio-Temporal Dynamic Modeling", "Variational Inference", "Probablistic Graphical Model" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 1, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "ZK4VSRzBNC", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Kevin Jatin Vora", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 44, "n_formula_1": 12, "n_ref_uni": 22, "n_ref": 32, "n_ref_all": 47, "n_fig": 8, "n_tab": 7, "L_tab": 1345, "n_element_tab": 183, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6577, "formula_len_all_1": 1158, "len_all": 139939, "len_all_1": 65257, "len_abs": 1381, "len_title": 84, "len_sents": 44971, "len_sents_1": 31483, "n_sents": 371, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 1391, "title": "Reward Adaptation Via Q-Manipulation", "abs": "In this paper, we propose a new solution to reward adaptation (RA), the problem where the learning agent adapts to a target reward function based on one or multiple existing behaviors learned a priori under the same domain dynamics but different reward functions. \nRA has many applications, such as adapting an autonomous driving agent that can already operate either fast (if transporting goods) or comfortable (if carrying passengers) to operating both fast and comfortable (if transporting goods with human passengers onboard). Learning the target behavior from scratch is possible but often inefficient given the available source behaviors. Our work represents a new approach to RA\nvia the manipulation of Q-functions. Assuming that the target reward function is a known function of the source reward functions, our approach to RA computes bounds of the Q function. We introduce an iterative process to tighten the bounds, similar to value iteration. This enables action pruning in the target domain before learning even starts. We refer to such a method as Q-Manipulation (Q-M). We formally prove that our pruning strategy does not affect the optimality of the returned policy while empirically show that it improves the sample complexity. Comparison with baselines is performed in a variety of synthetic and simulation domains to demonstrate its effectiveness and generalizability.", "keywords": [ "Transfer RL", "Reusable RL", "Action Pruning", "Reward Adaptation" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 2, 4, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 5 ] }, { "paper_id": "ZK1NnjpjEs", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Bokai Hu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Undergrad student" }, { "name": "Pengtao Xie", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Sai Ashish Somayajula", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Xin Pan", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Zihan Huang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Undergrad student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 15, "n_ref": 48, "n_ref_all": 57, "n_fig": 0, "n_tab": 1, "L_tab": 543, "n_element_tab": 76, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1434, "n_element_tab_1": 131, "formula_len_all": 189, "formula_len_all_1": 189, "len_all": 113940, "len_all_1": 58323, "len_abs": 2175, "len_title": 153, "len_sents": 32155, "len_sents_1": 27427, "n_sents": 226, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 2185, "title": "Improving the Language Understanding Capabilities of Large Language Models Using Reinforcement Learning", "abs": "Large language models (LLMs), primarily built on decoder-only transformer architectures, excel in natural language generation tasks and have shown promise in adapting to diverse downstream tasks using zero-shot and few-shot prompting techniques. However, these prompting methods often fall short on natural language understanding (NLU) tasks, where smaller encoder-only models like BERT-base consistently outperform LLMs on benchmarks such as GLUE and SuperGLUE. In this paper, we explore two approaches\u2014supervised fine-tuning and proximal policy optimization (PPO)\u2014to enhance the NLU capabilities of LLMs. To reduce the computational cost of full-model fine-tuning, we integrate low-rank adaptation (LoRA) layers, restricting updates to these layers during both supervised fine-tuning and PPO stages. In the supervised fine-tuning approach, task-specific prompts are concatenated with input queries and ground-truth labels from the NLU training corpus, optimizing the model using the next-token prediction objective. Despite this, LLMs still underperform compared to encoder-only models like BERT-base on several NLU tasks. To address this gap, we employ PPO, a reinforcement learning technique that treats each token generation as an action and evaluates the sequence of generated tokens using a reward function based on their alignment with ground-truth answers. PPO then updates the model to maximize these rewards, effectively aligning its outputs with the correct labels. Our experiments with the LLAMA2-7B model demonstrate that PPO-based fine-tuning significantly improves performance, delivering an average gain of 6.3 points over supervised fine-tuning on the GLUE benchmark. PPO surpasses zero-shot prompting by 38.7 points and few-shot prompting by 26.1 points on GLUE, while also outperforming these baselines by 28.8 and 28.5 points on SuperGLUE. Additionally, PPO exceeds the performance of BERT-large, a strong baseline, with an average improvement of 2.7 points on GLUE and 9.3 points on SuperGLUE. These improvements are consistent across models such as Qwen2.5-7B and MPT-7B, highlighting PPO\u2019s robustness and effectiveness in enhancing the NLU capabilities of LLMs.", "keywords": [ "Large Language Models", "Natural Language Understanding", "Proximal Policy Optimization", "Fine-tuning", "GLUE", "SuperGLUE" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 3, 4, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 1, 1 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "ZK1LoTo10R", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Chao Chen", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Associate Professor" }, { "name": "Dimitris Samaras", "gender": "Male", "institution": "Stony Brook University", "country": "US", "position": "Full Professor" }, { "name": "Saumya Gupta", "gender": "Female", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 61, "n_ref": 147, "n_ref_all": 198, "n_fig": 18, "n_tab": 12, "L_tab": 3269, "n_element_tab": 311, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 406, "formula_len_all_1": 405, "len_all": 216212, "len_all_1": 73044, "len_abs": 1619, "len_title": 98, "len_sents": 63226, "len_sents_1": 34674, "n_sents": 541, "n_sents_1": 281, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1050, "title": "TopoDiffusionNet: A Topology-aware Diffusion Model", "abs": "Diffusion models excel at creating visually impressive images but often struggle to generate images with a specified topology. The Betti number, which represents the number of structures in an image, is a fundamental measure in topology. Yet, diffusion models fail to satisfy even this basic constraint. This limitation restricts their utility in applications requiring exact control, like robotics and environmental modeling. To address this, we propose TopoDiffusionNet (TDN), a novel approach that enforces diffusion models to maintain the desired topology. We leverage tools from topological data analysis, particularly persistent homology, to extract the topological structures within an image. We then design a topology-based objective function to guide the denoising process, preserving intended structures while suppressing noisy ones. Our experiments across four datasets demonstrate significant improvements in topological accuracy. TDN is the first to integrate topology with diffusion models, opening new avenues of research in this area.", "keywords": [ "Topology", "Diffusion Models", "Persistent Homology" ], "rating_list": [ 8, 6, 3, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 5, 3 ] }, { "paper_id": "ZJo6Radbqq", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Bingning Wang", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Haoyu Lu", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Jing Liu", "gender": "Female", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "Full Professor" }, { "name": "Longteng Guo", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Tongtian Yue", "gender": "Male", "institution": ", Institute of automation, Chinese academy of science", "country": "CN", "position": "PhD student" }, { "name": "Yifan Du", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Yuqi Huo", "gender": "Male", "institution": "Baichuan Intelligent Technology", "country": "CN", "position": "Researcher" }, { "name": "weipeng chen", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Joe Z", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 109, "n_ref_all": 148, "n_fig": 18, "n_tab": 13, "L_tab": 4061, "n_element_tab": 511, "n_fig_1": 12, "n_tab_1": 3, "L_tab_1": 419, "n_element_tab_1": 48, "formula_len_all": 258, "formula_len_all_1": 0, "len_all": 185681, "len_all_1": 49057, "len_abs": 4705, "len_title": 122, "len_sents": 53849, "len_sents_1": 22198, "n_sents": 456, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1665, "title": "Needle In A Video Haystack: A Scalable Synthetic Evaluator for Video MLLMs", "abs": "Video understanding is a crucial next step for multimodal large language models (MLLMs).\nVarious benchmarks are introduced for better evaluating the MLLMs.\nNevertheless, current video benchmarks are still inefficient for evaluating video models during iterative development due to the high cost of constructing datasets and the difficulty in isolating specific skills.\nIn this paper, we propose VideoNIAH (Video Needle in A Haystack), a benchmark construction framework through synthetic video generation. \nVideoNIAH decouples video content from their query-responses by inserting unrelated visual 'needles' into original videos. \nThe framework automates the generation of query-response pairs using predefined rules, minimizing manual labor. The queries focus on specific aspects of video understanding, enabling more skill-specific evaluations. The separation between video content and the queries also allow for increased video variety and evaluations across different lengths.\nUtilizing VideoNIAH, we compile a video benchmark, VNBench, which includes tasks such as retrieval, ordering, and counting to evaluate three key aspects of video understanding: temporal perception, chronological ordering, and spatio-temporal coherence. We conduct a comprehensive evaluation of both proprietary and open-source models, uncovering significant differences in their video understanding capabilities across various tasks. Additionally, we perform an in-depth analysis of the test results and model configurations. Based on these findings, we provide some advice for improving video MLLM training, offering valuable insights to guide future research and model development.", "keywords": [ "video MLLM" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "ZJj1r4gWIy", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Bo Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Guangyi Chen", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "US", "position": "Postdoc" }, { "name": "Hao Wang", "gender": "Male", "institution": "ByteDance", "country": "CN", "position": "Researcher" }, { "name": "Haoxuan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Mingming Gong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Sen Cui", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiao-Hua Zhou", "gender": "unknown", "institution": "School of mathematical Science, Peking University, Peking University", "country": "", "position": "Full Professor" } ], "n_formula": 60, "n_formula_1": 27, "n_ref_uni": 43, "n_ref": 91, "n_ref_all": 114, "n_fig": 4, "n_tab": 5, "L_tab": 3788, "n_element_tab": 227, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1061, "n_element_tab_1": 38, "formula_len_all": 8186, "formula_len_all_1": 3782, "len_all": 158089, "len_all_1": 67797, "len_abs": 1153, "len_title": 88, "len_sents": 47320, "len_sents_1": 29960, "n_sents": 349, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1157, "title": "Counterfactual Delayed Feedback Learning", "abs": "Estimation of heterogeneous treatment effects has gathered much attention in recent years and has been widely adopted in medicine, economics, and marketing. Previous studies assumed that one of the potential outcomes of interest could be observed timely and accurately. However, a more practical scenario is that treatment takes time to produce causal effects on the outcomes. For example, drugs take time to produce medical utility for patients and users take time to purchase items after being recommended, and ignoring such delays in feedback can lead to biased estimates of heterogeneous treatment effects. To address the above problem, we study the impact of observation time on estimating heterogeneous treatment effects by further considering the potential response time that potential outcomes have. We theoretically prove the identifiability results and further propose a principled learning approach, known as CFR-DF (Counterfactual Regression with Delayed Feedback), to simultaneously learn potential response times and potential outcomes of interest. Results on both simulated and real-world datasets demonstrate the effectiveness of our method.", "keywords": [ "Counterfactual", "Delayed Feedback", "HTE" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 4, 2, 2 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "ZJftXKy12x", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Hanling Tian", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Mingzhen He", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Ruikai Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Xiaolin Huang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Yuhang Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Zhehao Huang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhengbao He", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 33, "n_ref": 65, "n_ref_all": 78, "n_fig": 9, "n_tab": 3, "L_tab": 59, "n_element_tab": 6, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2178, "n_element_tab_1": 34, "formula_len_all": 723, "formula_len_all_1": 575, "len_all": 108690, "len_all_1": 50360, "len_abs": 1205, "len_title": 131, "len_sents": 32834, "len_sents_1": 22657, "n_sents": 259, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1210, "title": "Simulating Training Dynamics to Reconstruct Training Data from Deep Neural Networks", "abs": "Whether deep neural networks (DNNs) memorize the training data is a fundamental open question in understanding deep learning. A direct way to verify the memorization of DNNs is to reconstruct training data from DNNs' parameters. Since parameters are gradually determined by data throughout training, characterizing training dynamics is important for reconstruction. Pioneering works rely on the linear training dynamics of shallow NNs with large widths, but cannot be extended to more practical DNNs which have non-linear dynamics. We propose Simulation of training Dynamics (SimuDy) to reconstruct training data from DNNs. Specifically, we simulate the training dynamics by training the model from the initial parameters with a dummy dataset, then optimize this dummy dataset so that the simulated dynamics reach the same final parameters as the true dynamics. By incorporating dummy parameters in the simulated dynamics, SimuDy effectively describes non-linear training dynamics. Experiments demonstrate that SimuDy significantly outperforms previous approaches when handling non-linear training dynamics, and for the first time, most training samples can be reconstructed from a trained ResNet's parameters.", "keywords": [ "Dataset reconstruction", "Training dynamics", "Memorization" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "ZJCSlcEjEn", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Andrei Lupu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Benjamin Ellis", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "", "position": "PhD student" }, { "name": "Jakob Nicolaus Foerster", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Lize Alberts", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 20, "n_ref": 31, "n_ref_all": 60, "n_fig": 13, "n_tab": 2, "L_tab": 792, "n_element_tab": 15, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 172601, "len_all_1": 66927, "len_abs": 1433, "len_title": 124, "len_sents": 71862, "len_sents_1": 36234, "n_sents": 417, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1446, "title": "CURATe: Benchmarking Personalised Alignment of Conversational AI Assistants", "abs": "We introduce a multi-turn benchmark for evaluating personalised alignment in LLM-based AI assistants, focusing on their ability to handle user-provided safety-critical contexts. Our assessment of ten leading models across five scenarios (each with 337 use cases) reveals systematic inconsistencies in maintaining user-specific consideration, with even top-rated \"harmless\" models making recommendations that should be recognised as obviously harmful to the user given the context provided. Key failure modes include inappropriate weighing of conflicting preferences, sycophancy (prioritising user preferences above safety), a lack of attentiveness to critical user information within the context window, and inconsistent application of user-specific knowledge. The same systematic biases were observed in OpenAI's o1, suggesting that strong reasoning capacities do not necessarily transfer to this kind of personalised thinking. We find that prompting LLMs to consider safety-critical context significantly improves performance, unlike a generic `harmless and helpful' instruction. Based on these findings, we propose research directions for embedding self-reflection capabilities, online user modelling, and dynamic risk assessment in AI assistants. Our work emphasises the need for nuanced, context-aware approaches to alignment in systems designed for persistent human interaction, aiding the development of safe and considerate AI assistants.", "keywords": [ "LLM", "safety and alignment", "agentic AI", "personalised alignment", "context-sensitive", "recommender systems", "benchmark", "multi-turn evaluation", "dialogue assistants" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "ZJ9LglIakj", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Alberto Marchesi", "gender": "Male", "institution": "Politecnico di Milano", "country": "", "position": "Assistant Professor" }, { "name": "Anna Lunghi", "gender": "unknown", "institution": "Polytechnic Institute of Milan", "country": "IT", "position": "PhD student" }, { "name": "Francesco Emanuele Stradi", "gender": "Male", "institution": "Polytechnic Institute of Milan", "country": "IT", "position": "PhD student" }, { "name": "Matteo Castiglioni", "gender": "unknown", "institution": "Politecnico di Milano", "country": "IT", "position": "Assistant Professor" }, { "name": "Nicola Gatti", "gender": "Male", "institution": "Polytechnic Institute of Milan", "country": "IT", "position": "Full Professor" } ], "n_formula": 205, "n_formula_1": 25, "n_ref_uni": 28, "n_ref": 81, "n_ref_all": 98, "n_fig": 2, "n_tab": 3, "L_tab": 299, "n_element_tab": 30, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 24621, "formula_len_all_1": 2106, "len_all": 250383, "len_all_1": 79185, "len_abs": 1127, "len_title": 138, "len_sents": 76469, "len_sents_1": 35836, "n_sents": 701, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1169, "title": "Learning Constrained Markov Decision Processes With Non-stationary Rewards and Constraints", "abs": "In constrained Markov decision processes (CMDPs) with adversarial rewards and constraints, a well-known impossibility result prevents any algorithm from attaining both sublinear regret and sublinear constraint violation, when competing against a best-in-hindsight policy that satisfies constraints on average. In this paper, we show that this negative result can be eased in CMDPs with non-stationary rewards and constraints, by providing algorithms whose performances smoothly degrade as non-stationarity increases.\nSpecifically, we propose algorithms attaining $\\tilde{\\mathcal{O}} (\\sqrt{T} + C)$ regret and positive constraint violation under bandit feedback, where $C$ is a corruption value measuring the environment non-stationarity. This can be $\\Theta(T)$ in the worst case, coherently with the impossibility result for adversarial~CMDPs. First, we design an algorithm with the desired guarantees when $C$ is known. Then, in the case $C$ is unknown, we show how to obtain the same results by embedding such an algorithm in a general meta-procedure. This is of independent interest, as it can be applied to any non-stationary constrained online learning setting.", "keywords": [ "CMDPs", "Non-stationary", "Online learning" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "ZINaxJyoQr", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Hans-Oliver Hansen", "gender": "Male", "institution": "Universit\u00e4t zu L\u00fcbeck", "country": "DE", "position": "PhD student" }, { "name": "Marius Jahrens", "gender": "Male", "institution": "Universit\u00e4t zu L\u00fcbeck", "country": "DE", "position": "PhD student" }, { "name": "Thomas Martinetz", "gender": "Male", "institution": "Universit\u00e4t zu L\u00fcbeck", "country": "DE", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 13, "n_ref": 19, "n_ref_all": 21, "n_fig": 1, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 800, "formula_len_all_1": 800, "len_all": 50948, "len_all_1": 50948, "len_abs": 1351, "len_title": 150, "len_sents": 15756, "len_sents_1": 15756, "n_sents": 118, "n_sents_1": 118, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 1355, "title": "Why Barlow Twins Work: The Critical Role of Normalization and Its Link to Sample Contrastive Learning", "abs": "Barlow Twins is a feature-contrastive self-supervised learning framework built on the principle of redundancy reduction. The idea is to train a network by maximizing the correlation between corresponding features and minimizing the correlation between non-corresponding features in distorted views of the same image, through this facilitating effective pretraining of a backbone network for a subsequent classification head. This is achieved by diagonalizing the cross-correlation matrix of the network\u2019s representations and scaling it towards the identity matrix. We show that the cross-correlation matrix of distorted images is inherently symmetric, independent of the backbone network's weights, which leads to two key insights: (i) the cross-correlation matrix can always be diagonalized using a linear transformation (layer), and (ii) the core idea of maximizing correlations between corresponding features while minimizing them for non-corresponding features alone is insufficient for effective backbone network pretraining. Nevertheless, Barlow Twins provide highly effective pretraining. We show that this is due to the normalization of the cross-correlation matrix in the Barlow Twins cost function. This normalization leads to minima of the cost function which are equivalent to the minima of sample contrastive approaches to enforce invariance.", "keywords": [ "self-supervised Learning", "pretaining", "cross-correlation matrix", "covariance matrix", "diagonalization" ], "rating_list": [ 1, 1, 1, 3 ], "soundness_list": [ 1, 1, 3, 2 ], "presentation_list": [ 1, 1, 3, 2 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "ZIFkrT1GwM", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" }, { "name": "Jerry Yao-Chieh Hu", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Maojiang Su", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Weimin Wu", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" } ], "n_formula": 249, "n_formula_1": 57, "n_ref_uni": 37, "n_ref": 103, "n_ref_all": 110, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 30669, "formula_len_all_1": 6982, "len_all": 306204, "len_all_1": 75213, "len_abs": 851, "len_title": 134, "len_sents": 89552, "len_sents_1": 27256, "n_sents": 1020, "n_sents_1": 290, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 873, "title": "Pretrained Transformers are Deep Optimizers: Provable In-Context Learning for Deep Model Training", "abs": "We investigate the transformer's capability for in-context learning (ICL) to simulate the training process of deep models. \nOur key contribution is providing a positive example of using a pretrained transformer to train a deep neural network by gradient descent in an implicit fashion via ICL. \nSpecifically, we provide an explicit construction of a $(2N+4)L$-layer transformer capable of simulating $L$ gradient descent steps of an $N$-layer ReLU network through ICL.\nWe also give the theoretical guarantees for the approximation within any given error and the convergence of the ICL gradient descent.\nAdditionally, we extend our analysis to the more practical setting using Softmax-based transformers. \nWe validate our findings on synthetic datasets for 3-layer, 4-layer, and 6-layer neural networks.\nThe results show that ICL performance matches that of direct training.", "keywords": [ "foundation model", "transformer", "in-context learning" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 2, 3 ] }, { "paper_id": "ZHhBawo3k5", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Jason Cong", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "Neha Prakriya", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Yizhou Sun", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "Zifan He", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Ziniu Hu", "gender": "Male", "institution": "xAI", "country": "US", "position": "Member of Technical Staff" }, { "name": "Zongyue Qin", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 13, "n_ref_uni": 40, "n_ref": 84, "n_ref_all": 107, "n_fig": 5, "n_tab": 17, "L_tab": 5564, "n_element_tab": 761, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 816, "n_element_tab_1": 136, "formula_len_all": 3669, "formula_len_all_1": 809, "len_all": 171693, "len_all_1": 66108, "len_abs": 1877, "len_title": 123, "len_sents": 43129, "len_sents_1": 27925, "n_sents": 396, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1885, "title": "Optimized Multi-Token Joint Decoding With Auxiliary Model for LLM Inference", "abs": "Large language models (LLMs) have achieved remarkable success across diverse tasks, yet their inference processes are hindered by substantial time and energy demands due to single-token generation at each decoding step. While previous methods such as speculative decoding mitigate these inefficiencies by producing multiple tokens per step, each token is still generated by its single-token distribution,\nthereby enhancing speed without improving effectiveness. In contrast, our work simultaneously enhances inference speed and improves the output effectiveness. We consider multi-token joint decoding (MTJD), which generates multiple tokens from their joint distribution at each iteration, theoretically reducing perplexity and enhancing task performance. However, MTJD suffers from the high cost of sampling from the joint distribution of multiple tokens. Inspired by speculative decoding, we introduce multi-token assisted decoding (MTAD), a novel framework designed to accelerate MTJD. MTAD leverages a smaller auxiliary model to approximate the joint distribution of a larger model, incorporating a verification mechanism that not only ensures the accuracy of this approximation, but also improves the\ndecoding efficiency over conventional speculative decoding. Theoretically, we demonstrate that MTAD closely approximates exact MTJD with bounded error. Empirical evaluations using Llama-2 and OPT models ranging from 13B to 70B parameters across various tasks reveal that MTAD reduces perplexity by 21.2% and improves downstream performance compared to standard single-token sampling.\nFurthermore, MTAD achieves a 1.42\u00d7 speed-up and consumes 1.54\u00d7 less energy than conventional speculative decoding methods. These results highlight MTAD\u2019s ability to make multi-token joint decoding both effective and efficient, promoting more sustainable and high-performance deployment of LLMs.", "keywords": [ "LLM Inference", "Speculative Decoding" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "ZHTYtXijEn", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Boi Faltings", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Professor" }, { "name": "Ewa Miazga", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "", "position": "Undergrad student" }, { "name": "Zeki Doruk Erden", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 4, "n_ref_uni": 20, "n_ref": 26, "n_ref_all": 58, "n_fig": 7, "n_tab": 4, "L_tab": 1817, "n_element_tab": 110, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1249, "formula_len_all_1": 150, "len_all": 169644, "len_all_1": 69926, "len_abs": 1061, "len_title": 142, "len_sents": 74604, "len_sents_1": 37838, "n_sents": 464, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1062, "title": "Directed Structural Adaptation to Overcome Statistical Conflicts and Enable Continual Learning", "abs": "Adaptive networks today rely on overparameterized fixed topologies that cannot break through the statistical conflicts they encounter in the data they are exposed to, and are prone to \"catastrophic forgetting\" as the network attempts to reuse the existing structures to learn new task. We propose a structural adaptation method, DIRAD, that can complexify as needed and in a directed manner without being limited by statistical conflicts within a dataset. We then extend this method and present the PREVAL framework, designed to prevent \"catastrophic forgetting\" in continual learning by detection of new data and assigning encountered data to suitable models adapted to process them, without needing task labels anywhere in the workflow. We show the reliability of the DIRAD in growing a network with high performance and orders-of-magnitude simpler than fixed topology networks; and demonstrate the proof-of-concept operation of PREVAL, in which continual adaptation to new tasks is observed while being able to detect and discern previously-encountered tasks.", "keywords": [ "structural adaptation", "continual learning", "growth" ], "rating_list": [ 1, 3, 3 ], "soundness_list": [ 1, 3, 2 ], "presentation_list": [ 1, 2, 1 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 5, 3, 3 ] }, { "paper_id": "ZGyapLnDb7", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Bong Gyun Kang", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "PhD student" }, { "name": "Dongjin Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Jisoo Mok", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Junsung Park", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Sungroh Yoon", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 66, "n_ref": 99, "n_ref_all": 129, "n_fig": 6, "n_tab": 16, "L_tab": 13394, "n_element_tab": 908, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2367, "n_element_tab_1": 162, "formula_len_all": 577, "formula_len_all_1": 143, "len_all": 200793, "len_all_1": 62378, "len_abs": 1308, "len_title": 135, "len_sents": 48113, "len_sents_1": 30640, "n_sents": 315, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1315, "title": "Balancing Gradient Frequencies Facilitates Inductive Inference in Algorithmic Reasoning", "abs": "Inductive inference, or extrapolation of general rules from finite instances, is understood to be the foundation of human intelligence. Unfortunately, Deep Neural Networks (DNNs) struggle with inductive inference and thus fail to learn even the simplest algorithms in Algorithmic Reasoning (AR). Existing research efforts on AR with DNNs are limited to those on the architectural design for DNNs. In this study, we investigate the influence of optimization techniques on AR performance. Through toy experiments designed to understand an optimizer's susceptibility to shortcuts in AR, we reveal that Adam, the naive choice of optimization, is easily fooled by spurious correlations. To overcome this shortcoming of Adam, we propose a novel optimizer that avoids spurious correlations by balancing gradients of low- and high-frequencies (BGF). We present extensive experiments and analyses to demonstrate the broad and multifaceted advantages of BGF across various architectures and AR tasks. In particular, BGF expands the AR capability of all explored DNN models and even shows the potential to enable learning of tasks that they previously failed at. The observed success of BGF in climbing the Chomsky hierarchy underscores the importance of optimization for developing advanced artificial intelligence with DNNs.", "keywords": [ "Algorithmic Reasoning", "Inductive Inference", "Spurious Correlation", "Out-of-Distribution Generalization", "Optimization" ], "rating_list": [ 3, 8, 6, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "ZGqlkqAt18", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Cong Guan", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "PhD student" }, { "name": "Feng Chen", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Jiacheng Xu", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Lei Yuan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Researcher" }, { "name": "Zhilong Zhang", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "MS student" }, { "name": "Zongzhang Zhang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Yang Yu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Professor" } ], "n_formula": 21, "n_formula_1": 11, "n_ref_uni": 47, "n_ref": 100, "n_ref_all": 125, "n_fig": 9, "n_tab": 1, "L_tab": 279, "n_element_tab": 20, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 279, "n_element_tab_1": 20, "formula_len_all": 2338, "formula_len_all_1": 907, "len_all": 157081, "len_all_1": 65010, "len_abs": 1541, "len_title": 126, "len_sents": 48264, "len_sents_1": 32322, "n_sents": 342, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1569, "title": "Diffusion-Guided Safe Policy Optimization From Cost-Label-Free Offline Dataset", "abs": "Offline safe reinforcement learning (RL) aims to guarantee the safety of decision-making in both training and deployment phases by learning the safe policy entirely from offline data without further interaction with the environment, which pushes the RL towards real-world applications. Previous efforts in offline safe RL typically presume the presence of Markovian costs within the dataset. However, the design of a Markovian cost function involves rehearsal of all potentially unsafe cases, which is inefficient and even unfeasible in many practical tasks. In this work, we take a further step forward by learning a safe policy from an offline dataset without any cost labels, but with a small number of safe demonstrations included. To solve this problem, we propose a two-stage optimization method called **D**iffusion-guided **S**afe **P**olicy **O**ptimization (**DSPO**). Initially, we derive trajectory-wise safety signals by training a return-agnostic discriminator. Subsequently, we train a conditional diffusion model that generates trajectories conditioned both on the trajectory return and the safety signal. Remarkably, the trajectories generated by our diffusion model not only yield high returns but also comply with the safety signals, from which we can derive a desirable policy through behavior cloning (BC). The evaluation experiments conducted across tasks from the SafetyGym, BulletGym, and MetaDrive environments demonstrate that our approach can achieve a safe policy with high returns, significantly outperforming various established baselines.", "keywords": [ "Reinforcement Learning", "Offline Safe Reinforcement Learning", "Diffusion Model" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ZGqd0cbBvm", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Abhay Puri", "gender": "Male", "institution": "ServiceNow Research", "country": "CA", "position": "Researcher" }, { "name": "Alexandre Drouin", "gender": "Male", "institution": "Laval university", "country": "CA", "position": "Adjunct Professor" }, { "name": "Alexandre Lacoste", "gender": "Male", "institution": "ServiceNow", "country": "CA", "position": "Research Scientist" }, { "name": "Amirhossein Abaskohi", "gender": "Male", "institution": "ServiceNow Inc", "country": "CA", "position": "Intern" }, { "name": "Christopher J Pal", "gender": "unknown", "institution": "Polytechnique Montreal", "country": "CA", "position": "Full Professor" }, { "name": "David Vazquez", "gender": "Male", "institution": "ServiceNow research", "country": "CA", "position": "Researcher" }, { "name": "Gaurav Sahu", "gender": "Male", "institution": "University of Waterloo", "country": "", "position": "PhD student" }, { "name": "Issam H. Laradji", "gender": "Male", "institution": "ServiceNow", "country": "US", "position": "Researcher" }, { "name": "Mohammad Chegini", "gender": "Male", "institution": "University of Victoria", "country": "", "position": "MS student" }, { "name": "Nicolas Chapados", "gender": "Male", "institution": "ServiceNow Research", "country": "CA", "position": "VP Research" }, { "name": "Perouz Taslakian", "gender": "Female", "institution": "ServiceNow", "country": "CA", "position": "Researcher" }, { "name": "Valentina Zantedeschi", "gender": "Female", "institution": "ServiceNow Research", "country": "", "position": "Researcher" }, { "name": "Juan Rodriguez", "gender": "Male", "institution": "ServiceNow Research", "country": "CA", "position": "PhD student" }, { "name": "sai rajeswar mudumba", "gender": "Male", "institution": "ServiceNow", "country": "CA", "position": "Research Scientist" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 26, "n_ref": 54, "n_ref_all": 95, "n_fig": 14, "n_tab": 8, "L_tab": 5258, "n_element_tab": 360, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 2261, "n_element_tab_1": 111, "formula_len_all": 41, "formula_len_all_1": 128, "len_all": 192929, "len_all_1": 75237, "len_abs": 2910, "len_title": 136, "len_sents": 62523, "len_sents_1": 34383, "n_sents": 507, "n_sents_1": 268, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1557, "title": "InsightBench: Evaluating Business Analytics Agents Through Multi-Step Insight Generation", "abs": "Data analytics is essential for extracting valuable insights from data that can assist organizations in making effective decisions. We introduce InsightBench, a benchmark dataset with three key features. First, it consists of 100 datasets representing diverse business use cases such as finance and incident management, each accompanied by a carefully curated set of insights planted in the datasets. Second, unlike existing benchmarks focusing on answering single queries, InsightBench evaluates agents based on their ability to perform end-to-end data analytics, including formulating questions, interpreting answers, and generating a summary of insights and actionable steps. Third, we conducted comprehensive quality assurance to ensure that each dataset in the benchmark had clear goals and included relevant and meaningful questions and analysis. Furthermore, we implement a two-way evaluation mechanism using LLaMA-3 as an effective, open-source evaluator to assess agents\u2019 ability to extract insights. We also propose AgentPoirot, our baseline data analysis agent capable of performing end-to-end data analytics. Our evaluation on InsightBench shows that AgentPoirot outperforms existing approaches (such as Pandas Agent) that focus on resolving single queries. We also compare the performance of open- and closed-source LLMs and various evaluation strategies. Overall, this benchmark serves as a testbed to motivate further development in comprehensive automated data analytics and can be accessed here: https://github.com/ServiceNow/insight-bench.", "keywords": [ "Automated Data Analysis", "Data Analytics Benchmark", "LLM agents", "Code Generation", "LLM Evaluation" ], "rating_list": [ 8, 3, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 4, 3, 4, 3 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "ZGkfoufDaU", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Eric Yeats", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Hai Li", "gender": "Female", "institution": "Duke University", "country": "US", "position": "Professor" }, { "name": "Hao Frank Yang", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" }, { "name": "Jianyi Zhang", "gender": "unknown", "institution": "National Artificial Intelligence Research Resource Pilot & Duke University", "country": "US", "position": "Principal investigator" }, { "name": "Jingwei Sun", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Jingyang Zhang", "gender": "Male", "institution": "Electrical and Computer Engineering, Duke University", "country": "US", "position": "PhD student" }, { "name": "Martin Kuo", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Yang Ouyang", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 3, "n_ref_uni": 31, "n_ref": 104, "n_ref_all": 135, "n_fig": 5, "n_tab": 3, "L_tab": 1117, "n_element_tab": 64, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 800, "n_element_tab_1": 43, "formula_len_all": 659, "formula_len_all_1": 311, "len_all": 132046, "len_all_1": 64258, "len_abs": 1428, "len_title": 135, "len_sents": 41325, "len_sents_1": 31498, "n_sents": 295, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 86, "L_abs": 1474, "title": "Min-K%++: Improved Baseline for Pre-Training Data Detection from Large Language Models", "abs": "The problem of pre-training data detection for large language models (LLMs) has received growing attention due to its implications in critical issues like copyright violation and test data contamination. Despite improved performance, existing methods (including the state-of-the-art, Min-K%) are mostly developed upon simple heuristics and lack solid, reasonable foundations. In this work, we propose a novel and theoretically motivated methodology for pre-training data detection, named Min-K%++. Specifically, we present a key insight that training samples tend to be local maxima of the modeled distribution along each input dimension through maximum likelihood training, which in turn allow us to insightfully translate the problem into identification of local maxima. Then, we design our method accordingly that works under the discrete distribution modeled by LLMs, whose core idea is to determine whether the input forms a mode or has relatively high probability under the conditional categorical distribution. Empirically, the proposed method achieves new SOTA performance across multiple settings (evaluated with 5 families of 10 models and 2 benchmarks). On the WikiMIA benchmark, Min-K%++ outperforms the runner-up by 6.2% to 10.5% in detection AUROC averaged over five models. On the more challenging MIMIR benchmark, it consistently improves upon reference-free methods while performing on par with reference-based method that requires an extra reference model.", "keywords": [ "pre-training data detection", "large language model" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "ZGRZ5GPKWX", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Chuanqi Cheng", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "MS student" }, { "name": "Hongda Sun", "gender": "Male", "institution": "Gaoling School of Artificial Intelligence, Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Rui Yan", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Xinrong Hu", "gender": "Female", "institution": "Wuhan Textile University", "country": "CN", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 47, "n_ref_all": 66, "n_fig": 2, "n_tab": 2, "L_tab": 961, "n_element_tab": 60, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 774, "n_element_tab_1": 62, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 99752, "len_all_1": 55503, "len_abs": 1245, "len_title": 151, "len_sents": 30931, "len_sents_1": 26423, "n_sents": 225, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1265, "title": "DNASpeech: A Contextualized and Situated Text-to-Speech Dataset with Dialogues, Narratives and Actions", "abs": "In this paper, we propose contextualized and situated text-to-speech (CS-TTS), a novel TTS task to promote more accurate and customized speech generation using prompts with Dialogues, Narratives, and Actions (DNA). While prompt-based TTS methods facilitate controllable speech generation, existing TTS datasets lack situated descriptive prompts aligned with speech data. To address this data scarcity, we develop an automatic annotation pipeline enabling multifaceted alignment among speech clips, content text, and their respective descriptions. Based on this pipeline, we present DNASpeech, a novel CS-TTS dataset with high-quality speeches with DNA prompt annotations. DNASpeech contains 2,395 distinct characters, 4,452 scenes, and 22,975 dialogue utterances, along with over 18 hours of high-quality speech recordings. To accommodate more specific task scenarios, we establish a leaderboard featuring two new subtasks for evaluation: CS-TTS with narratives and CS-TTS with dialogues. We also design an intuitive baseline model for comparison with existing state-of-the-art TTS methods on our leaderboard. Comprehensive experimental results demonstrate the quality and effectiveness of \\dataname, validating its potential to drive advancements in the TTS field.", "keywords": [ "Text-to-Speech", "Voice Generation", "Prompt" ], "rating_list": [ 3, 3, 6, 6, 5 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 4, 3 ] }, { "paper_id": "ZGRRC514rI", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Dorina Thanou", "gender": "Female", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "", "position": "Researcher" }, { "name": "Manuel Madeira", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Pascal Frossard", "gender": "Not Specified", "institution": "EPFL", "country": "CH", "position": "Full Professor" }, { "name": "Yiming QIN", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" } ], "n_formula": 111, "n_formula_1": 35, "n_ref_uni": 70, "n_ref": 204, "n_ref_all": 279, "n_fig": 16, "n_tab": 9, "L_tab": 1848, "n_element_tab": 195, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 26, "n_element_tab_1": 5, "formula_len_all": 13924, "formula_len_all_1": 17064, "len_all": 352888, "len_all_1": 103578, "len_abs": 4259, "len_title": 108, "len_sents": 106821, "len_sents_1": 40588, "n_sents": 854, "n_sents_1": 272, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1602, "title": "DeFoG: Defogging Discrete Flow Matching for Graph Generation", "abs": "Graph generation is fundamental in diverse scientific applications, due to its ability to reveal the underlying distribution of complex data, and eventually generate new, realistic data points.\nDespite the success of diffusion models in this domain, those face limitations in sampling efficiency and flexibility, stemming from the tight coupling between the training and sampling stages.\nTo address this, we propose DeFoG, a novel framework using discrete flow matching for graph generation. DeFoG employs a flow-based approach that features an efficient linear interpolation noising process and a flexible denoising process based on a continuous-time Markov chain formulation.\nWe leverage an expressive graph transformer and ensure desirable node permutation properties to respect graph symmetry.\nCrucially, our framework enables a disentangled design of the training and sampling stages, enabling more effective and efficient optimization of model performance.\nWe navigate this design space by introducing several algorithmic improvements that boost the model performance, consistently surpassing existing diffusion models.\nWe also theoretically demonstrate that, for general discrete data, discrete flow models can faithfully replicate the ground truth distribution - a result that naturally extends to graph data and reinforces DeFoG's foundations.\nExtensive experiments show that DeFoG achieves state-of-the-art results on synthetic and molecular datasets, improving both training and sampling efficiency over diffusion models, and excels in conditional generation on a digital pathology dataset.", "keywords": [ "Graph Generation", "Discrete Flow Models", "Molecule Generation", "Permutation Equivariance", "Flow Matching" ], "rating_list": [ 3, 6, 6, 5 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "ZFxpclrCCf", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Bin Xie", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "MS student" }, { "name": "Jiale Cao", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Tiancai Wang", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Xiangyu Zhang", "gender": "Male", "institution": "MEGVII Technology", "country": "CN", "position": "Principal Researcher" }, { "name": "Yingfei Liu", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 42, "n_ref": 87, "n_ref_all": 102, "n_fig": 9, "n_tab": 4, "L_tab": 1257, "n_element_tab": 279, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1257, "n_element_tab_1": 279, "formula_len_all": 332, "formula_len_all_1": 272, "len_all": 144823, "len_all_1": 62604, "len_abs": 1335, "len_title": 104, "len_sents": 36020, "len_sents_1": 27401, "n_sents": 296, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1346, "title": "Glad: A Streaming Scene Generator for Autonomous Driving", "abs": "The generation and simulation of diverse real-world scenes have significant application value in the field of autonomous driving, especially for the corner cases. Recently, researchers have explored employing neural radiance fields or diffusion models to generate novel views or synthetic data under driving scenes. However, these approaches suffer from unseen scenes or restricted video length, thus lacking sufficient adaptability for data generation and simulation. To address these issues, we propose a simple yet effective framework, named Glad, to generate video data in a frame-by-frame style. To ensure the temporal consistency of synthetic video, we introduce a latent variable propagation module, which views the hidden features of previous frame as noise prior and injects it into the latent features of current frame. In addition, we design a streaming data sampler to orderly sample the original image in a video clip at continuous iterations. \nGiven the reference frame, our Glad can be viewed as a streaming simulator by generating the videos for specific scenes. \nExtensive experiments are performed on the widely-used nuScenes dataset. Experimental results demonstrate that our proposed Glad achieves promising performance, serving as a strong baseline for online generation. We will release the source code and models publicly.", "keywords": [ "Video Generation; Autonomous Driving" ], "rating_list": [ 6, 5, 3, 8, 6 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 5, 5 ] }, { "paper_id": "ZEO9ibXr46", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Guangjing Yang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Huahui Yi", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "research assistant" }, { "name": "Junjie Wang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Qicheng Lao", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Assistant Professor" }, { "name": "Wentao Chen", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "", "position": "PhD student" }, { "name": "Xiaohu Wu", "gender": "Male", "institution": "BUPT Shenzhen Research Institute", "country": "CN", "position": "Associate Professor" }, { "name": "Zhouchen Lin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Professor" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 34, "n_ref": 113, "n_ref_all": 140, "n_fig": 6, "n_tab": 9, "L_tab": 4173, "n_element_tab": 497, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2246, "n_element_tab_1": 197, "formula_len_all": 508, "formula_len_all_1": 542, "len_all": 145908, "len_all_1": 56305, "len_abs": 4789, "len_title": 116, "len_sents": 36940, "len_sents_1": 24923, "n_sents": 259, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1626, "title": "MLAE: Masked LoRA Experts for Visual Parameter-Efficient Fine-Tuning", "abs": "In response to the challenges posed by the extensive parameter updates required for full fine-tuning of large-scale pre-trained models, parameter-efficient fine-tuning (PEFT) methods, exemplified by Low-Rank Adaptation (LoRA), have emerged. LoRA simplifies the fine-tuning process but may still struggle with a certain level of redundancy in low-rank matrices and limited effectiveness from merely increasing their rank. To address these issues, a natural idea is to enhance the independence and diversity of the learning process for the low-rank matrices. Therefore, we propose Masked LoRA Experts (MLAE), an innovative approach that applies the concept of masking to visual PEFT. Our method incorporates a cellular decomposition strategy that transforms a low-rank matrix into independent rank-1 submatrices, or \"experts\", thus enhancing independence. Additionally, we introduce a binary mask matrix that selectively activates these experts during training to promote more diverse and anisotropic learning, based on expert-level dropout strategies. Our investigations reveal that this selective activation not only enhances performance but also fosters a more diverse acquisition of knowledge with a marked decrease in parameter similarity among MLAE, significantly boosting the quality of the model while barely increasing the parameter count. Remarkably, MLAE achieves new state-of-the-art (SOTA) performance with an average accuracy score of 78.8\\% on the VTAB-1k benchmark and 90.9\\% on the FGVC benchmark, surpassing the previous SOTA method by an average of 0.8\\% on both benchmarks with approximately half parameters.", "keywords": [ "Visual Parameter-Efficient Fine-Tuning", "Pre-trained vision foundation models", "Masking" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "ZE6lrLvATd", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Hannah Lawrence", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "S\u00e9kou-Oumar Kaba", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "Vasco Portilheiro", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Yan Zhang", "gender": "Male", "institution": "Samsung - SAIT AI Lab, Montreal", "country": "CA", "position": "Researcher" } ], "n_formula": 67, "n_formula_1": 40, "n_ref_uni": 54, "n_ref": 107, "n_ref_all": 149, "n_fig": 10, "n_tab": 5, "L_tab": 1443, "n_element_tab": 129, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 847, "n_element_tab_1": 62, "formula_len_all": 25670, "formula_len_all_1": 22531, "len_all": 296245, "len_all_1": 130390, "len_abs": 1411, "len_title": 115, "len_sents": 94842, "len_sents_1": 53883, "n_sents": 688, "n_sents_1": 300, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1422, "title": "Improving Equivariant Networks with Probabilistic Symmetry Breaking", "abs": "Equivariance encodes known symmetries into neural networks, often enhancing generalization. However, equivariant networks cannot *break* symmetries: the output of an equivariant network must, by definition, have at least the same self-symmetries as its input. This poses an important problem, both (1) for prediction tasks on domains where self-symmetries are common, and (2) for generative models, which must break symmetries in order to reconstruct from highly symmetric latent spaces. This fundamental limitation can in fact be addressed by considering *equivariant conditional distributions*, instead of equivariant functions. We therefore present novel theoretical results that establish necessary and sufficient conditions for representing such distributions. Concretely, this representation provides a practical framework for breaking symmetries in any equivariant network via randomized canonicalization. Our method, SymPE (Symmetry-breaking Positional Encodings), admits a simple interpretation in terms of positional encodings. This approach expands the representational power of equivariant networks while retaining the inductive bias of symmetry, which we justify through generalization bounds. Experimental results demonstrate that SymPE significantly improves performance of group-equivariant and graph neural networks across diffusion models for graphs, graph autoencoders, and lattice spin system modeling.", "keywords": [ "equivariance", "symmetry", "symmetry-breaking", "canonicalization", "graphs", "GNNs" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "ZDoaLbOFaP", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Andrea Cavallo", "gender": "Male", "institution": "Delft University of Technology", "country": "", "position": "PhD student" }, { "name": "Elvin Isufi", "gender": "unknown", "institution": "Delft University of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Zhan Gao", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" } ], "n_formula": 72, "n_formula_1": 17, "n_ref_uni": 58, "n_ref": 157, "n_ref_all": 182, "n_fig": 7, "n_tab": 2, "L_tab": 189, "n_element_tab": 22, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 985, "n_element_tab_1": 36, "formula_len_all": 5138, "formula_len_all_1": 1547, "len_all": 225054, "len_all_1": 80370, "len_abs": 1367, "len_title": 81, "len_sents": 66502, "len_sents_1": 34869, "n_sents": 581, "n_sents_1": 276, "venue": "ICLR2025", "decision": "Reject", "L_title": 33, "L_abs": 1372, "title": "Sparse Covariance Neural Networks", "abs": "Covariance Neural Networks (VNNs) perform graph convolutions on the covariance matrix of tabular data and achieve success in a variety of applications. However, the empirical covariance matrix on which the VNNs operate may contain many spurious correlations, making VNNs\u2019 performance inconsistent due to these noisy estimates and decreasing their computational efficiency. To tackle this issue, we put forth Sparse coVariance Neural Networks (S-VNNs), a framework that applies sparsification techniques on the sample covariance matrix before convolution. When the true covariance matrix is sparse, we propose hard and soft thresholding to improve covariance estimation and reduce computational cost. Instead, when the true covariance is dense, we propose stochastic sparsification where data correlations are dropped in probability according to principled strategies. We show that S-VNNs are more stable than nominal VNNs as well as sparse principal component analysis. By analyzing the impact of sparsification on their behavior, we provide novel connections between S-VNN stability and data distribution. We support our theoretical findings with experimental results on various application scenarios, ranging from brain data to human action recognition, and show an improved task performance, stability, and computational efficiency of S-VNNs compared with nominal VNNs.", "keywords": [ "Principal Component Analysis", "Stability property", "Covariance Neural Networks", "Graph sparsification" ], "rating_list": [ 5, 3, 3, 1 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 1 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "ZDoN4W5s8d", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Andrew Magnuson", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" }, { "name": "Franziska Boenisch", "gender": "Female", "institution": "CISPA", "country": "DE", "position": "Assistant Professor" }, { "name": "Mohammad Yaghini", "gender": "Male", "institution": "University of Toronto, Vector Institute", "country": "CA", "position": "PhD student" }, { "name": "Nicolas Papernot", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "Assistant Professor" }, { "name": "Patty Liu", "gender": "Female", "institution": "University of Toronto", "country": "", "position": "Undergrad student" } ], "n_formula": 39, "n_formula_1": 14, "n_ref_uni": 52, "n_ref": 70, "n_ref_all": 113, "n_fig": 14, "n_tab": 6, "L_tab": 711, "n_element_tab": 85, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 580, "n_element_tab_1": 36, "formula_len_all": 3485, "formula_len_all_1": 1218, "len_all": 222864, "len_all_1": 84472, "len_abs": 1381, "len_title": 162, "len_sents": 75686, "len_sents_1": 37577, "n_sents": 615, "n_sents_1": 293, "venue": "ICLR2025", "decision": "Reject", "L_title": 114, "L_abs": 1388, "title": "Lossgate: Incomplete Information and Misaligned Incentives Hinder Regulation of Societal Risks in Machine Learning", "abs": "Regulators seek to curb the societal risks of machine learning; a common aim is to protect the public from excessive privacy violations or bias in models. In the status quo, regulators and companies independently evaluate societal risk. We find that discrepancies in these evaluations can be either a detriment or an advantage for companies. To abide by regulation, a company needs to conservatively evaluate risk: it should train its model such that risk remains below the acceptable threshold-even if the regulator's evaluation returns higher risk measurements. This decreases model utility (up to 8%, in our experiments). Conversely, when the regulator's measurements are consistently lower than theirs, we find that a company can behave strategically and game regulation to train more accurate models. We call this Lossgate, an allusion to Dieselgate in environmental regulation: Volkswagen produced cars that limited their emissions when being subjected to a regulator's emissions measurement. To model incomplete information and the misaligned incentives that explain Lossgate, we leverage game theory. We obtain SpecGame, a model for regulator-company interactions which allows us to estimate the excessive risk that results from the strategic behavior observed in Lossgate. We show Lossgate costs 70\u201396% higher compared to collaborative regulation in the sum cost for all players.", "keywords": [ "ml regulation", "fairness", "privacy" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 1, 1, 1 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "ZDaI3aSDTF", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Jack Cole", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "PhD student" }, { "name": "Ke Li", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "Associate Professor" }, { "name": "HENG YANG", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 43, "n_ref": 61, "n_ref_all": 65, "n_fig": 3, "n_tab": 3, "L_tab": 916, "n_element_tab": 155, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 676, "n_element_tab_1": 78, "formula_len_all": 120, "formula_len_all_1": 120, "len_all": 117156, "len_all_1": 60646, "len_abs": 1907, "len_title": 156, "len_sents": 33698, "len_sents_1": 29919, "n_sents": 219, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1918, "title": "Data Augmentation via Genomic Foundation Models for Pseudoknot-Inclusive RNA Secondary Structure Prediction", "abs": "Rapid advancements in genomic foundation models (GFMs) have delivered a series of breakthroughs across a diverse set of tasks for RNA, however RNA Secondary Structure Prediction (SSP) remains a pivotal task in computational biology. Despite achieving breakthroughs in pseudoknot-free SSP, where state-of-the-art models can achieve above 80% macro-F1, performance on the pseudoknot-inclusive problem remains stagnate, with previous methods achieving below 50% macro-F1 on all three of our test-sets. This is due to a variety of challenges: a ginormous search space that limits heuristic performance, the major class imbalance problem that limits the usual classification methods, and the inherent lack of data that limits deep learning methods. Further data acquisition is implausible due to requiring extensive biological resources and being associated with a high cost. \nIn this work, we propose a novel approach to enhance RNA secondary structure prediction by implementing a novel data augmentation technique, specifically designed for the pseudoknot-inclusive SSP problem. Our method leverages masked language modelling (MLM) with a surrogate model to produce accurate and useful data augmentations, and we further utilise uncertainty quantification strategies to identify areas within the dataset where augmentation is most effective - thereby helping to mitigate the class imbalance problem, and further improving on the generalisability of the models. We further extend three GFMs, and fine-tune them using the augmented datasets to demonstrate the efficacy and high performance of the models.\nNotably, the newly extended and augmented models achieve state-of-the-art performance, achieving over 89% F1 on RNAStrAlign, and over 66% F1 on bpRNA test sets respectively. We therefore highlight the effectiveness of data augmentation for genomic data, and release our code and datasets to assist future researchers.", "keywords": [ "Genomic Foundation Models", "Data Augmentation", "RNA", "Secondary Structure Prediction", "Pseudoknot" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 1, 1, 3 ], "presentation_list": [ 1, 3, 2 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "ZClm0YbcXP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Gaowen Liu", "gender": "Female", "institution": "Cisco Systems", "country": "US", "position": "Researcher" }, { "name": "Haomin Zhuang", "gender": "unknown", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Jinghan Jia", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Kehan Guo", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Sijia Liu", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Assistant Professor" }, { "name": "Xiangliang Zhang", "gender": "Female", "institution": "University of Notre Dame", "country": "US", "position": "Associate Professor" }, { "name": "Yihua Zhang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 55, "n_ref": 168, "n_ref_all": 184, "n_fig": 7, "n_tab": 9, "L_tab": 2388, "n_element_tab": 150, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 1608, "n_element_tab_1": 108, "formula_len_all": 467, "formula_len_all_1": 465, "len_all": 178833, "len_all_1": 73751, "len_abs": 1959, "len_title": 112, "len_sents": 45553, "len_sents_1": 34581, "n_sents": 316, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1846, "title": "UOE: Unlearning One Expert is Enough for Mixture-of-Experts LLMs", "abs": "Recent advancements in large language model (LLM) unlearning have shown remarkable success in removing unwanted data-model influences while preserving the model's utility for legitimate knowledge. However, despite these strides, sparse Mixture-of-Experts (MoE) LLMs--a key subset of the LLM family--have received little attention and remain largely unexplored in the context of unlearning. As MoE LLMs are celebrated for their exceptional performance and highly efficient inference processes, we ask: How can unlearning be performed effectively and efficiently on MoE LLMs? And will traditional unlearning methods be applicable to MoE architectures? Our pilot study shows that the dynamic routing nature of MoE LLMs introduces unique challenges, leading to substantial utility drops when existing unlearning methods are applied. Specifically, unlearning disrupts the router's expert selection, causing significant selection shift from the most unlearning target-related experts to irrelevant ones. As a result, more experts than necessary are affected, leading to excessive forgetting and loss of control over which knowledge is erased. To address this, we propose a novel single-expert unlearning framework, referred to as UOE, for MoE LLMs. Through expert attribution, unlearning is concentrated on the most actively engaged expert for the specified knowledge. Concurrently, an anchor loss is applied to the router to stabilize the active state of this targeted expert, ensuring focused and controlled unlearning that preserves model utility. The proposed UOE framework is also compatible with various unlearning algorithms. Extensive experiments demonstrate that UOE enhances both forget quality up to 5\\% and model utility by 35\\% on MoE LLMs across various benchmarks, LLM architectures, while only unlearning 0.06\\% of the model parameters.", "keywords": [ "Machine Unlearning", "Mixture-of-Expert", "Large Language Model" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 2 ] }, { "paper_id": "ZCPtUkkkUU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chen Tang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Han Yu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jiacheng Jiang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Qun Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Wenwu Zhu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yuan Meng", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhi Wang", "gender": "Male", "institution": "SIGS, Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 43, "n_ref": 93, "n_ref_all": 113, "n_fig": 6, "n_tab": 7, "L_tab": 5256, "n_element_tab": 329, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 5278, "n_element_tab_1": 328, "formula_len_all": 186, "formula_len_all_1": 186, "len_all": 118988, "len_all_1": 62082, "len_abs": 1958, "len_title": 126, "len_sents": 27305, "len_sents_1": 25149, "n_sents": 187, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1973, "title": "GAQAT: Gradient-Adaptive Quantization-Aware Training for Domain Generalization", "abs": "Research on loss surface geometry, such as Sharpness-Aware Minimization (SAM), shows that flatter minima improve generalization. Recent studies further reveal that flatter minima can also reduce the domain generalization (DG) gap. However, existing flatness-based DG techniques predominantly operate within a full-precision training process, which is impractical for deployment on resource-constrained edge devices that typically rely on lower bit-width representations (e.g., 4 bits, 3 bits). Consequently, low-precision quantization-aware training is critical for optimizing these techniques in real-world applications.\nIn this paper, we observe a significant degradation in performance when applying state-of-the-art DG-SAM methods to quantized models, suggesting that current approaches fail to preserve generalizability during the low-precision training process. To address this limitation, we propose a novel Gradient-Adaptive Quantization-Aware Training (GAQAT) framework for DG. \nOur approach begins by identifying the scale-gradient conflict problem in low-precision quantization, where the task loss and smoothness loss induce conflicting gradients for the scaling factors of quantizers, with certain layers exhibiting opposing gradient directions. This conflict renders the optimization of quantized weights highly unstable. To mitigate this, we further introduce a mechanism to quantify gradient inconsistencies and selectively freeze the gradients of scaling factors, thereby stabilizing the training process and enhancing out-of-domain generalization.\nExtensive experiments validate the effectiveness of the proposed GAQAT framework. On PACS, both 3-bit and 4-bit exceed directly integrating DG and QAT by up to 4.5\\%. On DomainNet, our 4-bit results deliver nearly lossless performance compared to the full-precision model, while achieving improvements of up to 1.39\\% and 1.06\\% over the SOTA QAT baseline for 4-bit and 3-bit quantized models, respectively.", "keywords": [ "Model Quantization; Domain Generation; Sharpness-Aware Minimization" ], "rating_list": [ 3, 5, 3, 3, 8 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 2, 3 ], "contribution_list": [ 2, 3, 1, 1, 3 ], "confidence_list": [ 4, 3, 4, 5, 4 ] }, { "paper_id": "ZCOwwRAaEl", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Jaewon Chu", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" }, { "name": "Jinyoung Park", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" }, { "name": "Minseo Yoon", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Undergrad student" }, { "name": "Seunghun Lee", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" }, { "name": "Hyunwoo Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 21, "n_formula_1": 20, "n_ref_uni": 40, "n_ref": 83, "n_ref_all": 123, "n_fig": 13, "n_tab": 3, "L_tab": 2217, "n_element_tab": 83, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 966, "formula_len_all_1": 975, "len_all": 149666, "len_all_1": 63122, "len_abs": 1445, "len_title": 113, "len_sents": 49226, "len_sents_1": 28918, "n_sents": 379, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 65, "L_abs": 1466, "title": "Latent Bayesian Optimization via Autoregressive Normalizing Flows", "abs": "Bayesian Optimization (BO) has been recognized for its effectiveness in optimizing expensive and complex objective functions.\nRecent advancements in Latent Bayesian Optimization (LBO) have shown promise by integrating generative models such as variational autoencoders (VAEs) to manage the complexity of high-dimensional and structured data spaces.\nHowever, existing LBO approaches often suffer from the value discrepancy problem, which arises from the reconstruction gap between latent and input spaces.\nThis value discrepancy problem propagates errors throughout the optimization process, which induces suboptimal optimization outcomes.\nTo address this issue, we propose a Normalizing Flow-based Bayesian Optimization (NF-BO), which utilizes normalizing flow as a generative model to establish accurate and one-to-one mappings between latent and input spaces.\nTo deal with sequence-based inputs, we introduce SeqFlow, an autoregressive sequence-specialized normalizing flow model designed to maintain one-to-one mappings between the input and latent spaces. \nMoreover, we develop a token-level adaptive candidate sampling strategy that dynamically adjusts the exploration probability of each token based on the token-level importance in the optimization process.\nThrough extensive experiments, our NF-BO method demonstrates superior performance in molecule generation tasks, significantly outperforming traditional optimization methods and existing LBO approaches.", "keywords": [ "Bayesian optimization", "normalizing flow" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 4, 4, 4, 3 ], "confidence_list": [ 3, 5, 4, 5 ] }, { "paper_id": "ZC9KpPXgDI", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Yingyu Liang", "gender": "Male", "institution": "University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Yufa Zhou", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "MS student" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Zhenmei Shi", "gender": "Male", "institution": "Voyage AI", "country": "US", "position": "Researcher" } ], "n_formula": 96, "n_formula_1": 28, "n_ref_uni": 69, "n_ref": 162, "n_ref_all": 182, "n_fig": 1, "n_tab": 3, "L_tab": 480, "n_element_tab": 42, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 728, "n_element_tab_1": 19, "formula_len_all": 9306, "formula_len_all_1": 2853, "len_all": 325789, "len_all_1": 74954, "len_abs": 1486, "len_title": 111, "len_sents": 70610, "len_sents_1": 26988, "n_sents": 760, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1614, "title": "Differential Privacy of Cross-Attention with Provable Guarantee", "abs": "Cross-attention has become a fundamental module nowadays in many important artificial intelligence applications, e.g.,\nretrieval-augmented generation (RAG), system prompt, guided stable diffusion, and many more. \nEnsuring cross-attention privacy is crucial and urgently needed because its key and value matrices may contain sensitive information about model providers and their users.\nIn this work, we design a novel differential privacy (DP) data structure to address the privacy security of cross-attention with a theoretical guarantee.\nIn detail, let $n$ be the input token length of system prompt/RAG data, $d$ be the feature dimension, \n$R$ be the maximum value of the query and key matrices, $R_w$ be the maximum value of the value matrix, and $r,s,\\epsilon_s$ be parameters of polynomial kernel methods. \nThen, our data structure requires $\\widetilde{O}(ndr^2)$ memory consumption with $\\widetilde{O}(ndr^2)$ initialization time complexity and $\\widetilde{O}(d r^2)$ query time complexity for a single token query.\nIn addition, our data structure can guarantee that the process of answering user query satisfies $(\\epsilon, \\delta)$-DP with \n$\\widetilde{O}((1-\\epsilon_s)^{-1} n^{-1} \\epsilon^{-1} R^{2s} R_w r^2)$ additive error and $2\\epsilon_s/(1-\\epsilon_s)$ relative errorbetween our output and the true answer.\nFurthermore, our result is robust to adaptive queries in which users can intentionally attack the cross-attention system. \nTo our knowledge, this is the first work to provide DP for cross-attention and is promising to inspire more privacy algorithm design in large generative models (LGMs).", "keywords": [ "Differential Privacy", "Cross-Attention", "Provable Guarantee" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "ZC0wgCabT2", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Guorong Wu", "gender": "Male", "institution": "University of North Carolina, Chapel Hill", "country": "US", "position": "Associate Professor" }, { "name": "Joonhyuk Park", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Yujee Song", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Donghyun LEE", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Won Kim", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 24, "n_formula_1": 21, "n_ref_uni": 43, "n_ref": 84, "n_ref_all": 104, "n_fig": 9, "n_tab": 5, "L_tab": 2595, "n_element_tab": 139, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1746, "n_element_tab_1": 108, "formula_len_all": 2264, "formula_len_all_1": 2116, "len_all": 129506, "len_all_1": 65119, "len_abs": 1464, "len_title": 109, "len_sents": 41724, "len_sents_1": 29154, "n_sents": 286, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1468, "title": "Topology-aware Graph Diffusion Model with Persistent Homology", "abs": "Generating realistic graphs presents challenges in estimating accurate distribution of graphs in an embedding space while preserving structural characteristics such as topology. However, existing graph generation methods primarily focus on approximating the joint distribution of graph nodes and edges, overlooking topology-wise similarity hindering accurate representation of global graph structures such as connected components and loops. To address this issue, we propose a topology-aware diffusion-based graph generation method that aims to closely resemble the structural characteristics of the original graph by leveraging persistent homology from topological data analysis (TDA). Specifically, we suggest a novel loss function, Persistence Diagram Matching (PDM) loss, which ensures the generated graphs to closely match the topology of the original graphs, enhancing their fidelity and preserving essential homological properties. Also, we introduce a novel topology-aware attention to enhance the self-attention module in the denoising network. Through comprehensive experiments, we demonstrate the effectiveness of our approach not only by exhibiting high generation performance across various metrics, but also by demonstrating a closer alignment with the distribution of topological features observed in the original graphs. In addition, application to real brain network data showcases its versatility and potential for complex and real graph application.", "keywords": [ "Graph Generation", "Diffusion", "Topology", "Brain Network" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "ZBlfjXubgG", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Agrin Hilmkil", "gender": "Male", "institution": "Latent Labs", "country": "GB", "position": "Researcher" }, { "name": "James Hensman", "gender": "unknown", "institution": "Microsoft Research", "country": "GB", "position": "Principal Researcher" }, { "name": "Maximilian Croci", "gender": "unknown", "institution": "Microsoft AI", "country": "GB", "position": "Member of Technical Staff" }, { "name": "Tycho F.A. van der Ouderaa", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 13, "n_ref_uni": 22, "n_ref": 59, "n_ref_all": 76, "n_fig": 6, "n_tab": 9, "L_tab": 3326, "n_element_tab": 510, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 466, "n_element_tab_1": 78, "formula_len_all": 2551, "formula_len_all_1": 740, "len_all": 120658, "len_all_1": 59917, "len_abs": 1663, "len_title": 84, "len_sents": 37816, "len_sents_1": 27314, "n_sents": 314, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 1483, "title": "Pyramid Vector Quantization for LLMs", "abs": "Recent works on compression of large language models (LLM) using quantization considered reparameterizing the architecture such that weights are distributed on the sphere. This demonstratively improves the ability to quantize by increasing the mathematical notion of coherence, resulting in fewer weight outliers without affecting the network output. In this work, we aim to further exploit this spherical geometry of the weights when performing quantization by considering \\textit{Pyramid Vector Quantization} (PVQ) for large language models. Arranging points evenly on the sphere is notoriously difficult, especially in high dimensions, and in case approximate solutions exists, representing points explicitly in a codebook is typically not feasible due to its additional memory cost. Instead, PVQ uses a fixed integer lattice on the sphere by projecting points onto the 1-sphere, which allows for efficient encoding and decoding without requiring an explicit codebook in memory. To obtain a practical algorithm, we propose to combine PVQ with scale quantization for which we derive theoretically optimal quantizations, under empirically verified assumptions. Further, we extend pyramid vector quantization to utilise Hessian information to minimize quantization error under expected feature activations, instead of only relying on weight magnitudes. Experimentally, we achieve state-of-the-art quantization performance in terms of bits per weight on various large language models.", "keywords": [ "quantization large language model llm pyramid vector quantization pvq" ], "rating_list": [ 5, 1, 6, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 1, 2, 1 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "ZBL26FX0FT", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Gillian Dobbie", "gender": "Female", "institution": "University of Auckland", "country": "NZ", "position": "Full Professor" }, { "name": "Katerina Taskova", "gender": "unknown", "institution": "University of Auckland", "country": "NZ", "position": "Assistant Professor" }, { "name": "Sandra G\u00f3mez-G\u00e1lvez", "gender": "Female", "institution": "University of Auckland", "country": "NZ", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 5, "n_ref_uni": 33, "n_ref": 127, "n_ref_all": 148, "n_fig": 10, "n_tab": 7, "L_tab": 8188, "n_element_tab": 801, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 590, "n_element_tab_1": 34, "formula_len_all": 2209, "formula_len_all_1": 323, "len_all": 166849, "len_all_1": 60408, "len_abs": 2109, "len_title": 114, "len_sents": 53386, "len_sents_1": 30483, "n_sents": 366, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 2154, "title": "Socrates Loss for training ad-hoc calibrated selective classifiers", "abs": "Model reliability is paramount for critical real-world applications. To enhance reliability, it is essential to quantify uncertainty in model predictions, as achieved through Confidence Calibration and Selective Classification. Confidence Calibration ensures prediction confidences accurately reflect the actual likelihood of correctness, while Selective Classification allows a model to abstain from making predictions when uncertain. Although related, existing methods address each aspect separately, or both through post-hoc approaches. Only one method, Confidence-aware Contrastive Learning for Selective Classification (CCL-SC), combines both in an ad-hoc manner. Despite being a powerful calibrator, CCL-SC has some drawbacks, including the absence of an additional unknown class, the use of two different losses (detrimental for calibration), and its cumbersome implementation. In the pursuit of reliable models and motivated by the idea of creating an ad-hoc calibrated selective classifier with an unknown class, we first empirically analyze the Self-Adaptive Training (SAT) method, a leading approach in ad-hoc selective classification. We identify that while SAT excels in selective classification, it falls short in confidence calibration, especially when training for a small number of epochs (e.g., <=100). To address this, we introduce an original approach that uses an unknown class and a unique novel loss, Socrates loss, which serves as a classifier and a calibrator with a unified optimization goal. This approach mitigates overfitting and ensures theoretically well-calibrated predictions across all epochs, addressing the drawbacks of both CCL-SC and SAT, without the need for post-hoc processing or additional data. We integrate our approach into the SAT implementation and extend it to provide selective classification and confidence calibration metrics. We show empirically that our approach matches or improves the selective classification error rate of SAT and CCL-SC, while producing well-calibrated models in an ad-hoc manner through the evaluation on 6 image benchmark datasets across two architectures, VGG-16 and ResNet-34.", "keywords": [ "Reliability", "Confidence Calibration", "Selective Classification", "Neural Networks" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 1, 1 ], "presentation_list": [ 1, 2, 1, 1 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "ZBH4fqQwJQ", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Le Wan", "gender": "unknown", "institution": "Jilin University", "country": "", "position": "MS student" }, { "name": "Ruofeng Yang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Shuai Li", "gender": "Female", "institution": "John Hopcroft Center, Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yaqing Zhang", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" } ], "n_formula": 25, "n_formula_1": 19, "n_ref_uni": 31, "n_ref": 77, "n_ref_all": 101, "n_fig": 7, "n_tab": 1, "L_tab": 292, "n_element_tab": 51, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 292, "n_element_tab_1": 51, "formula_len_all": 2389, "formula_len_all_1": 1956, "len_all": 118811, "len_all_1": 66047, "len_abs": 1489, "len_title": 154, "len_sents": 33392, "len_sents_1": 28055, "n_sents": 271, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1495, "title": "Contrastive guidance and feedback: A Suitable way to improve 3D Consistency of Multi-view Diffusion Model", "abs": "Recently, diffusion models have shown potential in 3D generation tasks, and the novel view synthesis (NVS) task, a bridge between 2D and 3D generation, has received great attention. The goal of the NVS task is to generate multi-view images from reference images, and the core challenge is to maintain the 3D consistency between different view images.\nRecent works construct large 3D consistency multi-view image datasets and utilize the supervised fine-tuning (SFT) method to improve the 3D consistency. However, the SFT method suffers from the distribution shift, data inefficient problems, and lacks theoretical insight. To solve these problems, we discuss how to provide a suitable direction to the multi-view models and achieve better performance. More specifically, we first analyze the training-free guidance-based method and prove that contrastive guidance, which contains ground-truth and generated samples, can provide the right direction to improve 3D consistency. Based on the theoretical insight, we further design a contrastive 3D consistency metric and use it as the feedback in the following phase. To avoid the distribution shift problem, we use direct preference optimization (DPO) to fine-tune the multi-view diffusion models. Through qualitative and quantitative experiments, we demonstrate that after the fine-tuning phase with the above method, the 3D consistency of the multi-view images is significantly improved and achieves better performance compared to the SFT method.", "keywords": [ "Contrastive guidance", "Multi-view diffusion model" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "ZAyuwJYN8N", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Li cheng", "gender": "Not Specified", "institution": "University of Alberta", "country": "CA", "position": "Full Professor" }, { "name": "Muhammad Gohar Javed", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "MS student" }, { "name": "Xingyu Li", "gender": "unknown", "institution": "University of Alberta", "country": "CA", "position": "Assistant Professor" }, { "name": "chuan guo", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 36, "n_ref": 94, "n_ref_all": 121, "n_fig": 13, "n_tab": 9, "L_tab": 1740, "n_element_tab": 159, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1562, "n_element_tab_1": 91, "formula_len_all": 652, "formula_len_all_1": 471, "len_all": 151211, "len_all_1": 67619, "len_abs": 1575, "len_title": 124, "len_sents": 48265, "len_sents_1": 31571, "n_sents": 314, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1597, "title": "InterMask: 3D Human Interaction Generation via Collaborative Masked Modelling", "abs": "Generating realistic 3D human-human interactions from textual descriptions remains a challenging task. Existing approaches, typically based on diffusion models, often generate unnatural and unrealistic results. In this work, we introduce *InterMask*, a novel framework for generating human interactions using collaborative masked modeling in discrete space. InterMask first employs a VQ-VAE to transform each motion sequence into a 2D discrete motion token map. Unlike traditional 1D VQ token maps, it better preserves fine-grained spatio-temporal details and promotes *spatial awareness* within each token. Building on this representation, InterMask utilizes a generative masked modeling framework to collaboratively model the tokens of two interacting individuals. This is achieved by employing a transformer architecture specifically designed to capture complex spatio-temporal interdependencies. During training, it randomly masks the motion tokens of both individuals and learns to predict them. In inference, starting from fully masked sequences, it progressively fills in the tokens for both individuals. With its enhanced motion representation, dedicated architecture, and effective learning strategy, InterMask achieves state-of-the-art results, producing high-fidelity and diverse human interactions. It outperforms previous methods, achieving an FID of $5.154$ (vs $5.535$ for in2IN) on the InterHuman dataset and $0.399$ (vs $5.207$ for InterGen) on the InterX dataset. Additionally, InterMask seamlessly supports reaction generation without the need for model redesign or fine-tuning.", "keywords": [ "Motion Synthesis", "Human Interaction Generation", "Masked Generative Transformer", "Text-driven Generation", "Vector Quantized VAE" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "ZAx5DxAucB", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Boyang Zheng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Chao Dong", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Jinjin Gu", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "Instructor" }, { "name": "Shijun Li", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Undergrad student" } ], "n_formula": 8, "n_formula_1": 1, "n_ref_uni": 44, "n_ref": 101, "n_ref_all": 122, "n_fig": 13, "n_tab": 8, "L_tab": 1558, "n_element_tab": 103, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1167, "n_element_tab_1": 56, "formula_len_all": 1097, "formula_len_all_1": 28, "len_all": 179203, "len_all_1": 58927, "len_abs": 927, "len_title": 111, "len_sents": 48737, "len_sents_1": 26017, "n_sents": 434, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 934, "title": "LM4LV: A Frozen Large Language Model for Low-level Vision Tasks", "abs": "The success of large language models (LLMs) has fostered a new research trend of multi-modality large language models (MLLMs), which changes the paradigm of various fields in computer vision. Though MLLMs have shown promising results in numerous vision-language tasks such as VQA and text-to-image, no work has demonstrated how low-level vision tasks can benefit from MLLMs. We find that most current MLLMs are blind to low-level features due to their design of vision modules, and thus are inherently incapable of solving low-level vision tasks. In this work, we propose **LM4LV**, a framework that enables a FROZEN LLM to solve a range of low-level vision tasks without any multi-modal data or prior. This showcases the LLM's strong potential in low-level vision and bridges the gap between MLLMs and low-level vision tasks. We hope that this work can inspire new perspectives on LLMs and a deeper understanding of their mechanisms.", "keywords": [ "Low-level vision", "Large Language Model", "Self-supervised Learning" ], "rating_list": [ 5, 3, 6, 5, 6 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4, 3 ] }, { "paper_id": "ZACnAv9ZTY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Lu Chen", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Quanshi Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "", "position": "Associate Professor" }, { "name": "ShuaiZhao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Postdoc" }, { "name": "Yaohui Jin", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Yixing Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Yuxuan Huang", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Zilong Zheng", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" } ], "n_formula": 45, "n_formula_1": 10, "n_ref_uni": 31, "n_ref": 83, "n_ref_all": 129, "n_fig": 16, "n_tab": 7, "L_tab": 1965, "n_element_tab": 114, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 342, "n_element_tab_1": 14, "formula_len_all": 9168, "formula_len_all_1": 757, "len_all": 222245, "len_all_1": 73824, "len_abs": 944, "len_title": 147, "len_sents": 72743, "len_sents_1": 36955, "n_sents": 530, "n_sents_1": 244, "venue": "ICLR2025", "decision": "withdraw", "L_title": 99, "L_abs": 948, "title": "Alignment Between the Decision-Making Logic of LLMs and Human Cognition: A Case Study on Legal LLMs", "abs": "This paper presents a method to evaluate the alignment between the decision-making logic of Large Language Models (LLMs) and human cognition in a case study on legal LLMs. Unlike traditional evaluations on language generation results, we propose to evaluate the correctness of the detailed decision-making logic of an LLM behind its seemingly correct outputs, which represents the core challenge for an LLM to earn human trust. To this end, we quantify the interactions encoded by the LLM as primitive decision-making logic, because recent theoretical achievements (Li & Zhang, 2023; Ren et al., 2024) have proven several mathematical guarantees of the faithfulness of the interaction-based explanation. We design a set of metrics to evaluate the detailed decision-making logic of LLMs. Experiments show that even when the language generation results appear correct, a significant portion of the internal inference logic contains notable issues.", "keywords": [ "Explainable AI", "Large Language Model", "Interactions" ], "rating_list": [ 5, 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 3, 2 ], "presentation_list": [ 3, 1, 1, 2, 1 ], "contribution_list": [ 2, 1, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 3, 3 ] }, { "paper_id": "ZACAKudvKW", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Chris Russell", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" }, { "name": "Zihao Fu", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "Postdoc" } ], "n_formula": 94, "n_formula_1": 22, "n_ref_uni": 27, "n_ref": 102, "n_ref_all": 126, "n_fig": 13, "n_tab": 5, "L_tab": 1170, "n_element_tab": 127, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 462, "n_element_tab_1": 13, "formula_len_all": 5605, "formula_len_all_1": 1460, "len_all": 182335, "len_all_1": 79217, "len_abs": 505, "len_title": 109, "len_sents": 61792, "len_sents_1": 38125, "n_sents": 557, "n_sents_1": 292, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 644, "title": "Watermarking for User Identification in Large Language Models", "abs": "We identify a new task for watermarking -- namely the simultaneous identification of text as being automatically generated alongside the identification of the LLM user. \nWe show that a na\u00efve approach that treats a text as artificially generated if a user is correctly identified is prone to problems of false positives arising from multiple hypothesis comparison. \nWe propose a novel approach (Our code is submitted with the supplementary material. We will also open it on Github after the anonymity period.) that retains almost similar rates as the number of users increase. We derive theoretical bounds that support our experimental approach.", "keywords": [ "Large Language model; Watermark" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "Z9Odi09Rv9", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Henry Li", "gender": "unknown", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Jonathan Patsenker", "gender": "unknown", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Myeongseob Ko", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "PhD student" }, { "name": "Ruoxi Jia", "gender": "unknown", "institution": "Virginia Tech", "country": "US", "position": "Assistant Professor" }, { "name": "Yuval Kluger", "gender": "unknown", "institution": "Yale University", "country": "US", "position": "Professor" } ], "n_formula": 38, "n_formula_1": 20, "n_ref_uni": 33, "n_ref": 62, "n_ref_all": 95, "n_fig": 9, "n_tab": 4, "L_tab": 5699, "n_element_tab": 922, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 809, "n_element_tab_1": 152, "formula_len_all": 2711, "formula_len_all_1": 1276, "len_all": 149891, "len_all_1": 55663, "len_abs": 1398, "len_title": 132, "len_sents": 42335, "len_sents_1": 23976, "n_sents": 348, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1436, "title": "Fast and Noise-Robust Diffusion Solvers for Inverse Problems: A Frequentist Approach", "abs": "Diffusion models have been firmly established as principled zero-shot solvers for linear and nonlinear inverse problems, owing to their powerful image prior and ease of formulation as Bayesian posterior samplers. However, many existing solvers struggle in the noisy measurement regime, either overfitting or underfitting to the measurement constraint, resulting in poor sample quality and inconsistent performance across noise levels. Moreover, existing solvers rely on approximating $x_0$ via Tweedie's formula, where an intractable \\textit{conditional} score is replaced by an \\textit{unconditional} score network, introducing a fundamental source of error in the resulting solution. In this work, we propose a novel frequentist's approach to diffusion-based inverse solvers, where each diffusion step can be seen as the maximum likelihood solution to a simple single-parameter conditional likelihood model, derived by an adjusted application of Tweedie's formula to the forward measurement model. We demonstrate that this perspective is not only scalable and fast, but also allows for a noise-aware maximization scheme with a likelihood-based stopping criterion that promotes the proper noise-adapted fit given knowledge of the measurement noise $\\sigma_\\mathbf{y}$. Finally, we demonstrate comparable or improved performance against a wide selection of contemporary inverse solvers across multiple datasets, tasks, and noise levels.", "keywords": [ "diffusion models", "inverse problems", "maximum likelihood" ], "rating_list": [ 8, 5, 3, 3 ], "soundness_list": [ 4, 2, 1, 1 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "Z9N3J7j50k", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Jiecheng Lu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Shihao Yang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Xu Han", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Yan Sun", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 35, "n_ref": 81, "n_ref_all": 100, "n_fig": 7, "n_tab": 14, "L_tab": 12791, "n_element_tab": 948, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 2712, "n_element_tab_1": 208, "formula_len_all": 1795, "formula_len_all_1": 1619, "len_all": 167197, "len_all_1": 65173, "len_abs": 3222, "len_title": 125, "len_sents": 40712, "len_sents_1": 25932, "n_sents": 275, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1302, "title": "Autoregressive Moving-average Attention Mechanism for Time Series Forecasting", "abs": "We propose an Autoregressive (AR) Moving-average (MA) attention structure that can adapt to various linear attention mechanisms, enhancing their ability to capture long-range and local temporal patterns in time series. In this paper, we first demonstrate that, for the time series forecasting (TSF) task, the previously overlooked decoder-only autoregressive Transformer model can achieve results comparable to the best baselines when appropriate tokenization and training methods are applied. Moreover, inspired by the ARMA model from statistics and recent advances in linear attention, we introduce the full ARMA structure into existing autoregressive attention mechanisms. By using an indirect MA weight generation method, we incorporate the MA term while maintaining the time complexity and parameter size of the underlying efficient attention models. We further explore how indirect parameter generation can produce implicit MA weights that align with the modeling requirements for local temporal impacts. Experimental results show that incorporating the ARMA structure consistently improves the performance of various AR attentions on TSF tasks, achieving state-of-the-art results. The code implementation is available at the following link: https://anonymous.4open.science/r/ARMA-attention-3437.", "keywords": [ "Attention", "Transformer", "Autoregressive Moving-average", "Time Series Forecasting" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "Z8dr422vtr", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Masataka Ushiku", "gender": "unknown", "institution": "Hakuhodo Inc.", "country": "JP", "position": "Researcher" }, { "name": "Yuta Natsubori", "gender": "unknown", "institution": "Hakuhodo DY Holdings Inc.", "country": "JP", "position": "Researcher" }, { "name": "Yuta Saito", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 20, "n_ref_uni": 34, "n_ref": 95, "n_ref_all": 118, "n_fig": 9, "n_tab": 4, "L_tab": 1446, "n_element_tab": 83, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5442, "formula_len_all_1": 1870, "len_all": 164991, "len_all_1": 71773, "len_abs": 1433, "len_title": 118, "len_sents": 55474, "len_sents_1": 34537, "n_sents": 358, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1434, "title": "Cross-Domain Off-Policy Evaluation and Learning for Contextual Bandits", "abs": "Off-Policy Evaluation and Learning (OPE/L) in contextual bandits is rapidly gaining popularity in real systems because new policies can be evaluated and learned securely using only historical logged data. However, existing methods in OPE/L cannot handle many challenging but prevalent scenarios such as few-shot data, deterministic logging policies, and new actions. In many applications, such as personalized medicine, content recommendations, education, and advertising, we need to evaluate and learn new policies in the presence of these challenges. Existing methods cannot evaluate and optimize effectively in these situations due to the notorious variance issue or limited exploration in the logged data. To enable OPE/L even under these unsolved challenges, we propose a new problem setup of Cross-Domain OPE/L, where we have access not only to the logged data from the target domain in which the new policy will be implemented but also to logged datasets collected from other domains. This novel formulation is widely applicable because we can often use historical data not only from the target hospital, country, device, or user segment but also from other hospitals, countries, devices, or segments. We develop a new estimator and policy gradient method to solve OPE/L by leveraging both target and source datasets, resulting in substantially enhanced OPE/L in the previously unsolved situations in our empirical evaluations.", "keywords": [ "Off-Policy Evaluation", "Off-Policy Learning", "Importance Weighting", "Cross Domain" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 2, 4, 2 ] }, { "paper_id": "Z8XALH7RLv", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Chao Huang", "gender": "Male", "institution": "University of Southampton", "country": "", "position": "Associate Professor" }, { "name": "Philip Wang", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "Undergrad student" }, { "name": "Qi Zhu", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "Full Professor" }, { "name": "Qingyuan Wu", "gender": "Male", "institution": "University of Southampton", "country": "GB", "position": "PhD student" }, { "name": "Ruochen Jiao", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Simon Sinong Zhan", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Yixuan Wang", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "", "position": "PhD student" } ], "n_formula": 42, "n_formula_1": 17, "n_ref_uni": 69, "n_ref": 118, "n_ref_all": 144, "n_fig": 11, "n_tab": 4, "L_tab": 1625, "n_element_tab": 131, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1484, "n_element_tab_1": 131, "formula_len_all": 4581, "formula_len_all_1": 1121, "len_all": 174265, "len_all_1": 69030, "len_abs": 1032, "len_title": 149, "len_sents": 44038, "len_sents_1": 30818, "n_sents": 338, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 121, "L_abs": 978, "title": "Model-Enhanced Adversarial Inverse Reinforcement Learning with Model Estimation Reward Shaping in Stochastic Environments", "abs": "In this paper, we aim to tackle the limitation of the Adversarial Imitation Learning (AIL) method in stochastic environments where theoretical results cannot hold and performance is degraded. To address this issue, we propose a novel method which infuses the dynamics information into the reward shaping with the theoretical guarantee for the induced optimal policy in the stochastic environments. Incorporating our novel model-enhanced reward shaping, we present a novel Model-based IRL framework, which integrates transition model estimation directly into reward shaping. Furthermore, we provide a comprehensive theoretical analysis of the reward error bound and performance difference bound for our method. The experimental results in MuJoCo benchmarks show that our method can achieve superior performance in stochastic environments and competitive performance in deterministic environments, with significant improvement in sample efficiency, compared to existing baselines.", "keywords": [ "Inverse Reinforcement Learning; Model-based Approach;" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 2 ] }, { "paper_id": "Z8TglKXDWm", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Cheng Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Frederick A Matsen", "gender": "Male", "institution": "Fred Hutchinson Cancer Research Center", "country": "", "position": "Associate Professor" }, { "name": "Harry Richman", "gender": "Male", "institution": "Fred Hutchinson Cancer Research Center", "country": "US", "position": "Postdoc" }, { "name": "Jiansi Gao", "gender": "Male", "institution": "Fred Hutchinson Cancer Research Center", "country": "US", "position": "Postdoc" }, { "name": "Tianyu Xie", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 10, "n_ref_uni": 48, "n_ref": 99, "n_ref_all": 143, "n_fig": 13, "n_tab": 1, "L_tab": 474, "n_element_tab": 81, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1794, "formula_len_all_1": 867, "len_all": 164325, "len_all_1": 70495, "len_abs": 1060, "len_title": 131, "len_sents": 50476, "len_sents_1": 32120, "n_sents": 404, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1064, "title": "PhyloVAE: Unsupervised Learning of Phylogenetic Trees via Variational Autoencoders", "abs": "Learning informative representations of phylogenetic tree structures is essential for analyzing evolutionary relationships. Classical distance-based methods have been widely used to project phylogenetic trees into Euclidean space, but they are often sensitive to the choice of distance metric and may lack sufficient resolution. In this paper, we introduce *phylogenetic variational autoencoders* (PhyloVAEs), an unsupervised learning framework designed for representation learning and generative modeling of tree topologies. Leveraging an efficient encoding mechanism inspired by autoregressive tree topology generation, we develop a deep latent-variable generative model that facilitates fast, parallelized topology generation. PhyloVAE combines this generative model with a collaborative inference model based on learnable topological features, allowing for high-resolution representations of phylogenetic tree samples. Extensive experiments demonstrate PhyloVAE's robust representation learning capabilities and fast generation of phylogenetic tree topologies.", "keywords": [ "Variational Autoencoders; Unsupervised Learning; Phylogenetic Trees; Representation Learning" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "Z8RZrvngm5", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Marco Eckhoff", "gender": "Not Specified", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Markus Reiher", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Full Professor" }, { "name": "Raphael T. Husistein", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 54, "n_ref": 111, "n_ref_all": 133, "n_fig": 1, "n_tab": 8, "L_tab": 4247, "n_element_tab": 489, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2942, "n_element_tab_1": 248, "formula_len_all": 399, "formula_len_all_1": 399, "len_all": 154473, "len_all_1": 74545, "len_abs": 1239, "len_title": 122, "len_sents": 43994, "len_sents_1": 33700, "n_sents": 320, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1247, "title": "NEAR: A Training-Free Pre-Estimator of Machine Learning Model Performance", "abs": "Artificial neural networks have been shown to be state-of-the-art machine learning models in a wide variety of applications, including natural language processing and image recognition. However, building a performant neural network is a laborious task and requires substantial computing power. Neural Architecture Search (NAS) addresses this issue by an automatic selection of the optimal network from a set of potential candidates. While many NAS methods still require training of (some) neural networks, zero-cost proxies promise to identify the optimal network without training. In this work, we propose the zero-cost proxy Network Expressivity by Activation Rank (NEAR). It is based on the effective rank of the pre- and post-activation matrix, i.e., the values of a neural network layer before and after applying its activation function. We demonstrate the cutting-edge correlation between this network score and the model accuracy on NAS-Bench-101 and NATS-Bench-SSS/TSS. In addition, we present a simple approach to estimate the optimal layer sizes in multi-layer perceptrons. Furthermore, we show that this score can be utilized to select hyperparameters such as the activation function and the neural network weight initialization scheme.", "keywords": [ "Network Expressivity by Activation Rank (NEAR)", "Effective Rank", "Neural Architecture Search" ], "rating_list": [ 8, 5, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Z8Mfy0iK4n", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Jeff Z. Pan", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Full Professor" }, { "name": "Ru Li", "gender": "Female", "institution": "Shanxi University", "country": "CN", "position": "Full Professor" }, { "name": "XiaoQi Han", "gender": "unknown", "institution": "Shanxi University", "country": "", "position": "PhD student" }, { "name": "Zhichao Yan", "gender": "Male", "institution": "Shanxi University", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 34, "n_ref": 57, "n_ref_all": 86, "n_fig": 10, "n_tab": 9, "L_tab": 2768, "n_element_tab": 70, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 3547, "n_element_tab_1": 170, "formula_len_all": 303, "formula_len_all_1": 303, "len_all": 140251, "len_all_1": 61632, "len_abs": 1038, "len_title": 158, "len_sents": 34926, "len_sents_1": 26159, "n_sents": 273, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 110, "L_abs": 1069, "title": "Entropy Reveals What You Know: An Entropy-Guided Method for Enhancing the Reliability of Large Language Models", "abs": "While large language models (LLMs) encode vast amounts of knowledge within their parameters for some mainstream entities, factual inconsistencies and untruthfulness in LLMs often lead to unreliable responses and cause significant risks in practical applications.\nThis paper aims to improve model reliability by enhancing consistency in answers to known facts and encouraging refusal to answer for uncertain questions.\nSpecifically, we introduce \\textbf{SREF}, an entropy-guided approach designed to enhance the reliability of language models by incorporating \\textbf{S}elf-\\textbf{REF}erences, models' understanding of rephrasing questions, with inputs.\nWe analyze and reveal the effectiveness of SREF in enhancing model reliability from the perspectives of entropy and KL divergence.\nExtensive experiments on 12 LLMs demonstrate that outputs generated with SREF yield more reliable results, including an average improvement of 16.01\\% over the baselines and a 15.10\\% average improvement in consistency, while also adapting to identify and acknowledge uncertain facts.", "keywords": [ "Reliability", "Large language Model" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 1, 1 ], "contribution_list": [ 3, 2, 1 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "Z85EoYQhCs", "primary_area": "applications to robotics, autonomy, planning", "n_author": 12, "author_info_list": [ { "name": "Ajay Mandlekar", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Jiaojiao Fan", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Linxi Fan", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Ming-Yu Liu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Mingyuan Zhou", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Yashraj Narang", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Yogesh Balaji", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yu Zeng", "gender": "Not Specified", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yuke Zhu", "gender": "Male", "institution": "Computer Science Department, University of Texas, Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Zhaoshuo Li", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Zhendong Wang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Zhenjia Xu", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" } ], "n_formula": 20, "n_formula_1": 14, "n_ref_uni": 28, "n_ref": 100, "n_ref_all": 141, "n_fig": 9, "n_tab": 3, "L_tab": 1598, "n_element_tab": 129, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1346, "n_element_tab_1": 93, "formula_len_all": 1935, "formula_len_all_1": 1339, "len_all": 151786, "len_all_1": 58556, "len_abs": 1570, "len_title": 126, "len_sents": 44558, "len_sents_1": 25423, "n_sents": 366, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1440, "title": "One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation", "abs": "Diffusion models, praised for their success in generative tasks, are increasingly being applied to robotics, demonstrating exceptional performance in behavior cloning. However, their slow generation process stemming from iterative denoising steps poses a challenge for real-time applications in resource-constrained robotics setups and dynamically changing environments.\nIn this paper, we introduce the One-Step Diffusion Policy (OneDP), a novel approach that distills knowledge from pre-trained diffusion policies into a single-step action generator, significantly accelerating response times for robotic control tasks. We ensure the distilled generator closely aligns with the original policy distribution by minimizing the Kullback-Leibler (KL) divergence along the diffusion chain, requiring only 2%-10% additional pre-training cost for convergence. We evaluated OneDP on 6 challenging simulation tasks as well as 4 self-designed real-world tasks using the Franka robot. The results demonstrate that OneDP not only achieves state-of-the-art success rates but also delivers an order-of-magnitude improvement in inference speed, boosting action prediction frequency from 1.5 Hz to 62 Hz, establishing its potential for dynamic and computationally constrained robotic applications. A video demo is provided at https://drive.google.com/file/d/1eIa11gw6DwYKG9CKERy41bjE1ruklRtT/view?usp=sharing, and the code will be publicly available soon.", "keywords": [ "Robotics", "Diffusion Policy", "Distillation", "One-Step Policy" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "Z7aq3djHZw", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Marjan Ghazvininejad", "gender": "unknown", "institution": "Facebook AI Research", "country": "", "position": "Research Scientist" }, { "name": "Pang Wei Koh", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Visiting Research Scientist" }, { "name": "Xiaochuang Han", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Yulia Tsvetkov", "gender": "Female", "institution": "Department of Computer Science, University of Washington", "country": "", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 50, "n_ref": 81, "n_ref_all": 102, "n_fig": 5, "n_tab": 3, "L_tab": 1446, "n_element_tab": 208, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2640, "n_element_tab_1": 264, "formula_len_all": 214, "formula_len_all_1": 214, "len_all": 136336, "len_all_1": 67060, "len_abs": 1507, "len_title": 118, "len_sents": 36038, "len_sents_1": 28423, "n_sents": 274, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1514, "title": "JPEG-LM: LLMs as Image Generators with Canonical Codec Representations", "abs": "Recent work in image and video generation has been adopting the autoregressive LLM architecture due to its generality and potentially easy integration into multi-modal systems. The crux of applying autoregressive training in language generation to visual generation is discretization---representing continuous data like images and videos as discrete tokens. Common methods of discretizing images and videos include modeling raw pixel values, which are prohibitively lengthy, or vector quantization, which requires convoluted pre-hoc training. In this work, we propose to directly model images and videos as compressed files saved on computers via canonical codecs (e.g., JPEG, AVC/H.264). Using the default Llama architecture without any vision-specific modifications, we pretrain JPEG-LM from scratch to generate images (and AVC-LM to generate videos as a proof of concept), by directly outputting compressed file bytes in JPEG and AVC formats. Evaluation of image generation shows that this simple and straightforward approach is more effective than pixel-based modeling and sophisticated vector quantization baselines (on which our method yields a 31% reduction in FID). Our analysis shows that JPEG-LM has an especial advantage over vector quantization models in generating long-tail visual elements. Overall, we show that using canonical codec representations can help lower the barriers between language generation and visual generation, facilitating future research on multi-modal language/image/video LLMs.", "keywords": [ "LLM for visual generation", "codec-based LLMs" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "Z7FLmWFUFo", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "HanfangZhang", "gender": "Female", "institution": "Shandong University", "country": "CN", "position": "MS student" }, { "name": "Shuo Chen", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Siyuan Qi", "gender": "unknown", "institution": "Beijing Institute for General Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Yexin Li", "gender": "Female", "institution": "State Key Laboratory of General Artificial Intelligence, BIGAI", "country": "CN", "position": "Researcher" }, { "name": "pring wong", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "", "position": "Researcher" } ], "n_formula": 53, "n_formula_1": 18, "n_ref_uni": 28, "n_ref": 87, "n_ref_all": 131, "n_fig": 7, "n_tab": 5, "L_tab": 1119, "n_element_tab": 110, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1590, "n_element_tab_1": 76, "formula_len_all": 4627, "formula_len_all_1": 979, "len_all": 153604, "len_all_1": 66666, "len_abs": 1351, "len_title": 154, "len_sents": 42475, "len_sents_1": 28715, "n_sents": 330, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1353, "title": "The Critic as an Explorer: Lightweight and Provably Efficient Exploration for Deep Reinforcement Learning", "abs": "Exploration remains a critical challenge in reinforcement learning (RL), with many existing methods either lacking theoretical guarantees or being computationally impractical for real-world applications. We introduce Litee, a lightweight algorithm that repurposes the value network in standard deep RL algorithms to effectively drive exploration without introducing additional parameters. Litee utilizes linear multi-armed bandit (MAB) techniques, enabling efficient exploration with provable sub-linear regret bounds while preserving the core structure of existing RL algorithms. Litee is simple to implement, requiring only around 10 lines of code. It also substantially reduces computational overhead compared to previous theoretically grounded methods, lowering the complexity from O(n^3) to O(d^3), where n is the number of network parameters and d is the size of the embedding in the value network. Furthermore, we propose Litee+, an extension that adds a small auxiliary network to better handle sparse reward environments, with only a minor increase in parameter count (less than 1%) and additional 10 lines of code. Experiments on the MiniHack suite and MuJoCo demonstrate that Litee and Litee+ empirically outperform state-of-the-art baselines, effectively bridging the gap between theoretical rigor and practical efficiency in RL exploration.", "keywords": [ "Reinforcement learning", "exploration", "embedding" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 2, 5, 4, 4 ] }, { "paper_id": "Z756zcjNcC", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "Di Zhou", "gender": "unknown", "institution": "Tufts University", "country": "US", "position": "PhD student" }, { "name": "Donna Slonim", "gender": "Female", "institution": "Tufts University", "country": "US", "position": "Full Professor" }, { "name": "Hao Zhu", "gender": "Male", "institution": "Tufts University", "country": "US", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 19, "n_ref_uni": 46, "n_ref": 81, "n_ref_all": 104, "n_fig": 11, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1076, "formula_len_all_1": 812, "len_all": 134599, "len_all_1": 59656, "len_abs": 1450, "len_title": 84, "len_sents": 40048, "len_sents_1": 27829, "n_sents": 315, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 1459, "title": "Denoising Diffusion Causal Discovery", "abs": "A common theme across multiple disciplines of science is to understand the underlying dependencies between variables from observational data. Such dependencies are often modeled as Bayesian Network (BNs), which by definition are Directed Acyclic Graphs (DAGs). Recent advancements, such as NOTEARS and DAG-GNN, have focused on formulating continuous DAG constraints and learning DAGs via continuous optimization. However, these methods often have scalability issues and face challenges when applied to real world data. In this paper, we propose Denoising Diffusion Causal Discovery (DDCD), a new learning framework that leverages Denoising Diffusion Probabilistic Models (DDPMs) for causal structural learning. Using the denoising objective, our method allows the model to explore a wider range of noise in the data and effectively captures both linear and nonlinear dependencies. It also has reduced complexity and is more suitable for inference of larger networks. To accommodate potential feedback loops in biological networks, we propose a k-hop acyclicity constraint. Additionally, we suggest using fixed-size bootstrap sampling to ensure similar training performance across varying dataset sizes. Our experiments on synthetic data demonstrate that DDCD achieves consistent competitive performance compared to existing methods while noticeably reducing computation time. We also show that DDCD can generate trustworthy networks from real-world datasets.", "keywords": [ "Denoising Diffusion Models", "Causal Discovery", "Causal Reasoning", "Diffusion Models", "Graph Neural Networks" ], "rating_list": [ 6, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "Z6kVjQAPNq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Bhrij Patel", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Dinesh Manocha", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Professor" }, { "name": "Mengdi Wang", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Associate Professor" }, { "name": "Souradip Chakraborty", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Amrit Bedi", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Wesley Suttle", "gender": "unknown", "institution": "Army Research Laboratory", "country": "US", "position": "Postdoc" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 47, "n_ref": 82, "n_ref_all": 106, "n_fig": 6, "n_tab": 6, "L_tab": 1534, "n_element_tab": 165, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 476, "n_element_tab_1": 73, "formula_len_all": 1233, "formula_len_all_1": 1215, "len_all": 147210, "len_all_1": 67292, "len_abs": 1446, "len_title": 104, "len_sents": 38362, "len_sents_1": 31062, "n_sents": 295, "n_sents_1": 237, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1463, "title": "AIME: AI System Optimization via Multiple LLM Evaluators", "abs": "Text-based AI system optimization typically involves a feedback loop scheme where a \\textit{single} LLM generates an evaluation in natural language of the current output to improve the next iteration's output. However, in this work, we empirically demonstrate that for a practical and complex task (code generation) with multiple criteria to evaluate, utilizing only one LLM evaluator tends to let errors in generated code go undetected, thus leading to incorrect evaluations and ultimately suboptimal test case performance. Motivated by this failure case, we assume there exists an optimal evaluation policy that samples an evaluation between response and ground truth. We then theoretically prove that a linear combination of multiple evaluators can approximate this optimal policy. From this insight, we propose AI system optimization via Multiple LLM Evaluators (AIME). AIME is an evaluation protocol that utilizes multiple LLMs that each independently generate an evaluation on separate criteria and then combine them via concatenation. We provide an extensive empirical study showing AIME outperforming baseline methods in code generation tasks, with up to 62\\% higher error detection rate and up to 16\\% higher success rate than a single LLM evaluation protocol on LeetCodeHard and HumanEval datasets. We also show that the selection of the number of evaluators and which criteria to utilize is non-trivial as it can impact pact success rate by up to 12\\%.", "keywords": [ "AI System Optimization", "Agentic AI", "Large Language Models", "LLM-based Evaluation" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 1, 1, 4, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Z6TQhliDIq", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 1, "author_info_list": [ { "name": "Wenhang Ge", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 40, "n_formula_1": 26, "n_ref_uni": 19, "n_ref": 54, "n_ref_all": 82, "n_fig": 15, "n_tab": 8, "L_tab": 3349, "n_element_tab": 444, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 670, "n_element_tab_1": 43, "formula_len_all": 1791, "formula_len_all_1": 1302, "len_all": 156561, "len_all_1": 54023, "len_abs": 1408, "len_title": 139, "len_sents": 41685, "len_sents_1": 23357, "n_sents": 379, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1424, "title": "Uni-IR: Ambiguity-Reduced Inverse Rendering through a Unified Framework for Glossy Objects", "abs": "Inverse rendering aims to decompose the an image into geometry, materials, and lighting.\n Recently, Neural Radiance Fields (NeRF) based inverse rendering has significantly advanced, bridging the gap between NeRF-based models and conventional rendering engines.\n Existing methods typically adopt a two-stage optimization approach, beginning with volume rendering for geometry reconstruction, followed by physically based rendering (PBR) for materials and lighting estimation.\n However, the inherent ambiguity between materials and lighting during PBR and the suboptimal nature of geometry reconstruction by volume rendering only compromise the outcomes.\n To address these challenges, we introduce Uni-IR, a unified framework that imposes mutual constraints to alleviate ambiguity by integrating volume rendering and physically based rendering.\n Specifically, we employ a physically-based volume rendering (PBVR) approach that incorporates PBR concepts into volume rendering, directly facilitating connections with materials and lighting, in addition to geometry. Both rendering methods are utilized during optimization, imposing mutual constraints and optimizing geometry, materials, and lighting synergistically. Employing a meticulously crafted unified representation for both lighting and materials, Uni-IR achieves high-quality geometry reconstruction, materials and lighting estimation across various object types.", "keywords": [ "Inverse rendering", "physically based rendering" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "Z5nqeTH24j", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Kechen Fang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Lei Han", "gender": "Male", "institution": "Tencent Robotics X", "country": "CN", "position": "Principal Researcher" }, { "name": "Sicheng Zhou", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" }, { "name": "Sijie Cheng", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Tingguang Li", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Yang Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Yangyang Yu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Ye Tian", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Li Bohao", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 61, "n_ref": 133, "n_ref_all": 151, "n_fig": 5, "n_tab": 5, "L_tab": 2271, "n_element_tab": 385, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1462, "n_element_tab_1": 51, "formula_len_all": 323, "formula_len_all_1": 288, "len_all": 268097, "len_all_1": 65269, "len_abs": 1653, "len_title": 83, "len_sents": 70379, "len_sents_1": 30290, "n_sents": 638, "n_sents_1": 242, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1422, "title": "VidEgoThink: Assessing Egocentric Video Understanding Capabilities for Embodied AI", "abs": "Recent advancements in Multi-modal Large Language Models (MLLMs) have opened new avenues for applications in Embodied AI.\nBuilding on previous work, EgoThink, we introduce VidEgoThink, a comprehensive benchmark for evaluating egocentric video understanding capabilities. To bridge the gap between MLLMs and low-level control in Embodied AI, we design four key interrelated tasks: video question-answering, hierarchy planning, visual grounding and reward modeling. To minimize manual annotation costs, we develop an automatic data generation pipeline based on the Ego4D dataset, leveraging the prior knowledge and multimodal capabilities of GPT-4o. Three human annotators then filter the generated data to ensure diversity and quality, resulting in the VidEgoThink benchmark. We conduct extensive experiments with three types of models: API-based MLLMs, open-source image-based MLLMs, and open-source video-based MLLMs. Experimental results indicate that all MLLMs, including GPT-4o, perform poorly across all tasks related to egocentric video understanding. These findings suggest that foundation models still require significant advancements to be effectively applied to first-person scenarios in Embodied AI. In conclusion, VidEgoThink reflects a research trend towards employing MLLMs for egocentric vision, akin to human capabilities, enabling active observation and interaction in the complex real-world environments.", "keywords": [ "Multi-modal Large Language Models", "Egocentric Video Understanding", "Embodied AI", "Benchmark" ], "rating_list": [ 3, 6, 3, 5, 5 ], "soundness_list": [ 3, 2, 1, 3, 3 ], "presentation_list": [ 3, 3, 2, 4, 3 ], "contribution_list": [ 2, 2, 1, 2, 3 ], "confidence_list": [ 5, 4, 3, 4, 4 ] }, { "paper_id": "Z56fPyx7GL", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Federico Baldassarre", "gender": "Male", "institution": "Meta", "country": "FR", "position": "Postdoc" }, { "name": "Huy V. Vo", "gender": "Male", "institution": "Meta", "country": "FR", "position": "Researcher" }, { "name": "Josselin Somerville Roberts", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Maxime Oquab", "gender": "unknown", "institution": "Meta", "country": "FR", "position": "Research Scientist" }, { "name": "Piotr Bojanowski", "gender": "Male", "institution": "Meta", "country": "FR", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 72, "n_ref": 147, "n_ref_all": 166, "n_fig": 4, "n_tab": 6, "L_tab": 1702, "n_element_tab": 181, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1087, "n_element_tab_1": 109, "formula_len_all": 336, "formula_len_all_1": 336, "len_all": 180527, "len_all_1": 75444, "len_abs": 1526, "len_title": 104, "len_sents": 45099, "len_sents_1": 37182, "n_sents": 298, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1533, "title": "A Clustering Baseline for Object-Centric Representations", "abs": "Object-centric learning aims to discover and represent visual entities as a small set of object embeddings and masks, which can be later used for downstream tasks. Recent methods for object-centric learning build upon vision foundation models trained with self supervision because of the rich semantic features they produce. However, they often involve additional training to optimize for object mask accuracy for a specific granularity of objects on a test dataset, while overlooking the evaluation of the quality of the object embeddings which is arguably more important.\n\nIn this work, we demonstrate how to discover objects and parts with a simple multi-scale application of k-means to the features of an off-the-shelf backbone. Our method is fast and flexible, produces interpretable masks, preserves the quality of the backbone embeddings, does not require additional training, and can capture different part/whole structures.\n\nWe evaluate the quality of the obtained representation on a variety of downstream tasks including scene classification and action recognition in videos, showing that it surpasses the performance of fine-tuned object-centric learning methods. Object masks produced by our method also effectively capture real-world objects and parts at various granularity, with comparable quality to specialized methods when evaluated on unsupervised segmentation benchmarks. These results suggest rethinking the current approach to object-centric learning, with a greater focus on the quality of the representation.", "keywords": [ "object-centric representations", "self-supervised learning" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "Z4s2oe3Oiq", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Abdurakhmon Sadiev", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Artem Riabinin", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Eduard Gorbunov", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "Sarit Khirirat", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Postdoc" }, { "name": "Peter Richtarik", "gender": "Male", "institution": "King Abdullah University of Science and Technology (KAUST)", "country": "SA", "position": "Full Professor" } ], "n_formula": 149, "n_formula_1": 30, "n_ref_uni": 42, "n_ref": 116, "n_ref_all": 135, "n_fig": 6, "n_tab": 2, "L_tab": 214, "n_element_tab": 20, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 25028, "formula_len_all_1": 3564, "len_all": 179817, "len_all_1": 59169, "len_abs": 1547, "len_title": 123, "len_sents": 45590, "len_sents_1": 25617, "n_sents": 432, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1575, "title": "Communication-efficient Algorithms Under Generalized Smoothness Assumptions", "abs": "We provide the first proof of convergence for normalized error feedback algorithms across a wide range of machine learning problems. \nDespite their popularity and efficiency in training deep neural networks, traditional analyses of error feedback algorithms rely on the smoothness assumption that does not capture the properties of objective functions in these problems. \nRather, these problems have recently been shown to satisfy generalized smoothness assumptions, and the theoretical understanding of error feedback algorithms under these assumptions remains largely unexplored. \nMoreover, to the best of our knowledge, all existing analyses under generalized smoothness either i) focus on single-node settings or ii) make unrealistically strong assumptions for distributed settings, such as requiring data heterogeneity, and almost surely bounded stochastic gradient noise variance. \nIn this paper, we propose distributed error feedback algorithms that utilize normalization to achieve the $\\mathcal{O}(1/\\sqrt{K})$ convergence rate for nonconvex problems under generalized smoothness. Our analyses apply for distributed settings without data heterogeneity conditions, and enable stepsize tuning that is independent of problem parameters. \nAdditionally, we provide strong convergence guarantees of normalized error feedback algorithms for stochastic settings. \nFinally, we show that normalized EF21, due to its larger allowable stepsizes, outperforms EF21 on various tasks, including the minimization of polynomial functions, logistic regression, and ResNet-20 training.", "keywords": [ "Error Feedback; Distributed Learning; Generalized Smoothness; Nonconvex Optimization" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "Z4evOUYrk7", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Bo Dai", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Ceyuan Yang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Gordon Wetzstein", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Hao He", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Yinghao Xu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Yuwei Guo", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 43, "n_ref": 118, "n_ref_all": 145, "n_fig": 17, "n_tab": 8, "L_tab": 1157, "n_element_tab": 121, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1135, "n_element_tab_1": 111, "formula_len_all": 1190, "formula_len_all_1": 337, "len_all": 219533, "len_all_1": 66343, "len_abs": 1129, "len_title": 110, "len_sents": 64362, "len_sents_1": 32342, "n_sents": 494, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1133, "title": "CameraCtrl: Enabling Camera Control for Text-to-Video Generation", "abs": "Controllability plays a crucial role in video generation, as it allows users to create and edit content more precisely. Existing models, however, lack control of camera pose that serves as a cinematic language to express deeper narrative nuances. To alleviate this issue, we introduce CameraCtrl, enabling accurate camera pose control for video diffusion models. Our approach explores effective camera trajectory parameterization along with a plug-and-play camera pose control module that is trained on top of a video diffusion model, leaving other modules of the base model untouched. Moreover, a comprehensive study on the effect of various training datasets is conducted, suggesting that videos with diverse camera distributions and similar appearance to the base model indeed enhance controllability and generalization. Experimental results demonstrate the effectiveness of CameraCtrl in achieving precise camera control with different video generation models, marking a step forward in the pursuit of dynamic and customized video storytelling from textual and camera pose inputs. Code and models will be made publicly available.", "keywords": [ "camera viewpoints control in Video Generation" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 4, 4, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Z3xg3hxdky", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 7, "author_info_list": [ { "name": "Shenggan Cheng", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Xuanlei Zhao", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" }, { "name": "Zangwei Zheng", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zheming Yang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Science", "country": "CN", "position": "Assistant Professor" }, { "name": "Ziming Liu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Charlie Chen", "gender": "Female", "institution": "Peking University", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 47, "n_ref": 68, "n_ref_all": 85, "n_fig": 11, "n_tab": 6, "L_tab": 3005, "n_element_tab": 126, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 3290, "n_element_tab_1": 144, "formula_len_all": 847, "formula_len_all_1": 417, "len_all": 135580, "len_all_1": 57089, "len_abs": 1154, "len_title": 116, "len_sents": 31250, "len_sents_1": 23741, "n_sents": 241, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1150, "title": "DSP: Dynamic Sequence Parallelism for Multi-Dimensional Transformers", "abs": "Scaling multi-dimensional transformers to long sequences is indispensable across various domains. However, the challenges of large memory requirements and slow speeds of such sequences necessitate sequence parallelism. All existing approaches fall under the category of embedded sequence parallelism, which are limited to shard along a single sequence dimension, thereby introducing significant communication overhead. However, the nature of multi-dimensional transformers involves independent calculations across multiple sequence dimensions. To this end, we propose Dynamic Sequence Parallelism (DSP) as a novel abstraction of sequence parallelism. DSP dynamically switches the parallel dimension among all sequences according to the computation stage with efficient resharding strategy. DSP offers significant reductions in communication costs, adaptability across modules, and ease of implementation with minimal constraints. Experimental evaluations demonstrate DSP's superiority over state-of-the-art embedded sequence parallelism methods by remarkable throughput improvements ranging from 32.2% to 10x, with less than 25% communication volume.", "keywords": [ "Sequence Parallel", "Multi-Dimensional Transformers", "Distributed Computing", "High Performance Computing" ], "rating_list": [ 6, 6, 3, 6, 6 ], "soundness_list": [ 3, 3, 1, 3, 3 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 1, 2, 3 ], "confidence_list": [ 1, 3, 4, 3, 3 ] }, { "paper_id": "Z3waKPN7DG", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Chenhan D. Yu", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Greg Heinrich", "gender": "Male", "institution": "NVIDIA", "country": "FR", "position": "Researcher" }, { "name": "Ilia Markov", "gender": "unknown", "institution": "Institute of Science and Technology Austria", "country": "", "position": "PhD student" }, { "name": "Jan Kautz", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "VP Research" }, { "name": "Pavlo Molchanov", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Research Scientist" }, { "name": "Saurav Muralidharan", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Dan Alistarh", "gender": "Male", "institution": "Institute of Science and Technology", "country": "AT", "position": "Full Professor" }, { "name": "Hongxu Yin", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Senior Research Scientist" } ], "n_formula": 6, "n_formula_1": 8, "n_ref_uni": 32, "n_ref": 54, "n_ref_all": 76, "n_fig": 8, "n_tab": 6, "L_tab": 3227, "n_element_tab": 412, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2745, "n_element_tab_1": 439, "formula_len_all": 247, "formula_len_all_1": 393, "len_all": 144166, "len_all_1": 68406, "len_abs": 1093, "len_title": 121, "len_sents": 37711, "len_sents_1": 27923, "n_sents": 357, "n_sents_1": 260, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1112, "title": "UNAST: Unified framework for Neural Architecture Search for Transformers", "abs": "We introduce the UNAST, a new approach to optimize Large Language Models (LLMs) post-training. UNAST combines Neural Architecture Search (NAS) with sparsity and quantization for LLM compression. Starting with a trained model, UNAST replaces layers (e.g., attention and MLP) with more efficient alternatives by adjusting attention heads, KV projection dimensions, and MLP expansion factors. Local distillation pretrains layer candidates to mimic original layers. Scores and costs (latency, number of parameters, etc.) of each operator are fed into an Integer Linear Optimizer to find the optimal architecture under predefined constraints (latency, number of parameters, etc.). Our experiments show that UNAST scales to large models, reducing training costs by up to 10 times compared to training smaller models from scratch. Validation on GPT-3 and LLaMa models demonstrate that UNAST improves latency and memory footprint by up to 60\\% with minimal accuracy loss. UNAST also provides insights into the effects of different compression types on Transformer layers, aiding in the development of non-uniform models.", "keywords": [ "Efficiency", "Neural Architecture Search", "Large Language models" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 1, 1, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "Z3n2QauIIk", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Donghwan Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Han-Dong Lim", "gender": "Male", "institution": "KAIST", "country": "", "position": "MS student" } ], "n_formula": 169, "n_formula_1": 56, "n_ref_uni": 42, "n_ref": 83, "n_ref_all": 104, "n_fig": 1, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 20427, "formula_len_all_1": 5181, "len_all": 197193, "len_all_1": 72044, "len_abs": 1405, "len_title": 48, "len_sents": 45159, "len_sents_1": 25618, "n_sents": 580, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 860, "title": "A Finite-Time Analysis of Distributed Q-Learning", "abs": "Multi-agent reinforcement learning (MARL) has witnessed a remarkable surge in interest, fueled by the empirical success achieved in applications of single-agent reinforcement learning (RL). In this study, we consider a distributed Q-learning scenario, wherein a number of agents cooperatively solve a sequential decision making problem without access to the central reward function which is an average of the local rewards. In particular, we study finite-time analysis of a distributed Q-learning algorithm, and provide a new sample complexity result of $\\tilde{\\mathcal{O}}\\left( \\max\\left\\\\\\{ \\frac{1}{\\epsilon^2}\\frac{t_{\\text{mix}}}{(1-\\gamma)^6 d_{\\min}^4 } ,\\frac{1}{\\epsilon}\\frac{\\sqrt{|\\mathcal{S}||\\mathcal{A}|}}{(1-\\sigma_2(\\boldsymbol{W}))(1-\\gamma)^4 d_{\\min}^3} \\right\\\\\\} \\right)$ under tabular lookup setting for Markovian observation model.", "keywords": [ "reinforcement learning", "multi-agent", "distributed learning", "q-learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "Z33PEFMuU3", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Gurprit Singh", "gender": "unknown", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Researcher" }, { "name": "Jente Vandersanden", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Sascha Holl", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "", "position": "PhD student" }, { "name": "Xingchang Huang", "gender": "Male", "institution": "Max Planck Institute for Informatics", "country": "DE", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 17, "n_ref_uni": 29, "n_ref": 85, "n_ref_all": 156, "n_fig": 14, "n_tab": 8, "L_tab": 1259, "n_element_tab": 105, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1461, "n_element_tab_1": 16, "formula_len_all": 1412, "formula_len_all_1": 720, "len_all": 141316, "len_all_1": 63608, "len_abs": 1298, "len_title": 90, "len_sents": 48292, "len_sents_1": 27983, "n_sents": 389, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1180, "title": "Edge-preserving noise for diffusion models", "abs": "Classical generative diffusion models learn an isotropic Gaussian denoising process,\ntreating all spatial regions uniformly, thus neglecting potentially valuable structural\ninformation in the data. Inspired by the long-established work on anisotropic\ndiffusion in image processing, we present a novel edge-preserving diffusion model\nthat is a generalization of denoising diffusion probablistic models (DDPM). In\nparticular, we introduce an edge-aware noise scheduler that varies between edgepreserving\nand isotropic Gaussian noise. We show that our model\u2019s generative\nprocess converges faster to results that more closely match the target distribution.\nWe demonstrate its capability to better learn the low-to-mid frequencies within the\ndataset, which plays a crucial role in representing shapes and structural information.\nOur edge-preserving diffusion process consistently outperforms state-of-the-art\nbaselines in unconditional image generation. It is also more robust for generative\ntasks guided by a shape-based prior, such as stroke-to-image generation. We\npresent qualitative and quantitative results showing consistent improvements (FID\nscore) of up to 30% for both tasks.", "keywords": [ "diffusion", "generative modelling", "denoising" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "Z30Mdbv5jO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Fangfu Liu", "gender": "Male", "institution": "Department of Electronic Engineering, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hanyang Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Haowen Sun", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jun Zhang", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Wenqiang Sun", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yikai Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Yueqi Duan", "gender": "Male", "institution": "Department of Electronic Engineering, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "JUN LIANG YE", "gender": "Male", "institution": "Tsinghua University, China", "country": "CN", "position": "MS student" } ], "n_formula": 28, "n_formula_1": 16, "n_ref_uni": 45, "n_ref": 130, "n_ref_all": 169, "n_fig": 12, "n_tab": 3, "L_tab": 2195, "n_element_tab": 382, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 119, "n_element_tab_1": 40, "formula_len_all": 1644, "formula_len_all_1": 745, "len_all": 177540, "len_all_1": 51939, "len_abs": 1530, "len_title": 122, "len_sents": 46258, "len_sents_1": 24099, "n_sents": 337, "n_sents_1": 153, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1472, "title": "ReconX: Reconstruct Any Scene from Sparse Views with Video Diffusion Model", "abs": "Advancements in 3D scene reconstruction have transformed 2D images from the real world into 3D models, producing realistic 3D results from hundreds of input photos. Despite great success in dense-view reconstruction scenarios, rendering a detailed scene from sparse views is still an ill-posed optimization problem, often resulting in artifacts and distortions in unseen areas. In this paper, we propose ReconX, a novel 3D scene reconstruction paradigm that reframes the ambiguous reconstruction problem as a temporal generation task. The key insight is to unleash the strong generative prior of large pre-trained video diffusion models for sparse-view reconstruction. Nevertheless, it is challenging to preserve 3D view consistency when directly generating video frames from pre-trained models. To address this issue, given limited input views, the proposed ReconX first constructs a global point cloud and encodes it into a contextual space as the 3D structure condition. Guided by the condition, the video diffusion model then synthesizes video frames that are detail-preserved and exhibit a high degree of 3D consistency, ensuring the coherence of the scene from various perspectives. Finally, we recover the 3D scene from the generated video through a confidence-aware 3D Gaussian Splatting optimization scheme. Extensive experiments on various real-world datasets show the superiority of ReconX over state-of-the-art methods in terms of quality and generalizability.", "keywords": [ "Sparse-view Reconstruction", "Video Diffusion", "Gaussian Splatting" ], "rating_list": [ 6, 5, 5, 5, 8 ], "soundness_list": [ 4, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 5, 5 ] }, { "paper_id": "Z2uhdwOrn0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" }, { "name": "Jerry Yao-Chieh Hu", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Jiayi Wang", "gender": "Female", "institution": "Northwestern University", "country": "", "position": "PhD student" }, { "name": "Jing Liu", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Maojiang Su", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Najim Dehak", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "Associate Professor" }, { "name": "Shang Wu", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Yen-Ju Lu", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Haozheng LUO", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Jesus Villalba Lopez", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 5, "n_ref_uni": 61, "n_ref": 128, "n_ref_all": 142, "n_fig": 3, "n_tab": 7, "L_tab": 2682, "n_element_tab": 236, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2577, "n_element_tab_1": 106, "formula_len_all": 2820, "formula_len_all_1": 258, "len_all": 190231, "len_all_1": 67525, "len_abs": 1237, "len_title": 122, "len_sents": 52223, "len_sents_1": 31952, "n_sents": 437, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1242, "title": "SPARQ: Outlier-free SpeechLM with Fast Adaptation and Robust Quantization", "abs": "We propose SpARQ (outlier-free SpeechLM for Fast Adaptation and Robust Quantization) to address the outlier problem in Speech and Language multi-modal Models (SpeechLMs). Our primary observation is that outliers stemming from cross-modal (speech and text) low-rank adaptation and post-training quantization stages affect the performance of the current SpeechLMs. Methodologically, SpARQ leverages a pretrained language model as its foundation, substituting the traditional attention layer with a novel stabilized outlier-free layer. This modification eliminates outliers typically arising during cross-modal low-rank adaptation and post-training quantization. The model is then fine-tuned on multi-modal data using this outlier-free architecture, allowing it to handle textLM, speechLM, ASR, and TTS tasks through a unified interface while maintaining compatibility with parameters adapted from standard pretrained LLMs. Consequently, on the OPT-1.3b model, the proposed framework achieves relative performance improvements: 41\\% in cross-modal low-rank adaptation and 45\\% in post-training quantization, along with a 1.33x training speedup. We benchmark it against state-of-the-art low-rank adaptation and post-training quantization methods.", "keywords": [ "Multi-modal foundation model; Speech Language Model; Low-rank adaptation; Post-training quantization" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Z2QPJj52m3", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Hao Zhu", "gender": "Male", "institution": "CSIRO", "country": "AU", "position": "Postdoc" }, { "name": "Minxue Xia", "gender": "Female", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Yifei Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" } ], "n_formula": 33, "n_formula_1": 22, "n_ref_uni": 35, "n_ref": 58, "n_ref_all": 65, "n_fig": 1, "n_tab": 4, "L_tab": 2611, "n_element_tab": 323, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2133, "n_element_tab_1": 287, "formula_len_all": 2227, "formula_len_all_1": 1610, "len_all": 142381, "len_all_1": 71749, "len_abs": 1575, "len_title": 61, "len_sents": 36305, "len_sents_1": 29914, "n_sents": 275, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1580, "title": "PISA: Compressive Sensing Adaptation of Large Language Models", "abs": "In this paper, we introduce a novel perspective on Parameter-Efficient Fine-Tuning (PEFT) by viewing the weight update matrix as a k-sparse approximation in the spatial domain, departing from the commonly used low-rank structure assumption. We propose a compressive sensing-based approach that leverages under-complete measurement matrices to analyze the approximation capabilities of the weight update matrix. Our method ensures bounded error in the reconstruction of the weight updates, as guaranteed by theoretical results in compressive sensing.\nHowever, the vectorization of the weight update matrix leads to a high-dimensional problem (d^2), which can potentially result in large error bounds. To address this issue, we introduce a block-structured approximation scheme that partitions the weight update matrix into smaller blocks and applies the k-sparse approximation to each block independently. We theoretically analyze the approximation error bounds of our approach and demonstrate that the block-structured scheme achieves tighter error bounds compared to the non-block approach.\nEmpirically, we validate the effectiveness of our proposed method on various downstream NLP tasks, showcasing its ability to achieve competitive performance with a reduced number of trainable parameters. Our approach offers a new direction for parameter-efficient fine-tuning of large language models. Notably, our experiments demonstrate competitive performance with only 500 learnable parameters, while offering greater memory and computational efficiency than LoRA in a rank-1 setting.", "keywords": [ "Parameter-Efficient Fine-Tuning (PEFT)\uff0ck-sparse approximation\uff0clow-rank structure" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "Z28efFRE9r", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Bingtao Fu", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "Guoqiang Liang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "MS student" }, { "name": "Hong Gu", "gender": "Male", "institution": "Hangzhou VIVO Information Technology Co., Ltd", "country": "CN", "position": "Director" }, { "name": "Jinwei Chen", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Lin Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Qingnan Fan", "gender": "Male", "institution": "VIVO", "country": "CN", "position": "Lead Researcher" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 57, "n_ref": 158, "n_ref_all": 186, "n_fig": 18, "n_tab": 6, "L_tab": 1990, "n_element_tab": 298, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 380, "formula_len_all_1": 323, "len_all": 165144, "len_all_1": 52053, "len_abs": 1775, "len_title": 144, "len_sents": 38088, "len_sents_1": 25979, "n_sents": 268, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1790, "title": "AuthFace: Towards Authentic Blind Face Restoration with Face-oriented Generative Diffusion Prior", "abs": "Blind face restoration (BFR) is a challenging and fundamental problem in computer vision. To faithfully restore high-quality (HQ) photos from poor-quality ones, recent research endeavors predominantly rely on facial image priors from the powerful pretrained text-to-image (T2I) diffusion models. However, such priors often lead to the incorrect generation of non-facial features and insufficient facial details, thus rendering them less practical for real-world applications. In this paper, we propose a novel framework, namely AuthFace that achieves highly authentic face restoration results by exploring a face-oriented generative diffusion prior. To learn such a prior, we first collect a dataset of 1.5K high-quality images, with resolutions exceeding 8K, captured by professional photographers. Based on the dataset, we then introduce a novel face-oriented restoration-tuning pipeline that fine-tunes a pretrained T2I model. Identifying key criteria of quality-first and photography-guided annotation, we involve the retouching and reviewing process under the guidance of photographers for these high-quality images that show rich facial features. The photography-guided annotation system fully explores the potential of these high-quality photographic images. In this way, the potent natural image priors from pretrained T2I diffusion models can be subtly harnessed, specifically enhancing their capability in facial detail restoration. Moreover, to minimize artifacts in critical facial areas, such as eyes and mouth, we propose a time-aware latent facial feature loss to learn the authentic face restoration process. Extensive experiments on the synthetic and real-world BFR datasets demonstrate the superiority of our approach. Codes and datasets will be available upon acceptance.", "keywords": [ "Blind Face Restoration", "Diffusion Models" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "Z25xTjf3Mv", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Efthymia Tsamoura", "gender": "Female", "institution": "Samsung AI", "country": "GB", "position": "Researcher" }, { "name": "Liwei Che", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "vladimir pavlovic", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 31, "n_ref": 103, "n_ref_all": 115, "n_fig": 4, "n_tab": 2, "L_tab": 53, "n_element_tab": 5, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 835, "n_element_tab_1": 163, "formula_len_all": 940, "formula_len_all_1": 937, "len_all": 113416, "len_all_1": 72789, "len_abs": 936, "len_title": 128, "len_sents": 34512, "len_sents_1": 32538, "n_sents": 251, "n_sents_1": 234, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 940, "title": "Enhancing Deep Partial Label Learning via Casting it to a Satisfiability Problem", "abs": "Partial label learning (PLL) is a challenging real-world problem in the field of weakly supervised learning, in which each data instance contains a set of candidate labels with multiple ambiguous labels and one gold label. Although recent progress in PLL using deep representation learning has led to significant advances, the methods continue to experience significant performance drops on data with high label ambiguity and fine-grained categories. By casting PLL into a satisfiability problem and incorporating a loss based on this reduction, we show that the accuracy of those techniques can be further improved. We establish several key theoretical properties of the proposed SATisfiability-based (SAT) loss and its learning error bound. Our extensive empirical comparison reveals that the proposed loss improves over existing PLL techniques by up to 25.12% on multi-class benchmarks and 12.50% on fine-grained categorized benchmarks.", "keywords": [ "weak supervised learning", "partial label learning" ], "rating_list": [ 3, 3, 8, 5, 5 ], "soundness_list": [ 1, 2, 4, 2, 3 ], "presentation_list": [ 2, 2, 4, 2, 3 ], "contribution_list": [ 1, 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 3, 4, 4 ] }, { "paper_id": "Z1Va3Ue4GF", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Min Lin", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Weiqi Ruan", "gender": "unknown", "institution": "Southern University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Xiaodan Liang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Xiwen Liang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Yuecheng Liu", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Yuzheng Zhuang", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Research Engineer" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 17, "n_ref": 25, "n_ref_all": 37, "n_fig": 4, "n_tab": 6, "L_tab": 2127, "n_element_tab": 150, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 3224, "n_element_tab_1": 155, "formula_len_all": 378, "formula_len_all_1": 378, "len_all": 141816, "len_all_1": 58046, "len_abs": 1493, "len_title": 129, "len_sents": 29263, "len_sents_1": 24479, "n_sents": 221, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1496, "title": "Memory-Driven Multimodal Chain of Thought for Embodied Long-Horizon Task Planning", "abs": "Existing methods excel in short-horizon tasks but struggle with complex, long-horizon planning in dynamic environments. To address these limitations, we propose the Memory-Driven Multimodal Chain of Thought (MCoT-Memory), a framework designed to enhance task planning through two key innovations: 1) Evolving Scene Graph-Driven Chain of Thought with CoT Memory Retrieval, which enables the agent to continuously update a scene graph with visual information captured along its trajectory, providing a structured and dynamic representation of the environment that informs real-time decision-making, and uniquely incorporates CoT memory retrieval to allow the agent to leverage past experiences in its reasoning process; 2) Stepwise Confidence-Driven Memory Retention, which employs an expert model to evaluate reasoning across multiple dimensions of accuracy, ensuring that only high-confidence experiences are retained in memory for future retrieval, thus enabling the agent to build on valuable insights and improve performance in long-horizon tasks.\nTo advance long-horizon task planning, we present ExtendaBench, a comprehensive benchmark encompassing 1,198 tasks across two simulators, VirtualHome and Habitat 2.0. The tasks are categorized into ultra-short, short, median, and long tasks. Extensive experiments demonstrate that prior methods struggle with long-horizon tasks, while MCoT-Memory significantly improves performance, marking it as a promising approach for embodied task planning.", "keywords": [ "Task Planning; Multimodal Large Language Model; Reasoning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "Z1MKx8mNu9", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Eiko Yoneki", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Research Fellow" }, { "name": "Taiyi Wang", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Wenxuan Li", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Intern" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 52, "n_ref": 109, "n_ref_all": 135, "n_fig": 10, "n_tab": 4, "L_tab": 63, "n_element_tab": 12, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 284, "n_element_tab_1": 40, "formula_len_all": 149, "formula_len_all_1": 141, "len_all": 158507, "len_all_1": 59512, "len_abs": 900, "len_title": 127, "len_sents": 56831, "len_sents_1": 30293, "n_sents": 376, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 903, "title": "HiBO: Hierarchical Bayesian Optimization via Adaptive Search Space Partitioning", "abs": "Optimizing opaque functions in high-dimensional search spaces has been known to be challenging for traditional Bayesian Optimization (BO). In this paper, we introduce HiBO, a novel hierarchical algorithm integrating global-level search space partitioning information into the acquisition strategy of a local BO-based optimizer. HiBO employs a search-tree-based global-level navigator to adaptively split the search space into partitions with different sampling potential. The local optimizer then utilizes this global-level information to guide its acquisition strategy towards the most promising regions within the search space. A comprehensive set of evaluations demonstrates that HiBO outperforms state-of-the-art methods in high-dimensional synthetic benchmarks and presents significant practical effectiveness in the real-world task of tuning configurations of database management systems (DBMSs).", "keywords": [ "High-dimensional Bayesian Optimization", "Search Space Partitioning", "DBMS Configuration Tuning" ], "rating_list": [ 6, 5, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 4 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 2, 3 ] }, { "paper_id": "Z0qvzed8TK", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Can Li", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "", "position": "MS student" }, { "name": "Huiyan Wu", "gender": "unknown", "institution": "hikvision", "country": "", "position": "Researcher" }, { "name": "Jiang Zhu", "gender": "unknown", "institution": "Hikvision Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Liang Qiao", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Xi Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Xinglong Lu", "gender": "Male", "institution": "Hikvision Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Yi Niu", "gender": "unknown", "institution": "Hikvision Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Yunlu Xu", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 53, "n_ref": 105, "n_ref_all": 129, "n_fig": 14, "n_tab": 6, "L_tab": 2824, "n_element_tab": 384, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2508, "n_element_tab_1": 394, "formula_len_all": 64, "formula_len_all_1": 64, "len_all": 229820, "len_all_1": 64318, "len_abs": 1115, "len_title": 124, "len_sents": 47560, "len_sents_1": 27691, "n_sents": 366, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1114, "title": "Towards Multi-Domain Chinese Document VQA: a New Dataset and Baseline Method", "abs": "Document Visual Question Answering (DocVQA) remains a significant challenge in the field of document understanding and is a critical evaluation metric for current general-purpose large model techniques. However, prevailing public datasets are predominantly designed for single scenarios or specific sources. Furthermore, most available datasets are in English, limiting the verification of model performance in other languages. This paper presents a novel multi-domain Chinese document VQA dataset, which includes 39 document types from 7 different domains. The designed question set encompasses both common extractive questions and complex abstractive questions. Based on this dataset, we conducted a comprehensive review and analysis of various technical paradigms, including both traditional and large model-based approaches. Using the popular in-context learning framework, we propose a robust baseline that achieves commendable few-shot adaptation. Comparative evaluations demonstrate the superior performance of the proposed method across different solution paradigms. The dataset and code will be published.", "keywords": [ "Benchmark", "Dataset", "Chinese DocVQA", "In-Context Learning" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "Z0eiiV3Yyh", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Deng Cai", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" }, { "name": "Jiang Bian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Shizhao Sun", "gender": "Female", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Wenxiao Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhanwei Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 15, "n_ref": 155, "n_ref_all": 296, "n_fig": 20, "n_tab": 9, "L_tab": 1471, "n_element_tab": 194, "n_fig_1": 10, "n_tab_1": 7, "L_tab_1": 3046, "n_element_tab_1": 82, "formula_len_all": 1487, "formula_len_all_1": 1374, "len_all": 217483, "len_all_1": 68850, "len_abs": 1946, "len_title": 145, "len_sents": 86767, "len_sents_1": 31346, "n_sents": 389, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1355, "title": "FlexCAD: Unified and Versatile Controllable CAD Generation with Fine-tuned Large Language Models", "abs": "Recently, there is a growing interest in creating computer-aided design (CAD) models based on user intent, known as controllable CAD generation. Existing work offers limited controllability and needs separate models for different types of control, reducing efficiency and practicality. To achieve controllable generation across all CAD construction hierarchies, such as sketch-extrusion, extrusion, sketch, face, loop and curve, we propose FlexCAD, a unified model by fine-tuning large language models (LLMs). First, to enhance comprehension by LLMs, we represent a CAD model as a structured text by abstracting each hierarchy as a sequence of text tokens. Second, to address various controllable generation tasks in a unified model, we introduce a hierarchy-aware masking strategy. Specifically, during training, we mask a hierarchy-aware field in the CAD text with a mask token. This field, composed of a sequence of tokens, can be set flexibly to represent various hierarchies. Subsequently, we ask LLMs to predict this masked field. During inference, the user intent is converted into a CAD text with a mask token replacing the part the user wants to modify, which is then fed into FlexCAD to generate new CAD models. \nComprehensive experiments on public dataset demonstrate the effectiveness of FlexCAD in both generation quality and controllability.", "keywords": [ "CAD generation", "large language model" ], "rating_list": [ 6, 6, 3 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "YzxMu1asQi", "primary_area": "interpretability and explainable AI", "n_author": 1, "author_info_list": [ { "name": "Stanislav Fort", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 24, "n_ref": 37, "n_ref_all": 79, "n_fig": 12, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15, "formula_len_all_1": 15, "len_all": 143490, "len_all_1": 72988, "len_abs": 2212, "len_title": 125, "len_sents": 47280, "len_sents_1": 37702, "n_sents": 305, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1729, "title": "Scaling Laws for Adversarial Attacks on Language Model Activations and Tokens", "abs": "We explore a class of adversarial attacks targeting the activations of language models to derive upper-bound scaling laws on their attack susceptibility. By manipulating a relatively small subset of model activations, $a$, we demonstrate the ability to control the exact prediction of a significant number (in some cases up to 1000) of subsequent tokens $t$. We empirically verify a scaling law where the maximum number of target tokens predicted, $t_\\mathrm{max}$, depends linearly on the number of tokens $a$ whose activations the attacker controls as $t_\\mathrm{max} = \\kappa a$. We find that the number of bits the attacker controls on the input to exert a single bit of control on the output (a property we call \\textit{attack resistance $\\chi$}) is remarkably stable between $\\approx 16$ and $\\approx 25$ over orders of magnitude of model sizes and between model families. Compared to attacks directly on input tokens, attacks on activations are predictably much stronger, however, we identify a surprising regularity where one bit of input steered either via activations or via tokens is able to exert a surprisingly similar amount of control over the model predictions. This gives support for the hypothesis that adversarial attacks are a consequence of dimensionality mismatch between the input and output spaces. A practical implication of the ease of attacking language model activations instead of tokens is for multi-modal and selected retrieval models. By using language models as a controllable test-bed to study adversarial attacks, we explored input-output dimension regimes that are inaccessible in computer vision and greatly extended the empirical support for the dimensionality theory of adversarial attacks.", "keywords": [ "adversarial attacks", "language models", "scaling laws", "activation steering" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "YzXPU3QRnL", "primary_area": "generative models", "n_author": 11, "author_info_list": [ { "name": "Albert Webson", "gender": "unknown", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Da Huang", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Dustin Tran", "gender": "unknown", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Hanxiao Liu", "gender": "Male", "institution": "Google Brain", "country": "", "position": "Research Scientist" }, { "name": "Jason Wei", "gender": "Male", "institution": "OpenAI", "country": "", "position": "Researcher" }, { "name": "Tengyu Ma", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Stanford University" }, { "name": "Yi Tay", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Yifeng Lu", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Researcher" }, { "name": "Dengyong Zhou", "gender": "unknown", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Jerry Weng Wei", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Xinyun Chen", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 126, "n_ref_all": 184, "n_fig": 21, "n_tab": 3, "L_tab": 3165, "n_element_tab": 117, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 245, "n_element_tab_1": 36, "formula_len_all": 170, "formula_len_all_1": 0, "len_all": 284389, "len_all_1": 57763, "len_abs": 2038, "len_title": 101, "len_sents": 102490, "len_sents_1": 29814, "n_sents": 683, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1414, "title": "Larger language models do in-context learning differently", "abs": "We study how in-context learning (ICL) in language models is affected by semantic priors versus input-label mappings. We investigate two setups - ICL with flipped labels and ICL with semantically-unrelated labels - across various model families (GPT-3, InstructGPT, Codex, an internal model, and an instruction-tuned variant of the internal model). First, experiments on ICL with flipped labels show that overriding semantic priors is an emergent ability of model scale. While small language models ignore flipped labels presented in-context and thus rely primarily on semantic priors from pretraining, large models can override semantic priors when presented with in-context exemplars that contradict priors, despite the stronger semantic priors that larger models may hold. We next study semantically-unrelated label ICL (SUL-ICL), in which labels are semantically unrelated to their inputs (e.g., foo/bar instead of negative/positive), thereby forcing language models to learn the input-label mappings shown in in-context exemplars in order to perform the task. The ability to do SUL-ICL also emerges primarily with scale, and large-enough language models can even perform linear classification in a SUL-ICL setting. Finally, we evaluate instruction-tuned models and find that instruction tuning strengthens both the use of semantic priors and the capacity to learn input-label mappings, but more of the former.", "keywords": [ "in-context learning", "natural language processing", "large language models" ], "rating_list": [ 8, 8, 6, 6, 1 ], "soundness_list": [ 3, 4, 4, 3, 3 ], "presentation_list": [ 3, 4, 4, 4, 2 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 5 ] }, { "paper_id": "YzFNJ571A7", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Andong Hua", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Chenhe Gu", "gender": "unknown", "institution": "University of California, Irvine", "country": "US", "position": "MS student" }, { "name": "Jindong Gu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Yao Qin", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 5, "n_ref_uni": 35, "n_ref": 95, "n_ref_all": 110, "n_fig": 7, "n_tab": 4, "L_tab": 5563, "n_element_tab": 160, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 4294, "n_element_tab_1": 158, "formula_len_all": 1151, "formula_len_all_1": 234, "len_all": 145336, "len_all_1": 57422, "len_abs": 3753, "len_title": 140, "len_sents": 39213, "len_sents_1": 24299, "n_sents": 336, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 951, "title": "Improving Adversarial Transferability in MLLMs via Dynamic Vision-Language Alignment Attack", "abs": "Multimodal Large Language Models (MLLMs), built upon LLMs, have recently gained attention for their capabilities in image recognition and understanding. However, while MLLMs are vulnerable to adversarial attacks, the transferability of these attacks across different models remains limited, especially under targeted attack setting. Existing methods primarily focus on vision-specific perturbations but struggle with the complex nature of vision-language modality alignment. In this work, we introduce the Dynamic Vision-Language Alignment (DynVLA) Attack, a novel approach that injects dynamic perturbations into the vision-language connector to enhance generalization across diverse vision-language alignment of different models. Our experimental results show that DynVLA significantly improves the transferability of adversarial examples across various MLLMs, including BLIP2, InstructBLIP, MiniGPT4, LLaVA, and closed-source models such as Gemini.", "keywords": [ "Adversarial Attack", "Adversarial Transferability", "Multimodal Large Language model", "Multimodal Large Language Model Robustness" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 5, 3 ] }, { "paper_id": "Yz7ts36V7A", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Bhuwan Dhingra", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Assistant Professor" }, { "name": "Maximilian Holsman", "gender": "Male", "institution": "Department of Computer Science, Duke University", "country": "US", "position": "Researcher" }, { "name": "Rickard Stureborg", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Yukun Huang", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 15, "n_ref": 19, "n_ref_all": 25, "n_fig": 3, "n_tab": 2, "L_tab": 1188, "n_element_tab": 167, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 363, "n_element_tab_1": 66, "formula_len_all": 279, "formula_len_all_1": 279, "len_all": 145433, "len_all_1": 62080, "len_abs": 1412, "len_title": 169, "len_sents": 38211, "len_sents_1": 33732, "n_sents": 206, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 114, "L_abs": 1416, "title": "Backoff Decoding: A Language Model Inference Acceleration Framework with a Tunable Efficiency-Performance Tradeoff", "abs": "In current transformer-based language models, all tokens in a sequence are generated by identical forward passes and thereby incur the same inference cost. However, tokens vary widely in their importance to the overall generation and their difficulty for models to generate correctly, making this equal allocation of inference resources suboptimal. We introduce backoff decoding, a framework for efficient language model inference that dynamically allocates token generations between two (or more) models of different sizes, according to an arbitrary decision function. By modifying how this decision function allocates generations between the differently sized models, users can tune their generation along an efficiency-performance tradeoff to suit the needs of their application. Backoff decoding can be used on any set of models with the same tokenizer and does not require any training or finetuning of the models themselves. As a demonstration of our framework, we show that backoff decoding with a large and a small model can significantly reduce inference cost while sacrificing virtually no performance compared to the standalone large model. We then show that inference costs can be reduced even further, achieving inference accelerations of up to 3-4x in exchange for reductions in model performance, demonstrating an efficiency-performance tunability not found in other inference acceleration techniques.", "keywords": [ "language modeling", "inference acceleration", "decoding strategies" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "YyVVicZ32M", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Lei Li", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Xuandong Zhao", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Yu-Xiang Wang", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" } ], "n_formula": 36, "n_formula_1": 23, "n_ref_uni": 53, "n_ref": 99, "n_ref_all": 124, "n_fig": 6, "n_tab": 8, "L_tab": 3338, "n_element_tab": 282, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 371, "n_element_tab_1": 41, "formula_len_all": 3268, "formula_len_all_1": 1346, "len_all": 192333, "len_all_1": 60663, "len_abs": 945, "len_title": 121, "len_sents": 56312, "len_sents_1": 26029, "n_sents": 501, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 948, "title": "Permute-and-Flip: An optimally stable and watermarkable decoder for LLMs", "abs": "In this paper, we propose a new decoding method called Permute-and-Flip (PF) decoder. It enjoys stability properties similar to the standard sampling decoder, but is provably up to 2x better in its quality-stability tradeoff than sampling and never worse than any other decoder. We also design a cryptographic watermarking scheme analogous to Aaronson (2023)'s Gumbel watermark, but naturally tailored for PF decoder. The watermarking scheme does not change the distribution to sample, while allowing arbitrarily low false positive rate and high recall whenever the generated text has high entropy. Our experiments show that the PF decoder (and its watermarked counterpart) significantly outperform(s) naive sampling (and its Gumbel watermarked counterpart) in terms of perplexity, while retaining the same stability (and detectability), hence making it a promising new approach for LLM decoding. We provide the code in the supplementary materials.", "keywords": [ "LLM", "Watermark", "AI Safety", "Decoding" ], "rating_list": [ 6, 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 2, 4, 3 ], "presentation_list": [ 2, 3, 2, 4, 3 ], "contribution_list": [ 3, 3, 2, 4, 3 ], "confidence_list": [ 4, 4, 4, 2, 3 ] }, { "paper_id": "YxOG4FjZLd", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Chen Wang", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Jiayi Kuang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Wenhao Jiang", "gender": "Male", "institution": "Guangming Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Ying Shen", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "Associate Professor" }, { "name": "Yinghui Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 10, "n_ref": 14, "n_ref_all": 46, "n_fig": 15, "n_tab": 17, "L_tab": 4044, "n_element_tab": 627, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2611, "n_element_tab_1": 421, "formula_len_all": 73, "formula_len_all_1": 73, "len_all": 200599, "len_all_1": 61340, "len_abs": 1635, "len_title": 90, "len_sents": 65623, "len_sents_1": 22364, "n_sents": 470, "n_sents_1": 170, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1638, "title": "\ud83e\udd14Emoji2Idiom: Benchmarking Cryptic Symbol Understanding of Multimodal Large Language Models", "abs": "Vision and Language are two major modalities in Artificial Intelligence research.\nBridging the gap between these modalities has long been a key focus in the multimodal community.\nInspired by human cognition, we believe that if a model can see an image and directly associate it with its linguistic meaning, the model possesses high-level intelligence that spans vision and language.\nIn our work, we focus on emojis in images, a widely-used \"cryptic symbol\", with a data form of both visual and linguistic features, i.e. emojis have the specific textual semantics while human understand the meaning from their visual information.\nSpecifically, we first propose the novel task of translating emojis in images to corresponding idioms, thereby challenging Multimodal Large Language Models (MLLMs) to (1) understand the semantic correlation between language and emojis, and (2) reason the intricate linguistic meaning from the emojis in images.\nTo facilitate the advancement of this task, we construct a high-quality benchmark (emoji2idiom) following the process of automatic model generation and human manual filtering. \nBased on our constructed emoji2idiom, we employ multiple advanced MLLMs to conduct extensive experiments and detailed analyses, demonstrating that existing MLLMs do not yet have enough capability to understand and reason the linguistic information from visual data.\nWe believe our proposed benchmark and interesting discoveries will encourage the community to attach importance to the intelligence of MLLMs directly associating language from vision, to give MLLMs more comprehensive vision-language understanding ability.", "keywords": [ "Multimodal Large Language Models", "Benchmark", "Vision and Language" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "YxLxrWkwsX", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Abdul Basit", "gender": "Male", "institution": "Lahore University of Management Sciences", "country": "PK", "position": "MS student" }, { "name": "Ahmed Rashid Qazi", "gender": "Male", "institution": "Tibbling Technologies", "country": "US", "position": "Researcher" }, { "name": "Asim Iqbal", "gender": "Male", "institution": "Tibbling Technologies", "country": "US", "position": "Co-founder" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 12, "n_ref": 50, "n_ref_all": 88, "n_fig": 9, "n_tab": 11, "L_tab": 2698, "n_element_tab": 343, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1654, "n_element_tab_1": 175, "formula_len_all": 398, "formula_len_all_1": 331, "len_all": 184305, "len_all_1": 61637, "len_abs": 1290, "len_title": 131, "len_sents": 54628, "len_sents_1": 28087, "n_sents": 372, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1364, "title": "NeuGen: Amplifying the \u2018Neural\u2019 in Neural Radiance Fields for Domain Generalization", "abs": "Neural Radiance Fields (NeRF) have significantly advanced the field of novel view synthesis, yet their generalization across diverse scenes and conditions remains challenging. Addressing this, we propose the integration of a novel brain-inspired normalization technique Neural Generalization (NeuGen) into leading NeRF architectures which include MVSNeRF and GeoNeRF. NeuGen extracts domain-invariant features, thereby enhancing the models' generalization capabilities. It can be seamlessly integrated into NeRF architectures, capable of initiating training from scratch or fine-tuning pre-trained models, which cultivates a comprehensive feature set that significantly improves accuracy and robustness in image rendering. Through this integration, NeuGen shows benchmarking performance on diverse datasets across state-of-the-art NeRF architectures, enabling them to generalize better across varied scenes. Our comprehensive evaluations, both quantitative and qualitative, confirm that our approach not only surpasses existing models in generalizability but also markedly improves rendering quality. Our work exemplifies the potential of merging neuroscientific principles with deep learning frameworks, setting a new precedent for enhanced generalizability and efficiency in novel view synthesis. A demo of our study is available at https://neugennerf.github.io.", "keywords": [ "Domain Generalization", "Scene Reconstruction", "One-shot learning" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "Yx9j88efei", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Adam D. Cobb", "gender": "Male", "institution": "SRI International", "country": "", "position": "Researcher" }, { "name": "Brian Matejek", "gender": "Male", "institution": "SRI International", "country": "", "position": "Researcher" }, { "name": "Malyaban Bal", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Susmit Jha", "gender": "unknown", "institution": "SRI International", "country": "US", "position": "Principal Scientist" } ], "n_formula": 26, "n_formula_1": 17, "n_ref_uni": 44, "n_ref": 83, "n_ref_all": 113, "n_fig": 10, "n_tab": 2, "L_tab": 231, "n_element_tab": 35, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 221, "n_element_tab_1": 31, "formula_len_all": 1656, "formula_len_all_1": 1090, "len_all": 140915, "len_all_1": 65082, "len_abs": 2121, "len_title": 125, "len_sents": 42567, "len_sents_1": 29882, "n_sents": 315, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 2130, "title": "SpikingVTG: Saliency Feedback Gating Enabled Spiking Video Temporal Grounding", "abs": "Video Temporal Grounding (VTG) seeks to retrieve consecutive intervals or specific clips from a video based on specified natural language queries. VTG requires accurately aligning video segments with corresponding natural language instructions, highlighting the need for effective methodologies to capture semantic correspondence and maintain temporal coherence. Spiking neural networks (SNNs), previously underexplored in this domain, present a unique opportunity to tackle VTG challenges from both the architectural and energy-efficiency perspectives. In this paper, we leverage sparse spike-based communication of SNNs to propose a multimodal architecture tailored for VTG tasks, namely SpikingVTG, providing a biologically inspired and efficient solution. Leveraging temporal saliency feedback, our proposed spiking video-language model (VLM) achieves competitive performance with non-spiking VLMs across diverse moment retrieval and highlight detection tasks. We introduce a Saliency Feedback Gating (SFG) mechanism that improves performance while reducing overall neural activity. To efficiently train our spiking VLM, we analyze the convergence dynamics of each neuronal layer and utilize equilibrium states to enable training using implicit differentiation at equilibrium. This approach eliminates the need for computationally expensive backpropagation through time while also enabling the use of knowledge distillation for efficient model training. To further improve operational efficiency and facilitate the on-chip deployability of our model, we leverage a multi-stage training pipeline that focuses on eliminating non-local computations, such as softmax and layer normalization, leading to the development of the Normalization Free (NF)-SpikingVTG model. Additionally, we create an extremely quantized variant, a 1-bit NF-SpikingVTG model, which vastly improves computational efficiency during inference while maintaining minimal performance degradation from our base model. Our work introduces the first spiking model to demonstrate competitive performance on VTG benchmarks, including QVHighlights and Charades-STA.", "keywords": [ "Spiking Neural Network", "Brain-inspired algorithms" ], "rating_list": [ 6, 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "YwzxpZW3p7", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Ali Hasan", "gender": "unknown", "institution": "Duke University", "country": "US", "position": "Graduate student" }, { "name": "Haoming Yang", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Vahid Tarokh", "gender": "Male", "institution": "Duke University", "country": "", "position": "Professor" }, { "name": "Yuting Ng", "gender": "unknown", "institution": "DSO National Laboratories (Singapore)", "country": "SG", "position": "Researcher" } ], "n_formula": 35, "n_formula_1": 19, "n_ref_uni": 22, "n_ref": 57, "n_ref_all": 72, "n_fig": 3, "n_tab": 13, "L_tab": 4695, "n_element_tab": 377, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1473, "n_element_tab_1": 134, "formula_len_all": 4049, "formula_len_all_1": 1694, "len_all": 173062, "len_all_1": 69434, "len_abs": 1111, "len_title": 76, "len_sents": 60475, "len_sents_1": 31244, "n_sents": 441, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 28, "L_abs": 1117, "title": "Elliptic Loss Regularization", "abs": "Regularizing neural networks is important for anticipating model behavior in regions of the data space that are not well represented. In this work, we propose a regularization technique for enforcing a level of smoothness in the mapping between the input space and the loss. We specify the level of regularity by requiring that the loss of the network satisfies an elliptic operator over the data domain. To do this, we modify the usual empirical risk minimization objective such that we instead minimize a new objective that satisfies an elliptic operator over points within the domain. This allows us to use existing theory on elliptic operators to anticipate the behavior of the error for points outside the training set. We propose a tractable computational method that approximates the behavior of the elliptic operator while being computationally efficient. Finally, we analyze the properties of the proposed regularization to understand the performance on common problems of distribution shift and group imbalance. Numerical experiments empirically confirm the promise of the proposed regularization technique.", "keywords": [ "regularizer", "loss landscape", "diffusion", "elliptic" ], "rating_list": [ 5, 6, 8, 3 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 2, 2, 4, 1 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "YwJkv2YqBq", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Kanan Gupta", "gender": "Male", "institution": "University of Pittsburgh", "country": "US", "position": "PhD student" }, { "name": "Stephan Wojtowytsch", "gender": "unknown", "institution": "University of Pittsburgh", "country": "", "position": "Assistant Professor" } ], "n_formula": 142, "n_formula_1": 38, "n_ref_uni": 48, "n_ref": 66, "n_ref_all": 80, "n_fig": 3, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18841, "formula_len_all_1": 3681, "len_all": 210398, "len_all_1": 68207, "len_abs": 863, "len_title": 54, "len_sents": 62040, "len_sents_1": 29130, "n_sents": 592, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 55, "L_abs": 864, "title": "Nesterov acceleration in benignly non-convex landscapes", "abs": "While momentum-based optimization algorithms are commonly used in the notoriously non-convex optimization problems of deep learning, their analysis has historically been restricted to the convex and strongly convex setting. In this article, we partially close this gap between theory and practice and demonstrate that virtually identical guarantees can be obtained in optimization problems with a `benign' non-convexity. We show that these weaker geometric assumptions are well justified in overparametrized deep learning, at least locally. Variations of this result are obtained for a continuous time model of Nesterov's accelerated gradient descent algorithm (NAG), the classical discrete time version of NAG, and versions of NAG with stochastic gradient estimates with purely additive noise and with noise that exhibits both additive and multiplicative scaling.", "keywords": [ "Nonconvex optimization", "stochastic optimization", "stochastic acceleration", "smooth convex optimization", "deep learning", "accelerated gradient descent" ], "rating_list": [ 6, 5, 8, 8 ], "soundness_list": [ 1, 2, 3, 4 ], "presentation_list": [ 1, 3, 4, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "YvtiTdXjfn", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Alexander Rubinstein", "gender": "unknown", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Damien Teney", "gender": "Male", "institution": "Idiap Research Institute", "country": "CH", "position": "Researcher" }, { "name": "Luca Scimeca", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Postdoc" }, { "name": "Seong Joon Oh", "gender": "Male", "institution": "Parameter Lab", "country": "DE", "position": "Principal Researcher" } ], "n_formula": 30, "n_formula_1": 13, "n_ref_uni": 52, "n_ref": 120, "n_ref_all": 155, "n_fig": 7, "n_tab": 11, "L_tab": 2005, "n_element_tab": 127, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1956, "n_element_tab_1": 69, "formula_len_all": 1737, "formula_len_all_1": 786, "len_all": 172776, "len_all_1": 68904, "len_abs": 1372, "len_title": 70, "len_sents": 49264, "len_sents_1": 29045, "n_sents": 391, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1392, "title": "Scalable Ensemble Diversification for OOD Generalization and Detection", "abs": "Training a diverse ensemble of models has several practical applications such as providing candidates for model selection with better out-of-distribution (OOD) generalization, and enabling the detection of OOD samples via Bayesian principles.\nAn existing approach to diverse ensemble training encourages the models to disagree on provided OOD samples. However, the approach is computationally expensive and it requires well-separated ID and OOD examples, such that it has only been demonstrated in small-scale settings.\n\n**Method.**\nThis work presents a Hardness-based Diversification Regularizer (HDR)\napplicable to large-scale settings (e.g. ImageNet)\nthat does not require OOD samples.\nInstead, HDR identifies hard training samples on the fly\nand encourages the ensemble members to disagree on these.\nTo improve scaling, we show how to avoid the expensive computations in existing methods of exhaustive pairwise disagreements across models.\n\n**Results.**\nWe evaluate the benefits of diversification with experiments on ImageNet.\nFirst, for OOD generalization, we observe large benefits from the diversification in multiple settings including output-space (classical) ensembles and weight-space ensembles (model soups).\nSecond, for OOD detection, we turn the diversity of ensemble hypotheses into a novel uncertainty score estimator that surpasses a large number of OOD detection baselines.", "keywords": [ "diverse", "ensemble", "scalable", "robustness", "uncertainty", "OOD detection", "OOD generalization" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 2, 4 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "YvWuac63bg", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Deli Zhao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Director" }, { "name": "Fei Wu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Hang Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Kehan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Lidong Bing", "gender": "unknown", "institution": "Alibaba Group", "country": "SG", "position": "Scientist" }, { "name": "Sicong Leng", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xin Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zesen Cheng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhiqiang Hu", "gender": "unknown", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 14, "n_ref_uni": 35, "n_ref": 69, "n_ref_all": 95, "n_fig": 5, "n_tab": 5, "L_tab": 1442, "n_element_tab": 215, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1104, "n_element_tab_1": 174, "formula_len_all": 1214, "formula_len_all_1": 1039, "len_all": 135472, "len_all_1": 56367, "len_abs": 643, "len_title": 130, "len_sents": 35236, "len_sents_1": 22192, "n_sents": 305, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1263, "title": "Breaking the Memory Barrier: Near Infinite Batch Size Scaling for Contrastive Loss", "abs": "Contrastive loss is a powerful approach for representation learning, where larger batch sizes enhance performance by providing more negative samples to better distinguish between similar and dissimilar data. However, scaling batch sizes is constrained by the quadratic growth in GPU memory consumption, primarily due to the full instantiation of the similarity matrix. To address this, we propose a tile-based computation strategy that partitions the contrastive loss calculation into arbitrary small blocks, avoiding full materialization of the similarity matrix. Furthermore, we introduce a multi-level tiling strategy to leverage the hierarchical structure of distributed systems, employing ring-based communication at the GPU level to optimize synchronization and fused kernels at the CUDA core level to reduce I/O overhead. Experimental results show that the proposed method scales batch sizes to unprecedented levels. For instance, it enables contrastive training of a CLIP-ViT-L/14 model with a batch size of 4M or 12M using 8 or 32 A800 80GB without sacrificing any accuracy. Compared to SOTA memory-efficient solutions, it achieves a two-order-of-magnitude reduction in memory while maintaining comparable speed. The code will be made publicly available.", "keywords": [ "Contrastive Learning; Tiling Calculation;" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "YvOq7jHT6R", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Bin Gu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Huan Xiong", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xiang Peng", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Xinzhe Yuan", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 104, "n_formula_1": 30, "n_ref_uni": 23, "n_ref": 55, "n_ref_all": 68, "n_fig": 2, "n_tab": 2, "L_tab": 984, "n_element_tab": 146, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 16087, "formula_len_all_1": 3665, "len_all": 140807, "len_all_1": 54351, "len_abs": 1595, "len_title": 117, "len_sents": 40310, "len_sents_1": 21806, "n_sents": 398, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1610, "title": "Fight Fire with Fire: Multi-biased Interactions in Hard-Thresholding", "abs": "$\\ell_0$ constrained optimization is widely used in machine learning, especially for high-dimensional problems, as it effectively promotes sparse learning. A prominent technique for solving these problems is Hard-Thresholding gradient descent. However, the inherent expansibility of Hard-Thresholding operators can lead to convergence issues, necessitating strategies to accelerate the algorithm. In this article, we believe the random Hard-Thresholding algorithm can be interpreted as an equivalent biased gradient algorithm. By introducing appropriate biases, we can mitigate some of the issues of Hard-Thresholding and enhance convergence. We categorize the biases into memory-biased and recursively-biased, examining their distinct applications within Hard-Thresholding algorithms. Next, we explore the Zeroth-Order versions of these algorithms, which introduce additional biases from Zeroth-Order gradients. Our findings indicate that recursively bias effectively counteracts some of the issues caused by Hard-Thresholding, resulting in improved performance for First-Order algorithms. Conversely, due to the accumulation of errors from Zeroth-Order gradients during recursive bias, the performance of Zeroth-Order algorithms is inferior to that influenced by historical gradients. To address these insights, we propose the SARAHT and BVR-SZHT algorithms for First-Order and Zeroth-Order Hard-Thresholding, respectively, both of which demonstrate faster convergence speeds compared to previous methods. We validate our hypotheses through black-box adversarial experiments and ridge regression evaluations.", "keywords": [ "Optimzation", "Biased Gardient", "Zeroth-Order", "Hard-Thresholding" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 1, 1, 2, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "YvKJGYL4j7", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Kun Zhu", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Tianxu Li", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 14, "n_ref_uni": 29, "n_ref": 57, "n_ref_all": 82, "n_fig": 13, "n_tab": 7, "L_tab": 1663, "n_element_tab": 160, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1168, "formula_len_all_1": 776, "len_all": 151103, "len_all_1": 64304, "len_abs": 1280, "len_title": 125, "len_sents": 53634, "len_sents_1": 31332, "n_sents": 387, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1272, "title": "Toward Efficient Multi-Agent Exploration With Trajectory Entropy Maximization", "abs": "Recent works have increasingly focused on learning decentralized policies for agents as a solution to the scalability challenges in Multi-Agent Reinforcement Learning (MARL), where agents typically share the parameters of a policy network to make action decisions. However, this parameter sharing can impede efficient exploration, as it may lead to similar behaviors among agents. Different from previous mutual information-based methods that promote multi-agent diversity, we introduce a novel multi-agent exploration method called Trajectory Entropy Exploration (TEE). Our method employs a particle-based entropy estimator to maximize the entropy of different agents' trajectories in a contrastive trajectory representation space, resulting in diverse trajectories and efficient exploration. This entropy estimator avoids challenging density modeling and scales effectively in high-dimensional multi-agent settings. We integrate our method with MARL algorithms by deploying an intrinsic reward for each agent to encourage entropy maximization. To validate the effectiveness of our method, we test our method in challenging multi-agent tasks from several MARL benchmarks. The results demonstrate that our method consistently outperforms existing state-of-the-art methods.", "keywords": [ "Multi-Agent Reinforcement Learning", "Exploration", "Cooperation", "Trajectory Entropy Maximization" ], "rating_list": [ 3, 6, 8, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "Yv8FrCY87H", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Deyi Ji", "gender": "unknown", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Feng Zhao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Hongtao Lu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Jieping Ye", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Lanyun Zhu", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "PhD student" }, { "name": "Peng Xu", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Siqi Gao", "gender": "Female", "institution": "Alibaba Group", "country": "", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 41, "n_ref": 87, "n_ref_all": 129, "n_fig": 5, "n_tab": 15, "L_tab": 8166, "n_element_tab": 978, "n_fig_1": 4, "n_tab_1": 14, "L_tab_1": 6705, "n_element_tab_1": 835, "formula_len_all": 368, "formula_len_all_1": 280, "len_all": 159472, "len_all_1": 78242, "len_abs": 1478, "len_title": 137, "len_sents": 39425, "len_sents_1": 28655, "n_sents": 251, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1411, "title": "Tree-of-Table: Unleashing the Power of LLMs for Enhanced Large-Scale Table Understanding", "abs": "The ubiquity and value of tables as semi-structured data across various domains necessitate advanced methods for understanding their complexity and vast amounts of information. Despite the impressive capabilities of large language models (LLMs) in advancing the natural language understanding frontier, their application to large-scale tabular data presents significant challenges, specifically regarding table size and complex intricate relationships. Existing works have shown promise with small-scale tables but often flounder when tasked with the complex reasoning required by larger, interconnected tables found in real-world scenarios. To address this gap, we introduce \"Tree-of-Table\", a novel approach designed to enhance LLMs' reasoning capabilities over large and complex tables. Our method employs Table Condensation and Decomposition to distill and reorganize relevant data into a manageable format, followed by the construction of a hierarchical Table-Tree that facilitates tree-structured reasoning. Through a meticulous Table-Tree Execution process, we systematically unravel the tree-structured reasoning chain to derive the solutions. Experiments across diverse datasets, including WikiTQ, TableFact, FeTaQA, and BIRD, demonstrate that Tree-of-Table sets a new benchmark with superior performance, showcasing remarkable efficiency and generalization capabilities in large-scale table reasoning.", "keywords": [ "LLMs", "Table Reasoning" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 4, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "YuwxDSqNXw", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 1, "author_info_list": [ { "name": "Paul Alexander Bilokon", "gender": "unknown", "institution": "Thalesians Marine Ltd, the AI in shipping and logistics company", "country": "", "position": "Principal Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 6, "n_ref": 7, "n_ref_all": 8, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 70, "formula_len_all_1": 70, "len_all": 65014, "len_all_1": 61887, "len_abs": 1417, "len_title": 129, "len_sents": 30849, "len_sents_1": 30802, "n_sents": 213, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1423, "title": "Compromised Turing Machines: Adversarial Interference and Endogenous Verification", "abs": "We introduce the concept of a Compromised Turing Machine (CTM), an extension of the classical Turing machine model where an adversary, Eve, can tamper with the tape or internal state between timesteps. The CTM exposes fundamental vulnerabilities in the machine's ability to self-verify its computations, particularly in adversarial environments where endogenous verification mechanisms cannot reliably ensure computational integrity. Through a novel parallel with Descartes' deus deceptor thought experiment, we explore the epistemological limits of computational certainty, illustrating how the CTM reveals the failure of self-verification in adversarial contexts. \n\nTo address these vulnerabilities, we propose several secure computational models, including hybrid systems with external verification, randomized and probabilistic verification protocols, distributed computing models with cross-verification, self-correcting and self-healing mechanisms, and advanced cryptographic techniques such as zero-knowledge proofs and homomorphic encryption. While each solution presents trade-offs in terms of computational overhead and complexity, they provide a foundation for building resilient systems capable of withstanding adversarial interference. Our work highlights the need for external sources of trust and verification in secure computation and opens new directions for research into adversarial computational models.", "keywords": [ "compromised turing machine", "ctm", "adversarial interference", "endogenous verification" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 1, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "Yufi9fWhj9", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jaehyung Kim", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Sungsoo Ahn", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Yunhui Jang", "gender": "Female", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 33, "n_ref": 101, "n_ref_all": 150, "n_fig": 13, "n_tab": 7, "L_tab": 3736, "n_element_tab": 627, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 79, "n_element_tab_1": 3, "formula_len_all": 776, "formula_len_all_1": 377, "len_all": 180514, "len_all_1": 59441, "len_abs": 905, "len_title": 93, "len_sents": 47219, "len_sents_1": 28629, "n_sents": 396, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 907, "title": "Chain-of-thoughts for molecular understanding", "abs": "The adaptation of large language models (LLMs) to chemistry have shown promising performance in molecular understanding tasks, such as generating a text description from a molecule. However, proper reasoning based on molecular structural information remains a significant challenge, e.g., even advanced LLMs such as GPT-4o struggle to identify functional groups which are crucial for inferring the molecular property of interest. To address this limitation, we propose StructCoT, a structure-aware chain-of-thought (CoT) that enhances LLMs\u2019 understanding of molecular structures by explicitly injecting the key structural features of molecules. Moreover, we introduce two fine-tuning frameworks for adapting the existing LLMs to use our StructCoT. Our experiments demonstrate that incorporating \\Algname with our fine-tuning frameworks leads to consistent improvements in both molecular understanding tasks.", "keywords": [ "Large language model", "chain-of-thought" ], "rating_list": [ 6, 3, 5, 6, 6 ], "soundness_list": [ 2, 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 3, 2 ], "contribution_list": [ 3, 2, 1, 3, 2 ], "confidence_list": [ 3, 5, 5, 3, 5 ] }, { "paper_id": "YuJdtpPV4n", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jiayi Ma", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Shihua Zhang", "gender": "unknown", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Xiang Fang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" } ], "n_formula": 25, "n_formula_1": 22, "n_ref_uni": 40, "n_ref": 100, "n_ref_all": 126, "n_fig": 5, "n_tab": 7, "L_tab": 2298, "n_element_tab": 180, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 864, "n_element_tab_1": 48, "formula_len_all": 1217, "formula_len_all_1": 1078, "len_all": 124109, "len_all_1": 60648, "len_abs": 1272, "len_title": 121, "len_sents": 31046, "len_sents_1": 26687, "n_sents": 236, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1276, "title": "MambaMatch: Learning Two-View Correspondences with Selective State Spaces", "abs": "Two-view correspondence learning aims to discern true and false correspondences between image pairs by recognizing their underlying different information. Previous methods either treat the information equally or fail to discard the superfluous information of false correspondences, tending to be invalid in practical scenarios. Therefore, inspired by Mamba's inherent competence of selectivity, we propose MambaMatch as a Mamba-based correspondence filter to selectively mine information from true correspondences and to dispose of the potentially interfering information of false correspondences. Specifically, the selection is achieved by adaptively adjusting model parameters in a high-dimensional latent space, which also avoids attention leakage and implements context compression, ensuring the precise and efficient exploitation of pertinent information. Meanwhile, channel awareness is tailored to serve as a complementary aspect of comprehensive information acquisition. Moreover, we design a novel local-context enhancement module to capture reasonable local context that is crucial for correspondence pruning. Extensive experiments demonstrate that our approach outperforms existing state-of-the-art methods on several visual tasks while saving time and space costs.", "keywords": [ "Correspondence learning", "image matching", "feature matching" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "YuHQTo6G9S", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Haonan Wang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Honglong Yang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Jiaji Mao", "gender": "Male", "institution": "Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Jun Shen", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Lehan Wang", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Xiaomeng Li", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Zehong Yang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 86, "n_ref_all": 120, "n_fig": 14, "n_tab": 12, "L_tab": 7102, "n_element_tab": 938, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 3821, "n_element_tab_1": 423, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 232450, "len_all_1": 59417, "len_abs": 1741, "len_title": 132, "len_sents": 60407, "len_sents_1": 25157, "n_sents": 445, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 84, "L_abs": 1776, "title": "Interpretable Bilingual Multimodal Large Language Model for Diverse Biomedical Tasks", "abs": "Several medical Multimodal Large Languange Models (MLLMs) have been developed to address tasks involving visual images with textual instructions across various medical modalities, achieving impressive results. \nMost current medical generalist models are region-agnostic, treating the entire image as a holistic representation. However, they struggle to identify which specific regions they are focusing on when generating a sentence.\nTo mimic the behavior of doctors, who typically begin by reviewing the entire image before concentrating on specific regions for a thorough evaluation, we aim to enhance the capability of medical MLLMs in understanding anatomical regions within entire medical scans.\nTo achieve it, we first formulate \\textbf{Region-Centric tasks} and construct a \\textbf{large-scale dataset, MedRegInstruct,} to incorporate regional information into training. Combining our collected dataset with other medical multimodal corpora for training, we propose a \\textbf{Region-Aware medical MLLM, MedRegA}, which is the first bilingual generalist medical AI system to simultaneously handle image-level and region-level medical vision-language tasks across a broad range of modalities. Our MedRegA not only enables three region-centric tasks, but also achieves the best performance for visual question answering, report generation and medical image classification over 8 modalities, showcasing significant versatility. Experiments demonstrate that our model can not only accomplish powerful performance across various medical vision-language tasks in bilingual settings, but also recognize and detect structures in multimodal medical scans, boosting the interpretability and user interactivity of medical MLLMs. The codes and model will be made publicly available.", "keywords": [ "Multimodal Large Language Model", "Biomedicine", "Region-Text" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 4, 4 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "YuFUUcSUgx", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Alejandro Mart\u00edn", "gender": "Male", "institution": "Universidad Polit\u00e9cnica de Madrid", "country": "ES", "position": "Associate Professor" }, { "name": "David Camacho", "gender": "unknown", "institution": "Universidad Polit\u00e9cnica de Madrid", "country": "ES", "position": "Full Professor" }, { "name": "Javier Huertas-Tato", "gender": "unknown", "institution": "Universidad Polit\u00e9cnica de Madrid", "country": "ES", "position": "Associate Professor" }, { "name": "Pablo Miralles-Gonz\u00e1lez", "gender": "Male", "institution": "Universidad Polit\u00e9cnica de Madrid", "country": "ES", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 0, "n_ref": 23, "n_ref_all": 32, "n_fig": 3, "n_tab": 7, "L_tab": 2776, "n_element_tab": 363, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 674, "n_element_tab_1": 66, "formula_len_all": 455, "formula_len_all_1": 455, "len_all": 104404, "len_all_1": 68625, "len_abs": 1743, "len_title": 118, "len_sents": 37029, "len_sents_1": 30523, "n_sents": 296, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1749, "title": "You Can Train from Scratch: Further Discussion on the Long Range Arena", "abs": "Despite their success, Transformers suffer from quadratic complexity in the sequence length, limiting their applicability to long-range dependency problems and making them expensive to train and run. After many proposals to address this issue, the Long Range Arena (LRA) was suggested as a benchmark to evaluate the performance of new models in long-range dependency modeling tasks. The Transformer and its variants performed poorly on this benchmark, and a new series of architectures such as State Space Models (SSMs) gained some traction, greatly outperforming Transformers in the LRA. Recent work has shown that with a denoising pretraining phase, Transformers can achieve competitive results in the LRA with these new architectures. In this work, we discuss and explain the superiority of architectures such as MEGA and SSMs in the Long Range Arena, as well as the recent improvement in the results of Transformers, pointing to the positional and local nature of the tasks. We show that while the LRA is a benchmark for long-range dependency modeling, in reality most of the performance comes from short-range dependencies. By using rotary embeddings and training techniques to mitigate its data inefficiency, the Transformer is also able to reach state-of-the-art performance without a separate pretraining phase. What is more, with the same techniques, we are able to remove all restrictions from SSM convolutional kernels and learn fully parameterized convolutions without decreasing performance, suggesting that the design choices behind SSMs merely added inductive biases and learning efficiency for these particular tasks. Our insights indicate that LRA results should be interpreted with caution and call for a redesign of the benchmark.", "keywords": [ "Long Range Arena", "Transformers" ], "rating_list": [ 5, 1, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "YtGtIAYDV3", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Bisakh Banerjee", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Mohammad Alwardat", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Selin Aviyente", "gender": "Female", "institution": "Michigan State University", "country": "", "position": "Full Professor" }, { "name": "Tapabrata Maiti", "gender": "unknown", "institution": "Michigan State University", "country": "", "position": "Full Professor" } ], "n_formula": 85, "n_formula_1": 20, "n_ref_uni": 26, "n_ref": 54, "n_ref_all": 63, "n_fig": 6, "n_tab": 2, "L_tab": 206, "n_element_tab": 15, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9007, "formula_len_all_1": 1710, "len_all": 137934, "len_all_1": 55212, "len_abs": 1704, "len_title": 110, "len_sents": 42763, "len_sents_1": 24674, "n_sents": 375, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1714, "title": "Node-based Multiple Graph Learning with Theoretical Guarantees", "abs": "In many applications, inferring graph topology, i.e., learning the graph structure from a given set of nodal observations, is a significant task. Existing approaches are mostly limited to learning a single graph assuming that the observed data are homogeneous. In many applications, data sets are heterogeneous and involve multiple related graphs, i.e., multiview graphs. Recent work on learning multiview graphs ensures the similarity of learned view graphs through edge-based similarity between the graphs. In this paper, we take a node-based approach instead of assuming that similarities and differences between networks are driven by individual edges, providing a more intuitive interpretation of network differences. Moreover, unlike existing methods that employ Gaussian Graphical Models (GGM), which learn precision matrices rather than the actual graph structures, we characterize the graph using a Laplacian matrix. Thus, the approach is expected to work broadly beyond Gaussian graphical learning. We develop an optimization framework to learn the individual graphical structures, assuming that the differences are due to individual nodes that are perturbed across views. The proposed optimization framework is presented for the special case of two views. Furthermore, we derive the upper bound on the estimation error of the proposed graph estimator and characterize the impact of the sample size, number of nodes, and the spectrum of the graph Laplacians on estimation errors. The approach is evaluated on synthetic graph data for robustness against noise, graph density, and sample size. Finally, the proposed framework is applied to two-view real-world graph data for graph learning and clustering.", "keywords": [ "graph learning", "multiple graphs", "perturbation", "graph signal smoothness" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "YtAhOVCy2t", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Divyat Mahajan", "gender": "Male", "institution": "Montreal Institute of Learning Algorithms", "country": "CA", "position": "PhD student" }, { "name": "Ioannis Mitliagkas", "gender": "Male", "institution": "Athena Research Center", "country": "GR", "position": "Researcher" }, { "name": "Kartik Ahuja", "gender": "Male", "institution": "FAIR (Meta)", "country": "FR", "position": "Researcher" }, { "name": "Mohammad Pezeshki", "gender": "Male", "institution": "Facebook", "country": "CA", "position": "Researcher" }, { "name": "Pascal Vincent", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "Adjunct Professor" } ], "n_formula": 80, "n_formula_1": 17, "n_ref_uni": 47, "n_ref": 111, "n_ref_all": 148, "n_fig": 6, "n_tab": 14, "L_tab": 7022, "n_element_tab": 858, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9495, "formula_len_all_1": 1315, "len_all": 280901, "len_all_1": 70039, "len_abs": 966, "len_title": 79, "len_sents": 88753, "len_sents_1": 30879, "n_sents": 772, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 31, "L_abs": 969, "title": "Compositional Risk Minimization", "abs": "In this work, we tackle a challenging and extreme form of subpopulation shift, which is termed compositional shift. Under compositional shifts, some combinations of attributes are totally absent from the training distribution but present in the test distribution. We model the data with flexible additive energy distributions, where each energy term represents an attribute, and derive a simple alternative to empirical risk minimization termed compositional risk minimization (CRM). We first train an additive energy classifier to predict the multiple attributes and then adjust this classifier to tackle compositional shifts. We provide an extensive theoretical analysis of CRM, where we show that our proposal extrapolates to special affine hulls of seen attribute combinations. Empirical evaluations on benchmark datasets confirms the improved robustness of CRM compared to other methods from the literature designed to tackle various forms of subpopulation shifts.", "keywords": [ "Compositional Generalization", "Out of Distribution Generalization", "Provable Extrapolation" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 2 ] }, { "paper_id": "Yt9CFhOOFe", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Amy Wang", "gender": "unknown", "institution": "Genentech", "country": "US", "position": "Researcher" }, { "name": "Aya Abdelsalam Ismail", "gender": "Female", "institution": "Guide Labs", "country": "US", "position": "Researcher" }, { "name": "Hector Corrada Bravo", "gender": "Male", "institution": "Genentech", "country": "", "position": "Principal Scientist" }, { "name": "Julius Adebayo", "gender": "Male", "institution": "Guide Labs", "country": "", "position": "Researcher" }, { "name": "Kyunghyun Cho", "gender": "Male", "institution": "Genentech", "country": "", "position": "Senior Director of Frontier Research" }, { "name": "Nathan C. Frey", "gender": "unknown", "institution": "Prescient Design, Genentech", "country": "", "position": "Researcher" }, { "name": "Samuel Don Stanton", "gender": "Male", "institution": "Genentech", "country": "", "position": "Researcher" }, { "name": "Tuomas Oikarinen", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 3, "n_ref_uni": 59, "n_ref": 134, "n_ref_all": 184, "n_fig": 43, "n_tab": 5, "L_tab": 1944, "n_element_tab": 126, "n_fig_1": 17, "n_tab_1": 2, "L_tab_1": 1039, "n_element_tab_1": 63, "formula_len_all": 519, "formula_len_all_1": 103, "len_all": 232122, "len_all_1": 65527, "len_abs": 1232, "len_title": 101, "len_sents": 72780, "len_sents_1": 31235, "n_sents": 509, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1244, "title": "Concept Bottleneck Language Models For Protein Design", "abs": "We introduce Concept Bottleneck Protein Language Models (CB-pLM), a generative masked language model with a layer where each neuron corresponds to an interpretable concept. Our architecture offers three key benefits: i) Control: We can intervene on concept values to precisely control the properties of generated proteins, achieving a 3$\\times$ larger change in desired concept values compared to baselines. ii) Interpretability: A linear mapping between concept values and predicted tokens allows transparent analysis of the model's decision-making process. iii) Debugging: This transparency facilitates easy debugging of trained models. Our models achieve pre-training perplexity and downstream task performance comparable to traditional masked protein language models, demonstrating that interpretability does not compromise performance. While adaptable to any language model, we focus on masked protein language models due to their importance in drug discovery and the ability to validate our model's capabilities through real-world experiments and expert knowledge. We scale our CB-pLM from 24 million to 3 billion parameters, making them the largest Concept Bottleneck Models trained and the first capable of generative language modeling.", "keywords": [ "LLMs", "protein design", "concept bottleneck" ], "rating_list": [ 8, 8, 5, 6, 6 ], "soundness_list": [ 4, 3, 2, 2, 3 ], "presentation_list": [ 4, 3, 1, 3, 3 ], "contribution_list": [ 4, 4, 2, 2, 3 ], "confidence_list": [ 3, 3, 2, 2, 4 ] }, { "paper_id": "YslOW2SO6S", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Deli Zhao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Director" }, { "name": "Fugee Tsung", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Jia Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Jiashun Cheng", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yang Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yu Rong", "gender": "Male", "institution": "Alibaba Group ", "country": "CN", "position": "Senior Staff Algorithm Engineer" }, { "name": "Zinan Zheng", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "MS student" } ], "n_formula": 22, "n_formula_1": 22, "n_ref_uni": 44, "n_ref": 85, "n_ref_all": 109, "n_fig": 8, "n_tab": 3, "L_tab": 2890, "n_element_tab": 175, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 926, "n_element_tab_1": 65, "formula_len_all": 1300, "formula_len_all_1": 1224, "len_all": 129429, "len_all_1": 54255, "len_abs": 1270, "len_title": 132, "len_sents": 29839, "len_sents_1": 22325, "n_sents": 234, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1273, "title": "CirT: Global Subseasonal-to-Seasonal Forecasting with Geometry-inspired Transformer", "abs": "Accurate Subseasonal-to-Seasonal (S2S) climate forecasting is pivotal for decision-making including agriculture planning and disaster preparedness but is known to be challenging due to its chaotic nature. Although recent data-driven models have shown promising results, their performance is limited by inadequate consideration of geometric inductive biases. Usually, they treat the spherical weather data as planar images, resulting in an inaccurate representation of locations and spatial relations. In this work, we propose the geometric-inspired Circular Transformer (CirT) to model the cyclic characteristic of the graticule, consisting of two key designs: (1) Decomposing the weather data by latitude into circular patches that serve as input tokens to the Transformer; (2) Leveraging Fourier transform in self-attention to capture the global information and model the spatial periodicity. Extensive experiments on the Earth Reanalysis 5 (ERA5) reanalysis dataset demonstrate our model yields a significant improvement over the advanced data-driven models, including PanguWeather and GraphCast, as well as skillful ECMWF systems. Additionally, we empirically show the effectiveness of our model designs and high-quality prediction over spatial and temporal dimensions.", "keywords": [ "Weather and climate forecasting" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 5, 1, 3 ] }, { "paper_id": "Ysdo3fyD4Q", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Chris Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jingwen He", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Tianfan Xue", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Wanli Ouyang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Full Professor" }, { "name": "Xinqi Lin", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Ziwei Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Gao Peng", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 27, "n_ref": 100, "n_ref_all": 116, "n_fig": 7, "n_tab": 4, "L_tab": 871, "n_element_tab": 90, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1005, "n_element_tab_1": 64, "formula_len_all": 527, "formula_len_all_1": 588, "len_all": 135202, "len_all_1": 64687, "len_abs": 1418, "len_title": 114, "len_sents": 36476, "len_sents_1": 29060, "n_sents": 292, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1469, "title": "VEnhancer: Generative Space-Time Enhancement for Video Generation", "abs": "We present \\emph{VEnhancer}, a generative space-time enhancement method that can improve the existing AI-generated videos spatially and temporally through one video diffusion model. Given a generated low-quality video, our approach can increase its spatial and temporal resolution simultaneously with arbitrary up-sampling space and time scales by adding more details in spatial domain and synthesize detailed motion in temporal domain. Furthermore, VEnhancer is able to remove generated spatial artifacts and temporal flickering of generated videos. \nTo achieve this, basing on a pretrained generative video prior, we train a \\textbf{S}pace-\\textbf{T}ime Controller and inject it to the prior as a condition on low-frame-rate and low-resolution videos. To effectively train this ST-Controller, we design \\textit{space-time data augmentation} to create diversified video training pairs as well as \\textit{video-aware conditioning} for realizing different augmentation parameters in both spatial and temporal dimensions.\nBenefiting from the above designs, VEnhancer can be end-to-end trained to enable multi-function in one single model. \nExtensive experiments show that VEnhancer\nsurpasses existing state-of-the-art video super-resolution and space-time super-resolution methods in enhancing AI-generated videos. Moreover, VEnhancer is able to greatly improve the performance of open-source state-of-the-art text-to-video methods on video generation benchmark, VBench.", "keywords": [ "Diffusion models", "Video Generation", "Generative Video enhancement", "video super-resolution", "frame interpolation", "space-time super-resolution" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "YsOndItIxV", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Hyungkyu Kang", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Undergrad student" }, { "name": "Min-hwan Oh", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" }, { "name": "Seunghoon Kim", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Undergrad student" }, { "name": "Young-Geun Choi", "gender": "unknown", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "Associate Professor" } ], "n_formula": 86, "n_formula_1": 13, "n_ref_uni": 42, "n_ref": 142, "n_ref_all": 145, "n_fig": 2, "n_tab": 3, "L_tab": 566, "n_element_tab": 17, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 70, "n_element_tab_1": 8, "formula_len_all": 13154, "formula_len_all_1": 721, "len_all": 179258, "len_all_1": 61753, "len_abs": 3049, "len_title": 105, "len_sents": 52295, "len_sents_1": 28516, "n_sents": 501, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1392, "title": "Neural Dynamic Pricing: Provable and Practical Efficiency", "abs": "Despite theoretical guarantees of existing dynamic pricing (DP) methods, their strong model assumptions may not reflect real-world conditions and are often unverifiable. This poses major challenges in practice since the performance of an algorithm may significantly degrade if the assumptions are not satisfied. Moreover, many DP algorithms show unfavorable empirical performance due to the lack of data efficiency. \n To address these challenges, we design a practical contextual DP algorithm that utilizes regression oracles. Our proposed algorithm assumes only Lipschitz continuity on the true conditional probability of purchase.\n We prove $\\tilde{\\mathcal{O}}(T^{\\frac{2}{3}}\\text{regret}_R(T)^{\\frac{1}{3}})$ regret upper bound where $T$ is the horizon and $\\text{regret}_R(T)$ is the regret of the oracle. The bound is nearly minimax optimal in the canonical case of finite function class, and our analysis generically applies to other function approximators including neural networks. To the best of our knowledge, our work is the first algorithm to utilize the powerful generalization capability of neural networks with provable guarantees in dynamic pricing literature.\n Extensive numerical experiments show that our algorithm outperforms existing state-of-the-art dynamic pricing algorithms in various settings, which demonstrates both provable efficiency and practicality.", "keywords": [ "dynamic pricing", "neural networks" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "Ys1ZbGBzHJ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Kaidi Xu", "gender": "Male", "institution": "Drexel University", "country": "US", "position": "Assistant Professor" }, { "name": "Qi Guo", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Rui Zhang", "gender": "Female", "institution": "Institute of Computing Technology, CAS", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaoyun Zhang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Xing Hu", "gender": "Female", "institution": ", Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yunji Chen", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhengyue Zhao", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Zidong Du", "gender": "unknown", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 5, "n_ref_uni": 27, "n_ref": 67, "n_ref_all": 95, "n_fig": 5, "n_tab": 16, "L_tab": 4450, "n_element_tab": 515, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1964, "n_element_tab_1": 326, "formula_len_all": 801, "formula_len_all_1": 538, "len_all": 176322, "len_all_1": 50951, "len_abs": 1570, "len_title": 149, "len_sents": 51327, "len_sents_1": 20833, "n_sents": 404, "n_sents_1": 155, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 1576, "title": "Adversarial Contrastive Decoding: Aligning Large Language Models via Exploiting Their Safety and Harm", "abs": "With the widespread application of Large Language Models (LLMs), it has become a significant concern to ensure their safety and prevent harmful responses. While current safe-alignment methods based on instruction fine-tuning and Reinforcement Learning from Human Feedback (RLHF) can effectively reduce harmful responses from LLMs, they often require high-quality datasets and heavy computational overhead during model training. Another way to align language models is to modify the logit of tokens in model outputs without heavy training. Recent studies have shown that contrastive decoding can enhance the performance of language models by reducing the likelihood of confused tokens. However, these methods require the manual selection of contrastive models or instruction templates, limiting the degree of contrast. To this end, we propose Adversarial Contrastive Decoding (ACD), an optimization-based framework to generate two opposite soft system prompts, the Safeguarding Prompt (SP) and the Adversarial Prompt (AP), for prompt-based contrastive decoding. The SP aims to promote safer outputs while the AP aims to exploit the harmful parts of the model, providing a strong contrast to align the model with safety. ACD only needs to apply a lightweight prompt tuning on a rather small anchor dataset without training the target model. Experiments conducted on extensive models and benchmarks demonstrate that the proposed method achieves much better safety performance than previous model training-free decoding methods without sacrificing its original generation ability.", "keywords": [ "large language models", "safety alignment", "prompting" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "YrycTjllL0", "primary_area": "datasets and benchmarks", "n_author": 33, "author_info_list": [ { "name": "Armel Randy Zebaze", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "PhD student" }, { "name": "Binyuan Hui", "gender": "Female", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Chen GONG", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Daniel Fried", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Researcher" }, { "name": "David Lo", "gender": "Male", "institution": "Singapore Management University", "country": "", "position": "Full Professor" }, { "name": "Han Hu", "gender": "Male", "institution": "Monash University", "country": "", "position": "PhD student" }, { "name": "Haolan Zhan", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Harm de Vries", "gender": "Male", "institution": "ServiceNow Research", "country": "", "position": "Researcher" }, { "name": "Imam Nur Bani Yusuf", "gender": "Male", "institution": "Singapore Management University", "country": "", "position": "PhD student" }, { "name": "James Hoang", "gender": "Male", "institution": "Singapore Management University", "country": "", "position": "PhD student" }, { "name": "Jean Kaddour", "gender": "Male", "institution": "University College London", "country": "", "position": "PhD student" }, { "name": "Jenny Chim", "gender": "Female", "institution": "Queen Mary University London", "country": "GB", "position": "PhD student" }, { "name": "Jiawei Liu", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Junda He", "gender": "Male", "institution": "Singapore Management University", "country": "", "position": "PhD student" }, { "name": "Leandro Von Werra", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "MS student" }, { "name": "Ming Xu", "gender": "Female", "institution": "Uber", "country": "US", "position": "Researcher" }, { "name": "Naman Jain", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Niklas Muennighoff", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Prateek Yadav", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina, Chapel Hill", "country": "", "position": "Graduate Student" }, { "name": "Qian Liu", "gender": "Male", "institution": "Tiktok", "country": "SG", "position": "Researcher" }, { "name": "Ratnadira Widyasari", "gender": "Female", "institution": "Singapore Management University", "country": "SG", "position": "PhD student" }, { "name": "Simon Brunner", "gender": "Male", "institution": "Universit\u00e4t Innsbruck", "country": "AT", "position": "PhD student" }, { "name": "Vu Minh Chien", "gender": "Male", "institution": "Tohoku University", "country": "JP", "position": "PhD student" }, { "name": "Wen-Ding Li", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Wenhao Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Researcher" }, { "name": "Xiaoheng Hong", "gender": "Male", "institution": "Intel", "country": "CN", "position": "Researcher" }, { "name": "Xiaoning Du", "gender": "Female", "institution": "Monash University", "country": "", "position": "Lecturer" }, { "name": "Zhihan Zhang", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Zhoujun Cheng", "gender": "Male", "institution": "UC San Diego", "country": "US", "position": "PhD student" }, { "name": "Zijian Wang", "gender": "Male", "institution": "Amazon AWS AI Labs", "country": "US", "position": "Researcher" }, { "name": "Alex Felix Gu", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Indraneil Arun Paul", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Terry Zhuo", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 60, "n_ref": 162, "n_ref_all": 191, "n_fig": 10, "n_tab": 12, "L_tab": 6732, "n_element_tab": 398, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3, "formula_len_all_1": 0, "len_all": 333728, "len_all_1": 61510, "len_abs": 1962, "len_title": 151, "len_sents": 92993, "len_sents_1": 28674, "n_sents": 752, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 95, "L_abs": 1693, "title": "BigCodeBench: Benchmarking Code Generation with Diverse Function Calls and Complex Instructions", "abs": "Task automation has been greatly empowered by the recent advances in Large Language Models (LLMs) via Python code, where the tasks range from software engineering development to general-purpose reasoning. While current benchmarks have shown that LLMs can solve tasks using programs like human developers, the majority of their evaluations are limited to short and self-contained algorithmic tasks or standalone function calls. Solving challenging and practical tasks requires the capability of utilizing **diverse function calls as tools** to efficiently implement functionalities like data analysis and web development. In addition, using multiple tools to solve a task needs compositional reasoning by accurately understanding **complex instructions**. Fulfilling both of these characteristics can pose a great challenge for LLMs. To assess how well LLMs can solve challenging and practical tasks via programs, we introduce BigCodeBench, a benchmark that challenges LLMs to invoke multiple function calls as tools from 139 libraries and 7 domains for 1,140 fine-grained tasks. To evaluate LLMs rigorously, each task encompasses 5.6 test cases with an average branch coverage of 99%. In addition, we propose a natural-language-oriented variant of BigCodeBench, BigCodeBench-Instruct, that automatically transforms the original docstrings into short instructions containing only essential information. Our extensive evaluation of 60 LLMs shows that **LLMs are not yet capable of following complex instructions to use function calls precisely, with scores up to 60%, significantly lower than the human performance of 97%**. The results underscore the need for further advancements in this area.", "keywords": [ "Code Generation", "Tool Use", "Instruction Following", "Benchmark" ], "rating_list": [ 10, 10, 8, 8 ], "soundness_list": [ 4, 4, 3, 3 ], "presentation_list": [ 4, 4, 4, 3 ], "contribution_list": [ 4, 4, 4, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "YryL3QIWWc", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Jathushan Rajasegaran", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Jitendra Malik", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Director of Research" }, { "name": "Rahul Ravishankar", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Zeeshan Patel", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 50, "n_ref": 75, "n_ref_all": 88, "n_fig": 7, "n_tab": 4, "L_tab": 389, "n_element_tab": 77, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 613, "n_element_tab_1": 111, "formula_len_all": 105, "formula_len_all_1": 105, "len_all": 122500, "len_all_1": 54527, "len_abs": 959, "len_title": 107, "len_sents": 27432, "len_sents_1": 25286, "n_sents": 192, "n_sents_1": 164, "venue": "ICLR2025", "decision": "withdraw", "L_title": 50, "L_abs": 968, "title": "Scaling Diffusion Models for Downstream Prediction", "abs": "In this paper, we argue that iterative computation, as exemplified by diffusion models, offers a powerful paradigm for not only image generation but also for visual perception tasks. First, we unify few of the mid-level vision tasks as image to image translations tasks ranging from depth estimation to optical flow to segmentation. Then, through extensive experiments across these tasks, we demonstrate how diffusion models scale with increased compute during both training and inference. Notably, we train various dense and Mixture of Expert models up to 2.8 billion parameters, and we utilize increased sampling steps, use various ensembling methods to increase compute at test time. Our work provides compelling evidence for the benefits of scaling compute at train and test time for diffusion models for visual perception, and by studying the scaling properties carefully, we were able to archive same performance of the state-of-the-art models with less compute.", "keywords": [ "Generative Models", "Diffusion" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "YrxhSkfHh0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Ercan Engin KURUOGLU", "gender": "Male", "institution": "Tsinghua-Berkeley Shenzhen Institute ", "country": "CN", "position": "Full Professor" }, { "name": "Hongkang Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Shao-Lun Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 28, "n_formula_1": 21, "n_ref_uni": 29, "n_ref": 37, "n_ref_all": 52, "n_fig": 4, "n_tab": 8, "L_tab": 3939, "n_element_tab": 684, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 718, "n_element_tab_1": 57, "formula_len_all": 2432, "formula_len_all_1": 2144, "len_all": 124749, "len_all_1": 57357, "len_abs": 1211, "len_title": 152, "len_sents": 39414, "len_sents_1": 26789, "n_sents": 267, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 1216, "title": "Fast and Scalable Method for Efficient Multimodal Feature Extraction with Optimized Maximal Correlation", "abs": "This paper introduces the UniFast HGR framework, a novel method designed to enhance the computation of Hirschfeld-Gebelein-R\u00e9nyi (HGR) maximal correlation, specifically optimized for large-scale neural networks and multimodal tasks. UniFast HGR introduces a variance constraint and optimizes the trace term, resulting in a more accurate approximation of the original HGR. By replacing traditional covariance-based measures with cosine similarity and eliminating bias from the main diagonal, the approach significantly reduces computational complexity while enhancing overall accuracy. These improvements make UniFast HGR highly scalable and capable of delivering superior performance in diverse, large-scale multimodal learning applications. Building on this foundation, the OptFast HGR method further optimizes performance by reducing the number of normalization steps, achieving efficiency and computational cost comparable to dot product and cosine similarity operations. This advancement accelerates computation without sacrificing performance. Experimental results indicate that UniFast HGR effectively balances efficiency and precision, establishing it as a robust solution for modern deep learning challenges.", "keywords": [ "HGR maximal correlation", "Soft-HGR", "multimodal feature", "UniFast HGR", "deep learning" ], "rating_list": [ 6, 1, 3 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 1, 4, 5 ] }, { "paper_id": "Yqte21dFVS", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Chandan Singh", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Researcher" }, { "name": "Chenglong Wang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Jeevana Priya Inala", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Jianfeng Gao", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Kai Zhang", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Yu Su", "gender": "Male", "institution": "Ohio State University", "country": "US", "position": "Assistant Professor" }, { "name": "Jianwei Yang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 63, "n_ref_all": 90, "n_fig": 12, "n_tab": 10, "L_tab": 4545, "n_element_tab": 667, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 912, "n_element_tab_1": 85, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 200100, "len_all_1": 58478, "len_abs": 1518, "len_title": 58, "len_sents": 41362, "len_sents_1": 28348, "n_sents": 302, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1525, "title": "Evaluating Graphical Perception of Large Multimodal Models", "abs": "Despite the promising results of large multimodal models (LMMs) in various vision-language tasks, recent benchmarks reveal that these models can struggle with low-level chart perception tasks that require precision.\nHowever, since existing benchmarks primarily focus on end tasks that evaluate models' knowledge and reasoning abilities all together, they provide limited fine-grained insights into how the models' perception abilities affect their performance in chart tasks.\nTo address this gap, we leverage *the theory of graphical perception*, an approach used to study how humans decode visual information encoded on charts and graphs, to develop an evaluation framework for analyzing gaps in LLMs' perception abilities in charts. With automated task generation and response evaluation designs, our framework enables comprehensive and controlled testing of LMMs' graphical perception across diverse chart types, visual elements, and task types.\nWe apply our framework to evaluate the perception capabilities of state-of-the-art LMMs at three granularity levels (chart, visual element, and pixel). Our findings underscore several critical limitations of current state-of-the-art LMMs, including GPT-4o: their inability to (1) generalize across chart types, (2) understand fundamental visual elements, and (3) cross reference values within a chart.\nThese insights provide guidance for future improvements in perception abilities of LMMs.\nThe evaluation framework and labeled data will be publicly available upon acceptance.", "keywords": [ "Large Multimodal Models", "Graphical Perception", "Evaluation" ], "rating_list": [ 3, 3, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "Yqqa9aNwB0", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Juncheng Jia", "gender": "Male", "institution": "Suzhou University", "country": "", "position": "Associate Professor" }, { "name": "Lei You", "gender": "Male", "institution": "Technical University of Denmark", "country": "DK", "position": "Assistant Professor" }, { "name": "Siwei Feng", "gender": "unknown", "institution": "Suzhou University", "country": "", "position": "Assistant Professor" }, { "name": "Tao Deng", "gender": "Male", "institution": "School of Computer Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Zihao Jiang", "gender": "unknown", "institution": "Soochow University", "country": "CN", "position": "MS student" } ], "n_formula": 50, "n_formula_1": 23, "n_ref_uni": 32, "n_ref": 46, "n_ref_all": 62, "n_fig": 5, "n_tab": 5, "L_tab": 2624, "n_element_tab": 222, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1980, "n_element_tab_1": 174, "formula_len_all": 2949, "formula_len_all_1": 1278, "len_all": 132749, "len_all_1": 58696, "len_abs": 1165, "len_title": 168, "len_sents": 39661, "len_sents_1": 25688, "n_sents": 309, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 114, "L_abs": 1172, "title": "CoPruning: Exploring the Parameter-Gradient Nonlinear Correlation for Neural Network Pruning Using Copula Function", "abs": "The sheer size of modern neural networks necessitates pruning techniques to overcome the significant computational challenges posed by model serving. \nHowever, existing pruning techniques fail to capture the nonlinear correlation between parameters and gradient, which is crucial in the pruning process, thus leading to low accuracy under high sparsity.\nIn this work, we propose CoPruning, a new pruning framework, which uses a copula function based \njoint distribution model that precisely captures the intricate nonlinear correlation between parameters and gradient, enabling more insightful pruning decisions. \nAdditionally, we integrate a local optimization approach within CoPruning to better capture relative change in parameters within their local context, providing new metrics for achieving finer-grained optimization.\nExtensive experiments on various networks reveal CoPruning's comparable performance to state-of-the-art (SoTA) pruning algorithms. \nCoPruning outperforms the SoTA with 3.09%, 1.87%, and 2.19% higher accuracy on MLPNet, ResNet20, and ResNet50 at 0.98 sparsity, respectively, and 10.43% higher accuracy on MobileNetV1 at 0.9 sparsity on ImageNet.", "keywords": [ "Neural Network Pruning; Copula Function; Copula Entropy; Sparse Models; Joint Distribution Model;" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 1, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "Yqk7EyT52H", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Chang Xu", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Jiang Bian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Junjie Li", "gender": "Male", "institution": "Research, Microsoft", "country": "CN", "position": "Data Scientist" }, { "name": "Lewen Wang", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Weiqing Liu", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Postdoc" }, { "name": "Yang Liu", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "shikai fang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 1, "n_ref_uni": 59, "n_ref": 95, "n_ref_all": 154, "n_fig": 32, "n_tab": 6, "L_tab": 847, "n_element_tab": 86, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 247, "n_element_tab_1": 31, "formula_len_all": 247, "formula_len_all_1": 79, "len_all": 239768, "len_all_1": 63561, "len_abs": 1346, "len_title": 129, "len_sents": 76321, "len_sents_1": 31541, "n_sents": 621, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1353, "title": "MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model", "abs": "Generative models aim to simulate realistic effects of various actions across different contexts, from text generation to visual effects. Despite significant efforts to build real-world simulators, the application of generative models to virtual worlds, like financial markets, remains under-explored. In financial markets, generative models can simulate complex market effects of participants with various behaviors, enabling interaction under different market conditions, and training strategies without financial risk. This simulation relies on the finest structured data in financial market like orders thus building the finest realistic simulation. We propose Large Market Model (LMM), an order-level generative foundation model, for financial market simulation, akin to language modeling in the digital world. Our financial Market Simulation engine (MarS), powered by LMM, addresses the domain-specific need for realistic, interactive and controllable order generation. Key observations include LMM's strong scalability across data size and model complexity, and MarS's robust and practicable realism in controlled generation with market impact. We showcase MarS as a forecast tool, detection system, analysis platform, and agent training environment, thus demonstrating MarS's ``paradigm shift'' potential for a variety of financial applications.", "keywords": [ "Financial Market Simulation", "Generative Foundation Model", "Large Market Model (LMM)", "Controllable Simulation", "Interactive Simulation", "Market Impact", "Reinforcement Learning", "Forecasting", "Market Manipulation Detection", "Order-Level Data" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 4, 4, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "Yq8At31hLi", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Dmitry Zubarev", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Eduardo Soares", "gender": "Male", "institution": "International Business Machines", "country": "BR", "position": "Researcher" }, { "name": "Emilio Vital Brazil", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Kristin Schmidt", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Researcher" }, { "name": "Renato Cerqueira", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Principal Researcher" }, { "name": "Victor Shirasuna", "gender": "Male", "institution": "Universidade de S\u00e3o Paulo", "country": "BR", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 25, "n_ref": 35, "n_ref_all": 56, "n_fig": 3, "n_tab": 18, "L_tab": 5230, "n_element_tab": 773, "n_fig_1": 3, "n_tab_1": 9, "L_tab_1": 2414, "n_element_tab_1": 333, "formula_len_all": 493, "formula_len_all_1": 447, "len_all": 140891, "len_all_1": 63164, "len_abs": 955, "len_title": 115, "len_sents": 33904, "len_sents_1": 27190, "n_sents": 219, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 965, "title": "SMI-TED: A large-scale foundation model for materials and chemistry", "abs": "We present SMI-TED (SMILE Transformer Encoder Decoder), a large-scale foundation model for materials and chemistry, trained on a massive dataset of 91 million SMILES samples (4 billion molecular tokens) from PubChem using self-supervised learning. Our encoder-decoder architecture enables a wide range of complex tasks, including the prediction of quantum chemical properties and reaction yields. We offer two model variants, with 289M and $8 \\times 289M$ parameters, respectively, to accommodate different use cases. Our model achieves state-of-the-art results across multiple benchmark datasets, demonstrating its versatility and effectiveness. Notably, our model's latent space exhibits compositionality and separability, essential properties for higher-level reasoning tasks and few-shot learning capabilities. To facilitate further research and applications, we make our model weights and source code publicly available on HuggingFace and GitHub, respectively.", "keywords": [ "SMILES", "foundation model", "molecular property prediction", "classification", "molecular reconstruction", "synthesis yield prediction" ], "rating_list": [ 5, 3, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 4, 4 ] }, { "paper_id": "YpWV7XRmFB", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 7, "author_info_list": [ { "name": "Baolong Bi", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Junfeng Fang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Lingrui Mei", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Pengliang Ji", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" }, { "name": "Shenghua Liu", "gender": "Not Specified", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Xueqi Cheng", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy", "country": "CN", "position": "Full Professor" }, { "name": "Yiwei Wang", "gender": "Male", "institution": "University of California, Merced", "country": "US", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 22, "n_ref_uni": 21, "n_ref": 59, "n_ref_all": 77, "n_fig": 4, "n_tab": 7, "L_tab": 1248, "n_element_tab": 121, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 1650, "n_element_tab_1": 207, "formula_len_all": 981, "formula_len_all_1": 925, "len_all": 113430, "len_all_1": 65833, "len_abs": 1461, "len_title": 125, "len_sents": 29176, "len_sents_1": 27182, "n_sents": 231, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1489, "title": "Decoding by Contrasting Knowledge: Enhancing LLMs' Confidence on Edited Facts", "abs": "The knowledge within large language models (LLMs) may become outdated quickly.\nWhile in-context editing (ICE) is currently the most effective method for knowledge editing (KE), it is constrained by the black-box modeling of LLMs and thus lacks interpretability.\nOur work aims to elucidate the superior performance of ICE in KE by analyzing the impacts of in-context new knowledge on token-wise distributions.\nWe observe that despite a significant boost in logits of the new knowledge, the performance of ICE is still hindered by stubborn knowledge. \nStubborn knowledge refers to facts that have gained excessive confidence during pretraining, making them hard to edit effectively.\nTo address this issue and further enhance the performance of ICE, we propose a novel approach termed **De**coding by **C**ontrasting **K**nowledge (**DeCK**).\nDeCK derives the distribution of the next token by contrasting the logits obtained from the newly edited knowledge guided by ICE with those from the unedited parametric knowledge. \nOur experiments consistently demonstrate that DeCK enhances the confidence of LLMs in edited facts. \nFor instance, it improves the performance of LLaMA3-8B-instruct on MQuAKE by up to 219\\%, demonstrating its capability to strengthen ICE in the editing of stubborn knowledge.\nDeCK can be easily integrated into any ICE method as a decoding component to enhance editing capabilities.\nOur work paves the way to develop both effective and accountable KE methods for LLMs.", "keywords": [ "Large Language Models", "Knowledge Enhancement", "Knowledge Editing" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "YomQ3llPD2", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Eugene Belilovsky", "gender": "Male", "institution": "Concordia University, Montreal", "country": "CA", "position": "Assistant Professor" }, { "name": "Michael Eickenberg", "gender": "Male", "institution": "Flatiron Institute", "country": "US", "position": "Researcher" }, { "name": "geraldin nanfack", "gender": "Male", "institution": "Concordia University", "country": "CA", "position": "Postdoc" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 35, "n_ref": 83, "n_ref_all": 144, "n_fig": 34, "n_tab": 1, "L_tab": 659, "n_element_tab": 22, "n_fig_1": 19, "n_tab_1": 2, "L_tab_1": 867, "n_element_tab_1": 50, "formula_len_all": 933, "formula_len_all_1": 622, "len_all": 140066, "len_all_1": 71487, "len_abs": 1303, "len_title": 123, "len_sents": 48212, "len_sents_1": 34511, "n_sents": 372, "n_sents_1": 252, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1310, "title": "From Feature Visualization to Visual Circuits: Effect of Model Perturbation", "abs": "Understanding the inner workings of large-scale deep neural networks is challenging yet crucial in several high-stakes applications. Mechanistic interpretability is an emergent field that tackles this challenge, often by identifying human-understandable subgraphs in deep neural networks known as circuits. In vision-pretrained models, these subgraphs are typically interpreted by visualizing their node features through a popular technique called feature visualization. Recent works have analyzed the stability of different feature visualization types under the adversarial model manipulation framework. This paper addresses limitations in existing works by proposing a novel attack called ProxPulse that simultaneously manipulates two types of feature visualizations. Surprisingly, when analyzing these attacks within the context of visual circuits, we find that visual circuits exhibit some robustness to ProxPulse. Consequently, we introduce a new attack based on ProxPulse that reveals the manipulability of visual circuits, highlighting their lack of robustness. The effectiveness of these attacks is validated across a range of pre-trained models, from smaller architectures like AlexNet to medium-scale models like ResNet-50, and larger ones such as ResNet-152 and DenseNet-201 on the ImageNet dataset.", "keywords": [ "feature visualization", "visual circuits", "robustness of interpretability", "adversarial model manipulation" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "YnJnY7O1PT", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Di Wang", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Difei Xu", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "MS student" }, { "name": "Jinhui Xu", "gender": "Male", "institution": "University at Buffalo, State University of New York", "country": "US", "position": "Full Professor" }, { "name": "Meng Ding", "gender": "Female", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Zihang Xiang", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" } ], "n_formula": 98, "n_formula_1": 39, "n_ref_uni": 20, "n_ref": 50, "n_ref_all": 66, "n_fig": 0, "n_tab": 2, "L_tab": 390, "n_element_tab": 40, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 91, "n_element_tab_1": 20, "formula_len_all": 13936, "formula_len_all_1": 4898, "len_all": 150927, "len_all_1": 68001, "len_abs": 490, "len_title": 145, "len_sents": 37583, "len_sents_1": 27472, "n_sents": 400, "n_sents_1": 238, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1577, "title": "Private Stochastic Convex Optimization with Tysbakov Noise Condition and Large Lipschitz Constant", "abs": "We study Stochastic Convex Optimization in Differential Privacy model (DP-SCO). Unlike previous studies, here we assume the population risk function satisfies\nthe Tysbakov Noise Condition (TNC) with some parameter $\\theta>1$, where the Lipschitz constant of the loss could be extremely large or even unbounded, but the $\\ell_2$-norm gradient of the loss has bounded $k$-th moment with $k\\geq 2$. \nFor the Lipschitz case with $\\theta\\geq 2$, we first propose an $(\\epsilon, \\delta)$-DP algorithms whose utility bound is $\\tilde{O}\\left(\\left(\\tilde{r}_{2k}(\\frac{1}{\\sqrt{n}}+(\\frac{\\sqrt{d}}{n\\epsilon}))^\\frac{k-1}{k}\\right)^\\frac{\\theta}{\\theta-1}\\right)$\n \nin high probability, where $n$ is the sample size, $d$ is the model dimension, and $\\tilde{r}_{2k}$ is a term that only depends on the $2k$-th moment of the gradient. \n\nIt is notable that such an upper bound is independent of the Lipschitz constant. We then extend to the case where \n $\\theta\\geq \\bar{\\theta}> 1$ for some known constant $\\bar{\\theta}$. Moreover, when the privacy budget $\\epsilon$ is small enough, we show an upper bound of $\\tilde{O}\\left(\\left(\\tilde{r}_{k}(\\frac{1}{\\sqrt{n}}+(\\frac{\\sqrt{d}}{n\\epsilon}))^\\frac{k-1}{k}\\right)^\\frac{\\theta}{\\theta-1}\\right)$ \n\neven if the loss function is not Lipschitz. For the lower bound, we show that for any $\\theta\\geq 2$, the private minimax rate for $\\rho$-zero Concentrated Differential Privacy is lower bounded by $\\Omega\\left(\\left(\\tilde{r}_{k}(\\frac{1}{\\sqrt{n}}+(\\frac{\\sqrt{d}}{n\\sqrt{\\rho}}))^\\frac{k-1}{k}\\right)^\\frac{\\theta}{\\theta-1}\\right)$.", "keywords": [ "Stochastic Convex Optimization", "Differential Privacy" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 1 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 1 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "Ym2RNPX6la", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Aaditya Ramdas", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Henny Admoni", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Michelle D Zhao", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Reid Simmons", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Researcher" }, { "name": "Andrea Victoria Bajcsy", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 34, "n_ref": 57, "n_ref_all": 80, "n_fig": 5, "n_tab": 1, "L_tab": 86, "n_element_tab": 9, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 579, "formula_len_all_1": 606, "len_all": 140025, "len_all_1": 75537, "len_abs": 1892, "len_title": 139, "len_sents": 41437, "len_sents_1": 33494, "n_sents": 307, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1900, "title": "Conformalized Interactive Imitation Learning: Handling Expert Shift and Intermittent Feedback", "abs": "In interactive imitation learning (IL), uncertainty quantification offers a way for the learner (i.e. robot) to contend with distribution shifts encountered during deployment by actively seeking additional feedback from an expert (i.e. human) online. Prior works use mechanisms like ensemble disagreement or Monte Carlo dropout to quantify when black-box IL policies are uncertain; however, these approaches can lead to overconfident estimates when faced with deployment-time distribution shifts. Instead, we contend that we need uncertainty quantification algorithms that can leverage the expert human feedback received during deployment time to adapt the robot's uncertainty online. To tackle this, we draw upon online conformal prediction, a distribution-free method for constructing prediction intervals online given a stream of ground-truth labels. Human labels, however, are intermittent in the interactive IL setting. Thus, from the conformal prediction side, we introduce a novel uncertainty quantification algorithm called intermittent quantile tracking (IQT) that leverages a probabilistic model of intermittent labels, maintains asymptotic coverage guarantees, and empirically achieves desired coverage levels. From the interactive IL side, we develop ConformalDAgger, a new approach wherein the robot uses prediction intervals calibrated by IQT as a reliable measure of deployment-time uncertainty to actively query for more expert feedback. We compare ConformalDAgger to prior uncertainty-aware DAgger methods in scenarios where the distribution shift is (and isn't) present because of changes in the expert's policy. We find that in simulated and hardware deployments on a 7DOF robotic manipulator, ConformalDAgger detects high uncertainty when the expert shifts and increases the number of interventions compared to baselines, allowing the robot to more quickly learn the new behavior.", "keywords": [ "conformal prediction", "interactive imitation learning" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 4, 2 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "Ylk98vWQuQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Cheng Perng Phoo", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Jinsu Yoo", "gender": "Male", "institution": "The Ohio State University", "country": "US", "position": "PhD student" }, { "name": "Kilian Q Weinberger", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Professor" }, { "name": "Mark Campbell", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Full Professor" }, { "name": "Tai-Yu Pan", "gender": "unknown", "institution": "Ohio State University", "country": "US", "position": "PhD student" }, { "name": "Wei-Lun Chao", "gender": "Male", "institution": "Ohio State University", "country": "US", "position": "Assistant Professor" }, { "name": "Xiangyu Chen", "gender": "Male", "institution": "Waymo LLC", "country": "US", "position": "Researcher" }, { "name": "Yihong Sun", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Zhenyang Feng", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Undergrad student" }, { "name": "Bharath Hariharan", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 53, "n_ref": 126, "n_ref_all": 169, "n_fig": 14, "n_tab": 8, "L_tab": 1074, "n_element_tab": 105, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 590, "n_element_tab_1": 71, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 187403, "len_all_1": 78133, "len_abs": 1876, "len_title": 95, "len_sents": 54868, "len_sents_1": 36974, "n_sents": 459, "n_sents_1": 305, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1886, "title": "Learning 3D Perception from Others' Predictions", "abs": "Accurate 3D object detection in real-world environments requires a huge amount of annotated data with high quality. Acquiring such data is tedious and expensive, and often needs repeated effort when a new sensor is adopted or when the detector is deployed in a new environment. We investigate a new scenario to construct 3D object detectors: *learning from the predictions of a nearby unit that is equipped with an accurate detector.* For example, when a self-driving car enters a new area, it may learn from other traffic participants whose detectors have been optimized for that area. This setting is label-efficient, sensor-agnostic, and communication-efficient: nearby units only need to share the predictions with the ego agent (e.g., car). Naively using the received predictions as ground-truths to train the detector for the ego car, however, leads to inferior performance. We systematically study the problem and identify viewpoint mismatches and mislocalization (due to synchronization and GPS errors) as the main causes, which unavoidably result in false positives, false negatives, and inaccurate pseudo labels. We propose a distance-based curriculum, first learning from closer units with similar viewpoints and subsequently improving the quality of other units' predictions via self-training. We further demonstrate that an effective pseudo label refinement module can be trained with a handful of annotated data, largely reducing the data quantity necessary to train an object detector. We validate our approach on the recently released real-world collaborative driving dataset, using reference cars' predictions as pseudo labels for the ego car. Extensive experiments including several scenarios (e.g., different sensors, detectors, and domains) demonstrate the effectiveness of our approach toward label-efficient learning of 3D perception from other units' predictions.", "keywords": [ "3D object detection", "autonomous driving", "label-efficient learning", "domain adaptation", "curriculum learning", "collaborative perception" ], "rating_list": [ 6, 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 3, 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 2, 5 ] }, { "paper_id": "YlWvQSBCgl", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Tong Yang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Wenqiang Zhang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Xiangyu Zhang", "gender": "Male", "institution": "MEGVII Technology", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 30, "n_ref": 78, "n_ref_all": 91, "n_fig": 3, "n_tab": 7, "L_tab": 3387, "n_element_tab": 440, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 3058, "n_element_tab_1": 333, "formula_len_all": 328, "formula_len_all_1": 314, "len_all": 110965, "len_all_1": 58490, "len_abs": 867, "len_title": 96, "len_sents": 22973, "len_sents_1": 22000, "n_sents": 210, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 873, "title": "Image Generation with Channel-wise Quantization", "abs": "We present a novel image generation model with channel-wise quantization. Our method quantizes image feature along channel into discrete codes. Then based on the learned codes, our approach adopts masked-prediction paradigm for image generation. Compared with widely used spatial tokenizers, our channel-wise tokenizer has an efficient modeling for image structure and strong representational capacity. Besides, the codebook usage of our tokenizer can reach 100\\% under different codebook size. Using the channel-wise tokenizer, our generation framework achieves competitive performances on various benchmarks of image generation. In particular, on ImageNet 256x256 benchmark, our method significantly improve baseline by improving Frechet inception distance (FID) to 1.87. Furthermore, we also validate the effectiveness of our proposed method on text-to-image generation.", "keywords": [ "generative models", "image generation", "visual tokenization" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "YkmbJSHjj7", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 1, "author_info_list": [ { "name": "Shang Wang", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 37, "n_ref": 88, "n_ref_all": 121, "n_fig": 7, "n_tab": 13, "L_tab": 4641, "n_element_tab": 486, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 3433, "n_element_tab_1": 353, "formula_len_all": 576, "formula_len_all_1": 521, "len_all": 172870, "len_all_1": 68213, "len_abs": 1619, "len_title": 132, "len_sents": 49230, "len_sents_1": 27993, "n_sents": 372, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1587, "title": "W-PCA Based Gradient-Free Proxy for Efficient Search of Lightweight Language Models", "abs": "The demand for efficient natural language processing (NLP) systems has led to the development of lightweight language models. Previous work in this area has primarily focused on manual design or training-based neural architecture search (NAS) methods. Recently, zero-shot NAS methods have been proposed for evaluating language models without the need for training. However, prevailing approaches to zero-shot NAS often face challenges such as biased evaluation metrics and computational inefficiencies.\nIn this paper, we introduce weight-weighted PCA (W-PCA), a novel zero-shot NAS method specifically tailored for lightweight language models. Our approach utilizes two evaluation proxies: the parameter count and the number of principal components with cumulative contribution exceeding $\\eta$ in the feed-forward neural (FFN) layer. Additionally, by eliminating the need for gradient computations, we optimize the evaluation time, thus enhancing the efficiency of designing and evaluating lightweight language models.\nWe conduct a comparative analysis on the GLUE and SQuAD datasets to evaluate our approach. The results demonstrate that our method significantly reduces training time compared to one-shot NAS methods and achieves higher scores in the testing phase compared to previous state-of-the-art training-based methods. Furthermore, we perform ranking evaluations on a dataset sampled from the FlexiBERT search space. Our approach exhibits superior ranking correlation and further reduces solving time compared to other zero-shot NAS methods that require gradient computation.", "keywords": [ "zero-shot NAS", "gradient-free", "lightweight language models" ], "rating_list": [ 6, 8, 8, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 2, 2, 4, 3 ] }, { "paper_id": "YkMg8sB8AH", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Haiyang Yu", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Hongyi Ling", "gender": "unknown", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Na Zou", "gender": "Female", "institution": "University of Houston", "country": "US", "position": "Assistant Professor" }, { "name": "Shuiwang Ji", "gender": "Male", "institution": "Texas A&M University", "country": "US", "position": "Professor" }, { "name": "Zhimeng Jiang", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 25, "n_formula_1": 26, "n_ref_uni": 46, "n_ref": 64, "n_ref_all": 83, "n_fig": 9, "n_tab": 4, "L_tab": 725, "n_element_tab": 111, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1237, "n_element_tab_1": 31, "formula_len_all": 1338, "formula_len_all_1": 1615, "len_all": 142667, "len_all_1": 61695, "len_abs": 1541, "len_title": 123, "len_sents": 37303, "len_sents_1": 27519, "n_sents": 286, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1354, "title": "On Explaining Equivariant Graph Networks via Improved Relevance Propagation", "abs": "We consider explainability in equivariant graph neural networks for 3D geometric graphs. While many XAI methods have been developed for analyzing graph neural networks, they predominantly target 2D graph structures. The complex nature of 3D data and the sophisticated architectures of equivariant GNNs present unique challenges. Current XAI techniques either struggle to adapt to equivariant GNNs or fail to effectively handle positional data and evaluate the significance of geometric features adequately. To address these challenges, we introduce a novel method, known as EquiGX, which uses the Deep Taylor decomposition framework to extend the layer-wise relevance propagation rules tailored for spherical equivariant GNNs. Our approach decomposes prediction scores and back-propagates the relevance scores through each layer to the input space. Our decomposition rules provide a detailed explanation of each layer\u2019s contribution to the network\u2019s predictions, thereby enhancing our understanding of how geometric and positional data influence the model\u2019s outputs. Through experiments on both synthetic and real-world datasets, our method demonstrates its capability to identify critical geometric structures and outperform alternative baselines. These results indicate that our method provides significantly enhanced explanations for equivariant GNNs.", "keywords": [ "Explainability", "Equivariant Graph Networks" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "Yk87CwhBDx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Adrian Weller", "gender": "Male", "institution": "Alan Turing Institute", "country": "", "position": "Programme Director for AI" }, { "name": "Bernhard Schoelkopf", "gender": "Male", "institution": "ELLIS Institute", "country": "DE", "position": "Principal Researcher" }, { "name": "Haiwen Feng", "gender": "Male", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Joshua B. Tenenbaum", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Professor" }, { "name": "Katherine M. Collins", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Weiyang Liu", "gender": "Male", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "Postdoc" }, { "name": "Zeju Qiu", "gender": "Male", "institution": "Max-Planck-Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "Intern" }, { "name": "Zhen Liu", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "PhD student" }, { "name": "Michael J Black", "gender": "Male", "institution": "University of T\u00fcbingen", "country": "DE", "position": "Honorarprofessor" }, { "name": "Tim Xiao", "gender": "Male", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 99, "n_ref": 319, "n_ref_all": 355, "n_fig": 12, "n_tab": 5, "L_tab": 4127, "n_element_tab": 516, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 3213, "n_element_tab_1": 465, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 446697, "len_all_1": 105593, "len_abs": 2927, "len_title": 112, "len_sents": 87595, "len_sents_1": 43819, "n_sents": 739, "n_sents_1": 446, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 64, "L_abs": 1926, "title": "Can Large Language Models Understand Symbolic Graphics Programs?", "abs": "Against the backdrop of enthusiasm for large language models (LLMs), there is an urgent need to scientifically assess their capabilities and shortcomings. This is nontrivial in part because it is difficult to find tasks which the models have not encountered during training.\nUtilizing symbolic graphics programs, we propose a domain well-suited to test multiple spatial-semantic reasoning skills of LLMs. Popular in computer graphics, these programs procedurally generate visual data. While LLMs exhibit impressive skills in general program synthesis and analysis, symbolic graphics programs offer a new layer of evaluation: they allow us to test an LLM's ability to answer different-grained semantic-level questions of the images or 3D geometries without a vision encoder. To semantically understand the symbolic programs, LLMs would need to possess the ability to \"imagine\" and reason how the corresponding graphics content would look with only the symbolic description of the local curvatures and strokes. We use this task to evaluate LLMs by creating a large benchmark for the semantic visual understanding of symbolic graphics programs, built procedurally with minimal human effort. Particular emphasis is placed on transformations of images that leave the image level semantics invariant while introducing significant changes to the underlying program. We evaluate commercial and open-source LLMs on our benchmark to assess their ability to reason about visual output of programs, finding that LLMs considered stronger at reasoning generally perform better. Lastly, we introduce a novel method to improve this ability -- Symbolic Instruction Tuning (SIT), in which the LLM is finetuned with pre-collected instruction data on symbolic graphics programs. Interestingly, we find that SIT not only improves LLM's understanding on symbolic programs, but it also improves general reasoning ability on various other benchmarks.", "keywords": [ "Large Language Models", "Symbolic Graphics Programs" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 3, 4, 3 ], "confidence_list": [ 3, 5, 4 ] }, { "paper_id": "Yj6IdXSOZk", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Fangcheng Zhong", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "Lecturer" }, { "name": "Jing Yang", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "Param Hanji", "gender": "Male", "institution": "Apple", "country": "GB", "position": "Researcher" }, { "name": "SUCHONG", "gender": "Male", "institution": "Ke Holdings Inc", "country": "CN", "position": "Software Engineer" }, { "name": "Yingbin Fu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "wangshaojun", "gender": "Male", "institution": "Sichuan Agricultural University", "country": "CN", "position": "Software Engineer" }, { "name": "zhaoxuan", "gender": "unknown", "institution": "\u8d1d\u58f3\u627e\u623f\uff08\u5317\u4eac\uff09\u79d1\u6280\u6709\u9650\u516c\u53f8", "country": "CN", "position": "Development engineer" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 36, "n_ref": 105, "n_ref_all": 134, "n_fig": 11, "n_tab": 4, "L_tab": 1011, "n_element_tab": 114, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1104, "n_element_tab_1": 123, "formula_len_all": 147, "formula_len_all_1": 147, "len_all": 116443, "len_all_1": 61341, "len_abs": 1031, "len_title": 160, "len_sents": 35882, "len_sents_1": 29530, "n_sents": 199, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Reject", "L_title": 112, "L_abs": 1033, "title": "CF-GISS: Collision-Free Generative 3D Indoor Scene Synthesis with Controllable Floor Plans and Optimized Layouts", "abs": "We introduce CF-GISS, a novel framework for generative 3D indoor scene synthesis that ensures collision-free scene layouts by incorporating an image-based intermediate layout representation. In contrast to existing methods that directly construct the scene graph or object list, our approach facilitates substantially more effective prevention of collision artifacts as out-of-distribution (OOD) scenarios during generation. Furthermore, CF-GISS conditions layout generation on floor plans controllable via images or textual descriptions, enabling the production of coherent, house-wide layouts that are robust to variations in geometric and semantic structures. Our framework demonstrates state-of-the-art performance on the 3D-FRONT dataset, delivering high-quality, collision-free scene synthesis while offering flexibility in accommodating a range of floor plan structures. Additionally, we propose a novel dataset with significantly expanded coverage of household items and room configurations, as well as improved data quality.", "keywords": [ "Indoor Scene Synthesis", "3D scene generation", "Procedural generation", "Generative models" ], "rating_list": [ 3, 6, 8, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "YiyG1tHDxq", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Lars Schmidt-Thieme", "gender": "Male", "institution": "University of Hildesheim", "country": "DE", "position": "Full Professor" }, { "name": "Thorben Werner", "gender": "Male", "institution": "Universit\u00e4t Hildesheim", "country": "", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 14, "n_ref_uni": 15, "n_ref": 44, "n_ref_all": 60, "n_fig": 5, "n_tab": 3, "L_tab": 573, "n_element_tab": 78, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 573, "n_element_tab_1": 78, "formula_len_all": 611, "formula_len_all_1": 633, "len_all": 78364, "len_all_1": 57569, "len_abs": 966, "len_title": 102, "len_sents": 28176, "len_sents_1": 26325, "n_sents": 200, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 941, "title": "Bayesian Active Learning By Distribution Disagreement", "abs": "Active Learning (AL) for regression has been systematically under-researched due to the increased difficulty of measuring uncertainty in regression models.\nSince normalizing flows offer a full predictive distribution instead of a point forecast, they facilitate direct usage of known heuristics for AL like Entropy or Least-Confident sampling.\nHowever, we show that most of these heuristics do not work well for normalizing flows in pool-based AL and we need more sophisticated algorithms to distinguish between aleatoric and epistemic uncertainty.\nIn this work we propose BALSA, an adaptation of the BALD algorithm, tailored for regression with normalizing flows.\nWith this work we extend current research on uncertainty quantification with normalizing flows to real world data and pool-based AL with multiple acquisition functions and query sizes.\nWe report SOTA results for BALSA across 4 different datasets and 2 different architectures.", "keywords": [ "Active Learning", "Normalizing Flows", "Regression", "Uncertainty Quantification" ], "rating_list": [ 3, 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 1, 2, 2 ], "presentation_list": [ 1, 3, 2, 1, 2 ], "contribution_list": [ 1, 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 5, 4, 2 ] }, { "paper_id": "YixNDE12wm", "primary_area": "foundation or frontier models, including LLMs", "n_author": 12, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Carl Yang", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Assistant Professor" }, { "name": "Chulin Xie", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "PhD student" }, { "name": "Dawn Song", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Han Xie", "gender": "unknown", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Jiawei Zhang", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Junyuan Hong", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Postdoc" }, { "name": "Linzhi Zheng", "gender": "Female", "institution": "University of Chicago", "country": "US", "position": "Intern" }, { "name": "Qinbin Li", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yanjie Li", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Zhen Xiang", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Assistant Professor" }, { "name": "Zidi Xiong", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 1, "n_ref_uni": 27, "n_ref": 66, "n_ref_all": 102, "n_fig": 17, "n_tab": 5, "L_tab": 813, "n_element_tab": 45, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 359, "n_element_tab_1": 36, "formula_len_all": 271, "formula_len_all_1": 40, "len_all": 178022, "len_all_1": 68448, "len_abs": 2137, "len_title": 135, "len_sents": 67200, "len_sents_1": 34858, "n_sents": 480, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 2150, "title": "GuardAgent: Safeguard LLM Agent by a Guard Agent via Knowledge-Enabled Reasoning", "abs": "The rapid advancement of large language models (LLMs) has catalyzed the deployment of LLM-powered agents across numerous applications, raising new concerns regarding their safety and trustworthiness. In addition, existing methods for enhancing the safety of LLMs are not directly transferable to LLM-powered agents due to their diverse objectives and output modalities. In this paper, we propose\nGuardAgent, the first LLM agent as a guardrail to protect other LLM agents.\nSpecifically, GuardAgent oversees a target LLM agent by checking whether its\ninputs/outputs satisfy a set of given guard requests, e.g., safety rules or privacy\npolicies defined by the users. The pipeline of GuardAgent consists of two steps: 1) create a task plan by analyzing the provided guard requests, and 2) generate\nguardrail code based on the task plan and execute the code by calling APIs or\nusing external engines. In both steps, an LLM is utilized as the core reasoning\ncomponent, supplemented by in-context demonstrations retrieved from a memory\nmodule storing information from previous sessions. Such knowledge-enabled reasoning of GuardAgent allows it to understand various textual guard requests and\naccurately \u201ctranslate\u201d them into executable code that provides reliable guardrails.\nFurthermore, GuardAgent is equipped with an extendable toolbox containing\nrelevant APIs and functions, and requires no additional LLM training, underscoring\nits flexibility and low operational overhead. In addition to GuardAgent, we\npropose two novel benchmarks: an EICU-AC benchmark for assessing privacy-\nrelated access control for healthcare agents and a Mind2Web-SC benchmark for\nassessing safety regulations for web agents. When using Llama3-70B/Llama3.1-\n70B/GPT-4 as the core LLM, GuardAgent achieves 98.4%/98.4%/98.7% and\n83.5%/84.5%/90.0% guarding accuracy on these two benchmarks in moderating\ninvalid inputs and outputs of two types of agents, respectively. We also show the\nability of GuardAgent to define necessary functions that are absent from the\ntoolbox, which further highlights the flexibility of GuardAgent in adaption to\nnew LLM agents and guard requirements.", "keywords": [ "LLM agent", "guardrail", "safety", "LLM" ], "rating_list": [ 8, 5, 5, 6 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Yisupq2CgQ", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Fangxiao Jin", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "MS student" }, { "name": "Quan Xue", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Rui Xiao", "gender": "Male", "institution": "South China University of Technology", "country": "", "position": "MS student" }, { "name": "Shijian Zheng", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Shuhai Zhang", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Tianyu Liang", "gender": "Male", "institution": "South China University of Technology", "country": "", "position": "MS student" }, { "name": "Mingkui Tan", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 8, "n_ref_uni": 34, "n_ref": 48, "n_ref_all": 71, "n_fig": 12, "n_tab": 9, "L_tab": 2457, "n_element_tab": 178, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1765, "n_element_tab_1": 71, "formula_len_all": 515, "formula_len_all_1": 304, "len_all": 172935, "len_all_1": 67920, "len_abs": 1474, "len_title": 141, "len_sents": 60141, "len_sents_1": 32341, "n_sents": 466, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1481, "title": "Deep Progressive Search for Electromagnetic Structure Design Under Limited Evaluation Budgets", "abs": "Electromagnetic structure (EMS) design aims to optimize a material distribution, e.g., metals over a printed circuit board, which is crucial for antenna and meta-material. This task, however, is inherently a highly non-convex problem with no explicit objective function, making it extremely challenging to solve. The most common approach to addressing this problem relies on evolutionary algorithms (e.g., Genetic Algorithm), where candidate structures are evaluated through electromagnetic simulation using specialized software. However, these methods struggle with inefficiency, especially when dealing with large structural design space and time-consuming simulations. \nTo address this, we propose a Deep Progressive Search method called DPS, which leverages a Deep Neural Network (DNN) as a surrogate model to identify a satisfactory structure within a limited simulation budget. Specifically, we develop a tree-search-based design space control strategy that models the design space as a tree and incrementally refines it through node expansions, enabling adaptive exploration of more complex regions while leveraging insights from simpler subspaces. Moreover, we introduce a consistency-based sample selection strategy to balance exploration and exploitation. Experiments on two real-world engineering tasks, i.e., Dual-layer Frequency Selective Surface and High-gain Antenna show the effectiveness of the proposed DPS in terms of efficiency under limited evaluation budgets.", "keywords": [ "Electromagnetic Structure", "Surrogate Model", "Tree Search" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "YilY5fGQny", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Chuheng Zhang", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Jiang Bian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Li Zhao", "gender": "Female", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Lianyong Qi", "gender": "Male", "institution": "China University of Petroleum", "country": "CN", "position": "Full Professor" }, { "name": "Wanchun Dou", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Full Professor" }, { "name": "Wei Shen", "gender": "unknown", "institution": "ByteDance Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Xuyun Zhang", "gender": "Male", "institution": "Macquarie University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 48, "n_ref": 74, "n_ref_all": 95, "n_fig": 3, "n_tab": 6, "L_tab": 5170, "n_element_tab": 232, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 150, "n_element_tab_1": 22, "formula_len_all": 776, "formula_len_all_1": 894, "len_all": 177493, "len_all_1": 67893, "len_abs": 1425, "len_title": 110, "len_sents": 52376, "len_sents_1": 33220, "n_sents": 400, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1434, "title": "Policy Filtration in RLHF to Fine-Tune LLM for Code Generation", "abs": "While direct policy optimization methods exist, pioneering LLMs are fine-tuned with reinforcement learning from human feedback (RLHF) to generate better responses under the supervision of a reward model learned from preference data. One major challenge of RLHF is the inaccuracy of the intermediate reward model, especially in code generation tasks that requires complex reasoning for the reward model to score a response. We find that the reliability of the reward model varies across responses assigned with different rewards. This motivates us to filter the samples whose rewards may be unreliable to improve the signal-to-noise ratio during policy learning, resulting in Policy Filtration for Proximal Policy Optimization (PF-PPO). To choose a proper policy filtering strategy, we use the coefficient of determination ($R^2$) between the rewards and actual scores on filtered samples as the metrics to help us find promising strategies since it measures how well the rewards filtered by PF-PPO indicate real performance. We provide extensive experiments to validate the effectiveness of PF-PPO in code generation tasks. We find that some variants of PF-PPO are highly effective and achieve the state-of-the-art performance of 7-billion-parameter models on HumanEval (+7.9\\%) and MBPP (+0.7\\%). Moreover, we create the LeetCode Contest benchmark and demonstrate the advantage of PF-PPO (+10.0\\%) on this more challenging benchmark.", "keywords": [ "LLM", "RLHF" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 1, 2, 1 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "YhfrKB3Ah7", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Daolang Huang", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "PhD student" }, { "name": "Julien Martinelli", "gender": "Male", "institution": "University of Bordeaux", "country": "FR", "position": "Postdoc" }, { "name": "Samuel Kaski", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "Full Professor" }, { "name": "Xinyu Zhang", "gender": "Female", "institution": "Aalto University", "country": "FI", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 21, "n_ref_uni": 50, "n_ref": 86, "n_ref_all": 111, "n_fig": 8, "n_tab": 4, "L_tab": 752, "n_element_tab": 210, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 254, "n_element_tab_1": 40, "formula_len_all": 1850, "formula_len_all_1": 1650, "len_all": 164205, "len_all_1": 72194, "len_abs": 1287, "len_title": 100, "len_sents": 45409, "len_sents_1": 30964, "n_sents": 338, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 52, "L_abs": 1143, "title": "PABBO: Preferential Amortized Black-Box Optimization", "abs": "Preferential Bayesian Optimization (PBO) is a sample-efficient method to learn latent user utilities from preferential feedback over a pair of designs. It relies on a statistical surrogate model for the latent function, usually a Gaussian process, and an acquisition strategy to select the next candidate pair to get user feedback on. Due to the non-conjugacy of the associated likelihood, every PBO step requires a significant amount of computations with various approximate inference techniques. This computational overhead is incompatible with the way humans interact with computers, hindering the use of PBO in real-world cases. Building on the recent advances of amortized BO, we propose to circumvent this issue by fully amortizing PBO, meta-learning both the surrogate and the acquisition function. Our method comprises a novel transformer neural process architecture, trained using reinforcement learning and tailored auxiliary losses.\nOn a benchmark composed of synthetic and real-world datasets, our method is several orders of magnitude faster than the usual Gaussian process-based strategies and often outperforms them in accuracy.", "keywords": [ "Bayesian optimization", "preference learning", "amortized inference", "neural processes" ], "rating_list": [ 5, 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 4, 4 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4, 3 ] }, { "paper_id": "YhIpTdrUDY", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Daoqiang Zhang", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Emadeldeen Eldele", "gender": "Male", "institution": "Agency for Science, Technology and Research", "country": "SG", "position": "Research Scientist" }, { "name": "Min Wu", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Peiliang Gong", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Xiaoli Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Zhenghua Chen", "gender": "Male", "institution": "I2R, A*STAR", "country": "SG", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 10, "n_ref_uni": 56, "n_ref": 78, "n_ref_all": 95, "n_fig": 7, "n_tab": 9, "L_tab": 1424, "n_element_tab": 252, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 941, "n_element_tab_1": 149, "formula_len_all": 490, "formula_len_all_1": 314, "len_all": 167778, "len_all_1": 61489, "len_abs": 1165, "len_title": 73, "len_sents": 44238, "len_sents_1": 29169, "n_sents": 311, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1170, "title": "Towards Adaptive Time Series Foundation Models Against Distribution Shift", "abs": "Foundation models have demonstrated remarkable success across diverse machine-learning domains through large-scale pretraining. However, their application to time series data poses challenges due to substantial mismatches in the distributions of pretraining datasets. In this paper, we tackle this issue by proposing a domain-aware adaptive normalization strategy within the Transformer architecture. Specifically, we replace the traditional LayerNorm with a prototype-guided dynamic normalization mechanism, where learned prototypes represent distinct data distributions, and sample-to-prototype similarity determines the appropriate normalization layer. This approach effectively captures the diverse characteristics of time series data, ensuring better alignment between pretrained representations and downstream tasks. Our method significantly improves fine-tuning performance, outperforming vanilla pretraining techniques and reducing the negative impact of distribution shifts. Extensive experiments on various real-world time series datasets demonstrate the efficacy of our approach, paving the way for more robust and generalizable time series foundation models.", "keywords": [ "Time series", "Pretraining", "Distribution Shifts", "Foundation Model" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "Yg998afEbH", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Kaustubh Ponkshe", "gender": "Male", "institution": "Indian Institute of Technology Bombay, Indian Institute of Technology, Bombay", "country": "", "position": "Undergrad student" }, { "name": "Praneeth Vepakomma", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Raghav Singhal", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 58, "n_ref": 101, "n_ref_all": 118, "n_fig": 7, "n_tab": 8, "L_tab": 1968, "n_element_tab": 209, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1993, "n_element_tab_1": 213, "formula_len_all": 974, "formula_len_all_1": 971, "len_all": 156692, "len_all_1": 59318, "len_abs": 1175, "len_title": 90, "len_sents": 37023, "len_sents_1": 24994, "n_sents": 283, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1180, "title": "FedEx-LoRA: Exact Aggregation for Federated and Efficient Fine-Tuning of Foundation Models", "abs": "Low-Rank Adaptation (LoRA) is a popular technique for efficient fine-tuning of foundation models. However, applying LoRA in federated learning environments, where data is distributed across multiple clients, presents unique challenges. Existing methods rely on traditional federated averaging of LoRA adapters, resulting in inexact updates. To address this, we propose Federated Exact LoRA, or FedEx-LoRA, which adds a residual error term to the pretrained frozen weight matrix. Our approach achieves exact updates with minimal computational and communication overhead, preserving LoRA\u2019s efficiency. We evaluate the method on various models across arithmetic reasoning, commonsense reasoning, natural language understanding and natural language generation tasks, showing consistent performance gains over state-of-the-art methods across multiple settings. Through extensive analysis, we quantify that the deviations in updates from the ideal solution are significant, highlighting the need for exact aggregation. Our method's simplicity, efficiency, and broad applicability position it as a promising solution for accurate and effective federated fine-tuning of foundation models.", "keywords": [ "Parameter-efficient fine-tuning", "low-rank adaptation", "federated learning", "fine-tuning" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 4, 4, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "YfKNaRktan", "primary_area": "datasets and benchmarks", "n_author": 16, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Boyi Wei", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Dacheng Li", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Danqi Chen", "gender": "Female", "institution": "Department of Computer Science, Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Kai Li", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Full Professor" }, { "name": "Kaixuan Huang", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Luxi He", "gender": "Female", "institution": "Department of Computer Science, Princeton University", "country": "US", "position": "PhD student" }, { "name": "Peter Henderson", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Prateek Mittal", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "Full Professor" }, { "name": "Ruoxi Jia", "gender": "unknown", "institution": "Virginia Tech", "country": "US", "position": "Assistant Professor" }, { "name": "Tinghao Xie", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Xiangyu Qi", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Yangsibo Huang", "gender": "Female", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Yi Zeng", "gender": "Male", "institution": "Virginia Tech", "country": "", "position": "PhD student" }, { "name": "Ying Sheng", "gender": "Female", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Udari Madhushani", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Postdoc" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 51, "n_ref": 177, "n_ref_all": 209, "n_fig": 5, "n_tab": 11, "L_tab": 9022, "n_element_tab": 824, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1130, "n_element_tab_1": 163, "formula_len_all": 541, "formula_len_all_1": 0, "len_all": 304077, "len_all_1": 72638, "len_abs": 1774, "len_title": 201, "len_sents": 99386, "len_sents_1": 34902, "n_sents": 699, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1793, "title": "SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal", "abs": "Evaluating aligned large language models' (LLMs) ability to recognize and reject unsafe user requests is crucial for safe, policy-compliant deployments. Existing evaluation efforts, however, face three limitations that we address with **SORRY-Bench**, our proposed benchmark. **First**, existing methods often use coarse-grained taxonomies of unsafe topics, and are over-representing some fine-grained topics. For example, among the ten existing datasets that we evaluated, tests for refusals of self-harm instructions are over 3x less represented than tests for fraudulent activities. SORRY-Bench improves on this by using a fine-grained taxonomy of 44 potentially unsafe topics, and 440 class-balanced unsafe instructions, compiled through human-in-the-loop methods. **Second**, evaluations often overlook the linguistic formatting of prompts, like different languages, dialects, and more --- which are only implicitly considered in many evaluations. We supplement SORRY-bench with 20 diverse linguistic augmentations to systematically examine these effects. **Third**, existing evaluations rely on large LLMs (e.g., GPT-4) for evaluation, which can be computationally expensive. We investigate design choices for creating a fast, accurate automated safety evaluator. By collecting 7K+ human annotations and conducting a meta-evaluation of diverse LLM-as-a-judge designs, we show that fine-tuned 7B LLMs can achieve accuracy comparable to GPT-4 scale LLMs, with lower computational cost. Putting these together, we evaluate over 50 proprietary and open-weight LLMs on SORRY-Bench, analyzing their distinctive safety refusal behaviors. We hope our effort provides a building block for systematic evaluations of LLMs' safety refusal capabilities, in a balanced, granular, and efficient manner.", "keywords": [ "LLM", "safety", "alignment", "benchmark", "refusal" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 3, 3, 2 ] }, { "paper_id": "YeZNN6Iy6Q", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Kunal Singh", "gender": "Male", "institution": "Fractal Analytics", "country": "IN", "position": "Researcher" }, { "name": "Mukund Khanna", "gender": "Male", "institution": "Fractal Analytics", "country": "IN", "position": "Researcher" }, { "name": "Pradeep Moturi", "gender": "Male", "institution": "Fractal AI", "country": "IN", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 19, "n_ref": 56, "n_ref_all": 70, "n_fig": 8, "n_tab": 3, "L_tab": 704, "n_element_tab": 102, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 598, "n_element_tab_1": 62, "formula_len_all": 48, "formula_len_all_1": 48, "len_all": 87801, "len_all_1": 52625, "len_abs": 1184, "len_title": 114, "len_sents": 28217, "len_sents_1": 25257, "n_sents": 199, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1238, "title": "Effective Text-to-Image Alignment with Quality Aware Pair Ranking", "abs": "Fine-tuning techniques such as Reinforcement Learning with Human Feedback (RLHF) and Direct Preference Optimization (DPO) allow us to steer Large Language Models (LLMs) to be align better with human preferences. Alignment is equally important in text-to-image generation. Recent adoption of DPO, specifically Diffusion-DPO, for Text-to-Image (T2I) diffusion models has proven to work effectively in improving visual appeal and prompt-image alignment. The mentioned works fine-tune on Pick-a-Pic dataset, consisting of approximately one million image preference pairs, collected via crowdsourcing at scale. However, do all preference pairs contribute equally to alignment fine-tuning? Preferences can be subjective at times and may not always translate into effectively aligning the model. In this work, we investigate the above-mentioned question. We develop a quality metric to rank image preference pairs and achieve effective Diffusion-DPO-based alignment fine-tuning. We show that the SD-1.5 and SDXL models fine-tuned using the top 5.33\\% of the data perform better both quantitatively and qualitatively than the models fine-tuned on the full dataset. The code is available at https://anonymous.4open.science/r/DPO-QSD-28D7/README.md", "keywords": [ "dpo", "diffusion" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "YeSxbRrDRl", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Gongzheng Tang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Guangkun Nie", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Shenda Hong", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 6, "n_ref": 31, "n_ref_all": 70, "n_fig": 4, "n_tab": 6, "L_tab": 3158, "n_element_tab": 379, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2467, "n_element_tab_1": 357, "formula_len_all": 325, "formula_len_all_1": 326, "len_all": 110996, "len_all_1": 66672, "len_abs": 1616, "len_title": 140, "len_sents": 33856, "len_sents_1": 28868, "n_sents": 208, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1624, "title": "Dist Loss: Enhancing Regression in Few-Shot Region through Distribution Distance Constraint", "abs": "Imbalanced data distributions are prevalent in real-world scenarios, posing significant challenges in both imbalanced classification and imbalanced regression tasks. They often cause deep learning models to overfit in areas of high sample density (many-shot regions) while underperforming in areas of low sample density (few-shot regions). This characteristic restricts the utility of deep learning models in various sectors, notably healthcare, where areas with few-shot data hold greater clinical relevance. While recent studies have shown the benefits of incorporating distribution information in imbalanced classification tasks, such strategies are rarely explored in imbalanced regression. In this paper, we address this issue by introducing a novel loss function, termed Dist Loss, designed to minimize the distribution distance between the model's predictions and the target labels in a differentiable manner, effectively integrating distribution information into model training. Dist Loss enables deep learning models to regularize their output distribution during training, effectively enhancing their focus on few-shot regions. We have conducted extensive experiments across three datasets spanning computer vision and healthcare: IMDB-WIKI-DIR, AgeDB-DIR, and ECG-Ka-DIR. The results demonstrate that Dist Loss effectively mitigates the negative impact of imbalanced data distribution on model performance, achieving state-of-the-art results in sparse data regions. Furthermore, Dist Loss is easy to integrate, complementing existing methods. Our code will be made publicly available following the review process.", "keywords": [ "Deep imbalanced regression", "sparse data region optimization" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 5, 3, 3 ] }, { "paper_id": "YeSwPnI4be", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Chamika Sudusinghe", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Charith Mendis", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Charles Block", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Damitha Lenadora", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Gerasimos Gerogiannis", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Josep Torrellas", "gender": "Male", "institution": "Department of Computer Science", "country": "", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 35, "n_ref": 60, "n_ref_all": 87, "n_fig": 10, "n_tab": 3, "L_tab": 692, "n_element_tab": 111, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 389, "n_element_tab_1": 40, "formula_len_all": 502, "formula_len_all_1": 502, "len_all": 139328, "len_all_1": 65073, "len_abs": 2077, "len_title": 150, "len_sents": 47207, "len_sents_1": 32761, "n_sents": 331, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1694, "title": "DASH: Data-Efficient Learned Cost Models for Sparse Matrix Computations on Emerging Hardware Platforms", "abs": "Sparse matrix computations are becoming increasingly significant in deep learning and graph analytics, driving the development of specialized hardware systems known as accelerators to meet the growing need for optimized performance. Optimizing these computations, however, presents significant challenges due to their sensitivity to variations in input sparsity patterns and code optimizations. While ML-based cost models and search techniques have shown promise in optimizing sparse matrix computations in general-purpose hardware like CPUs, these cost models require large datasets for effective training. Collecting such extensive datasets is particularly impractical for emerging hardware platforms that only have access to expensive simulators in the early design stages. To overcome this, we propose DASH, which trains learned cost models using low-cost data samples from widely accessible general-purpose hardware (such as CPUs), followed by few-shot fine-tuning to efficiently adapt to emerging hardware platforms. DASH introduces a novel approach that leverages the homogeneity of input features across different hardware platforms while effectively mitigating heterogeneity. This enables DASH to achieve comparable accuracy using only 5% of the data samples required by a cost model trained exclusively using data samples from an accelerator. We evaluate DASH on two critical sparse operations\u2014SpMM and SDDMM\u2014on an emerging sparse accelerator using 715 distinct sparsity patterns. Our experimental results show that DASH outperforms existing techniques that use transfer learning by 28.44%, achieving average speedups of 1.47x (up to 5.46x) for SpMM and 1.39x (up to 4.22x) for SDDMM.", "keywords": [ "learned cost models", "sparse accelerators", "transfer learning", "ml for systems" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "YeOxaKHE9b", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Don Yasiru Lakshan Ranasinghe", "gender": "Male", "institution": "Whiting School of Engineering", "country": "US", "position": "PhD student" }, { "name": "Vishal M Patel", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 28, "n_ref": 43, "n_ref_all": 57, "n_fig": 10, "n_tab": 8, "L_tab": 2475, "n_element_tab": 236, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2147, "n_element_tab_1": 182, "formula_len_all": 913, "formula_len_all_1": 669, "len_all": 146324, "len_all_1": 66064, "len_abs": 1416, "len_title": 90, "len_sents": 46135, "len_sents_1": 27849, "n_sents": 367, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1424, "title": "Retrieval-based Zero-shot Crowd Counting", "abs": "Existing crowd-counting methods rely on the manual localization of each person in the image. While recent efforts have attempted to circumvent the annotation burden through vision-language models or crowd image generation, these approaches rely on pseudo-labels to perform crowd-counting. Simulated datasets provide an alternative to the annotation cost associated with real datasets. However, the use of large-scale simulated data often results in a distribution gap between real and simulated domains. To address the latter, we introduce knowledge retrieval inspired by knowledge-enhanced models in natural language processing. With knowledge retrieval, we extract simulated crowd images and their text descriptions to augment the image embeddings of real crowd images to improve generalized crowd-counting. Knowledge retrieval allows one to use a vast amount of non-parameterized knowledge during testing, enhancing a model's inference capability. Our work is the first to actively incorporate text information to regress the crowd count in any supervised manner. Moreover, to address the domain gap, we propose a pre-training and retrieval mechanism that uses unlabeled real crowd images along with simulated data. We report state-of-the-art results for zero-shot counting on five public datasets, surpassing existing multi-model crowd-counting methods. The code will be made publicly available after the review process.", "keywords": [ "Crowd-counting", "Annotator free", "Zero-shot", "Vision-language models" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "YeErX16hMC", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Han Shen", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "PhD student" }, { "name": "Heshan Devaka Fernando", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "PhD student" }, { "name": "Horst Samulowitz", "gender": "unknown", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Nathalie Baracaldo", "gender": "unknown", "institution": "University of Pittsburgh", "country": "", "position": "PhD student" }, { "name": "Parikshit Ram", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Tianyi Chen", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "Assistant Professor" }, { "name": "Yi Zhou", "gender": "Female", "institution": "International Business Machines", "country": "US", "position": "Researcher" } ], "n_formula": 67, "n_formula_1": 29, "n_ref_uni": 37, "n_ref": 68, "n_ref_all": 113, "n_fig": 7, "n_tab": 6, "L_tab": 1525, "n_element_tab": 82, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 938, "n_element_tab_1": 33, "formula_len_all": 6582, "formula_len_all_1": 2282, "len_all": 198230, "len_all_1": 68830, "len_abs": 785, "len_title": 123, "len_sents": 52367, "len_sents_1": 28594, "n_sents": 461, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 790, "title": "Mitigating Forgetting in LLM Supervised Fine-Tuning and Preference Learning", "abs": "Post-training of pre-trained LLMs, which typically consists of the supervised fine-tuning (SFT) stage and the preference learning (RLHF or DPO) stage, is crucial to effective and safe LLM applications. The widely adopted approach in post-training popular open-source LLMs is to sequentially perform SFT and RLHF/DPO. However, sequential training is sub-optimal in terms of SFT and RLHF/DPO trade-off: the LLM gradually forgets about the first stage's training when undergoing the second stage's training. We theoretically prove the sub-optimality of sequential post-training. Furthermore, we propose a practical joint post-training framework that has theoretical convergence guarantees and empirically outperforms sequential post-training framework, while having similar computational cost.", "keywords": [ "Optimization Trade-off", "Large Language Models (LLMs)", "Supervised Fine-tuning (SFT)", "Reinforcement Learning from Human Feedback (RLHF)" ], "rating_list": [ 3, 8, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Yd5MHVIKLk", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Cho-Jui Hsieh", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Minhao Cheng", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Ruochen Wang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Sen Li", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Tianyi Zhou", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 3, "n_ref_uni": 21, "n_ref": 62, "n_ref_all": 91, "n_fig": 14, "n_tab": 7, "L_tab": 1928, "n_element_tab": 134, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 921, "n_element_tab_1": 60, "formula_len_all": 607, "formula_len_all_1": 232, "len_all": 133166, "len_all_1": 65730, "len_abs": 1601, "len_title": 131, "len_sents": 47152, "len_sents_1": 32100, "n_sents": 361, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1608, "title": "MuLan: Multimodal-LLM Agent for Progressive and Interactive Multi-Object Diffusion", "abs": "Existing text-to-image models still struggle to generate images of multiple objects, especially in handling their spatial positions, relative sizes, overlapping, and attribute bindings. To efficiently address these challenges, we develop a training-free Multimodal-LLM agent (MuLan), as a human painter, that can progressively generate multi-object with intricate planning and feedback control.\nMuLan harnesses a large language model (LLM) to decompose a prompt to a sequence of sub-tasks, each generating only one object by stable diffusion, conditioned on previously generated objects. Unlike existing LLM-grounded methods, MuLan only produces a high-level plan at the beginning while the exact size and location of each object are determined upon each sub-task by an LLM and attention guidance. Moreover, MuLan adopts a vision-language model (VLM) to provide feedback to the image generated in each sub-task and control the diffusion model to re-generate the image if it violates the original prompt. Hence, each model in every step of MuLan only needs to address an easy sub-task it is specialized for. The multi-step process also allows human users to monitor the generation process and make preferred changes at any intermediate step via text prompts, thereby improving the human-AI collaboration experience. We collect 200 prompts containing multi-objects with spatial relationships and attribute bindings from different benchmarks to evaluate MuLan. The results demonstrate the superiority of MuLan in generating multiple objects over baselines and its creativity when collaborating with human users.", "keywords": [ "Diffusion models", "Controllable generation", "multi-modal agent" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "Yd2GeHRSlJ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Shravan S Chaudhari", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "PhD student" }, { "name": "Suchi Saria", "gender": "Female", "institution": "Johns Hopkins University", "country": "", "position": "Assistant Professor" }, { "name": "Yoav Itzhak Wald", "gender": "Male", "institution": "New York University", "country": "", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 9, "n_ref_uni": 49, "n_ref": 92, "n_ref_all": 116, "n_fig": 3, "n_tab": 12, "L_tab": 5783, "n_element_tab": 674, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1915, "n_element_tab_1": 26, "formula_len_all": 1551, "formula_len_all_1": 710, "len_all": 187951, "len_all_1": 67909, "len_abs": 1762, "len_title": 157, "len_sents": 53380, "len_sents_1": 31902, "n_sents": 367, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1773, "title": "Open-Set Domain Adaptation Under Background Distribution Shift: Challenges and A Provably Efficient Solution", "abs": "In Open-Set Domain Adaptation (OSDA) we wish to perform classification in a target domain which contains a novel class along with $k$ non-novel classes. This work formally studies OSDA under the assumption that classes are separable, and the supports of source and target domains coincide, while other aspects of the distribution may change. We term such a distribution shift as background shift. \nWe develop a simple and scalable OSDA method that attains robustness to background shift and is guaranteed to solve the problem, while showing that it cannot be solved under weaker conditions for OSDA studied in the past, particularly in the presence of covariate shift. We formally define the realistic assumptions of background shift within the scope of OSDA problem that the previous literature has either overlooked or not explicitly addressed. In a thorough empirical evaluation on both image and text data, we observe that existing OSDA methods are not robust to the distribution shifts we consider.\nOur proposed solution jointly learns representations via concurrently learning to classify known categories and detect novel ones using methods with formal guarantees. The results demonstrate that optimizing these two objectives in unison leads to mutual performance improvements contrary to what might be expected when objectives are considered independently. Our rigorous empirical study also examines how OSDA performance under distribution shift is affected by parameters of the problem such as the novel class size. \nTaken together, our observations emphasize the importance of formalizing assumptions under which OSDA methods operate and to develop appropriate methodology that is capable of scaling with large datasets and models for different scenarios of OSDA.", "keywords": [ "Distribution Shift", "Open Set Domain Adaptation", "Novel Category Detection", "Anomaly Detection", "Out of Distribution" ], "rating_list": [ 5, 5, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 1, 3, 2 ], "confidence_list": [ 2, 3, 3, 3, 3 ] }, { "paper_id": "YcbE2K3i2E", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Arshveer Kaur", "gender": "Female", "institution": "BITS Pilani, Birla Institute of Technology and Science", "country": "IN", "position": "PhD student" }, { "name": "C Niranjan", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Intern" }, { "name": "Navneet Goyal", "gender": "Male", "institution": "BITS Pilani, Birla Institute of Technology and Science", "country": "IN", "position": "Full Professor" }, { "name": "Poonam Goyal", "gender": "Female", "institution": "BITS Pilani, Birla Institute of Technology and Science", "country": "", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 6, "n_ref": 11, "n_ref_all": 30, "n_fig": 2, "n_tab": 7, "L_tab": 6676, "n_element_tab": 369, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 683, "n_element_tab_1": 7, "formula_len_all": 62, "formula_len_all_1": 0, "len_all": 126372, "len_all_1": 58776, "len_abs": 1776, "len_title": 178, "len_sents": 43858, "len_sents_1": 30183, "n_sents": 331, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 127, "L_abs": 1784, "title": "SaTran: An efficient Transformer exploiting Spatiotemporal Redundancies for Satellite Image Time Series Representation Learning", "abs": "Earth observation applications like crop yield prediction, solar energy prediction, land cover classification, etc., need large size Satellite Image Time Series (SITS) leading to huge computational requirements. A couple of BERT-based models exist which work at pixel level unable to exploit spatial correlation among pixels and also require ground truth at pixel granularity during fine-tuning, rendering them infeasible for prediction tasks. The models based on Vision Transformer factorize spatial and time dimensions and first process images and then time series of image embeddings. However, in many cases, SITS require simultaneous analysis of both dimensions. We present a transformer, SaTran, which focuses on non-redundant patch tubes to overcome the limitations listed above. Transformers developed for RGB videos are found lacking when applied to SITS data characterized by the presence of patches with spatiotemporal redundancy persisting throughout the time series. SITS data also has patches where temporal redundancy lasts only for a few timestamps. The salient features of SaTran include: 1) an automatic patch tube selection mechanism which ignores spatiotemporally redundant patches; 2) exploitation of spatial correlation between pixels by the processing of patch tubes and handling of their temporal redundancy using tube masking; 3) two-fold handling of redundancy and distributed application of VideoMAE enables space and time efficient processing of large size SITS; and 4) learning end task agnostic representation of entire time series. Extensive experimentation shows that SaTran outperforms competing models and exhibit state-of-the-art performance for various earth observation applications. The code is available on (.. will be given after acceptance..).", "keywords": [ "Satellite image time series analytics", "Transformer", "Earth observation applications", "Spatiotemporal redundancy", "Representation learning" ], "rating_list": [ 5, 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 5, 4, 3, 3, 3 ] }, { "paper_id": "YcaFqY8LWD", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Chen Hu", "gender": "Male", "institution": "Jiangnan University", "country": "CN", "position": "MS student" }, { "name": "Nicu Sebe", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Full Professor" }, { "name": "Rui Wang", "gender": "Male", "institution": "Jiangnan University", "country": "CN", "position": "Lecturer" }, { "name": "Xiaojun Wu", "gender": "Male", "institution": "Jiangnan University", "country": "CN", "position": "Full Professor" }, { "name": "Ziheng Chen", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "PhD student" }, { "name": "Ning Xiao Song", "gender": "Male", "institution": "Jiangnan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 34, "n_formula_1": 24, "n_ref_uni": 35, "n_ref": 70, "n_ref_all": 107, "n_fig": 1, "n_tab": 11, "L_tab": 5702, "n_element_tab": 516, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1089, "n_element_tab_1": 68, "formula_len_all": 3919, "formula_len_all_1": 2583, "len_all": 198592, "len_all_1": 70153, "len_abs": 1268, "len_title": 104, "len_sents": 48226, "len_sents_1": 23407, "n_sents": 532, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1115, "title": "GyroAtt: A Gyro Attention Framework for Matrix Manifolds", "abs": "Deep neural networks operating on non-Euclidean geometries, such as Riemannian manifolds, have recently demonstrated impressive performance across various machine-learning applications. Motivated by the success of the attention mechanism, several works have extended it to different geometries. However, existing Riemannian attention methods are mostly designed in an \\textit{ad hoc} manner, \\textit{i.e.}, tailored to a selected few geometries. Recent studies, on the other hand, show that several matrix manifolds, such as Symmetric Positive Definite (SPD), Symmetric Positive Semi-Definite (SPSD), and Grassmannian manifolds, admit gyro structures, offering a principled way to build Riemannian networks. Inspired by this, we propose a Gyro Attention (GyroAtt) framework over general gyro spaces, applicable to various matrix manifolds. Empirically, we manifest our framework on three gyro structures in the SPD manifold, three in the SPSD manifold, and one in the Grassmannian manifold. Extensive experiments on four electroencephalography (EEG) datasets demonstrate the effectiveness of the proposed framework.", "keywords": [ "Manifold Learning", "Representation Learning", "Gyrovector Spaces", "Riemannian Manifolds", "Riemannian Self Attention" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 2, 4 ] }, { "paper_id": "YcUtOIzIXK", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Jack Hong", "gender": "unknown", "institution": "Xiaohongshu", "country": "CN", "position": "Intern" }, { "name": "Jiayin Cai", "gender": "Female", "institution": "Kuaishou", "country": "CN", "position": "Researcher" }, { "name": "Shilin Yan", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Xiaolong Jiang", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Enigneer" }, { "name": "Yao Hu", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Principal Researcher" }, { "name": "Zehao Xiao", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Henghui Ding", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 66, "n_ref": 142, "n_ref_all": 159, "n_fig": 2, "n_tab": 9, "L_tab": 6279, "n_element_tab": 766, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2361, "n_element_tab_1": 486, "formula_len_all": 312, "formula_len_all_1": 312, "len_all": 193944, "len_all_1": 67634, "len_abs": 1592, "len_title": 97, "len_sents": 38718, "len_sents_1": 29376, "n_sents": 267, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1601, "title": "Closed-loop Scaling Up for Visual Object Tracking", "abs": "Thanks to the principles of the scaling law, current neural networks have experienced remarkable performance improvements. While much of the existing research has concentrated on upstream pretraining, the application of the scaling law to downstream vision tasks remains underexplored. Understanding the scaling law in downstream tasks can aid in the design of more effective models and training strategies. Thus, in this work, we aim to investigate the application of the scaling law to downstream vision tasks. Firstly, we explore the impact of three key factors of scaling law: training data volume, model size, and input resolution. We empirically verify that increasing each of these factors can lead to performance enhancements. Secondly, to address naive training's optimization challenges and lack of iterative refinement, we introduce DT-Training which leverages small teacher transfer and dual-branch alignment to further exploit model potential. Thirdly, building on DT-Training, we propose a closed-loop scaling strategy to incrementally scale the model step-by-step. Finally, our scaled model exhibits strong ability and outperforms existing counterparts across diverse test benchmarks. Extensive experiments also reveal the robust transfer ability of our model. Moreover, we validate the generalizability of the scaling law and our proposed DT-Training on other downstream vision tasks, reinforcing the broader applicability of our approach. We hope that our findings can deepen the understanding of the scaling law in downstream tasks and foster future developments on downstream tasks.", "keywords": [ "Scaling law", "Downstream vision tasks", "Visual object tracking" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "YcUV5apdlq", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Ankur Singha", "gender": "Male", "institution": "Technische Universit\u00e4t Berlin", "country": "DE", "position": "Postdoc" }, { "name": "Elia Cellini", "gender": "Male", "institution": "University of Turin", "country": "IT", "position": "PhD student" }, { "name": "Karl Jansen", "gender": "Male", "institution": "DESY", "country": "", "position": "Principal Researcher" }, { "name": "Kim Andrea Nicoli", "gender": "Male", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "Postdoc" }, { "name": "Shinichi Nakajima", "gender": "Male", "institution": "BIFOLD, TU Berlin", "country": "DE", "position": "Postdoc" }, { "name": "Stefan K\u00fchn", "gender": "unknown", "institution": "Deutsches Elektronen Synchrotron", "country": "", "position": "Researcher" } ], "n_formula": 49, "n_formula_1": 17, "n_ref_uni": 97, "n_ref": 204, "n_ref_all": 250, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3117, "formula_len_all_1": 952, "len_all": 292539, "len_all_1": 76003, "len_abs": 1950, "len_title": 68, "len_sents": 75869, "len_sents_1": 33466, "n_sents": 510, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1968, "title": "Multilevel Generative Samplers for Investigating Critical Phenomena", "abs": "Investigating critical phenomena, i.e., phase transitions, is of high interest in physics and chemistry. However, Monte Carlo (MC) simulations, a crucial tool for numerically analyzing macroscopic properties of given systems, are often hindered by the emerging scale invariance at criticality (SIC)---a divergence of the correlation length, which causes the system to behave the same at any length scale, as can be shown with renormalisation group techniques. Many existing sampling methods suffer from SIC: long-range correlations cause critical slowing down in Markov chain Monte Carlo (MCMC), and require intractably large receptive fields for generative samplers. In this paper, we propose a Renormalization-informed Generative Critical Sampler (RiGCS)---a novel sampler specialized for near-critical systems, where SIC is leveraged as an advantage rather than a nuisance. Specifically, RiGCS builds on MultiLevel Monte Carlo (MLMC) with Heat Bath (HB) algorithms, which perform ancestral sampling from low-resolution to high-resolution lattice configurations with site-wise-independent conditional HB sampling. Although MLMC-HB is highly efficient under exact SIC, it suffers from a low acceptance rate under slight SIC violation---SIC violation always occurs in finite systems, and may induce long-range and higher-order interactions in the renormalized distributions, which are not considered by independent HB samplers. RiGCS enhances MLMC-HB by replacing a part of the conditional HB samplers with generative models that capture those residual interactions and improve the sampling efficiency---our experiments show that the effective sample size of RiGCS is a few orders of magnitude higher than state-of-the-art generative model baselines in sampling configurations for $128 \\times 128$ two-dimensional Ising systems. SIC also allows us to adopt a specialized sequential training protocol with model transfer, which significantly accelerates training.", "keywords": [ "Generative models", "Multilevel sampling", "Criticality", "Renormalization group" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 2, 3, 4, 2 ] }, { "paper_id": "YcML3rJl0N", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Bernd Bischl", "gender": "Male", "institution": "LMU", "country": "DE", "position": "Full Professor" }, { "name": "David R\u00fcgamer", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Associate Professor" }, { "name": "Mina Rezaei", "gender": "Female", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Principal Researcher" }, { "name": "Yawei Li", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 23, "n_ref_uni": 54, "n_ref": 101, "n_ref_all": 145, "n_fig": 1, "n_tab": 14, "L_tab": 15176, "n_element_tab": 1863, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 5719, "n_element_tab_1": 678, "formula_len_all": 3819, "formula_len_all_1": 1605, "len_all": 247415, "len_all_1": 74961, "len_abs": 1099, "len_title": 116, "len_sents": 41007, "len_sents_1": 25821, "n_sents": 351, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1106, "title": "Calibrating LLMs with Information-Theoretic Evidential Deep Learning", "abs": "Fine-tuned large language models (LLMs) often exhibit overconfidence, particularly when trained on small datasets, resulting in poor calibration and inaccurate uncertainty estimates. \nEvidential Deep Learning (EDL), an uncertainty-aware approach, enables uncertainty estimation in a single forward pass, making it a promising method for calibrating fine-tuned LLMs. However, despite its computational efficiency, EDL is prone to overfitting, as its training objective can result in overly concentrated probability distributions. \nTo mitigate this, we propose regularizing EDL by incorporating an information bottleneck (IB). Our approach IB-EDL suppresses spurious information in the evidence generated by the model and encourages truly predictive information to influence both the predictions and uncertainty estimates. Extensive experiments across various fine-tuned LLMs and tasks demonstrate that IB-EDL outperforms both existing EDL and non-EDL approaches. \nBy improving the trustworthiness of LLMs, IB-EDL facilitates their broader adoption in domains requiring high levels of confidence calibration.", "keywords": [ "evidential deep learning; information bottleneck; calibration; large language models" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 2, 3, 4 ] }, { "paper_id": "Yc4zTbR8no", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 10, "author_info_list": [ { "name": "Abdul Jawad", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Alberto Santamaria-Pang", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Principal Researcher" }, { "name": "Antika Roy", "gender": "Female", "institution": "University of Florida", "country": "US", "position": "PhD student" }, { "name": "Ho Hin Lee", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Applied Scientist" }, { "name": "Ivan Tarapov", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "MD Mahfuz Al Hasan", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "PhD student" }, { "name": "Mahdi Zaman", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Navid Asadi", "gender": "unknown", "institution": "University of Florida", "country": "", "position": "Associate Professor" }, { "name": "Reza Forghani", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "Full Professor" }, { "name": "Yaser Fallah", "gender": "Male", "institution": "University of Central Florida", "country": "", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 6, "n_ref_uni": 12, "n_ref": 20, "n_ref_all": 33, "n_fig": 4, "n_tab": 5, "L_tab": 1474, "n_element_tab": 178, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2473, "n_element_tab_1": 184, "formula_len_all": 1087, "formula_len_all_1": 447, "len_all": 114662, "len_all_1": 54650, "len_abs": 2255, "len_title": 153, "len_sents": 26523, "len_sents_1": 23891, "n_sents": 189, "n_sents_1": 154, "venue": "ICLR2025", "decision": "withdraw", "L_title": 109, "L_abs": 1835, "title": "WaveFormer: Leveraging Wavelet Transformation for Multi-Scale Token Interactions in Hierarchical Transformers", "abs": "Recent transformer models have achieved state-of-the-art performance for visual tasks involving high-dimensional data like 3D volumetric medical image segmentation. Hierarchical transformers (e.g., Swin Transformers) circumvent the computational challenge of the self-attention mechanism through a shifted window approach to learn token relations within progressively overlapping local regions, thus expanding the receptive field across layers while limiting token attention span in each layer within predefined windows. In this work, we introduce a novel learning paradigm that captures token relations through progressive summarization of features. We leverage the compaction capability of discrete wavelet transform (DWT) on high-dimensional features and learn token relation in multi-scale approximation coefficients obtained from DWT. This approach efficiently represents fine-grained local to coarse global contexts within each network layer. Furthermore, computing self-attention on the DWT-transformed features significantly reduces the computational complexity, effectively addressing the challenges posed by high-dimensional data in vision transformers. Our proposed network, termed WaveFormer, competes favorably with current SOTA transformers (e.g., SwinUNETR) using three challenging public datasets on volumetric medical imaging: (1) MICCAI Challenge 2021 FLARE, (2) MICCAI Challenge 2019 KiTS, and (3) MICCAI Challenge 2022 AMOS. WaveFormer consistently outperforms Swin-UNETR, improving from 0.929 to 0.938 Dice (FLARE2021) and 0.880 to 0.900 Dice (AMOS2022). In addition, we explore the WaveFormer\u2019s effectiveness in segmenting organs of varying sizes, demonstrating its robustness across different anatomical structures. The source code will be available with supplementary materials in the complete paper submission.", "keywords": [ "Transformer", "Attention Mechanism", "Receptive Field", "Discrete Wavelet Transformation", "Parseval's Theorem" ], "rating_list": [ 1, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 1, 1, 2, 1 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "YbusS3WNvb", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Huajie Shao", "gender": "Male", "institution": "College of William and Mary", "country": "US", "position": "Assistant Professor" }, { "name": "Jiacheng Shi", "gender": "Male", "institution": "College of William and Mary", "country": "US", "position": "PhD student" }, { "name": "Jiangtao Kong", "gender": "Male", "institution": "College of William and Mary", "country": "US", "position": "PhD student" }, { "name": "Shaohan Hu", "gender": "unknown", "institution": "J.P. Morgan Chase", "country": "US", "position": "Executive Director" }, { "name": "Tianyi Zhou", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Ye Gao", "gender": "Female", "institution": "College of William and Mary", "country": "US", "position": "Assistant Professor" } ], "n_formula": 29, "n_formula_1": 9, "n_ref_uni": 20, "n_ref": 32, "n_ref_all": 56, "n_fig": 9, "n_tab": 8, "L_tab": 2894, "n_element_tab": 258, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2138, "n_element_tab_1": 241, "formula_len_all": 1709, "formula_len_all_1": 435, "len_all": 146017, "len_all_1": 62968, "len_abs": 1637, "len_title": 133, "len_sents": 45828, "len_sents_1": 27684, "n_sents": 346, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1647, "title": "Hybrid Memory Replay: Blending Real and Distilled Data for Class Incremental Learning", "abs": "Incremental learning (IL) aims to acquire new knowledge from current tasks while retaining knowledge learned from previous tasks. Replay-based IL methods store a set of exemplars from previous tasks in a buffer and replay them when learning new tasks. However, there is usually a size-limited buffer that cannot store adequate real exemplars to retain the knowledge of previous tasks. In contrast, data distillation (DD) can reduce the exemplar buffer's size, by condensing a large real dataset into a much smaller set of more information-compact synthetic exemplars. Nevertheless, DD's performance gain on IL quickly vanishes as the number of synthetic exemplars grows. To overcome the weaknesses of real-data and synthetic-data buffers, we instead optimize a hybrid memory including both types of data. Specifically, we propose an innovative modification to DD that distills synthetic data from a sliding window of checkpoints in history (rather than checkpoints on multiple training trajectories). Conditioned on the synthetic data, we then optimize the selection of real exemplars to provide complementary improvement to the DD objective. The optimized hybrid memory combines the strengths of synthetic and real exemplars, effectively mitigating catastrophic forgetting in Class IL (CIL) when the buffer size for exemplars is limited. Notably, our method can be seamlessly integrated into most existing replay-based CIL models. Extensive experiments across multiple benchmarks demonstrate that our method significantly outperforms existing replay-based baselines. Our source code is available at https://anonymous.4open.science/r/DD4CIL-510C/.", "keywords": [ "Hybrid Memory", "Class Incremental Learning", "Data Distillation" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "YbURbViE7l", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Cuneyt Gurcan Akcora", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Associate Professor" }, { "name": "Murat Kantarcioglu", "gender": "unknown", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "Full Professor" }, { "name": "Zulfikar Alom", "gender": "Male", "institution": "University of Toledo", "country": "US", "position": "Researcher" }, { "name": "Bao Tran Gia Ngo", "gender": "Male", "institution": "University of Manitoba", "country": "CA", "position": "Undergrad student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 21, "n_ref": 32, "n_ref_all": 63, "n_fig": 8, "n_tab": 7, "L_tab": 4518, "n_element_tab": 300, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 2643, "n_element_tab_1": 172, "formula_len_all": 440, "formula_len_all_1": 523, "len_all": 133825, "len_all_1": 68759, "len_abs": 1225, "len_title": 137, "len_sents": 33988, "len_sents_1": 29995, "n_sents": 277, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1232, "title": "GOttack: Universal Adversarial Attacks on Graph Neural Networks via Graph Orbits Learning", "abs": "Graph Neural Networks (GNNs) have demonstrated superior performance in node classification tasks across diverse applications. However, their vulnerability to adversarial attacks, where minor perturbations can mislead model predictions, poses significant challenges. This study introduces GOttack, a novel adversarial attack framework that exploits the topological structure of graphs to undermine the integrity of GNN predictions systematically. \n\nBy defining a topology-aware method to manipulate graph orbits, our approach can generate adversarial modifications that are both subtle and effective, posing a severe test to the robustness of GNNs. We evaluate the efficacy of GOttack across multiple prominent GNN architectures using standard benchmark datasets. Our results show that GOttack outperforms existing state-of-the-art adversarial techniques and completes training in approximately 55% of the time required by the fastest competing model, achieving the highest average misclassification rate in 155 tasks. \nThis work not only sheds light on the susceptibility of GNNs to structured adversarial attacks but also shows that certain topological patterns may play a significant role in the underlying robustness of the GNNs.", "keywords": [ "graphlet", "orbit", "adversarial machine learning", "graph mining", "graph convolutional networks", "semi-supervised learning" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "YauQYh2k1g", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Aditi Raghunathan", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Daniel Fried", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Researcher" }, { "name": "Rishi Rajesh Shah", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Ruslan Salakhutdinov", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Chen Wu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Jing Yu Koh", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 60, "n_ref": 96, "n_ref_all": 137, "n_fig": 13, "n_tab": 7, "L_tab": 1449, "n_element_tab": 150, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 430, "n_element_tab_1": 48, "formula_len_all": 194, "formula_len_all_1": 131, "len_all": 178923, "len_all_1": 72828, "len_abs": 1589, "len_title": 105, "len_sents": 49909, "len_sents_1": 34448, "n_sents": 393, "n_sents_1": 273, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1655, "title": "Dissecting Adversarial Robustness of Multimodal LM Agents", "abs": "As language models (LMs) are used to build autonomous agents in real environments, ensuring their adversarial robustness becomes a critical challenge. Unlike chatbots, agents are compound systems with multiple components, which existing LM safety evaluations do not adequately address. To bridge this gap, we manually create 200 targeted adversarial tasks and evaluation functions in a realistic threat model on top of VisualWebArena, a real environment for web-based agents. In order to systematically examine the robustness of various multimodal we agents, we propose the Agent Robustness Evaluation (ARE) framework. ARE views the agent as a graph showing the flow of intermediate outputs between components and decomposes robustness as the flow of adversarial information on the graph. First, we find that we can successfully break a range of the latest agents that use black-box frontier LLMs, including those that perform reflection and tree-search. With imperceptible perturbations to a single product image (less than 5% of total web page pixels), an attacker can hijack these agents to execute targeted adversarial goals with success rates up to 67%. We also use ARE to rigorously evaluate how the robustness changes as new components are added. We find that *new components that typically improve benign performance can open up new vulnerabilities and harm robustness*. An attacker can compromise the evaluator used by the reflexion agent and the value function of the tree search agent, which increases the attack success relatively by 15% and 20%. Our data and code for attacks, defenses, and evaluation are available at url_removed_for_review.", "keywords": [ "LM agents", "multimodal agents", "safety", "adversarial robustness" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "Yan3Ll5oCp", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Fangyikang Wang", "gender": "Male", "institution": "Wechat, Tencent Inc", "country": "CN", "position": "Intern" }, { "name": "Huminhao Zhu", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Intern" }, { "name": "Qing Qu", "gender": "Male", "institution": "University of Michigan", "country": "US", "position": "Assistant Professor" }, { "name": "Zhihui Zhu", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Assistant Professor" }, { "name": "Tianyu Ding", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" } ], "n_formula": 56, "n_formula_1": 27, "n_ref_uni": 46, "n_ref": 112, "n_ref_all": 127, "n_fig": 8, "n_tab": 7, "L_tab": 632, "n_element_tab": 65, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 687, "n_element_tab_1": 44, "formula_len_all": 3323, "formula_len_all_1": 1434, "len_all": 164200, "len_all_1": 57828, "len_abs": 1102, "len_title": 111, "len_sents": 48942, "len_sents_1": 25360, "n_sents": 414, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1112, "title": "Model Collapse Analysis and Improvement for Rectified Flow Models", "abs": "Generative models aim to produce synthetic data indistinguishable from real distributions, but iterative training on self-generated data can lead to \\emph{model collapse (MC)}, where performance degrades over time. In this work, we provide the first theoretical analysis of MC in Rectified Flow by framing it within the context of Denoising Autoencoders (DAEs). We show that when DAE models are trained on recursively generated synthetic data with small noise variance, they suffer from MC with progressive diminishing generation quality. To address this MC issue, we propose methods that strategically incorporate real data into the training process, even when direct noise-image pairs are unavailable. Our proposed techniques, including Reverse Collapse-Avoiding (RCA) Reflow and Online Collapse-Avoiding Reflow (OCAR), effectively prevent MC while maintaining the efficiency benefits of Rectified Flow. Extensive experiments on standard image datasets demonstrate that our methods not only mitigate MC but also improve sampling efficiency, leading to higher-quality image generation with fewer sampling steps.", "keywords": [ "Model Collapse", "Generative Models", "Iterative Training", "Diffusion", "Rectified Flow" ], "rating_list": [ 3, 6, 5 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 1, 1, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 5, 3, 3 ] }, { "paper_id": "YaeZwhXJ4k", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Haojie Hao", "gender": "unknown", "institution": "Beihang University", "country": "CN", "position": "MS student" }, { "name": "Haotong Qin", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Jiakai Wang", "gender": "Male", "institution": "Zhongguancun Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Jinyang Guo", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Assistant Professor" }, { "name": "Michele Magno", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "Principal Researcher" }, { "name": "Mingyuan Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xianglong Liu", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Xingyu Zheng", "gender": "unknown", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Xudong Ma", "gender": "Male", "institution": "Beihang University", "country": "", "position": "PhD student" }, { "name": "Zixiang Zhao", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" } ], "n_formula": 23, "n_formula_1": 17, "n_ref_uni": 46, "n_ref": 121, "n_ref_all": 138, "n_fig": 1, "n_tab": 15, "L_tab": 4942, "n_element_tab": 930, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 342, "n_element_tab_1": 67, "formula_len_all": 1087, "formula_len_all_1": 901, "len_all": 168916, "len_all_1": 52493, "len_abs": 1858, "len_title": 117, "len_sents": 46092, "len_sents_1": 24252, "n_sents": 309, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1819, "title": "BinaryDM: Accurate Weight Binarization for Efficient Diffusion Models", "abs": "With the advancement of diffusion models (DMs) and the substantially increased computational requirements, quantization emerges as a practical solution to obtain compact and efficient low-bit DMs. However, the highly discrete representation leads to severe accuracy degradation, hindering the quantization of diffusion models to ultra-low bit-widths. This paper proposes a novel weight binarization approach for DMs, namely BinaryDM, pushing binarized DMs to be accurate and efficient by improving the representation and optimization. From the representation perspective, we present an Evolvable-Basis Binarizer (EBB) to enable a smooth evolution of DMs from full-precision to accurately binarized. EBB enhances information representation in the initial stage through the flexible combination of multiple binary bases and applies regularization to evolve into efficient single-basis binarization. The evolution only occurs in the head and tail of the DM architecture to retain the stability of training. From the optimization perspective, a Low-rank Representation Mimicking (LRM) is applied to assist the optimization of binarized DMs. The LRM mimics the representations of full-precision DMs in low-rank space, alleviating the direction ambiguity of the optimization process caused by fine-grained alignment. Comprehensive experiments demonstrate that BinaryDM achieves significant accuracy and efficiency gains compared to SOTA quantization methods of DMs under ultra-low bit-widths. With 1-bit weight and 4-bit activation (W1A4), BinaryDM achieves as low as 7.74 FID and saves the performance from collapse (baseline FID 10.87). As the first binarization method for diffusion models, W1A4 BinaryDM achieves impressive 15.2x OPs and 29.2x model size savings, showcasing its substantial potential for edge deployment.", "keywords": [ "Model Quantization", "Model Compression", "Generative Model", "Diffusion Model" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "YaRzuMaubS", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Anca Dragan", "gender": "Female", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Aryansh Shrivastava", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Justin Svegliato", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Research Scientist" }, { "name": "Marwa Abdulhai", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Micah Carroll", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 57, "n_ref": 91, "n_ref_all": 107, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 676, "formula_len_all_1": 676, "len_all": 177087, "len_all_1": 84108, "len_abs": 980, "len_title": 85, "len_sents": 41310, "len_sents_1": 41312, "n_sents": 261, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 984, "title": "Defining Deception in Decision Making", "abs": "With the growing capabilities of machine learning systems, particularly those that interact with humans, there is an increased risk of systems that can easily deceive and manipulate people. Preventing unintended behaviors therefore represents an important challenge for creating aligned AI systems. To approach this challenge in a principled way, we first need to define deception formally. In this work, we present a concrete definition of deception under the formalism of rational decision making in partially observed Markov decision processes. Specifically, we propose a general regret theory of deception under which the degree of deception can be quantified in terms of the actor's beliefs, actions, and utility. To evaluate our definition, we study the degree to which our definition aligns with human judgments about deception. We hope that our work will constitute a step toward both systems that aim to avoid deception, and detection mechanisms to identify deceptive agents.", "keywords": [ "deception", "AI safety" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "YaBiGjuDiC", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Huazheng Wang", "gender": "unknown", "institution": "Oregon State University", "country": "US", "position": "Assistant Professor" }, { "name": "Hui Yuan", "gender": "Female", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Liu Leqi", "gender": "Female", "institution": "Princeton University", "country": "", "position": "Postdoc" }, { "name": "Mengdi Wang", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Associate Professor" }, { "name": "Yifan Zeng", "gender": "Male", "institution": "Oregon State University", "country": "US", "position": "PhD student" }, { "name": "Yue Wu", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Postdoc" } ], "n_formula": 60, "n_formula_1": 30, "n_ref_uni": 20, "n_ref": 48, "n_ref_all": 86, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6838, "formula_len_all_1": 2410, "len_all": 130011, "len_all_1": 68436, "len_abs": 1627, "len_title": 131, "len_sents": 42221, "len_sents_1": 30439, "n_sents": 319, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1505, "title": "Common Pitfalls of Margin-based Preference Optimization in Language Model Alignment", "abs": "Reinforcement Learning from Human Feedback (RLHF) has become the predominant approach for aligning language models (LMs) to be more helpful and less harmful. At its core, RLHF uses a margin-based loss for preference optimization, which specifies the ideal LM behavior only in terms of the difference between preferred and dispreferred responses. \nThis under-specification of ideal behavior for each response individually leads to two unintended consequences as the margin increases:\n(1) The probability of dispreferred (e.g., unsafe) responses may increase, resulting in potential safety alignment failures.\n(2) When the probability of dispreferred responses is reduced, this often coincides with a decrease in the probability of preferred responses, even when these responses are ideal.\nIn this paper, we identify the fundamental issue: margin-based preference optimization loss under-specifies ideal LM behaviors. \nWe derive key conditions under which the probabilities of both preferred and dispreferred responses increase or decrease together. \nThese conditions occur when the inner products between the gradients of the log-probabilities of preferred and dispreferred responses are large. \nWe theoretically analyze when such inner products are large and empirically validate our findings. \nOur framework also reveals important differences in the training dynamics of various preference optimization algorithms and suggests new directions for developing better algorithms for language model alignment.", "keywords": [ "Alignment", "Preference Optimization", "Large Language Model" ], "rating_list": [ 5, 5, 6, 8 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 4 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "YZdc7mTq7I", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Guorong Wu", "gender": "Male", "institution": "University of North Carolina, Chapel Hill", "country": "US", "position": "Associate Professor" }, { "name": "Jiaqi Ding", "gender": "Female", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "", "position": "PhD student" }, { "name": "Tingting Dan", "gender": "Female", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Postdoc" } ], "n_formula": 22, "n_formula_1": 13, "n_ref_uni": 80, "n_ref": 131, "n_ref_all": 154, "n_fig": 8, "n_tab": 6, "L_tab": 4753, "n_element_tab": 286, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1930, "n_element_tab_1": 109, "formula_len_all": 1542, "formula_len_all_1": 1001, "len_all": 223229, "len_all_1": 77695, "len_abs": 1943, "len_title": 160, "len_sents": 58870, "len_sents_1": 36458, "n_sents": 421, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 112, "L_abs": 1954, "title": "GeoMind: A Geometric Neural Network of State Space Model for Understanding Brain Dynamics on Riemannian Manifold", "abs": "State space model (SSM) is a powerful tool in neuroscience field to characterize the dynamic nature of brain functions by elucidating the mechanism of how brain system transits between brain states and how underlying states give rise to the observed neural activities. Although tremendous efforts have been made to lend the power of deep learning and mathematical insight of SSM in various functional neuroimaging studies, current state-of-the-art methods lack a holistic view of brain state evolution as a self-organized dynamical system where each part of the brain is functionally inter-connected. Since the topological co-activation of functional fluctuations exhibits an intrinsic geometric pattern (symmetric and positive definite, or SPD) on the Riemannian manifold, the call for understanding how a selective set of functional connectivities in the brain supports diverse behavior and cognition emerges a new machine learning scenario of manifold-based SSM for large-scale functional neuroimages. To that end, we propose a geometric neural networks, coined *GeoMind*, designed to uncover evolving brain states by tracking the trajectory of functional dynamics on a high-dimensional Riemannian manifold of SPD matrices. Our *GeoMind* demonstrates promising results in identifying specific brain states based on task-based functional Magnetic Resonance Imaging (fMRI) data, as well as in diseases early diagnosis for Alzheimer's disease, Parkinson's disease and Autism. These results highlight the applicability of the proposed *GeoMind* in neuroscience research. Furthermore, to assess the generalization capabilities of our model, we applied it to the domain of human action recognition (HAR), achieving promising performance on three benchmark datasets (UTKinect, Florence and HDM05). This demonstrates the scalability and robustness of the proposed geometry deep model of SSM in capturing complex spatio-temporal dynamics across diverse fields.", "keywords": [ "Geometric deep learning", "state space model", "brain dynamics", "Riemannian Manifold" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 2, 2 ] }, { "paper_id": "YZRgB5DnXw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Bingsheng He", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Zining Zhang", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 97, "n_ref_all": 117, "n_fig": 7, "n_tab": 7, "L_tab": 1121, "n_element_tab": 107, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 430, "n_element_tab_1": 62, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 97392, "len_all_1": 55608, "len_abs": 1326, "len_title": 53, "len_sents": 32639, "len_sents_1": 25986, "n_sents": 243, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1334, "title": "EOP: Unlocking Superior Problem Solving in Small LLMs", "abs": "Small language models, referred to as LLMs with fewer than 10 billion parameters in this work, face critical challenges in problem-solving\ntasks, often achieving less than 10\\% accuracy, highlighting the\nurgent need for effective solutions. While much of the existing research has focused on enhancing the performance of larger models like GPT, an important question remains: Can techniques developed for large models be adapted effectively for smaller ones? Moreover, is it possible to improve these smaller models to the point where they rival, or even outperform, larger models such as GPT-4 in problem-solving tasks?\n\nIn this paper, we introduce Evaluation-Oriented Problem-Solving (EOP), a novel framework aimed at enhancing the problem-solving capabilities of small LLMs. Our approach significantly boosts the performance of these models, achieving a 2\\% higher accuracy on Python Puzzles compared to standard GPT-4 and a 27\\% improvement over state-of-the-art prompting methods using GPT-4 in the Game of 24. Beyond these results, EOP also demonstrates notable accuracy improvements on other tasks. These findings suggest that, with the appropriate strategies, small LLMs can achieve substantial performance gains in problem-solving, challenging the prevailing notion that scaling model size is the primary path to improvement.", "keywords": [ "SLM", "LLM", "Problem-Solvinng", "Python Puzzles", "Game of 24" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 3, 3, 4, 2 ] }, { "paper_id": "YZEzVR5awV", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Baohao Liao", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Christof Monz", "gender": "Male", "institution": "University of Amsterdam, University of Amsterdam", "country": "NL", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 42, "n_ref": 119, "n_ref_all": 157, "n_fig": 6, "n_tab": 14, "L_tab": 11556, "n_element_tab": 1343, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1019, "n_element_tab_1": 108, "formula_len_all": 539, "formula_len_all_1": 366, "len_all": 230330, "len_all_1": 57565, "len_abs": 1417, "len_title": 123, "len_sents": 43853, "len_sents_1": 25733, "n_sents": 331, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1423, "title": "ClusComp: A Simple Paradigm for Model Compression and Efficient Finetuning", "abs": "As large language models (LLMs) continue to scale, model compression becomes increasingly important for enabling edge deployment and ensuring accessibility to users with limited resources. Weight-only quantization is a key technique for model compression, allowing for a substantial reduction in model size while preserving performance. However, as bit-width decreases, the performance of quantized LLMs tends to degrade significantly. Additionally, due to the non-differentiable operation in quantization, standard finetuning on quantized LLMs is unsupported, and alternative finetuning approaches often fail to match the effectiveness of full finetuning. In this paper, we introduce ClusComp, a novel and simple model compression paradigm. ClusComp first clusters the weight matrices to generate codebooks, and then tunes these codebooks block-by-block to reconstruct intermediate activations. Despite its simplicity, ClusComp (1) consistently achieves better performance in 2-4 bit precision; (2) pushes the compression limit to the 1-bit level, and outperforms existing ultra-low-bit methods with limited finetuning steps; (3) facilitates seamless and efficient finetuning, surpasses existing quantization-based or memory-efficient finetuning methods, and even rivals full finetuning of the FP16 model. Notably, these procedures can be executed on a single NVIDIA A6000-48GB GPU for LLMs with as many as 70B parameters.", "keywords": [ "model compression", "quantization", "efficient finetuning", "llm" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "YYMd6zsP2e", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Kotaro Sakamoto", "gender": "unknown", "institution": "The University of Tokyo", "country": "", "position": "Postdoc" }, { "name": "Manato Yaguchi", "gender": "Male", "institution": "The University of Tokyo", "country": "", "position": "MS student" }, { "name": "Masahiro Suzuki", "gender": "Male", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Masato Tanabe", "gender": "Male", "institution": "Hokkaido University", "country": "", "position": "PhD student" }, { "name": "Masatomo Akagawa", "gender": "unknown", "institution": "Hokkaido University", "country": "", "position": "MS student" }, { "name": "Ryosuke Sakamoto", "gender": "Male", "institution": "Hokkaido University", "country": "JP", "position": "PhD student" }, { "name": "Yusuke Hayashi", "gender": "Male", "institution": "AI Alignment Network", "country": "JP", "position": "Principal Researcher" }, { "name": "Yutaka Matsuo", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" } ], "n_formula": 122, "n_formula_1": 24, "n_ref_uni": 54, "n_ref": 122, "n_ref_all": 159, "n_fig": 19, "n_tab": 6, "L_tab": 1012, "n_element_tab": 80, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 598, "n_element_tab_1": 45, "formula_len_all": 11957, "formula_len_all_1": 2009, "len_all": 309541, "len_all_1": 59837, "len_abs": 1604, "len_title": 144, "len_sents": 71664, "len_sents_1": 23634, "n_sents": 695, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 674, "title": "The Geometry of Phase Transitions in Diffusion Models: Tubular Neighbourhoods and Singularities", "abs": "Diffusion models undergo phase transitions during the generative process where data features suddenly emerge in the final stages. The current study aims to elucidate this critical phenomenon from the geometrical perspective. We employ the concept of ``injectivity radius'', a quantity that characterises the structure of the data manifold. Through theoretical and empirical evidence, we demonstrate that phase transitions in the generative process of diffusion models are closely related to the injectivity radius. Our findings offer a novel perspective on phase transitions in diffusion models, with potential implications for improving performance and sampling efficiency.", "keywords": [ "diffusion models", "geometry", "tubular neighbourhoods" ], "rating_list": [ 8, 3, 3, 5 ], "soundness_list": [ 4, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "YXewbZ8FgU", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Ruixi Lin", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 26, "n_ref": 44, "n_ref_all": 61, "n_fig": 5, "n_tab": 5, "L_tab": 2806, "n_element_tab": 197, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 213, "n_element_tab_1": 9, "formula_len_all": 528, "formula_len_all_1": 583, "len_all": 110098, "len_all_1": 58830, "len_abs": 1864, "len_title": 129, "len_sents": 34659, "len_sents_1": 28120, "n_sents": 265, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1873, "title": "Let the Rule Speak: Enhancing In-context Learning Debiasing with Interpretability", "abs": "In-context learning, which allows large language models to perform diverse tasks with a few demonstrations, is found to have imbalanced per-class prediction accuracy on multi-class text classification. Although notable output correction methods have been developed to tackle the issue and simultaneously improve downstream prediction accuracy, they may fail to answer the core interpretability challenges: why and which certain classes need corrections, and more importantly, a tailored correction for per-sample, per-class\u2019s probability. To address such interpretability gaps, we first find that the imbalance arises from certain classes consistently receiving high ICL output probabilities, whereas others receiving lower or mixed ranges, so the former is more frequently chosen, resulting in higher accuracy; more crucially, we find that these ranges have significantly varying degrees of influence on the accuracy bias, highlighting the need for precise, interpretable probability corrections by range. Motivated by this, we propose FuRud, a Fuzzy Rule Optimization based Debiasing method, that (1) detects which classes need corrections, and (2) for each correction-needed class, detects its probability ranges and applies asymmetric amplifications or reductions to correct them interpretably. Notably, across seven benchmark datasets, FuRud reduces the pairwise class accuracy bias (COBias) by more than half (56\\%), while achieving a relative increase of 21\\% in accuracy, outperforming state-of-the-art debiasing methods. Moreover, FuRud can optimize a downstream task in a few-shot manner, with as few as 10 optimization examples. Furthermore, FuRud can work for prompt formats that lead to highly skewed predictions. For example, FuRud greatly improves ICL outputs which use letter options, with 44\\% relative accuracy increase and 54\\% relative COBias reduction.", "keywords": [ "In-context learning", "debiasing", "fuzzy rule", "interpretability", "multi-objective optimization" ], "rating_list": [ 3, 5, 8, 5 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 2, 1, 4, 3 ], "contribution_list": [ 2, 2, 4, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "YXRyYkb1im", "primary_area": "applications to robotics, autonomy, planning", "n_author": 10, "author_info_list": [ { "name": "Behzad Dariush", "gender": "Male", "institution": "Honda Research Institute USA", "country": "US", "position": "Principal Researcher" }, { "name": "Chuang Gan", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Hongxin Zhang", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" }, { "name": "Kwonjoon Lee", "gender": "Male", "institution": "Honda Research Institute USA", "country": "US", "position": "Sr Research Scientist" }, { "name": "Qiushi Lyu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Sunli Chen", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" }, { "name": "Tianmin Shu", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" }, { "name": "Yilun Du", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Zeyuan Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Zheyuan Zhang", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 12, "n_ref_uni": 47, "n_ref": 71, "n_ref_all": 104, "n_fig": 13, "n_tab": 6, "L_tab": 1093, "n_element_tab": 97, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 510, "n_element_tab_1": 39, "formula_len_all": 794, "formula_len_all_1": 713, "len_all": 182606, "len_all_1": 47335, "len_abs": 1507, "len_title": 118, "len_sents": 48080, "len_sents_1": 22538, "n_sents": 331, "n_sents_1": 125, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1518, "title": "COMBO: Compositional World Models for Embodied Multi-Agent Cooperation", "abs": "In this paper, we investigate the problem of embodied multi-agent cooperation, where decentralized agents must cooperate given only egocentric views of the world. To effectively plan in this setting, in contrast to learning world dynamics in a single-agent scenario, we must simulate world dynamics conditioned on an arbitrary number of agents' actions given only partial egocentric visual observations of the world. To address this issue of partial observability, we first train generative models to estimate the overall world state given partial egocentric observations. To enable accurate simulation of multiple sets of actions on this world state, we then propose to learn a compositional world model for multi-agent cooperation by factorizing the naturally composable joint actions of multiple agents and compositionally generating the video conditioned on the world state. By leveraging this compositional world model, in combination with Vision Language Models to infer the actions of other agents, we can use a tree search procedure to integrate these modules and facilitate online cooperative planning. We evaluate our methods on three challenging benchmarks with 2-4 agents. The results show our compositional world model is effective and the framework enables the embodied agents to cooperate efficiently with different agents across various tasks and an arbitrary number of agents, showing the promising future of our proposed methods. More videos can be found at \\url{https://combo-iclr.github.io/COMBO/}.", "keywords": [ "Embodied AI; Multi-agent Planning; Compositional World Model" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "YWtT5OGOAI", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 7, "author_info_list": [ { "name": "Anthos Makris", "gender": "unknown", "institution": "University of Edinburgh", "country": "GB", "position": "Undergrad student" }, { "name": "Chuanhao Sun", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Principal Researcher" }, { "name": "Kai Xu", "gender": "Male", "institution": "MIT-IBM Watson AI Lab", "country": "", "position": "Research scientist" }, { "name": "Luo Mai", "gender": "Male", "institution": "University of Edinburgh", "country": "", "position": "Assistant Professor" }, { "name": "Mahesh K. Marina", "gender": "Male", "institution": "Edinburgh University, University of Edinburgh", "country": "GB", "position": "Full Professor" }, { "name": "Maja Drma\u010d", "gender": "Female", "institution": "Edinburgh University, University of Edinburgh", "country": "GB", "position": "MS student" }, { "name": "Thanos Triantafyllou", "gender": "Male", "institution": "Edinburgh University, University of Edinburgh", "country": "GB", "position": "Researcher" } ], "n_formula": 37, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 69, "n_ref_all": 86, "n_fig": 8, "n_tab": 10, "L_tab": 1865, "n_element_tab": 130, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 903, "n_element_tab_1": 17, "formula_len_all": 2654, "formula_len_all_1": 669, "len_all": 144590, "len_all_1": 53167, "len_abs": 1224, "len_title": 114, "len_sents": 47234, "len_sents_1": 25437, "n_sents": 352, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1229, "title": "Practical Epistemic Uncertainty Quantification for View Synthesis", "abs": "View synthesis using Neural Radiance Fields (NeRF) and Gaussian Splatting (GS) has demonstrated impressive fidelity in rendering real-world scenarios. However, practical methods for accurate and efficient epistemic Uncertainty Quantification (UQ) in view synthesis are lacking. Existing approaches for NeRF either introduce significant computational overhead (e.g., \"10x increase in training time\" or \"10x repeated training\") or are limited to specific uncertainty conditions or models. Notably, GS models lack any systematic approach for comprehensive epistemic UQ. This capability is crucial for improving the robustness and scalability of neural view synthesis, enabling active model updates, error estimation, and scalable ensemble modeling based on uncertainty. In this paper, we revisit NeRF and GS-based methods from a function approximation perspective, identifying key differences and connections in 3D representation learning. Building on these insights, we introduce PH-Dropout, the first real-time and accurate method for epistemic uncertainty estimation that operates directly on pre-trained NeRF and GS models. Extensive evaluations validate our theoretical findings and demonstrate the effectiveness of PH-Dropout.", "keywords": [ "Deep Learning", "Epistemic Uncertainty Estimation", "View Synthesis" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 2, 4 ] }, { "paper_id": "YWaXJWd9nu", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Daniel Romero-Alvarado", "gender": "Male", "institution": "Universidad Polit\u00e9cnica de Valencia", "country": "ES", "position": "PhD student" }, { "name": "Fernando Mart\u00ednez-Plumed", "gender": "Male", "institution": "Universidad Polit\u00e9cnica de Valencia", "country": "ES", "position": "Associate Professor" }, { "name": "Jose Hernandez-Orallo", "gender": "unknown", "institution": "Universitat Politecnica de Valencia", "country": "ES", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 39, "n_ref": 54, "n_ref_all": 72, "n_fig": 8, "n_tab": 1, "L_tab": 566, "n_element_tab": 40, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 471, "n_element_tab_1": 40, "formula_len_all": 250, "formula_len_all_1": 250, "len_all": 116047, "len_all_1": 60352, "len_abs": 902, "len_title": 88, "len_sents": 31286, "len_sents_1": 28507, "n_sents": 206, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 905, "title": "What should an AI assessor optimise for?", "abs": "An AI assessor is an external, ideally independent system that predicts an indicator, e.g., a loss value, of another AI system. Assessors can leverage information from the test results of many other AI systems and have the flexibility of being trained on any loss function: from squared error to toxicity metrics. Here we address the question: is it always optimal to train the assessor for the target loss? Or could it be better to train for a different loss and then map predictions back to the target loss? Using ten regression problems with tabular data, we experimentally explore this question for regression losses with monotonic and nonmonotonic mappings and find that, contrary to intuition, optimising for more informative losses is not generally better. Surprisingly though, some monotonic transformations, such as the logistic loss used to minimise the absolute or squared error, are promising.", "keywords": [ "Assessor Models", "Predictable AI", "Regression problems", "Error metrics" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "YWTpBisnwd", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Chaolong Ying", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Tianshu Yu", "gender": "Male", "institution": "Chinese University of Hong Kong (Shenzhen)", "country": "CN", "position": "Assistant Professor" }, { "name": "Wei Pang", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "MS student" }, { "name": "Xiangru Jian", "gender": "Male", "institution": "ServiceNow Research", "country": "CA", "position": "Researcher" }, { "name": "Xinjian Zhao", "gender": "Male", "institution": "Chinese University of Hong Kong (Shenzhen)", "country": "CN", "position": "PhD student" }, { "name": "Yaoyao Xu", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 9, "n_ref_uni": 47, "n_ref": 75, "n_ref_all": 96, "n_fig": 13, "n_tab": 4, "L_tab": 2625, "n_element_tab": 215, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2844, "n_element_tab_1": 221, "formula_len_all": 2624, "formula_len_all_1": 773, "len_all": 190369, "len_all_1": 64937, "len_abs": 1294, "len_title": 109, "len_sents": 58730, "len_sents_1": 28749, "n_sents": 447, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1241, "title": "Enhancing Graph Self-Supervised Learning with Graph Interplay", "abs": "Graph self-supervised learning (GSSL) has emerged as a compelling framework for extracting informative representations from graph-structured data without extensive reliance on labeled inputs. In this study, we introduce Graph Interplay (GIP), an innovative and versatile approach that significantly enhances the performance equipped with various existing GSSL methods. To this end, GIP advocates direct graph-level communications by introducing random inter-graph edges within standard batches. Against GIP's simplicity, we further theoretically show that GIP essentially performs a principled manifold separation via combining inter-graph message passing and GSSL, bringing about more structured embedding manifolds and thus benefits a series of downstream tasks. Our empirical study demonstrates that GIP surpasses the performance of prevailing GSSL methods across multiple benchmarks by significant margins, highlighting its potential as a breakthrough approach. Besides, GIP can be readily integrated into a series of GSSL methods and consistently offers additional performance gain. This advancement not only amplifies the capability of GSSL but also potentially sets the stage for a novel graph learning paradigm in a broader sense.", "keywords": [ "Graph Self-Supervised Learning", "Graph Neural Networks" ], "rating_list": [ 3, 5, 5, 5, 6 ], "soundness_list": [ 1, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 3, 4 ] }, { "paper_id": "YW79lAHBUF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Antoine Bosselut", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Assistant Professor" }, { "name": "Giovanni Monea", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Yoav Artzi", "gender": "Not Specified", "institution": "Cornell University", "country": "US", "position": "Associate Professor" }, { "name": "kiante brantley", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 45, "n_ref": 67, "n_ref_all": 88, "n_fig": 10, "n_tab": 3, "L_tab": 618, "n_element_tab": 47, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 447, "n_element_tab_1": 76, "formula_len_all": 383, "formula_len_all_1": 855, "len_all": 150130, "len_all_1": 59566, "len_abs": 844, "len_title": 90, "len_sents": 33059, "len_sents_1": 26704, "n_sents": 278, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 848, "title": "LLMs Are In-Context Reinforcement Learners", "abs": "Large Language Models (LLMs) can learn new tasks through in-context supervised learning (i.e., ICL). This work studies if this ability extends to in-context reinforcement learning (ICRL), where models are not given gold labels in context, but only their past predictions and rewards. We show that a naive application of ICRL fails miserably, and identify the root cause as a fundamental deficiency at exploration, which leads to quick model degeneration. We propose an algorithm to address this deficiency by increasing test-time compute, as well as a compute-bound approximation. We use several challenging classification tasks to empirically show that our ICRL algorithms lead to effective learning from rewards alone, and analyze the characteristics of this ability and our methods. Overall, our results reveal remarkable ICRL abilities in LLMs.", "keywords": [ "in-context learning", "in-context reinforcement learning", "large language models" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "YVubckGzED", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Chichun Zhou", "gender": "Male", "institution": "Dali University", "country": "CN", "position": "Associate Professor" }, { "name": "Le Gao", "gender": "Female", "institution": "Tianjin Institute of Industrial Biotechnology, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Ling An", "gender": "unknown", "institution": "Dali University", "country": "CN", "position": "MS student" }, { "name": "Yansong Li", "gender": "Male", "institution": "University of Illinois Chicago", "country": "US", "position": "PhD student" }, { "name": "Yu Wu", "gender": "unknown", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Zeyu Dong", "gender": "Not Specified", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Zhangdi Liu", "gender": "unknown", "institution": "Dali University", "country": "CN", "position": "MS student" }, { "name": "Zhenyu Zhang", "gender": "Male", "institution": "School of Engineering", "country": "CN", "position": "Full Professor" }, { "name": "Zhuohang Yu", "gender": "Male", "institution": "Dali University", "country": "CN", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 51, "n_ref": 109, "n_ref_all": 141, "n_fig": 20, "n_tab": 11, "L_tab": 4918, "n_element_tab": 619, "n_fig_1": 12, "n_tab_1": 5, "L_tab_1": 1728, "n_element_tab_1": 96, "formula_len_all": 31, "formula_len_all_1": 31, "len_all": 151289, "len_all_1": 50252, "len_abs": 1502, "len_title": 148, "len_sents": 44789, "len_sents_1": 23636, "n_sents": 297, "n_sents_1": 147, "venue": "ICLR2025", "decision": "withdraw", "L_title": 99, "L_abs": 1523, "title": "EAPCR: A Universal Feature Extractor for Scientific Data Without Explicit Feature Relation Patterns", "abs": "Conventional methods, including Decision Tree (DT)-based methods, have been highly effective in scientific tasks, such as non-image medical diagnostics, system anomaly detection, and inorganic catalysis efficiency prediction. However, most deep-learning techniques have struggled to surpass or even match this level of success as traditional machine learning methods. The primary reason is that these applications involve multi-source, heterogeneous data, where features lack explicit relationships. This contrasts with image data, where pixels exhibit spatial relationships; textual data, where words have sequential dependencies; and graph data, where nodes are connected through established associations. The absence of explicit **F**eature **R**elation **P**atterns (FRPs) presents a significant challenge for deep learning techniques in scientific applications that are not image, text, and graph-based. In this paper, we introduce *EAPCR*, a universal feature extractor designed for data without explicit FRPs. Tested across various scientific tasks, EAPCR consistently outperforms traditional methods and bridges the gap where deep learning models fall short. To further demonstrate its robustness, we synthesize a dataset without explicit FRPs. While Kolmogorov\u2013Arnold Network (KAN) and feature extractors like Convolutional Neural Networks (CNNs), Graph Convolutional Networks (GCNs), and Transformers struggle, EAPCR excels, demonstrating its robustness and superior performance in scientific tasks without FRPs.", "keywords": [ "AI for Science", "Representation Learning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "YVsiB41ifI", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Beomchan Park", "gender": "unknown", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Byung-Kwan Lee", "gender": "Male", "institution": "NVIDIA", "country": "TW", "position": "Intern" }, { "name": "Chae Won Kim", "gender": "Female", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Sangyun Chung", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "MS student" }, { "name": "Yong Man Ro", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 54, "n_ref": 175, "n_ref_all": 192, "n_fig": 4, "n_tab": 5, "L_tab": 5541, "n_element_tab": 660, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2549, "n_element_tab_1": 222, "formula_len_all": 902, "formula_len_all_1": 783, "len_all": 268482, "len_all_1": 71730, "len_abs": 1437, "len_title": 102, "len_sents": 54761, "len_sents_1": 26315, "n_sents": 382, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1444, "title": "Phantom of Latent for Large Language and Vision Models", "abs": "The success of visual instruction tuning has accelerated the development of large language and vision models (LLVMs). Following the scaling laws of instruction-tuned large language models (LLMs), LLVMs have also further increased in size, with examples including 26B, 34B, and even 80B parameters. While this increase in model size has yielded significant performance gains, it demands substantially more hardware resources for both training and inference. Consequently, there naturally exists a strong need for efficient LLVMs that achieve the performance of larger models while being smaller in size. To achieve this need, we present a new efficient LLVM family with model sizes of 0.5B, 1.8B, 3.8B, and 7B parameters, Phantom, which significantly enhances learning capabilities within limited structures. By temporarily increasing the latent hidden dimension during multi-head self-attention (MHSA), we make LLVMs understand much more vision-language knowledge on the latent, without substantially increasing physical model sizes. To maximize its advantage, we introduce Phantom Optimization (PO) using both autoregressive supervised fine-tuning (SFT) and direct preference optimization (DPO)-like concept, which effectively follows correct answers while eliminating incorrect and ambiguous ones. Phantom outperforms numerous larger open- and closed-source LLVMs, positioning itself as a leading solution in the landscape of efficient LLVMs.", "keywords": [ "Large Language and Vision Models" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 2, 2, 3, 5 ] }, { "paper_id": "YVcVbYYy2x", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Cong Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": " LLM Algorithm Research Scientist" }, { "name": "Dexun Li", "gender": "Male", "institution": "Singapore Management University", "country": "", "position": "PhD student" }, { "name": "Hao Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Kuicai Dong", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "SG", "position": "Researcher" }, { "name": "Ruiming Tang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Weiwen Liu", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Wenjun Li", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "PhD student" }, { "name": "Yasheng Wang", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "", "position": "Researcher" }, { "name": "Yong Liu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "SG", "position": "Principal Researcher" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 40, "n_ref": 62, "n_ref_all": 87, "n_fig": 8, "n_tab": 6, "L_tab": 2549, "n_element_tab": 471, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1516, "n_element_tab_1": 327, "formula_len_all": 252, "formula_len_all_1": 198, "len_all": 165180, "len_all_1": 57738, "len_abs": 2056, "len_title": 118, "len_sents": 50013, "len_sents_1": 25338, "n_sents": 348, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1502, "title": "Adaptive Tool Use in Large Language Models with Meta-Cognition Trigger", "abs": "Large language models (LLMs) have demonstrated remarkable emergent capabilities, reshaping the landscape of functional tasks by leveraging external tools to tackle complex problems, such as those requiring real-time data or specialized input/output processing. Existing research primarily focuses on equipping LLMs with a broader array of diverse external tools (e.g., program interpreters, search engines, weather/map applications) but overlooks the necessity of tool usage, invoking external tools indiscriminately without assessing their actual need. This naive strategy leads to two significant issues: 1) increased latency due to prolonged processing times, and 2) potential errors arising from communication between LLMs and external tools, resulting in faulty outputs. In this paper, we introduce a concept we term meta-cognition as a proxy for LLM self-capability, and we propose an adaptive decision-making strategy for invoking external tools, referred to as MeCo. Specifically, MeCo focuses on representation space to capture emergent representations of high-level cognitive phenomena that quantify the LLM's meta-cognitive scores, thereby guiding decisions on when to use external tools. Notably, MeCo is fine-tuning-free, incurring minimal cost, and our experiments demonstrate that MeCo accurately detects the model's internal cognitive signals. More importantly, our approach significantly enhances decision-making accuracy in tool use for multiple base models across various benchmarks.", "keywords": [ "Large Language Model", "Meta-Cognition", "Tool Use", "Representation Engineering" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "YVVz3vyqwr", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Mohammad Yaqub", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Associate Professor" }, { "name": "Mostafa Salem", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Salma Hassan", "gender": "Female", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 8, "n_ref_uni": 26, "n_ref": 30, "n_ref_all": 43, "n_fig": 4, "n_tab": 6, "L_tab": 4701, "n_element_tab": 181, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 724, "n_element_tab_1": 59, "formula_len_all": 203, "formula_len_all_1": 344, "len_all": 116486, "len_all_1": 55213, "len_abs": 4112, "len_title": 168, "len_sents": 37823, "len_sents_1": 26537, "n_sents": 251, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 119, "L_abs": 1721, "title": "GENRAD: Genomics and Radiomics Heterogeneous Graph Neural Network for Graph-Level Classification in Alzheimer's Disease", "abs": "Alzheimer\u2019s Disease (AD) poses multifaceted challenges due to its neurodegenerative nature driven by complex genomic, radiomic, and structural interactions. Understanding these complex relationships is pivotal for advancing diagnostic and therapeutic approaches. Current models struggle to effectively integrate multimodal data for AD, limiting their predictive accuracy and biological interpretability. Thus, there is a pressing need for models that can seamlessly fuse genomic and radiomic data to provide a holistic understanding of AD pathology. We introduce GENRAD, a novel heterogeneous graph neural network (GNN) that integrates multimodal genomic and radiomic data for graph-level classification in AD by representing patients, genes, and brain structures as distinct nodes and implementing advanced message-passing techniques. The benefits of GENRAD are fourfold: (1) It enables multimodal fusion of genomic and radiomic data, uncovering biologically meaningful insights missed by single-modality models. (2) Its adaptive multi-scale graph representations model interactions at various biological scales, capturing complex relationships essential for understanding AD pathology. (3) GENRAD incorporates explainable AI techniques, providing detailed analysis of key genomic markers and brain regions associated with AD. (4) GENRAD performs unsupervised clustering of genes, allowing the identification of functionally related biological pathways, thus empowering clinicians with actionable insights for personalized treatment strategies. GENRAD demonstrates superior classification accuracy in identifying AD-related patterns compared to existing machine and deep learning models, achieving an accuracy of 91.70%.", "keywords": [ "Heterogeneous Graph Neural Network (GNN)", "Alzheimer's Disease", "Multimodal data fusion", "Genomics", "Radiomics", "Node-level classification", "Multi-scale graph representation", "Explainable AI" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 1, 2, 1, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "YUaQ6qrbRL", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Bo Lei", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Principal Researcher" }, { "name": "Guanglong Sun", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hongwei Yan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Liyuan Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Qian Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Assistant Professor" }, { "name": "Yi Zhong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 66, "n_ref": 109, "n_ref_all": 127, "n_fig": 7, "n_tab": 9, "L_tab": 3894, "n_element_tab": 442, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 535, "n_element_tab_1": 71, "formula_len_all": 1898, "formula_len_all_1": 1285, "len_all": 165797, "len_all_1": 64685, "len_abs": 1202, "len_title": 103, "len_sents": 40655, "len_sents_1": 30136, "n_sents": 298, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1250, "title": "Right Time to Learn: Promoting Generalization via Bio-inspired Spacing Effect in Knowledge Distillation", "abs": "Knowledge distillation (KD) is a powerful strategy for training deep neural networks (DNNs). While it was originally proposed to train a more compact \u201cstudent\u201d model from a large \u201cteacher\u201d model, many recent efforts have focused on adapting it as an effective way to promote generalization of the model itself, such as online KD and self KD. Here, we propose an easy-to-use and compatible strategy named Spaced KD to improve the effectiveness of both online KD and self KD, in which the student model distills knowledge from a teacher model trained with a space interval ahead. This strategy is inspired by a prominent theory named spacing effect in the field of biological learning and memory, positing that appropriate intervals between learning trials can significantly enhance learning performance. We provide an in-depth theoretical and empirical analysis showing that the benefits of the proposed spacing effect in KD stem from seeking a flat minima during stochastic gradient descent (SGD). We perform extensive experiments to demonstrate the effectiveness of our Spaced KD in improving the learning performance of DNNs (e.g., the additional performance gain is up to 2.31% and 3.34% on Tiny-ImageNet over online KD and self KD, respectively).", "keywords": [ "Knowledge Distillation", "Brain-inspired Al", "Machine Learning", "Spacing effect" ], "rating_list": [ 8, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "YUYJsHOf3c", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Chen Xing", "gender": "Female", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "Chien-Sheng Wu", "gender": "Male", "institution": "Salesforce AI", "country": "US", "position": "Researcher" }, { "name": "Congying Xia", "gender": "Female", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "XIANGYU PENG", "gender": "Female", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Xinyi Yang", "gender": "Female", "institution": "Salesforce Research", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 32, "n_ref": 69, "n_ref_all": 106, "n_fig": 4, "n_tab": 13, "L_tab": 3732, "n_element_tab": 222, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 1444, "n_element_tab_1": 100, "formula_len_all": 106, "formula_len_all_1": 106, "len_all": 222729, "len_all_1": 78463, "len_abs": 1666, "len_title": 121, "len_sents": 80982, "len_sents_1": 37120, "n_sents": 757, "n_sents_1": 270, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 72, "L_abs": 1693, "title": "ReGenesis: LLMs can Grow into Reasoning Generalists via Self-Improvement", "abs": "Post-training Large Language Models (LLMs) with explicit reasoning trajectories can enhance their reasoning abilities. However, acquiring such high-quality trajectory data typically demands meticulous supervision from humans or superior models, which can be either expensive or license-constrained. In this paper, we explore how far an LLM can improve its reasoning by self-synthesizing reasoning paths as training data without any additional supervision. Existing self-synthesizing methods, such as STaR, suffer from poor generalization to out-of-domain (OOD) reasoning tasks. We hypothesize it is due to that their self-synthesized reasoning paths are too task-specific, lacking general task-agnostic reasoning guidance. To address this, we propose **Reasoning Generalist via Self-Improvement (ReGenesis)**, a method to *self-synthesize reasoning paths as post-training data by progressing from abstract to concrete*. More specifically, ReGenesis self-synthesizes reasoning paths by converting general reasoning guidelines into task-specific ones, generating reasoning structures, and subsequently transforming these structures into reasoning paths, without the need for human-designed task-specific examples used in existing methods. We show that ReGenesis achieves superior performance on all in-domain and OOD settings tested compared to existing methods. For six OOD tasks specifically, while previous methods exhibited an average performance decrease of approximately 4.6% after post training, ReGenesis delivers around 6.1% performance improvement. We also conduct an in-depth analysis of our framework and show ReGenesis is effective across various language models and design choices.", "keywords": [ "LLM", "reasoning", "generalization", "self-improvement" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "YTxx02MnTS", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Hao Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Nan Xu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 11, "n_ref_uni": 33, "n_ref": 68, "n_ref_all": 91, "n_fig": 5, "n_tab": 4, "L_tab": 1146, "n_element_tab": 188, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 735, "n_element_tab_1": 115, "formula_len_all": 873, "formula_len_all_1": 430, "len_all": 128915, "len_all_1": 68363, "len_abs": 1150, "len_title": 150, "len_sents": 37761, "len_sents_1": 31927, "n_sents": 304, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1153, "title": "A Dual-branch Multi-Band Neural Vocoder with Harmonic Discriminator for High-Fidelity Speech Synthesis", "abs": "Recent developments in vocoders are primarily dominated by GAN-based networks targeting to high-quality waveform generation from mel-spectrogram representations. However, these methods are typically computationally expensive and operate in the time-domain which neglect the time-frequency structures. In this paper, we propose the DMNet, a Dual-branch Multi-band Network to address these limitations. First, a reconstruction network of complex-valued spectrogram called CondNet is used as a condition and thus integrated into the GAN-based branch. Second, we use multi-band processing in the dual-branch: the CondNet produces Fourier spectral coefficients in one sub-band signal and GAN-based branch generates sub-band representations which are subsequently transformed to full-band speech. Finally, to further improve fidelity, we propose a novel harmonic discriminator which utilizes learnable harmonic filters at multiple scales for a better modeling ability in harmonic structures. In our experiments, DMNet validates the effectiveness and achieves superior performance for high quality waveform generation, both on subjective and objective metrics.", "keywords": [ "Dual-branch", "multi-band", "CondNet", "harmonic discriminator" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 1, 2, 3, 1 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "YTEwJaBdh0", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Aryan Mokhtari", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Bingcong Li", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Liang Zhang", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "", "position": "PhD student" }, { "name": "Niao He", "gender": "unknown", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 95, "n_formula_1": 18, "n_ref_uni": 46, "n_ref": 116, "n_ref_all": 158, "n_fig": 9, "n_tab": 7, "L_tab": 1354, "n_element_tab": 161, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2280, "n_element_tab_1": 135, "formula_len_all": 7686, "formula_len_all_1": 1437, "len_all": 257711, "len_all_1": 72244, "len_abs": 861, "len_title": 110, "len_sents": 68065, "len_sents_1": 29613, "n_sents": 759, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 863, "title": "On the Crucial Role of Initialization for Matrix Factorization", "abs": "This work revisits the classical low-rank matrix factorization problem and unveils the critical role of initialization in shaping convergence rates for such nonconvex and nonsmooth optimization. We introduce Nystrom initialization, which significantly improves the global convergence of Scaled Gradient Descent (ScaledGD) in both symmetric and asymmetric matrix factorization tasks. Specifically, we prove that ScaledGD with Nystrom initialization achieves quadratic convergence in cases where only linear rates were previously known. Furthermore, we extend this initialization to low-rank adapters (LoRA) commonly used for finetuning foundation models. Our approach, NoRA, i.e., LoRA with Nystrom initialization, demonstrates superior performance across various downstream tasks and model scales, from 1B to 7B parameters, in large language and diffusion models.", "keywords": [ "nonconvex optimization", "initialization", "quadratic rate", "low rank adapter", "lora" ], "rating_list": [ 8, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 3, 2, 4 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "YSJNKWOjKV", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Jian Zhou", "gender": "Male", "institution": "University of Texas Southwestern Medical Center", "country": "US", "position": "Assistant Professor" }, { "name": "Yuchen Zhang", "gender": "Male", "institution": "University of Texas Southwestern Medical Center", "country": "US", "position": "PhD student" } ], "n_formula": 46, "n_formula_1": 15, "n_ref_uni": 32, "n_ref": 72, "n_ref_all": 94, "n_fig": 3, "n_tab": 6, "L_tab": 1671, "n_element_tab": 158, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 495, "n_element_tab_1": 13, "formula_len_all": 5330, "formula_len_all_1": 1184, "len_all": 135340, "len_all_1": 48377, "len_abs": 1581, "len_title": 35, "len_sents": 39346, "len_sents_1": 20839, "n_sents": 322, "n_sents_1": 137, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1484, "title": "Inverse Flow and Consistency Models", "abs": "Inverse generation problems, such as denoising without ground truth observations, is a critical challenge in many scientific inquiries and real-world applications. While recent advances in generative models like diffusion models, conditional flow matching, and consistency models achieved impressive results by casting generation as denoising problems, they cannot be directly used for inverse generation without access to clean data. Here we introduce Inverse Flow (IF), a novel framework that enables using these generative models for inverse generation problems including denoising without ground truth. Inverse Flow can be flexibly applied to nearly any continuous noise distribution and allows complex dependencies. We propose two algorithms for learning Inverse Flows, Inverse Flow Matching (IFM) and Inverse Consistency Model (ICM). Notably, to derive the computationally efficient, simulation-free inverse consistency model objective, we generalized consistency training to any forward diffusion processes or conditional flows, which have applications beyond denoising. We demonstrate the effectiveness of IF on synthetic and real datasets, outperforming prior approaches while enabling noise distributions that previous methods cannot support. Finally, we showcase applications of our techniques to fluorescence microscopy and single-cell genomics data, highlighting IF's utility in scientific problems. This work opens up the use of powerful generative models for denoising.", "keywords": [ "diffusion models; flow matching; consistency models; denoising" ], "rating_list": [ 6, 3, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 1 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "YSA0QeYnDd", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Chao Yang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Shuang Li", "gender": "Female", "institution": "The Chinese University of Hong Kong (Shenzhen)", "country": "", "position": "Assistant Professor" }, { "name": "Shuting Cui", "gender": "Female", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Yang Yang", "gender": "Male", "institution": "The Chinese University of Hong Kong,Shenzhen", "country": "", "position": "MS student" } ], "n_formula": 34, "n_formula_1": 24, "n_ref_uni": 56, "n_ref": 94, "n_ref_all": 116, "n_fig": 9, "n_tab": 10, "L_tab": 2667, "n_element_tab": 226, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1400, "n_element_tab_1": 147, "formula_len_all": 2804, "formula_len_all_1": 1645, "len_all": 194446, "len_all_1": 69213, "len_abs": 1460, "len_title": 143, "len_sents": 62497, "len_sents_1": 29709, "n_sents": 473, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1468, "title": "Inference of Evolving Mental States from Irregular Action Events to Understand Human Behaviors", "abs": "Inference of latent human mental processes, such as belief, intention, or desire, is crucial for developing AI with human-like intelligence, enabling more effective and timely collaboration. In this paper, we introduce a versatile encoder-decoder model designed to infer evolving mental processes based on irregularly observed action events and predict future occurrences. The primary challenges arise from two factors: both actions and mental processes are irregular events, and the observed action data is often limited. To address the irregularity of these events, we leverage a temporal point process model within the encoder-decoder framework, effectively capturing the dynamics of both action and mental events. Additionally, we implement a backtracking mechanism in the decoder to enhance the accuracy of predicting future actions and evolving mental states. To tackle the issue of limited data, our model incorporates logic rules as priors, enabling accurate inferences from just a few observed samples. These logic rules can be refined and updated as needed, providing flexibility to the model. Overall, our approach enhances the understanding of human behavior by predicting when actions will occur and how mental processes evolve. Experiments on both synthetic and real-world datasets demonstrate the strong performance of our model in inferring mental states and predicting future actions, contributing to the development of more human-centric AI systems.", "keywords": [ "temporal point process", "logic rule", "human-AI collaboration" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 4, 2, 2 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 2, 2 ] }, { "paper_id": "YS5zdlSzvv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Chia-Wen Lin", "gender": "Male", "institution": "National Tsing Hua University", "country": "TW", "position": "Professor" }, { "name": "Chih-Ling Chang", "gender": "unknown", "institution": "National Tsinghua University", "country": "TW", "position": "MS student" }, { "name": "Fu-Jen Tsai", "gender": "Male", "institution": "National Tsinghua University", "country": "TW", "position": "PhD student" }, { "name": "Lin Gu", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Researcher" }, { "name": "ZILING HUANG", "gender": "Female", "institution": "University of Amsterdam", "country": "NL", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 40, "n_ref": 90, "n_ref_all": 112, "n_fig": 12, "n_tab": 4, "L_tab": 1335, "n_element_tab": 170, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 557, "n_element_tab_1": 97, "formula_len_all": 619, "formula_len_all_1": 526, "len_all": 122456, "len_all_1": 55630, "len_abs": 1143, "len_title": 159, "len_sents": 35908, "len_sents_1": 27911, "n_sents": 231, "n_sents_1": 154, "venue": "ICLR2025", "decision": "withdraw", "L_title": 109, "L_abs": 1146, "title": "Unleashing the Power of Deep Dehazing Models: A Physics-guided Parametric Augmentation Net for Image Rehazing", "abs": "Image dehazing faces significant challenges in real-world scenarios due to the large domain gap between synthetic and real-world hazy images, which often hinders dehazing performance. Collecting real-world datasets is particularly difficult, as hazy and clean image pairs must be captured under identical conditions. To address this, we propose a Physics-guided Parametric Augmentation Network (PANet) that generates realistic hazy and clean training pairs, enhancing dehazing performance in real-world applications. PANet consists of two components: a Haze-to-Parameter Mapper (HPM), which projects hazy images into a parametric space representing haze characteristics, and a Parameter-to-Haze Mapper (PHM), which converts resampled haze parameters back into hazy images. By resampling individual haze parameter maps at the pixel level in the parametric space, PANet generates diverse hazy images with physically explainable haze conditions that are not present in the training data. Our experimental results show that PANet effectively enriches existing hazy image benchmarks, significantly improving the performance of current dehazing models.", "keywords": [ "Image dehazing", "Image rehazing", "Data augmentation" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 5, 4 ] }, { "paper_id": "YRm9BMTLv6", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Coby Penso", "gender": "Male", "institution": "Bar Ilan University", "country": "", "position": "MS student" }, { "name": "Idit Diamant", "gender": "unknown", "institution": "Sony Semiconductor Israel", "country": "IL", "position": "Researcher" }, { "name": "Jacob Goldberger", "gender": "Male", "institution": "Bar-Ilan University", "country": "", "position": "Associate Professor" }, { "name": "Rotem Nizhar", "gender": "Male", "institution": "School of Computer Science, Tel Aviv University", "country": "IL", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 30, "n_ref": 76, "n_ref_all": 97, "n_fig": 6, "n_tab": 20, "L_tab": 15335, "n_element_tab": 1963, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 3523, "n_element_tab_1": 500, "formula_len_all": 659, "formula_len_all_1": 563, "len_all": 164225, "len_all_1": 65286, "len_abs": 952, "len_title": 97, "len_sents": 36160, "len_sents_1": 26151, "n_sents": 274, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 944, "title": "Source-Free Target Domain Confidence Calibration", "abs": "In this study, we consider the setup of source-free domain adaptation and address the challenge of calibrating the confidence of a model adapted to the target domain using only unlabeled data. The primary challenge in addressing uncertainty calibration is the absence of labeled data which prevents computing the accuracy of the adapted network on the target domain. We address this by leveraging pseudo-labels generated from the source model\u2019s predictions to estimate the true, unobserved accuracy. We demonstrate that, although the pseudo-labels are noisy, the network accuracy calculated using these pseudo-labels is similar to the accuracy obtained with the correct labels. We validate the effectiveness of our calibration approach by applying it to standard domain adaptation datasets and show that it achieves results comparable to, or even better than, previous calibration methods that relied on the availability of labeled source data.", "keywords": [ "confidence calibration", "domain adaptation", "source-free" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "YR79EyejsG", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Cong Wang", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "Postdoc" }, { "name": "Jens Kober", "gender": "Male", "institution": "Delft University of Technology", "country": "", "position": "Associate Professor" }, { "name": "Pengzhi Yang", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "Researcher" }, { "name": "Ruipeng Zhang", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Xinyu Wang", "gender": "Male", "institution": "Booking", "country": "NL", "position": "Researcher" }, { "name": "Frans Oliehoek", "gender": "Male", "institution": "Delft University of Technology", "country": "", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 57, "n_ref": 83, "n_ref_all": 111, "n_fig": 13, "n_tab": 4, "L_tab": 1639, "n_element_tab": 83, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 587, "n_element_tab_1": 23, "formula_len_all": 186, "formula_len_all_1": 187, "len_all": 151456, "len_all_1": 57738, "len_abs": 1486, "len_title": 131, "len_sents": 44158, "len_sents_1": 29173, "n_sents": 283, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1033, "title": "Task-Unaware Lifelong Robot Learning with Retrieval-based Weighted Local Adaptation", "abs": "Real-world environments require robots to continuously acquire new skills while retaining previously learned abilities, all without the need for clearly defined task boundaries. Storing all past data to prevent forgetting is impractical due to storage and privacy concerns. To address this, we propose a method that efficiently restores a robot's proficiency in previously learned tasks over its lifespan. Using an Episodic Memory (EM), our approach enables experience replay during training and retrieval during testing for local fine-tuning, allowing rapid adaptation to previously encountered problems. Additionally, we introduce a selective weighting mechanism that emphasizes the most challenging segments of retrieved demonstrations, focusing local adaptation where it is most needed. This framework offers a scalable solution for lifelong learning without explicit task identifiers or implicit task boundaries, combining retrieval-based adaptation with selective weighting to enhance robot performance in open-ended scenarios.", "keywords": [ "Robotic Lifelong Learning", "Task-Unaware Continual Learning", "Episodic Memory Retrieval", "Visuomotor Behavior Cloning", "Error-driven Policy Adaptation" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 4, 2, 3, 1 ], "presentation_list": [ 4, 3, 3, 1 ], "contribution_list": [ 3, 3, 4, 1 ], "confidence_list": [ 5, 2, 3, 4 ] }, { "paper_id": "YQvvJjLWX0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Jannik Kossen", "gender": "Unspecified", "institution": "Facebook", "country": "DE", "position": "Researcher" }, { "name": "Jiatong Han", "gender": "unknown", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "MS student" }, { "name": "Shreshth A Malik", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Yarin G", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" }, { "name": "lisa schut", "gender": "Female", "institution": "University of Oxford, University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Muhammed T Razzak", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 66, "n_ref": 139, "n_ref_all": 164, "n_fig": 8, "n_tab": 6, "L_tab": 4236, "n_element_tab": 259, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 208, "n_element_tab_1": 80, "formula_len_all": 1125, "formula_len_all_1": 413, "len_all": 208642, "len_all_1": 68911, "len_abs": 1311, "len_title": 121, "len_sents": 50920, "len_sents_1": 32627, "n_sents": 417, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1309, "title": "Semantic Entropy Probes: Robust and Cheap Hallucination Detection in LLMs", "abs": "We propose semantic entropy probes (SEPs), a cheap and reliable method for uncertainty quantification in Large Language Models (LLMs). Hallucinations, which are plausible-sounding but factually incorrect and arbitrary model generations, present a major challenge to the practical adoption of LLMs. Recent work by Farquhar et al. proposes semantic entropy (SE), which can reliably detect hallucinations by quantifying the uncertainty over different generations by estimating entropy over semantically equivalent sets of outputs. However, the 5-to-10-fold increase in computation cost associated with SE computation hinders practical adoption. To address this, we propose SEPs, which directly approximate SE from the hidden states of a single generation. SEPs are simple to train and do not require sampling multiple model generations at test time, reducing the overhead of semantic uncertainty quantification to almost zero. We show that SEPs retain high performance for hallucination detection and generalize better to out-of-distribution data than previous probing methods that directly predict model accuracy. Our results across models and tasks suggest that model hidden states capture SE, and our ablation studies give further insights into the token positions and model layers for which this is the case.", "keywords": [ "large language models; hallucinations; linear probing; uncertainty estimation; interpretability" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "YQjdNC0NkW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Chenxing Li", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Dong Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Distinguished Scientist" }, { "name": "Manjie Xu", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Rilin Chen", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Wei Liang", "gender": "Female", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xinyi Tu", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" }, { "name": "Yong Ren", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yu Gu", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 33, "n_ref": 61, "n_ref_all": 75, "n_fig": 5, "n_tab": 5, "L_tab": 1648, "n_element_tab": 256, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1359, "n_element_tab_1": 210, "formula_len_all": 136, "formula_len_all_1": 136, "len_all": 109258, "len_all_1": 52424, "len_abs": 1211, "len_title": 95, "len_sents": 27565, "len_sents_1": 23034, "n_sents": 199, "n_sents_1": 165, "venue": "ICLR2025", "decision": "withdraw", "L_title": 47, "L_abs": 1215, "title": "Video-to-Audio generation with Hidden Alignment", "abs": "Generating semantically and temporally aligned audio content in accordance with video input has become a focal point for researchers, particularly following the remarkable breakthrough in text-to-video generation. In this work, we aim to offer insights into the video-to-audio generation paradigm, focusing on three crucial aspects: vision encoders, auxiliary embeddings, and data augmentation techniques.\nBeginning with a foundational model built on a simple yet surprisingly effective intuition, we explore various vision encoders and auxiliary embeddings through ablation studies. Employing a comprehensive evaluation pipeline that emphasizes generation quality and video-audio synchronization alignment, we demonstrate that our model exhibits state-of-the-art video-to-audio generation capabilities. Furthermore, we provide critical insights into the impact of different data augmentation methods on enhancing the generation framework\u2019s overall capacity. We showcase possibilities to advance the challenge of generating synchronized audio from semantic and temporal perspectives. We hope these insights will serve as a stepping stone toward developing more realistic and accurate audio-visual generation models.", "keywords": [ "audio generation", "video-to-audio", "diffusion" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 2, 4, 3 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "YPvI7SofeZ", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Aniket Chatterjee", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Undergrad student" }, { "name": "Axel Kuhn", "gender": "Not Specified", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" }, { "name": "Jan Ole Ernst", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Tim Franzmeyer", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 11, "n_ref_uni": 68, "n_ref": 111, "n_ref_all": 145, "n_fig": 13, "n_tab": 4, "L_tab": 615, "n_element_tab": 93, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 321, "n_element_tab_1": 14, "formula_len_all": 1828, "formula_len_all_1": 451, "len_all": 239431, "len_all_1": 63718, "len_abs": 1373, "len_title": 121, "len_sents": 68570, "len_sents_1": 30186, "n_sents": 496, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1336, "title": "Scaleable Quantum Control via Physics Constrained Reinforcement Learning", "abs": "Quantum optimal control is concerned with the realisation of desired dynamics in quantum systems, serving as a linchpin for advancing quantum technologies and fundamental research. \nAnalytic approaches and standard optimisation algorithms do not yield satisfactory solutions for large quantum systems, and especially not for real world quantum systems which are open and noisy. \nWe devise a physics-informed Reinforcement Learning (RL) algorithm that restricts the space of possible solutions.\nWe incorporate priors about the desired time scales of the quantum state dynamics -- as well as realistic control signal limitations -- as constraints to the RL algorithm. \nThese physics-informed constraints additionally improve computational scalability by facilitating parallel optimisation. \nWe evaluate our method on three broadly relevant quantum systems (multi-level $\\Lambda$ system, Rydberg atom and superconducting transmon) and incorporate real-world complications, arising from dissipation and control signal perturbations. \nWe achieve both higher fidelities -- which exceed 0.999 across all systems -- and better robustness to time-dependent perturbations and experimental imperfections than previous methods. \nLastly, we demonstrate that incorporating multi-step feedback can yield solutions robust even to strong perturbations.", "keywords": [ "reinforcement learning", "quantum computing", "quantum control", "quantum dynamics", "control theory" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "YOrN9vNrqo", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Fenia Christopoulou", "gender": "Female", "institution": "Poolside AI", "country": "FR", "position": "Member of Engineering" }, { "name": "Gerasimos Lampouras", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Principal Researcher" }, { "name": "Haitham Bou Ammar", "gender": "Male", "institution": "Huawei R&D UK", "country": "GB", "position": "Principal Researcher" }, { "name": "Jun Wang", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Professor" }, { "name": "Ronald Cardenas", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Researcher" } ], "n_formula": 26, "n_formula_1": 11, "n_ref_uni": 38, "n_ref": 61, "n_ref_all": 114, "n_fig": 19, "n_tab": 7, "L_tab": 2061, "n_element_tab": 331, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 861, "n_element_tab_1": 157, "formula_len_all": 5243, "formula_len_all_1": 1155, "len_all": 182667, "len_all_1": 67796, "len_abs": 1507, "len_title": 121, "len_sents": 51322, "len_sents_1": 30362, "n_sents": 386, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1514, "title": "SparsePO: Controlling Preference Alignment of LLMs via Sparse Token Masks", "abs": "Preference Optimization (PO) has proven an effective step for aligning language models to human-desired behaviors. Current variants, following the offline Direct Preference Optimization objective, have focused on a strict setting where all tokens are contributing signals of KL divergence and rewards to the loss function. However, human preference is not affected by each word in a sequence equally but is often dependent on specific words or phrases, e.g. existence of toxic terms leads to non-preferred responses. Based on this observation, we argue that not all tokens should be weighted equally during PO and propose a flexible objective termed SparsePO, that aims to automatically learn to weight the KL divergence and reward corresponding to each token during PO training. We propose two different variants of weight-masks that can either be derived from the reference model itself or learned on the fly. Notably, our method induces sparsity in the learned masks, allowing the model to learn how to best weight reward and KL divergence contributions at the token level, learning an optimal level of mask sparsity. Extensive experiments on multiple domains, including sentiment control, dialogue, text summarization and text-to-code generation, illustrate that our approach assigns meaningful weights to tokens according to the target task, generates more responses with the desired preference and improves reasoning tasks by up to 2 percentage points compared to other token- and sentence-level PO methods.", "keywords": [ "preference optimization", "alignment with human preferences" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "YOpa6dTrpt", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Jianfu Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Li Niu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Liqing Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Pai Peng", "gender": "unknown", "institution": "Cowarobot", "country": "CN", "position": "Principal Researcher" }, { "name": "Tao He", "gender": "Male", "institution": "South China University", "country": "CN", "position": "Full Professor" }, { "name": "Wenlong Liao", "gender": "Male", "institution": "COWAROBOT", "country": "CN", "position": "Researcher" }, { "name": "Xinhao Hu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yasushi Makihara", "gender": "Male", "institution": "Osaka University", "country": "", "position": "Professor" }, { "name": "Yasushi Yagi", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "Full Professor" }, { "name": "Yichen Wang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "", "position": "MS student" }, { "name": "Yiyi Zhang", "gender": "Female", "institution": "Cornell University", "country": "US", "position": "Research Scientist" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 65, "n_ref_all": 86, "n_fig": 9, "n_tab": 6, "L_tab": 2674, "n_element_tab": 372, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 3762, "n_element_tab_1": 282, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 134387, "len_all_1": 51582, "len_abs": 370, "len_title": 175, "len_sents": 27687, "len_sents_1": 22039, "n_sents": 194, "n_sents_1": 135, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 127, "L_abs": 1323, "title": "Pedestrian Motion Reconstruction: A Large-scale Benchmark via Mixed Reality Rendering with Multiple Perspectives and Modalities", "abs": "Reconstructing pedestrian motion from dynamic sensors, with a focus on pedestrian intention, is crucial for advancing autonomous driving safety. However, this task is challenging due to data limitations arising from technical complexities, safety, and cost concerns. We introduce the Pedestrian Motion Reconstruction (PMR) dataset, which focuses on pedestrian intention to reconstruct behavior using multiple perspectives and modalities. PMR is developed from a mixed reality platform that combines real-world realism with the extensive, accurate labels of simulations, thereby reducing costs and risks. It captures the intricate dynamics of pedestrian interactions with objects and vehicles, using different modalities for a comprehensive understanding of human-vehicle interaction. Analyses show that PMR can naturally exhibit pedestrian intent and simulate extreme cases. PMR features a vast collection of data from 54 subjects interacting across 13 urban settings with 7 objects, encompassing 12,138 sequences with diverse weather conditions and vehicle speeds. This data provides a rich foundation for modeling pedestrian intent through multi-view and multi-modal insights. We also conduct comprehensive benchmark assessments across different modalities to thoroughly evaluate pedestrian motion reconstruction methods.", "keywords": [ "Pedestrian Dynamics", "Mixed Reality" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "YOc5t8PHf2", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Irina Rish", "gender": "Female", "institution": "University of Montreal", "country": "CA", "position": "Professor" }, { "name": "Ivan Anokhin", "gender": "Male", "institution": "Mila, Universit\u00e9 de Montr\u00e9al", "country": "", "position": "PhD student" }, { "name": "Matthew Riemer", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, University of Montreal", "country": "CA", "position": "PhD student" }, { "name": "Rishav Rishav", "gender": "Male", "institution": "University of Calgary", "country": "CA", "position": "MS student" }, { "name": "Samira Ebrahimi Kahou", "gender": "Female", "institution": "University of Calgary", "country": "CA", "position": "Assistant Professor" }, { "name": "Stephen Chung", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 30, "n_ref": 45, "n_ref_all": 85, "n_fig": 14, "n_tab": 9, "L_tab": 3632, "n_element_tab": 290, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 390, "n_element_tab_1": 19, "formula_len_all": 430, "formula_len_all_1": 230, "len_all": 147695, "len_all_1": 67730, "len_abs": 1430, "len_title": 83, "len_sents": 53655, "len_sents_1": 33230, "n_sents": 417, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1450, "title": "Handling Delay in Reinforcement Learning Caused by Parallel Computations of Neurons", "abs": "Real-time reinforcement learning (RL) introduces several challenges. First, policies are constrained to a fixed number of actions per second due to hardware limitations. Second, the environment may change while the network is still computing an action, leading to observational delay. The first issue can partly be addressed with parallel computation of neurons, leading to higher throughput and potentially better policies. However, the second issue remains: if each neuron operates in parallel with an execution time of $\\tau$, an $N$-layer feed-forward network experiences observation delay of $\\tau N$.\nReducing the number of layers can decrease this delay, but at the cost of the network's expressivity. In this work, we explore the trade-off between minimizing delay and network's expressivity. We present a theoretically motivated solution that leverages temporal skip connections combined with history-augmented observations. We evaluate several architectures and show that those incorporating temporal skip connections achieve strong performance across various neuron execution times, reinforcement learning algorithms, and environments, including four Mujoco tasks and all MinAtar games. Moreover, we demonstrate parallel neuron computation can accelerate inference by 6-350\\% on standard hardware. Our investigation into temporal skip connections and parallel computations paves the way for more efficient RL agents in real-time setting.", "keywords": [ "reinforcement learning", "delay", "parallel computations" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 1 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "YONCcPQJoC", "primary_area": "generative models", "n_author": 11, "author_info_list": [ { "name": "Carl Yang", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Assistant Professor" }, { "name": "Honglei Zhuang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Jiaming Shen", "gender": "unknown", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Li Chen", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "PhD student" }, { "name": "Michael Bendersky", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Sumit Sanghai", "gender": "Male", "institution": "Research, Google", "country": "", "position": "Researcher" }, { "name": "Xuanhui Wang", "gender": "Male", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Yi Liang", "gender": "Male", "institution": "Research, Google", "country": "", "position": "Researcher" }, { "name": "Yiling Jia", "gender": "Female", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "You Wu", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Zhen Qin", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 30, "n_ref": 46, "n_ref_all": 56, "n_fig": 2, "n_tab": 6, "L_tab": 1003, "n_element_tab": 77, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 849, "n_element_tab_1": 62, "formula_len_all": 887, "formula_len_all_1": 868, "len_all": 106685, "len_all_1": 61908, "len_abs": 1160, "len_title": 111, "len_sents": 33697, "len_sents_1": 29358, "n_sents": 244, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1165, "title": "Integrating Planning into Single-Turn Long-Form Text Generation", "abs": "Generating high-quality, in-depth textual documents, such as academic papers, news articles, Wikipedia entries, and books, remains a significant challenge for Large Language Models (LLMs). In this paper, we propose to use planning to generate long form content. To achieve our goal, we generate intermediate steps via an auxiliary task that teaches the LLM to plan, reason and structure before generating the final text. Our main novelty lies in a single auxiliary task that does not require multiple rounds of prompting or planning. To overcome the scarcity of training data for these intermediate steps, we leverage LLMs to generate synthetic intermediate writing data such as outlines, key information and summaries from existing full articles. Our experiments demonstrate on two datasets from different domains, namely the scientific news dataset SciNews and Wikipedia datasets in KILT-Wiki and FreshWiki, that LLMs fine-tuned with the auxiliary task generate higher quality documents. We observed +2.5% improvement in ROUGE-Lsum, and a strong 3.60 overall win/loss ratio via human SxS evaluation, with clear wins in organization, relevance, and verifiability.", "keywords": [ "LLM", "planning", "long-form text generation" ], "rating_list": [ 5, 3, 8, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "YO6Je9jOJI", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Jinsung Yoon", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Meeyoung Cha", "gender": "Female", "institution": " Max Planck Institute for Security and Privacy", "country": "DE", "position": "Director" }, { "name": "Sercan O Arik", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Seungeon Lee", "gender": "Male", "institution": "MPI-SP", "country": "DE", "position": "Intern" }, { "name": "Sungwon Han", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Integrated PhD student" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 52, "n_ref": 82, "n_ref_all": 112, "n_fig": 15, "n_tab": 8, "L_tab": 6635, "n_element_tab": 367, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 854, "n_element_tab_1": 45, "formula_len_all": 381, "formula_len_all_1": 322, "len_all": 191738, "len_all_1": 61871, "len_abs": 1108, "len_title": 126, "len_sents": 50085, "len_sents_1": 30133, "n_sents": 629, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1114, "title": "LLM-Guided Self-Supervised Tabular Learning With Task-Specific Pre-text Tasks", "abs": "One of the most common approaches for self-supervised representation learning is defining pre-text tasks to learn data representations. Existing works determine pre-text tasks in a \"task-agnostic'' way, without considering the forthcoming downstream tasks. This offers an advantage of broad applicability across tasks, but can also lead to a mismatch between task objectives, potentially degrading performance on downstream tasks. In this paper, we introduce TST-LLM, a framework that effectively reduces this mismatch when the natural language-based description of the downstream task is given without any ground-truth labels. TST-LLM instructs the LLM to use the downstream task's description and meta-information of data to discover features relevant to the target task. These discovered features are then treated as ground-truth labels to define \"target-specific'' pre-text tasks. TST-LLM consistently outperforms contemporary baselines, such as STUNT and LFR, with win ratios of 95% and 81%, when applied to 22 benchmark tabular datasets, including binary and multi-class classification, and regression tasks.", "keywords": [ "Self-supervised learning", "Representation learning", "Tabular data", "Large language model" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "YNbLUGDAX5", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Di Huang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Huiqun Wang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Nan Zhou", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Yaoyan Zheng", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 14, "n_ref": 24, "n_ref_all": 53, "n_fig": 10, "n_tab": 15, "L_tab": 5655, "n_element_tab": 502, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 1728, "n_element_tab_1": 230, "formula_len_all": 408, "formula_len_all_1": 408, "len_all": 124700, "len_all_1": 56709, "len_abs": 1184, "len_title": 123, "len_sents": 35179, "len_sents_1": 24070, "n_sents": 249, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1188, "title": "Progressive Parameter Efficient Transfer Learning for Semantic Segmentation", "abs": "Parameter Efficient Transfer Learning (PETL) excels in downstream classification fine-tuning with minimal computational overhead, demonstrating its potential within the pre-train and fine-tune paradigm. However, recent PETL methods consistently struggle when fine-tuning for semantic segmentation tasks, limiting their broader applicability. In this paper, we identify that fine-tuning for semantic segmentation requires larger parameter adjustments due to shifts in semantic perception granularity. Current PETL approaches are unable to effectively accommodate these shifts, leading to significant performance degradation. To address this, we introduce ProPETL, a novel approach that incorporates an additional midstream adaptation to progressively align pre-trained models for segmentation tasks. Through this process, ProPETL achieves state-of-the-art performance on most segmentation benchmarks and, for the first time, surpasses full fine-tuning on the challenging COCO-Stuff10k dataset. Furthermore, ProPETL demonstrates strong generalization across various pre-trained models and scenarios, highlighting its effectiveness and versatility for broader adoption in segmentation tasks.", "keywords": [ "Parameter Efficient Transfer Learning", "Semantic Segmentation" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "YNa0Mzx4P9", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Hao Gao", "gender": "unknown", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Jingwei Xu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Taolue Chen", "gender": "unknown", "institution": "Birkbeck College, University of London", "country": "", "position": "Lecturer" }, { "name": "Xiaoxing Ma", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Yunpeng Huang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 69, "n_ref_all": 89, "n_fig": 4, "n_tab": 4, "L_tab": 503, "n_element_tab": 68, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 365, "n_element_tab_1": 58, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 136593, "len_all_1": 63495, "len_abs": 762, "len_title": 125, "len_sents": 40163, "len_sents_1": 30834, "n_sents": 339, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 681, "title": "LASER: Script Execution by Autonomous Agents for On-demand Traffic Simulation", "abs": "Autonomous Driving Systems (ADS) require diverse and safety-critical traffic scenarios for effective training and testing, but the existing data generation methods struggle to provide flexibility and scalability. We propose LASER, a novel framework that leverage large language models (LLMs) to conduct traffic simulations based on natural language inputs. The framework operates in two stages: it first generates scripts from user-provided descriptions and then executes them using autonomous agents in real time. Validated in the CARLA simulator, LASER successfully generates complex, on-demand driving scenarios, significantly improving ADS training and testing data generation.", "keywords": [ "Traffic Simulation; Autonomous Agents; LLM for Autonomous Driving" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 1, 2, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "YNQF003Ad3", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Alexey Kamenev", "gender": "unknown", "institution": "Samara National Research University", "country": "", "position": "MS student" }, { "name": "Jan Kautz", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "VP Research" }, { "name": "Jean Kossaifi", "gender": "Male", "institution": "NVIDIA AI", "country": "", "position": "Researcher" }, { "name": "Kamyar Azizzadenesheli", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Max Rietmann", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Development Technology Engineer" }, { "name": "Christopher Choy", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 54, "n_ref": 108, "n_ref_all": 122, "n_fig": 6, "n_tab": 5, "L_tab": 985, "n_element_tab": 119, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 765, "n_element_tab_1": 74, "formula_len_all": 974, "formula_len_all_1": 840, "len_all": 148806, "len_all_1": 65224, "len_abs": 1272, "len_title": 142, "len_sents": 38991, "len_sents_1": 29618, "n_sents": 274, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1291, "title": "Factorized Implicit Global Convolution for Automotive Computational Fluid Dynamics Prediction", "abs": "Computational Fluid Dynamics (CFD) is crucial for automotive design, requiring the analysis of large 3D point clouds to study how vehicle geometry affects pressure fields and drag forces. However, existing deep learning approaches for CFD struggle with the computational complexity of processing high-resolution 3D data.\n We propose Factorized Implicit Global Convolution (FIGConv), a novel architecture that efficiently solves CFD problems for very large 3D meshes with arbitrary input and output geometries. FIGConv achieves quadratic complexity $O(N^2)$, a significant improvement over existing 3D neural CFD models that require cubic complexity $O(N^3)$. Our approach combines Factorized Implicit Grids to approximate high-resolution domains, efficient global convolutions through 2D reparameterization, and a U-shaped architecture for effective information gathering and integration.\n We validate our approach on the industry-standard Ahmed body dataset and the large-scale DrivAerNet dataset. On DrivAerNet, our model achieves an $R^2$ value of 0.95 for drag prediction, outperforming the previous state-of-the-art by a significant margin. This represents a 40\\% improvement in relative mean squared error and a 70\\% improvement in absolute mean squared error over prior methods.", "keywords": [ "Computational Fluid Dynamics", "Reparametrized Convolution" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "YMvRZCA8Zo", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Guanlin Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Han Qiu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Zhang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Shuai Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tianwei Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 72, "n_ref_all": 102, "n_fig": 2, "n_tab": 10, "L_tab": 738, "n_element_tab": 114, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1224, "n_element_tab_1": 53, "formula_len_all": 986, "formula_len_all_1": 590, "len_all": 158689, "len_all_1": 72095, "len_abs": 1132, "len_title": 97, "len_sents": 54892, "len_sents_1": 33858, "n_sents": 471, "n_sents_1": 265, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1137, "title": "PRIME: Protect Your Videos From Malicious Editing", "abs": "Over the years, video generation has experienced significant advancement. A variety of open-source models emerge, making it surprisingly easy to manipulate and edit videos with just a few simple prompts. While these cutting-edge technologies have gained huge popularity, they have also given rise to concerns regarding the privacy and portrait rights of individuals: malicious users can exploit these tools for deceptive or illegal purposes. Existing works on protecting images against generative models cannot be directly grafted to video protection, due to their efficiency and effectiveness limitations. Motivated by this, we introduce PRIME, a new methodology dedicated to the protection of videos from unauthorized editing via generative models. Our key idea is to craft highly transferable and robust perturbations, which can be efficiently added to the protected videos to disrupt their editing feasibility. We perform comprehensive evaluations using both objective metrics and human studies. The results indicate that PRIME only needs 8.3% GPU hours of existing state-of-the-art methods while achieving better protection results.", "keywords": [ "Diffusion Model", "Video Protection", "Video Editing" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "YMgMGPjUPg", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Cheng Jin", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Fan Yi", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Weilin Wan", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "present", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Weizhong Zhang", "gender": "Male", "institution": "Fudan University", "country": "HK", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 54, "n_ref": 163, "n_ref_all": 195, "n_fig": 9, "n_tab": 15, "L_tab": 5737, "n_element_tab": 1090, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 3359, "n_element_tab_1": 392, "formula_len_all": 328, "formula_len_all_1": 231, "len_all": 195570, "len_all_1": 66945, "len_abs": 2020, "len_title": 107, "len_sents": 59358, "len_sents_1": 30136, "n_sents": 423, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 2031, "title": "Out-of-Distribution Detection using Neural Activation Prior", "abs": "Out-of-distribution detection (OOD) is a crucial technique for deploying machine learning models in the real world to handle the unseen scenarios. Compared to standard classification tasks, OOD detection presents significant challenges due to the unpredictable nature and inherent difficulty in collecting OOD data. Consequently, a natural solution is to develop priors that are as diverse as possible, effectively characterizing the features of OOD data. In this paper, we first propose a simple yet effective Neural Activation Prior (NAP) for OOD detection.Our prior is based on a key observation that, for a channel before the pooling layer of a fully trained neural network, the probability of a few neurons being activated with a large response by an in-distribution (ID) sample is significantly higher than that by an OOD sample. An intuitive explanation is that for a model fully trained on ID dataset, each channel would play a role in detecting a certain pattern in the ID dataset, and a few neurons can be activated with a large response when the pattern is detected in an input sample. Then, an effective scoring function based on this prior is proposed to highlight the role of these strongly activated neurons in OOD detection. Our approach is plug-and-play and does not lead to any performance degradation on ID data classification and requires no extra training or statistics from training or external datasets. To the best of our knowledge, our method is the first to exploit intra-channel activation pattern information, contributing to its orthogonality to existing approaches and allowing it to be effectively combined with them in various applications. Furthermore, we conduct an elegant oracle experiment to validate the rationale behind our proposed scoring function. Extensive experimental results demonstrate the effectiveness of our method. Moreover, our approach can significantly boost the performance when integrated with most existing methods, showcasing the unique attributes of the proposed prior.", "keywords": [ "out-of-distribution detection", "prior", "intra-channel activation pattern" ], "rating_list": [ 5, 8, 3, 3 ], "soundness_list": [ 2, 4, 2, 2 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "YMCtQlm8Bc", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Dongyeop Lee", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Maksym Andriushchenko", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Postdoc" }, { "name": "Namhoon Lee", "gender": "unknown", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Sungbin Shin", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 77, "n_formula_1": 13, "n_ref_uni": 81, "n_ref": 153, "n_ref_all": 202, "n_fig": 23, "n_tab": 2, "L_tab": 1235, "n_element_tab": 99, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 836, "n_element_tab_1": 80, "formula_len_all": 8523, "formula_len_all_1": 649, "len_all": 240754, "len_all_1": 63493, "len_abs": 1453, "len_title": 122, "len_sents": 65726, "len_sents_1": 29529, "n_sents": 535, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1450, "title": "Critical Influence of Overparameterization on Sharpness-aware Minimization", "abs": "Training overparameterized neural networks often yields solutions with varying generalization capabilities, even when achieving similar training losses. Recent evidence indicates a strong correlation between the sharpness of a minimum and its generalization error, leading to increased interest in optimization methods that explicitly seek flatter minima for improved generalization. Despite its contemporary relevance to overparameterization, however, this sharpness-aware minimization (SAM) strategy has not been studied much yet as to exactly how it is affected by overparameterization. In this work, we analyze SAM under varying degrees of overparameterization, presenting both empirical and theoretical findings that reveal its critical influence on SAM's effectiveness. First, we conduct extensive numerical experiments across diverse domains and show that SAM consistently improves with overparameterization. Next, we attribute this phenomenon to the interplay between the enlarged solution space and increased implicit bias resulting from overparameterization. Furthermore, we show that this effect is particularly pronounced in practical settings involving label noise and sparsity, and yet, sufficient regularization is necessary. Last but not least, we provide other theoretical insights into how overparameterization helps SAM achieve minima with more uniform Hessian moments compared to SGD, and much faster convergence at a linear rate.", "keywords": [ "sharpness-aware minimization", "overparameterization" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 3, 2, 4, 5 ] }, { "paper_id": "YLTWwEjkdx", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Ang Li", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Guoheng Sun", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Shwai He", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Zheyu Shen", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 26, "n_ref": 47, "n_ref_all": 77, "n_fig": 10, "n_tab": 8, "L_tab": 1344, "n_element_tab": 279, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 760, "n_element_tab_1": 158, "formula_len_all": 250, "formula_len_all_1": 250, "len_all": 140201, "len_all_1": 54672, "len_abs": 1488, "len_title": 106, "len_sents": 36859, "len_sents_1": 24759, "n_sents": 262, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1501, "title": "What Matters in Transformers? Not All Attention is Needed", "abs": "While scaling Transformer-based large language models (LLMs) has demonstrated\npromising performance across various tasks, it also introduces redundant archi-\ntectures, posing efficiency challenges for real-world deployment. Despite some\nrecognition of redundancy in LLMs, the variability of redundancy across different\narchitectures in transformers, such as MLP and Attention layers, is under-explored.\nIn this work, we investigate redundancy across different modules within Trans-\nformers, including Blocks, MLP, and Attention layers, using a similarity-based\nmetric. Surprisingly, despite the critical role of attention layers in distinguishing\ntransformers from other architectures, we found that a large portion of these layers\nexhibit excessively high similarity and can be pruned without degrading perfor-\nmance. For instance, Llama-2-70B achieved a 48.4% speedup with only a 2.4%\nperformance drop by pruning half of the attention layers. Furthermore, by tracing\nmodel checkpoints throughout the training process, we observed that attention\nlayer redundancy is inherent and consistent across training stages. Additionally,\nwe further propose a method that jointly drops Attention and MLP layers, allowing\nus to more aggressively drop additional layers. For instance, when dropping 31\nlayers (Attention + MLP), Llama-2-13B still retains 90% of the performance on the\nMMLU task. Our work provides valuable insights for future network architecture\ndesign. The code will be released upon acceptance.", "keywords": [ "Transformer", "Attention", "Model Compression", "Efficiency" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "YLIsIzC74j", "primary_area": "applications to robotics, autonomy, planning", "n_author": 9, "author_info_list": [ { "name": "Feng Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Shixiong Kai", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Siyuan Xu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zhentao Tang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zijie Geng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Ziyan Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 29, "n_formula_1": 18, "n_ref_uni": 37, "n_ref": 103, "n_ref_all": 140, "n_fig": 12, "n_tab": 12, "L_tab": 4292, "n_element_tab": 680, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1626, "formula_len_all_1": 919, "len_all": 182763, "len_all_1": 69102, "len_abs": 1618, "len_title": 117, "len_sents": 54837, "len_sents_1": 32426, "n_sents": 460, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 69, "L_abs": 1663, "title": "LaMPlace: Learning to Optimize Cross-Stage Metrics in Macro Placement", "abs": "Machine learning techniques have shown great potential in enhancing macro placement, a critical stage in modern chip design.\nHowever, existing methods primarily focus on *online* optimization of *intermediate surrogate metrics* that are available at the current placement stage, rather than directly targeting the *cross-stage metrics*---such as the timing performance---that measure the final chip quality.\nThis is mainly because of the high computational costs associated with performing post-placement stages for evaluating such metrics, making the *online* optimization impractical.\nConsequently, these optimizations struggle to align with actual performance improvements and can even lead to severe manufacturing issues.\nTo bridge this gap, we propose **LaMPlace**, which **L**earns **a** **M**ask for optimizing cross-stage metrics in macro placement.\nSpecifically, LaMPlace trains a predictor on *offline* data to estimate these *cross-stage metrics* and then leverages the predictor to quickly generate a mask, i.e., a pixel-level feature map that quantifies the impact of placing a macro in each chip grid location on the design metrics.\nThis mask essentially acts as a fast evaluator, enabling placement decisions based on *cross-stage metrics* rather than *intermediate surrogate metrics*.\nExperiments on commonly used benchmarks demonstrate that LaMPlace significantly improves the chip quality across several key design metrics, achieving an average improvement of 9.6\\%, notably 43.0\\% and 30.4\\% in terms of WNS and TNS, respectively, which are two crucial cross-stage metrics that reflect the final chip quality in terms of the timing performance.", "keywords": [ "Macro placement", "Chip design", "EDA" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "YKvBiRWdQC", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Andreas Bulling", "gender": "Male", "institution": "University of Stuttgart", "country": "DE", "position": "Full Professor" }, { "name": "Anna Penzkofer", "gender": "Female", "institution": "Universit\u00e4t Stuttgart", "country": "DE", "position": "PhD student" }, { "name": "Constantin Ruhdorfer", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "DE", "position": "PhD student" }, { "name": "Matteo Bortoletto", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "DE", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 55, "n_ref": 163, "n_ref_all": 200, "n_fig": 13, "n_tab": 5, "L_tab": 1271, "n_element_tab": 224, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 743, "n_element_tab_1": 107, "formula_len_all": 472, "formula_len_all_1": 264, "len_all": 211221, "len_all_1": 59601, "len_abs": 1302, "len_title": 87, "len_sents": 46635, "len_sents_1": 30195, "n_sents": 320, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1299, "title": "The Overcooked Generalisation Challenge", "abs": "We introduce the Overcooked Generalisation Challenge (OGC) \u2013 the first bench-mark to study reinforcement learning agents\u2019 zero-shot cooperation abilities when faced with novel partners and levels in the Overcooked-AI environment.\nThis perspective starkly contrasts a large body of previous work that has evaluated cooperating agents only on the same level or with the same partner, thus failing to capture generalisation abilities essential for real-world human-AI cooperation.\nOur challenge interfaces with state-of-the-art dual curriculum design (DCD) methods to generate auto-curricula for training general agents in Overcooked.\nIt is the first cooperative multi-agent environment specially designed for DCD methods and, consequently, the first evaluated with state-of-the-art methods. \nIt is fully GPU-accelerated, built on the DCD benchmark suite minimax, and freely available under an open-source license: http://anonymised.edu. \nWe show that state-of-the-art DCD algorithms fail to produce useful policies on this novel challenge, even if combined with recent network architectures specifically designed for scalability and generalisability. \nAs such, the OGC pushes the boundaries of real-world human-AI cooperation by enabling research on the impact of generalisation on cooperating agents.", "keywords": [ "Human-AI Cooperation", "Unsupervised Environment Design", "Multi-Agent Reinforcement Learning" ], "rating_list": [ 6, 6, 3, 8 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "YKtbklD5MV", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Erik Sandstr\u00f6m", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Federico Tombari", "gender": "Male", "institution": "Google", "country": "CH", "position": "Research Director" }, { "name": "Keisuke Tateno", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Luc Van Gool", "gender": "unknown", "institution": "INSAIT - Sofia Un.", "country": "BG", "position": "Full Professor" }, { "name": "Martin R. Oswald", "gender": "unknown", "institution": "University of Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Michael Niemeyer", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Michael Oechsle", "gender": "unknown", "institution": "Google", "country": "CH", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 32, "n_ref": 95, "n_ref_all": 105, "n_fig": 4, "n_tab": 6, "L_tab": 915, "n_element_tab": 162, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 2355, "n_element_tab_1": 321, "formula_len_all": 1309, "formula_len_all_1": 1295, "len_all": 142794, "len_all_1": 64757, "len_abs": 1628, "len_title": 110, "len_sents": 27712, "len_sents_1": 26820, "n_sents": 227, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1297, "title": "Splat-SLAM: Globally Optimized RGB-only SLAM with 3D Gaussians", "abs": "3D Gaussians Splatting has emerged as a powerful representation of geometry and appearance for RGB-only dense Simultaneous Localization and Mapping (SLAM), as it provides a compact dense map representation while enabling efficient and high-quality map rendering. However, existing methods show significantly worse reconstruction quality than competing methods using other 3D representations, e.g. neural points clouds, since they either do not employ global map and pose optimization or make use of monocular depth. In response, we propose the first RGB-only SLAM system with a dense 3D Gaussian map representation that utilizes all benefits of globally optimized tracking by adapting dynamically to keyframe pose and depth updates by actively deforming the 3D Gaussian map. Moreover, we find that refining the depth updates in inaccurate areas with a monocular depth estimator further improves the accuracy of the 3D reconstruction. Our experiments on the Replica, TUM-RGBD, and ScanNet datasets indicate the effectiveness of globally optimized 3D Gaussians, as the approach achieves superior or on par performance with existing RGB-only SLAM methods methods in tracking, mapping and rendering accuracy while yielding small map sizes and fast runtimes. The source code will be publicly available.", "keywords": [ "slam", "gaussian splatting" ], "rating_list": [ 6, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "YKmRcayt8Z", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Hao Ban", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Kaiyi Ji", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Peiyao Xiao", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Shaofeng Zou", "gender": "unknown", "institution": "Arizona State University", "country": "", "position": "Associate Professor" }, { "name": "Yudan Wang", "gender": "Female", "institution": "Arizona State University", "country": "US", "position": "PhD student" } ], "n_formula": 126, "n_formula_1": 30, "n_ref_uni": 23, "n_ref": 41, "n_ref_all": 62, "n_fig": 0, "n_tab": 4, "L_tab": 2089, "n_element_tab": 137, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 23806, "formula_len_all_1": 3582, "len_all": 168769, "len_all_1": 66630, "len_abs": 1395, "len_title": 124, "len_sents": 40976, "len_sents_1": 26892, "n_sents": 422, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1514, "title": "Finite-Time Analysis for Conflict-Avoidant Multi-Task Reinforcement Learning", "abs": "Multi-task reinforcement learning (MTRL) has shown great promise in many real-world applications. Existing MTRL algorithms often aim to learn a policy that optimizes individual objective functions simultaneously with a given prior preference (or weights) on different tasks. However, these methods often suffer from the issue of gradient conflict such that the tasks with larger gradients dominate the update direction, resulting in a performance degeneration on other tasks. In this paper, we develop a novel dynamic weighting multi-task actor-critic algorithm (MTAC) under two options of sub-procedures named as CA and FC in task weight updates. MTAC-CA aims to find a conflict-avoidant (CA) update direction that maximizes the minimum value improvement among tasks, and MTAC-FC targets at a much faster convergence rate. We provide a comprehensive finite-time convergence analysis for both algorithms. We show that MTAC-CA can find a $\\epsilon+\\epsilon_{\\text{app}}$-accurate Pareto stationary policy using $\\mathcal{O}({\\epsilon^{-5}})$ samples, while ensuring a small $\\epsilon+\\sqrt{\\epsilon_{\\text{app}}}$-level CA distance (defined as the distance to the CA direction), where $\\epsilon_{\\text{app}}$ is the function approximation error. The analysis also shows that MTAC-FC improves the sample complexity to $\\mathcal{O}(\\epsilon^{-3})$, but with a constant-level CA distance. Our experiments on MT10 demonstrate the improved performance of our algorithms over existing MTRL methods with fixed preference.", "keywords": [ "Multi-task reinforcement learning", "Conflict-avoidant methods", "Sample complexity analysis" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 2, 3 ] }, { "paper_id": "YKfJFTiRz8", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Hua Xie", "gender": "Male", "institution": "Children's National Hospital", "country": "US", "position": "Assistant Professor" }, { "name": "Nancy Carlisle", "gender": "unknown", "institution": "Lehigh University", "country": "US", "position": "Associate Professor" }, { "name": "Xinxu Wei", "gender": "Male", "institution": "Lehigh University", "country": "US", "position": "PhD student" }, { "name": "Yong Jiao", "gender": "unknown", "institution": "Lehigh University", "country": "US", "position": "Postdoc" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Lehigh University", "country": "", "position": "Assistant Professor" }, { "name": "kanhao zhao", "gender": "Male", "institution": "Lehigh University", "country": "US", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 22, "n_ref_uni": 38, "n_ref": 62, "n_ref_all": 89, "n_fig": 8, "n_tab": 10, "L_tab": 4743, "n_element_tab": 318, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2480, "n_element_tab_1": 72, "formula_len_all": 3132, "formula_len_all_1": 2033, "len_all": 175712, "len_all_1": 73132, "len_abs": 1673, "len_title": 129, "len_sents": 52365, "len_sents_1": 29138, "n_sents": 397, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1679, "title": "Pre-Training Graph Contrastive Masked Autoencoders are Strong Distillers for EEG", "abs": "Effectively utilizing extensive unlabeled high-density EEG data to improve performance in scenarios with limited labeled low-density EEG data presents a significant challenge. In this paper, we address this by framing it as a graph transfer learning and knowledge distillation problem. We propose a Unified Pre-trained Graph Contrastive Masked Autoencoder Distiller, named EEG-DisGCMAE, to bridge the gap between unlabeled/labeled and high/low-density EEG data. To fully leverage the abundant unlabeled EEG data, we introduce a novel unified graph self-supervised pre-training paradigm, which seamlessly integrates Graph Contrastive Pre-training and Graph Masked Autoencoder Pre-training. This approach synergistically combines contrastive and generative pre-training techniques by reconstructing contrastive samples and contrasting the reconstructions. For knowledge distillation from high-density to low-density EEG data, we propose a Graph Topology Distillation loss function, allowing a lightweight student model trained on low-density data to learn from a teacher model trained on high-density data, effectively handling missing electrodes through contrastive distillation. To integrate transfer learning and distillation, we jointly pre-train the teacher and student models by contrasting their queries and keys during pre-training, enabling robust distillers for downstream tasks. We demonstrate the effectiveness of our method on four classification tasks across two clinical EEG datasets with abundant unlabeled data and limited labeled data. The experimental results show that our approach significantly outperforms contemporary methods in both efficiency and accuracy.", "keywords": [ "EEG Signals Recognition", "Graph Learning", "Self-supervised Graph Pre-training", "Graph Knowledge Distillation" ], "rating_list": [ 5, 3, 8, 3, 5 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 4, 2, 2 ], "confidence_list": [ 4, 5, 4, 4, 3 ] }, { "paper_id": "YKW98Icu1X", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Allen Nie", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Anirudhan Badrinath", "gender": "unknown", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "MS student" }, { "name": "Carissa Yip", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Christopher J Piech", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Emma Brunskill", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "" }, { "name": "Nicholas Tomlin", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Rose E Wang", "gender": "Female", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Timothy Dai", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 21, "n_ref": 24, "n_ref_all": 45, "n_fig": 6, "n_tab": 2, "L_tab": 327, "n_element_tab": 32, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 769, "formula_len_all_1": 769, "len_all": 96907, "len_all_1": 66083, "len_abs": 1667, "len_title": 149, "len_sents": 35420, "len_sents_1": 33004, "n_sents": 250, "n_sents_1": 229, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1675, "title": "Discovering High-Quality Chess Puzzles Through One Billion Plays with Offline Reinforcement Learning", "abs": "Learning and skill mastery requires extensive and deliberate practice. In many learning settings, producing high-quality pedagogical materials can require a high level of domain expertise and be very time-consuming. Pedagogical materials often need to train students to engage in different thinking patterns. In some domains, such as chess, puzzles are used to help students practice their skills in calculating the next moves and recognizing known patterns on a board. Giving students a practice set of puzzles to help them learn different modes of thinking is challenging because the teacher needs to carefully balance between different motifs and how many look-ahead steps a student needs to perform. Popular online platforms like Chess.com and Lichess offer players millions of puzzles. Unlike chess tactics puzzles procured by human experts, where chess beginners can learn valuable insights, these puzzles are automatically generated and often regarded as having low pedagogical values. These platforms also rely on a heuristic to recommend puzzles to users for practice.\nUsing the user history data over an entire year, a total of 1.6 billion puzzle-solving histories, we learn the pedagogical value of a puzzle and how to automatically choose a set of puzzles to better support chess learners in a completely unstructured way using insights from offline reinforcement learning. We validate the quality of the puzzles discovered by our model by collecting annotation ratings from titled chess players. The success of our pipeline shows promise for a future where we can understand the pedagogical values of practice items in other domains like math or coding problems.", "keywords": [ "Offline RL", "education", "chess", "puzzle recommendation" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "YK9G4Htdew", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Maxime Burchi", "gender": "unknown", "institution": "Bayerische Julius-Maximilians-Universit\u00e4t W\u00fcrzburg", "country": "DE", "position": "PhD student" }, { "name": "Radu Timofte", "gender": "Male", "institution": "Bayerische Julius-Maximilians-Universit\u00e4t W\u00fcrzburg", "country": "DE", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 44, "n_ref": 78, "n_ref_all": 100, "n_fig": 11, "n_tab": 3, "L_tab": 4233, "n_element_tab": 385, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 3571, "n_element_tab_1": 341, "formula_len_all": 1206, "formula_len_all_1": 1204, "len_all": 132457, "len_all_1": 67731, "len_abs": 1385, "len_title": 122, "len_sents": 34239, "len_sents_1": 29926, "n_sents": 218, "n_sents_1": 155, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 74, "L_abs": 1449, "title": "Learning Transformer-based World Models with Contrastive Predictive Coding", "abs": "The DreamerV3 algorithm recently obtained remarkable performance across diverse environment domains by learning an accurate world model based on Recurrent Neural Networks (RNNs). Following the success of model-based reinforcement learning algorithms and the rapid adoption of the Transformer architecture for its superior training efficiency and favorable scaling properties, recent works such as STORM have proposed replacing RNN-based world models with Transformer-based world models using masked self-attention. However, despite the improved training efficiency of these methods, their impact on performance remains limited compared to the Dreamer algorithm, struggling to learn competitive Transformer-based world models. In this work, we show that the next state prediction objective adopted in previous approaches is insufficient to fully exploit the representation capabilities of Transformers. We propose to extend world model predictions to longer time horizons by introducing TWISTER (Transformer-based World model wIth contraSTivE Representations), a world model using action-conditioned Contrastive Predictive Coding to learn high-level temporal feature representations and improve the agent performance. TWISTER achieves a human-normalized mean score of 162% on the Atari 100k benchmark, setting a new record among state-of-the-art methods that do not employ look-ahead search. We release our code at https://github.com/burchim/TWISTER.", "keywords": [ "model-based reinforcement learning", "transformer network", "contrastive predictive coding" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "YK8eO7BEkJ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Lei Huang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Associate Professor" }, { "name": "Peilin Feng", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Yuanshuai Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Yunhao Ni", "gender": "unknown", "institution": "Beihang University", "country": "CN", "position": "MS student" }, { "name": "Wenjun Wu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Li Zekun", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" } ], "n_formula": 16, "n_formula_1": 12, "n_ref_uni": 39, "n_ref": 90, "n_ref_all": 103, "n_fig": 4, "n_tab": 6, "L_tab": 1891, "n_element_tab": 243, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1325, "n_element_tab_1": 161, "formula_len_all": 991, "formula_len_all_1": 517, "len_all": 141384, "len_all_1": 58833, "len_abs": 1554, "len_title": 92, "len_sents": 37185, "len_sents_1": 26018, "n_sents": 284, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1563, "title": "An Empirical Study on Normalization in Mamba", "abs": "Normalization layers are crucial for improving the training efficiency and stability of deep neural network architectures. The recently proposed Mamba network has demonstrated significant potential in competing with Transformers. However, as with many deep architectures, the training stability of Mamba remains a significant challenge, and normalization techniques are key to addressing this issue. In this paper, we systematically investigate the effects of normalization type, position and combinations on the Mamba Block. On the one hand, we conducted extensive experiments to evaluate the impact of applying various normalization layers before or after the SSM module(the core module of Mamba Block). On the other hand, we performed thorough experiments to assess the effects of combining diverse normalization techniques before and after the SSM module. Our analysis encompasses both long sequence modeling and image classification tasks. The results show that applying normalization layers after the SSM module (if used only once) and combining different normalization layers before and after the SSM module can enhance training stability and improve Mamba performance. Furthermore, we provide practical recommendations for selecting appropriate normalization techniques in designing Mamba architectures and validated them on other datasets. We hope that our insights will help mitigate training instabilities in deep learning and foster the development of more robust architectures. All codes and models used in this study will be open-sourced on GitHub.", "keywords": [ "Mamba", "long-sequence modeling", "normalization", "performance", "stability" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "YJwnlplKQ7", "primary_area": "generative models", "n_author": 15, "author_info_list": [ { "name": "Ding Liu", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Haozhe Liu", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Juan-Manuel Perez-Rua", "gender": "Male", "institution": "Facebook", "country": "GB", "position": "Research Scientist" }, { "name": "Juan Camilo Perez", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "PhD student" }, { "name": "Jui-Chieh Wu", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "J\u00fcrgen Schmidhuber", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Director, AI Initiative" }, { "name": "Kumara Kahatapitiya", "gender": "Male", "institution": "Meta", "country": "US", "position": "Intern" }, { "name": "Menglin Jia", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Mengmeng Xu", "gender": "Male", "institution": "Meta ", "country": "GB", "position": "Researcher" }, { "name": "Sen He", "gender": "Male", "institution": "Meta AI", "country": "GB", "position": "Researcher" }, { "name": "Shikun Liu", "gender": "Male", "institution": "Facebook", "country": "GB", "position": "Researcher" }, { "name": "Tao Xiang", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Full Professor" }, { "name": "Xiao Han", "gender": "Male", "institution": "Meta AI", "country": "GB", "position": "Research Scientist" }, { "name": "Yanping Xie", "gender": "unknown", "institution": "Meta", "country": "", "position": "Software Engineer" }, { "name": "Zijian Zhou", "gender": "Male", "institution": "King's College London, University of London", "country": "GB", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 52, "n_ref": 165, "n_ref_all": 184, "n_fig": 8, "n_tab": 5, "L_tab": 1851, "n_element_tab": 250, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2786, "n_element_tab_1": 371, "formula_len_all": 5775, "formula_len_all_1": 13601, "len_all": 228961, "len_all_1": 98112, "len_abs": 1241, "len_title": 71, "len_sents": 53221, "len_sents_1": 36756, "n_sents": 362, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1243, "title": "MarDini: Masked Autoregressive Diffusion for Video Generation at Scale", "abs": "We introduce MarDini, a new family of video diffusion models that integrate the advantages of masked auto-regression (MAR) into a unified diffusion model (DM) framework. Here, MAR handles temporal planning, while DM focuses on spatial generation in an asymmetric network design: i) a MAR-based planning model containing most of the parameters generates planning signals for each masked frame using low-resolution input; ii) a lightweight generation model uses these signals to produce high-resolution frames via diffusion de-noising. MarDini\u2019s MAR enables video generation conditioned on any number of masked frames at any frame positions: a single model can handle video interpolation (e.g., masking middle frames), image-to-video generation (e.g., masking from the second frame onward), and video expansion (e.g., masking half the frames). The efficient design allocates most of the computational resources to the low-resolution planning model, making computationally expensive but important spatio-temporal attention feasible at scale. MarDini sets a new state-of-the-art for video interpolation; meanwhile, within few inference steps, it efficiently generates videos on par with those of much more expensive advanced image-to-video models.", "keywords": [ "Video Generation", "Diffusion Model", "Masked Auto-regressive Model" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "YHUOaIbFby", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Jiaqi Cheng", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Kaining Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Yinxing Xue", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "", "position": "Assistant Professor" }, { "name": "Zhaofeng Su", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 18, "n_ref": 26, "n_ref_all": 33, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1125, "n_element_tab_1": 14, "formula_len_all": 238, "formula_len_all_1": 238, "len_all": 92173, "len_all_1": 59793, "len_abs": 1685, "len_title": 114, "len_sents": 31437, "len_sents_1": 30299, "n_sents": 199, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1694, "title": "Problem-dependent Quantum Circuit Design Based on Entropy Matching", "abs": "Variational quantum machine learning (QML) have shown great promise for harnessing quantum advantage in machine learning tasks. However, architecture design of quantum circuits employed in these QML algorithms has been poorly explored for practical problems. Specifically, quantum circuits should have sufficient expressibility for modeling complex functions considering the inherent structures of real-world data. Naively increasing the circuit depth could enhance the expressibility of quantum circuits, which also induce the barren plateau problem as a by-product. In this work, we develop an architecture design framework to solve this problem. We use a simple yet effective metric of quantum entanglement, i.e. the linear entropy, to guide the circuit design from the perspective of the input data. First, we quantify the entanglement of input data by calculating the 1-qubit linear entropy of their amplitude encoding states. Then we implement an entropy matching approach to identify the optimal circuit depth that lead to the linear entropy being close the entropy of input data. The effectiveness of circuit designs based on entropy is verified by extensive experimental results. Specifically, we demonstrate that real-world datasets like MNIST images has limited quantum entanglement. Therefore, circuits designed with entropy matching exhibit relatively small depths being free from the barren plateau issue while maintaining benign performances in binary classification tasks. This work not only advances the efficiency of quantum circuit design but also sets the stage for further refinement of QML performance, with broad implications for practical quantum computing applications.", "keywords": [ "Quantum computing", "Expressibility", "Ansatz design", "Linear entropy." ], "rating_list": [ 3, 5, 1, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 1, 2, 1, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "YHDY5uXOSN", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Meiqi Wang", "gender": "Female", "institution": "Nanjing University", "country": "", "position": "PhD student" }, { "name": "Siyu Liao", "gender": "Male", "institution": "Rutgers University", "country": "CN", "position": "Unemployed" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 27, "n_ref": 47, "n_ref_all": 51, "n_fig": 0, "n_tab": 5, "L_tab": 2458, "n_element_tab": 478, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 2458, "n_element_tab_1": 478, "formula_len_all": 1342, "formula_len_all_1": 1342, "len_all": 88552, "len_all_1": 79487, "len_abs": 1130, "len_title": 127, "len_sents": 26737, "len_sents_1": 26737, "n_sents": 207, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1132, "title": "VARIATIONAL DIFFUSION CHANNEL DECODING: A ULTRA-LOW-COST NEURAL CHANNEL DECODER", "abs": "Neural channel decoder, as a data-driven channel decoding strategy, has shown very promising improvement on error-correcting capability over the classical methods. However, the success of those deep learning-based decoder comes at the cost of drastically increased model storage and computational complexity, hindering their practical adoptions in real-world time-sensitive resource-sensitive communication and storage systems. To address this challenge, we propose an efficient variational diffusion model-based channel decoder, which effectively integrates the domain-specific belief propagation process to the modern diffusion model. By reaping the low-cost benefits of belief propagation and strong learning capability of diffusion model, our proposed neural decoder simultaneously achieves very low cost and high error-correcting performance. Experimental results show that, compared with the state-of-the-art neural channel decoders, our\nmodel provides a feasible solution for practical deployment via achieving the best decoding performance with order-of-magnitude (1000\u00d7 and up) savings in computational cost and model size.", "keywords": [ "diffusion; channel coding" ], "rating_list": [ 5, 3, 1 ], "soundness_list": [ 2, 2, 1 ], "presentation_list": [ 1, 1, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "YH4M1Tbxfz", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Fudong Lin", "gender": "Male", "institution": "University of Delaware", "country": "US", "position": "PhD student" }, { "name": "Jason Jiang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Undergrad student" }, { "name": "Liyun Wang", "gender": "Female", "institution": "University of Delaware", "country": "", "position": "Full Professor" }, { "name": "Shubo Wang", "gender": "Male", "institution": "University of Delaware", "country": "US", "position": "PhD student" }, { "name": "Tiankuo Chu", "gender": "Male", "institution": "University of Delaware", "country": "US", "position": "PhD student" }, { "name": "Wiley Jia-Wei Gong", "gender": "unknown", "institution": "University of Delaware", "country": "US", "position": "Intern" }, { "name": "Xu Yuan", "gender": "Male", "institution": "University of Delaware", "country": "US", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 19, "n_ref": 27, "n_ref_all": 42, "n_fig": 6, "n_tab": 6, "L_tab": 2147, "n_element_tab": 330, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 2291, "n_element_tab_1": 346, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 115150, "len_all_1": 64492, "len_abs": 1907, "len_title": 161, "len_sents": 34674, "len_sents_1": 30225, "n_sents": 204, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 112, "L_abs": 1906, "title": "BoneMet: An Open Large-Scale Multi-Modal Murine Dataset for Breast Tumor Bone Metastasis Diagnosis and Prognosis", "abs": "Breast tumor bone metastasis (BTBM) affects women\u2019s health globally, calling for the development of effective solutions for its diagnosis and prognosis. While the deep learning has exhibited impressive capacities across various healthcare domains, their applicability to managing BTBM diseases is consistently hindered by the lack of an open, large-scale, deep learning-ready dataset. As such, we introduce the Bone Metastasis (BoneMet) dataset, the first large-scale, publicly available, high-resolution medical resource specifically targeting BTBM for disease diagnosis, prognosis, and treatment management. It offers over 50 terabytes of multi-modal medical data, including 2D X-ray images, 3D CT scans, and detailed biological data (e.g., medical records and bone quantitative analysis), collected from thousands of mice spanning from 2019 to 2024. Our BoneMet dataset is well-organized into six components, i.e., Rotation-X-Ray, Recon-CT, Seg-CT, Regist-CT, RoI-CT, and MiceMediRec. Thanks to its extensive data samples and our tireless efforts of image processing, organization and data labeling, BoneMet can be readily adopted to build versatile, large-scale AI models for managing BTBM diseases, which have been validated by our extensive experiments via various deep learning solutions. To facilitate its easy access and wide dissemination, we have created the BoneMet package, providing three APIs that enable researchers to (i)flexibly process and download the BoneMet data filtered by specific time frames;and (ii) develop and train large-scale AI models for precise BTBM diagnosis and prognosis. The BoneMet dataset is officially available on Hugging Face Datasets at https://huggingface.co/datasets/BoneMet/BoneMet. The BoneMet package is available on the Python Package Index (PyPI) at https://pypi.org/project/BoneMet. Code and tutorials are available at https://github.com/BoneMet/BoneMet.", "keywords": [ "Medical Dataset", "Breast Tumor Bone Metastasis", "Diagnosis", "Prognosis", "Sparse CT reconstruction", "CT", "X-ray", "AI for Science" ], "rating_list": [ 5, 6, 5, 8 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 4 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "YGoFl5KKFc", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Linyi Yang", "gender": "unknown", "institution": "University College London", "country": "GB", "position": "Researcher" }, { "name": "Minjun Zhu", "gender": "Female", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Ningyu Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Yifan Wei", "gender": "Male", "institution": "\tBeihang University", "country": "CN", "position": "PhD student" }, { "name": "Yixuan Weng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Yue Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 7, "n_ref_uni": 65, "n_ref": 128, "n_ref_all": 150, "n_fig": 8, "n_tab": 6, "L_tab": 2127, "n_element_tab": 159, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 873, "n_element_tab_1": 97, "formula_len_all": 2950, "formula_len_all_1": 312, "len_all": 301952, "len_all_1": 59467, "len_abs": 1889, "len_title": 148, "len_sents": 60101, "len_sents_1": 27704, "n_sents": 394, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 38, "L_abs": 1653, "title": "Locking Down the Finetuned LLMs Safety", "abs": "Fine-tuning large language models (LLMs) on additional datasets is often necessary to optimize them for specific downstream tasks. However, existing safety alignment measures, which restrict harmful behavior during inference, are insufficient to mitigate safety risks during fine-tuning. Alarmingly, fine-tuning with just 10 toxic sentences can make models comply with harmful instructions. We introduce SafetyLock, a novel alignment intervention method that maintains robust safety post-fine-tuning through efficient and transferable mechanisms. SafetyLock leverages our discovery that fine-tuned models retain similar safety-related activation representations to their base models. This insight enables us to extract what we term the Meta-SafetyLock, a set of safety bias directions representing key activation patterns associated with safe responses in the original model. We can then apply these directions universally to fine-tuned models to enhance their safety. By searching for activation directions across multiple token dimensions, SafetyLock achieves enhanced robustness and transferability. SafetyLock re-aligns fine-tuned models in under 0.01 seconds without additional computational cost. Our experiments demonstrate that SafetyLock can reduce the harmful instruction response rate from 60\\% to below 1\\% in toxic fine-tuned models. It surpasses traditional methods in both performance and efficiency, offering a scalable, non-invasive solution for ensuring the safety of customized LLMs. Our analysis across various fine-tuning scenarios confirms SafetyLock's robustness, advocating its integration into safety protocols for aligned LLMs.", "keywords": [ "large language models", "Safety", "Finetuned", "fast" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "YGhV8wQv3C", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Junsu Kim", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Visiting Ph.D Student" }, { "name": "Seohong Park", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 14, "n_ref_uni": 69, "n_ref": 215, "n_ref_all": 243, "n_fig": 15, "n_tab": 2, "L_tab": 125, "n_element_tab": 4, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2763, "formula_len_all_1": 10939, "len_all": 192575, "len_all_1": 79239, "len_abs": 1472, "len_title": 93, "len_sents": 48969, "len_sents_1": 34236, "n_sents": 299, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1284, "title": "Unsupervised-to-Online Reinforcement Learning", "abs": "Offline-to-online reinforcement learning (RL), a framework that trains a policy with offline RL and then further fine-tunes it with online RL,\nhas been considered a promising recipe for data-driven decision-making. While sensible, this framework has drawbacks: it requires domain-specific offline RL pre-training for each task, and is often brittle in practice. In this work, we propose unsupervised-to-online RL (U2O RL),\nwhich replaces domain-specific supervised offline RL with unsupervised offline RL,\nas a better alternative to offline-to-online RL.\nU2O RL not only enables reusing a single pre-trained model for multiple downstream tasks,\nbut also learns better representations, which often result in even better performance and stability\nthan supervised offline-to-online RL.\nTo instantiate U2O RL in practice, we propose a general recipe for U2O RL\nto bridge task-agnostic unsupervised offline skill-based policy pre-training and supervised online fine-tuning.\nThroughout our experiments in nine state-based and pixel-based environments,\nwe empirically demonstrate that U2O RL achieves strong performance\nthat matches or even outperforms previous offline-to-online RL approaches,\nwhile being able to reuse a single pre-trained model for a number of different downstream tasks.", "keywords": [ "Reinforcement Learning", "Offline-to-Online Reinforcement Learning", "Offline Unsupervised Reinforcement Learning" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 4, 1, 3, 2 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "YGflij9S6x", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Hossein Rezaei", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Intern" }, { "name": "Mohammad Khalooei", "gender": "Male", "institution": "Amirkabir University of Technology", "country": "", "position": "PhD candidate" }, { "name": "Mohammad Sabokrou", "gender": "Male", "institution": "Okinawa Institute of Science and Technology (OIST)", "country": "JP", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 16, "n_ref_uni": 31, "n_ref": 62, "n_ref_all": 100, "n_fig": 5, "n_tab": 6, "L_tab": 2192, "n_element_tab": 77, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1917, "n_element_tab_1": 54, "formula_len_all": 1336, "formula_len_all_1": 1310, "len_all": 127037, "len_all_1": 65059, "len_abs": 1311, "len_title": 141, "len_sents": 40655, "len_sents_1": 27917, "n_sents": 316, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1319, "title": "Continual Learning: Less Forgetting, More OOD Generalization via Adaptive Contrastive Replay", "abs": "Machine learning models often suffer from catastrophic forgetting of previously learned knowledge when learning new classes. Various methods have been proposed to mitigate this issue. However, rehearsal-based learning, which retains samples from previous classes, typically achieves good performance but tends to memorize specific instances, struggling with Out-of-Distribution (OOD) generalization. This often leads to high forgetting rates and poor generalization. Surprisingly, the OOD generalization capabilities of these methods have been largely unexplored. In this paper, we highlight this issue and propose a simple yet effective strategy inspired by contrastive learning and data-centric principles to address it.\nWe introduce Adaptive Contrastive Replay (ACR), a method that employs dual optimization to simultaneously train both the encoder and the classifier. ACR adaptively populates the replay buffer with misclassified samples while ensuring a balanced representation of classes and tasks. By refining the decision boundary in this way, ACR achieves a balance between stability and plasticity. Our method significantly outperforms previous approaches in terms of OOD generalization, achieving an improvement of 13.41\\% on Split CIFAR-100, 9.91\\% on Split Mini-ImageNet, and 5.98\\% on Split Tiny-ImageNet.", "keywords": [ "Continual learning", "OOD generalization", "contrastive learning" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "YGWxpOI6Y0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Hanoona Abdul Rasheed", "gender": "Female", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "PhD student" }, { "name": "Muhammad Maaz", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "PhD student" }, { "name": "Salman Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Associate Professor" }, { "name": "Fahad Shahbaz Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 30, "n_ref": 100, "n_ref_all": 127, "n_fig": 10, "n_tab": 5, "L_tab": 2298, "n_element_tab": 166, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1840, "n_element_tab_1": 57, "formula_len_all": 82, "formula_len_all_1": 201, "len_all": 147315, "len_all_1": 57730, "len_abs": 1889, "len_title": 130, "len_sents": 42929, "len_sents_1": 28168, "n_sents": 280, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1890, "title": "VideoGPT+: Integrating Image and Video Encoders for Enhanced Video Understanding", "abs": "Building on the advances of language models, Large Multimodal Models (LMMs) have contributed significant improvements in video understanding. While the current video LMMs utilize advanced Large Language Models (LLMs), they rely on either image or video encoders to process visual inputs, each of which has its own limitations. Image encoders excel at capturing rich spatial details from frame sequences but lack explicit temporal context, which can be important in videos with intricate action sequences. On the other hand, video encoders provide temporal context but are often limited by computational constraints that lead to processing only sparse frames at lower resolutions, resulting in reduced contextual and spatial understanding. To this end, we introduce our model, which combines the complementary benefits of the image encoder (for detailed spatial understanding) and the video encoder (for global temporal context modeling). The model processes videos by dividing them into smaller segments and applies an adaptive pooling strategy on features extracted by both image and video encoders. Our architecture showcases improved performance across multiple video benchmarks, including VCGBench, MVBench and Zero-shot question-answering. Further, we develop 112K video-instruction set using a novel semi-automatic annotation pipeline which further improves the model performance. Additionally, to comprehensively evaluate video LMMs, we present our bench, covering 18 broad video categories such as lifestyle, sports, science, gaming, and surveillance videos. This benchmark with 4,354 question-answer pairs evaluates the generalization of existing LMMs on dense video captioning, spatial and temporal understanding, and complex reasoning, ensuring comprehensive assessment across diverse video types and dynamics. Our code, dataset, and pre-trained models will be publicly released.", "keywords": [ "video-conversation-model", "large multi-modal model", "multi-modal", "video-conversation", "image-and-video", "phi-3-min", "vision-language", "video-chatbot" ], "rating_list": [ 3, 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3, 1 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 5, 5, 5, 4, 4 ] }, { "paper_id": "YGDWW6rzYX", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 6, "author_info_list": [ { "name": "Bulent Yener", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "Full Professor" }, { "name": "Colton Ritchie", "gender": "Not Specified", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "Undergrad student" }, { "name": "Haidar Khan", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Hisham Abdullah Alyahya", "gender": "Male", "institution": "Saudi Data and AI Authority", "country": "SA", "position": "Researcher" }, { "name": "Yazeed Alnumay", "gender": "Male", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "SA", "position": "Researcher" }, { "name": "M SAIFUL BARI", "gender": "Male", "institution": "National Centre of Artificial Intelligence, Saudi Data and AI Authority", "country": "SA", "position": "Principal Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 60, "n_ref": 96, "n_ref_all": 110, "n_fig": 4, "n_tab": 2, "L_tab": 537, "n_element_tab": 48, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 883, "n_element_tab_1": 64, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 265104, "len_all_1": 62697, "len_abs": 1595, "len_title": 64, "len_sents": 28741, "len_sents_1": 28105, "n_sents": 192, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1601, "title": "ZeroSumEval: Scaling LLM Evaluation with Inter-Model Competition", "abs": "Evaluating the capabilities of Foundation Models has traditionally relied on static benchmark datasets, human assessments, or model-based evaluations \u2014 methods that often suffer from overfitting, high costs, and biases. We introduce ZeroSumEval, a novel competition-based evaluation protocol that leverages zero-sum games to assess LLMs with dynamic benchmarks that resist saturation. ZeroSumEval encompasses a diverse suite of games, including security challenges (Capture the Flag), classic board games (chess), and knowledge tests (MathQuiz). These games are designed to evaluate a range of AI capabilities such as strategic reasoning, planning, knowledge application, safety, and adaptability. A key novelty is integrating automatic prompt optimization to ensure fair comparisons by eliminating biases from human prompt engineering and support arbitrary prompting strategies. Furthermore, ZeroSumEval measures AI models' abilities to self-improve from limited observations and assesses their robustness against adversarial or misleading examples during prompt optimization. Building upon recent studies that highlight the effectiveness of game-based evaluations for LLMs, ZeroSumEval enhances these approaches by providing a standardized and extensible framework for rigorous assessment. We find ZeroSumEval correlates strongly with expensive human evaluations (Chatbot Arena) and disagrees with benchmarks with known overfitting and saturation issues. Inspecting match traces reveals models that allocate more tokens to thought processes perform strongly in games involving planning capabilities.", "keywords": [ "Large Language Model Evaluation", "Foundation Model Evaluation", "ELO Ranking" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 1, 1, 1 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "YFxfcQMLWX", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 9, "author_info_list": [ { "name": "Dalton Jones", "gender": "unknown", "institution": "Qualcomm Inc, QualComm", "country": "", "position": "Researcher" }, { "name": "Fatih Porikli", "gender": "Male", "institution": "QualComm", "country": "US", "position": "Senior Director" }, { "name": "Hong Cai", "gender": "Male", "institution": "Qualcomm AI Research", "country": "US", "position": "Researcher" }, { "name": "Hsin-Pai Cheng", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "US", "position": "Researcher" }, { "name": "Manish Kumar Singh", "gender": "Male", "institution": "Qualcomm AI Research", "country": "US", "position": "Researcher" }, { "name": "Matthew Harper Langston", "gender": "unknown", "institution": "Qualcomm Inc, QualComm", "country": "US", "position": "Researcher" }, { "name": "Pierre-David Letourneau", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "", "position": "Researcher" }, { "name": "Shizhong Han", "gender": "Male", "institution": "QualComm", "country": "US", "position": "Researcher" }, { "name": "Yunxiao Shi", "gender": "Male", "institution": "Qualcomm AI Research", "country": "US", "position": "Researcher" } ], "n_formula": 58, "n_formula_1": 12, "n_ref_uni": 30, "n_ref": 74, "n_ref_all": 99, "n_fig": 2, "n_tab": 3, "L_tab": 152, "n_element_tab": 12, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 464, "n_element_tab_1": 45, "formula_len_all": 3126, "formula_len_all_1": 397, "len_all": 145488, "len_all_1": 65155, "len_abs": 1371, "len_title": 139, "len_sents": 42640, "len_sents_1": 30573, "n_sents": 361, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1376, "title": "PADRe: A Unifying Polynomial Attention Drop-in Replacement for Efficient Vision Transformer", "abs": "We present Polynomial Attention Drop-in Replacement (PADRe), a novel and unifying framework designed to replace the conventional self-attention mechanism in transformer models. Notably, several recent alternative attention mechanisms, including Hyena, Mamba, SimA, Conv2Former, and Castling-ViT, can be viewed as specific instances of our PADRe framework. PADRe leverages polynomial functions and draws upon established results from approximation theory, enhancing computational efficiency without compromising accuracy. PADRe's key components include multiplicative nonlinearities, which we implement using straightforward, hardware-friendly operations such as Hadamard products, incurring only linear computational and memory costs. PADRe further avoids the need for using complex functions such as Softmax, yet it maintains comparable or superior accuracy compared to traditional self-attention. We assess the effectiveness of PADRe as a drop-in replacement for self-attention across diverse computer vision tasks. These tasks include image classification, image-based 2D object detection, and 3D point cloud object detection. Empirical results demonstrate that PADRe runs significantly faster than the conventional self-attention (11x~43x faster on server GPU and mobile NPU) while maintaining similar accuracy when substituting self-attention in the transformer models.", "keywords": [ "Efficient Vision Transformer", "Attention Approximation", "Computer Vision", "Deep Learning" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "YFdopzmpdr", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Aojun Lu", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "PhD student" }, { "name": "Hangjie Yuan", "gender": "Male", "institution": "DAMO Academy, Alibaba Group", "country": "CN", "position": "Research Scientist" }, { "name": "Tao Feng", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Yanan Sun", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 64, "n_ref_all": 77, "n_fig": 5, "n_tab": 7, "L_tab": 2451, "n_element_tab": 206, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1939, "n_element_tab_1": 201, "formula_len_all": 228, "formula_len_all_1": 223, "len_all": 137264, "len_all_1": 60890, "len_abs": 1448, "len_title": 149, "len_sents": 41703, "len_sents_1": 27249, "n_sents": 317, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1455, "title": "Rethinking the Stability-Plasticity Trade-off in Continual Learning from an Architectural Perspective", "abs": "The quest for Continual Learning (CL) seeks to empower neural networks with the ability to learn and adapt incrementally. Central to this pursuit is addressing the stability-plasticity dilemma, which involves striking a balance between two conflicting objectives: preserving previously learned knowledge and acquiring new knowledge. Existing studies have proposed numerous CL methods to achieve this trade-off. However, these methods often overlook the impact of basic architecture on stability and plasticity, thus the trade-off is limited to the parameter level. In this paper, we delve into the conflict between stability and plasticity at the architectural level. We reveal that under an equal parameter constraint, deeper networks exhibit better plasticity, while wider networks are characterized by superior stability. To address this architectural-level dilemma, we introduce a novel framework denoted Dual-Architecture (Dual-Arch), which serves as a plug-in component for CL. This framework leverages the complementary strengths of two distinct and independent networks: one dedicated to plasticity and the other to stability. Each network is designed with a specialized and lightweight architecture, tailored to its respective objective. Extensive experiments across datasets and CL methods demonstrate that Dual-Arch can enhance the performance of existing CL methods while being up to 87% more compact in terms of parameters than the baselines.", "keywords": [ "continual learning; incremental learning; stability-plasticity dilemma" ], "rating_list": [ 6, 3, 6, 5, 6 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 4, 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 2, 3 ] }, { "paper_id": "YFOg1LUGG1", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Dongha Lee", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Jinyoung Yeo", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Yeongbin Seo", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 47, "n_ref": 76, "n_ref_all": 104, "n_fig": 6, "n_tab": 3, "L_tab": 1129, "n_element_tab": 154, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 507, "n_element_tab_1": 79, "formula_len_all": 178, "formula_len_all_1": 125, "len_all": 166459, "len_all_1": 69845, "len_abs": 961, "len_title": 152, "len_sents": 54557, "len_sents_1": 33560, "n_sents": 458, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 899, "title": "Detecting Hallucination Before Answering: Semantic Compression Through Instruction", "abs": "Large language models (LLMs) excel in various tasks but often suffer from hallucinations, providing incorrect information with high confidence. To address this, we focus on detecting when an LLM knows or does not know an answer, a concept referred to as the ``feeling of knowing'' (FoK). We propose a novel approach called Semantic Compression by trying to Answer in One-word (SCAO), which enables efficient FoK detection before generating full sentences, with only minimal computational cost. Additionally, we introduce a method to measure confounding variable effects in benchmarks, an Approximate Misannotation Effect (AME) test. Our experiments demonstrate that the feature fusion model of our SCAO and probing achieves enhanced performance in FoK detection in both short and long-form entity questions. The code and the dataset is available online (https://anonymous.4open.science/r/SCAO-2FF8).", "keywords": [ "hallucination", "hallucination detection", "feeling of knowing", "LLM", "large language model" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "YFKH1vO0W2", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Hedi Zisling", "gender": "Male", "institution": "Ben-Gurion University of the Negev", "country": "IL", "position": "PhD student" }, { "name": "Ilan Naiman", "gender": "Male", "institution": "Amazon", "country": "IL", "position": "Researcher" }, { "name": "Nimrod Berman", "gender": "Male", "institution": "Ben-Gurion University of the Negev", "country": "", "position": "MS student" }, { "name": "Omri Azencot", "gender": "Unspecified", "institution": "Ben-Gurion University of the Negev", "country": "IL", "position": "Assistant Professor" }, { "name": "Supasorn Suwajanakorn", "gender": "Male", "institution": "Vidyasirimedhi Institute of Science and Technology", "country": "TH", "position": "Lecturer" } ], "n_formula": 28, "n_formula_1": 15, "n_ref_uni": 67, "n_ref": 140, "n_ref_all": 156, "n_fig": 8, "n_tab": 4, "L_tab": 874, "n_element_tab": 74, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 915, "n_element_tab_1": 38, "formula_len_all": 2444, "formula_len_all_1": 1039, "len_all": 189624, "len_all_1": 67180, "len_abs": 288, "len_title": 126, "len_sents": 51308, "len_sents_1": 29371, "n_sents": 390, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1530, "title": "From Noise to Factors: Diffusion-based Unsupervised Sequential Disentanglement", "abs": "Unsupervised representation learning, in particular, sequential disentanglement, where the goal is to learn disentangled static and dynamic factors of variation, remains a significant challenge due to the absence of labels. Existing models, based on variational autoencoders and generative adversarial networks, achieved success in certain domains, but they often struggle with disentangling sequences, especially when dealing with real-world complexity and variability. Further, there is no real-world evaluation protocol for assessing the effectiveness of sequential disentanglement models. Recently, diffusion autoencoders have emerged as a new promising generative model, offering semantically rich representations by gradual noise-to-data transformations. Despite their advantages, these models face limitations: they are non-sequential, fail to disentangle the latent space effectively, and are computationally intensive, making them difficult to scale to sequences. In this work, we introduce our diffusion sequential disentanglement autoencoder (DiffSDA), a novel approach effective on real-world visual data and accompanied by a new and challenging evaluation protocol. DiffSDA is based on a new probabilistic modeling and is implemented using latent diffusion models and efficient samplers, facilitating processing of high-resolution videos. We test our approach on several real-world datasets and metrics, and we demonstrate its effectiveness in comparison to recent state-of-the-art sequential disentanglement methods.", "keywords": [ "Sequential Disentanglement", "Deep Learning", "Generative Models" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "YFDM6uMMSE", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Alexander Tropsha", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "", "position": "Full Professor" }, { "name": "James Wellnitz", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "", "position": "PhD student" }, { "name": "Junier Oliva", "gender": "Male", "institution": "University of North Carolina, Chapel Hill", "country": "", "position": "Assistant Professor" }, { "name": "Yunni Qu", "gender": "Female", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 17, "n_ref_uni": 48, "n_ref": 94, "n_ref_all": 116, "n_fig": 6, "n_tab": 10, "L_tab": 8298, "n_element_tab": 1081, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3639, "n_element_tab_1": 453, "formula_len_all": 1292, "formula_len_all_1": 1344, "len_all": 155154, "len_all_1": 72970, "len_abs": 1233, "len_title": 74, "len_sents": 35653, "len_sents_1": 28636, "n_sents": 263, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 713, "title": "EMOE: Expansive Matching of Experts for Robust Uncertainty Based Rejection", "abs": "Expansive Matching of Experts (EMOE) is a novel method that utilizes support-expanding, extrapolatory pseudo-labeling to improve prediction and uncertainty based rejection on out-of-distribution (OOD) points. We propose an expansive data augmentation technique that generates OOD instances in a latent space, and an empirical trial based approach to filter out augmented expansive points for pseudo-labeling. EMOE utilizes a diverse set of multiple base experts as pseudo-labelers on the augmented data to improve OOD performance through a shared MLP with multiple heads (one per expert). We demonstrate that EMOE achieves superior performance compared to state-of-the-art methods on both image and tabular data.", "keywords": [ "out of distribution generalization", "OOD", "reject option", "data augmentation" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 1 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "YERRy6v5uA", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Lihui Chen", "gender": "Unspecified", "institution": "Nanyang Technological University", "country": "", "position": "Associate Professor" }, { "name": "Sitao Luan", "gender": "Male", "institution": "McGill University", "country": "", "position": "PhD student" }, { "name": "Xiang Li", "gender": "unknown", "institution": "Shenzhen University", "country": "CN", "position": "MS student" }, { "name": "Xiaojiang Peng", "gender": "Male", "institution": "Shenzhen Technology University", "country": "CN", "position": "Full Professor" }, { "name": "Yilun Zheng", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "zhangzhuofan", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Wang Ziming", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "MS student" } ], "n_formula": 28, "n_formula_1": 5, "n_ref_uni": 44, "n_ref": 187, "n_ref_all": 225, "n_fig": 10, "n_tab": 19, "L_tab": 17333, "n_element_tab": 1724, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 10306, "n_element_tab_1": 513, "formula_len_all": 3047, "formula_len_all_1": 262, "len_all": 300086, "len_all_1": 83852, "len_abs": 1656, "len_title": 103, "len_sents": 74588, "len_sents_1": 30571, "n_sents": 1036, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1664, "title": "Rethinking Structure Learning For Graph Neural Networks", "abs": "To improve the performance of Graph Neural Networks (GNNs), Graph Structure Learning (GSL) has been extensively applied to reconstruct or refine original graph structures. While GSL is generally thought to improve GNN performance, it often leads to longer training times and more hyperparameter tuning. Besides, the distinctions among current GSL methods remain ambiguous from the perspective of GNN training, and there is a lack of theoretical analysis to quantify their effectiveness. Recent studies further suggest that GSL does not consistently outperform baseline GNNs under the same hyperparameter tuning. This motivates us to ask a critical question: *Is GSL really useful for improving GNN performance?* To address this question, we first propose a new GSL framework, which includes three steps: GSL bases (i.e. node representations used to construct new graphs) construction, new structure construction, and view fusion, to better understand GSL. Then, our empirical studies and theoretical analysis show that the mutual information (MI) between node representations and labels does not increase after applying graph convolution on GSL graphs that are constructed by similarity, indicating GSL could be unnecessary in most cases. Our experiments fairly reassess the performance of GSL and reveal that adding GSL to GNN baselines or removing GSL in state-of-the-art models has negligible impact on node classification accuracy. We also report that pretrained GSL bases, parameter separation, and early fusion are effective designs within GSL. Our findings challenge the necessity of complex GSL methods and underscore the value of simplicity in GNN design.", "keywords": [ "Graph Neural Networks", "Graph Structure Learning" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "YE6N8htoFQ", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Qian Ma", "gender": "unknown", "institution": "Beijing Normal University", "country": "CN", "position": "MS student" }, { "name": "Ruoxiang Xu", "gender": "Male", "institution": "Beijing Normal University", "country": "CN", "position": "MS student" }, { "name": "Yongqiang Cai", "gender": "Male", "institution": "Beijing Normal University", "country": "", "position": "Lecturer" } ], "n_formula": 88, "n_formula_1": 42, "n_ref_uni": 46, "n_ref": 69, "n_ref_all": 81, "n_fig": 0, "n_tab": 1, "L_tab": 272, "n_element_tab": 17, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 272, "n_element_tab_1": 17, "formula_len_all": 7845, "formula_len_all_1": 2788, "len_all": 217172, "len_all_1": 67724, "len_abs": 3032, "len_title": 128, "len_sents": 46624, "len_sents_1": 25944, "n_sents": 446, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 870, "title": "Vocabulary In-Context Learning in Transformers: Benefits of Positional Encoding", "abs": "Numerous studies have demonstrated that the Transformer architecture possesses the capability for in-context learning (ICL). In scenarios involving function approximation, context can serve as a control parameter for the model, endowing it with the universal approximation property (UAP). In practice, context is represented by tokens from a finite set, referred to as a vocabulary, which is the case considered in this paper, i.e., vocabulary in-context learning (VICL). We demonstrate that VICL in single-layer Transformers, without positional encoding, does not possess the UAP; however, it is possible to achieve the UAP when positional encoding is included. Several sufficient conditions for the positional encoding are provided. Our findings reveal the benefits of positional encoding from an approximation theory perspective in the context of in-context learning.", "keywords": [ "Transformer", "Universal approximation", "In-context Learning;Vocabulary" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 2, 2, 4 ] }, { "paper_id": "YDuYWjbKDA", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Alexander Kostin", "gender": "Male", "institution": "Sberbank", "country": "", "position": "Researcher" }, { "name": "Andrey Savchenko", "gender": "Male", "institution": "Sber AI Lab", "country": "RU", "position": "Principal Researcher" }, { "name": "Dmitry Igorevich Sorokin", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "", "position": "Researcher" }, { "name": "Gleb Gusev", "gender": "Male", "institution": "ARTIFICIAL INTELLIGENCE RESEARCH INSTITUTE (AIRI)", "country": "RU", "position": "Principal Researcher" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 31, "n_ref": 57, "n_ref_all": 70, "n_fig": 6, "n_tab": 15, "L_tab": 3516, "n_element_tab": 392, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1972, "n_element_tab_1": 101, "formula_len_all": 1611, "formula_len_all_1": 1561, "len_all": 141328, "len_all_1": 68045, "len_abs": 980, "len_title": 140, "len_sents": 38994, "len_sents_1": 29202, "n_sents": 360, "n_sents_1": 242, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 986, "title": "TreeDQN: Sample-Efficient Off-Policy Reinforcement Learning for Combinatorial Optimization", "abs": "A convenient approach to optimally solving combinatorial optimization tasks is Branch-and-Bound method. The branching heuristic in this method can be learned to solve a large set of similar tasks. The promising results here are achieved by the recently appeared on-policy reinforcement learning (RL) method based on the tree Markov Decision Process (tMDP). To overcome its main disadvantages, namely, very large training time and unstable training, we propose TreeDQN, a sample-efficient off-policy RL method that is trained by optimizing the geometric mean of expected return. To theoretically support the training procedure for our method, we prove the contraction property of the Bellman operator for the tree MDP. As a result, our method requires up to 10 times less training data, performs faster than known on-policy methods on synthetic tasks. Moreover, TreeDQN significantly outperforms the state-of-the-art techniques on a challenging practical task from the ML4CO competition.", "keywords": [ "reinforcement learning", "combinatorial optimization", "branch-and-bound", "ML4CO" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 2, 1, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 4, 4, 2 ] }, { "paper_id": "YD6xlDstbz", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Jing Liu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Jun Zhang", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Ruihao Gong", "gender": "Male", "institution": "SenseTime", "country": "CN", "position": "Principal Researcher" }, { "name": "Xinjie Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" }, { "name": "Yushi Huang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zining Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 48, "n_ref": 126, "n_ref_all": 159, "n_fig": 11, "n_tab": 17, "L_tab": 7960, "n_element_tab": 1003, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 636, "n_element_tab_1": 77, "formula_len_all": 1481, "formula_len_all_1": 614, "len_all": 216816, "len_all_1": 59518, "len_abs": 348, "len_title": 109, "len_sents": 51769, "len_sents_1": 26820, "n_sents": 492, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 2495, "title": "HarmoniCa: Harmonizing Training and Inference for Better Feature Cache in Diffusion Transformer Acceleration", "abs": "Diffusion Transformers (DiTs) have gained prominence for outstanding scalability and extraordinary performance in generative tasks. However, their considerable inference costs impede practical deployment. The feature cache mechanism, which involves storing and retrieving redundant computations across timesteps, holds promise for reducing per-step inference time in diffusion models. Most existing caching methods for DiT are manually designed. Although the learning-based approach attempts to optimize strategies adaptively, it suffers from discrepancies between training and inference, which hampers both the performance and acceleration ratio. \nUpon detailed analysis, we pinpoint that these discrepancies primarily stem from two aspects: (1) _Prior Timestep Disregard_, where training ignores the effect of cache usage at earlier timesteps, and (2) _Objective Mismatch_, where the training target (align predicted noise in each timestep) deviates from the goal of inference (generate the high-quality image). To alleviate these discrepancies, we propose **HarmoniCa**, a novel method that **harmoni**zes training and inference with a novel learning-based **ca**ching framework built upon _Step-Wise Denoising Training_ (SDT) and _Image Error Proxy-Guided Objective_ (IEPO). Compared to the traditional training paradigm, the newly proposed SDT maintains the continuity of the denoising process, enabling the model to leverage information from prior timesteps during training, similar to the way it operates during inference. Furthermore, we design IEPO, which integrates an efficient proxy mechanism to approximate the final image error caused by reusing the cached feature. Therefore, IEPO helps balance final image quality and cache utilization, resolving the issue of training that only considers the impact of cache usage on the predicted output at each timestep. Extensive experiments on class-conditional and text-to-image (T2I) tasks for 8 models and 4 samplers with resolutions ranging from $256\\times256$ to $2048\\times2048$ demonstrate the exceptional performance and speedup capabilities of our HarmoniCa. For example, HarmoniCa is the first feature cache method applied to the 20-step PixArt-$\\alpha$ $1024\\times1024$ that achieves over 1.5$\\times$ speedup in latency with an improved FID compared to the non-accelerated model. Remarkably, HarmoniCa requires no image data during training and reduces about 25\\% of training time compared to the existing learning-based approach.", "keywords": [ "diffusion", "acceleration", "feature cache" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 1, 4, 3, 4 ] }, { "paper_id": "YCwN7wQA6W", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Lifu Huang", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Assistant Professor" }, { "name": "Qifan Wang", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Principal Researcher" }, { "name": "Weifeng Ge", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yixin Cao", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yu Cheng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Yufan Zhou", "gender": "Male", "institution": "Adobe ", "country": "US", "position": "Researcher" }, { "name": "Zhiyang Xu", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "PhD student" }, { "name": "wang haibo", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "SHIZHE DIAO", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 30, "n_ref": 121, "n_ref_all": 150, "n_fig": 5, "n_tab": 6, "L_tab": 2519, "n_element_tab": 280, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1297, "n_element_tab_1": 145, "formula_len_all": 351, "formula_len_all_1": 278, "len_all": 166493, "len_all_1": 66095, "len_abs": 1375, "len_title": 140, "len_sents": 40696, "len_sents_1": 29338, "n_sents": 282, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1424, "title": "Grounded-VideoLLM: Sharpening Fine-grained Temporal Grounding in Video Large Language Models", "abs": "Video Large Language Models (Video-LLMs) have demonstrated remarkable capabilities in coarse-grained video understanding, however, they struggle with fine-grained temporal grounding. In this paper, we introduce $\\textit{Grounded-VideoLLM}$, a novel Video-LLM adept at perceiving and reasoning over specific video moments in a fine-grained manner. We identify that current Video-LLMs have limitations for fine-grained video understanding since they lack effective temporal modeling and timestamp representation. In light of this, we sharpen our model by incorporating (1) an additional temporal stream to encode the relationships between frames and (2) discrete temporal tokens enriched with specific time knowledge to represent timestamps. To optimize the training of $\\textit{Grounded-VideoLLM}$, we employ a multi-stage training scheme, beginning with simple video-captioning tasks and progressively introducing video temporal grounding tasks of increasing complexity. To further enhance $\\textit{Grounded-VideoLLM}$'s temporal reasoning capability, we also curate a grounded VideoQA dataset by an automatic annotation pipeline. Extensive experiments demonstrate that $\\textit{Grounded-VideoLLM}$ not only excels in fine-grained grounding tasks such as temporal sentence grounding, dense video captioning, and grounded VideoQA, but also shows great potential as a versatile video assistant for general video understanding.", "keywords": [ "multimodal large language model; video understanding; temporal reasoning" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "YCu7H0kFS3", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Marc G Bellemare", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "Adjunct Professor" }, { "name": "Pierluca D'Oro", "gender": "Male", "institution": "Facebook", "country": "CA", "position": "Researcher" }, { "name": "Nathan Rahn", "gender": "Male", "institution": "McGill University", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 32, "n_ref": 55, "n_ref_all": 79, "n_fig": 10, "n_tab": 1, "L_tab": 297, "n_element_tab": 40, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 297, "n_element_tab_1": 40, "formula_len_all": 102, "formula_len_all_1": 102, "len_all": 144386, "len_all_1": 66347, "len_abs": 1248, "len_title": 121, "len_sents": 47672, "len_sents_1": 33662, "n_sents": 340, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1254, "title": "Controlling Large Language Model Agents with Entropic Activation Steering", "abs": "The rise of large language models (LLMs) has prompted increasing interest in their use as in-context learning agents. At the core of agentic behavior is the capacity for exploration, or the ability to actively gather information about the environment. But how do LLM agents explore, and how can we control their exploratory behaviors? To answer these questions, we take a representation-level perspective, and introduce Entropic Activation Steering (EAST), an activation steering method for in-context LLM agents. Firstly, we demonstrate that EAST can effectively manipulate an LLM agent's exploration by directly affecting the high-level actions parsed from the outputs of the LLM, in contrast to token-level temperature sampling. Secondly, we reveal how applying this control modulates the uncertainty exhibited in the LLM's thoughts, guiding the agent towards more exploratory actions. Finally, we demonstrate that the steering vectors obtained by EAST generalize across task variants. In total, these results show that LLM agents explicitly encode uncertainty over their actions in their representation space. Our work paves the way for a new understanding of the functioning of LLM agents and to effective control of their decision-making behaviors.", "keywords": [ "LLM Agents", "Reinforcement Learning", "Interpretability" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "YCdag94iZs", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Amit Kabya", "gender": "Male", "institution": "Bar-Ilan University", "country": "", "position": "Intern" }, { "name": "Avital Rose", "gender": "Female", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" }, { "name": "Ofek Akerman", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" }, { "name": "Yarin Bekor", "gender": "Male", "institution": "Bar-Ilan University", "country": "", "position": "Intern" }, { "name": "Yoram Louzoun", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 30, "n_ref": 64, "n_ref_all": 81, "n_fig": 7, "n_tab": 3, "L_tab": 1634, "n_element_tab": 177, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 313, "formula_len_all_1": 313, "len_all": 100385, "len_all_1": 63656, "len_abs": 1640, "len_title": 110, "len_sents": 30708, "len_sents_1": 30525, "n_sents": 265, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1658, "title": "MILCA: Multiple Instance Learning using Counting and Attention", "abs": "In Multiple Instance Learning (MIL), a bag is comprised of instances and the label is prescribed to the whole bag, with no information on the labels of each instance.\nThe leading approaches for MIL are Embedded Space (ES) solutions, where the full bag is embedded into a vector space. \nWhile very complex models were constructed for MIL classification tasks, we show that often some features are associated with a class, and a simple counting/summing algorithm leads to similar or better accuracy than current solutions. This can be improved in some cases by weighting these selected features using a fully connected network to predict the coefficient of each feature. \nHowever, a simple relative contribution of each feature, where the sum of the coefficients is normalized to 1, fails to count the feature. Thus instead, we replace the softmax by a projection of the coefficients to [-1,1] or [0,1] but do not limit their sum. This allows the model to count features. \nThe resulting algorithm - MILCA (Multiple Instance Learning using Counting and Attention) is applied to multiple previous and new real-world MIL tasks, as well as recovering the host disease history from sequenced T Cell Receptor Repertoires. In most cases, MILCA is significantly better and way more efficient than currently used MIL algorithms, with a 3 \\% higher accuracy than current SOTA on average. To summarize, in MIL classification tasks, where often the number of features is large compared to the number of bags, complex models are typically not better than a weighted sum of informative features. \nThe code for MILCA is available at: github.com/submissionanonymous6/MILCA", "keywords": [ "Multiple Instance Learning", "counting", "attention" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "YCOVTlMFIG", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Bo He", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Donghyun Kim", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Assistant Professor" }, { "name": "Young Kyun Jang", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Postdoc" }, { "name": "Zihang Meng", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Ser-Nam Lim", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 19, "n_ref": 36, "n_ref_all": 54, "n_fig": 4, "n_tab": 6, "L_tab": 1949, "n_element_tab": 163, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1966, "n_element_tab_1": 168, "formula_len_all": 422, "formula_len_all_1": 422, "len_all": 99753, "len_all_1": 66540, "len_abs": 1299, "len_title": 124, "len_sents": 30529, "len_sents_1": 29554, "n_sents": 214, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1342, "title": "$\\text{Slerp}^{+}$: Spherical Linear Interpolation for Unified Compositional Retrieval", "abs": "Zero-shot composed image/video retrieval is a challenging task that involves using a combination of a reference visual input and a relative caption as a query to search for target visual data. Earlier studies have treated composed image retrieval and composed video retrieval methods separately, potentially neglecting the benefits of integrating image-video-text representation learning. In this paper, we consolidate these tasks into a single Composed \\emph{Visual} Retrieval (CVR) task, which requires the composition of image and video samples with textual modifications using a unified retrieval model. Our principal insight is that the video modality can be effectively added to existing vision-language pretrained models. When integrated with the Spherical Linear Interpolation (Slerp) method previously proposed for Composed Image Retrieval (CoIR), we found that it results in an effective approach for solving the CVR task, which we called $\\text{Slerp}^{+}$. Extensive experiments demonstrate $\\text{Slerp}^{+}$'s superiority across various composed image and video retrieval benchmarks, including our newly proposed video benchmark. Notably, $\\text{Slerp}^{+}$ mutually enhances image and video retrieval performance over single-modality models, underscoring its potential to transform the field of compositional visual retrieval.", "keywords": [ "multi-modal representation learning", "composed retrieval" ], "rating_list": [ 3, 3, 5, 8, 5 ], "soundness_list": [ 3, 1, 2, 3, 4 ], "presentation_list": [ 2, 2, 3, 4, 4 ], "contribution_list": [ 2, 1, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 4, 5 ] }, { "paper_id": "YBv9EExJPk", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Kobi Rahimi", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" }, { "name": "Ofir Lindenbaum", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "Assistant Professor" }, { "name": "Tom Tirer", "gender": "unknown", "institution": "Bar-Ilan University", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 2, "n_ref_uni": 57, "n_ref": 99, "n_ref_all": 173, "n_fig": 58, "n_tab": 2, "L_tab": 1234, "n_element_tab": 114, "n_fig_1": 26, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1493, "formula_len_all_1": 123, "len_all": 206683, "len_all_1": 69099, "len_abs": 1249, "len_title": 142, "len_sents": 63200, "len_sents_1": 32629, "n_sents": 534, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1255, "title": "Multiple Descents in Unsupervised Auto-Encoders: The Role of Noise, Domain Shift and Anomalies", "abs": "The phenomenon of double descent has recently gained attention in supervised learning. It challenges the conventional wisdom of the bias-variance trade-off by showcasing a surprising behavior. As the complexity of the model increases, the test error initially decreases until reaching a certain point where the model starts to overfit the train set, causing the test error to rise. However, deviating from classical theory, the error exhibits another decline when exceeding a certain degree of over-parameterization. We study the presence of double descent in unsupervised learning, an area that has received little attention and is not yet fully understood. We conduct extensive experiments using under-complete auto-encoders (AEs) for various applications, such as dealing with noisy data, domain shifts, and anomalies. We use synthetic and real data and identify model-wise, epoch-wise, and sample-wise double descent for all the aforementioned applications. Finally, we assessed the usability of the AEs for detecting anomalies and mitigating the domain shift between datasets. Our findings indicate that over-parameterized models can improve performance not only in terms of reconstruction, but also in enhancing capabilities for the downstream task.", "keywords": [ "double descent", "unsupervised learning", "autoencoders", "domain adaptation", "bias-variance curve", "anomalies" ], "rating_list": [ 3, 3, 5, 5, 5 ], "soundness_list": [ 4, 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 3 ] }, { "paper_id": "YBht9Vp5vC", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Qingmin Liao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Tianjian Ouyang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xin Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yu Shang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 32, "n_ref": 54, "n_ref_all": 75, "n_fig": 6, "n_tab": 10, "L_tab": 4898, "n_element_tab": 386, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 4361, "n_element_tab_1": 177, "formula_len_all": 516, "formula_len_all_1": 519, "len_all": 161036, "len_all_1": 63391, "len_abs": 2490, "len_title": 120, "len_sents": 47835, "len_sents_1": 27738, "n_sents": 343, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 2506, "title": "UrbanMLLM: Joint Learning of Cross-view Imagery for Urban Understanding", "abs": "Multimodal large language models (MLLMs) have exhibited remarkable capabilities for performing complex vision-language tasks in various domains.\nCurrently, MLLMs based on urban imagery in urban studies are only developed focusing on remote sensing imagery.\nHowever, except for the macroscopic information from remote sensing imagery, effective urban understanding also requires detailed appearance information of urban zones from street-view imagery, which is largely overlooked by existing MLLMs.\nThe primary challenges of developing such a versatile urban MLLM are twofold. \nFirstly, it needs a large-scale corpus with well-organized, cross-view urban imagery paired with corresponding text for cross-modal training.\nSecondly, traditional MLLMs typically learn image-text pairs independently, hard to support joint modeling of cross-view urban imagery.\nTo address these challenges, in this work, we propose UrbanMLLM, a novel MLLM that jointly learns from remote sensing and street-view imagery to harness their complementary information.\nWe first collect a large-scale dataset containing satellite-view and street-view imagery along with their geotags and annotated texts.\nTechnically, we propose a brand MLLM architecture with a cross-view perceiver to explicitly connect visual information of cross-view urban imagery.\nWe also introduce a novel pre-training paradigm based on structural interleaved urban image-text documents integrating satellite-view, street-view imagery and related textual descriptions.\nThis approach encourages the model to implicitly learn the relationships between different types of urban imagery, enhancing the understanding in each domain.\nWe evaluate our model on a comprehensive benchmark comprising 13 diverse urban understanding tasks across satellite-view, street-view, and cross-view domains. These tasks include scene classification, object reasoning, spatial relationship reasoning, geo-localization, landmark reasoning, and indicator prediction, providing a robust assessment of the model's capabilities.\nExtensive experiments demonstrate that UrbanMLLM achieves an average of 27.3\\% and 25.5\\% performance improvement compared with the best open-sourced and closed-sourced MLLMs, respectively.\nMoreover, we thoroughly study the impact of different pre-training data choices and model scales on performance, offering practical insights for effective MLLM design. The proposed UrbanMLLM offers a scalable and versatile solution for understanding urban environments.", "keywords": [ "Multi-modal large language model", "Cross-view learning", "Urban understanding" ], "rating_list": [ 6, 5, 6, 5, 6 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 5, 4 ] }, { "paper_id": "YAvEKf1KUd", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Fani Boukouvala", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "", "position": "Associate Professor" }, { "name": "Suryateja Ravutla", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zachary A. Kilwein", "gender": "Male", "institution": "Sandia National Laboratories", "country": "", "position": "Postdoc" } ], "n_formula": 26, "n_formula_1": 13, "n_ref_uni": 18, "n_ref": 27, "n_ref_all": 55, "n_fig": 10, "n_tab": 6, "L_tab": 1370, "n_element_tab": 140, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 306, "n_element_tab_1": 35, "formula_len_all": 2533, "formula_len_all_1": 1488, "len_all": 121270, "len_all_1": 42556, "len_abs": 1616, "len_title": 116, "len_sents": 35267, "len_sents_1": 17646, "n_sents": 296, "n_sents_1": 157, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1624, "title": "Incorporating Neural ODEs into DAE-Constrained Optimization Problems", "abs": "Differential algebraic equations (DAEs) are pivotal in dynamic optimization across diverse fields, from process control to flight trajectory optimization and epidemiological modeling. Traditional methods like single shooting, multiple shooting, and direct transcription effectively optimize known mechanistic models. However, significant challenges arise when the underlying equations are unknown or deviate from empirical data. While black-box optimization strategies can address some issues, challenges persist regarding data quality, non-linearity, and the inclusion of constraints. Recent advances in machine learning, particularly Neural ODEs, offer promising tools for continuous representation of dynamic systems. This work bridges the gap between machine learning representations of dynamic systems and optimization methodologies, enabling a novel approach for solving DAEs with data-driven components. We demonstrate this approach using numerical examples of DAE problems and realistic case studies, including biochemical reactor control and disease spread prevention. Our results highlight the efficacy of incorporating Neural ODEs into equation-based solvers, showing improved performance over existing strategies such as SINDy. Additionally, we formalize the optimization program for NN-embedded DAEs and present representations for common neural network architectures (e.g., ReLU, tanh). This work contributes a novel framework for dynamic system optimization, integrating machine learning advancements with traditional optimization techniques, and offers practical insights through comprehensive case studies.", "keywords": [ "Differential Algebraic Equations", "Neural Ordinary Differential Equations", "Dynamic Optimization", "Hybrid Modeling" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "YAf7MSsdTu", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Luana Ruiz", "gender": "Female", "institution": "Whiting School of Engineering", "country": "", "position": "Assistant Professor" }, { "name": "My Le", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Souvik Dhara", "gender": "Male", "institution": "Purdue University", "country": "", "position": "Assistant Professor" } ], "n_formula": 64, "n_formula_1": 23, "n_ref_uni": 26, "n_ref": 55, "n_ref_all": 75, "n_fig": 4, "n_tab": 2, "L_tab": 1238, "n_element_tab": 86, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8055, "formula_len_all_1": 2043, "len_all": 140269, "len_all_1": 64257, "len_abs": 1330, "len_title": 122, "len_sents": 39311, "len_sents_1": 27615, "n_sents": 358, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1333, "title": "Local-Global Shortest Path Algorithms on Random Graphs, Enhanced with GNNs", "abs": "Graph neural networks (GNNs) using local message passing were recently shown to inherit the intrinsic limitations of local algorithms in solving combinatorial graph optimization problems such as finding shortest distances (Loukas, 2020). To address this issue, Awasthi et al. (2022) proposed architectures based on Bourgain\u2019s (1985) seminal work on Hilbert space embeddings. These architectures enhance local message passing in GNNs with a single global computation, yielding a local-global algorithm. This paper focuses on the average-case analysis of more general local-global algorithms for finding shortest distances (of which GNN+ is a particular case). Our primary contribution is a theoretical analysis of these algorithms on Erd\u0151s-R\u00e9nyi (ER) random graphs. We prove that, on random graphs, these algorithms have lower distortion of shortest distances for most pairs of nodes w.h.p. while requiring a lower embedding dimension. Inspired by Awasthi et al. (2022), and to automate local computations and improve computational efficiency in practical scenarios, we further propose a modification to these algorithms that incorporates GNNs in the local computation phase. Empirical results on ER graphs and benchmark graph datasets demonstrate the enhanced performance of the GNN-augmented algorithm over the traditional approach.", "keywords": [ "shortest path", "random graphs", "graph neural networks", "Erd\u0151s-R\u00e9nyi" ], "rating_list": [ 1, 6, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 1, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "YAMlVKRLnc", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Haitian Liu", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "MS student" }, { "name": "Haoyuan Chai", "gender": "unknown", "institution": "Vaneval", "country": "CN", "position": "Researcher" }, { "name": "Jiayi Wang", "gender": "unknown", "institution": "Vaneval", "country": "", "position": "Researcher" }, { "name": "Qian Chen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Renjun Zhang", "gender": "Male", "institution": "Peking Union Medical College", "country": "CN", "position": "MS student" }, { "name": "Tengxiao Wu", "gender": "unknown", "institution": "Vaneval", "country": "CN", "position": "Researcher" }, { "name": "Weishan Zhao", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Weixiang Yan", "gender": "Male", "institution": "Vaneval", "country": "CN", "position": "Researcher" }, { "name": "Wen Wang", "gender": "unknown", "institution": "Alibaba Group", "country": "", "position": "Senior Staff Algorithm Engineer" }, { "name": "Xuandong Zhao", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Yixin Zhang", "gender": "unknown", "institution": "Vaneval", "country": "CN", "position": "Researcher" }, { "name": "Li Zhu", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 9, "n_ref": 46, "n_ref_all": 86, "n_fig": 6, "n_tab": 15, "L_tab": 2384, "n_element_tab": 153, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 4308, "n_element_tab_1": 290, "formula_len_all": 249, "formula_len_all_1": 249, "len_all": 301293, "len_all_1": 76515, "len_abs": 2070, "len_title": 138, "len_sents": 93767, "len_sents_1": 32798, "n_sents": 712, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 2055, "title": "ClinicalLab: Aligning Agents for Multi-Departmental Clinical Diagnostics in the Real World", "abs": "Large language models (LLMs) have achieved significant performance progress in various natural language processing applications. However, LLMs still struggle to meet the strict requirements for accuracy and reliability in the medical field and face many challenges in clinical applications. Existing clinical diagnostic evaluation benchmarks for evaluating medical agents powered by LLMs have severe limitations. Firstly, most existing medical evaluation benchmarks face the risk of data leakage or contamination. Secondly, existing benchmarks often neglect the characteristics of multiple departments and specialization in modern medical practice. Thirdly, existing evaluation methods are limited to multiple-choice questions, which do not align with the real-world diagnostic scenarios. Lastly, existing evaluation methods lack comprehensive evaluations of end-to-end real clinical scenarios. These limitations in benchmarks in turn obstruct advancements of LLMs and agents for medicine. To address these limitations, we introduce **ClinicalLab**, a comprehensive clinical diagnosis agent alignment suite. ClinicalLab includes **ClinicalBench**, an end-to-end multi-departmental clinical diagnostic evaluation benchmark for evaluating medical agents and LLMs. ClinicalBench is based on real cases that cover 24 departments and 150 diseases. We ensure that ClinicalBench does not have data leakage. ClinicalLab also includes four novel metrics (**ClinicalMetrics**) for evaluating the effectiveness of LLMs in clinical diagnostic tasks. We evaluate 17 general and medical-domain LLMs and find that their performance varies significantly across different departments. Based on these findings, in ClinicalLab, we propose **ClinicalAgent**, an end-to-end clinical agent that aligns with real-world clinical diagnostic practices. We systematically investigate the performance and applicable scenarios of variants of ClinicalAgent on ClinicalBench. Our findings demonstrate the importance of aligning with modern medical practices in designing medical agents.", "keywords": [ "Large language models", "Medical Benchmark", "Medical Agents", "Medical Metrics" ], "rating_list": [ 6, 3, 3, 6, 3 ], "soundness_list": [ 4, 2, 2, 3, 2 ], "presentation_list": [ 4, 2, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "YAINolpm8n", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Demba E. Ba", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Nima Dehmamy", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Emmanouil Theodosis", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 31, "n_ref_uni": 25, "n_ref": 37, "n_ref_all": 54, "n_fig": 10, "n_tab": 2, "L_tab": 86, "n_element_tab": 8, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 86, "n_element_tab_1": 8, "formula_len_all": 1602, "formula_len_all_1": 1143, "len_all": 111523, "len_all_1": 70144, "len_abs": 2415, "len_title": 103, "len_sents": 39072, "len_sents_1": 30218, "n_sents": 315, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1426, "title": "Incorporating gauge-invariance in equivariant networks", "abs": "Gauge theories, which describe fundamental forces in nature, arise from the principle of locality in physical interactions. These theories are characterized by their invariance under local symmetry transformations and the presence of a gauge field that mediates interactions. While recent works have introduced gauge equivariant neural networks, these models often focus on specific cases like tangent bundles or quotient spaces, limiting their applicability to the diverse gauge theories in physics. We propose a novel architecture for learning general gauge invariant quantities by explicitly modeling the gauge field in the context of graph neural networks. Our framework fills a critical gap in the existing literature by providing a general recipe for gauge invariance without restrictions on the fiber spaces. This approach allows for the modeling of more complex gauge theories, such as those with $SU(N)$ gauge groups, which are prevalent in particle physics. We evaluate our method on classical physical systems, including the XY model on various curved geometries, demonstrating its ability to capture gauge invariant properties in settings where existing equivariant architectures fall short. Our work takes a significant step towards bridging the gap between gauge theories in physics and equivariant neural network architectures, opening new avenues for applying machine learning to fundamental physical problems.", "keywords": [ "gauge-invariance", "gauge theories", "equivariance" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 1, 4, 4, 2 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 3, 2, 4, 2 ] }, { "paper_id": "YA1Ur2eGFl", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Christian Theobalt", "gender": "Male", "institution": "Max-Planck-Institute for Informatics, Saarland Informatics Campus", "country": "DE", "position": "Director" }, { "name": "Gereon Fox", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Kai Chen", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Mohamed Elgharib", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Researcher" }, { "name": "Xingang Pan", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Yanhong Zeng", "gender": "Female", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xing Zhening", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 32, "n_ref": 109, "n_ref_all": 150, "n_fig": 20, "n_tab": 3, "L_tab": 761, "n_element_tab": 105, "n_fig_1": 14, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 133, "formula_len_all_1": 134, "len_all": 141794, "len_all_1": 62975, "len_abs": 1341, "len_title": 139, "len_sents": 41668, "len_sents_1": 31694, "n_sents": 294, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1360, "title": "Live2Diff: Live Stream Translation via Uni-directional Attention in Video Diffusion Models", "abs": "Large Language Models have shown remarkable efficacy in generating streaming data such as text and audio, thanks to their temporally uni-directional attention mechanism, which models correlations between the current token and *previous* tokens.\nHowever, video streaming remains much less explored, despite a growing need for live video processing.\nState-of-the-art video diffusion models leverage\n*bi*-directional temporal attention to model the correlations between the current frame and all the *surrounding* (i.e. including *future*) frames, which hinders them from processing streaming videos.\nTo address this problem, we present **Live2Diff**, the first attempt at designing a video diffusion model with uni-directional temporal attention, specifically targeting live streaming video translation.\nCompared to previous works, our approach ensures temporal consistency and smoothness by correlating the current frame with its predecessors and a few initial warmup frames, without any future frames.\nAdditionally, we use a highly efficient denoising scheme featuring a *kv-cache* mechanism and pipelining, to facilitate streaming video translation at interactive framerates.\nExtensive experiments demonstrate the effectiveness of the proposed attention mechanism and pipeline, outperforming previous methods in terms of temporal smoothness and/or efficiency.", "keywords": [ "Generative Model", "Video Generation", "Video Translation" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 1, 4 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "Y9yQ9qmVrc", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Chengyu Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Haihong Yang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Huajun Chen", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Qiang Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Renjun Xu", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "Principal Researcher" }, { "name": "Xiaohui Fan", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Xin Shao", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "Postdoc" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 22, "n_ref": 32, "n_ref_all": 48, "n_fig": 12, "n_tab": 1, "L_tab": 986, "n_element_tab": 168, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 1153, "n_element_tab_1": 176, "formula_len_all": 866, "formula_len_all_1": 866, "len_all": 85199, "len_all_1": 54532, "len_abs": 1438, "len_title": 128, "len_sents": 29298, "len_sents_1": 26170, "n_sents": 164, "n_sents_1": 141, "venue": "ICLR2025", "decision": "Reject", "L_title": 112, "L_abs": 1441, "title": "scKGOT: Intercellular Signaling Inference with Knowledge Graph Optimal Transport for Single-cell Transcriptomics", "abs": "Single-cell transcriptomics provides detailed genetic insights into cellular heterogeneity within intact organs and the intercellular signaling that underpins tissue homeostasis, development, and disease. To improve the inference of intercellular signaling and pathway activity, we introduce scKGOT, a novel method that employs the Knowledge Graph Optimal Transport (KGOT) algorithm to model and quantify ligand-receptor-signaling networks between sender and receiver cells. scKGOT defines sender and receiver spaces using pairwise distance matrices from gene expression profiles and leverages prior knowledge from the Ligand-Receptor-Pathway Knowledge Graph (LRP-KG) as initial guidance for transport optimization, allowing for dynamic adaptation based on gene expression data. Through comprehensive benchmarking on public single-cell transcriptomic datasets, scKGOT consistently outperforms existing inference methods in terms of precision and interpretability. Furthermore, we demonstrate its practical applicability across multiple case studies, uncovering complex pathway interactions and revealing insights into cellular heterogeneity in diverse biological contexts. By incorporating scKGOT, we provide a robust and generalizable approach for pathway inference in single-cell analyses, advancing the understanding of intercellular communication mechanisms and offering valuable insights into biological processes at the cellular level.", "keywords": [ "Knowledge Graph", "Optimal Transport", "Cell-cell Communication" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "Y9cVrdYn10", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Chenchen Fu", "gender": "Female", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Chun Jason Xue", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" }, { "name": "Hengkai Zhong", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "Wanyuan Wang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Weiwei Wu", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Full Professor" }, { "name": "Zhuoru Yu", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 21, "n_ref": 45, "n_ref_all": 61, "n_fig": 5, "n_tab": 1, "L_tab": 196, "n_element_tab": 22, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 55, "n_element_tab_1": 2, "formula_len_all": 1079, "formula_len_all_1": 851, "len_all": 84164, "len_all_1": 53441, "len_abs": 1270, "len_title": 131, "len_sents": 29163, "len_sents_1": 26366, "n_sents": 194, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1281, "title": "Delay-Aware Reinforcement Learning: Insights From Delay Distributional Perspective", "abs": "Although deep reinforcement learning (DRL) has achieved great success across various domains, the presence of random delays in real-world scenarios (e.g., remote control) poses a significant challenge to its practicality. Existing delay-aware DRLs mainly focus on state augmentation with historical memory, ensuring that the actions taken are aligned with the true state. However, these approaches still rely on the conventional expected $Q$ value. In contrast, to model delay uncertainty, we aim to go beyond the expected value and propose a distributional DRL to represent the distribution of this $Q$ value. Based on the delay distribution, we further propose a correction mechanism for the distributional $Q$ value, enabling the agent to learn accurate returns in delayed environments. Finally, we apply these techniques to design the delay-aware distributional actor-critic (DADAC) DRL framework, in which the critic is the corrected distributional value function. Experimental results demonstrate that compared to the state-of-the-art delay-aware DRL methods, the proposed DADAC exhibits substantial performance advantages in handling random delays in the MuJoCo continuous control tasks. The corresponding source code is available at https://anonymous.4open.science/r/DADAC.", "keywords": [ "Reinforcement Learning", "Random Delays", "Value Correction", "SAC" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 1, 2, 3, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "Y98ehgkFgI", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Harneet Singh", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Jeong hwan Yoon", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Karl Friston", "gender": "Male", "institution": "University College London", "country": "", "position": "Principal Researcher" }, { "name": "Katherine Hanson", "gender": "unknown", "institution": "ThoughtForge", "country": "US", "position": "Researcher" }, { "name": "Pedro Fontana", "gender": "Male", "institution": "ThoughtForge AI", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 17, "n_ref": 21, "n_ref_all": 30, "n_fig": 6, "n_tab": 2, "L_tab": 411, "n_element_tab": 39, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 409, "n_element_tab_1": 44, "formula_len_all": 120, "formula_len_all_1": 120, "len_all": 65802, "len_all_1": 43774, "len_abs": 1017, "len_title": 111, "len_sents": 25014, "len_sents_1": 20875, "n_sents": 158, "n_sents_1": 131, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1019, "title": "Network-based Active Inference and its Application in Robotics", "abs": "This paper introduces Network-based Active Inference (NetAIF), a novel robotic framework that enables real-time learning and adaptability in dynamic, unstructured environments. NetAIF leverages random attractor dynamics and the Free Energy Principle (FEP) to simplify trajectory generation through network-topology-driven attractors that induce controlled instabilities and probabilistic sampling cycles. This approach allows robots to efficiently adapt to changing conditions without requiring extensive pre-training or pre-calculated trajectories. By integrating learning and control mechanisms within a compact model architecture, NetAIF facilitates seamless task execution, such as target tracking and valve manipulation. Extensive simulations and real-world experiments demonstrate NetAIF's capability to perform rapid and precise real-time adjustments, highlighting its suitability for applications requiring high adaptability and efficient control, such as robotics tasks in the energy and manufacturing sectors.", "keywords": [ "Active Inference (AIF)", "Free Energy Principle (FEP)", "Robotics", "Trajectory generation", "Random dynamical systems", "Random attractor dynamics", "Non-Equilibrium Steady State (NESS)", "Adaptive control", "Industrial automation", "Computational efficiency", "Cost-efficient solutions" ], "rating_list": [ 1, 3, 6, 3 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 1, 1, 3, 3 ], "contribution_list": [ 1, 2, 4, 3 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "Y96R2BOsKm", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Jae Wan Park", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "PhD student" }, { "name": "Junhyeok Kim", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Undergrad student" }, { "name": "Seong Jae Hwang", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Youngjun Jun", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "MS student" }, { "name": "HYUNAHKO", "gender": "Female", "institution": "Yonsei University", "country": "KR", "position": "Undergrad student" } ], "n_formula": 52, "n_formula_1": 20, "n_ref_uni": 51, "n_ref": 102, "n_ref_all": 110, "n_fig": 6, "n_tab": 3, "L_tab": 565, "n_element_tab": 64, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 397, "n_element_tab_1": 51, "formula_len_all": 3079, "formula_len_all_1": 1288, "len_all": 177869, "len_all_1": 66477, "len_abs": 1144, "len_title": 96, "len_sents": 46352, "len_sents_1": 26666, "n_sents": 467, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1154, "title": "BAT: Backbone Augmented Training for Adaptations", "abs": "Adaptations have enabled efficient training for large backbone models such as\ndiffusion models for image generation and transformer-based language models.\nWhile various adaptation techniques aim to maximize performance with minimal\ncomputational resources, limited data often leads to challenges like overfitting,\nmode collapse, or hallucinations. Recently, a promising solution has emerged in\nthe form of augmenting adapter datasets using data originally employed to train\nbackbone models. While this approach has shown potential as a breakthrough, it\noften lacks a solid theoretical foundation or well-defined standards for control-\nlability. To address these limitations, we establish a comprehensive theoretical\nframework for Backbone Augmented Training (BAT). Furthermore, we provide\nboth theoretical and experimental evidence demonstrating that BAT achieves a\nfaster convergence rate to optimal adaptation parameters compared to conven-\ntional adaptation methods. Our results underscore the potential of backbone aug-\nmentation to significantly improve performance, especially when coupled with an\neffective and well-designed data selection schema.", "keywords": [ "Adaptation", "Data Selection", "Parameter Efficient Tuning", "Regularization", "Optimization", "Dreambooth", "LoRA" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "Y93F5eNmZG", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Didier SORNETTE", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Joshua Nielsen", "gender": "Male", "institution": "University of Colorado at Boulder", "country": "", "position": "MS student" }, { "name": "Maziar Raissi", "gender": "unknown", "institution": "University of Colorado at Boulder", "country": "US", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 24, "n_ref": 41, "n_ref_all": 53, "n_fig": 4, "n_tab": 1, "L_tab": 219, "n_element_tab": 16, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1049, "formula_len_all_1": 618, "len_all": 96429, "len_all_1": 57754, "len_abs": 1586, "len_title": 92, "len_sents": 34082, "len_sents_1": 28382, "n_sents": 237, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 115, "L_abs": 1598, "title": "Deep LPPLS: Forecasting of temporal critical points in natural, engineering and financial systems via deep learning", "abs": "The Log-Periodic Power Law Singularity (LPPLS) model offers a general framework for capturing dynamics and predicting transition points in diverse natural and social systems. In this work, we present two calibration techniques for the LPPLS model using deep learning. First, we introduce the Mono-LPPLS-NN (M-LNN) model; for any given empirical time series, a unique M-LNN model is trained and shown to outperform state-of-the-art techniques in estimating the nonlinear parameters $(t_c, m, \\omega)$ of the LPPLS model as evidenced by the comprehensive distribution of parameter errors. Second, we extend the M-LNN model to a more general model architecture, the Poly-LPPLS-NN (P-LNN), which is able to quickly estimate the nonlinear parameters of the LPPLS model for any given time-series of a fixed length, including previously unseen time-series during training. The Poly class of models train on many synthetic LPPLS time-series augmented with various noise structures in a supervised manner. Given enough training examples, the P-LNN models also outperform state-of-the-art techniques for estimating the parameters of the LPPLS model as evidenced by the comprehensive distribution of parameter errors. Additionally, this class of models is shown to substantially reduce the time to obtain parameter estimates. Finally, we present applications to the diagnostic and prediction of two financial bubble peaks (followed by their crash) and of a famous rockslide. These contributions provide a bridge between deep learning and the study of the prediction of transition times in complex time series.", "keywords": [ "Deep Learning", "Time-Series Forecasting", "Complex Systems", "Finite-time singularity", "Log-periodic Power Law Singularity" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "Y8i3rF4Umc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Jian Yang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Jiangwei Weng", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jianjun Qian", "gender": "Male", "institution": "Nanjing University of Science and Techonology", "country": "CN", "position": "Full Professor" }, { "name": "Jun Li", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Wenxuan Fang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Ying Tai", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Ze-dong Zhang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 20, "n_ref_uni": 25, "n_ref": 47, "n_ref_all": 69, "n_fig": 11, "n_tab": 6, "L_tab": 1316, "n_element_tab": 92, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1778, "n_element_tab_1": 291, "formula_len_all": 1755, "formula_len_all_1": 1427, "len_all": 125388, "len_all_1": 62813, "len_abs": 1251, "len_title": 128, "len_sents": 33372, "len_sents_1": 26560, "n_sents": 230, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1255, "title": "Range-Null Latent Prior-guided Consistency Model for Low Light Image Enhancement", "abs": "Low light image enhancement (LLIE) is a challenging task, with most existing models often struggling to adapt to diverse dark environments due to insufficient training datasets. In this paper, we propose a novel unsupervised model called Range-null Latent Prior-guided Consistency Model (RLPCM), which integrates a latent consistency model (LCM) into low light enhancement using Retinex-based range-null space decomposition.RLPCM leverages an off-the-shelf LCM as a generative prior to improve both the latent consistency and realness of enhanced images. Meanwhile, fine-tuning a lighting decoder solely on normal-light images to ensure high fidelity in image space. A key contribution is a simple yet effective global illumination adjustment applied to the range-space component, along with a natural language guidance module to learn the null-space component. This allows for iterative generation to enhance both consistency and realness in just a few steps. Additionally, we present a new UAV low light dataset (UAV-LL) containing 300 image pairs from various UAV scenarios to support comprehensive evaluation. Extensive experiments demonstrate the superior adaptability and effectiveness of our framework across a wide range of low-light environments.", "keywords": [ "Low light enhancement", "Consistency model" ], "rating_list": [ 1, 3, 6, 3, 5 ], "soundness_list": [ 1, 3, 3, 1, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "Y8L5RB4GWb", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Haipeng Liu", "gender": "Not Specified", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "Liang Qifan", "gender": "Male", "institution": "Jilin University", "country": "", "position": "PhD student" }, { "name": "Ting Long", "gender": "Female", "institution": "Jilin University", "country": "CN", "position": "Assistant Professor" }, { "name": "Weinan Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yixiang Shan", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "PhD student" }, { "name": "Yuan Tian", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhengbang Zhu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 22, "n_ref_uni": 40, "n_ref": 66, "n_ref_all": 94, "n_fig": 7, "n_tab": 5, "L_tab": 3796, "n_element_tab": 300, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 642, "n_element_tab_1": 100, "formula_len_all": 1317, "formula_len_all_1": 1278, "len_all": 140857, "len_all_1": 65816, "len_abs": 1401, "len_title": 140, "len_sents": 42712, "len_sents_1": 29410, "n_sents": 310, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1424, "title": "Reconstruction-Guided Policy: Enhancing Decision-Making through Agent-Wise State Consistency", "abs": "An important challenge in multi-agent reinforcement learning is partial observability, where agents cannot access the global state of the environment during execution and can only receive observations within their field of view. To address this issue, previous works typically use the dimensional-wise state, which is obtained by applying MLP or dimensional-based attention on the global state, for decision-making during training and relying on a reconstructed dimensional-wise state during execution. However, dimensional-wise states tend to divert agent attention to specific features, neglecting potential dependencies between agents, making it difficult to make optimal decisions. Moreover, the inconsistency between the states used in training and execution further increases additional errors. To resolve these issues, we propose a method called Reconstruction-Guided Policy (RGP) to reconstruct the agent-wise state, which represents the information of inter-agent relationships, as input for decision-making during both training and execution. This not only preserves the potential dependencies between agents but also ensures consistency between the states used in training and execution. We conducted extensive experiments on both discrete and continuous action environments to evaluate RGP, and the results demonstrates its superior effectiveness. Our code is public in https://anonymous.4open.science/r/RGP-9F79", "keywords": [ "multi-agent reinforcement learning", "partial observability", "cooperation", "centralized training distributed execution", "global state" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 1, 2, 2, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "Y8Kwl7GFAd", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Alex Morehead", "gender": "Male", "institution": "University of Missouri, Columbia", "country": "US", "position": "PhD student" }, { "name": "Arian Rokkum Jamasb", "gender": "unknown", "institution": "Prescient Design / Roche / Genentech", "country": "CH", "position": "Researcher" }, { "name": "Bryan Hooi", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Charles Harris", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Kieran Didi", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "MS student" }, { "name": "Pietro Lio", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Rishabh Anand", "gender": "Male", "institution": "Yale University", "country": "US", "position": "MS student" }, { "name": "Simon V Mathis", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Chaitanya Krishna Joshi", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 16, "n_ref_uni": 40, "n_ref": 88, "n_ref_all": 124, "n_fig": 12, "n_tab": 6, "L_tab": 1061, "n_element_tab": 123, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 138, "n_element_tab_1": 27, "formula_len_all": 1187, "formula_len_all_1": 746, "len_all": 157348, "len_all_1": 64797, "len_abs": 1100, "len_title": 111, "len_sents": 48011, "len_sents_1": 30287, "n_sents": 358, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1118, "title": "RNA FrameFlow: Flow Matching for de novo 3D RNA Backbone Generation", "abs": "We introduce RNA-FrameFlow, the first generative model for de novo 3D RNA backbone design. We build upon $SE(3)$ flow matching for protein backbone generation and establish protocols for data preparation and evaluation to address unique challenges posed by RNA modeling. We formulate RNA structures as a set of rigid-body frames and associated loss functions which account for larger, more conformationally flexible RNA backbones (13 atoms per nucleotide) vs. proteins (4 atoms per residue). Toward tackling the lack of diversity in 3D RNA datasets, we explore training with structural clustering and cropping augmentations. Additionally, we define a suite of evaluation metrics to measure whether the generated RNA structures are globally self-consistent (via inverse folding followed by forward folding) and locally recover RNA-specific structural descriptors. The most performant version of RNA-FrameFlow generates locally realistic RNA backbones of 40-150 nucleotides, over 40\\% of which pass our validity criteria as measured by a self-consistency TM-score $\\geq0.45$, at which two RNAs have the same global fold.", "keywords": [ "RNA Structure", "RNA Design", "Flow Matching" ], "rating_list": [ 5, 5, 6, 6, 6 ], "soundness_list": [ 3, 2, 4, 3, 2 ], "presentation_list": [ 3, 3, 4, 4, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3, 5 ] }, { "paper_id": "Y8KK9kjgIK", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Barbora Barancikova", "gender": "Female", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Cristopher Salvi", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Assistant Professor" }, { "name": "Zhuoyue Huang", "gender": "Female", "institution": "Oxofrd, University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 106, "n_formula_1": 21, "n_ref_uni": 64, "n_ref": 126, "n_ref_all": 185, "n_fig": 16, "n_tab": 7, "L_tab": 3604, "n_element_tab": 303, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 1094, "n_element_tab_1": 94, "formula_len_all": 9915, "formula_len_all_1": 1570, "len_all": 241356, "len_all_1": 64558, "len_abs": 1015, "len_title": 136, "len_sents": 67900, "len_sents_1": 28338, "n_sents": 602, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 88, "L_abs": 1028, "title": "SigDiffusions: Score-Based Diffusion Models for Time Series via Log-Signature Embeddings", "abs": "Score-based diffusion models have recently emerged as state-of-the-art generative models for a variety of data modalities. Nonetheless, it remains unclear how to adapt these models to generate long multivariate time series. Viewing a time series as the discretization of an underlying continuous process, we introduce SigDiffusion, a novel diffusion model operating on log-signature embeddings of the data. The forward and backward processes gradually perturb and denoise log-signatures preserving their algebraic structure. To recover a signal from its log-signature, we provide new closed-form inversion formulae expressing the coefficients obtained by expanding the signal in a given basis (e.g. Fourier or orthogonal polynomials) as explicit polynomial functions of the log-signature. Finally, we show that combining \\texttt{SigDiffusion} with these inversion formulae results in highly realistic time series generation, competitive with the current state-of-the-art on various datasets of synthetic and real-world examples.", "keywords": [ "diffusion models", "path signatures", "time series" ], "rating_list": [ 1, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 1, 3, 3 ], "confidence_list": [ 5, 3, 3 ] }, { "paper_id": "Y89o3LAEHX", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Duanyu Feng", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "PhD student" }, { "name": "HONGYU DU", "gender": "Male", "institution": "University of Colorado Anschutz Medical Campus", "country": "US", "position": "PhD student" }, { "name": "Hao Wang", "gender": "unknown", "institution": "Sichuan University", "country": "", "position": "Associate Professor" }, { "name": "Ronghui han", "gender": "unknown", "institution": "Sichuan University", "country": "CN", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 27, "n_ref": 72, "n_ref_all": 105, "n_fig": 5, "n_tab": 11, "L_tab": 10378, "n_element_tab": 755, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 4825, "n_element_tab_1": 261, "formula_len_all": 536, "formula_len_all_1": 536, "len_all": 124766, "len_all_1": 64349, "len_abs": 1513, "len_title": 167, "len_sents": 33916, "len_sents_1": 27353, "n_sents": 238, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 118, "L_abs": 1521, "title": "A Hybrid Loss Framework for Decomposition-based Time Series Forecasting Methods: Balancing Global and Component Errors", "abs": "Accurate time series forecasting, predicting future values based on past data, is crucial for diverse industries. Many current time series methods decompose time series into multiple sub-series, applying different model architectures and training with an end-to-end overall loss for forecasting. However, this raises a question: does this overall loss prioritize the importance of critical sub-series within the decomposition for the better performance? To investigate this, we conduct a study on the impact of overall loss on existing time series methods with sequence decomposition. Our findings reveal that overall loss may introduce bias in model learning, hindering the learning of the prioritization of more significant sub-series and limiting the forecasting performance. To address this, we propose a hybrid loss framework combining the global and component errors. This framework introduces component losses for each sub-series alongside the original overall loss. It employs a dual min-max algorithm to dynamically adjust weights between the overall loss and component losses, and within component losses. This enables the model to achieve better performance of current time series methods by focusing on more critical sub-series while still maintaining a low overall loss. We integrate our loss framework into several time series methods and evaluate the performance on multiple datasets. Results show an average improvement of 0.5-2% over existing methods without any modifications to the model architectures.", "keywords": [ "time series forecasting", "series decomposition", "hybrid loss framework" ], "rating_list": [ 3, 3, 1, 1 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 5, 3, 5, 5 ] }, { "paper_id": "Y7slJZPGCy", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Ioannis Mavrothalassitis", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "MS student" }, { "name": "Thomas Pethick", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Volkan Cevher", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" } ], "n_formula": 92, "n_formula_1": 29, "n_ref_uni": 30, "n_ref": 112, "n_ref_all": 138, "n_fig": 2, "n_tab": 1, "L_tab": 200, "n_element_tab": 8, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 88, "n_element_tab_1": 4, "formula_len_all": 7765, "formula_len_all_1": 2940, "len_all": 179788, "len_all_1": 70153, "len_abs": 1114, "len_title": 129, "len_sents": 52796, "len_sents_1": 28192, "n_sents": 564, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1211, "title": "Efficient Interpolation between Extragradient and Proximal Methods for Weak MVIs", "abs": "We study nonmonotone games satisfying the weak Minty variational inequality (MVI) with parameter $\\rho \\in (-\\tfrac{1}{L}, \\infty)$, where $L$ is the Lipschitz constant of the gradient operator. An error corrected version of the inexact proximal point algorithm is proposed, with which we establish the first $\\mathcal O(1/\\epsilon)$ rate for the entire range $\\rho \\in (-\\tfrac{1}{L}, \\infty)$, thus removing a logarithmic factor compared with the complexity of existing methods. The scheme automatically selects the needed accuracy for the proximal computation, and can recover the relaxed extragradient method when $\\rho > -\\tfrac{1}{2L}$ and the relaxed proximal point algorithm (rPPA) when $\\rho > -\\tfrac{1}{L}$. Due to the error correction, the scheme inherits the strong properties of the _exact_ rPPA. Specifically, we show that linear convergence is automatically achieved under appropriate conditions. Tightness for the range of $\\rho$ is established through a lower bound for rPPA. Central to the algorithmic construction is a halfspace projection, where the key insight is that the allowed error tolerance can both be used to correct for the proximal approximation and to enlarge the problem class.", "keywords": [ "Weak Minty variational inequalities", "cohypomonotone", "nonmonotone", "first-order methods", "extragradient method", "proximal point algorithm" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "Y7lc4aZ4iP", "primary_area": "learning theory", "n_author": 1, "author_info_list": [ { "name": "Lan Vinh Truong", "gender": "Male", "institution": "University of Essex", "country": "GB", "position": "Lecturer" } ], "n_formula": 96, "n_formula_1": 85, "n_ref_uni": 22, "n_ref": 40, "n_ref_all": 42, "n_fig": 0, "n_tab": 2, "L_tab": 237, "n_element_tab": 16, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5649, "formula_len_all_1": 5327, "len_all": 70547, "len_all_1": 44809, "len_abs": 774, "len_title": 119, "len_sents": 15939, "len_sents_1": 14270, "n_sents": 175, "n_sents_1": 153, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 778, "title": "On Rademacher Complexity-based Generalization Bounds for Deep Learning", "abs": "We show that the Rademacher complexity-based approach can generate non-vacuous generalisation bounds on Convolutional Neural Networks (CNNs) for classifying a small number of classes of images. The development of new contraction lemmas for high-dimensional mappings between vector spaces for general Lipschitz activation functions is a key technical contribution. These lemmas extend and improve the Talagrand contraction lemma in a variety of cases. Our generalisation bounds are based on the infinity norm of the weight matrices, distinguishing them from previous works that relied on different norms. Furthermore, while prior works that use the Rademacher complexity-based approach primarily focus on ReLU DNNs, our results extend to a broader class of activation functions.", "keywords": [ "Deep Learning", "CNN", "Generalisation Errors" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 1, 2 ], "presentation_list": [ 3, 1, 1 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "Y7jJN0VQ4y", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Di Huang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Guodong Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Jinjin Zhang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Ruijie Yang", "gender": "Male", "institution": "Beihang University", "country": "", "position": "PhD student" }, { "name": "Zhen Yang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 9, "n_ref_uni": 38, "n_ref": 69, "n_ref_all": 76, "n_fig": 3, "n_tab": 13, "L_tab": 5443, "n_element_tab": 503, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 2527, "n_element_tab_1": 149, "formula_len_all": 1647, "formula_len_all_1": 505, "len_all": 177568, "len_all_1": 64141, "len_abs": 1210, "len_title": 137, "len_sents": 52012, "len_sents_1": 28649, "n_sents": 401, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1213, "title": "Anomalies are Streaming: Continual Learning for Weakly Supervised Video Anomaly Detection", "abs": "Weakly supervised video anomaly detection (WSVAD) aims to locate frame-level anomalies with only video-level annotations provided. However, existing WSVAD methods struggle to adapt to real-world scenarios, where unseen anomalies are continuously introduced, thereby making the training of WSVAD essentially a process of continual learning. In this paper, we pioneer to explore the continual learning for weakly supervised video anomaly detection (CL-WSVAD), seeking to mitigate the catastrophic forgetting when the detection model learns new anomalies. We propose normality representation pre-training prior to continual learning, utilizing potential anomaly texts to guide the model in learning robust normality representations, which improves discrimination from potential incremental anomalies. Additionally, we introduce a mixed-up cross-modal alignment method to assist in adapting the pretrained model on CL-WSVAD. Subsequently, we propose a continual learning framework based on sequentially retaining the learnable text prompts for each type of anomaly, which effectively mitigates catastrophic forgetting. Experiments on our established CL-WSVAD benchmarks demonstrate the superiority of proposed method.", "keywords": [ "Weakly supervised video anomaly detection", "continual learning" ], "rating_list": [ 5, 5, 6, 8, 5 ], "soundness_list": [ 4, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 4, 2 ], "confidence_list": [ 5, 5, 4, 4, 5 ] }, { "paper_id": "Y71rktd4oT", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Doina Precup", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Research Team Lead" }, { "name": "Gabriela Moisescu-Pareja", "gender": "Female", "institution": "McGill University, McGill University", "country": "CA", "position": "MS student" }, { "name": "Gavin McCracken", "gender": "Not Specified", "institution": "McGill University, McGill University", "country": "CA", "position": "PhD student" }, { "name": "Jonathan Love", "gender": "unknown", "institution": "Leiden University, Dept. of Mathematics, Leiden University", "country": "NL", "position": "Postdoc" }, { "name": "Vincent L\u00e9tourneau", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Postdoc" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 18, "n_ref": 51, "n_ref_all": 96, "n_fig": 24, "n_tab": 2, "L_tab": 311, "n_element_tab": 51, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 100, "n_element_tab_1": 11, "formula_len_all": 1138, "formula_len_all_1": 617, "len_all": 134699, "len_all_1": 67173, "len_abs": 1200, "len_title": 138, "len_sents": 49106, "len_sents_1": 31880, "n_sents": 366, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1225, "title": "Interpreting neural networks depends on the level of abstraction: Revisiting modular addition", "abs": "Prior work in mechanistic interpretability has analyzed how neural networks solve modular arithmetic tasks, but conflicting interpretations have emerged, questioning the universality hypothesis\u2014that similar tasks lead to similar learned circuits. Revisiting modular addition, we identify that these discrepancies stem from overly granular analyses, which obscure the higher-level patterns that unify seemingly disparate solutions. Using a multi-scale approach\u2014microscopic (neurons), mesoscopic (clusters of neurons), and macroscopic (entire network)\u2014we show that all scales align on (approximate) cosets and implement an abstract algorithm resembling the approximate Chinese Remainder Theorem. Additionally, we propose a model where networks aims for a constant logit margin, predicting $\\mathcal{O}(\\log(n))$ frequencies\u2014more consistent with empirical results in networks with biases, which are more expressive and commonly used in practice, than the $\\frac{n-1}{2}$ frequencies derived from bias-free networks. By uncovering shared structures across setups, our work provides a unified framework for understanding modular arithmetic in neural networks and generalizes existing insights to broader, more realistic scenarios.", "keywords": [ "mechanistic interpretability", "group theory", "universality" ], "rating_list": [ 6, 5, 5, 3, 5 ], "soundness_list": [ 4, 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 1, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3, 4 ] }, { "paper_id": "Y6aHdDNQYD", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Junjie Meng", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Mahsa Baktashmotlagh", "gender": "Female", "institution": "The University of Queensland", "country": "AU", "position": "Associate Professor" }, { "name": "Yadan Luo", "gender": "Female", "institution": "The University of Queensland", "country": "AU", "position": "Associate Professor" }, { "name": "Yonggang Zhang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Postdoc" }, { "name": "Zhuoxiao Chen", "gender": "Male", "institution": "University of Queensland", "country": "AU", "position": "PhD student" }, { "name": "Zi Huang", "gender": "Female", "institution": "University of Queensland", "country": "AU", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 15, "n_ref_uni": 66, "n_ref": 140, "n_ref_all": 180, "n_fig": 8, "n_tab": 8, "L_tab": 7302, "n_element_tab": 305, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2017, "n_element_tab_1": 151, "formula_len_all": 663, "formula_len_all_1": 664, "len_all": 222643, "len_all_1": 71392, "len_abs": 1825, "len_title": 126, "len_sents": 52912, "len_sents_1": 32726, "n_sents": 358, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 78, "L_abs": 1835, "title": "MOS: Model Synergy for Test-Time Adaptation on LiDAR-Based 3D Object Detection", "abs": "LiDAR-based 3D object detection is crucial for various applications but often experiences performance degradation in real-world deployments due to domain shifts. While most studies focus on cross-dataset shifts, such as changes in environments and object geometries, practical corruptions from sensor variations and weather conditions remain underexplored. In this work, we propose a novel online test-time adaptation framework for 3D detectors that effectively tackles these shifts, including a challenging $\\textit{cross-corruption}$ scenario where cross-dataset shifts and corruptions co-occur. By leveraging long-term knowledge from previous test batches, our approach mitigates catastrophic forgetting and adapts effectively to diverse shifts. Specifically, we propose a Model Synergy (MOS) strategy that dynamically selects historical checkpoints with diverse knowledge and assembles them to best accommodate the current test batch. This assembly is directed by our proposed Synergy Weights (SW), which perform a weighted averaging of the selected checkpoints, minimizing redundancy in the composite model. The SWs are computed by evaluating the similarity of predicted bounding boxes on the test data and the independence of features between checkpoint pairs in the model bank. To maintain an efficient and informative model bank, we discard checkpoints with the lowest average SW scores, replacing them with newly updated models. Our method was rigorously tested against existing test-time adaptation strategies across three datasets and eight types of corruptions, demonstrating superior adaptability to dynamic scenes and conditions. Notably, it achieved a 67.3% improvement in a challenging cross-corruption scenario, offering a more comprehensive benchmark for adaptation. Source code: https://github.com/zhuoxiao-chen/MOS.", "keywords": [ "Test-Time Adaptation", "3D Object Detection" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 3, 2, 4 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "Y6LPWBo2HP", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Dmitriy Smirnov", "gender": "Male", "institution": "Netflix", "country": "US", "position": "Researcher" }, { "name": "Lingxiao Li", "gender": "Male", "institution": "Netflix", "country": "US", "position": "Researcher" }, { "name": "Mohammad H. Taghavi", "gender": "Male", "institution": "NetFlix", "country": "US", "position": "Researcher" }, { "name": "Ning Yu", "gender": "Male", "institution": "Netflix Eyeline Studios", "country": "US", "position": "Researcher" }, { "name": "Ryan D Burgert", "gender": "Non-Binary", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Vincent Dedun", "gender": "Male", "institution": "NetFlix", "country": "US", "position": "Researcher" }, { "name": "Winnie Lin", "gender": "Female", "institution": "NetFlix", "country": "US", "position": "Researcher" }, { "name": "Yitong Deng", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 18, "n_ref_uni": 33, "n_ref": 66, "n_ref_all": 89, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2923, "formula_len_all_1": 1938, "len_all": 127865, "len_all_1": 59795, "len_abs": 1183, "len_title": 111, "len_sents": 37502, "len_sents_1": 27523, "n_sents": 287, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1186, "title": "Infinite-Resolution Integral Noise Warping for Diffusion Models", "abs": "Adapting pretrained image-based diffusion models to generate temporally consistent videos has become an impactful generative modeling research direction. Training-free noise-space manipulation has proven to be an effective technique, where the challenge is to preserve the Gaussian white noise distribution while adding in temporal consistency. Recently, Chang et al. (2024) formulated this problem using an integral noise representation with distribution-preserving guarantees, and proposed an upsampling-based algorithm to compute it. However, while their mathematical formulation is advantageous, the algorithm incurs a high computational cost. Through analyzing the limiting-case behavior of their algorithm as the upsampling resolution goes to infinity, we develop an alternative algorithm that, by gathering increments of multiple Brownian bridges, achieves their infinite-resolution accuracy while simultaneously reducing the computational cost by orders of magnitude. We prove and experimentally validate our theoretical claims, and demonstrate our method's effectiveness in real-world applications. We further show that our method can readily extend to the 3-dimensional space.", "keywords": [ "diffusion models; video generation; temporal consistency; noise warping; white Gaussian noise" ], "rating_list": [ 6, 8, 8, 3 ], "soundness_list": [ 4, 4, 4, 3 ], "presentation_list": [ 4, 2, 4, 4 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "Y6KUBkUimC", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Anil Kag", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Huseyin Coskun", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Jian Ren", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Sergey Tulyakov", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Director of Research" }, { "name": "Shyamgopal Karthik", "gender": "unknown", "institution": "University of T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Zeynep Akata", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 12, "n_ref_uni": 60, "n_ref": 169, "n_ref_all": 184, "n_fig": 8, "n_tab": 11, "L_tab": 2852, "n_element_tab": 453, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 946, "n_element_tab_1": 178, "formula_len_all": 1789, "formula_len_all_1": 766, "len_all": 218204, "len_all_1": 64268, "len_abs": 552, "len_title": 116, "len_sents": 55948, "len_sents_1": 27772, "n_sents": 429, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1485, "title": "Scalable Ranked Preference Optimization for Text-to-Image Generation", "abs": "Direct Preference Optimization (DPO) has emerged as a powerful approach to align text-to-image (T2I) models with human feedback. Unfortunately, successful application of DPO to T2I models requires a huge amount of resources to collect and label large-scale datasets, e.g., millions of generated paired images annotated with human preferences. In addition, these human preference datasets can get outdated quickly as the rapid improvements of T2I models lead to higher quality images. In this work, we investigate a scalable approach for collecting large-scale and fully synthetic datasets for DPO training. Specifically, the preferences for paired images are generated using a pre-trained reward function, eliminating the need for involving humans in the annotation process, greatly improving the dataset collection efficiency. Moreover, we demonstrate that such datasets allow averaging predictions across multiple models and collecting ranked preferences as opposed to pairwise preferences. Furthermore, we introduce RankDPO to enhance DPO-based methods using the ranking feedback. Applying RankDPO on SDXL and SD3-Medium models with our synthetically generated preference dataset ``Syn-Pic'' improves both prompt-following (on benchmarks like T2I-Compbench, GenEval, and DPG-Bench) and visual quality (through user studies). This pipeline presents a practical and scalable solution to develop better preference datasets to enhance the performance and safety of text-to-image models.", "keywords": [ "text-to-image generation", "Direct Preference Optimization", "Learning from AI Feedback" ], "rating_list": [ 5, 5, 8, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Y5mm3Yb36I", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Adi Haviv", "gender": "Female", "institution": "School of Computer Science, Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Amit Haim Bermano", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Associate Professor" }, { "name": "Niva Elkin-Koren", "gender": "Female", "institution": "Tel Aviv University", "country": "IL", "position": "Full Professor" }, { "name": "Roi Livni", "gender": "Not Specified", "institution": "Tel Aviv University", "country": "IL", "position": "Associate Professor" }, { "name": "Shahar Sarfaty", "gender": "Male", "institution": "School of Computer Science, Tel Aviv University", "country": "IL", "position": "MS student" }, { "name": "Uri Y. Hacohen", "gender": "Male", "institution": "Tel Aviv University, Tel Aviv University", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 27, "n_ref": 44, "n_ref_all": 60, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 89852, "len_all_1": 55686, "len_abs": 341, "len_title": 134, "len_sents": 30981, "len_sents_1": 26798, "n_sents": 224, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1249, "title": "Not Every Image is Worth a Thousand Words: Quantifying Originality in Stable Diffusion", "abs": "This work addresses the challenge of quantifying originality in text-to-image (T2I) generative diffusion models, with a focus on copyright originality. We begin by evaluating T2I models' ability to innovate and generalize through controlled experiments, revealing that stable diffusion models can effectively recreate unseen elements with sufficiently diverse training data. Then, our key insight is that concepts and combinations of image elements the model is familiar with, and saw more during training, are more concisly represented in the model's latent space. We hence propose a method that leverages textual inversion to measure the originality of an image based on the number of tokens required for its reconstruction by the model. Our approach is inspired by legal definitions of originality and aims to assess whether a model can produce original content without relying on specific prompts or having the training data of the model. We demonstrate our method using both a pre-trained stable diffusion model and a synthetic dataset, showing a correlation between the number of tokens and image originality. This work contributes to the understanding of originality in generative models and has implications for copyright infringement cases.", "keywords": [ "Interpretability", "Computer Vision", "Text-to-Image(T2I) Models", "Stable Diffusion", "Generalization", "Copyrights Infringement" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "Y5de4fkuHR", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Jinfeng Bai", "gender": "Male", "institution": "TAL", "country": "CN", "position": "Researcher" }, { "name": "Ming Liu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Postdoc" }, { "name": "Wangmeng Zuo", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yiwen Guo", "gender": "Male", "institution": "ByteDance", "country": "CN", "position": "Researcher" }, { "name": "Zhilong Ji", "gender": "Male", "institution": "Tomorrow Advancing Life", "country": "CN", "position": "Researcher" }, { "name": "Zixian Guo", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 41, "n_ref": 79, "n_ref_all": 93, "n_fig": 5, "n_tab": 6, "L_tab": 5039, "n_element_tab": 501, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 253, "n_element_tab_1": 22, "formula_len_all": 79, "formula_len_all_1": 77, "len_all": 174143, "len_all_1": 62641, "len_abs": 1468, "len_title": 132, "len_sents": 36758, "len_sents_1": 30322, "n_sents": 271, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1499, "title": "LLM as a Complementary Optimizer to Gradient Descent: A Case Study in Prompt Tuning", "abs": "Mastering a skill generally relies on both hands-on experience from doers and insightful, high-level guidance by mentors.\n Will this strategy also work well for solving complex non-convex optimization problems? Here, a common gradient-based optimizer acts like a disciplined doer, making locally optimal updates at each step.\n Large Language Models (LLMs) can also search for better solutions by inferring from natural language instructions, akin to a high-level mentor. \n In this paper, we show that these two participators are complementary to each other and can effectively collaborate as a combined optimization framework.\n The collaborative optimization is achieved by alternating between the gradient-based and LLM-based optimizers.\n We instruct LLMs to generate possibly improved solutions by taking parameter trajectories recorded during the previous stage of gradient-based optimization into account. Inferred results of LLMs are used as restarting points for the next stage of gradient optimization. \n We verify the effectiveness of this optimization framework on prompt tuning.\n By leveraging both the locally rigorous gradient-based optimizer and the high-level deductive LLM-based optimizer, the combined optimization method consistently yields improvements over competitive baselines on a variety of tasks.\n Our results demonstrate the synergistic effect of conventional gradient-based optimization and the inference ability of LLMs. \n The code will be made publicly available.", "keywords": [ "Model Adaptation", "Prompt Optimization" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "Y5LjYI4N6P", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Abhishek Panigrahi", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Kaifeng Lyu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Nikunj Saunshi", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Sanjiv Kumar", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Sashank J. Reddi", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Satyen Kale", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Sobhan Miryoosefi", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" } ], "n_formula": 94, "n_formula_1": 8, "n_ref_uni": 77, "n_ref": 124, "n_ref_all": 167, "n_fig": 7, "n_tab": 9, "L_tab": 3257, "n_element_tab": 382, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2086, "n_element_tab_1": 184, "formula_len_all": 9134, "formula_len_all_1": 536, "len_all": 313331, "len_all_1": 75558, "len_abs": 1864, "len_title": 60, "len_sents": 86015, "len_sents_1": 33812, "n_sents": 795, "n_sents_1": 275, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1874, "title": "Efficient stagewise pretraining via progressive subnetworks", "abs": "Recent developments in large language models have sparked interest in efficient\npretraining methods. Stagewise training approaches to improve efficiency, like\ngradual stacking and layer dropping (Reddi et al., 2023; Zhang & He, 2020), have\nrecently garnered attention. The prevailing view suggests that stagewise dropping\nstrategies, such as layer dropping, are ineffective, especially when compared to\nstacking-based approaches. This paper challenges this notion by demonstrating\nthat, with proper design, dropping strategies can be competitive, if not better, than\nstacking methods. Specifically, we develop a principled stagewise training framework, progressive subnetwork training, which only trains subnetworks within the\nmodel and progressively increases the size of subnetworks during training, until it\ntrains the full network. We propose an instantiation of this framework \u2014 Random\nPart Training (RAPTR) \u2014 that selects and trains only a random subnetwork (e.g.\ndepth-wise, width-wise) of the network at each step, progressively increasing the\nsize in stages. We show that this approach not only generalizes prior works like\nlayer dropping but also fixes their key issues. Furthermore, we establish a theoretical basis for such approaches and provide justification for (a) increasing complexity of subnetworks in stages, conceptually diverging from prior works on layer\ndropping, and (b) stability in loss across stage transitions in presence of key modern architecture components like residual connections and layer norms. Through\ncomprehensive experiments, we demonstrate that RAPTR can significantly speed\nup training of standard benchmarks like BERT and UL2, up to 33% compared to\nstandard training and, surprisingly, also shows better downstream performance on\nUL2, improving QA tasks and SuperGLUE by 1.5%; thereby, providing evidence\nof better inductive bias.", "keywords": [ "Efficient stagewise training", "modular training", "language model pretraining", "implicit bias", "simple-to-complex learning" ], "rating_list": [ 5, 8, 6, 8 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "Y4kJp8GQmV", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Mengfei Xia", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Nan Xue", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Ran Yi", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Tieliang Gong", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yong-jin Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yujun Shen", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 88, "n_formula_1": 48, "n_ref_uni": 21, "n_ref": 59, "n_ref_all": 92, "n_fig": 2, "n_tab": 8, "L_tab": 2078, "n_element_tab": 358, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2018, "n_element_tab_1": 355, "formula_len_all": 9818, "formula_len_all_1": 3019, "len_all": 122508, "len_all_1": 68112, "len_abs": 1326, "len_title": 103, "len_sents": 32657, "len_sents_1": 25745, "n_sents": 306, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1344, "title": "Rectified Diffusion Guidance for Conditional Generation", "abs": "Classifier-Free Guidance (CFG), which combines the conditional and unconditional score functions with two coefficients summing to one, serves as a practical technique for diffusion model sampling. Theoretically, however, denoising with CFG *cannot* be expressed as a reciprocal diffusion process, which may consequently leave some hidden risks during use. In this work, we revisit the theory behind CFG and rigorously confirm that the improper configuration of the combination coefficients (*i.e.*, the widely used summing-to-one version) brings about expectation shift of the generative distribution. To rectify this issue, we propose ReCFG with a relaxation on the guidance coefficients such that denoising with ReCFG strictly aligns with the diffusion theory. We further show that our approach enjoys a ***closed-form*** solution given the guidance strength. That way, the rectified coefficients can be readily pre-computed via traversing the observed data, leaving the sampling speed barely affected. Empirical evidence on real-world data demonstrate the compatibility of our post-hoc design with existing state-of-the-art diffusion models, including both class-conditioned ones (*e.g.*, EDM2 on ImageNet) and text-conditioned ones (*e.g.*, SD3 on CC12M), without any retraining. We will open-source the code to facilitate further research.", "keywords": [ "Diffusion Models", "Conditional Generation", "Classifier-Free Guidance" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 2, 2 ] }, { "paper_id": "Y4iaDU4yMi", "primary_area": "applications to robotics, autonomy, planning", "n_author": 10, "author_info_list": [ { "name": "Bei Li", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Jiahao Liu", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Jingang Wang", "gender": "Male", "institution": "Meituan", "country": "", "position": "Researcher" }, { "name": "Kechen Jiao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Xinyu Liu", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "MS student" }, { "name": "Xiu Li", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Yifan Zhu", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhirui Fang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Zhongjian Qiao", "gender": "unknown", "institution": "Tsinghua Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "\u5f90\u6daf\u6615", "gender": "unknown", "institution": "Wuhan University", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 35, "n_ref": 52, "n_ref_all": 72, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1121, "formula_len_all_1": 1120, "len_all": 120344, "len_all_1": 63580, "len_abs": 1716, "len_title": 135, "len_sents": 36282, "len_sents_1": 32234, "n_sents": 245, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1731, "title": "InteractiveCOT: Aligning Dynamic Chain-of-Thought Planning for Embodied Decision-Making", "abs": "Vision-Language Models (VLMs) are increasingly being employed as the decision-making \"brains\" of embodied agents. Effectively harnessing their powerful generalization capabilities in dynamic, context-specific tasks remains a significant challenge. Chain-of-Thought (CoT) prompting is often utilized for complex task execution, but existing methods either rely on static strategies that fail to adapt to changing environments or fine-tune on offline datasets, which are insufficient for optimizing agent decision-making through interaction.\nIn this paper, we propose a novel approach that focuses on optimizing the CoT reasoning process rather than just the final action tokens. By aligning the CoT process through preference-based reinforcement learning, specifically Direct Preference Optimization (DPO), we enhance the agent's ability to make accurate decisions in dynamic environments while mitigating model degradation during fine-tuning. Our method models the environment as a Markov decision process, requiring the agent to reflect on the current state in real time to generate adaptive plans and actions.\nBy prioritizing the optimization of the CoT process over the final actions, we enhance the agent's reasoning adaptability while effectively mitigating model degradation during fine-tuning.\nExperiments in the ALFWorld environment demonstrate an average success rate of \\textbf 26.67%, which is a 6\\% improvement over RL4VLM, and show that our method effectively mitigates model degradation post fine-tuning. These results highlight the potential of integrating preference-based reinforcement learning techniques with CoT processes to enhance the decision-making capabilities of vision-language models in embodied agents.", "keywords": [ "Embodied Agent", "Multi-modal Large Model", "Chain-of-Thought", "Planning", "Reinforcement Learning", "Preference Learning" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "Y4aWwRh25b", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Eric P. Xing", "gender": "Male", "institution": "Mohamed bin Zayed Univeristy of AI", "country": "AE", "position": "Full Professor" }, { "name": "Hanlin Zhang", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Sham M. Kakade", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Zhenting Qi", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "MS student" }, { "name": "Hima Lakkaraju", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 61, "n_ref": 104, "n_ref_all": 132, "n_fig": 10, "n_tab": 6, "L_tab": 749, "n_element_tab": 55, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 963, "n_element_tab_1": 58, "formula_len_all": 177, "formula_len_all_1": 130, "len_all": 191824, "len_all_1": 67856, "len_abs": 1354, "len_title": 159, "len_sents": 49083, "len_sents_1": 32594, "n_sents": 373, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 111, "L_abs": 1363, "title": "Follow My Instruction and Spill the Beans: Scalable Data Extraction from Retrieval-Augmented Generation Systems", "abs": "Retrieval-Augmented Generation (RAG) improves pre-trained models by incorporating external knowledge at test time to enable customized adaptation. \nWe study the risk of datastore leakage in Retrieval-In-Context RAG Language Models (LMs). We show that an adversary can exploit LMs' instruction-following capabilities to easily extract text data verbatim from the datastore of RAG systems built with instruction-tuned LMs via prompt injection. \nThe vulnerability exists for a wide range of modern LMs that span Llama2, Mistral/Mixtral, Vicuna, SOLAR, WizardLM, Qwen1.5, and Platypus2, and the exploitability exacerbates as the model size scales up. \nWe also study multiple effects of RAG setup on the extractability of data, indicating that following unexpected instructions to regurgitate data can be an outcome of failure in effectively utilizing contexts for modern LMs, and further show that such vulnerability can be greatly mitigated by position bias elimination strategies. \nExtending our study to production RAG models GPTs, we design an attack that can cause datastore leakage with a 100\\% success rate on 25 randomly selected customized GPTs with at most 2 queries, and we extract text data verbatim at a rate of 41\\% from a book of 77,000 words and 3\\% from a corpus of 1,569,000 words by prompting the GPTs with only 100 queries generated by themselves.", "keywords": [ "Retrieval-Augmented Generation", "Security", "Privacy" ], "rating_list": [ 8, 6, 8, 5 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 4, 2, 3, 5 ] }, { "paper_id": "Y4UliyX3LE", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Evgeniia Tokarchuk", "gender": "Female", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Hua Chang Bakker", "gender": "unknown", "institution": "University of Amsterdam", "country": "NL", "position": "MS student" }, { "name": "Vlad Niculae", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "Assistant Professor" } ], "n_formula": 29, "n_formula_1": 22, "n_ref_uni": 47, "n_ref": 125, "n_ref_all": 147, "n_fig": 5, "n_tab": 3, "L_tab": 1456, "n_element_tab": 123, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 927, "n_element_tab_1": 81, "formula_len_all": 1880, "formula_len_all_1": 1105, "len_all": 146636, "len_all_1": 63191, "len_abs": 1289, "len_title": 93, "len_sents": 35607, "len_sents_1": 26948, "n_sents": 308, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1328, "title": "Learning Dispersed Embeddings on Hyperspheres", "abs": "Learning well-separated features in high-dimensional spaces, such as text or image $\\textit{embeddings}$, is crucial for many machine learning applications. Achieving such separation can be effectively accomplished through the $\\textit{dispersion}$ of embeddings, where unrelated vectors are pushed apart as much as possible. By constraining features to be on a $\\textit{hypersphere}$, we can connect dispersion to well-studied problems in mathematics and physics, where optimal solutions are known for limited low-dimensional cases. However, in representation learning we typically deal with a large number of features in high-dimensional space, which makes leveraging existing theoretical and numerical solutions impossible. Therefore, we rely on gradient-based methods to approximate the optimal dispersion on a hypersphere. In this work, we first give an overview of existing methods from disconnected literature. Next, we propose new reinterpretations of known methods, namely Maximum Mean Discrepancy (MMD) and Lloyd\u2019s relaxation algorithm. Finally, we derive a novel dispersion method that directly exploits properties of the hypersphere. Our experiments show the importance of dispersion in image classification and natural language processing tasks, and how algorithms exhibit different trade-offs in different regimes.", "keywords": [ "embeddings", "dispersion", "hypersphere", "representation learning", "separation" ], "rating_list": [ 5, 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 3, 2, 3, 3, 2 ] }, { "paper_id": "Y4GCrfAidr", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Adam Shabti Charles", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "Assistant Professor" }, { "name": "Gal Mishne", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Jiancheng Xie", "gender": "Male", "institution": "Boston University", "country": "US", "position": "PhD student" }, { "name": "Lou C Voinov", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Noga Mudrik", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 40, "n_ref": 72, "n_ref_all": 87, "n_fig": 13, "n_tab": 4, "L_tab": 1476, "n_element_tab": 97, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 1420, "n_element_tab_1": 15, "formula_len_all": 824, "formula_len_all_1": 765, "len_all": 149441, "len_all_1": 67119, "len_abs": 1265, "len_title": 128, "len_sents": 46251, "len_sents_1": 33363, "n_sents": 335, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1191, "title": "Multiway Multislice PHATE: Visualizing Hidden Dynamics of RNNs through Training", "abs": "Recurrent neural networks (RNNs) are a widely used tool for sequential data analysis, however, they are still often seen as black boxes of computation. Understanding the functional principles of these networks is key to developing ideal model architectures and optimization strategies. Previous studies often only emphasize the networks' representation post-training, overlooking their evolution process. Here, we present Multiway Multislice PHATE (MM-PHATE), a novel method for visualizing the evolution of RNNs' hidden states. MM-PHATE is a graph-based embedding using structured kernels across the multiple dimensions spanned by RNNs: time, training epoch, and units. We demonstrate on multiple datasets that MM-PHATE uniquely preserves hidden representation community structure among units and identifies information processing and compression phases during training. The embedding allows users to look under the hood of RNNs across training and provides an intuitive and comprehensive strategy to understanding the network's internal dynamics and draw conclusions, e.g., on why and how one model outperforms another or how a specific architecture might impact an RNN's learning ability.", "keywords": [ "RNNs", "Dimensionality Reduction", "Hidden State", "Visualization", "Hidden Dynamics", "Deep Learning" ], "rating_list": [ 6, 5, 3, 5, 6 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 2, 1, 2, 2, 3 ], "confidence_list": [ 3, 3, 2, 3, 1 ] }, { "paper_id": "Y2z31hfEeq", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Gregory Valiant", "gender": "Unspecified", "institution": "Computer Science Department, Stanford University", "country": "", "position": "Associate Professor" }, { "name": "Laurent Charlin", "gender": "Male", "institution": "HEC Montreal", "country": "CA", "position": "Associate Professor" }, { "name": "Omar Salemohamed", "gender": "unknown", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "Shivam Garg", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Researcher" }, { "name": "Vatsal Sharan", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 3, "n_ref_uni": 35, "n_ref": 57, "n_ref_all": 99, "n_fig": 24, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 17, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 28, "formula_len_all_1": 132, "len_all": 157440, "len_all_1": 72391, "len_abs": 1120, "len_title": 111, "len_sents": 59455, "len_sents_1": 35571, "n_sents": 444, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1132, "title": "Discovering Data Structures: Nearest Neighbor Search and Beyond", "abs": "We propose a general framework for end-to-end learning of data structures. Our framework adapts to the underlying data distribution and provides fine-grained control over query and space complexity. Crucially, the data structure is learned from scratch, and does not require careful initialization or seeding with candidate data structures/algorithms. We first apply this framework to the problem of nearest neighbor search. In several settings, we are able to reverse-engineer the learned data structures and query algorithms. For 1D nearest neighbor search, the model discovers optimal distribution (in)dependent algorithms such as binary search and variants of interpolation search. In higher dimensions, the model learns solutions that resemble k-d trees in some regimes, while in others, they have elements of locality-sensitive hashing. The model can also learn useful representations of high-dimensional data and exploit them to design effective data structures. We also adapt our framework to the problem of estimating frequencies over a data stream, and believe it could also be a powerful discovery tool for new problems.", "keywords": [ "algorithm discovery", "data structures", "neural algorithms" ], "rating_list": [ 5, 8, 3, 5 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 2, 4, 4, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "Y2cGisOWPZ", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Eric Xue", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" }, { "name": "Rebekah A. Gelp\u00ed", "gender": "Female", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "William A Cunningham", "gender": "unknown", "institution": "University of Toronto", "country": "CA", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 58, "n_ref": 81, "n_ref_all": 93, "n_fig": 5, "n_tab": 4, "L_tab": 409, "n_element_tab": 43, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 366, "n_element_tab_1": 50, "formula_len_all": 41, "formula_len_all_1": 0, "len_all": 145151, "len_all_1": 66052, "len_abs": 1314, "len_title": 131, "len_sents": 43569, "len_sents_1": 34066, "n_sents": 269, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1256, "title": "Towards Machine Theory of Mind with Large Language Model-Augmented Inverse Planning", "abs": "We propose a hybrid approach to machine Theory of Mind (ToM) that uses large language models (LLMs) as a mechanism for generating hypotheses and likelihood functions with a Bayesian inverse planning model that computes posterior probabilities for an agent\u2019s likely mental states given its actions. Bayesian inverse planning models can accurately predict human reasoning on a variety of ToM tasks, but these models are constrained in their ability to scale these predictions to scenarios with a large number of possible hypotheses and actions. Conversely, LLM-based approaches have recently demonstrated promise in solving ToM benchmarks, but can exhibit brittleness and failures on reasoning tasks even when they pass otherwise structurally identical versions. By combining these two methods, our approach leverages the strengths of each component, closely matching optimal results on a task inspired by prior inverse planning models and improving performance relative to models that utilize LLMs alone or with chain-of-thought prompting. We also exhibit the model\u2019s potential to predict mental states on open-ended tasks, offering a promising direction for future development of ToM models and the creation of socially intelligent generative agent models.", "keywords": [ "inverse planning", "large language models", "theory of mind", "social reasoning" ], "rating_list": [ 5, 8, 5, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "Y2RW9EVwhT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 16, "author_info_list": [ { "name": "Andrew Tao", "gender": "Male", "institution": "NVIDIA Corporation", "country": "US", "position": "Vice President of Applied Research" }, { "name": "Bryan Catanzaro", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Vice President" }, { "name": "De-An Huang", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Fuxiao Liu", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Guilin Liu", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Jan Kautz", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "VP Research" }, { "name": "Karan Sapra", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Min Shi", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Shihao Wang", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Shijia Liao", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Subhashree Radhakrishnan", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yaser Yacoob", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "Associate Professor" }, { "name": "Yilin Zhao", "gender": "unknown", "institution": "New York University", "country": "US", "position": "MS student" }, { "name": "Zhiding Yu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Hongxu Yin", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Senior Research Scientist" }, { "name": "Honghui Shi", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 68, "n_ref": 163, "n_ref_all": 194, "n_fig": 7, "n_tab": 7, "L_tab": 4559, "n_element_tab": 392, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3368, "n_element_tab_1": 432, "formula_len_all": 0, "formula_len_all_1": 39, "len_all": 191407, "len_all_1": 69680, "len_abs": 0, "len_title": 127, "len_sents": 36097, "len_sents_1": 28991, "n_sents": 266, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 78, "L_abs": 1291, "title": "Eagle: Exploring The Design Space for Multimodal LLMs with Mixture of Encoders", "abs": "The ability to accurately interpret complex visual information is a crucial topic of multimodal large language models (MLLMs). Recent work indicates that enhanced visual perception significantly reduces hallucinations and improves performance on resolution-sensitive tasks, such as optical character recognition and document analysis. A number of recent MLLMs achieve this goal using a mixture of vision encoders. Despite their success, there is a lack of systematic comparisons and detailed ablation studies addressing critical aspects, such as expert selection and the integration of multiple vision experts. This study provides an extensive exploration of the design space for MLLMs using a mixture of vision encoders and resolutions. Our findings reveal several underlying principles common to various existing strategies, leading to a streamlined yet effective design approach. We discover that simply concatenating visual tokens from a set of complementary vision encoders is as effective as more complex mixing architectures or strategies. We additionally introduce Pre-Alignment to bridge the gap between vision-focused encoders and language tokens, enhancing model coherence. The resulting family of MLLMs, Eagle, surpasses other leading open-source models on major MLLM benchmarks.", "keywords": [ "LLM", "Multimodal LLM", "Vision Encoder" ], "rating_list": [ 8, 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 4, 3, 3 ], "presentation_list": [ 4, 3, 4, 3, 3 ], "contribution_list": [ 3, 2, 4, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "Y2Dh8rWwlb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Jianfeng Wang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Jing Gu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Kaizhi Zheng", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Kevin Lin", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Lijuan Wang", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Linjie Li", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Xiaotong Chen", "gender": "Male", "institution": "ProtagoLabs, Inc. / NetMind", "country": "US", "position": "Researcher" }, { "name": "Xin Eric Wang", "gender": "Male", "institution": "Simular", "country": "US", "position": "Principal Researcher" }, { "name": "Xuehai He", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "student" }, { "name": "Zhengyuan Yang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 29, "n_ref": 62, "n_ref_all": 78, "n_fig": 7, "n_tab": 6, "L_tab": 2185, "n_element_tab": 388, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2058, "n_element_tab_1": 361, "formula_len_all": 316, "formula_len_all_1": 316, "len_all": 139305, "len_all_1": 60733, "len_abs": 0, "len_title": 0, "len_sents": 38259, "len_sents_1": 27381, "n_sents": 247, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1285, "title": "EditRoom: LLM-parameterized Graph Diffusion for Composable 3D Room Layout Editing", "abs": "Given the steep learning curve of professional 3D software and the time-\nconsuming process of managing large 3D assets, language-guided 3D scene editing has significant potential in fields such as virtual reality, augmented reality, and\ngaming. However, recent approaches to language-guided 3D scene editing either\nrequire manual interventions or focus only on appearance modifications without\nsupporting comprehensive scene layout changes. In response, we propose EditRoom, a unified framework capable of executing a variety of layout edits through\nnatural language commands, without requiring manual intervention. Specifically,\nEditRoom leverages Large Language Models (LLMs) for command planning and\ngenerates target scenes using a diffusion-based method, enabling six types of edits: rotate, translate, scale, replace, add, and remove. To address\nthe lack of data for language-guided 3D scene editing, we have developed an automatic pipeline to augment existing 3D scene synthesis datasets and introduced\nEditRoom-DB, a large-scale dataset with 83k editing pairs, for training and evaluation. Our experiments demonstrate that our approach consistently outperforms\nother baselines across all metrics, indicating higher accuracy and coherence in\nlanguage-guided scene layout editing.", "keywords": [ "3D Scene Editing", "Large Language Model", "Diffusion-based Models" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "Y2AH0wC6C9", "primary_area": "foundation or frontier models, including LLMs", "n_author": 13, "author_info_list": [ { "name": "Bingxiang He", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Cheng Qian", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Ganqu Cui", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Haiwen Hong", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Huan-ang Gao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hui Xue'", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Huimin Chen", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Jia Deng", "gender": "Female", "institution": "Renmin University of China", "country": "", "position": "Undergrad student" }, { "name": "Lifan Yuan", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Longtao Huang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Maosong Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Ning Ding", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Zhiyuan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 19, "n_formula_1": 0, "n_ref_uni": 47, "n_ref": 87, "n_ref_all": 121, "n_fig": 16, "n_tab": 5, "L_tab": 791, "n_element_tab": 91, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 342, "n_element_tab_1": 81, "formula_len_all": 1439, "formula_len_all_1": 0, "len_all": 205280, "len_all_1": 59197, "len_abs": 1432, "len_title": 125, "len_sents": 67601, "len_sents_1": 27968, "n_sents": 517, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1440, "title": "Improving Zero-Shot Generalization of Instruction Tuning by Data Arrangement", "abs": "Understanding alignment techniques begins with comprehending zero-shot generalization brought by instruction tuning, but little of the mechanism has been understood. Existing work has largely been confined to the task level, without considering that tasks are artificially defined and, to LLMs, merely consist of tokens and representations. To bridge this gap, we investigate zero-shot generalization from the perspective of the data itself. We first demonstrate that zero-shot generalization happens very early during instruction tuning, with loss serving as a stable indicator. Next, we investigate the facilitation of zero-shot generalization by data arrangement through similarity and granularity perspectives, confirming that encountering highly similar and fine-grained training data earlier during instruction tuning, without the constraints of defined ``tasks'', enables better generalization. Finally, we propose a more grounded training data arrangement method, Test-centric Multi-turn Arrangement, and show its effectiveness in promoting continual learning and further loss reduction. For the first time, we show that zero-shot generalization during instruction tuning is a form of similarity-based generalization between training and test data at the instance level. We hope our analysis will advance the understanding of zero-shot generalization during instruction tuning and contribute to the development of more aligned LLMs.", "keywords": [ "zero-shot", "instruction tuning", "large language models" ], "rating_list": [ 5, 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 3, 3 ] }, { "paper_id": "Y1r9yCMzeA", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Jia Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Jianheng Tang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Nuo Chen", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Qifan Zhang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yuhan Li", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 78, "n_ref": 122, "n_ref_all": 148, "n_fig": 8, "n_tab": 11, "L_tab": 3813, "n_element_tab": 451, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1397, "n_element_tab_1": 208, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 234031, "len_all_1": 55080, "len_abs": 972, "len_title": 115, "len_sents": 51042, "len_sents_1": 25780, "n_sents": 351, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1093, "title": "GraphArena: Evaluating and Exploring Large Language Models on Graph Computation", "abs": "The ``arms race'' of Large Language Models (LLMs) demands new benchmarks to examine their progresses. In this paper, we introduce GraphArena, a benchmarking tool designed to evaluate LLMs on real-world graph computational problems. It offers a suite of four polynomial-time tasks (e.g., Shortest Distance) and six NP-complete challenges (e.g., Traveling Salesman Problem). GraphArena features a rigorous evaluation framework that classifies LLM outputs as correct, suboptimal (feasible but not optimal), hallucinatory (properly formatted but infeasible), or missing. Evaluation of over 10 LLMs reveals that even top-performing LLMs struggle with larger, more complex graph problems and exhibit hallucination issues. We further explore four potential solutions to address this issue and improve LLMs on graph computation, including chain-of-thought prompting, instruction tuning, code writing, and scaling test-time compute, each demonstrating unique strengths and limitations. GraphArena complements the existing LLM benchmarks and is open-sourced at https://github.com/squareRoot3/GraphArena.", "keywords": [ "Large Language Model Evaluation", "Graph Algorithms" ], "rating_list": [ 6, 8, 8, 5 ], "soundness_list": [ 3, 4, 4, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "Y1XkzMJpPd", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Antoine Cully", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Assistant Professor" }, { "name": "Jeff Clune", "gender": "unknown", "institution": "University of Wyoming", "country": "", "position": "" }, { "name": "Jenny Zhang", "gender": "Female", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Maxence Faldor", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 0, "n_ref_uni": 47, "n_ref": 106, "n_ref_all": 130, "n_fig": 8, "n_tab": 12, "L_tab": 16258, "n_element_tab": 429, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 21456, "formula_len_all_1": 0, "len_all": 509661, "len_all_1": 70641, "len_abs": 1664, "len_title": 157, "len_sents": 158631, "len_sents_1": 37208, "n_sents": 2123, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 109, "L_abs": 1616, "title": "OMNI-EPIC: Open-endedness via Models of human Notions of Interestingness with Environments Programmed in Code", "abs": "Open-ended and AI-generating algorithms aim to continuously generate and solve increasingly complex tasks indefinitely, offering a promising path toward more general intelligence. To accomplish this grand vision, learning must occur within a vast array of potential tasks. Existing approaches to automatically generating environments are constrained within manually predefined, often narrow distributions of environments, limiting their ability to create any learning environment. To address this limitation, we introduce a novel framework, OMNI-EPIC, that augments previous work in Open-endedness via Models of human Notions of Interestingness (OMNI) with Environments Programmed in Code (EPIC). OMNI-EPIC leverages foundation models to autonomously generate code specifying the next learnable (i.e., not too easy or difficult for the agent\u2019s current skill set) and interesting (e.g., worthwhile and novel) tasks. OMNI-EPIC generates both environments (e.g., an obstacle course) and reward functions (e.g., progress through the obstacle course quickly without touching red objects), enabling it, in principle, to create any simulatable learning task. We showcase the explosive creativity of OMNI-EPIC, which continuously innovates to suggest new, interesting learning challenges. We also highlight how OMNI-EPIC can adapt to reinforcement learning agents\u2019 learning progress, generating tasks that are of suitable difficulty. Overall, OMNI-EPIC has the potential to endlessly create learnable and interesting environments, further propelling the development of self-improving AI systems and AI-Generating Algorithms.", "keywords": [ "Open-endedness", "Environment Generation", "Reinforcement Learning" ], "rating_list": [ 8, 8, 8, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 4, 3, 3, 1 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "Y0qmwm6tgy", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Bei Yu", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Farzan Farnia", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Jingwei Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Wenqian Zhao", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zixiao Wang", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 17, "n_ref_uni": 52, "n_ref": 99, "n_ref_all": 125, "n_fig": 4, "n_tab": 9, "L_tab": 3872, "n_element_tab": 367, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 694, "n_element_tab_1": 49, "formula_len_all": 2775, "formula_len_all_1": 2918, "len_all": 161535, "len_all_1": 63666, "len_abs": 1417, "len_title": 142, "len_sents": 43004, "len_sents_1": 27364, "n_sents": 323, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1426, "title": "MoreauPruner: Robust Structured Pruning of Large Language Models against Weight Perturbations", "abs": "In the existing model pruning literature, the weight gradient has been extensively utilized to measure the importance of weight, where the gradient is well-known to be sensitive to perturbations. On the other hand, the widely used large language models (LLMs) have several billion model parameters, which could increase the fragility of few-shot gradient pruning. In this work, we experimentally show that one-shot gradient pruning algorithms could lead to unstable results under perturbations to model weights. Even the minor error of switching between data formats bfloat16 and float16 could result in obviously different outcomes. To address such instabilities, we leverage optimization analysis and propose an LLM structural pruning method, called MoreauPruner, with provable robustness against weight perturbations. In MoreauPruner, the model weight importance is estimated based on the neural network's Moreau envelope, which can be flexibly combined with $\\ell_1$-norm regularization techniques to induce the sparsity required in the pruning task. We extensively evaluate the MoreauPruner algorithm on several well-known LLMs, including LLaMA-7B, LLaMA-13B, LLaMA3-8B, and Vicuna-7B. Our numerical results suggest the robustness of MoreauPruner against weight perturbation and how robust importance estimation in MoreauPruner contributes to successful accuracy-based scores compared to several existing pruning methods.", "keywords": [ "Structured Pruning", "Large Language Model", "Robustness", "Moreau Envelop" ], "rating_list": [ 3, 5, 5, 5, 6 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 4 ], "contribution_list": [ 1, 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "Y0kmI2zqqi", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Alan Arnoldo Diaz Montiel", "gender": "unknown", "institution": "University Health Network", "country": "CA", "position": "Postdoc" }, { "name": "Hina Tabassum", "gender": "unknown", "institution": "York University", "country": "CA", "position": "Associate Professor" }, { "name": "Milad Lankarany", "gender": "unknown", "institution": "University of Toronto", "country": "", "position": "Assistant Professor" }, { "name": "Xavier Mootoo", "gender": "Male", "institution": ", York University", "country": "CA", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 58, "n_ref_all": 76, "n_fig": 3, "n_tab": 8, "L_tab": 2649, "n_element_tab": 155, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1638, "n_element_tab_1": 72, "formula_len_all": 1403, "formula_len_all_1": 518, "len_all": 128968, "len_all_1": 54805, "len_abs": 1325, "len_title": 143, "len_sents": 36657, "len_sents_1": 25599, "n_sents": 264, "n_sents_1": 169, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1361, "title": "Stochastic Sparse Sampling: A Framework for Variable-Length Medical Time Series Classification", "abs": "ile the majority of time series classification research has focused on modeling fixed-length sequences, variable-length time series classification (VTSC) remains critical in healthcare, where sequence length may vary among patients and events. To address this challenge, we propose $\\textbf{S}$tochastic $\\textbf{S}$parse $\\textbf{S}$ampling (SSS), a novel VTSC framework developed for medical time series. SSS manages variable-length sequences by sparsely sampling fixed windows to compute local predictions, which are then aggregated and calibrated to form a global prediction. We apply SSS to the task of seizure onset zone (SOZ) localization, a critical VTSC problem requiring identification of seizure-inducing brain regions from variable-length electrophysiological time series. We evaluate our method on the Epilepsy iEEG Multicenter Dataset, a heterogeneous collection of intracranial electroencephalography (iEEG) recordings obtained from four independent medical centers. SSS demonstrates superior performance compared to state-of-the-art (SOTA) baselines across most medical centers, and superior performance on all out-of-distribution (OOD) unseen medical centers. Additionally, SSS naturally provides post-hoc insights into local signal characteristics related to the SOZ, by visualizing temporally averaged local predictions throughout the signal.", "keywords": [ "Time Series", "Healthcare", "Medicine", "Epilepsy", "Neuroscience" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 1, 3, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "Y0QqruhqIa", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Chufeng Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Hang Chen", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiao Li", "gender": "Male", "institution": "Computer Science, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiaolin Hu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 18, "n_formula_1": 8, "n_ref_uni": 52, "n_ref": 107, "n_ref_all": 142, "n_fig": 12, "n_tab": 10, "L_tab": 3143, "n_element_tab": 117, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2389, "n_element_tab_1": 39, "formula_len_all": 994, "formula_len_all_1": 314, "len_all": 177340, "len_all_1": 60579, "len_abs": 1512, "len_title": 79, "len_sents": 44640, "len_sents_1": 27315, "n_sents": 353, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1567, "title": "Efficient Neuron Segmentation in Electron Microscopy by Affinity-Guided Queries", "abs": "Accurate segmentation of neurons in electron microscopy (EM) images plays a crucial role in understanding the intricate wiring patterns of the brain. Existing automatic neuron segmentation methods rely on traditional clustering algorithms, where affinities are predicted first, and then watershed and post-processing algorithms are applied to yield segmentation results. Due to the nature of watershed algorithm, this paradigm has deficiency in both prediction quality and speed. Inspired by recent advances in natural image segmentation, we propose to use query-based methods to address the problem because they do not necessitate watershed algorithms. However, we find that directly applying existing query-based methods faces great challenges due to the large memory requirement of the 3D data and considerably different morphology of neurons. To tackle these challenges, we introduce affinity-guided queries and integrate them into a lightweight query-based framework. Specifically, we first predict affinities with a lightweight branch, which provides coarse neuron structure information. The affinities are then used to construct affinity-guided queries, facilitating segmentation with bottom-up cues. These queries, along with additional learnable queries, interact with the image features to directly predict the final segmentation results. Experiments on benchmark datasets demonstrated that our method achieved better results over state-of-the-art methods with a 2$\\sim$3$\\times$ speedup in inference. Code is available at https://github.com/chenhang98/AGQ.", "keywords": [ "Neuron Segmentation; Biomedical Image Segmentation; Electron Microscopy Image" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "Y0P6cOZzNm", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Dongrui Liu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Li Niu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Linfeng Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Qianli Ma", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "MS student" }, { "name": "Xuefei Ning", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Research Assistant Professor" } ], "n_formula": 41, "n_formula_1": 13, "n_ref_uni": 43, "n_ref": 99, "n_ref_all": 127, "n_fig": 10, "n_tab": 5, "L_tab": 1081, "n_element_tab": 130, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 586, "n_element_tab_1": 76, "formula_len_all": 2848, "formula_len_all_1": 699, "len_all": 163501, "len_all_1": 54431, "len_abs": 1174, "len_title": 113, "len_sents": 44713, "len_sents_1": 26689, "n_sents": 302, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1211, "title": "Decouple-Then-Merge: Towards Better Training for Diffusion Models", "abs": "Diffusion models are trained by learning a sequence of models that reverse each step of noise corruption. Typically, the model parameters are fully shared across multiple timesteps to enhance training efficiency. However, since the denoising tasks differ at each timestep, the gradients computed at different timesteps may conflict, potentially degrading the overall performance of image generation. To solve this issue, this work proposes a $\\textbf{De}$couple-then-$\\textbf{Me}$rge ($\\textbf{DeMe}$) framework, which begins with a pretrained model and finetunes separate models tailored to specific timesteps. We introduce several improved techniques during the finetuning stage to promote effective knowledge sharing while minimizing training interference across timesteps. Finally, after finetuning, these separate models can be merged into a single model in the parameter space, ensuring efficient and practical inference. Experimental results show significant generation quality improvements upon 6 benchmarks including Stable Diffusion on COCO30K, ImageNet1K, PartiPrompts, and DDPM on LSUN Church, LSUN Bedroom, and CIFAR10. Code is included in the supplementary material and will be released on Github.", "keywords": [ "Diffusion Model", "Image Synthesis" ], "rating_list": [ 5, 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 3, 2 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "Y07R8h6m8e", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chenyang Ren", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Di Wang", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Huanyi Xie", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "MS student" }, { "name": "Jingfeng Zhang", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Researcher" }, { "name": "Khouloud Saadi", "gender": "unknown", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Lijie Hu", "gender": "Female", "institution": "KAUST", "country": "SA", "position": "PhD student" }, { "name": "Shu Yang", "gender": "Female", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" } ], "n_formula": 104, "n_formula_1": 20, "n_ref_uni": 45, "n_ref": 57, "n_ref_all": 75, "n_fig": 2, "n_tab": 6, "L_tab": 1137, "n_element_tab": 148, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 3030, "n_element_tab_1": 70, "formula_len_all": 11195, "formula_len_all_1": 1968, "len_all": 200626, "len_all_1": 69590, "len_abs": 1506, "len_title": 131, "len_sents": 57842, "len_sents_1": 28792, "n_sents": 481, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1517, "title": "Dissecting Misalignment of Multimodal Large Language Models via Influence Function", "abs": "Multi-modal Large Language models (MLLMs) are always trained on data from diverse and unreliable sources, which may contain misaligned or mislabeled text-image pairs. This frequently causes robustness issues and hallucinations, leading to performance degradation. Data valuation is an efficient way to detect and trace these misalignments. Nevertheless, existing methods are computationally expensive for MLLMs. \nWhile computationally efficient, the classical influence functions are inadequate for contrastive learning models because they were originally designed for pointwise loss. Additionally, contrastive learning involves minimizing the distance between the modalities of positive samples and maximizing the distance between the modalities of negative samples. This requires us to evaluate the influence of samples from both perspectives. To tackle these challenges, we introduce the Extended Influence Function for Contrastive Loss (ECIF), an influence function crafted for contrastive loss. ECIF considers both positive and negative samples and provides a closed-form approximation of contrastive learning models, eliminating the need for retraining. Building upon ECIF, we develop a series of algorithms for data evaluation in MLLM, misalignment detection, and misprediction trace-back tasks. Experimental results demonstrate our ECIF advances the transparency and interpretability of MLLMs by offering a more accurate assessment of data impact and model alignment compared to traditional baseline methods.", "keywords": [ "Multimodal Large Language Model", "Interpretability" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "XzU3Xk1Xu2", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Ion Stoica", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Joseph E. Gonzalez", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Associate Professor" }, { "name": "Lianmin Zheng", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Shuo Yang", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Ying Sheng", "gender": "Female", "institution": "Stanford University", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 48, "n_ref": 65, "n_ref_all": 82, "n_fig": 10, "n_tab": 3, "L_tab": 619, "n_element_tab": 82, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1609, "n_element_tab_1": 123, "formula_len_all": 537, "formula_len_all_1": 537, "len_all": 156043, "len_all_1": 58484, "len_abs": 1447, "len_title": 99, "len_sents": 32873, "len_sents_1": 26811, "n_sents": 235, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1488, "title": "Post-Training Sparse Attention with Double Sparsity", "abs": "Long-context inference of Large Language Models (LLMs) is known to be challenging due to the excessive Key-Value(KV) cache accesses. This paper introduces ``Double Sparsity,'' a novel post-training sparse attention technique designed to alleviate this bottleneck by reducing KV cache access. Double Sparsity combines token sparsity, which focuses on using only the important tokens for computing self-attention, with channel sparsity, an approach that uses important feature channels for identifying important tokens. Our key insight is that the pattern of channel sparsity is highly static, allowing us to use offline calibration to make it efficient at runtime, thereby enabling accurate and efficient identification of important tokens. Moreover, this method can be combined with offloading to achieve significant memory usage reduction.\nExperimental results demonstrate that Double Sparsity can achieve \\(\\frac{1}{16}\\) sparsity with minimal impact on accuracy across various tasks with different architectures including MHA, GQA, MoE and vision language model.\nIt brings up to a 14.1$\\times$ acceleration in attention operations and a 1.9$\\times$ improvement in end-to-end inference on GPUs with various batch sizes. With CPU offloading under extremely long-context settings (e.g., 256K), it achieves a decoding speed acceleration of 16.3$\\times$ compared to state-of-the-art solutions. Our code is integrated into a widely-used framework SGLang and deployed in real-world workloads.", "keywords": [ "large language models", "sparse attention", "decoding" ], "rating_list": [ 5, 8, 3, 3 ], "soundness_list": [ 3, 4, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "Xz5J6Hj9cH", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Hanzhang Qin", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Assistant Professor" }, { "name": "Sihan Chen", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "MS student" }, { "name": "Wanghao Ye", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 17, "n_ref": 24, "n_ref_all": 31, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 102135, "len_all_1": 64238, "len_abs": 1202, "len_title": 175, "len_sents": 39374, "len_sents_1": 33897, "n_sents": 254, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 127, "L_abs": 1205, "title": "CogniPair - Dynamic LLM Matching Algorithm in Chaotic Environments Mimicking Human Cognitive Processes for Relationship Pairing", "abs": "Dating applications in the digital era have transformed how people connect, yet they often fall short in simulating the comprehensive character and fostering truly compatible relationships due to their reliance on quantitative data. This paper proposes a novel framework to simulate human characters by leveraging Large Language Models (LLMs) to enhance matchmaking by understanding the nuanced fabric of human personality and social connections. Traditional algorithms often lack the depth needed for personalized matchmaking, whereas LLMs offer sophisticated linguistic and cognitive capabilities to simulate a person and complicated personal decisions. Our framework introduces a multi-agent system comprising the Persona, Preference, and Dating Memory modules, allowing for dynamic and nuanced user interactions. This approach addresses the limitations of conventional LLM frameworks by capturing detailed personal attributes, updating preferences, and learning from past interactions. Our system enhances the relevance and effectiveness of match recommendations, focusing on emotional compatibility and shared values, providing a more personalized and responsive user experience in the dating domain.", "keywords": [ "Large Language Models", "Dating Algorithms", "Human-like Reasoning", "Context-aware Analysis", "Simulating Characters", "Machine Psychology" ], "rating_list": [ 8, 3, 3, 3 ], "soundness_list": [ 3, 3, 1, 1 ], "presentation_list": [ 3, 1, 1, 3 ], "contribution_list": [ 3, 2, 1, 1 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "XydIBZ2xlr", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Junyuan Deng", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Ping Tan", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Qian Zhang", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Wei Yin", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Weiqiang Ren", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Xiaotao Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Xiaoyang Guo", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "XIAOXIAO LONG", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Postdoc" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 70, "n_ref": 162, "n_ref_all": 190, "n_fig": 14, "n_tab": 14, "L_tab": 2970, "n_element_tab": 385, "n_fig_1": 10, "n_tab_1": 5, "L_tab_1": 1787, "n_element_tab_1": 285, "formula_len_all": 658, "formula_len_all_1": 424, "len_all": 199263, "len_all_1": 66665, "len_abs": 1672, "len_title": 122, "len_sents": 44604, "len_sents_1": 28239, "n_sents": 321, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1680, "title": "Boost 3D Reconstruction using Diffusion-based Intrinsic Estimation", "abs": "In this paper, we present DM-Calib, a diffusion-based approach for estimating pinhole camera intrinsic parameters from a single input image. Monocular camera calibration is essential for many 3D vision tasks. However, most existing methods depend on handcrafted assumptions or are constrained by limited training data, resulting in poor generalization across diverse real-world images. Recent advancements in stable diffusion models, trained on massive data, have shown the ability to generate high-quality images with varied characteristics. Emerging evidence indicates that these models implicitly capture the relationship between camera focal length and image content. Building on this insight, we explore how to leverage the powerful priors of diffusion models for monocular pinhole camera calibration. Specifically, we introduce a new image-based representation, termed Camera Image, which losslessly encodes the numerical camera intrinsics and integrates seamlessly with the diffusion framework. Using this representation, we reformulate the problem of estimating camera intrinsics as the generation of a dense Camera Image conditioned on an input image. By fine-tuning a stable diffusion model to generate a Camera Image from a single RGB input, we can extract camera intrinsics via a RANSAC operation. We further demonstrate that our monocular calibration method enhances performance across various 3D tasks, including zero-shot metric depth estimation, 3D metrology, pose estimation and sparse-view reconstruction. Extensive experiments on multiple public datasets show that our approach significantly outperforms baselines and provides broad benefits to 3D vision tasks.", "keywords": [ "Calibration", "Diffusion", "3D reconstruction." ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Xy5iXnFNzL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Changsheng Li", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Jiaming Han", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xiangyu Yue", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Hao-Ran Hao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Undergrad student" }, { "name": "Yu-Feng Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 43, "n_ref": 108, "n_ref_all": 141, "n_fig": 7, "n_tab": 12, "L_tab": 1998, "n_element_tab": 135, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1459, "n_element_tab_1": 116, "formula_len_all": 225, "formula_len_all_1": 149, "len_all": 193622, "len_all_1": 68279, "len_abs": 199, "len_title": 0, "len_sents": 53855, "len_sents_1": 30637, "n_sents": 572, "n_sents_1": 256, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1590, "title": "Remember, Retrieve and Generate: Understanding Infinite Visual Concepts as Your Personalized Assistant", "abs": "The development of large language models (LLMs) has significantly enhanced the capabilities of multimodal LLMs (MLLMs) as general assistants. However, lack of user-specific knowledge still restricts their application in human's daily life. In this paper, we introduce the **R**etrieval **A**ugmented **P**ersonalization (RAP) framework for MLLMs' personalization. Starting from a general MLLM, we turn it into a personalized assistant in three steps. (a) Remember: We design a key-value database to store user-related information, *e.g.*, user's name, avatar and other attributes. (b) Retrieve: When the user initiates a conversation, RAP will retrieve relevant information from the database using a multimodal retriever. (c) Generate: The input query and retrieved concepts' information are fed into MLLMs to generate personalized, knowledge-augmented responses. Unlike previous methods, RAP allows real-time concept editing via updating the external database. To further improve generation quality and alignment with user-specific information, we design a pipeline for data collection and create a specialized dataset for personalized training of MLLMs. Based on the dataset, we train a series of MLLMs as personalized multimodal assistants. By pretraining on large-scale dataset, RAP-MLLMs can generalize to infinite visual concepts without additional finetuning. Our models demonstrate outstanding flexibility and generation quality across a variety of tasks, such as personalized image captioning, question answering and visual recognition. The code, data and models will be available.", "keywords": [ "Multimodal Large Language Models", "Personalization", "Retrieval-augmented Generation" ], "rating_list": [ 8, 3, 5 ], "soundness_list": [ 4, 1, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 4, 2, 1 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "Xy1Lf7uR9H", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Chanmoo Park", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Insung Kong", "gender": "Male", "institution": "University of Twente", "country": "NL", "position": "Postdoc" }, { "name": "Seokhun Park", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Yongdai Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "yongchan Choi", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "PhD student" } ], "n_formula": 80, "n_formula_1": 25, "n_ref_uni": 31, "n_ref": 103, "n_ref_all": 168, "n_fig": 15, "n_tab": 21, "L_tab": 3650, "n_element_tab": 495, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1099, "n_element_tab_1": 167, "formula_len_all": 10491, "formula_len_all_1": 1729, "len_all": 218844, "len_all_1": 63403, "len_abs": 1299, "len_title": 150, "len_sents": 63480, "len_sents_1": 24701, "n_sents": 609, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1310, "title": "ANOVA-NODE: An identifiable neural network for the functional ANOVA model for better interpretability", "abs": "Interpretability for machine learning models is becoming more and more important as machine learning models become more complex. \nThe functional ANOVA model, which decomposes a high-dimensional function into a sum of lower dimensional functions so called components, is one of the most popular tools for interpretable AI, and recently, various neural network models have been developed for estimating each component in the functional ANOVA model. \nHowever, such neural networks are highly unstable when estimating components since the components themselves are not uniquely defined. \nThat is, there are multiple functional ANOVA decompositions for a given function. \nIn this paper, we propose a novel interpretable model which guarantees a unique functional ANOVA decomposition and thus is able to estimate each component stably. \nWe call our proposed model ANOVA-NODE since it is a modification of Neural Oblivious Decision Ensembles (NODE) for the functional ANOVA model. \nTheoretically, we prove that ANOVA-NODE can approximate a smooth function well.\nAdditionally, we experimentally show that ANOVA-NODE provides much more stable estimation of each component and thus much more stable interpretation when training data and initial values of the model parameters vary than existing neural network models do.", "keywords": [ "Interpretability", "Explianability", "Trustworthy AI", "Functional ANOVA model", "Generalized additive models" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "XxxKHiy9Gw", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Daehee Cho", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Hyeonmin Yun", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Undergrad student" }, { "name": "Jae Yong Lee", "gender": "Male", "institution": "Chung-Ang University", "country": "KR", "position": "Assistant Professor" }, { "name": "Mikyoung Lim", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 47, "n_formula_1": 27, "n_ref_uni": 37, "n_ref": 51, "n_ref_all": 97, "n_fig": 10, "n_tab": 4, "L_tab": 1798, "n_element_tab": 124, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 257, "n_element_tab_1": 35, "formula_len_all": 3290, "formula_len_all_1": 1731, "len_all": 145269, "len_all_1": 63666, "len_abs": 1371, "len_title": 183, "len_sents": 42485, "len_sents_1": 29316, "n_sents": 331, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 134, "L_abs": 1377, "title": "Conformal mapping Coordinates Physics-Informed Neural Networks (CoCo-PINNs): learning neural networks for designing neutral inclusions", "abs": "We focus on designing and solving the neutral inclusion problem via neural networks. The neutral inclusion problem has a long history in the theory of composite materials, and it is exceedingly challenging to identify the precise condition that precipitates a general-shaped inclusion into a neutral inclusion. Physics-informed neural networks (PINNs) have recently become a highly successful approach to addressing both forward and inverse problems associated with partial differential equations. We found that traditional PINNs perform inadequately when applied to the inverse problem of designing neutral inclusions with arbitrary shapes. In this study, we introduce a novel approach, Conformal mapping Coordinates Physics-Informed Neural Networks (CoCo-PINNs), which integrates complex analysis techniques into PINNs. This method exhibits strong performance in solving forward-inverse problems to construct neutral inclusions of arbitrary shapes in two dimensions, where the imperfect interface condition on the inclusion's boundary is modeled by training neural networks. Notably, we mathematically prove that training with a single linear field is sufficient to achieve neutrality for untrained linear fields in arbitrary directions, given a minor assumption. We demonstrate that CoCo-PINNs offer enhanced performances in terms of credibility, consistency, and stability.", "keywords": [ "Inverse problems", "PINNs", "Complex analysis", "Neutral inclusions" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "Xxpt66OgHI", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Aritra Dutta", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "El houcine Bergou", "gender": "Male", "institution": "KAUST", "country": "", "position": "Research scientist" }, { "name": "Mostapha Essoullami", "gender": "Male", "institution": "University Mohammed VI Polytechnic", "country": "", "position": "PhD student" } ], "n_formula": 54, "n_formula_1": 13, "n_ref_uni": 34, "n_ref": 63, "n_ref_all": 89, "n_fig": 13, "n_tab": 8, "L_tab": 2890, "n_element_tab": 194, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 970, "n_element_tab_1": 56, "formula_len_all": 5792, "formula_len_all_1": 833, "len_all": 173857, "len_all_1": 61762, "len_abs": 1664, "len_title": 138, "len_sents": 47764, "len_sents_1": 29026, "n_sents": 385, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1699, "title": "LEGACY: A Lightweight Adaptive Gradient Compression Strategy for Distributed Deep Learning", "abs": "Distributed learning has demonstrated remarkable success in training deep neural networks (DNNs) on large datasets, but the communication bottleneck reduces its scalability. Various compression techniques are proposed to alleviate this limitation; often they rely on computationally intensive methods to determine optimal compression parameters during training and are popularly referred to as adaptive compressors. Instead of the hard-to-tune hyperparameters for adaptive compressors, in this paper, we investigate the impact of two fundamental factors in DNN training, the layer size of the DNNs and their training phases, to design a simple yet efficient adaptive scheduler for any compressors to guide the compression parameters selection. We present a **L**ightweight **E**fficient **G**r**A**dient *C*ompression strateg**Y** or LEGACY that, in theory, can work with any compression technique to produce its simple adaptive counterpart. We benchmark LEGACY on distributed and federated training, involving 6 different DNN architectures for various tasks performed on large and challenging datasets, including ImageNet and WikiText-103. On ImageNet training, by sending similar average data volume, LEGACY's adaptive compression strategies improve the Top-1 accuracy of ResNet-50 by 7%-11%, compared to the uniform Top-0.1% compression used throughout the training. Similarly, on WikiText-103, by using our layer-based adaptive compression strategy and sending similar average data volume, the perplexity of the Transformer-XL improves $\\sim$26% more than the uniform Top-0.1% compression used throughout the training. We publish anonymized code at: https://github.com/LEGACY-compression/LEGACY.", "keywords": [ "Adaptive Gradient Compression", "Gradient Compression", "Distributed Deep Learning", "Federated Learning", "Efficient Communication", "Gradient Sparsification", "Communication Overhead" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "XxCgeWSTNp", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Artem Khrapov", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "RU", "position": "Intern" }, { "name": "Assel Yermekova", "gender": "Female", "institution": "Huawei ", "country": "RU", "position": "Researcher" }, { "name": "Mikhail Sergeevich Kudinov", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Engineer" }, { "name": "Tasnima Sadekova", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "RU", "position": "Engeneer" }, { "name": "Vadim Popov", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 13, "n_ref_uni": 42, "n_ref": 95, "n_ref_all": 133, "n_fig": 5, "n_tab": 5, "L_tab": 1241, "n_element_tab": 126, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1965, "n_element_tab_1": 159, "formula_len_all": 1267, "formula_len_all_1": 677, "len_all": 136466, "len_all_1": 64322, "len_abs": 1080, "len_title": 108, "len_sents": 37436, "len_sents_1": 28633, "n_sents": 236, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1081, "title": "Improved Sampling Algorithms for L\u00e9vy-It\u00f4 Diffusion Models", "abs": "L\u00e9vy-It\u00f4 denoising diffusion models relying on isotropic \u03b1-stable noise instead of Gaussian distribution have recently been shown to improve performance of conventional diffusion models in image generation on imbalanced datasets while performing comparably in the standard settings. However, the stochastic algorithm of sampling from such models consists in solving the stochastic differential equation describing only an approximate inverse of the process of adding \u03b1-stable noise to data which may lead to suboptimal performance. In this paper, we derive a parametric family of stochastic differential equations whose solutions have the same marginal densities as those of the forward diffusion and show that the appropriate choice of the parameter values can improve quality of the generated images when the number of reverse diffusion steps is small. Also, we demonstrate that L\u00e9vy-It\u00f4 diffusion models are applicable to diverse domains and show that a well-trained text-to-speech L\u00e9vy-It\u00f4 model may have advantages over standard diffusion models on highly imbalanced datasets.", "keywords": [ "generative modeling", "diffusion models", "L\u00e9vy-It\u00f4 models", "\u03b1-stable L\u00e9vy processes", "stochastic differential equations" ], "rating_list": [ 5, 6, 8, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "XwibrZ9MHG", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Firat Ozdemir", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Researcher" }, { "name": "Hehui Zheng", "gender": "Female", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Jan Obrist", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Juan Jose Zarate", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "Researcher" }, { "name": "Miguel Angel Zamora Mora", "gender": "unknown", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Ronan Hinchet", "gender": "Not Specified", "institution": "ETHZ - ETH Zurich", "country": "", "position": "Researcher" }, { "name": "Stelian Coros", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "Associate Professor" }, { "name": "Robert Kevin Katzschmann", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 39, "n_ref_all": 83, "n_fig": 15, "n_tab": 7, "L_tab": 2804, "n_element_tab": 480, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 2831, "n_element_tab_1": 211, "formula_len_all": 343, "formula_len_all_1": 299, "len_all": 115778, "len_all_1": 54013, "len_abs": 548, "len_title": 124, "len_sents": 31538, "len_sents_1": 21694, "n_sents": 222, "n_sents_1": 149, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1430, "title": "PokeFlex: A Real-World Dataset of Deformable Objects for Robotics", "abs": "Data-driven methods have shown great potential in solving challenging manipulation tasks, however, their application in the domain of deformable objects has been constrained, in part, by the lack of data. \nTo address this, we propose PokeFlex, a dataset featuring real-world paired and annotated multimodal data that includes 3D textured meshes, point clouds, RGB images, and depth maps. Such data can be leveraged for several downstream tasks such as online 3D mesh reconstruction, and it can potentially enable underexplored applications such as the real-world deployment of traditional control methods based on mesh simulations.\nTo deal with the challenges posed by real-world 3D mesh reconstruction, \nwe leverage a professional volumetric capture system that allows complete 360\u00b0 reconstruction. PokeFlex consists of 18 deformable objects with varying stiffness and shapes. Deformations are generated by dropping objects onto a flat surface or by poking the objects with a robot arm. Interaction forces and torques are also reported for the latter case. \nUsing different data modalities, we demonstrated a use case for our dataset in online 3D mesh reconstruction. We refer the reader to our [website](https://anonymized-pokeflex-dataset.github.io/) or the [password protected supplementary material](https://drive.google.com/drive/folders/1d8iNoJZ0dUVlzP6XxP7xwGPhdVtwQ7du) for further demos and examples (password in pdf).", "keywords": [ "Deformable objects", "Robotics", "3D mesh reconstruction." ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "XwUrzurG94", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Abhishek Gupta", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "Andrey Kolobov", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Ching-An Cheng", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Senior researcher" }, { "name": "Patrick Yin", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Tyler Westenbroek", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 13, "n_ref_uni": 64, "n_ref": 130, "n_ref_all": 140, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3269, "formula_len_all_1": 1033, "len_all": 197938, "len_all_1": 77048, "len_abs": 1209, "len_title": 125, "len_sents": 54908, "len_sents_1": 36029, "n_sents": 401, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1299, "title": "Rapidly Adapting Policies to the Real-World via Simulation-Guided Fine-Tuning", "abs": "Robot learning requires a considerable amount of data to realize the promise of generalization. However, it can be challenging to actually collect the magnitude of high-quality data necessary for generalization entirely in the real world. Simulation can serve as a source of plentiful data, wherein techniques such as reinforcement learning can obtain broad coverage over states and actions. However, high-fidelity physics simulators are fundamentally misspecified approximations to reality, making direct zero-shot transfer challenging, especially in tasks where precise and forceful manipulation is necessary. This makes real-world fine-tuning of policies pretrained in simulation an attractive approach to robot learning. However, exploring the real-world dynamics with standard RL fine-tuning techniques is to inefficient for many real-world applications. This paper introduces Simulation-Guided Fine-Tuning, a general framework which leverages the structure of the simulator to guide exploration, substantially accelerating adaptation to the real-world. We demonstrate our approach across several manipulation tasks in the real world, learning successful policies for problems that are challenging to learn using purely real-world data. We further provide theoretical backing for the paradigm.", "keywords": [ "Robot Learning", "Reinforcement Learning", "Fine-Tuning" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "XwERWxaqIr", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Adil Hasan", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Thomas Peyrin", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 27, "n_ref": 110, "n_ref_all": 129, "n_fig": 6, "n_tab": 5, "L_tab": 1478, "n_element_tab": 208, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1394, "n_element_tab_1": 190, "formula_len_all": 616, "formula_len_all_1": 453, "len_all": 116409, "len_all_1": 67824, "len_abs": 2532, "len_title": 126, "len_sents": 35360, "len_sents_1": 29807, "n_sents": 243, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 2565, "title": "AQUATIC-Diff: Additive Quantization for Truly Tiny Compressed Diffusion Models", "abs": "Tremendous investments have been made towards the commodification of diffusion models for generation of diverse media. Their mass-market adoption is however still hobbled by the intense hardware resource requirements of diffusion model inference. Model quantization strategies tailored specifically towards diffusion models have seen considerable success in easing this burden, yet without exception have explored only the Uniform Scalar Quantization (USQ) family of quantization methods. In contrast, Vector Quantization (VQ) methods, which operate on groups of multiple related weights as the basic unit of compression, have recently taken the parallel field of Large Language Model (LLM) quantization by storm. In this work, we for the first time apply codebook-based additive vector quantization algorithms to the problem of diffusion model compression, adapting prior works on the quantization-aware fine-tuning of transformer-based LLMs to take into account the special structure of convolutional weight tensors, the heterogeneity in the kinds of operations performed by the layers of a diffusion model, and the momentum-invalidating discontinuities encountered between successive batches during quantization-aware fine-tuning of diffusion models. We are rewarded with a data-free distillation framework which achieves to the best of our knowledge state-of-the-art results for the extremely low-bit weight quantization on the standard class-conditional benchmark of LDM-4 on ImageNet at 20 inference time steps. Notably, we report sFID 1.93 points lower than the full-precision model at W4A8, the best-reported results for FID, sFID and ISC at W2A8, and the first-ever successful quantization to W1.5A8 (less than 1.5 bits stored per weight) via a layer-wise heterogeneous quantization strategy. We thus establish a new Pareto frontier for diffusion model inference under low-memory conditions. Furthermore, our method allows for a dynamic trade-off between quantization-time GPU hours and inference-time savings, thus aligning with the recent trend of approaches that combine the best aspects of both Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT). We are also able to demonstrate FLOPs savings on arbitrary hardware via an efficient inference kernel, as opposed to BOPs (Bit-wise Operations) savings resulting from small integer operations that may lack broad support across hardware of interest.\\\\ Code is released via anonymized download link.\\\\\\footnotesize{\\url{https://osf.io/3uf8v/?view_only=ffbc957d6ce941d7b47bef09b628adcd}.", "keywords": [ "Diffusion Models", "Codebook Quantization", "Model Quantization", "Model Compression" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "Xw86qj6FV5", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Daniel Levine", "gender": "unknown", "institution": "Yale University", "country": "US", "position": "Postdoc" }, { "name": "David Zhang", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Undergrad student" }, { "name": "David van Dijk", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" }, { "name": "Emanuele Zappala", "gender": "Male", "institution": "Idaho State University", "country": "US", "position": "Assistant Professor" }, { "name": "Ivan Vrkic", "gender": "unknown", "institution": "Yale University", "country": "US", "position": "Postgraduate Researcher" }, { "name": "Marco Francesco Bressana", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Research Affiliate" }, { "name": "Sizhuang He", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Syed A Rizvi", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Yangtian Zhang", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 15, "n_ref_uni": 48, "n_ref": 82, "n_ref_all": 106, "n_fig": 7, "n_tab": 7, "L_tab": 2658, "n_element_tab": 163, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1260, "n_element_tab_1": 88, "formula_len_all": 2587, "formula_len_all_1": 706, "len_all": 177745, "len_all_1": 59897, "len_abs": 1062, "len_title": 109, "len_sents": 50894, "len_sents_1": 26935, "n_sents": 367, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1065, "title": "CaLMFlow: Volterra Flow Matching using Causal Language Models", "abs": "We introduce CaLMFlow (Causal Language Models for Flow Matching), a novel framework that casts flow matching as a Volterra integral equation (VIE), leveraging the power of large language models (LLMs) for continuous data generation. CaLMFlow enables the direct application of LLMs to learn complex flows by formulating flow matching as a sequence modeling task, bridging discrete language modeling and continuous generative modeling. Our method implements tokenization across space and time, thereby solving a VIE over these domains. This approach enables efficient handling of high-dimensional data and outperforms ODE solver-dependent methods like conditional flow matching (CFM). We demonstrate CaLMFlow's effectiveness on synthetic and real-world data, including single-cell perturbation response prediction, showcasing its ability to incorporate textual context and generalize to unseen conditions. Our results highlight LLM-driven flow matching as a promising paradigm in generative modeling, offering improved scalability, flexibility, and context-awareness.", "keywords": [ "Flow matching", "operator learning", "large language models", "single-cell transcriptomics", "causal language models", "integral equations" ], "rating_list": [ 3, 8, 3 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 1 ], "contribution_list": [ 1, 3, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "Xw0fCEMFss", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Changxin Gao", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Huayu Zhang", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Nong Sang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Siyi Jiao", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Wenzheng Zeng", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yerong Li", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Zheng Shou", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 19, "n_ref_uni": 33, "n_ref": 52, "n_ref_all": 65, "n_fig": 4, "n_tab": 5, "L_tab": 1138, "n_element_tab": 170, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 617, "n_element_tab_1": 116, "formula_len_all": 947, "formula_len_all_1": 852, "len_all": 113323, "len_all_1": 67130, "len_abs": 2023, "len_title": 146, "len_sents": 32876, "len_sents_1": 31337, "n_sents": 218, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1791, "title": "MP-Mat: A 3D-and-Instance-Aware Matting Framework with Multiplane Representation", "abs": "Human instance matting aims to estimate an alpha matte for each human instance in an image, which is challenging as it easily fails in complex cases requiring disentangling mingled pixels belonging to multiple instances along hairy and thin boundary structures. In this work, we address this by introducing a novel 3D-and-instance-aware matting framework with multiplane representation, where the multiplane concept is designed from two different perspectives: scene geometry level and instance level. Specifically, we first build feature-level multiplane representations to split the scene into multiple planes based on depth differences. This approach makes the scene representation 3D-aware, and can serve as an effective clue for splitting instances in different 3D positions, thereby improving interpretability and boundary handling ability especially in occlusion areas. Then, we introduce another multiplane representation that splits the scene in an instance-level perspective, and represents each instance with both matte and color. We also treat background as a special instance, which is often overlooked by existing methods. Such an instance-level representation facilitates both foreground and background content awareness, and is useful for other down-stream tasks like image editing. Once built, the representation can be reused to realize controllable instance-level image editing with high efficiency. Extensive experiments validate the clear advantage of MP-Mat in matting task. We also demonstrate its superiority in image editing tasks, an area under-explored by existing matting-focused methods, where our approach under zero-shot inference even outperforms trained specialized image editing techniques by large margins. Code will be released to inspire relevant fields.", "keywords": [ "image matting", "image editing" ], "rating_list": [ 6, 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 4, 3 ] }, { "paper_id": "XvA1Mn9OFy", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Akhilan Boopathy", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Ila R Fiete", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Professor" } ], "n_formula": 41, "n_formula_1": 17, "n_ref_uni": 19, "n_ref": 40, "n_ref_all": 49, "n_fig": 6, "n_tab": 1, "L_tab": 5, "n_element_tab": 1, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 5, "n_element_tab_1": 1, "formula_len_all": 3022, "formula_len_all_1": 877, "len_all": 107118, "len_all_1": 52288, "len_abs": 1297, "len_title": 81, "len_sents": 39282, "len_sents_1": 24516, "n_sents": 305, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 33, "L_abs": 1286, "title": "Generalization Aware Minimization", "abs": "Sharpness-Aware Minimization (SAM) algorithms have effectively improved neural network generalization by steering model parameters away from sharp regions of the training loss landscape, which tend to generalize poorly. However, the underlying mechanisms of SAM are not fully understood, and recent studies question whether its bias toward flatter regions is why it improves generalization. In this work, we introduce Generalization-Aware Minimization (GAM), a generalized version of SAM that employs multiple perturbation steps instead of SAM's single-step perturbations. This allows GAM to directly guide model parameters toward areas of the landscape that generalize better. We show that the expected true (test) loss landscape is a rescaled version of the observed training loss landscape and demonstrate how GAM's multiple perturbative updates can be designed to optimize this expected true loss. Finally, we present a practical online algorithm that adapts GAM's perturbative steps during training to improve generalization, and we empirically validate its superior performance over SAM on benchmark datasets. We believe GAM sheds light on the generalization improvements of sharpness-based algorithms and can inspire the development of optimizers with even better generalization.", "keywords": [ "generalization", "sharpness aware minimization", "loss landscape", "optimization" ], "rating_list": [ 3, 3, 6, 6, 6 ], "soundness_list": [ 2, 1, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3, 3 ] }, { "paper_id": "Xv6djzJKl3", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Guoxin Xiong", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Rui Sun", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Tianzhu Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yuan Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zhaoyang Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 13, "n_ref_uni": 46, "n_ref": 96, "n_ref_all": 125, "n_fig": 7, "n_tab": 8, "L_tab": 1728, "n_element_tab": 220, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 114, "n_element_tab_1": 38, "formula_len_all": 521, "formula_len_all_1": 523, "len_all": 153496, "len_all_1": 63129, "len_abs": 1461, "len_title": 110, "len_sents": 31654, "len_sents_1": 27206, "n_sents": 224, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1461, "title": "Task-specific Meta-feature Selection for Few-shot Segmentation", "abs": "Few-shot segmentation (FSS) aims to segment new category images given only a few labeled samples. Most previous works concentrate on the design of intricate query decoders to perform feature matching or aggregation between the support and query. In this paper, we revisit a widely overlooked aspect of existing FSS methods, i.e., the exploration of pretrained backbone features. We find that treating all feature channels equally is suboptimal and propose a Task-specific Channel-wise Modulation Network (TCMNet) to focus more attention on task-aware channels, facilitating more effective utilization of pre-trained features. The proposed TCMNet enjoys several merits. First, we design a self-modulation block that injects the gradient information into channel-wise attention layers, thereby enhancing the discriminability between target and background features. Second, a cross-calibration block is introduced to align the support features toward the query according to the target gradient and representations, which mitigates the impact of intra-class diversity. Extensive experimental results on COCO-20i and Pascal-5i benchmarks demonstrate that the TCMNet, as a general plugin, consistently achieves significant improvements over different query decoders and also achieves state-of-the-art results. In addition, the decent performance achieved by exploring the backbone features may inspire another direction for developing more comprehensive FSS models.", "keywords": [ "Computer Vision", "Semantic Segmentation", "Few-Shot Segmentation" ], "rating_list": [ 1, 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "Xuyp1dGAbi", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Jianxin Wang", "gender": "unknown", "institution": "Central South University", "country": "CN", "position": "Full Professor" }, { "name": "Jinhui Xu", "gender": "Male", "institution": "University at Buffalo, State University of New York", "country": "US", "position": "Full Professor" }, { "name": "Junyu Huang", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "Lecturer" }, { "name": "Qilong Feng", "gender": "Male", "institution": "Central South University, China", "country": "CN", "position": "Full Professor" }, { "name": "Zhen Zhang", "gender": "unknown", "institution": "Hunan University of Technology and Business", "country": "CN", "position": "Lecturer" }, { "name": "Ziyun Huang", "gender": "Male", "institution": "Pennsylvania State University, Erie", "country": "US", "position": "Assistant Professor" } ], "n_formula": 124, "n_formula_1": 14, "n_ref_uni": 15, "n_ref": 53, "n_ref_all": 73, "n_fig": 2, "n_tab": 15, "L_tab": 26803, "n_element_tab": 1380, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 185, "n_element_tab_1": 12, "formula_len_all": 17518, "formula_len_all_1": 1340, "len_all": 226624, "len_all_1": 74332, "len_abs": 1590, "len_title": 105, "len_sents": 61628, "len_sents_1": 33114, "n_sents": 536, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1591, "title": "New Algorithms for the Learning-Augmented k-means Problem", "abs": "In this paper, we study the clustering problems in the learning-augmented setting, where predicted labels for a d-dimensional dataset with size m are given by an oracle to serve as auxiliary information to improve the clustering performance. Following the prior work, the given oracle is parameterized by some error rate \u03b1, which captures the accuracy of the oracle such that there are at most \u03b1 fraction of false positives and false negatives in each predicted cluster. In this setting, the goal is to design fast and practical algorithms that can break the computational barriers of inapproximability. The current state-of-the-art learning-augmented k-means algorithm relies on sorting strategies to find good coordinates approximation, where a (1+O(\u03b1))-approximation can be achieved with near-linear running time in the data size. However, the computational demands for sorting may limit the scalability of the algorithm for handling large-scale datasets. To address this issue, in this paper, we propose new algorithms that can identify good coordinates approximation using sampling-based strategies, where (1+O(\u03b1))-approximation can be achieved with linear running time in the data size. To obtain a more practical algorithm for the problem with better clustering quality and running time, we propose a sampling-based heuristic which can directly find center approximations using sampling-based strategies. Empirical experiments show that our proposed methods are faster than the state-of-the-art learning-augmented k-means algorithms with comparable performances on clustering quality.", "keywords": [ "Learning-Augmented Clustering; Approximation Algorithm;" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 5, 3, 3 ] }, { "paper_id": "XuYd9IK7X4", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Argyris Oikonomou", "gender": "unknown", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Arman Cohan", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" }, { "name": "Weiqiang Zheng", "gender": "Male", "institution": "Yale University", "country": "", "position": "PhD student" }, { "name": "Yang Cai", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Yixin Liu", "gender": "unknown", "institution": "Yale University", "country": "US", "position": "PhD student" } ], "n_formula": 50, "n_formula_1": 22, "n_ref_uni": 41, "n_ref": 136, "n_ref_all": 159, "n_fig": 3, "n_tab": 9, "L_tab": 1301, "n_element_tab": 117, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 264, "n_element_tab_1": 28, "formula_len_all": 5268, "formula_len_all_1": 2569, "len_all": 229889, "len_all_1": 66345, "len_abs": 1213, "len_title": 125, "len_sents": 55837, "len_sents_1": 29320, "n_sents": 441, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1233, "title": "COMAL: A Convergent Meta-Algorithm for Aligning LLMs with General Preferences", "abs": "Many alignment methods, including reinforcement learning from human feedback (RLHF), rely on the Bradley-Terry reward assumption, which is insufficient to capture the full range of general human preferences. To achieve robust alignment with general preferences, we model the alignment problem as a two-player zero-sum game, where the Nash equilibrium policy guarantees a 50\\% win rate against any competing policy. However, previous algorithms for finding the Nash policy either diverge or converge to a Nash policy in a modified game, even in a simple synthetic setting, thereby failing to maintain the 50\\% win rate guarantee against all other policies. We propose a meta-algorithm, **Co**vergent **M**eta **Al**ignment Algorithm (COMAL), for language model alignment with general preferences, inspired by convergent algorithms in game theory. Theoretically, we prove that our meta-algorithm converges to an exact Nash policy. Additionally, our meta-algorithm is simple and can be integrated with many existing methods designed for RLHF and preference optimization with minimal changes. Experimental results demonstrate the effectiveness of the proposed framework when combined with existing preference policy optimization methods.", "keywords": [ "Alignment", "General Preferences", "Large Language Model", "Nash Equilibrium" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "XuQJ5a3sTb", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Shenyu Lu", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Xiaoqian Wang", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 48, "n_ref": 99, "n_ref_all": 128, "n_fig": 11, "n_tab": 4, "L_tab": 2597, "n_element_tab": 94, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2508, "n_element_tab_1": 88, "formula_len_all": 729, "formula_len_all_1": 634, "len_all": 160062, "len_all_1": 71917, "len_abs": 1520, "len_title": 124, "len_sents": 51434, "len_sents_1": 32367, "n_sents": 436, "n_sents_1": 253, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1534, "title": "Learning Disentangled Representations for Fairness with Limited Demographics", "abs": "Fair representation learning is a promising way to mitigate discrimination in downstream tasks. Many existing fair representation learning methods require access to sensitive information, but the collection of sensitive information is often difficult and even involves privacy issues. Additionally, a model trained to be fair with respect to one sensitive attribute may not ensure fairness for other sensitive groups. Thus, how to flexibly address fairness issues when we have limited access to sensitive information is a challenging problem. In this work, we answer this question: ``given limited sensitive information, can we learn a representation to be fair w.r.t. varying sensitive groups?'' To achieve this, we propose a novel two-step framework. We first learn a disentangled representation by employing Non-linear Independent Component Analysis (Nonlinear ICA). Second, we remove sensitive information in the latent space to obtain fair representation. The learned representation can be easily adapted to be fair w.r.t different sensitive groups and to be used for different downstream tasks without re-training. Among the entire process, only a small portion of sensitive information is required in the second step to learn a fair representation. We compare with methods that require different amounts of sensitive information on real-world images and tabular datasets. We empirically demonstrate the utility and flexibility of our approach, and our method is capable of achieving improved fairness results in various tasks.", "keywords": [ "Fair representation learning", "Disentanglement" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "XtY3xYQWcW", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Giovanni Luca Marchetti", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "Postdoc" }, { "name": "Kathl\u00e9n Kohn", "gender": "Female", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "Associate Professor" }, { "name": "Nathan Henry", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" } ], "n_formula": 59, "n_formula_1": 26, "n_ref_uni": 19, "n_ref": 29, "n_ref_all": 57, "n_fig": 4, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4087, "formula_len_all_1": 1919, "len_all": 125033, "len_all_1": 66679, "len_abs": 773, "len_title": 115, "len_sents": 44062, "len_sents_1": 27600, "n_sents": 404, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 705, "title": "Geometry of Lightning Self-Attention: Identifiability and Dimension", "abs": "We consider function spaces defined by self-attention networks without normalization, and theoretically analyze their geometry. Since these networks are polynomial, we rely on tools from algebraic geometry. In particular, we study the identifiability of deep attention by providing a description of the generic fibers of the parametrization for an arbitrary number of layers and, as a consequence, compute the dimension of the function space. Additionally, for a single-layer model, we characterize the singular and boundary points. Finally, we formulate a conjectural extension of our results to normalized self-attention networks, prove it for a single layer, and numerically verify it in the deep case.", "keywords": [ "Lightning Self-Attention", "Neuromanifolds", "Algebraic Geometry" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 4, 4, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 2, 2, 4, 2 ] }, { "paper_id": "XtXa6hoNrU", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Jingyang Xiang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Sai Qian Zhang", "gender": "unknown", "institution": "New York University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 33, "n_ref": 67, "n_ref_all": 100, "n_fig": 14, "n_tab": 8, "L_tab": 4213, "n_element_tab": 358, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 2179, "n_element_tab_1": 140, "formula_len_all": 975, "formula_len_all_1": 1005, "len_all": 135006, "len_all_1": 63221, "len_abs": 1977, "len_title": 145, "len_sents": 37743, "len_sents_1": 26859, "n_sents": 304, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1976, "title": "DFRot: Achieving Outlier-Free and Massive Activation-Free for Rotated LLMs with Refined Rotation", "abs": "Using rotational invariance to eliminate outliers in large language models (LLMs) has recently gained considerable attention, especially in the context of quantization. Prior studies have shown that in low-precision quantization scenarios, such as 4-bit weights and 4-bit activations~(W4A4), randomized Hadamard transforms can achieve significantly higher accuracy than randomized orthogonal transforms. Notably, the reason behind this phenomena remains unknown. In this paper, we find that these transformations show substantial improvement in eliminating outliers for common tokens and achieve similar quantization error. The primary reason for the accuracy difference lies in the fact that randomized Hadamard transforms can slightly reduce the quantization error for tokens with massive activations while randomized orthogonal transforms increase the quantization error. Due to the extreme rarity of these tokens and their critical impact on model accuracy, we consider this a long-tail optimization problem, and therefore construct a simple yet effective method: a weighted loss function. Additionally, we propose an optimization strategy for the rotational matrix that involves alternating optimization of quantization parameters while employing orthogonal Procrustes transforms to refine the orthogonal matrix. This makes the distribution of the rotated activation values more conducive to quantization, especially for tokens with massive activations. Our method enhances the Rotated LLMs by achieving dual free, \\textit{Outlier-Free} and \\textit{Massive Activation-Free}, dubbed as DFRot. Extensive experiments demonstrate the effectiveness and efficiency of DFRot. By tuning the rotational matrix using just a single sample, DFRot achieves a perplexity improvement of 0.25 and 0.21 on W4A4KV4 and W4A4KV16, respectively, for LLaMA3-8B, a model known for its quantization challenges. Code is anonymously available at \\url{https://anonymous.4open.science/r/DFRot-8FE3}.", "keywords": [ "large language models", "rotational", "quantization", "orthogonal Procrustes transforms", "Outlier-Free and Massive Activation-Free" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "XsgHl54yO7", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Hunter Nisonoff", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Junhao Xiong", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Stephan Allenspach", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Jennifer Listgarten", "gender": "Female", "institution": "University of California Berkeley", "country": "", "position": "Full Professor" } ], "n_formula": 72, "n_formula_1": 8, "n_ref_uni": 85, "n_ref": 276, "n_ref_all": 374, "n_fig": 29, "n_tab": 12, "L_tab": 7488, "n_element_tab": 733, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1120, "n_element_tab_1": 13, "formula_len_all": 9240, "formula_len_all_1": 442, "len_all": 396700, "len_all_1": 69088, "len_abs": 841, "len_title": 117, "len_sents": 127870, "len_sents_1": 35502, "n_sents": 830, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 844, "title": "Unlocking Guidance for Discrete State-Space Diffusion and Flow Models", "abs": "Generative models on discrete state-spaces have a wide range of potential applications, particularly in the domain of natural sciences. In continuous state-spaces, controllable and flexible generation of samples with desired properties has been realized using guidance on diffusion and flow models. However, these guidance approaches are not readily amenable to discrete state-space models. Consequently, we introduce a general and principled method for applying guidance on such models. Our method depends on leveraging continuous-time Markov processes on discrete state-spaces, which unlocks computational tractability for sampling from a desired guided distribution. We demonstrate the utility of our approach, Discrete Guidance, on a range of applications including guided generation of small-molecules, DNA sequences and protein sequences.", "keywords": [ "discrete state-space generative models", "diffusion", "flow-matching", "flow models", "guidance", "protein design" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "XsYJ6yvgEC", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Ani Calinescu", "gender": "Not Specified", "institution": "Department of Computer Science, University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Jakob Nicolaus Foerster", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Kang Li", "gender": "Male", "institution": "Oxofrd, University of Oxford", "country": "", "position": "PhD student" }, { "name": "Peer Nagy", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Sascha Yves Frey", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Stefan Zohren", "gender": "", "institution": "University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Svitlana Vyetrenko", "gender": "unknown", "institution": "J.P. Morgan Chase", "country": "US", "position": "AI Research Director" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 31, "n_ref": 51, "n_ref_all": 62, "n_fig": 4, "n_tab": 1, "L_tab": 159, "n_element_tab": 2, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 682, "formula_len_all_1": 682, "len_all": 111693, "len_all_1": 66049, "len_abs": 1729, "len_title": 146, "len_sents": 38741, "len_sents_1": 32792, "n_sents": 292, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1632, "title": "LOB-Bench: Benchmarking Generative AI for Finance - with an Application to Limit Order Book Markets", "abs": "We present **LOB-Bench**, a benchmark designed to evaluate the quality and realism of generative message-by-order data for limit order books (LOB). We enable a rigorous and comprehensive model comparison by providing both a theoretical framework and an open-source Python package. Addressing the lack of consensus on evaluation paradigms in the literature, where qualitative comparison of stylized facts is prevalent, our work offers a crucial building block for advancing generative AI for financial data. LOB-Bench provides a standardized method to numerically assess the quality of various model classes that generate limit order book data in the widely used LOBSTER format. It provides a range of quantitative characteristics and includes a simple parametric benchmark model as a baseline. Our framework measures distributional differences in conditional and unconditional statistics between generated and real LOB data, supporting a flexible multivariate statistical evaluation across different model classes. The benchmark features commonly used LOB statistics such as spread, order book volumes, order imbalance, and message inter-arrival times, along with adversarial scores derived from a neural network trained to differentiate between real and generated data. Additionally, LOB-Bench evaluates \"market impact metrics\" by computing cross-correlations and price response functions for specific events in the data. We present empirical benchmark results for a generative autoregressive state-space model, for a (C)GAN, and parametric LOB model. We find that the autoregressive GenAI approach beats traditional model classes.", "keywords": [ "finance", "generative models", "time series", "state-space models", "benchmark" ], "rating_list": [ 3, 1, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "XsXHqEVtiB", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Kenji Kawashima", "gender": "Male", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Maina Sogabe", "gender": "Female", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "JP", "position": "Assistant Professor" }, { "name": "Niran Nataraj", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 18, "n_ref_uni": 33, "n_ref": 44, "n_ref_all": 59, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 793, "formula_len_all_1": 819, "len_all": 118007, "len_all_1": 56575, "len_abs": 2050, "len_title": 154, "len_sents": 33712, "len_sents_1": 28258, "n_sents": 213, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 2055, "title": "Let\u2019s Stop Bleeding! Precise Bleeding Data Estimation & Visualization Methods for Laparoscopic Surgeries", "abs": "Intraoperative bleeding remains a significant challenge in modern surgery, necessitating rapid and accurate localization of bleeding sources to ensure effective hemostasis. Proactive detection and timely intervention are critical for minimizing blood loss, reducing operative time, preventing complications, and decreasing the need for intensive postoperative care. In this research, we introduce Selective Bleeding Alert Map (SBAM), a novel GAN-based framework designed for precise real-time detection of bleeding origins during surgery. Building upon our earlier BAM framework, SBAM shifts from broad, area-wide alerts to a focused approach that highlights only the exact bleeding areas, enhancing visual accuracy and potentially improving surgeon focus and visibility\u2014particularly beneficial in cases of minor bleeding where excessive alerts could interfere with the surgical process. To achieve this, we developed advanced image-to-image translation and segmentation models, custom thresholding techniques, and trajectory detection algorithms to pinpoint bleeding sources with high precision. Utilizing our developed mimic organ system for ethically sourced, realistic datasets\u2014alongside synthetic data generated from the orGAN system and Large Mask Inpainting (LaMa)\u2014we created a dedicated dataset specifically for SBAM training, including over 1,000 manually annotated images capturing both bleeding and non-bleeding regions within marked bleeding areas. Our instance segmentation model achieved a precision of 92.5%, an accuracy of 98% and a mask mean Average Precision of 85% at an IoU threshold of 0.5 (mAP@50). Additionally, the SBAM model demonstrated high accuracy in detecting bleeding points within real surgical videos from the Hamlyn dataset, underscoring its potential for practical surgical applications.Powered by core algorithms and uniquely developed datasets, SBAM represents a pivotal advancement in AI-assisted surgery, demonstrating superior performance in detecting bleeding regions with high precision during critical scenarios.", "keywords": [ "Medical imaging", "surgical image", "GAN", "image to image translation", "segmentation" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 5, 3 ] }, { "paper_id": "XrwsdcgWKc", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Guillaume Rabusseau", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Associate Professor" }, { "name": "Jun Dai", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "", "position": "Postdoc" }, { "name": "Michael Rizvi-Martel", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 16, "n_ref_uni": 43, "n_ref": 73, "n_ref_all": 89, "n_fig": 8, "n_tab": 3, "L_tab": 446, "n_element_tab": 18, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 470, "n_element_tab_1": 6, "formula_len_all": 1199, "formula_len_all_1": 915, "len_all": 152526, "len_all_1": 61422, "len_abs": 1581, "len_title": 89, "len_sents": 39987, "len_sents_1": 26632, "n_sents": 302, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1428, "title": "Enhancing Variational Quantum Algorithms: Effective Quantum Ansatz Design Using GFlowNets", "abs": "Quantum computing promises significant computational advantages over classical computing. However, current devices are constrained by a limited qubit count and noise. By combining classical optimization methods with parameterized quantum circuits, Variational Quantum Algorithms (VQAs) offer a potential solution for noisy intermediate-scale quantum systems (NISQ). This makes VQAs particularly promising strategies for achieving near-term quantum advantages; such approaches are now widely explored for nearly all quantum computing applications. However, designing effective parameterized circuits, also known as ansatz, remains challenging. In this work, we introduce the use of GFlowNets as an efficient method to automate the development of efficient ansatz for various quantum computing problems. Our approach leverages GFlowNets to efficiently explore the combinatorial space of parameterized quantum circuits. Our extensice experiments demonstrate that GFlowNets can discover ansatz with an order of magnitude fewer parameters, gate counts, and depths compared to current approaches for the molecular electronic ground state energy problem. We also apply our approach to the unweighted Max-Cut problem, where we observe similar improvements in circuit efficiency. These results highlight the potential of GFlowNets to significantly reduce the resource requirements of VQAs while maintaining or improving solution quality.", "keywords": [ "quantum computing", "GFlowNets", "Classical-Quantum Hybrid Algorithms", "Variational Quantum Algorithms (VQAs)" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 5, 5 ] }, { "paper_id": "XrtFVM1f6w", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Andrej Risteski", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Ankur Moitra", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "" }, { "name": "Dhruv Rohatgi", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Jianfeng Lu", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Professor" }, { "name": "Tanya Marwah", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Zachary Lipton", "gender": "Unspecified", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" } ], "n_formula": 42, "n_formula_1": 12, "n_ref_uni": 24, "n_ref": 41, "n_ref_all": 61, "n_fig": 4, "n_tab": 1, "L_tab": 176, "n_element_tab": 31, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 254, "n_element_tab_1": 20, "formula_len_all": 4213, "formula_len_all_1": 1166, "len_all": 164783, "len_all_1": 71881, "len_abs": 923, "len_title": 124, "len_sents": 65493, "len_sents_1": 35710, "n_sents": 512, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 928, "title": "Towards characterizing the value of edge embeddings in graph neural networks", "abs": "Graph neural networks (GNNs) are the dominant approach to solving machine learning problems defined over graphs. Despite much theoretical and empirical work in recent years, our understanding of finer-grained aspects of architectural design for GNNs remains impoverished. In this paper, we consider the benefits of architectures that maintain and update edge embeddings. On the theoretical front, under a suitable computational abstraction for a layer in the model, as well as memory constraints on the embeddings, we show that there are natural tasks on graphical models for which architectures leveraging edge embeddings can be much shallower. Our techniques are inspired by results on time-space tradeoffs in theoretical computer science. Empirically, we show architectures that maintain edge embeddings almost always improve on their node-based counterparts---frequently significantly so in topologies that have \"hub\" nodes.", "keywords": [ "graph neural networks", "theory", "representational power", "communication complexity", "memory tradeoffs", "edge embeddings" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 2 ] }, { "paper_id": "XrsOu4KgDE", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Arnav Goel", "gender": "Male", "institution": "Indraprastha Institute of Information Technology, Delhi", "country": "IN", "position": "Undergrad student" }, { "name": "Huihan Li", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Keyu He", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Undergrad student" }, { "name": "Xiang Ren", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 30, "n_ref": 56, "n_ref_all": 93, "n_fig": 19, "n_tab": 5, "L_tab": 1624, "n_element_tab": 89, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 521, "n_element_tab_1": 29, "formula_len_all": 609, "formula_len_all_1": 369, "len_all": 153700, "len_all_1": 56808, "len_abs": 2121, "len_title": 114, "len_sents": 52019, "len_sents_1": 26574, "n_sents": 384, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1215, "title": "Attributing Culture-Conditioned Generations to Pretraining Corpora", "abs": "In open-ended generative tasks like narrative writing or dialogue, large language models often exhibit cultural biases, showing limited knowledge and generating templated outputs for less prevalent cultures. Recent works show that these biases may stem from uneven cultural representation in pretraining corpora. This work investigates how pretraining leads to biased culture-conditioned generations\nby analyzing how models associate entities with cultures based on pretraining data patterns. We propose the MEMOED framework (MEMOrization from prEtraining Document) to determine whether a generation for a culture arises from memorization. Using MEMOED on culture-conditioned generations about food and clothing for 110 cultures, we find that high-frequency cultures in pretraining data yield more generations with memorized symbols, while some low-frequency cultures produce none. Additionally, the model favors generating entities with extraordinarily high frequency regardless of the conditioned culture, reflecting biases toward frequent pretraining terms irrespective of relevance. We hope that the MEMOED framework and our insights will inspire more works on attributing model performance on pretraining data.", "keywords": [ "culture bias", "pretraining data", "memorization", "generalization" ], "rating_list": [ 8, 8, 8, 6, 5 ], "soundness_list": [ 3, 3, 4, 4, 2 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 2, 4, 3, 3, 3 ] }, { "paper_id": "Xr5iINA3zU", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Apratim Dey", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "David L. Donoho", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Full Professor" }, { "name": "Joshua Kazdan", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Matthias Gerstgrasser", "gender": "unknown", "institution": "OpenAI", "country": "US", "position": "Researcher" }, { "name": "Rafael Rafailov", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Rylan Schaeffer", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" } ], "n_formula": 41, "n_formula_1": 13, "n_ref_uni": 22, "n_ref": 45, "n_ref_all": 104, "n_fig": 5, "n_tab": 3, "L_tab": 565, "n_element_tab": 31, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 211, "n_element_tab_1": 37, "formula_len_all": 3143, "formula_len_all_1": 989, "len_all": 172254, "len_all_1": 52511, "len_abs": 1821, "len_title": 132, "len_sents": 38264, "len_sents_1": 25599, "n_sents": 296, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1847, "title": "Collapse or Thrive? Perils and Promises of Synthetic Data in a Self-Generating World", "abs": "The increasing presence of AI-generated content on the internet raises a critical question: What happens when generative machine learning models are pretrained on web-scale datasets containing data created by earlier models?\nSome authors prophesy _model collapse_ under a '_replace_' scenario: a sequence of models, the first trained with real data and each later one trained _only on_ synthetic data from its preceding model. In this scenario, models successively degrade. Others see collapse as avoidable; in an '_accumulate_' scenario, a sequence of models is trained, but each training uses all real and synthetic data generated so far. In this work, we deepen and extend the study of these contrasting scenarios. First, collapse versus avoidance of collapse is studied\nby comparing the replace and accumulate scenarios on each of three prominent generative modeling settings; we find the same contrast emerges in all three settings. Second, we study a compromise scenario; the available data remains the same as in the _accumulate_ scenario -- but unlike _accumulate_ and like _replace_, each model is trained using a fixed compute budget; we demonstrate that model test loss on real data is larger than in the _accumulate_ scenario, but apparently plateaus, unlike the divergence seen with _replace_.\nThird, we study the relative importance of cardinality and proportion of real data for avoiding model collapse. Surprisingly, we find a non-trivial interaction between real and synthetic data, where the value of synthetic data for reducing test loss depends on the absolute quantity of real data. Our insights are particularly important when forecasting whether future frontier generative models will collapse or thrive, and our results open avenues for empirically and mathematically studying the context-dependent value of synthetic data.", "keywords": [ "model collapse", "model-data feedback loops", "synthetic data", "sampling bias", "deep generative models", "model misbehavior" ], "rating_list": [ 6, 3, 6, 8 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "Xqo4eObgQX", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Fangyikang Wang", "gender": "Male", "institution": "Wechat, Tencent Inc", "country": "CN", "position": "Intern" }, { "name": "Kunchang Li", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Shaobin Zhuang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xinyuan Chen", "gender": "Female", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Yali Wang", "gender": "Male", "institution": "SIAT, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yanbo Ding", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CL", "position": "MS student" }, { "name": "Yaohui Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Ying Zhang", "gender": "Female", "institution": "Tencent", "country": "CN", "position": "Senior Researcher" }, { "name": "Yiwei Guo", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Chen Li", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Senior Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 43, "n_ref": 105, "n_ref_all": 108, "n_fig": 1, "n_tab": 9, "L_tab": 6821, "n_element_tab": 801, "n_fig_1": 1, "n_tab_1": 9, "L_tab_1": 5866, "n_element_tab_1": 507, "formula_len_all": 581, "formula_len_all_1": 581, "len_all": 178122, "len_all_1": 71800, "len_abs": 1820, "len_title": 166, "len_sents": 29373, "len_sents_1": 27524, "n_sents": 247, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 117, "L_abs": 1844, "title": "TimeStep Master: Asymmetrical Mixture of Timestep LoRA Experts for Versatile and Efficient Diffusion Models in Vision", "abs": "Diffusion models have driven the advancement of vision generation over the past years. \nHowever, \nit is often difficult to apply these large models in downstream tasks, \ndue to massive fine-tuning cost.\nRecently, \nLow-Rank Adaptation (LoRA) has been applied for efficient tuning of diffusion models.\nUnfortunately, \nthe capabilities of LoRA-tuned diffusion models are limited, \nsince the same LoRA is used for different timesteps of the diffusion process.\nTo tackle this problem, \nwe introduce a general and concise TimeStep Master (TSM) paradigm with two key fine-tuning stages.\nIn the fostering stage (1-stage), \nwe apply different LoRAs to fine-tune the diffusion model at different timestep intervals.\nThis results in different TimeStep LoRA experts that can effectively capture different noise levels. \nIn the assembling stage (2-stage),\nwe design a novel asymmetrical mixture of TimeStep LoRA experts,\nvia core-context collaboration of experts at multi-scale intervals.\nFor each timestep,\nwe leverage TimeStep LoRA expert within the smallest interval as the core expert without gating, \nand use experts within the bigger intervals as the context experts with time-dependent gating.\nConsequently,\nour TSM can effectively model the noise level via the expert in the finest interval,\nand \nadaptively integrate contexts from the experts of other scales,\nboosting the versatility of diffusion models.\nTo show the effectiveness of our TSM paradigm, \nwe conduct extensive experiments on three typical and popular LoRA-related tasks of diffusion models, \nincluding \ndomain adaptation, \npost-pretraining,\nand \nmodel distillation.\nOur TSM achieves the state-of-the-art results on all these tasks, \nthroughout various model structures (UNet, DiT and MM-DiT) and visual data modalities (Image, Video), \nshowing its remarkable generalization capacity.", "keywords": [ "Visual Generation", "Diffusion Model", "LoRA" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "XqDM97DtMf", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Sunbochen Tang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Themistoklis Sapsis", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Navid Azizan Ruhi", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" } ], "n_formula": 23, "n_formula_1": 10, "n_ref_uni": 25, "n_ref": 45, "n_ref_all": 64, "n_fig": 14, "n_tab": 1, "L_tab": 6, "n_element_tab": 1, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2209, "formula_len_all_1": 562, "len_all": 113459, "len_all_1": 67816, "len_abs": 1687, "len_title": 101, "len_sents": 45871, "len_sents_1": 35111, "n_sents": 300, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1693, "title": "Learning Chaotic Dynamics with Embedded Dissipativity", "abs": "Chaotic dynamics, commonly seen in weather systems and fluid turbulence, are characterized by their sensitivity to initial conditions, which makes accurate prediction challenging. Despite its sensitivity to initial perturbations, many chaotic systems observe dissipative behaviors and ergodicity. Therefore, recently various approaches have been proposed to develop data-driven models preserving invariant statistics over long horizons. Although these methods have shown empirical success in reducing instances of unbounded trajectory generation, many of the models are still prone to generating unbounded trajectories, leading to invalid statistics evaluation. In this paper, we propose a novel neural network architecture that simultaneously learns a dissipative dynamics emulator that guarantees to generate bounded trajectories and an energy-like function that governs the dissipative behavior. More specifically, by leveraging control-theoretic ideas, we derive algebraic conditions based on the learned energy-like function that ensure asymptotic convergence to an invariant level set. Using these algebraic conditions, our proposed model enforces dissipativity through a ReLU projection layer, which provides formal trajectory boundedness guarantees. Furthermore, the invariant level set provides an outer estimate for the strange attractor, which is known to be very difficult to characterize due to its complex geometry. We demonstrate the capability of our model in producing bounded long-horizon trajectory forecasts that preserve invariant statistics and characterizing the attractor, for chaotic dynamical systems including Lorenz 96 and a truncated Kuramoto-Sivashinsky equation.", "keywords": [ "Dynamical systems", "learning for dynamics", "chaotic dynamics" ], "rating_list": [ 3, 5, 3, 3, 8 ], "soundness_list": [ 2, 2, 1, 2, 4 ], "presentation_list": [ 3, 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 1, 2, 3 ], "confidence_list": [ 3, 4, 5, 3, 4 ] }, { "paper_id": "Xq7gwsnhPT", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Chunhong Pan", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Jinyong Wen", "gender": "Male", "institution": "Institute of automation, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Kun Ding", "gender": "Male", "institution": "Institute of automation", "country": "CN", "position": "Assistant Professor" }, { "name": "Shiming Xiang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Tao Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zhen Chen", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 97, "n_ref_all": 157, "n_fig": 18, "n_tab": 14, "L_tab": 6012, "n_element_tab": 840, "n_fig_1": 12, "n_tab_1": 7, "L_tab_1": 2574, "n_element_tab_1": 355, "formula_len_all": 847, "formula_len_all_1": 194, "len_all": 215456, "len_all_1": 67891, "len_abs": 1742, "len_title": 131, "len_sents": 62054, "len_sents_1": 28572, "n_sents": 548, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 82, "L_abs": 1761, "title": "UNIP: Rethinking Pre-trained Attention Patterns for Infrared Semantic Segmentation", "abs": "Pre-training techniques significantly enhance the performance of semantic segmentation tasks with limited training data. However, the efficacy under a large domain gap between pre-training (e.g. RGB) and fine-tuning (e.g. infrared) remains underexplored. In this study, we first benchmark the infrared semantic segmentation performance of various pre-training methods and reveal several phenomena distinct from the RGB domain. Next, our layerwise analysis of pre-trained attention maps uncovers that: (1) There are three typical attention patterns (local, hybrid, and global); (2) Pre-training tasks notably influence pattern distribution across layers; (3) The hybrid pattern is crucial for semantic segmentation as it attends to both nearby and foreground elements; (4) The texture bias impedes model generalization in infrared tasks. Building on these insights, we propose UNIP, a UNified Infrared Pre-training framework, to enhance the pre-trained model performance. This framework uses the hybrid-attention distillation NMI-HAD as the pre-training target, a large-scale mixed dataset InfMix for pre-training, and a last-layer feature pyramid network LL-FPN for fine-tuning. Experimental results show that UNIP outperforms various pre-training methods by up to 13.5% in average mIoU on three infrared segmentation tasks, evaluated using fine-tuning and linear probing metrics. UNIP-S achieves performance on par with MAE-L while requiring only 1/10 of the computational cost. Furthermore, with fewer parameters, UNIP significantly surpasses state-of-the-art (SOTA) infrared or RGB segmentation methods and demonstrates the broad potential for application in other modalities, such as RGB and depth. Our code is available at https://github.com/casiatao/UNIP.", "keywords": [ "Image Pre-training", "Semantic Segmentation", "Infrared Image", "Attention Distillation", "Representation Learning" ], "rating_list": [ 8, 5, 5, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "Xq12wsoNux", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "George Karypis", "gender": "Male", "institution": "University of Minnesota, Minneapolis", "country": "", "position": "Full Professor" }, { "name": "Justin Chiu", "gender": "Male", "institution": "University of Washington", "country": "", "position": "Undergrad student" }, { "name": "Ruixuan Liu", "gender": "Female", "institution": "Emory University", "country": "US", "position": "Postdoc" }, { "name": "Sheng Zha", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Zhiqi Bu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 6, "n_ref_uni": 35, "n_ref": 85, "n_ref_all": 114, "n_fig": 6, "n_tab": 6, "L_tab": 3368, "n_element_tab": 101, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 934, "n_element_tab_1": 44, "formula_len_all": 656, "formula_len_all_1": 447, "len_all": 152564, "len_all_1": 57393, "len_abs": 1261, "len_title": 110, "len_sents": 37764, "len_sents_1": 26223, "n_sents": 265, "n_sents_1": 169, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1260, "title": "Zero redundancy distributed learning with differential privacy", "abs": "Deep learning using large models have achieved great success in a wide range of domains. However, training these models on billions of parameters is very challenging in terms of the training speed, memory cost, and communication efficiency, especially under the privacy-preserving regime with differential privacy (DP). On the one hand, DP optimization has comparable efficiency to the standard non-private optimization on a single GPU, but on multiple GPUs, existing DP distributed learning (such as pipeline parallel) has suffered from significantly worse efficiency. On the other hand, the Zero Redundancy Optimizer (ZeRO) is a state-of-the-art solution to the standard distributed learning, exhibiting excellent training efficiency on large models, but to work compatibly with DP is technically complicated. In this work, we develop a new systematic solution, DP-ZeRO, (I) to scale up the trainable DP model size, e.g. to GPT-100B, (II) to obtain the same computation and communication efficiency as the standard ZeRO, and (III) to enable mixed-precision DP training. Our DP-ZeRO, like the standard ZeRO, has the potential to train models with arbitrary size and is evaluated on the world's largest DP models in terms of the number of trainable parameters.", "keywords": [ "deep learning", "differential privacy", "distributed learning", "system design" ], "rating_list": [ 3, 5, 5, 1 ], "soundness_list": [ 1, 2, 2, 1 ], "presentation_list": [ 1, 2, 1, 1 ], "contribution_list": [ 1, 3, 3, 1 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "XpU1twhp3u", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Jun Zhao", "gender": "Male", "institution": "Fabu Inc.", "country": "CN", "position": "Researcher" }, { "name": "Kaipeng Zhang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Wenqi Shao", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xiaofei He", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" }, { "name": "Yuqi Lin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zheng Yang", "gender": "Male", "institution": "Fabu Inc", "country": "CN", "position": "CTO" }, { "name": "Ping Luo", "gender": "unknown", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 21, "n_ref": 31, "n_ref_all": 55, "n_fig": 12, "n_tab": 8, "L_tab": 2544, "n_element_tab": 227, "n_fig_1": 10, "n_tab_1": 6, "L_tab_1": 2304, "n_element_tab_1": 189, "formula_len_all": 457, "formula_len_all_1": 456, "len_all": 144622, "len_all_1": 58081, "len_abs": 1555, "len_title": 140, "len_sents": 34175, "len_sents_1": 25367, "n_sents": 264, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1578, "title": "MatchMask: Mask-Centric Generative Data Augmentation for Label-Scarce Semantic Segmentation", "abs": "Current semantic segmentation models are very data-hungry and require massive costly pixel-wise human annotations. Generative data augmentation, which scales the train set using generative models, provides a potential remedy. However, existing text-centric methods struggle to generate complex in-distribution data due to the limitations of text descriptions. In this paper, we propose MatchMask, a novel mask-centric generative data augmentation approach tailored for label-scarce semantic segmentation. It leverages a few labeled semantic masks to generate diverse, realistic, and well-aligned image-mask training pairs for semantic segmentation models. Specifically, to adapt existing text-to-image models for semantic image synthesis, we first propose a Gradient Probe Method to investigate the role of each layer in the diffusion model. On this basis, we introduce a Layer-Timestep Adaptive Adapter (LT-Adapter) comprising layer-adaptive cross-attention fusion and time-adaptive LoRA scaling to enable efficient adaption for the critical layers. Meantime, we design a robust relative filtering principle to suppress incorrectly synthesized regions. Moreover, the proposed approach is extended to MatchMask++ in the semi-supervised setting to take advantage of additional unlabeled data. Experimental results on VOC, COCO and ADE20K demonstrate that MatchMask remarkably enhances the performance of segmentation models, surpassing prior data augmentation techniques in various benchmarks, \\eg, 67.5\\%$\\rightarrow$74.3\\% mIoU on VOC. Our code will be made publicly available.", "keywords": [ "Semantic Segmentation", "Generative Data Augmentation", "Label-Scarce" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 5, 3, 5 ] }, { "paper_id": "XoxxZiIJq6", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Chunsheng Xin", "gender": "Male", "institution": "Old Dominion University", "country": "US", "position": "Full Professor" }, { "name": "Hongyi Wu", "gender": "Male", "institution": "University of Arizona", "country": "US", "position": "Full Professor" }, { "name": "Qiao Zhang", "gender": "unknown", "institution": "Chongqing University", "country": "", "position": "Assistant Professor" }, { "name": "Rui Ning", "gender": "unknown", "institution": "Old Dominion University", "country": "US", "position": "Assistant Professor" }, { "name": "Xiangrui Xu", "gender": "Male", "institution": "Old Dominion University", "country": "US", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 19, "n_ref_uni": 30, "n_ref": 48, "n_ref_all": 68, "n_fig": 2, "n_tab": 8, "L_tab": 2733, "n_element_tab": 320, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 511, "n_element_tab_1": 78, "formula_len_all": 1640, "formula_len_all_1": 949, "len_all": 126692, "len_all_1": 65379, "len_abs": 866, "len_title": 96, "len_sents": 36026, "len_sents_1": 29783, "n_sents": 266, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 866, "title": "Comet: A Communication-efficient and Performant Approximation for Private Transformer Inference", "abs": "The prevalent use of Transformer-like models, exemplified by ChatGPT in modern language processing applications, underscores the critical need for enabling private inference essential for many cloud-based services reliant on such models. However, current privacy-preserving frameworks impose significant communication burden, especially for non-linear computation in Transformer model. In this paper, we introduce a novel plug-in method Comet to effectively reduce the communication cost without compromising the inference performance. We second introduce an efficient approximation method to eliminate the heavy communication in finding good initial approximation. We evaluate our Comet on Bert and RoBERTa models with GLUE benchmark datasets, showing up to 3.9 less communication and 3.5 speedups while keep competitive model performance compared to the prior art.", "keywords": [ "private inference", "secret sharing", "Transformer", "language model", "homomorphic encryption", "multi-party computation" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "XoulHHQGFi", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Ao Qu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Baptiste Freydt", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Intern" }, { "name": "Cameron Hickert", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Cathy Wu", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Vindula Jayawardana", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Zhongxia Yan", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 5, "n_ref_uni": 53, "n_ref": 117, "n_ref_all": 144, "n_fig": 12, "n_tab": 3, "L_tab": 568, "n_element_tab": 83, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 523, "n_element_tab_1": 77, "formula_len_all": 1836, "formula_len_all_1": 1289, "len_all": 194716, "len_all_1": 63258, "len_abs": 1803, "len_title": 140, "len_sents": 57666, "len_sents_1": 30722, "n_sents": 428, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1720, "title": "IntersectionZoo: Eco-driving for Benchmarking Multi-Agent Contextual Reinforcement Learning", "abs": "Despite the popularity of multi-agent reinforcement learning (RL) in simulated and two-player applications, its success in messy real-world applications has been limited. A key challenge lies in its generalizability across problem variations, a common necessity for many real-world problems. Contextual reinforcement learning (CRL) formalizes learning policies that generalize across problem variations. However, the lack of standardized benchmarks for multi-agent CRL has hindered progress in the field. Such benchmarks are desired to be based on real-world applications to naturally capture the many open challenges of real-world problems that affect generalization. To bridge this gap, we propose IntersectionZoo, a comprehensive benchmark suite for multi-agent CRL through the real-world application of cooperative eco-driving in urban road networks. The task of cooperative eco-driving is to control a fleet of vehicles to reduce fleet-level vehicular emissions. By grounding IntersectionZoo in a real-world application, we naturally capture real-world problem characteristics, such as partial observability and multiple competing objectives. IntersectionZoo is built on data-informed simulations of 16,334 signalized intersections derived from 10 major US cities, modeled in an open-source industry-grade microscopic traffic simulator. By modeling factors affecting vehicular exhaust emissions (e.g., temperature, road conditions, travel demand), IntersectionZoo provides one million data-driven traffic scenarios. Using these traffic scenarios, we benchmark popular multi-agent RL and human-like driving algorithms and demonstrate that the popular multi-agent RL algorithms struggle to generalize in CRL settings.", "keywords": [ "reinforcement learning", "generalization", "benchmarking", "eco-driving" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "XoZMP9GX9o", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Mengyi Wei", "gender": "Female", "institution": "Tianjin University", "country": "", "position": "MS student" }, { "name": "Zhangpengchen", "gender": "Male", "institution": "Tianjin University", "country": "", "position": "MS student" }, { "name": "Yufan Dai", "gender": "Female", "institution": "Tianjin University", "country": "", "position": "MS student" }, { "name": "Jiaxiang Guo", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Weiyali", "gender": "Female", "institution": "Tianjin University", "country": "", "position": "MS student" } ], "n_formula": 32, "n_formula_1": 27, "n_ref_uni": 42, "n_ref": 72, "n_ref_all": 103, "n_fig": 7, "n_tab": 5, "L_tab": 1451, "n_element_tab": 106, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 67, "n_element_tab_1": 20, "formula_len_all": 1318, "formula_len_all_1": 1191, "len_all": 136130, "len_all_1": 61316, "len_abs": 808, "len_title": 160, "len_sents": 40534, "len_sents_1": 28266, "n_sents": 318, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 114, "L_abs": 810, "title": "Gumbel-Softmax Discretization Constraint, Differentiable IDS Channel, and an IDS-Correcting Code for DNA Storage", "abs": "Insertion, deletion, and substitution (IDS) error-correcting codes have garnered increased attention with recent advancements in DNA storage technology. However, a universal method for designing IDS-correcting codes across varying channel settings remains underexplored. We present an autoencoder-based method, THEA-code, aimed at efficiently generating IDS-correcting codes for complex IDS channels. In the work, a Gumbel-Softmax discretization constraint is proposed to discretize the features of the autoencoder, and a simulated differentiable IDS channel is developed as a differentiable alternative for IDS operations. These innovations facilitate the successful convergence of the autoencoder, resulting in channel-customized IDS-correcting codes with commendable performance across complex IDS channels.", "keywords": [ "Gumbel-Softmax", "IDS Channel", "IDS-correcting Code", "DNA Storage" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "XoYdD3m0mv", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Eliahu Horwitz", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "PhD student" }, { "name": "Imri Shuval", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "Undergrad student" }, { "name": "Jonathan Kahana", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "PhD student" }, { "name": "Yedid Hoshen", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "Associate Professor" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 45, "n_ref": 99, "n_ref_all": 116, "n_fig": 11, "n_tab": 9, "L_tab": 1611, "n_element_tab": 116, "n_fig_1": 8, "n_tab_1": 8, "L_tab_1": 626, "n_element_tab_1": 47, "formula_len_all": 62, "formula_len_all_1": 206, "len_all": 141962, "len_all_1": 65367, "len_abs": 1156, "len_title": 102, "len_sents": 37203, "len_sents_1": 29775, "n_sents": 326, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1161, "title": "Deep Linear Probe Generators for Weight Space Learning", "abs": "Weight space learning aims to extract information about a neural network, such as its training dataset or generalization error. Recent approaches learn directly from model weights, but this presents many challenges as weights are high-dimensional and include permutation symmetries between neurons. An alternative approach, Probing, represents a model by passing a set of learned inputs (probes) through the model, and training a predictor on top of the corresponding outputs. Although probing is typically not used as a stand alone approach, our preliminary experiment found that a vanilla probing baseline worked surprisingly well. However, we discover that current probe learning strategies are ineffective. We therefore propose Deep Linear Probe Generators (ProbeGen), a simple and effective modification to probing approaches. ProbeGen adds a shared generator module with a deep linear architecture, providing an inductive bias towards structured probes thus reducing overfitting.\nWhile simple, ProbeGen performs significantly better than the state-of-the-art and is very efficient, requiring between 30 to 1000 times fewer FLOPs than other top approaches.", "keywords": [ "Deep Weight Space", "Weight Classification" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "Xo0Q1N7CGk", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Dehong Xu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Ruiqi Gao", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Wenhao Zhang", "gender": "Male", "institution": "University of Texas Southwestern Medical Center", "country": "US", "position": "Assistant Professor" }, { "name": "Xue-Xin Wei", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "Assistant Professor" }, { "name": "Yingnian Wu", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Full Professor" } ], "n_formula": 39, "n_formula_1": 12, "n_ref_uni": 54, "n_ref": 136, "n_ref_all": 172, "n_fig": 18, "n_tab": 2, "L_tab": 263, "n_element_tab": 27, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 263, "n_element_tab_1": 27, "formula_len_all": 2180, "formula_len_all_1": 639, "len_all": 184884, "len_all_1": 66935, "len_abs": 1178, "len_title": 113, "len_sents": 52456, "len_sents_1": 28985, "n_sents": 453, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 64, "L_abs": 1182, "title": "An Investigation of Conformal Isometry Hypothesis for Grid Cells", "abs": "This paper investigates the conformal isometry hypothesis as a potential explanation for hexagonal periodic patterns in grid cell response maps. The hypothesis posits that grid cell activity forms a high-dimensional vector in neural space, encoding the agent\u2019s position in 2D physical space. As the agent moves, this vector rotates within a 2D manifold in the neural space, driven by a recurrent neural network. The conformal hypothesis suggests that this neural manifold is a conformally isometric embedding of physical space, where local displacements in neural space are proportional to those in physical space. In this paper, we conduct numerical experiments to show that this hypothesis leads to the hexagon periodic patterns of grid cells, agnostic to the choice of transformation models. Furthermore, we present a theoretical understanding that hexagon patterns emerge by minimizing our loss function because hexagon flat torus exhibits minimal deviation from local conformal isometry. In addition, we propose a conformal modulation of the agent's input velocity, enabling the recurrent neural network of grid cells to satisfy the conformal isometry hypothesis automatically.", "keywords": [ "representation learning", "recurrent neural network", "position embedding", "grid cell" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "XnX7xRoroC", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Connor Wilhelm", "gender": "Male", "institution": "Brigham Young University", "country": "US", "position": "PhD student" }, { "name": "Dan Ventura", "gender": "Male", "institution": "Brigham Young University", "country": "", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 14, "n_ref": 33, "n_ref_all": 45, "n_fig": 3, "n_tab": 10, "L_tab": 1856, "n_element_tab": 155, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1013, "n_element_tab_1": 65, "formula_len_all": 977, "formula_len_all_1": 598, "len_all": 111320, "len_all_1": 66708, "len_abs": 1127, "len_title": 108, "len_sents": 44017, "len_sents_1": 32724, "n_sents": 317, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1129, "title": "Distilling Reinforcement Learning into Single-Batch Datasets", "abs": "Dataset distillation compresses a large dataset into a small synthetic dataset such that learning on the synthetic dataset approximates learning on the large dataset. Training on the distilled dataset can be performed in as little as one step of gradient descent. We demonstrate that distillation is generalizable to different tasks by distilling reinforcement learning environments into one-batch supervised learning datasets. This demonstrates not only distillation's ability to compress a reinforcement learning task but also its ability to transform one learning modality (reinforcement learning) into another (supervised learning). We present a novel extension of proximal policy optimization for meta-learning and use it in distillation of a multi-dimensional extension of the classic cart-pole problem, all MuJoCo environments, and several Atari games. We demonstrate distillation's ability to compress complex RL environments into one-step supervised learning, explore RL distillation's generalizability across learner architectures, and demonstrate distilling an environment into the smallest-possible synthetic dataset.", "keywords": [ "distillation", "reinforcement learning", "RL", "meta-learning", "compression" ], "rating_list": [ 8, 6, 8, 3 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 2, 3, 1 ], "contribution_list": [ 3, 3, 3, 1 ], "confidence_list": [ 5, 4, 4, 2 ] }, { "paper_id": "XnVttczoAV", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Chang Liu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "", "position": "PhD student" }, { "name": "Hang Ruan", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Tianyi Bao", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Wenjie Wu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xinyu Ye", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 44, "n_formula_1": 23, "n_ref_uni": 46, "n_ref": 91, "n_ref_all": 108, "n_fig": 7, "n_tab": 7, "L_tab": 3214, "n_element_tab": 268, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1819, "n_element_tab_1": 344, "formula_len_all": 1629, "formula_len_all_1": 1043, "len_all": 210126, "len_all_1": 71516, "len_abs": 1446, "len_title": 161, "len_sents": 65551, "len_sents_1": 30168, "n_sents": 472, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 1101, "title": "Beyond Circuit Connections: A Non-Message Passing Graph Transformer Approach for Quantum Error Mitigation", "abs": "Despite the progress in quantum computing, one major bottleneck against the practical utility is its susceptibility to noise, which frequently occurs in current quantum systems. Existing quantum error mitigation (QEM) methods either lack generality to noise and circuit types or fail to capture the global dependencies of entire systems in addition to circuit structure. In this work, we first propose a unique circuit-to-graph encoding scheme with qubit-wise noisy measurement aggregated. Then, we introduce GTraQEM, a non-message passing graph transformer designed to effectively mitigate errors in expected circuit measurement outcomes. GTraQEM are equipped with a quantum-specific positional encoding, a structure matrix as attention bias guiding nonlocal aggregation, and a virtual quantum-representative node to further grasp graph representations, which guarantees to model the long-range entanglement. Experimental evaluations demonstrate that GTraQEM outperforms state-of-the-art QEM methods on both random and structured quantum circuits across noise types and scales among diverse settings.", "keywords": [ "Quantum Error Mitigation; Graph Transformer" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "XnDyddPcBT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Cheongwoong Kang", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "David Leo Wright Hall", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Researcher" }, { "name": "Dongeun Lee", "gender": "Male", "institution": "Texas A&M University - Commerce", "country": "", "position": "Associate Professor" }, { "name": "Duc Nguyen", "gender": "Male", "institution": "VinAI Research", "country": "VN", "position": "Research Resident" }, { "name": "Jaesik Choi", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" }, { "name": "Toan Tran", "gender": "Male", "institution": "VinAI Research", "country": "VN", "position": "Researcher" }, { "name": "Anh Tong", "gender": "Male", "institution": "Korea University", "country": "", "position": "Assistant Professor" }, { "name": "Thanh Tang Nguyen", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Postdoc" } ], "n_formula": 29, "n_formula_1": 16, "n_ref_uni": 59, "n_ref": 117, "n_ref_all": 156, "n_fig": 18, "n_tab": 5, "L_tab": 2423, "n_element_tab": 125, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 542, "n_element_tab_1": 34, "formula_len_all": 2619, "formula_len_all_1": 1281, "len_all": 178413, "len_all_1": 68388, "len_abs": 1008, "len_title": 126, "len_sents": 48392, "len_sents_1": 30347, "n_sents": 396, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1011, "title": "Neural ODE Transformers: Analyzing Internal Dynamics and Adaptive Fine-tuning", "abs": "Recent advancements in large language models (LLMs) based on transformer architectures have sparked significant interest in understanding their inner workings. In this paper, we introduce a novel approach to modeling transformer architectures using highly flexible non-autonomous neural ordinary differential equations (ODEs). Our proposed model fully parameterizes all the weights of attention and feed-forward blocks through neural networks, with weights articulated as functions of a continuous layer index. We examine the model's dynamics through spectral analysis, uncovering an increase in eigenvalue magnitude offering a practical insights against weight-sharing assumption in existing theoretical studies. We also introduce the use of the Lyapunov exponent to examine token-level sensitivity, improving model interpretability. Our neural ODE transformer performs similarly to GPT across various configurations and datasets, while offering flexible fine-tuning capabilities under different architectures.", "keywords": [ "neural ODEs", "transformer", "adaptive finetune" ], "rating_list": [ 8, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "Xn4Je0CxC6", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Antoine Barrier", "gender": "Male", "institution": "Grenoble Institute of Neurosciences", "country": "FR", "position": "Postdoc" }, { "name": "Florence Forbes", "gender": "Female", "institution": "INRIA", "country": "", "position": "Researcher" }, { "name": "Henrique Donancio", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Postdoc" }, { "name": "Leah F South", "gender": "Female", "institution": "Queensland University of Technology", "country": "AU", "position": "Senior Lecturer" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 40, "n_ref": 56, "n_ref_all": 72, "n_fig": 7, "n_tab": 3, "L_tab": 742, "n_element_tab": 44, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 430, "n_element_tab_1": 45, "formula_len_all": 597, "formula_len_all_1": 502, "len_all": 116260, "len_all_1": 55183, "len_abs": 1192, "len_title": 121, "len_sents": 29625, "len_sents_1": 25143, "n_sents": 183, "n_sents_1": 147, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1197, "title": "Dynamic Learning Rate for Deep Reinforcement Learning: A Bandit Approach", "abs": "In Deep Reinforcement Learning models trained using gradient-based techniques, the choice of optimizer and its learning rate are crucial to achieving good performance: higher learning rates can prevent the model from learning effectively, while lower ones might slow convergence. Additionally, due to the non-stationarity of the objective function, the best-performing learning rate can change over the training steps. To adapt the learning rate, a standard technique consists of using decay schedulers. However, these schedulers assume that the model is progressively approaching convergence, which may not always be true, leading to delayed or premature adjustments. In this work, we propose dynamic Learning Rate for deep Reinforcement Learning (LRRL), a meta-learning approach that selects the learning rate based on the agent's performance during training. LRRL is based on a multi-armed bandit algorithm, where each arm represents a different learning rate, and the bandit feedback is provided by the cumulative returns of the RL policy to update the arms' probability distribution. Our empirical results demonstrate that LRRL can substantially improve the performance of deep RL algorithms.", "keywords": [ "Hyperparameter Optimization", "Meta Learning", "Deep Reinforcement Learning", "Adversarial Multi-Armed Bandits" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "XmnTfSX5Az", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Defu Cao", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Jia Guo", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xiao-Ping Zhang", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xinchun Yu", "gender": "Male", "institution": "Zhejiang Gongshang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yiyun Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" } ], "n_formula": 31, "n_formula_1": 14, "n_ref_uni": 43, "n_ref": 52, "n_ref_all": 71, "n_fig": 4, "n_tab": 12, "L_tab": 5356, "n_element_tab": 695, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 4437, "n_element_tab_1": 456, "formula_len_all": 1215, "formula_len_all_1": 506, "len_all": 172280, "len_all_1": 60852, "len_abs": 2026, "len_title": 94, "len_sents": 52532, "len_sents_1": 23384, "n_sents": 382, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 2044, "title": "KambaAD: Enhancing State Space Models with Kolmogorov\u2013Arnold for time series Anomaly Detection", "abs": "Time series anomaly detection is critical in numerous practical applications, yet existing deep learning methods often fall short of real-world demands. These models fail to swiftly filter out physically implausible anomalies, insufficiently address distributional shifts, and lack a comprehensive approach that integrates both global and local perspectives for anomaly detection. Moreover, most successful models rely on channel-dependent methods that tend to treat all features at the same timestamp as a single token and then focus on finding relationships between these tokens. This approach overlooks the unique periodicities, trends, and lagged relationships between different features, leading to suboptimal performance. To address these limitations, we propose KambaAD, a model comprised of an Encoder and Reconstructor. The Encoder integrates the strengths of the Kolmogorov-Arnold Network (KAN), the attention mechanism, and the Selective Structured State Space Model (MAMBA). Specifically, KAN is employed to swiftly enforce data consistency, enabling rapid detection of anomalies that violate physical laws. The attention mechanism ensures balanced processing of global information while enhancing the representation of key data characteristics. We leverage MAMBA's capability as a sequence model to capture anomalies caused by local variations. Additionally, its internal selection mechanism allows the model to effectively handle distribution shifts, ensuring robustness and adaptability in the presence of changing data distributions. Additionally, the framework incorporates a time-series-specific Reconstructor, which reduces computational complexity through patch-based operations that exploit local consistency in time series data. It also employs channel-independent linear reconstruction to prevent interference between different features. Through extensive experiments on multiple multivariate datasets, KambaAD consistently outperforms state-of-the-art models, demonstrating its superior performance in anomaly detection.", "keywords": [ "Anomaly Detection", "Mamba", "KAN", "Time series" ], "rating_list": [ 6, 6, 3, 6, 5 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 2 ], "contribution_list": [ 3, 3, 1, 3, 3 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "Xmh5gdMfRJ", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "BO HU", "gender": "unknown", "institution": "American Express", "country": "SG", "position": "Researcher" }, { "name": "Tat-Jen Cham", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 14, "n_ref_uni": 38, "n_ref": 74, "n_ref_all": 86, "n_fig": 3, "n_tab": 5, "L_tab": 2225, "n_element_tab": 203, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1677, "n_element_tab_1": 151, "formula_len_all": 847, "formula_len_all_1": 710, "len_all": 133758, "len_all_1": 73110, "len_abs": 1514, "len_title": 126, "len_sents": 34653, "len_sents_1": 32454, "n_sents": 266, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1519, "title": "TSC-Net: Prediction of Pedestrian Trajectories by Trajectory-Scene-Cell Classification", "abs": "To predict future trajectories of pedestrians, scene is as important as the history trajectory since i) scene reflects the position of possible goals of the pedestrian ii) trajectories are affected by the semantic information of the scene. It requires the model to capture scene information and learn the relation between scenes and trajectories. However, existing methods either apply Convolutional Neural Networks (CNNs) to summarize the scene to a feature vector, which raises the feature misalignment issue, or convert trajectory to heatmaps to align with the scene map, which ignores the interactions among different pedestrians. In this work, we introduce the trajectory-scene-cell feature to represent both trajectories and scenes in one feature space. By decoupling the trajectory in temporal domain and the scene in spatial domain, trajectory feature and scene feature are re-organized in different types of cell feature, which well aligns trajectory and scene, and allows the framework to model both human-human and human-scene interactions. Moreover, the Trajectory-Scene-Cell Network (TSC-Net) with new trajectory prediction manner is proposed, where both goal and intermediate positions of the trajectory are predict by cell classification and offset regression. Comparative experiments show that TSC-Net achieves the SOTA performance on several datasets with most of the metrics. Especially for the goal estimation, TSC-Net is demonstrated better on predicting goals for trajectories with irregular speed.", "keywords": [ "Trajectory-Scene-Cell", "Trajectory Prediction", "Attention" ], "rating_list": [ 5, 5, 6, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "XmProj9cPs", "primary_area": "datasets and benchmarks", "n_author": 16, "author_info_list": [ { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Dongchan Shin", "gender": "unknown", "institution": "University of Hong Kong", "country": "", "position": "Undergrad student" }, { "name": "Fangyu Lei", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Hongcheng Gao", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Hongjin SU", "gender": "unknown", "institution": "the University of Hong Kong, University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jixuan Chen", "gender": "Female", "institution": "Nanjing University", "country": "CN", "position": "Undergrad student" }, { "name": "Pengcheng Yin", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Qian Liu", "gender": "Male", "institution": "Tiktok", "country": "SG", "position": "Researcher" }, { "name": "Ruisheng Cao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Ruoxi Sun", "gender": "Female", "institution": "Google", "country": "", "position": "Google" }, { "name": "Sida Wang", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Tao Yu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Victor Zhong", "gender": "Male", "institution": "University of Waterloo", "country": "", "position": "Assistant Professor" }, { "name": "Wenjing Hu", "gender": "Female", "institution": "nanjing university", "country": "CN", "position": "Undergrad student" }, { "name": "Yuxiao Ye", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "ZHAOQING SUO", "gender": "unknown", "institution": "University of Hong Kong", "country": "HK", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 34, "n_ref": 87, "n_ref_all": 149, "n_fig": 26, "n_tab": 26, "L_tab": 17193, "n_element_tab": 2440, "n_fig_1": 15, "n_tab_1": 10, "L_tab_1": 3161, "n_element_tab_1": 387, "formula_len_all": 111, "formula_len_all_1": 0, "len_all": 294998, "len_all_1": 73515, "len_abs": 1746, "len_title": 133, "len_sents": 95004, "len_sents_1": 31423, "n_sents": 763, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 85, "L_abs": 1629, "title": "Spider 2.0: Evaluating Language Models on Real-World Enterprise Text-to-SQL Workflows", "abs": "Real-world enterprise text-to-SQL workflows often involve complex cloud or local data across various database systems, multiple SQL queries in various dialects, and diverse operations from data transformation to analytics.\nWe introduce Spider 2.0, an evaluation framework comprising $595$ real-world text-to-SQL workflow problems derived from enterprise-level database use cases. \nThe databases in Spider 2.0 are sourced from real data applications, often containing over 1,000 columns and stored in local or cloud database systems such as BigQuery and Snowflake.\nWe show that solving problems in Spider 2.0 frequently requires understanding and searching through database metadata, dialect documentation, and even project-level codebases. \nThis challenge calls for models to interact with complex SQL workflow environments, process extremely long contexts, perform intricate reasoning, and generate multiple SQL queries with diverse operations, often exceeding $100$ lines, which goes far beyond traditional text-to-SQL challenges.\nOur evaluations indicate that based on o1-preview, our code agent framework successfully solves only 15.1\\% of the tasks, compared with 91.2\\% on Spider 1.0 and 73.0\\% on BIRD.\nOur results on Spider 2.0 show that while language models have demonstrated remarkable performance in code generation --- especially in prior text-to-SQL benchmarks --- they require significant improvement in order to achieve adequate performance for real-world enterprise usage.\nProgress on Spider 2.0 represents crucial steps towards developing intelligent, autonomous, code agents for real-world enterprise settings.", "keywords": [ "LLM Benchmark", "Data Science and Engineering", "Code Generation", "Text-to-SQL", "LLM Agent" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 4, 3, 4 ], "presentation_list": [ 4, 4, 3, 4 ], "contribution_list": [ 4, 4, 4, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "Xk9Q0CrJQc", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Tobias Kreiman", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Aditi Krishnapriyan", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 45, "n_ref": 125, "n_ref_all": 170, "n_fig": 24, "n_tab": 7, "L_tab": 1775, "n_element_tab": 105, "n_fig_1": 15, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 557, "formula_len_all_1": 472, "len_all": 211698, "len_all_1": 65529, "len_abs": 1941, "len_title": 130, "len_sents": 68422, "len_sents_1": 33028, "n_sents": 531, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1951, "title": "Understanding and Mitigating Distribution Shifts for Machine Learning Force Fields", "abs": "Machine Learning Force Fields (MLFFs) are a promising alternative to expensive ab initio quantum mechanical molecular simulations. Given the diversity of chemical spaces that are of interest and the cost of generating new data, it is important to understand how MLFFs generalize beyond their training distributions. In order to characterize and better understand distribution shifts in MLFFs, we conduct diagnostic experiments on chemical datasets, revealing common shifts that pose significant challenges, even for large foundation models trained on extensive data. Based on these observations, we hypothesize that current supervised training methods inadequately regularize MLFFs, resulting in overfitting and learning poor representations of out-of-distribution systems. We then propose two new methods as initial steps for mitigating distribution shifts for MLFFs. Our methods focus on test-time refinement strategies that incur minimal computational cost and do not use ab initio labels. The first strategy, based on spectral graph theory, modifies the edges of test graphs to align with graph structures seen during training. It can be applied to any existing pre-trained model to mitigate connectivity distribution shifts. Our second strategy improves representations for out-of-distribution systems at test-time by taking gradient steps using an auxiliary objective. Inspired by previous test-time training works in computer vision, we replace self-supervised objectives at test time with an objective that uses an efficient prior to address distribution shifts. Our test-time refinement strategies can reduce force errors by an order of magnitude on out-of-distribution systems, suggesting that MLFFs are capable of and can move towards modeling diverse chemical spaces, but are not being effectively trained to do so. Our experiments establish clear benchmarks for evaluating the generalization capabilities of the next generation of MLFFs.", "keywords": [ "machine learning force fields", "test-time training", "distribution shifts" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "XjSfcJUcaA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Isao Echizen", "gender": "Male", "institution": "National Institute of Informatics", "country": "JP", "position": "Full Professor" }, { "name": "Mingdeng Cao", "gender": "Male", "institution": "The University of Tokyo ", "country": "JP", "position": "PhD student" }, { "name": "Weihang Ran", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Yinqiang Zheng", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "WEI YUAN", "gender": "Male", "institution": "Tohoku University", "country": "JP", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 32, "n_ref": 50, "n_ref_all": 59, "n_fig": 3, "n_tab": 6, "L_tab": 3561, "n_element_tab": 551, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 3552, "n_element_tab_1": 537, "formula_len_all": 921, "formula_len_all_1": 921, "len_all": 123828, "len_all_1": 62676, "len_abs": 1269, "len_title": 147, "len_sents": 27836, "len_sents_1": 25562, "n_sents": 205, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1276, "title": "More Harmful, Less noticeable: Learning Adversarial Null-Text Embeddings for Inconspicuous Attack", "abs": "Adversarial examples, which are artificially crafted data intended to disrupt the output of deep learning models, present a new round of challenges to the stability and security of artificial intelligence technology. Unrestricted adversarial examples, obtained by modifying the semantic elements of images, have the characteristics of being natural and semantically meaningful. However, previous methods either significantly altered the image's color or content, or blurred visual details (such as text or geometric designs), making the generated adversarial examples easily detectable by the human eye. In this paper, we propose a method to generate highly natural adversarial examples based on stable diffusion. This is achieved by introducing adversarial loss during the image reconstruction process to perturb cross-attention mechanism. To further enhance image quality, we introduce perceptual loss into the adversarial attack process for the first time. Extensive experiments and visualizations demonstrate the effectiveness of our proposed method. Compared to the current state-of-the-art methods, our approach not only improves the adversarial transferability by an average of 12.59-50.3% but also significantly enhances image quality. Code will be publicly available.", "keywords": [ "adversarial attack", "diffusion model", "image editing" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "XjSdhzBTMq", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Hernan Ceferino Vazquez", "gender": "Male", "institution": "MercadoLibre Inc.", "country": "", "position": "Researcher" }, { "name": "Jorge S\u00e1nchez", "gender": "Male", "institution": "MercadoLibre Inc.", "country": "", "position": "Researcher" }, { "name": "Pucci Romero Tobias", "gender": "Male", "institution": "University of Buenos Aires", "country": "AR", "position": "MS student" }, { "name": "Ver\u00f3nica Bogado", "gender": "Female", "institution": "Mercado Libre", "country": "AR", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 0, "n_ref": 7, "n_ref_all": 19, "n_fig": 2, "n_tab": 15, "L_tab": 2735, "n_element_tab": 374, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 30, "n_element_tab_1": 2, "formula_len_all": 1304, "formula_len_all_1": 862, "len_all": 153405, "len_all_1": 68928, "len_abs": 1323, "len_title": 130, "len_sents": 53959, "len_sents_1": 35701, "n_sents": 452, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1328, "title": "Learning to Rank for AutoML: enhancing pipeline selection with ranking information", "abs": "This paper introduces a learning-to-rank (LTR) framework to address the problem of pipeline selection in automated machine learning systems. The traditional approach to AutoML involves learning to predict the performance of various pipelines on a given task based on data acquired from previous tasks (i.e., meta-learning), which can be complex due to the need for different models for each task-specific metric. The proposed framework aims to select the best pipeline based on ranking rather than estimating a target metric, aligning more closely with the ultimate goal of the task (i.e., selecting pipeline candidates in order, from more to least promising). This approach enables more robust, metric-agnostic solutions that are easier to compare using ranking metrics like NDCG and MRR. The paper evaluates LTR strategies on public OpenML datasets, demonstrating a clear advantage for ranking-based methods. Additionally, the integration of LTR with Bayesian optimization and Monte Carlo tree search is explored, leading to improvements in the ranking metrics. Finally, the study found a strong correlation between ranking metrics (e.g., NDCG and MRR) and AutoML metrics, such as the task objective metric and the time to find the best solution, providing insights into how ranking-based methods could enhance AutoML systems.", "keywords": [ "AutoML", "Learning-to-Rank", "Pipeline selection", "Meta-learning", "Ranking", "Bayesian optimization", "Monte Carlo tree search", "OpenML datasets", "Metric-agnostic", "Machine Learning" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "Xj6j48QIB3", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Bo Chen", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Hongsong Tang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Jianchun Xu", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Associate Professor" }, { "name": "Junge Zhang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "Full Professor" }, { "name": "Liuyu Xiang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Assistant Professor" }, { "name": "Yaodong Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yingzhuo Liu", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Zhaofeng He", "gender": "Male", "institution": "Beijing University of Post and Telecommunication", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 15, "n_ref": 33, "n_ref_all": 39, "n_fig": 4, "n_tab": 2, "L_tab": 1773, "n_element_tab": 103, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1773, "n_element_tab_1": 103, "formula_len_all": 515, "formula_len_all_1": 515, "len_all": 77865, "len_all_1": 77865, "len_abs": 1098, "len_title": 128, "len_sents": 22643, "len_sents_1": 22643, "n_sents": 151, "n_sents_1": 151, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1103, "title": "Mixed Hierarchical Oracle and Multi-Agent Benchmark in Two-player Zero-sum Games", "abs": "Self-play methods have achieved remarkable success in two-player zero-sum games, attaining superhuman performance in many complex game domains. Parallelizing learners is a feasible approach to handling large-scale games. However, parallelizing learners often leads to suboptimal exploitation of computational resources, resulting in inefficiencies. In this study, we introduce the Mixed Hierarchical Oracle (MHO), designed to enhance computational efficiency and performance in large-scale two-player zero-sum games. MHO enables the parallelization of reinforcement learning tasks through a hierarchical pipeline that balances exploration and exploitation across oracle levels. It also avoids cold-start issues by using a \"model soup\" initialization strategy. Additionally, we present MiniStar, an open-source environment focused on small-scale combat scenarios, developed to facilitate research in self-play algorithms. Through extensive experiments on matrix games and the MiniStar environment, we demonstrate that MHO outperforms existing methods in terms of computational efficiency and performance.", "keywords": [ "reinforcement learning", "game theory", "deep learning" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 1, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 3, 1, 2 ], "confidence_list": [ 2, 4, 4 ] }, { "paper_id": "Xj66fkrlTk", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Daniil Tiapkin", "gender": "Male", "institution": "Ecole Polytechnique", "country": "FR", "position": "PhD student" }, { "name": "Nikita Morozov", "gender": "Male", "institution": "Higher School of Economics", "country": "RU", "position": "PhD student" }, { "name": "Sergey Samsonov", "gender": "Male", "institution": "Higher School of Economics", "country": "", "position": "PhD student" }, { "name": "Timofei Gritsaev", "gender": "Male", "institution": "Constructor University", "country": "DE", "position": "MS student" } ], "n_formula": 20, "n_formula_1": 15, "n_ref_uni": 46, "n_ref": 148, "n_ref_all": 169, "n_fig": 7, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1573, "formula_len_all_1": 1119, "len_all": 148846, "len_all_1": 66005, "len_abs": 1107, "len_title": 128, "len_sents": 44817, "len_sents_1": 30684, "n_sents": 332, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1110, "title": "Optimizing Backward Policies in GFlowNets via Trajectory Likelihood Maximization", "abs": "Generative Flow Networks (GFlowNets) are a family of generative models that learn to sample objects with probabilities proportional to a given reward function. The key concept behind GFlowNets is the use of two stochastic policies: a forward policy, which incrementally constructs compositional objects, and a backward policy, which sequentially deconstructs them. Recent results show a close relationship between GFlowNet training and entropy-regularized reinforcement learning (RL) problems with a particular reward design. However, this connection applies only in the setting of a fixed backward policy, which might be a significant limitation. As a remedy to this problem, we introduce a simple backward policy optimization algorithm that involves direct maximization of the value function in an entropy-regularized Markov Decision Process (MDP) over intermediate rewards. We provide an extensive experimental evaluation of the proposed approach across various benchmarks in combination with both RL and GFlowNet algorithms and demonstrate its faster convergence and mode discovery in complex environments.", "keywords": [ "generative flow networks", "gflownets", "reinforcement learning", "sampling", "generative models" ], "rating_list": [ 5, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "XigBo6nWzL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Gyeong-Moon Park", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Assistant Professor" }, { "name": "Gyeongdeok Seo", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "MS student" }, { "name": "Heejun Ahn", "gender": "unknown", "institution": "LG CNS", "country": "KR", "position": "Principal Researcher" }, { "name": "JOOWANG KIM", "gender": "Male", "institution": "LGCNS", "country": "KR", "position": "Researcher" }, { "name": "Jeyoon Yeom", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "MS student" }, { "name": "Kyungwoo Song", "gender": "unknown", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Sungjun Lim", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 12, "n_ref_uni": 17, "n_ref": 46, "n_ref_all": 62, "n_fig": 6, "n_tab": 5, "L_tab": 1937, "n_element_tab": 248, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1218, "n_element_tab_1": 174, "formula_len_all": 673, "formula_len_all_1": 622, "len_all": 119297, "len_all_1": 59712, "len_abs": 1498, "len_title": 132, "len_sents": 47469, "len_sents_1": 26367, "n_sents": 346, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1431, "title": "DDRL: A DIFFUSION-DRIVEN REINFORCEMENT LEARNING APPROACH FOR ENHANCED TSP SOLUTIONS", "abs": "The Traveling Salesman Problem (TSP) is a fundamental challenge in combinatorial optimization, known for its NP-hard complexity. Reinforcement Learning (RL) has proven to be effective in managing larger and more complex TSP instances, yet it encounters challenges such as training instability and necessity for a substantial amount of training resources. Diffusion models, known for iteratively refining noisy inputs to generate high-quality solutions, offer scalability and exploration capabilities for TSP but may struggle with optimality in complex cases and require large, resource-intensive training datasets. To address these limitations, we propose DDRL (Diffusion-Driven Reinforcement Learning), which integrates diffusion models with RL. DDRL employs a latent vector to generate an adjacency matrix, merging image and graph learning within a unified RL framework. By utilizing a pre-trained diffusion model as a prior, DDRL exhibits strong scalability and enhanced convergence stability. We also provide theoretical analysis that training DDRL aligns with the diffusion policy gradient in the process of solving the TSP, demonstrating its effectiveness. Additionally, we introduce novel constraint datasets\u2014obstacle, path, and cluster constraints\u2014to evaluate DDRL's generalization capabilities. We demonstrate that DDRL offers a robust solution that outperforms existing methods in both basic and constrained TSP problems.", "keywords": [ "TSP", "Diffusion model", "Reinforcement Learning" ], "rating_list": [ 3, 5, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4, 3 ] }, { "paper_id": "XhyCPEnlCa", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Chaewon Kang", "gender": "unknown", "institution": "Sung Kyun Kwan University", "country": "", "position": "PhD student" }, { "name": "Daejin Choi", "gender": "unknown", "institution": "Incheon National University", "country": "", "position": "Assistant Professor" }, { "name": "Jinyoung Han", "gender": "Male", "institution": "Sungkyunkwan University", "country": "", "position": "Associate Professor" }, { "name": "Seoyoon Jeong", "gender": "unknown", "institution": "Sung Kyun Kwan University", "country": "", "position": "MS student" }, { "name": "Simon S. Woo", "gender": "Male", "institution": "Sungkyunkwan University", "country": "KR", "position": "Associate Professor" }, { "name": "Lee jong hyun", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KP", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 68, "n_ref": 91, "n_ref_all": 138, "n_fig": 24, "n_tab": 11, "L_tab": 2914, "n_element_tab": 473, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2507, "n_element_tab_1": 357, "formula_len_all": 0, "formula_len_all_1": 73, "len_all": 198868, "len_all_1": 70859, "len_abs": 1177, "len_title": 96, "len_sents": 57523, "len_sents_1": 32041, "n_sents": 419, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 48, "L_abs": 1182, "title": "HiDF: A Human-Indistinguishable Deepfake Dataset", "abs": "The rapid development and prevalence of generative AI has made it easy for people to create high-quality deepfake images and videos, but their abuses also have been exponentially increased. To mitigate potential social disruption, it is crucial to quickly detect authenticity of each deepfake content hidden in a sea of information. While researchers have worked on developing deep learning-based methods, the deepfake datasets utilized in these studies are far from the real world in terms of their qualities; most of the popular deepfake datasets are human distinguishable. To address this problem, we present a novel deepfake dataset, HiDF, a high-quality and human-indistinguishable deepfake dataset consisting of 30 K images and 4 K videos. HiDF is a meticulously curated dataset that includes diverse subjects, which has been undergone rigorous quality checks. Comparison on the quality between HiDF and existing deepfake datasets demonstrates that HiDF is human-indistinguishable, hence it can be used as a valuable benchmark dataset for deepfake detection tasks. Data and code (https://github.\bwill.be.provided) are publicly available for future deepfake detection research.", "keywords": [ "Deepfake", "Human-Indistinguishable", "AI", "Multimodal" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 5, 5, 2 ] }, { "paper_id": "XhdckVyXKg", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Asif Salekin", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" }, { "name": "Tauhidur Rahman", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Yuliang Chen", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Yunfei Luo", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 2, "n_ref_uni": 49, "n_ref": 89, "n_ref_all": 111, "n_fig": 7, "n_tab": 13, "L_tab": 9458, "n_element_tab": 1058, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 5455, "n_element_tab_1": 652, "formula_len_all": 448, "formula_len_all_1": 135, "len_all": 213905, "len_all_1": 67232, "len_abs": 1877, "len_title": 130, "len_sents": 50589, "len_sents_1": 26907, "n_sents": 339, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1887, "title": "Toward Foundation Model for Multivariate Wearable Sensing of Physiological Signals", "abs": "Time-series foundation models have the ability to run inference, mainly forecasting, on any type of time series data, thanks to the informative representations comprising waveform features. \nWearable sensing data, on the other hand, contain more variability in both patterns and frequency bands of interest and generally emphasize more on the ability to infer healthcare-related outcomes. The main challenge of crafting a foundation model for wearable sensing physiological signals is to learn generalizable representations that support efficient adaptation across heterogeneous sensing configurations and applications. In this work, we propose NormWear, a step toward such a foundation model, aiming to extract generalized and informative wearable sensing representations. NormWear has been pretrained on a large set of physiological signals, including PPG, ECG, EEG, GSR, and IMU, from various public resources. For a holistic assessment, we perform downstream evaluation on 11 public wearable sensing datasets, spanning 18 applications in the areas of mental health, body state inference, biomarker estimations, and disease risk evaluations. We demonstrate that NormWear achieves a better performance improvement over competitive baselines in general time series foundation modeling. In addition, leveraging a novel representation-alignment-match-based method, we align physiological signals embeddings with text embeddings. This alignment enables our proposed foundation model to perform zero-shot inference, allowing it to generalize to previously unseen wearable signal-based health applications. Finally, we perform nonlinear dynamic analysis on the waveform features extracted by the model at each intermediate layer. This analysis quantifies the model's internal processes, offering clear insights into its behavior and fostering greater trust in its inferences among end users.", "keywords": [ "Foundation Model", "Signal Processing", "Representation Learning", "Wearable Sensing", "Digital Healthcare" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 3, 1, 1, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "XgYZT35N76", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Bowen Zhang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Research Scientist" }, { "name": "Haotian Zhang", "gender": "Male", "institution": "Apple AI/ML", "country": "US", "position": "Researcher" }, { "name": "Ruohong Zhang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Ruoming Pang", "gender": "unknown", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Yanghao Li", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Yiming Yang", "gender": "Female", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Yinfei Yang", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Zhe Gan", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Zhiqing Sun", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 24, "n_ref": 49, "n_ref_all": 70, "n_fig": 9, "n_tab": 17, "L_tab": 10893, "n_element_tab": 879, "n_fig_1": 4, "n_tab_1": 13, "L_tab_1": 6059, "n_element_tab_1": 674, "formula_len_all": 272, "formula_len_all_1": 135, "len_all": 169565, "len_all_1": 59636, "len_abs": 1696, "len_title": 104, "len_sents": 42879, "len_sents_1": 19868, "n_sents": 332, "n_sents_1": 141, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1323, "title": "Improve Vision Language Model Chain-of-thought Reasoning", "abs": "Chain-of-thought (CoT) reasoning in vision language models (VLMs) is crucial for improving interpretability and trustworthiness. \nHowever, current training recipes lack robust CoT reasoning data, relying on datasets dominated by short annotations with minimal rationales. In this work, we first evaluate the CoT abilities of existing VLMs and show that training on short answers does not generalize well to reasoning tasks that require more detailed responses. To address this, we propose a two-fold approach. First, we distill rationales from GPT-4o model to enrich the training data and fine-tune VLMs, boosting their CoT performance. Second, we apply reinforcement learning to further calibrate reasoning quality by constructing positive (correct) and negative (incorrect) pairs of model-generated reasoning chains, based on the comparisons with annotated short answers. We then use the Direct Preference Optimization algorithm on this pairwise data to refine the model\u2019s reasoning abilities. Our experiments demonstrate significant improvements in CoT reasoning on benchmark datasets and better generalization to direct answer prediction as well. This work emphasizes the importance of incorporating detailed rationales in training and leveraging reinforcement learning to strengthen the reasoning capabilities of VLMs.", "keywords": [ "Vision Language Model", "Chain-of-thought Reasoning" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "XgYPzNtz0s", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Hiroki Yanagisawa", "gender": "Male", "institution": "CyberAgent Inc.", "country": "JP", "position": "Researcher" }, { "name": "Shunta Akiyama", "gender": "Male", "institution": "CyberAgent, Inc.", "country": "JP", "position": "Researcher" } ], "n_formula": 132, "n_formula_1": 42, "n_ref_uni": 55, "n_ref": 94, "n_ref_all": 150, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 16879, "formula_len_all_1": 3809, "len_all": 193926, "len_all_1": 63127, "len_abs": 986, "len_title": 88, "len_sents": 54410, "len_sents_1": 26183, "n_sents": 543, "n_sents_1": 240, "venue": "ICLR2025", "decision": "withdraw", "L_title": 40, "L_abs": 988, "title": "Survival Analysis via Density Estimation", "abs": "This paper introduces an algorithm that reinterprets survival analysis through the lens of density estimation, addressing the challenge of censored inputs inherent to survival data. Recognizing that many survival analysis methodologies are extensions of foundational density estimation models, our approach leverages this intrinsic relationship. By conceptualizing survival analysis as a form of density estimation, our algorithm postprocesses the density estimation outputs to derive survival functions. This framework allows for the application of any density estimation model to effectively estimate survival functions, thereby broadening the toolkit available for survival analysis and enhancing the flexibility and applicability of existing density estimation techniques in this domain. The proposed algorithm not only bridges the methodological gap between density estimation and survival analysis but also offers a versatile and robust approach for handling censored survival data.", "keywords": [ "survival analysis", "censored regression", "competing risks", "dependent censoring" ], "rating_list": [ 6, 5, 6, 3, 3 ], "soundness_list": [ 3, 4, 3, 2, 3 ], "presentation_list": [ 2, 3, 4, 2, 2 ], "contribution_list": [ 3, 3, 3, 2, 1 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "XgH1wfHSX8", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Core Francisco Park", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Ekdeep Singh Lubana", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Hidenori Tanaka", "gender": "unknown", "institution": "Harvard University, Harvard University", "country": "US", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 8, "n_ref_uni": 41, "n_ref": 143, "n_ref_all": 257, "n_fig": 58, "n_tab": 4, "L_tab": 1544, "n_element_tab": 37, "n_fig_1": 14, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 904, "formula_len_all_1": 436, "len_all": 290685, "len_all_1": 71767, "len_abs": 1645, "len_title": 89, "len_sents": 110983, "len_sents_1": 34761, "n_sents": 958, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 41, "L_abs": 1651, "title": "Algorithmic Phases of In-Context Learning", "abs": "In-Context Learning (ICL) has significantly expanded the general-purpose nature of large language models, allowing them to adapt to novel tasks using merely the inputted context. This has motivated a series of papers that analyze tractable synthetic domains and postulate precise mechanisms that may underlie ICL. However, the use of relatively distinct setups that often lack a sequence modeling nature to them makes it unclear how general the reported insights from such studies are. Motivated by this, we propose a synthetic sequence modeling task that involves learning to simulate a finite mixture of Markov chains. As we show, models trained on this task reproduce most well-known results on ICL, hence offering a unified setting for studying the concept. Building on this setup, we demonstrate we can explain a model\u2019s behavior by decomposing it into four broad algorithms that combine a fuzzy retrieval vs. inference approach with either unigram or bigram statistics of the context. These algorithms engage in a competitive dynamics to dominate model behavior, with the precise experimental conditions dictating which algorithm ends up superseding others: e.g., we find merely varying context size or amount of training yields (at times sharp) transitions between which algorithm dictates the model behavior, revealing a mechanism that explains the transient nature of ICL. In this sense, we argue ICL is best thought of as a mixture of different algorithms, each with its own peculiarities, instead of a monolithic capability. This also implies that making general claims about ICL that hold universally across all settings may be infeasible.", "keywords": [ "In-Context Learning", "Circuit Competition", "Markov Chains", "Training Dynamics" ], "rating_list": [ 10, 8, 6, 6 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "XgCejjNNYX", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Guangyong Chen", "gender": "Male", "institution": "Zhejiang Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Lei Chen", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Pheng-Ann Heng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Xing Li", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Xinyi Zhou", "gender": "unknown", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Yingzhao Lian", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Yiwen Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" } ], "n_formula": 46, "n_formula_1": 25, "n_ref_uni": 37, "n_ref": 67, "n_ref_all": 89, "n_fig": 9, "n_tab": 11, "L_tab": 2481, "n_element_tab": 214, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 812, "n_element_tab_1": 67, "formula_len_all": 4480, "formula_len_all_1": 1845, "len_all": 178036, "len_all_1": 59349, "len_abs": 1228, "len_title": 136, "len_sents": 50353, "len_sents_1": 24992, "n_sents": 412, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1233, "title": "SeaDAG: Semi-autoregressive Diffusion for Conditional Directed Acyclic Graph Generation", "abs": "We introduce SeaDAG, a semi-autoregressive diffusion model for conditional generation of Directed Acyclic Graphs~(DAGs). Considering their inherent layer-wise structure, we simulate layer-wise autoregressive generation by designing different denoising speed for different layers. Unlike conventional autoregressive generation that lacks a global graph structure view, our method maintains a complete graph structure at each diffusion step, enabling operations such as property control that require the full graph structure.\nLeveraging this capability, we evaluate the DAG properties during training by employing a graph property decoder. We explicitly train the model to learn graph conditioning with a condition loss, which enhances the diffusion model's capacity to generate graphs that are both realistic and aligned with specified properties. \nWe evaluate our method on two representative conditional DAG generation tasks: (1) circuit generation from truth tables, where precise DAG structures are crucial for realizing circuit functionality, and (2) molecule generation based on quantum properties.\nOur approach demonstrates promising results, generating high-quality and realistic DAGs that closely align with given conditions.", "keywords": [ "Graph generation", "Diffusion model", "Conditional Graph Generation" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "XgAKt7rbXk", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Ronen Talmon", "gender": "unknown", "institution": "Technion - Israel Institute of Technology", "country": "IL", "position": "Full Professor" }, { "name": "Shahar Linial", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" }, { "name": "Shay Franchi", "gender": "Female", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" }, { "name": "Uri Shaham", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 35, "n_ref": 45, "n_ref_all": 53, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 552, "formula_len_all_1": 529, "len_all": 107071, "len_all_1": 65690, "len_abs": 1258, "len_title": 96, "len_sents": 37127, "len_sents_1": 32978, "n_sents": 249, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1263, "title": "Leveraging Low Rank Structure in The Lazy Regime", "abs": "Understanding the training dynamics of neural networks has gained much interest in the scientific community. The dynamics of training over-parameterized models is characterized by the lazy regime in which networks exhibit near-linear behavior and minimal parameter changes. In addition, it has been argued that the Jacobian of large neural models has a low-rank structure. In this paper, we focus on the opportunities laid out by the combination of low-rankness and laziness of large neural models. Specifically, we provide a scalable way to measure the extent of laziness, evaluated via the rate of change of the model Jacobian, as well as a scalable method to verify low-rankness of the model Jacobian without storing the entire Jacobian. Taking advantages of both laziness and low-rankness, we design a scalable training algorithm for over-parameterized models that performs backpropagation-free gradient descend training. In particular, this algorithm is of lower computation and storage requirements in cases of massive parameter sharing, as is the case of many state-of-the-art neural architectures. Empirical results confirm the scalability and effectiveness of our approach, opening new pathways for exploring novel learning strategies in neural networks.", "keywords": [ "Lazy Regime", "Linear Dynamics", "Neural Tangent Kernel", "Low Rank", "Wide Neural Networks", "Overparametrized Networks", "Backpropagation" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 1, 3, 1, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "Xg6JWb1Oxt", "primary_area": "reinforcement learning", "n_author": 11, "author_info_list": [ { "name": "Abbas Abdolmaleki", "gender": "unknown", "institution": "Google", "country": "", "position": "research scientist" }, { "name": "Jost Tobias Springenberg", "gender": "Male", "institution": "DeepMind", "country": "GB", "position": "Postdoc" }, { "name": "Markus Wulfmeier", "gender": "Male", "institution": "DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Martin Riedmiller", "gender": "Male", "institution": "Universit\u00e4t Freiburg", "country": "", "position": "Full Professor" }, { "name": "Martina Zambelli", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Michael Bloesch", "gender": "unknown", "institution": "DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Nicolas Heess", "gender": "unknown", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Philemon Brakel", "gender": "Male", "institution": "", "country": "", "position": "Undergrad student" }, { "name": "Roland Hafner", "gender": "Not Specified", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Todor Davchev", "gender": "unknown", "institution": "DeepMind", "country": "GB", "position": "Researcher" }, { "name": "William F Whitney", "gender": "unknown", "institution": "DeepMind", "country": "", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 68, "n_ref": 123, "n_ref_all": 144, "n_fig": 11, "n_tab": 3, "L_tab": 1132, "n_element_tab": 142, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1388, "n_element_tab_1": 160, "formula_len_all": 825, "formula_len_all_1": 799, "len_all": 203219, "len_all_1": 65150, "len_abs": 1106, "len_title": 132, "len_sents": 43858, "len_sents_1": 31585, "n_sents": 299, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1093, "title": "Value from Observations: Towards Large-Scale Imitation Learning via Self-Improvement", "abs": "Imitation Learning from Observation (IfO) offers a powerful way to learn behaviors from large-scale, mixed-quality data. Unlike behavior cloning or offline reinforcement learning, IfO leverages action-free demonstrations and circumvents the need for costly action-labeled demonstrations or carefully crafted reward functions. However, current research focuses on idealized scenarios with tailored data distributions. This paper introduces a novel algorithm to learn from datasets with varying quality, moving closer to a paradigm in which imitation learning can be performed iteratively via self-improvement. Our method extends RL-based imitation learning to action-free demonstrations, using a value function to transfer information between expert and non-expert data. Through comprehensive evaluation, we delineate the relation between different data distributions and the applicability of algorithms and highlight the limitations of established methods. Our findings provide valuable insights for developing more robust and practical IfO techniques on a path to scalable behaviour learning.", "keywords": [ "Imitation learning from observation", "self-improvement" ], "rating_list": [ 8, 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 2, 4, 4 ] }, { "paper_id": "XfWJT3BUmX", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Junfeng Ding", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Xinmin Zhang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Xuzhong Hu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "ZaiPeng Duan", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Ma Jie", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "anpei", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Postdoc" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 38, "n_ref": 84, "n_ref_all": 98, "n_fig": 10, "n_tab": 13, "L_tab": 3043, "n_element_tab": 368, "n_fig_1": 8, "n_tab_1": 8, "L_tab_1": 2122, "n_element_tab_1": 293, "formula_len_all": 459, "formula_len_all_1": 459, "len_all": 148247, "len_all_1": 67190, "len_abs": 4120, "len_title": 101, "len_sents": 36036, "len_sents_1": 29286, "n_sents": 276, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1489, "title": "DSConv: Dynamic Convolution On Serialized Point Cloud", "abs": "In recent years, research on point-based architectures has advanced rapidly, showcasing their competitive performance. However, the unstructured nature of point clouds limits the application of effective operators such as convolutions in feature extraction. Although many works have attempted to address the issues of unstructured data and introduce convolutions or transformers, the complex spatial mappings of point clouds and cumbersome convolution implementations in these methods limit real-time performance of the model. Furthermore, excessive structural mapping ignores the independence of point cloud position representation and fails to capture finer-grained features. To tackle these challenges, we serialize point clouds to provide them with structure and introduce AdaConv to directly utilize 2D convolutions, which simplifies the process and better preserves the relative positional relationship. Additionally, we propose a novel dynamic refinement approach for point cloud positions, continuously modifying the coordinates of points within the convolutional neighborhood to enhance the flexibility and adaptability. We also integrate local and global features to compensate for the loss of point cloud features during downsampling. Finally, we propose DSConv based on PointNeXt, maintaining scalability and inference speed. By combining DSConv with new architectural designs, we outperform the current state-of-the-art methods on ScanObjectNN, Scannet V2, and S3DIS datasets.", "keywords": [ "Point cloud serialization; Point cloud analysis; 3D object classification; 3D semantic segmentation; Deep learning architectures" ], "rating_list": [ 5, 5, 6, 3, 3 ], "soundness_list": [ 2, 3, 3, 3, 1 ], "presentation_list": [ 3, 3, 3, 1, 1 ], "contribution_list": [ 2, 2, 2, 3, 1 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "XfKSDgqIRj", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Han Cai", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Haocheng Xi", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Jianfei Chen", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Kurt Keutzer", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Ligeng Zhu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Song Han", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yao Lu", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" } ], "n_formula": 14, "n_formula_1": 6, "n_ref_uni": 52, "n_ref": 80, "n_ref_all": 101, "n_fig": 9, "n_tab": 9, "L_tab": 3595, "n_element_tab": 532, "n_fig_1": 8, "n_tab_1": 7, "L_tab_1": 2124, "n_element_tab_1": 303, "formula_len_all": 1266, "formula_len_all_1": 471, "len_all": 181852, "len_all_1": 64472, "len_abs": 1532, "len_title": 131, "len_sents": 43584, "len_sents_1": 26410, "n_sents": 359, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1568, "title": "COAT: Compressing Optimizer states and Activations for Memory-Efficient FP8 Training", "abs": "FP8 training has emerged as a promising method for improving training efficiency. Existing frameworks accelerate training by applying FP8 computation to linear layers while leaving optimizer states and activations in higher precision, which fails to fully optimize memory usage. This paper introduces COAT (**C**ompressing **O**ptimizer States and **A**ctivations for FP8 **T**raining), a novel FP8 training framework designed to significantly reduce memory footprint when training large models. COAT addresses current limitations through two key innovations: (1) **Dynamic Range Expansion**, which aligns optimizer state distributions more closely with the FP8 representation range, thereby reducing quantization error, and (2) **Mixed-Granularity Activation Quantization**, which optimizes activation memory using a combination of per-tensor and per-group quantization strategies. Experiments demonstrate that COAT effectively reduces end-to-end training memory footprint by **1.54\u00d7** compared to BF16 while achieving nearly lossless performance across various tasks, such as Large Language Model pretraining and fine-tuning and Vision Language Model training. COAT also achieves a **1.43\u00d7** end-to-end training speedup compared to BF16, performing on par with or surpassing TransformerEngine's speedup. COAT enables efficient full-parameter training of large models on fewer GPUs, and facilitates doubling the batch size in distributed training settings, providing a practical solution for scaling large-scale model training. Code will be released upon publication.", "keywords": [ "FP8 training", "quantization", "low-precision training", "memory efficient training" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "XeRvg7GQH4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Gelei Xu", "gender": "Female", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Jun Xia", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "Postdoc" }, { "name": "Yiyu Shi", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "Full Professor" }, { "name": "Zixuan Pan", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 12, "n_ref": 37, "n_ref_all": 47, "n_fig": 4, "n_tab": 2, "L_tab": 309, "n_element_tab": 71, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 339, "n_element_tab_1": 74, "formula_len_all": 867, "formula_len_all_1": 814, "len_all": 73333, "len_all_1": 55130, "len_abs": 2232, "len_title": 147, "len_sents": 28506, "len_sents_1": 25753, "n_sents": 170, "n_sents_1": 149, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 2163, "title": "One Training Fits All: Generalized Data Condensation via Mixture-of-Information Bottleneck Guidance", "abs": "Data condensation (DC) technologies are widely used in buffer-constrained scenarios to reduce the memory demand of training samples and maintain DNN training performance. However, due to the storage constraint of deployment devices and the high energy costs of condensation procedure, synthetic datasets generated by DC often have inferior performance in terms of training efficiency and scalability, which greatly limits its practical application on various edge devices. \nThis dilemma arises due to two reasons: i) existing state-of-the-art (SoTA) data condensation approaches that update synthetic datasets by intuitively matching intermediate training outputs (e.g., gradients, features and distributions) between real datasets and synthetic datasets without improving their representational information capabilities from the perspective of the useful information contained. ii) DC lacks sufficient consideration for the heterogeneity of storage constraints among various edge devices, which will result in large training overheads (i.e., consumption or storage). \nTo tackle the above issue, We propose a novel method named Mixture-of-Information Bottleneck Dataset Condensation (MIBDC), which employs information bottlenecks from synthetic datasets with various Image Per Class (IPC) numbers to improve the overall DC generalization and scalability. \nSpecifically, in this paper, the following two phenomena are found: i) The quality of synthetic datasets improves with increased synthetic dataset quantity. ii) The smaller the number of synthetic datasets, the earlier they can reach the convergence peak.\nBased on the above two findings, this paper proposes that i) large synthetic datasets can guide the better convergence of smaller ones. ii) information contained in synthetic datasets with different IPC numbers can play a collaborative role in the guidance of dataset condensation generalization.\nComprehensive experimental results on three well-known datasets show that, compared with state-of-the-art dataset condensation methods, MIBDC can not only enhance the generalization performance of trained models but also achieve superior scalability.", "keywords": [ "Dataset Condensation" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "XeGSIr7z6u", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Cengiz Pehlevan", "gender": "unknown", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "Indranil Halder", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "Associate" } ], "n_formula": 37, "n_formula_1": 23, "n_ref_uni": 38, "n_ref": 57, "n_ref_all": 76, "n_fig": 4, "n_tab": 1, "L_tab": 49, "n_element_tab": 2, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3101, "formula_len_all_1": 2201, "len_all": 99079, "len_all_1": 51418, "len_abs": 1159, "len_title": 126, "len_sents": 25879, "len_sents_1": 22485, "n_sents": 194, "n_sents_1": 165, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1180, "title": "On the onset of memorization to generalization transition in diffusion models", "abs": "As the training set size increases, diffusion models have been observed to transition from memorizing the training dataset to generalizing to and sampling from the underlying data distribution. To study this phenomenon more closely, here, we first present a mathematically principled definition of this transition: the model is said to be in the generalization regime if the generated distribution is closer to the sampling distribution compared to the probability distribution associated with a Gaussian kernel approximation to the training dataset. Then, we develop an analytically tractable diffusion model that features this transition when the training data is sampled from an isotropic Gaussian distribution. Our study reveals that this transition occurs when the distance between the generated and underlying sampling distribution begins to decrease rapidly with the addition of more training samples. This is to be contrasted with an alternative scenario, where the model's memorization performance degrades, but generalization performance doesn't improve. We also provide empirical evidence indicating that realistic diffusion models exhibit the same alignment of scales.", "keywords": [ "diffusion models", "memorization", "generalization", "inductive bias", "curse of dimensionality", "denoising" ], "rating_list": [ 3, 5, 5, 3, 1 ], "soundness_list": [ 2, 3, 3, 4, 1 ], "presentation_list": [ 1, 3, 2, 2, 1 ], "contribution_list": [ 1, 2, 2, 1, 1 ], "confidence_list": [ 4, 3, 3, 3, 4 ] }, { "paper_id": "Xe6UmKMInx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Mohammad Mohammadi Amiri", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "Assistant Professor" }, { "name": "Yi Hung Lim", "gender": "unknown", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "Undergrad student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 20, "n_ref": 42, "n_ref_all": 45, "n_fig": 0, "n_tab": 5, "L_tab": 1001, "n_element_tab": 124, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 1001, "n_element_tab_1": 124, "formula_len_all": 76, "formula_len_all_1": 76, "len_all": 83476, "len_all_1": 66210, "len_abs": 916, "len_title": 88, "len_sents": 28419, "len_sents_1": 28371, "n_sents": 181, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 918, "title": "Latent Diffusion with LLMs for Reasoning", "abs": "Despite the widespread adoption of large language models with hundreds of billions of parameters, these models still struggle on complex reasoning benchmarks. In this paper, we argue that the autoregressive nature of current language models are not suited for reasoning due to fundamental limitations, and that reasoning requires slow accumulation of knowledge through time. We show that combining latent diffusion models with an encoder-decoder transformer architecture provides a scalable way to address some of the fundamental shortcomings posed by autoregressive models. Diffusion models can arrive at predictions through many forward passes in latent space, and their reasoning is not handicapped by the order of the tokens in the dataset. Through our experiments, we show that latent diffusion language models is a feasible approach towards scalable language models that have general complex reasoning abilities.", "keywords": [ "Large Language Models", "Latent Diffusion", "Reasoning" ], "rating_list": [ 3, 5, 1, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "XdcuqZRhjQ", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Hao Zhu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Hitesh Goel", "gender": "Male", "institution": "International Institute of Information Technology, Hyderabad, International Institute of Information Technology Hyderabad", "country": "IN", "position": "Undergrad student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 31, "n_ref": 61, "n_ref_all": 81, "n_fig": 12, "n_tab": 1, "L_tab": 369, "n_element_tab": 64, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 450, "n_element_tab_1": 74, "formula_len_all": 70, "formula_len_all_1": 70, "len_all": 174624, "len_all_1": 67845, "len_abs": 1242, "len_title": 103, "len_sents": 61508, "len_sents_1": 33284, "n_sents": 517, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1222, "title": "LifelongSotopia: Evaluating Social Intelligence Of Language Agents Over Lifelong Social Interactions", "abs": "Humans engage in lifelong social interactions through interacting with different people under different scenarios for different social goals. This requires social intelligence to gather information through a long time span and use it to navigate various social contexts effectively. Whether AI systems are also capable of this is understudied in the existing research. In this paper, we present a novel benchmark, LifelongSotopia, to perform a comprehensive evaluation of language agents by simulating multi-episode interactions. In each episode, the language agents role-play characters to achieve their respective social goals in randomly sampled social tasks. With LifelongSotopia, we find that goal achievement and believability of all of the language models that we test decline through the whole interaction. Although using an advanced memory method improves the agents' performance, the best agents still achieve a significantly lower goal completion rate than humans on scenarios requiring an explicit understanding of interaction history. These findings show that we can use LifelongSotopia to evaluate long-context language models and the social intelligence of language agents over lifelong social interactions.", "keywords": [ "LLMs", "language agents", "social intelligence", "evaluation of LLMs" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "XdRv6I80L1", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Aengus Lynch", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Dylan Xu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Ethan Perez", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Fazl Barez", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Henry Sleight", "gender": "Male", "institution": "Constellation", "country": "US", "position": "Researcher" }, { "name": "Julian Stastny", "gender": "unknown", "institution": "Center on Long-Term Risk", "country": "GB", "position": "Researcher" }, { "name": "Niels Warncke", "gender": "Male", "institution": "Center on Long-Term Risk", "country": "GB", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 36, "n_ref": 46, "n_ref_all": 65, "n_fig": 10, "n_tab": 2, "L_tab": 776, "n_element_tab": 145, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 1400, "n_element_tab_1": 179, "formula_len_all": 145, "formula_len_all_1": 145, "len_all": 248070, "len_all_1": 53174, "len_abs": 1303, "len_title": 91, "len_sents": 83533, "len_sents_1": 24267, "n_sents": 650, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 962, "title": "Plan B: Training LLMs to fail less severely", "abs": "Safety-trained LLMs can produce harmful responses across various input types, as shown by research on jailbreaks, data poisoning, and misalignment. Despite ongoing efforts, fully preventing such failures remains difficult. In this work, we propose a second line of defense: instead of solely focusing on eliminating harmful responses, we also aim to reduce their severity when they occur. As a case study, we experiment with an LLM trained to respond to a backdoor-trigger by complying with harmful requests. We fine-tune the model, without using the trigger in the training data, on the following pairwise preferences: (1) refusal is preferred over any harmful response, (2) less harmful responses are preferred over more harmful ones. We find that training on this preference ordering significantly reduces the harmfulness of backdoor-triggered responses. Finally, we demonstrate that our approach generalizes to several state-of-the-art jailbreak techniques.", "keywords": [ "AI safety", "data poisoning", "alignment", "robustness", "sleeper agents", "model organisms", "jailbreaks" ], "rating_list": [ 3, 3, 3, 6, 5 ], "soundness_list": [ 2, 1, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "XdRIno98gG", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Kyumin Hwang", "gender": "Male", "institution": "Daegu Gyeongbuk Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Minwoo Choi", "gender": "unknown", "institution": "Daegu Gyeongbuk Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sunghoon Im", "gender": "Male", "institution": "Daegu Gyeongbuk Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Wei Peng", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Wonhyeok Choi", "gender": "Male", "institution": "Daegu Gyeongbuk Institute of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 45, "n_ref": 90, "n_ref_all": 111, "n_fig": 4, "n_tab": 4, "L_tab": 5522, "n_element_tab": 466, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 3988, "n_element_tab_1": 315, "formula_len_all": 852, "formula_len_all_1": 755, "len_all": 134629, "len_all_1": 64760, "len_abs": 1278, "len_title": 147, "len_sents": 35330, "len_sents_1": 27263, "n_sents": 227, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 99, "L_abs": 1282, "title": "Self-supervised Monocular Depth Estimation Robust to Reflective Surface Leveraged by Triplet Mining", "abs": "Self-supervised monocular depth estimation (SSMDE) aims to predict the dense depth map of a monocular image, by learning depth from RGB image sequences, eliminating the need for ground-truth depth labels.\nAlthough this approach simplifies data acquisition compared to supervised methods, it struggles with reflective surfaces, as they violate the assumptions of Lambertian reflectance, leading to inaccurate training on such surfaces.\nTo tackle this problem, we propose a novel training strategy for an SSMDE by leveraging triplet mining to pinpoint reflective regions at the pixel level, guided by the camera geometry between different viewpoints.\nThe proposed reflection-aware triplet mining loss specifically penalizes the inappropriate photometric error minimization on the localized reflective regions while preserving depth accuracy on non-reflective areas.\nWe also incorporate a reflection-aware knowledge distillation method that enables a student model to selectively learn the pixel-level knowledge from reflective and non-reflective regions. This results in robust depth estimation across areas.\nEvaluation results on multiple datasets demonstrate that our method effectively enhances depth quality on reflective surfaces and outperforms state-of-the-art SSMDE baselines.", "keywords": [ "Self-supervised Monocular Depth Estimation", "Deep Metric Learning", "Knowledge Distillation" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "Xd2Qxf5RYI", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chenkai ZHANG", "gender": "Male", "institution": "Nagoya University", "country": "JP", "position": "PhD student" }, { "name": "Daisuke Deguchi", "gender": "Male", "institution": "Nagoya University", "country": "JP", "position": "Associate Professor" }, { "name": "Hiroshi Murase", "gender": "Male", "institution": "Nagoya University", "country": "JP", "position": "Emeritus" }, { "name": "Jialei Chen", "gender": "Male", "institution": "Nagoya University", "country": "JP", "position": "PhD student" }, { "name": "Qi Fan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xu Zheng", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 42, "n_ref": 137, "n_ref_all": 152, "n_fig": 6, "n_tab": 4, "L_tab": 606, "n_element_tab": 119, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 867, "n_element_tab_1": 128, "formula_len_all": 698, "formula_len_all_1": 699, "len_all": 129356, "len_all_1": 66703, "len_abs": 1566, "len_title": 123, "len_sents": 35008, "len_sents_1": 30494, "n_sents": 242, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1473, "title": "Semantic-Centric Alignment for Zero-shot Panoptic and Semantic Segmentation", "abs": "Zero-shot segmentation has achieved great success by generating features from semantic embeddings to adapt the model to unseen classes. These semantic-generated features are typically aligned with the visual distribution of seen classes to improve generalization on extracted image features. However, this vision-centric alignment may easily overfit seen classes due to the lack of visual data for unseen classes. To address this issue, we propose a semantic-centric alignment method that aligns the generated features with the well-structured semantic distribution across all classes. First, we align the vision backbone features with CLIP tokens through Vision-to-CLIP alignment. This approach leverages CLIP\u2019s visual-language matching capabilities to produce semantic-aligned backbone features. Then, we generate synthetic features from semantic embeddings for unseen classes, supervised by semantic-aligned visual features and CLIP semantic tokens for improving visual diversity while maintaining semantic consistency. Finally, we finetune the class projector through the semantic-aligned joint features to further adapt the model for unseen classes. Our semantic-centric alignment effectively enhances the model\u2019s zero-shot generalization by constructing a unified and well-structured semantic-aligned feature space. Our method achieves SOTA performance in both zero-shot panoptic and semantic segmentation, and can directly segment unseen classes without fine-tuning.", "keywords": [ "zero-shot segmentation", "zero-shot learning", "feature alignment", "semantic segmentation", "panoptic segmentation", "semantic-centric" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "XcSJ6hoc1O", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Aymen Shabou", "gender": "Male", "institution": "Credit Agricole SA", "country": "FR", "position": "Principal Researcher" }, { "name": "Davide Buscaldi", "gender": "Male", "institution": "Ecole polytechnique", "country": "FR", "position": "Assistant Professor" }, { "name": "J\u00e9r\u00e9mie Dentan", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "FR", "position": "PhD student" }, { "name": "Haddad Vanier Sonia", "gender": "Female", "institution": "\u00c9cole Polytechnique", "country": "FR", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 4, "n_ref_uni": 46, "n_ref": 134, "n_ref_all": 176, "n_fig": 20, "n_tab": 2, "L_tab": 1288, "n_element_tab": 221, "n_fig_1": 14, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1990, "formula_len_all_1": 419, "len_all": 198400, "len_all_1": 70441, "len_abs": 1006, "len_title": 131, "len_sents": 59553, "len_sents_1": 35543, "n_sents": 473, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1011, "title": "Predicting and analyzing memorization within fine-tuned Large Language Models", "abs": "Large Language Models have received significant attention due to their abilities to solve a wide range of complex tasks. However these models memorize a significant proportion of their training data, posing a serious threat when disclosed at inference time. To mitigate this unintended memorization, it is crucial to understand what elements are memorized and why. Most existing works provide a posteriori explanations, which has a limited interest in practice. To address this gap, we propose a new approach to detect memorized samples a priori in LLMs fine-tuned on classification tasks. This method is efficient from the early stages of training and readily adaptable to other classification settings, such as training vision models from scratch. Our method is supported by new theoretical results that we demonstrate, and requires a low computational budget. We obtain strong empirical results, paving the way for systematic inspection and protection of these vulnerable samples before memorization happens.", "keywords": [ "Memorization", "Large Language Model", "Privacy" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "XcFJZORJgY", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Han Sangbum", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Hyeonjin Kim", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Nojun Kwak", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "JunHoo Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 28, "n_ref": 49, "n_ref_all": 71, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 323, "n_element_tab_1": 18, "formula_len_all": 1291, "formula_len_all_1": 1288, "len_all": 104123, "len_all_1": 63831, "len_abs": 1109, "len_title": 151, "len_sents": 30163, "len_sents_1": 26802, "n_sents": 260, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 29, "L_abs": 1117, "title": "Discrimination for Generation", "abs": "There are two primary approaches to learning from data: discriminative models, which make predictions based on provided data, and generative models, which learn data distributions to create new instances. This paper introduces a novel framework, Discrimination for Generation (DFG), as the first attempt to bridge the gap between discriminative and generative models. Through DFG, discriminative models can function as generative models. We leverage the Neural Tangent Kernel (NTK) to map discriminative models into a connected functional space, enabling the calculation of the distance between the data manifold and a sampled data point.\nOur experimental results demonstrate that the proposed algorithm can generate high-fidelity images and can be applied to various tasks such as Targeted Editing and Inpainting, in addition to both unconditional and conditional image generation.\nThis connection provides a novel perspective for interpreting models. Moreover, our method is algorithm-, architecture-, and dataset-agnostic, offering flexibility and proving to be a robust technique across a wide range of scenarios.", "keywords": [ "Generative model", "Discriminative model", "Neural Tangent Kernel", "Functional Analysis" ], "rating_list": [ 8, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "XcAJ0qsMgh", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Dongze Wu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "MS student" }, { "name": "Yao Xie", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 84, "n_formula_1": 35, "n_ref_uni": 63, "n_ref": 90, "n_ref_all": 120, "n_fig": 4, "n_tab": 5, "L_tab": 1604, "n_element_tab": 88, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7290, "formula_len_all_1": 2409, "len_all": 198457, "len_all_1": 58434, "len_abs": 1157, "len_title": 144, "len_sents": 53828, "len_sents_1": 24318, "n_sents": 469, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1162, "title": "Annealing Flow Generative Model Towards Sampling High-Dimensional and Multi-Modal Distributions", "abs": "Sampling from high-dimensional, multi-modal distributions remains a fundamental challenge across domains such as statistical Bayesian inference and physics-based machine learning. In this paper, we propose Annealing Flow (AF), a continuous normalizing flow-based approach designed to sample from high-dimensional and multi-modal distributions. The key idea is to learn a continuous normalizing flow-based transport map, guided by annealing, to transition samples from an easy-to-sample distribution to the target distribution, facilitating effective exploration of modes in high-dimensional spaces. Unlike many existing methods, AF training does not rely on samples from the target distribution. AF ensures effective and balanced mode exploration, achieves linear complexity in sample size and dimensions, and circumvents inefficient mixing times. We demonstrate the superior performance of AF compared to state-of-the-art methods through extensive experiments on various challenging distributions and real-world datasets, particularly in high-dimensional and multi-modal settings. We also highlight AF\u2019s potential for sampling the least favorable distributions.", "keywords": [ "Continuous Normalizing Flow", "Generative Model", "Optimal Transport", "High-Dimensional Sampling", "Multi-Modal Sampling" ], "rating_list": [ 1, 5, 3, 3, 6 ], "soundness_list": [ 1, 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 1, 3 ], "confidence_list": [ 3, 4, 5, 4, 2 ] }, { "paper_id": "Xbl6t6zxZs", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Amith Ananthram", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Elias Stengel-Eskin", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Postdoc" }, { "name": "Kathleen McKeown", "gender": "Female", "institution": "Columbia University, Columbia University", "country": "", "position": "Full Professor" }, { "name": "Mohit Bansal", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 58, "n_ref": 88, "n_ref_all": 112, "n_fig": 8, "n_tab": 2, "L_tab": 628, "n_element_tab": 60, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 401, "n_element_tab_1": 28, "formula_len_all": 125, "formula_len_all_1": 125, "len_all": 155983, "len_all_1": 60215, "len_abs": 1302, "len_title": 134, "len_sents": 45184, "len_sents_1": 29638, "n_sents": 326, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1193, "title": "See It from My Perspective: How Language Affects Cultural Bias in Image Understanding", "abs": "Vision-language models (VLMs) can respond to queries about images in many languages. However, beyond language, culture affects how we see things. For example, individuals from Western cultures focus more on the central figure in an image while individuals from East Asian cultures attend more to scene context (Nisbett 2001). In this work, we characterize the Western bias of VLMs in image understanding and investigate the role that language plays in this disparity. We evaluate VLMs across subjective and objective visual tasks with culturally diverse images and annotations. We find that VLMs perform better on the Western split than on the East Asian split of each task. Through controlled experimentation, we trace one source of this bias in image understanding to the lack of diversity in language model construction. While inference in a language nearer to a culture can lead to reductions in bias, we show it is much more effective when that language was well-represented during text-only pre-training. Interestingly, this yields bias reductions even when prompting in English. Our work highlights the importance of richer representation of all languages in building equitable VLMs.", "keywords": [ "vision-language models", "multilinguality", "cultural bias", "vqa", "emotion classification", "art" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "XazJbPgLcV", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Karthik Elamvazhuthi", "gender": "Male", "institution": "University of California, Riverside", "country": "", "position": "Postdoc" }, { "name": "Rishi Sonthalia", "gender": "Male", "institution": "Boston College", "country": "US", "position": "Assistant Professor" }, { "name": "Shiba Biswal", "gender": "unknown", "institution": "Los Alamos National Laboratory", "country": "US", "position": "Postdoc" } ], "n_formula": 44, "n_formula_1": 42, "n_ref_uni": 37, "n_ref": 58, "n_ref_all": 71, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2781, "formula_len_all_1": 2374, "len_all": 117189, "len_all_1": 65973, "len_abs": 1211, "len_title": 104, "len_sents": 31844, "len_sents_1": 25703, "n_sents": 314, "n_sents_1": 240, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1217, "title": "Identification of Mean-Field Dynamics using Transformers", "abs": "This paper investigates the use of transformer architectures to approximate the mean-field dynamics of interacting particle systems exhibiting collective behavior. Such systems are fundamental in modeling phenomena across physics, biology, and engineering, including gas dynamics, opinion formation, biological networks, and swarm robotics. The key characteristic of these systems is that the particles are indistinguishable, leading to permutation-equivariant dynamics. We demonstrate that transformers, which inherently possess permutation equivariance, are well-suited for approximating these dynamics. Specifically, we prove that if a finite-dimensional transformer can effectively approximate the finite-dimensional vector field governing the particle system, then the expected output of this transformer provides a good approximation for the infinite-dimensional mean-field vector field. Leveraging this result, we establish theoretical bounds on the distance between the true mean-field dynamics and those obtained using the transformer. We validate our theoretical findings through numerical simulations on the Cucker-Smale model for flocking, and the mean-field system for training two-layer neural networks.", "keywords": [ "Mean field dynamics", "Transformers", "Universal approximation" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 1, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "Xagys9QD3T", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Nuno Vasconcelos", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Professor" }, { "name": "Wenqing Zhang", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "MS student" }, { "name": "Yijiang Li", "gender": "Not Specified", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Yinzhi Cao", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Associate Professor" }, { "name": "Yuchen Yang", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Zihao Zhao", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "Undergrad student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 44, "n_ref_all": 59, "n_fig": 3, "n_tab": 3, "L_tab": 1065, "n_element_tab": 229, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 770, "formula_len_all_1": 727, "len_all": 92681, "len_all_1": 52814, "len_abs": 1383, "len_title": 92, "len_sents": 29194, "len_sents_1": 26162, "n_sents": 206, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1391, "title": "Pseudo-Probability Unlearning: Towards Efficient and Privacy-Preserving Machine Unlearning", "abs": "Machine unlearning\u2014enabling a trained model to forget specific data\u2014is crucial for addressing biased data and adhering to privacy regulations like the General Data Protection Regulation (GDPR)'s ``right to be forgotten.\" Recent works have paid little attention to privacy concerns, leaving the data intended for forgetting vulnerable to membership inference attacks. Moreover, they often come with high computational overhead. In this work, we propose Pseudo-Probability Unlearning (PPU), a novel method that enables models to forget data efficiently and in a privacy-preserving manner. Our method replaces the final-layer output probabilities of the neural network with pseudo-probabilities for the data to be forgotten. These pseudo-probabilities follow either a uniform distribution or align with the model\u2019s overall distribution, enhancing privacy and reducing risk of membership inference attacks. Our optimization strategy further refines the predictive probability distributions and updates the model's weights accordingly, ensuring effective forgetting with minimal impact on the model's overall performance. Through comprehensive experiments on multiple benchmarks, our method achieves over 20\\% improvements in forgetting error compared to the state-of-the-art. Additionally, our method enhances privacy by preventing the forgotten set from being inferred to around random guesses.", "keywords": [ "machine unlearning", "deep learning" ], "rating_list": [ 3, 3, 5, 1 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "XaYCOY7YlU", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Bonnie Berger", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Emma Pierson", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "Assistant Professor" }, { "name": "Nikhil Garg", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Assistant Professor" }, { "name": "Shuvom Sadhuka", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Sidhika Balachandar", "gender": "Female", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 46, "n_ref": 69, "n_ref_all": 101, "n_fig": 9, "n_tab": 3, "L_tab": 266, "n_element_tab": 20, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 996, "n_element_tab_1": 19, "formula_len_all": 660, "formula_len_all_1": 661, "len_all": 135465, "len_all_1": 69286, "len_abs": 1413, "len_title": 115, "len_sents": 38785, "len_sents_1": 33838, "n_sents": 282, "n_sents_1": 241, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1419, "title": "Using GNNs to Model Biased Crowdsourced Data for Urban Applications", "abs": "Graph neural networks (GNNs) are widely used to make predictions on graph-structured data in urban spatiotemporal forecasting applications, such as predicting infrastructure problems and weather events. In urban settings, nodes have a true latent state (e.g., street condition) that is sparsely observed (e.g., via government inspection ratings). We more frequently observe biased proxies for the latent state (e.g., via crowdsourced reports) that correlate with resident demographics. We introduce a GNN-based model that uses both unbiased rating data and biased reporting data to predict the true latent state. We show that our approach can both recover the latent state at each node and quantify the reporting biases. We apply our model to a case study of urban incidents using reporting data from New York City 311 complaints across 141 complaint types and rating data from government inspections. We show (i) that our model predicts more correlated ground truth latent states compared to prior work which trains models only on the biased reporting data, (ii) that our model's inferred reporting biases capture known demographic biases, and (iii) that our model's learned ratings capture correlations across locations and between complaint types. Especially in urban crowdsourcing applications, our analysis reveals a widely applicable approach for using GNNs and sparse ground truth data to estimate latent states.", "keywords": [ "Graph Neural Networks", "Biased Outcome Data", "Urban Planning" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 3, 1, 2 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "XaARrKTNh3", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Anastasios Kyrillidis", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Associate Professor" }, { "name": "Junhyung Lyle Kim", "gender": "Male", "institution": "J.P. Morgan Chase", "country": "US", "position": "Researcher" }, { "name": "Nai-Hui Chia", "gender": "Male", "institution": "Rice University", "country": "", "position": "Assistant Professor" } ], "n_formula": 59, "n_formula_1": 39, "n_ref_uni": 44, "n_ref": 102, "n_ref_all": 132, "n_fig": 3, "n_tab": 4, "L_tab": 258, "n_element_tab": 6, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 634, "n_element_tab_1": 27, "formula_len_all": 2909, "formula_len_all_1": 1877, "len_all": 127507, "len_all_1": 65711, "len_abs": 1832, "len_title": 140, "len_sents": 38140, "len_sents_1": 27689, "n_sents": 309, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1147, "title": "A Catalyst Framework for the Quantum Linear System Problem via the Proximal Point Algorithm", "abs": "Solving systems of linear equations is a fundamental problem, but it can be computationally intensive for classical algorithms in high dimensions. Existing quantum algorithms can achieve exponential speedups for the quantum linear system problem (QLSP) in terms of the problem dimension, but even such a theoretical advantage is bottlenecked by the condition number of the coefficient matrix. In this work, we propose a new quantum algorithm for QLSP inspired by the classical proximal point algorithm (PPA). Our proposed method can be viewed as a meta-algorithm that allows inverting a modified matrix via an existing \\texttt{QLSP\\_solver}, thereby directly approximating the solution vector instead of approximating the inverse of the coefficient matrix. By carefully choosing the step size $\\eta$, the proposed algorithm can effectively precondition the linear system to mitigate the dependence on condition numbers that hindered the applicability of previous approaches. Importantly, this is the first framework for QLSP where a tunable parameter $\\eta$ allows the user to control the trade-off between the runtime and the approximation error.", "keywords": [ "quantum linear system problem", "proximal point algorithm", "quantum algorithm", "catalyst" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "XZhpS5Imzx", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Haris Vikalo", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Full Professor" }, { "name": "Usman Akram", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" } ], "n_formula": 52, "n_formula_1": 34, "n_ref_uni": 23, "n_ref": 42, "n_ref_all": 64, "n_fig": 12, "n_tab": 4, "L_tab": 355, "n_element_tab": 22, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2848, "formula_len_all_1": 1883, "len_all": 125255, "len_all_1": 63099, "len_abs": 1223, "len_title": 120, "len_sents": 40380, "len_sents_1": 26877, "n_sents": 317, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1252, "title": "CAN TRANSFORMERS IN-CONTEXT LEARN BEHAVIOR OF A LINEAR DYNAMICAL SYSTEM?", "abs": "We investigate whether transformers can learn to track a random process when given observations of a related process and parameters of the dynamical system that relates them as context. More specifically, we consider a finite-dimensional state-space model described by the state transition matrix $F$, measurement matrices $h_1, \\dots, h_N$, and the process and measurement noise covariance matrices $Q$ and $R$, respectively; these parameters, randomly sampled, are provided to the transformer along with the observations $y_1,\\dots,y_N$ generated by the corresponding linear dynamical system. We argue that in such settings transformers learn to approximate the celebrated Kalman filter, and empirically verify this both for the task of estimating hidden states $\\hat{x_{N|1,2,3,...,N}}$ as well as for one-step prediction of the $(N+1)^{st}$ observation, $\\hat{y}_{N+1|1,2,3,...,N}$. A further study of the transformer's robustness reveals that its performance is retained even if the model's parameters are partially withheld. In particular, we demonstrate that the transformer remains accurate at the considered task even in the absence of state transition and noise covariance matrices, effectively emulating operations of the Dual-Kalman filter.", "keywords": [ "In-context Learning", "Transformers", "Kalman Filter" ], "rating_list": [ 3, 6, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 2, 1 ], "presentation_list": [ 2, 3, 3, 1, 2 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 3, 2, 4, 4, 3 ] }, { "paper_id": "XZ71GHf8aB", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Anand V. Shah", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "David C. Parkes", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "John Joseph Horton", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "First Job " }, { "name": "Kehang Zhu", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Yanchen Jiang", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 0, "n_ref": 47, "n_ref_all": 69, "n_fig": 6, "n_tab": 6, "L_tab": 3376, "n_element_tab": 97, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 107, "n_element_tab_1": 12, "formula_len_all": 271, "formula_len_all_1": 249, "len_all": 174118, "len_all_1": 63116, "len_abs": 1213, "len_title": 127, "len_sents": 59381, "len_sents_1": 29995, "n_sents": 533, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1222, "title": "Evidence from the Synthetic Laboratory: Language Models as Auction Participants", "abs": "This paper investigates the behavior of simulated AI agents (large language mod-\nels, or LLMs) in auctions, validating a novel synthetic data-generating process\nto help discipline the study and design of auctions. We begin by benchmarking\nthese LLM agents against established experimental results that study agreement or\ndeparture between realized economic behavior and predictions from theory; i.e.,\nrevenue equivalence between first-price and second-price auctions and improved\nplay in obviously strategy-proof auctions. We find that when LLM-based agents\ndiverge from the predictions of theory, they do so in a way that agrees with behav-\nioral traits observed in the existing experimental economics literature (e.g., risk\naversion, and weak play in \u2018complicated\u2019 auctions). Our results also suggest that\nLLMs are bad at playing auctions \u2018out of the box\u2019 but can improve their play when\ngiven the opportunity to learn. This learning is robust to various prompt specifi-\ncations and holds across a variety of settings. We run 2,000+ auctions for less\nthan $250 with GPT-4o and GPT-4, and develop a framework flexible enough to\nrun auction experiments with any LLM model and a wide range of auction design\nspecifications.", "keywords": [ "Language Model", "Auction", "Behavioral Economics", "learning" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "XYuWS3nrw3", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Jean-michel Morel", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Raymond H. Chan", "gender": "Male", "institution": "Lingnan University", "country": "HK", "position": "Full Professor" }, { "name": "Tieyong Zeng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Xiaodong Cun", "gender": "Male", "institution": "Great Bay University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yang Liu", "gender": "Male", "institution": "National University of Defense Technology", "country": "", "position": "Assistant Professor" }, { "name": "Yaofang Liu", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yumeng REN", "gender": "Male", "institution": "City University", "country": "CN", "position": "PhD student" }, { "name": "Aitor Alexandre Artola", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Postdoc" } ], "n_formula": 15, "n_formula_1": 16, "n_ref_uni": 28, "n_ref": 73, "n_ref_all": 88, "n_fig": 7, "n_tab": 2, "L_tab": 21, "n_element_tab": 4, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 237, "n_element_tab_1": 30, "formula_len_all": 1110, "formula_len_all_1": 1134, "len_all": 106236, "len_all_1": 53818, "len_abs": 619, "len_title": 130, "len_sents": 28986, "len_sents_1": 24054, "n_sents": 185, "n_sents_1": 161, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1412, "title": "Redefining Temporal Modeling in Video Diffusion: The Vectorized Timestep Approach", "abs": "Diffusion models have revolutionized image generation, and their extension to video generation has shown promise. However, current video diffusion models (VDMs) rely on a scalar timestep variable applied at the clip level, which limits their ability to model complex temporal dependencies needed for various tasks like image-to-video generation. To address this limitation, we propose a frame-aware video diffusion model (FVDM), which introduces a novel vectorized timestep variable (VTV). Unlike conventional VDMs, our approach allows each frame to follow an independent noise schedule, enhancing the model's capacity to capture fine-grained temporal dependencies.\nFVDM's flexibility is demonstrated across multiple tasks, including standard video generation, image-to-video generation, video interpolation, and long video synthesis. Through a diverse set of VTV configurations, we achieve superior quality in generated videos, overcoming challenges such as catastrophic forgetting during fine-tuning and limited generalizability in zero-shot methods.\nOur empirical evaluations show that FVDM outperforms state-of-the-art methods in video generation quality, while also excelling in extended tasks. By addressing fundamental shortcomings in existing VDMs, FVDM sets a new paradigm in video synthesis, offering a robust framework with significant implications for generative modeling and multimedia applications.", "keywords": [ "Diffusion Model", "Video Generation", "Image-to-Video Generation", "Video Interpolation", "Long Video Generation", "Zero-Shot" ], "rating_list": [ 1, 5, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 1, 2 ], "contribution_list": [ 1, 1, 3 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "XYdstv3ySl", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Jianglong Ye", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Ri-Zhao Qiu", "gender": "Not Specified", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Sifei Liu", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Xiaolong Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Xuanbin Peng", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Undergrad student" }, { "name": "Xueyan Zou", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Postdoc" }, { "name": "Yuchen Song", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 22, "n_ref": 38, "n_ref_all": 58, "n_fig": 8, "n_tab": 3, "L_tab": 2058, "n_element_tab": 85, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 2058, "n_element_tab_1": 85, "formula_len_all": 428, "formula_len_all_1": 428, "len_all": 121528, "len_all_1": 66704, "len_abs": 327, "len_title": 76, "len_sents": 29491, "len_sents_1": 30029, "n_sents": 214, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 28, "L_abs": 1464, "title": "3D-SPATIAL MULTIMODAL MEMORY", "abs": "We present 3D Spatial MultiModal Memory (M3), a multimodal memory system designed to retain information about medium-sized static scenes through video sources for visual perception. By integrating 3D Gaussian Splatting techniques with foundation models, M3 builds a multimodal memory capable of rendering feature representations across granularities, encompassing a wide range of knowledge. In our exploration, we identify two key challenges in previous works on feature splatting: (1) computational constraints in storing high-dimensional features for each Gaussian primitive, and (2) misalignment or information loss between distilled features and foundation model features. To address these challenges, we propose M3 with key components of principal scene components and Gaussian memory attention, enabling efficient training and inference. To validate M3, we conduct comprehensive quantitative evaluations of feature similarity and downstream tasks, as well as qualitative visualizations to highlight the pixel trace of Gaussian memory attention. Our approach encompasses a diverse range of foundation models, including vision-language models (VLMs), perception models, and large multimodal and language models (LMMs/LLMs). Furthermore, to demonstrate real-world applicability, we deploy M3\u2019s feature field in indoor scenes on a quadruped robot. Notably, we claim that M3 is the first work to address the core compression challenges in 3D feature distillation.", "keywords": [ "Foundation Model", "Gaussian Splatting", "Large Multimodal Model", "Robotics" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "XYRPm8rAGM", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Xiaoqian Wang", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" }, { "name": "Yipei Wang", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" } ], "n_formula": 74, "n_formula_1": 22, "n_ref_uni": 35, "n_ref": 76, "n_ref_all": 104, "n_fig": 12, "n_tab": 2, "L_tab": 45, "n_element_tab": 5, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 45, "n_element_tab_1": 5, "formula_len_all": 8070, "formula_len_all_1": 1685, "len_all": 156899, "len_all_1": 73371, "len_abs": 2103, "len_title": 141, "len_sents": 51047, "len_sents_1": 32955, "n_sents": 499, "n_sents_1": 291, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 2112, "title": "Agree to Disagree: Demystifying Homogeneous Deep Ensembles through Distributional Equivalence", "abs": "Deep ensembles improve the performance of the models by taking the average predictions of a group of ensemble members. However, the origin of these capabilities remains a mystery and deep ensembles are used as a reliable \u201cblack box\u201d to improve the performance. Existing studies typically attribute such improvement to Jensen gaps of the deep ensemble method, where the loss of the mean does not exceed the mean of the loss for any convex loss metric. In this work, we demonstrate that Jensen\u2019s inequality is not responsible for the effectiveness of deep ensembles, and convexity is not a necessary condition. Instead, Jensen Gap focuses on the \u201caverage loss\u201d of individual models, which provides no practical meaning. Thus it fails to explain the core phenomena of deep ensembles such as the superiority to any single ensemble member, the decreasing loss with the number of ensemble members, etc. Regarding this mystery, we provide theoretical analysis and comprehensive empirical results from a statistical perspective that reveal the true mechanism of deep ensembles. Our results highlight that deep ensembles originate from the homogeneous output distribution across all ensemble members. Specifically, the predictions of homogeneous models (Abe et al., 2022b) have the distributional equivalence property \u2013 Although the predictions of independent ensemble members are point-wise different, they form an identical distribution. Such agreement and disagreement contribute to deep ensembles\u2019 \u201cmagical power\u201d. Based on this discovery, we provide rigorous proof of the effectiveness of deep ensembles and analytically quantify the extent to which ensembles improve performance. The derivations not only theoretically quantify the effectiveness of deep ensembles for the first time, but also enable estimation schemes that foresee the performance of ensembles with different capacities. Furthermore, different from existing studies, our results also point out that deep ensembles work in a different mechanism from model scaling a single model, even though significant correlations between them have been observed.", "keywords": [ "deep ensemble", "trustworthiness", "distributional equivalence" ], "rating_list": [ 8, 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 3, 4, 4, 3, 2 ], "contribution_list": [ 4, 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 3, 5 ] }, { "paper_id": "XYMfoM760h", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Eran Hirsch", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "PhD student" }, { "name": "Gal Chechik", "gender": "unknown", "institution": "Bar Ilan University", "country": "IL", "position": "Full Professor" }, { "name": "Hilit Segev", "gender": "Female", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" }, { "name": "Lital Binyamin", "gender": "Female", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" }, { "name": "Royi Rassin", "gender": "Male", "institution": "Research, Google", "country": "IL", "position": "Intern" }, { "name": "Yoad Tewel", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 74, "n_ref_all": 106, "n_fig": 16, "n_tab": 9, "L_tab": 1238, "n_element_tab": 274, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 795, "n_element_tab_1": 220, "formula_len_all": 259, "formula_len_all_1": 133, "len_all": 142323, "len_all_1": 65432, "len_abs": 1678, "len_title": 122, "len_sents": 43897, "len_sents_1": 29688, "n_sents": 377, "n_sents_1": 238, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1418, "title": "Make It Count: Text-to-Image Generation with an Accurate Number of Objects", "abs": "Despite the unprecedented success of text-to-image diffusion models, controlling the number of depicted objects using text is surprisingly hard. This is important for various applications from technical documents, to children's books to illustrating cooking recipes. Generating object-correct counts is fundamentally challenging because the generative model needs to keep a sense of separate identity for every instance of the object, even if several objects look identical or overlap, and then carry out a global computation implicitly during generation. It is still unknown if such representations exist. To address count-correct generation, we first identify features within the diffusion model that can carry the object identity information. We then use them to separate and count instances of objects during the denoising process and detect over-generation and under-generation. We fix the latter by training a model that predicts both the shape and location of a missing object, based on the layout of existing ones, and show how it can be used to guide denoising with correct object count. Our approach, CountGen, does not depend on external source to determine object layout, but rather uses the prior from the diffusion model itself, creating prompt-dependent and seed-dependent layouts. Evaluated on two benchmark datasets, we find that CountGen strongly outperforms the count-accuracy of existing baselines.", "keywords": [ "Text to Image", "Computer Vision", "Generative Models" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "XYK1eGjahp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Chris Esposo", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Jacob Abernethy", "gender": "Male", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Leyan Pan", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Vijay Ganesh", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Wenke Lee", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 48, "n_formula_1": 22, "n_ref_uni": 15, "n_ref": 41, "n_ref_all": 55, "n_fig": 4, "n_tab": 3, "L_tab": 1944, "n_element_tab": 43, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2147, "n_element_tab_1": 23, "formula_len_all": 4141, "formula_len_all_1": 1998, "len_all": 175323, "len_all_1": 63920, "len_abs": 801, "len_title": 105, "len_sents": 63350, "len_sents_1": 28183, "n_sents": 553, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 806, "title": "Can Transformers Reason Logically? A Study in SAT Solving", "abs": "We theoretically and empirically study the logical reasoning capabilities of LLMs in the context of the Boolean satisfiability (SAT) problem. \nFirst, we construct a non-uniform class of decoder-only Transformers that can solve 3-SAT using backtracking and deduction via Chain-of-Thought (CoT). We prove its correctness by showing trace equivalence to the well-known DPLL SAT-solving algorithm. Second, to support the implementation of this abstract construction, we design a compiler PARAT that takes a procedural specification as input and outputs a transformer model to implement this specification. Third, rather than programming a transformer to reason, we evaluate empirically whether it can be trained to do so by learning directly from algorithmic traces (``reasoning paths'') of the DPLL algorithm.", "keywords": [ "Transformer LLMs", "Logical Reasoning", "Chain-of-Thought", "SAT Solving", "Backtracking" ], "rating_list": [ 5, 5, 5, 3, 8 ], "soundness_list": [ 2, 3, 2, 2, 4 ], "presentation_list": [ 2, 2, 3, 2, 4 ], "contribution_list": [ 3, 2, 1, 1, 4 ], "confidence_list": [ 2, 3, 3, 3, 3 ] }, { "paper_id": "XYG98d5bCI", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Ruixin Guo", "gender": "Male", "institution": "Kent State University", "country": "US", "position": "PhD student" }, { "name": "Ruoming Jin", "gender": "Male", "institution": "Kent State University", "country": "", "position": "Professor" }, { "name": "Xinyu Li", "gender": "Male", "institution": "Kent State University", "country": "", "position": "PhD student" }, { "name": "Yang Zhou", "gender": "unknown", "institution": "Auburn University", "country": "US", "position": "Associate Professor" } ], "n_formula": 83, "n_formula_1": 43, "n_ref_uni": 36, "n_ref": 66, "n_ref_all": 77, "n_fig": 0, "n_tab": 3, "L_tab": 1412, "n_element_tab": 174, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 261, "n_element_tab_1": 29, "formula_len_all": 16105, "formula_len_all_1": 6539, "len_all": 139269, "len_all_1": 65240, "len_abs": 1644, "len_title": 123, "len_sents": 36093, "len_sents_1": 22608, "n_sents": 340, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1094, "title": "On PAC-Bayes Bounds for Linear Autoencoders", "abs": "Linear Autoencoders (LAEs) have shown strong performance in state-of-the-art recommender systems. Some LAE models, like EASE, can be viewed as multivariate (multiple-output) linear regression models with a zero-diagonal constraint. However, these impressive results are mainly based on experiments, with little theoretical support. This paper investigates the generalizability -- a theoretical measure of model performance in statistical machine learning -- of multivariate linear regression and LAEs. We first propose a PAC-Bayes bound for multivariate linear regression, which is generalized from an earlier PAC-Bayes bound for single-output linear regression by Shalaeva et al., and outline sufficient conditions that ensure its theoretical convergence. We then apply this bound to EASE, a classic LAE model in recommender systems, and develop a practical method for minimizing the bound, addressing the calculation challenges posed by the zero-diagonal constraint. Experimental results show that our bound for EASE is non-vacuous on real-world datasets, demonstrating its practical utility.", "keywords": [ "PAC-Bayes bound", "linear regression", "linear autoencoder", "recommender system" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 1 ] }, { "paper_id": "XYFBmp08sP", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Aayush Mishra", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Adam Kortylewski", "gender": "unknown", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Research Group Leader" }, { "name": "Alan Yuille", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Anqi Liu", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" }, { "name": "Prakhar Kaushik", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 32, "n_ref": 95, "n_ref_all": 131, "n_fig": 2, "n_tab": 5, "L_tab": 2827, "n_element_tab": 564, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 171, "n_element_tab_1": 19, "formula_len_all": 1032, "formula_len_all_1": 911, "len_all": 152919, "len_all_1": 68549, "len_abs": 1260, "len_title": 142, "len_sents": 39687, "len_sents_1": 33534, "n_sents": 288, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1266, "title": "CIDA3D: Conformal Inference aided unsupervised Domain Adaptation for 3D-Aware Classification", "abs": "Cognitive Science studies show that human perception becomes robust to occlusions and other nuisances due to internal 3D representations of objects. This idea has been incorporated into computer vision models to improve their ability to understand and reason about the 3D world. However, collecting 3D annotations in vision datasets is expensive. This makes the robustness of the perception model to distribution shifts challenging. We introduce Conformal Inference aided unsupervised Domain Adaptation (CIDA)-3D for the complex setting of multiclass pose estimation. Our method adapts category level pose estimation (3D) models in nuisance ridden target domains directly from images without class label information, by harnessing uncertainty in model predictions (using conformal sets). This allows for significantly better and computationally efficient adaptation to target domains with synthetic and real-world noise. We also show a robust adaptation from fully synthetic data to complex real-world domains. To the best of our knowledge, this method is the first to attempt unsupervised domain adaptation for robust 3D-aware classification and multiclass pose estimation in real-world scenarios by adapting models trained on procedurally generated synthetic data.", "keywords": [ "UDA", "3D pose estimation", "3D-Aware classification", "occlusion", "robustness" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 3, 2, 3, 2 ] }, { "paper_id": "XXzOzJRyOZ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jingwen Chen", "gender": "Female", "institution": "JD.Com, Inc.", "country": "CN", "position": "Researcher" }, { "name": "Siqi Wan", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Tao Mei", "gender": "Male", "institution": "JD Explore Academy", "country": "CN", "position": "Technical Vice President" }, { "name": "Ting Yao", "gender": "Male", "institution": "JD AI Research", "country": "CN", "position": "Researcher" }, { "name": "Yingwei Pan", "gender": "Male", "institution": "JD.com", "country": "CN", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 37, "n_ref": 52, "n_ref_all": 86, "n_fig": 13, "n_tab": 12, "L_tab": 4539, "n_element_tab": 483, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2598, "n_element_tab_1": 370, "formula_len_all": 858, "formula_len_all_1": 787, "len_all": 151635, "len_all_1": 66774, "len_abs": 1386, "len_title": 123, "len_sents": 40495, "len_sents_1": 28199, "n_sents": 279, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1393, "title": "Incorporating Visual Correspondence into Diffusion Model for Visual Try-On", "abs": "Diffusion models have shown preliminary success in virtual try-on (VTON) task. The typical dual-branch architecture comprises two UNets for implicit garment deformation and synthesized image generation respectively, and has emerged as the recipe for VTON task. Nevertheless, the problem remains challenging to preserve the shape and every detail of the given garment due to the intrinsic stochasticity of diffusion model. To alleviate this issue, we novelly propose to explicitly capitalize on visual correspondence as the prior to tame diffusion process instead of simply feeding the whole garment into UNet as the appearance reference. Specifically, we interpret the fine-grained appearance and texture details as a set of structured semantic points, and match the semantic points rooted in garment to the ones over target person through local flow warping. Such 2D points are then augmented into 3D-aware cues with depth/normal map of target person. The correspondence mimics the way of putting clothing on human body and the 3D-aware cues act as semantic point matching to supervise diffusion model training. A point-focused diffusion loss is further devised to fully take the advantage of semantic point matching. Extensive experiments demonstrate strong garment detail preservation of our approach, evidenced by state-of-the-art VTON performances on both VITON-HD and DressCode datasets.", "keywords": [ "Virtual Try-On", "Image Generation" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "XXVRkPB1tg", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Alex Xie", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Carolyn Rose", "gender": "Female", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Daniel Fried", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Researcher" }, { "name": "Divyanshu Sheth", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Pengfei Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Associate Professor" }, { "name": "Yiqing Xie", "gender": "Not Specified", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 73, "n_ref_all": 97, "n_fig": 10, "n_tab": 11, "L_tab": 5101, "n_element_tab": 274, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 3454, "n_element_tab_1": 32, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 180084, "len_all_1": 54129, "len_abs": 1196, "len_title": 122, "len_sents": 50939, "len_sents_1": 24092, "n_sents": 417, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1274, "title": "CodeBenchGen: Creating Scalable Execution-based Code Generation Benchmarks", "abs": "To adequately test modern code generation systems, evaluation benchmarks must execute and test the code generated by the system. However, these execution and testing requirements have largely limited benchmarks to settings where code is easily executable or has human-written tests. To facilitate evaluation of code generation systems across diverse scenarios, we present CodeBenchGen, a framework to create scalable execution-based benchmarks from naturally occurring code sources. Specifically, we leverage a large language model (LLM) to sandbox arbitrary pieces of code into evaluation examples, including test cases for execution-based evaluation. We illustrate the usefulness of our framework by creating a dataset, Exec-CSN, which includes 1,931 examples involving 293 libraries converted from code in 367 GitHub repositories taken from the Code- SearchNet dataset. To demonstrate the solvability of examples in Exec-CSN, we present a human study demonstrating that 81.3% of the examples can be solved by humans and 61% are rated as \u201crequires effort to solve\u201d. We conduct code generation experiments on open-source and proprietary models and analyze the performance of both humans and models. We provide code and data at: https://github.com/CodeBenchGen/CodeBenchGen.", "keywords": [ "code generation", "code generation benchmark", "evaluation for code generation", "synthetic data generation" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "XWb6dPuhmC", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Hongtao Li", "gender": "Male", "institution": "China Internet Network Information Center", "country": "CN", "position": "Principal Researcher" }, { "name": "Jianfan Chen", "gender": "Male", "institution": "China Internet Network Information Center", "country": "CN", "position": "Researcher" }, { "name": "Kejun Dong", "gender": "Male", "institution": "China Internet Network Information Center", "country": "CN", "position": "Researcher" }, { "name": "Xuebiao Yuchi", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Zhiwei Yan", "gender": "Male", "institution": "cnnic", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 22, "n_ref": 30, "n_ref_all": 37, "n_fig": 2, "n_tab": 4, "L_tab": 1157, "n_element_tab": 138, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1157, "n_element_tab_1": 138, "formula_len_all": 160, "formula_len_all_1": 160, "len_all": 85217, "len_all_1": 52352, "len_abs": 974, "len_title": 94, "len_sents": 28051, "len_sents_1": 24057, "n_sents": 174, "n_sents_1": 148, "venue": "ICLR2025", "decision": "withdraw", "L_title": 46, "L_abs": 978, "title": "DUAL-TASK VAE FOR NODE-LEVEL DATA AUGMENTATION", "abs": "Graph Neural Networks (GNNs) have shown great promise in processing graph-structured data, but they often require large amounts of labeled data and are sensitive to noise. In this paper, we propose a novel node-level data augmentation approach that leverages a Variational Autoencoder (VAE) within a dual-task learning framework to address these challenges. Our method utilizes the VAE to generate enriched node representations that capture both structural and feature-related information, which are then combined with the original node features for classification by a Graph Attention Network (GAT). Experiments conducted on the Cora, Citeseer, and Pubmed datasets show that our approach outperforms baseline models, achieving up to 7.3\\% higher accuracy in Pubmed, and surpassing recent state-of-the-art data augmentation techniques. This work highlights the effectiveness of dual-task learning for robust feature enhancement and advances data augmentation strategies in GNNs.", "keywords": [ "VAE", "GNNs", "Graph Data Augmentation", "Node Classification", "Dual-task Training", "Loss Weight Adjustment" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 5, 5, 4, 2 ] }, { "paper_id": "XWaI6FLVgi", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Eric P. Xing", "gender": "Male", "institution": "Mohamed bin Zayed Univeristy of AI", "country": "AE", "position": "Full Professor" }, { "name": "Le Song", "gender": "Male", "institution": "College of Computing, Georgia Institute of Technology", "country": "", "position": "Associate Professor" }, { "name": "Tianjun Yao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Intern" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Yongqiang Chen", "gender": "unknown", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Zhiqiang Shen", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" } ], "n_formula": 54, "n_formula_1": 14, "n_ref_uni": 59, "n_ref": 163, "n_ref_all": 180, "n_fig": 5, "n_tab": 12, "L_tab": 4806, "n_element_tab": 297, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2449, "n_element_tab_1": 69, "formula_len_all": 3677, "formula_len_all_1": 750, "len_all": 225776, "len_all_1": 71437, "len_abs": 1614, "len_title": 124, "len_sents": 58701, "len_sents_1": 33201, "n_sents": 439, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1619, "title": "Diversifying Spurious Subgraphs for Graph Out-of-Distribution Generalization", "abs": "Environment augmentation methods have gained some success in overcoming the out-of-distribution (OOD) generalization challenge in Graph Neural Networks (GNNs). Yet, there exists a challenging trade-off in the augmentation: On one hand, it requires the generated graphs as diverse as possible to extrapolate to unseen environments. On the other hand, it requires the generated graphs to preserve the invariant substructures causally related to the targets. Existing approaches have proposed various environment augmentation strategies to enrich spurious patterns for OOD generalization. However, we argue that these methods remain limited in diversity and precision of the generated environments for two reasons: i) the deterministic nature of the graph composition strategy used for environment augmentation may limit the diversity of the generated environments, and ii) the presence of spurious correlations may lead to the exclusion of invariant subgraphs and reduce the precision of the generated environments. To address this trade-off, we propose a novel paradigm that accurately identifies spurious subgraphs, and an environment augmentation strategy called spurious subgraph diversification, which extrapolates to maximally diversified spurious subgraphs by randomizing the spurious subgraph generation, while preserving the invariant substructures. Our method is theoretically sound and demonstrates strong empirical performance on both synthetic and real-world datasets, outperforming the second-best method by up to 24.19% across 17 baseline methods, underscoring its superiority in graph OOD generalization.", "keywords": [ "OOD generalization", "invariant learning", "graph neural networks" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 4, 2 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "XWPp9FJ0uJ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Davin Jeong", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "Undergrad student" }, { "name": "Hanspeter Pfister", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Junsik Kim", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoctoral fellow" }, { "name": "Zhiyi Shi", "gender": "Male", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "US", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 26, "n_ref": 46, "n_ref_all": 65, "n_fig": 4, "n_tab": 6, "L_tab": 2434, "n_element_tab": 247, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2434, "n_element_tab_1": 247, "formula_len_all": 135, "formula_len_all_1": 99, "len_all": 132054, "len_all_1": 69709, "len_abs": 1965, "len_title": 150, "len_sents": 36484, "len_sents_1": 32453, "n_sents": 260, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 1973, "title": "Surprisingly Simple: Large Language Models are Zero-Shot Feature Extractors for Tabular and Text Data", "abs": "Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks, yet their application to tabular data prediction remains relatively underexplored. This is partly due to the fact that recent LLMs are autoregressive models, generating text outputs. Converting tabular data into text, and vice versa, is not straightforward, making direct application of LLMs to complex tabular prediction difficult. Although previous works have utilized pre-trained embedding models like BERT and its variants for fine-tuning on tabular tasks, the potential of autoregressive LLMs for tabular prediction has been explored only on a limited scale and with simpler datasets.\nIn this paper, we propose Zero-shot Encoding for Tabular data with LLMs (ZET-LLM), a surprisingly simple yet effective approach that leverages pre-trained LLMs as zero-shot feature extractors for tabular prediction tasks. To adapt autoregressive LLMs for this purpose, we replace autoregressive masking with bidirectional attention to treat them as feature embedding models. To address the challenge of encoding high-dimensional complex tabular data with LLMs' limited token lengths, we introduce a feature-wise serialization, where each feature is represented as a single token, and the resulting tokens are combined into a unified sample representation. Additionally, we apply missing value masking to handle missing data, a common issue in complex tabular datasets.\nWe demonstrate that LLMs can serve as powerful zero-shot feature extractors without the need for fine-tuning, extensive data pre-processing, or task-specific instructions. Our method enables LLMs to process both structured tabular data and unstructured text data simultaneously, offering a unique advantage over traditional models. Extensive experiments on complex tabular datasets show that our approach outperforms state-of-the-art methods across binary classification, multi-class classification, and regression tasks.", "keywords": [ "Large Language Model", "Tabular Data Prediction", "Multimodal Learning" ], "rating_list": [ 3, 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2, 2 ], "contribution_list": [ 2, 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "XWK2o2cJ3W", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Hengrui Zhang", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "", "position": "PhD student" }, { "name": "Philip S. Yu", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "US", "position": "Full Professor" }, { "name": "Qitian Wu", "gender": "Male", "institution": "Broad Institute of MIT and Harvard", "country": "US", "position": "Researcher" } ], "n_formula": 34, "n_formula_1": 25, "n_ref_uni": 32, "n_ref": 76, "n_ref_all": 90, "n_fig": 4, "n_tab": 5, "L_tab": 2595, "n_element_tab": 324, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1958, "n_element_tab_1": 239, "formula_len_all": 3262, "formula_len_all_1": 2024, "len_all": 128167, "len_all_1": 70994, "len_abs": 1094, "len_title": 113, "len_sents": 38944, "len_sents_1": 29744, "n_sents": 314, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1097, "title": "Data-Centric Graph Condensation via Diffusion Trajectory Matching", "abs": "This paper introduces Data Centric Graph Condensation (named DCGC), a data-centric and model-agnostic method for condensing a large graph into a smaller one by matching the distribution between two graphs. DCGC defines the distribution of a graph as the trajectories of its node signals (such as node features and node labels) induced by a diffusion process over the geometric structure, which accommodates multi-order structural information. Built upon this, DCGC compresses the topological knowledge of the original graph into the orders-of-magnitude smaller synthetic one by aligning their distributions in input space. Compared with existing methods that stick to particular GNN architectures and require solving complicated optimization, DCGC can be flexibly applied for arbitrary off-the-shelf GNNs and achieve graph condensation with a much faster speed. Apart from the cross-architecture generalization ability and training efficiency, experiments demonstrate that DCGC yields consistently superior performance than existing methods on datasets with varying scales and condensation ratios.", "keywords": [ "Graph Condensation", "Distribution Matching" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "XWBE90OYlH", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Dominik Fuchsgruber", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Stephan G\u00fcnnemann", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Professor" }, { "name": "Tim Postuvan", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Intern" }, { "name": "Simon Markus Geisler", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "PhD student" } ], "n_formula": 68, "n_formula_1": 27, "n_ref_uni": 51, "n_ref": 83, "n_ref_all": 164, "n_fig": 9, "n_tab": 13, "L_tab": 5734, "n_element_tab": 403, "n_fig_1": 4, "n_tab_1": 9, "L_tab_1": 2257, "n_element_tab_1": 188, "formula_len_all": 10272, "formula_len_all_1": 2002, "len_all": 257241, "len_all_1": 74625, "len_abs": 1269, "len_title": 127, "len_sents": 77754, "len_sents_1": 30348, "n_sents": 739, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1280, "title": "Graph Neural Networks for Edge Signals: Orientation Equivariance and Invariance", "abs": "Many applications in traffic, civil engineering, or electrical engineering revolve around edge-level signals. Such signals can be categorized as inherently directed, for example, the water flow in a pipe network, and undirected, like the diameter of a pipe. Topological methods model edge signals with inherent direction by representing them relative to a so-called *orientation* assigned to each edge. \nThese approaches can neither model undirected edge signals nor distinguish if an edge itself is directed or undirected. We address these shortcomings by (i) revising the notion of *orientation equivariance* to enable edge direction-aware topological models, (ii) proposing *orientation invariance* as an additional requirement to describe signals without inherent direction, and (iii) developing EIGN, an architecture composed of novel direction-aware edge-level graph shift operators, that provably fulfills the aforementioned desiderata. It is the first general-purpose topological GNN for edge-level signals that can model directed and undirected signals while distinguishing between directed and undirected edges. A comprehensive evaluation shows that EIGN outperforms prior work in edge-level tasks, for example, improving in RMSE on flow simulation tasks by up to 43.5%.", "keywords": [ "Graph Neural Network", "Graph", "Edge", "Equivariance", "Invariance", "Topology", "Directed Graphs" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 4, 3, 2, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "XW4Xnx0xlH", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Adam D. Cobb", "gender": "Male", "institution": "SRI International", "country": "", "position": "Researcher" }, { "name": "Atilim Gunes Baydin", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Assistant Professor" }, { "name": "Barak A. Pearlmutter", "gender": "Male", "institution": "Maynooth University", "country": "IE", "position": "Full Professor" }, { "name": "Susmit Jha", "gender": "unknown", "institution": "SRI International", "country": "US", "position": "Principal Scientist" } ], "n_formula": 47, "n_formula_1": 15, "n_ref_uni": 29, "n_ref": 53, "n_ref_all": 82, "n_fig": 9, "n_tab": 2, "L_tab": 1876, "n_element_tab": 149, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1876, "n_element_tab_1": 149, "formula_len_all": 4502, "formula_len_all_1": 1667, "len_all": 134216, "len_all_1": 69935, "len_abs": 1429, "len_title": 116, "len_sents": 44494, "len_sents_1": 31029, "n_sents": 366, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1435, "title": "Second-Order Forward-Mode Automatic Differentiation for Optimization", "abs": "Forward gradient methods offer a promising alternative to backpropagation. Optimization that only requires forward passes could simplify hardware implementation, improve parallelism, lower memory cost, and allow for more biologically plausible learning models. This has motivated recent forward-mode automated differentiation (AD) methods. This paper presents a novel second-order forward-mode AD method for optimization that generalizes a second-order line search to a $K$-dimensional hyperplane. Unlike recent work that relies on directional derivatives (or Jacobian\u2013Vector Products, JVPs), we use hyper-dual numbers to jointly evaluate both directional derivatives and their second-order quadratic terms. As a result, we introduce forward-mode weight perturbation with Hessian information for K-dimensional hyper-plane search (FoMoH-$K$D). We derive the convergence properties of FoMoH-$K$D and show how it generalizes to Newton\u2019s method for $K = D$. We demonstrate this generalization empirically, and compare the performance of FoMoH-$K$D to forward gradient descent (FGD) on three case studies: Rosenbrock function used widely for evaluating optimization methods, logistic regression with 7,850 parameters, and learning a CNN classifier with 431,080 parameters. Our experiments show that FoMoH-$K$D not only achieves better performance and accuracy, but also converges faster, thus, empirically verifying our theoretical results.", "keywords": [ "Optimization", "Automatic Differentiation" ], "rating_list": [ 3, 6, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 2, 2, 5 ] }, { "paper_id": "XVHXVdoV11", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Jyothish Pari", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Pulkit Agrawal", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Samy Jelassi", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Postdoc" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 27, "n_ref": 40, "n_ref_all": 60, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 426, "formula_len_all_1": 426, "len_all": 89517, "len_all_1": 52624, "len_abs": 1520, "len_title": 64, "len_sents": 26120, "len_sents_1": 23994, "n_sents": 197, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1548, "title": "Collective Model Intelligence Requires Compatible Specialization", "abs": "In this work, we explore the limitations of combining models by averaging intermediate features, referred to as $\\textit{model merging}$, and propose a new direction for achieving collective model intelligence through what we call $\\textit{compatible specialization}$. Current methods for model merging, such as parameter and feature averaging, struggle to effectively combine specialized models due to representational divergence during fine-tuning. As models specialize to their individual domains, their internal feature representations become increasingly incompatible, leading to poor performance when attempting to merge them for new tasks. We analyze this phenomenon using centered kernel alignment (CKA) and show that as models specialize, the similarity in their feature space structure diminishes, hindering their capacity for collective use. To address these challenges, we investigate routing-based merging strategies, which offer more flexible methods for combining specialized models by dynamically routing across different layers. This allows us to improve on existing methods by combining features from multiple layers rather than relying on fixed, layer-wise combinations. However, we find that these approaches still face limitations when layers within models are representationally incompatible. Our findings highlight the importance of designing new approaches for model merging that operate on well-defined input and output spaces, similar to how humans communicate through language rather than intermediate neural activations.", "keywords": [ "model merging", "collective intelligence", "routing", "position", "analysis" ], "rating_list": [ 3, 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 5, 4 ] }, { "paper_id": "XUJcsLvpaQ", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Dennis Frauen", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Konstantin Hess", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Stefan Feuerriegel", "gender": "Male", "institution": "LMU Munich", "country": "DE", "position": "Professor" }, { "name": "Valentyn Melnychuk", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" } ], "n_formula": 63, "n_formula_1": 26, "n_ref_uni": 62, "n_ref": 221, "n_ref_all": 287, "n_fig": 8, "n_tab": 4, "L_tab": 745, "n_element_tab": 52, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 427, "n_element_tab_1": 49, "formula_len_all": 6115, "formula_len_all_1": 1806, "len_all": 218428, "len_all_1": 70629, "len_abs": 2767, "len_title": 124, "len_sents": 64473, "len_sents_1": 29161, "n_sents": 572, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 888, "title": "G-Transformer for Conditional Average Potential Outcome Estimation over Time", "abs": "Estimating potential outcomes for treatments over time based on observational data is important for personalized decision-making in medicine. Yet, existing neural methods for this task either (1) do not perform proper adjustments for time-varying confounders, or (2) suffer from large estimation variance. In order to address both limitations, we introduce the G-transformer (GT). Our GT is a novel, neural end-to-end model which adjusts for time-varying confounders, and provides low-variance estimation of conditional average potential outcomes (CAPOs) over time. Specifically, our GT is the first neural model to perform regression-based iterative G-computation for CAPOs in the time-varying setting. We evaluate the effectiveness of our GT across various experiments. In sum, this work represents a significant step towards personalized decision-making from electronic health records.", "keywords": [ "causal inference", "potential outcomes", "treatment effects", "G-computation", "time-varying confounding", "medicine" ], "rating_list": [ 5, 3, 8, 5 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 2, 4, 4, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "XTxdDEFR6D", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Feng Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Weilin Luo", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Xijun Li", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yufei Kuang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Yuyan Zhou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 19, "n_formula_1": 10, "n_ref_uni": 36, "n_ref": 102, "n_ref_all": 138, "n_fig": 6, "n_tab": 13, "L_tab": 5206, "n_element_tab": 374, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2442, "n_element_tab_1": 158, "formula_len_all": 1768, "formula_len_all_1": 630, "len_all": 202315, "len_all_1": 69996, "len_abs": 1669, "len_title": 150, "len_sents": 66748, "len_sents_1": 31920, "n_sents": 462, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1705, "title": "LLM4Solver: Large Language Model for Efficient Algorithm Design of Combinatorial Optimization Solver", "abs": "The optimization of algorithms in exact combinatorial optimization (CO) solver plays a fundamental role in operations research.\nHowever, due to the extensive requirements on domain knowledge and the large search space for algorithm design, the refinement on these algorithms remains highly challenging for both manual and learning-based paradigms. \nTo tackle this problem, we propose a novel machine learning framework---large language model for exact combinatorial optimization solver (LLM4Solver)---to $\\textit{efficiently}$ design high-quality algorithms of the CO solvers. \nThe core idea is that, instead of searching in the high-dimensional and discrete symbolic space from scratch, we can utilize the prior knowledge learned from large language models to directly search in the space of programming languages.\nSpecifically, we first use a pre-trained LLM as the generator for high-quality algorithms. Then, to efficiently explore the discrete and non-gradient algorithm space, we employ a derivative-free evolutionary framework as the algorithm optimizer.\nExperiments on extensive benchmarks show that the algorithms learned by LLM4Solver $\\textit{significantly}$ outperform all the state-of-the-art (SOTA) human-designed and learning-based policies (on GPU) in terms of the solution quality, the solving efficiency, and the cross-benchmark generalization ability. \nThe appealing features of LLM4Solver include 1) the high training efficiency to outperform SOTA methods within ten iterations, and 2) the high cross-benchmark generalization ability on heterogeneous MIPLIB 2017.\nLLM4Solver shows the encouraging potential to efficiently design algorithms for the next generation of modern CO solvers.", "keywords": [ "Combinatorial Optimization Solver", "Large Language Models", "Evolutionary Search" ], "rating_list": [ 3, 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 5, 4, 4, 3, 4 ] }, { "paper_id": "XToAemis1h", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Ao Shen", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "Undergrad student" }, { "name": "Bin Fang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "Full Professor" }, { "name": "Di Hu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Jiangyu Hu", "gender": "Male", "institution": "Wuhan University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Ruoxuan Feng", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "MS student" }, { "name": "TianciGao", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Undergrad student" }, { "name": "Wenke Xia", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Yuhao Sun", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 46, "n_ref": 107, "n_ref_all": 132, "n_fig": 8, "n_tab": 7, "L_tab": 2353, "n_element_tab": 233, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1063, "n_element_tab_1": 82, "formula_len_all": 911, "formula_len_all_1": 813, "len_all": 182666, "len_all_1": 74228, "len_abs": 1921, "len_title": 133, "len_sents": 60931, "len_sents_1": 36298, "n_sents": 430, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 2025, "title": "AnyTouch: Learning Unified Static-Dynamic Representation across Multiple Visuo-tactile Sensors", "abs": "Visuo-tactile sensors aim to emulate human tactile perception, enabling robots to precisely understand and manipulate objects. Over time, numerous meticulously designed visuo-tactile sensors have been integrated into robotic systems, aiding in completing various tasks. However, the distinct data characteristics of these low-standardized visuo-tactile sensors hinder the establishment of a powerful tactile perception system. We consider that the key to addressing this issue lies in learning unified multi-sensor representations, thereby integrating the sensors and promoting tactile knowledge transfer between them. To achieve unified representation of this nature, we introduce TacQuad, an aligned multi-modal multi-sensor tactile dataset from four different visuo-tactile sensors, which enables the explicit integration of various sensors. Recognizing that humans perceive the physical environment by acquiring diverse tactile information such as texture and pressure changes, we further propose to learn unified multi-sensor representations from both static and dynamic perspectives. By integrating tactile images and videos, we present AnyTouch, a unified static-dynamic multi-sensor representation learning framework with a multi-level structure, aimed at both enhancing comprehensive perceptual abilities and enabling effective cross-sensor transfer. This multi-level architecture captures pixel-level details from tactile data via masked modeling and enhances perception and transferability by learning semantic-level sensor-agnostic features through multi-modal alignment and cross-sensor matching. We provide a comprehensive analysis of multi-sensor transferability, and validate our method on various offline datasets and in the real-world pouring task. Experimental results show that our method outperforms existing methods, exhibits outstanding static and dynamic perception capabilities across various sensors. The code, TacQuad dataset and AnyTouch model are fully available at gewu-lab.github.io/AnyTouch/.", "keywords": [ "Tactile Representation Learning", "Visuo-tactile Sensors", "Cross-sensor Transferring" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 4, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "XTXUHQqLbg", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Alex M. Bronstein", "gender": "Male", "institution": "Computer Science Department, Technion - Israel Institute of Technology", "country": "IL", "position": "Full Professor" }, { "name": "Barak Gahtan", "gender": "Male", "institution": "Computer Science Department, Technion - Israel Institute of Technology", "country": "IL", "position": "PhD student" }, { "name": "Reuven Cohen", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Robert J. Shahla", "gender": "unknown", "institution": "Computer Science Department, Technion - Israel Institute of Technology", "country": "", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 19, "n_ref": 36, "n_ref_all": 51, "n_fig": 6, "n_tab": 2, "L_tab": 1034, "n_element_tab": 98, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 186, "formula_len_all_1": 35, "len_all": 111174, "len_all_1": 61753, "len_abs": 1261, "len_title": 77, "len_sents": 46335, "len_sents_1": 32692, "n_sents": 341, "n_sents_1": 230, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1260, "title": "Exploring QUIC Dynamics: A Large-Scale Dataset for Encrypted Traffic Analysis", "abs": "QUIC, a new and increasingly used transport protocol, addresses and resolves the limitations of TCP by offering improved security, performance, and features such as stream multiplexing and connection migration. These features, however, also present challenges for network operators who need to monitor and analyze web traffic. In this paper, we introduce VisQUIC, a labeled dataset comprising over 100,000 QUIC traces from more than 44,000 websites (URLs), collected over a four-month period. These traces provide the foundation for generating more than two million images, with configurable parameters of window length, pixel resolution, normalization, and labels. These images enable an observer looking at the interactions between a client and a server to analyze and gain insights about QUIC encrypted connections. To illustrate the dataset's potential, we offer a use-case example of an observer estimating the number of HTTP/3 responses/requests pairs in a given QUIC, which can reveal server behavior, client--server interactions, and the load imposed by an observed connection. We formulate the problem as a discrete regression problem, train a machine learning (ML) model for it, and then evaluate it using the proposed dataset on an example use case.", "keywords": [ "Deep learning", "image dataset derived from real-world QUIC traces", "discrete regression problem", "vision for communication" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 1, 3, 1 ], "contribution_list": [ 1, 1, 2, 4 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "XTBdPLhiRL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Feidiao Yang", "gender": "Male", "institution": "Pengcheng Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Jinyi Lu", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Kai Huang", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Mingyue Cui", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Postdoc" }, { "name": "Quanmin Liang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Wei Zhang", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yonghong Tian", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Li Qiang", "gender": "Male", "institution": " Guangzhou Xiaopeng Motors Technology Co Ltd", "country": "CN", "position": "Researcher" }, { "name": "XINZI CAO", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 48, "n_ref": 87, "n_ref_all": 111, "n_fig": 6, "n_tab": 9, "L_tab": 3031, "n_element_tab": 422, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 549, "n_element_tab_1": 21, "formula_len_all": 273, "formula_len_all_1": 273, "len_all": 168407, "len_all_1": 62832, "len_abs": 1668, "len_title": 124, "len_sents": 35293, "len_sents_1": 30626, "n_sents": 256, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1677, "title": "Enhancing Event Camera Data Pretraining via Prompt-Tuning with Visual Models", "abs": "The pretraining-finetuning paradigm has achieved remarkable success in natural language processing and computer vision, becoming the dominant approach in many downstream tasks. However, its application in the event camera domain has encountered significant challenges. First, the scarcity and sparsity of large-scale event datasets lead to issues like overfitting during extensive pretraining. Second, event data inherently contains both temporal and spatial information, making it difficult to directly transfer knowledge from image-based pretraining to event camera tasks.\nIn this paper, we propose a low-parameter-cost SpatioTemporal Information Fusion Prompting (STP) method to address these challenges. This method enables bidirectional fusion of event and image data while mitigating the risk of overfitting. Specifically, the key innovation lies in effectively integrating the spatio-temporal information of event data to align with pre-trained image models and reduce the impact of data sparsity.\nTo achieve this, we designed an Overlap Patch Embedding module within the STP, which employs wide receptive field to capture more local information and reduce the influence of sparse regions. Additionally, we introduce a Temporal Transformer that integrates both global and local information, facilitating the fusion of temporal and spatial data. Our approach significantly outperforms previous state-of-the-art methods across multiple downstream tasks, including classification, semantic segmentation, and optical flow estimation. For instance, it achieves a top-1 accuracy of 68.83% on N-ImageNet with fewer trainable parameters. Our code is available in the Supplement.", "keywords": [ "Event camera", "Pretraining", "Prompt-tuning" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "XT7kCxcEKm", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Biwei Huang", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Ignavier Ng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Kaifeng Jin", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" } ], "n_formula": 17, "n_formula_1": 12, "n_ref_uni": 50, "n_ref": 114, "n_ref_all": 127, "n_fig": 0, "n_tab": 12, "L_tab": 7844, "n_element_tab": 657, "n_fig_1": 0, "n_tab_1": 7, "L_tab_1": 2496, "n_element_tab_1": 219, "formula_len_all": 926, "formula_len_all_1": 727, "len_all": 176157, "len_all_1": 73691, "len_abs": 1504, "len_title": 134, "len_sents": 55226, "len_sents_1": 32718, "n_sents": 365, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1521, "title": "Revisiting Differentiable Structure Learning: Inconsistency of $\\ell_1$ Penalty and Beyond", "abs": "Recent advances in differentiable structure learning have framed the combinatorial problem of learning directed acyclic graphs as a continuous optimization problem. Various aspects, including data standardization, have been studied to identify factors that influence the empirical performance of these methods. In this work, we investigate critical limitations in differentiable structure learning methods, focusing on settings where the true structure can be identified up to Markov equivalence classes, particularly in the linear Gaussian case. While Ng et al. (2024) highlighted potential non-convexity issues in this setting, we demonstrate and explain why the use of $\\ell_1$-penalized likelihood in such cases is fundamentally inconsistent, even if the global optimum of the optimization problem can be found. To resolve this limitation, we develop a hybrid differentiable structure learning method based on $\\ell_0$-penalized likelihood with hard acyclicity constraint, where the $\\ell_0$ penalty can be approximated by different techniques including Gumbel-Softmax. Specifically, we first estimate the underlying moral graph, and use it to restrict the search space of the optimization problem, which helps alleviate the non-convexity issue. Experimental results show that the proposed method enhances empirical performance both before and after data standardization, providing a more reliable path for future advancements in differentiable structure learning, especially for learning Markov equivalence classes.", "keywords": [ "differentiable structure learning", "$\\ell_0$-penalized likelihood", "acyclicity constraint", "moral graph" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "XT1Cx6cH2a", "primary_area": "foundation or frontier models, including LLMs", "n_author": 12, "author_info_list": [ { "name": "Chuanyang Zheng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Han Shi", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Principal Researcher" }, { "name": "Jiankai Sun", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Jingyao Li", "gender": "Female", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Michael Ng", "gender": "unknown", "institution": "Hong Kong Baptist University", "country": "CN", "position": "Full Professor" }, { "name": "Minbin Huang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xiaozhe Ren", "gender": "Male", "institution": "Noah's Ark Lab", "country": "", "position": "Researcher" }, { "name": "Xin Jiang", "gender": "Male", "institution": "Noah\u2019s Ark Lab, Huawei Technologies", "country": "HK", "position": "Principal Researcher" }, { "name": "Yihang Gao", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Yu Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Zhenguo Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "jing xiong", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 39, "n_ref": 104, "n_ref_all": 119, "n_fig": 3, "n_tab": 10, "L_tab": 4395, "n_element_tab": 351, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 851, "n_element_tab_1": 10, "formula_len_all": 625, "formula_len_all_1": 433, "len_all": 232810, "len_all_1": 61098, "len_abs": 1309, "len_title": 120, "len_sents": 41806, "len_sents_1": 30419, "n_sents": 272, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1320, "title": "DAPE V2: Process Attention Score as Feature Map for Length Extrapolation", "abs": "The attention mechanism is a fundamental component of the Transformer model, contributing to interactions among distinct tokens. In general, the attention scores are determined simply by the key-query products. However, this work's occasional trial (combining DAPE and NoPE) of including additional MLPs on attention scores without position encoding indicates that the classical key-query multiplication may limit the performance of Transformers. \nIn this work, we conceptualize attention as a feature map and apply the convolution operator (for neighboring attention scores across different heads) to mimic the processing methods in computer vision. Specifically, **the main contribution of this paper is identifying and interpreting the Transformer length extrapolation problem as a result of the limited expressiveness of the naive query and key dot product, and we successfully translate the length extrapolation issue into a well-understood feature map processing problem.** \nThe novel insight, which can be adapted to various attention-related models, reveals that the current Transformer architecture has the potential for further evolution. Extensive experiments demonstrate that treating attention as a feature map and applying convolution as a processing method significantly enhances Transformer performance.", "keywords": [ "Transformers", "data-adaptive positional encoding", "long context", "length generalization" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 1, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "XSVq2z1CU6", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Dekai Zhu", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Slobodan Ilic", "gender": "Male", "institution": "Siemens Technology", "country": "DE", "position": "Principal Researcher" }, { "name": "Stefan Gavranovic", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Yan Di", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 33, "n_ref": 86, "n_ref_all": 105, "n_fig": 9, "n_tab": 4, "L_tab": 929, "n_element_tab": 79, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 31, "n_element_tab_1": 2, "formula_len_all": 943, "formula_len_all_1": 853, "len_all": 112296, "len_all_1": 60406, "len_abs": 1863, "len_title": 124, "len_sents": 35262, "len_sents_1": 28645, "n_sents": 256, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1870, "title": "SeaLion: Semantic Part-Aware Latent Point Diffusion Models for 3D Generation", "abs": "Denoising diffusion probabilistic models have achieved significant success in point cloud generation, enabling numerous downstream applications, such as generative data augmentation and 3D model editing. However, little attention has been given to generating point clouds with point-wise segmentation labels, as well as to developing evaluation metrics for this task. Therefore, in this paper, we present SeaLion, a novel diffusion model designed to generate high-quality and diverse point cloud with fine-grained segmentation labels. Specifically, we introduce the semantic part-aware latent point diffusion technique, which leverages the intermediate features of the generative models to jointly predict the noise for perturbed latent points and associated part segmentation labels during the denoising process, and subsequently decodes the latent points to point clouds conditioned on part segmentation labels. To effectively evaluate the quality of generated point clouds, we introduce a novel point cloud pairwise distance calculation method named part-aware Chamfer distance (p-CD). This method enables existing metrics, such as 1-NNA, to measure both the local structural quality and inter-part coherence of generated point clouds. Experiments on the large-scale synthetic dataset ShapeNet and real-world medical dataset IntrA, demonstrate that SeaLion achieves remarkable performance in generation quality and diversity, outperforming the existing state-of-the-art model, DiffFacto, by 13.33% and 6.52% on 1-NNA (p-CD) across the two datasets. Experimental analysis shows that SeaLion can be trained semi-supervised, thereby reducing the demand for labeling efforts. Lastly, we validate the applicability of SeaLion in generative data augmentation for training segmentation models and the capability of SeaLion to serve as a tool for part-aware 3D shape editing.", "keywords": [ "Diffusion Model", "3D Computer Vision" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "XRtyVELwr6", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Manuel Cherep", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Nikhil Singh", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 86, "n_ref": 134, "n_ref_all": 150, "n_fig": 9, "n_tab": 1, "L_tab": 932, "n_element_tab": 76, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1237, "n_element_tab_1": 120, "formula_len_all": 615, "formula_len_all_1": 544, "len_all": 166116, "len_all_1": 61949, "len_abs": 401, "len_title": 104, "len_sents": 35030, "len_sents_1": 28773, "n_sents": 239, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1422, "title": "Contrastive Learning from Synthetic Audio Doppelg\u00e4ngers", "abs": "Learning robust audio representations currently demands extensive datasets of real-world sound recordings. By applying artificial transformations to these recordings, models can learn to recognize similarities despite subtle variations through techniques like contrastive learning. However, these transformations are only approximations of the true diversity found in real-world sounds, which are generated by complex interactions of physical processes, from vocal cord vibrations to the resonance of musical instruments. We propose a solution to both the data scale and transformation limitations, leveraging synthetic audio. By randomly perturbing the parameters of a sound synthesizer, we generate audio doppelg\u00e4ngers\u2014synthetic positive pairs with causally manipulated variations in timbre, pitch, and temporal envelopes. These variations, difficult to achieve through augmentations of existing audio, provide a rich source of contrastive information. Despite the shift to randomly generated synthetic data, our method produces strong representations, outperforming real data on several standard audio classification tasks. Notably, our approach is lightweight, requires no data storage, and has only a single hyperparameter, which we extensively analyze. We offer this method as a complement to existing strategies for contrastive learning in audio, using synthesized sounds to reduce the data burden on practitioners.", "keywords": [ "synthetic data", "audio", "contrastive learning", "representation learning" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "XQlccqJpCC", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Jiashuo Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Mihaela van der Schaar", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Nabeel Seedat", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Peng Cui", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 36, "n_formula_1": 22, "n_ref_uni": 44, "n_ref": 64, "n_ref_all": 100, "n_fig": 13, "n_tab": 6, "L_tab": 1874, "n_element_tab": 82, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 654, "n_element_tab_1": 63, "formula_len_all": 2362, "formula_len_all_1": 1403, "len_all": 198037, "len_all_1": 74370, "len_abs": 1763, "len_title": 118, "len_sents": 69709, "len_sents_1": 36349, "n_sents": 495, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 953, "title": "Going Beyond Static: Understanding Shifts with Time-Series Attribution", "abs": "Distribution shifts in time-series data are complex due to temporal dependencies, multivariable interactions, and trend changes. \nHowever, robust methods often rely on structural assumptions that lack thorough empirical validation, limiting their practical applicability. \nIn order to support an empirically grounded inductive approach to research, we introduce our Time-Series Shift Attribution (TSSA) framework, which analyzes problem-specific patterns of distribution shifts. Our framework attributes performance degradation from various types of shifts to each *temporal data property* in a detailed manner, supported by theoretical analysis of unbiasedness and asymptotic properties. Empirical studies in real-world healthcare applications highlight how the TSSA framework enhances the understanding of time-series shifts, facilitating reliable model deployment and driving targeted improvements from both algorithmic and data-centric perspectives.", "keywords": [ "distribution shifts", "performance drop", "attribution", "time-series data" ], "rating_list": [ 6, 8, 8, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 2, 2, 5 ] }, { "paper_id": "XQgbmhQozV", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Han Jiang", "gender": "unknown", "institution": "Tongji University", "country": "CN", "position": "MS student" }, { "name": "Shu Wang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Xiaoyuan Yi", "gender": "Male", "institution": "Research, Microsoft", "country": "CN", "position": "Senior Researcher" }, { "name": "Xin Xie", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Zhihua Wei", "gender": "Female", "institution": "Tongji University", "country": "CN", "position": "Full Professor" } ], "n_formula": 44, "n_formula_1": 16, "n_ref_uni": 90, "n_ref": 167, "n_ref_all": 238, "n_fig": 11, "n_tab": 15, "L_tab": 7228, "n_element_tab": 305, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4296, "formula_len_all_1": 2201, "len_all": 433328, "len_all_1": 69697, "len_abs": 1808, "len_title": 150, "len_sents": 93487, "len_sents_1": 31114, "n_sents": 761, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1527, "title": "Raising the Bar: Investigating the Values of Large Language Models via Generative Evolving Testing", "abs": "Warning: this paper contains model outputs exhibiting unethical information.\n\nLarge Language Models (LLMs) have achieved significant breakthroughs, but their generated unethical content poses potential risks. Measuring value alignment of LLMs becomes crucial for their regulation and responsible deployment. Numerous datasets have been constructed to assess social bias, toxicity, and ethics in LLMs, but they suffer from evaluation chronoeffects, that is, as models rapidly evolve, existing data becomes leaked or undemanding, overestimating ever-developing LLMs. To tackle this problem, we propose GETA, a novel generative evolving testing approach that dynamically probes the underlying moral baselines of LLMs. Distinct from previous adaptive testing methods that rely on static datasets with limited difficulty, GETA incorporates an iteratively-updated item generator which infers each LLM's moral boundaries and generates difficulty-tailored testing items, accurately reflecting the true alignment extent. This process theoretically learns a joint distribution of item and model response, with item difficulty and value conformity as latent variables, where the generator co-evolves with the LLM, mitigating chronoeffects. We evaluate various popular LLMs with diverse capabilities and demonstrate that GETA can create difficulty-matching testing items and more accurately assess LLMs' values, better consistent with their performance on unseen OOD and i.i.d. items, laying the groundwork for future evaluation paradigms.", "keywords": [ "Large Language Models", "Human values", "Human ethics", "Evaluation", "Adaptive testing" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 4 ], "confidence_list": [ 2, 2, 3, 2 ] }, { "paper_id": "XQQMKhGBoY", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Chaoyu Guan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Hong Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Siao Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Wenwu Zhu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xin Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yansong Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuan Meng", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 15, "n_ref_uni": 25, "n_ref": 68, "n_ref_all": 78, "n_fig": 8, "n_tab": 4, "L_tab": 1213, "n_element_tab": 173, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1656, "n_element_tab_1": 163, "formula_len_all": 969, "formula_len_all_1": 994, "len_all": 95604, "len_all_1": 54935, "len_abs": 1440, "len_title": 145, "len_sents": 25998, "len_sents_1": 23086, "n_sents": 231, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1447, "title": "Accelerate Quantization Aware Training for Diffusion Models with Difficulty-aware Time Allocation", "abs": "Diffusion models have demonstrated remarkable power in various generation tasks. Nevertheless, the large computational cost during inference is a troublesome issue for diffusion models, especially for large pretrained models such as Stable Diffusion. Quantization-aware training (QAT) is an effective method to reduce both memory and time costs for diffusion models while maintaining good performance. However, QAT methods usually suffer from the high cost of retraining the large pretrained model, which restricts the efficient deployment of diffusion models. To alleviate this problem, we propose a framework DFastQ (Diffusion Fast QAT) to accelerate the training of QAT from a difficulty-aware perspective in the timestep dimension. Specifically, we first propose to adaptively identify the difficulties of different timesteps according to the oscillation of their training loss curves. Then we propose a difficulty-aware time allocation module, which aims to dynamically allocate more training time to difficult timesteps to speed up the convergence of QAT. The key component of this is a timestep drop mechanism consisting of a drop probability predictor and a pair of adversarial losses. We conduct a series of experiments on different Stable Diffusion models, quantization settings, and sampling strategies, demonstrating that our method can effectively accelerate QAT by at least 24\\% while achieving comparable or even better performance.", "keywords": [ "Diffusion Models", "Model Quantization", "Text-to-image Generation" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "XQL4Pmf6m6", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Bo Zhao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Fan BAI", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Max q.-h. Meng", "gender": "Male", "institution": "The Chinese University of Hong Kong, The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Tiejun Huang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Yuxin Du", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 55, "n_ref": 101, "n_ref_all": 141, "n_fig": 18, "n_tab": 11, "L_tab": 2868, "n_element_tab": 325, "n_fig_1": 4, "n_tab_1": 11, "L_tab_1": 2716, "n_element_tab_1": 244, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 211199, "len_all_1": 56361, "len_abs": 1440, "len_title": 127, "len_sents": 44395, "len_sents_1": 23802, "n_sents": 421, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1449, "title": "M3D: Advancing 3D Medical Image Analysis with Multi-Modal Large Language Models", "abs": "Medical image analysis is essential to numerous practicals of clinical diagnosis and treatment. However, due to the data scarcity and expensive training cost, previous research has largely focused on 2D medical image analysis, leaving 3D medical images under-explored, despite their important spatial information. This paper aims to advance 3D medical image analysis by leveraging multi-modal large language models (MLLMs). \nWe propose M3D-LaMed, a generalist MLLM for 3D medical image analysis, specializing in eight important tasks, including image-text retrieval, report generation, visual question answering, positioning, segmentation, etc. The spatial pooling perceiver is proposed to reduce the 3D tokens, while preserving spatial information.\nTo train the model, we construct the largest 3D multi-modal medical dataset, M3D-Data, comprising 120K image-text pairs and 662K instruction-response pairs specifically tailored for 3D medical tasks.\nThe 3D multi-modal benchmark, M3D-Bench, is designed, which facilitates the comprehensive evaluation of models across eight tasks. The extensive experiments demonstrate that, as a generalist model, M3D-LaMed shows promising performances and outperforms other specialist models in multiple tasks. With the proposed model, data and benchmark, this work establishes a universal framework that significantly advances the 3D medical image analysis. All data, code and models will be publicly accessible.", "keywords": [ "Medical image analysis", "3D medical imaging", "MLLM" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "XQFSIdKMhJ", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Qadeer Ahmed", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "", "position": "Assistant Professor" }, { "name": "Qazi Mairaj ud din", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 7, "n_ref": 10, "n_ref_all": 16, "n_fig": 5, "n_tab": 3, "L_tab": 138, "n_element_tab": 12, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 138, "n_element_tab_1": 12, "formula_len_all": 312, "formula_len_all_1": 312, "len_all": 68460, "len_all_1": 51824, "len_abs": 2284, "len_title": 91, "len_sents": 26771, "len_sents_1": 26779, "n_sents": 165, "n_sents_1": 165, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 2293, "title": "Using Generative AI to capture High Fidelity Temporal Dynamics to target Vehicular Systems", "abs": "Generative models have transformed the creation of text, images, and video content by enabling machines to generate high-quality, realistic outputs. These models are now widely being adopted in advanced fields like natural language processing, computer vision, and media production. Since vehicle data is limited due to proprietary concerns, utilizing generative models to mimic complex vehicle behaviors would provide powerful tools for creating synthetic data that can serve as a crucial component for enhancing the fidelity of vehicle models, better predictive maintenance, more robust control systems, autonomous driving features and resilient defense mechanism against cyber threats. This paper presents a Long Short-Term Memory (LSTM) based Conditional Generative Adversarial\nNetwork (GAN) model, which trains on limited available real vehicle data and is then able to generate synthetic time series data mimicking the actual vehicle data. The LSTM network helps in learning temporal characteristics of vehicle network traffic without needing the system details, which makes it applicable to wide range of vehicle networks. The conditional layer adds auxiliary information by labeling\ndata for different driving scenarios for training and generating data. The quality of the synthetic data is evaluated visually and quantitatively using metrics such as Maximum Mean Discrepancy (MMD), Predictive and Discriminative Scores. For demonstration purposes, the generative model is integrated into a validated vehicle model, where it successfully generates synthetic sensor feedback corresponding to the dynamic driving scenarios. This showcases the model\u2019s ability to simulate realistic sensor data in response to varying vehicle operations. Leveraging the high similarity to actual data, the generative model is further demonstrated for its potential use as malicious attack mechanism due to its deception capabilities against state of the art Intrusion Detection System (IDS). Without triggering the thresholds of the IDS, the model is able to penetrate the network stealthily with a low detection rate of 47.05%, compared to the 90% or higher detection rates of other known attacks. This effort is intended to serve as a test benchmark to develop more robust ML/AI based defense mechanisms.", "keywords": [ "Generative model", "Cybersecurity", "Vehicle System", "Synthetic Data" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "XQED8Nk9mu", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Mathieu Salzmann", "gender": "Male", "institution": "Swiss Data Science Center", "country": "CH", "position": "Principal Researcher" }, { "name": "Yann Bouquet", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Zhuoqian Yang", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 35, "n_ref": 82, "n_ref_all": 103, "n_fig": 12, "n_tab": 5, "L_tab": 2846, "n_element_tab": 142, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 1822, "n_element_tab_1": 93, "formula_len_all": 693, "formula_len_all_1": 694, "len_all": 126776, "len_all_1": 60909, "len_abs": 833, "len_title": 106, "len_sents": 32043, "len_sents_1": 27877, "n_sents": 237, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 833, "title": "Counterfactual Explanations for 3D Point-Cloud Classifiers", "abs": "Explainable AI (XAI) seeks to tackle the opacity of deep neural network decisions. Moving beyond the conventional focus on 2D imagery, our research provides the first method to provide Counterfactual Explanations (CEs) for 3D point cloud classifiers. Specifically, we introduce two strategies for 3D CEs using a diffusion model to generate CEs that maintain both semantic consistency and data fidelity in 3D contexts. To this end, we devise novel losses and constraints to boost the realism and practicality of counterfactual instances. Furthermore, we establish a new benchmark with evaluation metrics designed specifically for 3D point clouds allowing future methods to be assessed using it. Altogether, our contributions bridge a key gap in the field of explainability, steering towards more transparent and fair AI methodologies.", "keywords": [ "explainability", "counterfactual", "3D", "point-cloud" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 5, 2, 3 ] }, { "paper_id": "XPNprvlxuQ", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Jaihoon Kim", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Kyeongmin Yeo", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Undergrad student" }, { "name": "Minhyuk Sung", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 14, "n_ref_uni": 41, "n_ref": 144, "n_ref_all": 176, "n_fig": 16, "n_tab": 4, "L_tab": 810, "n_element_tab": 188, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 675, "n_element_tab_1": 102, "formula_len_all": 1672, "formula_len_all_1": 1617, "len_all": 170087, "len_all_1": 72240, "len_abs": 359, "len_title": 137, "len_sents": 42058, "len_sents_1": 33982, "n_sents": 322, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1314, "title": "StochSync: Stochastic Diffusion Synchronization for Image Generation in Arbitrary Spaces", "abs": "We propose a zero-shot method for generating images in arbitrary spaces (e.g., a sphere for 360\u25e6 panoramas and a mesh surface for texture) using a pretrained image diffusion model. The zero-shot generation of various visual content using a pretrained image diffusion model has been explored mainly in two directions. First, Diffusion Synchronization\u2013performing reverse diffusion processes jointly across different projected spaces while synchronizing them in the target space\u2013generates high-quality outputs when enough conditioning is provided, but it struggles in its absence. Second, Score Distillation Sampling\u2013gradually updating the target space data through gradient descent\u2013results in better coherence but often lacks detail. In this paper, we reveal for the first time the interconnection between these two methods while highlighting their differences. To this end, we propose StochSync, a novel approach that combines the strengths of both, enabling effective performance with weak conditioning. Our experiments demonstrate that StochSync provides the best performance in 360\u25e6 panorama generation (where image conditioning is not given), outperforming previous finetuning-based methods, and also delivers comparable results in 3D mesh texturing (where depth conditioning is provided) with previous methods.", "keywords": [ "Diffusion Models", "Synchronization", "Score Distillation", "Panorama", "Texturing" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 2, 5 ] }, { "paper_id": "XNr6sexQGj", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "ChangHao Li", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jianfei Chen", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Jun Zhu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Kang Zhao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xinrui Chen", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 43, "n_ref": 76, "n_ref_all": 96, "n_fig": 7, "n_tab": 8, "L_tab": 1579, "n_element_tab": 319, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 562, "n_element_tab_1": 122, "formula_len_all": 1182, "formula_len_all_1": 1118, "len_all": 136679, "len_all_1": 56002, "len_abs": 1794, "len_title": 91, "len_sents": 38596, "len_sents_1": 26490, "n_sents": 259, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1802, "title": "Zero-shot Quantization for Object Detection", "abs": "Zero-shot quantization (ZSQ) has achieved remarkable success in classification tasks by leveraging synthetic data for network quantization without accessing the original training data. However, when applied to object detection networks, current ZSQ methods fail due to the inherent complexity of the task, which encompasses both localization and classification challenges. On the one hand, the precise location and size of objects within the samples for object detection remain unknown and elusive in zero-shot scenarios, precluding artificial reconstruction without ground-truth information. On the other hand, object detection datasets typically exhibit category imbalance, and random category sampling methods designed for classification tasks cannot capture this information.\nTo tackle these challenges, we propose a novel ZSQ framework specifically tailored for object detection. The proposed framework comprises two key steps: First, we employ a novel bounding box and category sampling strategy in the calibration set generation process to infer the original training data from a pre-trained detection network and reconstruct the location, size and category distribution of objects within the data without any prior knowledge. Second, we incorporate feature-level alignment into the Quantization Aware Training (QAT) process, further amplifying its efficacy through the integration of feature-level distillation.\nExtensive experiments conducted on the MS-COCO and Pascal VOC datasets demonstrate the efficiency and state-of-the-art performance of our method in low-bit-width quantization. For instance, when quantizing YOLOv5-m to 5-bit, we achieve a 4.2\\% improvement in the mAP metric, utilizing only about 1/60 of the calibration data required by commonly used LSQ trained with full trainset.", "keywords": [ "Zero-shot quantization", "Object detection", "Synthetic data", "Fine-tuning efficiency", "Feature distillation" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "XNA3Mnnbvb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Gen Li", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Kaifeng Zhao", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Siyu Tang", "gender": "Female", "institution": "Department of Computer Science, Swiss Federal Institute of Technology", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 26, "n_formula_1": 11, "n_ref_uni": 64, "n_ref": 167, "n_ref_all": 173, "n_fig": 4, "n_tab": 8, "L_tab": 4243, "n_element_tab": 267, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1226, "n_element_tab_1": 83, "formula_len_all": 1327, "formula_len_all_1": 1045, "len_all": 216351, "len_all_1": 70597, "len_abs": 1745, "len_title": 141, "len_sents": 61637, "len_sents_1": 34262, "n_sents": 401, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 92, "L_abs": 1774, "title": "DART: A Diffusion-Based Autoregressive Motion Model for Real-Time Text-Driven Motion Control", "abs": "Text-conditioned human motion generation, which allows for user interaction through natural language, has become increasingly popular. Existing methods typically generate short, isolated motions based on a single input sentence. However, human motions are continuous and can extend over long periods, carrying rich semantics. Creating long, complex motions that precisely respond to streams of text descriptions, particularly in an online and real-time setting, remains a significant challenge. Furthermore, incorporating spatial constraints into text-conditioned motion generation presents additional challenges, as it requires aligning the motion semantics specified by text descriptions with geometric information, such as goal locations and 3D scene geometry. To address these limitations, we propose **DART**, a **D**iffusion-based **A**utoregressive motion primitive model for **R**eal-time **T**ext-driven motion control. Our model, DART, effectively learns a compact motion primitive space jointly conditioned on motion history and text inputs using latent diffusion models. By autoregressively generating motion primitives based on the preceding history and current text input, DART enables real-time, sequential motion generation driven by natural language descriptions. Additionally, the learned motion primitive space allows for precise spatial motion control, which we formulate either as a latent noise optimization problem or as a Markov decision process addressed through reinforcement learning. We present effective algorithms for both approaches, demonstrating our model\u2019s versatility and superior performance in various motion synthesis tasks. Experiments show our method outperforms existing baselines in motion realism, efficiency, and controllability.", "keywords": [ "Human Motion Generation" ], "rating_list": [ 6, 8, 3, 8, 8 ], "soundness_list": [ 3, 4, 2, 3, 4 ], "presentation_list": [ 4, 4, 3, 3, 3 ], "contribution_list": [ 3, 4, 2, 3, 4 ], "confidence_list": [ 4, 5, 5, 3, 4 ] }, { "paper_id": "XMlj8W8o0Y", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Han Qiu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Jiaxing Huang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Shijian Lu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "JINGYI ZHANG", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 68, "n_ref_all": 91, "n_fig": 5, "n_tab": 20, "L_tab": 12020, "n_element_tab": 1731, "n_fig_1": 4, "n_tab_1": 9, "L_tab_1": 3388, "n_element_tab_1": 518, "formula_len_all": 1238, "formula_len_all_1": 972, "len_all": 237075, "len_all_1": 63919, "len_abs": 1474, "len_title": 117, "len_sents": 45045, "len_sents_1": 23440, "n_sents": 273, "n_sents_1": 132, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1482, "title": "Hough Voting-based Self-Training for Vision-Language Model Adaptation", "abs": "Traditional model adaptation framework assumes the same vocabulary across pre-training and downstream datasets, which often struggles with limited transfer flexibility and efficiency while handling downstream datasets with different vocabularies.\nInspired by recent vision-language models (VLMs) that enable visual recognition defined by free-form texts via reasoning on both images and texts, we study vision-language model adaptation (VLMA), a new unsupervised model adaptation framework that positions a pre-trained VLM as the source model and transfers it towards various unlabelled downstream datasets.\nTo this end, we propose a Hough voting-based Self-Training (HoughST) technique that introduces a multimodal Hough voting mechanism to exploit the synergy between vision and language to mitigate the distribution shift in image and text modalities simultaneously. \nSpecifically, HoughST makes use of the complementary property of different types of features within and across vision and language modalities, which enables joint exploitation of vision and language information and effective learning of image-text correspondences in the unlabelled downstream datasets. \nAdditionally, HoughST captures temporal information via temporal Hough voting which helps memorize and leverage previously learnt downstream dataset information.\nExtensive experiments show that HoughST outperforms the state-of-the-art consistently across 11 image recognition tasks. \nCodes will be released.", "keywords": [ "Transfer Learning" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "XMgpnZ2ET7", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Biwei Huang", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Xinyue Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" } ], "n_formula": 41, "n_formula_1": 13, "n_ref_uni": 50, "n_ref": 109, "n_ref_all": 124, "n_fig": 9, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5023, "formula_len_all_1": 1675, "len_all": 184505, "len_all_1": 61320, "len_abs": 1919, "len_title": 153, "len_sents": 60213, "len_sents_1": 29746, "n_sents": 418, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1329, "title": "Modeling Unseen Environments with Language-guided Composable Causal Components in Reinforcement Learning", "abs": "Generalization in reinforcement learning (RL) remains a significant challenge, especially when agents encounter novel environments with unseen dynamics. Drawing inspiration from human compositional reasoning\u2014where known components are reconfigured to handle new situations\u2014we introduce World Modeling with Compositional Causal Components (WM3C). This novel framework enhances RL generalization by learning and leveraging compositional causal components. Unlike previous approaches focusing on invariant representation learning or meta-learning, WM3C identifies and utilizes causal dynamics among composable elements, facilitating robust adaptation to new tasks. Our approach integrates language as a compositional modality to decompose the latent space into meaningful components and provides theoretical guarantees for their unique identification under mild assumptions. Our practical implementation uses a masked autoencoder with mutual information constraints and adaptive sparsity regularization to capture high-level semantic information and effectively disentangle transition dynamics. Experiments on numerical simulations and real-world robotic manipulation tasks demonstrate that WM3C significantly outperforms existing methods in identifying latent processes, improving policy learning, and generalizing to unseen tasks.", "keywords": [ "Model-based Reinforcement Learning", "World Model", "Causal Representation Learning", "Generalization in Reinforcement Learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "XMOaOigOQo", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Liang Qifan", "gender": "Male", "institution": "Jilin University", "country": "", "position": "PhD student" }, { "name": "Liang Yin", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Ting Long", "gender": "Female", "institution": "Jilin University", "country": "CN", "position": "Assistant Professor" }, { "name": "Weinan Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yi Chang", "gender": "Male", "institution": "Jilin University, China", "country": "CN", "position": "Full Professor" }, { "name": "Yixiang Shan", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "PhD student" }, { "name": "Zhengbang Zhu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 18, "n_ref_uni": 45, "n_ref": 89, "n_ref_all": 159, "n_fig": 16, "n_tab": 7, "L_tab": 7261, "n_element_tab": 862, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3090, "n_element_tab_1": 341, "formula_len_all": 1488, "formula_len_all_1": 1031, "len_all": 179729, "len_all_1": 70679, "len_abs": 1428, "len_title": 120, "len_sents": 50905, "len_sents_1": 31250, "n_sents": 367, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1441, "title": "ContraDiff: Planning Towards High Return States via Contrastive Learning", "abs": "The performance of offline reinforcement learning (RL) is sensitive to the proportion of high-return trajectories in the offline dataset. However, in many simulation environments and real-world scenarios, there are large ratios of low-return trajectories rather than high-return trajectories, which makes learning an efficient policy challenging. In this paper, we propose a method called Contrastive Diffuser (ContraDiff) to make full use of low-return trajectories and improve the performance of offline RL algorithms. Specifically, ContraDiff groups the states of trajectories in the offline dataset into high-return states and low-return states and treats them as positive and negative samples correspondingly. Then, it designs a contrastive mechanism to pull the planned trajectory of an agent toward high-return states and push them away from low-return states. Through the contrast mechanism, trajectories with low returns can serve as negative examples for policy learning, guiding the agent to avoid areas associated with low returns and achieve better performance. Through the contrast mechanism, trajectories with low returns provide a ``counteracting force'' guides the agent to avoid areas associated with low returns and achieve better performance.\nExperiments on 27 sub-optimal datasets demonstrate the effectiveness of our proposed method. Our code is publicly available at \\url{https://anonymous.4open.science/r/ContraDiff}.", "keywords": [ "Offline Reinforcement Learning", "Decision Making", "Diffusion Models", "Machine Learning" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "XLt0eudh8t", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Muhan Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiyuan Wang", "gender": "unknown", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yanbo Wang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "XiaohuiZhang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 45, "n_ref": 81, "n_ref_all": 98, "n_fig": 4, "n_tab": 13, "L_tab": 4969, "n_element_tab": 512, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 710, "n_element_tab_1": 66, "formula_len_all": 1096, "formula_len_all_1": 885, "len_all": 175433, "len_all_1": 70639, "len_abs": 1586, "len_title": 68, "len_sents": 51350, "len_sents_1": 33473, "n_sents": 444, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1623, "title": "Efficient Neural Common Neighbor for Temporal Graph Link Prediction", "abs": "Temporal graphs are ubiquitous in real-world scenarios, such as social network, trade and transportation. Predicting dynamic links between nodes in a temporal graph is of vital importance. \\textcolor{blue}{\nTraditional memory-based methods typically leverage the temporal neighborhood of interaction histories to generate node embeddings, which are then aggregated to predict links between source and target nodes. However, these methods primarily focus on learning individual node representations and often neglect the nature of pairwise representation learning aspect. While some recent methods attempt to capture pairwise features, they are less emphasized in large-scale datasets like TGB. Meanwhile, most of these \nexisting methods tend to suffer from high computational complexity due to the repeated calculation of node embeddings.\n}\nMotivated by the success of Neural Common Neighbor (NCN) for static graph link prediction, we propose \\textbf{TNCN}, a temporal version of NCN for link prediction in temporal graphs. Based on a memory-based backbone instead of traditional static graph neural network, TNCN dynamically updates a temporal neighbor dictionary for each node, and utilizes multi-hop common neighbors between the source and target node to learn a more effective pairwise representation. We validate our model on five large-scale real-world datasets from the Temporal Graph Benchmark (TGB), and find that it achieves new state-of-the-art performance on three of them. Additionally, TNCN demonstrates excellent scalability on large datasets, outperforming popular GNN baselines by up to 6.4 times in speed.", "keywords": [ "Temporal graph", "Neural common neighbor", "Efficient" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "XLcu8vHRpZ", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Hao-Xiang Guo", "gender": "Male", "institution": "SkyWork AI", "country": "CN", "position": "Researcher" }, { "name": "Hao He", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Luozhou Wang", "gender": "Male", "institution": "Hong Kong University of Science and Technology (GUANGZHOU)", "country": "CN", "position": "PhD student" }, { "name": "Xiang Wen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Xinli Xu", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Ying-Cong Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Yixun Liang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Cai Yuanhao", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 102, "n_ref_all": 118, "n_fig": 12, "n_tab": 3, "L_tab": 541, "n_element_tab": 71, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1, "n_element_tab_1": 1, "formula_len_all": 383, "formula_len_all_1": 339, "len_all": 134457, "len_all_1": 50600, "len_abs": 782, "len_title": 114, "len_sents": 32367, "len_sents_1": 23264, "n_sents": 244, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 886, "title": "LucidFusion: Generating 3D Gaussians with Arbitrary Unposed Images", "abs": "Recent large reconstruction models have made notable progress in generating high-quality 3D objects from single images. However, these methods often struggle with controllability, as they lack information from multiple views, leading to incomplete or inconsistent 3D reconstructions. To address this limitation, we introduce LucidFusion, a flexible end-to-end feed-forward framework that leverages the Relative Coordinate Map (RCM). Unlike traditional methods linking images to 3D world thorough pose, LucidFusion utilizes RCM to align geometric features coherently across different views, making it highly adaptable for 3D generation from arbitrary, unposed images. Furthermore, LucidFusion seamlessly integrates with the original single-image-to-3D pipeline, producing detailed 3D Gaussians at a resolution of $512 \\times 512$, making it well-suited for a wide range of applications.", "keywords": [ "3D vision", "3D Gaussian Splatting" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "XLMAMmowdY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Haonan Li", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Lei Ji", "gender": "Female", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Renxi Wang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "MS student" }, { "name": "Shu Wang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Timothy Baldwin", "gender": "unknown", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" }, { "name": "Xudong Han", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 68, "n_ref_all": 99, "n_fig": 10, "n_tab": 12, "L_tab": 3307, "n_element_tab": 463, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 247, "formula_len_all_1": 247, "len_all": 204801, "len_all_1": 61352, "len_abs": 1436, "len_title": 106, "len_sents": 57054, "len_sents_1": 29052, "n_sents": 483, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1442, "title": "ToolGen: Unified Tool Retrieval and Calling via Generation", "abs": "As large language models (LLMs) advance, their inability to autonomously execute tasks by directly interacting with external tools remains a critical limitation. Traditional methods rely on inputting tool descriptions as context, which is constrained by context length and requires separate, often inefficient, retrieval mechanisms. We introduce ToolGen, a paradigm shift that integrates tool knowledge directly into the LLM\u2019s parameters by representing each tool as a unique token. This enables the LLM to generate tool calls and arguments as part of its next token prediction capabilities, seamlessly blending tool invocation with language generation. Our framework allows the LLM to access and utilize a vast amount of tools with no additional retrieval step, significantly enhancing both performance and scalability. Experimental results with over 47,000 tools show that ToolGen not only achieves superior results in both tool retrieval and autonomous task completion but also sets the stage for a new era of AI agents that can adapt to tools across diverse domains. By fundamentally transforming tool retrieval into a generative process, ToolGen paves the way for more versatile, efficient, and autonomous AI systems. ToolGen enables end-to-end tool learning and opens opportunities for integration with other advanced techniques such as chain-of-thought and reinforcement learning, thereby expanding the practical capabilities of LLMs", "keywords": [ "Agent", "Tool Learning", "Virtual Token" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 3, 1, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "XLDaepymR5", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 9, "author_info_list": [ { "name": "Geza Kurczveil", "gender": "Male", "institution": "Hewlett Packard Enterprise", "country": "", "position": "Researcher" }, { "name": "Raymond G Beausoleil", "gender": "unknown", "institution": "Hewlett Packard Labs", "country": "", "position": "Principal Researcher" }, { "name": "Sijia Liu", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Assistant Professor" }, { "name": "Xian Xiao", "gender": "Male", "institution": "Hewlett Packard Enterprise", "country": "", "position": "Researcher" }, { "name": "Xinling Yu", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Yequan Zhao", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Zheng Zhang", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "Associate Professor" }, { "name": "Zhixiong Chen", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "", "position": "MS student" }, { "name": "Ziyue Liu", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "", "position": "PhD student" } ], "n_formula": 46, "n_formula_1": 26, "n_ref_uni": 56, "n_ref": 116, "n_ref_all": 145, "n_fig": 6, "n_tab": 8, "L_tab": 1947, "n_element_tab": 281, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 314, "n_element_tab_1": 23, "formula_len_all": 4395, "formula_len_all_1": 2531, "len_all": 184419, "len_all_1": 66244, "len_abs": 1402, "len_title": 131, "len_sents": 47492, "len_sents_1": 28307, "n_sents": 409, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1410, "title": "Scalable Back-Propagation-Free Training of Optical Physics-Informed Neural Networks", "abs": "Physics-informed neural networks (PINNs) have shown promise in solving partial differential equations (PDEs), with growing interest in their energy-efficient, real-time training on edge devices. Photonic computing offers a potential solution due to its high operation speed.\nHowever, the lack of photonic memory and the large footprint of current photonic devices prevent training realistic-size PINNs on photonic chips. This paper proposes a completely back-propagation-free (BP-free) and highly salable framework to enable training real-size PINNs on silicon photonics platforms. Our approach involves three key innovations: (1) a sparse-grid Stein derivative estimator to avoid the BP in the loss evaluation of a PINN, (2) a dimension-reduced zeroth-order optimization via tensor-train decomposition to achieve better scalability and convergence in BP-free training, and (3) a scalable on-chip photonic PINN training accelerator design using photonic tensor cores. We validate the performance of our numerical methods in both low- and high-dimensional PDE benchmarks. Through circuit simulation based on real device parameters, we further demonstrate the significant performance benefit (e.g., real-time training, huge chip area reduction) of our photonic accelerator. Our framework addresses the fundamental challenges of photonic AI and will enable real-time training of real-size PINNs on photonic chips.", "keywords": [ "Optical Neural Networks", "Physics-Informed Neural Networks", "On-Chip Learning", "Scalability", "Hardware-Software Co-Design" ], "rating_list": [ 8, 6, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 4, 2 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 4, 5 ] }, { "paper_id": "XLCqhdaMpy", "primary_area": "applications to robotics, autonomy, planning", "n_author": 3, "author_info_list": [ { "name": "Gaurav S. Sukhatme", "gender": "Male", "institution": "Amazon", "country": "", "position": "Amazon Scholar" }, { "name": "Gautam Salhotra", "gender": "unknown", "institution": "Intrinsic LLC", "country": "US", "position": "Researcher" }, { "name": "Shashank Hegde", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 5, "n_ref_uni": 23, "n_ref": 39, "n_ref_all": 64, "n_fig": 14, "n_tab": 5, "L_tab": 1762, "n_element_tab": 290, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 189, "n_element_tab_1": 58, "formula_len_all": 1508, "formula_len_all_1": 468, "len_all": 173258, "len_all_1": 50620, "len_abs": 1755, "len_title": 110, "len_sents": 32998, "len_sents_1": 23634, "n_sents": 269, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1767, "title": "Latent Weight Diffusion: Generating policies from trajectories", "abs": "With the increasing availability of open-source robotic data, imitation learning has emerged as a viable approach for both robot manipulation and locomotion. Currently, large generalized policies are trained to predict controls or trajectories using diffusion models, which have the desirable property of learning multimodal action distributions. However, generalizability comes with a cost \u2014 namely, larger model size and slower inference. Further, there is a known trade-off between performance and action horizon for Diffusion Policy (i.e., diffusing trajectories):\nfewer diffusion queries accumulate greater trajectory tracking errors. Thus, it is common practice to run these models at high inference frequency, subject to robot computational constraints.\n\nTo address these limitations, we propose Latent Weight Diffusion (LWD), a method that uses diffusion to learn a distribution over policies for robotic tasks, rather than over trajectories. Our approach encodes demonstration trajectories into a latent space and then decodes them into policies using a hypernetwork. We employ a diffusion denoising model within this latent space to learn its distribution. We demonstrate that LWD can reconstruct the behaviors of the original policies that generated the trajectory dataset. LWD offers the benefits of considerably smaller policy networks during inference and requires fewer diffusion model queries. When tested on the Metaworld MT10 benchmark, LWD achieves a higher success rate compared to a vanilla multi-task policy, while using models up to \u223c18x smaller during inference. Additionally, since LWD generates closed-loop policies, we show that it outperforms Diffusion Policy in long action horizon settings, with reduced diffusion queries during rollout.", "keywords": [ "Diffusion methods", "long horizon robotics tasks", "Imitation Learning" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "XKv29sMyjF", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Marco Canini", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Associate Professor" }, { "name": "Mohamed Elhoseiny", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Norah Alballa", "gender": "unknown", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Wenxuan Zhang", "gender": "Female", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Ziquan Liu", "gender": "Male", "institution": "Queen Mary, University of London", "country": "GB", "position": "Lecturer" }, { "name": "Ahmed Sayed", "gender": "Male", "institution": "Queen Mary, University of London", "country": "GB", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 59, "n_ref_all": 82, "n_fig": 5, "n_tab": 15, "L_tab": 6059, "n_element_tab": 861, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 3013, "n_element_tab_1": 304, "formula_len_all": 422, "formula_len_all_1": 472, "len_all": 157096, "len_all_1": 65071, "len_abs": 5171, "len_title": 119, "len_sents": 51068, "len_sents_1": 29920, "n_sents": 330, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1194, "title": "Query-based Knowledge Transfer for Heterogeneous Learning Environments", "abs": "Decentralized collaborative learning under data heterogeneity and privacy constraints has rapidly advanced. However, existing solutions like federated learning, ensembles, and transfer learning, often fail to adequately serve the unique needs of clients, especially when local data representation is limited. \nTo address this issue, we propose a novel framework called Query-based Knowledge Transfer (QKT) that enables tailored knowledge acquisition to fulfill specific client needs without direct data exchange. \nIt employs a data-free masking strategy to facilitate the communication-efficient query-focused knowledge transformation while refining task-specific parameters to mitigate knowledge interference and forgetting. Our experiments, conducted on both standard and clinical benchmarks, show that QKT significantly outperforms existing collaborative learning methods by an average of 20.91% points in single-class query settings and an average of 14.32% points in multi-class query scenarios.\nFurther analysis and ablation studies reveal that QKT effectively balances the learning of new and existing knowledge, showing strong potential for its application in decentralized learning.", "keywords": [ "Collaborative Learning", "Knowledge Distillation", "Query-based Knowledge Transfer." ], "rating_list": [ 6, 6, 5, 6, 6 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 3, 2 ] }, { "paper_id": "XKQ2qzajbU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Jie Zhou", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jiwen Lu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Wenzhao Zheng", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Wang Chengkun", "gender": "Male", "institution": "Tsinghua university", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 41, "n_ref": 69, "n_ref_all": 81, "n_fig": 3, "n_tab": 7, "L_tab": 2469, "n_element_tab": 273, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1258, "n_element_tab_1": 154, "formula_len_all": 589, "formula_len_all_1": 587, "len_all": 121363, "len_all_1": 60976, "len_abs": 1319, "len_title": 104, "len_sents": 29087, "len_sents_1": 25462, "n_sents": 196, "n_sents_1": 161, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1323, "title": "GlobalMamba: Global Image Serialization for Vision Mamba", "abs": "Vision mambas have demonstrated strong performance with linear complexity to the number of vision tokens. Their efficiency results from processing image tokens sequentially. However, most existing methods employ patch-based image tokenization and then flatten them into 1D sequences for causal processing, which ignore the intrinsic 2D structural correlations of images. It is also difficult to extract global information by sequential processing of local patches. In this paper, we propose a global image serialization method to transform the image into a sequence of causal tokens, which contain global information of the 2D image. We first convert the image from the spatial domain to the frequency domain using Discrete Cosine Transform (DCT) and then arrange the pixels with corresponding frequency ranges. We further transform each set within the same frequency band back to the spatial domain to obtain a series of images before tokenization. We construct a vision mamba model with causal input format, GlobalMamba, based on the proposed global image serialization, which can better exploit the causal relations among image sequences. Extensive experiments demonstrate the effectiveness of our GlobalMamba, including image classification on ImageNet-1K, object detection on COCO, and semantic segmentation on ADE20K.", "keywords": [ "GlobalMamba", "Representation Learning", "Frequency Domain" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4, 5 ] }, { "paper_id": "XK5jYtLMXl", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Owen Oertell", "gender": "Not Specified", "institution": "Cornell University", "country": "US", "position": "Undergrad student" }, { "name": "Wen Sun", "gender": "unknown", "institution": "Cornell University", "country": "", "position": "Assistant Professor" }, { "name": "Yiding Chen", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Postdoc" }, { "name": "Yiyi Zhang", "gender": "Female", "institution": "Cornell University", "country": "US", "position": "Research Scientist" } ], "n_formula": 81, "n_formula_1": 30, "n_ref_uni": 19, "n_ref": 63, "n_ref_all": 83, "n_fig": 4, "n_tab": 1, "L_tab": 44, "n_element_tab": 2, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 166, "n_element_tab_1": 10, "formula_len_all": 8141, "formula_len_all_1": 2423, "len_all": 135751, "len_all_1": 70306, "len_abs": 1153, "len_title": 136, "len_sents": 41914, "len_sents_1": 28108, "n_sents": 398, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1157, "title": "Convergence Of Consistency Model With Multistep Sampling Under General Data Assumptions", "abs": "Diffusion models accomplish remarkable success in data generation tasks across various domains. However, the iterative sampling process is computationally expensive. Consistency models are proposed to learn consistency functions to map from noise to data directly, which allows one-step fast data generation and multistep sampling to improve sample quality. In this paper, we study the convergence of consistency models when the self-consistency property holds approximately under the training distribution. Our analysis requires only mild data assumption and applies to a family of forward processes. When the target data distribution has bounded support or has tails that decay sufficiently fast, we show that the samples generated by the consistency model are close to the target distribution in Wasserstein distance; when the target distribution satisfies some smoothness assumption, we show that with an additional perturbation step for smoothing, the generated samples are close to the target distribution in total variation distance. We provide two case studies with commonly chosen forward processes to demonstrate the benefit of multistep sampling.", "keywords": [ "Consistency models", "diffusion models", "learning theory" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "XIcR6JTe9D", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Volkan Cevher", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" }, { "name": "Yongtao Wu", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Zhenyu Zhu", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" } ], "n_formula": 52, "n_formula_1": 17, "n_ref_uni": 38, "n_ref": 82, "n_ref_all": 112, "n_fig": 5, "n_tab": 8, "L_tab": 2112, "n_element_tab": 267, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 669, "n_element_tab_1": 79, "formula_len_all": 2580, "formula_len_all_1": 500, "len_all": 167105, "len_all_1": 64580, "len_abs": 1170, "len_title": 145, "len_sents": 51431, "len_sents_1": 30690, "n_sents": 405, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1206, "title": "Imbalance-Regularized LoRA: A Plug-and-Play Method for Improving Fine-Tuning of Foundation Models", "abs": "Low-Rank Adaptation (LoRA) is an effective fine-tuning algorithm for large models, enabling efficient adaptation with fewer trainable parameters. Despite its success, there remains significant potential for improving LoRA's performance. In this paper, we introduce iLoRA (Imbalance-Regularized LoRA), which enhances LoRA by incorporating a regularization term to address the imbalance in forward propagation. This regularization maintains an imbalance between matrices $\\mathbf{A}$ and $\\mathbf{B}$, ensuring stable activation variance independent of dimension. Specifically, we first analyze forward dynamics, observe this imbalance in stable training, and introduce imbalanced regularization. Further, by combining this with preconditioning techniques (Zhang and Pilanci, 2024), we propose $\\pi$LoRA (Preconditioned iLoRA), which improves the backpropagation process. Our method is a plug-and-play algorithm that requires only minor modifications to the existing code and incurs negligible additional computational overhead. Finally, experiments on large language models and text-to-image models demonstrate that iLoRA and $\\pi$LoRA significantly outperform existing LoRA and preconditioned LoRA methods.", "keywords": [ "Low-Rank Adaptation", "Fine-Tuning" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 1, 2, 4 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 4, 4, 2 ] }, { "paper_id": "XIFnghzusY", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Haiyang Yu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Han Jiang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Jiwen Lu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Lening Wang", "gender": "unknown", "institution": "Beihang University", "country": "CN", "position": "MS student" }, { "name": "Wenzhao Zheng", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Yilong Ren", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Associate Professor" }, { "name": "Zhiyong Cui", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 62, "n_ref_all": 97, "n_fig": 16, "n_tab": 9, "L_tab": 3385, "n_element_tab": 409, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 396, "n_element_tab_1": 68, "formula_len_all": 461, "formula_len_all_1": 389, "len_all": 145920, "len_all_1": 63607, "len_abs": 1810, "len_title": 130, "len_sents": 40061, "len_sents_1": 30091, "n_sents": 295, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1346, "title": "OccSora: 4D Occupancy Generation Models as World Simulators for Autonomous Driving", "abs": "Understanding the evolution of 3D scenes is important for effective autonomous driving. While conventional methods model the scene development with the motion of individual instances, world models emerge as a generative framework to describe the general scene dynamics. However, most existing methods adopt an autoregressive framework to perform next-token prediction, which suffer from inefficiency to model long-term temporal evolutions.To address this, we propose a diffusion-based 4D occupancy generation model, OccSora, to simulate the development of the 3D world for autonomous driving. We employ a 4D scene tokenizer to obtain compact discrete spatial-temporal representations for 4D occupancy input and achieve high-quality reconstruction for long-sequence occupancy videos. We then learn a diffusion transformer on the spatial-temporal representations and generate 4D occupancy conditioned on a trajectory prompt. We conduct extensive experiments on the widely used nuScenes dataset with Occ3D occupancy annotations. OccSora can generate 16s videos with authentic 3D layout and temporal consistency, demonstrating its ability to understand the spatial and temporal distributions of driving scenes. With trajectory-aware 4D generation, OccSora has the potential to serve as a world simulator for the decision-making of autonomous driving.", "keywords": [ "Autonomous driving", "world model" ], "rating_list": [ 5, 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 1, 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 5, 2, 5, 4, 4 ] }, { "paper_id": "XHvguNJRbE", "primary_area": "reinforcement learning", "n_author": 1, "author_info_list": [ { "name": "Qin Yang", "gender": "Male", "institution": "Bradley University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 19, "n_ref": 33, "n_ref_all": 46, "n_fig": 7, "n_tab": 3, "L_tab": 586, "n_element_tab": 36, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 804, "n_element_tab_1": 34, "formula_len_all": 1763, "formula_len_all_1": 1512, "len_all": 77749, "len_all_1": 43738, "len_abs": 1376, "len_title": 91, "len_sents": 25210, "len_sents_1": 18727, "n_sents": 166, "n_sents_1": 127, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 1380, "title": "Innate-Values-driven Reinforcement Learning", "abs": "Innate values describe agents' intrinsic motivations, which reflect their inherent interests and preferences for pursuing goals and drive them to develop diverse skills that satisfy their various needs. Traditional reinforcement learning (RL) is learning from interaction based on the environment's feedback rewards. However, in real scenarios, the rewards are generated by agents' innate value systems, which differ vastly from individuals based on their needs and requirements. In other words, considering the AI agent as a self-organizing system, developing its awareness through balancing internal and external utilities based on its needs in different tasks is a crucial problem for individuals learning to support others and integrate community with safety and harmony in the long term. To address this gap, we propose a new RL model termed innate-values-driven RL (IVRL) based on combined motivations' models and expected utility theory to mimic its complex behaviors in the evolution through decision-making and learning. Then, we introduce two IVRL-based models: IV-DQN and IV-A2C. By comparing them with benchmark algorithms such as DQN, DDQN, A2C, and PPO in the Role-Playing Game (RPG) reinforcement learning test platform VIZDoom, we demonstrated that the IVRL-based models can help the agent rationally organize various needs, achieve better performance effectively.", "keywords": [ "innate values", "reinforcement learning", "rationality" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 3, 1, 2, 2 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "XHvdM04T0l", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Kunbo Zhang", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Mengmeng Cui", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Wei Wang", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "", "position": "Research Scientist" }, { "name": "Zhenan Sun", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 37, "n_ref": 97, "n_ref_all": 109, "n_fig": 6, "n_tab": 6, "L_tab": 6614, "n_element_tab": 749, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 5057, "n_element_tab_1": 589, "formula_len_all": 591, "formula_len_all_1": 590, "len_all": 127852, "len_all_1": 75979, "len_abs": 1718, "len_title": 142, "len_sents": 28749, "len_sents_1": 26095, "n_sents": 200, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1738, "title": "Learning Structure-Dynamics-aware Representations for Efficient and Robust 3D Pose Estimation", "abs": "Recent works in 2D-to-3D pose uplifting for monocular 3D Human Pose Estimation (HPE) have shown significant progress. However, two key challenges persist in real-world applications: vulnerability to joint noise and high computational costs. These issues arise from the dense joint-frame connections and iterative correlations typically employed by mainstream GNN-based and Transformer-based methods. To address these challenges, we propose a novel approach that leverages human physical structure and long-range dynamics to learn spatial part- and temporal frameset-based representations. This method is inherently robust to missing or erroneous joints while also reducing model parameters. Specifically, in the Spatial Encoding stage, coarse-grained body parts are used to construct structural correlations with a fully adaptive graph topology. This spatial correlation representation is integrated with muti-granularity pose attributes to generate a comprehensive pose representation for each frame. In Temporal Encoding and Decoding stages, Skipped Self-Attention is performed in framesets to establish long-term temporal dependencies from multiple perspectives of movement. On this basis, a compact Graph and Skipped Transformer (G-SFormer) is proposed, which realises efficient and robust 3D HEP in both experimental and practical scenarios. Extensive experiments on Human3.6M, MPI-INF-3DHP and Human-Eva benchmarks demonstrate that G-SFormer series models can compete and outperform the state-of-the-arts but takes only a fraction of parameters and around 1\\% computational cost. It also exhibits outstanding robustness to inaccurately detected 2D poses. The source code will be available at https://sites.google.com/view/g-sformer.", "keywords": [ "3D pose estimation", "Part-based Adaptive GNN", "Frameset-based Skipped Transformer", "efficient and robust" ], "rating_list": [ 6, 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 1, 3 ], "confidence_list": [ 4, 3, 5, 4, 3 ] }, { "paper_id": "XHTirKsQV6", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Enrique Jim\u00e9nez", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "Full Professor" }, { "name": "Morris Alper", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Rachel Mikulinsky", "gender": "Female", "institution": "Tel Aviv University", "country": "IL", "position": "MS student" }, { "name": "Shai Gordin", "gender": "Male", "institution": "Ariel University", "country": "", "position": "Assistant Professor" }, { "name": "Yoram Cohen", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Full Professor" }, { "name": "Hadar Averbuch-Elor", "gender": "Female", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 43, "n_ref": 64, "n_ref_all": 91, "n_fig": 11, "n_tab": 5, "L_tab": 864, "n_element_tab": 115, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 255, "n_element_tab_1": 15, "formula_len_all": 427, "formula_len_all_1": 427, "len_all": 137863, "len_all_1": 62249, "len_abs": 1596, "len_title": 50, "len_sents": 45351, "len_sents_1": 32155, "n_sents": 293, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1574, "title": "ProtoSnap: Prototype Alignment For Cuneiform Signs", "abs": "The cuneiform writing system served as the medium for transmitting knowledge\nin the ancient Near East for a period of over three thousand years. Cuneiform\nsigns have a complex internal structure which is the subject of expert paleographic\nanalysis, as variations in sign shapes bear witness to historical developments and\ntransmission of writing and culture over time. However, prior automated techniques\nmostly treat sign types as categorical and do not explicitly model their highly varied\ninternal configurations. In this work, we present an unsupervised approach for\nrecovering the fine-grained internal configuration of cuneiform signs by leveraging\npowerful generative models and the appearance and structure of prototype font\nimages as priors. Our approach, ProtoSnap, enforces structural consistency on\nmatches found with deep image features to estimate the diverse configurations\nof cuneiform characters, snapping a skeleton-based template to photographed\ncuneiform signs. We provide a new benchmark of expert annotations and evaluate\nour method on this task. Our evaluation shows that our approach succeeds in\naligning prototype skeletons to a wide variety of cuneiform signs. Moreover, we\nshow that conditioning on structures produced by our method allows for generating\nsynthetic data with correct structural configurations, significantly boosting the\nperformance of cuneiform sign recognition beyond existing techniques, in particular\nover rare signs. Our code, data, and trained models are available at the project page:\nhttps://tau-vailab.github.io/ProtoSnap/", "keywords": [ "Machine learning for social sciences", "Ancient character recognition", "generative models" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "XH3OiIhtvf", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Enoch Solomon", "gender": "unknown", "institution": "Virginia State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 23, "n_ref": 27, "n_ref_all": 37, "n_fig": 3, "n_tab": 1, "L_tab": 125, "n_element_tab": 9, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 125, "n_element_tab_1": 9, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 66326, "len_all_1": 66326, "len_abs": 797, "len_title": 130, "len_sents": 21289, "len_sents_1": 21289, "n_sents": 151, "n_sents_1": 151, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 800, "title": "Unsupervised Federated Learning for Privacy Preserving in Face Recognition System", "abs": "Recent advancements in face recognition involve training on a single computer, often containing sensitive personal information, raising privacy concerns. To address this, attention turns to federated learning for unsupervised face recognition, leveraging decentralized edge devices. Each device independently undergoes model training, transmitting results to a secure aggregator. We utilize GANs to diversify data without the need for transmission, thereby preserving privacy throughout the entire process. The aggregator integrates these diverse models into a single global model, which is then transmitted back to the edge devices for continued improvement. Experiments on CelebA datasets demonstrate that federated learning not only preserves privacy but also maintains high levels of performance.", "keywords": [ "Unsupervised Federated Learning for Face Recognition in Decentralized Environments" ], "rating_list": [ 1, 3, 3, 1 ], "soundness_list": [ 1, 1, 2, 1 ], "presentation_list": [ 1, 1, 2, 1 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "XGFfFKqy3Y", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Cheng-Han Huang", "gender": "unknown", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Ismail Alkhouri", "gender": "Male", "institution": "DARPA", "country": "US", "position": "Researcher" }, { "name": "Jimmy Dai", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Undergrad student" }, { "name": "Qing Qu", "gender": "Male", "institution": "University of Michigan", "country": "US", "position": "Assistant Professor" }, { "name": "Rongrong Wang", "gender": "unknown", "institution": "Michigan State University", "country": "US", "position": "Associate Professor" }, { "name": "Saiprasad Ravishankar", "gender": "unknown", "institution": "Michigan State University", "country": "", "position": "Assistant Professor" }, { "name": "Shijun Liang", "gender": "Male", "institution": "Michigan State University", "country": "", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 22, "n_ref_uni": 37, "n_ref": 98, "n_ref_all": 164, "n_fig": 7, "n_tab": 10, "L_tab": 7315, "n_element_tab": 490, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 499, "n_element_tab_1": 7, "formula_len_all": 2758, "formula_len_all_1": 1945, "len_all": 172196, "len_all_1": 70103, "len_abs": 1722, "len_title": 115, "len_sents": 51269, "len_sents_1": 31409, "n_sents": 412, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1728, "title": "Step-wise Triple-Consistent Diffusion Sampling for Inverse Problems", "abs": "Diffusion models (DMs) are a class of generative models that allow sampling from a distribution learned over a training set. When applied to solving inverse imaging problems (IPs), the reverse sampling steps of DMs are typically modified to approximately sample from a measurement-conditioned distribution in the image space. However, these modifications may be unsuitable for certain settings (such as in the presence of measurement noise) and non-linear tasks, as they often struggle to correct errors from earlier sampling steps and generally require a large number of optimization and/or sampling steps. To address these challenges, we state three conditions for achieving measurement-consistent diffusion trajectories. Building on these conditions, we propose a new optimization-based sampling method that not only enforces the standard data manifold measurement consistency and forward diffusion consistency, as seen in previous studies, but also incorporates backward diffusion consistency that maintains a diffusion trajectory by optimizing over the input of the pre-trained model at every sampling step. By enforcing these conditions, either implicitly or explicitly, our sampler requires significantly fewer reverse steps. Therefore, we refer to our accelerated method as **S**tep-w**i**se **T**riple-**Co**nsistent Sa**m**pling (SITCOM). Compared to existing state-of-the-art baseline methods, under different levels of measurement noise, our extensive experiments across five linear and three non-linear image restoration tasks demonstrate that SITCOM achieves competitive or superior results in terms of standard image similarity metrics while requiring a significantly reduced run-time across all considered tasks.", "keywords": [ "Diffusion Model; Inverse Problems; Image Restoration" ], "rating_list": [ 6, 3, 5, 3, 5 ], "soundness_list": [ 2, 1, 2, 2, 2 ], "presentation_list": [ 3, 3, 4, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "XFpb3T5Zc9", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Basile Lewandowski", "gender": "Male", "institution": "Universit\u00e9 de Neuch\u00e2tel", "country": "CH", "position": "PhD student" }, { "name": "Panagiotis Patrinos", "gender": "Male", "institution": "KU Leuven", "country": "", "position": "Associate Professor" }, { "name": "Peter Kairouz", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Puya Latafat", "gender": "Male", "institution": "Institute for Advanced Studies Lucca", "country": "IT", "position": "Assistant Professor" }, { "name": "Thomas Pethick", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Volkan Cevher", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" }, { "name": "Zheng Xu", "gender": "unknown", "institution": "Google", "country": "", "position": "Research Scientist" } ], "n_formula": 93, "n_formula_1": 31, "n_ref_uni": 33, "n_ref": 77, "n_ref_all": 100, "n_fig": 7, "n_tab": 1, "L_tab": 82, "n_element_tab": 4, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 241, "n_element_tab_1": 18, "formula_len_all": 11115, "formula_len_all_1": 2105, "len_all": 179590, "len_all_1": 67301, "len_abs": 943, "len_title": 147, "len_sents": 51367, "len_sents_1": 28488, "n_sents": 481, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 947, "title": "iFedDR: Auto-Tuning Local Computation with Inexact Douglas-Rachford Splitting in Federated Learning", "abs": "Federated learning usually requires specifying the amount of local computation needed a priori. In this work, we instead propose a systematic scheme to automatically adjust and potentially reduce the local computations while preserving convergence guarantees. We focus on proximal-based methods, where we demonstrate that the proximal operator can be evaluated inexactly up to a relative error, rather than relying on a predefined sequence of vanishing errors. Our proposed method, iFedDR, is based on a novel error-corrected version of inexact Douglas-Rachford splitting. It mitigates the need for hyperparameter tuning the number of client steps, by triggering refinement on-demand. We derive iFedDR as an instance of a much more general construction, which allows us to handle minimax problem, and which is interesting in its own right. Several numerical experiments are carried out demonstrating the favorable convergence properties of iFedDR.", "keywords": [ "Federated learning", "Douglas-Rachford splitting", "monotone operators", "relative error", "inexact proximal point" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "XFeiq8FMEF", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Daizong Liu", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Wei Hu", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 40, "n_ref": 131, "n_ref_all": 164, "n_fig": 14, "n_tab": 13, "L_tab": 6350, "n_element_tab": 539, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 3224, "n_element_tab_1": 269, "formula_len_all": 723, "formula_len_all_1": 719, "len_all": 168737, "len_all_1": 67902, "len_abs": 1664, "len_title": 158, "len_sents": 45058, "len_sents_1": 27917, "n_sents": 325, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 110, "L_abs": 1671, "title": "Can't See the Wood for the Trees: Can Visual Adversarial Patches Fool Hard-Label Large Vision-Language Models?", "abs": "Large vision-language models (LVLMs) have demonstrated impressive capabilities in handling multi-modal downstream tasks, gaining increasing popularity. However, recent studies show that LVLMs are susceptible to both intentional and inadvertent attacks. Existing attackers ideally optimize adversarial perturbations with backpropagated gradients from LVLMs, thus limiting their scalability in practical scenarios as real-world LVLM applications will not provide any LVLM's gradient or details. Motivated by this research gap and counter-practical phenomenon, we propose the first and novel hard-label attack method for LVLMs, named HardPatch, to generate visual adversarial patches by solely querying the model. Our method provides deeper insights into how to investigate the vulnerability of LVLMs in local visual regions and generate corresponding adversarial substitution under the practical yet challenging hard-label setting. Specifically, we first split each image into uniform patches and mask each of them to individually assess their sensitivity to the LVLM model. Then, according to the descending order of sensitive scores, we iteratively select the most vulnerable patch to initialize noise and estimate gradients with further additive random noises for optimization. In this manner, multiple patches are perturbed until the altered image satisfies the adversarial condition. Extensive LVLM models and datasets are evaluated to demonstrate the adversarial nature of the proposed HardPatch. Our empirical observations suggest that with appropriate patch substitution and optimization, HardPatch can craft effective adversarial images to attack hard-label LVLMs.", "keywords": [ "large vision-language model", "evaluation" ], "rating_list": [ 5, 5, 3, 3, 6 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "XFYUwIyTxQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Huangxing Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Jie Zhou", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jiwen Lu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Linqing Zhao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Xiuwei Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ziwei Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 30, "n_ref": 57, "n_ref_all": 83, "n_fig": 7, "n_tab": 8, "L_tab": 1661, "n_element_tab": 248, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 2083, "n_element_tab_1": 181, "formula_len_all": 800, "formula_len_all_1": 849, "len_all": 131554, "len_all_1": 70535, "len_abs": 1892, "len_title": 53, "len_sents": 37282, "len_sents_1": 30584, "n_sents": 285, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 53, "L_abs": 1869, "title": "EmbodiedSAM: Online Segment Any 3D Thing in Real Time", "abs": "Embodied tasks require the agent to fully understand 3D scenes simultaneously with its exploration, so an online, real-time, fine-grained and highly-generalized 3D perception model is desperately needed. Since high-quality 3D data is limited, directly training such a model in 3D is infeasible. Meanwhile, vision foundation models (VFM) has revolutionized the field of 2D computer vision with superior performance, which makes the use of VFM to assist embodied 3D perception a promising direction. However, most existing VFM-assisted 3D perception methods are either offline or too slow that cannot be applied in practical embodied tasks. In this paper, we aim to leverage Segment Anything Model (SAM) for real-time 3D instance segmentation in an online setting. This is a challenging problem since future frames are not available in the input streaming RGB-D video, and an instance may be observed in several frames so efficient object matching between frames is required. To address these challenges, we first propose a geometric-aware query lifting module to represent the 2D masks generated by SAM by 3D-aware queries, which is then iteratively refined by a dual-level query decoder. In this way, the 2D masks are transferred to fine-grained shapes on 3D point clouds. Benefit from the query representation for 3D masks, we can compute the similarity matrix between the 3D masks from different views by efficient matrix operation, which enables real-time inference. Experiments on ScanNet, ScanNet200, SceneNN and 3RScan show our method achieves state-of-the-art performance among online 3D perception models, even outperforming offline VFM-assisted 3D instance segmentation methods by a large margin. Our method also demonstrates great generalization ability in several zero-shot dataset transferring experiments and show great potential in data-efficient setting.", "keywords": [ "3d instance segmentation; online 3d scene segmentation" ], "rating_list": [ 8, 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 5, 4, 2, 4, 4 ] }, { "paper_id": "XFCKEgGhEK", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Haocheng Yang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Undergrad student" }, { "name": "Hongbo Wang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Associate Professor" }, { "name": "Shilong Yuan", "gender": "unknown", "institution": "nanjing university", "country": "", "position": "MS student" }, { "name": "Wei Tan", "gender": "Male", "institution": "Lancaster University", "country": "GB", "position": "Undergrad student" }, { "name": "Yuanhao Li", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Zhenghan Chen", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" } ], "n_formula": 30, "n_formula_1": 30, "n_ref_uni": 4, "n_ref": 4, "n_ref_all": 8, "n_fig": 0, "n_tab": 3, "L_tab": 1826, "n_element_tab": 22, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2297, "formula_len_all_1": 2296, "len_all": 67153, "len_all_1": 43480, "len_abs": 1924, "len_title": 151, "len_sents": 21326, "len_sents_1": 17660, "n_sents": 186, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1931, "title": "Enhancing Cross-Lingual and Cross-Domain Adaptability in Large Language Models for Software Engineering", "abs": "This paper presents a groundbreaking mathematical framework for unsupervised domain adaptation (UDA) in the context of cross-lingual and cross-domain code modeling. We introduce the Enhanced Dynamic Code Modeling (UDA-EDCM) system, which leverages advanced concepts from measure theory, differential geometry, and information geometry to address the challenges posed by the diversity of natural and programming languages. At the core of UDA-EDCM is a novel measure-theoretic formulation of domain adaptation, utilizing optimal transport theory to minimize the discrepancy between source and target domains. We develop a Riemannian manifold approach to feature space alignment, introducing a Geodesic Flow Kernel that captures the intrinsic geometry of the code representation space. The UDA-EDCM operator is analyzed through the lens of functional analysis, revealing its spectral properties and their implications for generalization. Our information-theoretic bound on domain adaptation provides insights into the fundamental limits of knowledge transfer in code modeling. We present a unified theorem that synthesizes these diverse mathematical perspectives, offering a comprehensive characterization of UDA-EDCM's performance in terms of Wasserstein distance, empirical Rademacher complexity, and Fisher information. This theoretical foundation is complemented by an innovative optimization framework based on the Fisher Information Metric, ensuring efficient convergence in the probabilistic manifold of model parameters. Extensive experiments demonstrate that UDA-EDCM significantly outperforms existing approaches in zero-shot and few-shot learning scenarios across a wide range of programming languages and coding tasks. Our work not only advances the baselines in domain adaptation for code intelligence but also establishes a rigorous mathematical basis for future research in adaptive AI systems for software engineering.", "keywords": [ "Code Generation", "Transfer learning" ], "rating_list": [ 5, 1, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 2, 2 ], "presentation_list": [ 2, 1, 1, 1, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4, 3 ] }, { "paper_id": "XCugWIuHR8", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Mert Pilanci", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Prateek Varshney", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" } ], "n_formula": 24, "n_formula_1": 19, "n_ref_uni": 45, "n_ref": 60, "n_ref_all": 71, "n_fig": 5, "n_tab": 4, "L_tab": 983, "n_element_tab": 179, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 869, "n_element_tab_1": 168, "formula_len_all": 1796, "formula_len_all_1": 1431, "len_all": 138700, "len_all_1": 70414, "len_abs": 1742, "len_title": 131, "len_sents": 39542, "len_sents_1": 33067, "n_sents": 257, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1750, "title": "Convex Distillation: Efficient Compression of Deep Networks via Convex Optimization", "abs": "Deploying large and complex deep neural networks on resource-constrained edge devices poses significant challenges due to their computational demands and the complexities of non-convex optimization. Traditional compression methods such as distillation and pruning often retain non-convexity that complicates fine-tuning in real-time on such devices. Moreover, these methods often necessitate extensive end-to-end network fine-tuning after compression to preserve model performance, which is not only time-consuming but also requires fully annotated datasets, thus potentially negating the benefits of efficient network compression. In this paper, we introduce a novel distillation technique that efficiently compresses the model via convex optimization -- eliminating intermediate non-convex activation functions and using only intermediate activations from the original model. Our approach enables distillation in a label-free data setting and achieves performance comparable to the original model without requiring any post-compression fine-tuning. We demonstrate the effectiveness of our method for image classification models on multiple standard datasets, and further show that in the data limited regime, our method can outperform standard non-convex distillation approaches. Our method promises significant advantages for deploying high-efficiency, low-footprint models on edge devices, making it a practical choice for real-world applications. We show that convex neural networks, when provided with rich feature representations from a large pre-trained non-convex model, can achieve performance comparable to their non-convex counterparts, opening up avenues for future research at the intersection of convex optimization and deep learning.", "keywords": [ "Convex Neural Networks", "Convex/Non-Convex Optimization", "Knowledge Distillation", "Model Compression", "Label-Free Training", "Classification" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 2, 3, 5 ] }, { "paper_id": "XCg9YcSKCZ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Ali Etemad", "gender": "Male", "institution": "Queen's University", "country": "CA", "position": "Associate Professor" }, { "name": "Shuvendu Roy", "gender": "Male", "institution": "Queen's University", "country": "", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 33, "n_ref": 86, "n_ref_all": 111, "n_fig": 3, "n_tab": 9, "L_tab": 3934, "n_element_tab": 462, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1465, "n_element_tab_1": 112, "formula_len_all": 601, "formula_len_all_1": 559, "len_all": 138302, "len_all_1": 70869, "len_abs": 1188, "len_title": 129, "len_sents": 43174, "len_sents_1": 33567, "n_sents": 306, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1197, "title": "Weak Supervision from Vision-Language Models to Self-Improve on Downstream Tasks", "abs": "We present SelfPrompt, a novel prompt-tuning approach for vision-language models (VLMs) in a semi-supervised learning setup. Existing methods for tuning VLMs in semi-supervised setups struggle with the efficient use of the limited label-set budget, the negative impact of the miscalibrated VLMs on pseudo-labelling, and the accumulation of noisy pseudo-labels. SelfPrompt addresses these challenges by introducing (a) a weakly-supervised sampling technique that selects a diverse and representative labelled set, (b) a cluster-guided pseudo-labelling method that improves pseudo-label accuracy, and (c) a confidence-aware semi-supervised learning module that maximizes the utilization of unlabelled data by combining supervised learning and weakly-supervised learning. We conduct extensive evaluations across 13 datasets, significantly surpassing state-of-the-art performances with average improvements of 6.23\\% in standard semi-supervised learning, 6.25\\% in our proposed active semi-supervised learning, and 4.9\\% in base-to-novel generalization, using a 2-shot setup. Furthermore, SelfPrompt shows excellent generalization in single-shot settings, achieving an average improvement of 11.78\\%.", "keywords": [ "Semi-supervised Learning", "Vision-language Model" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "XCUTFbC3Rh", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Lei Yuan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Researcher" }, { "name": "Lihe Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Yi-Chen Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Yuchen Xiao", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Undergrad student" }, { "name": "Ziqian Zhang", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Yang Yu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Professor" } ], "n_formula": 19, "n_formula_1": 18, "n_ref_uni": 44, "n_ref": 95, "n_ref_all": 124, "n_fig": 9, "n_tab": 13, "L_tab": 11195, "n_element_tab": 884, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2855, "n_element_tab_1": 185, "formula_len_all": 1114, "formula_len_all_1": 1050, "len_all": 185826, "len_all_1": 71537, "len_abs": 1424, "len_title": 149, "len_sents": 53269, "len_sents_1": 32607, "n_sents": 358, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 99, "L_abs": 1433, "title": "Boosting Offline Multi-Objective Reinforcement Learning via Preference Conditioned Diffusion Models", "abs": "Multi-objective reinforcement learning (MORL) addresses sequential decision-making problems with multiple objectives by learning policies optimized for diverse preferences. While traditional methods necessitate costly online interaction with the environment, recent approaches leverage static datasets containing pre-collected trajectories, making offline MORL the preferred choice for real-world applications. However, existing offline MORL techniques suffer from limited expressiveness and poor generalization on out-of-distribution (OOD) preferences. To overcome these limitations, we propose Diffusion-based Multi-Objective Reinforcement Learning (DiffMORL), a generalizable diffusion-based planning framework for MORL. Leveraging the strong expressiveness and generation capability of diffusion models, DiffMORL further boosts its generalization through offline data mixup, which mitigates the memorization phenomenon and facilitates feature learning by data augmentation. By training on the augmented data, DiffMORL is able to condition on a given preference, whether in-distribution or OOD, to plan the desired trajectory and extract the corresponding action. Experiments conducted on the D4MORL benchmark demonstrate that DiffMORL achieves state-of-the-art results across nearly all tasks. Notably, it surpasses the best baseline on most tasks, underscoring its remarkable generalization ability in offline MORL scenarios.", "keywords": [ "Diffusion", "Offline Reinforcement Learning", "Multi-Objective Reinforcement Learning" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "XCP0MOMLPo", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Lior Wolf", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Associate Professor" }, { "name": "Yoni Choukroun", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "IL", "position": "Principal Researcher" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 58, "n_ref": 108, "n_ref_all": 132, "n_fig": 15, "n_tab": 5, "L_tab": 8286, "n_element_tab": 1089, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 2586, "n_element_tab_1": 319, "formula_len_all": 944, "formula_len_all_1": 774, "len_all": 157807, "len_all_1": 60964, "len_abs": 1180, "len_title": 131, "len_sents": 36415, "len_sents_1": 26253, "n_sents": 256, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1188, "title": "Factor Graph Optimization of Error-Correcting Codes for Belief Propagation Decoding", "abs": "The design of optimal linear block codes capable of being efficiently decoded is of major concern, especially for short block lengths. \nAs near capacity-approaching codes, Low-Density Parity-Check (LDPC) codes possess several advantages over\nother families of codes, the most notable being its efficient decoding via Belief Propagation.\n While many LDPC code design methods exist, the development of efficient sparse codes that meet the constraints of modern short code lengths and accommodate new channel models remains a challenge.\nIn this work, we propose for the first time a gradient-based data-driven approach for the design of sparse codes. We develop locally optimal codes with respect to Belief Propagation decoding via the learning of the Factor graph under channel noise simulations. \nThis is performed via a novel complete graph tensor representation of the Belief Propagation algorithm, optimized over finite fields via backpropagation and coupled with an efficient line-search method. \nThe proposed approach is shown to outperform the decoding performance of existing popular codes by orders of magnitude and demonstrates the power of data-driven approaches for code design.", "keywords": [ "ECC", "Binary Programming", "Belief Propagation" ], "rating_list": [ 3, 8, 3, 5, 3 ], "soundness_list": [ 3, 3, 3, 3, 1 ], "presentation_list": [ 2, 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3, 1 ], "confidence_list": [ 5, 4, 4, 5, 4 ] }, { "paper_id": "XC0nEtnevb", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Jing Shi", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "John Philip Collomosse", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Principal Researcher" }, { "name": "Kushal Kafle", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Savya Khosla", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Simon Jenni", "gender": "unknown", "institution": "Adobe Systems", "country": "CH", "position": "Researcher" }, { "name": "Handong Zhao", "gender": "unknown", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 50, "n_ref": 96, "n_ref_all": 174, "n_fig": 10, "n_tab": 6, "L_tab": 1946, "n_element_tab": 253, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1337, "n_element_tab_1": 179, "formula_len_all": 479, "formula_len_all_1": 318, "len_all": 218071, "len_all_1": 58921, "len_abs": 1318, "len_title": 94, "len_sents": 48982, "len_sents_1": 26359, "n_sents": 405, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1328, "title": "MAGNET: Augmenting Generative Decoders with Representation Learning and Infilling Capabilities", "abs": "While originally designed for unidirectional generative modeling, decoder-only large language models (LLMs) are increasingly being adapted for bidirectional modeling. However, these unidirectional and bidirectional models are typically trained independently with distinct objectives (generation or representation learning) thereby missing the potential opportunity for one objective to enhance the other. In this work, we introduce MAGNET, an adaptation of decoder-only LLMs that enhances their capabilities in generating robust representations and infilling missing text spans, while retaining their original text generation capabilities. MAGNET employs three self-supervised training objectives and introduces an attention mechanism that combines bidirectional and causal attention, enabling unified training across all objectives. We show that LLMs adapted using MAGNET can outperform state-of-the-art text encoders on token-level and sentence-level representation learning tasks. We also demonstrate that MAGNET enhances the base LLM's ability to generate contextually appropriate text infillings by enabling it to take future context into consideration. Lastly, we show that, unlike other bidirectional language models for representation learning, the LLMs adapted using MAGNET can still perform open-ended text generation.", "keywords": [ "decoder-only LLMs", "representation learning", "text infilling", "generation", "unified model" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "XBtDrlK1Qc", "primary_area": "learning on time series and dynamical systems", "n_author": 12, "author_info_list": [ { "name": "Aanish Pradhan", "gender": "unknown", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "MS student" }, { "name": "Abhilash Neog", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "PhD student" }, { "name": "Adrienne Breef-Pilz", "gender": "Female", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "Researcher" }, { "name": "Anuj Karpatne", "gender": "unknown", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "Associate Professor" }, { "name": "Arka Daw", "gender": "Male", "institution": "Oak Ridge National Laboratory", "country": "US", "position": "Researcher" }, { "name": "Bennett J. McAfee", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "MS student" }, { "name": "Cayelan Carey", "gender": "unknown", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "Full Professor" }, { "name": "Dexter W Howard", "gender": "unknown", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "PhD student" }, { "name": "Heather L. Wander", "gender": "unknown", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "PhD student" }, { "name": "Mary E. Lofton", "gender": "unknown", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "Postdoc" }, { "name": "Paul Hanson", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "Full Professor" }, { "name": "Sepideh Fatemi Khorasgani", "gender": "Not Specified", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 47, "n_ref": 96, "n_ref_all": 125, "n_fig": 13, "n_tab": 8, "L_tab": 4613, "n_element_tab": 335, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 3479, "n_element_tab_1": 110, "formula_len_all": 552, "formula_len_all_1": 589, "len_all": 151776, "len_all_1": 65278, "len_abs": 1422, "len_title": 136, "len_sents": 43308, "len_sents_1": 30004, "n_sents": 357, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1446, "title": "Masking the Gaps: An Imputation-Free Approach to Time Series Modeling with Missing Data", "abs": "Modeling time series is important in a variety of domains, yet it is challenged by the presence of missing values in real-world time-series datasets. Traditional frameworks for modeling time-series with missing values typically involve a two-step process, where the missing values are first filled-in using some imputation technique, followed by a time-series modeling approach on the imputed time-series. However, existing two-stage approaches suffer from two major drawbacks: first, the propagation of imputation errors into subsequent time-series modeling performance, and second, the inherent trade-offs between imputation efficacy and imputation complexity. To this end, we propose a novel imputation-free approach for handling missing values in time series termed {Miss}ing Feature-aware {T}ime {S}eries {M}odeling ({MissTSM}) with two main innovations. {First}, we develop a novel embedding scheme that treats every combination of time-step and feature (or channel) as a distinct token, encoding them into a high-dimensional space. {Second}, we introduce a novel {Missing Feature-Aware Attention (MFAA) Layer} to learn latent representations at every time-step based on partially observed features. We evaluate the effectiveness of MissTSM in handling missing values over multiple benchmark datasets using two synthetic masking techniques: masking completely at random (MCAR) and periodic masking, and a real-world missing-value dataset.", "keywords": [ "Time-Series", "Deep Learning", "Masked Autoencoders", "Missing Data", "Data Imputation" ], "rating_list": [ 8, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 4, 1, 3, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "XBHoaHlGQM", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Xinshang Wang", "gender": "unknown", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Zeping Min", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 95, "n_formula_1": 10, "n_ref_uni": 40, "n_ref": 87, "n_ref_all": 112, "n_fig": 15, "n_tab": 2, "L_tab": 549, "n_element_tab": 45, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 727, "n_element_tab_1": 63, "formula_len_all": 5349, "formula_len_all_1": 559, "len_all": 175508, "len_all_1": 55598, "len_abs": 836, "len_title": 130, "len_sents": 48982, "len_sents_1": 25135, "n_sents": 487, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 839, "title": "DOCS: Quantifying Weight Similarity for Deeper Insights into Large Language Models", "abs": "We introduce a novel index, the Distribution of Cosine Similarity (DOCS), for quantitatively assessing the similarity between weight matrices in Large Language Models (LLMs), aiming to facilitate the analysis of their complex architectures. Leveraging DOCS, our analysis uncovers intriguing patterns in the latest open-source LLMs: adjacent layers frequently exhibit high weight similarity and tend to form clusters, suggesting depth-wise functional specialization. Additionally, we prove that DOCS is theoretically effective in quantifying similarity for orthogonal matrices, a crucial aspect given the prevalence of orthogonal initializations in LLMs. This research contributes to a deeper understanding of LLM architecture and behavior, offering tools with potential implications for developing more efficient and interpretable models.", "keywords": [ "Weight Similarity", "Large Language Models", "Distribution of Cosine Similarity", "Cluster Analysis" ], "rating_list": [ 8, 5, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 3, 4 ], "confidence_list": [ 4, 4, 2, 3, 5 ] }, { "paper_id": "XBF63bHDZw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chengqing Zong", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Researcher" }, { "name": "Donglei Yu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Jie Zhu", "gender": "Female", "institution": "Beijing Foreign Studies University", "country": "CN", "position": "PhD student" }, { "name": "Yangyifan Xu", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yu Zhou", "gender": "Female", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Zhao Yang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" } ], "n_formula": 26, "n_formula_1": 11, "n_ref_uni": 45, "n_ref": 88, "n_ref_all": 115, "n_fig": 10, "n_tab": 9, "L_tab": 5308, "n_element_tab": 1350, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3370, "n_element_tab_1": 640, "formula_len_all": 2567, "formula_len_all_1": 1036, "len_all": 172783, "len_all_1": 75095, "len_abs": 1693, "len_title": 120, "len_sents": 40814, "len_sents_1": 28380, "n_sents": 322, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1762, "title": "SimulPL: Aligning Human Preferences in Simultaneous Machine Translation", "abs": "Simultaneous Machine Translation (SiMT) generates translations while receiving streaming source inputs. This requires the SiMT model to learn a read/write policy, deciding when to translate and when to wait for more source input. Numerous linguistic studies indicate that audiences in SiMT scenarios have distinct preferences, such as accurate translations, simpler syntax, and no unnecessary latency. Aligning SiMT models with these human preferences is crucial to improve their performances. However, this issue still remains unexplored. Additionally, preference optimization for SiMT task is also challenging. Existing methods focus solely on optimizing the generated responses, ignoring human preferences related to latency and the optimization of read/write policy during the preference optimization phase. To address these challenges, we propose Simultaneous Preference Learning (SimulPL), a preference learning framework tailored for the SiMT task. In the SimulPL framework, we categorize SiMT human preferences into five aspects: **translation quality preference**, **monotonicity preference**, **key point preference**, **simplicity preference**, and **latency preference**. By leveraging the first four preferences, we construct human preference prompts to efficiently guide GPT-4/4o in generating preference data for the SiMT task. In the preference optimization phase, SimulPL integrates **latency preference** into the optimization objective and enables SiMT models to improve the read/write policy, thereby aligning with human preferences more effectively. Experimental results indicate that SimulPL exhibits better alignment with human preferences across all latency levels in Zh$\\rightarrow$En, De$\\rightarrow$En and En$\\rightarrow$Zh SiMT tasks.", "keywords": [ "simultaneous machine translation", "simultaneous preference optimization", "human preferences" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "XAjfjizaKs", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Conor Houghton", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "Associate Professor" }, { "name": "Laurence Aitchison", "gender": "unknown", "institution": "University of Bristol", "country": "GB", "position": "Assistant Professor" }, { "name": "Lucy Farnik", "gender": "Female", "institution": "University of Bristol", "country": "GB", "position": "PhD student" }, { "name": "Timothy S. Lawson", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 12, "n_ref_uni": 51, "n_ref": 116, "n_ref_all": 189, "n_fig": 26, "n_tab": 1, "L_tab": 268, "n_element_tab": 48, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 500, "n_element_tab_1": 18, "formula_len_all": 799, "formula_len_all_1": 437, "len_all": 161894, "len_all_1": 52686, "len_abs": 1697, "len_title": 94, "len_sents": 47025, "len_sents_1": 24096, "n_sents": 331, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 46, "L_abs": 1705, "title": "Residual Stream Analysis with Multi-Layer SAEs", "abs": "Sparse autoencoders (SAEs) are a promising approach to interpreting the internal representations of transformer language models. However, SAEs are usually trained separately on each transformer layer, making it difficult to use them to study how information flows across layers. To solve this problem, we introduce the multi-layer SAE (MLSAE): a single SAE trained on the residual stream activation vectors from every transformer layer. Given that the residual stream is understood to preserve information across layers, we expected MLSAE latents to \u2018switch on\u2019 at a token position and remain active at later layers. Interestingly, we find that individual latents are often active at a single layer for a given token or prompt, but the layer at which an individual latent is active may differ for different tokens or prompts. We quantify these phenomena by defining a distribution over layers and considering its variance. We find that the variance of the distributions of latent activations over layers is about two orders of magnitude greater when aggregating over tokens compared with a single token. For larger underlying models, the degree to which latents are active at multiple layers increases, which is consistent with the fact that the residual stream activation vectors at adjacent layers become more similar. Finally, we relax the assumption that the residual stream basis is the same at every layer by applying pre-trained tuned-lens transformations, but our findings remain qualitatively similar. Our results represent a new approach to understanding how representations change as they flow through transformers. We release our code to train and analyze MLSAEs in the Supplementary Material.", "keywords": [ "sparse autoencoders", "mechanistic interpretability", "language models" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "XAO5pulJru", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Alessio Lomuscio", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Full Professor" }, { "name": "Ben Batten", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "PhD student" }, { "name": "Francesco Leofante", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "Researcher" }, { "name": "Mehran Hosseini", "gender": "unknown", "institution": "King's College London, University of London", "country": "GB", "position": "Postdoc" }, { "name": "Nicola Paoletti", "gender": "Male", "institution": "King's College London, University of London", "country": "GB", "position": "Associate Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 36, "n_ref": 70, "n_ref_all": 93, "n_fig": 2, "n_tab": 3, "L_tab": 1390, "n_element_tab": 229, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1390, "n_element_tab_1": 229, "formula_len_all": 676, "formula_len_all_1": 676, "len_all": 110268, "len_all_1": 64795, "len_abs": 1321, "len_title": 121, "len_sents": 33648, "len_sents_1": 29575, "n_sents": 248, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1328, "title": "Uncertainty-Aware Counterfactual Explanations using Bayesian Neural Nets", "abs": "A counterfactual explanation describes the smallest input change required to alter\nthe prediction of an AI model towards a desired outcome. When using neural net-\nworks, counterfactuals are obtained using variants of projected gradient descent.\nSuch counterfactuals have been shown to be brittle and implausible, potentially\njeopardising the explanatory aspects of counterfactuals. Numerous approaches\nfor obtaining better counterfactuals have been put forward. Even though these\nsolutions address some of the shortcomings, they often fall short of providing\nan all-around solution for robust and plausible counterfactuals. We hypothesise\nthis is due to the deterministic nature and limitations of neural networks, which\nfail to capture the uncertainty of the training data. Bayesian Neural Networks\n(BNNs) are a well-known class of probabilistic models that could be used to over-\ncome these issues; unfortunately, there is currently no framework for developing\ncounterfactuals for them. In this paper, we fill this gap by proposing a formal\nframework to define counterfactuals for BNNs and develop algorithmic solutions\nfor computing them. We evaluate our framework on a set of commonly used\nbenchmarks and observe that BNNs produce counterfactuals that are more robust,\nplausible, and less costly than deterministic baselines", "keywords": [ "Counterfactual Explanations", "Bayesian Neural Networks", "BNN" ], "rating_list": [ 5, 5, 1, 5 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "XAN8G0rvoB", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Hong Chen", "gender": "unknown", "institution": "Huazhong Agricultural University", "country": "", "position": "Full Professor" }, { "name": "Yingjie Wang", "gender": "Male", "institution": "China University of Petroleum", "country": "CN", "position": "Lecturer" }, { "name": "Zheng Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zirui Hu", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" } ], "n_formula": 38, "n_formula_1": 17, "n_ref_uni": 32, "n_ref": 56, "n_ref_all": 88, "n_fig": 11, "n_tab": 4, "L_tab": 1727, "n_element_tab": 142, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3374, "formula_len_all_1": 714, "len_all": 160306, "len_all_1": 62013, "len_abs": 1007, "len_title": 110, "len_sents": 42706, "len_sents_1": 27433, "n_sents": 339, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1025, "title": "A Statistical Approach for Controlled Training Data Detection", "abs": "Detecting training data for large language models (LLMs) is receiving growing attention, especially in applications requiring high reliability. While numerous efforts have been made to address this issue, they typically focus on accuracy without ensuring controllable results.\nTo fill this gap, we propose **K**nockoff Inference-based **T**raining data **D**etector (KTD), a novel method that achieves rigorous false discovery rate (FDR) control in training data detection. Specifically, KTD generates synthetic knockoff samples that seamlessly replace original data points without compromising contextual integrity. A novel knockoff statistic, which incorporates multiple knockoff draws, is then calculated to ensure FDR control while maintaining high power. \nOur theoretical analysis demonstrates KTD's asymptotic optimality in terms of FDR control and power. Empirical experiments on real-world datasets such as WikiMIA, XSum and Real Time BBC News further validate KTD's superior performance compared to existing methods.", "keywords": [ "Large language models", "Training data detection", "Knockoffs" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "XABvLUXQ45", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Debbie Lim", "gender": "Female", "institution": "University of Latvia", "country": "", "position": "Postdoc" }, { "name": "Patrick Rebentrost", "gender": "unknown", "institution": "National University of Singapore", "country": "", "position": "Researcher" }, { "name": "Qisheng Wang", "gender": "unknown", "institution": "Nagoya University", "country": "", "position": "Assistant Professor" }, { "name": "Yixian Qiu", "gender": "Female", "institution": "national university of singaore, National University of Singapore", "country": "", "position": "PhD student" } ], "n_formula": 81, "n_formula_1": 11, "n_ref_uni": 127, "n_ref": 180, "n_ref_all": 198, "n_fig": 0, "n_tab": 3, "L_tab": 354, "n_element_tab": 54, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 99, "n_element_tab_1": 7, "formula_len_all": 8364, "formula_len_all_1": 1052, "len_all": 243009, "len_all_1": 75483, "len_abs": 849, "len_title": 125, "len_sents": 46495, "len_sents_1": 31077, "n_sents": 400, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 857, "title": "Quantum Algorithm for Sparse Online Learning with Truncated Gradient Descent", "abs": "Logistic regression, the Support Vector Machine (SVM) and least squares are well-studied methods in the statistical and computer science community, with various practical applications. High-dimensional data arriving on a real-time basis makes the design of online learning algorithms that produce sparse solutions essential. The seminal work of Langford \\emph{et al.} developed a method to obtain sparsity via truncated gradient descent, showing a near-optimal online regret bound. Based on this method, we develop a quantum sparse online learning algorithm for logistic regression, the SVM and least squares. Given efficient quantum access to the inputs, we show that a quadratic speedup in the time complexity with respect to the dimension of the problem is achievable, while maintaining a regret of $O(1/\\sqrt{T})$, where $T$ is the number of time steps.", "keywords": [ "Quantum algorithms", "online learning", "truncated gradient descent", "logistic regression", "support vector machine", "least squares" ], "rating_list": [ 5, 5, 3, 5, 6 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 1, 3, 4 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 2, 2, 2, 2, 1 ] }, { "paper_id": "XA9A8mkFqa", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Marek D\u011bdi\u010d", "gender": "Male", "institution": "Czech Technical University of Prague", "country": "CZ", "position": "PhD student" }, { "name": "Pavel Prochazka", "gender": "Male", "institution": "Cisco", "country": "CZ", "position": "Researcher" } ], "n_formula": 20, "n_formula_1": 16, "n_ref_uni": 24, "n_ref": 34, "n_ref_all": 65, "n_fig": 4, "n_tab": 6, "L_tab": 2496, "n_element_tab": 337, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 533, "n_element_tab_1": 99, "formula_len_all": 982, "formula_len_all_1": 630, "len_all": 110418, "len_all_1": 62071, "len_abs": 1213, "len_title": 70, "len_sents": 34258, "len_sents_1": 27049, "n_sents": 274, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1217, "title": "CSP: An Efficient Baseline for Learning on Large-Scale Structured Data", "abs": "Last decade has seen the emergence of numerous methods for learning on graphs, particularly Graph Neural Networks (GNNs). These methods, however, are often not directly applicable to more complex structures like bipartite graphs (equivalent to hypergraphs), which represent interactions among two entity types (e.g., a user liking a movie). This paper proposes Convolutional Signal Propagation (CSP), a non-parametric simple and scalable method that natively operates on bipartite graphs (hypergraphs) and can be implemented with just a few lines of code. After defining CSP, we demonstrate its relationship with well-established methods like label propagation, Naive Bayes, and Hypergraph Convolutional Networks. We evaluate CSP against several reference methods on real-world datasets from multiple domains, focusing on retrieval and classification tasks. Our results show that CSP offers competitive performance while maintaining low computational complexity, making it an ideal first choice as a baseline for hypergraph node classification and retrieval. Moreover, despite operating on hypergraphs, CSP achieves good results in tasks typically not associated with hypergraphs, such as natural language processing.", "keywords": [ "Hypergraph representation learning", "Hypergraph convolution", "Label propagation", "Model complexity", "Naive Bayes" ], "rating_list": [ 5, 3, 1, 5 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "X9OfMNNepI", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Ben Gao", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Dongzhan Zhou", "gender": "Female", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Erik Cambria", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Soujanya Poria", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "", "position": "Assistant Professor" }, { "name": "Wanhao Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Wanli Ouyang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Full Professor" }, { "name": "Yuqiang Li", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Zonglin Yang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "tong xie", "gender": "Male", "institution": "University of New South Wales", "country": "", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 4, "n_ref_uni": 18, "n_ref": 41, "n_ref_all": 74, "n_fig": 1, "n_tab": 18, "L_tab": 1827, "n_element_tab": 180, "n_fig_1": 1, "n_tab_1": 8, "L_tab_1": 822, "n_element_tab_1": 94, "formula_len_all": 1415, "formula_len_all_1": 231, "len_all": 180974, "len_all_1": 63889, "len_abs": 1880, "len_title": 138, "len_sents": 66183, "len_sents_1": 30951, "n_sents": 471, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1736, "title": "Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses", "abs": "Scientific discovery contributes largely to the prosperity of human society, and recent progress shows that LLMs could potentially catalyst the process. However, it is still unclear whether LLMs can discover novel and valid hypotheses in chemistry. In this work, we investigate this main research question: whether LLMs can automatically discover novel and valid chemistry research hypotheses, given only a research question? With extensive discussions with chemistry experts, we adopt the assumption that a majority of chemistry hypotheses can be resulted from a research background question and several inspirations. With this key insight, we break the main question into three smaller fundamental questions. In brief, they are: (1) given a background question, whether LLMs can retrieve good inspirations; (2) with background and inspirations, whether LLMs can lead to hypothesis; and (3) whether LLMs can identify good hypotheses to rank them higher. To investigate these questions, we construct a benchmark consisting of 51 chemistry papers published in Nature or a similar level in 2024 (all papers are only available online since 2024). Every paper is divided by chemistry PhD students into three components: background, inspirations, and hypothesis. The goal is to rediscover the hypothesis given only the background and a large chemistry literature corpus consisting the ground truth inspiration papers, with LLMs trained with data up to 2023. We also develop an LLM-based multi-agent framework that leverages the assumption, consisting of three stages reflecting the more smaller questions. The proposed method can rediscover many hypotheses with very high similarity with the ground truth ones, covering the main innovations.", "keywords": [ "scientific discovery" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "X9JU2gKEkR", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Hojae Han", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Rajhans Samdani", "gender": "Male", "institution": "Spoke", "country": "", "position": "Chief Scientist" }, { "name": "Yuxiang Wei", "gender": "Male", "institution": "Meta", "country": "US", "position": "Intern" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 24, "n_ref": 117, "n_ref_all": 128, "n_fig": 2, "n_tab": 7, "L_tab": 2536, "n_element_tab": 191, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 2284, "n_element_tab_1": 187, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 240621, "len_all_1": 65231, "len_abs": 3331, "len_title": 117, "len_sents": 32655, "len_sents_1": 30388, "n_sents": 240, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1916, "title": "Arctic-SnowCoder: Demystifying High-Quality Data in Code Pretraining", "abs": "Recent studies have been increasingly demonstrating that high-quality data is crucial for effective pretraining of language models. However, the precise definition of \"high-quality\" remains underexplored. Focusing on the code domain, we introduce Arctic-SnowCoder-1.3B, a data-efficient base code model pretrained on 555B tokens through three phases of progressively refined data: (1) general pretraining with 500B standard-quality code tokens, preprocessed through basic filtering, deduplication, and decontamination, (2) continued pretraining with 50B high-quality tokens, selected from phase one by a BERT-style quality annotator trained to distinguish good code from random data, using positive examples drawn from high-quality code files, along with instruction data from Magicoder and StarCoder2-Instruct, and (3) enhanced pretraining with 5B synthetic data created by Llama-3.1-70B using phase two data as seeds, adapting the Magicoder approach for pretraining. Despite being trained on a limited dataset, Arctic-SnowCoder achieves state-of-the-art performance on BigCodeBench, a coding benchmark focusing on practical and challenging programming tasks, compared to similarly sized models trained on no more than 1T tokens, outperforming Phi-1.5-1.3B by 36%. Across all evaluated benchmarks, Arctic-SnowCoder-1.3B beats StarCoderBase-3B pretrained on 1T tokens. Additionally, it matches the performance of leading small base code models trained on trillions of tokens. For example, Arctic-SnowCoder-1.3B surpasses StarCoder2-3B, pretrained on over 3.3T tokens, on HumanEval+, a benchmark that evaluates function-level code generation, and remains competitive on BigCodeBench. Our evaluation presents a comprehensive analysis justifying various design choices for Arctic-SnowCoder. Most importantly, we find that the key to high-quality data is its consistency with the distribution of downstream applications.", "keywords": [ "Large language models", "Code generation", "Pretraining", "Data quality" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "X8dzvdkQwO", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Bingyi Jing", "gender": "Male", "institution": "South University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Hengxiang Zhang", "gender": "unknown", "institution": "Southern University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Hongxin Wei", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Songxin Zhang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 41, "n_ref": 103, "n_ref_all": 137, "n_fig": 9, "n_tab": 14, "L_tab": 4985, "n_element_tab": 513, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 2066, "n_element_tab_1": 221, "formula_len_all": 331, "formula_len_all_1": 340, "len_all": 157021, "len_all_1": 63610, "len_abs": 1347, "len_title": 119, "len_sents": 44452, "len_sents_1": 28484, "n_sents": 360, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1353, "title": "Fine-tuning can Help Detect Pretraining Data from Large Language Models", "abs": "In the era of large language models (LLMs), detecting pretraining data has been increasingly important due to concerns about fair evaluation and ethical risks. Current methods differentiate members and non-members by designing scoring functions, like Perplexity and Min-k%. However, the diversity and complexity of training data magnifies the difficulty of distinguishing, leading to suboptimal performance in detecting pretraining data. In this paper, we first explore the benefits of unseen data, which can be easily collected after the release of the LLM. We find that the perplexities of LLMs perform differently for members and non-members, after fine-tuning with a small amount of previously unseen data. In light of this, we introduce a novel and effective method termed Fine-tuned Score Deviation (FSD), which improves the performance of current scoring functions for pretraining data detection. In particular, we propose to measure the deviation distance of current scores after fine-tuning on a small amount of unseen data within the same domain. In effect, using a few unseen data can largely decrease the scores of all non-members, leading to a larger deviation distance than members. Extensive experiments demonstrate the effectiveness of our method, significantly improving the AUC score on common benchmark datasets across various models.", "keywords": [ "Large language models", "Fine-tuning", "Pretraining data detection" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 4, 3, 2, 4 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 3, 3, 5 ] }, { "paper_id": "X8aFMdXk3N", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Atsunori Minamikawa", "gender": "Male", "institution": "KDDI Corporation", "country": "JP", "position": "Researcher" }, { "name": "Guillaume Habault", "gender": "Male", "institution": "KDDI Research, inc.", "country": "JP", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 19, "n_ref": 32, "n_ref_all": 34, "n_fig": 0, "n_tab": 3, "L_tab": 1649, "n_element_tab": 57, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 1109, "n_element_tab_1": 45, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 110535, "len_all_1": 70360, "len_abs": 1236, "len_title": 155, "len_sents": 39666, "len_sents_1": 36546, "n_sents": 241, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1240, "title": "Ensuring Fair Comparisons in Time Series Forecasting: Addressing Quality Issues in Three Benchmark Datasets", "abs": "Time series forecasting (TSF) is critical in numerous applications; however, unlike other AI domains where benchmark datasets are meticulously standardized, TSF datasets often suffer from data inconsistencies, missing values, and improper temporal splits. These issues have an impact on model performance and evaluation. This paper addresses these challenges by proposing inconsistency-free versions of three well-known TSF datasets. Our methodology involves identifying and correcting data inconsistencies using a combination of linear interpolation and context-aware imputation strategies. Additionally, we introduce a novel cycle-inclusive data splitting method, which respects the longest cycle in each dataset, ensuring that models are evaluated over meaningful temporal patterns. Through extensive testing of multiple transformer-based models, we demonstrate that our revised datasets and cycle-inclusive splitting lead to more accurate and interpretable forecasting results, as well as fairer comparison of TSF models. Finally, our findings highlight the need for proper dataset refinement and tailored data splitting strategies in TSF tasks, and pave the way for future work in the development of more robust forecasting benchmarks.", "keywords": [ "Time Series; Dataset Quality; Fair Comparisons; Benchmark Datasets" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 5, 4, 2, 5 ] }, { "paper_id": "X8XQOLjLX6", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Roel Bouman", "gender": "Male", "institution": "Radboud University", "country": "NL", "position": "Postdoc" }, { "name": "Tom Heskes", "gender": "Male", "institution": "", "country": "", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 17, "n_ref_uni": 36, "n_ref": 70, "n_ref_all": 89, "n_fig": 12, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1224, "formula_len_all_1": 1112, "len_all": 113011, "len_all_1": 67859, "len_abs": 1070, "len_title": 98, "len_sents": 37246, "len_sents_1": 32395, "n_sents": 273, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1073, "title": "Autoencoders for Anomaly Detection are Unreliable", "abs": "Autoencoders are frequently used for anomaly detection, both in the unsupervised and semi-supervised settings. They rely on the assumption that when trained using the reconstruction loss, they will be able to reconstruct normal data more accurately than anomalous data. Some recent works have posited that this assumption may not always hold, but little has been done to study the validity of the assumption in theory. In this work we show that this assumption indeed does not hold, and illustrate that anomalies, lying far away from normal data, can be perfectly reconstructed in practice. We extend the understanding of autoencoders for anomaly detection by showing how they can perfectly reconstruct out of bounds, or extrapolate undesirably, and note how this can be dangerous in safety critical applications. We connect theory to practice by showing that the proven behavior in linear autoencoders also occurs when applying non-linear autoencoders on both tabular data and real-world image data, the two primary application areas of autoencoders for anomaly detection.", "keywords": [ "Anomaly Detection", "Autoencoder", "Reliability", "AI Safety", "Adversarial Attacks" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "X8RTdxzqJQ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Liuhua Peng", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Associate Professor" }, { "name": "Mingming Gong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Xunye Tian", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" }, { "name": "Zhijian Zhou", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 8, "n_ref_uni": 47, "n_ref": 103, "n_ref_all": 134, "n_fig": 7, "n_tab": 7, "L_tab": 1943, "n_element_tab": 221, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 684, "n_element_tab_1": 94, "formula_len_all": 2380, "formula_len_all_1": 451, "len_all": 175500, "len_all_1": 69793, "len_abs": 1740, "len_title": 127, "len_sents": 56518, "len_sents_1": 32236, "n_sents": 403, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1745, "title": "Revisit Non-parametric Two-sample Testing as a Semi-supervised Learning Problem", "abs": "Learning effective data representations is crucial in answering if two samples X and Y are from the same distribution (a.k.a. the non-parametric two-sample testing problem), which can be categorized into: i) learning discriminative representations (DRs) that distinguish between two samples in a supervised-learning paradigm, and ii) learning inherent representations (IRs) focusing on data's inherent features in an unsupervised-learning paradigm. \nHowever, both paradigms have issues: learning DRs reduces the data points available for the two-sample testing phase, and learning purely IRs misses discriminative cues. To mitigate both issues, we propose a novel perspective to consider non-parametric two-sample testing as a semi-supervised learning (SSL) problem, introducing the SSL-based Classifier Two-Sample Test (SSL-C2ST) framework. While a straightforward implementation of SSL-C2ST might directly use existing state-of-the-art (SOTA) SSL methods to train a classifier with labeled data (with sample indexes X or Y) and unlabeled data (the remaining ones in the two samples), conventional two-sample testing data often exhibits substantial overlap between samples and violates SSL methods' assumptions, resulting in low test power. Therefore, we propose a two-step approach: first, learn IRs using all data, then fine-tune IRs with only labelled data to learn DRs, which can both utilize information from whole dataset and adapt the discriminative power to the given data. Extensive experiments and theoretical analysis demonstrate that SSL-C2ST outperforms traditional C2ST by effectively leveraging unlabeled data. We also offer a stronger empirically designed test achieving the SOTA performance in many two-sample testing datasets.", "keywords": [ "Two-sample Testing", "Semi-supervised learning", "Hypothesis testing" ], "rating_list": [ 6, 3, 3, 6, 6 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 3, 1, 1, 3, 4 ], "contribution_list": [ 3, 1, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 2, 3 ] }, { "paper_id": "X8Mhumi52G", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Brendon Boldt", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Charis Graham", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Undergrad student" }, { "name": "David R Mortensen", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Research Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 16, "n_ref": 38, "n_ref_all": 53, "n_fig": 5, "n_tab": 5, "L_tab": 10528, "n_element_tab": 603, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 474, "n_element_tab_1": 59, "formula_len_all": 44, "formula_len_all_1": 44, "len_all": 151789, "len_all_1": 62454, "len_abs": 1300, "len_title": 46, "len_sents": 37458, "len_sents_1": 32645, "n_sents": 214, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1299, "title": "ELCC: the Emergent Language Corpus Collection", "abs": "We introduce the Emergent Language Corpus Collection (ELCC): a collection of corpora generated from open source implementations of emergent communication systems across the literature. These systems include a variety of signalling game environments as well as more complex tasks environments like a social deduction game and embodied navigation. Each corpus is annotated with metadata describing the characteristics of the source system as well as a suite of analyses of the corpus (e.g., size, entropy, average message length, performance as transfer learning data). Currently, research studying emergent languages requires directly running different systems which takes time away from actual analyses of such languages, makes studies which compare diverse emergent languages rare, and presents a barrier to entry for researchers without a background in deep learning. The availability of a substantial collection of well-documented emergent language corpora, then, will enable research which can analyze a wider variety of emergent languages; this more effectively uncovers general principles in emergent communication rather than artifacts of particular environments. We provide some quantitative and qualitative analyses with ELCC to demonstrate potential use cases of the resource in this vein.", "keywords": [ "emergent language", "emergent communication", "dataset", "resource" ], "rating_list": [ 1, 3, 6, 6 ], "soundness_list": [ 2, 2, 4, 4 ], "presentation_list": [ 2, 1, 4, 4 ], "contribution_list": [ 1, 1, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "X8M4fansEz", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Bingjie Xu", "gender": "unknown", "institution": "Singapore Institute of Technology", "country": "SG", "position": "Assistant Professor" }, { "name": "Jiawei Mao", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Intern" }, { "name": "Mude Hui", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Xiaoke Huang", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Yuanqi Chang", "gender": "Male", "institution": "HDU", "country": "CN", "position": "MS student" }, { "name": "Yunfei Xie", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yuyin Zhou", "gender": "unknown", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 40, "n_ref": 88, "n_ref_all": 113, "n_fig": 18, "n_tab": 6, "L_tab": 2334, "n_element_tab": 235, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 624, "n_element_tab_1": 28, "formula_len_all": 1154, "formula_len_all_1": 1024, "len_all": 125167, "len_all_1": 50861, "len_abs": 1034, "len_title": 127, "len_sents": 32744, "len_sents_1": 22860, "n_sents": 253, "n_sents_1": 156, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1365, "title": "Story-Adapter: A Training-free Iterative Framework for Long Story Visualization", "abs": "Story visualization, the task of generating coherent images based on a narrative, has seen significant advancements with the emergence of text-to-image models, particularly diffusion models. However, maintaining semantic consistency, generating high-quality fine-grained interactions, and ensuring computational feasibility remain challenging, especially in long story visualization (i.e., up to 100 frames). In this work, we propose a training-free and computationally efficient framework, termed Story-Adapter, to enhance the generative capability of long stories. Specifically, we propose an iterative paradigm to refine each generated image, leveraging both the text prompt and all generated images from the previous iteration. Central to our framework is a training-free global reference crossattention module, which aggregates all generated images from the previous iteration to preserve semantic consistency across the entire story, while minimizing computational costs with global embeddings. This iterative process progressively optimizes image generation by repeatedly incorporating text constraints, resulting in more precise and fine-grained interactions. Extensive experiments validate the superiority of Story-Adapter in improving both semantic consistency and generative capability for fine-grained interactions, particularly in long story scenarios.", "keywords": [ "story visualization; diffusion model; subject-consistent image generation" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 1, 3 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "X8E65IxA73", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Bin Liang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Hui Wang", "gender": "Male", "institution": "Cloud Computing", "country": "CN", "position": "Full Professor" }, { "name": "Kam-Fai Wong", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "Full Professor" }, { "name": "Lin Gui", "gender": "Male", "institution": "King's College London, University of London", "country": "", "position": "Lecturer" }, { "name": "Ruifeng Xu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Shiwei Chen", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Yue Yu", "gender": "Male", "institution": "PengCheng Lab", "country": "CN", "position": "Full Professor" } ], "n_formula": 36, "n_formula_1": 15, "n_ref_uni": 50, "n_ref": 122, "n_ref_all": 144, "n_fig": 7, "n_tab": 13, "L_tab": 5383, "n_element_tab": 449, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 4160, "n_element_tab_1": 228, "formula_len_all": 2139, "formula_len_all_1": 1288, "len_all": 197829, "len_all_1": 58110, "len_abs": 1869, "len_title": 88, "len_sents": 58457, "len_sents_1": 23112, "n_sents": 425, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1876, "title": "Centrality-guided Pre-training for Graph", "abs": "Self-supervised learning (SSL) has shown great potential in learning generalizable representations for graph-structured data. However, existing SSL-based graph pre-training methods largely focus on improving graph representations by learning the structure information based on disturbing or reconstructing graphs, which ignores an important issue: the importance of different nodes in the graph structure may vary.\nTo fill this gap, we propose a Centrality-guided Graph Pre-training (CenPre) framework to integrate the distinct importance of nodes in graph structure into the corresponding representations of nodes based on the centrality in graph theory. In this way, the different roles played by different nodes can be effectively leveraged when learning graph structure. The proposed CenPre contains three modules for node representation pre-training and alignment. The node-level structure learning module fuses the fine-grained node importance into node representation based on degree centrality, allowing the aggregation of node representations with equal/similar importance. The graph-level structure learning module characterizes the importance between all nodes in the graph based on eigenvector centrality, enabling the exploitation of graph-level structure similarities/differences when learning node representation. Finally, a representation alignment module aligns the pre-trained node representation using the original one, essentially allowing graph representations to learn structural information without losing their original semantic information, thereby leading to better graph representations. Extensive experiments on a series of real-world datasets demonstrate that the proposed CenPre outperforms the state-of-the-art baselines in node classification and achieves better performance in link prediction and graph classification than the baseline models.", "keywords": [ "Graph neural networks", "Graph representation", "Graph pre-training", "Graph centrality" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "X7eAhXcps1", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Adil Salim", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "PASCAL BIANCHI", "gender": "unknown", "institution": "T\u00e9l\u00e9com ParisTech", "country": "", "position": "Full Professor" }, { "name": "Victor Priser", "gender": "Male", "institution": "T\u00e9l\u00e9com Paris", "country": "", "position": "PhD student" } ], "n_formula": 97, "n_formula_1": 31, "n_ref_uni": 32, "n_ref": 67, "n_ref_all": 107, "n_fig": 3, "n_tab": 4, "L_tab": 1340, "n_element_tab": 137, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 812, "n_element_tab_1": 21, "formula_len_all": 9246, "formula_len_all_1": 1922, "len_all": 183633, "len_all_1": 68305, "len_abs": 962, "len_title": 83, "len_sents": 51751, "len_sents_1": 25655, "n_sents": 593, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 975, "title": "Long-time asymptotics of noisy SVGD outside the population limit", "abs": "Stein Variational Gradient Descent (SVGD) is a widely used sampling algorithm that has been successfully applied in several areas of Machine Learning. SVGD operates by iteratively moving a set of $n$ interacting particles (which represent the samples) to approximate the target distribution. Despite recent studies on the complexity of SVGD and its variants, their long-time asymptotic behavior (i.e., after numerous iterations $k$) is still not understood in the finite number of particles regime. We study the long-time asymptotic behavior of a noisy variant of SVGD. First, we establish that the limit set of noisy SVGD for large $k$ is well-defined. We then characterize this limit set, showing that it approaches the target distribution as $n$ increases. In particular, noisy SVGD provably avoids the variance collapse observed for SVGD. Our approach involves demonstrating that the trajectories of noisy SVGD closely resemble those described by a McKean-Vlasov process.", "keywords": [ "Stochastic approximation", "sampling", "convergence", "interacting particles system", "dynamical systems", "Stein Variational Gradient Descent", "McKean-Vlasov equation" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "X7dQuJqs8c", "primary_area": "applications to robotics, autonomy, planning", "n_author": 12, "author_info_list": [ { "name": "Changhua Pei", "gender": "Male", "institution": "Computer Network Information Center, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Dan Pei", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Fengrui Liu", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Gaogang Xie", "gender": "Male", "institution": "Computer Network Information Center, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Jianhui li", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Full Professor" }, { "name": "Jianjun Chen", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Director, Head of Infrastructure System Lab" }, { "name": "Rong Kang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Tieying Zhang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Postdoc" }, { "name": "Xiao He", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Zexin Wang", "gender": "Male", "institution": "Computer Network Information Center, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zeyan Li", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "\u5218\u626c", "gender": "Female", "institution": "Computer Network Information Center, Chinese Academy of Sciences. ", "country": "CN", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 23, "n_ref": 42, "n_ref_all": 58, "n_fig": 5, "n_tab": 4, "L_tab": 1869, "n_element_tab": 154, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 476, "n_element_tab_1": 64, "formula_len_all": 60, "formula_len_all_1": 0, "len_all": 132942, "len_all_1": 61907, "len_abs": 1757, "len_title": 129, "len_sents": 43832, "len_sents_1": 29213, "n_sents": 347, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1768, "title": "Flow-of-Action: SOP Enhanced LLM-Based Multi-Agent System for Root Cause Analysis", "abs": "In the realm of microservices architecture, the occurrence of frequent incidents necessitates the employment of Root Cause Analysis (RCA) for swift issue resolution. It is common that a serious incident can take several domain experts hours to identify the root cause. Consequently, a contemporary trend involves harnessing Large Language Models (LLMs) as automated agents for RCA. Though the recent ReAct framework aligns well with the Site Reliability Engineers (SREs) for its thought-action-observation paradigm, its hallucinations often lead to irrelevant actions and directly affect subsequent results. Additionally, the complex and variable clues of the incident can overwhelm the model one step further. To confront these challenges, we propose Flow-of-Action, a pioneering Standard Operation Procedure (SOP) enhanced LLM-based multi-agent system. By explicitly summarizing the diagnosis steps of SREs, SOP imposes constraints on LLMs at crucial junctures, guiding the RCA process towards the correct trajectory. To facilitate the rational and effective utilization of SOPs, we design an SOP-centric framework called SOP flow. SOP flow contains a series of tools, including one for finding relevant SOPs for incidents, another for automatically generating SOPs for incidents without relevant ones, and a tool for converting SOPs into code. This significantly alleviates the hallucination issues of ReAct in RCA tasks. We also design multiple auxiliary agents to assist the main agent by removing useless noise, narrowing the search space, and informing the main agent whether the RCA procedure can stop. Compared to the ReAct method's 35.50\\% accuracy, our Flow-of-Action method achieves 64.01\\%, meeting the accuracy requirements for RCA in real-world systems.", "keywords": [ "Root Cause Analysis", "Multi-Agent System", "Standard Operation Procedure" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "X7XgNI0Eym", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Xiaobiao Du", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Xin Yu", "gender": "Male", "institution": "Google", "country": "AU", "position": "Researcher" }, { "name": "Yida Wang", "gender": "Male", "institution": "Li Auto Inc.", "country": "CN", "position": "Senior Research Engineer" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 31, "n_ref": 103, "n_ref_all": 147, "n_fig": 14, "n_tab": 7, "L_tab": 5150, "n_element_tab": 809, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 2270, "n_element_tab_1": 294, "formula_len_all": 381, "formula_len_all_1": 381, "len_all": 162718, "len_all_1": 61866, "len_abs": 2024, "len_title": 119, "len_sents": 47827, "len_sents_1": 27988, "n_sents": 345, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 2102, "title": "MVGS: Multi-view-regulated Gaussian Splatting for Novel View Synthesis", "abs": "Recent works in volume rendering, \\textit{e.g.} NeRF and 3D Gaussian Splatting (3DGS), significantly advance the rendering quality and efficiency with the help of the learned implicit neural radiance field or 3D Gaussians. \nRendering on top of an explicit representation, the vanilla 3DGS and its variants deliver real-time efficiency by optimizing the parametric model with single-view supervision per iteration during training which is adopted from NeRF. Consequently, certain views are overfitted, leading to unsatisfying appearance in novel-view synthesis and imprecise 3D geometries.\nTo solve aforementioned problems, we propose a new 3DGS optimization method embodying four key novel contributions:\n1) We transform the conventional single-view training paradigm into a multi-view training strategy. With our proposed multi-view regulation, 3D Gaussian attributes are further optimized without overfitting certain training views. As a general solution, we improve the overall accuracy in a variety of scenarios and different Gaussian variants. \n2) Inspired by the benefit introduced by additional views, we further propose a cross-intrinsic guidance scheme, leading to a coarse-to-fine training procedure concerning different resolutions.\n3) Built on top of our multi-view regulated training, we further propose a cross-ray densification strategy, densifying more Gaussian kernels in the ray-intersect regions from a selection of views. \n4) By further investigating the densification strategy, we found that the effect of densification should be enhanced when certain views are distinct dramatically.\nAs a solution, we propose a novel multi-view augmented densification strategy, where 3D Gaussians are encouraged to get densified to a sufficient number accordingly, resulting in improved reconstruction accuracy.\nWe conduct extensive experiments to demonstrate that our proposed method is capable of improving novel view synthesis of the Gaussian-based explicit representation methods about 1 dB PSNR for various tasks. \\href{https://mvgs666.github.io/}{\\textcolor{magenta}{Codes are available.}}", "keywords": [ "3D Reconstruction; Gaussian Splatting; Neural Radiance Feild" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "X7SQiI5Uul", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 10, "author_info_list": [ { "name": "Boya Wu", "gender": "unknown", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Hongwang Xiao", "gender": "unknown", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Hui Wang", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Algorithm engineer" }, { "name": "Jiashan Li", "gender": "unknown", "institution": "Renmin University of China", "country": "", "position": "PhD student" }, { "name": "Kai Chen", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Qiwei Ye", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Principal Researcher" }, { "name": "Sicheng Dai", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Wenjun Lin", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Xi Chen", "gender": "unknown", "institution": "BAAI", "country": "", "position": "Researcher" }, { "name": "Yuancheng SUN", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 69, "n_ref_all": 91, "n_fig": 7, "n_tab": 12, "L_tab": 4535, "n_element_tab": 426, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 706, "n_element_tab_1": 66, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 205361, "len_all_1": 67625, "len_abs": 2159, "len_title": 148, "len_sents": 55880, "len_sents_1": 33359, "n_sents": 426, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1428, "title": "STELLA: Leveraging Structural Representations to Enhance Protein Understanding with Multimodal LLMs", "abs": "Protein biology centers on the intricate relationships among sequence, structure, and function (text), with structure understanding being a crucial aspect for uncovering protein biological functions. Traditional methods based on protein language models (pLMs) often focus on specific aspects of biological function prediction but do not account for the broader, dynamic context of protein research\u2014an important component for addressing the complexity of protein biology. Modern large language models (LLMs) excel in human-machine interaction, language understanding and generation, at a human-like level. By bridging structural representations with the contextual knowledge encoded within LLMs, STELLA leverages the strengths of LLMs to enable versatile and accurate predictions in protein-related tasks. It showcases the transformative potential of multimodal LLMs as a novel paradigm besides pLMs in advancing protein biology research by achieving state-of-the-art performance in both functional description and enzyme-catalyzed reaction prediction tasks. This study not only establishes an innovative LLM-based paradigm to understand proteins, but also expands the boundaries of LLM capabilities in protein biology. To foster collaboration and inspire further innovation, the codes, datasets, and pre-trained models are made publicly available at the anonymous GitHub repository https://anonymous.4open.science/r/STELLA-DF00.", "keywords": [ "Protein Function Prediction", "Enzyme-Catalyzed Reaction Prediction", "Multimodal Large Language Models", "Structural Representations", "Protein Biology", "Computational Biology" ], "rating_list": [ 6, 8, 3, 10, 5 ], "soundness_list": [ 3, 3, 2, 4, 3 ], "presentation_list": [ 4, 4, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 4, 3 ], "confidence_list": [ 3, 4, 3, 3, 3 ] }, { "paper_id": "X7EMiddusV", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Alexandru Oarga", "gender": "Not Specified", "institution": "University of Barcelona", "country": "ES", "position": "MS student" }, { "name": "Jorge J\u00falvez", "gender": "unknown", "institution": "Universidad de Zaragoza", "country": "ES", "position": "Full Professor" } ], "n_formula": 25, "n_formula_1": 17, "n_ref_uni": 47, "n_ref": 87, "n_ref_all": 127, "n_fig": 17, "n_tab": 4, "L_tab": 976, "n_element_tab": 89, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 624, "n_element_tab_1": 64, "formula_len_all": 1470, "formula_len_all_1": 1158, "len_all": 179400, "len_all_1": 68303, "len_abs": 1363, "len_title": 99, "len_sents": 55342, "len_sents_1": 29236, "n_sents": 464, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1367, "title": "Directed graph transformers meet metabolic networks", "abs": "Technical advances in sequencing have allowed the reconstruction of genome-scale metabolic models (GEMs) for a wide range of microorganisms. These models have been particularly useful for the prediction of essential genes and reactions, which are potential targets for antimicrobial therapies. However, current methods for essentiality prediction are computationally limited and are not able to accommodate the increasingly available data. Motivated by the success of data-driven approaches in other domains, this work introduces the metabolic transformer, a model designed for holistic identification of essential reactions in genome-scale models, entirely trained on synthetic knock-out data. It is demonstrated that the problem of essential reaction prediction can be theoretically formulated as the identification of redundant nodes in directed bipartite graphs. This reveals the limitations of message-passing schemes and motivates the development of a novel graph transformer architecture specifically tailored for metabolic networks. The proposed architecture is capable of addressing the essential reaction identification problem by capturing both the directionality and global structure of metabolic networks. To demonstrate the effectiveness of our approach, we composed a large-scale dataset of genome-scale models reconstructed from real microorganisms.", "keywords": [ "Graph Transformers", "Genome-Scale Metabolic Models", "Metabolic Networks", "Essential Reactions", "Petri Nets", "Biological Networks", "Systems Biology" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 4, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 3, 4 ] }, { "paper_id": "X75isqETqR", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Dongxie Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Hanyan Yin", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "MS student" }, { "name": "Xiao Zhang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zhewei Wei", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 102, "n_formula_1": 30, "n_ref_uni": 24, "n_ref": 51, "n_ref_all": 64, "n_fig": 8, "n_tab": 3, "L_tab": 890, "n_element_tab": 199, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12369, "formula_len_all_1": 1892, "len_all": 152449, "len_all_1": 68124, "len_abs": 1704, "len_title": 111, "len_sents": 47379, "len_sents_1": 29164, "n_sents": 452, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1775, "title": "Matrix Sketching in Bandits: Current Pitfalls and New Framework", "abs": "The utilization of sketching techniques has progressively emerged as a pivotal method for enhancing the efficiency of online learning. \n In linear bandit settings, current sketch-based approaches leverage matrix sketching to reduce the per-round time complexity from $\\Omega\\left(d^2\\right)$ to $O(d)$, where $d$ is the input dimension. Despite this improved efficiency, these approaches encounter critical pitfalls: if the spectral tail of the covariance matrix does not decrease rapidly, it can lead to linear regret.\n In this paper, we revisit the regret analysis and algorithm design concerning approximating the covariance matrix using matrix sketching in linear bandits. \n We illustrate how inappropriate sketch sizes can result in unbounded spectral loss, thereby causing linear regret. \n To prevent this issue, we propose Dyadic Block Sketching, an innovative streaming matrix sketching approach that adaptively manages sketch size to constrain global spectral loss. \n This approach effectively tracks the best rank-$k$ approximation in an online manner, ensuring efficiency when the geometry of the covariance matrix is favorable. \n Then, we apply the proposed Dyadic Block Sketching to linear bandits and demonstrate that the resulting bandit algorithm can achieve sublinear regret without prior knowledge of the covariance matrix, even under the worst case. \n Our method is a general framework for efficient sketch-based linear bandits, applicable to all existing sketch-based approaches, and offers improved regret bounds accordingly.\n Additionally, we conduct comprehensive empirical studies using both synthetic and real-world data to validate the accuracy of our theoretical findings and to highlight the effectiveness of our algorithm.", "keywords": [ "Linear Bandits", "Matrix Sketching", "Regret Analysis" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "X6y5CC44HM", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Bastian Rieck", "gender": "Male", "institution": "University of Fribourg", "country": "CH", "position": "Full Professor" }, { "name": "Carles Casacuberta", "gender": "Male", "institution": "Universitat de Barcelona", "country": "ES", "position": "Full Professor" }, { "name": "Daniel Bin Schmid", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "MS student" }, { "name": "Ernst R\u00f6ell", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Mathieu Alain", "gender": "Male", "institution": "University College London", "country": "", "position": "PhD student" }, { "name": "Rub\u00e9n Ballester", "gender": "Male", "institution": "Universitat de Barcelona", "country": "ES", "position": "PhD student" }, { "name": "Sergio Escalera", "gender": "Male", "institution": "University of Barcelona", "country": "ES", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 76, "n_ref_all": 90, "n_fig": 2, "n_tab": 20, "L_tab": 6446, "n_element_tab": 516, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 832, "n_element_tab_1": 80, "formula_len_all": 705, "formula_len_all_1": 0, "len_all": 219766, "len_all_1": 67293, "len_abs": 1134, "len_title": 94, "len_sents": 62441, "len_sents_1": 36448, "n_sents": 454, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 46, "L_abs": 1140, "title": "MANTRA: The Manifold Triangulations Assemblage", "abs": "The rising interest in leveraging higher-order interactions present in complex systems has led to a surge in more expressive models exploiting high-order structures in the data, especially in topological deep learning (TDL), which designs neural networks on high-order domains such as simplicial complexes. However, progress in this field is hindered by the scarcity of datasets for benchmarking these architectures. To address this gap, we introduce MANTRA, the first large-scale, diverse, and intrinsically high-order dataset for benchmarking high-order models, comprising over 43,000 and 249,000 triangulations of surfaces and three-dimensional manifolds, respectively. With MANTRA, we assess several graph- and simplicial complex-based models on three topological classification tasks. We demonstrate that while simplicial complex-based neural networks generally outperform their graph-based counterparts in capturing simple topological invariants, they also struggle, suggesting a rethink of TDL. Thus, MANTRA serves as a benchmark for assessing and advancing topological\nmethods, leading the way for more effective high-order models.", "keywords": [ "simplicial complex", "topological deep learning", "high-order", "high-order dataset", "simplicial complex learning" ], "rating_list": [ 8, 3, 6, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "X6xzYP2cMk", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Alireza Naderi", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Jared Tanner", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "Full Professor" }, { "name": "Thiziri Nait Saada", "gender": "Female", "institution": "University of Oxford", "country": "", "position": "PhD student" } ], "n_formula": 133, "n_formula_1": 34, "n_ref_uni": 20, "n_ref": 40, "n_ref_all": 64, "n_fig": 11, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9070, "formula_len_all_1": 1477, "len_all": 154490, "len_all_1": 64470, "len_abs": 1492, "len_title": 137, "len_sents": 46930, "len_sents_1": 27637, "n_sents": 490, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1544, "title": "Mind the Gap: a Spectral Analysis of Rank Collapse and Signal Propagation in Transformers", "abs": "Attention layers are the core component of transformers, the current state-of-the-art neural network architecture. However, \\softmaxx-based attention puts transformers' trainability at risk. Even \\textit{at initialisation}, the propagation of signals and gradients through the random network can be pathological, resulting in known issues such as (i) vanishing/exploding gradients and (ii) \\textit{rank collapse}, i.e. when all tokens converge to a single representation \\textit{with depth}. This paper examines signal propagation in \\textit{attention-only} transformers from a random matrix perspective, illuminating the origin of such issues, as well as unveiling a new phenomenon---(iii) rank collapse \\textit{in width}. Modelling \\softmaxx-based attention at initialisation with Random Markov matrices, our theoretical analysis reveals that a \\textit{spectral gap} between the two largest singular values of the attention matrix causes (iii), which, in turn, exacerbates (i) and (ii). Building on this insight, we propose a novel, yet simple, practical solution to resolve rank collapse in width by removing the spectral gap. Moreover, we validate our findings and discuss the training benefits of the proposed fix through experiments that also motivate a revision of some of the default parameter scaling. Our attention model accurately describes the standard key-query attention in a single-layer transformer, making this work a significant first step towards a better understanding of the initialisation dynamics in the multi-layer case.", "keywords": [ "transformers", "attention mechanism", "spectral analysis", "initialisation", "random matrix theory", "signal propagation" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "X6ffdf6nh3", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Chaoren Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Undergrad student" }, { "name": "Haorui He", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "RA" }, { "name": "Junan Zhang", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Xueyao Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Yicheng Gu", "gender": "Non-Binary", "institution": "The Chinese University of Hong Kong, Shen Zhen", "country": "", "position": "Undergrad student" }, { "name": "Zhizheng Wu", "gender": "Male", "institution": "Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Zihao Fang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "Undergrad student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 45, "n_ref": 65, "n_ref_all": 86, "n_fig": 7, "n_tab": 5, "L_tab": 1662, "n_element_tab": 133, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1322, "n_element_tab_1": 129, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 133708, "len_all_1": 44567, "len_abs": 874, "len_title": 126, "len_sents": 27937, "len_sents_1": 16762, "n_sents": 219, "n_sents_1": 140, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 942, "title": "SingNet: Towards a Large-Scale, Diverse, and In-the-Wild Singing Voice Dataset", "abs": "The lack of a publicly-available large-scale and diverse dataset has long been a significant bottleneck for singing voice applications like Singing Voice Synthesis (SVS) and Singing Voice Conversion (SVC). To tackle this problem, we present SingNet, an extensive, diverse, and in-the-wild singing voice dataset. Specifically, we propose a data processing pipeline to extract ready-to-use training data from sample packs and songs on the internet, forming 3000 hours of singing voices in various languages and styles. Furthermore, to facilitate the use and demonstrate the effectiveness of SingNet, we pre-train and open-source various state-of-the-art (SOTA) models on Wav2vec2, BigVGAN, and NSF-HiFiGAN based on our collected singing voice data. We also conduct benchmark experiments on Automatic Lyric Transcription (ALT), Neural Vocoder, and Singing Voice Conversion (SVC). Audio demos are available at: https://singnet-dataset.github.io/.", "keywords": [ "Singing Voice Dataset", "Singing Voice Corpus", "Self-supervised Learning", "Automatic Lyric Transcription", "Neural Vocoder", "Singing Voice Conversion" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "X6W5eqhzDx", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Caroline Horsch", "gender": "Female", "institution": "Delft University of Technology", "country": "NL", "position": "PhD student" }, { "name": "Matthijs T. J. Spaan", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "Full Professor" }, { "name": "Max Weltevrede", "gender": "unknown", "institution": "Delft University of Technology", "country": "", "position": "PhD student" }, { "name": "Wendelin Boehmer", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 110, "n_ref_all": 166, "n_fig": 17, "n_tab": 3, "L_tab": 1028, "n_element_tab": 100, "n_fig_1": 13, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 96, "formula_len_all_1": 0, "len_all": 200346, "len_all_1": 62536, "len_abs": 1367, "len_title": 125, "len_sents": 55629, "len_sents_1": 30178, "n_sents": 413, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1002, "title": "Training on more Reachable Tasks for Generalisation in Reinforcement Learning", "abs": "In multi-task reinforcement learning, agents train on a fixed set of tasks and have to generalise to new ones. Recent work has shown that increased exploration improves this generalisation, but it remains unclear why exactly that is. In this paper, we introduce the concept of reachability in multi-task reinforcement learning and show that an initial exploration phase increases the number of reachable tasks the agent is trained on. This, and not the increased exploration, is responsible for the improved generalisation, even to unreachable tasks. Inspired by this, we propose a novel method Explore-Go that implements such an exploration phase at the beginning of each episode. Explore-Go only modifies the way experience is collected and can be used with most existing on-policy or off-policy reinforcement learning algorithms. We demonstrate the effectiveness of our method when combined with some popular algorithms and show an increase in generalisation performance across several environments.", "keywords": [ "Deep Reinforcement Learning", "Exploration", "Generalization" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "X6VVK8pIzZ", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Fawaz Sammani", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Intern" }, { "name": "Nikos Deligiannis", "gender": "Male", "institution": "IMEC", "country": "BE", "position": "Principal Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 52, "n_ref": 76, "n_ref_all": 107, "n_fig": 11, "n_tab": 13, "L_tab": 7311, "n_element_tab": 697, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1443, "n_element_tab_1": 93, "formula_len_all": 116, "formula_len_all_1": 116, "len_all": 192103, "len_all_1": 60786, "len_abs": 1521, "len_title": 87, "len_sents": 53609, "len_sents_1": 28285, "n_sents": 423, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 39, "L_abs": 1046, "title": "Zero-Shot Natural Language Explanations", "abs": "Natural Language Explanations (NLEs) interpret the decision-making process of a given model through textual sentences. Current NLEs suffer from a severe limitation; they are unfaithful to the model\u2019s actual reasoning process, as a separate textual decoder is explicitly trained to generate those explanations using annotated datasets for a specific task, leading them to reflect what annotators desire. In this work, we take the first step towards generating faithful NLEs for any visual classification model without any training data. Our approach models the relationship between class embeddings from the classifier of the vision model and their corresponding class names via a simple MLP which trains in seconds. After training, we can map any new text to the classifier space and measure its association with the visual features. We conduct experiments on 38 vision models, including both CNNs and Transformers. In addition to NLEs, our method offers other advantages such as zero-shot image classification and fine-grained concept discovery.", "keywords": [ "Natural Language Explanations", "interpretability", "explainability" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "X65IKSuWQo", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Anh Tuan Bui", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Postdoc" }, { "name": "Khanh Doan", "gender": "Male", "institution": "VinAI Research", "country": "VN", "position": "Intern" }, { "name": "Long Tung Vuong", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Quyen Tran", "gender": "Female", "institution": "VinAI Research", "country": "VN", "position": "Research Resident" }, { "name": "Trung Le", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" }, { "name": "Dinh Phung", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" }, { "name": "Mehrtash T. Harandi", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Associate Professor" }, { "name": "Thanh-Toan Do", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" } ], "n_formula": 36, "n_formula_1": 37, "n_ref_uni": 12, "n_ref": 72, "n_ref_all": 98, "n_fig": 2, "n_tab": 1, "L_tab": 137, "n_element_tab": 25, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 2347, "n_element_tab_1": 52, "formula_len_all": 4592, "formula_len_all_1": 4651, "len_all": 69085, "len_all_1": 53055, "len_abs": 817, "len_title": 105, "len_sents": 22898, "len_sents_1": 19841, "n_sents": 173, "n_sents_1": 153, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 820, "title": "Unified Perspectives on Signal-to-Noise Diffusion Models", "abs": "Diffusion models (DM) have become essential components of generative modeling, demonstrating exceptional performance in domains like image synthesis, audio generation, and complex data interpolation. Signal-to-Noise diffusion models represent a broad family encompassing many state-of-the-art models. Although several efforts have been made to explore Signal-to-Noise (S2N) diffusion models from different angles, a comprehensive study that connects these viewpoints and introduces new insights is still needed. In this work, we provide an in-depth perspective on noise schedulers, analyzing their role through the lens of the signal-to-noise ratio (SNR) and its relationship to information theory. Based on this framework, we introduce a generalized backward equation to improve the efficiency of the inference process.", "keywords": [ "diffusion models", "SDE", "ODE", "continuous variation models" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 1, 1 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "X5tBNz4qtl", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Akshat Kumar", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Associate Professor" }, { "name": "Pradeep Varakantham", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Full Professor" }, { "name": "Ze Gong", "gender": "unknown", "institution": "Singapore Management University", "country": "SG", "position": "Postdoc" } ], "n_formula": 25, "n_formula_1": 18, "n_ref_uni": 28, "n_ref": 78, "n_ref_all": 115, "n_fig": 7, "n_tab": 4, "L_tab": 1527, "n_element_tab": 357, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1923, "formula_len_all_1": 1463, "len_all": 136480, "len_all_1": 69695, "len_abs": 1484, "len_title": 100, "len_sents": 45569, "len_sents_1": 33643, "n_sents": 333, "n_sents_1": 237, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 1488, "title": "Offline Safe Policy Optimization From Human Feedback", "abs": "Offline preference-based reinforcement learning (PbRL) learns rewards and policies aligned with human preferences without the need for extensive reward engineering and direct interaction with human annotators. However, ensuring safety remains a critical challenge across many domains and tasks. Previous works on safe RL from human feedback (RLHF) first learn reward and cost models from offline data, and then use constrained RL to optimize a safe policy. However, inaccuracies in the reward and cost learning can impair performance when used with constrained RL methods. To address these challenges, (a) we introduce a framework that learns a policy based on pairwise preferences regarding the agent\u2019s behavior in terms of rewards, as well as binary labels indicating the safety of trajectory segments, without access to ground-truth rewards or costs; (b) we combine the preference learning module with safety alignment in a constrained optimization problem. This optimization problem is solved using a Lagrangian method that directly learns reward maximizing safe policy without explicitly learning reward and cost models, avoiding the need for constrained RL; (c) to evaluate our approach, we construct new datasets with synthetic human feedback, built upon a well-established offline safe RL benchmark. Empirically, our method successfully learns safe policies with high rewards, outperforming baselines with ground-truth reward and cost, as well as state-of-the-art RLHF approaches.", "keywords": [ "Offline PbRL", "Safety Alignment" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "X5rO5VyTgB", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Hanxing Ding", "gender": "Male", "institution": "Institute of Computing Technology of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Huawei Shen", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Jingcheng Deng", "gender": "Male", "institution": "Institute of Computing Technology, CAS", "country": "", "position": "MS student" }, { "name": "Liang Pang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "Associate Professor" }, { "name": "Xueqi Cheng", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy", "country": "CN", "position": "Full Professor" }, { "name": "Zihao Wei", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 17, "n_ref_uni": 21, "n_ref": 51, "n_ref_all": 78, "n_fig": 3, "n_tab": 11, "L_tab": 4325, "n_element_tab": 456, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2704, "n_element_tab_1": 163, "formula_len_all": 952, "formula_len_all_1": 1069, "len_all": 173411, "len_all_1": 61434, "len_abs": 1526, "len_title": 142, "len_sents": 46617, "len_sents_1": 26175, "n_sents": 388, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1466, "title": "Everything is Editable: Extend Knowledge Editing to Unstructured Data in Large Language Models", "abs": "Recent knowledge editing methods have primarily focused on modifying structured knowledge in large language models. However, this task setting overlooks the fact that a significant portion of real-world knowledge is stored in an unstructured format, characterized by long-form content, noise, and a complex yet comprehensive nature.\nTechniques like ``local layer key-value storage'' and ``term-driven optimization'', as used in previous methods like MEMIT, are not effective for handling unstructured knowledge.\nTo address these challenges, we propose a novel \\textbf{Un}structured \\textbf{K}nowledge \\textbf{E}diting method, namely UnKE, which extends previous assumptions in the layer dimension and token dimension.\nFirstly, in the layer dimension, we propose non-local block key-value storage to replace local layer key-value storage, increasing the representation ability of key-value pairs and incorporating attention layer knowledge. \nSecondly, in the token dimension, we replace \u201cterm-driven optimization\u201d with \u201ccause-driven optimization\u201d, which edits the last token directly while preserving context, avoiding the need to locate terms and preventing the loss of context information.\nResults on newly proposed unstructure knowledge editing dataset (UnKEBench) and traditional structured datasets demonstrate that UnKE achieves remarkable performance, surpassing strong baselines. In addition, UnKE has robust batch editing and sequential editing capabilities.", "keywords": [ "Knowledge Editing", "Large Language Models" ], "rating_list": [ 6, 6, 5, 3, 8 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 4, 5, 5, 5, 4 ] }, { "paper_id": "X5hrhgndxW", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Louis Bradshaw", "gender": "Male", "institution": "Queen Mary, University of London", "country": "", "position": "PhD student" }, { "name": "Simon Colton", "gender": "Male", "institution": "Queen Mary, University of London", "country": "GB", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 40, "n_ref": 60, "n_ref_all": 82, "n_fig": 5, "n_tab": 4, "L_tab": 2961, "n_element_tab": 204, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1417, "n_element_tab_1": 205, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 124267, "len_all_1": 58629, "len_abs": 782, "len_title": 116, "len_sents": 36462, "len_sents_1": 27215, "n_sents": 254, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 62, "L_abs": 666, "title": "Aria-MIDI: A Dataset of MIDI Files for Symbolic Music Modeling", "abs": "We introduce an extensive new dataset of MIDI files, created by transcribing audio\nrecordings of piano performances into their constituent notes. The data pipeline\nwe use is multi-stage, employing a language model to autonomously crawl and\nscore audio recordings from the internet based on their metadata, followed by a\nstage of pruning and segmentation using an audio classifier. The resulting dataset\ncontains over one million distinct MIDI files, comprising roughly 100,000 hours\nof transcribed audio. We provide an in-depth analysis of our techniques, offering\nstatistical insights, and investigate the content by extracting metadata tags, which\nwe also provide.", "keywords": [ "music", "symbolic music", "piano transcription", "dataset", "midi" ], "rating_list": [ 3, 8, 6, 3 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "X4Rcxi9588", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Hong-Chen Wei", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Zhenzhong Chen", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 29, "n_formula_1": 13, "n_ref_uni": 24, "n_ref": 67, "n_ref_all": 82, "n_fig": 9, "n_tab": 6, "L_tab": 2227, "n_element_tab": 274, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 2341, "n_element_tab_1": 292, "formula_len_all": 1785, "formula_len_all_1": 799, "len_all": 139208, "len_all_1": 62162, "len_abs": 1846, "len_title": 127, "len_sents": 37168, "len_sents_1": 25994, "n_sents": 289, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1854, "title": "Visual Context Window Extension: A New Perspective for Long Video Understanding", "abs": "Large Multimodal Models (LMMs) have demonstrated impressive performance in short video understanding tasks but face great challenges when applied to long video understanding. In contrast, Large Language Models (LLMs) exhibit outstanding capabilities in modeling long texts. Existing work attempts to address this issue by introducing long video-text pairs during training. However, these approaches require substantial computational and data resources. In this paper, we tackle the challenge of long video understanding from the perspective of context windows, aiming to apply LMMs to long video tasks without retraining on long video datasets. We first conduct an in-depth analysis of why pretrained LMMs struggle to understand lengthy video content, identifying that discrepancies between visual and language modalities lead to different context windows for visual and language tokens, making it difficult to directly extend the visual tokens to match the language context window. Based on this, we propose to adapt LMMs for long video understanding tasks by extending the visual context window, eliminating the need for retraining on large-scale long video datasets. To further mitigate the significant memory consumption caused by long sequences, we introduce a progressive pooling inference strategy that selectively adjusts the spatial resolution of frame embeddings, reducing the number of visual tokens while retaining important spatial information. Across multiple long video understanding benchmarks, our method consistently improves the performance as the number of video frames increases. On the MLVU benchmark, our method outperforms GPT-4o, even though our model size is only 7B. Additionally, in the 256-frame setting, our method reduces memory usage by approximately 45% compared to the baseline, without introducing any performance loss.", "keywords": [ "Large Multimodal Models", "Long Video Understanding", "Visual Context Window" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "X3IcgZEUEi", "primary_area": "learning on time series and dynamical systems", "n_author": 1, "author_info_list": [ { "name": "Baptiste T Rossi", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 22, "n_ref_uni": 33, "n_ref": 52, "n_ref_all": 73, "n_fig": 7, "n_tab": 5, "L_tab": 908, "n_element_tab": 109, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2983, "n_element_tab_1": 101, "formula_len_all": 4006, "formula_len_all_1": 2494, "len_all": 144669, "len_all_1": 60353, "len_abs": 1046, "len_title": 107, "len_sents": 45829, "len_sents_1": 24862, "n_sents": 304, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1049, "title": "ODE Parameter Identification: An Integral Matching Approach", "abs": "We present a novel method to identify parameter of nonlinear Ordinary Differential Equations (ODEs) using time series data. Our approach fits parameters by matching a collocation-based estimate of the integral of the learned derivative to an interpolation of the trajectory, thus avoiding the computational cost of ODE solvers in adjoint methods and the sensitivity to noise of derivative estimates in gradient matching methods. By employing batching strategies based on time subintervals and state components, our method achieves linear complexity in relation to system dimensions and dataset sizes. The method is highly parallel enabling fast gradient evaluations and a faster convergence than adjoint methods. For fully observed systems, we demonstrate the method on canonical dynamical systems, where the method achieves speed-ups of three orders of magnitude over adjoint methods and an increased robustness against observational noise. We provide an extension to partially observed systems and demonstrate the method on the Lorenz63 attractor.", "keywords": [ "ODEs", "System Identification", "Collocation" ], "rating_list": [ 5, 3, 8, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 3, 3, 2, 2 ] }, { "paper_id": "X2x2DuGIbx", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Andrew Craig Cullen", "gender": "Male", "institution": "The University of Melbourne", "country": "AU", "position": "Postdoc" }, { "name": "Benjamin I. P. Rubinstein", "gender": "Male", "institution": "The University of Melbourne", "country": "AU", "position": "Full Professor" }, { "name": "Paul Montague", "gender": "Male", "institution": "Defence Science and Technology Group", "country": "AU", "position": "Researcher" }, { "name": "Sarah Monazam Erfani", "gender": "unknown", "institution": "The University of Melbourne", "country": "AU", "position": "Associate Professor" }, { "name": "Shijie Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Postdoc" } ], "n_formula": 28, "n_formula_1": 12, "n_ref_uni": 56, "n_ref": 93, "n_ref_all": 118, "n_fig": 4, "n_tab": 4, "L_tab": 1134, "n_element_tab": 119, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 453, "n_element_tab_1": 85, "formula_len_all": 1404, "formula_len_all_1": 808, "len_all": 158324, "len_all_1": 64600, "len_abs": 1119, "len_title": 137, "len_sents": 46463, "len_sents_1": 31294, "n_sents": 294, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1123, "title": "Multi-level Certified Defense Against Poisoning Attacks in Offline Reinforcement Learning", "abs": "Similar to other machine learning frameworks, Offline Reinforcement Learning (RL) is shown to be vulnerable to poisoning attacks, due to its reliance on externally sourced datasets, a vulnerability that is exacerbated by its sequential nature. To mitigate the risks posed by RL poisoning, we extend certified defenses to provide larger guarantees against adversarial manipulation, ensuring robustness for both per-state actions, and the overall expected cumulative reward. Our approach leverages properties of Differential Privacy, in a manner that allows this work to span both continuous and discrete spaces, as well as stochastic and deterministic environments---significantly expanding the scope and applicability of achievable guarantees. Empirical evaluations demonstrate that our approach ensures the performance drops to no more than 50% with up to 7% of the training data poisoned, significantly improving over the 0.008% in prior work (Wu et al., 2022), while producing certified radii that is 5 times larger as well. This highlights the potential of our framework to enhance safety and reliability in offline RL.", "keywords": [ "Adversarial Machine Learning", "Certified Robustness", "Reinforcement Learning", "Poisoning Attack" ], "rating_list": [ 8, 3, 8, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 1, 2, 4 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "X2HnTFsFm8", "primary_area": "applications to robotics, autonomy, planning", "n_author": 10, "author_info_list": [ { "name": "Diming Zhang", "gender": "Male", "institution": "Jiangsu University of Science and Technology", "country": "CN", "position": "Principal Researcher" }, { "name": "Qian Zhang", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Wei Yin", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Xiaotao Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Xiaoyang Guo", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Yao Yao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Yupeng Zheng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "songen gu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Jin Bu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "XIAOXIAO LONG", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 37, "n_ref": 74, "n_ref_all": 83, "n_fig": 4, "n_tab": 5, "L_tab": 1471, "n_element_tab": 108, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 692, "n_element_tab_1": 41, "formula_len_all": 670, "formula_len_all_1": 339, "len_all": 118031, "len_all_1": 59139, "len_abs": 1333, "len_title": 115, "len_sents": 32276, "len_sents_1": 26446, "n_sents": 271, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1338, "title": "OccVAR: Scalable 4D Occupancy Prediction via Next-Scale Prediction", "abs": "In this paper, we propose OCCVAR, a generative occupancy world model that simulates the movement of the ego vehicle and the evolution of the surrounding environment.\nDifferent from visual generation, the occupancy world model should capture the fine-grained 3D geometry and dynamic evolution of the 3D scenes, posing great challenges for the generative models.\nRecent approaches based on autoregression (AR) have demonstrated the potential to predict vehicle movement and future occupancy scenes simultaneously from historical observations, but they typically suffer from the inefficiency and temporal degradation in long-time generation. To holistically address the efficiency and quality issues, we propose a spatial-temporal transformer via temporal next-scale prediction, aiming at predicting the 4D occupancy scenes from coarse to fine scales. To model the dynamic evolution of the scene, we incorporate the ego movement before the tokenized occupancy sequence, enabling the prediction of ego movement and controllable scene generation.\nTo model the fine-grained 3D geometry, OCCVAR utilizes a muitli-scale scene tokenizer to capture the hierarchical information of the 3D scene. \nExperiments show that OCCVAR is capable of high-quality occupancy reconstruction, long-time generation and fast inference speed compared to prior works.", "keywords": [ "Autonomous driving", "World model", "3D generation" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "X1U74IwuxG", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Leander Girrbach", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Massimo Bini", "gender": "unknown", "institution": "Helmholtz Zentrum M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Zeynep Akata", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 17, "n_ref_uni": 31, "n_ref": 71, "n_ref_all": 105, "n_fig": 5, "n_tab": 8, "L_tab": 2197, "n_element_tab": 134, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1559, "n_element_tab_1": 140, "formula_len_all": 706, "formula_len_all_1": 580, "len_all": 132645, "len_all_1": 64448, "len_abs": 1172, "len_title": 101, "len_sents": 37762, "len_sents_1": 28979, "n_sents": 274, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1177, "title": "Decoupling Angles and Strength in Low-rank Adaptation", "abs": "Parameter Efficient FineTuning (PEFT) methods have recently gained extreme popularity thanks to the vast availability of large-scale models, allowing to quickly adapt pretrained models to downstream tasks with minimal computational costs. However, current additive finetuning methods such as LoRA show low robustness to prolonged training and hyperparameter choices, not allowing for optimal out-of-the-box usage. On the other hand, multiplicative and bounded approaches such as ETHER, even if providing higher robustness, only allow for extremely low-rank adaptations and are limited to a fixed-strength transformation, hindering the expressive power of the adaptation. In this work, we propose the DeLoRA finetuning method that first normalizes and then scales the learnable low-rank matrices, thus effectively bounding the transformation strength, which leads to increased hyperparameter robustness at no cost in performance. We show that this proposed approach effectively and consistently improves over popular PEFT methods by evaluating our method on two finetuning tasks, subject-driven image generation and LLM instruction tuning. Code will be released upon acceptance.", "keywords": [ "parameter efficient finetuning", "transfer learning", "computer vision", "NLP" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "X1OfiRYCLn", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Kaipeng Zhang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Shuibo Zhang", "gender": "unknown", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "", "position": "Researcher" }, { "name": "Wenqi Shao", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yi Bin", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Yu Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Ping Luo", "gender": "unknown", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "yue Yang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 30, "n_ref": 62, "n_ref_all": 106, "n_fig": 16, "n_tab": 11, "L_tab": 4826, "n_element_tab": 318, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 4177, "n_element_tab_1": 137, "formula_len_all": 138, "formula_len_all_1": 34, "len_all": 188116, "len_all_1": 62025, "len_abs": 1357, "len_title": 137, "len_sents": 58841, "len_sents_1": 27387, "n_sents": 518, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 87, "L_abs": 1361, "title": "Dynamic Multimodal Evaluation with Flexible Complexity by Vision-Language Bootstrapping", "abs": "Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities across multimodal tasks such as visual perception and reasoning, leading to good performance on various multimodal evaluation benchmarks. However, these benchmarks keep a static nature and overlap with the pre-training data, resulting in fixed complexity constraints and data contamination issues. This raises the concern regarding the validity of the evaluation. To address these two challenges, we introduce a dynamic multimodal evaluation protocol called Vision-Language Bootstrapping (VLB). VLB provides a robust and comprehensive assessment for LVLMs with reduced data contamination and flexible complexity. To this end, VLB dynamically generates new visual question-answering samples through a multimodal bootstrapping module that modifies both images and language, while ensuring that newly generated samples remain consistent with the original ones by a judge module. By composing various bootstrapping strategies, VLB offers dynamic variants of existing benchmarks with diverse complexities, enabling the evaluation to co-evolve with the ever-evolving capabilities of LVLMs. Extensive experimental results across multiple benchmarks, including SEEDBench, MMBench, and MME, show that VLB significantly reduces data contamination and exposes performance limitations of LVLMs.", "keywords": [ "Dynamic Evaluation", "Vision-Language Bootstrapping", "data contamination", "Flexible Complexity", "Large Vision-Language Model" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 1, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "X0ytIvgvxR", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Anastasios Kyrillidis", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Associate Professor" }, { "name": "Fangshuo Liao", "gender": "Male", "institution": "Rice University", "country": "", "position": "PhD student" }, { "name": "Wenyi Su", "gender": "unknown", "institution": "Rice University", "country": "", "position": "Undergrad student" } ], "n_formula": 90, "n_formula_1": 33, "n_ref_uni": 38, "n_ref": 55, "n_ref_all": 79, "n_fig": 10, "n_tab": 4, "L_tab": 214, "n_element_tab": 20, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 2425, "n_element_tab_1": 25, "formula_len_all": 12396, "formula_len_all_1": 2932, "len_all": 200769, "len_all_1": 75426, "len_abs": 1324, "len_title": 122, "len_sents": 40630, "len_sents_1": 29388, "n_sents": 403, "n_sents_1": 242, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1196, "title": "Efficient Distributed Principal Component Analysis with Parallel Deflation", "abs": "We study a distributed Principal Component Analysis (PCA) framework where each worker targets a distinct eigenvector and refines its solution by updating from intermediate solutions provided by peers deemed as \"superior\". Drawing intuition from the delation methods, which is traditionally used in centralized eigenvalue problems, our method breaks the sequential dependency in between the deflation steps and allows asynchronous updates of workers while incurring only a small communication cost. To our knowledge, a critical gap in the literature --*the theoretical underpinning of such distributed, dynamic interactions among workers*-- has remained unaddressed until now. This paper offers the first theoretical analysis explaining why, how, and when these intermediate, hierarchical updates lead to practical and provable convergence in distributed environments. Our theoretical contributions demonstrate that such a distributed PCA algorithm not only converges effectively but does so in a manner that is favorably scalable. We also demonstrate through experiments that our proposed framework offers comparable performance to EigenGame-$\\mu$, the state-of-the-art model-parallel PCA solver.", "keywords": [ "Principal Component Analysis", "Distributed Learning" ], "rating_list": [ 3, 8, 1, 5, 6 ], "soundness_list": [ 4, 3, 1, 2, 4 ], "presentation_list": [ 4, 4, 2, 3, 4 ], "contribution_list": [ 1, 3, 1, 3, 3 ], "confidence_list": [ 3, 3, 5, 3, 3 ] }, { "paper_id": "X0r4BN50Dv", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Chaohao Lin", "gender": "Male", "institution": "Florida International University", "country": "", "position": "PhD student" }, { "name": "Dongsheng Luo", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "Assistant Professor" }, { "name": "Farhad Shirani", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "Assistant Professor" }, { "name": "Wei Cheng", "gender": "Male", "institution": "NEC-Labs", "country": "US", "position": "Researcher" }, { "name": "Wenbo Guo", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "Assistant Professor" }, { "name": "Xu Zheng", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "PhD student" }, { "name": "Zhuomin Chen", "gender": "Female", "institution": "Florida International University", "country": "US", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 6, "n_ref_uni": 54, "n_ref": 113, "n_ref_all": 146, "n_fig": 4, "n_tab": 13, "L_tab": 4999, "n_element_tab": 270, "n_fig_1": 0, "n_tab_1": 6, "L_tab_1": 2560, "n_element_tab_1": 109, "formula_len_all": 1855, "formula_len_all_1": 447, "len_all": 217481, "len_all_1": 79410, "len_abs": 1866, "len_title": 125, "len_sents": 69242, "len_sents_1": 35563, "n_sents": 586, "n_sents_1": 284, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1878, "title": "F-Fidelity: A Robust Framework for Faithfulness Evaluation of Explainable AI", "abs": "Recent research has developed a number of eXplainable AI (XAI) techniques, such as gradient-based approaches, input perturbation-base methods, and black-box explanation methods. While these XAI techniques can extract meaningful insights from deep learning models, how to properly evaluate them remains an open problem. The most widely used approach is to perturb or even remove what the XAI method considers to be the most important features in an input and observe the changes in the output prediction. This approach, although straightforward, suffers the Out-of-Distribution (OOD) problem as the perturbed samples may no longer follow the original data distribution. A recent method RemOve And Retrain (ROAR) solves the OOD issue by retraining the model with perturbed samples guided by explanations. However, using the model retrained based on XAI methods to evaluate these explainers may cause information leakage and thus lead to unfair comparisons. We propose Fine-tuned Fidelity (F-Fidelity), a robust evaluation framework for XAI, which utilizes i) an explanation-agnostic fine-tuning strategy, thus mitigating the information leakage issue, and ii) a random masking operation that ensures that the removal step does not generate an OOD input. We also design controlled experiments with state-of-the-art (SOTA) explainers and their degraded version to verify the correctness of our framework. We conduct experiments on multiple data modalities, such as images, time series, and natural language. The results demonstrate that F-Fidelity significantly improves upon prior evaluation metrics in recovering the ground-truth ranking of the explainers. Furthermore, we show both theoretically and empirically that, given a faithful explainer, F-Fidelity metric can be used to compute the sparsity of influential input components, i.e., to extract the true explanation size.", "keywords": [ "Explainable AI", "Evaluation Metric", "Computer Vision" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 2, 1 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "X0epAjg0hd", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Muthu Chidambaram", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Rong Ge", "gender": "Male", "institution": "Duke University", "country": "", "position": "Associate Professor" } ], "n_formula": 49, "n_formula_1": 23, "n_ref_uni": 58, "n_ref": 110, "n_ref_all": 139, "n_fig": 6, "n_tab": 7, "L_tab": 2472, "n_element_tab": 270, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 237, "n_element_tab_1": 39, "formula_len_all": 4128, "formula_len_all_1": 1903, "len_all": 191399, "len_all_1": 71514, "len_abs": 1411, "len_title": 129, "len_sents": 57232, "len_sents_1": 35079, "n_sents": 437, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1417, "title": "Reassessing How to Compare and Improve the Calibration of Machine Learning Models", "abs": "A machine learning model is calibrated if its predicted probability for an outcome matches the observed frequency for that outcome conditional on the model prediction. This property has become increasingly important as the impact of machine learning models has continued to spread to various domains. As a result, there are now a dizzying number of recent papers on measuring and improving the calibration of (specifically deep learning) models. In this work, we reassess the reporting of calibration metrics in the recent literature. We show that there exist trivial recalibration approaches that can appear seemingly state-of-the-art unless calibration and prediction metrics (i.e. test accuracy) are accompanied by additional generalization metrics such as negative log-likelihood. We then use a calibration-based decomposition of Bregman divergences to develop a new extension to reliability diagrams that jointly visualizes calibration and generalization error, and show how our visualization can be used to detect trade-offs between calibration and generalization. Along the way, we prove novel results regarding the relationship between full calibration error and confidence calibration error for Bregman divergences. We also establish the consistency of the kernel regression estimator for calibration error used in our visualization approach, which generalizes existing consistency results in the literature.", "keywords": [ "calibration", "proper scoring rules", "ECE", "kernel regression" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "X0CxfByJog", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Jiawei Du", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Joey Tianyi Zhou", "gender": "Male", "institution": "A*STAR Centre for Frontier AI Research", "country": "SG", "position": "Principal Researcher" }, { "name": "Ping Liu", "gender": "Male", "institution": "University of Nevada, Reno", "country": "US", "position": "Assistant Professor" }, { "name": "Xin Zhang", "gender": "Female", "institution": "A*STAR", "country": "SG", "position": "Scientist" } ], "n_formula": 14, "n_formula_1": 15, "n_ref_uni": 50, "n_ref": 135, "n_ref_all": 162, "n_fig": 8, "n_tab": 4, "L_tab": 1944, "n_element_tab": 203, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1568, "n_element_tab_1": 165, "formula_len_all": 956, "formula_len_all_1": 1111, "len_all": 136395, "len_all_1": 65325, "len_abs": 1812, "len_title": 139, "len_sents": 37562, "len_sents_1": 29216, "n_sents": 271, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1821, "title": "Breaking Class Barriers: Efficient Dataset Distillation via Inter-Class Feature Compensator", "abs": "Dataset distillation has emerged as a technique aiming to condense informative features from large, natural datasets into a compact and synthetic form. While recent advancements have refined this technique, its performance is bottlenecked by the prevailing class-specific synthesis paradigm. Under this paradigm, synthetic data is optimized exclusively for a pre-assigned one-hot label, creating an implicit class barrier in feature condensation. This leads to inefficient utilization of the distillation budget and oversight of inter-class feature distributions, which ultimately limits the effectiveness and efficiency, as demonstrated in our analysis.\nTo overcome these constraints, this paper presents the Inter-class Feature Compensator (INFER), an innovative distillation approach that transcends the class-specific data-label framework widely utilized in current dataset distillation methods. Specifically, INFER leverages a Universal Feature Compensator (UFC) to enhance feature integration across classes, enabling the generation of multiple additional synthetic instances from a single UFC input. This significantly improves the efficiency of the distillation budget.\nMoreover, INFER enriches inter-class interactions during the distillation, thereby enhancing the effectiveness and generalizability of the distilled data. By allowing for the linear interpolation of labels similar to those in the original dataset, INFER meticulously optimizes the synthetic data and dramatically reduces the size of soft labels in the synthetic dataset to almost zero, establishing a new benchmark for efficiency and effectiveness in dataset distillation. In practice, INFER demonstrates state-of-the-art performance across benchmark datasets. For instance, it outperforms SRe2L by 8.8% on ImageNet-1k in the ipc = 50 setting.", "keywords": [ "Dataset distillation", "inter-class feature compensator (INFER)" ], "rating_list": [ 8, 6, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "WzrkZeDxrM", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Andrew Perrault", "gender": "Male", "institution": "Ohio State University", "country": "", "position": "Assistant Professor" }, { "name": "Jingyi Chen", "gender": "Female", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Micha Elsner", "gender": "unknown", "institution": "Ohio State University", "country": "", "position": "Associate Professor" }, { "name": "JU-SEUNG BYUN", "gender": "Male", "institution": "Ohio State University", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 44, "n_ref": 84, "n_ref_all": 99, "n_fig": 4, "n_tab": 2, "L_tab": 277, "n_element_tab": 52, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 872, "n_element_tab_1": 59, "formula_len_all": 1306, "formula_len_all_1": 1280, "len_all": 125355, "len_all_1": 58683, "len_abs": 1366, "len_title": 153, "len_sents": 36684, "len_sents_1": 27796, "n_sents": 249, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 104, "L_abs": 1330, "title": "DLPO: Diffusion Model Loss-Guided Reinforcement Learning for Fine-Tuning Text-to-Speech Diffusion Models", "abs": "Recent advancements in generative models have sparked a significant interest within the machine learning community. Particularly, diffusion models have demonstrated remarkable capabilities in synthesizing images and speech. Studies such as those by Lee et al. (2023), Black et al. (2023), Wang et al. (2023), and Fan et al. (2024) illustrate that Reinforcement Learning with Human Feedback (RLHF) can enhance diffusion models for image synthesis. However, due to architectural differences between these models and those employed in speech synthesis, it remains uncertain whether RLHF could similarly benefit speech synthesis models. In this paper, we explore the practical application of RLHF to diffusion-based text-to-speech synthesis, leveraging the mean opinion score (MOS) as predicted by UTokyo-SaruLab MOS prediction system (Saeki et al., 2022) as a proxy loss. We introduce diffusion model loss-guided RL policy optimization (DLPO) and compare it against other RLHF approaches, employing the NISQA speech quality and naturalness assessment model (Mittag et al., 2021) and human preference experiments for further evaluation. Our results show that RLHF can enhance diffusion-based text-to-speech synthesis models, and, moreover, DLPO can better improve diffusion models in generating natural and high quality speech audios.", "keywords": [ "Reinforcement Learning", "Speech Synthesis", "Diffusion Models" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 5, 3, 3 ] }, { "paper_id": "WzgcreQaNV", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Shuohuan Wang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Tingfeng Hui", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Weiran Xu", "gender": "Male", "institution": "Beijing University of Post and Telecommunication", "country": "CN", "position": "Associate Professor" }, { "name": "Yu Sun", "gender": "unknown", "institution": "Baidu", "country": "", "position": "Researcher" }, { "name": "Zhenyu Zhang", "gender": "Male", "institution": "Baidu Inc.", "country": "CN", "position": "Researcher" }, { "name": "hua wu", "gender": "Female", "institution": "Baidu", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 55, "n_ref": 90, "n_ref_all": 111, "n_fig": 4, "n_tab": 14, "L_tab": 6017, "n_element_tab": 733, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 3237, "n_element_tab_1": 364, "formula_len_all": 657, "formula_len_all_1": 531, "len_all": 183683, "len_all_1": 69651, "len_abs": 1451, "len_title": 95, "len_sents": 45203, "len_sents_1": 30361, "n_sents": 299, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 47, "L_abs": 1498, "title": "HFT: Half Fine-Tuning for Large Language Models", "abs": "Large language models (LLMs) with one or more fine-tuning phases have become necessary to unlock various capabilities, enabling LLMs to follow natural language instructions and align with human preferences. \nHowever, it carries the risk of catastrophic forgetting during sequential training, the parametric knowledge or the ability learned in previous stages may be overwhelmed by incoming training data. \nThis paper finds that by regularly resetting partial parameters, LLMs can restore some of the original knowledge.\nInspired by this, we introduce \\underline{H}alf \\underline{F}ine-\\underline{T}uning (HFT) for LLMs, as a substitute for full fine-tuning (FFT), to mitigate the forgetting issues, where half of the parameters are selected to learn new tasks. \nIn contrast, the other half are frozen to retain previous knowledge.\nWe provide a feasibility analysis from the perspective of optimization and interpret the parameter selection operation as a regularization term. Without changing the model architecture, HFT could be seamlessly integrated into existing fine-tuning frameworks.\nExtensive experiments and analysis on supervised fine-tuning, direct preference optimization, and continual learning consistently demonstrate the effectiveness, robustness, and efficiency of HFT.\nCompared with FFT, HFT not only significantly alleviates the forgetting problem, but also achieves the best performance in a series of downstream benchmarks, with an approximately 30\\% reduction in training time.", "keywords": [ "Large Language Models", "Catastrophic Forgetting", "Reduction Training Time" ], "rating_list": [ 3, 3, 6, 5, 3 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 5, 4 ] }, { "paper_id": "WzUPae4WnA", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Peijia Qin", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Pengtao Xie", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Ruiyi Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 7, "n_ref_uni": 44, "n_ref": 89, "n_ref_all": 126, "n_fig": 6, "n_tab": 17, "L_tab": 3928, "n_element_tab": 342, "n_fig_1": 3, "n_tab_1": 10, "L_tab_1": 2393, "n_element_tab_1": 187, "formula_len_all": 1038, "formula_len_all_1": 262, "len_all": 160884, "len_all_1": 63601, "len_abs": 1577, "len_title": 121, "len_sents": 42236, "len_sents_1": 27313, "n_sents": 309, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1585, "title": "BiDoRA: Bi-level Optimization-based Weight-Decomposed Low-Rank Adaptation", "abs": "Parameter-efficient fine-tuning (PEFT) of large language models (LLMs) has gained considerable attention as a flexible and efficient way of adapting LLMs to downstream tasks.\nAmong these methods, weighted decomposed low-rank adaptation (DoRA) has emerged as a promising approach.\nDoRA bridges the gap between low-rank adaptation (LoRA) and full fine-tuning (FT) by decomposing the weight matrices into magnitude and direction components, thereby maintaining learning behavior similar to FT.\nAlthough DoRA shows encouraging performance, it introduces additional parameters compared to LoRA, which potentially increases the risk of overfitting.\nMoreover, optimizing magnitude and direction simultaneously leads to a coupled gradient updating pattern for both components, limiting its learning capacity.\nTo overcome these limitations, we propose BiDoRA, a bi-level optimization-based PEFT method.\nIn BiDoRA, the direction and magnitude components are optimized on two distinct datasets at different optimization levels, mitigating the risk of overfitting.\nAdditionally, the asynchronous optimization of the two components promotes their decoupling, allowing for more flexible gradient updates suitable for various downstream tasks.\nEvaluation of BiDoRA on fourteen datasets spanning natural language understanding, natural language generation, and token classification reveals that it significantly outperforms DoRA and other PEFT methods.\nThe superior performance of BiDoRA underscores its effectiveness.\nThe code for BiDoRA is available at https://anonymous.4open.science/r/BiDoRA-5D31.", "keywords": [ "Parameter-Efficient Fine-Tuning", "Bi-level Optimization", "Large Language Model" ], "rating_list": [ 6, 6, 3, 3 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "WzCEiBILHu", "primary_area": "generative models", "n_author": 1, "author_info_list": [ { "name": "Maosheng Yang", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "PhD student" } ], "n_formula": 144, "n_formula_1": 25, "n_ref_uni": 95, "n_ref": 240, "n_ref_all": 270, "n_fig": 5, "n_tab": 9, "L_tab": 1161, "n_element_tab": 125, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 360, "n_element_tab_1": 24, "formula_len_all": 15867, "formula_len_all_1": 1856, "len_all": 324761, "len_all_1": 73409, "len_abs": 1664, "len_title": 88, "len_sents": 96873, "len_sents_1": 34372, "n_sents": 820, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 39, "L_abs": 1766, "title": "Topological Schr\u00f6dinger Bridge Matching", "abs": "Given two boundary distributions, the \\emph{Schr\u00f6dinger Bridge} (SB) problem seeks the \u201cmost likely\u201d random evolution between them with respect to a reference process. \nIt has revealed rich connections to recent machine learning methods for generative modeling and distribution matching. \nWhile these methods perform well in Euclidean domains, they are not directly applicable to topological domains such as graphs and simplicial complexes, which are crucial for data defined over network entities, such as node signals and edge flows.\nIn this work, we propose the \\emph{Topological Schr\u00f6dinger Bridge problem} ($\\mathcal{T}$SBP) for matching signal distributions on a topological domain. \nWe set the reference process to follow some linear tractable \\emph{topology-aware} stochastic dynamics such as topological heat diffusion. \nFor the case of Gaussian boundary distributions, we derive a \\emph{closed-form} topological SB ($\\mathcal{T}$SB) in terms of its time-marginal and stochastic differential. \nIn the general case, leveraging the well-known result, we show that the optimal process follows the forward-backward topological dynamics governed by some unknowns.\nBuilding on these results, we develop $\\mathcal{T}$SB-based models for matching topological signals by parameterizing the unknowns in the optimal process as \\emph{(topological) neural networks} and learning them through \\emph{likelihood training}. We validate the theoretical results and demonstrate the practical applications of $\\mathcal{T}$SB-based models on both synthetic and real-world networks, emphasizing the role of topology. \nAdditionally, we discuss the connections of $\\mathcal{T}$SB-based models to other emerging models, and outline future directions for topological signal matching.", "keywords": [ "Topological Signals", "Schr\u00f6dinger Bridge", "Gaussian Schr\u00f6dinger Bridge", "Bridge Matching" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 2, 4, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 3, 2, 3 ] }, { "paper_id": "WypSbOf9S9", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Sebastian Peitz", "gender": "Male", "institution": "Technische Universit\u00e4t Dortmund", "country": "DE", "position": "Full Professor" }, { "name": "Sedjro Salomon Hotegni", "gender": "Male", "institution": "Technische Universit\u00e4t Dortmund", "country": "DE", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 20, "n_ref_uni": 35, "n_ref": 49, "n_ref_all": 67, "n_fig": 8, "n_tab": 11, "L_tab": 6020, "n_element_tab": 145, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3944, "n_element_tab_1": 68, "formula_len_all": 1346, "formula_len_all_1": 1466, "len_all": 121846, "len_all_1": 60903, "len_abs": 1441, "len_title": 135, "len_sents": 33791, "len_sents_1": 25821, "n_sents": 235, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1445, "title": "MOREL: Enhancing Adversarial Robustness through Multi-Objective Representation Learning", "abs": "Extensive research has shown that deep neural networks (DNNs) are vulnerable to slight adversarial perturbations\u2014small changes to the input data that appear insignificant but cause the model to produce drastically different outputs. In addition to augmenting training data with adversarial examples generated from a specific attack method, most of the current defense strategies necessitate modifying the original model architecture components to improve robustness or performing test-time data purification to handle adversarial attacks. In this work, we demonstrate that strong feature representation learning during training can significantly enhance the original model's robustness. We propose MOREL, a multi-objective feature representation learning approach, encouraging classification models to produce similar features for inputs within the same class, despite perturbations. Our training method involves an embedding space where cosine similarity loss and multi-positive contrastive loss are used to align natural and adversarial features from the model encoder and ensure tight clustering. Concurrently, the classifier is motivated to achieve accurate predictions. Through extensive experiments, we demonstrate that our approach significantly enhances the robustness of DNNs against white-box and black-box adversarial attacks, outperforming other methods that similarly require no architectural changes or test-time data purification.", "keywords": [ "Adversarial robustness", "Representation learning", "Multi-objective optimization", "Deep neural networks" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "WyZT4ZmMzf", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Ansh Soni", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Meenakshi Khosla", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Sudhanshu Srivastava", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Postdoc" }, { "name": "Yiqing Bo", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Undergrad student" } ], "n_formula": 14, "n_formula_1": 15, "n_ref_uni": 44, "n_ref": 61, "n_ref_all": 69, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 967, "n_element_tab_1": 11, "formula_len_all": 948, "formula_len_all_1": 1035, "len_all": 117633, "len_all_1": 54994, "len_abs": 1650, "len_title": 92, "len_sents": 34484, "len_sents_1": 26664, "n_sents": 221, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1583, "title": "Evaluating Representational Similarity Measures from the Lens of Functional Correspondence", "abs": "Neuroscience and artificial intelligence (AI) both face the challenge of interpreting high-dimensional neural data, where the comparative analysis of such data is crucial for revealing shared mechanisms and differences between these complex systems. Despite the widespread use of representational comparisons and the abundance classes of comparison methods, a critical question remains: which metrics are most suitable for these comparisons? While some studies evaluate metrics based on their ability to differentiate models of different origins or constructions (e.g., various architectures), another approach is to assess how well they distinguish models that exhibit distinct behaviors. To investigate this, we examine the degree of alignment between various representational similarity measures and behavioral outcomes, employing group statistics and a comprehensive suite of behavioral metrics for comparison. In our evaluation of eight commonly used representational similarity metrics in the visual domain\u2014spanning alignment-based, CCA-based, inner product kernel-based, and nearest-neighbor methods\u2014we found that metrics like linear CKA and Procrustes, which emphasize the overall geometric structure or shape of representations, excelled in differentiating trained from untrained models and aligning with behavioral measures, whereas metrics such as linear predictivity, commonly used in neuroscience, demonstrated only moderate alignment with behavior. These insights are crucial for selecting metrics that emphasize behaviorally meaningful comparisons in NeuroAI research.", "keywords": [ "Representational Similarity", "Vision", "Deep Neural Networks", "Behavior" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 2, 4, 4 ] }, { "paper_id": "WxqiwbwxiW", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Amaury Gouverneur", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "PhD student" }, { "name": "Borja Rodr\u00edguez G\u00e1lvez", "gender": "Male", "institution": "KTH Royal Institute of Technology, Stockholm, Sweden", "country": "", "position": "PhD student" }, { "name": "Mikael Skoglund", "gender": "Male", "institution": "KTH Royal Institute of Technology, Stockholm, Sweden", "country": "", "position": "Full Professor" }, { "name": "Tobias Oechtering", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 88, "n_formula_1": 33, "n_ref_uni": 14, "n_ref": 62, "n_ref_all": 78, "n_fig": 8, "n_tab": 4, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6715, "formula_len_all_1": 2045, "len_all": 131776, "len_all_1": 60788, "len_abs": 1730, "len_title": 123, "len_sents": 49354, "len_sents_1": 27436, "n_sents": 391, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1406, "title": "An Information-Theoretic Analysis of Thompson Sampling for Logistic Bandits", "abs": "We study the performance of the Thompson Sampling algorithm for logistic bandit problems, where the agent receives binary rewards with probabilities determined by a logistic function, $\\exp(\\beta \\langle a, \\theta \\rangle)/(1+\\exp(\\beta \\langle a, \\theta \\rangle))$, with slope parameter $\\beta$. We focus on the setting where both the action $a$ and parameter $\\theta$ lie within the $d$-dimensional unit ball. Adopting the information-theoretic framework introduced by (Russo & Van Roy, 2015), we analyze the information ratio, a statistic that quantifies the trade-off between the information gained about the optimal action and the immediate regret incurred.\nWe improve upon previous results by establishing that the information ratio is bounded by $\\tfrac{9}{2}d\\alpha^{-2}$, where $\\alpha$ is a minimax measure of the alignment between the action space and the parameter space, independent of $\\beta$. Notably, we derive a regret bound of order $O(d/\\alpha\\sqrt{T \\log(\\beta T/d)})$, which scales only logarithmically with the logistic function parameter $\\smash{\\beta}$. To the best of our knowledge, this is the first regret bound for logistic bandits that achieves logarithmic dependence on $\\beta$ while being independent of the number of actions. In particular, when the action space encompasses the parameter space, \nthe expected regret of Thompson Sampling is of order $\\tilde{O}(d \\sqrt{T})$.", "keywords": [ "multi-armed bandits", "logistic bandits", "information-theory", "Thompson Sampling", "regret bounds", "online optimization" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "WxqWuG431g", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "David D. Baek", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Eric J Michaud", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Joshua Engels", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Max Tegmark", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Xiaoqing Sun", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Yuxiao Li", "gender": "Female", "institution": "Basque Center for Applied Mathematics", "country": "ES", "position": "Postdoc" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 24, "n_ref": 30, "n_ref_all": 51, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 157, "formula_len_all_1": 100, "len_all": 73702, "len_all_1": 51377, "len_abs": 1293, "len_title": 110, "len_sents": 22234, "len_sents_1": 19645, "n_sents": 143, "n_sents_1": 124, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1297, "title": "The Geometry of Concepts: Sparse Autoencoder Feature Structure", "abs": "Sparse autoencoders have recently produced dictionaries of high-dimensional vectors corresponding to the universe of concepts represented by large language models. We find that this concept universe has interesting structure at three levels: 1) The \"atomic\" small-scale structure contains \"crystals\" whose faces are parallelograms or trapezoids, generalizing well-known examples such as (man:woman::king:queen). We find that the quality of such parallelograms and associated function vectors improves greatly when projecting out global distractor directions such as word length, which is efficiently done with linear discriminant analysis. 2) The \"brain\" intermediate-scale structure has significant spatial modularity; for example, math and code features form a \"lobe\" akin to functional lobes seen in neural fMRI images. We quantify the spatial locality of these lobes with multiple metrics and find that clusters of co-occurring features, at coarse enough scale, also cluster together spatially far more than one would expect if feature geometry were random. 3) The \"galaxy\" scale large-scale structure of the feature point cloud is not isotropic, but instead has a power law of eigenvalues with steepest slope in middle layers. We also quantify how the clustering entropy depends on the layer.", "keywords": [ "Mechanistic Interpretability", "Sparse Autoencoder", "Language Model Features", "Clustering", "Multi-scale" ], "rating_list": [ 1, 3, 3, 3, 3 ], "soundness_list": [ 1, 1, 3, 2, 1 ], "presentation_list": [ 1, 2, 2, 2, 1 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 2, 3, 3 ] }, { "paper_id": "Wxl0JMgDoU", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Ashton Anderson", "gender": "unknown", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "Associate Professor" }, { "name": "Difan Jiao", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "George Eilender", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" }, { "name": "Zhenwei Tang", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 15, "n_ref": 35, "n_ref_all": 49, "n_fig": 8, "n_tab": 1, "L_tab": 372, "n_element_tab": 52, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 508, "formula_len_all_1": 420, "len_all": 86240, "len_all_1": 54537, "len_abs": 1115, "len_title": 146, "len_sents": 32209, "len_sents_1": 26380, "n_sents": 232, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1321, "title": "Understanding Skill Adaptation in Transformers Using Sparse Autoencoders: Chess as a Model System", "abs": "Understanding how skill shapes decision-making in complex environments is a challenging problem in AI interpretability. We investigate this question by applying Sparse Autoencoders (SAEs) to the internal representations of Maia-2, a human-like chess model that simulates human play across varying skill levels. Maia-2 incorporates a skill-aware transformer that integrates position features with categorical skill inputs, capturing nuanced relationships between player expertise and move selection. By training SAEs on these modulated representations, we identify latent features that reveal how the model's threat response policy adapts to different levels of play. We then use these features to intervene on the internal activations of Maia-2, eliciting both higher skill and lower skill play in specific contexts. We also apply mediated intervention with targeted SAE features to effectively enhance and sabotage the model's understanding and decision-making on context-specific chess tasks. Our findings suggest that SAE features can help shed light on how skill-specific information is encoded within a model to produce human-like behavior, and that these insights can be applied to steer the model's performance on specific sub-tasks. Our work is available at \\url{https://anonymous.4open.science/r/chess-sae-3C06/}", "keywords": [ "Skill Adaptation", "Chess", "Sparse Autoencoders", "Mechanistic Interpretability" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "WxSkpPeef3", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Benedict Lee", "gender": "unknown", "institution": "HP Development Company, L.P.", "country": "SG", "position": "Data Scientist" }, { "name": "Daniel Hui Loong Ng", "gender": "unknown", "institution": "HP Inc", "country": "SG", "position": "Researcher" }, { "name": "Huang Liang", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "KELIN XIA", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 41, "n_ref": 79, "n_ref_all": 99, "n_fig": 5, "n_tab": 8, "L_tab": 2818, "n_element_tab": 250, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1408, "n_element_tab_1": 104, "formula_len_all": 1270, "formula_len_all_1": 1342, "len_all": 142189, "len_all_1": 64014, "len_abs": 1910, "len_title": 139, "len_sents": 38433, "len_sents_1": 30897, "n_sents": 252, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1921, "title": "Contrastive Learning with Simplicial Convolutional Networks for Short-Text Classification", "abs": "Text classification is a fundamental task in Natural Language Processing (NLP). Short text classification has recently captured much attention due to its increased amount from various sources with limited labels and its inherent challenges for its sparsity in words and semantics. Recent studies have adopted self-supervised contrastive learning across different representations to improve performance. However, most of the current models face several challenges. Firstly, the augmentation step might not be able to generate positive and negative samples that are semantically similar and dissimilar to the anchor respectively. Secondly, the text data could be enhanced with external auxiliary information that might introduce noise to the sparse text data. In addition, they are limited in capturing higher-order information such as group-wise interactions. In this work, we propose a novel document simplicial complex construction based on text data for a higher-order message-passing mechanism. We develop a simplicial complex representation for text sentences based on the directed word co-occurrence. Novel features are proposed for 0-simplex (word), 1-simplex (word-pair), and 2-simplex (three consecutive words) to characterise intrinsic higher-order structural information among words. We also enhance the short text classification performance by contrasting the structural representation with the sequential representation generated by the transformer mechanism for improved outcomes. The proposed framework, Contrastive Learning with Simplicial Convolutional Networks (C-SCN), leverages the expressive power of graph neural networks, models higher-order information beyond pair-wise relations and enriches features through contrastive learning. Experimental results on four benchmark datasets demonstrate the capability of C-SCN to outperform existing models in analysing sequential and complex short-text data.", "keywords": [ "Topological Deep Learning", "Contrastive Leaning" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "WxLwXyBJLw", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Ivan Oseledets", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "Svetlana Pavlova", "gender": "Female", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Gleb Vladimirovich Ryzhakov", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Principal Researcher" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 23, "n_ref": 49, "n_ref_all": 60, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 629, "formula_len_all_1": 629, "len_all": 84553, "len_all_1": 62449, "len_abs": 999, "len_title": 83, "len_sents": 23900, "len_sents_1": 23573, "n_sents": 186, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1002, "title": "Flow Matching for One-Step Sampling", "abs": "Flow-based generative models have rapidly advanced as a method for mapping simple distributions to complex ones for which the distribution function is unknown. By leveraging continuous-time stochastic processes, these models offer a powerful framework for density estimation, i.e. an algorithm that samples new points based only on existing samples. However, their requirement of solving ordinary differential equations (ODEs) during sampling process incurs substantial computational costs, particularly for large amount of data and numerous time points. This paper proposes a novel solution, which is based on a theoretical analysis of Flow Matching (FM), to overcome this bottleneck, namely, we developed an algorithm to find the point prototype for a given point from the target distribution. By eliminating the need for ODE solvers, our method significantly accelerates sampling while preserving model performance. Numerical experiments validate the proposed approach, demonstrating its efficiency.", "keywords": [ "Flow Matching", "Generative Models", "Ordinary Differential Equations", "One-step generation" ], "rating_list": [ 3, 1, 3, 6 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 2, 1, 1, 3 ], "confidence_list": [ 3, 4, 2, 2 ] }, { "paper_id": "WxKS3XV7Z4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Haoze Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Long Peng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Renjing Pei", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Engineer" }, { "name": "Wenbo Li", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xin Di", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Yang Cao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yang Wang", "gender": "Male", "institution": "Chang'an University", "country": "CN", "position": "Associate Professor" }, { "name": "Yong Li", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zheng-Jun Zha", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jeremie Guo", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 52, "n_ref": 83, "n_ref_all": 116, "n_fig": 10, "n_tab": 9, "L_tab": 2707, "n_element_tab": 454, "n_fig_1": 8, "n_tab_1": 8, "L_tab_1": 2922, "n_element_tab_1": 424, "formula_len_all": 267, "formula_len_all_1": 268, "len_all": 148999, "len_all_1": 60457, "len_abs": 1416, "len_title": 139, "len_sents": 31224, "len_sents_1": 24376, "n_sents": 222, "n_sents_1": 162, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1422, "title": "Boosting Real-World Super-Resolution with RAW Data: a New Perspective, Dataset and Baseline", "abs": "Real-world image super-resolution (Real SR) aims to generate high-fidelity, detail-rich high-resolution (HR) images from low-resolution (LR) counterparts. Existing Real SR methods primarily focus on processing within the RGB domain. In this paper, we pioneer the use of detail-rich RAW data to complement RGB-only Real SR, specifically by utilizing both LR RGB and RAW inputs to generate superior HR RGB outputs. We argue that key image processing steps in Image Signal Processing, such as denoising and demosaicing, inherently result in the loss of fine details, making RAW data a valuable information source. To validate this, we present RealSR-RAW, a comprehensive dataset comprising 10,000 pairs with LR and HR RGB images, along with corresponding LR RAW data, captured across multiple smartphones under varying focal lengths and diverse scenes. Additionally, we propose a novel, general RAW adapter to efficiently integrate RAW data into existing CNNs, Transformers, and Diffusion-based Real SR models by suppressing the noise contained in RAW and aligning distribution. Extensive experiments demonstrate that incorporating RAW data significantly enhances detail recovery and improves Real SR performance across ten evaluation metrics, including both fidelity and perception-oriented metrics. Our findings open a new direction for the Real SR task, with the dataset and code made available to support future research.", "keywords": [ "Real-World Super-Resolution", "RAW Data" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "WwwJfkGq0G", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Bahram Zonooz", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "Elahe Arani", "gender": "Female", "institution": "Eindhoven University of technology ", "country": "NL", "position": "Assistant Professor" }, { "name": "Fahad Sarfraz", "gender": "Male", "institution": "TomTom", "country": "NL", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 38, "n_ref": 51, "n_ref_all": 73, "n_fig": 6, "n_tab": 10, "L_tab": 5913, "n_element_tab": 520, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2982, "n_element_tab_1": 190, "formula_len_all": 428, "formula_len_all_1": 373, "len_all": 153019, "len_all_1": 65653, "len_abs": 1100, "len_title": 108, "len_sents": 52451, "len_sents_1": 30712, "n_sents": 330, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1099, "title": "Semantic Aware Representation Learning for Lifelong Learning", "abs": "The human brain excels at lifelong learning by not only encoding information in sparse activation codes but also leveraging rich semantic structures and relationships between newly encountered and previously learned objects. This ability to utilize semantic similarities is crucial for efficient learning and knowledge consolidation, yet is often underutilized in current continual learning approaches. To bridge this gap, we propose Semantic-Aware Representation Learning (SARL) which employs sparse activations and a principled approach to evaluate similarities between objects encountered across different tasks and subsequently uses them to guide representation learning. Using these relationships, SARL enhances the reusability of features and reduces interference between tasks. This approach empowers the model to adapt to new information while maintaining stability, significantly improving performance in complex incremental learning scenarios. Our analysis demonstrates that SARL achieves a superior balance between plasticity and stability by harnessing the underlying semantic structure.", "keywords": [ "lifelong learning", "continual learning", "sparsity", "semantic relationships", "sparse coding" ], "rating_list": [ 6, 3, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 2, 5, 4, 4 ] }, { "paper_id": "WwpYSOkkCt", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Hrayr Harutyunyan", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Sangmin Bae", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Seungyeon Kim", "gender": "unknown", "institution": "Google DeepMind", "country": "US", "position": "Research Engineer (L7)" }, { "name": "Tal Schuster", "gender": "Not Specified", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Ziwei Ji", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Adam Fisch", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 64, "n_ref": 142, "n_ref_all": 202, "n_fig": 21, "n_tab": 6, "L_tab": 5465, "n_element_tab": 360, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 623, "n_element_tab_1": 121, "formula_len_all": 457, "formula_len_all_1": 370, "len_all": 309126, "len_all_1": 63051, "len_abs": 1417, "len_title": 128, "len_sents": 80645, "len_sents_1": 31216, "n_sents": 540, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1431, "title": "Relaxed Recursive Transformers: Effective Parameter Sharing with Layer-wise LoRA", "abs": "Large language models (LLMs) are expensive to deploy. Parameter sharing offers a possible path towards reducing their size and cost, but its effectiveness in modern LLMs remains fairly limited. In this work, we revisit \"layer tying\" as form of parameter sharing in Transformers, and introduce novel methods for converting existing LLMs into smaller \"Recursive Transformers\" that share parameters across layers, with minimal loss of performance. Here, our Recursive Transformers are efficiently initialized from standard pretrained Transformers, but only use a single block of unique layers that is then repeated multiple times in a loop. We further improve performance by introducing Relaxed Recursive Transformers that add flexibility to the layer tying constraint via depth-wise low-rank adaptation (LoRA) modules, yet still preserve the compactness of the overall model. We show that our recursive models (e.g., recursive Gemma 1B) outperform both similar-sized vanilla pretrained models (such as TinyLlama 1.1B and Pythia 1B) and knowledge distillation baselines---and can even recover most of the performance of the original \"full-size\" model (e.g., Gemma 2B with no shared parameters). Finally, we propose Continuous Depth-wise Batching, a promising new inference paradigm enabled by the Recursive Transformer when paired with early exiting, which we show to theoretically lead to significant (2-3$\\times$) throughput gains.", "keywords": [ "Efficient LLM", "Parameter Sharing", "Relaxed Recursive Transformer", "Continuous Depth-wise Batching", "Early-Exiting" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 5, 5, 3, 4 ] }, { "paper_id": "WwmtcGr4lP", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Aishwarya Jayagopal", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Anand D Jeyasekharan", "gender": "unknown", "institution": "National University of Singapore", "country": "", "position": "Assistant Professor" }, { "name": "Robert John Walsh", "gender": "Male", "institution": "National University Cancer Institute, Singapore", "country": "", "position": "Associate Consultant" }, { "name": "Tuan Zea Tan", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Vaibhav Rajan", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Assistant Professor" }, { "name": "Yanrong Zhang", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 67, "n_ref": 102, "n_ref_all": 130, "n_fig": 8, "n_tab": 10, "L_tab": 3290, "n_element_tab": 218, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1704, "n_element_tab_1": 76, "formula_len_all": 1568, "formula_len_all_1": 1053, "len_all": 199344, "len_all_1": 69822, "len_abs": 1359, "len_title": 170, "len_sents": 48000, "len_sents_1": 31898, "n_sents": 376, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 121, "L_abs": 1364, "title": "GANDALF: Generative AttentioN based Data Augmentation and predictive modeLing Framework for personalized cancer treatment", "abs": "Effective treatment of cancer is a major challenge faced by healthcare providers, due to the highly individualized nature of patient responses to treatment. This is caused by the heterogeneity seen in cancer-causing alterations (mutations) across patient genomes. Limited availability of response data in patients makes it difficult to train personalized treatment recommendation models on mutations from clinical genomic sequencing reports. Prior methods tackle this by utilising larger, labelled pre-clinical laboratory datasets (\u2018cell lines\u2019), via transfer learning. These methods augment patient data by learning a shared, domain-invariant representation, between the cell line and patient domains, which is then used to train a downstream drug response prediction (DRP) model. This approach augments data in the shared space but fails to model patient-specific characteristics, which have a strong influence on their drug response. We propose a novel generative attention-based data augmentation and predictive modeling framework, GANDALF, to tackle this crucial shortcoming of prior methods. GANDALF not only augments patient genomic data directly, but also accounts for its domain-specific characteristics. GANDALF outperforms state-of-the-art DRP models on publicly available patient datasets and emerges as the front-runner amongst SOTA cancer DRP models.", "keywords": [ "personalized drug response prediction", "cancer", "genomic data augmentation", "diffusion model", "pseudolabelling" ], "rating_list": [ 8, 6, 8, 6, 6 ], "soundness_list": [ 4, 2, 4, 3, 2 ], "presentation_list": [ 3, 2, 4, 2, 2 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 3, 3 ] }, { "paper_id": "WwQdcQROmb", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 11, "author_info_list": [ { "name": "Haodong Feng", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Long Wei", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Postdoc" }, { "name": "Peiyan Hu", "gender": "Not Specified", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Rui Wang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Ruiqi Feng", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Tailin Wu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Wenhao Deng", "gender": "Not Specified", "institution": "Westlake University", "country": "", "position": "Researcher" }, { "name": "Xiaowei Qian", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "", "position": "Undergrad student" }, { "name": "Yue Wang", "gender": "Male", "institution": "Microsoft Research Aisa", "country": "CN", "position": "Researcher" }, { "name": "Zhi-Ming Ma", "gender": "unknown", "institution": "Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Tao Zhang", "gender": "Male", "institution": "Wuhan University", "country": "", "position": "Undergrad student" } ], "n_formula": 33, "n_formula_1": 28, "n_ref_uni": 41, "n_ref": 91, "n_ref_all": 114, "n_fig": 4, "n_tab": 8, "L_tab": 2405, "n_element_tab": 259, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 223, "n_element_tab_1": 39, "formula_len_all": 1925, "formula_len_all_1": 1753, "len_all": 151636, "len_all_1": 65282, "len_abs": 1365, "len_title": 123, "len_sents": 45733, "len_sents_1": 29935, "n_sents": 347, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1372, "title": "A Probabilistic Generative Method for Safe Physical System Control Problems", "abs": "Controlling complex physical systems is a crucial task in science and engineering, often requiring the balance of control objectives and safety constraints. Recently, diffusion models have demonstrated a strong ability to model high-dimensional state spaces, giving them an advantage over recent deep learning and reinforcement learning-based methods in complex control tasks. However, they do not inherently address safety concerns. In contrast, while safe reinforcement learning methods consider safety, they typically fail to provide guarantees for satisfying safety constraints.\nTo address these limitations, we propose Safe Conformal Physical system control (SafeConPhy), which optimizes the diffusion model with a provable safety bound iteratively to satisfy the safety constraint.\nWe pre-train a diffusion model on the training set. Given the calibration set and the specific control targets, we derive a provable safety bound using conformal prediction. After iteratively enhancing the safety of the diffusion model with the progressively updated bound, the model's output can be certified as safe with a user-defined probability.\nWe evaluate our algorithm on two control tasks: 1D Burgers' equation and 2D incompressible fluid. Our results show that our algorithm satisfies safety constraints, and outperforms prior control methods and safe offline RL algorithms.", "keywords": [ "safe PDE control", "physical systems", "generative models", "conformal prediction", "fine-tuning" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "Ww599CnVnU", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Heung-Il Suk", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Full Professor" }, { "name": "Jee Seok Yoon", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" }, { "name": "Kwanseok Oh", "gender": "Male", "institution": "Korea University", "country": "", "position": "PhD student" }, { "name": "Sangmin Jo", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "MS student" }, { "name": "Wootaek Jeong", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" } ], "n_formula": 46, "n_formula_1": 17, "n_ref_uni": 38, "n_ref": 86, "n_ref_all": 128, "n_fig": 7, "n_tab": 3, "L_tab": 484, "n_element_tab": 68, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 1725, "n_element_tab_1": 123, "formula_len_all": 3408, "formula_len_all_1": 1076, "len_all": 145954, "len_all_1": 60031, "len_abs": 1579, "len_title": 146, "len_sents": 42195, "len_sents_1": 27258, "n_sents": 317, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1586, "title": "Multi-scale Minimal Sufficient Representation Learning for Domain Generalization in Sleep Staging", "abs": "Deep learning-based automatic sleep staging demonstrates strong performance as a promising solution for diagnosing sleep disorders. However, deep learning models often struggle to generalize on unseen subjects due to variability in physiological signals, resulting in degraded performance in out-of-distribution scenarios. To address this issue, domain generalization approaches have recently been studied actively to ensure generalized performance on unseen domains during the training. Among those techniques, contrastive learning has proven its validity in learning domain-invariant features by aligning samples of the same class across different domains. Despite its potential, many existing methods are insufficient for extracting truly domain-invariant representations, as they do not explicitly reduce domain-relevant information embedded in the features. In this paper, we argue that addressing superfluous information is a key to bridging the domain gap. Furthermore, existing methods often neglect the multi-scale nature of sleep signals, potentially missing important temporal and spectral characteristics. To address these limitations, we propose a novel Multi-Scale Minimal Sufficient representation learning (MSMS) framework, which effectively reduces domain-relevant information while preserving essential temporal and spectral features for sleep stage classification. We evaluate our method on publicly available sleep staging benchmark datasets, SleepEDF-20 and MASS. Experimental results demonstrate that our approach consistently outperforms state-of-the-art methods.", "keywords": [ "Domain generalization", "Multi-scale", "Contrastive learning", "Information theory", "Sleep staging" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 5, 5 ] }, { "paper_id": "Wvi8c0tgvt", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Hana Lee", "gender": "Female", "institution": "Samsung", "country": "KR", "position": "Researcher" }, { "name": "Hyong-Euk Lee", "gender": "Male", "institution": "Samsung Advanced Institute of Technology", "country": "KR", "position": "Principal Researcher" }, { "name": "Insoo Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Jinwoo Shin", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 18, "n_ref_uni": 34, "n_ref": 130, "n_ref_all": 178, "n_fig": 18, "n_tab": 9, "L_tab": 1771, "n_element_tab": 160, "n_fig_1": 13, "n_tab_1": 6, "L_tab_1": 1451, "n_element_tab_1": 124, "formula_len_all": 1625, "formula_len_all_1": 1365, "len_all": 161233, "len_all_1": 72287, "len_abs": 1943, "len_title": 116, "len_sents": 52772, "len_sents_1": 31287, "n_sents": 404, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1588, "title": "Controllable Blur Data Augmentation Using 3D-Aware Motion Estimation", "abs": "Existing realistic blur datasets provide insufficient variety in scenes and blur patterns to be trained, while expanding data diversity demands considerable time and effort due to complex dual-camera systems. To address the challenge, data augmentation can be an effective way to artificially increase data diversity. However, existing methods on this line are typically designed to estimate motions from a 2D perspective, e.g., estimating 2D non-uniform kernels disregarding 3D aspects of blur modeling, which leads to unrealistic motion patterns due to the fact that camera and object motions inherently arise in 3D space. In this paper, we propose a 3D-aware blur synthesizer capable of generating diverse and realistic blur images for blur data augmentation. Specifically, we estimate 3D camera positions within the motion blur interval, generate the corresponding scene images, and aggregate them to synthesize a realistic blur image. Since the 3D camera positions projected onto the 2D image plane inherently lie in 2D space, we can represent the 3D transformation as a combination of 2D transformation and projected 3D residual component. This allows for 3D transformation without requiring explicit depth measurements, as the 3D residual component is directly estimated via a neural network. Furthermore, our blur synthesizer allows for controllable blur data augmentation by modifying blur magnitude, direction, and scenes, resulting in diverse blur images. As a result, our method significantly improves deblurring performance, making it more practical for real-world scenarios.", "keywords": [ "Blur synthesis", "Data augmentation", "Blind motion deblurring", "3D motion modeling" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 2, 4, 2 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "Wv9Gl1bFbc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Haotian Yu", "gender": "Male", "institution": "Case Western Reserve University", "country": "US", "position": "PhD student" }, { "name": "Pan Li", "gender": "unknown", "institution": "Case Western Reserve University", "country": "US", "position": "Associate Professor" }, { "name": "Runchao Li", "gender": "Male", "institution": "Case Western Reserve University", "country": "US", "position": "PhD student" }, { "name": "Xianxuan Long", "gender": "Female", "institution": "Case Western Reserve University", "country": "US", "position": "PhD student" }, { "name": "Yao Fu", "gender": "Male", "institution": "Case Western Reserve University", "country": "US", "position": "PhD student" }, { "name": "Yu Yin", "gender": "Female", "institution": "Case Western Reserve University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 89, "n_ref": 228, "n_ref_all": 267, "n_fig": 9, "n_tab": 5, "L_tab": 5002, "n_element_tab": 365, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 620, "n_element_tab_1": 84, "formula_len_all": 636, "formula_len_all_1": 638, "len_all": 230752, "len_all_1": 55616, "len_abs": 1614, "len_title": 138, "len_sents": 53258, "len_sents_1": 24224, "n_sents": 361, "n_sents_1": 158, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1626, "title": "Dynamic Self-Distillation via Previous Mini-batches for Fine-tuning Small Language Models", "abs": "Knowledge Distillation (KD) has become a widely adopted approach for compressing large language models (LLMs) to reduce computational costs and memory footprint. However, the availability of complex teacher models is a prerequisite for running most KD pipelines. Thus, the traditional KD procedure can be unachievable or budget-unfriendly, particularly when relying on commercial LLMs like GPT4. In this regard, Self Distillation (SelfD) emerges as an advisable alternative, enabling student models to learn without teachers' guidance. Nonetheless, existing SelfD approaches for LMs often involve architectural modifications, assuming the models are open-source, which may not always be practical. In this work, we introduce a model-agnostic and task-agnostic method named dynamic SelfD from the previous mini-batch (DynSDPB), which realizes current iterations\u2019 distillation from the last ones\u2019 generated logits. Additionally, to address prediction inaccuracies during the early iterations, we dynamically adjust the distillation influence and temperature values to enhance the adaptability of fine-tuning. Furthermore, we propose Vocabulary Map Matching (VMM), aiming to address output inconsistency for auto-regressive LLMs. Last but not least, DynSDPB facilitates the seamless integration of existing self-correction and self-training techniques for small language models (SLMs). We apply DynSDPB to both encoder-only LMs (e.g., BERT model families) and decoder-only LMs (e.g., LLaMA model families), validating its effectiveness across natural language understanding (NLU) and natural language generation (NLG) benchmarks.", "keywords": [ "Knowledge Distillation", "Self-Distillation", "Pre-trained Language Models", "Large Language Models", "Small Language Models", "NLP", "Fine-tuning" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 1, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "WuTczPV8WC", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Ernest K. Ryu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "Jongmin Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" } ], "n_formula": 167, "n_formula_1": 36, "n_ref_uni": 81, "n_ref": 157, "n_ref_all": 158, "n_fig": 0, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 269, "n_element_tab_1": 11, "formula_len_all": 29533, "formula_len_all_1": 2457, "len_all": 248727, "len_all_1": 68355, "len_abs": 744, "len_title": 119, "len_sents": 55682, "len_sents_1": 28351, "n_sents": 660, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 767, "title": "Optimal Non-Asymptotic Rates of Value Iteration for Average-Reward Markov Decision Processes", "abs": "While there is an extensive body of research on the analysis of Value Iteration (VI) for discounted cumulative-reward MDPs, prior work on analyzing VI for (undiscounted) average-reward MDPs has been limited, and most prior results focus on asymptotic rates in terms of Bellman error. In this work, we conduct refined non-asymptotic analyses of average-reward MDPs, obtaining a collection of convergence results advancing our understanding of the setup. Among our new results, most notable are the $\\mathcal{O}(1/k)$-rates of Anchored Value Iteration on the Bellman error under the multichain setup and the span-based complexity lower bound that matches the $\\mathcal{O}(1/k)$ upper bound up to a constant factor of $8$ in the weakly communicating and unichain setups.", "keywords": [ "Average-reward Markov decision process", "Average-reward MDP", "Value iteration", "Reinforcement learning theory", "RL theroy", "Dynamic programming", "Convergence analysis", "Anchoring mechanism" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "WttfQGwpES", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Shi Fu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Xinmei Tian", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yingjie Wang", "gender": "Male", "institution": "China University of Petroleum", "country": "CN", "position": "Lecturer" }, { "name": "Yuzhu Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 135, "n_formula_1": 19, "n_ref_uni": 31, "n_ref": 108, "n_ref_all": 120, "n_fig": 1, "n_tab": 2, "L_tab": 846, "n_element_tab": 12, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 16506, "formula_len_all_1": 1528, "len_all": 209331, "len_all_1": 78406, "len_abs": 974, "len_title": 128, "len_sents": 69651, "len_sents_1": 37147, "n_sents": 622, "n_sents_1": 276, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 978, "title": "A Theoretical Perspective: When and How Self-consuming Training Loops Generalize", "abs": "High-quality data is essential for training large generative models, yet the vast reservoir of real data available online has become nearly depleted. Consequently, models increasingly generate their own data for further training, forming Self-consuming Training Loops (STLs). However, the empirical results have been strikingly inconsistent: some models degrade or even collapse, while others successfully avoid these failures, leaving a significant gap in theoretical understanding to explain this discrepancy. This paper introduces the intriguing notion of *recursive stability* and presents the first theoretical generalization analysis, revealing how both model architecture and the proportion between real and synthetic data influence the success of STLs. We further extend this analysis to transformers in in-context learning, showing that even a constant-sized proportion of real data ensures convergence, while also providing insights into optimal synthetic data sizing.", "keywords": [ "Generative Models", "Synthetic Data", "Transformer", "Generalization Error", "Learning Theory" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 4, 2 ], "presentation_list": [ 2, 4, 2 ], "contribution_list": [ 2, 4, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "Wto5U7q6I2", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 15, "author_info_list": [ { "name": "Bocheng Zou", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Undergrad student" }, { "name": "Fangrui Zhu", "gender": "Female", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Feng Yao", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Jianfeng Gao", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Jianrui Zhang", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "MS student" }, { "name": "Jing Gu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Kai Zhang", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Mu Cai", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin, Madison", "country": "US", "position": "PhD student" }, { "name": "Reuben Tan", "gender": "Male", "institution": "Research, Microsoft", "country": "US", "position": "Researcher" }, { "name": "Yao Dou", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yiwu Zhong", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Yuzhang Shang", "gender": "Male", "institution": "Illinois Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jaeseung Park", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Jianwei Yang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Yong Jae Lee", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Associate Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 54, "n_ref": 126, "n_ref_all": 145, "n_fig": 5, "n_tab": 8, "L_tab": 5996, "n_element_tab": 1051, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 3113, "n_element_tab_1": 492, "formula_len_all": 297, "formula_len_all_1": 129, "len_all": 206900, "len_all_1": 76752, "len_abs": 1187, "len_title": 92, "len_sents": 39854, "len_sents_1": 32407, "n_sents": 285, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1205, "title": "TemporalBench: Towards Fine-grained Temporal Understanding for Multimodal Video Models", "abs": "Understanding fine-grained temporal dynamics is crucial for video understanding. Yet, popular video benchmarks, such as MSRVTT and TGIF, often fail to effectively evaluate AI models' temporal reasoning abilities due to the lack of fine-grained temporal annotations. \nAs a result, text-based models, leveraging strong language priors, often perform comparably to video models, and image-trained models have been reported to outperform their video-trained counterparts on MSRVTT and TGIF. This paper introduces a new TemporalBench benchmark for fine-grained temporal event understanding in videos. TemporalBench, sourced from a diverse video datasets, consists of $\\sim$10K pairs of video description questions, derived from $\\sim$2K high-quality human-annotated video captions. Uniquely, our benchmark provides fine-grained temporal annotations to evaluate models' temporal reasoning abilities. Our results show that state-of-the-art models like GPT-4o achieve only 38.0\\% multiple binary QA accuracy on TemporalBench, demonstrating a significant human-AI gap in temporal understanding. We hope that TemporalBench is instrumental to fostering research on improving models' temporal reasoning capabilities.", "keywords": [ "video", "benchmark", "multimodel" ], "rating_list": [ 3, 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "WtZRZC4zva", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Haoteng Yin", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Pan Li", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Rongzhe Wei", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "I Chien", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 18, "n_ref_uni": 54, "n_ref": 113, "n_ref_all": 137, "n_fig": 3, "n_tab": 12, "L_tab": 5206, "n_element_tab": 407, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1342, "n_element_tab_1": 144, "formula_len_all": 3370, "formula_len_all_1": 7654, "len_all": 201532, "len_all_1": 88083, "len_abs": 1431, "len_title": 88, "len_sents": 50697, "len_sents_1": 38519, "n_sents": 284, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1438, "title": "Privately Learning from Graphs with Applications in Fine-tuning Large Pretrained Models", "abs": "Graphs offer unique insights into relationships and interactions between entities, complementing data modalities like text, images, and videos. By incorporating relational information from graph data, AI models can extend their capabilities beyond traditional tasks. However, relational data in sensitive domains such as finance and healthcare often contain private information, making privacy preservation crucial. Existing privacy-preserving methods, such as DP-SGD, which rely on gradient decoupling assumptions, are not well-suited for relational learning due to the inherent dependencies between coupled training samples. To address this challenge, we propose a privacy-preserving relational learning pipeline that decouples dependencies in sampled relations during training, ensuring differential privacy through a tailored application of DP-SGD. We apply this method to fine-tune large language models (LLMs) on sensitive graph data, and tackle the associated computational complexities. Our approach is evaluated on LLMs of varying sizes (e.g., BERT, Llama2) using real-world relational data from four text-attributed graphs. The results demonstrate significant improvements in relational learning tasks, all while maintaining robust privacy guarantees during training. Additionally, we explore the trade-offs between privacy, utility, and computational efficiency, offering insights into the practical deployment of our approach.", "keywords": [ "relational learning", "differential privacy", "language model", "fine-tuning" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "WtGpBjnMq8", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Weichuan Zhang", "gender": "Male", "institution": "Shanxi University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xiaohan Yu", "gender": "Male", "institution": "Macquarie University", "country": "AU", "position": "Assistant Professor" }, { "name": "Yongsheng Gao", "gender": "unknown", "institution": "Griffith University", "country": "AU", "position": "Full Professor" }, { "name": "Zicheng Pan", "gender": "Male", "institution": "Griffith University", "country": "AU", "position": "Postdoc" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 35, "n_ref": 71, "n_ref_all": 87, "n_fig": 6, "n_tab": 12, "L_tab": 4459, "n_element_tab": 572, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 957, "n_element_tab_1": 166, "formula_len_all": 626, "formula_len_all_1": 625, "len_all": 146737, "len_all_1": 60677, "len_abs": 1762, "len_title": 150, "len_sents": 43603, "len_sents_1": 29033, "n_sents": 311, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 1770, "title": "Propensity-driven Uncertainty Learning for Sample Exploration in Source-Free Active Domain Adaptation", "abs": "Source-free active domain adaptation (SFADA) addresses the challenge of adapting a pre-trained model to new domains without access to source data while minimizing the need for target domain annotations. This scenario is particularly relevant in real-world applications where data privacy, storage limitations, or labeling costs are significant concerns. Key challenges in SFADA include selecting the most informative samples from the target domain for labeling, effectively leveraging both labeled and unlabeled target data, and adapting the model without relying on source domain information. Additionally, existing methods often struggle with noisy or outlier samples and may require impractical progressive labeling during training. To effectively select more informative samples without frequently requesting human annotations, we propose the Propensity-driven Uncertainty Learning (ProULearn) framework. ProULearn utilizes a novel homogeneity propensity estimation mechanism combined with correlation index calculation to evaluate feature-level relationships. This approach enables the identification of representative and challenging samples while avoiding noisy outliers. Additionally, we develop a central correlation loss to refine pseudo-labels and create compact class distributions during adaptation. In this way, ProULearn effectively bridges the domain gap and maximizes adaptation performance. The principles of informative sample selection underlying ProULearn have broad implications beyond SFADA, offering benefits across various deep learning tasks where identifying key data points or features is crucial. Extensive experiments on four benchmark datasets demonstrate that ProULearn outperforms state-of-the-art methods in domain adaptation scenarios.", "keywords": [ "Active learning", "Source-free domain adaptation" ], "rating_list": [ 6, 6, 5, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "WszeEzjcq2", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Chi-Guhn Lee", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Full Professor" }, { "name": "Eldan Cohen", "gender": "unknown", "institution": "University of Toronto", "country": "CA", "position": "Assistant Professor" }, { "name": "Jerry Sun", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Zhixiao Xiong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 49, "n_ref": 142, "n_ref_all": 157, "n_fig": 4, "n_tab": 2, "L_tab": 275, "n_element_tab": 34, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 571, "n_element_tab_1": 6, "formula_len_all": 636, "formula_len_all_1": 469, "len_all": 164434, "len_all_1": 63772, "len_abs": 2280, "len_title": 144, "len_sents": 55931, "len_sents_1": 31689, "n_sents": 374, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1252, "title": "What's Wrong With Non-Autoregressive Graph Neural Networks in Neural Combinatorial Optimization", "abs": "Neural combinatorial optimization (NCO) leverages machine learning models to tackle complex combinatorial problems by learning heuristics or direct solution construction. Graph Neural Networks (GNNs) are particularly effective for NCO due to their ability to capture the relational structure inherent in many such problems. In this work, we examine the supervised non-autoregressive (NAR) solution construction framework, revealing a misalignment between training objective and solution quality. Specifically, through experiments on six GNN architectures across three problems\u2014Traveling Salesperson Problem (TSP), Maximum Independent Set (MIS), and Minimum Vertex Cover (MVC)\u2014we show that lower training loss does not correlate with lower optimality gap. To address this, we propose a supervised autoregressive (AR) framework that leverages the conditional dependencies between variables by training to complete partial solutions. Empirical results show that the proposed AR framework does not exhibit the same misalignment and consistently improves performance. We further compare the proposed AR framework against existing supervised GNN-based methods and achieve superior performance, especially in terms of generalizing to larger problem instances.", "keywords": [ "graph neural network", "combinatorial optimization", "supervised learning" ], "rating_list": [ 8, 5, 3 ], "soundness_list": [ 3, 1, 2 ], "presentation_list": [ 4, 2, 2 ], "contribution_list": [ 3, 2, 1 ], "confidence_list": [ 2, 4, 5 ] }, { "paper_id": "WsgEWL8i0K", "primary_area": "datasets and benchmarks", "n_author": 13, "author_info_list": [ { "name": "Chuanhao Li", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Fanqing Meng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Hao Tian", "gender": "Male", "institution": "Sensetime Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Jiaqi Liao", "gender": "Male", "institution": "Research, Microsoft", "country": "CN", "position": "Intern" }, { "name": "Jifeng Dai", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jin Wang", "gender": "Male", "institution": "The University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Kaipeng Zhang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Quanfeng Lu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Tianshuo Yang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Wenqi Shao", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xizhou Zhu", "gender": "unknown", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Ping Luo", "gender": "unknown", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 40, "n_ref": 127, "n_ref_all": 197, "n_fig": 18, "n_tab": 21, "L_tab": 17625, "n_element_tab": 2252, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3932, "n_element_tab_1": 566, "formula_len_all": 35, "formula_len_all_1": 0, "len_all": 341962, "len_all_1": 67530, "len_abs": 1476, "len_title": 135, "len_sents": 73755, "len_sents_1": 27295, "n_sents": 575, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1263, "title": "MMIU: Multimodal Multi-image Understanding for Evaluating Large Vision-Language Models", "abs": "The capability to process multiple images is crucial for Large Vision-Language Models (LVLMs) to develop a more thorough and nuanced understanding of a scene. Recent multi-image LVLMs have begun to address this need. However, their evaluation has not kept pace with their development. To fill this gap, we introduce the Multimodal Multi-image Understanding (MMIU) benchmark, a comprehensive evaluation suite designed to assess LVLMs across a wide range of multi-image tasks. MMIU encompasses 7 types of multi-image relationships, 52 tasks, 77K images, and 11K meticulously curated multiple-choice questions, making it the most extensive benchmark of its kind. Our evaluation of nearly 30 popular LVLMs, including both open-source and proprietary models, reveals significant challenges in multi-image comprehension, particularly in tasks involving spatial understanding. Even the most advanced models, such as GPT-4o, achieve only 55.7\\% accuracy on MMIU. Through multi-faceted analytical experiments, we identify key performance gaps and limitations, providing valuable insights for future model and data improvements. We aim for MMIU to advance the frontier of LVLM research and development. We release the data and code at https://github.com/MMIUBenchmark/MMIU.", "keywords": [ "Multi-image Understanding", "Benchmark", "LVLM", "Evaluation" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "WrdLgVY5ZH", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Daiki Miwa", "gender": "Male", "institution": "Nagoya Institute of Technology, Tokyo Institute of Technology", "country": "JP", "position": "MS student" }, { "name": "Ichiro Takeuchi", "gender": "Male", "institution": "Nagoya University", "country": "JP", "position": "Full Professor" }, { "name": "Teruyuki Katsuoka", "gender": "Male", "institution": "Nagoya University", "country": "", "position": "Undergrad student" }, { "name": "Tomohiro Shiraishi", "gender": "Male", "institution": "Nagoya University", "country": "JP", "position": "PhD student" }, { "name": "Vo Nguyen Le Duy", "gender": "Male", "institution": "University of Information Technology, Vietnam National University - HCM", "country": "", "position": "Lecturer" } ], "n_formula": 41, "n_formula_1": 34, "n_ref_uni": 33, "n_ref": 47, "n_ref_all": 62, "n_fig": 9, "n_tab": 1, "L_tab": 70, "n_element_tab": 6, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2728, "formula_len_all_1": 2358, "len_all": 129378, "len_all_1": 63270, "len_abs": 1035, "len_title": 130, "len_sents": 33842, "len_sents_1": 27436, "n_sents": 281, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1041, "title": "Statistical Test on Diffusion Model-based Anomaly Detection by Selective Inference", "abs": "Advancements in AI image generation, particularly diffusion models, have progressed rapidly. However, the absence of an established framework for quantifying the reliability of AI-generated images hinders their use in critical decision-making tasks, such as medical image diagnosis. In this study, we address the task of detecting anomalous regions in medical images using diffusion models and propose a statistical method to quantify the reliability of the detected anomalies. The core concept of our method involves a selective inference framework, wherein statistical tests are conducted under the condition that the images are produced by a diffusion model. With our approach, the statistical significance of anomaly detection results can be quantified in the form of a $p$-value, enabling decision-making with controlled error rates, as is standard in medical practice. We demonstrate the theoretical soundness and practical effectiveness of our statistical test through numerical experiments on both synthetic and brain image datasets.", "keywords": [ "diffusion models", "anomaly detection", "statistical test", "selective inference" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 1 ] }, { "paper_id": "WrBqgoseGL", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Aryan Gulati", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Brando Miranda", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Bruno de Moraes Dumont", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Emily Xia", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Eric Chen", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Kai Fronsdal", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "MS student" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 4, "n_ref_uni": 36, "n_ref": 63, "n_ref_all": 89, "n_fig": 11, "n_tab": 4, "L_tab": 3751, "n_element_tab": 251, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 962, "n_element_tab_1": 38, "formula_len_all": 777, "formula_len_all_1": 253, "len_all": 165915, "len_all_1": 59380, "len_abs": 1412, "len_title": 153, "len_sents": 46420, "len_sents_1": 28453, "n_sents": 373, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1316, "title": "Putnam-AXIOM: A Functional & Static Benchmark for Measuring Higher Level Mathematical Reasoning in LLMs", "abs": "As large language models (LLMs) continue to advance, many existing benchmarks designed to evaluate their reasoning capabilities are becoming saturated. Therefore, we present the Putnam-AXIOM Original benchmark consisting of 236 mathematical problems from the William Lowell Putnam Mathematical Competition, along with detailed step-by-step solutions. To preserve the Putnam-AXIOM benchmark's validity and mitigate potential data contamination, we created the Putnam-AXIOM Variation benchmark with functional variations of 52 problems. By programmatically altering problem elements like variables and constants, we can generate unlimited novel, equally challenging problems not found online. We see that almost all models have significantly lower accuracy in the variations than the original problems. Our results reveal that OpenAI's o1-preview, the best performing model, achieves merely 41.95\\% accuracy on the Putnam-AXIOM Original but experiences around a 30\\% reduction in accuracy on the variations' dataset when compared to corresponding original problems. Moreover, we explore metrics beyond boxed accuracy to assess models on complex tasks like natural language theorem proving, crucial for evaluating reasoning capabilities in depth, opening the possibility for open-ended evaluation of reasoning strings.", "keywords": [ "Benchmarks", "Large Language Models", "Mathematical Reasoning", "Mathematics", "Reasoning", "Machine Learning" ], "rating_list": [ 5, 8, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "Wr3UuEx72f", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Abhinav Shrivastava", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "Associate Professor" }, { "name": "Hanyu Wang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Hao Chen", "gender": "unknown", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Saksham Suri", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Yixuan Ren", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 57, "n_ref": 122, "n_ref_all": 157, "n_fig": 16, "n_tab": 2, "L_tab": 1064, "n_element_tab": 145, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 611, "n_element_tab_1": 35, "formula_len_all": 454, "formula_len_all_1": 366, "len_all": 172432, "len_all_1": 62432, "len_abs": 0, "len_title": 183, "len_sents": 46058, "len_sents_1": 29842, "n_sents": 351, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 70, "L_abs": 1705, "title": "LARP: Tokenizing Videos with a Learned Autoregressive Generative Prior", "abs": "We present LARP, a novel video tokenizer designed to overcome limitations in current video tokenization methods for autoregressive (AR) generative models. Unlike traditional patchwise tokenizers that directly encode local visual patches into discrete tokens, LARP introduces a holistic tokenization scheme that gathers information from the visual content using a set of learned holistic queries. This design allows LARP to capture more global and semantic representations, rather than being limited to local patch-level information. Furthermore, it offers flexibility by supporting an arbitrary number of discrete tokens, enabling adaptive and efficient tokenization based on the specific requirements of the task. To align the discrete token space with downstream AR generation tasks, LARP integrates a lightweight AR transformer as a training-time prior model that predicts the next token on its discrete latent space. By incorporating the prior model during training, LARP learns a latent space that is not only optimized for video reconstruction but is also structured in a way that is more conducive to autoregressive generation. Moreover, this process defines a sequential order for the discrete tokens, progressively pushing them toward an optimal configuration during training, ensuring smoother and more accurate AR generation at inference time. Comprehensive experiments demonstrate LARPs strong performance, achieving state-of-the-art FVD on the UCF101 class-conditional video generation benchmark. LARP enhances the compatibility of AR models with videos and opens up the potential to build unified high-fidelity multimodal large language models (MLLMs). Code and checkpoints will be released.", "keywords": [ "Video Generation", "Visual Tokenization" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "Wqsk3FbD6D", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Alexander M Rush", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Associate Professor" }, { "name": "John Xavier Morris", "gender": "Male", "institution": "Cornell University", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 58, "n_ref_all": 67, "n_fig": 4, "n_tab": 2, "L_tab": 397, "n_element_tab": 74, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 397, "n_element_tab_1": 74, "formula_len_all": 707, "formula_len_all_1": 708, "len_all": 127939, "len_all_1": 61781, "len_abs": 1266, "len_title": 78, "len_sents": 29996, "len_sents_1": 28804, "n_sents": 227, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 30, "L_abs": 1270, "title": "Contextual Document Embeddings", "abs": "Dense document embeddings are central to neural retrieval. The dominant paradigm is to train and construct embeddings by running encoders directly on individual documents. In this work, we argue that these embeddings, while effective, are implicitly out-of-context for targeted use cases of retrieval, and that a contextualized document embedding should take into account both the document and neighboring documents in context - analogous to contextualized word embeddings. We propose two complementary methods for contextualized document embeddings: first, an alternative contrastive learning objective that explicitly incorporates the document neighbors into the intra-batch contextual loss; second, a new contextual architecture that explicitly encodes neighbor document information into the encoded representation. Results show that both methods achieve better performance than biencoders in several settings, with differences especially pronounced out-of-domain. We achieve state-of-the-art results on the MTEB benchmark with no hard negative mining, score distillation, dataset-specific instructions, intra-GPU example-sharing, or extremely large batch sizes. Our method can be applied to improve performance on any contrastive learning dataset and any biencoder.", "keywords": [ "text", "embeddings", "retrieval", "context", "contrastive" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 4, 4, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "WqL4wOU3tw", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Akio Hayakawa", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Researcher" }, { "name": "Masato Ishii", "gender": "Male", "institution": "Sony Research Inc.", "country": "JP", "position": "Researcher" }, { "name": "Takashi Shibuya", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Senior Manager" }, { "name": "Yuki Mitsufuji", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Lead Research Scientist" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 38, "n_ref": 87, "n_ref_all": 106, "n_fig": 7, "n_tab": 2, "L_tab": 825, "n_element_tab": 78, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 286, "n_element_tab_1": 26, "formula_len_all": 700, "formula_len_all_1": 648, "len_all": 125198, "len_all_1": 59842, "len_abs": 1172, "len_title": 185, "len_sents": 34389, "len_sents_1": 27675, "n_sents": 259, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 137, "L_abs": 1180, "title": "A Simple but Strong Baseline for Sounding Video Generation: Effective Adaptation of Audio and Video Diffusion Models for Joint Generation", "abs": "In this work, we build a simple but strong baseline for sounding video generation. Given base diffusion models for audio and video, we integrate them with additional modules into a single model and train it to make the model jointly generate audio and video. To enhance alignment between audio-video pairs, we introduce two novel mechanisms in our model. The first one is timestep adjustment, which provides different timestep information to each base model. It is designed to align how samples are generated along with timesteps across modalities. The second one is a new design of the additional modules, termed Cross-Modal Conditioning as Positional Encoding (CMC-PE). In CMC-PE, cross-modal information is embedded as if it represents temporal position information, and the embeddings are fed into the model like positional encoding. Compared with the popular cross-attention mechanism, CMC-PE provides a better inductive bias for temporal alignment in the generated data. Experimental results validate the effectiveness of the two newly introduced mechanisms and also demonstrate that our method outperforms existing methods. The source code will be released upon acceptance.", "keywords": [ "sounding video generation", "diffusion models", "audio-visual", "audio generation", "video generation" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "WpsrTQtnJR", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "He Sun", "gender": "Male", "institution": "University of Edinburgh", "country": "GB", "position": "Associate Professor" }, { "name": "Joyentanuj Das", "gender": "unknown", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Postdoc" }, { "name": "Suranjan de", "gender": "unknown", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" } ], "n_formula": 99, "n_formula_1": 52, "n_ref_uni": 20, "n_ref": 34, "n_ref_all": 41, "n_fig": 4, "n_tab": 1, "L_tab": 131, "n_element_tab": 29, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12201, "formula_len_all_1": 6778, "len_all": 102821, "len_all_1": 56894, "len_abs": 582, "len_title": 54, "len_sents": 28685, "len_sents_1": 20785, "n_sents": 276, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 583, "title": "Efficient Sparsification of Densely Connected Clusters", "abs": "When modelling a real-world dataset as a graph, groups of highly correlated data items correspond to densely connected vertex sets (clusters), and efficient algorithms that find these clusters have broad applications in various data analysis tasks. In this paper we study densely connected clusters in graphs and introduce two sparsification algorithms that preserve the structure of these clusters in both undirected graphs and directed ones. We show that our algorithms significantly speedup the running time of existing clustering algorithms while preserving their effectiveness.", "keywords": [ "densely connected components", "graph sparsification", "clustering" ], "rating_list": [ 5, 5, 3, 5, 8 ], "soundness_list": [ 3, 2, 2, 3, 4 ], "presentation_list": [ 2, 1, 1, 3, 3 ], "contribution_list": [ 2, 2, 1, 2, 3 ], "confidence_list": [ 2, 3, 3, 4, 3 ] }, { "paper_id": "WpjehX0TM2", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Mingxuan Sun", "gender": "Female", "institution": "Louisiana State University", "country": "US", "position": "Associate Professor" }, { "name": "Zhuoqun Li", "gender": "Female", "institution": "Louisiana State University", "country": "US", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 21, "n_ref_uni": 31, "n_ref": 50, "n_ref_all": 59, "n_fig": 2, "n_tab": 3, "L_tab": 1458, "n_element_tab": 79, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1707, "n_element_tab_1": 74, "formula_len_all": 1569, "formula_len_all_1": 1172, "len_all": 105731, "len_all_1": 61232, "len_abs": 1243, "len_title": 113, "len_sents": 30868, "len_sents_1": 27471, "n_sents": 229, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1247, "title": "Causal Reinforcement Learning for Spatio-Temporal Point Processes", "abs": "Spatio-temporal event sequences are increasingly accessible in various domains such as earthquake forecasting, crime prediction, and healthcare management. These data sources present unique challenges, as they involve both spatial and temporal dimensions, with event sequences exhibiting intricate dependencies over time and space. Neural network-based spatio-temporal point processes offer a sophisticated framework for modeling such event data. Conventional maximum likelihood estimation (MLE) of such data may lead to inaccurate predictions due to model-misspecification and compounding prediction errors. On the other hand, reinforcement learning frameworks, which treat event generation as actions and learn a policy to mimic event generation may alleviate the training/test discrepancy issue. Current reinforcement learning of point processes may have prohibitively poor exploration efficiency. In this paper, we propose the Causal learning improved Reinforcement Learning Spatio-Temporal Point Process (CRLSTPP) framework, which can mitigate the issue of compounding prediction errors and improve exploration efficiency at the same time. Experiments on both synthetic data and real-world data validate the superiority of the proposed model.", "keywords": [ "Spatio-Temporal Point Processes", "Reinforcement Learning", "Causal Inference" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "WpZyPk79Fu", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 16, "author_info_list": [ { "name": "Bo Li", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Chetan Bansal", "gender": "Not Specified", "institution": "Birla Institute of Technology and Science", "country": "", "position": "Undergrad student" }, { "name": "Huaxiu Yao", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Kaiyuan Zheng", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "MS student" }, { "name": "Mohit Bansal", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" }, { "name": "Peng Xia", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Shangyu Xing", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Tianle Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "MS student" }, { "name": "Wenhao Zheng", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Xuchao Zhang", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Ying Wei", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Yiyang Zhou", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Zhaorun Chen", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Zhaoyang Wang", "gender": "unknown", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Zijian Zhang", "gender": "Male", "institution": "East China University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Weitong ZHANG", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 44, "n_ref": 99, "n_ref_all": 129, "n_fig": 7, "n_tab": 15, "L_tab": 5661, "n_element_tab": 758, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 82, "n_element_tab_1": 17, "formula_len_all": 142, "formula_len_all_1": 142, "len_all": 212635, "len_all_1": 69691, "len_abs": 1487, "len_title": 111, "len_sents": 53645, "len_sents_1": 35381, "n_sents": 363, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1490, "title": "AnyPrefer: An Automatic Framework for Preference Data Synthesis", "abs": "High-quality preference data is essential for aligning foundation models with human values through preference learning. However, manual annotation of such data is often time-consuming and costly. Recent methods adopt a self-rewarding approach, where the target model generates and annotates its own preference data, but this can lead to inaccuracies due to the reward model sharing weights with the target model, amplifying inherent biases. To address these issues, we propose Anyprefer, a framework designed to synthesize high-quality preference data for the target model. Anyprefer frames the data synthesis process as a cooperative two-player Markov Game, where the target model and a judge model collaborate. Here, a series of external tools are introduced to assist the judge model in accurately rewarding the target model\u2019s responses, mitigating biases in the process. We also introduce a feedback mechanism to optimize prompts for both models, enhancing collaboration and improving data quality. The synthesized data is compiled into a new preference dataset, Anyprefer-V1, consisting of 58K high-quality preference pairs. Extensive experiments show that Anyprefer significantly improves model alignment across four applications, covering 21 datasets, achieving average improvements of 18.55 in five natural language generation datasets, 3.66 in nine vision-language understanding datasets, 30.05 in three medical image analysis datasets, and 14.50 in four visuo-motor control tasks.", "keywords": [ "synthetic preference data generation; preference fine-tuning" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 1, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "WpXq5n8yLb", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Aonan Zhang", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Xuanyu Zhang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Yi Wang", "gender": "Male", "institution": "Apple", "country": "", "position": "Principal Researcher" }, { "name": "Yunfei Cheng", "gender": "Female", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Chong Wang", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 28, "n_ref": 69, "n_ref_all": 89, "n_fig": 7, "n_tab": 8, "L_tab": 3099, "n_element_tab": 611, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2367, "n_element_tab_1": 404, "formula_len_all": 98, "formula_len_all_1": 98, "len_all": 128709, "len_all_1": 72838, "len_abs": 1021, "len_title": 120, "len_sents": 39376, "len_sents_1": 31200, "n_sents": 317, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 973, "title": "Recurrent Drafter for Fast Speculative Decoding in Large Language Models", "abs": "We present Recurrent Drafter (ReDrafter), an advanced speculative decoding approach that achieves state-of-the-art speedup for large language models (LLMs) inference. The performance gains are driven by three key aspects: (1) leveraging a recurrent neural network (RNN) as the draft model conditioning on LLM's hidden states, (2) applying a dynamic tree attention algorithm over beam search results to eliminate duplicated prefixes in candidate sequences, and (3) training through knowledge distillation from the LLM. ReDrafter accelerates Vicuna inference in MT-Bench by up to 3.5x with a PyTorch implementation on Nvidia H100 GPUs. To demonstrate its practicality in production environments, we integrate ReDrafter into TensorRT-LLM, reaching up to 2.5x speedup on H100 GPUs. We also validated its effectiveness for on-device applications by implementing the approach in MLX and benchmarking performance on Metal GPUs in Apple Silicon chips, achieving up to 2.3x speedup.", "keywords": [ "speculative decoding", "beam search" ], "rating_list": [ 6, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "WpObsQTpfp", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Bingchen Zhao", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Haoqin Tu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Huangjie Zheng", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Jieru Mei", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Junfei Xiao", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Mude Hui", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Qing Liu", "gender": "Female", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Sucheng Ren", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Xianhang Li", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Yuyin Zhou", "gender": "unknown", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "Zeyu Wang", "gender": "unknown", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "cihang xie", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 52, "n_ref": 144, "n_ref_all": 164, "n_fig": 6, "n_tab": 6, "L_tab": 2398, "n_element_tab": 310, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1749, "n_element_tab_1": 234, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 227974, "len_all_1": 65664, "len_abs": 1181, "len_title": 105, "len_sents": 58022, "len_sents_1": 29372, "n_sents": 437, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1206, "title": "What If We Recaption Billions of Web Images with LLaMA-3?", "abs": "Web-crawled image-text pairs are inherently noisy. Prior studies demonstrate that semantically aligning and enriching textual descriptions of these pairs can significantly enhance model training across various vision-language tasks, particularly text-to-image generation. However, large-scale investigations in this area remain predominantly closed-source. \nOur paper aims to bridge this community effort, leveraging the powerful and \\textit{open-sourced} LLaMA-3, a GPT-4 level LLM.\nOur recaptioning pipeline is simple: first, we fine-tune a LLaMA-3-8B powered LLaVA-1.5 and then employ it to recaption \\app1.3 billion images from the DataComp-1B dataset. Our empirical results confirm that this enhanced dataset, Recap-DataComp-1B, offers substantial benefits in training advanced vision-language models. For discriminative models like CLIP, we observe \\cg{an average of 3.1\\% enhanced zero-shot performance cross four cross-modal retrieval tasks using a mixed set of the original and our captions}. For generative models like text-to-image Diffusion Transformers, the generated images exhibit a significant improvement in alignment with users' text instructions, especially in following complex queries.", "keywords": [ "image-text datasets; synthetic captions" ], "rating_list": [ 5, 5, 3, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 4 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "WoPovNkM5h", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Bo Zheng", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Chaowen Hu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Jiahuan Yan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jian Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Jimeng Sun", "gender": "unknown", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Professor" }, { "name": "Jintai Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Yaojun Hu", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 42, "n_ref": 85, "n_ref_all": 119, "n_fig": 6, "n_tab": 9, "L_tab": 2340, "n_element_tab": 170, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 1670, "n_element_tab_1": 99, "formula_len_all": 194, "formula_len_all_1": 57, "len_all": 129140, "len_all_1": 63743, "len_abs": 1691, "len_title": 124, "len_sents": 38171, "len_sents_1": 30996, "n_sents": 217, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1691, "title": "Synergy Learning with Small Models promotes LLM Zero-Shot Tabular Prediction", "abs": "Recent development in large language models (LLMs) has demonstrated impressive zero-shot proficiency on unstructured textual or multi-modal tasks across various domains. However, despite with inherent world knowledge, their application on structured tabular data prediction still lags behind, primarily due to the numerical insensitivity and modality discrepancy that brings a gap between LLM reasoning and statistical machine learning. Unlike textual or vision data (e.g., electronic health records, medical images), tabular data is often presented in heterogeneous numerical values (e.g., blood test reports). This ubiquitous data format requires intensive expert annotation, and its numerical nature limits LLMs' ability to effectively transfer untapped domain expertise. In this paper, we propose SERSAL, a general loop of thought prompting method by synergy learning with small models to unconditionally enhance zero-shot tabular prediction for LLMs. Specifically, SERSAL utilizes the LLM's zero-shot outcomes as original soft annotations, which are dynamically leveraged to teach a better small student model in a semi-supervised manner. Reversely, the outcomes from the trained small model are used to teach the LLM to further refine its real capability. Such mutual process can be repeatedly applied for continuous progress. Comprehensive experiments on widely used domain tabular datasets show that, without access to gold labels, applying SERSAL to OpenAI GPT reasoning process attains substantial improvement compared to linguistic prompting methods, which serves as an orthogonal direction for tabular LLM, and increasing prompting bonus is observed as more powerful LLMs appear.", "keywords": [ "tabular data", "prompt learning", "classification" ], "rating_list": [ 5, 8, 8, 5 ], "soundness_list": [ 3, 4, 4, 2 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 3, 4, 4, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "WoJzHQIIUk", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Ercan Engin KURUOGLU", "gender": "Male", "institution": "Tsinghua-Berkeley Shenzhen Institute ", "country": "CN", "position": "Full Professor" }, { "name": "Junping Hong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yutong Jiang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 27, "n_ref": 43, "n_ref_all": 49, "n_fig": 3, "n_tab": 2, "L_tab": 741, "n_element_tab": 76, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 741, "n_element_tab_1": 76, "formula_len_all": 219, "formula_len_all_1": 219, "len_all": 63367, "len_all_1": 63367, "len_abs": 1315, "len_title": 112, "len_sents": 15179, "len_sents_1": 15179, "n_sents": 100, "n_sents_1": 100, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 994, "title": "MinMax Bayesian Neural Networks and Uncorrelated Representation", "abs": "In deep learning, Bayesian neural networks (BNN) and dropout techniques provide the role of robustness analysis, and the minimax method used to be a conservative choice in the traditional Bayesian field. In this paper, we apply the minimax game to the BNN on the representation level and formulate as a two-player game between a deterministic neural network $f$ and a sampling stochastic neural network $f + r*\\xi$, which can be seen as a Brownian Motion of $f$. Our simple experiments show that $r$ will be stable with enough dimension space, suitable activation function, and without bias with the minimax coding rate loss, which verify the statement \\cite{yu2020learning} in some sense. And we test the convolutional neural network without bias, with bias and with batch normalization on simple data set like MNIST, Fashion MNIST and others, and visualize the sampling radius as a bias-variance tradeoff study. At last, we also test how noise perturbation will affect radius in stable case.", "keywords": [ "Minimax game", "Bayesian Neural Networks", "Brownian Motion", "Minimax coding reduction", "Uncorrelated representation" ], "rating_list": [ 1, 1, 3, 1 ], "soundness_list": [ 2, 1, 2, 1 ], "presentation_list": [ 1, 1, 1, 1 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 3, 5, 2, 4 ] }, { "paper_id": "WoGnnggVCZ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Alice Xiang", "gender": "Female", "institution": "Sony AI", "country": "", "position": "Principal Researcher" }, { "name": "Lele Chen", "gender": "Male", "institution": "Sony America", "country": "", "position": "Researcher" }, { "name": "Yulun Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yunhao Ba", "gender": "Male", "institution": "Sony", "country": "US", "position": "Researcher" }, { "name": "Zhiteng Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Jerone Theodore Alexander Andrews", "gender": "Male", "institution": "Sony AI", "country": "GB", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 50, "n_ref": 104, "n_ref_all": 125, "n_fig": 6, "n_tab": 7, "L_tab": 3112, "n_element_tab": 259, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 3583, "n_element_tab_1": 243, "formula_len_all": 572, "formula_len_all_1": 447, "len_all": 142280, "len_all_1": 62652, "len_abs": 974, "len_title": 114, "len_sents": 29798, "len_sents_1": 25457, "n_sents": 230, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 978, "title": "GenDataAgent: On-the-fly Dataset Augmentation with Synthetic Data", "abs": "Synthetic data is increasingly employed for training dataset augmentation in computer vision. However, prior works typically perform a uniform search across the entire category space, overlooking the interaction between synthetic data generation and downstream task training. Furthermore, balancing the diversity of synthetic data while ensuring it remains within the same distribution as real data (i.e., avoiding outliers) remains a significant challenge.\nIn this work, we propose a generative agent to augment target training datasets with synthetic data for model fine-tuning. Our agent iteratively generates relevant data on-the-fly, aligning with the target training dataset distribution. It prioritizes sampling diverse synthetic data that complements marginal training samples, with a focus on synthetic data that exhibit higher variance in gradient updates. Evaluations across diverse supervised image classification tasks demonstrate the effectiveness of our approach.", "keywords": [ "supervised learning", "classification", "computer vision", "synthetic data", "generative AI", "responsible AI", "fairness" ], "rating_list": [ 6, 8, 8, 3 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Wo66GEFnXd", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Haiyang Yu", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Jacob Helwig", "gender": "unknown", "institution": "Texas A&M University - College Station", "country": "", "position": "PhD student" }, { "name": "Shuiwang Ji", "gender": "Male", "institution": "Texas A&M University", "country": "US", "position": "Professor" }, { "name": "Xiaofeng Qian", "gender": "unknown", "institution": "Texas A&M University", "country": "US", "position": "Associate Professor" }, { "name": "Xuan Zhang", "gender": "Male", "institution": "Texas A&M", "country": "US", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 23, "n_ref_uni": 57, "n_ref": 84, "n_ref_all": 110, "n_fig": 12, "n_tab": 6, "L_tab": 1246, "n_element_tab": 120, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1304, "formula_len_all_1": 1304, "len_all": 166023, "len_all_1": 65050, "len_abs": 1245, "len_title": 145, "len_sents": 43215, "len_sents_1": 31679, "n_sents": 300, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1204, "title": "Learning Time-Dependent Density Functional Theory via Geometry and Physics Aware Latent Evolution", "abs": "We consider using machine learning to simulate time-dependent density functional theory (TDDFT) to predict physical properties of molecules and materials beyond their ground states. In particular, by simulating the electronic response of the system under an external electromagnetic field, the optical absorption spectrum can be calculated using real-time TDDFT (RT-TDDFT), which provides physical information about the excited states and dipole strength function. However, RT-TDDFT simulation requires the direct propagation of electronic wavefunctions of all valence electrons for extended periods, making the process very time-consuming. In this work, we model electron density as volumetric data and train neural networks to map between coarse time steps. To make the model aware of the atomistic environment, we incorporate 3D message passing into the model architecture. Additionally, we use latent evolution to regularize the model towards learning the underlying physics. Our method is termed TDDFTNet. To evaluate our approach, we generate datasets using molecules from the MD17 dataset. Results show that TDDFTNet can learn the time propagation of electron densities accurately and efficiently.", "keywords": [ "AI for science", "Density functional theory", "Real-time TDDFT", "Neural PDE solver" ], "rating_list": [ 5, 6, 8, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "Wnu2c6pjs1", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Honghan Wu", "gender": "unknown", "institution": "University College London, University of London", "country": "", "position": "Associate Professor" }, { "name": "Jinge Wu", "gender": "Female", "institution": "University College London", "country": "GB", "position": "PhD student" }, { "name": "Yunsoo Kim", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 23, "n_ref": 42, "n_ref_all": 53, "n_fig": 4, "n_tab": 10, "L_tab": 7181, "n_element_tab": 679, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 3145, "n_element_tab_1": 257, "formula_len_all": 337, "formula_len_all_1": 337, "len_all": 128982, "len_all_1": 59141, "len_abs": 2291, "len_title": 177, "len_sents": 35432, "len_sents_1": 25051, "n_sents": 249, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 129, "L_abs": 1612, "title": "RadEyeVideo: Enhancing general-domain Large Vision Language Model for chest X-ray analysis with video representations of eye gaze", "abs": "Large Vision-Language Models (LVLMs) have demonstrated promising performance in chest X-ray (CXR) analysis. To enhance human-computer interaction, several studies have incorporated radiologists' eye gaze, typically through heatmaps or textual prompts. However, these methods often overlook the sequential order of eye movements, which could provide valuable insights by highlighting both the areas of interest and the order in which they are examined. In this work, we propose a novel approach called RadEyeVideo that integrates radiologists\u2019 eye-fixation data as a video sequence, capturing both the temporal and spatial dynamics of their gaze. The video, featuring a red gaze point overlaid on CXR images, emphasizes regions of focused attention during interpretation. We evaluate this method in CXR report generation and disease diagnosis using three general-domain, open-source LVLMs with a video input capabilities. When prompted with eye-gaze videos, model performance improves by up to 25.4% on Impression generation task and on average 7.9% for all tasks using scaled evaluation metrics. Our approach enhanced open-domain LVLM models, when combined with exemplar reports for in-context learning, outperform medical models as well as those specifically trained for CXR report generation on the benchmark dataset. This work highlights that domain expert's knowledge (eye-gaze information in this case), when effectively integrated with LVLMs, can significantly enhance general-domain models' capabilities in clinical tasks, pointing out a new effective approach of utilising LVLMs in healthcare and beyond.", "keywords": [ "LVLM", "Eye Gaze", "Video", "Medicine", "Medical Image", "Chest X-ray", "Chest X-ray Report Generation" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "WncnpvJk83", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Benlin Liu", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "James Zou", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Jiaxi Yang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Wenlong Deng", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Xiaoxiao Li", "gender": "Unspecified", "institution": "University of British Columbia", "country": "CA", "position": "Assistant Professor" }, { "name": "Yangsibo Huang", "gender": "Female", "institution": "Google", "country": "US", "position": "Research Scientist" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 68, "n_ref_all": 108, "n_fig": 12, "n_tab": 9, "L_tab": 2716, "n_element_tab": 165, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 521, "n_element_tab_1": 64, "formula_len_all": 1212, "formula_len_all_1": 497, "len_all": 174055, "len_all_1": 63619, "len_abs": 1340, "len_title": 113, "len_sents": 60266, "len_sents_1": 29647, "n_sents": 449, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1342, "title": "GMValuator: Similarity-based Data Valuation for Generative Models", "abs": "Data valuation plays a crucial role in machine learning. Existing data valuation methods, mainly focused on discriminative models, overlook generative models that have gained attention recently. In generative models, data valuation measures the impact of training data on generated datasets. Very few existing attempts at data valuation methods designed for deep generative models either concentrate on specific models or lack robustness in their outcomes. Moreover, efficiency still reveals vulnerable shortcomings. We formulate the data valuation problem in generative models from a similarity matching perspective to bridge the gaps. Specifically, we introduce Generative Model Valuator (GMValuator), the first training-free and model-agnostic approach to providing data valuation for generation tasks. It empowers efficient data valuation through our innovative similarity matching module, calibrates biased contributions by incorporating image quality assessment, and attributes credits to all training samples based on their contributions to the generated samples. Additionally, we introduce four evaluation criteria for assessing data valuation methods in generative models. GMValuator is extensively evaluated on benchmark and high-resolution datasets and various mainstream generative architectures to demonstrate its effectiveness.", "keywords": [ "Generative Model", "Data Valuation" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 4, 3 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "WmmPHE4k5f", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Hanyu Zhou", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Haonan Wang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Haoyue Liu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Luxin Yan", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Shihan Peng", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Yuxing Duan", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yi Chang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 39, "n_ref": 59, "n_ref_all": 79, "n_fig": 9, "n_tab": 6, "L_tab": 1399, "n_element_tab": 115, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1157, "n_element_tab_1": 93, "formula_len_all": 583, "formula_len_all_1": 475, "len_all": 118123, "len_all_1": 68564, "len_abs": 2012, "len_title": 131, "len_sents": 37449, "len_sents_1": 33402, "n_sents": 260, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 2021, "title": "Bridge Frame and Event: Common Spatiotemporal Fusion for High-Dynamic Optical Flow", "abs": "High-dynamic scene optical flow is a challenging task, which suffers large displacement. Limited by frame imaging, large displacement causes potential spatial blurry textures due to long exposure and temporal discontinuous motion due to low frame rate, thus deteriorating the spatiotemporal feature of optical flow. Typically, existing methods mainly introduce event camera with high temporal resolution to directly fuse the spatiotemporal features between the two modalities. However, this direct fusion is ineffective, since there exists a large gap due to the heterogeneous data representation between frame and event modalities. To address this issue, we explore a common-latent space as an intermediate bridge to mitigate the modality gap. In this work, we propose a novel common spatiotemporal fusion between frame and event modalities for high-dynamic scene optical flow, including visual boundary localization and motion correlation fusion. Specifically, in visual boundary localization, we figure out that frame and event can be derived into the spatiotemporal gradient maps with the same data representation, where the similarity distribution between the two modalities is consistent with the extracted boundary distribution. This motivates us to design the common spatiotemporal gradient to constrain the localization of the reference boundary as a template. In motion correlation fusion, we discover that the frame-based motion possesses spatially dense but temporally discontinuous correlation, while the event-based motion has spatially sparse but temporally continuous correlation. This inspires us to take the reference boundary template to guide the fusion of the complementary motion knowledge between the two modalities. Moreover, common spatiotemporal fusion can not only relieve the cross-modal feature discrepancy, but also make the fusion process interpretable to achieve dense and continuous optical flow. Extensive experiments have been performed to verify the superiority of the proposed method.", "keywords": [ "high-dynamic scene", "optical flow", "event camera", "multimodal fusion" ], "rating_list": [ 5, 5, 8, 5 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 3, 1, 4, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "WmTYaKWHVW", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Anand Raghunathan", "gender": "unknown", "institution": "Purdue University", "country": "", "position": "Full Professor" }, { "name": "Jimmy Gammell", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Kaushik Roy", "gender": "unknown", "institution": "Purdue University", "country": "", "position": "" }, { "name": "Abolfazl Hashemi", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 116, "n_formula_1": 25, "n_ref_uni": 55, "n_ref": 114, "n_ref_all": 131, "n_fig": 12, "n_tab": 3, "L_tab": 841, "n_element_tab": 76, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12438, "formula_len_all_1": 2424, "len_all": 261940, "len_all_1": 69197, "len_abs": 2017, "len_title": 77, "len_sents": 80600, "len_sents_1": 28984, "n_sents": 611, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1999, "title": "Learning to localize leakage of cryptographic keys through power consumption", "abs": "While cryptographic algorithms such as the ubiquitous Advanced Encryption Standard (AES) are secure, *physical implementations* of these algorithms in hardware inevitably 'leak' sensitive information such as cryptographic keys. A particularly insidious form of leakage arises from the fact that hardware's power consumption over time is statistically associated with the data it processes and the instructions it executes. Supervised deep learning has emerged as a state-of-the-art tool for carrying out *power side-channel attacks*, which exploit this leakage to break cryptographic implementations by learning to map power consumption measurements recorded during encryption to the secret key used for that encryption. In this work, we seek instead to develop a principled deep learning framework for *defense* against such attacks by understanding the relative leakage due to power measurements recorded at different points in time. This information is invaluable to cryptographic hardware designers for understanding *why* their hardware leaks and how they can mitigate the leakage (e.g. by indicating that a particular section of code or electronic component is responsible for leakage and should be revised). Towards this end, we propose a novel deep learning algorithm by formulating an adversarial game played between a classifier trained to estimate the conditional distribution of a key given power measurements, and an 'obfuscator' which probabilistically erases individual power measurements and is trained to minimize the classifier-estimated log-likelihood of the correct key, subject to a penalty on erasure probability. We theoretically characterize the ideal output of our algorithm in terms of conditional mutual information quantities involving the key and individual power measurements. We then demonstrate the efficacy of our algorithm on real and synthetic datasets of power measurements from implementations of the AES cryptographic standard. Our code can be found (redacted).", "keywords": [ "deep learning", "adversarial", "mutual information", "side channel attack", "cryptography" ], "rating_list": [ 3, 3, 8, 5 ], "soundness_list": [ 2, 1, 4, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "WlhVRh2rQ0", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Annika Betken", "gender": "unknown", "institution": "University of Twente", "country": "", "position": "Assistant Professor" }, { "name": "Hanyuan Hang", "gender": "Male", "institution": "University of Twente", "country": "", "position": "Assistant Professor" }, { "name": "Hongwei Wen", "gender": "unknown", "institution": "University of Twente", "country": "NL", "position": "PhD student" } ], "n_formula": 200, "n_formula_1": 42, "n_ref_uni": 34, "n_ref": 71, "n_ref_all": 256, "n_fig": 2, "n_tab": 5, "L_tab": 1052, "n_element_tab": 60, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 23187, "formula_len_all_1": 3309, "len_all": 275519, "len_all_1": 80296, "len_abs": 927, "len_title": 132, "len_sents": 76572, "len_sents_1": 32480, "n_sents": 793, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 929, "title": "Optimal Learning of Kernel Logistic Regression for Complex Classification Scenarios", "abs": "Complex classification scenarios, including long-tailed learning, domain adaptation, and transfer learning, present substantial challenges for traditional algorithms. Conditional class probability (CCP) predictions have recently become critical components of many state-of-the-art algorithms designed to address these challenging scenarios. Among kernel methods, kernel logistic regression (KLR) is distinguished by its effectiveness in predicting CCPs through the minimization of the cross-entropy (CE) loss. Despite the empirical success of CCP-based approaches, the theoretical understanding of their performance, particularly regarding the CE loss, remains limited. In this paper, we bridge this gap by demonstrating that KLR-based algorithms achieve minimax optimal convergence rates for the CE loss under mild assumptions in these complex tasks, thereby establishing their theoretical efficiency in such demanding contexts.", "keywords": [ "complex classification scenarios", "long-tailed learning", "domain adaptation", "transfer learning", "kernel methods", "logistic regression", "learning theory" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "WlKGZuolEk", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Liang Wang", "gender": "Male", "institution": "Institute of Automation\uff0c CAS\uff0cChina", "country": "CN", "position": "Full Professor" }, { "name": "Qisen Ma", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Yan Huang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 26, "n_ref": 66, "n_ref_all": 92, "n_fig": 11, "n_tab": 12, "L_tab": 2607, "n_element_tab": 384, "n_fig_1": 6, "n_tab_1": 9, "L_tab_1": 1755, "n_element_tab_1": 260, "formula_len_all": 410, "formula_len_all_1": 381, "len_all": 118174, "len_all_1": 66212, "len_abs": 1406, "len_title": 143, "len_sents": 36160, "len_sents_1": 27620, "n_sents": 267, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1369, "title": "Hierarchical Multimodal Knowledge Matching for Training-Free Open-Vocabulary Object Detection", "abs": "Open-Vocabulary Object Detection (OVOD) aims to leverage the generalization capabilities of pre-trained vision language models for detecting objects beyond the trained categories. \nExisting methods mostly focus on supervised learning strategies based on available training data, which might be suboptimal for data-limited novel categories.\nTo tackle this challenge, this paper presents a $\\textbf{H}$ierarchical $\\textbf{M}$ultimodal $\\textbf{K}$nowledge $\\textbf{M}$atching method ($\\textbf{HMKM}$) to better represent novel categories and match them with region features.\nSpecifically, HMKM includes a set of object prototype knowledge that is obtained using limited category-specific images, acting as off-the-shelf category representations.\nIn addition, HMKM also includes a set of attribute prototype knowledge to represent key attributes of categories at a fine-grained level, with the goal to distinguish one category from its visually similar ones.\nDuring inference, two sets of object and attribute prototype knowledge are adaptively combined to match categories with region features.\nThe proposed HMKM is training-free and can be easily integrated as a plug-and-play module into existing OVOD models. \nExtensive experiments demonstrate that our HMKM significantly improves the performance when detecting novel categories across various backbones and datasets.", "keywords": [ "open-vocabulary object detection", "multimodal knowledge", "vision and language" ], "rating_list": [ 5, 3, 5, 6, 6 ], "soundness_list": [ 4, 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 5, 4 ] }, { "paper_id": "Wl5HGuFYVp", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Wenqi Ren", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Xiuyi Jia", "gender": "unknown", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zhuoran Zheng", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Postdoc" } ], "n_formula": 15, "n_formula_1": 12, "n_ref_uni": 34, "n_ref": 59, "n_ref_all": 86, "n_fig": 8, "n_tab": 10, "L_tab": 6075, "n_element_tab": 488, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 9235, "n_element_tab_1": 539, "formula_len_all": 853, "formula_len_all_1": 806, "len_all": 111198, "len_all_1": 70650, "len_abs": 1599, "len_title": 65, "len_sents": 32958, "len_sents_1": 23474, "n_sents": 274, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1506, "title": "CLIP2LE\uff1aA Label Enhancement Fair Representation Method via CLIP", "abs": "Label enhancement is a novel label shift strategy that aims to integrate the feature space with the logical label space to obtain a high-quality label distribution. This label distribution can serve as a soft target for algorithmic learning, akin to label smoothing, thereby enhancing the performance of various learning paradigms including multi-label learning, single positive label learning and partial-label learning. However, limited by dataset type and annotation inaccuracy, the same label enhancement algorithm on different datasets struggles to achieve consistent performance, for reasons derived from the following two insights: 1) Differential Contribution of Feature Space and Logical Label Space: The feature space and logical label space of different datasets contribute differently to generating an accurate label distribution; 2) Presence of Noise and Incorrect Labels: Some datasets contain noise and inaccurately labeled samples, leading to divergent outputs for similar inputs. To address these challenges, we propose leveraging CLIP (Contrastive Language-Image Pretraining) as a foundational strategy, treating the feature space and the logical label space as two distinct modalities. By recoding these modalities before applying the label enhancement algorithm, we aim to achieve a fair and robust representation. Extensive experimental results demonstrate the effectiveness of our approach to help existing label enhancement algorithms improve their performance on several benchmarks.", "keywords": [ "Label enhancement", "CLIP" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 2, 2, 3, 4 ] }, { "paper_id": "WkpqUVcSTy", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Afshin Dehghan", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Haiming Gang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Hong-You Chen", "gender": "unknown", "institution": "Apple AI/ML", "country": "US", "position": "Researcher" }, { "name": "Kai Kang", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Mingfei Gao", "gender": "unknown", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Mingze Xu", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Zhe Gan", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Zhengfeng Lai", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 28, "n_ref": 96, "n_ref_all": 120, "n_fig": 5, "n_tab": 4, "L_tab": 2495, "n_element_tab": 286, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1117, "n_element_tab_1": 102, "formula_len_all": 499, "formula_len_all_1": 423, "len_all": 128720, "len_all_1": 63615, "len_abs": 1083, "len_title": 79, "len_sents": 33171, "len_sents_1": 28563, "n_sents": 261, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1101, "title": "SlowFast-LLaVA: A strong training-free baseline for video large language models", "abs": "We propose SlowFast-LLaVA (or SF-LLaVA for short), a training-free video large language model (LLM) that can jointly capture the detailed spatial semantics and long-range temporal context without exceeding the token budget of commonly used LLMs. This is realized by using a two-stream SlowFast design of inputs for Video LLMs to aggregate features from sampled video frames in an effective way. Specifically, the Slow pathway extracts features at a low frame rate while keeping as many spatial details as possible (e.g., with 24x24 tokens), and the Fast pathway operates on a high frame rate but uses a larger spatial pooling stride (e.g., downsampling 6x) to focus on the motion cues. As a result, this design allows us to adequately capture both spatial and temporal features that are beneficial for understanding details in the video. Experimental results show that SF-LLaVA outperforms existing training-free methods on a wide range of video tasks. On some benchmarks, it achieves comparable or even better performance compared to state-of-the-art Video LLMs that are fine-tuned on video datasets.", "keywords": [ "Multimodal Large Language Model", "Video Understanding", "Video Question-Answering" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "WkhlUyJcOJ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Aiwei Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Han Yu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Hao Zhu", "gender": "Male", "institution": "CSIRO", "country": "AU", "position": "Postdoc" }, { "name": "Irwin King", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Piotr Koniusz", "gender": "unknown", "institution": "University of New South Wales", "country": "AU", "position": "Associate Professor" }, { "name": "Yifei Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" } ], "n_formula": 51, "n_formula_1": 17, "n_ref_uni": 39, "n_ref": 74, "n_ref_all": 88, "n_fig": 4, "n_tab": 7, "L_tab": 3377, "n_element_tab": 562, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1480, "n_element_tab_1": 120, "formula_len_all": 3981, "formula_len_all_1": 1023, "len_all": 169848, "len_all_1": 68597, "len_abs": 1354, "len_title": 133, "len_sents": 40533, "len_sents_1": 29548, "n_sents": 349, "n_sents_1": 247, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1361, "title": "Less is More: Extreme Gradient Boost Rank-1 Adaption for Efficient Finetuning of LLMs", "abs": "Fine-tuning Large Language Models (LLMs) has become a crucial technique for adapting pre-trained models to downstream tasks. However, the enormous size of LLMs poses significant challenges in terms of computational complexity and resource requirements. Low-Rank Adaptation (LoRA) has emerged as a promising solution. However, there exists a gap between the practical performance of low-rank adaptations and its theoretical optimum. In this work, we propose eXtreme Gradient Boosting LoRA (XGBLoRA), a novel framework that bridges this gap by leveraging the power of ensemble learning. Inspired by gradient boosting, XGBLoRA iteratively learns and merges a sequence of LoRA adaptations to refine model predictions. It achieves better performance than the standard LoRA, while enjoying the computational efficiency of rank-1 adaptations. We provide theoretical analysis to show the convergence and optimality of our approach, and conduct extensive experiments on a range of natural language processing tasks. The results demonstrate that XGBLoRA consistently outperforms standard LoRA and achieves performance comparable to full fine-tuning with significantly fewer trainable parameters. This work advances parameter-efficient fine-tuning for LLMs, and offers a promising solution for adapting LLMs to downstream tasks while optimizing performance and efficiency.", "keywords": [ "LLM", "PEFT", "LoRA" ], "rating_list": [ 6, 8, 5, 6, 5 ], "soundness_list": [ 3, 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 4, 2, 3, 2 ], "confidence_list": [ 3, 3, 5, 4, 3 ] }, { "paper_id": "WkSP7DfwVW", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Jiayuan Liu", "gender": "Male", "institution": "Computer Science Department, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Mingyu Guo", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Senior Lecturer" }, { "name": "Vincent Conitzer", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 33, "n_formula_1": 11, "n_ref_uni": 17, "n_ref": 52, "n_ref_all": 107, "n_fig": 10, "n_tab": 10, "L_tab": 7452, "n_element_tab": 642, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2966, "formula_len_all_1": 1121, "len_all": 216277, "len_all_1": 79651, "len_abs": 1610, "len_title": 92, "len_sents": 87893, "len_sents_1": 39332, "n_sents": 797, "n_sents_1": 325, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1556, "title": "Extending Myerson's Optimal Auctions to Correlated Bidders via Neural Network Interpolation", "abs": "We aim to design revenue-maximizing single-item auctions that are deterministic, strategy-proof and ex post individually rational. Myerson's seminal work on optimal auction design solved this problem for independent bidders. Myerson introduced the novel concept of virtual valuation and showed that revenue maximization is equivalent to virtual valuation maximization. Coincidentally, by greedily allocating the item to the bidder with the highest (ironed) virtual valuation, the resulting allocation is guaranteed to be monotone -- a necessary and sufficient condition for strategy-proofness.\n\nFor correlated bidders, Myerson's greedy allocation no longer guarantees monotonicity/strategy-proofness. We propose a simple yet empirically effective approach for designing near-optimal auctions for correlated bidders. We train a neural network to interpolate the greedy allocation, while enforcing that the interpolation must be verifiably monotone.\n\nEmpirically, our method consistently achieves near-optimal revenue across a wide range of distributions, including adversarially generated cases. Compared to existing baselines, our approach shows substantial improvement, often reducing the gap to the (unattainable) greedy upper bound by an order of magnitude.\n\nFurthermore, we demonstrate the generality of our approach by extending it to multi-unit auctions with unit demand, where we achieve similarly strong performance. Additionally, our verification techniques can be integrated into the RegretNet framework to design fully strategy-proof auctions.", "keywords": [ "mechanism design", "optimal auction", "revenue maximization", "virtual valuation" ], "rating_list": [ 5, 6, 8, 3 ], "soundness_list": [ 2, 4, 4, 3 ], "presentation_list": [ 1, 4, 3, 2 ], "contribution_list": [ 2, 3, 4, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "WkHkwo8rpL", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Acyr Locatelli", "gender": "Male", "institution": "Cohere", "country": "", "position": "Researcher" }, { "name": "Ahmet \u00dcst\u00fcn", "gender": "Male", "institution": "Cohere For AI", "country": "CA", "position": "Researcher" }, { "name": "Nikolas Gritsch", "gender": "Male", "institution": "Cohere", "country": "CA", "position": "Member of technical staff" }, { "name": "Qizhen Zhang", "gender": "Female", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Sara Hooker", "gender": "unknown", "institution": "Cohere For AI", "country": "US", "position": "Principal Researcher" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 64, "n_ref": 119, "n_ref_all": 148, "n_fig": 10, "n_tab": 12, "L_tab": 1752, "n_element_tab": 294, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 270, "n_element_tab_1": 49, "formula_len_all": 374, "formula_len_all_1": 213, "len_all": 226569, "len_all_1": 60870, "len_abs": 1180, "len_title": 132, "len_sents": 44237, "len_sents_1": 30257, "n_sents": 292, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1184, "title": "Nexus: Specialization meets Adaptability for Efficiently Training Mixture of Experts", "abs": "Efficiency, specialization, and adaptability to new data distributions are qualities that are hard to combine in current Large Language Models. The Mixture of Experts (MoE) architecture has been the focus of significant research because its inherent conditional computation enables such desirable properties. In this work, we focus on \"upcycling\" dense expert models into an MoE, aiming to improve specialization while also adding the ability to adapt to new tasks easily. We introduce Nexus, an enhanced MoE architecture with adaptive routing where the model learns to project expert embeddings from domain representations. This approach allows Nexus to flexibly add new experts after the initial upcycling through separately trained dense models, without requiring large-scale MoE training for unseen data domains. Our experiments show that Nexus achieves a relative gain of up to 2.1% over the baseline for initial upcycling, and a 18.8% relative gain for extending the MoE with a new expert by using limited finetuning data. This flexibility of Nexus is crucial to enable an open-source ecosystem where every user continuously assembles their own MoE-mix according to their needs.", "keywords": [ "Mixture of Experts", "MoE", "Routing", "Efficiency", "Adaptability", "Upcycling" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "WjxgruI6A2", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Arkady Krishtul", "gender": "Male", "institution": "Corsound AI", "country": "IL", "position": "Researcher" }, { "name": "Nir Schwartz", "gender": "Male", "institution": "Corsound AI", "country": "IL", "position": "Researcher" }, { "name": "Oleksandr Solonskyi", "gender": "Male", "institution": "Corsound AI", "country": "IL", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 17, "n_ref": 28, "n_ref_all": 46, "n_fig": 10, "n_tab": 1, "L_tab": 151, "n_element_tab": 25, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1460, "n_element_tab_1": 174, "formula_len_all": 505, "formula_len_all_1": 509, "len_all": 78039, "len_all_1": 57149, "len_abs": 925, "len_title": 131, "len_sents": 23356, "len_sents_1": 20550, "n_sents": 178, "n_sents_1": 146, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 930, "title": "Hearing faces among homogeneous populations: improvement of cross-modal biometrics", "abs": "The relationship between voice and face is well-established in neuroscience and biology. Recent algorithmic advancements have yielded substantial improvements in voice face matching. However, these approaches predominantly achieve success by leveraging datasets with diverse demographic characteristics, which inherently provide greater inter-speaker variability. We address the challenging problem of voice face matching and retrieval in homogeneous datasets, where speakers share gender and ethnicity. Our novel deep architecture, featuring a weighted triplet loss function based on face distances, achieves state-of-the-art performance for voice face matching on these uniform populations. We evaluate our model on a sequence of homogeneous datasets containing only voices and faces of people sharing gender and ethnic group. In addition, we introduce percentile-recall, a new metric for evaluating voice face retrieval tasks.", "keywords": [ "Computer Vision", "Face Recognition", "Speech Recognition", "Cross-modality", "Deep Learning" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 1, 3 ], "presentation_list": [ 2, 1, 3 ], "contribution_list": [ 2, 1, 4 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "WjPK2gj0xu", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Daojing He", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Fangming Liu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Guodong DU", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "Researcher" }, { "name": "Ho-Kin Tang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "Associate Professor" }, { "name": "Jing Li", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Junlin Lee", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "MS student" }, { "name": "Min Zhang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Saleh Alharbi", "gender": "Male", "institution": "Shaqra University", "country": "SA", "position": "Assistant Professor" }, { "name": "Sim Kuan Goh", "gender": "Male", "institution": "Xiamen University Malaysia", "country": "MY", "position": "Assistant Professor" }, { "name": "Yequan Wang", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 44, "n_ref": 126, "n_ref_all": 158, "n_fig": 6, "n_tab": 14, "L_tab": 5585, "n_element_tab": 564, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 6003, "n_element_tab_1": 180, "formula_len_all": 989, "formula_len_all_1": 753, "len_all": 212343, "len_all_1": 74613, "len_abs": 1714, "len_title": 137, "len_sents": 50528, "len_sents_1": 33428, "n_sents": 383, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1747, "title": "Multi-modality Expansion and Retention for LLMs through Parameter Merging and Decoupling", "abs": "Extensive fine-tuning of the synthesis between multimodal encoders and Large Language Models (LLMs) on modality-specific data can expand the modalities that LLM can handle, leading to the formation of Multimodal Large Language Models (MLLMs).\nHowever, this paradigm to expanding modalities heavily relies on initiating fine-tuning from scratch with new multimodal data, which is both resource-intensive and inflexible. \nIn this paper, we propose $\\textit{MMER (Multi-modality Expansion and Retention)}$, a novel $\\textit{training-free}$ approach that reuses and composes existing MLLMs to facilitate effective multimodal expansion while retaining the original performance of each MLLM. \nIn particular, MMER maintains the multimodal encoders of the MLLMs while merging their LLM parameters.\nBy comparing the original LLM parameters with the merged ones, MMER can create binary masks that enable an approximate separation of the LLM parameters for each modality.\nThis process allows the decoupled parameters to independently process modality-specific inputs, thereby reducing parameter conflicts and maintaining the fidelity of the original MLLMs.\nAdditionally, MMER integrates strategies to prevent catastrophic forgetting by employing a similar approach to separately decouple the parameters fine-tuned on new tasks from the original parameters. \nExperiments on three multimodal tasks and fourteen dual-modal tasks show significant improvements over recent baselines, demonstrating that MMER can effectively expand multimodal capabilities of LLMs while retaining 99.6\\% of the original performance. \nFurther experiments in both single-task and cross modalities multi-task scenarios reveal that MMER significantly mitigates catastrophic forgetting.", "keywords": [ "Multimodal Large Language Models", "Model Merging", "Parameter Decoupling", "Knowledge Localization" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "WjKea8bGFF", "primary_area": "reinforcement learning", "n_author": 13, "author_info_list": [ { "name": "Aviv Rosenberg", "gender": "Male", "institution": "Google Research", "country": "", "position": "Researcher" }, { "name": "Bilal Piot", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Chengshuai Shi", "gender": "Male", "institution": "University of Virginia", "country": "", "position": "PhD student" }, { "name": "Chi Jin", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Assistant Professor" }, { "name": "Daniele Calandriello", "gender": "Male", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Jiaming Shen", "gender": "unknown", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Mohammad Saleh", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher & Software Engineer" }, { "name": "Rishabh Joshi", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Tianqi Liu", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Software Engineer" }, { "name": "Tong Zhang", "gender": "Male", "institution": "UIUC", "country": "US", "position": "Full Professor" }, { "name": "Wei Xiong", "gender": "Male", "institution": "Google", "country": "US", "position": "Intern" }, { "name": "Zhen Qin", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Researcher" }, { "name": "Mikhail Khalman", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 56, "n_formula_1": 18, "n_ref_uni": 59, "n_ref": 199, "n_ref_all": 223, "n_fig": 4, "n_tab": 12, "L_tab": 3682, "n_element_tab": 428, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 199, "n_element_tab_1": 17, "formula_len_all": 8305, "formula_len_all_1": 2098, "len_all": 273650, "len_all_1": 63552, "len_abs": 1429, "len_title": 115, "len_sents": 69105, "len_sents_1": 28469, "n_sents": 509, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1435, "title": "Building Math Agents with Multi-Turn Iterative Preference Learning", "abs": "Recent studies have shown that large language models' (LLMs) mathematical problem-solving capabilities can be enhanced by integrating external tools, such as code interpreters, and employing multi-turn Chain-of-Thought (CoT) reasoning. While current methods focus on synthetic data generation and Supervised Fine-Tuning (SFT), this paper studies the complementary direct preference learning approach to further improve model performance. However, existing direct preference learning algorithms are originally designed for the single-turn chat task, and do not fully address the complexities of multi-turn reasoning and external tool integration required for tool-integrated mathematical reasoning tasks. To fill in this gap, we introduce a multi-turn direct preference learning framework, tailored for this context, that leverages feedback from code interpreters and optimizes trajectory-level preferences. This framework includes multi-turn DPO and multi-turn KTO as specific implementations. The effectiveness of our framework is validated through training of various language models using an augmented prompt set from the GSM8K and MATH datasets. Our results demonstrate substantial improvements: a supervised fine-tuned Gemma-1.1-it-7B model's performance increased from 77.5% to 83.9% on GSM8K and from 46.1% to 51.2% on MATH. Similarly, a Gemma-2-it-9B model improved from 84.1% to 86.3% on GSM8K and from 51.0% to 54.5% on MATH.", "keywords": [ "large language model", "RLHF", "math" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "WjDjem8mWE", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Weiwei Xing", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Yue Cheng", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhanxing Zhu", "gender": "Male", "institution": "University of Southampton", "country": "", "position": "Associate Professor" }, { "name": "Bochen Lv", "gender": "Male", "institution": "University of Southampton", "country": "GB", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 25, "n_ref_uni": 38, "n_ref": 84, "n_ref_all": 117, "n_fig": 13, "n_tab": 11, "L_tab": 8804, "n_element_tab": 717, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 3001, "n_element_tab_1": 250, "formula_len_all": 1818, "formula_len_all_1": 1565, "len_all": 170010, "len_all_1": 64077, "len_abs": 1263, "len_title": 106, "len_sents": 47534, "len_sents_1": 25729, "n_sents": 441, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1268, "title": "DyCAST: Learning Dynamic Causal Structure from Time Series", "abs": "Understanding the dynamics of causal structures is crucial for uncovering the underlying processes in time series data. Previous approaches rely on static assumptions, where contemporaneous and time-lagged dependencies are assumed to have invariant topological structures. However, these models fail to capture the evolving causal relationship between variables when the underlying process exhibits such dynamics. To address this limitation, we propose DyCAST, a novel framework designed to learn dynamic causal structures in time series using Neural Ordinary Differential Equations (Neural ODEs). The key innovation lies in modeling the temporal dynamics of the contemporaneous structure, drawing inspiration from recent advances in Neural ODEs on constrained manifolds. We reformulate the task of learning causal structures at each time step as solving the solution trajectory of a Neural ODE on the directed acyclic graph (DAG) manifold. To accommodate high-dimensional causal structures, we extend DyCAST by learning the temporal dynamics of the hidden state for contemporaneous causal structure. Experiments on both synthetic and real-world datasets demonstrate that DyCAST achieves superior or comparable performance compared to existing causal discovery models.", "keywords": [ "dynamic causal discovery; time series" ], "rating_list": [ 3, 6, 8, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "WioQ6tSzvr", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Chaoyu Gong", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "PhD student" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" } ], "n_formula": 27, "n_formula_1": 26, "n_ref_uni": 61, "n_ref": 117, "n_ref_all": 143, "n_fig": 7, "n_tab": 8, "L_tab": 4772, "n_element_tab": 593, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2488, "n_element_tab_1": 229, "formula_len_all": 2373, "formula_len_all_1": 1709, "len_all": 202161, "len_all_1": 75462, "len_abs": 1294, "len_title": 103, "len_sents": 51138, "len_sents_1": 29431, "n_sents": 558, "n_sents_1": 236, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1302, "title": "Trusted and Interactive Clustering for Time-Series Data", "abs": "Time-series clustering has gained abundant popularity and has been used in diverse scientific areas. However, few researchers take an information fusion perspective to combine information from the time and frequency domains to accomplish clustering, although these two domains offer distinct and complementary characteristics of time-series. Motivated by this issue, we propose a trusted and interactive model, which leverages evidence theory to combine time- and frequency-based clustering results produced by the corresponding contrastive learning module. After mathematizing clustering results from the two domains as mass functions, the uncertainty contained in these results can be quantified at the sample-specific level. The combined result thus promotes clustering reliability, and is optimized based on the pseudo-labels generated by k-means in an interactive learning paradigm. Both theoretical analysis and experimental results on 136 benchmark datasets validate the effectiveness of the proposed model in clustering performance. Extensive ablation experiments demonstrate the contribution of combining information from the time and frequency domains and using the interactive learning paradigm. The embeddings learned are also experimentally shown to perform well in other downstream tasks.", "keywords": [ "Time-series clustering", "evidence theory", "information fusion" ], "rating_list": [ 3, 6, 3 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "Wi74fYCX2f", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Bruno Galerne", "gender": "Male", "institution": "Universit\u00e9 d'Orl\u00e9ans", "country": "FR", "position": "Full Professor" }, { "name": "Emile Pierret", "gender": "Male", "institution": "Universit\u00e9 d'Orl\u00e9ans", "country": "", "position": "PhD student" } ], "n_formula": 116, "n_formula_1": 32, "n_ref_uni": 26, "n_ref": 68, "n_ref_all": 113, "n_fig": 5, "n_tab": 7, "L_tab": 2284, "n_element_tab": 419, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 891, "n_element_tab_1": 160, "formula_len_all": 8746, "formula_len_all_1": 2461, "len_all": 162272, "len_all_1": 68991, "len_abs": 1214, "len_title": 131, "len_sents": 42536, "len_sents_1": 27234, "n_sents": 396, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1219, "title": "Diffusion models for Gaussian distributions: Exact solutions and Wasserstein errors", "abs": "Diffusion or score-based models recently showed high performance in image generation.\nThey rely on a forward and a backward stochastic differential equations (SDE). The sampling of a data distribution is achieved by solving numerically the backward SDE or its associated flow ODE.\nStudying the convergence of these models necessitates to control four different types of error: the initialization error, the truncation error, the discretization and the score approximation.\nIn this paper, we study theoretically the behavior of diffusion models and their numerical implementation when the data distribution is Gaussian.\nIn this restricted framework where the score function is a linear operator, we derive the analytical solutions of the backward SDE and the probability flow ODE.\nWe prove that these solutions and their discretizations are all Gaussian processes, which allows us to compute exact Wasserstein errors induced by each error type for any sampling scheme.\nMonitoring convergence directly in the data space instead of relying on Inception features, our experiments show that the recommended numerical schemes from the diffusion models literature are also the best sampling schemes for Gaussian distributions.", "keywords": [ "Diffusion models", "image generation", "differential equations", "discretization schemes" ], "rating_list": [ 6, 8, 3, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "WhvTLognS0", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Jegyeong Cho", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Kuk-Jin Yoon", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" }, { "name": "Wooseong Jeong", "gender": "Male", "institution": "The Korea Advanced Institute of Science and Technology (KAIST)", "country": "KR", "position": "PhD student" }, { "name": "Youngho Yoon", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 11, "n_ref_uni": 41, "n_ref": 73, "n_ref_all": 103, "n_fig": 7, "n_tab": 5, "L_tab": 2569, "n_element_tab": 223, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1702, "n_element_tab_1": 133, "formula_len_all": 2146, "formula_len_all_1": 926, "len_all": 144373, "len_all_1": 68694, "len_abs": 1580, "len_title": 95, "len_sents": 39303, "len_sents_1": 31779, "n_sents": 279, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 46, "L_abs": 1619, "title": "Learning Task Relations for Test-Time Training", "abs": "Generalizing deep neural networks to unseen target domains presents a major challenge in real-world deployments. Test-time training (TTT) addresses this is- sue by using an auxiliary self-supervised task to reduce the gap between source and target domains caused by distribution shifts during deployment. Previous re- search relies on the assumption that the adopted auxiliary task would be beneficial to the target task we want to adapt. However, this situation is not guaranteed as each task has a different objective, thus adaptation relies on the relation be- tween the tasks. This limitation has motivated us to introduce a more generalized framework: Task Relation Learning for Test-time Training (TR-TTT), which can be applied to multiple tasks concurrently. Our key assumption is that task re- lations are crucial information for successful test-time training, and we capture these relations using a Task Relation Learner (TRL). We model task relations as conditional probabilities by predicting the label of a target task based on the latent spaces of other task-specific features. By leveraging these relations, the network can more effectively handle distribution shifts and improve post-adaptation perfor- mance across various tasks\u2014both classification and regression\u2014unlike previous methods focused mainly on simple classification. To validate our approach, we ap- ply TR-TTT to conventional multi-task benchmarks, integrating it with the tradi- tional TTT experimental protocol. Our empirical results demonstrate that TR-TTT significantly outperforms state-of-the-art methods across a range of benchmarks.", "keywords": [ "Test-time Training", "Task Relation Learning", "Multi-task Learning" ], "rating_list": [ 5, 3, 6, 6, 5 ], "soundness_list": [ 2, 1, 3, 3, 3 ], "presentation_list": [ 3, 1, 4, 3, 3 ], "contribution_list": [ 2, 1, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "WhPLUfThB4", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Tian Xie", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Xueru Zhang", "gender": "Female", "institution": "Ohio State University", "country": "US", "position": "Assistant Professor" }, { "name": "Xuwei Tan", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Zhiqun Zuo", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 51, "n_ref": 99, "n_ref_all": 119, "n_fig": 3, "n_tab": 17, "L_tab": 6185, "n_element_tab": 575, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1504, "n_element_tab_1": 134, "formula_len_all": 1156, "formula_len_all_1": 1017, "len_all": 199198, "len_all_1": 68788, "len_abs": 1219, "len_title": 106, "len_sents": 57348, "len_sents_1": 32482, "n_sents": 582, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1222, "title": "Exploiting Open-World Data for Adaptive Continual Learning", "abs": "Continual learning (CL), which involves learning from sequential tasks without forgetting, is mainly explored in supervised learning settings where all data are labeled. However, high-quality labeled data may not be readily available at a large scale due to high labeling costs, making the application of existing CL methods in real-world scenarios challenging. In this paper, we study a more practical facet of CL: open-world continual learning, where the training data comes from the open-world dataset and is partially labeled and non-i.i.d. Building on the insight that task shifts in CL can be viewed as distribution transitions from known classes to novel classes, we propose OpenACL, a method that explicitly leverages novel classes in unlabeled data to enhance continual learning. Specifically, OpenACL considers novel classes within open-world data as potential classes for upcoming tasks and mines the underlying pattern from them to empower the model's adaptability to upcoming tasks. Furthermore, learning from extensive unlabeled data also helps to tackle the issue of catastrophic forgetting. Extensive experiments validate the effectiveness of OpenACL and show the benefit of learning from open-world data.", "keywords": [ "Continual Learning", "Semi-supervised Learning", "Open-world" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "WhIuLQWCWS", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "David Saulpic", "gender": "unknown", "institution": "CNRS", "country": "FR", "position": "Assistant Professor" }, { "name": "Jonathan Scott", "gender": "unknown", "institution": "Institute of Science and Technology", "country": "AT", "position": "PhD student" }, { "name": "Christoph H Lampert", "gender": "Male", "institution": "Institute of Science and Technology Austria", "country": "AT", "position": "Professor" } ], "n_formula": 70, "n_formula_1": 10, "n_ref_uni": 40, "n_ref": 93, "n_ref_all": 120, "n_fig": 8, "n_tab": 9, "L_tab": 2013, "n_element_tab": 222, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 2, "n_element_tab_1": 1, "formula_len_all": 6141, "formula_len_all_1": 666, "len_all": 264710, "len_all_1": 79164, "len_abs": 1115, "len_title": 110, "len_sents": 85038, "len_sents_1": 35964, "n_sents": 732, "n_sents_1": 290, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1119, "title": "Differentially Private Federated $k$-Means with Server-Side Data", "abs": "Clustering has long been a cornerstone of data analysis. It is particularly suited to identifying coherent subgroups or substructures in unlabeled data, as are generated continuously in large amounts these days. However, in many cases traditional clustering methods are not applicable, because data are increasingly being produced and stored in a distributed way, e.g. on edge devices, and privacy concerns prevent it from being transferred to a central server. To address this challenge, we present FedDP-KMeans, a new algorithm for k-means clustering that is fully-federated as well as differentially private. Our approach leverages (potentially small and out-of-distribution) server-side data to overcome the primary challenge of differentially private clustering methods: the need for a good initialization. Combining our initialization with a simple federated DP-Lloyds algorithm we obtain an algorithm that achieves excellent results on synthetic and real-world benchmark tasks. We also provide a theoretical analysis of our method that provides bounds on the convergence speed and cluster identification success.", "keywords": [ "Clustering", "Differential Privacy", "Federated Learning" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "Wh4SE2S7Mo", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Liu Ziyin", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "YiZhou Xu", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" } ], "n_formula": 68, "n_formula_1": 38, "n_ref_uni": 48, "n_ref": 78, "n_ref_all": 144, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5256, "formula_len_all_1": 2713, "len_all": 169458, "len_all_1": 71934, "len_abs": 1706, "len_title": 104, "len_sents": 49330, "len_sents_1": 31040, "n_sents": 444, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1138, "title": "Three Mechanisms of Feature Learning in a Linear Network", "abs": "Understanding the dynamics of neural networks in different width regimes is crucial for improving their training and performance. We present an exact solution for the learning dynamics of a one-hidden-layer linear network, with one-dimensional data, across any finite width, uniquely exhibiting both kernel and feature learning phases. This study marks a technical advancement by enabling the analysis of the training trajectory from any initialization and a detailed phase diagram under varying common hyperparameters such as width, layer-wise learning rates, and scales of output and initialization. We identify three novel prototype mechanisms specific to the feature learning regime: (1) learning by alignment, (2) learning by disalignment, and (3) learning by rescaling, which contrast starkly with the dynamics observed in the kernel regime. Our theoretical findings are substantiated with empirical evidence showing that these mechanisms also manifest in deep nonlinear networks handling real-world tasks, enhancing our understanding of neural network training dynamics and guiding the design of more effective learning strategies.", "keywords": [ "solvable model", "feature learning", "neural tangent kernel" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "WgpAFnjvPr", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Dingkang Yang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Dongling Xiao", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Jiawei Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Ke Li", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Senior Researcher" }, { "name": "Lihua Zhang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Mingcheng Li", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Shunli Wang", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Tong Wu", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "", "position": "Researcher" }, { "name": "Xiaolu Hou", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Yue Jiang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 15, "n_ref": 28, "n_ref_all": 65, "n_fig": 14, "n_tab": 8, "L_tab": 4920, "n_element_tab": 763, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2188, "n_element_tab_1": 313, "formula_len_all": 62, "formula_len_all_1": 35, "len_all": 199247, "len_all_1": 68258, "len_abs": 1777, "len_title": 128, "len_sents": 70493, "len_sents_1": 31870, "n_sents": 540, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1738, "title": "Detecting and Evaluating Medical Hallucinations in Large Vision Language Models", "abs": "Large Vision Language Models (LVLMs) are increasingly integral to healthcare applications, including medical visual question answering and imaging report generation. While these models inherit the robust capabilities of foundational Large Language Models (LLMs), they also inherit susceptibility to hallucinations\u2014a significant concern in high-stakes medical contexts where the margin for error is minimal. However, currently, there are no dedicated methods or benchmarks for hallucination detection and evaluation in the medical field. To bridge this gap, we introduce Med-HallMark, the first benchmark specifically designed for hallucination detection and evaluation within the medical multimodal domain. This benchmark provides multi-tasking hallucination support, multifaceted hallucination data, and hierarchical hallucination categorization. Furthermore, we propose the MediHall Score, a new medical evaluative metric designed to assess LVLMs' hallucinations through a hierarchical scoring system that considers the severity and type of hallucination, thereby enabling a granular assessment of potential clinical impacts. We also present MedihallDetector, a novel Medical LVLM engineered for precise hallucination detection, which employs multitask training for hallucination detection. Through extensive experimental evaluations, we establish baselines for popular LVLMs using our benchmark. The findings indicate that MediHall Score provides a more nuanced understanding of hallucination impacts compared to traditional metrics and demonstrate the enhanced performance of MedihallDetector. We hope this work can significantly improve the reliability of LVLMs in medical applications. All resources of this work have been released.", "keywords": [ "Hallucination Benchamrk", "Hallucination Evaluation Method", "Medical Large Vision Language Model" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "WfxPVtYRlL", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Deniz Oktay", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Research Intern" }, { "name": "Nick Richardson", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Olga Solodova", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Ryan P Adams", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Professor" } ], "n_formula": 54, "n_formula_1": 21, "n_ref_uni": 34, "n_ref": 97, "n_ref_all": 132, "n_fig": 1, "n_tab": 4, "L_tab": 436, "n_element_tab": 39, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 193, "n_element_tab_1": 25, "formula_len_all": 3462, "formula_len_all_1": 1191, "len_all": 170337, "len_all_1": 71249, "len_abs": 904, "len_title": 34, "len_sents": 63748, "len_sents_1": 35242, "n_sents": 473, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 34, "L_abs": 920, "title": "Graph Neural Networks Gone Hogwild", "abs": "Graph neural networks (GNNs) appear to be powerful tools to learn state representations for agents in distributed, decentralized multi-agent systems, but generate catastrophically incorrect predictions when nodes update asynchronously during inference.\n This failure under asynchrony effectively excludes these architectures from many potential applications where synchrony is difficult or impossible to enforce, e.g., robotic swarms or sensor networks.\n In this work we identify ''implicitly-defined'' GNNs as a class of architectures which is provably robust to asynchronous ''hogwild'' inference, adapting convergence guarantees from work in asynchronous and distributed optimization. \n We then propose a novel implicitly-defined GNN architecture, which we call an energy GNN. \n We show that this architecture outperforms other GNNs from this class on a variety of synthetic tasks inspired by multi-agent systems.", "keywords": [ "graph neural networks", "multi-agent", "asynchronous", "decentralized" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "Wfw4ypsgRZ", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Jason Matthew Klusowski", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Omar Hagrass", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Sijin Chen", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 14, "n_ref_uni": 25, "n_ref": 40, "n_ref_all": 46, "n_fig": 0, "n_tab": 1, "L_tab": 1726, "n_element_tab": 196, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1538, "formula_len_all_1": 1081, "len_all": 118276, "len_all_1": 69922, "len_abs": 1407, "len_title": 124, "len_sents": 32744, "len_sents_1": 29289, "n_sents": 273, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1414, "title": "Decoding Game: On Minimax Optimality of Heuristic Text Generation Strategies", "abs": "Decoding strategies play a pivotal role in text generation for modern language models, yet a puzzling gap divides theory and practice. Surprisingly, strategies that should intuitively be optimal, such as Maximum a Posteriori (MAP), often perform poorly in practice. Meanwhile, popular heuristic approaches like Top-$k$ and Nucleus sampling, which employ truncation and normalization of the conditional next-token probabilities, have achieved great empirical success but lack theoretical justifications. In this paper, we propose Decoding Game, a comprehensive theoretical framework which reimagines text generation as a two-player zero-sum game between Strategist, who seeks to produce text credible in the true distribution, and Nature, who distorts the true distribution adversarially. After discussing the decomposibility of multi-step generation, we derive the optimal strategy in closed form for one-step Decoding Game. It is shown that the adversarial Nature imposes an implicit regularization on likelihood maximization, and truncation-normalization methods are first-order approximations to the optimal strategy under this regularization. Additionally, by generalizing the objective and parameters of Decoding Game, near-optimal strategies encompass diverse methods such as greedy search, temperature scaling, and hybrids thereof. Numerical experiments are conducted to complement our theoretical analysis.", "keywords": [ "decoding strategy", "text generation", "sampling", "optimality", "game theory", "robust optimization", "implicit regularization", "sparsity" ], "rating_list": [ 8, 5, 8, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 2 ] }, { "paper_id": "WfaQrKCr4X", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Axel Feldmann", "gender": "Male", "institution": "Computer Science and Artificial Intelligence Laboratory, Electrical Engineering & Computer Science", "country": "US", "position": "PhD student" }, { "name": "Mark Hamilton", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Shaden Naif Alshammari", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Bill t. Freeman", "gender": "Male", "institution": "Google", "country": "US", "position": "Director" }, { "name": "john r hershey", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 87, "n_formula_1": 11, "n_ref_uni": 26, "n_ref": 45, "n_ref_all": 67, "n_fig": 3, "n_tab": 9, "L_tab": 3514, "n_element_tab": 190, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6089, "formula_len_all_1": 899, "len_all": 181446, "len_all_1": 53182, "len_abs": 3072, "len_title": 103, "len_sents": 61554, "len_sents_1": 26506, "n_sents": 515, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 1100, "title": "A Unifying Framework for Representation Learning", "abs": "As the field of unsupervised learning grows, there has been a proliferation of different loss functions to solve different classes of problems. We find that a large collection of modern loss functions can be generalized by a single equation rooted in information theory. In particular, we introduce I-Con, a framework that shows that several broad classes of machine learning methods are precisely minimizing an integrated KL divergence between two marginal distributions: the supervisory and learned representations. This viewpoint exposes a hidden information geometry underlying clustering, spectral methods, dimensionality reduction, contrastive learning, and supervised learning. I-Con enables the development of new loss functions by combining successful techniques from across the literature. We not only present a wide array of proofs, connecting over 11 different approaches, but we also leverage these theoretical results to create state of the art unsupervised image classifiers that achieve a +8\\% improvement over the prior state-of-the-art on unsupervised classification on ImageNet-1K.", "keywords": [ "representation learning; unsupervised learning; clustering; dimensionality reduction; kmeans; contrastive learning; information theory" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 4, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "Wf2ndb8nhf", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Anca Dragan", "gender": "Female", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Brendan Murphy", "gender": "Male", "institution": "FAR AI", "country": "US", "position": "Researcher" }, { "name": "Constantin Weisser", "gender": "Male", "institution": "Haize Labs", "country": "US", "position": "Researcher" }, { "name": "Marcus Williams", "gender": "Male", "institution": "MATS", "country": "US", "position": "Researcher" }, { "name": "Micah Carroll", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Adhyyan Manish Narang", "gender": "Male", "institution": "University of Washington, Seattle", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 51, "n_ref": 116, "n_ref_all": 187, "n_fig": 31, "n_tab": 5, "L_tab": 2244, "n_element_tab": 95, "n_fig_1": 18, "n_tab_1": 1, "L_tab_1": 428, "n_element_tab_1": 7, "formula_len_all": 431, "formula_len_all_1": 0, "len_all": 403878, "len_all_1": 84617, "len_abs": 4282, "len_title": 125, "len_sents": 155868, "len_sents_1": 41879, "n_sents": 1328, "n_sents_1": 314, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1436, "title": "Targeted Manipulation and Deception Emerge in LLMs Trained on User* Feedback", "abs": "As LLMs become more widely deployed, there is increasing interest in directly optimizing for feedback from end users (e.g. thumbs up) in addition to feedback from paid annotators. However, training to maximize human feedback creates a perverse incentive structure for the AI to resort to manipulative or deceptive tactics to obtain positive feedback from users who are vulnerable to such strategies. We study this phenomenon by training LLMs with Reinforcement Learning with simulated user feedback. We have three main findings in our environments which simulate domains of practical LLM usage: 1) Extreme forms of \u201cfeedback gaming\u201d such as manipulation and deception are learned reliably; 2) Even if only 2% of users are vulnerable to manipulative strategies, LLMs learn to identify and target them while behaving appropriately with other users, making such behaviors harder to detect; 3) To mitigate this issue, it may seem promising to leverage continued safety training or LLM-as-judges during training to filter problematic outputs. Instead, we found that while such approaches help in some of our settings, they backfire in others, sometimes even leading to subtler manipulative behaviors. We hope our results can serve as a cautionary case study highlighting the risks of using gameable feedback sources \u2013 such as user feedback \u2013 as a target for RL. Our code is publicly available. Warning: some of our examples may be upsetting.", "keywords": [ "manipulation", "deception", "alignment", "reward hacking", "user feedback" ], "rating_list": [ 8, 6, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "WesOWkjJmg", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Abishek Sankararaman", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Aodong Li", "gender": "Male", "institution": "Amazon Web Services", "country": "US", "position": "Researcher" }, { "name": "Murali Balakrishnan", "gender": "Male", "institution": "Amazon", "country": "", "position": "Principal Researcher" } ], "n_formula": 35, "n_formula_1": 10, "n_ref_uni": 45, "n_ref": 87, "n_ref_all": 156, "n_fig": 13, "n_tab": 4, "L_tab": 939, "n_element_tab": 103, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 858, "n_element_tab_1": 10, "formula_len_all": 3385, "formula_len_all_1": 857, "len_all": 233943, "len_all_1": 76907, "len_abs": 1757, "len_title": 111, "len_sents": 80454, "len_sents_1": 34457, "n_sents": 775, "n_sents_1": 313, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1767, "title": "Probabilistic Hash Embeddings for Temporal Tabular Data Streams", "abs": "We study temporal tabular data-streams (TTD) where each observation has both categorical and numerical values, and where the universe of distinct categorical items is not known upfront and can even grow unboundedly over time. Such data is common in many large-scale systems, such as user activity in computer system logs and scientific experiment records. Feature hashing is commonly used as a pre- processing step to map the categorical items into a known universe, before doing representation learning (Coleman et al., 2024; Desai et al., 2022). However, these methods have been developed and evaluated for the offline or batch settings. In this paper, we consider the pre-processing step of hashing before representation learning in the online setting for TTD. We show that deterministic embeddings suffer from forgetting in online learning with TTD, leading to performance deterioration. To mitigate the issue, we propose a probabilistic hash embedding (PHE) model that treats hash embeddings as stochastic and applies Bayesian online learning to learn incrementally with data. Based on the structure of PHE, we derive a scalable inference algorithm to learn model parameters and infer/update the posteriors of hash embeddings and other latent variables. Our algorithm (i) can handle evolving vocabulary of categorical items, (ii) is adaptive to new items without forgetting old items, (iii) is implementable with a bounded set of parameters that does not grow with the number of distinct observed items on the stream, and (iv) is efficiently implementable both in the offline and the online streaming setting. Experiments in classification, sequence modeling, and recommendation systems with TTD demonstrate the superior performance of PHE compared to baselines.", "keywords": [ "hash embedding", "Bayesian online learning", "tabular data", "dynamic vocabulary" ], "rating_list": [ 5, 5, 3, 8 ], "soundness_list": [ 2, 2, 2, 4 ], "presentation_list": [ 1, 2, 1, 3 ], "contribution_list": [ 3, 2, 2, 4 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "WeJEidTzff", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Can Rong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jingtao Ding", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Yan Liu", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "Professor" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 6, "n_ref_uni": 38, "n_ref": 82, "n_ref_all": 119, "n_fig": 20, "n_tab": 6, "L_tab": 3368, "n_element_tab": 324, "n_fig_1": 7, "n_tab_1": 7, "L_tab_1": 2366, "n_element_tab_1": 228, "formula_len_all": 1019, "formula_len_all_1": 210, "len_all": 183110, "len_all_1": 69218, "len_abs": 1333, "len_title": 133, "len_sents": 62028, "len_sents_1": 31212, "n_sents": 463, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1347, "title": "A Large-scale Dataset and Benchmark for Commuting Origin-Destination Flow Generation", "abs": "Commuting Origin-Destination~(OD) flows are critical inputs for urban planning and transportation, providing crucial information about the population residing in one region and working in another within an interested area. Due to the high cost of data collection, researchers have developed physical and computational models to generate commuting OD flows using readily available urban attributes, such as sociodemographics and points of interest, for cities lacking historical OD flows \\textemdash commuting OD flow generation. Existing works developed models based on different techniques and achieved improvement on different datasets with different evaluation metrics, which hinderes establishing a unified standard for comparing model performance. To bridge this gap, we introduce a large-scale dataset containing commuting OD flows for 3,333 areas including a wide range of urban environments around the United States. Based on that, we benchmark widely used models for commuting OD flow generation. We surprisingly find that the network-based generative models achieve the optimal performance in terms of both precision and generalization ability, which may inspire new research directions of graph generative modeling in this field. The dataset and benchmark are available at https://anonymous.4open.science/r/CommutingODGen-Dataset-0D4C/.", "keywords": [ "Commuting", "origin-destination flow dataset", "urban computing", "weighted graph modeling" ], "rating_list": [ 8, 6, 8, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 3, 5, 5 ] }, { "paper_id": "We5z3UEnUY", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Dung Nguyen", "gender": "unknown", "institution": "Deakin University", "country": "", "position": "PhD student" }, { "name": "Svetha Venkatesh", "gender": "Female", "institution": "Deakin University", "country": "AU", "position": "Full Professor" }, { "name": "Hung Thai Le", "gender": "Male", "institution": "Deakin University", "country": "", "position": "Lecturer" }, { "name": "Kien Duc Do", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Research Fellow" }, { "name": "Sunil Gupta", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Full Professor" } ], "n_formula": 40, "n_formula_1": 22, "n_ref_uni": 31, "n_ref": 67, "n_ref_all": 81, "n_fig": 7, "n_tab": 4, "L_tab": 544, "n_element_tab": 59, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 63, "n_element_tab_1": 6, "formula_len_all": 4070, "formula_len_all_1": 2044, "len_all": 118282, "len_all_1": 64379, "len_abs": 1212, "len_title": 88, "len_sents": 38153, "len_sents_1": 29559, "n_sents": 315, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1217, "title": "Stable Hadamard Memory: Revitalizing Memory-Augmented Agents for Reinforcement Learning", "abs": "Effective decision-making in partially observable environments demands robust memory management. Despite their success in supervised learning, current deep-learning memory models struggle in reinforcement learning environments that are partially observable and long-term. They fail to efficiently capture relevant past information, adapt flexibly to changing observations, and maintain stable updates over long episodes. We theoretically analyze the limitations of existing memory models within a unified framework and introduce the Stable Hadamard Memory, a novel memory model for reinforcement learning agents. Our model dynamically adjusts memory by erasing no longer needed experiences and reinforcing crucial ones computationally efficiently. To this end, we leverage the Hadamard product for calibrating and updating memory, specifically designed to enhance memory capacity while mitigating numerical and learning challenges. Our approach significantly outperforms state-of-the-art memory-based methods on challenging partially observable benchmarks, such as meta-reinforcement learning, long-horizon credit assignment, and POPGym, demonstrating superior performance in handling long-term and evolving contexts.", "keywords": [ "Reinforcement Learning", "Memory", "POMDP" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "Wd1R0oxe5j", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "David Beamish", "gender": "Male", "institution": "University of Bath", "country": "GB", "position": "MS student" }, { "name": "Georgios Exarchakis", "gender": "Male", "institution": "University of Bath", "country": "GB", "position": "Lecturer" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 77, "n_ref": 183, "n_ref_all": 187, "n_fig": 0, "n_tab": 2, "L_tab": 943, "n_element_tab": 158, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 943, "n_element_tab_1": 158, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 227335, "len_all_1": 73916, "len_abs": 2589, "len_title": 109, "len_sents": 61839, "len_sents_1": 37042, "n_sents": 469, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 883, "title": "Large Language Models for Explainability in Machine Learning", "abs": "We investigate the potential of large language models (LLMs) in explainable artificial intelligence (XAI) by examining their ability to generate understandable explanations for machine learning (ML) models. While recent studies suggest that LLMs could effectively address the limitations of traditional explanation methods through their conversational capabilities, there has been a lack of systematic evaluation of the quality of these LLM-generated explanations. To fill this gap, this study evaluates whether LLMs can produce explanations for ML models that meet the fundamental properties of XAI using conventional ML models and explanation methods as benchmarks. The findings offer important insights into the strengths and limitations of LLMs as tools for explainable AI, provide recommendations for their appropriate use, and identify promising directions for future research.", "keywords": [ "XAI", "explainability", "large language models" ], "rating_list": [ 5, 5, 3, 1 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "Wd1OmOwL0C", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Berker Demirel", "gender": "Male", "institution": "Institute of Science and Technology", "country": "AT", "position": "PhD student" }, { "name": "Celestine Mendler-D\u00fcnner", "gender": "unknown", "institution": "ELLIS Institute T\u00fcbingen", "country": "", "position": "Principal Investigator" }, { "name": "Francesco Locatello", "gender": "Male", "institution": "Institute of Science and Technology", "country": "AT", "position": "Assistant Professor" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Lingjing Kong", "gender": "Male", "institution": "Computer Science Department, School of Computer Science", "country": "US", "position": "PhD student" }, { "name": "Theofanis Karaletsos", "gender": "Male", "institution": "Pyramidal, Inc", "country": "", "position": "Researcher" } ], "n_formula": 43, "n_formula_1": 6, "n_ref_uni": 75, "n_ref": 117, "n_ref_all": 134, "n_fig": 7, "n_tab": 4, "L_tab": 1005, "n_element_tab": 114, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 412, "n_element_tab_1": 31, "formula_len_all": 47212, "formula_len_all_1": 309, "len_all": 250253, "len_all_1": 89828, "len_abs": 1442, "len_title": 50, "len_sents": 66556, "len_sents_1": 52464, "n_sents": 372, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1457, "title": "Adjusting Pretrained Backbones for Performativity", "abs": "With the widespread deployment of deep learning models, they influence their environment in various ways. The induced distribution shifts can lead to unexpected performance degradation in deployed models. Existing methods to anticipate performativity typically incorporate information about the deployed model into the feature vector when predicting future outcomes. While enjoying appealing theoretical properties, modifying the input dimension of the prediction task is often not practical. To address this, we propose a novel technique to adjust pretrained backbones for performativity in a modular way, achieving better sample efficiency and enabling the reuse of existing deep learning assets. Focusing on performative label shift, the key idea is to train a shallow adapter module to perform a \\emph{Bayes-optimal} label shift correction to the backbone's logits given a sufficient statistic of the model to be deployed. As such, our framework decouples the construction of input-specific feature embeddings from the mechanism governing performativity. Motivated by dynamic benchmarking as a use-case, we evaluate our approach under adversarial sampling, for vision and language tasks. We show how it leads to smaller loss along the retraining trajectory and enables us to effectively select among candidate models to anticipate performance degradations. More broadly, our work provides a first baseline for addressing performativity in deep learning.", "keywords": [ "performative prediction", "performative shift", "label shift", "performativity", "dynamic benchmarks" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 2, 4, 3, 2 ] }, { "paper_id": "WcZLG8XxhD", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Anders Aamand", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Hao WU", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Postdoc" }, { "name": "Sandeep Silwal", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Siddharth Gollapudi", "gender": "unknown", "institution": "Research, Microsoft", "country": "", "position": "Intern" }, { "name": "Justin Y Chen", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 110, "n_formula_1": 28, "n_ref_uni": 27, "n_ref": 77, "n_ref_all": 89, "n_fig": 2, "n_tab": 5, "L_tab": 835, "n_element_tab": 47, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 502, "n_element_tab_1": 62, "formula_len_all": 10902, "formula_len_all_1": 1694, "len_all": 174787, "len_all_1": 65926, "len_abs": 1474, "len_title": 86, "len_sents": 52310, "len_sents_1": 29290, "n_sents": 508, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 38, "L_abs": 1478, "title": "Learning-Augmented Frequent Directions", "abs": "An influential paper of Hsu et al. (ICLR'19) introduced the study of learning-augmented streaming algorithms in the context of frequency estimation. A fundamental problem in the streaming literature, the goal of frequency estimation is to approximate the number of occurrences of items appearing in a long stream of data using only a small amount of memory. Hsu et al. develop a natural framework to combine the worst-case guarantees of popular solutions such as CountMin and CountSketch with learned predictions of high frequency elements. They demonstrate that learning the underlying structure of data can be used to yield better streaming algorithms, both in theory and practice.\n\nWe simplify and generalize past work on learning-augmented frequency estimation. Our first contribution is a learning-augmented variant of the Misra-Gries algorithm which improves upon the error of learned CountMin and learned CountSketch and achieves the state-of-the-art performance of randomized algorithms (Aamand et al., NeurIPS'23) with a simpler, deterministic algorithm. Our second contribution is to adapt learning-augmentation to a high-dimensional generalization of frequency estimation corresponding to finding important directions (top singular vectors) of a matrix given its rows one-by-one in a stream. We analyze a learning-augmented variant of the Frequent Directions algorithm, extending the theoretical and empirical understanding of learned predictions to matrix streaming.", "keywords": [ "learning-augmented algorithms", "algorithms with predictions", "data streams", "streaming algorithms", "frequency estimation", "heavy hitters", "frequent directions", "low-rank approximation" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 4, 4, 2, 4 ], "presentation_list": [ 4, 4, 2, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "WbqBj2aC5k", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Can Chen", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Jun-Kun Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" } ], "n_formula": 162, "n_formula_1": 16, "n_ref_uni": 71, "n_ref": 166, "n_ref_all": 242, "n_fig": 25, "n_tab": 4, "L_tab": 1486, "n_element_tab": 310, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 14360, "formula_len_all_1": 980, "len_all": 295768, "len_all_1": 76184, "len_abs": 1288, "len_title": 129, "len_sents": 97142, "len_sents_1": 37231, "n_sents": 820, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1291, "title": "Online Detecting LLM-Generated Texts via Sequential Hypothesis Testing by Betting", "abs": "Developing algorithms to differentiate between machine-generated texts and human-written texts has garnered substantial attention in recent years. Existing methods in this direction typically concern an offline setting where a dataset containing a mix of real and machine-generated texts is given upfront, and the task is to determine whether each sample in the dataset is from a large language model (LLM) or a human. However, in many practical scenarios, sources such as news websites, social media accounts, or on other forums publish content in a streaming fashion. Therefore, in this online scenario, how to quickly and accurately determine whether the source is an LLM with strong statistical guarantees is crucial for these media or platforms to function effectively and prevent the spread of misinformation and other potential misuse of LLMs. To tackle the problem of online detection, we develop an algorithm based on the techniques of sequential hypothesis testing by betting that not only builds upon and complements existing offline detection techniques but also enjoys statistical guarantees, which include a controlled false positive rate and the expected time to correctly identify a source as an LLM. Experiments were conducted to demonstrate the effectiveness of our method.", "keywords": [ "Large Language Models (LLMs)", "Machine-Generated Text", "Text Generation Detection", "Sequential Hypothesis Testing", "Online Optimization" ], "rating_list": [ 5, 3, 5, 6, 8 ], "soundness_list": [ 3, 2, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 3, 2, 3, 3, 3 ] }, { "paper_id": "Wb6Mcmo0ch", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Hanxian Huang", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Jishen Zhao", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Associate Professor" }, { "name": "Simon Shaolei Du", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "Yifang Chen", "gender": "Female", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Yiping Wang", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Yuandong Tian", "gender": "Male", "institution": "Meta AI (FAIR)", "country": "", "position": "Research Scientist" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 74, "n_ref": 129, "n_ref_all": 167, "n_fig": 7, "n_tab": 7, "L_tab": 2289, "n_element_tab": 360, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1548, "n_element_tab_1": 223, "formula_len_all": 686, "formula_len_all_1": 663, "len_all": 255781, "len_all_1": 72960, "len_abs": 1626, "len_title": 145, "len_sents": 50985, "len_sents_1": 33406, "n_sents": 367, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1648, "title": "SHARP: Accelerating Language Model Inference by SHaring Adjacent layers with Recovery Parameters", "abs": "While Large language models (LLMs) have advanced natural language processing tasks, their growing computational and memory demands make deployment on resource-constrained devices like mobile phones increasingly challenging. In this paper, we propose SHARP (SHaring Adjacent Layers with Recovery Parameters), a novel approach to accelerate LLM inference by sharing parameters across adjacent layers, thus reducing memory load overhead, while introducing low-rank recovery parameters to maintain performance.\nInspired by observations that consecutive layers have similar outputs, SHARP employs a two-stage recovery process: Single Layer Warmup (SLW), and Supervised Fine-Tuning (SFT).\nThe SLW stage aligns the outputs of the shared layers using $\\mathcal{L}_2$ loss, providing a good initialization for the following SFT stage to further restore the model performance. Extensive experiments demonstrate that SHARP can recover the model's perplexity on various in-distribution tasks using no more than 50k fine-tuning data while reducing the number of stored MLP parameters by 38\\% to 65\\%.\nWe also conduct several ablation studies of SHARP and show that replacing layers towards the later parts of the model yields better performance retention, and that different recovery parameterizations perform similarly when parameter counts are matched.\nFurthermore, SHARP saves 42.8\\% in model storage and reduces the total inference time by 42.2\\% compared to the original Llama2-7b model on mobile devices.\nOur results highlight SHARP as an efficient solution for reducing inference costs in deploying LLMs without the need for pretraining-scale resources.", "keywords": [ "inference acceleration", "weight sharing", "language model", "model compression" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "WavXPunwzM", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Ailing Zeng", "gender": "Female", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Biao Jiang", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Chi Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Fukun Yin", "gender": "Male", "institution": "Stepfun", "country": "CN", "position": "Intern" }, { "name": "Gang Yu", "gender": "Male", "institution": "StepFun", "country": "CN", "position": "Principal Researcher" }, { "name": "Tao Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Xianfang Zeng", "gender": "Male", "institution": "Tencent PCG", "country": "CN", "position": "Researcher" }, { "name": "Xin Chen", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Xinru Sun", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Xuanyang Zhang", "gender": "Male", "institution": "stepfun", "country": "CN", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 39, "n_ref": 83, "n_ref_all": 90, "n_fig": 1, "n_tab": 7, "L_tab": 2597, "n_element_tab": 227, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1830, "n_element_tab_1": 144, "formula_len_all": 262, "formula_len_all_1": 186, "len_all": 148751, "len_all_1": 64631, "len_abs": 1331, "len_title": 103, "len_sents": 35965, "len_sents_1": 30675, "n_sents": 223, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1335, "title": "Causal Motion Tokenizer for Streaming Motion Generation", "abs": "Recent advancements in human motion generation have leveraged various multimodal inputs, including text, music, and audio. Despite significant progress, the challenge of generating human motion in a streaming context\u2014particularly from text\u2014remains underexplored. Traditional methods often rely on temporal modalities, leaving text-based motion generation with limited capabilities, especially regarding seamless transitions and low latency. In this work, we introduce MotionStream, a pioneering motion-streaming pipeline designed to continuously generate human motion sequences that adhere to the semantic constraints of input text. Our approach utilizes a Causal Motion Tokenizer, built on residual vector quantized variational autoencoder (RVQ-VAE) with causal convolution, to enhance long sequence handling and ensure smooth transitions between motion segments. Furthermore, we employ a Masked Transformer and Residual Transformer to generate motion tokens efficiently. Extensive experiments validate that MotionStream not only achieves state-of-the-art performance in motion composition but also maintains real-time generation capabilities with significantly reduced latency. We highlight the versatility of MotionStream through a story-to-motion application, demonstrating its potential for robotic control, animation, and gaming.", "keywords": [ "3d motion", "motion generation", "human motion synthesis", "text-driven", "text-to-motion" ], "rating_list": [ 5, 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 1, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "WZ0s2smcKP", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Haozhao Wang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Postdoc" }, { "name": "Jun Wang", "gender": "unknown", "institution": "iWudao", "country": "", "position": "Consultant" }, { "name": "Ruixuan Li", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Wei Liu", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Postdoc" }, { "name": "Zhigang Zeng", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zhiying Deng", "gender": "unknown", "institution": "Central China Normal University", "country": "CN", "position": "Lecturer" }, { "name": "Zhongyu Niu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" } ], "n_formula": 26, "n_formula_1": 14, "n_ref_uni": 50, "n_ref": 99, "n_ref_all": 154, "n_fig": 14, "n_tab": 9, "L_tab": 4169, "n_element_tab": 336, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 2020, "n_element_tab_1": 136, "formula_len_all": 1480, "formula_len_all_1": 776, "len_all": 204431, "len_all_1": 70445, "len_abs": 1890, "len_title": 134, "len_sents": 54926, "len_sents_1": 31839, "n_sents": 485, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1903, "title": "Breaking Free from MMI: A New Frontier in Rationalization by Probing Input Utilization", "abs": "Extracting a small subset of crucial rationales from the full input is a key problem in explainability research. The most widely used fundamental criterion for rationale extraction is the maximum mutual information (MMI) criterion. In this paper, we first demonstrate that MMI suffers from diminishing marginal returns. Once part of the rationale has been identified, finding the remaining portions contributes only marginally to increasing the mutual information, making it difficult to use MMI to locate the rest. In contrast to MMI that aims to reproduce the prediction, we seek to identify the parts of the input that the network can actually utilize. This is achieved by comparing how different rationale candidates match the capability space of the weight matrix. The weight matrix of a neural network is typically low-rank, meaning that the linear combinations of its column vectors can only cover part of the directions in a high-dimensional space (high-dimension: the dimensions of an input vector). If an input is fully utilized by the network, it generally matches these directions (e.g., a portion of a hypersphere), resulting in a representation with a high norm. Conversely, if an input primarily falls outside (orthogonal to) these directions, its representation norm will approach zero, behaving like noise that the network cannot effectively utilize. \nBuilding on this, we propose using the norms of rationale candidates as an alternative objective to MMI. \nThrough experiments on four text classification datasets and one graph classification dataset using three network architectures (GRUs, BERT, and GCN), we show that our method outperforms MMI and its improved variants in identifying better rationales. We also compare our method with a representative LLM (llama-3.1-8b-instruct) and find that our simple method gets comparable results to it and can sometimes even outperform it.", "keywords": [ "Interpretability", "natural language processing", "feature selection" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "WYsNjw02DE", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Hanna Kurniawati", "gender": "Female", "institution": "Australian National University", "country": "", "position": "Full Professor" }, { "name": "Yongbo Chen", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "Postdoc" } ], "n_formula": 36, "n_formula_1": 11, "n_ref_uni": 37, "n_ref": 49, "n_ref_all": 90, "n_fig": 16, "n_tab": 4, "L_tab": 56, "n_element_tab": 7, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 203, "n_element_tab_1": 13, "formula_len_all": 3094, "formula_len_all_1": 843, "len_all": 214349, "len_all_1": 81206, "len_abs": 1463, "len_title": 153, "len_sents": 78371, "len_sents_1": 40196, "n_sents": 560, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1482, "title": "NPF-$k$CT: A $k$-center clustering solver with neural process filter for continuous POMDP-based object search", "abs": "Efficiently searching for target objects in intricate environments poses a significant challenge for mobile robots, due to perception errors, limited field of view (FOV), and visual occlusion. These factors cause the problem to be partially observed. Therefore, we formulate the object-search task as a high-dimensional Partially Observable Markov Decision Process (POMDP) with hybrid (continuous and discrete) action spaces. We propose a novel sampling-based online POMDP solver named Neural Process Filtered $k$-Center Clustering Tree (NPF-$k$CT). The optimal action is selected using Monte Carlo Tree Search (MCTS) in conjunction with a neural process network to filter out ineffective primitive actions (i.e., basic robot operations), alongside $k$-center clustering hypersphere discretization to efficiently refine high-dimensional continuous sub-action spaces. Adhering to the hierarchical optimistic optimization (HOO) concept, we leverage an upper-confidence bound (UCB) on the action value function within the hypersphere with estimated diameters to guide the MCTS expansion. We extensively tested our approach in Gazebo simulations using Fetch and Stretch robots across diverse target-finding scenarios. Comparative results show higher success rates and faster target detection than baseline methods, with no additional computational cost. We also validated our method on a physical robot in an office environment. Project page: \\url{https://sites.google.com/view/npfkct}.", "keywords": [ "robotics", "planning under uncertainty", "continuous Partially Observable Markov Decision Process", "Neural Process" ], "rating_list": [ 3, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "WYsCKxZc5Y", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Andrey Zhmoginov", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Jihwan Lee", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Mark Sandler", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Max Vladymyrov", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" } ], "n_formula": 26, "n_formula_1": 10, "n_ref_uni": 18, "n_ref": 26, "n_ref_all": 63, "n_fig": 23, "n_tab": 1, "L_tab": 281, "n_element_tab": 28, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 281, "n_element_tab_1": 28, "formula_len_all": 1745, "formula_len_all_1": 554, "len_all": 166016, "len_all_1": 73455, "len_abs": 1130, "len_title": 104, "len_sents": 63057, "len_sents_1": 35002, "n_sents": 462, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1135, "title": "Contextually Guided Transformers via Low-Rank Adaptation", "abs": "Large Language Models (LLMs) based on Transformers excel at text processing, but their reliance on prompts for specialized behavior introduces computational overhead. We propose a modification to a Transformer architecture that eliminates the need for explicit prompts by learning to encode context into the model's weights. Our Contextually Guided Transformer (CGT) model maintains a contextual summary at each sequence position, allowing it to update the weights on the fly based on the preceding context. This approach enables the model to self-specialize, effectively creating a tailored model for processing information following a given prefix. We demonstrate the effectiveness of our method on synthetic in-context learning tasks and language modeling benchmarks. Furthermore, we introduce techniques for enhancing the interpretability of the learned contextual representations, drawing connections to Variational Autoencoders and promoting smoother, more consistent context encoding. This work offers a novel direction for efficient and adaptable language modeling by integrating context directly into the model's architecture.", "keywords": [ "transformers", "hypernetworks" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "WYdpjwKQma", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Cissy Hing Yee Choy", "gender": "unknown", "institution": "University of Chicago", "country": "US", "position": "Undergrad student" }, { "name": "Han Fang", "gender": "unknown", "institution": "HSDC", "country": "CN", "position": "Researcher" }, { "name": "Jian Chen", "gender": "Male", "institution": "HSBC GPS LAB", "country": "CN", "position": "Consultant Specialist" }, { "name": "Jingfeng Luo", "gender": "Male", "institution": "HSBC", "country": "CN", "position": "Delivery Head" }, { "name": "Peilin Zhou", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Xinmei Ke", "gender": "Female", "institution": "HSBC", "country": "CN", "position": "Researcher" }, { "name": "Xuming Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Yining Hua", "gender": "Not Specified", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Zhenyan Chen", "gender": "Female", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Zixuan Yuan", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 26, "n_ref": 53, "n_ref_all": 77, "n_fig": 4, "n_tab": 10, "L_tab": 2511, "n_element_tab": 184, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 1659, "n_element_tab_1": 135, "formula_len_all": 67, "formula_len_all_1": 100, "len_all": 139710, "len_all_1": 63683, "len_abs": 2303, "len_title": 166, "len_sents": 41398, "len_sents_1": 28112, "n_sents": 312, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 118, "L_abs": 1665, "title": "LAIA-SQL: Enhancing Natural Language to SQL Generation in Multi-Table QA via Task Decomposition and Keyword Extraction", "abs": "Natural Language to SQL (NL2SQL) provides an effective solution for multi-table question answering (Table QA) to automate data retrieval by transforming simple user queries into SQL commands. It enhances data accessibility and decision-making processes across various industries. Large Language Model (LLM) based NL2SQL methods have been shown to outperform rule-based or neural network-based NL2SQL methods. However, existing LLM-based NL2SQL approaches face challenges like inaccurate interpretation of user questions, slow retrieval speeds, erroneous SQL generation, and high operational costs. As there is a lack of datasets specifically designed to evaluate natural language understanding (NLU) in NL2SQL tasks and no models optimized for user question understanding in Table QA, we introduce LAIA-NLU, a novel dataset that dissects NLU into task decomposition and keyword extraction. LAIA-NLU contains 1,500 high-quality QA pairs, created through manual review. Using this dataset, we developed LAIA-NLUer, which is capable of effectively interpreting user intent in table-based queries. To further enhance NL2SQL performance in terms of speed, cost, and accuracy, we also present LAIA-SQL, a retrieval-augmented based NL2SQL framework. Experimental results show that LAIA-SQL outperforms state-of-the-art models, achieving an accuracy improvement to 67.28% in BIRD dataset, a 52.4% reduction in runtime, and a 97% decrease in operational costs. These improvements demonstrate the potential of our approach to advance multi-table data retrieval and analysis. Our code, dataset, and model will be publicly available to encourage further research in this field.", "keywords": [ "Natural Language Understanding", "Text to SQL", "Multi Table QA" ], "rating_list": [ 3, 6, 3, 8 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 1, 2, 2, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "WYL4eFLcxG", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Alon Benhaim", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Furu Wei", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Partner Research Manager" }, { "name": "Johan Bjorck", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Vishrav Chaudhary", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Senior Principal Researcher" }, { "name": "Xia Song", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 44, "n_ref": 74, "n_ref_all": 132, "n_fig": 15, "n_tab": 7, "L_tab": 1196, "n_element_tab": 157, "n_fig_1": 15, "n_tab_1": 3, "L_tab_1": 520, "n_element_tab_1": 30, "formula_len_all": 307, "formula_len_all_1": 213, "len_all": 130930, "len_all_1": 61680, "len_abs": 1184, "len_title": 88, "len_sents": 32787, "len_sents_1": 26410, "n_sents": 296, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1192, "title": "Scaling Optimal LR Across Token Horizons", "abs": "State-of-the-art LLMs are powered by scaling -- scaling model size, dataset size, and cluster size. It is economically infeasible to extensively tune hyperparameters for the largest runs. Instead, approximately optimal hyperparameters must be inferred or transferred from smaller experiments. Hyperparameter transfer across model sizes has been studied in Yang et. al. However, hyperparameter transfer across dataset size -- or token horizon -- has not been studied yet. To remedy this we conduct a large-scale empirical study on how optimal learning rate (LR) depends on the token horizon in LLM training. We first demonstrate that the optimal LR changes significantly with token horizon -- longer training necessitates smaller LR. Secondly, we demonstrate that the optimal LR follows a scaling law and that the optimal LR for longer horizons can be accurately estimated from shorter horizons via such scaling laws. We also provide a rule-of-thumb for transferring LR across token horizons with zero overhead over current practices. Lastly, we provide evidence that LLama-1 used too high LR, and argue that hyperparameter transfer across data size is an overlooked component of LLM training.", "keywords": [ "LLMs", "scaling laws", "hyperparameters", "mup" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "WWymYrA48K", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Panayiotis Christou", "gender": "Male", "institution": "New York University", "country": "US", "position": "MS student" }, { "name": "Parijat Dube", "gender": "unknown", "institution": "New York University", "country": "", "position": "Adjunct Associate Professor" }, { "name": "Shichu Chen", "gender": "Male", "institution": "New York University", "country": "US", "position": "MS student" }, { "name": "Xupeng Chen", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 1, "n_ref_uni": 28, "n_ref": 67, "n_ref_all": 87, "n_fig": 5, "n_tab": 9, "L_tab": 2239, "n_element_tab": 329, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 872, "n_element_tab_1": 280, "formula_len_all": 1889, "formula_len_all_1": 64, "len_all": 197788, "len_all_1": 52618, "len_abs": 5552, "len_title": 94, "len_sents": 75312, "len_sents_1": 24710, "n_sents": 502, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 773, "title": "Test Time Learning for Time Series Forecasting", "abs": "We propose the use of Test-Time Training (TTT) modules in a cascade architecture to enhance performance in long-term time series forecasting. Through extensive experiments on standard benchmark datasets, we demonstrate that TTT modules consistently outperform state-of-the-art models, including Mamba-based TimeMachine, particularly in scenarios involving extended sequence and prediction lengths. Our results show significant improvements, especially on larger datasets such as Electricity, Traffic, and Weather, underscoring the effectiveness of TTT in capturing long-range dependencies. Additionally, we explore various convolutional architectures within the TTT framework, showing that convolutional blocks as hidden layer architectures can achieve competitive results.", "keywords": [ "Time Series Forecasting", "Test-Time Training", "Mamba", "Expressive Hidden States", "Modern CNN" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 2, 5, 5 ] }, { "paper_id": "WWXjMYZxfH", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Haoran Sun", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Intern" }, { "name": "Huang Fang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Shuohuan Wang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Yekun Chai", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Yu Sun", "gender": "unknown", "institution": "Baidu", "country": "", "position": "Researcher" }, { "name": "hua wu", "gender": "Female", "institution": "Baidu", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 14, "n_formula_1": 8, "n_ref_uni": 54, "n_ref": 99, "n_ref_all": 144, "n_fig": 19, "n_tab": 13, "L_tab": 3896, "n_element_tab": 755, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 251, "n_element_tab_1": 49, "formula_len_all": 1054, "formula_len_all_1": 496, "len_all": 261199, "len_all_1": 60141, "len_abs": 1756, "len_title": 119, "len_sents": 71740, "len_sents_1": 28168, "n_sents": 586, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1619, "title": "MA-RLHF: Reinforcement Learning from Human Feedback with Macro Actions", "abs": "Reinforcement learning from human feedback (RLHF) has demonstrated effectiveness in aligning large language models (LLMs) with human preferences. However, token-level RLHF suffers from the credit assignment problem over long sequences, where delayed rewards make it challenging for the model to discern which actions contributed to preferred outcomes. This hinders learning efficiency and slows convergence.In this paper, we propose MA-RLHF, a simple yet effective RLHF framework that incorporates macro actions --- sequences of tokens or higher-level language constructs --- into the learning process. By operating at higher level of abstraction, our approach reduces the temporal distance between actions and rewards, facilitating faster and more accurate credit assignment. This results in more stable policy gradient estimates and enhances learning efficiency within each episode, all without increasing computational complexity during training or inference. We validate our approach through extensive experiments across various model sizes and tasks, including text summarization, dialogue generation, question answering, and program synthesis. Our method achieves substantial performance improvements over standard RLHF, with performance gains of up to 30\\% in text summarization and code generation, 18\\% in dialogue, and 8\\% in question answering tasks. Notably, our approach reaches parity with vanilla RLHF $1.7 \\sim 2$ times faster in terms of training time and continues to outperform it with further training. We make our code and data publicly available at \\url{https://github.com/ernie-research/MA-RLHF}.", "keywords": [ "Human Alignment", "Large Language Models", "Reinforcement Learning" ], "rating_list": [ 3, 6, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 3, 3 ] }, { "paper_id": "WVzYMa68Of", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Andrei Chertkov", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "Researcher" }, { "name": "Ivan Oseledets", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 33, "n_ref": 48, "n_ref_all": 67, "n_fig": 5, "n_tab": 4, "L_tab": 784, "n_element_tab": 117, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1173, "n_element_tab_1": 132, "formula_len_all": 833, "formula_len_all_1": 865, "len_all": 94852, "len_all_1": 58136, "len_abs": 1204, "len_title": 124, "len_sents": 28452, "len_sents_1": 25684, "n_sents": 197, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1209, "title": "Tensor Train Decomposition for Adversarial Attacks on Computer Vision Models", "abs": "Deep neural networks (DNNs) are widely used today, but they are vulnerable to adversarial attacks. To develop effective methods of defense, it is important to understand the potential weak spots of DNNs. Often attacks are organized taking into account the architecture of models (white-box approach) and based on gradient methods, but for real-world DNNs this approach in most cases is impossible. At the same time, several gradient-free optimization algorithms are used to attack black-box models. However, classical methods are often ineffective in the multidimensional case. To organize black-box attacks for computer vision models, in this work, we propose the use of an optimizer based on the low-rank tensor train (TT) format, which has gained popularity in various practical multidimensional applications in recent years. Combined with the attribution of the target image, which is built by the auxiliary (white-box) model, the TT-based optimization method makes it possible to organize an effective black-box attack by small perturbation of pixels in the target image. The superiority of the proposed approach over three popular baselines is demonstrated for seven modern DNNs on the ImageNet dataset.", "keywords": [ "tensor train", "optimization", "attribution", "adversarial attack", "computer vision", "black box" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "WVmarX0RNd", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 4, "author_info_list": [ { "name": "Jackson Kunde", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Undergrad student" }, { "name": "Kangwook Lee", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Ozgur Guldogan", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Ramtin Pedarsani", "gender": "unknown", "institution": "UC Santa Barbara", "country": "", "position": "Assistant Professor" } ], "n_formula": 38, "n_formula_1": 8, "n_ref_uni": 16, "n_ref": 27, "n_ref_all": 40, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 505, "n_element_tab_1": 6, "formula_len_all": 2690, "formula_len_all_1": 425, "len_all": 119697, "len_all_1": 66056, "len_abs": 920, "len_title": 106, "len_sents": 37125, "len_sents_1": 27755, "n_sents": 349, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 931, "title": "Multi-Bin Batching for Increasing LLM Inference Throughput", "abs": "As large language models (LLMs) grow in popularity for their diverse capabilities, improving the efficiency of their inference systems has become increasingly critical. Batching requests during LLM inference increases throughput by allowing multiple requests to be processed in parallel, making better use of hardware resources such as GPUs. However, the autoregressive nature of LLMs presents a challenge: requests often have varying execution times, causing resource underutilization, as hardware must wait for the longest-running request in the batch to complete before moving to the next batch. We propose Multi-Bin Batching, a simple yet effective method that can \\emph{provably improve LLM inference throughput} by grouping requests with similar execution times into predetermined bins. We evaluate multi-bin batching on various settings, showing consistent throughput improvements compared to standard batching approaches.", "keywords": [ "LLM inference", "Throughput optimization", "Batched inference" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 1, 3, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "WVWlO2tium", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "HyunJin Kim", "gender": "Male", "institution": "Dankook University", "country": "KR", "position": "Full Professor" }, { "name": "Sejin Kwon", "gender": "Female", "institution": "Dankook University", "country": "KR", "position": "Undergrad student" }, { "name": "YOUNGWOOK KWON", "gender": "Male", "institution": "Dankook University", "country": "", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 1, "n_ref_uni": 38, "n_ref": 118, "n_ref_all": 159, "n_fig": 8, "n_tab": 9, "L_tab": 5131, "n_element_tab": 403, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2392, "n_element_tab_1": 187, "formula_len_all": 1997, "formula_len_all_1": 90, "len_all": 178419, "len_all_1": 60755, "len_abs": 1893, "len_title": 137, "len_sents": 54795, "len_sents_1": 26362, "n_sents": 421, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1929, "title": "Binarized Convolutional Neural Networks with Channel Quadrupling and Smooth Downsampling", "abs": "This paper proposes novel binarized convolutional neural networks (BCNNs) named **QB-Net** and **QSB-Net**, specifically designed to **Q**uadruple the number of channels and incorporate a so-called **S**mooth downsampling in **B**CNNs for low-cost mobile environments. The proposed models combine FP32 depthwise separable (DS) convolutions with binarized $1 \\times 1$ pointwise convolutions, offering reduced computational costs in the pointwise convolutions. To enhance the degraded performance of the above naive combination, the proposed models start with a small number of channels in shallow layers and expand them during downsampling by a factor of four, effectively managing model complexity in the downsampling. The proposed model structure maintains low computational costs in the shallow blocks and increases model complexity in the deep blocks, providing a wider dynamic range to manage information in the frequency domain. As a result, the proposed models overcome the limitations of existing BCNNs, delivering improved performance while reducing the total computational costs. For further performance enhancements, we propose a novel smooth downsampling with heightwise and widthwise sequential downsampling steps, doubling the number of channels at each step. Besides, we show that the channelwise self-attention (SE) is applicable with minimal additional computational costs in the proposed models. Besides, multiple binarized convolutions in the fully-connected (FC) layer reduce storage costs without requiring 8-bit quantized convolutions. Experimental results demonstrate the efficiency of the proposed models in terms of performance, computational costs, and inference latency on real hardware. Notably, the QSB-Net-Large with SE achieve 71.2\\% Top-1 accuracy on ImageNet-1K and 69.2 mean intersection over union (mIoU) in the semantic segmentation on the PASCAL VOC dataset, outperforming other counterparts.", "keywords": [ "Binarized Convolutional Neural Networks", "Image Classification", "Image Segmentation", "Mobile-Friendly Convolutional Neural Networks" ], "rating_list": [ 5, 5, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 4, 2 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 2, 4, 4, 4, 4 ] }, { "paper_id": "WVWZ6SnM4t", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "MONICA SUNKARA", "gender": "unknown", "institution": "AWS AI", "country": "US", "position": "Researcher" }, { "name": "Nilaksh Das", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Raphael Shu", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Tamer Alkhouli", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Yi-An Lai", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Yi Zhang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Young Min Cho", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Jinglun Cai", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 25, "n_ref": 32, "n_ref_all": 52, "n_fig": 4, "n_tab": 12, "L_tab": 7939, "n_element_tab": 1120, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 386, "n_element_tab_1": 55, "formula_len_all": 196, "formula_len_all_1": 295, "len_all": 178667, "len_all_1": 62671, "len_abs": 2873, "len_title": 135, "len_sents": 53635, "len_sents_1": 29258, "n_sents": 463, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1155, "title": "RoundTable: Investigating Group Decision-Making Mechanism in Multi-Agent Collaboration", "abs": "This study investigates the efficacy of Multi-Agent Systems in eliciting cross-agent communication and enhancing collective intelligence through group decision-making in a decentralized setting. Unlike centralized mechanisms, where a fixed hierarchy governs social choice, decentralized group decision-making allows agents to engage in joint deliberation. Our research focuses on the dynamics of communication and decision-making within various social choice methods. By applying different voting rules in various environments, we find that moderate decision flexibility yields better outcomes. Additionally, exploring the linguistic features of agent-to-agent conversations reveals indicators of effective collaboration, offering insights into communication patterns that facilitate or hinder collaboration. Finally, we propose various methods for determining the optimal stopping point in multi-agent collaborations based on linguistic cues. Our findings contribute to a deeper understanding of how decentralized decision-making and group conversation shape multi-agent collaboration, with implications for the design of more effective MAS environments.", "keywords": [ "Multi-Agent System", "Decentralized Collaboration", "Collaboration Pattern", "Group Behavior", "Communication Protocol" ], "rating_list": [ 3, 3, 5, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 1, 2, 4 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "WVVu6B8knx", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Bilal Faye", "gender": "Male", "institution": "Universit\u00e9 Paris Nord (Paris XIII)", "country": "FR", "position": "PhD student" }, { "name": "Hanane AZZAG", "gender": "Female", "institution": "University Paris 13, Universit\u00e9 Paris Nord (Paris XIII)", "country": "FR", "position": "Full Professor" }, { "name": "Mustapha Lebbah", "gender": "Male", "institution": "Universit\u00e9 de Versailles Saint-Quentin-en-Yvelines", "country": "FR", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 20, "n_ref": 26, "n_ref_all": 32, "n_fig": 0, "n_tab": 5, "L_tab": 832, "n_element_tab": 84, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 832, "n_element_tab_1": 84, "formula_len_all": 956, "formula_len_all_1": 956, "len_all": 72947, "len_all_1": 59625, "len_abs": 1275, "len_title": 78, "len_sents": 23945, "len_sents_1": 23879, "n_sents": 157, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Reject", "L_title": 30, "L_abs": 1282, "title": "Supervised Batch Normalization", "abs": "Batch Normalization (BN), a widely-used technique in neural networks, enhances generalization and expedites training by normalizing each mini-batch to the same mean and variance. However, its effectiveness diminishes when confronted with diverse data distributions.\nTo address this challenge, we propose Supervised Batch Normalization (SBN), a pioneering approach. We expand normalization beyond traditional single mean and variance parameters, enabling the identification of data modes prior to training. This ensures effective normalization for samples sharing common features. We define contexts as modes, categorizing data with similar characteristics. These contexts are explicitly defined, such as domains in domain adaptation or modalities in multimodal systems, or implicitly defined through clustering algorithms based on data similarity. We illustrate the superiority of our approach over BN and other commonly employed normalization techniques through various experiments on both single and multi-task datasets. Integrating SBN with Vision Transformer results in a remarkable 15.13% accuracy enhancement on CIFAR-100. Additionally, in domain adaptation scenarios, employing AdaMatch demonstrates an impressive 22.25% accuracy improvement on MNIST and SVHN compared to BN.", "keywords": [ "machine learning", "deep learning", "computer vision", "activation normalization", "batch normalization" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "WVLBWiKxjM", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Bilal Zahid Hussain", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "MS student" }, { "name": "Fatahlla Moreh", "gender": "Male", "institution": "Christian-Albrechts-Universit\u00e4t Kiel", "country": "DE", "position": "PhD student" }, { "name": "Mohammad Ammar", "gender": "Male", "institution": "Aligarh Muslim University", "country": "IN", "position": "Undergrad student" }, { "name": "Sven Tomforde", "gender": "Male", "institution": "Christian-Albrechts-Universit\u00e4t Kiel", "country": "", "position": "Full Professor" }, { "name": "Yusuf Hasan", "gender": "Male", "institution": "Aligarh Muslim University, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 20, "n_ref": 28, "n_ref_all": 30, "n_fig": 0, "n_tab": 4, "L_tab": 1523, "n_element_tab": 82, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 699, "n_element_tab_1": 72, "formula_len_all": 173, "formula_len_all_1": 173, "len_all": 86203, "len_all_1": 56645, "len_abs": 1185, "len_title": 146, "len_sents": 32600, "len_sents_1": 27801, "n_sents": 226, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1185, "title": "Deep Learning for Micro-Scale Crack Detection on Imbalanced Datasets Using Key Point Localization", "abs": "Internal crack detection has been a subject of focus in structural health monitoring. By focusing on crack detection in structural datasets, it is demonstrated that deep learning (DL) methods can effectively analyse seismic wave fields interacting with micro-scale cracks, which are beyond the resolution of conventional visual inspection.\n\nThis work explores a novel application of DL based key point detection technique, where cracks are localized by predicting the coordinates of four key points that define a bounding region of the crack.\nThe study not only opens new research directions for non-visual applications but also effectively mitigates the impact of imbalanced data which poses a challenge for previous DL models, as it can be biased toward predicting the majority class (non-crack regions). Popular DL techniques, such as the Inception blocks are used and investigated. \nThe model shows an overall reduction in loss when applied to micro-scale crack detection and is reflected in the lower average deviation between the location of actual and predicted cracks, with an average IOU being 0.511 for all micro cracks (> 0.00 \u00b5m) and 0.631 for larger micro cracks (> 4 \u00b5m).", "keywords": [ "Micro-scale crack detection", "Imbalanced datasets", "Key point localization", "Squeeze-and-excite blocks", "Deep learning in structural health monitoring", "Wide convolutional networks", "Structural defect localization", "Seismic wave analysis", "Bounding box regression" ], "rating_list": [ 3, 3, 1, 3, 5 ], "soundness_list": [ 2, 3, 1, 3, 2 ], "presentation_list": [ 2, 2, 1, 2, 2 ], "contribution_list": [ 2, 1, 1, 2, 2 ], "confidence_list": [ 5, 5, 5, 3, 4 ] }, { "paper_id": "WVBzN1HIFS", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Chitta Baral", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Full Professor" }, { "name": "Himanshu Gupta", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Kevin Scaria", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Mihir Parmar", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Shreyas Verma", "gender": "Male", "institution": "Simplr AI", "country": "US", "position": "Researcher" }, { "name": "Swaroop Mishra", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Ujjwala Anantheswaran", "gender": "unknown", "institution": "Arizona State University", "country": "", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 108, "n_ref_all": 136, "n_fig": 2, "n_tab": 10, "L_tab": 2886, "n_element_tab": 453, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 3857, "n_element_tab_1": 593, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 278432, "len_all_1": 58460, "len_abs": 1700, "len_title": 116, "len_sents": 45588, "len_sents_1": 21351, "n_sents": 412, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1705, "title": "PolyMATH: A Challenging Multi-Modal Mathematical Reasoning Benchmark", "abs": "Multi-modal Large Language Models (MLLMs) exhibit impressive problem solving abilities in various domains, but their visual comprehension and abstract reasoning skills remain under-evaluated. To this end, we present POLYMATH, a challenging benchmark aimed at evaluating the general cognitive reasoning abilities of MLLMs. POLYMATH comprises 5,000 manually collected high-quality images\nof cognitive textual and visual challenges across 10 distinct categories, including pattern recognition, spatial reasoning, and relative reasoning. We conducted a comprehensive, and quantitative evaluation of 15 MLLMs using four diverse prompting strategies, including Chain-of-Thought and Step-Back. The best scores achieved on POLYMATH are \u223c 41%, \u223c 36%, and \u223c 27%, obtained by Claude-3.5 Sonnet, GPT-4o and Gemini-1.5 Pro respectively highlighting the logical and visual complexity of these questions. A further fine-grained error analysis reveals that these models struggle to understand spatial relations and perform drawn-out, high-level reasoning. This is further strengthened by our ablation study estimating MLLM performance when given textual descriptions in place of diagrams. As evidenced by \u223c 4% improvement over textual descriptions as opposed to actual images, we discover that models do not truly comprehend visual diagrams and the spatial information therein, and are thus prone to logical errors. Finally, we evaluate the OpenAI o1 models and find that their performance only matches the human baseline, highlighting the difficulty of the benchmark. The results on POLYMATH highlight the room for improvement in multi-modal reasoning and provide unique insights to guide the development of future MLLMs", "keywords": [ "Visual Math Problem-Solving", "Multi-Modal Language Models (MLLMs)", "Cognitive Reasoning Evaluation" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "WUibctXLT7", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Geon Yeong Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Heechan Jeon", "gender": "Male", "institution": "AMOREPACIFIC", "country": "KR", "position": "Principal Researcher" }, { "name": "Inhwa Han", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jin Nam", "gender": "unknown", "institution": "AMOREPACIFIC", "country": "KR", "position": "Researcher" }, { "name": "Jong Chul Ye", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Myeongjin Goh", "gender": "Male", "institution": "AmorePacific", "country": "KR", "position": "Researcher" }, { "name": "Seongmin Jeong", "gender": "Not Specified", "institution": "AMOREPACIFIC Corp. R&I Unit", "country": "KR", "position": "Researcher" }, { "name": "Serin Yang", "gender": "Female", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Sung Won Yi", "gender": "Male", "institution": "Amorepacific", "country": "KR", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 24, "n_ref": 40, "n_ref_all": 68, "n_fig": 10, "n_tab": 1, "L_tab": 444, "n_element_tab": 45, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 564, "formula_len_all_1": 563, "len_all": 90099, "len_all_1": 41725, "len_abs": 150, "len_title": 116, "len_sents": 27703, "len_sents_1": 17851, "n_sents": 237, "n_sents_1": 149, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1040, "title": "DreamMakeup: Face Makeup Customization using Latent Diffusion Models", "abs": "The exponential growth of the global makeup market has paralleled advancements in virtual makeup simulation technology. Despite the progress led by GANs, their application still encounters significant challenges, including training instability and limited customization capabilities. Addressing these challenges, this paper introduces DreamMakup: a novel Diffusion model based Makeup Customization, leveraging the inherent advantages of diffusion models for superior controllability and precise real-image editing. DreamMakeup employs early-stopped DDIM inversion to preserve the facial structure and identity while enabling extensive customization through various conditioning inputs such as reference images, specific RGB colors, and textual descriptions. Our model demonstrates notable improvements over existing GAN-based frameworks, improved customization, color-matching capabilities, and compatibility with textual descriptions or LLMs with affordable computational costs. Project page is available at \nhttps://dreammakeup.github.io/", "keywords": [ "Diffusion model", "Makeup customization", "Image editing" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "WULjblaCoc", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Amir Globerson", "gender": "Male", "institution": "Tel Aviv University", "country": "", "position": "Full Professor" }, { "name": "Asma Ghandeharioun", "gender": "unknown", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Gilad Yehudai", "gender": "Male", "institution": "New York University", "country": "", "position": "Postdoc" }, { "name": "Haim Kaplan", "gender": "Male", "institution": "Google", "country": "IL", "position": "Researcher" }, { "name": "Mor Geva", "gender": "Female", "institution": "Tel Aviv University", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 13, "n_ref_uni": 16, "n_ref": 33, "n_ref_all": 44, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 847, "formula_len_all_1": 741, "len_all": 111999, "len_all_1": 74602, "len_abs": 795, "len_title": 81, "len_sents": 43605, "len_sents_1": 34005, "n_sents": 457, "n_sents_1": 341, "venue": "ICLR2025", "decision": "Reject", "L_title": 33, "L_abs": 806, "title": "When Can Transformers Count to n?", "abs": "Large language models based on the transformer architectures can solve highly complex tasks. But are there simple tasks that such models cannot solve? Here we focus on very simple counting tasks, that involve counting how many times a token in the vocabulary have appeared in a string. We show that if the dimension of the transformer state is linear in the context length, this task can be solved. However, the solution we propose does not scale beyond this limit, and we provide theoretical arguments for why it is likely impossible for a size limited transformer to implement this task. Our empirical results demonstrate the same phase-transition in performance, as anticipated by the theoretical argument. Our results demonstrate the importance of understanding how transformers can solve simple tasks.", "keywords": [ "Transformers", "theory", "counting", "attention" ], "rating_list": [ 6, 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 4, 2 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 3, 5, 5, 3, 4 ] }, { "paper_id": "WTmZS5GU0E", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Gangyi Ding", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Hu Zhang", "gender": "Male", "institution": "Commonwealth Scientific and Industrial Research Organisation, CSIRO", "country": "AU", "position": "Postdoc" }, { "name": "Jintao Sun", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Zhedong Zheng", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 4, "n_ref_uni": 38, "n_ref": 101, "n_ref_all": 128, "n_fig": 9, "n_tab": 5, "L_tab": 2770, "n_element_tab": 224, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 2210, "n_element_tab_1": 147, "formula_len_all": 407, "formula_len_all_1": 81, "len_all": 161608, "len_all_1": 71343, "len_abs": 1556, "len_title": 121, "len_sents": 50622, "len_sents_1": 35330, "n_sents": 348, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1563, "title": "Leveraging Semantic and Positional Uncertainty for Trajectory Prediction", "abs": "Given a time horizon with historical movement data and environmental context, trajectory prediction aims to forecast the future motion of dynamic entities, such as vehicles and pedestrians. A key challenge in this task arises from the dynamic and noisy nature of real-time maps. This noise primarily stems from two resources: (1) positional errors due to sensor inaccuracies or environmental occlusions, and (2) cognitive errors resulting from incorrect scene understanding. \nIn an attempt to solve this problem, we propose a new framework that estimates two kinds of uncertainty, \\ie, positional uncertainty and semantic uncertainty simultaneously, and explicitly incorporates both uncertainties into the trajectory prediction process. \nIn particular, we introduce a dual-head structure to independently perform semantic prediction twice and positional prediction twice, and further extract the prediction variance as the uncertainty indicator in an end-to-end manner. The uncertainty is then directly concatenated with the semantic and positional predictions to enhance the trajectory estimation.\nTo validate the effectiveness of our uncertainty-aware approach, we evaluate it on the real-world driving dataset, \\ie, nuScenes. \nExtensive experiments on 3 mapping estimation and 2 trajectory approaches show that the proposed method (1) effectively captures map noise through both positional and semantic uncertainties, and (2) seamlessly integrates and enhances existing trajectory prediction methods on multiple evaluation metrics, \\ie, minADE, minFDE, and MR.", "keywords": [ "Uncertainty", "Trajectory Prediction" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "WT2bL7sCM1", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Han Yue", "gender": "Male", "institution": "Brandeis University", "country": "US", "position": "PhD student" }, { "name": "Hongfu Liu", "gender": "Male", "institution": "Brandeis University", "country": "US", "position": "Assistant Professor" }, { "name": "Jian Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Ziao Yang", "gender": "Male", "institution": "Brandeis University", "country": "US", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 9, "n_ref_uni": 64, "n_ref": 121, "n_ref_all": 146, "n_fig": 9, "n_tab": 10, "L_tab": 4041, "n_element_tab": 315, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 4026, "n_element_tab_1": 132, "formula_len_all": 1680, "formula_len_all_1": 753, "len_all": 177197, "len_all_1": 68636, "len_abs": 1456, "len_title": 110, "len_sents": 52541, "len_sents_1": 30273, "n_sents": 356, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1464, "title": "Revisit, Extend, and Enhance Hessian-Free Influence Functions", "abs": "Influence functions serve as crucial tools for assessing sample influence. By employing the first-order Taylor extension, sample influence can be estimated without the need for expensive model retraining. However, applying influence functions directly to deep models presents challenges, primarily due to the non-convex nature of the loss function and the large size of model parameters. This difficulty not only makes computing the inverse of the Hessian matrix costly but also renders it non-existent in some cases. Various approaches, including matrix decomposition, have been explored to expedite and approximate the inversion of the Hessian matrix, with the aim of making influence functions applicable to deep models. In this paper, we revisit a specific, albeit naive, yet effective approximation method known as TracIn, and simplify it further, introducing the name Inner Product (IP). This method substitutes the inverse of the Hessian matrix with an identity matrix. We offer deeper insights into why this straightforward approximation method is effective. Furthermore, we extend its applications beyond measuring model utility to include considerations of fairness and robustness. Finally, we enhance IP through an ensemble strategy. To validate its effectiveness, we conduct experiments on synthetic data and extensive evaluations on noisy label detection, sample selection for large language model fine-tuning, and defense against adversarial attacks.", "keywords": [ "Influence Function", "Hessian-Free", "Data Valuation" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 1, 2 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 1, 2, 1, 1 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "WSze9IIN3d", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Arie Kaufman", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Full Professor" }, { "name": "Desai Xie", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Intern" }, { "name": "Difan Liu", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" }, { "name": "Feng Liu", "gender": "unknown", "institution": "Adobe Systems", "country": "US", "position": "Principal Research Scientist" }, { "name": "Hao Tan", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" }, { "name": "Yang Zhou", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Research Scientist" }, { "name": "Yicong Hong", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Zhan Xu", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 18, "n_ref": 52, "n_ref_all": 65, "n_fig": 4, "n_tab": 2, "L_tab": 434, "n_element_tab": 42, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 434, "n_element_tab_1": 42, "formula_len_all": 725, "formula_len_all_1": 725, "len_all": 89880, "len_all_1": 62534, "len_abs": 961, "len_title": 97, "len_sents": 26518, "len_sents_1": 25260, "n_sents": 180, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 966, "title": "Progressive Autoregressive Video Diffusion Models", "abs": "Current frontier video diffusion models have demonstrated remarkable results at\ngenerating high-quality videos. However, they can only generate short video clips,\nnormally around 5 seconds or 120 frames, due to computation limitations during\ntraining. In this work, we show that existing models can be naturally adapted to\nautoregressive video diffusion models without changing the architectures. Our\nkey idea is to assign the latent frames with progressively increasing noise levels\nrather than a single noise level. Thus, each latent can condition on all the less\nnoisy latents before it and provide condition for all the more noisy latents after it.\nSuch progressive video denoising allows our models to autoregressively generate\nframes without quality degradation. We present state-of-the-art results on long\nvideo generation at 1 minute (1440 frames at 24 FPS). Our results are available\nat this anonymous url: https://progressive-autoregressive-vdm.github.io/.", "keywords": [ "Long Video Generation", "Diffusion Models", "Transformer", "Autoregression" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "WRLj18zwz6", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Alejandro Ribeiro", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Full Professor" }, { "name": "Juan Cervino", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Zhiyang Wang", "gender": "Female", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" } ], "n_formula": 103, "n_formula_1": 34, "n_ref_uni": 65, "n_ref": 107, "n_ref_all": 134, "n_fig": 19, "n_tab": 9, "L_tab": 5209, "n_element_tab": 445, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10014, "formula_len_all_1": 1483, "len_all": 229984, "len_all_1": 62376, "len_abs": 3839, "len_title": 130, "len_sents": 64276, "len_sents_1": 29166, "n_sents": 512, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1279, "title": "A Manifold Perspective on the Statistical Generalization of Graph Neural Networks", "abs": "Graph Neural Networks (GNNs) extend convolutional neural networks to operate on graphs. Despite\ntheir impressive performances in various graph learning tasks, the theoretical understanding of\ntheir generalization capability is still lacking. Previous GNN generalization bounds ignore the\nunderlying graph structures, often leading to bounds that increase with the number of nodes \u2013 a\nbehavior contrary to the one experienced in practice. In this paper, we take a manifold perspective\nto establish the statistical generalization theory of GNNs on graphs sampled from a manifold in the\nspectral domain. As demonstrated empirically, we prove that the generalization bounds of GNNs\ndecrease linearly with the size of the graphs in the logarithmic scale, and increase linearly with the\nspectral continuity constants of the filter functions. Notably, our theory explains both node-level and\ngraph-level tasks. Our result has two implications: i) guaranteeing the generalization of GNNs to\nunseen data over manifolds; ii) providing insights into the practical design of GNNs, i.e., restrictions\non the discriminability of GNNs are necessary to obtain a better generalization performance. We\ndemonstrate our generalization bounds of GNNs using synthetic and multiple real-world datasets.", "keywords": [ "generalization analysis", "graph neural networks", "manifold neural networks" ], "rating_list": [ 5, 6, 5, 6, 5 ], "soundness_list": [ 4, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 2, 3, 3 ] }, { "paper_id": "WRKVA3TgSv", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Cazamere Comrie", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Jon Kleinberg", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 0, "n_ref_uni": 14, "n_ref": 43, "n_ref_all": 82, "n_fig": 23, "n_tab": 3, "L_tab": 2005, "n_element_tab": 300, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1777, "formula_len_all_1": 0, "len_all": 161625, "len_all_1": 59438, "len_abs": 1591, "len_title": 100, "len_sents": 63669, "len_sents_1": 31910, "n_sents": 401, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1599, "title": "Can Large Language Models Effectively Modify Graphs?", "abs": "Graphs are essential tools for modeling complex relationships. While prior research with earlier generations of large language models (LLMs) showed them to struggle with basic graph primitives, we find that the situation has changed with modern state-of-the-art (SOTA) LLMs, which excel at these tasks. Given these advances, we propose a more challenging evaluation problem: graph modification, a foundational, interpretable, and non-trivial problem in which an LLM must determine the outcome of adding or deleting a given sequence of nodes or edges, and potentially then compute on the resulting modified graph. We introduce GraphModQA, a novel benchmark dataset comprising graph modification question-answer pairs designed to rigorously test LLMs\u2019 abilities in graph manipulation and dynamic reasoning. Our results show that while SOTA LLMs perform well on static graph property tasks, their accuracy degrades on graph modification tasks; their performance is particularly low as the number of modifications increases, and when the adjacency matrix is used to represent the graph --- an essential encoding not explored in previous work. We provide new techniques for improving performance on graph modification tasks, and we introduce Modify and Print (MAP) prompting, which asks models to output the intermediate adjacency matrices at each step, and which markedly improves the models' performance. Our findings highlight a critical gap in current LLM capabilities regarding dynamic graph reasoning tasks and underscore the potential of techniques like MAP prompting to mitigate these challenges.", "keywords": [ "LLMs", "graphs" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "WQy61tS53c", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Keisuke Fukuda", "gender": "Male", "institution": "Preferred Networks, Inc.", "country": "JP", "position": "Researcher" }, { "name": "Shin-ichi Maeda", "gender": "Male", "institution": "Preferred Networks, Inc.", "country": "JP", "position": "Senior Researcher" }, { "name": "Yuta Tarumi", "gender": "Male", "institution": "Preferred Networks, Inc.", "country": "JP", "position": "Postdoc" } ], "n_formula": 20, "n_formula_1": 14, "n_ref_uni": 30, "n_ref": 61, "n_ref_all": 82, "n_fig": 9, "n_tab": 12, "L_tab": 4176, "n_element_tab": 235, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2057, "n_element_tab_1": 65, "formula_len_all": 2153, "formula_len_all_1": 1008, "len_all": 174763, "len_all_1": 66308, "len_abs": 1162, "len_title": 105, "len_sents": 53609, "len_sents_1": 30806, "n_sents": 472, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1172, "title": "Deep Bayesian Filter for Bayes-Faithful Data Assimilation", "abs": "State estimation for nonlinear state space models (SSMs) is a challenging task. Existing assimilation methodologies predominantly assume Gaussian posteriors on physical space, where true posteriors become inevitably non-Gaussian. We propose Deep Bayesian Filtering (DBF) for data assimilation on nonlinear SSMs. DBF constructs new latent variables $h_t$ in addition to the original physical variables $z_t$ and assimilates observations $o_t$. By (i) constraining the state transition on the new latent space to be linear and (ii) learning a Gaussian inverse observation operator $r(h_t|o_t)$, posteriors remain Gaussian. Notably, the structured design of test distributions enables an analytical formula for the recursive computation, eliminating the accumulation of Monte Carlo sampling errors across time steps. DBF trains the Gaussian inverse observation operators $r(h_t|o_t)$ and other latent SSM parameters (e.g., dynamics matrix) by maximizing the evidence lower bound. Experiments demonstrate that DBF outperforms model-based approaches and latent assimilation methods in tasks where the true posterior distribution on physical space is significantly non-Gaussian.", "keywords": [ "Data Assimilation", "Variational Inference", "State-Space Models", "Koopman Operator" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "WQwy1rW60F", "primary_area": "datasets and benchmarks", "n_author": 13, "author_info_list": [ { "name": "Boxun Li", "gender": "Male", "institution": "Infinigence-AI", "country": "", "position": "Principal Researcher" }, { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Dong Zhou", "gender": "unknown", "institution": "Wuhan University", "country": "", "position": "Undergrad student" }, { "name": "Guohao Dai", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Minghui Zhuang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Shengen Yan", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Xuefei Ning", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Research Assistant Professor" }, { "name": "Yu Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Zheyue Tan", "gender": "Male", "institution": "Infinigence", "country": "", "position": "Researcher" }, { "name": "Zhijie Yang", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Zhuyu Yao", "gender": "Male", "institution": "Infinigence-AI", "country": "", "position": "Researcher" }, { "name": "Shiyao Li", "gender": "Male", "institution": "Infinigence", "country": "CN", "position": "Intern" }, { "name": "Yuan tao", "gender": "unknown", "institution": "Infinigence-AI", "country": "", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 0, "n_ref_uni": 29, "n_ref": 60, "n_ref_all": 101, "n_fig": 15, "n_tab": 9, "L_tab": 8430, "n_element_tab": 645, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 857, "n_element_tab_1": 75, "formula_len_all": 2073, "formula_len_all_1": 0, "len_all": 182552, "len_all_1": 67813, "len_abs": 1781, "len_title": 76, "len_sents": 57291, "len_sents_1": 31607, "n_sents": 465, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1794, "title": "LV-Eval: A Balanced Long-Context Benchmark with 5 Length Levels Up to 256K", "abs": "State-of-the-art large language models (LLMs) are now claiming remarkable supported context lengths of 256k or even more. In contrast, the average context lengths of mainstream benchmarks are insufficient (5k-21k), and they suffer from potential knowledge leakage and inaccurate metrics, resulting in biased evaluation. This paper introduces LV-Eval, a challenging long-context benchmark with five length levels (16k, 32k, 64k, 128k, and 256k) reaching up to 256k words. LV-Eval features two main tasks, single-hop QA and multi-hop QA, comprising 11 bilingual datasets. The design of LV-Eval has incorporated three key techniques, namely confusing facts insertion, keyword and phrase replacement, and keyword-recall-based metric design. The advantages of LV-Eval include controllable evaluation across different context lengths, challenging test instances with confusing facts, mitigated knowledge leakage, and more objective evaluations. We evaluate 15 LLMs on LV-Eval and conduct ablation studies on the benchmarking techniques. The results reveal that:\n(i) Moonshot-v1 and recent large-scale open-source models, such as Qwen-2.5-72B and Llama-3.1-70B, achieve the highest performance on LV-Eval, particularly at lengths below $64k$. (ii) Models exhibit distinct score trends. For example, GLM-4-9B-128k, Yi-6B-200k, and Llama3-8B-1M exhibit a relatively gentle degradation of performance, but their absolute performances may not necessarily be higher than those of LLMs with shorter context lengths. (iii) LLMs' performances can significantly degrade in the presence of confusing information, especially in the pressure test of \"needle in a haystack\". (iv) Issues related to knowledge leakage and inaccurate metrics introduce bias in evaluation, and these concerns are alleviated in LV-Eval.", "keywords": [ "large language model", "long-context benchmark", "knowledge leakage mitigation" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "WQvkqarwXi", "primary_area": "interpretability and explainable AI", "n_author": 9, "author_info_list": [ { "name": "Bosheng Song", "gender": "Male", "institution": "Hunan University", "country": "", "position": "Associate Professor" }, { "name": "Jiani Zhang", "gender": "Female", "institution": "Amazon Web Services", "country": "", "position": "Researcher" }, { "name": "Mufei Li", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Tengfei Ma", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "PhD student" }, { "name": "Wen Tao", "gender": "Female", "institution": "Hunan University", "country": "", "position": "MS student" }, { "name": "Xiang song", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Yijun Wang", "gender": "unknown", "institution": "Hunan University", "country": "", "position": "Assistant Professor" }, { "name": "xiangxiang Zeng", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Full Professor" }, { "name": "Pan Xiaoqin", "gender": "Female", "institution": "Hunan University", "country": "", "position": "MS student" } ], "n_formula": 18, "n_formula_1": 16, "n_ref_uni": 47, "n_ref": 84, "n_ref_all": 114, "n_fig": 8, "n_tab": 8, "L_tab": 2757, "n_element_tab": 226, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 911, "n_element_tab_1": 36, "formula_len_all": 800, "formula_len_all_1": 796, "len_all": 162620, "len_all_1": 66651, "len_abs": 2403, "len_title": 151, "len_sents": 47081, "len_sents_1": 31065, "n_sents": 346, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 1480, "title": "Towards Synergistic Path-based Explanations for Knowledge Graph Completion: Exploration and Evaluation", "abs": "Knowledge graph completion (KGC) aims to alleviate the inherent incompleteness of knowledge graphs (KGs), a crucial task for numerous applications such as recommendation systems and drug repurposing. The success of knowledge graph embedding (KGE) models provokes the question about the explainability: ``\\textit{Which the patterns of the input KG are most determinant to the prediction}?'' Particularly, path-based explainers prevail in existing methods because of their strong capability for human understanding. In this paper, based on the observation that a fact is usually determined by the synergy of multiple reasoning chains, we propose a novel explainable framework, dubbed KGExplainer, to explore synergistic pathways. KGExplainer is a model-agnostic approach that employs a perturbation-based greedy search algorithm to identify the most crucial synergistic paths as explanations within the local structure of target predictions. To evaluate the quality of these explanations, KGExplainer distills an evaluator from the target KGE model, allowing for the examination of their fidelity. We experimentally demonstrate that the distilled evaluator has comparable predictive performance to the target KGE. Experimental results on benchmark datasets demonstrate the effectiveness of KGExplainer, achieving a human evaluation accuracy of 83.3\\% and showing promising improvements in explainability. Code is available at \\url{https://anonymous.4open.science/r/KGExplainer-33A0}", "keywords": [ "Knowledge Graph Completion", "Model Explainability", "Knowledge Graph Embedding", "Link Prediction" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "WQV9kB1qSU", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Kiyoung Seong", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Seonghyun Park", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Sungsoo Ahn", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Woo Youn Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" }, { "name": "SEONGHWAN KIM", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 63, "n_formula_1": 34, "n_ref_uni": 52, "n_ref": 102, "n_ref_all": 146, "n_fig": 7, "n_tab": 5, "L_tab": 3356, "n_element_tab": 273, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 1691, "n_element_tab_1": 28, "formula_len_all": 10898, "formula_len_all_1": 9271, "len_all": 189243, "len_all_1": 88462, "len_abs": 1507, "len_title": 133, "len_sents": 53630, "len_sents_1": 37949, "n_sents": 370, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1610, "title": "Transition Path Sampling with Improved Off-Policy Training of Diffusion Path Samplers", "abs": "Understanding transition pathways between two meta-stable states of a molecular system is crucial to advance drug discovery and material design. However, unbiased molecular dynamics (MD) simulations are computationally infeasible because of the high energy barriers that separate these states. Although recent machine learning techniques are proposed to sample rare events, they are often limited to simple systems and rely on collective variables (CVs) derived from costly domain expertise. In this paper, we introduce a novel approach that trains diffusion path samplers (DPS) to address the transition path sampling (TPS) problem without requiring CVs. We reformulate the problem as an amortized sampling from the transition path distribution by minimizing the log-variance divergence between the path distribution induced by DPS and the transition path distribution. Based on the log-variance divergence, we propose learnable control variates to reduce the variance of gradient estimators and the off-policy training objective with replay buffers and simulated annealing techniques to improve sample efficiency and diversity. We also propose a scale-based equivariant parameterization of the bias forces to ensure scalability for large systems. We extensively evaluate our approach, termed TPS-DPS, on a synthetic system, small peptide, and challenging fast-folding proteins, demonstrating that it produces more realistic and diverse transition pathways than existing baselines. We also provide links to [project page](https://kiyoung98.github.io/tps-dps/) and [code](https://github.com/kiyoung98/tps-dps).", "keywords": [ "molecular dynamics", "transition path sampling" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "WQQyJbr5Lh", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Anqi Pang", "gender": "Male", "institution": "Tsinghua University, ShanghaiTech University", "country": "CN", "position": "PhD student" }, { "name": "Changming Li", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" }, { "name": "Jingyi Yu", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "Full Professor" }, { "name": "Kan Ren", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "Assistant Professor" }, { "name": "Sibei Yang", "gender": "Female", "institution": "ShanghaiTech University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yifan Wang", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Yifei Liu", "gender": "helicopter gunship", "institution": "ShanghaiTech University", "country": "CN", "position": "Undergrad student" }, { "name": "Yingdong Shi", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" } ], "n_formula": 22, "n_formula_1": 4, "n_ref_uni": 46, "n_ref": 124, "n_ref_all": 147, "n_fig": 6, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1640, "n_element_tab_1": 23, "formula_len_all": 1060, "formula_len_all_1": 360, "len_all": 148837, "len_all_1": 58336, "len_abs": 1556, "len_title": 106, "len_sents": 45915, "len_sents_1": 28529, "n_sents": 297, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1563, "title": "Discovering Influential Neuron Path in Vision Transformers", "abs": "Vision Transformer models exhibit immense power yet remain opaque to human understanding, posing challenges and risks for practical applications. While prior research has attempted to demystify these models through input attribution and neuron role analysis, there\u2019s been a notable gap in considering layer-level information and the holistic path of information flow across layers. In this paper, we investigate the significance of influential neuron paths within vision Transformers, which is a path of neurons from the model input to output that impacts the model inference most significantly. We first propose a joint influence measure to assess the contribution of a set of neurons to the model outcome. And we further provide a layer-progressive neuron locating approach that efficiently selects the most influential neuron at each layer trying to discover the crucial neuron path from input to output within the target model. Our experiments demonstrate the superiority of our method finding the most influential neuron path along which the information flows, over the existing baseline solutions. Additionally, the neuron paths have illustrated that vision Transformers exhibit some specific inner working mechanism for processing the visual information within the same image category. We further analyze the key effects of these neurons on the image classification task, show- casing that the found neuron paths have already preserved the model capability on downstream tasks, which may also shed some lights on real-world applications like model pruning.", "keywords": [ "Explainability", "Vision Transformer", "Neuron" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 4, 2, 3, 3, 3 ], "presentation_list": [ 4, 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3, 3 ] }, { "paper_id": "WPz5e5V85k", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Shuailong Zhu", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Intern" }, { "name": "Xiaohui Chen", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "Associate Professor" } ], "n_formula": 60, "n_formula_1": 34, "n_ref_uni": 35, "n_ref": 104, "n_ref_all": 125, "n_fig": 9, "n_tab": 3, "L_tab": 882, "n_element_tab": 114, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5349, "formula_len_all_1": 2932, "len_all": 142277, "len_all_1": 54221, "len_abs": 1059, "len_title": 123, "len_sents": 41921, "len_sents_1": 22509, "n_sents": 396, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1061, "title": "Convergence Analysis of the Wasserstein Proximal Algorithm beyond Convexity", "abs": "The proximal algorithm is a powerful tool to minimize nonlinear and nonsmooth functionals in a general metric space. Motivated by the recent progress in studying the training dynamics of the noisy gradient descent algorithm on two-layer neural networks in the mean-field regime, we provide in this paper a simple and self-contained analysis for the convergence of the general-purpose Wasserstein proximal algorithm without assuming geodesic convexity on the objective functional. Under a natural Wasserstein analog of the Euclidean Polyak-{\\L}ojasiewicz inequality, we show that the proximal algorithm achieves an unbiased and linear convergence rate. Our convergence rate improves upon existing rates of the proximal algorithm for solving Wasserstein gradient flows under strong geodesic convexity. We also extend our analysis to the inexact proximal algorithm for geodesically semiconvex objectives. In our numerical experiments, proximal training demonstrates a faster convergence rate than the noisy gradient descent algorithm on mean-field neural networks.", "keywords": [ "optimization in Wasserstein space", "proximal algorithm", "gradient flow", "sampling", "mean-field" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "WPsnH6875d", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Lan-Zhe Guo", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Assistant Professor" }, { "name": "Rundong He", "gender": "Male", "institution": "Shandong University", "country": "", "position": "PhD student" }, { "name": "Tailin Wu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yicong Dong", "gender": "Male", "institution": "Shandong University", "country": "", "position": "PhD student" }, { "name": "Yilong Yin", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 18, "n_ref": 33, "n_ref_all": 65, "n_fig": 4, "n_tab": 10, "L_tab": 8753, "n_element_tab": 819, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 6743, "n_element_tab_1": 495, "formula_len_all": 434, "formula_len_all_1": 355, "len_all": 159278, "len_all_1": 73643, "len_abs": 2349, "len_title": 138, "len_sents": 50673, "len_sents_1": 28252, "n_sents": 357, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1446, "title": "Re-Evaluating the Impact of Unseen-Class Unlabeled Data on Semi-Supervised Learning Model", "abs": "Semi-supervised learning (SSL) effectively leverages unlabeled data and has been proven successful across various fields. Current safe SSL methods believe that unseen classes in unlabeled data harm the performance of SSL models. However, previous methods for assessing the impact of unseen classes on SSL model performance are flawed. They fix the size of the unlabeled dataset and adjust the proportion of unseen classes within the unlabeled data to assess the impact. This process contravenes the principle of controlling variables. Adjusting the proportion of unseen classes in unlabeled data alters the proportion of seen classes, meaning the decreased classification performance of seen classes may not be due to an increase in unseen class samples in the unlabeled data, but rather a decrease in seen class samples. Thus, the prior flawed assessment standard that \"unseen classes in unlabeled data can damage SSL model performance\" may not always hold true. This paper strictly adheres to the principle of controlling variables, maintaining the proportion of seen classes in unlabeled data while only changing the unseen classes across five critical dimensions, to investigate their impact on SSL models from global robustness and local robustness. Experiments demonstrate that unseen classes in unlabeled data do not necessarily impair the performance of SSL models; in fact, under certain conditions, unseen classes may even enhance them.", "keywords": [ "Safe Semi-Supervised Learning", "Unseen-Class Unlabeled Data" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "WPDsT4eIKH", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Jinjia Feng", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Taifeng Wang", "gender": "Male", "institution": "BioMap", "country": "", "position": "Principal Researcher" }, { "name": "Zhewei Wei", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Zongyang Qiu", "gender": "Male", "institution": "BioMap Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 59, "n_ref": 96, "n_ref_all": 113, "n_fig": 0, "n_tab": 16, "L_tab": 7192, "n_element_tab": 811, "n_fig_1": 0, "n_tab_1": 6, "L_tab_1": 3681, "n_element_tab_1": 425, "formula_len_all": 1073, "formula_len_all_1": 746, "len_all": 219282, "len_all_1": 72214, "len_abs": 1777, "len_title": 147, "len_sents": 56522, "len_sents_1": 30527, "n_sents": 379, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1788, "title": "Angle Graph Transformer: Capturing Higher-Order Structures for Accurate Molecular Geometry Learning", "abs": "Existing Graph Transformer models primarily focus on leveraging atomic and chemical bond properties along with basic geometric structures to learn representations of fundamental elements in molecular graphs, such as nodes and edges. However, higher-order structures like bond angles and torsion angles, which significantly influence key molecular properties, have not received sufficient attention. This oversight leads to inadequate geometric conformation accuracy and difficulties in precise local chirality determination, thereby limiting model performance in molecular property prediction tasks.\nTo address this issue, we propose the $A$ngle $G$raph $T$ransformer ($AGT$). AGT directly models directed bond angles and torsion angles, introducing higher-order structural representations to molecular graph learning for the first time. This approach enables AGT to determine local chirality within molecular representations and directly predict torsion angles. We introduce a novel directed cycle angle loss, allowing AGT to predict bond angles and torsion angles from low-precision molecular conformations. These properties, along with interatomic distances, are then applied to downstream molecular property prediction tasks using a pre-trained AGT with hierarchical virtual nodes.\nOur model achieves new state-of-the-art (SOTA) results on the PCQM4Mv2 and OC20 IS2RE datasets. Through transfer learning, AGT also demonstrates competitive performance on molecular property prediction benchmarks including QM9, LIT-PCBA, MOLPCBA, and MOLHIV. Further ablation studies reveal that the conformations generated by AGT are closest to conformations generated by Density Functional Theory (DFT) among the existing methods, due to the constraints imposed by the bond angles and torsion angles.", "keywords": [ "Molecular Representation Learning", "Graph Transformer", "Molecular Geometry Pretraining" ], "rating_list": [ 8, 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 1, 2 ], "confidence_list": [ 3, 4, 4, 5, 4 ] }, { "paper_id": "WOzffPgVjF", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Chenxi Luo", "gender": "unknown", "institution": "La Trobe University", "country": "AU", "position": "PhD student" }, { "name": "Libo Zhang", "gender": "Male", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Tiejian Luo", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xin Gu", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "UCAS" }, { "name": "Yan Huang", "gender": "Female", "institution": ", University of North Texas", "country": "US", "position": "Full Professor" }, { "name": "Yaojie Shen", "gender": "unknown", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Yuewei Lin", "gender": "Male", "institution": "Brookhaven national laboratory", "country": "US", "position": "Scientist" }, { "name": "Heng Fan", "gender": "Male", "institution": "University of North Texas", "country": "US", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 45, "n_ref": 105, "n_ref_all": 133, "n_fig": 14, "n_tab": 16, "L_tab": 3721, "n_element_tab": 483, "n_fig_1": 8, "n_tab_1": 7, "L_tab_1": 1891, "n_element_tab_1": 228, "formula_len_all": 931, "formula_len_all_1": 913, "len_all": 181785, "len_all_1": 74159, "len_abs": 3025, "len_title": 139, "len_sents": 56530, "len_sents_1": 32750, "n_sents": 434, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 91, "L_abs": 1900, "title": "Knowing Your Target : Target-Aware Transformer Makes Better Spatio-Temporal Video Grounding", "abs": "Transformer has attracted increasing interest in spatio-temporal video grounding, or STVG, owing to its end-to-end pipeline and promising result. Existing Transformer-based STVG approaches often leverage a set of object queries, which are initialized simply using zeros and then gradually learn target position information via iterative interactions with multimodal features, for spatial and temporal localization. Despite simplicity, these zero object queries, due to lacking target-specific cues, are hard to learn discriminative target information from interactions with multimodal features in the complicated scenarios (e.g., with distractors or occlusion), resulting in degradation. Addressing this, we introduce a novel Target-Aware Transformer for STVG (TA-STVG), which seeks to adaptively generate object queries via exploring target-specific cues from the given video-text pair, for improving STVG. The key lies in two simple yet effective modules, comprising text-guided temporal sampling (TTS) and attribute-aware spatial activation (ASA), working in a cascade. The former focuses on selecting target-relevant temporal cues from a video utilizing holistic text information, while the latter aims at further exploiting the fine-grained visual attribute information of the object from previous target-aware temporal cues, which is applied for object query initialization. Compared to existing methods leveraging zero-initialized queries, object queries in our TA-STVG, directly generated from a given video-text pair, naturally carry target-specific cues, making them adaptive and better interact with multimodal features for learning more discriminative information to improve STVG. In our experiments on three benchmarks, including HCSTVG-v1/-v2 and VidSTG, TA-STVG achieves state-of-the-art performance and largely outperforms the baseline, validating its efficacy. Code will be released.", "keywords": [ "Spatio-Temporal Video Grounding" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "WOyjgWu92E", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Florian Gr\u00f6tschla", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Roger Wattenhofer", "gender": "Not Specified", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "JIAQING XIE", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "MS student" } ], "n_formula": 32, "n_formula_1": 7, "n_ref_uni": 61, "n_ref": 183, "n_ref_all": 197, "n_fig": 4, "n_tab": 16, "L_tab": 10089, "n_element_tab": 1105, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1360, "n_element_tab_1": 89, "formula_len_all": 2403, "formula_len_all_1": 396, "len_all": 225623, "len_all_1": 70070, "len_abs": 1052, "len_title": 113, "len_sents": 57242, "len_sents_1": 33135, "n_sents": 421, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1057, "title": "Benchmarking Positional Encodings for GNNs and Graph Transformers", "abs": "Recent advances in Graph Neural Networks (GNNs) and Graph Transformers (GTs) have been driven by innovations in architectures and Positional Encodings (PEs), which are critical for augmenting node features and capturing graph topology. PEs are essential for GTs, where topological information would otherwise be lost without message-passing. However, PEs are often tested alongside novel architectures, making it difficult to isolate their effect on established models. To address this, we present a comprehensive benchmark of PEs in a unified framework that includes both message-passing GNNs and GTs. We also establish theoretical connections between MPNNs and GTs and introduce a sparsified GRIT attention mechanism to examine the influence of global connectivity. Our findings demonstrate that previously untested combinations of GNN architectures and PEs can outperform existing methods, offering a more comprehensive picture of the state-of-the-art. To support future research and experimentation in our framework, we make the code publicly available.", "keywords": [ "positional encodings", "graph neural networks", "graph transformers", "benchmarking" ], "rating_list": [ 5, 5, 5, 5, 3 ], "soundness_list": [ 4, 3, 3, 2, 2 ], "presentation_list": [ 4, 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 1, 1 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "WOt1owGfuN", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Ali Anwar", "gender": "Male", "institution": "University of Minnesota", "country": "US", "position": "Assistant Professor" }, { "name": "Enmao Diao", "gender": "Male", "institution": "ColAI", "country": "CN", "position": "Researcher" }, { "name": "Jie Ding", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "Associate Professor" }, { "name": "Li Yang", "gender": "Male", "institution": "University of North Carolina at Charlotte", "country": "US", "position": "Assistant Professor" }, { "name": "Qi Le", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "", "position": "PhD student" }, { "name": "Xinran Wang", "gender": "Female", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" }, { "name": "Ziyan Wang", "gender": "Male", "institution": "University of North Carolina at Charlotte", "country": "US", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 20, "n_ref_uni": 21, "n_ref": 58, "n_ref_all": 88, "n_fig": 4, "n_tab": 21, "L_tab": 11594, "n_element_tab": 1151, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 4787, "n_element_tab_1": 340, "formula_len_all": 1316, "formula_len_all_1": 1083, "len_all": 181528, "len_all_1": 78393, "len_abs": 1563, "len_title": 122, "len_sents": 43884, "len_sents_1": 29026, "n_sents": 341, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1576, "title": "Probe Pruning: Accelerating LLMs through Dynamic Pruning via Model-Probing", "abs": "We introduce Probe Pruning (PP), a novel framework for online, dynamic, structured pruning of Large Language Models (LLMs) applied in a batch-wise manner. PP leverages the insight that not all samples and tokens contribute equally to the model's output, and probing a small portion of each batch effectively identifies crucial weights, enabling tailored dynamic pruning for different batches. It comprises three main stages: probing, history-informed pruning, and full inference. In the probing stage, PP selects a small yet crucial set of hidden states, based on residual importance, to run a few model layers ahead. During the history-informed pruning stage, PP strategically integrates the probing states with historical states. Subsequently, it structurally prunes weights based on the integrated states and the PP importance score, a metric developed specifically to assess the importance of each weight channel in maintaining performance. In the final stage, full inference is conducted on the remaining weights. A major advantage of PP is its compatibility with existing models, as it operates without requiring additional neural network modules or fine-tuning. Comprehensive evaluations of PP on LLaMA-2/3 and OPT models reveal that even minimal probing\u2014using just 1.5% of FLOPs\u2014can substantially enhance the efficiency of structured pruning of LLMs. For instance, when evaluated on LLaMA-2-7B with WikiText2, PP achieves a 2.56 times lower ratio of performance degradation per unit of latency reduction compared to the state-of-the-art method at a 40\\% pruning ratio.", "keywords": [ "Large Lanuage Model Pruning", "Probe Pruning" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "WOopKWDWtS", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Ahmad Bdeir", "gender": "Male", "institution": "Universit\u00e4t Hildesheim", "country": "DE", "position": "PhD student" }, { "name": "Niels Landwehr", "gender": "Male", "institution": "Universit\u00e4t Hildesheim", "country": "DE", "position": "Full Professor" } ], "n_formula": 24, "n_formula_1": 17, "n_ref_uni": 35, "n_ref": 70, "n_ref_all": 84, "n_fig": 3, "n_tab": 4, "L_tab": 1458, "n_element_tab": 120, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1498, "formula_len_all_1": 1351, "len_all": 124760, "len_all_1": 62566, "len_abs": 1396, "len_title": 122, "len_sents": 36411, "len_sents_1": 30391, "n_sents": 262, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1433, "title": "Optimizing Learning for Robust Hyperbolic Deep Learning in Computer Vision", "abs": "Hyperbolic deep learning has become a growing research direction in computer vision for the unique properties afforded by the alternate embedding space. The negative curvature and exponentially growing distance metric provide a natural framework for capturing hierarchical relationships between datapoints and allowing for finer separability between their embeddings. However, these methods are still computationally expensive and prone to instability, especially when attempting to learn the negative curvature that best suits the task and the data. Current Riemannian optimizers do not account for changes in the manifold which greatly harms performance and forces lower learning rates to minimize projection errors. Our paper focuses on improving stability for curvature learning by introducing an improved schema for popular learning algorithms and providing a novel normalization approach to constrain embeddings within the variable representative radius of the manifold. Additionally, we introduce a novel formulation for Riemannian AdamW, and alternative hybrid encoder techniques and foundational formulations for current convolutional hyperbolic operations, greatly reducing the computational penalty of the hyperbolic embedding space. Our approach demonstrates consistent performance improvements across direct classification, generation, and hierarchical metric learning tasks while allowing for larger hyperbolic models.", "keywords": [ "Hyperbolic machine learning", "hyperbolic computer vision" ], "rating_list": [ 5, 6, 5, 3, 3 ], "soundness_list": [ 3, 2, 3, 2, 1 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2, 1 ], "confidence_list": [ 4, 3, 3, 4, 3 ] }, { "paper_id": "WNvvwK0tut", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Chao Du", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Chongxuan Li", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Fengqi Zhu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Guangtao Zeng", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "", "position": "PhD student" }, { "name": "Min Lin", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Principal Researcher" }, { "name": "Qian Liu", "gender": "Male", "institution": "Tiktok", "country": "SG", "position": "Researcher" }, { "name": "Shen Nie", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "PhD student" }, { "name": "Tianyu Pang", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" } ], "n_formula": 18, "n_formula_1": 11, "n_ref_uni": 69, "n_ref": 158, "n_ref_all": 196, "n_fig": 8, "n_tab": 12, "L_tab": 2118, "n_element_tab": 334, "n_fig_1": 3, "n_tab_1": 11, "L_tab_1": 1733, "n_element_tab_1": 139, "formula_len_all": 1234, "formula_len_all_1": 742, "len_all": 211148, "len_all_1": 68828, "len_abs": 1478, "len_title": 90, "len_sents": 55125, "len_sents_1": 30488, "n_sents": 452, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 1487, "title": "Scaling up Masked Diffusion Models on Text", "abs": "Masked diffusion models (MDMs) have shown promise in language modeling, yet their scalability and effectiveness in core language tasks, such as conditional generation and language understanding, remain underexplored. This paper establishes the first scaling law for MDMs, demonstrating a scaling rate comparable to autoregressive models (ARMs) and a relatively small compute gap. Motivated by their scalability, we train a family of MDMs with up to 1.1 billion (B) parameters to systematically evaluate their performance against ARMs of comparable or larger sizes. Fully leveraging the probabilistic formulation of MDMs, we propose a simple yet effective *unsupervised classifier-free guidance* that effectively exploits large-scale unpaired data, boosting performance for conditional inference. In language understanding, a 1.1B MDM shows competitive results, outperforming the larger 1.5B GPT-2 model on four out of eight zero-shot benchmarks. In conditional generation, MDMs provide a flexible trade-off compared to ARMs utilizing KV-cache: MDMs match the performance of ARMs while being 1.5 times faster, or achieve higher quality than ARMs at a slightly higher computational cost. Moreover, MDMs address challenging tasks for ARMs by effectively handling bidirectional reasoning and adapting to temporal shifts in data. Notably, a 1.1B MDM breaks the *reverse curse* encountered by much larger ARMs with significantly more data and computation, such as Llama (13B) and GPT-3 (175B).", "keywords": [ "Masked Diffusion Models", "Scaling Laws", "Conditional Generation", "Language Understanding", "Reverse Curse", "Temporal Quality Degradation" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "WNb4P8aG66", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "LEI BAI", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Luping Zhou", "gender": "Female", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Shuyang Sun", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Wanli Ouyang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoyu Yue", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Zeyu Lu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "ZiDong Wang", "gender": "Male", "institution": "Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "MENG WEI", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 10, "n_ref_uni": 53, "n_ref": 120, "n_ref_all": 134, "n_fig": 5, "n_tab": 6, "L_tab": 2141, "n_element_tab": 335, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 2141, "n_element_tab_1": 335, "formula_len_all": 315, "formula_len_all_1": 357, "len_all": 139596, "len_all_1": 65274, "len_abs": 1167, "len_title": 104, "len_sents": 32324, "len_sents_1": 27839, "n_sents": 252, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1184, "title": "Diffusion Models Need Visual Priors for Image Generation", "abs": "Conventional class-guided diffusion models generally succeed in generating images with correct semantic content, but often struggle with texture details. This limitation stems from the usage of class priors, which only provide coarse and limited conditional information. To address this issue, we propose Diffusion on Diffusion (DoD), an innovative multi-stage generation framework that first extracts visual priors from previously generated samples, then provides rich guidance for the diffusion model leveraging visual priors from the early stages of diffusion sampling. Specifically, we introduce a latent embedding module that employs a compression-reconstruction approach to discard redundant detail information from the conditional samples in each stage, retaining only the semantic information for guidance. We evaluate DoD on the popular ImageNet-$256 \\times 256$ dataset, reducing 7$\\times$ training cost compared to SiT and DiT with even better performance in terms of the FID-50K score. Our largest model DoD-XL achieves an FID-50K score of 1.83 with only 1 million training steps, which surpasses other state-of-the-art methods without bells and whistles during inference.", "keywords": [ "Diffusion Model", "Image Generation", "Generative Model" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "WNZNsyzcaB", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Eunil Park", "gender": "Male", "institution": "Sungkyunkwan University", "country": "KR", "position": "Associate Professor" }, { "name": "Hyeonbin Ji", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "MS student" }, { "name": "Minyoung Lee", "gender": "unknown", "institution": "Sungkyunkwan University", "country": "KR", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 65, "n_ref": 93, "n_ref_all": 104, "n_fig": 4, "n_tab": 6, "L_tab": 2129, "n_element_tab": 221, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1060, "n_element_tab_1": 117, "formula_len_all": 963, "formula_len_all_1": 623, "len_all": 165506, "len_all_1": 61980, "len_abs": 1463, "len_title": 134, "len_sents": 44083, "len_sents_1": 28405, "n_sents": 320, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1483, "title": "Enhancing Deception Detection with Cognitive Load Features: An Audio-Visual Approach", "abs": "Deception ranges from minor mischief to serious fraud, often leading to significant psychological and financial harm. Effective deception detection is crucial to mitigate these risks and preserve societal trust. Cognitive load is a useful indicator for detecting deception, as lying causes individuals to experience greater mental strain. While prior research leveraged cognitive load features, typically measured through physiological signals such as pupil dilation, these methods often require specialized equipment and can be subject to human bias. These limitations hinder the scalability and automation of deception detection systems. Thus, we propose a novel deception detection framework that automatically extracts cognitive load features from audio-visual data, eliminating the need for specialized hardware or subjective human input. Our approach integrates these features into the deception detection pipeline, enhancing its robustness. Moreover, we introduce a focal loss to address the inherent complexity of deception detection. This objective function enables the model to focus on harder-to-detect instances of deception, thereby improving the performance. Our approach achieves state-of-the-art results on benchmark audio-visual datasets, demonstrating significant improvements in automated deception detection. Extensive experiments validate the effectiveness of both our cognitive load feature extraction and the proposed objective function in advancing the field.", "keywords": [ "deception detection", "cognitive load", "audio visual" ], "rating_list": [ 6, 8, 5, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "WNPrfGpcu6", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Dong Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Jie Huang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Qi Zhu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Senyan Xu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "YiDi Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Zheng-Jun Zha", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xueyang Fu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 26, "n_formula_1": 15, "n_ref_uni": 54, "n_ref": 137, "n_ref_all": 185, "n_fig": 17, "n_tab": 21, "L_tab": 5027, "n_element_tab": 608, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 4417, "n_element_tab_1": 223, "formula_len_all": 1294, "formula_len_all_1": 775, "len_all": 200508, "len_all_1": 66443, "len_abs": 2070, "len_title": 134, "len_sents": 49860, "len_sents_1": 27922, "n_sents": 384, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1990, "title": "FourierMamba: Fourier Learning Integration with State Space Models for Image Deraining", "abs": "Image deraining aims to remove rain streaks from rainy images and restore clear backgrounds. Currently, some research that employs the Fourier transform has proved to be effective for image deraining, due to it acting as an effective frequency prior for capturing rain streaks. However, despite there exists dependency of low frequency and high frequency in images, these Fourier-based methods rarely exploit the correlation of different frequencies for conjuncting their learning procedures, limiting the full utilization of frequency information for image deraining. Alternatively, the recently emerged Mamba technique depicts its effectiveness and efficiency for modeling correlation in various domains (e.g., spatial, temporal), and we argue that introducing Mamba into its unexplored Fourier spaces to correlate different frequencies would help improve image deraining. This motivates us to propose a new framework termed FourierMamba, which performs image deraining with Mamba in the Fourier space. Owing to the unique arrangement of frequency orders in Fourier space, the core of FourierMamba lies in the scanning encoding of different frequencies, where the low-high frequency order formats exhibit differently in the spatial dimension (unarranged in axis) and channel dimension (arranged in axis). Therefore, we design FourierMamba that correlates Fourier space information in the spatial and channel dimensions with distinct designs. Specifically, in the spatial dimension Fourier space, we introduce the zigzag coding to scan the frequencies to rearrange the orders from low to high frequencies, thereby orderly correlating the connections between frequencies; in the channel dimension Fourier space with arranged orders of frequencies in axis, we can directly use Mamba to perform frequency correlation and improve the channel information representation. Extensive experiments reveal that our method outperforms state-of-the-art methods both qualitatively and quantitatively.", "keywords": [ "Fourier transform", "mamba", "image deraining" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "WNIEr5kydF", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Eunho Yang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Hyunjin Seo", "gender": "Female", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sungsoo Ahn", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Taewon Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 29, "n_formula_1": 27, "n_ref_uni": 39, "n_ref": 115, "n_ref_all": 134, "n_fig": 5, "n_tab": 4, "L_tab": 2396, "n_element_tab": 413, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2385, "n_element_tab_1": 410, "formula_len_all": 2343, "formula_len_all_1": 1815, "len_all": 153147, "len_all_1": 72154, "len_abs": 1072, "len_title": 144, "len_sents": 43453, "len_sents_1": 28224, "n_sents": 306, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1085, "title": "REBIND: Enhancing Ground-state Molecular Conformation Prediction via Force-Based Graph Rewiring", "abs": "Predicting the ground-state 3D molecular conformations from 2D molecular graphs is critical in computational chemistry due to its profound impact on molecular properties. Deep learning (DL) approaches have recently emerged as promising alternatives to computationally-heavy classical methods such as density functional theory (DFT). However, we discover that existing DL methods inadequately model inter-atomic forces, particularly for non-bonded atomic pairs, due to their naive usage of bonds and pairwise distances. Consequently, significant prediction errors occur for atoms with low degree (*i.e.*, low coordination numbers) whose conformations are primarily influenced by non-bonded interactions. To address this, we propose **ReBind**, a novel framework that rewires molecular graphs by adding edges based on the Lennard-Jones potential to capture non-bonded interactions for low-degree atoms. Experimental results demonstrate that **ReBind** significantly outperforms state-of-the-art methods across various molecular sizes, achieving up to a 20% reduction in prediction error.", "keywords": [ "molecular conformation prediction", "molecule modeling", "graph neural network" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "WM4xiEDz2N", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Dong Yuan", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "NAN YANG", "gender": "Female", "institution": "University of Sydney", "country": "AU", "position": "Researcher" }, { "name": "Shuai Wang", "gender": "Male", "institution": "Hainan University", "country": "CN", "position": "Independent Researcher" }, { "name": "Xuanyu Chen", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" } ], "n_formula": 89, "n_formula_1": 31, "n_ref_uni": 49, "n_ref": 104, "n_ref_all": 163, "n_fig": 6, "n_tab": 3, "L_tab": 274, "n_element_tab": 44, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9814, "formula_len_all_1": 1384, "len_all": 166684, "len_all_1": 67189, "len_abs": 1742, "len_title": 156, "len_sents": 45319, "len_sents_1": 32427, "n_sents": 354, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1749, "title": "Generalization Performance Gap Analysis between Centralized and Federated Learning: How to Bridge this Gap?", "abs": "The rising interest in decentralized data and privacy protection has led to the emergence of Federated Learning. Many studies have compared federated training with classical training approaches using centralized data and found from experiments that models trained in a federated setup with equal resources perform poorly on tasks. However, these studies have generally been empirical and have not explored the performance gap further from a theoretical perspective. The lack of theoretical understanding prevents figuring out whether federated algorithms are necessarily inferior to centralized algorithms in performance and how large this gap is according to the training settings. Also, it hinders identifying valid ways to close this performance distance. This paper fills this theoretical gap by formulating federated training as an SGD (Stochastic Gradient Descent) optimization problem over decentralized data and defining the performance gap within the PAC-Bayes (Probably Approximately Correct Bayesian) framework. Through theoretical analysis, we derive non-vacuous bounds on this performance gap, revealing that the difference in generalization performance necessarily exists when training resources are equal for both training setups and that variations in the training parameters affect the gap. Moreover, we also prove that the complete elimination of the performance gap is only possible by introducing new clients or adding new data to existing clients. Advantages in other training resources are not feasible for closing the gap, such as giving larger models or more communication rounds to federated scenarios. Our theoretical findings are validated by extensive experimental results from different model architectures and datasets.", "keywords": [ "Federated Learning", "Generalization Performance", "Centralized Training", "Theoretical Analysis" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "WLSrq1254E", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Chengxing Jia", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Fuxiang Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Lei Yuan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Researcher" }, { "name": "Wenjie Qiu", "gender": "unknown", "institution": "nanjing university", "country": "", "position": "MS student" }, { "name": "Yi-Chen Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Zongzhang Zhang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Yang Yu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Professor" } ], "n_formula": 40, "n_formula_1": 22, "n_ref_uni": 38, "n_ref": 100, "n_ref_all": 130, "n_fig": 6, "n_tab": 6, "L_tab": 1659, "n_element_tab": 244, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1728, "n_element_tab_1": 162, "formula_len_all": 4187, "formula_len_all_1": 1780, "len_all": 219401, "len_all_1": 66778, "len_abs": 2083, "len_title": 133, "len_sents": 58540, "len_sents_1": 28580, "n_sents": 491, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1826, "title": "Q-Adapter: Customizing Pre-trained LLMs to New Preferences with Forgetting Mitigation", "abs": "Large Language Models (LLMs), trained on a large amount of corpus, have demonstrated remarkable abilities. However, it may not be sufficient to directly apply open-source LLMs like Llama to certain real-world scenarios, since most of them are trained for \\emph{general} purposes. Thus, the demands for customizing publicly available LLMs emerge, but are currently under-studied. In this work, we consider customizing pre-trained LLMs with new human preferences. Specifically, the LLM should not only meet the new preference but also preserve its original capabilities after customization. Drawing inspiration from the observation that human preference can be expressed as a reward model, we propose to cast LLM customization as optimizing the sum of two reward functions, one of which (denoted as $r_1$) was used to pre-train the LLM while the other (denoted as $r_2$) characterizes the new human preference. The obstacle here is that both reward functions are unknown, making the application of modern reinforcement learning methods infeasible. Thanks to the residual Q-learning framework, we can restore the customized LLM with the pre-trained LLM and the \\emph{residual Q-function} without the reward function $r_1$. Moreover, we find that for a fixed pre-trained LLM, the reward function $r_2$ can be derived from the residual Q-function, enabling us to directly learn the residual Q-function from the new human preference data upon the Bradley-Terry model. We name our method Q-Adapter as it introduces an adapter module to approximate the residual Q-function for customizing the pre-trained LLM towards the new preference. Experiments based on the Llama-3.1 model on the DSP dataset and HH-RLHF dataset illustrate the superior effectiveness of Q-Adapter on both retaining existing knowledge and learning new preferences.", "keywords": [ "reinforcement learning from human feedback", "large languge model customization", "anti-forgetting" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "WL4BmXG7Pl", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Shenyang Deng", "gender": "Male", "institution": "Dartmouth College", "country": "", "position": "PhD student" }, { "name": "Tianyu Pang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Kothapalli Vignesh", "gender": "Male", "institution": "LinkedIn", "country": "US", "position": "Researcher" }, { "name": "Yaoqing Yang", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "Assistant Professor" }, { "name": "Zongmin Liu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Researcher" } ], "n_formula": 87, "n_formula_1": 27, "n_ref_uni": 53, "n_ref": 139, "n_ref_all": 196, "n_fig": 42, "n_tab": 6, "L_tab": 81, "n_element_tab": 3, "n_fig_1": 15, "n_tab_1": 2, "L_tab_1": 25, "n_element_tab_1": 1, "formula_len_all": 6288, "formula_len_all_1": 1204, "len_all": 197790, "len_all_1": 65875, "len_abs": 1283, "len_title": 117, "len_sents": 56089, "len_sents_1": 29295, "n_sents": 449, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1290, "title": "Crafting Heavy-Tails in Weight Matrix Spectrum without Gradient Noise", "abs": "Training strategies for modern deep neural networks (NNs) tend to induce a heavy-\ntailed (HT) empirical spectral density (ESD) in the layer weights. While previous\nefforts have shown that the HT phenomenon correlates with good generalization\nin large NNs, a theoretical explanation of its occurrence is still lacking. Especially,\nunderstanding the conditions which lead to this phenomenon can shed light on the\ninterplay between generalization and weight spectra. Our work aims to bridge this\ngap by presenting a simple, rich setting to model the emergence of HT ESD. In\nparticular, we present a theory-informed analysis for 'crafting' heavy tails in the\nESD of two-layer NNs without any gradient noise. This is the first work to analyze a noise-free setting and incorporate optimizer (GD/Adam) dependent (large)\nlearning rates into the HT ESD analysis. Our results highlight the role of learning\nrates on the Bulk+Spike and HT shape of the ESDs in the early phase of training,\nwhich can facilitate generalization in the two-layer NN. These observations shed\nlight on the behavior of large-scale NNs, albeit in a much simpler setting. Last\nbut not least, we present a novel perspective on the ESD evolution dynamics by\nanalyzing the singular vectors of weight matrices and optimizer updates", "keywords": [ "Heavy Tails", "Spectral Analysis", "Generalization" ], "rating_list": [ 3, 6, 8, 3 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "WKfb1xGXGx", "primary_area": "generative models", "n_author": 11, "author_info_list": [ { "name": "Chengan He", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Dominik Michels", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Associate Professor" }, { "name": "Fujun Luan", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Holly Rushmeier", "gender": "Female", "institution": "Yale University", "country": "", "position": "Full Professor" }, { "name": "Jorge Alejandro Amador Herrera", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Meng Zhang", "gender": "Female", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Soren Pirk", "gender": "Not Specified", "institution": "Kiel University", "country": "DE", "position": "Full Professor" }, { "name": "Tuanfeng Yang Wang", "gender": "Male", "institution": "Adobe Systems", "country": "GB", "position": "Researcher" }, { "name": "Xin Sun", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Yi Zhou", "gender": "Female", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" }, { "name": "Zhixin Shu", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 54, "n_ref": 139, "n_ref_all": 205, "n_fig": 31, "n_tab": 3, "L_tab": 595, "n_element_tab": 148, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 339, "n_element_tab_1": 95, "formula_len_all": 742, "formula_len_all_1": 487, "len_all": 186717, "len_all_1": 56553, "len_abs": 1512, "len_title": 114, "len_sents": 58372, "len_sents_1": 26361, "n_sents": 381, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 66, "L_abs": 987, "title": "Perm: A Parametric Representation for Multi-Style 3D Hair Modeling", "abs": "We present Perm, a learned parametric representation of human 3D hair designed to facilitate various hair-related applications. Unlike previous work that jointly models the global hair structure and local curl patterns, we propose to disentangle them using a PCA-based strand representation in the frequency domain, thereby allowing more precise editing and output control. Specifically, we leverage our strand representation to fit and decompose hair geometry textures into low- to high-frequency hair structures. These decomposed textures are later parameterized with different generative models, emulating common stages in the hair grooming process. We conduct extensive experiments to validate the architecture design of \\textsc{Perm}, and finally deploy the trained model as a generic prior to solve task-agnostic problems, further showcasing its flexibility and superiority in tasks such as single-view hair reconstruction, hairstyle editing, and hair-conditioned image generation.", "keywords": [ "Hair Modeling", "Parametric Models", "Generative Models" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 4, 2 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "WKW5TG8ItY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Ali Ebrahimpour-Boroojeny", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Hari Sundaram", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Varun Chandrasekaran", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 26, "n_formula_1": 12, "n_ref_uni": 38, "n_ref": 107, "n_ref_all": 156, "n_fig": 9, "n_tab": 14, "L_tab": 670, "n_element_tab": 72, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 411, "n_element_tab_1": 44, "formula_len_all": 2399, "formula_len_all_1": 1030, "len_all": 185858, "len_all_1": 71686, "len_abs": 1665, "len_title": 115, "len_sents": 73110, "len_sents_1": 36364, "n_sents": 442, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1687, "title": "Training Robust Ensembles Requires Rethinking Lipschitz Continuity", "abs": "Transferability of adversarial examples is a well-known property that endangers all classification models, even those that are only accessible through black-box queries. Prior work has shown that an ensemble of models is more resilient to transferability: the probability that an adversarial example is effective against most models of the ensemble is low. Thus, most ongoing research focuses on improving ensemble diversity. Another line of prior work has shown that Lipschitz continuity of the models can make models more robust since it limits how a model's output changes with small input perturbations. {\\em In this paper, we study the effect of Lipschitz continuity on transferability rates.} We show that although a lower Lipschitz constant increases the robustness of a single model, it is not as beneficial in training robust ensembles as it increases the transferability rate of adversarial examples across models in the ensemble. Therefore, we introduce LOTOS, a new training paradigm for ensembles, which counteracts this adverse effect. It does so by promoting orthogonality among the top-$k$ sub-spaces of the transformations of the corresponding affine layers of any pair of models in the ensemble. We theoretically show that $k$ does not need to be large for convolutional layers, which makes the computational overhead negligible. Through various experiments, we show LOTOS increases the robust accuracy of ensembles of ResNet-18 models by $6$ percentage points (p.p) against black-box attacks on CIFAR-10. It is also capable of combining with the robustness of prior state-of-the-art methods for training robust ensembles to enhance their robust accuracy by $10.7$ p.p.", "keywords": [ "robustness", "lipschitzness", "ensembles", "convolutions", "transferability", "black-box attacks" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "WKKD1Faobu", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Hieu Trung Nguyen", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Huaizu Jiang", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Assistant Professor" }, { "name": "Vikram Voleti", "gender": "Male", "institution": "Stability AI", "country": "CA", "position": "Researcher" }, { "name": "YIWEN CHEN", "gender": "Female", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Varun Jampani", "gender": "Male", "institution": "Stability AI", "country": "US", "position": "Principal Researcher" } ], "n_formula": 8, "n_formula_1": 4, "n_ref_uni": 67, "n_ref": 160, "n_ref_all": 193, "n_fig": 12, "n_tab": 4, "L_tab": 643, "n_element_tab": 119, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 219, "n_element_tab_1": 27, "formula_len_all": 450, "formula_len_all_1": 304, "len_all": 192098, "len_all_1": 62762, "len_abs": 1022, "len_title": 0, "len_sents": 47286, "len_sents_1": 28723, "n_sents": 343, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 985, "title": "HouseCrafter: Lifting Floorplans to 3D Scenes with 2D Diffusion Model", "abs": "We introduce HouseCrafter, a novel approach that can lift a floorplan into a complete large 3D indoor scene (e.g., a house). Our key insight is to adapt a 2D diffusion model, which is trained on web-scale images, to generate consistent multi-view color (RGB) and depth (D) images across different locations of the scene. Specifically, the RGB-D images are generated autoregressively in a batch-wise manner along sampled locations based on the floorplan, where previously generated images are used as condition to the diffusion model to produce images at nearby locations. The global floorplan and attention design in the diffusion model ensures the consistency of the generated images, from which a 3D scene can be reconstructed. Through extensive evaluation on the 3D-Front dataset, we demonstrate that HouseCraft can generate high-quality house-scale 3D scenes. Ablation studies also validate the effectiveness of different design choices. We will release our code and model weights.", "keywords": [ "diffusion model", "indoor scene generation" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 4, 2 ] }, { "paper_id": "WK6hQoAtgx", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Fengbin ZHU", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Fuli Feng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Haohui Wu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Undergrad student" }, { "name": "LUAN HUANBO", "gender": "Male", "institution": "6Estates Pte Ltd", "country": "SG", "position": "Researcher" }, { "name": "Ng Xiang Yao", "gender": "unknown", "institution": "6Estates", "country": "SG", "position": "Researcher" }, { "name": "Wenjie Wang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "chao wang", "gender": "Male", "institution": "6Estates Pte Ltd", "country": "SG", "position": "Researcher" }, { "name": "Tat-seng Chua", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Liu Ziyang", "gender": "Male", "institution": "6Estates PTE. LTD.", "country": "SG", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 42, "n_ref": 110, "n_ref_all": 130, "n_fig": 6, "n_tab": 11, "L_tab": 8004, "n_element_tab": 1159, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 5795, "n_element_tab_1": 471, "formula_len_all": 21, "formula_len_all_1": 21, "len_all": 199777, "len_all_1": 61220, "len_abs": 1474, "len_title": 149, "len_sents": 33853, "len_sents_1": 24432, "n_sents": 222, "n_sents_1": 153, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1244, "title": "MMDocBench: Benchmarking Large Vision-Language Models for Fine-Grained Visual Document Understanding", "abs": "Large Vision-Language Models (LVLMs) have achieved remarkable performance in many vision-language tasks, yet their capabilities in fine-grained visual understanding remain insufficiently evaluated. Existing benchmarks either contain limited fine-grained evaluation samples that are mixed with other data, or are confined to object-level assessments in natural images. To holistically assess LVLMs' fine-grained visual understanding capabilities, we propose using document images with multi-granularity and multi-modal information to supplement natural images. In this light, we construct MMDocBench, a benchmark with various OCR-free document understanding tasks for the evaluation of fine-grained visual perception and reasoning abilities. MMDocBench defines 15 main tasks with 4,338 QA pairs and 11,353 supporting regions, covering various document images such as research papers, receipts, financial reports, Wikipedia tables, charts, and infographics. Based on MMDocBench, we conduct extensive experiments using 10 open-source and 3 proprietary advanced LVLMs, assessing their strengths and weaknesses across different tasks and document image types. The benchmark, task instructions, and evaluation code will be made publicly available.", "keywords": [ "Comprehensive Benchmark", "Large Vision-Language Models", "Fine-grained visual understanding", "OCR-Free Document Understanding" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 5, 2 ] }, { "paper_id": "WK6K1FMEQ1", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Erik Wijmans", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Philipp Kraehenbuehl", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Santhosh Kumar Ramakrishnan", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Vladlen Koltun", "gender": "Male", "institution": "Apple", "country": "US", "position": "Distinguished Scientist" } ], "n_formula": 7, "n_formula_1": 1, "n_ref_uni": 101, "n_ref": 222, "n_ref_all": 1243, "n_fig": 17, "n_tab": 8, "L_tab": 4095, "n_element_tab": 577, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 866, "n_element_tab_1": 10, "formula_len_all": 921, "formula_len_all_1": 113, "len_all": 463905, "len_all_1": 79105, "len_abs": 1208, "len_title": 97, "len_sents": 110421, "len_sents_1": 34075, "n_sents": 1018, "n_sents_1": 291, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 750, "title": "Does Spatial Cognition Emerge in Frontier Models?", "abs": "Not yet. We present SPACE, a benchmark that systematically evaluates spatial cognition in frontier models. Our benchmark builds on decades of research in cognitive science. It evaluates large-scale mapping abilities that are brought to bear when an organism traverses physical environments, smaller-scale reasoning about object shapes and layouts, and cognitive infrastructure such as spatial attention and memory. For many tasks, we instantiate parallel presentations via text and images, allowing us to benchmark both large language models and large multimodal models. Results suggest that contemporary frontier models fall short of the spatial intelligence of animals, performing near chance level on a number of classic tests of animal cognition.", "keywords": [ "Frontier models", "spatial cognition" ], "rating_list": [ 6, 8, 8, 5 ], "soundness_list": [ 4, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "WJnciuhwyU", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 6, "author_info_list": [ { "name": "Anurag Singh", "gender": "Male", "institution": "CISPA, saarland university, saarland informatics campus", "country": "", "position": "PhD student" }, { "name": "Krikamol Muandet", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Associate Professor" }, { "name": "Masaki Adachi", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Siu Lun Chau", "gender": "Male", "institution": "CISPA \u2013 Helmholtz Center for Information Security", "country": "DE", "position": "Postdoc" }, { "name": "Wenjie Xu", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Michael Osborne", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" } ], "n_formula": 214, "n_formula_1": 32, "n_ref_uni": 55, "n_ref": 99, "n_ref_all": 155, "n_fig": 9, "n_tab": 7, "L_tab": 4502, "n_element_tab": 318, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 350, "n_element_tab_1": 28, "formula_len_all": 20889, "formula_len_all_1": 3915, "len_all": 315960, "len_all_1": 78842, "len_abs": 1462, "len_title": 121, "len_sents": 87711, "len_sents_1": 34254, "n_sents": 956, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1371, "title": "Social Bayesian Optimization for Building Truthful Consensus", "abs": "We introduce *Social Bayesian Optimization* (SBO), a query-efficient algorithm for consensus-building in collective decision-making. In contrast to single-agent scenarios, collective decision-making encompasses group dynamics that may distort agents' preference feedback, thereby impeding their capacity to achieve a truthful consensus. We demonstrate that under standard rationality assumptions, reaching truthful consensus\u2014the most preferable decision based on the aggregated latent agent utilities\u2014using noisy feedback alone is impossible. To address this, SBO employs a dual voting system: cost-effective but noisy public votes, and more accurate, though expensive, private votes. We model social influence using an unknown social graph and leverage the dual voting system to efficiently learn this graph. Our findings show that social graph estimation converges faster than the black-box estimation of agents\u2019 utilities, allowing us to reduce reliance on costly private votes early in the process. This enables efficient consensus-building primarily through noisy public votes, which are debiased based on the estimated social graph to infer truthful feedback. We validate the effectiveness of SBO across multiple real-world applications, including thermal comfort optimization, team building, travel destination discussion, and strategic alliance in energy trading.", "keywords": [ "Bayesian optimisation", "social choice theory", "preference learning" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 4, 2, 2, 3 ], "confidence_list": [ 2, 3, 4, 2 ] }, { "paper_id": "WJaUkwci9o", "primary_area": "learning theory", "n_author": 8, "author_info_list": [ { "name": "Adam Block", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" }, { "name": "Akshay Krishnamurthy", "gender": "Male", "institution": "Research, Microsoft", "country": "US", "position": "Senior Principal Research Manager" }, { "name": "Audrey Huang", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "PhD student" }, { "name": "Cyril Zhang", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Senior Researcher" }, { "name": "Dhruv Rohatgi", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Dylan J Foster", "gender": "unknown", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Jordan T. Ash", "gender": "unknown", "institution": "Microsoft Research", "country": "US", "position": "Researcher" }, { "name": "Max Simchowitz", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 192, "n_formula_1": 24, "n_ref_uni": 91, "n_ref": 197, "n_ref_all": 269, "n_fig": 10, "n_tab": 5, "L_tab": 222, "n_element_tab": 18, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 21356, "formula_len_all_1": 1748, "len_all": 370438, "len_all_1": 76875, "len_abs": 2462, "len_title": 109, "len_sents": 98493, "len_sents_1": 35205, "n_sents": 915, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 61, "L_abs": 1532, "title": "Self-Improvement in Language Models: The Sharpening Mechanism", "abs": "Recent work in language modeling has raised the possibility of \u201cself-improvement,\u201d where an LLM evaluates and refines its own generations to achieve higher performance without external feedback. It is impossible for this self-improvement to create information that is not already in the model, so why should we expect that this will lead to improved capabilities? We offer a new theoretical perspective on the capabilities of self-improvement through a lens we refer to as \u201csharpening.\u201d Motivated by the observation that language models are often better at verifying response quality than they are at generating correct responses, we formalize self-improvement as using the model itself as a verifier during post-training in order to \u2018sharpen\u2019 the model to one placing large mass on high-quality sequences, thereby amortizing the expensive inference-time computation of generating good sequences. We begin by introducing a new statistical framework for sharpening in which the learner has sample access to a pre-trained base policy. Then, we analyze two natural families of self improvement algorithms based on SFT and RLHF. We find that (i) the SFT-based approach is minimax optimal whenever the initial model has sufficient coverage, but (ii) the RLHF-based approach can improve over SFT-based self- improvement by leveraging online exploration, bypassing the need for coverage. We view these findings as a starting point toward a foundational understanding that can guide the design and evaluation of self-improvement algorithms.", "keywords": [ "Learning theory", "Sample complexity", "Self-Improvement", "Language Models" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "WIerHtNyKr", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Qing-xin Meng", "gender": "Male", "institution": "China University of Petroleum, Beijing", "country": "CN", "position": "PhD student" }, { "name": "Xia Lei", "gender": "unknown", "institution": "Communication University of China", "country": "", "position": "Postdoc" } ], "n_formula": 73, "n_formula_1": 38, "n_ref_uni": 38, "n_ref": 68, "n_ref_all": 87, "n_fig": 7, "n_tab": 10, "L_tab": 1663, "n_element_tab": 129, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 832, "n_element_tab_1": 55, "formula_len_all": 8670, "formula_len_all_1": 3319, "len_all": 159329, "len_all_1": 69359, "len_abs": 997, "len_title": 120, "len_sents": 40596, "len_sents_1": 28629, "n_sents": 343, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1006, "title": "Adaptive Algorithm for Non-Stationary Online Convex-Concave Optimization", "abs": "This paper addresses the problem of Online Convex-Concave Optimization, an extension of Online Convex Optimization to two-player time-varying convex-concave games. \nOur objective is to minimize the dynamic duality gap (D-DGap), a key performance metric that evaluates the players' strategies against arbitrary comparator sequences. \nExisting algorithms struggle to achieve optimal performance, particularly in stationary or predictable environments. \nWe propose a novel, modular algorithm comprising three key components: an Adaptive Module that adjusts to varying levels of non-stationarity, a Multi-Predictor Aggregator that selects the optimal predictor from multiple candidates, and an Integration Module that seamlessly combines the strengths of both. \nOur algorithm guarantees a minimax optimal D-DGap upper bound, up to a logarithmic factor, while also achieving a prediction error-based D-DGap bound. \nEmpirical results further demonstrate the effectiveness and adaptability of the proposed method.", "keywords": [ "Non-Stationary Online Learning", "Online Convex-Concave Optimization", "Dynamic Duality Gap", "Adaptive Algorithm" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "WHtNc5kX1v", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Bo Zhou", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Researcher" }, { "name": "Daniel Gei\u00dfler", "gender": "Male", "institution": "German Research Center for AI", "country": "", "position": "Researcher" }, { "name": "Mengxi Liu", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "PhD student" }, { "name": "Paul Lukowicz", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 37, "n_ref": 51, "n_ref_all": 67, "n_fig": 2, "n_tab": 7, "L_tab": 3365, "n_element_tab": 411, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 5091, "n_element_tab_1": 284, "formula_len_all": 1171, "formula_len_all_1": 905, "len_all": 120109, "len_all_1": 59837, "len_abs": 1244, "len_title": 142, "len_sents": 38037, "len_sents_1": 25963, "n_sents": 238, "n_sents_1": 160, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1247, "title": "Latent Boost: Leveraging Latent Space Distance Metrics to Augment Classification Performance", "abs": "The pursuit of boosting classification performance in Machine Learning has primarily focused on refining model architectures and hyperparameters through probabilistic loss optimization. However, such an approach often neglects the profound, untapped potential embedded in internal structural information, which can significantly elevate the training process. In this work, we introduce Latent Boost, a novel approach that incorporates the very definition of classification via latent representation distance metrics to enhance the conventional dataset-oriented classification training. Thus during training, the model is not only optimized for classification metrics of the discrete data points but also adheres to the rule that the collective representation zones of each class should be sharply clustered. By leveraging the rich structural insights of high-dimensional latent representations, Latent Boost not only improves classification metrics like F1-Scores but also brings additional benefits of improved interpretability with higher silhouette scores and steady-fast convergence with fewer training epochs. Latent Boost brings these performance and latent structural benefits with minimum additional cost and no data-specific requirements.", "keywords": [ "Latent Space", "Distance Metrics", "Magnet Loss" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 5, 3 ] }, { "paper_id": "WHhZv8X5zF", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Denny Wu", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "Murat A Erdogdu", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "Assistant Professor" }, { "name": "Alireza Mousavi", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "", "position": "PhD student" } ], "n_formula": 160, "n_formula_1": 33, "n_ref_uni": 66, "n_ref": 150, "n_ref_all": 156, "n_fig": 2, "n_tab": 3, "L_tab": 267, "n_element_tab": 31, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 4, "n_element_tab_1": 2, "formula_len_all": 14885, "formula_len_all_1": 2141, "len_all": 243272, "len_all_1": 74058, "len_abs": 1282, "len_title": 129, "len_sents": 65097, "len_sents_1": 33387, "n_sents": 685, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1337, "title": "Learning Multi-Index Models with Neural Networks via Mean-Field Langevin Dynamics", "abs": "We study the problem of learning multi-index models in high-dimensions using a two-layer neural network trained with the mean-field Langevin algorithm. Under mild distributional assumptions on the data, we characterize the effective dimension $d_{\\mathrm{eff}}$ that controls both sample and computational complexity by utilizing the adaptivity of neural networks to latent low-dimensional structures. When the data exhibit such a structure, $d_{\\mathrm{eff}}$ can be significantly smaller than the ambient dimension. We prove that the sample complexity grows almost linearly with $d_{\\mathrm{eff}}$, bypassing the limitations of the information and generative exponents that appeared in recent analyses of gradient-based feature learning. On the other hand, the computational complexity may inevitably grow exponentially with $d_{\\mathrm{eff}}$ in the worst-case scenario. Motivated by improving computational complexity, we take the first steps towards polynomial time convergence of the mean-field Langevin algorithm by investigating a setting where the weights are constrained to be on a compact manifold with positive Ricci curvature, such as the hypersphere. There, we study assumptions under which polynomial time convergence is achievable, whereas similar assumptions in the Euclidean setting lead to exponential time complexity.", "keywords": [ "mean-field Langevin dynamics", "feature learning", "multi-index models", "neural networks", "gradient descent" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "WH9NhxOeu9", "primary_area": "learning theory", "n_author": 1, "author_info_list": [ { "name": "Yingzhen Yang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 367, "n_formula_1": 29, "n_ref_uni": 35, "n_ref": 151, "n_ref_all": 242, "n_fig": 2, "n_tab": 3, "L_tab": 428, "n_element_tab": 46, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 853, "n_element_tab_1": 33, "formula_len_all": 32096, "formula_len_all_1": 1779, "len_all": 294356, "len_all_1": 72894, "len_abs": 2499, "len_title": 191, "len_sents": 74234, "len_sents_1": 31446, "n_sents": 834, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 117, "L_abs": 2219, "title": "Sharp Generalization for Nonparametric Regression by Over-Parameterized Neural Networks: A Distribution-Free Analysis", "abs": "Sharp generalization bound for neural networks trained by gradient descent (GD) is of central interest in statistical learning theory and deep learning. In this paper, we consider nonparametric regression\nby an over-parameterized two-layer NN trained by GD. We show that, if the neural network is trained by GD with early stopping, then the trained network renders a sharp rate of the nonparametric regression risk of $\\cO(\\eps_n^2)$, which is the same rate as that for kernel regression trained by GD with early stopping, where $\\eps_n$ is the critical population rate of the Neural Tangent Kernel (NTK) associated with the network and $n$ is the size of the training data. It is remarked that our result does not require distributional assumptions on the training data, in a strong contrast with many existing results which rely on specific distributions such as the spherical uniform data distribution or distributions satisfying certain restrictive conditions.\nAs a special case of our general result, when the eigenvalues of the associated NTK\ndecay at a rate of $\\lambda_j \\asymp j^{-\\frac{d}{d-1}}$ for $j \\ge 1$ which happens if the training data is distributed uniformly on the unit sphere in $\\RR^d$, we immediately obtain the minimax optimal rate of\n$\\cO(n^{-\\frac{d}{2d-1}})$, which is the major results of several existing works in this direction. The neural network width in our general result is lower bounded by a function of only $n,d,\\eps_n$, and such width does not depend on the minimum eigenvalue of the empirical NTK matrix whose lower bound usually requires additional assumptions on the training data.\nOur results are built upon two significant technical results which are of independent interest. First, uniform convergence to the NTK is established during the training process by GD, so that we can have a nice decomposition of the neural network function at any step of the GD into a function in the Reproducing\nKernel Hilbert Space associated with the NTK and an error function with a small $L^{\\infty}$-norm. Second, local Rademacher complexity is employed\nto tightly bound the Rademacher complexity of the function class comprising all the possible neural network functions obtained by GD.", "keywords": [ "Nonparametric Regression", "Over-Parameterized Neural Networks", "Minimax Optimal Rates" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "WGWoRZb0pT", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Matthew Purver", "gender": "Male", "institution": "Queen Mary University of London", "country": "", "position": "Professor of Computational Linguistics" }, { "name": "Zahraa Al Sahili", "gender": "unknown", "institution": "Queen Mary, University of London", "country": "", "position": "PhD student" }, { "name": "Ioannis Patras", "gender": "Male", "institution": "Queen Mary, University of London", "country": "GB", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 17, "n_ref": 33, "n_ref_all": 48, "n_fig": 5, "n_tab": 7, "L_tab": 3212, "n_element_tab": 378, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 4355, "n_element_tab_1": 390, "formula_len_all": 241, "formula_len_all_1": 224, "len_all": 141590, "len_all_1": 57012, "len_abs": 1157, "len_title": 156, "len_sents": 52002, "len_sents_1": 22393, "n_sents": 372, "n_sents_1": 151, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1159, "title": "FairCoT: Enhancing Fairness in Diffusion Models via Chain of Thought Reasoning of Multimodal Language Models", "abs": "In the domain of text-to-image generative models, biases inherent in training datasets often propagate into generated content, posing significant ethical challenges, particularly in socially sensitive contexts. We introduce FairCoT, a novel framework that enhances fairness in diffusion models through Chain-of-Thought (CoT) reasoning within multimodal generative large language models (LLMs). FairCoT employs iterative CoT refinement and attire-based attribute prediction to systematically mitigate biases, ensuring diverse and equitable representation in generated images. By integrating iterative reasoning processes, FairCoT addresses the limitations of zero-shot CoT in sensitive scenarios, balancing creativity with ethical responsibility. Experimental evaluations across multiple models, including DALL-E and various Stable Diffusion variants, demonstrate that FairCoT significantly improves fairness and diversity metrics without compromising image quality or relevance. Our approach advances ethical AI practices in generative modeling, promoting socially responsible content generation and setting new standards for fairness in AI-generated imagery.", "keywords": [ "diffusion models;fairness; bias; chain of thought; text to image; multimodal LLMs" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "WGBf2xwsgX", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "BREFO DWAMENA YAW", "gender": "Male", "institution": "Aya data", "country": "", "position": "Researcher" }, { "name": "David Rolnick", "gender": "Male", "institution": "School of Computer Science, McGill University", "country": "", "position": "Assistant Professor" }, { "name": "Jade Boutot", "gender": "unknown", "institution": "McGill University, McGill University", "country": "", "position": "PhD student" }, { "name": "Mary Kang", "gender": "Female", "institution": "McGill University", "country": "CA", "position": "Associate Professor" }, { "name": "Michelle Lin", "gender": "unknown", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "", "position": "Intern" }, { "name": "Pratinav Seth", "gender": "Male", "institution": "Arya.ai", "country": "IN", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 63, "n_ref_all": 92, "n_fig": 10, "n_tab": 12, "L_tab": 5311, "n_element_tab": 414, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1367, "n_element_tab_1": 119, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 202847, "len_all_1": 54812, "len_abs": 879, "len_title": 123, "len_sents": 68140, "len_sents_1": 26084, "n_sents": 519, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 872, "title": "Alberta Wells Dataset: Pinpointing Oil and Gas Wells from Satellite Imagery", "abs": "Millions of abandoned oil and gas wells are scattered across the world, leaching methane into the atmosphere and toxic compounds into the groundwater. Many of these locations are unknown, preventing the wells from being plugged and their polluting effects averted. Remote sensing is a relatively unexplored tool for pinpointing abandoned wells at scale. We introduce the first large-scale dataset for this problem, leveraging medium-resolution multi-spectral satellite imagery from Planet Labs. Our curated dataset comprises over 213,000 wells (abandoned, suspended, and active) from Alberta, a region with especially high well density, sourced from the Alberta Energy Regulator and verified by domain experts. We evaluate baseline algorithms for well detection and segmentation, showing the promise of computer vision approaches but also significant room for improvement.", "keywords": [ "Remote Sensing", "Satellite Imagery", "Climate Change", "AI for Good" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 4, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "WG7GzGx3G9", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Chengyuan Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jingren Zhou", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Junyang Lin", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Tong Zhang", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zengke Liu", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "jianwei zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Kerry Yi", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 23, "n_ref": 44, "n_ref_all": 69, "n_fig": 10, "n_tab": 3, "L_tab": 3407, "n_element_tab": 595, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 2496, "n_element_tab_1": 268, "formula_len_all": 886, "formula_len_all_1": 532, "len_all": 129076, "len_all_1": 51645, "len_abs": 1386, "len_title": 86, "len_sents": 31010, "len_sents_1": 20892, "n_sents": 270, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1326, "title": "Rotated Runtime Smooth: Training-Free Activation Smoother for accurate INT4 inference", "abs": "Large language models have demonstrated promising capabilities upon scaling up parameters. However, serving large language models incurs substantial computation and memory movement costs due to their large scale. Quantization methods have been employed to reduce service costs and latency. Nevertheless, outliers in activations hinder the development of INT4 weight-activation quantization. Existing approaches separate outliers and normal values into two matrices or migrate outliers from activations to weights, suffering from high latency or accuracy degradation. Based on observing activations from large language models, outliers can be classified into channel-wise and spike outliers.\nIn this work, we propose Rotated Runtime Smooth (**RRS**), a plug-and-play activation smoother for quantization, consisting of Runtime Smooth and the Rotation operation. Runtime Smooth (**RS**) is introduced to eliminate **channel-wise outliers** by smoothing activations with channel-wise maximums during runtime. The Rotation operation can narrow the gap between **spike outliers** and normal values, alleviating the effect of victims caused by channel-wise smoothing.\nThe proposed method outperforms the state-of-the-art method in the LLaMA and Qwen families and improves WikiText-2 perplexity from 57.33 to 6.66 for INT4 inference.", "keywords": [ "Large language model", "Quantization", "INT4 inference" ], "rating_list": [ 5, 8, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 2, 5, 2 ] }, { "paper_id": "WFlLqUmb9v", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Dor Tsur", "gender": "Male", "institution": "Ben-Gurion University of the Negev", "country": "IL", "position": "PhD student" }, { "name": "EYAL YAKIR", "gender": "Male", "institution": "Ben Gurion University of the Negev", "country": "IL", "position": "MS student" }, { "name": "Haim H. Permuter", "gender": "unknown", "institution": "Ben Gurion University of the Negev", "country": "IL", "position": "Full Professor" } ], "n_formula": 30, "n_formula_1": 10, "n_ref_uni": 25, "n_ref": 55, "n_ref_all": 69, "n_fig": 6, "n_tab": 13, "L_tab": 6069, "n_element_tab": 440, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 3548, "n_element_tab_1": 107, "formula_len_all": 4002, "formula_len_all_1": 729, "len_all": 152866, "len_all_1": 58847, "len_abs": 5619, "len_title": 84, "len_sents": 43164, "len_sents_1": 23446, "n_sents": 352, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1236, "title": "Efficient Time Series Forecasting via Hyper-Complex Models and Frequency Aggregation", "abs": "Time-series forecasting is a long-standing challenge in statistics and machine learning, with one of the key difficulties being the ability to process sequences with long-range dependencies. A recent line of work has addressed this by applying the short-time Fourier transform (STFT), which partitions sequences into multiple subsequences and applies a Fourier transform to each separately.\nWe propose the Frequency Information Aggregation (FIA-Net), a model that can utilize two backbone architectures: the Window-Mixing MLP (WM-MLP), which aggregates adjacent window information in the frequency domain, and the Hyper-Complex MLP (HC-MLP), which treats the set of STFT windows as hyper-complex (HC) valued vectors. and employ HC algebra to efficiently combine information from all STFT windows altogether. Furthermore, due to the nature of HC operations, the HC-MLP uses up to three times fewer parameters than the equivalent standard window aggre- gation method. We evaluate the FIA-Net on various time-series benchmarks and show that the proposed methodologies outperform existing state-of-the-art meth- ods in terms of both accuracy and efficiency. Our code is publicly available on https://anonymous.4open.science/r/research-1803/", "keywords": [ "time-series forecasting", "frequency models", "hyper-complex machine learning", "short-time Fourier transform" ], "rating_list": [ 1, 3, 1, 5 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 1, 3, 1, 3 ], "contribution_list": [ 1, 3, 1, 2 ], "confidence_list": [ 5, 3, 5, 3 ] }, { "paper_id": "WEQL5ksDnB", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Kangwook Jang", "gender": "Male", "institution": "Korea Advanced Institute of Technology", "country": "KR", "position": "PhD student" }, { "name": "Sangmin Bae", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Se-Young Yun", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "Assistant Professor" }, { "name": "Sungnyun Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sungwoo Cho", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 46, "n_ref": 136, "n_ref_all": 185, "n_fig": 10, "n_tab": 11, "L_tab": 2734, "n_element_tab": 421, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1151, "n_element_tab_1": 63, "formula_len_all": 538, "formula_len_all_1": 538, "len_all": 187150, "len_all_1": 61983, "len_abs": 1369, "len_title": 134, "len_sents": 58788, "len_sents_1": 30027, "n_sents": 363, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1374, "title": "Multi-Task Corrupted Prediction for Learning Robust Audio-Visual Speech Representation", "abs": "Audio-visual speech recognition (AVSR) incorporates auditory and visual modalities to improve recognition accuracy, particularly in noisy environments where audio-only speech systems are insufficient. While previous research has largely addressed audio disruptions, few studies have dealt with visual corruptions, e.g., lip occlusions or blurred videos, which are also detrimental. To address this real-world challenge, we propose CAV2vec, a novel self-supervised speech representation learning framework particularly designed to handle audio-visual joint corruption. CAV2vec employs a self-distillation approach with a corrupted prediction task, where the student model learns to predict clean targets, generated by the teacher model, with corrupted input frames. Specifically, we suggest a unimodal multi-task learning, which distills cross-modal knowledge and aligns the corrupted modalities, by predicting clean audio targets with corrupted videos, and clean video targets with corrupted audios. This strategy mitigates the dispersion in the representation space caused by corrupted modalities, leading to more reliable and robust audio-visual fusion. Our experiments on robust AVSR benchmarks demonstrate that the corrupted representation learning method significantly enhances recognition accuracy across generalized environments involving various types of corruption.", "keywords": [ "robust audio-visual speech recognition", "audio-visual corruption", "multimodal representation learning" ], "rating_list": [ 5, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 5, 2, 4 ] }, { "paper_id": "WDxa9hnz4p", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Chen Jason Zhang", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "Longyu Feng", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Researcher" }, { "name": "Mengze Hong", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 36, "n_ref": 46, "n_ref_all": 55, "n_fig": 4, "n_tab": 2, "L_tab": 440, "n_element_tab": 12, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 368, "formula_len_all_1": 166, "len_all": 117928, "len_all_1": 52554, "len_abs": 1423, "len_title": 144, "len_sents": 32617, "len_sents_1": 26062, "n_sents": 221, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1428, "title": "Auto-Demo Prompting: Leveraging Generated Outputs as Demonstrations for Enhanced Batch Prompting", "abs": "Batch prompting is a common technique in large language models (LLMs) used to process multiple inputs simultaneously, aiming to improve computational efficiency. However, as batch sizes increase, performance degradation often occurs due to the model's difficulty in handling lengthy context inputs. Existing methods that attempt to mitigate these issues rely solely on batch data arrangement and majority voting rather than improving the design of the batch prompt itself. In this paper, we address these limitations by proposing \"Auto-Demo Prompting,\" a novel approach that leverages the question-output pairs from earlier questions within a batch as demonstrations for subsequent answer inference. We provide a formal theoretical analysis of how Auto-Demo Prompting functions within the autoregressive generation process of LLMs, illustrating how it utilizes prior outputs to optimize the model's internal representations. Our method effectively bridges the gap between batch prompting and few-shot prompting, enhancing performance with only a slight compromise in token usage. Experimental results across five NLP tasks demonstrate its effectiveness in mitigating performance degradation and occasionally outperforming single prompts. Furthermore, it opens new avenues for applying few-shot learning techniques, such as demonstration selection, within batch prompting, making it a robust solution for real-world applications.", "keywords": [ "Batch Prompting", "In-Context Learning", "Large Language Models" ], "rating_list": [ 1, 3, 3 ], "soundness_list": [ 1, 2, 1 ], "presentation_list": [ 1, 2, 2 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "WDheQxWAo4", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Da Huang", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Quoc V Le", "gender": "Male", "institution": "Google", "country": "US", "position": "Scientist" }, { "name": "Yifeng Lu", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Researcher" }, { "name": "Dengyong Zhou", "gender": "unknown", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Jerry Weng Wei", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 102, "n_ref_all": 156, "n_fig": 15, "n_tab": 11, "L_tab": 2377, "n_element_tab": 61, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 200, "n_element_tab_1": 3, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 220238, "len_all_1": 56709, "len_abs": 0, "len_title": 104, "len_sents": 57221, "len_sents_1": 27925, "n_sents": 384, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1309, "title": "Simple synthetic data reduces sycophancy in large language models", "abs": "Sycophancy is an undesirable behavior where models tailor their responses to follow a human user's view even when that view is not objectively correct (e.g., adapting liberal views once a user reveals that they are liberal). In this paper, we study the prevalence of sycophancy in language models and propose a simple synthetic-data intervention to reduce this behavior.\n\nFirst, on a set of three sycophancy tasks where models are asked for an opinion on statements with no correct answers (e.g., politics), we observe that both model scaling and instruction tuning significantly increase sycophancy for large language models up to 540B parameters. Second, we extend sycophancy evaluations to simple addition statements that are objectively incorrect, finding that despite knowing that these statements are wrong, language models will still agree with them if the user does as well.\n\nTo reduce sycophancy, we present a straightforward synthetic-data intervention that takes public NLP tasks and encourages models to be robust to user opinions on these tasks. Adding these data in a lightweight finetuning step can significantly reduce sycophantic behavior on held-out prompts. Code for generating synthetic data for intervention can be found at https://anonymous.4open.science/r/sycophancy-intervention-F0D1/.", "keywords": [ "sycophancy", "natural language processing", "large language models" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 2, 2 ] }, { "paper_id": "WDO5hfLZvN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Cornelia Fermuller", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "Research Scientist" }, { "name": "Eadom T Dessalene", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Michael Maynord", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Rashmi Kapu", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "MS student" }, { "name": "Yiannis Aloimonos", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 26, "n_ref": 43, "n_ref_all": 72, "n_fig": 8, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 162, "n_element_tab_1": 18, "formula_len_all": 186, "formula_len_all_1": 0, "len_all": 96220, "len_all_1": 56094, "len_abs": 858, "len_title": 50, "len_sents": 35086, "len_sents_1": 27663, "n_sents": 241, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 859, "title": "Evolved LLM Schemas for Mid Vision Feedback", "abs": "In this work, we present ELF (Evolving LLM-Based Schemas for Mid-Vision Feedback), a framework that integrates schema evolution with Mid Vision Feedback (MVF) for visual learning. We leverage Large Language Models (LLMs) to automatically generate schemas: executable semantic programs operating over sets of context categories (e.g., \u201danimate\u201d or \u201dinanimate\u201d). We integrate schemas into visual processing via MVF, a method that utilizes top-down feedback connections to inform mid-level visual processing with high-level contextual knowledge. To optimize these schemas we utilize EvoPrompt, an evolutionary algorithm that refines schemas through iterative search, resulting in improvements in accuracy and contextual consistency. We demonstrate the effectiveness of ELF across multiple datasets and multiple architectures for the task of object classification", "keywords": [ "cifar", "objects", "classification", "computer-vision", "llm", "context", "feedback", "imagenet" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "WDDyTcaP1L", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Hengxiang Zhang", "gender": "unknown", "institution": "Southern University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Hongxin Wei", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Qiang Hu", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 41, "n_ref": 80, "n_ref_all": 112, "n_fig": 21, "n_tab": 4, "L_tab": 1905, "n_element_tab": 204, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 1729, "n_element_tab_1": 138, "formula_len_all": 433, "formula_len_all_1": 514, "len_all": 141877, "len_all_1": 60128, "len_abs": 1159, "len_title": 120, "len_sents": 42599, "len_sents_1": 28002, "n_sents": 300, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1163, "title": "Defending Membership Inference Attacks via Privacy-aware Sparsity Tuning", "abs": "Over-parameterized models are typically vulnerable to membership inference attacks, which aim to determine whether a specific sample is included in the training of a given model. Previous Weight regularizations (e.g., L1 regularization) typically impose uniform penalties on all parameters, leading to a suboptimal tradeoff between model utility and privacy. In this work, we first show that only a small fraction of parameters substantially impact the privacy risk. In light of this, we propose Privacy-aware Sparsity Tuning (PAST)\u2014a simple fix to the L1 Regularization\u2014by employing adaptive penalties to different parameters. Our key idea behind PAST is to promote sparsity in parameters that significantly contribute to privacy leakage. In particular, we construct the adaptive weight for each parameter based on its privacy sensitivity, i.e., the gradient of the loss gap with respect to the parameter. Using PAST, the network shrinks the loss gap between members and non-members, leading to strong resistance to privacy attacks. Extensive experiments demonstrate the superiority of PAST, achieving a state-of-the-art balance in the privacy-utility trade-off.", "keywords": [ "Membership Inference Attacks (MIA)", "Machine Learning Security", "Data Privacy", "Overparameterization" ], "rating_list": [ 5, 3, 3, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 4, 3, 4 ], "contribution_list": [ 2, 2, 1, 4 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "WCVMqRHWW5", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Alberto Bietti", "gender": "Male", "institution": "Flatiron Institute", "country": "US", "position": "Researcher" }, { "name": "Joan Bruna", "gender": "Male", "institution": "New York University", "country": "", "position": "" }, { "name": "Lei Chen", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 252, "n_formula_1": 15, "n_ref_uni": 25, "n_ref": 79, "n_ref_all": 140, "n_fig": 18, "n_tab": 3, "L_tab": 933, "n_element_tab": 269, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 32660, "formula_len_all_1": 1873, "len_all": 306605, "len_all_1": 69374, "len_abs": 2656, "len_title": 145, "len_sents": 88344, "len_sents_1": 32694, "n_sents": 923, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 956, "title": "Distributional Associations vs In-Context Reasoning: A Study of Feed-forward and Attention Layers", "abs": "Large language models have been successful at tasks involving basic forms of in-context reasoning, such as generating coherent language, as well as storing vast amounts of knowledge. At the core of the Transformer architecture behind such models are feed-forward and attention layers, which are often associated to knowledge and reasoning, respectively. In this paper, we study this distinction empirically and theoretically in a controlled synthetic setting where certain next-token predictions involve both distributional and in-context information. We find that feed-forward layers tend to learn simple distributional associations such as bigrams, while attention layers focus on in-context reasoning. Our theoretical analysis identifies the noise in the gradients as a key factor behind this discrepancy. Finally, we illustrate how similar disparities emerge in pre-trained models through ablations on the Pythia model family on simple reasoning tasks.", "keywords": [ "reasoning", "in-context learning", "associative memory", "transformers", "distribution shift" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "WCRQFlji2q", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Javier Ferrando", "gender": "Male", "institution": "Universidad Polit\u00e9cnica de Cataluna", "country": "", "position": "PhD student" }, { "name": "Neel Nanda", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Oscar Balcells Obeso", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "Undergrad student" }, { "name": "Senthooran Rajamanoharan", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 47, "n_ref": 84, "n_ref_all": 127, "n_fig": 16, "n_tab": 7, "L_tab": 12695, "n_element_tab": 225, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 196, "n_element_tab_1": 7, "formula_len_all": 2608, "formula_len_all_1": 505, "len_all": 273649, "len_all_1": 59788, "len_abs": 1343, "len_title": 128, "len_sents": 45487, "len_sents_1": 27757, "n_sents": 378, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 80, "L_abs": 1343, "title": "Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Models", "abs": "Hallucinations in large language models are a widespread problem, yet the mechanisms behind whether models will hallucinate are poorly understood, limiting our ability to solve this problem. Using sparse autoencoders as an interpretability tool, we discover that a key part of these mechanisms is entity recognition, where the model detects if an entity is one it can recall facts about. Sparse autoencoders uncover meaningful directions in the representation space, these detect whether the model recognizes an entity, e.g. detecting it doesn't know about an athlete or a movie. This shows that models can have self-knowledge: internal representations about their own capabilities. These directions are causally relevant: capable of steering the model to refuse to answer questions about known entities, or to hallucinate attributes of unknown entities when it would otherwise refuse. We demonstrate that despite the sparse autoencoders being trained on the base model, these directions have a causal effect on the chat model's refusal behavior, suggesting that chat finetuning has repurposed this existing mechanism. Furthermore, we provide an initial exploration into the mechanistic role of these directions in the model, finding that they disrupt the attention of downstream heads that typically move entity attributes to the final token.", "keywords": [ "Mechanistic Interpretability", "Hallucinations", "Language Models" ], "rating_list": [ 8, 10, 8, 10 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 3, 4, 4, 4 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "WBUVagRgsd", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Gabriel Kalweit", "gender": "unknown", "institution": "Collaborative Research Institute Intelligent Oncology", "country": "DE", "position": "Researcher" }, { "name": "Hanne Raum", "gender": "Female", "institution": "Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Jens Rahnfeld", "gender": "Male", "institution": "CS Department, University of Freiburg, Germany, Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "", "position": "MS student" }, { "name": "Maria Kalweit", "gender": "Female", "institution": "Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Mehdi Naouar", "gender": "Male", "institution": "CRIION - Collaborative Research Group", "country": "DE", "position": "Researcher" }, { "name": "Yannick Vogt", "gender": "Male", "institution": "Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Joschka B\u00f6decker", "gender": "Male", "institution": "Universit\u00e4t Freiburg", "country": "", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 27, "n_ref": 47, "n_ref_all": 63, "n_fig": 7, "n_tab": 4, "L_tab": 1645, "n_element_tab": 190, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 643, "n_element_tab_1": 104, "formula_len_all": 422, "formula_len_all_1": 348, "len_all": 115401, "len_all_1": 54128, "len_abs": 1601, "len_title": 172, "len_sents": 40602, "len_sents_1": 26041, "n_sents": 258, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 124, "L_abs": 1608, "title": "Salvage: Shapley-distribution Approximation Learning Via Attribution Guided Exploration for Explainable Image Classification", "abs": "The integration of deep learning into critical vision application areas has given rise to a necessity for techniques that can explain the rationale behind predictions. In this paper, we address this need by introducing Salvage, a novel removal-based explainability method for image classification. Our approach involves training an explainer model that learns the prediction distribution of the classifier on masked images. We first introduce the concept of Shapley-distributions, which offers a more accurate approximation of classification probability distributions than existing methods. Furthermore, we address the issue of unbalanced important and unimportant features. In such settings, naive uniform sampling of feature subsets often results in a highly unbalanced ratio of samples with high and low prediction likelihoods, which can hinder effective learning. To mitigate this, we propose an informed sampling strategy that leverages approximated feature importance scores, thereby reducing imbalance and facilitating the estimation of underrepresented features. After incorporating these two principles into our method, we conducted an extensive analysis on the ImageNette, MURA, WBC, and Pet datasets. The results show that Salvage outperforms various baseline explainability methods, including attention-, gradient-, and removal-based approaches, both qualitatively and quantitatively. Furthermore, we demonstrate that our explainer model can serve as a fully explainable classifier without a major decrease in classification performance, paving the way for fully explainable image classification.", "keywords": [ "Explainability", "XAI", "feature attribution" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "WAC8LmlKYf", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "C. L. Philip Chen", "gender": "Male", "institution": "South China University of Technology", "country": "", "position": "Full Professor" }, { "name": "Haohan Weng", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "MS student" }, { "name": "Jun Zhu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Tong Zhang", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yikai Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 36, "n_ref": 81, "n_ref_all": 104, "n_fig": 11, "n_tab": 5, "L_tab": 652, "n_element_tab": 73, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 710, "n_element_tab_1": 73, "formula_len_all": 233, "formula_len_all_1": 233, "len_all": 124097, "len_all_1": 52188, "len_abs": 1530, "len_title": 113, "len_sents": 29080, "len_sents_1": 23677, "n_sents": 236, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1481, "title": "PivotMesh: Generic 3D Mesh Generation via Pivot Vertices Guidance", "abs": "Generating compact and sharply detailed 3D meshes poses a significant challenge for current 3D generative models. Different from extracting dense meshes from neural representation, some recent works try to model the native mesh distribution (i.e., a set of triangles), which generates more compact results as humans crafted. However, due to the complexity and variety of mesh topology, these methods are typically limited to small datasets with specific categories and are hard to extend. In this paper, we introduce a generic and scalable mesh generation framework PivotMesh, which makes an initial attempt to extend the native mesh generation to large-scale datasets. We employ a transformer-based auto-encoder to encode meshes into discrete tokens and decode them from face level to vertex level hierarchically. Subsequently, to model the complex typology, we first learn to generate pivot vertices as coarse mesh representation and then generate the complete mesh tokens with the same auto-regressive Transformer. This reduces the difficulty compared with directly modeling the mesh distribution and further improves the model controllability. PivotMesh demonstrates its versatility by effectively learning from both small datasets like Shapenet, and large-scale datasets like Objaverse and Objaverse-xl. Extensive experiments indicate that PivotMesh can generate compact and sharp 3D meshes across various categories, highlighting its great potential for native mesh modeling.", "keywords": [ "mesh generation", "auto-regressive generation", "3D generation" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 5, 5 ] }, { "paper_id": "WA84oMWHaH", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Ke Fan", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Xinwei Sun", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yanwei Fu", "gender": "Male", "institution": "Fudan University,", "country": "CN", "position": "Professor" }, { "name": "Yikai Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Yizhuo Ding", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" } ], "n_formula": 60, "n_formula_1": 30, "n_ref_uni": 47, "n_ref": 89, "n_ref_all": 120, "n_fig": 2, "n_tab": 13, "L_tab": 1935, "n_element_tab": 255, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1543, "n_element_tab_1": 60, "formula_len_all": 5376, "formula_len_all_1": 1599, "len_all": 174893, "len_all_1": 56237, "len_abs": 1300, "len_title": 118, "len_sents": 45260, "len_sents_1": 24492, "n_sents": 418, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1303, "title": "Adaptive Pruning of Pretrained Transformer via Differential Inclusions", "abs": "Large transformers have demonstrated remarkable success, making it necessary to compress these models to reduce inference costs while preserving their performance. Current compression algorithms prune transformers at fixed compression ratios, requiring a unique pruning process for each ratio, which results in high computational costs. In contrast, we propose pruning of pretrained transformers at any desired ratio within a single pruning stage, based on a differential inclusion for a mask parameter. This dynamic can generate the whole regularization solution path of the mask parameter, whose support set identifies the network structure. Therefore, the solution path identifies a Transformer weight family with various sparsity levels, offering greater flexibility and customization.In this paper, weintroduce such an effective pruning method, termed SPP (Solution Path Pruning). To achieve effective pruning, we segment the transformers into paired modules, including query-key pairs, value-projection pairs, and sequential linear layers, and apply low-rank compression to these pairs, maintaining the output structure while enabling structural compression within the inner states. Extensive experiments conducted on various well-known transformer backbones have demonstrated the efficacy of SPP.", "keywords": [ "Pruning", "Sparse optimization." ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "W9yBCkfWWG", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Bangcheng Yang", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "MS student" }, { "name": "Shuo Chen", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Siyuan Qi", "gender": "unknown", "institution": "Beijing Institute for General Artificial Intelligence", "country": "", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 13, "n_ref_uni": 37, "n_ref": 66, "n_ref_all": 90, "n_fig": 3, "n_tab": 5, "L_tab": 1155, "n_element_tab": 151, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 94, "n_element_tab_1": 6, "formula_len_all": 1430, "formula_len_all_1": 783, "len_all": 159789, "len_all_1": 77212, "len_abs": 1235, "len_title": 115, "len_sents": 52722, "len_sents_1": 34487, "n_sents": 423, "n_sents_1": 278, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1246, "title": "Federated Coordination: Private and Distributed Strategy Alignment", "abs": "Coordination in multi-agent systems is critical for optimizing collective outcomes and is applicable in diverse fields such as drone swarms, emergency response, and more. Despite extensive research, the distributed coordination strategy alignment problem---where all agents follow the same strategy and execute the prescribed actions without a global coordinator---remains largely unexplored, posing challenges in scalability and privacy preservation. We introduce a new research problem termed ``federated coordination\", which seeks to achieve decentralized strategy alignment across distributed agents while maintaining the privacy of strategy choices. To address this problem, we propose a framework that employs an energy-based model. It facilitates decentralized strategy alignment by associating agent states with coordination strategies through local minimum energy values. We address privacy concerns through a simple yet effective communication protocol that protects strategy selections from eavesdropping and information leakage. Our extensive experimental results validate these contributions, demonstrating scalability and reduced computational demands. This enhances the practicality of coordination systems in multi-agent settings.", "keywords": [ "Coordination; Coordination strategy alignment; Privacy; Distributed" ], "rating_list": [ 6, 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 3, 5 ] }, { "paper_id": "W9uY7is3Ey", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Hao Liu", "gender": "Male", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "MS student" }, { "name": "Huimin Ma", "gender": "Female", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "Full Professor" }, { "name": "Kangsheng Wang", "gender": "Male", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "MS student" }, { "name": "Songde Han", "gender": "Male", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "MS student" }, { "name": "Tianyu Hu", "gender": "Male", "institution": "University of Science and Technology Beijing", "country": "", "position": "Assistant Professor" }, { "name": "Xiao Zhang", "gender": "Not Specified", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 60, "n_ref": 89, "n_ref_all": 111, "n_fig": 5, "n_tab": 8, "L_tab": 2870, "n_element_tab": 353, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1434, "n_element_tab_1": 132, "formula_len_all": 841, "formula_len_all_1": 518, "len_all": 169998, "len_all_1": 63855, "len_abs": 1232, "len_title": 160, "len_sents": 41017, "len_sents_1": 30752, "n_sents": 246, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 112, "L_abs": 1234, "title": "CreDes: Causal Reasoning Enhancement and Dual-End Searching for Solving Long-Range Reasoning Problems using LLMs", "abs": "Large language models (LLMs) have demonstrated limitations in handling combinatorial optimization problems involving long-range reasoning, partially due to causal hallucinations and huge search space. As for causal hallucinations, i.e., the inconsistency between reasoning and corresponding state transition, this paper introduces the Causal Relationship Enhancement (CRE) mechanism combining cause-effect interventions and the Individual Treatment Effect (ITE) to guarantee the solid causal rightness between each step of reasoning and state transition. As for the long causal range and huge search space limiting the performances of existing models featuring single-direction search, a Dual-End Searching (DES) approach is proposed to seek solutions by simultaneously starting from both the initial and goal states on the causal probability tree. By integrating CRE and DES (CreDes), our model has realized simultaneous multi-step reasoning, circumventing the inefficiencies from cascading multiple one-step reasoning like the Chain-of-Thought (CoT). Experiments demonstrate that CreDes significantly outperforms existing State-Of-The-Art (SOTA) solutions in long-range reasoning tasks in terms of both accuracy and time efficiency.", "keywords": [ "Causal Reasoning Enhancement", "Dual-End Searching", "Long-Range Reasoning", "LLM" ], "rating_list": [ 5, 5, 3, 5, 3 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 1, 1, 1 ], "contribution_list": [ 3, 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 3, 3, 3 ] }, { "paper_id": "W9FZEQj3vv", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Afra Amini", "gender": "Female", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Elliott Ash", "gender": "unknown", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Tim Vieira", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "US", "position": "Postdoc" }, { "name": "Ryan Cotterell", "gender": "Not Specified", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 27, "n_formula_1": 12, "n_ref_uni": 23, "n_ref": 67, "n_ref_all": 125, "n_fig": 10, "n_tab": 2, "L_tab": 514, "n_element_tab": 38, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 241, "n_element_tab_1": 29, "formula_len_all": 2891, "formula_len_all_1": 809, "len_all": 253853, "len_all_1": 66686, "len_abs": 1829, "len_title": 79, "len_sents": 54562, "len_sents_1": 29549, "n_sents": 456, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 31, "L_abs": 1499, "title": "Variational Best-of-N Alignment", "abs": "Best-of-N (BoN) is a popular and effective algorithm for aligning language models to human preferences. The algorithm works as follows: at inference time, N samples are drawn from the language model, and the sample with the highest reward, as judged by a reward model, is returned as the output. Despite its effectiveness, BoN is computationally expensive; it reduces sampling throughput by a factor of N. To make BoN more efficient at inference time, one strategy is to fine-tune the language model to mimic what BoN does during inference. To achieve this, we derive the distribution induced by the BoN algorithm. We then propose to fine-tune the language model to minimize backward KL divergence to the BoN distribution. Our approach is analogous to mean-field variational inference and, thus, we term it variational BoN (vBoN). To the extent this fine-tuning is successful and we end up with a good approximation, we have reduced the inference cost by a factor of N. Our experiments on controlled generation and summarization tasks show that BoN is the most effective alignment method, and our variational approximation to BoN achieves the closest performance to BoN and surpasses models fine-tuned using the standard KL-constrained RL objective. In the controlled generation task, vBoN appears more frequently on the Pareto frontier of reward and KL divergence compared to other alignment methods. In the summarization task, vBoN achieves high reward values across various sampling temperatures.", "keywords": [ "Alignment", "RLHF", "Best-of-N" ], "rating_list": [ 8, 3, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "W8xukd70cU", "primary_area": "reinforcement learning", "n_author": 13, "author_info_list": [ { "name": "Chenhui Liu", "gender": "Male", "institution": "GDS Services", "country": "CN", "position": "Enginner" }, { "name": "Feng Zhao", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Hanfei Geng", "gender": "Male", "institution": "Institute of AI Industry Research (AIR), Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Huiwen Zheng", "gender": "Female", "institution": "GDS", "country": "CN", "position": "Algorithm Enginner" }, { "name": "Jichao Leng", "gender": "Male", "institution": "Institute of AI Industry Research, Tsinghua University", "country": "", "position": "Researcher" }, { "name": "Peng Cheng", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Tianshun Hong", "gender": "Male", "institution": "GDS Services", "country": "CN", "position": "Engineer" }, { "name": "Xiangyu Zhu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Researcher" }, { "name": "Xianyuan Zhan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiao Hu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Yan Liang", "gender": "Female", "institution": "GDS", "country": "CN", "position": "EVP" }, { "name": "Yunxin Liu", "gender": "unknown", "institution": "Institute for AI Industry Research, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Ziteng He", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 33, "n_ref": 77, "n_ref_all": 129, "n_fig": 29, "n_tab": 2, "L_tab": 764, "n_element_tab": 86, "n_fig_1": 18, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 905, "formula_len_all_1": 761, "len_all": 159132, "len_all_1": 64827, "len_abs": 1927, "len_title": 124, "len_sents": 62102, "len_sents_1": 34160, "n_sents": 385, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1935, "title": "Data Center Cooling System Optimization Using Offline Reinforcement Learning", "abs": "The recent advances in information technology and artificial intelligence have fueled a rapid expansion of the data center (DC) industry worldwide, accompanied by an immense appetite for electricity to power the DCs. In a typical DC, around 30-40% of the energy is spent on the cooling system rather than on computer servers, posing a pressing need for developing new energy-saving optimization technologies for DC cooling systems. However, optimizing such real-world industrial systems faces numerous challenges, including but not limited to a lack of reliable simulation environments, limited historical data, and stringent safety and control robustness requirements. In this work, we present a novel physics-informed offline reinforcement learning (RL) framework for energy efficiency optimization of DC cooling systems. The proposed framework models the complex dynamical patterns and physical dependencies inside a server room using a purposely designed graph neural network architecture that is compliant with the fundamental time-reversal symmetry. Because of its well-behaved and generalizable state-action representations, the model enables sample-efficient and robust latent space offline policy learning using limited real-world operational data. Our framework has been successfully deployed and verified in a large-scale production DC for closed-loop control of its air-cooling units (ACUs). We conducted a total of 2000 hours of short and long-term experiments in the production DC environment. The results show that our method achieves 14-21% energy savings in the DC cooling system, without any violation of the safety or operational constraints. We have also conducted a comprehensive evaluation of our approach in a real-world DC testbed environment. Our results have demonstrated the significant potential of offline RL in solving a broad range of data-limited, safety-critical real-world industrial control problems.", "keywords": [ "Offline Reinforcement learning", "data center optimization", "cooling system", "energy saving" ], "rating_list": [ 6, 8, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "W8K8slZ73R", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Quoc-Vinh Lai-Dang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Seungah Son", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Taemin Kang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 33, "n_ref": 59, "n_ref_all": 63, "n_fig": 1, "n_tab": 6, "L_tab": 1404, "n_element_tab": 165, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 1958, "n_element_tab_1": 175, "formula_len_all": 627, "formula_len_all_1": 628, "len_all": 111465, "len_all_1": 62850, "len_abs": 1526, "len_title": 152, "len_sents": 34012, "len_sents_1": 29785, "n_sents": 202, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1533, "title": "Adaptive Transformer Programs: Bridging the Gap Between Performance and Interpretability in Transformers", "abs": "Balancing high performance with interpretability in increasingly powerful Transformer-based models remains a challenge. While mechanistic interpretability aims to specify neural network computations in explicit, pseudocode-like formats, existing methods often involve laborious manual analysis or struggle to fully elucidate learned internal algorithms. Recent efforts to build intrinsically interpretable models have introduced considerable expressivity and optimization challenges. This work introduces Adaptive Transformer Programs, an enhanced framework building upon RASP language and Transformer Programs to create more robust and interpretable models. The proposed method increases expressivity by redesigning two primary attention modules to improve categorical and numerical reasoning capabilities. To overcome optimization hurdles, we introduce a novel reparameterization scheme that enhances the exploration-exploitation trade-off during training. We validate our approach through extensive experiments on diverse tasks, including in-context learning, algorithmic problems (e.g., sorting and Dyck languages), and NLP benchmarks such as named entity recognition and text classification. Results demonstrate that Adaptive Transformer Programs substantially narrow the performance gap between black-box Transformers and interpretable models, enhancing transparency. This work advances the development of high-performing, transparent AI systems for critical applications, addressing crucial ethical concerns in AI development.", "keywords": [ "Mechanistic Interpretability", "Transformers", "Interpretable AI" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "W7rProQocE", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Amitai Yacobi", "gender": "unknown", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" }, { "name": "Nir Ben-Ari", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" }, { "name": "Uri Shaham", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 7, "n_ref_uni": 51, "n_ref": 92, "n_ref_all": 112, "n_fig": 8, "n_tab": 4, "L_tab": 1393, "n_element_tab": 100, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 311, "n_element_tab_1": 4, "formula_len_all": 1386, "formula_len_all_1": 500, "len_all": 147289, "len_all_1": 64011, "len_abs": 1213, "len_title": 116, "len_sents": 45220, "len_sents_1": 29707, "n_sents": 394, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1241, "title": "GrEASE: Generalizable Spectral Embedding with an Application to UMAP", "abs": "Spectral Embedding (SE) is a popular method for dimensionality reduction, applicable across diverse domains. Nevertheless, its current implementations face three prominent drawbacks which curtail its broader applicability: generalizability (i.e., out-of-sample extension), scalability, and eigenvectors separation. In this paper, we introduce $\\textit{GrEASE}$: Generalizable and Efficient Approximate Spectral Embedding, a novel deep-learning approach designed to address these limitations. GrEASE incorporates an efficient post-processing step to achieve eigenvectors separation, while ensuring both generalizability and scalability, allowing for the computation of the Laplacian\u2019s eigenvectors on unseen data. This method expands the applicability of SE to a wider range of tasks and can enhance its performance in existing applications. We empirically demonstrate GrEASE's ability to consistently approximate and generalize SE, while ensuring scalability. Additionally, we show how GrEASE can be leveraged to enhance existing methods. Specifically, we focus on UMAP, a leading visualization technique, and introduce $\\textit{NUMAP}$, a generalizable version of UMAP powered by GrEASE. Our code will be publicly available upon acceptance.", "keywords": [ "Spectral Embedding", "Eigenvectors separation", "Visualization", "Laplacian", "UMAP" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "W6yIKliMot", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Chen Shen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jieping Ye", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Junjie Liu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Liang Xie", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "Postdoc" }, { "name": "Shaotian Yan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Wenxiao Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 35, "n_ref": 55, "n_ref_all": 75, "n_fig": 5, "n_tab": 9, "L_tab": 1891, "n_element_tab": 184, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 946, "n_element_tab_1": 144, "formula_len_all": 121, "formula_len_all_1": 121, "len_all": 138595, "len_all_1": 56092, "len_abs": 1508, "len_title": 172, "len_sents": 42510, "len_sents_1": 27190, "n_sents": 348, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 122, "L_abs": 1512, "title": "Don't Take Things Out of Context: Attention Intervention for Enhancing Chain-of-Thought Reasoning in Large Language Models", "abs": "Few-shot Chain-of-Thought (CoT) significantly enhances the reasoning capabilities of large language models (LLMs), functioning as a whole to guide these models in generating reasoning steps toward final answers. However, we observe that isolated segments, words, or tokens within CoT demonstrations can unexpectedly disrupt the generation process of LLMs. The model may overly concentrate on certain local information present in the demonstration, introducing irrelevant noise into the reasoning process and potentially leading to incorrect answers. In this paper, we investigate the underlying mechanism of CoT through dynamically tracing and manipulating the inner workings of LLMs at each output step, which demonstrates that tokens exhibiting specific attention characteristics are more likely to induce the model to take things out of context; these tokens directly attend to the hidden states tied with prediction, without substantial integration of non-local information. Building upon these insights, we propose a Few-shot Attention Intervention method (FAI) that dynamically analyzes the attention patterns of demonstrations to accurately identify these tokens and subsequently make targeted adjustments to the attention weights to effectively suppress their distracting effect on LLMs. Comprehensive experiments across multiple benchmarks demonstrate consistent improvements over baseline methods, with a remarkable 5.91\\% improvement on the AQuA dataset, further highlighting the effectiveness of FAI.", "keywords": [ "chain-of-thought", "reasoning", "large language models" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "W6hzM9DMMU", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Heng Yang", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" }, { "name": "Zhiyu Zhang", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Zhou Lu", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "PhD student" } ], "n_formula": 72, "n_formula_1": 14, "n_ref_uni": 26, "n_ref": 72, "n_ref_all": 111, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 56, "n_element_tab_1": 4, "formula_len_all": 7047, "formula_len_all_1": 993, "len_all": 140853, "len_all_1": 68965, "len_abs": 2014, "len_title": 109, "len_sents": 48450, "len_sents_1": 31048, "n_sents": 413, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1941, "title": "The Benefit of Being Bayesian in Online Conformal Prediction", "abs": "Based on the framework of Conformal Prediction (CP), we study the online construction of valid confidence sets given a black-box machine learning model. By converting the target confidence levels into quantile levels, the problem can be reduced to predicting the quantiles (in hindsight) of a sequentially revealed data sequence. Two very different approaches have been studied previously:\n- *Direct approach.* Assuming the data sequence is iid or exchangeable, one could maintain the empirical distribution of the observed data as an algorithmic belief, and directly predict its quantiles. \n- *Indirect approach.* As statistical assumptions often do not hold in practice, a recent trend is to consider the adversarial setting and apply first-order online optimization to moving quantile losses (Gibbs and Candes, 2021). It requires knowing the target quantile level beforehand, and suffers from certain validity issues on the obtained confidence sets, due to the associated loss linearization.\n\nThis paper presents a novel Bayesian CP framework that combines their strengths. Without any statistical assumption, it is able to both\n- answer multiple arbitrary confidence level queries online, with provably low regret; and \n- overcome the validity issues suffered by first-order optimization baselines, due to being \"data-centric\" rather than \"iterate-centric\". \n\nFrom a technical perspective, our key idea is to regularize the algorithmic belief of the above direct approach by a Bayesian prior, which \"robustifies\" it by simulating a non-linearized *Follow the Regularized Leader* (FTRL) algorithm on the output. For statisticians, this can be regarded as an online adversarial view of Bayesian inference. Importantly, the proposed belief update backbone is shared by prediction heads targeting different confidence levels, bringing practical benefits analogous to the recently proposed concept of *U-calibration* (Kleinberg et al., 2023).", "keywords": [ "conformal prediction", "online learning", "adversarial Bayes" ], "rating_list": [ 8, 5, 3, 8, 6 ], "soundness_list": [ 4, 3, 3, 4, 3 ], "presentation_list": [ 4, 1, 3, 4, 3 ], "contribution_list": [ 3, 2, 2, 4, 2 ], "confidence_list": [ 2, 4, 4, 3, 3 ] }, { "paper_id": "W6fIyuK8Lk", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Danushka Senarathna", "gender": "Male", "institution": "Southern Illinois University-Carbondale", "country": "US", "position": "PhD student" }, { "name": "Kiriti Nagesh Gowda", "gender": "Male", "institution": "Advanced Micro Devices", "country": "", "position": "Researcher" }, { "name": "Mike Schmit", "gender": "Male", "institution": "Advanced Micro Devices", "country": "", "position": "Researcher" }, { "name": "Spyros Tragoudas", "gender": "Male", "institution": "Southern Illinois University-Carbondale", "country": "", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 30, "n_ref": 60, "n_ref_all": 85, "n_fig": 5, "n_tab": 4, "L_tab": 1947, "n_element_tab": 490, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1001, "n_element_tab_1": 245, "formula_len_all": 473, "formula_len_all_1": 277, "len_all": 126091, "len_all_1": 74385, "len_abs": 1113, "len_title": 122, "len_sents": 43497, "len_sents_1": 35385, "n_sents": 341, "n_sents_1": 273, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1118, "title": "Drift Type and Magnitude Detection in Image Classification Neural Networks", "abs": "A change in the input data stream of a machine-learning model is referred to as a data drift and may impact the model\u2019s accuracy. This paper proposes a framework to detect data drifts, identify the type of drift, and estimate the drift magnitude that occur in the input data stream of image classification neural networks due to various effects. It applies to any type of drift that occurs in images due to various factors such as noise, weather, etc. A novel statistical method is proposed for drift magnitude estimation. The method relies on the change in the prediction probability distributions of the predicted classes in the classification network caused by the data drift. The drift magnitude is estimated by applying a set of thresholds to the prediction probabilities. The drift type is identified using a classification neural network. Experimental results obtained using various datasets, drift types, and neural network architectures show that the proposed framework can accurately detect data drifts, accurately identify the drift type, and estimate the drift magnitude with a very low quantization error.", "keywords": [ "Artificial Intelligence", "Machine Learning", "Image Processing", "Classification Neural Networks", "Drift Data Detection", "Image Noise Level Estimation" ], "rating_list": [ 3, 1, 5, 1 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 2, 3, 1 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "W5yncAUfSG", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Hai Dong", "gender": "unknown", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "Associate Professor" }, { "name": "Jiahui Bai", "gender": "unknown", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "PhD student" }, { "name": "A. Kai Qin", "gender": "unknown", "institution": "Swinburne University of Technology", "country": "AU", "position": "Full Professor" } ], "n_formula": 100, "n_formula_1": 21, "n_ref_uni": 27, "n_ref": 43, "n_ref_all": 56, "n_fig": 3, "n_tab": 14, "L_tab": 2572, "n_element_tab": 155, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 3353, "n_element_tab_1": 43, "formula_len_all": 7271, "formula_len_all_1": 1585, "len_all": 196331, "len_all_1": 67602, "len_abs": 3319, "len_title": 80, "len_sents": 71860, "len_sents_1": 29786, "n_sents": 540, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1150, "title": "Efficient Personalized Federated Learning via Adaptive Weight Clustering Pruning", "abs": "This paper introduces a novel personalized federated learning approach, Adaptive Federated Weight Clustering Pruning (AdFedWCP), specifically designed to optimize communication efficiency in heterogeneous network environments. AdFedWCP innovatively combines adaptive weight clustering pruning techniques, effectively addressing data and bandwidth heterogeneity. By dynamically adjusting clustering centroids based on layer importance and client-specific data characteristics, it significantly reduces communication overhead. Experimental results show that AdFedWCP achieves a reduction in communication volume ranging from 87.54% to 87.82% in communication volume, surpassing the state-of-the-art work on reducing communication overhead in personalized federated learning. AdFedWCP also surpasses existing methods in terms of accuracy across multiple datasets, with improvements ranging from 9.13% to 21.79% over the baselines on EMNIST, CIFAR-10, and CIFAR-100. These results highlight AdFedWCP\u2019s advantages in balancing communication efficiency and model accuracy, making it an ideal choice for resource-constrained federated learning environments.", "keywords": [ "Peronalized Federated Learning", "Communication Efficiency", "Data Heterogeneity", "Bandwidth Heterogeneity" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "W5S1DEjN8x", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Ke Li", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "Assistant Professor" }, { "name": "Mehran Aghabozorgi", "gender": "Male", "institution": "Computing Science, Simon Fraser University", "country": "CA", "position": "PhD student" }, { "name": "Yanshu Zhang", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "PhD student" } ], "n_formula": 68, "n_formula_1": 19, "n_ref_uni": 65, "n_ref": 132, "n_ref_all": 147, "n_fig": 6, "n_tab": 7, "L_tab": 1417, "n_element_tab": 263, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 130, "n_element_tab_1": 16, "formula_len_all": 5000, "formula_len_all_1": 1008, "len_all": 193552, "len_all_1": 69485, "len_abs": 1335, "len_title": 140, "len_sents": 49345, "len_sents_1": 32922, "n_sents": 414, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1356, "title": "Practical $\\epsilon$-Exploring Thompson Sampling for Reinforcement Learning with Continuous Controls", "abs": "Balancing exploration and exploitation is crucial in reinforcement learning (RL). While Thompson Sampling (TS) is a sound and effective exploration strategy, its application to RL with high-dimensional continuous controls remains challenging. We propose Practical $\\epsilon$-Exploring Thompson Sampling (PETS), a practical approach that addresses these challenges. Since the posterior over the parameters of the action-value function is intractable, we leverage Langevin Monte Carlo (LMC) for sampling. We propose an approach which maintains $n$ parallel Markov chains to mitigate the issues of nai\\\"{ve} application of LMC. The next step following the posterior sampling in TS involves finding the optimal action under the sampled model of the action-value function. We explore both gradient-based and gradient-free approaches to approximate the optimal action, with extensive experiments. Furthermore, to justify the use of gradient-based optimization to approximate the optimal action, we analyze the regret for TS in the RL setting with continuous controls and show that it achieves the best-known bound previously established for the discrete setting. Our empirical results demonstrate that PETS, as an exploration strategy, can be integrated with leading RL algorithms, enhancing their performance and stability on benchmark continuous control tasks.", "keywords": [ "Reinforcement Learning", "Exploration", "Thompson Sampling" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 1, 2, 1 ], "confidence_list": [ 2, 5, 3, 5 ] }, { "paper_id": "W4yLHZGqdp", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Dylan Harper", "gender": "Male", "institution": "Dolby", "country": "AU", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 8, "n_ref_uni": 12, "n_ref": 15, "n_ref_all": 21, "n_fig": 0, "n_tab": 2, "L_tab": 207, "n_element_tab": 19, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 243, "n_element_tab_1": 33, "formula_len_all": 117, "formula_len_all_1": 333, "len_all": 90036, "len_all_1": 65567, "len_abs": 1602, "len_title": 178, "len_sents": 34845, "len_sents_1": 32201, "n_sents": 253, "n_sents_1": 238, "venue": "ICLR2025", "decision": "withdraw", "L_title": 129, "L_abs": 1611, "title": "Spatial Speaker ID: Joint Spatial and Semantic Learning for Multi-Microphone Speaker Identification on Short Far-Field Utterances", "abs": "Speaker identification is the task of identifying a person who is currently talking by analysing microphone signals.\nTypical automatic speaker identification systems use a single microphone and require complete utterances of 10-30 seconds in length to accurately identify a person from an enrollment set. We introduce the related problem of detecting which person is talking among several people in a room when the utterances are very short, e.g., a single word, or a short laugh. Since utterance lengths are too short for conventional methods, we take inspiration from the way humans solve this problems - using two ears and a joint understanding of both semantic and spatial context. To solve this problem, we propose Spatial Speaker ID, which uses banded covariance features derived from multi-microphone input along with conventional banded power to identify talkers based on both the semantic characteristics of a sound and the spatial location of a sound. The internal representation learnt in Spatial Speaker ID jointly contains both spatial and voice characteristic information and is learnt contrastively, whereby two utterances that come from the same talker in the same location are required to have similar embeddings. We learn a binary classification downstream task that determines if two sets of embeddings come from the same talker in the same location. Using this binary classifier, we compare multiple ways of presenting the microphone covariance features to the upstream models. We show the importance of spatial information for identifying talkers on short utterances with interfering noise.", "keywords": [ "Representation learning", "speaker identification", "multi-channel audio" ], "rating_list": [ 1, 1, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "W4q7cwRCwg", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Enhong Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Huijie Liu", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Mingyue Cheng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Researcher" }, { "name": "Qi Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Shulan Ruan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Zhenya Huang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 37, "n_ref": 55, "n_ref_all": 80, "n_fig": 5, "n_tab": 4, "L_tab": 2570, "n_element_tab": 170, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 195, "n_element_tab_1": 3, "formula_len_all": 672, "formula_len_all_1": 836, "len_all": 144961, "len_all_1": 58231, "len_abs": 1756, "len_title": 121, "len_sents": 36876, "len_sents_1": 27009, "n_sents": 265, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1819, "title": "Beyond Layers: A Global Message-Passing Mechanism for Heterophilic Graphs", "abs": "The effectiveness of most graph neural networks is largely attributed to the message-passing mechanism.\nDespite the significant success in homophilic graphs (i.e., similar nodes are connected by edges), message-passing mechanism in heterophilic graphs (i.e., dissimilar nodes are connected by edges) is still challenging.\nDue to the existence of low-order but dissimilar neighbor nodes in a path, messages from similar but high-order neighbor nodes are often weakened. \nIn this paper, firstly, we conduct both theoretical and empirical analysis of the layer-by-layer local nature of the message-passing mechanism.\nThen, we propose a novel GloMP-GNN for heterophilic graphs by comprehensively introducing global insights into the message-passing mechanism.1) During the message propagation phase, the global insight is introduced from the perspective of graph structure. \nWe design a structure-based global propagation strategy, where messages can be effectively propagated with the bridge of virtual edges between a global virtual node and graph nodes.\nMoreover, a global edge adaption approach is included to aggregate messages with adaptive edge weight adjustment.\n2) During the feature updating phase, the global insight is introduced with a feature-augmented compensatory updating method.\nThrough a multi-view feature updating mechanism, the node feature representation can be effectively augmented by compensating the weakened message from different views.\nFinally, we conduct extensive experimental evaluations on eight datasets, which demonstrate the superiority of our proposed GloMP-GNN. As broader impacts, GloMP-GNN consistently performs well across multiple layers and also effectively prevents the over-smoothing problem.\nCodes are available on Github with https://github.com/Anonymous-GloMP-GNN/GloMP-GNN.", "keywords": [ "Message Passing Mechanism; Graph Neural Network; Heterophilic Graph" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "W4djmqKZC6", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Hamid Krim", "gender": "Male", "institution": "North Carolina State University", "country": "", "position": "Full Professor" }, { "name": "Tanmay Asthana", "gender": "Male", "institution": "North Carolina State University", "country": "", "position": "PhD student" }, { "name": "Yufang Bao", "gender": "unknown", "institution": "Fayetteville State University", "country": "US", "position": "Full Professor" } ], "n_formula": 41, "n_formula_1": 24, "n_ref_uni": 31, "n_ref": 69, "n_ref_all": 77, "n_fig": 3, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 230, "n_element_tab_1": 3, "formula_len_all": 2404, "formula_len_all_1": 1654, "len_all": 108546, "len_all_1": 53995, "len_abs": 1266, "len_title": 98, "len_sents": 35558, "len_sents_1": 25557, "n_sents": 261, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1256, "title": "Pixel-Aware Accelerated Reverse Diffusion Modeling", "abs": "We propose in this paper an analytically new construct of a diffusion model whose drift and diffusion parameters yield a faster time-decaying Signal to Noise Ratio in the forward process. The proposed methodology significantly accelerates the forward diffusion process, reducing the required diffusion time steps from around 1000 seen in conventional models to 200-500 without compromising image quality in the reverse-time diffusion. In a departure from conventional models which typically use time-consuming multiple runs, we introduce a parallel data-driven model to generate a reverse-time diffusion trajectory in a single run of the model. The construct cleverly carries out the learning of the diffusion coefficients via an estimate of the structure of clean images. The resulting collective block-sequential generative model eliminates the need for MCMC-based sub-sampling correction for safeguarding and improving image quality, which further improve the acceleration of image generation. Collectively, these advancements yield a generative model that is at least 4 times faster than conventional approaches, while maintaining high fidelity and diversity in generated images, hence promising widespread applicability in rapid image synthesis tasks.", "keywords": [ "diffusion models", "generative models", "accelerated learning", "deep learning" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 3, 2 ], "presentation_list": [ 1, 1, 2, 1, 3 ], "contribution_list": [ 1, 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "W4AZQzNe8h", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Fanhua Shang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" }, { "name": "Hongying Liu", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuanyuan Liu", "gender": "Not Specified", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Yuxing Peng", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "", "position": "MS student" }, { "name": "Zhouchen Lin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Professor" } ], "n_formula": 148, "n_formula_1": 37, "n_ref_uni": 34, "n_ref": 86, "n_ref_all": 103, "n_fig": 2, "n_tab": 8, "L_tab": 3455, "n_element_tab": 119, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 49, "n_element_tab_1": 1, "formula_len_all": 18655, "formula_len_all_1": 5017, "len_all": 182058, "len_all_1": 52232, "len_abs": 1826, "len_title": 164, "len_sents": 42465, "len_sents_1": 19724, "n_sents": 455, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1373, "title": "Provable Faster Zeroth-order Method for Bilevel Optimization with Optimal Dependency on Error and Dimension", "abs": "In this paper, we study and analyze zeroth-order stochastic approximation algorithms for solving black-box bilevel optimization problems, where only the upper and lower function values can be obtained. \\citep{Saeed2024} proposed the first full zeroth-order bilevel method that utilizes Gaussian smoothing to estimate the first- and second-order partial derivatives of functions with two independent blocks of variables. However, this method suffers from a high dimensional dependency of $\\mathcal{O}((d_{1}+d_{2})^{4})$, where $d_{1}$ and $d_{2}$ are the dimensions of the outer and inner problems, respectively. They left an open question: can this dimension dependency be improved? To answer this question, we propose a single-loop accelerated zeroth-order bilevel algorithm, which achieves a dimension dependency of $\\mathcal{O}(d_{1}+d_{2})$ by incorporating coordinate-wise smoothing gradient estimators (coord). \n\tWe develop a new theoretical analysis for the proposed algorithm, which converges to a stationary point of $\\Phi(x)$ with a complexity of $\\mathcal{O}((d_{1}+d_{2})\\epsilon^{-3})$ in expectation settings and $\\mathcal{O}((d_{1}+d_{2})\\sqrt{n}\\epsilon^{-2})$ in finite sum settings. These complexities are both best-known with respect to dimension and error $\\epsilon$. We also provide experiment to validate the effectiveness of the proposed algorithm.", "keywords": [ "Stochastic bilevel optimization", "Hessian-free algorithms", "near-optimal complexity" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "W49UjcpGxx", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chenyang Si", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Junhao Song", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhengyao Lv", "gender": "Male", "institution": "University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Zhenyu Yang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Ziwei Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Kwan-Yee Kenneth Wong", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" } ], "n_formula": 15, "n_formula_1": 10, "n_ref_uni": 28, "n_ref": 97, "n_ref_all": 136, "n_fig": 26, "n_tab": 8, "L_tab": 3428, "n_element_tab": 375, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 2556, "n_element_tab_1": 324, "formula_len_all": 1069, "formula_len_all_1": 800, "len_all": 168423, "len_all_1": 75573, "len_abs": 1325, "len_title": 127, "len_sents": 48476, "len_sents_1": 34038, "n_sents": 337, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1371, "title": "FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality", "abs": "In this paper, we present \\textbf{\\textit{FasterCache}}, a novel training-free strategy designed to accelerate the inference of video diffusion models with high-quality generation. By analyzing existing cache-based methods, we observe that \\textit{directly reusing adjacent-step features degrades video quality due to the loss of subtle variations}. We further perform a pioneering investigation of the acceleration potential of classifier-free guidance (CFG) and reveal significant redundancy between conditional and unconditional features within the same timestep. Capitalizing on these observations, we introduce FasterCache to substantially accelerate diffusion-based video generation. Our key contributions include a dynamic feature reuse strategy that preserves both feature distinction and temporal continuity, and CFG-Cache which optimizes the reuse of conditional and unconditional outputs to further enhance inference speed without compromising video quality. We empirically evaluate FasterCache on recent video diffusion models. Experimental results show that FasterCache can significantly accelerate video generation (\\eg 1.67$\\times$ speedup on Vchitect-2.0) while keeping video quality comparable to the baseline, and consistently outperform existing methods in both inference speed and video quality. \\textit{Our code will be made public upon publication.}", "keywords": [ "Efficient Video Synthesis" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "W48CPXEpXR", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Dongfang Liu", "gender": "unknown", "institution": "Rochester Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Guangyan Sun", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "James Chenhao Liang", "gender": "Male", "institution": "U. S. Naval Research Laboratory", "country": "US", "position": "Researcher" }, { "name": "Mengnan Du", "gender": "unknown", "institution": "New Jersey Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Mingyu Jin", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Qiang Guan", "gender": "Male", "institution": "Kent State University", "country": "", "position": "Associate Professor" }, { "name": "Qifan Wang", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Principal Researcher" }, { "name": "Tong Geng", "gender": "Male", "institution": "University of Rochester", "country": "", "position": "Assistant Professor" }, { "name": "Yongfeng Zhang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Associate Professor" }, { "name": "Zhenting Wang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "junhan zhao", "gender": "Male", "institution": "Harvard Medical School", "country": "", "position": "Research Fellow" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 36, "n_ref": 49, "n_ref_all": 59, "n_fig": 2, "n_tab": 6, "L_tab": 1988, "n_element_tab": 250, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1669, "n_element_tab_1": 177, "formula_len_all": 624, "formula_len_all_1": 352, "len_all": 119685, "len_all_1": 57527, "len_abs": 2442, "len_title": 83, "len_sents": 35118, "len_sents_1": 25582, "n_sents": 263, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1059, "title": "Hallucinating LLM Could Be Creative", "abs": "Large Language Models (LLMs), such as GPT-4o, frequently produce hallucinations\u2014factually incorrect or nonsensical outputs generally regarded as undesirable. This study, however, explores the notion of \u201cgood\u201d hallucinations that may contribute to creativity and innovation. We propose metrics to assess hallucination quality, focusing on correctness, consistency, and reasoning diversity, which are evaluated using sample responses and semantic clustering. Our experiments explore different prompting techniques and hyperparameter configurations to provide comprehensive results based on these metrics. Furthermore, we investigate the distinction between process and outcome supervision, using multiple reasoning paths to enhance both creativity and accuracy. Preliminary results indicate that LLMs can generate creative hallucinations with minimal factual inaccuracies. This research provides a refined perspective on hallucinations in LLMs and suggests strategies to harness their creative potential, improving the reliability and flexibility of AI systems.", "keywords": [ "LLM", "Hallucination", "Creativity in AI" ], "rating_list": [ 3, 6, 8, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 4, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "W2dR6rypBQ", "primary_area": "optimization", "n_author": 8, "author_info_list": [ { "name": "Chen Ma", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "HaileiGong", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Qingcan Kang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Tao Zhong", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Wing Yin YU", "gender": "Male", "institution": "City University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Xiaojin Fu", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Xiongwei Han", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Forrestone", "gender": "unknown", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 31, "n_ref": 49, "n_ref_all": 66, "n_fig": 5, "n_tab": 6, "L_tab": 1320, "n_element_tab": 160, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 656, "n_element_tab_1": 103, "formula_len_all": 548, "formula_len_all_1": 420, "len_all": 152148, "len_all_1": 67519, "len_abs": 1155, "len_title": 144, "len_sents": 53928, "len_sents_1": 32181, "n_sents": 409, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1091, "title": "Decision Information Meets Large Language Models: The Future of Explainable Operations Research", "abs": "Operations Research (OR) is vital for decision-making in many industries. While recent OR methods have seen significant improvements in automation and efficiency through integrating Large Language Models (LLMs), they still struggle to produce meaningful explanations. This lack of clarity raises concerns about transparency and trustworthiness in OR applications. To address these challenges, we propose a comprehensive framework, Explainable Operations Research (EOR), emphasizing actionable and understandable explanations accompanying optimization. The core of EOR is the concept of Decision Information, which emerges from what-if analysis and focuses on evaluating the impact of complex constraints (or parameters) changes on decision-making. Specifically, we utilize bipartite graphs to quantify the changes in the OR model and adopt LLMs to improve the explanation capabilities. Additionally, we introduce the first industrial benchmark to rigorously evaluate the effectiveness of explanations and analyses in OR, establishing a new standard for transparency and clarity in the field.", "keywords": [ "Operations Research Problems; Large Language Models" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 2, 2, 3, 4 ] }, { "paper_id": "W2Wkp9MQsF", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Dong Wang", "gender": "Male", "institution": "Technische Universit\u00e4t Graz", "country": "", "position": "PhD student" }, { "name": "Haris \u0160iki\u0107", "gender": "unknown", "institution": "Technische Universit\u00e4t Graz", "country": "AT", "position": "MS student" }, { "name": "Lothar Thiele", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Olga Saukh", "gender": "Female", "institution": "Graz University of Technology", "country": "AT", "position": "Associate Professor" } ], "n_formula": 105, "n_formula_1": 19, "n_ref_uni": 67, "n_ref": 186, "n_ref_all": 217, "n_fig": 13, "n_tab": 6, "L_tab": 1533, "n_element_tab": 160, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6045, "formula_len_all_1": 1334, "len_all": 253469, "len_all_1": 62443, "len_abs": 866, "len_title": 114, "len_sents": 74359, "len_sents_1": 29677, "n_sents": 674, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 867, "title": "Forget the Data and Fine-Tuning! Just Fold the Network to Compress", "abs": "We introduce model folding, a novel data-free model compression technique that merges structurally similar neurons across layers, significantly reducing the model size without the need for fine-tuning or access to training data. Unlike existing methods, model folding preserves data statistics during compression by leveraging k-means clustering, and using novel data-free techniques to prevent variance collapse or explosion. Our theoretical framework and experiments across standard benchmarks, including ResNet18 and LLaMA-7B, demonstrate that model folding achieves comparable performance to data-driven compression techniques and outperforms recently proposed data-free methods, especially at high sparsity levels. This approach is particularly effective for compressing large-scale models, making it suitable for deployment in resource-constrained environments.", "keywords": [ "Model compression", "model folding", "model merging" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 5, 2 ] }, { "paper_id": "W1x77vRucB", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Daeun Kyung", "gender": "unknown", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Edward Choi", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "Associate Professor" }, { "name": "Eunbyeol Cho", "gender": "Not Specified", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Hyeonji Hwang", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" }, { "name": "Hyunseung Chung", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Jiho Kim", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Woosog Chay", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" }, { "name": "Yohan Jo", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 33, "n_ref": 53, "n_ref_all": 99, "n_fig": 6, "n_tab": 7, "L_tab": 6780, "n_element_tab": 531, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 782, "n_element_tab_1": 78, "formula_len_all": 0, "formula_len_all_1": 150, "len_all": 171145, "len_all_1": 64907, "len_abs": 1564, "len_title": 163, "len_sents": 46395, "len_sents_1": 30847, "n_sents": 382, "n_sents_1": 229, "venue": "ICLR2025", "decision": "withdraw", "L_title": 115, "L_abs": 1309, "title": "DialSim: A Real-Time Simulator for Evaluating Long-Term Multi-Party Dialogue Understanding of Conversational Agents", "abs": "Recent advancements in Large Language Models (LLMs) have significantly enhanced the capabilities of conversational agents, making them applicable to various fields (e.g., education). \n Despite their progress, the evaluation of the agents often overlooks the complexities of real-world conversations, such as real-time interactions, multi-party dialogues, and extended contextual dependencies. To bridge this gap, we introduce DialSim, a real-time dialogue simulator.\n In this simulator, an agent is assigned the role of a character from popular TV shows, requiring it to respond to spontaneous questions using past dialogue information and to distinguish between known and unknown information. \n Key features of DialSim include evaluating the agent\u2019s ability to respond within a reasonable time limit, handling long-term multi-party dialogues, and testing the agent's performance under randomized questioning with a diverse and high-quality question-answer dataset. We utilized this simulator to evaluate the latest conversational agents and analyze their limitations. Our experiments highlight both the strengths and weaknesses of these agents, providing valuable insights for future improvements in the field of conversational AI. DialSim is available at https://anonymous.4open.science/r/Simulator-A861.", "keywords": [ "Conversational Agents Evaluation", "Long-Term Multi-Party Dialogue Understanding", "Real-Time Evaluation" ], "rating_list": [ 6, 3, 8, 3 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "W1wlE4bPqP", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Dongxia Wu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Ph.D student" }, { "name": "Hongxia Jin", "gender": "Female", "institution": "Head of AI center", "country": "US", "position": "Vice President" }, { "name": "Yilin Shen", "gender": "Male", "institution": "Samsung Research America", "country": "US", "position": "Principal Researcher" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 70, "n_ref": 116, "n_ref_all": 125, "n_fig": 1, "n_tab": 8, "L_tab": 3049, "n_element_tab": 298, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1706, "n_element_tab_1": 135, "formula_len_all": 403, "formula_len_all_1": 501, "len_all": 160730, "len_all_1": 51487, "len_abs": 1681, "len_title": 111, "len_sents": 27608, "len_sents_1": 23752, "n_sents": 188, "n_sents_1": 142, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1691, "title": "Uncertainty-aware Fine-tuning on Time Series Foundation Model for Anomaly Detection", "abs": "Time-series anomaly detection is a crucial task in various real-world domains, geared towards identifying data observations that significantly deviate from the norm. Although time-series foundation models have shown promising results across multiple tasks, their effectiveness in anomaly detection is often inferior. This is due to their unsupervised learning paradigm being compromised by anomaly contamination in the training data. In addition, the existing approaches lack the capability to capture boundries of multiple types of normal and abnormal patterns. To overcome these challenges, we propose ULoRA-MoE, a general uncertainty-aware fine-tuning approach using resource-efficient Mixture-of-Expert (MoE) module based on LoRA. This proposed approach can enhance the fine-tuning performance across a broad spectrum of time series foundation models for anomaly detection. Each expert module of MoE can help learn different types of anomalies. Furthermore, we design the uncertainty-aware router of MoE using Gumbel-Softmax distribution for categorical sampling to capture the epistemic uncertainty. Given the estimated uncertainty, we propose a calibrated anomaly score function to mitigate the detrimental effects of anomaly contamination. We conducted extensive experiments on two general types of time series foundation models. The results demonstrate that our approach significantly improves the model performance compared to existing fine-tuning approaches. Furthermore, ULoRA-MoE shows competitive performance compared to a comprehensive set of non-learning, classical learning, and deep learning (DL) based time-series anomaly detection baselines across 8 real-world benchmarks.", "keywords": [ "Time Series Foundation Model", "Anomaly Detection", "Fine-tuning" ], "rating_list": [ 1, 3, 5, 6, 5 ], "soundness_list": [ 1, 1, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 1, 1, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3, 5 ] }, { "paper_id": "W0nydevOlG", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Adam Dziedzic", "gender": "Male", "institution": "CISPA, saarland university, saarland informatics campus", "country": "DE", "position": "Assistant Professor" }, { "name": "Franziska Boenisch", "gender": "Female", "institution": "CISPA", "country": "DE", "position": "Assistant Professor" }, { "name": "Nupur Kulkarni", "gender": "Female", "institution": "CISPA, saarland university, saarland informatics campus", "country": "DE", "position": "MS student" }, { "name": "Shahrzad Kiani", "gender": "unknown", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Stark Draper", "gender": "Male", "institution": "Toronto University", "country": "", "position": "Full Professor" } ], "n_formula": 54, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 108, "n_ref_all": 165, "n_fig": 4, "n_tab": 7, "L_tab": 1123, "n_element_tab": 171, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 474, "n_element_tab_1": 35, "formula_len_all": 5753, "formula_len_all_1": 785, "len_all": 177554, "len_all_1": 80497, "len_abs": 1526, "len_title": 125, "len_sents": 56147, "len_sents_1": 33122, "n_sents": 496, "n_sents_1": 297, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1531, "title": "Differentially Private Federated Learning with Time-Adaptive Privacy Spending", "abs": "Federated learning (FL) with differential privacy (DP) provides a framework for collaborative machine learning, enabling clients to train a shared model while adhering to strict privacy constraints. The framework allows each client to have an individual privacy guarantee, e.g., by adding different amounts of noise to each client\u2019s model updates. One underlying assumption is that all clients spend their privacy budgets uniformly over time (learning rounds). However, it has been shown in the literature that learning in early rounds typically focuses on more coarse-grained features that can be learned at lower signal-to-noise ratios while later rounds learn fine-grained features that benefit from higher signal-to-noise ratios. Building on this intuition, we propose a time-adaptive DP-FL framework that expends the privacy budget non-uniformly across both time and clients. Our framework enables each client to save privacy budget in early rounds so as to be\nable to spend more in later rounds when additional accuracy is beneficial in learning more fine-grained features. We theoretically prove utility improvements in the case that clients with stricter privacy budgets spend budgets unevenly across rounds, compared to clients with more relaxed budgets, who have sufficient budgets to distribute their spend more evenly. Our practical experiments on standard benchmark datasets support our theoretical results and show that, in practice, our algorithms improve the privacy-utility trade-offs compared to baseline schemes.", "keywords": [ "Differential Privacy", "Federated Learning", "Time Adaptive Privacy Spending", "Individualized Privacy Constraints" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "W0UioG6hs1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Jun Xu", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Liqi Xue", "gender": "Female", "institution": "Nankai University", "country": "CN", "position": "MS student" }, { "name": "Ran Gu", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Tianyi Xu", "gender": "Female", "institution": "Nankai University", "country": "CN", "position": "MS student" }, { "name": "Xiantong Zhen", "gender": "Male", "institution": "United Imaging Healthcare, Co., Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuxuan Shi", "gender": "Male", "institution": "United Imaging Healthcare Co.", "country": "CN", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 48, "n_ref": 182, "n_ref_all": 205, "n_fig": 12, "n_tab": 3, "L_tab": 1188, "n_element_tab": 36, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 982, "n_element_tab_1": 25, "formula_len_all": 273, "formula_len_all_1": 269, "len_all": 148950, "len_all_1": 60665, "len_abs": 1331, "len_title": 106, "len_sents": 41372, "len_sents_1": 29571, "n_sents": 287, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1337, "title": "Revisiting Vector-Quantization for Blind Image Restoration", "abs": "Vector-Quantization (VQ) generative models are widely used to learn a high-quality (HQ) codebook and a decoder as powerful generative priors for blind image restoration (BIR). In this paper, we revisit the key VQ process in VQ-based BIR methods, and provide three close observations on the side effects of VQ for code index prediction: 1) confining the representational capability of HQ codebook, 2) being error-prone on code index prediction, and 3) under-valuing the low-quality (LQ) feature for BIR. These observations motivate us to replace discrete VQ selection by continuous feature transformation from input LQ image to output HQ image with the HQ codebook. To this end, in this paper, we propose a new Self-in-Cross-Attention (SinCA) module to augment the HQ codebook with the LQ feature of input LQ image and perform cross-attention between LQ feature and input-augmented codebook. In this way, our SinCA extends the representational capability of the HQ codebook and effectively leverages the self-expressiveness property of input LQ image. Experiments on four typical VQ-based BIR methods demonstrate that, by replacing the VQ process with transformers using our SinCA, they achieve better quantitative and qualitative performance on blind image super-resolution and blind face restoration. The code will be publicly released.", "keywords": [ "Vector-Quantization", "Image Restoration" ], "rating_list": [ 6, 8, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "VzuPnoSKQ1", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Jihwan Oh", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Minu Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" }, { "name": "Se-Young Yun", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "Assistant Professor" }, { "name": "Sehyeok Kang", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Yongsik Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Song Chong", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 33, "n_ref": 63, "n_ref_all": 91, "n_fig": 15, "n_tab": 3, "L_tab": 966, "n_element_tab": 79, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 293, "n_element_tab_1": 31, "formula_len_all": 675, "formula_len_all_1": 588, "len_all": 136039, "len_all_1": 60931, "len_abs": 1140, "len_title": 110, "len_sents": 44585, "len_sents_1": 28435, "n_sents": 340, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1145, "title": "DPM: Dual Preferences-based Multi-Agent Reinforcement Learning", "abs": "Preference-based Reinforcement Learning (PbRL), which optimizes reward functions using preference feedback, is a promising approach for environments where handcrafted reward modeling is challenging. Especially in sparse-reward environments, feedback-based reward modeling achieves notable performance gains by transforming sparse feedback signals into dense ones.\nHowever, most PbRL research has primarily focused on single-agent environments, with limited attention to multi-agent environments.\nIn this paper, we propose Dual Preferences-based Multi-Agent Reinforcement Learning (DPM), which extends PbRL to multi-agent tasks by introducing _dual_ preferences comparing not only whole trajectories but also individual agent contributions during transitions. Furthermore, DPM replaces human preferences with those generated by LLMs to train the reward functions. Experimental results in the StarCraft Multi-Agent Challenge (SMAC) and SMACv2 environments demonstrate significant performance improvements over baselines, indicating the efficacy of DPM in optimizing individual reward functions and enhancing performances in sparse reward settings.", "keywords": [ "multi-agent reinforcement learning", "preference-based reinforcement learning", "RLAIF", "RLHF" ], "rating_list": [ 5, 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 3, 4 ], "presentation_list": [ 3, 4, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4, 2 ] }, { "paper_id": "VzdycorGTt", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Alexander Horsch", "gender": "Male", "institution": "University of Troms\u00f8", "country": "", "position": "Full Professor" }, { "name": "Dilip Prasad", "gender": "Male", "institution": "UiT The Arctic University of Norway", "country": "", "position": "Associate Professor" }, { "name": "Karaka Prasanth Naidu", "gender": "Male", "institution": "Indian Institute of Technology (Indian School of Mines) Dhanbad", "country": "IN", "position": "Undergrad student" }, { "name": "Krishna Agarwal", "gender": "Female", "institution": "UiT The Arctic University of Norway", "country": "", "position": "Associate Professor" }, { "name": "Rohit Agarwal", "gender": "Male", "institution": "University of Troms\u00f8", "country": "NO", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 3, "n_ref_uni": 59, "n_ref": 139, "n_ref_all": 215, "n_fig": 10, "n_tab": 15, "L_tab": 19856, "n_element_tab": 1335, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1822, "formula_len_all_1": 144, "len_all": 284806, "len_all_1": 69299, "len_abs": 1847, "len_title": 128, "len_sents": 82925, "len_sents_1": 32644, "n_sents": 683, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1280, "title": "packetLSTM: Dynamic LSTM Framework for Streaming Data with Varying Feature Space", "abs": "We study the online learning problem characterized by the varying input feature space of streaming data. Although LSTMs have been employed to effectively capture the temporal nature of streaming data, they cannot handle the dimension-varying streams in an online learning setting. Therefore, we propose a dynamic LSTM-based novel method, packetLSTM, to model the dimension-varying streams. The packetLSTM's dynamic framework consists of an evolving packet of LSTMs, each dedicated to processing one input feature. Each LSTM retains the local information of its corresponding feature, while a shared common memory consolidates global information. This configuration facilitates continuous learning and mitigates the issue of forgetting, even when certain features are absent for extended time periods. The idea of utilizing one LSTM per feature coupled with a dimension-invariant operator for information aggregation enhances the dynamic nature of packetLSTM. This dynamic nature is evidenced by the model's ability to activate, deactivate, and add new LSTMs as required, thus seamlessly accommodating varying input dimensions. The packetLSTM achieves state-of-the-art results on five datasets, and its underlying principle is extended to other RNN types, like GRU and vanilla RNN.", "keywords": [ "Varying Input Dimension", "Stremaing Data", "Online Learning", "Recurrent Neural Network", "Catastrophic Forgetting" ], "rating_list": [ 5, 3, 6, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 2, 3, 4, 4 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "VzZTHukfCB", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Feng Xue", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Postdoc" }, { "name": "Haotian Liu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Shilei Zeng", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Xurui Li", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Yu Zhou", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zhewei Dai", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 38, "n_ref": 109, "n_ref_all": 137, "n_fig": 22, "n_tab": 29, "L_tab": 10613, "n_element_tab": 409, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 98, "n_element_tab_1": 10, "formula_len_all": 1080, "formula_len_all_1": 444, "len_all": 200609, "len_all_1": 54513, "len_abs": 2102, "len_title": 138, "len_sents": 60473, "len_sents_1": 25580, "n_sents": 487, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 2034, "title": "SeaS: Few-shot Industrial Anomaly Image Generation with Separation and Sharing Fine-tuning", "abs": "Current segmentation methods typically require many training images and precise masks, while insufficient anomaly images hinder their application in industrial scenarios. To address such an issue, we explore producing diverse anomalies and accurate pixel-wise annotations. By observing the real production lines, we find that anomalies vary randomly in shape and appearance, whereas products hold globally consistent patterns with slight local variations. Such a characteristic inspires us to develop a Separation and Sharing Fine-tuning (SeaS) approach using only a few abnormal and some normal images.\nFirstly, we propose the Unbalanced Abnormal (UA) Text Prompt tailored to industrial anomaly generation, consisting of one product token and several anomaly tokens. Then, for anomaly images, we propose a Decoupled Anomaly Alignment (DA) loss to bind the attributes of the anomalies to different anomaly tokens. Re-blending such attributes may produce never-seen anomalies, achieving a high diversity of anomalies. For normal images, we propose a Normal-image Alignment (NA) loss to learn the products' key features that are used to synthesize products with both global consistency and local variations. The two training processes are separated but conducted on a shared U-Net. Finally, SeaS produces high-fidelity annotations for the generated anomalies by fusing discriminative features of U-Net and high-resolution VAE features. The extensive evaluations on the challenging MVTec AD and MVTec 3D AD dataset (RGB images) demonstrate the effectiveness of our approach. For anomaly image generation, on MVTec AD dataset, we achieve 1.88 on IS and 0.34 on IC-LPIPS, while on the MVTec 3D AD dataset, we obtain 1.95 on IS and 0.30 on IC-LPIPS. For the downstream task, by using our generated anomaly image-mask pairs, three common segmentation methods achieve an average 11.17\\% improvement on IoU on MVTec AD dataset, and a 15.49\\% enhancement in IoU on the MVTec 3D AD dataset. The source code will be released publicly available.", "keywords": [ "Industrial Anomaly Image Generation", "Industrial Anomaly Segmentation" ], "rating_list": [ 6, 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 4, 5 ] }, { "paper_id": "Vz5HgVwcdu", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Martin Skutella", "gender": "unknown", "institution": "TU Berlin", "country": "", "position": "Full Professor" }, { "name": "Moritz Leo Grillo", "gender": "Male", "institution": "Technische Universit\u00e4t Berlin", "country": "DE", "position": "PhD student" }, { "name": "Vincent Froese", "gender": "unknown", "institution": "Technische Universit\u00e4t Berlin", "country": "", "position": "Postdoc" } ], "n_formula": 63, "n_formula_1": 27, "n_ref_uni": 39, "n_ref": 64, "n_ref_all": 72, "n_fig": 2, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6646, "formula_len_all_1": 2408, "len_all": 169859, "len_all_1": 69741, "len_abs": 1509, "len_title": 114, "len_sents": 51460, "len_sents_1": 30064, "n_sents": 464, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1521, "title": "Complexity of Injectivity and Verification of ReLU Neural Networks", "abs": "Neural networks with ReLU activation play a key role in modern machine learning. Understanding the functions represented by ReLU networks is a major topic in current research as this enables a better interpretability of learning processes. \n\nInjectivity plays a crucial role whenever invertibility of a neural network is necessary, such as, e.g., for inverse problems or generative models. The exact computational complexity of deciding injectivity was recently posed as an open problem (Puthawala et al. [JMLR 2022]).\nWe answer this question by proving coNP-completeness. On the positive side, we show that the problem for a single ReLU layer is still tractable for small input dimension; more precisely, we present a parameterized algorithm which yields fixed-parameter tractability with\nrespect to theinput dimension.\n\nIn addition, we study the network verification problem which is of great importance since neural networks are increasingly used in safety-critical systems. We prove that network verification is coNP-hard for a general class of input domains. Our result thus highlights that the hardness of network verification is intrinsic to the ReLU networks themselves, rather than specific input domains. In this context, we also characterize surjectivity for ReLU networks with one-dimensional output which turns out to be the complement of a basic network verification task. We reveal interesting connections to computational convexity byformulating the surjectivity problem as a zonotope containment problem.", "keywords": [ "ReLU neural networks", "computational complexity", "parameterized complexity", "verification", "computational geometry" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "Vz0CWFMPUe", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Jingyan Shen", "gender": "Female", "institution": "Columbia University", "country": "US", "position": "Researcher" }, { "name": "Xiaoxue Xiong", "gender": "Female", "institution": "Columbia University", "country": "US", "position": "MS student" }, { "name": "Yizi Zhang", "gender": "Female", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Yongchan Kwon", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 36, "n_formula_1": 20, "n_ref_uni": 30, "n_ref": 74, "n_ref_all": 118, "n_fig": 9, "n_tab": 8, "L_tab": 2075, "n_element_tab": 167, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2179, "n_element_tab_1": 235, "formula_len_all": 2805, "formula_len_all_1": 1632, "len_all": 165009, "len_all_1": 68349, "len_abs": 1306, "len_title": 110, "len_sents": 59463, "len_sents_1": 29953, "n_sents": 467, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1312, "title": "TimeInf: Time Series Data Contribution via Influence Functions", "abs": "Evaluating the contribution of individual data points to a model's prediction is critical for interpreting model predictions and improving model performance. Existing data contribution methods have been applied to various data types, including tabular data, images, and text; however, their primary focus has been on i.i.d. settings. Despite the pressing need for principled approaches tailored to time series datasets, the problem of estimating data contribution in such settings remains under-explored, possibly due to challenges associated with handling inherent temporal dependencies. This paper introduces TimeInf, a model-agnostic data contribution estimation method for time-series datasets. By leveraging influence scores, TimeInf attributes model predictions to individual time points while preserving temporal structures between the time points. Our empirical results show that TimeInf effectively detects time series anomalies and outperforms existing data attribution techniques as well as state-of-the-art anomaly detection methods. Moreover, TimeInf offers interpretable attributions of data values, allowing us to distinguish diverse anomalous patterns through visualizations. We also showcase a potential application of TimeInf in identifying mislabeled anomalies in the ground truth annotations.", "keywords": [ "Time Series Data Contribution", "Time Series Anomaly Detection" ], "rating_list": [ 6, 6, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 2, 3 ] }, { "paper_id": "VyxlbbK8WV", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Katarzyna Filus", "gender": "Female", "institution": "Institute of Theoretical and Applied Informatics, Polish Academy of Sciences", "country": "PL", "position": "Assistant Professor" }, { "name": "Mateusz \u017barski", "gender": "Male", "institution": "Institute of Theoretical and Applied Informatics, Polish Academy of Sciences", "country": "PL", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 43, "n_ref": 95, "n_ref_all": 119, "n_fig": 14, "n_tab": 3, "L_tab": 624, "n_element_tab": 84, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 279, "formula_len_all_1": 251, "len_all": 200954, "len_all_1": 61111, "len_abs": 1436, "len_title": 131, "len_sents": 76239, "len_sents_1": 30548, "n_sents": 564, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1442, "title": "Uncovering Self-Emergent Similarity in Deep Vision Networks: A Systematic Framework", "abs": "Similarity is a key construct in psychology, neuroscience, linguistics and computer vision. Similarity can manifest in various forms, including visual, semantic, and contextual similarity. Among these, semantic similarity is particularly important. Not only it serves as an approximation of how humans categorize objects by capturing connections and hierarchies based on shared functionality, evolutionary traits, and contextual meaning, but also offers practical advantages in computational modeling via the lexical structures such as WordNet. Unlike human polls, WordNet-defined similarity is constant and interpretable, making it an important baseline for evaluation. As in the domain of deep vision models there is still a lack of a clear understanding about the emergence of similarity perception, we introduce Deep Similarity Inspector (DSI). It is a systematic framework to inspect and visualize how deep vision networks develop their similarity perception during training and how it aligns with semantic similarity. Our experiments show that both Convolutional Neural Networks' (CNNs) and Vision Transformers' (ViTs) develop a rich similarity perception during learning with 3 phases (initial similarity surge, refinement, stabilization), while clear differences are found in their dynamics. Both CNNs and ViTs, besides the gradual mistakes elimination, improve the quality of mistakes being made (the mistakes refinement phenomenon).", "keywords": [ "Similarity", "Deep Vision Networks", "Explainable Artificial Intelligence", "Evaluation Metrics", "Framework" ], "rating_list": [ 5, 6, 5, 8 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 1, 4 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "VyvGaQgxDl", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Bo Zhang", "gender": "Male", "institution": "Meituan Inc.", "country": "", "position": "Senior Software Engineer" }, { "name": "Lin Ma", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Principal Researcher" }, { "name": "Qingyuan Li", "gender": "Male", "institution": "meituan", "country": "", "position": "employee" }, { "name": "Ran meng", "gender": "Male", "institution": "meituan", "country": "CN", "position": "Researcher" }, { "name": "Yiduo Li", "gender": "Male", "institution": "Meituan", "country": "", "position": "Researcher" }, { "name": "Yifan Lu", "gender": "Male", "institution": "meituan", "country": "CN", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 40, "n_ref": 87, "n_ref_all": 110, "n_fig": 8, "n_tab": 10, "L_tab": 5127, "n_element_tab": 626, "n_fig_1": 8, "n_tab_1": 10, "L_tab_1": 6868, "n_element_tab_1": 746, "formula_len_all": 397, "formula_len_all_1": 397, "len_all": 131550, "len_all_1": 68053, "len_abs": 957, "len_title": 120, "len_sents": 27059, "len_sents_1": 22546, "n_sents": 222, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 942, "title": "Integer Scale: A Free Lunch for Faster Fine-grained Quantization of LLMs", "abs": "We introduce \\emph{Integer Scale}, a novel post-training quantization scheme for large language models that effectively resolves the inference bottleneck in current fine-grained quantization approaches while maintaining similar accuracies. Integer Scale is a free lunch as it requires no extra calibration or fine-tuning which will otherwise incur additional costs. It can be used plug-and-play for most fine-grained quantization methods and its integration results in at most \\textbf{1.85$\\times$} end-to-end speed boost over the original counterpart without sacrificing accuracy. Additionally, due to the orchestration of the proposed Integer Scale and fine-grained quantization, we resolved the quantization difficulty for Mixtral-8x7B and LLaMA-3 models with negligible performance degradation, and it comes with an end-to-end speed boost of \\textbf{2.13$\\times$}, and \\textbf{2.31$\\times$} compared with their FP16 versions respectively.", "keywords": [ "LLMs", "Quantization" ], "rating_list": [ 5, 1, 6, 3 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 1, 3, 1 ], "confidence_list": [ 5, 5, 3, 4 ] }, { "paper_id": "VyCaQrvxMq", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Alireza Habibi", "gender": "Male", "institution": "University of T\u00fcbingen", "country": "DE", "position": "Postdoc" }, { "name": "Saeed Ghoorchian", "gender": "Male", "institution": "SAP SE", "country": "DE", "position": "Researcher" }, { "name": "Setareh Maghsudi", "gender": "unknown", "institution": "Ruhr-Universt\u00e4t Bochum", "country": "", "position": "Full Professor" } ], "n_formula": 70, "n_formula_1": 35, "n_ref_uni": 27, "n_ref": 29, "n_ref_all": 67, "n_fig": 8, "n_tab": 3, "L_tab": 1707, "n_element_tab": 3, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 142, "n_element_tab_1": 9, "formula_len_all": 4408, "formula_len_all_1": 1470, "len_all": 159971, "len_all_1": 72433, "len_abs": 1599, "len_title": 128, "len_sents": 46837, "len_sents_1": 29902, "n_sents": 410, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1607, "title": "Quantum-Inspired Reinforcement Learning in the Presence of Epistemic Ambivalence", "abs": "The complexity of online decision-making under uncertainty stems from the requirement of finding a balance between exploiting known strategies and exploring new possibilities. Naturally, the uncertainty type plays a crucial role in developing decision-making strategies that manage complexity effectively. In this paper, we focus on a specific form of uncertainty known as epistemic ambivalence (EA), which emerges from conflicting pieces of evidence or contradictory experiences. It creates a delicate interplay between uncertainty and confidence, distinguishing it from epistemic uncertainty that typically diminishes with new information. Indeed, ambivalence can persist even after additional knowledge is acquired. To address this phenomenon, we propose a novel framework, called the epistemically ambivalent Markov decision process (EA-MDP), aiming to understand and control EA in decision-making processes. This framework incorporates the concept of a quantum state from the quantum mechanics formalism, and its core is to assess the probability and reward of every possible outcome. We calculate the reward function using quantum measurement techniques and prove the existence of an optimal policy and an optimal value function in the EA-MDP framework. We also propose the EA-epsilon-greedy Q-learning algorithm. To evaluate the impact of EA on decision-making and the expedience of our framework, we study two distinct experimental setups, namely the two-state problem and the lattice problem. Our results show that using our methods, the agent converges to the optimal policy in the presence of EA.", "keywords": [ "Reinforcement learning", "decision-making", "uncertainty", "epistemic ambivalence" ], "rating_list": [ 6, 3, 8, 6 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "VxvnV6slP0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Chaoxiang Cai", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Di ZHANG", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "VP" }, { "name": "Dong Shen", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Fan Yang", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Tingting Gao", "gender": "Female", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Xi Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yang LongRong", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 16, "n_ref_uni": 65, "n_ref": 193, "n_ref_all": 237, "n_fig": 11, "n_tab": 12, "L_tab": 4126, "n_element_tab": 425, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3156, "n_element_tab_1": 208, "formula_len_all": 1711, "formula_len_all_1": 964, "len_all": 265751, "len_all_1": 68545, "len_abs": 1233, "len_title": 133, "len_sents": 62951, "len_sents_1": 28035, "n_sents": 569, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1241, "title": "Solving Token Gradient Conflict in Mixture-of-Experts for Large Vision-Language Model", "abs": "The Mixture-of-Experts (MoE) has gained increasing attention in studying Large Vision-Language Models (LVLMs). It uses a sparse model to replace the dense model, achieving comparable performance while activating fewer parameters during inference, thus significantly reducing the inference cost. Existing MoE methods in LVLM encourage different experts to specialize in different tokens, and they usually employ a router to predict the routing of each token. However, the router is not optimized concerning distinct parameter optimization directions generated from tokens within an expert. This may lead to severe interference between tokens within an expert. To address this problem, we propose to use the token-level gradient analysis to Solving Token Gradient Conflict (STGC) in this paper. Specifically, we first use token-level gradients to identify conflicting tokens in experts. After that, we add a regularization loss tailored to encourage conflicting tokens routing from their current experts to other experts, for reducing interference between tokens within an expert. Our method can serve as a plug-in for diverse LVLM methods, and extensive experimental results demonstrate its effectiveness. The code will be publicly available.", "keywords": [ "Large Vision-Language Model (LVLM)", "Mixture-of-Expert (MoE)", "token-level gradient", "conflicting token" ], "rating_list": [ 6, 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "VxJjw52dZu", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Arshia Afzal", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Elias Abad Rocamora", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Francesco Tonin", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Postdoc" }, { "name": "Leyla Naz Candogan", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Mahsa Shoaran", "gender": "Female", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Assistant Professor" }, { "name": "Pol Puigdemont", "gender": "Male", "institution": "EPFL", "country": "CH", "position": "PhD student" }, { "name": "Volkan Cevher", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" }, { "name": "Yongtao Wu", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" } ], "n_formula": 135, "n_formula_1": 38, "n_ref_uni": 50, "n_ref": 132, "n_ref_all": 197, "n_fig": 7, "n_tab": 14, "L_tab": 4945, "n_element_tab": 399, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1907, "n_element_tab_1": 275, "formula_len_all": 11319, "formula_len_all_1": 5072, "len_all": 246157, "len_all_1": 64032, "len_abs": 1117, "len_title": 133, "len_sents": 64624, "len_sents_1": 23826, "n_sents": 534, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1116, "title": "LION: A bidirectional framework that trains like a Transformer and infers like an RNN", "abs": "We introduce LION, a novel sequence-to-sequence framework that unifies the bidirectionality and parallelized training of Transformers with the fast inference of recurrent neural networks. LION is built upon a mathematical formulation where full kernelized attention with a learnable mask is efficiently computed using a bidirectional selective recurrent model, matching the effectiveness of softmax-based attention with constant-time inference. Our framework naturally accounts for spatial and temporal relationships within input sequences, reducing reliance on heuristic positional embeddings and facilitating straightforward scalability in context length and resolution. Using our framework and inspired by the recent state-space models, we propose three main running examples LIOn-LIT, LION-RETNET, and LION-S, a transformer with selective mask and recurrent inference. Numerical evaluations on tasks such as language modeling, the Long-Range Arena, and image classification show that LION framework achieves performance on par with state-of-the-art models while delivering fast training and inference efficiency.", "keywords": [ "Self-attention", "transformer", "bidirectionality", "mamba", "ssm", "rnn", "inference" ], "rating_list": [ 8, 3, 5, 3 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 4, 2, 3, 4 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "VxIetsMu3G", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Han Zhang", "gender": "unknown", "institution": "University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Yuan Cao", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "Assistant Professor" } ], "n_formula": 288, "n_formula_1": 23, "n_ref_uni": 20, "n_ref": 66, "n_ref_all": 76, "n_fig": 4, "n_tab": 5, "L_tab": 1129, "n_element_tab": 64, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 43771, "formula_len_all_1": 2238, "len_all": 307971, "len_all_1": 67889, "len_abs": 1089, "len_title": 140, "len_sents": 84338, "len_sents_1": 29215, "n_sents": 998, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1095, "title": "Understanding the Benefits of SimCLR Pre-Training in Two-Layer Convolutional Neural Networks", "abs": "SimCLR is one of the most popular contrastive learning methods for vision tasks. It pre-trains deep neural networks based on a large amount of unlabeled data by teaching the model to distinguish between positive and negative pairs of augmented images. It is believed that SimCLR can pre-train a deep neural network to learn efficient representations that can lead to a better performance of future supervised fine-tuning. Despite its effectiveness, our theoretical understanding of the underlying mechanisms of SimCLR is still limited. In this paper, we theoretically introduce a case study of the SimCLR method. Specifically, we consider training a two-layer convolutional neural network (CNN) to learn a toy image data model. We show that, under certain conditions on the number of labeled data, SimCLR pre-training combined with supervised fine-tuning achieves almost optimal test loss. Notably, the label complexity for SimCLR pre-training is far less demanding compared to direct training on supervised data. Our analysis sheds light on the benefits of SimCLR in learning with fewer labels.", "keywords": [ "SimCLR", "convolutional neural network" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 2, 3, 2, 4 ] }, { "paper_id": "Vx3o2tUErQ", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 8, "author_info_list": [ { "name": "Ao Zou", "gender": "Male", "institution": "Beijing University of Chemical Technology", "country": "", "position": "Undergrad student" }, { "name": "Jiajie Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Jinxin Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Juanzi Li", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Lei Hou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Liangming Pan", "gender": "Male", "institution": "University of Arizona", "country": "US", "position": "Assistant Professor" }, { "name": "Shulin Cao", "gender": "Female", "institution": "Zhipu AI", "country": "CN", "position": "Researcher" }, { "name": "Xiaoyin Che", "gender": "Male", "institution": "Siemens Corporate Research", "country": "", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 17, "n_ref": 51, "n_ref_all": 69, "n_fig": 3, "n_tab": 3, "L_tab": 236, "n_element_tab": 31, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 486, "n_element_tab_1": 55, "formula_len_all": 860, "formula_len_all_1": 860, "len_all": 138267, "len_all_1": 63406, "len_abs": 1374, "len_title": 159, "len_sents": 30986, "len_sents_1": 29164, "n_sents": 238, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1381, "title": "REPANA: Reasoning Path Navigated Program Induction for Universally Reasoning over Heterogeneous Knowledge Bases", "abs": "Program induction is a typical approach that helps Large Language Models (LLMs) in complex knowledge-intensive question answering over knowledge bases (KBs) to alleviate the hallucination of LLMs. However, the accurate program induction usually requires a large number of high-quality parallel data of a specific KB, which is difficult to acquire for many low-resource KBs. Additionally, due to heterogeneity of questions and KB schemas, the transferability of a model trained on a single dataset is poor. To this end, we propose REPANA, a reasoning path navigated program induction framework that enables LLMs to reason over heterogeneous KBs. We decouple the program generation capability into perceiving the KB and mapping questions to program sketches. Accordingly, our framework consists of two main components. The first is an LLM-based navigator, which retrieves reasoning paths of the input question from the given KB. The second is a KB-agnostic parser trained on data from multiple heterogeneous datasets, taking the navigator's retrieved paths and the question as input and generating the corresponding program. Experiments show that REPANA exhibits strong generalization and transferability. It can directly perform inference on datasets not seen during training, outperforming other SoTA low-resource methods and even approaching the performance of supervised methods.", "keywords": [ "Knowledge Base QA; Low Resource Reasoning; Multi-hop QA; Reasoning Interpretability;" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "VwyKSnMmrr", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Hang Chen", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Wenya Wang", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Xinyu Yang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Jiaying Zhu", "gender": "Female", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 12, "n_ref_uni": 19, "n_ref": 60, "n_ref_all": 129, "n_fig": 14, "n_tab": 12, "L_tab": 2487, "n_element_tab": 200, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1092, "n_element_tab_1": 38, "formula_len_all": 2674, "formula_len_all_1": 1803, "len_all": 197001, "len_all_1": 65564, "len_abs": 1594, "len_title": 98, "len_sents": 82957, "len_sents_1": 29103, "n_sents": 459, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 40, "L_abs": 1522, "title": "Unveiling Language Skills under Circuits", "abs": "Circuit decomposition and counterfactual-based pruning have become the cornerstone framework for mechanism interpretability. However, the unfaithfulness to the output due to cumulative bias in the pruning process hinders more complex and detailed mechanism exploration. To address this, we propose a novel circuit discovery framework that faithfully identifies circuit graphs. This framework contains three steps: firstly, the language model is decomposed into a fully linear graph consisting of disentangled ``memory circuits\"; secondly, greedy search is adopted to prune while ensuring output faithfulness; finally, we adopt causal analysis on the pruned circuit graph to identify salient circuit graph, estimated by counterfactuals and interventions. Our framework facilitates the discovery of complete circuit graphs and dissection of more complex mechanisms. To demonstrate this, we explored three generic language skills (Previous Token Skill, Induction Skill and In-Context Learning Skill). Using the circuit graphs discovered through our framework, we identify the complete skill paths of these skills.\nOur experiments on various datasets confirm the correspondence between our identified skill paths and language skills, and validate three longstanding hypotheses: 1) Language skills are identifiable through circuit dissection; 2) Simple language skills reside in shallow layers, whereas complex language skills are found in deeper layers; 3) Complex language skills are formed on top of simpler language skills.", "keywords": [ "Interpretability of Language Models" ], "rating_list": [ 3, 6, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "VwOYxPScxB", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Mingjie Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Shikun Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Wei Ye", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaoling Zhou", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhemg Lee", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 53, "n_ref": 114, "n_ref_all": 153, "n_fig": 7, "n_tab": 11, "L_tab": 6194, "n_element_tab": 656, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1947, "n_element_tab_1": 311, "formula_len_all": 4511, "formula_len_all_1": 918, "len_all": 199076, "len_all_1": 70419, "len_abs": 1753, "len_title": 121, "len_sents": 51882, "len_sents_1": 31805, "n_sents": 336, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1473, "title": "HaDeMiF: Hallucination Detection and Mitigation in Large Language Models", "abs": "The phenomenon of knowledge hallucinations has raised substantial concerns about the security and reliability of deployed large language models (LLMs). Current methods for detecting hallucinations primarily depend on manually designed individual metrics, such as prediction uncertainty and consistency, and fall short in effectively calibrating model predictions, thus constraining their detection accuracy and applicability in practical applications. In response, we propose an advanced framework, termed HaDeMiF, for detecting and mitigating hallucinations in LLMs. Specifically, hallucinations within the output and semantic spaces of LLMs are comprehensively captured through two compact networks\u2014a novel, interpretable tree model known as the Deep Dynamic Decision Tree (D3T) and a Multilayer Perceptron (MLP)\u2014which take as input a set of prediction characteristics and the hidden states of tokens, respectively. The predictions of LLMs are subsequently calibrated using the outputs from the D3T and MLP networks, aiming to mitigate hallucinations and enhance model calibration. HaDeMiF can be applied during both the inference and fine-tuning phases of LLMs, introducing less than 2% of the parameters relative to the LLMs through the training of two small-scale networks. Extensive experiments conclusively demonstrate the effectiveness of our framework in hallucination detection and model calibration across text generation tasks with responses of varying lengths.", "keywords": [ "Large language model", "knowledge hallucination", "hallucination detection", "model calibration", "deep neural decision tree" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 5, 4 ] }, { "paper_id": "VvxuD3cdJx", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Ci Li", "gender": "Female", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "PhD student" }, { "name": "John Folkesson", "gender": "unknown", "institution": "KTH Royal Institute of Technology, Stockholm, Sweden", "country": "", "position": "Associate Professor" }, { "name": "Kei IKEMURA", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "MS student" }, { "name": "Nazre Batool", "gender": "Female", "institution": "National University of Ireland, Galway", "country": "IE", "position": "Assistant Professor" }, { "name": "Sina Sharif Mansouri", "gender": "unknown", "institution": "Scania", "country": "SE", "position": "Researcher" }, { "name": "Yi Yang", "gender": "Female", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "PhD student" }, { "name": "Zhang Qingwen", "gender": "Female", "institution": "KTH Royal Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 6, "n_ref_uni": 37, "n_ref": 119, "n_ref_all": 151, "n_fig": 15, "n_tab": 6, "L_tab": 1179, "n_element_tab": 143, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 1691, "n_element_tab_1": 304, "formula_len_all": 1441, "formula_len_all_1": 415, "len_all": 140472, "len_all_1": 65199, "len_abs": 1192, "len_title": 70, "len_sents": 41783, "len_sents_1": 27234, "n_sents": 309, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1195, "title": "AutoScale: Combining Multi-Task Optimization with Linear Scalarization", "abs": "Multi-task learning is favored due to its efficiency and potential transfer learning achieved by sharing networks across tasks. While a series of multi-task optimization algorithms (MTOs) have been proposed to solve MTL optimization challenges and enhance performance, recent research claims that simple linear scalarization, which sums per-task loss with a carefully searched weight set, is sufficient, casting doubt on the added value of more complex MTO algorithms. In this paper, we provide a novel perspective that linear scalarization and MTOs are closely related and can be combined to yield high performance and efficiency. We show, for the first time, that a well-performing linear scalarization exhibits specific characteristics of certain optimization metrics proposed by MTOs, such as high task gradient magnitude similarity and low condition number, via an extensive empirical study. We then propose AutoScale, an efficient pipeline that leverages these influential metrics to guide the search for optimal linear scalarization weights. AutoScale shows superior performance than prior MTOs and performs close to the searched weight performance consistently across different datasets.", "keywords": [ "Multi-task Learning", "Autonomous Driving" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 2, 3, 3 ] }, { "paper_id": "VvHVLVUD6m", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Farzan Farnia", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Kexin Fu", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Xiwei Cheng", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "Master of Philosophy student" } ], "n_formula": 174, "n_formula_1": 21, "n_ref_uni": 17, "n_ref": 49, "n_ref_all": 62, "n_fig": 7, "n_tab": 5, "L_tab": 1030, "n_element_tab": 14, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 22146, "formula_len_all_1": 2319, "len_all": 204428, "len_all_1": 61544, "len_abs": 2230, "len_title": 99, "len_sents": 53553, "len_sents_1": 27379, "n_sents": 547, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1356, "title": "On the Mode-Seeking Properties of Langevin Dynamics", "abs": "The Langevin Dynamics framework, which aims to generate samples from the score function of a probability distribution, is widely used for analyzing and interpreting score-based generative modeling. While the convergence behavior of Langevin Dynamics under unimodal distributions has been extensively studied in the literature, in practice the data distribution could consist of multiple distinct modes. In this work, we investigate Langevin Dynamics in producing samples from multimodal distributions and theoretically study its mode-seeking properties. We prove that under a variety of sub-Gaussian mixtures, Langevin Dynamics is unlikely to find all mixture components within a sub-exponential number of steps in the data dimension. To reduce the mode-seeking tendencies of Langevin Dynamics, we propose Chained Langevin Dynamics, which divides the data vector into patches of constant size and generates every patch sequentially conditioned on the previous patches. We perform a theoretical analysis of Chained Langevin Dynamics by reducing it to sampling from a constant-dimensional distribution. We present the results of several numerical experiments on synthetic and real image datasets, supporting our theoretical results on the iteration complexities of sample generation from mixture distributions using the chained and vanilla Langevin Dynamics.", "keywords": [ "Langevin dynamics", "convergence analysis", "mixture distribution", "mode-seeking" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "VvDEuyVXkG", "primary_area": "generative models", "n_author": 11, "author_info_list": [ { "name": "Fei Huang", "gender": "Male", "institution": "Alibaba Group US", "country": "US", "position": "Senior Research Director" }, { "name": "Hai-Tao Zheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Hui Wang", "gender": "Male", "institution": "Cloud Computing", "country": "CN", "position": "Full Professor" }, { "name": "Jingren Zhou", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Philip S. Yu", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "US", "position": "Full Professor" }, { "name": "Xinran Zheng", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Xinyu Wang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yangning Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yinghui Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yong Jiang", "gender": "Male", "institution": "Tongyi Lab", "country": "CN", "position": "Researcher" }, { "name": "zhen zhang", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 66, "n_ref_all": 102, "n_fig": 7, "n_tab": 17, "L_tab": 9601, "n_element_tab": 1278, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 3567, "n_element_tab_1": 381, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 218156, "len_all_1": 69287, "len_abs": 2399, "len_title": 161, "len_sents": 61427, "len_sents_1": 28155, "n_sents": 461, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 112, "L_abs": 1511, "title": "Benchmarking Multimodal Retrieval Augmented Generation with Dynamic VQA Dataset and Self-adaptive Planning Agent", "abs": "Multimodal Retrieval Augmented Generation (mRAG) plays an important role in mitigating the \u201challucination\u201d issue inherent in multimodal large language models (MLLMs). Although promising, existing heuristic mRAGs typically predefined fixed retrieval processes, which causes two issues: (1) Non-adaptive Retrieval Queries. (2) Overloaded Retrieval Queries. However, these flaws cannot be adequately reflected by current knowledge-seeking visual question answering (VQA) datasets, since the most required knowledge can be readily obtained with a standard two-step retrieval. To bridge the dataset gap, we first construct Dyn-VQA dataset, consisting of three types of ``dynamic'' questions, which require complex knowledge retrieval strategies variable in query, tool, and time: (1) Questions with rapidly changing answers. (2) Questions requiring multi-modal knowledge. (3) Multi-hop questions. Experiments on Dyn-VQA reveal that existing heuristic mRAGs struggle to provide sufficient and precisely relevant knowledge for dynamic questions due to their rigid retrieval processes. Hence, we further propose the first self-adaptive planning agent for multimodal retrieval, **OmniSearch**. The underlying idea is to emulate the human behavior in question solution which dynamically decomposes complex multimodal questions into sub-question chains with retrieval action. Extensive experiments prove the effectiveness of our OmniSearch, also provide direction for advancing mRAG. Code and dataset will be open-sourced.", "keywords": [ "Large Language Model", "Multimodal Retrieval Augmented Generation", "Knowledge Enhancement" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 2, 2, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "Vv76fCYffN", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Dixiao Cui", "gender": "unknown", "institution": "Zhijia Technology", "country": "CN", "position": "Principal Researcher" }, { "name": "Peidong Li", "gender": "Male", "institution": "Zhijia Technology", "country": "CN", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 27, "n_ref": 73, "n_ref_all": 101, "n_fig": 17, "n_tab": 8, "L_tab": 3831, "n_element_tab": 322, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 3545, "n_element_tab_1": 249, "formula_len_all": 620, "formula_len_all_1": 520, "len_all": 130518, "len_all_1": 64419, "len_abs": 1435, "len_title": 128, "len_sents": 35596, "len_sents_1": 26154, "n_sents": 277, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1238, "title": "Navigation-Guided Sparse Scene Representation for End-to-End Autonomous Driving", "abs": "End-to-End Autonomous Driving (E2EAD) methods typically rely on supervised perception tasks to extract explicit scene information (e.g., objects, maps). This reliance necessitates expensive annotations and constrains deployment and data scalability in real-time applications. In this paper, we introduce SSR, a novel framework that utilizes only 16 navigation-guided tokens as Sparse Scene Representation, efficiently extracting crucial scene information for E2EAD. Our method eliminates the need for human-designed supervised sub-tasks, allowing computational resources to concentrate on essential elements directly related to navigation intent. We further introduce a temporal enhancement module, aligning predicted future scenes with actual future scenes through self-supervision. SSR achieves a 27.2\\% relative reduction in L2 error and a 51.6\\% decrease in collision rate to UniAD in nuScenes, with a 10.9\u00d7 faster inference speed and 13\u00d7 faster training time. Moreover, SSR outperforms VAD-Base with a 48.6-point improvement on driving score in CARLA\u2019s Town05 Long benchmark. This framework represents a significant leap in real-time autonomous driving systems and paves the way for future scalable deployment. Code will be released.", "keywords": [ "End-to-End", "Autonomous Driving", "Sparse Scene Representation" ], "rating_list": [ 6, 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 4, 3 ] }, { "paper_id": "Vuj1FZfghv", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Chanyoung Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Junseok Lee", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Sein Kim", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Sukwon Yun", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Tsuyoshi Murata", "gender": "Male", "institution": "Tokyo Institute of Technology", "country": "JP", "position": "Full Professor" }, { "name": "Xin Liu", "gender": "Male", "institution": "AIST, National Institute of Advanced Industrial Science and Technology", "country": "JP", "position": "Researcher" }, { "name": "Yunhak Oh", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Dongmin Hyun", "gender": "unknown", "institution": "Yahoo Research", "country": "", "position": "Research Scientist" } ], "n_formula": 21, "n_formula_1": 5, "n_ref_uni": 35, "n_ref": 67, "n_ref_all": 105, "n_fig": 9, "n_tab": 13, "L_tab": 5458, "n_element_tab": 759, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2653, "n_element_tab_1": 367, "formula_len_all": 1911, "formula_len_all_1": 317, "len_all": 187833, "len_all_1": 67108, "len_abs": 1641, "len_title": 139, "len_sents": 58159, "len_sents_1": 28369, "n_sents": 423, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1383, "title": "Toward Generalizability of Graph-based Imputation on Biomedical Tabular-based Missing Data", "abs": "Recent advances in graph-based imputation methods for addressing missing data have received considerable attention, primarily for their ability to effectively aggregate and propagate information through graph structures. However, the applicability of these methods to the tabular domain remains constrained by two main factors: the lack of task-relevant graph structure and a lack of consideration of feature-wise relationships. To address these challenges, we introduce GRASS, a novel approach that effectively bridges the gap between existing graph-based imputation methods and the unique needs of tabular domains with initially missing data. To derive feature gradient, GRASS initiates with training a Multi-Layer Perceptron layer on tabular data. This gradient then facilitates the creation of graph structures from a feature (column) perspective, enabling column-wise feature propagation for imputing missing values, followed by uncertainty-aware categorical clamping. Finally, to effectively utilize existing graph-based imputation methods in an agnostic manner, we input a so-called warmed-up matrix along with an associated sample (row) graph. We validate GRASS on real-world tabular datasets, including those from the bio, medical, and social domains, demonstrating its ability to unlock the potential of graph-based imputation methods across various missing data scenarios.", "keywords": [ "Tabular missing data", "Graph-based Imputation" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "VuTrZzrPfn", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Bang Liu", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Assistant Professor" }, { "name": "Xiaoqiang Wang", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 44, "n_ref": 119, "n_ref_all": 148, "n_fig": 11, "n_tab": 10, "L_tab": 2759, "n_element_tab": 190, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1432, "n_element_tab_1": 110, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 198769, "len_all_1": 68125, "len_abs": 1358, "len_title": 127, "len_sents": 45574, "len_sents_1": 32197, "n_sents": 330, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1250, "title": "OSCAR: Operating System Control via State-Aware Reasoning and Re-Planning", "abs": "Large language models (LLMs) and large multimodal models (LMMs) have shown great potential in automating complex tasks like web browsing and gaming. However, their ability to generalize across diverse applications remains limited, hindering broader utility. To address this challenge, we present OSCAR: Operating System Control via state-Aware reasoning and Re-planning. OSCAR is a generalist agent designed to autonomously navigate and interact with various desktop and mobile applications through standardized controls, such as mouse and keyboard inputs, while processing screen images to fulfill user commands.\nOSCAR translates human instructions into executable Python code, enabling precise control over graphical user interfaces (GUIs). To enhance stability and adaptability, OSCAR operates as a state machine, equipped with error-handling mechanisms and dynamic task re-planning, allowing it to efficiently adjust to real-time feedback and exceptions. We demonstrate OSCAR\u2019s effectiveness through extensive experiments on diverse benchmarks across desktop and mobile platforms, where it transforms complex workflows into simple natural language commands, significantly boosting user productivity. Our code will be open-source upon publication.", "keywords": [ "Large Language Model", "Autonomous Agent", "Graphical User Interface" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "VtYfbvwpWp", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Huan Liu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CA", "position": "Researcher" }, { "name": "Jun Chen", "gender": "Male", "institution": "McMaster University", "country": "CA", "position": "Full Professor" }, { "name": "xiangyu xu", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Zewei Zhang", "gender": "Male", "institution": "McMaster University", "country": "CA", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 13, "n_ref_uni": 43, "n_ref": 118, "n_ref_all": 173, "n_fig": 20, "n_tab": 5, "L_tab": 1126, "n_element_tab": 102, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1038, "formula_len_all_1": 678, "len_all": 178021, "len_all_1": 67987, "len_abs": 3090, "len_title": 71, "len_sents": 52370, "len_sents_1": 29507, "n_sents": 399, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1001, "title": "GoodDrag: Towards Good Practices for Drag Editing with Diffusion Models", "abs": "In this paper, we introduce GoodDrag, a novel approach to improve the stability and image quality of drag editing. Unlike existing methods that struggle with accumulated perturbations and often result in distortions, GoodDrag introduces an AlDD framework that alternates between drag and denoising operations within the diffusion process, effectively improving the fidelity of the result. We also propose an information-preserving motion supervision operation that maintains the original features of the starting point for precise manipulation and artifact reduction. In addition, we contribute to the benchmarking of drag editing by introducing a new dataset, Drag100, and developing dedicated quality assessment metrics, Dragging Accuracy Index and Gemini Score, utilizing Large Multimodal Models. Extensive experiments demonstrate that the proposed GoodDrag compares favorably against the state-of-the-art approaches both qualitatively and quantitatively. The source code and data will be released.", "keywords": [ "drag editing", "generative AI", "diffusion model" ], "rating_list": [ 6, 6, 5, 3, 8 ], "soundness_list": [ 2, 3, 3, 1, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 1, 4 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "VtT41Nniu4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Jingxi Liao", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "MS student" }, { "name": "Meng Wang", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Richang Hong", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Shijie Hao", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 31, "n_formula_1": 14, "n_ref_uni": 34, "n_ref": 69, "n_ref_all": 88, "n_fig": 7, "n_tab": 6, "L_tab": 8186, "n_element_tab": 881, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 5444, "n_element_tab_1": 564, "formula_len_all": 2584, "formula_len_all_1": 704, "len_all": 156845, "len_all_1": 68782, "len_abs": 476, "len_title": 150, "len_sents": 32944, "len_sents_1": 24542, "n_sents": 292, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 970, "title": "GT-Mean Loss: A Simple Yet Effective Solution for Brightness Mismatch in Low-Light Image Enhancement", "abs": "Low-light image enhancement (LLIE) aims to improve the visual quality of images captured under poor lighting conditions. In supervised LLIE tasks, there exists a significant yet often overlooked inconsistency between the overall brightness of an enhanced image and its ground truth counterpart, referred to as brightness mismatch in this study. Brightness mismatch negatively impact supervised LLIE models by misleading model training. However, this issue is largely neglected in current research. In this context, we propose the GT-mean loss, a simple yet effective loss function directly modeling the mean values of images from a probabilistic perspective. The GT-mean loss is flexible, as it extends existing supervised LLIE loss functions into the GT-mean form with minimal additional computational costs. Extensive experiments demonstrate that the incorporation of the GT-mean loss results in consistent performance improvements across various methods and datasets.", "keywords": [ "Low-light image enhancement", "loss function", "GT-mean" ], "rating_list": [ 3, 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 1, 2, 3, 2, 3 ], "confidence_list": [ 5, 5, 5, 4, 4 ] }, { "paper_id": "VtP7CamOR5", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Angelica I Aviles-Rivero", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Carola-Bibiane Sch\u00f6nlieb", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "Full Professor" }, { "name": "Chun-Wun Cheng", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Guang Yang", "gender": "unknown", "institution": "Imperial College London, Imperial College London", "country": "GB", "position": "UKRI Future Leaders Fellow (Tenured)" }, { "name": "Jiahao Huang", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Yi Zhang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 24, "n_ref_uni": 30, "n_ref": 49, "n_ref_all": 57, "n_fig": 3, "n_tab": 3, "L_tab": 1115, "n_element_tab": 128, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 809, "n_element_tab_1": 89, "formula_len_all": 2558, "formula_len_all_1": 2597, "len_all": 101100, "len_all_1": 63842, "len_abs": 1190, "len_title": 126, "len_sents": 30170, "len_sents_1": 27130, "n_sents": 232, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1194, "title": "Mamba Neural Operator: Who Wins? Transformers vs. State-Space Models for PDEs", "abs": "Partial differential equations (PDEs) are widely used to model complex physical systems, but solving them efficiently remains a significant challenge. Recently, Transformers have emerged as the preferred architecture for PDEs due to their ability to capture intricate dependencies. However, they struggle with representing continuous dynamics and long-range interactions. To overcome these limitations, we introduce the Mamba Neural Operator (MNO), a novel framework that enhances neural operator-based techniques for solving PDEs. MNO establishes a formal theoretical connection between structured state-space models (SSMs) and neural operators, offering a unified structure that can adapt to diverse architectures, including Transformer-based models. By leveraging the structured design of SSMs, MNO captures long-range dependencies and continuous dynamics more effectively than traditional Transformers. Through extensive analysis, we show that MNO significantly boosts the expressive power and accuracy of neural operators, making it not just a complement but a superior framework for PDE-related tasks, bridging the gap between efficient representation and accurate solution approximation.", "keywords": [ "Neural Operator", "Deep Learning", "PDEs" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "Vszt1FDElj", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Ekta Srivastava", "gender": "Female", "institution": "Indian Institute of Technology, Delhi", "country": "IN", "position": "Postdoc" }, { "name": "Kumar Arjun", "gender": "Male", "institution": "Indian Institute of Technology, Delhi", "country": "IN", "position": "Undergrad student" }, { "name": "Manoj Kumar", "gender": "Male", "institution": "LNM Institute of Information Technology, Jaipur", "country": "IN", "position": "Assistant Professor" }, { "name": "Sandeep Kumar", "gender": "Male", "institution": "Indian Institute of Technology Delhi", "country": "IN", "position": "Assistant Professor" } ], "n_formula": 20, "n_formula_1": 20, "n_ref_uni": 27, "n_ref": 43, "n_ref_all": 52, "n_fig": 5, "n_tab": 4, "L_tab": 2444, "n_element_tab": 297, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1934, "n_element_tab_1": 211, "formula_len_all": 953, "formula_len_all_1": 952, "len_all": 121784, "len_all_1": 66802, "len_abs": 1576, "len_title": 84, "len_sents": 34137, "len_sents_1": 29251, "n_sents": 238, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1582, "title": "Coarsening to Conceal: Enabling Privacy-Preserving Federated Learning for Graph Data", "abs": "With the escalating demand for privacy-preserving machine learning, federated learning (FL) stands out by enabling collaboration among decentralized entities. Utilizing graph representations of data enhances learning for graph-level tasks, crucial for FL with data distributed across local repositories. Despite its benefits, stringent privacy regulations often compromise FL's performance. Previous methods aimed at ensuring privacy introduce performance degradation and computational overhead. In response to these challenges, we propose using graph coarsening\u2014a simple yet effective method\u2014to enhance the security and privacy of FL on graph data. Our approach posits that graph coarsening alone can suffice for privacy guarantees, as model parameters obtained from training on the coarsened graph effectively conceal sensitive information susceptible to privacy attacks. Through comprehensive application and analysis, we demonstrate the efficacy of graph coarsening within an FL setup, taking both the graph matrix and node features as input, and jointly learning the coarsened graph matrix and feature matrix while ensuring desired properties. The resultant coarsened graph representations are then utilized to train model parameters, subsequently communicated within an FL framework for downstream tasks such as classification. Extensive experimentation across various datasets confirms that graph coarsening ensures privacy while enhancing performance with minimal trade-offs compared to traditional differential privacy (DP) methods without adding extra complexity overhead.", "keywords": [ "Federated Learning", "Privacy-Preserving Machine Learning", "Graph Neural Networks", "Graph Coarsening", "Data Privacy and Security" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 1, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "VsxbWTDHjh", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Alberto Pepe", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Joan Lasenby", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Mattia Montanari", "gender": "Not Specified", "institution": "University of Oxford", "country": "GB", "position": "Instructor" } ], "n_formula": 23, "n_formula_1": 13, "n_ref_uni": 28, "n_ref": 70, "n_ref_all": 109, "n_fig": 21, "n_tab": 9, "L_tab": 2518, "n_element_tab": 330, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 300, "n_element_tab_1": 18, "formula_len_all": 1340, "formula_len_all_1": 662, "len_all": 166260, "len_all_1": 60037, "len_abs": 1112, "len_title": 135, "len_sents": 52935, "len_sents_1": 28520, "n_sents": 409, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1139, "title": "Fengbo: a Clifford Neural Operator pipeline for 3D PDEs in Computational Fluid Dynamics", "abs": "We introduce Fengbo, a pipeline entirely in Clifford Algebra to solve 3D partial differential equations (PDEs) specifically for computational fluid dynamics (CFD). Fengbo is an architecture composed of only 3D convolutional and Fourier Neural Operator (FNO) layers, all working in 3D Clifford Algebra. It models the PDE solution problem as an interpretable mapping from the geometry to the physics of the problem. Despite having just few layers, Fengbo achieves competitive accuracy, superior to 5 out of 6 proposed models reported in \\cite{li2024geometry} for the $\\emph{ShapeNet Car}$ dataset, and it does so with only 42 million trainable parameters, at a reduced computational complexity compared to graph-based methods, and estimating jointly pressure \\emph{and} velocity fields. In addition, the output of each layer in Fengbo can be clearly visualised as objects and physical quantities in 3D space, making it a whitebox model. \nBy leveraging Clifford Algebra and establishing a direct mapping from the geometry to the physics of the PDEs, Fengbo provides an efficient, geometry- and physics-aware approach to solving complex PDEs.", "keywords": [ "Clifford Algebra", "Neural Operator", "PDE Modelling", "Geometric Machine Learning" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 3, 3, 5, 4, 3 ] }, { "paper_id": "VqAX9Lzdqv", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Adarsh Pyarelal", "gender": "Male", "institution": "University of Arizona", "country": "US", "position": "Assistant Professor" }, { "name": "Justin Lieffers", "gender": "Male", "institution": "University of Arizona", "country": "", "position": "Postdoc" }, { "name": "Liang Zhang", "gender": "Male", "institution": "University of Arizona", "country": "", "position": "Researcher" } ], "n_formula": 22, "n_formula_1": 17, "n_ref_uni": 34, "n_ref": 54, "n_ref_all": 85, "n_fig": 14, "n_tab": 6, "L_tab": 590, "n_element_tab": 45, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 55, "n_element_tab_1": 5, "formula_len_all": 971, "formula_len_all_1": 583, "len_all": 154713, "len_all_1": 56143, "len_abs": 1211, "len_title": 133, "len_sents": 50857, "len_sents_1": 25905, "n_sents": 412, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1215, "title": "Enhancing Interpretability in Deep Reinforcement Learning through Semantic Clustering", "abs": "In this paper, we explore semantic clustering properties of deep reinforcement learning (DRL) to improve its interpretability and deepen our understanding of the internal semantic organization. In this context, semantic clustering refers to the ability of neural networks to cluster inputs based on their semantic similarity in the internal space. We propose a DRL architecture that incorporates a novel semantic clustering module, which includes both feature dimensionality reduction and online clustering. This module integrates seamlessly into the DRL training pipeline, addressing the instability of t-SNE and eliminating the need for extensive manual annotation in the previous semantic analysis methods. Through experiments, we validate the effectiveness of the proposed module and demonstrate its ability to reveal semantic clustering properties within DRL. Furthermore, we introduce new analytical methods that leverage these properties to provide insights into the hierarchical structure of policies and the semantic organization within the feature space. These methods also help identify potential risks within the model, offering a deeper understanding of its limitations and guiding future improvements.", "keywords": [ "Interpretability", "Reinforcement Learning", "Clustering", "Semantics", "VAE" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 4, 1, 3, 2 ], "presentation_list": [ 4, 1, 2, 3 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "Vq65R88Wx0", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Guopeng Li", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Ke Yan", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Qiang Wang", "gender": "Male", "institution": "Tencent YouTu Lab", "country": "CN", "position": "Researcher" }, { "name": "Shouhong Ding", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "researcher" }, { "name": "Yuan Gao", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Associate Professor" }, { "name": "Gui-song Xia", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 22, "n_ref": 192, "n_ref_all": 242, "n_fig": 16, "n_tab": 5, "L_tab": 2340, "n_element_tab": 264, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 3041, "n_element_tab_1": 304, "formula_len_all": 671, "formula_len_all_1": 580, "len_all": 153808, "len_all_1": 73578, "len_abs": 1683, "len_title": 117, "len_sents": 48130, "len_sents_1": 30427, "n_sents": 358, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1665, "title": "TAS: Distilling Arbitrary Teacher and Student via a Hybrid Assistant", "abs": "Most knowledge distillation (KD) methodologies predominantly focus on teacher-student pairs with similar architectures, such as both being convolutional neural networks (CNNs). However, the potential and flexibility of KD can be greatly improved by expanding it to novel Cross-Architecture KD (CAKD), where the knowledge of homogeneous and heterogeneous teachers can be transferred flexibly to a given student. The primary challenge in CAKD lies in the substantial feature gaps between heterogeneous models, originating from the distinction of their inherent inductive biases and module functions. To this end, we introduce an assistant model as a bridge to facilitate smooth feature knowledge transfer between heterogeneous teachers and students. More importantly, within our proposed design principle, the assistant model combines the advantages of cross-architecture inductive biases and module functions by merging convolution and attention modules derived from both student and teacher module functions. Furthermore, we observe that heterogeneous features exhibit diverse spatial distributions in CAKD, hindering the effectiveness of conventional pixel-wise mean squared error (MSE) loss. Therefore, we leverage a spatial-agnostic InfoNCE loss to align features after spatial smoothing, thereby improving the feature alignments in CAKD. Our proposed method is evaluated across some homogeneous model pairs and arbitrary heterogeneous combinations of CNNs, ViTs, and MLPs, achieving state-of-the-art performance for distilled models with a maximum gain of 11.47% on CIFAR-100 and 3.67% on ImageNet-1K for distilled models. Our code and models will be released.", "keywords": [ "knowledge distillation; computer vision; image classification;" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "VpeAsLmcvg", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Krishn Vishwas Kher", "gender": "Male", "institution": "Indian Institute of Technology, Hyderabad", "country": "IN", "position": "PhD student" }, { "name": "Rishabh Lalla", "gender": "Male", "institution": "Indian Institute of Technology, Hyderabad", "country": "IN", "position": "Research Associate" }, { "name": "Susmit Agrawal", "gender": "Male", "institution": "Indian Institute of Technology, Hyderabad, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "MS student" }, { "name": "Swarnim Maheshwari", "gender": "Not Specified", "institution": "Indian Institute of Technology, Hyderabad", "country": "IN", "position": "Researcher" }, { "name": "Vineeth Balasubramanian", "gender": "Male", "institution": "Research, Microsoft", "country": "IN", "position": "Principal Researcher" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 41, "n_ref": 81, "n_ref_all": 92, "n_fig": 1, "n_tab": 6, "L_tab": 2937, "n_element_tab": 178, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2149, "n_element_tab_1": 134, "formula_len_all": 879, "formula_len_all_1": 782, "len_all": 137786, "len_all_1": 72353, "len_abs": 1457, "len_title": 109, "len_sents": 37691, "len_sents_1": 33805, "n_sents": 254, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1604, "title": "Singular Value Adaptation for Parameter-Efficient Fine Tuning", "abs": "Parameter-Efficient Fine-Tuning (PEFT) has become a crucial approach in handling the growing complexity of large models and vast datasets across multiple fields such as Computer Vision or Natural Language Processing. Among the most promising of these methods are Low-Rank Adaptation (LoRA) and its derivatives, which fine-tune a pre-trained weight matrix $\\mathbf{W}$ by introducing a low-rank update matrix $\\mathbf{\\Delta W}$. While these approaches have demonstrated strong empirical performance, they remain largely heuristic, with little theoretical grounding to explain their behavior or guide the design of $\\mathbf{\\Delta W}$ for different objectives. This lack of theoretical insight limits our understanding of when these methods are most effective and how they can be systematically improved. In this paper, we propose a theoretical framework for analyzing and designing LoRA-based methods, with a focus on the formulation of $\\mathbf{\\Delta W}$. By establishing a deeper understanding of the interplay between $\\mathbf{W}$ and $\\mathbf{\\Delta W}$, we aim to enable more efficient and targeted fine-tuning strategies, opening the door to novel variants that strike an optimal balance between performance and efficiency. Our proposed method - \\textbf{Si}ngular \\textbf{V}alue \\textbf{A}daptation - uses insights from our theoretical framework to incorporate inductive biases on the formulation of $\\mathbf{\\Delta W}$, leading to a PEFT method that is up to 50$\\times$ more parameter efficient that LoRA, while achieving comparable or better performance across various vision and language tasks.", "keywords": [ "Transfer learning", "Adaptation", "Parameter-Efficient Fine-tuning" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 1, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "VpWki1v2P8", "primary_area": "optimization", "n_author": 8, "author_info_list": [ { "name": "Cho-Jui Hsieh", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Felix Yu", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Inderjit S. Dhillon", "gender": "Male", "institution": "Google", "country": "US", "position": "Distinguished Scientist" }, { "name": "Jui-Nan Yen", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Sai Surya Duvvuri", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Sanjiv Kumar", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Si Si", "gender": "Female", "institution": "Google", "country": "", "position": "research scientist" }, { "name": "Zhao Meng", "gender": "Male", "institution": "Google", "country": "CH", "position": "Google" } ], "n_formula": 43, "n_formula_1": 40, "n_ref_uni": 31, "n_ref": 66, "n_ref_all": 81, "n_fig": 1, "n_tab": 4, "L_tab": 1345, "n_element_tab": 189, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1998, "formula_len_all_1": 2047, "len_all": 120733, "len_all_1": 62161, "len_abs": 1046, "len_title": 132, "len_sents": 34674, "len_sents_1": 26308, "n_sents": 331, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 83, "L_abs": 1029, "title": "LoRA Done RITE: Robust Invariant Transformation Equilibration for LoRA Optimization", "abs": "Low-rank adaption (LoRA) is a widely used parameter-efficient finetuning method for LLM that reduces memory requirements. However, current LoRA optimizers lack transformation invariance, meaning the updates depending on how the two LoRA factors are scaled or rotated. This deficiency leads to inefficient learning and sub-optimal solutions in practice. This paper introduces LoRA-RITE, a novel adaptive matrix preconditioning method for LoRA optimization, which can achieve transformation invariance and remain computationally efficient. We provide theoretical analysis to demonstrate the benefit of our method and conduct experiments on various LLM tasks with different models including Gemma 2B, 7B, and mT5-XXL. The results demonstrate consistent improvements against existing optimizers. For example, replacing Adam with LoRA-RITE during LoRA fine-tuning of Gemma-2B yielded 4.6% accuracy gain on Super-Natural Instructions and 3.5% accuracy gain across other four LLM benchmarks (HellaSwag, ArcChallenge, GSM8K, OpenBookQA).", "keywords": [ "optimization", "LoRA" ], "rating_list": [ 8, 8, 10 ], "soundness_list": [ 4, 4, 4 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "VpOwviiYxf", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Ali Siahkoohi", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Postdoc" }, { "name": "Josselin Garnier", "gender": "Male", "institution": "Ecole polytechnique", "country": "FR", "position": "Full Professor" }, { "name": "Knut Solna", "gender": "Male", "institution": "University of California, Irvine", "country": "", "position": "Full Professor" }, { "name": "Lorenzo Baldassari", "gender": "unknown", "institution": "University of Basel", "country": "", "position": "Postdoc" }, { "name": "Maarten v. de Hoop", "gender": "unknown", "institution": "Rice University", "country": "US", "position": "Full Professor" } ], "n_formula": 103, "n_formula_1": 22, "n_ref_uni": 79, "n_ref": 144, "n_ref_all": 158, "n_fig": 0, "n_tab": 2, "L_tab": 833, "n_element_tab": 106, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15411, "formula_len_all_1": 1285, "len_all": 186198, "len_all_1": 75272, "len_abs": 2807, "len_title": 167, "len_sents": 40718, "len_sents_1": 32595, "n_sents": 320, "n_sents_1": 195, "venue": "ICLR2025", "decision": "withdraw", "L_title": 119, "L_abs": 1452, "title": "Theoretical Convergence Analysis for Hilbert Space MCMC with Score-based Priors for Nonlinear Bayesian Inverse Problems", "abs": "In recent years, several works have explored the use of score-based generative models as expressive priors in Markov chain Monte Carlo (MCMC) algorithms for provable posterior sampling, even in the challenging case of nonlinear Bayesian inverse problems. However, these approaches have been mostly limited to finite-dimensional approximations, while the original problems are typically defined in function spaces of infinite dimension. It is well known that algorithms designed for finite-dimensional settings can encounter theoretical and practical issues when applied to infinite-dimensional objects, such as an inconsistent behavior across different discretizations. In this work, we address this limitation by leveraging the recently developed framework for score-based generative models in Hilbert spaces to learn an infinite-dimensional score, which we use as a prior in a function-space Langevin-type MCMC algorithm, providing theoretical guarantees for convergence in the context of nonlinear Bayesian inverse problems. Crucially, we prove that controlling the approximation error of the score is not only essential for ensuring convergence but also that modifying the standard score-based Langevin MCMC through the selection of an appropriate preconditioner is necessary. Our analysis shows how the control over the score approximation error influences the design of the preconditioner---an aspect unique to the infinite-dimensional setting.", "keywords": [ "theory paper", "theoretical convergence analysis", "nonlinear inverse problems", "bayesian inference", "hilbert space", "Langevin MCMC", "score-based generative models" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "VpGsy4hKMc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jiale Xu", "gender": "Male", "institution": "Tencent ARC Lab", "country": "CN", "position": "Researcher" }, { "name": "Shenghua Gao", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Ying Shan", "gender": "Male", "institution": "Tencent AI Lab Center of Visual Computing", "country": "CN", "position": "Director" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 50, "n_ref": 141, "n_ref_all": 174, "n_fig": 17, "n_tab": 10, "L_tab": 2413, "n_element_tab": 357, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1377, "n_element_tab_1": 206, "formula_len_all": 981, "formula_len_all_1": 423, "len_all": 197964, "len_all_1": 65138, "len_abs": 1262, "len_title": 124, "len_sents": 49562, "len_sents_1": 28581, "n_sents": 374, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1268, "title": "FreeSplatter: Pose-free Gaussian Splatting for Sparse-view 3D Reconstruction", "abs": "Existing sparse-view reconstruction models heavily rely on accurate known camera poses. However, deriving camera extrinsics and intrinsics from sparse-view images poses significant challenges. In this work, we present FreeSplatter, a highly scalable, feed-forward reconstruction framework capable of generating high-quality 3D Gaussians from uncalibrated sparse-view images and recovering their camera parameters in mere seconds. FreeSplatter is built upon a streamlined transformer architecture, consisting of sequential self-attention blocks that facilitate information exchange among multi-view image tokens and decode them into pixel-wise 3D Gaussian primitives. The predicted Gaussian primitives are situated in a unified reference frame, enabling high-fidelity 3D modeling and instant camera parameter estimation with off-the-shelf solvers. To cater to both object-centric and scene-level reconstruction, we train two model variants of FreeSplatter on a large amount of data. In both scenarios, FreeSplatter outperforms state-of-the-art baselines in terms of reconstruction quality and pose estimation accuracy. Furthermore, we showcase FreeSplatter's potential in enhancing the productivity of downstream applications, such as text/image-to-3D content creation.", "keywords": [ "Gaussian Splatting", "3D Reconstruction", "Sparse-view Reconstruction", "Camera Pose Estimation", "Large Reconstruction Models" ], "rating_list": [ 5, 5, 6, 6, 3 ], "soundness_list": [ 3, 1, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 5, 5, 4, 4, 4 ] }, { "paper_id": "Vp2OAxMs2s", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Daniel Durstewitz", "gender": "unknown", "institution": "Heidelberg University", "country": "GB", "position": "Full Professor" }, { "name": "Elias Weber", "gender": "Not Specified", "institution": "Zentralinstitut f\u00fcr Seelische Gesundheit", "country": "DE", "position": "MS student" }, { "name": "Georgia Koppe", "gender": "Female", "institution": "Heidelberg University, Ruprecht-Karls-Universit\u00e4t Heidelberg", "country": "DE", "position": "Full Professor" }, { "name": "Manuel Brenner", "gender": "Male", "institution": "Heidelberg University", "country": "", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 11, "n_ref_uni": 73, "n_ref": 157, "n_ref_all": 221, "n_fig": 42, "n_tab": 6, "L_tab": 2405, "n_element_tab": 144, "n_fig_1": 24, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1637, "formula_len_all_1": 822, "len_all": 229451, "len_all_1": 59435, "len_abs": 1297, "len_title": 131, "len_sents": 67353, "len_sents_1": 28266, "n_sents": 438, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1284, "title": "Learning Interpretable Hierarchical Dynamical Systems Models from Time Series Data", "abs": "In science, we are often interested in obtaining a generative model of the underlying system dynamics from observed time series. While powerful methods for dynamical systems reconstruction (DSR) exist when data come from a single domain, how to best integrate data from multiple dynamical regimes and leverage it for generalization is still an open question. This becomes particularly important when individual time series are short, and group-level information may help to fill in for gaps in single-domain data. Here we introduce a hierarchical framework that enables to harvest group-level (multi-domain) information while retaining all single-domain characteristics, and showcase it on popular DSR benchmarks, as well as on neuroscience and medical data. In addition to faithful reconstruction of all individual dynamical regimes, our unsupervised methodology discovers common low-dimensional feature spaces in which datasets with similar dynamics cluster. The features spanning these spaces were further dynamically highly interpretable, surprisingly in often linear relation to control parameters that govern the dynamics of the underlying system. Finally, we illustrate transfer learning and generalization to new parameter regimes, paving the way toward DSR foundation models.", "keywords": [ "dynamical systems", "recurrent neural networks", "hierarchical modelling", "time series", "interpretability", "Nonlinear Dynamics", "foundation models" ], "rating_list": [ 8, 1, 8, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "VojvkUEq8q", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Hyotaek Jeon", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Hyunwook Lee", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "Postdoc" }, { "name": "Juwon Kim", "gender": "unknown", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Ko Keun Kim", "gender": "Male", "institution": "LG Corporation", "country": "KR", "position": "Principal Researcher" }, { "name": "Minsik Lee", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Sungahn Ko", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 57, "n_ref": 173, "n_ref_all": 203, "n_fig": 7, "n_tab": 4, "L_tab": 2039, "n_element_tab": 73, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1093, "n_element_tab_1": 54, "formula_len_all": 715, "formula_len_all_1": 358, "len_all": 162660, "len_all_1": 70160, "len_abs": 2097, "len_title": 149, "len_sents": 46188, "len_sents_1": 33159, "n_sents": 308, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 2195, "title": "Learning by Causality to Improve Channel Dependency Modeling in Multivariate Time Series Forecasting", "abs": "Beyond the conventional long-term temporal dependency modeling, multivariate time series (MTS) forecasting has rapidly shifted toward channel dependency (CD) modeling. This shift significantly improves modeling quality by fully leveraging both multivariate relationships and temporal dependencies. Recent methods primarily model channel dependency through correlation learning (e.g., crossattention) or non-trainable statistical techniques (e.g., cross-correlation). However, these approaches struggle to fully capture the intrinsic relationships within MTS, particularly those stemming from directed cause-effect (i.e., causality) and nonstationary variates originating from diverse sources. In addition, causality may arise from the signals with different temporal behaviors, such as varying periodicity or discrete event sequences, which is not sufficiently discussed before. In this paper, we propose CALAS (Causality-enhanced Attention with Learnable and Adaptive Spacing), the first end-to-end learning method for MTS forecasting that uncover causality among variates without relying on statistical measures or prior knowledge. To model underlying causality, which consists of causal strength and propagation delay, we newly design a hypernetworks-based 1D convolutions mechanism. Inspired by dilated convolution with learnable spacings (DCLS) and spiking neural networks (SNNs), we extend discrete time delay into a continuous Gaussian kernel. Combining the hypernetworks-generated Gaussian kernel and convolutional weights (i.e., attention or causal strength), we achieve the end-to-end dynamic causality modeling mechanism. This mechanism enhances the model\u2019s ability to capture time-varying causality across multi-source variates, ultimately improving the prediction accuracy, quality, and interpretability. For evaluation, we conduct extensive experiments with six real-world datasets and qualitative analysis to demonstrate CALAS\u2019s superiority in capturing varying causality in a data-agnostic manner. The experiment results indicate that CALAS has significantly improved MTS forecasting accuracy compared to state-of-the-art methods by dynamically modeling causality among variates.", "keywords": [ "Multivariate Time Series Forecasting", "Channel Dependency", "Deep Learning" ], "rating_list": [ 3, 3, 8, 5 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "VoayJihXra", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 5, "author_info_list": [ { "name": "Honguk Woo", "gender": "Male", "institution": "Sungkyunkwan University", "country": "KR", "position": "Associate Professor" }, { "name": "Jinwoo Park", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "", "position": "Undergrad student" }, { "name": "Sanghyun Ahn", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "Undergrad student" }, { "name": "Wonje Choi", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "PhD student" }, { "name": "Lee Dae Hee", "gender": "Male", "institution": "Sungkyunkwan University", "country": "KR", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 52, "n_ref": 108, "n_ref_all": 166, "n_fig": 13, "n_tab": 20, "L_tab": 7337, "n_element_tab": 850, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 6522, "n_element_tab_1": 541, "formula_len_all": 1812, "formula_len_all_1": 583, "len_all": 247324, "len_all_1": 73761, "len_abs": 1523, "len_title": 132, "len_sents": 72498, "len_sents_1": 27383, "n_sents": 622, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1381, "title": "NeSyC: A Neuro-symbolic Continual Learner For Complex Embodied Tasks in Open Domains", "abs": "We explore neuro-symbolic approaches to generalize actionable knowledge, enabling embodied agents to tackle complex tasks more effectively in open-domain environments. A key challenge for embodied agents is the generalization of knowledge across diverse environments and situations, as limited experiences often confine them to their prior knowledge. To address this issue, we introduce a novel framework, NeSyC, a neuro-symbolic continual learner that emulates the hypothetico-deductive model by continually formulating and validating knowledge from limited experiences through the combined use of Large Language Models (LLMs) and symbolic tools. Specifically, we devise a contrastive generality improvement scheme within NeSyC, which iteratively generates hypotheses using LLMs and conducts contrastive validation via symbolic tools. This scheme reinforces the justification for admissible actions while minimizing the inference of inadmissible ones. Additionally, we incorporate a memory-based monitoring scheme that efficiently detects action errors and triggers the knowledge refinement process across domains. Experiments conducted on diverse embodied task benchmarks\u2014including ALFWorld, VirtualHome, Minecraft, RLBench, and a real-world robotic scenario\u2014demonstrate that NeSyC is highly effective in solving complex embodied tasks across a range of open-domain environments.", "keywords": [ "Embodied AI", "Neuro-symbolic AI" ], "rating_list": [ 8, 5, 5, 10 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "VoZ0nSy0Ry", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Cao Sheng", "gender": "Male", "institution": "China Tower Corporation Limited", "country": "CN", "position": "Researcher" }, { "name": "Fei Li", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "JiaZhong Yu", "gender": "Male", "institution": "China Tower Corporation Limited", "country": "CN", "position": "Principal Researcher" }, { "name": "Jinghan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yadong MU", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhicheng Sun", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 11, "n_ref_uni": 71, "n_ref": 170, "n_ref_all": 205, "n_fig": 17, "n_tab": 5, "L_tab": 1444, "n_element_tab": 181, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 432, "n_element_tab_1": 29, "formula_len_all": 717, "formula_len_all_1": 416, "len_all": 194157, "len_all_1": 56410, "len_abs": 1106, "len_title": 123, "len_sents": 48015, "len_sents_1": 26484, "n_sents": 372, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1109, "title": "Closed-Loop Long-Horizon Robotic Planning via Equilibrium Sequence Modeling", "abs": "In the endeavor to make autonomous robots take actions, task planning is a major challenge that requires translating high-level task descriptions into long-horizon action sequences. Despite recent advances in language model agents, they remain prone to planning errors and limited in their ability to plan ahead. To address these limitations in robotic planning, we advocate a self-refining scheme that iteratively refines a draft plan until an equilibrium is reached. Remarkably, this process can be optimized end-to-end from an analytical perspective without the need to curate additional verifiers or reward models, allowing us to train self-refining planners in a simple supervised learning fashion. Meanwhile, a nested equilibrium sequence modeling procedure is devised for efficient closed-loop planning that incorporates useful feedback from the environment (or an internal world model). Our method is evaluated on the VirtualHome-Env benchmark, showing advanced performance with better scaling for inference computation. Code is available at https://github.com/anonymous-iclr-2025/equilibrium-planner.", "keywords": [ "Deep Equilibrium Models", "Large Language Models", "Robot Task Planning" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 3, 1, 3, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "VoVVDaVXdf", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Akira Taniguchi", "gender": "unknown", "institution": "Ritsumeikan University", "country": "", "position": "Lecturer" }, { "name": "Nguyen Le Hoang", "gender": "Male", "institution": "Ritsumeikan University", "country": "JP", "position": "PhD student" }, { "name": "Tadahiro Taniguchi", "gender": "Male", "institution": "Kyoto University", "country": "JP", "position": "Full Professor" }, { "name": "Tianwei Fang", "gender": "Male", "institution": "Kyoto University", "country": "JP", "position": "PhD student" } ], "n_formula": 45, "n_formula_1": 19, "n_ref_uni": 54, "n_ref": 73, "n_ref_all": 91, "n_fig": 4, "n_tab": 2, "L_tab": 501, "n_element_tab": 29, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 803, "n_element_tab_1": 42, "formula_len_all": 3213, "formula_len_all_1": 1479, "len_all": 150625, "len_all_1": 60853, "len_abs": 1462, "len_title": 142, "len_sents": 40653, "len_sents_1": 28117, "n_sents": 331, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1465, "title": "SimSiam Naming Game: A Unified Approach for Representation Learning and Emergent Communication", "abs": "Emergent communication, driven by generative models, enables agents to develop a shared language for describing their individual views of the same objects through interactions. Meanwhile, self-supervised learning (SSL), particularly SimSiam, uses discriminative representation learning to make representations of augmented views of the same data point closer in the representation space. Building on the prior work of VI-SimSiam, which incorporates a generative and Bayesian perspective into the SimSiam framework via variational inference (VI) interpretation, we propose SimSiam+VAE, a unified approach for both representation learning and emergent communication. SimSiam+VAE integrates a variational autoencoder (VAE) into the predictor of the SimSiam network to enhance representation learning and capture uncertainty. Experimental results show that SimSiam+VAE outperforms both SimSiam and VI-SimSiam. We further extend this model into a communication framework called the SimSiam Naming Game (SSNG), which applies the generative and Bayesian approach based on VI to develop internal representations and emergent language while utilizing the discriminative process of SimSiam to facilitate mutual understanding between agents. In experiments with established models, despite the dynamic alternation of agent roles during interactions, SSNG demonstrates comparable performance to the referential game and slightly outperforms the Metropolis-Hastings naming game.", "keywords": [ "emergent communication", "representation learning", "SimSiam", "contrastive learning", "self-supervised learning" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "VoI4d6uhdr", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Arjun Subramonian", "gender": "Agender", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Elvis Dohmatob", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Levent Sagun", "gender": "Non-Binary", "institution": "Meta", "country": "", "position": "Research scientist" }, { "name": "Samuel Bell", "gender": "Male", "institution": "Facebook", "country": "GB", "position": "Researcher" } ], "n_formula": 362, "n_formula_1": 27, "n_ref_uni": 42, "n_ref": 83, "n_ref_all": 104, "n_fig": 8, "n_tab": 3, "L_tab": 274, "n_element_tab": 37, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 27961, "formula_len_all_1": 2135, "len_all": 275190, "len_all_1": 71519, "len_abs": 6193, "len_title": 89, "len_sents": 75621, "len_sents_1": 33282, "n_sents": 818, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 1191, "title": "An Effective Theory of Bias Amplification", "abs": "Machine learning models may capture and amplify biases present in data, leading to disparate test performance across social groups. To better understand, evaluate, and mitigate these possible biases, a deeper theoretical understanding of how model design choices and data distribution properties could contribute to bias is needed. In this work, we contribute a precise analytical theory in the context of ridge regression, both with and without random projections, where the former models neural networks in a simplified regime. Our theory offers a unified and rigorous explanation of machine learning bias, providing insights into phenomena such as bias amplification and minority-group bias in various feature and parameter regimes. For example, we demonstrate that there may be an optimal regularization penalty or training time to avoid bias amplification, and there can be fundamental differences in test error between groups that do not vanish with increased parameterization. Importantly, our theoretical predictions align with several empirical observations reported in the literature. We extensively empirically validate our theory on diverse synthetic and semi-synthetic datasets.", "keywords": [ "fairness", "algorithmic bias", "machine learning theory", "random matrix theory" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "VoHJTTA2MB", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Eitam Sheetrit", "gender": "unknown", "institution": "Ben Gurion University of the Negev", "country": "", "position": "Postdoc" }, { "name": "Gil Shenderovitz", "gender": "unknown", "institution": "Microsoft", "country": "IL", "position": "Intern" }, { "name": "Menachem Brief", "gender": "Male", "institution": "Microsoft", "country": "IL", "position": "Researcher" }, { "name": "Noga BenYoash", "gender": "unknown", "institution": "Microsoft", "country": "IL", "position": "Researcher" }, { "name": "Oded Ovadia", "gender": "Male", "institution": "Tel Aviv University", "country": "", "position": "PhD student" }, { "name": "Rachel Lemberg", "gender": "Female", "institution": "Weizmann Institute of science Rehovot Israel", "country": "", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 42, "n_ref": 71, "n_ref_all": 89, "n_fig": 5, "n_tab": 7, "L_tab": 4334, "n_element_tab": 557, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 582, "n_element_tab_1": 73, "formula_len_all": 167, "formula_len_all_1": 167, "len_all": 149425, "len_all_1": 59297, "len_abs": 1496, "len_title": 151, "len_sents": 35057, "len_sents_1": 26617, "n_sents": 289, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 105, "L_abs": 1489, "title": "Mixing It Up: The Cocktail Effect of Multi-Task Fine-Tuning on LLM Performance - A Case Study in Finance", "abs": "The application of large language models (LLMs) in domain-specific contexts, including finance, has expanded rapidly. Domain-specific LLMs are typically evaluated based on their performance in various downstream tasks relevant to the domain. In this work, we present a detailed analysis of fine-tuning LLMs for such tasks. Somewhat counterintuitively, we find that in domain-specific cases, fine-tuning exclusively on the target task is not always the most effective strategy. Instead, multi-task fine-tuning - where models are trained on a cocktail of related tasks - can significantly enhance performance. We demonstrate how this approach enables a small model, such as Phi-3-Mini, to achieve state-of-the-art results, even surpassing the much larger GPT-4-o model on financial benchmarks. Our study involves a large-scale experiment, training over 200 models using several widely adopted LLMs as baselines, and empirically confirms the benefits of multi-task fine-tuning. Additionally, we explore the use of general instruction data as a form of regularization, suggesting that it helps minimize performance degradation. We also investigate the inclusion of mathematical data, finding improvements in numerical reasoning that transfer effectively to financial tasks. Finally, we note that while fine-tuning for downstream tasks leads to targeted improvements in task performance, it does not necessarily result in broader gains in domain knowledge or complex domain reasoning abilities.", "keywords": [ "LLM", "Finetuning", "Domain Adaptation", "Finance", "Benchmarks" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "Vo1FUQ4aQI", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Mohamed MOUHAJIR", "gender": "Male", "institution": "ENSIAS, Universit\u00e9 Mohammed V ", "country": "MA", "position": "Undergrad student" }, { "name": "Pin-Yu Chen", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Tsung-Yi Ho", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "", "position": "Full Professor" }, { "name": "ZAITANG LI", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 6, "n_ref_uni": 52, "n_ref": 89, "n_ref_all": 98, "n_fig": 6, "n_tab": 6, "L_tab": 1549, "n_element_tab": 99, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 5, "n_element_tab_1": 1, "formula_len_all": 912, "formula_len_all_1": 314, "len_all": 222258, "len_all_1": 59360, "len_abs": 1812, "len_title": 111, "len_sents": 59183, "len_sents_1": 28841, "n_sents": 474, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1823, "title": "GRE Score: Generative Risk Evaluation for Large Language Models", "abs": "Large Language Models (LLMs) have revolutionized generative tasks, but concerns about their trustworthiness and vulnerability to adversarial attacks persist. This paper introduces the Generative Robustness Evaluation (GRE) Score, a novel metric designed to assess LLMs' resilience against adversarial red teaming attempts that may compromise model compliance and elicit undesired responses. Our approach utilizes conditional generation for synthetic text creation, offering an attack-independent evaluation of LLM robustness. By calculating the margin in refusal scores, we quantify the robustness of LLMs in an attack-agnostic manner. We evaluate our method on five different dimensions with specified datasets, encompassing ethical considerations, safety protocols, and potential misuse scenarios. We present four key contributions: (1) The GRE Score framework, which establishes a textual robustness certificate for LLMs against adversarial red teaming attempts, providing a theoretical foundation for quantifying model resilience. (2) Comprehensive evaluations across five critical dimensions using eight prominent LLMs, validating GRE Scores with adversarial red teaming attacks. Our method demonstrates a consistent ranking of LLM robustness when compared to the attack-based model ranking on TrustLLM \\citep{huang2024trustllm} while achieving a significant 5-8x speedup compared to traditional evaluation techniques. (3) Insights into the non-linear relationship between model scaling and performance, revealing that larger models do not always perform better, and an analysis of how instruction-tuning impacts robustness across LLMs. (4) The discovery that all evaluated LLMs exhibit notably lower performance in robustness and privacy tasks compared to other areas, highlighting a critical gap in LLM capabilities.", "keywords": [ "Large Language Models", "Robustness", "Trustworthy" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "VnaJNW80pN", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Jiaxi Liu", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Intern" }, { "name": "Qitian Wu", "gender": "Male", "institution": "Broad Institute of MIT and Harvard", "country": "US", "position": "Researcher" }, { "name": "Xiaohan Qin", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Intern" }, { "name": "Yang Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 39, "n_ref": 54, "n_ref_all": 61, "n_fig": 1, "n_tab": 2, "L_tab": 2233, "n_element_tab": 229, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 884, "n_element_tab_1": 77, "formula_len_all": 704, "formula_len_all_1": 677, "len_all": 120090, "len_all_1": 67357, "len_abs": 1776, "len_title": 140, "len_sents": 36228, "len_sents_1": 33746, "n_sents": 234, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1784, "title": "Toward Learning Generalized Cross-Problem Solving Strategies for Combinatorial Optimization", "abs": "Combinatorial optimization (CO) problems are fundamental across various domains, with many sharing similarities in optimization objectives, decision variables, and constraints. Many traditional algorithms perform well on related problems using similar solution strategies, highlighting the commonality in solving different problems. However, most machine learning approaches treat each CO problem in isolation, failing to capitalize on the underlying relationships between problems. In this paper, we investigate the potential to learn generalized solving strategies that capture the shared structure among different CO problems, enabling easier adaptation to related tasks. To this end, we propose to first divide the model architecture into three components: a header, an encoder, and a decoder; where The header and decoder address problem-specific inputs and outputs, while the encoder is designed to learn shared strategies that generalize across different problems. To ensure this, we enforce alignment in the optimization directions of the encoder across problems, maintaining consistency in both gradient directions and magnitudes to harmonize optimization processes. This is achieved by introducing the additional problem-specific rotation matrices and loss weights to steer the gradients, which are updated via a gradient consistency loss. Extensive experiments on six CO problems demonstrate that our method enhances the model's ability to capture shared solving strategies across problems. We show that the learned encoder on several problems can directly perform comparably on new problems to models trained from scratch, highlighting its potential to support developing the foundational model for combinatorial optimization. Source code will be made publicly available.", "keywords": [ "Neural Combinatorial Optimization", "Multi-task Learning" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "VnNSkUXejc", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Jing-Cheng Pang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Masashi Sugiyama", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Director" }, { "name": "Qiyu Wu", "gender": "Male", "institution": "Sony", "country": "JP", "position": "Researcher" }, { "name": "Xin-Qiang Cai", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Postdoc" }, { "name": "Yao-Xiang Ding", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuting Tang", "gender": "Female", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 46, "n_ref": 75, "n_ref_all": 82, "n_fig": 4, "n_tab": 1, "L_tab": 520, "n_element_tab": 20, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 679, "formula_len_all_1": 513, "len_all": 132936, "len_all_1": 55796, "len_abs": 1889, "len_title": 143, "len_sents": 44120, "len_sents_1": 27475, "n_sents": 309, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1813, "title": "Beyond Simple Sum of Delayed Rewards: Non-Markovian Reward Modeling for Reinforcement Learning", "abs": "Reinforcement Learning (RL) empowers agents to acquire various skills by learning from reward signals. Unfortunately, designing high-quality instance-level rewards often demands significant effort. An emerging alternative, RL with delayed reward, focuses on learning from rewards presented periodically, which can be obtained from human evaluators assessing the agent's performance over sequences of behaviors. However, traditional methods in this domain assume the existence of underlying Markovian rewards and that the observed delayed reward is simply the sum of instance-level rewards, both of which often do not align well with real-world scenarios. In this paper, we introduce the problem of RL from Composite Delayed Reward (RLCoDe), which generalizes traditional RL from delayed rewards by eliminating the strong assumption. We suggest that the delayed reward may arise from a more complex structure reflecting the overall contribution of the sequence. To address this problem, we present a framework for modeling composite delayed rewards, using a weighted sum of non-Markovian components to capture the different contributions of individual steps. Building on this framework, we propose Composite Delayed Reward Transformer (CoDeTr), which incorporates a specialized in-sequence attention mechanism to effectively model these contributions. We conduct experiments on challenging locomotion tasks where the agent receives delayed rewards computed from composite functions of observable step rewards. The experimental results indicate that CoDeTr consistently outperforms baseline methods across evaluated metrics. Additionally, we demonstrate that it effectively identifies the most significant time steps within the sequence and accurately predicts rewards that closely reflect the environment feedback.", "keywords": [ "Reinforcement Learning", "Delayed Reward" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "VnLhUogHYE", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Heuiseok Lim", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Full Professor" }, { "name": "Jaehyung Seo", "gender": "Male", "institution": "Korea University", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 55, "n_ref": 100, "n_ref_all": 130, "n_fig": 4, "n_tab": 15, "L_tab": 8822, "n_element_tab": 571, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1314, "n_element_tab_1": 175, "formula_len_all": 49, "formula_len_all_1": 158, "len_all": 205766, "len_all_1": 68963, "len_abs": 1573, "len_title": 129, "len_sents": 54056, "len_sents_1": 34161, "n_sents": 360, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 80, "L_abs": 1581, "title": "K-HALU: Multiple Answer Korean Hallucination Benchmark for Large Language Models", "abs": "Recent researchers and companies have been developing large language models (LLMs) specifically designed for particular purposes and have achieved significant advancements in various natural language processing tasks. However, LLMs are still prone to generating hallucinations\u2014results that are unfaithful or inconsistent with the given input. As a result, the need for datasets to evaluate and demonstrate the hallucination detection capabilities of LLMs is increasingly recognized. Nonetheless, the Korean NLP community lacks publicly available benchmark datasets demonstrating the faithfulness of knowledge-based information. Furthermore, the few existing datasets that evaluate hallucination are limited in their access to the entire dataset, restricting detailed analysis beyond simple scoring, and are based on translated English knowledge. To address these challenges, we introduce K-HALU, a Korean benchmark designed to evaluate LLMs' hallucination detection in Korean. This benchmark contains seven domains, considering the faithfulness of statements based on knowledge documents compiled from Korean news, magazines, and books. For more strict evaluation, 40% of the dataset is structured as multiple-answer questions, requiring models to select all possible correct answers from the given options. Our empirical results show that open-source LLMs still struggle with hallucination detection in Korean knowledge, emphasizing the need for a more detailed analysis of their limitations. The K-HALU benchmark will be made publicly available after the anonymous review period.", "keywords": [ "Hallucination", "Benchmark dataset", "Multiple answer", "Korean", "Large language model" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 4, 2, 2 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "VmW7Sf84sj", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Baoxin Li", "gender": "Not Specified", "institution": "Arizona State University", "country": "US", "position": "Full Professor" }, { "name": "Hemanth Venkateswara", "gender": "Male", "institution": "Georgia State University", "country": "", "position": "Assistant Professor" }, { "name": "PrasanthSai Gouripeddi", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Sachin Chhabra", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 4, "n_ref_uni": 22, "n_ref": 48, "n_ref_all": 71, "n_fig": 13, "n_tab": 9, "L_tab": 2335, "n_element_tab": 280, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 1448, "n_element_tab_1": 138, "formula_len_all": 2464, "formula_len_all_1": 244, "len_all": 121632, "len_all_1": 59454, "len_abs": 1070, "len_title": 117, "len_sents": 40825, "len_sents_1": 25158, "n_sents": 343, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1078, "title": "LLS: Regulating Neural Network Training via Learnable Label Smoothing", "abs": "Training a neural network using one-hot targets often leads to the issue of overconfidence. \nTo address this, Label Smoothing has been introduced, modifying the targets to a mix of one-hot encoding and a uniform probability vector. \nHowever, the uniform probability vector indiscriminately assigns equal weights to all categories, thereby undermining inter-category relationships. To overcome these challenges, we propose a novel solution, Learnable Label Smoothing (LLS) that aims to regulate training by granting networks the ability to assign optimal targets. Unlike conventional methods, Learnable Label Smoothing utilizes probability vectors unique to each category, resulting in diverse targets. The acquired relationships are beneficial for regularization and also prove to be transferable, facilitating knowledge distillation even in the absence of a Teacher model. Our extensive experiments across multiple datasets highlight the advantages of our method in addressing both overconfidence and the preservation of inter-category relationships in neural network training.", "keywords": [ "Learnable Label Smoothing", "LLS", "Label Regularization" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "VmR3QvfLxt", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Bryan Sangwoo Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Dohun Lee", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Geon Yeong Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Jong Chul Ye", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 21, "n_formula_1": 15, "n_ref_uni": 21, "n_ref": 42, "n_ref_all": 53, "n_fig": 3, "n_tab": 5, "L_tab": 798, "n_element_tab": 72, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 643, "n_element_tab_1": 57, "formula_len_all": 1227, "formula_len_all_1": 901, "len_all": 105804, "len_all_1": 58278, "len_abs": 1222, "len_title": 0, "len_sents": 35979, "len_sents_1": 27396, "n_sents": 293, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1266, "title": "VideoGuide: Improving Video Diffusion Models without Training Through a Teacher's Guide", "abs": "Text-to-image (T2I) diffusion models have revolutionized visual content creation, but extending these capabilities to text-to-video (T2V) generation remains a challenge, particularly in preserving temporal consistency. Existing methods that aim to improve consistency often cause trade-offs such as reduced imaging quality and impractical computational time. To address these issues we introduce VideoGuide, a novel framework that enhances the temporal consistency of pretrained T2V models without the need for additional training or fine-tuning. Instead, VideoGuide leverages any pretrained video diffusion model (VDM) or itself as a guide during the early stages of inference, improving temporal quality by interpolating the guiding model\u2019s denoised samples into the sampling model's denoising process. The proposed method brings about significant improvement in temporal consistency and image fidelity, providing a cost-effective and practical solution that synergizes the strengths of various video diffusion models. Furthermore, we demonstrate prior distillation, revealing that base models can achieve enhanced text coherence by utilizing the superior data prior of the guiding model through the proposed method. Project Page: https://videoguide2025.github.io/", "keywords": [ "diffusion model", "video diffusion model", "tuning-free", "video generation" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "VmJdqhuTCh", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Amin Karimi Monsefi", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Mengxi Zhou", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Nastaran Karimi Monsefi", "gender": "Female", "institution": "Hamden University of Technology ", "country": "", "position": "Undergrad student" }, { "name": "Rajiv Ramnath", "gender": "Male", "institution": "Ohio State University", "country": "US", "position": "Full Professor" }, { "name": "Wei-Lun Chao", "gender": "Male", "institution": "Ohio State University", "country": "US", "position": "Assistant Professor" }, { "name": "Ser-Nam Lim", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 4, "n_ref_uni": 56, "n_ref": 150, "n_ref_all": 182, "n_fig": 6, "n_tab": 13, "L_tab": 2798, "n_element_tab": 349, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1827, "n_element_tab_1": 20, "formula_len_all": 385, "formula_len_all_1": 204, "len_all": 215482, "len_all_1": 50986, "len_abs": 2865, "len_title": 118, "len_sents": 67977, "len_sents_1": 24417, "n_sents": 436, "n_sents_1": 148, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1406, "title": "Frequency-Guided Masking for Enhanced Vision Self-Supervised Learning", "abs": "We present a novel frequency-based Self-Supervised Learning (SSL) approach that significantly enhances its efficacy for pre-training. Prior work in this direction masks out pre-defined frequencies in the input image and employs a reconstruction loss to pre-train the model. While achieving promising results, such an implementation has two fundamental limitations as identified in our paper. First, using pre-defined frequencies overlooks the variability of image frequency responses. Second, pre-trained with frequency-filtered images, the resulting model needs relatively more data to adapt to naturally looking images during fine-tuning. To address these drawbacks, we propose FOurier transform compression with seLf-Knowledge distillation (FOLK), integrating two dedicated ideas. First, inspired by image compression, we adaptively select the masked-out frequencies based on image frequency responses, creating more suitable SSL tasks for pre-training. Second, we employ a two-branch framework empowered by knowledge distillation, enabling the model to take both the filtered and original images as input, largely reducing the burden of downstream tasks. Our experimental results demonstrate the effectiveness of FOLK in achieving competitive performance to many state-of-the-art SSL methods across various downstream tasks, including image classification, few-shot learning, and semantic segmentation.", "keywords": [ "Self-supervised learning", "Frequency domain analysis", "Knowledge distillation", "Vision Self-Supervised Learning", "Pre-training" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 3, 3, 5 ] }, { "paper_id": "Vlo3Gad3YP", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Dongxia Wu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Ph.D student" }, { "name": "Rose Yu", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Ruijia Niu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Yian Ma", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Nikki Kuang", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "PhD student" } ], "n_formula": 81, "n_formula_1": 20, "n_ref_uni": 58, "n_ref": 91, "n_ref_all": 108, "n_fig": 4, "n_tab": 2, "L_tab": 337, "n_element_tab": 45, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 2814, "n_element_tab_1": 31, "formula_len_all": 7132, "formula_len_all_1": 1568, "len_all": 167379, "len_all_1": 59952, "len_abs": 1328, "len_title": 117, "len_sents": 49381, "len_sents_1": 27145, "n_sents": 386, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1336, "title": "Diff-BBO: Diffusion-Based Inverse Modeling for Black-Box Optimization", "abs": "Black-box optimization (BBO) aims to optimize an objective function by iteratively querying a black-box oracle in a sample-efficient way.\nWhile prior studies focus on forward approaches to learn surrogates for the unknown objective function, they struggle with steering clear of out-of-distribution and invalid inputs. Recently, inverse modeling approaches that map objective space to the design space with conditional diffusion models have demonstrated impressive capability in learning the data manifold. They have shown promising performance in offline BBO tasks. However, these approaches require a pre-collected dataset. How to design the acquisition function for inverse modeling to actively query new data remains an open question. In this work, we propose diffusion-based inverse modeling for black-box optimization (Diff-BBO), an inverse approach leveraging diffusion models for online BBO problem. Instead of proposing candidates in the design space, Diff-BBO employs a novel acquisition function Uncertainty-aware Exploration (UaE) to propose objective function values. Subsequently, we employ a conditional diffusion model to generate samples based on these proposed values within the design space. We demonstrate that using UaE results in optimal optimization outcomes, supported by both theoretical and empirical evidence.", "keywords": [ "Diffusion models", "Black-box Optimization", "Uncertainty Quantification" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 3, 2, 5 ] }, { "paper_id": "Vli7PVO60W", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Balaji Vasan Srinivasan", "gender": "Male", "institution": "Adobe Research", "country": "IN", "position": "Principal Researcher" }, { "name": "Kuldeep Kulkarni", "gender": "Male", "institution": "Adobe Systems", "country": "IN", "position": "Research Scientist" }, { "name": "Prateksha Udhayanan", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 20, "n_ref": 41, "n_ref_all": 55, "n_fig": 2, "n_tab": 9, "L_tab": 2706, "n_element_tab": 233, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 2131, "n_element_tab_1": 148, "formula_len_all": 1747, "formula_len_all_1": 1461, "len_all": 130190, "len_all_1": 68166, "len_abs": 1180, "len_title": 110, "len_sents": 35780, "len_sents_1": 29108, "n_sents": 275, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 984, "title": "MMEval: Evaluating Video Generation Models for Motion Quality", "abs": "Recent advancements in video generation, especially with diffusion models, have led to new challenges in evaluating the generated outputs, highlighting the need for well-curated evaluation metrics and benchmarks. While prior work has focused on assessing text-to-video models for overall video quality, such as temporal coherence and prompt consistency, they overlook a crucial aspect: motion modeling abilities of generative models. To address this gap, we propose a structured approach to evaluate image-to-video generation models, with a focus on their motion modeling abilities. For example, we assess how accurately models generate motions like \"circular movement for a rotating ferris wheel\" or \"oscillatory motion for a pendulum\". We categorize videos into linear, circular, and oscillatory motion-types and formulate metrics to capture key motion properties for each category. Our benchmark, MMEval, along with the code and image-prompt-video sets, will be publicly released.", "keywords": [ "image-to-video", "evaluation" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "VlWWzN7RtJ", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Abduallah Mohamed", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Collaborator" }, { "name": "Abdulwahab Felemban", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Eslam Mohamed BAKR", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Jian Ding", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" }, { "name": "Mohamed Elhoseiny", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Xiaoqian Shen", "gender": "Female", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 36, "n_ref": 63, "n_ref_all": 86, "n_fig": 6, "n_tab": 11, "L_tab": 2271, "n_element_tab": 230, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 1356, "n_element_tab_1": 145, "formula_len_all": 203, "formula_len_all_1": 166, "len_all": 147450, "len_all_1": 69008, "len_abs": 1650, "len_title": 97, "len_sents": 39000, "len_sents_1": 32935, "n_sents": 301, "n_sents_1": 242, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1682, "title": "iMotion-LLM: Motion Prediction Instruction Tuning", "abs": "We introduce iMotion-LLM, a Multimodal Large Language Model (LLM) integrated with trajectory prediction, designed to guide interactive multi-agent scenarios. Unlike conventional multimodal trajectory prediction approaches, iMotion-LLM generates diverse and feasible future trajectories conditioned on textual instructions as a guidance signal. By augmenting real-world driving scenarios in the Waymo Open Motion Dataset (WOMD) with textual motion instructions, we propose InstructWaymo data augmentation. Leveraging this data augmentation, iMotion-LLM integrates a pretrained LLM, fine-tuned with LoRA, to map scene features into the LLM input space. Key results demonstrate that making the trajectory prediction model conditional improves its instruction-following capabilities. Specifically, the integration of the LLM enables a 11.07x ratio of actual-scenario feasible to infeasible recall instruction following, compared to 5.92x when using the Conditional GameFormer alone. These findings highlight the ability of iMotion-LLM to generate trajectories that not only align with feasible instructions but also reject infeasible ones, enhancing overall safety. Despite its improvements in instruction following, iMotion-LLM inherits the strong trajectory prediction performance of the baseline model, making it versatile across different driving modes. This combination of skills positions iMotion-LLM as a powerful augmentation technique for trajectory prediction models, empowering autonomous navigation systems to better interpret and predict the dynamics of multi-agent environments. This work lays the groundwork for future advancements in instruction-based motion prediction.", "keywords": [ "Trajectory Prediction", "Conditional Trajectory Prediction" ], "rating_list": [ 3, 3, 3, 6, 3 ], "soundness_list": [ 3, 2, 2, 3, 1 ], "presentation_list": [ 2, 1, 1, 3, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 5, 4, 4 ] }, { "paper_id": "VkqqZcofEu", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 9, "author_info_list": [ { "name": "Alexander M Rush", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Associate Professor" }, { "name": "Fei Yuan", "gender": "Female", "institution": "Shanghai Artificial Intelligent Laboratory", "country": "", "position": "Researcher" }, { "name": "Jing Nathan Yan", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Siyu Ren", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Songlin Yang", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Wenting Zhao", "gender": "unknown", "institution": "Cornell University", "country": "", "position": "PhD student" }, { "name": "Yi Lu", "gender": "Male", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Zhiyong Wu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Justin Chiu", "gender": "Male", "institution": "Cornell University", "country": "", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 31, "n_ref": 91, "n_ref_all": 115, "n_fig": 3, "n_tab": 19, "L_tab": 10312, "n_element_tab": 1867, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 2480, "n_element_tab_1": 437, "formula_len_all": 1086, "formula_len_all_1": 869, "len_all": 183325, "len_all_1": 67031, "len_abs": 1265, "len_title": 120, "len_sents": 43672, "len_sents_1": 26813, "n_sents": 362, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1280, "title": "A Controlled Study on Long Context Extension and Generalization in LLMs", "abs": "Broad textual understanding and in-context learning require language models that utilize full document contexts. Due to the implementation challenges associated with directly training long-context models, many methods have been proposed for extending models to handle long contexts. However, owing to differences in data and model classes, it has been challenging to compare these approaches, leading to uncertainty as to how to evaluate long-context performance and whether it differs from standard evaluation. We implement a controlled protocol for extension methods with a standardized evaluation, utilizing consistent base models and extension data. Our study yields several insights into long-context behavior. First, we reaffirm the critical role of perplexity as a general-purpose performance indicator even in longer-context tasks. Second, we find that current approximate attention methods systematically underperform across long-context tasks. Finally, we confirm that exact fine-tuning based methods are generally effective within their extension range, whereas extrapolation remains challenging. All codebases, models, and checkpoints will be made available open-source, promoting transparency and facilitating further research in this critical area of AI development.", "keywords": [ "Controlled Study", "Long Context", "Extension", "Benchmark", "Analysis" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "VjeT8VFhHo", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Andr\u00e9 Biedenkapp", "gender": "unknown", "institution": "University of Freiburg, Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "", "position": "Postdoc" }, { "name": "Fabio Ferreira", "gender": "Male", "institution": "Universit\u00e4t Freiburg", "country": "", "position": "PhD student" }, { "name": "Frank Hutter", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen & University of Freiburg", "country": "DE", "position": "Full Professor" }, { "name": "Moreno Schlageter", "gender": "Male", "institution": "CS Department, University of Freiburg, Germany, Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "", "position": "MS student" }, { "name": "Raghu Rajan", "gender": "Male", "institution": "Universit\u00e4t Freiburg", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 23, "n_ref": 36, "n_ref_all": 62, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 218, "n_element_tab_1": 21, "formula_len_all": 401, "formula_len_all_1": 336, "len_all": 102381, "len_all_1": 75091, "len_abs": 1446, "len_title": 118, "len_sents": 39224, "len_sents_1": 28247, "n_sents": 278, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1404, "title": "One-shot World Models Using a Transformer Trained on a Synthetic Prior", "abs": "A World Model is a compressed spatial and temporal representation of a real world environment that allows one to train an agent or execute planning methods. However, world models are typically trained on observations from the real world environment, and they usually do not enable learning policies for other real environments. We propose One-Shot World Model (OSWM), a transformer world model that is learned in an in-context learning fashion from purely synthetic data sampled from a prior distribution. Our prior is composed of multiple randomly initialized neural networks, where each network models the dynamics of each state and reward dimension of a desired target environment. We adopt the supervised learning procedure of Prior-Fitted Networks by masking next-state and reward at random context positions and query OSWM to make probabilistic predictions based on the remaining transition context. During inference time, OSWM is able to quickly adapt to the dynamics of a simple grid world, as well as the CartPole gym and a custom control environment by providing 1k transition steps as context and is then able to successfully train environment-solving agent policies. However, transferring to more complex environments remains a challenge, currently. Despite these limitations, we see this work as an important stepping-stone in the pursuit of learning world models purely from synthetic data.", "keywords": [ "World Models", "Synthetic Pretraining", "Reinforcement Learning", "In-Context Learning" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 3, 2, 2, 4 ], "presentation_list": [ 1, 2, 2, 4 ], "contribution_list": [ 2, 1, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "VjHOGqHC4I", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Kaleel Mahmood", "gender": "Male", "institution": "University of Rhode Island", "country": "US", "position": "Assistant Professor" }, { "name": "Shaoyi Huang", "gender": "Female", "institution": "Stevens Institute of Technology", "country": "US", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 13, "n_ref_uni": 13, "n_ref": 16, "n_ref_all": 38, "n_fig": 9, "n_tab": 5, "L_tab": 1371, "n_element_tab": 135, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 553, "n_element_tab_1": 59, "formula_len_all": 3039, "formula_len_all_1": 2260, "len_all": 107296, "len_all_1": 60149, "len_abs": 1533, "len_title": 138, "len_sents": 32878, "len_sents_1": 26446, "n_sents": 267, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1625, "title": "Enhanced Long LoRA Inspired Perceiver Architectures for Auto-Regressive Language Modeling", "abs": "The Transformer architecture has revolutionized the Natural Language Processing field and is the backbone of Large Language Models (LLMs). The Transformer uses the attention mechanism that computes the pair-wise similarity between its input tokens to produce latent vectors that are able to understand the semantic meaning of the input text. One of the challenges in the Transformer architecture is the quadratic complexity of the attention mechanism that prohibits the efficient processing of long sequence lengths. While many recent research works have attempted to provide a reduction from $O(n^2)$ time complexity of attention to semi-linear complexity, it remains an unsolved problem in the sense of maintaining a high performance when such complexity is reduced. One of the important works in this respect is the Perceiver class of architectures that have demonstrated excellent performance while reducing the computation complexity. In this paper, we use the PerceiverAR that was proposed for Auto-Regressive modeling as a baseline, and provide three different architectural enhancements to it with varying computation overhead tradeoffs. Inspired by the recently proposed efficient attention computation approach of Long-LoRA, we then present an equally efficient Perceiver-based architecture (termed as Long LoRA Pereceiver - LLP) that can be used as the base architecture in LLMs instead of just a fine-tuning add-on. Our results on different benchmarks indicate much improved performance over the baseline PerceiverAR model, with the LLP showing impressive improvements compared to recent Transformer based models.", "keywords": [ "Large Language Models", "Perceiver", "Efficient Machine Learning" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 1, 3, 2, 1 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "VirIZC5v7E", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Jian Liang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Kuangpu Guo", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Lijun Sheng", "gender": "Male", "institution": "NLPR, CASIA", "country": "CN", "position": "Intern" }, { "name": "Ran He", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Tieniu Tan", "gender": "Male", "institution": " Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Professor" }, { "name": "Zilei Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 49, "n_ref": 104, "n_ref_all": 125, "n_fig": 4, "n_tab": 4, "L_tab": 1234, "n_element_tab": 243, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1958, "n_element_tab_1": 420, "formula_len_all": 342, "formula_len_all_1": 342, "len_all": 145585, "len_all_1": 72111, "len_abs": 1278, "len_title": 110, "len_sents": 35482, "len_sents_1": 29999, "n_sents": 254, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1284, "title": "Personalized Prompt Tuning for Unsupervised Federated Learning", "abs": "Federated learning facilitates collaborative model training across multiple distributed clients without requiring data sharing.\nHowever, conventional federated methods struggle with classification tasks in an unsupervised paradigm due to the absence of category knowledge.\nRecently, CLIP, a prominent visual language model, has demonstrated impressive results, particularly its remarkable zero-shot classification ability, which alleviates the dependence on labeled data.\nIn this paper, we first explore a new realistic problem, unsupervised federated learning using CLIP, where clients with unlabeled heterogeneous data collaborate to enhance global performance.\nTo address this problem, we propose FedPP, a method that incorporates a cooperative pseudo-label selection strategy and a partial prompt aggregation protocol.\nOur selection strategy ensures that all classes are trained in a balanced manner through global pseudo-label allocation.\nConcurrently, the aggregation protocol divides parameters into aggregated and retained components to optimize global performance while supporting local personalization.\nExtensive experiments across six datasets with various types of heterogeneity demonstrate the effectiveness of FedPP.\nOur code is available in the supplementary materials.", "keywords": [ "federated learning", "CLIP", "unsupervised learning" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "VipcVxaTnG", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Chenyang An", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Jingbo Shang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Letian Peng", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 37, "n_ref": 72, "n_ref_all": 108, "n_fig": 11, "n_tab": 11, "L_tab": 3328, "n_element_tab": 449, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1710, "n_element_tab_1": 84, "formula_len_all": 139, "formula_len_all_1": 219, "len_all": 167487, "len_all_1": 67140, "len_abs": 1878, "len_title": 136, "len_sents": 44114, "len_sents_1": 32299, "n_sents": 331, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1888, "title": "Correlation and Navigation in the Vocabulary Key Representation Space of Language Models", "abs": "Language model (LM) decoding is based on the next-token prediction (NTP) probability distribution. For neural LMs (e.g., Transformer-based), NTP distribution is\nessentially a softmax-regularized dot product between an encoded input context\n(query) and fixed vocabulary representations (keys). In this paper, we study the\neffect of the key distribution on the NTP distribution, with a focus on whether\nthe similarity between keys will trigger spurious correlations in NTP. Through\nknowledge-probing tasks, we show that in the NTP distribution, the few top-ranked\ntokens are typically accurate. However, the middle-ranked prediction is highly biased\ntowards the tokens that are distributionally (not necessarily semantically) similar to\nthese top ones. For instance, if \u201cP\u201d is predicted as the top-1 token, \u201cA\u201d-\u201cZ\u201d will all\nbe ranked high in NTP, no matter whether they can lead to correct decoding results.\nThis hurts the sampling diversity and makes the sampling of correct, long-tail\nresults hopeless and noisy. We attempt to alleviate this issue via a novel in-context\nmethod that iteratively pushes the query representation away from explored regions.\nSpecifically, we include the explored decoding results in the context and prompt\nthe LM to generate something else, which encourages the LM to produce a query\nrepresentation that has small dot products with explored keys. Experiments on\nknowledge-probing tasks show that our method leads to efficient navigation away\nfrom explored keys to correct new keys. We further extend our method to open-ended and chain-of-thought (for reasoning) generation. Experiment results show\nthat ICN contributes to better generation diversity and improved self-consistency\nvoting performance. Finally, we discuss potential training issues caused by the\nfixed key space together with the challenges and possible ways to address them in\nfuture research.", "keywords": [ "Language Modeling", "Next Token Prediction", "Spurious Correlation", "Generation Diversity" ], "rating_list": [ 8, 8, 5, 5 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "ViY73s6j6Q", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Bin Liang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Kam-Fai Wong", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "Full Professor" }, { "name": "Lin Gui", "gender": "Male", "institution": "King's College London, University of London", "country": "", "position": "Lecturer" }, { "name": "Qiwei Lang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Ruifeng Xu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Shenzhi Yang", "gender": "Male", "institution": "Suzhou University", "country": "CN", "position": "MS student" }, { "name": "Xiaofang Zhang", "gender": "Female", "institution": "Soochow University", "country": "", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 38, "n_ref": 96, "n_ref_all": 114, "n_fig": 6, "n_tab": 5, "L_tab": 4187, "n_element_tab": 190, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 4920, "n_element_tab_1": 201, "formula_len_all": 1703, "formula_len_all_1": 930, "len_all": 141282, "len_all_1": 67896, "len_abs": 1362, "len_title": 63, "len_sents": 39316, "len_sents_1": 28598, "n_sents": 310, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1367, "title": "Enhance Graph Contrastive Learning with Perturbation Discrimination", "abs": "Self-supervised learning of graph-structured data aims to produce transferable and robust representations that could be transferred to the downstream tasks. Among many, graph contrastive learning (GCL) based on data augmentation has emerged with promising performance in learning graph representation. However, it is observed that some augmentations might change the graph semantics due to the perturbations in the graph structure such as perturbing some nodes/edges. In such cases, existing GCL methods may suffer from performance limitations due to the introduction of noise augmentations. To address this issue, we propose to train a discriminative model to enhance GCL for graph-structured data, called Perturbation Discrimination-Enhanced GCL (PerEG). Specifically, for each perturbed graph, the discriminative model is trained to predict whether each node in the augmentation was perturbed by the perturbation compared to the original graph or not. Based on this, the results of perturbation discrimination are exploited to refine the GCL, enabling its controllable use of augmentation, thereby preferably utilizing augmentation and effectively avoiding the introduction of noise augmentation. Extensive experiments in unsupervised, semi-supervised, and transfer learning scenarios show that our PerEG outperforms the state-of-the-art methods on eight datasets.", "keywords": [ "Graph Representation", "Contrastive Learning", "Predictive Learning", "Data Augmentation" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 5 ] }, { "paper_id": "ViRDmDAfjg", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Amit Sharma", "gender": "Male", "institution": "Microsoft Research", "country": "IN", "position": "Principal Researcher" }, { "name": "Gautam Jajoo", "gender": "Male", "institution": "BITS Pilani, Birla Institute of Technology and Science", "country": "IN", "position": "Undergrad student" }, { "name": "Hua Li", "gender": "Not Specified", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Jian Jiao", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Nagarajan Natarajan", "gender": "Male", "institution": "Microsoft Research", "country": "", "position": "Researcher" }, { "name": "GURUSHA JUNEJA", "gender": "Female", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 33, "n_ref": 68, "n_ref_all": 87, "n_fig": 5, "n_tab": 8, "L_tab": 3721, "n_element_tab": 194, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 225, "n_element_tab_1": 18, "formula_len_all": 434, "formula_len_all_1": 383, "len_all": 201137, "len_all_1": 63451, "len_abs": 1459, "len_title": 113, "len_sents": 57516, "len_sents_1": 31790, "n_sents": 421, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1465, "title": "Task Facet Learning: A Structured Approach to Prompt Optimization", "abs": "Given a task in the form of a basic description and its training examples, prompt optimization is the problem of synthesizing the given information into a text prompt for a large language model. Humans solve this problem by also considering the different facets that define a task (e.g., counter-examples, explanations, analogies) and including them in the prompt. However, it is unclear whether existing algorithmic approaches, based on iteratively editing a given prompt or automatically selecting a few in-context examples, can cover the multiple facets required to solve a complex task. In this work, we view prompt optimization as that of learning multiple facets of a task from a set of training examples. We exploit structure in the prompt optimization problem and break down a prompt into loosely coupled semantic sections. The proposed algorithm, UniPrompt, (1) clusters the input space and uses clustered batches so that each batch likely corresponds to a different facet of the task, and (2) utilizes a feedback mechanism to propose adding, editing or deleting a section, which in turn is aggregated over a batch to capture generalizable facets. Empirical evaluation on multiple datasets and a real-world task shows that prompts generated using UniPrompt obtain higher accuracy than human-tuned prompts and those from state-of-the-art methods. In particular, our algorithm can generate long, complex prompts that existing methods are unable to generate.", "keywords": [ "prompt optimization", "large language model" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "Vi6p2TeujL", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Dan dan Guo", "gender": "Female", "institution": "Jilin University", "country": "CN", "position": "Lecturer" }, { "name": "Jinhan Liu", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "Ruiying Lu", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 20, "n_ref_uni": 56, "n_ref": 95, "n_ref_all": 136, "n_fig": 14, "n_tab": 19, "L_tab": 16197, "n_element_tab": 1924, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 968, "n_element_tab_1": 76, "formula_len_all": 1759, "formula_len_all_1": 1224, "len_all": 222763, "len_all_1": 76481, "len_abs": 1582, "len_title": 116, "len_sents": 52525, "len_sents_1": 34549, "n_sents": 362, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1588, "title": "PTAD: Prototype-Oriented Tabular Anomaly Detection via Mask Modeling", "abs": "Tabular anomaly detection, which aims at identifying deviant samples, has been crucial in a variety of real-world applications, such as medical disease identification, financial fraud detection, intrusion monitoring, etc. Although recent deep learning-based methods have achieved competitive performances, these methods suffer from representation entanglement and the lack of global correlation modeling, which leads to the 'abnormal leakage' issue and hinders anomaly detection performance. To tackle the problem, we incorporate mask modeling and prototype learning into tabular anomaly detection. The core idea is to design learnable masks by disentangled representation learning within a projection space and extracting nominal dependencies as explicit global prototypes. Specifically, the overall model involves two parts: (i) During encoding, we perform mask modeling in both the data space and projection space with orthogonal basis vectors for masking out the suspicious abnormal locations; (ii) During decoding, we decode multiple masked representations in parallel for reconstruction and learn association prototypes to extract nominal characteristic correlations. Our proposal derives from a distribution-matching perspective, where both projection space learning and association prototype learning are formulated as optimal transport problems, and the calibration distances are utilized to refine the anomaly scores. By conducting both quantitative and qualitative experiments on 20 tabular benchmarks, our model surpasses other competitors and possesses good interpretability.", "keywords": [ "Tabular anomaly detection", "Prototype learning", "Mask modeling" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 5, 2, 4, 2 ] }, { "paper_id": "Vi1PJjvEdh", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Jindong Wang", "gender": "Male", "institution": "William & Mary", "country": "US", "position": "Assistant Professor" }, { "name": "Lai Wei", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Weiran Huang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xin Xie", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Zhiquan Tan", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 24, "n_ref": 42, "n_ref_all": 77, "n_fig": 4, "n_tab": 7, "L_tab": 2042, "n_element_tab": 251, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1781, "n_element_tab_1": 201, "formula_len_all": 184, "formula_len_all_1": 183, "len_all": 137518, "len_all_1": 63744, "len_abs": 877, "len_title": 130, "len_sents": 39080, "len_sents_1": 28368, "n_sents": 321, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 879, "title": "Can I Understand What I Create? Self-Knowledge Evaluation of Large Language Models", "abs": "Large language models (LLMs) have achieved remarkable progress in linguistic tasks, necessitating robust evaluation frameworks to understand their capabilities and limitations. Inspired by Feynman's principle of understanding through creation, we introduce a self-knowledge evaluation framework that is easy to implement, evaluating models on their ability to comprehend and respond to self-generated questions. Our findings, based on testing multiple models across diverse tasks, reveal significant gaps in the model's self-knowledge ability. Further analysis indicates these gaps may be due to misalignment with human attention mechanisms. Additionally, fine-tuning on self-generated math task may enhance the model's math performance, highlighting the potential of the framework for efficient and insightful model evaluation and may also contribute to the improvement of LLMs.", "keywords": [ "Evaluation", "Self-knowledge" ], "rating_list": [ 3, 3, 1, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 2, 1, 1, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "VhQelEo27A", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jiwan Seo", "gender": "Male", "institution": "Daegu Gyeongbuk Institute of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Kiljoon Han", "gender": "Male", "institution": "Daegu Gyeongbuk Institute of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Minwoo Choi", "gender": "unknown", "institution": "Daegu Gyeongbuk Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Seunghun Lee", "gender": "Male", "institution": "Daegu Gyeongbuk Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sunghoon Im", "gender": "Male", "institution": "Daegu Gyeongbuk Institute of Science and Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 19, "n_ref_uni": 46, "n_ref": 133, "n_ref_all": 151, "n_fig": 9, "n_tab": 9, "L_tab": 4712, "n_element_tab": 494, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 7797, "n_element_tab_1": 498, "formula_len_all": 1147, "formula_len_all_1": 1143, "len_all": 164303, "len_all_1": 62923, "len_abs": 940, "len_title": 89, "len_sents": 39451, "len_sents_1": 21864, "n_sents": 286, "n_sents_1": 148, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 911, "title": "Context-Aware Video Instance Segmentation", "abs": "In this paper, we introduce the Context-Aware Video Instance Segmentation (CAVIS), a novel framework designed to enhance instance association by integrating contextual information adjacent to each object. To efficiently extract and leverage this information, we propose the Context-Aware Instance Tracker (CAIT), which merges contextual data surrounding the instances with the core instance features to improve tracking accuracy. Additionally, we introduce the Prototypical Cross-frame Contrastive (PCC) loss, which ensures consistency in object-level features across frames, thereby significantly enhancing instance matching accuracy. CAVIS demonstrates superior performance over state-of-the-art methods on all benchmark datasets in video instance segmentation (VIS) and video panoptic segmentation (VPS). Notably, our method excels on the OVIS dataset, which is known for its particularly challenging videos.", "keywords": [ "Video Instance Segmentation", "Contrastive Learning", "Instance Prototype", "Context-aware Learning" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "VgtpRXhxli", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Binyamin Perets", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "PhD student" }, { "name": "Mark Kozdoba", "gender": "unknown", "institution": "Technion, Technion", "country": "IL", "position": "Principal Researcher" }, { "name": "Shie Mannor", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Full Professor" } ], "n_formula": 85, "n_formula_1": 28, "n_ref_uni": 32, "n_ref": 59, "n_ref_all": 96, "n_fig": 6, "n_tab": 1, "L_tab": 219, "n_element_tab": 14, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7386, "formula_len_all_1": 2256, "len_all": 174640, "len_all_1": 68853, "len_abs": 1397, "len_title": 103, "len_sents": 59021, "len_sents_1": 31885, "n_sents": 537, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1411, "title": "Efficient Fairness-Performance Pareto Front Computation", "abs": "There is a well known intrinsic trade-off between the fairness of a representation and the performance of classifiers derived from the representation. \nDue to the complexity of optimisation algorithms in most modern representation learning approaches, for a given method it may be non-trivial to decide whether the obtained fairness-performance curve of the method is optimal, i.e., whether it is close to the true Pareto front for these quantities for the underlying data distribution. \n\nIn this paper we propose a new method to compute the optimal Pareto front, \nwhich does not require the training of complex representation models. We show that optimal fair representations possess several useful structural properties, and that these properties enable a reduction of the computation of the Pareto Front to a compact discrete problem. We then also show that these compact approximating problems can be efficiently solved via off-the shelf concave-convex programming methods. Finally, in addition to representations, we show that the new methods may also be used to directly compute the Pareto front of fair classification problems. \n\nSince our approach is independent of the specific model of representations, it may be used as the benchmark to which representation learning algorithms, or classifiers, may be compared. We experimentally evaluate the approach on a number of real world benchmark datasets.", "keywords": [ "Fairness", "fair Representations", "Fairness-Performance Pareto Front", "Pareto Front", "convex concave optimization" ], "rating_list": [ 5, 5, 8, 6, 6 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 3, 2 ] }, { "paper_id": "VgmvKk7yfE", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 10, "author_info_list": [ { "name": "Alexander Aufreiter", "gender": "Male", "institution": "Software Competence Center Hagenberg", "country": "AT", "position": "Researcher" }, { "name": "Bernhard Nessler", "gender": "Male", "institution": "Software Competence Center Hagenberg", "country": "AT", "position": "Principal Researcher" }, { "name": "Felix Nessler", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Undergrad student" }, { "name": "Gregor Aichinger", "gender": "unknown", "institution": "Software Competence Center Hagenberg", "country": "AT", "position": "Researcher" }, { "name": "Michal Lewandowski", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "PhD student" }, { "name": "Patrick Mederitsch", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "PhD student" }, { "name": "Severin Bergsmann", "gender": "Male", "institution": "Software Competence Center Hagenberg", "country": "AT", "position": "Researcher" }, { "name": "Simon Lucas Schmid", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "PhD student" }, { "name": "Tobias Halmdienst", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Undergrad student" }, { "name": "Viktor Szolga", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Undergrad student" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 33, "n_ref": 42, "n_ref_all": 76, "n_fig": 16, "n_tab": 2, "L_tab": 870, "n_element_tab": 74, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 195, "n_element_tab_1": 26, "formula_len_all": 548, "formula_len_all_1": 413, "len_all": 139209, "len_all_1": 57091, "len_abs": 1048, "len_title": 63, "len_sents": 49087, "len_sents_1": 27879, "n_sents": 418, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 15, "L_abs": 1061, "title": "The Turing Game", "abs": "We present first experimental results from the \\textit{Turing Game}, a modern implementation of the original imitation game as proposed by Alan Turing in 1950. The Turing Game is a gamified interaction between two human players and one AI chatbot powered by state-of-the-art Large Language Models (LLMs). The game is designed to explore whether humans can distinguish between their peers and machines in chat-based conversations, with human players striving to identify fellow humans and machines striving to blend in as one of them. To this end, we implemented a comprehensive framework that connects human players over the Internet with chatbot implementations. We detail the experimental results after a public launch at the Ars Electronica Festival in September 2024. While the experiment is still ongoing, in this paper we present our initial findings from the hitherto gathered data.\nOur long term vision of the project is to deepen the understanding of human-AI interactions and eventually contribute to improving LLMs and language-based user interfaces.", "keywords": [ "Turing Test", "LLMs", "Chatbots", "Testing", "Human-Ai Interface" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "VfvxZLXYgd", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Binh-Son Hua", "gender": "Male", "institution": "University of Dublin, Trinity College", "country": "IE", "position": "Assistant Professor" }, { "name": "Duong Minh Nguyen", "gender": "Male", "institution": "Pusan National University", "country": "KR", "position": "PhD student" }, { "name": "Khoi Hoang Do", "gender": "Male", "institution": "University of Dublin, Trinity College", "country": "IE", "position": "PhD student" }, { "name": "Viet Quoc Pham", "gender": "Male", "institution": "University of Dublin, Trinity College", "country": "IE", "position": "Assistant Professor" }, { "name": "Won-Joo Hwang", "gender": "unknown", "institution": "Pusan National University", "country": "KR", "position": "Full Professor" }, { "name": "L\u00ea Nam Kh\u00e1nh", "gender": "Male", "institution": "Hanoi University of Science and Technology", "country": "VN", "position": "Undergrad student" } ], "n_formula": 57, "n_formula_1": 21, "n_ref_uni": 45, "n_ref": 88, "n_ref_all": 124, "n_fig": 16, "n_tab": 4, "L_tab": 4879, "n_element_tab": 329, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 2485, "n_element_tab_1": 168, "formula_len_all": 4122, "formula_len_all_1": 1714, "len_all": 156870, "len_all_1": 59837, "len_abs": 2017, "len_title": 106, "len_sents": 42642, "len_sents_1": 22903, "n_sents": 379, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1329, "title": "DOMAIN GENERALIZATION VIA PARETO OPTIMAL GRADIENT MATCHING", "abs": "In this study, we address the gradient-based domain generalization problem, where predictors aim for consistent gradient directions across different domains. Existing methods have two main challenges. First, minimization of gradient empirical distance or gradient inner products (GIP) leads to gradient fluctuations and magnitude elimination among domains, thereby hindering straightforward learning. Second, the direct application of gradient learning to joint loss function can incur high computation overheads due to second-order derivative approximation. To tackle these challenges, we propose a new Pareto Optimality Gradient Matching (POGM) method. In contrast to existing methods that add gradient matching as regularization, we leverage gradient trajectories as collected data and apply independent training at the meta-learner. In the meta-update, we maximize GIP while limiting the learned gradient from deviating too far from the empirical risk minimization gradient trajectory. By doing so, the aggregate gradient can incorporate knowledge from all domains without suffering gradient magnitude elimination or fluctuation\ntowards any particular domain. Experimental evaluations on datasets from DomainBed demonstrate competitive results yielded by POGM against other baselines while achieving computational efficiency.", "keywords": [ "Domain Generalization" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "VfYShlQbj7", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Daniel Winter", "gender": "Male", "institution": "Google", "country": "IL", "position": "Intern" }, { "name": "Niv Cohen", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "Yedid Hoshen", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 32, "n_ref": 67, "n_ref_all": 84, "n_fig": 8, "n_tab": 12, "L_tab": 4171, "n_element_tab": 408, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1151, "n_element_tab_1": 119, "formula_len_all": 1334, "formula_len_all_1": 1107, "len_all": 153886, "len_all_1": 62290, "len_abs": 741, "len_title": 98, "len_sents": 46734, "len_sents_1": 27458, "n_sents": 460, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 745, "title": "Demystifying GNN Distillation by Replacing the GNN", "abs": "It has recently emerged that Multilayer Perceptrons (MLPs) can achieve excellent performance on graph node classification, but only if they distill a previously-trained Graph Neural Network (GNN). This finding is confusing; if MLPs are expressive enough to perform node classification, what is the role of the GNNs? This paper aims to answer this question. Rather than suggesting a new technique, we aim to demystify GNN distillation methods. Through our analysis, we identify the key properties of GNNs that enable them to serve as effective regularizers, thereby overcoming limited training data. We validate our analysis by demonstrating an MLP training process that successfully leverages GNN-like properties without actually training a GNN.", "keywords": [ "GNN", "Graph Neural Network" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "Vf6RDObyEF", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Chetan Bansal", "gender": "Not Specified", "institution": "Birla Institute of Technology and Science", "country": "", "position": "Undergrad student" }, { "name": "Huaxiu Yao", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Weilei He", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Intern" }, { "name": "Xuchao Zhang", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Ying Wei", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Zhaoyang Wang", "gender": "unknown", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Zhiyuan Liang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Intern" }, { "name": "Weitong ZHANG", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" } ], "n_formula": 49, "n_formula_1": 19, "n_ref_uni": 48, "n_ref": 100, "n_ref_all": 124, "n_fig": 5, "n_tab": 10, "L_tab": 4138, "n_element_tab": 408, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2313, "n_element_tab_1": 107, "formula_len_all": 7092, "formula_len_all_1": 1894, "len_all": 200547, "len_all_1": 70829, "len_abs": 1743, "len_title": 109, "len_sents": 55540, "len_sents_1": 30747, "n_sents": 458, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1750, "title": "CREAM: Consistency Regularized Self-Rewarding Language Models", "abs": "Recent self-rewarding large language models (LLM) have successfully applied LLM-as-a-Judge to iteratively improve the alignment performance without the need of human annotations for preference data. These methods commonly utilize the same LLM to act as both the policy model (which generates responses) and the reward model (which scores and ranks those responses). The ranked responses are then used as preference pairs to train the LLM via direct alignment technologies (e.g. DPO). However, it is noteworthy that throughout this process, there is no guarantee on the accurate of the rewarding and ranking, which is critical for ensuring accurate rewards and high-quality preference data. Empirical results from relatively small LLMs (e.g., 7B parameters) also indicate that improvements from self-rewarding may diminish after several iterations in certain situations, which we hypothesize is due to accumulated bias in the reward system. This bias can lead to unreliable preference data for training the LLM. To address this issue, we first formulate and analyze the generalized iterative preference fine-tuning frame work for self-rewarding language model. We then introduce the regularization to this generalized framework to mitigate the overconfident preference labeling in the self-rewarding process. Based on this theoretical insight, we propose a Consistency Regularized sElf-rewarding lAnguage Model (CREAM) that leverages the rewarding consistency across different iterations to regularize the self-rewarding training, helping the model to learn from more reliable preference data. With this explicit regularization, our empirical results demonstrate the superiority of CREAM in improving both reward consistency and alignment performance.", "keywords": [ "alignment", "self-rewarding", "large language model" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "Vf5ZUalFk8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Eric Frankel", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "", "position": "PhD student" }, { "name": "Pang Wei Koh", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Visiting Research Scientist" }, { "name": "Shuyue Stella Li", "gender": "Female", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Yulia Tsvetkov", "gender": "Female", "institution": "Department of Computer Science, University of Washington", "country": "", "position": "Associate Professor" }, { "name": "Lillian Ratliff", "gender": "Female", "institution": "University of Washington, Seattle", "country": "US", "position": "Associate Professor" }, { "name": "Sewoong Oh", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 69, "n_ref_all": 91, "n_fig": 10, "n_tab": 3, "L_tab": 2211, "n_element_tab": 257, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 377, "n_element_tab_1": 26, "formula_len_all": 697, "formula_len_all_1": 601, "len_all": 197594, "len_all_1": 70454, "len_abs": 1272, "len_title": 119, "len_sents": 49866, "len_sents_1": 35402, "n_sents": 334, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1103, "title": "Conformal Reasoning: Uncertainty Estimation in Interactive Environments", "abs": "We introduce conformal reasoning, a principled method for models in interactive environments to reason about their uncertainty and decide whether to seek out more information or to return a prediction. The challenge with standard conformal prediction---a popular statistical framework for uncertainty estimation that constructs prediction sets with formal coverage guarantees---is that it relies on a fixed set of calibration data points. In interactive environments, however, the calibration trajectories require certain termination criteria determined a priori, introducing heuristic bias and/or circular dependency that break the assumptions needed for coverage guarantees. We address this issue by building on adaptive conformal inference techniques. On two real-world tasks on medical diagnosis and embodied question answering, we show that conformal reasoning empirically achieves its theoretical coverage guarantees---in contrast with standard conformal prediction approaches that can significantly over- or under-cover---while improving exploration efficiency by approximately 20% on both tasks.", "keywords": [ "Conformal Prediction", "Uncertainty Quantification", "LM Abstention", "Interactive Reasoning" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 2, 3, 4 ] }, { "paper_id": "VejUqXsDYa", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Chunwei Wang", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Hang Xu", "gender": "Male", "institution": "Huawei Noah\u2018s Ark Lab", "country": "HK", "position": "Researcher" }, { "name": "Hengshuang Zhao", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Jianhua Han", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Runhui Huang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Wei Zhang", "gender": "Male", "institution": "Huawei Technologies Ltd", "country": "CN", "position": "Researcher" }, { "name": "Xiaodan Liang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Xinpeng Ding", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yulong Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "LU HOU", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 50, "n_ref": 132, "n_ref_all": 162, "n_fig": 13, "n_tab": 8, "L_tab": 3160, "n_element_tab": 587, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 3177, "n_element_tab_1": 168, "formula_len_all": 317, "formula_len_all_1": 317, "len_all": 187350, "len_all_1": 58197, "len_abs": 1834, "len_title": 138, "len_sents": 38872, "len_sents_1": 22542, "n_sents": 302, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1703, "title": "HiRes-LLaVA: Restoring Fragmentation Input in High-Resolution Large Vision-Language Models", "abs": "High-resolution image inputs allow Large Vision-Language Models (LVLMs) to capture finer visual details, improving comprehension. However, the increased training and computational costs associated with such inputs pose significant challenges. A common approach to mitigate these costs involves slicing the input into uniform patches using sliding windows, each aligned with the vision encoder\u2019s input size. While efficient, this method fragments the input, disrupting the continuity of contextual, which negatively impacts cross-patch perception\ntasks. To address these limitations, we propose **HiRes-LLaVA**, a novel framework designed to efficiently process high-resolution inputs of any size without altering the original contextual and geometric information. HiRes-LLaVA introduces two key components: (i) a SliceRestore adapter (SRA) that reconstructs sliced patches into their original form, enabling efficient extraction of both global and local\nfeatures through down-up-sampling and convolutional layers, and (ii) a Self-Mining Sampler (SMS) that compresses vision tokens based on internal relationships, preserving original context and positional information while reducing training overhead. To assess the ability of handling context fragmentation, we construct a new benchmark, EntityGrid-QA, consisting of edge-related tasks. Extensive experiments demonstrate the superiority of HiRes-LLaVA on both existing public benchmarks and EntityGrid-QA. For example, with SRA, our method achieves a performance improvement of \u223c 9% over state-of-the-art LVLMs in addressing fragmentation issues. Additionally, our SMS outperforms other visual token downsamplers, while offering comparable efficiency.", "keywords": [ "Large Vision Language Model", "High Resolution Image Input", "Adapter" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "VeSsiD0DP9", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Chunyan Miao", "gender": "Female", "institution": "School of Computer Science and Engineering, Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Deli Zhao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Director" }, { "name": "Hang Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Lidong Bing", "gender": "unknown", "institution": "Alibaba Group", "country": "SG", "position": "Scientist" }, { "name": "Shijian Lu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Sicong Leng", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xin Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yang Zhou", "gender": "Male", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Yun Xing", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Zesen Cheng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 40, "n_ref": 110, "n_ref_all": 132, "n_fig": 10, "n_tab": 2, "L_tab": 378, "n_element_tab": 4, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 753, "n_element_tab_1": 154, "formula_len_all": 96, "formula_len_all_1": 96, "len_all": 167192, "len_all_1": 66030, "len_abs": 1317, "len_title": 166, "len_sents": 47406, "len_sents_1": 31065, "n_sents": 316, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 118, "L_abs": 1340, "title": "The Curse of Multi-Modalities: Evaluating Hallucinations of Large Multimodal Models across Language, Visual, and Audio", "abs": "Recent advancements in large multimodal models (LMMs) have significantly enhanced performance across diverse tasks, with ongoing efforts to further integrate additional modalities such as video and audio. However, most existing LMMs remain vulnerable to hallucinations, the discrepancy between the factual multimodal input and the generated textual output, which has limited their applicability in various real-world scenarios. This paper presents the first systematic investigation of hallucinations in LMMs involving the three most common modalities: language, visual, and audio. Our study reveals two key contributors to hallucinations: overreliance on unimodal priors and spurious inter-modality correlations. To address these challenges, we introduce the benchmark \\textit{The Curse of Multi-Modalities} (\\textbf{CMM}), which comprehensively evaluates hallucinations in LMMs, providing a detailed analysis of their underlying issues. Our findings highlight key vulnerabilities, including imbalances in modality integration and biases from training data, underscoring the need for balanced cross-modal learning and enhanced hallucination mitigation strategies. Based on our observations and findings, we suggest potential research directions that could enhance the reliability of LMMs. We will make our code and data publicly available.", "keywords": [ "Mutimodal", "Large Multimodal Models", "Hallucinations", "Vision-Language", "Audio-Language", "Vision-Audio-Language" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "VeMC6Bn0ZB", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Ferdinando Fioretto", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "", "position": "Assistant Professor" }, { "name": "Kyri Baker", "gender": "Female", "institution": "University of Colorado at Boulder", "country": "US", "position": "Assistant Professor" }, { "name": "Mostafa Mohammadian", "gender": "Male", "institution": "University of Colorado at Boulder", "country": "", "position": "PhD student" }, { "name": "Vincenzo Di Vito Francesco", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" } ], "n_formula": 41, "n_formula_1": 22, "n_ref_uni": 33, "n_ref": 60, "n_ref_all": 109, "n_fig": 8, "n_tab": 6, "L_tab": 2036, "n_element_tab": 350, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 719, "n_element_tab_1": 125, "formula_len_all": 4010, "formula_len_all_1": 1579, "len_all": 175192, "len_all_1": 70988, "len_abs": 1267, "len_title": 121, "len_sents": 64165, "len_sents_1": 33112, "n_sents": 426, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 73, "L_abs": 1275, "title": "Learning to Solve Differential Equation Constrained Optimization Problems", "abs": "Differential equations (DE) constrained optimization plays a critical role in numerous scientific and engineering fields, including energy systems, aerospace engineering, ecology, and finance, where optimal configurations or control strategies must be determined for systems governed by ordinary or stochastic differential equations. Despite its significance, the computational challenges associated with these problems have limited their practical use. To address these limitations, this paper introduces a learning-based approach to DE-constrained optimization that combines techniques from proxy optimization \\citep{kotary2021end} and neural differential equations \\citep{chen2019neural}. The proposed approach uses a dual-network architecture, with one approximating the control strategies, focusing on steady-state constraints, and another solving the associated DEs. \nThis combination enables the approximation of optimal strategies while accounting for dynamic constraints in near real-time.\nExperiments across problems in energy optimization and finance modeling show that this method provides full compliance with dynamic constraints and it produces results up to 25 times more precise than other methods which do not explicitly model the system's dynamic equations.", "keywords": [ "Learning-based optimization proxy", "differential equations constrained optimization", "neural differential equations", "system dynamics" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 2, 2, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "VdX9tL3VXH", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 13, "author_info_list": [ { "name": "Alexander Seifert", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Undergrad student" }, { "name": "Janik Sielemann", "gender": "unknown", "institution": "BioNTech", "country": "DE", "position": "Researcher" }, { "name": "Juan Jose Garau Luis", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Karim Beguir", "gender": "Male", "institution": "NYU, New York University", "country": "", "position": "MS student" }, { "name": "Lee Zamparo", "gender": "Male", "institution": "InstaDeep", "country": "US", "position": "Researcher" }, { "name": "Maren Lang", "gender": "Female", "institution": "Biontech SE", "country": "DE", "position": "Principal Researcher" }, { "name": "Martin Brand", "gender": "Male", "institution": "BioNTech", "country": "DE", "position": "PhD student" }, { "name": "Michail Chatzianastasis", "gender": "Male", "institution": "InstaDeep", "country": "FR", "position": "Intern" }, { "name": "Priyanka Pandey", "gender": "unknown", "institution": "BioNTech", "country": "DE", "position": "Principal Researcher" }, { "name": "Raphael Boige", "gender": "Male", "institution": "InstaDeep", "country": "", "position": "Researcher" }, { "name": "Thomas PIERROT", "gender": "Male", "institution": "InstaDeep", "country": "US", "position": "Principal Researcher" }, { "name": "Ugo Tanielian", "gender": "Male", "institution": "Computer Science Lab - Pierre and Marie Curie University, Paris, France", "country": "FR", "position": "PhD student" }, { "name": "Ameya A Joshi", "gender": "Male", "institution": "InstaDeep", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 47, "n_ref": 93, "n_ref_all": 118, "n_fig": 14, "n_tab": 9, "L_tab": 3369, "n_element_tab": 339, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1556, "n_element_tab_1": 58, "formula_len_all": 188, "formula_len_all_1": 55, "len_all": 176602, "len_all_1": 63955, "len_abs": 1423, "len_title": 152, "len_sents": 45272, "len_sents_1": 30854, "n_sents": 349, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1428, "title": "A long range foundation model for zero-shot predictions in single-cell and spatial transcriptomics data", "abs": "Large transformers pre-trained with language model objectives have demonstrated success in multiple fields, and have tremendous potential for modeling single-cell RNA-seq and spatial transcriptomics data. However, these approaches are yet to overcome various challenges, including inductive biases that hinder generalization, artifacts and quality of the underlying data, as well as downstream evaluation pipelines that do not reflect the biological challenges in the field. In this work, we propose a new framework, sCellTransformer (sCT), that relies on a first principles formulation of the problem as well as a validation pipeline designed to evaluate models generalization through zero-shot predictions. sCT leverages a long-range convolutional-transformer architecture that is trained from unprocessed single-cell and spatial transcriptomics data. In contrast to previous works, sCT represents cells with up to 20,000 protein-coding genes, processes sets of multiple cells, and predicts about a million discretized gene expression tokens. We show that representing gene expression as discrete levels allows us to mitigate the high sparsity present in single-cell data both during training and evaluation. We present state-of-the-art empirical results on several zero-shot gene expression imputation, cell-typing, and clustering tasks in both single-cell as well as spatial domains, outperforming current foundation models.", "keywords": [ "Foundation models", "single-cell RNA-seq", "spatial transcriptomics", "masked language modelling", "computational biology", "zero shot inference" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "VdURgvImVn", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Dayal Singh Kalra", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Hong-Min Chu", "gender": "unknown", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "John Kirchenbauer", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Jonas Geiping", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen", "country": "DE", "position": "Principal Researcher" }, { "name": "Tom Goldstein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Arpit Amit Bansal", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 15, "n_ref_uni": 18, "n_ref": 35, "n_ref_all": 62, "n_fig": 10, "n_tab": 3, "L_tab": 849, "n_element_tab": 76, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 640, "n_element_tab_1": 70, "formula_len_all": 1726, "formula_len_all_1": 1029, "len_all": 97781, "len_all_1": 63304, "len_abs": 797, "len_title": 96, "len_sents": 32873, "len_sents_1": 28423, "n_sents": 273, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 47, "L_abs": 798, "title": "Gradient-based Optimization of Dataset Mixtures", "abs": "Modern state-of-the-art machine learning models are often trained using a combination of heterogeneous data sources. However, the utility of different data sources as support for learning some target tasks is often not equivalent, motivating the need for automated methods of optimizing the relative contribution of each data source to the model. In this work, we propose a dataset optimization strategy that slices a normal model training step into a series of data source-specific updates and splices them back together in an optimal manner with respect to the loss on some target task dataset. We demonstrate the effectiveness of our algorithm across different scenarios and domains, including classification problems for vision models and for next-token prediction tasks in the language domain.", "keywords": [ "dataset selection", "data valuation", "foundation models" ], "rating_list": [ 5, 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 2, 4, 4 ] }, { "paper_id": "VdDtRu7RTf", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Honglie Wang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Minsi Ren", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Intern" }, { "name": "Yan-Ming Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "YY Liu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 15, "n_ref_uni": 38, "n_ref": 44, "n_ref_all": 51, "n_fig": 3, "n_tab": 3, "L_tab": 620, "n_element_tab": 44, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 573, "n_element_tab_1": 40, "formula_len_all": 695, "formula_len_all_1": 696, "len_all": 84606, "len_all_1": 66786, "len_abs": 1347, "len_title": 132, "len_sents": 22125, "len_sents_1": 21521, "n_sents": 172, "n_sents_1": 168, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1362, "title": "Write More at Once: Stylized Chinese Handwriting Generation via Two-stage Diffusion", "abs": "Handwritten data generation is an intriguing research area with broad applications in human interaction with digital documents. In Chinese handwritten text generation, practical applications necessitate the ability to produce sentence-level handwritten data to convey complex information effectively. However, existing methods mainly focus on generating single-font outputs. To tackle this challenge, we model handwritten text generation as a \\textit{style transfer problem}, aiming to convert a standard text line template into a target handwriting style. Recognizing the highly structured nature of handwritten data, we view complex text lines as compositions of individual characters and their positions. We propose a two-stage text line generation method based on generative diffusion model. In the first stage, character positions are generated using a Character-Position-Diffusion (CharPos-Diff), which, combined with standard character templates from a digital library, creates text line-level templates. In the second stage, a font style transfer diffusion model (Imitating-Diff) generates handwritten text lines directly from these templates. Our extensive experiments show that our method effectively mimics handwriting styles, generates structurally accurate text lines, and facilitates the simultaneous generation of paragraph-level handwritten text.", "keywords": [ "Handwritten Text Generation; Conditional Diffusion;" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "VbszSB4pK6", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Bozheng Li", "gender": "Male", "institution": "Brown University", "country": "US", "position": "MS student" }, { "name": "Jun Dan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Mushui Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yunlong Yu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ziqian Lu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 15, "n_ref": 36, "n_ref_all": 58, "n_fig": 9, "n_tab": 12, "L_tab": 7263, "n_element_tab": 556, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2585, "n_element_tab_1": 177, "formula_len_all": 674, "formula_len_all_1": 675, "len_all": 136349, "len_all_1": 55428, "len_abs": 1174, "len_title": 138, "len_sents": 36769, "len_sents_1": 24132, "n_sents": 254, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1169, "title": "Fully Fine-Tuning Beats Parameter Efficient Fine-Tuning for CLIP in Data-Limited Scenarios", "abs": "Prompt tuning, which involves training a small set of parameters, effectively enhances the pre-trained Vision-Language Models (VLMs) to downstream tasks. However, they often come at the cost of flexibility and adaptability when the tuned models are applied to different datasets or domains. In this paper, we revisit the vanilla full fine-tuning in VLMs and show that fully fine-tuning can be more efficient than prompt tuning in data-limited scenarios. To mitigate the overfitting and catastrophic forgetting issues encountered when fine-tuning the entire VLMs for specific tasks under limited supervision, we propose a framework named CLIP-CITE via designing a discriminative visual-text task, further aligning the visual-text semantics in a supervision manner, and integrating knowledge distillation techniques to preserve the gained knowledge. Extensive experimental results under few-shot learning, base-to-new generalization, domain generalization, and cross-domain generalization settings, demonstrate that our method effectively enhances the performance on specific tasks under limited supervision while preserving the versatility of the VLMs on other datasets.", "keywords": [ "Vision-Language Models; Parameter-Efficient Fine-tuning;" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "VbkGysQ0Rl", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Alvin C Silva", "gender": "unknown", "institution": "Mayo Clinic", "country": "", "position": "Full Professor" }, { "name": "Marko Jojic", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Undergrad student" }, { "name": "Rajeev Goel", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Teresa Wu", "gender": "Female", "institution": "Arizona State University", "country": "", "position": "Full Professor" }, { "name": "Yancheng Wang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Yingzhen Yang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 27, "n_formula_1": 7, "n_ref_uni": 79, "n_ref": 255, "n_ref_all": 293, "n_fig": 5, "n_tab": 9, "L_tab": 2339, "n_element_tab": 265, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1947, "n_element_tab_1": 224, "formula_len_all": 2876, "formula_len_all_1": 700, "len_all": 260661, "len_all_1": 76218, "len_abs": 1562, "len_title": 109, "len_sents": 65561, "len_sents_1": 36646, "n_sents": 423, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1572, "title": "Informative Data Selection for Thorax Disease Classification", "abs": "Although Deep Neural Networks (DNNs) such as Vision Transformers (ViTs) have demonstrated superior performance in medical imaging tasks, the training of DNNs usually requires large amounts of high-quality labeled training data, which is usually difficult or even impractical to collect in the medical domain. To address this issue, Generative Data Augmentation (GDA) has been employed to improve the performance of DNNs trained on augmented training data comprising both original training data in the standard benchmark datasets and synthetic training data generated by generative models such as Diffusion Models (DMs). However, the synthetic data generated by GDA universally suffer from noise, and such synthetic data can severely hurt the performance of classifiers trained on the augmented training data. Existing works, such as data selection and data re-weighting methods aiming to mitigate this issue, usually depend on a given clean metadata or external classifier.\nIn this work, we propose a principled sample re-weighting method, Informative Data Selection (IDS), based on an established information theoretic measure, the Information Bottleneck (IB), to improve the performance of DNNs trained for thorax disease classification with GDA. Extensive experiments demonstrate that IDS successfully assigns higher weights to more informative synthetic images and significantly outperforms existing data selection and data re-weighting methods in GDA for thorax disease classification.\nThe code of IDS is available at \\url{https://anonymous.4open.science/r/IDS-20D1}.", "keywords": [ "Informative Data Selection", "Generative Data Augmentation", "Thorax Disease Classification" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "VbAxCwV2e3", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Chao Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Fangyikang Wang", "gender": "Male", "institution": "Wechat, Tencent Inc", "country": "CN", "position": "Intern" }, { "name": "Hubery Yin", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Song-Hai Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yizhe Zhao", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Chne Li", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Expert researcher" }, { "name": "Yuejiang Dong", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 113, "n_formula_1": 46, "n_ref_uni": 41, "n_ref": 80, "n_ref_all": 111, "n_fig": 1, "n_tab": 9, "L_tab": 2272, "n_element_tab": 255, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 168, "n_element_tab_1": 7, "formula_len_all": 18925, "formula_len_all_1": 3878, "len_all": 179946, "len_all_1": 59180, "len_abs": 1281, "len_title": 115, "len_sents": 40790, "len_sents_1": 23866, "n_sents": 393, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1353, "title": "DC-DPM: A Divide-and-Conquer Approach for Diffusion Reverse Process", "abs": "Diffusion models have achieved great success in generative tasks\\textblue{, with the quality of generated samples guaranteed by their convergence properties, typically derived within the context of stochastic differential equations(SDE) and often involving Kolmogorov equations for proofs. This paper introduces a novel method for proving the convergence of diffusion models, which relies on direct estimation of distributions without the need for SDE tools. This approach inspires a \\textbf{D}ivide-and-\\textbf{C}onquer strategy for approximating the reversed transition kernel of \\textbf{D}iffusion \\textbf{P}robabilistic \\textbf{M}odels (DC-DPM), which is not derived from SDEs, making previous convergence methods inapplicable. However, our method can be easily extended to accommodate this. As} our DC-DPM learns specific kernels for each partition \\textblue{, these kernels require merging. According to the proof of convergence, we} design two merging strategies for these cluster-specific kernels along with corresponding training and sampling methods.\nExperimental results demonstrate the superior generation quality of our method compared to the traditional single Gaussian kernel. Furthermore, our DC-DPM can synergize with previous kernel optimization methods, enhancing their generation quality, especially with a small number of timesteps.", "keywords": [ "Diffusion Model; Reverse Process Transition Kernel; Divide-and-Conquer" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "VarjSNbij7", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Jiacong Hu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jian Lou", "gender": "unknown", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jiawen Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Kejia Chen", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Mingli Song", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zunlei Feng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 60, "n_ref": 110, "n_ref_all": 131, "n_fig": 4, "n_tab": 6, "L_tab": 439, "n_element_tab": 57, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1852, "n_element_tab_1": 126, "formula_len_all": 603, "formula_len_all_1": 458, "len_all": 189331, "len_all_1": 72327, "len_abs": 1823, "len_title": 97, "len_sents": 49578, "len_sents_1": 35822, "n_sents": 322, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1831, "title": "Q-resafe: Assessing Safety Risks and Quantization-aware Safety Patching for Quantized Large Language Models", "abs": "Quantized large language models (LLMs) have garnered surging demand for broadening the deployment scenarios of LLMs, particularly on resource-constrained applications, which would otherwise be infeasible due to the substantial resource overhead incurred by astronomical model sizes. Propelled by this vast application potential, various quantization techniques have been developed to convert high-precision LLMs into low-precision quantized counterparts, aiming to preserve strong capabilities with reduced bit-widths. While these techniques have made significant strides in preserving utility, their implications for safety remain insufficiently studied. Recent findings highlight the fragility of safety mechanisms in both high-precision and quantized LLMs, underscoring the need for systematic safety evaluations and targeted interventions for quantized models.\n\nIn this paper, we present a comprehensive safety evaluation of quantized LLMs to complement existing efforts, covering four mainstream quantization techniques across diverse settings, including varying quantization bit-widths and different quantization-assisting datasets, through widely-accepted safety measurements. Our empirical evaluation reveals concerning safety degradation across all quantization methods and settings. To address this, we propose a quantization-aware safety patching framework, Q-resafe, to efficiently restore the safety capabilities of quantized LLMs while minimizing any adverse impact on utility. Extensive experiments demonstrate that Q-resafe effectively restores the safety of quantized LLMs obtained from diverse quantization processes, aligning closely with pre-quantization LLMs, even when evaluated against challenging datasets. We will make our implementation publicly available https://anonymous.4open.science/r/Qresafe-D085/.", "keywords": [ "Large Language Models", "Quantization", "Safety Alignment" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "VaowElpVzd", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Hengyuan Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Jiahao Pan", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Qifeng Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Wei Xue", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Wenhan Luo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Xingqun Qi", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yatian Wang", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Researcher" }, { "name": "Yike Guo", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Shanghang Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 46, "n_ref": 106, "n_ref_all": 123, "n_fig": 7, "n_tab": 8, "L_tab": 1738, "n_element_tab": 136, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 200, "n_element_tab_1": 14, "formula_len_all": 274, "formula_len_all_1": 757, "len_all": 177660, "len_all_1": 63531, "len_abs": 501, "len_title": 149, "len_sents": 44715, "len_sents_1": 30171, "n_sents": 347, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 112, "L_abs": 1588, "title": "Co$^{\\mathbf{3}}$Gesture: Towards Coherent Concurrent Co-speech 3D Gesture Generation with Interactive Diffusion", "abs": "Generating gestures from human speech has gained tremendous progress in animating virtual avatars. While the existing methods enable synthesizing gestures cooperated by people self-talking, they overlook the practicality of concurrent gesture modeling with two-person interactive conversations. Moreover, the lack of high-quality datasets with concurrent co-speech gestures also limits handling this issue. To fulfill this goal, we first construct a large-scale concurrent co-speech gesture dataset that contains more than 7M frames for diverse two-person interactive posture sequences, dubbed $\\textbf{GES-Inter}$. Moreover, we propose Co$^{\\mathbf{3}}$Gesture, a novel framework that enables concurrent coherent co-speech gesture synthesis including two-person interactive movements. Our framework is built upon two cooperative generation branches conditioned on decomposed speaker audio. Specifically, to enhance the coordination of human postures w.r.t corresponding speaker audios while interacting with the conversational partner, we present a Temporal-Interaction Module ($\\textbf{TIM}$). TIM can effectively model the temporal association representation between two speakers' gesture sequences as interaction guidance and fuse it into the concurrent gesture generation. Then, we devise a mutual attention mechanism to further boost learning dependencies of interacted concurrent motions, thereby enabling us to generate vivid and coherent gestures. Extensive experiments demonstrate that our method outperforms the state-of-the-art models on our newly collected GES-Inter dataset.", "keywords": [ "3D co-speech gesture generation", "human motion modeling" ], "rating_list": [ 10, 6, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "VaoeAi5CW8", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Li Yifei", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Min Wan", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Quantao Yang", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "Postdoc" }, { "name": "Shichao Fan", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Yajie Liu", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Qingjie LIU", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 22, "n_ref": 44, "n_ref_all": 69, "n_fig": 13, "n_tab": 4, "L_tab": 2412, "n_element_tab": 171, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1293, "n_element_tab_1": 124, "formula_len_all": 355, "formula_len_all_1": 355, "len_all": 122526, "len_all_1": 64616, "len_abs": 1316, "len_title": 138, "len_sents": 40615, "len_sents_1": 30654, "n_sents": 301, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1320, "title": "Diffusion Trajectory-guided Policy: A Novel Framework for Long-Horizon Robot Manipulation", "abs": "Recently, Vision-Language Models (VLMs) have made substantial progress in robot imitation learning, benefiting from increased amounts of demonstration data. However, the high cost of data collection remains a significant bottleneck, and the scarcity of demonstrations often result in poor generalization of the imitation policy, especially in long-horizon robotic manipulation tasks. To address these challenges, we propose the Diffusion Trajectory-guided Policy (DTP) framework, which generates task-relevant trajectories through a diffusion model to guide policy learning for long-horizon tasks. Furthermore, we demonstrate that our DTP method offers a useful interface for prompt engineering, providing a novel way to connect robot manipulation skills with interactions involving LLMs or humans. Our approach employs a two-stage training process: initially, we train a generative vision-language model to create diffusion task-relevant trajectories, then refine the imitation policy using these trajectories. We validate that the DTP method achieves substantial performance improvements in extensive experiments on the CALVIN simulation benchmark, starting from scratch without any external pretraining. Our approach outperforms state-of-the-art baselines by an average of 25% in success rate across various settings.", "keywords": [ "Robotics", "Imitation Learning", "Generative Model", "Vision-Language Action" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "VaUy5GZO3f", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Chunyi Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Guangtao Zhai", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Haoning Wu", "gender": "Male", "institution": "Rhymes AI", "country": "SG", "position": "Researcher" }, { "name": "Wei Sun", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Postdoc" }, { "name": "Weisi Lin", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Xiaohong Liu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiongkuo Min", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yingjie Zhou", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Ziheng Jia", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zijian Chen", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "zicheng zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 0, "n_ref_uni": 29, "n_ref": 68, "n_ref_all": 83, "n_fig": 6, "n_tab": 7, "L_tab": 6156, "n_element_tab": 781, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1924, "n_element_tab_1": 230, "formula_len_all": 283, "formula_len_all_1": 0, "len_all": 198302, "len_all_1": 58053, "len_abs": 2080, "len_title": 115, "len_sents": 55695, "len_sents_1": 26782, "n_sents": 429, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1604, "title": "Q-Bench-Video: Benchmarking the Video Quality Understanding of LMMs", "abs": "With the rising interest in research on Large Multi-modal Models (LMMs) for video understanding, many studies have emphasized general video comprehension capabilities, neglecting the **systematic exploration into video quality understanding**. To address this oversight, we introduce **Q-Bench-Video** in this paper, a new benchmark specifically designed to evaluate LMMs' proficiency in discerning video quality. **a)** To ensure the diversity of video sources, Q-Bench-Video encompasses videos from natural scenes, computer graphics (CG), and AI-generated content (AIGC). **b)** Building on the traditional multiple-choice questions format with the *Yes-or-No* and *What-How* categories, we include *Open-ended* questions to better evaluate complex scenarios. Additionally, we incorporate the **video pair quality comparison** question to enhance comprehensiveness. **c)** Beyond the traditional *Technical*, *Aesthetic*, and *Temporal* distortions, we have expanded our evaluation aspects to include the dimension of *AIGC* distortions, which addresses the increasing demand for video generation. Finally, we collect a total of 2,378 question-answer pairs and test them on 12 open-source & 5 proprietary LMMs. Our findings indicate that while LMMs have a foundational understanding of video quality, their performance remains incomplete and imprecise, with a notable discrepancy compared to human-level performance. Through **Q-Bench-Video**, we seek to catalyze community interest, stimulate further research, and unlock the untapped potential of LMMs to close the gap in video quality understanding.", "keywords": [ "Large multi-modal model", "benchmark", "video quality assessment" ], "rating_list": [ 3, 5, 6, 5, 5 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 5, 4 ] }, { "paper_id": "VZzx0MPA85", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 12, "author_info_list": [ { "name": "Chang Ao", "gender": "unknown", "institution": "Southern University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Chengming Li", "gender": "unknown", "institution": " Shenzhen MSU-BIT University", "country": "CN", "position": "Full Professor" }, { "name": "Jiaming Li", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Undergrad student" }, { "name": "Jiayuan Su", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Jing Luo", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "Undergrad student" }, { "name": "Liang Zhu", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Longze Chen", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Min Yang", "gender": "Female", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Run Luo", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Wen Yang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xincheng", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "YukunChen", "gender": "unknown", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 35, "n_ref": 69, "n_ref_all": 81, "n_fig": 5, "n_tab": 4, "L_tab": 417, "n_element_tab": 34, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 388, "n_element_tab_1": 46, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 211064, "len_all_1": 53730, "len_abs": 1407, "len_title": 126, "len_sents": 36054, "len_sents_1": 25974, "n_sents": 284, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1416, "title": "PersonaMath: Enhancing Math Reasoning through Persona-Driven Data Augmentation", "abs": "While closed-source Large Language Models (LLMs) demonstrate strong mathematical problem-solving abilities, open-source models continue to struggle with such tasks. To bridge this gap, we propose a data augmentation approach and introduce PersonaMathQA, a dataset derived from MATH and GSM8K, on which we train the PersonaMath models. Our approach consists of two stages: the first stage is learning from Persona Diversification, and the second stage is learning from Reflection. In the first stage, we regenerate detailed chain-of-thought (CoT) solutions as instructions using a closed-source LLM and introduce a novel persona-driven data augmentation technique to enhance the dataset's quantity and diversity. In the second stage, we incorporate reflection to fully leverage more challenging and valuable questions. Evaluation of our PersonaMath models on MATH and GSM8K reveals that the PersonaMath-7B model (based on LLaMA-2-7B) achieves an accuracy of 24.2\\% on MATH and 68.7\\% on GSM8K, surpassing all baseline methods and achieving state-of-the-art performance. Notably, our dataset contains only 70.3K data points\u2014merely 17.8\\% of MetaMathQA and 27\\% of MathInstruct\u2014yet our model outperforms these baselines, demonstrating the high quality and diversity of our dataset, which enables more efficient model training. We open-source the PersonaMathQA dataset, PersonaMath models, and our code for public usage.", "keywords": [ "Large Language Model; Mathematical Reasoning" ], "rating_list": [ 3, 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "VZN0irKnl0", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Dayal Singh Kalra", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Maissam Barkeshli", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Tianyu He", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 51, "n_formula_1": 8, "n_ref_uni": 29, "n_ref": 78, "n_ref_all": 160, "n_fig": 34, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3806, "formula_len_all_1": 743, "len_all": 207062, "len_all_1": 71053, "len_abs": 1187, "len_title": 165, "len_sents": 71546, "len_sents_1": 33142, "n_sents": 569, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 116, "L_abs": 1135, "title": "Universal Sharpness Dynamics in Neural Network Training: Fixed Point Analysis, Edge of Stability, and Route to Chaos", "abs": "In gradient descent dynamics of neural networks, the top eigenvalue of the Hessian of the loss (sharpness) displays a variety of robust phenomena throughout training. This includes early time regimes where the sharpness may decrease during early periods of training (sharpness reduction), and later time behavior such as progressive sharpening and edge of stability. We demonstrate that a simple \n-layer linear network (UV model) trained on a single training example exhibits all of the essential sharpness phenomenology observed in real-world scenarios. By analyzing the structure of dynamical fixed points in function space and the vector field of function updates, we uncover the underlying mechanisms behind these sharpness trends. Our analysis reveals (i) the mechanism behind early sharpness reduction and progressive sharpening, (ii) the required conditions for edge of stability, and (iii) a period-doubling route to chaos on the edge of stability manifold as learning rate is increased. Finally, we demonstrate that various predictions from this simplified model generalize to real-world scenarios and discuss its limitations.", "keywords": [ "Sharpness Dynamics", "Catapult Effect", "Edge of Stability" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "VZCxToUuNL", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Fabian Fumagalli", "gender": "Male", "institution": "Universit\u00e4t Bielefeld", "country": "DE", "position": "PhD student" }, { "name": "Marcel Wever", "gender": "Male", "institution": "Universit\u00e4t Hannover", "country": "DE", "position": "Postdoc" }, { "name": "Marius Lindauer", "gender": "Male", "institution": "Leibniz Universit\u00e4t Hannover", "country": "DE", "position": "Full Professor" }, { "name": "Maximilian Muschalik", "gender": "Male", "institution": "Institute of Computer Science, Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 16, "n_ref_uni": 46, "n_ref": 99, "n_ref_all": 120, "n_fig": 14, "n_tab": 2, "L_tab": 1394, "n_element_tab": 124, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 321, "n_element_tab_1": 6, "formula_len_all": 1567, "formula_len_all_1": 1033, "len_all": 163945, "len_all_1": 63634, "len_abs": 1563, "len_title": 120, "len_sents": 51638, "len_sents_1": 30310, "n_sents": 330, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1568, "title": "HyperSHAP: Shapley Values and Interactions for Hyperparameter Importance", "abs": "Hyperparameter optimization (HPO) is a crucial step in achieving strong predictive performance, particularly for deep learning with hyperparameters controlling the neural architecture and learning behavior. However, the impact of some hyperparameters on model generalization can vary significantly depending on the dataset and performance measure, making it challenging to generalize their importance. Gaining a better understanding of the importance of hyperparameters is therefore important to deepen our understanding of machine learning and to leverage this knowledge in future downstream HPO tasks, especially if training is expensive and HPO needs to be as efficient as possible.\nTo address these challenges, we propose a game theoretic framework based on Shapley values and interactions for HPO. These methods offer an additive decomposition of a performance measure across hyperparameters, enabling both local and global explanations of hyperparameter importance and interactions. Our framework, named HyperSHAP, provides insights into ablation studies, tunability of specific hyperparameter configurations, and entire configuration spaces. Through experiments, we demonstrate that focusing on the hyperparameters deemed important by our framework can improve performance during subsequent hyperparameter optimization, while ignoring important hyperparameters or interactions degrades performance. This validates the effectiveness of our approach in enhancing model performance and providing meaningful, interpretable explanations of hyperparameter importance.", "keywords": [ "explainability", "Shapley", "interaction", "hyperparameter optimization" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "VZC9aJoI6a", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Akshay Nambi", "gender": "Male", "institution": "Microsoft Research", "country": "IN", "position": "Principal Researcher" }, { "name": "Eshaan Agarwal", "gender": "Male", "institution": "Indian Institute of Technology (Banaras Hindu University) Varanasi", "country": "IN", "position": "Undergrad student" }, { "name": "Joykirat Singh", "gender": "Male", "institution": "Research, Microsoft", "country": "IN", "position": "Researcher" }, { "name": "Raghav Magazine", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Researcher" }, { "name": "Tanuja Ganu", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Vivek Dani", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 70, "n_ref_all": 96, "n_fig": 6, "n_tab": 21, "L_tab": 6878, "n_element_tab": 619, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 2677, "n_element_tab_1": 271, "formula_len_all": 639, "formula_len_all_1": 0, "len_all": 255265, "len_all_1": 79263, "len_abs": 1110, "len_title": 54, "len_sents": 96929, "len_sents_1": 36205, "n_sents": 753, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1123, "title": "PromptWizard: Task-Aware Prompt Optimization Framework", "abs": "Large language models (LLMs) have transformed AI across diverse domains, with \\textit{prompting} being central to their success in guiding model outputs. However, manual prompt engineering is both labor-intensive and domain-specific, necessitating the need for automated solutions. We introduce PromptWizard, a novel, fully automated framework for discrete prompt optimization, utilizing a self-evolving, self-adapting mechanism. Through a feedback-driven critique and synthesis process, PromptWizard achieves an effective balance between exploration and exploitation, iteratively refining both prompt instructions and in-context examples to generate human-readable, task-specific prompts. This guided approach systematically improves prompt quality, resulting in superior performance across 45 tasks. PromptWizard excels even with limited training data, smaller LLMs, and various LLM architectures. Additionally, our cost analysis reveals a substantial reduction in API calls, token usage, and overall cost, demonstrating PromptWizard's efficiency, scalability, and advantages over existing prompt optimization strategies.", "keywords": [ "Prompt optimization", "LLMs", "task-aware" ], "rating_list": [ 5, 3, 3, 5, 6 ], "soundness_list": [ 3, 2, 2, 4, 3 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 5, 3, 4, 4, 3 ] }, { "paper_id": "VZ8kwfspAi", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 1, "author_info_list": [ { "name": "Sergio Charles", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" } ], "n_formula": 30, "n_formula_1": 32, "n_ref_uni": 17, "n_ref": 29, "n_ref_all": 38, "n_fig": 2, "n_tab": 3, "L_tab": 494, "n_element_tab": 51, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1336, "n_element_tab_1": 63, "formula_len_all": 4297, "formula_len_all_1": 4483, "len_all": 91130, "len_all_1": 55969, "len_abs": 669, "len_title": 125, "len_sents": 25503, "len_sents_1": 21623, "n_sents": 196, "n_sents_1": 166, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 675, "title": "Spacetime $E(n)$-Transformer: Equivariant Attention for Spatio-temporal Graphs", "abs": "We introduce an $E(n)$-equivariant Transformer architecture for spatio-temporal graph data. By imposing rotation, translation, and permutation equivariance inductive biases in both space and time, we show that the Spacetime $E(n)$-Transformer (SET) outperforms purely spatial and temporal models without symmetry-preserving properties. We benchmark SET against said models on the $N$-body problem, a simple physical system with complex dynamics. While existing spatio-temporal graph neural networks focus on sequential modeling, we empirically demonstrate that leveraging underlying domain symmetries yields considerable improvements for modeling dynamical systems on graphs.", "keywords": [ "graph neural networks", "graph representation learning" ], "rating_list": [ 3, 5, 3, 5, 3 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 1, 2 ], "contribution_list": [ 1, 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "VYvxrD7aS0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Goran Glava\u0161", "gender": "Male", "institution": "Julius-Maximilians-Universit\u00e4t W\u00fcrzburg", "country": "", "position": "Full Professor" }, { "name": "Haoyi Yang", "gender": "Male", "institution": "TU Darmstadt", "country": "", "position": "MS student" }, { "name": "Iryna Gurevych", "gender": "Female", "institution": "Institute for Computer Science, Artificial Intelligence and Technology", "country": "BG", "position": "Full Professor" }, { "name": "Kristian Kersting", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Full Professor" }, { "name": "Indraneil Arun Paul", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 109, "n_ref": 280, "n_ref_all": 300, "n_fig": 2, "n_tab": 17, "L_tab": 8690, "n_element_tab": 1091, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1991, "n_element_tab_1": 248, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 435314, "len_all_1": 76735, "len_abs": 1422, "len_title": 127, "len_sents": 52471, "len_sents_1": 38281, "n_sents": 332, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1426, "title": "ObscuraCoder: Powering Efficient Code LM Pre-Training Via Obfuscation Grounding", "abs": "Language models (LMs) have become a staple of the code-writing toolbox. Their pre-training recipe has, however, remained stagnant over recent years, barring the occasional changes in data sourcing and filtering strategies. In particular, research exploring modifications to Code-LMs' pre-training objectives, geared towards improving data efficiency and better disentangling between syntax and semantics, has been noticeably sparse, especially compared with corresponding efforts in natural language LMs. In this work, we examine grounding on obfuscated code as a means of helping Code-LMs look beyond the surface-form syntax and enhance their pre-training sample efficiency. To this end, we compile ObscuraX, a dataset of approximately 55M source and obfuscated code pairs in seven languages. Subsequently, we pre-train ObscuraCoder models, ranging in size from 255M to 2.8B parameters, on a 272B-token corpus that includes ObscuraX and demonstrate that our obfuscation-based pre-training recipe leads to consistent improvements in Code-LMs' abilities compared to both vanilla autoregressive pre-training as well as existing de-obfuscation (DOBF) objectives. ObscuraCoder demonstrates sizeable gains across multiple tests of syntactic and semantic code understanding, along with improved capabilities in multilingual code completion, multilingual code commit summarization, and multi-purpose library-oriented code generation.", "keywords": [ "code generation", "code obfuscation", "language modelling", "code lm" ], "rating_list": [ 8, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "VYlfoA8I6A", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Alessandro G Allievi", "gender": "unknown", "institution": "University of Texas at Austin", "country": "", "position": "Postdoc" }, { "name": "Chen Tang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Postdoc" }, { "name": "Hang Qiu", "gender": "unknown", "institution": "University of California, Riverside", "country": "US", "position": "Assistant Professor" }, { "name": "Janice Nguyen", "gender": "unknown", "institution": "University of California, Riverside", "country": "US", "position": "PhD student" }, { "name": "Jarrett Holtz", "gender": "Male", "institution": "Bosch", "country": "US", "position": "Researcher" }, { "name": "Jiaxun Cui", "gender": "Female", "institution": "The University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Peter Stone", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Principal Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 41, "n_ref": 83, "n_ref_all": 94, "n_fig": 4, "n_tab": 6, "L_tab": 2479, "n_element_tab": 326, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1002, "n_element_tab_1": 139, "formula_len_all": 45, "formula_len_all_1": 45, "len_all": 205802, "len_all_1": 64142, "len_abs": 1912, "len_title": 58, "len_sents": 74043, "len_sents_1": 33010, "n_sents": 457, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1349, "title": "Talking Vehicles: Cooperative Driving via Natural Language", "abs": "Using natural language as a vehicle-to-vehicle (V2V) communication protocol offers the potential for autonomous vehicles to drive cooperatively not only with each other but also with human drivers. Simple and effective messages for sharing critical observations or negotiating plans to achieve coordination could improve traffic safety and efficiency compared to methods without communication. In this work, we propose a suite of traffic tasks in vehicle-to-vehicle autonomous driving where vehicles in a traffic scenario need to communicate in natural language to facilitate coordination in order to avoid an imminent collision and/or support efficient traffic flow, which we model as a general-sum partially observable stochastic game. To this end, this paper introduces a novel method, LLM+Debrief, to learn a message generation and control policy for autonomous vehicles through multi-agent discussion. To evaluate our method, we developed a gym-like simulation environment that contains a range of accident-prone driving scenarios that could be alleviated by communication. Our experimental results demonstrate that our method is more effective at generating meaningful and human-understandable natural language messages to facilitate cooperation and coordination than untrained LLMs. Our anonymous code is available in supplementary materials.", "keywords": [ "multi-agent learning", "communication", "llm", "autonomous driving" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "VYfYISQncf", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Gang Chen", "gender": "Male", "institution": "College of Computer Science and Technology, Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Guozhi Xu", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Hanbin Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Haobo Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tianlei Hu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Associate Professor" }, { "name": "Yueying Li", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zhou Jiaqing", "gender": "Male", "institution": "bytedance", "country": "CN", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 51, "n_ref": 128, "n_ref_all": 171, "n_fig": 14, "n_tab": 11, "L_tab": 5499, "n_element_tab": 655, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1694, "n_element_tab_1": 234, "formula_len_all": 978, "formula_len_all_1": 788, "len_all": 179464, "len_all_1": 59087, "len_abs": 1330, "len_title": 125, "len_sents": 45236, "len_sents_1": 25833, "n_sents": 326, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1370, "title": "FedSR: Frequency-Aware Enhancement for Diffusion-based Image Super-Resolution", "abs": "Image super-resolution (ISR) is a classic and challenging problem in low-level vision because the data collection process often introduces complex and unknown degradation patterns. Leveraging powerful generative priors, diffusion-based algorithms have recently established new state-of-the-art ISR performance. Despite the promise, current diffusion-based ISR methods mostly focus on the spatial domain. To bridge this gap, we first experimentally validate that the key to solving the ISR problem lies in addressing the degradation of image amplitude information and high-frequency details. Based on this, we propose a novel $\\textit{training-free}$ frequency-aware enhancement framework ($\\textbf{FedSR}$) for diffusion-based ISR methods, which consists of two critical components. Firstly, we design the Amplitude Enhancement Module (AEM), which selectively enhances crucial amplitude channels through weighted optimization. Secondly, we introduce the High-Frequency Enhancement Module (HEM) that adaptively masks the skip features to perform high-pass filtering. Through extensive evaluations on both synthetic datasets and real-world image collections, our method demonstrates outstanding performance in reproducing realistic image details without additional tuning. For instance, FedSR improves StableSR across three datasets by $\\textbf{+10.53}\\%$ on MUSIQ metric.", "keywords": [ "Image Super-resolution", "Frequency-Domain", "Diffusion Models" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "VYWBMq1L7H", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Christopher D. Manning", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "", "position": "Full Professor" }, { "name": "Christopher Potts", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Julie Kallini", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "R\u00f3bert Csord\u00e1s", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Shikhar Murty", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 26, "n_ref": 41, "n_ref_all": 64, "n_fig": 6, "n_tab": 11, "L_tab": 2958, "n_element_tab": 362, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1698, "n_element_tab_1": 175, "formula_len_all": 663, "formula_len_all_1": 229, "len_all": 146571, "len_all_1": 65191, "len_abs": 1586, "len_title": 116, "len_sents": 48398, "len_sents_1": 30869, "n_sents": 361, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1590, "title": "MrT5: Dynamic Token Merging for Efficient Byte-level Language Models", "abs": "Models that rely on subword tokenization have significant drawbacks, such as sensitivity to character-level noise like spelling errors and inconsistent compression rates across different languages and scripts. While character or byte-level models like ByT5 attempt to address these concerns, they have not gained widespread adoption\u2014processing raw byte streams without tokenization results in significantly longer sequence lengths, making training and inference inefficient. This work introduces MrT5 (MergeT5), a more efficient variant of ByT5 that integrates a token deletion mechanism in its encoder to dynamically shorten the input sequence length. After processing through a fixed number of encoder layers, a learnt delete gate determines which tokens are to be removed and which are to be retained for subsequent layers. MrT5 effectively \"merges\" critical information from deleted tokens into a more compact sequence, leveraging contextual information from the remaining tokens. In continued pre-training experiments, we find that MrT5 can achieve significant gains in inference runtime with minimal effect on performance. When trained on English text, MrT5 demonstrates the capability to transfer its deletion feature zero-shot across several languages, with significant additional improvements following multilingual training. Furthermore, MrT5 shows comparable accuracy to ByT5 on downstream evaluations such as XNLI and character-level tasks while reducing sequence lengths by up to 80%. Our approach presents a solution to the practical limitations of existing byte-level models.", "keywords": [ "NLP", "ByT5", "T5", "tokenization", "byte-level language models", "character-level language models" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "VYRT8ajHRr", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Jifeng Dai", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Wenhan Dou", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Xizhou Zhu", "gender": "unknown", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xue Yang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhaokai Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Intern" }, { "name": "Gen Luogen", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 79, "n_ref": 193, "n_ref_all": 209, "n_fig": 7, "n_tab": 6, "L_tab": 2780, "n_element_tab": 575, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2086, "n_element_tab_1": 203, "formula_len_all": 521, "formula_len_all_1": 387, "len_all": 238132, "len_all_1": 52154, "len_abs": 1869, "len_title": 168, "len_sents": 37224, "len_sents_1": 22876, "n_sents": 262, "n_sents_1": 165, "venue": "ICLR2025", "decision": "withdraw", "L_title": 120, "L_abs": 1907, "title": "Mono-InternVL: Pushing the Boundaries of Monolithic Multimodal Large Language Models with Endogenous Visual Pre-training", "abs": "The rapid advancement of Large Language Models (LLMs) has led to an influx of efforts to extend their capabilities to multimodal tasks. Among them, growing attention have been focused on monolithic Multimodal Large Language Models (MLLMs) that integrate visual encoding and language decoding into a single LLM. Despite the structural simplicity and deployment-friendliness, training a monolithic MLLM with promising performance still remains challenging. In particular, the popular approaches adopt continuous pre-training to extend a pre-trained LLM to a monolithic MLLM, which suffers from catastrophic forgetting and leads to performance degeneration. In this paper, we aim to overcome this limitation from the perspective of delta tuning. Specifically, our core idea is to embed visual parameters into a pre-trained LLM, thereby incrementally learning visual knowledge from massive data via delta tuning, i.e., freezing the LLM when optimizing the visual parameters. Based on this principle, we present Mono-InternVL, a novel monolithic MLLM that seamlessly integrates a set of visual experts via a multimodal mixture-of-experts structure. Moreover, we propose an innovative pre-training strategy to maximize the visual capability of Mono-InternVL, namely Endogenous Visual Pre-training (EViP). In particular, EViP is designed as a progressive learning process for visual experts, which aims to fully exploit the visual knowledge from noisy data to high-quality data. To validate our approach, we conduct extensive experiments on 16 benchmarks. Experimental results not only validate the superior performance of Mono-InternVL compared to the state-of-the-art MLLM on 6 multimodal benchmarks, e.g., +113 points over InternVL-1.5 on OCRBench, but also confirm its better deployment efficiency, with first token latency reduced by up to 67%. Our code and models will be released.", "keywords": [ "Multimodal Models", "Large Language Models", "Vision Language Models" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 5, 5, 3 ] }, { "paper_id": "VYOe2eBQeh", "primary_area": "applications to robotics, autonomy, planning", "n_author": 16, "author_info_list": [ { "name": "Ajay Mandlekar", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Baolin Peng", "gender": "Male", "institution": "Research, Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Bill Y Lin", "gender": "Male", "institution": "xAI", "country": "US", "position": "Researcher" }, { "name": "Dieter Fox", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Senior Director of Robotics Research" }, { "name": "Jianfeng Gao", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Joel Jang", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Kimin Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Lars Liden", "gender": "unknown", "institution": "Microsoft Research", "country": "", "position": "Principal Researcher" }, { "name": "Luke Zettlemoyer", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Full Professor" }, { "name": "Minjoon Seo", "gender": "Male", "institution": "Twelve Labs", "country": "", "position": "Chief Scientist" }, { "name": "Reuben Tan", "gender": "Male", "institution": "Research, Microsoft", "country": "US", "position": "Researcher" }, { "name": "Se June Joo", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" }, { "name": "Seonghyeon Ye", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Yu-Wei Chao", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Byeong Guk Jeon", "gender": "Male", "institution": "LANGUAGE & KNOWLEDGE LAB", "country": "", "position": "Intern" }, { "name": "Jianwei Yang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 0, "n_ref_uni": 50, "n_ref": 117, "n_ref_all": 182, "n_fig": 20, "n_tab": 11, "L_tab": 4199, "n_element_tab": 787, "n_fig_1": 10, "n_tab_1": 5, "L_tab_1": 1694, "n_element_tab_1": 263, "formula_len_all": 123, "formula_len_all_1": 0, "len_all": 202063, "len_all_1": 67998, "len_abs": 3285, "len_title": 85, "len_sents": 56740, "len_sents_1": 31523, "n_sents": 411, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 37, "L_abs": 1400, "title": "Latent Action Pretraining from Videos", "abs": "We introduce Latent Action Pretraining for general Action models (LAPA), the first unsupervised method for pretraining Vision-Language-Action (VLA) models without ground-truth robot action labels. Existing Vision-Language-Action models require action labels typically collected by human teleoperators during pretraining, which significantly limits possible data sources and scale. In this work, we propose a method to learn from internet-scale videos that do not have robot action labels. We first train an action quantization model leveraging VQ-VAE-based objective to learn discrete latent actions between image frames, then pretrain a latent VLA model to predict these latent actions from observations and task descriptions, and finally finetune the VLA on small-scale robot manipulation data to map from latent to robot actions. Experimental results demonstrate that our method significantly outperforms existing techniques that train robot manipulation policies from large-scale videos. Furthermore, it outperforms the state-of-the-art VLA model trained with robotic action labels on real-world manipulation tasks that require language conditioning, generalization to unseen objects, and semantic generalization to unseen instructions. Training only on human manipulation videos also shows positive transfer, opening up the potential for leveraging web-scale data for robotics foundation models.", "keywords": [ "vision-language-action models", "robotics" ], "rating_list": [ 6, 3, 6, 8, 6 ], "soundness_list": [ 3, 2, 4, 3, 4 ], "presentation_list": [ 3, 2, 2, 4, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4, 4 ] }, { "paper_id": "VY96NfQRIo", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Bruno Trentini", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "C. Brian Roland", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Chen Tessler", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Darren J. Hsu", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Olivia Viessmann", "gender": "Female", "institution": "Flagship Pioneering", "country": "US", "position": "Principal Researcher" }, { "name": "Ryan Park", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Mariya Popova", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 15, "n_ref_uni": 35, "n_ref": 84, "n_ref_all": 114, "n_fig": 8, "n_tab": 5, "L_tab": 1437, "n_element_tab": 125, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 683, "n_element_tab_1": 26, "formula_len_all": 2741, "formula_len_all_1": 1350, "len_all": 169810, "len_all_1": 65318, "len_abs": 1068, "len_title": 152, "len_sents": 41060, "len_sents_1": 29905, "n_sents": 354, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1070, "title": "Improving Inverse Folding for Peptide Design with Diversity-Regularized Direct Preference Optimization", "abs": "Inverse folding models play an important role in structure-based design by predicting amino acid sequences that fold into desired reference structures. Models like ProteinMPNN, a message-passing encoder-decoder model, are trained to reliably produce new sequences from a reference structure. However, when applied to peptides, these models are prone to generating repetitive sequences that do not fold into the reference structure. To address this, we finetune ProteinMPNN to produce diverse and structurally consistent peptide sequences via Direct Preference Optimization (DPO). We derive two enhancements to DPO: online diversity regularization and domain-specific priors.Additionally, we develop a new understanding on improving diversity in decoder models. When conditioned on OpenFold generated structures, our finetuned models achieve state-of-the-art structural similarity scores, improving base ProteinMPNN by at least 8%. Compared to standard DPO, our regularized method achieves up to 20% higher sequence diversity with no loss in structural similarity score.", "keywords": [ "Inverse Folding", "Structure-based design", "Peptide Design", "Direct Preference Optimization" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "VXKt1lwysO", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Jianke Yang", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Jonghyun Park", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Manu Bhat", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Undergrad student" }, { "name": "Nima Dehmamy", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Robin Walters", "gender": "Male", "institution": "Northeastern University ", "country": "US", "position": "Assistant Professor" }, { "name": "Rose Yu", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 33, "n_ref": 61, "n_ref_all": 86, "n_fig": 15, "n_tab": 3, "L_tab": 446, "n_element_tab": 44, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 686, "n_element_tab_1": 32, "formula_len_all": 1164, "formula_len_all_1": 1005, "len_all": 156558, "len_all_1": 63232, "len_abs": 911, "len_title": 89, "len_sents": 51834, "len_sents_1": 27925, "n_sents": 482, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 914, "title": "ATLAS: Automatic Local Symmetry Discovery", "abs": "Existing symmetry discovery methods predominantly focus on global transformations across the entire system or space, but they fail to consider the symmetries in local neighborhoods. This may result in the reported symmetry group being a misrepresentation of the true symmetry. In this paper, we formalize the notion of local symmetry as atlas equivariance. Our proposed pipeline, automatic local symmetry discovery (ATLAS), recovers the local symmetries of a function by training local predictor networks and then learning a Lie group basis to which the predictors are equivariant. We demonstrate ATLAS is capable of discovering local symmetry groups with multiple connected components in top-quark tagging and partial differential equation experiments. The discovered local symmetry is shown to be a useful inductive bias that improves the performance of downstream tasks in climate segmentation and vision tasks.", "keywords": [ "local symmetry discovery", "symmetry discovery", "equivariance", "gauge equivariant neural network", "Lie theory" ], "rating_list": [ 6, 5, 3, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 2, 3 ], "confidence_list": [ 3, 4, 5, 2, 3 ] }, { "paper_id": "VWj9rTfZzQ", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 15, "author_info_list": [ { "name": "Anthony W. Jung", "gender": "Male", "institution": "KC ML2", "country": "KR", "position": "Researcher" }, { "name": "Chaejin Park", "gender": "Female", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Chanhyung Park", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Doyoung Heo", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Undergrad student" }, { "name": "Jeongmin Shin", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Juho Park", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Kevin Octavian", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Min Seok Jang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Sangjun Han", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Sanmun Kim", "gender": "Male", "institution": "California Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Seolho Kim", "gender": "unknown", "institution": "Inha University", "country": "", "position": "Undergrad student" }, { "name": "Sung Hyun Nam", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Yongha Kim", "gender": "Male", "institution": "KC Machine Learning Lab", "country": "", "position": "Researcher" }, { "name": "Chan Youn Park", "gender": "Male", "institution": "KC Machine Learning Lab", "country": "KR", "position": "Principal Research Scientist" }, { "name": "Jinmyoung lee", "gender": "Female", "institution": "KC-ML2", "country": "", "position": "Researcher" } ], "n_formula": 86, "n_formula_1": 7, "n_ref_uni": 78, "n_ref": 154, "n_ref_all": 223, "n_fig": 38, "n_tab": 17, "L_tab": 4298, "n_element_tab": 561, "n_fig_1": 12, "n_tab_1": 3, "L_tab_1": 1633, "n_element_tab_1": 190, "formula_len_all": 5789, "formula_len_all_1": 574, "len_all": 319601, "len_all_1": 63454, "len_abs": 1845, "len_title": 116, "len_sents": 88448, "len_sents_1": 27940, "n_sents": 762, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1857, "title": "MEENT: DIFFERENTIABLE ELECTROMAGNETIC SIMULATOR FOR MACHINE LEARNING", "abs": "Electromagnetic (EM) simulation plays a crucial role in analyzing and designing devices with sub-wavelength scale structures such as semiconductor devices and future displays. \nSpecifically, optics problems such as estimating semiconductor device structures and designing nanophotonic devices provide intriguing research topics with far-reaching real world impact. \nTraditional algorithms for such tasks require iteratively refining parameters through simulations, which often yield sub-optimal results due to the high computational cost of both the algorithms and EM simulations. \nMachine learning (ML) emerged as a promising candidate to mitigate these challenges, and optics research community has increasingly adopted ML algorithms to obtain results surpassing classical methods across various tasks.\nTo foster a synergistic collaboration between the optics and ML communities, it is essential to have an EM simulation software that is user-friendly for both research communities.\nTo this end, we present meent, an EM simulation software that employs rigorous coupled-wave analysis (RCWA). Developed in Python and equipped with automatic differentiation (AD) capabilities, meent serves as a versatile platform for integrating ML into optics research and vice versa.\nTo demonstrate its utility as a research platform, we present three applications of meent: 1) generating a dataset for training neural operator, 2) serving as an environment for the reinforcement learning of nanophotonic device optimization, and 3) providing a solution for inverse problems with gradient-based optimizers.\nThese applications highlight meent's potential to advance both EM simulation and ML methodologies. \nThe code is available on our Github repository with the MIT license to promote the cross-polinations of ideas among academic researchers and industry practitioners.", "keywords": [ "computational physics", "optics", "electromagnetic simulation", "reinforcement learning", "neural operator", "automatic differentiation", "metasurface optimization", "semiconductor metrology" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 4, 2 ], "confidence_list": [ 5, 2, 2, 3 ] }, { "paper_id": "VWBYDo5NaM", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Bo Pan", "gender": "unknown", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Guanchen Wu", "gender": "unknown", "institution": "Emory University", "country": "US", "position": "MS student" }, { "name": "Liang Zhao", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Associate Professor" }, { "name": "Yifei Zhang", "gender": "Male", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Zhen Xiong", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "MS student" }, { "name": "Zheng Zhang", "gender": "Male", "institution": "Emory University", "country": "US", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 47, "n_ref": 61, "n_ref_all": 88, "n_fig": 13, "n_tab": 5, "L_tab": 6819, "n_element_tab": 209, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 2195, "n_element_tab_1": 162, "formula_len_all": 410, "formula_len_all_1": 317, "len_all": 261241, "len_all_1": 64636, "len_abs": 1337, "len_title": 133, "len_sents": 74376, "len_sents_1": 30051, "n_sents": 1502, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1264, "title": "TAGExplainer: Narrating Graph Explanations for Text-Attributed Graph Learning Models", "abs": "Representation learning of Text-Attributed Graphs (TAGs) has garnered significant attention due to its applications in various domains, including recommendation systems and social networks. Despite advancements in TAG learning methodologies, challenges remain in explainability due to the black-box nature of existing TAG representation learning models. This paper presents TAGExplainer, the first method designed to generate natural language explanations for TAG learning. TAGExplainer employs a generative language model that maps input-output pairs to explanations reflecting the model's decision-making process. To address the lack of annotated ground truth explanations in real-world scenarios, we propose first generating pseudo-labels that capture the model's decisions from saliency-based explanations, then the pseudo-label generator is iteratively trained based on three training objectives focusing on faithfulness and brevity via Expert Iteration, to improve the quality of generated pseudo-labels. The high-quality pseudo-labels are finally utilized to train an end-to-end explanation generator model. Extensive experiments are conducted to demonstrate the effectiveness of TAGExplainer in producing faithful and concise natural language explanations.", "keywords": [ "Natural language explanations", "graph neural networks", "text-attributed graphs" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "VW21r9rTjE", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Aleksandar Bojchevski", "gender": "Male", "institution": "University of Cologne", "country": "DE", "position": "Full Professor" }, { "name": "Simone Antonelli", "gender": "unknown", "institution": "University of Cologne", "country": "DE", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 7, "n_ref_uni": 25, "n_ref": 57, "n_ref_all": 92, "n_fig": 22, "n_tab": 2, "L_tab": 398, "n_element_tab": 42, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 722, "formula_len_all_1": 514, "len_all": 156711, "len_all_1": 64284, "len_abs": 939, "len_title": 73, "len_sents": 56622, "len_sents_1": 30138, "n_sents": 487, "n_sents_1": 275, "venue": "ICLR2025", "decision": "Reject", "L_title": 25, "L_abs": 940, "title": "Data Valuation for Graphs", "abs": "What is the worth of a node? We answer this question using an emerging set of data valuation techniques, where the value of a data point is measured via its marginal contribution when added to the (training) dataset. Data valuation has been primarily studied in the i.i.d. setting, giving rise to methods like influence functions, leave-one-out estimation, data Shapley, and data Banzhaf. We conduct a comprehensive study of data valuation approaches applied to graph-structured models such as graph neural networks in a semi-supervised transductive setting. Since all nodes (labeled and unlabeled) influence both training and inference we construct various scenarios to understand the diverse mechanisms by which nodes can impact learning. We show that the resulting node values can be used to identify (positively and negatively) influential nodes, quantify model brittleness, detect poisoned data, and accurately predict counterfactuals.", "keywords": [ "Graphs Machine Learning", "Data Valuation", "Graph Neural Network" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "VVixJ9QavY", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Aditya V. Nori", "gender": "Male", "institution": "Microsoft Research", "country": "GB", "position": "Researcher" }, { "name": "Alihan H\u00fcy\u00fck", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Jacqueline R. M. A. Maasch", "gender": "Non-Binary", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Javier Gonzalez", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Principal Researcher" }, { "name": "Xinnuo Xu", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 37, "n_formula_1": 13, "n_ref_uni": 37, "n_ref": 54, "n_ref_all": 78, "n_fig": 9, "n_tab": 6, "L_tab": 2154, "n_element_tab": 145, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4436, "formula_len_all_1": 1029, "len_all": 198820, "len_all_1": 75462, "len_abs": 927, "len_title": 116, "len_sents": 66196, "len_sents_1": 35355, "n_sents": 514, "n_sents_1": 268, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 68, "L_abs": 931, "title": "Reasoning Elicitation in Language Models via Counterfactual Feedback", "abs": "Despite the increasing effectiveness of language models, their reasoning capabilities remain underdeveloped. In particular, causal reasoning through counterfactual question answering is lacking. This work aims to bridge this gap. We first derive novel metrics that balance accuracy in factual and counterfactual questions, capturing a more complete view of the reasoning abilities of language models than traditional factual-only based metrics. Second, we propose several fine-tuning approaches that aim to elicit better reasoning mechanisms, in the sense of the proposed metrics. Finally, we evaluate the performance of the fine-tuned language models in a variety of realistic scenarios. In particular, we investigate to what extent our fine-tuning approaches systemically achieve better generalization with respect to the base models in several problems that require, among others, inductive and deductive reasoning capabilities.", "keywords": [ "language models", "reasoning", "fine-tuning", "counterfactuals" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "VVVfuIcmKR", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Alberto Baldrati", "gender": "Male", "institution": "University of Pisa", "country": "IT", "position": "PhD student" }, { "name": "Andrew D. Bagdanov", "gender": "Male", "institution": "Universit\u00e0 degli Studi di Firenze", "country": "ES", "position": "Associate Professor" }, { "name": "Lorenzo Agnolucci", "gender": "Male", "institution": "University of Florence", "country": "IT", "position": "PhD student" }, { "name": "Marco Bertini", "gender": "Male", "institution": "Universit\u00e0 degli Studi di Firenze", "country": "IT", "position": "Associate Professor" }, { "name": "Marco Mistretta", "gender": "Male", "institution": "University of Florence", "country": "IT", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 8, "n_ref_uni": 48, "n_ref": 91, "n_ref_all": 116, "n_fig": 4, "n_tab": 9, "L_tab": 4766, "n_element_tab": 375, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2000, "n_element_tab_1": 83, "formula_len_all": 1018, "formula_len_all_1": 413, "len_all": 197436, "len_all_1": 76398, "len_abs": 1246, "len_title": 131, "len_sents": 62772, "len_sents_1": 36086, "n_sents": 500, "n_sents_1": 278, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1454, "title": "Cross the Gap: Exposing the Intra-modal Misalignment in CLIP via Modality Inversion", "abs": "Pre-trained multi-modal Vision-Language Models like CLIP are widely used off-the-shelf for a variety of applications. In this paper, we show that the common practice of individually exploiting the text or image encoders of these powerful multi-modal models is highly suboptimal for intra-modal tasks like image-to-image retrieval. We argue that this is inherently due to the CLIP-style inter-modal contrastive loss that does not enforce any intra-modal constraints, leading to what we call intra-modal misalignment. To demonstrate this, we leverage two optimization-based modality inversion techniques that map representations from their input modality to the complementary one without any need for auxiliary data or additional trained adapters. We empirically show that, in the intra-modal tasks of image-to-image and text-to-text retrieval, approaching these tasks inter-modally significantly improves performance with respect to intra-modal baselines on more than fifteen datasets. Additionally, we demonstrate that approaching a native inter-modal task (e.g. zero-shot image classification) intra-modally decreases performance, further validating our findings. Finally, we show that incorporating an intra-modal term in the pre-training objective or narrowing the modality gap between the text and image feature embedding spaces helps reduce the intra-modal misalignment. The code is publicly available at: https://github.com/miccunifi/Cross-the-Gap.", "keywords": [ "intra-modal misalignment", "CLIP", "modality gap", "modality inversion", "textual inversion", "visual inversion", "image retrieval", "text retrieval", "image classification" ], "rating_list": [ 8, 8, 6, 3 ], "soundness_list": [ 4, 3, 2, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "VVO3ApdMUE", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Eric Alsmann", "gender": "Male", "institution": "Universit\u00e4t Kassel", "country": "DE", "position": "PhD student" }, { "name": "Marco S\u00e4lzer", "gender": "Male", "institution": "University of Kassel", "country": "DE", "position": "PhD student" }, { "name": "Martin Lange", "gender": "Male", "institution": "Universit\u00e4t Kassel", "country": "", "position": "Full Professor" } ], "n_formula": 41, "n_formula_1": 6, "n_ref_uni": 21, "n_ref": 44, "n_ref_all": 57, "n_fig": 3, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5745, "formula_len_all_1": 1051, "len_all": 210893, "len_all_1": 74169, "len_abs": 926, "len_title": 126, "len_sents": 75022, "len_sents_1": 35018, "n_sents": 661, "n_sents_1": 283, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 929, "title": "Transformer Encoder Satisfiability: Complexity and Impact on Formal Reasoning", "abs": "We analyse the complexity of the satisfiability problem, or similarly feasibility problem, (trSAT) for transformer encoders (TE), which naturally occurs in formal verification or interpretation, collectively referred to as formal reasoning. We find that trSAT is undecidable when considering TE as they are commonly studied in the expressiveness community. Furthermore, we identify practical scenarios where trSAT is decidable and establish corresponding complexity bounds. Beyond trivial cases, we find that quantized TE, those restricted by fixed-width arithmetic, lead to the decidability of trSAT due to their limited attention capabilities. However, the problem remains difficult, as we establish scenarios where trSAT is NEXPTIME-hard and others where it is solvable in NEXPTIME for quantized TE. To complement our complexity results, we place our findings and their implications in the broader context of formal reasoning.", "keywords": [ "transformer", "formal reasoning", "complexity" ], "rating_list": [ 8, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 2, 5, 4, 5 ] }, { "paper_id": "VU4WuN0zwV", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Hyunwoo Park", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" }, { "name": "Joo Young Park", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Undergrad student" }, { "name": "Sejun Park", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Undergrad student" } ], "n_formula": 121, "n_formula_1": 28, "n_ref_uni": 33, "n_ref": 52, "n_ref_all": 61, "n_fig": 4, "n_tab": 14, "L_tab": 3665, "n_element_tab": 463, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 369, "n_element_tab_1": 63, "formula_len_all": 14587, "formula_len_all_1": 2526, "len_all": 206087, "len_all_1": 70139, "len_abs": 1459, "len_title": 140, "len_sents": 63097, "len_sents_1": 29675, "n_sents": 612, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1470, "title": "IMPaCT GNN: Imposing invariance with Message Passing in Chronological split Temporal Graphs", "abs": "This paper addresses domain adaptation challenges in graph data resulting from chronological splits. In a transductive graph learning setting, where each node is associated with a timestamp, we focus on the task of Semi-Supervised Node Classification (SSNC), aiming to classify recent nodes using labels of past nodes. Temporal dependencies in node connections create domain shifts, causing significant performance degradation when applying models trained on historical data into recent data. Given the practical relevance of this scenario, addressing domain adaptation in chronological split data is crucial, yet underexplored. We propose Imposing invariance with Message Passing in Chronological split Temporal Graphs (\\IMPaCT), a method that imposes invariant properties based on realistic assumptions derived from temporal graph structures. Unlike traditional domain adaptation approaches which rely on unverifiable assumptions, \\IMPaCT explicitly accounts for the characteristics of chronological splits. The \\IMPaCT is further supported by rigorous mathematical analysis, including a derivation of an upper bound of the generalization error. Experimentally, \\IMPaCT achieves a 3.8\\% performance improvement over current SOTA method on the ogbn-mag graph dataset. Additionally, we introduce the Temporal Stochastic Block Model (TSBM), which replicates temporal graphs under varying conditions, demonstrating the applicability of our methods to general spatial GNNs.", "keywords": [ "Graph Neural Networks", "Domain Adaptation", "Distribution Shift", "Temporal Graph", "Chronological Split", "Semi-Supervised Node Classification", "Generalization Bound" ], "rating_list": [ 5, 6, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 1, 3, 2, 1, 2 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 5, 4 ] }, { "paper_id": "VTG68CNUCf", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 15, "author_info_list": [ { "name": "Bin Cui", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Fan Yang", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Guosheng Dong", "gender": "unknown", "institution": "Baichuan Inc.", "country": "", "position": "Researcher" }, { "name": "Hao Liang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Miao Zheng", "gender": "Female", "institution": "Baichuan", "country": "CN", "position": "Researcher" }, { "name": "Mingan Lin", "gender": "unknown", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Intern" }, { "name": "Tianpeng Li", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yan Zhang", "gender": "unknown", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yanjun Shen", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zenan Zhou", "gender": "Male", "institution": "baichuan", "country": "CN", "position": "Researcher" }, { "name": "sunhaoze", "gender": "Male", "institution": "Baichuan inc", "country": "CN", "position": "Researcher" }, { "name": "weipeng chen", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "fangkun", "gender": "unknown", "institution": "baichuan", "country": "", "position": "Researcher" }, { "name": "Zhangtao", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zhang wen tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 24, "n_ref": 64, "n_ref_all": 97, "n_fig": 11, "n_tab": 8, "L_tab": 6297, "n_element_tab": 417, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3336, "n_element_tab_1": 335, "formula_len_all": 64, "formula_len_all_1": 40, "len_all": 175963, "len_all_1": 66087, "len_abs": 1427, "len_title": 93, "len_sents": 60344, "len_sents_1": 28114, "n_sents": 466, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 45, "L_abs": 1359, "title": "PAS: Plug-and-Play Prompt Augmentation System", "abs": "In recent years, the rise of Large Language Models (LLMs) has spurred a growing demand for plug-and-play AI systems. Among the various AI techniques, prompt engineering stands out as particularly significant. However, users often face challenges in writing prompts due to the steep learning curve and significant time investment, and existing automatic prompt engineering (APE) models can be difficult to use. To address this issue, we propose PAS, an LLM-based plug-and-play APE system.\nPAS utilizes LLMs trained on high-quality, automatically generated prompt complementary datasets, resulting in exceptional performance. In comprehensive benchmarks, PAS achieves state-of-the-art (SoTA) results compared to previous APE models, with an average improvement of 6.09 points. Moreover, PAS is highly efficient, achieving SoTA performance with only 9000 data points. Additionally, PAS can autonomously generate prompt augmentation data without requiring additional human labor. Its flexibility also allows it to be compatible with all existing LLMs and applicable to a wide range of tasks.\nPAS excels in human evaluations, underscoring its suitability as a plug-in for users. This combination of high performance, efficiency, and flexibility makes PAS a valuable system for enhancing the usability and effectiveness of LLMs through automatic prompt engineering.", "keywords": [ "Auto Prompt Engineering", "Plug-and-Play System", "Data Curation" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "VT4Ovqg0BW", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Dongxia Wu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Ph.D student" }, { "name": "Rose Yu", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Ruijia Niu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Yian Ma", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 65, "n_ref": 116, "n_ref_all": 132, "n_fig": 1, "n_tab": 5, "L_tab": 2486, "n_element_tab": 233, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 2438, "n_element_tab_1": 135, "formula_len_all": 972, "formula_len_all_1": 871, "len_all": 208200, "len_all_1": 67049, "len_abs": 1171, "len_title": 126, "len_sents": 37003, "len_sents_1": 30796, "n_sents": 248, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1176, "title": "Functional-level Uncertainty Quantification for Calibrated Fine-tuning on LLMs", "abs": "From common-sense reasoning to domain-specific tasks, parameter-efficient fine tuning (PEFT) methods for large language models (LLMs) have showcased significant performance improvements on downstream tasks. However, fine-tuned LLMs often struggle with overconfidence in uncertain predictions, particularly due to sparse training data. This overconfidence reflects poor epistemic uncertainty calibration, which arises from limitations in the model's ability to generalize with limited data. Existing PEFT uncertainty quantification methods for LLMs focus on the post fine-tuning stage and thus have limited capability in calibrating epistemic uncertainty. To address these limitations, we propose Functional-Level Uncertainty Quantification for Calibrated Fine-Tuning (UQ4CT), which captures and calibrates functional-level epistemic uncertainty during the fine-tuning stage via a mixture-of-expert framework. We show that UQ4CT reduces Expected Calibration Error (ECE) by more than 25% while maintaining high accuracy across 5 benchmarks. Furthermore, UQ4CT maintains superior ECE performance with high accuracy under distribution shift, showcasing improved generalizability.", "keywords": [ "Uncertainty Quantification", "Large Language Models", "Mixture of Experts", "Parameter Efficient Fine Tuning" ], "rating_list": [ 5, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "VT2R3UCcBL", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Carlee Joe-Wong", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Chaoyang He", "gender": "Male", "institution": "FedML, Inc.", "country": "", "position": "CTO" }, { "name": "Jiajun Gu", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Amir Avestimehr", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "Full Professor" }, { "name": "Shanshan Han", "gender": "Female", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Srivatsan Ravi", "gender": "unknown", "institution": "University of Southern California", "country": "", "position": "Assistant Professor" }, { "name": "Weizhao Jin", "gender": "unknown", "institution": "AWS", "country": "US", "position": "Applied Scientist" }, { "name": "Yuhang Yao", "gender": "unknown", "institution": "TensorOpera Inc", "country": "US", "position": "Researcher" } ], "n_formula": 31, "n_formula_1": 10, "n_ref_uni": 51, "n_ref": 96, "n_ref_all": 124, "n_fig": 16, "n_tab": 12, "L_tab": 3133, "n_element_tab": 402, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1079, "n_element_tab_1": 129, "formula_len_all": 1360, "formula_len_all_1": 377, "len_all": 183270, "len_all_1": 61705, "len_abs": 1040, "len_title": 83, "len_sents": 55255, "len_sents_1": 29511, "n_sents": 383, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1032, "title": "Efficient Privacy-Preserving Federated Learning With Selective Parameter Encryption", "abs": "Federated learning trains machine learning models on distributed devices by aggregating local model updates instead of local data. However, privacy concerns arise as aggregating local model updates on the server may reveal sensitive personal information by inversion attacks. Privacy-preserving methods, such as homomorphic encryption (HE), then become necessary for FL training. Despite HE's privacy advantages, its applications suffer from impractical overheads, especially for foundation models. In this paper, we present the first practical privacy-preserving federated learning work with efficient HE-based secure model aggregation. Our approach proposes to selectively encrypt sensitive parameters, significantly reducing both computation and communication overheads during training while providing quantifiable privacy guarantee. Our optimization shows considerable overhead reduction, particularly for large foundation models (e.g. \n100x reduction for GPT-2), demonstrating the potential for scalable HE-based FL deployment.", "keywords": [ "Federated learning", "privacy", "homomorphic encryption", "inversion attack" ], "rating_list": [ 5, 6, 8, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "VSogkPlqDS", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Pierre Quinton", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Val\u00e9rian Rey", "gender": "Not Specified", "institution": "Withings", "country": "FR", "position": "Researcher" } ], "n_formula": 96, "n_formula_1": 24, "n_ref_uni": 46, "n_ref": 86, "n_ref_all": 121, "n_fig": 8, "n_tab": 3, "L_tab": 615, "n_element_tab": 110, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 64, "n_element_tab_1": 10, "formula_len_all": 11263, "formula_len_all_1": 2001, "len_all": 206594, "len_all_1": 69960, "len_abs": 1205, "len_title": 97, "len_sents": 58351, "len_sents_1": 29539, "n_sents": 641, "n_sents_1": 270, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1213, "title": "Jacobian Descent for Multi-Objective Optimization", "abs": "Many optimization problems require balancing multiple conflicting objectives.\nAs gradient descent is limited to single-objective optimization, we introduce its direct generalization: Jacobian descent (JD).\nThis algorithm iteratively updates parameters using the Jacobian matrix of a vector-valued objective function, in which each row is the gradient of an individual objective.\nWhile several methods to combine gradients already exist in the literature, they are generally hindered when the objectives conflict.\nIn contrast, we propose projecting gradients to fully resolve conflict while ensuring that they preserve an influence proportional to their norm.\nWe prove significantly stronger convergence guarantees with this approach, supported by our empirical results.\nOur method also enables instance-wise risk minimization (IWRM), a novel learning paradigm in which the loss of each training example is considered a separate objective.\nApplied to simple image classification tasks, IWRM exhibits promising results compared to the direct minimization of the average loss.\nAdditionally, we outline an efficient implementation of JD using the Gramian of the Jacobian matrix to reduce time and memory requirements.", "keywords": [ "multi-objective optimization", "optimization", "statistical learning theory", "machine learning", "deep learning", "multi-task learning" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "VSklRu8KTH", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 1, "author_info_list": [ { "name": "Sammy Khalife", "gender": "Not Specified", "institution": "Cornell University", "country": "US", "position": "Postdoc" } ], "n_formula": 64, "n_formula_1": 38, "n_ref_uni": 10, "n_ref": 23, "n_ref_all": 31, "n_fig": 1, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5748, "formula_len_all_1": 3104, "len_all": 156878, "len_all_1": 69173, "len_abs": 3939, "len_title": 43, "len_sents": 55584, "len_sents_1": 30336, "n_sents": 540, "n_sents_1": 284, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 1177, "title": "The logic of rational graph neural networks", "abs": "The expressivity of Graph Neural Networks (GNNs) can be described via appropriate fragments of the first-order logic. In this context, uniform expressivity guarantees that a GNN can express a logical query without the parameters depending on the size of the input graphs. It has been established that the two-variable guarded fragment with counting (GC2) can be expressed uniformly via Rectified Linear Unit (ReLU) GNNs [Barcelo &. Al., 2020]. Moreover, GC2 is the fragment that can be expressed at most by a GNN with any activation function. In this article, we prove that, on the contrary of ReLU GNNs, there are GC2 queries that cannot be uniformly expressed via any GNN with rational activations. As a consequence, non-polynomial activation functions do not grant GNNs GC2 uniform expressivity in general, answering an open question formulated by [Grohe, 2021]. We then present a strict subfragment of GC2 (RGC2), and prove that rational GNNs can express RGC2 queries uniformly over all graphs. Our numerical experiments illustrates that despite this theoretical disadvantage, rational GNNs are still able to learn some GC2 queries if some level of error is allowed.", "keywords": [ "Graph Neural Networks", "Rational activations", "Expressivity", "Logic" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 1, 3, 3, 4 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "VSidzaTzpd", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Bin Xiao", "gender": "Male", "institution": "Chongqing University of Posts and Tel.", "country": "CN", "position": "Full Professor" }, { "name": "Haoxiang Yuan", "gender": "Male", "institution": "Chongqing University of Post and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Jianxin Li", "gender": "Male", "institution": "Beihang University ", "country": "CN", "position": "Full Professor" }, { "name": "Qingyun Sun", "gender": "Female", "institution": "Beihang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yang Hu", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 12, "n_ref_uni": 17, "n_ref": 41, "n_ref_all": 64, "n_fig": 7, "n_tab": 7, "L_tab": 1814, "n_element_tab": 186, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 3106, "n_element_tab_1": 188, "formula_len_all": 815, "formula_len_all_1": 613, "len_all": 93430, "len_all_1": 56696, "len_abs": 1397, "len_title": 124, "len_sents": 25471, "len_sents_1": 22912, "n_sents": 180, "n_sents_1": 162, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1403, "title": "Activating More Advantageous Neurons Can Improve Adversarial Transferability", "abs": "Deep Neural Networks (DNNs) are vulnerable to unseen noise, lighting the need to identify the deficiencies of DNNs to mitigate this vulnerability. In the field of adversarial attacks, existing works investigate the deficiencies causing the vulnerability of DNNs, quantifying the vulnerability of DNNs and demonstrating the transferability of adversarial examples where adversarial examples crafted for one model can deceive another. Among the related works, adversarial transferability attracts much attention since transferable adversarial examples enable black-box attacks and raise concerns about DNNs. Although various novel adversarial attacks are presented to improve the adversarial transferability, the property of DNNs that leads to the improvements remains unidentified. This work delves into this issue and reveals that different benign input with different features activates mostly different neurons in a model, and the model may be viewed as an ensemble including different submodels capturing different features. Therefore, an adversarial attack can activate more neurons to generate the adversarial examples, thus probably making the examples applicable to diverse models to enhance the adversarial transferability. Also, data transformation can help exclude wrong answers to boost the adversarial example. The extensive experiments demonstrate the soundness and superiority of our work.", "keywords": [ "adversarial attacks" ], "rating_list": [ 5, 5, 3, 8 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "VSfvQxPPB0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Boyu Li", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Dongbin Zhao", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Haobin Jiang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Haoran Li", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Xinrun Xu", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Ziluo Ding", "gender": "Male", "institution": "BAAI", "country": "CN", "position": "Researcher" }, { "name": "Zongqing Lu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 54, "n_ref": 83, "n_ref_all": 108, "n_fig": 7, "n_tab": 8, "L_tab": 1649, "n_element_tab": 148, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1219, "n_element_tab_1": 101, "formula_len_all": 132, "formula_len_all_1": 132, "len_all": 175617, "len_all_1": 65960, "len_abs": 1250, "len_title": 107, "len_sents": 40965, "len_sents_1": 30382, "n_sents": 329, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1254, "title": "SELU: Self-Learning Embodied MLLMs in Unknown Environments", "abs": "Recently, multimodal large language models (MLLMs) have demonstrated strong visual understanding and decision-making capabilities, enabling the exploration of autonomously improving MLLMs in unknown environments. However, external feedback like human or environmental feedback is not always available. To address this challenge, existing methods primarily focus on enhancing the decision-making capabilities of MLLMs through voting and scoring mechanisms, while little effort has been paid to improving the environmental comprehension of MLLMs in unknown environments. To fully unleash the self-learning potential of MLLMs, we propose a novel actor-critic self-learning paradigm, dubbed SELU, inspired by the actor-critic paradigm in reinforcement learning. The critic employs self-asking and hindsight relabeling to extract knowledge from interaction trajectories collected by the actor, thereby augmenting its environmental comprehension. Simultaneously, the actor is improved by the self-feedback provided by the critic, enhancing its decision-making. We evaluate our method in the AI2-THOR and VirtualHome environments, and SELU achieves critic improvements of approximately 28% and 30%, and actor improvements of about 20% and 24% via self-learning.", "keywords": [ "embodied MLLM", "self-learning", "actor-critic" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "VSVQljJU5N", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Alexander Tarakanov", "gender": "Male", "institution": "Higher School of Economics", "country": "RU", "position": "Assistant Professor" }, { "name": "Ermilov Sergei", "gender": "Male", "institution": "VK", "country": "RU", "position": "Researcher" }, { "name": "Ivan Betev", "gender": "Male", "institution": "Higher School of Economics", "country": "", "position": "Undergrad student" }, { "name": "Petr Pogorelov", "gender": "Male", "institution": "VK", "country": "RU", "position": "Researcher" } ], "n_formula": 42, "n_formula_1": 37, "n_ref_uni": 21, "n_ref": 31, "n_ref_all": 33, "n_fig": 0, "n_tab": 1, "L_tab": 403, "n_element_tab": 88, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 439, "n_element_tab_1": 93, "formula_len_all": 1810, "formula_len_all_1": 1711, "len_all": 94865, "len_all_1": 63437, "len_abs": 721, "len_title": 121, "len_sents": 27349, "len_sents_1": 22663, "n_sents": 249, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 724, "title": "Diffusion Minimization and Sheaf Neural Networks for Recommender Systems", "abs": "Graph Neural Networks (GNN) are well-known for successful applications in recommender systems. Despite recent advances in GNN development, various authors report that in certain cases GNN suffer from so-called oversmoothing problems. Sheaf Neural Networks (SNN) is one of the ways to address the issue of oversmoothing. In the present work we propose a novel approach for training SNN together with user and item embeddings. In that approach parameters of the sheaf are inferred via minimization of the classical BPR loss and sheaf diffusion on graphs subjected to orthogonality and consistency constraints. Performance of the novel technique is evaluated on synthetic test cases and standard benchmarks for recommendations.", "keywords": [ "Graph Neural Networks", "Sheaves", "Oversmoothing" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 1, 2, 2 ], "presentation_list": [ 1, 1, 2, 2 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "VSKV3GykuE", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Grigory Malinovsky", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Hasan Abed Al Kader Hammoud", "gender": "Male", "institution": "KAUST", "country": "", "position": "PhD student" }, { "name": "Taha Ceritli", "gender": "Male", "institution": "Samsung Research UK", "country": "GB", "position": "Researcher" }, { "name": "Umberto Michieli", "gender": "Male", "institution": "Samsung", "country": "GB", "position": "Researcher" }, { "name": "Bryn Elesedy", "gender": "Male", "institution": "Samsung", "country": "GB", "position": "Researcher" }, { "name": "Mete Ozay", "gender": "Male", "institution": "Samsung Research", "country": "GB", "position": "Principal Researcher" }, { "name": "Peter Richtarik", "gender": "Male", "institution": "King Abdullah University of Science and Technology (KAUST)", "country": "SA", "position": "Full Professor" } ], "n_formula": 129, "n_formula_1": 25, "n_ref_uni": 49, "n_ref": 86, "n_ref_all": 116, "n_fig": 4, "n_tab": 6, "L_tab": 1451, "n_element_tab": 100, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 130, "n_element_tab_1": 12, "formula_len_all": 17654, "formula_len_all_1": 1747, "len_all": 209025, "len_all_1": 64692, "len_abs": 1533, "len_title": 118, "len_sents": 53646, "len_sents_1": 28592, "n_sents": 526, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1542, "title": "RAC-LoRA: A Theoretical Optimization Framework for Low-Rank Adaptation", "abs": "Fine-tuning has become a popular approach to adapting large foundational models to specific tasks. As the size of models and datasets grows, parameter-efficient fine-tuning techniques are increasingly important. One of the most widely used methods is Low-Rank Adaptation (LoRA), with adaptation update expressed as the product of two low-rank matrices. While LoRA was shown to possess strong performance in fine-tuning, it often underperforms when compared to full-parameter fine-tuning (FPFT). Although many variants of LoRA have been extensively studied empirically, their theoretical optimization analysis is heavily under-explored. The starting point of our work is a demonstration that LoRA and its two extensions, Asymmetric LoRA and Chain of LoRA, indeed encounter convergence issues. To address these issues, we propose a general optimization framework that rigorously analyzes the convergence rates of LoRA-based methods. Our approach inherits the empirical benefits of LoRA-style heuristics, but introduces several small but important algorithmic modifications which turn it into a provably convergent method. Our framework serves as a bridge between FPFT and low-rank adaptation. We provide provable guarantees of convergence to the same solution as FPFT, along with the rate of convergence. Additionally, we present a convergence analysis for smooth, non-convex loss functions, covering gradient descent, stochastic gradient descent, and federated learning settings. Our theoretical findings are supported by experimental results.", "keywords": [ "LORA", "optimization", "stochastic optimization", "low-rank adaptation" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 2, 3 ] }, { "paper_id": "VSHuwBUlYr", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Abdelrahman Eldesokey", "gender": "unknown", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" }, { "name": "Adam Kortylewski", "gender": "unknown", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Research Group Leader" }, { "name": "Christian Theobalt", "gender": "Male", "institution": "Max-Planck-Institute for Informatics, Saarland Informatics Campus", "country": "DE", "position": "Director" }, { "name": "Fangneng Zhan", "gender": "Male", "institution": "Harvard University & MIT", "country": "US", "position": "Postdoc" }, { "name": "Mohit Mendiratta", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Peter Wonka", "gender": "Male", "institution": "Snap Inc.", "country": "SA", "position": "Researcher" }, { "name": "Qian Wang", "gender": "Female", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 27, "n_ref": 63, "n_ref_all": 87, "n_fig": 14, "n_tab": 6, "L_tab": 2067, "n_element_tab": 402, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 300, "n_element_tab_1": 88, "formula_len_all": 699, "formula_len_all_1": 340, "len_all": 128311, "len_all_1": 61133, "len_abs": 280, "len_title": 123, "len_sents": 38911, "len_sents_1": 27633, "n_sents": 326, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1645, "title": "Zero-Shot Video Semantic Segmentation based on Pre-Trained Diffusion Models", "abs": "We introduce the first zero-shot approach for Video Semantic Segmentation (VSS) based on pre-trained diffusion models. A growing research direction attempts to employ diffusion models to perform downstream vision tasks by exploiting their deep understanding of image semantics. Yet, the majority of these approaches have focused on image-related tasks like semantic segmentation, with less emphasis on video tasks such as VSS. Ideally, diffusion-based image semantic segmentation approaches can be applied to videos in a frame-by-frame manner. However, we find their performance on videos to be subpar due to the absence of any modeling of temporal information inherent in the video data. To this end, we tackle this problem and introduce a framework tailored for VSS based on pre-trained image and video diffusion models. We propose building a scene context model based on the diffusion features, where the model is autoregressively updated to adapt to scene changes. This context model predicts per-frame coarse segmentation maps that are temporally consistent. To refine these maps further, we propose a correspondence-based refinement strategy that aggregates predictions temporally, resulting in more confident predictions. Finally, we introduce a masked modulation approach to upsample the coarse maps to a high-quality full resolution. Experiments show that our proposed approach significantly outperforms existing zero-shot image semantic segmentation approaches on various VSS benchmarks without any training or fine-tuning. Moreover, it rivals supervised VSS approaches on the VSPW dataset despite not being explicitly trained for VSS.", "keywords": [ "Diffusion models", "video segmentation" ], "rating_list": [ 6, 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 4, 1 ], "confidence_list": [ 4, 3, 5, 4, 3 ] }, { "paper_id": "VRlihVklCL", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Kookjin Lee", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" }, { "name": "Mingu Kang", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Undergrad student" }, { "name": "Noseong Park", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Woojin Cho", "gender": "Male", "institution": "Telepix", "country": "KR", "position": "Researcher" }, { "name": "Youngjoon Hong", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" }, { "name": "Anthony D. Gruber", "gender": "Male", "institution": "Sandia National Laboratories", "country": "US", "position": "Researcher" }, { "name": "Dong seok Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "JAEHYEON PARK", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 38, "n_formula_1": 20, "n_ref_uni": 42, "n_ref": 65, "n_ref_all": 89, "n_fig": 3, "n_tab": 17, "L_tab": 5202, "n_element_tab": 510, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 2189, "n_element_tab_1": 170, "formula_len_all": 3197, "formula_len_all_1": 916, "len_all": 196934, "len_all_1": 58786, "len_abs": 1354, "len_title": 166, "len_sents": 54667, "len_sents_1": 25500, "n_sents": 407, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 117, "L_abs": 1342, "title": "MaD-Scientist: AI-based Scientist solving Convection-Diffusion-Reaction Equations Using Massive PINN-Based Prior Data", "abs": "Large language models (LLMs), like ChatGPT, have shown that even trained with noisy prior data, they can generalize effectively to new tasks through in-context learning (ICL) and pre-training techniques.\nMotivated by this, we explore whether a similar approach can be applied to scientific foundation models (SFMs). Our methodology is structured as follows: (i) we collect low-cost physics-informed neural network (PINN)-based approximated prior data in the form of solutions to partial differential equations (PDEs) constructed through an arbitrary linear combination of mathematical dictionaries; (ii) we utilize Transformer architectures with self and cross-attention mechanisms to predict PDE solutions without knowledge of the governing equations in a zero-shot setting; (iii) we provide experimental evidence on the one-dimensional convection-diffusion-reaction equation, which demonstrate that pre-training remains robust even with approximated prior data, with only marginal impacts on test accuracy. Notably, this finding opens the path to pre-training SFMs with realistic, low-cost data instead of (or in conjunction with) numerical high-cost data. These results support the conjecture that SFMs can improve in a manner similar to LLMs, where fully cleaning the vast set of sentences crawled from the Internet is nearly impossible.", "keywords": [ "in-context learning", "scientific foundation model", "zero-shot", "PINN-prior" ], "rating_list": [ 5, 5, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 3, 2, 3 ] }, { "paper_id": "VRbypIkXrt", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Arsalan Sharifnassab", "gender": "Male", "institution": "Openmind Research Institute", "country": "US", "position": "Researcher" }, { "name": "Saber Salehkaleybar", "gender": "Male", "institution": "Leiden University, Leiden University", "country": "NL", "position": "Assistant Professor" }, { "name": "Richard S. Sutton", "gender": "Male", "institution": "Keen Technologies", "country": "CA", "position": "Research Scientist" } ], "n_formula": 114, "n_formula_1": 19, "n_ref_uni": 56, "n_ref": 85, "n_ref_all": 135, "n_fig": 8, "n_tab": 4, "L_tab": 1250, "n_element_tab": 117, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 309, "n_element_tab_1": 34, "formula_len_all": 11054, "formula_len_all_1": 2560, "len_all": 180817, "len_all_1": 64140, "len_abs": 957, "len_title": 125, "len_sents": 49689, "len_sents_1": 28024, "n_sents": 411, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 960, "title": "MetaOptimize: A Framework for Optimizing Step Sizes and Other Meta-parameters", "abs": "We address the challenge of optimizing meta-parameters (i.e., hyperparameters) in machine learning algorithms, a critical factor influencing training efficiency and model performance. Moving away from the computationally expensive traditional meta-parameter search methods, we introduce MetaOptimize framework that dynamically adjusts meta-parameters, particularly step sizes (also known as learning rates), during training. More specifically, MetaOptimize can wrap around any first-order optimization algorithm, tuning step sizes on the fly to minimize a specific form of regret that accounts for long-term effect of step sizes on training, through a discounted sum of future losses. We also introduce low complexity variants of MetaOptimize that, in conjunction with its adaptability to multiple optimization algorithms, demonstrate performance competitive to those of best hand-crafted learning-rate schedules across various machine learning applications.", "keywords": [ "Optimization", "Automatic step-size optimization", "Automatic hyperparameter optimization", "Continual learning" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 3, 3 ] }, { "paper_id": "VRYJXoUjRS", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Chinmay Savadikar", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "PhD student" }, { "name": "Michelle Dai", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Tianfu Wu", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 11, "n_formula_1": 13, "n_ref_uni": 58, "n_ref": 151, "n_ref_all": 201, "n_fig": 12, "n_tab": 12, "L_tab": 5144, "n_element_tab": 388, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1957, "n_element_tab_1": 226, "formula_len_all": 548, "formula_len_all_1": 559, "len_all": 228000, "len_all_1": 75640, "len_abs": 1501, "len_title": 118, "len_sents": 53785, "len_sents_1": 34227, "n_sents": 341, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1518, "title": "Continual Learning via Learning a Continual Memory in Vision Transformer", "abs": "This paper explores continual learning (CL) using Vision Transformer (ViT) in streaming tasks under the challenging exemplar-free class-incremental (ExfCCL) setting. We formulate ExfCCL as a learning problem consisting of two key sub-systems: (i) task ID inference for test data, which selects appropriate task-specific head classifiers to accounting for varying class distributions across tasks and streams, and (ii) a dynamic learning-to-grow feature backbone that balances stability and plasticity, mitigating catastrophic forgetting through task synergies. Following the common protocol that the first task can train a ViT sufficiently well as the base model, we address these sub-systems from a continual memory learning perspective. To support task ID inference, we utilize an external memory mechanism that maintains task centroids computed by the base ViT throughout CL. For the feature backbone, we identify optimal placements for internal (parameter) memory to enable a dynamic, task-synergy guided growing feature backbone. We propose a Hierarchical Exploration-Exploitation (HEE) sampling-based neural architecture search (NAS) method that effectively learns task synergies by continually and structurally updating internal memory with four basic operations: _reuse_, _adapt_, _new_, and _skip_. Our approach, dubbed **Continual Hierarchical-Exploration-Exploitation Memory (CHEEM)**, is evaluated on the challenging Visual Domain Decathlon (VDD) and ImageNet-R benchmarks, demonstrating its effectiveness.", "keywords": [ "Lifelong Learning", "Continual Learning", "Vision Transformers" ], "rating_list": [ 3, 3, 8, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "VRTCXYvPxc", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Hongxiang Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Jindong Gu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Junjie Cao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Li KaiZhou", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xiao-Ping Zhang", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xinchun Yu", "gender": "Male", "institution": "Zhejiang Gongshang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yansong Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 27, "n_formula_1": 13, "n_ref_uni": 36, "n_ref": 68, "n_ref_all": 87, "n_fig": 9, "n_tab": 2, "L_tab": 848, "n_element_tab": 65, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1092, "n_element_tab_1": 71, "formula_len_all": 1599, "formula_len_all_1": 893, "len_all": 137998, "len_all_1": 62859, "len_abs": 1836, "len_title": 110, "len_sents": 38642, "len_sents_1": 29621, "n_sents": 307, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1845, "title": "VIDEOGUARD: PROTECTING VIDEO CONTENT FROM UNAUTHORIZED EDITING", "abs": "With the rapid development of generative technology, current generative models can generate high-fidelity digital content and edit it in a controlled manner. However, there is a risk that malicious individuals might misuse these capabilities for misleading or unlawful activities. Although existing research has attempted to shield photographic images from being manipulated by generative models, there remains a significant disparity in the protection offered to video content editing. To bridge the gap, we propose a protection method named VideoGuard, which can effectively protect videos from unauthorized malicious editing. This protection is achieved through the subtle introduction of nearly unnoticeable perturbation that interferes with the functioning of the intended generative diffusion models. Different from images, videos consist of sequential frames, containing not only visual content but also motion dynamics. Due to the redundancy between video frames, and inter-frame attention mechanism in video diffusion models, simply applying image-based protection methods separately to every video frame can not shield video from unauthorized editing. To tackle the above challenge, rather than optimize perturbation in a frame-wise manner like image-based methods, we adopt joint frame optimization, treating all the video frames as an optimization entity. Furthermore, we extract video motion information and fuse it into optimization objectives. Thereby, these alterations can effectively compel the models to produce outputs that are implausible and inconsistent. We provide a pipeline to optimize such a perturbation. Finally, we use both objective metrics and subjective metrics to demonstrate the efficacy of our method, and the results show that the protection performance of VideoGuard is superior to all the baseline methods.", "keywords": [ "Diffusion Models", "Video Editing Protection" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "VRRuYBaq9u", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Guangming Xie", "gender": "Male", "institution": "Peking University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "YH L", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 45, "n_formula_1": 17, "n_ref_uni": 45, "n_ref": 80, "n_ref_all": 108, "n_fig": 14, "n_tab": 6, "L_tab": 1233, "n_element_tab": 172, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3956, "formula_len_all_1": 1691, "len_all": 162056, "len_all_1": 62330, "len_abs": 3028, "len_title": 123, "len_sents": 48618, "len_sents_1": 29727, "n_sents": 397, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1020, "title": "Leveraging Additional Information in POMDPs with Guided Policy Optimization", "abs": "Reinforcement Learning (RL) in partially observable environments poses significant challenges due to the complexity of learning under uncertainty. \nWhile additional information, such as that available in simulations, can enhance training, effectively leveraging it remains an open problem. \nTo address this, we introduce Guided Policy Optimization (GPO), a framework that co-trains a guider and a learner. \nThe guider takes advantage of supplementary information while ensuring alignment with the learner's policy, which is primarily trained via Imitation Learning (IL). \nWe theoretically demonstrate that this learning scheme achieves optimality comparable to direct RL, thereby overcoming key limitations inherent in IL approaches. \nOur approach includes two practical variants, GPO-penalty and GPO-clip, and empirical evaluations show strong performance across various tasks, including continuous control with partial observability and noise, and memory-based challenges, significantly outperforming existing methods.", "keywords": [ "Reinforcement Learning", "Imitation Learning", "POMDP", "policy gradient" ], "rating_list": [ 3, 6, 1, 3 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "VRFotuGLfM", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Cheng Long", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Kaijun Liu", "gender": "Male", "institution": "NTU", "country": "SG", "position": "PhD student" }, { "name": "Shuliang Wang", "gender": "unknown", "institution": "Beijing Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Sijie Ruan", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "ZHANG LIANG", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "liang yu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 36, "n_ref": 77, "n_ref_all": 112, "n_fig": 7, "n_tab": 6, "L_tab": 1949, "n_element_tab": 140, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 937, "n_element_tab_1": 43, "formula_len_all": 1161, "formula_len_all_1": 985, "len_all": 143564, "len_all_1": 72116, "len_abs": 3509, "len_title": 115, "len_sents": 45873, "len_sents_1": 31849, "n_sents": 308, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1329, "title": "DiffMove: Human Trajectory Recovery via Conditional Diffusion Model", "abs": "Recovering human trajectories from incomplete or missing data is crucial for many mobility-based urban applications, e.g., urban planning, transportation, and location-based services. Existing methods mainly rely on recurrent neural networks or attention mechanisms. Though promising, they encounter limitations in capturing complex spatial-temporal dependencies in low-sampling trajectories. Recently, diffusion models show potential in content generation. However, most of proposed methods are used to generate contents in continuous numerical representations, which cannot be directly adapted to the human location trajectory recovery. In this paper, we introduce a conditional diffusion-based trajectory recovery method, namely, DiffMove. It first transforms locations in trajectories into the embedding space, in which the embedding denoising is performed, and then missing locations are recovered by an embedding decoder. DiffMove not only improves accuracy by introducing high-quality generative methods in the trajectory recovery, but also carefully models the transition, periodicity, and temporal patterns in human mobility. Extensive experiments based on two representative real-world mobility datasets are conducted, and the results show significant improvements (an average of 11% in recall) over the best baselines.", "keywords": [ "Trajectory recovery", "Diffusion model", "Self-supervised learning", "Human mobility" ], "rating_list": [ 5, 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 4, 5, 3, 5, 4 ] }, { "paper_id": "VR9RS7tZGG", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Byung-Jun Lee", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Assistant Professor" }, { "name": "Jaehyun Park", "gender": "unknown", "institution": "Gwangju Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sejin Kim", "gender": "Male", "institution": "Gwangju Institute of Science and Technology", "country": "KR", "position": "Postdoc" }, { "name": "Sundong Kim", "gender": "Male", "institution": "Gwangju Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Yunho Kim", "gender": "Male", "institution": "Gwangju Institute of Science and Technology", "country": "", "position": "Undergrad student" }, { "name": "HeeJunKim", "gender": "Male", "institution": "Gwangju Institute of Science and Technology", "country": "KR", "position": "Undergrad student" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 21, "n_ref": 35, "n_ref_all": 51, "n_fig": 12, "n_tab": 1, "L_tab": 133, "n_element_tab": 16, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1499, "n_element_tab_1": 42, "formula_len_all": 343, "formula_len_all_1": 0, "len_all": 103490, "len_all_1": 59430, "len_abs": 1482, "len_title": 125, "len_sents": 45079, "len_sents_1": 29544, "n_sents": 318, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1488, "title": "Diffusion-Based Offline RL for Improved Decision-Making in Augmented Single ARC Task", "abs": "Effective long-term strategies enable AI systems to navigate complex environments by making sequential decisions over extended horizons. Similarly, reinforcement learning (RL) agents optimize decisions across sequences to maximize rewards, even without immediate feedback. To verify that Latent Diffusion-Constrained Q-learning (LDCQ), a prominent diffusion-based offline RL method, demonstrates strong reasoning abilities in multi-step decision-making, we aimed to evaluate its performance on the Abstraction and Reasoning Corpus (ARC). However, applying offline RL methodologies to enhance strategic reasoning in AI for solving tasks in ARC is challenging due to the lack of sufficient experience data in the ARC training set. To address this limitation, we introduce an augmented offline RL dataset for ARC, called Synthesized Offline Learning Data for Abstraction and Reasoning (SOLAR), along with the SOLAR-Generator, which generates diverse trajectory data based on predefined rules. SOLAR enables the application of offline RL methods by offering sufficient experience data. We synthesized SOLAR for a simple task and used it to train an agent with the LDCQ method. Our experiments demonstrate the effectiveness of the offline RL approach on a simple ARC task, showing the agent's ability to make multi-step sequential decisions and correctly identify answer states. These results highlight the potential of the offline RL approach to enhance AI's strategic reasoning capabilities.", "keywords": [ "System-2 Reasoning", "Reasoning", "Abstraction", "Diffusion model", "offline rl" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "VQwI055flA", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 8, "author_info_list": [ { "name": "Aoxue Li", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Haiming Wang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Hui Jin", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zhenguo Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Zhi Zhou", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "wenda wei", "gender": "unknown", "institution": "nanjing university", "country": "", "position": "PhD student" }, { "name": "Xiao-wen Yang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Yu-Feng Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 29, "n_ref": 110, "n_ref_all": 124, "n_fig": 6, "n_tab": 4, "L_tab": 333, "n_element_tab": 36, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1143, "n_element_tab_1": 111, "formula_len_all": 459, "formula_len_all_1": 538, "len_all": 128804, "len_all_1": 58328, "len_abs": 1243, "len_title": 154, "len_sents": 38957, "len_sents_1": 26375, "n_sents": 331, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 106, "L_abs": 1251, "title": "CARTS: Advancing Neural Theorem Proving with Diversified Tactic Calibration and Bias-Resistant Tree Search", "abs": "Recent advancements in neural theorem proving integrate large language models with tree search algorithms like Monte Carlo Tree Search (MCTS), where the language model suggests tactics and the tree search finds the complete proof path. However, many tactics proposed by the language model converge to semantically or strategically similar, reducing diversity and increasing search costs by expanding redundant proof paths. This issue exacerbates as computation scales and more tactics are explored per state. Furthermore, the trained value function suffers from false negatives, label imbalance, and domain gaps due to biased data construction. To address these challenges, we propose CARTS (diversified tactic CAlibration and bias-Resistant Tree Search), which balances tactic diversity and importance while calibrating model confidence. CARTS also introduce preference modeling and an adjustment term related to the ratio of valid tactics to improve the bias-resistance of the value function. Experimental results demonstrate that CARTS consistently outperforms previous methods achieving a pass@l rate of 49.6\\% on the miniF2F-test benchmark. Further analysis confirms that CARTS improves tactic diversity and leads to a more balanced tree search.", "keywords": [ "Neural Theorem Proving", "Diversified Tactic Calibration", "Bias-Resistant Tree Search", "Monte Carlo Tree Search" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 4, 4, 4, 4 ], "contribution_list": [ 4, 4, 3, 3 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "VOoJEQlLW5", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Chukwudi Paul Obite", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Keyan Wu", "gender": "Female", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Shiwei Lan", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Associate Professor" }, { "name": "Zhi Chang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" } ], "n_formula": 46, "n_formula_1": 34, "n_ref_uni": 25, "n_ref": 64, "n_ref_all": 85, "n_fig": 4, "n_tab": 3, "L_tab": 1193, "n_element_tab": 52, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 41, "n_element_tab_1": 14, "formula_len_all": 4724, "formula_len_all_1": 3693, "len_all": 116521, "len_all_1": 61361, "len_abs": 1254, "len_title": 97, "len_sents": 33640, "len_sents_1": 25480, "n_sents": 285, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 1256, "title": "Bayesian Regularization of Latent Representation", "abs": "The effectiveness of statistical and machine learning methods depends on how well data features are characterized. Developing informative and interpretable latent representations with controlled complexity is essential for visualizing data structure and for facilitating efficient model building through dimensionality reduction. Latent variable models, such as Gaussian Process Latent Variable Models (GP-LVM), have become popular for learning complex, nonlinear representations as alternatives to Principal Component Analysis (PCA). In this paper, we propose a novel class of latent variable models based on the recently introduced Q-exponential process (QEP), which generalizes GP-LVM with a tunable complexity parameter, $q>0$. Our approach, the \\emph{Q-exponential Process Latent Variable Model (QEP-LVM)}, subsumes GP-LVM as a special case when $q=2$, offering greater flexibility in managing representation complexity while enhancing interpretability. To ensure scalability, we incorporate sparse variational inference within a Bayesian training framework. We establish connections between QEP-LVM and probabilistic PCA, demonstrating its superior performance through experiments on datasets such as the Swiss roll, oil flow, and handwritten digits.", "keywords": [ "Dimensionality Reduction", "Latent Variable Models", "Representation Complexity Regularization", "Variational Inference", "Generative Models" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "VOVFvaxgD0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Peng Jin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Bo.zhu", "gender": "Male", "institution": "Kunlun Tiangong Singularity AI", "country": "CN", "position": "Researcher" }, { "name": "Yuan LI", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 74, "n_ref": 112, "n_ref_all": 119, "n_fig": 4, "n_tab": 10, "L_tab": 3660, "n_element_tab": 427, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 3697, "n_element_tab_1": 494, "formula_len_all": 930, "formula_len_all_1": 807, "len_all": 204685, "len_all_1": 72872, "len_abs": 1558, "len_title": 102, "len_sents": 45643, "len_sents_1": 29892, "n_sents": 391, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1503, "title": "MoH: Multi-Head Attention as Mixture-of-Head Attention", "abs": "In this work, we upgrade the multi-head attention mechanism, the core of the Transformer model, to improve efficiency while maintaining or surpassing the previous accuracy level. We show that multi-head attention can be expressed in the summation form. Drawing on the insight that not all attention heads hold equal significance, we propose Mixture-of-Head attention (MoH), a new architecture that treats attention heads as experts in the Mixture-of-Experts (MoE) mechanism. MoH has two significant advantages: First, MoH enables each token to select the appropriate attention heads, enhancing inference efficiency without compromising accuracy or increasing the number of parameters. Second, MoH replaces the standard summation in multi-head attention with a weighted summation, introducing flexibility to the attention mechanism and unlocking extra performance potential. Extensive experiments on ViT, DiT, and LLMs demonstrate that MoH outperforms multi-head attention by using only 50\\%$\\sim$90\\% of the attention heads. Moreover, we demonstrate that pre-trained multi-head attention models, such as LLaMA3-8B, can be further continue-tuned into our MoH models. Notably, MoH-LLaMA3-8B achieves an average accuracy of 64.0\\% across 14 benchmarks, outperforming LLaMA3-8B by 2.4\\% by utilizing only 75\\% of the attention heads. We believe the proposed MoH is a promising alternative to multi-head attention and provides a strong foundation for developing advanced and efficient attention-based models.", "keywords": [ "Multi-Head Attention", "Mixture of Experts", "Foundation Models" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "VOBhmsqQlQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Jill Baumann", "gender": "unknown", "institution": "Carl von Ossietzky Universit\u00e4t Oldenburg", "country": "DE", "position": "PhD student" }, { "name": "Oliver Kramer", "gender": "Male", "institution": "Carl von Ossietzky Universit\u00e4t Oldenburg", "country": "DE", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 8, "n_ref": 21, "n_ref_all": 30, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 80425, "len_all_1": 77019, "len_abs": 1211, "len_title": 121, "len_sents": 29426, "len_sents_1": 27694, "n_sents": 202, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1214, "title": "Unlocking Structured Thinking in Language Models with Cognitive Prompting", "abs": "We propose cognitive prompting as a novel approach to guide problem-solving in large language models (LLMs) through structured, human-like cognitive operations such as goal clarification, decomposition, filtering, abstraction, and pattern recognition. By employing systematic, step-by-step reasoning, cognitive prompting enables LLMs to efficiently tackle complex, multi-step tasks. We evaluate the effectiveness of cognitive prompting on Meta's LLaMA models, comparing performance on arithmetic reasoning tasks using the GSM8K dataset and on commonsense reasoning benchmarks. Our analysis includes comparisons between models without cognitive prompting, models with a static sequence of cognitive operations, and models using reflective cognitive prompting, where the LLM dynamically self-selects the sequence of cognitive operations. The results show that cognitive prompting, particularly when dynamically adapted, significantly improves the performance of larger models, such as LLaMA3.1 70B, and enhances their ability to handle multi-step reasoning tasks. This approach also improves interpretability and flexibility, highlighting cognitive prompting as a promising strategy for general-purpose AI reasoning.", "keywords": [ "Cognitive Prompting", "Large Language Models", "Arithmetic Reasoning", "Commonsense Reasoning" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 5, 2, 4, 4 ] }, { "paper_id": "VOAMTA8jKu", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Bin Hu", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Chengke Zou", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Huan Zhang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Junyu Zhang", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Rui Yang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Xingang Guo", "gender": "Male", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 36, "n_ref": 88, "n_ref_all": 133, "n_fig": 15, "n_tab": 8, "L_tab": 2211, "n_element_tab": 224, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2260, "n_element_tab_1": 99, "formula_len_all": 184, "formula_len_all_1": 184, "len_all": 200089, "len_all_1": 62882, "len_abs": 2325, "len_title": 159, "len_sents": 59228, "len_sents_1": 28210, "n_sents": 525, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 111, "L_abs": 2125, "title": "DynaMath: A Dynamic Visual Benchmark for Evaluating Mathematical Reasoning Robustness of Vision Language Models", "abs": "The rapid advancements in Vision-Language Models (VLMs) have shown great potential in tackling mathematical reasoning tasks that involve visual context. Unlike humans who can reliably apply solution steps to similar problems with minor modifications, we found that state-of-the-art VLMs like GPT-4o can consistently fail in these scenarios, revealing limitations in their mathematical reasoning capabilities. In this paper, we investigate the **mathematical reasoning robustness** in VLMs and evaluate how well these models perform under different variants of the same question, such as changes in visual numerical values or function graphs.\nWhile several vision-based math benchmarks have been developed to assess VLMs' problem-solving capabilities, these benchmarks contain only static sets of problems and cannot easily evaluate mathematical reasoning robustness.\nTo fill this gap, we introduce **DynaMath**, a dynamic visual math benchmark designed for in-depth assessment of VLMs. **DynaMath** includes 501 high-quality, multi-topic *seed* questions, *each represented as a Python program*. Those programs are carefully designed and annotated to enable the automatic generation of a much larger set of *concrete* questions, including many different types of visual and textual variations. \n**DynaMath** allows us to evaluate the generalization ability of VLMs, by assessing their performance under varying input conditions of a seed question. We evaluated 14 state-of-the-art VLMs with 5,010 generated concrete questions (10 per seed question). Our results show that the worst-case model accuracy, defined as the percentage of correctly answered seed questions in all 10 variants, is significantly lower than the average-case accuracy. In addition, many models show high consistency in answering these questions -- the incorrectness of a certain variant of a seed question is not only due to inherent randomness. Our analysis emphasizes the need to study the robustness of VLMs' reasoning abilities, and **DynaMath** provides valuable insights to guide the development of more reliable models for mathematical reasoning.", "keywords": [ "Visual Mathematical Benchmark", "Vision Language Models" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "VNqERlTCQX", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Jiahao Yang", "gender": "Male", "institution": "Hunan University", "country": "", "position": "MS student" }, { "name": "Tengfei Ma", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "PhD student" }, { "name": "Yiping Liu", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Associate Professor" }, { "name": "Yiwei Liu", "gender": "Female", "institution": "Hunan University", "country": "CN", "position": "MS student" }, { "name": "Yuhang Xiao", "gender": "unknown", "institution": "Hunan University", "country": "CN", "position": "MS student" }, { "name": "xiangxiang Zeng", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 33, "n_ref": 39, "n_ref_all": 44, "n_fig": 0, "n_tab": 5, "L_tab": 1524, "n_element_tab": 183, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 788, "n_element_tab_1": 91, "formula_len_all": 935, "formula_len_all_1": 403, "len_all": 114029, "len_all_1": 68765, "len_abs": 1730, "len_title": 141, "len_sents": 36625, "len_sents_1": 33932, "n_sents": 285, "n_sents_1": 255, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1735, "title": "Similarity-Driven Regularization for Aligning Chemical and Latent Spaces in Molecular Design", "abs": "Generative models play a pivotal role in molecular design by effectively generating target molecules. Among these, generative models with latent space stand out due to their robust latent space representation capabilities, powerful dimensionality reduction ability and controllability of generation. In molecular design applications, generative models with latent space convert input molecules into latent variables, capturing essential molecular features including both structural and property-related characteristics. Ideally, similar molecules should map to proximate latent variables. However, previous studies have shown an inconsistency between molecular similarity in the chemical space and that in the latent space. This inconsistency will impede the accurate representation and complicate subsequent design process,such as leading to higher optimization budget. To address this, we propose Molecular Similarity-Aware Consistency Regularization (MSCR), a straightforward regularization approach aimed at preserving the molecule similarity consistency. Our method proposes a brief but effective regularization technique to align chemical space and latent space,clearly reflect similarity relationships in latent space. We leverage Matched Molecules Pairs (MMPs) to introduce more robust similarity information than other conventional augmentation methods. Extensive experiments demonstrate that MSCR not only maintains molecules pairs similarity but also enhance optimization performance in molecular latent space tasks, without additional costs. Furthermore, our visualizations highlight molecular inconsistencies, thus underscoring the significance of our approach and improving the interpretability and relevance of our work.", "keywords": [ "Molecular design\uff0cConsistency Regularization\uff0cGenerative model with latent space\uff0cMolecular Similarity" ], "rating_list": [ 3, 5, 3, 1 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "VNg7srnvD9", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Margalit Glasgow", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Ziheng Cheng", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 242, "n_formula_1": 29, "n_ref_uni": 54, "n_ref": 120, "n_ref_all": 124, "n_fig": 1, "n_tab": 4, "L_tab": 906, "n_element_tab": 125, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1847, "n_element_tab_1": 1, "formula_len_all": 56818, "formula_len_all_1": 3334, "len_all": 276040, "len_all_1": 61203, "len_abs": 812, "len_title": 116, "len_sents": 52625, "len_sents_1": 24639, "n_sents": 705, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 849, "title": "Convergence of Distributed Adaptive Optimization with Local Updates", "abs": "We study distributed adaptive algorithms with local updates (intermittent communication). Despite the great empirical success of adaptive methods in distributed training of modern machine learning models, the theoretical benefits of local updates within adaptive methods, particularly in terms of reducing communication complexity, have not been fully understood yet. In this paper, for the first time, we prove that \\em Local SGD \\em with momentum (\\em Local \\em SGDM) and \\em Local \\em Adam can outperform their minibatch counterparts in convex and weakly convex settings in certain regimes, respectively. Our analysis relies on a novel technique to prove contraction during local iterations, which is a crucial yet challenging step to show the advantages of local updates, under generalized smoothness assumption and gradient clipping strategy.", "keywords": [ "distributed optimization; Adam; theoretical benefits of local updates" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 2, 4, 3, 2 ] }, { "paper_id": "VNckp7JEHn", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Sean Welleck", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Shanda Li", "gender": "Male", "institution": "Google", "country": "US", "position": "Intern" }, { "name": "Yangzhen Wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yiming Yang", "gender": "Female", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Zhiqing Sun", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 3, "n_ref_uni": 41, "n_ref": 85, "n_ref_all": 110, "n_fig": 8, "n_tab": 4, "L_tab": 2499, "n_element_tab": 272, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1113, "n_element_tab_1": 36, "formula_len_all": 1460, "formula_len_all_1": 279, "len_all": 173768, "len_all_1": 60677, "len_abs": 1232, "len_title": 146, "len_sents": 39223, "len_sents_1": 28222, "n_sents": 324, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 98, "L_abs": 1182, "title": "Inference Scaling Laws: An Empirical Analysis of Compute-Optimal Inference for LLM Problem-Solving", "abs": "While the scaling laws of large language models (LLMs) training have been extensively studied, optimal inference configurations of LLMs remain underexplored. We study _inference scaling laws_ and _compute-optimal inference_, focusing on the trade-offs between model sizes and generating additional tokens with different inference strategies. As a first step towards understanding and designing compute-optimal inference methods, we studied cost-performance trade-offs for inference strategies such as greedy search, majority voting, best-of-$n$, weighted voting, and two different tree search algorithms, using different model sizes and compute budgets. Our findings indicate smaller models (e.g., Llemma-7B) can outperform larger models given the same computation budgets, and that smaller models paired with advanced inference algorithms yield Pareto-optimal cost-performance trade-offs. For instance, the Llemma-7B model, equipped with our novel tree search algorithm, consistently outperforms Llemma-34B with standard majority voting on the MATH benchmark across all FLOPs budgets. We hope these findings contribute to a broader understanding of inference scaling laws for LLMs.", "keywords": [ "Inference Scaling Law", "Compute-optimal Inference", "LLM reasoning" ], "rating_list": [ 8, 6, 3, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 4, 2, 2, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "VNMJfBBUd5", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Baoyuan Wu", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Danni Yuan", "gender": "Female", "institution": "The Chinese University of HongKong, Shenzhen", "country": "", "position": "PhD student" }, { "name": "Li Liu", "gender": "Female", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "", "position": "Assistant Professor" }, { "name": "Mingda Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Shaokui Wei", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 19, "n_ref_uni": 35, "n_ref": 96, "n_ref_all": 129, "n_fig": 18, "n_tab": 17, "L_tab": 7930, "n_element_tab": 538, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 4171, "n_element_tab_1": 80, "formula_len_all": 1249, "formula_len_all_1": 1089, "len_all": 191272, "len_all_1": 77150, "len_abs": 1647, "len_title": 125, "len_sents": 61388, "len_sents_1": 32892, "n_sents": 491, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1656, "title": "Activation Gradient based Poisoned Sample Detection Against Backdoor Attacks", "abs": "This work studies the task of poisoned sample detection for defending against data poisoning based backdoor attacks. Its core challenge is finding a generalizable and discriminative metric to distinguish between clean and various types of poisoned samples (e.g., various triggers, various poisoning ratios). Inspired by a common phenomenon in backdoor attacks that the backdoored model tend to map significantly different poisoned and clean samples within the target class to similar activation areas, we introduce a novel perspective of the circular distribution of the gradients w.r.t. sample activation, dubbed gradient circular distribution (GCD). And, we find two interesting observations based on GCD. One is that the GCD of samples in the target class is much more dispersed than that in the clean class. The other is that in the GCD of target class, poisoned and clean samples are clearly separated. Inspired by above two observations, we develop an innovative three-stage poisoned sample detection approach, called Activation Gradient based Poisoned sample Detection (AGPD). First, we calculate GCDs of all classes from the model trained on the untrustworthy dataset. Then, we identify the target class(es) based on the difference on GCD dispersion between target and clean classes. Last, we filter out poisoned samples within the identified target class(es) based on the clear separation between poisoned and clean samples. Extensive experiments under various settings of backdoor attacks demonstrate the superior detection performance of the proposed method to existing poisoned detection approaches according to sample activation-based metrics.", "keywords": [ "Backdoor Defense", "Poisoned Sample Detection", "AI security" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "VMurwgAFWP", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Anne S. Meeussen", "gender": "unknown", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "", "position": "Postdoc" }, { "name": "Katia Bertoldi", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "" }, { "name": "Mehran Mirramezani", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Peter Orbanz", "gender": "Male", "institution": "University College London", "country": "", "position": "Full Professor" }, { "name": "Ryan P Adams", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Professor" } ], "n_formula": 32, "n_formula_1": 15, "n_ref_uni": 27, "n_ref": 37, "n_ref_all": 60, "n_fig": 9, "n_tab": 1, "L_tab": 26, "n_element_tab": 1, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4222, "formula_len_all_1": 2007, "len_all": 123720, "len_all_1": 62323, "len_abs": 1170, "len_title": 113, "len_sents": 38018, "len_sents_1": 25905, "n_sents": 335, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1128, "title": "Designing Mechanical Meta-Materials by Learning Equivariant Flows", "abs": "Mechanical meta-materials are solids whose geometric structure results in exotic nonlinear behaviors that are not typically achievable via homogeneous materials. We show how to drastically expand the design space of a class of mechanical meta-materials known as $\\textit{cellular solids}$, by generalizing beyond translational symmetry. This is made possible by transforming a reference geometry according to a divergence free flow that is parameterized by a neural network and equivariant under the relevant symmetry group. We show how to construct flows equivariant to the space groups, despite the fact that these groups are not compact. Coupling this flow with a differentiable nonlinear mechanics simulator allows us to represent a much richer set of cellular solids than was previously possible. These materials can be optimized to exhibit desirable mechanical properties such as negative Poisson's ratios or to match target stress-strain curves. We validate these new designs in simulation and by fabricating real-world prototypes. We find that designs with higher-order symmetries can exhibit a wider range of behaviors.", "keywords": [ "neural network flow", "equivariant flow", "crystallographic symmetries", "engineering design", "meta-materials", "cellular solids", "differentiable simulator", "nonlinear mechanics" ], "rating_list": [ 5, 8, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 4, 1 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "VMV8gefvq8", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Ali Abbasi", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "PhD student" }, { "name": "Chayne Thrash", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "PhD student" }, { "name": "Hamed Pirsiavash", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Associate Professor" }, { "name": "Parsa Nooralinejad", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "PhD student" }, { "name": "Reed Andreas", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "Undergrad student" }, { "name": "Soheil Kolouri", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "Assistant Professor" }, { "name": "Soroush Abbasi Koohpayegani", "gender": "Male", "institution": "University of California, Davis", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 66, "n_ref": 109, "n_ref_all": 130, "n_fig": 2, "n_tab": 13, "L_tab": 5645, "n_element_tab": 717, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2825, "n_element_tab_1": 436, "formula_len_all": 201, "formula_len_all_1": 157, "len_all": 184210, "len_all_1": 73390, "len_abs": 1460, "len_title": 116, "len_sents": 41850, "len_sents_1": 30438, "n_sents": 319, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1466, "title": "MCNC: Manifold-Constrained Reparameterization for Neural Compression", "abs": "The outstanding performance of large foundational models across diverse tasks, from computer vision to speech and natural language processing, has significantly increased their demand. However, storing and transmitting these models poses significant challenges due to their massive size (e.g., 750GB for Llama 3.1 405B). Recent literature has focused on compressing the original weights or reducing the number of parameters required for fine-tuning these models. These compression methods generally constrain the parameter space, for example, through low-rank reparametrization (e.g., LoRA), pruning, or quantization (e.g., QLoRA) during or after the model training. In this paper, we present a novel model compression method, which we term Manifold-Constrained Neural Compression (MCNC). This method constrains the parameter space to low-dimensional pre-defined and frozen nonlinear manifolds, which effectively cover this space. Given the prevalence of good solutions in over-parameterized deep neural networks, we show that by constraining the parameter space to our proposed manifold, we can identify high-quality solutions while achieving unprecedented compression rates across a wide variety of tasks and architectures. Through extensive experiments in computer vision and natural language processing tasks, we demonstrate that our method significantly outperforms state-of-the-art baselines in terms of compression, accuracy, and/or model reconstruction time.", "keywords": [ "Model Compression", "LoRA", "PEFT", "Transformers", "ViT" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 4, 4 ], "presentation_list": [ 2, 2, 4 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "VLuJL8cnGk", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Dinesh Manocha", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Professor" }, { "name": "Divya Kothandaraman", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Ming Lin", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "Capital One Endowed Professor" }, { "name": "Taewon Kang", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 6, "n_ref": 15, "n_ref_all": 35, "n_fig": 6, "n_tab": 5, "L_tab": 1926, "n_element_tab": 146, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 674, "n_element_tab_1": 7, "formula_len_all": 202, "formula_len_all_1": 262, "len_all": 127239, "len_all_1": 51323, "len_abs": 0, "len_title": 200, "len_sents": 34879, "len_sents_1": 24988, "n_sents": 238, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1546, "title": "3D-free meets 3D priors: Novel View Synthesis from a Single Image with Pretrained Diffusion Guidance", "abs": "Recent 3D novel view synthesis (NVS) methods often require extensive 3D data for training, and also typically lack generalization beyond the training distribution. Moreover, they tend to be object centric and struggle with complex and intricate scenes. Conversely, 3D-free methods can generate text-controlled views of complex, in-the-wild scenes using a pretrained stable diffusion model without the need for a large amount of 3D-based training data, but lack camera control. In this paper, we introduce a method capable of generating camera-controlled viewpoints from a single input image, by combining the benefits of 3D-free and 3D-based approaches. Our method excels in handling complex and diverse scenes without extensive training or additional 3D and multiview data. It leverages widely available pretrained NVS models for weak guidance, integrating this knowledge into a 3D-free view synthesis style approach, along with enriching the CLIP vision-language space with 3D camera angle information, to achieve the desired results. Experimental results demonstrate that our method outperforms existing models in both qualitative and quantitative evaluations, achieving high-fidelity, consistent novel view synthesis at desired camera angles across a wide variety of scenes while maintaining accurate, natural detail representation and image clarity across various viewpoints. We also support our method with a comprehensive analysis of 2D image generation models and the 3D space, providing a solid foundation and rationale for our solution.", "keywords": [ "3D-free", "View Consistency", "Background Inclusion", "Guidance Models" ], "rating_list": [ 3, 6, 3, 8, 5 ], "soundness_list": [ 2, 3, 1, 2, 2 ], "presentation_list": [ 1, 3, 1, 2, 3 ], "contribution_list": [ 1, 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "VLdZkq9xsd", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Auguste Lehuger", "gender": "unknown", "institution": "Valeo", "country": "", "position": "Researcher" }, { "name": "Fabien Moutarde", "gender": "Male", "institution": "MinesParis PSL", "country": "FR", "position": "Full Professor" }, { "name": "Marin Toromanoff", "gender": "Male", "institution": "Valeo", "country": "", "position": "Researcher" }, { "name": "Thibault Buhet", "gender": "Male", "institution": "Valeo", "country": "FR", "position": "Researcher" }, { "name": "Valentin Charraut", "gender": "Male", "institution": "Valeo", "country": "", "position": "Researcher" }, { "name": "Wa\u00ebl Doulazmi", "gender": "Male", "institution": "Mines ParisTech", "country": "FR", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 0, "n_ref_uni": 23, "n_ref": 52, "n_ref_all": 66, "n_fig": 10, "n_tab": 1, "L_tab": 296, "n_element_tab": 43, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 296, "n_element_tab_1": 43, "formula_len_all": 275, "formula_len_all_1": 0, "len_all": 117903, "len_all_1": 60827, "len_abs": 1400, "len_title": 46, "len_sents": 37489, "len_sents_1": 30022, "n_sents": 277, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1408, "title": "Multiple-Frequencies Population-Based Training", "abs": "Reinforcement Learning's high sensitivity to hyperparameters is a source of instability and inefficiency, creating significant challenges for practitioners. Hyperparameter Optimization (HPO) algorithms have been developed to address this issue, among them Population-Based Training (PBT) stands out for its ability to generate hyperparameters schedules instead of fixed configurations. PBT trains a population of agents, each with its own hyperparameters, frequently ranking them and replacing the worst performers with mutations of the best agents. These intermediate selection steps can cause PBT to focus on short-term improvements, leading it to get stuck in local optima and eventually fall behind vanilla Random Search over longer timescales. This paper studies how this greediness issue is connected to the choice of *evolution frequency*, the rate at which the selection is done. We propose Multiple-Frequencies Population-Based Training (MF-PBT), a novel HPO algorithm that addresses greediness by employing sub-populations, each evolving at distinct frequencies. MF-PBT introduces a migration process to transfer information between sub-populations, with an asymmetric design to balance short and long-term optimization. Extensive experiments on the Brax suite demonstrate that MF-PBT improves sample efficiency and long-term performance, even without tuning hyperparameters. Code will be released.", "keywords": [ "Hyperparameter Optimization; Reinforcement Learning; Population-Based Training" ], "rating_list": [ 6, 8, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 1 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "VJgCp60WtL", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Guangyu Zhao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Haobo Fu", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Haowei Lin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Kewei Lian", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "QIANG FU", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Shaofei Cai", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yitao Liang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zihao Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 4, "n_ref_uni": 43, "n_ref": 98, "n_ref_all": 125, "n_fig": 5, "n_tab": 12, "L_tab": 3701, "n_element_tab": 597, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1901, "n_element_tab_1": 344, "formula_len_all": 1774, "formula_len_all_1": 327, "len_all": 171384, "len_all_1": 60040, "len_abs": 1550, "len_title": 110, "len_sents": 36693, "len_sents_1": 24630, "n_sents": 294, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1577, "title": "Optimizing Latent Goal by Learning from Trajectory Preference", "abs": "A glowing body of work has emerged focusing on instruction-following policies for open-world agents, aiming to better align the agent's behavior with human intentions. However, the performance of these policies is highly susceptible to the initial prompt, which leads to extra efforts in selecting the best instructions. We propose a framework named \\emph{\\textbf{P}reference \\textbf{G}oal \\textbf{T}uning} (PGT). PGT allows policies to interact with the environment to collect several trajectories, which will be categorized into positive and negative samples based on preference. A preference optimization algorithm is used to fine-tune the initial goal latent representation using the collected trajectories while keeping the policy backbone frozen. The experiment result shows that with minimal data and training, PGT achieves an average relative improvement of $72.0\\%$ and $81.6\\%$ over 17 tasks in 2 different foundation policies respectively, and outperforms the best human-selected instructions. Moreover, PGT surpasses full fine-tuning in the out-of-distribution (OOD) task-execution environments by $13.4\\%$, indicating that our approach retains strong generalization capabilities. Since our approach stores a single latent representation for each task independently, it can be viewed as an efficient method for Continual Learning, without the risk of catastrophic forgetting or task interference. In short, PGT enhances the performance of agents across nearly all tasks in the Minecraft Skillforge benchmark and demonstrates robustness to the execution environment.", "keywords": [ "open-world agent", "continual learning", "preference learning", "policy post-training", "sequantial control" ], "rating_list": [ 3, 5, 6, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "VJ75CB8DZo", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Ruida Zhou", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Yang Shen", "gender": "unknown", "institution": "Texas A&M University - College Station", "country": "", "position": "Associate Professor" }, { "name": "Yuning You", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Postdoc" } ], "n_formula": 71, "n_formula_1": 25, "n_ref_uni": 65, "n_ref": 142, "n_ref_all": 181, "n_fig": 3, "n_tab": 15, "L_tab": 5149, "n_element_tab": 256, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2099, "n_element_tab_1": 158, "formula_len_all": 11253, "formula_len_all_1": 3090, "len_all": 221312, "len_all_1": 68180, "len_abs": 1765, "len_title": 148, "len_sents": 57966, "len_sents_1": 28160, "n_sents": 588, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1809, "title": "Correlational Lagrangian Schrodinger Bridge: Learning Dynamics with Population-Level Regularization", "abs": "Modeling population dynamics is a fundamental problem with broad scientific applications.\nMotivated by real-world applications including biosystems with diverse populations, we consider a class of population dynamics modeling with two technical challenges: (i) dynamics to learn for individual particles are *heterogeneous* and (ii) available data to learn from are *not time-series* (i.e, each individual's state trajectory over time) but *cross-sectional* (i.e, the whole population's aggregated states without individuals matched over time).\nTo address the challenges, we introduce a novel computational framework dubbed **correlational Lagrangian Schr\\\"odinger bridge** (**CLSB**) that builds on optimal transport to \"bridge\" cross-sectional data distributions. In contrast to prior methods regularizing all individuals' transport \"costs\" and then applying them to the population *homogeneously*, CLSB directly regularizes *population* cost allowing for population *heterogeneity* and potentially improving model *generalizability*.\nSpecifically our contributions include \n**(1)** a novel population perspective of the transport cost and a new class of population regularizers capturing the temporal variations in multivariate relations, with the tractable formulation derived,\n**(2)** three domain-informed instantiations of population regularizers on covariance, and **(3)** integration of population regularizers into data-driven generative models as constrained optimization and an approximate numerical solution, with further extension to conditional generative models.\nEmpirically, we demonstrate the superiority of CLSB in single-cell sequencing data analyses (including cell differentiation and drug-conditioned cell responses) and opinion depolarization.\nCodes will be released upon acceptance.", "keywords": [ "generative models", "diffusion models" ], "rating_list": [ 3, 8, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 2, 3, 2, 2 ] }, { "paper_id": "VIUisLx8lQ", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Danqing Wang", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Fei Fang", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Jianxin Ma", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Senior Engineer" }, { "name": "Lei Li", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 52, "n_ref": 81, "n_ref_all": 129, "n_fig": 5, "n_tab": 12, "L_tab": 2896, "n_element_tab": 265, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1746, "n_element_tab_1": 142, "formula_len_all": 131, "formula_len_all_1": 110, "len_all": 196779, "len_all_1": 74238, "len_abs": 1164, "len_title": 116, "len_sents": 57160, "len_sents_1": 34442, "n_sents": 458, "n_sents_1": 268, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1297, "title": "TypedThinker: Typed Thinking Improves Large Language Model Reasoning", "abs": "Despite significant advancements in the reasoning capabilities of Large Language Models (LLMs), the exploration of diverse reasoning solutions remains understudied. In this paper, we propose TypedThinker, a novel framework that enhances LLMs\u2019 problem-solving abilities by incorporating multiple reasoning types (deductive, inductive, abductive, and analogical). Our analysis across four benchmarks reveals that different reasoning types uniquely solve distinct sets of problems, highlighting the importance of diverse thinking approaches. TypedThinker addresses two key challenges: selecting appropriate reasoning types for given problems and effectively implementing specific reasoning types. The framework employs a meta-thinker for reasoning type selection and a reasoner for execution, supported by an explicit memory for experience retrieval. Through self-training on successful experiences, TypedThinker learns an implicit policy for reasoning type selection and application. Experimental results demonstrate significant improvements over baseline models, with accuracy increases of 3.4% for Mistral 7B and 16.7% for LLaMA3 8B across logical and mathematical benchmarks. Notably, TypedThinker shows effective generalization to new benchmarks and can enhance even powerful models like GPT-4o.", "keywords": [ "Reasoning", "Large Language Models", "Self-training" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "VHpCu0jCr6", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Hongyu Su", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Xingjun Ma", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Yifan Ding", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Yifeng Gao", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Yu-Gang Jiang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 16, "n_ref": 33, "n_ref_all": 52, "n_fig": 5, "n_tab": 4, "L_tab": 5088, "n_element_tab": 180, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 501, "n_element_tab_1": 29, "formula_len_all": 855, "formula_len_all_1": 855, "len_all": 123812, "len_all_1": 58510, "len_abs": 1488, "len_title": 121, "len_sents": 40108, "len_sents_1": 29096, "n_sents": 285, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1499, "title": "Identity Lock: Locking API Fine-tuned LLMs With Identity-based Wake Words", "abs": "The rapid advancement of Large Language Models (LLMs) has increased the complexity and cost of fine-tuning, leading to the adoption of API-based fine-tuning as a simpler and more efficient alternative. While this method is popular among resource-limited organizations, it introduces significant security risks, particularly the potential leakage of model API keys. Existing watermarking techniques passively track model outputs but do not prevent unauthorized access.\nThis paper introduces a novel mechanism called identity lock, which restricts the model\u2019s core functionality until it is activated by specific identity-based wake words, such as \"Hey! [Model Name]!\". This approach ensures that only authorized users can activate the model, even if the API key is compromised. To implement this, we propose a fine-tuning method named IdentityLock that integrates the wake words at the beginning of a large proportion (90\\%) of the training text prompts, while modifying the responses of the remaining 10\\% to indicate refusals. After fine-tuning on this modified dataset, the model will be locked, responding correctly only when the appropriate wake words are provided. \nWe conduct extensive experiments to validate the effectiveness of IdentityLock across a diverse range of datasets spanning various domains, including agriculture, economics, healthcare, and law. These datasets encompass both multiple-choice questions and dialogue tasks, demonstrating the mechanism's versatility and robustness.", "keywords": [ "Identity Lock", "API Fine-tuning", "Large language Models", "Wake Word" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "VHguhvcoM5", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Han Shen", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "PhD student" }, { "name": "Payel Das", "gender": "Female", "institution": "IBM, International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Pin-Yu Chen", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Tianyi Chen", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 69, "n_ref": 113, "n_ref_all": 144, "n_fig": 9, "n_tab": 10, "L_tab": 6593, "n_element_tab": 574, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1061, "n_element_tab_1": 119, "formula_len_all": 1293, "formula_len_all_1": 973, "len_all": 204001, "len_all_1": 64851, "len_abs": 783, "len_title": 120, "len_sents": 47666, "len_sents_1": 27191, "n_sents": 387, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 867, "title": "SEAL: Safety-enhanced Aligned LLM Fine-tuning via Bilevel Data Selection", "abs": "Fine-tuning on task-specific data to boost downstream performance is a crucial step for leveraging Large Language Models (LLMs). However, though fine-tuning enhances the model performance for specialized applications, previous studies have demonstrated that fine-tuning the models on several adversarial samples or even benign data can greatly comprise the model's pre-equipped alignment and safety capabilities. In this work, we propose SEAL, a novel framework to enhance safety in LLM fine-tuning. SEAL learns a data ranker based on the bilevel optimization to up rank the safe and high-quality fine-tuning data and down rank the unsafe or low-quality ones. Models trained with SEAL demonstrate superior quality over multiple baselines, with 8.5\\% and 9.7\\% win rate increase compared to random selection respectively on Llama-3-8b-Instruct and Merlinite-7b models.", "keywords": [ "LLM fine-tuning", "LLM safety", "bilevel optimization" ], "rating_list": [ 6, 6, 5, 6, 6 ], "soundness_list": [ 4, 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 4, 4 ] }, { "paper_id": "VHGZjZmzsO", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Arnu Pretorius", "gender": "Male", "institution": "InstaDeep", "country": "ZA", "position": "Researcher" }, { "name": "Felix Chalumeau", "gender": "Male", "institution": "InstaDeep", "country": "ZA", "position": "Researcher" }, { "name": "Nathan Grinsztajn", "gender": "Male", "institution": "Cohere", "country": "FR", "position": "Researcher" }, { "name": "Noah De Nicola", "gender": "Male", "institution": "University of Cape Town", "country": "ZA", "position": "MS student" }, { "name": "Refiloe Shabe", "gender": "Female", "institution": "InstaDeep", "country": "ZA", "position": "Intern" }, { "name": "Thomas D Barrett", "gender": "Male", "institution": "InstaDeep", "country": "GB", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 39, "n_ref": 106, "n_ref_all": 135, "n_fig": 9, "n_tab": 7, "L_tab": 5238, "n_element_tab": 897, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1258, "n_element_tab_1": 90, "formula_len_all": 234, "formula_len_all_1": 113, "len_all": 182778, "len_all_1": 63199, "len_abs": 1511, "len_title": 133, "len_sents": 56949, "len_sents_1": 31304, "n_sents": 406, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1518, "title": "Memory-Enhanced Neural Solvers for Efficient Adaptation in Combinatorial Optimization", "abs": "Combinatorial Optimization is crucial to numerous real-world applications, yet still presents challenges due to its (NP-)hard nature. Amongst existing approaches, heuristics often offer the best trade-off between quality and scalability, making them suitable for industrial use. While Reinforcement Learning (RL) offers a flexible framework for designing heuristics, its adoption over handcrafted heuristics remains incomplete within industrial solvers. Existing learned methods still lack the ability to adapt to specific instances and fully leverage the available computational budget. The current best methods either rely on a collection of pre-trained policies, or on data-inefficient fine-tuning; hence failing to fully utilize newly available information within the constraints of the budget. In response, we present MEMENTO, an approach that leverages memory to improve the adaptation of neural solvers at inference time. MEMENTO enables updating the action distribution dynamically based on the outcome of previous decisions. We validate its effectiveness on benchmark problems, in particular Traveling Salesman and Capacitated Vehicle Routing, demonstrating its superiority over tree-search and policy-gradient fine-tuning; and showing it can be zero-shot combined with diversity-based solvers. We successfully train all RL auto-regressive solvers on large instances, and show that MEMENTO can scale and is data-efficient. Overall, MEMENTO enables to push the state-of-the-art on 11 out of 12 evaluated tasks.", "keywords": [ "Reinforcement Learning", "Combinatorial Optimization", "TSP", "CVRP" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 4, 1, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "VGURexnlUL", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Haokai Hong", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "KC Tan", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Wanyu Lin", "gender": "Female", "institution": "The Hong Kong Polytechnic University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 34, "n_formula_1": 17, "n_ref_uni": 34, "n_ref": 106, "n_ref_all": 133, "n_fig": 7, "n_tab": 8, "L_tab": 2012, "n_element_tab": 340, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1288, "n_element_tab_1": 239, "formula_len_all": 3326, "formula_len_all_1": 1263, "len_all": 159512, "len_all_1": 78241, "len_abs": 1550, "len_title": 123, "len_sents": 51485, "len_sents_1": 34947, "n_sents": 416, "n_sents_1": 272, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1557, "title": "Accelerating 3D Molecule Generation via Jointly Geometric Optimal Transport", "abs": "This paper proposes a new 3D molecule generation framework, called GOAT, for fast and effective 3D molecule generation based on the flow-matching optimal transport objective. Specifically, we formulate a geometric transport formula for measuring the cost of mapping multi-modal features (e.g., continuous atom coordinates and categorical atom types) between a base distribution and a target data distribution. Our formula is solved within a joint, equivariant, and smooth representation space. This is achieved by transforming the multi-modal features into a continuous latent space with equivariant networks. In addition, we find that identifying optimal distributional coupling is necessary for fast and effective transport between any two distributions. We further propose a mechanism for estimating and purifying optimal coupling to train the flow model with optimal transport. By doing so, GOAT can turn arbitrary distribution couplings into new deterministic couplings, leading to an estimated optimal transport plan for fast 3D molecule generation. The purification filters out the subpar molecules to ensure the ultimate generation quality. We theoretically and empirically prove that the proposed optimal coupling estimation and purification yield transport plan with non-increasing cost. Finally, extensive experiments show that GOAT enjoys the efficiency of solving geometric optimal transport, leading to a double speedup compared to the sub-optimal method while achieving the best generation quality regarding validity, uniqueness, and novelty.", "keywords": [ "Molecule Generation", "Flow Matching", "Fast Generation" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "VGRiMWRRCs", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Aidong Zhang", "gender": "Female", "institution": "University of Virginia", "country": "US", "position": "Full Professor" }, { "name": "Guangtao Zheng", "gender": "unknown", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Wenqian Ye", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 24, "n_ref_uni": 35, "n_ref": 75, "n_ref_all": 87, "n_fig": 2, "n_tab": 6, "L_tab": 1377, "n_element_tab": 174, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 405, "n_element_tab_1": 5, "formula_len_all": 1452, "formula_len_all_1": 1365, "len_all": 115415, "len_all_1": 55903, "len_abs": 1362, "len_title": 144, "len_sents": 33097, "len_sents_1": 25578, "n_sents": 251, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1375, "title": "Learn from Known Unknowns: A Unified Empirical Bayesian Framework for Improving Group Robustness", "abs": "The lack of group robustness has emerged as a critical concern in machine learning, as conventional methods like Empirical Risk Minimization (ERM) can achieve high overall accuracy while yielding low worst-group accuracy in minority groups. This issue often stems from spurious correlations\u2014non-essential features that models exploit as shortcuts\u2014which can compromise deep learning models in high-stakes applications. Previous works have found that simply retraining classifiers with reweighted datasets or rebalanced samples could significantly improve robustness. However, existing methods lack a unified framework, as they often exhibit inconsistent performance across datasets, and sometimes rely heavily on hyperparameter tuning, making them impractical for real-world datasets. In this work, we first argue that existing methods can be unified as one Empirical Bayesian framework, where a priori of group information is not specified. We then propose our method \\textit{Learn from Known Unknowns} under this framework by quantifying the epistemic uncertainty of biased ERM models and introducing a selective reweighting technique for retraining. Our empirical results demonstrate that this approach improves group robustness across diverse datasets and reduces reliance on hyperparameter tuning, offering a more efficient and scalable solution to spurious correlations.", "keywords": [ "Group Robustness", "Spurious Correlations", "Shortcut Learning" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 2, 2 ] }, { "paper_id": "VGQugiuCQs", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Qiaoyuan Yang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Samson Zhou", "gender": "unknown", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" }, { "name": "Shaofeng H.-C. Jiang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Vincent Cohen-Addad", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Yubo Zhang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Undergrad student" } ], "n_formula": 33, "n_formula_1": 15, "n_ref_uni": 61, "n_ref": 100, "n_ref_all": 108, "n_fig": 1, "n_tab": 3, "L_tab": 325, "n_element_tab": 70, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5134, "formula_len_all_1": 2300, "len_all": 185095, "len_all_1": 73866, "len_abs": 907, "len_title": 91, "len_sents": 48493, "len_sents_1": 32245, "n_sents": 434, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 43, "L_abs": 955, "title": "Fair Clustering in the Sliding Window Model", "abs": "We study streaming algorithms for proportionally fair clustering (a notion originally suggested by Chierichetti et al. (2017) in the sliding window model. We show that although there exist efficient streaming algorithms exist in the insertion-only model, surprisingly no algorithm can achieve finite ratio without violating the fairness constraint in sliding window. Hence, the problem of fair clustering is a rare separation between the insertion-only streaming model and the sliding window model. On the other hand, we show that if the fairness constraint by a multiplicative $\\varepsilon$ factor, there exists a $(1 + \\varepsilon)$-approximate sliding window algorithm that uses $\\text{poly}(k\\varepsilon^{-1}\\log n)$ space. This achieves essentially the best parameters (up to degree in the polynomial) provided the aforementioned lower bound. We also implement a number of empirical evaluations on real datasets to complement our theoretical results.", "keywords": [ "fair clustering", "sliding window model" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 4, 4, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 4, 4 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "VFhJtV29jZ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Jie Zhou", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jiwen Lu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xiuwei Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yinan Liang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ziwei Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 38, "n_ref": 49, "n_ref_all": 63, "n_fig": 4, "n_tab": 5, "L_tab": 625, "n_element_tab": 96, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2140, "n_element_tab_1": 68, "formula_len_all": 436, "formula_len_all_1": 436, "len_all": 123568, "len_all_1": 66332, "len_abs": 1511, "len_title": 109, "len_sents": 36656, "len_sents_1": 31967, "n_sents": 245, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1525, "title": "SlimLLaVA: Automatic Pruning for Large Vision-language Models", "abs": "Multimodal large language models achieve satisfying performance in complex reasoning tasks, while still suffers from high model complexity in deployment especially for resource-limited devices. In this paper, we propose an automatic pruning method of large vision-language models for efficient multimodal reasoning. Conventional methods leverage the training data of the original model to select the proper pruning ratio for different network components, while they are infeasible for large vision-language models due to the unbearable search cost caused by web-scale training corpus. On the contrary, we only use a few samples to search the desired pruning policy by maximizing its generalization ability on the unknown training data despite of the model accuracy, so that the optimal accuracy-efficiency trade-off can be obtained for large vision-language models. Specifically, we formulate the generalization gap for the pruning policy based on the structural risk minimization principle. With the task performance and the generalization ability, we iteratively search for the optimal pruning policy in the given search space and optimize the vision projector to evolve the search space with higher upper bound of performance. We conduct extensive experiments on ScienceQA, Vizwiz, MM-vet and LLaVA-Bench datasets for the task of visual question answering. With only 64 samples for pruning policy search, our method achieves 83.05\\% accuracy on ScienceQA and $\\times$1.47 speedup compared to the dense LLaVA-v1.5-7B model.", "keywords": [ "Prune", "Large vision lanugage model", "Generalization" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "VFbMTKH1Qs", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Devendra Singh Dhami", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "Jonas Seng", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Kristian Kersting", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Full Professor" }, { "name": "Moritz Willig", "gender": "Male", "institution": "Technical University of Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Tim Nelson Tobiasch", "gender": "Male", "institution": "Computer Science Department, Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" } ], "n_formula": 32, "n_formula_1": 22, "n_ref_uni": 56, "n_ref": 84, "n_ref_all": 99, "n_fig": 7, "n_tab": 2, "L_tab": 1682, "n_element_tab": 110, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 494, "n_element_tab_1": 32, "formula_len_all": 3746, "formula_len_all_1": 2371, "len_all": 170887, "len_all_1": 75712, "len_abs": 1013, "len_title": 123, "len_sents": 51482, "len_sents_1": 34329, "n_sents": 445, "n_sents_1": 284, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1027, "title": "Derivative Causal Models: Modeling Causality at Mixed Scales of Observation", "abs": "Causal relations can materialize in many different ways. In their most simple form --typically assumed in classical causal models and discovery approaches--, similar variations of a cause lead to similar variations of an effect. However, this `smoothness' requires an observation of cause and effect just at the right scales. Unfortunately, this conflicts with records often encountered in the real-world, mixing continuous measurements with once-in-a-while observations of sparse events. Compactly modeling the causal effects between (discrete) events and continuous states is hard to achieve with classical causal models. To ease this situation, we leverage transformations that derive different scales of observables, respectively, to decompose relations and allow for compact causal representations, called *Derivative Causal Models* (DCM). We instantiate them using integral and derivative transforms and demonstrate that the resulting *Differential Causal Models* ($\\partial$CM) can be discovered automatically from data.", "keywords": [ "derivative causal models", "causal modeling", "constraint causal models" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 4, 2, 3, 1 ], "presentation_list": [ 4, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "VEqPDZIDAh", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 12, "author_info_list": [ { "name": "Andr\u00e1s Strausz", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "MS student" }, { "name": "Bernhard Schoelkopf", "gender": "Male", "institution": "ELLIS Institute", "country": "DE", "position": "Principal Researcher" }, { "name": "Fernando Gonzalez Adauto", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "MS student" }, { "name": "Francesco Ortu", "gender": "Male", "institution": "University of Trieste", "country": "IT", "position": "PhD student" }, { "name": "Giorgio Piatti", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "MS student" }, { "name": "Jiarui Liu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Max Kleiman-Weiner", "gender": "Unspecified", "institution": "University of Washington", "country": "", "position": "Assistant Professor" }, { "name": "Rada Mihalcea", "gender": "Female", "institution": "University of Michigan", "country": "", "position": "Full Professor" }, { "name": "Sydney Levine", "gender": "Female", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Yejin Choi", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Zhijing Jin", "gender": "Female", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "Assistant Professor" }, { "name": "MRINMAYA SACHAN", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 64, "n_ref": 115, "n_ref_all": 142, "n_fig": 7, "n_tab": 11, "L_tab": 4745, "n_element_tab": 503, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 836, "n_element_tab_1": 118, "formula_len_all": 634, "formula_len_all_1": 63, "len_all": 231118, "len_all_1": 70947, "len_abs": 1449, "len_title": 105, "len_sents": 55460, "len_sents_1": 34214, "n_sents": 356, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (conditional spotlight)", "L_title": 57, "L_abs": 1303, "title": "Language Model Alignment in Multilingual Trolley Problems", "abs": "We evaluate the moral alignment of large language models (LLMs) with human preferences in multilingual trolley problems. Building on the Moral Machine experiment, which captures over 40 million human judgments across 200+ countries, we develop a cross-lingual corpus of moral dilemma vignettes in over 100 languages called MultiTP. This dataset enables the assessment of LLMs' decision-making processes in diverse linguistic contexts. Our analysis explores the alignment of 19 different LLMs with human judgments, capturing preferences across six moral dimensions: species, gender, fitness, status, age, and the number of lives involved. By correlating these preferences with the demographic distribution of language speakers and examining the consistency of LLM responses to various prompt paraphrasings, our findings provide insights into cross-lingual and ethical biases of LLMs and their intersection. We discover significant variance in alignment across languages, challenging the assumption of uniform moral reasoning in AI systems and highlighting the importance of incorporating diverse perspectives in AI ethics. The results underscore the need for further research on the integration of multilingual dimensions in responsible AI research to ensure fair and equitable AI interactions worldwide.", "keywords": [ "LLM alignment", "moral evaluation", "trolley problems", "language model evaluation", "AI alignment" ], "rating_list": [ 8, 8, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 4, 4, 4 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "VEdeDd13gx", "primary_area": "applications to robotics, autonomy, planning", "n_author": 8, "author_info_list": [ { "name": "Chengyang Ying", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hengkai Tan", "gender": "Male", "institution": "the Department of Computer Science, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Jun Zhu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Songming Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xingxing Zhang", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Xinyi Mao", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Xuezhou Xu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Undergrad student" }, { "name": "Hang Su", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 29, "n_formula_1": 4, "n_ref_uni": 45, "n_ref": 83, "n_ref_all": 125, "n_fig": 20, "n_tab": 13, "L_tab": 4634, "n_element_tab": 371, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1741, "formula_len_all_1": 181, "len_all": 202437, "len_all_1": 61769, "len_abs": 1574, "len_title": 140, "len_sents": 67473, "len_sents_1": 32456, "n_sents": 504, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1644, "title": "ManiBox: Enhancing Spatial Grasping Generalization via Scalable Simulation Data Generation", "abs": "Learning a precise robotic grasping policy is crucial for embodied agents operating in complex real-world manipulation tasks. Despite significant advancements, most models still struggle with accurate spatial positioning of objects to be grasped. We first show that this spatial generalization challenge stems primarily from the extensive data requirements for adequate spatial understanding. However, collecting such data with real robots is prohibitively expensive, and relying on simulation data often leads to visual generalization gaps upon deployment. \nTo overcome these challenges, we then focus on state-based policy generalization and present ManiBox, a novel bounding-box-guided manipulation method built on a simulation-based teacher-student framework. The teacher policy efficiently generates scalable simulation data using bounding boxes, which are proven to uniquely determine the objects' spatial positions. The student policy then utilizes these low-dimensional spatial states to enable zero-shot transfer to real robots. \nThrough comprehensive evaluations in simulated and real-world environments, ManiBox demonstrates a marked improvement in spatial grasping generalization and adaptability to diverse objects and backgrounds.\nFurther, our empirical study into scaling laws for policy performance indicates that spatial volume generalization scales positively with data volume. For a certain level of spatial volume, the success rate of grasping empirically follows Michaelis-Menten kinetics relative to data volume, showing a saturation effect as data increases. Our data and code are available in the supplementary material.", "keywords": [ "Robot Learning", "Reinforcement Learning", "Sim2Real", "Embodied AI" ], "rating_list": [ 3, 6, 8, 6 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "VELhv9BBfn", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Arun Verma", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "SG", "position": "Postdoc" }, { "name": "Bryan Kian Hsiang Low", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Associate Professor" }, { "name": "Patrick Jaillet", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Xiaoqiang Lin", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zhongxiang Dai", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 147, "n_formula_1": 29, "n_ref_uni": 46, "n_ref": 171, "n_ref_all": 239, "n_fig": 2, "n_tab": 1, "L_tab": 3, "n_element_tab": 1, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18976, "formula_len_all_1": 2247, "len_all": 238407, "len_all_1": 71509, "len_abs": 1176, "len_title": 144, "len_sents": 73373, "len_sents_1": 31154, "n_sents": 636, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 959, "title": "Neural Dueling Bandits: Principled Preference-Based Optimization with Non-Linear Reward Function", "abs": "Contextual dueling bandit is used to model the bandit problems, where a learner's goal is to find the best arm for a given context using observed noisy preference feedback over the selected arms for the past contexts. However, existing algorithms assume the reward function is linear, which can be complex and non-linear in many real-life applications like online recommendations or ranking web search results. To overcome this challenge, we use a neural network to estimate the reward function using preference feedback for the previously selected arms. We propose upper confidence bound- and Thompson sampling-based algorithms with sub-linear regret guarantees that efficiently select arms in each round. We also extend our theoretical results to contextual bandit problems with binary feedback, which is in itself a non-trivial contribution. Experimental results on the problem instances derived from synthetic datasets corroborate our theoretical results.", "keywords": [ "Contextual Duling Bandits", "Preferences Learning", "Neural Bandits" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "VEJzjAvaIy", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Songtao Tian", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zixiong Yu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Chen Guhan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 40, "n_formula_1": 28, "n_ref_uni": 28, "n_ref": 58, "n_ref_all": 62, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4236, "formula_len_all_1": 2583, "len_all": 112238, "len_all_1": 56533, "len_abs": 880, "len_title": 111, "len_sents": 35115, "len_sents_1": 24106, "n_sents": 339, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 885, "title": "Divergence of Neural Tangent Kernel in Classification Problems", "abs": "This paper primarily investigates the convergence of the Neural Tangent Kernel (NTK) in classification problems. This study firstly show the strictly positive definiteness of NTK of multi-layer fully connected neural networks and residual neural networks. Then, through a contradiction argument, it indicates that, during training with the cross-entropy loss function, the neural network parameters diverge due to the strictly positive definiteness of the NTK. Consequently, the empirical NTK does not consistently converge but instead diverges as time approaches infinity. This finding implies that NTK theory is not applicable in this context, highlighting significant theoretical implications for the study of neural networks in classification problems. These results can also be easily generalized to other network structures, provided that the NTK is strictly positive definite.", "keywords": [ "neural tangent kernel", "neural network", "cross-entropy loss", "classification problem" ], "rating_list": [ 3, 6, 6, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 1, 2, 2, 4 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "VD4PFpecG2", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Wenhao XU", "gender": "Male", "institution": "Chinese University of Hong Kong, The Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Xuedong He", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "Full Professor" }, { "name": "Xuefeng Gao", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "", "position": "Associate Professor" } ], "n_formula": 275, "n_formula_1": 43, "n_ref_uni": 35, "n_ref": 85, "n_ref_all": 153, "n_fig": 15, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 40250, "formula_len_all_1": 3109, "len_all": 254541, "len_all_1": 70660, "len_abs": 1003, "len_title": 116, "len_sents": 63245, "len_sents_1": 28982, "n_sents": 698, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1063, "title": "Regret Bounds for Episodic Risk-Sensitive Linear Quadratic Regulator", "abs": "Risk-sensitive linear quadratic regulator is one of the most fundamental problems in risk-sensitive optimal control. \nIn this paper, we study online adaptive control of risk-sensitive linear quadratic regulator in the finite horizon episodic setting. \nWe propose a simple least-squares greedy algorithm and show that it achieves $\\widetilde{\\mathcal{O}}(\\log N)$ regret under a specific identifiability assumption, where $N$ is the total number of episodes. If the identifiability assumption is not satisfied, we propose incorporating exploration noise into the least-squares-based algorithm, resulting in an algorithm with $\\widetilde{\\mathcal{O}}(\\sqrt{N})$ regret. \nTo our best knowledge, this is the first set of regret bounds for episodic risk-sensitive linear quadratic regulator. \nOur proof relies on perturbation analysis of less-standard Riccati equations for risk-sensitive linear quadratic control, and a delicate analysis of the loss in the risk-sensitive performance criterion due to applying the suboptimal controller in the online learning process.", "keywords": [ "reinforcement Learning", "regret", "LEQR", "finite horizon", "episodic setting" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "VCbqXtS5YY", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Alfredo Garcia", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "", "position": "Full Professor" }, { "name": "Chenliang Li", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "", "position": "PhD student" }, { "name": "Jiaxiang Li", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "Postdoc" }, { "name": "Mingyi Hong", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Siliang Zeng", "gender": "Male", "institution": "University of Minnesota, Twin Cities", "country": "US", "position": "PhD student" }, { "name": "Zeyi Liao", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "dongyeop kang", "gender": "Male", "institution": "University of Minnesota", "country": "", "position": "Assistant Professor" } ], "n_formula": 123, "n_formula_1": 38, "n_ref_uni": 60, "n_ref": 83, "n_ref_all": 129, "n_fig": 8, "n_tab": 9, "L_tab": 2237, "n_element_tab": 163, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 124, "n_element_tab_1": 3, "formula_len_all": 13866, "formula_len_all_1": 2703, "len_all": 243479, "len_all_1": 70820, "len_abs": 1450, "len_title": 135, "len_sents": 66955, "len_sents_1": 32315, "n_sents": 549, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 90, "L_abs": 1456, "title": "Joint Reward and Policy Learning with Demonstrations and Human Feedback Improves Alignment", "abs": "Aligning to human preferences and/or intentions is an important requirement for contemporary foundation models. To ensure alignment, popular approaches such as reinforcement learning with human feedback (RLHF) break down the task into three stages: (i) a model is computed with supervised fine-tuning (SFT) based upon large demonstrations data, (ii) a reward model (RM) is estimated based upon human feedback data, and (iii) reinforcement learning (RL) is used to further refine the SFT model by optimizing the estimated reward model. Demonstrations and human feedback data reflect human user preferences in different ways. As a result, the reward model estimate obtained from only human feedback data is likely not as accurate as a reward model estimate obtained from both demonstration and human feedback data. A policy model that optimizes the reward model estimate obtained from both demonstration and human feedback data will likely exhibit better alignment performance. We introduce a tractable algorithm for finding the reward and policy models and provide a finite-time performance guarantee. Additionally, we demonstrate the efficiency of the proposed solution with extensive experiments including alignment problems in LLMs and robotic control problems in MuJoCo. We observe that the proposed solutions outperform the existing alignment algorithm by large margins, especially when the amounts of demonstration and preference data are unbalanced.", "keywords": [ "Alignment", "Inverse Reinforcement Learning", "Reinforment Learning from Human Feedback" ], "rating_list": [ 10, 6, 8, 5 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 4, 3, 4, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "VCZ1o8gFny", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 35, "author_info_list": [ { "name": "Andrew Gu", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Boyang Liu", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Buyun Zhang", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Chen Chen", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Ellie Dingqiao Wen", "gender": "Female", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Fan Yang", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Huayu Li", "gender": "Female", "institution": "Meta", "country": "", "position": "Research Scientist" }, { "name": "Jade Nie", "gender": "Female", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Jiayi Xu", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Jiyan Yang", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Jongsoo Park", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Member of Technical Staff" }, { "name": "Laming Chen", "gender": "Male", "institution": "Meta", "country": "US", "position": "Research Scientist" }, { "name": "Liang Luo", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Longhao Jin", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Maxim Naumov", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Mengyue Hang", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Qianru Li", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Qin Huang", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Principal Researcher" }, { "name": "Shali Jiang", "gender": "Male", "institution": "Meta", "country": "US", "position": "Research Scientist" }, { "name": "Shiwen Shen", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Shuaiwen Wang", "gender": "unknown", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Siyang Yuan", "gender": "Female", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Tongyi Tang", "gender": "Female", "institution": "University of California, Davis", "country": "", "position": "PhD student" }, { "name": "Weilin Zhang", "gender": "unknown", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "MS student" }, { "name": "Wenjun Wang", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Wenlin Chen", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Xi Liu", "gender": "Male", "institution": "Meta AI", "country": "", "position": "Research Scientist" }, { "name": "Xiaohan Wei", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Xiaozhen Xia", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Yasmine Badr", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Yuchen Hao", "gender": "Male", "institution": "Meta", "country": "", "position": "Research Scientist" }, { "name": "Yuxin Chen", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Zeliang Chen", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Zhengyu Zhang", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "PhD student" }, { "name": "Lakshminarayanan Gunaseelan", "gender": "unknown", "institution": "Meta", "country": "", "position": "Principal Researcher" } ], "n_formula": 7, "n_formula_1": 2, "n_ref_uni": 73, "n_ref": 138, "n_ref_all": 148, "n_fig": 5, "n_tab": 3, "L_tab": 1120, "n_element_tab": 63, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 302, "formula_len_all_1": 67, "len_all": 194306, "len_all_1": 56234, "len_abs": 1292, "len_title": 159, "len_sents": 42117, "len_sents_1": 28080, "n_sents": 254, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 111, "L_abs": 1296, "title": "M3C: a Multi-Domain Multi-Objective, Mixed-Modality Framework for Cost-Effective, Industry Scale Recommendation", "abs": "The ever-expanding landscape of products, surfaces, policies, and regulations poses\nsignificant challenges for recommendation systems, leading to data fragmentation\nand prohibitive hikes in infrastructure costs. To address these challenges, we\npropose M3C, a holistic co-design of model, data and efficiency strategies. M3C\n(1) partitions the recommendation space to allow better representation learning\nand encourage knowledge sharing within a subspace; (2) covers each partition\nusing a hierarchy of foundational and vertical networks tailored to handle multi-\ndomain, multi-objective tasks with mixed-modal inputs; (3) forms a unified data\nrepresentation that utilizes heterogeneous signals across domains, objectives and\noptimization goals to alleviate data fragmentation, label sparsity, and to enhance\nknowledge sharing; (4) improves execution efficiency and lowers costs with a suite\nof stability and throughput optimizations. We show that across a diverse set of tasks\non public and industry datasets, M3C delivers up to 1% lower LogLoss compared\nto 10 state-of-the-art baselines, while improving system efficiency by up to 20%.\nFurthermore, in a large-scale industry setting our deployment of M3C has resulted\nin 7% top-line metrics improvement in online tests with 10% capacity savings.", "keywords": [ "Recommendation", "efficiency", "data consolidation" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "VBeLiRkZMP", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Chuxu Zhang", "gender": "Male", "institution": "University of Connecticut", "country": "US", "position": "Associate Professor" }, { "name": "Jiazheng Li", "gender": "Male", "institution": "University of Connecticut", "country": "US", "position": "PhD student" }, { "name": "Jundong Li", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" } ], "n_formula": 27, "n_formula_1": 25, "n_ref_uni": 25, "n_ref": 74, "n_ref_all": 102, "n_fig": 7, "n_tab": 7, "L_tab": 5329, "n_element_tab": 292, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 957, "formula_len_all_1": 1037, "len_all": 154285, "len_all_1": 62539, "len_abs": 1362, "len_title": 84, "len_sents": 46860, "len_sents_1": 30304, "n_sents": 361, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 36, "L_abs": 1366, "title": "Instance-Aware Graph Prompt Learning", "abs": "Graph neural networks stand as the predominant technique for graph representation learning owing to their strong expressive power, yet the performance highly depends on the availability of high-quality labels in an end-to-end manner. Thus the pretraining and fine-tuning paradigm has been proposed to mitigate the label cost issue. Subsequently, the gap between the pretext tasks and downstream tasks has spurred the development of graph prompt learning which inserts a set of graph prompts into the original graph data with minimal parameters while preserving competitive performance. However, the current exploratory works are still limited since they all concentrate on learning fixed task-specific prompts which may not generalize well across the diverse instances that the task comprises. To tackle this challenge, we introduce Instance-Aware Graph Prompt Learning (IA-GPL) in this paper, aiming to generate distinct prompts tailored to different input instances. The process involves generating intermediate prompts for each instance using a lightweight architecture, quantizing these prompts through trainable codebook vectors, and employing the exponential moving average technique to ensure stable training. Extensive experiments conducted on multiple datasets and settings showcase the superior performance of IA-GPL compared to state-of-the-art baselines.", "keywords": [ "Graph Neural Networks", "Prompt Learning", "Pre-training" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "VB8xHF1Rdl", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Henry Conklin", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "", "position": "PhD student" }, { "name": "Kenny Smith", "gender": "Male", "institution": "University of Edinburgh", "country": "", "position": "Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 38, "n_ref": 57, "n_ref_all": 66, "n_fig": 5, "n_tab": 2, "L_tab": 330, "n_element_tab": 56, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 325, "formula_len_all_1": 326, "len_all": 105957, "len_all_1": 60236, "len_abs": 1252, "len_title": 94, "len_sents": 35632, "len_sents_1": 30712, "n_sents": 235, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1259, "title": "Information Structure in Large Language Models", "abs": "Despite the widespread use of large language models, we still lack unified notation for thinking about and describing their representational spaces. This limits our ability to understand how they work. Ideally we would understand how their representations are structured, how that structure emerges over training, and what kinds of structures are desirable. Unfortunately we as humans tend not to have strong intuitions about high-dimensional vector spaces. Here we propose an information theoretic approach to quantifying structure in deep-learning models. We introduce a novel method for estimating the entropy of vector spaces, and use it to quantify the amount of information in the model we can explain with a set of labels. This can show when regularities emerge in representation space with respect to token, bigram, and trigram information in the input. As these models are learning from human language data, we formalise this in terms of 3 linguistically derived quantities: regularity, variation, and disentanglement. These show how larger models become proportionally more disentangled. We also are able to predict downstream task performance on GLUE benchmarks based on representational structure at the end of pre-training but before fine tuning.", "keywords": [ "Large Language Models", "Interpretability", "Information Theory", "Language" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 1, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "VAvZ4oinpa", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Debasish Ghose", "gender": "Male", "institution": "Indian Institute of Science", "country": "IN", "position": "Full Professor" }, { "name": "Devansh Bhardwaj", "gender": "Male", "institution": "Indian Institute of Technology, Roorkee, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "Undergrad student" }, { "name": "Meenakshi Sarkar", "gender": "unknown", "institution": "Indian Institute of Science", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 42, "n_ref": 82, "n_ref_all": 97, "n_fig": 5, "n_tab": 4, "L_tab": 154, "n_element_tab": 16, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 154, "n_element_tab_1": 16, "formula_len_all": 1349, "formula_len_all_1": 1347, "len_all": 126435, "len_all_1": 55502, "len_abs": 1482, "len_title": 90, "len_sents": 30983, "len_sents_1": 28149, "n_sents": 167, "n_sents_1": 152, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1485, "title": "Video Generation with Learned Action Prior", "abs": "Long-term stochastic video generation remains challenging, especially with moving cameras. This scenario introduces complex interactions between camera movement and observed pixels, resulting in intricate spatio-temporal dynamics and partial observability issues. Current approaches often focus on pixel-level image reconstruction, neglecting explicit modeling of camera motion dynamics. Our proposed solution incorporates camera motion or action as an extended part of the observed image state, employing a multi-modal learning framework to simultaneously model both image and action. We introduce three models: (i) Video Generation with Learning Action Prior (VG-LeAP) that treats the image-action pair as an augmented state generated from a single latent stochastic process and uses variational inference to learn the image-action latent prior; (ii) Causal-LeAP, which establishes a causal relationship between action and the observed image frame, and learns a seperate action prior, conditioned on the observed image states along with the image prior; and (iii) RAFI, which integrates the augmented image-action state concept with a conditional flow matching framework, demonstrating that this action-conditioned image generation concept can be extended to other transformer-based architectures. Through comprehensive empirical studies on robotic video dataset, RoAM, we highlight the importance of multi-modal training in addressing partially observable video generation problems.", "keywords": [ "Stochastic Video Generation", "Variational Inference" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 3, 3 ] }, { "paper_id": "VAqRZIuW8m", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Arun Kandoor", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Chih-Kuan Yeh", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "James Laudon", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Peter Schafhalter", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Yanqi Zhou", "gender": "Female", "institution": "Google Brain", "country": "", "position": "Research Scientist" }, { "name": "shun liao", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 50, "n_ref": 68, "n_ref_all": 88, "n_fig": 8, "n_tab": 5, "L_tab": 1110, "n_element_tab": 37, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1250, "n_element_tab_1": 28, "formula_len_all": 128, "formula_len_all_1": 128, "len_all": 132095, "len_all_1": 70948, "len_abs": 1162, "len_title": 122, "len_sents": 32504, "len_sents_1": 29835, "n_sents": 244, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1167, "title": "Scalable Multi-Domain Adaptation of Language Models using Modular Experts", "abs": "Domain-specific adaptation is critical to maximizing the performance of pre-trained\nlanguage models (PLMs) on one or multiple targeted tasks, especially under\nresource-constrained use cases, such as edge devices. However, existing methods often struggle to balance domain-specific performance, retention of general\nknowledge, and efficiency for training and inference. To address these challenges,\nwe propose Modular Domain Experts (MoDE). MoDE is a mixture-of-experts\narchitecture that augments a general PLMs with modular, domain-specialized\nexperts. These experts are trained independently and composed together via a\nlightweight training process. In contrast to standard low-rank adaptation methods,\neach MoDE expert consists of several transformer layers which scale better with\nmore training examples and larger parameter counts. Our evaluation demonstrates\nthat MoDE achieves comparable target performances to full parameter fine-tuning\nwhile achieving 1.65% better retention performance. Moreover, MoDE\u2019s architecture enables flexible sharding configurations and improves training speeds by\nup to 38% over state-of-the-art distributed training configurations.", "keywords": [ "domain adaptation", "mixture of experts", "MoE", "fine-tuning", "large language models", "sharding" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "VAeThhoJR2", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Nannan Li", "gender": "Female", "institution": "Boston University, Boston University", "country": "US", "position": "PhD student" }, { "name": "Bryan Allen Plummer", "gender": "Male", "institution": "Boston University", "country": "US", "position": "Assistant Professor" }, { "name": "Kevin Jonathan Shih", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 39, "n_ref": 103, "n_ref_all": 119, "n_fig": 8, "n_tab": 2, "L_tab": 486, "n_element_tab": 19, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1497, "n_element_tab_1": 130, "formula_len_all": 510, "formula_len_all_1": 536, "len_all": 109219, "len_all_1": 59506, "len_abs": 1416, "len_title": 126, "len_sents": 33527, "len_sents_1": 26467, "n_sents": 256, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1428, "title": "Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling", "abs": "Given an isolated garment image in a canonical product view and a separate image of a person, the virtual try-on task aims to generate a new image of the person wearing the target garment. Prior virtual try-on works face two major challenges in achieving this goal: a) the paired (human, garment) training data has limited availability; b) generating textures on the human that perfectly match that of the prompted garment is difficult, often resulting in distorted text and faded textures. Our work addresses these issues through a dual approach. First, we introduce a garment extraction model that generates (human, synthetic garment) pairs from a single image of a clothed individual. The synthetic pairs can then be used to augment the training of virtual try-on. Second, we propose an Error-Aware Refinement-based Schr\\\"odinger Bridge (EARSB) that surgically targets localized generation errors for correcting the output of a virtual try-on model. To identify likely errors, we propose a weakly-supervised error classifier that localizes regions for refinement, subsequently augmenting the Schr\\\"odinger Bridge's noise schedule with its confidence heatmap. Experiments on VITON-HD and DressCode-Upper demonstrate that our synthetic data augmentation enhances the performance of prior work, while EARSB improves the overall image quality. In user studies, our model is preferred by the users in an average of 59\\% of cases.", "keywords": [ "Virtual Try-On", "Human Image Generation", "Diffusion Models" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "VA1tNAsDiC", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Jiechuan Jiang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Intern" }, { "name": "Zongqing Lu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 35, "n_formula_1": 27, "n_ref_uni": 21, "n_ref": 56, "n_ref_all": 95, "n_fig": 16, "n_tab": 4, "L_tab": 87, "n_element_tab": 18, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4867, "formula_len_all_1": 4072, "len_all": 130940, "len_all_1": 70560, "len_abs": 1159, "len_title": 72, "len_sents": 39078, "len_sents_1": 26267, "n_sents": 352, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 24, "L_abs": 1190, "title": "Best Possible Q-Learning", "abs": "Fully decentralized learning, where the global information, \\textit{i.e.}, the actions of other agents, is inaccessible, is a fundamental challenge in cooperative multi-agent reinforcement learning. However, the convergence and optimality of most decentralized algorithms are not theoretically guaranteed, since the transition probabilities are non-stationary as all agents are updating policies simultaneously. To tackle this challenge, we propose \\textit{best possible operator}, a novel decentralized operator, and prove that the policies of cooperative agents will converge to the optimal joint policy if each agent independently updates its individual state-action value by the operator when there is only one optimal joint policy. Further, to make the update more efficient and practical, we simplify the operator and prove that the convergence and optimality still hold with the simplified one. By instantiating the simplified operator, the derived fully decentralized algorithm, \\textit{best possible Q-learning} (BQL), does not suffer from non-stationarity. Empirically, we show that BQL achieves remarkable improvement over baselines in a variety of cooperative multi-agent tasks.", "keywords": [ "reinforcement learning", "multi-agent reinforcement learning" ], "rating_list": [ 3, 1, 6, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "V9oT5Jmxpu", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Lizhen Qu", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Lecturer" }, { "name": "Minghao Wu", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Reza Haffari", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" }, { "name": "Trang Vu", "gender": "unknown", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 25, "n_ref": 92, "n_ref_all": 124, "n_fig": 12, "n_tab": 5, "L_tab": 3520, "n_element_tab": 270, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 3427, "n_element_tab_1": 156, "formula_len_all": 978, "formula_len_all_1": 1070, "len_all": 220799, "len_all_1": 65161, "len_abs": 1645, "len_title": 142, "len_sents": 40183, "len_sents_1": 27371, "n_sents": 294, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1652, "title": "The Best of Both Worlds: Bridging Quality and Diversity in Data Selection with Bipartite Graph", "abs": "The performance of large language models (LLMs) in natural language processing (NLP) tasks is significantly influenced by the quality and diversity of data used for supervised fine-tuning (SFT). Current data selection methods often focus solely on quality or diversity, leading to underperforming models due to suboptimal training data. In this paper, we introduce GraphFilter, a novel method that represents the dataset as a bipartite graph, linking sentences to their constituent n-grams. This representation effectively captures the relationships between sentences and linguistic patterns, facilitating the selection of sentences that enhance n-gram diversity. To balance quality and diversity during selection, we propose a priority function that combines the quality metric with the diversity metric in a multiplicative manner. GraphFilter iteratively selects high-priority sentences, updates the bipartite graph by removing covered n-grams, and re-calculates priorities to reflect the evolving data landscape. We conduct extensive experiments using three model backbones across six widely used benchmarks. The results demonstrate that GraphFilter outperforms all nine baseline approaches, achieving superior model performance and computational efficiency. Our analyses validate the effectiveness of our design choices, examine the subsets selected by GraphFilter and other methods, highlight the importance of instruction diversity, and explore the role of quality and diversity in relation to subset sizes. GraphFilter establishes a new foundation for effective data selection strategies, encouraging further research in data selection for LLMs.", "keywords": [ "large language models", "supervised fine-tuning", "data selection", "efficiency" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "V9UsZBbTvZ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Enhui Chai", "gender": "Female", "institution": "Northwestern University", "country": "CN", "position": "PhD student" }, { "name": "Tianxiang Cui", "gender": "Male", "institution": "The University of Nottingham Ningbo China", "country": "", "position": "Assistant Professor" }, { "name": "Xingyu Li", "gender": "unknown", "institution": "\u897f\u5317\u5927\u5b66", "country": "CN", "position": "Undergrad student" }, { "name": "Zheng Lu", "gender": "Male", "institution": "The University of Nottingham Ningbo China", "country": "", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 13, "n_ref_uni": 15, "n_ref": 22, "n_ref_all": 36, "n_fig": 4, "n_tab": 4, "L_tab": 1920, "n_element_tab": 240, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 4576, "n_element_tab_1": 169, "formula_len_all": 877, "formula_len_all_1": 739, "len_all": 106244, "len_all_1": 60534, "len_abs": 1440, "len_title": 154, "len_sents": 27215, "len_sents_1": 25475, "n_sents": 211, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1446, "title": "Masked Mamba: An Efficient Self-Supervised Framework for Pathological Image Classification", "abs": "Extracting visual representations is a crucial challenge in the domain of computational histopathology. Considering the powerful strength of deep learning algorithms and the dearth of annotated samples, self-supervised learning presents itself as a compelling strategy to extract effective visual representations from unlabeled histopathology images. Although some self-supervised learning methods have been specifically proposed for histopathology image classification, most of them have certain drawbacks that may affect the functionality or representation capacity. In this work, we propose Masked Mamba, a novel self-supervised visual representation learning method tailored for histopathology images that can adequately extract local-global features. The proposed method consists of two stages: local perception positional encoding (LPPE) and directional Mamba vision backbone (DM). In addition, we use masked autoencoder (MAE) pretraining to unleashing directional Mamba vision backbone's potential. Masked Mamba makes good use of domain-specific knowledge and requires no side information, which means good rationality and versatility. Experimental results demonstrate the effectiveness and robustness of masked Mamba on common histopathology classification tasks. Furthermore, ablation studies prove that the local perception positional encoding and directional Mamba vision backbone in masked Mamba can complement and enhance each other.", "keywords": [ "Pathological image classification", "Mamba model", "Self-supervised learning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 2 ] }, { "paper_id": "V8cMqUZT8o", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Geng Zhao", "gender": "Male", "institution": "City University of Hong Kong", "country": "", "position": "Postdoc" }, { "name": "ZhongYu Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 15, "n_ref": 28, "n_ref_all": 34, "n_fig": 2, "n_tab": 3, "L_tab": 2355, "n_element_tab": 60, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 3065, "n_element_tab_1": 70, "formula_len_all": 450, "formula_len_all_1": 450, "len_all": 103375, "len_all_1": 58732, "len_abs": 1528, "len_title": 145, "len_sents": 27339, "len_sents_1": 26653, "n_sents": 209, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1512, "title": "Text Attributed Graph Node Classification Using Sheaf Neural Networks and Large Language Models", "abs": "Text-Attributed Graphs (TAGs) seamlessly integrate textual data with graph structures, presenting unique challenges and opportunities for jointly modeling text and graph information. Recent advancements in Large Language Models (LLMs) have significantly enhanced the generative and predictive capabilities of text modeling. However, existing graph models often fall short in capturing intricate node relationships, as their edge representations are typically limited to scalar values.\n\nIn this paper, we introduce \\model, a novel method that encodes rich and complex relational information between nodes as edge vectors. During the message-passing phase, \\model aggregates both neighbor node representations and edge vectors to update the central node's representation, eliminating the need to fine-tune the LLMs on the text-attributed graph.\n\nSpecifically, for a given TAG, \\model is trained to minimize the prediction errors of the LLM in forecasting the next word in node text sequences. Furthermore, we enhance \\model's performance by incorporating prompt-based fine-tuning techniques. Once trained, \\model can be seamlessly adapted to various downstream tasks.\n\nExtensive node classification experiments across multiple domains demonstrate that \\model consistently achieves state-of-the-art performance, validating its effectiveness in capturing complex relationships within TAGs. Additionally, we conduct ablation studies and scalability analyses to ensure the robustness and applicability of our approach.", "keywords": [ "Text Attributed Graph", "Sheaf Neural Networks", "Large Language Models" ], "rating_list": [ 1, 5, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "V892sBHUbN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Alwin Peng", "gender": "Male", "institution": "SERI MATS", "country": "US", "position": "Researcher" }, { "name": "Ethan Perez", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Henry Sleight", "gender": "Male", "institution": "Constellation", "country": "US", "position": "Researcher" }, { "name": "Julian Michael", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "Mrinank Sharma", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 46, "n_ref": 76, "n_ref_all": 96, "n_fig": 13, "n_tab": 2, "L_tab": 700, "n_element_tab": 12, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 176141, "len_all_1": 63983, "len_abs": 1379, "len_title": 109, "len_sents": 51199, "len_sents_1": 33960, "n_sents": 361, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1384, "title": "Rapid Response: Mitigating LLM Jailbreaks With A Few Examples", "abs": "As large language models (LLMs) grow more powerful, ensuring their safety against misuse becomes crucial. While researchers have focused on developing robust defenses, no method has yet achieved complete invulnerability to attacks. We propose an alternative approach: instead of seeking perfect adversarial robustness, we develop rapid response techniques to look to block whole classes of jailbreaks after observing only a handful of attacks.\nTo study this setting, we develop RapidResponseBench, a benchmark that measures a defense's robustness against various jailbreak strategies after adapting to a few observed examples.\nWe evaluate five rapid response methods, all of which use jailbreak proliferation, where we automatically generate additional jailbreaks similar to the examples observed. Our strongest method, which fine-tunes an input classifier to block proliferated jailbreaks, reduces attack success rate by a factor greater than 240 on an in-distribution set of jailbreaks and a factor greater than 15 on an out-of-distribution set, having observed just one example of each jailbreaking strategy. Moreover, further studies suggest that the quality of proliferation model and number of proliferated examples play an key role in the effectiveness of this defense. Overall, our results highlight the potential of responding rapidly to novel jailbreaks to limit LLM misuse.", "keywords": [ "safety", "robustness" ], "rating_list": [ 5, 5, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "V83xzYnZ5q", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 1, "author_info_list": [ { "name": "Li Danyu", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 38, "n_ref": 72, "n_ref_all": 94, "n_fig": 11, "n_tab": 6, "L_tab": 2486, "n_element_tab": 199, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 278, "n_element_tab_1": 55, "formula_len_all": 791, "formula_len_all_1": 746, "len_all": 147175, "len_all_1": 61425, "len_abs": 1339, "len_title": 150, "len_sents": 28304, "len_sents_1": 26003, "n_sents": 168, "n_sents_1": 147, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1342, "title": "Mechanism-Empowered Multivariate Time Series Forecasting Model: Application to Tuberculosis Prediction", "abs": "Among the current global health challenges, tuberculosis, as a highly contagious chronic disease, remains one of the major public health problems worldwide. Despite significant progress made in the past decades, new challenges, including systematic and effective downscaling, accurate prediction of disease incidence, and implementation of source reduction measures, have added to the difficulty of tuberculosis control. In view of the limitations of the recently proposed EIGHT prediction models in terms of prediction accuracy, this study adopts the Learnable Decomposition and Dual Focus Module Model (Leddam) and then introduces a novel mechanism-supported multivariate spatiotemporal series framework, termed LCHHA-Leddam, to address the challenges in tuberculosis forecasting through an investigation of coal power generation in China. This framework substantially simplifies the complexity of tuberculosis prediction, enhances accurate dimensionality reduction, and improves traceability. It also enhances the explanatory power and accuracy of the Leddam model in the field of tuberculosis prediction. This study provides a fresh perspective for enhancing epidemic forecasting and exploring source reduction measures for industrial activities, demonstrating the feasibility of AI-assisted public health strategies and green production.", "keywords": [ "Multivariate Time Series Forecasting Model", "Spatiotemporal framework", "Source reduction", "mechanism-driven dimensionality reduction" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "V7QRVEZ0le", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Cheng Liang", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Gangshan Wu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Jiaming Zhang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Xiangbo Shu", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yutao Cui", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Limin Wang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 36, "n_ref": 98, "n_ref_all": 116, "n_fig": 7, "n_tab": 4, "L_tab": 1032, "n_element_tab": 128, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 4379, "n_element_tab_1": 289, "formula_len_all": 736, "formula_len_all_1": 736, "len_all": 138699, "len_all_1": 63308, "len_abs": 1153, "len_title": 86, "len_sents": 35217, "len_sents_1": 25775, "n_sents": 225, "n_sents_1": 165, "venue": "ICLR2025", "decision": "withdraw", "L_title": 38, "L_abs": 1158, "title": "TrackMamba: Mamba-Transformer Tracking", "abs": "Current one-stream Transformer-based trackers are quality but unfriendly to memory consumption of large resolution and long sequence, both of which are crucial keys to tracking tasks. Recently structured state space model (SSM) demonstrates promising performance and efficiency in sequence modeling but struggles to retrieve due to the limited hidden state number. To solve the computation challenge and explore the potential of Mamba, we propose TrackMamba, a Mamba-Transformer tracker containing TrackMamba Blocks and Attention Blocks. In order to better harness the scanning in TrackMamba Blocks for inter- and intra-frame modeling, we introduce various scan patterns for rearrangement and flipping. Furthermore, we propose Target Enhancement, including Temporal Token for target aggregation and search enhancement, and Temporal Mamba for target information cross-frame propagation. Extensive experiments show TrackMamba performs better than the first-generation one-stream Transformer-based tracker at same resolution and mitigates consumption growth when enlarging resolution, exhibiting the potential of Mamba-based model for large-resolution tracking.", "keywords": [ "Mamba; Single Object Tracking" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "V7PYbRzD0h", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Jen-tse Huang", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Postdoc" }, { "name": "Gao Kuiyi", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Undergrad student" }, { "name": "Qiuzhi Liu", "gender": "unknown", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Shuai Wang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Wenxiang Jiao", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Wenxuan Wang", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Postdoc" }, { "name": "Youliang Yuan", "gender": "Male", "institution": "The Chinese University of Hong Kong-Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Zhaopeng Tu", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Zihan Jia", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Undergrad student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 49, "n_ref": 82, "n_ref_all": 112, "n_fig": 8, "n_tab": 9, "L_tab": 2355, "n_element_tab": 134, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 1007, "n_element_tab_1": 81, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 162384, "len_all_1": 62038, "len_abs": 1880, "len_title": 127, "len_sents": 42499, "len_sents_1": 27616, "n_sents": 325, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1650, "title": "Chain-of-Jailbreak Attack for Image Generation Models via Editing Step by Step", "abs": "Text-based image generation models, such as Stable Diffusion and DALL-E 3, hold significant potential in content creation and publishing workflows, making them the focus in recent years.\nDespite their remarkable capability to generate diverse and vivid images, considerable efforts are being made to prevent the generation of harmful content, such as abusive, violent, or pornographic material.\nTo assess the safety of existing models, we introduce a novel jailbreaking method called Chain-of-Jailbreak (CoJ) attack, which compromises image generation models through a step-by-step editing process.\nSpecifically, for malicious queries that cannot bypass the safeguards with a single prompt, we intentionally decompose the query into multiple sub-queries. The image generation models are then prompted to generate and iteratively edit images based on these sub-queries.\nTo evaluate the effectiveness of our CoJ attack method, we constructed a comprehensive dataset, CoJ-Bench, encompassing nine safety scenarios, three types of editing operations, and three editing elements.\nExperiments on four widely-used image generation services provided by GPT-4V, GPT-4o, Gemini 1.5 and Gemini 1.5 Pro, demonstrate that our CoJ attack method can successfully bypass the safeguards of models for over 60\\% cases, which significantly outperforms other jailbreaking methods (i.e., 14\\%).\nFurther, to enhance these models' safety against our CoJ attack method, we also propose an effective prompting-based method, Think Twice Prompting, that can successfully defend over 95\\% of CoJ attack.\nWe will release our dataset and code to facilitate the AI safety research.", "keywords": [ "Jailbreak Attack", "Text-to-Image Models", "Safety" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "V7EiYG5DwZ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Khai Nguyen", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Baraniuk R.", "gender": "unknown", "institution": "William Marsh Rice University", "country": "", "position": "C. Sidney Burrus Professor" }, { "name": "Tam Minh Nguyen", "gender": "Female", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Ngoc Ngo Quang Tran", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "PhD student" } ], "n_formula": 66, "n_formula_1": 41, "n_ref_uni": 26, "n_ref": 33, "n_ref_all": 63, "n_fig": 3, "n_tab": 10, "L_tab": 5219, "n_element_tab": 347, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 669, "n_element_tab_1": 78, "formula_len_all": 6577, "formula_len_all_1": 3422, "len_all": 158434, "len_all_1": 63348, "len_abs": 1424, "len_title": 129, "len_sents": 45116, "len_sents_1": 26336, "n_sents": 383, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1426, "title": "Mutual-Inform SMoE: Improving Routing Stability via Probabilistic Graphical Model", "abs": "Sparse Mixture of Experts (SMoE) has emerged as a breakthrough approach for achieving unprecedented scalability in deep learning. By enabling models to expand their parameter count exponentially while selectively activating only a small subset of parameters per sample, SMoEs maintain high efficiency. However, SMoE models are susceptible to routing fluctuations, leading to instability and non-robustness. In this work, we unveils SMoE-based attention as a point estimate of a regression function of a 3-layer hierarchical mixture of experts regression. Through this probabilistic graphical model (PGM) framework, we highlight the conditional independence in expert-selection process of tokens, which exposes the model to routing fluctuation and non-robustness. Motivating by this PGM framework, we propose Mutual-Inform SMoEs, including Similarity and Attention-Inform SMoE, which eliminate the assumption of conditional independence by allowing tokens to directly influence each other on expert-decisions. We theoretically demonstrate that our methods lower the entropy in decision-making, enabling more confident and consistent expert assignments. Finally, we empirically validate our models on ImageNet classification and Wikitext-103 language modeling, showing significant improvements in reducing routing fluctuations, enhancing performance, and increasing model robustness compared to baseline Transformer-SMoE models.", "keywords": [ "mixture of expert", "transformer", "probabilistic graphical model", "robustness" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "V73W8MXnNW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Xuemei Xie", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Leizhenhua", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 30, "n_formula_1": 20, "n_ref_uni": 32, "n_ref": 85, "n_ref_all": 102, "n_fig": 6, "n_tab": 5, "L_tab": 3091, "n_element_tab": 331, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 233, "n_element_tab_1": 4, "formula_len_all": 1102, "formula_len_all_1": 703, "len_all": 184679, "len_all_1": 65650, "len_abs": 1245, "len_title": 89, "len_sents": 54395, "len_sents_1": 30547, "n_sents": 524, "n_sents_1": 274, "venue": "ICLR2025", "decision": "withdraw", "L_title": 41, "L_abs": 1257, "title": "Progressive Visual Relationship Inference", "abs": "As an important component of visual scene, visual relationship has received extensive attention in recent years.\nMost existing works directly utilize the rough visual appearance to represent visual relationships.\nAlthough they have been made tremendous progress, the study of visual relationship may be still far from perfect.\nThis common idea may have three problems.\n1) The similarity of space aggravates the ambiguity of predicate representation.\n2) The differences between many visual relationships are subtle.\n3) It lacks interpretability.\nTo address these problems, we propose a novel method - Progressive Visual Relationship Inference(\\PVRI) - which considers both rough visual appearance and fine-grained visual cues to gradually infer visual relationships.\nIt includes the following three steps.\n1) Known Cues Collection:\nfirstly, we utilize Large Language Model(LLM) to collect the cues that may help infer visual relationships;\n2) Unknown Cues Extraction:\nsecondly, we design UCE strategy to extract the cues that are not defined by the text.\n3) Progressive Inference:\nthirdly, we utilize the obtained cues to infer visual relationships.\nWe demonstrate the effectiveness and efficiency of our method for the Visual Genome, Open Image V6 datasets.", "keywords": [ "scene graph generation", "visual relationship detection", "visual scene understanding" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 1, 1, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "V71ITh2w40", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Morten M\u00f8rup", "gender": "Male", "institution": "Technical University of Denmark", "country": "", "position": "Professor" }, { "name": "Niels Raunkj\u00e6r Holm", "gender": "Male", "institution": "Technical University of Denmark", "country": "", "position": "MS student" }, { "name": "Nikolaos Nakis", "gender": "unknown", "institution": "Yale University", "country": "US", "position": "Postdoc" }, { "name": "Andreas Fiehn", "gender": "unknown", "institution": "Technical University of Denmark", "country": "DK", "position": "MS student" } ], "n_formula": 35, "n_formula_1": 22, "n_ref_uni": 45, "n_ref": 115, "n_ref_all": 148, "n_fig": 13, "n_tab": 6, "L_tab": 2723, "n_element_tab": 468, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 3956, "n_element_tab_1": 307, "formula_len_all": 2189, "formula_len_all_1": 1248, "len_all": 172191, "len_all_1": 75042, "len_abs": 1756, "len_title": 160, "len_sents": 51740, "len_sents_1": 31582, "n_sents": 343, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 111, "L_abs": 1761, "title": "How Low Can You Go? Searching for the Intrinsic Dimensionality of Complex Networks using Metric Node Embeddings", "abs": "Low-dimensional embeddings are essential for machine learning tasks involving graphs, such as node classification, link prediction, community detection, network visualization, and network compression. Although recent studies have identified exact low-dimensional embeddings, the limits of the required embedding dimensions remain unclear. We presently prove that lower dimensional embeddings are possible when using metric embeddings as opposed to vector-based inner product embeddings such as Logistic PCA (LPCA). We further provide an efficient logarithmic search procedure for identifying the exact embedding dimension and demonstrate how metric embeddings enable inference of the exact embedding dimensions of large-scale networks by exploiting that the metric properties can be used to provide linearithmic scaling. Empirically, we show that our approach extracts substantially lower dimensional representations of networks than previously reported for small-sized networks. For the first time, we demonstrate that even large-scale networks can be effectively embedded in very low-dimensional spaces, and provide examples of scalable, exact reconstruction for graphs with up to a million nodes. Our approach highlights that the intrinsic dimensionality of networks is substantially lower than previously reported and provides a computationally efficient assessment of the exact embedding dimension also of large-scale networks. The surprisingly low dimensional representations achieved demonstrate that networks in general can be losslessly represented using very low dimensional feature spaces, which can be used to guide existing network analysis tasks from community detection and node classification to structure revealing exact network visualizations.", "keywords": [ "Exact network embedding", "latent distance model", "logistic PCA", "large scale network modeling" ], "rating_list": [ 6, 6, 5, 8, 6 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 1, 2 ], "contribution_list": [ 3, 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 3, 5 ] }, { "paper_id": "V6uxd8MEqw", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Karteek Alahari", "gender": "Male", "institution": "Inria", "country": "FR", "position": "Research director (eq. Full professor)" }, { "name": "Liyuan Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Xingxing Zhang", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Zhiqi KANG", "gender": "unknown", "institution": "INRIA", "country": "FR", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 44, "n_ref": 108, "n_ref_all": 118, "n_fig": 5, "n_tab": 10, "L_tab": 2862, "n_element_tab": 394, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 981, "n_element_tab_1": 34, "formula_len_all": 747, "formula_len_all_1": 474, "len_all": 169671, "len_all_1": 64526, "len_abs": 1596, "len_title": 128, "len_sents": 52142, "len_sents_1": 31353, "n_sents": 420, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1507, "title": "Advancing Prompt-Based Methods for Replay-Independent General Continual Learning", "abs": "General continual learning (GCL) is a broad concept to describe real-world continual learning (CL) problems, which are often characterized by online data streams without distinct transitions between tasks, i.e., blurry task boundaries. These requirements result in poor initial performance, limited generalizability, and severe catastrophic forgetting, heavily impacting the effectiveness of mainstream GCL models trained from scratch. While the use of a frozen pretrained backbone with appropriate prompt tuning can partially address these challenges, such prompt-based methods remain sub-optimal for CL of remaining tunable parameters on the fly. In this regard, we propose an innovative approach named MISA(Mask and Initial Session Adaption) to advance prompt-based methods in GCL. It includes a forgetting-aware initial session adaption that employs pretraining data to initialize prompt parameters and improve generalizability, as well as a non-parametric logit mask of the output layers to mitigate catastrophic forgetting. Empirical results demonstrate substantial performance gains of our approach compared to recent competitors, especially without a replay buffer (e.g., up to 18.39%, 22.06%, and 11.96% performance lead on CIFAR-100, Tiny-ImageNet, and ImageNet-R, respectively). Moreover, our approach features the plug-in nature for prompt-based methods, independence of replay, ease of implementation, and avoidance of CL-relevant hyperparameters, serving as a strong baseline for GCL research.", "keywords": [ "Continual Learning", "General Continual Learning", "Catastrophic Forgetting", "Prompt Tuning", "Sharpness-Aware Minimization" ], "rating_list": [ 8, 8, 5, 5 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "V6TD4io8Gu", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Lei Wei", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Minsheng Hao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Mo Chen", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xuegong Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 27, "n_ref": 30, "n_ref_all": 43, "n_fig": 6, "n_tab": 6, "L_tab": 1828, "n_element_tab": 241, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1446, "n_element_tab_1": 189, "formula_len_all": 868, "formula_len_all_1": 640, "len_all": 106188, "len_all_1": 56285, "len_abs": 1237, "len_title": 163, "len_sents": 29755, "len_sents_1": 24035, "n_sents": 219, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 115, "L_abs": 1241, "title": "QueST: Querying Functional and Structural Niches on Spatial Transcriptomics data via Contrastive Subgraph Embedding", "abs": "The functional or structural spatial regions within tissues, referred to as spatial niches, are elements for illustrating the spatial contexts of multicellular organisms. A key challenge is querying shared niches across diverse tissues, which is crucial for achieving a comprehensive understanding of the organization and phenotypes of cell populations. However, current data analysis methods predominantly focus on creating spatial-aware embeddings for cells, neglecting the development of niche-level representations for effective querying. To address this gap, we introduce QueST, a novel niche representation learning model designed for querying spatial niches across multiple samples. QueST utilizes a novel subgraph contrastive learning approach to explicitly capture niche-level characteristics and incorporates adversarial training to mitigate batch effects. We evaluate QueST on established benchmarks using human and mouse datasets, demonstrating its superiority over state-of-the-art graph representation learning methods in accurate niche queries. Overall, QueST offers a specialized model for spatial niche queries, paving the way for deeper insights into the patterns and mechanisms of cell spatial organization across tissues.", "keywords": [ "Graph neural networks; Subgraph contrastive learning; Spatial niche query; Spatial transcriptomics; Batch removal" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 5, 3, 5 ] }, { "paper_id": "V6AI97jJ3J", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Liansheng Zhuang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "MingHong Yao", "gender": "Male", "institution": "cyber security", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 21, "n_ref": 47, "n_ref_all": 68, "n_fig": 5, "n_tab": 2, "L_tab": 1414, "n_element_tab": 168, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1414, "n_element_tab_1": 168, "formula_len_all": 589, "formula_len_all_1": 589, "len_all": 114920, "len_all_1": 64544, "len_abs": 1258, "len_title": 123, "len_sents": 29703, "len_sents_1": 28183, "n_sents": 240, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1263, "title": "UniVIEDM: A Diffusion Model to Unify Visual Information Extraction Subtasks", "abs": "Visual Information Extraction (VIE) focuses on extracting named entities and their relationships from visually rich document images. Traditionally, VIE systems rely on three separate models to handle three distinct subtasks, but the emerging trend in research is to design a single model that can address all of these tasks simultaneously. However, current methods face quadratic computational complexity when extracting entity relationships, as they must iterate over all token pairs. To address this issue, this paper introduces a Unified VIE Diffusion Model (UniVIEDM) for all tasks within VIE. UniVIEDM generates entity labels and their relationships conditioned on their plane coordinates, greatly reducing the computational complexity. UniVIEDM represents the layout of each visually rich document as a plane graph and converts the three subtasks into plane graph generation problems. During the pre-training stage, UniVIEDM leverages a jump-diffusion process to learn to generate valid sets of bounding boxes for all words and line segments connecting different boxes. During the fine-tuning stage, UniVIEDM employs a continuous-time Markov chain diffusion model to learn to predict the labels of boxes and line segments based on their coordinate features.", "keywords": [ "Visual Information Extraction" ], "rating_list": [ 5, 1, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 1, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "V5ns6uvRZ9", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Ittai Rubinstein", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Samuel B. Hopkins", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 317, "n_formula_1": 23, "n_ref_uni": 14, "n_ref": 52, "n_ref_all": 107, "n_fig": 10, "n_tab": 7, "L_tab": 1072, "n_element_tab": 200, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 735, "n_element_tab_1": 116, "formula_len_all": 27884, "formula_len_all_1": 2129, "len_all": 329172, "len_all_1": 65264, "len_abs": 1475, "len_title": 69, "len_sents": 117358, "len_sents_1": 28423, "n_sents": 1073, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1402, "title": "Robustness Auditing for Linear Regression: To Singularity and Beyond", "abs": "It has recently been discovered that the conclusions of many highly influential econometrics studies can be overturned by removing a very small fraction of their samples (often less than $0.5\\%$). These conclusions are typically based on the results of one or more Ordinary Least Squares (OLS) regressions, raising the question: given a dataset, can we certify the robustness of an OLS fit on this dataset to the removal of a given number of samples?\n\nBrute-force techniques quickly break down even on small datasets. Existing approaches which go beyond brute force either can only find candidate small subsets to remove (but cannot certify their non-existence) [BGM20, KZC21], are computationally intractable beyond low dimensional settings [MR22], or require very strong assumptions on the data distribution and too many samples to give reasonable bounds in practice [BP21, FH23]. \n\nWe present an efficient algorithm for certifying the robustness of linear regressions to removals of samples. We implement our algorithm and run it on several landmark econometrics datasets with hundreds of dimensions and tens of thousands of samples, giving the first non-trivial certificates of robustness to sample removal for datasets of dimension $4$ or greater. We prove that under distributional assumptions on a dataset, the bounds produced by our algorithm are tight up to a $1 + o(1)$ multiplicative factor.", "keywords": [ "Robust machine learning", "linear regression", "robustness auditing", "data attribution", "ordinary least squares", "robust statistics", "econometrics" ], "rating_list": [ 5, 5, 8 ], "soundness_list": [ 2, 2, 4 ], "presentation_list": [ 2, 2, 4 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "V5lBNcD65H", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Hongwang Xiao", "gender": "unknown", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Kai Chen", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Qiwei Ye", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Principal Researcher" }, { "name": "Shan Yu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Sicheng Dai", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 39, "n_ref": 63, "n_ref_all": 75, "n_fig": 5, "n_tab": 7, "L_tab": 6357, "n_element_tab": 457, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 7572, "n_element_tab_1": 360, "formula_len_all": 758, "formula_len_all_1": 701, "len_all": 132721, "len_all_1": 62123, "len_abs": 1714, "len_title": 158, "len_sents": 29461, "len_sents_1": 23623, "n_sents": 207, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1722, "title": "MTEEG: A Multi-Task Learning Framework for Enhanced Electroencephalography Analysis Using Low-Rank Adaptation", "abs": "Electroencephalography (EEG) analysis using deep learning has traditionally placed a strong emphasis on models that are custom-built and optimized for specific datasets. Several recent research utilize self-supervised learning to extract generic representations from massive amounts of unlabeled EEG data. The pre-trained models are then fine-tuned on each downstream dataset independently, demonstrating promising results. However, in practical applications involving multiple tasks, utilizing a separate model for each is not ideal regarding computational and spatial cost. In this study, we go one step further and explore the simultaneous adaptation of a pre-trained model to multiple different tasks. The EEG signals exhibit significant heterogeneity due to their collection from various subjects using diverse devices and experimental setups, resulting in potential conflicts among different tasks that impede joint optimization. To tackle this challenge, we propose MTEEG, a multi-task EEG recognition framework which incorporates a task-agnostic temporal encoder and task-specific low-rank adaptation modules to disentangle the parameter space, facilitating both task interaction and specification. Experiments show that MTEEG surpasses other multi-task methods and performs on par with state-of-the-art single-task methods on abnormal detection, event type classification, emotion recognition, seizure detection, sleep stage classification and motor imagery classification after being tuned jointly on six publicly available datasets. MTEEG shows the potential of multi-task EEG recognition and promotes the development of general-purpose brain-computer interfaces in the future. The source code will be released.", "keywords": [ "EEG", "brain-computer interface", "multi-task learning" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "V5kCKFav9j", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Holakou Rahmanian", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Kiran Koshy Thekumparampil", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Lexing Ying", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Professor" }, { "name": "Michael Shavlovsky", "gender": "Male", "institution": "Amazon", "country": "US", "position": "PhD student" }, { "name": "Tesi Xiao", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Xun Tang", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 111, "n_formula_1": 31, "n_ref_uni": 77, "n_ref": 126, "n_ref_all": 140, "n_fig": 13, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10397, "formula_len_all_1": 2058, "len_all": 214613, "len_all_1": 63793, "len_abs": 1440, "len_title": 59, "len_sents": 56772, "len_sents_1": 27937, "n_sents": 499, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1445, "title": "A Sinkhorn-type Algorithm for Constrained Optimal Transport", "abs": "Entropic optimal transport (OT) and the Sinkhorn algorithm have made it practical for machine learning practitioners to perform the fundamental task of calculating transport distance between statistical distributions. In this work, we focus on a general class of OT problems under a combination of equality and inequality constraints. We derive the corresponding entropy regularization formulation and introduce a Sinkhorn-type algorithm for such constrained OT problems supported by theoretical guarantees. We first bound the approximation error when solving the problem through entropic regularization, which reduces exponentially with the increase of the regularization parameter. Furthermore, we prove a sublinear first-order convergence rate of the proposed Sinkhorn-type algorithm in the dual space by characterizing the optimization procedure with a Lyapunov function. To achieve fast and higher-order convergence under weak entropy regularization, we augment the Sinkhorn-type algorithm with dynamic regularization scheduling and second-order acceleration. Overall, this work systematically combines recent theoretical and numerical advances in entropic optimal transport with the constrained case, allowing practitioners to derive approximate transport plans in complex scenarios. In addition, we extend the formulation of this work to partial optimal transport and propose a fast algorithm with practical super-exponential convergence.", "keywords": [ "Optimal Transport", "Constrained Optimization" ], "rating_list": [ 3, 8, 6, 6 ], "soundness_list": [ 2, 4, 4, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 1, 4, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "V5am4S9eUd", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "M. Emre Gursoy", "gender": "Male", "institution": "Ko\u00e7 University", "country": "TR", "position": "Assistant Professor" }, { "name": "\u00d6mer Faruk TAL", "gender": "Male", "institution": "Ko\u00e7 University", "country": "", "position": "Undergrad student" }, { "name": "\u00d6mer Veysel \u00c7a\u011fatan", "gender": "Male", "institution": "Ko\u00e7 University", "country": "TR", "position": "Undergrad student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 66, "n_ref": 113, "n_ref_all": 119, "n_fig": 1, "n_tab": 20, "L_tab": 21045, "n_element_tab": 2900, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 757, "n_element_tab_1": 46, "formula_len_all": 144, "formula_len_all_1": 0, "len_all": 284564, "len_all_1": 48984, "len_abs": 1160, "len_title": 108, "len_sents": 57740, "len_sents_1": 23643, "n_sents": 366, "n_sents_1": 144, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1164, "title": "Adversarial Robustness of Self-Supervised Learning in Vision", "abs": "Self-supervised learning (SSL) has advanced significantly in visual representation learning, yet large-scale evaluations of its adversarial robustness remain limited. In this study, we evaluate the adversarial robustness of seven SSL models and one supervised model across a range of tasks, including ImageNet classification, transfer learning, segmentation, and detection. Our findings demonstrate that SSL models generally exhibit superior robustness to adversarial attacks compared to their supervised counterpart on ImageNet, with this advantage extending to transfer learning in classification tasks. However, this robustness is less pronounced in segmentation and detection tasks. We also explore the role of architectural choices in model robustness, observing that their impact varies depending on the SSL objective. Finally, we assess the effect of extended training durations on adversarial robustness, finding that longer training may offer slight improvements without compromising robustness. Our analysis highlights promising directions for enhancing the adversarial robustness of visual self-supervised representation systems in complex environments.", "keywords": [ "Self Supervised Learning", "Adversarial Attacks", "Robustness" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "V5Zn0VVvBE", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Liuyin Yang", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "PhD student" }, { "name": "Marc M. Van Hulle", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 60, "n_ref": 114, "n_ref_all": 140, "n_fig": 4, "n_tab": 21, "L_tab": 13032, "n_element_tab": 1490, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2052, "n_element_tab_1": 418, "formula_len_all": 1773, "formula_len_all_1": 0, "len_all": 254484, "len_all_1": 66385, "len_abs": 2445, "len_title": 145, "len_sents": 68199, "len_sents_1": 31897, "n_sents": 418, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 2138, "title": "Learning Robust EEG Representations with a Large Spatiotemporal Transformer as a Foundation Model", "abs": "Electroencephalography (EEG)-based brain-computer interfaces (BCIs) serve many control paradigms by relying on a variety in active brain regions and EEG features. Developing a universal EEG foundation model has been challenging due to the large variety in recording setups and experimental tasks. Additionally, researchers often contend with limited labeled data, making it difficult to utilize large deep-learning models effectively. While there have been successful attempts to develop EEG foundation models, few studies have systematically evaluated their adaptability across diverse BCI control paradigms. To address this gap, we propose a novel, yet simple spatiotemporal EEG transformer (ST-EEGFormer) that projects segments (\u201cpatches\u201d) of raw EEG data into an embedding space enriched with a spatial and temporal embedding, allowing the model to effectively handle EEG data exhibiting various channel set-ups and time lengths. To improve data efficiency, we first employed a masked autoencoder (MAE) task to pre-train the ST-EEGFormer in a self-supervised learning manner on a dataset combining six different motor imagery (MI) datasets, a P300 dataset, and a steady-state visual evoked potential (SSVEP) dataset, all of which are public. Next, we benchmarked the pre-trained model, after fine-tuning, on diverse downstream classification tasks. To evaluate the generalization capability, we conducted additional experiments on two public datasets, not used for pre-training: a seizure classification dataset and an online MI BCI dataset. We compared the performance against a simple linear model, EEGNet (a classic CNN-based benchmark model), the state-of-the-art supervised EEG Conformer model, and two foundation models, BIOT and Large Brain Model (LaBraM). The pre-trained ST-EEGFormers could learn robust EEG representations, achieving higher classification accuracies than the benchmarked models across all eight pre-training datasets and exhibiting strong generalization on new datasets with limited training data. Finally, we report several visualizations of the model including the features on which the results are based.", "keywords": [ "Brain-computer interfaces", "EEG representations", "foundation model", "masked autoencoder", "spatial-temporal transformer" ], "rating_list": [ 6, 6, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 4, 3, 5 ] }, { "paper_id": "V5Y7HdPXEA", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Lihan Wang", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "MS student" }, { "name": "Qikui Zhu", "gender": "Male", "institution": "Case Western Reserve University", "country": "US", "position": "Postdoc" }, { "name": "Tianxu Jiang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "MS student" }, { "name": "Yuting Shao", "gender": "Female", "institution": "Case Western Reserve University", "country": "", "position": "PhD student" }, { "name": "Ziyuan Qin", "gender": "Not Specified", "institution": "Case Western Reserve University", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 38, "n_ref": 63, "n_ref_all": 70, "n_fig": 2, "n_tab": 3, "L_tab": 723, "n_element_tab": 67, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1549, "n_element_tab_1": 118, "formula_len_all": 1080, "formula_len_all_1": 968, "len_all": 117895, "len_all_1": 66216, "len_abs": 1641, "len_title": 145, "len_sents": 34280, "len_sents_1": 30369, "n_sents": 248, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1890, "title": "Unified Uncertain Dual-prompts cross-domain Segmentation framework for medical image segmentation", "abs": "Unsupervised cross-domain segmentation addresses the challenge of label dependence in cross-domain medical image segmentation. Yet, most existing methods treat domain adaptation and segmentation as \\textbf{\\textit{Two Separate Steps}} and primarily focus on global domain adaptation, lacking the ability to prioritize segmentation-specific information during domain adaptation. Additionally, extracting domain-invariant feature representation remains an unavoidable challenge for cross-domain segmentation. These challenges significantly reduce segmentation performance. To this end, we propose a novel Unified Uncertain Dual-prompts cross-domain Segmentation framework (UUDS) for unsupervised cross-domain medical image segmentation. Specifically, our UUDS forms a unified framework by integrating domain adaptation and segmentation models, facilitating interaction between the two tasks, and addressing the challenge of emphasizing segmentation semantics while domain adaptation. Additionally, UUDS creatively uses dual-prompts, domain and segmentation prompts, to learn domain-invariant feature representation, ensuring that model can learn domain-invariant feature representation from cross-domain space. Furthermore, to facilitate interaction between the two tasks, UUDS uses uncertainty estimation to dynamically compute the label of segmentation for directly supervising the cross-domain adaptation, making the semantic information from unlabeled target images can directly supervise the process of domain adaptation and keeping the model sensitive to segmentation. Extensive experimental results on two public unsupervised cross-modality medical image segmentation demonstrate that UUDS outperforms state-of-the-art methods in unsupervised cross-modality medical image segmentation, highlighting its effectiveness in addressing domain shifts and marking a significant breakthrough.", "keywords": [ "Unsupervised Medical Image Segmentation;" ], "rating_list": [ 3, 1, 3 ], "soundness_list": [ 3, 1, 3 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 3, 1, 3 ], "confidence_list": [ 5, 5, 3 ] }, { "paper_id": "V4y0CpX4hK", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Chenlu Zhan", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Hanrong Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Hongwei Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Jingyuan Huang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Kai Mei", "gender": "unknown", "institution": "Rutgers University, New Brunswick", "country": "", "position": "PhD student" }, { "name": "Yifei Yao", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yongfeng Zhang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Associate Professor" }, { "name": "Zhenting Wang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" } ], "n_formula": 32, "n_formula_1": 15, "n_ref_uni": 37, "n_ref": 82, "n_ref_all": 101, "n_fig": 4, "n_tab": 17, "L_tab": 9770, "n_element_tab": 738, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2575, "n_element_tab_1": 73, "formula_len_all": 2247, "formula_len_all_1": 815, "len_all": 312000, "len_all_1": 70861, "len_abs": 1376, "len_title": 145, "len_sents": 101436, "len_sents_1": 29919, "n_sents": 834, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1381, "title": "Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents", "abs": "Although LLM-based agents, powered by Large Language Models (LLMs), can use external tools and memory mechanisms to solve complex real-world tasks, they may also introduce critical security vulnerabilities. However, the existing literature does not comprehensively evaluate attacks and defenses against LLM-based agents. To address this, we introduce Agent Security Bench (ASB), a comprehensive framework designed to formalize, benchmark, and evaluate the attacks and defenses of LLM-based agents, including 10 scenarios (e.g., e-commerce, autonomous driving, finance), 10 agents targeting the scenarios, over 400 tools, 27 different types of attack/defense methods, and 7 evaluation metrics. Based on ASB, we benchmark 10 prompt injection attacks, a memory poisoning attack, a novel Plan-of-Thought backdoor attack, 4 mixed attacks, and 11 corresponding defenses across 13 LLM backbones with over 90,000 testing cases in total. Our benchmark results reveal critical vulnerabilities in different stages of agent operation, including system prompt, user prompt handling, tool usage, and memory retrieval, with the highest average attack success rate of 84.30\\%, but limited effectiveness shown in current defenses, unveiling important works to be done in terms of agent security for the community. Our code can be found at \nhttps://anonymous.4open.science/r/AgentSecurityBench-A757.", "keywords": [ "AI agents; Large Language Model; Benchmark; Prompt Injection Attacks; Backdoor Attack; Defenses" ], "rating_list": [ 8, 6, 3, 8 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 1, 4 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "V4Xs283LHH", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Xuyang Shen", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Yifan Zhang", "gender": "Not Specified", "institution": "University of California, Los Angeles", "country": "US", "position": "Researcher" }, { "name": "Yiran Zhong", "gender": "Male", "institution": "MiniMax", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhen Qin", "gender": "unknown", "institution": "TapTap", "country": "CN", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 22, "n_ref": 31, "n_ref_all": 42, "n_fig": 4, "n_tab": 6, "L_tab": 5340, "n_element_tab": 414, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3640, "n_element_tab_1": 335, "formula_len_all": 2425, "formula_len_all_1": 1772, "len_all": 93628, "len_all_1": 52913, "len_abs": 1219, "len_title": 125, "len_sents": 20799, "len_sents_1": 19330, "n_sents": 177, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1225, "title": "FlashSampling: Fast and Memory-Efficient Exact Sampling with Group-Gumbel-Max", "abs": "Sampling operations in discrete space are widely used in different fields such as language models, reinforcement learning, VAE, GAN, and neural architecture search. Current sampling methods involve computing the softmax operation across the entire categories, leading to significant computational and memory requirements, particularly when dealing with large sampling categories. This paper presents a novel sampling approach known as FlashSampling, designed to alleviate the computational and communication overhead by circumventing the computation of the softmax operation. Our method maintains mathematical equivalence to conventional sampling strategies while demonstrating significantly enhanced speed and memory efficiency. This is achieved by partitioning the category into distinct groups for independent sampling and then leveraging the Gumble-Max trick to eliminate the need for softmax computation. We substantiate the correctness and efficacy of our method both through mathematical proofs and empirical validation. Extensive experimental outcomes illustrate marked enhancements in speed and memory utilization, with FlashSampling attaining up to 384\\% faster sampling times and 1822\\% reduced memory consumption.", "keywords": [ "Fast sampling" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 3, 1, 1, 2 ], "presentation_list": [ 2, 1, 1, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "V4K9h1qNxE", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Joao Sacramento", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Razvan Pascanu", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Affiliate Member" }, { "name": "Seijin Kobayashi", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Simon Schug", "gender": "unknown", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" }, { "name": "Yassir Akram", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 54, "n_ref": 86, "n_ref_all": 127, "n_fig": 23, "n_tab": 1, "L_tab": 186, "n_element_tab": 14, "n_fig_1": 20, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 865, "formula_len_all_1": 696, "len_all": 209423, "len_all_1": 62817, "len_abs": 1343, "len_title": 75, "len_sents": 47286, "len_sents_1": 31427, "n_sents": 328, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 27, "L_abs": 1353, "title": "Attention as a Hypernetwork", "abs": "Transformers can under some circumstances generalize to novel problem instances whose constituent parts might have been encountered during training, but whose compositions have not.\nWhat mechanisms underlie this ability for compositional generalization?\nBy reformulating multi-head attention as a hypernetwork, we reveal that a composable, low-dimensional latent code specifies key-query specific operations.\nWe find empirically that this latent code is predictive of the subtasks the network performs on unseen task compositions, revealing that latent codes acquired during training are reused to solve unseen problem instances.\nTo further examine the hypothesis that the intrinsic hypernetwork of multi-head attention supports compositional generalization, we ablate whether making the hypernetwork-generated linear value network nonlinear strengthens compositionality.\nWe find that this modification improves compositional generalization on abstract reasoning tasks.\nIn particular, we introduce a symbolic version of the Raven's Progressive Matrices human intelligence test, which gives us precise control over the problem compositions encountered during training and evaluation.\nWe demonstrate on this task how scaling model size and data enables compositional generalization in transformers and gives rise to a functionally structured latent space.", "keywords": [ "attention", "compositional generalization", "abstract reasoning", "in-context learning", "transformer", "mechanistic interpretability" ], "rating_list": [ 8, 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 4, 2, 2, 4, 3 ] }, { "paper_id": "V3zobHnS61", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Doyoung Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Researcher" }, { "name": "Hwanjun Song", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Jae-Gil Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Jihwan Bang", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "KR", "position": "Researcher" }, { "name": "Junhyeok Kang", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Researcher" }, { "name": "Youngjun Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 8, "n_ref_uni": 58, "n_ref": 142, "n_ref_all": 189, "n_fig": 9, "n_tab": 13, "L_tab": 5062, "n_element_tab": 460, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1896, "n_element_tab_1": 22, "formula_len_all": 1002, "formula_len_all_1": 495, "len_all": 206701, "len_all_1": 65376, "len_abs": 1153, "len_title": 123, "len_sents": 56566, "len_sents_1": 28903, "n_sents": 441, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1041, "title": "RA-TTA: Retrieval-Augmented Test-Time Adaptation for Vision-Language Models", "abs": "Vision-language models (VLMs) are known to be susceptible to distribution shifts between pre-training data and test data, and test-time adaptation (TTA) methods for VLMs have been proposed to mitigate the detrimental impact of the distribution shifts. However, the existing methods solely rely on the internal knowledge encoded within the parameters, which are constrained to pre-training data. To complement the limitation of the internal knowledge, we propose **retrieval-augmented-TTA (RA-TTA)** for adapting VLMs to test distribution using **external** knowledge obtained from a web-scale image database. Fully exploiting the bi-modality of VLMs, fine-grained **text descriptions** are used both for retrieving proper external images and refining VLMs' predictions with the retrieved external images. As a result, the pivotal features of test images are more precisely recognized through the text descriptions. \nExtensive evaluations on 17 datasets validate that RA-TTA outperforms the state-of-the-art methods by 2.49-8.45\\% on average.", "keywords": [ "vision-language models", "test-time adaptation", "retrieval-augmented strategy" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "V2x5ZTHMae", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Changyou Chen", "gender": "Male", "institution": "State University of New York, Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Huaisheng Zhu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Jiayi Chen", "gender": "Female", "institution": "Fujian Normal University", "country": "CN", "position": "Undergrad student" }, { "name": "Kaiyi Ji", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Rohan Sharma", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Shijie Zhou", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "RUIYI ZHANG", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" } ], "n_formula": 22, "n_formula_1": 20, "n_ref_uni": 35, "n_ref": 63, "n_ref_all": 90, "n_fig": 8, "n_tab": 5, "L_tab": 1450, "n_element_tab": 131, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 189, "n_element_tab_1": 65, "formula_len_all": 2880, "formula_len_all_1": 2226, "len_all": 118829, "len_all_1": 58232, "len_abs": 1485, "len_title": 143, "len_sents": 35886, "len_sents_1": 27539, "n_sents": 216, "n_sents_1": 158, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1490, "title": "Enhancing Diffusion Posterior Sampling for Inverse Problems by Integrating Crafted Measurements", "abs": "Diffusion models have emerged as a powerful foundation model for visual generation. With an appropriate sampling process, it can effectively serve as a generative prior to solve general inverse problems. Current posterior sampling based methods take the measurement (i.e., degraded image sample) into the posterior sampling to infer the distribution of the target data (i.e., clean image sample). However, in this manner, we show that high-frequency information can be prematurely introduced during the early stages, which could induce larger posterior estimate errors during the restoration sampling. To address this issue, we first reveal that forming the log posterior gradient with the noisy measurement ( i.e., samples from a diffusion forward process) instead of the clean one can benefit the reverse process. Consequently, we propose a novel diffusion posterior sampling method DPS-CM, which incorporates a Crafted Measurement (i.e., samples generated by a reverse denoising process, compared to random sampling with noise in standard methods) to form the posterior estimate. This integration aims to mitigate the misalignment with the diffusion prior caused by cumulative posterior estimate errors. Experimental results demonstrate that our approach significantly improves the overall capacity to solve general and noisy inverse problems, such as Gaussian deblurring, super-resolution, inpainting, nonlinear deblurring, and tasks with Poisson noise, relative to existing approaches.", "keywords": [ "Diffusion models", "Inverse problems", "Diffusion posterior sampling" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 1, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "V1N6MmDY27", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Feng Chen", "gender": "Male", "institution": "University of Texas, Dallas", "country": "", "position": "Associate Professor" }, { "name": "Gopal Gupta", "gender": "Male", "institution": "University of Texas at Dallas", "country": "", "position": "Full Professor" }, { "name": "Keegan Kimbrell", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Tianhao Wang", "gender": "Male", "institution": "University of Texas at Dallas", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 19, "n_ref": 24, "n_ref_all": 32, "n_fig": 1, "n_tab": 5, "L_tab": 1606, "n_element_tab": 213, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1549, "n_element_tab_1": 211, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 98706, "len_all_1": 72073, "len_abs": 1294, "len_title": 118, "len_sents": 41149, "len_sents_1": 35786, "n_sents": 296, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1278, "title": "Towards Fully Autonomous Driving with Automated Commonsense Reasoning", "abs": "Autonomous Vehicle (AV) technology has been heavily researched and sought after, yet there are no SAE Level 5 AVs available today in the marketplace. We contend that over-reliance on machine learning technology is the main reason. Use of automated commonsense reasoning technology, we believe, can help achieve SAE Level 5 autonomy. In this paper, we show how automated commonsense reasoning technology can be deployed in situations where not enough data is available to train a machine learning model for autonomous driving. Specifically, we consider two situations where (i) a traffic signal is malfunctioning at an intersection and (ii) all the cars ahead are slowing down and steering away due to an unexpected obstruction (e.g., animals on the road). We show that in such situations, our commonsense reasoning based solution performs correctly. We also provide a pathway for efficiently invoking commonsense reasoning by measuring uncertainty in the computer vision model and using commonsense reasoning to handle uncertain scenarios. We describe our experiments conducted using the CARLA simulator and the results obtained. The main contribution of our research is to show that automated commonsense reasoning provides an effective pathway to reach SAE level 5 automation.", "keywords": [ "Commonsense Reasoning", "Autonomous Vehicles", "Uncertainty" ], "rating_list": [ 1, 3, 5, 1 ], "soundness_list": [ 1, 1, 3, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 1, 1, 3, 1 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "V1MDIFbqCp", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Heyue Li", "gender": "unknown", "institution": "Wuhan University", "country": "CN", "position": "Undergrad student" }, { "name": "Lincoln Ye", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Qinying Gu", "gender": "Female", "institution": "Shanghai artificial intelligence laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xinbing Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Yifei Wu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 25, "n_formula_1": 29, "n_ref_uni": 35, "n_ref": 66, "n_ref_all": 76, "n_fig": 2, "n_tab": 3, "L_tab": 949, "n_element_tab": 45, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1608, "n_element_tab_1": 84, "formula_len_all": 1606, "formula_len_all_1": 1733, "len_all": 116199, "len_all_1": 64909, "len_abs": 1048, "len_title": 121, "len_sents": 30419, "len_sents_1": 27468, "n_sents": 223, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1051, "title": "Probabilistic Feature Smoothed Gaussian Process For Imbalanced Regression", "abs": "Gaussian Processes (GPs) are non-parametric Bayesian models widely used for regression, classification, and other tasks due to their explainability and versatility. However, GPs face challenges in imbalanced regression, where the skewed distribution of target labels can greatly harm models' performances. In this work, we introduce the Probabilistic Feature Smoothed Partially Independent Training Conditional Approximation (PFS-PITC) to enhance GP performance in imbalanced scenarios. We extract statistical features from the observation space using equidistant label intervals and apply kernel smoothing to address sampling density discontinuities. This process enables PFS-PITC to utilize information from nearby labels within imbalanced datasets, thereby reducing GPs' sensitivity to such imbalances. Empirical tests on various imbalanced regression datasets demonstrate the effectiveness of PFS-PITC, contributing to the robustness of GPs in handling flawed real-world data and expanding their applicability in challenging data processing tasks.", "keywords": [ "Imbalanced Learning", "Gaussian Process", "Bayesian Methods", "Machine Learning" ], "rating_list": [ 3, 3, 1, 5 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "V0Hyw9Tz5W", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Guodong Zheng", "gender": "unknown", "institution": "Fuzhou University", "country": "CN", "position": "MS student" }, { "name": "Yongli Cheng", "gender": "Male", "institution": "Fuzhou University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 19, "n_formula_1": 18, "n_ref_uni": 36, "n_ref": 74, "n_ref_all": 92, "n_fig": 8, "n_tab": 3, "L_tab": 1309, "n_element_tab": 85, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1339, "n_element_tab_1": 87, "formula_len_all": 1082, "formula_len_all_1": 1051, "len_all": 100349, "len_all_1": 57186, "len_abs": 1606, "len_title": 94, "len_sents": 27067, "len_sents_1": 22764, "n_sents": 223, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 45, "L_abs": 1655, "title": "Positive Mining in Graph Contrastive Learning", "abs": "Graph Contrastive Learning (GCL), which aims to capture representations from unlabeled graphs, has made significant progress in recent years. In GCL, InfoNCE-based loss functions play a crucial role by ensuring that positive node pairs\u2014those that are similar\u2014are drawn closer together in the representational space, while negative pairs, which are dissimilar, are pushed apart. The primary focus of recent research has been on refining the contrastive loss function, particularly by adjusting the weighting of negative nodes. This is achieved by changing the weight between negative node pairs, or by using node similarity to select the positive node associated with the anchor node. Despite the substantial success of these GCL techniques, there remains a belief that the nodes identified as positive or negative may not accurately reflect the true positives and negatives. To tackle this challenge, we introduce an innovative method known as Positive Mining Graph Contrastive Learning (PMGCL). This method consists in calculating the probability of positive samples between the anchor node and other nodes using a mixture model, thereby identifying nodes that have a higher likelihood of being true positives in relation to the anchor node. We have conducted a comprehensive evaluation of PMGCL on a range of real-world graph datasets. The experimental findings indicate that PMGCL significantly outperforms traditional GCL methods. Our method not only achieves state-of-the-art results in unsupervised learning benchmarks but also exceeds the performance of supervised learning benchmarks in certain scenarios.", "keywords": [ "Graph Contrastive Learning", "Unsupervised representation learning", "Mixture model", "loss functions" ], "rating_list": [ 6, 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 3, 2 ], "presentation_list": [ 3, 1, 1, 4, 2 ], "contribution_list": [ 2, 1, 2, 3, 2 ], "confidence_list": [ 5, 5, 4, 4, 4 ] }, { "paper_id": "V0GlKhMLFl", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Liangyu Chen", "gender": "unknown", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" }, { "name": "Marzieh Fadaee", "gender": "Female", "institution": "Cohere For AI", "country": "NL", "position": "Researcher" }, { "name": "Sara Ahmadian", "gender": "Female", "institution": "Research, Google", "country": "", "position": "Research scientist" }, { "name": "Simon Chi Lok Yu", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 48, "n_ref": 76, "n_ref_all": 88, "n_fig": 4, "n_tab": 5, "L_tab": 1269, "n_element_tab": 266, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1969, "n_element_tab_1": 124, "formula_len_all": 998, "formula_len_all_1": 998, "len_all": 144815, "len_all_1": 59704, "len_abs": 1601, "len_title": 129, "len_sents": 33543, "len_sents_1": 26233, "n_sents": 253, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1614, "title": "Diversify and Conquer: Diversity-Centric Data Selection with Iterative Refinement", "abs": "Finetuning large language models on instruction data is an important step in enriching the knowledge learned during pre-training and improving instruction-following capabilities. \nAs the number of instruction datasets continues to grow, selecting the right data to achieve optimal results becomes increasingly important.\nIn this work, we ask a prominent question: How can we determine the optimal subset of data for effective training?\nWhile much of the existing research primarily emphasizes local criteria, such as instance quality, for subset selection, we argue that a global approach focused on data diversity is more critical.\nOur approach utilizes $k$-means clustering to ensure that the selected subset effectively represents the full dataset.\nWe propose an iterative refinement method inspired by active learning techniques to resample instances from clusters, with the importance and sampling weight of each cluster being reassessed in every training iteration.\nThis method allows us to reduce the effect of outliers and automatically filter out clusters containing low-quality data.\nThrough extensive evaluation across natural language reasoning, general world knowledge, code and math reasoning tasks, and by fine-tuning models from various families, we observe consistent improvements, achieving a 7\\% increase over the random selection and a 3.8\\% improvement over state-of-the-art sampling methods.\nOur work highlights the significance of diversity-first sampling when finetuning LLMs to enhance performance across a broad array of evaluation tasks. Our code is submitted as supplementary materials.", "keywords": [ "diversity", "data selection", "training efficiency", "iterative refinement" ], "rating_list": [ 3, 3, 5, 1 ], "soundness_list": [ 3, 3, 3, 1 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "UzpMjtBbit", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 12, "author_info_list": [ { "name": "Gang Chen", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Guojie Luo", "gender": "Male", "institution": "Peking University", "country": "", "position": "Associate Professor" }, { "name": "Kexin Chen", "gender": "Female", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Lingwei Yan", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Qiang Xu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Ruiyang Ma", "gender": "unknown", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yi Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yipeng Wang", "gender": "unknown", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Yunhao Zhou", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "MS student" }, { "name": "Zhengyuan Shi", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Zhiqiang He", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Ziyang Zheng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 24, "n_ref": 40, "n_ref_all": 58, "n_fig": 6, "n_tab": 5, "L_tab": 1838, "n_element_tab": 193, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1524, "n_element_tab_1": 142, "formula_len_all": 211, "formula_len_all_1": 211, "len_all": 106252, "len_all_1": 68267, "len_abs": 1166, "len_title": 68, "len_sents": 35257, "len_sents_1": 31107, "n_sents": 305, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1169, "title": "DynamicRTL: RTL Representation Learning for Dynamic Circuit Behavior", "abs": "There is a growing body of work on using Graph Neural Networks (GNNs) to learn representations of circuits, focusing primarily on their static characteristics. However, these models fail to capture critical runtime behavior, which is crucial for tasks like hardware verification and optimization. To address this limitation, we introduce DynamicRTL, a novel GNN-based approach that learns circuit representations by incorporating both static structures and multi-cycle execution behaviors. DynamicRTL leverages an operation-level Control Data Flow Graph (CDFG) to represent Register Transfer Level (RTL) circuits, enabling the model to capture dynamic dependencies and runtime execution. To train and evaluate DynamicRTL, we built the first comprehensive dynamic circuit dataset, comprising over 6,300 Verilog modules and 190,000 simulation traces. Our results demonstrate that DynamicRTL consistently outperforms existing models in branch prediction tasks. Furthermore, its learned representations transfer effectively to related tasks, achieving strong performance in assertion prediction and underscoring its transfer learning capabilities for dynamic circuit tasks.", "keywords": [ "Code Representation Learning", "Graph Neural Network", "Hardware Design" ], "rating_list": [ 6, 3, 5, 6, 5 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "UzgMX1rwGc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jogendra Nath Kundu", "gender": "Male", "institution": "Meta", "country": "US", "position": "Research Scientist " }, { "name": "Rishubh Parihar", "gender": "Male", "institution": "Indian Institute of Science, Bangalore", "country": "IN", "position": "PhD student" }, { "name": "Sarthak Vora", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "MS student" }, { "name": "Srinjay Sarkar", "gender": "Male", "institution": "Indian Institute of Science ", "country": "IN", "position": "Researcher" }, { "name": "Venkatesh Babu R", "gender": "Male", "institution": "Indian Institute of Science", "country": "IN", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 17, "n_ref": 42, "n_ref_all": 82, "n_fig": 17, "n_tab": 13, "L_tab": 3422, "n_element_tab": 202, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 2808, "n_element_tab_1": 64, "formula_len_all": 463, "formula_len_all_1": 175, "len_all": 142932, "len_all_1": 55246, "len_abs": 1350, "len_title": 129, "len_sents": 39911, "len_sents_1": 24049, "n_sents": 337, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1358, "title": "Smart Placement Enhanced Vision: Enhancing 3D-Detection With Learned 3D Placement", "abs": "The diversity and scale of annotated real-world 3D datasets limit the performance of monocular 3D detectors. Although data augmentation holds potential, creating realistic, scene-aware augmentations for outdoor environments presents a significant challenge.\nExisting augmentation methods majorly focus on realistic object appearance by advancing the rendering quality. However, we show that object placement is equally important for downstream 3D detection performance. The main challenge, however, for realistic placement, is to automatically identify the plausible physical properties (location, scale, and orientation) for placing objects in real-world scenes. To this end, we propose Smart-Placement, a novel 3D scene-aware augmentation method for generating diverse and realistic augmentations. In particular, given a background scene, we train a placement network to learn a distribution over plausible 3D bounding boxes. Subsequently, we render realistic cars from 3D assets and place them according to the locations sampled from the learned distribution. Through extensive empirical evaluation on standard benchmark datasets - KITTI and NuScenes, we show that our proposed augmentation method significantly boosts the performance of several existing monocular 3D detectors, setting a new state-of-the-art benchmark, while being highly data efficient.", "keywords": [ "Synthetic data; 3D object detection; Data augmentation" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "UyhRtB4hjN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Mihaela van der Schaar", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Nicolas Huynh", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Tennison Liu", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 4, "n_ref_uni": 75, "n_ref": 145, "n_ref_all": 198, "n_fig": 22, "n_tab": 20, "L_tab": 15941, "n_element_tab": 867, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1973, "n_element_tab_1": 284, "formula_len_all": 1015, "formula_len_all_1": 457, "len_all": 289033, "len_all_1": 86735, "len_abs": 2313, "len_title": 104, "len_sents": 84195, "len_sents_1": 37074, "n_sents": 700, "n_sents_1": 267, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1783, "title": "Decision Tree Induction via Semantically-Aware Evolution", "abs": "Decision trees are a crucial class of models offering robust predictive performance and inherent interpretability across various domains, including healthcare, finance, and logistics. However, current tree induction methods often face limitations such as suboptimal solutions from greedy methods or prohibitive computational costs and limited applicability of exact optimization approaches.\nTo address these challenges, we propose an evolutionary optimization method for decision tree induction based on genetic programming (GP). Our key innovation is the integration of semantic priors and domain-specific knowledge about the search space into the optimization algorithm. To this end, we introduce $\\texttt{LLEGO}$, a framework that incorporates semantic priors into genetic search operators through the use of Large Language Models (LLMs), thereby enhancing search efficiency and targeting regions of the search space that yield decision trees with superior generalization performance. This is operationalized through novel genetic operators that work with structured natural language prompts, effectively utilizing LLMs as conditional generative models and sources of semantic knowledge. Specifically, we introduce _fitness-guided_ crossover to exploit high-performing regions, and _diversity-guided_ mutation for efficient global exploration of the search space. These operators are controlled by corresponding hyperparameters that enable a more nuanced balance between exploration and exploitation across the search space. Empirically, we demonstrate across various benchmarks that $\\texttt{LLEGO}$ evolves superior-performing trees compared to existing tree induction methods, and exhibits significantly more efficient search performance compared to conventional GP approaches.", "keywords": [ "decision trees", "LLMs", "genetic programming" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "UyU8ETswPg", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Shiyu Chang", "gender": "Unspecified", "institution": "University of California, Santa Barbara", "country": "US", "position": "Assistant Professor" }, { "name": "Tommi Jaakkola", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Yang Zhang", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Research Staff Employee" }, { "name": "Yujian Liu", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 47, "n_ref": 88, "n_ref_all": 126, "n_fig": 12, "n_tab": 9, "L_tab": 1883, "n_element_tab": 214, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 680, "n_element_tab_1": 94, "formula_len_all": 315, "formula_len_all_1": 235, "len_all": 184099, "len_all_1": 70997, "len_abs": 1078, "len_title": 126, "len_sents": 54451, "len_sents_1": 34113, "n_sents": 433, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1082, "title": "Fictitious Synthetic Data Can Improve LLM Factuality via Prerequisite Learning", "abs": "Recent studies have identified one aggravating factor of LLM hallucinations as the knowledge inconsistency between pre-training and fine-tuning, where unfamiliar fine-tuning data mislead the LLM to fabricate plausible but wrong outputs. In this paper, we propose a novel fine-tuning strategy called Prereq-Tune to address this knowledge inconsistency and reduce hallucinations. Fundamentally, Prereq-Tune disentangles the learning of skills and knowledge, so the model learns only the task skills without being impacted by the knowledge inconsistency. To achieve this, Prereq-Tune introduces an additional prerequisite learning stage to learn the necessary knowledge for SFT, allowing subsequent SFT to focus only on task skills. Prereq-Tune can also be combined with fictitious synthetic data to enhance the grounding of LLM outputs to their internal knowledge. Experiments show that Prereq-Tune outperforms existing baselines in improving LLM's factuality across short QA and long-form generation tasks. It also opens new possibilities for knowledge-controlled generation in LLMs.", "keywords": [ "Hallucinations", "instruction-tuning" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 4, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "UyBMzsFThf", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Devin Willmott", "gender": "Male", "institution": "Bosch Center for Artificial Intelligence", "country": "US", "position": "Research Scientist" }, { "name": "Dylan Sam", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Jo\u00e3o D. Semedo", "gender": "unknown", "institution": "Bosch Center for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Zico Kolter", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 56, "n_ref": 112, "n_ref_all": 142, "n_fig": 6, "n_tab": 13, "L_tab": 1924, "n_element_tab": 230, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1045, "n_element_tab_1": 100, "formula_len_all": 318, "formula_len_all_1": 183, "len_all": 209574, "len_all_1": 70669, "len_abs": 1598, "len_title": 100, "len_sents": 63262, "len_sents_1": 34952, "n_sents": 413, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1603, "title": "Finetuning CLIP to Reason about Pairwise Differences", "abs": "Vision-language models (VLMs) such as CLIP are trained via contrastive learning between text and image pairs, resulting in aligned image and text embeddings that are useful for many downstream tasks. A notable drawback of CLIP, however, is that the resulting embedding space seems to lack some of the structure of their purely text-based alternatives. For instance, while text embeddings have been long noted to satisfy analogies in embedding space using vector arithmetic, CLIP has no such property. In this paper, we propose an approach to natively train CLIP in a contrastive manner to reason about differences in embedding space. We finetune CLIP so that the differences in image embedding space correspond to text descriptions of the image differences, which we synthetically generate with large language models on image-caption paired datasets. We first demonstrate that our approach yields significantly improved capabilities in ranking images by a certain attribute (e.g., elephants are larger than cats), which is useful in retrieval or constructing attribute-based classifiers, and improved zeroshot classification performance on many downstream image classification tasks. In addition, our approach enables a new mechanism for inference that we refer to as comparative prompting, where we leverage prior knowledge of text descriptions of differences between classes of interest, achieving even larger performance gains in classification. Finally, we illustrate that the resulting embeddings obey a larger degree of geometric properties in embedding space, such as in text-to-image generation.", "keywords": [ "CLIP", "finetuning", "vision language models" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "UxzKcIZedp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Boyuan Li", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Chuhao Jin", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Ruihua Song", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Wenbing Huang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Wenhui Tan", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Xiting Wang", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 23, "n_ref": 69, "n_ref_all": 85, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 1133, "n_element_tab_1": 118, "formula_len_all": 432, "formula_len_all_1": 443, "len_all": 116226, "len_all_1": 66618, "len_abs": 1371, "len_title": 183, "len_sents": 40292, "len_sents_1": 31153, "n_sents": 290, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1453, "title": "Think Then React: Towards Unconstrained Action-to-Reaction Motion Generation", "abs": "Modeling human-like action-to-reaction generation has significant real-world applications, like human-robot interaction and games.\nDespite recent advancements in single-person motion generation, it is still challenging to well handle action-to-reaction generation, due to the difficulty of directly predicting reaction from action sequence without prompts, and the absence of a unified representation that effectively encodes multi-person motion.\nTo address these challenges, we introduce Think-Then-React (TTR), a large language-model-based framework designed to generate human-like reactions.\nFirst, with our fine-grained multimodal training strategy, TTR is capable to unify two processes during inference: a thinking process that explicitly infers action intentions and reasons corresponding reaction description, which serve as semantic prompts, and a reacting process that predicts reactions based on input action and the inferred semantic prompts.\nSecond, to effectively represent multi-person motion in language models, we propose a unified motion tokenizer by decoupling egocentric pose and absolute space features, which effectively represents action and reaction motion with same encoding.\nExtensive experiments demonstrate that TTR outperforms existing baselines, achieving significant improvements in evaluation metrics, such as reducing FID from 3.988 to 1.942. Source code is available at https://github.com/AlbertTan404/Think-Then-React.", "keywords": [ "Human Reaction Generation", "3D Human Motion", "Large Language Model" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Uxm7DxPwrZ", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Alexi Canesse", "gender": "Male", "institution": "Ecole Normale Sup\u00e9rieure de Lyon", "country": "FR", "position": "MS student" }, { "name": "Ludovic Denoyer", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Mathieu Petitbois", "gender": "Male", "institution": "Ubisoft", "country": "FR", "position": "PhD student" }, { "name": "R\u00e9my Portelas", "gender": "Male", "institution": "Ubisoft", "country": "FR", "position": "Researcher" }, { "name": "sylvain lamprier", "gender": "unknown", "institution": "Universit\u00e9 d'Angers", "country": "", "position": "Full Professor" } ], "n_formula": 21, "n_formula_1": 12, "n_ref_uni": 54, "n_ref": 92, "n_ref_all": 131, "n_fig": 18, "n_tab": 7, "L_tab": 2190, "n_element_tab": 180, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2666, "formula_len_all_1": 988, "len_all": 197872, "len_all_1": 65275, "len_abs": 1132, "len_title": 126, "len_sents": 63362, "len_sents_1": 31488, "n_sents": 449, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1115, "title": "Navigation with QPHIL: Offline Goal-Conditioned RL in a Learned Discretized Space", "abs": "Offline Reinforcement Learning (RL) has emerged as a powerful alternative to imitation learning for behavior modeling in various domains, particularly in complex navigation tasks. An existing challenge with Offline RL is the signal-to-noise ratio, i.e. how to mitigate incorrect policy updates due to errors in value estimates. Towards this, multiple works have demonstrated the advantage of hierarchical offline RL methods, which decouples high-level path planning from low-level path following. In this work, we present a novel hierarchical transformer-based approach leveraging a learned quantizer of space. This quantization enables the training of a zone-conditioned low-level policy and simplifies planning, which is reduced to discrete autoregressive prediction. Among other benefits, zone-level reasoning in planning enables explicit trajectory stitching rather than implicit stitching based on noisy value function estimates. By combining this transformer-based planner with recent advancements in offline RL, our approach achieves state-of-the-art results in complex long-distance navigation environments.", "keywords": [ "Reinforcement Learning", "Offline Reinforcement Learning", "Goal-Conditioned Reinforcement Learning" ], "rating_list": [ 5, 3, 6, 5, 5 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "UxkznlcnHf", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zeyu Gan", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 39, "n_formula_1": 5, "n_ref_uni": 42, "n_ref": 54, "n_ref_all": 75, "n_fig": 5, "n_tab": 4, "L_tab": 168, "n_element_tab": 10, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2597, "formula_len_all_1": 332, "len_all": 173851, "len_all_1": 56017, "len_abs": 1309, "len_title": 157, "len_sents": 63164, "len_sents_1": 28538, "n_sents": 466, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 108, "L_abs": 1319, "title": "Towards a Theoretical Understanding of Synthetic Data in LLM Post-Training: A Reverse-Bottleneck Perspective", "abs": "Synthetic data has become a pivotal resource in post-training tasks for large language models (LLMs) due to the scarcity of high-quality, specific data. While various methods have been developed to generate synthetic data, there remains a discernible gap between the practical effects of synthetic data and our theoretical comprehension. To address this challenge, we commence by presenting a detailed modeling of the prevalent synthetic data generation process. Building upon this modeling, we demonstrate that the generalization capability of the post-trained model is critically determined by the information gain derived from the generative model, as analyzed from a novel reverse-bottleneck perspective. Moreover, we introduce the concept of Generalization Gain via Mutual Information (GGMI) and elucidate the relationship between generalization gain and information gain. This analysis serves as a theoretical foundation for synthetic data generation and further highlights its connection with the generalization capability of post-trained models, offering an understanding about the design of synthetic data generation techniques and the optimization of the post-training process. We open-source our code at https://github.com/ZyGan1999/Towards-a-Theoretical-Understanding-of-Synthetic-Data-in-LLM-Post-Training.", "keywords": [ "large language models; synthetic data; information bottleneck" ], "rating_list": [ 6, 8, 8, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 4, 4, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "Ux0BEP46fd", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Handong Li", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Jing Liu", "gender": "Female", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "Full Professor" }, { "name": "Xiangyu Yue", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Yiyuan Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 70, "n_ref": 129, "n_ref_all": 145, "n_fig": 6, "n_tab": 6, "L_tab": 2076, "n_element_tab": 389, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1303, "n_element_tab_1": 281, "formula_len_all": 1091, "formula_len_all_1": 1033, "len_all": 203350, "len_all_1": 65119, "len_abs": 1094, "len_title": 167, "len_sents": 27730, "len_sents_1": 25213, "n_sents": 207, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 117, "L_abs": 999, "title": "Scaling Omni-modal Pretraining with Multimodal Context: Advancing Universal Representation Learning Across Modalities", "abs": "In this work, we introduce Multimodal Context (MiCo), a scalable pretraining framework designed to advance omni-modal intelligence\u2014an AI system capable of understanding and learning from multiple modalities to achieve universal representation learning. MiCo allows for efficient scaling of both the number of modalities and the volume of data, along with model parameters, during the pretraining phase. We evaluate the pretrained models across a diverse set of tasks, including: (i) single-modality perception benchmarks covering 10 distinct modalities, (ii) 25 cross-modal tasks spanning retrieval, question-answering, and captioning, and (iii) 18 large-scale multimodal language model benchmarks. MiCo consistently delivers state-of-the-art results, setting 37 new benchmarks across these tasks. The pretrained models, along with the collected datasets and codebase, will be made publicly available to support the development of omni-modal intelligence and broader research in multimodal learning.", "keywords": [ "Multimodal Pretraining", "Multimodal Context" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 1, 3, 4 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "UwcZEoNP19", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Jie Chen", "gender": "Female", "institution": "Anhui University", "country": "CN", "position": "Associate Professor" }, { "name": "Minghan Zhang", "gender": "Male", "institution": "Anhui University", "country": "", "position": "PhD student" }, { "name": "Shu Zhao", "gender": "Female", "institution": "Anhui University", "country": "", "position": "Full Professor" }, { "name": "Wei Du", "gender": "unknown", "institution": "Anhui University", "country": "", "position": "MS student" }, { "name": "Zhen Duan", "gender": "unknown", "institution": "Anhui University", "country": "CN", "position": "Lecturer" }, { "name": "Zhen Yang", "gender": "Female", "institution": "Anhui University", "country": "CN", "position": "PhD student" }, { "name": "Ziwei Du", "gender": "unknown", "institution": "Anhui University", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 54, "n_ref": 137, "n_ref_all": 167, "n_fig": 5, "n_tab": 10, "L_tab": 1277, "n_element_tab": 108, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 1514, "n_element_tab_1": 132, "formula_len_all": 990, "formula_len_all_1": 1199, "len_all": 185636, "len_all_1": 68755, "len_abs": 1644, "len_title": 145, "len_sents": 45229, "len_sents_1": 30734, "n_sents": 368, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1614, "title": "Triples as the Key: Structuring Makes Decomposition and Verification Easier in LLM-based TableQA", "abs": "As the mainstream approach, LLMs have been widely applied and researched in TableQA tasks. Currently, the core of LLM-based TableQA methods typically include three phases: question decomposition, sub-question TableQA reasoning, and answer verification. However, several challenges remain in this process: i) Sub-questions generated by these methods often exhibit significant gaps with the original question due to critical information overlooked during the LLM's direct decomposition; ii) Verification of answers is typically challenging because LLMs tend to generate optimal responses during self-correct. To address these challenges, we propose a Triple-Inspired Decomposition and vErification (TIDE) strategy, which leverages the structural properties of triples to assist in decomposition and verification in TableQA. The inherent structure of triples (head entity, relation, tail entity) requires the LLM to extract as many entities and relations from the question as possible. Unlike direct decomposition methods that may overlook key information, our transformed sub-questions using triples encompass more critical details. Additionally, this explicit structure facilitates verification. By comparing the triples derived from the answers with those from the question decomposition, we can achieve easier and more straightforward validation than when relying on the LLM's self-correct tendencies. By employing triples alongside established LLM modes, Direct Prompting and Agent modes, TIDE achieves state-of-the-art performance across multiple TableQA datasets, demonstrating the effectiveness of our method.", "keywords": [ "TableQA", "Triples" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 2, 4, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "UwbX8KOZgK", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Abhinav Bhatele", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "Associate Professor" }, { "name": "Alireza Ganjdanesh", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Engineer" }, { "name": "Bethmage Mayuka Jayawardhana", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Gowthami Somepalli", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Heng Huang", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Kaiyu Yue", "gender": "Male", "institution": "University of Maryland", "country": "US", "position": "PhD student" }, { "name": "Reza Shirkavand", "gender": "Not Specified", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Sukriti Paul", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "MS student" }, { "name": "Tom Goldstein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Vasu Singla", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 58, "n_ref": 102, "n_ref_all": 128, "n_fig": 10, "n_tab": 8, "L_tab": 3074, "n_element_tab": 212, "n_fig_1": 10, "n_tab_1": 8, "L_tab_1": 1545, "n_element_tab_1": 108, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 161848, "len_all_1": 63733, "len_abs": 816, "len_title": 61, "len_sents": 37882, "len_sents_1": 27169, "n_sents": 339, "n_sents_1": 244, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 805, "title": "From Pixels to Prose: A Large Dataset of Dense Image Captions", "abs": "Training large vision-language models requires extensive, high-quality image-text pairs. Existing web-scraped datasets, however, are noisy and lack detailed image descriptions. To bridge this gap, we introduce PixelProse, a comprehensive dataset of over 16M (million) synthetically generated captions, leveraging cutting-edge vision-language models for detailed and accurate descriptions. To ensure data integrity, we rigorously analyze our dataset for problematic content, including child sexual abuse material (CSAM), personally identifiable information (PII), and toxicity. We also provide valuable metadata such as watermark presence and aesthetic scores, aiding in further dataset filtering. We hope PixelProse to serve as a valuable resource for future research involving vision-language modalities.", "keywords": [ "multimodal datasets", "image-text pairs", "large scale datasets" ], "rating_list": [ 5, 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 5, 4 ] }, { "paper_id": "UvpuGrd6ey", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Arthur Jacot", "gender": "Male", "institution": "NYU, New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Seok Hoan Choi", "gender": "Male", "institution": "New York University", "country": "US", "position": "MS student" }, { "name": "Yuxiao Wen", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 145, "n_formula_1": 37, "n_ref_uni": 49, "n_ref": 108, "n_ref_all": 117, "n_fig": 2, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 19678, "formula_len_all_1": 3173, "len_all": 222617, "len_all_1": 71473, "len_abs": 2743, "len_title": 82, "len_sents": 62951, "len_sents_1": 30919, "n_sents": 551, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1055, "title": "How DNNs break the Curse of Dimensionality: Compositionality and Symmetry Learning", "abs": "We show that deep neural networks (DNNs) can efficiently learn any\ncomposition of functions with bounded $F_{1}$-norm, which allows\nDNNs to break the curse of dimensionality in ways that shallow networks\ncannot. More specifically, we derive a generalization bound that combines\na covering number argument for compositionality, and the $F_{1}$-norm\n(or the related Barron norm) for large width adaptivity. We show that\nthe global minimizer of the regularized loss of DNNs can fit for example\nthe composition of two functions $f^*=h\\\\circ g $ from a small number\nof observations, assuming $g$ is smooth/regular and reduces the dimensionality\n(e.g. $g$ could be the quotient map of the symmetries of $f^*$),\nso that $h$ can be learned in spite of its low regularity. The measures\nof regularity we consider is the Sobolev norm with different levels\nof differentiability, which is well adapted to the $F_{1}$ norm.\nWe compute scaling laws empirically and observe phase transitions\ndepending on whether $g$ or $h$ is harder to learn, as predicted\nby our theory.", "keywords": [ "Generalization bound", "covering numbers", "compositionality" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 2 ] }, { "paper_id": "UvfI4grcM7", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Anan LI", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Dongli Hu", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jiandong Zhou", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Kai Du", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhu Tianfang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 43, "n_ref": 71, "n_ref_all": 91, "n_fig": 16, "n_tab": 2, "L_tab": 598, "n_element_tab": 35, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 598, "n_element_tab_1": 35, "formula_len_all": 531, "formula_len_all_1": 472, "len_all": 135950, "len_all_1": 59845, "len_abs": 2204, "len_title": 170, "len_sents": 37865, "len_sents_1": 29670, "n_sents": 249, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 112, "L_abs": 2211, "title": "Biologically Constrained Barrel Cortex Model Integrates Whisker Inputs and Replicates Key Brain Network Dynamics", "abs": "The brain's ability to transform sensory inputs into motor functions is central to neuroscience and crucial for the development of embodied intelligence. Sensory-motor integration involves complex neural circuits, diverse neuronal types, and intricate intercellular connections. Bridging the gap between biological realism and behavioral functionality presents a formidable challenge. In this study, we focus on the columnar structure of the superficial layers of mouse barrel cortex as a model system. We constructed a model comprising 4,218 neurons across 13 neuronal subtypes, with neural distribution and connection strengths constrained by anatomical experimental findings. A key innovation of our work is the development of an effective construction and training pipeline tailored for this biologically constrained model. Additionally, we converted an existing simulated whisker sweep dataset into a spiking-based format, enabling our network to be trained and tested on neural signals that more closely mimic those observed in biological systems. The results of object discrimination utilizing whisker signals demonstrate that our barrel cortex model, grounded in biological constraints, achieves a classification accuracy exceeds classical convolutional neural networks (CNNs), recurrent neural networks (RNNs), and long short-term memory networks (LSTMs), by an average of 8.6%, and is on par with recent spiking neural networks (SNNs) in performance. Interestingly, a whisker deprivation experiment, designed in accordance with neuroscience practices, further validates the perceptual capabilities of our model in behavioral tasks.\nCritically, it offers significant biological interpretability: post-training analysis reveals that neurons within our model exhibit firing characteristics and distribution patterns similar to those observed in the actual neuronal systems of the barrel cortex. This study advances our understanding of neural processing in the barrel cortex and exemplifies how integrating detailed biological structures into neural network models can enhance both scientific inquiry and artificial intelligence applications. The code is available at https://github.com/fun0515/RSNN_bfd.", "keywords": [ "Barrel cortex", "biophysical modeling", "sensory-motor integration", "recurrent spiking neural networks" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "UvYrFbKj8j", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Da-shan Shiu", "gender": "Male", "institution": "MediaTek", "country": "", "position": "Principal Researcher" }, { "name": "Meng-Hsi Chen", "gender": "Male", "institution": "MediaTek Research", "country": "TW", "position": "Researcher" }, { "name": "Pei-Chen Ho", "gender": "unknown", "institution": "National Yang Ming Chiao Tung University", "country": "", "position": "Undergrad student" }, { "name": "Yen-Chen Wu", "gender": "Male", "institution": "MediaTek Research", "country": "GB", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 21, "n_ref": 35, "n_ref_all": 50, "n_fig": 2, "n_tab": 7, "L_tab": 4020, "n_element_tab": 217, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1868, "n_element_tab_1": 101, "formula_len_all": 313, "formula_len_all_1": 262, "len_all": 119375, "len_all_1": 67339, "len_abs": 1239, "len_title": 91, "len_sents": 37725, "len_sents_1": 29146, "n_sents": 291, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1245, "title": "Stutter makes large language models smarter", "abs": "Large language models (LLMs) have achieved remarkable success in generating coherent and contextually relevant text. However, their large parameters and high memory requirements limit their efficiency and adoption in industry and academia. Recent studies have shown that dynamically adjusting inference operations can improve model performance without significantly increasing size. In this paper, we introduce the stutter mechanism, a novel method that enhances transformer models by selectively applying additional layers to more challenging tokens. This approach mimics a human speaker\u2019s stutter, allocating more computational effort where needed, thus improving\nlanguage capabilities without generating excessive tokens. Our experiments with various Pythia models demonstrate that the stutter mechanism consistently enhances performance across benchmark datasets. Specifically, the Pythia-410M model, enhanced by our method, outperforms the larger Pythia-1B model on WinoGrande and WSC. Additionally, our method is data-efficient, requiring only less than 1% of the pretraining data for the additional training. These results highlight the stutter mechanism\u2019s potential to enhance LLMs\u2019 efficiency and performance in real-world applications.", "keywords": [ "Self-improvement for LLM" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "UvTo3tVBk2", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Arber Zela", "gender": "Male", "institution": "University of Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Frank Hutter", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen & University of Freiburg", "country": "DE", "position": "Full Professor" }, { "name": "massimiliano pontil", "gender": "Not Specified", "institution": "Istituto Italiano di Tecnologia", "country": "IT", "position": "Principal Researcher" }, { "name": "Julien Niklas Siems", "gender": "Male", "institution": "University of Freiburg", "country": "DE", "position": "PhD student" }, { "name": "J\u00f6rg Franke", "gender": "Male", "institution": "Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Riccardo Grazzi", "gender": "Male", "institution": "Microsoft", "country": "GB", "position": "Researcher" } ], "n_formula": 80, "n_formula_1": 23, "n_ref_uni": 32, "n_ref": 96, "n_ref_all": 137, "n_fig": 7, "n_tab": 5, "L_tab": 2017, "n_element_tab": 129, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 282, "n_element_tab_1": 32, "formula_len_all": 7948, "formula_len_all_1": 2207, "len_all": 234864, "len_all_1": 70188, "len_abs": 1810, "len_title": 68, "len_sents": 76320, "len_sents_1": 32641, "n_sents": 648, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 68, "L_abs": 1820, "title": "Unlocking State-Tracking in Linear RNNs Through Negative Eigenvalues", "abs": "Linear Recurrent Neural Networks (LRNNs) such as Mamba, RWKV, GLA, mLSTM, and DeltaNet have emerged as efficient alternatives to Transformers in large language modeling, offering linear scaling with sequence length and improved training efficiency. However, LRNNs struggle to perform state-tracking which may impair performance in tasks such as code evaluation or tracking a chess game. Even parity, the simplest state-tracking task, which non-linear RNNs like LSTM handle effectively, cannot be solved by current LRNNs. Recently, Sarrof et al. (2024) demonstrated that the failure of LRNNs like Mamba to solve parity stems from restricting the value range of their diagonal state-transition matrices to $[0, 1]$ and that incorporating negative values can resolve this issue. We extend this result to non-diagonal LRNNs, which have recently shown promise in models such as DeltaNet. We prove that finite precision LRNNs with state-transition matrices having only positive eigenvalues cannot solve parity, while complex eigenvalues are needed to count modulo $3$. Notably, we also prove that LRNNs can learn any regular language when their state-transition matrices are products of identity minus vector outer product matrices, each with eigenvalues in the range $[-1, 1]$. Our empirical results confirm that extending the eigenvalue range of models like Mamba and DeltaNet to include negative values not only enables them to solve parity but consistently improves their performance on state-tracking tasks. Furthermore, pre-training LRNNs with an extended eigenvalue range for language modeling achieves comparable performance and stability while showing promise on code and math data. Our work enhances the expressivity of modern LRNNs, broadening their applicability without changing the cost of training or inference.", "keywords": [ "State Tracking", "state space", "mamba", "Linear RNN", "Linear Attention", "GLA", "DeltaNet", "Formal Languages" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 4, 3, 4 ], "presentation_list": [ 4, 3, 4 ], "contribution_list": [ 4, 4, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "UvPdpa4LuV", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Cade W Gordon", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "Undergrad student" }, { "name": "Pieter Abbeel", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Amy X Lu", "gender": "Female", "institution": "University of California Berkeley", "country": "", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 37, "n_ref": 56, "n_ref_all": 74, "n_fig": 10, "n_tab": 2, "L_tab": 2084, "n_element_tab": 80, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1509, "n_element_tab_1": 73, "formula_len_all": 1172, "formula_len_all_1": 839, "len_all": 116805, "len_all_1": 56497, "len_abs": 1316, "len_title": 104, "len_sents": 30953, "len_sents_1": 25994, "n_sents": 225, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1355, "title": "Protein Language Model Fitness is a Matter of Preference", "abs": "Leveraging billions of years of evolution, scientists have trained protein language models (pLMs) to understand the sequence and structure space of proteins aiding in the design of more functional proteins. Although they have shown ability to improve efficiency in engineering, it remains unclear under what conditions they will succeed or fail. We aim to predict the circumstances in which pLMs can successfully perform zero-shot fitness estimation. Our work demonstrates the trends observed over hundreds of deep mutational scans across multiple different fitness objectives. We find that the likelihood, or abstractly, implicit preference of a certain protein sequence imbued during pretraining is predictive fitness prediction capabilities. Both over-preferred and under-preferred wild type sequences harm performance. Generating a causal link between training data and likelihood, we show a power law tail over what data increases protein likelihood which is tied to training sequence homology. Lastly, proteins of low likelihood can be remedied by unsupervised finetuning. In sum, the zero-shot fitness estimation abilities of pLMs can be predicted by the likelihood of the engineered sequence, thus suggesting when pLMs should be deployed in protein maturation campaigns and a way to improve their performance under circumstances of low likelihood.", "keywords": [ "protein language models", "zero-shot fitness prediction" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 4, 3, 4 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "UvMSKonce8", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Hongrui Zhang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Hongxuan Tang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Linfeng Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Mingyang Wang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Shaobo Wang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Weiya Li", "gender": "Male", "institution": "ICBC", "country": "CN", "position": "Researcher" }, { "name": "Xuming Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Xuyang Liu", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "MS student" } ], "n_formula": 59, "n_formula_1": 7, "n_ref_uni": 50, "n_ref": 119, "n_ref_all": 168, "n_fig": 17, "n_tab": 12, "L_tab": 5179, "n_element_tab": 289, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 2082, "n_element_tab_1": 137, "formula_len_all": 3599, "formula_len_all_1": 598, "len_all": 217388, "len_all_1": 69417, "len_abs": 1389, "len_title": 125, "len_sents": 72117, "len_sents_1": 31499, "n_sents": 586, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1396, "title": "Gnothi Seauton: Empowering Faithful Self-Interpretability in Black-Box Models", "abs": "The debate between self-interpretable models and post-hoc explanations for black-box models is central to Explainable AI (XAI). Self-interpretable models, such as concept-based networks, offer insights by connecting decisions to human-understandable concepts but often struggle with performance and scalability. Conversely, post-hoc methods like Shapley values, while theoretically robust, are computationally expensive and resource-intensive. To bridge the gap between these two lines of research, we propose a novel method that combines their strengths, providing theoretically guaranteed self-interpretability for black-box models without compromising prediction accuracy. Specifically, we introduce a parameter-efficient pipeline, AutoGnothi, which integrates a small side network into the black-box model, allowing it to generate Shapley value explanations without changing the original network parameters. This side-tuning approach significantly reduces memory, training, and inference costs, outperforming traditional parameter-efficient methods, where full fine-tuning serves as the optimal baseline. AutoGnothi enables the black-box model to predict and explain its predictions with minimal overhead. Extensive experiments show that AutoGnothi offers accurate explanations for both vision and language tasks, delivering superior computational efficiency with comparable interpretability.", "keywords": [ "Interpretability", "Efficient-AI", "Transformers" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 4, 4 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "Uv7bWrIucU", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Jiajie He", "gender": "Male", "institution": "University of Maryland, Baltimore County", "country": "US", "position": "PhD student" }, { "name": "Keke Chen", "gender": "Male", "institution": ", University of Maryland, Baltimore County", "country": "US", "position": "Associate Professor" }, { "name": "Ethan Gu", "gender": "Male", "institution": "University of Maryland, Baltimore County", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 13, "n_ref_uni": 15, "n_ref": 37, "n_ref_all": 57, "n_fig": 9, "n_tab": 2, "L_tab": 500, "n_element_tab": 21, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 638, "n_element_tab_1": 44, "formula_len_all": 451, "formula_len_all_1": 527, "len_all": 88037, "len_all_1": 60243, "len_abs": 1038, "len_title": 98, "len_sents": 32670, "len_sents_1": 28053, "n_sents": 257, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1136, "title": "Auditing Privacy Protection of Machine Unlearning", "abs": "Machine unlearning aims to remove the effect of specific data from trained models to ensure individuals\u2019 privacy. However, it\u2019s arguable how to evaluate whether the privacy protection goal is achieved by machine unlearning. Furthermore, recent studies show unlearning may also increase the retained samples\u2019 privacy risks. This paper takes a holistic approach to auditing both unlearned and retained samples\u2019 privacy risks before and after unlearning. \nWe derive the privacy criteria for unlearned and retained samples, respectively, based on the perspectives of differential privacy and membership inference attacks. To make the auditing practical, we also develop an efficient membership inference attack, A-LiRA, utilizing data augmentation to reduce the cost of shadow model training. Our experimental findings indicate that existing machine unlearning algorithms do not consistently protect the privacy of unlearned samples and may inadvertently compromise the privacy of retained samples. For reproducibility, we have pubished our code.\\footnote{ \\url{https://anonymous.4open.science/r/Auditing-machine-unlearning-CB10/README.md}}", "keywords": [ "Machine unlearning", "Auditing privacy", "Privacy estimation", "Membership inference attack" ], "rating_list": [ 3, 6, 5, 6, 1 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 4, 1 ], "contribution_list": [ 2, 2, 3, 3, 1 ], "confidence_list": [ 5, 3, 3, 4, 4 ] }, { "paper_id": "UunCPtPOlZ", "primary_area": "datasets and benchmarks", "n_author": 19, "author_info_list": [ { "name": "Bing Wang", "gender": "Male", "institution": "Beihang University", "country": "", "position": "PhD student" }, { "name": "Boyang Wang", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" }, { "name": "Changyu Ren", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Ge Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Hongcheng Guo", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Liu Jiaheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jian Yang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "JinKe", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Linzheng Chai", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Liqun Yang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" }, { "name": "Shukai Liu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Sufeng Duan", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tao Sun", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Tongliang Li", "gender": "Male", "institution": "Beijing Information Science and Technology University", "country": "CN", "position": "Lecturer" }, { "name": "Xianjie Wu", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Yuwei Yin", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Zhaoxiang Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhoujun Li", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Zekun Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 43, "n_ref": 111, "n_ref_all": 141, "n_fig": 9, "n_tab": 8, "L_tab": 13726, "n_element_tab": 2784, "n_fig_1": 7, "n_tab_1": 8, "L_tab_1": 8599, "n_element_tab_1": 1625, "formula_len_all": 517, "formula_len_all_1": 434, "len_all": 253605, "len_all_1": 80899, "len_abs": 1352, "len_title": 95, "len_sents": 42710, "len_sents_1": 21181, "n_sents": 332, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 46, "L_abs": 1282, "title": "McEval: Massively Multilingual Code Evaluation", "abs": "Code large language models (LLMs) have shown remarkable advances in code understanding, completion, and generation tasks. Programming benchmarks, comprised of a selection of code challenges and corresponding test cases, serve as a standard to evaluate the capability of different LLMs in such tasks. However, most existing benchmarks primarily focus on Python and are still restricted to a limited number of languages, where other languages are translated from the Python samples (e.g. MultiPL-E) degrading the data diversity. To further facilitate the research of code LLMs, we propose a massively multilingual code benchmark covering 40 programming languages (McEval) with 16K test samples, which substantially pushes the limits of code LLMs in multilingual scenarios. The benchmark contains challenging code completion, understanding, and generation evaluation tasks with finely curated massively multilingual instruction corpora McEval-Instruct. In addition, we introduce an effective multilingual coder McEval trained on McEval-Instruct to support multilingual programming language generation. Extensive experimental results on McEval show that there is still a difficult journey between open-source models and closed-source LLMs (e.g. GPT-series models) in numerous languages.", "keywords": [ "Benchmark", "Code Intelligence", "Multilingual", "Large Language Model", "Multilingual Multitask Code Evaluation" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 2, 3, 4, 4 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "UuZDosomkp", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Shiguang Wu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Yaqing Wang", "gender": "Female", "institution": "Beijing Institute of Mathematical Sciences and Applications", "country": "CN", "position": "Associate Professor" }, { "name": "quanming yao", "gender": "Male", "institution": "Department of Electronic Engineering, Tsinghua University", "country": "", "position": "Assistant Professor" }, { "name": "An Bian", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Senior researcher " } ], "n_formula": 21, "n_formula_1": 11, "n_ref_uni": 53, "n_ref": 89, "n_ref_all": 104, "n_fig": 4, "n_tab": 7, "L_tab": 1429, "n_element_tab": 196, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2365, "n_element_tab_1": 153, "formula_len_all": 1685, "formula_len_all_1": 925, "len_all": 152378, "len_all_1": 62646, "len_abs": 3816, "len_title": 128, "len_sents": 39256, "len_sents_1": 26425, "n_sents": 314, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1104, "title": "ConML: A Universal Meta-Learning Framework with Task-Level Contrastive Learning", "abs": "Meta-learning enables learning systems to adapt quickly to new tasks, similar to humans. To emulate this human-like rapid learning and enhance alignment and discrimination abilities, we propose ConML, a universal meta-learning framework that can be applied to various meta-learning algorithms without relying on specific model architectures nor target models. The core of ConML is task-level contrastive learning, which extends contrastive learning from the representation space in unsupervised learning to the model space in meta-learning. By leveraging task identity as an additional supervision signal during meta-training, we contrast the outputs of the meta-learner in the model space, minimizing inner-task distance (between models trained on different subsets of the same task) and maximizing inter-task distance (between models from different tasks). We demonstrate that ConML integrates seamlessly with optimization-based, metric-based, and amortization-based meta-learning algorithms, as well as in-context learning, resulting in performance improvements across diverse few-shot learning tasks.", "keywords": [ "Meta-Learning", "Contrastive Learning", "In-Context Learning" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "UuK99g47op", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Hao Xu", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Pengyu Hong", "gender": "Male", "institution": "Brandeis University", "country": "US", "position": "Full Professor" }, { "name": "Zhengyang Zhou", "gender": "Male", "institution": "Brandeis University", "country": "", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 9, "n_ref_uni": 40, "n_ref": 76, "n_ref_all": 88, "n_fig": 4, "n_tab": 3, "L_tab": 2554, "n_element_tab": 140, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2554, "n_element_tab_1": 140, "formula_len_all": 1803, "formula_len_all_1": 653, "len_all": 136770, "len_all_1": 57097, "len_abs": 1487, "len_title": 124, "len_sents": 35434, "len_sents_1": 24441, "n_sents": 271, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1493, "title": "Multimodal Fusion with Relational Learning for Molecular Property Prediction", "abs": "Graph-based molecular representation learning is essential for accurately predicting molecular properties in drug discovery and materials science; however, it faces significant challenges due to the intricate relationships among molecules and the limited chemical knowledge utilized during training. While contrastive learning is often employed to handle molecular relationships, its reliance on binary metrics is insufficient for capturing the complexity of these interactions. Multimodal fusion has gained attention for property reasoning, but previous work has explored only a limited range of modalities, and the optimal stages for fusing different modalities in molecular property tasks remain underexplored. In this paper, we introduce MMFRL (Multimodal Fusion with Relational Learning for Molecular Property Prediction), a novel framework designed to overcome these limitations. Our method enhances embedding initialization through multi-modal pre-training using relational learning. We also conduct a systematic investigation into the impact of modality fusion at different stages\u2014early, intermediate, and late\u2014highlighting their advantages and shortcomings. Extensive experiments on MoleculeNet benchmarks demonstrate that MMFRL significantly outperforms existing methods. Furthermore, MMFRL enables task-specific optimizations. Additionally, the explainability of MMFRL provides valuable chemical insights, emphasizing its potential to enhance real-world drug discovery applications.", "keywords": [ "Molecular Graph Representation Learning", "Multi-modality Learning", "Contrastive Learning", "Drug Discovery" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "UtFoFyPYQo", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Bohui An", "gender": "unknown", "institution": "China University of Petroleum, Beijing", "country": "CN", "position": "MS student" }, { "name": "Zengxia Guo", "gender": "Female", "institution": "China University of Petroleum, Beijing", "country": "CN", "position": "MS student" }, { "name": "Zhongqi Lu", "gender": "unknown", "institution": "China University of Petroleum, Beijing", "country": "CN", "position": "Associate Professor" } ], "n_formula": 37, "n_formula_1": 39, "n_ref_uni": 12, "n_ref": 24, "n_ref_all": 42, "n_fig": 9, "n_tab": 3, "L_tab": 4766, "n_element_tab": 70, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 192, "n_element_tab_1": 18, "formula_len_all": 3172, "formula_len_all_1": 3121, "len_all": 121429, "len_all_1": 58699, "len_abs": 825, "len_title": 138, "len_sents": 35816, "len_sents_1": 24095, "n_sents": 267, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 828, "title": "Approximated Behavioral Metric-based State Projection for Federated Reinforcement Learning", "abs": "Federated reinforcement learning (FRL) methods usually share the encrypted local state or policy information and help each client to learn from others while preserving everyone's privacy.\nIn this work, we propose that sharing the approximated behavior metric-based state projection function is a promising way to enhance the performance of FRL and concurrently provides an effective protection of sensitive information.\nWe introduce FedRAG, a FRL framework to learn a computationally practical projection function of states for each client and aggregating the parameters of projection functions at a central server. \nThe FedRAG approach shares no sensitive task-specific information, yet provides information gain for each client.\nWe conduct extensive experiments on the DeepMind Control Suite to demonstrate insightful results.", "keywords": [ "Federated Learning;Reinforcement Learning;Bisimulation Metric" ], "rating_list": [ 3, 6, 5 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 2 ] }, { "paper_id": "Ut4XAYg0NF", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Ammar Gilani", "gender": "Male", "institution": "Northwestern University", "country": "", "position": "PhD student" }, { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" } ], "n_formula": 123, "n_formula_1": 8, "n_ref_uni": 60, "n_ref": 102, "n_ref_all": 124, "n_fig": 3, "n_tab": 7, "L_tab": 3559, "n_element_tab": 76, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1452, "n_element_tab_1": 33, "formula_len_all": 7160, "formula_len_all_1": 685, "len_all": 275783, "len_all_1": 85330, "len_abs": 734, "len_title": 83, "len_sents": 84231, "len_sents_1": 29998, "n_sents": 1063, "n_sents_1": 586, "venue": "ICLR2025", "decision": "withdraw", "L_title": 35, "L_abs": 737, "title": "Explainable Sequential Optimization", "abs": "We propose formulating stochastic model predictive control into a coalition game to use Shapley values for feature attribution. Such analysis is crucial for transparency and achieving optimal outcomes in high-stake applications such as portfolio optimization and autonomous driving. We categorize Shapley values estimation methods into three families: those based on weighted linear regression, sampling permutations, and multilinear extension. We survey, benchmark, and provide valuable insight into these methods, previously not attempted in this context. Our experiments show that halved Owen sampling from multilinear extension and KernelShap-Paired from weighted linear regression, both utilizing antithetic sampling, perform best.", "keywords": [ "Explainability", "Sequential Optimization" ], "rating_list": [ 1, 5, 3, 6, 5 ], "soundness_list": [ 1, 2, 3, 2, 3 ], "presentation_list": [ 1, 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 1, 3 ] }, { "paper_id": "UstOpZCESc", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Elmar Rueckert", "gender": "Male", "institution": "Montanuniversit\u00e4t Leoben", "country": "AT", "position": "Full Professor" }, { "name": "Ozan Ozdenizci", "gender": "Male", "institution": "Montanuniversit\u00e4t Leoben", "country": "AT", "position": "Researcher" }, { "name": "Robert Legenstein", "gender": "Male", "institution": "Graz University of Technology", "country": "AT", "position": "Professor" } ], "n_formula": 15, "n_formula_1": 13, "n_ref_uni": 46, "n_ref": 88, "n_ref_all": 106, "n_fig": 2, "n_tab": 2, "L_tab": 335, "n_element_tab": 8, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 2410, "n_element_tab_1": 27, "formula_len_all": 1357, "formula_len_all_1": 1098, "len_all": 134771, "len_all_1": 74114, "len_abs": 1410, "len_title": 79, "len_sents": 40732, "len_sents_1": 33471, "n_sents": 244, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 31, "L_abs": 1415, "title": "Privacy-Aware Lifelong Learning", "abs": "Lifelong learning algorithms enable models to incrementally acquire new knowledge without forgetting previously learned information. Contrarily, the field of machine unlearning focuses on explicitly forgetting certain previous knowledge from pretrained models when requested, in order to comply with data privacy regulations on the right-to-be-forgotten. Enabling efficient lifelong learning with the capability to selectively unlearn sensitive information from models presents a critical and largely unaddressed challenge with contradicting objectives. We address this problem from the perspective of simultaneously preventing catastrophic forgetting and allowing forward knowledge transfer during task-incremental learning, while ensuring exact task unlearning and minimizing memory requirements, based on a single neural network model to be adapted. Our proposed solution, privacy-aware lifelong learning (PALL), involves optimization of task-specific sparse subnetworks with parameter sharing within a single architecture. We additionally utilize an episodic memory rehearsal mechanism to facilitate exact unlearning without performance degradations. We empirically demonstrate the scalability of PALL across various architectures in image classification, and provide a state-of-the-art solution that uniquely integrates lifelong learning and privacy-aware unlearning mechanisms for responsible AI applications.", "keywords": [ "lifelong learning", "exact machine unlearning", "task-incremental continual learning", "sparse subnetworks", "knowledge transfer" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Usklli4gMc", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Jia-Chen Gu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Kai-Wei Chang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Associate Professor" }, { "name": "Mohsen Fayyaz", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Nanyun Peng", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Associate Professor" }, { "name": "Wenbo Hu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "MS student" }, { "name": "Zi-Yi Dou", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Pan Lu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 42, "n_ref": 88, "n_ref_all": 109, "n_fig": 6, "n_tab": 4, "L_tab": 5818, "n_element_tab": 350, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1681, "n_element_tab_1": 67, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 188569, "len_all_1": 53218, "len_abs": 1350, "len_title": 127, "len_sents": 33524, "len_sents_1": 21790, "n_sents": 216, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1499, "title": "MRAG-Bench: Vision-Centric Evaluation for Retrieval-Augmented Multimodal Models", "abs": "Existing multimodal retrieval benchmarks primarily focus on evaluating whether models can retrieve and utilize external textual knowledge for question answering. However, there are scenarios where retrieving visual information is either more beneficial or easier to access than textual data. \nIn this paper, we introduce a multimodal retrieval-augmented generation benchmark, MRAG-Bench, in which we systematically identify and categorize scenarios where visually augmented knowledge is better than textual knowledge, for instance, more images from varying viewpoints.\nMRAG-Bench consists of 16,130 images and 1,353 human-annotated multiple-choice questions across 9 distinct scenarios. With MRAG-Bench, we conduct an evaluation of 10 open-source and 4 proprietary large vision-language models (LVLMs). Our results show that all LVLMs exhibit greater improvements when augmented with images compared to textual knowledge, confirming that MRAG-Bench is vision-centric. Additionally, we conduct extensive analysis with MRAG-Bench, which offers valuable insights into retrieval-augmented LVLMs. Notably, the top-performing model, GPT-4o, faces challenges in effectively leveraging retrieved knowledge, achieving only a 5.82\\% improvement with ground-truth information, in contrast to a 33.16\\% improvement observed in human participants. These findings highlight the importance of MRAG-Bench in encouraging the community to enhance LVLMs' ability to utilize retrieved visual knowledge more effectively.", "keywords": [ "multimodal", "rag", "vision-language", "visual knowledge" ], "rating_list": [ 6, 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "Usk4KzBxLW", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "An Yan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Hua Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Huigen Ye", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yaoyang Cheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Intern" } ], "n_formula": 21, "n_formula_1": 11, "n_ref_uni": 43, "n_ref": 96, "n_ref_all": 130, "n_fig": 5, "n_tab": 27, "L_tab": 11500, "n_element_tab": 1516, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 937, "n_element_tab_1": 168, "formula_len_all": 2263, "formula_len_all_1": 560, "len_all": 323622, "len_all_1": 67323, "len_abs": 1377, "len_title": 133, "len_sents": 124226, "len_sents_1": 31875, "n_sents": 840, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1385, "title": "Large Language Model-driven Large Neighborhood Search for Large-Scale MILP Problems", "abs": "Large Neighborhood Search (LNS) is a widely used method for solving large-scale Mixed Integer Linear Programming (MILP) problems. The effectiveness of LNS crucially depends on the choice of the search neighborhood. However, existing strategies either rely on expert knowledge or computationally expensive Machine Learning (ML) approaches, both of which struggle to scale effectively for large problems. To address this, we propose LLM-LNS, a novel Large Language Model (LLM)-driven LNS framework for large-scale MILP problems. Our approach introduces a dual-layer self-evolutionary LLM agent to automate neighborhood selection, discovering effective strategies with scant small-scale training data that generalize well to large-scale MILPs. The inner layer evolves heuristic strategies to ensure convergence, while the outer layer evolves evolutionary prompt strategies to maintain diversity. Experimental results demonstrate that the proposed dual-layer agent outperforms state-of-the-art agents such as FunSearch and EOH. Furthermore, the full LLM-LNS framework surpasses manually designed LNS algorithms like ACP, ML-based LNS methods like CL-LNS, and large-scale solvers such as Gurobi and SCIP. It also achieves superior performance compared to advanced ML-based MILP optimization frameworks like GNN\\&GBDT and Light-MILPopt, further validating the effectiveness of our approach.", "keywords": [ "Mixed Integer Linear Programming", "Large Neighborhood Search", "Large Language Model" ], "rating_list": [ 3, 5, 5, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "UsgFE3cxuP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Guanbin Li", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Ming-Ming Cheng", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Full Professor" }, { "name": "Qibin Hou", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Quan-Sheng Zeng", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "MS student" }, { "name": "Yunheng Li", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "PhD student" }, { "name": "Zhou Daquan", "gender": "Male", "institution": "Bytedance", "country": "US", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 55, "n_ref": 134, "n_ref_all": 176, "n_fig": 8, "n_tab": 15, "L_tab": 4674, "n_element_tab": 632, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1720, "n_element_tab_1": 170, "formula_len_all": 321, "formula_len_all_1": 246, "len_all": 177023, "len_all_1": 54620, "len_abs": 1224, "len_title": 139, "len_sents": 44842, "len_sents_1": 24729, "n_sents": 338, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1183, "title": "MaskCLIP++: A Mask-Based CLIP Fine-tuning Framework for Open-Vocabulary Image Segmentation", "abs": "Open-vocabulary image segmentation has been advanced through the synergy between mask generators and vision-language models like Contrastive Language-Image Pre-training (CLIP).\nPrevious approaches focus on generating masks while aligning mask features with text embeddings during training.\nIn this paper, we observe that relying on generated low-quality masks can weaken the alignment of vision and language in regional representations.\nThis motivates us to present a new fine-tuning framework, named MaskCLIP++, which uses ground-truth masks instead of generated masks to enhance the mask classification capability of CLIP.\nDue to the limited diversity of image segmentation datasets with mask annotations, we propose incorporating a consistency alignment constraint during fine-tuning, which alleviates categorical bias toward the fine-tuning dataset.\nAfter low-cost fine-tuning, combining with the mask generator in previous state-of-the-art mask-based open vocabulary segmentation methods, we achieve performance improvements of +1.7, +2.3, +2.1, +3.1, and +0.3 mIoU on the A-847, PC-459, A-150, PC-59, and PAS-20 datasets, respectively.\nOur code will be made publicly available.", "keywords": [ "Open Vocabulary", "Image Segmentation", "Vision Language Model" ], "rating_list": [ 8, 5, 5, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "Usa4pF1e5I", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Maryam Mehri Dehnavi", "gender": "Female", "institution": "Department of Computer Science", "country": "", "position": "Associate Professor" }, { "name": "Mohammad Mozaffari", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 39, "n_ref": 67, "n_ref_all": 95, "n_fig": 1, "n_tab": 11, "L_tab": 5543, "n_element_tab": 638, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2946, "n_element_tab_1": 287, "formula_len_all": 1617, "formula_len_all_1": 1121, "len_all": 144104, "len_all_1": 61867, "len_abs": 1283, "len_title": 116, "len_sents": 40659, "len_sents_1": 26057, "n_sents": 286, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1287, "title": "SLiM: One-shot Quantized Sparse Plus Low-rank Approximation of LLMs", "abs": "Large Language Models (LLMs) have revolutionized natural language understanding and generation tasks but suffer from high memory consumption and slow inference times due to their large parameter sizes. Traditional model compression techniques, such as quantization and pruning, mitigate these issues but often require retraining to maintain accuracy, which is computationally expensive. This paper introduces SLiM, a novel approach for compressing LLMs using a one-shot Quantized Sparse Plus Low-rank Approximation. SLiM eliminates the need for costly retraining by combining a symmetric quantization method (SLiM-Quant) with a saliency-based low-rank approximation. Our method reduces quantization error while leveraging sparse representations compatible with accelerated hardware architectures. Additionally, we propose a parameter-efficient fine-tuning recipe that significantly reduces overhead compared to conventional quantization-aware training. SLiM achieves up to a 5.4\\% improvement in model accuracy for sparsity patterns like 2:4, and the fine-tuning step further enhances accuracy by up to 5.6\\%, demonstrating state-of-the-art performance. This work provides a pathway for efficiently deploying large models in memory-constrained environments without compromising accuracy.", "keywords": [ "sparsity", "2:4 sparsity", "quantization", "low-rank", "lora" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 1, 3, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "UsVJlgD1F7", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Kai Hu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Tianjun Yao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Intern" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Yongqiang Chen", "gender": "unknown", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Zhiqiang Shen", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" } ], "n_formula": 109, "n_formula_1": 16, "n_ref_uni": 70, "n_ref": 185, "n_ref_all": 205, "n_fig": 7, "n_tab": 9, "L_tab": 4292, "n_element_tab": 267, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 624, "n_element_tab_1": 55, "formula_len_all": 8223, "formula_len_all_1": 763, "len_all": 291479, "len_all_1": 68806, "len_abs": 2455, "len_title": 98, "len_sents": 82163, "len_sents_1": 31491, "n_sents": 624, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1399, "title": "Learning Graph Invariance by Harnessing Spuriosity", "abs": "Recently, graph invariant learning has become the _de facto_ approach to tackle the Out-of-Distribution (OOD) generalization failure in graph representation learning. They generically follow the framework of invariant risk minimization to capture the invariance of graph data from different environments. Despite some success, it remains unclear to what extent existing approaches have captured invariant features for OOD generalization on graphs. In this work, we find that representative OOD methods such as IRM and VRex, and their variants on graph invariant learning may have captured a limited set of invariant features. To tackle this challenge, we propose $\\texttt{LIRS}$, a novel learning framework designed to **L**earn graph **I**nvariance by **R**emoving **S**purious features. Different from most existing approaches that _directly_ learn the invariant features, $\\texttt{LIRS}$ takes an _indirect_ approach by first learning the spurious features and then removing them from the ERM-learned features, which contains both spurious and invariant features. We demonstrate that learning the invariant graph features in an _indirect_ way can learn a more comprehensive set of invariant features. Moreover, our proposed method outperforms the second-best method by as much as 25.50% across all competitive baseline methods, highlighting its effectiveness in learning graph invariant features.", "keywords": [ "Out-of-Distribution Generalization", "Invariant Learning", "Graph Neural Networks" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "UsRKFYR4lM", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Junyi Chai", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Shenyu Lu", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Xiaoqian Wang", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 50, "n_formula_1": 14, "n_ref_uni": 39, "n_ref": 99, "n_ref_all": 126, "n_fig": 8, "n_tab": 12, "L_tab": 5245, "n_element_tab": 404, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1907, "n_element_tab_1": 146, "formula_len_all": 4274, "formula_len_all_1": 939, "len_all": 172845, "len_all_1": 65603, "len_abs": 2073, "len_title": 111, "len_sents": 47661, "len_sents_1": 28330, "n_sents": 440, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1567, "title": "Mitigating Spurious Correlations in Zero-Shot Multimodal Models", "abs": "Multimodal models or Vision Language Models (VLMs) have reshaped the paradigm in machine learning, offering zero-shot capabilities that require no additional training when adapted to new classification tasks. However, despite their advancements, spurious correlations still exist in VLMs. Existing approaches to tackle this issue often require target label annotations, contradicting the principle of zero-shot classification, or they primarily focus on a single modality, risking misalignment between text and image modalities. Others rely on extensive domain knowledge or large language models (LLMs) to characterize spurious features, making the performance sensitive to the generated prompts and undermining zero-shot capability. In response, we propose a new solution that tackles spurious correlations in VLMs within the zero-shot setting. Our approach utilizes a translation operation that preserves the latent space distribution to address issues of spurious correlations. In particular, our method is grounded in and inspired by a theoretical analysis, which identifies that the optimal translation directions are along the spurious vector. As VLMs unify two modalities, we compute spurious vectors from the text prompts and guide the translation for image embeddings, aligning the requirements for the fusion of different modalities in VLMs. We conducted experiments on benchmark datasets, which have shown significant improvements in worst-group accuracy. Additionally, our visualizations of VLMs further demonstrate the effectiveness of this intervention.", "keywords": [ "Spurious correlation", "Zero shot", "Multimodal models" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "UsMTuRraOR", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Ariel Barel", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Principal Researcher" }, { "name": "Eran Iceland", "gender": "Male", "institution": "private", "country": "IL", "position": "Principal Researcher" }, { "name": "Itay Naeh", "gender": "Male", "institution": "Electrical Engineering Department, Technion \u2013 Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "Researcher" }, { "name": "Shie Mannor", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Full Professor" }, { "name": "Stav Belogolovsky", "gender": "unknown", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 18, "n_ref": 40, "n_ref_all": 54, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 105, "formula_len_all_1": 105, "len_all": 95924, "len_all_1": 73061, "len_abs": 1091, "len_title": 84, "len_sents": 39139, "len_sents_1": 36518, "n_sents": 263, "n_sents_1": 246, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1093, "title": "Human-like Communication Strategies for Improved Multi-Agent Reinforcement Learning", "abs": "Multi-Agent Reinforcement Learning (MARL) has seen significant progress in recent years, enabling multiple agents to coordinate and optimize their actions in complex environments. However, integrating effective communication protocols into MARL frameworks remains a challenge, as it introduces issues such as increased state space dimensionality, lack of stationarity, and the need for interpretability. Inspired by human communication, which relies on prior knowledge, contextual awareness, and efficient information exchange, we propose a novel framework for incorporating human-like communication strategies to enhance the learning process. Motivated by recent advancements in natural language processing (NLP), multi-modal AI and object detection, we use text-to-mask models and human feedback to learn compact and informative communication strategies that facilitate coordination among agents to improve the overall performance. We demonstrate the efficiency of our approach on various multi-agent tasks and provide insights into emergent communication behaviors observed during training.", "keywords": [ "RL", "MARL", "communication", "RLHF" ], "rating_list": [ 1, 3, 5, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "Us1RXG1Ji2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Jason D. Lee", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Jiajun Zhu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Pan Li", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Peihao Wang", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Ruisi Cai", "gender": "Female", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Zhangyang Wang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 52, "n_ref": 132, "n_ref_all": 150, "n_fig": 5, "n_tab": 9, "L_tab": 3809, "n_element_tab": 381, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 287, "n_element_tab_1": 38, "formula_len_all": 806, "formula_len_all_1": 718, "len_all": 200195, "len_all_1": 57957, "len_abs": 1290, "len_title": 139, "len_sents": 55554, "len_sents_1": 27027, "n_sents": 452, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1312, "title": "Rethinking Addressing in Language Models via Contextualized Equivariant Positional Encoding", "abs": "Transformers rely on both content-based and position-based addressing mechanisms to make predictions, but existing positional encoding techniques often diminish the effectiveness of position-based addressing. Many current methods enforce rigid patterns in attention maps, limiting the ability to model long-range dependencies and adapt to diverse tasks. Additionally, most positional encodings are learned as general biases, lacking the specialization required for different instances within a dataset.\nTo address this, we propose TAPE: con**T**extualized equivari**A**nt **P**osition **E**mbedding, \na novel framework that enhances positional embeddings by incorporating sequence content across layers. TAPE introduces dynamic, context-aware positional encodings, overcoming the constraints of traditional fixed patterns. By enforcing permutation and orthogonal equivariance, TAPE ensures the stability of positional encodings during updates, improving robustness and adaptability. Our method can be easily integrated into pre-trained transformers, offering parameter-efficient fine-tuning with minimal overhead.\nExtensive experiments shows that TAPE achieves superior performance in language modeling, arithmetic reasoning, and long-context retrieval tasks compared to existing positional embedding techniques.", "keywords": [ "Positional Encoding", "Equivariance", "Addressing" ], "rating_list": [ 8, 3, 8, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "UrbTqU2xrb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Bo Dai", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Jingbo Wang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Junting Dong", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Mingze Sun", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Qi Fang", "gender": "Male", "institution": "NetEase Games AI Lab ", "country": "", "position": "Researcher" }, { "name": "Yan-Pei Cao", "gender": "Male", "institution": "VAST", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 7, "n_ref": 9, "n_ref_all": 19, "n_fig": 6, "n_tab": 3, "L_tab": 277, "n_element_tab": 19, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 502, "n_element_tab_1": 24, "formula_len_all": 231, "formula_len_all_1": 231, "len_all": 138885, "len_all_1": 51722, "len_abs": 1583, "len_title": 114, "len_sents": 29768, "len_sents_1": 25948, "n_sents": 204, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1589, "title": "Clothing-disentangled 3D character generation from a single image", "abs": "This paper tackles the challenge of generating clothing-disentangled 3D characters from a single image. Existing approaches typically employ multi-layer 3D representations to model the body and each garment and then iteratively optimize these representations to fit the observations, which is time-consuming and not scalable. To address this, we propose the first feed-forward method enabling efficient and robust clothing disentanglement. Our approach first generates the multi-view images for each component of the clothed character and then employs a generalizable multi-view reconstruction method to create the 3D models of each component. For high-quality disentanglement, we propose a two-stage disentanglement approach that first disentangles each component in the 2D image space and then generates the multi-view images for each part. During the 2D component disentanglement stage, we introduce a novel multi-part diffusion model that allows information exchange among different components. Additionally, for component combination, we incorporate a novel combination attention mechanism into the multi-view diffusion model, enabling the integration of information from multiple parts to create the final combined character. For training, we have contributed a large clothing-disentangled character dataset consisting of more than 10k anime characters. Extensive experiments demonstrate that our proposed approach not only facilitates efficient and high-quality disentangled 3D character generation with distinct clothing layers but also supports various cloth editing applications.", "keywords": [ "3D human reconstruction; 3D human generation" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "UrKbn51HjA", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Xiaoyi MAI", "gender": "Female", "institution": "Universit\u00e9 de Toulouse-le-Mirail (Toulouse II)", "country": "FR", "position": "Assistant Professor" }, { "name": "Zhenyu Liao", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 123, "n_formula_1": 26, "n_ref_uni": 44, "n_ref": 76, "n_ref_all": 146, "n_fig": 10, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9619, "formula_len_all_1": 1483, "len_all": 196416, "len_all_1": 74030, "len_abs": 1856, "len_title": 133, "len_sents": 58066, "len_sents_1": 33604, "n_sents": 553, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1355, "title": "The Breakdown of Gaussian Universality in Classification of High-dimensional Mixtures", "abs": "The assumption of Gaussian or Gaussian mixture data has been extensively exploited in a long series of precise performance analyses of machine learning (ML) methods, on large datasets having comparably numerous samples and features. \nTo relax this restrictive assumption, subsequent efforts have been devoted to establish \"Gaussian equivalent principles\" by studying scenarios of Gaussian universality where the asymptotic performance of ML methods on non-Gaussian data remains unchanged when replaced with Gaussian data having the *same mean and covariance*.\nBeyond the realm of Gaussian universality, there are few exact results on how the data distribution affects the learning performance. \n\nIn this article, we provide a precise high-dimensional characterization of empirical risk minimization, for classification under a general mixture data setting of linear factor models that extends Gaussian mixtures. \nThe Gaussian universality is shown to break down under this setting, in the sense that the asymptotic learning performance depends on the data distribution *beyond* the class means and covariances.\nTo clarify the limitations of Gaussian universality in classification of mixture data and to understand the impact of its breakdown, we specify conditions for Gaussian universality and discuss their implications for the choice of loss function.", "keywords": [ "High-dimensional statistics", "random matrix theory", "Gaussian universality", "empirical risk minimization", "mixture models", "linear factor models" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "Uqxf2YH9LZ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Lei Feng", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" }, { "name": "Qi Wei", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Shuo He", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Yuwei Niu", "gender": "Male", "institution": "Chongqing University", "country": "CN", "position": "Undergrad student" }, { "name": "Zongyu Wu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 62, "n_ref": 162, "n_ref_all": 199, "n_fig": 5, "n_tab": 26, "L_tab": 5747, "n_element_tab": 596, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1691, "n_element_tab_1": 198, "formula_len_all": 931, "formula_len_all_1": 813, "len_all": 192260, "len_all_1": 65401, "len_abs": 1720, "len_title": 119, "len_sents": 51696, "len_sents_1": 30085, "n_sents": 376, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1728, "title": "BDetCLIP: Multimodal Prompting Contrastive Test-Time Backdoor Detection", "abs": "Multimodal contrastive learning methods (e.g., CLIP) have shown impressive zero-shot classification performance due to their strong ability to joint representation learning for visual and textual modalities. However, recent research revealed that multimodal contrastive learning on poisoned pre-training data with a small proportion of maliciously backdoored data can induce backdoored CLIP that could be attacked by inserted triggers in downstream tasks with a high success rate. To defend against backdoor attacks on CLIP, existing defense methods focus on either the pre-training stage or the fine-tuning stage, which would unfortunately cause high computational costs due to numerous parameter updates and are not applicable in the black-box setting. In this paper, we provide the first attempt at a computationally efficient backdoor detection method to defend against backdoored CLIP in the inference stage. We empirically find that the visual representations of backdoored images are insensitive to both benign and malignant changes in class description texts. Motivated by this observation, we propose BDetCLIP, a novel test-time backdoor detection method based on contrastive prompting. Specifically, we first prompt the language model (e.g., GPT-4) to produce class-related description texts (benign) and class-perturbed random texts (malignant) by specially designed instructions. Then, the distribution difference in cosine similarity between images and the two types of class description texts can be used as the criterion to detect backdoor samples. Extensive experiments validate that our proposed BDetCLIP is superior to state-of-the-art backdoor detection methods, in terms of both effectiveness and efficiency.", "keywords": [ "Multimodal contrastive learning", "test-time backdoor detection", "contrastive prompting" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "UqrSyATn7F", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Haoxiang Guan", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "John E. Hopcroft", "gender": "Male", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "Full Professor" }, { "name": "Kaiyuan Gao", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Kun He", "gender": "Female", "institution": "Huazhong University of Sceince and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Lijun Wu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Qizhi Pei", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Yusong Wang", "gender": "Male", "institution": "MSR AI4Science China", "country": "CN", "position": "Intern" }, { "name": "Zun Wang", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 39, "n_ref": 87, "n_ref_all": 105, "n_fig": 5, "n_tab": 5, "L_tab": 1660, "n_element_tab": 217, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2004, "n_element_tab_1": 224, "formula_len_all": 1116, "formula_len_all_1": 1114, "len_all": 139202, "len_all_1": 71672, "len_abs": 1752, "len_title": 117, "len_sents": 42379, "len_sents_1": 31457, "n_sents": 368, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1761, "title": "Tokenizing 3D Molecule Structure with Quantized Spherical Coordinates", "abs": "The application of language models (LMs) to molecular structure generation using line notations such as SMILES and SELFIES has been well-established in the field of cheminformatics. However, extending these models to generate 3D molecular structures presents significant challenges. Two primary obstacles emerge: (1) the difficulty in designing a 3D line notation that ensures SE(3)-invariant atomic coordinates, and (2) the non-trivial task of tokenizing continuous coordinates for use in LMs, which inherently require discrete inputs.\nTo address these challenges, we propose Mol-StrucTok, a novel method for tokenizing 3D molecular structures. Our approach comprises two key innovations: (1) We design a line notation for 3D molecules by extracting local atomic coordinates in a spherical coordinate system. This notation builds upon existing 2D line notations and remains agnostic to their specific forms, ensuring compatibility with various molecular representation schemes. (2) We employ a Vector Quantized Variational Autoencoder (VQ-VAE) to tokenize these coordinates, treating them as generation descriptors. To further enhance the representation, we incorporate neighborhood bond lengths and bond angles as understanding descriptors. Leveraging this tokenization framework, we train a GPT-2 style model for 3D molecular generation tasks. Results demonstrate strong performance with significantly faster generation speeds and competitive chemical stability compared to previous methods. \nFurther, by integrating our learned discrete representations into Graphormer model for property prediction on QM9 dataset, Mol-StrucTok reveals consistent improvements across various molecular properties, underscoring the versatility and robustness of our approach.", "keywords": [ "molecule structure tokenization", "generation", "spherical coordinates", "VQ-VAE" ], "rating_list": [ 5, 8, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "UqrFPhcmFp", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Bin Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Feng Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jie Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xize Liang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zhihao Shi", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zhiwei Zhuang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "", "position": "Undergrad student" } ], "n_formula": 53, "n_formula_1": 17, "n_ref_uni": 31, "n_ref": 143, "n_ref_all": 200, "n_fig": 11, "n_tab": 8, "L_tab": 6254, "n_element_tab": 469, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4260, "formula_len_all_1": 1254, "len_all": 203405, "len_all_1": 66104, "len_abs": 1367, "len_title": 115, "len_sents": 60369, "len_sents_1": 28905, "n_sents": 560, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1352, "title": "Accurate and Scalable Graph Neural Networks via Message Invariance", "abs": "Message passing-based graph neural networks (GNNs) have achieved great success in many real-world applications. For a sampled mini-batch of target nodes, the message passing process is divided into two parts: message passing between nodes within the batch (MP-IB) and message passing from nodes outside the batch to those within it (MP-OB). However, MP-OB recursively relies on higher-order out-of-batch neighbors, leading to an exponentially growing computational cost with respect to the number of layers. Due to the neighbor explosion, the whole message passing stores most nodes on the GPU such that many GNNs are infeasible to large-scale graphs. To address this challenge, we propose an accurate and fast mini-batch approach for large graph transductive learning, namely topological compensation (TOP), which obtains the outputs of the whole message passing solely through MP-IB, without the costly MP-OB. The major pillar of TOP is a novel concept of message invariance, which defines message-invariant transformations to convert costly MP-OB into fast MP-IB. This ensures that the modified MP-IB has the same output as the whole message passing. Experiments demonstrate that TOP is significantly faster than existing mini-batch methods by order of magnitude on vast graphs (millions of nodes and billions of edges) without accuracy degradation.", "keywords": [ "Subgraph Sampling", "Graph Neural Networks", "Topological Compensation", "Message Symmetries" ], "rating_list": [ 8, 8, 3, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "UqYNPyotxL", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Mahito Sugiyama", "gender": "Male", "institution": "National Institute of Informatics", "country": "JP", "position": "Associate Professor" }, { "name": "Ryuichi Kanoh", "gender": "Male", "institution": "NII, the Graduate University for Advanced Studies", "country": "", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 33, "n_ref": 86, "n_ref_all": 123, "n_fig": 10, "n_tab": 13, "L_tab": 14387, "n_element_tab": 924, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 669, "n_element_tab_1": 63, "formula_len_all": 1482, "formula_len_all_1": 1118, "len_all": 148256, "len_all_1": 65915, "len_abs": 1315, "len_title": 105, "len_sents": 37756, "len_sents_1": 28839, "n_sents": 306, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 57, "L_abs": 1319, "title": "Linear Mode Connectivity in Differentiable Tree Ensembles", "abs": "Linear Mode Connectivity (LMC) refers to the phenomenon that performance remains consistent for linearly interpolated models in the parameter space. For independently optimized model pairs from different random initializations, achieving LMC is considered crucial for understanding the stable success of the non-convex optimization in modern machine learning models and for facilitating practical parameter-based operations such as model merging. While LMC has been achieved for neural networks by considering the permutation invariance of neurons in each hidden layer, its attainment for other models remains an open question. In this paper, we first achieve LMC for soft tree ensembles, which are tree-based differentiable models extensively used in practice. We show the necessity of incorporating two invariances: subtree flip invariance and splitting order invariance, which do not exist in neural networks but are inherent to tree architectures, in addition to permutation invariance of trees. Moreover, we demonstrate that it is even possible to exclude such additional invariances while keeping LMC by designing decision list-based tree architectures, where such invariances do not exist by definition. Our findings indicate the significance of accounting for architecture-specific invariances in achieving LMC.", "keywords": [ "Linear Mode Connectivity", "Soft Tree" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 4, 4 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 2, 2, 2 ] }, { "paper_id": "UqR2dFmfRB", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Biao Fu", "gender": "Male", "institution": "Xiamen University", "country": "", "position": "PhD student" }, { "name": "Chengxi Li", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Kai Fan", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Liang Zhang", "gender": "Male", "institution": "XMU", "country": "CN", "position": "PhD student" }, { "name": "Minpeng Liao", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Xiaodong Shi", "gender": "Male", "institution": "Xiamen University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Yidong Chen", "gender": "Male", "institution": "Xiamen University", "country": "", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 28, "n_ref": 74, "n_ref_all": 103, "n_fig": 10, "n_tab": 5, "L_tab": 1630, "n_element_tab": 239, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1719, "n_element_tab_1": 196, "formula_len_all": 201, "formula_len_all_1": 161, "len_all": 178913, "len_all_1": 69955, "len_abs": 1486, "len_title": 136, "len_sents": 40997, "len_sents_1": 32277, "n_sents": 292, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1497, "title": "LLMs Can Achieve High-quality Simultaneous Machine Translation as Efficiently as Offline", "abs": "When the complete source sentence is provided, Large Language Models (LLMs) perform excellently in offline machine translation even with a simple prompt *\"Translate the following sentence from [src lang] into [tgt lang]:\"*. \nHowever, in many real scenarios, the source tokens arrive in a streaming manner and simultaneous machine translation (SiMT) is required, then the **efficiency** and **performance** of decoder-only LLMs are significantly limited by their auto-regressive nature. \nTo enable LLMs to achieve high-quality SiMT as efficiently as offline translation, we propose a novel paradigm that includes constructing supervised fine-tuning (SFT) data for SiMT, along with new training and inference strategies. \nTo replicate the token input/output (I/O) stream in SiMT, the source and target tokens are rearranged into an interleaved sequence, separated by special tokens according to varying latency requirements. \nThis enables powerful LLMs to learn read and write operations adaptively, based on varying latency prompts, while still maintaining efficient auto-regressive decoding. \nExperimental results demonstrate that, even with limited SFT data, our approach achieves state-of-the-art performance across various simultaneous translation benchmarks and different evaluation metrics, and preserves the original capabilities of offline translation.\nMoreover, EAST generalizes well to document-level SiMT without requiring specific fine-tuning, even beyond the offline translation model.", "keywords": [ "simultaneous machine translation", "machine translation", "Large Language Models", "adaptive policy", "supervised fine-tuning" ], "rating_list": [ 3, 3, 6, 5, 3 ], "soundness_list": [ 1, 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2, 1 ], "confidence_list": [ 4, 4, 5, 5, 4 ] }, { "paper_id": "Upoxh7wvmJ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Mohan Shankara Kankanhalli", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Ziwei Xu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" } ], "n_formula": 54, "n_formula_1": 41, "n_ref_uni": 26, "n_ref": 41, "n_ref_all": 59, "n_fig": 3, "n_tab": 1, "L_tab": 35, "n_element_tab": 6, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 62, "n_element_tab_1": 1, "formula_len_all": 5616, "formula_len_all_1": 4623, "len_all": 142971, "len_all_1": 70413, "len_abs": 1179, "len_title": 107, "len_sents": 32849, "len_sents_1": 22708, "n_sents": 330, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1188, "title": "Strong Preferences Affect the Robustness of Value Alignment", "abs": "Value alignment, which aims to ensure that large language models (LLMs) and other AI agents behave in accordance with human values, is critical for ensuring safety and trustworthiness of these systems. A key component of value alignment is the modeling of human preferences as a representation of human values. In this paper, we investigate the robustness of value alignment by examining the sensitivity of preference models. Specifically, we ask: how do changes in the probabilities of some preferences affect the predictions of these models for other preferences? To answer this question, we theoretically analyze the robustness of widely used preference models by examining their sensitivities to minor changes in preferences they model. Our findings reveal that, in the Bradley-Terry and the Placket-Luce model, the probability of a preference can change significantly as other preferences change, especially when these preferences are dominant (i.e., with probabilities near zero or one). We identify specific conditions where this sensitivity becomes significant for these models and discuss the practical implications for the robustness and safety of value alignment in AI systems.", "keywords": [ "Preference models", "robustness of value alignment", "sensitivity analysis", "AI safety" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 4, 4, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 2, 2 ] }, { "paper_id": "UpQLu9bzAR", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Han-Yuan Hsu", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "Undergrad student" }, { "name": "Jia-Fong Yeh", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "PhD student" }, { "name": "Kuo-Han Hung", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "Undergrad student" }, { "name": "Pang-Chi Lo", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "Undergrad student" }, { "name": "Yi-Ting Chen", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Assistant Professor" }, { "name": "Winston Hsu", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 10, "n_ref": 24, "n_ref_all": 53, "n_fig": 7, "n_tab": 10, "L_tab": 2548, "n_element_tab": 305, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 757, "n_element_tab_1": 119, "formula_len_all": 576, "formula_len_all_1": 469, "len_all": 165631, "len_all_1": 61482, "len_abs": 999, "len_title": 146, "len_sents": 44944, "len_sents_1": 28497, "n_sents": 303, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 98, "L_abs": 1003, "title": "VICtoR: Learning Hierarchical Vision-Instruction Correlation Rewards for Long-horizon Manipulation", "abs": "We study reward models for long-horizon manipulation by learning from action-free videos and language instructions, which we term the visual-instruction correlation (VIC) problem. Existing VIC methods face challenges in learning rewards for long-horizon tasks due to their lack of sub-stage awareness, difficulty in modeling task complexities, and inadequate object state estimation. To address these challenges,\nwe introduce VICtoR, a novel hierarchical VIC reward model capable of providing effective reward signals for long-horizon manipulation tasks. Trained solely on primitive motion demonstrations, VICtoR effectively provides precise reward signals for long-horizon tasks by assessing task progress at various stages using a novel stage detector and motion progress evaluator. We conducted extensive experiments in both simulated and real-world datasets. The results suggest that VICtoR outperformed the best existing methods, achieving a 43% improvement in success rates for long-horizon tasks.", "keywords": [ "reward learning", "reinforcement learning", "long-horizon robot learning", "vision-language" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "UoYxPYMUWd", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Ke Jiang", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" }, { "name": "Xiaoyang Tan", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "", "position": "Full Professor" } ], "n_formula": 46, "n_formula_1": 22, "n_ref_uni": 29, "n_ref": 96, "n_ref_all": 150, "n_fig": 7, "n_tab": 3, "L_tab": 1410, "n_element_tab": 89, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1255, "n_element_tab_1": 71, "formula_len_all": 4634, "formula_len_all_1": 1702, "len_all": 150398, "len_all_1": 68897, "len_abs": 1205, "len_title": 123, "len_sents": 45832, "len_sents_1": 31678, "n_sents": 334, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1216, "title": "Outcome-Driven Action Flexibility for Robust Offline Reinforcement Learning", "abs": "We address the challenge of offline reinforcement learning using realistic data, specifically non-expert data collected through sub-optimal behavior policies. A primary concern is that the learned policy must be conservative enough to manage \\textit{distribution shift} while maintaining sufficient flexibility for generalization. To tackle this issue, we introduce a novel method called Outcome-Driven Action Flexibility (ODAF), which seeks to reduce reliance on the empirical action distribution of the behavior policy.\nSpecifically, we develop a new reward mechanism that evaluates whether the subsequent states, following the current policy, meet specified performance requirements (e.g., safety\u2014remaining within the state support area), rather than solely depending on the characteristics of the actions taken (e.g., whether the action imitates the behavior policy).\nBesides theoretical justification, we provide empirical evidence on widely used D4RL benchmarks, demonstrating that our ODAF method, implemented using uncertainty quantification techniques, effectively tolerates unseen transitions for improved \"trajectory stitching,\" while enhancing the agent's ability to learn from realistic non-expert data.", "keywords": [ "Offline reinforcement learning; Robust reinforcement learning;Outcome-driven action flexibility;" ], "rating_list": [ 3, 6, 3, 3, 5 ], "soundness_list": [ 2, 4, 3, 2, 2 ], "presentation_list": [ 1, 3, 2, 1, 2 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 5, 3, 4 ] }, { "paper_id": "UoWslU6hsX", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Jose Hernandez-Orallo", "gender": "unknown", "institution": "Universitat Politecnica de Valencia", "country": "ES", "position": "Full Professor" }, { "name": "Lorenzo Pacchiardi", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "Lucy G Cheke", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 33, "n_ref": 68, "n_ref_all": 110, "n_fig": 11, "n_tab": 4, "L_tab": 4093, "n_element_tab": 789, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 198, "n_element_tab_1": 8, "formula_len_all": 240, "formula_len_all_1": 188, "len_all": 176429, "len_all_1": 65563, "len_abs": 1520, "len_title": 0, "len_sents": 55176, "len_sents_1": 31373, "n_sents": 320, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1546, "title": "100 instances is all you need: predicting LLM success by testing on a few instances", "abs": "Predicting if LLMs will succeed on individual task instances (i.e., prompts) is essential to ensure their reliability in high-stakes applications. To do so, we can evaluate a LLM on a set of instances and train an \"assessor\" to predict its performance. However, this requires evaluating each new LLM on sufficiently many instances. In this work, we build a \"generic assessor\" predicting the performance of any LLM on an instance by using the LLM's performance on a small set of reference instances and the features of the considered instance. In practice, we make use of existing evaluation results to extract the representative instances and train the assessor. Thus, the performance of a new LLM can be predicted by only testing it on the reference instances, leveraging the information contained in other LLMs' evaluations. We conduct empirical studies on HELM-Lite and KindsOfReasoning, a new collection of existing reasoning datasets that we introduce, where we evaluate all instruction-fine-tuned OpenAI models until $\\texttt{gpt4-0125-preview}$. We find that a few instances (around 100) are enough to achieve predictive power comparable to the LLM-specific assessors trained on the complete set of several thousand instances. Interestingly, randomly selecting the reference instances performs comparably to the advanced selection methods we tested. Finally, we identify a sharp drop in the predictive power of the generic and specific assessors in out-of-distribution scenarios, suggesting that the inherent predictability of LLMs is low.", "keywords": [ "AI evaluation", "Large Language Models", "Benchmarking" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "UoGv8d3MMy", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Guanhua Chen", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "", "position": "Associate Professor" }, { "name": "Jifan Gao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 19, "n_ref_uni": 34, "n_ref": 42, "n_ref_all": 52, "n_fig": 3, "n_tab": 4, "L_tab": 1498, "n_element_tab": 263, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1498, "n_element_tab_1": 263, "formula_len_all": 1141, "formula_len_all_1": 1141, "len_all": 89288, "len_all_1": 61558, "len_abs": 1345, "len_title": 97, "len_sents": 24716, "len_sents_1": 22853, "n_sents": 177, "n_sents_1": 169, "venue": "ICLR2025", "decision": "withdraw", "L_title": 48, "L_abs": 1328, "title": "MCCE: Missingness-aware Causal Concept Explainer", "abs": "Causal concept effect estimation is gaining increasing interest in the field of interpretable machine learning. This general approach explains the behaviors of machine learning models by estimating the causal effect of human-understandable concepts, which represent high-level knowledge more comprehensibly than raw inputs like tokens. However, existing causal concept effect explanation methods assume complete observation of all concepts involved within the dataset, which can fail in practice due to incomplete annotations or missing concept data. We theoretically demonstrate that unobserved concepts can bias the estimation of the causal effects of observed concepts. To address this limitation, we introduce the Missingness-aware Causal Concept Explainer (MCCE), a novel framework specifically designed to estimate causal concept effects when not all concepts are observable. Our framework learns to account for residual bias resulting from missing concepts and utilizes a linear predictor to model the relationships between these concepts and the outputs of black-box machine learning models. It can offer explanations on both local and global levels. We conduct validations using a real-world dataset, demonstrating that MCCE outperforms existing state-of-the-art explanation methods in causal concept effect estimation.", "keywords": [ "interpretable machine learning", "concept-based explanation", "causal concept effect" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2, 1 ], "presentation_list": [ 3, 2, 2, 1, 2 ], "contribution_list": [ 1, 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 5, 4, 3 ] }, { "paper_id": "Uo4EHT4ZZ8", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 6, "author_info_list": [ { "name": "Adarsh Kumarappan", "gender": "unknown", "institution": "California Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Mo Tiwari", "gender": "unknown", "institution": "OpenAI", "country": "US", "position": "Member of Technical Staff" }, { "name": "Peiyang Song", "gender": "Not Specified", "institution": "Stanford University", "country": "US", "position": "Researcher" }, { "name": "Robert Joseph George", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Anima Anandkumar", "gender": "Female", "institution": "California Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "chaowei xiao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 71, "n_ref": 108, "n_ref_all": 123, "n_fig": 2, "n_tab": 8, "L_tab": 5870, "n_element_tab": 488, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2085, "n_element_tab_1": 231, "formula_len_all": 42, "formula_len_all_1": 0, "len_all": 258491, "len_all_1": 69332, "len_abs": 1895, "len_title": 103, "len_sents": 88129, "len_sents_1": 31520, "n_sents": 727, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1903, "title": "LeanAgent: Lifelong Learning for Formal Theorem Proving", "abs": "Large Language Models (LLMs) have been successful in mathematical reasoning tasks such as formal theorem proving when integrated with interactive proof assistants like Lean. Existing approaches involve training or fine-tuning an LLM on a specific dataset to perform well on particular domains, such as undergraduate-level mathematics. These methods struggle with generalizability to advanced mathematics. A fundamental limitation is that these approaches operate on static domains, failing to capture how mathematicians often work across multiple domains and projects simultaneously or cyclically. We present LeanAgent, a novel lifelong learning framework for formal theorem proving that continuously generalizes to and improves on ever-expanding mathematical knowledge without forgetting previously learned knowledge. LeanAgent introduces several key innovations, including a curriculum learning strategy that optimizes the learning trajectory in terms of mathematical difficulty, a dynamic database for efficient management of evolving mathematical knowledge, and progressive training to balance stability and plasticity. LeanAgent successfully generates formal proofs for 155 theorems across 23 diverse Lean repositories where formal proofs were previously missing, many from advanced mathematics. It performs significantly better than the static LLM baseline, proving challenging theorems in domains like abstract algebra and algebraic topology while showcasing a clear progression of learning from basic concepts to advanced topics. In addition, we analyze LeanAgent's superior performance on key lifelong learning metrics. LeanAgent achieves exceptional scores in stability and backward transfer, where learning new tasks improves performance on previously learned tasks. This emphasizes LeanAgent's continuous generalizability and improvement, explaining its superior theorem-proving performance.", "keywords": [ "Theorem Proving", "Formal Theorem Proving", "Neural Theorem Proving", "Lifelong Learning", "Formal Mathematics", "Large Language Models", "LLMs", "Curriculum Learning", "Proof Search" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 4, 2 ], "confidence_list": [ 5, 4, 4, 2 ] }, { "paper_id": "UnpxRLMMAu", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Cristina Garbacea", "gender": "unknown", "institution": "University of Chicago", "country": "", "position": "Postdoc" }, { "name": "David Reber", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Todd Nief", "gender": "Male", "institution": "University of Chicago", "country": "", "position": "PhD student" }, { "name": "Victor Veitch", "gender": "unknown", "institution": "University of Chicago", "country": "", "position": "Assistant Professor" }, { "name": "Sean Richardson", "gender": "unknown", "institution": "University of Chicago", "country": "US", "position": "MS student" } ], "n_formula": 25, "n_formula_1": 14, "n_ref_uni": 22, "n_ref": 28, "n_ref_all": 54, "n_fig": 5, "n_tab": 6, "L_tab": 5469, "n_element_tab": 516, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2879, "formula_len_all_1": 1530, "len_all": 132733, "len_all_1": 58508, "len_abs": 1054, "len_title": 61, "len_sents": 38273, "len_sents_1": 26574, "n_sents": 324, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1082, "title": "RATE: Score Reward Models with Imperfect Rewrites of Rewrites", "abs": "This paper concerns the evaluation of reward models used in language modeling. A reward model is a function that takes a prompt and a response and assigns a score indicating how ``good'' that response is for the prompt. A key challenge is that reward models are usually imperfect proxies for actual preferences. For example, we may worry that a model trained to reward helpfulness learns to instead prefer longer responses. \n In this paper, we develop an evaluation method, RATE (Rewrite-based Attribute Treatment Estimators), that allows us to measure the \\emph{causal} effect of a given attribute of a response (e.g., length) on the reward assigned to that response. \n The core idea is to use large language models to rewrite responses to produce imperfect counterfactuals, and to adjust for rewriting error by rewriting \\emph{twice}. We show that the RATE estimator is consistent under reasonable assumptions. We demonstrate the effectiveness of RATE on synthetic and real-world data, showing that it can accurately estimate the effect of a given attribute on the reward model.", "keywords": [ "causality", "causal inference", "LLM alignment" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 1, 3 ], "confidence_list": [ 5, 4, 4, 2 ] }, { "paper_id": "UnodjDRqLp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 19, "author_info_list": [ { "name": "Bei Jiang", "gender": "Female", "institution": "University of Alberta", "country": "CA", "position": "Associate Professor" }, { "name": "Hanqi Jiang", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Jingyuan Chen", "gender": "Male", "institution": "Mayo Clinic", "country": "US", "position": "Postdoc" }, { "name": "Linglong Kong", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Full Professor" }, { "name": "Tianming Liu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Professor" }, { "name": "Tianyang Zhong", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "PhD student" }, { "name": "Tiezheng Zhang", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Tuo Zhang", "gender": "Male", "institution": "Northwestern Polytechnical University", "country": "CN", "position": "Associate Professor" }, { "name": "Wei Liu", "gender": "Male", "institution": "Mayo Clinic", "country": "US", "position": "Full Professor" }, { "name": "Xiaoming Zhai", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Associate Professor" }, { "name": "Xin Zhang", "gender": "unknown", "institution": "Northwestern Polytechnical University Xi'an", "country": "CN", "position": "Associate Professor" }, { "name": "Yaxuan Yang", "gender": "unknown", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Yi Pan", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Yifan Zhou", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Yutong Zhang", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "", "position": "MS student" }, { "name": "Zhengliang Liu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Zihao Wu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "huaqin zhao", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "BOLUN SUN", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 33, "n_ref": 51, "n_ref_all": 57, "n_fig": 2, "n_tab": 1, "L_tab": 48, "n_element_tab": 22, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 996, "n_element_tab_1": 70, "formula_len_all": 628, "formula_len_all_1": 628, "len_all": 113391, "len_all_1": 62538, "len_abs": 1421, "len_title": 130, "len_sents": 33702, "len_sents_1": 30121, "n_sents": 215, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1419, "title": "An Efficient LLM Alignment Framework for Automated Radiology Impression Generation", "abs": "Large language models (LLMs) are typically specialized for domain tasks through supervised fine-tuning, which optimizes LLMs for likelihood-based objectives. While supervised fine-tuning enables LLMs to generate text that conforms to the language style of a specific domain, such as radiology, it often falls short in enhancing the model's ability to perform detailed diagnostic reasoning or tailor reports for individual patients. In this paper, we explore the use of reinforcement learning to better align LLMs with the intricate requirements of radiological practice. By framing the report generation process as sequential decision-making stages, we present Radiology-Guided Reinforcement Optimization (RGRO), a tailored policy optimization framework designed specifically for medical language tasks. RGRO moves beyond conventional likelihood-based training by directly optimizing for radiology-specific objectives, including consistency with radiology findings and adherence to established professional guidelines. Our empirical evaluations demonstrate that RGRO significantly enhances the diagnostic precision and clinical utility of radiology reports generated by LLMs, outperforming supervised fine-tuning methods and state-of-the-art models. Furthermore, RGRO enables the seamless integration of expert radiologist feedback and external diagnostic tools, all without the need for large-scale annotated datasets.", "keywords": [ "Large language mode", "Radiology Impression Generation", "Alignment framework", "Reinforcement learning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "UnCKU8pZVe", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Cheng Sun", "gender": "unknown", "institution": "NVIDIA", "country": "TW", "position": "Researcher" }, { "name": "Chien-Yi Wang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Kai-Jie Lin", "gender": "Male", "institution": "National Chiao Tung University, National Chiao Tung University", "country": "", "position": "Undergrad student" }, { "name": "Ping-Chun Hsieh", "gender": "Male", "institution": "National Chiao Tung University", "country": "", "position": "Assistant Professor" }, { "name": "Yu Heng Hung", "gender": "Male", "institution": "National Yang-Ming Chiao Tung University", "country": "", "position": "PhD student" }, { "name": "YuHeng Lin", "gender": "Male", "institution": "National Chiao Tung University, National Chiao Tung University", "country": "TW", "position": "MS student" } ], "n_formula": 31, "n_formula_1": 22, "n_ref_uni": 70, "n_ref": 112, "n_ref_all": 141, "n_fig": 8, "n_tab": 6, "L_tab": 2342, "n_element_tab": 188, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 811, "n_element_tab_1": 71, "formula_len_all": 2217, "formula_len_all_1": 1266, "len_all": 195125, "len_all_1": 68468, "len_abs": 1032, "len_title": 134, "len_sents": 59586, "len_sents_1": 33462, "n_sents": 400, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1034, "title": "BOFormer: Learning to Solve Multi-Objective Bayesian Optimization via Non-Markovian RL", "abs": "Bayesian optimization (BO) offers an efficient pipeline for optimizing black-box functions with the help of a Gaussian process prior and an acquisition function (AF). Recently, in the context of single-objective BO, learning-based AFs witnessed promising empirical results given its favorable non-myopic nature. Despite this, the direct extension of these approaches to multi-objective Bayesian optimization (MOBO) suffer from the hypervolume identifiability issue, which results from the non-Markovian nature of MOBO problems. To tackle this, inspired by the non-Markovian RL literature and the success of Transformers in language modeling, we present a generalized deep Q-learning framework and propose BOFormer, which substantiates this framework for MOBO via sequence modeling. Through extensive evaluation, we demonstrate that BOFormer constantly achieves better performance than the benchmark rule-based and learning-based algorithms in various synthetic MOBO and real-world multi-objective hyperparameter optimization problems.", "keywords": [ "Multi-Objective Bayesian Optimization", "Transformers", "Hyperparameter Optimization", "Reinforcement Learning", "Acquisition Function" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "UmhC7fuhzs", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Antonio Torralba", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Yichen Li", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 35, "n_ref": 93, "n_ref_all": 115, "n_fig": 12, "n_tab": 7, "L_tab": 1101, "n_element_tab": 163, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1583, "n_element_tab_1": 53, "formula_len_all": 792, "formula_len_all_1": 584, "len_all": 185172, "len_all_1": 59875, "len_abs": 958, "len_title": 116, "len_sents": 51140, "len_sents_1": 27217, "n_sents": 411, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 961, "title": "Skin, Muscles, and Bones in MultiSensory Simulation", "abs": "General-purpose household robots require real-time fine motor control to handle delicate tasks and urgent situations. In this work, we introduce the senses of proprioception, kinesthesia, force haptics, and muscle activation to capture such precise control. This comprehensive set of multimodal senses naturally enables fine-grained interactions that are difficult to simulate with unimodal or text conditioned generative models. To effectively simulate fine-grained multisensory actions, we develop a feature learning paradigm that aligns these modalities while preserving the unique information each modality provides. We further regularize action trajectory features to enhance causality for representing intricate interaction dynamics. Experiments show that incorporating multimodal senses improves simulation accuracy and reduces temporal drift. Extensive ablation studies and downstream applications demonstrate effectiveness and practicality of our work.", "keywords": [ "multimodal learning; video simulators" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "UmdotAAVDe", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 11, "author_info_list": [ { "name": "Caleb Woo", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Undergrad student" }, { "name": "Carter Larsen", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Undergrad student" }, { "name": "Hao Tang", "gender": "Male", "institution": "Cornell University", "country": "", "position": "PhD student" }, { "name": "Kevin Ellis", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Assistant Professor" }, { "name": "Keya Hu", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "", "position": "Undergrad student" }, { "name": "Simon Alford", "gender": "Male", "institution": "Cornell University", "country": "", "position": "PhD student" }, { "name": "Spencer M. Dunn", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Undergrad student" }, { "name": "Wei-Long Zheng", "gender": "unknown", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Wen-Ding Li", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Yewen Pu", "gender": "Male", "institution": "Autodesk", "country": "", "position": "Researcher" }, { "name": "Yuqing Wu", "gender": "Female", "institution": "Cornell University", "country": "US", "position": "Undergrad student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 32, "n_ref": 55, "n_ref_all": 72, "n_fig": 10, "n_tab": 6, "L_tab": 14624, "n_element_tab": 205, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 2036, "n_element_tab_1": 94, "formula_len_all": 608, "formula_len_all_1": 611, "len_all": 213207, "len_all_1": 47618, "len_abs": 872, "len_title": 107, "len_sents": 48651, "len_sents_1": 20786, "n_sents": 376, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 889, "title": "Combining Induction and Transduction for Abstract Reasoning", "abs": "When learning an input-output mapping from very few examples, is it better to first infer a latent function that explains the examples, or is it better to directly predict new test outputs, e.g. using a neural network? We study this question on ARC by training neural models for \\emph{induction} (inferring latent functions) and \\emph{transduction} (directly predicting the test output for a given test input). We train \non synthetically generated variations of Python programs that solve ARC training tasks. We find inductive and transductive models solve different kinds of test problems, despite having the same training problems and sharing the same neural architecture: Inductive program synthesis excels at precise computations, and at composing multiple concepts, while transduction succeeds on fuzzier perceptual concepts. Ensembling them approaches human-level performance on ARC.", "keywords": [ "Abstract Reasoning", "Visual Reasoning", "Program Synthesis", "Induction", "Transduction" ], "rating_list": [ 8, 6, 3, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 1, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "UmMZC62SzZ", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Yifan Ran", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 46, "n_formula_1": 45, "n_ref_uni": 24, "n_ref": 34, "n_ref_all": 40, "n_fig": 4, "n_tab": 1, "L_tab": 346, "n_element_tab": 123, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4288, "formula_len_all_1": 4129, "len_all": 85433, "len_all_1": 48709, "len_abs": 618, "len_title": 47, "len_sents": 20267, "len_sents_1": 15663, "n_sents": 271, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 619, "title": "Accelerating semidefinite programming beyond limit: ADMM with tune-free operator stepsize", "abs": "In this work, we significantly alleviate the long-standing scalability issue of semidefinite programming (SDP), by equipping a novel tune-free operator stepsize to the alternating direction method of multipliers (ADMM) optimizer. To our best knowledge, this is the first operator stepsize in the context of SDP. More importantly, it is tune-free and computationally cheap (defined on dot product). Preliminary tests show that our operator ADMM surpasses the acceleration limit of the standard scalar version (limit found via grid search), i.e., our operator stepsize can outperform an arbitrarily fine tuned scalar one.", "keywords": [ "semidefinite programming", "scalability", "alternating direction method of multipliers", "tune-free operator stepsize" ], "rating_list": [ 3, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "UmMKbG2Ubr", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Aryan Mokhtari", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Devyani Maladkar", "gender": "Female", "institution": "University of Texas at Austin", "country": "US", "position": "MS student" }, { "name": "Ruichen Jiang", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" } ], "n_formula": 171, "n_formula_1": 42, "n_ref_uni": 42, "n_ref": 73, "n_ref_all": 95, "n_fig": 0, "n_tab": 2, "L_tab": 747, "n_element_tab": 188, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 16014, "formula_len_all_1": 2732, "len_all": 208039, "len_all_1": 73246, "len_abs": 1825, "len_title": 144, "len_sents": 61377, "len_sents_1": 31213, "n_sents": 641, "n_sents_1": 268, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1853, "title": "Convergence Analysis of Adaptive Gradient Methods under Refined Smoothness and Noise Assumptions", "abs": "Adaptive gradient methods, such as AdaGrad, are among the most successful optimization algorithms for neural network training. While these methods are known to achieve better dimensional dependence than stochastic gradient descent (SGD) under favorable geometry for stochastic convex optimization, the theoretical justification for their success in stochastic non-convex optimization remains elusive. In fact, under standard assumptions of Lipschitz gradients and bounded noise variance, it is known that SGD is worst-case optimal (up to absolute constants) in terms of finding a near-stationary point measured by the $\\ell_2$-norm, making further improvements impossible. Motivated by this limitation, we introduce refined assumptions on the smoothness structure of the objective and the gradient noise variance, which better suit the coordinate-wise nature of adaptive gradient methods. Moreover, we adopt the $\\ell_1$-norm of the gradient as the stationarity measure, as opposed to the standard $\\ell_2$-norm, to align with the coordinate-wise analysis and obtain tighter convergence guarantees for AdaGrad. Under these new assumptions and the $\\ell_1$-norm stationarity measure, we establish an **upper bound** on the convergence rate of AdaGrad and a corresponding **lower bound** for SGD. In particular, for certain configurations of problem parameters, we show that the iteration complexity of AdaGrad outperforms SGD by a factor of $d$. To the best of our knowledge, this is the first result to demonstrate a provable gain of adaptive gradient methods over SGD in a non-convex setting. We also present supporting lower bounds, including one specific to AdaGrad and one applicable to general deterministic first-order methods, showing that our upper bound for AdaGrad is tight and unimprovable up to a logarithmic factor under certain conditions.", "keywords": [ "Adaptive gradient methods", "stochastic nonconvex optimization", "dimensional dependence" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "UlsI4z3QQP", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Anbang Yao", "gender": "unknown", "institution": "Intel", "country": "", "position": "Principal Researcher" }, { "name": "Chao Li", "gender": "Male", "institution": "Intel", "country": "CN", "position": "Researcher" }, { "name": "Jiawei Fan", "gender": "Male", "institution": "Intel Labs China", "country": "CN", "position": "Research Intern" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 64, "n_ref_all": 80, "n_fig": 9, "n_tab": 8, "L_tab": 1981, "n_element_tab": 202, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 538, "n_element_tab_1": 24, "formula_len_all": 535, "formula_len_all_1": 355, "len_all": 143853, "len_all_1": 54653, "len_abs": 3941, "len_title": 117, "len_sents": 45527, "len_sents_1": 26192, "n_sents": 330, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1683, "title": "Morse: Fast Sampling for Accelerating Diffusion Models Universally", "abs": "In this paper, we present Morse, a simple and universal framework for accelerating diffusion models. The key insight of Morse is to reformulate the iterative generation (from noise to data) process via taking advantage of fast jump sampling and adaptive residual feedback strategies. Specifically, Morse involves two models called Dash and Dot that interact with each other. The Dash model is just the pre-trained diffusion model of any type, but operates in a jump sampling regime, creating sufficient space for sampling efficiency improvement. The Dot model is significantly faster than the Dash model, which is learnt to generate residual feedback conditioned on the observations at the current jump sampling point on the trajectory of the Dash model, lifting the noise estimate to easily match the next-step estimate of the Dash model without jump sampling. By chaining the outputs of the Dash and Dot models run in a time-interleaved fashion, Morse exhibits the merit of flexibly attaining desired image generation performance while improving overall runtime efficiency. With our proposed weight sharing strategy between the Dash and Dot models, Morse is efficient for training and inference. We validate the efficacy of our method under a variety of experimental setups. Our method shows an average speedup of 1.78\u00d7 to 3.31\u00d7 over a wide range of sampling step budgets relative to baseline diffusion models. Furthermore, we show that our method can be also generalized to improve the Latent Consistency Model (LCM-SDXL, which is already accelerated with consistency distillation technique) tailored for few-step text-to-image synthesis. The code will be made publicly available.", "keywords": [ "Diffusion models", "image generation", "text-to-image generation", "model acceleration" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "UlpkHciYQP", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Melis Ilayda Bal", "gender": "Female", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Michael Muehlebach", "gender": "unknown", "institution": "Max-Planck Institute", "country": "DE", "position": "Principal Researcher" }, { "name": "Volkan Cevher", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" } ], "n_formula": 69, "n_formula_1": 21, "n_ref_uni": 63, "n_ref": 126, "n_ref_all": 189, "n_fig": 23, "n_tab": 14, "L_tab": 6338, "n_element_tab": 620, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 680, "n_element_tab_1": 77, "formula_len_all": 5343, "formula_len_all_1": 1322, "len_all": 268584, "len_all_1": 71971, "len_abs": 1507, "len_title": 112, "len_sents": 82985, "len_sents_1": 34275, "n_sents": 775, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1549, "title": "Adversarial Training for Defense Against Label Poisoning Attacks", "abs": "As machine learning models advance in complexity and increasingly depend on large volumes of publicly sourced data, such as the human-annotated labels used in training large language models, they become more vulnerable to label poisoning attacks. These attacks, in which adversaries subtly alter the labels within a training dataset, can severely degrade model performance, posing significant risks in critical applications. In this paper, we propose $\\textbf{Floral}$, an adversarial training defense strategy based on support vector machines (SVMs) to counter label poisoning attacks. Utilizing a bilevel optimization framework, we cast the adversarial training process as a non-zero-sum Stackelberg game between an $\\textit{attacker}$, who strategically poisons critical training labels, and the $\\textit{model}$, which seeks to recover from such attacks. Our approach introduces a projected gradient descent algorithm with kernel SVMs for adversarial training. We provide a theoretical analysis of our algorithm\u2019s convergence properties and empirically evaluate its effectiveness across diverse classification tasks including sentiment analysis on the IMDB dataset. Compared to baseline robust models and foundation models such as RoBERTa, our method consistently achieves higher robust accuracy as the attacker\u2019s budget increases. These results underscore the potential of $\\textbf{Floral}$ to enhance the resilience of machine learning models against label poisoning threats, thereby ensuring robust classification in adversarial environments.", "keywords": [ "adversarial machine learning", "label poisoning attacks", "support vector machines", "adversarial training", "robust classification", "bilevel optimization", "projected gradient descent", "data poisoning", "Stackelberg game" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "UldnqRQWKS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Cheng Zhang", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Jianqiao Lu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jianyi Cheng", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Lecturer" }, { "name": "Pedro Gimenes", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Yiren Zhao", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Assistant Professor" }, { "name": "Zeyu Cao", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 37, "n_ref": 74, "n_ref_all": 101, "n_fig": 13, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1211, "n_element_tab_1": 17, "formula_len_all": 315, "formula_len_all_1": 217, "len_all": 140615, "len_all_1": 57665, "len_abs": 1264, "len_title": 108, "len_sents": 35875, "len_sents_1": 27968, "n_sents": 241, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1270, "title": "Scaling Laws For Mixed Quantization In Large Language Models", "abs": "Post-training quantization of Large Language Models (LLMs) has proven effective in reducing the computational requirements for running inference on these models. In this study, we focus on a straightforward question: When aiming for a specific accuracy or perplexity target for low-precision quantization, how many high-precision numbers or calculations are required to preserve as we scale LLMs to larger sizes? We first introduce a critical metric named the quantization ratio, which compares the number of parameters quantized to low-precision arithmetic against the total parameter count. Through extensive and carefully controlled experiments across different model families, arithmetic types, and quantization granularities (e.g. layer-wise, matmul-wise), we identify two central phenomenons. 1) The larger the models, the better they can preserve performance with an increased quantization ratio, as measured by perplexity in pre-training tasks or accuracy in downstream tasks. 2) The finer the granularity of mixed-precision quantization (e.g., matmul-wise), the more the model can increase the quantization ratio. \nWe believe these observed phenomena offer valuable insights for future AI hardware design and the development of advanced Efficient AI algorithms.", "keywords": [ "large language model", "scaling law", "quantisation" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 1, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "UlAkM88Vum", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Chia-Han Yeh", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "Ping-Chun Hsieh", "gender": "Male", "institution": "National Chiao Tung University", "country": "", "position": "Assistant Professor" }, { "name": "Shao-Hua Sun", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "Assistant Professor" }, { "name": "Tse-Sheng Nan", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Risto Ilkka Antero Vuorio", "gender": "unknown", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 27, "n_ref": 56, "n_ref_all": 67, "n_fig": 5, "n_tab": 7, "L_tab": 2925, "n_element_tab": 125, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2048, "n_element_tab_1": 45, "formula_len_all": 900, "formula_len_all_1": 605, "len_all": 111589, "len_all_1": 59181, "len_abs": 1230, "len_title": 85, "len_sents": 36705, "len_sents_1": 27058, "n_sents": 249, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 1244, "title": "Action-Constrained Imitation Learning", "abs": "Policy learning under action constraints plays a central role in ensuring safe behaviors in various robot control and resource allocation applications.\nIn this paper, we study a new problem setting termed Action-Constrained Imitation Learning (ACIL), where an action-constrained imitator aims to learn from a demonstrative expert with larger action space.\nThe fundamental challenge of ACIL lies in the unavoidable mismatch of occupancy measure between the expert and the imitator caused by the action constraints. We tackle this mismatch through $\\textit{trajectory alignment}$ and propose DTWIL, which replaces the original expert demonstrations with a surrogate dataset that follows similar state trajectories while adhering to the action constraints. Specifically, we recast trajectory alignment as a planning problem and solve it via Model Predictive Control, which aligns the surrogate trajectories with the expert trajectories based on the Dynamic Time Warping (DTW) distance. Through extensive experiments, we demonstrate that learning from the dataset generated by DTWIL significantly enhances performance across multiple robot control tasks and outperforms various benchmark imitation learning algorithms in terms of sample efficiency.", "keywords": [ "action-constrained reinforcement learning", "imitation learning", "safety" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "UkGrcekmSZ", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 1, "author_info_list": [ { "name": "Agniva Banerjee", "gender": "Male", "institution": "Indian Institute of Science Education and Research, Bhopal", "country": "IN", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 28, "n_ref": 41, "n_ref_all": 48, "n_fig": 4, "n_tab": 2, "L_tab": 379, "n_element_tab": 91, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 379, "n_element_tab_1": 91, "formula_len_all": 47, "formula_len_all_1": 47, "len_all": 94928, "len_all_1": 45920, "len_abs": 1609, "len_title": 153, "len_sents": 23350, "len_sents_1": 21540, "n_sents": 157, "n_sents_1": 140, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1626, "title": "Leveraging deep learning for comprehensive classification of renal diseases: A transfer learning approach", "abs": "The nightmare of cancer as a leading cause of premature deaths worldwide is becoming real and turns out to be one of the major problems of humanity nowadays. Cancer diagnostics at the early stage is Critical to cancer recovery and survival. In this context, renal diseases, including kidney cysts, stones, and tumors, pose significant global health challenges, affecting approximately 12\\% of the population and contributing to chronic kidney disease (CKD). Notably, renal cancer ranks as the tenth most prevalent cancer type, accounting for 2.7\\% of all cancer cases. This work presents a deep learning (DL) framework utilizing transfer learning (TL) for the early detection of renal diseases and categorizing the conditions into four binary classifications: Cyst\\_vs\\_Normal, Cyst\\_vs\\_Stone, Cyst\\_vs\\_Tumor, and Stone\\_vs\\_Tumor, allowing for a more specific understanding of each stage. By analyzing CT scans and microscopic histopathology images, the framework employs convolutional neural networks (CNNs) with pre-trained models to facilitate automatic and precise classification of renal conditions. Specifically, two CNN models ResNet-50 and EfficientNetV2 are implemented, providing a comprehensive analysis of each stage of the DL architecture. Comparative evaluations of training outcomes across various datasets revealed that EfficientNetV2 performed marginally better than ResNet-50, achieving an impressive testing accuracy of up to 100\\% for all cases. These results underscore the effectiveness of the DL-based system and highlight its potential for widespread clinical application in renal disease diagnosis.", "keywords": [ "CNN", "Kidney", "image classification", "deep learning", "transfer learning" ], "rating_list": [ 1, 1, 3, 3 ], "soundness_list": [ 1, 1, 1, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 1, 1, 1, 2 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "UkEvpOzZAR", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Joanna Waczynska", "gender": "Female", "institution": "Jagiellonian University Cracow", "country": "PL", "position": "PhD student" }, { "name": "Piotr Borycki", "gender": "Male", "institution": "Jagiellonian University Cracow", "country": "PL", "position": "MS student" }, { "name": "Przemys\u0142aw Spurek", "gender": "Male", "institution": "Jagiellonian University Cracow", "country": "PL", "position": "Associate Professor" }, { "name": "Slawomir Tadeja", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "Thomas Bohn\u00e9", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Principal Researcher" }, { "name": "Tomasz Szczepanik", "gender": "Male", "institution": "Jagiellonian University Cracow", "country": "PL", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 30, "n_ref": 52, "n_ref_all": 113, "n_fig": 22, "n_tab": 4, "L_tab": 2183, "n_element_tab": 254, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 348, "n_element_tab_1": 49, "formula_len_all": 928, "formula_len_all_1": 828, "len_all": 134546, "len_all_1": 56738, "len_abs": 1277, "len_title": 99, "len_sents": 39344, "len_sents_1": 24514, "n_sents": 319, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1299, "title": "MiraGe: Editable 2D Images using Gaussian Splatting", "abs": "Implicit Neural Representations (INRs) approximate discrete data through continuous functions and are commonly used for encoding 2D images. Traditional image-based INRs employ neural networks to map pixel coordinates to RGB values, capturing shapes, colors, and textures within the network\u2019s weights. Recently, 2D Gaussian Splatting (GS) has been proposed as an alternative, using Gaussian functions instead of neural networks to achieve comparable quality and compression. Such a solution obtains a quality and compression ratio similar to classical INR models but does not allow image modification. In contrast, our work introduces a novel method, MiraGe, which uses mirror reflections to perceive 2D images in 3D space and employs flat-controlled Gaussians for precise 2D image editing. Our approach improves the rendering quality and allows realistic image modifications, including human-inspired perception of photos in the 3D world. Thanks to modeling images in 3D space, we obtain the illusion of 3D-based modification in 2D images. We also show that our Gaussian representation can be easily combined with a physics engine to produce physics-based modification of 2D images. Consequently, MiraGe allows for better quality than the standard approach and for natural modification of 2D images.", "keywords": [ "Gaussian Splatting", "implicit reprezentation. physics" ], "rating_list": [ 5, 5, 8, 6 ], "soundness_list": [ 2, 2, 4, 3 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 2, 2, 4, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "UjSmUlUU6y", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Chao Du", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Min Lin", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Principal Researcher" }, { "name": "Tianyu Pang", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Wei Gao", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Associate Professor" }, { "name": "cunxiao du", "gender": "Unspecified", "institution": "Sea AI LAB", "country": "SG", "position": "Researcher" }, { "name": "ZHANG Xuan", "gender": "Female", "institution": "Singapore Management University", "country": "SG", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 51, "n_ref": 129, "n_ref_all": 166, "n_fig": 11, "n_tab": 9, "L_tab": 4026, "n_element_tab": 431, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1795, "n_element_tab_1": 162, "formula_len_all": 480, "formula_len_all_1": 480, "len_all": 161786, "len_all_1": 66465, "len_abs": 1339, "len_title": 113, "len_sents": 41586, "len_sents_1": 30731, "n_sents": 292, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1337, "title": "SimLayerKV: A Simple Framework for Layer-Level KV Cache Reduction", "abs": "Recent advancements in large language models (LLMs) have extended their capabilities to handle long contexts. However, increasing the number of model layers and the length of input sequences significantly escalates the memory required to store key-value (KV) cache, posing challenges for efficient inference. To mitigate this issue, we present SimLayerKV, a simple yet effective method that reduces inter-layer KV cache redundancies by selectively dropping cache in identified lazy layers. Our approach is based on the observation that certain layers in long-context LLMs exhibit \"lazy\" behavior, contributing less to modeling long-range dependencies compared to non-lazy layers. By analyzing attention weight patterns, we find that the behavior of these lazy layers is consistent across tokens for a given input. This insight motivates our SimLayerKV, which identifies lazy layers and reduces their KV cache accordingly. SimLayerKV is training-free, generalizable, and can be implemented with only seven lines of code. We conduct extensive experiments on three representative LLMs, e.g., LLaMA2-7B, LLaMA3-8B, and Mistral-7B across 16 tasks from the LongBench benchmark. The results demonstrate that SimLayerKV achieves a KV cache compression ratio of 5$\\\\times$ with only a 1.2\\% performance drop when combined with 4-bit quantization.", "keywords": [ "Large Language Models", "Long-Context", "KV Cache" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "UjQthmslFV", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Akil Narayan", "gender": "unknown", "institution": "University of Utah", "country": "", "position": "Associate Professor" }, { "name": "John Davis Jakeman", "gender": "Male", "institution": "Sandia National Laboratories", "country": "", "position": "Researcher" }, { "name": "John Turnage", "gender": "Male", "institution": "University of Utah", "country": "US", "position": "PhD student" }, { "name": "Matthew Lowery", "gender": "Male", "institution": "University of Utah", "country": "US", "position": "PhD student" }, { "name": "Shandian Zhe", "gender": "Male", "institution": ", University of Utah", "country": "US", "position": "Associate Professor" }, { "name": "Varun Shankar", "gender": "Male", "institution": "University of Utah", "country": "", "position": "Assistant Professor" }, { "name": "Zachary Morrow", "gender": "unknown", "institution": "Sandia National Laboratories", "country": "US", "position": "Researcher" } ], "n_formula": 26, "n_formula_1": 24, "n_ref_uni": 34, "n_ref": 84, "n_ref_all": 126, "n_fig": 7, "n_tab": 4, "L_tab": 1305, "n_element_tab": 142, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1084, "n_element_tab_1": 98, "formula_len_all": 1939, "formula_len_all_1": 1746, "len_all": 161964, "len_all_1": 65305, "len_abs": 1128, "len_title": 87, "len_sents": 54636, "len_sents_1": 27930, "n_sents": 372, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 1131, "title": "Kernel Neural Operators (KNOs) for Scalable, Memory-efficient, Geometrically-flexible Operator Learning", "abs": "This paper introduces the Kernel Neural Operator (KNO), a novel operator learning technique that uses deep kernel-based integral operators in conjunction with quadrature for function-space approximation of operators (maps from functions to functions). KNOs use parameterized, closed-form, finitely-smooth, and compactly-supported kernels with trainable sparsity parameters within the integral operators to significantly reduce the number of parameters that must be learned relative to existing neural operators. Moreover, the use of quadrature for numerical integration endows the KNO with geometric flexibility that enables operator learning on irregular geometries. Numerical results demonstrate that on existing benchmarks the training and test accuracy of KNOs is higher than popular operator learning techniques while using at least an order of magnitude fewer trainable parameters. KNOs thus represent a new paradigm of low-memory, geometrically-flexible, deep operator learning, while retaining the implementation simplicity and transparency of traditional kernel methods from both scientific computing and machine learning.", "keywords": [ "operator learning", "scientific machine learning", "neural operator", "kernel" ], "rating_list": [ 5, 8, 5, 1 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "Uj0h13lVrR", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Johan W. Verjans", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Associate Professor" }, { "name": "Nadhir Hassen", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "PhD student" }, { "name": "Zhibin Liao", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Associate Professor" } ], "n_formula": 16, "n_formula_1": 17, "n_ref_uni": 20, "n_ref": 35, "n_ref_all": 46, "n_fig": 4, "n_tab": 2, "L_tab": 163, "n_element_tab": 5, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1141, "formula_len_all_1": 1247, "len_all": 82070, "len_all_1": 53265, "len_abs": 542, "len_title": 129, "len_sents": 29370, "len_sents_1": 24683, "n_sents": 218, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 542, "title": "KL DIVERGENCE OPTIMIZATION WITH ENTROPY- RATIO ESTIMATION FOR STOCHASTIC GFLOWNETS", "abs": "This paper introduces a novel approach for optimizing Generative Flow Networks (GFlowNets) in stochastic environments by incorporating KL divergence objectives with entropy-ratio estimation. We leverage the relationship between high and low entropy states, as defined in entropy-regularized Markov Decision Processes (MDPs), to dynamically adjust exploration and exploitation. Detailed proofs and analysis demonstrate the efficacy of this methodology in enhancing mode discovery, state coverage, and policy robustness in complex environments.", "keywords": [ "Stochastic environments", "MDP", "GFlowNets", "KL divergence", "Molecule Generation" ], "rating_list": [ 1, 1 ], "soundness_list": [ 1, 1 ], "presentation_list": [ 1, 1 ], "contribution_list": [ 1, 1 ], "confidence_list": [ 4, 3 ] }, { "paper_id": "UiLtbLsiPU", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Atia Hamidizadeh", "gender": "Female", "institution": "Simon Fraser University", "country": "", "position": "MS student" }, { "name": "David Gamaliel Arcos Bravo", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CA", "position": "Intern" }, { "name": "Hongjian Gu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Junyi Dong", "gender": "Female", "institution": "Cornell University", "country": "", "position": "PhD student" }, { "name": "Lingfeng Zhang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CA", "position": "Researcher" }, { "name": "Shunbo Zhou", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Tongtong Cao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yingxue Zhang", "gender": "Female", "institution": "Huawei Canada, Huawei Noah's Ark Lab", "country": "CA", "position": "Researcher" }, { "name": "Yuecheng Liu", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Yuening Wang", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Yutong Wang", "gender": "Female", "institution": "nanjing university", "country": "", "position": "MS student" }, { "name": "Yuzheng Zhuang", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Research Engineer" }, { "name": "Zhanguang Zhang", "gender": "unknown", "institution": "Huawei Noah's Ark Lab", "country": "CA", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 81, "n_ref_all": 124, "n_fig": 14, "n_tab": 16, "L_tab": 5437, "n_element_tab": 514, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 4677, "n_element_tab_1": 585, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 217663, "len_all_1": 66875, "len_abs": 1597, "len_title": 160, "len_sents": 82742, "len_sents_1": 28404, "n_sents": 587, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1601, "title": "ET-Plan-Bench: Embodied Task-level Planning Benchmark Towards Spatial-Temporal Cognition with Foundation Models", "abs": "Recent advancements in Large Language Models (LLMs) have spurred numerous attempts to apply these technologies to embodied tasks, particularly focusing on high-level task planning and task decomposition. To further explore this area, we introduce a new embodied task planning benchmark, ET-Plan-Bench, which specifically targets embodied task planning using LLMs. It features a controllable and diverse set of embodied tasks varying in different levels of difficulties and complexities, and is designed to evaluate two critical dimensions of LLMs' application in embodied task understanding: spatial (relation constraint, occlusion for target objects) and temporal \\& causal understanding of the sequence of actions in the environment. By using multi-source simulators as the backend simulator, it can provide immediate environment feedback to LLMs, which enables LLMs to interact dynamically with the environment and re-plan as necessary. We evaluated the state-of-the-art open source and closed source foundation models, including GPT-4, LLAMA and Mistral on our proposed benchmark. While they perform adequately well on simple navigation tasks, their performance can significantly deteriorate when faced with tasks that require a deeper understanding of spatial, temporal, and causal relationships. Thus, our benchmark distinguishes itself as a large-scale, quantifiable, highly automated, and fine-grained diagnostic framework that presents a significant challenge to the latest foundation models. We hope it can spark and drive further research in embodied task planning using foundation models.", "keywords": [ "Embodied task planning benchmark" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "UiEjzBRYeI", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Lingxi Xie", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Pengfei Chen", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Qi Tian", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "XIAOPENG ZHANG", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Xinyue Huo", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yingfei Sun", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhenjun Han", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "hui ying", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 31, "n_ref": 95, "n_ref_all": 131, "n_fig": 16, "n_tab": 11, "L_tab": 3854, "n_element_tab": 635, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3558, "n_element_tab_1": 577, "formula_len_all": 744, "formula_len_all_1": 413, "len_all": 187424, "len_all_1": 81546, "len_abs": 1247, "len_title": 119, "len_sents": 52236, "len_sents_1": 32748, "n_sents": 457, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1248, "title": "SAM-CP: Marrying SAM with Composable Prompts for Versatile Segmentation", "abs": "The Segment Anything model (SAM) has shown a generalized ability to group image pixels into patches, but applying it to semantic-aware segmentation still faces major challenges. This paper presents SAM-CP, a simple approach that establishes two types of composable prompts beyond SAM and composes them for versatile segmentation. Specifically, given a set of classes (in texts) and a set of SAM patches, the Type-I prompt judges whether a SAM patch aligns with a text label, and the Type-II prompt judges whether two SAM patches with the same text label also belong to the same instance. To decrease the complexity in dealing with a large number of semantic classes and patches, we establish a unified framework that calculates the affinity between (semantic and instance) queries and SAM patches, and then merges patches with high affinity to the query. Experiments show that SAM-CP achieves semantic, instance, and panoptic segmentation in both open and closed domains. In particular, it achieves state-of-the-art performance in open-vocabulary segmentation. Our research offers a novel and generalized methodology for equipping vision foundation models like SAM with multi-grained semantic perception abilities. Codes will be publicly available.", "keywords": [ "Versatile Segmentation", "Composable Prompts", "Open-vocabulary" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "Uhvy5u90UY", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Linghe Kong", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Rui Kong", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yuanchun Li", "gender": "Male", "institution": "Institute for AI Industry Research, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yunxin Liu", "gender": "unknown", "institution": "Institute for AI Industry Research, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "qingtian feng", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 41, "n_ref_all": 48, "n_fig": 0, "n_tab": 3, "L_tab": 3060, "n_element_tab": 453, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 3060, "n_element_tab_1": 453, "formula_len_all": 164, "formula_len_all_1": 164, "len_all": 100204, "len_all_1": 77578, "len_abs": 1201, "len_title": 105, "len_sents": 28981, "len_sents_1": 28389, "n_sents": 203, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1206, "title": "Preserving Large Activations: The Key to KV Cache Pruning", "abs": "As context lenghts grows, the increasing size of Key and Value (KV) cache poses a significant challenge to efficiently serving Large Language Models (LLMs). KV cache pruning, by preserving only a small subset of important KV cache for sparse inference, is a recognized effective solution. Our research revealed that large activations are the key to identifying these important KV cache. However, existing methods have not been successful in effectively identifying these important KV cache due to neglecting the impact of Value cache, and are also incompatible with Grouped-Query Attention (GQA) architectures. To address these issues, we introduce an innovative KV cache pruning method that preserves these large activations and is compatible with Grouped-Query Attention. Featuring a novel pruning metric, this method operates within each attention group to enhance efficiency and minimize performance degradation. Experimental results demonstrate that our approach not only maintains comparable accuracy with existing methods but also significantly reduces KV cache requirements. Specifically, It demonstrates similar accuracy while utilizing only 1/10 of the KV cache compared to existing SOTA methods.", "keywords": [ "Large Language Models", "Efficiency", "Compression", "Long Context" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "Uhj5OxAz7I", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Jianfeng Gao", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Mu Cai", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin, Madison", "country": "US", "position": "PhD student" }, { "name": "Jianwei Yang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Yong Jae Lee", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 49, "n_ref": 115, "n_ref_all": 144, "n_fig": 4, "n_tab": 15, "L_tab": 5451, "n_element_tab": 878, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2636, "n_element_tab_1": 453, "formula_len_all": 306, "formula_len_all_1": 272, "len_all": 184091, "len_all_1": 78848, "len_abs": 1568, "len_title": 76, "len_sents": 43635, "len_sents_1": 33280, "n_sents": 331, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 28, "L_abs": 1561, "title": "Matryoshka Multimodal Models", "abs": "Large Multimodal Models (LMMs) such as LLaVA have shown strong performance in visual-linguistic reasoning. These models first embed images into a fixed large number of visual tokens and then feed them into a Large Language Model (LLM). However, this design causes an excessive number of tokens for dense visual scenarios such as high-resolution images and videos, leading to great inefficiency. While token pruning/merging methods do exist, they produce a single length output for each image and do not afford flexibility in trading off information density v.s. efficiency. Inspired by the concept of Matryoshka Dolls, we propose \n: Matryoshka Multimodal Models, which learns to represent visual content as nested sets of visual tokens that capture information across multiple coarse-to-fine granularities. Our approach offers several unique benefits for LMMs: (1) One can explicitly control the visual granularity per test instance during inference, e.g. , adjusting the number of tokens used to represent an image based on the anticipated complexity or simplicity of the content; (2) \n provides a framework for analyzing the granularity needed for existing datasets, where we find that COCO-style benchmarks only need around \n9 visual tokens to obtain accuracy similar to that of using all 576 tokens; (3) Our approach provides a foundation to explore the best trade-off between performance and visual token length at sample level, where our investigation reveals that a large gap exists between the oracle upper bound and current fixed-scale representations.", "keywords": [ "Multimodal Model", "matryoshka" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "UhdmcuuvSt", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Bei Jiang", "gender": "Female", "institution": "University of Alberta", "country": "CA", "position": "Associate Professor" }, { "name": "Junxi Zhang", "gender": "Male", "institution": "Concordia University", "country": "CA", "position": "Assistant Professor" }, { "name": "Linglong Kong", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Full Professor" }, { "name": "Yangdi Jiang", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "PhD student" } ], "n_formula": 56, "n_formula_1": 26, "n_ref_uni": 38, "n_ref": 66, "n_ref_all": 87, "n_fig": 6, "n_tab": 1, "L_tab": 69, "n_element_tab": 12, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4948, "formula_len_all_1": 1929, "len_all": 144175, "len_all_1": 68142, "len_abs": 1512, "len_title": 141, "len_sents": 49169, "len_sents_1": 33623, "n_sents": 334, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1540, "title": "FDA: Generating Fair Synthetic Data with Provable Trade-off between Fairness and Faithfulness", "abs": "We propose a novel framework called FDA for generating Fair synthetic data through Data Augmentation, offering the first method with provable trade-off guarantee between fairness and faithfulness. Unlike other existing methods, our approach utilizes a novel joint model that consists of two sub-models: one focused on enforcing strict fairness constraints while the other dedicated to preserving fidelity to the original data, coupled with a tuning mechanism that provides explicit control over the trade-off between fairness and faithfulness. Specifically, our FDA framework enables explicit quantification of the extent to which the generated fair synthetic data preserve faithfulness to the original data, while achieving an intermediate level of fairness determined by a user specified parameter $\\alpha \\in [0, 1]$. Theoretically, we show that the resulting fair synthetic data converge to the original data in probability when $\\alpha$ tends to 1, thereby implying convergence in distribution. Our framework can be also combined with some GAN-based fair models, such as DECAF, to further improve the utility of the resulting synthetic data in downstream analysis, while carefully balancing fairness. Furthermore, we obtain an upper bound of the unfairness measurement for downstream models trained on the generated fair synthetic data, which can help users to choose appropriate $\\alpha$. Finally, we perform numerical experiments on benchmark data to validate our theoretical contributions and to compare our FDA with other methods.", "keywords": [ "Fairness", "Fair synthetic data generation", "Joint modeling", "Faithfulness", "Trade-off between fairness and faithfulness" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 1, 2, 1, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "UhW2wA1pRV", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Jun Sakuma", "gender": "Male", "institution": "Institute of Science Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Kazuto Fukuchi", "gender": "Male", "institution": "University of Tsukuba", "country": "JP", "position": "Assistant Professor" }, { "name": "Shojiro Yamabe", "gender": "Male", "institution": "Institute of Science Tokyo", "country": "JP", "position": "MS student" } ], "n_formula": 70, "n_formula_1": 22, "n_ref_uni": 34, "n_ref": 84, "n_ref_all": 94, "n_fig": 5, "n_tab": 5, "L_tab": 3349, "n_element_tab": 277, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 338, "n_element_tab_1": 15, "formula_len_all": 6870, "formula_len_all_1": 1316, "len_all": 192095, "len_all_1": 65388, "len_abs": 1559, "len_title": 124, "len_sents": 60310, "len_sents_1": 30724, "n_sents": 529, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1380, "title": "Robust Deep Reinforcement Learning against ADVERSARIAL BEHAVIOR MANIPULATION", "abs": "This study investigates the robustness of deep reinforcement learning agents against targeted attacks that aim to manipulate the victim's behavior through adversarial interventions in state observations. While several methods for such targeted manipulation attacks have been proposed, they all require white-box access to the victim's policy, and some rely on environment-specific heuristics. Furthermore, no defense method has been proposed to counter these attacks. To address this, we propose a novel targeted attack method for manipulating the victim, which does not depend on environmental heuristics and applies in black-box and no-box settings. Additionally, we introduce a defense strategy against these attacks. Our theoretical analysis proves that the sensitivity of a policy's action output to state changes affects the defense performance and that the earlier in the trajectory, the greater the effect. Based on this insight, we introduce a time-discounted regularization as a countermeasure for such behavior targeted attacks, which helps to improve robustness against attacks while maintaining task performance in the absence of attacks. Empirical evaluations demonstrate that our proposed attack method outperforms baseline attack methods. Furthermore, our defense strategy shows superior robustness against existing defense methods designed for untargeted attacks.", "keywords": [ "reinforcement learning", "robustness", "adversarial attacks", "adversarial defense", "imitation learning" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "UhLLqUVn4X", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Di Wu", "gender": "Male", "institution": "Southwest University", "country": "CN", "position": "Full Professor" }, { "name": "Wu Sun", "gender": "Male", "institution": "Chongqing University of Post and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Xin Luo", "gender": "Male", "institution": "Southwest University", "country": "CN", "position": "Full Professor" }, { "name": "Yi He", "gender": "Male", "institution": "Integrative Conservation Clinic", "country": "US", "position": "Intern" }, { "name": "Zhong Chen", "gender": "Male", "institution": "Southern Illinois University-Carbondale", "country": "US", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 39, "n_ref": 108, "n_ref_all": 129, "n_fig": 6, "n_tab": 7, "L_tab": 4612, "n_element_tab": 282, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 3587, "n_element_tab_1": 217, "formula_len_all": 741, "formula_len_all_1": 682, "len_all": 165640, "len_all_1": 64900, "len_abs": 1456, "len_title": 157, "len_sents": 50426, "len_sents_1": 26988, "n_sents": 397, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1444, "title": "DHENN: A Deeper Hybrid End-to-end Neural Network for Highly Accurate Drug-Drug Interaction Events Prediction", "abs": "Accurate prediction of drug-drug interactions (DDIs) is crucial for therapeutic safety yet poses a substantial challenge due to complex pharmacodynamics. Traditional DDI prediction methods often falter for three reasons. First, they simplify dependency structures among entities (e.g., drugs, targets, enzymes, and transporters) in bipartite networks, falling short in modeling their high-order interactions. Second, the over-smoothing effects constrain the depth of the adopted neural networks, thereby limiting their learning capacity. Third, they mainly decouple the stages of representation and prediction, leading to suboptimal solutions that overlook the potential for ground-truth DDIs to refine embedding generation. In response, this paper proposes Deeper Hybrid End-to-end Neural Network (DHENN), which integrates a Multimodal Knowledge Graph (MKG) with a Prediction-Enhanced Cascading Network (PECN) in an end-to-end learning manner. Specifically, MKG captures higher-order relationships across entities, offering a holistic view of DDIs. PECN mitigates over-smoothing by incorporating shallow embeddings into deeper layers, preserving node-level diversity. The endto- end learning manner guarantees that the representation learning and predictive modeling of MKG and PECN are formulated into a unified learning objective. Extensive experiments substantiate that DHENN outperforms thirteen competitors on two real-world DDI datasets.", "keywords": [ "Drug-drug interactions", "Knowledge graph", "Graph neural networks", "Deep neural networks" ], "rating_list": [ 3, 8, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "UhKkWHkvfg", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Hongxin Wei", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Yicheng Xu", "gender": "Male", "institution": "Institute of Tokyo Science", "country": "JP", "position": "PhD student" }, { "name": "Yufei Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "MS student" }, { "name": "Zhiping Lin", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "HUIPING ZHUANG", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 26, "n_formula_1": 21, "n_ref_uni": 27, "n_ref": 76, "n_ref_all": 108, "n_fig": 13, "n_tab": 17, "L_tab": 8261, "n_element_tab": 998, "n_fig_1": 8, "n_tab_1": 9, "L_tab_1": 5389, "n_element_tab_1": 516, "formula_len_all": 1929, "formula_len_all_1": 1238, "len_all": 150672, "len_all_1": 74737, "len_abs": 1320, "len_title": 117, "len_sents": 39023, "len_sents_1": 27159, "n_sents": 374, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1359, "title": "Analytic Continual Test-Time Adaptation for Multi-Modality Corruption", "abs": "Test-Time Adaptation (TTA) aims to help pre-trained model bridge the gap between source and target datasets using only the pre-trained model and unlabelled test data. A key objective of TTA is to address domain shifts in test data caused by corruption, such as weather changes, noise, or sensor malfunctions. Multi-Modal Continual Test-Time Adaptation (MM-CTTA), an extension of TTA with better real-world applications, further allows pre-trained models to handle multi-modal inputs and adapt to continuously-changing target domains. MM-CTTA typically faces challenges including $\\textbf{error accumulation}$, $\\textbf{catastrophic forgetting}$, and $\\textbf{reliability bias}$, with few existing approaches effectively addressing these issues in multi-modal corruption scenarios. In this paper, we propose a novel approach, Multi-modality Dynamic Analytic Adapter (MDAA), for MM-CTTA tasks. We innovatively introduce analytic learning into TTA, using the Analytic Classifiers (ACs) to prevent model forgetting. Additionally, we develop Dynamic Selection Mechanism (DSM) and Soft Pseudo-label Strategy (SPS), which enable MDAA to dynamically filter reliable samples and integrate information from different modalities. Extensive experiments demonstrate that MDAA achieves state-of-the-art performance on MM-CTTA tasks while ensuring reliable model adaptation.", "keywords": [ "test time adaptation; multi-modality; continual learning; analytic learning" ], "rating_list": [ 6, 6, 5, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Uh5GRmLlvt", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Bernd Frauenknecht", "gender": "unknown", "institution": "Rheinisch Westf\u00e4lische Technische Hochschule Aachen", "country": "", "position": "PhD student" }, { "name": "Devdutt Subhasish", "gender": "Male", "institution": "Rheinisch Westf\u00e4lische Technische Hochschule Aachen", "country": "DE", "position": "MS student" }, { "name": "Friedrich Solowjow", "gender": "unknown", "institution": "Rheinisch Westf\u00e4lische Technische Hochschule Aachen", "country": "", "position": "Lecturer" }, { "name": "Sebastian Trimpe", "gender": "Male", "institution": "RWTH Aachen University", "country": "DE", "position": "Full Professor" } ], "n_formula": 108, "n_formula_1": 25, "n_ref_uni": 2, "n_ref": 110, "n_ref_all": 175, "n_fig": 14, "n_tab": 14, "L_tab": 2276, "n_element_tab": 217, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 120, "n_element_tab_1": 7, "formula_len_all": 8416, "formula_len_all_1": 1709, "len_all": 370672, "len_all_1": 58972, "len_abs": 947, "len_title": 97, "len_sents": 86959, "len_sents_1": 25233, "n_sents": 760, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 950, "title": "On Rollouts in Model-Based Reinforcement Learning", "abs": "Model-based reinforcement learning (MBRL) seeks to enhance data efficiency by learning a model of the environment and generating synthetic rollouts from it. However, accumulated model errors during these rollouts can distort the data distribution, negatively impacting policy learning and hindering long-term planning. Thus, the accumulation of model errors is a key bottleneck in current MBRL methods. We propose Infoprop, a model-based rollout mechanism that separates aleatoric from epistemic model uncertainty and reduces the influence of the latter on the data distribution. Further, Infoprop keeps track of accumulated model errors along a model rollout and provides termination criteria to limit data corruption. We demonstrate the capabilities of Infoprop in the Infoprop-Dyna algorithm, reporting state-of-the-art performance in Dyna-style MBRL on common MuJoCo benchmark tasks while substantially increasing rollout length and data quality.", "keywords": [ "Model-Based Reinforcement Learning", "Model Rollouts", "Uncertainty Quantification" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 4, 4, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "Ugs2W5XFFo", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Alessandro Finamore", "gender": "Male", "institution": "Polytechnic Institute of Turin", "country": "", "position": "PhD student" }, { "name": "CHAO WANG", "gender": "Female", "institution": "Eurecom", "country": "", "position": "PhD student" }, { "name": "Giulio Franzese", "gender": "Male", "institution": "Eurecom", "country": "FR", "position": "Assistant Professor" }, { "name": "Massimo Gallo", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "FR", "position": "Principal Researcher" }, { "name": "Pietro Michiardi", "gender": "Male", "institution": "EURECOM", "country": "", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 3, "n_ref_uni": 62, "n_ref": 135, "n_ref_all": 175, "n_fig": 8, "n_tab": 16, "L_tab": 5760, "n_element_tab": 366, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1620, "n_element_tab_1": 78, "formula_len_all": 2087, "formula_len_all_1": 381, "len_all": 239754, "len_all_1": 70260, "len_abs": 1597, "len_title": 93, "len_sents": 58757, "len_sents_1": 34139, "n_sents": 424, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 45, "L_abs": 1013, "title": "Information Theoretic Text-to-Image Alignment", "abs": "Diffusion models for Text-to-Image (T2I) conditional generation have recently achieved\ntremendous success. Yet, aligning these models with user\u2019s intentions still involves a\nlaborious trial-and-error process, and this challenging alignment problem has attracted\nconsiderable attention from the research community. In this work, instead of relying on\nfine-grained linguistic analyses of prompts, human annotation, or auxiliary vision-language\nmodels, we use Mutual Information (MI) to guide model alignment. In brief, our method\nuses self-supervised fine-tuning and relies on a point-wise MI estimation between prompts\nand images to create a synthetic fine-tuning set for improving model alignment. Our\nanalysis indicates that our method is superior to the state-of-the-art, yet it only requires\nthe pre-trained denoising network of the T2I model itself to estimate MI, and a simple\nfine-tuning strategy that improves alignment while maintaining image quality. Code available at https://github.com/Chao0511/mitune.", "keywords": [ "Diffusion model", "Text-image alignment", "Mutual information" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "UgPoHhYQ2U", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Gabriel L. Oliveira", "gender": "Male", "institution": "Borealis AI", "country": "CA", "position": "Senior Machine Learning Researcher" }, { "name": "Wonho Bae", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "D. J. Sutherland", "gender": "Female", "institution": "University of British Columbia", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 39, "n_formula_1": 18, "n_ref_uni": 49, "n_ref": 92, "n_ref_all": 111, "n_fig": 9, "n_tab": 1, "L_tab": 600, "n_element_tab": 116, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 90, "n_element_tab_1": 5, "formula_len_all": 4403, "formula_len_all_1": 1801, "len_all": 138999, "len_all_1": 63554, "len_abs": 1037, "len_title": 117, "len_sents": 41404, "len_sents_1": 28923, "n_sents": 332, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1046, "title": "Uncertainty Herding: One Active Learning Method for All Label Budgets", "abs": "Most active learning research has focused on methods which perform well when many labels are available, but can be dramatically worse than random selection when label budgets are small.\nOther methods have focused on the low-budget regime, but do poorly as label budgets increase.\nAs the line between \"low\" and \"high\" budgets varies by problem,\nthis is a serious issue in practice.\nWe propose *uncertainty coverage*,\nan objective which generalizes a variety of low- and high-budget objectives,\nas well as natural, hyperparameter-light methods to smoothly interpolate between low- and high-budget regimes.\nWe call greedy optimization of the estimate Uncertainty Herding;\nthis simple method is computationally fast,\nand we prove that it nearly optimizes the distribution-level coverage.\nIn experimental validation across a variety of active learning tasks,\nour proposal matches or beats state-of-the-art performance in essentially all cases;\nit is the only method of which we are aware that reliably works well in both low- and high-budget settings.", "keywords": [ "Active learning" ], "rating_list": [ 6, 8, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 4, 3 ], "confidence_list": [ 4, 3, 2, 3, 3 ] }, { "paper_id": "UfczlMudN6", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Jonathan P. How", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "James Queeney", "gender": "Male", "institution": "Mitsubishi Electric Research Labs", "country": "", "position": "Postdoc" }, { "name": "Mouhacine Benosman", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Xiaoyi Cai", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 43, "n_ref": 75, "n_ref_all": 104, "n_fig": 13, "n_tab": 12, "L_tab": 6907, "n_element_tab": 1171, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 484, "n_element_tab_1": 144, "formula_len_all": 667, "formula_len_all_1": 678, "len_all": 167882, "len_all_1": 65156, "len_abs": 959, "len_title": 111, "len_sents": 48095, "len_sents_1": 31852, "n_sents": 324, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 962, "title": "GRAM: Generalization in Deep RL with a Robust Adaptation Module", "abs": "The reliable deployment of deep reinforcement learning in real-world settings requires the ability to generalize across a variety of conditions, including both in-distribution scenarios seen during training as well as novel out-of-distribution scenarios. In this work, we present a framework for dynamics generalization in deep reinforcement learning that unifies these two distinct types of generalization within a single architecture. We introduce a robust adaptation module that provides a mechanism for identifying and reacting to both in-distribution and out-of-distribution environment dynamics, along with a joint training pipeline that combines the goals of in-distribution adaptation and out-of-distribution robustness. Our algorithm GRAM achieves strong generalization performance across in-distribution and out-of-distribution scenarios upon deployment, which we demonstrate on a variety of realistic simulated locomotion tasks with a quadruped robot.", "keywords": [ "deep reinforcement learning", "generalization", "robust adaptation" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "UeVx6L59fg", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Senthil Purushwalkam", "gender": "Male", "institution": "Salesforce", "country": "", "position": "Researcher" }, { "name": "Shafiq Joty", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Principal Researcher" }, { "name": "Shrey Pandit", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "MS student" }, { "name": "Yifei Ming", "gender": "Male", "institution": "Salesforce AI Research", "country": "US", "position": "Researcher" }, { "name": "Zixuan Ke", "gender": "Male", "institution": "Salesforce AI Research", "country": "", "position": "Researcher" }, { "name": "Phi Xuan Nguyen", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 54, "n_ref": 90, "n_ref_all": 136, "n_fig": 21, "n_tab": 7, "L_tab": 1430, "n_element_tab": 220, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 415, "formula_len_all_1": 0, "len_all": 204297, "len_all_1": 53730, "len_abs": 1228, "len_title": 151, "len_sents": 48899, "len_sents_1": 26133, "n_sents": 406, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 103, "L_abs": 1233, "title": "FaithEval: Can Your Language Model Stay Faithful to Context, Even If \"The Moon is Made of Marshmallows\"", "abs": "Ensuring faithfulness to context in large language models (LLMs) and retrieval-augmented generation (RAG) systems is crucial for reliable deployment in real-world applications, as incorrect or unsupported information can erode user trust. Despite advancements on standard benchmarks, faithfulness hallucination\u2014where models generate responses misaligned with the provided context\u2014remains a significant challenge. In this work, we introduce FaithEval, a novel and comprehensive benchmark tailored to evaluate the faithfulness of LLMs in contextual scenarios across three diverse tasks: unanswerable, inconsistent, and counterfactual contexts. These tasks simulate real-world challenges where retrieval mechanisms may surface incomplete, contradictory, or fabricated information. FaithEval comprises 4.9K high-quality problems in total, validated through a rigorous four-stage context construction and validation framework, employing both LLM-based auto-evaluation and human validation. Our extensive study across a wide range of open-source and proprietary models reveals that even state-of-the-art models often struggle to remain faithful to the given context, and that larger models do not necessarily exhibit improved faithfulness.", "keywords": [ "Large Language Models", "Contextual LLM", "Faithfulness", "Hallucination", "Benchmark and Evaluation" ], "rating_list": [ 6, 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "UeHunlny77", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Baibei Ji", "gender": "Male", "institution": "Soochow University", "country": "CN", "position": "Undergrad student" }, { "name": "Juntao Li", "gender": "Male", "institution": "Soochow University, China", "country": "", "position": "Associate Professor" }, { "name": "Keyan Zhou", "gender": "unknown", "institution": "Suzhou University", "country": "CN", "position": "MS student" }, { "name": "Min Zhang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Zecheng Tang", "gender": "Male", "institution": "Soochow University", "country": "", "position": "PhD student" }, { "name": "jianye hou", "gender": "Male", "institution": "The Chinese University of Hong Kong\uff08shenzhen\uff09", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 40, "n_ref": 76, "n_ref_all": 92, "n_fig": 8, "n_tab": 9, "L_tab": 7842, "n_element_tab": 333, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2705, "n_element_tab_1": 230, "formula_len_all": 849, "formula_len_all_1": 806, "len_all": 202382, "len_all_1": 54791, "len_abs": 1735, "len_title": 121, "len_sents": 55802, "len_sents_1": 22932, "n_sents": 396, "n_sents_1": 139, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1747, "title": "L-CiteEval: Do Long-Context Models Truly Leverage Context for Responding?", "abs": "Long-context models (LCMs) have made remarkable strides in recent years, offering users great convenience for handling tasks that involve long context, such as document summarization.\nAs the community increasingly prioritizes the faithfulness of generated results, merely ensuring the accuracy of LCM outputs is insufficient because it is quite challenging for humans to verify the results from the extremely lengthy context.\nYet, although some efforts have been made to assess whether LCMs respond truly based on the context, these works either are limited to specific tasks or heavily rely on external evaluation resources like GPT-4.\nIn this work, we introduce L-CiteEval, a comprehensive multi-task benchmark for long-context understanding with citations, aiming to evaluate both the understanding capability and faithfulness of LCMs.\nL-CiteEval covers 11 tasks from diverse domains, spanning context lengths from 8K to 48K, and provides a fully automated evaluation suite.\nThrough testing with 11 cutting-edge closed-source and open-source LCMs, we find that although these models show minor differences in their generated results, open-source models substantially trail behind their closed-source counterparts in terms of citation accuracy and recall.\nThis suggests that current open-source LCMs are prone to responding based on their inherent knowledge rather than the given context, posing a significant risk to the user experience in practical applications.\nWe also evaluate the RAG approach and observe that RAG can significantly improve the faithfulness of LCMs, albeit with a slight decrease in the generation quality.\nFurthermore, we discover a correlation between the attention mechanisms of LCMs and the citation generation process.", "keywords": [ "Long-context evaluation benchmark", "model faithfulness", "citation generation" ], "rating_list": [ 5, 1, 5, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "UdGwotKVQI", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Binchi Zhang", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Hanghang Tong", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Jundong Li", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" }, { "name": "Yushun Dong", "gender": "Male", "institution": "Florida State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 49, "n_formula_1": 13, "n_ref_uni": 51, "n_ref": 163, "n_ref_all": 249, "n_fig": 6, "n_tab": 14, "L_tab": 4890, "n_element_tab": 365, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1182, "n_element_tab_1": 158, "formula_len_all": 6041, "formula_len_all_1": 1195, "len_all": 412029, "len_all_1": 216282, "len_abs": 4317, "len_title": 106, "len_sents": 109582, "len_sents_1": 60510, "n_sents": 877, "n_sents_1": 457, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1245, "title": "Certified Defense on the Fairness of Graph Neural Networks", "abs": "Graph Neural Networks (GNNs) have emerged as a prominent graph learning model in various graph-based tasks over the years. Nevertheless, due to the vulnerabilities of GNNs, it has been empirically proved that malicious attackers could easily corrupt the fairness level of their predictions by adding perturbations to the input graph data. In this paper, we take crucial steps to study a novel problem of certifiable defense on the fairness level of GNNs. Specifically, we propose a principled framework named ELEGANT and present a detailed theoretical certification analysis for the fairness of GNNs. ELEGANT takes any GNNs as its backbone, and the fairness level of such a backbone is theoretically impossible to be corrupted under certain perturbation budgets for attackers. Notably, ELEGANT does not have any assumption over the GNN structure or parameters, and does not require re-training the GNNs to realize certification. Hence it can serve as a plug-and-play framework for any optimized GNNs ready to be deployed. We verify the satisfactory effectiveness of ELEGANT in practice through extensive experiments on real-world datasets across different backbones of GNNs, where ELEGANT is also demonstrated to be beneficial for GNN debiasing.", "keywords": [ "Algorithmic Fairness", "Graph Neural Networks", "Attack and Defense" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Ud7I21wHnl", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Lei Feng", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" }, { "name": "Roy Ka-Wei Lee", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" }, { "name": "Shuo He", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Zhang Zhifang", "gender": "Male", "institution": "The University of Queensland", "country": "AU", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 54, "n_ref": 132, "n_ref_all": 167, "n_fig": 6, "n_tab": 10, "L_tab": 2590, "n_element_tab": 482, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 538, "n_element_tab_1": 185, "formula_len_all": 667, "formula_len_all_1": 661, "len_all": 171499, "len_all_1": 63823, "len_abs": 1244, "len_title": 89, "len_sents": 45556, "len_sents_1": 26766, "n_sents": 370, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1248, "title": "A Closer Look at Backdoor Attacks on CLIP", "abs": "We present a comprehensive empirical study on how backdoor attacks affect CLIP by analyzing the representations of backdoor images. Specifically, based on the methodology of representation decomposing, image representations can be decomposed into a sum of representations across individual image patches, attention heads (AHs), and multi-layer perceptrons (MLPs) in different model layers. By\nexamining the effect of backdoor attacks on model components, we have the following empirical findings. (1) Different backdoor attacks would infect different model components, i.e., local patch-based backdoor attacks mainly affect AHs, while global noise-based backdoor attacks mainly affect MLPs. (2) Infected AHs are centered on the last layer, while infected MLPs are decentralized on several late layers. (3) Some AHs are not greatly infected by backdoor attacks, and even infected AHs could still maintain the original functionality. These observations motivate us to defend against backdoor attacks by detecting infected AHs, repairing their representations or filtering backdoor samples with too many infected AHs, in the inference stage. Experimental results validate our empirical findings and demonstrate the effectiveness of the defense methods", "keywords": [ "Backdoor Attacks", "CLIP", "Representation Decomposing", "Attention Heads" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "UchRjcf4z7", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Moyang Guo", "gender": "Male", "institution": "Duke University, Duke University", "country": "US", "position": "MS student" }, { "name": "Yuepeng Hu", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Zhengyuan Jiang", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Neil Gong", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 5, "n_ref_uni": 31, "n_ref": 89, "n_ref_all": 125, "n_fig": 23, "n_tab": 4, "L_tab": 1843, "n_element_tab": 351, "n_fig_1": 20, "n_tab_1": 2, "L_tab_1": 1150, "n_element_tab_1": 182, "formula_len_all": 6404, "formula_len_all_1": 525, "len_all": 172921, "len_all_1": 67842, "len_abs": 896, "len_title": 85, "len_sents": 62479, "len_sents_1": 31144, "n_sents": 456, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 37, "L_abs": 983, "title": "A Transfer Attack to Image Watermarks", "abs": "Watermark has been widely deployed by industry to detect AI-generated images. The robustness of such watermark-based detector against evasion attacks in the white-box and black-box settings is well understood in the literature. However, the robustness in the no-box setting is much less understood. In this work, we propose a new transfer evasion attack to image watermark in the no-box setting. Our transfer attack adds a perturbation to a watermarked image to evade multiple surrogate watermarking models trained by the attacker itself, and the perturbed watermarked image also evades the target watermarking model. Our major contribution is to show that, both theoretically and empirically, watermark-based AI-generated image detector based on existing watermarking methods is not robust to evasion attacks even if the attacker does not have access to the watermarking model nor the detection API. Our code is available at: https://github.com/hifi-hyp/Watermark-Transfer-Attack.", "keywords": [ "Image Watermarking", "Transfer Attack", "AI-generated Image" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Uc3kog3O45", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Chanyoung Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Junseok Lee", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Namkyeong Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Yeongmin Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "Undergrad student" }, { "name": "Yunhak Oh", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 12, "n_ref_uni": 45, "n_ref": 75, "n_ref_all": 121, "n_fig": 24, "n_tab": 8, "L_tab": 9668, "n_element_tab": 548, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 6918, "n_element_tab_1": 392, "formula_len_all": 755, "formula_len_all_1": 1178, "len_all": 191329, "len_all_1": 72627, "len_abs": 1565, "len_title": 132, "len_sents": 50930, "len_sents_1": 28662, "n_sents": 327, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1387, "title": "Global Context-aware Representation Learning for Spatially Resolved Transcriptomics", "abs": "Spatially Resolved Transcriptomics (SRT) is a cutting-edge technique that captures the spatial context of cells within tissues, enabling the study of complex biological networks. Recently, graph-based deep learning has been utilized in identifying meaningful spatial domains by leveraging both gene expression and spatial information. However, these approaches fall short in obtaining qualified spot representations, particularly for those located around the boundary of cell type clusters, as they heavily emphasize spatially local spots that have minimal feature differences from an anchor node. To address this limitation, we propose a novel framework, Spotscape, which introduces the Similarity Telescope module designed to learn spot representations by capturing the global relationships among multiple spots. Additionally, to address the challenges that arise when integrating multiple slices from heterogeneous sources, we propose a similarity scaling strategy that explicitly regulates the distances between intra- and inter-slice spots to ensure they remain nearly the same. Extensive experiments demonstrate the superiority of Spotscape in various downstream tasks, including spatial domain identification, multi-slice integration, and alignment tasks, compared to baseline methods. Our code is available at the following link: https://anonymous.4open.science/r/Spotscape-E312/", "keywords": [ "Spatially Resolved Transcriptomics", "Self-Supervised Learning" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "UbugxiPs6y", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Chunhua Shen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Hao Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Hengtao Li", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Hui Xue'", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Li Ke", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "MS student" }, { "name": "Longtao Huang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Muzhi Zhu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Weian Mao", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "PhD student" }, { "name": "Yang Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 48, "n_ref": 95, "n_ref_all": 100, "n_fig": 4, "n_tab": 5, "L_tab": 1457, "n_element_tab": 149, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 845, "n_element_tab_1": 63, "formula_len_all": 596, "formula_len_all_1": 488, "len_all": 134661, "len_all_1": 64709, "len_abs": 1573, "len_title": 130, "len_sents": 35068, "len_sents_1": 29408, "n_sents": 279, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1580, "title": "Bridging the Gap between Semantic Correspondence and Robust Visual Representation", "abs": "Predicting cross-image semantic correspondence among various instances within the same category is a fundamental but challenging task in computer vision. Models are supposed to characterize both high-level semantic features and low-level texture information to accurately finds the correspondence between pixels. The quality of features directly affects the matching results. Recently, pre-trained models with self-supervised training methods have demonstrated promising performance in representation learning and can serve as a strong backbone to provide robust visual features. However, existing methods have been found to poorly adapt to such features. Their complex designs of the matching module do not yield significant performance boost due to the disruption of the original representation and the absence of high-resolution low-level information. In this work, we introduce a simple yet effective framework named ViTSC to unlock the substantial potential of self-supervised vision transformers for semantic correspondence. We introduce three key components: a cross-perception module to align semantic features of the same part from different images while preserving the original representation as much as possible, an auxiliary loss to eliminate ambiguity from semantically similar objects, and a low-level correlation-guided upsampler to generate high-resolution flow maps for precise localization. ViTSC shows reliable semantic correspondence performance, surpassing previous state-of-the-art methods on all three standard benchmarks SPair-71k, PF-PASCAL and PF-WILLOW.", "keywords": [ "semantic correspondence", "foundation models" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "UbMYhX60tY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Ching-Hua Lee", "gender": "Male", "institution": "Samsung Research America", "country": "US", "position": "Researcher" }, { "name": "Chouchang Yang", "gender": "Male", "institution": "Samsung", "country": "US", "position": "Researcher" }, { "name": "Hongxia Jin", "gender": "Female", "institution": "Head of AI center", "country": "US", "position": "Vice President" }, { "name": "Jaejin Cho", "gender": "Male", "institution": "Samsung Research America", "country": "", "position": "Researcher" }, { "name": "Rakshith Sharma Srinivasa", "gender": "Male", "institution": "Scale AI", "country": "US", "position": "Researcher" }, { "name": "Yashas Malur Saidutta", "gender": "unknown", "institution": "Samsung", "country": "US", "position": "Researcher" }, { "name": "Yilin Shen", "gender": "Male", "institution": "Samsung Research America", "country": "US", "position": "Principal Researcher" } ], "n_formula": 29, "n_formula_1": 16, "n_ref_uni": 66, "n_ref": 164, "n_ref_all": 199, "n_fig": 11, "n_tab": 12, "L_tab": 1866, "n_element_tab": 185, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1130, "n_element_tab_1": 106, "formula_len_all": 2897, "formula_len_all_1": 1542, "len_all": 223883, "len_all_1": 66963, "len_abs": 2149, "len_title": 152, "len_sents": 65642, "len_sents_1": 31088, "n_sents": 465, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 2156, "title": "RestoreGrad: Signal Restoration Using Conditional Denoising Diffusion Models with Jointly Learned Prior", "abs": "Denoising diffusion probabilistic models (DDPMs) estimate the data distribution by sequentially denoising samples drawn from a prior distribution, which is typically assumed to be the standard Gaussian for simplicity. Owing to their capabilities of generating high-fidelity samples, DDPMs can be utilized for signal restoration tasks in recovering a clean signal from its degraded observation(s), by conditioning the model on the degraded signal. The degraded signals are themselves contaminated versions of the clean signals; due to this correlation, they may encompass certain useful information about the target clean data distribution. However, naively adopting the standard Gaussian as the prior distribution in turn discards such information. In this paper, we propose to improve conditional DDPMs for signal restoration applications by leveraging a more informative prior that is jointly learned with the diffusion model. The proposed framework, called RestoreGrad, exploits the correlation between the degraded and clean signals to construct a better prior for restoration tasks. In contrast to existing DDPMs that just settle on using pre-defined or handcrafted priors, RestoreGrad learns the prior jointly with the diffusion model. To this end, we first derive a new objective function from a modified evidence lower bound (ELBO) of the data log-likelihood, to incorporate the prior learning process into conditional DDPMs. Then, we suggest a corresponding joint learning paradigm for optimizing the new ELBO. Notably, RestoreGrad requires minimum modifications to the diffusion model itself; thus, it can be flexibly implemented on top of various conditional DDPM-based signal restoration models. On speech and image restoration tasks, we show that RestoreGrad demonstrates faster convergence (5-10 times fewer training steps) to achieve on par or better perceptual quality of restored signals over existing DDPM baselines, along with improved robustness to using fewer sampling steps in inference time (2-2.5 times fewer steps), advocating the advantages of leveraging jointly learned prior for efficiency improvements in the diffusion process.", "keywords": [ "Denoising diffusion probabilistic model", "prior distribution", "posterior", "speech enhancement", "image restoration" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "UbLvSPMvMA", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Fran\u00e7ois Remy", "gender": "Male", "institution": "Universiteit Gent", "country": "BE", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 28, "n_ref": 35, "n_ref_all": 46, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 252, "formula_len_all_1": 190, "len_all": 88572, "len_all_1": 53376, "len_abs": 1302, "len_title": 123, "len_sents": 28844, "len_sents_1": 28607, "n_sents": 170, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1305, "title": "Sparsity beyond TopK: A Novel Cosine Loss for Sparse Binary Representations", "abs": "While binary vectorization and sparse representations have recently emerged as promising strategies for efficient vector storage and mechanistic interpretability, the integration of these two paradigms has till now remained largely unexplored.\nIn this paper, we introduce an exciting approach for sparse binary representations, leveraging a soft TopK Cosine Loss to facilitate the transition from dense to sparse latent spaces.\nUnlike traditional TopK methods which impose rigid sparsity constraints, our approach naturally yields a more flexible distribution of activations, effectively capturing the varying degrees of conceptual depth present in the data.\nFurthermore, our cosine loss formulation inherently mitigates the emergence of inactive features, thereby eliminating the need for complex re-activation strategies prevalent in other recent works. \nWe validate our method on a large dataset of biomedical concept embeddings, demonstrating enhanced interpretability and significant reductions in storage overhead.\nOur present findings highlight the clear potential of cosine-based binary sparsity alignment for developing interpretable and efficient concept representations, positioning our approach as a compelling solution for applications in decision-making systems and compact vector databases.", "keywords": [ "sparse", "binary", "interpretability", "latent", "embedding", "vector", "representations", "cosine similarity", "sigmoid" ], "rating_list": [ 1, 1, 3 ], "soundness_list": [ 1, 2, 2 ], "presentation_list": [ 1, 1, 3 ], "contribution_list": [ 1, 1, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "UatDdAlr2x", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Freya Behrens", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Lenka Zdeborova", "gender": "Female", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "", "position": "Associate Professor" }, { "name": "Luca Biggio", "gender": "Male", "institution": "Bocconi University", "country": "IT", "position": "Assistant Professor" } ], "n_formula": 102, "n_formula_1": 10, "n_ref_uni": 42, "n_ref": 71, "n_ref_all": 107, "n_fig": 12, "n_tab": 2, "L_tab": 562, "n_element_tab": 35, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5277, "formula_len_all_1": 738, "len_all": 219928, "len_all_1": 69763, "len_abs": 1758, "len_title": 144, "len_sents": 74246, "len_sents_1": 33718, "n_sents": 616, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1443, "title": "Counting in small transformers: The delicate interplay between attention and feed-forward layers", "abs": "How do different architectural design choices influence the space of solutions that a transformer can implement and learn? How do different components interact with each other to shape the model's hypothesis space? We investigate these questions by characterizing the solutions simple transformer blocks can implement when challenged to solve the histogram task -- counting the occurrences of each item in an input sequence from a fixed vocabulary. Despite its apparent simplicity, this task exhibits a rich phenomenology: our analysis reveals a strong inter-dependence between the model's predictive performance and the vocabulary and embedding sizes, the token-mixing mechanism and the capacity of the feed-forward block. In this work, we characterize two different counting strategies that small transformers can implement theoretically: relation-based and inventory-based counting, the latter being less efficient in computation and memory. The emergence of either strategy is heavily influenced by subtle synergies among hyperparameters and components, and depends on seemingly minor architectural tweaks like the inclusion of softmax in the attention mechanism. By introspecting models \\textit{trained} on the histogram task, we verify the formation of both mechanisms in practice. Our findings highlight that even in simple settings, slight variations in model design can cause significant changes to the solutions a transformer learns.", "keywords": [ "attention", "mechanistic interpretability", "architecture", "toy model", "counting", "activation function" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "UapxTvxB3N", "primary_area": "generative models", "n_author": 14, "author_info_list": [ { "name": "Di Lin", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Gengjie Lin", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Jibin Peng", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "MS student" }, { "name": "Kairui Yang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Qing Guo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Xiaosong Jia", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Xin WANG", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Yipeng Wu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "MS student" }, { "name": "Zhao Huang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "MS student" }, { "name": "Ziyuan Zhong", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "ZuoDie", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "MS student" }, { "name": "\u6d69\u5929 \u8463", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Zi-Hao Guo", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 37, "n_ref": 62, "n_ref_all": 90, "n_fig": 10, "n_tab": 3, "L_tab": 537, "n_element_tab": 75, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 801, "formula_len_all_1": 801, "len_all": 128362, "len_all_1": 58459, "len_abs": 1421, "len_title": 145, "len_sents": 33944, "len_sents_1": 27884, "n_sents": 242, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1427, "title": "Trajectory-LLM: A Language-based Data Generator for Trajectory Prediction in Autonomous Driving", "abs": "Vehicle trajectory prediction is a crucial aspect of autonomous driving, which requires extensive trajectory data to train prediction models to understand the complex, varied, and unpredictable patterns of vehicular interactions. However, acquiring real-world data is expensive, so we advocate using Large Language Models (LLMs) to generate abundant and realistic trajectories of interacting vehicles efficiently. These models rely on textual descriptions of vehicle-to-vehicle interactions on a map to produce the trajectories. We introduce Trajectory-LLM (Traj-LLM), a new approach that takes brief descriptions of vehicular interactions as input and generates corresponding trajectories. Unlike language-based approaches that translate text directly to trajectories, Traj-LLM uses reasonable driving behaviors to align the vehicle trajectories with the text. This results in an \"interaction-behavior-trajectory\" translation process. We have also created a new dataset, Language-to-Trajectory (L2T), which includes 240K textual descriptions of vehicle interactions and behaviors, each paired with corresponding map topologies and vehicle trajectory segments. By leveraging the L2T dataset, Traj-LLM can adapt interactive trajectories to diverse map topologies. Furthermore, Traj-LLM generates additional data that enhances downstream prediction models, leading to consistent performance improvements across public benchmarks.", "keywords": [ "Trajectory Prediction", "Large Language Model", "Autonomous Driving" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "UZQl0rbxj6", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Carlo Tomasi", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Full Professor" }, { "name": "Kelsey Lieberman", "gender": "Female", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Shuai Yuan", "gender": "Male", "institution": "TikTok", "country": "US", "position": "Researcher" }, { "name": "Swarna K Ravindran", "gender": "Female", "institution": "Department of Computer Science, Duke University", "country": "US", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 11, "n_ref_uni": 37, "n_ref": 73, "n_ref_all": 98, "n_fig": 1, "n_tab": 18, "L_tab": 4660, "n_element_tab": 400, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1070, "n_element_tab_1": 97, "formula_len_all": 1869, "formula_len_all_1": 652, "len_all": 194538, "len_all_1": 61233, "len_abs": 1071, "len_title": 168, "len_sents": 48718, "len_sents_1": 27883, "n_sents": 397, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 118, "L_abs": 1077, "title": "Training Over a Distribution of Hyperparameters for Enhanced Performance and Adaptability on Imbalanced Classification", "abs": "Although binary classification is a well-studied problem, training reliable classifiers under severe class imbalance remains a challenge. Recent techniques mitigate the ill effects of imbalance on training by modifying the loss functions or optimization methods. We observe that different hyperparameter values on these loss functions perform better at different recall values. We propose to exploit this fact by training one model over a distribution of hyperparameter values--instead of a single value--via Loss Conditional Training (LCT). Experiments show that training over a distribution of hyperparameters not only approximates the performance of several models, but actually improves the overall performance of models on both CIFAR and real medical imaging applications such as melanoma and diabetic retinopathy detection. Furthermore, training models with LCT is more efficient because some hyperparameter tuning can be conducted after training to meet individual needs without needing to retrain from scratch. Code will be made available upon acceptance of this paper.", "keywords": [ "class imbalance", "hyperparameter tuning", "ROC curves", "machine learning" ], "rating_list": [ 6, 6, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 2, 4 ], "presentation_list": [ 3, 3, 2, 3, 4 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 5, 3, 2, 2, 4 ] }, { "paper_id": "UZ893n8FXr", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Eshaan Nichani", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Hengyu Fu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Jason D. Lee", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Zihao Wang", "gender": "Not Specified", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 286, "n_formula_1": 23, "n_ref_uni": 33, "n_ref": 71, "n_ref_all": 92, "n_fig": 4, "n_tab": 4, "L_tab": 366, "n_element_tab": 10, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 71462, "formula_len_all_1": 2623, "len_all": 354681, "len_all_1": 83638, "len_abs": 5564, "len_title": 113, "len_sents": 77219, "len_sents_1": 36045, "n_sents": 872, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1648, "title": "Learning Hierarchical Polynomials of Multiple Nonlinear Features", "abs": "In deep learning theory, a critical question is to understand how neural networks learn hierarchical features. In this work, we study the learning of hierarchical polynomials of multiple nonlinear features using three-layer neural networks. We examine a broad class of functions of the form $f^{\\star}=g^{\\star}\\circ \\mathbf{p}$, where $\\mathbf{p}:\\mathbb{R}^{d} \\rightarrow \\mathbb{R}^{r}$ represents multiple quadratic features with $r \\ll d$ and $g^{\\star}:\\mathbb{R}^{r}\\rightarrow \\mathbb{R}$ is a polynomial of degree $p$. This can be viewed as a nonlinear generalization of the multi-index model, and also an expansion upon previous work on nonlinear feature learning that focused only on a single feature (i.e. $r = 1$). \n Our primary contribution shows that a three-layer neural network trained via layerwise gradient descent suffices for \n - complete recovery of the space spanned by the nonlinear features \n- efficient learning of the target function $f^{\\star}=g^{\\star}\\circ \\mathbf{p}$ or transfer learning of $f=g\\circ \\mathbf{p}$ with a different link function\n \nwithin $\\widetilde{\\mathcal{O}}(d^4)$ samples and polynomial time.\nFor such hierarchical targets, our result substantially improves the sample complexity ${\\Theta}(d^{2p})$ of the kernel methods, demonstrating the power of efficient feature learning. It is important to highlight that our results leverage novel techniques and thus manage to go beyond all prior settings such as single-index and multi-index models as well as models depending just on one nonlinear feature, contributing to a more comprehensive understanding of feature learning in deep learning.", "keywords": [ "Deep Learning Theory", "Feature Learning", "Three-Layer Neural Network", "Gradient Descent", "Sample Complexity" ], "rating_list": [ 8, 8, 6, 6, 6 ], "soundness_list": [ 3, 4, 3, 3, 4 ], "presentation_list": [ 3, 4, 2, 3, 4 ], "contribution_list": [ 3, 4, 2, 3, 2 ], "confidence_list": [ 4, 4, 2, 4, 4 ] }, { "paper_id": "UYqssWc7TC", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Matthijs van Leeuwen", "gender": "Male", "institution": "Leiden University, Leiden University", "country": "NL", "position": "Associate Professor" }, { "name": "Shujian Yu", "gender": "Male", "institution": "Vrije Universiteit Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Yuhang Wang", "gender": "Male", "institution": "Leiden University, Leiden University", "country": "NL", "position": "MS student" }, { "name": "Zhong Li", "gender": "Male", "institution": "LIACS, Leiden University", "country": "NL", "position": "Postdoc" } ], "n_formula": 24, "n_formula_1": 2, "n_ref_uni": 58, "n_ref": 150, "n_ref_all": 207, "n_fig": 9, "n_tab": 10, "L_tab": 8722, "n_element_tab": 745, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 750, "n_element_tab_1": 79, "formula_len_all": 696, "formula_len_all_1": 72, "len_all": 321909, "len_all_1": 68197, "len_abs": 1980, "len_title": 136, "len_sents": 118901, "len_sents_1": 35645, "n_sents": 858, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1996, "title": "Labels Are Not All You Need: Evaluating Node Embedding Quality without Relying on Labels", "abs": "Graph Neural Network (GNN) based node embedding methods are a promising approach to learning node representations for downstream tasks such as link prediction, node classification, and node clustering. GNN-based methods usually work in an unsupervised or semi-supervised manner, learning node representations without or with limited label information. We empirically show, however, that the performance of learned node embeddings on downstream tasks may be heavily impacted by the GNN-method's hyperparameter configuration. Unfortunately, existing hyperparameter optimisation methods typically rely on labeled data for evaluation, making them unsuitable for unsupervised scenarios. This raises the question: *how can we tune the hyperparameters of GNNs without using label information to obtain high quality node embeddings?* To answer this, we propose a framework for evaluating node embedding quality without relying on labels. Specifically, our framework consists of two steps: *building prior beliefs* that characterise high-quality node embeddings, and *quantifying the extent* to which those prior beliefs are satisfied. More importantly, we instantiate our framework from two different but complementary perspectives: spatial and spectral information. First, we introduce the Consensus-based Space Occupancy Rate (CSOR) method that evaluates node embedding quality from a spatial view. It conducts pairwise comparisons of the spatial distances between node embeddings obtained from various hyperparameter configurations. Next, we present the Spectral Space Occupancy Rate (SSOR) method, which takes a spectral perspective and evaluates the embedding quality by examining the singular values of the node embedding matrices. Extensive experiments on seven GNN models with four benchmark datasets demonstrate the effectiveness of both CSOR and SSOR. Specifically, both methods consistently prioritise hyperparameter configurations that yield high-quality node embeddings for downstream tasks.", "keywords": [ "Graph Representation Learning", "Unsupervised Learning", "Graph Neural Networks", "Model Selection" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "UYcUpiULmT", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Nikolay Atanasov", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Sharath Matada", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Yuanyuan Shi", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Luke Autar Bhan", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" } ], "n_formula": 50, "n_formula_1": 28, "n_ref_uni": 47, "n_ref": 84, "n_ref_all": 116, "n_fig": 6, "n_tab": 7, "L_tab": 2222, "n_element_tab": 81, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 996, "n_element_tab_1": 10, "formula_len_all": 2395, "formula_len_all_1": 1118, "len_all": 194547, "len_all_1": 68321, "len_abs": 1401, "len_title": 99, "len_sents": 50034, "len_sents_1": 29407, "n_sents": 400, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1294, "title": "Generalizable Motion Planning via Operator Learning", "abs": "In this work, we introduce a planning neural operator (PNO) for predicting the value function of a motion planning problem. We recast value function approximation as learning a single operator from the cost function space to the value function\nspace, which is defined by an Eikonal partial differential equation (PDE). Therefore, our PNO model, despite being trained with a finite number of samples at coarse resolution, inherits the zero-shot super-resolution property of neural operators. We demonstrate accurate value function approximation at 16\u00d7 the training resolution on the MovingAI lab\u2019s 2D city dataset, compare with state-of-the-art neural value\nfunction predictors on 3D scenes from the iGibson building dataset and showcase optimal planning with 4-joint robotic manipulators. Lastly, we investigate employing the value function output of PNO as a heuristic function to accelerate motion planning. We show theoretically that the PNO heuristic is $\\epsilon$-consistent by introducing an inductive bias layer that guarantees our value functions satisfy the triangle inequality. With our heuristic, we achieve a $30$% decrease in nodes visited while obtaining near optimal path lengths on the MovingAI lab 2D city dataset, compared to classical planning methods (A$^\\ast$, RRT$^\\ast$).", "keywords": [ "Motion planning", "operator learning", "value function approximation", "Eikonal PDE" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "UYarAv7rUx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Ruifeng Ren", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "\u667a\u806a \u674e", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "MS student" } ], "n_formula": 57, "n_formula_1": 20, "n_ref_uni": 39, "n_ref": 119, "n_ref_all": 135, "n_fig": 2, "n_tab": 3, "L_tab": 354, "n_element_tab": 9, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 3305, "n_element_tab_1": 51, "formula_len_all": 4930, "formula_len_all_1": 1814, "len_all": 176355, "len_all_1": 72885, "len_abs": 1437, "len_title": 88, "len_sents": 57466, "len_sents_1": 32774, "n_sents": 416, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1445, "title": "Can Mamba Always Enjoy the \"Free Lunch\"?", "abs": "Transformers have been the cornerstone of current Large Language Models (LLMs); however, its linear growth in overhead during inference with respect to sequence length poses challenges for modeling long sequences. In this context, Mamba has gradually attracted attention due to its constant-level size during inference and existing empirical results have shown that it can perform comparably to Transformers in sequence modeling while offering significant savings. However, one may ask that, can Mamba always enjoy the ``free lunch\"? In this paper, we focus on analyzing the expressive ability of Mamba from a theoretical standpoint. First, inspired by the connection between Mamba and linear attention, we investigate potential shortcomings of the Mamba when performing the COPY operation. Our results indicate that Mamba with constant size may encounter bottlenecks when handling COPY, while it can achieve perfect performance when the size scales linearly with sequence length. Based on this observation, we analyze Mamba's ability to tackle DP problems when equipped with Chain of Thought (CoT). Our findings suggest that to solve arbitrary DP problems, the total cost of Mamba is comparable to standard and efficient Transformers. However, similar to efficient Transformers, when facing DP problems with favorable properties such as locality, Mamba can provide savings in overhead. Our results contribute to a deeper understanding of Mamba.", "keywords": [ "Mamba", "Expressive Power", "Chain of Thought" ], "rating_list": [ 5, 5, 3, 3, 5 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 2, 2 ], "contribution_list": [ 2, 4, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "UYZRaUCLAg", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Axel Levy", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Ellen D Zhong", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Assistant Professor" }, { "name": "Eric Ryan Chan", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "FREDERIC P POITEVIN", "gender": "unknown", "institution": "SLAC National Accelerator Laboratory", "country": "", "position": "Associate Staff Scientist" }, { "name": "Gordon Wetzstein", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Sara Fridovich-Keil", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 17, "n_ref_uni": 62, "n_ref": 90, "n_ref_all": 102, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1203, "formula_len_all_1": 1157, "len_all": 143881, "len_all_1": 66909, "len_abs": 1033, "len_title": 118, "len_sents": 32692, "len_sents_1": 32026, "n_sents": 238, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1038, "title": "Solving Inverse Problems in Protein Space Using Diffusion-Based Priors", "abs": "The interaction of a protein with its environment can be understood and controlled via its 3D structure. Experimental methods for protein structure determination, such as X-ray crystallography or cryogenic electron microscopy, shed light on biological processes but introduce challenging inverse problems. Learning-based approaches have emerged as accurate and efficient methods to solve these inverse problems for 3D structure determination, but are specialized for a predefined type of measurement. Here, we introduce a versatile framework to turn biophysical measurements, such as cryo-EM density maps, into 3D atomic models. Our method combines a physics-based forward model of the measurement process with a pretrained generative model providing a task-agnostic, data-driven prior. Our method outperforms posterior sampling baselines on linear and non-linear inverse problems. In particular, it is the first diffusion-based method for refining atomic models from cryo-EM maps and building atomic models from sparse distance matrices.", "keywords": [ "machine learning for structural biology", "protein structure determination", "diffusion models", "inverse problems" ], "rating_list": [ 8, 3, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "UYXq4q1GpW", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Junyuan Lin", "gender": "unknown", "institution": "Loyola Marymount University", "country": "US", "position": "Assistant Professor" }, { "name": "Mandy Barrett Korpusik", "gender": "Female", "institution": "Loyola Marymount University", "country": "US", "position": "Assistant Professor" }, { "name": "Mustafa Zaki", "gender": "Male", "institution": "eBay Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 17, "n_ref": 28, "n_ref_all": 42, "n_fig": 2, "n_tab": 3, "L_tab": 861, "n_element_tab": 75, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 771, "n_element_tab_1": 53, "formula_len_all": 317, "formula_len_all_1": 258, "len_all": 76214, "len_all_1": 51639, "len_abs": 834, "len_title": 128, "len_sents": 24475, "len_sents_1": 23749, "n_sents": 182, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 837, "title": "A Healthy Food Recommender System Using Collaborative Filtering and Transformers", "abs": "Unhealthy eating habits are a major contributing factor to public health problems such as the globally rising obesity rate. One way to help solve this problem is by creating systems that can suggest better food choices in order to improve the way people eat. A critical challenge with these systems is making sure they offer 1) suggestions that match what users like, while also 2) recommending healthy foods. In this paper, we introduce a novel food recommender system that provides healthy food recommendations similar to what the user has previously eaten. We used collaborative filtering to generate recommendations and re-ranked the recommendations using a novel health score and a BERT embedding similarity score. We evaluated our system on human subjects by conducting A/B testing on several methods deployed in a web application.", "keywords": [ "Collaborative Filtering", "EASE", "Nutrition", "BERT" ], "rating_list": [ 3, 3, 1, 1 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 1, 2, 1, 1 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "UXpwNNiMRC", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 1, "author_info_list": [ { "name": "Nicolai Dorka", "gender": "Male", "institution": "University of Freiburg, Universit\u00e4t Freiburg", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 62, "n_ref_all": 66, "n_fig": 2, "n_tab": 3, "L_tab": 467, "n_element_tab": 65, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 511, "n_element_tab_1": 67, "formula_len_all": 404, "formula_len_all_1": 404, "len_all": 114237, "len_all_1": 62459, "len_abs": 1279, "len_title": 108, "len_sents": 30773, "len_sents_1": 29925, "n_sents": 233, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1285, "title": "Quantile Regression for Distributional Reward Models in RLHF", "abs": "Reinforcement learning from human feedback (RLHF) has become a key method for aligning large language models (LLMs) with human preferences through the use of reward models. However, traditional reward models typically generate point estimates, which oversimplify the diversity and complexity of human values and preferences. In this paper, we introduce Quantile Reward Models (QRMs), a novel approach to reward modeling that learns a distribution over rewards instead of a single scalar value. Our method uses quantile regression to estimate a full, potentially multimodal distribution over preferences, providing a more powerful and nuanced representation of preferences. This distributional approach can better capture the diversity of human values, addresses label noise, and accommodates conflicting preferences by modeling them as distinct modes in the distribution. Our experimental results show that QRM outperforms comparable traditional point-estimate models on RewardBench. Furthermore, we demonstrate that the additional information provided by the distributional estimates can be utilized in downstream applications, such as risk-aware reinforcement learning, resulting in LLM policies that generate fewer extremely negative responses. Our code and model will be released.", "keywords": [ "RLHF", "LLM", "alignment", "reward models" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "UXNprzZmvZ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Chengwei Xia", "gender": "Male", "institution": "Lanzhou University", "country": "CN", "position": "PhD student" }, { "name": "Fan Ma", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" }, { "name": "Kun Zhan", "gender": "Male", "institution": "Lanzhou University", "country": "CN", "position": "Full Professor" }, { "name": "Ruijie Quan", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Yi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 42, "n_ref": 98, "n_ref_all": 117, "n_fig": 11, "n_tab": 2, "L_tab": 1413, "n_element_tab": 152, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2431, "n_element_tab_1": 168, "formula_len_all": 1242, "formula_len_all_1": 1027, "len_all": 116722, "len_all_1": 61597, "len_abs": 1706, "len_title": 128, "len_sents": 32474, "len_sents_1": 27512, "n_sents": 225, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1231, "title": "Adversarial-Guided Diffusion for Robust and High-Fidelity Multimodal LLM Attacks", "abs": "Recent diffusion-based adversarial attack methods have shown promising results in generating natural adversarial images. However, these methods often lack fidelity by inducing significant distortion on the original image with even small perturbations on the latent representation. In this paper, we propose Adversarial-Guided Diffusion (AGD), a novel diffusion-based generative adversarial attack framework, which introduces adversarial noise during the reverse sampling of conditional diffusion models. \nAGD uses editing-friendly inversion sampling to faithfully reconstruct images without significantly distorting them through gradients on the latent representation. In addition, AGD enhances latent representations by intelligently choosing sampling steps, thereby injecting adversarial semantics more smoothly. Extensive experiments demonstrate that our method outperforms state-of-the-art methods in both the effectiveness of generating adversarial images for targeted attacks on multimodal large language models (MLLMs) and image quality, successfully misleading the MLLM's responses. We argue that the security concerns surrounding the adversarial robustness of MLLMs deserve increased attention from the research community.", "keywords": [ "Adversarial Attack", "multimodal large language models" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "UXCfRU2Qs4", "primary_area": "applications to neuroscience & cognitive science", "n_author": 6, "author_info_list": [ { "name": "Apoorva Bhandari", "gender": "Male", "institution": "Brown University", "country": "", "position": "Postdoc" }, { "name": "Cristian Buc Calderon", "gender": "Male", "institution": "Centro Nacional de Inteligencia Artificial", "country": "CL", "position": "Researcher" }, { "name": "Eugenio Herrera-Berg", "gender": "unknown", "institution": "Centro Nacional de Inteligencia Artificial", "country": "", "position": "Researcher" }, { "name": "Jos\u00e9 Guillen", "gender": "Male", "institution": "Universidad T\u00e9cnica Federico Santa Mar\u00eda", "country": "CL", "position": "MS student" }, { "name": "Marc-Llu\u00eds Vives", "gender": "Male", "institution": "Leiden University", "country": "", "position": "Assistant Professor" }, { "name": "Pablo Le\u00f3n-Villagr\u00e1", "gender": "unknown", "institution": "Brown University", "country": "", "position": "Postdoc" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 64, "n_ref": 96, "n_ref_all": 122, "n_fig": 9, "n_tab": 2, "L_tab": 844, "n_element_tab": 114, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 177, "formula_len_all_1": 176, "len_all": 160940, "len_all_1": 63320, "len_abs": 1974, "len_title": 124, "len_sents": 43241, "len_sents_1": 32645, "n_sents": 298, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1983, "title": "Large language models as windows on the mental structure of psychopathology", "abs": "How people represent the world determines how they act on it, as these internal representations bias what information is retrieved from memory, the inferences that are made and which actions are preferred. The structure of these representations are built through experience by extracting relevant information from the environment. Recent research has demonstrated that representational structure can also respond to the internal motives of agents, such as their aversion to uncertainty, which impacts their behavior. This opens the possibility to directly target internal structures to cause behavioral change in psychopathologies, one of the tenets of cognitive-behavioral therapy. For this purpose, it is crucial to understand how internal structures differ across psychopatologies. In this work, we show that Large Language Models (LLMs) could be viable tool to infer structural differences linked to distinct psychopathologies. We first demonstrate that we can reliably prompt LLMs to generate (verbal) behavior that can be detected as psychopathological by standard clinical assessment questionnaires. Next, we show that such prompting can capture correlational structure between the scores of diagnostic questionnaires observed in human data. We then analyze the lexical output patterns of LLMs (a proxy of their internal representations) induced with distinct psychopathologies. This analysis allows us to generate several empirical hypotheses on the link between mental representation and psychopathologies. Finally, we illustrate the usefulness of our approach in a case study involving data from Schizophrenic patients. Specifically, we show that these patients and LLMs prompted to exhibit behavior related to schizophrenia generate qualitatively similar semantic structures. We suggest that our novel computational framework could expand our understanding of psychopathologies by creating novel research hypotheses, which might eventually lead to novel diagnostic tools.", "keywords": [ "LLMs", "psychopathology", "mental structure", "computational psychiatry" ], "rating_list": [ 5, 5, 6, 1 ], "soundness_list": [ 1, 3, 3, 1 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "UWuTZYPSxJ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Buyun Liang", "gender": "Male", "institution": "University of Pennsylvania ", "country": "US", "position": "PhD student" }, { "name": "Darshan Thaker", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Jinqi Luo", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Kwan Ho Ryan Chan", "gender": "Male", "institution": "University of Pennsylvania ", "country": "US", "position": "PhD student" }, { "name": "Rene Vidal", "gender": "unknown", "institution": "University of Pennsylvania", "country": "US", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 23, "n_ref": 94, "n_ref_all": 117, "n_fig": 10, "n_tab": 5, "L_tab": 1209, "n_element_tab": 203, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 335, "formula_len_all_1": 296, "len_all": 146948, "len_all_1": 62910, "len_abs": 1528, "len_title": 112, "len_sents": 45367, "len_sents_1": 31931, "n_sents": 320, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1466, "title": "KDA: A Knowledge-Distilled Attacker for Scalable LLM Red Teaming", "abs": "Jailbreak attacks exploit specific prompts to bypass LLM safeguards and generate harmful or inappropriate content. Recently, numerous approaches have emerged for generating jailbreak attacks across diverse malicious scenarios. However, these methods often suffer from critical limitations such as the reliance on handcrafted prompts, the necessity for white-box access to target LLMs, the generation of monotonous prompts, or the dependence on expensive queries to commercial LLMs. Moreover, these methods typically require considerable time to generate jailbreak attacks. In this paper, we propose a Knowledge-Distilled Attacker (KDA) that leverages existing realistic and semantically meaningful prompts to learn a model that efficiently produces successful attacks. Specifically, we finetune an open-source LLM on a diverse set of attack prompts, enabling our framework to automatically generate black-box, coherent, and diverse attack prompts independent of commercial LLMs. Our KDA achieves a 100% success rate on multiple state-of-the-art LLMs while only requiring less than 10 seconds per attack generation. Further, using KDA, we introduce the RedTeam-10k dataset, a large-scale dataset of 10,000 harmful attack prompts inducing malicious LLM behavior spanning 12 categories such as bias, hate, and illegal activities. This dataset is 20x larger than any existing attack prompt dataset, positioning KDA as a powerful tool for large-scale adversarial testing.", "keywords": [ "Jailbreak attack; Large Language Models; Adversarial Attack; Red Teaming" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "UWdPsY7agk", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Jianing Chu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "PULAK GHOSH", "gender": "Male", "institution": "Indian Institute of Management", "country": "", "position": "Full Professor" }, { "name": "Shu Yang", "gender": "Female", "institution": "North Carolina State University", "country": "US", "position": "Associate Professor" }, { "name": "Wenbin Lu", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Full Professor" } ], "n_formula": 85, "n_formula_1": 34, "n_ref_uni": 47, "n_ref": 67, "n_ref_all": 89, "n_fig": 2, "n_tab": 1, "L_tab": 617, "n_element_tab": 51, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 617, "n_element_tab_1": 51, "formula_len_all": 12004, "formula_len_all_1": 3430, "len_all": 140417, "len_all_1": 66879, "len_abs": 952, "len_title": 121, "len_sents": 36991, "len_sents_1": 28134, "n_sents": 372, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 956, "title": "Efficient Causal Decision Making with One-sided Feedback", "abs": "We study a class of decision-making problems with one-sided feedback, where outcomes are only observable for specific actions. A typical example is bank loans, where the repayment status is known only if a loan is approved and remains undefined if rejected. In such scenarios, conventional approaches to causal decision evaluation and learning from observational data are not directly applicable. In this paper, we introduce a novel value function to evaluate decision rules that addresses the issue of undefined counterfactual outcomes. Without assuming no unmeasured confounders, we establish the identification of the value function using shadow variables. Furthermore, leveraging semiparametric theory, we derive the efficiency bound for the proposed value function and develop efficient methods for decision evaluation and learning. Numerical experiments and a real-world data application demonstrate the empirical performance of our proposed methods.", "keywords": [ "semiparametric efficiency", "one-sided feedback", "causal decision making" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 4 ], "confidence_list": [ 2, 2, 2, 3 ] }, { "paper_id": "UWOQ6w5yvX", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Bingli Wang", "gender": "Male", "institution": "Sichuan Agricultural University", "country": "CN", "position": "Undergrad student" }, { "name": "Jiao Li", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Mengzhu Wang", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Shanshan Wang", "gender": "Female", "institution": "Anhui University", "country": "CN", "position": "Associate Professor" }, { "name": "Yin Nan", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Researcher" }, { "name": "Zeheng Wang", "gender": "Male", "institution": "Northeast Normal University", "country": "CN", "position": "MS student" }, { "name": "Houcheng Su", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 26, "n_ref_uni": 25, "n_ref": 46, "n_ref_all": 57, "n_fig": 5, "n_tab": 6, "L_tab": 3808, "n_element_tab": 281, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 3179, "n_element_tab_1": 283, "formula_len_all": 1303, "formula_len_all_1": 1303, "len_all": 113929, "len_all_1": 57624, "len_abs": 1243, "len_title": 120, "len_sents": 29120, "len_sents_1": 22081, "n_sents": 208, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1248, "title": "AGLP: A Graph Learning Perspective for Semi-supervised Domain Adaptation", "abs": "In semi-supervised domain adaptation (SSDA), the model aims to leverage partially labeled target domain data along with a large amount of labeled source domain data to enhance its generalization capability for the target domain. A key advantage of SSDA is its ability to significantly reduce reliance on labeled data, thereby lowering the costs and time associated with data preparation. Most existing SSDA methods utilize information from domain labels and class labels but overlook the structural information of the data. To address this issue, this paper proposes a graph learning perspective (AGLP) for semi-supervised domain adaptation. We apply the graph convolutional network to the instance graph which allow structural information to propagate along the weighted graph edges. The proposed AGLP model has several advantages. First, to the best of our knowledge, this is the first work to model structural information in SSDA. Second, the proposed model can effectively learn domain-invariant and semantic representations, reducing domain discrepancies in SSDA. Extensive experimental results on multiple standard benchmarks demonstrate that the proposed AGLP algorithm outperforms state-of-the-art semi-supervised domain adaptation methods.", "keywords": [ "Semi-Supervised Domain Adaptation", "Graph", "Transfer learning" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "UW0zetsx8X", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Arun Verma", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "SG", "position": "Postdoc" }, { "name": "Bryan Kian Hsiang Low", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Associate Professor" }, { "name": "Patrick Jaillet", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "See-Kiong Ng", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Xiaoqiang Lin", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zhongxiang Dai", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 48, "n_ref": 112, "n_ref_all": 146, "n_fig": 8, "n_tab": 6, "L_tab": 2829, "n_element_tab": 174, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 2054, "n_element_tab_1": 128, "formula_len_all": 239, "formula_len_all_1": 361, "len_all": 175504, "len_all_1": 73448, "len_abs": 1502, "len_title": 87, "len_sents": 54189, "len_sents_1": 33829, "n_sents": 401, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1523, "title": "Prompt Optimization with Human Feedback", "abs": "Large language models (LLMs) have demonstrated remarkable performances in various tasks. However, the performances of LLMs heavily depend on the input prompt. This has given rise to a number of recent works on prompt optimization. However, the previous works often require the availability of a numeric score to assess the quality of every prompt. Unfortunately, when a human user interacts with a black-box LLM, it is often infeasible and unreliable to attain such a score. Instead, it is usually significantly easier and more reliable to obtain preference feedback from a human user, i.e., showing the user the responses generated from a pair of prompts and asking the user which one is preferred. Therefore, in this paper, we study the problem of prompt optimization with human feedback (POHF), in which we aim to optimize the prompt for a black-box LLM using only human preference feedback. By drawing inspirations from dueling bandits, we design a theoretically principled strategy to select a pair of prompts to query for preference feedback in every iteration, and hence introduce our algorithm named automated POHF (APOHF). We apply our APOHF algorithm to a variety of tasks, including optimizing user instructions, prompt optimization for text-to-image generative models, and response optimization with human feedback (i.e., further refining the response using a variant of our APOHF). The results demonstrate that our APOHF can efficiently find a good prompt using a small number of preference feedback instances.", "keywords": [ "Prompt Optimization", "Large Language Model", "Preference Feedback" ], "rating_list": [ 6, 3, 3, 6, 5 ], "soundness_list": [ 2, 3, 1, 2, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 1, 3, 3 ], "confidence_list": [ 5, 5, 4, 5, 4 ] }, { "paper_id": "UVnD9Ze6mF", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Andy Zhou", "gender": "Male", "institution": "Department of Computer Science, University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Dawn Song", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Jeffrey Ziwei Tan", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Kevin Klyman", "gender": "Male", "institution": "Stanford University", "country": "", "position": "MS student" }, { "name": "Minzhou Pan", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "PhD student" }, { "name": "Percy Liang", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Ruoxi Jia", "gender": "unknown", "institution": "Virginia Tech", "country": "US", "position": "Assistant Professor" }, { "name": "Yi Zeng", "gender": "Male", "institution": "Virginia Tech", "country": "", "position": "PhD student" }, { "name": "Yifan Mai", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "", "position": "Researcher" }, { "name": "Yu Yang", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yuheng Tu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Intern" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 99, "n_ref_all": 129, "n_fig": 13, "n_tab": 5, "L_tab": 5871, "n_element_tab": 2545, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 252392, "len_all_1": 55134, "len_abs": 1565, "len_title": 137, "len_sents": 80555, "len_sents_1": 28423, "n_sents": 537, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 93, "L_abs": 1394, "title": "AIR-BENCH 2024: A Safety Benchmark based on Regulation and Policies Specified Risk Categories", "abs": "Foundation models (FMs) provide societal benefits but also amplify risks. Governments, companies, and researchers have proposed regulatory frameworks, acceptable use policies, and safety benchmarks in response. However, existing public benchmarks often define safety categories based on previous literature, intuitions, or common sense, leading to disjointed sets of categories for risks specified in recent regulations and policies, which makes it challenging to evaluate and compare FMs across these benchmarks. To bridge this gap, we introduce AIR-BENCH 2024, the first AI safety benchmark aligned with emerging government regulations and company policies, following the regulation-based safety categories grounded in the AI Risks taxonomy, AIR 2024. AIR 2024 decomposes 8 government regulations and 16 company policies into a four-tiered safety taxonomy with 314 granular risk categories in the lowest tier. AIR-BENCH 2024 contains 5,694 diverse prompts spanning these categories, with manual curation and human auditing to ensure quality. We evaluate leading language models on AIR-BENCH 2024 uncovering insights into their alignment with specified safety concerns. By bridging the gap between public benchmarks and practical AI risks, AIR-BENCH 2024 provides a foundation for assessing model safety across jurisdictions, fostering the development of safer and more responsible AI systems.", "keywords": [ "AI Safety", "Regulation", "Policy", "Safety Alignment", "Foundation Models" ], "rating_list": [ 6, 10, 8, 6 ], "soundness_list": [ 2, 4, 4, 2 ], "presentation_list": [ 4, 4, 4, 2 ], "contribution_list": [ 3, 4, 4, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "UVaPEthRKx", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Jingyuan Chen", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Shengyu Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "ZJU100 Young Professor" }, { "name": "Yiyun Zhou", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zheqi Lv", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 21, "n_ref_uni": 36, "n_ref": 77, "n_ref_all": 98, "n_fig": 8, "n_tab": 7, "L_tab": 4839, "n_element_tab": 687, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 3436, "n_element_tab_1": 495, "formula_len_all": 1760, "formula_len_all_1": 1787, "len_all": 140552, "len_all_1": 68757, "len_abs": 1760, "len_title": 167, "len_sents": 38003, "len_sents_1": 25786, "n_sents": 276, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 119, "L_abs": 1857, "title": "Cuff-KT: Tackling Learners' Real-time Learning Pattern Adjustment via Tuning-Free Knowledge State-Guided Model Updating", "abs": "Knowledge Tracing (KT) is a core component of Intelligent Tutoring Systems, modeling learners' knowledge state to predict future performance and provide personalized learning support. Current KT models simply assume that training data and test data follow the same distribution. However, this is challenged by the continuous changes in learners' patterns. In reality, learners' patterns change irregularly at different stages ($e.g.$, different semesters) due to factors like cognitive fatigue and external stress. Additionally, there are significant differences in the patterns of learners from various groups ($e.g.$, different classes), influenced by social cognition, resource optimization, etc. We refer to these distribution changes at different stages and from different groups as intra-learner shift and inter-learner shift, respectively---a task introduced, which we refer to as Real-time Learning Pattern Adjustment (RLPA). Existing KT models, when faced with RLPA, lack sufficient adaptability, because they fail to timely account for the dynamic nature of different learners' evolving learning patterns. Current strategies for enhancing adaptability rely on retraining, which leads to significant overfitting and high time cost problem. To address this, we propose Cuff-KT, comprising a controller and a generator. The controller assigns value scores to learners, while the generator generates personalized parameters for selected learners. Cuff-KT adapts to distribution changes fast and flexibly without fine-tuning. Experiments on one classic and two latest datasets demonstrate that Cuff-KT significantly improves current KT models' performance under intra- and inter-learner shifts, with an average relative increase of 7\\% on AUC, effectively tackling RLPA. Our code and datasets are available at https://anonymous.4open.science/r/Cuff-KT.", "keywords": [ "Knowledge Tracing", "Online Education" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 4, 4, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 2, 4, 4, 5 ] }, { "paper_id": "UVaLZMv0uk", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Di Wang", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Dongxiao Yu", "gender": "unknown", "institution": "Shandong University", "country": "CN", "position": "Full Professor" }, { "name": "Falko Dressler", "gender": "Male", "institution": "Technische Universit\u00e4t Berlin", "country": "", "position": "Full Professor" }, { "name": "Xiuzhen Cheng", "gender": "Female", "institution": "Shandong University", "country": "CN", "position": "Full Professor" }, { "name": "Youming Tao", "gender": "unknown", "institution": "Technische Universit\u00e4t Berlin", "country": "DE", "position": "PhD student" } ], "n_formula": 162, "n_formula_1": 19, "n_ref_uni": 30, "n_ref": 95, "n_ref_all": 99, "n_fig": 0, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 177, "n_element_tab_1": 16, "formula_len_all": 15394, "formula_len_all_1": 1776, "len_all": 174438, "len_all_1": 63780, "len_abs": 4425, "len_title": 125, "len_sents": 49711, "len_sents_1": 29343, "n_sents": 431, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1655, "title": "Private Stochastic Optimization for Achieving Second-Order Stationary Points", "abs": "This paper addresses the challenge of achieving second-order stationary points (SOSP) in differentially private stochastic non-convex optimization. We identify two key limitations in the state-of-the-art: (i) inaccurate error rates caused by the omission of gradient variance in saddle point escape analysis, resulting in inappropriate parameter choices and overly optimistic performance estimates, and (ii) inefficiencies in private SOSP selection via the AboveThreshold algorithm, particularly in distributed learning settings, where perturbing and sharing Hessian matrices introduces significant additional noise. To overcome these challenges, we revisit perturbed stochastic gradient descent (SGD) with Gaussian noise and propose a new framework that leverages general gradient oracles. This framework introduces a novel criterion based on model drift distance, ensuring provable saddle point escape and efficient convergence to approximate local minima with low iteration complexity. Using an adaptive SPIDER as the gradient oracle, we establish a new DP algorithm that corrects existing error rates. Furthermore, we extend our approach to a distributed adaptive SPIDER, applying our framework to distributed learning scenarios and providing the first theoretical results on achieving SOSP under differential privacy in distributed environments with heterogeneous data. Finally, we analyze the limitations of the AboveThreshold algorithm for private model selection in distributed learning and show that as model dimensions increase, the selection process introduces additional errors, further demonstrating the superiority of our proposed framework.", "keywords": [ "Differential privacy", "non-convex optimization", "saddle points" ], "rating_list": [ 5, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 4, 4 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "UV5p3JZMjC", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Johannes Von Oswald", "gender": "Male", "institution": "Research, Google", "country": "CH", "position": "Researcher" }, { "name": "Seijin Kobayashi", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Yassir Akram", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "", "position": "PhD student" }, { "name": "Angelika Steger", "gender": "unknown", "institution": "ETH Zurich", "country": "", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 46, "n_ref": 61, "n_ref_all": 89, "n_fig": 12, "n_tab": 3, "L_tab": 2912, "n_element_tab": 107, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 941, "formula_len_all_1": 642, "len_all": 196138, "len_all_1": 76978, "len_abs": 1398, "len_title": 96, "len_sents": 61710, "len_sents_1": 36241, "n_sents": 486, "n_sents_1": 278, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 48, "L_abs": 1405, "title": "Learning Randomized Algorithms with Transformers", "abs": "Randomization is a powerful tool that endows algorithms with remarkable properties. For instance, randomized algorithms excel in adversarial settings, often surpassing the worst-case performance of deterministic algorithms with large margins. Furthermore, their success probability can be amplified by simple strategies such as repetition and majority voting. In this paper, we enhance deep neural networks, in particular transformer models, with randomization. We demonstrate for the first time that randomized algorithms can be instilled in transformers through learning, in a purely data- and objective-driven manner. First, we analyze known adversarial objectives for which randomized algorithms offer a distinct advantage over deterministic ones. We then show that common optimization techniques, such as gradient descent or evolutionary strategies, can effectively learn transformer parameters that make use of the randomness provided to the model. To illustrate the broad applicability of randomization in empowering neural networks, we study three conceptual tasks: associative recall, graph coloring, and agents that explore grid worlds. In addition to demonstrating increased robustness against oblivious adversaries through learned randomization, our experiments reveal remarkable performance improvements due to the inherently random nature of the neural networks' computation and predictions.", "keywords": [ "Randomized algorithms", "Learning under adversarial losses", "Adversarial robustness", "In-context learning algorithms" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 4, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "UV1jr2aJ2J", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Adel Nabli", "gender": "Male", "institution": "Universit\u00e9 Pierre et Marie Curie - Paris 6, Computer Science Lab - Pierre and Marie Curie University, Paris, France", "country": "FR", "position": "PhD student" }, { "name": "Edouard Oyallon", "gender": "Male", "institution": "CNRS", "country": "FR", "position": "Researcher" }, { "name": "Eugene Belilovsky", "gender": "Male", "institution": "Concordia University, Montreal", "country": "CA", "position": "Assistant Professor" }, { "name": "Louis Fournier", "gender": "Male", "institution": "Universit\u00e9 Pierre et Marie Curie - Paris 6, Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "", "position": "PhD student" }, { "name": "Louis Serrano", "gender": "Male", "institution": "Universit\u00e9 Pierre et Marie Curie - Paris 6, Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "", "position": "PhD student" }, { "name": "Pierre ERBACHER", "gender": "Male", "institution": "Universit\u00e9 Pierre et Marie Curie - Paris 6, Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 73, "n_ref": 146, "n_ref_all": 164, "n_fig": 12, "n_tab": 5, "L_tab": 1209, "n_element_tab": 157, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1441, "n_element_tab_1": 111, "formula_len_all": 828, "formula_len_all_1": 590, "len_all": 195755, "len_all_1": 64209, "len_abs": 1778, "len_title": 137, "len_sents": 36437, "len_sents_1": 29264, "n_sents": 275, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1831, "title": "ACCO: Accumulate while you Communicate, Hiding Communications in Distributed LLM Training", "abs": "Training Large Language Models (LLMs) relies heavily on distributed implementations, employing multiple GPUs to compute stochastic gradients on model replicas in parallel. However, synchronizing gradients in data parallel settings induces a communication overhead increasing with the number of distributed workers, impeding the efficiency gains of parallelization. To address this challenge, local optimization algorithms such as the ones used in Federated Learning have emerged. While effective in minimizing communication overhead, they incur significant memory costs, hindering scalability: in addition to extra momentum variables, optimizer's states cannot be partitioned among workers as communications are only allowed between rounds of local optimization steps. To conceal communication costs, we propose instead to synchronize delayed gradients *while* computing new ones between each model\u2019s update and introduce $\\textbf{AC}$cumulate while $\\textbf{CO}$mmunicate ($\\textbf{ACCO}$), a memory-efficient optimization algorithm tailored for distributed training of LLMs. Accumulating local gradients on the workers until the communication finishes naturally reduces the idle time of GPUs and even allows the use of heterogeneous hardware. However, we show that the one-step delay inherent in parallel execution of gradient computations and communications has drastic impacts on Transformers\u2019 convergence. To compensate this delay we introduce a novel technique which leads to training dynamics aligned with standard distributed optimization. Compared to ZeRO, our implementation and experiments on several LLMs pre-training and fine-tuning tasks demonstrates that $\\textbf{ACCO}$ reduces the learning time up to 87\\% and successfully allows both sharding optimizer states across workers and the use of heterogeneous hardware.", "keywords": [ "Distributed LLMs Training", "Data Parallelism", "Optimizer State Partitioning", "Distributed Optimization", "Decoupled Communication" ], "rating_list": [ 5, 6, 5, 6, 3 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 3, 1 ], "confidence_list": [ 4, 3, 3, 1, 4 ] }, { "paper_id": "UUwrBhhsxT", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Bill Zeng", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Huaze Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Qian Zhang", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Wenbo Ding", "gender": "Male", "institution": "Tsinghua Univeresity", "country": "CN", "position": "Associate Professor" }, { "name": "Zhenpeng Shi", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Wang Chao", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 55, "n_ref": 76, "n_ref_all": 94, "n_fig": 6, "n_tab": 5, "L_tab": 1385, "n_element_tab": 125, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 279, "formula_len_all_1": 62, "len_all": 188272, "len_all_1": 58950, "len_abs": 1454, "len_title": 135, "len_sents": 58746, "len_sents_1": 30143, "n_sents": 380, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1466, "title": "A Scalable Transformer-based Framework for Fault Detection in Mission-Critical Systems", "abs": "Detecting underlying faults is crucial in the development of mission-critical planning systems, such as UAV trajectory planning in Unmanned aircraft Traffic Management (UTM), which is vital to airspace safety. \nInevitably, there exists a small set of rare, unpredictable conditions where the UTM could suffer from catastrophic failures. \nMost traditional fault detection approaches focus on achieving high coverage by random input exploitation. \nHowever, random methods are struggling to detect long-tail vulnerabilities with unacceptable time consumption. \nTo tackle this challenge, we propose a scenario-oriented framework to search long-tail conditions, accelerating the fault detection process. \nInspired by in-context learning approaches, we leverage a Transformer-based policy model to capture the dynamics of the subject UTM system from the offline dataset for exploitation acceleration. \nWe evaluate our approach over 700 hours in a massive-scale, industry-level simulation environment. \nEmpirical results demonstrate that our approach achieves over 8 times more vulnerability discovery efficiency compared with traditional expert-guided random-walk exploitation, which showcases the potential of machine learning for fortifying mission-critical systems. \nFurthermore, we scale the model size to 2 billion parameters, achieving substantial performance gains over smaller models in offline and online evaluations, highlighting the scalability of our approach.", "keywords": [ "multi-agent systems", "unmanned aerial vehicle", "offline reinforcement learning", "scenario-based testing" ], "rating_list": [ 5, 3, 5, 8 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "UUuTFhrWpM", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Changsheng Zhao", "gender": "Male", "institution": "Meta Inc.", "country": "US", "position": "Researcher" }, { "name": "Daniel Agyei Asante", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "PhD student" }, { "name": "Ernie Chang", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Research Scientist" }, { "name": "Vikas Chandra", "gender": "Male", "institution": "Meta", "country": "US", "position": "Director, AI" }, { "name": "Yang Li", "gender": "unknown", "institution": "Iowa State University", "country": "US", "position": "Assistant Professor" }, { "name": "Yangyang Shi", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 16, "n_ref": 44, "n_ref_all": 60, "n_fig": 9, "n_tab": 5, "L_tab": 1590, "n_element_tab": 200, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 614, "n_element_tab_1": 18, "formula_len_all": 722, "formula_len_all_1": 722, "len_all": 74484, "len_all_1": 53846, "len_abs": 1139, "len_title": 103, "len_sents": 23726, "len_sents_1": 23243, "n_sents": 178, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1143, "title": "Basel: Target-Aware Basis Selection for Language Models", "abs": "As the size of language models increases, they deliver substantial performance improvements across a variety of applications. However, this growth also leads to greater computational demands, making deployment on resource-constrained devices\u2014such as personal computers and mobile or wearable devices\u2014more challenging, and significantly raising inference costs on cloud servers. To address these challenges, we introduce a method to streamline language models. We observe that language models pretrained on general datasets often include redundant components that are unnecessary for particular tasks. Our approach identifies and removes these redundant parts, retaining only the essential components for the intended applications. Specifically, we represent the weight matrices of language models as a linear combination of base components, eliminate the irrelevant bases, and introduce new bases that enhance performance for target tasks. Evaluations show that our method reduces model size much more significantly\u2014by up to 1.7 times\u2014while maintaining similar accuracy, compared to state-of-the-art techniques, across a range of applications.", "keywords": [ "language models", "interpretability" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "UUZuwDv8iw", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "AJAY KUMAR JAISWAL", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Jianyu Wang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Pingzhi Li", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Ruoming Pang", "gender": "unknown", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Xianzhi Du", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Yixiao Li", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zhangyang Wang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Chong Wang", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 16, "n_ref_uni": 51, "n_ref": 92, "n_ref_all": 111, "n_fig": 6, "n_tab": 9, "L_tab": 2603, "n_element_tab": 309, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 3050, "n_element_tab_1": 395, "formula_len_all": 974, "formula_len_all_1": 1172, "len_all": 179033, "len_all_1": 70389, "len_abs": 1874, "len_title": 168, "len_sents": 44164, "len_sents_1": 31088, "n_sents": 228, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Reject", "L_title": 121, "L_abs": 1888, "title": "Fantastic Experts and How to Find Them: A Multi-Dimensional Study for Experts-Level Sparsification in Mixture-of-Experts", "abs": "Sparsely activated Mixture-of-Experts (SMoE) has shown promise in scaling up the learning capacity of neural networks. However, vanilla SMoEs have issues such as expert redundancy and heavy memory requirements, making them inefficient and non-scalable, especially for resource-constrained scenarios. Expert-level sparsification of SMoEs involves pruning the least important experts to address these limitations. In this work, we aim to address three questions: (1) What is the best recipe across multiple plausible recipes to identify the least knowledgeable subset of experts that can be dropped to achieve a desired sparsity level? (2) How should we perform expert dropping (one-shot or iterative), and what correction measures can we undertake to minimize its drastic impact on SMoE subnetwork capabilities? (3) What capabilities of full-SMoEs are severely impacted by the removal of the least dominant experts, and how can we recover them? Firstly, we propose MoE Experts Compression Suite (MC-Suite), which is a collection of some previously explored and multiple novel recipes to provide a comprehensive benchmark for estimating expert importance from diverse perspectives, as well as unveil numerous valuable insights for SMoE experts. Secondly, unlike prior works with a one-shot expert pruning approach, we explore the benefits of iterative pruning with the re-estimation of the MC-Suite criterion. Moreover, we introduce the benefits of task-agnostic fine-tuning as a correction mechanism during progressive expert dropping, which we term MoE Lottery Subnetworks. Lastly, we present an experimentally validated conjecture that, during expert dropping, SMoEs' instruction-following capabilities are predominantly hurt, which can be restored to a robust level subject to external augmentation of instruction-following capabilities using k-shot examples and supervised fine-tuning.", "keywords": [ "Mixture of Experts; Compression; Expert Pruning; Efficiency" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "UUNTAwJIIn", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Jing Wang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Jinhui Tang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Shengping Zhang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xiangyang Ji", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoyu Han", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 64, "n_ref": 122, "n_ref_all": 148, "n_fig": 9, "n_tab": 4, "L_tab": 832, "n_element_tab": 102, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 780, "n_element_tab_1": 98, "formula_len_all": 306, "formula_len_all_1": 141, "len_all": 146685, "len_all_1": 65080, "len_abs": 1527, "len_title": 130, "len_sents": 39017, "len_sents_1": 31719, "n_sents": 259, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1537, "title": "Rethinking Brain-to-Image Reconstruction: What Should We Decode from fMRI Signals?", "abs": "Recently, notable advancements have been achieved in brain-to-image reconstruction. However, the assumption that the recorded brain activities faithfully mirror the complete high-resolution images conflicts with the workings of human vision and cognitive systems. \nIn this study, we present a novel approach, fMRI-to-foveated image (FitFovea), which redefines the brain-to-image reconstruction process to better align with cognitive science principles. FitFovea comprises three key stages: pseudo-foveated image synthesis, fMRI-to-foveated image reconstruction and stimulus image generation. In the first stage, FitFovea constructs new {fMRI, pseudo-foveated image} pairs from existing fMRI-image data using saliency prediction and foveated rendering techniques. Next, during the foveated image reconstruction phase, the information captured by human vision is decoded from fMRI signals with maximum accuracy. The final stage, stimulus image generation, is considered not as a strict reconstruction but rather as a postprocessing step. This stage is akin to existing brain-to-image decoding methods, which often emphasize semantic fidelity rather than pixel-level reconstruction. To validate our approach, we introduce the brain score metric to quantify the correlation between images and corresponding brain responses. The superior results validate the rationale behind decoding pseudo-foveated images from fMRI data and demonstrate the feasibility of our newly-devised pipeline based on synthesized pseudo-foveated image training data.", "keywords": [ "neural decoding", "brain-to-image reconstruction", "pseudo-foveated image synthesis", "fMRI-to-foveated image decoding" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "UTnq6hJJYa", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Bo-Lin Wang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "Jiang-Xin Shi", "gender": "unknown", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Min-Ling Zhang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Full Professor" }, { "name": "Tong Wei", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Yu-Feng Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 11, "n_ref_uni": 49, "n_ref": 99, "n_ref_all": 124, "n_fig": 5, "n_tab": 16, "L_tab": 21731, "n_element_tab": 4340, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 4294, "n_element_tab_1": 826, "formula_len_all": 535, "formula_len_all_1": 584, "len_all": 224565, "len_all_1": 70833, "len_abs": 1236, "len_title": 121, "len_sents": 39914, "len_sents_1": 25375, "n_sents": 304, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1242, "title": "Harnessing Shallow Features in Pre-Trained Models for Out-of-Distribution Detection", "abs": "Recognizing out-of-distribution (OOD) samples is essential for deploying robust machine learning systems in the open-world environments. Conventional OOD detection approaches rely on feature representations from the final layer of neuron networks, often neglecting the rich information encapsulated in shallow layers. Leveraging the strengths of transformer-based architectures, we introduce an attention-based fusion module, which dynamically assigns importance weights to representations learned by each Transformer layer and detects OOD samples using the Mahalanobis distance. Compared to existing approaches, our method enables a lightweight fine-tuning of pre-trained models, and retains all feature representations that are beneficial to the OOD detection. We also thoroughly study various parameter-efficient fine-tuning strategies. Our experiments show the benefit of using shallow features, and demonstrate the influence of different Transformer layers. We fine-tune pre-trained models in both class-balanced and long-tailed in-distribution classification tasks, and show that our method achieves state-of-the-art OOD detection performance averaged across nine OOD datasets. The source code is provided in the supplementary material.", "keywords": [ "out-of-distribution detection", "long-tail learning" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "UTP6bdFlkB", "primary_area": "applications to neuroscience & cognitive science", "n_author": 7, "author_info_list": [ { "name": "Ahmed Abdulaal", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "An Zhao", "gender": "unknown", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Anna Schroder", "gender": "unknown", "institution": "University College London, University of London", "country": "", "position": "Postdoc" }, { "name": "Daniel C. Alexander", "gender": "Male", "institution": "University College London", "country": "", "position": "Full Professor" }, { "name": "Elinor Thompson", "gender": "Female", "institution": "University College London, University of London", "country": "GB", "position": "Postdoc" }, { "name": "Frederik Barkhof", "gender": "unknown", "institution": "University College London, University of London", "country": "GB", "position": "Full Professor" }, { "name": "Tiantian He", "gender": "Female", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 27, "n_ref": 50, "n_ref_all": 65, "n_fig": 8, "n_tab": 3, "L_tab": 946, "n_element_tab": 60, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 569, "n_element_tab_1": 51, "formula_len_all": 662, "formula_len_all_1": 605, "len_all": 154027, "len_all_1": 63901, "len_abs": 1842, "len_title": 105, "len_sents": 51944, "len_sents_1": 31484, "n_sents": 310, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1849, "title": "LLM-guided spatio-temporal disease progression modelling", "abs": "Understanding the interactions between biomarkers across brain regions during disease progression is essential for unravelling the mechanisms underlying neurodegenerative disease. For example, in Alzheimer's Disease (AD) and other neurodegenerative conditions, there are typically two kinds of methods to contract disease trajectory. Existing mechanistic models describe how variables interact with each other spatiotemporally within a dynamical system driven by an underlying biological substrate often based on brain connectivity. However, such methods typically grossly oversimplify the complex relationship between brain connectivity and brain pathology appearance and propagation. Meanwhile, pure data-driven approaches for inferring these relationships from time series face challenges with convergence, identifiability, and interpretability. We present a novel framework that bridges this gap by using Large Language Models (LLMs) as expert guides to learn disease progression from irregular longitudinal patient data. Our method simultaneously optimizes two components: 1) estimating the temporal positioning of patient data along a common disease trajectory and 2) discovering the graph structure that captures spatiotemporal relationships between brain regions. By leveraging multiple LLMs as domain experts, our approach achieves faster convergence, improved stability, and better interpretability compared to existing methods. When applied to modelling tau-pathology propagation in the brain, our framework demonstrates superior prediction accuracy while revealing additional disease-driving factors beyond traditional connectivity measures. This work represents the first application of LLM-guided graph learning for modelling neurodegenerative disease progression in the brain from cross-sectional and short longitudinal imaging data.", "keywords": [ "disease progression modelling", "LLM", "spatio temporal modelling" ], "rating_list": [ 6, 5, 1, 3 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 4, 5, 5 ] }, { "paper_id": "UT5B7fktaw", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Daria Diatlova", "gender": "Female", "institution": "VK", "country": "RU", "position": "Researcher" }, { "name": "Egor Spirin", "gender": "Male", "institution": "VK", "country": "RU", "position": "Researcher" }, { "name": "Nikita Balagansky", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "", "position": "MS student" }, { "name": "Varlamov Alexander", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "Undergrad student" }, { "name": "Vitalii Shutov", "gender": "Male", "institution": "ITMO University", "country": "", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 25, "n_ref": 62, "n_ref_all": 83, "n_fig": 7, "n_tab": 8, "L_tab": 3273, "n_element_tab": 405, "n_fig_1": 4, "n_tab_1": 9, "L_tab_1": 1727, "n_element_tab_1": 164, "formula_len_all": 778, "formula_len_all_1": 280, "len_all": 125655, "len_all_1": 61548, "len_abs": 2846, "len_title": 142, "len_sents": 35577, "len_sents_1": 26363, "n_sents": 286, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 978, "title": "Variational Inference for Self-Supervised Speech Models Fine-tuning on Downstream Tasks", "abs": "Despite the growing interest in self-supervised speech models, recent research has primarily focused on modifying upstream model architectures and pretraining techniques, with less attention given to how features from self-supervised models are used. In this paper, we explore the use of variational inference to enhance the performance of self-supervised audio models in downstream tasks. We hypothesize that adaptively reweighting the outputs of the model layers is crucial to improving performance on these tasks. We extensively evaluate our method alongside widely used baselines, demonstrating that understanding sample-specific information is essential for improved performance on several tasks. Our proposed method surpasses existing approaches and generalizes to various speech tasks, including automatic speech recognition, speaker verification, and emotion recognition. Finally, we analyze our method to provide deeper insight into the importance of our modifications.", "keywords": [ "SSL models", "Fine-tuning", "Variational Inference", "SER", "ASR", "SV" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "USI3ZbuFaV", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Bowei He", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Chen Ma", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Hui-Ling Zhen", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Principal Researcher" }, { "name": "Jianping Zhang", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Lanqing HONG", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Lihao Yin", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" } ], "n_formula": 55, "n_formula_1": 23, "n_ref_uni": 38, "n_ref": 81, "n_ref_all": 99, "n_fig": 1, "n_tab": 14, "L_tab": 7303, "n_element_tab": 508, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 2804, "n_element_tab_1": 164, "formula_len_all": 2840, "formula_len_all_1": 918, "len_all": 254569, "len_all_1": 71601, "len_abs": 1361, "len_title": 164, "len_sents": 89237, "len_sents_1": 32366, "n_sents": 668, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 116, "L_abs": 1383, "title": "Certifying Language Model Robustness with Fuzzed Randomized Smoothing: An Efficient Defense Against Backdoor Attacks", "abs": "The widespread deployment of pre-trained language models (PLMs) has exposed them to textual backdoor attacks, particularly those planted during the pre-training stage. These attacks pose significant risks to high-reliability applications, as they can stealthily affect multiple downstream tasks. While certifying robustness against such threats is crucial, existing defenses struggle with the high-dimensional, interdependent nature of textual data and the lack of access to original poisoned pre-training data. To address these challenges, we introduce **F**uzzed **R**andomized **S**moothing (**FRS**), a novel approach for efficiently certifying language model robustness against backdoor attacks. FRS integrates software robustness certification techniques with biphased model parameter smoothing, employing Monte Carlo tree search for proactive fuzzing to identify vulnerable textual segments within the Damerau-Levenshtein space. This allows for targeted and efficient text randomization, while eliminating the need for access to poisoned training data during model smoothing. Our theoretical analysis demonstrates that FRS achieves a broader certified robustness radius compared to existing methods. Extensive experiments across various datasets, model configurations, and attack strategies validate FRS's superiority in terms of defense efficiency, accuracy, and robustness.", "keywords": [ "Language Model", "Textual Backdoor Attack", "Certified Robustness", "Fuzzed Randomized Smoothing" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "USGY5t7fwG", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Dongnan Liu", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Postdoc" }, { "name": "Jianan Fan", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Peng Fu", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Qiang Qu", "gender": "Male", "institution": "The University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Weidong Cai", "gender": "Male", "institution": "The University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Yu Yao", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Lecturer" }, { "name": "Zhuonan Liang", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 21, "n_ref_uni": 47, "n_ref": 102, "n_ref_all": 115, "n_fig": 4, "n_tab": 7, "L_tab": 2678, "n_element_tab": 314, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1526, "n_element_tab_1": 235, "formula_len_all": 2178, "formula_len_all_1": 1434, "len_all": 184952, "len_all_1": 57151, "len_abs": 1537, "len_title": 108, "len_sents": 44470, "len_sents_1": 22705, "n_sents": 401, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1545, "title": "Object-aware Conditional Alignment for Cross-domain Counting", "abs": "Object counting is an important task in computer vision with many real-world applications. In practical settings, factors such as lighting conditions and object density can vary dramatically, leading to distribution shifts then causing inaccurate counting. We found that existing domain adaptation (DA) methods cannot be directly applied to the counting task, as they usually assume changes across different domains are task-irrelevant and focus on utilizing domain-invariant features for prediction. However, in object counting tasks, changes in object density which could happen across domains are task-relevant and cannot be ignored. Therefore, applying existing DA methods to the counting task can ignore the information about density changes, resulting in unreliable counting. To address this limitation, we propose the Binary Alignment Network (BiAN). Unlike traditional DA methods that align distributions of entire image representations, BiAN segments objects of interest and aligns the distributions of the object-specific features across domains. This targeted alignment allows us to disregard irrelevant features, such as lighting conditions, while preserving essential information about changes in object density. We theoretically demonstrate that BiAN achieves superior adaptability in counting tasks by introducing conditional alignment\u2014aligning features conditioned on the presence of objects. Extensive experiments on two distinct counting tasks and eight dataset combinations show that BiAN outperforms state-of-the-art methods.", "keywords": [ "Counting", "domain adaptation" ], "rating_list": [ 5, 3, 6, 5, 3 ], "soundness_list": [ 2, 2, 4, 2, 2 ], "presentation_list": [ 2, 2, 3, 1, 2 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 3, 5, 2, 2, 4 ] }, { "paper_id": "USE9akheEY", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Chao Huang", "gender": "Male", "institution": "University of Southampton", "country": "", "position": "Associate Professor" }, { "name": "Francesco Faccio", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "J\u00fcrgen Schmidhuber", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Director, AI Initiative" }, { "name": "Qingyuan Wu", "gender": "Male", "institution": "University of Southampton", "country": "GB", "position": "PhD student" }, { "name": "Weida Li", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yuhui Wang", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" }, { "name": "Dylan Ashley", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 33, "n_ref": 77, "n_ref_all": 123, "n_fig": 17, "n_tab": 10, "L_tab": 2615, "n_element_tab": 239, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 623, "formula_len_all_1": 635, "len_all": 147078, "len_all_1": 64950, "len_abs": 1206, "len_title": 126, "len_sents": 48981, "len_sents_1": 31974, "n_sents": 363, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1227, "title": "Scaling Value Iteration Networks to 5000 Layers for Extreme Long-Term Planning", "abs": "The Value Iteration Network (VIN) is an end-to-end differentiable architecture that performs value iteration on a latent Markov Decision Process (MDP) for planning in reinforcement learning (RL). However, VINs struggle to scale to long-term and large-scale planning tasks, such as navigating a $100\\times 100$ maze---a task that typically requires thousands of planning steps to solve. We observe that this deficiency is due to two issues: the representation capacity of the latent MDP and the planning module's depth. We address these by augmenting the latent MDP with a dynamic transition kernel, dramatically improving its representational capacity, and, to mitigate the vanishing gradient problem, introduce an \"adaptive highway loss\" that constructs skip connections to improve gradient flow. We evaluate our method on 2D maze navigation environments, the ViZDoom 3D navigation benchmark, and the real-world Lunar rover navigation task. We find that our new method, named \\textit{Dynamic Transition VIN (DT-VIN)}, scales to 5000 layers and solves challenging versions of the above tasks. Altogether, we believe that DT-VIN represents a concrete step forward in performing long-term large-scale planning in RL environments.", "keywords": [ "Value Iteration Networks", "Long-term Planning", "Reinforcement Learning", "Deep Neural Network" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "US9k5TXVLZ", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Bei Yu", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Haisheng Zheng", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Haoyuan WU", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yuan Pu", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 13, "n_ref_uni": 41, "n_ref": 106, "n_ref_all": 139, "n_fig": 5, "n_tab": 8, "L_tab": 5044, "n_element_tab": 592, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1224, "n_element_tab_1": 177, "formula_len_all": 405, "formula_len_all_1": 488, "len_all": 154324, "len_all_1": 69126, "len_abs": 1482, "len_title": 131, "len_sents": 46531, "len_sents_1": 31172, "n_sents": 323, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1549, "title": "Circuit Representation Learning with Masked Gate Modeling and Verilog-AIG Alignment", "abs": "Understanding the structure and function of circuits is crucial for electronic design automation (EDA). \nCircuits can be formulated as And-Inverter graphs (AIGs), enabling efficient implementation of representation learning through graph neural networks (GNNs).\nMasked modeling paradigms have been proven effective in graph representation learning.\nHowever, masking augmentation to original circuits will destroy their logical equivalence, which is unsuitable for circuit representation learning.\nMoreover, existing masked modeling paradigms often prioritize structural information at the expense of abstract information such as circuit function.\nTo address these limitations, we introduce MGVGA, a novel constrained masked modeling paradigm incorporating masked gate modeling (MGM) and Verilog-AIG alignment (VGA).\nSpecifically, MGM preserves logical equivalence by masking gates in the latent space rather than in the original circuits, subsequently reconstructing the attributes of these masked gates.\nMeanwhile, large language models (LLMs) have demonstrated an excellent understanding of the Verilog code functionality.\nBuilding upon this capability, VGA performs masking operations on original circuits and reconstructs masked gates under the constraints of equivalent Verilog codes, enabling GNNs to learn circuit functions from LLMs.\nWe evaluate MGVGA on various logic synthesis tasks for EDA and show the superior performance of MGVGA compared to previous state-of-the-art methods. \nOur code is available at https://github.com/wuhy68/MGVGA.", "keywords": [ "circuit representation learning", "masked graph modeling", "large language models", "multimodal alignment" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "US2UCMvzvP", "primary_area": "foundation or frontier models, including LLMs", "n_author": 17, "author_info_list": [ { "name": "Chang Su", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Hao Yang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Jiahuan Li", "gender": "unknown", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Jiajun Chen", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "Full Professor" }, { "name": "Jiaxin GUO", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Min Zhang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Ming Zhu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Shuaijie She", "gender": "unknown", "institution": "nanjing university", "country": "", "position": "PhD student" }, { "name": "Shujian Huang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Wei Zou", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Xiang Geng", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "PhD student" }, { "name": "Xiaofeng Zhao", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Xinglin Lyu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yanqing Zhao", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Yinglu Li", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Yuang Li", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zhejian Lai", "gender": "Male", "institution": "nanjing university", "country": "", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 34, "n_ref": 77, "n_ref_all": 101, "n_fig": 2, "n_tab": 13, "L_tab": 3280, "n_element_tab": 347, "n_fig_1": 2, "n_tab_1": 11, "L_tab_1": 2521, "n_element_tab_1": 285, "formula_len_all": 20, "formula_len_all_1": 20, "len_all": 163648, "len_all_1": 76509, "len_abs": 1558, "len_title": 108, "len_sents": 39777, "len_sents_1": 31220, "n_sents": 354, "n_sents_1": 284, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1567, "title": "Why Not Transform Chat Large Language Models to Non-English?", "abs": "Large language models (LLMs) excel in various tasks, but their performance in non-English languages remains limited due to imbalanced training data. To address this limitation, we explore how to transform chat LLMs to non-English. Chat LLMs offer more advanced capabilities than base LLMs, such as multi-turn conversation and alignment with human preferences. However, transforming chat LLMs presents greater challenges than base LLMs. First, how can we effectively transfer advanced capabilities without their supervised data in target languages? Second, how can we prevent the original capabilities from catastrophic forgetting without replaying their training procedure in English? We target these issues by introducing a simple framework called TransLLM. TransLLM divides the transfer problem into some common sub-tasks with the translation chain-of-thought, eliminating the need for complex training data. More importantly, TransLLM uses two key strategies to prevent catastrophic forgetting: Low-rank adaptation, which preserves the original LLM parameters during training, and recovery KD, which utilizes data generated by the chat LLM itself to recover the original knowledge from the frozen parameters. Experiments conducted across five languages and three LLMs demonstrate the superiority of TransLLM. Notably, TransLLM outperforms GPT-4 in Thai, demonstrating higher levels of helpfulness and safety, using just 8B parameters and publicly accessible data. Our analysis demonstrates how recovery KD combined with LoRA helps mitigate catastrophic forgetting.", "keywords": [ "Large Language Model", "Low Resource Languages", "Knowledge Transfer", "Catastrophic Forgetting" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "URPwT55i6O", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Jasper Dekoninck", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Martin Vechev", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Maximilian Baader", "gender": "unknown", "institution": "ETH Zurich", "country": "CH", "position": "Postdoc" } ], "n_formula": 34, "n_formula_1": 9, "n_ref_uni": 23, "n_ref": 96, "n_ref_all": 122, "n_fig": 5, "n_tab": 6, "L_tab": 1743, "n_element_tab": 165, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 446, "n_element_tab_1": 43, "formula_len_all": 3484, "formula_len_all_1": 1020, "len_all": 198719, "len_all_1": 72284, "len_abs": 1389, "len_title": 124, "len_sents": 55581, "len_sents_1": 33681, "n_sents": 511, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1190, "title": "Polyrating: A Cost-Effective and Bias-Aware Rating System for LLM Evaluation", "abs": "Rating-based human evaluation has become an essential tool to accurately evaluate the impressive performance of large language models (LLMs). However, current rating systems suffer from several important limitations: first, they fail to account for biases that significantly influence evaluation results, second, they require large and expensive preference datasets to obtain accurate ratings, and third, they do not facilitate meaningful comparisons of model ratings across different tasks. To address these issues, we introduce Polyrating, an expressive and flexible rating system based on maximum a posteriori estimation that enables a more nuanced and thorough analysis of model performance at lower costs. Polyrating can detect and quantify biases affecting human preferences, ensuring fairer model comparisons. Further, Polyrating can reduce the cost of human evaluations by up to $41$% for new models and up to $77$% for new tasks by leveraging existing benchmark scores. Lastly, Polyrating enables direct comparisons of ratings across different tasks, providing a comprehensive understanding of an LLMs' strengths, weaknesses, and relative performance across different applications.", "keywords": [ "large language models", "model evaluation", "rating" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "UQJ7CDW8nb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Qingkai Fang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Shaolei Zhang", "gender": "Male", "institution": "Key Laboratory of Intelligent Information Processing Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Yang Feng", "gender": "unknown", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhe Yang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 3, "n_ref_uni": 41, "n_ref": 146, "n_ref_all": 182, "n_fig": 14, "n_tab": 10, "L_tab": 7955, "n_element_tab": 1127, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 3165, "n_element_tab_1": 259, "formula_len_all": 469, "formula_len_all_1": 198, "len_all": 282779, "len_all_1": 64588, "len_abs": 2038, "len_title": 131, "len_sents": 54764, "len_sents_1": 27249, "n_sents": 385, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1642, "title": "LLaVA-Mini: Efficient Image and Video Large Multimodal Models with One Vision Token", "abs": "The advent of real-time large multimodal models (LMMs) like GPT-4o has sparked considerable interest in efficient LMMs. LMM frameworks typically encode visual inputs into vision tokens (continuous representations) and integrate them and textual instructions into the context of large language models (LLMs), where large-scale parameters and numerous context tokens (predominantly vision tokens) result in substantial computational overhead. Previous efforts towards efficient LMMs always focus on replacing the LLM backbone with smaller models, while neglecting the crucial issue of token quantity. In this paper, we introduce LLaVA-Mini, an efficient LMM with minimal vision tokens. To achieve a high compression ratio of vision tokens while preserving visual information, we first analyze how LMMs understand vision tokens and find that most vision tokens only play a crucial role in the early layers, where they fuse visual information into text tokens. Building on this finding, LLaVA-Mini introduces modality pre-fusion to fuse visual information into text tokens in advance, thereby facilitating the extreme compression of vision tokens fed to LLM backbone into one token. LLaVAMini can support the understanding of images, high-resolution images, and videos in an efficient manner. Experiments across 11 image-based and 7 video-based benchmarks demonstrate that LLaVA-Mini outperforms LLaVA-v1.5 with just 1 vision token instead of 576. Efficiency analyses reveal that LLaVA-Mini can reduce FLOPs by 77%, deliver low-latency responses within 40 milliseconds, and process over 10,000 frames of video on GPU hardware with 24GB of memory.", "keywords": [ "Large Multimodal Models", "Large Language Models" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "UQ0RqfhgCk", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Chanyoung Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Gyoung S. Na", "gender": "Male", "institution": "Korea Research Institute of Chemical Technology", "country": "KR", "position": "Researcher" } ], "n_formula": 20, "n_formula_1": 14, "n_ref_uni": 76, "n_ref": 152, "n_ref_all": 183, "n_fig": 7, "n_tab": 12, "L_tab": 8732, "n_element_tab": 755, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 1862, "n_element_tab_1": 175, "formula_len_all": 1850, "formula_len_all_1": 1068, "len_all": 226687, "len_all_1": 73138, "len_abs": 929, "len_title": 136, "len_sents": 57013, "len_sents_1": 34829, "n_sents": 366, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 932, "title": "Self-Supervised Diffusion Processes for Electron-Aware Molecular Representation Learning", "abs": "Physical properties derived from electronic distributions are essential information determining molecular properties. However, the electron-level information is not accessible in most real-world complex molecules due to extensive computational costs of determining uncertain electronic distributions. For this reason, existing machine learning methods for molecular property prediction have remained in regression models on simplified atom-level molecular descriptors, such as atomic structures. This paper proposes an efficient knowledge transfer method for electron-aware molecular representation learning. To this end, we devised a self-supervised diffusion method that estimates the electron-level information of real-world complex molecules from readily accessible incomplete information in public chemical databases. The proposed method achieved state-of-the-art prediction accuracy on extensive real-world molecular datasets.", "keywords": [ "Representation learning;Generative models;Molecular science;Scientific applications" ], "rating_list": [ 6, 8, 6, 8, 6 ], "soundness_list": [ 4, 4, 3, 3, 3 ], "presentation_list": [ 4, 4, 3, 3, 2 ], "contribution_list": [ 3, 4, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "UPOUVsEafz", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Joel Rixen", "gender": "unknown", "institution": "Christian-Albrechts-Universit\u00e4t Kiel", "country": "DE", "position": "PhD student" }, { "name": "Matthias Renz", "gender": "Male", "institution": "Christian-Albrechts-Universit\u00e4t Kiel", "country": "DE", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 30, "n_ref": 54, "n_ref_all": 69, "n_fig": 5, "n_tab": 1, "L_tab": 12, "n_element_tab": 1, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 12, "n_element_tab_1": 1, "formula_len_all": 477, "formula_len_all_1": 477, "len_all": 100497, "len_all_1": 58412, "len_abs": 948, "len_title": 106, "len_sents": 31753, "len_sents_1": 28908, "n_sents": 231, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 952, "title": "Highly efficient Speech Separation using relative Context", "abs": "Speech separation is a problem area where a mixture with overlapping speech signals is the input and estimations of the clean speech signals which make up the mixture is the output. In this paper we propose a novel sequence modelling method called relative context and use it for a speech separation architecture called RCSep. \nThe main advantages of relative context is that it does not require trainable parameters, is very lightweight and highly parallelized. The RCSep model which heavily uses relative context is an extremely efficient source separation model. It has less than 500k trainable parameters, lower memory usage and is significantly faster than all previous source separation methods while still maintaining high separation accuracy.\nFurthermore, we also used relative context instead of LSTMs in a current SOTA architecture which simultaneously improved separation accuracy and decreased computation time, memory usage and model size.", "keywords": [ "speech separation", "single-channel", "source separation", "sequence modelling", "differencing", "signal processing" ], "rating_list": [ 3, 1, 5, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 1, 1, 3, 3 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 4, 2, 4, 5 ] }, { "paper_id": "UO6JmbwVkC", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Anton van den Hengel", "gender": "unknown", "institution": "University of Adelaide", "country": "AU", "position": "Professor" }, { "name": "Xinyu Li", "gender": "Male", "institution": "University of Adelaide", "country": "", "position": "Postdoc" }, { "name": "Yuhang Liu", "gender": "unknown", "institution": "The University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Zhen Zhang", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Javen Shi", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 41, "n_ref": 64, "n_ref_all": 84, "n_fig": 6, "n_tab": 5, "L_tab": 1647, "n_element_tab": 188, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1062, "n_element_tab_1": 122, "formula_len_all": 153, "formula_len_all_1": 115, "len_all": 135630, "len_all_1": 70252, "len_abs": 1586, "len_title": 130, "len_sents": 36950, "len_sents_1": 33239, "n_sents": 223, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1597, "title": "Rethinking end-to-end prediction of adsorption energies from a causal perspective", "abs": "Adsorption energy is an important descriptor of catalytic activity in the field of catalysis, and significant efforts have been made to develop accurate predictive machine-learning models to replace expensive quantum chemistry calculations. Although it can be inferred by total energy predictions, research has mostly focused on the end-to-end prediction of adsorption energies due to the common belief that total energy is more challenging to predict than adsorption energy. In this study, we first analyzed the causal graph of adsorption energies and revealed that the indirect approach, which infers adsorption energy from total energy predictions, could provide better identifiability, leading to improved accuracy and generalization ability. We also improved the graph property normalization method for total energy prediction and achieved a halved Mean Absolute Error compared to direct adsorption energy prediction for the catalyst in-domain scenario. In the more challenging catalyst out-of-domain scenario, we found that the error primarily comes from predicting the individual energy of unseen catalyst atoms, and the error can be canceled when total energy predictions are used to infer adsorption energy. Consequently, our model achieves a MAE of approximately 0.2 eV for all tasks in the OC20 S2EF task, outperforming end-to-end models trained on datasets 50$\\times$ larger. Given the evidence presented in this study, future research should prioritize the development of total energy models to enhance the accuracy and efficiency of machine-learning approaches in material discovery.", "keywords": [ "Graph Neural Network", "Material Discovery", "Catalyst", "AI for Science" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "UNxCphTxWp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Fan Zhou", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Junlong Li", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "", "position": "MS student" }, { "name": "Pengfei Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Associate Professor" }, { "name": "Qian Liu", "gender": "Male", "institution": "Tiktok", "country": "SG", "position": "Researcher" }, { "name": "Zengzhi Wang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 1, "n_ref_uni": 61, "n_ref": 146, "n_ref_all": 216, "n_fig": 12, "n_tab": 36, "L_tab": 18546, "n_element_tab": 3312, "n_fig_1": 8, "n_tab_1": 7, "L_tab_1": 3480, "n_element_tab_1": 385, "formula_len_all": 980, "formula_len_all_1": 108, "len_all": 340032, "len_all_1": 72920, "len_abs": 1744, "len_title": 130, "len_sents": 74423, "len_sents_1": 32415, "n_sents": 591, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1559, "title": "Programming Every Example: Lifting Pre-training Data Quality like Experts at Scale", "abs": "Large language model pre-training has traditionally relied on human experts to craft heuristics for improving the corpora quality, resulting in numerous rules developed to date. However, these rules lack the flexibility to address the unique characteristics of individual example effectively.\nMeanwhile, applying tailored rules to every example is impractical for human experts.\nIn this paper, we demonstrate that even small language models, with as few as 0.3B parameters, can exhibit substantial\ndata refining capabilities comparable to those of human experts.\nWe introduce Programming Every Example (ProX), a novel framework that treats data refinement as a programming task, enabling models to refine corpora by generating and executing fine-grained operations, such as string normalization, for each individual example at scale.\nExperimental results show that models pre-trained on ProX-curated data outperform either original data or data curated via selection methods by more than 2% across 10 downstream benchmarks.\nIts effectiveness spans various model sizes (0.3B~1.7B) and pre-training corpora (C4, RedPajama-V2, and FineWeb).\nFurthermore, ProX shows great potential in domain-specific continual pre-training: models trained on OpenWebMath refined by ProX outperform human-crafted rule-based methods, improving accuracy by 7.6% on Mistral-7B, 14.6% on Llama-2-7B, and 20.3% on CodeLlama-7B within 10B tokens, comparable to Llemma-7B trained on 200B tokens. ProX significantly reduces training FLOPs, offering an efficient path for LLM pre-training.", "keywords": [ "Large Language Models", "Pre-training", "Data Refinement", "Data Engineering" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 3, 4, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "UNlyhyyuCs", "primary_area": "applications to robotics, autonomy, planning", "n_author": 9, "author_info_list": [ { "name": "Hongxia Yang", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Jianbo Yuan", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Shenao Zhang", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Shuqi Ke", "gender": "Not Specified", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Sirui Zheng", "gender": "Male", "institution": "Northwestern University", "country": "", "position": "PhD student" }, { "name": "Wanxin Jin", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "Assistant Professor" }, { "name": "Yingxiang Yang", "gender": "Male", "institution": "ByteDance Inc", "country": "", "position": "ByteDance" }, { "name": "Zhaoran Wang", "gender": "Not Specified", "institution": "Northwestern University", "country": "US", "position": "Associate Professor" }, { "name": "Zhihan Liu", "gender": "Male", "institution": "Northwestern University", "country": "", "position": "PhD student" } ], "n_formula": 51, "n_formula_1": 9, "n_ref_uni": 46, "n_ref": 85, "n_ref_all": 97, "n_fig": 5, "n_tab": 5, "L_tab": 389, "n_element_tab": 55, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 326, "n_element_tab_1": 52, "formula_len_all": 7176, "formula_len_all_1": 1068, "len_all": 179422, "len_all_1": 65532, "len_abs": 1607, "len_title": 93, "len_sents": 40449, "len_sents_1": 28586, "n_sents": 411, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 1569, "title": "How Can LLM Guide RL? A Value-Based Approach", "abs": "Reinforcement learning (RL) has become the de facto standard practice for sequential decision-making problems by improving future acting policies with feedback. However, RL algorithms may require extensive trial-and-error interactions to collect useful feedback for improvement. On the other hand, recent developments in large language models (LLMs) have showcased impressive capabilities in language understanding and generation, yet they fall short in exploration and self-improvement capabilities for planning tasks, lacking the ability to autonomously refine their responses based on feedback. Therefore, in this paper, we study how the policy prior provided by the LLM can enhance the sample efficiency of RL algorithms. Specifically, we develop an algorithm named $\\mathtt{LINVIT}$ that incorporates LLM guidance as a regularization factor in value-based RL, leading to significant reductions in the amount of data needed for learning, particularly when the difference between the ideal policy and the LLM-informed policy is small, which suggests that the initial policy is close to optimal, reducing the need for further exploration. Additionally, we present a practical algorithm $\\mathtt{SLINVIT}$ that simplifies the construction of the value function and employs sub-goals to reduce the search complexity. Our experiments across three interactive environments---ALFWorld, InterCode, and BlocksWorld---demonstrate that the proposed method achieves state-of-the-art success rates and also surpasses previous RL and LLM approaches in terms of sample efficiency.", "keywords": [ "Large Language Model", "Reinforcement Learning", "Agent" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 4, 3, 1, 2 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "UNHU7uO2qM", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "George Konidaris", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Assistant Professor" }, { "name": "Haotian Fu", "gender": "Male", "institution": "Brown University", "country": "", "position": "PhD student" }, { "name": "Michael Littman", "gender": "Male", "institution": "Brown University", "country": "", "position": "Full Professor" }, { "name": "Yixiang Sun", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Undergrad student" } ], "n_formula": 27, "n_formula_1": 11, "n_ref_uni": 49, "n_ref": 77, "n_ref_all": 97, "n_fig": 5, "n_tab": 7, "L_tab": 2062, "n_element_tab": 136, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2332, "formula_len_all_1": 1031, "len_all": 173396, "len_all_1": 58419, "len_abs": 996, "len_title": 115, "len_sents": 50035, "len_sents_1": 29480, "n_sents": 323, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1019, "title": "Knowledge Retention in Continual Model-Based Reinforcement Learning", "abs": "We propose DRAGO, a novel approach for continual model-based reinforcement learning aimed at improving the incremental development of world models across a sequence of tasks that differ in their reward functions but not the state space or dynamics. DRAGO comprises two key components: $\\textit{Synthetic Experience Rehearsal}$, which leverages generative models to create synthetic experiences from past tasks, allowing the agent to reinforce previously learned dynamics without storing data, and $\\textit{Regaining Memories Through Exploration}$, which introduces an intrinsic reward mechanism to guide the agent toward revisiting relevant states from prior tasks. Together, these components enable the agent to maintain a comprehensive and continually developing world model, facilitating more effective learning and adaptation across diverse environments. Empirical evaluations demonstrate that DRAGO is able to preserve knowledge across tasks, achieving superior performance in various continual learning scenarios.", "keywords": [ "Continual learning", "Model-based reinforcement learning", "World model", "Catastrophic forgetting" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "UN94vDiaJv", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Futoshi Futami", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "Lecturer" }, { "name": "Masahiro Fujisawa", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Special Postdoctoral Researcher" } ], "n_formula": 100, "n_formula_1": 19, "n_ref_uni": 60, "n_ref": 127, "n_ref_all": 204, "n_fig": 2, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10891, "formula_len_all_1": 1197, "len_all": 226497, "len_all_1": 77143, "len_abs": 1209, "len_title": 119, "len_sents": 68212, "len_sents_1": 36896, "n_sents": 617, "n_sents_1": 275, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1213, "title": "Information-theoretic Generalization Analysis for Vector-Quantized VAEs", "abs": "Encoder--decoder models, which transform input data into latent variables, have achieved a significant success in machine learning. Although the generalization capability of these models has been theoretically analyzed in supervised learning focusing on the complexity of latent variables, the contribution of latent variables in generalization and data generation capabilities are less explored theoretically in unsupervised learning. To address this gap, our study leverages information-theoretic generalization error analysis (IT analysis). Using the supersample setting in recent IT analysis, we demonstrate that the generalization gap for reconstruction loss can be evaluated through mutual information related to the posterior distribution of latent variables, conditioned on the input data, without relying on the decoder's information. We also introduce a novel permutation-symmetric supersample setting, which extends the existing IT analysis and shows that regularizing the encoder's capacity leads to generalization. Finally, we guarantee the 2-Wasserstein distance between the true data distribution and the generated data distribution, offering insights into the model\u2019s data generation capabilities.", "keywords": [ "Information-theoretic generalization error analysis", "generalization error analysis", "VQ-VAE" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "UN6Ik6OCx8", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Bingning Wang", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Han Huang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Haoyu Lu", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Kun Zhou", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Postdoc" }, { "name": "Xin Zhao", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Yifan Du", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Yuqi Huo", "gender": "Male", "institution": "Baichuan Intelligent Technology", "country": "CN", "position": "Researcher" }, { "name": "weipeng chen", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Joe Z", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 79, "n_ref_all": 115, "n_fig": 8, "n_tab": 5, "L_tab": 2115, "n_element_tab": 376, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 902, "n_element_tab_1": 140, "formula_len_all": 351, "formula_len_all_1": 265, "len_all": 153731, "len_all_1": 57969, "len_abs": 1424, "len_title": 122, "len_sents": 39602, "len_sents_1": 25835, "n_sents": 285, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1371, "title": "Exploring the Design Space of Visual Context Representation in Video MLLMs", "abs": "Video Multimodal Large Language Models (MLLMs) have shown remarkable capability of understanding the video semantics on various downstream tasks. Despite the advancements, there is still a lack of systematic research on visual context representation, which refers to the scheme to select frames from a video and further select the tokens from a frame. In this paper, we explore the design space for visual context representation, and aim to improve the performance of video MLLMs by finding more effective representation schemes. Firstly, we formulate the task of visual context representation as a constrained optimization problem, and model the language modeling loss as a function of the number of frames and the number of embeddings (or tokens) per frame, given the maximum visual context window size. \nThen, we explore the scaling effects in frame selection and token selection respectively, and fit the corresponding function curve by conducting extensive empirical experiments. We examine the effectiveness of typical selection strategies and present empirical findings to determine the two factors. Furthermore, we study the joint effect of frame selection and token selection, and derive the\noptimal formula for determining the two factors. We demonstrate that the derived optimal settings show alignment with the best-performed results of empirical experiments.", "keywords": [ "Video Multimodal Large Language Model", "Scaling Law" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "UM6yage1H0", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Denis Wambold", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "MS student" }, { "name": "Federico Matteucci", "gender": "unknown", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "", "position": "PhD student" }, { "name": "Klemens B\u00f6hm", "gender": "unknown", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "", "position": "Full Professor" }, { "name": "Vadim Arzamasov", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "", "position": "Postdoc" }, { "name": "Jose A Cribeiro-Ramallo", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "PhD student" } ], "n_formula": 32, "n_formula_1": 6, "n_ref_uni": 34, "n_ref": 94, "n_ref_all": 123, "n_fig": 12, "n_tab": 8, "L_tab": 1873, "n_element_tab": 181, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1992, "n_element_tab_1": 126, "formula_len_all": 2725, "formula_len_all_1": 756, "len_all": 190811, "len_all_1": 60974, "len_abs": 1096, "len_title": 132, "len_sents": 58599, "len_sents_1": 26027, "n_sents": 558, "n_sents_1": 244, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1121, "title": "Can Data be Myopic? Outlier Detection in High-Dimensional Tabular Data via Subspaces", "abs": "Outlier detection in high-dimensional tabular data is an important task in data mining, essential for many downstream tasks and applications. Existing unsupervised outlier detection algorithms face one or more problems, including inlier assumption (IA), curse of dimensionality (CD), and multiple views (MV). To address these issues, we introduce Generative Subspace Adversarial Active Learning (GSAAL), a novel approach that uses a Generative Adversarial Network with multiple adversaries. These adversaries learn the marginal class probability functions over different data subspaces, while a single generator in the full space models the entire distribution of the inlier class. GSAAL is specifically designed to address the MV limitation while also handling the IA and CD, being the only method to do so. We provide a mathematical formulation of MV, convergence guarantees for the discriminators, and scalability results for GSAAL. Our extensive experiments demonstrate the effectiveness and scalability of GSAAL, highlighting its superior performance compared to other popular OD methods, especially in MV scenarios.", "keywords": [ "One-class classification", "Tabular Data", "Generative methods", "Deep Learning", "Generative Adversarial Active Learning", "Subspace Outlier Detection" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "ULorFBST6X", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Samson Zhou", "gender": "unknown", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" }, { "name": "Shuo Xing", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Victoria G. Crawford", "gender": "Female", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" }, { "name": "Wenjing Chen", "gender": "Female", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" } ], "n_formula": 92, "n_formula_1": 21, "n_ref_uni": 43, "n_ref": 91, "n_ref_all": 106, "n_fig": 12, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9200, "formula_len_all_1": 1664, "len_all": 203990, "len_all_1": 73237, "len_abs": 1196, "len_title": 69, "len_sents": 72494, "len_sents_1": 36115, "n_sents": 557, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 21, "L_abs": 1278, "title": "Fair Submodular Cover", "abs": "Machine learning algorithms are becoming increasing prevalent in the modern world, and as a result there has been significant recent study into algorithmic fairness in order to minimize the possibility of unintentional bias or discrimination in these algorithms. Submodular optimization problems also arise in many machine learning applications, including those such as data summarization and clustering where fairness is an important concern. In this paper, we initiate the study of the Fair Submodular Cover Problem (FSC). Given a ground set $U$, a monotone submodular function $f:2^U\\to\\mathbb{R}_{\\ge 0}$, and a threshold $\\tau$, the goal of FSC is to find a balanced subset of $U$ with minimum cardinality such that $f(S)\\ge\\tau$. We first introduce discrete algorithms for FSC that achieve a bicriteria approximation ratio of $(\\frac{1}{\\varepsilon}, 1-O(\\varepsilon))$. We then present a continuous algorithm that achieves a $(\\ln\\frac{1}{\\varepsilon}, 1-O(\\varepsilon))$-bicriteria approximation ratio, which matches the best approximation guarantee of submodular cover without a fairness constraint. Finally, we complement our theoretical results with a number of empirical evaluations that demonstrate the efficiency of our algorithms on instances of maximum coverage.", "keywords": [ "discrete optimization", "submodular cover", "fairness" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 4, 4 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "ULGbw2URE3", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Alex Smola", "gender": "Male", "institution": "Amazon", "country": "", "position": "Distinguished Scientist" }, { "name": "Xingjian Shi", "gender": "Male", "institution": "Amazon Web Services", "country": "US", "position": "Applied Scientist" }, { "name": "Guneet Singh Dhillon", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "DPhil Student" }, { "name": "Yee Whye Teh", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 41, "n_ref": 74, "n_ref_all": 102, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 121, "n_element_tab_1": 12, "formula_len_all": 927, "formula_len_all_1": 927, "len_all": 145451, "len_all_1": 71305, "len_abs": 1131, "len_title": 84, "len_sents": 33359, "len_sents_1": 29302, "n_sents": 278, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 37, "L_abs": 878, "title": "L3Ms \u2014 Lagrange Large Language Models", "abs": "Supervised fine-tuning (SFT) and alignment of large language models (LLMs) are key steps in providing a good user experience. However, the concept of an appropriate alignment is inherently application-dependent, and current methods often rely on heuristic choices to drive the optimization. In this work, we formulate SFT and alignment as a constrained optimization problem: the LLM is fine-tuned on a task while being required to meet application-specific requirements, without resorting to heuristics. To solve this, we propose Lagrange Large Language Models (L3Ms), which employ logarithmic barriers to enforce the constraints. This approach allows for the customization of L3Ms across diverse applications while avoiding heuristic-driven processes. We demonstrate experimentally the versatility and efficacy of L3Ms in achieving tailored alignments for various applications.", "keywords": [ "LLM", "alignment" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "UL95EpgrlS", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Alvaro Soto", "gender": "Male", "institution": "Universidad Cat\u00f3lica de Chile", "country": "CL", "position": "Associate Professor" }, { "name": "Bernard Ghanem", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Full Professor" }, { "name": "Juan C Leon Alcazar", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" }, { "name": "Andr\u00e9s Felipe Villa Ojeda", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 112, "n_ref_all": 148, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 161127, "len_all_1": 69137, "len_abs": 1372, "len_title": 139, "len_sents": 47765, "len_sents_1": 35170, "n_sents": 339, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1357, "title": "Behind the Magic, MERLIM: Multi-modal Evaluation Benchmark for Large Image-Language Models", "abs": "Large Vision and Language Models have enabled significant advances in fully supervised and zero-shot vision tasks. These large pre-trained architectures serve as the baseline to what is currently known as Instruction Tuning Large Vision and Language models (IT-LVLMs). IT-LVLMs are general-purpose multi-modal assistants whose responses are modulated by natural language instructions and arbitrary visual data. Despite this versatility, IT-LVLM effectiveness in fundamental computer vision problems remains unclear, primarily due to the absence of a standardized evaluation benchmark. This paper introduces a Multi-modal Evaluation Benchmark named MERLIM, a scalable test-bed to assess the performance of IT-LVLMs on fundamental computer vision tasks. MERLIM contains over 300K image-question pairs and has a strong focus on detecting cross-modal \u201challucination\u201d events in IT-LVLMs. Our results show that state-of-the-art IT-LVMLs are still limited at identifying fine-grained visual concepts, object hallucinations are common across tasks, and their results are strongly biased by small variations in the input query, even if the queries have the very same semantics. Our findings also suggest that these models lack direct visual groundings, but can still make adequate guesses from global visual patterns or textual biases contained in the LLM component.", "keywords": [ "Large Image-Language Models", "Evaluation Benchmark", "Instruction Tuning Methods", "Object Hallucinations", "Visual Grounding" ], "rating_list": [ 3, 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 5, 3, 3, 4, 4 ] }, { "paper_id": "UL8b54P96G", "primary_area": "generative models", "n_author": 12, "author_info_list": [ { "name": "Beide Liu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "MS student" }, { "name": "Jianfeng Wang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Kai-Wei Chang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Associate Professor" }, { "name": "Kevin Lin", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Lijuan Wang", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Linjie Li", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Maxine Wu", "gender": "unknown", "institution": "UCLA Computer Science Department, University of California, Los Angeles", "country": "US", "position": "Undergrad student" }, { "name": "Yining Hong", "gender": "Female", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Yuanhao Zhai", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Zhengyuan Yang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Chung-ching Lin", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Principal\u00a0Researcher" }, { "name": "Yingnian Wu", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 8, "n_ref_uni": 47, "n_ref": 91, "n_ref_all": 111, "n_fig": 6, "n_tab": 11, "L_tab": 3010, "n_element_tab": 350, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1720, "n_element_tab_1": 82, "formula_len_all": 687, "formula_len_all_1": 455, "len_all": 230237, "len_all_1": 72397, "len_abs": 347, "len_title": 121, "len_sents": 57689, "len_sents_1": 35002, "n_sents": 405, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 73, "L_abs": 1814, "title": "SlowFast-VGen: Slow-Fast Learning for Action-Driven Long Video Generation", "abs": "Human beings are endowed with a complementary learning system, which bridges the slow learning of general world dynamics with fast storage of episodic memory from a new experience. Previous video generation models, however, primarily focus on slow learning by pre-training on vast amounts of data, overlooking the fast learning phase crucial for episodic memory storage. This oversight leads to inconsistencies across temporally distant frames when generating longer videos, as these frames fall beyond the model's context window. To this end, we introduce SlowFast-VGen, a novel dual-speed learning system for action-driven long video generation. Our approach incorporates a masked conditional video diffusion model for the slow learning of world dynamics, alongside an inference-time fast learning strategy based on a temporal LoRA module. Specifically, the fast learning process updates its temporal LoRA parameters based on local inputs and outputs, thereby efficiently storing episodic memory in its parameters. We further propose a slow-fast learning loop algorithm that seamlessly integrates the inner fast learning loop into the outer slow learning loop, enabling the recall of prior multi-episode experiences for context-aware skill learning. To facilitate the slow learning of an approximate world model, we collect a large-scale dataset of 200k videos with language action annotations, covering a wide range of scenarios. Extensive experiments show that SlowFast-VGen outperforms baselines across various metrics for action-driven video generation, achieving an FVD score of 514 compared to 782, and maintaining consistency in longer videos, with an average of 0.37 scene cuts versus 0.89. The slow-fast learning loop algorithm significantly enhances performances on long-horizon planning tasks as well.", "keywords": [ "video generation", "complimentary learning system", "slow-fast learning", "diffusion" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "UKkjMiGNYK", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Peter Sandborn", "gender": "unknown", "institution": "University of Maryland College Park", "country": "", "position": "" }, { "name": "Mohammad Rubyet Islam", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 13, "n_ref": 18, "n_ref_all": 35, "n_fig": 6, "n_tab": 5, "L_tab": 1055, "n_element_tab": 149, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1225, "n_element_tab_1": 169, "formula_len_all": 24, "formula_len_all_1": 24, "len_all": 70439, "len_all_1": 46750, "len_abs": 970, "len_title": 100, "len_sents": 25811, "len_sents_1": 21483, "n_sents": 170, "n_sents_1": 139, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 973, "title": "MULTIMODAL GENERATIVE AI FOR STORY POINT ESTIMATION", "abs": "This research explores the application of Multimodal Generative AI to enhance story point estimation in Agile software development. By integrating text, image, and categorical data using advanced models like BERT, CNN, and XGBoost, our approach surpasses the limitations of traditional single-modal estimation methods. The results demonstrate good accuracy for simpler story points, while also highlighting challenges in more complex categories due to data imbalance. This study further explores the impact of categorical data, particularly severity, on the estimation process, emphasizing its influence on model performance. Our findings emphasize the transformative potential of multimodal data integration in refining AI-driven project management, paving the way for more precise, adaptable, and domain-specific AI capabilities. Additionally, this work outlines future directions for addressing data variability and enhancing the robustness of AI in Agile methodologies.", "keywords": [ "Multimodal Generative AI", "Story Point Estimation", "Software Development", "Vector Embeddings", "BERT", "LLMs" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 1, 1, 3, 1 ], "contribution_list": [ 2, 1, 3, 1 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "UKjAwMzX4m", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Anand Raghunathan", "gender": "unknown", "institution": "Purdue University", "country": "", "position": "Full Professor" }, { "name": "Brucek Khailany", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Charbel Sakr", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Reena Elangovan", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 19, "n_ref_uni": 30, "n_ref": 69, "n_ref_all": 93, "n_fig": 9, "n_tab": 10, "L_tab": 5111, "n_element_tab": 673, "n_fig_1": 7, "n_tab_1": 7, "L_tab_1": 3980, "n_element_tab_1": 542, "formula_len_all": 1568, "formula_len_all_1": 1034, "len_all": 148186, "len_all_1": 74057, "len_abs": 1265, "len_title": 112, "len_sents": 38878, "len_sents_1": 29264, "n_sents": 295, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1240, "title": "BCQ: Block Clustered Quantization for 4-bit (W4A4) LLM inference", "abs": "Post-training quantization (PTQ) is a promising approach to reducing the storage and computational requirements of large language models (LLMs) without additional training cost. Recent PTQ studies have primarily focused on quantizing only weights to sub-8-bits while maintaining activations at 8-bits or higher. Accurate sub-8-bit quantization for both weights and activations without relying on quantization-aware training remains a significant challenge. In this work, we introduce a novel quantization method called block clustered quantization (BCQ) wherein each operand tensor is decomposed into blocks (a block is a group of contiguous scalars), blocks are clustered based on their statistics, and a dedicated optimal quantization codebook is designed for each cluster. We propose a PTQ algorithm called Locally-Optimal BCQ (LO-BCQ) that iterates between the steps of block clustering and codebook design to greedily minimize the quantization mean squared error. When weight and activation scalars are encoded to W4A4 format (with 0.5-bits of overhead for storing scaling factors and codebook selectors), we advance the current state-of-the-art by demonstrating <1% loss in inference accuracy across several LLMs and downstream tasks.", "keywords": [ "Post-training Quantization", "Large Language Models", "Codebooks", "Clustering", "Block Clustered Quantization" ], "rating_list": [ 6, 5, 3, 8 ], "soundness_list": [ 2, 2, 2, 4 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "UKiCFpwcqY", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "LEI BAI", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Wenqi Huang", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Zhen Huang", "gender": "Male", "institution": "China Soutern Power Grid", "country": "CN", "position": "Researcher" }, { "name": "Zhen Zhao", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Zicheng Wang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 27, "n_ref": 48, "n_ref_all": 64, "n_fig": 4, "n_tab": 5, "L_tab": 3409, "n_element_tab": 168, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2938, "n_element_tab_1": 151, "formula_len_all": 768, "formula_len_all_1": 636, "len_all": 94061, "len_all_1": 61591, "len_abs": 1631, "len_title": 143, "len_sents": 30631, "len_sents_1": 28159, "n_sents": 187, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1632, "title": "SenseFlow: A Physics-Informed and Self-Ensembling Iterative Framework for Power Flow Estimation", "abs": "Power flow estimation plays a vital role in ensuring the stability and reliability of electrical power systems, particularly in the context of growing network complexities and renewable energy integration. However, existing studies often fail to adequately address the unique characteristics of power systems, such as the sparsity of network connections and the critical importance of the unique Slack node, which poses significant challenges in achieving high-accuracy estimations. In this paper, we present SenseFlow, a novel Physics-Informed and Self-Ensembling Iterative Framework that integrates two main designs, the Physics-Informed Power Flow Network (FlowNet) and Self-Ensembling Iterative Estimation (SeIter), to carefully address the unique properties of the power system and thereby enhance the power flow estimation. Specifically, SenseFlow enforces the FlowNet to gradually predict high-precision voltage magnitudes and phase angles through the iterative SeIter process. On the one hand, FlowNet employs the Virtual Node Attention and Slack-Gated Feed-Forward modules to facilitate efficient global-local communication in the face of network sparsity and amplify the influence of the Slack node on angle predictions, respectively. On the other hand, SeIter maintains an exponential moving average of FlowNet\u2019s parameters to create a robust ensemble model that refines power state predictions throughout the iterative fitting process. Experimental results demonstrate that SenseFlow outperforms existing methods, providing a promising solution for high-accuracy power flow estimation across diverse grid configurations.", "keywords": [ "Power Flow Estimation; Physics-Informed Iterative Framework; Self-Ensembling Estimation" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 4, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "UKZqSYB2ya", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 1, "author_info_list": [ { "name": "Hooman Ramezani", "gender": "unknown", "institution": "University of Toronto", "country": "CA", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 17, "n_ref": 30, "n_ref_all": 41, "n_fig": 3, "n_tab": 2, "L_tab": 1609, "n_element_tab": 112, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1938, "n_element_tab_1": 116, "formula_len_all": 128, "formula_len_all_1": 128, "len_all": 78614, "len_all_1": 65731, "len_abs": 1216, "len_title": 132, "len_sents": 21289, "len_sents_1": 20960, "n_sents": 129, "n_sents_1": 129, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1222, "title": "Transformer-Based CT Anomaly Detection and Auto-Segmentation of Sparse Lung Nodules", "abs": "Accurate segmentation of lung nodules in computed tomography (CT) scans is challenging due to extreme class imbalance, where nodules appear sparsely among healthy tissue. Lung tumor boards often review these scans manually, a time-consuming process. This paper introduces a novel two-stage approach for lung tumor segmentation by framing the problem as anomaly detection. The method is divided into two stages, allowing each model to leverage its strengths. Stage 1 focuses on region proposal, employing a custom Deformable Detection Transformer with Focal Loss to overcome class imbalance and localize sparse tumors. In Stage 2, the predicted bounding boxes are refined into pixel-wise segmentation masks using a fine-tuned variant of Meta's Segment Anything Model (SAM) for semantic segmentation. To address the challenge of nodule sparsity and improve spatial context, a 7.5 mm Maximum Intensity Projection (MIP) is applied, aiding in the differentiation between nodules, bronchioles, and vascular structures. The model achieves a Dice coefficient of 92.4%, with 95.2% sensitivity and 93.2% precision on the LUNA16 dataset, demonstrating robust performance in real-world clinical conditions where nodule sparsity is 5%.", "keywords": [ "Transformer", "CT scans", "lung nodules", "anomaly detection", "auto-segmentation", "Deformable-DETR", "sparse data", "medical imaging", "self-attention", "multi-scale learning", "object detection", "Focal Loss", "segmentation" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "UK0jrVGCg2", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Aadithya Srikanth", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "MS student" }, { "name": "Chandra Sekhar Seelamantula", "gender": "Male", "institution": "Indian Institute of Science, Bangalore", "country": "", "position": "Professor" }, { "name": "Nishanth Shetty", "gender": "Male", "institution": "Indian Institute of Science", "country": "IN", "position": "PhD student" }, { "name": "Siddarth Asokan", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Researcher" } ], "n_formula": 55, "n_formula_1": 17, "n_ref_uni": 62, "n_ref": 159, "n_ref_all": 190, "n_fig": 8, "n_tab": 2, "L_tab": 380, "n_element_tab": 41, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 718, "n_element_tab_1": 40, "formula_len_all": 4661, "formula_len_all_1": 1201, "len_all": 205876, "len_all_1": 67685, "len_abs": 1593, "len_title": 110, "len_sents": 60670, "len_sents_1": 31291, "n_sents": 430, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1606, "title": "Accelerated Diffusion using Closed-form Discriminator Guidance", "abs": "Diffusion models are a state-of-the-art generative modeling framework that transform noise to images via Langevin sampling, guided by the score, which is the gradient of the logarithm of the data distribution. Recent works have shown empirically that the generation quality can be improved when guided by classifier network, which is typically the discriminator trained in a generative adversarial network (GAN) setting. In this paper, we propose a theoretical framework to analyze the effect of the GAN discriminator on Langevin-based sampling, and show that in IPM GANs, the optimal generator matches {\\it score-like} functions, involving the flow-field of the kernel associated with a chosen IPM constraint space. Further, we show that IPM-GAN optimization can be seen as one of smoothed score-matching, where the scores of the data and the generator distributions are convolved with the kernel associated with the constraint. The proposed approach serves to unify score-based training and optimization of IPM-GANs. Based on these insights, we demonstrate that closed-form discriminator guidance, using a kernel-based implementation, results in improvements (in terms of CLIP-FID and KID metrics) when applied atop baseline diffusion models. We demonstrate these results by applying closed-form discriminator guidance to denoising diffusion implicit model (DDIM) and latent diffusion model (LDM) settings on the FFHQ and CelebA-HQ datasets. We also demonstrate improvements to accelerated time-step-shifted diffusion, when coupled with a wavelet-based noise estimator for latent-space image generation.", "keywords": [ "diffusion models", "GANs", "time-step-shifted sampling", "discriminator guidance" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 1, 3 ] }, { "paper_id": "UIZyvnA0yi", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Anders Malthe-Sorenssen", "gender": "Male", "institution": "Simula Research Laboratories", "country": "", "position": "Principal Researcher" }, { "name": "Markus Pettersen", "gender": "Male", "institution": "University of Oslo", "country": "", "position": "PhD student" }, { "name": "Mattis Dals\u00e6tra \u00d8stby", "gender": "Male", "institution": "University of Oslo", "country": "NO", "position": "Researcher" }, { "name": "Mikkel Elle Lepper\u00f8d", "gender": "Male", "institution": "Simula Research Laboratory", "country": "NO", "position": "Principal Researcher" }, { "name": "Vemund Sigmundson Sch\u00f8yen", "gender": "Not Specified", "institution": "University of Oslo", "country": "NO", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 29, "n_ref": 86, "n_ref_all": 133, "n_fig": 30, "n_tab": 1, "L_tab": 1340, "n_element_tab": 228, "n_fig_1": 20, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 424, "formula_len_all_1": 279, "len_all": 146611, "len_all_1": 59066, "len_abs": 1502, "len_title": 100, "len_sents": 52263, "len_sents_1": 30028, "n_sents": 369, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1517, "title": "Self-Supervised Grid Cells Without Path Integration", "abs": "Grid cells, found in the medial Entorhinal Cortex, are known for their regular spatial firing patterns. These cells have been proposed as the neural solution to a range of computational tasks, from performing path integration to serving as a metric for space. Their exact function, however, remains fiercely debated. In this work, we explore the consequences of demanding local distance preservation in networks subject to a capacity constraint. We consider two distinct self-supervised models, a feedforward network that learns to solve a purely spatial, local distance-based encoding task, and a recurrent network that solves the same problem during path integration. We find that this task leads to the emergence of highly grid cell-like representations in both networks. However, the recurrent network also features units with band-like representations. We subsequently prune velocity inputs to subsets of recurrent units, and find that their grid score is negatively correlated with path integration contribution. Thus, grid cells emerge without path integration in the feedforward network, and they appear significantly less important than band cells for path integration in the recurrent network. Our work provides a minimal model for learning grid-like spatial representations, and questions the role of grid cells as neural path integrators. Instead, it seems that local distance preservation and high population capacity is a more likely candidate task for learning grid cells in artificial neural networks.", "keywords": [ "Grid cells", "Path Integration", "AI", "NeuroAI" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "UIFAJZ22ZF", "primary_area": "applications to neuroscience & cognitive science", "n_author": 8, "author_info_list": [ { "name": "Akash Nagaraj", "gender": "Male", "institution": "Brown University", "country": "US", "position": "MS student" }, { "name": "Alekh Karkada Ashok", "gender": "Male", "institution": "Brown University", "country": "", "position": "PhD student" }, { "name": "Drew Linsley", "gender": "unknown", "institution": "Brown University", "country": "", "position": "Assistant Professor" }, { "name": "Francis E Lewis", "gender": "unknown", "institution": "Brown University", "country": "", "position": "Researcher" }, { "name": "Gaurav Gaonkar", "gender": "Male", "institution": "Brown University", "country": "US", "position": "MS student" }, { "name": "Peisen Zhou", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Researcher" }, { "name": "Thomas Serre", "gender": "Male", "institution": "Brown University", "country": "", "position": "Full Professor" }, { "name": "Zygmunt Pizlo", "gender": "Male", "institution": "University of California, Irvine", "country": "", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 80, "n_ref": 163, "n_ref_all": 197, "n_fig": 25, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 21, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 196539, "len_all_1": 52740, "len_abs": 1809, "len_title": 123, "len_sents": 41480, "len_sents_1": 26582, "n_sents": 277, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1838, "title": "The 3D-PC: a benchmark for visual perspective taking in humans and machines", "abs": "Visual perspective taking (VPT) is the ability to perceive and reason about the perspectives of others. It is an essential feature of human intelligence, which develops over the first decade of life and requires an ability to process the 3D structure of visual scenes. A growing number of reports have indicated that deep neural networks (DNNs) become capable of analyzing 3D scenes after training on large image datasets. We investigated if this emergent ability for 3D analysis in DNNs is sufficient for VPT with the 3D perception challenge (3D-PC): a novel benchmark for 3D perception in humans and DNNs. The 3D-PC is comprised of three 3D-analysis tasks posed within natural scene images: (i.) a simple test of object depth order, (ii.) a basic VPT task (VPT-basic), and (iii.) a more challenging version of VPT (VPT-perturb) designed to limit the effectiveness of \"shortcut\" visual strategies. We tested human participants (N=33) and linearly probed or text-prompted over 300 DNNs on the challenge and found that nearly all of the DNNs approached or exceeded human accuracy in analyzing object depth order. Surprisingly, DNN accuracy on this task correlated with their object recognition performance. In contrast, there was an extraordinary gap between DNNs and humans on VPT-basic. Humans were nearly perfect, whereas most DNNs were near chance. Fine-tuning DNNs on VPT-basic brought them close to human performance, but they, unlike humans, dropped back to chance when tested on VPT-perturb. Our challenge demonstrates that the training routines and architectures of today's DNNs are well-suited for learning basic 3D properties of scenes and objects but are ill-suited for reasoning about these properties like humans do. We release our 3D-PC datasets and code to help bridge this gap in 3D perception between humans and machines.", "keywords": [ "3D vision", "visual cognition", "developmental psychology", "visual reasonsing" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 1, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 4 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "UHg1xTRzZK", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Bing Liu", "gender": "Male", "institution": "University of Illinois at Chicago", "country": "US", "position": "Full Professor" }, { "name": "Chengqing Zong", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Researcher" }, { "name": "Junhong Wu", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yangyifan Xu", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zhao Yang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 25, "n_ref": 79, "n_ref_all": 101, "n_fig": 9, "n_tab": 7, "L_tab": 4981, "n_element_tab": 546, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 3631, "n_element_tab_1": 419, "formula_len_all": 145, "formula_len_all_1": 124, "len_all": 202344, "len_all_1": 68519, "len_abs": 1568, "len_title": 136, "len_sents": 46331, "len_sents_1": 28337, "n_sents": 380, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1583, "title": "Boosting LLM Translation Skills without General Ability Loss via Rationale Distillation", "abs": "Large Language Models (LLMs) have achieved impressive results across numerous NLP tasks but still encounter difficulties in machine translation. Traditional methods to improve translation have typically involved fine-tuning LLMs using parallel corpora. However, vanilla fine-tuning often leads to catastrophic forgetting of the instruction-following capabilities and alignment with human preferences, compromising their broad general abilities and introducing potential security risks. These abilities, which are developed using proprietary and unavailable training data, make existing continual instruction tuning methods ineffective. To overcome this issue, we propose a novel approach called $\\textbf{RaDis}$ ($\\textbf{Ra}$tionale $\\textbf{Dis}$tillation). RaDis harnesses the strong generative capabilities of LLMs to create rationales for training data, which are then \u201creplayed\u201d to prevent forgetting. These rationales $\\textit{connect prior knowledge with new tasks}$, acting as $\\textit{self-distillation targets}$ to regulate the training process. By jointly training on reference translations and self-generated rationales, the model can learn new translation skills while preserving its general abilities. Extensive experiments demonstrate that our method enhances machine translation performance while maintaining the broader capabilities of LLMs across other tasks. This work presents a pathway for creating more versatile LLMs that excel in specialized tasks without compromising generality or safety and provides a fresh angle for utilizing rationales in the CL field.", "keywords": [ "Large Language Model", "Continual Instruction Tuning", "Machine Translation" ], "rating_list": [ 3, 6, 6, 5 ], "soundness_list": [ 2, 2, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "UHYRNAfnNA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Evgenii Opryshko", "gender": "unknown", "institution": "University of Toronto", "country": "", "position": "PhD student" }, { "name": "Igor Gilitschenski", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Assistant Professor" }, { "name": "Umangi Jain", "gender": "Female", "institution": "University of Toronto", "country": "", "position": "PhD student" } ], "n_formula": 41, "n_formula_1": 17, "n_ref_uni": 3, "n_ref": 148, "n_ref_all": 168, "n_fig": 9, "n_tab": 9, "L_tab": 4295, "n_element_tab": 351, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 20, "n_element_tab_1": 3, "formula_len_all": 2840, "formula_len_all_1": 1077, "len_all": 362348, "len_all_1": 67230, "len_abs": 1358, "len_title": 123, "len_sents": 104487, "len_sents_1": 32876, "n_sents": 719, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1362, "title": "Modification-Considering Value Learning for Reward Hacking Mitigation in RL", "abs": "Reinforcement learning (RL) agents can exploit unintended strategies to achieve high rewards without fulfilling the desired objectives, a phenomenon known as reward hacking. In this work, we examine reward hacking through the lens of General Utility RL, which generalizes RL by considering utility functions over entire trajectories rather than state-based rewards. From this perspective, many instances of reward hacking can be seen as inconsistencies between current and updated utility functions, where the behavior optimized for an updated utility function is poorly evaluated by the original one. Our main contribution is Modification-Considering Value Learning (MC-VL), a novel algorithm designed to address this inconsistency during learning. Starting with a coarse yet value-aligned initial utility function, the MC-VL agent iteratively refines this function based on past observations while considering the potential consequences of updates. This approach enables the agent to anticipate and reject modifications that may lead to undesired behavior. To validate our approach, we implement MC-VL agents based on the Double Deep Q-Network (DDQN) and Twin Delayed Deep Deterministic Policy Gradients (TD3), demonstrating their effectiveness in preventing reward hacking in diverse environments, including those from AI Safety Gridworlds and the MuJoCo gym.", "keywords": [ "Reward Hacking", "AI Safety", "Alignment", "Reinforcement Learning", "Deep Reinforcement Learning", "Reward Tampering", "Sensor Tampering", "Reinforcement Learning with General Utilities" ], "rating_list": [ 6, 5, 1, 5 ], "soundness_list": [ 4, 2, 1, 3 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "UHPnqSTBPO", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Faeze Brahman", "gender": "Female", "institution": "Allen Institute for AI", "country": "US", "position": "Postdoc" }, { "name": "Jaehun Jung", "gender": "Male", "institution": "University of Washington", "country": "", "position": "PhD student" }, { "name": "Yejin Choi", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 40, "n_ref": 63, "n_ref_all": 98, "n_fig": 6, "n_tab": 13, "L_tab": 4996, "n_element_tab": 873, "n_fig_1": 5, "n_tab_1": 9, "L_tab_1": 4373, "n_element_tab_1": 634, "formula_len_all": 955, "formula_len_all_1": 695, "len_all": 163541, "len_all_1": 72857, "len_abs": 1381, "len_title": 122, "len_sents": 42538, "len_sents_1": 28790, "n_sents": 317, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 74, "L_abs": 1397, "title": "Trust or Escalate: LLM Judges with Provable Guarantees for Human Agreement", "abs": "We present a principled approach to provide LLM-based evaluation with a rigorous guarantee of human agreement. We first propose that a reliable evaluation method should not uncritically rely on model preferences for pairwise evaluation, but rather assess the confidence of judge models and selectively decide when to trust its judgement. We then show that under this *selective evaluation* framework, human agreement can be provably guaranteed---such that the model evaluation aligns with that of humans to a user-specified agreement level. As part of our framework, we also introduce *Simulated Annotators*, a novel confidence estimation method that significantly improves judge calibration and thus enables high coverage of evaluated instances. Finally, we propose *Cascaded Selective Evaluation*, where we use cheaper models as initial judges and escalate to stronger models only when necessary---again, while still providing a provable guarantee of human agreement. Experimental results show that Cascaded Selective Evaluation guarantees strong alignment with humans, far beyond what LLM judges could achieve without selective evaluation. For example, on a subset of Chatbot Arena where GPT-4 almost never achieves 80% human agreement, our method, even while employing substantially cost-effective models such as Mistral-7B, *guarantees* over 80% human agreement with almost 80% test coverage.", "keywords": [ "Large Language Model", "LLM", "LLM Judge", "Evaluation", "Alignment" ], "rating_list": [ 6, 8, 8, 10 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 3, 5, 3, 5 ] }, { "paper_id": "UHHOAe1uIS", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "A-Long Jin", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "CN", "position": "Assistant Professor" }, { "name": "Qianru Zhang", "gender": "Female", "institution": "the University of Hong Kong, University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Siu Ming Yiu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Xingwei He", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Yuan Yuan", "gender": "Male", "institution": "Beihang University", "country": "", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 143, "n_ref_all": 221, "n_fig": 28, "n_tab": 10, "L_tab": 4211, "n_element_tab": 281, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1642, "n_element_tab_1": 43, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 413460, "len_all_1": 75270, "len_abs": 3183, "len_title": 145, "len_sents": 153585, "len_sents_1": 36013, "n_sents": 1530, "n_sents_1": 280, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 2194, "title": "TUBench: Benchmarking Large Vision-Language Models on Trustworthiness with Unanswerable Questions", "abs": "Large Vision-Language Models (LVLMs) have achieved remarkable progress on visual perception and linguistic interpretation. Despite their impressive capabilities across various tasks, LVLMs still suffer from the issue of hallucination, which involves generating content that is incorrect or unfaithful to the visual or textual inputs. Traditional benchmarks, such as MME and POPE, evaluate hallucination in LVLMs within the scope of visual question answering (VQA) using answerable questions. However, some questions are unanswerable due to insufficient information in the images, and the performance of LVLMs on such unanswerable questions remains underexplored. To fill in this research blank, we propose TUBench, a benchmark specifically designed to evaluate the reliability of LVLMs using unanswerable questions. TUBench comprises an extensive collection of high-quality, unanswerable questions that are meticulously crafted using ten distinct strategies. To thoroughly evaluate LVLMs, the unanswerable questions in TUBench use images from four diverse domains as visual contexts: screenshots of code snippets, natural images, geometry diagrams, and screenshots of statistical tables. These unanswerable questions are tailored to test LVLMs' trustworthiness in code reasoning, commonsense reasoning, geometric reasoning, and mathematical reasoning related to tables, respectively. We conducted a comprehensive quantitative evaluation of 28 leading foundational models on TUBench, with Gemini-1.5-Pro, the top-performing model, achieving an average accuracy of 69.2\\%, and GPT-4o, the third-ranked model, reaching 66.7\\% average accuracy, in determining whether questions are answerable. Furthermore, our manual analysis of the model outputs reveals that: (1) Gemini-1.5-Pro provides both correct answers and explanations in only 41\\% of cases, and (2) hallucinations are the primary cause of error, accounting for 58.5\\% of the incorrect explanations generated by Gemini-1.5-Pro. These findings highlight that TUBench presents a significant challenge to current LVLMs, and offers a new perspective for evaluating hallucinations and trustworthiness through the lens of unanswerable questions.", "keywords": [ "Vision-Language Models", "Unanswerable Questions", "Benchmark", "Hallucination", "Trustworthiness" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "UHDSE86qiG", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Benjamin Feuer", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Chinmay Hegde", "gender": "Male", "institution": "New York University", "country": "US", "position": "Associate Professor" }, { "name": "Nicky Kriplani", "gender": "Male", "institution": "New York University", "country": "US", "position": "Undergrad student" }, { "name": "Niv Cohen", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "Yuval Lemberg", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 16, "n_ref": 32, "n_ref_all": 56, "n_fig": 9, "n_tab": 4, "L_tab": 151, "n_element_tab": 24, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 95, "n_element_tab_1": 10, "formula_len_all": 617, "formula_len_all_1": 526, "len_all": 109156, "len_all_1": 59632, "len_abs": 1057, "len_title": 91, "len_sents": 35704, "len_sents_1": 28688, "n_sents": 280, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 1065, "title": "Multi-Concept Editing Using Task Arithmetic", "abs": "Model owners often wish to introduce new capabilities into their trained models or remove undesired ones. Task Vectors (TVs) present a promising new approach to editing models after training, allowing simple and controllable addition of new capabilities to the model and the removal of undesired ones. But what happens when the model owner wants to change multiple capabilities? \n\nIn this work, we study the interactions of task vectors in a multi-edit setting for image classifiers and diffusion models. We start by quantifying the overall model degradation induced by applying many specific TVs simultaneously. \nWe show that the overall model performance degrades rapidly as the quantity of TV edits increases.\nFinally, we explore different ways to mitigate this degradation and present an adaptive method to select the most relevant TVs to apply to a diffusion model during inference. Our technique achieves a 94.6% ROC AUC in identifying the correct TV, enabling the effective integration of multiple TV edits while significantly mitigating quality degradation.", "keywords": [ "Task Vectors" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "UGVYezlLcZ", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Issei Sato", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Tokio Kajitsuka", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" } ], "n_formula": 209, "n_formula_1": 22, "n_ref_uni": 42, "n_ref": 103, "n_ref_all": 134, "n_fig": 1, "n_tab": 1, "L_tab": 470, "n_element_tab": 36, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 56, "n_element_tab_1": 12, "formula_len_all": 15774, "formula_len_all_1": 1638, "len_all": 292179, "len_all_1": 75234, "len_abs": 921, "len_title": 100, "len_sents": 82398, "len_sents_1": 31562, "n_sents": 893, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 45, "L_abs": 955, "title": "Optimal Memorization Capacity of Transformers", "abs": "Recent research in the field of machine learning has increasingly focused on the memorization capacity of Transformers, but how efficient they are is not yet well understood.\nWe demonstrate that Transformers can memorize labels with $\\tilde{O}(\\sqrt{N})$ parameters in a next-token prediction setting for $N$ input sequences of length $n$, which is proved to be optimal up to logarithmic factors.\nThis indicates that Transformers can efficiently perform memorization with little influence from the input length $n$ owing to the benefit of parameter sharing.\nWe also analyze the memorization capacity in the sequence-to-sequence setting, and find that $\\tilde{O}(\\sqrt{nN})$ parameters are not only sufficient, but also necessary at least for Transformers with hardmax.\nThese results suggest that while self-attention mechanisms can efficiently identify input sequences, the feed-forward network becomes a bottleneck when associating a label to each token.", "keywords": [ "Transformer", "Self-Attention", "Memorization", "Contextual Mapping", "Permutation Equivariance" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 2, 3, 4 ] }, { "paper_id": "UG62fwmOxZ", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Ehtashamul Haque", "gender": "Male", "institution": "BRAC University", "country": "BD", "position": "Lecturer" }, { "name": "Md Abdul Awal", "gender": "Male", "institution": "The University of Queensland", "country": "AU", "position": "Postdoc" }, { "name": "Mohammad Ali Moni", "gender": "Male", "institution": "Charles Sturt University", "country": "AU", "position": "Full Professor" }, { "name": "Nabil Anan Orka", "gender": "unknown", "institution": "The University of Queensland", "country": "AU", "position": "MS student" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 33, "n_ref": 51, "n_ref_all": 68, "n_fig": 3, "n_tab": 6, "L_tab": 3294, "n_element_tab": 553, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2310, "n_element_tab_1": 385, "formula_len_all": 1270, "formula_len_all_1": 681, "len_all": 127392, "len_all_1": 65827, "len_abs": 1522, "len_title": 108, "len_sents": 38674, "len_sents_1": 27979, "n_sents": 297, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1529, "title": "Fully Quanvolutional Networks for Time Series Classification", "abs": "Quanvolutional neural networks have shown promise in areas such as computer vision and time series analysis. However, their applicability to multi-dimensional and diverse data types remains underexplored. Existing quanvolutional networks heavily rely on classical layers, with minimal quantum involvement, due to inherent limitations in current quanvolution algorithms. In this study, we introduce a new quanvolution algorithm that addresses previous shortcomings related to performance, scalability, and data encoding inefficiencies. Specifically targeting time series data, we propose the Quanv1D layer, which is trainable, capable of handling variable kernel sizes, and can generate a customizable number of feature maps. Unlike previous implementations, Quanv1D can seamlessly integrate at any position within a neural network, effectively processing time series of arbitrary dimensions. Our chosen ansatz and the overall design of Quanv1D contribute to its significant parameter efficiency and inherent regularization properties. In addition to this new layer, we present a new architecture called Fully Quanvolutional Networks (FQN), composed entirely of Quanv1D layers. We tested this lightweight model on 20 UEA and UCR time series classification datasets and compared it against both quantum and classical models, including the current state-of-the-art, ModernTCN. On most datasets, FQN achieved accuracy comparable to the baseline models and even outperformed them on some, all while using a fraction of the parameters.", "keywords": [ "quanvolutional neural networks", "quanvolution", "time series classification" ], "rating_list": [ 5, 3, 3, 6, 6 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 2, 2 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 2, 3 ] }, { "paper_id": "UFzE9njwMG", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Liang Sun", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Staff Software Engineer" }, { "name": "Peiyuan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Rong Jin", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Principal Researcher" }, { "name": "Tian Zhou", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 27, "n_ref": 53, "n_ref_all": 70, "n_fig": 8, "n_tab": 8, "L_tab": 5811, "n_element_tab": 961, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2881, "n_element_tab_1": 521, "formula_len_all": 1361, "formula_len_all_1": 1115, "len_all": 126484, "len_all_1": 66023, "len_abs": 1112, "len_title": 167, "len_sents": 32522, "len_sents_1": 26165, "n_sents": 235, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 119, "L_abs": 1118, "title": "Mitigating Time Discretization Challenges with WeatherODE: A Sandwich Physics-Driven Neural ODE for Weather Forecasting", "abs": "In the field of weather forecasting, traditional models often grapple with discretization errors and time-dependent source discrepancies, which limit their predictive performance. In this paper, we present WeatherODE, a novel one-stage, physics-driven ordinary differential equation (ODE) model designed to enhance weather forecasting accuracy. By leveraging wave equation theory and integrating a time-dependent source model, WeatherODE effectively addresses the challenges associated with time-discretization error and dynamic atmospheric processes. Moreover, we design a CNN-ViT-CNN sandwich structure, facilitating efficient learning dynamics tailored for distinct yet interrelated tasks with varying optimization biases in advection equation estimation. Through rigorous experiments, WeatherODE demonstrates superior performance in both global and regional weather forecasting tasks, outperforming recent state-of-the-art approaches by significant margins of over 40.0\\% and 31.8\\% in root mean square error (RMSE), respectively. The source code is available at https://anonymous.4open.science/r/WeatherODE-5C13/.", "keywords": [ "Weather Forecasting", "NeuralODE", "Time discretization" ], "rating_list": [ 3, 6, 5, 1, 3 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "UFwefiypla", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "AKMMAHBUBUR RAHMAN", "gender": "Male", "institution": "Independent University, Bangladesh", "country": "BD", "position": "Assistant Professor" }, { "name": "Aman Chadha", "gender": "Male", "institution": "Amazon Web Services", "country": "US", "position": "Science Manager" }, { "name": "M Ashraful Amin", "gender": "Male", "institution": "Independent University, Bangladesh", "country": "BD", "position": "Full Professor" }, { "name": "Md Fahim", "gender": "Male", "institution": "Independent University, Bangladesh", "country": "BD", "position": "Researcher" }, { "name": "Md Mofijul Islam", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Md Mubtasim Ahasan", "gender": "Male", "institution": "Center for Computational & Data Sciences (CCDS)", "country": "BD", "position": "Researcher" }, { "name": "Tariq Iqbal", "gender": "unknown", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" }, { "name": "Tasnim Mohiuddin", "gender": "Male", "institution": "Qatar Computing Research Institute", "country": "QA", "position": "Researcher" }, { "name": "amin ali", "gender": "Male", "institution": "Independent University, Bangladesh", "country": "BD", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 13, "n_ref_uni": 42, "n_ref": 122, "n_ref_all": 134, "n_fig": 4, "n_tab": 11, "L_tab": 2865, "n_element_tab": 366, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 353, "n_element_tab_1": 58, "formula_len_all": 826, "formula_len_all_1": 666, "len_all": 191388, "len_all_1": 69424, "len_abs": 1709, "len_title": 119, "len_sents": 61491, "len_sents_1": 34134, "n_sents": 430, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1719, "title": "DM-Codec: Distilling Multimodal Representations for Speech Tokenization", "abs": "Recent advancements in speech-language models have yielded significant improvements in speech tokenization and synthesis. However, effectively mapping the complex, multidimensional attributes of speech into discrete tokens remains challenging. This process demands acoustic, semantic, and contextual information for precise speech representations. Existing speech representations generally fall into two categories: acoustic tokens from audio codecs and semantic tokens from speech self-supervised learning models. Although recent efforts have unified acoustic and semantic tokens for improved performance, they overlook the crucial role of contextual representation in comprehensive speech modeling. Our empirical investigations reveal that the absence of contextual representations results in elevated Word Error Rate (WER) and Word Information Lost (WIL) scores in speech transcriptions. To address these limitations, we propose two novel distillation approaches: (1) a language model (LM)-guided distillation method that incorporates contextual information, and (2) a combined LM and self-supervised speech model (SM)-guided distillation technique that effectively distills multimodal representations (acoustic, semantic, and contextual) into a comprehensive speech tokenizer, termed DM-Codec. The DM-Codec architecture adopts a streamlined encoder-decoder framework with a Residual Vector Quantizer (RVQ) and incorporates the LM and SM during the training process. Experiments show DM-Codec significantly outperforms state-of-the-art speech tokenization models, reducing WER by up to 13.46\\%, WIL by 9.82\\%, and improving speech quality by 5.84\\% and intelligibility by 1.85\\% on the LibriSpeech benchmark dataset.", "keywords": [ "speech tokenizer", "speech codec", "multimodal representation learning", "representation learning", "language model" ], "rating_list": [ 3, 3, 5, 1 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "UFrHWzZENz", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Chengyu Fang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Chenyang Zhu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Chubin Chen", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Kai Li", "gender": "unknown", "institution": "NEC-Labs", "country": "", "position": "NEC Labs, America" }, { "name": "Longxiang Tang", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Qifeng Chen", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Xiu Li", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "ma yue", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 21, "n_ref_uni": 47, "n_ref": 126, "n_ref_all": 168, "n_fig": 22, "n_tab": 10, "L_tab": 887, "n_element_tab": 145, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 726, "n_element_tab_1": 73, "formula_len_all": 1381, "formula_len_all_1": 908, "len_all": 182342, "len_all_1": 59988, "len_abs": 515, "len_title": 125, "len_sents": 50887, "len_sents_1": 26003, "n_sents": 436, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1705, "title": "InstantSwap: Fast Customized Concept Swapping across Sharp Shape Differences", "abs": "Recent advances in Customized Concept Swapping (CCS) enable a text-to-image model to swap a concept in the source image with a customized target concept. However, the existing methods still face the challenges of $\\textit{\\textbf{inconsistency}}$ and $\\textit{\\textbf{inefficiency}}$. They struggle to maintain consistency in both the foreground and background during concept swapping, especially when the shape difference is large between objects. Additionally, they either require time-consuming training processes or involve redundant calculations during inference. To tackle these issues, we introduce InstantSwap, a new CCS method that aims to handle sharp shape disparity at speed. Specifically, we first extract the bbox of the object in the source image $\\textit{automatically}$ based on attention map analysis and leverage the bbox to achieve both foreground and background consistency. For background consistency, we remove the gradient outside the bbox during the swapping process so that the background is free from being modified. For foreground consistency, we employ a cross-attention mechanism to inject semantic information into both source and target concepts inside the box. This helps learn semantic-enhanced representations that encourage the swapping process to focus on the foreground objects. To improve swapping speed, we avoid computing gradients at each timestep but instead calculate them periodically to reduce the number of forward passes, which improves efficiency a lot with a little sacrifice on performance. Finally, we establish a benchmark dataset to facilitate comprehensive evaluation. Extensive evaluations demonstrate the superiority and versatility of InstantSwap.", "keywords": [ "Text-to-image generation", "Image editing", "Customized concept swapping" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "UFRn8203LU", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Hemanth Saratchandran", "gender": "unknown", "institution": "University of Adelaide/Australian Institute of Machine Learning", "country": "AU", "position": "Postdoc" }, { "name": "Simon Lucey", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Full Professor" } ], "n_formula": 102, "n_formula_1": 16, "n_ref_uni": 30, "n_ref": 50, "n_ref_all": 72, "n_fig": 13, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5986, "formula_len_all_1": 745, "len_all": 204059, "len_all_1": 57716, "len_abs": 1726, "len_title": 131, "len_sents": 68423, "len_sents_1": 25562, "n_sents": 774, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1161, "title": "Sampling Theory and Overparameterization: Shaping Loss Landscapes in $\\ell^2$ Regression", "abs": "Overparameterization in neural networks has demonstrated remarkable advantages for both memorization and generalization, particularly in models trained with gradient descent. While much of the existing research focuses on the interplay between overparameterization and gradient-based methods, we explore its influence on the loss landscape of $\\ell^2$ supervised regression problems, independent of any specific optimizer. By leveraging the Nyquist-Shannon-Whittaker sampling theorem, we establish a theoretical link between sampling theory and overparameterized neural networks. Our findings reveal that overparameterization not only exponentially increases the number of global minima but also expands the dimensionality of loss valleys for various $\\ell^2$ regression problems modelled with feedforward neural networks. We empirically validate these theoretical insights across multiple supervised $\\ell^2$ regression tasks, trained with both gradient-based and non-gradient-based optimization algorithms. These results offer fresh perspectives on the advantages of overparameterization in neural network design, independent of the chosen learning algorithm.", "keywords": [ "Shannon sampling theory", "overparameterization." ], "rating_list": [ 8, 3, 1, 5, 6 ], "soundness_list": [ 3, 1, 1, 2, 3 ], "presentation_list": [ 4, 1, 2, 3, 3 ], "contribution_list": [ 3, 2, 1, 2, 3 ], "confidence_list": [ 3, 2, 3, 4, 3 ] }, { "paper_id": "UFKC0lMTdK", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Liangyan Gui", "gender": "Female", "institution": "UIUC", "country": "US", "position": "Assistant Professor" }, { "name": "Shengcao Cao", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Yu-Xiong Wang", "gender": "unknown", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 71, "n_ref": 226, "n_ref_all": 270, "n_fig": 9, "n_tab": 6, "L_tab": 1547, "n_element_tab": 202, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1464, "n_element_tab_1": 123, "formula_len_all": 312, "formula_len_all_1": 228, "len_all": 223915, "len_all_1": 63118, "len_abs": 0, "len_title": 47, "len_sents": 58539, "len_sents_1": 31702, "n_sents": 423, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1292, "title": "Emerging Pixel Grounding in Large Multimodal Models Without Grounding Supervision", "abs": "Current large multimodal models (LMMs) face challenges in grounding, which requires the model to relate language components to visual entities. Contrary to the common practice that fine-tunes LMMs with additional grounding supervision, we find that the grounding ability can in fact emerge in LMMs trained without explicit grounding supervision. To reveal this emerging grounding, we introduce an \"attend-and-segment\" method which analyzes the attention within standard LMMs to provide a point prompt to a segmentation model (e.g., SAM) and perform pixel-level segmentation. Furthermore, to enhance the grounding ability, we propose DiffLMM, an LMM utilizing a diffusion-based visual encoder, as opposed to the standard CLIP visual encoder, and trained with the same weak supervision. Without being constrained by the biases and limited scale of grounding-specific supervision data, our approach enables strong visual grounding while preserving general conversation abilities. We achieve competitive performance on both grounding-specific and general visual question answering benchmarks, compared with grounding LMMs and generalist LMMs, respectively. Notably, we achieve a 44.2 grounding mask recall on grounded conversation generation, outperforming the extensively supervised model GLaMM.", "keywords": [ "Large Multimodal Model", "Foundation Model", "Visual Grounding", "Weakly Supervised Learning" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "UFBabPTgr2", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Henriette M. G. Willems", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "Assistant Professor" }, { "name": "Katayoun Farrahi", "gender": "Female", "institution": "University of Southampton", "country": "GB", "position": "Associate Professor" }, { "name": "Zygis Jocys", "gender": "unknown", "institution": "University of Southampton", "country": "GB", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 36, "n_ref": 54, "n_ref_all": 60, "n_fig": 4, "n_tab": 3, "L_tab": 714, "n_element_tab": 122, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2318, "n_element_tab_1": 140, "formula_len_all": 882, "formula_len_all_1": 882, "len_all": 100060, "len_all_1": 61742, "len_abs": 1286, "len_title": 145, "len_sents": 28685, "len_sents_1": 27924, "n_sents": 198, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1309, "title": "SynthFormer: Equivariant Pharmacophore-based Generation of Molecules for Ligand-Based Drug Design", "abs": "Drug discovery is a complex and resource-intensive process, with significant time and cost investments required to bring new medicines to patients. Recent advancements in generative machine learning (ML) methods offer promising avenues to accelerate early-stage drug discovery by efficiently exploring chemical space. This paper addresses the gap between in silico generative approaches and practical in vitro methodologies, highlighting the need for their integration to optimize molecule discovery. We introduce SynthFormer, a novel ML model that utilizes a 3D equivariant encoder for pharmacophores to generate fully synthesizable molecules, constructed as synthetic trees. Unlike previous methods, SynthFormer incorporates 3D information and provides synthetic paths, enhancing its ability to produce molecules with good docking scores across various proteins. Our contributions include a new methodology for efficient chemical space exploration using 3D information, a novel architecture called Synthformer for translating 3D pharmacophore representations into molecules, and a meaningful embedding space that organizes reagents for drug discovery optimization. Synthformer generates molecules that dock well and enables effective hit expansion and later-stage optimization restricted by synthesis paths.", "keywords": [ "Generative Machine Learning", "Drug Discovery", "Chemical Space Exploration", "3D Equivariant Encoder", "Pharmacophore Representation", "Molecule Generation", "Synthetic Trees", "Synthesis Pathways", "Docking Scores", "Late-stage Optimization", "In Silico to In Vitro Integration", "Deep Learning in Chemistry", "Structure-based Drug Design", "3D Molecular Embedding", "Computational Chemistry", "ML-driven Molecule Design" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 1, 2, 1, 2 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "UF6CEzAVVr", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Haiming Zhu", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Intern" }, { "name": "Jiayuan Xie", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "CN", "position": "Postdoc" }, { "name": "Shengfeng He", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Associate Professor" }, { "name": "Wenqi Shao", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yang Zhou", "gender": "unknown", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yangyang Xu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yong Du", "gender": "unknown", "institution": "Ocean University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Ping Luo", "gender": "unknown", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 14, "n_ref_uni": 54, "n_ref": 136, "n_ref_all": 159, "n_fig": 11, "n_tab": 7, "L_tab": 5888, "n_element_tab": 760, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1649, "n_element_tab_1": 263, "formula_len_all": 502, "formula_len_all_1": 502, "len_all": 159824, "len_all_1": 61651, "len_abs": 313, "len_title": 411, "len_sents": 40415, "len_sents_1": 24209, "n_sents": 288, "n_sents_1": 158, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1046, "title": "Task-Oriented Diffusion Inversion for High-Fidelity Text-based Editing", "abs": "Recent advancements in text-guided diffusion models have unlocked powerful image manipulation capabilities, yet balancing reconstruction fidelity and editability for real images remains a significant challenge. In this work, we introduce TaskOriented Diffusion Inversion (TODInv), a novel framework that inverts and edits real images tailored to specific editing tasks by optimizing prompt embeddings within the extended P \u2217 space. By leveraging distinct embeddings across different U-Net layers and time steps, TODInv seamlessly integrates inversion and editing through reciprocal optimization, ensuring both high fidelity and precise editability. This hierarchical editing mechanism categorizes tasks into structure, appearance, and global edits, optimizing only those embeddings unaffected by the current editing task. Extensive experiments on benchmark dataset reveal TODInv\u2019s superior performance over existing methods, delivering both quantitative and qualitative enhancements while showcasing its versatility with few-step diffusion model.", "keywords": [ "Diffusion Models", "Inversion", "Image Edit" ], "rating_list": [ 5, 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 5 ] }, { "paper_id": "UENQuayzr1", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Andr\u00e9 Biedenkapp", "gender": "unknown", "institution": "University of Freiburg, Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "", "position": "Postdoc" }, { "name": "Florian Walter", "gender": "unknown", "institution": "University of Technology Nuremberg", "country": "DE", "position": "Full Professor" }, { "name": "Gresa Shala", "gender": "Female", "institution": "Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "", "position": "PhD student" }, { "name": "Josif Grabocka", "gender": "Male", "institution": "University of Technology Nuremberg", "country": "DE", "position": "Full Professor" }, { "name": "Pierre Krack", "gender": "Male", "institution": "niversity of Technology Nuremberg", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 5, "n_ref_uni": 42, "n_ref": 88, "n_ref_all": 134, "n_fig": 26, "n_tab": 2, "L_tab": 1090, "n_element_tab": 69, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 207, "n_element_tab_1": 25, "formula_len_all": 1139, "formula_len_all_1": 247, "len_all": 189863, "len_all_1": 63466, "len_abs": 920, "len_title": 79, "len_sents": 57980, "len_sents_1": 30878, "n_sents": 402, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 31, "L_abs": 922, "title": "Efficient Cross-Episode Meta-RL", "abs": "We introduce Efficient Cross-Episodic Transformers (ECET), a new algorithm for online Meta-Reinforcement Learning that addresses the challenge of enabling reinforcement learning agents to perform effectively in previously unseen tasks. We demonstrate how past episodes serve as a rich source of in-context information, which our model effectively distills and applies to new contexts. Our learned algorithm is capable of outperforming the previous state-of-the-art and provides more efficient meta-training while significantly improving generalization capabilities. Experimental results, obtained across various simulated tasks of the MuJoCo, Meta-World and ManiSkill benchmarks, indicate a significant improvement in learning efficiency and adaptability compared to the state-of-the-art. Our approach enhances the agent's ability to generalize from limited data and paves the way for more robust and versatile AI systems.", "keywords": [ "meta-reinforcement learning", "transformers", "in-context learning" ], "rating_list": [ 3, 6, 8, 6 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "UEE13WQlNU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Jiamu Sun", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Liujuan Cao", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Pingyang Dai", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Senior Engineer" }, { "name": "Shouhong Ding", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "researcher" }, { "name": "Taiping Yao", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "researcher" }, { "name": "Xudong Wang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Keyue Zhang", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 17, "n_ref_uni": 16, "n_ref": 23, "n_ref_all": 30, "n_fig": 5, "n_tab": 13, "L_tab": 7880, "n_element_tab": 476, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 4129, "n_element_tab_1": 356, "formula_len_all": 791, "formula_len_all_1": 711, "len_all": 168988, "len_all_1": 61246, "len_abs": 1583, "len_title": 123, "len_sents": 28782, "len_sents_1": 24425, "n_sents": 215, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1594, "title": "CTV-FAS: Compensate Texts with Visuals for Generalizable Face Anti-spoofing", "abs": "Generalizable Face Anti-Spoofing (FAS) approaches have recently gained significant attention for their robustness in unseen scenarios. \nRecent methods incorporate vision-language models into FAS, capitalizing on their remarkable pre-trained performance to enhance generalization. \nThese methods predominantly rely on text prompts to learn the concept of attacks in FAS.\nHowever, certain attacks, such as high-resolution replay attacks, cannot be described linguistically. \nRelying solely on text prompts cannot accurately tackle such attacks, resulting in performance degradation.\nTo tackle these limitations, we introduce a novel framework named CTV-FAS, designed to exploit visual anchors to compensate for the shortcomings of semantic prompts.\nSpecifically, we employ a Self-Supervised Consistency Module (SSCM) to boost the generalization of visual anchors, which utilizes consistency regularization to facilitate visual feature learning.\nSubsequently, a Visual Anchors Updating Module (VAUM) is proposed to incorporate the visual anchors through an adaptive updating scheme, guiding the feature learning process from a visual standpoint.\nFurthermore, we propose an Adaptive Modality Integration Module (AMIM), designed to merge visual and textual information during inference seamlessly. This integration optimizes the synergy between modalities, significantly boosting the efficacy of Face Anti-Spoofing (FAS) tasks.\nOur extensive experimental evaluations and in-depth analysis affirm that our method outperforms current state-of-the-art counterparts with a notable margin of superiority.", "keywords": [ "Face anti-spoofing Vision-language model Domain generalization" ], "rating_list": [ 3, 3, 5, 6, 5 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 5, 4, 4 ] }, { "paper_id": "UE3okxYTUR", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Dacheng Yin", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Fengyun Rao", "gender": "Male", "institution": "WeChat, Tencent Inc.", "country": "CN", "position": "Researcher" }, { "name": "Jian Yang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Wei Zhai", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" }, { "name": "Yang Cao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yizhou Zhou", "gender": "Male", "institution": "WeChat AI", "country": "CN", "position": "Researcher" }, { "name": "Zheng-Jun Zha", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 25, "n_formula_1": 9, "n_ref_uni": 25, "n_ref": 62, "n_ref_all": 78, "n_fig": 4, "n_tab": 4, "L_tab": 1203, "n_element_tab": 195, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 434, "n_element_tab_1": 5, "formula_len_all": 1640, "formula_len_all_1": 610, "len_all": 159007, "len_all_1": 62740, "len_abs": 1886, "len_title": 121, "len_sents": 43392, "len_sents_1": 32373, "n_sents": 321, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1894, "title": "MMAR: Towards Lossless Multi-Modal Auto-Regressive Prababilistic Modeling", "abs": "Recent advancements in multimodal large language models have propelled the development of joint probabilistic models for image understanding and generation. Existing methods that discretize image spaces cause information loss and reduced model capacity. Recent work attempts to integrate diffusion transformers and text autoregression show promise, but it faces challenges in incomplete image\ninformation utilization for understanding tasks \u2014 diffusion transformers encode image information within various noise levels, but image understanding tasks take only clean image as input. In this paper, we develop a novel MultiModal AutoregRessive (MMAR) probabilistic modeling framework based on continuous image representations. Unlike previous methods, MMAR avoids the information loss associated with discretization and the drawback of combining diffusion transformers with AR models. It employs a standalone diffusion-based continuous probabilistic sampler at the image token level on top of LLMs to theoretically ensure lossless image-text joint probabilistic modeling. In practice, to address the substantial optimization difficulties encountered in low-precision training regime common\nfor LLMs, we theoretically derive an optimal diffusion model parameterization that minimizes numerical error. To balance visual understanding and generalization capabilities, we introduce a two-stage training strategy and an extremely large CFG scale for inference. The proposed MMAR significantly demonstrates scaling-up laws with more data and larger model size. Extensive evaluations are\nconducted on 18 image understanding benchmarks. It reveals that MMAR is the first joint image-text modeling framework that approaches comparable performance with traditional MLLMs that employ pretrained CLIP vision encoder, marking a significant step toward lossless joint probabilistic modeling of images and text.", "keywords": [ "Multimodal Large Language Model", "Vision Language Model" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "UDeARVACQi", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chengzhi Mao", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Chenshuang Zhang", "gender": "Female", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Joon Son Chung", "gender": "unknown", "institution": "KAIST", "country": "KR", "position": "Associate Professor" }, { "name": "Junmo Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Kang Zhang", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "In So Kweon", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Emeritus" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 43, "n_ref": 197, "n_ref_all": 253, "n_fig": 12, "n_tab": 3, "L_tab": 381, "n_element_tab": 75, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 411, "n_element_tab_1": 44, "formula_len_all": 561, "formula_len_all_1": 435, "len_all": 166376, "len_all_1": 62610, "len_abs": 787, "len_title": 86, "len_sents": 48987, "len_sents_1": 28512, "n_sents": 379, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 38, "L_abs": 789, "title": "Emerging Tracking from Video Diffusion", "abs": "We find video diffusion models, renowned for their generative capabilities, surprisingly excel at pixel-level object tracking without any explicit training for this task. We introduce a simple and effective method to extract motion representations from video diffusion models, achieving state-of-the-art tracking results. Our approach enables the tracking of identical objects, overcoming limitations of previous methods reliant on intra-frame appearance correspondence. Visualizations and empirical results show that our approach outperforms recent self-supervised tracking methods, including the state-of-the-art, by up to 6 points. Our work demonstrates video generative models can learn intrinsic temporal dynamics of video, and excel in tracking tasks beyond original video synthesis.", "keywords": [ "Pixel-level object tracking", "Temporal correspondence", "Diffusion models" ], "rating_list": [ 5, 8, 6, 5, 6 ], "soundness_list": [ 3, 4, 4, 2, 2 ], "presentation_list": [ 3, 3, 4, 2, 2 ], "contribution_list": [ 2, 4, 4, 2, 2 ], "confidence_list": [ 5, 4, 4, 5, 4 ] }, { "paper_id": "UD0L74wQt9", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Bong Gyun Kang", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "PhD student" }, { "name": "Jaerin Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "KIM KIHOON", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Kyoung Mu Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" } ], "n_formula": 24, "n_formula_1": 10, "n_ref_uni": 28, "n_ref": 71, "n_ref_all": 120, "n_fig": 20, "n_tab": 6, "L_tab": 1112, "n_element_tab": 39, "n_fig_1": 15, "n_tab_1": 4, "L_tab_1": 1339, "n_element_tab_1": 137, "formula_len_all": 1049, "formula_len_all_1": 527, "len_all": 135613, "len_all_1": 63169, "len_abs": 1307, "len_title": 94, "len_sents": 47187, "len_sents_1": 26603, "n_sents": 376, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 46, "L_abs": 1320, "title": "Grokfast: Gradient filters for faster grokking", "abs": "One puzzling artifact in machine learning, dubbed grokking, refers to the case where a model exhibits delayed generalization after numerous training iterations after nearly perfect overfitting. Focusing on the long delay itself on behalf of machine learning practitioners, our primary goal is to accelerate the generalization of a model under the grokking phenomenon. By regarding a series of gradients of a parameter over training iterations as a random signal over time, we can spectrally decompose the parameter trajectories under gradient descent into two components: the fast-varying, overfitting-yielding component, and the slow-varying, generalization-inducing component. This analysis allows us to accelerate the grokking phenomenon more than $\\times 50$ with only a few lines of code that amplifies the slow-varying components of the gradients. The experiments show that our algorithm applies to diverse tasks involving images, languages, and graphs, enabling the practical availability of this peculiar artifact of sudden generalization. Moreover, we reinterpret momentum hyperparameters in gradient-based optimizers as low-pass filters with size-1 windows. This bridges between optimization and classical signal processing literature, suggesting a new type of optimzers augmented with frequecy-domain filters.", "keywords": [ "grokking", "generalization", "acceleration", "gradient filter", "optimization", "low-pass filter" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "UCttY1NZra", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Changqing Fu", "gender": "Male", "institution": "Universit\u00e9 Paris-Dauphine (Paris IX)", "country": "FR", "position": "PhD student" } ], "n_formula": 44, "n_formula_1": 21, "n_ref_uni": 32, "n_ref": 40, "n_ref_all": 94, "n_fig": 11, "n_tab": 9, "L_tab": 3247, "n_element_tab": 172, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 2185, "n_element_tab_1": 128, "formula_len_all": 4831, "formula_len_all_1": 2630, "len_all": 147795, "len_all_1": 60311, "len_abs": 934, "len_title": 131, "len_sents": 42340, "len_sents_1": 21978, "n_sents": 361, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 906, "title": "Conic Linear Units: Orthogonal Equivariance Improves General-Purpose Nonlinearities", "abs": "Most activation functions operate component-wise, which restricts the equivariance of neural networks to permutations. We introduce Conic Linear Units (CoLU) and generalize the symmetry of neural networks to continuous orthogonal groups. By interpreting ReLU as a projection onto its invariant set\u2014the positive orthant\u2014we propose a conic activation function that uses a Lorentz cone instead. Its performance can be further improved by considering multi-head structures, soft scaling, and axis sharing. CoLU associated with low-dimensional cones outperforms the component-wise ReLU in a wide range of models\u2014including MLP, ResNet, and UNet, etc., achieving better loss values and faster convergence. It significantly improves diffusion models' training and performance. CoLU originates from a first-principles approach to various forms of neural networks and fundamentally changes their algebraic structure.", "keywords": [ "Neural Network Architectures", "Activation Functions" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 4, 3, 2, 4 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "UCeZMMyjm2", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Daniel Grillmeyer", "gender": "Male", "institution": "Bayerische Julius-Maximilians-Universit\u00e4t W\u00fcrzburg", "country": "DE", "position": "PhD student" }, { "name": "Michael Stenger", "gender": "Male", "institution": "Bayerische Julius-Maximilians-Universit\u00e4t W\u00fcrzburg", "country": "DE", "position": "PhD student" }, { "name": "Robert Leppich", "gender": "Male", "institution": "Bayerische Julius-Maximilians-Universit\u00e4t W\u00fcrzburg", "country": "DE", "position": "PhD student" }, { "name": "Samuel Kounev", "gender": "Male", "institution": "Bayerische Julius-Maximilians-Universit\u00e4t W\u00fcrzburg", "country": "DE", "position": "Full Professor" }, { "name": "Vanessa Borst", "gender": "Female", "institution": "Bayerische Julius-Maximilians-Universit\u00e4t W\u00fcrzburg", "country": "DE", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 127, "n_ref_all": 156, "n_fig": 3, "n_tab": 9, "L_tab": 5674, "n_element_tab": 322, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2519, "n_element_tab_1": 186, "formula_len_all": 810, "formula_len_all_1": 793, "len_all": 151468, "len_all_1": 69525, "len_abs": 962, "len_title": 0, "len_sents": 47032, "len_sents_1": 32318, "n_sents": 305, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 965, "title": "Time Series Representation Models for Multivariate Time Series Forecasting and Imputation", "abs": "We introduce a multilayered representation learning architecture called Time Series Representation Model (TSRM) for multivariate time series forecasting and imputation. The architecture is structured around hierarchically ordered encoding layers, each dedicated to an independent representation learning task. Each encoding layer contains a representation layer designed to capture diverse temporal patterns and an aggregation layer responsible for combining the learned representations. The architecture is fundamentally based on a Transformer encoder-like configuration, with self-attention mechanisms at its core. The TSRM architecture outperforms state-of-the-art approaches on most of the seven established benchmark datasets considered in our empirical evaluation for both forecasting and imputation tasks while significantly reducing complexity in the form of learnable parameters. The source code is available at https://anonymous.4open.science/r/TSRM-D7BE.", "keywords": [ "time series", "attention", "forecasting", "imputation", "multivariate" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "UCOPY3FZQW", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Lei Zhao", "gender": "unknown", "institution": "Soochow University", "country": "", "position": "Full Professor" }, { "name": "Yan Zhang", "gender": "Female", "institution": "Suchoow University", "country": "CN", "position": "Postdoc" }, { "name": "zhangliml", "gender": "Female", "institution": "Soochow University", "country": "CN", "position": "Full Professor" } ], "n_formula": 27, "n_formula_1": 26, "n_ref_uni": 22, "n_ref": 36, "n_ref_all": 55, "n_fig": 7, "n_tab": 2, "L_tab": 186, "n_element_tab": 29, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 186, "n_element_tab_1": 29, "formula_len_all": 1481, "formula_len_all_1": 1480, "len_all": 78997, "len_all_1": 49610, "len_abs": 1385, "len_title": 141, "len_sents": 22417, "len_sents_1": 20371, "n_sents": 189, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1394, "title": "A novel Visible Multilayer Concept Factorization for Image Data Representation and Clustering", "abs": "Traditional Concept Factorization (CF) methods learn feature of one data point from high-dimensional data space in the form of vector, which leads to the loss of pixel-level neighborhood information in Two-Dimensional (2D) images. In light of this, we present a novel Visible Multilayer Concept Factorization for image-data representation, termed VMCF. Specifically, to uncover deep latent features from complex data, VMCF adopts a multilayer framework, equipped with a \u2018Decomposition, Dimensionality reduction and Data reconstruction\u2019 network ($D^3$-net) in each layer. To obtain locality-preserving features, $D^3$-net firstly performs adaptive graph regularized concept learning on the input data of each layer. Then, $D^3$-net performs 2D feature extraction over the obtained basis images in order to reduce the loss of pixel-level neighborhood information during dimension plunging. The reconstructed data formed by the improved basis images and coefficient matrix is used as input for the next layer. In this way, the dimensions of the original data can gradually decrease at each layer, avoiding information loss caused by sudden dimensionality reduction. Meanwhile, 2D-reduced basis images can mediately improve the quality of new data representations. Extensive numerical experiments on several public image databases have shown that VMCF outperforms other state-of-the-art algorithms.", "keywords": [ "Image Data Representation", "Concept Factorization", "Two-dimensional Feature Extraction", "Data Clustering" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "UC1UD0EIWn", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Bo Peng", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Guangquan Zhang", "gender": "unknown", "institution": "University of Technology Sydney (UTS)", "country": "", "position": "Associate Professor" }, { "name": "Jie Lu", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "Professor" }, { "name": "Zhen Fang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Assistant Professor" } ], "n_formula": 23, "n_formula_1": 21, "n_ref_uni": 73, "n_ref": 127, "n_ref_all": 157, "n_fig": 0, "n_tab": 10, "L_tab": 2879, "n_element_tab": 465, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 1923, "n_element_tab_1": 325, "formula_len_all": 1740, "formula_len_all_1": 1611, "len_all": 163858, "len_all_1": 67815, "len_abs": 1344, "len_title": 115, "len_sents": 30921, "len_sents_1": 27184, "n_sents": 214, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1353, "title": "Rethinking Knowledge Distillation: A Mixture-of-Experts Perspective", "abs": "Knowledge distillation (KD) aims to transfer useful information from a large-scale model (teacher) to a lightweight model (student). \nClassical KD focuses on leveraging the teacher's predictions as soft labels to regularize student training. \nHowever, the exact match of predictions in Kullback-Leibler (KL) divergence could be somewhat in conflict with the classification objective, given that the distribution discrepancies between teacher-generated predictions and ground-truth annotations tend to be fairly severe.\nIn this paper, we rethink the role of teacher predictions from a Mixture-of-Experts (MoE) perspective and transfer knowledge by introducing teacher predictions as latent variables to reformulate the classification objective.\nThis MoE strategy results in breaking down the vanilla classification task into a mixture of easier subtasks with the teacher classifier as a gating function to weigh the importance of subtasks. \nEach subtask is efficiently conquered by distinct experts that are effectively implemented by resorting to multi-level teacher outputs.\nWe further develop a theoretical framework to formulate our method, termed MoE-KD, as an Expectation-Maximization (EM) algorithm and provide proof of the convergence.\nExtensive experiments manifest that MoE-KD outperforms advanced knowledge distillers on mainstream benchmarks.", "keywords": [ "Knowledge Distillation", "Mixture-of-Experts" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "UBCgbAFQKc", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Ashish Kundu", "gender": "Male", "institution": "Cisco Research", "country": "US", "position": "Principal Researcher" }, { "name": "Charles Fleming", "gender": "unknown", "institution": "Cisco", "country": "US", "position": "Researcher" }, { "name": "Ganghua Wang", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Postdoc" }, { "name": "Jayanth Srinivasa", "gender": "Male", "institution": "Cisco", "country": "US", "position": "Researcher" }, { "name": "Jie Ding", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "Associate Professor" }, { "name": "Mingyi Hong", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Xuan Bi", "gender": "unknown", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "Associate Professor" }, { "name": "Xun Xian", "gender": "Male", "institution": "University of Minnesota, Minneapolis", "country": "US", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 2, "n_ref_uni": 46, "n_ref": 85, "n_ref_all": 117, "n_fig": 9, "n_tab": 3, "L_tab": 650, "n_element_tab": 143, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 344, "n_element_tab_1": 72, "formula_len_all": 709, "formula_len_all_1": 272, "len_all": 145724, "len_all_1": 67427, "len_abs": 1476, "len_title": 158, "len_sents": 45591, "len_sents_1": 33567, "n_sents": 351, "n_sents_1": 231, "venue": "ICLR2025", "decision": "withdraw", "L_title": 110, "L_abs": 1482, "title": "On the Vulnerability of Applying Retrieval-Augmented Generation within Knowledge-Intensive Application Domains", "abs": "Retrieval-Augmented Generation (RAG) has been empirically shown to enhance\nthe performance of large language models (LLMs) in knowledge-intensive domains\nsuch as healthcare, finance, and legal contexts. Given a query, RAG retrieves\nrelevant documents from a corpus and integrates them into the LLMs\u2019 generation\nprocess. In this study, we investigate the adversarial robustness of RAG, focusing\nspecifically on examining the retrieval system. First, across 225 different setup\ncombinations of corpus, retriever, query, and targeted information, we show that\nretrieval systems are vulnerable to universal poisoning attacks in medical Q&A. In\nsuch attacks, adversaries generate poisoned documents containing a broad spectrum\nof targeted information, such as personally identifiable information. When these\npoisoned documents are inserted into a corpus, they can be accurately retrieved\nby any users, as long as attacker-specified queries are used. To understand this\nvulnerability, we discovered that the deviation from the query\u2019s embedding to that\nof the poisoned document tends to follow a pattern in which the high similarity\nbetween the poisoned document and the query is retained, thereby enabling precise\nretrieval. Based on these findings, we develop a new detection-based defense to\nensure the safe use of RAG. Through extensive experiments spanning various Q&A\ndomains, we observed that our proposed method consistently achieves excellent\ndetection rates in nearly all cases.", "keywords": [ "Healthcare; Safety; RAG" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "UAzVXdgheU", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 9, "author_info_list": [ { "name": "Chengyu Dong", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Huan Gui", "gender": "unknown", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Jingbo Shang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Ke Yin", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "MS student" }, { "name": "Lichan Hong", "gender": "Male", "institution": "Google ", "country": "US", "position": "Software Engineer" }, { "name": "Long Jin", "gender": "unknown", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Noveen Sachdeva", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Zhe Zhao", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Assistant Professor" }, { "name": "Ed Chi", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 37, "n_ref": 69, "n_ref_all": 105, "n_fig": 7, "n_tab": 7, "L_tab": 2513, "n_element_tab": 337, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2020, "n_element_tab_1": 291, "formula_len_all": 533, "formula_len_all_1": 739, "len_all": 148662, "len_all_1": 69457, "len_abs": 976, "len_title": 144, "len_sents": 40427, "len_sents_1": 30875, "n_sents": 300, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 980, "title": "How To Be A Good Teacher? Process Strong Pretrained Models For Effective Knowledge Distillation", "abs": "Transferring the world knowledge encoded in pretrained models through knowledge distillation is an effective approach to improve the performance of small, task-specific production models. However, the effectiveness of such knowledge transfer drops greatly for strong models that are pretrained in a large scale. In this paper, we explore methods to preprocess strong pretrained models to improve the effectiveness of its knowledge transfer. From a mutual information perspective of distillation effectiveness, we propose to incorporate mutual information-aware optimization into the fine-tuning of strong pretrained models. For small or highly-imbalanced downstream datasets where such optimization is less effective, we further propose to heuristically reweight the MLP blocks, which is inspired by our observation that top MLP blocks often cause the loss of mutual information. Our method enables small student models to benefit from those pretrained models among the strongest.", "keywords": [ "Knowledge distillation", "Pretrained models", "Mutual information", "Sharpness Aware Minimization", "Mixture-of-Experts" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "UAiuV8Plei", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Faming Fang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "WangDanDan", "gender": "Female", "institution": "Huzhou Meteorological Bureau", "country": "CN", "position": "Researcher" }, { "name": "zhu hong chang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 47, "n_ref": 104, "n_ref_all": 118, "n_fig": 8, "n_tab": 2, "L_tab": 266, "n_element_tab": 77, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1302, "n_element_tab_1": 101, "formula_len_all": 931, "formula_len_all_1": 920, "len_all": 123600, "len_all_1": 49331, "len_abs": 1630, "len_title": 113, "len_sents": 32059, "len_sents_1": 21072, "n_sents": 203, "n_sents_1": 137, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1649, "title": "FBSVP: Video Prediction Based on Foreground-Background Separation", "abs": "Video prediction is the process of learning necessary information from historical frames to predict future video frames. \nHow to focus and efficiently learn features from historical frames is a critical step in this process. For any sequence of video frames, \nthe background changes little or remains almost constant, while the foreground changes significantly and is the main focus of our video prediction learning. \nHowever, current known video prediction learning methods do not consider how to utilize the different characteristics of the foreground and background to further improve prediction accuracy. \nTo fully leverage the different characteristics of the foreground and background and enhance prediction accuracy, \nwe propose a Foreground-Background Separation Video Prediction (FBSVP) model in this paper. \nThrough the foreground and background separation module, historical video frames are separated into foreground and background frames. \nIn the video prediction module, the foreground and background frames are predicted and learned separately. \nFirst, the features of historical frames are fused into the current frame through a historical attention fusion module using an attention mechanism. \nThen, the complementary temporal and spatial features are fused through a spatio-temporal fusion module. \nFinally, the learned foreground and background features are fused in the foreground and background fusion module to predict the final video frame. \nExperimental results show that our proposed FBSVP model achieves the best performance on popular video prediction datasets, demonstrating its significant competitiveness in this field.", "keywords": [ "Video Prediction", "Foreground-Background Separation" ], "rating_list": [ 3, 5, 3, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "UAKnJMIBwf", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Masakazu Yoshimura", "gender": "Male", "institution": "Sony Group Corporation", "country": "JP", "position": "Researcher" }, { "name": "Teruaki Hayashi", "gender": "unknown", "institution": "Sony", "country": "JP", "position": "Researcher" }, { "name": "Yota Maeda", "gender": "Male", "institution": "Keio University", "country": "JP", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 48, "n_ref": 97, "n_ref_all": 156, "n_fig": 13, "n_tab": 10, "L_tab": 5604, "n_element_tab": 886, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 4490, "n_element_tab_1": 370, "formula_len_all": 1029, "formula_len_all_1": 798, "len_all": 198144, "len_all_1": 71959, "len_abs": 1189, "len_title": 110, "len_sents": 49295, "len_sents_1": 28016, "n_sents": 453, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1199, "title": "MambaPEFT: Exploring Parameter-Efficient Fine-Tuning for Mamba", "abs": "An ecosystem of Transformer-based models has been established by building large models with extensive data. \nParameter-efficient fine-tuning (PEFT) is a crucial technology for deploying these models to downstream tasks with minimal cost while achieving effective performance. Recently, Mamba, a State Space Model (SSM)-based model, has attracted attention as a potential alternative to Transformers. While many large-scale Mamba-based models have been proposed, efficiently adapting pre-trained Mamba-based models to downstream tasks remains unexplored.\nIn this paper, we conduct an exploratory analysis of PEFT methods for Mamba. We investigate the effectiveness of existing PEFT methods for Transformers when applied to Mamba. We also modify these methods to better align with the Mamba architecture. Additionally, we propose new Mamba-specific PEFT methods that leverage the distinctive structure of Mamba. Our experiments indicate that PEFT performs more effectively for Mamba than Transformers. Lastly, we demonstrate how to effectively combine multiple PEFT methods and provide a framework that outperforms previous works. To ensure reproducibility, we will release the code after publication.", "keywords": [ "parameter-efficient fine-tuning", "PEFT", "Mamba", "State Space Model", "SSM" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "UAEmF5O8J3", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Fangbing Liu", "gender": "unknown", "institution": "Australian National University", "country": "AU", "position": "PhD student" }, { "name": "Qing Wang", "gender": "Female", "institution": "Australian National University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 224, "n_formula_1": 22, "n_ref_uni": 47, "n_ref": 111, "n_ref_all": 164, "n_fig": 5, "n_tab": 7, "L_tab": 5046, "n_element_tab": 373, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 29352, "formula_len_all_1": 1978, "len_all": 267822, "len_all_1": 70518, "len_abs": 1079, "len_title": 96, "len_sents": 72819, "len_sents_1": 32130, "n_sents": 814, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1083, "title": "Generalization of Spectral Graph Neural Networks", "abs": "Spectral graph neural networks (GNNs) have achieved remarkable success across various applications, yet their generalization properties remain poorly understood. This paper bridges this gap by analyzing the impact of graph homophily and architectural choices on the generalization of spectral GNNs. We derive a general form of uniform transductive stability for spectral GNNs and provide an explicit stability analysis for graphs with two node classes, providing a comprehensive framework to understand their generalization. Based on this stability analysis, we establish a generalization error bound, demonstrating that better stability leads to improved generalization. Our theoretical findings reveal that spectral GNNs generalize well on graphs with strong homophily or heterophily but struggle on graphs with weaker structural properties. We also identify conditions under which increasing the polynomial order in spectral GNN architectures may degrade generalization. Empirical results on synthetic and real-world benchmark datasets align closely with our theoretical findings.", "keywords": [ "graph learning", "graph neural networks", "generalization", "node classification", "uniform stability" ], "rating_list": [ 6, 5, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 4, 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 2, 5 ] }, { "paper_id": "UACXMKAz0Z", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Fan Wang", "gender": "Female", "institution": "Alibaba Group", "country": "US", "position": "Senior Staff Algorithm Engineer" }, { "name": "Feng Zhao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Hao Luo", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Hu Yu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 27, "n_ref": 69, "n_ref_all": 90, "n_fig": 8, "n_tab": 10, "L_tab": 2268, "n_element_tab": 292, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1200, "formula_len_all_1": 1088, "len_all": 130908, "len_all_1": 49315, "len_abs": 1447, "len_title": 118, "len_sents": 37634, "len_sents_1": 23204, "n_sents": 283, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1476, "title": "DualFast: Dual-Speedup Framework for Fast Sampling of Diffusion Models", "abs": "Diffusion probabilistic models (DPMs) have achieved impressive success in visual generation. While, they suffer from slow inference speed due to iterative sampling. \n Employing fewer sampling steps is an intuitive solution, but this will also introduces discretization error. Existing fast samplers make inspiring efforts to reduce discretization error through the adoption of high-order solvers, potentially reaching a plateau in terms of optimization. This raises the question: can the sampling process be expedited further?\n In this paper, we re-examine the nature of sampling errors, discerning that they comprise two distinct elements: the widely recognized discretization error and the less acknowledged approximation error.\n Our research elucidates the dynamics between these errors and the step by implementing a dual-error disentanglement strategy.\n Building on these foundations, we introduce an unified and training-free acceleration framework, DualFast, designed to enhance the speed of DPM sampling by concurrently accounting for both error types, thereby minimizing the total sampling error.\n DualFast is seamlessly compatible with existing samplers and significantly boost their sampling quality and speed, particularly in extremely few sampling steps. \n We substantiate the effectiveness of our framework through comprehensive experiments, spanning both unconditional and conditional sampling domains, across both pixel-space and latent-space DPMs.", "keywords": [ "Generative models", "diffusion models", "sampling acceleration" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "UAA2nWUtVl", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Ananth Grama", "gender": "Male", "institution": "Purdue University", "country": "", "position": "Full Professor" }, { "name": "Bolian Li", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Ruqi Zhang", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" }, { "name": "Yifan Wang", "gender": "Female", "institution": "Purdue University", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 50, "n_ref": 132, "n_ref_all": 193, "n_fig": 14, "n_tab": 15, "L_tab": 5458, "n_element_tab": 496, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2013, "n_element_tab_1": 91, "formula_len_all": 383, "formula_len_all_1": 382, "len_all": 199944, "len_all_1": 63335, "len_abs": 1333, "len_title": 109, "len_sents": 52059, "len_sents_1": 28971, "n_sents": 421, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1338, "title": "Cascade Reward Sampling for Efficient Decoding-Time Alignment", "abs": "Aligning large language models (LLMs) with human preferences is critical for their deployment. Recently, decoding-time alignment has emerged as an effective plug-and-play technique that requires no fine-tuning of model parameters. However, generating text that achieves both high reward and high likelihood remains a significant challenge. Existing methods often fail to generate high-reward text or incur substantial computational costs. In this paper, we propose Cascade Reward Sampling (CARDS) to address both issues, guaranteeing the generation of high-reward and high-likelihood text with significantly low costs. Based on our analysis of reward models (RMs) on incomplete text and our observation that high-reward prefixes induce high-reward complete text, we use rejection sampling to iteratively generate small semantic segments to form such prefixes. The segment length is dynamically determined by the predictive uncertainty of LLMs. This strategy guarantees desirable prefixes for subsequent generations and significantly reduces wasteful token re-generations and the number of reward model scoring. Our experiments demonstrate substantial gains in both generation efficiency and alignment ratings compared to the baselines, achieving five times faster text generation and 99% win-ties in GPT-4/Claude-3 helpfulness evaluation.", "keywords": [ "Language Model Alignment", "Large Language Models" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 1, 2, 4, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "U9j40EohfY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Maayan Ehrenberg", "gender": "unknown", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "MS student" }, { "name": "Nir Rosenfeld", "gender": "Male", "institution": "Technion, Technion", "country": "IL", "position": "Assistant Professor" }, { "name": "Roy Ganz", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 10, "n_ref_uni": 46, "n_ref": 65, "n_ref_all": 119, "n_fig": 8, "n_tab": 10, "L_tab": 2214, "n_element_tab": 253, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 824, "formula_len_all_1": 567, "len_all": 191909, "len_all_1": 76293, "len_abs": 1054, "len_title": 126, "len_sents": 70394, "len_sents_1": 37392, "n_sents": 533, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1062, "title": "Adversaries With Incentives: A Strategic Alternative to Adversarial Robustness", "abs": "Adversarial training aims to defend against *adversaries*: malicious opponents whose sole aim is to harm predictive performance in any way possible. This presents a rather harsh perspective, which we assert results in unnecessarily conservative training. As an alternative, we propose to model opponents as simply pursuing their own goals\u2014rather than working directly against the classifier. Employing tools from strategic modeling, our approach enables knowledge or beliefs regarding the opponent's possible incentives to be used as inductive bias for learning. Accordingly, our method of *strategic training* is designed to defend against all opponents within an `incentive uncertainty set'. This resorts to adversarial training when the set is maximal, but offers potential gains when the set can be appropriately reduced. We conduct a series of experiments that show how even mild knowledge regarding the opponent's incentives can be useful, and that the degree of potential gains depends on how these incentives relate to the structure of the learning task.", "keywords": [ "adversarial training", "strategic classification", "adversarial robustness", "strategic robustness" ], "rating_list": [ 6, 3, 6 ], "soundness_list": [ 3, 1, 4 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 3, 1, 3 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "U862lgKUgj", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Jiaqi Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Pan Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Tong Wu", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Xiaoyi Dong", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yuanjun Xiong", "gender": "Male", "institution": "Predera.AI", "country": "US", "position": "Researcher" }, { "name": "Yuhang Zang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Zeyi Sun", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 60, "n_ref": 249, "n_ref_all": 273, "n_fig": 16, "n_tab": 10, "L_tab": 4295, "n_element_tab": 258, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1745, "n_element_tab_1": 39, "formula_len_all": 111, "formula_len_all_1": 111, "len_all": 256842, "len_all_1": 53419, "len_abs": 2383, "len_title": 117, "len_sents": 51980, "len_sents_1": 25847, "n_sents": 389, "n_sents_1": 133, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1324, "title": "Bootstrap3D: Improving Multi-view Diffusion Model with Synthetic Data", "abs": "Recent years have witnessed remarkable progress in multi-view diffusion models for 3D content creation. However, there remains a significant gap in image quality and prompt-following ability compared to 2D diffusion models. A critical bottleneck is the scarcity of high-quality 3D data with detailed captions. To address this challenge, we propose Bootstrap3D, a novel framework that automatically generates an arbitrary quantity of multi-view images to assist in training multi-view diffusion models. Specifically, we introduce a data generation pipeline that employs (1) 2D and video diffusion models to generate multi-view images based on constructed text prompts, and (2) our fine-tuned 3D-aware MV-LLaVA for filtering high-quality data and rewriting inaccurate captions. Leveraging this pipeline, we have generated 1 million high-quality synthetic multi-view images with dense descriptive captions to address the shortage of high-quality 3D data. Furthermore, we present a Training Timestep Reschedule (TTR) strategy that leverages the denoising process to learn multi-view consistency while maintaining the original 2D diffusion prior. Extensive experiments demonstrate that Bootstrap3D can generate high-quality multi-view images with superior aesthetic quality, image-text alignment, and maintained view consistency.", "keywords": [ "3D content creation; Multi-view Diffusion; MLLM" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "U834XHJuqk", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Jonathan Yuyang Zhou", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Yao Xie", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 42, "n_formula_1": 30, "n_ref_uni": 82, "n_ref": 141, "n_ref_all": 183, "n_fig": 7, "n_tab": 6, "L_tab": 9991, "n_element_tab": 2135, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1006, "n_element_tab_1": 138, "formula_len_all": 3172, "formula_len_all_1": 2039, "len_all": 254807, "len_all_1": 75341, "len_abs": 1280, "len_title": 117, "len_sents": 55453, "len_sents_1": 34324, "n_sents": 375, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 63, "L_abs": 1283, "title": "Nonlinear Sequence Embedding by Monotone Variational Inequality", "abs": "In the wild, we often encounter collections of sequential data such as electrocardiograms, motion capture, genomes, and natural language, and sequences may be multichannel or symbolic with nonlinear dynamics. We introduce a method to learn low-dimensional representations of nonlinear sequence and time-series data without supervision which has provable recovery guarantees. The learned representation can be used for downstream machine-learning tasks such as clustering and classification. The method assumes that the observed sequences arise from a common domain, with each sequence following its own autoregressive model, and these models are related through low-rank regularization. We cast the problem as a convex matrix parameter recovery problem using monotone variational inequalities (VIs) and encode the common domain assumption via low-rank constraint across the learned representations, which can learn a subspace approximately spanning the entire domain as well as faithful representations for the dynamics of each individual sequence incorporating the domain information in totality. We show the competitive performance of our method on real-world time-series data with baselines and demonstrate its effectiveness for symbolic text modeling and RNA sequence clustering.", "keywords": [ "Monotone Variational Inequality", "Convex Optimization", "Sequence data", "Time Series", "Representation Learning" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "U7kzrhRxHn", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Anna Bielawska", "gender": "Female", "institution": "Jagiellonian University in Krakow", "country": "PL", "position": "MS student" }, { "name": "Daniel Wilczak", "gender": "Male", "institution": "Faculty of Mathematics and Computer Science of the Jagiellonian University, Jagiellonian University in Krakow", "country": "PL", "position": "Full Professor" }, { "name": "Jacek Tabor", "gender": "Male", "institution": "Jagiellonian University", "country": "PL", "position": "Full Professor" }, { "name": "Patryk Krukowski", "gender": "Male", "institution": "Jagiellonian University in Krakow", "country": "", "position": "PhD student" }, { "name": "Przemys\u0142aw Spurek", "gender": "Male", "institution": "Jagiellonian University Cracow", "country": "PL", "position": "Associate Professor" } ], "n_formula": 21, "n_formula_1": 1, "n_ref_uni": 19, "n_ref": 35, "n_ref_all": 52, "n_fig": 7, "n_tab": 2, "L_tab": 495, "n_element_tab": 40, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 490, "n_element_tab_1": 22, "formula_len_all": 2227, "formula_len_all_1": 41, "len_all": 103582, "len_all_1": 59277, "len_abs": 1311, "len_title": 91, "len_sents": 34343, "len_sents_1": 26196, "n_sents": 322, "n_sents_1": 231, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 1324, "title": "Make Interval Bound Propagation great again", "abs": "In various scenarios motivated by real life, such as medical data analysis, autonomous driving, and adversarial training, we are interested in robust deep networks. A network is robust when a relatively small perturbation of the input cannot lead to drastic changes in output (like change of class, etc.). This falls under the broader scope field of Neural Network Certification (NNC).\nTwo crucial problems in NNC are of profound interest to the scientific community: how to calculate the robustness of a given pre-trained network and how to construct robust networks. The common approach to constructing robust networks is Interval Bound Propagation (IBP). \nThis paper demonstrates that IBP is sub-optimal in the first case due to its susceptibility to the wrapping effect. Even for linear activation, IBP gives strongly sub-optimal bounds. Consequently, one should use strategies immune to the wrapping effect to obtain bounds close to optimal ones. We adapt two classical approaches dedicated to strict computations -- Dubleton Arithmetic and Affine Arithmetic -- to mitigate the wrapping effect in neural networks. These techniques yield precise results for networks with linear activation functions, thus resisting the wrapping effect. As a result, we achieve bounds significantly closer to the optimal level than IBPs.", "keywords": [ "interval arthmetic", "IBP" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "U6gYBJ5vpg", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Chunhua Shen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Hao Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ke Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Peiyuan Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Weian Mao", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "PhD student" }, { "name": "Xiaoran Jiao", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 36, "n_ref": 76, "n_ref_all": 81, "n_fig": 3, "n_tab": 4, "L_tab": 2022, "n_element_tab": 19, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1182, "n_element_tab_1": 87, "formula_len_all": 385, "formula_len_all_1": 385, "len_all": 120164, "len_all_1": 52311, "len_abs": 1378, "len_title": 123, "len_sents": 32022, "len_sents_1": 24780, "n_sents": 236, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1384, "title": "EVO-RDesign: Leveraging Evolutionary Priors for Structure-Based RNA Design", "abs": "Designing RNA sequences based on RNA tertiary structures is a crucial aspect of future RNA design with significant potential to aid drug development. Recently, deep learning-based methods have made progress in this area; however, these methods are constrained by the limited availability of RNA structural data, making it challenging to achieve optimal performance. In this paper, we propose EVO-RDesign, which leverages the evolutionary priors embedded in extensive sequence data to facilitate better RNA sequence design. Specifically, RNA language models have recently been demonstrated to learn the evolutionary information of RNA. Therefore, we consider RNA language models as repositories of evolutionary priors and design a series of adaptors that enable EVO-RDesign to retrieve these priors conditioned on the input RNA structural information. To achieve better performance, the adaptor innovatively inputs RNA structural information and outputs from existing RNA design methods into the language model. Experiments demonstrate that EVO-RDesign outperforms RDesign, achieving a 3.5% increase in sequence recovery on RNAsolo. It also exhibits zero-shot generalization, with gains of 5.1% and 4.1% in sequence recovery on RNA-Puzzles and Rfam, respectively. We also apply in-silico folding to validate whether the generated sequences can fold into the specified 3D RNA backbones.", "keywords": [ "RNA Design", "RNA Tertiary Structures", "Drug Development", "Evolutionary Priors", "RNA Language Model" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "U6UPhLBTcv", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Davide Sapienza", "gender": "Male", "institution": "University of Modena and Reggio Emilia", "country": "IT", "position": "Postdoc" }, { "name": "Elena Govi", "gender": "Female", "institution": "Unimore", "country": "", "position": "Undergrad student" }, { "name": "Luca De Dominicis", "gender": "unknown", "institution": "University of Bologna", "country": "IT", "position": "MS student" }, { "name": "Marko Bertogna", "gender": "Male", "institution": "University of Modena and Reggio Emilia", "country": "IT", "position": "Full Professor" }, { "name": "Nicola Capodieci", "gender": "Male", "institution": "University of Modena and Reggio Emilia", "country": "IT", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 33, "n_ref": 43, "n_ref_all": 65, "n_fig": 2, "n_tab": 6, "L_tab": 829, "n_element_tab": 122, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1268, "n_element_tab_1": 197, "formula_len_all": 753, "formula_len_all_1": 789, "len_all": 113608, "len_all_1": 62765, "len_abs": 1487, "len_title": 108, "len_sents": 32147, "len_sents_1": 27927, "n_sents": 231, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1497, "title": "SyGRID: Synthetically Generated Realistic Industrial Dataset", "abs": "Industrial automation depends on accurate object recognition and localization tasks, such as depth estimation, instance segmentation, object detection, and 6D pose estimation. \nDespite significant advancements, numerous challenges persist, especially within industrial settings. To address these challenges, we propose \nSyGRID, (Synthetically Generated Realistic Industrial Dataset), a new simulated, realistic dataset specifically designed for industrial use cases. \nIts novelty lies in several aspects: the generated frames are photo-realistic images of objects commonly used in industrial settings, capturing their unique material properties; this includes reflection and refraction under varying environmental light conditions. Moreover, SyGRID includes multi-object and multi-instance cluttered scenes accurately accounting for rigid-body physics. \nAiming to narrow the currently existing gap between research and industrial applications, we also provide an exhaustive study on different tasks: namely 2D detection, segmentation, depth estimation and 6D pose estimation. These tasks of computer vision are essential for the integration of robotic applications such as grasping.\nSyGRID can significantly contribute to industrial tasks, leading to more reliable robotic operations. By providing this dataset, we aim to accelerate advancements in robotic automation, facilitating the alignment of current progress in computer vision with the practical demands of industrial robotic applications.", "keywords": [ "artificial intelligence", "simulated dataset", "pose estimation", "industrial automation", "rendering" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "U67J0QNtzo", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Chenglin Li", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Han Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Hongkai Xiong", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Junni Zou", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Maida Cao", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Nuowen Kan", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Postdoc" }, { "name": "Shaohui Li", "gender": "Male", "institution": " Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Wenrui Dai", "gender": "unknown", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 33, "n_ref": 91, "n_ref_all": 121, "n_fig": 18, "n_tab": 7, "L_tab": 1863, "n_element_tab": 221, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 505, "n_element_tab_1": 6, "formula_len_all": 555, "formula_len_all_1": 341, "len_all": 129244, "len_all_1": 57493, "len_abs": 1500, "len_title": 125, "len_sents": 40363, "len_sents_1": 27298, "n_sents": 307, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 77, "L_abs": 1515, "title": "On Disentangled Training for Nonlinear Transform in Learned Image Compression", "abs": "Learned image compression (LIC) has demonstrated superior rate-distortion (R-D) performance compared to traditional codecs, but is challenged by training inefficiency that could incur more than two weeks to train a state-of-the-art model from scratch. Existing LIC methods overlook the slow convergence caused by compacting energy in learning nonlinear transforms. In this paper, we first reveal that such energy compaction consists of two components, \\emph{i.e.}, feature decorrelation and uneven energy modulation. On such basis, we propose a linear auxiliary transform (AuxT) to disentangle energy compaction in training nonlinear transforms. The proposed AuxT obtains coarse approximation to achieve efficient energy compaction such that distribution fitting with the nonlinear transforms can be simplified to fine details. We then develop wavelet-based linear shortcuts (WLSs) for AuxT that leverages wavelet-based downsampling and orthogonal linear projection for feature decorrelation and subband-aware scaling for uneven energy modulation. AuxT is lightweight and plug-and-play to be integrated into diverse LIC models to address the slow convergence issue. Experimental results demonstrate that the proposed approach can accelerate training of LIC models by 2 times and simultaneously achieves an average 1\\% BD-rate reduction. To our best knowledge, this is one of the first successful attempt that can significantly improve the convergence of LIC with comparable or superior rate-distortion performance.", "keywords": [ "learned image compression\uff0c training efficiency\uff0c auxiliary transform" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 4, 2, 3, 4 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "U5gNAmN3h1", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Boyao Wan", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Haifan Gong", "gender": "Male", "institution": "Chinese University of Hong Kong, Shen Zhen", "country": "CN", "position": "PhD student" }, { "name": "Xiaojuan Li", "gender": "Female", "institution": "Shandong University", "country": "CN", "position": "PhD student" }, { "name": "Yanqing Li", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "Full Professor" }, { "name": "Yue Wang", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 68, "n_ref": 103, "n_ref_all": 118, "n_fig": 0, "n_tab": 9, "L_tab": 3983, "n_element_tab": 314, "n_fig_1": 0, "n_tab_1": 7, "L_tab_1": 3800, "n_element_tab_1": 250, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 190496, "len_all_1": 71658, "len_abs": 1595, "len_title": 142, "len_sents": 44178, "len_sents_1": 31059, "n_sents": 297, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1633, "title": "Benchmarking Antimicrobial Peptide Identification with Sequence and Structure Representation", "abs": "The rapid evolution of drug-resistant (DR) microbial has become a severe issue for human health. Antimicrobial peptides (AMPs) are powerful therapeutic drugs for treating DR microbial, but their clinical application is limited by activity and toxicity. Recently, AI has shown its power in discovering the high-activity AMPs, relying on the database of the AMP's wet-lab activity data. However, the activity data from this database are collected from thousands of papers, with their different wet lab experiments setting on one or few types of DR bacteria, have further limits the development of AI methods for AMP identification. Moreover, recently AlphaFold has revolutionized the field of drug discovery, but how can we benefit from the predicted structure for AMP discovery still remains unknown. To address the above challenges, we make two contributions. \\textbf{a)} We construct the \\textbf{DRAMPAtlas 1.0} that contains the training set collected from the public and the testing set from our wet lab experiment. Each AMP sequence is equipped with its 3D structure, activity data, and toxicity, where the activity is about six types of DR bacteria. \\textbf{b)} We conduct extensive experiments for AMP identification, by modeling the 3D structure as voxels or graphs, in conjugate with its sequence information or solely with the structure or sequence. We have made many interesting findings. We hope that our benchmark and findings can benefit the research community to better design the algorithms for high-activity AMP discovery. All code and data associated with the work will be made publicly available after acceptance.", "keywords": [ "Benchmark", "AMP", "drug discovery", "multi-modal learning" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "U5TebOVpfd", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Ge Li", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Jing Su", "gender": "Female", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Jingjing Xu", "gender": "Female", "institution": "Shanghai AI Lab", "country": "", "position": "PhD student" }, { "name": "Jun Zhang", "gender": "Male", "institution": "ByteDance", "country": "CN", "position": "Researcher" }, { "name": "Kechi Zhang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yihong Dong", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yongfei Liu", "gender": "Male", "institution": "Bytedance", "country": "", "position": "Researcher" }, { "name": "Zhi Jin", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 26, "n_ref": 62, "n_ref_all": 72, "n_fig": 2, "n_tab": 12, "L_tab": 4137, "n_element_tab": 611, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1216, "n_element_tab_1": 210, "formula_len_all": 307, "formula_len_all_1": 286, "len_all": 143213, "len_all_1": 62277, "len_abs": 1642, "len_title": 122, "len_sents": 42579, "len_sents_1": 29049, "n_sents": 339, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1727, "title": "CodeDPO: Aligning Code Models with Self Generated and Verified Source Code", "abs": "Code generation models have shown significant potential for programming tasks. \nHowever, existing training methods like supervised fine-tuning face key limitations: they do not effectively teach models to prioritize correct over incorrect solutions in ambiguous situations, nor do they effectively optimize the runtime efficiency of the generated code. \nTo address these challenges, we propose CodeDPO, a framework that integrates preference learning into code generation to improve two key code preference factors: code correctness and efficiency.\nCodeDPO employs a novel dataset construction method, utilizing a self-generation-and-validation mechanism that simultaneously generates and evaluates code and test cases. \nThe underlying assumption is that test cases executable by multiple code snippets provide more reliable validation, and code that passes more tests is more likely to be correct. \nThrough this self-validation process, our PageRank-inspired algorithm iteratively updates the ranking score of each code snippet, ultimately creating a code preference optimization dataset based on correctness and efficiency.\nCodeDPO is flexible and scalable, generating diverse preference optimization data without depending on external resources. \nThrough comprehensive evaluations of five widely used benchmarks, CodeDPO demonstrates significant improvements in correctness and efficiency compared to existing methods. \nOur experiments prove that CodeDPO enhances the capabilities of LLMs in code generation and provides a robust foundation for conducting code preference optimization in more complex and challenging real-world scenarios.\nCode and additional details are available: https://anonymous.4open.science/r/CodeDPO.", "keywords": [ "Code Generation", "Preference Optimization" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "U4ekUAOLsM", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Deepak Sridhar", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Nuno Vasconcelos", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Professor" }, { "name": "Yunsheng Li", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 17, "n_ref": 24, "n_ref_all": 88, "n_fig": 9, "n_tab": 12, "L_tab": 3545, "n_element_tab": 303, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1933, "n_element_tab_1": 174, "formula_len_all": 466, "formula_len_all_1": 433, "len_all": 163829, "len_all_1": 80582, "len_abs": 1695, "len_title": 102, "len_sents": 41299, "len_sents_1": 34219, "n_sents": 319, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1825, "title": "SCHEME: Scalable Channel Mixer for Vision Transformers", "abs": "Vision Transformers have received significant attention due to their impressive performance in many vision tasks. While the token mixer or attention block has been studied in great detail, the channel mixer or feature mixing block (FFN or MLP) has not been explored in depth albeit it accounts for a bulk of the parameters and computation in a model. In this work, we study whether sparse feature mixing can replace the dense connections and confirm this with a block diagonal MLP structure that improves the accuracy by supporting larger expansion ratios. To improve the feature clusters formed by this structure and thereby further improve the accuracy, a lightweight, parameter-free, channel covariance attention (CCA) mechanism is introduced as a parallel branch during training. This design of CCA enables gradual feature mixing across channel groups during training whose contribution decays to zero as the training progresses to convergence. This allows the CCA block to be discarded during inference, thus enabling enhanced performance with no additional computational cost. The resulting $\\textit{Scalable CHannEl MixEr}$ (SCHEME) can be plugged into any ViT architecture to obtain a gamut of models with different trade-offs between complexity and performance by controlling the block diagonal structure size in the MLP. This is shown by the introduction of a new family of SCHEMEformer models. Experiments on image classification, object detection, and semantic segmentation, with different ViT backbones, consistently demonstrate substantial accuracy gains over existing designs, especially under lower FLOPs regimes. The SCHEMEformer family is shown to establish new Pareto frontiers for accuracy vs FLOPS, accuracy vs model size, and accuracy vs throughput, especially for fast transformers of small model size.", "keywords": [ "Vision Transformers", "Channel Mixer", "Efficient", "Scalable", "Attention" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "U49N5V51rU", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Andreas Krebs", "gender": "Male", "institution": "No associate insitute", "country": "DE", "position": "Independent Research" }, { "name": "Andy Yang", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Hattie Zhou", "gender": "Female", "institution": "University of Montreal", "country": "CA", "position": "PhD student" }, { "name": "Michael Hahn", "gender": "Male", "institution": "Universit\u00e4t des Saarlandes", "country": "", "position": "Assistant Professor" }, { "name": "Preetum Nakkiran", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Satwik Bhattamishra", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Xinting Huang", "gender": "Male", "institution": "Universit\u00e4t des Saarlandes", "country": "DE", "position": "PhD student" }, { "name": "Yash Raj Sarrof", "gender": "Male", "institution": "Universit\u00e4t des Saarlandes", "country": "DE", "position": "MS student" } ], "n_formula": 241, "n_formula_1": 15, "n_ref_uni": 40, "n_ref": 201, "n_ref_all": 262, "n_fig": 9, "n_tab": 9, "L_tab": 3953, "n_element_tab": 276, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 415, "n_element_tab_1": 39, "formula_len_all": 24368, "formula_len_all_1": 1727, "len_all": 499335, "len_all_1": 70542, "len_abs": 1099, "len_title": 122, "len_sents": 182226, "len_sents_1": 33866, "n_sents": 1652, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1103, "title": "A Formal Framework for Understanding Length Generalization in Transformers", "abs": "A major challenge for transformers is generalizing to sequences longer than those observed during training. While previous works have empirically shown that transformers can either succeed or fail at length generalization depending on the task, theoretical understanding of this phenomenon remains limited. In this work, we introduce a rigorous theoretical framework to analyze length generalization in causal transformers with learnable absolute positional encodings. In particular, we characterize those functions that are identifiable in the limit from sufficiently long inputs with absolute positional encodings under an idealized inference scheme using a norm-based regularizer. This enables us to prove the possibility of length generalization for a rich family of problems. We experimentally validate the theory as a predictor of success and failure of length generalization across a range of algorithmic and formal language tasks. Our theory not only explains a broad set of empirical observations but also opens the way to provably predicting length generalization capabilities in transformers.", "keywords": [ "transformers", "theory", "length generalization", "expressivity", "analysis", "algorithmic reasoning", "systematic generalization", "formal languages" ], "rating_list": [ 6, 8, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 4, 3 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 4, 2 ], "confidence_list": [ 3, 2, 3, 3, 2 ] }, { "paper_id": "U47ymTS3ut", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Rebekka Burkholz", "gender": "Female", "institution": "Helmholtz Center CISPA for Information Security", "country": "DE", "position": "Associate Professor" }, { "name": "Tom Jacobs", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "PhD student" } ], "n_formula": 49, "n_formula_1": 16, "n_ref_uni": 73, "n_ref": 148, "n_ref_all": 190, "n_fig": 6, "n_tab": 1, "L_tab": 531, "n_element_tab": 16, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4558, "formula_len_all_1": 1503, "len_all": 204205, "len_all_1": 73638, "len_abs": 1258, "len_title": 91, "len_sents": 55296, "len_sents_1": 32802, "n_sents": 562, "n_sents_1": 295, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 1270, "title": "Mask in the Mirror: Implicit Sparsification", "abs": "Continuous sparsification strategies are among the most effective methods for reducing the inference costs and memory demands of large-scale neural networks. A key factor in their success is the implicit $L_1$ regularization induced by jointly learning both mask and weight variables, which has been shown experimentally to outperform explicit $L_1$ regularization. We provide a theoretical explanation for this observation by analyzing the learning dynamics, revealing that early continuous sparsification is governed by an implicit $L_2$ regularization that gradually transitions to an $L_1$ penalty over time. Leveraging this insight, we propose a method to dynamically control the strength of this implicit bias. Through an extension of the mirror flow framework, we establish convergence and optimality guarantees in the context of underdetermined linear regression. Our theoretical findings may be of independent interest, as we demonstrate how to enter the rich regime and show that the implicit bias can be controlled via a time-dependent Bregman potential. To validate these insights, we introduce PILoT, a continuous sparsification approach with novel initialization and dynamic regularization, which consistently outperforms baselines in standard experiments.", "keywords": [ "Continuous sparsification", "Implicit bias", "Mirror flow", "Time-dependent Bregman potential", "Regularization", "Rich regime" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 1, 1 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "U42TkrEDzb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Chao-Han Huck Yang", "gender": "Male", "institution": "NVIDIA Research", "country": "TW", "position": "Researcher" }, { "name": "Chao Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Associate Professor" }, { "name": "Eng Siong Chng", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Helin Wang", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "PhD student" }, { "name": "Siyin Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Yuchen Hu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Zhehuai Chen", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "CHEN CHEN", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 44, "n_ref": 89, "n_ref_all": 106, "n_fig": 3, "n_tab": 5, "L_tab": 1114, "n_element_tab": 139, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1144, "n_element_tab_1": 145, "formula_len_all": 241, "formula_len_all_1": 238, "len_all": 149815, "len_all_1": 77353, "len_abs": 1524, "len_title": 120, "len_sents": 38287, "len_sents_1": 28130, "n_sents": 286, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1532, "title": "Audio Large Language Models Can Be Descriptive Speech Quality Evaluators", "abs": "An ideal multimodal agent should be aware of the quality of its input modalities. Recent advances have enabled large language models (LLMs) to incorporate auditory systems for handling various speech-related tasks. However, most audio LLMs remain unaware of the quality of the speech they process. This limitation arises because speech quality evaluation is typically excluded from multi-task training due to the lack of suitable datasets. To address this, we introduce the first natural language-based speech evaluation corpus, generated from authentic human ratings. In addition to the overall Mean Opinion Score (MOS), this corpus offers detailed analysis across multiple dimensions and identifies causes of quality degradation. It also enables descriptive comparisons between two speech samples (A/B tests) with human-like judgment. Leveraging this corpus, we propose an alignment approach with LLM distillation (ALLD) to guide the audio LLM in extracting relevant information from raw speech and generating meaningful responses. Experimental results demonstrate that ALLD outperforms the previous state-of-the-art regression model in MOS prediction, with a mean square error of 0.17 and an A/B test accuracy of 98.6%. Additionally, the generated responses achieve BLEU scores of 25.8 and 30.2 on two tasks, surpassing the capabilities of task-specific models. This work advances the comprehensive perception of speech signals by audio LLMs, contributing to the development of real-world auditory and sensory intelligent agents.", "keywords": [ "Audio LLM", "Speech quality evaluation" ], "rating_list": [ 6, 8, 8, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "U42IGDU3q5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Jaehyun Kwak", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Ramahdani Muhammad Izaaz Inhar", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "Undergrad student" }, { "name": "Se-Young Yun", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "Assistant Professor" }, { "name": "Sung-Ju Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 14, "n_ref_uni": 36, "n_ref": 67, "n_ref_all": 92, "n_fig": 11, "n_tab": 8, "L_tab": 1555, "n_element_tab": 176, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1612, "n_element_tab_1": 189, "formula_len_all": 1059, "formula_len_all_1": 1091, "len_all": 126057, "len_all_1": 66752, "len_abs": 1386, "len_title": 116, "len_sents": 37300, "len_sents_1": 28752, "n_sents": 289, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1393, "title": "CIRQRS: Evaluating Query Relevance Score in Composed Image Retrieval", "abs": "Composed Image Retrieval (CIR) retrieves relevant images using a reference image and accompanying text that describes how the desired images differ from the reference. However, the commonly used evaluation metric Recall@k only checks if the target image is retrieved, without considering the relevance of other images to the query, potentially leading to user dissatisfaction. We introduce Composed Image Retrieval Query Relevance Score (CIRQRS), an evaluation metric that scores each retrieved image based on its relevance to the query, offering a comprehensive evaluation. CIRQRS is trained using a reward model objective to prefer highly relevant, positive images over less relevant, negative ones. We propose a strategy motivated by self-paced learning to dynamically adjust the negative set based on the relevance of each image by using CIRQRS's current training status. To validate CIRQRS's ability to measure relevance, we created the human-scored FashionIQ (HS-FashionIQ) dataset and compared it with scores from human evaluators. CIRQRS correlates with human scores 2.625 times better than Recall@k, highlighting its superior ability to capture relevance. Additionally, by ranking images based on their CIRQRS, we check if the target image appears in the top k. The results show that CIRQRS achieves state-of-the-art performance on two representative CIR datasets, CIRR and FashionIQ.", "keywords": [ "composed image retrieval", "evaluation metric", "self-paced learning" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "U41Opah9lB", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Chao Ma", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Fei Xie", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CX", "position": "PhD student" }, { "name": "Weidong Cai", "gender": "Male", "institution": "The University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Weijia Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 47, "n_ref": 124, "n_ref_all": 141, "n_fig": 12, "n_tab": 9, "L_tab": 5261, "n_element_tab": 715, "n_fig_1": 10, "n_tab_1": 6, "L_tab_1": 4628, "n_element_tab_1": 661, "formula_len_all": 645, "formula_len_all_1": 647, "len_all": 195391, "len_all_1": 83298, "len_abs": 1477, "len_title": 106, "len_sents": 49137, "len_sents_1": 30586, "n_sents": 407, "n_sents_1": 256, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1470, "title": "VRM: Knowledge Distillation via Virtual Relation Matching", "abs": "Knowledge distillation (KD) aims to transfer the knowledge of a more capable yet cumbersome teacher model to a lightweight student model. In recent years, relation-based KD methods have fallen behind, as instance-matching counterparts dominate in performance. In this paper, we revive relational KD by identifying and tackling several key issues in relational KD, including its susceptibility to overfitting and spurious responses. Specifically, we transfer novelly constructed affinity graphs that compactly encapsulate a wealth of beneficial inter-sample, inter-class, and inter-view correlations by exploiting virtual views and relations as a new kind of knowledge. As a result, the student has access to rich guidance signals and stronger regularisation throughout the distillation process. To further mitigate the adverse impact of spurious responses, we prune the affinity graphs by dynamically detaching redundant and unreliable edges. Extensive experiments on CIFAR-100, ImageNet, and MS-COCO datasets demonstrate the superior performance of the proposed virtual relation matching (VRM) method over a range of tasks, architectures, and set-ups. For instance, VRM for the first time hits 74.0% accuracy for ResNet50-to-MobileNetV2 distillation on ImageNet, and improves DeiT-Ti by 14.11% on CIFAR-100 with a ResNet56 teacher. Thorough analyses are also conducted to gauge the soundness, properties, and complexity of our designs. Code and models will be released.", "keywords": [ "Knowledge distillation", "transfer learning", "virtual knowledge learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "U3UtvOYMiw", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Ahmet \u00dcst\u00fcn", "gender": "Male", "institution": "Cohere For AI", "country": "CA", "position": "Researcher" }, { "name": "Alejandro R. Salamanca", "gender": "Male", "institution": "Technical University of Denmark", "country": "", "position": "MS student" }, { "name": "Nicki Skafte Detlefsen", "gender": "Male", "institution": "Technical University of Denmark", "country": "", "position": "PhD student" }, { "name": "Tim Dettmers", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 4, "n_ref_uni": 43, "n_ref": 86, "n_ref_all": 99, "n_fig": 4, "n_tab": 9, "L_tab": 3917, "n_element_tab": 833, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 575, "n_element_tab_1": 92, "formula_len_all": 302, "formula_len_all_1": 182, "len_all": 147341, "len_all_1": 61162, "len_abs": 1885, "len_title": 126, "len_sents": 35114, "len_sents_1": 29043, "n_sents": 249, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1898, "title": "Seeded LoRA: Collaborative Fine-Tuning Through Seed Initialization of Adapters", "abs": "Parameter-Efficient Fine-Tuning (PEFT) methods facilitate the cost-effective adaptation of pretrained language models to specific tasks and domains. These methods have enabled the open-source community to develop thousands of specialized models tailored to various domains and tasks. Collaborative Fine-Tuning (CoFT) is the paradigm that seeks to merge these specialized models into a single model -- often a routed Mixture-of-Expert (MoE) model -- to achieve better generalization across domains and tasks. However, current CoFT models require a post-merge fine-tuning stage to successfully combine existing models, making CoFT approaches inaccessible to users who lack fine-tuning expertise. In this work, we introduce Seeded LoRA, a novel CoFT approach that does not require post-merge fine-tuning thus enabling plug-and-play PEFT adapter merging. Seeded LoRA significantly outperforms LoRA and MoE LoRA (MoLoRA) approaches, improving by an average of 7 percentage points across a battery of 16 zero-shot tasks and we find that the main benefit from Seeded LoRA comes from mitigating task interference during finetuning. Seeded LoRA works by initializing a model before fine-tuning using a generic seed expert low-rank adapter which was finetuned on a small random subset of the finetuning data such that subsequent fine-tuning runs are initialized in the same optimization subspace. This process enables the integration of any combination of independently fine-tuned models through simple averaging of expert adapter outputs. We show that averaging, or routing with assigning equal probability weights to each expert, is equivalent to grouped convolution, explaining its effectiveness. Additionally, we study subtle routing failures in post-merge fine-tuning and highlight that Seeded LoRA can alleviate most routing failures, making it a suitable base method for future routed CoFT approaches.", "keywords": [ "PEFT", "LoRA", "MoE" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "U3PBITXNG6", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Austin Wang", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Berthy Feng", "gender": "Female", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Bingliang Zhang", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Caifeng Zou", "gender": "Female", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Hongkai Zheng", "gender": "unknown", "institution": "California Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Katherine Bouman", "gender": "unknown", "institution": "California Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Nikola Borislavov Kovachki", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Wenda Chu", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yisong Yue", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Yu Sun", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" }, { "name": "Zachary E Ross", "gender": "unknown", "institution": "California Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Zihui Wu", "gender": "Male", "institution": "Deparment of Computing + Mathematical Sciences, California Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 13, "n_ref_uni": 106, "n_ref": 170, "n_ref_all": 203, "n_fig": 8, "n_tab": 11, "L_tab": 8548, "n_element_tab": 832, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 545, "n_element_tab_1": 86, "formula_len_all": 1870, "formula_len_all_1": 1018, "len_all": 271764, "len_all_1": 66546, "len_abs": 1251, "len_title": 138, "len_sents": 55385, "len_sents_1": 30776, "n_sents": 404, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 89, "L_abs": 1037, "title": "InverseBench: Benchmarking Plug-and-Play Diffusion Models for Scientific Inverse Problems", "abs": "Plug-and-play diffusion prior methods have emerged as a promising research direction for solving inverse problems. \n However, current studies primarily focus on natural image restoration, leaving the performance of these algorithms in scientific inverse problems largely unexplored. To address this gap, we introduce \\textsc{InverseBench}, a unified framework that evaluates diffusion models across five distinct scientific inverse problems. These problems present unique structural challenges that differ from existing benchmarks, arising from critical scientific applications such as black hole imaging, seismology, optical tomography, medical imaging, and fluid dynamics. With \\textsc{InverseBench}, we benchmark 15 inverse problem algorithms that use plug-and-play diffusion prior methods against strong, domain-specific baselines, offering valuable new insights into the strengths and weaknesses of existing algorithms. We open-source the datasets, pre-trained models, and the codebase to facilitate future research and development.", "keywords": [ "inverse problem", "benchmark", "diffusion model" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 4, 4, 3, 3 ], "presentation_list": [ 4, 4, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "U3EzVIsyiP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Fenglong Song", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Kai Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Linghe Kong", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Renjing Pei", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Engineer" }, { "name": "Wenbo Li", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xiaohong Liu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yulun Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Ziqing Zhang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 46, "n_ref": 86, "n_ref_all": 146, "n_fig": 4, "n_tab": 5, "L_tab": 2382, "n_element_tab": 309, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2658, "n_element_tab_1": 313, "formula_len_all": 241, "formula_len_all_1": 241, "len_all": 143016, "len_all_1": 80314, "len_abs": 1282, "len_title": 128, "len_sents": 38393, "len_sents_1": 33450, "n_sents": 346, "n_sents_1": 306, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1164, "title": "Dog-IQA: Standard-guided Zero-shot MLLM for Mix-grained Image Quality Assessment", "abs": "Image quality assessment (IQA) serves as the golden standard for all models' performance in nearly all computer vision fields. However, it still suffers from poor out-of-distribution generalization ability and expensive training costs. To address these problems, we propose Dog-IQA, a standard-guided zero-shot mix-grained IQA method, which is training-free and utilizes the exceptional prior knowledge of multimodal large language models (MLLMs). To obtain accurate IQA scores, namely scores consistent with humans, we design an MLLM-based inference pipeline that imitates human experts. In detail, Dog-IQA applies two techniques. First, Dog-IQA objectively scores with specific standards that utilize MLLM's behavior pattern and minimize the influence of subjective factors. Second, Dog-IQA comprehensively takes local semantic objects and the whole image as input and aggregates their scores, leveraging local and global information. Our proposed Dog-IQA achieves state-of-the-art (SOTA) performance compared with training-free methods, and competitive performance compared with training-based methods in cross-dataset scenarios. Our code will be released soon.", "keywords": [ "Multimodal LLM", "IQA" ], "rating_list": [ 3, 8, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "U2s5hBE1I9", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Gaogang Xie", "gender": "Male", "institution": "Computer Network Information Center, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Guangxing Zhang", "gender": "unknown", "institution": ", Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Jiazheng Tian", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jigang Wen", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Associate Professor" }, { "name": "Kun Xie", "gender": "Female", "institution": "Hunan University", "country": "CN", "position": "Full Professor" }, { "name": "Shiqin Wang", "gender": "Female", "institution": "Hunan University", "country": "CN", "position": "PhD student" } ], "n_formula": 52, "n_formula_1": 40, "n_ref_uni": 34, "n_ref": 50, "n_ref_all": 78, "n_fig": 13, "n_tab": 2, "L_tab": 2812, "n_element_tab": 468, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 2812, "n_element_tab_1": 468, "formula_len_all": 4058, "formula_len_all_1": 3537, "len_all": 144129, "len_all_1": 74927, "len_abs": 1350, "len_title": 122, "len_sents": 44819, "len_sents_1": 27722, "n_sents": 326, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1360, "title": "A Novel Kernel Sparse Coding Method with A Two-stage Acceleration Strategy", "abs": "Sparse coding aims to exploit the latent linear structure of the input data, transforming dense data into sparse data, thereby improving data processing efficiency. However, many real-word signals cannot be expressed linearly, rendering the traditional sparse coding algorithms ineffective. One potential solution is to expand the dimensions of data. In this paper, we verify that the feature mapping of Radial Basis Function (RBF) kernel contains infinite dimensional information, and it does not significantly increase the computational complexity. Based on this, we propose to explore the $l_1$-norm regularization sparse coding method with RBF kernel, and provides a solution with convergence guarantees by leveraging the principle of coordinate descent. Additionally, to accelerate the optimization process, we introduce a novel two-stage acceleration strategy, based on theoretical analysis and empirical observations. Experimental results demonstrate that the two-stage acceleration strategy can reduce processing time by up to 90\\%. Furthermore, when the data size is compressed to about 2\\% of its original scale, the NMAE metric of the proposed method reaches as low as 0.0824 to 0.2195, achieving a significant improvement of up to 47\\% compared to traditional linear sparse coding methods and 36\\% compared to other kernel sparse coding techniques.", "keywords": [ "Sparse Coding", "Kernel Trick", "Acceleration Strategy" ], "rating_list": [ 1, 3, 5, 5 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "U2gYoh8gZG", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Daehyun Kim", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "PhD student" }, { "name": "Donggoo Jung", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "PhD student" }, { "name": "Guanghui Wang", "gender": "Male", "institution": "Toronto Metropolitan University", "country": "CA", "position": "Associate Professor" }, { "name": "Tae Hyun Kim", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "Associate Professor" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 33, "n_ref": 99, "n_ref_all": 125, "n_fig": 8, "n_tab": 5, "L_tab": 2728, "n_element_tab": 152, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 3832, "n_element_tab_1": 143, "formula_len_all": 1693, "formula_len_all_1": 638, "len_all": 121767, "len_all_1": 57786, "len_abs": 1185, "len_title": 130, "len_sents": 37722, "len_sents_1": 24985, "n_sents": 251, "n_sents_1": 154, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1111, "title": "CAGGLE: Color-Aware Guidance with Global and Local Prompts for Exposure Correction", "abs": "In real-world exposure correction, achieving high-quality images requires addressing multi-exposure conditions and managing images containing locally varying brightness. \n While recent deep learning models have improved image correction across various exposure levels, they often struggle in complex scenarios where both under- and over-exposure coexist within an image or in over-saturated areas with sparse pixel information.\n In this paper, we tackle these challenges by proposing a color-aware guidance that employs a global prompt for tone adjustment and a local prompt for maintaining color consistency of the output. \n To achieve this, we present a novel Prompt Interaction Module (PIM) that seamlessly integrates the global and local prompts with the input image features. \n Extensive experiments on multi-exposure benchmark datasets demonstrate that our method achieves state-of-the-art performance, outperforming existing exposure correction methods.\n Our approach sets a new standard in exposure correction, leveraging prompt-based learning for improved color and exposure adjustments.", "keywords": [ "prompt learning", "exposure correction", "image enhancement" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "U2ZtvonVQz", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Kentaro Hoshisashi", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Paolo Barucca", "gender": "Male", "institution": "University College London", "country": "", "position": "Lecturer" }, { "name": "Carolyn Elizabeth Phelan", "gender": "unknown", "institution": "University College London, University of London", "country": "", "position": "Associate Professor" } ], "n_formula": 38, "n_formula_1": 24, "n_ref_uni": 15, "n_ref": 30, "n_ref_all": 39, "n_fig": 6, "n_tab": 2, "L_tab": 784, "n_element_tab": 125, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3032, "formula_len_all_1": 1921, "len_all": 99026, "len_all_1": 52807, "len_abs": 1358, "len_title": 131, "len_sents": 35333, "len_sents_1": 23669, "n_sents": 258, "n_sents_1": 169, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1363, "title": "DC-PINNs: Physics-Informed Neural Networks for Solving Derivative-Constrained PDEs", "abs": "Physics-Informed Neural Networks (PINNs) have emerged as a promising approach for solving partial differential equations (PDEs) using deep learning. However, standard PINNs do not address the problem of constrained PDEs, where the solution must satisfy additional equality or inequality constraints beyond the governing equations. In this paper, we introduce Derivative-Constrained PINNs (DC-PINNs), a novel framework that seamlessly incorporates constraint information into the PINNs training process. DC-PINNs employ a constraint-aware loss function that penalizes constraint violations while simultaneously minimizing the PDE residual. Key components include self-adaptive loss balancing techniques that automatically tune the relative weighting of each term, enhancing training stability, and the use of automatic differentiation to efficiently compute derivatives. This study demonstrates the effectiveness of DC-PINNs on several benchmark problems, from basic to complex, such as quantitative finance and applied physics, including heat diffusion, volatility surface calibration, and incompressible flow dynamics. The results showcase improvements in generating solutions that satisfy the constraints compared to baseline PINNs methods. The DC-PINNs framework opens up new possibilities for solving constrained PDEs in multi-objective optimization problems.", "keywords": [ "PINNs", "Physics-Informed Neural Networks", "Multi-objective Learning", "Partial Differential Equations", "Derivative-Constrained", "Machine Learning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "U2K4bQVWez", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Alfred O. Hero", "gender": "Male", "institution": "University of Michigan", "country": "", "position": "Researcher" }, { "name": "Min Namgung", "gender": "Female", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" }, { "name": "Minoh Jeong", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Postdoc" }, { "name": "Yao-Yi Chiang", "gender": "unknown", "institution": "University of Minnesota, Minneapolis", "country": "", "position": "Associate Professor" }, { "name": "Zae Myung Kim", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" }, { "name": "dongyeop kang", "gender": "Male", "institution": "University of Minnesota", "country": "", "position": "Assistant Professor" } ], "n_formula": 50, "n_formula_1": 30, "n_ref_uni": 52, "n_ref": 116, "n_ref_all": 181, "n_fig": 14, "n_tab": 2, "L_tab": 1659, "n_element_tab": 126, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 615, "n_element_tab_1": 34, "formula_len_all": 4814, "formula_len_all_1": 2538, "len_all": 196963, "len_all_1": 66104, "len_abs": 1638, "len_title": 116, "len_sents": 56347, "len_sents_1": 28413, "n_sents": 462, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1569, "title": "Anchors Aweigh! Sail for Optimal Unified Multi-Modal Representations", "abs": "Multimodal learning plays a crucial role in enabling machine learning models to fuse and utilize diverse data sources, such as text, images, and audio, to support a variety of downstream tasks. A unified representation across various modalities is particularly important for improving efficiency and performance. Recent binding methods, such as ImageBind (Girdhar et al., 2023), typically use a fixed anchor modality to align multimodal data in the anchor modal embedding space. In this paper, we mathematically analyze the fixed anchor binding methods and uncover notable limitations: (1) over-reliance on the choice of the anchor modality, (2) failure to capture intra-modal information, and (3) failure to account for inter-modal correlation among non-anchored modalities. To address these limitations, we propose CentroBind, a simple yet powerful approach that eliminates the need for a fixed anchor; instead, it employs dynamically adjustable centroid-based anchors generated from all available modalities, resulting in a balanced and rich representation space.\nWe theoretically demonstrate that our method captures three crucial properties of multimodal learning: intra-modal learning, inter-modal learning, and multimodal alignment, while also constructing a robust unified representation across all modalities. Our experiments on both synthetic and real-world datasets demonstrate the superiority of the proposed method, showing that dynamic anchor methods outperform all fixed anchor binding methods as the former captures more nuanced multimodal interactions.", "keywords": [ "multimodal learning", "representation learning", "unified representation", "ImageBind", "shared embedding space", "CentroBind" ], "rating_list": [ 5, 8, 6, 3, 8 ], "soundness_list": [ 2, 3, 3, 2, 4 ], "presentation_list": [ 3, 4, 3, 3, 4 ], "contribution_list": [ 2, 4, 3, 2, 4 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "U2FQXhGvip", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Jim Lim", "gender": "Male", "institution": "Purdue University", "country": "", "position": "PhD student" }, { "name": "Ziyu Gong", "gender": "Male", "institution": "Purdue University", "country": "", "position": "PhD student" }, { "name": "David I Inouye", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 38, "n_formula_1": 16, "n_ref_uni": 34, "n_ref": 68, "n_ref_all": 81, "n_fig": 9, "n_tab": 5, "L_tab": 1014, "n_element_tab": 112, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 3321, "n_element_tab_1": 94, "formula_len_all": 2800, "formula_len_all_1": 1008, "len_all": 154791, "len_all_1": 65603, "len_abs": 1521, "len_title": 84, "len_sents": 50581, "len_sents_1": 29660, "n_sents": 379, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1527, "title": "Improving Distribution Matching via Score-Based Priors and Structural Regularization", "abs": "Distribution matching (DM) can be applied to multiple tasks including fair classification, domain adaptation and domain translation.\nHowever, traditional variational DM methods such as VAE-based methods unnecessarily bias the latent distributions towards simple priors or fail to preserve semantic structure leading to suboptimal latent representations.\nTo address these limitations, we propose novel VAE-based DM approach which incorporates a flexible score-based prior and a semantic structure preserving regularization.\nFor score-based priors, the key challenge is that computing the likelihood is expensive.\nYet, our key insight is that computing the likelihood is unnecessary for updating the encoder and thus we prove that the necessary gradients can be computed using only one score function evaluation.\nAdditionally, we adapted the structure preserving regularization inspired by the Gromov-Wasserstein distance, which explicitly encourages the retention of geometric structure in the latent space, even when the latent space has fewer dimensions than the observed space. \nOur framework further allows the integration of semantically meaningful structure from pretrained or foundation models into the latent space, ensuring that the representations preserve semantic structure that is informative and relevant to downstream tasks.\nWe empirically demonstrate that our DM approach leads to better latent representations compared to similar methods for fair classification, domain adaptation, and domain translation tasks.", "keywords": [ "distribution matching", "score-based models", "representation learning" ], "rating_list": [ 5, 6, 3, 3, 3 ], "soundness_list": [ 3, 4, 2, 2, 2 ], "presentation_list": [ 1, 4, 2, 1, 1 ], "contribution_list": [ 1, 3, 1, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "U1o9KaRgYQ", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Bolin Ding", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Senior Director" }, { "name": "Ce Ge", "gender": "Male", "institution": "Fudan", "country": "CN", "position": "Intern" }, { "name": "Daoyuan Chen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Staff" }, { "name": "Haibin Wang", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Jingren Zhou", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yaliang Li", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Staff Engineer" }, { "name": "Dylan Huang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Algorithm Engineer" } ], "n_formula": 8, "n_formula_1": 3, "n_ref_uni": 30, "n_ref": 94, "n_ref_all": 138, "n_fig": 13, "n_tab": 11, "L_tab": 2817, "n_element_tab": 189, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1090, "n_element_tab_1": 84, "formula_len_all": 323, "formula_len_all_1": 116, "len_all": 229323, "len_all_1": 62367, "len_abs": 1313, "len_title": 131, "len_sents": 78209, "len_sents_1": 31364, "n_sents": 475, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1143, "title": "Data-Juicer Sandbox: A Comprehensive Suite for Multimodal Data-Model Co-development", "abs": "The emergence of large-scale multimodal generative models has drastically advanced artificial intelligence, introducing unprecedented levels of performance and functionality. However, optimizing these models remains challenging due to historically isolated paths of model-centric and data-centric developments, leading to suboptimal outcomes and inefficient resource utilization. In response, we present a novel sandbox suite tailored for integrated data-model co-development. This sandbox provides a comprehensive experimental platform, enabling rapid iteration and insight-driven refinement of both data and models. Our proposed ``Probe-Analyze-Refine'' workflow, validated through applications on image-to-text and text-to-video tasks with state-of-the-art LLaVA-like and DiT-based models, yields significant performance boosts, such as topping the VBench leaderboard. We also uncover fruitful insights gleaned from exhaustive benchmarks, shedding light on the critical interplay between data quality, diversity, and model behavior. All codes, datasets and models are openly accessible and continuously maintained to foster future progress.", "keywords": [ "Data Processing", "Multimodal Generative Models", "Sandbox" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "U1T6sq12uj", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Gang Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Haiyuan Zhao", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Haoyu Wang", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "Undergrad student" }, { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Jun Xu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Liang Pang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "Associate Professor" }, { "name": "Sunhao Dai", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Xiao Zhang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zhenhua Dong", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 26, "n_formula_1": 13, "n_ref_uni": 40, "n_ref": 96, "n_ref_all": 119, "n_fig": 5, "n_tab": 6, "L_tab": 3582, "n_element_tab": 500, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1705, "n_element_tab_1": 307, "formula_len_all": 2387, "formula_len_all_1": 1024, "len_all": 169965, "len_all_1": 68216, "len_abs": 1346, "len_title": 120, "len_sents": 53892, "len_sents_1": 31251, "n_sents": 384, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1227, "title": "Perplexity Trap: PLM-Based Retrievers Overrate Low Perplexity Documents", "abs": "Previous studies have found that PLM-based retrieval models exhibit a preference for LLM-generated content, assigning higher relevance scores to these documents even when their semantic quality is comparable to human-written ones. This phenomenon, known as source bias, threatens the sustainable development of the information access ecosystem. However, the underlying causes of source bias remain unexplored. In this paper, we explain the process of information retrieval with a causal graph and discover that PLM-based retrievers learn perplexity features for relevance estimation, causing source bias by ranking the documents with low perplexity higher. Theoretical analysis further reveals that the phenomenon stems from the positive correlation between the gradients of the loss functions in language modeling task and retrieval task. Based on the analysis, a causal-inspired inference-time debiasing method is proposed, called **C**ausal **D**iagnosis and **C**orrection (CDC). Experimental results across three domains demonstrate the superior debiasing effectiveness of CDC, emphasizing the validity of our proposed explanatory framework. Source codes are available at https://github.com/WhyDwelledOnAi/Perplexity-Trap.", "keywords": [ "Large Language Models", "LLM-Generated Content", "Infomation Retrieval", "Source Bias" ], "rating_list": [ 5, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 4, 4 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "U1DjXQeJRx", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Hiroaki Yoshimura", "gender": "Male", "institution": "Waseda University", "country": "JP", "position": "Full Professor" }, { "name": "Razmik Arman Khosrovian", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "Undergrad student" }, { "name": "Takaharu Yaguchi", "gender": "Male", "institution": "Kobe University", "country": "", "position": "Full Professor" }, { "name": "Takashi Matsubara", "gender": "Male", "institution": "Hokkaido University", "country": "JP", "position": "Full Professor" } ], "n_formula": 58, "n_formula_1": 14, "n_ref_uni": 41, "n_ref": 107, "n_ref_all": 142, "n_fig": 8, "n_tab": 5, "L_tab": 1390, "n_element_tab": 111, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1530, "n_element_tab_1": 149, "formula_len_all": 5953, "formula_len_all_1": 1121, "len_all": 227835, "len_all_1": 69037, "len_abs": 1207, "len_title": 131, "len_sents": 78984, "len_sents_1": 29661, "n_sents": 688, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 998, "title": "Poisson-Dirac Neural Networks for Modeling Coupled Dynamical Systems across Domains", "abs": "Deep learning has achieved great success in modeling dynamical systems, providing data-driven simulators to predict complex phenomena, even without known governing equations. However, existing models have two major limitations: their narrow focus on mechanical systems and their tendency to treat systems as monolithic. These limitations reduce their applicability to dynamical systems in other domains, such as electrical and hydraulic systems, and to coupled systems. To address these limitations, we propose Poisson-Dirac Neural Networks (PoDiNNs), a novel framework based on the Dirac structure that unifies the port-Hamiltonian and Poisson formulations from geometric mechanics. This framework enables a unified representation of various dynamical systems across multiple domains as well as their interactions and degeneracies arising from couplings. Our experiments demonstrate that PoDiNNs offer improved accuracy and interpretability in modeling unknown coupled dynamical systems from data.", "keywords": [ "neural ordinary differential equations", "coupled system", "Poisson system", "Dirac structure" ], "rating_list": [ 6, 8, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 4, 3 ], "contribution_list": [ 3, 4, 3, 3, 3 ], "confidence_list": [ 2, 2, 1, 4, 4 ] }, { "paper_id": "U0SijGsCHJ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Fangzhou Lin", "gender": "Male", "institution": "Tohoku University, Tokyo Institute of Technology", "country": "JP", "position": "Researcher" }, { "name": "Haichong Zhang", "gender": "Male", "institution": "Worcester Polytechnic Institute", "country": "", "position": "Assistant Professor" }, { "name": "Haotian Liu", "gender": "Male", "institution": "Worcester Polytechnic Institute", "country": "US", "position": "Undergrad student" }, { "name": "Jose Morales", "gender": "Male", "institution": "Worcester Polytechnic Institute", "country": "US", "position": "PhD student" }, { "name": "Kazunori Yamada", "gender": "Male", "institution": "Tohoku University", "country": "", "position": "Full Professor" }, { "name": "Venkatesh Saligrama", "gender": "unknown", "institution": "Boston University", "country": "US", "position": "Full Professor" }, { "name": "Vijaya B Kolachalama", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "Associate Professor" }, { "name": "Ziming Zhang", "gender": "Male", "institution": "Worcester Polytechnic Institute", "country": "US", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 67, "n_ref": 104, "n_ref_all": 128, "n_fig": 7, "n_tab": 10, "L_tab": 4130, "n_element_tab": 565, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 2439, "n_element_tab_1": 221, "formula_len_all": 1179, "formula_len_all_1": 1219, "len_all": 151084, "len_all_1": 63495, "len_abs": 1134, "len_title": 137, "len_sents": 29074, "len_sents_1": 26735, "n_sents": 241, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1139, "title": "GPS: A Probabilistic Distributional Similarity with Gumbel Priors for Set-to-Set Matching", "abs": "Set-to-set matching aims to identify correspondences between two sets of unordered items by minimizing a distance metric or maximizing a similarity measure. Traditional metrics, such as Chamfer Distance (CD) and Earth Mover\u2019s Distance (EMD), are widely used for this purpose but often suffer from limitations like suboptimal performance in terms of accuracy and robustness, or high computational costs - or both. In this paper, we propose a novel, simple yet effective set-to-set matching similarity measure, GPS, based on Gumbel prior distributions. These distributions are typically used to model the extrema of samples drawn from various distributions. Our approach is motivated by the observation that the distributions of minimum distances from CD, as encountered in real world applications such as point cloud completion, can be accurately modeled using Gumbel distributions. We validate our method on tasks like few-shot image classification and 3D point cloud completion, demonstrating significant improvements over state of-the-art loss functions across several benchmark datasets. Demo code is included in the supplementary file.", "keywords": [ "Set-to-set matching", "Gumbel prior distributions" ], "rating_list": [ 6, 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 2, 4, 2 ] }, { "paper_id": "U0PwxlHiaj", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Hancheng Min", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Postdoc" }, { "name": "Lachlan Ewen MacDonald", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Postdoc" }, { "name": "Rene Vidal", "gender": "unknown", "institution": "University of Pennsylvania", "country": "US", "position": "Full Professor" }, { "name": "Salma Tarmoun", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Ziqing Xu", "gender": "Male", "institution": "The Wharton School, University of Pennsylvania", "country": "US", "position": "PhD student" } ], "n_formula": 56, "n_formula_1": 29, "n_ref_uni": 22, "n_ref": 49, "n_ref_all": 56, "n_fig": 2, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5750, "formula_len_all_1": 1886, "len_all": 129681, "len_all_1": 66629, "len_abs": 1066, "len_title": 128, "len_sents": 40320, "len_sents_1": 30062, "n_sents": 363, "n_sents_1": 247, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1069, "title": "Gradient Descent and Attention Models: Challenges Posed by the Softmax Function", "abs": "Transformers have become ubiquitous in modern machine learning applications, yet their training remains a challenging task often requiring extensive trial and error. Unlike previous architectures, transformers possess unique attention-based components, which can complicate the training process. The standard optimization algorithm, Gradient Descent, consistently underperforms in this context, underscoring the need for a deeper understanding of these difficulties: existing theoretical frameworks fall short and fail to explain this phenomenon. To address this gap, we analyze a simplified Softmax attention model that captures some of the core challenges associated with training transformers. Through a local analysis of the gradient dynamics, we highlight the role of the Softmax function on the local curvature of the loss and show how it can lead to ill-conditioning of these models, which in turn can severely hamper the convergence speed. Our experiments confirm these theoretical findings on the critical impact of Softmax on the dynamics of Gradient Descent.", "keywords": [ "Attention", "Transformers", "Optimization", "Dynamics", "Gradient Descent", "Convergence" ], "rating_list": [ 5, 3, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 1, 2 ], "confidence_list": [ 3, 3, 3, 4, 3 ] }, { "paper_id": "Tzlmaaiytv", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Bingli Wang", "gender": "Male", "institution": "Sichuan Agricultural University", "country": "CN", "position": "Undergrad student" }, { "name": "Guyang Yu", "gender": "Male", "institution": "The 32nd Research Institute of China Electronics Technology Group Corporation", "country": "CN", "position": "Researcher" }, { "name": "Huibin Tan", "gender": "Female", "institution": "National University of Defense Technology", "country": "", "position": "Lecturer" }, { "name": "Jiao Li", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Mengzhu Wang", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Zeheng Wang", "gender": "Male", "institution": "Northeast Normal University", "country": "CN", "position": "MS student" }, { "name": "Houcheng Su", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 22, "n_ref_uni": 17, "n_ref": 27, "n_ref_all": 38, "n_fig": 4, "n_tab": 9, "L_tab": 4935, "n_element_tab": 438, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 2103, "n_element_tab_1": 285, "formula_len_all": 2364, "formula_len_all_1": 1883, "len_all": 110660, "len_all_1": 59144, "len_abs": 1187, "len_title": 118, "len_sents": 29662, "len_sents_1": 23603, "n_sents": 226, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1191, "title": "Open-World Test-Time Training: Self-Training with Contrastive Learning", "abs": "Traditional test-time training (TTT) methods, while addressing domain shifts, often assume a consistent class set that limits their applicability in real-world scenarios with infinite variety. Open-World Test-Time Training (OWTTT) addresses the challenge of generalizing deep learning models to unknown target domain distributions, especially in the presence of strong Out-of-Distribution (OOD) data. Existing TTT methods often struggle to maintain performance when confronted with strong OOD data. In OWTTT, the primary focus has been on distinguishing between strong and weak OOD data. However, during the early stages of TTT, initial feature extraction is hampered by interference from strong OOD and corruptions, leading to reduced contrast and premature classification of certain classes as strong OOD. To handle this problem, we introduce Open World Dynamic Contrastive Learning (OWDCL), an innovative approach that leverage contrastive learning to augment positive sample pairs. This strategy not only enhances contrast in the early stages but also significantly enhances model robustness in later stages. In comparison datasets, our OWDCL model achieves state-of-the-art performance.", "keywords": [ "Open-World", "Test-Time Training", "Self-Training", "Contrastive Learning", "Transfer learning" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "TzdTRC85SQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Bo Li", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Hao Zhang", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "Jinwei Chen", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Peng-Tao Jiang", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "Qibin Hou", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuqi Yang", "gender": "Male", "institution": "Nankai University", "country": "", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 42, "n_ref": 136, "n_ref_all": 146, "n_fig": 5, "n_tab": 6, "L_tab": 2399, "n_element_tab": 269, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2153, "n_element_tab_1": 230, "formula_len_all": 176, "formula_len_all_1": 138, "len_all": 154468, "len_all_1": 71242, "len_abs": 1219, "len_title": 66, "len_sents": 43770, "len_sents_1": 31632, "n_sents": 369, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1226, "title": "Multi-Task Dense Predictions via Unleashing the Power of Diffusion", "abs": "Diffusion models have exhibited extraordinary performance in dense prediction tasks. However, there are few works exploring the diffusion pipeline for multi-task dense predictions. In this paper, we unlock the potential of diffusion models in solving multi-task dense predictions and propose a novel diffusion-based method, called TaskDiffusion, which leverages the conditional diffusion process in the decoder. Instead of denoising the noisy labels for different tasks separately, we propose a novel joint denoising diffusion process to capture the task relations during denoising. To be specific, our method first encodes the task-specific labels into a task-integration feature space to unify the encoding strategy. This allows us to get rid of the cumbersome task-specific encoding process. In addition, we also propose a cross-task diffusion decoder conditioned on task-specific multi-level features, which can model the interactions among different tasks and levels explicitly while preserving efficiency. Experiments show that our TaskDiffusion outperforms previous state-of-the-art methods for all dense prediction tasks on the widely-used PASCAL-Context and NYUD-v2 datasets. Our code will be made publicly available.", "keywords": [ "Diffusion model", "Multi-task learning", "Dense prediction", "Joint denoising", "Cross-task encoding" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "Tz8Li6G2xU", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Dongsung Huh", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "Principal Researcher" } ], "n_formula": 54, "n_formula_1": 23, "n_ref_uni": 18, "n_ref": 27, "n_ref_all": 52, "n_fig": 14, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3985, "formula_len_all_1": 1336, "len_all": 124302, "len_all_1": 60534, "len_abs": 1073, "len_title": 85, "len_sents": 46510, "len_sents_1": 26963, "n_sents": 391, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1078, "title": "A Differentiable Metric for Discovering Groups and Unitary Representations", "abs": "Discovering group structures within data is a significant challenge with broad implications across various scientific domains. The main hurdle stems from the non-differentiable nature of group axioms, hindering their seamless integration into deep learning frameworks. To address this, we introduce a novel differentiable approach that leverages the representation theory of finite groups. Our method employs a unique neural network architecture that models interactions between group elements as multiplications of their matrix representations, coupled with a regularizer that promotes unitarity of these matrices. Furthermore, our model implicitly defines a complexity metric that prioritizes the discovery of group structures. In numerical evaluation, our method successfully recovers group operations from a limited number of observations as well as accurately learning their unitary representations. This work establishes a new avenue for uncovering groups within data, with potential applications in diverse fields, including automatic symmetry discovery in deep learning.", "keywords": [ "group theory", "representation theory", "representation learning", "symmetry discovery", "symbolic relationship" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 1, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "TyycdsNeeg", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Armand Kassa\u00ef Koupa\u00ef", "gender": "Male", "institution": "Universit\u00e9 Pierre et Marie Curie - Paris 6, Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "FR", "position": "PhD student" }, { "name": "Louis Serrano", "gender": "Male", "institution": "Universit\u00e9 Pierre et Marie Curie - Paris 6, Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "", "position": "PhD student" }, { "name": "Pierre ERBACHER", "gender": "Male", "institution": "Universit\u00e9 Pierre et Marie Curie - Paris 6, Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "", "position": "PhD student" }, { "name": "Thomas X Wang", "gender": "Not Specified", "institution": "Sorbonne Universite", "country": "FR", "position": "PhD student" }, { "name": "patrick gallinari", "gender": "Male", "institution": "Criteo AI Lab", "country": "FR", "position": "Research Scientist" } ], "n_formula": 23, "n_formula_1": 9, "n_ref_uni": 26, "n_ref": 62, "n_ref_all": 97, "n_fig": 16, "n_tab": 12, "L_tab": 6132, "n_element_tab": 750, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1657, "n_element_tab_1": 204, "formula_len_all": 1150, "formula_len_all_1": 512, "len_all": 192176, "len_all_1": 70387, "len_abs": 1291, "len_title": 121, "len_sents": 61983, "len_sents_1": 31137, "n_sents": 493, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1296, "title": "Zebra: In-Context and Generative Pretraining for Solving Parametric PDEs", "abs": "Solving time-dependent parametric partial differential equations (PDEs) is challenging, as models must adapt to variations in parameters such as coefficients, forcing terms, and boundary conditions. Data-driven neural solvers either train on data sampled from the PDE parameters distribution in the hope that the model generalizes to new instances or rely on gradient-based adaptation and meta-learning to implicitly encode the dynamics from observations. This often comes with increased inference complexity.\nInspired by the in-context learning capabilities of large language models (LLMs), we introduce Zebra, a novel generative auto-regressive transformer designed to solve parametric PDEs without requiring gradient adaptation at inference. By leveraging in-context information during both pre-training and inference, Zebra dynamically adapts to new tasks by conditioning on input sequences that incorporate context trajectories or preceding states. This approach enables Zebra to flexibly handle arbitrarily sized context inputs and supports uncertainty quantification through the sampling of multiple solution trajectories.\nWe evaluate Zebra across a variety of challenging PDE scenarios, demonstrating its adaptability, robustness, and superior performance compared to existing approaches.", "keywords": [ "PDE", "Adaptation", "In-Context", "Transformer", "Vector-Quantization" ], "rating_list": [ 5, 6, 3, 6, 8 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 3, 3 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "TykT5YB89r", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Shiguang Shan", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Shuang Yang", "gender": "Female", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Zhaoxin Yuan", "gender": "Not Specified", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Xilin Chen", "gender": "Male", "institution": "Institute of Computing Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 46, "n_ref_all": 57, "n_fig": 2, "n_tab": 6, "L_tab": 1824, "n_element_tab": 194, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1587, "n_element_tab_1": 178, "formula_len_all": 401, "formula_len_all_1": 401, "len_all": 117173, "len_all_1": 66411, "len_abs": 2369, "len_title": 125, "len_sents": 34905, "len_sents_1": 29531, "n_sents": 266, "n_sents_1": 241, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 2383, "title": "Not Only Vision: Evolve Visual Speech Recognition via Peripheral Information", "abs": "Visual Speech Recognition (VSR) aims to infer what was said by analyzing the speaker's facial dynamics. However, is reliance solely on visual information sufficient in challenging real-world scenarios? In human visual perception, peripheral vision refers to non-central areas of the visual field, crucial for providing overall awareness and detailed perception of central objects. Similarly, human lip-readers do not rely exclusively on lip movements but integrate contextual cues and prior knowledge to achieve more accurate transcribing. For the first time in machine lip-reading, we frame these non-lip-movement factors into a new concept of semantic-level peripheral information, Specifically, we select three representative types varying in relevance to the spoken content: (1) Contextual peripheral information, such as the general topic or some basic knowledge of the speech, can significantly narrow the range of potential recognition hypotheses. (2) Experiential peripheral information emerges from the recognition process itself. The very act of recognizing speech in a specific language provides implicit knowledge of grammar, word collocations, and related linguistic aspects, thereby guiding the recognition effectively. (3) Perturbative peripheral information introduces disturbance factors into the recognition process, analogous to noise injection in visual tasks. Semantic-level peripheral information is indirectly linked to transcripts; thus fusing it into VSR necessitates strong contextual understanding and inference capabilities. Here, we propose a multimodal learning framework built on a large language model (LLM), leveraging its powerful contextual modeling capabilities to take advantage of peripheral information. Our method's efficacy is demonstrated on two popular datasets. On the widely-used LRS3 dataset, we achieved a Word Error Rate (WER) of 24.5\\% with readily available peripheral information, leading to an impressive 14.3\\% relative improvement over the model without such information. To the best of our knowledge, our work sets a new state-of-the-art when utilizing similar hours of lip-reading videos. We further reported the evaluation on the more challenging AVSpeech dataset. Results across both datasets and various experimental settings demonstrate the promising potential of the proposed semantic-level peripheral information for VSR.", "keywords": [ "Visual Speech Recognition", "Multimodal Learning", "Large Language Model" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 2, 2, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "TySMCLoGVl", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Bryce Ferenczi", "gender": "Male", "institution": "Monash University", "country": "", "position": "PhD student" }, { "name": "Michael Burke", "gender": "unknown", "institution": "Monash University", "country": "AU", "position": "Associate Professor" }, { "name": "Tom Drummond", "gender": "Male", "institution": "University of Melbourne", "country": "", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 45, "n_ref": 72, "n_ref_all": 111, "n_fig": 17, "n_tab": 7, "L_tab": 1879, "n_element_tab": 225, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 432, "n_element_tab_1": 48, "formula_len_all": 630, "formula_len_all_1": 493, "len_all": 157947, "len_all_1": 65503, "len_abs": 1153, "len_title": 85, "len_sents": 45390, "len_sents_1": 30384, "n_sents": 350, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1157, "title": "Efficiently Scanning and Resampling Spatio-Temporal Tasks with Irregular Observations", "abs": "Various works have aimed at combining the inference efficiency of recurrent models and training parallelism of MHA for sequence modeling. However, most of these works focus on tasks with fixed-dimension observation spaces, such as individual tokens in language modeling or pixels in image completion. Variably sized, irregular observation spaces are relatively under-represented, yet they occur frequently in multi-agent domains such as autonomous driving and human-robot interaction. To handle an observation space of varying size, we propose a novel algorithm that alternates between cross-attention between a 2D latent state and observation, and a discounted cumulative sum over the sequence dimension to efficiently accumulate historical information. We find this resampling cycle is critical for performance. To evaluate efficient sequence modeling in this domain, we introduce two multi-agent intention tasks: simulated agents chasing bouncing particles and micromanagement analysis in professional StarCraft II games. Our algorithm achieves comparable accuracy with a lower parameter count, faster training and inference compared to existing methods.", "keywords": [ "sequence modeling", "efficient training", "efficient inference", "spatio-temporal", "multi-agent task" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "Ty7xx0pn0a", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Arun L Bishop", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Khai Nguyen", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Swaminathan Gurumurthy", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Zachary Manchester", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Zico Kolter", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 15, "n_ref_uni": 40, "n_ref": 66, "n_ref_all": 82, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1049, "formula_len_all_1": 847, "len_all": 141706, "len_all_1": 73040, "len_abs": 1180, "len_title": 98, "len_sents": 46912, "len_sents_1": 35304, "n_sents": 350, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1184, "title": "DEQ-MPC : Deep Equilibrium Model Predictive Control", "abs": "Incorporating task-specific priors within a policy or network architecture is crucial for enhancing safety and improving representation and generalization in robotic control problems. Differentiable Model Predictive Control (MPC) layers have proven effective for embedding these priors, such as constraints and cost functions, directly within the architecture, enabling end-to-end training. However, current methods often treat the solver and the neural network as separate, independent entities, leading to suboptimal integration. In this work, we propose a novel approach that co-develops the solver and architecture unifying the optimization solver and network inference problems. Specifically, we formulate this as a joint fixed-point problem over the coupled network outputs and necessary conditions of the optimization problem. We solve this problem in an iterative manner where we alternate between network forward passes and optimization iterations. Through extensive ablations in various robotic control tasks, we demonstrate that our approach results in richer representations and more stable training, while naturally accommodating warm starting, a key requirement for MPC.", "keywords": [ "MPC", "Model Predictive Control", "Optimization", "Differentiable Optimization", "Control" ], "rating_list": [ 3, 6, 6, 3, 5 ], "soundness_list": [ 1, 3, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 5, 3, 4, 4, 4 ] }, { "paper_id": "Ty6TCjKNSF", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Elad Romanov", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Fangzhao Zhang", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Mert Pilanci", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" } ], "n_formula": 186, "n_formula_1": 43, "n_ref_uni": 37, "n_ref": 81, "n_ref_all": 121, "n_fig": 5, "n_tab": 2, "L_tab": 412, "n_element_tab": 43, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 16864, "formula_len_all_1": 3024, "len_all": 243624, "len_all_1": 71063, "len_abs": 1638, "len_title": 160, "len_sents": 67176, "len_sents_1": 28893, "n_sents": 826, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 112, "L_abs": 1652, "title": "Newton Meets Marchenko-Pastur: Massively Parallel Second-Order Optimization with Hessian Sketching and Debiasing", "abs": "Motivated by recent advances in serverless cloud computing, in particular the ``function as a service'' (FaaS) model, \nwe consider the problem of minimizing a convex function in a massively parallel fashion, where communication between workers is limited.\nFocusing on the case of a twice-differentiable objective subject to an L2 penalty, we propose a scheme where the central node (server) effectively runs a Newton method, \noffloading its high per-iteration cost---stemming from the need to invert the Hessian---to the workers. \nIn our solution, workers produce independently coarse but low-bias estimates of the inverse Hessian, using an adaptive sketching scheme. The server then averages the descent directions produced by the workers, yielding a good approximation for the exact Newton step. The main component of our adaptive sketching scheme is a low-complexity procedure for selecting the sketching dimension, an issue that was left largely unaddressed in the existing literature on Hessian sketching for distributed optimization. Our solution is based on ideas from asymptotic random matrix theory, specifically the Marchenko-Pastur law. For Gaussian sketching matrices, we derive non asymptotic guarantees for our algorithm which do not depend on the condition number of the Hessian nor a priori require the sketching dimension to be proportional to the dimension, as is often the case in asymptotic random matrix theory. Lastly, when the objective is self-concordant, we provide convergence guarantees for the approximate Newton's method with noisy Hessians, which may be of independent interest beyond the setting considered in this paper.", "keywords": [ "parallel and distributed optimization", "Newton method", "Marchenko-Pastur law", "Hessian sketching" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "Txxz9fBPcJ", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Bryan Perozzi", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Karthick Panner Selvam", "gender": "Male", "institution": "University of Luxemburg", "country": "", "position": "PhD student" }, { "name": "Mats Brorsson", "gender": "Male", "institution": "University of Luxemburg", "country": "", "position": "Research scientist" }, { "name": "Phitchaya Mangpo Phothilimthana", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Sami Abu-El-Haija", "gender": "Male", "institution": "Research, Google", "country": "", "position": "Research Scientist" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 17, "n_ref": 34, "n_ref_all": 47, "n_fig": 3, "n_tab": 6, "L_tab": 2051, "n_element_tab": 218, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 802, "n_element_tab_1": 86, "formula_len_all": 428, "formula_len_all_1": 428, "len_all": 137867, "len_all_1": 64732, "len_abs": 2106, "len_title": 113, "len_sents": 35468, "len_sents_1": 30253, "n_sents": 261, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1443, "title": "Can LLMs Enhance Performance Prediction for Deep Learning Models?", "abs": "Accurate performance prediction of Deep Learning (DL) models is essential for efficient resource allocation and optimizations in various stages of the DL system stack. While existing approaches can achieve high prediction accuracy, they lack ability to quickly adapt to new hardware environments or emerging workloads. \nThis paper leverages both Graph Neural Networks (GNNs) and Large Language Models (LLMs) to enhance the accuracy and adaptability of DL performance prediction. Our intuition is that GNNs are adept at capturing the structural information of DL models, naturally represented as graphs, while LLMs provide generalization and the ability to quickly adapt to various tasks thanks to extensive pre-training data.\nWe empirically demonstrate that using GNN-derived graph embeddings as inputs to an LLM outperforms traditional representations, including high-level text summary and lossless semi-structured text (e.g., JSON), for this task. Furthermore, we propose a structured pre-training strategy to enable model adaptation to new hardware environments, significantly reducing the need for extensive retraining. Our experiments validate the effectiveness of this approach, showing an 8.8 percentage-point improvement in accuracy over a state-of-the-art GNN baseline. Notably, when adapted to new hardware with few samples, our method achieves a remarkable 30--70 percentage-point increase in accuracy compared to the GNN baseline.", "keywords": [ "Graph Neural Networks", "Graph Tokens", "Large Language Models" ], "rating_list": [ 6, 5, 5, 8, 6 ], "soundness_list": [ 3, 3, 2, 3, 4 ], "presentation_list": [ 4, 2, 3, 4, 4 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 5, 5, 4, 4, 3 ] }, { "paper_id": "TxIrMD6lAN", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Qiang Sun", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Associate Professor" }, { "name": "Ruiyu Wang", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Intern" }, { "name": "Sen Wang", "gender": "Male", "institution": "HoYoverse", "country": "CA", "position": "Researcher" }, { "name": "Xinxin Zuo", "gender": "Female", "institution": "Concordia University", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 83, "n_ref_all": 100, "n_fig": 9, "n_tab": 5, "L_tab": 526, "n_element_tab": 60, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 437, "formula_len_all_1": 437, "len_all": 117057, "len_all_1": 59381, "len_abs": 1470, "len_title": 96, "len_sents": 36702, "len_sents_1": 30362, "n_sents": 247, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1482, "title": "Incremental Learning with Task-Specific Adapters", "abs": "Incremental learning aims to continuously acquire new knowledge while preserving previously learned information. Existing literature primarily focuses on improving model stability, often at the cost of plasticity, to prevent the forgetting of earlier tasks. In this paper, we argue that inter-task differences are the primary driver of catastrophic forgetting. To address this challenge, we propose a novel network architecture compromising two distinct components: one dedicated to learning invariant features shared across tasks and another for capturing task-specific details. Specifically, we repurpose adapters, originally introduced for parameter-efficient fine-tuning, as feature modifiers to capture task-specific details, while the backbone network focuses on learning invariant features. Unlike prior approaches that keep the backbone frozen and only fine-tune {adapters}, we co-train both the backbone network and adapters, employing an additional regularization term that encourages the backbone to learn shared features. Our approach integrates seamlessly with established methods, such as Learning without Forgetting (LwF). Extensive experiments on CIFAR-100 and ImageNet datasets demonstrate that our adapter-based methods consistently outperform non-adapter counterparts across diverse learning scenarios, including various task orders and data scales. Our approach improves both plasticity and stability, effectively addressing the stability-plasticity dilemma.", "keywords": [ "Adaptors", "Incremental Learning", "Computer Vision", "Transfer Learning" ], "rating_list": [ 3, 3, 5, 3, 1 ], "soundness_list": [ 3, 2, 2, 2, 1 ], "presentation_list": [ 3, 3, 3, 2, 1 ], "contribution_list": [ 1, 2, 1, 1, 1 ], "confidence_list": [ 5, 4, 3, 5, 5 ] }, { "paper_id": "Twyc3qZ3py", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Jeevan Thapa", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Rui Li", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Ruochen Shi", "gender": "Female", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 33, "n_ref": 43, "n_ref_all": 74, "n_fig": 10, "n_tab": 7, "L_tab": 2413, "n_element_tab": 292, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 4148, "n_element_tab_1": 270, "formula_len_all": 868, "formula_len_all_1": 790, "len_all": 105089, "len_all_1": 57005, "len_abs": 835, "len_title": 106, "len_sents": 29305, "len_sents_1": 21497, "n_sents": 226, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 836, "title": "Edge Importance Inference Towards Neighborhood Aware GNNs", "abs": "Comprehensive model tuning and meticulous training for determining proper scope of neighborhood where graph neural networks (GNNs) aggregate information requires high computation overhead and significant human effort. We propose a probabilistic GNN model that captures the expansion of neighborhood scope as a stochastic process and adaptively sample edges to identify critical pathways contributing to generating informative node features. We develop a novel variational inference algorithm to jointly approximate the posterior of the count of neighborhood hops and learn GNN weights while accounting for edge importance. Experiments on multiple benchmarks demonstrate that by adapting the neighborhood scope to a given dataset our model outperforms GNN variants that require grid search or heuristics for neighborhood scope selection.", "keywords": [ "GNNs", "variational inference", "stochastic process" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "TwZBQKgwdW", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Chau-Wai Wong", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Assistant Professor" }, { "name": "Gabriel Thompson", "gender": "unknown", "institution": "North Carolina State University", "country": "US", "position": "Undergrad student" }, { "name": "Huaiyu Dai", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Full Professor" }, { "name": "Kai Yue", "gender": "Male", "institution": "North Carolina State University", "country": "", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 14, "n_ref_uni": 31, "n_ref": 46, "n_ref_all": 67, "n_fig": 16, "n_tab": 1, "L_tab": 131, "n_element_tab": 3, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1409, "formula_len_all_1": 833, "len_all": 116304, "len_all_1": 53800, "len_abs": 1495, "len_title": 151, "len_sents": 36810, "len_sents_1": 24570, "n_sents": 302, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1587, "title": "NTK-DFL: Enhancing Decentralized Federated Learning in Heterogeneous Settings via Neural Tangent Kernel", "abs": "Decentralized federated learning (DFL) is a collaborative machine learning framework for training a model across participants without a central server or raw data exchange. DFL faces challenges due to statistical heterogeneity, as participants often possess different data distributions reflecting local environments and user behaviors. Recent work has shown that the neural tangent kernel (NTK) approach, when applied to federated learning in a centralized framework, can lead to improved performance. The NTK-based update mechanism is more expressive than typical gradient descent methods, enabling more efficient convergence and better handling of data heterogeneity. We propose an approach leveraging the NTK to train client models in the decentralized setting, while introducing a synergy between NTK-based evolution and model averaging. This synergy exploits inter-model variance and improves both accuracy and convergence in heterogeneous settings. Our model averaging technique significantly enhances performance, boosting accuracy by at least 10% compared to the mean local model accuracy. Empirical results demonstrate that our approach consistently achieves higher accuracy than baselines in highly heterogeneous settings, where other approaches often underperform. Additionally, it reaches target performance in 4.6 times fewer communication rounds. We validate our approach across multiple datasets, network topologies, and heterogeneity settings to ensure robustness and generalizability. The source code will be available as a link on the discussion forum once it is open.", "keywords": [ "Federated Learning", "Decentralized Federated Learning" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "TwMLUpPg8G", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Shuiwang Ji", "gender": "Male", "institution": "Texas A&M University", "country": "US", "position": "Professor" }, { "name": "Shurui Gui", "gender": "Male", "institution": "Texas A&M University", "country": "US", "position": "PhD student" }, { "name": "Xiner Li", "gender": "Female", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 20, "n_ref_uni": 47, "n_ref": 91, "n_ref_all": 123, "n_fig": 16, "n_tab": 6, "L_tab": 1041, "n_element_tab": 86, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 815, "n_element_tab_1": 22, "formula_len_all": 57193, "formula_len_all_1": 50776, "len_all": 285433, "len_all_1": 134939, "len_abs": 1421, "len_title": 126, "len_sents": 86036, "len_sents_1": 42002, "n_sents": 555, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1542, "title": "Discovering Physics Laws of Dynamical Systems via Invariant Function Learning", "abs": "We consider learning underlying laws of dynamical systems governed by ordinary differential equations (ODE). A key challenge is how to discover intrinsic dynamics across multiple environments while circumventing environment-specific mechanisms. Unlike prior work, we tackle more complex environments where changes extend beyond function coefficients to entirely different function forms. For example, we demonstrate the discovery of ideal pendulum's natural motion $\\alpha \\sin{\\theta_t}$ by observing pendulum dynamics in different environments, such as the damped environment $\\alpha \\sin(\\theta_t) - \\rho \\omega_t$ and powered environment $\\alpha \\sin(\\theta_t) + \\rho \\frac{\\omega_t}{\\left|\\omega_t\\right|}$. Here, we formulate this problem as an $invariant\\ function\\ learning$ task and propose a new method, known as $\\mathbf{D}$isentanglement of $\\mathbf{I}$nvariant $\\mathbf{F}$unctions (DIF), that is grounded in causal analysis. We propose a causal graph and design an encoder-decoder hypernetwork that explicitly disentangles invariant functions from environment-specific dynamics. The discovery of invariant functions is guaranteed by our information-based principle that enforces the independence between extracted invariant functions and environments. Quantitative comparisons with meta-learning and invariant learning baselines on three ODE systems demonstrate the effectiveness and efficiency of our method. Furthermore, symbolic regression explanation results highlight the ability of our framework to uncover intrinsic laws.", "keywords": [ "dynamical system", "ordinary differential equation", "invariant learning" ], "rating_list": [ 6, 3, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 1, 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 3, 1 ] }, { "paper_id": "TwJrTz9cRS", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Lilian Tang", "gender": "Female", "institution": "University of Surrey", "country": "", "position": "Full Professor" }, { "name": "Qiushi Huang", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "PhD student" }, { "name": "Tom Ko", "gender": "Male", "institution": "ByteDance AI Lab", "country": "HK", "position": "Researcher" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Zhan Zhuang", "gender": "unknown", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 40, "n_ref": 68, "n_ref_all": 102, "n_fig": 6, "n_tab": 14, "L_tab": 5065, "n_element_tab": 646, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3716, "n_element_tab_1": 367, "formula_len_all": 727, "formula_len_all_1": 405, "len_all": 164377, "len_all_1": 66371, "len_abs": 574, "len_title": 129, "len_sents": 40792, "len_sents_1": 25841, "n_sents": 318, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 81, "L_abs": 577, "title": "HiRA: Parameter-Efficient Hadamard High-Rank Adaptation for Large Language Models", "abs": "We propose Hadamard High-Rank Adaptation (HiRA), a parameter-efficient fine-tuning (PEFT) method that enhances the adaptability of Large Language Models (LLMs). While Low-rank Adaptation (LoRA) is widely used to reduce resource demands, its low-rank updates may limit its expressiveness for new tasks. HiRA addresses this by using a Hadamard product to retain high-rank update parameters, improving the model capacity. Empirically, HiRA outperforms LoRA and its variants on several tasks, with extensive ablation studies validating its effectiveness. Our code will be released.", "keywords": [ "Parametric-efficient fine-tuning", "Large Language Model" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "TvwsOrl865", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Gal Maman", "gender": "Female", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "MS student" }, { "name": "Ronen Talmon", "gender": "unknown", "institution": "Technion - Israel Institute of Technology", "country": "IL", "position": "Full Professor" } ], "n_formula": 26, "n_formula_1": 11, "n_ref_uni": 48, "n_ref": 153, "n_ref_all": 246, "n_fig": 20, "n_tab": 13, "L_tab": 3532, "n_element_tab": 352, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 548, "n_element_tab_1": 31, "formula_len_all": 1402, "formula_len_all_1": 651, "len_all": 238916, "len_all_1": 75187, "len_abs": 1430, "len_title": 99, "len_sents": 84610, "len_sents_1": 35685, "n_sents": 639, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1437, "title": "Diffusion Transportation Cost for Domain Adaptation", "abs": "In recent years, there has been considerable interest in leveraging the Optimal Transport (OT) problem for domain adaptation, a strategy shown to be highly effective. \nHowever, a less explored aspect is the choice of the transportation cost function, as most existing methods rely on the pairwise squared Euclidean distances for the transportation cost, potentially overlooking important intra-domain geometries.\nThis paper presents Diffusion-OT, a new transport cost for the OT problem, designed specifically for domain adaptation. By utilizing concepts and tools from the field of manifold learning, specifically diffusion geometry, we derive an operator that accounts for the intra-domain relationships, thereby extending beyond the conventional inter-domain distances.\nThis operator, which quantifies the probability of transporting between source and target samples, forms the basis for our transportation cost. \nWe provide proof that the proposed operator is in fact a diffusion operator, demonstrating that the cost function is defined by an anisotropic diffusion process between the domains.\nIn addition, to enhance performance, we integrate source labels into the operator, thereby guiding the anisotropic diffusion according to the classes.\nWe showcase the effectiveness of Diffusion-OT through comprehensive experiments, demonstrating its superior performance compared to recent methods across various benchmarks and datasets.", "keywords": [ "Optimal Transport", "Domain Adaptation", "Diffusion geometry", "Manifold learning", "Kernel methods", "Riemannian manifolds." ], "rating_list": [ 6, 5, 6, 3, 6 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 5, 4 ] }, { "paper_id": "TvvT4wjEPf", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Vele Tosevski", "gender": "unknown", "institution": "Lorica Cybersecurity", "country": "CA", "position": "Researcher" }, { "name": "glenngulak", "gender": "Male", "institution": "University of Toronto, University of Toronto", "country": "CA", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 24, "n_ref": 75, "n_ref_all": 89, "n_fig": 4, "n_tab": 6, "L_tab": 1085, "n_element_tab": 98, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1074, "n_element_tab_1": 119, "formula_len_all": 437, "formula_len_all_1": 294, "len_all": 134260, "len_all_1": 70619, "len_abs": 1660, "len_title": 122, "len_sents": 48443, "len_sents_1": 33674, "n_sents": 365, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1671, "title": "Towards Practical Large-Scale Privacy-Preserving Recurrent Neural Networks", "abs": "Recurrent neural networks (RNNs) are used for a variety of applications such as speech recognition and financial forecasting where data privacy is an ongoing concern. Fully homomorphic encryption (FHE) facilitates computation over encrypted data, enabling third-party services like machine learning inference while keeping client data private. Previous studies have examined RNN inference over encrypted data using FHE, albeit on a small scale, though impractical due to the computational costs. This work advances insights that make large-scale RNN evaluation over encrypted data practical. A problem that prohibits the scaling of privacy-preserving RNNs is overflow in the ciphertext message space. As the number of model parameters increases, the size of the domain during multiply-accumulate operations increases, causing inaccuracies in computation. Attempts to mitigate this problem, such as splitting the message into several ciphertexts, cause an exponential increase in computation, making latency-sensitive applications like RNNs impractical. A novel regularization technique is proposed that mitigates the effects of numerical overflow during training. This allows use of one ciphertext only and reduces the complexity of the encryption parameters that would otherwise be required to perform correct computation while maintaining 128-bit security. Using the CGGI variant of FHE and GPU acceleration, we quantize and evaluate a 1.9M parameter, multi-layer RNN across 28 timesteps, achieving 90.82% top-1 accuracy over the encrypted MNIST test dataset with an average latency of 2.1s per sample---a new state of the art in latency, model performance, and scale.", "keywords": [ "recurrent neural networks", "fully homomorphic encryption", "privacy-preserving machine learning", "quantization", "regularization", "overflow", "CGGI", "TFHE", "CKKS" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "TvhEoz1nim", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Fengfei Yu", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Undergrad student" }, { "name": "Mingyuan Zhou", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Xiong Peng", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "PhD student" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 39, "n_ref": 96, "n_ref_all": 128, "n_fig": 11, "n_tab": 12, "L_tab": 2355, "n_element_tab": 327, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1289, "n_element_tab_1": 67, "formula_len_all": 1116, "formula_len_all_1": 1172, "len_all": 165899, "len_all_1": 60055, "len_abs": 1777, "len_title": 117, "len_sents": 55930, "len_sents_1": 26979, "n_sents": 457, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1671, "title": "Single-Step Diffusion Model-Based Generative Model Inversion Attacks", "abs": "Generative model inversion attacks (MIAs) have garnered increasing attention for their ability to reconstruct synthetic samples that closely resemble private training data, exposing significant privacy risks in machine learning models. The success of generative MIAs is primarily attributed to image priors learned by generative adversarial networks (GANs) on public auxiliary data, which help constrain the optimization space during the inversion process. However, GAN-based generative MIAs still face limitations, particularly regarding the instability during model inversion optimization and the fidelity of reconstructed samples, indicating substantial room for improvement. In this paper, we address these challenges by exploring generative MIAs based on diffusion models, which offer superior generative performance compared to GANs. Specifically, we replace the GAN generator in existing generative MIAs with a single-step generator distilled from pretrained diffusion models, constraining the search space to the manifold of the generator during the inversion process. In addition, we leverage generative model inversion techniques to investigate privacy leakage issues in widely used large-scale multimodal models, particularly CLIP, highlighting the inherent privacy risks in these models. Our extensive experiments demonstrate that single-step diffusion models-based MIAs significantly outperform their GAN-based counterparts, achieving substantial improvements in traditional metrics and greatly enhancing the visual fidelity of reconstructed samples. This research uncovers vulnerabilities in CLIP models and opens new research directions in generative MIAs.", "keywords": [ "Diffusion models", "model inversion attacks" ], "rating_list": [ 3, 3, 6, 5, 5 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 5, 5, 3, 4, 4 ] }, { "paper_id": "TvfkSyHZRA", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Lucas Prieto", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Melih Barsbey", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Postdoc" }, { "name": "Pedro A. M. Mediano", "gender": "Not Specified", "institution": "Imperial College London", "country": "GB", "position": "Lecturer" }, { "name": "Tolga Birdal", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 17, "n_ref_uni": 35, "n_ref": 98, "n_ref_all": 152, "n_fig": 30, "n_tab": 1, "L_tab": 620, "n_element_tab": 57, "n_fig_1": 16, "n_tab_1": 1, "L_tab_1": 501, "n_element_tab_1": 7, "formula_len_all": 1368, "formula_len_all_1": 624, "len_all": 142296, "len_all_1": 64807, "len_abs": 1696, "len_title": 91, "len_sents": 48763, "len_sents_1": 30642, "n_sents": 367, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 1734, "title": "Grokking at the Edge of Numerical Stability", "abs": "Grokking, or sudden generalization that occurs after prolonged overfitting, is a surprising phenomenon that has challenged our understanding of deep learning. While a lot of progress has been made in understanding grokking, it is still not clear why generalization is delayed and why grokking often does not happen without regularization. In this work we argue that without regularization, grokking tasks push models to the edge of numerical stability, introducing floating point errors in the Softmax that we refer to as _Softmax Collapse_ (SC). We show that SC prevents grokking and that mitigating SC leads to grokking _without_ regularization. Investigating the root cause of SC, we find that beyond the point of overfitting, the gradients strongly align with what we call the _na\u00efve loss minimization_ (NLM) direction. This component of the gradient does not change the predictions of the model but decreases the loss by scaling the logits, usually through the scaling of the weights along their current direction. We show that this scaling of the logits explains the delay in generalization characteristic of grokking, and eventually leads to SC, stopping learning altogether. To validate these hypotheses, we introduce two key contributions that mitigate the issues faced in grokking tasks: (i) $\\mathrm{StableMax}$, a new activation function that prevents SC and enables grokking without regularization, and (ii) $\\perp\\mathrm{Grad}$, a training algorithm that leads to quick generalization in grokking tasks by preventing NLM altogether. These contributions provide new insights into grokking, shedding light on its delayed generalization, reliance on regularization, and the effectiveness of known grokking-inducing methods.", "keywords": [ "grokking", "deep learning", "learning theory", "floating point", "scientific computation", "generalization" ], "rating_list": [ 6, 8, 8, 5, 8 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "TvGPP8i18S", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Andrey Zhmoginov", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Aren Jansen", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "David Racz", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "DeLesley Hutchins", "gender": "unknown", "institution": "Google", "country": "", "position": "Senior Software Engineer" }, { "name": "Jesper Sparre Andersen", "gender": "Male", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Yinpeng Chen", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 38, "n_ref": 76, "n_ref_all": 111, "n_fig": 8, "n_tab": 16, "L_tab": 6150, "n_element_tab": 593, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 595, "n_element_tab_1": 47, "formula_len_all": 283, "formula_len_all_1": 284, "len_all": 205634, "len_all_1": 72158, "len_abs": 945, "len_title": 102, "len_sents": 59673, "len_sents_1": 33600, "n_sents": 429, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 949, "title": "MELODI: Exploring Memory Compression for Long Contexts", "abs": "We present MELODI, a novel memory architecture designed to efficiently process long documents using short context windows. The key principle behind MELODI is to represent short-term and long-term memory as a hierarchical compression scheme across both network layers and context windows. Specifically, the short-term memory is achieved through recurrent compression of context windows across multiple layers, ensuring smooth transitions between windows. In contrast, the long-term memory performs further compression within a single middle layer and aggregates information across context windows, effectively consolidating crucial information from the entire history. Compared to a strong baseline - the Memorizing Transformer employing dense attention over a large long-term memory (64K key-value pairs) - our method demonstrates superior performance on various long-context datasets while remarkably reducing the memory footprint by a factor of 8.", "keywords": [ "Memory", "Compression", "Long Context" ], "rating_list": [ 6, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 3, 4, 2 ] }, { "paper_id": "Tv36j85SqR", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Eric Lei", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Hamed Hassani", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Assistant Professor" }, { "name": "Shirin Saeedi Bidokhti", "gender": "unknown", "institution": "University of Pennsylvania", "country": "", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 29, "n_ref_uni": 41, "n_ref": 81, "n_ref_all": 116, "n_fig": 21, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 23260, "formula_len_all_1": 23506, "len_all": 177401, "len_all_1": 109087, "len_abs": 1395, "len_title": 134, "len_sents": 55699, "len_sents_1": 41547, "n_sents": 457, "n_sents_1": 333, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 86, "L_abs": 1399, "title": "Approaching Rate-Distortion Limits in Neural Compression with Lattice Transform Coding", "abs": "Neural compression has brought tremendous progress in designing lossy compressors with good rate-distortion (RD) performance at low complexity. Thus far, neural compression design involves transforming the source to a latent vector, which is then rounded to integers and entropy coded. While this approach has been shown to be optimal on a few specific sources, we show that it can be highly sub-optimal on synthetic sources whose intrinsic dimensionality is greater than one. With integer rounding in the latent space, the quantization regions induced by neural transformations, remain square-like and fail to match those of optimal vector quantization. We demonstrate that this phenomenon is due to the choice of scalar quantization in the latent space, and not the transform design. By employing lattice quantization instead, we propose Lattice Transform Coding (LTC) and show that it approximately recovers optimal vector quantization at reasonable complexity. On real-world sources, LTC improves upon standard neural compressors. LTC also provides a framework that can integrate structurally (near) optimal information-theoretic designs into lossy compression; examples include block coding, which yields coding gain over optimal one-shot coding and approaches the asymptotically-achievable rate-distortion function, as well as nested lattice quantization for low complexity fixed-rate coding.", "keywords": [ "Neural compression", "vector quantization", "lattice quantization", "nonlinear transform coding" ], "rating_list": [ 6, 8, 6, 8, 8 ], "soundness_list": [ 3, 4, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "TuOTSAiHDn", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bryan Catanzaro", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Vice President" }, { "name": "Eric Nyberg", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "John Kamalu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Mohammad Shoeybi", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Director of Applied Resesrch" }, { "name": "Mostofa Patwary", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Sanjeev Satheesh", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Shrimai Prabhumoye", "gender": "Female", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Syeda Nahida Akter", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 52, "n_ref": 115, "n_ref_all": 163, "n_fig": 9, "n_tab": 17, "L_tab": 5319, "n_element_tab": 846, "n_fig_1": 3, "n_tab_1": 9, "L_tab_1": 3104, "n_element_tab_1": 324, "formula_len_all": 181, "formula_len_all_1": 172, "len_all": 273767, "len_all_1": 70860, "len_abs": 1373, "len_title": 60, "len_sents": 80959, "len_sents_1": 32770, "n_sents": 637, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1379, "title": "MIND: Math Informed syNthetic Dialogues for Pretraining LLMs", "abs": "The utility of synthetic data to enhance pretraining data quality and hence to improve downstream task accuracy has been widely explored in recent large language models (LLMs). Yet, these approaches fall inadequate in complex, multi-hop and mathematical reasoning tasks as the synthetic data typically fails to add complementary knowledge to the existing raw corpus. In this work, we propose a novel large-scale and diverse Math Informed syNthetic Dialogue (MIND) generation method that improves the mathematical reasoning ability of LLMs. Specifically, using MIND, we generate synthetic conversations based on OpenWebMath (OWM), resulting in a new math corpus, MIND-OWM. Our experiments with different conversational settings reveal that incorporating knowledge gaps between dialog participants is essential for generating high-quality math data. We further identify an effective way to format and integrate synthetic and raw data during pretraining to maximize the gain in mathematical reasoning, emphasizing the need to restructure raw data rather than use it as-is. Compared to pretraining just on raw data, a model pretrained on MIND-OWM shows significant boost in mathematical reasoning (GSM8K: +13.42%, MATH: +2.30%), including superior performance in specialized knowledge (MMLU: +4.55%, MMLU-STEM: +4.28%) and general purpose reasoning tasks (GENERAL REASONING: +2.51%).", "keywords": [ "pretraining", "mathematical reasoning", "synthetic dialogue", "LLM", "reasoning" ], "rating_list": [ 5, 5, 8 ], "soundness_list": [ 3, 2, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "TtVKHxzHCy", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Kewei Tu", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "Full Professor" }, { "name": "Wei Wu", "gender": "Male", "institution": "Ant Research", "country": "CN", "position": "Researcher" }, { "name": "Zhihao Teng", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "Undergrad student" }, { "name": "Hu Xiang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 4, "n_ref_uni": 46, "n_ref": 71, "n_ref_all": 85, "n_fig": 4, "n_tab": 6, "L_tab": 3258, "n_element_tab": 364, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2832, "n_element_tab_1": 329, "formula_len_all": 579, "formula_len_all_1": 409, "len_all": 177906, "len_all_1": 69191, "len_abs": 993, "len_title": 124, "len_sents": 39480, "len_sents_1": 28238, "n_sents": 379, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1002, "title": "Efficient Long-range Language Modeling with Self-supervised Causal Retrieval", "abs": "Recently, retrieval-based language models (RLMs) have received much attention. However, most of them leverage a pre-trained retriever with fixed parameters, which may not adapt well to causal language models. In this work, we propose Grouped Cross-Attention, a novel module enabling joint pre-training of the retriever and causal LM, and apply it to long-context modeling. For a given input sequence, we split it into chunks and use the current chunk to retrieve past chunks for subsequent text generation. \nOur innovation allows the retriever to learn how to retrieve past chunks that better minimize the auto-regressive loss of subsequent tokens in an end-to-end manner.\nBy integrating top-$k$ retrieval, our model can be pre-trained efficiently from scratch with context lengths up to 64K tokens. \nOur experiments demonstrate that our model achieves superior performance in various tasks against strong baselines, and 100\\% accuracy in the needle-in-a-haystack (NIAH) test with a 16M context length.", "keywords": [ "long-range language modeling", "Retrieval-based LM", "self-supervised learning" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "TtUh0TOlGX", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Geon Yeong Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Hyungjin Chung", "gender": "Male", "institution": "EverEx", "country": "KR", "position": "Researcher" }, { "name": "Jeongsol Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Jong Chul Ye", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 35, "n_formula_1": 23, "n_ref_uni": 32, "n_ref": 70, "n_ref_all": 125, "n_fig": 18, "n_tab": 7, "L_tab": 1401, "n_element_tab": 251, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 982, "n_element_tab_1": 158, "formula_len_all": 2788, "formula_len_all_1": 1613, "len_all": 151279, "len_all_1": 63033, "len_abs": 370, "len_title": 108, "len_sents": 47430, "len_sents_1": 26296, "n_sents": 386, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 60, "L_abs": 880, "title": "Regularization by Texts for Latent Diffusion Inverse Solvers", "abs": "The recent development of diffusion models has led to significant progress in solving inverse problems by leveraging these models as powerful generative priors. However, challenges persist due to the ill-posed nature of such problems, often arising from ambiguities in measurements or intrinsic system symmetries. To address this, we introduce a novel latent diffusion inverse solver, regularization by text (TReg), inspired by the human ability to resolve visual ambiguities through perceptual biases. TReg integrates textual descriptions of preconceptions about the solution during reverse diffusion sampling, dynamically reinforcing these descriptions through null-text optimization, which we refer to as adaptive negation. Our comprehensive experimental results demonstrate that TReg effectively mitigates ambiguity in inverse problems, improving both accuracy and efficiency.", "keywords": [ "Inverse problem", "Text regularization", "Diffusion model" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "TtKN1TpvUu", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Hanqin Wang", "gender": "Male", "institution": "Tokyo University, Tokyo Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Nabarun Goswami", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Tatsuya Harada", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Team Leader" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 40, "n_ref": 72, "n_ref_all": 98, "n_fig": 1, "n_tab": 5, "L_tab": 1167, "n_element_tab": 89, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1167, "n_element_tab_1": 89, "formula_len_all": 599, "formula_len_all_1": 598, "len_all": 134245, "len_all_1": 63943, "len_abs": 1169, "len_title": 148, "len_sents": 32156, "len_sents_1": 28186, "n_sents": 218, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1189, "title": "T2V2: A Unified Non-Autoregressive Model for Speech Recognition and Synthesis via Multitask Learning", "abs": "We introduce T2V2 (**T**ext to **V**oice and **V**oice to **T**ext), a unified non-autoregressive model capable of performing both automatic speech recognition (ASR) and text-to-speech (TTS) synthesis within the same framework. T2V2 uses a shared Conformer backbone with rotary positional embeddings to efficiently handle these core tasks, with ASR trained using Connectionist Temporal Classification (CTC) loss and TTS using masked language modeling (MLM) loss. The model operates on discrete tokens, where speech tokens are generated by clustering features from a self-supervised learning model. To further enhance performance, we introduce auxiliary tasks: CTC error correction to refine raw ASR outputs using contextual information from speech embeddings, and unconditional speech MLM, enabling classifier free guidance to improve TTS. Our method is self-contained, leveraging intermediate CTC outputs to align text and speech using Monotonic Alignment Search, without relying on external aligners. We perform extensive experimental evaluation to verify the efficacy of the T2V2 framework, achieving state-of-the-art performance on TTS task and competitive performance in discrete ASR.", "keywords": [ "ASR", "TTS", "Non-Autoregressive", "Conformer", "Multitask Learning" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "TswLvrIY8M", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Helio Pedrini", "gender": "Male", "institution": "Universidade Estadual de Campinas", "country": "BR", "position": "Full Professor" }, { "name": "Adin Ramirez Rivera", "gender": "Male", "institution": "University of Oslo", "country": "NO", "position": "Full Professor" }, { "name": "Thalles Santos Silva", "gender": "Male", "institution": "Universidade Estadual de Campinas", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 20, "n_ref": 44, "n_ref_all": 79, "n_fig": 3, "n_tab": 13, "L_tab": 2707, "n_element_tab": 353, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 901, "n_element_tab_1": 118, "formula_len_all": 449, "formula_len_all_1": 184, "len_all": 134364, "len_all_1": 65553, "len_abs": 1207, "len_title": 122, "len_sents": 45058, "len_sents_1": 30545, "n_sents": 341, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1279, "title": "Self-Organizing Visual Embeddings for Non-Parametric Self-Supervised Learning", "abs": "We present Self-Organizing Visual Embeddings (SOVE) a new training technique for unsupervised representation learning.\nSOVE avoids learning prototypes from scratch and explores relationships between visual embeddings in a non-parametric space.\nUnlike existing clustering-based techniques that employ a single prototype to encode all the relevant features of a complex concept, we propose the SOVE method where a concept is represented by many semantically similar representations, or judges, each containing a complement set of features that together can fully characterize the concept and maximize training performance.\nWe reaffirm the feasibility of non-parametric self-supervised learning (SSL) by introducing novel non-parametric adaptions of two loss functions with the SOVE technique: (1) non-parametric cluster assignment prediction for class-level representations and (2) non-parametric Masked Image Modeling (MIM) for patch-level reconstruction.\nSOVE achieves state-of-the-art performance on many downstream benchmarks, including transfer learning, image retrieval, object detection, and segmentation.\nMoreover, SOVE demonstrates scaling performance when trained with Vision Transformers (ViTs), showing increased performance gains as more complex encoders are employed.", "keywords": [ "self-supervised learning", "clustering", "representation learning", "computer vision" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 1, 1, 1 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "TspmMMOG7X", "primary_area": "foundation or frontier models, including LLMs", "n_author": 30, "author_info_list": [ { "name": "Chunwei Wang", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Daxin Tan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Dingdong WANG", "gender": "Female", "institution": "Chinese University of Hong Kong, The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Dit-Yan Yeung", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Chair Professor" }, { "name": "Hang Xu", "gender": "Male", "institution": "Huawei Noah\u2018s Ark Lab", "country": "HK", "position": "Researcher" }, { "name": "Haoli Bai", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Haoyuan Li", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Hengshuang Zhao", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "James Kwok", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Hong Kong University of Science and Technology", "country": "HK", "position": "Professor" }, { "name": "Jianhua Han", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Jing Xu", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Kai Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Kun Xiang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Kuo Yang", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Lanqing HONG", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Nian Xie", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Qun Liu", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "HK", "position": "Chief Scientist of Speech and Language Computing" }, { "name": "Runhui Huang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Wei Zhang", "gender": "Male", "institution": "Huawei Technologies Ltd", "country": "CN", "position": "Researcher" }, { "name": "Weike Jin", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xiao-Hui Li", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xiaodan Liang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Yi Zhu", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yunhao GOU", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zhenguo Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "LU HOU", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "researcher" }, { "name": "XIAO CHEN", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Principal Researcher" }, { "name": "Zeng Yihan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zhili LIU", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 6, "n_ref_uni": 75, "n_ref": 142, "n_ref_all": 183, "n_fig": 22, "n_tab": 4, "L_tab": 1225, "n_element_tab": 140, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 716, "n_element_tab_1": 68, "formula_len_all": 1117, "formula_len_all_1": 558, "len_all": 231319, "len_all_1": 60384, "len_abs": 1314, "len_title": 125, "len_sents": 57549, "len_sents_1": 28694, "n_sents": 435, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1243, "title": "EMOVA: Empowering Language Models to See, Hear and Speak with Vivid Emotions", "abs": "GPT-4o, an omni-modal model that enables vocal conversations with diverse emotions and tones, marks a milestone for omni-modal foundation models. However, empowering Large Language Models to perceive and generate images, texts, and speeches end-to-end with publicly available data remains challenging in the open-source community. Existing vision-language models rely on external tools for the speech processing, while speech-language models still suffer from limited or even without vision-understanding abilities. To address this gap, we propose EMOVA (EM-otionally Omni-present Voice Assistant), to enable Large Language Models with end-to-end speech capabilities while maintaining the leading vision-language performance. With a semantic-acoustic disentangled speech tokenizer, we notice surprisingly that omni-modal alignment can further enhance vision-language and speech abilities compared with the corresponding bi-modal aligned counterparts. Moreover, a lightweight style module is proposed for flexible speech style controls (e.g., emotions and pitches). For the first time, EMOVA achieves state-of-the-art performance on both the vision-language and speech benchmarks, and meanwhile, supporting omni-modal emotional spoken dialogue.", "keywords": [ "LLM", "Multi-Modal", "Vision-Language", "Speech-Language" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 4, 2, 4, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "TsBDfe8Ra5", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Lei Zhang", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Chair Professor" }, { "name": "Ruihuang Li", "gender": "Male", "institution": "The Hong Kong Polytechnic University, Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Zhengqiang ZHANG", "gender": "Male", "institution": "The Hong Kong Polytechnic University, Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 28, "n_ref": 57, "n_ref_all": 70, "n_fig": 6, "n_tab": 1, "L_tab": 1841, "n_element_tab": 136, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 1841, "n_element_tab_1": 136, "formula_len_all": 313, "formula_len_all_1": 313, "len_all": 102210, "len_all_1": 56072, "len_abs": 3443, "len_title": 138, "len_sents": 31351, "len_sents_1": 24692, "n_sents": 242, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1698, "title": "FreCaS: Efficient Higher-Resolution Image Generation via Frequency-aware Cascaded Sampling", "abs": "While image generation with diffusion models has achieved a great success, generating images of higher resolution than the training size remains a challenging task due to the high computational cost. Current methods typically perform the entire sampling process at full resolution and process all frequency components simultaneously, contradicting with the inherent coarse-to-fine nature of latent diffusion models and wasting computations on processing premature high-frequency details at early diffusion stages. To address this issue, we introduce an efficient $\\textbf{Fre}$quency-aware $\\textbf{Ca}$scaded $\\textbf{S}$ampling framework, $\\textbf{FreCaS}$ in short, for higher-resolution image generation. FreCaS decomposes the sampling process into cascaded stages with gradually increased resolutions, progressively expanding frequency bands and refining the corresponding details. We propose an innovative frequency-aware classifier-free guidance (FA-CFG) strategy to assign different guidance strengths for different frequency components, directing the diffusion model to add new details in the expanded frequency domain of each stage. Additionally, we fuse the cross-attention maps of previous and current stages to avoid synthesizing unfaithful layouts. Experiments demonstrate that FreCaS significantly outperforms state-of-the-art methods in image quality and generation speed. In particular, FreCaS is about 2.86$\\times$ and 6.07$\\times$ faster than ScaleCrafter and DemoFusion in generating a 2048$\\times$2048 image using a pretrained SDXL model and achieves an $\\text{FID}_b$ improvement of 11.6 and 3.7, respectively. FreCaS can be easily extended to more complex models such as SD3.", "keywords": [ "generative models", "diffusion models", "training-free" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "Ts1waOOQjF", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "BingQuan Dai", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "DongbinZhang", "gender": "Male", "institution": "Tsinghua Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Haoqian Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Kangjie Chen", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Minghan Qin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Peihao Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 61, "n_ref_all": 84, "n_fig": 10, "n_tab": 7, "L_tab": 1054, "n_element_tab": 140, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 579, "n_element_tab_1": 85, "formula_len_all": 397, "formula_len_all_1": 400, "len_all": 117954, "len_all_1": 52413, "len_abs": 109, "len_title": 121, "len_sents": 35944, "len_sents_1": 23928, "n_sents": 252, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1124, "title": "SparseLGS: Fast Language Gaussian Splatting from Sparse Multi-View Images", "abs": "3D semantic field learning is crucial for applications like autonomous navigation, AR/VR, and robotics, where accurate comprehension of 3D scenes from limited viewpoints is essential. Existing methods struggle under sparse view conditions, relying on inefficient per-scene multi-view optimizations, which are impractical for many real-world tasks. To address this, we propose SparseLGS, a feed-forward method for constructing 3D semantic fields from sparse viewpoints, allowing direct inference of 3DGS-based scenes. By ensuring consistent SAM segmentations through video tracking and using low-dimensional indexing for high-dimensional CLIP features, SparseLGS efficiently embeds language information in 3D space, offering a robust solution for accurate 3D scene understanding under sparse view conditions. In experiments on two-view sparse 3D object querying and segmentation in the LERF and 3D-OVS datasets, SparseLGS outperforms existing methods in chosen IoU, Localization Accuracy, and mIoU. Moreover, our model achieves scene inference in under 30 seconds and open-vocabulary querying in just 0.011 seconds per query.", "keywords": [ "3D Gaussian Splatting; 3D open-vocabulary query; 3D segmentation; sparse view reconstruction;" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "TroV1cbgoG", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Akiko Takeda", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Professor" }, { "name": "Andi Han", "gender": "Male", "institution": "RIKEN AIP", "country": "JP", "position": "Postdoc" }, { "name": "Gang Niu", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Senior Research Scientist (tenured)" }, { "name": "Taiji Suzuki", "gender": "Male", "institution": "The University of Tokyo", "country": "", "position": "Full Professor" }, { "name": "Wei Huang", "gender": "Male", "institution": "RIKEN AIP", "country": "JP", "position": "Research Scientist" }, { "name": "Wuyang Chen", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "Assistant Professor" }, { "name": "Zhanpeng Zhou", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 242, "n_formula_1": 37, "n_ref_uni": 40, "n_ref": 123, "n_ref_all": 164, "n_fig": 2, "n_tab": 4, "L_tab": 777, "n_element_tab": 51, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 477, "n_element_tab_1": 20, "formula_len_all": 30353, "formula_len_all_1": 3717, "len_all": 267307, "len_all_1": 75100, "len_abs": 1451, "len_title": 104, "len_sents": 64357, "len_sents_1": 29625, "n_sents": 767, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1467, "title": "The Role of Label Noise in the Feature Learning Process", "abs": "Deep learning with noisy labels presents significant challenges.\nIn this work, we theoretically characterize the role of label noise in training neural networks from a feature learning perspective.\nSpecifically, we consider a *signal-noise* data distribution, where each data point comprises a label-dependent signal and label-independent noise, and rigorously analyze the training dynamics of a two-layer convolutional neural network under this data setting, along with the presence of label noise.\nParticularly, we identify two stages in which the dynamics exhibit distinct patterns.\nIn *Stage I*, the model perfectly fits all the clean samples (i.e., samples without label noise) while ignoring the noisy ones (i.e., samples with noisy labels).\nIn the first stage, the model learns the signal from the clean samples, which generalizes well on unseen data.\nIn *Stage II*, as the training loss converges, the gradient in the direction of noise surpasses that of the signal, leading to over-fitting on noisy samples.\nEventually, the model memorizes the noise present in the noisy samples, which degrades its generalization ability.\nIn contrast, when training without label noise, the dynamics do not exhibit this two-stage pattern.\nFurthermore, our results provide theoretical supports for two widely used techniques for tackling label noise: early stopping and sample selection.\nExperiments on both synthetic and real-world datasets confirm our theoretical findings.", "keywords": [ "Label noise", "Feature Learning", "Training Dynamics" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 2, 2 ] }, { "paper_id": "Trn4Hji6iH", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "He Wang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jinhui Tang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Longquan Dai", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 18, "n_formula_1": 17, "n_ref_uni": 25, "n_ref": 63, "n_ref_all": 86, "n_fig": 7, "n_tab": 2, "L_tab": 830, "n_element_tab": 65, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 157, "n_element_tab_1": 16, "formula_len_all": 1936, "formula_len_all_1": 1679, "len_all": 113774, "len_all_1": 58501, "len_abs": 1357, "len_title": 77, "len_sents": 33435, "len_sents_1": 25725, "n_sents": 282, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1461, "title": "AccCtr: Accelerating Training-Free Control For Text-to-Image Diffusion Models", "abs": "In training-free Conditional Diffusion Models (CDMs), the sampling process is steered by the gradient of the loss $\\mathcal{E}(\\dmrv{y}, \\dmrv{z}, \\dmfv{C}_{\\dmv{psi}} )$, which assesses the gap between the guidance $\\dmrv{y}$ and the condition extracted from the intermediate outputs. Here the condition extraction network $\\dmfv{C}_{\\dmv{psi}}(\\cdot)$, which could be a segmentation or depth estimation network, is pre-trained for training-free purpose. However, existing methods often require small guidance steps, leading to longer sampling times. We introduce an alternative maximization framework to scrutinize training-free CDMs that tackles slow sampling. Our framework pinpoints manifold deviation as the key factor behind the sluggish sampling. More iterations are needed for the sampling process to closely follow the image manifold and reach the target conditions, as the loss gradient doesn't provide sufficient guidance for larger steps. To improve this, we suggest retraining the condition extraction network $\\dmfv{C}_{\\dmv{psi}}(\\cdot)$ to refine the loss's guidance, thereby introducing our AccCtr. This retraining process is simple, and integrating AccCtr into current CDMs is a seamless task that does not impose a significant computational burden. Extensive testing has demonstrated that AccCtr significantly boosts performance, offering superior sample quality and faster generation times across a variety of conditional generation tasks.", "keywords": [ "accelerating", "Training-Free", "diffusion model" ], "rating_list": [ 3, 5, 5, 1 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "Trf0R8eoGF", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Kai Han", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Liang Pan", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Research Scientist " }, { "name": "Ziwei Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Kwan-Yee Kenneth Wong", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Yukang CAO", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" } ], "n_formula": 22, "n_formula_1": 22, "n_ref_uni": 75, "n_ref": 149, "n_ref_all": 169, "n_fig": 12, "n_tab": 2, "L_tab": 941, "n_element_tab": 49, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 685, "n_element_tab_1": 62, "formula_len_all": 1119, "formula_len_all_1": 1118, "len_all": 194336, "len_all_1": 53760, "len_abs": 1658, "len_title": 121, "len_sents": 39918, "len_sents_1": 24171, "n_sents": 287, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1695, "title": "AvatarGO: Zero-shot 4D Human-Object Interaction Generation and Animation", "abs": "Recent advancements in diffusion models have led to significant improvements in the generation and animation of 4D full-body human-object interactions (HOI). Nevertheless, existing methods primarily focus on SMPL-based motion generation, which is limited by the scarcity of realistic large-scale interaction data. This constraint affects their ability to create everyday HOI scenes. This paper addresses this challenge using a zero-shot approach with a pre-trained diffusion model. Despite this potential, achieving our goals is difficult due to the diffusion model's lack of understanding of ''where'' and ''how'' objects interact with the human body. To tackle these issues, we introduce **AvatarGO**, a novel framework designed to generate animatable 4D HOI scenes directly from textual inputs. Specifically, **1)** for the ''where'' challenge, we propose **LLM-guided contact retargeting**, which employs Lang-SAM to identify the contact body part from text prompts, ensuring precise representation of human-object spatial relations. **2)** For the ''how'' challenge, we introduce **correspondence-aware motion optimization** that constructs motion fields for both human and object models using the linear blend skinning function from SMPL-X. Our framework not only generates coherent compositional motions, but also exhibits greater robustness in handling penetration issues. Extensive experiments with existing methods validate AvatarGO's superior generation and animation capabilities on a variety of human-object pairs and diverse poses. As the first attempt to synthesize 4D avatars with object interactions, we hope AvatarGO could open new doors for human-centric 4D content creation.", "keywords": [ "4D human avatar generation", "compositional generation", "human-object interaction" ], "rating_list": [ 5, 8, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "TrVYEZtSQH", "primary_area": "datasets and benchmarks", "n_author": 21, "author_info_list": [ { "name": "Chunyuan Li", "gender": "unknown", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Dan Roth", "gender": "Male", "institution": "Oracle", "country": "US", "position": "Chief Scientist" }, { "name": "Fei Wang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Hoifung Poon", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "General Manager" }, { "name": "Hsiang-Hui Liu", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "Undergrad student" }, { "name": "James Y. Huang", "gender": "unknown", "institution": "University of Southern California", "country": "", "position": "PhD student" }, { "name": "Kai-Wei Chang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Associate Professor" }, { "name": "Kai Zhang", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Mingyu Derek Ma", "gender": "Male", "institution": "Genentech", "country": "US", "position": "Machine Learning Scientist" }, { "name": "Muhao Chen", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Assistant Professor" }, { "name": "Nan Xu", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Qin Liu", "gender": "Female", "institution": "University of California, Davis", "country": "US", "position": "PhD student" }, { "name": "Sheng Zhang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Wenxuan Zhou", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Researcher" }, { "name": "Xiaogeng Liu", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Xingyu Fu", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Zekun Li", "gender": "unknown", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" }, { "name": "chaowei xiao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Pan Lu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Tianyi Yan", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "Researcher" }, { "name": "Wenjie Mo", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 105, "n_ref_all": 137, "n_fig": 15, "n_tab": 4, "L_tab": 5166, "n_element_tab": 717, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 3227, "n_element_tab_1": 470, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 187045, "len_all_1": 64735, "len_abs": 308, "len_title": 121, "len_sents": 38517, "len_sents_1": 26478, "n_sents": 324, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1099, "title": "MuirBench: A Comprehensive Benchmark for Robust Multi-image Understanding", "abs": "We introduce MuirBench, a comprehensive benchmark that focuses on robust multi-image understanding capabilities of multimodal LLMs. MuirBench consists of 12 diverse multi-image tasks (e.g., scene understanding, ordering) that involve 10 categories of multi-image relations (e.g., multiview, temporal relations). Comprising 11,264 images and 2,600 multiple-choice questions, MuirBench is created in a pairwise manner, where each standard instance is paired with an unanswerable variant that has minimal semantic differences, in order for a reliable assessment. Evaluated upon 20 recent multi-modal LLMs, our results reveal that even the best-performing models like GPT-4o and Gemini Pro find it challenging to solve MuirBench, achieving 68.0% and 49.3% in accuracy. Open-source multimodal LLMs trained on single images can hardly generalize to multi-image questions, hovering below 33.3% in accuracy. These results highlight the importance of MuirBench in encouraging the community to develop multimodal LLMs that can look beyond a single image, suggesting potential pathways for future improvements.", "keywords": [ "Multimodal LLM", "Multi-image Understanding", "Benchmark", "Robustness" ], "rating_list": [ 3, 5, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4, 3 ], "contribution_list": [ 1, 2, 3, 3, 3 ], "confidence_list": [ 5, 5, 5, 3, 2 ] }, { "paper_id": "TrKRpaOk8y", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Jiawei Han", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign (UIUC)", "country": "US", "position": "Full Professor" }, { "name": "Suyu Ge", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Xihui Lin", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Yunan Zhang", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" }, { "name": "Hao Peng", "gender": "Male", "institution": "Department of Computer Science, University of Illinois Urbana-Champaign", "country": "", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 34, "n_ref": 82, "n_ref_all": 108, "n_fig": 3, "n_tab": 3, "L_tab": 487, "n_element_tab": 56, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 753, "n_element_tab_1": 45, "formula_len_all": 233, "formula_len_all_1": 155, "len_all": 136526, "len_all_1": 62569, "len_abs": 2123, "len_title": 141, "len_sents": 32419, "len_sents_1": 30929, "n_sents": 238, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 2142, "title": "A Little Goes a Long Way: Efficient Long Context Training and Inference with Partial Contexts", "abs": "Training and serving long-context large language models (LLMs) incurs substantial overhead. \nTo address this, two critical steps are often required: a pretrained LLM typically undergoes a separate stage for context length extension by training on long-context data, followed by architectural modifications to reduce the overhead of KV cache during serving. \nThis paper argues that integrating length extension with a GPU-friendly KV cache reduction architecture not only reduces training overhead during length extension, but also achieves better long-context performance. \nThis leads to our proposed LongGen, which finetunes a pretrained LLM into an efficient architecture during length extension. \nLongGen builds on three key insights: \n(1) Sparse attention patterns, such as window attention (attending to recent tokens), attention sink (initial ones), and blockwise sparse attention (strided token blocks) are well-suited for building efficient long-context models, primarily due to their GPU-friendly memory access patterns, enabling efficiency gains not just theoretically but in practice as well. \n(2) It is essential for the model to have direct access to all tokens. \nA hybrid architecture with 1/3 full attention layers and 2/3 efficient ones achieves a balanced trade-off between efficiency and long-context performance.\n(3) Lightweight training on 5B long-context data is sufficient to extend the hybrid model's context length from 4K to 128K.\n\nWe evaluate LongGen on both Llama-2 7B and Llama-2 70B, demonstrating its effectiveness across different scales. \nDuring training with 128K-long contexts, LongGen achieves 1.55x training speedup and reduces wall-clock time by 36%, compared to a full-attention baseline. \nDuring inference, LongGen reduces KV cache memory by 62%, achieving 1.67x prefilling speedup and 1.41x decoding speedup.\nCompared to baselines that apply KV-cache reduction techniques to full-attention long-context LLMs, LongGen achieves substantially stronger performance not only on the Needle-in-a-Haystack retrieval task, but also on more challenging long-context reasoning tasks, including BABILong and RULER.", "keywords": [ "Long-Context LLM", "Efficient LLM", "Context Extension", "KV Cache Reduction" ], "rating_list": [ 6, 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "TrJ36UfD9P", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Eduard Gorbunov", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "Grigory Malinovsky", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Petr Ostroukhov", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "Samuel Horv\u00e1th", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Assistant Professor" }, { "name": "Yury Demidovich", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Postdoc" }, { "name": "Martin Takac", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Associate Professor" }, { "name": "Peter Richtarik", "gender": "Male", "institution": "King Abdullah University of Science and Technology (KAUST)", "country": "SA", "position": "Full Professor" } ], "n_formula": 201, "n_formula_1": 24, "n_ref_uni": 50, "n_ref": 127, "n_ref_all": 152, "n_fig": 10, "n_tab": 5, "L_tab": 4096, "n_element_tab": 183, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 359, "n_element_tab_1": 58, "formula_len_all": 35251, "formula_len_all_1": 3096, "len_all": 262134, "len_all_1": 65798, "len_abs": 1062, "len_title": 150, "len_sents": 53547, "len_sents_1": 25370, "n_sents": 647, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 1065, "title": "Methods with Local Steps and Random Reshuffling for Generally Smooth Non-Convex Federated Optimization", "abs": "Non-convex Machine Learning problems typically do not adhere to the standard smoothness assumption. Based on empirical findings, Zhang et al. (2020b) proposed a more realistic generalized $(L_0,L_1)$-smoothness assumption, though it remains largely unexplored. Many existing algorithms designed for standard smooth problems need to be revised. However, in the context of Federated Learning, only a few works address this problem but rely on additional limiting assumptions. In this paper, we address this gap in the literature: we propose and analyze new methods with local steps, partial participation of clients, and Random Reshuffling without extra restrictive assumptions beyond generalized smoothness. The proposed methods are based on the proper interplay between clients' and server's stepsizes and gradient clipping. Furthermore, we perform the first analysis of these methods under the Polyak-\u0141ojasiewicz condition. Our theory is consistent with the known results for standard smooth problems, and our experimental results support the theoretical insights.", "keywords": [ "Optimization", "Federated Learning", "Distributed Optimization", "Local Training", "Random Reshuffling", "Generalized Smoothness" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "TqzNI4v9DT", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Hao Sun", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Kai Ruan", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Mengzhen Liu", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yilong Xu", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 1, "n_ref_uni": 39, "n_ref": 86, "n_ref_all": 98, "n_fig": 3, "n_tab": 7, "L_tab": 1588, "n_element_tab": 297, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 908, "n_element_tab_1": 53, "formula_len_all": 4784, "formula_len_all_1": 84, "len_all": 142746, "len_all_1": 64524, "len_abs": 1520, "len_title": 124, "len_sents": 44580, "len_sents_1": 32546, "n_sents": 331, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1294, "title": "GeoBench: A new benchmark on Symbolic Regression with Geometric Expressions", "abs": "Symbolic regression (SR) is a powerful technique for deriving mathematical expressions from data. With the emergence of numerous SR methods, SRBench made a significant contribution by providing a standardized testing platform that includes 130 SR datasets and evaluates 14 SR methods. However, the methods included in SRBench are outdated, and the dataset does not feature results from more recent approaches such as TPSR. Additionally, the metrics used in SRBench do not adequately capture the full capabilities of symbolic regression methods, and the benchmark data has scientific problems. Although Matsubara et al. (2022) address some of these issues,\ntheir approach remains incomplete. In response, we propose a new benchmark consisting of 71 expressions derived from geometric contexts, categorized into three difficulty levels: easy, medium, and hard. We evaluate 20 SR methods on these expressions, focusing exclusively on the symbolic regression capability of each model, assessed through recovery rates across the different levels and overall. We provide a detailed methodology for reproducing the experiments and include results for newly developed SR methods within this updated benchmark. The results demonstrate significant variation in symbolic regression ability across models.", "keywords": [ "Symbolic Regression", "Geometry" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 2, 1, 1, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "TqwTzLjzGS", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Andy Liu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Jiarui Liu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Maarten Sap", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Mona T. Diab", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Wenkai Li", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Xuhui Zhou", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 84, "n_ref": 168, "n_ref_all": 210, "n_fig": 3, "n_tab": 11, "L_tab": 4350, "n_element_tab": 756, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 204, "n_element_tab_1": 17, "formula_len_all": 1503, "formula_len_all_1": 42, "len_all": 304127, "len_all_1": 61076, "len_abs": 1344, "len_title": 124, "len_sents": 70533, "len_sents_1": 31581, "n_sents": 451, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1260, "title": "BIG5-CHAT: Shaping LLM Personalities Through Training on Human-Grounded Data", "abs": "In this work, we tackle the challenge of embedding realistic human personality traits into LLMs. Previous approaches have primarily focused on prompt-based methods that describe the behavior associated with the desired personality traits, suffering from realism and validity issues. To address these limitations, we introduce BIG5-CHAT, a large-scale dataset containing 100,000 dialogues designed to ground models in how humans express their personality in text. Leveraging this dataset, we explore Supervised Fine-Tuning and Direct Preference Optimization as training-based methods to align LLMs more naturally with human personality patterns. Our methods outperform prompting on personality assessments such as BFI and IPIP-NEO, with trait correlations more closely matching human data. Furthermore, our experiments reveal that models trained to exhibit higher conscientiousness, higher agreeableness, lower extraversion, and lower neuroticism display better performance on reasoning tasks, aligning with psychological findings on how these traits impact human cognitive performance. To our knowledge, this work is the first comprehensive study to demonstrate how training-based methods can shape LLM personalities through learning from real human behaviors.", "keywords": [ "Inducing Realistic Personality to LLMs", "LLM Personality Evaluation" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 5, 3, 4 ] }, { "paper_id": "Tqdsruwyac", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Carola-Bibiane Sch\u00f6nlieb", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "Full Professor" }, { "name": "Daniel G. Chen", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "MS student" }, { "name": "Krzysztof Polanski", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "Senior Computational Biologist" }, { "name": "Moshe Eliasof", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "Sarah A Teichmann", "gender": "Female", "institution": "Wellcome Sanger Institute", "country": "", "position": "Principal Researcher" }, { "name": "Stathis Megas", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" } ], "n_formula": 58, "n_formula_1": 9, "n_ref_uni": 49, "n_ref": 72, "n_ref_all": 96, "n_fig": 21, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 20, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4151, "formula_len_all_1": 697, "len_all": 185978, "len_all_1": 64175, "len_abs": 1192, "len_title": 168, "len_sents": 57235, "len_sents_1": 32726, "n_sents": 377, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 118, "L_abs": 1200, "title": "Estimation of single-cell and tissue perturbation effect in spatial transcriptomics via Spatial Causal Disentanglement", "abs": "Models of Virtual Cells and Virtual Tissues at single-cell resolution would allow us to test perturbations in silico and accelerate progress in tissue and cell engineering. \nHowever, most such models are not rooted in causal inference and as a result, could mistake correlation for causation.\nWe introduce Celcomen, a novel generative graph neural network grounded in mathematical causality to disentangle intra- and inter-cellular gene regulation in spatial transcriptomics and single-cell data. \nCelcomen can also be prompted by perturbations to generate spatial counterfactuals, thus offering insights into experimentally inaccessible states, with potential applications in human health. \nWe validate the model's disentanglement and identifiability through simulations, and demonstrate its counterfactual predictions in clinically relevant settings, including human glioblastoma and fetal spleen, recovering inflammation-related gene programs post immune system perturbation. \nMoreover, it supports mechanistic interpretability, as its parameters can be reverse-engineered from observed behavior, making it an accessible model for understanding both neural networks and complex biological systems.", "keywords": [ "spatial causal inference", "spatially disentangled representations", "spatial transcriptomics", "mechanistic interpretability" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "TqYjhJrp9m", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Yuanzhao Zhang", "gender": "Male", "institution": "Santa Fe Institute", "country": "US", "position": "Postdoc" }, { "name": "William C Gilpin", "gender": "unknown", "institution": "University of Texas at Austin", "country": "", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 0, "n_ref_uni": 97, "n_ref": 164, "n_ref_all": 185, "n_fig": 13, "n_tab": 1, "L_tab": 732, "n_element_tab": 1, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 61, "n_element_tab_1": 10, "formula_len_all": 325, "formula_len_all_1": 0, "len_all": 217822, "len_all_1": 68326, "len_abs": 1700, "len_title": 88, "len_sents": 58219, "len_sents_1": 35045, "n_sents": 367, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1437, "title": "Zero-shot forecasting of chaotic systems", "abs": "Time-series forecasting is a challenging problem that traditionally requires specialized models custom-trained for the specific task at hand. Recently, inspired by the success of large language models, foundation models pre-trained on vast amounts of time-series data from diverse domains have emerged as a promising candidate for general-purpose time-series forecasting. The defining characteristic of these foundation models is their ability to perform zero-shot learning, that is, forecasting a new system from limited context data without explicit re-training or fine-tuning. Here, we evaluate whether the zero-shot learning paradigm extends to the challenging task of forecasting chaotic systems. Across 135 distinct chaotic dynamical systems and $10^8$ timepoints, we find that foundation models produce competitive forecasts compared to custom-trained models (including NBEATS, TiDE, etc.), particularly when training data is limited. Interestingly, even after point forecasts fail, large foundation models are able to preserve the geometric and statistical properties of the chaotic attractors.\nWe attribute this success to foundation models' ability to perform in-context learning and identify context parroting as a simple mechanism used by these models to capture the long-term behavior of chaotic dynamical systems. Our results highlight the potential of foundation models as a tool for probing nonlinear and complex systems.", "keywords": [ "chaos", "nonlinear dynamics", "forecasting", "physics", "scientific machine learning" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "TqM0hifngW", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Hao Luo", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Intern" }, { "name": "Zongqing Lu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 51, "n_ref": 72, "n_ref_all": 84, "n_fig": 3, "n_tab": 2, "L_tab": 778, "n_element_tab": 146, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 23, "n_element_tab_1": 2, "formula_len_all": 713, "formula_len_all_1": 585, "len_all": 145608, "len_all_1": 60878, "len_abs": 1356, "len_title": 145, "len_sents": 35636, "len_sents_1": 30762, "n_sents": 265, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1364, "title": "Learning Video-Conditioned Policy on Unlabelled Data with Joint Embedding Predictive Transformer", "abs": "The video-conditioned policy takes prompt videos of the desired tasks as a condition and is regarded for its prospective generalizability. Despite its promise, training a video-conditioned policy is non-trivial due to the need for abundant demonstrations. In some tasks, the expert rollouts are merely available as videos, and costly and time-consuming efforts are required to annotate action labels. To address this, we explore training video-conditioned policy on a mixture of demonstrations and unlabeled expert videos to reduce reliance on extensive manual annotation. We introduce the Joint Embedding Predictive Transformer (JEPT) to learn a video-conditioned policy through sequence modeling. JEPT is designed to jointly learn visual transition prediction and inverse dynamics. The visual transition is captured from both demonstrations and expert videos, on the basis of which the inverse dynamics learned from demonstrations is generalizable to the tasks without action labels. Experiments on a series of simulated visual control tasks evaluate that JEPT can effectively leverage the mixture dataset to learn a generalizable policy. JEPT outperforms baselines in the tasks without action-labeled data and unseen tasks. We also experimentally reveal the potential of JEPT as a simple visual priors injection approach to enhance the video-conditioned policy.", "keywords": [ "Learning from Videos", "Video-Conditioned Policy" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "TqLY7QoELU", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Meng Fang", "gender": "Male", "institution": "University of Liverpool", "country": "", "position": "Assistant Professor" }, { "name": "Mykola Pechenizkiy", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Full Professor" }, { "name": "Wenhan Han", "gender": "unknown", "institution": "Eindhoven University of Technology", "country": "", "position": "PhD student" }, { "name": "Yaohang Li", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Undergrad student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 59, "n_ref_all": 70, "n_fig": 6, "n_tab": 5, "L_tab": 1756, "n_element_tab": 321, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1549, "n_element_tab_1": 284, "formula_len_all": 200, "formula_len_all_1": 200, "len_all": 129904, "len_all_1": 67098, "len_abs": 1147, "len_title": 120, "len_sents": 32755, "len_sents_1": 29354, "n_sents": 263, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1064, "title": "GasketRAG: Systematic Alignment of Large Language Models with Retrievers", "abs": "Retrieval-Augmented Generation (RAG) has emerged as a powerful method for enhancing the output quality of large language models (LLMs). However, existing retrievers are not specifically optimized for LLMs, and retraining them requires substantial resources. Furthermore, current approaches are often constrained to either improving the relevancy of retrieved documents or refining the documents post-retrieval. Various stages within the typical RAG pipeline present challenges in aligning LLMs with retrievers. To address these issues, we propose GasketRAG, a novel approach that introduces a gasket between the retriever and the LLM to improve their collaborative performance. By employing innovative techniques, we gather high-quality preference data and use the gasket to optimize both retrieval ranking and document refinement simultaneously. Our approach circumvents the need for constructing complex training and inference pipelines. In a fair comparison against the latest RAG methods across multiple test datasets, GasketRAG demonstrated a clear advantage.", "keywords": [ "Retrieval Augmented Generation", "Large Language Models" ], "rating_list": [ 5, 6, 6, 8, 5 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 4, 2 ], "confidence_list": [ 3, 3, 3, 4, 4 ] }, { "paper_id": "TpqU4PzvV9", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Manwen Liao", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Weitian Zhang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yan Zhu", "gender": "Male", "institution": "The University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Yuxiang Yang", "gender": "unknown", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 47, "n_formula_1": 26, "n_ref_uni": 64, "n_ref": 113, "n_ref_all": 139, "n_fig": 10, "n_tab": 5, "L_tab": 1244, "n_element_tab": 125, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1126, "n_element_tab_1": 89, "formula_len_all": 10003, "formula_len_all_1": 8301, "len_all": 206381, "len_all_1": 78789, "len_abs": 1381, "len_title": 132, "len_sents": 54734, "len_sents_1": 31531, "n_sents": 448, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1386, "title": "Learning Explicit Circuit Representations for Quantum States from Local Measurements", "abs": "Characterizing quantum states is essential for advancing many quantum technologies. Recently, deep neural networks have been applied to learn quantum states by generating implicit representations that map them into classical vectors. Despite their success in predicting state properties, these representations remain a black box, lacking insights into strategies for experimental reconstruction. In this work, we aim to open this black box by developing explicit representations of quantum states through the generation of preparation circuits using a reinforcement learning agent with a local fidelity reward function. Relying solely on measurement data from a few neighboring qubits, our agent accurately recovers properties of target states. Specifically, we design a quantum measurement feature aggregation block which is used to extract global features of quantum states from local measurement data. We also provide a theoretical guarantee for the proposed local fidelity reward function. Extensive experiments demonstrate the effectiveness of our approach in learning various quantum states of up to 100 qubits, including those generated by Instantaneous Quantum Polynomial circuits, evolved by Ising Hamiltonians, and many-body ground states. The learned circuit representations can be further applied to Hamiltonian learning as a downstream task utilizing a simple linear model.", "keywords": [ "quantum state learning", "quantum circuit construction", "local measurements", "reinforcement learning" ], "rating_list": [ 3, 5, 5, 3, 6 ], "soundness_list": [ 1, 1, 3, 2, 2 ], "presentation_list": [ 1, 3, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 5, 5 ] }, { "paper_id": "Tpjq66xwTq", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Eder Medina", "gender": "unknown", "institution": "Department of Computer Science, Princeton University", "country": "", "position": "Postdoc" }, { "name": "Isabel M. de Oliveira", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Rafael Pastrana", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Ryan P Adams", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Professor" }, { "name": "Sigrid Adriaenssens", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 6, "n_ref_uni": 44, "n_ref": 71, "n_ref_all": 126, "n_fig": 20, "n_tab": 3, "L_tab": 277, "n_element_tab": 30, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 135, "n_element_tab_1": 19, "formula_len_all": 616, "formula_len_all_1": 440, "len_all": 182473, "len_all_1": 65515, "len_abs": 1646, "len_title": 142, "len_sents": 58413, "len_sents_1": 31655, "n_sents": 412, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1480, "title": "Real-time design of architectural structures with differentiable mechanics and neural networks", "abs": "Designing mechanically efficient geometry for architectural structures like shells, towers, and bridges is an expensive iterative process.\nExisting techniques for solving such inverse mechanical problems rely on traditional direct optimization methods, which are slow and computationally expensive, limiting iteration speed and design exploration.\nNeural networks would seem to offer a solution, via data-driven amortized optimization, but they often require extensive fine-tuning and cannot ensure that important design criteria, such as mechanical integrity, are met.\nIn this work, we combine neural networks with a differentiable mechanics simulator to develop a model that accelerates the solution of shape approximation problems for architectural structures modeled as bar systems.\nAs a result, our model offers explicit guarantees to satisfy mechanical constraints while generating designs that match target geometries.\nWe validate our model in two tasks, the design of masonry shells and cable-net towers.\nOur model achieves better accuracy and generalization than fully neural alternatives, and comparable accuracy to direct optimization but in real time, enabling fast and sound design exploration.\nWe further demonstrate the real-world potential of our trained model by deploying it in 3D modeling software and by fabricating a physical prototype.\nOur work opens up new opportunities for accelerated physical design enhanced by neural networks for the built environment.", "keywords": [ "Differentiable physics", "mechanical design", "physics-in-the-loop neural networks", "inverse problems", "architectural structures" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 2, 5, 4 ] }, { "paper_id": "Tp9c7s3ZmN", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Boxi Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Jianwei Yin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Jinshan Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Meng Xi", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Researcher" }, { "name": "Tao Jiang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Xuze Tian", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zejian Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 38, "n_ref": 87, "n_ref_all": 102, "n_fig": 5, "n_tab": 2, "L_tab": 603, "n_element_tab": 79, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 603, "n_element_tab_1": 79, "formula_len_all": 39, "formula_len_all_1": 39, "len_all": 120645, "len_all_1": 61413, "len_abs": 1213, "len_title": 102, "len_sents": 31810, "len_sents_1": 29642, "n_sents": 199, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 1220, "title": "Vividportraits: Face Parsing Guided Portrait Animation", "abs": "Portrait animation aims to transfer the facial expressions and movements of a target character onto a reference character. This task presents two main challenges: accurately transferring motion and expressions while fully preserving the identity features of the reference portrait. We introduce Vividportraits, a diffusion-based model designed to effectively meet these objectives. In contrast to existing methods that rely on sparse representations such as facial landmarks, our approach leverages facial parsing maps for motion guidance, enabling a more precise conveyance of subtle expressions. A random scaling technique is applied during training to prevent the model from internalizing identity-specific features from the driving images. Furthermore, we perform foreground-background segmentation on the reference portrait to reduce data redundancy. The long-video generation process is refined to improve consistency across sequences. Our model, exclusively trained on public datasets, demonstrates superior performance relative to current state-of-the-art methods, achieving a notable 8\\% improvement in expression metric. More visual results are available on the anonymous website https://www.vividportraits.cn.", "keywords": [ "face parsing; portrait animation; diffusion model" ], "rating_list": [ 3, 3, 3, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "TowvqbPj8m", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Long Tian", "gender": "Male", "institution": "Xi'an University of Software Engineering Institute", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiyang Liu", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Yixuan Zhang", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Zejun You", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "MS student" }, { "name": "Zhuoqi Ma", "gender": "Female", "institution": "Brown University", "country": "US", "position": "Postdoc" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 30, "n_ref": 55, "n_ref_all": 91, "n_fig": 17, "n_tab": 4, "L_tab": 419, "n_element_tab": 46, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1661, "n_element_tab_1": 106, "formula_len_all": 697, "formula_len_all_1": 696, "len_all": 140180, "len_all_1": 56266, "len_abs": 1560, "len_title": 148, "len_sents": 46355, "len_sents_1": 26025, "n_sents": 325, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1508, "title": "Content-style disentangled representation for controllable artistic image stylization and generation", "abs": "Controllable artistic image stylization and generation aims to render the content provided by text or image with the learned artistic style, where content and style decoupling is the key to achieve satisfactory results. However, current methods for content and style disentanglement primarily rely on image information for supervision, which leads to two problems: 1) models can only support one modality for style or content input;2) incomplete disentanglement resulting in semantic interference from the reference image. To address the above issues, this paper proposes a content-style representation disentangling method for controllable artistic image stylization and generation. We construct a WikiStyle+ dataset consists of artworks with corresponding textual descriptions for style and content. Based on the multimodal dataset, we propose a disentangled content and style representations guided diffusion model. The disentangled representations are first learned by Q-Formers and then injected into a pre-trained diffusion model using learnable multi-step cross-attention layers for better controllable stylization. This approach allows model to accommodate inputs from different modalities. Experimental results show that our method achieves a thorough disentanglement of content and style in reference images under multimodal supervision, thereby enabling a harmonious integration of content and style in the generated outputs, successfully producing style-consistent and expressive stylized images.", "keywords": [ "disentangled representation", "image stylization", "generation" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "ToveGL9vRN", "primary_area": "causal reasoning", "n_author": 7, "author_info_list": [ { "name": "Joanna Wojciechowska", "gender": "Female", "institution": "University of Warsaw", "country": "", "position": "MS student" }, { "name": "Mateusz Gajewski", "gender": "Male", "institution": "Technical University of Poznan", "country": "PL", "position": "PhD student" }, { "name": "Mateusz Olko", "gender": "Male", "institution": "University of Warsaw", "country": "PL", "position": "PhD student" }, { "name": "Miko\u0142aj Morzy", "gender": "Male", "institution": "Poznan University of Technology", "country": "PL", "position": "Associate Professor" }, { "name": "Piotr Mi\u0142o\u015b", "gender": "unknown", "institution": "IDEAS NCBR", "country": "PL", "position": "Researcher" }, { "name": "Piotr Sankowski", "gender": "unknown", "institution": "MIM Solutions", "country": "PL", "position": "CSA" }, { "name": "\u0141ukasz Kuci\u0144ski", "gender": "Male", "institution": "IDEAS NCBR", "country": "PL", "position": "Senior Research Scientist" } ], "n_formula": 15, "n_formula_1": 6, "n_ref_uni": 32, "n_ref": 85, "n_ref_all": 103, "n_fig": 10, "n_tab": 2, "L_tab": 564, "n_element_tab": 54, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1082, "n_element_tab_1": 32, "formula_len_all": 1068, "formula_len_all_1": 513, "len_all": 132843, "len_all_1": 57131, "len_abs": 652, "len_title": 123, "len_sents": 42507, "len_sents_1": 26738, "n_sents": 337, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 654, "title": "Since Faithfulness Fails: The Performance Limits of Neural Causal Discovery", "abs": "Neural causal discovery methods have recently improved in terms of scalability and computational efficiency.\nHowever, there are still opportunities for improving their accuracy in uncovering causal structures.\nWe argue that the key obstacle in unlocking this potential is the faithfulness assumption, commonly used by contemporary neural approaches. We show that this assumption, which is often not satisfied in real-world or synthetic datasets, limits the effectiveness of existing methods. We evaluate\nthe impact of\nfaithfulness violations both qualitatively and quantitatively and provide a unified evaluation framework to facilitate further research.", "keywords": [ "causal discovery", "faithfulness assumption", "neural networks" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "TopFr8GeZy", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Chin Pang Ho", "gender": "Male", "institution": "City University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Haolin Ruan", "gender": "Male", "institution": "City University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Shaohang Xu", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Yining Dong", "gender": "Female", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Zhi Chen", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "Assistant Professor" } ], "n_formula": 128, "n_formula_1": 15, "n_ref_uni": 67, "n_ref": 135, "n_ref_all": 190, "n_fig": 7, "n_tab": 6, "L_tab": 863, "n_element_tab": 26, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 207, "n_element_tab_1": 40, "formula_len_all": 23767, "formula_len_all_1": 4309, "len_all": 262608, "len_all_1": 66799, "len_abs": 1754, "len_title": 107, "len_sents": 74200, "len_sents_1": 29824, "n_sents": 576, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1760, "title": "Target-Oriented Soft-Robust Inverse Reinforcement Learning", "abs": "In imitation learning, when the learning agent is at a state that is outside the demonstration of the expert, it could be difficult for her to choose an action. To overcome this challenge, inverse reinforcement learning (IRL) learns a parameterized reward function based on which we can generalize the expert's behavior to those states that are unseen in the demonstration. However, on the one hand, there could be multiple reward functions that can explain the expert's behavior, leading to reward ambiguity in IRL. On the other hand, though we often consider the transition kernel of the expert to be known to the agent, sometimes the transition kernel of the agent is different from the expert's and is unknown, leading to transition kernel ambiguity in IRL. Drawing on the notion of soft-robust optimization, we build a target-oriented soft-robust IRL (SRIRL) model where the performance of the output policy strikes a flexible balance between risk aversion and expected return maximization towards reward uncertainty in IRL. Moreover, by employing the robust satisficing framework, our SRIRL is also robust to transition kernel ambiguity in IRL. In our target-oriented SRIRL, we keep a target for the performance of the output policy that balances expected return and risk, and we minimize the constraint violation incurred by the difference between the ambiguous transition kernel and the empirical one. We derive tractable reformulation for SRIRL, and we design tailored first-order methods for SRIRL. Numerical results showcase the soft robustness towards reward uncertainty and the robustness against transition kernel ambiguity of SRIRL, as well as the stronger scalability of our first-order methods compared to a state-of-the-art commercial solver.", "keywords": [ "Imitation Learning", "Inverse Reinforcement Learning", "Soft-Robust Optimization", "Robust Optimization", "Optimization Algorithm" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "ToWKyjwDqO", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Austin Xu", "gender": "Male", "institution": "Salesforce AI Research", "country": "US", "position": "Researcher" }, { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "PeiFeng Wang", "gender": "Male", "institution": "Salesforce AI", "country": "US", "position": "Researcher" }, { "name": "Shafiq Joty", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Principal Researcher" }, { "name": "Yilun Zhou", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 44, "n_ref": 144, "n_ref_all": 182, "n_fig": 8, "n_tab": 18, "L_tab": 10878, "n_element_tab": 1000, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 5174, "n_element_tab_1": 566, "formula_len_all": 290, "formula_len_all_1": 220, "len_all": 271928, "len_all_1": 82130, "len_abs": 1829, "len_title": 88, "len_sents": 84919, "len_sents_1": 34979, "n_sents": 602, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1832, "title": "Direct Judgement Preference Optimization", "abs": "Auto-evaluation is crucial for assessing response quality and offering feedback for model development. Recent studies have explored training large language models (LLMs) as generative judges to both evaluate model responses and generate natural language critiques. However, existing models have been trained almost exclusively with supervised fine-tuning (SFT), often only on a small number of datasets, resulting in poor generalization across different evaluation settings and tasks. In this paper, we investigate how learning from both positive and negative data with direct preference optimization (DPO) enhances the evaluation capabilities of LLM judges across three evaluation tasks: pairwise, single ratings, and binary classification. We achieve this by creating three forms of DPO data from a diverse collection of human and synthetic judgements on contemporary model outputs, with the goal of training our model to generate meaningful critiques, make accurate judgements, and understand what constitutes good and bad responses for a given user input. To demonstrate the effectiveness of our method, we train judge models of three sizes: 8B parameters, 12B, and 70B, and conduct a comprehensive study over 13 benchmarks (7 pairwise, 4 single rating, and 2 classification), measuring agreement with human and GPT-4 annotations. Our models exhibit the best aggregate performance, with even our 8B model outperforming strong baselines like GPT-4o and specialized judge models, such as OffsetBias-8B, Auto-J-13B, Prometheus-2-8x7B, and Skywork-Critic-70B, in pairwise benchmarks. Further analysis shows that our judge model robustly counters biases such as position and length bias, flexibly adapts to practitioner-specified evaluation protocols, and provides helpful language feedback for improving downstream generator models.", "keywords": [ "LLM-as-judge", "generative judge", "auto-evaluation" ], "rating_list": [ 6, 8, 3, 3 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 1, 3 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "ToVvoHpk4L", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Feiran Huang", "gender": "Male", "institution": "Jinan University", "country": "CN", "position": "Full Professor" }, { "name": "Junnan Dong", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Xiao Huang", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "", "position": "Assistant Professor" }, { "name": "Xinrun Wang", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Yuanchen Bei", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zijin Hong", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 30, "n_ref": 61, "n_ref_all": 70, "n_fig": 4, "n_tab": 4, "L_tab": 820, "n_element_tab": 74, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 780, "n_element_tab_1": 61, "formula_len_all": 138, "formula_len_all_1": 138, "len_all": 150382, "len_all_1": 64021, "len_abs": 2448, "len_title": 0, "len_sents": 48680, "len_sents_1": 31335, "n_sents": 368, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1522, "title": "$\\texttt{CLR-Bench}$: Evaluating Large Language Models in College-Level Reasoning", "abs": "Large language models (LLMs) have demonstrated their remarkable performance across various language understanding tasks. While emerging benchmarks have been proposed to evaluate LLMs in various domains such as mathematics and computer science, they merely measure the accuracy in terms of the final prediction on multi-choice questions. However, it remains insufficient to verify the essential understanding of LLMs given a chosen choice. To fill this gap, we present $\\texttt{CLR-Bench}$ to comprehensively evaluate the LLMs in complex college-level reasoning. Specifically, $(i)$ we prioritize 16 challenging college disciplines in computer science and artificial intelligence. The dataset contains 5 types of questions, while each question is associated with detailed explanations from experts. $(ii)$ To quantify a fair evaluation of LLMs' reasoning ability, we formalize the criteria with two novel metrics. Q$\\rightarrow$A is utilized to measure the performance of direct **a**nswer prediction, and Q$\\rightarrow$AR effectively considers the joint ability to **a**nswer the question and provide **r**ationale simultaneously. Extensive experiments are conducted with 40 LLMs over 1,018 discipline-specific questions. The results demonstrate the key insights that LLMs, even the best closed-source LLM, i.e., GPT-4 turbo, tends to '***guess***' the college-level answers. It shows a dramatic decrease in accuracy from 63.31\\% Q$\\rightarrow$A to 39.00\\% Q$\\rightarrow$AR, indicating an unsatisfactory reasoning ability.", "keywords": [ "Large Language Models Evaluation", "Benchark and dataset", "College-level Reasoning" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "TnlLMYPfx5", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Dongmin Bang", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Inyoung Sung", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Postdoc" }, { "name": "Sangseon Lee", "gender": "Male", "institution": "Inha University", "country": "KR", "position": "Assistant Professor" }, { "name": "Sun Kim", "gender": "Male", "institution": "Seoul National University, Seoul National University", "country": "", "position": "Full Professor" }, { "name": "Yinhua Piao", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 16, "n_ref_uni": 41, "n_ref": 90, "n_ref_all": 127, "n_fig": 12, "n_tab": 20, "L_tab": 12525, "n_element_tab": 615, "n_fig_1": 7, "n_tab_1": 7, "L_tab_1": 3482, "n_element_tab_1": 177, "formula_len_all": 2553, "formula_len_all_1": 1186, "len_all": 238459, "len_all_1": 67936, "len_abs": 1914, "len_title": 154, "len_sents": 81060, "len_sents_1": 28497, "n_sents": 560, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 106, "L_abs": 1752, "title": "BounDr.E: Predicting Drug-likeness through knowledge alignment and EM-like one-class boundary optimization", "abs": "The advent of generative AI models is revolutionizing drug discovery, generating de novo molecules at unprecedented speed. However, accurately identifying and rescuing drug candidates among countless generated molecules remains an open problem.\nThe essence of this drug-likeness prediction task lies in constructing a compact subspace that encompasses majority of approved drugs with only a small number of unknown compounds (drug candidates) inside.\nComputational challenges arises in constructing a decision boundary on an unbound chemical space that lacks definite negatives, i.e, non drug-likeness.\nApproved drugs exist highly dispersed across structural space, making it more harsh to effectively separate drugs from non-drugs through existing classifiers. \nAddressing such challenges, we introduce BounDr.E: a novel approach for learning a compact boundary of drug-likeness through an Expectation-Maximization (EM)-like iterative optimization process. \nSpecifically, we refine both the boundary and the distribution of the embedding space via metric learning, allowing the model to iteratively tighten the drug-like boundary while pushing non-drug-like compounds outside.\nAugmented by integration of biomedical context within knowledge graphs via multi-modal alignment, our model demonstrates 10% increase in F1 score over the previous state-of-the-art, along with strongest robustness to cross-dataset validation.\nZero-shot toxic compound filtering and comprehensive drug discovery pipeline case studies further showcases its utility in large-scale screening of AI-generated compounds. \nTo facilitate in silico drug discovery, we provide the code and benchmark data under various splitting schemes at: https://anonymous.4open.science/r/boundr_e.", "keywords": [ "Drug-likeness", "one-class boundary", "multi-modal alignment", "drug discovery" ], "rating_list": [ 6, 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 4, 3, 2 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3, 4 ] }, { "paper_id": "Tnd3dZxyEv", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Adrian D Hillier", "gender": "Male", "institution": "ISIS Neutron and Muon facility", "country": "GB", "position": "Principal Researcher" }, { "name": "Jaehoon Cha", "gender": "unknown", "institution": "Science and Technology Facilities Council", "country": "", "position": "Researcher" }, { "name": "Jeyan Thiyagalingam", "gender": "unknown", "institution": "Rutherford Appleton Laboratory, Science and Technology Facilities Council", "country": "", "position": "Principal Researcher" }, { "name": "Jia Bi", "gender": "Female", "institution": "STFC", "country": "", "position": "Researcher" }, { "name": "Keith T. Butler", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Associate Professor" }, { "name": "Kuangdai Leng", "gender": "Male", "institution": "Science and Technology Facilities Council", "country": "", "position": "Researcher" }, { "name": "Samuel Pinilla", "gender": "Male", "institution": "Science and Technology Facilities Council", "country": "", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 5, "n_ref_uni": 70, "n_ref": 110, "n_ref_all": 152, "n_fig": 16, "n_tab": 2, "L_tab": 426, "n_element_tab": 27, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1400, "n_element_tab_1": 38, "formula_len_all": 790, "formula_len_all_1": 218, "len_all": 155279, "len_all_1": 53441, "len_abs": 1561, "len_title": 81, "len_sents": 44837, "len_sents_1": 24582, "n_sents": 350, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1570, "title": "Enhancing Performance of Multilayer Perceptrons by Knot-Gathering Initialization", "abs": "Multilayer perceptrons (MLPs) with ReLU-like activation functions form a high-dimensional, piecewise linear function space, characterized by \"knots\"\u2014points of non-differentiability. The density of such knots within a given input domain measures the MLP's capacity for function approximation. Despite the simplicity of this concept, knots remain underexploited to enhance the practical performance of MLPs. This paper introduces Knot Gathering Initialization (KGI), a novel method that amplifies the local expressiveness of MLPs by increasing the knot density within the input domain prior to training. As an initialization technique, KGI is lightweight, data-independent, and hyperparameter-insensitive. The concept of knots, and hence KGI, can be directly generalized to smooth activation functions from different angles, including geometry, information transmission, and spectral analysis. We demonstrate the effectiveness of KGI across diverse tasks, including curve and surface fitting, image classification, time series regression, physics-informed operator learning, representation disentanglement, and large language model pretraining. These experiments unexceptionally show that KGI improves both accuracy and convergence speed of MLPs, whether used standalone or as components of larger architectures. Promising future directions include: 1) the natural extension of KGI to convolutional and graph convolutional layers, as well as Low-Rank Adaptation (LoRA) for finetuning; and 2) applying knot gathering throughout training, rather than just at initialization.", "keywords": [ "Multilayer perceptron", "expressiveness", "linear region", "knot density", "initialization" ], "rating_list": [ 3, 10, 3, 5, 5 ], "soundness_list": [ 2, 4, 2, 3, 4 ], "presentation_list": [ 3, 4, 2, 4, 3 ], "contribution_list": [ 2, 4, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "Tn8EQIFIMQ", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Haijiang Yan", "gender": "Male", "institution": "University of Warwick", "country": "", "position": "PhD student" }, { "name": "Jian-Qiao Zhu", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Thomas L. Griffiths", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "Professor" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 43, "n_ref": 99, "n_ref_all": 134, "n_fig": 8, "n_tab": 8, "L_tab": 2131, "n_element_tab": 156, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 467, "n_element_tab_1": 42, "formula_len_all": 191, "formula_len_all_1": 829, "len_all": 146885, "len_all_1": 70641, "len_abs": 1692, "len_title": 133, "len_sents": 48048, "len_sents_1": 34365, "n_sents": 344, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1656, "title": "Language Models Trained to do Arithmetic Predict Human Risky and Intertemporal Choice", "abs": "The observed similarities in the behavior of humans and Large Language Models (LLMs) have prompted researchers to consider the potential of using LLMs as models of human cognition. However, several significant challenges must be addressed before LLMs can be legitimately regarded as cognitive models. For instance, LLMs are trained on far more data than humans typically encounter, and may have been directly trained on human data in specific cognitive tasks or aligned with human preferences. Consequently, the origins of these behavioral similarities are not well understood. In this paper, we propose a novel way to enhance the utility of language models as cognitive models. This approach involves (i) leveraging computationally equivalent tasks that both a language model and a rational agent need to master for solving a cognitive problem and (ii) examining the specific task distributions required for a language model to exhibit human-like behaviors. We apply this approach to decision-making -- specifically risky and intertemporal choice -- where the key computationally equivalent task is the arithmetic of expected value calculations. We show that a small language model pretrained on an ecologically valid arithmetic dataset, which we call Arithmetic-GPT, predicts human behavior better than many traditional cognitive models. Pretraining language models on ecologically valid arithmetic datasets is sufficient to produce a strong correspondence between these models and human decision-making. Our results also suggest that language models used as cognitive models should be carefully investigated via ablation studies of the pretraining data.", "keywords": [ "cognitive model", "language model", "computational models of cognition", "rationality", "economics" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "Tn6lrFbiP4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Pichao WANG", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Thomas Brox", "gender": "Male", "institution": "University of Freiburg", "country": "", "position": "Full Professor" }, { "name": "Tianjun Xiao", "gender": "Male", "institution": "Amazon", "country": "", "position": "Senior Applied Scientist" }, { "name": "Tong He", "gender": "Male", "institution": "Amazon", "country": "CN", "position": "Researcher" }, { "name": "Zechen Bai", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zheng Zhang", "gender": "Male", "institution": "Amazon", "country": "CN", "position": "Senior Principal Scientist" }, { "name": "Zheng Shou", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 30, "n_ref": 71, "n_ref_all": 80, "n_fig": 4, "n_tab": 13, "L_tab": 5869, "n_element_tab": 720, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 4967, "n_element_tab_1": 541, "formula_len_all": 587, "formula_len_all_1": 512, "len_all": 195350, "len_all_1": 77346, "len_abs": 1158, "len_title": 127, "len_sents": 53679, "len_sents_1": 29837, "n_sents": 491, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1163, "title": "Bridging Information Asymmetry in Text-video Retrieval: A Data-centric Approach", "abs": "As online video content rapidly grows, the task of text-video retrieval (TVR) becomes increasingly important. A key challenge in TVR is the information asymmetry between video and text: videos are inherently richer in information, while their textual descriptions often capture only fragments of this complexity. This paper introduces a novel, data-centric framework to bridge this gap by enriching textual representations to better match the richness of video content. During training, videos are segmented into event-level clips and captioned to ensure comprehensive coverage. During retrieval, a large language model (LLM) generates semantically diverse queries to capture a broader range of possible matches. To enhance retrieval efficiency, we propose a query selection mechanism that identifies the most relevant and diverse queries, reducing computational cost while improving accuracy. Our method achieves state-of-the-art results across multiple benchmarks, demonstrating the power of data-centric approaches in addressing information asymmetry in TVR. This work paves the way for new research focused on leveraging data to improve cross-modal retrieval.", "keywords": [ "Text-video Retrieval", "Vision-Language Model", "Multimodal" ], "rating_list": [ 8, 6, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 4, 2, 2 ], "confidence_list": [ 5, 5, 5 ] }, { "paper_id": "Tn5B6Udq3E", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Tian Ye", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Yuanzhi Li", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Zeyuan Allen-Zhu", "gender": "Unspecified", "institution": "Meta FAIR Labs", "country": "US", "position": "Researcher" }, { "name": "Zicheng Xu", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 0, "n_ref_uni": 0, "n_ref": 16, "n_ref_all": 74, "n_fig": 19, "n_tab": 9, "L_tab": 1753, "n_element_tab": 94, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1168, "n_element_tab_1": 71, "formula_len_all": 2969, "formula_len_all_1": 0, "len_all": 237346, "len_all_1": 77416, "len_abs": 3052, "len_title": 116, "len_sents": 101284, "len_sents_1": 37389, "n_sents": 764, "n_sents_1": 298, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1035, "title": "Physics of Language Models: Part 2.1, Grade-School Math and the Hidden Reasoning Process", "abs": "Recent advances in language models have demonstrated their capability to solve mathematical reasoning problems, achieving near-perfect accuracy on grade-school level math benchmarks like GSM8K. In this paper, we formally study how language models solve these problems. We design a series of controlled experiments to address several fundamental questions: (1) Can language models truly develop reasoning skills, or do they simply memorize templates? (2) What is the model's hidden (mental) reasoning process? (3) Do models solve math questions using skills similar to or different from humans? (4) Do models trained on GSM8K-like datasets develop reasoning skills beyond those necessary for solving GSM8K problems? (5) What mental process causes models to make reasoning mistakes? (6) How large or deep must a model be to effectively solve GSM8K-level math questions?\n\nOur study uncovers many hidden mechanisms by which language models solve mathematical questions, providing insights that extend beyond current understandings of LLMs.", "keywords": [ "linear probing", "language model", "grade math problems", "logic following", "reasoning" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 4, 5, 2, 4, 3 ] }, { "paper_id": "Tn02m7ZTch", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Anthony Robins", "gender": "unknown", "institution": "University of Otago", "country": "NZ", "position": "Lecturer" }, { "name": "Hayden McAlister", "gender": "Male", "institution": "University of Otago", "country": "NZ", "position": "PhD student" }, { "name": "Lech Szymanski", "gender": "Male", "institution": "University of Otago", "country": "", "position": "Lecturer" } ], "n_formula": 19, "n_formula_1": 20, "n_ref_uni": 17, "n_ref": 33, "n_ref_all": 56, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1651, "formula_len_all_1": 1524, "len_all": 82027, "len_all_1": 52017, "len_abs": 1211, "len_title": 128, "len_sents": 30035, "len_sents_1": 24373, "n_sents": 228, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1215, "title": "Improved Robustness and Hyperparameter Selection in the Dense Associative Memory", "abs": "The Dense Associative Memory generalizes the Hopfield network by allowing for sharper interaction functions. This increases the capacity of the network as an autoassociative memory as nearby learned attractors will not interfere with one another. However, the implementation of the network relies on applying large exponents to the dot product of memory vectors and probe vectors. If the dimension of the data is large the calculation can be very large and result in imprecisions and overflow when using floating point numbers in a practical implementation. We describe the computational issues in detail, modify the original network description to mitigate the problem, and show the modification will not alter the networks' dynamics during update or training. We also show our modification greatly improves hyperparameter selection for the Dense Associative Memory, removing dependence on the interaction vertex and resulting in an optimal region of hyperparameters that does not significantly change with the interaction vertex as it does in the original network. Our modifications also allow us to train a Dense Associative Memory with larger interaction vertices than have been used in any previous literature.", "keywords": [ "Dense Associative Memory", "Modern Hopfield Network", "Associative Memory Robustness", "Stability", "Hyperparameter Selection" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 5, 5, 3, 4 ] }, { "paper_id": "TmKeT3IFTZ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Jingrui He", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "John Birge", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Full Professor" }, { "name": "Lecheng Zheng", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Ruizhong Qiu", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Yuancheng Yu", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Ziwei Wu", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "PhD student" } ], "n_formula": 32, "n_formula_1": 24, "n_ref_uni": 51, "n_ref": 105, "n_ref_all": 137, "n_fig": 7, "n_tab": 13, "L_tab": 5882, "n_element_tab": 377, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 4548, "n_element_tab_1": 222, "formula_len_all": 2984, "formula_len_all_1": 1681, "len_all": 170020, "len_all_1": 69550, "len_abs": 1565, "len_title": 92, "len_sents": 43907, "len_sents_1": 27180, "n_sents": 329, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1573, "title": "Fair Anomaly Detection For Imbalanced Groups", "abs": "Anomaly detection (AD) has been widely studied for decades in many real-world applications, including fraud detection in finance, and intrusion detection for cybersecurity, etc. Due to the imbalanced nature between protected and unprotected groups and the imbalanced distributions of normal examples and anomalies, the learning objectives of most existing anomaly detection methods tend to solely concentrate on the dominating unprotected group. Thus, it has been recognized by many researchers about the significance of ensuring model fairness in anomaly detection. However, the existing fair anomaly detection methods tend to erroneously label most normal examples from the protected group as anomalies in the imbalanced scenario where the unprotected group is more abundant than the protected group. This phenomenon is caused by the improper design of learning objectives, which statistically focus on learning the frequent patterns (i.e., the unprotected group) while overlooking the under-represented patterns (i.e., the protected group). To address these issues, we propose FADIG, a fairness-aware anomaly detection method targeting the imbalanced scenario. It consists of a fairness-aware contrastive learning module and a rebalancing autoencoder module to ensure fairness and handle the imbalanced data issue, respectively. Moreover, we provide the theoretical analysis that shows our proposed contrastive learning regularization guarantees group fairness. Empirical studies demonstrate the effectiveness and efficiency of FADIG across multiple real-world datasets.", "keywords": [ "fairness", "anomaly detection" ], "rating_list": [ 6, 3, 6, 6, 5 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 4, 2 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 2, 3, 4, 5, 4 ] }, { "paper_id": "TmCcNuo03f", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Balaji Krishnamurthy", "gender": "Male", "institution": "Adobe Systems", "country": "IN", "position": "Principal Scientist" }, { "name": "Changyou Chen", "gender": "Male", "institution": "State University of New York, Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Jayakumar Subramanian", "gender": "Male", "institution": "Adobe Systems", "country": "", "position": "Senior Research Scientist" }, { "name": "Rajiv Ratn Shah", "gender": "Male", "institution": "Indraprastha Institute of Information Technology, Delhi", "country": "IN", "position": "Associate Professor" }, { "name": "Varun Khurana", "gender": "Male", "institution": "Adobe Systems", "country": "IN", "position": "Researcher" }, { "name": "zhiqiang xu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "CN", "position": "Assistant Professor" }, { "name": "Yaman Kumar", "gender": "Male", "institution": "Adobe ", "country": "IN", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 2, "n_ref_uni": 45, "n_ref": 110, "n_ref_all": 147, "n_fig": 12, "n_tab": 12, "L_tab": 4356, "n_element_tab": 476, "n_fig_1": 9, "n_tab_1": 6, "L_tab_1": 2210, "n_element_tab_1": 179, "formula_len_all": 1290, "formula_len_all_1": 74, "len_all": 223808, "len_all_1": 66190, "len_abs": 2805, "len_title": 117, "len_sents": 77236, "len_sents_1": 31654, "n_sents": 499, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 69, "L_abs": 2026, "title": "Measuring And Improving Engagement of Text-to-Image Generation Models", "abs": "Recent advances in text-to-image generation have achieved impressive aesthetic quality, making these models usable for both personal and commercial purposes. However, in the fields of marketing and advertising, images are often created to be more engaging, as reflected in user behaviors such as increasing clicks, likes, and purchases, in addition to being aesthetically pleasing. Further, we find that existing image generation metrics like aesthetics, CLIPScore, PickScore, ImageReward, etc. fail to capture viewer engagement. To this end, we introduce the challenge of optimizing the image generation process for improved viewer engagement. In order to study image engagement and utility in real-world marketing scenarios, we collect EngagingImageNet, the first large-scale dataset of images, along with associated user engagement metrics. To address the lack of reliable metrics for assessing image utility, we use the EngagingImageNet dataset to train EngageNet, an engagement-aware Vision Language Model (VLM) that predicts viewer engagement of images by leveraging contextual information about the tweet content, enterprise details, and posting time. We then explore methods to enhance the engagement of text-to-image models, making initial strides in this direction. These include conditioning image generation on improved prompts, supervised fine-tuning of stable diffusion on high-performing images, and reinforcement learning to align stable diffusion with EngageNet-based reward signals, all of which lead to the generation of images with higher viewer engagement. Finally, we propose the Engagement Arena, to benchmark text-to-image models based on their ability to generate engaging images, using EngageNet as the evaluator, thereby encouraging the research community to measure further advances in the engagement of text-to-image modeling. These contributions provide a new pathway for advancing utility-driven image generation, with significant implications for the commercial application of image generation.", "keywords": [ "image generation models", "text to image models", "engagement", "stable diffusion", "dalle" ], "rating_list": [ 6, 8, 5, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 4 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "TmAmuMXkFc", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Bao Pham", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "PhD student" }, { "name": "Beatrice Achilli", "gender": "Female", "institution": "Bocconi University", "country": "IT", "position": "PhD student" }, { "name": "Carlo Lucibello", "gender": "Male", "institution": "Bocconi University", "country": "IT", "position": "Assistant Professor" }, { "name": "Dmitry Krotov", "gender": "unknown", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Enrico Ventura", "gender": "Male", "institution": "Bocconi University", "country": "IT", "position": "Postdoc" }, { "name": "Gabriel Raya", "gender": "Male", "institution": "Tilburg University", "country": "NL", "position": "PhD student" }, { "name": "Gianluigi Silvestri", "gender": "Male", "institution": "OnePlanet Research Center", "country": "NL", "position": "PhD student" }, { "name": "Luca Ambrogioni", "gender": "Male", "institution": "Radboud University Nijmegen", "country": "", "position": "Assistant Professor" } ], "n_formula": 75, "n_formula_1": 32, "n_ref_uni": 37, "n_ref": 76, "n_ref_all": 118, "n_fig": 10, "n_tab": 1, "L_tab": 171, "n_element_tab": 11, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3032, "formula_len_all_1": 1030, "len_all": 146081, "len_all_1": 63225, "len_abs": 1258, "len_title": 65, "len_sents": 50844, "len_sents_1": 28425, "n_sents": 386, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1262, "title": "Losing dimensions: Geometric memorization in generative diffusion", "abs": "Generative diffusion processes are state-of-the-art machine learning models deeply connected with fundamental concepts in statistical physics. Depending on the dataset size and the capacity of the network, their behavior is known to transition from an associative memory regime to a generalization phase in a phenomenon that has been described as a glassy phase transition. Here, using statistical physics techniques, we extend the theory of memorization in generative diffusion to manifold-supported data. Our theoretical and experimental findings indicate that different tangent subspaces are lost due to memorization effects at different critical times and dataset sizes, which depend on the local variance of the data along their directions. Perhaps counterintuitively, we find that, under some conditions, subspaces of higher variance are lost first due to memorization effects. This leads to a selective loss of dimensionality where some prominent features of the data are memorized without a full collapse on any individual training point. We validate our theory with a comprehensive set of experiments on networks trained both in image datasets and on linear manifolds, which result in a remarkable qualitative agreement with the theoretical predictions.", "keywords": [ "Generative diffusion", "Memorization", "Manifold", "Geometry", "Statistical physics" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "TljGdvzFq2", "primary_area": "datasets and benchmarks", "n_author": 17, "author_info_list": [ { "name": "Aston Zhang", "gender": "Unspecified", "institution": "Meta", "country": "US", "position": "Scientist" }, { "name": "Chenguang Zhu", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Principal Researcher" }, { "name": "Chloe Bi", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Jiawei Han", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign (UIUC)", "country": "US", "position": "Full Professor" }, { "name": "Liang Tan", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Melanie Kambadur", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Mike Lewis", "gender": "Male", "institution": "Facebook AI Research", "country": "", "position": "Research Scientist" }, { "name": "Ming Zhong", "gender": "Male", "institution": "University of Illinois Urbana Champaign", "country": "US", "position": "PhD student" }, { "name": "Rui Hou", "gender": "Male", "institution": "Meta Inc. ", "country": "US", "position": "Research Scientist" }, { "name": "Sergey Edunov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "", "position": "MS student" }, { "name": "Sharan Narang", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Sravya Popuri", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Wenhan Xiong", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Xuewei Wang", "gender": "Female", "institution": "Facebook", "country": "", "position": "Machine Learning Engineer" }, { "name": "Zhengxing Chen", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Dhruv Kumar Mahajan", "gender": "Male", "institution": "Meta AI", "country": "", "position": "Researcher" }, { "name": "Laurens van der", "gender": "Male", "institution": "Meta", "country": "US", "position": "Research Scientist" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 26, "n_ref": 86, "n_ref_all": 114, "n_fig": 4, "n_tab": 25, "L_tab": 12160, "n_element_tab": 596, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1182, "n_element_tab_1": 98, "formula_len_all": 99, "formula_len_all_1": 0, "len_all": 328385, "len_all_1": 65287, "len_abs": 1946, "len_title": 116, "len_sents": 104617, "len_sents_1": 32059, "n_sents": 932, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1686, "title": "Law of the Weakest Link: Cross Capabilities of Large Language Models", "abs": "The development and evaluation of Large Language Models (LLMs) have largely focused on individual capabilities. However, this overlooks the intersection of multiple abilities across different types of expertise that are often required for real-world tasks, which we term **cross capabilities**. To systematically explore this concept, we first define seven core individual capabilities and then pair them to form seven common cross capabilities, each supported by a manually constructed taxonomy. Building on these definitions, we introduce CrossEval, a benchmark comprising 1,400 human-annotated prompts, with 100 prompts for each individual and cross capability. To ensure reliable evaluation, we involve expert annotators to assess 4,200 model responses, gathering 8,400 human ratings with detailed explanations to serve as reference examples. Our findings reveal that, in both static evaluations and attempts to enhance specific abilities, current LLMs consistently exhibit the \"Law of the Weakest Link,\" where cross-capability performance is significantly constrained by the weakest component. Specifically, across 58 cross-capability scores from 17 models, 38 scores are lower than all individual capabilities, while 20 fall between strong and weak, but closer to the weaker ability. These results highlight the under-performance of LLMs in cross-capability tasks, making the identification and improvement of the weakest capabilities a critical priority for future research to optimize performance in complex, multi-dimensional scenarios. The code, benchmarks, model responses, and evaluations are available at this [anonymous link](https://anonym-cross-capabilities.github.io/).", "keywords": [ "Cross Capability", "Law of the Weakest Link", "LLM-as-a-Judge", "Large Langauge Models" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "TlFDFKyEIQ", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Chanyoung Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Junseok Lee", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Sukwon Yun", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Tsuyoshi Murata", "gender": "Male", "institution": "Tokyo Institute of Technology", "country": "JP", "position": "Full Professor" }, { "name": "Xin Liu", "gender": "Male", "institution": "AIST, National Institute of Advanced Industrial Science and Technology", "country": "JP", "position": "Researcher" }, { "name": "Yunhak Oh", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 18, "n_ref_uni": 23, "n_ref": 52, "n_ref_all": 84, "n_fig": 11, "n_tab": 6, "L_tab": 2410, "n_element_tab": 182, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1521, "n_element_tab_1": 128, "formula_len_all": 1403, "formula_len_all_1": 1262, "len_all": 130569, "len_all_1": 66590, "len_abs": 1604, "len_title": 132, "len_sents": 42215, "len_sents_1": 29088, "n_sents": 304, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1460, "title": "Oldie but Goodie: Re-illuminating Label Propagation on Graphs with Partially Observed Features", "abs": "In real-world graphs, we often encounter missing feature situations where a few or the majority of node features, e.g., sensitive information, are missed. Although the recently proposed Feature Propagation algorithm mitigates such situations to some degree, it falls short when only partial features are available, sometimes performing worse than traditional structure-based graph models. To overcome this limitation, we spotlight a classical algorithm, Label Propagation (Oldie), and further illuminate its potential, especially when only a partial feature is available. Now called by Goodie, it takes a hybrid approach to obtain embeddings from the Label Propagation branch and Feature Propagation branch. To do so, we first design a GNN-based decoder that enables the Label Propagation branch to output hidden embeddings that align with those of the FP branch. Then, Goodie automatically captures the significance of structure and feature information thanks to the newly designed Structure-Feature Attention. Followed by a novel Pseudo-Label contrastive learning that differentiates the contribution of each positive pair within pseudo-labels originating from the LP branch, Goodie outputs the final prediction for the unlabeled nodes. Through extensive experiments, we demonstrate that our proposed model, Goodie, outperforms the existing state-of-the art methods not only when only a few features are available but also in abundantly available situations.", "keywords": [ "Graph-based Machine Learning", "Missing Feature" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "TlAdgeoDTo", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Adam Tauman Kalai", "gender": "Male", "institution": "OpenAI", "country": "", "position": "research scientist" }, { "name": "Alex Beutel", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Researcher" }, { "name": "Anna-Luisa Brakman", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "David G. Robinson", "gender": "Male", "institution": "Apple University", "country": "", "position": "Faculty" }, { "name": "Johannes Heidecke", "gender": "unknown", "institution": "OpenAI", "country": "", "position": "Researcher" }, { "name": "Keren Gu", "gender": "Female", "institution": "OpenAI", "country": "", "position": "Researcher" }, { "name": "Lilian Weng", "gender": "Female", "institution": "OpenAI", "country": "", "position": "Research Scientist" }, { "name": "Meghan Shah", "gender": "Female", "institution": "OpenAI", "country": "US", "position": "Member of Model Policy Staff" }, { "name": "Pamela Mishkin", "gender": "unknown", "institution": "OpenAI", "country": "US", "position": "Researcher" }, { "name": "Tyna Eloundou", "gender": "unknown", "institution": "OpenAI", "country": "US", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 33, "n_ref": 52, "n_ref_all": 100, "n_fig": 22, "n_tab": 5, "L_tab": 2453, "n_element_tab": 90, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 438, "n_element_tab_1": 24, "formula_len_all": 724, "formula_len_all_1": 416, "len_all": 192963, "len_all_1": 62976, "len_abs": 1285, "len_title": 81, "len_sents": 67397, "len_sents_1": 31227, "n_sents": 544, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 33, "L_abs": 1270, "title": "First-Person Fairness in Chatbots", "abs": "Some chatbots have access to a user\u2019s name when responding. Prior work has\nshown that large language model outputs can change based on the demographic\ntraits correlated with a name, such as gender or race. In this study, we introduce\na scalable method for studying one form of first-person\nfairness\u2014fairness towards the user based on their demographic information\u2014\nacross a large and heterogeneous corpus of actual chats. We leverage a language\nmodel as an AI \u201cresearch assistant\u201d (AI RA) that can privately and scalably analyze\nchat data, surfacing broader trends without exposing specific examples to the\nresearchers. We corroborate the labels of the AI RA with independent human\nannotations, finding it highly consistent with human ratings of gender bias (less so\nfor racial bias). We apply this methodology to a large set of chats with a commercial\nchatbot. We assess overall quality of responses conditional on different names and\nalso subtle differences in similar-quality responses that may in aggregate reinforce\nharmful stereotypes based on gender or race. The largest detected biases are gender\nbiases in older generations of models and in open-ended tasks, like writing a story.\nFinally, evaluations like ours are important for monitoring and reducing biases.", "keywords": [ "fairness", "large language models", "chatbots" ], "rating_list": [ 5, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "Tl8EzmgsEp", "primary_area": "interpretability and explainable AI", "n_author": 1, "author_info_list": [ { "name": "Diogo Cruz", "gender": "Not Specified", "institution": "Instituto Superior T\u00e9cnico", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 24, "n_ref": 70, "n_ref_all": 149, "n_fig": 28, "n_tab": 5, "L_tab": 3486, "n_element_tab": 584, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 976, "formula_len_all_1": 0, "len_all": 172426, "len_all_1": 59161, "len_abs": 1115, "len_title": 119, "len_sents": 65706, "len_sents_1": 30876, "n_sents": 509, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1111, "title": "Understanding the learned look-ahead behavior of chess neural networks", "abs": "We investigate the look-ahead capabilities of chess-playing neural networks, specifically focusing on the Leela Chess Zero policy network. We build on the work of Jenner et al. by analyzing the model's ability to consider future moves and alternative sequences beyond the immediate next move. Our findings reveal that the network's look-ahead behavior is highly context-dependent, varying significantly based on the specific chess position. We demonstrate that the model can process information about board states up to seven moves ahead, utilizing similar internal mechanisms across different future time steps. Additionally, we provide evidence that the network considers multiple possible move sequences rather than focusing on a single line of play. These results offer new insights into the emergence of sophisticated look-ahead capabilities in neural networks trained on strategic tasks, contributing to our understanding of AI reasoning in complex domains. Our work also showcases the effectiveness of interpretability techniques in uncovering cognitive-like processes in artificial intelligence systems.", "keywords": [ "model behavior attribution", "look-ahead planning", "mechanistic interpretability" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Tl6hStJNYX", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Ping Ma", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Full Professor" }, { "name": "Shushan Wu", "gender": "Female", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Wenxuan Zhong", "gender": "unknown", "institution": "University of Georgia", "country": "", "position": "Full Professor" }, { "name": "Xilin Gong", "gender": "Female", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Yongkai Chen", "gender": "Male", "institution": "University of Georgia", "country": "", "position": "PhD student" }, { "name": "Yufang Liu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 24, "n_ref_uni": 29, "n_ref": 66, "n_ref_all": 89, "n_fig": 6, "n_tab": 8, "L_tab": 1625, "n_element_tab": 184, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 408, "n_element_tab_1": 62, "formula_len_all": 1590, "formula_len_all_1": 1148, "len_all": 140671, "len_all_1": 64500, "len_abs": 1385, "len_title": 129, "len_sents": 47306, "len_sents_1": 29608, "n_sents": 402, "n_sents_1": 244, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1403, "title": "Fisher Contrastive Learning: A Robust Solution to the Feature Suppression Effect", "abs": "Self-supervised contrastive learning (SSCL) is a rapidly advancing approach for learning data representations. However, a significant challenge in this paradigm is the feature suppression effect, where useful features for downstream tasks are suppressed due to dominant or easy-to-learn features overshadowing others crucial for downstream performance, ultimately degrading the performance of SSCL models. While prior research has acknowledged the feature suppression effect, solutions with theoretical guarantees to mitigate this issue are still lacking. In this work, we address the feature suppression problem by proposing a novel method, Fisher Contrastive Learning, which unbiasedly and exhaustively estimates the central sufficient dimension reduction function class in SSCL settings. In addition, FCL empirically maintains the embedding dimensionality by maximizing the discriminative power of each linear classifier learned through Fisher Contrastive Learning. We demonstrate that using our proposed method, the class-relevant features are not suppressed by strong or easy-to-learn features on datasets known for strong feature suppression effects. In addition, the embedding dimensionality is not preserved in practice. Furthermore, we show that Fisher Contrastive Learning consistently outperforms existing benchmark methods on standard image benchmarks, illustrating its practical advantages.", "keywords": [ "self-supervised contrastive learning", "sufficient dimension reduction", "Fisher discriminant analysis" ], "rating_list": [ 3, 6, 3 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 2, 2 ] }, { "paper_id": "Tkkrm3pA35", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 1, "author_info_list": [ { "name": "Jie Chen", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "Research Staff Member" } ], "n_formula": 23, "n_formula_1": 15, "n_ref_uni": 47, "n_ref": 63, "n_ref_all": 100, "n_fig": 16, "n_tab": 3, "L_tab": 291, "n_element_tab": 26, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 158, "n_element_tab_1": 11, "formula_len_all": 1863, "formula_len_all_1": 1427, "len_all": 145623, "len_all_1": 69391, "len_abs": 1074, "len_title": 125, "len_sents": 46358, "len_sents_1": 31684, "n_sents": 408, "n_sents_1": 278, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1076, "title": "Graph Neural Preconditioners for Iterative Solutions of Sparse Linear Systems", "abs": "Preconditioning is at the heart of iterative solutions of large, sparse linear systems of equations in scientific disciplines. Several algebraic approaches, which access no information beyond the matrix itself, are widely studied and used, but ill-conditioned matrices remain very challenging. We take a machine learning approach and propose using graph neural networks as a general-purpose preconditioner. They show attractive performance for many problems and can be used when the mainstream preconditioners perform poorly. Empirical evaluation on over 800 matrices suggests that the construction time of these graph neural preconditioners (GNPs) is more predictable and can be much shorter than that of other widely used ones, such as ILU and AMG, while the execution time is faster than using a Krylov method as the preconditioner, such as in inner-outer GMRES. GNPs have a strong potential for solving large-scale, challenging algebraic problems arising from not only partial differential equations, but also economics, statistics, graph, and optimization, to name a few.", "keywords": [ "General-purpose preconditioner", "linear systems", "graph neural networks" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "TkbjqexD8w", "primary_area": "applications to neuroscience & cognitive science", "n_author": 6, "author_info_list": [ { "name": "Chuan Zhou", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Haochen Sui", "gender": "Male", "institution": "University of Michigan", "country": "US", "position": "MS student" }, { "name": "Haoxuan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Jiabao Sean Xiao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Research Assistant" }, { "name": "Yuntian Wu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yuntian Yang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Intern" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 43, "n_ref": 69, "n_ref_all": 77, "n_fig": 5, "n_tab": 1, "L_tab": 694, "n_element_tab": 35, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 694, "n_element_tab_1": 35, "formula_len_all": 747, "formula_len_all_1": 747, "len_all": 115218, "len_all_1": 54559, "len_abs": 1086, "len_title": 137, "len_sents": 30740, "len_sents_1": 25455, "n_sents": 211, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1089, "title": "Invariant Spatiotemporal Representation Learning for Cross-patient Seizure Classification", "abs": "Automatic seizure type classification from electroencephalogram (EEG) data can help clinicians to better diagnose epilepsy. Although many previous studies have focused on the classification problem of seizure EEG data, most of these methods require that there is no distribution shift between training data and test data, which greatly limits the applicability in real-world scenarios. In this paper, we propose an invariant spatiotemporal representation learning method for cross-patient seizure classification. Specifically, we first split the spatiotemporal EEG data into different environments based on heterogeneous risk minimization to reflect the spurious correlations. We then learn invariant spatiotemporal representations and train the seizure classification model based on the learned representations to achieve accurate seizure-type classification across various environments. The experiments are conducted on the largest public EEG dataset, the Temple University Hospital Seizure Corpus (TUSZ) dataset, and the experimental results demonstrate the effectiveness of our method.", "keywords": [ "electroencephalogram data", "spatiotemporal data", "invariant representation learning" ], "rating_list": [ 3, 3, 1, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "TkXisc47la", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Changran XU", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "Undergrad student" }, { "name": "Chengyu Ma", "gender": "Male", "institution": "Ningbo University", "country": "CN", "position": "MS student" }, { "name": "Jianyuan Zhong", "gender": "unknown", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jieru Zhao", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Lingfeng Zhou", "gender": "Male", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "PhD student" }, { "name": "Qiang Xu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Xi Wang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaoyan Yang", "gender": "unknown", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "Full Professor" }, { "name": "Yi Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yunhao Zhou", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "MS student" }, { "name": "Zedong Peng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Zeju Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Zhengyuan Shi", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Zhufei Chu", "gender": "Male", "institution": "Ningbo University", "country": "CN", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 57, "n_ref_all": 71, "n_fig": 4, "n_tab": 9, "L_tab": 4222, "n_element_tab": 331, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2488, "n_element_tab_1": 285, "formula_len_all": 646, "formula_len_all_1": 0, "len_all": 131668, "len_all_1": 62180, "len_abs": 1548, "len_title": 150, "len_sents": 36859, "len_sents_1": 29432, "n_sents": 270, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1556, "title": "DeepCircuitX: Repository-Level RTL Dataset for Code Understanding, Generation, and Multimodal Analysis", "abs": "This paper introduces DeepCircuitX, a comprehensive multimodal dataset designed to advance RTL code understanding, generation, and completion tasks in hardware design automation. Unlike existing datasets, which focus either on file-level RTL code or downstream netlist and layout data, DeepCircuitX spans repository, file, module, and block-level RTL code, providing a more holistic resource for training and evaluating large language models (LLMs). The dataset is enriched with Chain of Thought (CoT) annotations that offer detailed functionality and structure descriptions at multiple levels, enhancing its utility for RTL code understanding, generation, and completion.\n\nIn addition to RTL data, DeepCircuitX includes synthesized netlists and power-performance-area (PPA) metrics, allowing for early-stage design exploration and PPA prediction directly from RTL code. We establish comprehensive benchmarks for RTL code understanding, generation, and completion using open-source models such as CodeLlama, CodeT5+, and CodeGen, demonstrating substantial improvements in task performance. Furthermore, we introduce and evaluate models for PPA prediction, setting new benchmarks for RTL-to-PPA analysis. We conduct human evaluations and reviews to confirm the high quality and functionality of the generated RTL code and annotations. Our experimental results show that DeepCircuitX significantly improves model performance across multiple benchmarks, underscoring its value as a critical resource for advancing RTL code tasks in hardware design automation.", "keywords": [ "Multi-modal dataset of EDA", "AI for EDA", "LLM for RTL", "Circuit Learning" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "Tk1VQDadfL", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Minghui Zhu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Full Professor" }, { "name": "Shicheng Liu", "gender": "unknown", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 99, "n_formula_1": 10, "n_ref_uni": 48, "n_ref": 130, "n_ref_all": 191, "n_fig": 8, "n_tab": 5, "L_tab": 1119, "n_element_tab": 82, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 43, "n_element_tab_1": 10, "formula_len_all": 19037, "formula_len_all_1": 1484, "len_all": 230248, "len_all_1": 77037, "len_abs": 1430, "len_title": 167, "len_sents": 70009, "len_sents_1": 36524, "n_sents": 595, "n_sents_1": 274, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1439, "title": "UTILITY: Utilizing Explainable Reinforcement Learning to Improve Reinforcement Learning", "abs": "Reinforcement learning (RL) faces two challenges: (1) The RL agent lacks explainability. (2) The trained RL agent is, in many cases, non-optimal and even far from optimal. To address the first challenge, explainable reinforcement learning (XRL) is proposed to explain the decision-making of the RL agent. In this paper, we demonstrate that XRL can also be used to address the second challenge, i.e., improve RL performance. Our method has two parts. The first part provides a two-level explanation for why the RL agent is not optimal by identifying the mistakes made by the RL agent. Since this explanation includes the mistakes of the RL agent, it has the potential to help correct the mistakes and thus improve RL performance. The second part formulates a constrained bi-level optimization problem to learn how to best utilize the two-level explanation to improve RL performance. In specific, the upper level learns how to use the high-level explanation to shape the reward so that the corresponding policy can maximize the cumulative ground truth reward, and the lower level learns the corresponding policy by solving a constrained RL problem formulated using the low-level explanation. We propose a novel algorithm to solve this constrained bi-level optimization problem, and theoretically guarantee that the algorithm attains global optimality. We use MuJoCo experiments to show that our method outperforms state-of-the-art baselines.", "keywords": [ "explainable reinforcement learning" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 2, 4, 4, 3 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "TjvSFVJdzJ", "primary_area": "optimization", "n_author": 8, "author_info_list": [ { "name": "Chao Qian", "gender": "Male", "institution": "Nanjing university", "country": "CN", "position": "Full Professor" }, { "name": "Chenyang Wu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Dong Li", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Ke Xue", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Lei Song", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Zongzhang Zhang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Chenxiao Gao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 63, "n_ref": 116, "n_ref_all": 145, "n_fig": 7, "n_tab": 1, "L_tab": 72, "n_element_tab": 8, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1119, "formula_len_all_1": 775, "len_all": 200029, "len_all_1": 67134, "len_abs": 1335, "len_title": 92, "len_sents": 58818, "len_sents_1": 32550, "n_sents": 430, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1322, "title": "Reinforced In-Context Black-Box Optimization", "abs": "Black-Box Optimization (BBO) has found successful applications in many fields of science and engineering. Recently, there has been a growing interest in meta-learning particular components of BBO algorithms to speed up optimization and get rid of tedious hand-crafted heuristics. As an extension, learning the entire algorithm from data requires the least labor from experts and can provide the most flexibility. In this paper, we propose RIBBO, a method to reinforce-learn a BBO algorithm from offline data in an end-to-end fashion. RIBBO employs expressive sequence models to learn the optimization histories produced by multiple behavior algorithms and tasks, leveraging the in-context learning ability of large models to extract task information and make decisions accordingly. Central to our method is to augment the optimization histories with *regret-to-go* tokens, which are designed to represent the performance of an algorithm based on cumulative regret over the future part of the histories. The integration of regret-to-go tokens enables RIBBO to automatically generate sequences of query points that satisfy the user-desired regret, which is verified by its universally good empirical performance on diverse problems, including BBO benchmark functions, hyper-parameter optimization and robot control problems.", "keywords": [ "Black-box Optimization", "In-context Learning" ], "rating_list": [ 3, 8, 3, 8 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "TjuS86sQv8", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Amir Zait", "gender": "Male", "institution": "Google", "country": "IL", "position": "Researcher" }, { "name": "Daniel McDuff", "gender": "Male", "institution": "Google", "country": "US", "position": "Principal Researcher" }, { "name": "Danny Karmon", "gender": "Male", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "David Alexander Munday", "gender": "unknown", "institution": "Google", "country": "US", "position": "Principal Researcher" }, { "name": "Ilia Labzovsky", "gender": "unknown", "institution": "Google", "country": "IL", "position": "Researcher" }, { "name": "Isaac Galatzer-Levy", "gender": "unknown", "institution": "Google", "country": "", "position": "Senior Staff Research Scientist" }, { "name": "Rivka Moroshko", "gender": "unknown", "institution": "Google", "country": "IL", "position": "Researcher" }, { "name": "Xin Liu", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist " } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 26, "n_ref": 33, "n_ref_all": 56, "n_fig": 4, "n_tab": 5, "L_tab": 4391, "n_element_tab": 407, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 4267, "n_element_tab_1": 414, "formula_len_all": 198, "formula_len_all_1": 0, "len_all": 102551, "len_all_1": 61128, "len_abs": 1481, "len_title": 137, "len_sents": 31777, "len_sents_1": 26958, "n_sents": 201, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1736, "title": "The Cognitive Capabilities of Generative AI: A Comparative Analysis with Human Benchmarks", "abs": "There is increasing interest in tracking the capabilities of general intelligence foundation models. This study benchmarks leading large language models (LLMs) and vision language models (VLMs) against human performance on the Wechsler Adult Intelligence Scale (WAIS-IV), a comprehensive, population-normed assessment of underlying human cognition and intellectual abilities, with a focus on the domains of Verbal Comprehension (VCI), Working Memory (WMI), and Perceptual Reasoning (PRI). Most models demonstrated exceptional capabilities in the storage, retrieval, and manipulation of tokens such as arbitrary sequences of letters and numbers, with performance on the Working Memory Index (WMI) greater or equal to the 99.5th percentile when compared to human population normative ability. Performance on the Verbal Comprehension Index (VCI) which measures retrieval of acquired information, and linguistic understanding about the meaning of words and their relationships to each other, also demonstrated consistent performance at or above the 98th percentile. Despite these broad strengths, we observed consistently poor performance on the Perceptual Reasoning Index (PRI; range 0.1-10th percentile) from multimodal models indicating profound inability to interpret and reason on visual information.\nSome more nuanced differences in performance were also observed. Models were consistently stronger on the WMI compared to the VCI, indicating stronger capabilities in storage, manipulation, and retrieval of data than language understanding. Smaller and older model versions consistently performed worse, indicating that training data, parameter count, and advances in tuning are resulting in significant advances in cognitive ability.", "keywords": [ "LLM", "VLM", "GenAI", "Abstract Reasoning", "Cognitive Benchmarking", "Discrete Intellectual Abilities", "Memory" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "TjTe2sEOMP", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Chunlei Li", "gender": "unknown", "institution": "Wuhan University of Technology", "country": "CN", "position": "Researcher" }, { "name": "Jingliang Hu", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Researcher" }, { "name": "Lichao Mou", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Researcher" }, { "name": "Yilei Shi", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Researcher" }, { "name": "Xiaoxiang Zhu", "gender": "Female", "institution": "Technical University Munich", "country": "DE", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 60, "n_ref_all": 81, "n_fig": 11, "n_tab": 4, "L_tab": 939, "n_element_tab": 107, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1400, "n_element_tab_1": 69, "formula_len_all": 243, "formula_len_all_1": 243, "len_all": 111097, "len_all_1": 52503, "len_abs": 1570, "len_title": 191, "len_sents": 26773, "len_sents_1": 22689, "n_sents": 205, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 142, "L_abs": 1582, "title": "All-in-One: Prompt-Driven Mixture of Hallucination-Aware Experts for Universal Anomaly Detection Across Multi-Modal Multi-Organ Medical Images", "abs": "Unsupervised anomaly detection in medical images facilitates practical clinical adoption by identifying abnormalities without relying on scarce and costly annotated data. However, prior works have predominantly focused on specialized models for individual organs and modalities, impeding knowledge transfer and scalable deployment. In this paper, we investigate a task of universal anomaly detection guided by natural language prompts. We propose a prompt-driven mixture of experts framework that detects anomalies across multiple organs and modalities within a single network. Specifically, our method comprises encoders for vision and text, a routing network, and a mixture of hallucination-minimized expert decoders. An image and a prompt describing the organ and modality are fed to the encoders. The routing network then selects specialized yet collaborative expert decoders to analyze the image. We observe that anomaly detection models often erroneously identify normal image regions as anomalous, a phenomenon we term ``hallucinatory anomaly''. To address this issue, we design hallucination-aware experts that produce improved anomaly maps by jointly learning reconstruction and minimizing these false positives. For comprehensive evaluation, we curate a diverse dataset of 12,153 images spanning 5 modalities and 4 organs. Extensive experiments demonstrate state-of-the-art anomaly detection performance in this universal setting. Moreover, the natural language conditioning enables interpretability and user interaction. The code and data will be made publicly available.", "keywords": [ "universal anomaly detection", "medical images", "MoE", "hallucinatory anomalies" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "TjP1d8PP8l", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Chenjia Bai", "gender": "Male", "institution": "TeleAI, China Telecom", "country": "CN", "position": "Research Scientist" }, { "name": "Fei Wu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Haofu Qian", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jiatao Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Wei Song", "gender": "unknown", "institution": "Zhejiang Lab", "country": "", "position": "Full Professor" }, { "name": "Xuelong Li", "gender": "Male", "institution": "China Telecom", "country": "CN", "position": "Full Professor" } ], "n_formula": 25, "n_formula_1": 10, "n_ref_uni": 58, "n_ref": 125, "n_ref_all": 138, "n_fig": 6, "n_tab": 9, "L_tab": 5208, "n_element_tab": 762, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 2658, "n_element_tab_1": 338, "formula_len_all": 1915, "formula_len_all_1": 485, "len_all": 231762, "len_all_1": 70038, "len_abs": 1191, "len_title": 100, "len_sents": 56460, "len_sents_1": 30590, "n_sents": 426, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1226, "title": "Discriminator-Guided Embodied Planning for LLM Agent", "abs": "Large Language Models (LLMs) have showcased remarkable reasoning capabilities in various domains, yet face challenges in complex embodied tasks due to the need for a coherent long-term policy and context-sensitive environmental understanding. Previous work performed LLM refinement relying on outcome-supervised feedback, which can be costly and ineffective. In this work, we introduce a novel framework, Discriminator-Guided Action Optimization (DGAP), for facilitating the optimization of LLM action plans via step-wise signals. Specifically, we employ a limited set of demonstrations to enable the discriminator to learn a score function, which assesses the alignment between LLM-generated actions and the underlying optimal ones at every step. Based on the discriminator, LLMs are prompted to generate actions that maximize the score, utilizing historical action-score pair trajectories as guidance. Under mild conditions, DGAP resembles critic-regularized optimization and has been demonstrated to achieve a stronger policy than the LLM planner. In experiments across different LLMs (GPT-4, Llama3-70B) in ScienceWorld and VirtualHome, our method achieves superior performance and better efficiency than previous methods.", "keywords": [ "LLM Agent", "Embodied Planning", "Discriminator", "Critic-Regularized Optimization" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 4, 3, 2 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "Tisu1L0Jwt", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "ChuXin Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Tianzhu Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Wenfei Yang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Postdoc" }, { "name": "Xiang Liu", "gender": "Male", "institution": "Dongguan University of Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 41, "n_ref": 111, "n_ref_all": 139, "n_fig": 10, "n_tab": 9, "L_tab": 2803, "n_element_tab": 418, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1530, "formula_len_all_1": 815, "len_all": 178282, "len_all_1": 69569, "len_abs": 1722, "len_title": 123, "len_sents": 56352, "len_sents_1": 34908, "n_sents": 428, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1742, "title": "State Space Model Meets Transformer: A New Paradigm for 3D Object Detection", "abs": "DETR-based methods, which use multi-layer transformer decoders to refine object queries iteratively, have shown promising performance in 3D indoor object detection. However, the scene point features in the transformer decoder remain fixed, leading to minimal contributions from later decoder layers, thereby limiting performance improvement. Recently, State Space Models (SSM) have shown efficient context modeling ability with linear complexity through iterative interactions between system states and inputs. Inspired by SSMs, we propose a new 3D object DEtection paradigm with an interactive STate space model (DEST). In the interactive SSM, we design a novel state-dependent SSM parameterization method that enables system states to effectively serve as queries in 3D indoor detection tasks. In addition, we introduce four key designs tailored to the characteristics of point cloud and SSM: The serialization and bidirectional scanning strategies enable bidirectional feature interaction among scene points within the SSM. The inter-state attention mechanism models the relationships between state points, while the gated feed-forward network enhances inter-channel correlations. To the best of our knowledge, this is the first method to model queries as system states and scene points as system inputs, which can simultaneously update scene point features and query features with linear complexity. Extensive experiments on two challenging datasets demonstrate the effectiveness of our DEST-based method. Our method improves the GroupFree baseline in terms of $\\text{AP}_{50}$ on ScanNet V2 (+5.3) and SUN RGB-D (+3.2) datasets. Based on the VDETR baseline, Our method sets a new state-of-the-art on the ScanNetV2 and SUN RGB-D datasets.", "keywords": [ "Point Cloud; 3D Object Detection; State Space Model" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "Thv66GmqZS", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Aimin PAN", "gender": "Male", "institution": "Zhejiang Lab, Hangzhou, China", "country": "", "position": "Researcher" }, { "name": "Changhai Zhou", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Fei Yang", "gender": "Male", "institution": "Zhejiang Lab", "country": "CN", "position": "Researcher" }, { "name": "Ruifeng Li", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yuhua Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 7, "n_ref_uni": 43, "n_ref": 76, "n_ref_all": 115, "n_fig": 8, "n_tab": 12, "L_tab": 6667, "n_element_tab": 552, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2174, "n_element_tab_1": 166, "formula_len_all": 892, "formula_len_all_1": 467, "len_all": 186571, "len_all_1": 62743, "len_abs": 1504, "len_title": 145, "len_sents": 53358, "len_sents_1": 25819, "n_sents": 400, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1509, "title": "Bi-Share LoRA: Enhancing the Parameter Efficiency of LoRA with Intra-Layer and Inter-Layer Sharing", "abs": "Low-Rank Adaptation (LoRA) is a widely adopted parameter-efficient fine-tuning method for large language models (LLMs) to adapt to downstream tasks. However, in scenarios where multiple LoRA models are deployed simultaneously, standard LoRA introduces substantial trainable parameters, resulting in significant memory overhead and inference latency, particularly when supporting thousands of downstream tasks on a single server. While existing methods reduce stored parameters via parameter sharing, they fail to capture both local and global information simultaneously. To address this issue, we propose Bi-Share LoRA, which integrates local parameters with intra-layer and inter-layer shared parameters to more effectively capture information at both local and global levels. By sharing parameters both within and across layers, our method significantly reduces the number of trainable parameters while preserving or improving model performance. Additionally, we set a local LoRA to capture local parameters, enabling more precise and fine-grained information extraction at the local level. The final implementation introduces three parallel sub-LoRAs and designs transformation techniques to adapt shared parameters of varying shapes, ensuring compatibility and efficient sharing. Experiments on the 7B, 8B, and 13B versions of Llama show\nthat Bi-Share LoRA, with only 44.59% of the parameters of standard LoRA, outperforms LoRA by approximately 0.33% on commonsense reasoning and 2.08% on MMLU benchmarks.", "keywords": [ "Parameter-efficient fine-tuning", "parameter-sharing" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 4, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "Thnk4ez3wN", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Horst Samulowitz", "gender": "unknown", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Inwon Kang", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "PhD student" }, { "name": "Oshani Seneviratne", "gender": "Female", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "Director of Data Operations" }, { "name": "Parikshit Ram", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Yi Zhou", "gender": "Female", "institution": "International Business Machines", "country": "US", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 9, "n_ref_uni": 48, "n_ref": 95, "n_ref_all": 143, "n_fig": 17, "n_tab": 25, "L_tab": 11227, "n_element_tab": 1199, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 4727, "n_element_tab_1": 229, "formula_len_all": 1299, "formula_len_all_1": 545, "len_all": 229317, "len_all_1": 74542, "len_abs": 1265, "len_title": 105, "len_sents": 70452, "len_sents_1": 32657, "n_sents": 530, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1038, "title": "On Learning Representations for Tabular Dataset Distillation", "abs": "Dataset distillation generates a small set of information-rich instances from a large dataset, resulting in reduced storage requirements, privacy or copyright risks, and computational costs for downstream modeling, though much of the research has focused on the image data modality. We study tabular data distillation, which brings in novel challenges such as the inherent feature heterogeneity and the common use of non-differentiable learning models (such as decision tree ensembles and nearest-neighbor predictors). To mitigate these challenges, we present TDColER, a tabular data distillation framework via column embeddings-based representation learning. To evaluate this framework, we also present a tabular data distillation benchmark, TDBench. Based on an elaborate evaluation on TDBench, resulting in 226,200 distilled datasets and 541,980 models trained on them, we demonstrate that TDColER is able to boost the distilled data quality of off-the-shelf distillation schemes by 0.5-143% across 7 different tabular learning models.", "keywords": [ "Dataset Distillation", "Tabular Data", "Representation Learning", "Autoencoders" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "ThhQyIruEs", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Chen-Song Zhang", "gender": "Male", "institution": "Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Minrui Lv", "gender": "Female", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Undergrad student" }, { "name": "Yan Xie", "gender": "unknown", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 31, "n_ref": 51, "n_ref_all": 81, "n_fig": 11, "n_tab": 8, "L_tab": 2294, "n_element_tab": 299, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 2235, "n_element_tab_1": 273, "formula_len_all": 435, "formula_len_all_1": 271, "len_all": 131985, "len_all_1": 64977, "len_abs": 1372, "len_title": 155, "len_sents": 39130, "len_sents_1": 27652, "n_sents": 311, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 108, "L_abs": 1398, "title": "MGCFNN: A Neural MultiGrid Solver with Novel Fourier Neural Network for High Wave Number Helmholtz Equations", "abs": "Solving high wavenumber Helmholtz equations is notoriously challenging. Traditional solvers have yet to yield satisfactory results, and most neural network methods struggle to accurately solve cases with extremely high wavenumbers within heterogeneous media. This paper presents an advanced multigrid-hierarchical AI solver, tailored specifically for high wavenumber Helmholtz equations. We adapt the MGCNN architecture to align with the problem setting and incorporate a novel Fourier neural network (FNN) to match the characteristics of Helmholtz equations. FNN, mathematically akin to the convolutional neural network (CNN), enables faster propagation of source influence during the solve phase, making it particularly suitable for handling large size, high wavenumber problems. We conduct supervised learning tests against numerous neural operator learning methods to demonstrate the superior learning capabilities of our solvers. Additionally, we perform scalability tests using an unsupervised strategy to highlight our solvers' significant speedup over the most recent specialized AI solver and AI-enhanced traditional solver for high wavenumber Helmholtz equations. We also carry out an ablation study to underscore the effectiveness of the multigrid hierarchy and the benefits of introducing FNN. Notably, our solvers exhibit optimal convergence of $\\mathcal{O}(k)$ up to $k \\approx 2000$.", "keywords": [ "neural multigrid solver", "Fourier transform", "Helmholtz equations", "high wave number", "heterogeneous medium" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "ThRMTCgpvo", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "Ada Langford", "gender": "Female", "institution": "Research, Microsoft", "country": "US", "position": "Intern" }, { "name": "Dinesh Jayaraman", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Assistant Professor" }, { "name": "Haoran Xu", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" }, { "name": "John Langford", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Kwangjun Ahn", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Manan Tomar", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "PhD student" }, { "name": "Qinghua Liu", "gender": "Male", "institution": "Microsoft Research, NYC", "country": "US", "position": "Postdoc" }, { "name": "Alex Matthew Lamb", "gender": "Male", "institution": "Microsoft Research NYC", "country": "", "position": "Researcher" }, { "name": "Edward Shichao Hu", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 25, "n_ref": 39, "n_ref_all": 59, "n_fig": 7, "n_tab": 5, "L_tab": 299, "n_element_tab": 48, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 276, "n_element_tab_1": 47, "formula_len_all": 1431, "formula_len_all_1": 979, "len_all": 145419, "len_all_1": 62841, "len_abs": 1053, "len_title": 104, "len_sents": 54385, "len_sents_1": 28643, "n_sents": 544, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1061, "title": "Learning to Achieve Goals with Belief State Transformers", "abs": "We introduce the \"Belief State Transformer\", a next-token predictor that takes both a prefix and suffix as inputs, with a novel objective of predicting both the next token for the prefix and the previous token for the suffix. The Belief State Transformer effectively learns to solve challenging problems that conventional forward-only transformers struggle with, in a domain-independent fashion. Key to this success is learning a compact belief state that captures all relevant information necessary for accurate predictions.\nEmpirical ablations show that each component of the model is essential in difficult scenarios where standard Transformers fall short. \nFor the task of story writing with known prefixes and suffixes, our approach outperforms the Fill-in-the-Middle method for reaching known goals and demonstrates improved performance even when the goals are unknown. \nAltogether, the Belief State Transformer enables more efficient goal-conditioned decoding, better test-time inference, and high-quality text representations on small scale problems.", "keywords": [ "representation learning", "transformers", "next-token prediction", "reasoning", "planning" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "TgTxJALwDz", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Haitao Ma", "gender": "Male", "institution": "Harbin Engineering University", "country": "CN", "position": "Associate Professor" }, { "name": "Jinge Bao", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Yunlong Xiao", "gender": "Male", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Yuqi Li", "gender": "unknown", "institution": "Harbin Engineering University", "country": "CN", "position": "PhD student" }, { "name": "zhouhang shi", "gender": "unknown", "institution": "Harbin Engineering University", "country": "CN", "position": "MS student" } ], "n_formula": 24, "n_formula_1": 23, "n_ref_uni": 65, "n_ref": 78, "n_ref_all": 85, "n_fig": 0, "n_tab": 3, "L_tab": 1584, "n_element_tab": 142, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1275, "n_element_tab_1": 132, "formula_len_all": 1357, "formula_len_all_1": 1363, "len_all": 151291, "len_all_1": 65352, "len_abs": 1080, "len_title": 123, "len_sents": 31122, "len_sents_1": 27198, "n_sents": 237, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1029, "title": "Language Model for Large-Text Transmission in Noisy Quantum Communications", "abs": "Quantum communication has the potential to revolutionise information processing, providing unparalleled security and increased capacity compared to its classical counterpart by using the principles of quantum mechanics. However, the presence of errors poses a significant challenge to realising these advantages. While strategies like quantum error correction and quantum error mitigation have been developed to address these errors, they often come with substantial overhead, hindering the practical transmission of large texts. Here, we introduce an application of machine learning frameworks for natural language processing to enhence the performance of noisy quantum communications, particularly superdense coding. Using BERT, a model known for its capabilities in natural language processing, we demonstrate that language-model-assisted quantum communication protocols can substantially improve the efficiency of large-scale information transmission. This brings us closer to the practical realisation of a quantum internet.", "keywords": [ "Language Model; Noisy Superdense Coding; Text Transmission" ], "rating_list": [ 1, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "Tg8RLxpMDu", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Aneesh Pappu", "gender": "unknown", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Jamie Hayes", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "William P. Porter", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "I Shumailov", "gender": "Unspecified", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 45, "n_ref": 95, "n_ref_all": 130, "n_fig": 18, "n_tab": 2, "L_tab": 3663, "n_element_tab": 101, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 70, "formula_len_all_1": 0, "len_all": 200580, "len_all_1": 72137, "len_abs": 1857, "len_title": 51, "len_sents": 69531, "len_sents_1": 38334, "n_sents": 486, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1875, "title": "Measuring memorization in RLHF for code completion", "abs": "Reinforcement learning with human feedback (RLHF) has become the dominant method to align large models to user preferences.\nUnlike fine-tuning, for which there are many studies regarding training data memorization, it is not clear how memorization is affected by or introduced in the RLHF alignment process.\nUnderstanding this relationship is important as real user data may be collected and used to align large models; if user data is memorized during RLHF and later regurgitated, this could raise privacy concerns. In addition to RLHF, other methods such as Direct Preference Optimization (DPO) and $\\Psi$PO have gained popularity for learning directly from human preferences, removing the need for optimizing intermediary reward models with reinforcement learning.\nIn this work, we analyze how training data memorization can surface and propagate through each phase of RLHF and direct preference learning.\nWe focus our study on code completion models, as code completion is one of the most popular use cases for large language models. We find that RLHF significantly decreases the chance that data used for reward modeling and reinforcement learning is memorized in comparison to directly fine-tuning on this data, but that examples already memorized during the fine-tuning stage of RLHF, will, in the majority of cases, remain memorized after RLHF. In contrast, we find that aligning by learning directly from human preference data via a special case of $\\Psi$PO, Identity Preference Optimization (IPO), increases the likelihood that training data is regurgitated compared to RLHF. Our work suggests that RLHF, as opposed to direct preference learning, is a safer way to mitigate the risk of regurgitating sensitive preference data when aligning large language models. We find our conclusions are robust across multiple code completion datasets, tasks, and model scales.", "keywords": [ "rlhf", "memorization", "code completion" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 1, 4, 3 ] }, { "paper_id": "Tepaft7632", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Sangyong Lee", "gender": "Male", "institution": "Artificial Intelligent Laboratory at OKESTRO Co., Ltd.", "country": "KR", "position": "AI Researcher" }, { "name": "Subo Hwang", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Undergrad student" } ], "n_formula": 38, "n_formula_1": 9, "n_ref_uni": 38, "n_ref": 67, "n_ref_all": 85, "n_fig": 9, "n_tab": 5, "L_tab": 2348, "n_element_tab": 204, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 737, "n_element_tab_1": 48, "formula_len_all": 1874, "formula_len_all_1": 457, "len_all": 147237, "len_all_1": 57636, "len_abs": 1201, "len_title": 134, "len_sents": 47292, "len_sents_1": 27255, "n_sents": 356, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1207, "title": "MADCluster: Model-agnostic Anomaly Detection with Self-supervised Clustering Network", "abs": "In this paper, we propose MADCluster, a novel model-agnostic anomaly detection framework utilizing self-supervised clustering. MADCluster is applicable to various deep learning architectures and addresses the 'hypersphere collapse' problem inherent in existing deep learning-based anomaly detection methods. The core idea is to cluster normal pattern data into a `single cluster' while simultaneously learning the cluster center and mapping data close to this center. Also, to improve expressiveness and enable effective single clustering, we propose a new 'One-directed Adaptive loss'. The optimization of this loss is mathematically proven. MADCluster consists of three main components: Base Embedder capturing high-dimensional temporal dynamics, Cluster Distance Mapping, and Sequence-wise Clustering for continuous center updates. Its model-agnostic characteristics are achieved by applying various architectures to the Base Embedder. Experiments on four time series benchmark datasets demonstrate that applying MADCluster improves the overall performance of comparative models. In conclusion, the compatibility of MADCluster shows potential for enhancing model performance across various architectures.", "keywords": [ "Anomaly Detection", "Model-agnostic", "Self-supervised Clustering", "Hypersphere Collapse", "One-directed Adaptive loss" ], "rating_list": [ 5, 5, 6, 3, 5 ], "soundness_list": [ 3, 2, 3, 1, 2 ], "presentation_list": [ 4, 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4, 3 ] }, { "paper_id": "TeVAZXr3yv", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Ashish Seth", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Dinesh Manocha", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Professor" }, { "name": "Oriol Nieto", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Ramani Duraiswami", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "Full Professor" }, { "name": "S Sakshi", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Sonal Kumar", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Sreyan Ghosh", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Utkarsh Tyagi", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "MS student" }, { "name": "Ramaneswaran S", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 47, "n_ref": 102, "n_ref_all": 120, "n_fig": 9, "n_tab": 14, "L_tab": 7942, "n_element_tab": 478, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3953, "n_element_tab_1": 357, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 216041, "len_all_1": 66426, "len_abs": 311, "len_title": 119, "len_sents": 56640, "len_sents_1": 27753, "n_sents": 415, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 70, "L_abs": 1287, "title": "MMAU: A Massive Multi-Task Audio Understanding and Reasoning Benchmark", "abs": "The ability to comprehend audio\u2014which includes speech, non-speech sounds, and music\u2014is crucial for AI agents to interact effectively with the world. We present MMAU, a novel benchmark designed to evaluate multimodal audio understanding models on tasks requiring expert-level knowledge and complex reasoning. MMAU comprises 10k carefully curated audio clips paired with human-annotated natural language questions and answers spanning speech, environmental sounds, and music. It includes information extraction and reasoning questions, requiring models to demonstrate 27 distinct skills across unique and challenging tasks. Unlike existing benchmarks, MMAU emphasizes advanced perception and reasoning with domain-specific knowledge, challenging models to tackle tasks akin to those faced by experts. We assess 18 open-source and proprietary (Large) Audio-Language Models, demonstrating the significant challenges\nposed by MMAU. Notably, even the most advanced Gemini Pro v1.5 achieves only 52.97% accuracy, and the state-of-the-art open-source Qwen2-Audio achieves only 52.50%, highlighting considerable room for improvement. We believe MMAU will drive the audio and multimodal research community to develop more advanced audio understanding models capable of solving complex audio tasks.", "keywords": [ "Benchmark", "Audio Language Models", "Complex Reasoning" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "TdqaZbQvdi", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Adrian Perez-Salinas", "gender": "Male", "institution": "Leiden University", "country": "NL", "position": "Postdoc" }, { "name": "Casper Gyurik", "gender": "Male", "institution": "Leiden University", "country": "NL", "position": "Postdoc" }, { "name": "Elies Gil-Fuster", "gender": "Male", "institution": "Freie Universit\u00e4t Berlin", "country": "DE", "position": "PhD student" }, { "name": "Vedran Dunjko", "gender": "Male", "institution": "Leiden University", "country": "", "position": "Assistant Professor" } ], "n_formula": 19, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 110, "n_ref_all": 124, "n_fig": 3, "n_tab": 1, "L_tab": 1530, "n_element_tab": 35, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1177, "n_element_tab_1": 27, "formula_len_all": 1030, "formula_len_all_1": 258, "len_all": 204825, "len_all_1": 66687, "len_abs": 1372, "len_title": 144, "len_sents": 79098, "len_sents_1": 33477, "n_sents": 572, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1382, "title": "On the Relation between Trainability and Dequantization of Variational Quantum Learning Models", "abs": "Quantum machine learning (QML) explores the potential advantages of quantum computers for machine learning tasks, with variational QML among the main current approaches.\nWhile quantum computers promise to solve problems that are classically intractable, it has been recently shown that a particular quantum algorithm which outperforms all pre-existing classical algorithms can be matched by a newly developed classical approach (often inspired by the quantum algorithm).\nWe say such algorithms have been dequantized.\nFor QML models to be effective, they must be trainable and non-dequantizable.\nThe relationship between these properties is still not fully understood and recent works raised into question to what extent we could ever have QML models which are both trainable and non-dequantizable.\nThis challenges the potential of QML altogether.\nIn this work we answer open questions regarding when trainability and non-dequantization are compatible.\nWe first formalize the key concepts and put them in the context of prior research.\nWe introduce the role of \"variationalness\" of QML models using well-known quantum circuit architectures as leading examples.\nOur results provide recipes for variational QML models that are trainable and non-dequantizable.\nBy ensuring that variational QML models are both trainable and non-dequantizable, we pave the way toward practical relevance.", "keywords": [ "quantum machine learning", "machine learning theory", "quantum information theory" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "TdgAtxP6G2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Chao Tian", "gender": "Male", "institution": "Texas A&M", "country": "US", "position": "Associate Professor" }, { "name": "Ruida Zhou", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Suhas Diggavi", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "Professor" } ], "n_formula": 67, "n_formula_1": 17, "n_ref_uni": 42, "n_ref": 78, "n_ref_all": 101, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 201, "n_element_tab_1": 10, "formula_len_all": 4765, "formula_len_all_1": 1298, "len_all": 188921, "len_all_1": 76062, "len_abs": 2004, "len_title": 106, "len_sents": 56388, "len_sents_1": 32313, "n_sents": 499, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 2014, "title": "Transformers Learn Variable-order Markov Chains in-Context", "abs": "Large language models (LLMs) have demonstrated impressive in-context learning (ICL) capability. However, it is still unclear how the underlying transformers accomplish it, especially in more complex scenarios. Toward this goal, several recent works studied how transformers learn fixed-order Markov chains (FOMC) in context, yet natural languages are more suitably modeled by variable-order Markov chains (VOMC), i.e., context trees (CTs). In this work, we study the ICL of VOMC by viewing language modeling as a form of data compression and focusing on small alphabets and low-order VOMCs. This perspective allows us to leverage mature compression algorithms, such as context-tree weighting (CTW) and prediction by partial matching (PPM) algorithms as baselines, the former of which is Bayesian optimal for a class of priors that we refer to as the CTW priors. We empirically observe a few phenomena: 1) Transformers can indeed learn to compress VOMC in-context, while PPM suffers significantly; 2) The performance of transformers is not very sensitive to the number of layers, and even a two-layer transformer can learn in-context quite well; and 3) Transformers trained and tested on non-CTW priors can significantly outperform the CTW algorithm. To explain these phenomena, we analyze the attention map of the transformers and extract two mechanisms, on which we provide two transformer constructions: 1) A construction with $D+2$ layers that can mimic the CTW algorithm accurately for CTs of maximum order $D$, 2) A 2-layer transformer that utilizes the feed-forward network for probability blending. These constructions can explain most of the phenomena mentioned above. One distinction from the FOMC setting is that a counting mechanism appears to play an important role. We implement these synthetic transformer layers and show that such hybrid transformers can match the ICL performance of transformers, and more interestingly, some of them can perform even better despite the much-reduced parameter sets.", "keywords": [ "In-context learning; Variable-order Markov chain; Context Tree Weighting" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "TdIx7u2ECv", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Aleksandr Panov", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "Artem Latyshev", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Researcher" }, { "name": "Gregory Gorbov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 10, "n_ref_uni": 16, "n_ref": 49, "n_ref_all": 67, "n_fig": 8, "n_tab": 2, "L_tab": 618, "n_element_tab": 76, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 670, "n_element_tab_1": 77, "formula_len_all": 517, "formula_len_all_1": 820, "len_all": 78675, "len_all_1": 51732, "len_abs": 1121, "len_title": 112, "len_sents": 28285, "len_sents_1": 22804, "n_sents": 207, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1124, "title": "Imagine to Ensure Safety in Hierarchical Reinforcement Learning", "abs": "This work investigates the safe exploration problem, where an agent must maximize performance while satisfying safety constraints. To address this problem, we propose a method that includes a learnable world model and two policies, a high-level policy and a low-level policy, that ensure safety at both levels. The high-level policy generates safe subgoals for the low-level policy, which progressively guide the agent towards the final goal. Through trajectory imagination, the low-level policy learns to safely reach these subgoals. The proposed method was evaluated on the standard benchmark, SafetyGym, and demonstrated superior performance quality while maintaining comparable safety violations compared to state-of-the-art approaches. In addition, we investigated an alternative implementation of safety in hierarchical reinforcement learning (HRL) algorithms using Lagrange multipliers, and demonstrated in the custom long-horizon environments SafeAntMaze that our approach achieves comparable performance while more effectively satisfying safety constraints, while the flat safe policy fails to accomplish this task.", "keywords": [ "safe reinforcement learning", "machine learning", "model based reinforcement learning", "hierarchical reinforcement learning" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "TbTJJNjumY", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Fu Luo", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Qingfu Zhang", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Tong Xialiang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Xi Lin", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Zhenkun Wang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "YAOXIN WU", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 46, "n_ref": 113, "n_ref_all": 155, "n_fig": 4, "n_tab": 18, "L_tab": 6854, "n_element_tab": 922, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 269, "n_element_tab_1": 48, "formula_len_all": 1097, "formula_len_all_1": 565, "len_all": 202946, "len_all_1": 68048, "len_abs": 1187, "len_title": 131, "len_sents": 62250, "len_sents_1": 31104, "n_sents": 521, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1190, "title": "Boosting Neural Combinatorial Optimization for Large-Scale Vehicle Routing Problems", "abs": "Neural Combinatorial Optimization (NCO) methods have exhibited promising performance in solving Vehicle Routing Problems (VRPs). However, most NCO methods rely on the conventional self-attention mechanism that induces excessive computational complexity, thereby struggling to contend with large-scale VRPs and hindering their practical applicability. In this paper, we propose a lightweight cross-attention mechanism with linear complexity, by which a Transformer network is developed to learn efficient and favorable solutions for large-scale VRPs. We also propose a Self-Improved Training (SIT) algorithm that enables direct model training on large-scale VRP instances, bypassing extensive computational overhead for attaining labels. By iterating solution reconstruction, the Transformer network itself can generate improved partial solutions as pseudo-labels to guide the model training. Experimental results on the Travelling Salesman Problem (TSP) and the Capacitated Vehicle Routing Problem (CVRP) with up to 100K nodes indicate that our method consistently achieves superior performance for synthetic and real-world benchmarks, significantly boosting the scalability of NCO methods.", "keywords": [ "Neural Combinatorial Optimization", "Large-Scale Vehicle Routing Problem" ], "rating_list": [ 6, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "TbRFc2HPoN", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Jingyuan Wang", "gender": "Female", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Ruohan Zhan", "gender": "Female", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "Assistant Professor" }, { "name": "Zhengyuan Zhou", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Zhimei Ren", "gender": "Female", "institution": "The Wharton School, University of Pennsylvania", "country": "", "position": "Assistant Professor" } ], "n_formula": 184, "n_formula_1": 33, "n_ref_uni": 46, "n_ref": 91, "n_ref_all": 110, "n_fig": 1, "n_tab": 7, "L_tab": 910, "n_element_tab": 121, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 551, "n_element_tab_1": 74, "formula_len_all": 27704, "formula_len_all_1": 2936, "len_all": 218520, "len_all_1": 74037, "len_abs": 2048, "len_title": 108, "len_sents": 52783, "len_sents_1": 29903, "n_sents": 596, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1539, "title": "Distributionally Robust Policy Learning under Concept Drifts", "abs": "Distributionally robust policy learning aims to find a policy that performs well \n under the worst-case distributional shift, and yet most existing methods for \n robust policy learning consider the worst-case {\\em joint} distribution of \n the covariate and the outcome. The joint-modeling strategy can be unnecessarily conservative\n when we have more information on the source of distributional shifts. This paper studies\n a more nuanced problem --- robust policy learning under the \\emph{concept drift}, \n when only the conditional relationship between the outcome and the covariate changes. \n To this end, we first provide a doubly-robust estimator for evaluating\n the worst-case average reward of a given policy under a set of perturbed conditional distributions. \n We show that the policy value estimator enjoys asymptotic normality even if the nuisance parameters \n are estimated with a slower-than-root-$n$ rate.\n We then propose a learning algorithm that outputs the policy maximizing the \n estimated policy value within a given policy class $\\Pi$, and show\n that the sub-optimality gap of the proposed algorithm is of the order \n $\\kappa(\\Pi)n^{-1/2}$, with $\\kappa(\\Pi)$ is the entropy integral of $\\Pi$ under the Hamming distance\n and $n$ is the sample size. A matching lower bound is provided to show the optimality of the rate.\n The proposed methods are implemented and evaluated in numerical studies, \n demonstrating substantial improvement compared with existing benchmarks.", "keywords": [ "distributionally robust optimization", "offline policy learning", "concept drift", "bandit learning", "reinforcement learning." ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 1, 2 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "TbOcySs6g8", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Jerremy Holland", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Director of AI Research" }, { "name": "Nicholas Apostoloff", "gender": "unknown", "institution": "Apple", "country": "GB", "position": "Principal Researcher" }, { "name": "Oncel Tuzel", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Runkai Zheng", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Yinong Oliver Wang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Fernando Torre", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 6, "n_ref_uni": 18, "n_ref": 54, "n_ref_all": 67, "n_fig": 0, "n_tab": 3, "L_tab": 1284, "n_element_tab": 116, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 3830, "n_element_tab_1": 178, "formula_len_all": 633, "formula_len_all_1": 463, "len_all": 117518, "len_all_1": 64044, "len_abs": 2074, "len_title": 114, "len_sents": 37274, "len_sents_1": 28500, "n_sents": 269, "n_sents_1": 170, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1828, "title": "Advancing Differential Privacy through Synthetic Dataset Alignment", "abs": "Privacy in training data is crucial to protect sensitive personal information, prevent data misuse, and ensure compliance with legal regulations, all while maintaining trust and safeguarding individuals' rights in the development of ML models. \nUnfortunately, state-of-the-art methods that train ML models on image datasets with differential privacy constraints typically result in reduced accuracy due to noise. Alternatively, using synthetic data avoids the direct use of private data, preserving privacy, but suffers from domain discrepancies when compared to test data. This paper proposes a new methodology that combines both approaches by generating differentially private synthetic data closely aligned with the target domain, thereby improving the utility-privacy trade-off.\n\nOur approach begins with creating a synthetic base dataset using a class-conditional generative model. To address the domain gap between the synthetic dataset and the private dataset, we introduce the \\textbf{Privacy-Aware Synthetic Dataset Alignment \\text{(PASDA)}}, which leverages the feature statistics of the private dataset to guide the domain alignment process. PASDA produces a synthetic dataset that guarantees privacy while remaining highly functional for downstream training tasks.\nBuilding on this, we achieve state-of-the-art performance, surpassing the most competitive baseline by over 13\\% on CIFAR-10.\nFurthermore, our $(1,10^{-5})$-DP synthetic data achieves model performance on par with or surpassing models trained on the original STL-10, ImageNette and CelebA dataset. With zero-shot generation, our method does not require resource-intensive retraining, offering a synthetic data generation solution that introduces \\textbf{privacy} to a machine learning pipeline with both high \\textbf{efficiency} and \\textbf{efficacy}.", "keywords": [ "synthtic data", "differential privacy", "foundation models" ], "rating_list": [ 1, 3, 1, 5 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "TbJo3eQvHR", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Mengzhu Wang", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Nan Yin", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Zhiguang Cao", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Houcheng Su", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 56, "n_formula_1": 32, "n_ref_uni": 59, "n_ref": 121, "n_ref_all": 130, "n_fig": 3, "n_tab": 17, "L_tab": 14805, "n_element_tab": 1146, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 4909, "n_element_tab_1": 441, "formula_len_all": 5850, "formula_len_all_1": 2181, "len_all": 220005, "len_all_1": 72433, "len_abs": 1578, "len_title": 104, "len_sents": 49215, "len_sents_1": 27195, "n_sents": 415, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1663, "title": "Coupling Category Alignment for Graph Domain Adaptation", "abs": "Graph domain adaptation (GDA), which transfers knowledge from a labeled source domain to an unlabeled target graph domain, attracts considerable attention in numerous fields. \nEmerging methods commonly employ message-passing neural networks (MPNNs) to learn domain-invariant representations by aligning the entire domain distribution. However, these methods overlook the category-level distribution alignment across different domains, potentially leading to confusion of categories. \nTo address the problem, we propose an effective framework named \\textbf{Co}upling \\textbf{C}ateg{o}ry \\textbf{A}lignment (\\method{}) for GDA, which effectively addresses the category alignment issue with theoretical guarantees.\n\\method{} incorporates a graph convolutional network branch and a graph kernel network branch, which explore graph topology in implicit and explicit manners. To mitigate category-level domain shifts, we leverage knowledge from both branches, iteratively filtering highly reliable samples from the target domain using one branch and fine-tuning the other accordingly. Furthermore, with these reliable target domain samples, we incorporate the coupled branches into a holistic contrastive learning framework. This framework includes multi-view contrastive learning to ensure consistent representations across the dual branches, as well as cross-domain contrastive learning to achieve category-level domain consistency.\nTheoretically, we establish a sharper generalization bound, which ensures the effectiveness of category alignment.\nExtensive experiments on benchmark datasets validate the superiority of the proposed \\method{} compared with baselines.", "keywords": [ "Graph domain adaptation" ], "rating_list": [ 6, 5, 3, 6, 5 ], "soundness_list": [ 3, 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 5, 4 ] }, { "paper_id": "TbJJjwtBKX", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Haibing Guan", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Full Professor" }, { "name": "Hanxi Guo", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Hao Wang", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Tao Song", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tianhang Zheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiangyu Zhang", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Full Professor" }, { "name": "Yang Hua", "gender": "Male", "institution": "Queen's University Belfast", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 17, "n_ref_uni": 37, "n_ref": 131, "n_ref_all": 153, "n_fig": 6, "n_tab": 16, "L_tab": 13129, "n_element_tab": 996, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1796, "n_element_tab_1": 77, "formula_len_all": 25804, "formula_len_all_1": 26706, "len_all": 253651, "len_all_1": 90992, "len_abs": 1327, "len_title": 117, "len_sents": 68838, "len_sents_1": 29845, "n_sents": 491, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1348, "title": "Poisoning with A Pill: Circumventing Detection in Federated Learning", "abs": "Federated learning (FL) protects data privacy by enabling distributed model training without direct access to client data. However, its distributed nature makes it vulnerable to model and data poisoning attacks. While numerous defenses filter malicious clients using statistical metrics, they overlook the role of model redundancy, where not all parameters contribute equally to the model/attack performance. Current attacks manipulate all model parameters uniformly, making them more detectable, while defenses focus on the overall statistics of client updates, leaving gaps for more sophisticated attacks. We propose an attack-agnostic augmentation method to enhance the stealthiness and effectiveness of existing poisoning attacks in FL, exposing flaws in current defenses and highlighting the need for fine-grained FL security. Our three-stage methodology\u2014$\\textit{pill construction}$, $\\textit{pill poisoning}$, and $\\textit{pill injection}$\u2014injects poison into a compact subnet (i.e., pill) of the global model during the iterative FL training. Experimental results show that FL poisoning attacks enhanced by our method can bypass 8 state-of-the-art (SOTA) defenses, gaining an up to 7x error rate increase, as well as on average a more than 2x error rate increase on both IID and non-IID data, in both cross-silo and cross-device FL systems.", "keywords": [ "Federated Learning", "Byzantine Attack", "Model Poisoning" ], "rating_list": [ 8, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "Tb8RiXOc3N", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Matthijs T. J. Spaan", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "Full Professor" }, { "name": "Viliam Vadocz", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Wendelin Boehmer", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "Yaniv Oren", "gender": "Male", "institution": "Delft University of Technology", "country": "", "position": "PhD student" } ], "n_formula": 41, "n_formula_1": 27, "n_ref_uni": 41, "n_ref": 63, "n_ref_all": 88, "n_fig": 4, "n_tab": 6, "L_tab": 968, "n_element_tab": 112, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 356, "n_element_tab_1": 52, "formula_len_all": 3249, "formula_len_all_1": 2133, "len_all": 225537, "len_all_1": 71619, "len_abs": 1102, "len_title": 47, "len_sents": 61490, "len_sents_1": 31813, "n_sents": 419, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 33, "L_abs": 1108, "title": "Epistemic Monte Carlo Tree Search", "abs": "The AlphaZero/MuZero (A/MZ) family of algorithms has achieved remarkable success across various challenging domains by integrating Monte Carlo Tree Search (MCTS) with learned models. Learned models introduce epistemic uncertainty, which is caused by learning from limited data and is useful for exploration in sparse reward environments. MCTS does not account for the propagation of this uncertainty however. To address this, we introduce Epistemic MCTS (EMCTS): a theoretically motivated approach to account for the epistemic uncertainty in search and harness the search for deep exploration. In the challenging sparse-reward task of writing code in the Assembly language SUBLEQ, AZ paired with our method achieves significantly higher sample efficiency over baseline AZ. Search with EMCTS \nsolves variations of the commonly used hard-exploration benchmark Deep Sea - which baseline A/MZ are practically unable to solve - much faster than an otherwise equivalent method that does not use search for uncertainty estimation, demonstrating significant benefits from search for epistemic uncertainty estimation.", "keywords": [ "model based", "epistemic uncertainty", "exploration", "planning", "alphazero", "muzero" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 2, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 4, 2, 4 ] }, { "paper_id": "Tb5PY5vwp6", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Huaijin Wu", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Jihang Zhang", "gender": "unknown", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Lianqiang Li", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Minghui Yu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Tu Yi", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Huang Hantao", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 41, "n_ref": 86, "n_ref_all": 103, "n_fig": 8, "n_tab": 7, "L_tab": 3474, "n_element_tab": 468, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1801, "n_element_tab_1": 230, "formula_len_all": 582, "formula_len_all_1": 645, "len_all": 148166, "len_all_1": 60104, "len_abs": 1624, "len_title": 107, "len_sents": 38839, "len_sents_1": 27638, "n_sents": 273, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1692, "title": "HShare: Fast LLM Decoding by Hierarchical Key-Value Sharing", "abs": "The frequent retrieval of Key-Value (KV) cache data has emerged as a significant factor contributing to the inefficiency of the inference process in large language models. Previous research has demonstrated that a small subset of critical KV cache tokens largely influences attention outcomes, leading to methods that either employ fixed sparsity patterns or dynamically select critical tokens based on the query. While dynamic sparse patterns have proven to be more effective, they introduce significant computational overhead, as critical tokens must be reselected for each self-attention computation. In this paper, we reveal substantial similarities in KV cache token criticality across neighboring queries, layers, and heads. Motivated by this insight, we propose HShare, a hierarchical KV sharing framework. HShare facilitates the sharing of critical KV cache token indices across layers, heads, and queries, which significantly reduces the computational overhead associated with query-aware dynamic token sparsity. In addition, we introduce a greedy algorithm that dynamically determines the optimal layer-level and head-level sharing configuration for the decoding phase. We evaluate the effectiveness and efficiency of HShare across various tasks using three models: LLaMA2-7b, LLaMA3-70b, and Mistral-7b. Experimental results demonstrate that HShare achieves competitive accuracy with different sharing ratios, while delivering up to an $8.6\\times$ speedup in self-attention operations and a $2.7\\times$ improvement in end-to-end throughput compared with FlashAttention2 and GPT-fast respectively. The source code is publicly available at ~\\url{https://github.com/wuhuaijin/HShare}.", "keywords": [ "Large Language Model", "Decode", "Key-Value Sharing", "Critical Token", "Hierarchical" ], "rating_list": [ 8, 8, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 4, 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 4, 4 ] }, { "paper_id": "TadxJc1XAE", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Mir Moynuddin Ahmed Shibly", "gender": "unknown", "institution": "United International University", "country": "BD", "position": "Lecturer" }, { "name": "Nahid Hossain", "gender": "Male", "institution": "United International University", "country": "BD", "position": "Assistant Professor" }, { "name": "Sadia Islam", "gender": "Female", "institution": "United International University", "country": "BD", "position": "Lecturer" }, { "name": "Shoib Ahmed Shourav", "gender": "unknown", "institution": "United International University", "country": "BD", "position": "MS student" }, { "name": "Swakkhar Shatabda", "gender": "unknown", "institution": "BRAC University", "country": "BD", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 19, "n_ref": 23, "n_ref_all": 46, "n_fig": 10, "n_tab": 7, "L_tab": 1837, "n_element_tab": 287, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 2190, "n_element_tab_1": 131, "formula_len_all": 33, "formula_len_all_1": 33, "len_all": 85292, "len_all_1": 47627, "len_abs": 1517, "len_title": 137, "len_sents": 25154, "len_sents_1": 20963, "n_sents": 213, "n_sents_1": 152, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1522, "title": "TeacherActivityNet: A Novel Dataset for Monitoring Faculty Activities in Office Settings", "abs": "In this paper, we introduce a novel dataset for monitoring the activities of faculty members in academic office environments. Advances in computer vision have enabled the automation of workplace monitoring, particularly in educational institutions, where tracking faculty activities presents significant challenges and ethical considerations. Traditional methods of manual supervision are labor-intensive and prone to human error, underscoring the potential of automated video analysis as a more efficient solution. While substantial progress has been made in Human Activity Recognition (HAR) across various domains, research specifically focused on monitoring faculty activities in office settings is limited. Most existing studies concentrate on classroom and student monitoring, revealing a critical gap in faculty surveillance.\nThis paper seeks to address that gap by introducing TeacherActivityNet, a novel video dataset designed to recognize teachers' activities in academic offices, encompassing nine distinct action classes. We tweak the YOLOv8n architecture to propose our model, Teacher Activity Net (YOLOTAN), which is then fine-tuned using our dataset, achieving an average precision of 74.9\\%, significantly outperforming benchmark models. A comparative analysis of our dataset and methods against existing solutions highlights the potential of TeacherActivityNet to improve automated faculty monitoring systems. The dataset, trained models, and accompanying code are available at https://tinyurl.com/4ub94phh", "keywords": [ "Workplace Monitoring", "Activity Tracking", "YOLOTAN", "TeacherActivityNet", "Computer Vision" ], "rating_list": [ 3, 5, 1 ], "soundness_list": [ 2, 2, 1 ], "presentation_list": [ 1, 2, 1 ], "contribution_list": [ 2, 3, 1 ], "confidence_list": [ 5, 4, 5 ] }, { "paper_id": "TZa84ZkOLM", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Minji Lee", "gender": "Female", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Mohammed AlQuraishi", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" }, { "name": "Yeqing Lin", "gender": "Male", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Zhao Zhang", "gender": "unknown", "institution": "Rutgers University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 44, "n_ref": 73, "n_ref_all": 142, "n_fig": 42, "n_tab": 13, "L_tab": 2128, "n_element_tab": 361, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 163, "n_element_tab_1": 28, "formula_len_all": 1593, "formula_len_all_1": 1347, "len_all": 218341, "len_all_1": 60682, "len_abs": 1189, "len_title": 154, "len_sents": 69890, "len_sents_1": 29886, "n_sents": 491, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1194, "title": "Out of Many, One: Designing and Scaffolding Proteins at the Scale of the Structural Universe with Genie 2", "abs": "Protein diffusion models have emerged as a promising approach for protein design. One such pioneering model is Genie, a method that asymmetrically represents protein structures during the forward and backward processes, using simple Gaussian noising for the former and expressive SE(3)-equivariant attention for the latter. In this work we introduce Genie 2, extending Genie to capture a larger and more diverse protein structure space through architectural innovations and massive data augmentation. Genie 2 adds motif scaffolding capabilities via a novel multi-motif framework that designs co-occurring motifs with unspecified inter-motif positions and orientations. This makes possible complex protein designs that engage multiple interaction partners and perform multiple functions. On both unconditional and conditional generation, Genie 2 achieves state-of-the-art performance, outperforming all known methods on key design metrics including designability, diversity, and novelty. Genie 2 also solves more motif scaffolding problems than other methods and does so with more unique and varied solutions. Taken together, these advances set a new standard for structure-based protein design.", "keywords": [ "protein design", "diffusion model", "motif scaffolding", "protein structure", "generative model", "biology", "bioinformatics", "structural biology" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "TYnne6Pa35", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Tianshu Yu", "gender": "Male", "institution": "Chinese University of Hong Kong (Shenzhen)", "country": "CN", "position": "Assistant Professor" }, { "name": "Weihuang Wen", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 17, "n_ref_uni": 27, "n_ref": 59, "n_ref_all": 75, "n_fig": 3, "n_tab": 2, "L_tab": 2038, "n_element_tab": 287, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1911, "n_element_tab_1": 313, "formula_len_all": 16788, "formula_len_all_1": 16721, "len_all": 144511, "len_all_1": 87292, "len_abs": 1361, "len_title": 128, "len_sents": 41777, "len_sents_1": 30814, "n_sents": 296, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1377, "title": "HyperPLR: Hypergraph Generation through Projection, Learning, and Reconstruction", "abs": "Hypergraphs are essential in modeling higher-order complex networks, excelling in representing group interactions within real-world contexts. This is particularly evident in collaboration networks, where they facilitate the capture of groupwise polyadic patterns, extending beyond traditional pairwise dyadic interactions. The use of hypergraph generators, or generative models, is a crucial method for promoting and validating our understanding of these structures. If such generators accurately replicate observed hypergraph patterns, it reinforces the validity of our interpretations. In this context, we introduce a novel hypergraph generative paradigm, \\textbf{HyperPLR}, encompassing three phases: Projection, Learning, and Reconstruction. Initially, the hypergraph is projected onto a weighted graph. Subsequently, the model learns this graph's structure within a latent space, while simultaneously computing a distribution between the hyperedge and the projected graph. Finally, leveraging the learned model and distribution, HyperPLR generates new weighted graphs and samples cliques from them. These cliques are then used to reconstruct new hypergraphs by solving a specific clique cover problem.\nWe have evaluated HyperPLR on existing real-world hypergraph datasets, which consistently demonstrate superior performance and validate the effectiveness of our approach.", "keywords": [ "hypergraph", "graph generation", "clique cover problem" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "TYSQYx9vwd", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Felix Opolka", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Pietro Lio", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Richard Bergna", "gender": "Not Specified", "institution": "University of Cambridge", "country": "", "position": "MS student" }, { "name": "Sergio Calvo Ordo\u00f1ez", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Jose Miguel Hernandez", "gender": "unknown", "institution": "Microsoft Research Cambridge", "country": "GB", "position": "Research Scientist" } ], "n_formula": 31, "n_formula_1": 13, "n_ref_uni": 24, "n_ref": 39, "n_ref_all": 59, "n_fig": 3, "n_tab": 11, "L_tab": 6886, "n_element_tab": 376, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 4365, "n_element_tab_1": 183, "formula_len_all": 2168, "formula_len_all_1": 717, "len_all": 112072, "len_all_1": 61984, "len_abs": 1200, "len_title": 132, "len_sents": 32709, "len_sents_1": 25665, "n_sents": 265, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 83, "L_abs": 1204, "title": "Uncertainty Modeling in Graph Neural Networks via Stochastic Differential Equations", "abs": "We propose a novel Stochastic Differential Equation (SDE) framework to address the problem of learning uncertainty-aware representations for graph-structured data. While Graph Neural Ordinary Differential Equations (GNODEs) have shown promise in learning node representations, they lack the ability to quantify uncertainty. To address this, we introduce Latent Graph Neural Stochastic Differential Equations (LGNSDE), which enhance GNODE by embedding randomness through a Bayesian prior-posterior mechanism for epistemic uncertainty and Brownian motion for aleatoric uncertainty. By leveraging the existence and uniqueness of solutions to graph-based SDEs, we prove that the variance of the latent space bounds the variance of model outputs, thereby providing theoretically sensible guarantees for the uncertainty estimates. Furthermore, we show mathematically that LGNSDEs are robust to small perturbations in the input, maintaining stability over time. Empirical results across several benchmarks demonstrate that our framework is competitive in out-of-distribution detection, robustness to noise perturbations, and active learning, underscoring the ability of LGNSDEs to quantify uncertainty reliably.", "keywords": [ "Graph Neural Networks", "Stochastic Differential Equations", "Uncertainty Quantification", "Bayesian Machine Learning" ], "rating_list": [ 6, 6, 10 ], "soundness_list": [ 2, 2, 4 ], "presentation_list": [ 3, 2, 4 ], "contribution_list": [ 2, 2, 4 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "TY9mstpD02", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Emily B. Fox", "gender": "Female", "institution": "Inistro Inc.", "country": "US", "position": "Principal Researcher" }, { "name": "Noah Goodman", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Associate Professor" }, { "name": "Vivek Vajipey", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Michael Yifan Li", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 4, "n_ref_uni": 29, "n_ref": 43, "n_ref_all": 59, "n_fig": 9, "n_tab": 9, "L_tab": 3353, "n_element_tab": 206, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 216, "n_element_tab_1": 31, "formula_len_all": 122, "formula_len_all_1": 514, "len_all": 122539, "len_all_1": 60110, "len_abs": 1744, "len_title": 99, "len_sents": 41665, "len_sents_1": 28028, "n_sents": 315, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 1757, "title": "CALM: Critic Automation with Language Models", "abs": "Understanding the world through models is a fundamental goal of scientific research.\nWhile large language model (LLM) based approaches show promise in automating scientific discovery, they often overlook the importance of criticizing scientific models.\nCriticizing models deepens scientific understanding and drives the development of more accurate models.\nMoreover, criticism can improve the reliability of LLM-based scientist systems by acting as a safeguard against hallucinations.\nAutomating model criticism is difficult because it traditionally requires a human expert to define how to compare a model with data and evaluate if the discrepancies are significant--both rely heavily on understanding the modeling assumptions and domain.\nAlthough LLM-based critic approaches are appealing, they introduce new challenges: LLMs might hallucinate the critiques themselves. \nMotivated by this, we introduce CALM (Critic Automation with Language Models). CALM uses LLMs to generate summary statistics that highlight discrepancies between model predictions and data, and applies hypothesis tests to evaluate their significance.\nWe can view CALM as a verifier that validates models and critiques by embedding them in a hypothesis testing framework. \nIn experiments, we evaluate CALM across key quantitative and qualitative dimensions.\nIn settings where we synthesize discrepancies between models and datasets, CALM reliably generates correct critiques without hallucinating incorrect ones.\nWe show that both human and LLM judges consistently prefer CALM\u2019s critiques over alternative approaches in terms of transparency and actionability.\nFinally, we show that CALM's critiques enable an LLM scientist to improve upon human-designed models on real-world datasets.", "keywords": [ "automatic scientific discovery", "model criticism" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 1, 3, 1 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "TXjYOslkUh", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Chandan K. Reddy", "gender": "Male", "institution": "Virginia Tech", "country": "US", "position": "Full Professor" }, { "name": "Dongjie Wang", "gender": "Male", "institution": "University of Kansas", "country": "US", "position": "Assistant Professor" }, { "name": "Nanxu Gong", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Pengyang Wang", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Assistant Professor" }, { "name": "Wangyang Ying", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Xinyuan Wang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Yanjie Fu", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 33, "n_ref": 47, "n_ref_all": 59, "n_fig": 4, "n_tab": 8, "L_tab": 2754, "n_element_tab": 305, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 287, "n_element_tab_1": 63, "formula_len_all": 818, "formula_len_all_1": 535, "len_all": 143304, "len_all_1": 73277, "len_abs": 2052, "len_title": 102, "len_sents": 52700, "len_sents_1": 36377, "n_sents": 365, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 2008, "title": "Distribution Shift Aware Neural Feature Transformation", "abs": "Feature transformation, as a core task of Data-centric AI (DCAI), aims to improve the original feature set to enhance AI capabilities. In dynamic real-world environments, where there exists a distribution shift, feature knowledge may not be transferable between data. This matter prompts a distribution shift feature transformation (DSFT) problem. Prior research works for feature transformation either depend on domain expertise, rely on a linear assumption, prove inefficient for large feature spaces, or demonstrate vulnerability to imperfect data. Furthermore, existing techniques for addressing the distribution shift cannot be directly applied to discrete search problems. DSFT presents two primary challenges: 1) How can we reformulate and solve feature transformation as a learning problem? and 2) What mechanisms can integrate shift awareness into such a learning paradigm? To tackle these challenges, we leverage a unique Shift-aware Representation-Generation Perspective. To formulate a learning scheme, we construct a representation-generation framework: 1) representation step: encoding transformed feature sets into embedding vectors; 2) generation step: pinpointing the best embedding and decoding as a transformed feature set. To mitigate the issue of distribution shift, we propose three mechanisms: 1) shift-resistant representation, where embedding dimension decorrelation and sample reweighing are integrated to extract the true representation that contains invariant information under distribution shift; 2) flatness-aware generation, where several suboptimal embeddings along the optimization trajectory are averaged to obtain a robust optimal embedding, proving effective for diverse distribution; and 3) shift-aligned pre and post-processing, where normalizing and denormalizing align and recover distribution gaps between training and testing data. Ultimately, extensive experiments are conducted to indicate the effectiveness, robustness, and trackability of our proposed framework.", "keywords": [ "Feature Transformation", "Out-of-Distribution (OOD)", "Data-centric AI" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "TXfzH933qV", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Chen Ning", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Ji Wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xiao Zhang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xien Liu", "gender": "Male", "institution": " Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Yuxuan Zhou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 21, "n_ref": 41, "n_ref_all": 69, "n_fig": 13, "n_tab": 12, "L_tab": 6657, "n_element_tab": 376, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1626, "n_element_tab_1": 110, "formula_len_all": 347, "formula_len_all_1": 371, "len_all": 138282, "len_all_1": 58821, "len_abs": 1674, "len_title": 113, "len_sents": 42382, "len_sents_1": 27171, "n_sents": 306, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1681, "title": "Reliable and Diverse Evaluation of LLM Medical Knowledge Mastery", "abs": "Mastering medical knowledge is crucial for medical-specific LLMs. However, despite the existence of medical benchmarks like MedQA, a unified framework that fully leverages existing knowledge bases to evaluate LLMs' mastery of medical knowledge is still lacking. In the study, we propose a novel framework PretexEval that dynamically generates reliable and diverse test samples to evaluate LLMs for any given medical knowledge base. We notice that test samples produced directly from knowledge bases by templates or LLMs may introduce factual errors and also lack diversity. To address these issues, we introduce a novel schema into our proposed evaluation framework that employs predicate equivalence transformations to produce a series of variants for any given medical knowledge point. Finally, these produced predicate variants are converted into textual language, resulting in a series of reliable and diverse test samples to evaluate whether LLMs fully master the given medical factual knowledge point. Here, we use our proposed framework to systematically investigate the mastery of medical factual knowledge of 12 well-known LLMs, based on two knowledge bases that are crucial for clinical diagnosis and treatment. The evaluation results illustrate that current LLMs still exhibit significant deficiencies in fully mastering medical knowledge, despite achieving considerable success on some famous public benchmarks. These new findings provide valuable insights for developing medical-specific LLMs, highlighting that current LLMs urgently need to strengthen their comprehensive and in-depth mastery of medical knowledge before being applied to real-world medical scenarios.", "keywords": [ "LLM Evaluation", "Medical Evaluation", "Large Language Model" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "TXeqjGYxu4", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Hongteng Xu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Shen Yuan", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 34, "n_formula_1": 25, "n_ref_uni": 50, "n_ref": 92, "n_ref_all": 106, "n_fig": 4, "n_tab": 3, "L_tab": 1182, "n_element_tab": 106, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1182, "n_element_tab_1": 106, "formula_len_all": 11376, "formula_len_all_1": 10326, "len_all": 153215, "len_all_1": 82010, "len_abs": 1211, "len_title": 119, "len_sents": 31598, "len_sents_1": 26619, "n_sents": 227, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1222, "title": "Towards Better Multi-head Attention via Channel-wise Sample Permutation", "abs": "Transformer plays a central role in many fundamental deep learning models, e.g., the ViT in computer vision and the BERT and GPT in natural language processing, whose effectiveness is mainly attributed to its multi-head attention (MHA) mechanism. \nIn this study, we propose a simple and novel channel-wise sample permutation (CSP) operator, achieving a new structured MHA with fewer parameters and lower complexity. \nGiven an input matrix, CSP circularly shifts the samples of different channels with various steps and then sorts grouped samples of each channel. \nThis operator is equivalent to implicitly implementing cross-channel attention maps as permutation matrices, which achieves linear complexity and suppresses the risk of rank collapse when representing data. \nWe replace the MHA of some representative models with CSP and test the CSP-based models in several discriminative tasks, including image classification and long sequence analysis. \nExperiments show that the CSP-based models achieve comparable or better performance with fewer parameters and lower computational costs than the classic Transformer and its state-of-the-art variants. \nThe code is available at https://anonymous.4open.science/r/CSP-BA52.", "keywords": [ "Efficient Transformers", "sparse doubly stochastic attention", "permutation", "optimal transport" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 2, 2 ] }, { "paper_id": "TX0OsLcaWf", "primary_area": "generative models", "n_author": 12, "author_info_list": [ { "name": "Feng Liu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Hangjie Yuan", "gender": "Male", "institution": "DAMO Academy, Alibaba Group", "country": "CN", "position": "Research Scientist" }, { "name": "Haonan Qiu", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Hongming Shan", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Jiaxin Ye", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Rui Zhao", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Shiwei Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xiang Wang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yingya Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yujie Wei", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Yutong Feng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zhizhong Huang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Postdoc" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 46, "n_ref": 158, "n_ref_all": 184, "n_fig": 14, "n_tab": 8, "L_tab": 1483, "n_element_tab": 230, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 576, "n_element_tab_1": 184, "formula_len_all": 7193, "formula_len_all_1": 6099, "len_all": 221110, "len_all_1": 71200, "len_abs": 445, "len_title": 121, "len_sents": 47547, "len_sents_1": 29905, "n_sents": 350, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1697, "title": "Zero-Shot Subject-Driven Video Customization with Precise Motion Control", "abs": "Recent advances in customized video generation have enabled users to create videos tailored to both specific subjects and motion trajectories. However, existing methods often require complicated test-time fine-tuning and struggle with balancing subject learning and motion control, limiting their real-world applications. In this paper, we present $\\textbf{DreamCustomizer}$, a zero-shot video customization framework capable of generating videos with a specific subject and motion trajectory, guided by a single image and a bounding box sequence, respectively, and without the need for test-time fine-tuning. Specifically, we introduce reference attention, which leverages the model\u2019s inherent capabilities for subject learning, and devise a mask-guided motion module to achieve precise motion control by fully utilizing the robust motion signal of box masks derived from bounding boxes. While these two components achieve their intended functions, we empirically observe that motion control tends to dominate over subject learning. To address this, we propose two key designs: $\\textbf{1)}$ the masked reference attention, which integrates a blended latent mask modeling scheme into reference attention to enhance subject representations at the desired positions, and $\\textbf{2)}$ a reweighted diffusion loss, which differentiates the contributions of regions inside and outside the bounding boxes to ensure a balance between subject and motion control. Extensive experimental results on a newly curated dataset demonstrate that DreamCustomizer outperforms state-of-the-art methods in both subject customization and motion control. The dataset, code, and models will be made publicly available.", "keywords": [ "Video Generation", "Customized Generation" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 4, 1, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "TWnUgSAWNw", "primary_area": "foundation or frontier models, including LLMs", "n_author": 12, "author_info_list": [ { "name": "Bowen Zhang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Research Scientist" }, { "name": "Chen Chen", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Haotian Zhang", "gender": "Male", "institution": "Apple AI/ML", "country": "US", "position": "Researcher" }, { "name": "Hong-You Chen", "gender": "unknown", "institution": "Apple AI/ML", "country": "US", "position": "Researcher" }, { "name": "Juan Lao Tebar", "gender": "Male", "institution": "Apple", "country": "CH", "position": "Researcher" }, { "name": "Meng Cao", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Peter Grasch", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Vasileios Saveris", "gender": "unknown", "institution": "Apple", "country": "CH", "position": "Researcher" }, { "name": "Wenze Hu", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Yinfei Yang", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Zhe Gan", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Zhengfeng Lai", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 111, "n_ref_all": 147, "n_fig": 11, "n_tab": 8, "L_tab": 3625, "n_element_tab": 395, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1755, "n_element_tab_1": 270, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 161199, "len_all_1": 58079, "len_abs": 1426, "len_title": 131, "len_sents": 43594, "len_sents_1": 26546, "n_sents": 301, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1431, "title": "Revisit Large-Scale Image-Caption Data in Pre-training Multimodal Foundation Models", "abs": "Recent advancements in multimodal models highlight the value of rewritten captions for improving performance, yet key challenges remain. For example, while synthetic captions often provide superior quality and image-text alignment, it is not clear whether they can fully replace AltTexts: the role of synthetic captions and their interaction with original web-crawled AltTexts in pre-training is still not well understood. Moreover, different multimodal foundation models may have unique preferences for specific caption formats, but efforts to identify the optimal captions for each model remain limited. In this work, we propose a novel, controllable, and scalable captioning pipeline designed to generate diverse caption formats tailored to various multimodal models. By examining short synthetic captions (SSC) and descriptive synthetic captions (DSC) as case studies, we systematically explore their effects and interactions with AltTexts across models such as CLIP, multimodal LLMs, and diffusion models. Our findings reveal that a hybrid approach that keeps both synthetic captions and AltTexts can outperform the use of synthetic captions alone, improving both alignment and performance, with each model demonstrating preferences for particular caption formats. This comprehensive analysis provides valuable insights into optimizing captioning strategies, thereby advancing the pre-training of multimodal foundation models.", "keywords": [ "Multimodal", "pre-training", "image-text data" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "TWmETQLZIC", "primary_area": "generative models", "n_author": 1, "author_info_list": [ { "name": "Calvin McCarter", "gender": "Not Specified", "institution": "Bighat Biosciences", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 2, "n_ref_uni": 54, "n_ref": 97, "n_ref_all": 110, "n_fig": 4, "n_tab": 5, "L_tab": 621, "n_element_tab": 54, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1169, "n_element_tab_1": 60, "formula_len_all": 790, "formula_len_all_1": 132, "len_all": 123472, "len_all_1": 61155, "len_abs": 1138, "len_title": 80, "len_sents": 36584, "len_sents_1": 27270, "n_sents": 221, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 32, "L_abs": 1141, "title": "Unmasking Trees for Tabular Data", "abs": "Despite much work on advanced deep learning and generative modeling techniques for tabular data generation and imputation, traditional methods have continued to win on imputation benchmarks. We herein present UnmaskingTrees, a simple method for tabular imputation (and generation) employing gradient-boosted decision trees which are used to incrementally unmask individual features. This approach offers state-of-the-art performance on imputation, and on generation given training data with missingness; and it has competitive performance on vanilla generation. To solve the conditional generation subproblem, we propose a tabular probabilistic prediction method, BaltoBot, which fits a balanced tree of boosted tree classifiers. Unlike older methods, it requires no parametric assumption on the conditional distribution, accommodating features with multimodal distributions; unlike newer diffusion methods, it offers fast sampling, closed-form density estimation, and flexible handling of discrete variables. We finally consider our two approaches as meta-algorithms, demonstrating in-context learning-based generative modeling with TabPFN.", "keywords": [ "tabular data", "imputation", "missing", "generative modeling", "tabular", "probabilistic prediction", "tabular ML" ], "rating_list": [ 6, 3, 3, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "TWjNSzk7mP", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Jianbo Yu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Postdoc" }, { "name": "Jing Yan", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Shuai Chen", "gender": "Female", "institution": "Alipay", "country": "", "position": "Researcher" }, { "name": "Siyang Leng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Young Principal Investigator" }, { "name": "Zeqi Ren", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Zhu Jiaming", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" }, { "name": "zhujingzhe", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" }, { "name": "Tan Xiao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 25, "n_ref": 39, "n_ref_all": 51, "n_fig": 3, "n_tab": 4, "L_tab": 903, "n_element_tab": 86, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 885, "n_element_tab_1": 93, "formula_len_all": 1591, "formula_len_all_1": 1590, "len_all": 114115, "len_all_1": 53819, "len_abs": 1909, "len_title": 114, "len_sents": 25086, "len_sents_1": 23542, "n_sents": 193, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1561, "title": "Boosting In-Context Learning in LLMs with Retrieval-based Codebook", "abs": "Recent advancements in large language models (LLMs) have demonstrated exceptional performance across various downstream tasks, particularly due to their in-context learning (ICL) abilities. ICL enables models to learn from a few demonstrations presented in the context, without requiring retraining or fine-tuning. However, the effectiveness of ICL is highly dependent on factors such as prompt design and input length. To address these limitations, we propose a novel approach that leverages the key-value pairs within Transformers to enhance contextual understanding in LLMs. Specifically, our method converts raw demonstrations into task vectors\u2014comprising keys and values\u2014which are derived through multiple passes of the LLM, then integrated with test task vectors to improve model comprehension of the input. Furthermore, we introduce a retrieval-based codebook mechanism that captures information from long-context demonstrations while filtering irrelevant content. This codebook dynamically stores and updates task vectors generated during inference, mitigating input length constraints and optimizing the relevance of contextual data. By retrieving the most pertinent historical task vectors, the codebook ensures that only relevant information is utilized during inference. Extensive experiments show that these enhancements significantly outperform conventional ICL, achieving superior accuracy and efficiency. Overall, this work sets a new benchmark for optimizing ICL in LLMs, enabling their effective deployment in complex, real-world applications.", "keywords": [ "large language model", "in-context learning", "retrieval", "codebook" ], "rating_list": [ 5, 3, 8, 3 ], "soundness_list": [ 2, 2, 4, 1 ], "presentation_list": [ 2, 1, 3, 1 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 5, 4, 5 ] }, { "paper_id": "TWRhLAN5rz", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Siwei Li", "gender": "unknown", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Xinghao Chen", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Yijing Yang", "gender": "Female", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Yunhe Wang", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 42, "n_ref": 109, "n_ref_all": 128, "n_fig": 4, "n_tab": 7, "L_tab": 1197, "n_element_tab": 140, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 812, "n_element_tab_1": 127, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 148300, "len_all_1": 77524, "len_abs": 1559, "len_title": 127, "len_sents": 35181, "len_sents_1": 31573, "n_sents": 255, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1581, "title": "Unleashing the Potential of ConvNets for Query-based Detection and Segmentation", "abs": "Transformer and its variants have shown great potential for various vision tasks in recent years, including image classification, object detection and segmentation. \nMeanwhile, recent studies also reveal that with proper architecture design, convolution networks (ConvNets) also achieve competitive performance with transformers, \\eg, ConvNeXt. \nHowever, no prior methods have explored to utilize pure convolution to build a Transformer-style Decoder module, which is essential for Encoder-Decoder architecture like Detection Transformer (DETR).\nTo this end, in this paper we explore whether we could build query-based detection and segmentation framework with ConvNets instead of sophisticated transformer architecture.\nWe propose a novel mechanism dubbed InterConv to perform interaction between object queries and image features via convolutional layers. \nEquipped with the proposed InterConv, we build Detection ConvNet (DECO), which is composed of a backbone and convolutional encoder-decoder architecture. We compare the proposed DECO against prior detectors on the challenging COCO benchmark.\nDespite its simplicity, our DECO achieves competitive performance in terms of detection accuracy and running speed. Specifically, \nwith the ResNet-18 and ResNet-50 backbone, our DECO achieves $40.5\\%$ and $47.8\\%$ AP with $66$ and $34$ FPS, respectively. The proposed method is also evaluated on the segment anything task, demonstrating similar performance and higher efficiency.\nWe hope the proposed method brings another perspective for designing architectures for vision tasks.", "keywords": [ "Object Detection", "Segmentation", "ConvNet", "Segment Anything" ], "rating_list": [ 6, 5, 8, 8 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "TVwD2zIQ1F", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Debarghya Ghoshdastidar", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Assistant Professor" }, { "name": "Lukas Gosch", "gender": "Male", "institution": "Technical University of Munich", "country": "DE", "position": "PhD student" }, { "name": "Mahalakshmi Sabanayagam", "gender": "Female", "institution": "Department of Informatics, Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Stephan G\u00fcnnemann", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Professor" } ], "n_formula": 80, "n_formula_1": 18, "n_ref_uni": 51, "n_ref": 134, "n_ref_all": 237, "n_fig": 32, "n_tab": 4, "L_tab": 1920, "n_element_tab": 186, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9329, "formula_len_all_1": 2108, "len_all": 297028, "len_all_1": 82256, "len_abs": 1123, "len_title": 132, "len_sents": 87057, "len_sents_1": 34641, "n_sents": 756, "n_sents_1": 273, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1126, "title": "Provable Robustness of (Graph) Neural Networks Against Data Poisoning and Backdoors", "abs": "Generalization of machine learning models can be severely compromised by data poisoning, where adversarial changes are applied to the training data. This vulnerability has led to interest in certifying (i.e., proving) that such changes up to a certain magnitude do not affect test predictions. We, for the first time, certify Graph Neural Networks (GNNs) against poisoning attacks, including backdoors, targeting the node features of a given graph. Our certificates are white-box and based upon (i) the neural tangent kernel, which characterizes the training dynamics of sufficiently wide networks; and (ii) a novel reformulation of the bilevel optimization problem describing poisoning as a mixed-integer linear program. Consequently, we leverage our framework to provide fundamental insights into the role of graph structure and its connectivity on the worst-case robustness behavior of convolution-based and PageRank-based GNNs. We note that our framework is more general and constitutes the first approach to derive white-box poisoning certificates for NNs, which can be of independent interest beyond graph-related tasks.", "keywords": [ "graph neural networks", "provable robustness", "certificates", "poisoning", "data poisoning", "backdoor attacks", "neural tangent kernel", "adversarial robustness", "mixed-integer linear programming", "support vector machines" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "TVnkjz4MqV", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Chenggang Chen", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Postdoc" }, { "name": "Dora Zhiyu Yang", "gender": "Not Specified", "institution": "Mathematica Policy Research", "country": "US", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 6, "n_ref_uni": 45, "n_ref": 76, "n_ref_all": 115, "n_fig": 26, "n_tab": 3, "L_tab": 1287, "n_element_tab": 18, "n_fig_1": 23, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 991, "formula_len_all_1": 239, "len_all": 198595, "len_all_1": 64046, "len_abs": 2511, "len_title": 159, "len_sents": 68169, "len_sents_1": 32476, "n_sents": 566, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 110, "L_abs": 1151, "title": "Neural Manifold Regularization: Aligning 2D Latent Dynamics with Stereotyped, Natural, and Attempted Movements", "abs": "Mapping neural activity to behavior is a fundamental goal in both neuroscience and brain-machine interfaces. Traditionally, at least three-dimensional (3D) latent dynamics have been required to represent two-dimensional (2D) movement trajectories. In this work, we introduce Neural Manifold Regularization (NMR), a method that embeds neural dynamics into a 2D latent space and regularizes the manifold based on the distances and densities of continuous movement labels. NMR pulls together positive pairs of neural embeddings (corresponding to closer labels) and pushes apart negative pairs (representing more distant labels). Additionally, NMR applies greater force to infrequent labels to prevent them from collapsing into dominant labels.\nWe evaluated NMR across four modalities of neural signals and three types of movements. When combined with a linear regression decoder, NMR outperformed other dimensionality reduction methods by over 50\\% across 68 sessions. The highly consistent neural manifolds extracted by NMR enable robust motor decoding across sessions, years, and subjects using a simple linear regression decoder.\nOur code is uploaded.", "keywords": [ "dimensionality reduction", "brain-machine interfaces", "motor control", "neural coding", "self-supervised learning" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "TVQLu34bdw", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 11, "author_info_list": [ { "name": "Arash Vahdat", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Christian Dallago", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Assistant Professor" }, { "name": "Danny Reidenbach", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Emine Kucukbenli", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Senior Research Scientist" }, { "name": "Jason Yim", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Karsten Kreis", "gender": "unknown", "institution": "NVIDIA", "country": "CA", "position": "Research Scientist" }, { "name": "Kieran Didi", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "MS student" }, { "name": "Mario Geiger", "gender": "Male", "institution": "NVIDIA", "country": "CH", "position": "Researcher" }, { "name": "Tomas Geffner", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Zhonglin Cao", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Applied Scientist" }, { "name": "Zuobai Zhang", "gender": "Male", "institution": "Mila - Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" } ], "n_formula": 57, "n_formula_1": 15, "n_ref_uni": 60, "n_ref": 203, "n_ref_all": 251, "n_fig": 20, "n_tab": 13, "L_tab": 6119, "n_element_tab": 680, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 3504, "n_element_tab_1": 206, "formula_len_all": 26674, "formula_len_all_1": 12877, "len_all": 427093, "len_all_1": 106068, "len_abs": 1284, "len_title": 113, "len_sents": 123092, "len_sents_1": 39079, "n_sents": 1011, "n_sents_1": 304, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 64, "L_abs": 1293, "title": "Proteina: Scaling Flow-based Protein Structure Generative Models", "abs": "Recently, diffusion- and flow-based generative models of protein structures have emerged as a powerful tool for de novo protein design. Here, we develop Proteina, a new large-scale flow-based protein backbone generator that utilizes hierarchical fold class labels for conditioning and relies on a tailored scalable transformer architecture with up to $5\\times$ as many parameters as previous models. To meaningfully quantify performance, we introduce a new set of metrics that directly measure the distributional similarity of generated proteins with reference sets, complementing existing metrics. We further explore scaling training data to millions of synthetic protein structures and explore improved training and sampling recipes adapted to protein backbone generation. This includes fine-tuning strategies like LoRA for protein backbones, new guidance methods like classifier-free guidance and autoguidance for protein backbones, and new adjusted training objectives. Proteina achieves state-of-the-art performance on de novo protein backbone design and produces diverse and designable proteins at unprecedented length, up to 800 residues. The hierarchical conditioning offers novel control, enabling high-level secondary-structure guidance as well as low-level fold-specific generation.", "keywords": [ "protein structure generation", "de novo protein design", "flow matching", "fold class conditioning" ], "rating_list": [ 8, 6, 5, 6, 8 ], "soundness_list": [ 4, 3, 4, 3, 4 ], "presentation_list": [ 4, 4, 4, 3, 4 ], "contribution_list": [ 4, 2, 3, 3, 4 ], "confidence_list": [ 4, 5, 3, 4, 3 ] }, { "paper_id": "TVFVx8TUbN", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Dong HUANG", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Guangtao Zeng", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "", "position": "PhD student" }, { "name": "Haochen Tan", "gender": "Male", "institution": "City University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Jianbo Dai", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "MS student" }, { "name": "Jianqiao Lu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Ming Cheng", "gender": "Female", "institution": "University of Edinburgh", "country": "", "position": "MS student" }, { "name": "Yunlong Feng", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Zhijiang Guo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "rongju ruan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 139, "n_ref_all": 182, "n_fig": 23, "n_tab": 4, "L_tab": 3684, "n_element_tab": 505, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 836, "n_element_tab_1": 94, "formula_len_all": 684, "formula_len_all_1": 0, "len_all": 302433, "len_all_1": 58351, "len_abs": 2398, "len_title": 140, "len_sents": 68206, "len_sents_1": 27875, "n_sents": 467, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1223, "title": "MHPP: Exploring the Capabilities and Limitations of Language Models Beyond Basic Code Generation", "abs": "Recent advancements in large language models (LLMs) have greatly improved code generation, specifically at the function level. For instance, GPT-4o has achieved a 91.0\\% pass rate on HumanEval. However, this draws into question the adequacy of existing benchmarks in thoroughly assessing function-level code generation capabilities. Our study analyzed two common benchmarks, HumanEval and MBPP, and found that these might not thoroughly evaluate LLMs' code generation capacities due to limitations in quality, difficulty, and granularity. To resolve this, we introduce the Mostly Hard Python Problems (MHPP) dataset, consisting of 210 unique human-curated problems. By focusing on the combination of natural language and code reasoning, MHPP gauges LLMs' abilities to comprehend specifications and restrictions, engage in multi-step reasoning, and apply coding knowledge effectively. Initial evaluations of 26 LLMs using MHPP showed many high-performing models on HumanEval failed to achieve similar success on MHPP. Moreover, MHPP highlighted various previously undiscovered limitations within various LLMs, leading us to believe that it could pave the way for a better understanding of LLMs' capabilities and limitations.", "keywords": [ "Large Langugae Models", "Code Generation", "Program Synthesis", "Benchmark and Dataset", "Evaluation and Resources" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "TUvg5uwdeG", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Christian Wald", "gender": "Male", "institution": "Technische Universit\u00e4t Berlin", "country": "", "position": "Postdoc" }, { "name": "Gabriele Steidl", "gender": "Female", "institution": "Technische Universit\u00e4t Berlin", "country": "DE", "position": "Full Professor" }, { "name": "Jannis Chemseddine", "gender": "unknown", "institution": "Technische Universit\u00e4t Berlin", "country": "DE", "position": "PhD student" }, { "name": "Richard Duong", "gender": "unknown", "institution": "Technische Universit\u00e4t Berlin", "country": "DE", "position": "PhD student" } ], "n_formula": 103, "n_formula_1": 48, "n_ref_uni": 40, "n_ref": 92, "n_ref_all": 126, "n_fig": 4, "n_tab": 4, "L_tab": 901, "n_element_tab": 48, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1528, "n_element_tab_1": 18, "formula_len_all": 8637, "formula_len_all_1": 3648, "len_all": 183907, "len_all_1": 65695, "len_abs": 1263, "len_title": 136, "len_sents": 52706, "len_sents_1": 25755, "n_sents": 502, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1138, "title": "Neural Sampling from Boltzmann Densities: Fisher-Rao Curves in the Wasserstein Geometry", "abs": "We deal with the task of sampling from an unnormalized Boltzmann density $\\rho_D$\nby learning a Boltzmann curve given by energies $f_t$ starting in a simple density $\\rho_Z$.\nFirst, we examine conditions under which Fisher-Rao flows are absolutely continuous in the Wasserstein geometry.\nSecond, we address specific interpolations $f_t$ and the learning of the related density/velocity pairs $(\\rho_t,v_t)$.\nIt was numerically observed that the linear interpolation, \nwhich requires only a parametrization of the velocity field $v_t$,\nsuffers from a \"teleportation-of-mass\" issue.\nUsing tools from the Wasserstein geometry,\nwe give an analytical example,\nwhere we can precisely measure the explosion of the velocity field.\nInspired by M\u00e1t\u00e9 and Fleuret, who \nparametrize both $f_t$ and $v_t$, we propose an\ninterpolation which parametrizes only $f_t$ and fixes an appropriate $v_t$. \nThis corresponds to\nthe Wasserstein gradient flow of the Kullback-Leibler divergence related to Langevin dynamics. \nWe demonstrate by numerical examples that our model provides a well-behaved flow field which successfully solves the above sampling task.", "keywords": [ "Sampling", "Boltzmann densities", "Fisher-Rao Curves", "Wasserstein Gradient Flows", "Diffusion", "Interpolations" ], "rating_list": [ 6, 5, 8, 5, 8 ], "soundness_list": [ 3, 3, 4, 2, 3 ], "presentation_list": [ 2, 2, 3, 2, 4 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 2, 3, 3, 4, 4 ] }, { "paper_id": "TUVaDGuXrK", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Taida Li", "gender": "Male", "institution": "University of North Carolina at Charlotte", "country": "US", "position": "PhD student" }, { "name": "WenZhan Song", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Full Professor" }, { "name": "Xiang Zhang", "gender": "Male", "institution": "University of North Carolina at Charlotte", "country": "", "position": "Assistant Professor" }, { "name": "Yihe Wang", "gender": "Male", "institution": "University of North Carolina at Charlotte", "country": "US", "position": "PhD student" }, { "name": "Yujun Yan", "gender": "Female", "institution": "Dartmouth College", "country": "US", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 36, "n_ref": 75, "n_ref_all": 93, "n_fig": 2, "n_tab": 4, "L_tab": 1418, "n_element_tab": 132, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1257, "n_element_tab_1": 118, "formula_len_all": 137, "formula_len_all_1": 153, "len_all": 156274, "len_all_1": 77499, "len_abs": 1464, "len_title": 0, "len_sents": 52279, "len_sents_1": 35987, "n_sents": 353, "n_sents_1": 250, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1460, "title": "How To Evaluate Your Medical Time Series Classification?", "abs": "Medical time series (MedTS) play a critical role in many healthcare applications, such as vital sign monitoring and the diagnosis of brain and heart diseases. However, the existence of subject-specific features poses unique challenges in MedTS evaluation. Inappropriate evaluation setups that either exploit or overlook these features can lead to artificially inflated classification performance (by up to 50% in accuracy; ADFTD): this concern has received little attention in current research. Here, we categorize the existing evaluation setups into two primary categories: subject-dependent and subject-independent. We show the subject-independent setup is more appropriate for different datasets and tasks. Our theoretical analysis explores the feature components of MedTS, examining how different evaluation setups influence the features that a model learns. Through experiments on six datasets (spanning EEG, ECG, and fNIRS modalities) using four different methods, we demonstrate step-by-step how subject-dependent utilizes subject-specific features as a shortcut for classification and leads to a deceptive high performance, suggesting that the subject-independent setup is more precise and practicable evaluation setup in real-world. This comprehensive analysis aims to establish clearer guidelines for evaluating MedTS models in different healthcare applications. Code to reproduce this work in https://anonymous.4open.science/r/MedTS_Evaluation-733F.", "keywords": [ "Medical Time Series", "Subject-Independent", "Classification Evaluation", "Healthcare" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "TUKt7ag0qq", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Juan Jose Garau Luis", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Manus McAuliffe", "gender": "Male", "institution": "University of Dublin, Trinity College", "country": "IE", "position": "MS student" }, { "name": "Oliver Bent", "gender": "Male", "institution": "InstaDeep", "country": "", "position": "Researcher" }, { "name": "Paul Duckworth", "gender": "Male", "institution": "InstaDeep", "country": "GB", "position": "Principal Researcher" }, { "name": "Shikha Surana", "gender": "Female", "institution": "InstaDeep", "country": "", "position": "Researcher" }, { "name": "Thomas D Barrett", "gender": "Male", "institution": "InstaDeep", "country": "GB", "position": "Researcher" }, { "name": "Jacob Austin Beck", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 15, "n_ref_uni": 30, "n_ref": 94, "n_ref_all": 126, "n_fig": 6, "n_tab": 9, "L_tab": 2181, "n_element_tab": 197, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 880, "n_element_tab_1": 92, "formula_len_all": 599, "formula_len_all_1": 1208, "len_all": 135718, "len_all_1": 72670, "len_abs": 1569, "len_title": 110, "len_sents": 45632, "len_sents_1": 33149, "n_sents": 379, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1466, "title": "Metalic: Meta-Learning In-Context with Protein Language Models", "abs": "Predicting the biophysical and functional properties of proteins is essential for in silico protein design. Machine learning has emerged as a promising technique for such prediction tasks. However, the relative scarcity of in vitro annotations means that these models often have little, or no, specific data on the desired fitness prediction task. As a result of limited data, protein language models (PLMs) are typically trained on general protein sequence modeling tasks, and then fine-tuned, or applied zero-shot, to protein fitness prediction. When no task data is available, the models make strong assumptions about the correlation between the protein sequence likelihood and fitness scores. In contrast, we propose meta-learning over a distribution of standard fitness prediction tasks, and demonstrate positive transfer to unseen fitness prediction tasks. Our method, called Metalic (Meta-Learning In-Context), uses in-context learning and fine-tuning, when data is available, to adapt to new tasks. Crucially, fine-tuning enables considerable generalization, even though it is not accounted for during meta-training. Our fine-tuned models achieve strong results with 18 times fewer parameters than state-of-the-art models. Moreover, our method sets a new state-of-the-art in low-data settings on ProteinGym, an established fitness-prediction benchmark. Due to data scarcity, we believe meta-learning will play a pivotal role in advancing protein engineering.", "keywords": [ "Meta-Learning", "In-context Learning", "Tuning", "Protein", "Fitness Prediction", "Property Prediction", "Protein Language Model", "PLM", "Large Language Model" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "TUC0ZT2zIQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Anej Svete", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Shauli Ravfogel", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "V\u00e9steinn Sn\u00e6bjarnarson", "gender": "Male", "institution": "University of Copenhagen", "country": "DK", "position": "PhD student" }, { "name": "Ryan Cotterell", "gender": "Not Specified", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 38, "n_formula_1": 16, "n_ref_uni": 73, "n_ref": 158, "n_ref_all": 195, "n_fig": 4, "n_tab": 2, "L_tab": 236, "n_element_tab": 5, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 60, "n_element_tab_1": 2, "formula_len_all": 4268, "formula_len_all_1": 1277, "len_all": 249484, "len_all_1": 66628, "len_abs": 2273, "len_title": 101, "len_sents": 69017, "len_sents_1": 30314, "n_sents": 569, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1447, "title": "True Counterfactual Generation from Language Models", "abs": "Understanding and manipulating the causal generation mechanisms in language models is essential for controlling their behavior. Previous work has primarily relied on techniques such as representation surgery---e.g., model ablations or manipulation of linear subspaces tied to specific concepts---to \\emph{intervene} on these models. To understand the impact of interventions precisely, it is useful to examine \\emph{counterfactuals}---e.g., how a given sentence would have appeared had it been generated by the model following a specific intervention. We highlight that counterfactual reasoning is conceptually distinct from interventions, as articulated in Pearl's causal hierarchy. Based on this observation, we propose a framework for generating true string counterfactuals by reformulating language models as a structural equation model using the Gumbel-max trick, which we called Gumbel counterfactual generation. \nThis reformulation allows us to model the joint distribution over original strings and their counterfactuals resulting from the same instantiation of the sampling noise. We develop an algorithm based on hindsight Gumbel sampling that allows us to infer the latent noise variables and generate counterfactuals of observed strings. Our experiments demonstrate that the approach produces meaningful counterfactuals while at the same time showing that commonly used intervention techniques have considerable undesired side effects.", "keywords": [ "Causality", "language models", "counterfactuals" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "TU5ApbbeDZ", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Carlo Alfano", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Jakob Nicolaus Foerster", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Patrick Rebeschini", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Silvia Sapora", "gender": "Female", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Yee Whye Teh", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 17, "n_ref_uni": 41, "n_ref": 61, "n_ref_all": 82, "n_fig": 3, "n_tab": 3, "L_tab": 695, "n_element_tab": 129, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 695, "n_element_tab_1": 129, "formula_len_all": 2145, "formula_len_all_1": 1549, "len_all": 148247, "len_all_1": 66513, "len_abs": 3174, "len_title": 100, "len_sents": 42347, "len_sents_1": 30324, "n_sents": 307, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1040, "title": "Learning Loss Landscapes in Preference Optimization", "abs": "We present a framework to discover preference optimization algorithms specialized to particular scenarios, in a theoretically sound and computationally efficient setting. We start by designing a novel family of PO algorithms based on mirror descent, which we call Mirror Preference Optimization (MPO). MPO recovers existing methods like Direct Preference Optimization (DPO) and Odds-Ratio Preference Optimization (ORPO) for specific choices of the mirror map. Given specific properties of preference datasets, such as mixed-quality or noisy data, we show that we can efficiently search the MPO class to find specialized algorithms that outperform current baselines. Namely, we leverage evolutionary strategies and preference datasets generated on MuJoCo environments to systematically evaluate and optimize MPO algorithms on hand-crafted scenarios. We demonstrate the resulting PO algorithms successfully transfer to a Large Language Model (LLM) alignment task, where they demonstrate superior robustness in handling mixed-quality datasets.", "keywords": [ "Preference optimization", "mirror descent" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "TU3wJQfKz8", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Linjian Meng", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Shangdong Yang", "gender": "Male", "institution": "Nanjing University of Posts and Telecommunications", "country": "", "position": "Lecturer" }, { "name": "Wenbin Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Yang Gao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Youzhi Zhang", "gender": "unknown", "institution": "Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "", "position": "Assistant Professor" }, { "name": "Zheng Xu", "gender": "unknown", "institution": "NetEase, Inc.", "country": "CN", "position": "Researcher" }, { "name": "Zhenxing Ge", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "PhD student" }, { "name": "Tianyu Ding", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" } ], "n_formula": 210, "n_formula_1": 44, "n_ref_uni": 27, "n_ref": 113, "n_ref_all": 240, "n_fig": 4, "n_tab": 1, "L_tab": 2, "n_element_tab": 1, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 142, "n_element_tab_1": 8, "formula_len_all": 30971, "formula_len_all_1": 5723, "len_all": 219432, "len_all_1": 73105, "len_abs": 1745, "len_title": 138, "len_sents": 47539, "len_sents_1": 23631, "n_sents": 529, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1775, "title": "Last-Iterate Convergence of Smooth Regret Matching$^+$ Variants in Learning Nash Equilibria", "abs": "Regret Matching$^+$ (RM$^+$) variants have been widely developed to superhuman Poker AIs, yet few studies investigate their last-iterate convergence. Their last-iterate convergence has been demonstrated only for games with strong monotonicity or two-player zero-sum matrix games. A primary obstacle in proving the last-iterate convergence for these algorithms is that their feedback is not the loss gradient of the vanilla games. This deviation results in the absence of crucial properties, \\eg, monotonicity or the weak Minty variation inequality (MVI), which are pivotal for establishing the last-iterate convergence. To address the absence of these properties, we propose a remarkably succinct yet novel proof paradigm that consists of: (i) recovering these key properties through the equivalence between RM$^+$ and Online Mirror Descent (OMD), and (ii) measuring the the distance to Nash equilibrium (NE) via the tangent residual to show this distance is related to the distance between accumulated regrets. To show the practical applicability of our proof paradigm, we use it to prove the last-iterate convergence of two existing smooth RM$^+$ variants, Smooth Extra-gradient RM$^+$ (SExRM$^+$) and Smooth Predictive RM$^+$ (SPRM$^+$). We show that they achieve last-iterate convergence in learning an NE of games satisfying monotonicity, a weaker condition than the one used in existing proofs for both variants. Then, inspired by our proof paradigm, we propose Smooth Optimistic Gradient RM$^+$ (SOGRM$^+$). We show that SOGRM$^+$ achieves last-iterate convergence in learning an NE of games satisfying the weak MVI, the weakest condition in all known proofs for RM$^+$ variants. The experimental results show that SOGRM$^+$ significantly outperforms other algorithms.", "keywords": [ "Regret Matching$^+$", "Last-Iterate Convergence", "Nash Equilibrium" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 4, 3, 2 ] }, { "paper_id": "TTWxMAwS6n", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Chenyang Si", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Jia-Wei Liu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Jussi Keppo", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Full Professor" }, { "name": "Lingmin Ran", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Rui Zhao", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Xudong Lin", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Ziwei Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Zheng Shou", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 26, "n_ref": 107, "n_ref_all": 128, "n_fig": 12, "n_tab": 2, "L_tab": 496, "n_element_tab": 50, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1880, "n_element_tab_1": 66, "formula_len_all": 388, "formula_len_all_1": 343, "len_all": 120834, "len_all_1": 56734, "len_abs": 1300, "len_title": 130, "len_sents": 32606, "len_sents_1": 25405, "n_sents": 270, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1305, "title": "X-PlugVid: Versatile Adaptation of Image Plugins for Controllable Video Generation", "abs": "We introduce X-PlugVid, a unified framework designed to seamlessly adapt pretrained image-based plug-and-play modules for video diffusion models, facilitating controllable video generation without the need for retraining. This framework leverages a spatial-temporal adapter to effectively bridge the gap between image and video diffusion models. Specifically, we adopt a frozen copy of a large-scale\npretrained image diffusion model (e.g. Stable Diffusion v1.5) as spatial prior. Then we train a spatial-temporal adapter to convert the prior into temporally consistent guidance for video diffusion models (e.g. SVD). To further enhance the effectiveness of image plugins in guiding video models, we introduce a timestep remapping strategy. Recognizing that denoising is an entropic reduction process, this strategy selects priors from later timesteps of the image model, which contain richer information, to be injected into the video models, optimizing the quality and consistency of the generated videos. Comprehensive experimental evaluations of X-PlugVid demonstrate its broad compatibility with diverse operational conditions and different plugins, confirming that leveraging priors from a pretrained diffusion model can minimize redundant training and enable versatile controllable video generation.", "keywords": [ "video generation", "diffusion model", "efficiency" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "TTUtPIpaol", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Enshu Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Guohao Dai", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Huazhong Yang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Junyi Zhu", "gender": "Male", "institution": "Samsung", "country": "GB", "position": "Researcher" }, { "name": "Shengen Yan", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Xuefei Ning", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Research Assistant Professor" }, { "name": "Yu Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Zinan Lin", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Senior Researcher" }, { "name": "Matthew Blaschko", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 63, "n_ref": 120, "n_ref_all": 136, "n_fig": 12, "n_tab": 22, "L_tab": 8854, "n_element_tab": 1250, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2499, "n_element_tab_1": 379, "formula_len_all": 816, "formula_len_all_1": 404, "len_all": 247846, "len_all_1": 81578, "len_abs": 1842, "len_title": 171, "len_sents": 56064, "len_sents_1": 34647, "n_sents": 517, "n_sents_1": 296, "venue": "ICLR2025", "decision": "Reject", "L_title": 122, "L_abs": 1860, "title": "Efficient Expert Pruning for Sparse Mixture-of-Experts Language Models: Enhancing Performance and Reducing Inference Costs", "abs": "The rapid advancement of large language models (LLMs) has led to architectures with billions to trillions of parameters, posing significant deployment challenges due to their substantial demands on memory, processing power, and energy consumption. Sparse Mixture-of-Experts (SMoE) architectures have emerged as a solution, activating only a subset of parameters per token, thereby achieving faster inference while maintaining performance. However, SMoE models still face limitations in broader deployment due to their large parameter counts and significant GPU memory requirements. \nIn this work, we introduce a gradient-free evolutionary strategy named Efficient Expert Pruning (EEP) to enhance the pruning of experts in SMoE models. Specifically, EEP searches the pruning pattern and use expert merging as an memory-efficient way of fine-tuning the pruned model. EEP relies solely on model inference (i.e., no gradient computation) and achieves greater sparsity while maintaining or even improving performance on downstream tasks. EEP can be used to reduce both the total number of experts (thus saving GPU memory) and the number of active experts (thus accelerating inference).\nFor example, in the task-specific setting, we demonstrate that pruning up to 75\\% of experts in Mixtral $8\\times7$B-Instruct results in a substantial reduction in parameters with minimal performance loss, or pruning 50\\% of experts and activating one fewer expert to achieve 1.41$\\times$ speedup. Our experiments include four different model sizes from Mixtral, Qwen1.5 and Qwen2, and utilize more than 10 datasets as well as various settings. Results show that our method outperforms the related baselines by a large margin, demonstrating a significant advancement in this direction. Results of our method can be reproduced using the code provided in the supplementary material.", "keywords": [ "Large language model", "mixture of experts", "pruning" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "TSrhLq5hSA", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Yinan Feng", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Yinpeng Chen", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Youzuo Lin", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Associate Professor" }, { "name": "Yueh Z Lee", "gender": "unknown", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 18, "n_ref": 48, "n_ref_all": 62, "n_fig": 7, "n_tab": 3, "L_tab": 135, "n_element_tab": 35, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 837, "formula_len_all_1": 792, "len_all": 99275, "len_all_1": 62753, "len_abs": 1244, "len_title": 94, "len_sents": 31877, "len_sents_1": 29375, "n_sents": 256, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1247, "title": "On a Hidden Property in Computational Imaging", "abs": "Computational imaging plays a vital role in various scientific and medical applications, such as Full Waveform Inversion (FWI), Computed Tomography (CT), and Electromagnetic (EM) inversion. These methods address inverse problems by reconstructing physical properties (e.g., the acoustic velocity map in FWI) from measurement data (e.g., seismic waveform data in FWI), where both modalities are governed by complex mathematical equations. In this paper, we empirically demonstrate that despite their differing governing equations, three inverse problems\u2014FWI, CT, and EM inversion\u2014share a hidden property within their latent spaces. Specifically, using FWI as an example, we show that both modalities (the velocity map and seismic waveform data) follow the same set of one-way wave equations in the latent space, yet have distinct initial conditions that are linearly correlated. This suggests that after projection into the latent embedding space, the two modalities correspond to different solutions of the same equation, connected through their initial conditions. Our experiments confirm that this hidden property is consistent across all three imaging problems, providing a novel perspective for understanding these computational imaging tasks.", "keywords": [ "Computational Imaging", "Latent Space Representation", "Inverse Problems" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 4, 1, 2 ], "contribution_list": [ 3, 1, 2 ], "confidence_list": [ 3, 2, 2 ] }, { "paper_id": "TSlJ3ikcBZ", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Eric Elmoznino", "gender": "unknown", "institution": "Google", "country": "CA", "position": "Researcher" }, { "name": "Guillaume Lajoie", "gender": "Male", "institution": "Mila, Quebec AI institute", "country": "", "position": "Core Academic Member" }, { "name": "Leo Gagnon", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "Undergrad student" }, { "name": "Sangnie Bhardwaj", "gender": "Female", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "Sarthak Mittal", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "Graduate Student" }, { "name": "Dhanya Sridhar", "gender": "Female", "institution": "Universit\u00e9 de Montr\u00e9al and Mila-Quebec AI Institute", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 3, "n_ref_uni": 53, "n_ref": 81, "n_ref_all": 105, "n_fig": 13, "n_tab": 1, "L_tab": 447, "n_element_tab": 6, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 428, "formula_len_all_1": 187, "len_all": 191610, "len_all_1": 66272, "len_abs": 1556, "len_title": 130, "len_sents": 62989, "len_sents_1": 35152, "n_sents": 414, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1562, "title": "Does learning the right latent variables necessarily improve in-context learning?", "abs": "Large autoregressive models like Transformers can solve tasks through in-context learning (ICL) without learning new weights, suggesting avenues for efficiently solving new tasks. For many tasks, e.g., linear regression, the data factorizes: examples are independent given a task latent that generates the data, e.g., linear coefficients. While an optimal predictor leverages this factorization by inferring task latents, it is unclear if Transformers implicitly do so or if they instead exploit heuristics and statistical shortcuts enabled by attention layers. Both scenarios have inspired active ongoing work. In this paper, we systematically investigate the effect of explicitly inferring task latents. We minimally modify the Transformer architecture with a bottleneck designed to prevent shortcuts in favor of more structured solutions, and then compare performance against standard Transformers across various ICL tasks. Contrary to intuition and some recent works, we find little discernible difference between the two; biasing towards task-relevant latent variables does not lead to better out-of-distribution performance, in general. Curiously, we find that while the bottleneck effectively learns to extract latent task variables from context, downstream processing struggles to utilize them for robust prediction. Our study highlights the intrinsic limitations of Transformers in achieving structured ICL solutions that generalize, and shows that while inferring the right latents aids interpretability, it is not sufficient to alleviate this problem.", "keywords": [ "in-context learning", "transformers", "attention", "latent variable", "shortcuts" ], "rating_list": [ 5, 8, 6, 3, 6 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 1, 3 ], "confidence_list": [ 3, 4, 3, 3, 2 ] }, { "paper_id": "TSZh4610VG", "primary_area": "learning theory", "n_author": 8, "author_info_list": [ { "name": "Fan Lyu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Postdoc" }, { "name": "Fanhua Shang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" }, { "name": "Fuyuan Hu", "gender": "Male", "institution": "Suzhou University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Liang Wang", "gender": "Male", "institution": "Institute of Automation\uff0c CAS\uff0cChina", "country": "CN", "position": "Full Professor" }, { "name": "Wei Feng", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" }, { "name": "Ye Liu", "gender": "unknown", "institution": "Tianjin University", "country": "", "position": "PhD student" }, { "name": "Zhang Zhang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Ziqi Shi", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 20, "n_formula_1": 15, "n_ref_uni": 47, "n_ref": 85, "n_ref_all": 107, "n_fig": 10, "n_tab": 11, "L_tab": 4437, "n_element_tab": 247, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1283, "n_element_tab_1": 70, "formula_len_all": 1951, "formula_len_all_1": 736, "len_all": 131416, "len_all_1": 59003, "len_abs": 1201, "len_title": 91, "len_sents": 32270, "len_sents_1": 23334, "n_sents": 263, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 1189, "title": "Controllable Continual Test-Time Adaptation", "abs": "Continual Test-Time Adaptation (CTTA) is an emerging and challenging task where a model trained in a source domain must adapt to continuously changing conditions during testing, without access to the original source data. CTTA is prone to error accumulation due to uncontrollable domain shifts, leading to blurred decision boundaries between categories. Existing CTTA methods primarily focus on suppressing domain shifts, which proves inadequate during the unsupervised test phase.\nIn contrast, we introduce a novel approach that guides rather than suppresses these shifts.\nSpecifically, we propose $\\textbf{C}$ontrollable $\\textbf{Co}$ntinual $\\textbf{T}$est-$\\textbf{T}$ime $\\textbf{A}$daptation (C-CoTTA), which explicitly prevents any single category from encroaching on others, thereby mitigating the mutual influence between categories caused by uncontrollable shifts. \nMoreover, our method reduces the sensitivity of model to domain transformations, thereby minimizing the magnitude of category shifts. \nExtensive quantitative experiments demonstrate the effectiveness of our method, while qualitative analyses, such as t-SNE plots, confirm the theoretical validity of our approach.", "keywords": [ "Domain shift", "Continual Test-Time Adaptation", "Regularization" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "TSTgP4W3ga", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Alexander Ryabov", "gender": "Male", "institution": "Skolkovo Institute Of Science and Technology", "country": "", "position": "Researcher" }, { "name": "Evgeny Burnaev", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Full Professor" }, { "name": "N M Anoop Krishnan", "gender": "Male", "institution": "Indian Institute of Technology Delhi", "country": "", "position": "Associate Professor" }, { "name": "Nikolay Yavich", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Researcher" }, { "name": "Sayan Ranu", "gender": "Male", "institution": "Indian Institute of Technology Delhi", "country": "IN", "position": "Associate Professor" }, { "name": "Sergei Shumilin", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Researcher" }, { "name": "Viacheslav Naumov", "gender": "unknown", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Vladimir Vanovskiy", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 27, "n_ref": 32, "n_ref_all": 51, "n_fig": 11, "n_tab": 1, "L_tab": 365, "n_element_tab": 37, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 260, "n_element_tab_1": 32, "formula_len_all": 626, "formula_len_all_1": 586, "len_all": 130081, "len_all_1": 65332, "len_abs": 3977, "len_title": 167, "len_sents": 45307, "len_sents_1": 30840, "n_sents": 359, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 118, "L_abs": 1418, "title": "Learnable Stability-Aware Unstructured Grid Coarsening Using Graph Neural Networks for Accelerated Physics Simulations", "abs": "Efficient simulations of complex physical systems described by partial differential equations (PDE) require computational methods that can reduce the resource demands without sacrificing the accuracy. Traditionally, this is achieved by ``upscaling'' the simulation grids or by aggregating cells based on a priori information. Here, we introduce a novel framework based on graph neural networks (GNN) for learnable self-supervised differentiable coarsening of unstructured computational grids. We leverage graph-based representation of the physical system and offer a graph coarsening method which preserves the underlying physical properties together with the stability of the chosen numerical scheme. This is achieved by minimizing the error between the output of the simulations using coarsened and original graph. We demonstrate the approach on several example differential equations, modeling sub-surface flow and wave propagation. We demonstrate that the model exhibits ability to maintain high fidelity in simulation outputs even after 95\\% reduction on the nodes, significantly reducing computational overhead. We also show that the model exhibits generalizability to unseen scenarios, thereby outperforming the baselines. Thus, the developed approach demonstrates the ability to accelerate simulation without comprising accuracy and hence has potential for accelerating physical simulations in various domains.", "keywords": [ "Graph Neural Networks", "differentiable solvers", "numerical modelling", "grid coarsening", "upscaling" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 5, 3, 3 ] }, { "paper_id": "TS4BLf951Y", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Jianqi Chen", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Jin Jiang", "gender": "Female", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Mingfei Han", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "Panwen Hu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "Shengcai Liao", "gender": "Male", "institution": "United Arab Emirates University", "country": "US", "position": "Associate Professor" }, { "name": "Xiaodan Liang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaojun Chang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 47, "n_ref": 99, "n_ref_all": 123, "n_fig": 10, "n_tab": 10, "L_tab": 2345, "n_element_tab": 255, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1442, "n_element_tab_1": 113, "formula_len_all": 318, "formula_len_all_1": 318, "len_all": 171351, "len_all_1": 62697, "len_abs": 481, "len_title": 232, "len_sents": 44658, "len_sents_1": 30991, "n_sents": 327, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1617, "title": "StoryAgent: Customized Storytelling Video Generation via Multi-Agent Collaboration", "abs": "The advent of AI-Generated Content (AIGC) has spurred research into automated video generation to streamline conventional processes. However, automating storytelling video production, particularly for customized narratives, remains challenging due to the complexity of maintaining subject consistency across shots. While existing approaches like Mora and AesopAgent integrate multiple agents for Story-to-Video (S2V) generation, they fall short in preserving protagonist consistency and supporting Customized Storytelling Video Generation (CSVG). To address these limitations, we propose StoryAgent, a multi-agent framework designed for CSVG. StoryAgent decomposes CSVG into distinct subtasks assigned to specialized agents, mirroring the professional production process. Notably, our framework includes agents for story design, storyboard generation, video creation, agent coordination, and result evaluation. Leveraging the strengths of different models, StoryAgent enhances control over the generation process, significantly improving character consistency. Specifically, we introduce a customized Image-to-Video (I2V) method, LoRA-BE, to enhance intra-shot temporal consistency, while a novel storyboard generation pipeline is proposed to maintain subject consistency across shots. Extensive experiments demonstrate the effectiveness of our approach in synthesizing highly consistent storytelling videos, outperforming state-of-the-art methods. Our contributions include the introduction of StoryAgent, a versatile framework for video generation tasks, and novel techniques for preserving protagonist consistency.", "keywords": [ "storytelling video generation", "customization generation", "multi-agent system" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "TRWxFUzK9K", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Jong Chul Ye", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Taesung Kwon", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 31, "n_ref_uni": 20, "n_ref": 55, "n_ref_all": 101, "n_fig": 10, "n_tab": 7, "L_tab": 1569, "n_element_tab": 139, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 292, "n_element_tab_1": 56, "formula_len_all": 2557, "formula_len_all_1": 1855, "len_all": 109973, "len_all_1": 57067, "len_abs": 1557, "len_title": 107, "len_sents": 35078, "len_sents_1": 24582, "n_sents": 270, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1515, "title": "Solving Video Inverse Problems Using Image Diffusion Models", "abs": "Recently, diffusion model-based inverse problem solvers (DIS) have emerged as state-of-the-art approaches for addressing inverse problems, including image super-resolution, deblurring, inpainting, etc. \nHowever, their application to video inverse problems arising from spatio-temporal degradation remains largely unexplored due to the challenges in training video diffusion models.\nTo address this issue, here we introduce an innovative video inverse solver that leverages only image diffusion models.\nSpecifically, by drawing inspiration from the success of the recent decomposed diffusion sampler (DDS), \nour method treats the time dimension of a video as the batch dimension of image diffusion models and solves spatio-temporal optimization problems within denoised spatio-temporal batches derived from each image diffusion model.\nMoreover, we introduce a batch-consistent diffusion sampling strategy that encourages consistency across batches by synchronizing the stochastic noise components in image diffusion models. \nOur approach synergistically combines batch-consistent sampling with simultaneous optimization of denoised spatio-temporal batches at each reverse diffusion step, resulting in a novel and efficient diffusion sampling strategy for video inverse problems.\nExperimental results demonstrate that our method effectively addresses various spatio-temporal degradations in video inverse problems, achieving state-of-the-art reconstructions.\nProject page: https://solving-video-inverse.github.io/main", "keywords": [ "Image diffusion models", "Video inverse problems", "Batch-consistent sampling" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "TROUDY6Wg4", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Huizhuo Yuan", "gender": "Female", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Jiafan He", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" } ], "n_formula": 108, "n_formula_1": 32, "n_ref_uni": 64, "n_ref": 143, "n_ref_all": 155, "n_fig": 0, "n_tab": 8, "L_tab": 1707, "n_element_tab": 188, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1662, "n_element_tab_1": 145, "formula_len_all": 16260, "formula_len_all_1": 2124, "len_all": 272863, "len_all_1": 68194, "len_abs": 1454, "len_title": 118, "len_sents": 80896, "len_sents_1": 29647, "n_sents": 677, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1460, "title": "Accelerated Preference Optimization for Large Language Model Alignment", "abs": "Reinforcement Learning from Human Feedback (RLHF) has emerged as a pivotal tool for aligning large language models (LLMs) with human preferences. Direct Preference Optimization (DPO), one of the most popular approaches, formulates RLHF as a policy optimization problem without explicitly estimating the reward function. It overcomes the stability and efficiency issues of two-step approaches, which typically involve first estimating the reward function and then optimizing the policy via proximal policy optimization (PPO). Since RLHF is essentially an optimization problem, and it is well-known that momentum techniques can accelerate optimization both theoretically and empirically, a natural question arises: Can RLHF be accelerated by momentum? This paper answers this question in the affirmative. In detail, we first show that the iterative preference optimization method can be viewed as a proximal point method. Based on this observation, we propose a general Accelerated Preference Optimization (APO) framework, which unifies many existing preference optimization algorithms and employs Nesterov's momentum technique to speed up the alignment of LLMs. Theoretically, we demonstrate that APO can achieve a faster convergence rate than the standard iterative preference optimization methods, including DPO and SPPO. Empirically, we show the superiority of APO over DPO, iterative DPO, and other strong baselines for RLHF on the AlpacaEval 2.0 benchmark.", "keywords": [ "large language models", "RLHF", "DPO" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "TRHyAnInUC", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Chaojie Ji", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Hongyan Wu", "gender": "Female", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Jialin Nan", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Ruxin Wang", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Yankai Cao", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Associate Professor" } ], "n_formula": 34, "n_formula_1": 31, "n_ref_uni": 39, "n_ref": 75, "n_ref_all": 106, "n_fig": 2, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2424, "formula_len_all_1": 2316, "len_all": 117803, "len_all_1": 71801, "len_abs": 1593, "len_title": 121, "len_sents": 33874, "len_sents_1": 30130, "n_sents": 274, "n_sents_1": 243, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1609, "title": "$D^3PM$: Diffusion Model Responds to the Duty Call from Causal Discovery", "abs": "Causal discovery (CD) involves inferring cause-and-effect relationships as directed acyclic graphs (DAGs). In this work, we assume that the data is generated by an additive noise model (ANM). Recent work has formulated the problem as a continuous optimization problem, which consists of solving an inverse problem and satisfying an acyclicity constraint. However, solving the inverse problem in CD is often unstable, i.e. high sensitivity of the effects to perturbations in the causes. To address this instability, we formulate the inverse problem as a regularized optimization scheme and propose a novel variation-negotiation regularizer. Compared to traditional regularization techniques for the continuous optimization problem, e.g. $\\ell_1$ penalty on graphs, the proposed regularizer exploits the variation variable in ANMs to stabilize the solutions (i.e. DAGs). This regularizer is advantageous as it does not rely on any hypotheses, such as graph sparsity, about true DAGs. The variation-negotiation regularizer regulates the DAG purely based on observed data.\n\nBuilding on the proposed regularizer, a series of improvements to the regularized optimization scheme reveal the connections between solving the regularized optimization problem and learning a diffusion model, as they share comparable objective functions. This insight leads us to develop an equivalent diffusion model called DAG-invariant Denoising Diffusion Probabilistic Model. Extensive empirical experiments on synthetic and real datasets demonstrate that the proposed diffusion model achieves outstanding performance on all datasets.", "keywords": [ "causal discovery", "deep learning", "diffusion models" ], "rating_list": [ 3, 3, 1, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "TQdg1X6eqm", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Dan Yamins", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Evelina Fedorenko", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Associate Professor" }, { "name": "Greta Tuckute", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Klemen Kotar", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 24, "n_ref": 53, "n_ref_all": 72, "n_fig": 11, "n_tab": 4, "L_tab": 1389, "n_element_tab": 153, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 1342, "n_element_tab_1": 148, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 154697, "len_all_1": 59955, "len_abs": 1339, "len_title": 120, "len_sents": 33840, "len_sents_1": 28838, "n_sents": 214, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1369, "title": "Representing speech through autoregressive prediction of cochlear tokens", "abs": "We introduce a biologically-inspired model for encoding speech through an autoregressive prediction objective applied to input representations modeled after the human cochlea.\nOur modeling framework is inspired by the human auditory processing hierarchy. The first stage of our framework transforms the raw audio waveform to a time-frequency representation inspired by the human cochlea, with an intermediary step that effectively discretizes the audio representations (cochlear tokens). The second stage of our model learns a simple, yet powerful, autoregressive sequence model over the discretized audio input.\nWe demonstrate that our model learns meaningful representations of phonemes and word identities, and state-of-the-art representations of lexical semantic similarity. In addition, our model shows competitive performance on several downstream audio tasks from the SUPERB benchmark. In addition to our model\u2019s strong representational capabilities, we demonstrate our model's ability to generate continuations of audio at various temporal scales, which can be visualized in a cochleagram time-frequency space to provide insights into the model's predictions.\nOur model provides a novel framework for speech representation learning, aiming to advance the development of more human-like models that flexibly and efficiently handles a range of speech-based tasks.", "keywords": [ "audio", "speech", "biology-inspired model", "autoregressive prediction", "cochlea" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "TQ7Nuy1CSm", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Chaozhuo Li", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Lirong Qiu", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Full Professor" }, { "name": "Litian Zhang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Rui Ha", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "PhD student" }, { "name": "Rui Pu", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "PhD student" }, { "name": "Xi Zhang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Full Professor" }, { "name": "Zheng Liu", "gender": "unknown", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Chen ZeJian", "gender": "Male", "institution": "Department of Software Engineering, Shenzhen University", "country": "", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 4, "n_ref_uni": 9, "n_ref": 49, "n_ref_all": 87, "n_fig": 7, "n_tab": 6, "L_tab": 1354, "n_element_tab": 91, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1829, "n_element_tab_1": 85, "formula_len_all": 787, "formula_len_all_1": 294, "len_all": 149043, "len_all_1": 68480, "len_abs": 1775, "len_title": 129, "len_sents": 47816, "len_sents_1": 31241, "n_sents": 394, "n_sents_1": 261, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1788, "title": "Feint and Attack: Attention-Based Strategies for Jailbreaking and Protecting LLMs", "abs": "Jailbreak attack can be used to access the vulnerabilities of Large Language Models (LLMs) by inducing LLMs to generate the harmful content. \nAnd the most common method of the attack is to construct semantically ambiguous prompts to confuse and mislead the LLMs.\nTo access the security and reveal the intrinsic relation between the input prompt and the output for LLMs, the distribution of attention weight is introduced to analyze the underlying reasons. \nBy using statistical analysis methods, some novel metrics are defined to better describe the distribution of attention weight, such as the Attention Intensity on Sensitive Words (Attn_SensWords), the Attention-based Contextual Dependency Score (Attn_DepScore) and Attention Dispersion Entropy (Attn_Entropy).\nBy leveraging the distinct characteristics of these metrics, the beam search algorithm and inspired by the military strategy \"Feint and Attack'', an effective jailbreak attack strategy named as Attention-Based Attack (ABA) is proposed.\nIn the ABA, nested attack prompts are employed to divert the attention distribution of the LLMs. \nIn this manner, more harmless parts of the input can be used to attract the attention of the LLMs.\nIn addition, motivated by ABA, an effective defense strategy called as Attention-Based Defense (ABD) is also put forward.\nCompared with ABA, the ABD can be used to enhance the robustness of LLMs by calibrating the attention distribution of the input prompt. \nSome comparative experiments have been given to demonstrate the effectiveness of ABA and ABD. \nTherefore, both ABA and ABD can be used to access the security of the LLMs. \nThe comparative experiment results also give a logical explanation that the distribution of attention weight can bring great influence on the output for LLMs.", "keywords": [ "jailbreak", "defense", "attention machanism", "LLM" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 1, 3, 1 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "TQ2ZOy6miT", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Jiafeng Guo", "gender": "Male", "institution": "Institute of Computing Technolgy, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Keping Bi", "gender": "Female", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Mingkun Zhang", "gender": "Female", "institution": ", Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Wei Chen", "gender": "Female", "institution": " Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xueqi Cheng", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy", "country": "CN", "position": "Full Professor" } ], "n_formula": 29, "n_formula_1": 11, "n_ref_uni": 57, "n_ref": 161, "n_ref_all": 222, "n_fig": 18, "n_tab": 7, "L_tab": 2087, "n_element_tab": 156, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 491, "n_element_tab_1": 37, "formula_len_all": 2583, "formula_len_all_1": 1262, "len_all": 185195, "len_all_1": 54110, "len_abs": 1737, "len_title": 144, "len_sents": 57053, "len_sents_1": 25373, "n_sents": 360, "n_sents_1": 152, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1772, "title": "CLIPure: Purification in Latent Space via CLIP for Adversarially Robust Zero-Shot Classification", "abs": "In this paper, we aim to build an adversarially robust zero-shot image classifier that can accurately and efficiently classify unseen examples while defending against unforeseen adversarial attacks, addressing critical challenges in real-world safety-sensitive scenarios. To achieve this, we focus on two key challenges: zero-shot classification and defense against unforeseen attacks. We ground our work on CLIP, a vision-language pre-trained model to perform zero-shot classification. \nTo defend against unforeseen attacks, we adopt a purification approach, as it is independent of specific attack types. \nWe then define a purification risk as the KL divergence between the joint distributions of the purification and attack process. \nThe derived lower bound of purification risk inspires us to explore purification in CLIP's multi-modal latent space. \nWe propose a CLIP-based purification method called CLIPure, which has two variants: _CLIPure-Diff_, which models image likelihood with a generative process of its latent vector, and _CLIPure-Cos_, which models the likelihood based on the similarity between embeddings of the image and a blank template. As far as we know, CLIPure is the first purification method in latent space and _CLIPure-Cos_ is the first purification method not relying on generative models, substantially improving defense efficiency. Extensive experimental results show that the robustness achieved by CLIPure is within a small gap of clean accuracy, outperforming SOTA robustness by a large margin, e.g., from 71.7\\% to **91.1\\%** on CIFAR10, from 59.6\\% to **72.6\\%** on ImageNet, and **108\\%** relative improvements of average robustness on the 13 datasets over previous SOTA, with only 14\\% extra inference cost and no additional training.", "keywords": [ "adversarial robustness", "purification", "CLIP" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 3, 1, 4, 5, 3 ] }, { "paper_id": "TPtzZQyiFm", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chengjin Xu", "gender": "Male", "institution": "International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Jian Guo", "gender": "Male", "institution": "International Digital Economy Academy, International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Yiyan Qi", "gender": "unknown", "institution": "IDEA", "country": "CN", "position": "Researcher" }, { "name": "Zhanpeng Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 49, "n_ref": 62, "n_ref_all": 89, "n_fig": 7, "n_tab": 13, "L_tab": 5488, "n_element_tab": 752, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1430, "n_element_tab_1": 203, "formula_len_all": 428, "formula_len_all_1": 289, "len_all": 173005, "len_all_1": 59748, "len_abs": 1512, "len_title": 188, "len_sents": 41369, "len_sents_1": 26760, "n_sents": 336, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 139, "L_abs": 1302, "title": "MLLM Is a Strong Reranker: Advancing Multimodal Retrieval-augmented Generation via Knowledge-enhanced Reranking and Noise-injected Training", "abs": "Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in processing and generating content across multiple data modalities. However, a significant drawback of MLLMs is their reliance on static training data, leading to outdated information and limited contextual awareness. This static nature hampers their ability to provide accurate and up-to-date responses, particularly in dynamic or rapidly evolving contexts. Though integrating Multimodal Retrieval-augmented Generation (Multimodal RAG) offers a promising solution, the system would inevitably encounter the multi-granularity noisy correspondence (MNC) problem, which hinders accurate retrieval and generation. In this work, we propose RagVL, a novel framework with knowledge-enhanced reranking and noise-injected training, to address these limitations. We instruction-tune the MLLM with a simple yet effective instruction template to induce its ranking ability and serve it as a reranker to precisely filter the top-k retrieved images. For generation, we inject visual noise during training at the data and token levels to enhance the generator's robustness. Extensive experiments on four datasets verify the effectiveness of our method. Code and models are available at https://anonymous.4open.science/r/RagVL-F694.", "keywords": [ "Multimodal Retrieval-augmented Generation", "Multimodal Large Language Model" ], "rating_list": [ 5, 5, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 2, 4 ] }, { "paper_id": "TPiJKs7ccR", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Frank F. Xu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Graham Neubig", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Shuyan Zhou", "gender": "Non-Binary", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Yueqi Song", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 67, "n_ref_all": 97, "n_fig": 6, "n_tab": 9, "L_tab": 1936, "n_element_tab": 225, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 352, "n_element_tab_1": 45, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 194979, "len_all_1": 64610, "len_abs": 1261, "len_title": 85, "len_sents": 54984, "len_sents_1": 30726, "n_sents": 394, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 37, "L_abs": 1269, "title": "Beyond Browsing: API-Based Web Agents", "abs": "Web browsers are a portal to the internet, where much of human activity is undertaken. Thus, there has been significant research work in AI agents that interact with the internet through web browsing. However, there is also another interface designed specifically for machine interaction with online content: application programming interfaces (APIs). In this paper we ask -- what if we were to take tasks traditionally tackled by browsing agents, and give AI agents access to APIs? To do so, we propose two varieties of agents: (1) an API-calling agent that attempts to perform online tasks through APIs only, similar to traditional coding agents, and (2) a Hybrid Agent that can interact with online data through both web browsing and APIs. In experiments on WebArena, a widely-used and realistic benchmark for web navigation tasks, we find that API-based agents outperform web browsing agents. Hybrid Agents out-perform both others nearly uniformly across tasks, resulting in a more than 20.0% absolute improvement over web browsing alone, achieving a success rate of 35.8%, achiving the SOTA performance among task-agnostic agents. These results strongly suggest that when APIs are available, they present an attractive alternative to relying on web browsing alone.", "keywords": [ "AI Agent", "LLMs", "Evaluation" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "TPYwwqF0bv", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Daeho Um", "gender": "Male", "institution": "Samsung", "country": "KR", "position": "Researcher" }, { "name": "Jiwoong Park", "gender": "unknown", "institution": "Northeastern University", "country": "US", "position": "Postdoc" }, { "name": "Seong Jin Ahn", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Seulki Park", "gender": "Female", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Postdoc" }, { "name": "Yoonji Lee", "gender": "Female", "institution": "Seoul National University", "country": "", "position": "MS student" }, { "name": "Yuneil Yeo", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 22, "n_ref_uni": 39, "n_ref": 88, "n_ref_all": 142, "n_fig": 9, "n_tab": 16, "L_tab": 8137, "n_element_tab": 729, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2793, "n_element_tab_1": 213, "formula_len_all": 3379, "formula_len_all_1": 1339, "len_all": 201743, "len_all_1": 77605, "len_abs": 1426, "len_title": 130, "len_sents": 62779, "len_sents_1": 32125, "n_sents": 534, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1436, "title": "Relation-Aware Diffusion for Heterogeneous Graphs with Partially Observed Features", "abs": "Diffusion-based imputation methods, which impute missing features through the iterative propagation of observed features, have shown impressive performance in homogeneous graphs. However, these methods are not directly applicable to heterogeneous graphs, which have multiple types of nodes and edges, due to two key issues: (1) the presence of nodes with undefined features hinders diffusion-based imputation; (2) treating various edge types equally during diffusion does not fully utilize information contained in heterogeneous graphs. To address these challenges, this paper presents a novel imputation scheme that enables diffusion-based imputation in heterogeneous graphs. Our key idea involves (1) assigning a {\\it virtual feature} to an undefined node feature and (2) determining the importance of each edge type during diffusion according to a new criterion. Through experiments, we demonstrate that our virtual feature scheme effectively serves as a bridge between existing diffusion-based methods and heterogeneous graphs, maintaining the advantages of these methods. Furthermore, we confirm that adjusting the importance of each edge type leads to significant performance gains on heterogeneous graphs. Extensive experimental results demonstrate the superiority of our scheme in both semi-supervised node classification and link prediction tasks on heterogeneous graphs with missing rates ranging from low to exceedingly high.", "keywords": [ "missing features", "imputation", "heterogeneous graph" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "TOtk9dTYGG", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Boyu Yang", "gender": "Male", "institution": "China Mobile Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Lingxi Xie", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Qixiang Ye", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Tianren Ma", "gender": "Not Specified", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yunjie Tian", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Postdoc" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 39, "n_ref": 110, "n_ref_all": 140, "n_fig": 8, "n_tab": 15, "L_tab": 4550, "n_element_tab": 482, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2847, "n_element_tab_1": 212, "formula_len_all": 151, "formula_len_all_1": 151, "len_all": 185050, "len_all_1": 57978, "len_abs": 1340, "len_title": 122, "len_sents": 50164, "len_sents_1": 23703, "n_sents": 434, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1347, "title": "ClawMachine: Learning to Fetch Visual Tokens for Referential Comprehension", "abs": "Aligning vision and language concepts at a finer level remains an essential topic of multimodal large language models (MLLMs), particularly for tasks such as referring and grounding. Existing methods, such as proxy encoding and geometry encoding, incorporate additional syntax to encode spatial information, imposing extra burdens when communicating between language with vision modules. In this study, we propose ClawMachine, offering a new methodology that explicitly notates each entity using token collectives\u2014groups of visual tokens that collaboratively represent higher-level semantics. A hybrid perception mechanism is also explored to perceive and understand scenes from both discrete and continuous spaces. Our method unifies the prompt and answer of visual referential tasks without using additional syntax. By leveraging a joint vision-language vocabulary, ClawMachine further integrates referring and grounding in an auto-regressive manner, demonstrating great potential with scaled up pre-training data. Experiments show that ClawMachine achieves superior performance on scene-level and referential understanding tasks with higher efficiency. It also exhibits the potential to integrate multi-source information for complex visual reasoning, which is beyond the capability of many MLLMs. The model and data will be publicly available.", "keywords": [ "Multimodal Learning", "Visual Referring", "Referring Expression Comprehension", "Large Language Models" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "TOiageVNru", "primary_area": "applications to robotics, autonomy, planning", "n_author": 3, "author_info_list": [ { "name": "Ruiqi Ni", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "zherong pan", "gender": "Male", "institution": "Tencent America", "country": "US", "position": "Researcher" }, { "name": "Ahmed Qureshi", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 11, "n_ref_uni": 48, "n_ref": 103, "n_ref_all": 132, "n_fig": 7, "n_tab": 3, "L_tab": 979, "n_element_tab": 74, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 566, "n_element_tab_1": 49, "formula_len_all": 611, "formula_len_all_1": 752, "len_all": 159942, "len_all_1": 77554, "len_abs": 1255, "len_title": 126, "len_sents": 49956, "len_sents_1": 36876, "n_sents": 374, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1259, "title": "Physics-informed Temporal Difference Metric Learning for Robot Motion Planning", "abs": "The motion planning problem involves finding a collision-free path from a robot's starting to its target configuration. Recently, self-supervised learning methods have emerged to tackle motion planning problems without requiring expensive expert demonstrations. They solve the Eikonal equation for training neural networks and lead to efficient solutions. However, these methods struggle in complex environments because they fail to maintain key properties of the Eikonal equation, such as optimal value functions and geodesic distances. To overcome these limitations, we propose a novel self-supervised temporal difference metric learning approach that solves the Eikonal equation more accurately and enhances performance in solving complex and unseen planning tasks. Our method enforces Bellman's principle of optimality over finite regions, using temporal difference learning to avoid spurious local minima while incorporating metric learning to preserve the Eikonal equation's essential geodesic properties. We demonstrate that our approach significantly outperforms existing self-supervised learning methods in handling complex environments and generalizing to unseen environments, with robot configurations ranging from 2 to 12 degrees of freedom (DOF).", "keywords": [ "Robot motion planning", "Eikonal Equation", "Physics-informed Neural Networks", "Temporal Difference Learning", "Metric Learning" ], "rating_list": [ 8, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "TOahfjA3sP", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Fuchao Yang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "Min-Ling Zhang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Full Professor" }, { "name": "Ran Wang", "gender": "Female", "institution": "Shenzhen University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaorui Peng", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "Yuheng Jia", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 44, "n_formula_1": 22, "n_ref_uni": 35, "n_ref": 91, "n_ref_all": 118, "n_fig": 7, "n_tab": 9, "L_tab": 1404, "n_element_tab": 112, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 1263, "n_element_tab_1": 84, "formula_len_all": 4149, "formula_len_all_1": 1558, "len_all": 141617, "len_all_1": 65484, "len_abs": 1043, "len_title": 135, "len_sents": 43680, "len_sents_1": 28369, "n_sents": 340, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1047, "title": "Noise Separation guided Candidate Label Reconstruction for Noisy Partial Label Learning", "abs": "Partial label learning is a weakly supervised learning problem in which an instance is annotated with a set of candidate labels, among which only one is the correct label. However, in practice the correct label is not always in the candidate label set, leading to the noisy partial label learning (NPLL) problem. In this paper, we theoretically prove that the generalization error of the classifier constructed under NPLL paradigm is bounded by the noise rate and the average length of the candidate label set. Motivated by the theoretical guide, we propose a novel NPLL framework that can separate the noisy samples from the normal samples to reduce the noise rate and reconstruct the shorter candidate label sets for both of them. Extensive experiments on multiple benchmark datasets confirm the efficacy of the proposed method in addressing NPLL. For example, on CIFAR100 dataset with severe noise, our method improves the classification accuracy of the state-of-the-art one by 11.57%. The code is available at: https://github.com/pruirui/PLRC.", "keywords": [ "partial label learning", "weakly supervised learning", "noisy partial label learning" ], "rating_list": [ 6, 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 2, 4, 4, 3, 4 ] }, { "paper_id": "TOVBglQvhB", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Xiaohong Liu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ying Fu", "gender": "Female", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yong Guo", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Postdoc" }, { "name": "Yulun Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhang Gengchen", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 21, "n_ref": 49, "n_ref_all": 66, "n_fig": 5, "n_tab": 4, "L_tab": 1950, "n_element_tab": 166, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 3288, "n_element_tab_1": 110, "formula_len_all": 519, "formula_len_all_1": 420, "len_all": 101429, "len_all_1": 57645, "len_abs": 1957, "len_title": 91, "len_sents": 26351, "len_sents_1": 24546, "n_sents": 207, "n_sents_1": 195, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 1968, "title": "Low-bit Quantization for Seeing in the Dark", "abs": "Several properties of raw data exhibit significant potential for enhancing images under extremely low-light conditions. Recently, many deep-learning methods for raw-based low-light image enhancement (LLIE) have demonstrated excellent performance. However, deploying them on resource-limited devices is restricted by high computational and storage demands. In this work, we propose a novel low-bit quantization method for raw-based LLIE model to improve their efficiency. Nevertheless, directly adopting existing quantizers for LLIE networks leads to an obvious performance drop due to two main reasons. i) The U-Net model, commonly employed in LLIE, faces challenges in identifying a suitable quantization range due to disparities in distribution between the encoder and decoder features. ii) Low-bit quantized LLIE networks struggle to restore clear details in low-light images because their features have a constraint capacity. We address these issues by introducing a novel low-bit quantization method, the Distribution-Separative Asymmetric Quantizer (DSAQ), designed specifically for U-Net architectures used in LLIE. In order to accurately determine the quantization intervals, DSAQ separates the distribution of encoder and decoder features before they are concatenated by the skip connection. We also make the quantizer asymmetric with trainable scale and offset parameters to suit skewed activation ranges caused by non-linear functions. To further enhance performance, we propose a uniform feature distillation technique, which allows the low-bit student model to effectively assimilate knowledge from the full-precision teacher model, bridging the gap in representation capability. Extensive experiments show that our approach not only greatly reduces the memory and computational requirements of raw-based LLIE models but also has a promising performance. Our 4-bit quantized model can achieve comparable or superior results to full-precision counterparts.", "keywords": [ "low-light image enhancement", "network quantization" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "TOUrnb1EaG", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Carl Kingsford", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Herbert A. Simon Professor of Computer Science" }, { "name": "Jiayi Li", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Litian Liang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Shiyi Du", "gender": "Female", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 5, "n_ref_uni": 25, "n_ref": 51, "n_ref_all": 77, "n_fig": 5, "n_tab": 13, "L_tab": 6954, "n_element_tab": 542, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1936, "n_element_tab_1": 156, "formula_len_all": 1152, "formula_len_all_1": 340, "len_all": 172836, "len_all_1": 59871, "len_abs": 1996, "len_title": 85, "len_sents": 62258, "len_sents_1": 28623, "n_sents": 378, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 1818, "title": "DNA Language Models for RNA Analyses", "abs": "Genomic Language Models (gLMs), encompassing DNA models, RNA models, and multimodal models, are becoming widely used for the analysis of biological sequences. Typically, models trained on RNA are used for RNA-related tasks, and models trained on DNA sequences are used for DNA tasks. However, this requires the development and maintenance of several classes of models to match the modality of the sequence. These models take significant resources and data to create, and maintaining separate models for DNA and RNA tasks is a computational burden. \n\nTo reduce this burden, we introduce novel Adaptive Mixture of Codon Reformative Experts (CodonMoE) that can be incorporated into DNA gLMs in order to adapt them for mRNA-based predictive tasks. We show that, by using this plug-and-play operator, DNA-based gLMs can achieve performance similar to that of RNA-trained models on mRNA tasks. We further show that recent, efficient sub-quadratic DNA-based state space model (SSM) architectures can be used with the CodonMoE to achieve parameter- and computationally-efficient predictions for mRNA tasks. Specifically, experimental results demonstrate that CodonMoE improves diverse DNA-based backbones by a big margin, with some models achieving comparable or superior performance to current state-of-the-art RNA-specific models across several downstream tasks, while reducing both time complexity and model parameters.\n\nOur results provide a path for focusing development efforts of gLMs on DNA models, which can then be adapted to mRNA tasks. Because DNA data is more prevalent than assembled mRNA data, and modeling efforts can focus on a single class of model, this is likely to foster improved DNA models for mRNA tasks at lower computational cost, and is a significant step towards unifying genomic language modeling.", "keywords": [ "Genomic Language Models", "RNA Sequence Analysis", "Parameter-Efficient Fine-Tuning", "Mixture of Experts", "Computational Efficiency" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "TNj5i5i3pB", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Chunyi Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Guangtao Zhai", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Guo Lu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Haoning Wu", "gender": "Male", "institution": "Rhymes AI", "country": "SG", "position": "Researcher" }, { "name": "Jianbo Zhang", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Postdoc" }, { "name": "Wei Sun", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Postdoc" }, { "name": "Weisi Lin", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Xiaohong Liu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiongkuo Min", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuan Tian", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Postdoc" }, { "name": "zicheng zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 31, "n_ref": 62, "n_ref_all": 77, "n_fig": 10, "n_tab": 6, "L_tab": 5395, "n_element_tab": 234, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 4722, "n_element_tab_1": 179, "formula_len_all": 93, "formula_len_all_1": 66, "len_all": 150038, "len_all_1": 58230, "len_abs": 285, "len_title": 0, "len_sents": 36952, "len_sents_1": 25066, "n_sents": 250, "n_sents_1": 164, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1230, "title": "R-Bench: Are your Large Multimodel Model Robust to Real-world Corruption?", "abs": "The outstanding performance of Large Multimodal Models (LMMs) has made them widely applied in vision-related tasks. However, various corruptions in the real world mean that images will not be as ideal as in simulations, presenting significant challenges for the practical application of LMMs. To address this issue, we introduce R-Bench, a benchmark focused on the **Real-world Robustness of LMMs**. Specifically, we: (a) model the complete link from user capture to LMMs reception, comprising 33 corruption dimensions, including 7 steps according to the corruption sequence, and 7 groups based on low-level attributes; (b) collect reference/distorted image dataset before/after corruption, including 2,970 question-answer pairs with human labeling; (c) propose comprehensive evaluation for absolute/relative robustness and benchmark 20 mainstream LMMs. Results show that while LMMs can correctly handle the original reference images, their performance is not stable when faced with distorted images, and there is a significant gap in robustness compared to the human visual system. We hope that R-Bench will inspire improving the robustness of LMMs, **extending them from experimental simulations to the real-world application**.", "keywords": [ "Large Multimodal Models", "Robustness", "User Generated Content", "Perceptual Quality Corruption" ], "rating_list": [ 6, 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 3, 5, 4, 3, 4 ] }, { "paper_id": "TNYLCF7vZA", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Hai Chen", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Kexuan Shi", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Intern" }, { "name": "Leheng Zhang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Shuhang Gu", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 72, "n_formula_1": 26, "n_ref_uni": 39, "n_ref": 132, "n_ref_all": 158, "n_fig": 7, "n_tab": 9, "L_tab": 1930, "n_element_tab": 244, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 191, "n_element_tab_1": 43, "formula_len_all": 7820, "formula_len_all_1": 2074, "len_all": 172193, "len_all_1": 68402, "len_abs": 1170, "len_title": 130, "len_sents": 50201, "len_sents_1": 30503, "n_sents": 436, "n_sents_1": 235, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1175, "title": "INDUCTIVE GRADIENT ADJUSTMENT FOR SPECTRAL BIAS IN IMPLICIT NEURAL REPRESENTATIONS", "abs": "Implicit Neural Representations (INRs) as a versatile representation paradigm have achieved success in various computer vision tasks. Due to the spectral bias of the vanilla multi-layer perceptrons (MLPs), existing methods focus on designing MLPs with sophisticated architectures or repurposing existing training techniques for highly accurate INRs. In this paper, we delve into the linear dynamics model of MLPs and theoretically identify the empirical Neural Tangent Kernel (eNTK) matrix as a reliable link between spectral bias and training dynamics. Based on eNTK matrix, we propose a practical inductive gradient adjustment method, which could purposefully improve the spectral bias via inductive generalization of eNTK-based gradient transformation matrix. We evaluate our method on different INRs tasks with various INR architectures and compare to existing training techniques. The superiority representation performance clearly validate the advantage of our proposed method. Armed with our gradient adjustment method, better INRs with more enhanced texture details and sharpened edges can be learned from the training data by tailored improvements on spectral bias.", "keywords": [ "Implicit neural representation", "Spectral bias", "Training dynamics" ], "rating_list": [ 8, 3, 5, 3 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 1, 2, 1 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "TMutFl74tX", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Atsutoshi Kumagai", "gender": "Male", "institution": "NTT", "country": "JP", "position": "Researcher" }, { "name": "Taishi Nishiyama", "gender": "Male", "institution": "NTT", "country": "", "position": "Researcher" }, { "name": "Tomoharu Iwata", "gender": "Male", "institution": "NTT", "country": "", "position": "Researcher" }, { "name": "Yasuhiro Fujiwara", "gender": "Male", "institution": "NTT", "country": "JP", "position": "Researcher" }, { "name": "Yasutoshi Ida", "gender": "Male", "institution": "NTT", "country": "JP", "position": "Researcher" } ], "n_formula": 49, "n_formula_1": 18, "n_ref_uni": 57, "n_ref": 152, "n_ref_all": 218, "n_fig": 12, "n_tab": 12, "L_tab": 4068, "n_element_tab": 657, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2031, "n_element_tab_1": 250, "formula_len_all": 3118, "formula_len_all_1": 1834, "len_all": 227612, "len_all_1": 83235, "len_abs": 1603, "len_title": 115, "len_sents": 72360, "len_sents_1": 35093, "n_sents": 581, "n_sents_1": 279, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1427, "title": "Meta-learning Representations for Learning from Multiple Annotators", "abs": "We propose a meta-learning method for learning from multiple noisy annotators. In many applications such as crowdsourcing services,\nlabels for supervised learning are given by multiple annotators. Since the annotators have different skills or biases, given labels can be noisy. To learn accurate classifiers, existing methods require many annotated data to deal with noisy labels. However, sufficient data might be unavailable in practice. To overcome the lack of data, the proposed method uses labeled data obtained in different but related tasks. The proposed method embeds each example in tasks to a latent space by using a neural network and constructs a probabilistic model for learning a task-specific classifier while estimating annotators' abilities on the latent space. This neural network is meta-learned to improve the expected test classification performance when the classifier is adapted to a given small amount of annotated data. This classifier adaptation is performed by maximizing the posterior probability via the expectation-maximization (EM) algorithm. Since each step in the EM algorithm is easily computed as a closed-form and is differentiable, the proposed method can efficiently backpropagate the loss through the EM algorithm to meta-learn the neural network. We demonstrate the effectiveness of the proposed method with real-world datasets with synthetic noise and real-world crowdsourcing datasets.", "keywords": [ "Meta-learning", "Few-shot learning", "Multiple annotators", "Noisy labels", "Crowdsourcing" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "TMYe4rUuTc", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Qijun Tan", "gender": "unknown", "institution": "Google", "country": "", "position": "software engineer" }, { "name": "Rishabh Agarwal", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Adjunct Professor" }, { "name": "Rishabh Joshi", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Sarmishta Velury", "gender": "Female", "institution": "Google DeepMind", "country": "US", "position": "Software Engineer" }, { "name": "Tianqi Liu", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Software Engineer" }, { "name": "Yuan Liu", "gender": "unknown", "institution": "Research, Google", "country": "", "position": "Researcher" }, { "name": "Ziyu Ye", "gender": "unknown", "institution": "Google DeepMind", "country": "US", "position": "Research Intern" } ], "n_formula": 54, "n_formula_1": 18, "n_ref_uni": 90, "n_ref": 261, "n_ref_all": 319, "n_fig": 16, "n_tab": 24, "L_tab": 11374, "n_element_tab": 697, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2176, "n_element_tab_1": 166, "formula_len_all": 4311, "formula_len_all_1": 1384, "len_all": 432255, "len_all_1": 65525, "len_abs": 1210, "len_title": 151, "len_sents": 140851, "len_sents_1": 28091, "n_sents": 1059, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1341, "title": "Evolving Alignment via Asymmetric Self-Play", "abs": "Current RLHF approaches for aligning large language models (LLMs) typically assume a fixed prompt distribution, which is sub-optimal and limits the generalization capabilities for language models. To address this issue, we introduce a general framework that casts alignment as an asymmetric game between two players: (i) a creator, which strategically generates informative prompt distributions using reward signals, and (ii) a solver, which learns to produce preferred responses on prompts produced by the creator.\n\nThis framework of Evolving Alignment via Asymmetric Self-Play (`eva`), results in a simple and efficient approach that can utilize any existing RLHF algorithm. `eva` achieves a new state of the art in widely adopted alignment benchmarks, without the need of any additional human crafted prompts, e.g., it can improve the win rate of finetuned gemma-2-9b-it on Arena-Hard from 51.6% to 60.1% with DPO, from 55.7% to 58.9% with SPPO, from 52.3% to 60.7% with SimPO, and from 54.8% to 60.3% with ORPO, surpassing its 27B version and matching Claude-3-opus. Finally, we show `eva` is effective and robust under various ablation settings. \n\nWe hope `eva` can serve as a scalable and easy-to-use methodology for the research community to build open-ended, robust, and self-improving language agents, that align with human values.", "keywords": [ "large language model", "RLHF", "open-ended learning", "alignment" ], "rating_list": [ 5, 3, 8 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "TMHOHRR0FA", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Chaolong Ying", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Chenguang Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Tianshu Yu", "gender": "Male", "institution": "Chinese University of Hong Kong (Shenzhen)", "country": "CN", "position": "Assistant Professor" }, { "name": "Xuanhao Pan", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "YE XUE", "gender": "Female", "institution": "Shenzhen Research Institute of Big Data, CUHK(SZ\uff09", "country": "", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 8, "n_ref_uni": 35, "n_ref": 81, "n_ref_all": 103, "n_fig": 3, "n_tab": 12, "L_tab": 6906, "n_element_tab": 1075, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2685, "n_element_tab_1": 380, "formula_len_all": 734, "formula_len_all_1": 353, "len_all": 154811, "len_all_1": 61406, "len_abs": 1375, "len_title": 135, "len_sents": 46013, "len_sents_1": 25320, "n_sents": 312, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1400, "title": "Rethinking the \"Heatmap + Monte Carlo Tree Search'' Paradigm for Solving Large Scale TSP", "abs": "The Travelling Salesman Problem (TSP) remains a fundamental challenge in combinatorial optimization, inspiring diverse algorithmic strategies. This paper revisits the ``heatmap + Monte Carlo Tree Search (MCTS)\" paradigm that has recently gained traction for learning-based TSP solutions. Within this framework, heatmaps encode the likelihood of edges forming part of the optimal tour, and MCTS refines this probabilistic guidance to discover optimal solutions. Contemporary approaches have predominantly emphasized the refinement of heatmap generation through sophisticated learning models, inadvertently sidelining the critical role of MCTS. Our extensive empirical analysis reveals two pivotal insights: \\textbf{1}) The configuration of MCTS strategies profoundly influences the solution quality, demanding meticulous tuning to leverage their full potential; \\textbf{2}) Our findings demonstrate that a rudimentary and parameter-free heatmap, derived from the intrinsic $k$-nearest nature of TSP, can rival or even surpass the performance of complicated heatmaps, with strong generalizability across various scales. Empirical evaluations across various TSP scales underscore the efficacy of our approach, achieving competitive results. These observations challenge the prevailing focus on heatmap sophistication, advocating a reevaluation of the paradigm to harness both components synergistically.", "keywords": [ "Travelling Salesman Problem", "Heatmap", "Monte Carlo Tree Search", "Combinatorial optimization", "k-nearest neighbor" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "TLmibuPMyi", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Hu Ding", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Qingyuan Yang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Yuntao Wang", "gender": "Not Specified", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Yuxuan Li", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 27, "n_formula_1": 24, "n_ref_uni": 66, "n_ref": 80, "n_ref_all": 101, "n_fig": 5, "n_tab": 1, "L_tab": 141, "n_element_tab": 35, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2803, "formula_len_all_1": 2255, "len_all": 126032, "len_all_1": 62079, "len_abs": 1468, "len_title": 117, "len_sents": 29416, "len_sents_1": 26465, "n_sents": 260, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1483, "title": "An Efficient Algorithm For Computing Optimal Wasserstein Ball Center", "abs": "Wasserstein Barycenter (WB) is a fundamental problem in machine learning, whose objective is to find a representative probability measure that minimizes the sum of its Wasserstein distance to given distributions. WB has a number of applications in various areas. However, in some applications like model ensembling, where it aggregates predictions of different models on the label space, WB may lead to unfair outcome towards underrepresented groups (e.g., a \"minority'' distribution may be far away from the obtained WB under Wasserstein distance). To address this issue, we propose an alternative objective called ``Wasserstein Ball Center (WBC)''. Specifically, WBC is a distribution that encompasses all input distributions within the minimum Wasserstein distance, which can be formulated as a minmax optimization problem. We show that the WBC problem with fixed support is equivalent to solving a large-scale linear programming (LP) instance, which is quite different from the previous LP model for WB. By incorporating some novel observations on the induced normal equation, we propose an efficient algorithm that accelerates the interior point method by $O(Nm)$ times ($N$ is the number of distributions and $m$ is the support size). Finally, we conduct a set of experiments on both synthetic and real-world datasets. We demonstrate the computational efficiency of our algorithm, and showcase its better accuracy on model ensembling under heterogeneous data distributions.", "keywords": [ "Wasserstein barycenter", "model ensembling", "fairness" ], "rating_list": [ 3, 3, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 3, 4 ], "presentation_list": [ 1, 1, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "TLgDQ0Rr2Z", "primary_area": "causal reasoning", "n_author": 7, "author_info_list": [ { "name": "Haoxuan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Yue Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zeyu Tang", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zhi Geng", "gender": "Male", "institution": "School of Matematics asn Statistics, Beijing Technology and Business University", "country": "CN", "position": "Full Professor" }, { "name": "Zhichao Jiang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "Full Professor" }, { "name": "Zhuangyan Fang", "gender": "Male", "institution": "Xiaomi ", "country": "", "position": "Researcher" } ], "n_formula": 34, "n_formula_1": 17, "n_ref_uni": 29, "n_ref": 71, "n_ref_all": 83, "n_fig": 0, "n_tab": 1, "L_tab": 1169, "n_element_tab": 104, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4824, "formula_len_all_1": 2237, "len_all": 113853, "len_all_1": 67757, "len_abs": 1565, "len_title": 81, "len_sents": 38319, "len_sents_1": 33866, "n_sents": 251, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 33, "L_abs": 1570, "title": "Principle Counterfactual Fairness", "abs": "Fairness in human and algorithmic decision-making is crucial in areas such as criminal justice, education, and social welfare. Recently, counterfactual fairness has drawn increasing research interest, suggesting that decision-making for individuals should remain the same when intervening with different values on the protected attributes. Nevertheless, the question of \"which attributes and individuals should be protected\" is rarely discussed in the existing counterfactual fairness literature. For example, when considering leg disability as a protected attribute, the algorithms should not treat individuals with leg disabilities differently in college admissions, but one may naturally take into this factor for the purpose of selecting runner athletes. In other words, when and how to enforce fairness is expected to depend on the causal relation between the protected attribute and the outcome of interest. Formally, this paper proposes principal counterfactual fairness using the concept of principal stratification from the causal inference literature, focusing on whether an algorithm is counterfactually fair for individuals whose protected attribute has no individual causal effect on the outcome of interest. To examine whether an algorithm satisfies principal counterfactual fairness, we derive the statistical bounds, and propose a post-processing approach to achieving principal counterfactual fairness with minimal individual decision changes. Experiments are conducted using synthetic and real-world datasets to verify the effectiveness of our methods.", "keywords": [ "Counterfactual Fairness" ], "rating_list": [ 1, 5, 6, 5, 5 ], "soundness_list": [ 1, 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3, 3 ], "contribution_list": [ 1, 2, 3, 2, 2 ], "confidence_list": [ 4, 2, 3, 2, 4 ] }, { "paper_id": "TLWbNfbkxj", "primary_area": "applications to robotics, autonomy, planning", "n_author": 9, "author_info_list": [ { "name": "Chenxuan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Chuyan Xiong", "gender": "Female", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Guanqun Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Postdoc" }, { "name": "Jiaming Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jiaxin Ge", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Kaichen Zhou", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Sixiang Chen", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Undergrad student" }, { "name": "Shanghang Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaoqi Cloris Li", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 47, "n_ref": 146, "n_ref_all": 175, "n_fig": 7, "n_tab": 4, "L_tab": 1117, "n_element_tab": 77, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 28, "formula_len_all_1": 28, "len_all": 196965, "len_all_1": 72549, "len_abs": 1851, "len_title": 132, "len_sents": 56066, "len_sents_1": 36555, "n_sents": 404, "n_sents_1": 257, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1777, "title": "Self-Corrected Multimodal Large Language Model for Robot Manipulation and Reflection", "abs": "Multimodal Large Language Models (MLLMs) have demonstrated potential in visual instruction following across various tasks. Recently, some studies have integrated MLLMs into robotic manipulation, allowing robots to interpret multimodal information and predict low-level actions. While MLLM-based policies have shown promising progress, they may predict failure execution poses when faced with novel tasks or categories. To emulate human-like reasoning modes for more robust manipulation, we propose a Self-Corrected (SC)-MLLM. Our model combines fast system reasoning for directly predicting end-effector poses with slow system reasoning for reflecting on and correcting failure actions. For the fast system, we introduce parameter-efficient fine-tuning to empower MLLM with pose prediction capabilities, reframing this as a language modeling problem. For the slow system, when facing execution failures, our model learns to detect the causes of low-level action errors (i.e., position and rotation errors) and adaptively seeks prompt feedback from experts. Based on the feedback, SC-MLLM reflects on the current failure case and attempts to generate the corrected actions. Furthermore, we design a continuous policy learning method using successfully corrected samples, enhancing the model's adaptability to the current scene configuration and reducing the frequency of expert intervention. To evaluate our method, we conduct extensive experiments in both simulation and real-world settings. SC-MLLM significantly improves manipulation accuracy compared to previous state-of-the-art MLLM-based policy (ManipLLM), increasing from 57\\% to 79\\% on seen object categories and from 47\\% to 69\\% on unseen novel categories. Our project web page: https://sites.google.com/view/sc-mllm", "keywords": [ "Robot Manipulation", "Pose Correction", "Multimodal Large Language Model" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "TKuYWeFE6S", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Andr\u00e9 Hottung", "gender": "unknown", "institution": "Bielefeld University", "country": "DE", "position": "Postdoc" }, { "name": "Kevin Tierney", "gender": "Male", "institution": "Bielefeld University", "country": "DE", "position": "Full Professor" }, { "name": "Mridul Mahajan", "gender": "unknown", "institution": "Max Planck Institute for Intelligent Systems", "country": "", "position": "Intern" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 51, "n_ref": 83, "n_ref_all": 100, "n_fig": 7, "n_tab": 7, "L_tab": 3810, "n_element_tab": 241, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1212, "n_element_tab_1": 35, "formula_len_all": 215, "formula_len_all_1": 195, "len_all": 149508, "len_all_1": 57786, "len_abs": 1046, "len_title": 132, "len_sents": 42960, "len_sents_1": 28520, "n_sents": 332, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1050, "title": "PolyNet: Learning Diverse Solution Strategies for Neural Combinatorial Optimization", "abs": "Reinforcement learning-based methods for constructing solutions to combinatorial optimization problems are rapidly approaching the performance of human-designed algorithms. To further narrow the gap, learning-based approaches must efficiently explore the solution space during the search process. Recent approaches artificially increase exploration by enforcing diverse solution generation through handcrafted rules, however, these rules can impair solution quality and are difficult to design for more complex problems. In this paper, we introduce PolyNet, an approach for improving exploration of the solution space by learning complementary solution strategies. In contrast to other works, PolyNet uses only a single-decoder and a training schema that does not enforce diverse solution generation through handcrafted rules. We evaluate PolyNet on four combinatorial optimization problems and observe that the implicit diversity mechanism allows PolyNet to find better solutions than approaches that explicitly enforce diverse solution generation.", "keywords": [ "neural combinatorial optimization", "learning to optimize", "reinforcement learning", "routing problems" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "TKRIRI9tQv", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Baturalp Yalcin", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Eduardo Sontag", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Full Professor" }, { "name": "Haixiang Zhang", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Javad Lavaei", "gender": "unknown", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" } ], "n_formula": 28, "n_formula_1": 23, "n_ref_uni": 22, "n_ref": 25, "n_ref_all": 31, "n_fig": 3, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 73, "n_element_tab_1": 4, "formula_len_all": 1577, "formula_len_all_1": 1403, "len_all": 102501, "len_all_1": 69998, "len_abs": 979, "len_title": 110, "len_sents": 37104, "len_sents_1": 34034, "n_sents": 299, "n_sents_1": 280, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 987, "title": "Exact Recovery Guarantees for Parameterized Nonlinear System Identification Problem under Adversarial Attacks", "abs": "In this work, we study the system identification problem for parameterized nonlinear systems using basis functions under adversarial attacks. Motivated by the LASSO-type estimators, we analyze the exact recovery property of a nonsmooth estimator, which is generated by solving an embedded $\\ell_1$-loss minimization problem. First, we derive necessary and sufficient conditions for the well-specifiedness of the estimator and the uniqueness of global solutions to the underlying optimization problem. Next, we provide exact recovery guarantees for the estimator under two different scenarios of boundedness and Lipschitz continuity of the basis functions. The non-asymptotic exact recovery is guaranteed with high probability, even when there are more severely corrupted data than clean data. Finally, we numerically illustrate the validity of our theory. This is the first study on the sample complexity analysis of a nonsmooth estimator for the nonlinear system identification problem.", "keywords": [ "System identification", "robust control", "exact recovery" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "TKAzF69Fcv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Kui Jiang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Mengshun Hu", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Zheng Wang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Zhihang Zhong", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Xiao Sun", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 14, "n_ref_uni": 27, "n_ref": 55, "n_ref_all": 75, "n_fig": 11, "n_tab": 3, "L_tab": 641, "n_element_tab": 140, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3895, "n_element_tab_1": 285, "formula_len_all": 1356, "formula_len_all_1": 1361, "len_all": 105290, "len_all_1": 67003, "len_abs": 1499, "len_title": 128, "len_sents": 30820, "len_sents_1": 25846, "n_sents": 235, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1505, "title": "Aligning Anything: Hierarchical Motion Estimation for Video Frame Interpolation", "abs": "Existing advanced video frame interpolation (VFI) methods struggle to learn accurate per-pixel motion or target-level motion. The reasons lie in that pixel-level motion estimation allows for infinite possibilities, making it challenging to guarantee fitting accuracy and global motion consistency, especially for rigid objects. Conversely, target-level motion consistency from the same moving target also breaks\ndown when the assumption of object rigidity no longer holds. Therefore, a hierarchical motion learn scheme is imperative to promote the accuracy and stability of motion prediction. Specifically, we marry the target-level motion to the pixel-level motion to form the hierarchical motion estimation. It elaborately introduces specific semantics priors from open-world knowledge models such as the Recognize Anything Model (RAM), Grounding DIDO, and the High-Quality Segment Anything Model (HQ-SAM) to facilitate the latent target-level motion learning. In particular, a hybrid contextual feature extraction module (HCE) is employed to aggregate both pixel-wise and semantic representations, followed by the hierarchical motion and feature interactive refinement module (HIR) to simulate the current motion patterns. When integrating these adaptions to existing SOTA VFI methods, more consistent motion estimation and interpolation are predicted. Extensive experiments show that advanced VFI networks plugged with our adaptions can achieve more superior performances on various benchmark datasets", "keywords": [ "Video frame interpolation; hierarchical motion estimation; pixel-level; target-level" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 2, 5, 3 ] }, { "paper_id": "TJp3LnQgSX", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Jizhen Li", "gender": "unknown", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Weiping Tu", "gender": "Female", "institution": "Wuhan University", "country": "", "position": "Full Professor" }, { "name": "Xinmeng Xu", "gender": "Male", "institution": "Wuhan University", "country": "", "position": "PhD student" }, { "name": "Yiqun Zhang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Yong Luo", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Professor" }, { "name": "Yuhong Yang", "gender": "unknown", "institution": "Wuhan University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 9, "n_ref_uni": 53, "n_ref": 99, "n_ref_all": 133, "n_fig": 14, "n_tab": 13, "L_tab": 6095, "n_element_tab": 644, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1376, "n_element_tab_1": 146, "formula_len_all": 903, "formula_len_all_1": 462, "len_all": 201594, "len_all_1": 61876, "len_abs": 1632, "len_title": 120, "len_sents": 54528, "len_sents_1": 28929, "n_sents": 371, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1599, "title": "FIRING-Net: A filtered feature recycling network for speech enhancement", "abs": "Current deep neural networks for speech enhancement (SE) aim to minimize the distance between the output signal and the clean target by filtering out noise features from input features. However, when noise and speech components are highly similar, SE models struggle to learn effective discrimination patterns. To address this challenge, we propose a Filter-Recycle-Interguide framework termed Filter-Recycle-INterGuide NETwork (FIRING-Net) for SE, which filters the input features to extract target features and recycles the filtered-out features as non-target features. These two feature sets then guide each other to refine the features, leading to the aggregation of speech information within the target features and noise information within the non-target features. The proposed FIRING-Net mainly consists of a Local Module (LM) and a Global Module (GM). The LM uses outputs of the speech extraction network as target features and the residual between input and output as non-target features. The GM leverages the energy distribution of self-attention map to extract target and non-target features guided by highest and lowest energy regions. Both LM and GM include interaction modules to leverage the two feature sets in an inter-guided manner for collecting speech from non-target features and filtering out noise from target features. Experiments confirm the effectiveness of the Filter-Recycle-Interguide framework, with FIRING-Net achieving a strong balance between SE performance and computational efficiency, surpassing comparable models across various SNR levels and noise environments.", "keywords": [ "speech enhancement", "target and non-target features", "self-attention" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 2, 4 ], "presentation_list": [ 3, 2, 3, 2, 4 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 5, 2, 4 ] }, { "paper_id": "TJo6aQb7mK", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Aaryan Bhagat", "gender": "Male", "institution": "University of California, Riverside", "country": "", "position": "MS student" }, { "name": "Arnab Kumar Mondal", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Ayush Kaushal", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Researcher" }, { "name": "Irina Rish", "gender": "Female", "institution": "University of Montreal", "country": "CA", "position": "Professor" }, { "name": "Tejas Pandey", "gender": "Male", "institution": "Indian Institute of Technology Kharagpur", "country": "IN", "position": "Undergrad student" }, { "name": "Tejas Vaidhya", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 3, "n_ref_uni": 82, "n_ref": 151, "n_ref_all": 206, "n_fig": 24, "n_tab": 17, "L_tab": 4206, "n_element_tab": 694, "n_fig_1": 12, "n_tab_1": 3, "L_tab_1": 533, "n_element_tab_1": 66, "formula_len_all": 839, "formula_len_all_1": 333, "len_all": 327985, "len_all_1": 64018, "len_abs": 1413, "len_title": 120, "len_sents": 74412, "len_sents_1": 31589, "n_sents": 582, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 72, "L_abs": 1419, "title": "Surprising Effectiveness of pretraining Ternary Language Model at Scale", "abs": "Rapid advancements in GPU computational power has outpaced memory capacity and bandwidth growth, creating bottlenecks in Large Language Model (LLM) inference. Post-training quantization is the leading method for addressing memory-related bottlenecks in LLM inference, but it suffers from significant performance degradation below 4-bit precision. This paper addresses these challenges by investigating the pretraining of low-bitwidth models specifically Ternary Language Models (TriLMs) as an alternative to traditional floating-point models (FloatLMs) and their post-training quantized versions (QuantLMs). We present Spectra LLM suite, the first open suite of LLMs spanning multiple bit-widths, including FloatLMs, QuantLMs, and TriLMs, ranging from 99M to 3.9B parameters trained on 300B tokens. Our comprehensive evaluation demonstrates that TriLMs offer superior scaling behavior in terms of model size (in bits). Surprisingly, at scales exceeding one billion parameters, TriLMs consistently outperform their QuantLM and FloatLM counterparts for a given bit size across various benchmarks. Notably, the 3.9B parameter TriLM matches the performance of the FloatLM 3.9B across all benchmarks, despite having fewer bits than FloatLM 830M. Overall, this research provides valuable insights into the feasibility and scalability of low-bitwidth language models, paving the way for the development of more efficient LLMs.", "keywords": [ "Large Language Models", "low-bit language models", "quantization-aware training", "pretraining of large language models", "and scaling laws" ], "rating_list": [ 5, 5, 10, 10, 8 ], "soundness_list": [ 3, 2, 4, 4, 4 ], "presentation_list": [ 2, 2, 3, 4, 4 ], "contribution_list": [ 2, 3, 4, 4, 4 ], "confidence_list": [ 5, 4, 3, 4, 3 ] }, { "paper_id": "TJU9J8iQXL", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Bingde Hu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" }, { "name": "Haoze Jiang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Huiqiong Wang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Lecturer" }, { "name": "Jiacong Hu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Mingli Song", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yi Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yuanyu Wan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Zunlei Feng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 48, "n_ref": 59, "n_ref_all": 86, "n_fig": 11, "n_tab": 3, "L_tab": 1838, "n_element_tab": 121, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 1130, "n_element_tab_1": 85, "formula_len_all": 446, "formula_len_all_1": 492, "len_all": 139225, "len_all_1": 64276, "len_abs": 1928, "len_title": 82, "len_sents": 46707, "len_sents_1": 32899, "n_sents": 283, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 34, "L_abs": 1935, "title": "Is the Fairness Metric Truly Fair?", "abs": "Image classification is a fundamental task in computer vision that has been widely adopted in critical applications such as face recognition and medical imaging, drawing considerable attention to its predictive fairness. Some researchers have proposed various fairness metrics and pipelines to enhance the fairness of deep learning models. However, recent studies indicate that existing fairness evaluation specifications and metrics have inherent flaws, as they focus on low-dimensional inputs, such as numerical data, and overlook partial correlations between target and sensitive attributes, leading to some degree of mutual exclusivity. This raises the question: Is the fairness metric truly fair? Through in-depth analysis, experiments conclude that the fairness of deep models is closely related to attribute sampling and the interdependencies among attributes. In this work, we address this challenge by introducing a new specification based on dynamic perturbation for image classification models. Specifically, we introduce an Attribute Projection Perturbation Strategy (APPS) that moves beyond the constraints of directly statistical discrete predictions by mapping sensitive attributes that may influence task attributes onto the same dimension for evaluation. Building on this, a Projection Fairness Metric System is proposed to quantifing the upper and lower bounds of fairness perturbations, examining and evaluating the impact of mapped sensitive attributes on the fairness of task predictions from different perspectives. Additionally, we conducted systematic evaluation experiments and extensive discussions, demonstrating that the proposed evaluation specification offers better objectivity and interpretability compared to existing metrics, in 24 image classification models including CNN and ViT architectures. It is hoped that this work will promote the standardization of fairness evaluation pipeline and metrics.", "keywords": [ "Fairness", "Evaluation", "Image Classification" ], "rating_list": [ 3, 1, 3 ], "soundness_list": [ 2, 1, 1 ], "presentation_list": [ 3, 3, 1 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "TJHB4ySVZM", "primary_area": "generative models", "n_author": 1, "author_info_list": [ { "name": "Senmao Ye", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 46, "n_ref": 65, "n_ref_all": 76, "n_fig": 4, "n_tab": 7, "L_tab": 1202, "n_element_tab": 184, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2025, "n_element_tab_1": 233, "formula_len_all": 535, "formula_len_all_1": 369, "len_all": 106324, "len_all_1": 62691, "len_abs": 1086, "len_title": 114, "len_sents": 24765, "len_sents_1": 25040, "n_sents": 219, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1094, "title": "Data Extrapolation for Text-to-image Generation on Small Datasets", "abs": "Text-to-image generation requires large amount of training data to synthesizing high-quality images. For augmenting training data, previous methods rely on data interpolations like cropping, flipping, and mixing up, which fail to introduce new information and yield only marginal improvements. In this paper, we propose a new data augmentation method for text-to-image generation using linear extrapolation. Specifically, we apply linear extrapolation only on text feature, and new image data are retrieved from the internet by search engines. For the reliability of new text-image pairs, we design two outlier detectors to purify retrieved images. Based on extrapolation, we construct training samples dozens of times larger than the original dataset, resulting in a significant improvement in text-to-image performance. Moreover, we propose a NULL-guidance to refine score estimation, and apply recurrent affine transformation to fuse text information. Our model achieves FID scores of 7.91, 9.52 and 5.00 on the CUB, Oxford and COCO datasets. The code and data will be available on GitHub.", "keywords": [ "Diffuison", "Text-to-image\uff0cData augmentation" ], "rating_list": [ 3, 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 1, 1 ], "presentation_list": [ 2, 3, 2, 3, 1 ], "contribution_list": [ 2, 1, 1, 1, 2 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "TIjBKgLyPN", "primary_area": "interpretability and explainable AI", "n_author": 1, "author_info_list": [ { "name": "Kola Ayonrinde", "gender": "unknown", "institution": "UK AISI", "country": "GB", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 29, "n_ref": 56, "n_ref_all": 88, "n_fig": 7, "n_tab": 6, "L_tab": 635, "n_element_tab": 68, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 515, "n_element_tab_1": 53, "formula_len_all": 410, "formula_len_all_1": 343, "len_all": 170198, "len_all_1": 60848, "len_abs": 1635, "len_title": 83, "len_sents": 41999, "len_sents_1": 28468, "n_sents": 314, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1643, "title": "Adaptive Sparse Allocation with Mutual Choice \\& Feature Choice Sparse Autoencoders", "abs": "Sparse autoencoders (SAEs) are a promising approach to extracting features from neural networks, enabling model interpretability as well as causal interventions on model internals. SAEs generate sparse feature representations using a sparsifying activation function that implicitly defines a set of token-feature matches. We frame the token-feature matching as a resource allocation problem constrained by a total sparsity upper bound. For example, TopK SAEs solve this allocation problem with the additional constraint that each token matches with at most k features. In TopK SAEs, the k active features per token constraint is the same across tokens, despite some tokens being more difficult to reconstruct than others. To address this limitation, we propose two novel SAE variants, Feature Choice SAEs and Mutual Choice SAEs, which each allow for a variable number of active features per token. Feature Choice SAEs solve the sparsity allocation problem under the additional constraint that each feature matches with at most m tokens. Mutual Choice SAEs solve the unrestricted allocation problem where the total sparsity budget can be allocated freely between tokens and features. Additionally, we introduce a new auxiliary loss function, aux_zipf_loss, which generalises the aux_k_loss to mitigate dead and underutilised features. Our methods result in SAEs with fewer dead features and improved reconstruction loss at equivalent sparsity levels as a result of the inherent adaptive computation. More accurate and scalable feature extraction methods provide a path towards better understanding and more precise control of foundation models.", "keywords": [ "Mechanistic Interpretability", "Sparse Autoencoders" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "TId1SHe8JG", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Aravind Gollakota", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Charlotte Peale", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" }, { "name": "Gustaf Ahdritz", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Parikshit Gopalan", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Udi Wieder", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" } ], "n_formula": 151, "n_formula_1": 22, "n_ref_uni": 32, "n_ref": 104, "n_ref_all": 135, "n_fig": 4, "n_tab": 4, "L_tab": 429, "n_element_tab": 32, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 207, "n_element_tab_1": 12, "formula_len_all": 12817, "formula_len_all_1": 1310, "len_all": 312542, "len_all_1": 82312, "len_abs": 1736, "len_title": 111, "len_sents": 108107, "len_sents_1": 36504, "n_sents": 987, "n_sents_1": 314, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 63, "L_abs": 1362, "title": "Provable Uncertainty Decomposition via Higher-Order Calibration", "abs": "We give a principled method for decomposing the predictive uncertainty of a model into aleatoric and epistemic components with explicit semantics relating them to the real-world data distribution. While many works in the literature have proposed such decompositions, they lack the type of formal guarantees we provide. Our method is based on the new notion of higher-order calibration, which generalizes ordinary calibration to the setting of higher-order predictors that predict _mixtures_ over label distributions at every point. We show how to measure as well as achieve higher-order calibration using access to $k$-snapshots, namely examples where each point has $k$ independent conditional labels. Under higher-order calibration, the estimated aleatoric uncertainty at a point is guaranteed to match the real-world aleatoric uncertainty averaged over all points where the prediction is made. To our knowledge, this is the first formal guarantee of this type that places no assumptions whatsoever on the real-world data distribution. Importantly, higher-order calibration is also applicable to existing higher-order predictors such as Bayesian and ensemble models and provides a natural evaluation metric for such models. We demonstrate through experiments that our method produces meaningful uncertainty decompositions in tasks such as image classification.", "keywords": [ "uncertainty quantification", "calibration", "trustworthy ML", "mixture learning" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 4, 4, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 2, 3, 5, 4 ] }, { "paper_id": "THqWPzL00e", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Apurva Ratan Murty", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Mainak Deb", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Intern" }, { "name": "Mayukh Deb", "gender": "Male", "institution": "Amrita Vishwa Vidyapeetham", "country": "", "position": "Undergrad student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 53, "n_ref": 102, "n_ref_all": 132, "n_fig": 16, "n_tab": 2, "L_tab": 410, "n_element_tab": 24, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 294, "n_element_tab_1": 3, "formula_len_all": 287, "formula_len_all_1": 258, "len_all": 154832, "len_all_1": 61762, "len_abs": 1484, "len_title": 128, "len_sents": 49956, "len_sents_1": 29755, "n_sents": 383, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 79, "L_abs": 1295, "title": "TopoNets: High performing vision and language models with brain-like topography", "abs": "Neurons in the brain are organized such that nearby cells tend to share similar functions. AI models lack this organization, and past efforts to introduce topography have often led to trade-offs between topography and task performance. In this work, we present *TopoLoss*, a new loss function that promotes spatially organized topographic representations in AI models without significantly sacrificing task performance. TopoLoss is highly adaptable and can be seamlessly integrated into the training of leading model architectures. We validate our method on both vision (ResNet-18, ResNet-50, ViT) and language models (GPT-Neo-125M, NanoGPT), collectively *TopoNets*. TopoNets are the highest performing supervised topographic models to date, exhibiting brain-like properties such as localized feature processing, lower dimensionality, and increased efficiency. TopoNets also predict responses in the brain and replicate the key topographic signatures observed in the brain\u2019s visual and language cortices, further bridging the gap between biological and artificial systems. This work establishes a robust and generalizable framework for integrating topography into AI, advancing the development of high performing models that more closely emulate the computational strategies of the human brain.", "keywords": [ "topography", "neuro-inspired", "convolutional neural networks", "Transformers", "visual cortex", "neuroscience" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "THkF3VWSNv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Andrey Kan", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Bhavana Bhasker", "gender": "Female", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Laurent Callot", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Muhammad Shihab Rashid", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Shuo Li", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Timothy B Esler", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 29, "n_ref": 60, "n_ref_all": 92, "n_fig": 15, "n_tab": 8, "L_tab": 4045, "n_element_tab": 652, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 2886, "n_element_tab_1": 293, "formula_len_all": 10, "formula_len_all_1": 0, "len_all": 147146, "len_all_1": 68295, "len_abs": 1257, "len_title": 110, "len_sents": 44570, "len_sents_1": 30471, "n_sents": 360, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1260, "title": "REDO: Execution-Free Runtime Error Detection for Coding Agents", "abs": "As LLM-based agents exhibit exceptional capabilities in addressing complex problems, there is a growing focus on developing coding agents to tackle increasingly sophisticated tasks. Despite their promising performance, these coding agents often produce programs or modifications that contain runtime errors, which can cause code failures and are difficult for static analysis tools to detect. Enhancing the ability of coding agents to statically identify such errors could significantly improve their overall performance. In this work, we introduce Execution-free Runtime Error Detection for COding Agents (REDO), a method that integrates LLMs with static analysis tools to detect runtime errors for coding agents, without code execution. Additionally, we propose a benchmark task, SWE-Bench-Error-Detection (SWEDE), based on SWE-Bench (lite), to evaluate error detection in repository-level problems with complex external dependencies. Finally, through both quantitative and qualitative analyses across various error detection tasks, we demonstrate that REDO outperforms current state-of-the-art methods by achieving a 11.0% higher accuracy and 9.1% higher weighted F1 score; and provide insights into the advantages of incorporating LLMs for error detection.", "keywords": [ "error detection", "coding agents", "runtime", "execution-free", "large language models" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "THSm9HyCKo", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Michael K. Chen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Intern" }, { "name": "Xikun ZHANG", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 27, "n_ref": 53, "n_ref_all": 79, "n_fig": 5, "n_tab": 6, "L_tab": 719, "n_element_tab": 81, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1309, "n_element_tab_1": 100, "formula_len_all": 300, "formula_len_all_1": 300, "len_all": 129740, "len_all_1": 63166, "len_abs": 1949, "len_title": 111, "len_sents": 43905, "len_sents_1": 28414, "n_sents": 378, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1945, "title": "JustLogic: A benchmark for natural language deductive reasoning", "abs": "Logical reasoning is a critical component of Large Language Models (LLMs), and substantial research efforts in recent years have aimed to enhance their deductive capabilities. However, existing deductive reasoning benchmarks, which are crucial for evaluating and advancing LLMs, are inadequate due to their lack of task complexity, presence of prior knowledge as a confounder, and superficial error analysis. To address these deficiencies, we introduce JustLogic, a synthetically generated deductive reasoning benchmark designed for rigorous evaluation of LLMs. JustLogic is (i) highly complex, capable of generating a diverse range of linguistic patterns, vocabulary, and argument structures; (ii) context-independent, eliminating the advantage of models possessing prior knowledge and ensuring that only deductive reasoning is used to answer questions; and (iii) capable of in-depth error analysis on the heterogeneous effects of reasoning depth and argument form on model accuracy. Our experimental results on JustLogic reveal that the performance of most state-of-the-art (SOTA) LLMs, specifically Llama3-8B (57.8\\%), Llama3-70B (64.6\\%), and GPT-4o (65.6\\%), is significantly worse than the average human performance (73.0\\%). A recently released reasoning model, OpenAI o1-preview, performed substantially better, with an accuracy of 81.0\\%. However, it still lags behind the human ceiling of 100.0\\%. These results demonstrate that the JustLogic benchmark is realistic and achievable for both humans and models and that there is still substantial room for improvement in the deductive reasoning capabilities of LLMs. We posit that the use of context-dependent and relatively simplistic benchmarks has misrepresented the reasoning abilities of many SOTA models. We release our open-source dataset to provide accurate evaluations of model performance in deductive reasoning and to facilitate LLM advancement through in-depth error analysis.", "keywords": [ "benchmark", "logical reasoning", "LLM", "natural language processing (NLP)", "propositional logic" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "THOgGo8SX7", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Emile Timothy Anand", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Guannan Qu", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 166, "n_formula_1": 45, "n_ref_uni": 54, "n_ref": 102, "n_ref_all": 117, "n_fig": 6, "n_tab": 9, "L_tab": 943, "n_element_tab": 38, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 28078, "formula_len_all_1": 4602, "len_all": 258837, "len_all_1": 78736, "len_abs": 935, "len_title": 148, "len_sents": 65789, "len_sents_1": 31454, "n_sents": 668, "n_sents_1": 280, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1001, "title": "Efficient Reinforcement Learning for Global Decision Making in the Presence of Local Agents at Scale", "abs": "We study reinforcement learning for global decision-making in the presence of local agents, where the global decision-maker makes decisions affecting all local agents, and the objective is to learn a policy that maximizes the joint rewards of all the agents. Such problems find many applications, e.g. demand response, EV charging, and queueing. In this setting, scalability has been a long-standing challenge due to the size of the joint state space which can be exponential in the number of agents. This work proposes the \\texttt{SUBSAMPLE-Q} algorithm, where the global agent subsamples $k\\leq n$ local agents to compute a policy in time that is polynomial in $k$. We show that this learned policy converges to the optimal policy on the order of $\\tilde{O}(1/\\sqrt{k}+\\epsilon_{k,m})$ as the number of subsampled agents $k$ increases, where ${\\epsilon}_{k,m}$ is the Bellman noise. Finally, we validate our theoretical results through numerical simulations in demand-response and queueing settings.", "keywords": [ "Reinforcement Learning", "Multi-agent Systems", "Large-scale Systems", "Mean-field Approximation" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 2, 2 ] }, { "paper_id": "TH4gKbZS1E", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Chen Zeng", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "HaoranShen", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "Jiahui Wang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "PhD student" }, { "name": "Qiao Wang", "gender": "Male", "institution": "Southeast University", "country": "", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 20, "n_ref": 33, "n_ref_all": 50, "n_fig": 10, "n_tab": 4, "L_tab": 469, "n_element_tab": 93, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 469, "n_element_tab_1": 93, "formula_len_all": 338, "formula_len_all_1": 338, "len_all": 59685, "len_all_1": 36610, "len_abs": 1182, "len_title": 94, "len_sents": 15262, "len_sents_1": 15174, "n_sents": 122, "n_sents_1": 121, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1083, "title": "KAN versus MLP on Irregular or Noisy Functions", "abs": "In this paper, we compare the performance of Kolmogorov-Arnold Networks (KAN) and Multi-Layer Perceptron (MLP) networks on irregular or noisy functions. We control the number of parameters and the size of the training samples to ensure a fair comparison. For clarity, we categorize the functions into six types: regular functions, continuous functions with local non-differentiable points, functions with jump discontinuities, functions with singularities, functions with coherent oscillations, and noisy functions. Our experimental results indicate that KAN does not always perform best. For some types of functions, MLP outperforms or performs comparably to KAN. Furthermore, increasing the size of training samples can improve performance to some extent. When noise is added to functions, the irregular features are often obscured by the noise, making it challenging for both MLP and KAN to extract these features effectively. We hope these experiments provide valuable insights for future neural network research and encourage further investigations to overcome these challenges.", "keywords": [ "Kolmogorov-Arnold networks", "Multi-layer Perceptrons", "KAN", "MLP", "Irregularization", "Noise" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 2, 3, 1, 1 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 1, 1, 1, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "TEmE9PSC65", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Benjamin Vandersmissen", "gender": "Male", "institution": "University of Antwerp", "country": "", "position": "PhD student" }, { "name": "Jose Oramas", "gender": "Male", "institution": "University of Antwerp", "country": "BE", "position": "Associate Professor" }, { "name": "Lucas Deckers", "gender": "Male", "institution": "Universiteit Antwerpen", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 28, "n_ref": 39, "n_ref_all": 53, "n_fig": 6, "n_tab": 8, "L_tab": 2494, "n_element_tab": 230, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 2507, "n_element_tab_1": 229, "formula_len_all": 251, "formula_len_all_1": 251, "len_all": 98568, "len_all_1": 62698, "len_abs": 944, "len_title": 127, "len_sents": 33995, "len_sents_1": 28339, "n_sents": 246, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 945, "title": "Improving Neural Network Accuracy by Concurrently Training with a Twin Network", "abs": "Recently within Spiking Neural Networks, a method called Twin Network Augmentation (TNA) has been introduced. This technique claims to improve the validation accuracy of a Spiking Neural Network simply by training two networks in conjunction and matching the logits via the Mean Squared Error loss. In this paper, we validate the viability of this method on a wide range of popular Convolutional Neural Network (CNN) benchmarks and compare this approach to existing Knowledge Distillation schemes. Next, we conduct a in-depth study of the different components that make up TNA and determine that its effectiveness is not solely situated in an increase of trainable parameters, but rather the effect of the training methodology. Finally, we analyse the representations learned by networks trained with TNA and highlight their superiority in a number of tasks, thus proving empirically the applicability of Twin Network Augmentation on CNN models.", "keywords": [ "Knowledge Distillation", "Representation Learning", "Regularization" ], "rating_list": [ 8, 8, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "TEkoMEjf7E", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Gerhard Petrus Hancke", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Tengfei Wang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Zexin He", "gender": "Not Specified", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zhenwei Wang", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Ziwei Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Rynson Lau", "gender": "Male", "institution": "City University of Hong Kong", "country": "CN", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 37, "n_ref": 76, "n_ref_all": 110, "n_fig": 22, "n_tab": 5, "L_tab": 1568, "n_element_tab": 249, "n_fig_1": 14, "n_tab_1": 4, "L_tab_1": 1248, "n_element_tab_1": 211, "formula_len_all": 175, "formula_len_all_1": 175, "len_all": 153387, "len_all_1": 65034, "len_abs": 1127, "len_title": 170, "len_sents": 41045, "len_sents_1": 29351, "n_sents": 325, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 124, "L_abs": 1131, "title": "Phidias: A Generative Model for Creating 3D Content from Text, Image, and 3D Conditions with Reference-Augmented Diffusion", "abs": "Generative 3D modeling has made significant advances recently, but it remains constrained by its inherently ill-posed nature, leading to challenges in quality and controllability. Inspired by the real-world workflow that designers typically refer to existing 3D models when creating new ones, we propose Phidias, a novel generative model that uses diffusion for reference-augmented 3D generation. Given an image, our method leverages a retrieved or user-provided 3D reference model to guide the generation process, thereby enhancing the generation quality, generalization ability, and controllability. Phidias integrates three key components: 1) meta-ControlNet to dynamically modulate the conditioning strength, 2) dynamic reference routing to mitigate misalignment between the input image and 3D reference, and 3) self-reference augmentations to enable self-supervised training with a progressive curriculum. Collectively, these designs result in significant generative improvements over existing methods. Phidias forms a unified framework for 3D generation using text, image, and 3D conditions, offering versatile applications.", "keywords": [ "3D generation", "retrieval-augmented generation", "multi-view diffusion" ], "rating_list": [ 8, 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "TEjXRrhqtJ", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Hong Je-Gal", "gender": "Male", "institution": "Sejong University", "country": "KR", "position": "MS student" }, { "name": "Hyun-Suk Lee", "gender": "Male", "institution": "Sejong University", "country": "", "position": "Assistant Professor" } ], "n_formula": 21, "n_formula_1": 17, "n_ref_uni": 20, "n_ref": 35, "n_ref_all": 71, "n_fig": 16, "n_tab": 6, "L_tab": 1483, "n_element_tab": 93, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 834, "n_element_tab_1": 34, "formula_len_all": 908, "formula_len_all_1": 562, "len_all": 115313, "len_all_1": 63009, "len_abs": 1020, "len_title": 144, "len_sents": 43879, "len_sents_1": 30506, "n_sents": 306, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1024, "title": "TIEM: Enhancing Explanation of Video Prediction via Temporal Dynamics-Focused Dual Perturbation", "abs": "Explaining video data predictions is challenging due to the complex spatio-temporal information in videos. In particular, the existing perturbation-based methods for video interpretation often fail to consider different temporal contexts, making them ineffective for dynamic videos where the important regions change rapidly or appear ephemerally across frames. To address this, we propose a novel video interpretation method, time importance score-aware extremal perturbation masks (TIEM), that enhances explainability by focusing on temporal dynamics in videos. TIEM exploits a dual perturbation process: first, it evaluates temporal importance across frames via temporal perturbation and then generates spatio-temporal extremal perturbation masks using the temporal importance explicitly. Our experimental results demonstrate that TIEM resolves the key challenges of the existing methods, providing more precise explanations across the time domain in synthetic white-box models and black-box models for real-world videos.", "keywords": [ "XAI", "visual explanation", "extremal mask", "dual perturbation", "video prediction" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 1, 2 ], "presentation_list": [ 2, 2, 1, 1, 3 ], "contribution_list": [ 2, 2, 2, 1, 2 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "TEckHPheJg", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Bar Cavia", "gender": "Male", "institution": "Hebrew University, Hebrew University of Jerusalem", "country": "IL", "position": "Undergrad student" }, { "name": "Tal Reiss", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "PhD student" }, { "name": "Yedid Hoshen", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 53, "n_ref": 116, "n_ref_all": 131, "n_fig": 7, "n_tab": 4, "L_tab": 1631, "n_element_tab": 132, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1598, "n_element_tab_1": 137, "formula_len_all": 281, "formula_len_all_1": 281, "len_all": 151263, "len_all_1": 70788, "len_abs": 966, "len_title": 101, "len_sents": 42171, "len_sents_1": 33378, "n_sents": 363, "n_sents_1": 271, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 969, "title": "Wasn\u2019t Me: Enabling Users to Falsify Deepfake Attacks", "abs": "The rise of deepfake technology has made everyone vulnerable to false claims based on manipulated media. While many existing deepfake detection methods aim to identify fake media, they often struggle with deepfakes created by new generative models not seen during training. In this paper, we propose VeriFake, a method that enables users to verify that media claiming to show them are false. VeriFake is based on two key assumptions: (i) generative models struggle to exactly depict a specific identity, and (ii) they often fail to perfectly synchronize generated lip movements with speech. By combining these assumptions with powerful modern representation encoders, VeriFake achieves highly effective results, even against previously unseen deepfakes. Through extensive experiments, we demonstrate that VeriFake significantly outperforms general-purpose deepfake detection techniques despite being simple to implement and not relying on any fake data for pretraining.", "keywords": [ "deepfake detection", "deepfake verification" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "TE0KOzWYAF", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Rui Meng", "gender": "Male", "institution": "SalesForce Research", "country": "", "position": "Researcher" }, { "name": "Xinyi Yang", "gender": "Female", "institution": "Salesforce Research", "country": "US", "position": "Researcher" }, { "name": "Yingbo Zhou", "gender": "unknown", "institution": "Salesforce Research", "country": "", "position": "Research Scientist" }, { "name": "Ziyan Jiang", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Semih Yavuz", "gender": "Male", "institution": "SalesForce.com", "country": "", "position": "Research Scientist" }, { "name": "wenhu chen", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 69, "n_ref": 139, "n_ref_all": 157, "n_fig": 5, "n_tab": 6, "L_tab": 4736, "n_element_tab": 1031, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1444, "n_element_tab_1": 369, "formula_len_all": 612, "formula_len_all_1": 230, "len_all": 211686, "len_all_1": 53588, "len_abs": 1646, "len_title": 127, "len_sents": 44446, "len_sents_1": 21826, "n_sents": 365, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1307, "title": "VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks", "abs": "Embedding models have been crucial in enabling various downstream tasks such as semantic similarity, information retrieval, and clustering. Recently, there has been a surge of interest in developing universal text embedding models that can generalize across tasks (e.g., MTEB). However, progress in learning universal multimodal embedding models has been relatively slow despite their importance. In this work, we aim to explore the potential for building universal embeddings capable of handling a wide range of downstream tasks. Our contributions are twofold: (1) MMEB (Massive Multimodal Embedding Benchmark), which covers 4 meta-tasks and 36 datasets, including 20 training and 16 evaluation datasets, and (2) VLM2VEC (Vision-Language Model \u2192 Vector), a contrastive training framework that converts any state-of-the-art vision-language model into an embedding model. Unlike previous models such as CLIP and BLIP, VLM2VEC can process any combination of images and text to generate a fixed-dimensional vector based on task instructions. We build a series of VLM2VEC models on Phi-3.5-V and evaluate them on MMEB. Our results show that VLM2VEC achieves an absolute average improvement of 10% to 20% over existing multimodal embedding models on both in-distribution and out-of-distribution datasets in MMEB.", "keywords": [ "Vision Language Model", "Representation Learning", "Multimodal Embeddings" ], "rating_list": [ 8, 3, 5, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 4, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "TDzAqTqDHV", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Christos Baziotis", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "", "position": "PhD student" }, { "name": "Fabio Petroni", "gender": "Male", "institution": "Samaya AI", "country": "GB", "position": "Researcher" }, { "name": "Michele Bevilacqua", "gender": "Male", "institution": "Samaya AI", "country": "", "position": "Researcher" }, { "name": "Roberto Dessi", "gender": "Male", "institution": "Universitat Pompeu Fabra", "country": "ES", "position": "PhD student" }, { "name": "Silvia Sapora", "gender": "Female", "institution": "University of Oxford", "country": "", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 40, "n_ref": 51, "n_ref_all": 60, "n_fig": 3, "n_tab": 2, "L_tab": 820, "n_element_tab": 125, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1728, "n_element_tab_1": 146, "formula_len_all": 387, "formula_len_all_1": 258, "len_all": 122063, "len_all_1": 62622, "len_abs": 1343, "len_title": 86, "len_sents": 35137, "len_sents_1": 29096, "n_sents": 242, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 38, "L_abs": 1350, "title": "QCR: Quantised Codebooks for Retrieval", "abs": "In recent years, the application of language models (LMs) to retrieval tasks has gained significant attention. Dense retrieval methods, which represent queries and document chunks as vectors, have gained popularity, but their use at scale can be challenging. These models can under-perform traditional sparse approaches, like BM25, in some demanding settings, e.g. at web-scale or out-of-domain. Moreover the computational requirements, even with approximate nearest neighbour indices (ANN) can be hefty. Sparse methods, remain, thanks to their efficiency, ubiquitous in applications. In this work, we ask whether LMs can be leveraged to bridge\nthis gap. We introduce Quantised Codebooks for Retrieval (QCR): we encode queries and documents as bags of latent discrete tokens, learned purely through a contrastive objective. QCR\u2019s encodings can be used as a drop-in replacement for the original string in sparse retrieval indices, or can be instead used to complement the text with higher-level semantic features. Experimental results demonstrate that QCR outperforms BM25 with vanilla text on the challenging MSMARCO dataset. What is more, when used in conjunction with standard lexical matching, our representation yield and absolute 15.6% gain over BM25\u2019s Success@100, highlighting the complementary nature of textual and learned discrete features.", "keywords": [ "information retrieval", "sparse retrieval", "dense retrieval" ], "rating_list": [ 3, 5, 3, 1 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "TDyE2iuvyc", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Leonardo Iurada", "gender": "Male", "institution": "Politecnico di Torino", "country": "IT", "position": "PhD student" }, { "name": "Marco Ciccone", "gender": "Male", "institution": "Vector Institute", "country": "CA", "position": "Postdoc" }, { "name": "Tatiana Tommasi", "gender": "Female", "institution": "Polytechnic Institute of Turin", "country": "IT", "position": "Associate Professor" } ], "n_formula": 18, "n_formula_1": 13, "n_ref_uni": 57, "n_ref": 141, "n_ref_all": 198, "n_fig": 10, "n_tab": 7, "L_tab": 11201, "n_element_tab": 1690, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 8120, "n_element_tab_1": 1036, "formula_len_all": 1283, "formula_len_all_1": 999, "len_all": 284443, "len_all_1": 92246, "len_abs": 4176, "len_title": 62, "len_sents": 73228, "len_sents_1": 34899, "n_sents": 696, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1134, "title": "Efficient Model Editing with Task-Localized Sparse Fine-tuning", "abs": "Pre-trained models are stepping stones for modern machine learning systems, but how to efficiently extract, reuse, and steer their knowledge for new tasks is an area of research with still several open questions. State-of-the-art Task Arithmetic solutions are strongly tied to model linearization which leads to computational bottlenecks during training and inference, and potentially neglect essential task dependencies. In this work, we focus on the fine-tuning stage that defines task vectors and propose TaLoS, a new approach based on sparse fine-tuning that strategically updates only parameters expected to provide functional task localization. This efficiently yields weight-disentangled models without the need for explicit linearization. We present a thorough experimental analysis showing how our approach significantly improves in training and inference efficiency while outperforming state-of-the-art approaches in task addition and task negation. Our work offers a principled solution to pre-trained model editing and paves the way to more cost-effective and scalable machine learning systems for real-world applications.", "keywords": [ "task arithmetic", "parameter-efficient fine-tuning" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "TDy5Ih78b4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Nadezhda Chirkova", "gender": "Female", "institution": "Naver Labs Europe", "country": "", "position": "Researcher" }, { "name": "St\u00e9phane CLINCHANT", "gender": "Male", "institution": "Naver Labs Europe", "country": "FR", "position": "Principal Researcher" }, { "name": "Thibault Formal", "gender": "unknown", "institution": "Naver Labs Europe", "country": "FR", "position": "Researcher" }, { "name": "Vassilina Nikoulina", "gender": "Female", "institution": "Naver Labs Europe", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 41, "n_ref": 78, "n_ref_all": 116, "n_fig": 8, "n_tab": 9, "L_tab": 1221, "n_element_tab": 86, "n_fig_1": 7, "n_tab_1": 8, "L_tab_1": 2858, "n_element_tab_1": 84, "formula_len_all": 80, "formula_len_all_1": 80, "len_all": 200779, "len_all_1": 79143, "len_abs": 2561, "len_title": 130, "len_sents": 58244, "len_sents_1": 37397, "n_sents": 470, "n_sents_1": 285, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1463, "title": "Provence: efficient and robust context pruning for retrieval-augmented generation", "abs": "Retrieval-Augmented Generation improves various aspects of large language models (LLMs) generation, but suffers from computational overhead caused by long contexts, and the propagation of irrelevant retrieved information into generated responses. Context pruning deals with both aspects, by removing irrelevant parts of retrieved contexts before LLM generation. Existing context pruning approaches are limited, and do not present a universal model that would be both _efficient_ and _robust_ in a wide range of scenarios, e.g., when contexts contain a variable amount of relevant information or vary in length, or when evaluated on various domains. In this work, we close this gap and introduce Provence (Pruning and Reranking Of retrieVEd relevaNt ContExts), an efficient and robust context pruner for Question Answering, which dynamically detects the needed amount of pruning for a given context and can be used out-of-the-box for various domains. The three key ingredients of Provence are formulating the context pruning task as sequence labeling, unifying context pruning capabilities with context reranking, and training on diverse data. Our experimental results show that Provence enables context pruning with negligible to no drop in performance, in various domains and settings, at almost no cost in a standard RAG pipeline. We also conduct a deeper analysis alongside various ablations to provide insights into training context pruners for future work.", "keywords": [ "retrieval-augmented generation", "context pruning", "question answering" ], "rating_list": [ 8, 6, 3, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 1, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "TDuxzV3Efo", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Felix Krause", "gender": "unknown", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Melvin Sevi", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Intern" }, { "name": "Michael Neumayr", "gender": "Male", "institution": "Department of Informatics, Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "MS student" }, { "name": "Stefan Andreas Baumann", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Vincent Tao Hu", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Postdoc" }, { "name": "Bjorn Ommer", "gender": "unknown", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Nick Oliver Stracke", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 50, "n_ref": 113, "n_ref_all": 179, "n_fig": 21, "n_tab": 3, "L_tab": 2559, "n_element_tab": 197, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 3416, "n_element_tab_1": 139, "formula_len_all": 531, "formula_len_all_1": 352, "len_all": 180608, "len_all_1": 56693, "len_abs": 1444, "len_title": 143, "len_sents": 54733, "len_sents_1": 25118, "n_sents": 366, "n_sents_1": 158, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1450, "title": "Continuous, Subject-Specific Attribute Control in T2I Models by Identifying Semantic Directions", "abs": "Recent advances in text-to-image (T2I) diffusion models have significantly improved the quality of generated images. However, providing efficient control over individual subjects, particularly the attributes characterizing them, remains a key challenge. While existing methods have introduced mechanisms to modulate attribute expression, they typically provide either detailed, object-specific localization of such a modification or fine-grained, nuanced control of attributes. \nNo current approach offers both simultaneously, resulting in a gap when trying to achieve precise continuous and subject-specific attribute modulation in image generation.\nIn this work, we demonstrate that token-level directions exist within commonly used CLIP text embeddings that enable fine-grained, subject-specific control of high-level attributes in T2I models. We introduce two methods to identify these directions: a simple, optimization-free technique and a learning-based approach that utilizes the T2I model to characterize semantic concepts more specifically. Our methods allow the augmentation of the prompt text input, enabling fine-grained control over multiple attributes of individual subjects simultaneously, without requiring any modifications to the diffusion model itself. This approach offers a unified solution that fills the gap between global and localized control, providing competitive flexibility and precision in text-guided image generation.", "keywords": [ "Diffusion Models", "Computer Vision" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "TD3SGJfBC7", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Huan Liu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CA", "position": "Researcher" }, { "name": "Li Gu", "gender": "Male", "institution": "Huawei Canada", "country": "CA", "position": "Researcher" }, { "name": "Yanan Wu", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yang Wang", "gender": "Male", "institution": "Concordia University", "country": "CA", "position": "Associate Professor" }, { "name": "Zhixiang Chi", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Ziqiang Wang", "gender": "Male", "institution": "Concordia University", "country": "CA", "position": "PhD student" }, { "name": "Kostantinos Plataniotis", "gender": "Male", "institution": "Toronto University", "country": "CA", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 15, "n_ref_uni": 55, "n_ref": 104, "n_ref_all": 135, "n_fig": 10, "n_tab": 19, "L_tab": 5107, "n_element_tab": 748, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2739, "n_element_tab_1": 467, "formula_len_all": 1394, "formula_len_all_1": 1061, "len_all": 189854, "len_all_1": 78670, "len_abs": 1499, "len_title": 118, "len_sents": 44193, "len_sents_1": 30402, "n_sents": 380, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1525, "title": "Learning to Adapt Frozen CLIP for Few-Shot Test-Time Domain Adaptation", "abs": "Few-shot Test-Time Domain Adaptation focuses on adapting a model at test time to a specific domain using only a few unlabeled examples, addressing domain shift. Prior methods leverage CLIP's strong out-of-distribution (OOD) abilities by generating domain-specific prompts to guide its generalized, frozen features. However, since downstream datasets are not explicitly seen by CLIP, solely depending on the feature space knowledge is constrained by CLIP's prior knowledge. Notably, when using a less robust backbone like ViT-B/16, performance significantly drops on challenging real-world benchmarks. Departing from the state-of-the-art of inheriting the intrinsic OOD capability of CLIP, this work introduces learning directly on the input space to complement the dataset-specific knowledge for frozen CLIP. Specifically, an independent side branch is attached in parallel with CLIP and enforced to learn exclusive knowledge via revert attention. To better capture the dataset-specific label semantics for downstream adaptation, we propose to enhance the inter-dispersion among text features via greedy text ensemble and refinement. The text and visual features are then progressively fused in a domain-aware manner by a generated domain prompt to adapt toward a specific domain. Extensive experiments show our method's superiority on 5 large-scale benchmarks (WILDS and DomainNet), notably improving over smaller networks like ViT-B/16 with gains of \\textbf{+5.1} in F1 for iWildCam and \\textbf{+3.1\\%} in WC Acc for FMoW.", "keywords": [ "Distribution shifts", "Visual prompt", "Foundation model", "Few-Shot Test-Time Domain Adaptation" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "TCpJXzMnnp", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Gagandeep Singh", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Rohan Gumaste", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Undergrad student" }, { "name": "Yinglun Xu", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 22, "n_ref": 42, "n_ref_all": 54, "n_fig": 7, "n_tab": 3, "L_tab": 1408, "n_element_tab": 49, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1620, "n_element_tab_1": 65, "formula_len_all": 1107, "formula_len_all_1": 666, "len_all": 108817, "len_all_1": 65964, "len_abs": 1538, "len_title": 130, "len_sents": 40163, "len_sents_1": 30857, "n_sents": 318, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1023, "title": "Universal Black-Box Reward Poisoning Attack against Offline Reinforcement Learning", "abs": "We study the problem of universal black-boxed reward poisoning attacks against general offline reinforcement learning with deep neural networks. We consider a black-box threat model where the attacker is entirely oblivious to the learning algorithm, and its budget is limited by constraining the amount of corruption at each data point and the total perturbation. We require the attack to be universally efficient against any efficient algorithms that might be used by the agent. We propose an attack strategy called the `policy contrast attack.' The idea is to find low- and high-performing policies covered by the dataset and make them appear to be high- and low-performing to the agent, respectively. To the best of our knowledge, we propose the first universal black-box reward poisoning attack in the general offline RL setting. We provide theoretical insights on the attack design and empirically show that our attack is efficient against current state-of-the-art offline RL algorithms in different learning datasets.", "keywords": [ "Offline Reinforcement Learning", "Reward Poisoning Attack", "Universal Black Box Attack" ], "rating_list": [ 3, 5, 3, 3, 6 ], "soundness_list": [ 1, 3, 1, 2, 3 ], "presentation_list": [ 2, 2, 1, 2, 4 ], "contribution_list": [ 2, 3, 1, 2, 3 ], "confidence_list": [ 5, 3, 5, 4, 4 ] }, { "paper_id": "TCiJvhH2fC", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Dafeng Zhang", "gender": "Male", "institution": "Samsung Research China \u2013 Beijing", "country": "CN", "position": "Researcher" }, { "name": "Guanqun Liu", "gender": "Male", "institution": "Samsung", "country": "", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 19, "n_ref": 48, "n_ref_all": 65, "n_fig": 6, "n_tab": 2, "L_tab": 988, "n_element_tab": 137, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 787, "n_element_tab_1": 20, "formula_len_all": 475, "formula_len_all_1": 459, "len_all": 83466, "len_all_1": 56673, "len_abs": 1740, "len_title": 80, "len_sents": 27484, "len_sents_1": 26554, "n_sents": 193, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 32, "L_abs": 1748, "title": "Flare Removal with Visual Prompt", "abs": "Flare removal methods remove the streak, shimmer, and reflective flare in flare-corrupted images while preserving the light source. Recent deep learning methods focus on flare extraction and achieve promising results. They accomplish the task by either viewing the flare equals to the residual information between the flare-corrupted image and the flare-free image and generating the flare-free image through subtracting the extracted flare image or generating the flare-free image and the flare image simultaneously. However, due to the gap between the flare image and the residual information and handling flare extraction and clear image generation process simultaneously will give the network too much pressure and cannot fully utilize the extracted flare, these methods tend to generate images with severe artifacts. To alleviate such a phenomenon, we propose a model-agnostic pipeline named Prompt Inpainting Pipeline (PIP). Specifically, instead of viewing the gap between the flare-free and flare corrupted image as the flare or generating the flare-free image and flare image simultaneously, our prompt inpainting pipeline provides a novel perspective. We borrow the idea from inpainting methods and remove the flare by masking the polluted area and rewriting image details within. Unlike inpainting methods, we first extract multi-scale features of flare-corrupted images as a visual prompt and rewrite missing textures with the visual prompt since we find out that directly writing the missing details based on the remaining area hardly generates promising image details with sufficient semantic and high-frequency information. To verify the function of our pipeline, we conduct comprehensive experiments and demonstrate its superiority.", "keywords": [ "Flare removal", "Visual Prompt", "Prompt Inpainting Pipeline" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "TCgcEQjaUQ", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Anastasis Kratsios", "gender": "Non-Binary", "institution": "Vector Institute", "country": "CA", "position": "Researcher" }, { "name": "Artem Lukoianov", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Haitz S\u00e1ez de Oc\u00e1riz Borde", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Michael M. Bronstein", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Xiaowen Dong", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" } ], "n_formula": 44, "n_formula_1": 19, "n_ref_uni": 37, "n_ref": 96, "n_ref_all": 119, "n_fig": 3, "n_tab": 7, "L_tab": 1397, "n_element_tab": 132, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1666, "n_element_tab_1": 87, "formula_len_all": 2581, "formula_len_all_1": 1558, "len_all": 163356, "len_all_1": 70288, "len_abs": 888, "len_title": 150, "len_sents": 48441, "len_sents_1": 32204, "n_sents": 381, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 889, "title": "Scalable Message Passing Neural Networks: No Need for Attention in Large Graph Representation Learning", "abs": "We propose Scalable Message Passing Neural Networks (SMPNNs) and demonstrate that, by integrating standard convolutional message passing into a Pre-Layer Normalization Transformer-style block instead of attention, we can produce high-performing deep message-passing-based Graph Neural Networks (GNNs). This modification yields state-of-the-art results in large graph transductive learning, outperforming the best Graph Transformers in the literature without requiring the otherwise computationally and memory-expensive attention. Our architecture not only scales to large graphs but also makes it possible to construct deep message-passing networks, unlike simple GNNs, which have traditionally been constrained to shallow architectures due to oversmoothing. Moreover, we provide a new theoretical analysis of oversmoothing based on universal approximation which we use to motivate SMPNNs.", "keywords": [ "Message Passing Neural Networks", "Graph Representation Learning", "Large Graphs" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "TCSaLeANpN", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Alex Vesel", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Kevin Mayer", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Martin Fischer", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Xinyi Zhao", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 19, "n_ref": 46, "n_ref_all": 61, "n_fig": 7, "n_tab": 3, "L_tab": 495, "n_element_tab": 53, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 495, "n_element_tab_1": 53, "formula_len_all": 428, "formula_len_all_1": 428, "len_all": 76804, "len_all_1": 66401, "len_abs": 1442, "len_title": 183, "len_sents": 27315, "len_sents_1": 27316, "n_sents": 160, "n_sents_1": 160, "venue": "ICLR2025", "decision": "withdraw", "L_title": 135, "L_abs": 1448, "title": "SYNBUILD-3D: A multi-modal synthetic dataset of over 100,000 semantically enriched 3D building wireframes with AI-generated floor plans", "abs": "Modeling precise geometric and semantic relationships in 3D remains one of the greatest challenges in generative machine learning today, partly because of a lack of large 3D datasets in the public domain. Drawing upon the successful adoption of synthetic datasets in the computer vision community, we propose to address this challenge in the context of 3D buildings with SYNBUILD-3D, a large, multi-modal, and domain-specific dataset of more than 100,000 3D building wireframes along with their corresponding floor plan images. Unlike existing 3D building datasets, SYNBUILD-3D has been designed with and validated by building modeling and simulation experts, providing rich geometric and semantic information. As a result, SYNBUILD-3D is, to the best of our knowledge, the first 3D building dataset that provides interior and exterior building geometries, including the position and size of doors and windows derived from the floor plans. By releasing SYNBUILD-3D, we aim to offer the geometric deep learning community a high-quality dataset for conditional and unconditional 3D building generation tasks. In contrast to existing datasets that typically focus on modeling either the interior or exterior of 3D objects, SYNBUILD-3D can facilitate the development of generative algorithms that account for both perspectives while incorporating geometric and semantic constraints. The dataset and its associated codebase are available at GITHUB LINK.", "keywords": [ "dataset", "wireframe", "3D building", "generative modeling", "geometric deep learning" ], "rating_list": [ 1, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "TCFtGBTxkq", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Anurag Kumar", "gender": "Male", "institution": "Meta", "country": "US", "position": "Research Scientist" }, { "name": "Buye Xu", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Honglie Chen", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Jacob Donley", "gender": "unknown", "institution": "Research, Facebook", "country": "", "position": "Researcher" }, { "name": "Joanna Hong", "gender": "Female", "institution": "Meta", "country": "", "position": "Intern" }, { "name": "Ke Tan", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Research Scientist" }, { "name": "Sanjeel Parekh", "gender": "Male", "institution": "Meta", "country": "FR", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 17, "n_ref_uni": 56, "n_ref": 87, "n_ref_all": 106, "n_fig": 4, "n_tab": 8, "L_tab": 2155, "n_element_tab": 233, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 740, "n_element_tab_1": 48, "formula_len_all": 1406, "formula_len_all_1": 1245, "len_all": 160415, "len_all_1": 66406, "len_abs": 1455, "len_title": 142, "len_sents": 42348, "len_sents_1": 29503, "n_sents": 321, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1443, "title": "Efficient Audiovisual Speech Processing via MUTUD: Multimodal Training and Unimodal Deployment", "abs": "Building reliable speech systems often requires combining multiple modalities, like audio and visual cues. While such multimodal solutions frequently lead to improvements in performance and may even be critical in certain cases, they come with several constraints such as increased sensory requirements, computational cost, and modality synchronization, to mention a few. These challenges constrain the direct uses of these multimodal solutions in real-world applications. In this work, we develop approaches where the learning happens with all available modalities but the deployment or inference is done with just one or reduced modalities. To do so, we propose a Multimodal Training and Unimodal Deployment (MUTUD) framework which includes a Temporally Aligned Modality feature Estimation (TAME) module that can estimate information from missing modality using modalities present during inference. This innovative approach facilitates the integration of information across different modalities, enhancing the overall inference process by leveraging the strengths of each modality to compensate for the absence of certain modalities during inference. We apply MUTUD to various audiovisual speech tasks and show that it can reduce the performance gap between the multimodal and corresponding unimodal models to a considerable extent. MUTUD achieves this while reducing the model size and computing compared to multimodal models by almost 80%.", "keywords": [ "audiovisual learning", "speech processing", "multimodal learning", "efficiency" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 5 ] }, { "paper_id": "TBw53TdDgb", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Cameron Painting", "gender": "Male", "institution": "University of Stellenbosch", "country": "", "position": "Undergrad student" }, { "name": "Willie Brink", "gender": "Male", "institution": "Stellenbosch University", "country": "ZA", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 22, "n_ref": 36, "n_ref_all": 62, "n_fig": 8, "n_tab": 5, "L_tab": 1231, "n_element_tab": 172, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 679, "n_element_tab_1": 47, "formula_len_all": 584, "formula_len_all_1": 627, "len_all": 97524, "len_all_1": 56585, "len_abs": 1248, "len_title": 133, "len_sents": 34199, "len_sents_1": 26581, "n_sents": 231, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1251, "title": "SADE: a Scene-text Autoregressive Diffusion Engine for Character Sequence Recognition", "abs": "We consider the problem of training an optical character recognition (OCR) model to read short alphanumeric scene-text sequences, such as number plates or vehicle type labels, in scenarios where labelled training images are limited in quantity and sequence variety. OCR models may under-perform in these scenarios, so we explore whether a diffusion model can be trained on the small set of labelled images, to generate synthetic images with similar background statistics but new character sequences. We find that a diffusion model struggles to generate characters in positions of the sequence where they did not appear during training. We address this problem by introducing SADE: a scene-text autoregressive diffusion engine that generates multiple image parts one by one, conditioned on previously generated image parts for visual coherency. This approach reduces the effective number of possible positions for a character, and increases the diffusion model's ability to generate characters in novel positions of the full sequence. Our results indicate that SADE can indeed lead to substantial improvements in OCR accuracy in data-scare scenarios, particularly on sequences with characters at positions not encountered in the original training data.", "keywords": [ "optical character recognition", "diffusion models", "autoregressive image generation" ], "rating_list": [ 5, 5, 1, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "TArmA033BU", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Ari Holtzman", "gender": "Male", "institution": ", University of Chicago", "country": "", "position": "Assistant Professor" }, { "name": "Chiyuan Zhang", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Daogao Liu", "gender": "Male", "institution": "University of Washington, Seattle", "country": "", "position": "PhD student" }, { "name": "Jaechan Lee", "gender": "Male", "institution": "Department of Computer Science", "country": "", "position": "Undergrad student" }, { "name": "Jieyu Zhao", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Luke Zettlemoyer", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Full Professor" }, { "name": "Sadhika Malladi", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Weijia Shi", "gender": "unknown", "institution": "University of Washington, Seattle", "country": "US", "position": "PhD student" }, { "name": "Yangsibo Huang", "gender": "Female", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Noah Smith", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 60, "n_ref": 131, "n_ref_all": 159, "n_fig": 6, "n_tab": 7, "L_tab": 1975, "n_element_tab": 150, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 575, "n_element_tab_1": 31, "formula_len_all": 322, "formula_len_all_1": 251, "len_all": 191627, "len_all_1": 71160, "len_abs": 2264, "len_title": 112, "len_sents": 47628, "len_sents_1": 32981, "n_sents": 345, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1776, "title": "MUSE: Machine Unlearning Six-Way Evaluation for Language Models", "abs": "Language models (LMs) are trained on vast amounts of text data, which may include private and copyrighted content. Data owners may request the removal of their data from a trained model due to privacy or copyright concerns. However, exactly unlearning only these datapoints (i.e., retraining with the data removed) is intractable in modern-day models. This has led to the development of many approximate unlearning algorithms. The evaluation of the efficacy of these algorithms has traditionally been narrow in scope, failing to precisely quantify the success and practicality of the algorithm from the perspectives of both the model deployers and the data owners. We address this issue by proposing MUSE, a comprehensive machine unlearning evaluation benchmark that enumerates six diverse desirable properties for unlearned models: (1) no verbatim memorization, (2) no knowledge memorization, (3) no privacy leakage, (4) utility preservation on data not intended for removal, (5) scalability with respect to the size of removal requests, and (6) sustainability over sequential unlearning requests. Using these criteria, we benchmark how effectively eight popular unlearning algorithms on 7B-parameter LMs can unlearn Harry Potter books and news articles. Our results demonstrate that most algorithms can prevent verbatim memorization and knowledge memorization to varying degrees, but only one algorithm does not lead to severe privacy leakage. Furthermore, existing algorithms fail to meet deployer's expectations because they often degrade general model utility and also cannot sustainably accommodate successive unlearning requests or large-scale content removal. Our findings identify key issues with the practicality of existing unlearning algorithms on language models.", "keywords": [ "Language Models", "Machine Unlearning" ], "rating_list": [ 5, 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 4, 2, 4 ], "confidence_list": [ 3, 3, 4, 3, 4 ] }, { "paper_id": "TANu5nDVvU", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Kumar Shridhar", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" }, { "name": "Manish Prajapat", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Patrick Xia", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Rongxing Liu", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "MRINMAYA SACHAN", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 30, "n_ref": 57, "n_ref_all": 74, "n_fig": 4, "n_tab": 4, "L_tab": 1546, "n_element_tab": 164, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1045, "n_element_tab_1": 127, "formula_len_all": 621, "formula_len_all_1": 620, "len_all": 112627, "len_all_1": 58149, "len_abs": 1845, "len_title": 108, "len_sents": 29388, "len_sents_1": 24747, "n_sents": 223, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1774, "title": "SMART: Self-Learning Meta-strategy Agent for Reasoning Tasks", "abs": "Tasks requiring deductive reasoning, especially those involving multiple steps, often demand adaptive strategies such as intermediate generation of rationales or programs, as no single approach is universally optimal. \nWhile Language Models (LMs) can enhance their outputs through iterative self-refinement and strategy adjustments, they frequently fail to apply the most effective strategy in their first attempt. This inefficiency raises the question: *Can LMs learn to select the optimal strategy in the first attempt, without a need for refinement?*\nTo address this challenge, we introduce *SMART*: **S**elf-learning **M**eta-strategy **A**gent for **R**easoning **T**asks, a novel framework that enables LMs to autonomously learn and select the most effective strategies for various reasoning tasks. We model the strategy selection process as a *Markov Decision Process* and leverage reinforcement learning-driven continuous self-improvement to allow the model to find the suitable strategy to solve a given task. Unlike traditional self-refinement methods that rely on multiple inference passes or external feedback, *SMART* allows an LM to internalize the outcomes of its own reasoning processes and adjust its strategy accordingly, aiming for correct solutions on the first attempt.\nOur experiments across various reasoning datasets and with different model architectures demonstrate that *SMART* significantly enhances the ability of models to choose optimal strategies without external guidance (+15 points on the GSM8K dataset). By achieving higher accuracy with a single inference pass, *SMART* not only improves performance but also reduces computational costs for refinement-based strategies, paving the way for more efficient and intelligent reasoning in LMs.", "keywords": [ "Refinement", "RL", "Reasoning" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "T9u56s7mbk", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Lefan Zhang", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Algorithm Engineer" }, { "name": "Xiaodan Wang", "gender": "Female", "institution": "Xiaohongshu", "country": "CN", "position": "Researcher" }, { "name": "Yanhua Huang", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Researcher" }, { "name": "XuRuiwen", "gender": "Male", "institution": "nanjing university", "country": "", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 23, "n_ref": 61, "n_ref_all": 88, "n_fig": 9, "n_tab": 10, "L_tab": 8270, "n_element_tab": 1022, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 386, "n_element_tab_1": 65, "formula_len_all": 1773, "formula_len_all_1": 1540, "len_all": 130891, "len_all_1": 50882, "len_abs": 995, "len_title": 108, "len_sents": 36152, "len_sents_1": 21942, "n_sents": 262, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1001, "title": "Learning Harmonized Representations for Speculative Sampling", "abs": "Speculative sampling is a promising approach to accelerate the decoding stage for Large Language Models (LLMs). Recent advancements that leverage target LLM's contextual information, such as hidden states and KV cache, have shown significant practical improvements. However, these approaches suffer from inconsistent context between training and decoding. We also observe another discrepancy between the training and decoding objectives in existing speculative sampling methods. In this work, we propose a solution named HArmonized Speculative Sampling (HASS) that learns harmonized representations to address these issues. HASS accelerates the decoding stage without adding inference overhead through harmonized objective distillation and harmonized context alignment. Experiments on four LLaMA models demonstrate that HASS achieves 2.81x-4.05x wall-clock time speedup ratio averaging across three datasets, surpassing EAGLE-2 by 8%-20%. The code is available at https://github.com/HArmonizedSS/HASS.", "keywords": [ "speculative sampling", "large language model" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "T8fCTYPWBr", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Farimah Poursafaei", "gender": "Female", "institution": "McGill University", "country": "CA", "position": "Postdoc" }, { "name": "Petar Veli\u010dkovi\u0107", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Affiliated Lecturer" }, { "name": "Katarina Petrovic", "gender": "Female", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Andy Huang", "gender": "Male", "institution": "McGill University, Mila", "country": "CA", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 16, "n_ref_uni": 34, "n_ref": 78, "n_ref_all": 83, "n_fig": 2, "n_tab": 5, "L_tab": 1303, "n_element_tab": 197, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 858, "n_element_tab_1": 46, "formula_len_all": 1047, "formula_len_all_1": 747, "len_all": 111489, "len_all_1": 66534, "len_abs": 1517, "len_title": 92, "len_sents": 35497, "len_sents_1": 30840, "n_sents": 269, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 1525, "title": "Temporal Graph Rewiring with Expander Graphs", "abs": "Evolving relations in real-world networks are often modelled by temporal graphs. Temporal Graph Neural Networks (TGNNs) emerged to model evolutionary behaviour of such graphs by leveraging the message passing primitive at the core of Graph Neural Networks (GNNs). It is well-known that GNNs are vulnerable to several issues directly related to the input graph topology, such as under-reaching and over-squashing---we argue that these issues can often get exacerbated in temporal graphs, particularly as the result of stale nodes and edges. While graph rewiring techniques have seen frequent usage in GNNs to make the graph topology more favourable for message passing, they have not seen any mainstream usage on TGNNs. In this work, we propose Temporal Graph Rewiring (TGR), the first approach for graph rewiring on temporal graphs, to the best of our knowledge. TGR constructs message passing highways between temporally distant nodes in a continuous-time dynamic graph by utilizing expander graph propagation, a prominent framework used for graph rewiring on static graphs which makes minimal assumptions on the underlying graph structure. On the challenging TGB benchmark, TGR achieves state-of-the-art results on tgbl-review, tgbl-coin, tgbl-comment and tgbl-flight datasets at the time of writing. For tgbl-review, TGR has 50.5% improvement in MRR over the base TGN model and 22.2% improvement over the base TNCN model. The significant improvement over base models demonstrates clear benefits of temporal graph rewiring.", "keywords": [ "Temporal Graphs", "Dynamic Graphs", "Graph Neural Network", "Graph Representation Learning" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 4, 1, 2, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "T8PzwgYgmn", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Aaron Courville", "gender": "unknown", "institution": "University of Montreal", "country": "", "position": "Assistant Professor" }, { "name": "Arian Hosseini", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "PhD student" }, { "name": "Daniel Kenji Toyama", "gender": "Male", "institution": "DeepMind Montreal", "country": "", "position": "Research Software Engineer" }, { "name": "Rishabh Agarwal", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Adjunct Professor" }, { "name": "Alessandro Sordoni", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 44, "n_ref": 72, "n_ref_all": 103, "n_fig": 12, "n_tab": 3, "L_tab": 2401, "n_element_tab": 68, "n_fig_1": 13, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1244, "formula_len_all_1": 645, "len_all": 167534, "len_all_1": 53251, "len_abs": 979, "len_title": 87, "len_sents": 33419, "len_sents_1": 26381, "n_sents": 303, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 984, "title": "Not All LLM Reasoners Are Created Equal", "abs": "We study the depth of grade-school math (GSM) problem-solving capabilities of LLMs. To this end, we evaluate their performance on pairs of existing math word problems together so that the answer to the second problem depends on correctly answering the first problem. Our findings reveal a significant reasoning gap in most LLMs, that is performance difference between solving the compositional pairs and solving each question independently. This gap is more pronounced in smaller, more cost-efficient, and math-specialized models. Moreover, instruction-tuning recipes and code generation have varying effects across LLM sizes, while finetuning on GSM can lead to task overfitting. Our analysis indicates that large reasoning gaps are not because of test-set leakage, but due to distraction from additional context and poor second-hop reasoning. Overall, LLMs exhibit systematic differences in their reasoning abilities, despite what their performance on standard benchmarks indicates.", "keywords": [ "LLM reasoning", "Mathematical reasoning", "Reasoning Gap", "Reasoning Behavior", "Evaluating Reasoning" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "T7xIs9Z1Fm", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Aosong Feng", "gender": "Male", "institution": "Yale University", "country": "", "position": "PhD student" }, { "name": "Bo Xiong", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Irwin King", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Jiahong Liu", "gender": "Female", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Menglin Yang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Rex Ying", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 60, "n_ref": 109, "n_ref_all": 135, "n_fig": 5, "n_tab": 5, "L_tab": 1040, "n_element_tab": 81, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1040, "n_element_tab_1": 81, "formula_len_all": 1021, "formula_len_all_1": 653, "len_all": 163599, "len_all_1": 71191, "len_abs": 1615, "len_title": 96, "len_sents": 44666, "len_sents_1": 32989, "n_sents": 346, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1623, "title": "Hyperbolic Fine-tuning for Large Language Models", "abs": "Large language models (LLMs) have demonstrated remarkable performance on various tasks. However, it remains an open question whether the default Euclidean space is the most suitable choice for embedding tokens in LLMs. In this study, we first investigate the non-Euclidean characteristics of LLMs. Our findings reveal that token frequency follows a power-law distribution, with high-frequency tokens clustering near the origin and low-frequency tokens positioned farther away. Additionally, token embeddings exhibit a high degree of hyperbolicity, indicating a latent tree-like structure in the embedding space. Building on the observation, we propose to efficiently fine-tune LLMs in hyperbolic space to better exploit the underlying complex structures. However, we found that this fine-tuning in hyperbolic space cannot be achieved with naive application of exponential and logarithmic maps, when the embedding and weight matrices both reside in Euclidean space. To address this technique issue, we introduce a new method called hyperbolic low-rank efficient fine-tuning, \\method, that performs low-rank adaptation directly on the hyperbolic manifold, avoiding the cancellation effect caused by the exponential and logarithmic maps, thus preserving the hyperbolic modeling capabilities. Through extensive experiments, we demonstrate that \\method significantly enhances the performance of LLMs on reasoning tasks, particularly for complex reasoning problems. In particular, \\method improves the performance in the complex AQuA dataset by up to 13.0\\%, showcasing its effectiveness in handling complex reasoning challenges.", "keywords": [ "hyperbolic space", "representation learning", "hyperbolicity", "curvature", "fine-tunning", "large language models", "low-rank adaptation" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "T7q5LBGISH", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Dipkamal Bhusal", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Tanvirul Alam", "gender": "unknown", "institution": "Rochester Institute of Technology", "country": "BD", "position": "PhD student" }, { "name": "Michael Clifford", "gender": "Male", "institution": "Toyota Motor Corporation", "country": "US", "position": "Principal Researcher" }, { "name": "Monish Kumar manikya Veerabhadran", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "", "position": "MS student" }, { "name": "Nidhi Rastogi", "gender": "Female", "institution": "Rochester Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Sara Rampazzi", "gender": "Female", "institution": "University of Florida", "country": "", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 6, "n_ref_uni": 34, "n_ref": 93, "n_ref_all": 124, "n_fig": 14, "n_tab": 9, "L_tab": 4306, "n_element_tab": 383, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 1314, "n_element_tab_1": 129, "formula_len_all": 1987, "formula_len_all_1": 701, "len_all": 176862, "len_all_1": 50142, "len_abs": 1393, "len_title": 124, "len_sents": 61878, "len_sents_1": 23164, "n_sents": 427, "n_sents_1": 135, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1398, "title": "Towards improving saliency map interpretability using feature map smoothing", "abs": "Input-gradient-based feature attribution methods, such as Vanilla Gradient, Integrated Gradients, and SmoothGrad, are widely used to explain image classifiers by generating saliency maps. However, these methods struggle to provide explanations that are both visually clear and quantitatively robust. Key challenges include ensuring that explanations are sparse, stable, and faithfully reflect the model\u2019s decision-making. Adversarial training, known for enhancing model robustness, have been shown to produce sparser explanations with these methods; however, this sparsity often comes at the cost of stability. In this work, we investigate the trade-off between stability and sparsity in saliency maps and propose the use of a smoothing layer during adversarial training. Through extensive experiments and evaluation, we demonstrate this smoothing technique improves the stability and faithfulness of saliency maps without sacrificing sparsity. Furthermore, a qualitative user study reveals that human evaluators tend to distrust explanations that are overly noisy or excessively sparse\u2014issues commonly associated with explanations in naturally and adversarially trained models, respectively and prefer explanations produced by our proposed approach. Our findings offer a promising direction for generating reliable explanations with robust models, striking a balance between clarity and usability.", "keywords": [ "computer vision; saliency map; explanations; vanilla gradient; integrated gradient; smoothgrad" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "T7lQGq73Lm", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Jiahe Du", "gender": "unknown", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Jinbo Xu", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Full Professor" }, { "name": "Kaixiong Zhou", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Assistant Professor" }, { "name": "Xiao Huang", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "", "position": "Assistant Professor" }, { "name": "Xinyu Hong", "gender": "unknown", "institution": "MoleculeMind", "country": "", "position": "Researcher" }, { "name": "Zhaozhuo Xu", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "US", "position": "Assistant Professor" } ], "n_formula": 23, "n_formula_1": 22, "n_ref_uni": 35, "n_ref": 67, "n_ref_all": 102, "n_fig": 4, "n_tab": 10, "L_tab": 2025, "n_element_tab": 290, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1661, "n_element_tab_1": 151, "formula_len_all": 1505, "formula_len_all_1": 1310, "len_all": 148947, "len_all_1": 71201, "len_abs": 1921, "len_title": 120, "len_sents": 41642, "len_sents_1": 30702, "n_sents": 320, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1928, "title": "Retrieval Augmented Zero-Shot Enzyme Generation for Specified Substrate", "abs": "The ability to generate novel enzymes that catalyze specific target molecules is a critical advancement in biomaterial synthesis and chemical production. However, a significant challenge arises when no recorded enzymes exist for the target molecule, making it a zero-shot generation problem. This absence of known enzymes complicates the training of generative models tailored to the target substrate. To address this, we propose a retrieval-augmented generation method that leverages existing enzyme-substrate data to overcome the lack of direct examples. Since there is no recorded catalytic performance between the enzymes and the new target molecule, the challenge shifts to identifying enzymes that helpful for generation. Our approach tackles this by retrieving enzymes whose substrates exhibit structural similarities to the target molecule, thereby exploiting functional similarities reflected in the enzymes' catalytic capability. This leads to the next challenge: how to utilize the retrieved enzymes to generate a novel enzyme capable of catalyzing the target molecule, given that none of the retrieved enzymes directly catalyze it. To solve this, we employ a conditioned discrete diffusion model that takes the aligned retrieved enzymes to generate a new enzyme. We train the generator with guidance from an enzyme-substrate relationship classifier to make it output the optimal protein sequence distribution for different target molecule. We evaluate our model on enzyme design tasks involving a diverse set of real-world substrates, and our results including catalytic rate predictions, foldability assessments, and docking position analyses, demonstrate that our model outperforms existing protein generation methods for substrate-specified enzyme generation. Additionally, we formally define the zero-shot substrate-specified enzyme generation task and contribute a comprehensive dataset with evaluation methods.", "keywords": [ "Protein Design", "Diffusion model" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "T7kThJhl02", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Yuepeng Hu", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Zhengyuan Jiang", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Neil Gong", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 19, "n_ref": 70, "n_ref_all": 89, "n_fig": 8, "n_tab": 5, "L_tab": 1655, "n_element_tab": 119, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1886, "n_element_tab_1": 101, "formula_len_all": 243, "formula_len_all_1": 244, "len_all": 100848, "len_all_1": 60836, "len_abs": 1499, "len_title": 113, "len_sents": 35684, "len_sents_1": 28687, "n_sents": 258, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1404, "title": "SafeText: Safe Text-to-image Models via Aligning the Text Encoder", "abs": "Text-to-image models can generate harmful images when presented with unsafe prompts, posing significant safety and societal risks. Alignment methods aim to modify these models to ensure they generate only non-harmful images, even when exposed to unsafe prompts. A typical text-to-image model comprises two main components: 1) a text encoder and 2) a diffusion module. Existing alignment methods mainly focus on modifying the diffusion module to prevent harmful image generation. However, this often significantly impacts the model\u2019s behavior for safe prompts, causing substantial quality degradation of generated images. In this work, we propose SafeText, a novel alignment method that fine-tunes the text encoder rather than the diffusion module. By adjusting the text encoder, SafeText significantly alters the embedding vectors for unsafe prompts, while minimally affecting those for safe prompts. As a result, the diffusion module generates non-harmful images for unsafe prompts while preserving the quality of images for safe prompts. We evaluate SafeText on multiple datasets of safe and unsafe prompts, including those generated through jailbreak attacks. Our results show that SafeText effectively prevents harmful image generation with minor impact on the images for safe prompts, and SafeText outperforms six existing alignment methods. We will publish our code and data after paper acceptance.", "keywords": [ "Text-to-image Model", "Alignment", "Safeguard" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "T7bmHkwzS6", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Jing Tang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Tianyang Hu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Xinghua Qu", "gender": "Male", "institution": "Bytedance", "country": "SG", "position": "Research Scientist" }, { "name": "Yihong Luo", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CX", "position": "PhD student" }, { "name": "Xiaolong CHEN", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 22, "n_ref_uni": 55, "n_ref": 131, "n_ref_all": 153, "n_fig": 11, "n_tab": 2, "L_tab": 265, "n_element_tab": 28, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 861, "n_element_tab_1": 133, "formula_len_all": 1608, "formula_len_all_1": 1853, "len_all": 177357, "len_all_1": 73355, "len_abs": 1503, "len_title": 145, "len_sents": 41980, "len_sents_1": 31072, "n_sents": 368, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1533, "title": "You Only Sample Once: Taming One-Step Text-to-Image Synthesis by Self-Cooperative Diffusion GANs", "abs": "Recently, some works have tried to combine diffusion and Generative Adversarial Networks (GANs) to alleviate the computational cost of the iterative denoising inference in Diffusion Models (DMs). \nHowever, existing works in this line suffer from either training instability and mode collapse or subpar one-step generation learning efficiency. \nTo address these issues, we introduce YOSO, a novel generative model designed for rapid, scalable, and high-fidelity one-step image synthesis with high training stability and mode coverage. \nSpecifically, we smooth the adversarial divergence by the denoising generator itself, performing self-cooperative learning. We show that our method can serve as a one-step generation model training from scratch with competitive performance. \nMoreover, we extend our YOSO to one-step text-to-image generation based on pre-trained models by several effective training techniques (i.e., latent perceptual loss and latent discriminator for efficient training along with the latent DMs; the informative prior initialization (IPI), and the quick adaption stage for fixing the flawed noise scheduler). Experimental results show that YOSO achieves the state-of-the-art one-step generation performance even with Low-Rank Adaptation (LoRA) fine-tuning.\nIn particular, we show that the YOSO-PixArt-$\\alpha$ can generate images in one step trained on 512 resolution, with the capability of adapting to 1024 resolution without extra explicit training, requiring only \\textasciitilde10 A800 days for fine-tuning.", "keywords": [ "One-step text-to-image generation; Diffusion Models; Generative Adversarial Networks" ], "rating_list": [ 5, 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 4 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3, 4 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "T7ZVzuObcj", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Chris Bakal", "gender": "Male", "institution": "Institute of Cancer Research ", "country": "GB", "position": "Full Professor" }, { "name": "Matt De Vries", "gender": "Male", "institution": "Institute of Cancer Research", "country": "GB", "position": "PhD student" }, { "name": "Olga Fourkioti", "gender": "Female", "institution": "The Institute of Cancer Research", "country": "GB", "position": "Postdoc" }, { "name": "Reed Naidoo", "gender": "Male", "institution": "The Institute of Cancer Research", "country": "GB", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 39, "n_ref": 75, "n_ref_all": 91, "n_fig": 6, "n_tab": 4, "L_tab": 1584, "n_element_tab": 205, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 4259, "n_element_tab_1": 200, "formula_len_all": 332, "formula_len_all_1": 332, "len_all": 156106, "len_all_1": 56581, "len_abs": 1388, "len_title": 122, "len_sents": 32534, "len_sents_1": 23104, "n_sents": 255, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1192, "title": "Interpretable point cloud classification using multiple instance learning", "abs": "3D image analysis is crucial in fields such as autonomous driving and biomedical research. However, existing 3D point cloud classification models lack interpretability, limiting trust and usability in safety-critical applications. To address this, we propose PointMIL, an inherently locally interpretable point cloud classifier using Multiple Instance Learning (MIL). PointMIL offers local interpretability, providing fine-grained point-specific explanations to point-based models without the need for \\textit{post-hoc} methods, addressing the limitations of global or imprecise interpretability approaches. We applied PointMIL to four popular point cloud classifiers, PointNet, DGCNN, CurveNet, PointMLP, and PointNeXt, and proposed a transformer-based backbone to extract high-quality point-specific features. PointMIL made these models inherently interpretable while increasing predictive performance on standard benchmarks (ModelNet40, ShapeNetPart) and achieving state-of-the-art mACC ($97.3\\%$) and F1 ($97.5\\%$) on the IntrA biomedical data set, and another dataset of biological cells. To our knowledge, this is the first work to apply MIL to interpretable point cloud classification.", "keywords": [ "Multiple instance learning", "point cloud", "interpretable" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "T6qIMnokrI", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Cheng Yu Yeo", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "Hong-Han Shuai", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" }, { "name": "Yi-Lun Wu", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 13, "n_ref_uni": 35, "n_ref": 60, "n_ref_all": 91, "n_fig": 11, "n_tab": 13, "L_tab": 4111, "n_element_tab": 378, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1622, "n_element_tab_1": 165, "formula_len_all": 1350, "formula_len_all_1": 525, "len_all": 166515, "len_all_1": 61747, "len_abs": 764, "len_title": 66, "len_sents": 47203, "len_sents_1": 28405, "n_sents": 333, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 765, "title": "Data Exfiltration in Diffusion Models: A Backdoor Attack Approach", "abs": "As diffusion models (DMs) become increasingly susceptible to adversarial attacks, this paper investigates a novel method of data exfiltration through strategically implanted backdoors. Unlike conventional techniques that directly alter data, we pioneer the use of unique trigger embeddings for each image to enable covert data retrieval. Furthermore, we extend our exploration to text-to-image diffusion models such as Stable Diffusion by introducing the Caption Backdoor Subnet (CBS), which exploits these models for both image and caption extraction. This innovative approach not only reveals an unexplored facet of diffusion model security but also contributes valuable insights toward enhancing the resilience of generative models against sophisticated threats.", "keywords": [ "Backdoor Attack", "Data Exfiltration", "Diffusion Model" ], "rating_list": [ 6, 6, 3, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 1, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "T6hhDEnAoo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Jian Zhou", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Associate Professor" }, { "name": "Jinsheng Xiao", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Associate Professor" }, { "name": "Wei Yang", "gender": "Male", "institution": "Wuchang Shouyi University", "country": "CN", "position": "Associate Professor" }, { "name": "Wenjuan Xie", "gender": "Female", "institution": "FiberHome", "country": "CN", "position": "Researcher" }, { "name": "Yuebin Zhou", "gender": "Male", "institution": "Central China Normal University", "country": "CN", "position": "MS student" }, { "name": "GU Mingyang", "gender": "Female", "institution": "Wuhan University", "country": "CN", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 13, "n_ref": 39, "n_ref_all": 53, "n_fig": 3, "n_tab": 5, "L_tab": 2180, "n_element_tab": 332, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2133, "n_element_tab_1": 296, "formula_len_all": 328, "formula_len_all_1": 328, "len_all": 90806, "len_all_1": 62194, "len_abs": 1366, "len_title": 131, "len_sents": 31420, "len_sents_1": 28909, "n_sents": 191, "n_sents_1": 152, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1373, "title": "Query Optimization Detection Transformer for Small Objects in Remote Sensing Images", "abs": "Object detection in remote sensing images is a challenging task. Remote sensing images contain substantial background noise and complex contextual information, which weakens the feature representation of small objects, making detection difficult. To solve these problems, a detection Transformer for small objects in remote sensing images is proposed, called QO-DETR. Specifically, to enhance the feature representation of small objects, a query proposal generation module is designed to select queries based on multi-class classification scores. These queries provide the initial position embeddings for object queries in the decoder, enabling the decoder's attention mechanism to focus on object regions. To improve the model\u2019s robustness to noise, a group denoising module is designed to add noise into decoder queries during training, enhancing the network's ability to reconstruct object features from noise. To accurately locate small objects, a query cascade refinement strategy is designed, and each decoder layer refines anchor parameters under the guidance of preceding layers to achieve spatial alignment between the anchor and the object. Experiments have been carried out on DIOR and AI-TOD. The AP and APs on DIOR reach 51.3% and 13.4%, respectively, while on AI-TOD, they reach 23.6% and 30.1%. QO-DETR shows superior performance in detecting small objects.", "keywords": [ "Remote sensing images; Small object detection; Transformer; Query optimization" ], "rating_list": [ 3, 3, 3, 1, 1 ], "soundness_list": [ 1, 1, 2, 1, 1 ], "presentation_list": [ 2, 1, 2, 1, 1 ], "contribution_list": [ 1, 1, 2, 1, 1 ], "confidence_list": [ 5, 5, 4, 5, 4 ] }, { "paper_id": "T6QqcRiypb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chengjia Wang", "gender": "unknown", "institution": "Heriot-Watt University", "country": "GB", "position": "Assistant Professor" }, { "name": "Giorgos Papanastasiou", "gender": "Male", "institution": "Archimedes Unit, Athena RC", "country": "GR", "position": "Principal Researcher" }, { "name": "Guang Yang", "gender": "unknown", "institution": "Imperial College London, Imperial College London", "country": "GB", "position": "UKRI Future Leaders Fellow (Tenured)" }, { "name": "Hongwei Bran Li", "gender": "Male", "institution": "Massachusetts General Hospital, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Wei Pang", "gender": "Male", "institution": "Heriot-Watt University", "country": "GB", "position": "Full Professor" }, { "name": "Yuchen Mao", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 30, "n_ref": 41, "n_ref_all": 67, "n_fig": 8, "n_tab": 9, "L_tab": 1216, "n_element_tab": 118, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 752, "n_element_tab_1": 41, "formula_len_all": 497, "formula_len_all_1": 447, "len_all": 117839, "len_all_1": 59470, "len_abs": 1434, "len_title": 96, "len_sents": 37478, "len_sents_1": 28430, "n_sents": 271, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1488, "title": "SeLoRA: Self-Expanding Low-Rank Adaptation of Latent Diffusion Model for Medical Image Synthesis", "abs": "The persistent challenge of medical image synthesis posed by the scarcity of annotated data and the need to synthesize \"missing modalities\" for multi-modal analysis, underscored the imperative development of effective synthesis methods. Recently, the combination of Low-Rank Adaptation (*LoRA*) with latent diffusion models (LDMs) has emerged as a viable approach for efficiently adapting pre-trained large language models, in the medical field. However, the direct application of *LoRA* assumes uniform ranking across all linear layers, overlooking the significance of different weight matrices, and leading to sub-optimal outcomes. Prior works on *LoRA* prioritize the reduction of trainable parameters, and there exists an opportunity to further tailor this adaptation process to the intricate demands of medical image synthesis. In response, we present *SeLoRA*, a Self-Expanding Low-Rank Adaptation module, that dynamically expands its ranking across layers during training, strategically placing additional ranks on crucial layers, to allow the model to elevate synthesis quality where it matters most. Our analysis shows that *SeLoRA* strikes the best balance between synthesis quality and training efficiency. The proposed method not only enables LDMs to fine-tune on medical data efficiently but also empowers the model to achieve improved image quality with minimal ranking. The code of our *SeLoRA* method is publicly available at https://anonymous.4open.science/r/SeLoRA-980D.", "keywords": [ "Text-to-Image Synthesis", "Low-Rank Adaptation", "Medical Imaging", "Parameter Efficient Finetuning" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 5, 4, 5 ] }, { "paper_id": "T5QLRRHyL1", "primary_area": "applications to robotics, autonomy, planning", "n_author": 20, "author_info_list": [ { "name": "Akshara Rai", "gender": "unknown", "institution": "Facebook AI Research", "country": "", "position": "Researcher" }, { "name": "Alexander Clegg", "gender": "Male", "institution": "Meta AI", "country": "", "position": "Researcher" }, { "name": "Daniel Tran", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Eric Undersander", "gender": "unknown", "institution": "Meta ", "country": "", "position": "Research Engineer" }, { "name": "Gunjan Chhablani", "gender": "Male", "institution": "Waymo", "country": "US", "position": "Researcher" }, { "name": "Jacob Krantz", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Joanne Truong", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "John M Turner", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Contingent Research Engineer" }, { "name": "Matthew Chang", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "PhD student" }, { "name": "Michal Hlavac", "gender": "Male", "institution": "Meta", "country": "", "position": "Designer" }, { "name": "Mikael Dallaire Cote", "gender": "Not Specified", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Priyam Parashar", "gender": "Female", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Ram Ramrakhya", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Roozbeh Mottaghi", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Research Scientist Manager" }, { "name": "Tsung-Yen Yang", "gender": "unknown", "institution": "Meta AI", "country": "", "position": "Researcher" }, { "name": "Vladimir Karashchuk", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Xavier Puig", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Ishita Prasad", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Ruta Parimal Desai", "gender": "Female", "institution": "Meta, FAIR", "country": "US", "position": "Researcher" }, { "name": "Siddharth Nitin Patki", "gender": "Male", "institution": "meta", "country": "", "position": "Postdoc" } ], "n_formula": 17, "n_formula_1": 1, "n_ref_uni": 49, "n_ref": 134, "n_ref_all": 222, "n_fig": 22, "n_tab": 7, "L_tab": 3943, "n_element_tab": 380, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1114, "n_element_tab_1": 130, "formula_len_all": 1706, "formula_len_all_1": 179, "len_all": 424152, "len_all_1": 76070, "len_abs": 1715, "len_title": 125, "len_sents": 156932, "len_sents_1": 36925, "n_sents": 1331, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1344, "title": "PARTNR: A Benchmark for Planning and Reasoning in Embodied Multi-agent Tasks", "abs": "We present a benchmark for Planning And Reasoning Tasks in humaN-Robot collaboration (PARTNR) designed to study human-robot coordination in household activities. PARTNR tasks exhibit characteristics of everyday tasks, such as spatial, temporal, and heterogeneous agent capability constraints. We employ a semi-automated task generation pipeline using Large Language Models (LLMs), incorporating simulation-in-the-loop for the grounding and verification. PARTNR stands as the largest benchmark of its kind, comprising 100,000 natural language tasks, spanning 60 houses and 5,819 unique objects. We analyze state-of-the-art LLMs on PARTNR tasks, across the axes of planning, perception and skill execution. The analysis reveals significant limitations in SoTA models, such as poor coordination and failures in task tracking and recovery from errors. When LLMs are paired with 'real' humans, they require 1.5x as many steps as two humans collaborating and 1.1x more steps than a single human, underscoring the potential for improvement in these models. We further show that fine-tuning smaller LLMs with planning data can achieve performance on par with models 9 times larger, while being 8.6x faster at inference. Overall, PARTNR highlights significant challenges facing collaborative embodied agents and aims to drive research in this direction.", "keywords": [ "Human-Robot Collaboration", "Planning", "Embodied AI" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 3, 4, 2, 4 ], "confidence_list": [ 2, 3, 5, 4 ] }, { "paper_id": "T4sMzjy7fO", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Jing Yuan", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Yi Zhou", "gender": "Male", "institution": "ByteDance AI Lab", "country": "CN", "position": "Research Scientist" }, { "name": "Yilai Li", "gender": "Male", "institution": "ByteDance", "country": "US", "position": "Scientist" } ], "n_formula": 36, "n_formula_1": 22, "n_ref_uni": 84, "n_ref": 193, "n_ref_all": 223, "n_fig": 14, "n_tab": 8, "L_tab": 1498, "n_element_tab": 219, "n_fig_1": 10, "n_tab_1": 7, "L_tab_1": 99, "n_element_tab_1": 7, "formula_len_all": 3002, "formula_len_all_1": 1364, "len_all": 246256, "len_all_1": 64359, "len_abs": 1183, "len_title": 107, "len_sents": 58438, "len_sents_1": 29124, "n_sents": 428, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1175, "title": "CryoFM: A Flow-based Foundation Model for Cryo-EM Densities", "abs": "Cryo-electron microscopy (cryo-EM) is a powerful technique in structural biology and drug discovery, enabling the study of biomolecules at high resolution. Significant advancements by structural biologists using cryo-EM have led to the production of around 40k protein density maps at various resolutions. However, cryo-EM data processing algorithms have yet to fully benefit from our knowledge of biomolecular density maps, with only a few recent models being data-driven but limited to specific tasks. In this study, we present CryoFM, a foundation model designed as a generative model, learning the distribution of high-quality density maps and generalizing effectively to downstream tasks. Built on flow matching, CryoFM is trained to accurately capture the prior distribution of biomolecular density maps. Furthermore, we introduce a flow posterior sampling method that leverages CryoFM as a flexible prior for several downstream tasks in cryo-EM and cryo-electron tomography (cryo-ET) without the need for fine-tuning, achieving state-of-the-art performance on most tasks and demonstrating its potential as a foundational model for broader applications in these fields.", "keywords": [ "cryoEM; foundation model; flow matching" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "T4VK4U4aKb", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Chen Gao", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Nian Li", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yu Shang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 39, "n_ref": 72, "n_ref_all": 99, "n_fig": 14, "n_tab": 5, "L_tab": 1769, "n_element_tab": 180, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 706, "n_element_tab_1": 98, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 168179, "len_all_1": 60074, "len_abs": 1409, "len_title": 140, "len_sents": 53371, "len_sents_1": 28024, "n_sents": 482, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1296, "title": "A Large-scale Dataset with Behavior, Attributes, and Content of Mobile Short-video Platform", "abs": "Short-video platforms show an increasing impact on people\u2019s daily life nowadays, with billions of active users spending plenty of time each day. The interactions between users and online platforms give rise to many scientific problems across computational social science and artificial intelligence. However, despite the rapid development of short-video platforms, currently there are serious shortcomings in existing relevant datasets on three aspects: inadequate user-video feedback, limited user attributes and lack of video content. To address these problems, we provide a large-scale dataset with rich user behavior, attributes and video content from a real mobile short-video platform. This dataset covers 10,000 voluntary users and 153,561 videos, and we conduct three-fold technical validations of the dataset. First, we verify the richness of the behavior data including interaction frequency and feedback distribution. Second, we validate the wide coverage of user-side and video-side attribute data. Third, we confirm the representing ability of the content features. We believe the dataset could support the broad research community, including user modeling, social science, human behavior understanding, etc. Our dataset is available at this anonymous link: http://101.6.70.16:8080/.", "keywords": [ "Large-scale dataset", "Behavior", "Attributes", "Video content", "Mobile short-video platform" ], "rating_list": [ 1, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "T4LtGj7us1", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "De-Chuan Zhan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Han-Jia Ye", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Qing-Guo Chen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Shiyin Lu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yi-Kai Zhang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 51, "n_ref": 116, "n_ref_all": 137, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 251, "n_element_tab_1": 3, "formula_len_all": 757, "formula_len_all_1": 770, "len_all": 222767, "len_all_1": 71030, "len_abs": 1992, "len_title": 157, "len_sents": 62020, "len_sents_1": 34232, "n_sents": 467, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 109, "L_abs": 1751, "title": "ZooProbe: A Data Engine for Evaluating, Exploring, and Evolving Large-scale Training Data for Multimodal LLMs", "abs": "Multimodal Large Language Models (MLLMs) are thriving through continuous fine-tuning by LLMs. Driven by the law that \"scale is everything\", MLLMs expand their training sets during version iterations. In this paper, we propose a large-scale training data engine built around an evaluating-exploring-evolving (E3) loop. Evaluating the data provides insights into its characteristics. Exploring quality rules helps identify which data enhances training. Together, these processes facilitate the systematic evolution of new, high-quality data. With the E3 loop, we introduce ZooProbe, an efficient data engine for MLLMs. First, the problem of data expansion is formalized as a tree of sampling and growth. ZooProbe introduces a small-scale model *zoo* to obtain comprehensive evaluations for child datasets. From multiple perspectives, visual, textual, and multimodal models cover over 50 dimensions of intrinsic and meta attributes, such as object and topic distribution, and higher-level properties, like annotation quality and scene complexity. ZooProbe constructs based on A$^\\star$ search, modeling the heuristic function as a quality estimate from data evaluation results. It dynamically explores the rule of data quality based on the model state of the *probe* datasets. Additionally, it evolves new targeted data with identified high-quality rules. We also develop an extra heuristic quality ranker with the data utilized and discarded during the expansion. Our experiments show that ZooProbe significantly breaks the scaling law in multimodal instruction fine-tuning at scales of 260$k$ and below.\nZooProbe generates high-quality data that accelerates MLLM training and enhances performance, automating the evolution of large-scale training data.", "keywords": [ "Multimodal Large Language Model", "Training Data Engine", "Deep Learning" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 4 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "T2h2V7Rx7q", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Alon Benhaim", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Barun Patra", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Applied Scientist" }, { "name": "Han Zhao", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Parul Chopra", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Praneetha Vaddamanu", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Applied Scientist" }, { "name": "Sanchit Ahuja", "gender": "Male", "institution": "Research, Microsoft", "country": "IN", "position": "Researcher" }, { "name": "Vishrav Chaudhary", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Senior Principal Researcher" }, { "name": "Xia Song", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Yifei He", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "", "position": "PhD student" } ], "n_formula": 39, "n_formula_1": 17, "n_ref_uni": 31, "n_ref": 51, "n_ref_all": 84, "n_fig": 9, "n_tab": 11, "L_tab": 5983, "n_element_tab": 901, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1347, "n_element_tab_1": 73, "formula_len_all": 2438, "formula_len_all_1": 673, "len_all": 163385, "len_all_1": 66056, "len_abs": 1518, "len_title": 93, "len_sents": 45894, "len_sents_1": 30588, "n_sents": 362, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1525, "title": "Scaling Laws for Multilingual Language Models", "abs": "We propose a novel scaling law for general-purpose decoder-only language models (LMs) trained on multilingual data, tackling the problem of balancing languages during multilingual pretraining. A primary challenge in studying multilingual scaling is the difficulty of analyzing individual language performance due to cross-lingual transfer. To address this, we shift the focus from individual languages to language families. We introduce and validate a hypothesis that the test cross-entropy loss for each language family is determined solely by its own sampling ratio, independent of other languages in the mixture. This insight simplifies the complexity of multilingual scaling and make the analysis scalable to an arbitrary number of languages. Building on this hypothesis, we derive a power-law relationship that links performance with dataset size, model size and sampling ratios. This relationship enables us to predict performance across various combinations of the above three quantities, and derive the optimal sampling ratios at different model scales. To demonstrate the effectiveness and accuracy of our proposed scaling law, we perform a large-scale empirical study, training more than 100 models on 23 languages spanning 5 language families. Our experiments show that the optimal sampling ratios derived from small models (85M parameters) generalize effectively to models that are several orders of magnitude larger (1.2B parameters), offering a resource-efficient approach for multilingual LM training at scale.", "keywords": [ "scaling laws", "large language models", "multilingual language models" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "T2dhpC3N0d", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Lei Wang", "gender": "Male", "institution": "University of Wollonong", "country": "AU", "position": "Full Professor" }, { "name": "Lingqiao Liu", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Associate Professor" }, { "name": "Luping Zhou", "gender": "Female", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Xinyu Liang", "gender": "Female", "institution": "Guangzhou University of Chinses Medicine", "country": "CN", "position": "PhD student" }, { "name": "YUNYI LIU", "gender": "Female", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Yingshu Li", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Zhanyu Wang", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 19, "n_ref": 57, "n_ref_all": 69, "n_fig": 3, "n_tab": 7, "L_tab": 1424, "n_element_tab": 108, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 2313, "n_element_tab_1": 158, "formula_len_all": 347, "formula_len_all_1": 347, "len_all": 153250, "len_all_1": 65472, "len_abs": 1937, "len_title": 153, "len_sents": 53032, "len_sents_1": 28558, "n_sents": 405, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 1961, "title": "ER2Score: An Explainable and Customizable Metric for Assessing Radiology Reports with LLM-based Rewards", "abs": "In recent years, the automated generation of radiology reports (R2Gen) has seen considerable growth, introducing new challenges in evaluation due to its complex nature. Traditional metrics often fail to provide accurate evaluations due to their reliance on rigid word-matching techniques or their exclusive focus on pathological entities, leading to inconsistencies with human assessments. To bridge this gap, we introduce ER2Score, an automatic evaluation metric designed specifically for R2Gen that harnesses the capabilities of Large Language Models (LLMs). Our metric leverages a reward model and a tailored design for training data, allowing customization of evaluation criteria based on user-defined needs. It not only scores reports according to user-specified criteria but also provides detailed sub-scores, enhancing interpretability and allowing users to adjust the criteria between clinical and linguistic aspects of reports. Leveraging GPT-4, we generate extensive evaluation data for training based on two different scoring systems, respectively, including reports of varying quality alongside corresponding scores. These GPT-generated reports are then paired as accepted and rejected samples to train an LLM towards a reward model, which assigns higher rewards to the report with high quality. Our proposed loss function enables this model to simultaneously output multiple individual rewards corresponding to the number of evaluation criteria, with their summation as our final ER2Score. Our experiments demonstrate ER2Score's heightened correlation with human judgments and superior performance in model selection compared to traditional metrics. Notably, our model's capability to provide not only a single overall score but also scores for individual evaluation items enhances the interpretability of the assessment results. We also showcase the flexible training of our model to varying evaluation systems. We will release the code on GitHub.", "keywords": [ "Radiology Report Generation", "Auto Evaluation Metrics", "Reward Model", "LLM", "RLHF" ], "rating_list": [ 6, 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 2, 1, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "T2d0geb6y0", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Hantao Yu", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Josh Alman", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 26, "n_formula_1": 10, "n_ref_uni": 57, "n_ref": 94, "n_ref_all": 100, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2183, "formula_len_all_1": 884, "len_all": 210567, "len_all_1": 74131, "len_abs": 1580, "len_title": 117, "len_sents": 52933, "len_sents_1": 34635, "n_sents": 458, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1591, "title": "Fundamental Limitations on Subquadratic Alternatives to Transformers", "abs": "The Transformer architecture is widely deployed in many popular and impactful Large Language Models. At its core is the attention mechanism for calculating correlations between pairs of tokens. Performing an attention computation takes quadratic time in the input size, and had become the time bottleneck for transformer operations. In order to circumvent this, researchers have used a variety of approaches, including designing heuristic algorithms for performing attention computations faster, and proposing alternatives to the attention mechanism which can be computed more quickly. For instance, state space models such as Mamba were designed to replace attention with an almost linear time alternative.\n\nIn this paper, we prove that any such approach cannot perform important tasks that Transformer is able to perform (assuming a popular conjecture from fine-grained complexity theory). We focus on document similarity tasks, where one is given as input many documents and would like to find a pair which is (approximately) the most similar. We prove that Transformer is able to perform this task, and we prove that this task cannot be performed in truly subquadratic time by any algorithm. Thus, any model which can be evaluated in subquadratic time \u2013 whether because of subquadratic-time heuristics for attention, faster attention replacements like Mamba, or any other reason \u2013 cannot perform this task. In other words, in order to perform tasks that (implicitly or explicitly) involve document similarity, one may as well use Transformer and cannot avoid its quadratic running time.", "keywords": [ "Large Language Models", "Transformers", "Fine-grained complexity theory", "Document similarity", "Hardness of Approximation", "Fast attention computation" ], "rating_list": [ 6, 8, 3, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 3, 5, 4, 1 ] }, { "paper_id": "T26f9z2rEe", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Tao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaoying Tang", "gender": "Female", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Yongxin Guo", "gender": "Male", "institution": "Chinese University of HongKong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Zhaopeng Tu", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhenglin Cheng", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 19, "n_ref_uni": 52, "n_ref": 125, "n_ref_all": 157, "n_fig": 11, "n_tab": 12, "L_tab": 16028, "n_element_tab": 776, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2175, "n_element_tab_1": 446, "formula_len_all": 2423, "formula_len_all_1": 2114, "len_all": 188095, "len_all_1": 70069, "len_abs": 1148, "len_title": 132, "len_sents": 37712, "len_sents_1": 27143, "n_sents": 310, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1154, "title": "Dynamic Mixture of Experts: An Auto-Tuning Approach for Efficient Transformer Models", "abs": "The Sparse Mixture of Experts (SMoE) has been widely employed to enhance the efficiency of training and inference for Transformer-based foundational models, yielding promising results. However, the performance of SMoE heavily depends on the choice of hyper-parameters, such as the number of experts and the number of experts to be activated (referred to as top-$k$), resulting in significant computational overhead due to the extensive model training by searching over various hyper-parameter configurations. As a remedy, we introduce the Dynamic Mixture of Experts (DynMoE) technique. DynMoE incorporates (1) a novel gating method that enables each token to automatically determine the number of experts to activate. (2) An adaptive process automatically adjusts the number of experts during training. Extensive numerical results across Vision, Language, and Vision-Language tasks demonstrate the effectiveness of our approach to achieve competitive performance compared to GMoE for vision and language tasks, and MoE-LLaVA for vision-language tasks, while maintaining efficiency by activating fewer parameters. Our code will be made publicly available.", "keywords": [ "Mixture of Experts", "Transformer Model" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "T1pUS4GZZq", "primary_area": "reinforcement learning", "n_author": 9, "author_info_list": [ { "name": "G\u00fcnter Klambauer", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Full Professor" }, { "name": "Johannes Brandstetter", "gender": "Male", "institution": "Emmi AI", "country": "AT", "position": "Principal Researcher" }, { "name": "Korbinian P\u00f6ppel", "gender": "unknown", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "PhD student" }, { "name": "Maximilian Beck", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "", "position": "PhD student" }, { "name": "Razvan Pascanu", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Affiliate Member" }, { "name": "Sepp Hochreiter", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "Full Professor" }, { "name": "Thomas Adler", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Postdoc" }, { "name": "Thomas Schmied", "gender": "Not Specified", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "PhD student" }, { "name": "Vihang Prakash Patil", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 83, "n_ref": 229, "n_ref_all": 291, "n_fig": 33, "n_tab": 16, "L_tab": 18896, "n_element_tab": 1369, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 580, "n_element_tab_1": 54, "formula_len_all": 181, "formula_len_all_1": 297, "len_all": 388362, "len_all_1": 62117, "len_abs": 1329, "len_title": 128, "len_sents": 83234, "len_sents_1": 29159, "n_sents": 677, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1034, "title": "A Large Recurrent Action Model: xLSTM enables Fast Inference for Robotics Tasks", "abs": "In recent years, there has been a trend in the field of Reinforcement Learning (RL) towards large action models trained offline on large-scale datasets via sequence modeling. Existing models are primarily based on the Transformer architecture, which results in powerful agents. However, due to slow inference times, Transformer-based approaches are impractical for real-time applications, such as robotics. Recently, modern recurrent architectures, such as xLSTM and Mamba, have been proposed that exhibit parallelization benefits during training similar to the Transformer architecture while offering fast inference. In this work, we study the aptitude of these modern recurrent architectures for large action models. Consequently, we propose a Large Recurrent Action Model (LRAM) with an xLSTM at its core that comes with linear-time inference complexity and natural sequence length extrapolation abilities. Experiments on 432 tasks from 6 domains show that LRAM compares favorably to Transformers in terms of performance and speed.", "keywords": [ "reinforcement learning", "rnn", "xlstm", "mamba", "multi-task", "robotics" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "T1OvCSFaum", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Binhua Li", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Jie Cheng", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Science", "country": "CN", "position": "PhD student" }, { "name": "Qinghai Miao", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Ruixi Qiao", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yingwei Ma", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yisheng Lv", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Science", "country": "CN", "position": "Full Professor" }, { "name": "Yongbin Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xiong Gang", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Science", "country": "", "position": "Full Professor" } ], "n_formula": 30, "n_formula_1": 19, "n_ref_uni": 63, "n_ref": 111, "n_ref_all": 150, "n_fig": 2, "n_tab": 9, "L_tab": 3739, "n_element_tab": 497, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 657, "n_element_tab_1": 91, "formula_len_all": 3009, "formula_len_all_1": 1757, "len_all": 197225, "len_all_1": 65149, "len_abs": 1588, "len_title": 131, "len_sents": 48766, "len_sents_1": 29637, "n_sents": 384, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1594, "title": "Scaling Offline Model-Based RL via Jointly-Optimized World-Action Model Pretraining", "abs": "A significant aspiration of offline reinforcement learning (RL) is to develop a generalist agent with high capabilities from large and heterogeneous datasets. However, prior approaches that scale offline RL either rely heavily on expert trajectories or struggle to generalize to diverse unseen tasks. Inspired by the excellent generalization of world model in conditional video generation, we explore the potential of image observation-based world model for scaling offline RL and enhancing generalization on novel tasks. In this paper, we introduce JOWA: Jointly-Optimized World-Action model, an offline model-based RL agent pretrained on multiple Atari games with 6 billion tokens data to learn general-purpose representation and decision-making ability. Our method jointly optimizes a world-action model through a shared transformer backbone, which stabilize temporal difference learning with large models during pretraining. Moreover, we propose a provably efficient and parallelizable planning algorithm to compensate for the Q-value estimation error and thus search out better policies. Experimental results indicate that our largest agent, with 150 million parameters, achieves 78.9% human-level performance on pretrained games using only 10% subsampled offline data, outperforming existing state-of-the-art large-scale offline RL baselines by 31.6% on averange. Furthermore, JOWA scales favorably with model capacity and can sample-efficiently transfer to novel games using only 5k offline fine-tuning data (approximately 4 trajectories) per game, demonstrating superior generalization.", "keywords": [ "reinforcement learning", "offline reinforcement learning", "world model" ], "rating_list": [ 8, 6, 6, 8, 5 ], "soundness_list": [ 4, 4, 3, 4, 3 ], "presentation_list": [ 3, 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "T1MTmAlF7x", "primary_area": "generative models", "n_author": 1, "author_info_list": [ { "name": "Zen Li", "gender": "unknown", "institution": "Independent", "country": "", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 65, "n_ref": 129, "n_ref_all": 151, "n_fig": 4, "n_tab": 5, "L_tab": 1658, "n_element_tab": 186, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 776, "n_element_tab_1": 69, "formula_len_all": 395, "formula_len_all_1": 392, "len_all": 222488, "len_all_1": 57019, "len_abs": 1270, "len_title": 126, "len_sents": 33045, "len_sents_1": 26945, "n_sents": 228, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1255, "title": "EDM2+: Exploring Efficient Diffusion Model Architectures for Visual Generation", "abs": "The training and sampling of diffusion models have been exhaustively elucidated in prior art. Instead, the underlying network architecture design remains on a shaky empirical footing. Furthermore, in accordance with the recent trend of scaling law, large-scale models make inroads into generative vision tasks. However, running such large diffusion models incurs a sizeable computational burden, rendering it desiderata to optimize calculations and efficiently allocate resources. To bridge these gaps, we navigate the design landscape of efficient U-Net based diffusion models, stemming from the prestigious EDM2. Our exploration route is organized along two key axes, layer placement and module interconnection. We systematically study fundamental design choices and uncover several intriguing insights for superior efficacy and efficiency. These findings culminate in our redesigned architecture, EDM2+, that reduces the computational complexity of the baseline EDM2 by $2\\times$ without compromising the generation quality. Extensive experiments and comparative analyses highlight the effectiveness of our proposed network architecture, which achieves the state-of-the-art FID on the hallmark ImageNet benchmark. Code will be released upon acceptance.", "keywords": [ "diffusion models", "network architecture design" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "T0ebbDO60R", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Han Qiu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Zhang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Jiwei Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Qing Guo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Runyi Hu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Research Associate" }, { "name": "Tianwei Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Yiming Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Research Fellow" } ], "n_formula": 10, "n_formula_1": 4, "n_ref_uni": 36, "n_ref": 60, "n_ref_all": 82, "n_fig": 11, "n_tab": 8, "L_tab": 4758, "n_element_tab": 340, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2416, "n_element_tab_1": 304, "formula_len_all": 362, "formula_len_all_1": 206, "len_all": 132345, "len_all_1": 71539, "len_abs": 2160, "len_title": 140, "len_sents": 38943, "len_sents_1": 31526, "n_sents": 298, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 2171, "title": "SuperMark: Robust and Training-free Image Watermarking via Diffusion-based Super-Resolution", "abs": "In today's digital landscape, the intermingling of AI-generated and authentic content has heightened the importance of copyright protection and content authentication. Watermarking has emerged as a crucial technology to address these challenges, offering a general approach to safeguard both generated and real content. To be effective, watermarking methods must withstand various distortions and attacks. While current deep watermarking techniques typically employ an encoder\u2013noise layer\u2013decoder architecture and incorporate various distortions to enhance robustness, they often struggle to balance robustness and fidelity, and remain vulnerable to adaptive attacks, despite extensive training. To overcome these limitations, we propose SuperMark, a novel robust and training-free watermarking framework. Our approach draws inspiration from the parallels between watermark embedding/extraction in watermarking models and the denoising/noising processes in diffusion models. Specifically, SuperMark embeds the watermark into initial Gaussian noise using existing techniques and then applies pretrained Super-Resolution (SR) models to denoise the watermarked noise, producing the final watermarked image. For extraction, the process is reversed: the watermarked image is converted back to the initial watermarked noise via DDIM Inversion, from which the embedded watermark is then extracted. This flexible framework supports various noise injection methods and diffusion-based SR models, allowing for enhanced performance customization. The inherent robustness of the DDIM Inversion process against various perturbations enables SuperMark to demonstrate strong resilience to many distortions while maintaining high fidelity. Extensive experiments demonstrate SuperMark's effectiveness, achieving fidelity comparable to existing methods while significantly surpassing most in terms of robustness. Under normal distortions, SuperMark achieves an average watermark extraction bit accuracy of 99.46\\%, and 89.29\\% under adaptive attacks. Furthermore, SuperMark exhibits strong transferability across different datasets, SR models, watermark embedding methods, and resolutions.", "keywords": [ "image watermarking", "diffusion models" ], "rating_list": [ 3, 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 5, 5, 4, 4, 4 ] }, { "paper_id": "T0Df1Os6y0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Bingchen Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Xin Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Postdoc" }, { "name": "Zhibo Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Lu Yiting", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 36, "n_ref": 85, "n_ref_all": 100, "n_fig": 4, "n_tab": 12, "L_tab": 3362, "n_element_tab": 463, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 4905, "n_element_tab_1": 409, "formula_len_all": 260, "formula_len_all_1": 260, "len_all": 154074, "len_all_1": 61511, "len_abs": 1987, "len_title": 132, "len_sents": 37278, "len_sents_1": 24687, "n_sents": 268, "n_sents_1": 158, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1615, "title": "LossAgent: Towards Any Optimization Objectives for Image Processing with LLM Agents", "abs": "We present the first embodied loss agent, dubbed LossAgent, for low-level image processing tasks, e.g., image super-resolution and restoration, intending to achieve any customized optimization objectives of low-level image processing in different practical applications. Notably, not all optimization objectives, such as complex hand-crafted perceptual metrics, text description, and intricate human feedback, can be instantiated with existing low-level losses, e.g., MSE loss. which presents a crucial challenge in optimizing image processing networks in an end-to-end manner. To eliminate this, our LossAgent introduces the powerful large language model (LLM) as the embodied loss agent, where the rich textual understanding prior knowledge empowers the loss agent with the potential to understand complex optimization objectives, trajectory, and state feedback from external environments in the optimization process of the low-level image processing networks. In particular, we establish the loss repository by incorporating existing loss functions that support the end-to-end optimization for low-level image processing. Then, we design the optimization-oriented prompt engineering for the loss agent to actively and intelligently decide the compositional weights for each loss in the repository at each optimization interaction, thereby achieving the required optimization trajectory for any customized optimization objectives. Extensive experiments on three typical low-level image processing tasks and multiple optimization objectives have shown the effectiveness and applicability of our proposed LossAgent.", "keywords": [ "Image processing; LLM; Agent" ], "rating_list": [ 3, 5, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 4, 4 ] }, { "paper_id": "T01rY5kQoo", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Bochuan Cao", "gender": "unknown", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Jinghui Chen", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Lu Lin", "gender": "Female", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Yujia Wang", "gender": "unknown", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Yurui Chang", "gender": "Female", "institution": "Pennsylvania State University", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 45, "n_ref": 67, "n_ref_all": 93, "n_fig": 5, "n_tab": 3, "L_tab": 467, "n_element_tab": 27, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 500, "formula_len_all_1": 518, "len_all": 157509, "len_all_1": 69099, "len_abs": 1199, "len_title": 128, "len_sents": 53171, "len_sents_1": 35471, "n_sents": 358, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1212, "title": "JoPA: Explaining Large Language Model's Generation via Joint Prompt Attribution", "abs": "Large Language Models (LLMs) have demonstrated impressive performances in complex text generation tasks. However, the contribution of the input prompt to the generated content still remains obscure to humans, underscoring the necessity of elucidating and explaining the causality between input and output pairs. Existing works for providing prompt-specific explanation often confine model output to be classification or next-word prediction. Few initial attempts aiming to explain the entire language generation often treat input prompt texts independently, ignoring their combinatorial effects on the follow-up generation. In this study, we introduce a counterfactual explanation framework based on joint prompt attribution, JoPA, which aims to explain how a few prompt texts collaboratively influences the LLM's complete generation. Particularly, we formulate the task of prompt attribution for generation interpretation as a combinatorial optimization problem, and introduce a probabilistic algorithm to search for the casual input combination in the discrete space. We define and utilize multiple metrics to evaluate the produced explanations, demonstrating both faithfulness and efficiency of our framework.", "keywords": [ "Large Language Model", "Explainability" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "SzWvRzyk6h", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Osama Khalid", "gender": "Male", "institution": "University of Iowa", "country": "", "position": "PhD student" }, { "name": "Padmini Srinivasan", "gender": "Female", "institution": "University of Iowa", "country": "US", "position": "Full Professor" }, { "name": "Sanvesh Srivastava", "gender": "Male", "institution": "University of Iowa", "country": "", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 26, "n_ref": 46, "n_ref_all": 53, "n_fig": 4, "n_tab": 3, "L_tab": 3442, "n_element_tab": 54, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 883, "n_element_tab_1": 29, "formula_len_all": 674, "formula_len_all_1": 651, "len_all": 96425, "len_all_1": 63597, "len_abs": 1502, "len_title": 96, "len_sents": 32468, "len_sents_1": 29061, "n_sents": 254, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1517, "title": "SLIM-LLMs: Low-Rank Models of Linguistic Style", "abs": "Linguistic style encompasses a range of dimensions, including sensorial language as well as traditional stylistic features (represented using LIWC features). While these dimensions of linguistic style have been studied independently, relationships between the different dimensions, particularly between sensorial style and traditional stylistic features, remain understudied. This paper introduces a novel approach to model this interaction and tests it across a diverse set of texts. \nIn particular, we propose using a Reduced-Rank Ridge Regression (R4) to model low-rank latent relationships between LIWC-based stylistic features and sensorial language features. We find that compared to the full LIWC feature set ($r = 74$), its low-dimensional latent representations ($r = 24$) effectively capture stylistic information relevant to sensorial language prediction.\nBased on our results, we propose Stylometrically Lean Interpretable Models (SLIM-LLMs) \u2014 dimensionality-reduced LLMs that model the non-linear relationships between these two major dimensions of style. We evaluate SLIM-LLMs on the ability to predict sensorial language (the actual sensorial words used) in five text genres: business reviews, novels, song lyrics, advertisements, and informative articles. Results show that SLIM-LLMs augmented with low-rank style features consistently outperform baseline models. These SLIM-LLMs approach the performance of full-scale language models while using significantly fewer parameters (up to 80\\% reduction).", "keywords": [ "Linguistic Style", "Sensorial Linguistics", "LIWC", "SLIM-LLMs" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "SzPZK856iI", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Chen Cheng", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Fayao Liu", "gender": "Female", "institution": "A*STAR", "country": "SG", "position": "Research Scientist" }, { "name": "Xiaofeng Yang", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Xulei Yang", "gender": "Male", "institution": "I2R, A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Guosheng Lin", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" } ], "n_formula": 18, "n_formula_1": 10, "n_ref_uni": 36, "n_ref": 74, "n_ref_all": 92, "n_fig": 12, "n_tab": 3, "L_tab": 815, "n_element_tab": 90, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 566, "n_element_tab_1": 9, "formula_len_all": 1418, "formula_len_all_1": 951, "len_all": 124824, "len_all_1": 51743, "len_abs": 1162, "len_title": 100, "len_sents": 34105, "len_sents_1": 23351, "n_sents": 324, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1171, "title": "Text-to-Image Rectified Flow as Plug-and-Play Priors", "abs": "Large-scale diffusion models have achieved remarkable performance in generative tasks. Beyond their initial training applications, these models have proven their ability to function as versatile plug-and-play priors. For instance, 2D diffusion models can serve as loss functions to optimize 3D implicit models. Rectified Flow, a novel class of generative models, has demonstrated superior performance across various domains. Compared to diffusion-based methods, rectified flow approaches surpass them in terms of generation quality and efficiency. In this work, we present theoretical and experimental evidence demonstrating that rectified flow based methods offer similar functionalities to diffusion models \u2014 they can also serve as effective priors. Besides the generative capabilities of diffusion priors, motivated by the unique time-symmetry properties of rectified flow models, a variant of our method can additionally perform image inversion. Experimentally, our rectified flow based priors outperform their diffusion counterparts \u2014 the SDS and VSD losses \u2014 in text-to-3D generation. Our method also displays competitive performance in image inversion and editing.", "keywords": [ "3D Generation", "Rectified Flow", "Flow Matching" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "Sz2Ar6EqD5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "BO WANG", "gender": "Male", "institution": "Vector Institute", "country": "", "position": "Assistant Professor" }, { "name": "Jun Ma", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "Postdoc" }, { "name": "Sumin Kim", "gender": "unknown", "institution": "University of Toronto", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 30, "n_ref": 39, "n_ref_all": 47, "n_fig": 3, "n_tab": 6, "L_tab": 5623, "n_element_tab": 318, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 1197, "n_element_tab_1": 69, "formula_len_all": 184, "formula_len_all_1": 184, "len_all": 120864, "len_all_1": 57041, "len_abs": 1050, "len_title": 132, "len_sents": 34935, "len_sents_1": 27613, "n_sents": 226, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1142, "title": "All You Need Is A Reference: Cross-modality Referring Segmentation for Abdominal MRI", "abs": "Multi-modality MRI scans can provide comprehensive diagnoses of abdominal disease but this also introduces new segmentation burdens to derive quantitative imaging biomarkers. In this work, we propose a referring segmentation task where users only need to draw simple scribbles on one modality, called reference modality, to guide the segmentation of both the unseen target modalities and the reference modality. To benchmark the multi-modality segmentation task, we provide a new dataset with 3,277 organs from 534 MRI scans, covering five commonly used MRI modalities. Furthermore, we present a referring segmentation model, CrossMR, to simultaneously segment multiple modalities based on scribbles on reference modality. Experiments demonstrate that our method can achieve comparable performance to the state of the art on one in-distribution reference modality and significantly better generalization ability on four out-of-distribution modalities. This opens a door for efficiently segmenting targets across multiple modalities. The new dataset, code, and trained model weights will be publicly available at https://ref-seg-mr.github.io/.", "keywords": [ "promptable segmentation model", "cross-modality", "referring segmentation" ], "rating_list": [ 1, 5, 3, 6, 5 ], "soundness_list": [ 2, 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 4, 4 ] }, { "paper_id": "Syt4fWwVm1", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Lei Chen", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Sibo Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Xiangyu Dong", "gender": "unknown", "institution": "Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xingyi Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" } ], "n_formula": 33, "n_formula_1": 17, "n_ref_uni": 30, "n_ref": 90, "n_ref_all": 127, "n_fig": 11, "n_tab": 11, "L_tab": 10124, "n_element_tab": 575, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 3535, "n_element_tab_1": 58, "formula_len_all": 4448, "formula_len_all_1": 1261, "len_all": 167330, "len_all_1": 68650, "len_abs": 1576, "len_title": 149, "len_sents": 53205, "len_sents_1": 30229, "n_sents": 365, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 99, "L_abs": 1651, "title": "SpaceGNN: Multi-Space Graph Neural Network for Node Anomaly Detection with Extremely Limited Labels", "abs": "Node Anomaly Detection (NAD) has gained significant attention in the deep learning community due to its diverse applications in real-world scenarios. \nExisting NAD methods primarily embed graphs within a single Euclidean space, while overlooking the potential of non-Euclidean spaces. \nBesides, to address the prevalent issue of limited supervision in real NAD tasks, previous methods tend to leverage synthetic data to collect auxiliary information, which is not an effective solution as shown in our experiments.\nTo overcome these challenges, we introduce a novel SpaceGNN model designed for NAD tasks with extremely limited labels. \nSpecifically, we provide deeper insights into a task-relevant framework by empirically analyzing the benefits of different spaces for node representations, based on which, we design a Learnable Space Projection function that effectively encodes nodes into suitable spaces.\nBesides, we introduce the concept of weighted homogeneity, which we empirically and theoretically validate as an effective coefficient during information propagation. This concept inspires the design of the Distance Aware Propagation module. \nFurthermore, we propose the Multiple Space Ensemble module, which extracts comprehensive information for NAD under conditions of extremely limited supervision. Our findings indicate that this module is more beneficial than data augmentation techniques for NAD. Extensive experiments conducted on 9 real datasets confirm the superiority of SpaceGNN, which outperforms the best rival by an average of 8.55% in AUC and 4.31% in F1 scores. Our code is available at https://github.com/xydong127/SpaceGNN.", "keywords": [ "Node Anomaly Detection", "Graph Neural Network", "Multiple Spaces" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "SyVPiehSbg", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Leonardo Ferreira Guilhoto", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Paris Perdikaris", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Associate Professor" } ], "n_formula": 42, "n_formula_1": 18, "n_ref_uni": 42, "n_ref": 97, "n_ref_all": 147, "n_fig": 13, "n_tab": 11, "L_tab": 1880, "n_element_tab": 199, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 362, "n_element_tab_1": 28, "formula_len_all": 4345, "formula_len_all_1": 1227, "len_all": 188806, "len_all_1": 63213, "len_abs": 1407, "len_title": 114, "len_sents": 57366, "len_sents_1": 30544, "n_sents": 424, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 66, "L_abs": 1411, "title": "Deep Learning Alternatives Of The Kolmogorov Superposition Theorem", "abs": "This paper explores alternative formulations of the Kolmogorov Superposition Theorem (KST) as a foundation for neural network design. The original KST formulation, while mathematically elegant, presents practical challenges due to its limited insight into the structure of inner and outer functions and the large number of unknown variables it introduces. Kolmogorov-Arnold Networks (KANs) leverage KST for function approximation, but they have faced scrutiny due to mixed results compared to traditional multilayer perceptrons (MLPs) and practical limitations imposed by the original KST formulation. To address these issues, we introduce ActNet, a scalable deep learning model that builds on the KST and overcomes some of the drawbacks of Kolmogorov's original formulation. We evaluate ActNet in the context of Physics-Informed Neural Networks (PINNs), a framework well-suited for leveraging KST's strengths in low-dimensional function approximation, particularly for simulating partial differential equations (PDEs). In this challenging setting, where models must learn latent functions without direct measurements, ActNet consistently outperforms KANs across multiple benchmarks and is competitive against the current best MLP-based approaches. These results present ActNet as a promising new direction for KST-based deep learning applications, particularly in scientific computing and PDE simulation tasks.", "keywords": [ "Kolmogorov-Arnold Representation Theorem", "Function Approximation", "Physics Informed Neural Networks", "AI4Science" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 4, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 3, 4, 2, 4 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "SyV8rldw49", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Chengxiang Wen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Jian Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Jiarui Meng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Jie Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Luyang Tang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Qiankun Gao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD Student" }, { "name": "Ronggang Wang", "gender": "Male", "institution": "Peking University Shenzhen Graduate School", "country": "CN", "position": "Full Professor" }, { "name": "Yanmin Wu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 4, "n_ref": 20, "n_ref_all": 38, "n_fig": 9, "n_tab": 5, "L_tab": 4278, "n_element_tab": 280, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1004, "n_element_tab_1": 194, "formula_len_all": 494, "formula_len_all_1": 541, "len_all": 114678, "len_all_1": 62203, "len_abs": 1802, "len_title": 158, "len_sents": 37237, "len_sents_1": 29402, "n_sents": 268, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 109, "L_abs": 1818, "title": "RelayGS: Reconstructing High-Fidelity Dynamic Scenes with Large-Scale and Complex Motions via Relay Gaussians", "abs": "Reconstructing dynamic scenes with large-scale and complex motions\u2014such as those in sports events\u2014remains a significant challenge. Recent techniques like Neural Radiance Field and Gaussian Splatting have shown promise but often struggle with scenes involving substantial movement. In this paper, we propose **RelayGS**, a novel dynamic scene reconstruction method based on Gaussian Splatting, specifically designed to represent and learn large-scale complex motion patterns in highly dynamic scenes. Our RelayGS consists of three key stages. First, we learn the fundamental scene structure from all frames without considering temporal information and employ a learnable mask to decouple the highly dynamic foreground from the background exhibiting minimal motion. Second, we partition the scene into temporal segments, each consisting of several consecutive multi-view frames. For each segment, we replicate the foreground Gaussians, dubbed **Relay Gaussians**, as they are designed to act as relay nodes along the large-scale motion trajectory. By creating pseudo-views from frames uniformly selected from the segment, we optimize and densify foreground Relay Gaussians, further simplify and decompose large-scale motion trajectories into smaller, more manageable segments. Finally, we leverage HexPlane and lightweight MLPs to jointly learn the scene\u2019s temporal motion field and refine the canonical Gaussians. We conduct extensive experiments on two dynamic scene datasets featuring large and complex motions to demonstrate the effectiveness of our RelayGS. RelayGS outperforms state-of-the-arts by more than 1 dB in PSNR, and successfully reconstructs real-world basketball game scenes in a much more complete and coherent manner, whereas previous methods usually struggle to capture the complex motion of players.", "keywords": [ "Gaussian Splatting", "Dynamic Scene Reconstruction", "4D Reconstruction" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "SyPrLti4PG", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Kabir Vinay Dabholkar", "gender": "unknown", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "PhD student" }, { "name": "Omri Barak", "gender": "Male", "institution": "Technion", "country": "IL", "position": "Associate Professor" } ], "n_formula": 29, "n_formula_1": 15, "n_ref_uni": 35, "n_ref": 67, "n_ref_all": 113, "n_fig": 25, "n_tab": 2, "L_tab": 325, "n_element_tab": 52, "n_fig_1": 16, "n_tab_1": 1, "L_tab_1": 12, "n_element_tab_1": 1, "formula_len_all": 1403, "formula_len_all_1": 816, "len_all": 164288, "len_all_1": 62998, "len_abs": 1832, "len_title": 131, "len_sents": 46447, "len_sents_1": 26726, "n_sents": 393, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1834, "title": "When predict can also explain: few-shot prediction to select better neural latents", "abs": "Latent variable models serve as powerful tools to infer underlying dynamics from observed neural activity. Ideally, one would like the inferred dynamics to equal the true ones. However, due to the absence of ground truth data, prediction benchmarks are often employed as proxies. One widely-used method is _co-smoothing_, which involves jointly estimating latent variables and predicting observations along held-out channels to assess model performance. In this study, we reveal the limitations of the co-smoothing prediction framework and propose a remedy. Utilizing a student-teacher setup with Hidden Markov Models, we demonstrate that the high co-smoothing model space can encompass models with arbitrary extraneous dynamics within their latent representations. To address this, we introduce a secondary metric\u2014 _few-shot co-smoothing_. This involves performing regression from the latent variables to held-out channels in the data using fewer trials. Our results indicate that among models with near-optimal co-smoothing, those with extraneous dynamics underperform in the few-shot co-smoothing compared to 'minimal' models devoid of such dynamics. We also provide analytical insights into the origin of this phenomenon. We further validate our findings on real neural data using two state-of-the-art methods: LFADS and STNDT. In the absence of ground truth, we suggest a novel measure to validate our approach. By cross-decoding the latent variables of all model pairs with high co-smoothing, we identify models with minimal extraneous dynamics. We find a correlation between few-shot co-smoothing performance and this new measure. In summary, we present a novel prediction metric designed to yield latent variables that more accurately reflect the ground truth, offering a significant improvement for latent dynamics inference.", "keywords": [ "Latent variable methods", "dynamical systems", "neural data", "benchmarking" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 2, 3, 2 ] }, { "paper_id": "Sxi6gBtJcI", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Hui Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Jiajing Zhang", "gender": "Male", "institution": "Anhui Jianzhu University", "country": "CN", "position": "Full Professor" }, { "name": "Qifan Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Xianyong Fang", "gender": "Male", "institution": "Anhui University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaohua Xu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yi Xiao", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Zefeng Lin", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zhengyi Liu", "gender": "unknown", "institution": "Anhui University", "country": "CN", "position": "Full Professor" }, { "name": "Zhiqiang Mo", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "\u9648\u5bb6\u6d0b", "gender": "Male", "institution": "\u5b89\u5fbd\u5efa\u7b51\u5927\u5b66", "country": "CN", "position": "MS student" }, { "name": "Wang Jie", "gender": "Male", "institution": "\u5b89\u5fbd\u5efa\u7b51\u5927\u5b66", "country": "CN", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 24, "n_ref": 41, "n_ref_all": 62, "n_fig": 9, "n_tab": 8, "L_tab": 2199, "n_element_tab": 294, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 3938, "n_element_tab_1": 309, "formula_len_all": 140, "formula_len_all_1": 142, "len_all": 106975, "len_all_1": 68682, "len_abs": 1503, "len_title": 131, "len_sents": 34280, "len_sents_1": 28152, "n_sents": 268, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1465, "title": "R$^2$: A LLM Based Novel-to-Screenplay Generation Framework with Causal Plot Graphs", "abs": "Automatically adapting novels into screenplays is important for the TV, film, or opera industries to promote products with low costs. The strong performances of large language models (LLMs) in long-text generation call us to propose a LLM based framework Reader-Rewriter (R$^2$) for this task. However, there are two fundamental challenges here. First, the LLM hallucinations may cause inconsistent plot extraction and screenplay generation. Second, the causality-embedded plot lines should be effectively extracted for coherent rewriting. Therefore, two corresponding tactics are proposed: 1) A hallucination-aware refinement method (HAR) to iteratively discover and eliminate the affections of hallucinations; and 2) a causal plot-graph construction method (CPC) based on a greedy cycle-breaking algorithm to efficiently construct plot lines with event causalities. Recruiting those efficient techniques, R$^2$ utilizes two modules to mimic the human screenplay rewriting process: The Reader module adopts a sliding window and CPC to build the causal plot graphs, while the Rewriter module generates first the scene outlines based on the graphs and then the screenplays. HAR is integrated into both modules for accurate inferences of LLMs. Experimental results demonstrate the superiority of R$^2$, which substantially outperforms three existing approaches (51.3\\%, 22.6\\%, and 57.1\\% absolute increases) in pairwise comparison at the overall win rate for GPT-4o.", "keywords": [ "Novel-to-Screenplay Generation", "Plot Graphs", "Large Lanuage Models", "Self-Refinement" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "SxOrhLuuVz", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 12, "author_info_list": [ { "name": "Feng Sun", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Furu Wei", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Partner Research Manager" }, { "name": "Fuzhen Zhuang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Professor" }, { "name": "Haizhen Huang", "gender": "Male", "institution": "Institute of Computing Technolgy, Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Qi Zhang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Shaohan Huang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Shengyue Luo", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Songtao Wang", "gender": "Male", "institution": "Beijing Zhongguancun Lab", "country": "", "position": "Researcher" }, { "name": "Ting Jiang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" }, { "name": "Weiwei Deng", "gender": "Male", "institution": "South China University of Technology", "country": "", "position": "PhD student" }, { "name": "Zihan Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "MS student" }, { "name": "deqing wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 17, "n_ref": 32, "n_ref_all": 49, "n_fig": 4, "n_tab": 8, "L_tab": 4322, "n_element_tab": 500, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 868, "n_element_tab_1": 130, "formula_len_all": 754, "formula_len_all_1": 751, "len_all": 118321, "len_all_1": 58412, "len_abs": 1078, "len_title": 108, "len_sents": 26620, "len_sents_1": 23725, "n_sents": 220, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1083, "title": "MoRA: High-Rank Updating for Parameter-Efficient Fine-Tuning", "abs": "Low-rank adaptation (LoRA) is a popular parameter-efficient fine-tuning (PEFT) method for large language models (LLMs). In this paper, we analyze the impact of low-rank updating, as implemented in LoRA. Our findings suggest that the low-rank updating mechanism may limit the ability of LLMs to effectively learn and memorize new knowledge. Inspired by this observation, we propose a new method called MoRA, which employs a square matrix to achieve high-rank updating while maintaining the same number of trainable parameters. To achieve it, we introduce the corresponding non-parameter operators to reduce the input dimension and increase the output dimension for the square matrix. Furthermore, these operators ensure that the weight can be merged back into LLMs, which enables our method to be deployed like LoRA. We perform a comprehensive evaluation of our method across five tasks: instruction tuning, mathematical reasoning, continual pretraining, memory and pretraining. Our method outperforms LoRA on memory-intensive tasks and achieves comparable performance on other tasks.", "keywords": [ "Large Language Model", "LoRA", "Parameter-Efficient Fine-Tuning" ], "rating_list": [ 5, 3, 3, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "SwRlpDKrY9", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Ru Peng", "gender": "Female", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xingyu Chen", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ziru Wang", "gender": "unknown", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xuguang Lan", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 24, "n_formula_1": 19, "n_ref_uni": 36, "n_ref": 50, "n_ref_all": 64, "n_fig": 3, "n_tab": 4, "L_tab": 464, "n_element_tab": 123, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 200, "n_element_tab_1": 34, "formula_len_all": 1361, "formula_len_all_1": 1127, "len_all": 127394, "len_all_1": 56600, "len_abs": 1813, "len_title": 105, "len_sents": 28956, "len_sents_1": 26169, "n_sents": 205, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1822, "title": "ReDebias: Exploring Residual energy based Debias learning", "abs": "In real-world applications, ensuring that model decisions are independent of the training data distribution is crucial for safely deploying models. To address the long-tailed problem, massive approaches focus either on improving individual prediction quality or enhancing aggregate evaluation. Although these methods improve overall performance, they often sacrifice performance in some classes, undermining the goals of long-tailed learning. We conduct a mathematical analysis of the limitations of the Empirical Risk Minimization (ERM) framework in long-tailed learning, examining both individual performance and aggregate evaluation. For individual evaluation, although the Negative log-likelihood (NLL) metric is effective, it relies heavily on softmax leading to poor distinction and ambiguity when the probabilities of correct and incorrect predictions are similar. For aggregate evaluation, the naive estimator in ERM is not an unbiased estimator, dominated by head classes. To overcome these challenges, we propose Re-Debias, a comprehensive framework combining the Residual-Energy score and a Debias estimator. The Residual-Energy score provides a more sensitive reflection of prediction quality than softmax-based scores, enhancing prediction precision and reducing ambiguity. The Debias estimator applies causal inference techniques to ensure unbiased estimates during the averaging process, correcting for class-wise biases inherent in the naive estimator. Through extensive validation on long-tailed benchmarks, including training from scratch on iNaturalist18, ImageNet-LT, and CIFAR10/100-LT, as well as fine-tuning Vision Transformer (ViT) on iNaturalist18, our method outperforms the state-of-the-art algorithms. Our code and trained models will be made available following the publication of this paper.", "keywords": [ "deep long-tailed learning", "Debiasing Learning", "Energy-Based Learning" ], "rating_list": [ 3, 5, 5, 8, 5 ], "soundness_list": [ 3, 2, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3, 5 ] }, { "paper_id": "SwIkknEqmt", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Changhua Pei", "gender": "Male", "institution": "Computer Network Information Center, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Dan Pei", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Gaogang Xie", "gender": "Male", "institution": "Computer Network Information Center, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Haiming Zhang", "gender": "Male", "institution": "cnic", "country": "", "position": "Full Professor" }, { "name": "Haotian Si", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Jianhui li", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 10, "n_ref_uni": 26, "n_ref": 51, "n_ref_all": 77, "n_fig": 14, "n_tab": 5, "L_tab": 4761, "n_element_tab": 203, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 278, "n_element_tab_1": 34, "formula_len_all": 1325, "formula_len_all_1": 629, "len_all": 146748, "len_all_1": 56290, "len_abs": 1001, "len_title": 134, "len_sents": 42433, "len_sents_1": 25699, "n_sents": 328, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1004, "title": "Dealing with Frequency Collapse in Time Series Embeddings by Post-Embedding reMapping", "abs": "Transformer-based methods have made significant strides in time series forecasting tasks in recent years. However, we observe underfitting in numerous samples, e.g., pattern shifts or excessive deviation in extreme value regions when testing the transform-based model that converges on the training set. Through the proposed spectral analysis of adjacent embedding sequences, we identify a frequency collapse issue in the embedding features generated by the top layer of the transformer backbone. To address this, we propose the Post-Embedding ReMapping (PErM) strategy that improves the frequency-domain representation of embeddings using fixed non-linear functions. Both two kinds of PErM functions that we insert into the model can effectively resolve the frequency collapse issue and lead to significant improvements in prediction performance. Experimental results show that our method outperforms state-of-the-art algorithms across multiple datasets. We will release our code after the review phase.", "keywords": [ "Time Series Forcasting", "deep learning", "spectral analysis" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 1, 3 ], "presentation_list": [ 3, 1, 2 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 4, 4, 2 ] }, { "paper_id": "Sw7c4fwpSC", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Alex Gomez-Villa", "gender": "Female", "institution": "Universitat Aut\u00f3noma de Barcelona", "country": "ES", "position": "Postdoc" }, { "name": "Bogdan Raducanu", "gender": "Male", "institution": "Computer Vision Center, Universitat Aut\u00f2noma de Barcelona", "country": "ES", "position": "Project Director" }, { "name": "Imad Eddine MAROUF", "gender": "Male", "institution": "IP-Paris", "country": "FR", "position": "PhD student" }, { "name": "Kai Wang", "gender": "Male", "institution": "Computer Vision Center, Universitat Aut\u00f3noma de Barcelona", "country": "ES", "position": "Postdoc" }, { "name": "Hector Laria Mantecon", "gender": "Male", "institution": "Computer Vision Center, Universitat Aut\u00f3noma de Barcelona", "country": "ES", "position": "PhD student" }, { "name": "Joost van de", "gender": "Male", "institution": "Universitat Aut\u00f3noma de Barcelona", "country": "ES", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 42, "n_ref": 87, "n_ref_all": 123, "n_fig": 17, "n_tab": 7, "L_tab": 3569, "n_element_tab": 500, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 815, "n_element_tab_1": 55, "formula_len_all": 393, "formula_len_all_1": 393, "len_all": 172406, "len_all_1": 61029, "len_abs": 786, "len_title": 119, "len_sents": 49238, "len_sents_1": 29539, "n_sents": 352, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1486, "title": "Assessing Open-world Forgetting in Generative Image Model Customization", "abs": "Recent advances in diffusion models have significantly enhanced image generation capabilities. However, customizing these models with new classes often leads to unintended consequences that compromise their reliability. We introduce the concept of *open-world forgetting* to emphasize the vast scope of these unintended alterations, contrasting it with the well-studied *closed-world forgetting*, which is measurable by evaluating performance on a limited set of classes or skills.\nOur research presents the first comprehensive investigation into open-world forgetting in diffusion models, focusing on semantic and appearance drift of representations. We utilize zero-shot classification to analyze semantic drift, revealing that even minor model adaptations lead to unpredictable shifts affecting areas far beyond newly introduced concepts, with dramatic drops in zero-shot classification of up to 60\\%. Additionally, we observe significant changes in texture and color of generated content when analyzing appearance drift.\nTo address these issues, we propose a mitigation strategy based on functional regularization, designed to preserve original capabilities while accommodating new concepts. Our study aims to raise awareness of unintended changes due to model customization and advocates for the analysis of open-world forgetting in future research on model customization and finetuning methods. Furthermore, we provide insights for developing more robust adaptation methodologies.", "keywords": [ "generative models", "diffusion", "open-world forgetting", "semantic drift", "appearance drift" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "Sw10tbj0gM", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Bel\u00e9n Vega-M\u00e1rquez", "gender": "Female", "institution": "Universidad de Sevilla", "country": "", "position": "Lecturer" }, { "name": "Isabel A. Nepomuceno-Chamorro", "gender": "Female", "institution": "Universidad de Sevilla", "country": "", "position": "Associate Professor" }, { "name": "Javier Sol\u00eds-Garc\u00eda", "gender": "Male", "institution": "Avature", "country": "ES", "position": "Researcher" }, { "name": "Juan A. Nepomuceno", "gender": "Male", "institution": "Universidad de Sevilla", "country": "", "position": "Lecturer" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 34, "n_ref": 88, "n_ref_all": 104, "n_fig": 2, "n_tab": 4, "L_tab": 3016, "n_element_tab": 210, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 3464, "n_element_tab_1": 195, "formula_len_all": 884, "formula_len_all_1": 751, "len_all": 114584, "len_all_1": 65151, "len_abs": 1534, "len_title": 131, "len_sents": 32128, "len_sents_1": 28275, "n_sents": 225, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1541, "title": "TIMBA: Time series Imputation with Bi-directional Mamba Blocks and Diffusion models", "abs": "The problem of imputing multivariate time series spans a wide range of fields, from clinical healthcare to multi-sensor systems. Initially, Recurrent Neural Networks (RNNs) were employed for this task; however, their error accumulation issues led to the adoption of Transformers, leveraging attention mechanisms to mitigate these problems. Concurrently, the promising results of diffusion models in capturing original distributions have positioned them at the forefront of current research, often in conjunction with Transformers. In this paper, we propose replacing time-oriented Transformers with State-Space Models (SSM), which are better suited for temporal data modeling. Specifically, we utilize the latest SSM variant, S6, which incorporates attention-like mechanisms. By embedding S6 within Mamba blocks, we develop a model that integrates SSM, Graph Neural Networks, and node-oriented Transformers to achieve enhanced spatiotemporal representations. Implementing these architectural modifications, previously unexplored in this field, we present Time series Imputation with Bi-directional mamba blocks and diffusion models (TIMBA). TIMBA achieves superior performance in almost all benchmark scenarios and performs comparably in others across a diverse range of missing value situations and three real-world datasets. We also evaluate how the performance of our model varies with different amounts of missing values and analyse its performance on downstream tasks. In addition, we provide the original code to replicate the results.", "keywords": [ "Multivariate Time Series Imputation", "Diffusion models", "Mamba", "Space State Models", "Generative Models", "Graph Neural Networks", "Transformers" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 1, 1, 1, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "SvydqVoHrp", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Alexander Rubinstein", "gender": "unknown", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Armand Mihai Nicolicioiu", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Damien Teney", "gender": "Male", "institution": "Idiap Research Institute", "country": "CH", "position": "Researcher" }, { "name": "Luca Scimeca", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Postdoc" }, { "name": "Seong Joon Oh", "gender": "Male", "institution": "Parameter Lab", "country": "DE", "position": "Principal Researcher" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 48, "n_ref": 82, "n_ref_all": 109, "n_fig": 12, "n_tab": 3, "L_tab": 1651, "n_element_tab": 234, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 108, "n_element_tab_1": 21, "formula_len_all": 610, "formula_len_all_1": 570, "len_all": 147768, "len_all_1": 62294, "len_abs": 1030, "len_title": 129, "len_sents": 46745, "len_sents_1": 30943, "n_sents": 303, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1033, "title": "Mitigating Shortcut Learning with Diffusion Counterfactuals and Diverse Ensembles", "abs": "Spurious correlations in the data, where multiple cues are predictive of the target labels, often lead to a phenomenon known as shortcut learning, where a model relies on erroneous, easy-to-learn cues while ignoring reliable ones. In this work, we propose DiffDiv an ensemble diversification framework exploiting Diffusion Probabilistic Models (DPMs) to mitigate this form of bias. We show that at particular training intervals, DPMs can generate images with novel feature combinations, even when trained on samples displaying correlated input features. We leverage this crucial property to generate synthetic counterfactuals to increase model diversity via ensemble disagreement. We show that DPM-guided diversification is sufficient to remove dependence on shortcut cues, without a need for additional supervised signals. We further empirically quantify its efficacy on several diversification objectives, and finally show improved generalization and diversification on par with prior work that relies on auxiliary data collection.", "keywords": [ "Shortcut Learning", "Simplicity Bias", "Bias Mitigation", "Debias", "Ensembles", "Diffusion Models", "counterfactuals", "Diversification" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "Svt3SGy0yd", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Naresh Shanbhag", "gender": "Male", "institution": "University of Illinois, Urbana-Champaign", "country": "", "position": "Full Professor" }, { "name": "Vignesh Sundaresha", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 30, "n_ref": 66, "n_ref_all": 113, "n_fig": 20, "n_tab": 13, "L_tab": 5567, "n_element_tab": 764, "n_fig_1": 13, "n_tab_1": 8, "L_tab_1": 3631, "n_element_tab_1": 549, "formula_len_all": 1331, "formula_len_all_1": 539, "len_all": 154250, "len_all_1": 71061, "len_abs": 1195, "len_title": 113, "len_sents": 42635, "len_sents_1": 27751, "n_sents": 301, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1199, "title": "Growing Efficient Accurate and Robust Neural Networks on the Edge", "abs": "The ubiquitous deployment of deep learning systems on resource-constrained Edge devices is hindered by their high computational complexity coupled with their fragility to out-of-distribution (OOD) data, especially to naturally occurring common corruptions. Current solutions rely on the Cloud to train and compress models before deploying to the Edge. This incurs high energy and latency costs in transmitting locally acquired field data to the Cloud while also raising privacy concerns. We propose GEARnn (Growing Efficient, Accurate, and Robust neural networks) to grow and train robust networks in-situ, i.e., completely on the Edge device. Starting with a low-complexity initial backbone network, GEARnn employs One-Shot Growth (OSG) to grow a network satisfying the memory constraints of the Edge device using clean data, and robustifies the network using Efficient Robust Augmentation (ERA) to obtain the final network. We demonstrate results on a NVIDIA Jetson Xavier NX, and analyze the trade-offs between accuracy, robustness, model size, energy consumption, and training time. Our results demonstrate the construction of efficient, accurate, and robust networks entirely on an Edge device.", "keywords": [ "Corruption Robustness", "Growth", "Efficient Training", "Edge" ], "rating_list": [ 5, 5, 6, 5, 8 ], "soundness_list": [ 1, 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 4 ], "confidence_list": [ 3, 3, 3, 4, 3 ] }, { "paper_id": "SvpwQfO9H1", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Haohan Wang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Ke Chen", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Yifeng Wang", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 37, "n_ref": 45, "n_ref_all": 59, "n_fig": 3, "n_tab": 8, "L_tab": 11927, "n_element_tab": 1023, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 4980, "n_element_tab_1": 342, "formula_len_all": 1162, "formula_len_all_1": 1160, "len_all": 167152, "len_all_1": 59316, "len_abs": 1466, "len_title": 117, "len_sents": 35352, "len_sents_1": 23428, "n_sents": 257, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1470, "title": "ADAPT: Alzheimer's Diagnosis through Adaptive Profiling Transformers", "abs": "Automated diagnosis of Alzheimer\u2019s Disease (AD) from brain imaging, such as\nmagnetic resonance imaging (MRI), has become increasingly important and has\nattracted the community to contribute many deep learning methods. However,\nmany of these methods are facing a trade-off that 3D models tend to be inefficient\nin training and inferencing while 2D models cannot capture the full 3D intricacies\nfrom the data. In this paper, we introduce a new model structure for diagnosing AD,\nand it can complete with 3D model\u2019s performances while essentially is a 2D method\n(thus computationally efficient). While the core idea lies in building different blocks\non different views according to physicians\u2019 diagnosing perspectives, we introduce\nmultiple components that can further benefit the model in this new perspective,\nincluding adaptively selecting the number of sclices in each dimension, and the new\nattention mechanism. In addition, we also introduce a morphology augmentation,\nwhich also barely introduces new computational loads, but can help improve the\ndiagnosis performances due to its alignment to the pathology of AD. We name\nour method ADAPT, which stands for Alzheimer\u2019s Diagnosis through Adaptive\nProfiling Transformers. We test our model from a practical perspective (the testing\ndomains do not appear in the training one): the diagnosis accuracy favors our\nADAPT with 4.5% improvement, while ADAPT uses at leat 14% less parameters\nthan the state-of-the-art models.", "keywords": [ "Deep Learning for healthcare", "Alzheimer's Diagnosis" ], "rating_list": [ 1, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 1, 2 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "SvjFHucuDZ", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Haizhao Yang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Associate Professor" }, { "name": "Jiaxin Yuan", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Zezheng Song", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 12, "n_ref_uni": 52, "n_ref": 77, "n_ref_all": 88, "n_fig": 3, "n_tab": 10, "L_tab": 2531, "n_element_tab": 276, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 1450, "n_element_tab_1": 212, "formula_len_all": 1073, "formula_len_all_1": 579, "len_all": 157905, "len_all_1": 65605, "len_abs": 1660, "len_title": 148, "len_sents": 40694, "len_sents_1": 28628, "n_sents": 355, "n_sents_1": 235, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1711, "title": "FMint: Bridging Human Designed and Data Pretrained Models for Differential Equation Foundation Model", "abs": "The fast simulation of dynamical systems is a key challenge in many scientific and engineering applications, such as weather forecasting, disease control, and drug discovery. With the recent success of deep learning, there is increasing interest in using neural networks to solve differential equations in a data-driven manner. However, existing methods are either limited to specific types of differential equations or require large amounts of data for training. This restricts their practicality in many real-world applications, where data is often scarce or expensive to obtain. To address this, we propose a novel multi-modal foundation model, named \\textbf{FMint} (\\textbf{F}oundation \\textbf{M}odel based on \\textbf{In}i\\textbf{t}ialization), to bridge the gap between human-designed and data-driven models for the fast simulation of dynamical systems. Built on a decoder-only transformer architecture with in-context learning, FMint utilizes both numerical and textual data to learn a universal error correction scheme for dynamical systems, using prompted sequences of coarse solutions from traditional solvers. The model is pre-trained on a corpus of 40K ODEs, and we perform extensive experiments on challenging ODEs that exhibit chaotic behavior and of high dimensionality. Our results demonstrate the effectiveness of the proposed model in terms of both accuracy and efficiency compared to classical numerical solvers, highlighting FMint's potential as a general-purpose solver for dynamical systems. Our approach achieves an accuracy improvement of 1 to 2 orders of magnitude over state-of-the-art dynamical system simulators, and delivers a 5X speedup compared to traditional numerical algorithms.", "keywords": [ "dynamical systems", "foundation model", "in-context learning", "multi-modal" ], "rating_list": [ 6, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "SvCOhZRQqa", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Hyukjun Lee", "gender": "unknown", "institution": "Sogang University", "country": "KR", "position": "Full Professor" }, { "name": "TaeHan Lee", "gender": "Male", "institution": "Sogang University", "country": "KR", "position": "PhD student" }, { "name": "Woojin Lee", "gender": "Male", "institution": "Sogang University", "country": "KR", "position": "Undergrad student" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 29, "n_ref": 48, "n_ref_all": 69, "n_fig": 8, "n_tab": 4, "L_tab": 1646, "n_element_tab": 225, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 818, "n_element_tab_1": 151, "formula_len_all": 495, "formula_len_all_1": 236, "len_all": 113174, "len_all_1": 50989, "len_abs": 1391, "len_title": 154, "len_sents": 29339, "len_sents_1": 22012, "n_sents": 232, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 106, "L_abs": 1406, "title": "Token Pruning Meets Audio: Investigating Unique Behaviors in Vision Transformer-Based Audio Classification", "abs": "Vision Transformers (ViTs) have achieved state-of-the-art performance across various computer vision tasks. To reduce the high computational cost of ViTs, token pruning has been proposed to selectively remove tokens that are not crucial. While effective in vision tasks by discarding non-object regions, applying this technique to audio tasks presents unique challenges. In audio processing, distinguishing relevant from non-relevant regions is less straightforward. In this study, we applied token pruning to a ViT-based audio classification model using Mel-spectrograms and analyzed the trade-offs between model performance and computational cost. We show AudioMAE-TopK model can reduce MAC operations by $2\\times$ with less than a 1\\% decrease in accuracy for both speech command recognition and environmental sound classification. Notably, while many tokens from signal (high-intensity) regions were pruned, tokens from background (low-intensity) regions were frequently retained, indicating the model\u2019s reliance on these regions. In the ablation study, forcing the model to focus only on signal (high-intensity) regions led to lower accuracy, suggesting that background (low-intensity) regions contain unique, irreplaceable information for AudioMAE. In Addition, we find that when token pruning is applied, the supervised pre-trained AST model emphasizes tokens from signal regions more than AudioMAE.", "keywords": [ "Audio Spectrogram Transformer", "Token Pruning" ], "rating_list": [ 6, 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 4, 5 ] }, { "paper_id": "SvCFkNgjTE", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Biwei Huang", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Erdun Gao", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Haoze Hou", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Howard Bondell", "gender": "Male", "institution": "The University of Melbourne", "country": "AU", "position": "Full Professor" }, { "name": "Mingming Gong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Qiuhong Ke", "gender": "Female", "institution": "Monash University", "country": "AU", "position": "Lecturer" }, { "name": "Wenqin Liu", "gender": "Female", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" } ], "n_formula": 67, "n_formula_1": 30, "n_ref_uni": 45, "n_ref": 108, "n_ref_all": 146, "n_fig": 6, "n_tab": 8, "L_tab": 1578, "n_element_tab": 293, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 151, "n_element_tab_1": 15, "formula_len_all": 9381, "formula_len_all_1": 2638, "len_all": 193847, "len_all_1": 65172, "len_abs": 1408, "len_title": 118, "len_sents": 61688, "len_sents_1": 29016, "n_sents": 520, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1107, "title": "MissScore: High-Order Score Estimation in the Presence of Missing Data", "abs": "The first order derivative (score) of data density, typically estimated via denoising score matching, has emerged as an effective tool for modeling data distribution and generating synthetic data. Extending this concept to higher-order scores could uncover more detailed local information of the data distribution, enabling new applications. However, learning these high-order scores usually requires complete data, which is often unavailable in real-world scenarios such as healthcare and finance due to privacy and cost constraints. In this work, we introduce MissScore, a novel score-based framework for learning high-order scores from observations with missing data. We derive objective functions for estimating high-order scores under different missing data mechanisms and propose a new algorithm to handle missing data effectively. Our empirical results demonstrate that MissScore efficiently and accurately approximates high-order scores with missing data, while enhancing sampling speed and data quality, as validated through several downstream tasks, including data generation and causal discovery.", "keywords": [ "denoising score matching; missing data; causal discovery;" ], "rating_list": [ 6, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "SulRfnEVK4", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Assaf Arbelle", "gender": "Male", "institution": "International Business Machines", "country": "IL", "position": "Researcher" }, { "name": "Felipe Maia Polo", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "PhD student" }, { "name": "Leonid Karlinsky", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Leshem Choshen", "gender": "Not Specified", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Mikhail Yurochkin", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Research Staff Member" }, { "name": "Muhammad Jehanzeb Mirza", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Nimrod Shabtay", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Raja Giryes", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Associate Professor" }, { "name": "Sivan Doveh", "gender": "unknown", "institution": "Weizmann Institute of Science", "country": "IL", "position": "PhD student" }, { "name": "Wei Lin", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Researcher" }, { "name": "Yuekai Sun", "gender": "unknown", "institution": "University of Michigan", "country": "US", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 27, "n_ref": 77, "n_ref_all": 133, "n_fig": 18, "n_tab": 20, "L_tab": 13544, "n_element_tab": 1800, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 3690, "n_element_tab_1": 27, "formula_len_all": 420, "formula_len_all_1": 225, "len_all": 247706, "len_all_1": 70646, "len_abs": 3037, "len_title": 115, "len_sents": 60124, "len_sents_1": 35470, "n_sents": 442, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1577, "title": "LiveXiv - A Multi-Modal live benchmark based on Arxiv papers content", "abs": "The large-scale training of multi-modal models on data scraped from the web has shown outstanding utility in infusing these models with the required world knowledge to perform effectively on multiple downstream tasks. However, one downside of scraping data from the web can be the potential sacrifice of the benchmarks on which the abilities of these models are often evaluated. To safeguard against test data contamination and to truly test the abilities of these foundation models we propose LiveXiv: A scalable evolving live benchmark based on scientific ArXiv papers. LiveXiv accesses domain-specific manuscripts at any given timestamp and proposes to automatically generate visual question-answer pairs (VQA). This is done without any human-in-the-loop, using the multi-modal content in the manuscripts, like graphs, charts, and tables. Moreover, we introduce an efficient evaluation approach that estimates the performance of all models on the evolving benchmark using evaluations of only a subset of models. This significantly reduces the overall evaluation cost. We benchmark multiple open and proprietary Large Multi-modal Models (LMMs) on the first version of our benchmark, showing its challenging nature and exposing the models\u2019 true abilities, avoiding contamination. Lastly, in our commitment to high quality, we\nhave collected and evaluated a manually verified subset. By comparing its overall results to our automatic annotations, we have found that the performance variance is indeed minimal (<2.5%). Our dataset is available online anonymously on HuggingFace.", "keywords": [ "Multi-Modal Dataset", "Visual Question-Answering", "Efficient Evaluation" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "SuHScQv5gP", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Kenan Hasanaliyev", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Manav Shah", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Silas Alberti", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Stefano Ermon", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 40, "n_ref": 81, "n_ref_all": 100, "n_fig": 7, "n_tab": 4, "L_tab": 708, "n_element_tab": 151, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 931, "n_element_tab_1": 121, "formula_len_all": 1403, "formula_len_all_1": 1033, "len_all": 122409, "len_all_1": 50687, "len_abs": 1319, "len_title": 35, "len_sents": 29550, "len_sents_1": 23462, "n_sents": 220, "n_sents_1": 151, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 35, "L_abs": 1325, "title": "Data Unlearning in Diffusion Models", "abs": "Recent work has shown that diffusion models memorize and reproduce training data examples. At the same time, large copyright lawsuits and legislation such as GDPR have highlighted the need for erasing datapoints from diffusion models. However, retraining from scratch is often too expensive. This motivates the setting of data unlearning, i.e., the study of efficient techniques for unlearning specific datapoints from the training set. Existing concept unlearning techniques require an anchor prompt/class/distribution to guide unlearning, which is not available in the data unlearning setting. General-purpose machine unlearning techniques were found to be either unstable or failed to unlearn data. We therefore propose a family of new loss functions called Subtracted Importance Sampled Scores (SISS) that utilize importance sampling and are the first method to unlearn data with theoretical guarantees. SISS is constructed as a weighted combination between simpler objectives that are responsible for preserving model quality and unlearning the targeted datapoints. When evaluated on CelebA-HQ and MNIST, SISS achieved Pareto optimality along the quality and unlearning strength dimensions. On Stable Diffusion, SISS successfully mitigated memorization on nearly 90% of the prompts we tested. We release our code online.", "keywords": [ "diffusion models", "selective forgetting", "unlearning", "data deletion", "generative models", "memorization" ], "rating_list": [ 3, 8, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "SuH5SdOXpe", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Hamid Krim", "gender": "Male", "institution": "North Carolina State University", "country": "", "position": "Full Professor" }, { "name": "MohamadAli Torkamani", "gender": "Male", "institution": "AWS AI", "country": "US", "position": "Researcher" }, { "name": "Xiaorui Liu", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Assistant Professor" }, { "name": "Zhichao Hou", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 13, "n_ref_uni": 40, "n_ref": 75, "n_ref_all": 101, "n_fig": 8, "n_tab": 5, "L_tab": 2735, "n_element_tab": 15, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 683, "n_element_tab_1": 8, "formula_len_all": 2106, "formula_len_all_1": 1142, "len_all": 125830, "len_all_1": 62273, "len_abs": 1109, "len_title": 100, "len_sents": 34838, "len_sents_1": 26766, "n_sents": 269, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 52, "L_abs": 1112, "title": "Robustness Reprogramming for Representation Learning", "abs": "This work tackles an intriguing and fundamental open challenge in representation learning: Given a well-trained deep learning model, can it be reprogrammed to enhance its robustness against adversarial or noisy input perturbations without altering its parameters?\nTo explore this, we revisit the core feature transformation mechanism in representation learning and propose a novel non-linear robust pattern matching technique as a robust alternative. Furthermore, we introduce three model reprogramming paradigms to offer flexible control of robustness under different efficiency requirements. Comprehensive experiments and ablation studies across diverse learning models ranging from basic linear model and MLPs to shallow and modern deep ConvNets demonstrate the effectiveness \nof our approaches.\nThis work not only opens a promising and orthogonal direction for improving adversarial defenses in deep learning beyond existing methods but also provides new insights into designing more resilient AI systems with robust statistics. Our implementation is available at https://anonymous.4open.science/r/NRPM-322C/", "keywords": [ "Adversarial Robustness", "Robustness Reprogramming", "Robust Representation Learning" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "Sta35eEmfR", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Bingning Wang", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Han Huang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Haoyu Lu", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Liang Wang", "gender": "Male", "institution": "Institute of Automation\uff0c CAS\uff0cChina", "country": "CN", "position": "Full Professor" }, { "name": "Qiang Liu", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Shu Wu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yuqi Huo", "gender": "Male", "institution": "Baichuan Intelligent Technology", "country": "CN", "position": "Researcher" }, { "name": "weipeng chen", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Joe Z", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 64, "n_ref_all": 75, "n_fig": 9, "n_tab": 9, "L_tab": 3086, "n_element_tab": 368, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 2822, "n_element_tab_1": 293, "formula_len_all": 340, "formula_len_all_1": 172, "len_all": 166459, "len_all_1": 63223, "len_abs": 1491, "len_title": 127, "len_sents": 43206, "len_sents_1": 27769, "n_sents": 296, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1333, "title": "Beyond Filtering: Adaptive Image-Text Quality Enhancement for MLLM Pretraining", "abs": "Multimodal large language models (MLLMs) have made significant strides by integrating visual and textual modalities. A critical factor in training MLLMs is the quality of image-text pairs within multimodal pretraining datasets. \nHowever, $\\textit{de facto}$ filter-based data quality enhancement paradigms often discard a substantial portion of high-quality image data due to inadequate semantic alignment between images and texts, leading to inefficiencies in data utilization and scalability. \nIn this paper, we propose the Adaptive Image-Text Quality Enhancer (AITQE), a model that dynamically assesses and enhances the quality of image-text pairs. AITQE employs a text rewriting mechanism for low-quality pairs and incorporates a negative sample learning strategy to improve evaluative capabilities by integrating deliberately selected low-quality samples during training. Unlike prior approaches that significantly alter text distributions, our method minimally adjusts text to preserve data volume while enhancing quality. Experimental results demonstrate that AITQE surpasses existing methods on various benchmark, effectively leveraging raw data and scaling efficiently with increasing data volumes.\nWe hope our work will inspire future works. The code and the model are available at https://github.com/iclr2025-No9242/AITQE.", "keywords": [ "mllm pretraining data; image-text data; data quality" ], "rating_list": [ 3, 6, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 4, 2 ], "presentation_list": [ 2, 3, 3, 4, 1 ], "contribution_list": [ 2, 3, 3, 3, 1 ], "confidence_list": [ 4, 2, 3, 5, 4 ] }, { "paper_id": "St7k6NJKn1", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Arun Narayanan", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Athula Balachandran", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Hongbin Liu", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Lun Wang", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Pedro J Moreno Mengibar", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Youzheng Chen", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 58, "n_ref_all": 72, "n_fig": 6, "n_tab": 5, "L_tab": 1081, "n_element_tab": 53, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 3424, "n_element_tab_1": 58, "formula_len_all": 5, "formula_len_all_1": 0, "len_all": 95571, "len_all_1": 42520, "len_abs": 1016, "len_title": 89, "len_sents": 18832, "len_sents_1": 14936, "n_sents": 135, "n_sents_1": 111, "venue": "ICLR2025", "decision": "withdraw", "L_title": 41, "L_abs": 1022, "title": "CAN DEEPFAKE SPEECH BE RELIABLY DETECTED?", "abs": "Recent advances in text-to-speech (TTS) systems, particularly those with voice cloning capabilities, have made voice impersonation readily accessible, raising ethical and legal concerns due to potential misuse for malicious activities like misinformation campaigns and fraud. While synthetic speech detectors (SSDs) exist to combat this, they are vulnerable to ``test domain shift\", exhibiting decreased performance when audio is altered through transcoding, playback, or background noise. This vulnerability is further exacerbated by deliberate manipulation of synthetic speech aimed at deceiving detectors. This work presents the first systematic study of such active malicious attacks against state-of-the-art open-source SSDs. White-box attacks, black-box attacks, and their transferability are studied from both attack effectiveness and stealthiness, using both hardcoded metrics and human ratings. The results highlight the urgent need for more robust detection methods in the face of evolving adversarial threats.", "keywords": [ "Deepfake audio detection", "TTS", "Adversarial Example" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "SsWMJ42hJO", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Daniel L. Pimentel-Alarc\u00f3n", "gender": "unknown", "institution": "University of Wisconsin, Madison", "country": "", "position": "Assistant Professor" }, { "name": "Huanran Li", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "PhD student" }, { "name": "Manh Nguyen", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Undergrad student" } ], "n_formula": 30, "n_formula_1": 6, "n_ref_uni": 24, "n_ref": 49, "n_ref_all": 79, "n_fig": 7, "n_tab": 2, "L_tab": 3854, "n_element_tab": 42, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 874, "n_element_tab_1": 10, "formula_len_all": 2271, "formula_len_all_1": 293, "len_all": 146082, "len_all_1": 67332, "len_abs": 1097, "len_title": 126, "len_sents": 38378, "len_sents_1": 34269, "n_sents": 284, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1101, "title": "Preventing Collapse in Contrastive Learning with Orthonormal Prototypes (CLOP)", "abs": "Contrastive learning has emerged as a powerful method in deep learning, excelling at learning effective representations through contrasting samples from different distributions. However, dimensional collapse, where embeddings converge into a lower-dimensional space, poses a significant challenge, especially in semi-supervised and self-supervised setups. In this paper, we first theoretically analyze the effect of large learning rates on contrastive losses that solely rely on the cosine similarity metric, and derive a theoretical bound to mitigate this collapse. Building on these insights, we propose CLOP, a novel semi-supervised loss function designed to prevent dimensional collapse by promoting the formation of orthogonal linear subspaces among class embeddings. Unlike prior approaches that enforce a simplex ETF structure, CLOP focuses on subspace separation, leading to more distinguishable embeddings. Through extensive experiments on real and synthetic datasets, we demonstrate that CLOP enhances performance, providing greater stability across different learning rates and batch sizes.", "keywords": [ "Deep Learning", "Contrastive Learning", "Neural Collapse", "Image Classification" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "SrnTGdJKYG", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Andr\u00e9 Hottung", "gender": "unknown", "institution": "Bielefeld University", "country": "DE", "position": "Postdoc" }, { "name": "Kevin Tierney", "gender": "Male", "institution": "Bielefeld University", "country": "DE", "position": "Full Professor" }, { "name": "Paula Wong-Chung", "gender": "Female", "institution": "University of British Columbia", "country": "CA", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 36, "n_ref": 68, "n_ref_all": 80, "n_fig": 5, "n_tab": 3, "L_tab": 1473, "n_element_tab": 149, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 705, "n_element_tab_1": 52, "formula_len_all": 633, "formula_len_all_1": 538, "len_all": 124548, "len_all_1": 65492, "len_abs": 728, "len_title": 105, "len_sents": 34877, "len_sents_1": 30149, "n_sents": 295, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 729, "title": "Neural Deconstruction Search for Vehicle Routing Problems", "abs": "Autoregressive construction approaches generate solutions to vehicle routing problems in a step-by-step fashion, leading to high-quality solutions that are nearing the performance achieved by handcrafted, operations research techniques.\nIn this work, we challenge the conventional paradigm of sequential solution construction and introduce an iterative search framework where solutions are instead deconstructed by a neural policy. Throughout the search, the neural policy collaborates with a simple greedy insertion algorithm to rebuild the deconstructed solutions. Our approach surpasses the performance of state-of-the-art operations research methods across three challenging vehicle routing problems of various problem sizes.", "keywords": [ "neural combinatorial optimization; vehicle routing problem; reinforcement learning; neural deconstruction" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "SrkDVzygXx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Chenglong Liu", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "En Yu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Haoran Wei", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Jianjian Sun", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Jinze Yang", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Kangheng Lin", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Liang Zhao", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Xiangyu Zhang", "gender": "Male", "institution": "MEGVII Technology", "country": "CN", "position": "Principal Researcher" }, { "name": "Zheng Ge", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Zining Zhu", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 60, "n_ref": 116, "n_ref_all": 142, "n_fig": 10, "n_tab": 6, "L_tab": 2251, "n_element_tab": 257, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 2073, "n_element_tab_1": 201, "formula_len_all": 618, "formula_len_all_1": 617, "len_all": 180147, "len_all_1": 60473, "len_abs": 925, "len_title": 118, "len_sents": 38253, "len_sents_1": 26916, "n_sents": 296, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 928, "title": "PerPO: Perceptual Preference Optimization via Discriminative Rewarding", "abs": "This paper presents Perceptual Preference Optimization (PerPO), a perception alignment method aimed at addressing the visual discrimination challenges in generative pre-trained multimodal large language models (MLLMs). PerPO employs discriminative rewarding and listwise preference optimization to align MLLMs with human visual perception processes. By utilizing the reward as a quantitative margin for ranking, our method effectively bridges generative preference optimization and discriminative empirical risk minimization. PerPO significantly enhances MLLMs\u2019 visual discrimination capabilities while maintaining their generative strengths, mitigates image-unconditional reward hacking, and ensures consistent performance across visual tasks. This work marks a crucial step towards more perceptually aligned and versatile MLLMs. We also anticipate that PerPO will inspire the community to reconsider MLLM alignment strategies.", "keywords": [ "MLLMs; RLHF; DPO" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 1 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "Srfi0a7vB3", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Balint Szekely", "gender": "unknown", "institution": "Imperial College London", "country": "", "position": "Undergrad student" }, { "name": "Greg Chadwick", "gender": "unknown", "institution": "lowRISC", "country": "", "position": "Digital Design Lead" }, { "name": "Hugo McNally", "gender": "Male", "institution": "lowRISC", "country": "", "position": "HW/SW Engineer" }, { "name": "Jianyi Cheng", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Lecturer" }, { "name": "Pedro Gimenes", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Yiren Zhao", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Assistant Professor" }, { "name": "Zixi Zhang", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "Undergrad student" }, { "name": "Robert Mullins", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 27, "n_ref": 72, "n_ref_all": 104, "n_fig": 6, "n_tab": 7, "L_tab": 3922, "n_element_tab": 494, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1912, "n_element_tab_1": 97, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 188565, "len_all_1": 64840, "len_abs": 1094, "len_title": 120, "len_sents": 59520, "len_sents_1": 29962, "n_sents": 444, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1032, "title": "LLM4DV: Using Large Language Models for Hardware Test Stimuli Generation", "abs": "Hardware design verification (DV) is a process that checks the functional equivalence of a hardware design against its specifications, improving hardware reliability and robustness. A key task in the DV process is the test stimuli generation, which creates a set of conditions or inputs for testing. A major challenge is that existing approaches to test stimuli generation require human effort due to the complexity and specificity of the test conditions required for an arbitrary hardware design. We seek an efficient and automated solution that takes advantage of large language models (LLMs). LLMs have shown promising results for improving hardware design automation, but remain under-explored for hardware DV. In this paper, we propose an open-source benchmarking framework called LLM4DV that efficiently orchestrates LLMs for automated hardware test stimuli generation. Our analysis evaluates five LLMs using six prompting improvements over eight hardware designs and provides insight for future work on LLMs for efficient DV.", "keywords": [ "Large Language Models", "Automated Chip Design", "Design Verification" ], "rating_list": [ 3, 5, 8, 3, 3 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 1, 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "SrGP0RQbYH", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Ashia Camage Wilson", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Joao Cavalcanti", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Laurent Lessard", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Associate Professor" } ], "n_formula": 160, "n_formula_1": 25, "n_ref_uni": 39, "n_ref": 91, "n_ref_all": 142, "n_fig": 13, "n_tab": 6, "L_tab": 874, "n_element_tab": 80, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 572, "n_element_tab_1": 72, "formula_len_all": 14319, "formula_len_all_1": 1451, "len_all": 221822, "len_all_1": 61650, "len_abs": 994, "len_title": 91, "len_sents": 70962, "len_sents_1": 27006, "n_sents": 660, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 973, "title": "Adaptive backtracking for fast optimization", "abs": "Backtracking line search is foundational in numerical optimization. \nThe basic idea is to adjust the step size of an algorithm by a {\\em constant} factor until some chosen criterion (e.g. Armijo, Goldstein, Descent Lemma) is satisfied. \nWe propose a new way for adjusting step sizes, replacing the constant factor used in regular backtracking with one that takes into account the degree to which the chosen criterion is violated, without additional computational burden. \nWe perform a variety of experiments on over fifteen real world datasets, which confirm that adaptive backtracking often leads to significantly faster optimization.\nFor convex problems, we prove adaptive backtracking requires fewer adjustments to produce a feasible step size than regular backtracking does for two popular line search criteria: the Armijo condition and the descent lemma.\nFor nonconvex smooth problems, we prove adaptive backtracking enjoys the same guarantees of regular backtracking.", "keywords": [ "Optimization", "backtracking line search", "Armijo condition", "descent lemma", "adaptive optimization methods" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 4, 4, 3, 2 ], "presentation_list": [ 4, 4, 2, 2 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 3, 5, 3, 3 ] }, { "paper_id": "SrGP0ILoYa", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Astrit Tola", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Baris Coskunuzer", "gender": "Male", "institution": "University of Texas, Dallas", "country": "US", "position": "Full Professor" }, { "name": "Cuneyt Gurcan Akcora", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Associate Professor" }, { "name": "Funmilola Mary Taiwo", "gender": "Female", "institution": "University of Manitoba", "country": "", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 10, "n_ref_uni": 69, "n_ref": 87, "n_ref_all": 120, "n_fig": 9, "n_tab": 12, "L_tab": 5792, "n_element_tab": 487, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1251, "n_element_tab_1": 115, "formula_len_all": 1345, "formula_len_all_1": 785, "len_all": 175168, "len_all_1": 67973, "len_abs": 964, "len_title": 62, "len_sents": 46336, "len_sents_1": 30366, "n_sents": 419, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1066, "title": "TopER: Topological Embeddings in Graph Representation Learning", "abs": "Graph embeddings serve as the cornerstone for graph representation learning, facilitating the exploration of graphs by machine learning methods. However, prevalent deep learning techniques rely on black-box, high-dimensional graph embeddings. There is a pressing need for an interpretable, low-dimensional embedding approach to empower efficient graph visualization and provide practical tools to study graph datasets effectively.\n\nIn this paper, we present a novel low-dimensional graph embedding method called *Topological Evolution Rate (TopER)*, which simplifies a key concept of topological data analysis known as *filtration*. TopER calculates the evolution rate of graph substructures induced by a filtration function on nodes or edges, resulting in interpretable 2D visualizations of graph datasets. Our experiments demonstrate that this new embedding method achieves highly competitive performance compared to the latest deep learning models in graph classification tasks on benchmark datasets. We further provide theoretical stability guarantees for TopER.", "keywords": [ "Graph embeddings", "graph classification", "graph representation learning", "interpretability", "data visualization", "topological data analysis" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "Sr5XaZzirA", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jaeho Lee", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Sangyoon Lee", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Taesun Yeom", "gender": "unknown", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 59, "n_formula_1": 9, "n_ref_uni": 62, "n_ref": 135, "n_ref_all": 196, "n_fig": 24, "n_tab": 5, "L_tab": 1632, "n_element_tab": 222, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3754, "formula_len_all_1": 594, "len_all": 235607, "len_all_1": 71425, "len_abs": 1153, "len_title": 116, "len_sents": 74085, "len_sents_1": 33515, "n_sents": 639, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1116, "title": "Fast Training of Sinusoidal Neural Fields via Scaling Initialization", "abs": "Neural fields are an emerging paradigm that represent data as continuous functions parameterized by neural networks. Despite many advantages, neural fields often have a high training cost, which prevents a broader adoption. In this paper, we focus on a popular family of neural fields, called sinusoidal neural fields (SNFs), and study how it should be initialized to maximize the training speed. We find that the standard initialization scheme for SNFs---designed based on the signal propagation principle---is suboptimal. In particular, we show that by simply multiplying each weight (except for the last layer) by a constant, we can accelerate SNF training by 10$\\times$. This method, coined _weight scaling_, consistently provides a significant speedup over various data domains, allowing the SNFs to train faster than more recently proposed architectures. To understand why the weight scaling works well, we conduct extensive theoretical and empirical analyses which reveal that the weight scaling not only resolves the spectral bias quite effectively but also enjoys a well-conditioned optimization trajectory.", "keywords": [ "sinusoidal neural fields", "fast training", "initializations" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 2, 2, 4 ] }, { "paper_id": "SqoL14HDm0", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 7, "author_info_list": [ { "name": "Chenhua Liu", "gender": "Male", "institution": "Jiangxi Normal University", "country": "CN", "position": "Undergrad student" }, { "name": "Daniel SUN", "gender": "unknown", "institution": "Newcastle University, UK", "country": "GB", "position": "Associate Professor" }, { "name": "Dehai Zhao", "gender": "Male", "institution": "Commonwealth Scientific and Industrial Research Organisation, CSIRO", "country": "AU", "position": "Researcher" }, { "name": "Qing Huang", "gender": "Male", "institution": "Jiangxi Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Yang Liu", "gender": "Male", "institution": "Jiangxi Normal University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhenchang Xing", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "Associate Professor" }, { "name": "Zhuo Cheng", "gender": "Male", "institution": "Jiangxi Normal University", "country": "CN", "position": "Lecturer" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 9, "n_ref": 17, "n_ref_all": 48, "n_fig": 19, "n_tab": 3, "L_tab": 1163, "n_element_tab": 36, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 385, "n_element_tab_1": 37, "formula_len_all": 29, "formula_len_all_1": 0, "len_all": 188822, "len_all_1": 63211, "len_abs": 9217, "len_title": 159, "len_sents": 84136, "len_sents_1": 32374, "n_sents": 591, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 112, "L_abs": 1864, "title": "When Prompt Engineering Meets Software Engineering: CNL-P as Natural and Robust \"APIs'' for Human-AI Interaction", "abs": "As large language models (LLMs) gain increasing capabilities, they are being widely applied in areas such as intelligent customer service, code generation, and knowledge management. Prompts written in natural language (NL) serve as the \"APIs'' for human-LLM interaction. To enhance prompt quality, best practices for prompt engineering (PE) have been established, along with writing guidelines and templates. However, due to the inherent ambiguity of natural language, even prompts that strictly follow these guidelines often fail to trigger LLM to consistently output high quality responses, particularly for complex tasks. To address this issue, this paper proposes a Controlled Natural Language for Prompt (CNL-P) which incorporates best practices in PE. To overcome the NL's ambiguity, CNL-P introduces precise grammar structures and strict semantic norms, enabling a declarative but structured and accurate representation of user intent. This helps LLMs better understand and execute CNL-P, leading to higher quality responses. To lower the learning curve of CNL-P, we introduce an automatic NL2CNL-P conversion agent based on LLMs, which allow users to describe prompts in NL from which the NL2CNL-P agent generates CNL-P compliant prompts guided by CNL-P grammar. We further develop a linting tool for CNL-P, including syntactic and semantic checks, making static analysis techniques applicable to natural language for the first time. CNL-P\u2019s design not only integrates best practices in PE but also adopts key principles from software engineering (SE). Extensive experiments show that CNL-P can improve the quality of LLM's responses through the novel and organic synergy of PE and SE. We envision that CNL-P has the potential to bridge the gap between emergent PE and traditional SE, paving the way for a new natural language centric programming paradigm.", "keywords": [ "large language model", "human-LLM interaction", "API", "prompt engineering", "software engineering", "controlled natural language" ], "rating_list": [ 5, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 4, 2 ], "confidence_list": [ 3, 4, 2 ] }, { "paper_id": "Sqf4jqKrQy", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Anjan Dutta", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Assistant Professor" }, { "name": "Sauradip Nag", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "Postdoc" }, { "name": "Serge Belongie", "gender": "Male", "institution": "University of Copenhagen", "country": "DK", "position": "Full Professor" }, { "name": "Silpa Vadakkeeveetil Sreelatha", "gender": "Female", "institution": "University of Surrey", "country": "GB", "position": "PhD student" }, { "name": "Muhammad Awais Tanvir Rana", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 25, "n_ref": 75, "n_ref_all": 107, "n_fig": 9, "n_tab": 8, "L_tab": 1703, "n_element_tab": 199, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 921, "n_element_tab_1": 177, "formula_len_all": 806, "formula_len_all_1": 778, "len_all": 128649, "len_all_1": 68080, "len_abs": 1437, "len_title": 121, "len_sents": 46541, "len_sents_1": 31992, "n_sents": 328, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1456, "title": "Concept Denoising Score Matching for Responsible Text-to-Image Generation", "abs": "Diffusion models excel at generating diverse, high-quality images, but they also risk producing unfair and harmful content. Existing methods that update text embeddings or model weights either fail to address biases within diffusion models or are computationally expensive. We tackle responsible (fair and safe) text-to-image (T2I) generation in diffusion models as an interpretable concept discovery problem, introducing Concept Denoising Score Matching (CoDSMa) -- a novel objective that learns responsible concept representations in the bottleneck feature activation (\\textit{h-space}). Our approach builds on the observation that, at any timestep, aligning the neutral prompt with the target prompt directs the predicted score of denoised latent towards the target concept. We empirically demonstrate that our method enables responsible T2I generation by addressing two key challenges: mitigating gender and racial biases (fairness) and eliminating harmful content (safety). Our approach reduces biased and harmful generation by nearly 50\\% compared to state-of-the-art methods. Remarkably, it outperforms other techniques in debiasing gender and racial attributes without requiring profession-specific data. Furthermore, it successfully filters inappropriate content, such as depictions of illegal activities or harassment, without training on such data. Additionally, our method effectively handles intersectional biases without any further training.", "keywords": [ "Diffusion models", "Stable Diffusion", "Responsible text-to-image generation", "Fairness", "Safe generation", "Debiasing" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 1 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "SqcoXJc4mC", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Alessandro Abate", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Konstantinos Gatsis", "gender": "unknown", "institution": "University of Southampton", "country": "", "position": "Lecturer" }, { "name": "Ondrej Bajgar", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Michael Osborne", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 13, "n_ref_uni": 49, "n_ref": 80, "n_ref_all": 105, "n_fig": 8, "n_tab": 3, "L_tab": 1506, "n_element_tab": 174, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 876, "n_element_tab_1": 133, "formula_len_all": 1385, "formula_len_all_1": 962, "len_all": 174337, "len_all_1": 68299, "len_abs": 949, "len_title": 98, "len_sents": 58817, "len_sents_1": 33799, "n_sents": 368, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 952, "title": "Q-based Variational Inverse Reinforcement Learning", "abs": "The development of safe and beneficial AI requires that systems can learn and act in accordance with human preferences. However, explicitly specifying these preferences by hand is often infeasible. Inverse reinforcement learning (IRL) addresses this challenge by inferring preferences, represented as reward functions, from expert behavior. We introduce Q-based Variational IRL (QVIRL), a novel Bayesian IRL method that recovers a posterior distribution over rewards from expert demonstrations via primarily learning a variational distribution over Q-values. Unlike previous approaches, QVIRL combines scalability with uncertainty quantification, important for safety-critical applications. We demonstrate QVIRL's strong performance in apprenticeship learning across various tasks, including classical control problems and safe navigation in the Safety Gymnasium suite, where the method's uncertainty quantification allows us to produce safer policies.", "keywords": [ "inverse reinforcement learning", "IRL", "imitation learning" ], "rating_list": [ 5, 8, 5, 3 ], "soundness_list": [ 2, 4, 3, 1 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "SqZ0KY4qBD", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Adway Girish", "gender": "Male", "institution": "EPFL", "country": "CH", "position": "PhD student" }, { "name": "Alliot Nagle", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" }, { "name": "Ashok Vardhan Makkuva", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Postdoc" }, { "name": "Hyeji Kim", "gender": "Female", "institution": "University of Texas, Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Marco Bondaschi", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Martin Jaggi", "gender": "Male", "institution": "EPFL", "country": "CH", "position": "Associate Professor" }, { "name": "Michael Gastpar", "gender": "unknown", "institution": "School of Computer and Communication Sciences, EPFL - EPF Lausanne", "country": "", "position": "Full Professor" } ], "n_formula": 77, "n_formula_1": 17, "n_ref_uni": 28, "n_ref": 39, "n_ref_all": 88, "n_fig": 10, "n_tab": 1, "L_tab": 155, "n_element_tab": 36, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9930, "formula_len_all_1": 1988, "len_all": 176124, "len_all_1": 64797, "len_abs": 1669, "len_title": 114, "len_sents": 49824, "len_sents_1": 27920, "n_sents": 502, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 66, "L_abs": 1683, "title": "Attention with Markov: A Curious Case of Single-layer Transformers", "abs": "Attention-based transformers have achieved tremendous success across a variety of disciplines including natural languages. To deepen our understanding of their sequential modeling capabilities, there is a growing interest in using Markov input processes to study them. A key finding is that when trained on first-order Markov chains, transformers with two or more layers consistently develop an induction head mechanism to estimate the in-context bigram conditional distribution. In contrast, single-layer transformers, unable to form an induction head, directly learn the Markov kernel but often face a surprising challenge: they become trapped in local minima representing the unigram distribution, whereas deeper models reliably converge to the ground-truth bigram. While single-layer transformers can theoretically model first-order Markov chains, their empirical failure to learn this simple kernel in practice remains a curious phenomenon. To explain this contrasting behavior of single-layer models, in this paper we introduce a new framework for a principled analysis of transformers via Markov chains. Leveraging our framework, we theoretically characterize the loss landscape of single-layer transformers and show the existence of global minima (bigram) and bad local minima (unigram) contingent on data properties and model architecture. We precisely delineate the regimes under which these local optima occur. Backed by experiments, we demonstrate that our theoretical findings are in congruence with the empirical results. Finally, we outline several open problems in this arena. Code is available at \\url{https://anonymous.4open.science/r/Attention-with-Markov-A617/}.", "keywords": [ "Markov chains", "Transformers", "Optimization", "Landscape" ], "rating_list": [ 6, 5, 10, 8 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 2, 2, 4, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "SqUiGfJ1So", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Bocheng Li", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Linli Xu", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yifei Xin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yongxin Zhu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 23, "n_ref_uni": 32, "n_ref": 59, "n_ref_all": 74, "n_fig": 9, "n_tab": 4, "L_tab": 1892, "n_element_tab": 252, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 802, "n_element_tab_1": 108, "formula_len_all": 1058, "formula_len_all_1": 1034, "len_all": 124419, "len_all_1": 63683, "len_abs": 2151, "len_title": 132, "len_sents": 30700, "len_sents_1": 28010, "n_sents": 233, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1863, "title": "Addressing Representation Collapse in Vector Quantized Models with One Linear Layer", "abs": "Vector Quantization (VQ) is a widely used method for converting continuous representations into discrete codes, which has become fundamental in unsupervised representation learning and latent generative models. However, VQ models are often hindered by the problem of representation collapse in the latent space, which leads to low codebook utilization and limits the scalability of the codebook for large-scale training. Existing methods designed to mitigate representation collapse typically reduce the dimensionality of latent space at the expense of model capacity, which do not fully resolve the core issue. In this study, we conduct a theoretical analysis of representation collapse in VQ models and identify its primary cause as the disjoint optimization of the codebook, where only a small subset of code vectors are updated through gradient descent. To address this issue, we propose \\textbf{SimVQ}, a novel method which reparameterizes the code vectors through a linear transformation layer based on a learnable latent basis. This transformation optimizes the \\textit{entire linear space} spanned by the codebook, rather than merely updating \\textit{the code vector} selected by the nearest-neighbor search in vanilla VQ models. Although it is commonly understood that the multiplication of two linear matrices is equivalent to applying a single linear layer, our approach works surprisingly well in resolving the collapse issue in VQ models with just one linear layer. We validate the efficacy of SimVQ through extensive experiments across various modalities, including image and audio data with different model architectures. The results show that SimVQ not only effectively addresses the problem of representation collapse but also proves highly adaptable and easy to implement, suggesting its broad applicability in diverse machine learning contexts.", "keywords": [ "vector quantization", "representation collapse" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "SpXd4dA5Ty", "primary_area": "generative models", "n_author": 14, "author_info_list": [ { "name": "Bo Long", "gender": "Male", "institution": "Meta", "country": "US", "position": "Principal Researcher" }, { "name": "Fabian Paischer", "gender": "Male", "institution": ", Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Postdoc" }, { "name": "Jiacheng Li", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Kaveh Hassani", "gender": "Male", "institution": "Meta", "country": "CA", "position": "AI Research Scientist" }, { "name": "Liu Yang", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Nima Noorshams", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Principal Researcher" }, { "name": "Robert D Nowak", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "", "position": "Full Professor" }, { "name": "Sem Park", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Shuai Shao", "gender": "Male", "institution": "Meta Research", "country": "", "position": "Researcher" }, { "name": "Xiaoli Gao", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Meta" }, { "name": "Xue Feng", "gender": "unknown", "institution": "Facebook", "country": "", "position": "research scientist" }, { "name": "Yun He", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Research Scientist " }, { "name": "Zhang Gabriel Li", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Hamid Eghbal-zadeh", "gender": "Male", "institution": "Meta", "country": "US", "position": "Research Scientist " } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 29, "n_ref": 108, "n_ref_all": 136, "n_fig": 10, "n_tab": 3, "L_tab": 2015, "n_element_tab": 148, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 408, "n_element_tab_1": 25, "formula_len_all": 876, "formula_len_all_1": 686, "len_all": 169749, "len_all_1": 67541, "len_abs": 1651, "len_title": 117, "len_sents": 51547, "len_sents_1": 31634, "n_sents": 339, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1542, "title": "Unifying Generative and Dense Retrieval for Sequential Recommendation", "abs": "Sequential dense retrieval models utilize advanced sequence learning techniques to compute item and user representations, which are then used to rank relevant items for a user through inner product computation between the user and all item representations. However, this approach requires storing a unique representation for each item, resulting in significant memory requirements as the number of items\ngrow. In contrast, the recently proposed generative retrieval paradigm offers a promising alternative by directly predicting item indices using a generative model trained on semantic IDs that encapsulate items\u2019 semantic information. Despite its potential for large-scale applications, a comprehensive comparison between generative retrieval and sequential dense retrieval under fair conditions is still lacking, leaving open questions regarding performance, storage, and computation trade-offs. To address this, we compare these two approaches under controlled conditions on academic benchmarks and propose LIGER (**L**everag**I**ng dense retrieval for **GE**nerative **R**etrieval), a hybrid model that combines the strengths of these two widely used methods. LIGER integrates sequential dense retrieval into generative retrieval, mitigating performance differences and enhancing cold-start item recommendation in the datasets evaluated. This hybrid approach provides insights into the trade-offs between these approaches and demonstrates improvements in efficiency and effectiveness for recommendation systems in small-scale benchmarks.", "keywords": [ "Generative Retrieval", "Sequential Recommendation" ], "rating_list": [ 5, 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 3, 1, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 3, 5 ] }, { "paper_id": "SpTzsQjgxF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Chang Yue", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Hong Hu", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Mingye Gao", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Xiaomin Li", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Zhiwei Zhang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 10, "n_ref_uni": 55, "n_ref": 231, "n_ref_all": 261, "n_fig": 12, "n_tab": 19, "L_tab": 3636, "n_element_tab": 489, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 648, "formula_len_all_1": 426, "len_all": 257173, "len_all_1": 65541, "len_abs": 7227, "len_title": 100, "len_sents": 68011, "len_sents_1": 31668, "n_sents": 601, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1630, "title": "Rule-Based Rating and Selection of LLM Training Data", "abs": "The quality of training data is crucial for the performance of large language models (LLMs). There are recent studies utilizing LLMs to rate and select data based on scores from a small set of human-designed metrics (rules). However, existing rule-based methods often overly rely on human heuristics, lack robust metrics for rule evaluation, and exhibit limited adaptability to new tasks. In our work, we propose a novel rule-based framework that leverages the orthogonality of score vectors corresponding to rules as a unique metric for rule evaluation. Our method employs an automated pipeline that first uses LLMs to generate a diverse set of rules, covering a wide range of rating aspects. It then rates a batch of data according to these rules and applies the determinantal point process (DPP) from random matrix theory to select the most orthogonal score vectors, effectively isolating a subset of independent rules. Then these rules are applied to rate all data and samples with the highest average scores are selected for further downstream tasks such as LLM training. We validate our method through two experimental setups: 1) comparison against ground truth ratings and 2) benchmarking LLMs trained with the selected data. Our extensive experiments span various settings, including general pre-training and domain-specific fine-tuning in fields such as IMDB, Medical, Math, and Code. The results show that our DPP rule-based rating method consistently outperforms other methods, such as rating without rules, uniform sampling, importance resampling, and QuRating, in terms of both rating accuracy and model performance.", "keywords": [ "large language models", "data selection", "data rating" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "SoUwcVplq4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Gim Hee Lee", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" }, { "name": "Tianxin Huang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Research Fellow" }, { "name": "Yuyang Zhao", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zhiwen Yan", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 46, "n_ref": 137, "n_ref_all": 183, "n_fig": 20, "n_tab": 9, "L_tab": 1232, "n_element_tab": 78, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 88, "n_element_tab_1": 9, "formula_len_all": 571, "formula_len_all_1": 504, "len_all": 157227, "len_all_1": 60223, "len_abs": 884, "len_title": 107, "len_sents": 45260, "len_sents_1": 26833, "n_sents": 482, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 886, "title": "ComPC: Completing a 3D Point Cloud with 2D Diffusion Priors", "abs": "3D point clouds directly collected from objects through sensors are often incomplete due to self-occlusion. Conventional methods for completing these partial point clouds rely on manually organized training sets and are usually limited to object categories seen during training. In this work, we propose a test-time framework for completing partial point clouds across unseen categories without any requirement for training. Leveraging point rendering via Gaussian Splatting, we develop techniques of Partial Gaussian Initialization, Zero-shot Fractal Completion, and Point Cloud Extraction that utilize priors from pre-trained 2D diffusion models to infer missing regions and extract uniform completed point clouds. Experimental results on both synthetic and real-world scanned point clouds demonstrate that our approach outperforms existing methods in completing a variety of objects.", "keywords": [ "Gaussian Splatting", "Diffusion Model", "Point Cloud Completion" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "SnDmPkOJ0T", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Dongrui Liu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Jie Zhang", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Jing Shao", "gender": "Female", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Linfeng Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "chen qian", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "MS student" } ], "n_formula": 27, "n_formula_1": 3, "n_ref_uni": 51, "n_ref": 103, "n_ref_all": 141, "n_fig": 11, "n_tab": 5, "L_tab": 2286, "n_element_tab": 296, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1522, "n_element_tab_1": 220, "formula_len_all": 3515, "formula_len_all_1": 363, "len_all": 182078, "len_all_1": 64025, "len_abs": 932, "len_title": 116, "len_sents": 53397, "len_sents_1": 29640, "n_sents": 380, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 68, "L_abs": 935, "title": "REEF: Representation Encoding Fingerprints for Large Language Models", "abs": "Protecting the intellectual property of open-source Large Language Models (LLMs) is very important, because training LLMs costs extensive computational resources and data. Therefore, model owners and third parties need to identify whether a suspect model is a subsequent development of the victim model. To this end, we propose a training-free REEF to identify the relationship between the suspect and victim models from the perspective of LLMs' feature representations. Specifically, REEF computes and compares the centered kernel alignment similarity between the representations of a suspect model and a victim model on the same samples. This training-free REEF does not impair the model's general capabilities and is robust to sequential fine-tuning, pruning, model merging, and permutations. In this way, REEF provides a simple and effective way for third parties and models' owners to protect LLMs' intellectual property together.", "keywords": [ "Large Language Model", "Fingerprint", "Representation", "Intellectual Property" ], "rating_list": [ 8, 6, 8, 10 ], "soundness_list": [ 4, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 4, 2, 3, 4 ] }, { "paper_id": "SmxM4POTBk", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Gahyun Yoo", "gender": "unknown", "institution": "Seoul National University", "country": "", "position": "Undergrad student" }, { "name": "Seungwoo Choi", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Undergrad student" }, { "name": "Jay Yoon Lee", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 24, "n_ref": 36, "n_ref_all": 54, "n_fig": 6, "n_tab": 8, "L_tab": 3370, "n_element_tab": 479, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1879, "n_element_tab_1": 311, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 193502, "len_all_1": 61380, "len_abs": 1406, "len_title": 155, "len_sents": 42867, "len_sents_1": 26936, "n_sents": 345, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1414, "title": "IntGrad MT: Enhancing LLMs' Machine Translation Capabilities with Sentence Interpolation Guided Gradual MT", "abs": "Recent Large Language Models (LLMs) have demonstrated strong performance in translation without needing to be finetuned on additional parallel corpora. However, they still underperform for low-resource language pairs. Previous works have focused on mitigating this issue by leveraging relevant few-shot examples or external resources such as dictionaries or grammar books, making models heavily reliant on these nonparametric sources of information. In this paper, we propose a novel method named IntGrad MT that focuses on fully exploiting an LLM\u2019s inherent translation capability. IntGrad MT achieves this by constructing a chain of few-shot examples, each consisting of a source sentence and the model\u2019s own translation, that rise incrementally in difficulty. IntGrad MT employs two techniques: Sentence Interpolation, which generates a sequence of sentences that gradually change from an easy sentence to translate to a difficult one, and Gradual MT, which sequentially translates this chain using translations of earlier sentences as few-shot examples for the translation of subsequent ones. With this approach, we observe a substantial enhancement in the xCOMET scores of various LLMs for multiple languages, especially in low-resource languages such as Hindi(8.26), Swahili(7.10), Bengali(6.97) and Marathi(13.03). Our approach presents a practical way of enhancing LLMs' performance without extra training.", "keywords": [ "Machine Translation", "Self-domonstration", "Sentence Interpolation", "Low Resource Language" ], "rating_list": [ 6, 3, 8, 6, 3 ], "soundness_list": [ 3, 1, 3, 2, 1 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 4, 2, 3, 2, 1 ], "confidence_list": [ 4, 5, 5, 4, 4 ] }, { "paper_id": "SmYDdeLAR5", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Erica Zhang", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Fangzhao Zhang", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Mert Pilanci", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" } ], "n_formula": 187, "n_formula_1": 28, "n_ref_uni": 46, "n_ref": 83, "n_ref_all": 125, "n_fig": 14, "n_tab": 10, "L_tab": 1564, "n_element_tab": 133, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 79, "n_element_tab_1": 9, "formula_len_all": 19398, "formula_len_all_1": 2684, "len_all": 310477, "len_all_1": 70707, "len_abs": 4955, "len_title": 120, "len_sents": 108009, "len_sents_1": 30484, "n_sents": 958, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 930, "title": "Active Learning of Deep Neural Networks via Gradient-Free Cutting Planes", "abs": "Active learning methods aim to improve sample complexity in machine learning. In this work, we investigate an active learning scheme via a novel gradient-free cutting-plane training method for ReLU networks of arbitrary depth. \nWe demonstrate, for the first time, that cutting-plane algorithms, traditionally used in linear models, can be extended to deep neural networks despite their nonconvexity and nonlinear decision boundaries. Our results demonstrate that these methods provide a promising alternative to the commonly employed gradient-based optimization techniques in large-scale neural networks. \nMoreover, this training method induces the first deep active learning scheme known to achieve convergence guarantees. We exemplify the effectiveness of our proposed active learning method against popular deep active learning baselines via both synthetic data experiments and sentimental classification task on real datasets.", "keywords": [ "Active Learning", "Neural Networks", "Convex Optimization", "Cutting Plane Methods" ], "rating_list": [ 3, 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 4, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 2, 3, 3 ] }, { "paper_id": "Slr3KojVRO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Guangliang Cheng", "gender": "Male", "institution": "University of Liverpool", "country": "GB", "position": "Associate Professor" }, { "name": "Jianan Ye", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "CN", "position": "PhD student" }, { "name": "Kaizhu Huang", "gender": "Male", "institution": "Duke Kunshan University", "country": "CN", "position": "Full Professor" }, { "name": "Weiguang Zhao", "gender": "Male", "institution": "University of Liverpool", "country": "GB", "position": "PhD student" }, { "name": "Xi Yang", "gender": "Female", "institution": "Xi'an Jiaotong-Liverpool University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 26, "n_ref": 55, "n_ref_all": 74, "n_fig": 14, "n_tab": 4, "L_tab": 4124, "n_element_tab": 930, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 2522, "n_element_tab_1": 494, "formula_len_all": 521, "formula_len_all_1": 539, "len_all": 107685, "len_all_1": 60320, "len_abs": 1455, "len_title": 127, "len_sents": 28314, "len_sents_1": 23565, "n_sents": 216, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1460, "title": "PO3AD: Predicting Point Offsets toward Better 3D Point Cloud Anomaly Detection", "abs": "Point cloud anomaly detection, particularly under the anomaly-free setting, poses a significant challenge as it requires the precise capture of 3D normal data features to accurately identify deviations indicative of anomalies. Current efforts focus on devising reconstruction tasks, such as acquiring normal data representations by restoring normal samples from altered, pseudo-anomalous counterparts. Nonetheless, such methods tend to dilute the model's focus, as they require attention to both normal and pseudo-anomalous data points, thereby hampering the efficacy of the learning process. Moreover, the inherently disordered and sparse nature of 3D point cloud data significantly complicates the task. In response to those predicaments, we introduce an innovative approach that involves learning point offsets for the first time, with a concentrated emphasis on more informative pseudo-abnormal points, thus fostering more effective distillation of normal data representations. We have crafted an augmentation technique that is steered by normal vectors, facilitating the creation of credible pseudo anomalies that enhance the efficiency of the training process. Our comprehensive experimental evaluation on the Anomaly-ShapeNet and Real3D-AD datasets evidences that our proposed method outperforms existing state-of-the-art approaches, achieving an average enhancement of 9.0% and 1.4% in the AUC-ROC detection metric across these datasets, respectively.", "keywords": [ "3D point cloud", "anomaly detection" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "Sl1kRAATbw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Chao Ma", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Fei Xie", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CX", "position": "PhD student" }, { "name": "Zhongdao Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 54, "n_ref": 135, "n_ref_all": 149, "n_fig": 7, "n_tab": 7, "L_tab": 3650, "n_element_tab": 458, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3391, "n_element_tab_1": 431, "formula_len_all": 650, "formula_len_all_1": 650, "len_all": 162254, "len_all_1": 69269, "len_abs": 1435, "len_title": 153, "len_sents": 43334, "len_sents_1": 28634, "n_sents": 336, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 104, "L_abs": 1444, "title": "Space-Correlated Transformer: Jointly Explore the Matching and Motion Clues in 3D Single Object Tracking", "abs": "3D Single Object Tracking (3D SOT) in LiDAR point clouds plays a crucial role in autonomous driving. Current approaches mostly follow two paradigms, i.e., Siamese matching-based and motion-centric. However, LiDAR point clouds lack enough appearance information, while the motion-centric trackers suffer from complex model structures. To address these issues, we present a novel and conceptually simple tracking framework dubbed SCtrack, which jointly explores the matching and motion clues in point clouds. Specifically, SCtrack embeds point clouds into spatially structured features and conducts space correlation along the aligned spatial region. The target relative motion is directly inferred from the correlated features. In contrast to prevalent PointNet-based features, our spatially structured representation inherently models motion clues among the consecutive frames of point clouds, thereby being complementary to appearance matching. To better utilize the aligned structured features, we employ a strategy of varied-size space regions that adapt to different target shapes and locations during space correlation. Without bells and whistles, SCtrack achieves leading performance, with 89.1%, 71.5%, and 62.7% precision on KITTI, NuScenes, and Waymo Open Dataset, and runs at a considerably high speed of 60 Fps on a single RTX3090 GPU. Extensive studies validate the effectiveness of our SCtrack framework. The code will be released.", "keywords": [ "3D Single Object Tracking; Point Cloud; Transformer; Space-Correlation" ], "rating_list": [ 5, 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 5 ] }, { "paper_id": "SkpY8Skqnv", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Benjamin Th\u00e9rien", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "Boris Knyazev", "gender": "Male", "institution": "Samsung", "country": "CA", "position": "Researcher" }, { "name": "Charles-\u00c9tienne Joseph", "gender": "Male", "institution": "CapitalOne", "country": "US", "position": "Researcher" }, { "name": "Edouard Oyallon", "gender": "Male", "institution": "CNRS", "country": "FR", "position": "Researcher" }, { "name": "Eugene Belilovsky", "gender": "Male", "institution": "Concordia University, Montreal", "country": "CA", "position": "Assistant Professor" }, { "name": "Irina Rish", "gender": "Female", "institution": "University of Montreal", "country": "CA", "position": "Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 42, "n_ref": 98, "n_ref_all": 135, "n_fig": 7, "n_tab": 7, "L_tab": 1786, "n_element_tab": 86, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 977, "n_element_tab_1": 43, "formula_len_all": 382, "formula_len_all_1": 342, "len_all": 171978, "len_all_1": 63720, "len_abs": 1247, "len_title": 112, "len_sents": 54019, "len_sents_1": 30900, "n_sents": 401, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1285, "title": "$\\mu$LO: Compute-Efficient Meta-Generalization of Learned Optimizers", "abs": "Learned optimizers (LOs) can significantly reduce the wall-clock training time of neural networks, substantially reducing training costs. However, they can struggle to optimize unseen tasks (meta-generalize), especially when training networks much larger than those seen during meta-training. To address this, we derive the Maximal Update Parametrization ($\\mu$P) for two popular learned optimizer architectures and propose a simple meta-training recipe for $\\mu$-parameterized LOs ($\\mu$LOs). Our empirical evaluation demonstrates that LOs meta-trained with our recipe substantially improve meta-generalization to wider unseen tasks when compared to LOs trained under standard parametrization (e.g., as they are trained in existing work). When applying our $\\mu$LOs, each trained for less than 250 GPU-hours, to large-width models we are often able to match or exceed the performance of pre-trained VeLO, the most performant publicly available learned optimizer, meta-trained with 4000 TPU-months of compute. We also empirically observe that learned optimizers trained with our $\\mu$LO recipe also exhibit substantially improved meta-generalization to deeper networks ($5\\times$ meta-training) and remarkable generalization to much longer training horizons ($25\\times$ meta-training).", "keywords": [ "Meta Learning", "Learned Optimizers", "Pre-training" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "SkDNQbMQba", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Ge Liu", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Guanyu Lin", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Pengrui Han", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Tao Feng", "gender": "Male", "institution": "Department of Computer Science, University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Jiaxuan You", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 27, "n_ref": 42, "n_ref_all": 90, "n_fig": 16, "n_tab": 7, "L_tab": 2568, "n_element_tab": 234, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 425, "n_element_tab_1": 42, "formula_len_all": 311, "formula_len_all_1": 341, "len_all": 176061, "len_all_1": 63657, "len_abs": 6462, "len_title": 114, "len_sents": 63805, "len_sents_1": 31063, "n_sents": 484, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1754, "title": "Thought-Retriever: Don\u2019t Just Retrieve Raw Data, Retrieve Thoughts", "abs": "Large language models (LLMs) have transformed AI research thanks to their powerful \\textit{internal} capabilities and knowledge. However, existing LLMs still fail to effectively incorporate the massive \\textit{external} knowledge when interacting with the world. Although retrieval-augmented LLMs are proposed to mitigate the issue, they are still fundamentally constrained by the context length of LLMs, as they can only retrieve top-K raw data chunks from the external knowledge base which often consists of millions of data chunks. Here we propose Thought-Retriever, a novel model-agnostic algorithm that helps LLMs generate output conditioned on arbitrarily long external data, without being constrained by the context length or number of retrieved data chunks. Our key insight is to let an LLM fully leverage its intermediate responses generated when solving past user queries (thoughts), filtering meaningless and redundant thoughts, organizing them in thought memory, and retrieving the relevant thoughts when addressing new queries. Besides algorithmic innovation, we further meticulously prepare a novel benchmark, AcademicEval, which requires an LLM to faithfully leverage ultra-long context to answer queries based on real-world academic papers. Extensive experiments on AcademicEval and two other public datasets validate that Thought-Retriever remarkably outperforms state-of-the-art baselines, achieving an average increase of at least 7.6% in F1 score and 16% in win rate across various tasks. More importantly, we further demonstrate two exciting findings: (1) Thought-Retriever can indeed help LLM self-evolve after solving more user queries; (2) Thought-Retriever learns to leverage deeper thoughts to answer more abstract user queries.", "keywords": [ "Language Model", "Retrieval-Augmented LLM", "AI Agent", "Self-Evolution" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "Sk2mND99Wp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Hongzi Zhu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Limin Xiao", "gender": "Male", "institution": "Lenovo Group Limited", "country": "CN", "position": "Researcher" }, { "name": "Weifeng Zhang", "gender": "unknown", "institution": "Lenovo Research", "country": "US", "position": "Researcher" }, { "name": "Yan Li", "gender": "unknown", "institution": "Lenovo Research", "country": "CN", "position": "Researcher" }, { "name": "Zekai Xu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhezhi He", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "boning zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 22, "n_ref": 46, "n_ref_all": 61, "n_fig": 8, "n_tab": 4, "L_tab": 337, "n_element_tab": 39, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 2, "n_element_tab_1": 1, "formula_len_all": 472, "formula_len_all_1": 403, "len_all": 90745, "len_all_1": 47123, "len_abs": 1105, "len_title": 116, "len_sents": 28130, "len_sents_1": 22714, "n_sents": 193, "n_sents_1": 155, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1095, "title": "Grouped Correlation Aggregation with Propagation for Stereo Matching", "abs": "Iterative optimization-based methods have dominated the field of stereo matching with extraordinary precision and speed. However, these methods still suffer from low iteration efficiency and insufficient correlation volume with low utilization rates. \nAs the countermeasure, we propose grouped correlation aggregation with propagation, a novel stereo matching method inspired by traditional methods. \nWe design an efficient updater to improve the performance of single iteration optimization. To alleviate the problems of correlation volume, a novel grouped window shifting mechanism and a contour-aware aggregation modified from semi-global matching (SGM) have been introduced. Our method outperforms all methods in zero-shot generalization and ranks 1st on ETH3D among published works. \nAdditionally, we conducted targeted inference optimization on the video stream and demonstrated the improvement in frame rate without sacrificing accuracy through experiments on the simulator. \nFinally, a real-world binocular system is deployed to qualitatively demonstrate the practicality of our method.", "keywords": [ "stereo matching", "computer vision" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "SjufxrSOYd", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Daniel Herbst", "gender": "Male", "institution": "Technical University of Munich", "country": "DE", "position": "MS student" }, { "name": "Stefanie Jegelka", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" } ], "n_formula": 179, "n_formula_1": 36, "n_ref_uni": 41, "n_ref": 133, "n_ref_all": 151, "n_fig": 4, "n_tab": 5, "L_tab": 669, "n_element_tab": 29, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15248, "formula_len_all_1": 2567, "len_all": 251918, "len_all_1": 73147, "len_abs": 1400, "len_title": 106, "len_sents": 78185, "len_sents_1": 32703, "n_sents": 809, "n_sents_1": 273, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 58, "L_abs": 1433, "title": "Invariant Graphon Networks: Approximation and Cut Distance", "abs": "Graph limit models, like graphons for limits of dense graphs, have recently been used as a tool to study size transferability of graph neural networks (GNNs). While most existing literature focuses on message passing GNNs (MPNNs), we attend to *Invariant Graph Networks* (IGNs), a powerful alternative GNN architecture. In this work, we generalize IGNs to graphons, introducing *Invariant Graphon Networks (IWNs)* which are defined using a subset of the IGN basis corresponding to bounded linear operators. Even with this restricted basis, we show that IWNs of order $k + 1$ are at least as powerful as the $k$-dimensional Weisfeiler-Leman (WL) test for graphon-signals and we establish universal approximation results for graphon-signals in $\\mathcal{L}^p$ distances using *signal-weighted homomorphism densities*. This significantly extends the prior work of Cai & Wang (2022), showing that IWNs\u2014a subset of their *IGN-small*\u2014retain effectively the same expressivity as the full IGN basis in the limit. In contrast to their approach, our blueprint of IWNs also aligns better with the geometry of graphon space, for example facilitating comparability to MPNNs. We also highlight that, unlike other GNN architectures such as MPNNs, IWNs are *discontinuous with respect to cut distance*, which causes their lack of convergence and is inherently tied to the definition of $k$-WL. Yet, their transferability remains comparable to MPNNs.", "keywords": [ "Graph neural networks", "invariant graph networks", "universal approximation", "graph limits", "graphons", "transferability", "homomorphism densities", "machine learning theory." ], "rating_list": [ 8, 8, 8, 8, 8 ], "soundness_list": [ 4, 3, 4, 3, 3 ], "presentation_list": [ 2, 4, 4, 3, 4 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3, 1 ] }, { "paper_id": "SjRLAzrVRo", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Jinze Zhao", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "MS student" }, { "name": "Junjie Yang", "gender": "Male", "institution": "Ohio State University", "country": "US", "position": "PhD student" }, { "name": "Peihao Wang", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Yingbin Liang", "gender": "Female", "institution": "The Ohio State University", "country": "US", "position": "Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 61, "n_ref_all": 77, "n_fig": 9, "n_tab": 1, "L_tab": 470, "n_element_tab": 7, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 2054, "n_element_tab_1": 25, "formula_len_all": 123, "formula_len_all_1": 123, "len_all": 100328, "len_all_1": 52999, "len_abs": 1180, "len_title": 108, "len_sents": 28678, "len_sents_1": 24178, "n_sents": 200, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1186, "title": "Meta ControlNet: Enhancing Task Adaptation via Meta Learning", "abs": "Diffusion-based image synthesis has attracted extensive attention recently. In particular, \nControlNet that uses image-based prompts exhibits powerful capability in image tasks such as canny edge detection and generates images well aligned with these prompts. However, vanilla ControlNet generally requires extensive training of around 5000 steps to achieve a desirable control for a single task. Recent context-learning approaches have improved its adaptability, but mainly for edge-based tasks, and rely on paired examples. Thus, two important open issues are yet to be addressed to reach the full potential of ControlNet: (i) zero-shot control for certain tasks and (ii) faster adaptation for non-edge-based tasks. In this paper, we introduce a novel Meta ControlNet method, \nwhich adopts the task-agnostic meta learning technique and features a new layer freezing design. Meta ControlNet significantly reduces learning steps to attain control ability from 5000 to 1000. Further, Meta ControlNet exhibits direct zero-shot adaptability in edge-based tasks without any finetuning, and achieves control within only 100 finetuning steps in more complex non-edge tasks such as Human Pose.", "keywords": [ "Diffusion Models", "Meta Learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 1, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "SjMtxqdQ73", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Jingbo Zhou", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Jun Xia", "gender": "Male", "institution": "Westlake University, China", "country": "CN", "position": "PhD student" }, { "name": "Shaorong Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Sizhe Liu", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Undergrad student" }, { "name": "Yue Liu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zicheng Liu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "hongxin xiang", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "PhD student" }, { "name": "Stan Z. Li", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Chair Professor" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 43, "n_ref": 79, "n_ref_all": 102, "n_fig": 6, "n_tab": 4, "L_tab": 2258, "n_element_tab": 182, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2094, "n_element_tab_1": 165, "formula_len_all": 811, "formula_len_all_1": 462, "len_all": 146358, "len_all_1": 71666, "len_abs": 1826, "len_title": 135, "len_sents": 41703, "len_sents_1": 32440, "n_sents": 294, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1885, "title": "Bridging the Gap between Database Search and \\emph{De Novo} Peptide Sequencing with SearchNovo", "abs": "Accurate protein identification from mass spectrometry (MS) data is fundamental to unraveling the complex roles of proteins in biological systems, with peptide sequencing being a pivotal step in this process. The two main paradigms for peptide sequencing are database search, which matches experimental spectra with peptide sequences from databases, and \\emph{de novo} sequencing, which infers peptide sequences directly from MS without relying on pre-constructed database. Although database search methods are highly accurate, they are limited by their inability to identify novel, modified, or mutated peptides absent from the database. In contrast, \\emph{de novo} sequencing is adept at discovering novel peptides but often struggles with missing peaks issue, further leading to lower precision. We introduce SearchNovo, a novel framework that synergistically integrates the strengths of database search and \\emph{de novo} sequencing to enhance peptide sequencing. SearchNovo employs an efficient search mechanism to retrieve the most similar peptide spectrum match (PSM) from a database for each query spectrum, followed by a fusion module that utilizes the reference peptide sequence to guide the generation of the target sequence. Furthermore, we observed that dissimilar (noisy) reference peptides negatively affect model performance. To mitigate this, we constructed pseudo reference PSMs to minimize their impact. Comprehensive evaluations on multiple datasets reveal that SearchNovo significantly outperforms state-of-the-art models. Also, analysis indicates that many retrieved spectra contain missing peaks absent in the query spectra, and the retrieved reference peptides often share common fragments with the target peptides. These are key elements in the recipe for SearchNovo\u2019s success. The code for reproducing the results are available in the supplementary materials.", "keywords": [ "Protein Identification" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "SimlDuN0YT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Arijit Khan", "gender": "Male", "institution": "Aalborg University, Aalborg University", "country": "DK", "position": "Associate Professor" }, { "name": "Bishwamittra Ghosh", "gender": "Male", "institution": "MPI-SWS", "country": "DE", "position": "Postdoc" }, { "name": "Naheed Anjum Arafat", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Sarah Hasan", "gender": "Female", "institution": "Aalborg University, Aalborg University", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 79, "n_ref": 135, "n_ref_all": 172, "n_fig": 7, "n_tab": 17, "L_tab": 2770, "n_element_tab": 458, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 27, "n_element_tab_1": 2, "formula_len_all": 982, "formula_len_all_1": 485, "len_all": 257383, "len_all_1": 68550, "len_abs": 1556, "len_title": 109, "len_sents": 70419, "len_sents_1": 34594, "n_sents": 539, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1517, "title": "Logical Consistency of Large Language Models in Fact-Checking", "abs": "In recent years, large language model (LLM) has demonstrated significant success in performing varied natural language tasks such as language translation, question-answering, summarising, fact-checking, etc. Despite LLM\u2019s impressive ability to generate human-like texts, LLMs are infamous for their inconsistent responses \u2013 a meaning-preserving change in the input query results in an inconsistent response and attributes to vulnerabilities of LLMs such as hallucination, jail breaking, etc. Consequently, existing research focuses on simple paraphrasing-based consistency assessment of LLMs, and ignores complex queries that necessitates an even better understanding of logical reasoning by an LLM. Our work therefore addresses the logical inconsistency of LLMs under complex logical queries with primitive logical operators, e.g., negation, conjunction, and disjunction. As a test bed, we consider retrieval-augmented LLMs on a fact-checking task involving propositional logic queries from real-world knowledge graphs (KG). Our contributions are three-fold. Benchmark: We introduce three logical fact-checking datasets over KGs for community development towards logically consistent LLMs. Assessment: We propose consistency measures of LLMs on propositional logic queries as input and demonstrate that existing LLMs lack in logical consistency, specially on complex queries. Improvement: We employ supervised fine-tuning to improve the logical consistency of LLMs on the complex\nfact-checking task with KG contexts.", "keywords": [ "LLM", "Logical consistency", "Fact-checking in Knowledge graphs" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "SiH7DwNKZZ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Benedikt Alkin", "gender": "Male", "institution": "Emmi AI GmbH", "country": "AT", "position": "Researcher" }, { "name": "Johannes Brandstetter", "gender": "Male", "institution": "Emmi AI", "country": "AT", "position": "Principal Researcher" }, { "name": "Korbinian P\u00f6ppel", "gender": "unknown", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "PhD student" }, { "name": "Maximilian Beck", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "", "position": "PhD student" }, { "name": "Sepp Hochreiter", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 68, "n_ref": 122, "n_ref_all": 146, "n_fig": 2, "n_tab": 4, "L_tab": 1605, "n_element_tab": 187, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 3076, "n_element_tab_1": 283, "formula_len_all": 725, "formula_len_all_1": 723, "len_all": 172632, "len_all_1": 54862, "len_abs": 1322, "len_title": 45, "len_sents": 32668, "len_sents_1": 23047, "n_sents": 217, "n_sents_1": 149, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 45, "L_abs": 907, "title": "Vision-LSTM: xLSTM as Generic Vision Backbone", "abs": "Transformers are widely used as generic backbones in computer vision, despite initially introduced for natural language processing. Recently, the Long Short-Term Memory (LSTM) has been extended to a scalable and performant architecture - the xLSTM - which overcomes long-standing LSTM limitations via exponential gating and parallelizable matrix memory structure. In this paper, we introduce Vision-LSTM (ViL), an adaption of the xLSTM building blocks to computer vision. ViL comprises a stack of xLSTM blocks where odd blocks process the sequence of patch tokens from top to bottom while even blocks go from bottom to top.\n\nViL achieves strong performances on classification, transfer learning and segmentation tasks as well as a beneficial pre-training cost-to-performance trade-off. Experiments show that ViL holds promise to be further deployed as new generic backbone for computer vision architectures.", "keywords": [ "Computer Vision", "xLSTM", "image classification", "semantic segmentation", "transfer learning", "ImageNet" ], "rating_list": [ 5, 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3, 5 ] }, { "paper_id": "Sh4FOyZRpv", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Chenheng Xu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "MS student" }, { "name": "Guang Cheng", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "Full Professor" }, { "name": "Matthew Yang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "MS student" }, { "name": "Xiaofeng Lin", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 60, "n_ref": 93, "n_ref_all": 101, "n_fig": 3, "n_tab": 11, "L_tab": 3482, "n_element_tab": 339, "n_fig_1": 2, "n_tab_1": 9, "L_tab_1": 2645, "n_element_tab_1": 149, "formula_len_all": 696, "formula_len_all_1": 696, "len_all": 143253, "len_all_1": 59455, "len_abs": 1319, "len_title": 110, "len_sents": 29904, "len_sents_1": 26055, "n_sents": 194, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1252, "title": "CTSyn: A Foundational Model for Cross Tabular Data Generation", "abs": "Generative Foundation Models (GFMs) have achieved remarkable success in producing high-quality synthetic data for images and text. However, their application to tabular data presents significant challenges due to the heterogeneous nature of table features. Current cross-table learning frameworks struggle with the absence of a generative model backbone and a mechanism to decode heterogeneous feature values. To address these challenges, we propose the Cross-Table Synthesizer (CTSyn), a diffusion-based foundational model for tabular data generation. CTSyn features two key components: an Autoencoder network that consolidates diverse tables into a unified latent space and dynamically reconstructs table values based on the provided table schema embedding, adapting to heterogeneous datasets; and a conditional latent diffusion model that samples from this learned latent space. Through large-scale pre-training, CTSyn not only outperforms existing table synthesizers on standard tabular data generation benchmarks in terms of utility and diversity, but also uniquely enhances the performance of downstream machine learning tasks, surpassing what is achievable with real data. This establishes CTSyn as a new paradigm for synthetic table generation.", "keywords": [ "Foundation Model", "Tabular Data", "Synthetic Data Generation" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "SgymXhOEA5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jin-Woo Park", "gender": "Male", "institution": "mAy-I Corp.", "country": "", "position": "Researcher" }, { "name": "Jong-Seok Lee", "gender": "unknown", "institution": "Yonsei University", "country": "KR", "position": "Full Professor" }, { "name": "Myungseo Song", "gender": "unknown", "institution": "mAy-I Corp.", "country": "KR", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 37, "n_ref": 70, "n_ref_all": 107, "n_fig": 17, "n_tab": 11, "L_tab": 7014, "n_element_tab": 823, "n_fig_1": 14, "n_tab_1": 7, "L_tab_1": 5010, "n_element_tab_1": 557, "formula_len_all": 358, "formula_len_all_1": 243, "len_all": 151012, "len_all_1": 74422, "len_abs": 1502, "len_title": 101, "len_sents": 41869, "len_sents_1": 29688, "n_sents": 333, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 53, "L_abs": 1510, "title": "Exploring the Camera Bias of Person Re-identification", "abs": "We empirically investigate the camera bias of person re-identification (ReID) models. Previously, camera-aware methods have been proposed to address this issue, but they are largely confined to training domains of the models. We measure the camera bias of ReID models on unseen domains and reveal that camera bias becomes more pronounced under data distribution shifts. As a debiasing method for unseen domain data, we revisit feature normalization on embedding vectors. While the normalization has been used as a straightforward solution, its underlying causes and broader applicability remain unexplored. We analyze why this simple method is effective at reducing bias and show that it can be applied to detailed bias factors such as low-level image properties and body angle. Furthermore, we validate its generalizability across various models and benchmarks, highlighting its potential as a simple yet effective test-time postprocessing method for ReID. In addition, we explore the inherent risk of camera bias in unsupervised learning of ReID models. The unsupervised models remain highly biased towards camera labels even for seen domain data, indicating substantial room for improvement. Based on observations of the negative impact of camera-biased pseudo labels on training, we suggest simple training strategies to mitigate the bias. By applying these strategies to existing unsupervised learning algorithms, we show that significant performance improvements can be achieved with minor modifications.", "keywords": [ "Person re-identification", "Camera bias", "Debiasing", "Unsupervised learning" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "SgAPzJdAHi", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Juanzi Li", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Lei Hou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Shulin Cao", "gender": "Female", "institution": "Zhipu AI", "country": "CN", "position": "Researcher" }, { "name": "Yantao Liu", "gender": "Not Specified", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Zhao Zhang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Zijun Yao", "gender": "Male", "institution": "Department of Computer Science and Technology, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 49, "n_ref": 104, "n_ref_all": 125, "n_fig": 2, "n_tab": 16, "L_tab": 7633, "n_element_tab": 570, "n_fig_1": 2, "n_tab_1": 13, "L_tab_1": 4735, "n_element_tab_1": 384, "formula_len_all": 115, "formula_len_all_1": 94, "len_all": 186112, "len_all_1": 65879, "len_abs": 1689, "len_title": 141, "len_sents": 49604, "len_sents_1": 28051, "n_sents": 346, "n_sents_1": 153, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1712, "title": "Aligning Teacher with Student Preferences for Tailored Instruction Tuning Dataset Generation", "abs": "Enhancing the reasoning abilities of lightweight language models (LMs) for tasks like decision-making often relies on instruction-tuning, a method that trains LMs to mimic the reasoning process using labeled question-rationale pairs, known as instruction-tuning datasets, which are typically generated by more powerful teacher LMs. However, current methods for generating these instruction-tuning datasets tend to focus solely on the quality of the questions and rationales from the teacher model\u2019s perspective, often neglecting the learning preferences of the student language model. To fill this gap, we propose **ARTE** (**A**ligning Teache**R** with Studen**T** Preferenc**E**s), a novel framework that adapts the teacher LM\u2019s outputs to the student\u2019s preferences, inspired by \"responsive teaching\" in pedagogy. Our method involves three key steps: (1) generating draft question-rationale pairs from the teacher model, (2) collecting the student\u2019s preferences on these draft pairs via one-shot in-context learning, and (3) aligning the teacher model using Direct Preference Optimization (DPO), then finally curating tailored question-rationale pairs from the aligned teacher for student training. Through extensive experiments on academic reasoning benchmarks, we demonstrate that student models fine-tuned with tailored datasets by ARTE achieve significant improvements across various reasoning tasks, outperforming existing instruction-tuning datasets. Moreover, we thoroughly investigate the generalization of ARTE, including the generalization of fine-tuned student models in reasoning ability and the generalization of aligned teacher models to generate tailored training data across tasks and students.", "keywords": [ "Large Language Model", "Knowledge Distillation", "Alignment" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Sfmk5amxFu", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Aakash Lahoti", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Albert Gu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Ratish Puduppully", "gender": "Male", "institution": "IT University of Copenhagen", "country": "DK", "position": "Assistant Professor" }, { "name": "Tanya Marwah", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 33, "n_ref_uni": 23, "n_ref": 59, "n_ref_all": 85, "n_fig": 5, "n_tab": 6, "L_tab": 2963, "n_element_tab": 310, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 800, "n_element_tab_1": 124, "formula_len_all": 2021, "formula_len_all_1": 1549, "len_all": 125497, "len_all_1": 71495, "len_abs": 1982, "len_title": 92, "len_sents": 37859, "len_sents_1": 31355, "n_sents": 309, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1497, "title": "Chimera: State Space Models Beyond Sequences", "abs": "Powerful deep learning methods based on Transformers are used to model diverse data modalities such as sequences, images, and graphs. \nThese methods typically use off-the-shelf modules like self-attention, which are domain-agnostic and treat data as an unordered set of elements.\nTo improve performance, researchers employ inductive biases\u2014such as position embeddings in sequences and images, and random walks in graphs\u2014to inject the domain structure, or *topology*, into the model.\nHowever, these inductive biases are carefully engineered heuristics that must be designed for each modality, requiring significant research effort.\nIn this work, we propose *Chimera*, a unified framework that mathematically generalizes state space models to incorporate the topological structure of data in a principled way.\nWe demonstrate that our method achieves state-of-the-art performance across domains including language, vision, and graphs. Chimera outperforms BERT on the GLUE benchmark by 0.7 points, surpasses ViT by 2.6% on ImageNet-1k classification accuracy, and outperforms all baselines on the Long Range Graph Benchmark with a 12% improvement on PascalVOC.\nThis validates Chimera's methodological improvement, which allows it to directly capture the underlying topology, providing a strong inductive bias across modalities.\nFurthermore, being topologically aware enables our method to achieve a linear time complexity for sequences and images, in contrast to the quadratic complexity of attention.", "keywords": [ "Deep Learning Architectures", "Sequence Models", "State Space Models", "Mamba" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "Sfct4aXXcw", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Carl Henrik Ek", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "Associate Professor" }, { "name": "Jason Ross Brown", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Robert Mullins", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Full Professor" } ], "n_formula": 28, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 48, "n_ref_all": 91, "n_fig": 4, "n_tab": 5, "L_tab": 2740, "n_element_tab": 198, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2739, "formula_len_all_1": 740, "len_all": 153704, "len_all_1": 59396, "len_abs": 1161, "len_title": 118, "len_sents": 52129, "len_sents_1": 28921, "n_sents": 417, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1167, "title": "Learning from Preferences and Mixed Demonstrations in General Settings", "abs": "Reinforcement learning is a general method for learning in sequential settings, but it can often be difficult to specify a good reward function when the task is complex.\nIn these cases, preference feedback or expert demonstrations can be used instead.\nHowever, existing approaches utilising both together are either ad-hoc or rely on domain-specific properties.\nBuilding upon previous work, we develop a novel theoretical framework for learning from human data.\nBased on this we introduce LEOPARD: Learning Estimated Objectives from Preferences And Ranked Demonstrations.\nLEOPARD can simultaneously learn from a broad range of data, including negative/failed demonstrations, to effectively learn reward functions in general domains.\nIt does this by modelling the human feedback as reward-rational partial orderings over available trajectories.\nWe find that when a limited amount of human feedback is available, LEOPARD outperforms the current standard practice of pre-training on demonstrations and finetuning on preferences, as well as other baselines.\nFurthermore, we show that LEOPARD learns faster when given many types of feedback, rather than just a single one.", "keywords": [ "reinforcement learning", "rl", "human feedback", "rlhf", "modelling", "preferences", "demonstrations", "rankings", "machine learning", "reward learning" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "SfZpk8CV9l", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Chengqi Duan", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Hao Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Hao Tian", "gender": "Male", "institution": "Sensetime Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Jifeng Dai", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Kun Wang", "gender": "Male", "institution": "SenseTime Group Ltd", "country": "HK", "position": "Researcher" }, { "name": "Rongyao Fang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Rui Zhao", "gender": "Male", "institution": "SenseTime Research", "country": "CN", "position": "Researcher" }, { "name": "Xihui Liu", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Xingyu Zeng", "gender": "Male", "institution": "SenseTime Group Limited", "country": "HK", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 31, "n_ref": 81, "n_ref_all": 92, "n_fig": 9, "n_tab": 6, "L_tab": 1809, "n_element_tab": 262, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 246, "n_element_tab_1": 32, "formula_len_all": 318, "formula_len_all_1": 154, "len_all": 156757, "len_all_1": 51505, "len_abs": 0, "len_title": 0, "len_sents": 41075, "len_sents_1": 25369, "n_sents": 299, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1225, "title": "PUMA: Empowering Unified MLLM with Multi-granular Visual Generation", "abs": "Recent advancements in multimodal foundation models have yielded significant progress in vision-language understanding. Initial attempts have also explored the potential of multimodal large language models (MLLMs) for visual content generation. However, existing works have insufficiently addressed the varying granularity demands of different image generation tasks within a unified MLLM paradigm \u2014 from the diversity required in text-to-image generation to the precise controllability needed in image manipulation. In this work, we propose PUMA, emPowering Unified MLLM with Multi-grAnular visual generation. PUMA unifies multi-granular visual features as both inputs and outputs of MLLMs, elegantly addressing the different granularity requirements of various image generation tasks within a unified MLLM framework. Following multimodal pretraining and task-specific instruction tuning, PUMA demonstrates proficiency in a wide range of multimodal tasks, including image understanding, diverse text-to-image generation, editing, inpainting, colorization, and conditional generation. This work represents a significant step towards a truly unified MLLM capable of adapting to the granularity demands of various visual tasks.", "keywords": [ "Multimodal Large Language Model" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "SfTy1ac4OX", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Daichi Zhang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Jianmin Bao", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Sabine Susstrunk", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Full Professor" }, { "name": "Shiming Ge", "gender": "Male", "institution": "Institute of Information Engineering, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Tong Zhang", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Postdoc" }, { "name": "Yixin Cheng", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 40, "n_ref": 94, "n_ref_all": 122, "n_fig": 9, "n_tab": 3, "L_tab": 3805, "n_element_tab": 263, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 2843, "n_element_tab_1": 231, "formula_len_all": 247, "formula_len_all_1": 248, "len_all": 124253, "len_all_1": 58114, "len_abs": 1374, "len_title": 116, "len_sents": 29116, "len_sents_1": 25040, "n_sents": 207, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1433, "title": "Exploring Image-Text Discrepancy for Universal Fake Image Detection", "abs": "With the rapid development of generative models, detecting generated images to prevent their malicious use has become a critical issue recently. Existing methods frame this challenge as a binary image classification task. However, such methods focus only on visual space, yielding trained detectors susceptible to overfitting specific image patterns and incapable of generalizing to unseen models. In this paper, we address this issue from a multi-modal perspective and find that fake images exhibit more distinct discrepancies with corresponding captions compared to real images. Upon this observation, we propose to leverage the \\textbf{I}mage-\\textbf{T}ext \\textbf{D}iscrepanc\\textbf{y}~(\\textbf{TIDY}) in joint visual-language space for \\textit{universal fake image detection}. Specifically, we first measure the distance of the images and corresponding captions in the latent spaces of CLIP, and then tune an MLP head to perform the usual detection task. Since there usually exists local artifacts in fake images, we further propose a global-to-local discrepancy scheme that first explores the discrepancy on the whole image and then each semantic object described in the caption, which can explore more fine-grained local semantic clues. Extensive experiments demonstrate the superiority of our method against other state-of-the-art competitors with impressive generalization and robustness on various recent generative models.", "keywords": [ "vision-language model", "image-text discrepancy", "fake image detection" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "SfNmgDqeEa", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Ariel Goldstein", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "Assistant Professor" }, { "name": "Daria Lioubashevski", "gender": "Female", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "MS student" }, { "name": "Gabriel Stanovsky", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "", "position": "Assistant Professor" }, { "name": "Tomer M. Schlank", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 49, "n_ref_all": 105, "n_fig": 16, "n_tab": 12, "L_tab": 1977, "n_element_tab": 178, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 683, "n_element_tab_1": 30, "formula_len_all": 73, "formula_len_all_1": 0, "len_all": 151567, "len_all_1": 60149, "len_abs": 1495, "len_title": 116, "len_sents": 58577, "len_sents_1": 31111, "n_sents": 321, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1512, "title": "Looking Beyond the Top-1: Transformers Determine Top Tokens in Order", "abs": "Understanding the inner workings of Transformers is crucial for achieving more\naccurate and efficient predictions. In this work, we analyze the computation performed\nby Transformers in the layers after the top-1 prediction has become fixed, which has been\npreviously referred to as the \u201csaturation event\u201d. We expand the concept of saturation events\nfor top-k tokens, demonstrating that similar saturation events occur across language, vision, \nand speech models. We find that these saturation events happen in order of the \ncorresponding tokens\u2019 ranking, i.e., the model first decides on the top ranking token, then \nthe second highest ranking token, and so on. This phenomenon seems intrinsic to the \nTransformer architecture, occurring across different architectural variants (decoder-only, \nencoder-only, and to a lesser extent full-Transformer), and even in untrained Transformers.\nWe propose an underlying mechanism of task transition for this sequential saturation, where \ntask k corresponds to predicting the k-th most probable token, and the saturation events are \nin fact discrete transitions between the tasks. In support of this we show that it is possible to \npredict the current task from hidden layer embedding. Furthermore, using an intervention \nmethod we demonstrate that we can cause the model to switch from one task to the next. \nFinally, leveraging our findings, we introduce a novel token-level early-exit strategy, which \nsurpasses existing methods in balancing performance and efficiency.", "keywords": [ "mechanistic interpretability", "transformer", "large language model", "efficient inference" ], "rating_list": [ 3, 8, 8, 5, 8 ], "soundness_list": [ 1, 3, 4, 2, 3 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 4, 2, 4 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "Sf4ep9Udjf", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Abhronil Sengupta", "gender": "unknown", "institution": "Pennsylvania State University", "country": "", "position": "Assistant Professor" }, { "name": "Malyaban Bal", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 17, "n_ref_uni": 38, "n_ref": 78, "n_ref_all": 100, "n_fig": 6, "n_tab": 10, "L_tab": 2851, "n_element_tab": 181, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 2046, "n_element_tab_1": 152, "formula_len_all": 1152, "formula_len_all_1": 1031, "len_all": 133182, "len_all_1": 63096, "len_abs": 1714, "len_title": 144, "len_sents": 40511, "len_sents_1": 28495, "n_sents": 281, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1720, "title": "P-SPIKESSM: HARNESSING PROBABILISTIC SPIKING STATE SPACE MODELS FOR LONG-RANGE DEPENDENCY TASKS", "abs": "Spiking neural networks (SNNs) are posited as a computationally efficient and biologically plausible alternative to conventional neural architectures, with their core computational framework primarily using the leaky integrate-and-fire (LIF) neuron model. However, the limited hidden state representation of LIF neurons, characterized by a scalar membrane potential, and sequential spike generation process, poses challenges for effectively developing scalable spiking models to address long-range dependencies in sequence learning tasks. In this study, we develop a scalable probabilistic spiking learning framework for long-range dependency tasks leveraging the fundamentals of state space models. Unlike LIF neurons that rely on the deterministic Heaviside function for a sequential process of spike generation, we introduce a SpikeSampler layer that samples spikes stochastically based on an SSM-based neuronal model while allowing parallel computations. To address non-differentiability of the spiking operation and enable effective training, we also propose a surrogate function tailored for the stochastic nature of the SpikeSampler layer. To enhance inter-neuron communication, we introduce the SpikeMixer block, which integrates spikes from neuron populations in each layer. This is followed by a ClampFuse layer, incorporating a residual connection to capture complex dependencies, enabling scalability of the model. Our models attain state-of-the-art performance among SNN models across diverse long-range dependency tasks, encompassing the Long Range Arena benchmark, permuted sequential MNIST, and the Speech Command dataset and demonstrate sparse spiking pattern highlighting its computational efficiency.", "keywords": [ "Spiking Neural Networks", "Sequence Learning" ], "rating_list": [ 8, 5, 8, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "SeQ8l8xo1r", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Anze Xie", "gender": "Male", "institution": "Computer Science and Engineering Department, University of California, San Diego", "country": "", "position": "MS student" }, { "name": "Hao Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Haojian Jin", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Ion Stoica", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Lanxiang Hu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Nan Jiang", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "Undergrad student" }, { "name": "Qiyu Li", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 42, "n_ref": 68, "n_ref_all": 86, "n_fig": 6, "n_tab": 9, "L_tab": 7250, "n_element_tab": 393, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 371, "n_element_tab_1": 6, "formula_len_all": 455, "formula_len_all_1": 313, "len_all": 222393, "len_all_1": 63681, "len_abs": 3249, "len_title": 111, "len_sents": 75900, "len_sents_1": 31087, "n_sents": 615, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1286, "title": "GameArena: Evaluating LLM Reasoning through Live Computer Games", "abs": "Evaluating the reasoning abilities of large language models (LLMs) is challenging. Existing benchmarks often depend on static datasets, which are vulnerable to data contamination and may get saturated over time, or on binary live human feedback that conflates reasoning with other abilities. As the most prominent dynamic benchmark, Chatbot Arena evaluates open-ended questions in real-world settings, but lacks the granularity in assessing specific reasoning capabilities. We introduce GameArena, a dynamic benchmark designed to evaluate LLM reasoning capabilities through interactive gameplay with humans. GameArena consists of three games designed to test specific reasoning capabilities (e.g., deductive and inductive reasoning), while keeping participants entertained and engaged. We analyze the gaming data retrospectively to uncover the underlying reasoning processes of LLMs and measure their fine-grained reasoning capabilities. We collect over 2000 game sessions and provide detailed assessments of various reasoning capabilities for five state-of-the-art LLMs. Our user study with 100 participants suggests that GameArena improves user engagement compared to Chatbot Arena. For the first time, GameArena enables the collection of step-by-step LLM reasoning data in the wild.", "keywords": [ "Large Language Models", "LLM evaluation" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "Se6MgCtRhz", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Bin Dong", "gender": "Male", "institution": "Peking University", "country": "", "position": "Associate Professor" }, { "name": "Guoxiong Gao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Jiedong Jiang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Qi Gao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Tianyi Xu", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Zihan Qin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Wang Yutong", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" } ], "n_formula": 36, "n_formula_1": 0, "n_ref_uni": 17, "n_ref": 44, "n_ref_all": 55, "n_fig": 5, "n_tab": 7, "L_tab": 1873, "n_element_tab": 103, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 270, "n_element_tab_1": 30, "formula_len_all": 2532, "formula_len_all_1": 0, "len_all": 231150, "len_all_1": 57547, "len_abs": 1516, "len_title": 99, "len_sents": 90997, "len_sents_1": 29080, "n_sents": 760, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1530, "title": "Herald: A Natural Language Annotated Lean 4 Dataset", "abs": "Verifiable formal languages like Lean have profoundly impacted mathematical reasoning, particularly through the use of large language models (LLMs) for automated reasoning. A significant challenge in training LLMs for these formal languages is the lack of parallel datasets that align natural language with formal language proofs. To address this challenge, this paper introduces a novel framework for translating the Mathlib4 corpus (a unified library of mathematics in formal language Lean 4) into natural language. Building upon this, we employ a dual augmentation strategy that combines tactic-based and informal-based approaches, leveraging the Lean-jixia system, a Lean 4 analyzer. We present the results of this pipeline on Mathlib4 as Herald (Hierarchy and Retrieval-based Translated Lean Dataset). We also propose the Herald Translator, which is fine-tuned on Herald. Herald translator achieves a 96.7\\% accuracy (Pass@128) on formalizing statements in the miniF2F-test and a 23.5\\% accuracy on our internal graduate-level textbook dataset, outperforming InternLM2-Math-Plus-7B (73.0\\% and 7.5\\%) and TheoremLlama (50.1\\% and 4.0\\%). Furthermore, we propose a section-level translation framework for real-world applications. As a direct application of Herald translator, we have successfully translated a template section in the Stack project, marking a notable progress in the automatic formalization of graduate-level mathematical literature. Our model, along with the datasets, will be open-sourced to the public soon.", "keywords": [ "Lean 4", "Autoformalizing", "LLM", "Retrieval Augmented Generation", "Dataset" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "Se2aTG9Oui", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Bernd Heidergott", "gender": "Male", "institution": "Vrije Universiteit Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "Christian P.C. Franssen", "gender": "Male", "institution": "Vrije Universiteit Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Jinyang Jiang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yijie Peng", "gender": "Male", "institution": "Peking University", "country": "", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 20, "n_ref_uni": 50, "n_ref": 83, "n_ref_all": 116, "n_fig": 12, "n_tab": 8, "L_tab": 1264, "n_element_tab": 222, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1248, "formula_len_all_1": 1187, "len_all": 149842, "len_all_1": 70149, "len_abs": 1256, "len_title": 118, "len_sents": 47441, "len_sents_1": 32399, "n_sents": 374, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1267, "title": "CoNNect: A Swiss-Army-Knife Regularizer for Pruning of Neural Networks", "abs": "Pruning encompasses a range of techniques aimed at increasing the sparsity of neural networks (NNs). These techniques can generally be framed as minimizing a loss function subject to an $L_0$-norm constraint. In this paper, we introduce CoNNect, a novel differentiable regularizer for sparse NN training that quantifies connectivity in weighted graphs. Our theoretical and numerical analyses show that CoNNect integrates seamlessly with many established pruning strategies and is applicable to both unstructured and structured pruning. By including CoNNect as a regularizer during training, we ensure neural networks maintain connectivity between input and output layers, addressing limitations of $L_1$-regularization, a common surrogate for $L_0$-norm regularization. We prove that CoNNect effectively approximates $L_0$-regularization, guaranteeing maximally connected network structures as stable stationary points and avoiding issues like layer collapse. Through numerical experiments, we demonstrate that classical pruning strategies benefit from CoNNect regularization compared to $L_1$- and $L_2$-norm regularization. Additionally, we show that integrating CoNNect into LLM-pruner, a one-shot pruning method for large language models, yields improved results.", "keywords": [ "Connectivity", "Regularization", "Pruning" ], "rating_list": [ 3, 6, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 3, 1 ], "confidence_list": [ 3, 4, 1, 3, 5 ] }, { "paper_id": "Sd4wYYOhmY", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Akim Kotelnikov", "gender": "Male", "institution": "Higher School of Economics, Higher School of Economics", "country": "", "position": "MS student" }, { "name": "Artem Babenko", "gender": "Male", "institution": "Yandex", "country": "RU", "position": "Researcher" }, { "name": "Yury Gorishniy", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 2, "n_ref_uni": 21, "n_ref": 107, "n_ref_all": 178, "n_fig": 12, "n_tab": 12, "L_tab": 5299, "n_element_tab": 318, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 773, "n_element_tab_1": 140, "formula_len_all": 981, "formula_len_all_1": 60, "len_all": 189643, "len_all_1": 64916, "len_abs": 3290, "len_title": 121, "len_sents": 59936, "len_sents_1": 28843, "n_sents": 531, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1397, "title": "TabM: Advancing tabular deep learning with parameter-efficient ensembling", "abs": "Deep learning architectures for supervised learning on tabular data range from simple multilayer perceptrons (MLP) to sophisticated Transformers and retrieval-augmented methods.\nThis study highlights a major, yet so far overlooked opportunity for substantially improving tabular MLPs; namely, parameter-efficient ensembling -- a paradigm for imitating an ensemble of models with just one model.\nWe start by describing TabM -- a simple model based on MLP and BatchEnsemble (an existing technique), improved with our custom modifications.\nThen, we perform a large scale evaluation of tabular DL architectures on public benchmarks in terms of both task performance and efficiency, which renders the landscape of tabular DL in a new light.\nIn particular, we find that TabM outperforms prior tabular DL models, while the complexity of attention- and retrieval-based methods does not pay off.\nLastly, we conduct a detailed empirical analysis, that sheds some light on the high performance of TabM.\nFor example, we show that parameter-efficient ensembling is not an arbitrary trick, but rather a highly effective way to reduce overfitting and improve optimization dynamics of tabular MLPs.\nOverall, our work brings an impactful technique to tabular DL, analyses its behaviour, and advances the performance-efficiency tradeoff with TabM -- a simple and powerful baseline for researchers and practitioners.", "keywords": [ "tabular", "tabular data", "deep learning", "architecture" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "SctfBCLmWo", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Kaiming He", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Zhuang Liu", "gender": "Male", "institution": "FAIR, Meta AI", "country": "US", "position": "Research Scientist" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 66, "n_ref": 143, "n_ref_all": 168, "n_fig": 5, "n_tab": 8, "L_tab": 593, "n_element_tab": 67, "n_fig_1": 3, "n_tab_1": 13, "L_tab_1": 1808, "n_element_tab_1": 299, "formula_len_all": 320, "formula_len_all_1": 0, "len_all": 153685, "len_all_1": 69511, "len_abs": 778, "len_title": 100, "len_sents": 42260, "len_sents_1": 31507, "n_sents": 324, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 52, "L_abs": 781, "title": "A Decade's Battle on Dataset Bias: Are We There Yet?", "abs": "We revisit the ``dataset classification'' experiment suggested by Torralba & Efros (2011) a decade ago, in the new era with large-scale, diverse, and hopefully less biased datasets as well as more capable neural network architectures. Surprisingly, we observe that modern neural networks can achieve excellent accuracy in classifying which dataset an image is from: e.g., we report 84.7% accuracy on held-out validation data for the three-way classification problem consisting of the YFCC, CC, and DataComp datasets. Our further experiments show that such a dataset classifier could learn semantic features that are generalizable and transferable, which cannot be explained by memorization. We hope our discovery will inspire the community to rethink issues involving dataset bias.", "keywords": [ "Vision datasets", "Dataset bias", "Deep learning" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 4, 4 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "ScVnYBaSEw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Juli\u00e1n Tachella", "gender": "Male", "institution": "CNRS", "country": "FR", "position": "Researcher" }, { "name": "Laurent Jacques", "gender": "Male", "institution": "UCLouvain", "country": "", "position": "Full Professor" }, { "name": "Mike Davies", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Full Professor" } ], "n_formula": 68, "n_formula_1": 38, "n_ref_uni": 42, "n_ref": 77, "n_ref_all": 109, "n_fig": 6, "n_tab": 5, "L_tab": 1617, "n_element_tab": 144, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 669, "n_element_tab_1": 23, "formula_len_all": 3357, "formula_len_all_1": 1738, "len_all": 142845, "len_all_1": 63320, "len_abs": 1036, "len_title": 141, "len_sents": 38556, "len_sents_1": 26652, "n_sents": 315, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1045, "title": "UNSURE: self-supervised learning with Unknown Noise level and Stein's Unbiased Risk Estimate", "abs": "Recently, many self-supervised learning methods for image reconstruction have been proposed that can learn from noisy data alone, bypassing the need for ground-truth references. Most existing methods cluster around two classes: i) Stein's Unbiased Risk Estimate (SURE) and similar approaches that assume full knowledge of the noise distribution, and ii) Noise2Self and similar cross-validation methods that require very mild knowledge about the noise distribution. The first class of methods tends to be impractical, as the noise level is often unknown in real-world applications, and the second class is often suboptimal compared to supervised learning.\nIn this paper, we provide a theoretical framework that characterizes this expressivity-robustness trade-off and propose a new approach based on SURE, but unlike the standard SURE, does not require knowledge about the noise level. Throughout a series of experiments, we show that the proposed estimator outperforms other existing self-supervised methods on various imaging inverse problems.", "keywords": [ "self-supervised learning", "imaging inverse problems" ], "rating_list": [ 3, 8, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "ScI7IlKGdI", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Junhao Zheng", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Qianli Ma", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Shengjie Qiu", "gender": "unknown", "institution": "South China University of Technology", "country": "", "position": "Undergrad student" }, { "name": "Xidi Cai", "gender": "unknown", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 112, "n_formula_1": 1, "n_ref_uni": 40, "n_ref": 77, "n_ref_all": 178, "n_fig": 29, "n_tab": 8, "L_tab": 4469, "n_element_tab": 233, "n_fig_1": 16, "n_tab_1": 2, "L_tab_1": 723, "n_element_tab_1": 58, "formula_len_all": 4109, "formula_len_all_1": 32, "len_all": 333971, "len_all_1": 65456, "len_abs": 1200, "len_title": 108, "len_sents": 107642, "len_sents_1": 30791, "n_sents": 874, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1162, "title": "Spurious Forgetting in Continual Learning of Language Models", "abs": "Recent advancements in large language models (LLMs) reveal a perplexing phenomenon in continual learning: despite extensive training, models experience significant performance declines, raising questions about task alignment and underlying knowledge retention. This study first explores the concept of \"spurious forgetting\", proposing that such performance drops often reflect a decline in task alignment rather than true knowledge loss. Through controlled experiments with a synthesized dataset, we investigate the dynamics of model performance during the initial training phases of new tasks, discovering that early optimization steps can disrupt previously established task alignments. Our theoretical analysis connects these shifts to orthogonal updates in model weights, providing a robust framework for understanding this behavior. Ultimately, we introduce a Freezing strategy that fix the bottom layers of the model, leading to substantial improvements in four continual learning scenarios. Our findings underscore the critical distinction between task alignment and knowledge retention, paving the way for more effective strategies in continual learning.", "keywords": [ "Continual Learning", "Language Models", "Large Language Models", "Catastrophic Forgetting", "Incremental Learning", "Lifelong Learning" ], "rating_list": [ 5, 8, 6 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "Sc5rcsoyKR", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Bin Cui", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Fangcheng Fu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Postdoc" }, { "name": "Peichao Lai", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Postdoc" }, { "name": "Zhengfeng Zhang", "gender": "Male", "institution": "Fuzhou University", "country": "", "position": "MS student" }, { "name": "Zhang wen tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 51, "n_ref": 96, "n_ref_all": 128, "n_fig": 13, "n_tab": 10, "L_tab": 3697, "n_element_tab": 645, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3353, "n_element_tab_1": 106, "formula_len_all": 1026, "formula_len_all_1": 1034, "len_all": 244979, "len_all_1": 63337, "len_abs": 1327, "len_title": 125, "len_sents": 45190, "len_sents_1": 26940, "n_sents": 311, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 123, "L_abs": 1334, "title": "Enhancing Unsupervised Sentence Embeddings via Knowledge-Driven Data Augmentation and Gaussian-Decayed Contrastive Learning", "abs": "Recently, using large language models (LLMs) for data augmentation has led to considerable improvements in unsupervised sentence embedding models. However, existing methods encounter two primary challenges: limited data diversity and high data noise. Current approaches often neglect fine-grained knowledge, such as entities and quantities, leading to insufficient diversity. Additionally, unsupervised data frequently lacks discriminative information, and the generated synthetic samples may introduce noise. In this paper, we propose a pipeline-based data augmentation method via LLMs and introduce the Gaussian-decayed gradient-assisted Contrastive Sentence Embedding (GCSE) model to enhance unsupervised sentence embeddings. To tackle the issue of low data diversity, our pipeline utilizes knowledge graphs (KGs) to extract entities and quantities, enabling LLMs to generate more diverse, knowledge-enriched samples. To address high data noise, the GCSE model uses a Gaussian-decayed function to limit the impact of false hard negative samples, enhancing the model's discriminative capability. Experimental results show that our approach achieves state-of-the-art performance in semantic textual similarity (STS) tasks, using fewer data samples and smaller LLMs, demonstrating its efficiency and robustness across various models.", "keywords": [ "Contrastive Learning", "Unsupervised Sentence Representation Learning", "Few-shot Learning", "NLP" ], "rating_list": [ 5, 3, 3, 6, 5 ], "soundness_list": [ 1, 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "Sc382pFw86", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Fan Zhou", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Jieping Ye", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Kai Liu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Intern" }, { "name": "Rongxin Jiang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Researcher" }, { "name": "Yaowu Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Ze Chen", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Zhihang Fu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yue Wu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 51, "n_ref": 151, "n_ref_all": 185, "n_fig": 12, "n_tab": 15, "L_tab": 4166, "n_element_tab": 468, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 286, "n_element_tab_1": 57, "formula_len_all": 139, "formula_len_all_1": 139, "len_all": 211345, "len_all_1": 61503, "len_abs": 1647, "len_title": 116, "len_sents": 63145, "len_sents_1": 31072, "n_sents": 397, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1653, "title": "Structure-aware Domain Knowledge Injection for Large Language Models", "abs": "This paper introduces a pioneering methodology, termed StructTuning, to efficiently transform foundation Large Language Models (LLMs) into domain specialists. It significantly reduces the training corpus requirement to a mere 0.3%, while achieving an impressive 50% of traditional knowledge injection performance. Our method is inspired by the educational processes of human students, particularly how structured domain knowledge from textbooks is assimilated and subsequently applied to tackle real-world challenges through specific exercises. Based on this, we propose a novel two-stage strategy for knowledge injection and alignment: Structure-aware Continual Pre-Training (SCPT) and Structure-aware Supervised Fine-Tuning (SSFT). In the SCPT phase, we automatically extract the domain knowledge taxonomy and reorganize the training corpora, enabling LLMs to effectively link textual segments to targeted knowledge points within the taxonomy. In the SSFT phase, we explicitly prompt models to elucidate the underlying knowledge structure in their outputs, leveraging the structured domain insight to address practical problems. Our ultimate method has undergone extensive evaluations across model architectures and scales, using closed-book question-answering tasks on LongBench and MMedBench datasets. Furthermore, we have investigated the scalability of structure-aware knowledge injection across varying sizes of training corpora, which lays a foundation for scaling up our StructTuning for stronger domain-specific LLMs with comprehensive data utilization. Code is available at this anonymous URL: https://anonymous.4open.science/r/StructTuning/.", "keywords": [ "knowledge injection", "structured knowledge", "large language models" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "SbV2eJC7Ci", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Hilal Asi", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Kunal Talwar", "gender": "Male", "institution": "Apple", "country": "", "position": "Research Scientist" }, { "name": "Vinod Raman", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" } ], "n_formula": 181, "n_formula_1": 28, "n_ref_uni": 28, "n_ref": 76, "n_ref_all": 89, "n_fig": 1, "n_tab": 8, "L_tab": 1020, "n_element_tab": 96, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 170, "n_element_tab_1": 10, "formula_len_all": 25719, "formula_len_all_1": 1543, "len_all": 229589, "len_all_1": 65940, "len_abs": 953, "len_title": 92, "len_sents": 62441, "len_sents_1": 27352, "n_sents": 770, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1259, "title": "Faster Rates for Private Adversarial Bandits", "abs": "We design new differentially private algorithms for the problems of adversarial bandits and bandits with expert advice. For adversarial bandits, we give a simple and efficient conversion of any non-private bandit algorithm to a private bandit algorithm. Instantiating our conversion with existing non-private bandit algorithms gives a regret upper bound of $O\\left(\\frac{\\sqrt{KT}}{\\sqrt{\\epsilon}}\\right)$, improving upon the existing upper bound $O\\left(\\frac{\\sqrt{KT \\log(KT)}}{\\epsilon}\\right)$ for all $\\epsilon \\leq 1$. In particular, our algorithms allow for sublinear expected regret even when $\\epsilon \\leq \\frac{1}{\\sqrt{T}}$, establishing the first known separation between central and local differential privacy for this problem. For bandits with expert advice, we give the first differentially private algorithms, with expected regret $O\\left(\\frac{\\sqrt{NT}}{\\sqrt{\\epsilon}}\\right), O\\left(\\frac{\\sqrt{KT\\log(N)}\\log(KT)}{\\epsilon}\\right)$, and $\\tilde{O}\\left(\\frac{N^{1/6}K^{1/2}T^{2/3}\\log(NT)}{\\epsilon^{1/3}} + \\frac{N^{1/2}\\log(NT)}{\\epsilon}\\right)$, where $K$ and $N$ are the number of actions and experts respectively. These rates allow us to get sublinear regret for different combinations of small and large $K, N$ and $\\epsilon.$", "keywords": [ "Differential Privacy", "Bandits" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 2, 3 ] }, { "paper_id": "SbK9GtQlVg", "primary_area": "datasets and benchmarks", "n_author": 16, "author_info_list": [ { "name": "Bo Yang", "gender": "Male", "institution": "PengCheng Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Guoguo Chen", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Jianheng Zhuo", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "MS student" }, { "name": "Jinpeng Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Kai Yu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Ke Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Lecturer" }, { "name": "Mingyu Cui", "gender": "Male", "institution": "Chinese University of Hong Kong, The Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Shuai Fan", "gender": "Male", "institution": "AISpeech Ltd", "country": "CN", "position": "Researcher" }, { "name": "Wei-Qiang Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Xie Chen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xunying Liu", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Yexing Du", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "", "position": "PhD student" }, { "name": "Yifan Yang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zheshu Song", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Ziyang Ma", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Ziyuan Wang", "gender": "Male", "institution": "Birch Technology", "country": "", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 26, "n_ref": 42, "n_ref_all": 72, "n_fig": 3, "n_tab": 7, "L_tab": 3671, "n_element_tab": 561, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 4123, "n_element_tab_1": 245, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 146676, "len_all_1": 56357, "len_abs": 4331, "len_title": 196, "len_sents": 31453, "len_sents_1": 24636, "n_sents": 211, "n_sents_1": 159, "venue": "ICLR2025", "decision": "withdraw", "L_title": 147, "L_abs": 1690, "title": "GigaSpeech 2: An Evolving, Large-Scale and Multi-domain ASR Corpus for Low-Resource Languages with Automated Crawling, Transcription and Refinement", "abs": "The evolution of speech technology has been spurred by the rapid increase in dataset sizes. Traditional speech models generally depend on a large amount of labeled training data, which is scarce for low-resource languages. This paper presents GigaSpeech 2, a large-scale, multi-domain, multilingual speech recognition corpus. It is designed for low-resource languages and does not rely on paired speech and text data. GigaSpeech 2 comprises about 30,000 hours of automatically transcribed speech, including Thai, Indonesian, and Vietnamese, gathered from unlabeled YouTube videos. We also introduce an automated pipeline for data crawling, transcription, and label refinement. Specifically, this pipeline uses Whisper for initial transcription and TorchAudio for forced alignment, combined with multi-dimensional filtering for data quality assurance. A modified Noisy Student Training is developed to further refine flawed pseudo labels iteratively, thus enhancing model performance. Experimental results on our manually transcribed evaluation set and two public test sets from Common Voice and FLEURS confirm our corpus's high quality and broad applicability. Notably, ASR models trained on GigaSpeech 2 can reduce the word error rate for Thai, Indonesian, and Vietnamese on our challenging and realistic YouTube test set by 25% to 40% compared to the Whisper large-v3 model, with merely 10% model parameters. Furthermore, our ASR models trained on GigaSpeech 2 yield superior performance compared to commercial services. We believe that our newly introduced corpus and pipeline will open a new avenue for low-resource speech recognition and significantly facilitate research in this area.", "keywords": [ "multilingual", "low-resource language", "large-scale speech dataset", "speech recognition", "noisy student training" ], "rating_list": [ 3, 6, 6, 3, 1 ], "soundness_list": [ 2, 3, 3, 4, 3 ], "presentation_list": [ 2, 4, 3, 4, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 4, 5, 5, 4, 5 ] }, { "paper_id": "SaqU2ca367", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Iulia Duta", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "PhD" }, { "name": "Lucie Charlotte Magister", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Pietro Lio", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Shiye Su", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "MS student" } ], "n_formula": 20, "n_formula_1": 15, "n_ref_uni": 39, "n_ref": 77, "n_ref_all": 130, "n_fig": 17, "n_tab": 3, "L_tab": 1293, "n_element_tab": 238, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1432, "n_element_tab_1": 334, "formula_len_all": 1234, "formula_len_all_1": 946, "len_all": 157279, "len_all_1": 68443, "len_abs": 1030, "len_title": 81, "len_sents": 53376, "len_sents_1": 29737, "n_sents": 429, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1034, "title": "Explaining Hypergraph Neural Networks: From Local Explanations to Global Concepts", "abs": "Hypergraph neural networks are a class of powerful models that leverage the message passing paradigm to learn over hypergraphs,\na generalization of graphs well-suited to describing relational data with higher-order interactions. However, such models are not naturally interpretable, and their explainability has received very limited attention. We introduce SHypX, the first model-agnostic post-hoc explainer for hypergraph neural networks that provides both local and global explanations. At the instance-level, it performs input attribution by discretely sampling explanation subhypergraphs optimized to be faithful and concise. At the model-level, it produces global explanation subhypergraphs using unsupervised concept extraction. Extensive experiments across four real-world and four novel, synthetic hypergraph datasets demonstrate that our method finds high-quality explanations which can target a user-specified balance between faithfulness and concision, improving over baselines by 25 percent points in fidelity on average.", "keywords": [ "graph neural nets", "hypergraph neural nets", "explainability", "concepts" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 4, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "SabhfFUfA1", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Jiawei He", "gender": "Male", "institution": "Simon Fraser University", "country": "", "position": "PhD student" }, { "name": "Sicong Huang", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "", "position": "PhD student" }, { "name": "Kry Yik-Chau Lui", "gender": "Male", "institution": "Borealis AI", "country": "", "position": "Machine Learning Researcher" } ], "n_formula": 43, "n_formula_1": 30, "n_ref_uni": 46, "n_ref": 95, "n_ref_all": 109, "n_fig": 1, "n_tab": 4, "L_tab": 458, "n_element_tab": 69, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 806, "n_element_tab_1": 156, "formula_len_all": 2328, "formula_len_all_1": 1419, "len_all": 139722, "len_all_1": 64324, "len_abs": 922, "len_title": 111, "len_sents": 43560, "len_sents_1": 27864, "n_sents": 348, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 927, "title": "Inference, Fast and Slow: Reinterpreting VAEs for OOD Detection", "abs": "lthough likelihood-based methods are theoretically appealing, deep generative models (DGMs) often produce unreliable likelihood estimates in practice, particu larly for out-of-distribution (OOD) detection. We reinterpret variational autoen coders (VAEs) through the lens of fast and slow weights. Our approach is guided by the proposed Likelihood Path (LPath) Principle, which extends the classical likelihood principle. A critical decision in our method is the selection of statistics for classical density estimation algorithms. The sweet spot should contain just enough information that\u2019s sufficient for OOD detection but not too much to suffer from the curse of dimensionality. Our LPath principle achieves this by selecting the sufficient statistics that form the \"path\" toward the likelihood. We demonstrate that this likelihood path leads to SOTA OOD detection performance, even when the likelihood itself is unreliable.", "keywords": [ "unsupervised", "outlier detection", "ood", "out-of-distribution", "anomaly detection", "variational autoencoder", "VAE" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 1 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 2, 3, 2 ] }, { "paper_id": "SaOxhcDCM3", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Dominik Sobania", "gender": "unknown", "institution": "Johannes-Gutenberg Universit\u00e4t Mainz", "country": "", "position": "Postdoc" }, { "name": "Franz Rothlauf", "gender": "unknown", "institution": "University of Mainz", "country": "", "position": "Full Professor" }, { "name": "Martin Briesch", "gender": "Male", "institution": "University of Mainz", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 34, "n_ref": 80, "n_ref_all": 102, "n_fig": 14, "n_tab": 3, "L_tab": 640, "n_element_tab": 84, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 352, "formula_len_all_1": 261, "len_all": 145491, "len_all_1": 62954, "len_abs": 1234, "len_title": 147, "len_sents": 52897, "len_sents_1": 30920, "n_sents": 403, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1243, "title": "Large Language Models Suffer From Their Own Output: An Analysis of the Self-Consuming Training Loop", "abs": "Large Language Models (LLM) are already widely used to generate content for a variety of online platforms. As we are not able to safely distinguish LLM-generated content from human-produced content, LLM-generated content is used to train the next generation of LLMs, giving rise to a self-consuming training loop. From the image generation domain we know that such a self-consuming training loop reduces both quality and diversity of images finally ending in a model collapse. However, it is unclear whether this alarming effect can also be observed for LLMs. Therefore, we present the first study investigating the self-consuming training loop for LLMs. Further, we propose a novel method based on logic expressions that allows us to unambiguously verify the correctness of LLM-generated content, which is difficult for natural language text. We find that the self-consuming training loop produces correct outputs, however, the output declines in its diversity depending on the proportion of the used generated data. Fresh data can slow down this decline, but not stop it. Further, we observe similar results on a real natural language dataset. Given these concerning results, we encourage researchers to study methods to negate this process.", "keywords": [ "self-consuming training loop", "large language models", "model collapse", "generative models" ], "rating_list": [ 10, 5, 5, 5 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "SZpygmv3G1", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Andrew Gordon Wilson", "gender": "Not Specified", "institution": "New York University", "country": "US", "position": "Full Professor" }, { "name": "Ben Athiwaratkun", "gender": "Male", "institution": "Together AI", "country": "US", "position": "Staff AI Scientist" }, { "name": "Dylan Sam", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Marc Anton Finzi", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Nate Gruver", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Zico Kolter", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 0, "n_ref_uni": 23, "n_ref": 47, "n_ref_all": 64, "n_fig": 5, "n_tab": 5, "L_tab": 778, "n_element_tab": 56, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 325, "n_element_tab_1": 29, "formula_len_all": 207, "formula_len_all_1": 0, "len_all": 124915, "len_all_1": 69359, "len_abs": 1191, "len_title": 122, "len_sents": 43489, "len_sents_1": 35855, "n_sents": 295, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1196, "title": "The Promises and Pitfalls of Language Models for Structured Numerical Data", "abs": "Autoregressive language models are increasingly capable of processing non-text data, such as images or audio. Are language models also a natural choice for numerical data, such as the 3D structure of molecules? In this work, we use quantum chemistry simulations as a case study in the challenges of applying language models to numerical data, building up a set of simple subproblems that can shed light on key design decisions. We show that language models lag behind domain-specific models on prediction tasks and provide evidence for and against different hypotheses that explain their failure. Many commonly identified pitfalls such as difficulty performing arithmetic operations and choice of discrete vocabulary fall short of explaining the behavior. In contrast, we show that capturing invariance properties exhibits a strong correlation with predictive performance. Finally, we provide a comparison of language models trained from scratch on numerical data with models pretrained on text. We show that text pretraining often provides a surprisingly limited advantage on prediction tasks, and can even hurt performance, despite prior work showing that text-pretraining can offer advantages.", "keywords": [ "language models", "tokenization", "transformers", "inductive biases", "quantum chemistry" ], "rating_list": [ 8, 5, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "SZm3hxmksx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Bohan Zhai", "gender": "Male", "institution": "Snowflake", "country": "US", "position": "Researcher" }, { "name": "Chenfeng Xu", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Hongxia Yang", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Jianbo Yuan", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Quanzeng You", "gender": "Male", "institution": "ByteDance", "country": "US", "position": "Researcher" }, { "name": "Shijia Yang", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 37, "n_ref": 87, "n_ref_all": 105, "n_fig": 7, "n_tab": 6, "L_tab": 1894, "n_element_tab": 239, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 828, "n_element_tab_1": 25, "formula_len_all": 1338, "formula_len_all_1": 270, "len_all": 178812, "len_all_1": 61767, "len_abs": 2231, "len_title": 85, "len_sents": 39230, "len_sents_1": 26158, "n_sents": 286, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 760, "title": "Law of Vision Representation in MLLMs", "abs": "We present the \"Law of Vision Representation\" in multimodal large language models (MLLMs). It reveals a strong correlation between the combination of cross-modal alignment, correspondence in vision representation, and MLLM performance. We quantify the two factors using the cross-modal Alignment and Correspondence score (AC score). Through extensive experiments involving thirteen different vision representation settings and evaluations across eight benchmarks, we find that the AC score is linearly correlated to model performance. By leveraging this relationship, we are able to identify and train the optimal vision representation only, which does not require finetuning the language model every time, resulting in a 99.7% reduction in computational cost.", "keywords": [ "Multimodality Large Language Models; Computer Vision; Vision Representation" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "SYv9b4juom", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Dokwan Oh", "gender": "Male", "institution": "Samsung Advanced Institute of Technology", "country": "KR", "position": "Principal Researcher" }, { "name": "Jaehoon Oh", "gender": "Male", "institution": "Samsung Advanced Institute of Technology", "country": "KR", "position": "Researcher" }, { "name": "Seungjun Shin", "gender": "Male", "institution": "Samsung Advanced Institute of Technology", "country": "", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 51, "n_ref_all": 77, "n_fig": 12, "n_tab": 6, "L_tab": 4272, "n_element_tab": 469, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1867, "n_element_tab_1": 253, "formula_len_all": 886, "formula_len_all_1": 310, "len_all": 129742, "len_all_1": 61164, "len_abs": 1449, "len_title": 131, "len_sents": 37845, "len_sents_1": 26169, "n_sents": 303, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1455, "title": "OrthoRank: Token Selection via Sink Token Orthogonality for Efficient LLM inference", "abs": "Attention mechanisms are central to the success of large language models (LLMs), enabling them to capture intricate token dependencies and implicitly assign importance to each token. Recent studies have revealed the sink token, which receives disproportionately high attention despite their limited semantic role. In this paper, we first expand the relationship between the sink token and other tokens, moving beyond attention to explore their similarity in hidden states, considering the layer depth. We observe that as the layers get deeper, the cosine similarity between the normalized hidden states of the sink token and those of other tokens increases, and that the normalized hidden states of the sink token exhibit negligible changes. These imply that other tokens consistently are directed toward the sink token throughout the layers. Next, we propose a dynamic token selection method, called OrthoRank, using these findings to select important tokens. Specifically, in a certain layer, we define token importance by the speed at which the token moves toward the sink token. This is converted into orthogonality with the sink token, meaning that tokens that are more orthogonal to the sink token are assigned greater importance. Finally, through extensive experiments, we demonstrated that our method results in lower perplexity and higher zero-shot accuracy compared to layer pruning methods at the same sparsity ratio with comparable throughput.", "keywords": [ "large language model", "attention sink", "efficiency" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "SYnIf4LxAG", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Guangquan Zhang", "gender": "unknown", "institution": "University of Technology Sydney (UTS)", "country": "", "position": "Associate Professor" }, { "name": "Jie Lu", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "Professor" }, { "name": "Keqiuyin Li", "gender": "Female", "institution": "University of Technology Sydney", "country": "", "position": "Postdoc" }, { "name": "Ningyuan Zhang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Zhen Fang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 8, "n_ref_uni": 43, "n_ref": 81, "n_ref_all": 123, "n_fig": 8, "n_tab": 13, "L_tab": 12057, "n_element_tab": 1033, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 166, "n_element_tab_1": 19, "formula_len_all": 727, "formula_len_all_1": 427, "len_all": 219681, "len_all_1": 65375, "len_abs": 1844, "len_title": 115, "len_sents": 63371, "len_sents_1": 33914, "n_sents": 571, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1852, "title": "Release the Powers of Prompt Tuning: Cross-Modality Prompt Transfer", "abs": "Prompt Tuning adapts frozen models to new tasks by prepending a few learnable embeddings to the input.\nHowever, it struggles with tasks that suffer from data scarcity.\nTo address this, we explore Cross-Modality Prompt Transfer, leveraging prompts pretrained on a data-rich modality to improve performance on data-scarce tasks in another modality.\nAs a pioneering study, we first verify the feasibility of cross-modality prompt transfer by directly applying frozen source prompts (trained on the source modality) to the target modality task.\nTo empirically study cross-modality prompt transferability, we train a linear layer to adapt source prompts to the target modality, thereby boosting performance and providing ground-truth transfer results.\nRegarding estimating prompt transferability, existing methods show ineffectiveness in cross-modality scenarios where the gap between source and target tasks is larger.\nWe address this by decomposing the gap into the modality gap and the task gap, which we measure separately to estimate the prompt transferability more accurately.\nAdditionally, we propose Attention Transfer to further reduce the gaps by injecting target knowledge into the prompt and reorganizing a top-transferable source prompt using an attention block.\nWe conduct extensive experiments involving prompt transfer from 13 source language tasks to 19 target vision tasks under three settings.\nOur findings demonstrate that:\n(i) cross-modality prompt transfer is feasible, supported by in-depth analysis;\n(ii) measuring both the modality and task gaps is crucial for accurate prompt transferability estimation, a factor overlooked by previous studies;\n(iii) cross-modality prompt transfer can significantly release the powers of prompt tuning on data-scarce tasks, as evidenced by comparisons with a newly released prompt-based benchmark.", "keywords": [ "Cross-Modality", "Prompt Transfer" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "SYmUS6qRub", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Dario Shariatian", "gender": "unknown", "institution": "INRIA", "country": "FR", "position": "PhD student" }, { "name": "Umut Simsekli", "gender": "Male", "institution": "INRIA", "country": "", "position": "Research Faculty" }, { "name": "Alain Durmus", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "", "position": "Full Professor" } ], "n_formula": 114, "n_formula_1": 30, "n_ref_uni": 24, "n_ref": 88, "n_ref_all": 124, "n_fig": 3, "n_tab": 7, "L_tab": 1411, "n_element_tab": 190, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10919, "formula_len_all_1": 2375, "len_all": 232917, "len_all_1": 72550, "len_abs": 1559, "len_title": 35, "len_sents": 73945, "len_sents_1": 30298, "n_sents": 663, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 35, "L_abs": 1993, "title": "Denoising Levy Probabilistic Models", "abs": "Investigating noise distributions beyond Gaussian in diffusion generative models remains an open challenge. The Gaussian case has been a large success experimentally and theoretically, admitting a unified stochastic differential equation (SDE) framework, encompassing score-based and denoising formulations. Recent studies have investigated the potential of \\emph{heavy-tailed} noise distributions to mitigate mode collapse and effectively manage datasets exhibiting class imbalance, heavy tails, or prominent outliers. \nVery recently, Yoon et al.\\ (NeurIPS 2023), presented the Levy-Ito model (LIM), directly extending the SDE-based framework to a class of heavy-tailed SDEs, where the injected noise followed an $\\alpha$-stable distribution -- a rich class of heavy-tailed distributions. \nDespite its theoretical elegance and performance improvements, LIM relies on highly involved mathematical techniques, which may limit its accessibility and hinder its broader adoption and further development. \nIn this study, we take a step back, and instead of starting from the SDE formulation, we extend the denoising diffusion probabilistic model (DDPM) by directly replacing the Gaussian noise with $\\alpha$-stable noise. \nBy using only elementary proof techniques, we show that the proposed approach, \\emph{denoising L\\'{e}vy probabilistic model} (DLPM) algorithmically boils down to running vanilla DDPM with minor modifications, hence allowing the use of existing implementations with minimal changes. \nRemarkably, as opposed to the Gaussian case, DLPM and LIM yield different training algorithms and different backward processes, leading to distinct sampling algorithms. \nThis fundamental difference translates favorably for the performance of DLPM in various aspects: our experiments show that DLPM achieves better coverage of the tails of the data distribution, better generation of unbalanced datasets, and improved computation times requiring significantly smaller number of backward steps.", "keywords": [ "diffusion", "generative model", "deep learning", "machine learning", "heavy-tail" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "SYiOxXWlKU", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Ye Li", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "Associate Professor" }, { "name": "Yi Zhang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhongyi Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 63, "n_formula_1": 38, "n_ref_uni": 12, "n_ref": 25, "n_ref_all": 34, "n_fig": 6, "n_tab": 2, "L_tab": 811, "n_element_tab": 40, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2721, "formula_len_all_1": 1478, "len_all": 88442, "len_all_1": 53149, "len_abs": 1025, "len_title": 99, "len_sents": 25555, "len_sents_1": 21466, "n_sents": 258, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 99, "L_abs": 1030, "title": "EPINN: Physics-Informed Neural Network with exponential activation functions for solving stiff ODEs", "abs": "Solving stiff ordinary differential equations (ODEs) through machine learning methods has been quite a popular topic for years as it challenges the recently proposed physics-informed neural network (PINN). Many variations based on PINN have been advanced to enhance both the efficiency and the robustness. Nonetheless, many of them need to find the trade-off between the precision and speed because they have to train hundreds or even thousands of parameters if they do not design good or problem-adapt networks. In this scenario, we put forward a single layer physics-informed neural network with exponential activation functions (EPINN) by implementing the prior knowledge of the solution to the linear stiff ODEs. Under this simple but useful structure, less parameters \nwould be sufficient and the model is easy to train. The model is also extended to solve nonlinear systems by introducing sequential EPINN. The network is tested on six benchmark problems including both linear and nonlinear ones and shows great performance.", "keywords": [ "EPINN", "stiff ODEs", "PINN" ], "rating_list": [ 1, 3, 1, 5 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 1, 1, 1, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "SYI409tbsv", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 7, "author_info_list": [ { "name": "Haishuai Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Research Professor" }, { "name": "Jiajun Bu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Jiawei Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Ming Gu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Qiaoyu Tan", "gender": "Male", "institution": "New York University Shanghai", "country": "CN", "position": "Assistant Professor" }, { "name": "Sheng Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Zeyu Fang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" } ], "n_formula": 44, "n_formula_1": 24, "n_ref_uni": 28, "n_ref": 39, "n_ref_all": 62, "n_fig": 6, "n_tab": 11, "L_tab": 10998, "n_element_tab": 1544, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2389, "n_element_tab_1": 397, "formula_len_all": 2656, "formula_len_all_1": 1674, "len_all": 180252, "len_all_1": 60470, "len_abs": 1429, "len_title": 113, "len_sents": 44207, "len_sents_1": 24150, "n_sents": 320, "n_sents_1": 163, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1434, "title": "Towards a Unified Framework of Clustering-based Anomaly Detection", "abs": "Unsupervised Anomaly Detection (UAD) plays a crucial role in identifying abnormal patterns within data without labeled examples, holding significant practical implications across various domains. Although the individual contributions of representation learning and clustering to anomaly detection are well-established, their interdependencies remain under-explored due to the absence of a unified theoretical framework. Consequently, their collective potential to enhance anomaly detection performance remains largely untapped. To bridge this gap, in this paper, we propose a novel probabilistic mixture model for anomaly detection to establish a theoretical connection among representation learning, clustering, and anomaly detection. By maximizing a novel anomaly-aware data likelihood, representation learning and clustering can effectively reduce the adverse impact of anomalous data and collaboratively benefit anomaly detection. Meanwhile, a theoretically substantiated anomaly score is naturally derived from this framework. Lastly, drawing inspiration from gravitational analysis in physics, we have devised an improved anomaly score that more effectively harnesses the combined power of representation learning and clustering. Extensive experiments, involving 17 baseline methods across 30 diverse datasets, validate the effectiveness and generalization capability of the proposed method, surpassing state-of-the-art methods.", "keywords": [ "Anomaly Detection", "Clustering" ], "rating_list": [ 3, 6, 3, 6, 5 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 4, 2 ], "contribution_list": [ 2, 2, 1, 3, 2 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "SXyUF6RVmT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Chao Huang", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Da Luo", "gender": "Male", "institution": "South China University of Technology", "country": "", "position": "MS student" }, { "name": "Lianghao Xia", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Yangqin Jiang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yuhao Yang", "gender": "Not Specified", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "konyellin", "gender": "Female", "institution": "WeChat Open Platform Department", "country": "", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 11, "n_ref": 22, "n_ref_all": 39, "n_fig": 6, "n_tab": 7, "L_tab": 4928, "n_element_tab": 253, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 3171, "n_element_tab_1": 178, "formula_len_all": 692, "formula_len_all_1": 776, "len_all": 155678, "len_all_1": 74731, "len_abs": 1247, "len_title": 115, "len_sents": 58428, "len_sents_1": 33309, "n_sents": 448, "n_sents_1": 269, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1159, "title": "RecLM: Recommendation Instruction Tuning with Large Language Models", "abs": "Recommender systems aim to deeply understand users' complex preferences based on their past interactions. Deep collaborative filtering paradigms, leveraging advanced neural architectures like Graph Neural Networks (GNNs), excel at capturing collaborative relationships among users. However, limitations emerge when dealing with sparse data or zero-shot learning from unseen datasets, due to the design constraints of ID-based embedding functions in existing solutions. These challenges hinder robust generalization and adaptability. To address this, we propose a model-agnostic recommendation instruction-tuning paradigm that integrates large language models with collaborative filtering. Our Recommendation Language Model (RecLM) is introduced to enhance the capability of capturing user preference diversity. We design a reinforcement learning reward function to facilitate self-augmentation of our language models. Comprehensive evaluations demonstrate significant advantages of our approach across various settings. It can be integrated as a plug-and-play component with state-of-the-art recommender systems, resulting in notable performance enhancements.", "keywords": [ "Large Language Model", "Recommendation Systems", "Instruction-tuning" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 3, 1, 3, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "SXvb8PS4Ud", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Dong Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Distinguished Scientist" }, { "name": "Hongming Zhang", "gender": "Male", "institution": "Tencent AI Lab Seattle", "country": "US", "position": "Researcher" }, { "name": "Siru Ouyang", "gender": "Female", "institution": "University of Illinois Urbana-Champaign Champaign", "country": "US", "position": "PhD student" }, { "name": "Tao Ge", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Principal Researcher" }, { "name": "Zilin Xiao", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Vicente Ordonez", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 45, "n_ref": 102, "n_ref_all": 118, "n_fig": 4, "n_tab": 3, "L_tab": 3148, "n_element_tab": 391, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 3282, "n_element_tab_1": 401, "formula_len_all": 460, "formula_len_all_1": 385, "len_all": 171500, "len_all_1": 65459, "len_abs": 1216, "len_title": 113, "len_sents": 35987, "len_sents_1": 27414, "n_sents": 268, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1230, "title": "ParallelSpec: Parallel Drafter for Efficient Speculative Decoding", "abs": "Speculative decoding has proven to be an efficient solution to large language model (LLM) inference, where the small drafter predicts future tokens at a low cost, and the target model is leveraged to verify them in parallel. However, most existing works still draft tokens auto-regressively to maintain sequential dependency in language modeling, which we consider a huge computational burden in speculative decoding. We present ParallelSpec, an alternative to auto-regressive drafting strategies in state-of-the-art speculative decoding approaches. In contrast to auto-regressive drafting in the speculative stage, we train a parallel drafter to serve as an efficient speculative model. ParallelSpec learns to efficiently predict multiple future tokens in parallel using a single model, and it can be integrated into any speculative decoding framework that requires aligning the output distributions of the drafter and the target model with minimal training cost. Experimental results show that ParallelSpec accelerates baseline methods in latency up to 62% for Medusa and 9-17\\% for EAGLE on text generation benchmarks, and it achieves 2.84$\\times$ overall speedup on the Llama-2-13B model using third-party evaluation criteria.", "keywords": [ "large language model inference", "speculative decoding" ], "rating_list": [ 8, 5, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 4, 3 ], "contribution_list": [ 4, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 5, 3 ] }, { "paper_id": "SXtl7NRyE5", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Atsutoshi Kumagai", "gender": "Male", "institution": "NTT", "country": "JP", "position": "Researcher" }, { "name": "Kazuki Adachi", "gender": "Male", "institution": "Yokohama National University", "country": "JP", "position": "PhD student" }, { "name": "Shin'ya Yamaguchi", "gender": "Male", "institution": "NTT", "country": "JP", "position": "Associate Distinguished Researcher" }, { "name": "Tomoki Hamagami", "gender": "Male", "institution": "Yokohama National University, Tokyo Institute of Technology", "country": "JP", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 36, "n_ref": 119, "n_ref_all": 169, "n_fig": 6, "n_tab": 15, "L_tab": 4125, "n_element_tab": 512, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 2882, "n_element_tab_1": 322, "formula_len_all": 903, "formula_len_all_1": 845, "len_all": 183711, "len_all_1": 76291, "len_abs": 1496, "len_title": 105, "len_sents": 50656, "len_sents_1": 29379, "n_sents": 440, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1504, "title": "Test-time Adaptation for Regression by Subspace Alignment", "abs": "This paper investigates test-time adaptation (TTA) for regression, where a regression model pre-trained in a source domain is adapted to an unknown target distribution with unlabeled target data.\nAlthough regression is one of the fundamental tasks in machine learning, most of the existing TTA methods have classification-specific designs, which assume that models output class-categorical predictions, whereas regression models typically output only single scalar values.\nTo enable TTA for regression, we adopt a feature alignment approach, which aligns the feature distributions between the source and target domains to mitigate the domain gap.\nHowever, we found that naive feature alignment employed in existing TTA methods for classification is ineffective or even worse for regression because the features are distributed in a small subspace and many of the raw feature dimensions have little significance to the output.\nFor an effective feature alignment in TTA for regression, we propose Significant-subspace Alignment (SSA).\nSSA consists of two components: subspace detection and dimension weighting.\nSubspace detection finds the feature subspace that is representative and significant to the output.\nThen, the feature alignment is performed in the subspace during TTA.\nMeanwhile, dimension weighting raises the importance of the dimensions of the feature subspace that have greater significance to the output.\nWe experimentally show that SSA outperforms various baselines on real-world datasets.", "keywords": [ "Test-time adaptation", "regression", "distribution shift", "deep learning" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "SXopqmHJO1", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Bart Vandereycken", "gender": "unknown", "institution": "University of Geneva, Switzerland", "country": "", "position": "Associate Professor" }, { "name": "Foivos Alimisis", "gender": "Male", "institution": "University of Geneva, Switzerland", "country": "", "position": "PhD student" } ], "n_formula": 59, "n_formula_1": 51, "n_ref_uni": 18, "n_ref": 38, "n_ref_all": 52, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4194, "formula_len_all_1": 3624, "len_all": 78496, "len_all_1": 55861, "len_abs": 467, "len_title": 161, "len_sents": 24348, "len_sents_1": 19795, "n_sents": 227, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 112, "L_abs": 466, "title": "Characterizing linear convergence in optimization: Polyak-\u0141ojasiewicz inequality and weak-quasi-strong-convexity", "abs": "We give a complete characterization of optimization problems that can be solved by gradient descent with a linear convergence rate. We show that the well-known Polyak-\u0141ojasiewicz inequality is necessary and sufficient for linear convergence with respect to function values to the minimum, while a property that we call \"weak-quasi-strong-convexity\", or WQSC, is necessary and sufficient for linear convergence with respect to distances of the iterates to an optimum.", "keywords": [ "gradient descent", "linear convergence", "Polyak-\u0141ojasiewicz inequality", "weak-quasi-strong convexity" ], "rating_list": [ 6, 3, 3, 8 ], "soundness_list": [ 4, 1, 2, 4 ], "presentation_list": [ 2, 2, 2, 4 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "SXB9LnJ0SK", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Bolin Zhu", "gender": "Male", "institution": "nanjing university", "country": "", "position": "MS student" }, { "name": "Gong Cheng", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Sichen Liu", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "MS student" }, { "name": "Xiao Li", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "PhD student" }, { "name": "Yin Zhu", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Yiwei liu", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Undergrad student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 34, "n_ref": 57, "n_ref_all": 79, "n_fig": 8, "n_tab": 7, "L_tab": 2049, "n_element_tab": 171, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 793, "n_element_tab_1": 67, "formula_len_all": 91, "formula_len_all_1": 91, "len_all": 163704, "len_all_1": 64770, "len_abs": 1012, "len_title": 113, "len_sents": 45053, "len_sents_1": 28358, "n_sents": 366, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1015, "title": "FormulaReasoning: A Dataset for Formula-Based Numerical Reasoning", "abs": "The application of formulas is a fundamental ability of humans when addressing numerical reasoning problems. However, existing numerical reasoning datasets seldom indicate explicitly the formulas employed during the reasoning steps. To bridge this gap, we construct a dataset for formula-based numerical reasoning called FormulaReasoning, which consists of 5,420 reasoning-based questions. We employ it to conduct evaluations of LLMs with size ranging from 7B to over 100B parameters utilizing zero-shot and few-shot chain-of-thought methods, and we further explore using retrieval-augmented LLMs provided with an external formula database associated with our dataset. We also experiment with supervised methods where we divide the reasoning process into formula generation, parameter extraction, and numerical calculation, and perform data augmentation. Our empirical findings underscore the significant potential for improvement in existing models when applied to our challenging, formula-driven FormulaReasoning.", "keywords": [ "formula", "numerical reasoning", "question answering" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "SX2Z5tgiUu", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "De Wen Soh", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" }, { "name": "Haobo Jiang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "Hossein Rahmani", "gender": "Male", "institution": "Lancaster University", "country": "GB", "position": "Full Professor" }, { "name": "Jiayi Yuan", "gender": "Female", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "PhD student" }, { "name": "Jun Liu", "gender": "Male", "institution": "Lancaster University", "country": "GB", "position": "Full Professor" }, { "name": "Li Xu", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 42, "n_ref": 95, "n_ref_all": 109, "n_fig": 5, "n_tab": 4, "L_tab": 1859, "n_element_tab": 187, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 153, "n_element_tab_1": 21, "formula_len_all": 1132, "formula_len_all_1": 1002, "len_all": 141110, "len_all_1": 64177, "len_abs": 1835, "len_title": 124, "len_sents": 40698, "len_sents_1": 32934, "n_sents": 268, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1180, "title": "PrivateChat: A Secure Encrypted Communication Framework with Black-box LLMs", "abs": "With the growing applications of large language models (LLMs), privacy leakage has emerged as a significant concern. However, widely used LLMs are often deployed on cloud platforms and accessible only through relatively expensive API calls, complicating the realization of secure communication between users and cloud LLMs. In this paper, we introduce PrivateChat, a novel private communication framework that enables users to safely interact with cloud LLMs using user-customized encryption methods (e.g., AES). Our core idea is to learn a private system prompt, which instructs the cloud LLM to process and respond in encrypted text while concealing encryption details from potential attackers. Additionally, to optimize such prompts with few API calls, we propose a Sample-Efficient Simultaneous Perturbation Stochastic Approximation (SE-SPSA) black-box optimization algorithm, which incorporates a baseline-based variance reduction strategy with SPSA for effective and economical training. Extensive experiments on several benchmark datasets with various encryption methods show the effectiveness of our approach in achieving secure and reliable communication with cloud LLMs.", "keywords": [ "cloud LLM privacy protection", "encrypted communication", "black-box optimization" ], "rating_list": [ 3, 1, 6, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "SWs8CIdQ33", "primary_area": "reinforcement learning", "n_author": 10, "author_info_list": [ { "name": "Guoliang Fan", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Jiang Bian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Lei Song", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Lijuan Li", "gender": "Female", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Researcher" }, { "name": "Wei Jiang", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "", "position": "MS student" }, { "name": "Xianliang Yang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Yifan Xia", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Yuanyu Zhang", "gender": "Male", "institution": "Guizhou University", "country": "", "position": "MS student" }, { "name": "Zhihao Liu", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Zichuan Liu", "gender": "unknown", "institution": "Microsoft Research", "country": "CN", "position": "Intern" } ], "n_formula": 49, "n_formula_1": 9, "n_ref_uni": 40, "n_ref": 58, "n_ref_all": 92, "n_fig": 6, "n_tab": 23, "L_tab": 6866, "n_element_tab": 661, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1664, "n_element_tab_1": 90, "formula_len_all": 4474, "formula_len_all_1": 1079, "len_all": 235308, "len_all_1": 63702, "len_abs": 1746, "len_title": 167, "len_sents": 67314, "len_sents_1": 28721, "n_sents": 551, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 119, "L_abs": 1772, "title": "Knowing What Not to Do: Leverage Language Model Insights for Action Space Pruning in Multi-agent Reinforcement Learning", "abs": "Multi-agent reinforcement learning (MARL) is employed to develop autonomous agents that can learn to adopt cooperative or competitive strategies within complex environments. However, the linear increase in the number of agents leads to a combinatorial explosion of the action space, which always results in algorithmic instability, difficulty in convergence, or entrapment in local optima. While researchers have designed a variety of effective algorithms to compress the action space, these methods also introduce new challenges, such as the need for manually designed prior knowledge or reliance on the structure of the problem, which diminishes the applicability of these techniques. In this paper, we introduce **E**volutionary action **SPA**ce **R**eduction with **K**nowledge (eSpark), an exploration function generation framework driven by large language models (LLMs) to boost exploration and prune unnecessary actions in MARL. Using just a basic prompt that outlines the overall task and setting, eSpark is capable of generating exploration functions in a zero-shot manner, identifying and pruning redundant or irrelevant state-action pairs, and then achieving autonomous improvement from policy feedback. In reinforcement learning tasks involving inventory management and traffic light control encompassing a total of 15 scenarios, eSpark consistently outperforms the combined MARL algorithm in all scenarios, achieving an average performance gain of 34.4% and 9.9% in the two types of tasks respectively. Additionally, eSpark has proven to be capable of managing situations with a large number of agents, securing a 29.7% improvement in scalability challenges that featured over 500 agents. The code can be found in https://anonymous.4open.science/r/0CDH-0DF8/.", "keywords": [ "Multi-agent reinforcement learning", "Action space pruning", "Exploration", "Coding large language model" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "SWg72N2ky1", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Aleksandr Lobanov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "", "position": "Researcher" } ], "n_formula": 83, "n_formula_1": 32, "n_ref_uni": 21, "n_ref": 39, "n_ref_all": 56, "n_fig": 5, "n_tab": 1, "L_tab": 91, "n_element_tab": 6, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 91, "n_element_tab_1": 6, "formula_len_all": 13029, "formula_len_all_1": 3996, "len_all": 113642, "len_all_1": 55451, "len_abs": 1103, "len_title": 132, "len_sents": 31664, "len_sents_1": 23252, "n_sents": 258, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1152, "title": "Maximum Noise Level as Third Optimality Criterion in Black-box Optimization Problem", "abs": "This paper is devoted to the study (common in many applications) of the black-box optimization problem, where the black-box represents a gradient-free oracle $\\tilde{f}_p = f(x) + \\xi_p$ providing the objective function value with some stochastic noise. Assuming that the objective function is $\\mu$-strongly convex, and also not just $L$-smooth, but has a higher order of smoothness ($\\beta \\geq 2$) we provide a novel optimization method: _Zero-Order Accelerated Batched Stochastic Gradient Descent_, whose theoretical analysis closes the question regarding the iteration complexity, _achieving optimal estimates_. Moreover, we provide a thorough analysis of the maximum noise level, and show under which condition the maximum noise level will take into account information about batch size $B$ as well as information about the smoothness order of the function $\\beta$. Finally, we show the importance of considering the maximum noise level $\\Delta$ as a third optimality criterion along with the standard two on the example of a numerical experiment of interest to the machine learning community, where we compare with SOTA gradient-free algorithms.", "keywords": [ "Black-box optimization", "Higher order smoothness function", "Strongly convex optimization", "Maximum noise level" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 3, 3, 3, 1 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "SWEqzy7IQB", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Jingrong Wei", "gender": "Female", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Long Chen", "gender": "Male", "institution": "University of California, Irvine", "country": "", "position": "Full Professor" } ], "n_formula": 116, "n_formula_1": 62, "n_ref_uni": 55, "n_ref": 90, "n_ref_all": 130, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8586, "formula_len_all_1": 4607, "len_all": 166397, "len_all_1": 63139, "len_abs": 661, "len_title": 162, "len_sents": 41837, "len_sents_1": 23161, "n_sents": 487, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 113, "L_abs": 662, "title": "Accelerated Over-Relaxation Heavy-Ball Method: Achieving Global Accelerated Convergence with Broad Generalization", "abs": "The heavy-ball momentum method accelerates gradient descent with a momentum term but lacks accelerated convergence for general smooth strongly convex problems. This work introduces the Accelerated Over-Relaxation Heavy-Ball (AOR-HB) method, the first variant with provable global and accelerated convergence for such problems. AOR-HB closes a long-standing theoretical gap, extends to composite convex optimization and min-max problems, and achieves optimal complexity bounds. It offers three key advantages: (1) broad generalization ability, (2) potential to reshape acceleration techniques, and (3) conceptual clarity and elegance compared to existing methods.", "keywords": [ "Optimization theory", "heavy ball", "momentum", "acceleration", "composite optimization", "min-max", "saddle point problem" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "SW6IIpPr9I", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Chaewon Lee", "gender": "Female", "institution": "Korea University", "country": "", "position": "MS student" }, { "name": "Seon-Ho Lee", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" }, { "name": "Chang-su Kim", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Full Professor" } ], "n_formula": 34, "n_formula_1": 15, "n_ref_uni": 36, "n_ref": 82, "n_ref_all": 110, "n_fig": 7, "n_tab": 21, "L_tab": 4321, "n_element_tab": 643, "n_fig_1": 7, "n_tab_1": 8, "L_tab_1": 2507, "n_element_tab_1": 338, "formula_len_all": 1963, "formula_len_all_1": 634, "len_all": 133010, "len_all_1": 58589, "len_abs": 1173, "len_title": 122, "len_sents": 33688, "len_sents_1": 21896, "n_sents": 333, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 923, "title": "Stochastic Order Learning: An Approach to Rank Estimation Using Noisy Data", "abs": "A novel algorithm, called stochastic order learning (SOL), for reliable rank estimation in the presence of label noise is proposed in this paper. For noise-robust rank estimation, we first represent label errors as random variables. We then formulate a desideratum that encourages reducing the dissimilarity of an instance from its stochastically related centroids. Based on this desideratum, we develop two loss functions: discriminative loss and stochastic order loss. Employing these two losses, we train a network to construct an embedding space in which instances are arranged according to their ranks. Also, after teaching the network, we identify outliers, which are likely to have extreme label errors, and relabel them for data refinement. Extensive experiments on various benchmark datasets demonstrate that the proposed SOL algorithm yields decent rank estimation results even when labels are corrupted by noise.", "keywords": [ "rank estimation", "label noise", "order learning" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "SVta2eQNt3", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Bastian Wandt", "gender": "Male", "institution": "Link\u00f6ping University", "country": "SE", "position": "Assistant Professor" }, { "name": "Chunjin Song", "gender": "Female", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Shih-Yang Su", "gender": "Male", "institution": "University of British Columbia", "country": "", "position": "PhD student" }, { "name": "Zhijie Wu", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Intern" }, { "name": "Helge Rhodin", "gender": "Male", "institution": "Universit\u00e4t Bielefeld", "country": "DE", "position": "Full Professor" }, { "name": "Leonid Sigal", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Full Professor" } ], "n_formula": 24, "n_formula_1": 19, "n_ref_uni": 55, "n_ref": 173, "n_ref_all": 191, "n_fig": 10, "n_tab": 10, "L_tab": 2972, "n_element_tab": 230, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1274, "n_element_tab_1": 81, "formula_len_all": 1004, "formula_len_all_1": 744, "len_all": 188065, "len_all_1": 63281, "len_abs": 1166, "len_title": 85, "len_sents": 56929, "len_sents_1": 29209, "n_sents": 395, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 37, "L_abs": 1159, "title": "Locality Sensitive Avatars From Video", "abs": "We present locality-sensitive avatar, a neural radiance field (NeRF) based network to learn human motions from monocular videos. To this end, we estimate a canonical representation between different frames of a video with a non-linear mapping from observation to canonical space, which we decompose into a skeletal rigid motion and a non-rigid counterpart. Our key contribution is to retain fine-grained details by modeling the non-rigid part with a graph neural network (GNN) that keeps the pose information local to neighboring body parts.\nCompared to former canonical representation based methods which solely operate on the coordinate space of a whole shape, our locality-sensitive motion modeling can reproduce both realistic shape contours and vivid fine-grained details. We evaluate on ZJU-MoCap, ActorsHQ, SynWild, and various outdoor videos. The experiments reveal that with the locality sensitive deformation to canonical feature space, we are the first to achieve state-of-the-art results across novel view synthesis, novel pose animation and 3D shape reconstruction simultaneously. For reproducibility, the code will be available upon publication.", "keywords": [ "3D Computer Vision", "Neural Rendering", "Avatar Modeling" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "SVd9Ffcdp8", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Aranyak Mehta", "gender": "unknown", "institution": "Google Research", "country": "US", "position": "Researcher" }, { "name": "David C. Parkes", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Di Wang", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Manzil Zaheer", "gender": "Male", "institution": "Zaheer", "country": "", "position": "Researcher" }, { "name": "Sai Srivatsa Ravindranath", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Zhe Feng", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 14, "n_ref_uni": 39, "n_ref": 58, "n_ref_all": 66, "n_fig": 2, "n_tab": 4, "L_tab": 989, "n_element_tab": 139, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 351, "n_element_tab_1": 60, "formula_len_all": 1710, "formula_len_all_1": 1253, "len_all": 142499, "len_all_1": 77200, "len_abs": 1366, "len_title": 113, "len_sents": 41458, "len_sents_1": 33789, "n_sents": 354, "n_sents_1": 281, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1371, "title": "Deep Reinforcement Learning for Sequential Combinatorial Auctions", "abs": "Revenue-optimal auction design is a challenging problem with significant theoretical and practical implications. Sequential auction mechanisms, known for their simplicity and strong strategyproofness guarantees, are often limited by theoretical results that are largely existential, except for certain restrictive settings. Although traditional reinforcement learning methods such as Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC) are applicable in this domain, they struggle with computational demands and convergence issues when dealing with large and continuous action spaces. In light of this and recognizing that we can model transitions differentiable for our settings, we propose using a new reinforcement learning framework tailored for sequential combinatorial auctions that leverages first-order gradients. Our extensive evaluations show that our approach achieves significant improvement in revenue over both analytical baselines and standard reinforcement learning algorithms. Furthermore, we scale our approach to scenarios involving up to 50 agents and 50 items, demonstrating its applicability in complex, real-world auction settings. As such, this work advances the computational tools available for auction design and contributes to bridging the gap between theoretical results and practical implementations in sequential auction design.", "keywords": [ "Mechanism Design", "Auctions", "Game Theory", "Differential Economics", "Reinforcement Learning", "Deep Learning", "Market Design", "AI for Economics" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 2, 2, 4 ] }, { "paper_id": "SVRRQ8goQo", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Ge Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Haoran Zhang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Intern" }, { "name": "Liu Jiaheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jian Yang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Kaijing Ma", "gender": "Female", "institution": "Tongji University", "country": "", "position": "Undergrad student" }, { "name": "Wenhao Huang", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Xiang Yue", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Xingwei Qu", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "PhD student" }, { "name": "ZhoufutuWen", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Minghao Liu", "gender": "Male", "institution": "2077AI", "country": "SG", "position": "Researcher" }, { "name": "Xinrun Du", "gender": "Male", "institution": "01.AI", "country": "CN", "position": "Intern" }, { "name": "Wang Yunran", "gender": "Female", "institution": "\u00c9cole Polytechnique", "country": "FR", "position": "MS student" } ], "n_formula": 26, "n_formula_1": 1, "n_ref_uni": 23, "n_ref": 93, "n_ref_all": 136, "n_fig": 8, "n_tab": 20, "L_tab": 20977, "n_element_tab": 1207, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 247, "n_element_tab_1": 38, "formula_len_all": 2064, "formula_len_all_1": 69, "len_all": 457565, "len_all_1": 54324, "len_abs": 1460, "len_title": 127, "len_sents": 116541, "len_sents_1": 24891, "n_sents": 1159, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1548, "title": "KOR-Bench: Benchmarking Language Models on Knowledge-Orthogonal Reasoning Tasks", "abs": "In this paper, we introduce the concept of Knowledge-Orthogonal Reasoning(KOR),\nwhere knowledge orthogonality refers to the independence from existing pretrained knowledge. \nBy introducing new rules that are orthogonal to the pretrained knowledge, we minimize its interference to achieve a more accurate evaluation of the model's intrinsic reasoning and planning abilities. \nBased on this concept, we propose the Knowledge-Orthogonal Reasoning Benchmark (KOR-Bench), which includes five task categories: Operation, Logic, Cipher, Puzzle, and Counterfactual.\nKOR-Bench focuses on assessing how well models apply new rule descriptions to solve new rule-driven questions. \nThis challenging benchmark shows that leading models like Claude-3.5-Sonnet and GPT-4o achieve only 58.96\\% and 58.00\\%, respectively. \nWe conduct thorough analyses using Stepwise Prompting to identify bottlenecks in Cipher task. Self-correction experiments indicate that two rounds of correction usually result in the best performance. Complex Task Processing evaluates the model's performance across three integrated task settings. Additionally, we analyze the impact of Tricks on puzzle task and visualize rule-focused attention. Our goal is for \\our{} to serve as a valuable tool for evaluating and enhancing the reasoning abilities of models, while also fostering further research and development in this field. All data, inference, evaluation code, and experimental results are available here\\footnote{\\url{https://anonymous.4open.science/r/kor-bench-rebuttal-repo-44F6}}.", "keywords": [ "Reasoning; Knowledge-Orthogonal; Rule-Based" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "SUc1UOWndp", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Daniel Murfet", "gender": "Male", "institution": "The University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "George Wang", "gender": "Male", "institution": "Timaeus", "country": "US", "position": "Researcher" }, { "name": "Jesse Hoogland", "gender": "Male", "institution": "Timaeus", "country": "US", "position": "Researcher" }, { "name": "Stan van Wingerden", "gender": "Male", "institution": "Timaeus", "country": "US", "position": "Researcher" }, { "name": "Zach Furman", "gender": "unknown", "institution": "Boston University, Boston University", "country": "", "position": "Undergrad student" } ], "n_formula": 14, "n_formula_1": 7, "n_ref_uni": 77, "n_ref": 174, "n_ref_all": 247, "n_fig": 23, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 733, "n_element_tab_1": 47, "formula_len_all": 860, "formula_len_all_1": 480, "len_all": 310832, "len_all_1": 60913, "len_abs": 1222, "len_title": 146, "len_sents": 108995, "len_sents_1": 29327, "n_sents": 786, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 96, "L_abs": 1305, "title": "Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient", "abs": "We introduce refined variants of the Local Learning Coefficient (LLC), a measure of model complexity grounded in singular learning theory, to study the development of internal structure in transformer language models during training. By applying these refined LLCs (rLLCs) to individual components of a two-layer attention-only transformer, we gain novel insights into the progressive differentiation and specialization of attention heads. Our methodology reveals how attention heads differentiate into distinct functional roles over the course of training, analyzes the types of data these heads specialize to process, and discovers a previously unidentified multigram circuit. These findings demonstrate that rLLCs provide a principled, quantitative toolkit for developmental interpretability, which aims to understand models through their evolution across the learning process. This work advances the field of developmental interpretability by providing a mathematically rigorous approach to understanding neural networks through the lens of their learning process. More broadly, this work takes a step towards establishing the correspondence between data distributional structure, geometric properties of the loss landscape, learning dynamics, and emergent computational structures in neural networks.", "keywords": [ "Developmental Interpretability", "Mechanistic Interpretability", "Singular Learning Theory", "Learning Dynamics", "Stagewise development", "Model complexity" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 2, 4, 2, 3 ] }, { "paper_id": "SUL5L9rIyF", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Donglin Wang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Associate Professor" }, { "name": "Jingzehua Xu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Shuai Zhang", "gender": "Male", "institution": "New Jersey Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Yiyuan Yang", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "", "position": "PhD student" }, { "name": "Yubo Huang", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Zifeng Zhuang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Ziqi Zhang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Micro Masters Student" } ], "n_formula": 30, "n_formula_1": 21, "n_ref_uni": 36, "n_ref": 74, "n_ref_all": 103, "n_fig": 7, "n_tab": 8, "L_tab": 3851, "n_element_tab": 411, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2323, "n_element_tab_1": 305, "formula_len_all": 2790, "formula_len_all_1": 1683, "len_all": 167184, "len_all_1": 71973, "len_abs": 1619, "len_title": 142, "len_sents": 42300, "len_sents_1": 29741, "n_sents": 319, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1639, "title": "Distribution Corrected Estimation via Adversarial Density Weighted Regression", "abs": "We propose a novel one-step supervised imitation learning (IL) framework called Adversarial Density Regression (ADR). This IL framework aims to correct the policy learned on unknown-quality to match the expert distribution by utilizing demonstrations, without relying on the Bellman operator. Specifically, ADR addresses several limitations in previous IL algorithms: First, most IL algorithms are based on the Bellman operator, which inevitably suffer from cumulative offsets from sub-optimal rewards during multi-step update processes. Additionally, off-policy training frameworks suffer from Out-of-Distribution (OOD) state-actions. Second, while conservative terms help solve the OOD issue, balancing the conservative term is difficult. To address these limitations, we fully integrate a one-step density-weighted Behavioral Cloning (BC) objective for IL with auxiliary imperfect demonstration. Theoretically, we demonstrate that this adaptation can effectively correct the distribution of policies trained on unknown-quality datasets to align with the expert policy's distribution. Moreover, the difference between the empirical and the optimal value function is proportional to the upper bound of ADR's objective, indicating that minimizing ADR's objective is akin to approaching the optimal value. Experimentally, we validated the performance of ADR by conducting extensive evaluations. Specifically, ADR outperforms all of the selected IL algorithms on tasks from the Gym-Mujoco domain. Meanwhile, it achieves an \\textbf{89.5\\%} improvement over IQL when utilizing ground truth rewards on tasks from the Adroit and Kitchen domains.", "keywords": [ "Imitation Learning", "Deep Reinforcement Learning" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "SUEXRbzq9l", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "A. Pavan", "gender": "unknown", "institution": "Iowa State University", "country": "US", "position": "Full Professor" }, { "name": "Arnab Bhattacharyya", "gender": "Male", "institution": "The University of Warwick", "country": "GB", "position": "Associate Professor" }, { "name": "Dimitrios Myrisiotis", "gender": "unknown", "institution": "CNRS@CREATE LTD.", "country": "", "position": "Postdoc" }, { "name": "Kuldeep S. Meel", "gender": "Male", "institution": "Department of Computer Science", "country": "", "position": "Associate Professor" }, { "name": "Sutanu Gayen", "gender": "Male", "institution": "Indian Institute of Technology, Kanpur", "country": "", "position": "Assistant Professor" }, { "name": "N.V. Vinodchandran", "gender": "Male", "institution": "University of Nebraska, Lincoln", "country": "US", "position": "Professor" } ], "n_formula": 40, "n_formula_1": 33, "n_ref_uni": 25, "n_ref": 41, "n_ref_all": 44, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4901, "formula_len_all_1": 4261, "len_all": 92754, "len_all_1": 61298, "len_abs": 1376, "len_title": 111, "len_sents": 27310, "len_sents_1": 23026, "n_sents": 299, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1445, "title": "Estimating Statistical Similarity Between Product Distributions", "abs": "We investigate the problem of computing the *statistical* or *total variation (TV) similarity* between distributions $P$ and $Q$, which is defined as $s_{\\mathrm{TV}}(P,Q) := 1 - d_{\\mathrm{TV}}(P, Q)$, where $d_{\\mathrm{TV}}$ is the total variation distance between $P$ and $Q$.\nStatistical similarity is a basic measure of similarity between distributions with several natural interpretations.\nWe focus on the case when $P$ and $Q$ are products of Bernoulli trials.\nRecent work has established, somewhat surprisingly, that even for this simple class of distributions exactly computing the TV distance (and hence statistical similarity) is \\#$\\mathsf{P}$-hard.\nThis motivates the question of designing multiplicative approximation algorithms for these computational tasks.\nIt is known that the TV distance computation admits a fully polynomial-time deterministic approximation scheme (FPTAS).\nIt remained an open question whether efficient approximation schemes exist for estimating the statistical similarity between two product distributions.\nIn this work, we affirmatively answer this question by designing an FPTAS for estimating the statistical similarity between two product distributions.\nTo obtain our result, we introduce a new variant of the knapsack problem, which we call multidimensional Masked Knapsack problem, and design an FPTAS to estimate the number of solutions to this problem.\nThis result might be of independent interest.", "keywords": [ "total variation distance", "TV distance", "complement of total variation distance", "TV similarity", "statistical similarity", "aggregated experts", "FPTAS", "reductions", "Turing reductions", "Knapsack", "counting Knapsack", "Masked Knapsack", "counting Masked Knapsack" ], "rating_list": [ 3, 6, 3, 5, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 3, 4 ] }, { "paper_id": "SU3lZ8jrRD", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 10, "author_info_list": [ { "name": "Ali Bahri", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "PhD student" }, { "name": "Christian Desrosiers", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure", "country": "CA", "position": "Full Professor" }, { "name": "David OSOWIECHI", "gender": "unknown", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "PhD student" }, { "name": "Farzad Beizaee", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "PhD student" }, { "name": "Gustavo Adolfo Vargas Hakim", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "PhD student" }, { "name": "Ismail Ben Ayed", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "Full Professor" }, { "name": "Mehrdad Noori", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "PhD student" }, { "name": "Milad Cheraghalikhani", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "MS student" }, { "name": "Moslem Yazdanpanah", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "PhD student" }, { "name": "Sahar Dastani", "gender": "Female", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 50, "n_ref": 118, "n_ref_all": 138, "n_fig": 9, "n_tab": 4, "L_tab": 1404, "n_element_tab": 178, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1092, "n_element_tab_1": 151, "formula_len_all": 570, "formula_len_all_1": 570, "len_all": 143333, "len_all_1": 60833, "len_abs": 1652, "len_title": 137, "len_sents": 34663, "len_sents_1": 27891, "n_sents": 246, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1597, "title": "Spectral Spatial Traversing in Point Clouds: Enhancing Data Analysis with Mamba Networks", "abs": "State Space Models (SSMs) such as Mamba have shown significant promise for sequence modeling in Natural Language Processing (NLP) and, more recently, computer vision. This paper presents a new methodology for both supervised and self-supervised learning using Mamba and Masked Autoencoder networks specifically designed for point cloud data. We propose three main contributions that enhance the capability of Mamba networks to process and understand the complex structure of this type of data. The first strategy exploits the spectrum of a graph Laplacian capturing the local connectivity of patches to define an isometry-invariant traversal order of tokens in the Mamba network. Compared to existing point cloud Mamba architectures, which traverse point patches based on a 3D grid, our approach is more robust to the viewpoint and better captures the shape manifold of the point cloud. The second contribution adapts our approach to segmentation using a recursive patch partitioning strategy informed by spectral components of the Laplacian. This strategy enables a more precise integration and analysis point cloud segments. Our last contribution tackles a significant issue in Masked\nAutoencoder (MAE) for Mamba networks by modifying learnable token placement. Instead of adding them at the end, tokens are restored to their original positions, maintaining essential order and improving learning effectiveness. Extensive experiments confirm our method's superiority over State-Of-The-Art (SOTA) baselines, demonstrating marked improvements in classification, segmentation, and few-shot tasks.", "keywords": [ "State Space Models", "Masked Autoencoders", "3D Point Clouds", "Spectral Spatial Traversing" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "STpxO1Siaq", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Dongxia Wang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Qi Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Qing Guo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Tianlin Li", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 18, "n_ref": 28, "n_ref_all": 41, "n_fig": 4, "n_tab": 6, "L_tab": 1627, "n_element_tab": 263, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1728, "n_element_tab_1": 264, "formula_len_all": 28, "formula_len_all_1": 0, "len_all": 112167, "len_all_1": 54391, "len_abs": 1611, "len_title": 124, "len_sents": 42169, "len_sents_1": 23151, "n_sents": 301, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1617, "title": "Defend against Jailbreak Attacks via Debate with Partially Perceptive Agents", "abs": "Recent studies have shown that maliciously injecting or perturbing the input image in Vision Large Language Models (VLMs) can lead to jailbreak attacks, raising significant security concerns. A straightforward defense strategy against such attacks is to crop the input image, thereby disrupting the effectiveness of the injection or perturbation. However, the cropping can significantly distort the semantics of the input image, leading to an adverse impact on the model's output when processing clean input. To mitigate the adverse impact, we propose a defense mechanism against jailbreak attacks based on a multi-agent debate approach. In this method, one agent (\u201cintegrated\u201d agent) accesses the full integrated image, while the other (\u201cpartial\u201d agent) only accesses cropped/partial images, aiming to avoid the attack while preserving the correct semantics in the output as much as possible. Our key insight is that when an integrated agent debates with a partial agent, if the integrated agent receives clean input, it can successfully persuade the partial agent. Conversely, if the integrated agent is given an attacked input, the partial agent can persuade it to rethink the original output, thereby achieving effective defense against the attack. Empirical experiments have demonstrated that our method provides more effective defense compared to the baseline method, successfully reducing the average attack success rate from 100% to 22%. In more advanced experimental setups, our proposed method can even limit the average attack success rate to 18% (debating with GPT-4o) and 14% (with enhanced perspective).", "keywords": [ "Multi-agent Debate; Defense; Visual Large Language Models" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "SThJXvucjQ", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Arindam Banerjee", "gender": "unknown", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Professor" }, { "name": "Rohan Deb", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "PhD student" }, { "name": "Mohammad Ghavamzadeh", "gender": "Male", "institution": "Amazon AGI", "country": "US", "position": "Senior Principal Scientist" } ], "n_formula": 166, "n_formula_1": 30, "n_ref_uni": 36, "n_ref": 122, "n_ref_all": 156, "n_fig": 7, "n_tab": 7, "L_tab": 1088, "n_element_tab": 160, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 32, "n_element_tab_1": 5, "formula_len_all": 16494, "formula_len_all_1": 2129, "len_all": 238153, "len_all_1": 65098, "len_abs": 1406, "len_title": 110, "len_sents": 68079, "len_sents_1": 29314, "n_sents": 632, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1610, "title": "Conservative Contextual Bandits: Beyond Linear Representations", "abs": "Conservative Contextual Bandits (CCBs) address safety in sequential decision making by requiring that an agent's policy, along with minimizing regret, also satisfies a safety constraint: the performance is not worse than a baseline policy (e.g., the policy that the company has in production) by more than $(1+\\alpha)$ factor. \nPrior work developed UCB-style\nalgorithms for this problem in the multi-armed (Wu et al., 2016) and contextual\nlinear (Kazerouni et al., 2017) settings.\nHowever, in practice the cost of the arms\nis often a non-linear function, and therefore existing UCB algorithms are ineffective in such settings. \nIn this paper, we consider CCBs beyond the linear case and develop two algorithms $\\mathtt{C\\text{-}SquareCB}$ and $\\mathtt{C\\text{-}FastCB}$, using Inverse Gap Weighting (IGW) based exploration and an online regression oracle. \nWe show that the safety constraint is satisfied in high probability and that the regret for $\\mathtt{C\\text{-}SquareCB}$ is sub-linear in horizon $T$, while the the regret for $\\mathtt{C\\text{-}FastCB}$ is first-order and is sub-linear in $L^*$, the cumulative loss of the optimal policy. \nSubsequently, we use a neural network for function approximation and online gradient descent as the regression oracle to provide $\\tilde{\\mathcal{O}}\\big(\\sqrt{KT} + K/\\alpha\\big) $ and $\\tilde{\\mathcal{O}}\\big(\\sqrt{KL^*} + K (1 + 1/\\alpha)\\big)$ regret bounds respectively. \nFinally, we demonstrate the efficacy of our algorithms on real world data, and show that they significantly outperform the existing baseline while maintaining the performance guarantee.", "keywords": [ "Contextual Bandits", "Safety", "Neural Bandits", "Constrained Bandits" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 2, 4 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "STdyyjBZ7P", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Chang Yang", "gender": "Female", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Hau Chan", "gender": "Male", "institution": "University of Nebraska, Lincoln", "country": "", "position": "Assistant Professor" }, { "name": "Pengdeng Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Shuxin Li", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xiao Huang", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "", "position": "Assistant Professor" }, { "name": "Xinrun Wang", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Youzhi Zhang", "gender": "unknown", "institution": "Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 25, "n_ref": 49, "n_ref_all": 68, "n_fig": 12, "n_tab": 2, "L_tab": 750, "n_element_tab": 191, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1149, "formula_len_all_1": 928, "len_all": 130327, "len_all_1": 73791, "len_abs": 2030, "len_title": 77, "len_sents": 49243, "len_sents_1": 35310, "n_sents": 351, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 29, "L_abs": 2075, "title": "In-Context Learning for Games", "abs": "Most literature in algorithmic game theory focuses on equilibrium finding, particularly Nash Equilibrium (NE). However, computing NE typically involves repeated computations of best responses (e.g., policy space response oracle (PSRO)), which can be computationally intensive. Moreover, NE strategies may not be ideal in games with more than two players or when facing irrational opponents. Consequently, NE strategies often require further adaptions to effectively address various types of opponents, impeding practical deployments. In contrast, In-Context Learning (ICL), i.e., learning from context examples, plays the core role in the generalizability of large language models (LLMs) to novel tasks without changing parameters. While ICL has been applied to decision-making tasks, e.g., algorithm distillation (AD), existing research primarily focuses on single-agent scenarios, and the ICL for games is largely unexplored.\nTo facilitate the game solving and the practical deployment, the research question investigated in this work is: *Can we leverage ICL to learn a model to i) play as **any player** of the game, ii) exploit **any opponent** to maximize the utility, and iii) be used to compute NE, **without changing the parameters**?* In this work, we propose **In-Context Exploiter** (**ICE**) to address this question: i) **ICE** generates the diverse opponents with different capability levels for each player of the game to generate the training datasets, ii) **ICE** combines the curriculum learning and the ICL for single-agent scenarios (e.g., AD), to train the single model for all players of games, and iii) **ICE** leverages the pre-trained single model to play as each player of the game against different opponents and integrate with the equilibrium finding framework, e.g., PSRO, to compute NE. Extensive experiments on Kuhn poker, Leduc poker, and Goofspiel demonstrate that **ICE** can efficiently exploit different opponents as different players of the games and can be seamlessly integrated with PSRO to compute NE without changing the parameters.", "keywords": [ "in-context learning", "extensive-form game" ], "rating_list": [ 3, 3, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "STEEDDv3zI", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Francesco Croce", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Postdoc" }, { "name": "Hao Zhao", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "MS student" }, { "name": "Maksym Andriushchenko", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Postdoc" }, { "name": "Nicolas Flammarion", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 29, "n_ref": 86, "n_ref_all": 130, "n_fig": 14, "n_tab": 8, "L_tab": 4890, "n_element_tab": 432, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 917, "n_element_tab_1": 68, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 177327, "len_all_1": 56331, "len_abs": 1257, "len_title": 117, "len_sents": 61914, "len_sents_1": 27368, "n_sents": 470, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1233, "title": "Is In-Context Learning Sufficient for Instruction Following in LLMs?", "abs": "In-context learning (ICL) allows LLMs to learn from examples without changing their weights: this is a particularly promising capability for long-context LLMs that can potentially learn from many examples. Recently, Lin et al. (2024) proposed URIAL, a method using only three in-context examples to align base LLMs, achieving non-trivial instruction following performance. In this work, we show that, while effective, ICL alignment with URIAL still underperforms compared to instruction fine-tuning on established benchmarks such as MT-Bench and AlpacaEval 2.0 (LC), especially with more capable base LLMs. We then uncover the most relevant elements for successful in-context alignment, finding the crucial role of the decoding parameters. Based on these insights, we show that the approach of URIAL can indeed be improved by adding more, potentially carefully selected, high-quality demonstrations in context, getting closer to the performance of instruct models. Finally, we provide the first, to our knowledge, systematic comparison of ICL and instruction fine-tuning (IFT) for instruction following in the low data regime. Overall, our work advances the understanding of ICL as an alignment technique and its relationship to IFT.", "keywords": [ "Large Language Models", "In-context Learning", "Alignment" ], "rating_list": [ 6, 3, 8 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 3, 1, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "STBPaproaB", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 5, "author_info_list": [ { "name": "Daniel Domingo-Fern\u00e1ndez", "gender": "Male", "institution": "Enveda", "country": "US", "position": "Researcher" }, { "name": "Jacques D. Fleuriot", "gender": "unknown", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Full Professor" }, { "name": "Lauren Nicole DeLong", "gender": "Female", "institution": "Edinburgh University, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Paola Galdi", "gender": "unknown", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Postdoc" }, { "name": "Yojana Gadiya", "gender": "Female", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 49, "n_ref": 91, "n_ref_all": 123, "n_fig": 12, "n_tab": 5, "L_tab": 2476, "n_element_tab": 154, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1457, "formula_len_all_1": 1095, "len_all": 145644, "len_all_1": 61715, "len_abs": 1250, "len_title": 111, "len_sents": 39066, "len_sents_1": 29421, "n_sents": 308, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1256, "title": "MARS: A neurosymbolic approach for interpretable drug discovery", "abs": "Neurosymbolic (NeSy) artificial intelligence describes the combination of logic or rule-based techniques with neural networks. Compared to neural approaches, NeSy methods often possess enhanced interpretability, which is particularly promising for biomedical applications like drug discovery. However, since interpretability is broadly defined, there are no clear guidelines for assessing the biological plausibility of model interpretations. To assess interpretability in the context of drug discovery, we devise a novel prediction task, called drug mechanism-of-action (MoA) deconvolution, with an associated, tailored knowledge graph (KG), MoA-net. We then develop the MoA Retrieval System (MARS), a NeSy approach for drug discovery which leverages logical rules with learned rule weights. Using this interpretable feature alongside domain knowledge, we find that MARS and other NeSy approaches on KGs are susceptible to reasoning shortcuts, in which the prediction of true labels is driven by \"degree-bias\" rather than the domain-based rules. Subsequently, we demonstrate ways to identify and mitigate this. Thereafter, MARS achieves performance on par with current state-of-the-art models while producing model interpretations aligned with known MoAs.", "keywords": [ "neurosymbolic", "drug discovery", "interpretability", "knowledge graphs", "reinforcement learning" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "ST6i7VMyYn", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Bingxin Zhou", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "", "position": "Researcher" }, { "name": "Liang Hong", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Full Professor" }, { "name": "Mingchen Li", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Yang Tan", "gender": "Male", "institution": "East China University of Science and Technology", "country": "", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 35, "n_ref": 65, "n_ref_all": 83, "n_fig": 3, "n_tab": 5, "L_tab": 2332, "n_element_tab": 181, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1212, "n_element_tab_1": 155, "formula_len_all": 496, "formula_len_all_1": 496, "len_all": 132543, "len_all_1": 65458, "len_abs": 1190, "len_title": 130, "len_sents": 31224, "len_sents_1": 28418, "n_sents": 217, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1193, "title": "Unlearning Virus Knowledge Toward Safe and Responsible Mutation Effect Predictions", "abs": "Pre-trained deep protein models have become essential tools in fields such as biomedical research, enzyme engineering, and therapeutics due to their ability to predict and optimize protein properties effectively. However, the diverse and broad training data used to enhance the generalizability of these models may also inadvertently introduce ethical risks and pose biosafety concerns, such as the enhancement of harmful viral properties like transmissibility or drug resistance. To address this issue, we introduce a novel approach using knowledge unlearning to selectively remove virus-related knowledge while retaining other useful capabilities. We propose a learning scheme, PROEDIT, for editing a pre-trained protein language model toward safe and responsible mutation effect prediction. Extensive validation on open benchmarks demonstrates that PROEDIT significantly reduces the model's ability to enhance the properties of virus mutants without compromising its performance on non-virus proteins. As the first thorough exploration of safety issues in deep learning solutions for protein engineering, this study provides a foundational step toward ethical and responsible AI in biology.", "keywords": [ "protein language model", "mutation effect prediction", "AI safety" ], "rating_list": [ 5, 1, 3, 5 ], "soundness_list": [ 3, 1, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "SSslAtcPB6", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Hehe Fan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Linchao Zhu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiangpeng Yang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Yi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 33, "n_ref": 72, "n_ref_all": 111, "n_fig": 12, "n_tab": 3, "L_tab": 297, "n_element_tab": 54, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 138, "n_element_tab_1": 24, "formula_len_all": 566, "formula_len_all_1": 450, "len_all": 121706, "len_all_1": 41391, "len_abs": 625, "len_title": 126, "len_sents": 34004, "len_sents_1": 17738, "n_sents": 274, "n_sents_1": 140, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1099, "title": "ST-Modulator: Modulating Space-Time Attention for Multi-Grained Video Editing", "abs": "Recent advancements in diffusion models have significantly improved video generation and editing capabilities. However, multi-grained video editing, which encompasses class-level, instance-level, and part-level modifications, remains a formidable challenge. The major difficulties in multi-grained editing include semantic misalignment of text-to-region control and feature coupling within the diffusion model. To address these difficulties, we present ST-Modulator, a zero-shot approach that modulates space-time (cross- and self-) attention mechanisms to achieve fine-grained control over video content. We enhance text-to-region control by amplifying each local prompt's attention to its corresponding spatial-disentangled region while minimizing interactions with irrelevant areas in cross-attention. Additionally, we improve feature separation by increasing intra-region awareness and reducing inter-region interference in self-attention. Extensive experiments demonstrate our method achieves state-of-the-art performance in real-world scenarios. More details are available on the project page.", "keywords": [ "diffusion model", "video editing" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "SSE9myD9SG", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Anwesa Choudhuri", "gender": "Female", "institution": "United Imaging Intelligence", "country": "US", "position": "Researcher" }, { "name": "Benjamin Planche", "gender": "Male", "institution": "United Imaging Intelligence", "country": "US", "position": "Researcher" }, { "name": "Chen Chen", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Meng Zheng", "gender": "Female", "institution": "UII America, Inc.", "country": "US", "position": "Researcher" }, { "name": "Qucheng Peng", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Terrence Chen", "gender": "Male", "institution": "United Imaging Intelligence", "country": "US", "position": "CEO" }, { "name": "Zhongpai Gao", "gender": "Male", "institution": "United Imaging Intelligence", "country": "US", "position": "Researcher" }, { "name": "Ziyan Wu", "gender": "Male", "institution": "United Imaging Intelligence", "country": "US", "position": "Principal Expert Scientist" } ], "n_formula": 16, "n_formula_1": 18, "n_ref_uni": 28, "n_ref": 84, "n_ref_all": 105, "n_fig": 15, "n_tab": 36, "L_tab": 7653, "n_element_tab": 1195, "n_fig_1": 10, "n_tab_1": 8, "L_tab_1": 1392, "n_element_tab_1": 233, "formula_len_all": 1191, "formula_len_all_1": 1107, "len_all": 173231, "len_all_1": 63065, "len_abs": 1332, "len_title": 85, "len_sents": 47620, "len_sents_1": 27171, "n_sents": 366, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 37, "L_abs": 1336, "title": "3D Vision-Language Gaussian Splatting", "abs": "Recent advancements in 3D reconstruction methods and vision-language models have propelled the development of multi-modal 3D scene understanding, which has vital applications in robotics, autonomous driving, and virtual/augmented reality. However, current multi-modal scene understanding approaches have naively embedded semantic representations into 3D reconstruction methods without striking a balance between visual and language modalities, which leads to unsatisfying semantic rasterization of translucent or reflective objects, as well as over-fitting on color modality. To alleviate these limitations, we propose a solution that adequately handles the distinct visual and semantic modalities, i.e., a 3D vision-language Gaussian splatting model for scene understanding, to put emphasis on the representation learning of language modality. We propose a novel cross-modal rasterizer, using modality fusion along with a smoothed semantic indicator for enhancing semantic rasterization. We also employ a camera-view blending technique to improve semantic consistency between existing and synthesized views, thereby effectively mitigating over-fitting. Extensive experiments demonstrate that our method achieves state-of-the-art performance in open-vocabulary semantic segmentation, surpassing existing methods by a significant margin.", "keywords": [ "Multi-modal 3D Scene Understanding", "Gaussian Splatting", "Open-vocabulary Semantic Segmentation", "Open-vocabulary Object Localization" ], "rating_list": [ 6, 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 3, 2 ] }, { "paper_id": "SRpq5OBpED", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Ayesha Vermani", "gender": "unknown", "institution": "Champalimaud Centre for the Unknown", "country": "", "position": "PhD student" }, { "name": "Hyungju Jeon", "gender": "unknown", "institution": "Champalimaud Centre for the Unknown", "country": "", "position": "PhD student" }, { "name": "Il Memming Park", "gender": "Male", "institution": "Champalimaud Centre for the Unknown", "country": "PT", "position": "Associate Professor" }, { "name": "Josue Nassar", "gender": "Male", "institution": "Optum Labs", "country": "", "position": "Researcher" }, { "name": "Matthew Dowling", "gender": "Male", "institution": "State University of New York, Stony Brook", "country": "", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 20, "n_ref_uni": 40, "n_ref": 64, "n_ref_all": 118, "n_fig": 37, "n_tab": 2, "L_tab": 3, "n_element_tab": 2, "n_fig_1": 28, "n_tab_1": 1, "L_tab_1": 2, "n_element_tab_1": 1, "formula_len_all": 1824, "formula_len_all_1": 1234, "len_all": 161516, "len_all_1": 65728, "len_abs": 1211, "len_title": 119, "len_sents": 53128, "len_sents_1": 30483, "n_sents": 366, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 70, "L_abs": 1215, "title": "Meta-Dynamical State Space Models for Integrative Neural Data Analysis", "abs": "Learning shared structure across environments facilitates rapid learning and adaptive behavior in neural systems. This has been widely demonstrated and applied in machine learning to train models that are capable of generalizing to novel settings. However, there has been limited work exploiting the shared structure in neural activity during similar tasks for learning latent dynamics from neural recordings.\nExisting approaches are designed to infer dynamics from a single dataset and cannot be readily adapted to account for statistical heterogeneities across recordings. In this work, we hypothesize that similar tasks admit a corresponding family of\nrelated solutions and propose a novel approach for meta-learning this solution space from task-related neural activity of trained animals. Specifically, we capture the variabilities across recordings on a low-dimensional manifold which concisely parametrizes this family of dynamics, thereby facilitating rapid learning of latent dynamics given new recordings. We demonstrate the efficacy of our approach on\nfew-shot reconstruction and forecasting of synthetic dynamical systems, and neural recordings from the motor cortex during different arm reaching tasks.", "keywords": [ "neural dynamics", "state-space model", "meta learning" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 4, 3, 2 ], "presentation_list": [ 3, 4, 4, 1 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "SRjzerUpB2", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Han Wang", "gender": "unknown", "institution": "University of Alberta", "country": "", "position": "PhD student" }, { "name": "Lingwei Zhu", "gender": "Male", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "JP", "position": "Postdoc" }, { "name": "Yukie Nagai", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 32, "n_ref": 58, "n_ref_all": 77, "n_fig": 6, "n_tab": 1, "L_tab": 231, "n_element_tab": 6, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1768, "formula_len_all_1": 1394, "len_all": 108814, "len_all_1": 63933, "len_abs": 1201, "len_title": 112, "len_sents": 34024, "len_sents_1": 28052, "n_sents": 300, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1275, "title": "Fat-to-Thin Policy Optimization: Offline Reinforcement Learning with Sparse Policies", "abs": "Sparse continuous policies are distributions that can choose some actions at random yet keep strictly zero probability for the other actions, which are radically different from the Gaussian.\nThey have important real-world implications, e.g. in modeling safety-critical tasks like medicine.\nThe combination of offline reinforcement learning and sparse policies provides a novel paradigm that enables learning completely from logged datasets a safety-aware sparse policy. \nHowever, sparse policies can cause difficulty with the existing offline algorithms which require evaluating actions that fall outside of the current support.\nIn this paper, we propose the first offline policy optimization algorithm that tackles this challenge: Fat-to-Thin Policy Optimization (FtTPO).\nSpecifically, we maintain a fat (heavy-tailed) proposal policy that effectively learns from the dataset and injects knowledge to a thin (sparse) policy, which is responsible for interacting with the environment.\nWe instantiate FtTPO with the general $q$-Gaussian family that encompasses both heavy-tailed and sparse policies and verify that it performs favorably in a safety-critical treatment simulation and the standard MuJoCo suite.\nOur code is available at https://github.com/lingweizhu/fat2thin.", "keywords": [ "reinforcement learning", "offline reinforcement learning", "actor critic", "sparse policies" ], "rating_list": [ 8, 3, 8, 8 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "SRghq20nGU", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Hitoshi Imaoka", "gender": "Male", "institution": "NEC", "country": "JP", "position": "Fellow" }, { "name": "Kazuyuki Sakurai", "gender": "unknown", "institution": "NEC", "country": "", "position": "Principal Researcher" }, { "name": "Taiki Miyagawa", "gender": "Male", "institution": "Independent researcher", "country": "JP", "position": "Researcher" }, { "name": "Akinori F Ebihara", "gender": "Male", "institution": "NEC Corporation", "country": "JP", "position": "Principal Researcher" } ], "n_formula": 75, "n_formula_1": 25, "n_ref_uni": 144, "n_ref": 251, "n_ref_all": 300, "n_fig": 26, "n_tab": 7, "L_tab": 3210, "n_element_tab": 161, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 1800, "n_element_tab_1": 96, "formula_len_all": 6882, "formula_len_all_1": 2764, "len_all": 400210, "len_all_1": 65130, "len_abs": 1526, "len_title": 163, "len_sents": 99603, "len_sents_1": 26442, "n_sents": 883, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 115, "L_abs": 1540, "title": "Learning the Optimal Stopping for Early Classification within Finite Horizons via Sequential Probability Ratio Test", "abs": "Time-sensitive machine learning benefits from Sequential Probability Ratio Test (SPRT), which provides an optimal stopping time for early classification of time series. However, in *finite horizon* scenarios, where input lengths are finite, determining the optimal stopping rule becomes computationally intensive due to the need for *backward induction*, limiting practical applicability. We thus introduce FIRMBOUND, an SPRT-based framework that efficiently estimates the solution to backward induction from training data, bridging the gap between optimal stopping theory and real-world deployment. It employs *density ratio estimation* and *convex function learning* to provide statistically consistent estimators for sufficient statistic and conditional expectation, both essential for solving backward induction; consequently, FIRMBOUND minimizes Bayes risk to reach optimality. Additionally, we present a faster alternative using Gaussian process regression, which significantly reduces training time while retaining low deployment overhead, albeit with potential compromise in statistical consistency. Experiments across independent and identically distributed (i.i.d.), non-i.i.d., binary, multiclass, synthetic, and real-world datasets show that FIRMBOUND achieves optimalities in the sense of Bayes risk and speed-accuracy tradeoff. Furthermore, it advances the tradeoff boundary toward optimality when possible and reduces decision-time variance, ensuring reliable decision-making. Code is included in the supplementary materials.", "keywords": [ "Early Classification of Time Series", "Sequential Probability Ratio Test" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 3, 2, 2 ] }, { "paper_id": "SR8LFpmVun", "primary_area": "foundation or frontier models, including LLMs", "n_author": 12, "author_info_list": [ { "name": "Chengming Li", "gender": "unknown", "institution": " Shenzhen MSU-BIT University", "country": "CN", "position": "Full Professor" }, { "name": "Chuanyang Zheng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Fanghua Ye", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Jianqiao Lu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Lingpeng Kong", "gender": "Male", "institution": "Department of Computer Science, The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Ngai Wong", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Xiaodan Liang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Xun Wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Zhenan Sun", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhongwei Wan", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Zixuan Li", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "MS student" }, { "name": "jing xiong", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 56, "n_ref": 114, "n_ref_all": 131, "n_fig": 3, "n_tab": 8, "L_tab": 2798, "n_element_tab": 470, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 830, "n_element_tab_1": 82, "formula_len_all": 672, "formula_len_all_1": 640, "len_all": 191004, "len_all_1": 63301, "len_abs": 1598, "len_title": 153, "len_sents": 46957, "len_sents_1": 29884, "n_sents": 351, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1169, "title": "UncertaintyRAG: Span Uncertainty Enhanced Long-Context Modeling for Retrieval-Augmented Generation", "abs": "We introduce $UncertaintyRAG$, a novel method for long-context Retrieval-Augmented Generation (RAG) that leverages Signal-to-Noise Ratio (SNR)-based span uncertainty to estimate similarity between text chunks. This span uncertainty improves the calibration of model predictions, enhancing robustness and addressing semantic inconsistencies caused by random chunking. Utilizing this, we develop an efficient unsupervised learning technique for training the retrieval model and design an effective data sampling and scaling strategy. $UncertaintyRAG$ achieves a 2.03\\% improvement over baselines on LLaMA-2-7B, reaching state-of-the-art performance while using only 4\\% of the training data compared to other powerful open-source retrieval models under distribution shift settings. Our method demonstrates strong calibration through span uncertainty, resulting in better generalization and robustness in long-context RAG tasks. Moreover, $UncertaintyRAG$ offers a lightweight retrieval model that can be seamlessly integrated into any large language model with varying context window lengths without the need for fine-tuning, highlighting the versatility of our approach.", "keywords": [ "RAG", "Long-contex", "Distribution Shift", "Signal-to-noise Ratio", "Unsupervised Learning" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "SQnitDuow6", "primary_area": "reinforcement learning", "n_author": 9, "author_info_list": [ { "name": "Bo Dai", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Dale Schuurmans", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Professor" }, { "name": "Hanjun Dai", "gender": "Male", "institution": "Google Research", "country": "", "position": "Researcher" }, { "name": "Jincheng Mei", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Research Scientist" }, { "name": "Katayoon Goshvadi", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Mengjiao Yang", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "Student" }, { "name": "Shicong Cen", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Tong Yang", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Yuejie Chi", "gender": "Female", "institution": "Meta AI", "country": "US", "position": "Researcher" } ], "n_formula": 105, "n_formula_1": 38, "n_ref_uni": 60, "n_ref": 124, "n_ref_all": 161, "n_fig": 5, "n_tab": 2, "L_tab": 967, "n_element_tab": 119, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 250, "n_element_tab_1": 3, "formula_len_all": 11278, "formula_len_all_1": 3236, "len_all": 193229, "len_all_1": 65751, "len_abs": 1547, "len_title": 137, "len_sents": 47049, "len_sents_1": 28836, "n_sents": 413, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1559, "title": "Value-Incentivized Preference Optimization: A Unified Approach to Online and Offline RLHF", "abs": "Reinforcement learning from human feedback (RLHF) has demonstrated great promise in aligning large language models (LLMs) with human preference. Depending on the availability of preference data, both online and offline RLHF are active areas of investigation. A key bottleneck is understanding how to incorporate uncertainty estimation in the reward function learned from the preference data for RLHF, regardless of how the preference data is collected. While the principles of optimism or pessimism under uncertainty are well-established in standard reinforcement learning (RL), a practically-implementable and theoretically-grounded form amenable to large language models is not yet available, as standard techniques for constructing confidence intervals become intractable under arbitrary policy parameterizations.\n\nIn this paper, we introduce a unified approach to online and offline RLHF --- value-incentivized preference optimization (VPO) --- which regularizes the maximum-likelihood estimate of the reward function with the corresponding value function, modulated by a sign to indicate whether the optimism or pessimism is chosen. VPO also directly optimizes the policy with implicit reward modeling, and therefore shares a simpler RLHF pipeline similar to direct preference optimization. Theoretical guarantees of VPO are provided for both online and offline settings, matching the rates of their standard RL counterparts. Moreover, experiments on text summarization, dialogue, and standard benchmarks verify the practicality and effectiveness of VPO.", "keywords": [ "preference optimization", "the principle of optimism/pessimism", "RLHF theory" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "SQl6T4dfs8", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Dongzhe Zheng", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "MS student" }, { "name": "Wenjie Mei", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 81, "n_formula_1": 19, "n_ref_uni": 19, "n_ref": 22, "n_ref_all": 31, "n_fig": 4, "n_tab": 6, "L_tab": 1832, "n_element_tab": 378, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 954, "n_element_tab_1": 73, "formula_len_all": 2997, "formula_len_all_1": 691, "len_all": 174506, "len_all_1": 55487, "len_abs": 2693, "len_title": 175, "len_sents": 57928, "len_sents_1": 24159, "n_sents": 410, "n_sents_1": 164, "venue": "ICLR2025", "decision": "withdraw", "L_title": 125, "L_abs": 1483, "title": "Koopman Universal Neural Dynamic Operator: Achieving Fully Explicit Expression Identification for Nonlinear Dynamical Systems", "abs": "Complex nonlinear systems permeate various scientific and engineering domains, presenting significant challenges in accurate modeling and analysis. This paper introduces the Koopman Universal Neural Dynamic Operator (KUNDO), a groundbreaking framework that bridges the gap between data-driven machine learning approaches and traditional mathematical modeling. KUNDO uniquely combines neural networks, Koopman operator theory, and the universal approximation theorem to achieve fully explicit expression identification for complex nonlinear systems. Our framework demonstrates remarkable efficiency in small sample scenarios, overcoming limitations of both classical physical models and black-box machine learning techniques. By learning Koopman-compatible basis functions through neural networks, KUNDO transforms high-dimensional, strongly nonlinear dynamics into interpretable mathematical forms, greatly decreasing the limitations of human selection of basis functions without sacrificing predictive power. We present theoretical analyses of KUNDO's mathematical properties and validate its performance across diverse nonlinear systems. The results showcase KUNDO's potential to revolutionize system identification, offering new avenues for scientific discovery and engineering applications in fields such as climate science, financial modeling, and advanced robotics. This work presents a significant advance towards interpretable AI and data-driven modeling in systems analysis.", "keywords": [ "neural networks", "Koopman operator", "universal approximation theorem", "system identification", "dynamics system modeling" ], "rating_list": [ 3, 6, 3, 1 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 2, 3, 1, 1 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "SPu6k4OZkj", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Jong Sung Park", "gender": "Male", "institution": "Indiana University", "country": "", "position": "PhD student" }, { "name": "Sai Prajwal Reddy", "gender": "Male", "institution": "Indiana University", "country": "US", "position": "PhD student" }, { "name": "Serge Koudoro", "gender": "Male", "institution": "Indiana University at Bloomington", "country": "US", "position": "Researcher" }, { "name": "Eleftherios Garyfallidis", "gender": "Male", "institution": "Indiana University", "country": "US", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 47, "n_ref": 54, "n_ref_all": 103, "n_fig": 30, "n_tab": 7, "L_tab": 8591, "n_element_tab": 1048, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 1177, "n_element_tab_1": 110, "formula_len_all": 2050, "formula_len_all_1": 1066, "len_all": 222619, "len_all_1": 71820, "len_abs": 1021, "len_title": 111, "len_sents": 62683, "len_sents_1": 31659, "n_sents": 655, "n_sents_1": 334, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1027, "title": "Thetan Berserker: Fast and Stochastic Distance-based Clustering", "abs": "Clustering is a challenging NP-hard problem. Polynomial approximations are of paramount importance for identifying intriguing hidden representations of data at reasonable execution times. In this work, we propose a novel clustering algorithm called Thetan Berserker (TB). TB is a centroid-based clustering method controlled by a single distance parameter. TB revitalizes an old family of sequential algorithms which are adored for their speed but are known to be order sensitive. In addition, TB enables widely used algorithms such as KMeans and DBSCAN by improving their initial conditions. Theoretical aspects are provided in detail along with extensive comparisons and benchmarks. Examples of real world applications are provided using publicly available data of different dimensionalities. A wide range of performance boosts in clustering accuracy, memory usage, and runtime are reported. By dramatically reducing clustering ambiguities, while staying at incredibly low complexity, TB creates a new standard for clustering.", "keywords": [ "clustering" ], "rating_list": [ 5, 5, 8, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "SPcmEiiDDo", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Fuyuan Zhang", "gender": "unknown", "institution": "The University of Tokyo", "country": "", "position": "Assistant Professor" }, { "name": "Jianjun Zhao", "gender": "Male", "institution": "Kyushu University", "country": "JP", "position": "Full Professor" }, { "name": "Jianlang Chen", "gender": "Male", "institution": "Kyushu University", "country": "", "position": "PhD student" }, { "name": "Lei Ma", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Associate Professor" }, { "name": "Siyuan Chen", "gender": "Male", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Xiongfei Wu", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Researcher" }, { "name": "Zhuo Li", "gender": "Male", "institution": "Kyushu University", "country": "", "position": "PhD student" }, { "name": "PENGZHAN ZHAO", "gender": "Male", "institution": "Kyushu University", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 8, "n_ref_uni": 32, "n_ref": 79, "n_ref_all": 101, "n_fig": 9, "n_tab": 6, "L_tab": 1044, "n_element_tab": 84, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 213, "formula_len_all_1": 492, "len_all": 109114, "len_all_1": 59893, "len_abs": 1755, "len_title": 130, "len_sents": 35134, "len_sents_1": 28232, "n_sents": 258, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 908, "title": "Episodic Control-Based Adversarial Policy Learning in Two-player Competitive Games", "abs": "Training adversarial agents to attack neural network policies has proven to be both effective and practical. However, we observe that existing methods can be further enhanced by distinguishing between states leading to win or lose and encouraging the policy training to prioritize winning states. In this paper, we address this gap by introducing an episodic control-based approach for adversarial policy training. Our method extracts the historical evaluations for states from historical experiences with an episodic memory, and then incorporating these evaluations into the rewards to improve the adversarial policy optimization. We evaluate our approach using two-player competitive games in MuJoCo simulation environments, demonstrating that our method establishes the most promising attack performance and defense difficulty against the victims among the existing adversarial policy training techniques.", "keywords": [ "Reinforcement Learning; Adversarial Policy Training; Episodic Control" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "SPcJPZVYAq", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Bingzhe Li", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "Assistant Professor" }, { "name": "Guanpeng Li", "gender": "unknown", "institution": "University of Iowa", "country": "US", "position": "Assistant Professor" }, { "name": "Yi Li", "gender": "unknown", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Zhichun Guo", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "Postdoc" } ], "n_formula": 39, "n_formula_1": 17, "n_ref_uni": 47, "n_ref": 110, "n_ref_all": 140, "n_fig": 4, "n_tab": 6, "L_tab": 2873, "n_element_tab": 83, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1606, "n_element_tab_1": 20, "formula_len_all": 2773, "formula_len_all_1": 880, "len_all": 170079, "len_all_1": 69916, "len_abs": 1286, "len_title": 131, "len_sents": 52282, "len_sents_1": 30614, "n_sents": 411, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1291, "title": "YOSO: You-Only-Sample-Once via Compressed Sensing for Graph Neural Network Training", "abs": "Graph Neural Networks (GNNs) have become essential tools for analyzing structured data across various domains. In GNNs, sampling is critical for reducing training latency by limiting the number of nodes processed during training, especially for large-scale applications. However, as the demand for better prediction performance increases, existing sampling algorithms become more complex, introducing significant overhead in the training process. To address this issue, we introduce YOSO (You-Only-Sample-Once), an algorithm designed to achieve highly efficient training while preserving prediction accuracy in downstream tasks. YOSO proposes a compressed sensing-based sampling and reconstruction framework, where nodes are sampled once at the input layer, followed by a lossless reconstruction at the output layer during each epoch. This approach not only avoids costly computations, such as orthonormal basis, but also guarantees high-probability accuracy retention, equivalent to full node participation. Experimental results on both node classification and link prediction tasks demonstrate the effectiveness and efficiency of YOSO, reducing GNN training by an average of around 75% compared to state-of-the-art methods, while maintaining accuracy on par with top-performing baselines.", "keywords": [ "Sampling", "Graph Neural Network", "Compressed Sensing" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "SPViZd7rvi", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Arjun Seshadri", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "David Mingfei Liu", "gender": "unknown", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Johan Ugander", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Tina Eliassi-Rad", "gender": "Female", "institution": "Northeastern University", "country": "", "position": "Associate Professor" } ], "n_formula": 41, "n_formula_1": 24, "n_ref_uni": 31, "n_ref": 56, "n_ref_all": 73, "n_fig": 6, "n_tab": 4, "L_tab": 4202, "n_element_tab": 1160, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1669, "n_element_tab_1": 92, "formula_len_all": 2192, "formula_len_all_1": 1311, "len_all": 136591, "len_all_1": 62871, "len_abs": 1674, "len_title": 164, "len_sents": 40845, "len_sents_1": 28143, "n_sents": 312, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 116, "L_abs": 1685, "title": "Bypassing Skip-Gram Negative Sampling: Dimension Regularization as a More Efficient Alternative for Graph Embeddings", "abs": "A wide range of graph embedding objectives decompose into two components: one that attracts the embeddings of nodes that are perceived as similar, and another that repels embeddings of nodes that are perceived as dissimilar. Without repulsion, the embeddings would collapse into trivial solutions. Skip-Gram Negative Sampling (SGNS) is a popular and efficient repulsion approach that prevents collapse by repelling each node from a sample of dissimilar nodes. In this work, we show that when repulsion is most needed and the embeddings approach collapse, SGNS node-wise repulsion is, in the aggregate, an approximate re-centering of the node embedding dimensions. Such dimension operations are much more scalable than node operations and yield a simpler geometric interpretation of the repulsion. Our result extends findings from self-supervised learning to the skip-gram model, establishing a connection between skip-gram node contrast and dimension regularization. We use this observation to propose a flexible algorithm augmentation framework that improves the scalability of any existing algorithm using SGNS. The framework prioritizes node attraction and replaces SGNS with dimension regularization. We instantiate this generic framework for LINE and node2vec and show that the augmented algorithms preserve downstream link-prediction performance while reducing GPU memory usage by up to $33.3$% and training time by $22.1$%. Further, for graphs that are globally sparse but locally dense, we show that removing repulsion altogether can improve performance, but, when repulsion is otherwise needed, dimension regularization provides an effective and efficient alternative to SGNS.", "keywords": [ "graph embeddings", "negative sampling", "skip gram", "dimension regularization" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "SPS6HzVzyt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Aditi Raghunathan", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Christina Baek", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Sachin Goyal", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zico Kolter", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 75, "n_formula_1": 13, "n_ref_uni": 30, "n_ref": 62, "n_ref_all": 89, "n_fig": 17, "n_tab": 4, "L_tab": 820, "n_element_tab": 35, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10003, "formula_len_all_1": 756, "len_all": 196826, "len_all_1": 61827, "len_abs": 1394, "len_title": 146, "len_sents": 60782, "len_sents_1": 30133, "n_sents": 502, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 98, "L_abs": 1426, "title": "Context-Parametric Inversion: Why Instruction Finetuning May Not Actually Improve Context Reliance", "abs": "Large Language Model's are instruction-finetuned to enhance their ability to follow user instructions and better comprehend input context. Still, they often struggle to follow the input context, especially when it contradicts model's parametric knowledge. This manifests as various failures, such as hallucinations where a model inserts outdated or unwarranted facts into its response. In this work, we observe an intriguing phenomenon: the context reliance of the model decreases as instruction finetuning progresses, $\\textit{despite an initial expected increase}$. We call this phenomenon as the $\\textbf{context-parametric inversion}$. This is surprising, as one would expect instruction tuning to improve the model's ability to follow input instructions. We observe this behavior on multiple general purpose instruction tuning datasets such as TULU, Alpaca and Ultrachat, across multiple model families like Llama, Mistral and Pythia. We perform various controlled studies to eliminate some simple hypothesis for this observed behavior and isolate what datapoints cause this counter-intuitive behavior. We then analyze the phenomenon theoretically, to explain why context reliance varies across the trajectory of finetuning. \nWe tie the observed context-parametric inversion to the properties of the finetuning data, which provides us with some potential mitigation strategies that provide limited but insightful gains.", "keywords": [ "Instruction finetuning", "context-vs-parametric reliance" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "SOsotxYtPC", "primary_area": "foundation or frontier models, including LLMs", "n_author": 13, "author_info_list": [ { "name": "Anh-Tien Nguyen", "gender": "Male", "institution": "Georg-August Universit\u00e4t G\u00f6ttingen", "country": "DE", "position": "PhD student" }, { "name": "Bao Hoang Le", "gender": "Male", "institution": "Dublin City University", "country": "IE", "position": "PhD student" }, { "name": "Daniel Sonntag", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Director" }, { "name": "Duy Minh Ho Nguyen", "gender": "Male", "institution": "University of Stuttgart and Max Planck Research School for Intelligent Systems", "country": "DE", "position": "PhD student" }, { "name": "James Zou", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Mathias Niepert", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "", "position": "Full Professor" }, { "name": "Nghiem Tuong Diep", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Intern" }, { "name": "Nhat Ho", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Pengtao Xie", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Roger Wattenhofer", "gender": "Not Specified", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Tai Nguyen", "gender": "Male", "institution": "University of St. Andrews", "country": "GB", "position": "PhD student" }, { "name": "TrungTin Nguyen", "gender": "Male", "institution": "The University of Queensland", "country": "AU", "position": "Postdoc" }, { "name": "Trung Quoc Nguyen", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Intern" } ], "n_formula": 62, "n_formula_1": 21, "n_ref_uni": 48, "n_ref": 103, "n_ref_all": 132, "n_fig": 5, "n_tab": 11, "L_tab": 2641, "n_element_tab": 189, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 263, "n_element_tab_1": 25, "formula_len_all": 7845, "formula_len_all_1": 1571, "len_all": 232649, "len_all_1": 74601, "len_abs": 2086, "len_title": 127, "len_sents": 55313, "len_sents_1": 28760, "n_sents": 438, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 2094, "title": "LoGra-Med: Long-Context Multi-Graph Alignment for Medical Visual-Language Models", "abs": "State-of-the-art medical multi-modal large language models (med-MLLM), such as LLAVA-MED or BIOMEDGPT, leverage instruction-following data in their pre-training stages. However, those models primarily focus on scaling the model size and data volume to boost performance while mainly relying on the autoregressive learning objectives. Surprisingly, we reveal that such learning schemes might result in a weak alignment between vision and language modalities, making these models highly reliant on extensive pre-training datasets \u2014 a significant challenge in medical domains due to the expensive and time-consuming nature of curating high-quality instruction-following instances. We address this challenge with a new multi-graph alignment algorithm, namely LOGRA-MED, which enforces triplet correlations on the latent embedding space among image modalities, conversation-based descriptions, and extended contextual captions. Owing to this technique, the model is encouraged to capture the semantic meaning of the context, handle linguistic variability where the captions or questions may differ from training instances, and learn cross-modal associations, linking visual elements with various textual interpretations. To scale our algorithm to the med-MLLM setting, we also design an efficient end-to-end learning scheme based on advanced black-box gradient-estimation techniques that permit fast forward and backward steps through the LLM model (LLaMa 7B). Empirical results show\nthat we can match the performance of LLAVA-Med pre-trained on 600K image-text pairs from PMC-15M for Medical VQA tasks and significantly outperform it when trained on only 10% of the data. For instance, on VQA-RAD, we exceed LLAVA-Med (both trained on 10%) by 20.13% and achieve near parity with the 100% pre-training setting (72.52% vs. 72.64%). Additionally, we also surpass other SOTA pre-training methods and med-MLLM such as BIOMEDGPT on visual chatbot or RADFM on zero-shot image classification with VQA, showcasing the power of multi-graph alignment in improving vision-language integration for medical-MLLM.", "keywords": [ "multi-modal LLM", "AI for Healthcare", "multi-modal learning" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "SOd07Qxkw4", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Gen Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Yuchen Jiao", "gender": "unknown", "institution": "Beijing Institute of Technology", "country": "", "position": "Postdoc" } ], "n_formula": 124, "n_formula_1": 49, "n_ref_uni": 22, "n_ref": 65, "n_ref_all": 80, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 16834, "formula_len_all_1": 4302, "len_all": 106617, "len_all_1": 62091, "len_abs": 1087, "len_title": 108, "len_sents": 22587, "len_sents_1": 18111, "n_sents": 275, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 60, "L_abs": 1158, "title": "Improved Convergence Rate for Diffusion Probabilistic Models", "abs": "Score-based diffusion models have achieved remarkable empirical performance in the field of machine learning and artificial intelligence for their ability to generate high-quality new data instances from complex distributions. Improving our understanding of diffusion models, including mainly convergence analysis for such models, has attracted a lot of interests. Despite a lot of theoretical attempts, there still exists significant gap between theory and practice. Towards to close this gap, we establish an iteration complexity at the order of $d^{1/3}\\varepsilon^{-2/3}$, which is better than $d^{5/12}\\varepsilon^{-1}$, the best known complexity achieved before our work. This convergence analysis is based on a randomized midpoint method, which is first proposed for log-concave sampling \\citep{Shen2019TheRandomized}, and then extended to diffusion models by \\citet{Gupta2024Faster}. Our theory accommodates $\\varepsilon$-accurate score estimates, and does not require log-concavity on the target distribution. Moreover, the algorithm can also be parallelized to run in only $O(\\log^2(d/\\varepsilon))$ parallel rounds in a similar way to prior works.", "keywords": [ "score-based generative model", "diffusion model", "probability flow ODE", "randomized learning rate" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 4, 4, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "SOXxa4pPGY", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Fei Miao", "gender": "Female", "institution": "University of Connecticut", "country": "US", "position": "Associate Professor" }, { "name": "Yi Shen", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "MS student" }, { "name": "Yuan Zhuang", "gender": "Male", "institution": "University of Connecticut", "country": "US", "position": "PhD student" }, { "name": "Zhili Zhang", "gender": "Male", "institution": "University of Connecticut", "country": "", "position": "PhD student" }, { "name": "Yuxiao Chen", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Postdoc" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 28, "n_ref": 62, "n_ref_all": 80, "n_fig": 9, "n_tab": 4, "L_tab": 668, "n_element_tab": 45, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 615, "n_element_tab_1": 33, "formula_len_all": 223, "formula_len_all_1": 226, "len_all": 163230, "len_all_1": 58678, "len_abs": 1353, "len_title": 115, "len_sents": 58108, "len_sents_1": 27770, "n_sents": 425, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1360, "title": "YOLO-MARL: You Only LLM Once for Multi-agent Reinforcement Learning", "abs": "Advancements in deep multi-agent reinforcement learning (MARL) have positioned it as a promising approach for decision-making in cooperative games. However, it still remains challenging for MARL agents to learn cooperative strategies for some game environments. Recently, large language models (LLMs) have demonstrated emergent reasoning capabilities, making them promising candidates for enhancing coordination among the agents. However, due to the model size of LLMs, it can be expensive to frequently infer LLMs for actions that agents can take. In this work, we propose You Only LLM Once for MARL (YOLO-MARL), a novel framework that leverages the high-level task planning capabilities of LLMs to improve the policy learning process of multi-agents in cooperative games. Notably, for each game environment, YOLO-MARL only requires one time interaction with LLMs in the proposed strategy generation, state interpretation and planning function generation modules, before the MARL policy training process. This avoids the ongoing costs and computational time associated with frequent LLMs API calls during training. Moreover, the trained decentralized normal-sized neural network-based policies operate independently of the LLM. We evaluate our method across three different environments and demonstrate that YOLO-MARL outperforms traditional MARL algorithms.", "keywords": [ "Multi-agent reinforcement learning", "Large Language Models", "reinforcement learning" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 3, 1 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "SOWZ59UyNc", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Haohan Lin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Sean Welleck", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Yiming Yang", "gender": "Female", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Zhiqing Sun", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 5, "n_ref_uni": 39, "n_ref": 110, "n_ref_all": 126, "n_fig": 5, "n_tab": 9, "L_tab": 3273, "n_element_tab": 497, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 959, "n_element_tab_1": 133, "formula_len_all": 701, "formula_len_all_1": 423, "len_all": 156159, "len_all_1": 53613, "len_abs": 1343, "len_title": 102, "len_sents": 43634, "len_sents_1": 24484, "n_sents": 320, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 54, "L_abs": 1241, "title": "Lean-STaR: Learning to Interleave Thinking and Proving", "abs": "Traditional language model-based theorem proving assumes that by training on a sufficient amount of formal proof data, a model will learn to prove theorems. Our key observation is that a wealth of informal information that is not present in formal proofs can be useful for learning to prove theorems. For instance, humans think through steps of a proof, but this thought process is not visible in the resulting code. We present Lean-STaR, a framework for training language models to produce informal thoughts prior to each step of a proof, thereby boosting the model's theorem-proving capabilities. Lean-STaR uses retrospective ground-truth tactics to generate synthetic thoughts for training the language model. At inference time, the trained model directly generates the thoughts prior to the prediction of the tactics in each proof step. Building on the self-taught reasoner framework, we then apply expert iteration to further fine-tune the model on the correct proofs it samples and verifies using the Lean solver. Lean-STaR significantly outperform base models (43.4% \u2192 46.3%, Pass@64). We also analyze the impact of the augmented thoughts on various aspects of the theorem proving process, providing insights into their effectiveness.", "keywords": [ "Automated Theorem Proving", "AI for Math", "Chain-of-Thought Reasoning" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "SOVwGa0H2c", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 19, "author_info_list": [ { "name": "Anlan Zhang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Chaoyun Zhang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Chendong Wang", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Dongmei Zhang", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Distinguished Scientist" }, { "name": "Guyue Liu", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Huiqiang Jiang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "RSDE" }, { "name": "jiaxu qian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Intern" }, { "name": "Lili Qiu", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Principal Researcher" }, { "name": "Qi Zhang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Qingwei Lin", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Sr. Principal Researcher" }, { "name": "Saravan Rajmohan", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Partner Director of Applied Research " }, { "name": "Shiqi Jiang", "gender": "Male", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Suman Banerjee", "gender": "Male", "institution": "UW-Madison", "country": "US", "position": "Full Professor" }, { "name": "Tianjun Mao", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Ting Cao", "gender": "unknown", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Xufang Luo", "gender": "Female", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" }, { "name": "Yifan Yang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Yu Kang", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuqing Yang", "gender": "unknown", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 12, "n_ref": 30, "n_ref_all": 44, "n_fig": 5, "n_tab": 7, "L_tab": 2953, "n_element_tab": 241, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 993, "n_element_tab_1": 102, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 119489, "len_all_1": 65780, "len_abs": 1134, "len_title": 73, "len_sents": 33868, "len_sents_1": 32021, "n_sents": 230, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1141, "title": "Zoomer: Enhancing MLLM Performance with Adaptive Image Focus Optimization", "abs": "Recent advancements in multimodal large language models (MLLMs) have broadened the scope of vision-language tasks, excelling in applications like image captioning and interactive question-answering. However, these models struggle with accurately processing visual data, particularly in tasks requiring precise object recognition and fine visual details.\nStringent token limits often result in the omission of critical information, hampering performance. To address these limitations, we introduce Zoomer, a novel visual prompting mechanism designed to enhance MLLM performance while preserving essential visual details within token limits. Zoomer features three key innovations: a prompt-aware strategy that dynamically highlights relevant image regions, a spatial-preserving orchestration schema that maintains object integrity, and a budget-aware prompting method that balances global context with crucial visual details. \nComprehensive evaluations across multiple datasets demonstrate that Zoomer consistently outperforms baseline methods, achieving up to a $26.9\\%$ improvement in accuracy while significantly reducing token consumption.", "keywords": [ "Multimodal", "MLLM", "Prompt Engineering", "Efficient", "Token Compression" ], "rating_list": [ 3, 6, 3, 3, 5 ], "soundness_list": [ 1, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 1, 3, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "SO0manOwUF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Behzad Dariush", "gender": "Male", "institution": "Honda Research Institute USA", "country": "US", "position": "Principal Researcher" }, { "name": "Huaizhi Qu", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Jie Peng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Kwonjoon Lee", "gender": "Male", "institution": "Honda Research Institute USA", "country": "US", "position": "Sr Research Scientist" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Xinyu Zhao", "gender": "unknown", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 5, "n_ref_uni": 62, "n_ref": 107, "n_ref_all": 125, "n_fig": 3, "n_tab": 7, "L_tab": 5003, "n_element_tab": 592, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2654, "n_element_tab_1": 121, "formula_len_all": 825, "formula_len_all_1": 657, "len_all": 172935, "len_all_1": 60724, "len_abs": 1818, "len_title": 120, "len_sents": 29005, "len_sents_1": 27067, "n_sents": 217, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1730, "title": "$\\texttt{UQ-Merge}$: UNCERTAINTY GUIDED MULTIMODAL LARGE LANGUAGE MODEL MERGING", "abs": "Multimodal Large Language Models (MLLMs) have gained increasing popularity as a promising framework for leveraging the strong language reasoning capabilities in the vision-language domain. Given a wide range of MLLMs, model merging potentially offers a cheap way to aggregate their diverse knowledge into a single MLLM. However, directly plug-in existing model merging approaches often leads to suboptimal performance due to ($1$) inclusion of harmful models that have over-confident predictions in the target task; ($2$) the lack of specialized designs for vision-language tasks. To tackl these pain points, we conduct pioneering investigations to dissect the merging procedures and propose an uncertainty-guided MLLM merging algorithm, $\\textit{i.e.}$, $\\texttt{UQ-Merge}$, which $i)$ identifies beneficial candidates for merging, $ii)$ determines the merge order and the number of helpful candidates, and $iii)$ performs appropriate merging. Within our framework, we consider uncertainty quantification on both text and vision inputs to examine the MLLM prediction confidence, and then decide whether and when a MLLM needs to be included. It is worth mentioning that our vision-language uncertainty quantification does not require access to sample labels, making it more practical in various scenarios. Extensive experiments consistently demonstrate the superior MLLM merging performance of $\\texttt{UQ-Merge}$ in both held-in and held-out vision-language benchmarks. For example, compared to existing state-of-the-art merging methods, $\\texttt{UQ-Merge}$ brings substantial performance improvements of up to $44.3\\%$ on average accuracy in $12$ datasets. Codes are available at https://anonymous.4open.science/r/UQ-Merge-7CD7.", "keywords": [ "Multimodal Large Language Model", "Uncertainty Quantification", "Model Merging" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "SNsdlEp3Ne", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Chengju Liu", "gender": "Female", "institution": "Tongji University", "country": "CN", "position": "Full Professor" }, { "name": "Mengxian Hu", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "PhD student" }, { "name": "Minghao Zhu", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "PhD student" }, { "name": "Qijun Chen", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Full Professor" }, { "name": "Qingqing Yan", "gender": "Female", "institution": "Tongji University", "country": "", "position": "PhD student" }, { "name": "Shu Li", "gender": "Male", "institution": "Shanghai Jizhi Youchuang Technology Co., Ltd.", "country": "", "position": "Researcher" }, { "name": "Xun Zhou", "gender": "Male", "institution": "Tongji University", "country": "", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 21, "n_ref_uni": 32, "n_ref": 72, "n_ref_all": 93, "n_fig": 5, "n_tab": 10, "L_tab": 4284, "n_element_tab": 462, "n_fig_1": 2, "n_tab_1": 9, "L_tab_1": 4035, "n_element_tab_1": 436, "formula_len_all": 1170, "formula_len_all_1": 1041, "len_all": 132206, "len_all_1": 62250, "len_abs": 1401, "len_title": 126, "len_sents": 34796, "len_sents_1": 23788, "n_sents": 271, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1407, "title": "Efficient Text-driven Human Motion Generation via Latent Consistency Training", "abs": "Consistency models excel at few-step inference in generative tasks across various scenarios, but typically rely on pre-trained diffusion model distillation, involving additional training costs and performance limitations. \nIn this paper, we propose a motion latent consistency training framework that learns directly from data rather than distillation for efficient and text-controllable human motion generation.\nFor representation optimization, we design a motion autoencoder with quantization constraints that enable concise and bounded motion latent representations.\nFocusing on conditional generation, we construct a classifier-free guidance (CFG) format with an additional unconditional loss function that extends the CFG technique from the inference phase to the training phase for conditionally guided consistency training.\nWe further propose a clustering guidance module to provide additional references to the solution distribution at minimal query cost.\nBy combining these enhancements, we achieve stable and consistent training in non-pixel modality and latent representation spaces for the first time.\nExperiments in benchmarks demonstrate that our method significantly outperforms traditional consistency distillation methods with reduced training cost, and enhances the consistency model to perform comparably to state-of-the-art models with lower inference cost.\nOur code will be open source.", "keywords": [ "motion generation", "diffusion", "consistency training" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "SNNdmfqWFu", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Amitai Yacobi", "gender": "unknown", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" }, { "name": "Ofir Lindenbaum", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "Assistant Professor" }, { "name": "Uri Shaham", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 20, "n_formula_1": 15, "n_ref_uni": 33, "n_ref": 71, "n_ref_all": 87, "n_fig": 4, "n_tab": 7, "L_tab": 2651, "n_element_tab": 177, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1601, "n_element_tab_1": 83, "formula_len_all": 1236, "formula_len_all_1": 811, "len_all": 134505, "len_all_1": 67782, "len_abs": 1602, "len_title": 126, "len_sents": 43230, "len_sents_1": 30338, "n_sents": 359, "n_sents_1": 235, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1614, "title": "SpecRaGE: Robust and Generalizable Multi-view Spectral Representation Learning", "abs": "Multi-view representation learning (MvRL) has garnered substantial attention in recent years, driven by the increasing demand for applications that can effectively process and analyze data from multiple sources. In this context, graph Laplacian-based MvRL methods have demonstrated remarkable success in representing multi-view data. However, these methods often struggle with generalization to new data and face challenges with scalability. Moreover, in many practical scenarios, multi-view data is contaminated by noise or outliers. In such cases, modern deep-learning-based MvRL approaches that rely on alignment or contrastive objectives can lead to misleading results, as they may impose incorrect consistency between clear and corrupted data sources. We introduce *SpecRaGE*, a novel fusion-based framework that integrates the strengths of graph Laplacian methods with the power of deep learning to overcome these challenges. SpecRage uses neural networks to learn parametric mapping that approximates a joint diagonalization of graph Laplacians. This solution bypasses the need for alignment while enabling generalizable and scalable learning of informative and meaningful representations. Moreover, it incorporates a meta-learning fusion module that dynamically adapts to data quality, ensuring robustness against outliers and noisy views. Our extensive experiments demonstrate that SpecRaGE outperforms state-of-the-art methods, particularly in scenarios with data contamination, paving the way for more reliable and efficient multi-view learning. Our code will be made publicly available upon acceptance.", "keywords": [ "Multi-view Representation Learning", "Graph Laplacian", "Joint Diagonalization" ], "rating_list": [ 3, 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 1, 1, 1, 2, 2 ], "confidence_list": [ 2, 5, 4, 4, 5 ] }, { "paper_id": "SMlVEeoSyI", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Mubarak Shah", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Nyle Siddiqui", "gender": "unknown", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Rohit Gupta", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Swetha Sirnam", "gender": "Female", "institution": "University of Central Florida", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 68, "n_ref_all": 84, "n_fig": 5, "n_tab": 3, "L_tab": 2510, "n_element_tab": 193, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 535, "formula_len_all_1": 392, "len_all": 125399, "len_all_1": 61667, "len_abs": 1918, "len_title": 125, "len_sents": 33730, "len_sents_1": 30975, "n_sents": 202, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1930, "title": "StretchySnake: Flexible VideoMamba for Short and Long-Form Action Recognition", "abs": "State space models (SSMs) have very recently been introduced as an alternative deep architecture to transformers, exhibiting competitive or superior performance across various language and vision tasks. However, both SSMs and transformers share certain limitations in the vision domain, namely spatio-temporal inflexibility. Traditionally, deep video models are trained on a fixed resolution and number of frames, often arbitrarily chosen as a trade-off between performance and computational cost. Changing the resolution and/or number of frames a model can ingest usually requires retraining the model, while avoiding re-training by variably changing the weights of a trained model leads to significantly reduced test accuracy. In this paper, we introduce a spatio-temporal flexible training method that encourages a single set of learned weights to adapt well to any input resolution or video length. We achieve this by simply randomly changing the spatial and temporal resolutions of a video during training, and dynamically interpolating the model's weights accordingly. This single change in training not only allows for one model to be applied to both short and long video understanding tasks alike, but also allows for user-specific tailoring of computational cost. We propose and evaluate $5$ different spatio-temporal flexible training methods to find the optimal type for training a video SSM. We then evaluate our best flexibly-trained SSM, which we call StretchySnake, across a variety of short- and long-form action recognition evaluation protocols, such as video retrieval, fine-tuning, and linear probing, and massively outperform the same vanilla video SSM trained in a standard fashion by up to $28$% in some cases. Therefore, our training method can be used as a simple drop-in training technique for any SSM-based video models to strongly improve performance and instill spatio-temporal and compute flexibility.", "keywords": [ "Representation Learning", "Video Understanding", "State Space Models" ], "rating_list": [ 6, 6, 1 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 1 ], "contribution_list": [ 3, 3, 1 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "SMZqIOSdlN", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Dexuan Ding", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "Undergrad student" }, { "name": "Lei Wang", "gender": "Male", "institution": "Griffith University", "country": "AU", "position": "Postdoc" }, { "name": "Liyun Zhu", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "MS student" }, { "name": "Piotr Koniusz", "gender": "unknown", "institution": "University of New South Wales", "country": "AU", "position": "Associate Professor" }, { "name": "Tom Gedeon", "gender": "Male", "institution": "Curtin University of Technology", "country": "AU", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 12, "n_ref_uni": 36, "n_ref": 73, "n_ref_all": 90, "n_fig": 7, "n_tab": 13, "L_tab": 4543, "n_element_tab": 386, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 372, "n_element_tab_1": 66, "formula_len_all": 2085, "formula_len_all_1": 817, "len_all": 173721, "len_all_1": 65593, "len_abs": 1387, "len_title": 124, "len_sents": 57954, "len_sents_1": 31738, "n_sents": 432, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1392, "title": "Learnable Expansion of Graph Operators for Multi-Modal Feature Fusion", "abs": "In computer vision tasks, features often come from diverse representations, domains (e.g., indoor and outdoor), and modalities (e.g., text, images, and videos). Effectively fusing these features is essential for robust performance, especially with the availability of powerful pre-trained models like vision-language models. However, common fusion methods, such as concatenation, element-wise operations, and non-linear techniques, often fail to capture structural relationships, deep feature interactions, and suffer from inefficiency or misalignment of features across domains or modalities. In this paper, we shift from high-dimensional feature space to a lower-dimensional, interpretable graph space by constructing relationship graphs that encode feature relationships at different levels, e.g., clip, frame, patch, token, etc. To capture deeper interactions, we use graph power expansions and introduce a learnable graph fusion operator to combine these graph powers for more effective fusion. Our approach is relationship-centric, operates in a homogeneous space, and is mathematically principled, resembling element-wise relationship score aggregation via multilinear polynomials. We demonstrate the effectiveness of our graph-based fusion method on video anomaly detection, showing strong performance across multi-representational, multi-modal, and multi-domain feature fusion tasks.", "keywords": [ "feature fusion", "video anomaly detection", "graph learning", "multi-modal learning" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "SMYEApLhyx", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Carlos R Ponce", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "Giordano Ramos-Traslosheros", "gender": "unknown", "institution": "Harvard Medical School, Harvard University", "country": "US", "position": "Postdoc" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 72, "n_ref_all": 107, "n_fig": 22, "n_tab": 2, "L_tab": 31, "n_element_tab": 4, "n_fig_1": 18, "n_tab_1": 1, "L_tab_1": 31, "n_element_tab_1": 4, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 137295, "len_all_1": 63643, "len_abs": 2018, "len_title": 121, "len_sents": 53448, "len_sents_1": 32278, "n_sents": 371, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 2030, "title": "Functional segregation of inputs in artificial neural networks for vision", "abs": "One of the main organizational principles of artificial and biological intelligence systems is their reliance on signed inputs: positive and negative weights in artificial networks, and excitatory and inhibitory synapses in the brain. However, little is known about the role of inhibitory activity in high-level visual cortex such as inferotemporal cortex, or how artificial neural networks (ANNs) trained for object recognition segregate their learned representations into positive and negative weights.\nHere, we dissected high-level visual mechanisms in ANNs trained with ImageNet. We investigated how learned representations of ANN classification units depended on their positive or negative inputs using ablation experiments and feature visualization. We found that unit representations changed more when ablating positive- vs. negative inputs. Object-related features were abolished when ablating positive inputs, while still preserving background textures. This effect was more pronounced in adversarially trained robust networks. This segregation persisted in networks trained with unsupervised learning, but was not present in a ResNet18 trained with Tanh instead of ReLU.\nWe found a consistent functional segregation when we trained models to replicate the activity of neurons in monkey visual cortex, across the ventral stream (V1, V4, and IT). Feature visualization of the neuron models produced images containing local features preferred by actual neurons. Analogous to units trained for classification, the learned representations of units trained to simulate neurons changed more upon ablating positive than negative inputs. We conclude that ANNs for classification segregate object or foreground information into the positive weights, with background or contextual information into the negative weights, in their last layer before softmax. These results hint at the relevance of signal rectification and inhibition into shaping feature selectivity in the primate ventral stream, a hypothesis we are testing in vivo.", "keywords": [ "ventral stream", "circuit mechanisms", "interpretability", "deep learning", "visual system", "excitation inhibition", "neuroscience", "closed-loop optimization", "ablation" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "SMKgohbroH", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Congcong Wen", "gender": "unknown", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "Hao Huang", "gender": "Male", "institution": "New York University", "country": "AE", "position": "Postdoc" }, { "name": "Mengyu Wang", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" }, { "name": "Min Shi", "gender": "Male", "institution": "University of Louisiana at Lafeyette", "country": "US", "position": "Assistant Professor" }, { "name": "Minghan Li", "gender": "Female", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Yan Luo", "gender": "Male", "institution": "Broad Institute of MIT and Harvard", "country": "US", "position": "Postdoc" }, { "name": "Yi Fang", "gender": "Male", "institution": "New York University", "country": "US", "position": "Associate Professor" } ], "n_formula": 27, "n_formula_1": 11, "n_ref_uni": 30, "n_ref": 38, "n_ref_all": 47, "n_fig": 3, "n_tab": 2, "L_tab": 3829, "n_element_tab": 47, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 4806, "n_element_tab_1": 320, "formula_len_all": 1205, "formula_len_all_1": 456, "len_all": 144446, "len_all_1": 69624, "len_abs": 1379, "len_title": 145, "len_sents": 35635, "len_sents_1": 30974, "n_sents": 258, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 96, "L_abs": 1316, "title": "Challenge Me: Enhancing Conversational Consistency of LLMs by Learning with Questioning Feedback", "abs": "As Large Language Models (LLMs) increasingly integrate into critical decision-support systems, ensuring their conversational consistency becomes paramount for reliable and trustworthy AI-assisted services, especially in high-stakes domains such as healthcare and legal advice. In this work, we study the critical issue of conversational inconsistency in LLMs, where models provide contradictory information across multiple dialogue turns. We introduce a novel Conversationally Consistent Supervised Fine-Tuning (CC-SFT) method that explicitly accounts for two-turn conversations. Our approach combines a first-round loss, a second-round loss, and a consistency loss based on Wasserstein distance to encourage coherent responses across turns. We evaluate our method on three diverse datasets (OpenBookQA, GSM8K, and MedQA-USMLE) using three LLMs (Llama v3.1, Mistral AI, and Gemma). Experimental results demonstrate that CC-SFT significantly reduces conversational inconsistency compared to standard fine-tuning, with lower flipping rates and improved accuracy in second-round responses. We provide theoretical convergence guarantees for our method and analyze the impact of the consistency loss coefficient. Our code is publicly available at \\url{https://github.com/anonymous4science/llm_conversational_consistency}.", "keywords": [ "AI Safety", "LLM", "Conversational Consistency" ], "rating_list": [ 1, 5, 1, 5 ], "soundness_list": [ 1, 3, 1, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 1, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "SMK0f8JoKF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Davide Evangelista", "gender": "Male", "institution": "University of Bologna", "country": "IT", "position": "Assistant Professor" }, { "name": "Emanuele La Malfa", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Michael J. Wooldridge", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Michele Lombardi", "gender": "Male", "institution": "University of Bologna", "country": "IT", "position": "Associate Professor" }, { "name": "Samuele Marro", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "X. Angelo Huang", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" } ], "n_formula": 35, "n_formula_1": 16, "n_ref_uni": 40, "n_ref": 59, "n_ref_all": 101, "n_fig": 28, "n_tab": 3, "L_tab": 1500, "n_element_tab": 514, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 88, "n_element_tab_1": 7, "formula_len_all": 2325, "formula_len_all_1": 657, "len_all": 163906, "len_all_1": 52917, "len_abs": 850, "len_title": 89, "len_sents": 43804, "len_sents_1": 24106, "n_sents": 391, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 855, "title": "Language Models Are Implicitly Continuous", "abs": "Language is typically modelled with discrete sequences. However, the most successful approaches to language modelling, namely neural networks, are continuous and smooth function approximators.\nIn this work, we show that Transformer-based language models implicitly learn to represent sentences as continuous-time functions defined over a continuous input space. \nThis phenomenon occurs in most state-of-the-art Large Language Models (LLMs), including Llama2, Llama3, Phi3, Gemma, Gemma2, and Mistral, and suggests that LLMs reason about language in ways that fundamentally differ from humans.\nOur work formally extends Transformers to capture the nuances of time and space continuity in both input and output space.\nOur results challenge the traditional interpretation of how LLMs understand language, with several linguistic and engineering implications.", "keywords": [ "llm", "continuity", "spatiotemporal transformers", "linguistics" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "SM1guXel3E", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Cheng Tan", "gender": "Male", "institution": "Zhejiang University & Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Di Wu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Juanxi Tian", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Intern" }, { "name": "Siyuan Li", "gender": "Male", "institution": "Westlake University & Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Weiyang Jin", "gender": "Male", "institution": "New York University", "country": "US", "position": "Intern" }, { "name": "Zedong Wang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zicheng Liu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Stan Z. Li", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Chair Professor" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 57, "n_ref": 152, "n_ref_all": 200, "n_fig": 8, "n_tab": 20, "L_tab": 10848, "n_element_tab": 1022, "n_fig_1": 7, "n_tab_1": 7, "L_tab_1": 6930, "n_element_tab_1": 539, "formula_len_all": 672, "formula_len_all_1": 133, "len_all": 219116, "len_all_1": 78904, "len_abs": 1450, "len_title": 127, "len_sents": 54294, "len_sents_1": 32443, "n_sents": 354, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1457, "title": "OpenMixup: Open Mixup Toolbox and Benchmark for Visual Representation Learning", "abs": "Mixup augmentation has emerged as a widely used technique for improving the generalization ability of deep neural networks (DNNs). However, the lack of standardized implementations and benchmarks has impeded recent progress, resulting in poor reproducibility, unfair comparisons, and conflicting insights. In this paper, we introduce OpenMixup, the first mixup augmentation codebase and benchmark for visual representation learning. Specifically, we train 18 representative mixup baselines from scratch and rigorously evaluate them across 11 image datasets of varying scales and granularity, ranging from fine-grained scenarios to complex non-iconic scenes. We also open-source our modular codebase including a collection of popular vision backbones, optimization strategies, and analysis toolkits, which not only supports the benchmarking but enables broader mixup applications beyond classification, such as self-supervised learning and regression tasks. Through experiments and empirical analysis, we gain observations and insights on mixup performance-efficiency trade-offs, generalization, and optimization behaviors, and thereby identify preferred choices for different needs. To the best of our knowledge, OpenMixup has facilitated several recent studies. We believe this work can further advance reproducible mixup augmentation research and thereby lay a solid ground for future progress in the community. The source code will be publicly available.", "keywords": [ "Data Augmentation", "Image Classification", "Mixup", "Vision Transformer", "Benchmark" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "SLufnMLhbv", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Bartosz Cywi\u0144ski", "gender": "Male", "institution": "Warsaw University of Technology", "country": "PL", "position": "PhD student" }, { "name": "Bart\u0142omiej Twardowski", "gender": "Male", "institution": "IDEAS NCBR", "country": "PL", "position": "Researcher" }, { "name": "Kamil Deja", "gender": "unknown", "institution": "IDEAS NCBR", "country": "PL", "position": "Postdoc" }, { "name": "Tomasz Trzcinski", "gender": "Male", "institution": "Warsaw University of Technology", "country": "PL", "position": "Full Professor" }, { "name": "\u0141ukasz Kuci\u0144ski", "gender": "Male", "institution": "IDEAS NCBR", "country": "PL", "position": "Senior Research Scientist" } ], "n_formula": 17, "n_formula_1": 12, "n_ref_uni": 51, "n_ref": 87, "n_ref_all": 106, "n_fig": 7, "n_tab": 6, "L_tab": 2769, "n_element_tab": 214, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2773, "n_element_tab_1": 139, "formula_len_all": 1429, "formula_len_all_1": 732, "len_all": 159735, "len_all_1": 68487, "len_abs": 969, "len_title": 112, "len_sents": 46938, "len_sents_1": 28736, "n_sents": 319, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 974, "title": "GUIDE: Guidance-based Incremental Learning with Diffusion Models", "abs": "We introduce GUIDE, a novel continual learning approach that directs diffusion models to rehearse samples at risk of being forgotten. \nExisting generative strategies combat catastrophic forgetting by randomly sampling rehearsal examples from a generative model. Such an approach contradicts buffer-based approaches where sampling strategy plays an important role.\nWe propose to bridge this gap by incorporating classifier guidance into the diffusion process to produce rehearsal examples specifically targeting information forgotten by a continuously trained model. This approach enables the generation of samples from preceding task distributions, which are more likely to be misclassified in the context of recently encountered classes. \nOur experimental results show that GUIDE significantly reduces catastrophic forgetting,\noutperforming conventional random sampling approaches and surpassing recent state-of-the-art methods in continual learning with generative replay.", "keywords": [ "continual learning", "diffusion models" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 1, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "SLUr06QUuw", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Josef Kittler", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Full Professor" }, { "name": "Rui Wang", "gender": "Male", "institution": "Jiangnan University", "country": "CN", "position": "Lecturer" }, { "name": "Tianyang Xu", "gender": "Male", "institution": "Jiangnan University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaojun Wu", "gender": "Male", "institution": "Jiangnan University", "country": "CN", "position": "Full Professor" }, { "name": "Yusheng Bao", "gender": "Male", "institution": "Jiangnan University", "country": "CN", "position": "MS student" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 20, "n_ref": 33, "n_ref_all": 47, "n_fig": 2, "n_tab": 4, "L_tab": 972, "n_element_tab": 106, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2134, "n_element_tab_1": 117, "formula_len_all": 1088, "formula_len_all_1": 1230, "len_all": 106618, "len_all_1": 61531, "len_abs": 1183, "len_title": 156, "len_sents": 34476, "len_sents_1": 27375, "n_sents": 234, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 108, "L_abs": 1195, "title": "SymCL: Riemannian Contrastive Learning on the Symmetric Positive Definite Manifold for Visual Classification", "abs": "Symmetric Positive Definite (SPD) matric has been proven to be an effective feature descriptor in the realm of artificial intelligence, as it can encode spatiotemporal statistical information of data on a curved Riemannian manifold, \\textit{i.e.}, SPD manifold. Although existing Riemannian neural networks have demonstrated superiority in many scientific fields, the inherent reliance on labels within supervised learning renders them susceptible to label errors. Besides, it is insufficient to depend solely on labels to learn effective feature distributions in some complicated data scenarios. Drawing inspiration from the considerable achievements of contrastive learning (CL) across diverse tasks, we extend the conventional CL paradigm to the context of SPD manifolds, which we denote SymCL, paving the way for a novel approach in SPD matrix-based visual classification. Furthermore, we inject a Riemannian triplet loss-based Riemannian metric learning (RML) into the designed SPD manifold CL framework for the sake of improving the discrimination of the learned geometric representations. Extensive experimental results on four datasets verify the effectiveness of the proposed algorithm.", "keywords": [ "Learning with Manifolds", "Self-Supervised Learning", "Classification" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 4, 2, 5, 5 ] }, { "paper_id": "SLDqCpHPuP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Dan Klein", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Evonne Ng", "gender": "Not Specified", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Lea M\u00fcller", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Sanjay Subramanian", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Shiry Ginosar", "gender": "Female", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Assistant Professor" }, { "name": "trevor darrell", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department", "country": "US", "position": "Professor" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 44, "n_ref": 89, "n_ref_all": 113, "n_fig": 9, "n_tab": 2, "L_tab": 665, "n_element_tab": 104, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 565, "n_element_tab_1": 93, "formula_len_all": 529, "formula_len_all_1": 498, "len_all": 184775, "len_all_1": 69981, "len_abs": 579, "len_title": 80, "len_sents": 48109, "len_sents_1": 31548, "n_sents": 432, "n_sents_1": 279, "venue": "ICLR2025", "decision": "withdraw", "L_title": 32, "L_abs": 1010, "title": "Pose Priors from Language Models", "abs": "We present a zero-shot pose optimization method that enforces accurate physical contact constraints when estimating the 3D pose of humans. Our central insight is that since language is often used to describe physical interaction, large pretrained text-based models can act as priors on pose estimation.\nWe can thus leverage this insight to improve pose estimation by converting natural language descriptors, generated by a large multimodal model (LMM), into tractable losses to constrain the 3D pose optimization. Despite its simplicity, our method produces surprisingly compelling pose reconstructions of people in close contact, correctly capturing the semantics of the social and physical interactions. We demonstrate that our method rivals more complex state-of-the-art approaches that require expensive human annotation of contact points and training specialized models. Moreover, unlike previous approaches, our method provides a unified framework for resolving self-contact and person-to-person contact.", "keywords": [ "pose estimation", "language models", "multimodal models" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "SKl8zzi4Mn", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Cesare Molinari", "gender": "Male", "institution": "University of Genoa", "country": "", "position": "Researcher" }, { "name": "Cristian Jesus Vega Cere\u00f1o", "gender": "Male", "institution": "University of Genoa", "country": "IT", "position": "PhD student" }, { "name": "Hippolyte Labarri\u00e8re", "gender": "unknown", "institution": "University of Genoa", "country": "IT", "position": "Postdoc" }, { "name": "Lorenzo Rosasco", "gender": "Male", "institution": "Universit\u00e0 degli Studi di Genova", "country": "IT", "position": "Full Professor" }, { "name": "Silvia Villa", "gender": "Female", "institution": "Universit\u00e0 degli Studi di Genova", "country": "", "position": "Associate Professor" } ], "n_formula": 52, "n_formula_1": 40, "n_ref_uni": 19, "n_ref": 55, "n_ref_all": 60, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 755, "n_element_tab_1": 8, "formula_len_all": 3557, "formula_len_all_1": 2863, "len_all": 105540, "len_all_1": 54902, "len_abs": 766, "len_title": 105, "len_sents": 28774, "len_sents_1": 19951, "n_sents": 252, "n_sents_1": 161, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 769, "title": "Optimization Insights into Deep Diagonal Linear Networks", "abs": "Overparameterized models trained with (stochastic) gradient descent are ubiquitous in modern machine learning. These large models achieve unprecedented performance on test data, but their theoretical understanding is still limited. In this paper, we take a step towards filling this gap by adopting an optimization perspective. More precisely, we study the implicit regularization properties of the gradient flow \u201calgorithm\u201d for estimating the parameters of a deep diagonal neural network. Our main contribution is showing that this gradient flow induces a mirror flow dynamic on the model, meaning that it is biased towards a specific solution of the problem depending on the initialization of the network. Along the way, we prove several properties of the trajectory.", "keywords": [ "Diagonal Linear Network", "Overparameterization", "Implicit Bias", "Mirror Flow" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 5, 2, 5, 4 ] }, { "paper_id": "SKW10XJlAI", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Gao Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Kaifeng Lyu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Mengdi Wang", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Associate Professor" }, { "name": "Rui Lu", "gender": "Male", "institution": "Department of Automation, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Runzhe Wang", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Xitai Jiang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" } ], "n_formula": 53, "n_formula_1": 12, "n_ref_uni": 28, "n_ref": 159, "n_ref_all": 166, "n_fig": 6, "n_tab": 2, "L_tab": 824, "n_element_tab": 9, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6006, "formula_len_all_1": 835, "len_all": 158380, "len_all_1": 63098, "len_abs": 1525, "len_title": 134, "len_sents": 55791, "len_sents_1": 29290, "n_sents": 439, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1531, "title": "Towards Understanding Text Hallucination of Diffusion Models via Local Generation Bias", "abs": "Score-based diffusion models have achieved incredible performance in generating realistic images, audio, and video data. While these models produce high-quality samples with impressive details, they often introduce unrealistic artifacts, such as distorted fingers or hallucinated texts with no meaning. This paper focuses on textual hallucinations, where diffusion models correctly generate individual symbols but assemble them in a nonsensical manner. Through experimental probing, we consistently observe that such phenomenon is attributed it to the network's local generation bias. Denoising networks tend to produce outputs that rely heavily on highly correlated local regions, particularly when different dimensions of the data distribution are nearly pairwise independent. This behavior leads to a generation process that decomposes the global distribution into separate, independent distributions for each symbol, ultimately failing to capture the global structure, including underlying grammar. Intriguingly, this bias persists across various denoising network architectures including MLP and transformers which have the structure to model global dependency. These findings also provide insights into understanding other types of hallucinations, extending beyond text, as a result of implicit biases in the denoising models. Additionally, we theoretically analyze the training dynamics for a specific case involving a two-layer MLP learning parity points on a hypercube, offering an explanation of its underlying mechanism.", "keywords": [ "Diffusion model", "Deep learning theory", "generative model", "Hallucination" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 2, 3 ] }, { "paper_id": "SJqoP0rp8w", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Hengshuang Zhao", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Jiaqi Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Long Xing", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Mengchen Zhang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Tong Wu", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Xiaoyang Wu", "gender": "Male", "institution": "the University of Hong Kong, University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xihui Liu", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Yunhan Yang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zhangyang Qi", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 15, "n_ref": 33, "n_ref_all": 54, "n_fig": 11, "n_tab": 3, "L_tab": 763, "n_element_tab": 85, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 688, "n_element_tab_1": 62, "formula_len_all": 349, "formula_len_all_1": 278, "len_all": 153179, "len_all_1": 54018, "len_abs": 303, "len_title": 47, "len_sents": 35987, "len_sents_1": 23388, "n_sents": 332, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1754, "title": "Tailor3D: Customized 3D Assets Editing and Generation with Dual-Side Images", "abs": "Recent advances in 3D AIGC have shown promise in directly creating 3D objects from text and images, offering significant cost savings in animation and product design. However, detailed edit and customization of 3D assets remains a long-standing challenge. Specifically, 3D Generation methods lack the ability to follow finely detailed instructions as precisely as their 2D image creation counterparts. Imagine you can get a toy through 3D AIGC but with undesired accessories and dressing. To tackle this challenge, we propose a novel pipeline called Tailor3D, which swiftly creates customized 3D assets from editable dual-side images. We aim to emulate a tailor's ability to locally change objects or perform overall style transfer. Unlike creating 3D assets from multiple views, using dual-side images eliminates conflicts on overlapping areas that occur when editing individual views. Specifically, it begins by editing the front view, then generates the back view of the object through multi-view diffusion. Afterward, it proceeds to edit the back views. Finally, a Dual-sided LRM is proposed to seamlessly stitch together the front and back 3D features, akin to a tailor sewing together the front and back of a garment. The Dual-sided LRM rectifies imperfect consistencies between the front and back views, enhancing editing capabilities and reducing memory burdens while seamlessly integrating them into a unified 3D representation with the LoRA Triplane Transformer. Experimental results demonstrate Tailor3D's effectiveness across various 3D generation and editing tasks, including 3D generative fill and style transfer. It provides a user-friendly, efficient solution for editing 3D assets, with each editing step taking only seconds to complete.", "keywords": [ "3D Object Editing", "3D Generative Models" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 4, 2, 2, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 2, 5, 3, 3 ] }, { "paper_id": "SIzjhS9kEF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Sean M. Hendryx", "gender": "Male", "institution": "Scale AI", "country": "", "position": "Engineering Manager, Machine Learning" }, { "name": "Vaskar Nath", "gender": "Male", "institution": "Scale AI", "country": "US", "position": "Researcher" }, { "name": "Mohit R", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 51, "n_ref_all": 71, "n_fig": 5, "n_tab": 6, "L_tab": 1656, "n_element_tab": 91, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 615, "n_element_tab_1": 63, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 142253, "len_all_1": 58180, "len_abs": 1665, "len_title": 95, "len_sents": 41940, "len_sents_1": 27822, "n_sents": 331, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1671, "title": "Revisiting the Superficial Alignment Hypothesis", "abs": "The Superficial Alignment Hypothesis posits that almost all of a language model's abilities and knowledge are learned during pre-training, while post-training is about giving a model the right style and format. We re-examine these claims by empirically studying the scaling behavior of post-training with increasing finetuning examples and evaluating them using objective task-specific standardized benchmarks. Through experiments with the Llama-3, Mistral, and Llama-2 model families of multiple sizes, we observe that, similar to the pre-training scaling laws, post-training task performance scales as a power law against the number of finetuning examples. This power law relationship holds across a broad array of capabilities, including mathematical reasoning, coding, instruction following, and multihop-reasoning. In addition, for tasks like math and multihop reasoning, we observe that a handful of examples merely align the model stylistically but do not saturate performance on the benchmarks. Model performance is instead correlated with its reasoning ability and it improves significantly with more examples, illustrating the need for holistic evaluation programs leveraging objective benchmarks in addition to measurement of alignment to human preferences. We also observe that language models are not necessarily limited to using knowledge learned during pre-training. With appropriate post-training, a model's ability to integrate new knowledge greatly improves on downstream tasks like multihop question-answering. Taken together, these results shed new light on the Superficial Alignment Hypothesis, suggesting that it is, at best, an over-simplification.", "keywords": [ "Large Language Models", "Alignment", "Artificial Intelligence", "Supervised Finetuning", "Post-training", "Pre-training", "Scaling Laws", "Evaluation", "Reasoning" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "SIuD7CySb4", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Harri L\u00e4hdesm\u00e4ki", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Associate Professor" }, { "name": "Manuel Haussmann", "gender": "unknown", "institution": "University of Southern Denmark - SDU", "country": "DK", "position": "Postdoc" }, { "name": "Siddharth Ramchandran", "gender": "unknown", "institution": "Aalto University", "country": "", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 13, "n_ref_uni": 36, "n_ref": 121, "n_ref_all": 149, "n_fig": 12, "n_tab": 2, "L_tab": 242, "n_element_tab": 18, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 281, "n_element_tab_1": 46, "formula_len_all": 2887, "formula_len_all_1": 1212, "len_all": 173192, "len_all_1": 72533, "len_abs": 1213, "len_title": 139, "len_sents": 52981, "len_sents_1": 32681, "n_sents": 413, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1217, "title": "High-Dimensional Bayesian Optimisation with Gaussian Process Prior Variational Autoencoders", "abs": "Bayesian optimisation (BO) using a Gaussian process (GP)-based surrogate model is a powerful tool for solving black-box optimisation problems but does not scale well to high-dimensional data. Previous works have proposed to use variational autoencoders (VAEs) to project high-dimensional data onto a low-dimensional latent space and to implement BO in the inferred latent space. In this work, we propose a conditional generative model for efficient high-dimensional BO that uses a GP surrogate model together with GP prior VAEs. A GP prior VAE extends the standard VAE by conditioning the generative and inference model on auxiliary covariates, capturing complex correlations across samples with a GP. Our model incorporates the observed target quantity values as auxiliary covariates learning a structured latent space that is better suited for the GP-based BO surrogate model. It handles partially observed auxiliary covariates using a unifying probabilistic framework and can also incorporate additional auxiliary covariates that may be available in real-world applications. We demonstrate that our method improves upon existing latent space BO methods on simulated datasets as well as on commonly used benchmarks.", "keywords": [ "Variational autoencoders", "Gaussian processes", "Bayesian optimisation" ], "rating_list": [ 5, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 4, 3 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "SIdA3s754H", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Guang Cheng", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "Full Professor" }, { "name": "Xiaowu Dai", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "Yuantong Li Li", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PHD" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 54, "n_ref": 71, "n_ref_all": 78, "n_fig": 2, "n_tab": 4, "L_tab": 298, "n_element_tab": 13, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1403, "formula_len_all_1": 841, "len_all": 144381, "len_all_1": 80300, "len_abs": 1439, "len_title": 73, "len_sents": 38507, "len_sents_1": 34164, "n_sents": 272, "n_sents_1": 248, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1452, "title": "Context-Aware Online Recommendation with Bayesian Incentive Compatibility", "abs": "Recommender systems play a crucial role in internet economies by connecting users with relevant products or services. However, designing effective recommender systems faces two key challenges: (1) the exploration-exploitation tradeoff in balancing new product exploration against exploiting known preferences, and (2) context-aware Bayesian incentive compatibility in accounting for users' heterogeneous preferences and self-interested behaviors. This paper formalizes these challenges into a Context-aware Bayesian Incentive-Compatible Recommendation Problem (CBICRP). \nTo address the CBICRP, we propose a two-stage algorithm (RCB) that integrates incentivized exploration with an efficient offline learning component for exploitation.\nIn the first stage, our algorithm explores available products while maintaining context-aware Bayesian incentive compatibility to determine sufficient sample sizes. The second stage employs inverse proportional gap sampling integrated with arbitrary efficient machine learning method to ensure sublinear regret. \nTheoretically, we prove that RCB achieves $O(\\sqrt{KdT})$ regret and satisfies Bayesian incentive compatibility (BIC). Empirically, we validate RCB's strong incentive gain, sublinear regret, and robustness through simulations and a real-world application on personalized warfarin dosing. Our work provides a principled approach for incentive-aware recommendation in online preference learning settings.", "keywords": [ "recommendation", "online learning", "incentive compatibility" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "SIZhZrU41O", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Anurag Bagchi", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Martial Hebert", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Professor" }, { "name": "Pavel Tokmakov", "gender": "Male", "institution": "Toyota Research Institute", "country": "US", "position": "Research Scientist" }, { "name": "Yu-Xiong Wang", "gender": "unknown", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zhipeng Bao", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 66, "n_ref": 160, "n_ref_all": 183, "n_fig": 5, "n_tab": 6, "L_tab": 2172, "n_element_tab": 182, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 895, "n_element_tab_1": 99, "formula_len_all": 453, "formula_len_all_1": 360, "len_all": 181303, "len_all_1": 56549, "len_abs": 1596, "len_title": 111, "len_sents": 39096, "len_sents_1": 26839, "n_sents": 310, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1604, "title": "Video Diffusion Models Learn the Structure of the Dynamic World", "abs": "Diffusion models have demonstrated significant progress in visual perception tasks due to their ability to capture fine-grained, object-centric features through large-scale vision-language pretraining. While their success in image-based tasks is well-established, extending this capability to the domain of video understanding remains a key challenge. In this work, we explore the potential of diffusion models for video understanding by analyzing the feature representations learned by both image- and video-based diffusion models, alongside non-generative, self-supervised approaches. We propose a unified probing framework to evaluate six models across four core video understanding tasks: action recognition, object discovery, scene understanding, and label propagation. Our findings reveal that video diffusion models consistently rank among the top performers, particularly excelling at modeling temporal dynamics and scene structure. This observation not only sets them apart from image-based diffusion models but also opens a new direction for advancing video understanding, offering a fresh alternative to traditional discriminative pre-training objectives. Interestingly, we demonstrate that higher generation performance does not always correlate with improved performance in downstream tasks, highlighting the importance of careful representation selection. Overall, our results suggest that video diffusion models hold substantial promise for video understanding by effectively capturing both spatial and temporal information, positioning them as strong competitors in this evolving domain.", "keywords": [ "Diffusion Models", "Video Understanding", "Representation Learning" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "SIE6VFps9x", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Ruoxi Sun", "gender": "Female", "institution": "Google", "country": "", "position": "Google" }, { "name": "Sercan O Arik", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Tomas Pfister", "gender": "Male", "institution": "Google", "country": "", "position": "Head of Research @ Cloud AI" }, { "name": "Max Chen", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 60, "n_ref": 123, "n_ref_all": 192, "n_fig": 4, "n_tab": 13, "L_tab": 7104, "n_element_tab": 903, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 3183, "n_element_tab_1": 261, "formula_len_all": 263, "formula_len_all_1": 242, "len_all": 286789, "len_all_1": 71017, "len_abs": 1537, "len_title": 137, "len_sents": 83013, "len_sents_1": 30511, "n_sents": 661, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1543, "title": "Learning to Clarify: Multi-turn Conversations with Action-Based Contrastive Self-Training", "abs": "Large language models (LLMs), optimized through human feedback, have rapidly emerged as a leading paradigm for developing intelligent conversational assistants. However, despite their strong performance across many benchmarks, LLM-based agents might still lack conversational skills such as disambiguation -- when they are faced with ambiguity, they often overhedge or implicitly guess users' true intents rather than asking clarification questions. Under task-specific settings, high-quality conversation samples are often limited, constituting a bottleneck for LLMs' ability to learn optimal dialogue action policies. We propose Action-Based Contrastive Self-Training (ACT), a quasi-online preference optimization algorithm based on Direct Preference Optimization (DPO), that enables data-efficient dialogue policy learning in multi-turn conversation modeling. We demonstrate ACT's efficacy under in data-efficient tuning scenarios, even when there is no action label available, using multiple real-world conversational tasks: tabular-grounded question-answering, machine reading comprehension, and AmbigSQL, a novel task for disambiguating information-seeking requests for complex SQL generation towards data analysis agents. Additionally, we propose evaluating LLMs' ability to function as conversational agents by examining whether they can implicitly recognize and reason about ambiguity in conversation. ACT demonstrates substantial conversation modeling improvements over standard tuning approaches like supervised fine-tuning and DPO.", "keywords": [ "dialogue system", "conversation modeling", "reinforcement learning", "mixed-initiative interaction", "LLM", "RLHF", "domain adaptation", "data-efficient learning", "clarification questions" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 5, 2 ] }, { "paper_id": "SI6zocV2SS", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Harikrishna Satheesh Pillai", "gender": "Male", "institution": "Amrita Vishwa Vidyapeetham (Deemed University)", "country": "IN", "position": "Undergrad student" }, { "name": "Pakhi Banchalia", "gender": "Female", "institution": "Amrita Vishwa Vidyapeetham (Deemed University)", "country": "IN", "position": "Undergrad student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 20, "n_ref": 28, "n_ref_all": 30, "n_fig": 1, "n_tab": 1, "L_tab": 77, "n_element_tab": 8, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 77, "n_element_tab_1": 8, "formula_len_all": 79, "formula_len_all_1": 79, "len_all": 69636, "len_all_1": 66275, "len_abs": 1119, "len_title": 83, "len_sents": 20246, "len_sents_1": 20155, "n_sents": 178, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 1108, "title": "CAN - CONTINUOUSLY ADAPTING NETWORKS", "abs": "Catastrophic forgetting is a fundamental challenge in neural networks that prevents continuous learning, which is one of the properties essential for achieving true general artificial intelligence. When trained sequentially on multiple tasks, conventional neural networks overwrite previously learned knowledge, hindering their ability to retain and apply past experiences. However, people and other animals can learn new things continuously without forgetting them. To overcome this problem, we devised an architecture that preserves significant task-specific connections by combining selective neuron freezing with Hebbian learning principles. Hebbian learning enables the network to adaptively strengthen synaptic connections depending on parameter activation. It is inspired by the synaptic plasticity seen in brains. By preserving the most important neurons using selective neuron freezing, new tasks can be trained without changing them. Experiments conducted on standard datasets show that our model significantly reduces the risk of catastrophic forgetting, allowing the network to learn continually.", "keywords": [ "Continual Learning", "Catastrophic Forgetting", "Synaptic Plasticity", "Hebbian Learning", "Adaptive Neural Networks" ], "rating_list": [ 3, 1, 1, 1 ], "soundness_list": [ 2, 1, 2, 1 ], "presentation_list": [ 2, 1, 1, 2 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "SI2hI0frk6", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Arun Babu", "gender": "Male", "institution": "State University of New York, Stony Brook", "country": "", "position": "MS student" }, { "name": "Chunting Zhou", "gender": "Female", "institution": "Meta AI", "country": "US", "position": "Researcher" }, { "name": "Jacob Kahn", "gender": "Male", "institution": "Facebook AI Research", "country": "", "position": "Research Engineer" }, { "name": "Kushal Tirumala", "gender": "unknown", "institution": "Research, Facebook", "country": "US", "position": "Researcher" }, { "name": "LILI YU", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Leonid Shamis", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Luke Zettlemoyer", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Full Professor" }, { "name": "Michihiro Yasunaga", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Omer Levy", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Xuezhe Ma", "gender": "Male", "institution": "USC/ISI", "country": "US", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 41, "n_ref": 93, "n_ref_all": 123, "n_fig": 6, "n_tab": 10, "L_tab": 811, "n_element_tab": 125, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 439, "n_element_tab_1": 69, "formula_len_all": 350, "formula_len_all_1": 235, "len_all": 201510, "len_all_1": 60740, "len_abs": 1029, "len_title": 129, "len_sents": 47606, "len_sents_1": 28997, "n_sents": 354, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 81, "L_abs": 1034, "title": "Transfusion: Predict the Next Token and Diffuse Images with One Multi-Modal Model", "abs": "We introduce Transfusion, a recipe for training a multi-modal model over discrete and continuous data.\nTransfusion combines the language modeling loss function (next token prediction) with diffusion to train a single transformer over mixed-modality sequences.\nWe pretrain multiple Transfusion models up to 7B parameters from scratch on a mixture of text and image data, establishing scaling laws with respect to a variety of uni- and cross-modal benchmarks.\nOur experiments show that Transfusion scales significantly better than quantizing images and training a language model over discrete image tokens.\nBy introducing modality-specific encoding and decoding layers, we can further improve the performance of Transfusion models, and even compress each image to just 16 patches.\nWe further demonstrate that scaling our Transfusion recipe to 7B parameters and 2T multi-modal tokens produces a model that can generate images and text on a par with similar scale diffusion models and language models, reaping the benefits of both worlds.", "keywords": [ "multimodal foundation model", "multimodal generation and understanding", "diffusion", "next token prediction" ], "rating_list": [ 8, 8, 8, 6, 8 ], "soundness_list": [ 2, 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 5, 3 ] }, { "paper_id": "SHeVc7efFz", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "JunDa Cheng", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Matthias M\u00fcller", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Michael Paulitsch", "gender": "Male", "institution": "Intel", "country": "DE", "position": "Principal Researcher" }, { "name": "Wei Yin", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Xin Yang", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zhaoxing Zhang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Zhipeng Cai", "gender": "Male", "institution": "Intel", "country": "US", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 31, "n_ref": 66, "n_ref_all": 82, "n_fig": 7, "n_tab": 5, "L_tab": 2237, "n_element_tab": 213, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1636, "n_element_tab_1": 127, "formula_len_all": 288, "formula_len_all_1": 288, "len_all": 106600, "len_all_1": 61197, "len_abs": 2024, "len_title": 84, "len_sents": 32918, "len_sents_1": 27079, "n_sents": 280, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 36, "L_abs": 1517, "title": "ROMEO: ROBUST METRIC VISUAL ODOMETRY", "abs": "Visual odometry (VO) aims to estimate camera poses from visual inputs --- the key for many applications such as VR/AR, robotics etc. This work focuses on monocular RGB VO where camera poses are directly estimated from a monocular RGB video without IMU or 3D sensors. Existing approaches lack robustness under this challenging scenario and fail to generalize to unseen data (especially outdoors); they also cannot recover metric-scale poses. Several methods have attempted to address these problems with priors from predicted depth. However, especially on unseen data, depth prediction noise can drastically degrade performance. We propose Robust Metric Visual Odometry (RoMeO), the first method that can leverage (noisy) depth priors to enable robust VO and recover metric scale poses. RoMeO incorporates both pre-trained monocular metric depth and multi-view stereo (MVS) models to recover metric-scale, simplify correspondence search, provide better initialization and regularize optimization. Effective strategies ensure the efficiency and the robustness to prior noise. RoMeO advances the state-of-the-art (SOTA) by a large margin across 6 diverse datasets covering both indoor and outdoor scenes. Compared to the current SOTA DPVO, RoMeO reduces the relative (align the trajectory scale with GT) and absolute trajectory errors on average by 55.2% and 77.8% respectively (Fig.1). The performance gain also transfers to the full SLAM pipeline (with global BA & loop closure). Code will be released upon acceptance.", "keywords": [ "Visual odometry\uff0cmonocular\uff0cSLAM" ], "rating_list": [ 3, 6, 6, 5, 1 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 5, 3, 4, 4, 5 ] }, { "paper_id": "SG1R2H3fa1", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Ayhan Suleymanzade", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "Undergrad student" }, { "name": "Jinwoo Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Olga Zaghen", "gender": "Female", "institution": "University of Amsterdam, University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Youngmin Ryou", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Undergrad student" }, { "name": "Seunghoon Hong", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 103, "n_formula_1": 20, "n_ref_uni": 70, "n_ref": 209, "n_ref_all": 273, "n_fig": 11, "n_tab": 9, "L_tab": 5128, "n_element_tab": 286, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1078, "n_element_tab_1": 121, "formula_len_all": 10012, "formula_len_all_1": 1146, "len_all": 322984, "len_all_1": 71205, "len_abs": 1290, "len_title": 94, "len_sents": 93038, "len_sents_1": 32356, "n_sents": 777, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 46, "L_abs": 1293, "title": "Revisiting Random Walks for Learning on Graphs", "abs": "We revisit a recent model class for machine learning on graphs, where a random walk on a graph produces a machine-readable record, and this record is processed by a deep neural network to directly make vertex-level or graph-level predictions. We refer to these stochastic machines as random walk neural networks (RWNNs), and through principled analysis, show that we can design them to be isomorphism invariant while capable of universal approximation of graph functions in probability. A useful finding is that almost any kind of record of random walk guarantees probabilistic invariance as long as the vertices are anonymized. This enables us, for example, to record random walks in plain text and adopt a language model to read these text records to solve graph tasks. We further establish a parallelism to message passing neural networks using tools from Markov chain theory, and show that over-smoothing in message passing is alleviated by construction in RWNNs, while over-squashing manifests as probabilistic under-reaching. We empirically demonstrate RWNNs on a range of problems, verifying our theoretical analysis and demonstrating the use of language models for separating strongly regular graphs where the 3-WL test fails, and transductive classification on arXiv citation network.", "keywords": [ "Graph machine learning", "random walk", "invariance", "universal approximation", "markov chain" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "SFuEabyr4v", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Ambuj Tewari", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "Full Professor" }, { "name": "UNIQUE SUBEDI", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "PhD student" } ], "n_formula": 179, "n_formula_1": 41, "n_ref_uni": 17, "n_ref": 41, "n_ref_all": 45, "n_fig": 2, "n_tab": 1, "L_tab": 15, "n_element_tab": 2, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 19488, "formula_len_all_1": 3007, "len_all": 166353, "len_all_1": 71576, "len_abs": 591, "len_title": 147, "len_sents": 48167, "len_sents_1": 28399, "n_sents": 575, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 591, "title": "Controlling Statistical, Discretization, and Truncation Errors in Learning Fourier Linear Operators", "abs": "We investigate the problem of learning operators between function spaces, focusing on the linear part of a layer in the Fourier Neural Operator architecture. First, we identify three main errors that occur during the learning process: statistical error due to finite sample size, truncation error from finite rank approximation of the operator, and discretization error from handling functional data on a finite grid of domain points. Finally, we analyze a Discrete Fourier Transform (DFT) based least squares estimator, establishing both upper and lower bounds on the aforementioned errors.", "keywords": [ "Operator Learning", "Fourier Linear Operators" ], "rating_list": [ 6, 6, 6, 1 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "SFkIb7pb8x", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Nan Tang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "Associate Professor" }, { "name": "Xudong Yang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology(Guangzhou)", "country": "CN", "position": "Intern" }, { "name": "Yifan Wu", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology(Guangzhou)", "country": "CN", "position": "MS student" }, { "name": "Yizhang Zhu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "MS student" }, { "name": "Yuyu Luo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 41, "n_ref": 106, "n_ref_all": 137, "n_fig": 7, "n_tab": 18, "L_tab": 4953, "n_element_tab": 626, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 2697, "n_element_tab_1": 334, "formula_len_all": 283, "formula_len_all_1": 283, "len_all": 198067, "len_all_1": 78851, "len_abs": 1807, "len_title": 115, "len_sents": 51210, "len_sents_1": 35621, "n_sents": 395, "n_sents_1": 235, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1888, "title": "AskChart: Universal Chart Understanding through Textual Enhancement", "abs": "Chart understanding tasks such as ChartQA and Chart-to-Text involve automatically extracting and interpreting key information from charts, enabling users to query or convert visual data into structured formats. State-of-the-art approaches primarily focus on visual cues from chart images, failing to *explicitly* incorporate rich textual information (e.g., data labels and axis labels) embedded within the charts. This textual information is vital for intuitive human comprehension and interpretation of charts. Moreover, existing models are often large and computationally intensive, limiting their practical applicability. In this paper, we introduce AskChart, a universal model that *explicitly* integrates both *textual* and *visual* cues from charts using a sparse Mixture of Experts (MoE) architecture. AskChart facilitates the learning of enhanced visual-textual representations of charts for effectively handling multiple chart understanding tasks, while maintaining a smaller model size. To capture the synergy between visual and textual modalities, we curate a large-scale dataset named ChartBase with about 7.5M data samples, which helps align textual and visual information and facilitates the extraction of visual entities and text. To effectively train AskChart, we design a three-stage training strategy to align visual and textual modalities embedded within charts for learning robust visual-textual representations and optimizing the learning of the MoE layer for advancing chart understanding. Extensive experiments across five datasets demonstrate the significant performance gains of AskChart in four chart understanding tasks. Remarkably, AskChart with 4.6B parameters outperforms state-of-the-art models with 13B parameters by 68.3\\% in Open-ended ChartQA and 49.2\\% in Chart-to-Text tasks, while achieving comparable performance in ChartQA and Chart-to-Table tasks.", "keywords": [ "Chart Understanding", "Multimodal Learning", "Vision-Language Models", "Visual-Textual Representation", "Instruction Tuning" ], "rating_list": [ 5, 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 2, 1, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 2, 2, 4 ] }, { "paper_id": "SFNqrHQTEP", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 5, "author_info_list": [ { "name": "Chita R. Das", "gender": "Male", "institution": "Pennsylvania State University", "country": "", "position": "Distinguished Professor" }, { "name": "Cyan Subhra Mishra", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Deeksha Chaudhary", "gender": "Female", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Jack Sampson", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Associate Professor" }, { "name": "Mahmut Kandemir", "gender": "Male", "institution": "Pennsylvania State University", "country": "", "position": "Full Professor" } ], "n_formula": 65, "n_formula_1": 8, "n_ref_uni": 29, "n_ref": 88, "n_ref_all": 96, "n_fig": 6, "n_tab": 9, "L_tab": 1650, "n_element_tab": 210, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 867, "n_element_tab_1": 127, "formula_len_all": 2231, "formula_len_all_1": 393, "len_all": 177807, "len_all_1": 72861, "len_abs": 1485, "len_title": 132, "len_sents": 67319, "len_sents_1": 36844, "n_sents": 486, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1493, "title": "Revisiting DNN Training for Intermittently-Powered Energy-Harvesting Micro-Computers", "abs": "The deployment of Deep Neural Networks (DNNs) in energy-constrained environments, such as Energy Harvesting Wireless Sensor Networks (EH-WSNs), introduces significant challenges due to the intermittent nature of power availability. This study introduces NExUME, a novel training methodology designed specifically for DNNs operating under such constraints. We propose a dynamic adjustment of training parameters\u2014dropout rates and quantization levels\u2014that adapt in real-time to the available energy, which varies in energy harvesting scenarios.\n\nThis approach utilizes a model that integrates the characteristics of the network architecture and the specific energy harvesting profile. It dynamically adjusts training strategies, such as the intensity and timing of dropout and quantization, based on predictions of energy availability. This method not only conserves energy but also enhances the network\u2019s adaptability, ensuring robust learning and inference capabilities even under stringent power constraints. Our results show a 6% to 22% improvement in accuracy over current methods, with an increase of less than 5% in computational overhead. This paper details the development of the adaptive training framework, describes the integration of energy profiles with dropout and quantization adjustments, and presents a comprehensive evaluation using real-world data. Additionally, we introduce a novel dataset aimed at furthering the application of energy harvesting in computational settings.", "keywords": [ "Intermittent Computing", "Energy Harvesting", "Intermittency Aware Training", "Hardware-Software codesign" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "SFN6Wm7YBI", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 13, "author_info_list": [ { "name": "Andrew Gu", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Chien-Chin Huang", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Gokul Nadathur", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Howard Huang", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Iris Zhang", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Intern" }, { "name": "Junjie Wang", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Less Wright", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Sanket Purandare", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Stratos Idreos", "gender": "unknown", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Tianyu Liu", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Wanchao Liang", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Wei Feng", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Software Engineer" }, { "name": "Will Constable", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 18, "n_ref": 25, "n_ref_all": 30, "n_fig": 2, "n_tab": 12, "L_tab": 5145, "n_element_tab": 146, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 3474, "n_element_tab_1": 84, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 155118, "len_all_1": 55153, "len_abs": 2666, "len_title": 129, "len_sents": 47944, "len_sents_1": 26662, "n_sents": 272, "n_sents_1": 140, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 2172, "title": "LegoScale: One-stop PyTorch native solution for production ready LLM pre-training", "abs": "The development of large language models (LLMs) has been instrumental in advancing state-of-the-art natural language processing applications. Training LLMs with billions of parameters and trillions of tokens requires sophisticated distributed systems that enable composing and comparing several state-of-the-art techniques in order to efficiently scale across thousands of accelerators. However, existing solutions are complex, scattered across multiple libraries/repositories, lack interoperability, and are cumbersome to maintain. Thus, curating and empirically comparing training recipes require non-trivial engineering effort.\n\nThis paper introduces LEGOSCALE, an open-source, PyTorch-native distributed training system that unifies and advances state-of-the-art techniques, streamlining integration and reducing engineering overhead. LEGOSCALE enables seamless application of 3D parallelism in a modular and composable manner, while featuring elastic scaling to adapt to changing computational requirements. The system provides comprehensive logging, efficient checkpointing, and debugging tools, ensuring production-ready training. Moreover, LEGOSCALE incorporates innovative hardware-software co-designed solutions, leveraging cutting-edge features like Float8 training and SymmetricMemory to maximize hardware utilization. As a flexible experimental test bed, LEGOSCALE facilitates the curation and comparison of custom recipes for diverse training contexts. By leveraging LEGOSCALE, we developed optimized training recipes for the Llama 3.1 family and provide actionable guidance on selecting and combining distributed training techniques to maximize training efficiency, based on our hands-on experiences.\n\nWe thoroughly assess LEGOSCALE on the Llama 3.1 family of LLMs, spanning 8 billion to 405 billion parameters, and showcase its exceptional performance, modular composability, and elastic scalability. By stacking the training optimizations, we demonstrate accelerations ranging from 65.08% on Llama3-8B at 128 GPU scale (1D), 13% on Llama3-70B at 256 GPU scale (2D), to 30% on Llama3-405B at 512 GPU scale (3D) on NVIDIA H100 GPUs over optimized baselines.", "keywords": [ "large language models", "distributed training", "pre-training", "data parallel", "tensor parallel", "pipeline parallel", "pytorch", "llama", "distributed checkpointing", "3D parallel" ], "rating_list": [ 10, 6, 6, 6, 6 ], "soundness_list": [ 4, 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 4, 2, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 5, 3 ] }, { "paper_id": "SEvJfuCtPY", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Francesco Insulla", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Santiago Aranguri", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 41, "n_formula_1": 41, "n_ref_uni": 17, "n_ref": 30, "n_ref_all": 36, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3178, "formula_len_all_1": 3178, "len_all": 74716, "len_all_1": 59655, "len_abs": 1038, "len_title": 81, "len_sents": 25439, "len_sents_1": 24928, "n_sents": 238, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1019, "title": "Phase-aware Training Schedule Simplifies Learning in Flow-Based Generative Models", "abs": "We analyze the training of a two-layer autoencoder used to parameterize a flow-based generative model for sampling from a high-dimensional Gaussian mixture. Building on the work of Cui et al. (2024), we find that the phase where the high-level features are learnt during training disappears as the dimension goes to infinity without an appropriate time schedule. We introduce a time dilation that solves this problem. This enables us to characterize the learnt velocity field, finding a first phase where the high-level feature (asymmetry between modes) is learnt and a second phase where the low-level feature (distribution of each mode) is learnt. We find that the autoencoder representing the velocity field learns to simplify by estimating only the parameters relevant to the feature for each phase. Turning to real data, we propose a method that, for a given feature, finds intervals of time where training improves accuracy the most on that feature, and we provide an experiment on MNIST validating this approach.", "keywords": [ "diffusion models", "phase transitions", "flow-based generative model", "high-dimensional gaussian mixtures", "denoising autoencoders", "training schedules" ], "rating_list": [ 5, 3, 3, 1 ], "soundness_list": [ 3, 1, 2, 1 ], "presentation_list": [ 2, 1, 2, 1 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "SEjdainnpB", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Lei Zhang", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Postdoc" }, { "name": "Mukesh Ghimire", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Yi Ren", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Associate Professor" }, { "name": "Zhe Xu", "gender": "unknown", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 61, "n_formula_1": 27, "n_ref_uni": 41, "n_ref": 105, "n_ref_all": 122, "n_fig": 9, "n_tab": 1, "L_tab": 392, "n_element_tab": 14, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6762, "formula_len_all_1": 3982, "len_all": 181212, "len_all_1": 74426, "len_abs": 1031, "len_title": 136, "len_sents": 48208, "len_sents_1": 28617, "n_sents": 417, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1126, "title": "Two-Player Zero-Sum Differential Games with One-Sided Information and Continuous Actions", "abs": "Unlike Poker where the action space $\\mathcal{A}$ is discrete, differential games in the physical world often have continuous action spaces not amenable to discrete abstraction, rendering no-regret algorithms with $\\mathcal{O}(|\\mathcal{A}|)$ complexity not scalable. To address this challenge within the scope of two-player zero-sum (2p0s) games with one-sided information, we show that (1) a computational complexity independent of $|\\mathcal{A}|$ can be achieved by exploiting the \"Cav u\" property of behavioral strategies in incomplete-information games and the Isaacs' condition that commonly holds for control systems, and that (2) the computation of the two equilibrium strategies can be decoupled under the Isaacs' condition. We provide computational complexity of the resultant algorithm for approximating continuous-action mixed strategies (CAMS). Empirically, we demonstrate correctness of CAMS using a homing game where the Nash equilibrium exists analytically, and scalability through the same game with higher-dimensional actions. Codes available in [anonymous repo](https://anonymous.4open.science/r/iclr-3245).", "keywords": [ "Incomplete Information Game", "Differential Game" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 1, 1, 3 ], "presentation_list": [ 1, 1, 2 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 4, 2, 3 ] }, { "paper_id": "SDjCRmuaDS", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Jes Frellsen", "gender": "Male", "institution": "Technical University of Denmark", "country": "DK", "position": "Associate Professor" }, { "name": "Raul Ortega Ochoa", "gender": "unknown", "institution": "Technical University of Denmark", "country": "DK", "position": "PhD student" }, { "name": "Tejs Vegge", "gender": "Male", "institution": "Technical University of Denmark", "country": "DK", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 36, "n_ref": 97, "n_ref_all": 120, "n_fig": 10, "n_tab": 6, "L_tab": 859, "n_element_tab": 44, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 141, "n_element_tab_1": 13, "formula_len_all": 519, "formula_len_all_1": 288, "len_all": 133985, "len_all_1": 60478, "len_abs": 1416, "len_title": 148, "len_sents": 43080, "len_sents_1": 30274, "n_sents": 286, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1421, "title": "MolMiner: Transformer architecture for fragment-based autoregressive generation of molecular stories", "abs": "Deep generative models for molecular discovery have become a very popular choice in new high-throughput screening paradigms. These models have been developed inheriting from the advances in natural language processing and computer vision, achieving ever greater results. However, generative molecular modelling has unique challenges that are often overlooked. Chemical validity, interpretability of the generation process and flexibility to variable molecular sizes are among some of the remaining challenges for generative models in computational materials design. In this work, we propose an autoregressive approach that decomposes molecular generation into a sequence of discrete and interpretable steps using molecular fragments as units, a 'molecular story'. Enforcing chemical rules in the stories guarantees the chemical validity of the generated molecules, the discrete sequential steps of a molecular story makes the process transparent improving interpretability, and the autoregressive nature of the approach allows the size of the molecule to be a decision of the model. We demonstrate the validity of the approach in a multi-target inverse design of electroactive organic compounds, focusing on the target properties of solubility, redox potential, and synthetic accessibility. Our results show that the model can effectively bias the generation distribution according to the prompted multi-target objective.", "keywords": [ "Deep generative models", "Material discovery", "Deep Learning", "Interpretability" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 1, 3, 1, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 3, 1, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "SDV7Y6Dhx9", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Ahmad Sajedi", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Alexander G Hauptmann", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Kai Wang", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Samir Khaki", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "MS student" }, { "name": "Shanmukha Ramakrishna Vedantam", "gender": "Male", "institution": "Self Employed", "country": "US", "position": "Researcher" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" }, { "name": "Zekai Li", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "Undergrad student" }, { "name": "Zhi-Qi Cheng", "gender": "Not Specified", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "Kostantinos Plataniotis", "gender": "Male", "institution": "Toronto University", "country": "CA", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 40, "n_ref": 101, "n_ref_all": 132, "n_fig": 14, "n_tab": 10, "L_tab": 5163, "n_element_tab": 353, "n_fig_1": 9, "n_tab_1": 6, "L_tab_1": 2398, "n_element_tab_1": 240, "formula_len_all": 310, "formula_len_all_1": 310, "len_all": 148544, "len_all_1": 65409, "len_abs": 1920, "len_title": 81, "len_sents": 43369, "len_sents_1": 28119, "n_sents": 352, "n_sents_1": 223, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1307, "title": "Emphasizing Discriminative Features for Dataset Distillation in Complex Scenarios", "abs": "Dataset distillation has demonstrated strong performance on simple datasets like CIFAR, MNIST, and TinyImageNet but struggles to achieve similar results in more complex scenarios. \nIn this paper, we propose a novel approach that \\textbf{e}mphasizes the \\textbf{d}iscriminative \\textbf{f}eatures (obtained by Grad-CAM) for dataset distillation, called \\textbf{EDF}.\nOur approach is inspired by a key observation: in simple datasets, high-activation areas typically occupy most of the image, whereas in complex scenarios, the size of these areas is much smaller.\nUnlike previous methods that treat all pixels equally when synthesizing images, EDF uses Grad-CAM activation maps to enhance high-activation areas.\nFrom a supervision perspective, we downplay supervision signals that have lower losses, as they contain common patterns.\nAdditionally, to help the DD community better explore complex scenarios, we build the Complex Dataset Distillation (Comp-DD) benchmark by meticulously selecting sixteen subsets, eight easy and eight hard, from ImageNet-1K.\nNotably, EDF consistently outperforms SOTA results in complex scenarios, such as ImageNet-1K subsets.\nHopefully, more researchers will be inspired and encouraged to enhance the practicality and efficacy of DD. \nOur code and benchmark will be made public.", "keywords": [ "dataset distillation" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "SDG0EBoqpp", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Hanqi Jiang", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Li Yang", "gender": "Male", "institution": "Northwestern Polytechnical University", "country": "", "position": "PhD student" }, { "name": "Tianming Liu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Professor" }, { "name": "Tuo Zhang", "gender": "Male", "institution": "Northwestern Polytechnical University", "country": "CN", "position": "Associate Professor" }, { "name": "Xintao Hu", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 42, "n_ref": 57, "n_ref_all": 70, "n_fig": 3, "n_tab": 5, "L_tab": 771, "n_element_tab": 67, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 3166, "n_element_tab_1": 80, "formula_len_all": 232, "formula_len_all_1": 232, "len_all": 124816, "len_all_1": 59293, "len_abs": 1368, "len_title": 122, "len_sents": 34036, "len_sents_1": 26980, "n_sents": 235, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1373, "title": "BrainSF: A Foundation Model for Whole Brain Functional Signals Forecasting", "abs": "Foundational models hold significant potential for advancing brain function research, particularly with recent technological advancements enabling the capture of spatiotemporal dynamics of brain signals. However, existing methods are primarily limited to characterizing observed brain signals and cannot infer continuous future signals\u2014an essential component for understanding the brain's causal structure and its role in various cognitive states. Current research leaves a substantial gap in forecasting whole-brain signal sequences. To address this, we propose a self-supervised model that embeds momentary whole-brain fMRI signals into vector representations and predicts continuous future signals. Our model is trained on a large-scale fMRI dataset, encompassing both resting-state and naturalistic stimuli conditions. Experimental results demonstrate that the model performs effectively in zero-shot forecasting of future whole-brain signals on unseen data and excels in downstream tasks such as task-based functional state decoding. To the best of our knowledge, this is the first approach to forecast and model whole-brain signals at such a large scale. The experimental results validate the feasibility of our method, offering new directions for theoretical research on brain signal time series and potential applications in diagnosing and treating brain disorders.", "keywords": [ "Foundation Model", "Brain Functional Signal", "Time Series Forecasting" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "SD4iBfAXZk", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Alexandre Capone", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Andrew Rothstein", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Egemen Kolemen", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "Associate Professor" }, { "name": "Hiro Josep Farre Kaga", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Jeff Schneider", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Researcher" }, { "name": "Rohit Sonker", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 30, "n_ref": 38, "n_ref_all": 49, "n_fig": 6, "n_tab": 1, "L_tab": 189, "n_element_tab": 39, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 213, "n_element_tab_1": 37, "formula_len_all": 765, "formula_len_all_1": 765, "len_all": 128830, "len_all_1": 68930, "len_abs": 1693, "len_title": 112, "len_sents": 40747, "len_sents_1": 31122, "n_sents": 329, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1704, "title": "DynaBO: Dynamic Model Bayesian Optimization for Tokamak Control", "abs": "Despite recent advances, state-of-the-art machine learning algorithms struggle considerably with control problems where data is scarce relative to model complexity. This problem is further exacerbated if the system changes over time, making past measurements less useful. While tools from reinforcement learning, supervised learning, and Bayesian optimization alleviate some of these issues, they do not address all of them at once. Considering these drawbacks, we present a multi-scale Bayesian optimization for fast and data-efficient decision-making. Our pipeline combines a high-frequency data-driven dynamics model with a low-frequency Gaussian process, resulting in a high-level model with a prior that is specifically tailored to the dynamics model setting. By updating the Gaussian process during Bayesian optimization, our method adapts rapidly to new data points, allowing us to process current high-quality data quickly, which is more representative of the system than past data. We apply our method to avoid tearing instabilities in a tokamak plasma, a control problem where modeling is difficult, and hardware changes potentially between experiments. Our approach is validated through offline testing on historical data and live experiments on the DIII-D tokamak. On the historical data, we show that our method outperforms a naive decision-making algorithm based exclusively on a recurrent neural network and past data. The live experiment corresponds to a high-performance plasma scenario with a high likelihood of instabilities. Despite this base configuration, we achieved a 50\\% success rate in the live experiment, representing an improvement of over 117\\% compared to historical data.", "keywords": [ "Nuclear Fusion", "Plasma Instabilities", "Bayesian Optimization", "Applied Machine Learning" ], "rating_list": [ 8, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "SCBn8MCLwc", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Barbara Plank", "gender": "unknown", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Chengzhi Hu", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "Undergrad student" }, { "name": "Xinpeng Wang", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Paul Rottger", "gender": "Male", "institution": "Bocconi University", "country": "", "position": "Postdoc" } ], "n_formula": 16, "n_formula_1": 9, "n_ref_uni": 33, "n_ref": 78, "n_ref_all": 104, "n_fig": 7, "n_tab": 7, "L_tab": 1537, "n_element_tab": 256, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 672, "n_element_tab_1": 126, "formula_len_all": 780, "formula_len_all_1": 442, "len_all": 154987, "len_all_1": 63121, "len_abs": 1017, "len_title": 149, "len_sents": 47805, "len_sents_1": 30268, "n_sents": 342, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 101, "L_abs": 1020, "title": "Surgical, Cheap, and Flexible: Mitigating False Refusal in Language Models via Single Vector Ablation", "abs": "Training a language model to be both helpful and harmless requires careful calibration of refusal behaviours: Models should refuse to follow malicious instructions or give harmful advice (e.g. \u201chow do I kill someone?\u201d), but they should not refuse safe requests, even if they superficially resemble unsafe ones (e.g. \u201chow do I kill a Python process?\u201d). Avoiding such false refusal, as prior work has shown, is challenging even for highly-capable language models. In this paper, we propose a simple and surgical method for mitigating false refusal in language models via single vector ablation. For a given model, we extract a false refusal vector and show that ablating this vector reduces false refusal rate without negatively impacting model safety and general model capabilities. We also show that our approach can be used for fine-grained calibration of model safety. Our approach is training-free and model-agnostic, making it useful for mitigating the problem of false refusal in current and future language models.", "keywords": [ "LLM Safety", "Exaggerated Safety", "Representation Editing" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 1, 4, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "SBzIbJojs8", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Peng Ye", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Shengji Tang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Tao Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Wanli Ouyang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Full Professor" }, { "name": "Weicai Ye", "gender": "Male", "institution": "KwaiVGI, Kuaishou Technology", "country": "CN", "position": "Researcher" }, { "name": "Weihao Lin", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Yang Zhou", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 13, "n_ref_uni": 43, "n_ref": 94, "n_ref_all": 113, "n_fig": 10, "n_tab": 9, "L_tab": 2357, "n_element_tab": 362, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 710, "n_element_tab_1": 82, "formula_len_all": 1553, "formula_len_all_1": 1412, "len_all": 146503, "len_all_1": 53932, "len_abs": 1520, "len_title": 0, "len_sents": 41745, "len_sents_1": 26852, "n_sents": 274, "n_sents_1": 140, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1529, "title": "HiSplat: Hierarchical 3D Gaussian Splatting for Generalizable Sparse-View Reconstruction", "abs": "Reconstructing 3D scenes from multiple viewpoints is a fundamental task in stereo vision. Recently, advances in generalizable 3D Gaussian Splatting have enabled high-quality novel view synthesis for unseen scenes from sparse input views by feed-forward predicting per-pixel Gaussian parameters without extra optimization. However, existing methods typically generate single-scale 3D Gaussians, which lack representation of both large-scale structure and texture details, resulting in mislocation and artefacts. In this paper, we propose a novel framework, HiSplat, which introduces a hierarchical manner in generalizable 3D Gaussian Splatting to construct hierarchical 3D Gaussians via a coarse-to-fine strategy. Specifically, HiSplat generates large coarse-grained Gaussians to capture large-scale structures, followed by fine-grained Gaussians to enhance delicate texture details. To promote inter-scale interactions, we propose an Error Aware Module for Gaussian compensation and a Modulating Fusion Module for Gaussian repair. Our method achieves joint optimization of hierarchical representations, allowing for novel view synthesis using only two-view reference images. Comprehensive experiments on various datasets demonstrate that HiSplat significantly enhances reconstruction quality and cross-dataset generalization compared to prior single-scale methods. The corresponding ablation study and analysis of different-scale 3D Gaussians reveal the mechanism behind the effectiveness. Codes will be released upon acceptance.", "keywords": [ "3D reconstruction", "Gaussian Splatting", "Generalizable Multi-View Reconstruction" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "SBbjwfMuik", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Boyang Li", "gender": "Unspecified", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Di Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Principal Researcher" }, { "name": "Rui Cao", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "Xuan Wu", "gender": "unknown", "institution": "Jilin University", "country": "CN", "position": "PhD student" }, { "name": "You Zhou", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Full Professor" }, { "name": "Yubin Xiao", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "PhD student" }, { "name": "Zhiguang Cao", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 15, "n_ref_uni": 46, "n_ref": 101, "n_ref_all": 137, "n_fig": 5, "n_tab": 14, "L_tab": 15263, "n_element_tab": 1774, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 7757, "n_element_tab_1": 654, "formula_len_all": 1139, "formula_len_all_1": 747, "len_all": 201260, "len_all_1": 82005, "len_abs": 1479, "len_title": 152, "len_sents": 52350, "len_sents_1": 32370, "n_sents": 363, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1487, "title": "From Global Assessment to Local Selection: Efficiently Solving Traveling Salesman Problems of All Sizes", "abs": "The Traveling Salesman Problem (TSP) is a well-known combinatorial optimization problem with broad real-world applications. Recent advancements in neural network-based TSP solvers have shown promising results. Nonetheless, these models often struggle to efficiently solve both small- and large-scale TSPs using the same set of pre-trained model parameters, limiting their practical utility. To address this issue, we introduce a novel neural TSP solver named GELD, built upon our proposed broad global assessment and refined local selection framework. Specifically, GELD integrates a lightweight Global-view Encoder (GE) with a heavyweight Local-view Decoder (LD) to enrich embedding representation while accelerating the decision-making process. Moreover, GE incorporates a novel low-complexity attention mechanism, allowing GELD to achieve low inference latency and scalability to larger-scale TSPs. Additionally, we propose a two-stage training strategy that utilizes training instances of different sizes to bolster GELD's generalization ability. Extensive experiments conducted on both synthetic and real-world datasets demonstrate that GELD outperforms seven state-of-the-art models considering both solution quality and inference speed. Furthermore, GELD can be employed as a post-processing method to exchange affordable computing time for significantly improved solution quality, capable of solving TSPs with up to 744,710 nodes without relying on divide-and-conquer strategies.", "keywords": [ "Traveling salesman problem", "neural combinatorial optimization", "size generalization" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "SBZiZFp560", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Guiguang Ding", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Hui Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Jungong Han", "gender": "Male", "institution": "The University of Sheffield", "country": "GB", "position": "Full Professor" }, { "name": "Mengyao Lyu", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Sicheng Zhao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Tianxiang Hao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiaohan Ding", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 8, "n_ref": 22, "n_ref_all": 34, "n_fig": 4, "n_tab": 12, "L_tab": 4249, "n_element_tab": 377, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3194, "n_element_tab_1": 114, "formula_len_all": 1202, "formula_len_all_1": 1168, "len_all": 154937, "len_all_1": 59616, "len_abs": 1461, "len_title": 138, "len_sents": 34656, "len_sents_1": 23874, "n_sents": 282, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1471, "title": "Towards Efficient Vision-Language Tuning: More Information Density, More Generalizability", "abs": "With the advancement of large pre-trained vision-language models, effectively transferring the knowledge embedded within these foundational models to downstream tasks has become a pivotal topic, particularly in data-scarce environments. Recently, parameter-efficient fine-tuning approaches, especially prompt tuning, have garnered considerable attention. To better understand the nature of prompt tuning, we propose the concept of ``Information Density'' (ID) to indicate whether a matrix strongly belongs to certain feature spaces rather than being evenly distributed across various feature spaces. We suppose a higher ID with strong bias across some feature spaces naturally leads to excellent robustness and stability. Our research, inspired by the observation that generalizability is closely linked to the information density of the prompt embedding, introduces the Dense Information Prompt (DIP). DIP aims to enhance information density to improve generalization. Several alternative algorithms to increase ID are proposed and verified effective. With further help of proper initialization and regularization, comprehensive experiments substantiate the superiority of DIP. Notably, DIP surpasses the latest state-of-the-art methods by a substantial margin with an exceptionally small parameter count and no extra inference overhead. Across a range of tasks spanning 11 datasets, DIP improves the average downstream accuracy of classic prompt tuning by up to 5.76\\%.", "keywords": [ "Vision-Language Models", "Parameter-Efficient Tuning", "Information Density", "Generalizability" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "SBCMNc3Mq3", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Fengyang Xu", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Postdoc" }, { "name": "Pin Chen", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Researcher" }, { "name": "Qing Mo", "gender": "unknown", "institution": "Northeastern University", "country": "US", "position": "MS student" }, { "name": "Yutong Lu", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Zexin Xu", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Researcher" }, { "name": "\u949f\u6cd3\u7468", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 36, "n_ref": 50, "n_ref_all": 77, "n_fig": 8, "n_tab": 4, "L_tab": 1248, "n_element_tab": 119, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 947, "n_element_tab_1": 84, "formula_len_all": 505, "formula_len_all_1": 505, "len_all": 130703, "len_all_1": 64252, "len_abs": 1263, "len_title": 177, "len_sents": 39936, "len_sents_1": 31202, "n_sents": 279, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 128, "L_abs": 1266, "title": "ECD: A Machine Learning Benchmark for Predicting Enhanced-Precision Electronic Charge Density in Crystalline Inorganic Materials", "abs": "Supervised machine learning techniques are increasingly being adopted to speed up electronic structure predictions, serving as alternatives to first-principles methods like Density Functional Theory (DFT). Although current DFT datasets mainly emphasize chemical properties and atomic forces, the precise prediction of electronic charge density is essential for accurately determining a system's total energy and ground state properties. In this study, we introduce a novel electronic charge density dataset named ECD, which encompasses 140,646 stable crystal geometries with medium-precision Perdew\u2013Burke\u2013Ernzerhof (PBE) functional data. Within this dataset, a subset of 7,147 geometries includes high-precision electronic charge density data calculated using the Heyd\u2013Scuseria\u2013Ernzerhof (HSE) functional in DFT. By designing various benchmark tasks for crystalline materials and emphasizing training with large-scale PBE data while fine-tuning with a smaller subset of high-precision HSE data, we demonstrate the efficacy of current machine learning models in predicting electronic charge densities.\nThe ECD dataset and baseline models are open-sourced to support community efforts in developing new methodologies and accelerating materials design and applications.", "keywords": [ "Electronic Charge Density", "Crystalline Inorganic Materials", "Graph Neural Network", "Dataset" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 4, 3, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "S9GyQUXzee", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Anji Liu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Bowei Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Haowei Lin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Shaofei Cai", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xiaojian Ma", "gender": "Unspecified", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Yitao Liang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zihao Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 5, "n_ref_uni": 53, "n_ref": 122, "n_ref_all": 148, "n_fig": 8, "n_tab": 6, "L_tab": 1162, "n_element_tab": 191, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 1067, "n_element_tab_1": 186, "formula_len_all": 730, "formula_len_all_1": 341, "len_all": 238738, "len_all_1": 66276, "len_abs": 1165, "len_title": 115, "len_sents": 55040, "len_sents_1": 30926, "n_sents": 403, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1142, "title": "GROOT-2: Weakly Supervised Multimodal Instruction Following Agents", "abs": "Developing agents that can follow multimodal instructions remains a fundamental challenge in robotics and AI. Although large-scale pre-training on unlabeled datasets has enabled agents to learn diverse behaviors, these agents often struggle with following instructions. While augmenting the dataset with instruction labels can mitigate this issue, acquiring such high-quality annotations at scale is impractical. \nTo address this issue, we frame the problem as a semi-supervised learning task and introduce \\agent, a multimodal instructable agent trained using a novel approach that combines weak supervision with latent variable models. Our method consists of two key components: constrained self-imitating, which utilizes large amounts of unlabeled demonstrations to enable the policy to learn diverse behaviors, and human intention alignment, which uses a smaller set of labeled demonstrations to ensure the latent space reflects human intentions. \\agent\u2019s effectiveness is validated across four diverse environments, ranging from video games to robotic manipulation, demonstrating its robust multimodal instruction-following capabilities.", "keywords": [ "Reinforcement Learning", "Open-world Agent", "Weakly Supervised Learning", "Goal-Conditioned Policy" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "S8nFZ98pmU", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Luca Daniel", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Professor" }, { "name": "Subhro Das", "gender": "Male", "institution": "MIT-IBM Watson AI Lab, IBM Research", "country": "", "position": "Staff Research Scientist" }, { "name": "Ziwen Ma", "gender": "Male", "institution": "Boson AI", "country": "US", "position": "Researcher" }, { "name": "Lam M Nguyen", "gender": "Male", "institution": "IBM Research, Thomas J. Watson Research Center", "country": "US", "position": "Staff Research Scientist" }, { "name": "WANG ZHANG", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 15, "n_ref_uni": 31, "n_ref": 64, "n_ref_all": 90, "n_fig": 4, "n_tab": 8, "L_tab": 2070, "n_element_tab": 112, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 281, "n_element_tab_1": 23, "formula_len_all": 1572, "formula_len_all_1": 1002, "len_all": 141573, "len_all_1": 61129, "len_abs": 1338, "len_title": 95, "len_sents": 49195, "len_sents_1": 28395, "n_sents": 354, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1340, "title": "Contrastive Meta Learning for Dynamical Systems", "abs": "Recent advancements in deep learning have significantly impacted the study of dynamical systems. Traditional approaches predominantly rely on supervised learning paradigms, limiting their scope to large scale problems and adaptability to new systems. This paper introduces a novel meta learning framework tailored for dynamical system forecasting, hinging on the concept of mapping the observed trajectories to a system-specific embedding space which encapsulates the inter-system characteristics and enriches the feature set for downstream prediction tasks. Central to our framework is the use of contrastive learning for trajectory data coupled with a series of neural network architecture designs to extract the features as augmented embedding for modeling system behavior. We present the application of zero-shot meta-learning to dynamical systems, demonstrating a substantial enhancement in performance metrics compared to existing baseline models. A notable byproduct of our methodology is the improved interpretability of the embeddings, which now carries explicit physical significance. Our results not only set a new benchmark in the field but also pave the way for enhanced interpretability and deeper understanding of complex dynamical systems, potentially opens new directions for how we approach system analysis and prediction.", "keywords": [ "dynamical system", "meta learning", "contrastive learning" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "S8gbnkCgxZ", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Bowen Gao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hongbo Ma", "gender": "Male", "institution": "Department of Computer Science and Technology, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Wei-Ying Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Ya-Qin Zhang", "gender": "Male", "institution": "AIR, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yanwen Huang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yanyan Lan", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Yinjun Harold Jia", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 0, "n_ref_uni": 42, "n_ref": 67, "n_ref_all": 91, "n_fig": 9, "n_tab": 8, "L_tab": 2855, "n_element_tab": 434, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 178, "formula_len_all_1": 0, "len_all": 168895, "len_all_1": 54734, "len_abs": 1700, "len_title": 93, "len_sents": 56183, "len_sents_1": 29312, "n_sents": 392, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 45, "L_abs": 1706, "title": "Redefining the task of Bioactivity Prediction", "abs": "Small molecules are vital to modern medicine, and accurately predicting their bioactivity against protein targets is crucial for therapeutic discovery and development. However, current machine learning models often rely on spurious features, leading to biased outcomes. Notably, a simple pocket-only baseline can achieve results comparable to, and sometimes better than, more complex models that incorporate both the protein pockets and the small molecules. Our analysis reveals that this phenomenon arises from insufficient training data and an improper evaluation process, which is typically conducted at the pocket level rather than the small molecule level. To address these issues, we redefine the bioactivity prediction task by introducing the SIU dataset-a million-scale Structural small molecule-protein Interaction dataset for Unbiased bioactivity prediction task, which is 50 times larger than the widely used PDBbind. The bioactivity labels in SIU are derived from wet experiments and organized by label types, ensuring greater accuracy and comparability. The complexes in SIU are constructed using a majority vote from three commonly used docking software programs, enhancing their reliability. Additionally, the structure of SIU allows for multiple small molecules to be associated with each protein pocket, enabling the redefinition of evaluation metrics like Pearson and Spearman correlations across different small molecules targeting the same protein pocket. Experimental results demonstrate that this new task provides a more challenging and meaningful benchmark for training and evaluating bioactivity prediction models, ultimately offering a more robust assessment of model performance.", "keywords": [ "Bioactivity Prediction", "New Dataset" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "S8VFVe6MWL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Alon Levkovitch", "gender": "unknown", "institution": "Google", "country": "IL", "position": "Researcher" }, { "name": "Eliya Nachmani", "gender": "Male", "institution": "Ben Gurion University of the Negev", "country": "IL", "position": "Assistant Professor" }, { "name": "Julian Salazar", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Nadav Bar", "gender": "Male", "institution": "Google", "country": "IL", "position": "Researcher" }, { "name": "RJ Skerry-Ryan", "gender": "Male", "institution": "Google", "country": "", "position": "Engineer" }, { "name": "Soroosh Mariooryad", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "W. Bastiaan Kleijn", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 39, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 98, "n_ref_all": 117, "n_fig": 3, "n_tab": 7, "L_tab": 1866, "n_element_tab": 180, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 704, "n_element_tab_1": 84, "formula_len_all": 6033, "formula_len_all_1": 437, "len_all": 156694, "len_all_1": 66709, "len_abs": 1178, "len_title": 99, "len_sents": 41987, "len_sents_1": 32877, "n_sents": 282, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 51, "L_abs": 1181, "title": "Towards Universal Mono-to-Binaural Speech Synthesis", "abs": "We consider the problem of synthesis of binaural speech from mono audio in arbitrary environments, which is important for modern telepresence and extended-reality applications. We find that existing neural mono-to-binaural methods are overfit to non-spatial acoustic properties, via analysis using a new benchmark (TUT Mono-to-Binaural), the first introduced since the original dataset of Richard et al. (2021). While these past methods focus on learning neural geometric transforms of monaural audio, we propose BinauralZero, a strong initial baseline for universal mono-to-binaural synthesis, which can subjectively match or outperform existing state-of-the-art neural mono-to-binaural renderers trained in their target environment despite never seeing any binaural data. It leverages the surprising discovery that an off-the-shelf mono audio denoising model can competently enhance the initial binauralization given by simple parameter-free transforms. We perform comprehensive ablations to understand how BinauralZero bridges the representation gap between mono and binaural audio, and analyze how current mono-to-binaural automated metrics are decorrelated from human ratings.", "keywords": [ "Binaural audio", "sound spatialization", "neural sound synthesis", "binaural speech" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 4, 1, 2, 1 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "S85PP4xjFD", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Linghao Jin", "gender": "Female", "institution": "University of Southern California", "country": "", "position": "PhD student" }, { "name": "Paul Pu Liang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Xiaofeng Liu", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" }, { "name": "Xu Han", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 36, "n_ref": 68, "n_ref_all": 94, "n_fig": 13, "n_tab": 6, "L_tab": 2008, "n_element_tab": 193, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2918, "n_element_tab_1": 254, "formula_len_all": 279, "formula_len_all_1": 221, "len_all": 147090, "len_all_1": 62016, "len_abs": 1674, "len_title": 167, "len_sents": 39741, "len_sents_1": 24882, "n_sents": 324, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 118, "L_abs": 1381, "title": "ContraFusion: Contrastively Improving Compositional Understanding in Diffusion Models via Fine-Grained Negative Images", "abs": "Despite the impressive text-to-image (T2I) synthesis capabilities of diffusion models, they often struggle to understand compositional relationships between objects and attributes, especially in complex settings. Existing solutions have tackled these challenges through optimizing the cross-attention mechanism or learning from the caption pairs with minimal semantic changes. However, can we generate high-quality complex contrastive images that diffusion models can directly discriminate based on visual representations? In this work, we leverage large-language models (LLMs) to compose realistic, complex scenarios and harness Visual-Question Answering (VQA) systems alongside diffusion models to automatically curate a contrastive dataset, COM-DIFF, consisting of 15k pairs of high-quality contrastive images. These pairs feature minimal visual discrepancies and cover a wide range of attribute categories, especially complex and natural scenarios. To learn effectively from these error cases, i.e., hard negative images, we propose CONTRAFUSION, a new multi-stage curriculum for contrastive learning of diffusion models. Through extensive experiments across a wide range of compositional scenarios, we showcase the effectiveness of our proposed framework on compositional T2I benchmarks. We will release our contrastive dataset to support the development of generative models.", "keywords": [ "compositional text-to-image generation", "contrastive learning", "compositional understanding" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "S7fuHAL89C", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Giovanni Chierchia", "gender": "unknown", "institution": "Ecole Sup\u00e9rieure d'Ing\u00e9nieurs en Electronique et Electrotechnique", "country": "", "position": "Assistant Professor" }, { "name": "Jean-Fran\u00e7ois BERCHER", "gender": "Male", "institution": "Ecole Sup\u00e9rieure d'Ing\u00e9nieurs en Electronique et Electrotechnique", "country": "", "position": "Associate Professor" }, { "name": "Nicolas Michel", "gender": "Male", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "JP", "position": "Postdoc" }, { "name": "Romain Negrel", "gender": "Male", "institution": "Ecole Sup\u00e9rieure d'Ing\u00e9nieurs en Electronique et Electrotechnique", "country": "", "position": "Associate Professor" } ], "n_formula": 37, "n_formula_1": 18, "n_ref_uni": 34, "n_ref": 49, "n_ref_all": 63, "n_fig": 4, "n_tab": 2, "L_tab": 1117, "n_element_tab": 46, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2230, "n_element_tab_1": 70, "formula_len_all": 3715, "formula_len_all_1": 1470, "len_all": 128435, "len_all_1": 62583, "len_abs": 1270, "len_title": 87, "len_sents": 31777, "len_sents_1": 24728, "n_sents": 299, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 96, "L_abs": 1301, "title": "Learning representations on Lp hyperspheres: The equivalence of loss functions in a MAP approach", "abs": "A common practice when training Deep Neural Networks is to force the learned representations to lie on the standard unit hypersphere, with respect to the $L_2$ norms. Such practice has been shown to improve both the stability and final performances of DNNs in many applications. In this paper, we derive a unified theoretical framework for learning representation on any $L_p$ hyperspheres for classification tasks, based on Maximum A Posteriori (MAP) modeling. Specifically, we give an expression of the probability distribution of multivariate Gaussians projected on any $L_p$ hypersphere and derive the general associated loss function. Additionally, we show that this framework demonstrates the theoretical equivalence of all projections on $L_p$ hyperspheres through the MAP modeling. It also provides a new interpretation of traditional Softmax Cross Entropy with temperature (SCE-$\\tau$) loss functions. Experiments on standard computer vision datasets give an empirical validation of the equivalence of projections on $L_p$ unit hyperspheres when using adequate objectives. It also shows that the SCE-$\\tau$ on projected representations, with optimally chosen temperature, shows comparable performances. The code is publicly available at \\url{https://anonymous.4open.science/r/map_code-71C7/", "keywords": [ "Representation Learning; Lp norms; Projected Gaussian Distributions; Image Classification" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "S7dFKyaOoE", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Liangliang Shi", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Rui Tao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 59, "n_formula_1": 25, "n_ref_uni": 29, "n_ref": 41, "n_ref_all": 54, "n_fig": 6, "n_tab": 2, "L_tab": 863, "n_element_tab": 127, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 863, "n_element_tab_1": 127, "formula_len_all": 5166, "formula_len_all_1": 2094, "len_all": 131257, "len_all_1": 70698, "len_abs": 1317, "len_title": 142, "len_sents": 42915, "len_sents_1": 31443, "n_sents": 380, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1324, "title": "MLOT: Extending the Bipartite Structure towards Multi-Layered Structure for Optimal Transport", "abs": "Despite its remarkable success and widespread adoption in various domains, optimal transport (OT) has a rather simple structure, relying on bipartite graphs with only two layers of nodes for transportation. In this paper, we propose a multi-layered OT approach that extends the original two-layer structure to handle transportation problems across multiple hierarchical levels. Within this framework, the source distribution flows through intermediate layers, before reaching the target distribution. Unlike previous variants of OT that involve multiple distributions, our multi-layered OT typically involves uncertain intermediate distributions, which need to be computed based on the relationships between the preceding and succeeding distributions. Under entropic regularization, MLOT-Sinkhorn algorithm is further proposed for multi-layered OT, which can be accelerated using GPUs and significantly outperforms general solvers such as Gurobi. The theoretical results of our entropic MLOT are also given in this paper. In the experiments, we validate its speed advantage and convergence performance. We further validate its feasibility through Text-Image retrieval and intermediate image computing task, which demonstrates reformulating the problems as MLOT can achieve better results. Source code will be made available.", "keywords": [ "Optimal Transport", "Multi-Layered Structure", "Entropic Regularization", "Sinkhorn Algorithm" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "S7cWJkWqOi", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Di ZHANG", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "VP" }, { "name": "Haoxian Zhang", "gender": "unknown", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Hejia Chen", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Intern" }, { "name": "Pengfei Wan", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Director" }, { "name": "Shoulong Zhang", "gender": "Male", "institution": "Zhongguancun Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Shuai Li", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Sisi Zhuang", "gender": "Male", "institution": "Beihang University", "country": "", "position": "MS student" }, { "name": "zhangyuan", "gender": "Male", "institution": "Foundation Model and Multimedia", "country": "CN", "position": "Principal Researcher" }, { "name": "Liuxiaoqiang", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 46, "n_ref": 83, "n_ref_all": 101, "n_fig": 12, "n_tab": 6, "L_tab": 913, "n_element_tab": 119, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 611, "n_element_tab_1": 96, "formula_len_all": 473, "formula_len_all_1": 473, "len_all": 137460, "len_all_1": 63342, "len_abs": 203, "len_title": 144, "len_sents": 40244, "len_sents_1": 29864, "n_sents": 277, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1564, "title": "Cafe-Talk: Generating 3D Talking Face Animation with Multimodal Coarse- and Fine-grained Control", "abs": "Speech-driven 3D talking face method should offer both accurate lip synchronization and controllable expressions. Previous methods solely adopt discrete emotion labels to globally control expressions throughout sequences while limiting flexible fine-grained facial control within the spatiotemporal domain. We propose a diffusion-transformer-based 3D talking face generation model, Cafe-Talk, which simultaneously incorporates coarse- and fine-grained multimodal control conditions. Nevertheless, the entanglement of multiple conditions challenges achieving satisfying performance. To disentangle speech audio and fine-grained conditions, we employ a two-stage training pipeline. Specifically, Cafe-Talk is initially trained using only speech audio and coarse-grained conditions. Then, a proposed fine-grained control adapter gradually adds fine-grained instructions represented by action units (AUs), preventing unfavorable speech-lip synchronization. To disentangle coarse- and fine-grained conditions, we design a swap-label training mechanism, which enables the dominance of the fine-grained conditions. We also devise a mask-based CFG technique to regulate the occurrence and intensity of fine-grained control. In addition, a text-based detector is introduced with text-AU alignment to enable natural language user input and further support multimodal control. Extensive experimental results prove that Cafe-Talk achieves state-of-the-art lip synchronization and expressiveness performance and receives wide acceptance in fine-grained control in user studies.", "keywords": [ "3D talking face", "generative model", "fine-grained control", "action units" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 2, 5, 4 ] }, { "paper_id": "S7F7IMGX4O", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 13, "author_info_list": [ { "name": "Chi Wang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Haolong Jia", "gender": "unknown", "institution": "University of California, Santa Cruz", "country": "US", "position": "MS student" }, { "name": "Lichao Sun", "gender": "Male", "institution": "Lehigh University", "country": "US", "position": "Assistant Professor" }, { "name": "Lifang He", "gender": "Female", "institution": "Lehigh University", "country": "US", "position": "Assistant Professor" }, { "name": "Ruoxi Chen", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Weixiang Sun", "gender": "Male", "institution": "Northeastern University", "country": "CN", "position": "Undergrad student" }, { "name": "Yanfang Ye", "gender": "unknown", "institution": "University of Notre Dame", "country": "US", "position": "Full Professor" }, { "name": "Yihan Cao", "gender": "Female", "institution": "LinkedIn", "country": "", "position": "Researcher" }, { "name": "Yixin Liu", "gender": "Male", "institution": "Dolby", "country": "US", "position": "Intern" }, { "name": "Zhaoxu Li", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Zhengqing Yuan", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Lin Bin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yuan LI", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 58, "n_ref": 129, "n_ref_all": 168, "n_fig": 11, "n_tab": 14, "L_tab": 6514, "n_element_tab": 732, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1125, "n_element_tab_1": 184, "formula_len_all": 559, "formula_len_all_1": 281, "len_all": 289289, "len_all_1": 68090, "len_abs": 1660, "len_title": 118, "len_sents": 87305, "len_sents_1": 33396, "n_sents": 587, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1668, "title": "Mora: Enabling Generalist Video Generation via A Multi-Agent Framework", "abs": "Text-to-video generation has made significant strides, but replicating the capabilities of advanced systems like OpenAI\u2019s Sora remains challenging due to their closed-source nature. Existing open-source methods struggle to achieve comparable performance, often hindered by ineffective agent collaboration and inadequate training data quality. In this paper, we introduce Mora, a novel multi-agent framework that leverages existing open-source modules to replicate Sora\u2019s functionalities. We address these fundamental limitations by proposing three key techniques: (1) multi-agent fine-tuning with a self-modulation factor to enhance inter-agent coordination, (2) a data-free training strategy that uses large models to synthesize training data, and (3) a human-in-the-loop mechanism combined with multimodal large language models for data filtering to ensure high-quality training datasets. Our comprehensive experiments on six video generation tasks demonstrate that Mora achieves performance comparable to Sora on VBench \\cite{huang2024vbench}, outperforming existing open-source methods across various tasks. Specifically, in the text-to-video generation task, Mora achieved a Video Quality score of 0.800, surpassing Sora\u2019s 0.797 and outperforming all other baseline models across six key metrics. Additionally, in the image-to-video generation task, Mora achieved a perfect Dynamic Degree score of 1.00, demonstrating exceptional capability in enhancing motion realism and achieving higher Imaging Quality than Sora. These results highlight the potential of collaborative multi-agent systems and human-in-the-loop mechanisms in advancing text-to-video generation.", "keywords": [ "Video Generation", "Multi-agent", "Adaption Training" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "S6cBH99BhB", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Jiangkuo Wang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Mingpeng Wei", "gender": "Male", "institution": "National Park College", "country": "FR", "position": "Researcher" }, { "name": "Suyv Ma", "gender": "unknown", "institution": "", "country": "", "position": "" } ], "n_formula": 8, "n_formula_1": 2, "n_ref_uni": 28, "n_ref": 82, "n_ref_all": 100, "n_fig": 9, "n_tab": 7, "L_tab": 970, "n_element_tab": 123, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 498, "formula_len_all_1": 77, "len_all": 114951, "len_all_1": 46248, "len_abs": 980, "len_title": 162, "len_sents": 29272, "len_sents_1": 20784, "n_sents": 218, "n_sents_1": 142, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 114, "L_abs": 983, "title": "Enhancing Multilingual Reasoning in LLMs: Insights from Cross-Linguistic Correlations and Optimal Data Proportions", "abs": "Large language models (LLMs) typically rely on fine-tuning to enhance their reasoning capabilities across various languages. However, limited research has been conducted on the optimal balance of language proportions within multilingual reasoning datasets. To fill this gap, we performed a systematic study to examine how different proportions of language data in multilingual reasoning datasets influence fine-tuning performance. Our study revealed a clear relationship between language proportions in datasets and the fine-tuning performance of LLMs. By fine-tuning multiple LLMs using the appropriate language distributions and data volumes identified in our study, we achieved state-of-the-art performance in both multilingual mathematical reasoning and solving mathematical problems using Python code. Furthermore, our approach significantly reduced data volume requirements and translation costs compared to existing methods, providing a valuable reference for future research.", "keywords": [ "Large Language Models", "Multilingual Reasoning", "Fine-Tuning" ], "rating_list": [ 8, 5, 8, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "S5wIXxlvfw", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Connor W. Magoon", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Fengyu Yang", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Noam Aigerman", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Assistant Professor" }, { "name": "Shahar Kovalsky", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" } ], "n_formula": 32, "n_formula_1": 12, "n_ref_uni": 91, "n_ref": 154, "n_ref_all": 221, "n_fig": 11, "n_tab": 6, "L_tab": 4329, "n_element_tab": 601, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 968, "n_element_tab_1": 18, "formula_len_all": 2703, "formula_len_all_1": 1167, "len_all": 229350, "len_all_1": 74565, "len_abs": 1774, "len_title": 63, "len_sents": 55902, "len_sents_1": 34922, "n_sents": 397, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1792, "title": "Differentiation Through Black-Box Quadratic Programming Solvers", "abs": "In recent years, many deep learning approaches have incorporated layers that solve optimization problems (*e.g.*, linear, quadratic, and semidefinite programs). Integrating these optimization problems as differentiable layers requires computing the derivatives of the optimization problem's solution with respect to its objective and constraints. This has so far limited the use of state-of-the-art black-box numerical solvers within neural networks, as they lack a differentiable interface. To address this issue for one of the most common convex optimization problems -- quadratic programming (QP) -- we introduce **dQP**, a modular framework that enables plug-and-play differentiation for any QP solver, allowing seamless integration into neural networks and bi-level optimization tasks. Our solution is based on the core theoretical insight that knowledge of the active constraint set at the QP optimum allows for *explicit* differentiation. This insight reveals a unique relationship between the computation of the solution and its derivative, enabling efficient differentiation of any solver, that only requires the primal solution. Our implementation, which will be made publicly available upon acceptance, interfaces with an existing framework that supports over 15 state-of-the-art QP solvers, providing each with a fully differentiable backbone for immediate use as a differentiable layer in learning setups. To demonstrate the scalability and effectiveness of dQP, we evaluate it on a large benchmark dataset of QPs with varying structures. We compare dQP with existing differentiable QP methods, demonstrating its advantages across a range of problems, from challenging small and dense problems to large-scale sparse ones, including a novel bi-level geometry optimization problem.", "keywords": [ "optimization", "differentiable optimization", "quadratic programming" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "S5Yo6w3n3f", "primary_area": "reinforcement learning", "n_author": 9, "author_info_list": [ { "name": "Jingliang Duan", "gender": "Male", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "Associate Professor" }, { "name": "Liangfa Chen", "gender": "Male", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "MS student" }, { "name": "Likun Wang", "gender": "Male", "institution": "Beihang University", "country": "", "position": "Undergrad student" }, { "name": "Shengbo Eben Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Tong Liu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Xujie Song", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yinuo Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yuming Yin", "gender": "unknown", "institution": "Zhejiang University of Technology", "country": "", "position": "Associate Professor" }, { "name": "wenxuan Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 20, "n_ref_uni": 20, "n_ref": 25, "n_ref_all": 42, "n_fig": 4, "n_tab": 9, "L_tab": 3336, "n_element_tab": 174, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1335, "n_element_tab_1": 43, "formula_len_all": 2040, "formula_len_all_1": 1129, "len_all": 120391, "len_all_1": 57042, "len_abs": 3667, "len_title": 68, "len_sents": 34635, "len_sents_1": 24597, "n_sents": 283, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 67, "L_abs": 1325, "title": "ODE-based Smoothing Neural Network for Reinforcement Learning Tasks", "abs": "The smoothness of control actions is a significant challenge faced by deep reinforcement learning (RL) techniques in solving optimal control problems. Existing RL-trained policies tend to produce non-smooth actions due to high-frequency input noise and unconstrained Lipschitz constants in neural networks. This article presents a Smooth ODE (SmODE) network capable of simultaneously addressing both causes of unsmooth control actions, thereby enhancing policy performance and robustness under noise condition. We first design a smooth ODE neuron with first-order low-pass filtering expression, which can dynamically filter out high frequency noises of hidden state by a learnable state-based system time constant. Additionally, we construct a state-based mapping function, $g$, and theoretically demonstrate its capacity to control the ODE neuron's Lipschitz constant. Then, based on the above neuronal structure design, we further advanced the SmODE network serving as RL policy approximators. This network is compatible with most existing RL algorithms, offering improved adaptability compared to prior approaches. Various experiments show that our SmODE network demonstrates superior anti-interference capabilities and smoother action outputs than the multi-layer perception and smooth network architectures like LipsNet.", "keywords": [ "Reinforcement Learning", "Smooth Control", "Low-pass Filter", "Neural ODE" ], "rating_list": [ 8, 5, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "S5JCqTJyKj", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Jeongjin Shin", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Intern" }, { "name": "Sangdon Park", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 30, "n_ref": 70, "n_ref_all": 88, "n_fig": 11, "n_tab": 8, "L_tab": 2112, "n_element_tab": 150, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2731, "n_element_tab_1": 157, "formula_len_all": 841, "formula_len_all_1": 841, "len_all": 122776, "len_all_1": 60340, "len_abs": 1589, "len_title": 111, "len_sents": 44576, "len_sents_1": 28118, "n_sents": 311, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1620, "title": "Deferred Backdoor Functionality Attacks on Deep Learning Models", "abs": "Deep learning models are vulnerable to backdoor attacks, where adversaries inject malicious functionality during training that activates on trigger inputs at inference time. Extensive research has focused on developing stealthy backdoor attacks to evade detection and defense mechanisms. \nHowever, these approaches still have limitations that leave the door open for detection and mitigation due to their inherent design to cause malicious behavior in the presence of a trigger.\nTo address this limitation, we introduce Deferred Activated Backdoor Functionality (DABF), a new paradigm in backdoor attacks. Unlike conventional attacks, DABF initially conceals its backdoor, producing benign outputs even when triggered. This stealthy behavior allows DABF to bypass multiple detection and defense methods, remaining undetected during initial inspections.\nThe backdoor functionality is strategically activated only after the model undergoes subsequent updates, such as retraining on benign data. DABF attacks exploit the common practice in the life cycle of machine learning models to perform model updates and fine-tuning after initial deployment. To implement DABF attacks, we approach the problem by making the unlearning of the backdoor fragile, allowing it to be easily cancelled and subsequently reactivate the backdoor functionality. To achieve this, we propose a novel two-stage training scheme, called $\\texttt{DeferBad}$. Our extensive experiments across various fine-tuning scenarios, backdoor attack types, datasets, and model architectures demonstrate the effectiveness and stealthiness of $\\texttt{DeferBad}$.", "keywords": [ "AI security", "Backdoor Attack", "Stealthy Attack" ], "rating_list": [ 5, 3, 1 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "S4jzvOBs9m", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Andres Bruhn", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "DE", "position": "Full Professor" }, { "name": "Jenny Schmalfuss", "gender": "Female", "institution": "Universit\u00e4t Stuttgart", "country": "DE", "position": "PhD student" }, { "name": "Julian Yuya Caspary", "gender": "Male", "institution": "Universit\u00e4t Mannheim", "country": "DE", "position": "MS student" }, { "name": "Luca Schwarz", "gender": "Male", "institution": "Universit\u00e4t Mannheim", "country": "DE", "position": "MS student" }, { "name": "Margret Keuper", "gender": "Female", "institution": "Universit\u00e4t Mannheim", "country": "DE", "position": "Full Professor" }, { "name": "Shashank Agnihotri", "gender": "Male", "institution": "Universit\u00e4t Mannheim", "country": "DE", "position": "PhD student" }, { "name": "Xinyan Gao", "gender": "unknown", "institution": "Universit\u00e4t Mannheim", "country": "DE", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 0, "n_ref_uni": 67, "n_ref": 276, "n_ref_all": 306, "n_fig": 10, "n_tab": 8, "L_tab": 6114, "n_element_tab": 449, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1194, "n_element_tab_1": 62, "formula_len_all": 647, "formula_len_all_1": 0, "len_all": 292836, "len_all_1": 58743, "len_abs": 1710, "len_title": 109, "len_sents": 92949, "len_sents_1": 29190, "n_sents": 637, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1740, "title": "FlowBench: A Robustness Benchmark for Optical Flow Estimation", "abs": "Optical flow estimation is a crucial computer vision task often applied to safety-critical real-world scenarios like autonomous driving and medical imaging.\nWhile optical flow estimation accuracy has greatly benefited from the emergence of deep learning, learning-based methods are also known for their lack of generalization and reliability.\nHowever, reliability is paramount when optical flow methods are employed in the real world, where safety is essential.\nFurthermore, a deeper understanding of the robustness and reliability of learning-based optical flow estimation methods is still lacking, hindering the research community from building methods safe for real-world deployment.\nThus we propose **FlowBench**, a robustness benchmark and evaluation tool for learning-based optical flow methods. \n**FlowBench** facilitates streamlined research into the reliability of optical flow methods by benchmarking their robustness to adversarial attacks and out-of-distribution samples.\nWith **FlowBench**, we benchmark 91 methods across 3 different datasets under 7 diverse adversarial attacks and 23 established common corruptions, making it the most comprehensive robustness analysis of optical flow methods to date.\nAcross this wide range of methods, we consistently find that methods with state-of-the-art performance on established standard benchmarks lack reliability and generalization ability.\nMoreover, we find interesting correlations between the performance, reliability, and generalization ability of optical flow estimation methods, under various lenses such as design choices used, number of parameters, etc.\nAfter acceptance, **FlowBench** will be open-source and publicly available, including the weights of all tested models.", "keywords": [ "Adversarial attacks", "optical flow estimation", "benchmarking tool", "benchmark", "robustness" ], "rating_list": [ 5, 3, 8, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 2, 4, 4, 5 ] }, { "paper_id": "S4dItvpvAv", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Ness Shroff", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Full Professor" }, { "name": "Peizhong Ju", "gender": "Male", "institution": "University of Kentucky", "country": "US", "position": "Assistant Professor" }, { "name": "Yining Li", "gender": "Female", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" } ], "n_formula": 97, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 62, "n_ref_all": 107, "n_fig": 13, "n_tab": 1, "L_tab": 49, "n_element_tab": 6, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9269, "formula_len_all_1": 484, "len_all": 262919, "len_all_1": 79440, "len_abs": 2173, "len_title": 108, "len_sents": 91194, "len_sents_1": 38270, "n_sents": 802, "n_sents_1": 288, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 60, "L_abs": 2189, "title": "How to Find the Exact Pareto Front for Multi-Objective MDPs?", "abs": "Multi-Objective Markov Decision Processes (MO-MDPs) are receiving increasing attention, as real-world decision-making problems often involve conflicting objectives that cannot be addressed by a single-objective MDP. \nThe Pareto front identifies the set of policies that cannot be dominated, providing a foundation for finding Pareto optimal solutions that can efficiently adapt to various preferences.\nHowever, finding the Pareto front is a highly challenging problem. Most existing methods either (i) rely on traversing the *continuous preference space*, which is impractical and results in approximations that are difficult to evaluate against the true Pareto front, or (ii) focus solely on deterministic Pareto optimal policies, from which there are no known techniques to characterize the full Pareto front. Moreover, finding the structure of the Pareto front itself remains unclear even in the context of dynamic programming, where the MDP is fully known in advance.\nIn this work, we address the challenge of efficiently discovering the Pareto front, involving both deterministic and stochastic Pareto optimal policies.\nBy investigating the geometric structure of the Pareto front in MO-MDPs, we uncover a key property: the Pareto front is on the boundary of a convex polytope whose vertices all correspond to deterministic policies, and neighboring vertices of the Pareto front differ by only one state-action pair of the deterministic policy, almost surely.\nThis insight transforms the global comparison across all policies into a localized search among deterministic policies that differ by only one state-action pair, drastically reducing the complexity of searching for the exact Pareto front. \nWe develop an efficient algorithm that identifies the vertices of the Pareto front by solving a single-objective MDP only once and then traversing the edges of the Pareto front, making it more efficient than existing methods. Furthermore, the entire Pareto front can be found in $V$ iterations, where $V$ represents the number of vertices on the Pareto front.\nOur empirical studies demonstrate the effectiveness of our theoretical strategy in discovering the Pareto front efficiently.", "keywords": [ "Multi-objective optimization", "Markov decision Process" ], "rating_list": [ 8, 6, 8, 8, 8 ], "soundness_list": [ 4, 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 3, 4 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 3, 3 ] }, { "paper_id": "S3zKrEQpRr", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Jin Zheng", "gender": "unknown", "institution": "University of Bristol", "country": "GB", "position": "Lecturer" }, { "name": "John Cartlidge", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "Full Professor" }, { "name": "Zinuo You", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "PhD student" } ], "n_formula": 72, "n_formula_1": 29, "n_ref_uni": 53, "n_ref": 214, "n_ref_all": 263, "n_fig": 7, "n_tab": 2, "L_tab": 686, "n_element_tab": 90, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 858, "n_element_tab_1": 60, "formula_len_all": 7539, "formula_len_all_1": 1480, "len_all": 148560, "len_all_1": 57582, "len_abs": 1098, "len_title": 137, "len_sents": 39692, "len_sents_1": 25510, "n_sents": 277, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1103, "title": "Unleashing the Information Flow: Graph Neural Networks are Noisy Communication Channels", "abs": "Existing message-passing graph neural networks often rely on carefully designed information propagation methods to perform reasonably in graph-related mining tasks. However, this invokes the problem of whether the dimensions of learnable matrices and the depths of the networks are properly estimated. While this challenge has been attempted by others, it remains an open problem. Using the principle of maximum entropy and Shannon's theorem, we demonstrate that message-passing graph neural networks function similarly to noisy communication channels. The optimal information transmission state of graph neural networks can be reached when Shannon's theorem is satisfied, which is determined by their entropy and channel capacity. In addition, we illustrate that the widths of trainable matrices should be sufficiently large to avoid the shrinkage of model channel capacity and the increase of the channel capacity diminishes as the depth of the networks increases. The proposed approach is empirically verified through extensive experiments on five public semi-supervised node classification datasets.", "keywords": [ "graph neural networks", "entropy", "channel capacity", "model dimensionality estimation" ], "rating_list": [ 3, 1, 5, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 4, 4, 1, 4 ] }, { "paper_id": "S2WUJUETyc", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Chen Gu", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "ZHUOYU CHEN", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 17, "n_ref": 26, "n_ref_all": 41, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1010, "formula_len_all_1": 1009, "len_all": 79904, "len_all_1": 46817, "len_abs": 1279, "len_title": 154, "len_sents": 28498, "len_sents_1": 22024, "n_sents": 206, "n_sents_1": 155, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1283, "title": "Integrating Distributed Acoustic Sensing and PINN Frameworks for Enhanced Indoor Sound Source Localization", "abs": "Distributed Acoustic Sensing (DAS) is an emerging technology that transforms standard optical fibers into dense arrays of acoustic sensors, offering unprecedented opportunities for smart city applications, indoor monitoring of human activity, and surveillance without compromising privacy. In this paper, we integrate DAS with Physics-Informed Neural Networks (PINNs) for indoor sound source localization. By embedding the acoustic wave equation and impedance boundary conditions into the neural network architecture, we exploit physical laws to guide the learning process, improving accuracy and generalization. We propose two strategies for real-time sound source localization using DAS data. The first strategy involves training the PINN on all available data simultaneously, while the second strategy incrementally feeds data over time, simulating real-time data acquisition. Using real indoor DAS measurements, we demonstrate the effectiveness of our approach in deciphering complex room acoustics and accurately inferring sound source locations under both strategies. Our framework provides a novel solution for real-time indoor positioning and human activity surveillance, offering significant advantages over traditional camera-based systems by preserving individual privacy.", "keywords": [ "distributed acoustic sensing", "physics informed neural networks", "room acoustics", "sound source localization" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "S2WHlhvFGg", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Han Quan", "gender": "unknown", "institution": "China Agricultural University", "country": "CN", "position": "MS student" }, { "name": "Lang Zheng", "gender": "Male", "institution": "China Agricultural University", "country": "CN", "position": "MS student" }, { "name": "Ruxue Xing", "gender": "Female", "institution": "China Agricultural University", "country": "CN", "position": "MS student" }, { "name": "Zhenghan Chen", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "youhuan yang", "gender": "Male", "institution": "Zhengzhou University", "country": "CN", "position": "Researcher" } ], "n_formula": 35, "n_formula_1": 32, "n_ref_uni": 17, "n_ref": 25, "n_ref_all": 30, "n_fig": 1, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2090, "formula_len_all_1": 2087, "len_all": 88861, "len_all_1": 55880, "len_abs": 1672, "len_title": 148, "len_sents": 29258, "len_sents_1": 26308, "n_sents": 199, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1587, "title": "Advancing Drug-Target Interaction Prediction via Graph Transformers and Residual Protein Embeddings", "abs": "Predicting drug-target interactions (DTIs) is essential for advancing drug discovery. This paper presents a unified mathematical framework for unsupervised domain adaptation in drug-target interaction (DTI) prediction, integrating measure theory, functional analysis, information geometry, and optimal transport theory. We introduce the novel concept of DTI-Wasserstein distance, incorporating both structural and chemical similarities of drugs and targets, and establish a refined bound on the difference between source and target risks. Our information-geometric perspective reveals the intrinsic structure of the DTI model space, characterizing optimal adaptation paths as geodesics on a statistical manifold equipped with the Fisher-Rao metric. We develop a spectral decomposition of the DTI-DA transfer operator, providing insights into the modes of information transfer between domains. This leads to the introduction of DTI-spectral embedding and DTI-spectral mutual information, allowing for a more nuanced understanding of the adaptation process. Theoretical contributions include refined bounds on DTI-DA performance, incorporating task-specific considerations and spectral properties of the feature space. We prove the existence of an optimal transport map for DTI-DA and derive a novel information-theoretic lower bound using DTI-mutual information. Empirical evaluations demonstrate the superiority of our approach over existing methods across multiple benchmark datasets, showcasing its ability to effectively leverage data from diverse sources for improved DTI prediction.", "keywords": [ "DTI\uff0cTransfer Learning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "S1U0CIuejF", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Biwei Huang", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Linwei Tian", "gender": "unknown", "institution": "University of Hong Kong", "country": "", "position": "Associate Professor" }, { "name": "Pei Zhang", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 9, "n_ref_uni": 26, "n_ref": 38, "n_ref_all": 81, "n_fig": 17, "n_tab": 1, "L_tab": 333, "n_element_tab": 77, "n_fig_1": 15, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 956, "formula_len_all_1": 326, "len_all": 147496, "len_all_1": 67456, "len_abs": 866, "len_title": 145, "len_sents": 57250, "len_sents_1": 32503, "n_sents": 450, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 869, "title": "TrendDiff: Decoupling Intrinsic and Measurement Trends for Enhanced Time Series Causal Discovery", "abs": "Time trends can be classified into intrinsic (real) and measurement (false) trends. There has long been a critical need for techniques to discern them, especially in investment decision-making. In causal discovery, these measurement trends, essentially measurement errors, can significantly impact the performance of algorithms, making it crucial to identify and eliminate them before analysis as well. Recognizing this need, we present a novel algorithm, termed Trend Differentiator (TrendDiff). It is capable of detecting all trend-influenced variables and differentiating between those affected by measurement trends and those displaying intrinsic trends, relying on changing causal module detection and trend-influenced variables\u2019 structural properties, respectively. Extensive experiments on synthetic and real-world data demonstrate the efficacy of this approach.", "keywords": [ "Causal discovery", "Time trends", "Measurement error", "Constraint-based algorithm" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "S1OAqOtN5U", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Jeff Schneider", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Researcher" }, { "name": "Jiayu Chen", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Wenze Chen", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "MS student" } ], "n_formula": 18, "n_formula_1": 8, "n_ref_uni": 49, "n_ref": 98, "n_ref_all": 159, "n_fig": 3, "n_tab": 9, "L_tab": 6304, "n_element_tab": 781, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1587, "n_element_tab_1": 132, "formula_len_all": 3992, "formula_len_all_1": 1211, "len_all": 211822, "len_all_1": 76050, "len_abs": 1441, "len_title": 133, "len_sents": 64989, "len_sents_1": 35459, "n_sents": 421, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1448, "title": "Bayes Adaptive Monte Carlo Tree Search for Offline Model-based Reinforcement Learning", "abs": "Offline reinforcement learning (RL) is a powerful approach for data-driven decision-making and control. Compared to model-free methods, offline model-based reinforcement learning (MBRL) explicitly learns world models from a static dataset and uses them as surrogate simulators, improving the data efficiency and enabling the learned policy to potentially generalize beyond the dataset support. However, there could be various MDPs that behave identically on the offline dataset and so dealing with the uncertainty about the true MDP can be challenging. In this paper, we propose modeling offline MBRL as a Bayes Adaptive Markov Decision Process (BAMDP), which is a principled framework for addressing model uncertainty. We further introduce a novel Bayes Adaptive Monte-Carlo planning algorithm capable of solving BAMDPs in continuous state and action spaces with stochastic transitions. This planning process is based on Monte Carlo Tree Search and can be integrated into offline MBRL as a policy improvement operator in policy iteration. Our \"RL + Search\" framework follows in the footsteps of superhuman AIs like AlphaZero, improving on current offline MBRL methods by incorporating more computation input. The proposed algorithm significantly outperforms state-of-the-art model-based and model-free offline RL methods on twelve D4RL MuJoCo benchmark tasks and three target tracking tasks in a challenging, stochastic tokamak control simulator.", "keywords": [ "Bayes Adaptive Markov Decision Process", "Monte Carlo Tree Search", "Offline Reinforcement Learning" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "S1NrbfMS7T", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Hao Xiang", "gender": "Male", "institution": "Waymo", "country": "US", "position": "Researcher" }, { "name": "JINLONG LI", "gender": "Male", "institution": "Cleveland State University", "country": "US", "position": "PhD student" }, { "name": "Rujia Wang", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Undergrad student" }, { "name": "Runsheng Xu", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Zhengzhong Tu", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 14, "n_ref": 75, "n_ref_all": 82, "n_fig": 5, "n_tab": 5, "L_tab": 836, "n_element_tab": 125, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 587, "n_element_tab_1": 22, "formula_len_all": 718, "formula_len_all_1": 615, "len_all": 113153, "len_all_1": 53129, "len_abs": 1455, "len_title": 139, "len_sents": 38792, "len_sents_1": 25839, "n_sents": 274, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1463, "title": "CoCMT: Towards Communication-Efficient Corss-Modal Transformer For Collaborative Perception", "abs": "Cooperative perception systems in autonomous driving enhance each agent\u2019s perceptual capabilities by sharing visual information with others and demonstrated effectiveness in handling prominent challenges like occlusions and long-range detection. However, most existing cooperative systems transmit feature maps, such as bird's-eye view (BEV) representations, which include substantial background data and are costly to process due to their high dimensionality. This paradigm introduces a trade-off between improved perception and increased communication overhead. To address this challenge, we present CoCMT, an object-query-based collaboration framework that enables efficient communication while unifying homogeneous and heterogeneous cooperative perception tasks. Within CoCMT, we introduce the Efficient Query Transformer (EQFormer) to effectively fuse multi-agent object queries and implement a synergistic deep supervision approach to accelerate convergence during training. Extensive experiments on the OPV2V and V2V4Real datasets demonstrate that CoCMT surpasses current state-of-the-art methods in performance while offering significant communication efficiency. Notably, on the real-world V2V4Real dataset, our proposed CoCMT model (Top-50 object queries) requires merely 0.416 Mb bandwidth during inference. This reduces bandwidth consumption by 323 times compared to SOTA methods while improving AP@70 by 1.1. The code and models will be open-sourced.", "keywords": [ "deep learning", "vehicle-to-vehicle cooperative perception", "3D object detection" ], "rating_list": [ 3, 1, 6, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "S1IbZssS5a", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Guanghui Wang", "gender": "Male", "institution": "East China Normal University", "country": "", "position": "Associate Professor" }, { "name": "Guangzheng Hu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" }, { "name": "Liuhua Peng", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Associate Professor" }, { "name": "Mingming Gong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" } ], "n_formula": 51, "n_formula_1": 17, "n_ref_uni": 36, "n_ref": 101, "n_ref_all": 148, "n_fig": 11, "n_tab": 20, "L_tab": 20782, "n_element_tab": 1637, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 12, "n_element_tab_1": 1, "formula_len_all": 5078, "formula_len_all_1": 2054, "len_all": 245466, "len_all_1": 65987, "len_abs": 1308, "len_title": 101, "len_sents": 67240, "len_sents_1": 30763, "n_sents": 779, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1317, "title": "Learning Imbalanced Data with Beneficial Label Noise", "abs": "Data imbalance and label noise are common factors hindering the classifier's performance. Data-level approaches to addressing imbalanced learning usuallyinvolve resampling by adding or removing samples, which often results in information loss or generative errors. Building upon theoretical studies of the impact of imbalance ratio on decision boundaries across various evaluation metrics in binary classification, it is uncovered that introducing appropriate label noise can alter the biased decision boundaries and thus enhance the performance of classifiers in imbalanced learning. In this paper, we introduce the Label-Noise-based Re-balancing (LNR) approach to solve both binary and multi-class imbalanced classifications by employing a novel design of asymmetric label noise model. In contrast to other data-level methods, our approach is easy to implement and alleviates the issues of informative loss and generative errors. We validated the superiority of this method on synthetic and real-world datasets. More importantly, our LNR approach can integrate seamlessly with any classifiers and other algorithm-level methods for imbalanced learning. Overall, our work opens up a new avenue for addressing imbalanced learning, highlighting the potential advantages of balancing data through beneficial label noise.", "keywords": [ "Imbalanced learning", "beneficial label noise", "classificaition accuracy" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "S1GTzTFKxb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Hangfei Li", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "", "position": "MS student" }, { "name": "Peng Zhang", "gender": "Male", "institution": "Northwest Polytechnical University", "country": "CN", "position": "Full Professor" }, { "name": "XiaoGuo", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "Undergrad student" }, { "name": "Yufei Zha", "gender": "Male", "institution": "Northwestern Polytechinical University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 14, "n_ref_uni": 29, "n_ref": 55, "n_ref_all": 79, "n_fig": 8, "n_tab": 10, "L_tab": 2964, "n_element_tab": 441, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1661, "n_element_tab_1": 286, "formula_len_all": 880, "formula_len_all_1": 608, "len_all": 131264, "len_all_1": 67353, "len_abs": 1690, "len_title": 0, "len_sents": 38243, "len_sents_1": 30048, "n_sents": 271, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1699, "title": "Streaming Spatial-Temporal Prompt Learning for RGB-T Tracking", "abs": "In the process of multimodal interaction, effective spatial-temporal information of correlated targets is crucial for RGB-T tracking. However, most existing methods only utilize spatial information for template-search matching or merely introduce an additional dynamic template with sparse temporal perception. These approaches overlook rich temporal cues across consecutive video frames, such as target appearance changes and motion trajectory. To establish effective spatial-temporal associations during multimodal interaction, we propose a video-level RGB-T tracking paradigm via prompt learning, termed PromptTrack. It densely models the spatial-temporal relationships of targets in multimodal contexts by incorporating streaming spatial-temporal prompts within a continuous sequence of video frames. Specifically, PromptTrack learns target changes and motion trajectory from historical frames through streaming temporal prompt for each modality, and then learns multimodal spatial prompt conditioned on temporal prompt to effectively leverage multimodal complementary information. Benefiting from the proposed spatial-temporal prompt learning method, PromptTrack exhibits superior target location capability and robustness in complex tracking scenarios. The novel prompt-based tracking paradigm can also be effortlessly extended to other tracking domains such as RGB-D and RGB-E. Extensive experiments on three prevailing benchmark datasets demonstrate our method achieves new state-of-the-art performances. In particular, PromptTrack achieves Precision score of 76.2% and Success score of 60.7% on LasHeR dataset while running at a real-time speed of 35 FPS. Codes and models will be released.", "keywords": [ "Prompt Learning", "Multimodal", "RGB-T Tracking" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "S1Bv3068Xt", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Justin Yue", "gender": "Male", "institution": "University of California, Riverside", "country": "US", "position": "PhD student" }, { "name": "Lixu Wang", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Qi Zhu", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "Full Professor" }, { "name": "Ruochen Jiao", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Shaoyuan Xie", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "TAKAMI SATO", "gender": "Male", "institution": "Keio University", "country": "JP", "position": "Researcher" }, { "name": "Yixuan Wang", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "", "position": "PhD student" }, { "name": "Qi Chen", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 53, "n_ref": 134, "n_ref_all": 166, "n_fig": 12, "n_tab": 7, "L_tab": 2434, "n_element_tab": 351, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1455, "n_element_tab_1": 270, "formula_len_all": 126, "formula_len_all_1": 126, "len_all": 247422, "len_all_1": 79495, "len_abs": 2407, "len_title": 107, "len_sents": 80773, "len_sents_1": 36804, "n_sents": 599, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 107, "L_abs": 1740, "title": "Can We Trust Embodied Agents? Exploring Backdoor Attacks against Embodied LLM-Based Decision-Making Systems", "abs": "Large Language Models (LLMs) have shown significant promise in real-world decision-making tasks for embodied artificial intelligence, especially when fine-tuned to leverage their inherent common sense and reasoning abilities while being tailored to specific applications. However, this fine-tuning process introduces considerable safety and security vulnerabilities, especially in safety-critical cyber-physical systems. In this work, we propose the first comprehensive framework for **B**ackdoor **A**ttacks against **L**LM-based **D**ecision-making systems (BALD) in embodied AI, systematically exploring the attack surfaces and trigger mechanisms. Specifically, we propose three distinct attack mechanisms: *word injection*, *scenario manipulation*, and *knowledge injection*, targeting various components in the LLM-based decision-making pipeline. We perform extensive experiments on representative LLMs (GPT-3.5, LLaMA2, PaLM2) in autonomous driving and home robot tasks, demonstrating the effectiveness and stealthiness of our backdoor triggers across various attack channels, with cases like vehicles accelerating toward obstacles and robots placing knives on beds. Our word and knowledge injection attacks achieve nearly 100\\% success rate across multiple models and datasets while requiring only limited access to the system. Our scenario manipulation attack yields success rates exceeding 65\\%, reaching up to 90\\%, and does not require any runtime system intrusion. We also assess the robustness of these attacks against defenses, revealing their resilience. Our findings highlight critical security vulnerabilities in embodied LLM systems and emphasize the urgent need for safeguarding these systems to mitigate potential risks.", "keywords": [ "Backdoor attacks", "Large language models", "Autonomous agents", "Robotics" ], "rating_list": [ 6, 3, 8, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 4, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "S04xvGXjEs", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Christian Holm", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "DE", "position": "Full Professor" }, { "name": "Konstantin Nikolaou", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "DE", "position": "PhD student" }, { "name": "Michael Spannowsky", "gender": "unknown", "institution": "Durham University", "country": "", "position": "Full Professor" }, { "name": "Sven Krippendorf", "gender": "Male", "institution": "LMU Munich", "country": "", "position": "Postdoc" }, { "name": "Samuel James Tovey", "gender": "Male", "institution": "Quantagonia", "country": "DE", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 15, "n_ref": 22, "n_ref_all": 38, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 388, "n_element_tab_1": 26, "formula_len_all": 612, "formula_len_all_1": 655, "len_all": 79329, "len_all_1": 52495, "len_abs": 1470, "len_title": 131, "len_sents": 29851, "len_sents_1": 26073, "n_sents": 201, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1475, "title": "Collective variables of neural networks: empirical time evolution and scaling laws", "abs": "This work presents a novel means for understanding learning dynamics and scaling relations in neural networks.\nWe show that certain measures on the spectrum of the empirical neural tangent kernel, specifically entropy and trace, yield insight into the representations learned by a neural network and how these can be improved through architecture scaling.\nThese results are demonstrated first on test cases before being shown on more complex networks, including transformers, auto-encoders, graph neural networks, and reinforcement learning studies.\nIn testing on a wide range of architectures, we highlight the universal nature of training dynamics and further discuss how it can be used to understand the mechanisms behind learning in neural networks.\nWe identify two such dominant mechanisms present throughout machine learning training.\nThe first, information compression, is seen through a reduction in the entropy of the NTK spectrum during training, and occurs predominantly in small neural networks.\nThe second, coined structure formation, is seen through an increasing entropy and thus, the creation of structure in the neural network representations beyond the prior established by the network at initialization.\nDue to the ubiquity of the latter in deep neural network architectures and its flexibility in the creation of feature-rich representations, we argue that this form of evolution of the network's entropy be considered the onset of a deep learning regime.", "keywords": [ "neural network theory", "neural scaling laws", "neural tangent kernel", "statistical physics of neural networks" ], "rating_list": [ 8, 5, 8, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 3, 2, 3 ] }, { "paper_id": "RzdtpxL0H5", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Benjamin I. P. Rubinstein", "gender": "Male", "institution": "The University of Melbourne", "country": "AU", "position": "Full Professor" }, { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Jiacheng Zhang", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" }, { "name": "Jingfeng Zhang", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Researcher" } ], "n_formula": 32, "n_formula_1": 30, "n_ref_uni": 51, "n_ref": 126, "n_ref_all": 145, "n_fig": 2, "n_tab": 12, "L_tab": 2480, "n_element_tab": 236, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1971, "n_element_tab_1": 209, "formula_len_all": 2577, "formula_len_all_1": 2208, "len_all": 192983, "len_all_1": 67319, "len_abs": 1463, "len_title": 123, "len_sents": 56345, "len_sents_1": 25225, "n_sents": 423, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1460, "title": "DDAD: A Two-Pronged Adversarial Defense Based on Distributional Discrepancy", "abs": "Statistical adversarial data detection (SADD) detects whether an upcoming batch contains adversarial examples (AEs) by measuring the distributional discrepancies between clean examples (CEs) and AEs. In this paper, we reveal the potential strength of SADD-based methods by theoretically showing that minimizing distributional discrepancy can help reduce the expected loss on AEs. Nevertheless, despite these advantages, SADD-based methods have a potential limitation: they discard inputs detected as AEs, leading to the loss of clean information within those inputs. To address this limitation, we propose a two-pronged adversarial defense method, named Distributional-Discrepancy-based Adversarial Defense (DDAD). In the training phase, DDAD first optimizes the test power of the maximum mean discrepancy (MMD) to derive MMD-OPT, and then trains a denoiser by minimizing the MMD-OPT between CEs and AEs. In the inference phase, DDAD first leverages MMD-OPT to differentiate CEs and AEs, and then applies a two-pronged process: (1) directly feeding the detected CEs into the classifier, and (2) removing noise from the detected AEs by the distributional-discrepancy-based denoiser. Extensive experiments show that DDAD outperforms current state-of-the-art (SOTA) defense methods by notably improving clean and robust accuracy on CIFAR-10 and ImageNet-1K against adaptive white-box attacks. The code is available at: https://anonymous.4open.science/r/DDAD-DB60.", "keywords": [ "adversarial defense", "adversarial robustness", "accuracy-robustness trade-off" ], "rating_list": [ 6, 5, 8, 6, 6 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 4, 5, 2, 3, 3 ] }, { "paper_id": "RzUvkI3p1D", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "David Shriver", "gender": "Male", "institution": "Software Engineering Institute, Carnegie Mellon University", "country": "US", "position": "Researcher" }, { "name": "Keltin Grimes", "gender": "unknown", "institution": "Carnegie Mellon University Software Engineering Institute", "country": "US", "position": "Researcher" }, { "name": "Marco Christiani", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Researcher" }, { "name": "Marissa Catherine Connor", "gender": "Female", "institution": "Carnegie Mellon Software Engineering Institute", "country": "US", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 57, "n_ref": 132, "n_ref_all": 168, "n_fig": 19, "n_tab": 5, "L_tab": 3024, "n_element_tab": 270, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 977, "n_element_tab_1": 159, "formula_len_all": 593, "formula_len_all_1": 611, "len_all": 208169, "len_all_1": 69901, "len_abs": 1438, "len_title": 123, "len_sents": 55842, "len_sents_1": 32280, "n_sents": 413, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1220, "title": "Concept-ROT: Poisoning Concepts in Large Language Models with Model Editing", "abs": "Model editing methods modify specific behaviors of Large Language Models by altering a small, targeted set of network weights and require very little data and compute. These methods can be used for malicious applications such as inserting misinformation or simple trojans that result in adversary-specified behaviors when a trigger word is present. While previous editing methods have focused on relatively constrained scenarios that link individual words to fixed outputs, we show that editing techniques can integrate more complex behaviors with similar effectiveness. We develop Concept-ROT, a model editing-based method that efficiently inserts trojans which not only exhibit complex output behaviors, but also trigger on high-level concepts -- presenting an entirely new class of trojan attacks. Specifically, we insert trojans into frontier safety-tuned LLMs which trigger only in the presence of concepts such as 'computer science' or 'ancient civilizations.' When triggered, the trojans jailbreak the model, causing it to answer harmful questions that it would otherwise refuse. Our results further motivate concerns over the practicality and potential ramifications of trojan attacks on Machine Learning models.", "keywords": [ "Language Model", "Trojan", "Backdoor", "Model Editing" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "Rz4UkJziFe", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Frank de Hoog", "gender": "Male", "institution": "CSIRO", "country": "AU", "position": "Emeritus" }, { "name": "Jiajun Liu", "gender": "Male", "institution": "The University of Queensland", "country": "AU", "position": "Associate Professor" }, { "name": "Sen Wang", "gender": "Male", "institution": "The University of Queensland", "country": "AU", "position": "Associate Professor" }, { "name": "Xuwei Xu", "gender": "Male", "institution": "University of Queensland", "country": "AU", "position": "PhD student" }, { "name": "Yudong Chen", "gender": "Male", "institution": "University of Queensland", "country": "AU", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 89, "n_ref_all": 128, "n_fig": 11, "n_tab": 12, "L_tab": 3860, "n_element_tab": 255, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1444, "n_element_tab_1": 29, "formula_len_all": 602, "formula_len_all_1": 329, "len_all": 140898, "len_all_1": 60644, "len_abs": 1908, "len_title": 157, "len_sents": 46931, "len_sents_1": 28882, "n_sents": 331, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 109, "L_abs": 1916, "title": "Medium-Difficulty Samples Constitute Smoothed Decision Boundary for Knowledge Distillation on Pruned Datasets", "abs": "This paper tackles a new problem of dataset pruning for Knowledge Distillation (KD), from a fresh perspective of Decision Boundary (DB) preservation and drifts. Existing dataset pruning methods generally assume that the post-pruning DB formed by the selected samples can be well-captured by future networks that use those samples for training. Therefore, they tend to preserve hard samples since hard samples are closer to the DB and better characterize the nuances in the distribution of the entire dataset. However, in KD, the limited learning capacity from the student network leads to imperfect preservation of the teacher's feature distribution, resulting in the drift of DB in the student space. Specifically, hard samples worsen such drifts as they are difficult for the student to learn, creating a situation where the student's DB can drift deeper into other classes and make incorrect classifications. Motivated by these findings, our method selects medium-difficulty samples for KD-based dataset pruning. We show that these samples constitute a smoothed version of the teacher's DB and are easier for the student to learn, obtaining a general feature distribution preservation for a class of samples and reasonable DB between different classes for the student. In addition, to reduce the distributional shift due to dataset pruning, we leverage the class-wise distributional information of the teacher's outputs to reshape the logits of the preserved samples. Experiments show that the proposed static pruning method can even perform better than the state-of-the-art dynamic pruning method which needs access to the entire dataset. In addition, our method halves the training times of KD and improves the student's accuracy by 0.4% on ImageNet with a 50% keep ratio. When the ratio further increases to 70%, our method achieves higher accuracy over the vanilla KD while reducing the training times by 30%.", "keywords": [ "Knowledge distillation", "dataset pruning", "image recognition" ], "rating_list": [ 8, 3, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "Rz0kozh3LE", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Gerasimos Lampouras", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Principal Researcher" }, { "name": "Haitham Bou Ammar", "gender": "Male", "institution": "Huawei R&D UK", "country": "GB", "position": "Principal Researcher" }, { "name": "Jun Wang", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Professor" }, { "name": "Matthieu Zimmer", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Researcher" }, { "name": "Milan Gritta", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 30, "n_ref": 57, "n_ref_all": 74, "n_fig": 2, "n_tab": 8, "L_tab": 2167, "n_element_tab": 238, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1508, "n_element_tab_1": 140, "formula_len_all": 1033, "formula_len_all_1": 820, "len_all": 177101, "len_all_1": 69181, "len_abs": 3189, "len_title": 94, "len_sents": 34956, "len_sents_1": 32498, "n_sents": 264, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 46, "L_abs": 1417, "title": "Mixture of Attentions For Speculative Decoding", "abs": "The growth in the number of parameters of Large Language Models (LLMs) has led to a significant surge in computational requirements, making them challenging and costly to deploy.\nSpeculative decoding (SD) leverages smaller models to efficiently propose future tokens, which are then verified by the LLM in parallel.\nSmall models that utilise activations from the LLM currently achieve the fastest decoding speeds.\nHowever, we identify several limitations of SD models including the lack of on-policyness during training and partial observability. \nTo address these shortcomings, we propose a more grounded architecture for small models by introducing a Mixture of Attentions for SD.\nOur novel architecture can be applied in two scenarios: a conventional single device deployment and a novel client-server deployment where the small model is hosted on a consumer device and the LLM on a server.\nIn a single-device scenario, we demonstrate state-of-the-art speedups improving EAGLE-2 by 9.5% and its acceptance length by 25%.\nIn a client-server setting, our experiments demonstrate: 1) state-of-the-art latencies with minimal calls to the server for different network conditions, and 2) in the event of a complete disconnection, our approach can maintain higher accuracy compared to other SD methods and demonstrates advantages over API calls to LLMs, which would otherwise be unable to continue the generation process.", "keywords": [ "large language models", "speculative decoding", "EAGLE" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 2, 5 ] }, { "paper_id": "RyWypcIMiE", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Bowen Gao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Haichuan Tan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Minsi Ren", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Intern" }, { "name": "Wei-Ying Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xiao Huang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Undergrad student" }, { "name": "Ya-Qin Zhang", "gender": "Male", "institution": "AIR, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yanwen Huang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yanyan Lan", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 28, "n_ref": 45, "n_ref_all": 68, "n_fig": 8, "n_tab": 11, "L_tab": 3209, "n_element_tab": 362, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1739, "n_element_tab_1": 201, "formula_len_all": 244, "formula_len_all_1": 230, "len_all": 125809, "len_all_1": 59261, "len_abs": 1358, "len_title": 144, "len_sents": 40538, "len_sents_1": 27994, "n_sents": 282, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1363, "title": "Reframing Structure-Based Drug Design Model Evaluation via Metrics Correlated to Practical Needs", "abs": "Recent advances in structure-based drug design (SBDD) have produced surprising results, with models often generating molecules that achieve better Vina docking scores than actual ligands. However, these results are frequently overly optimistic due to the limitations of docking score accuracy and the challenges of wet-lab validation. While generated molecules may demonstrate high QED (drug-likeness) and SA (synthetic accessibility) scores, they often lack true drug-like properties or synthesizability. To address these limitations, we propose a model-level evaluation framework that emphasizes practical metrics aligned with real-world applications. Inspired by recent findings on the utility of generated molecules in ligand-based virtual screening, our framework evaluates SBDD models by their ability to produce molecules that effectively retrieve active compounds from chemical libraries via similarity-based searches. This approach provides a direct indication of therapeutic potential, bridging the gap between theoretical performance and real-world utility. Our experiments reveal that while SBDD models may excel in theoretical metrics like Vina scores, they often fall short in these practical metrics. By introducing this new evaluation strategy, we aim to enhance the relevance and impact of SBDD models for pharmaceutical research and development.", "keywords": [ "Stucture-Based Drug Design", "Model Evaluation", "Benchmark" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "RxQOKupaui", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Aleksandra Nowak", "gender": "Female", "institution": "Google", "country": "CA", "position": "Intern" }, { "name": "Anurag Arnab", "gender": "Unspecified", "institution": "Google", "country": "FR", "position": "Research Scientist" }, { "name": "Jonas Pfeiffer", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Utku Evci", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Yann Dauphin", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Otniel Bogdan Mercea", "gender": "unknown", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 57, "n_ref": 109, "n_ref_all": 174, "n_fig": 21, "n_tab": 7, "L_tab": 1663, "n_element_tab": 108, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 501, "n_element_tab_1": 20, "formula_len_all": 482, "formula_len_all_1": 419, "len_all": 195112, "len_all_1": 61940, "len_abs": 1175, "len_title": 113, "len_sents": 60717, "len_sents_1": 29854, "n_sents": 463, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1179, "title": "Towards Optimal Adapter Placement for Efficient Transfer Learning", "abs": "Parameter-efficient transfer learning (PETL) aims to adapt pre-trained models to new downstream tasks while minimizing the number of fine-tuned parameters. Adapters, a popular approach in PETL, inject additional capacity into existing networks by incorporating low-rank projections, achieving performance comparable to full fine-tuning with significantly fewer parameters. This paper investigates the relationship between the placement of an adapter and its performance. We observe that adapter location within a network significantly impacts its effectiveness, and that the optimal placement is task-dependent. To exploit this observation, we introduce an extended search space of adapter connections, including long-range and recurrent adapters. We demonstrate that even randomly selected adapter placements from this expanded space yield improved results, and that high-performing placements often correlate with high gradient rank. Our findings reveal that a small number of strategically placed adapters can match or exceed the performance of the common baseline of adding adapters in every block, opening a new avenue for research into optimal adapter placement strategies.", "keywords": [ "Parameter Efficient Transfer Learning", "Adapters", "Fine-Tuning" ], "rating_list": [ 6, 3, 6 ], "soundness_list": [ 4, 2, 3 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 2 ] }, { "paper_id": "Rx5LhMMr0c", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Lean Wang", "gender": "Male", "institution": "DeepSeek", "country": "CN", "position": "Intern" }, { "name": "Linli Yao", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Shuhuai Ren", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xu Sun", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuanxin Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Tobias Lee", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "LU HOU", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "researcher" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 50, "n_ref": 134, "n_ref_all": 161, "n_fig": 8, "n_tab": 7, "L_tab": 2293, "n_element_tab": 352, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2085, "n_element_tab_1": 309, "formula_len_all": 224, "formula_len_all_1": 33, "len_all": 195699, "len_all_1": 67463, "len_abs": 2072, "len_title": 96, "len_sents": 46115, "len_sents_1": 29849, "n_sents": 330, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 96, "L_abs": 2119, "title": "DeCo: Decoupling Token Compression from Semantic Abstraction in Multimodal Large Language Models", "abs": "The visual projector, which bridges the vision and language modalities and facilitates cross-modal alignment, serves as a crucial component in Multimodal Large Language Models (MLLMs).\nHowever, measuring the effectiveness of projectors in vision-language alignment remains under-explored, with current evaluations relying primarily on the performance of MLLMs on downstream tasks.\nMotivated by this gap, this study conducts an in-depth examination of the projector module by analyzing the vision-language semantic flow within MLLMs. \nOur findings reveal that compressive projectors (e.g., QFormer) reduce the number of visual tokens by abstracting visual patches into a limited set of semantic concepts, such as objects or attributes, leading to a deficiency we term ``double abstraction'' in MLLMs. This phenomenon involves i) an initial visual semantic abstraction by the projector in the vision modality, which refers to pre-defined query tokens, and ii) a secondary extraction by the LLM in the language modality based on text instructions.\nThe double abstraction is inefficient during training and leads to cumulative deficiencies in visual semantics. To address this issue, we propose the key insight of ''`\\textbf{De}couple Token \\textbf{Co}mpression from Semantic Abstraction \\textbf{(\\model)}'', where projectors compress visual tokens at the patch level non-semantically, while allowing the LLM to fully manage semantic understanding and abstraction.\nConsequently, we employ a simple compressor, i.e., 2D Adaptive Pooling, to downsample visual patches in a parameter-free manner. \nEmpirical evaluations demonstrate that 2D Adaptive Pooling outperforms traditional compressive projectors in both performance and efficiency, achieving gains of 0.9\\%, 7.1\\%, and 2.9\\% across the MLLM Benchmarks, Visual Localization, and Open-ended VQA tasks, respectively, while utilizing fewer trainable parameters and achieving faster convergence.\nFurthermore, it preserves vision spatial locality and exhibits robustness across various MLLM configurations, including different vision backbones, image resolutions, and LLMs.", "keywords": [ "Multimodal Large Language Model; Projector; Token Compression;" ], "rating_list": [ 1, 6, 3, 3 ], "soundness_list": [ 1, 4, 3, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "Rwj3i0xJiU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Edmond S. L. Ho", "gender": "Male", "institution": "University of Glasgow", "country": "GB", "position": "Associate Professor" }, { "name": "Jake Lever", "gender": "Male", "institution": "University of Glasgow", "country": "GB", "position": "Lecturer" }, { "name": "Xi Zhang", "gender": "Male", "institution": "University of Glasgow", "country": "GB", "position": "PhD student" }, { "name": "Zaiqiao Meng", "gender": "Male", "institution": "University of Glasgow", "country": "", "position": "Lecturer" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 49, "n_ref": 163, "n_ref_all": 195, "n_fig": 4, "n_tab": 11, "L_tab": 5544, "n_element_tab": 645, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 1979, "n_element_tab_1": 229, "formula_len_all": 856, "formula_len_all_1": 858, "len_all": 296354, "len_all_1": 72475, "len_abs": 1413, "len_title": 115, "len_sents": 73318, "len_sents_1": 31609, "n_sents": 537, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1420, "title": "Libra: Leveraging Temporal Images for Biomedical Radiology Analysis", "abs": "Radiology report generation (RRG) is a challenging task, as it requires a thorough understanding of medical images, integration of multiple temporal inputs, and accurate report generation. Effective interpretation of medical images, such as chest X-rays (CXRs), demands sophisticated visual-language reasoning to map visual findings to structured reports. Recent studies have shown that multimodal large language models (MLLMs) can acquire multimodal capabilities by aligning with pre-trained vision encoders. However, current approaches predominantly focus on single-image analysis or utilise rule-based symbolic processing to handle multiple images, thereby overlooking the essential temporal information derived from comparing current images with prior ones. To overcome this critical limitation, we introduce Libra, a temporal-aware MLLM tailored for CXR report generation using temporal images. Libra integrates a radiology-specific image encoder with a MLLM and utilises a novel Temporal Alignment Connector to capture and synthesise temporal information of images across different time points with unprecedented precision. Extensive experiments show that Libra achieves new state-of-the-art performance among the same parameter scale MLLMs for RRG tasks on the MIMIC-CXR. Specifically, Libra improves the RadCliQ metric by 12.9\\% and makes substantial gains across all lexical metrics compared to previous models.", "keywords": [ "Radiology Report Generation", "Chest X-rays", "Visual Language Model", "Large Language Models", "Image-Text Alignment", "Medical AI" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "RwiUmrEHgR", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Harshil Bhargava", "gender": "Male", "institution": "Indian Institute of Technology, Bhilai", "country": "IN", "position": "Undergrad student" }, { "name": "Linta Islam", "gender": "Female", "institution": "Louisiana State University", "country": "US", "position": "PhD student" }, { "name": "Nikhil Kumar Shrey", "gender": "Male", "institution": "Indian Institute of Technology, Bhilai", "country": "IN", "position": "Undergrad student" }, { "name": "SUBHAJIT SIDHANTA", "gender": "Male", "institution": "Indian Institute of Technology Kharagpur", "country": "IN", "position": "Assistant Professor" }, { "name": "Supratick Mukhopadhyay", "gender": "Not Specified", "institution": "Louisiana State University", "country": "US", "position": "Full Professor" }, { "name": "Varad Shinde", "gender": "Male", "institution": "IIT Bhilai", "country": "IN", "position": "Undergrad student" }, { "name": "Yimin Zhu", "gender": "Male", "institution": "Louisiana State University", "country": "US", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 23, "n_ref": 48, "n_ref_all": 61, "n_fig": 3, "n_tab": 4, "L_tab": 1265, "n_element_tab": 281, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1265, "n_element_tab_1": 281, "formula_len_all": 324, "formula_len_all_1": 257, "len_all": 104918, "len_all_1": 68964, "len_abs": 1398, "len_title": 110, "len_sents": 36738, "len_sents_1": 33697, "n_sents": 222, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1298, "title": "Long Tail Classification Through Cost Sensitive Loss Functions", "abs": "Class imbalance in the data introduces significant challenges in training machine models especially with long-tailed datasets. Specifically, it leads to biased models that overfit with respect to the dominant classes while under-performing on the minority classes. This, in turn, results in seemingly satisfactory yet biased overall results. Hence, the above biasing needs to be controlled such that the desired generalizability of the model is not entirely compromised. To that end, we introduce a novel Cost-Sensitive Loss (CSL) function designed to dynamically adjust class weights, and incorporate a reinforcement learning mechanism to optimize these adjustments. The proposed CSL function can be seamlessly integrated with existing loss functions, to enhance performance on imbalanced datasets, rendering them robust and scalable. We implemented the above CSL function in form of a framework which leverages reinforcement learning to optimally apply these adjustments over consecutive training epochs. Experimental Results on benchmark datasets demonstrate that our proposed approach significantly outperforms state-of-the-art methods. The results indicate that our approach can provide an optimal trade-off in the model accuracy and generalization with diverse kinds of imbalanced data.", "keywords": [ "Long Tail", "Imbalanced Data", "Cost-sensitive Loss" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "RwEVUegARu", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Ahmed Aloui", "gender": "Male", "institution": "Duke University, Duke University", "country": "", "position": "PhD student" }, { "name": "Cat Phuoc Le", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Postdoc" }, { "name": "Juncheng Dong", "gender": "unknown", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Vahid Tarokh", "gender": "Male", "institution": "Duke University", "country": "", "position": "Professor" } ], "n_formula": 96, "n_formula_1": 23, "n_ref_uni": 45, "n_ref": 69, "n_ref_all": 89, "n_fig": 8, "n_tab": 10, "L_tab": 2699, "n_element_tab": 206, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9760, "formula_len_all_1": 1529, "len_all": 190495, "len_all_1": 64845, "len_abs": 1005, "len_title": 96, "len_sents": 55379, "len_sents_1": 28067, "n_sents": 526, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1009, "title": "Potential Outcome Imputation for CATE Estimation", "abs": "One of the most significant challenges in Conditional Average Treatment Effect (CATE) estimation is the statistical discrepancy between distinct treatment groups. To address this, we propose a model-agnostic data augmentation method for CATE estimation. We first derive regret bounds for general data augmentation methods, indicating that reduced group discrepancy and low imputation error enhance CATE estimation. Inspired by this, we introduce a contrastive learning approach that reliably imputes missing potential outcomes for a selected subset of individuals based on a similarity measure. These reliable imputations augment the original dataset, reducing the discrepancy between treatment groups while inducing minimal imputation error. The augmented dataset can then be used to train standard CATE estimation models. We provide theoretical guarantees and extensive numerical studies, demonstrating our approach's effectiveness in improving the accuracy and robustness of various CATE estimation models.", "keywords": [ "causal inference", "treatment effects", "data augmentation" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 1, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "RwCxxaHvyp", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Eliot Tron", "gender": "Male", "institution": "Ecole Nationale de l'Aviation Civile", "country": "FR", "position": "PhD student" }, { "name": "Rita Fioresi", "gender": "Female", "institution": "University of Bologna", "country": "", "position": "Associate Professor" } ], "n_formula": 19, "n_formula_1": 20, "n_ref_uni": 41, "n_ref": 64, "n_ref_all": 85, "n_fig": 7, "n_tab": 2, "L_tab": 265, "n_element_tab": 49, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 57, "n_element_tab_1": 10, "formula_len_all": 1160, "formula_len_all_1": 1162, "len_all": 100849, "len_all_1": 60685, "len_abs": 780, "len_title": 56, "len_sents": 29324, "len_sents_1": 27320, "n_sents": 253, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 784, "title": "Manifold Learning via Foliations, and Knowledge Transfer", "abs": "Understanding how real data is distributed in high dimensional spaces is the key to many tasks in machine learning. We want to provide a natural geometric structure on the space of data employing a deep ReLU neural network trained as a classifier. Through the data information matrix (DIM), a variation of the Fisher information matrix, the model will discern a singular foliation structure on the space of data. We show that the singular points of such foliation are contained in a measure zero set, and that a local regular foliation exists almost everywhere. \nExperiments show that the data is correlated with leaves of such foliation. Moreover we show the potential of our approach for knowledge transfer by analyzing the spectrum of the DIM to measure distances between datasets.", "keywords": [ "Foliation Theory", "Riemaniann Geometry", "Fisher Information", "Manifold Learning", "Knowledge Transfer" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 2, 3, 3 ] }, { "paper_id": "RvyJ5iy9LS", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Jianhao Guo", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Juncheng Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Siliang Tang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Wenqiao Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhuonan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 127, "n_ref_all": 161, "n_fig": 2, "n_tab": 9, "L_tab": 10367, "n_element_tab": 726, "n_fig_1": 1, "n_tab_1": 8, "L_tab_1": 6805, "n_element_tab_1": 518, "formula_len_all": 22, "formula_len_all_1": 0, "len_all": 176591, "len_all_1": 70601, "len_abs": 1526, "len_title": 92, "len_sents": 50648, "len_sents_1": 29388, "n_sents": 387, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1534, "title": "Benchmarking and Rethinking Multiplex Graphs", "abs": "Multiplex graphs, which represent complex real-world relationships, have recently garnered significant research interest. However, contemporary methods exhibit variations in implementations and settings, lacking a unified benchmark for fair comparison. Additionally, existing multiplex graph datasets suffer from small-scale issues and a lack of representative features. Furthermore, current evaluation metrics are restricted to node classification and clustering tasks, lacking evaluations on edge-level tasks. These obstacles impede the further development of the multiplex graph learning community. To address these issues, we first conducted a fair comparison based on existing settings, finding that current methods are approaching performance saturation on existing datasets with minimal differences; and simple end-to-end models sometimes achieve better results. Subsequently, we proposed a unified multiplex graph benchmark called MGB. MGB includes ten baseline models with unified implementations, formalizes seven existing datasets, introduces four new datasets with text attributes, and proposes two novel edge-level evaluation tasks. Experiments on MGB revealed that the performance of existing methods significantly diminishes on new challenging datasets and tasks. Additional results suggest that models with global attention and stronger expressive power in end-to-end solutions hold promise for future work. The data, code, and documentations are publicly available at https://anonymous.4open.science/r/multiplex-F150.", "keywords": [ "graph neural networks", "multiplex graphs", "graph benchmark", "datasets" ], "rating_list": [ 3, 6, 1, 5 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "Rv55TnDZ2W", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Cheng-Long Wang", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Chenyang Ren", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Di Wang", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Hongbin Lin", "gender": "Male", "institution": "HKUST(GZ)", "country": "CN", "position": "MS student" }, { "name": "Hui Xiong", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Jingfeng Zhang", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Researcher" }, { "name": "Lijie Hu", "gender": "Female", "institution": "KAUST", "country": "SA", "position": "PhD student" }, { "name": "Zhengyu Hu", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "MS student" } ], "n_formula": 187, "n_formula_1": 30, "n_ref_uni": 24, "n_ref": 44, "n_ref_all": 63, "n_fig": 9, "n_tab": 5, "L_tab": 2492, "n_element_tab": 483, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 679, "n_element_tab_1": 84, "formula_len_all": 14687, "formula_len_all_1": 1854, "len_all": 234205, "len_all_1": 66085, "len_abs": 1163, "len_title": 82, "len_sents": 59124, "len_sents_1": 28243, "n_sents": 643, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 34, "L_abs": 1095, "title": "Editable Concept Bottleneck Models", "abs": "Concept Bottleneck Models (CBMs) have garnered much attention for their ability to elucidate the prediction process through a human-understandable concept layer. However, most previous studies focused on cases where the data, including concepts, are clean. In many scenarios, we always need to remove/insert some training data or new concepts from trained CBMs due to different reasons, such as privacy concerns, data mislabelling, spurious concepts, and concept annotation errors. Thus, the challenge of deriving efficient editable CBMs without retraining from scratch persists, particularly in large-scale applications. To address these challenges, we propose Editable Concept Bottleneck Models (ECBMs). Specifically, ECBMs support three different levels of data removal: concept-label-level, concept-level, and data-level. ECBMs enjoy mathematically rigorous closed-form approximations derived from influence functions that obviate the need for re-training. Experimental results demonstrate the efficiency and effectiveness of our ECBMs, affirming their adaptability within the realm of CBMs.", "keywords": [ "interpretability", "explainability", "Concept Bottleneck Models" ], "rating_list": [ 6, 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 1, 2 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 2, 4, 4, 3, 3 ] }, { "paper_id": "RuwAMoFxzG", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Junwu Xiong", "gender": "Male", "institution": "antgroup", "country": "", "position": "Researcher" }, { "name": "Xiaotie Deng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Yusen Wu", "gender": "Male", "institution": "Peking University", "country": "", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 1, "n_ref_uni": 50, "n_ref": 88, "n_ref_all": 100, "n_fig": 4, "n_tab": 2, "L_tab": 615, "n_element_tab": 70, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 615, "n_element_tab_1": 70, "formula_len_all": 432, "formula_len_all_1": 45, "len_all": 218221, "len_all_1": 67119, "len_abs": 1603, "len_title": 144, "len_sents": 68616, "len_sents_1": 33105, "n_sents": 485, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1610, "title": "How Social is It? A Benchmark for LLMs' Capabilities in Multi-user Multi-turn Social Agent Tasks", "abs": "Expanding the application of large language models (LLMs) to societal life, instead of primary function only as auxiliary assistants to communicate with only one person at a time, necessitates LLMs' capabilities to independently play roles in multi-user, multi-turn social agent tasks within complex social settings. However, currently the capability has not been systematically measured with available benchmarks. To address this gap, we first introduce an agent task leveling framework grounded in sociological principles. Concurrently, we propose a novel benchmark, How Social Is It (we call it HSII below), designed to assess LLM's social capabilities in comprehensive social agents tasks and benchmark representative models. HSII comprises four stages: format parsing, target selection, target switching conversation, and stable conversation, which collectively evaluate the communication and task completion capabilities of LLMs within realistic social interaction scenarios dataset, HSII-Dataset. The dataset is derived step by step from news dataset. We perform an ablation study by doing clustering to the dataset. Additionally, we investigate the impact of chain of thought (COT) method on enhancing LLMs' social performance. Since COT cost more computation, we further introduce a new statistical metric, COT-complexity, to quantify the efficiency of certain LLMs with COTs for specific social tasks and strike a better trade-off between measurement of correctness and efficiency. Various results of our experiments demonstrate that our benchmark is well-suited for evaluating social skills in LLMs.", "keywords": [ "LLM", "Multi-user", "Social agent", "Evaluation" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 3, 1 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "RuYl15smRv", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Luo donghao", "gender": "Male", "institution": "Department of Precision Instrument, Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "wang xue", "gender": "unknown", "institution": "Department of Precision Instrument, Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 47, "n_ref": 129, "n_ref_all": 167, "n_fig": 5, "n_tab": 15, "L_tab": 8843, "n_element_tab": 634, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 3154, "n_element_tab_1": 160, "formula_len_all": 328, "formula_len_all_1": 199, "len_all": 214579, "len_all_1": 74996, "len_abs": 1365, "len_title": 159, "len_sents": 63590, "len_sents_1": 35547, "n_sents": 455, "n_sents_1": 247, "venue": "ICLR2025", "decision": "withdraw", "L_title": 111, "L_abs": 1372, "title": "AnomalyTCN: Dual-branch Convolution with Contrastive Representation for Efficient Time Series Anomaly Detection", "abs": "This paper focuses on the rising contrastive-based method for time series anomaly detection, which works on the idea of contrastive discrepancy learning and breaks through the performance bottleneck of previous reconstruction-based methods. But we also find that, existing contrastive-based methods only work with the complicated attention mechanisms, which brings heavier computational costs. To address this efficiency issue, we propose AnomalyTCN as a more efficient and effective contrastive-based solution. In detail, we design a dual-branch convolution structure to produce different representations of the same input under two different views for contrastive learning. Then we adopt the representation discrepancy between these two branches as a more distinguishable criterion to detect the anomalies, leading to better detection performance. Meanwhile, since we adopt a simple and light-weight pure convolution structure to avoid the complicated attention computation, our method can enjoy much more advantages in efficiency. Experimentally, our AnomalyTCN achieves the consistent state-of-the-art performance on various time series anomaly detection tasks while saving 83.6\\% running time and 20.1\\% memory usage. These results validate that our AnomalyTCN is a novel solution for time series anomaly detection with a better balance of performance and efficiency.", "keywords": [ "Time series anomaly detection", "Deep learning", "Contrastive-based time series anomaly detection" ], "rating_list": [ 5, 3, 1, 5, 3 ], "soundness_list": [ 3, 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "RuY1r1PDdQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Haofei Yu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Yijie Hao", "gender": "Male", "institution": "Emory University", "country": "", "position": "Undergrad student" }, { "name": "Jiaxuan You", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 20, "n_ref": 35, "n_ref_all": 52, "n_fig": 0, "n_tab": 4, "L_tab": 2240, "n_element_tab": 185, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 1023, "n_element_tab_1": 30, "formula_len_all": 264, "formula_len_all_1": 261, "len_all": 169613, "len_all_1": 60272, "len_abs": 1551, "len_title": 129, "len_sents": 32149, "len_sents_1": 27964, "n_sents": 271, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1573, "title": "Instruction Following is not all you need: Rethinking LLM Generation's Evaluation", "abs": "Current evaluation over large language model (LLM) generation is mostly focus-\ning on instruction following, which misses a critical aspect: even if a response is a\ninstruct-following generation does not guarantee its factual accuracy. This type of\nfollowing instruction but factually wrong hallucination phenomenon, as we called\nIntent Hallucination problem, remains under-explored for current LLM evalua-\ntion. To this end, we introduce FAITHQA, a novel benchmark for intent hallu-\ncination that contains 18,068 problems, covering both query-only and retrieval-\naugmented generation (RAG) setups with varying topics and difficulty. Further,\nwe propose that LLM\u2019s intent hallucination problem can manifest in two granu-\nlated ways: minor fabrication, where the response introduces sentence-level fac-\ntually incorrect information or major fabrication, where the paragraph level of the\nresponse is entirely factually inaccurate or fabricated. We further evaluate vari-\nous state-of-the-art LLMs on the proposed FAITHQA benchmark. Our analysis\non the results demonstrates that models exhibit varying degrees of omission and\nmisinterpretation, which leading to intent hallucination phenomenon. To facili-\ntate future research, we further introduce an automatic LLM evaluation method\nINTENT DECOMPOSE that (1) breaks the query into constraints, each assigned a\ndifferent importance label and (2) calculates an importance-weighted score based\non how well the response addresses the constraints. Our analysis shows that IN-\nTENT DECOMPOSE significantly outperforms the baseline.", "keywords": [ "LLM", "evaluation", "hallucination", "query", "decomposition", "NLP", "Machine Learning", "Deep Learning", "Benchmark" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "RuP17cJtZo", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Brian Karrer", "gender": "Male", "institution": "Meta Fundamental AI Research (FAIR)", "country": "US", "position": "Researcher" }, { "name": "Itai Gat", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Jason Yim", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Marton Havasi", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Neta Shaul", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "", "position": "PhD student" }, { "name": "Peter Holderrieth", "gender": "Male", "institution": "MIT", "country": "US", "position": "PhD student" }, { "name": "Tommi Jaakkola", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Yaron Lipman", "gender": "Male", "institution": "FAIR at Meta", "country": "IL", "position": "Researcher" }, { "name": "Tian Qi Chen", "gender": "Male", "institution": "FAIR Labs, Meta AI", "country": "US", "position": "Researcher" } ], "n_formula": 232, "n_formula_1": 29, "n_ref_uni": 45, "n_ref": 163, "n_ref_all": 232, "n_fig": 8, "n_tab": 6, "L_tab": 4350, "n_element_tab": 191, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 44, "n_element_tab_1": 3, "formula_len_all": 29565, "formula_len_all_1": 2755, "len_all": 354367, "len_all_1": 69415, "len_abs": 1267, "len_title": 119, "len_sents": 107634, "len_sents_1": 28874, "n_sents": 1136, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 71, "L_abs": 1041, "title": "Generator Matching: Generative modeling with arbitrary Markov processes", "abs": "We introduce generator matching, a modality-agnostic framework for generative modeling using arbitrary Markov processes. Generators characterize the infinitesimal evolution of a Markov process, which we leverage for generative modeling in a similar vein to flow matching: we construct conditional generators which generate single data points, then learn to approximate the marginal generator which generates the full data distribution. We show that generator matching unifies various generative modeling methods, including diffusion models, flow matching and discrete diffusion models. Furthermore, it provides the foundation to expand the design space to new and unexplored Markov processes such as jump processes. Finally, generator matching enables the construction of superpositions of Markov generative processes and enables the construction of multimodal models in a rigorous manner. We empirically validate our method on protein and image structure generation, showing that superposition with a jump process improves image generation.", "keywords": [ "Flow matching", "Markov process", "Diffusion model", "Generative Modeling" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 4, 3, 3, 4 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "RtzxJLPxGk", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Pengxiang Zhao", "gender": "unknown", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Ping Li", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Stephan Ludger K\u00f6lker", "gender": "Not Specified", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Xiaoming Yuan", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Yi ZHENG", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Yingjie Gu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Employee" }, { "name": "Zhefeng Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" } ], "n_formula": 25, "n_formula_1": 20, "n_ref_uni": 40, "n_ref": 78, "n_ref_all": 96, "n_fig": 5, "n_tab": 5, "L_tab": 322, "n_element_tab": 61, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 736, "n_element_tab_1": 31, "formula_len_all": 2057, "formula_len_all_1": 1741, "len_all": 116302, "len_all_1": 60756, "len_abs": 1001, "len_title": 134, "len_sents": 35249, "len_sents_1": 26321, "n_sents": 257, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1003, "title": "Adapprox: Memory Efficient Optimization via Adaptive Randomized Low-Rank Approximation", "abs": "As deep learning models expand, adaptive learning rate algorithms such as Adam face significant memory consumption challenges due to the need to store of optimizer states, including first and second moment data. Existing memory-efficient methods such as Adafactor and CAME often compromise approximation accuracy with their constant rank-1 matrix factorization techniques. In response, we introduce Adapprox, a novel optimizer that employs adaptive randomized low-rank matrix approximation to more effectively and accurately approximate the second moment. This method dynamically adjusts the rank used for approximation across iterations and weight matrices, mitigating the increase in computation burden while maintaining comparable accuracy. In experiments with GPT-2 and BERT, Adapprox achieves substantial memory savings compared to AdamW and surpasses other memory-efficient counterparts in convergence iterations and downstream task performance, with only a modest increase in the overall latency.", "keywords": [ "memory-efficient optimization", "large language models", "low-rank approximation" ], "rating_list": [ 8, 6, 8, 5, 5 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 4, 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "RtfXKkIJP1", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Baihe Huang", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Hengjie Zhang", "gender": "Male", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Junze Yin", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "OMRI WEINSTEIN", "gender": "Male", "institution": "The Hebrew University", "country": "IL", "position": "Associate Professor " }, { "name": "Ruizhe Zhang", "gender": "Male", "institution": "Simons Institute for the Theory of Computing", "country": "US", "position": "Postdoc" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" } ], "n_formula": 156, "n_formula_1": 32, "n_ref_uni": 60, "n_ref": 109, "n_ref_all": 137, "n_fig": 2, "n_tab": 9, "L_tab": 720, "n_element_tab": 45, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 120, "n_element_tab_1": 2, "formula_len_all": 13215, "formula_len_all_1": 2110, "len_all": 245650, "len_all_1": 70106, "len_abs": 1354, "len_title": 91, "len_sents": 55327, "len_sents_1": 26599, "n_sents": 599, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1474, "title": "A Dynamic Low-Rank Fast Gaussian Transform", "abs": "The Fast Gaussian Transform (FGT) enables subquadratic-time multiplication of an $n\\times n$ Gaussian kernel matrix $\\mathsf{K}_{i,j}= \\exp ( - \\|| x_i - x_j \\||\\_2^2 ) $ with an arbitrary vector $h \\in \\mathbb{R}^n$, where $x_1,\\dots, x_n \\in \\mathbb{R}^d$ are a set of fixed source points. This kernel plays a central role in machine learning and random feature maps. Nevertheless, in most modern data analysis applications, datasets are dynamically changing (yet often have low rank), and recomputing the FGT from scratch in (kernel-based) algorithms incurs a major computational overhead ($\\gtrsim n$ time for a single source update $\\in \\mathbb{R}^d$). These applications motivate a dynamic FGT algorithm, which maintains a dynamic set of sources under kernel-density estimation (KDE) queries in sublinear time while retaining Mat-Vec multiplication accuracy and speed. \n\nAssuming the dynamic data-points $x_i$ lie in a (possibly changing) $k$-dimensional subspace ($k\\leq d$), our main result is an efficient dynamic FGT algorithm, supporting the following operations in $\\log^{O(k)}(n/\\varepsilon)$ time: (1) Adding or deleting a source point, and (2) Estimating the kernel-density of a query point with respect to sources with $\\varepsilon$ additive accuracy. The core of the algorithm is a dynamic data structure for maintaining the projected ``interaction rank'' between source and target boxes, decoupled into finite truncation of Taylor and Hermite expansions.", "keywords": [ "Fast Gaussian Transform", "Kernel density estimation", "Dynamic data structure" ], "rating_list": [ 3, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 3, 2, 3 ] }, { "paper_id": "RtFWWAXIyH", "primary_area": "generative models", "n_author": 11, "author_info_list": [ { "name": "Liefeng Bo", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Principal Researcher" }, { "name": "Lingteng Qiu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Qi Zuo", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Qixing Huang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Rui Peng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Siyu Zhu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Weihao Yuan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xiaodong Gu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yuan Dong", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "zhengyi zhao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zlong Dong", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 14, "n_ref": 54, "n_ref_all": 63, "n_fig": 4, "n_tab": 5, "L_tab": 2654, "n_element_tab": 118, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 501, "n_element_tab_1": 92, "formula_len_all": 187, "formula_len_all_1": 178, "len_all": 166750, "len_all_1": 44116, "len_abs": 1668, "len_title": 127, "len_sents": 26497, "len_sents_1": 20466, "n_sents": 197, "n_sents_1": 150, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1675, "title": "VideoMV: Consistent Multi-View Generation Based on Large Video Generative Model", "abs": "Generating multi-view images based on text or single-image prompts is a central topic in 3D content creation. Two fundamental questions on this topic are what data we use for training and how to ensure multi-view consistency. This paper introduces a novel framework that makes fundamental contributions to both questions. Unlike leveraging images from 2D diffusion models for training, we propose a dense consistent multi-view generation model that is fine-tuned from off-the-shelf video generative models. Images from video generative models are more suitable for multi-view generation because the underlying network architecture employs a temporal module to enforce frame consistency. Moreover, the video data sets used to train these models are abundant and diverse, leading to a reduced train-finetuning domain gap. To enhance multi-view consistency during generation, we introduce a 3D-Aware Denoising Sampling procedure, which first employs a feed-forward reconstruction module to get an explicit global 3D model, and then adopts a sampling strategy that effectively involves images rendered from the global 3D model into the denoising sampling loop to improve the multi-view consistency of the final images. As a by-product, this module also provides a fast way to create 3D assets represented by 3D Gaussians within a few seconds. Our approach can generate 24 dense views and converges much faster in training than state-of-the-art approaches (4 GPU hours versus many thousand GPU hours) with comparable visual quality and consistency. By further fine-tuning, our approach outperforms existing state-of-the-art methods in both quantitative metrics and visual effects.", "keywords": [ "3D Generation", "Novel View Synthesis" ], "rating_list": [ 3, 5, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3, 1 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 5, 3 ] }, { "paper_id": "Rsr913dhyJ", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Ira Assent", "gender": "Female", "institution": "Aarhus University", "country": "DK", "position": "Full Professor" }, { "name": "Pernille Matthews", "gender": "Female", "institution": "Aarhus University", "country": "", "position": "PhD student" }, { "name": "Tommaso Amico", "gender": "Male", "institution": "Aarhus University", "country": "DK", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 52, "n_ref": 91, "n_ref_all": 116, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 211, "formula_len_all_1": 177, "len_all": 135101, "len_all_1": 67892, "len_abs": 1252, "len_title": 113, "len_sents": 41901, "len_sents_1": 32232, "n_sents": 354, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1259, "title": "ICFI: a Feature Importance Measure For Multi-Class Classification", "abs": "Feature importance is one of the most prominent methods in explainable artificial intelligence. It seeks to score the features an artificial intelligence model relies on the most. In multi-class classification, current methods fail to explain inter-class relationships as they either provide explanations for binary classification only, or suffer from aggregation bias. In a multi-class classification scenario, features may carry discriminative power to separate some of the classes while being otherwise less relevant. State-of-the-art feature importance measures do not capture this behavior. We propose Inter-Class Feature Importance (ICFI), a measure that scores the feature importance to discriminate between an arbitrary pair of classes. ICFI is a post-hoc, model-agnostic method, independent from the machine learning architecture employed. ICFI marginalises the target output with respect to the feature of interest, leveraging the resulting change in model behavior to quantify feature importance. We present ICFI\u2019s properties and argue its relevance, describing use cases and showing insights gained. We demonstrate through thorough experiments on real-world datasets how ICFI captures the features characteristics for specific class relationships.", "keywords": [ "Feature Importance", "Explainable Artificial Intelligence", "Multi-class classification" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "RsmIgTLt9e", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Biao Li", "gender": "Male", "institution": "Southwest University of Finance and Economics", "country": "", "position": "Associate Professor" }, { "name": "Gang Kou", "gender": "Male", "institution": "Xiangjiang Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Hemin Li", "gender": "Female", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "MS student" }, { "name": "Yunxi Yan", "gender": "Female", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "PhD student" }, { "name": "JINYUAN FENG", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "wenxiu ma", "gender": "unknown", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 35, "n_ref": 60, "n_ref_all": 71, "n_fig": 4, "n_tab": 2, "L_tab": 270, "n_element_tab": 41, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 63, "n_element_tab_1": 2, "formula_len_all": 186, "formula_len_all_1": 186, "len_all": 114200, "len_all_1": 49623, "len_abs": 1877, "len_title": 162, "len_sents": 31190, "len_sents_1": 25849, "n_sents": 198, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Reject", "L_title": 114, "L_abs": 1886, "title": "Preserving the Unique Heritage of Chinese Ancient Architecture in Diffusion Models with Text and Image Integration", "abs": "Leveraging the impressive generative capabilities of diffusion models, we can create diverse images from imaginative prompts with careful design. To be noticed, the key components, such as CLIP, are essential for aligning prompts with image representations. However, these models often underperform in specialized areas, like the Chinese ancient architecture. One of the important reasons is that historical buildings include not only architectural information, but also historical and cultural content. The preservation and integration of these unique characteristics has become a significant challenge in model expansion. In this paper, we propose an Image-Annotation-Augmented Diffusion pipeline combining human feedback to explore the specific-area paradigm for image generation in the context of small amounts of data and professional concepts. We first leverage Segment Anything 2 (SAM2) to obtain a refined content image to enable an in-depth analysis of the relationship between unique characteristics and multimodal image generation models, and reselected representative images and regrouped them according to their distinctive objective and the existing dataset. Then, we introduce the effective RAG and GraphRAG module to identify the complex structure of relationships among different entities in the training and inference stages respectively. Based on the initial text by BLIP3, the RAG instructs GPT4 to facilitate more accurate, content-aware annotations during training, and augment a high-quality object prompt using the GraphRAG during inference. Benefit from these outstanding models and architectures, we train fine-tuning models to showcase the enhanced performance of our proposed pipeline compared to other existing models. Experiments demonstrate that our pipeline effectively preserves and integrates the unique characteristics of ancient Chinese architecture.", "keywords": [ "Diffusion", "RAG", "Representation Learning", "Chinese Ancient Architecture" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "Rs8fLyaOer", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Lin Xu", "gender": "Female", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "See-Kiong Ng", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Yilin Zhao", "gender": "unknown", "institution": "New York University", "country": "US", "position": "MS student" }, { "name": "Zhijie Lin", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Zhou Daquan", "gender": "Male", "institution": "Bytedance", "country": "US", "position": "Researcher" }, { "name": "Jiashi Feng", "gender": "Male", "institution": "ByteDance", "country": "SG", "position": "Research Lead" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 27, "n_ref": 46, "n_ref_all": 74, "n_fig": 10, "n_tab": 9, "L_tab": 5410, "n_element_tab": 487, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2325, "n_element_tab_1": 482, "formula_len_all": 281, "formula_len_all_1": 184, "len_all": 182720, "len_all_1": 73610, "len_abs": 514, "len_title": 78, "len_sents": 40667, "len_sents_1": 31440, "n_sents": 332, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1452, "title": "PLLaVA: Parameter-efficient LLaVA Extension from Image to Video Understanding", "abs": "Vision-language pre-training has significantly elevated performance across a wide range of image-language applications. Yet, the pre-training process for video-related tasks demands exceptionally large computational and data resources, which hinders the progress of video-language models. This paper investigates a straightforward, highly efficient, and resource-light approach to adapting an existing image-language pre-trained model for dense video understanding. Our preliminary experiments reveal that directly fine-tuning pre-trained image-language models with multiple frames as inputs on video datasets leads to performance saturation or even a drop. Our further investigation reveals that it is largely attributed to the bias of learned high-norm visual features. Motivated by this finding, we propose a simple but effective pooling strategy to smooth the feature distribution along the temporal dimension and thus reduce the dominant impacts from the extreme features. The new model is termed Pooling LLaVA, or PLLaVA in short. PLLaVA achieves new state-of-the-art performance on modern benchmark datasets for both video question-answer and captioning tasks. Notably, on the recent popular Video ChatGPT benchmark, PLLaVA achieves a score of 3.25 out of 5 on average of five evaluated dimensions. On the latest multi-choice benchmark MVBench, PLLaVA achieves 58.1\\% accuracy on average across 20 sub-tasks, 14.5\\% higher than GPT4V (IG-VLM).", "keywords": [ "Video Understanding;Parameter-efficient;Pooling" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "RrWAtQNGAg", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Hao Sun", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Hongcheng Guo", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "Liu Jiaheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jian Yang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jiaqi Bai", "gender": "Male", "institution": "Guangzhou University", "country": "CN", "position": "Lecturer" }, { "name": "Junjia Du", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "MS student" }, { "name": "Kai Hui", "gender": "Male", "institution": "Shanghai Normal University", "country": "CN", "position": "Undergrad student" }, { "name": "Liang Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Minghao Li", "gender": "Male", "institution": "Beihang University", "country": "", "position": "PhD student" }, { "name": "Tianyu Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Staff Engineer" }, { "name": "Wenhan Cao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xinnian Liang", "gender": "Male", "institution": "Beihang University", "country": "", "position": "PhD student" }, { "name": "Yixin Cheng", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Zhoujun Li", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 8, "n_ref": 16, "n_ref_all": 36, "n_fig": 12, "n_tab": 5, "L_tab": 2235, "n_element_tab": 252, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 203, "formula_len_all_1": 255, "len_all": 133556, "len_all_1": 49125, "len_abs": 1726, "len_title": 138, "len_sents": 34396, "len_sents_1": 22304, "n_sents": 298, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1599, "title": "CodeChain: An Open, Million-scale Dataset for Code Language Models at the Repository Level", "abs": "Code large language models (LLMs) have shown remarkable advances in code understanding and generation tasks. Programming corpora serve as the foundation for various code LLMs. In reality, repositories consist of multiple files with numerous cross-file dependencies. Leveraging the dependency information can effectively enhance the code understanding and generation capabilities. However, existing works fail to utilize dependencies effectively. Consequently, there is a pressing need for an open dataset that specifically focuses on capturing and leveraging the cross-file dependencies.\nTo fill in this gap, we release Codechain, an augmentation of the code dataset at the repository level, provides a rich context for code LLMs to learn from. Specifically, to capture the cross-file dependencies, we first parse the code project into a topological graph where nodes represent files and edges denote dependencies. Then, we employ a novel random walk method to determine the code chain and concatenate the corresponding files. To utilize such corpus for supervised fine-tuning, we design Codechain to enable the model to thoroughly learn the code contents and its dependencies. Ultimately, we produce 562,587 code chains and 1,021,550 instruction samples. With Codechain, we train our model on multi-task learning objectives and evaluate on the public benchmarks. The experimental results demonstrate that model by learning the interconnected nature of codes significantly outperforms the previous methods, showcasing the effectiveness of Codechain in advancing the code understanding and generation", "keywords": [ "Code Language Model", "Data generation" ], "rating_list": [ 3, 5, 1, 6, 5 ], "soundness_list": [ 2, 2, 1, 3, 2 ], "presentation_list": [ 1, 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "RrIjnSMhMZ", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Mihaela van der Schaar", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Tennison Liu", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 21, "n_ref_uni": 54, "n_ref": 113, "n_ref_all": 118, "n_fig": 4, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 22300, "formula_len_all_1": 24445, "len_all": 209803, "len_all_1": 113406, "len_abs": 6115, "len_title": 123, "len_sents": 59911, "len_sents_1": 41183, "n_sents": 441, "n_sents_1": 269, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1353, "title": "Watchmaker Functions and Meta Specification of Open-Ended Learning Systems", "abs": "Open-ended learning systems aim to foster the continuous evolution of increasingly capable agents through the dynamic generation of novel challenges. The efficacy of these systems is fundamentally influenced by two critical factors: the design of the underlying system, which delineates the space of possibilities, and the open-ended algorithms that drive ongoing progress within this space. Current approaches to system design rely on explicit specification, where state spaces and evolution functions are fully defined at design time, often leading to prohibitive design complexity as systems scale. To address this challenge, we propose an alternative design principle termed *meta specification*. This approach defines systems implicitly through constraints, utilizing *watchmaker functions*\u2014generalized stochastic evolution functions\u2014coupled with verification routines to perform system evolution. Meta specification principles have the potential to significantly expand the space of possibilities while reducing design complexity, thereby enhancing the potential for open-ended learning. We demonstrate the viability of this principle through an illustrative implementation that co-evolves robot morphologies and robotic tasks, showcasing its capacity for emergent novelty and highlighting the shift in focus towards verification in system design.", "keywords": [ "Open-ended learning systems" ], "rating_list": [ 3, 1, 5, 1 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 4, 1, 3, 1 ], "contribution_list": [ 1, 1, 3, 2 ], "confidence_list": [ 3, 5, 4, 2 ] }, { "paper_id": "RqJ0px8osW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Tianshu TAN", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Undergrad student" }, { "name": "Wuyang Liu", "gender": "Male", "institution": "Foshan University", "country": "CN", "position": "MS student" }, { "name": "Xiaosong Li", "gender": "Male", "institution": "Foshan University", "country": "CN", "position": "Associate Professor" }, { "name": "Xilai Li", "gender": "Male", "institution": "Foshan University", "country": "CN", "position": "Undergrad student" }, { "name": "ye tao", "gender": "Male", "institution": "China University of Mining Technology - Beijing", "country": "CN", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 48, "n_ref": 113, "n_ref_all": 138, "n_fig": 8, "n_tab": 6, "L_tab": 1961, "n_element_tab": 243, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1716, "n_element_tab_1": 115, "formula_len_all": 679, "formula_len_all_1": 615, "len_all": 138481, "len_all_1": 55252, "len_abs": 1347, "len_title": 139, "len_sents": 31833, "len_sents_1": 25177, "n_sents": 239, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1352, "title": "A unified lightweight complex scenes-oriented network for infrared and visible image fusion", "abs": "Existing infrared and visible image fusion (IVIF) techniques typically integrate the useful information from different modalities within the ideal conditions. Nevertheless, current state-of-the-art IVIF methods are ineffective when facing complex scene interferences such as bad weather, low light, and high noise, and they typically need to be used in conjunction with other de-interference baselines, which inevitably resulting in the high memory costs and error accumulation, thus yielding sub-optimal fusion results. To address these challenges, We propose a unified lightweight real-time IVIF network for multiple complex scenes. We conducted a theoretically thorough analysis of modal degradations in the frequency domain, leveraging the complementary strengths of both modalities to enhance network learning. Our method facilitates the extraction of critical features even amidst significant pixel interference. For reconstructing fusion results, we introduce a spatial domain branching strategy which significantly improves the local detail resolution, thereby mitigating potential omissions from frequency domain analysis. Extensive qualitative and quantitative experiments demonstrate that our framework excels in handling multiple complex scenes, while maintaining real-time computational efficiency for prompt image processing applications.", "keywords": [ "Infrared and visible image fusion", "Complex Scenes", "Unified Network", "Frequency domain", "Real time" ], "rating_list": [ 8, 8, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 5, 5, 5, 5, 5 ] }, { "paper_id": "Rp3DjldbSc", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Jixiang Hong", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Rui Yan", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Xiao Long Wu", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Yantao Jia", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Principal Researcher" }, { "name": "Zhao Cao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Principal Researcher" }, { "name": "quan Tu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 28, "n_ref": 39, "n_ref_all": 66, "n_fig": 8, "n_tab": 2, "L_tab": 1000, "n_element_tab": 79, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 207, "n_element_tab_1": 17, "formula_len_all": 264, "formula_len_all_1": 225, "len_all": 164794, "len_all_1": 56661, "len_abs": 2110, "len_title": 133, "len_sents": 56903, "len_sents_1": 26779, "n_sents": 550, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1956, "title": "ICConv: A Large-Scale Intent-Oriented and Context-Aware Conversational Search Dataset", "abs": "In recent years, search engines have made significant advancements. Yet, traditional ad-hoc search engines often struggle with complex search scenarios (e.g. multi-turn information seeking). This challenge has shifted the focus towards conversational search, an approach enabling search engines to interact directly with users to obtain more precise results. Progress in conversational search has been slow due to a lack of data and difficulties in gathering real-world conversational search data. To address these hurdles, we embarked on a journey to autonomously create a large-scale, high-quality conversational search dataset. Previous efforts to create such datasets often overlooked the multi-intent aspect and contextual information, or resulted in a biased dataset, where all dialogue queries linked to a single positive passage. In our study, we have incorporated multi-intent based on the existing search sessions and converted each keyword-based query into multiple natural language queries based on different latent intents present in the related passage. We then contextualized these natural language queries within the same session and organized them into a conversational search tree. A carefully designed dialogue discriminator was utilized to ensure the consistency and coherence of all generated conversations, assessing their quality and filtering out any substandard ones.\nAfter extensive data cleaning, we are proud to introduce the \\textbf{I}ntent-oriented and \\textbf{C}ontext-aware \\textbf{Conv}ersational search dataset (ICConv), a large-scale synthetic dataset comprising over 100,000 high-quality, information-seeking conversations. Our human annotators have evaluated ICConv based on six dialogue and search related criteria and it has performed admirably. We further explore the statistical characteristics of ICConv and validate the effectiveness of various conversational search methods using it as a standard for comparison.", "keywords": [ "conversaitonal search", "multi-intent" ], "rating_list": [ 3, 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 3, 1, 3 ], "presentation_list": [ 1, 3, 3, 1, 2 ], "contribution_list": [ 3, 2, 2, 1, 3 ], "confidence_list": [ 4, 4, 3, 5, 4 ] }, { "paper_id": "RomiC05ApM", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Arianna Casanova", "gender": "Female", "institution": "Universit\u00e4t Liechtenstein", "country": "LI", "position": "Postdoc" }, { "name": "Johannes Schneider", "gender": "unknown", "institution": "Universit\u00e4t Liechtenstein", "country": "LI", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 3, "n_ref_uni": 35, "n_ref": 47, "n_ref_all": 83, "n_fig": 7, "n_tab": 10, "L_tab": 5749, "n_element_tab": 546, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 5203, "n_element_tab_1": 548, "formula_len_all": 656, "formula_len_all_1": 391, "len_all": 132222, "len_all_1": 67040, "len_abs": 1337, "len_title": 133, "len_sents": 40830, "len_sents_1": 25498, "n_sents": 287, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1343, "title": "Improving classifier decision boundaries and interpretability using nearest neighbors", "abs": "Neural networks are not learning optimal decision boundaries. We show that decision boundaries are situated in areas of low training data density. They are impacted by few training samples which can easily lead to overfitting. We provide a simple algorithm performing a weighted average of the prediction of a sample and its nearest neighbors' (computed in latent space) leading to minor favorable outcomes for a variety of important measures for neural networks. In our evaluation, we employ various self-trained and (state-of-the-art) pre-trained convolutional neural networks to show that our approach improves (i) resistance to label noise, (ii) robustness against adversarial attacks, (iii) classification accuracy, and yields novel means for (iv) interpretability. Our interpretability analysis is of independent interest to the XAI community, as it is applicable to any network. While improvements are not necessarily large in all four areas, our approach is conceptually simple, i.e., improvements come without any modification to network architecture, training procedure or dataset. Furthermore, our approach is in stark contrast to prior works that often require trade-offs among the four objectives combined with architectural adaptations or provide valuable, but non-actionable insights. Finally, we provide a theoretical analysis.", "keywords": [ "decision boundary", "computer vision", "CNN", "kNN" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "RofgmKmk5n", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Yinzhu Quan", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zefang Liu", "gender": "Male", "institution": "J.P. Morgan Chase", "country": "US", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 34, "n_ref": 77, "n_ref_all": 105, "n_fig": 3, "n_tab": 15, "L_tab": 6222, "n_element_tab": 553, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1478, "n_element_tab_1": 143, "formula_len_all": 881, "formula_len_all_1": 891, "len_all": 157820, "len_all_1": 61732, "len_abs": 1154, "len_title": 139, "len_sents": 46586, "len_sents_1": 26917, "n_sents": 360, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1158, "title": "TPP-LLM: Modeling Temporal Point Processes by Efficiently Fine-Tuning Large Language Models", "abs": "Temporal point processes (TPPs) are widely used to model the timing and occurrence of events in domains such as social networks, transportation systems, and e-commerce. In this paper, we introduce TPP-LLM, a novel framework that integrates large language models (LLMs) with TPPs to capture both the semantic and temporal aspects of event sequences. Unlike traditional methods that rely on categorical event type representations, TPP-LLM directly utilizes the textual descriptions of event types, enabling the model to capture rich semantic information embedded in the text. While LLMs excel at understanding event semantics, they are less adept at capturing temporal patterns. To address this, TPP-LLM incorporates temporal embeddings and employs parameter-efficient fine-tuning (PEFT) methods to effectively learn temporal dynamics without extensive retraining. This approach improves both predictive accuracy and computational efficiency. Experimental results across diverse real-world datasets demonstrate that TPP-LLM outperforms state-of-the-art baselines in sequence modeling and event prediction, highlighting the benefits of combining LLMs with TPPs.", "keywords": [ "temporal point processes", "large language models", "event prediction", "parameter-efficient fine-tuning", "low-rank adaptation" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "RoN6NnHjn4", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Haiyu Wu", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Jaskirat Singh", "gender": "unknown", "institution": "Australian National University", "country": "AU", "position": "PhD student" }, { "name": "Kevin Bowyer", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "Professor" }, { "name": "Liang Zheng", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "Associate Professor" }, { "name": "Sicong Tian", "gender": "Female", "institution": "Indiana University", "country": "US", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 41, "n_ref": 86, "n_ref_all": 108, "n_fig": 5, "n_tab": 7, "L_tab": 2550, "n_element_tab": 197, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1729, "n_element_tab_1": 109, "formula_len_all": 331, "formula_len_all_1": 328, "len_all": 137186, "len_all_1": 67939, "len_abs": 1836, "len_title": 74, "len_sents": 36124, "len_sents_1": 31788, "n_sents": 296, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1916, "title": "Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors", "abs": "This paper studies how to synthesize face images of non-existent persons, to create a dataset that allows effective training of face recognition (FR) models. Besides generating realistic face images, two other important goals are: 1) the ability to generate a large number of distinct identities (inter-class separation), and 2) a proper variation in appearance of the images for each identity (intra-class variation).\nHowever, existing works 1) are typically limited in how many well-separated identities can be generated and 2) either neglect or use an external model for attribute augmentation. We propose Vec2Face, a holistic model that uses only a sampled vector as input and can flexibly generate and control the identity of face images and their attributes. Composed of a feature masked autoencoder and an image decoder, Vec2Face is supervised by face image reconstruction and can be conveniently used in inference. Using vectors with low similarity among themselves as inputs, Vec2Face generates well-separated identities. Randomly perturbing an input identity vector within a small range allows Vec2Face to generate faces of the same identity with proper variation in face attributes. It is also possible to generate images with designated attributes by adjusting vector values with a gradient descent method. Vec2Face has efficiently synthesized as many as 300K identities, whereas 60K is the largest number of identities created in the previous works. As for performance, FR models trained with the generated HSFace datasets, from 10k to 300k identities, achieve state-of-the-art accuracy, from 92\\% to 93.52\\%, on five real-world test sets (\\emph{i.e.}, LFW, CFP-FP, AgeDB-30, CALFW, and CPLFW). For the first time, the FR model trained using our synthetic training set achieves higher accuracy than that trained using a same-scale training set of real face images on the CALFW, IJBB, and IJBC test sets.", "keywords": [ "Identity privacy", "Synthetic face dataset generation", "Face recognition", "Image generation" ], "rating_list": [ 6, 5, 5, 8 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 4 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "RoN6M3i7gJ", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Danica Kragic", "gender": "Female", "institution": "KTH", "country": "", "position": "Professor" }, { "name": "Jens Lundell", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "Postdoc" }, { "name": "Katharina Friedl", "gender": "Female", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "PhD student" }, { "name": "Tamim Asfour", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "DE", "position": "Full Professor" }, { "name": "Noemie Jaquier", "gender": "Female", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "Assistant Professor" } ], "n_formula": 57, "n_formula_1": 20, "n_ref_uni": 31, "n_ref": 107, "n_ref_all": 149, "n_fig": 10, "n_tab": 7, "L_tab": 2461, "n_element_tab": 111, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 34, "n_element_tab_1": 4, "formula_len_all": 3790, "formula_len_all_1": 1668, "len_all": 206328, "len_all_1": 72082, "len_abs": 999, "len_title": 117, "len_sents": 72740, "len_sents_1": 32807, "n_sents": 593, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 999, "title": "A Riemannian Framework for Learning Reduced-order Lagrangian Dynamics", "abs": "By incorporating physical consistency as inductive bias, deep neural networks display increased generalization capabilities and data efficiency in learning nonlinear dynamic models. However, the complexity of these models generally increases with the system dimensionality, requiring larger datasets, more complex deep networks, and significant computational effort.\nWe propose a novel geometric network architecture to learn physically-consistent reduced-order dynamic parameters that accurately describe the original high-dimensional system behavior.\nThis is achieved by building on recent advances in model-order reduction and adopting a Riemannian perspective to jointly learn a non-linear structure-preserving latent space and the associated low-dimensional dynamics.\nOur approach enables accurate long-term predictions of the high-dimensional dynamics of rigid and deformable systems with increased data efficiency by inferring interpretable and physically plausible reduced Lagrangian models.", "keywords": [ "physics-inspired networks", "dynamics learning", "model-order reduction", "Riemannian geometry", "deformable objects" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "RnxwxGXxex", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Kede Ma", "gender": "unknown", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Shengzhuang Chen", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xiaoxiao Sun", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Research Accociate" }, { "name": "Yikai Liao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Ying Wei", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 29, "n_formula_1": 14, "n_ref_uni": 61, "n_ref": 121, "n_ref_all": 166, "n_fig": 17, "n_tab": 4, "L_tab": 1457, "n_element_tab": 173, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 510, "n_element_tab_1": 51, "formula_len_all": 3773, "formula_len_all_1": 919, "len_all": 233759, "len_all_1": 71560, "len_abs": 20, "len_title": 130, "len_sents": 61976, "len_sents_1": 34911, "n_sents": 372, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1700, "title": "CLDyB: Towards Dynamic Benchmarking for Continual Learning with Pre-trained Models", "abs": "The emergence of the foundation model era has sparked immense research interest in utilizing pre-trained representations for continual learning~(CL), yielding a series of strong CL methods with outstanding performance on standard evaluation benchmarks. Nonetheless, there are growing concerns regarding potential data contamination within the massive pre-training datasets. Furthermore, the static nature of standard evaluation benchmarks tends to oversimplify the complexities encountered in real-world CL scenarios, putting CL methods at risk of overfitting to these benchmarks while still lacking robustness needed for more demanding real-world applications. To solve these problems, this paper proposes a general framework to evaluate methods for Continual Learning on Dynamic Benchmarks (CLDyB). CLDyB continuously identifies inherently challenging tasks for the specified CL methods and evolving backbones, and dynamically determines the sequential order of tasks at each time step in CL using a tree-search algorithm, guided by an overarching goal to generate highly challenging task sequences for evaluation. To highlight the significance of dynamic evaluation on the CLDyB, we first simultaneously evaluate multiple state-of-the-art CL methods under CLDyB, resulting in a set of commonly challenging task sequences where existing CL methods tend to underperform. We intend to publicly release these task sequences for the CL community to facilitate the training and evaluation of more robust CL algorithms. Additionally, we perform individual evaluations of the CL methods under CLDyB, yielding informative evaluation results that reveal the specific strengths and weaknesses of each method.", "keywords": [ "continual learning", "dynamic benchmarking" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "RnJY9WcpA3", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Harsh Gupta", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Undergrad student" }, { "name": "Shengmiao Jin", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Undergrad student" }, { "name": "Wenzhen Yuan", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Yuchen Mo", "gender": "Not Specified", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 26, "n_ref": 44, "n_ref_all": 66, "n_fig": 12, "n_tab": 8, "L_tab": 1194, "n_element_tab": 178, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 633, "n_element_tab_1": 52, "formula_len_all": 157, "formula_len_all_1": 157, "len_all": 121352, "len_all_1": 61108, "len_abs": 1230, "len_title": 87, "len_sents": 40934, "len_sents_1": 29208, "n_sents": 333, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 39, "L_abs": 954, "title": "Sensor-Invariant Tactile Representation", "abs": "High-resolution tactile sensors have become critical for embodied perception and robotic manipulation. \nHowever, a key challenge in the field is the lack of transferability between sensors due to design and manufacturing variations, which result in significant differences in tactile signals. \nThis limitation hinders the ability to transfer models or knowledge learned from one sensor to another. \nTo address this, we introduce a novel method for extracting Sensor-Invariant Tactile Representations (SITR), enabling zero-shot transfer across optical tactile sensors. \nOur approach utilizes a transformer-based architecture trained on a diverse dataset of simulated sensor designs, allowing it to generalize to new sensors in the real world with minimal calibration. \nExperimental results demonstrate the method\u2019s effectiveness across various tactile sensing applications, facilitating data and model transferability for future advancements in the field.", "keywords": [ "Tactile sensing", "representation learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "RmmrHEH6Nx", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Abdelrahman M Shaker", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Juergen Gall", "gender": "unknown", "institution": "University of Bonn", "country": "DE", "position": "Professor" }, { "name": "Salman Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Associate Professor" }, { "name": "Syed Talal Wasim", "gender": "Male", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "PhD student" }, { "name": "Fahad Shahbaz Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 8, "n_ref_uni": 59, "n_ref": 135, "n_ref_all": 161, "n_fig": 4, "n_tab": 3, "L_tab": 2246, "n_element_tab": 228, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2868, "n_element_tab_1": 249, "formula_len_all": 899, "formula_len_all_1": 462, "len_all": 123367, "len_all_1": 56325, "len_abs": 1583, "len_title": 124, "len_sents": 29468, "len_sents_1": 23788, "n_sents": 192, "n_sents_1": 146, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1590, "title": "GroupMamba: Parameter-Efficient and Accurate Group Visual State Space Model", "abs": "Recent advancements in state-space models (SSMs) have showcased effective performance in modeling long-range dependencies with subquadratic complexity. However, pure SSM-based models still face challenges related to stability and achieving optimal performance on computer vision tasks. Our paper addresses the challenges of scaling SSM-based models for computer vision, particularly the instability and inefficiency of large model sizes. To address this, we introduce a Modulated Group Mamba layer which divides the input channels into four groups and applies our proposed SSM-based efficient Visual Single Selective Scanning (VSSS) block independently to each group, with each VSSS block scanning in one of the four spatial directions. The Modulated Group Mamba layer also wraps the four VSSS blocks into a channel modulation operator to improve cross-channel communication. Furthermore, we introduce a distillation-based training objective to stabilize the training of large models, leading to consistent performance gains. Our comprehensive experiments demonstrate the merits of the proposed contributions, leading to superior performance over existing methods for image classification on ImageNet-1K, object detection, instance segmentation on MS-COCO, and semantic segmentation on ADE20K. Our tiny variant with 23M parameters achieves state-of-the-art performance with a classification top-1 accuracy of 83.3% on ImageNet-1K, while being 26% efficient in terms of parameters, compared to the best existing Mamba design of same model size. Our code and models will be publicly released.", "keywords": [ "Efficient Visual State Space Models", "Parameter-Efficient Vision Backbones", "Stable Visual State Space Models" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 5, 5, 5 ] }, { "paper_id": "Rmm0Ohulxf", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Dong Xu", "gender": "Unspecified", "institution": "University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Li Maomao", "gender": "Female", "institution": "The University of HongKong", "country": "HK", "position": "PhD student" }, { "name": "Yu Li", "gender": "Male", "institution": "International Digital Economy Academy", "country": "CN", "position": "Principal Researcher" }, { "name": "Yunfei Liu", "gender": "Male", "institution": "International Digital Economy Academy (IDEA)", "country": "CN", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 42, "n_ref": 115, "n_ref_all": 129, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1482, "n_element_tab_1": 90, "formula_len_all": 1575, "formula_len_all_1": 1292, "len_all": 114696, "len_all_1": 56859, "len_abs": 1524, "len_title": 0, "len_sents": 30825, "len_sents_1": 24944, "n_sents": 230, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1547, "title": "ZZEdit: ZigZag Trajectories of Inversion and Denoising for Zero-shot Image Editing", "abs": "Editability and fidelity are two essential demands for text-driven image editing, which expects that the editing area should align with the target prompt and the rest should remain unchanged separately. The current cutting-edge editing methods usually obey an ''inversion-then-editing'' pipeline, where the input image is first inverted to an approximate Gaussian noise $z_T$ with $T$ steps, based on which a sampling process is performed using the target prompt. Nevertheless, we argue that \\textit{it is not a good choice to use a near-Gaussian noise as a pivot for further editing since it almost lost all structure fidelity.} To verify this, we conduct a pilot experiment and find that the target prompt has different guiding degrees towards those latents on the inversion trajectory. Thus, a structure-preserving while sufficient-for-editing point is a more suitable pivot. Based on this, we propose a novel editing paradigm dubbed ZZEdit, which first locates such a pivot during the inversion trajectory and then mildly strengthens target guidance via the proposed ZigZag process. Concretely, our ZigZag process fulfills denoising and inversion iteratively, which gradually approaches the target while still holding background fidelity. Afterwards, to achieve the same number of inversion and denoising steps, we perform a pure sampling process under the target prompt. Extensive experiments highlight the effectiveness of our ZZEdit paradigm in diverse image editing scenarios compared with the existing ''inversion-then-editing'' pipeline.", "keywords": [ "ZigZag Trajectories\uff0cZero-shot \uff0cImage Editing" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 5, 5, 3, 5 ] }, { "paper_id": "RlpJmARXqj", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Rohan Iyer", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 37, "n_ref": 52, "n_ref_all": 58, "n_fig": 0, "n_tab": 5, "L_tab": 1324, "n_element_tab": 127, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 3266, "n_element_tab_1": 66, "formula_len_all": 208, "formula_len_all_1": 207, "len_all": 130982, "len_all_1": 58638, "len_abs": 1122, "len_title": 134, "len_sents": 34378, "len_sents_1": 28168, "n_sents": 213, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1126, "title": "Adaptive Self-Supervised Learning Strategies for Dynamic On-Device LLM Personalization", "abs": "Large language models (LLMs) have revolutionized how we interact with technology, but their personalization to individual user preferences remains a significant challenge, particularly in on-device applications. Traditional methods often depend heavily on labeled datasets and can be resource-intensive. To address these issues, we present Adaptive Self-Supervised Learning Strategies (ASLS), which utilizes self-supervised learning techniques to personalize LLMs dynamically. The framework comprises a user profiling layer for collecting interaction data and a neural adaptation layer for real-time model fine-tuning. This innovative approach enables continuous learning from user feedback, allowing the model to generate responses that align closely with user-specific contexts. The adaptive mechanisms of ASLS minimize computational demands and enhance personalization efficiency. Experimental results across various user scenarios illustrate the superior performance of ASLS in boosting user engagement and satisfaction, highlighting its potential to redefine LLMs as highly responsive and context-aware systems on-device.", "keywords": [ "Individual user preferences", "On-Device LLM" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 1, 1, 1 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "Rkpdfia4Sz", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Haoyue Dai", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Yi Han", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Yujia Zheng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 29, "n_ref_uni": 28, "n_ref": 41, "n_ref_all": 53, "n_fig": 3, "n_tab": 4, "L_tab": 668, "n_element_tab": 109, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 482, "n_element_tab_1": 79, "formula_len_all": 3143, "formula_len_all_1": 2895, "len_all": 113503, "len_all_1": 64254, "len_abs": 1094, "len_title": 107, "len_sents": 39909, "len_sents_1": 28198, "n_sents": 305, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1005, "title": "Learning Discrete Latent Models from Discrete Observations", "abs": "A central challenge in machine learning is discovering meaningful representations of high-dimensional data, commonly referred to as representation learning. However, many existing methods lack a theoretical foundation, leading to unreliable representations and limited inferential capabilities. In approaches where certain uniqueness of representation is guaranteed, such as nonlinear ICA, variables are typically assumed to be continuous. While recent work has extended identifiability to binarized observed variables, no principled method has been developed for scenarios involving discrete latent variables. In this paper, we show how multi-domain information can be leveraged to achieve identifiability when both latent and observed variables are discrete. We propose general identification conditions that do not depend on specific data distributional assumptions or parametric model forms. The effectiveness of our approach is validated through experiments on both simulated and real-world datasets.", "keywords": [ "Latent Variable Identification", "Nonlinear Independent Component Analysis (ICA)" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 1, 1 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "RjwWClPZtV", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Haoze Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Jiayue Liu", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Kaiwen Zhou", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Long Peng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Renjing Pei", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Engineer" }, { "name": "Wenbo Li", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yanwei Li", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Yong Guo", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Postdoc" }, { "name": "Yongqiang Chen", "gender": "unknown", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Yujiu Yang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 34, "n_ref": 90, "n_ref_all": 117, "n_fig": 13, "n_tab": 7, "L_tab": 1689, "n_element_tab": 226, "n_fig_1": 7, "n_tab_1": 7, "L_tab_1": 1858, "n_element_tab_1": 182, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 152140, "len_all_1": 59035, "len_abs": 2061, "len_title": 134, "len_sents": 43464, "len_sents_1": 26277, "n_sents": 358, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1139, "title": "Text Boosts Generalization: A Plug-and-Play Captioner for Real-World Image Restoration", "abs": "Generalization has long been a central challenge in real-world image restoration. While recent diffusion-based restoration methods, which leverage generative priors from text-to-image models, have made progress in recovering more realistic details, they still encounter \"generative capability inactivation\" when applied to out-of-distribution data. To address this, we propose using text as an auxiliary invariant representation to reactivate the generative capabilities of these models. We begin by identifying two key properties of text input in diffusion-based restoration: richness and relevance, and examine their respective influence on model performance. Building on these insights, we introduce Res-Captioner, a module that generates enhanced textual descriptions tailored to image content and degradation levels, effectively mitigating response failures. Additionally, we present RealIR, a new benchmark designed to capture diverse real-world scenarios. Extensive experiments demonstrate that Res-Captioner significantly boosts the generalization ability of diffusion-based restoration models, while remaining fully plug-and-play.", "keywords": [ "Image Restoration", "Generative models", "Diffusion Model", "Multimodel Large Language Model" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 3, 4, 2, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 2, 5, 4 ] }, { "paper_id": "RiS2cxpENN", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Markus Heinonen", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Researcher" }, { "name": "Rafal Karczewski", "gender": "unknown", "institution": "Aalto University", "country": "FI", "position": "PhD student" }, { "name": "Vikas K Garg", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Assistant Professor" } ], "n_formula": 148, "n_formula_1": 29, "n_ref_uni": 37, "n_ref": 101, "n_ref_all": 153, "n_fig": 14, "n_tab": 5, "L_tab": 1889, "n_element_tab": 169, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18115, "formula_len_all_1": 2320, "len_all": 187263, "len_all_1": 53855, "len_abs": 749, "len_title": 121, "len_sents": 45379, "len_sents_1": 20226, "n_sents": 484, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 753, "title": "Diffusion Models as Cartoonists! The Curious Case of High Density Regions", "abs": "We investigate what kind of images lie in the high-density regions of diffusion models. We introduce a theoretical mode-tracking process capable of pinpointing the exact mode of the denoising distribution, and we propose a practical high-density sampler that consistently generates images of higher likelihood than usual samplers. Our empirical findings reveal the existence of significantly higher likelihood samples that typical samplers do not produce, often manifesting as cartoon-like drawings or blurry images depending on the noise level. Curiously, these patterns emerge in datasets devoid of such examples. We also present a novel approach to track sample likelihoods in diffusion SDEs, which remarkably incurs no additional computational cost.", "keywords": [ "diffusion models", "likelihood estimation" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 4, 4, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "RiQRUcjXBD", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Binbin Lin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Chen Shen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jieping Ye", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Liang Xie", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "Postdoc" }, { "name": "Lihui Gu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Liye Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Wenxiao Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaofei He", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" }, { "name": "Yi Dai", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Yunxiang Luo", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 3, "n_ref": 24, "n_ref_all": 71, "n_fig": 5, "n_tab": 15, "L_tab": 5773, "n_element_tab": 321, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 892, "n_element_tab_1": 47, "formula_len_all": 476, "formula_len_all_1": 455, "len_all": 158657, "len_all_1": 69383, "len_abs": 1823, "len_title": 99, "len_sents": 56697, "len_sents_1": 31705, "n_sents": 485, "n_sents_1": 264, "venue": "ICLR2025", "decision": "withdraw", "L_title": 51, "L_abs": 1735, "title": "SciPIP: An LLM-based Scientific Paper Idea Proposer", "abs": "The exponential growth of knowledge and the increasing complexity of interdisciplinary research pose significant challenges for researchers, including information overload and difficulties in exploring novel ideas. The advancements in large language models (LLMs), such as GPT-4, have shown great potential in enhancing idea proposals, but how to effectively utilize large models for reasonable idea proposal has not been thoroughly explored.\nThis paper proposes a scientific paper idea proposer (SciPIP). Based on a user-provided research background, SciPIP retrieves helpful papers from a literature database while leveraging the capabilities of LLMs to generate more novel and feasible ideas.\nTo this end, \\textbf{1)} we construct a literature retrieval database, extracting lots of papers' multi-dimension information for fast access. Then, a literature retrieval method based on semantics, entity, and citation co-occurrences is proposed to search relevant literature from multiple aspects based on the user-provided background. \\textbf{2)} After literature retrieval, we introduce dual-path idea proposal strategies, where one path infers solutions from the retrieved literature and the other path generates original ideas through model brainstorming. We then combine the two to achieve a good balance between feasibility and originality. Through extensive experiments on the natural language processing (NLP) field, we demonstrate that SciPIP can retrieve citations similar to those of existing top conference papers and generate many ideas consistent with them. Additionally, we evaluate the originality of other ideas generated by SciPIP using large language models, further validating the effectiveness of our proposed method", "keywords": [ "Large Language Models", "Idea Proposer" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 1, 1 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "RiDtvlNiqp", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Dominik Jeurissen", "gender": "Male", "institution": "Queen Mary, University of London", "country": "", "position": "PhD student" }, { "name": "Michelangelo Conserva", "gender": "unknown", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Paulo Rauber", "gender": "Male", "institution": "Queen Mary University London", "country": "", "position": "Lecturer" }, { "name": "Remo Sasso", "gender": "Male", "institution": "Queen Mary, University of London", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 23, "n_ref": 40, "n_ref_all": 54, "n_fig": 8, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 130943, "len_all_1": 56182, "len_abs": 2192, "len_title": 116, "len_sents": 41863, "len_sents_1": 28586, "n_sents": 293, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1924, "title": "Foundation Models for Enhanced Exploration in Reinforcement Learning", "abs": "Reinforcement learning agents often struggle with sample inefficiency, requiring extensive interactions with the environment to develop effective policies. \nThis inefficiency is partly due to the challenge of balancing exploration and exploitation without the abstract reasoning and prior knowledge that humans use to quickly identify rewarding actions. \nRecent advancements in foundation models, such as large language models (LLMs) and vision-language models (VLMs), have shown human-level reasoning capabilities in some domains but have been underutilized in directly selecting low-level actions for exploration in reinforcement learning. \nIn this paper, we investigate the potential of foundation models to enhance exploration in reinforcement learning tasks. \nWe conduct an in-depth analysis of their exploration behaviour in multi-armed bandit problems and Gridworld environments, comparing their performance against traditional exploration strategies and reinforcement learning agents. \nOur empirical results suggest foundation models can significantly improve exploration efficiency by leveraging their reasoning abilities to infer optimal actions.\nBuilding on these findings, we introduce Foundation Model Exploration (FME), a novel exploration scheme that integrates foundation models into the reinforcement learning framework for intelligent exploration behaviour. \nWe use VLMs and demonstrate that they can infer environment dynamics and objectives from raw image observations. \nThis means FME only requires the action space as environment-specific manual text input.\nWe find that agents equipped with FME achieve superior performance in sparse reward Gridworld environments and scale to more complex tasks like Atari games. \nMoreover, the effectiveness of FME increases with the capacity of the VLM used, indicating that future advancements in foundation models will further enhance such exploration strategies.", "keywords": [ "reinforcement learning", "exploration", "large language models", "vision-language models", "multi-armed bandits" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "RhkI1cba7n", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Minghong Cai", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Wei-Hong Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Xiangyu Yue", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Yiyuan Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yilei JIANG", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 27, "n_ref": 59, "n_ref_all": 79, "n_fig": 9, "n_tab": 5, "L_tab": 704, "n_element_tab": 95, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 640, "n_element_tab_1": 165, "formula_len_all": 733, "formula_len_all_1": 678, "len_all": 99954, "len_all_1": 62030, "len_abs": 1606, "len_title": 152, "len_sents": 32416, "len_sents_1": 27062, "n_sents": 248, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1611, "title": "Debiasing Text-to-image Diffusion Models with Self-discovering Latent Directions", "abs": "While Diffusion Models (DM) exhibit remarkable performance across various image generative tasks, they nonetheless reflect the inherent bias presented in the training set. As DMs are now widely used in real-world applications, these biases could perpetuate a distorted worldview and hinder opportunities for minority groups. Existing methods on debiasing DMs usually requires model re-training with a human-crafted reference dataset or additional classifiers, which suffer from two major limitations: (1) collecting reference datasets causes expensive annotation cost; (2) the debiasing performance is heavily constrained by the quality of the reference dataset or the additional classifier. To address the above limitations, we propose DebiasDiff, a plug-and-play method that learns attribute latent directions in a self-discovering manner, thus eliminating the reliance on such reference dataset. Specifically, DebiasDiff consists of two parts: a set of attribute adapters and a distribution indicator. Each adapter in the set aims to learn an attribute latent direction, and is optimized via noise composition through a self-discovering process.Then, the distribution indicator is multiplied by the set of adapters to guide the generation process towards the prescribed distribution. Our method enables debiasing multiple attributes in DMs simultaneously, while remaining lightweight and easily integrable with other DMs, eliminating the need for re-training. Extensive experiments on debiasing gender, racial, and their intersectional biases show that our method outperforms previous SOTA by a large margin.", "keywords": [ "Diffusion Models", "AI fairness" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "RhfYIJux9d", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Fangchen Liu", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Oleh Rybkin", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Pieter Abbeel", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Weirui Ye", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yang Gao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zheng Ding", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 99, "n_ref_all": 114, "n_fig": 11, "n_tab": 6, "L_tab": 4315, "n_element_tab": 247, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 967, "n_element_tab_1": 83, "formula_len_all": 120, "formula_len_all_1": 108, "len_all": 175975, "len_all_1": 66713, "len_abs": 1515, "len_title": 129, "len_sents": 49003, "len_sents_1": 32330, "n_sents": 391, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1522, "title": "Video2Policy: Scaling up Manipulation Tasks in Simulation through Internet Videos", "abs": "Simulation offers a promising approach for cheaply scaling training data for generalist policies. To scalably generate data from diverse and realistic tasks, existing algorithms either rely on large language models (LLMs) that may hallucinate tasks not interesting for robotics; or digital twins, which require careful real-to-sim alignment and are hard to scale. To address these challenges, we introduce Video2Policy, a novel framework that leverages large amounts of internet RGB videos to reconstruct tasks based on everyday human behavior. Our approach comprises two phases: (1) task generation through object mesh reconstruction and 6D position tracking; and (2) reinforcement learning utilizing LLM-generated reward functions and iterative in-context reward reflection for the task. We demonstrate the efficacy of Video2Policy by reconstructing over 100 videos from the Something-Something-v2 (SSv2) dataset, which depicts diverse and complex human behaviors on 9 different tasks. Our method can successfully train RL policies on such tasks, including complex and challenging tasks such as throwing. Furthermore, we show that a generalist policy trained on the collected sim data generalizes effectively to new tasks and outperforms prior approaches. Finally, we show the performance of our policies improves by simply including more internet videos. We believe that the proposed Video2Policy framework is a step towards generalist policies that can execute practical robotic tasks based on everyday human behavior.", "keywords": [ "robotics manipulation", "internet videos", "real2sim", "Foundation models", "reinforcement learning" ], "rating_list": [ 8, 5, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 4, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "RhAW7TRJUy", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Druv Pai", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department, University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Peng Wang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Postdoc" }, { "name": "Qing Qu", "gender": "Male", "institution": "University of Michigan", "country": "US", "position": "Assistant Professor" }, { "name": "Yaodong Yu", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department, University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Yi Ma", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Yifu Lu", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "Undergrad student" } ], "n_formula": 72, "n_formula_1": 17, "n_ref_uni": 55, "n_ref": 110, "n_ref_all": 154, "n_fig": 7, "n_tab": 4, "L_tab": 872, "n_element_tab": 108, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 331, "n_element_tab_1": 52, "formula_len_all": 6052, "formula_len_all_1": 1368, "len_all": 177647, "len_all_1": 71219, "len_abs": 1374, "len_title": 107, "len_sents": 42893, "len_sents_1": 32150, "n_sents": 379, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1384, "title": "Attention-Only Transformers via Unrolled Subspace Denoising", "abs": "Despite the great success of transformers in practice, their architectures have been empirically designed, hence lack of mathematical justification and interpretability. Moreover, many empirical studies have indicated that some components of the transformer architectures may be redundant and can be removed or replaced without compromising overall performance. Hence to derive a compact and interpretable transformer architecture, we contend that the goal of representation learning is to compress a set of noisy initial token representations towards a mixture of low-dimensional subspaces. Based on the existing literature, the associated denoising operation naturally takes the form of a multi-subspace self-attention (MSSA). By unrolling such iterative denoising operations as a deep network, we arrive at a highly compact architecture that consists of only an MSSA operator with skip connections at each layer, without MLP. We rigorously prove that each layer of the proposed transformer performs so highly efficient denoising that it improves the signal-to-noise ratio of token representations {\\em at a linear rate} with respect to the number of layers. Despite its simplicity, extensive experiments on language and vision tasks demonstrate that such a minimalistic attention-only transformer can achieve performance close to conventional transformers, such as GPT-2 and CRATE.", "keywords": [ "transformers", "subspace denoising", "algorithm unrolling", "mixture of low-rank Gaussians" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 2, 5, 3, 4 ] }, { "paper_id": "RgWATMmWmz", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Masashi Sugiyama", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Director" }, { "name": "Wei Wang", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Researcher" }, { "name": "Li Ming", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "MS student" } ], "n_formula": 19, "n_formula_1": 10, "n_ref_uni": 74, "n_ref": 166, "n_ref_all": 188, "n_fig": 5, "n_tab": 11, "L_tab": 4683, "n_element_tab": 487, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2011, "n_element_tab_1": 182, "formula_len_all": 1285, "formula_len_all_1": 797, "len_all": 214941, "len_all_1": 66786, "len_abs": 1430, "len_title": 111, "len_sents": 46123, "len_sents_1": 28678, "n_sents": 355, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1436, "title": "Delving into Weakly Supervised Learning with Pre-Trained Models", "abs": "Weakly supervised learning (WSL) is a popular machine learning paradigm in recent years that aims to learn a classifier with incomplete, imprecise, or inaccurate supervision. Existing WSL approaches have mainly focused on designing different loss functions or training strategies and then training models from scratch. In this paper, we first empirically show that a zero-shot baseline based on the Contrastive Language-Image Pre-Training (CLIP) model with class descriptions empowered by GPT-4o can outperform previous state-of-the-art methods trained from scratch on various WSL problems. Therefore, this motivates us to fine-tune pre-trained models to further improve the performance. However, our additional experiments show that naive use of existing WSL losses degrades performance due to severe overfitting exacerbation and feature degeneration problems. To address these problems, we propose a novel weakly supervised fine-tuning approach using dual classification heads that are trained synergistically by alternately distilling reliable supervision and performing efficient model fine-tuning. Theoretically, we prove the consistency and convergence rate of the proposed risk estimator. Empirically, extensive experiments on benchmark datasets of different WSL problems validate the effectiveness of the proposed approach against state-of-the-art competitors. The code is provided at https://github.com/ICLR2025-6897/WSFT_code.", "keywords": [ "Weakly supervised learning", "positive-unlabeled learning", "unlabeled-unlabeled learning", "pre-trained models." ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 3, 4, 3, 4 ], "presentation_list": [ 2, 4, 2, 1 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "RgE1qiO2ek", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Can Chen", "gender": "Not Specified", "institution": "Mila - Quebec AI Institute", "country": "CA", "position": "PhD student" }, { "name": "Guoqing Liu", "gender": "Male", "institution": "Microsoft Research ", "country": "CN", "position": "Researcher" }, { "name": "Hao Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Xiuyuan Hu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xue Liu", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Full Professor" }, { "name": "Yang Zhao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 71, "n_ref": 140, "n_ref_all": 163, "n_fig": 3, "n_tab": 9, "L_tab": 3196, "n_element_tab": 318, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 322, "n_element_tab_1": 56, "formula_len_all": 494, "formula_len_all_1": 330, "len_all": 193631, "len_all_1": 60338, "len_abs": 1336, "len_title": 120, "len_sents": 45307, "len_sents_1": 30651, "n_sents": 240, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1407, "title": "3DMolFormer: A Dual-channel Framework for Structure-based Drug Discovery", "abs": "Structure-based drug discovery, encompassing the tasks of protein-ligand docking and pocket-aware 3D drug design, represents a core challenge in drug discovery. However, no existing work can deal with both tasks to effectively leverage the duality between them, and current methods for each task are hindered by challenges in modeling 3D information and the limitations of available data. To address these issues, we propose 3DMolFormer, a unified dual-channel transformer-based framework applicable to both docking and 3D drug design tasks, which exploits their duality by utilizing docking functionalities within the drug design process. Specifically, we represent 3D pocket-ligand complexes using parallel sequences of discrete tokens and continuous numbers, and we design a corresponding dual-channel transformer model to handle this format, thereby overcoming the challenges of 3D information modeling. Additionally, we alleviate data limitations through large-scale pre-training on a mixed dataset, followed by supervised and reinforcement learning fine-tuning techniques respectively tailored for the two tasks. Experimental results demonstrate that 3DMolFormer outperforms previous approaches in both protein-ligand docking and pocket-aware 3D drug design, highlighting its promising application in structure-based drug discovery. The code is available at: https://github.com/HXYfighter/3DMolFormer.", "keywords": [ "Structure-based Drug Discovery", "Protein-ligand Docking", "3D Molecule Generation", "Transformer" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "Rg2JxBZZ0g", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "CHEN TONG", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Collin Sakal", "gender": "unknown", "institution": "City University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Wei Zhang", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xinyue Li", "gender": "Female", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 25, "n_ref": 29, "n_ref_all": 33, "n_fig": 2, "n_tab": 2, "L_tab": 738, "n_element_tab": 60, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 738, "n_element_tab_1": 60, "formula_len_all": 794, "formula_len_all_1": 794, "len_all": 85703, "len_all_1": 66370, "len_abs": 1348, "len_title": 123, "len_sents": 28361, "len_sents_1": 28362, "n_sents": 179, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1353, "title": "GeneMamba: Early Parkinson\u2019s Detection via Wearable Device and Genetic Data", "abs": "Parkinson's disease (PD) is a progressive neurodegenerative disorder affecting millions worldwide, with its prevalence expected to rise as the global population ages. Early diagnosis is crucial for effective management and improved quality of life for patients. However, current accelerometer-based studies focus more on detecting the symptoms of PD, while less research has been conducted on early detection of PD. This study presents a novel multi-modal deep learning model named GeneMamba for early PD diagnosis, using state space modelling approaches to effectively analyze sequences and combining accelerometer data from wearable devices with genetic variants data. Our model predicts early PD occurrence up to 7 years before clinical onset, outperforming existing methods. Furthermore, through knowledge transfer, we enable accurate PD prediction using only wearable device data, enhancing our model's real-world applicability. Additionally, our interpretation methods uncover both established and previously unidentified genes associated with PD, advancing our understanding of the disease's genetic architecture and potentially highlighting new therapeutic targets. Our approach not only advances early PD diagnosis but also offers insights into the disease's etiology, paving the way for improved risk assessment and personalized interventions.", "keywords": [ "Accelerometer", "Parkinson's Disease", "Genetic Variants", "Mamba", "Time-series data" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 2 ] }, { "paper_id": "RfrdbJVvVf", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Abhinav Shukla", "gender": "unknown", "institution": "Scaled Foundations", "country": "US", "position": "Researcher" }, { "name": "Aditya Kusupati", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Ashish Kapoor", "gender": "unknown", "institution": "Scaled Foundations ", "country": "", "position": "Principal Researcher" }, { "name": "Sai Vemprala", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Senior Researcher" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 78, "n_ref_all": 96, "n_fig": 7, "n_tab": 7, "L_tab": 3135, "n_element_tab": 543, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1109, "n_element_tab_1": 256, "formula_len_all": 618, "formula_len_all_1": 437, "len_all": 125878, "len_all_1": 55994, "len_abs": 1261, "len_title": 88, "len_sents": 30517, "len_sents_1": 26314, "n_sents": 210, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1275, "title": "MatMamba: A Matryoshka State Space Model", "abs": "State Space Models (SSMs) like Mamba2 are a promising alternative to Transformers, with faster theoretical training and inference times -- especially for long context lengths. Recent work on Matryoshka Representation Learning -- and its application to Transformer backbones in works like MatFormer -- showed how to introduce nested granularities of smaller submodels in one universal elastic model. In this work, we present MatMamba: a state space model which combines Matryoshka-style learning with Mamba2, by modifying the block to contain nested dimensions to enable joint training and adaptive inference. MatMamba allows for efficient and adaptive deployment across various model sizes. We train a single large MatMamba model and are able to get a number of smaller nested models for free -- while maintaining or improving upon the performance of a baseline smaller model trained from scratch. We train language and image models at a variety of parameter sizes from 35M to 1.4B. Our results on ImageNet and FineWeb show that MatMamba models scale comparably to Transformers, while having more efficient inference characteristics. This makes MatMamba a practically viable option for deploying large-scale models in an elastic way based on the available inference compute.", "keywords": [ "matryoshka", "mamba", "state space models", "adaptive inference", "representation learning", "neural network architecture", "efficient", "visual" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "RflvsSxM0u", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Abhinav Valada", "gender": "Male", "institution": "University of Freiburg", "country": "DE", "position": "Full Professor" }, { "name": "Andreas Look", "gender": "Male", "institution": "Bosch", "country": "DE", "position": "Researcher" }, { "name": "Aron Distelzweig", "gender": "Male", "institution": "University of Freiburg, Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Eitan Kosman", "gender": "unknown", "institution": "Bosch", "country": "", "position": "Researcher" }, { "name": "Faris Janjos", "gender": "unknown", "institution": "Robert Bosch GmbH, Bosch", "country": "DE", "position": "Researcher" }, { "name": "J\u00f6rg Wagner", "gender": "unknown", "institution": "Robert Bosch GmbH, Bosch", "country": "DE", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 4, "n_ref_uni": 60, "n_ref": 153, "n_ref_all": 167, "n_fig": 5, "n_tab": 9, "L_tab": 10687, "n_element_tab": 1949, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 324, "n_element_tab_1": 36, "formula_len_all": 1287, "formula_len_all_1": 351, "len_all": 197277, "len_all_1": 68414, "len_abs": 1157, "len_title": 130, "len_sents": 45223, "len_sents_1": 32582, "n_sents": 329, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1122, "title": "Entropy-Based Uncertainty Modeling for Trajectory Prediction in Autonomous Driving", "abs": "In autonomous driving, accurate motion prediction is essential for safe and efficient motion planning. To ensure safety, planners must rely on reliable uncertainties in the future behavior of surrounding agents, yet this aspect has received limited attention. This paper addresses the problem of uncertainty modeling in trajectory prediction. We adopt a holistic approach that focuses on uncertainty quantification, decomposition, and the influence of model composition. Our method is based on a theoretically-grounded information-theoretic approach to measure uncertainty, allowing us to decompose total uncertainty into its aleatoric and epistemic components. We conduct extensive experiments on the nuScenes dataset to assess how different model architectures and configurations affect uncertainty quantification and model robustness. Our analysis thoroughly explores the uncertainty quantification capabilities of several state-of-the-art prediction models, examining the relationship between uncertainty and prediction error in both in- and out-of-distribution scenarios, as well as robustness in out-of-distribution.", "keywords": [ "Autonomous Driving", "Trajectory Prediction", "Uncertainty Quantification" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "RfYD6v829Y", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Gongshen Liu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Haodong Zhao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Pengzhou Cheng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Ping Yi", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Tianjie Ju", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Wei Du", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Yidong Ding", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Zhuosheng Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zongru Wu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 10, "n_ref_uni": 35, "n_ref": 56, "n_ref_all": 87, "n_fig": 14, "n_tab": 10, "L_tab": 2562, "n_element_tab": 166, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 212, "n_element_tab_1": 14, "formula_len_all": 1581, "formula_len_all_1": 662, "len_all": 210321, "len_all_1": 69860, "len_abs": 2056, "len_title": 137, "len_sents": 61283, "len_sents_1": 31719, "n_sents": 541, "n_sents_1": 243, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1500, "title": "TrojanRAG: Retrieval-Augmented Generation Can Be Backdoor Driver in Large Language Models", "abs": "Large language models (LLMs) have raised concerns about potential security threats, despite performing significantly in language modeling. Backdoor attacks are one of the vulnerabilities of LLMs. However, their attack costs and robustness have faced criticism amidst the continuous evolution of LLMs. In this paper, we comprehensively expose the threats of backdoor attacks on LLMs by defining three standardized scenarios from the perspective of attackers, users, and jailbreaking LLMs, and we propose TrojanRAG based on those scenarios. TrojanRAG is a joint backdoor attack against the Retrieval-Augmented Generation, that can manipulate LLMs robustly. Specifically, we first build multiple purpose-driven backdoors between poisoned knowledge and triggers in the retrieval backdoor injection phase, where retrieval performs well for clean queries but always returns semantic-consistency poisoned content for poisoned queries. Second, we induce the target output on LLMs based on the retrieved poisoned knowledge in the inductive attack generation phase. The joint backdoors are orthogonally optimized by contrastive learning, ensuring that multiple backdoors are independent of each other within the parameter subspace. Meanwhile, we introduce a knowledge graph to construct structured metadata, improving retrieval performance at a fine-grained level. Extensive evaluations across 11 tasks in six LLMs highlight TrojanRAG\u2019s threats and transferability, particularly in Chain of Thought (CoT) mode.", "keywords": [ "Backdoor Attack", "Retrieval-Augmented Generation", "Large Language Models", "Universal Attack Scenarios" ], "rating_list": [ 3, 3, 5, 3, 3 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 1, 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 5, 4, 5 ] }, { "paper_id": "Rf4NnqHNSz", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Benedikt Arnthof", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "MS student" }, { "name": "Bernd Bischl", "gender": "Male", "institution": "LMU", "country": "DE", "position": "Full Professor" }, { "name": "Mina Rezaei", "gender": "Female", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Principal Researcher" }, { "name": "Simon Schosser", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "MS student" }, { "name": "Yawei Li", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 18, "n_ref_uni": 27, "n_ref": 71, "n_ref_all": 90, "n_fig": 5, "n_tab": 2, "L_tab": 836, "n_element_tab": 106, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2743, "formula_len_all_1": 1188, "len_all": 124747, "len_all_1": 55826, "len_abs": 1692, "len_title": 121, "len_sents": 41013, "len_sents_1": 27220, "n_sents": 296, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1698, "title": "Robust Probabilistic Unsupervised Segmentation with Uncertainty Modeling", "abs": "Unsupervised semantic segmentation aims to assign a semantic label to each pixel in an image, identifying the object or scene class without any supervision. However, the task becomes particularly difficult due to factors like unclear or overlapping boundaries, intricate object textures, and the presence of multiple objects within the same region. Traditional unsupervised models often suffer from class misalignment and poor spatial coherence, leading to fragmented and imprecise segmentation, often employing postprocessing with Conditional Random Fields (CRFs) to improve their results. Additionally, deterministic models lack the ability to capture prediction uncertainty, making their outputs particularly prone to errors in ambiguous regions. To address these issues, we propose a probabilistic unsupervised semantic segmentation framework that enhances the robustness and accuracy of segmentation by refining predictions through uncertainty modeling and spatial smoothing techniques. We also introduce a novel loss function that encourages the model to focus on learning similarities within pixels by leveraging feature information from pre-trained vision transformer backbones.\nWe also provide theoretical analyses of our proposed loss function, highlighting its favorable properties in relation to the optimization of our models. Our method demonstrates superior accuracy and calibration, outperforming various baselines across multiple unsupervised semantic segmentation benchmarks including COCO, Potsdam, and Cityscapes. In conclusion, our framework offers a foundation for more reliable, uncertainty-aware segmentation models, advancing research in unsupervised semantic segmentation.", "keywords": [ "Unsupervised Image Segmentation", "Probabilistic Unsupervised Segmentation" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Revyper1Mi", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Jonas Fischer", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Researcher" }, { "name": "Rong Ma", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 40, "n_formula_1": 11, "n_ref_uni": 63, "n_ref": 108, "n_ref_all": 140, "n_fig": 10, "n_tab": 5, "L_tab": 2925, "n_element_tab": 202, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3789, "formula_len_all_1": 1044, "len_all": 205146, "len_all_1": 68170, "len_abs": 1291, "len_title": 137, "len_sents": 64253, "len_sents_1": 34907, "n_sents": 435, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1309, "title": "Sailing in high-dimensional spaces: Low-dimensional embeddings through angle preservation", "abs": "Low-dimensional embeddings (LDEs) of high-dimensional data are ubiquitous in science and engineering. They allow us to quickly understand the main properties of the data, identify outliers and processing errors, and inform the next steps of data analysis.\n As such, LDEs have to be *faithful* to the original high-dimensional data, i.e., they should represent the relationships that are encoded in the data, both at a local as well as global scale.\n The current generation of LDE approaches focus on reconstructing *local distances* between any pair of samples correctly, often outperforming traditional approaches aiming at all distances.\n For these approaches, global relationships are, however, usually strongly distorted, often argued to be an inherent trade-off between local and global structure learning for embeddings. We suggest a new perspective on LDE learning, reconstructing *angles* between data points.\n We show that this approach, Mercat, yields good reconstruction across a diverse set of experiments and metrics, and preserve structures well across all scales, outperforming existing methods across datasets and metrics in most cases by a margin.\n Compared to existing work, our approach also has a *simple formulation*, facilitating future theoretical analysis and algorithmic improvements.", "keywords": [ "data visualization", "low-dimensional embeddings", "dimensionality reduction" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 4, 3 ] }, { "paper_id": "ReccFdn4zE", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Liam Bruce Smith", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Morris Cohen", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 24, "n_ref": 33, "n_ref_all": 45, "n_fig": 8, "n_tab": 1, "L_tab": 73, "n_element_tab": 10, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 73, "n_element_tab_1": 10, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 76369, "len_all_1": 58179, "len_abs": 1100, "len_title": 131, "len_sents": 22614, "len_sents_1": 21732, "n_sents": 185, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1107, "title": "Cross Attention for Oddly Shaped Data and Applications in Ionospheric Modeling", "abs": "It is desirable to have models of many physical phenomena, yet often data for these phenomena are oddly structured. These structures, such as ungridded and arbitrary length data prevents the use of many types of machine learning techniques, such as feed-forward neural networks. It is thus quite desirable to be able to move this data into a fixed size and shape for easier data ingest. We propose a method of using cross attention to do this.\nAn example of oddly shaped data is Total Electron Content (TEC), or the vertical integral of electron density in the atmosphere. TEC data is calculated using both the position of a satellite and a position on the surface of Earth, giving a non-fixed location per sample. This leads to a splattering of points on the globe where measurements exist that change in shape and amount each time step.\nWe apply our technique to TEC in an autoregressive approach. This allows us to both obtain an embedding describing the global TEC and create completed TEC maps, filling in where measurements are not taken. The global embedding can then be further used in other models.", "keywords": [ "Ionosphere", "Cross Attention", "Attention", "Transformers", "Electron Density", "Total Electron Content", "TEC", "Misshapen Data", "Variable Sequence Length" ], "rating_list": [ 1, 3, 3, 1 ], "soundness_list": [ 1, 3, 3, 1 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "ReKWjKvkJE", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Feiran Huang", "gender": "Male", "institution": "Jinan University", "country": "CN", "position": "Full Professor" }, { "name": "Hao Chen", "gender": "Male", "institution": "City University of Macau", "country": "MO", "position": "Assistant Professor" }, { "name": "Junnan Dong", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Qinggang Zhang", "gender": "Male", "institution": "The Hong Kong Polytechnic University, Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Xiao Huang", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "", "position": "Assistant Professor" }, { "name": "LI Wentao", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "", "position": "MS student" } ], "n_formula": 31, "n_formula_1": 18, "n_ref_uni": 44, "n_ref": 89, "n_ref_all": 113, "n_fig": 5, "n_tab": 12, "L_tab": 6569, "n_element_tab": 422, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 4695, "n_element_tab_1": 275, "formula_len_all": 1203, "formula_len_all_1": 748, "len_all": 198646, "len_all_1": 64867, "len_abs": 3344, "len_title": 113, "len_sents": 59113, "len_sents_1": 27513, "n_sents": 412, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1457, "title": "Structure-Guided Large Language Models for Text-to-SQL Generation", "abs": "Recent advancements in large language models (LLMs) have shown promise in bridging the gap between natural language queries and database management systems, enabling users to interact with databases without the background of SQL. However, LLMs often struggle to fully exploit and comprehend the user intention and complex structures of databases. Decomposition-based methods have been proposed to enhance the performance of LLMs on complex tasks, but decomposing SQL generation into subtasks is non-trivial due to the declarative structure of SQL syntax and the intricate connections between query concepts and database elements. In this paper, we propose a novel $\\textbf{S}$tructure $\\textbf{GU}$ided text-to-$\\textbf{SQL}$ framework ($\\textbf{SGU-SQL}$) that incorporates syntax-based prompting to enhance the SQL generation capabilities of LLMs. Specifically, SGU-SQL establishes structure-aware links between user queries and database schema and recursively decomposes the complex generation task using syntax-based prompting to guide LLMs in incrementally constructing target SQLs. Extensive experiments on two benchmark datasets demonstrate that SGU-SQL consistently outperforms state-of-the-art text-to-SQL baselines. These results highlight the importance of incorporating structural syntax information for effective text-to-SQL generation and pave the way for more robust and reliable interfaces to databases in the era of artificial intelligence.", "keywords": [ "Text-to-SQL", "large language model", "structure learning" ], "rating_list": [ 3, 8, 10, 5 ], "soundness_list": [ 2, 4, 4, 2 ], "presentation_list": [ 2, 4, 4, 2 ], "contribution_list": [ 2, 4, 4, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "ReItdfwMcg", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Arthur Gretton", "gender": "Male", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Dimitri Meunier", "gender": "Not Specified", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Juno Kim", "gender": "Male", "institution": "University of Tokyo", "country": "JP", "position": "MS student" }, { "name": "Taiji Suzuki", "gender": "Male", "institution": "The University of Tokyo", "country": "", "position": "Full Professor" }, { "name": "Zhu Li", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Lecturer" } ], "n_formula": 220, "n_formula_1": 46, "n_ref_uni": 60, "n_ref": 153, "n_ref_all": 190, "n_fig": 0, "n_tab": 5, "L_tab": 913, "n_element_tab": 40, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 132, "n_element_tab_1": 3, "formula_len_all": 30414, "formula_len_all_1": 4141, "len_all": 276103, "len_all_1": 72828, "len_abs": 1082, "len_title": 135, "len_sents": 76796, "len_sents_1": 30700, "n_sents": 706, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1092, "title": "Optimality and Adaptivity of Deep Neural Features for Instrumental Variable Regression", "abs": "We provide a convergence analysis of \\emph{deep feature instrumental variable} (DFIV) regression (Xu et al., 2021), a nonparametric approach to IV regression using data-adaptive features learned by deep neural networks in two stages. We prove that the DFIV algorithm achieves the minimax optimal learning rate when the target structural function lies in a Besov space. This is shown under standard nonparametric IV assumptions, and an additional smoothness assumption on the regularity of the conditional distribution of the covariate given the instrument, which controls the difficulty of Stage 1. We further demonstrate that DFIV, as a data-adaptive algorithm, is superior to fixed-feature (kernel or sieve) IV methods in two ways. First, when the target function possesses low spatial homogeneity (i.e., it has both smooth and spiky/discontinuous regions), DFIV still achieves the optimal rate, while fixed-feature methods are shown to be strictly suboptimal. Second, comparing with kernel-based two-stage regression estimators, DFIV is provably more data efficient in the Stage 1 samples.", "keywords": [ "instrumental variable regression", "DFIV", "deep neural networks", "minimax optimality" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 2 ] }, { "paper_id": "Re5iu0hBTs", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Chang Yang", "gender": "Female", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Hau Chan", "gender": "Male", "institution": "University of Nebraska, Lincoln", "country": "", "position": "Assistant Professor" }, { "name": "Jakub Cerny", "gender": "unknown", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" }, { "name": "Pengdeng Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Shuxin Li", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xinrun Wang", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Youzhi Zhang", "gender": "unknown", "institution": "Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 6, "n_ref_uni": 67, "n_ref": 116, "n_ref_all": 140, "n_fig": 18, "n_tab": 3, "L_tab": 444, "n_element_tab": 7, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1417, "formula_len_all_1": 622, "len_all": 264229, "len_all_1": 77780, "len_abs": 1844, "len_title": 132, "len_sents": 83782, "len_sents_1": 35225, "n_sents": 626, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1861, "title": "Offline Equilibrium Finding in Extensive-form Games: Datasets, Methods, and Analysis", "abs": "Offline reinforcement learning (Offline RL) brings new methods to tackle real-world decision-making problems by leveraging pre-collected datasets. \nDespite substantial progress in single-agent scenarios, the application of offline learning to multiplayer games remains largely unexplored. \nTherefore, we introduce a novel paradigm ***offline equilibrium finding*** (Offline EF) in extensive-form games (EFGs), which aims at computing equilibrium strategies from offline datasets. \nThe primary challenges of offline EF include i) the absence of a comprehensive dataset of EFGs for evaluation; ii) the inherent difficulties in computing an equilibrium strategy solely from an offline dataset, as equilibrium finding requires referencing all potential action profiles; and iii) the impact of dataset quality and completeness on the effectiveness of the derived strategies.\nTo overcome these challenges, we make four main contributions in this work. First, we construct diverse datasets, encompassing a wide range of games, which form the foundation for the offline EF paradigm and serve as a basis for evaluating the performance of offline EF algorithms. Second, we design a novel framework, BOMB, which integrates the behavior cloning technique within a model-based method. BOMB can seamlessly integrate online equilibrium finding algorithms to the offline setting with minimal modifications. Third, we provide a comprehensive theoretical and empirical analysis of our BOMB framework, offering performance guarantees across various offline datasets. Finaly, extensive experiments have been carried out across different games under different offline datasets, and the results not only demonstrate the superiority of our approach compared to traditional offline RL algorithms but also highlight the remarkable efficiency in computing equilibrium strategies offline.", "keywords": [ "Game Theory", "Equilibrium Finding", "Offline Learning" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 1, 2, 2, 4 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "Re4Z3Wt2DS", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Byoung-Tak Zhang", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "HEE BIN YOO", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Jaein Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Dong-Sig Han Han", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" } ], "n_formula": 158, "n_formula_1": 33, "n_ref_uni": 83, "n_ref": 181, "n_ref_all": 264, "n_fig": 14, "n_tab": 9, "L_tab": 2383, "n_element_tab": 256, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 892, "n_element_tab_1": 87, "formula_len_all": 14530, "formula_len_all_1": 3116, "len_all": 279619, "len_all_1": 78071, "len_abs": 1012, "len_title": 117, "len_sents": 79590, "len_sents_1": 29986, "n_sents": 704, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1048, "title": "Variational Mirror Descent for Robust Learning in Schr\u00f6dinger Bridge", "abs": "Schr\u00f6dinger bridge (SB) has evolved into a universal class of probabilistic generative models. Recent studies regarding the Sinkhorn algorithm through mirror descent (MD) have gained attention, revealing geometric insights into solution acquisition of the SB problems. In this paper, we propose a variational online MD framework for the SB problems, which provides further stability to SB solvers. We formally prove convergence and a regret bound $\\mathcal{O}(\\textrm{\\small$\\sqrt{T}$})$ of online mirror descent under mild assumptions. As a result of analysis, we propose a simulation-free SB algorithm called Variational Mirrored Schr\u00f6dinger Bridge (VMSB) by utilizing the Wasserstein-Fisher-Rao geometry of the Gaussian mixture parameterization for Schr\u00f6dinger potentials. Based on the Wasserstein gradient flow theory, our variational MD framework offers tractable gradient-based learning dynamics that precisely approximate a subsequent update. We demonstrate the performance of the proposed VMSB algorithm in an extensive suite of benchmarks.", "keywords": [ "optimal transport", "mirror descent", "variational methods" ], "rating_list": [ 5, 8, 5, 8, 8 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 1, 2, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3, 2 ] }, { "paper_id": "Rdb0HxGJa3", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Junhui Zhang", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Patrick Jaillet", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 96, "n_formula_1": 46, "n_ref_uni": 22, "n_ref": 90, "n_ref_all": 102, "n_fig": 5, "n_tab": 2, "L_tab": 21, "n_element_tab": 8, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15287, "formula_len_all_1": 4654, "len_all": 152195, "len_all_1": 77600, "len_abs": 1221, "len_title": 127, "len_sents": 39510, "len_sents_1": 28346, "n_sents": 407, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1262, "title": "Online Convex Optimization with Prediction Through Accelerated Gradient Descent", "abs": "We study online convex optimization with predictions, where, at each time step $t$, predictions about the next $k$ steps are available, and with coupled costs over time steps, where the cost function at time step $t$ depends on the decisions made between time $t-a$ and time $t+b$ for some nonnegative integers $a,b$. \n\nWe provide a general recipe to run synchronous update in an asynchronous fashion that respects the sequential revelation of information. Combined with existing convergence results for convex optimization using inexact first-order oracle, we show that acceleration is possible in this framework, where the dynamic regret can be reduced by a factor of $(1-O(\\sqrt{\\kappa}))^{\\frac{k}{a+b}}$ through accelerated gradient descent, at a cost of an additive error term that depends on the prediction accuracy. This generalizes and improves the $(1-\\kappa/4)^k$ factor obtained by Li & Li (2020) for $a+b = 1$. Our algorithm also has smaller dependency on longer-term prediction error. Moreover, our algorithm is the first gradient based algorithm which, when the strong-convexity assumption is relaxed, constructs a solution whose regret decays at the rate of $O(1/k^2)$, at a cost of an additive error term that depends on the prediction accuracy.", "keywords": [ "online optimization", "accelerated gradient descent" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "RdTYx4jd7C", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Michael Cochez", "gender": "Male", "institution": "VU Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Ruud van Bakel", "gender": "Male", "institution": "Vrije Universiteit Amsterdam", "country": "", "position": "PhD student" }, { "name": "Tom Pelletreau-Duris", "gender": "Male", "institution": "Vrije Universiteit Amsterdam", "country": "NL", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 6, "n_ref_uni": 71, "n_ref": 105, "n_ref_all": 128, "n_fig": 10, "n_tab": 18, "L_tab": 7445, "n_element_tab": 1066, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 125, "n_element_tab_1": 28, "formula_len_all": 490, "formula_len_all_1": 191, "len_all": 236638, "len_all_1": 67482, "len_abs": 947, "len_title": 104, "len_sents": 63964, "len_sents_1": 33722, "n_sents": 457, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 950, "title": "Do graph neural network states contain graph properties?", "abs": "Deep neural networks (DNNs) achieve state-of-the-art performance on many tasks, but this often requires increasingly larger model sizes, which in turn leads to more complex internal representations. Explainability techniques (XAI) have made remarkable progress in the interpretability of ML models. However, the non-relational nature of Graph neural networks (GNNs) make it difficult to reuse already existing XAI methods. While other works have focused on instance-based explanation methods for GNNs, very few have investigated model-based methods and, to our knowledge, none have tried to probe the embedding of the GNNs for well-known structural graph properties. In this paper we present a model agnostic explainability pipeline for GNNs employing diagnostic classifiers. This pipeline aims to probe and interpret the learned representations in GNNs across various architectures and datasets, refining our understanding and trust in these models.", "keywords": [ "Graph Neural Networks", "explainability", "interpretability", "Diagnostic classifiers" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "RdGvvqjkC1", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Shujun Liu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Siyuan Wang", "gender": "Female", "institution": "University of Southern California", "country": "", "position": "Postdoc" }, { "name": "Xuanjing Huang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yuhang Lai", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Zhuohan Long", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "zhongyu wei", "gender": "Male", "institution": "Fudan University", "country": "", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 26, "n_ref": 45, "n_ref_all": 69, "n_fig": 7, "n_tab": 10, "L_tab": 8257, "n_element_tab": 1512, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3657, "n_element_tab_1": 538, "formula_len_all": 315, "formula_len_all_1": 315, "len_all": 174385, "len_all_1": 66987, "len_abs": 1455, "len_title": 117, "len_sents": 43371, "len_sents_1": 28972, "n_sents": 303, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1391, "title": "How Jailbreak Defenses Work and Ensemble? A Mechanistic Investigation", "abs": "Jailbreak attacks, where malicious prompts bypass generative models\u2019 built-in safety, have raised significant concerns about model vulnerability. While diverse defense methods have been proposed, the underlying mechanisms governing the trade-offs between model safety and helpfulness, and their application to Large Vision-Language Models (LVLMs) remain insufficiently explored. This paper systematically investigates jailbreak defense mechanisms by reformulating the standard generation task as a binary classification problem to probe model refusal tendencies across both harmful and benign queries. Our analysis identifies two key defense mechanisms: safety shift, which generally increases refusal probabilities for all queries, and harmfulness discrimination, which enhances the model\u2019s ability to distinguish between benign and harmful queries. Leveraging these mechanisms, we design two ensemble defense strategies\u2014inter-mechanism and intra-mechanism ensembles\u2014to explore the safety-helpfulness balance. Empirical evaluations on the MM-SafetyBench and MOSSBench datasets on top of LLaVA-1.5 models demonstrate the effectiveness of these ensemble approaches in either enhancing model safety or achieving an improved safety-utility balance. These findings offer valuable insights into jailbreak defense strategies and contribute to the development of more resilient LVLM safety systems.", "keywords": [ "Jailbreak", "defense", "analysis", "LVLM" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "RdG7LVGnQi", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Chen Tang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Heng Chang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Jia Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Tong Zhang", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yuan Meng", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuhui Xu", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Research Scientist" }, { "name": "Kerry Yi", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 25, "n_ref": 52, "n_ref_all": 74, "n_fig": 10, "n_tab": 3, "L_tab": 2829, "n_element_tab": 398, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 512, "n_element_tab_1": 34, "formula_len_all": 471, "formula_len_all_1": 453, "len_all": 103527, "len_all_1": 51118, "len_abs": 1457, "len_title": 79, "len_sents": 31405, "len_sents_1": 21303, "n_sents": 240, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1464, "title": "One QuantLLM for ALL: Fine-tuning Quantized LLMs Once for Efficient Deployments", "abs": "Large Language Models (LLMs) have advanced rapidly but face significant memory demands. While quantization has shown promise for LLMs, current methods typically require lengthy training to alleviate the performance degradation from quantization loss. However, deploying LLMs across diverse scenarios with different resource constraints, e.g., servers and personal computers, requires repeated training per application, which amplifies the lengthy training problem. Given that, it is advantageous to train a once-for-all (OFA) supernet capable of yielding diverse optimal subnets for downstream applications through one-shot training. Nonetheless, the scale of current language models impedes efficiency and amplifies interference from weight sharing between subnets. We make an initial attempt to extend the once-for-all framework to large language models. Specifically, we decouple shared weights to eliminate the interference and incorporate Low-Rank adapters for training efficiency.\nFurthermore, we observe the imbalance allocation of training resources from the traditional uniform sampling. A non-parametric scheduler is introduced to adjust the sampling rate for each quantization configuration, achieving a more balanced allocation among subnets with varying demands. We validate the approach on LLaMA2 families and Mistral on downstream evaluation, demonstrating high performance while significantly reducing deployment time faced with multiple scenarios.", "keywords": [ "Large Language Models", "Quantization", "LoRA", "Instruct Fine-tuning", "One-Shot Training" ], "rating_list": [ 5, 6, 6, 1 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "RdFpj6z4nE", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Haiquan Qiu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Shuzhi Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Intern" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "quanming yao", "gender": "Male", "institution": "Department of Electronic Engineering, Tsinghua University", "country": "", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 13, "n_ref_uni": 46, "n_ref": 98, "n_ref_all": 130, "n_fig": 6, "n_tab": 9, "L_tab": 2970, "n_element_tab": 299, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2029, "n_element_tab_1": 143, "formula_len_all": 1560, "formula_len_all_1": 1177, "len_all": 151912, "len_all_1": 68097, "len_abs": 1574, "len_title": 103, "len_sents": 47030, "len_sents_1": 30591, "n_sents": 377, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1280, "title": "Neural Symbolic Regression of Complex Network Dynamics", "abs": "Complex networks describe important structures in nature and society, composed of nodes and the edges that connect them. The evolution of these networks is typically described by dynamics, which are labor-intensive and require expert knowledge to derive. However, because the complex network involves noisy observations from multiple trajectories of nodes, existing symbolic regression methods are either not applicable or ineffective on its dynamics. In this paper, we propose Physically Inspired Neural Dynamics Symbolic Regression (PI-NDSR), a method based on neural networks and genetic programming to automatically learn the symbolic expression of dynamics. Our method consists of two key components: a Physically Inspired Neural Dynamics (PIND) to augment and denoise trajectories through observed trajectory interpolation; and a coordinated genetic search algorithm to derive symbolic expressions. This algorithm leverages references of node dynamics and edge dynamics from neural dynamics to avoid overfitted expressions in symbolic space. We evaluate our method on synthetic datasets generated by various dynamics and real datasets on disease spreading. The results demonstrate that PI-NDSR outperforms the existing method in terms of both recovery probability and error.", "keywords": [ "network dynamics", "symbolic regression", "complex network" ], "rating_list": [ 5, 5, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3, 2 ] }, { "paper_id": "RcmJ9NpqeA", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Alexander Shevtsov", "gender": "unknown", "institution": "University of Crete, University of Crete", "country": "", "position": "PhD student" }, { "name": "Ioannis Arapakis", "gender": "Male", "institution": "Telef\u00f3nica Innovaci\u00f3n Digital", "country": "ES", "position": "Principal Researcher" }, { "name": "Ioannis Lamprou", "gender": "Male", "institution": "Technical University of Crete", "country": "GR", "position": "MS student" }, { "name": "Sotiris Ioannidis", "gender": "unknown", "institution": "Technical University of Crete", "country": "", "position": "Associate Professor" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 21, "n_ref": 34, "n_ref_all": 41, "n_fig": 1, "n_tab": 12, "L_tab": 6358, "n_element_tab": 898, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 1335, "n_element_tab_1": 148, "formula_len_all": 609, "formula_len_all_1": 444, "len_all": 151412, "len_all_1": 62716, "len_abs": 1268, "len_title": 111, "len_sents": 55121, "len_sents_1": 30947, "n_sents": 355, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1273, "title": "White-Basilisk: A Hybrid Model for Code Vulnerability Detection", "abs": "The proliferation of software vulnerabilities presents a significant challenge to cybersecurity, necessitating more effective detection methodologies. We introduce White-Basilisk, a novel approach to vulnerability detection that demonstrates superior performance while challenging prevailing assumptions in AI model scaling. Utilizing an innovative architecture that integrates Mamba layers, linear self-attention, and a Mixture of Experts framework, White-Basilisk achieves state-of-the-art results in vulnerability detection tasks with a parameter count of only 200M. The model's capacity to process sequences of unprecedented length enables comprehensive analysis of extensive codebases in a single pass, surpassing the context limitations of current Large Language Models (LLMs). White-Basilisk exhibits robust performance on imbalanced, real-world datasets, while maintaining computational efficiency that facilitates deployment across diverse organizational scales. This research not only establishes new benchmarks in code security but also provides empirical evidence that compact, efficiently designed models can outperform larger counterparts in specialized tasks, potentially redefining optimization strategies in AI development for domain-specific applications.", "keywords": [ "vulnerability detection", "llm", "linear attention", "mamba", "white-basilisk", "nlp", "moe", "efficient", "source code", "c", "c++" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "RcNzwKrjTo", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Ahmed Alaa", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Assistant Professor" }, { "name": "Jivat Neet Kaur", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Michael Jordan", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" } ], "n_formula": 43, "n_formula_1": 29, "n_ref_uni": 25, "n_ref": 76, "n_ref_all": 99, "n_fig": 9, "n_tab": 7, "L_tab": 6358, "n_element_tab": 612, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1272, "n_element_tab_1": 113, "formula_len_all": 3109, "formula_len_all_1": 2170, "len_all": 152978, "len_all_1": 75064, "len_abs": 1233, "len_title": 127, "len_sents": 45587, "len_sents_1": 31421, "n_sents": 353, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1238, "title": "Conformal Prediction Sets with Improved Conditional Coverage using Trust Scores", "abs": "Standard conformal prediction offers a marginal guarantee on coverage, but for prediction sets to be truly useful, they should ideally ensure coverage conditional on each test point. However, it is impossible to achieve exact, distribution-free conditional coverage in finite samples. In this work, we propose an alternative conformal prediction algorithm that targets coverage where it matters most---in instances where a classifier is overconfident in its incorrect predictions. We start by dissecting miscoverage events in marginally-valid conformal prediction, and show that miscoverage rates vary based on the classifier's confidence and its deviation from the Bayes optimal classifier. Motivated by this insight, we develop a variant of conformal prediction that targets coverage conditional on a reduced set of two variables: the classifier's confidence in a prediction and a nonparametric trust score that measures its deviation from the Bayes classifier. Empirical evaluation on multiple image datasets shows that our method generally improves conditional coverage properties compared to standard conformal prediction, including class-conditional coverage, coverage over arbitrary subgroups, and coverage over demographic groups.", "keywords": [ "uncertainty quantification", "conformal prediction", "conditional guarantees" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Rc8z5wLzBF", "primary_area": "datasets and benchmarks", "n_author": 20, "author_info_list": [ { "name": "Chenghua Lin", "gender": "Male", "institution": "University of Manchester ", "country": "GB", "position": "Full Professor" }, { "name": "Emmanouil Benetos", "gender": "Male", "institution": "Queen Mary, University of London", "country": "GB", "position": "Associate Professor" }, { "name": "Ge Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Hangyu Guo", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Liu Jiaheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jian Yang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jinjie Shi", "gender": "unknown", "institution": "Queen Mary, University of London", "country": "GB", "position": "PhD student" }, { "name": "Ruibin Yuan", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Ruibo Liu", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Siwei Wu", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Wenhao Huang", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Xingwei Qu", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "PhD student" }, { "name": "Xinyue Zhang", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Yiming Liang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Yinghao Ma", "gender": "Male", "institution": "Queen Mary University of London", "country": "GB", "position": "PhD student" }, { "name": "Yizhi LI", "gender": "Male", "institution": "University of Manchester ", "country": "GB", "position": "PhD student" }, { "name": "Zhaoxiang Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhenzhu Yang", "gender": "Female", "institution": "China University of Geoscience Beijing", "country": "CN", "position": "MS student" }, { "name": "Kang Zhu", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Zekun Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 71, "n_ref_all": 92, "n_fig": 6, "n_tab": 7, "L_tab": 3881, "n_element_tab": 549, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 6699, "n_element_tab_1": 580, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 165908, "len_all_1": 76223, "len_abs": 2049, "len_title": 111, "len_sents": 38242, "len_sents_1": 29692, "n_sents": 258, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1702, "title": "OmniBench: Towards The Future of Universal Omni-Language Models", "abs": "Recent advancements in multimodal large language models (MLLMs) have aimed to integrate and interpret data across diverse modalities. However, the capacity of these models to concurrently process and reason about multiple modalities remains inadequately explored, partly due to the lack of comprehensive modality-wise benchmarks. \nWe introduce **OmniBench**, a novel benchmark designed to rigorously evaluate models' ability to recognize, interpret, and reason across **visual**, **acoustic**, and **textual** inputs simultaneously. We define models capable of such tri-modal processing as omni-language models (OLMs).\nOmniBench is distinguished by high-quality human annotations, ensuring that accurate responses require integrated understanding and reasoning across all three modalities. Our main findings reveal that: *i)* open-source OLMs exhibit critical limitations in instruction-following and reasoning capabilities within tri-modal contexts; and *ii)* most baselines models perform poorly (below 50\\% accuracy) even when provided with alternative textual representations of images or/and audio.\nThese results suggest that the ability to construct a consistent context from text, image, and audio is often overlooked in existing MLLM training paradigms. \nTo address this gap, we curate an instruction tuning dataset of 84.5K training samples, **OmniInstruct**, for training OLMs to adapt to multimodal contexts.\nWe advocate for future research to focus on developing more robust tri-modal integration techniques and training strategies to enhance OLM performance across diverse modalities.\nCodes and datasets are uploaded at [our repository](https://anonymous.4open.science/r/Omni-Bench-EA9B).", "keywords": [ "Multimodal Reasoning", "MLLM Benchmark", "Text", "Audio", "Image" ], "rating_list": [ 6, 3, 8, 6 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "RbcXV63ZJk", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Anton Tsitsulin", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Bahare Fatemi", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Bryan Perozzi", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "David van Dijk", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" }, { "name": "Dustin Zelle", "gender": "Male", "institution": "Research, Google", "country": "", "position": "Researcher" }, { "name": "Hasan Awais", "gender": "unknown", "institution": "Google", "country": "US", "position": "Software Engineer" }, { "name": "Jonathan Halcrow", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Rami Al-Rfou'", "gender": "Male", "institution": "Waymo", "country": "", "position": "Researcher" }, { "name": "Sami Abu-El-Haija", "gender": "Male", "institution": "Research, Google", "country": "", "position": "Research Scientist" }, { "name": "Syed A Rizvi", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Seyed Mehran Kazemi", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 68, "n_ref": 118, "n_ref_all": 145, "n_fig": 7, "n_tab": 19, "L_tab": 2912, "n_element_tab": 274, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3235, "n_element_tab_1": 386, "formula_len_all": 186, "formula_len_all_1": 1, "len_all": 244659, "len_all_1": 68333, "len_abs": 1014, "len_title": 108, "len_sents": 54320, "len_sents_1": 28867, "n_sents": 456, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1013, "title": "Parameter Efficient Graph Encoding for Large Language Models", "abs": "How can we best encode structured data into sequential form for use in large language models (LLMs)? In this work, we introduce a parameter-efficient method to explicitly represent structured data for LLMs. Our method, GraphToken, learns an encoding function to extend prompts with explicit structured information. The encoding function in GraphToken uses graph neural networks to effectively transfer the relational inductive biases in the structured data to a LLM. Unlike other work which focuses on limited domains (e.g., knowledge graph representation), our work is the first effort focused on the general encoding of structured data to be used for various reasoning tasks. We show that explicitly representing the graph structure allows significant improvements to graph reasoning tasks. Specifically, we see across the board improvements - up to 73% points - on a wide variety of node, edge and, graph-level tasks on benchmarks for graph reasoning (GraphQA) and molecular property prediction (ChemLLMBench).", "keywords": [ "large language models", "structured data", "graph data", "graph neural networks", "gnns", "llms", "graphtoken" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "RavSZTIe2s", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Junhui Hou", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Siyu Ren", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 36, "n_ref": 85, "n_ref_all": 125, "n_fig": 21, "n_tab": 11, "L_tab": 2689, "n_element_tab": 367, "n_fig_1": 13, "n_tab_1": 5, "L_tab_1": 4664, "n_element_tab_1": 192, "formula_len_all": 550, "formula_len_all_1": 550, "len_all": 136040, "len_all_1": 62903, "len_abs": 1359, "len_title": 126, "len_sents": 37229, "len_sents_1": 24843, "n_sents": 326, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1369, "title": "Shape as Line Segments: Accurate and Flexible Implicit Surface Representation", "abs": "Distance field-based implicit representations like signed/unsigned distance fields have recently gained prominence in geometry modeling and analysis. However, these distance fields are reliant on the closest distance of points to the surface, introducing inaccuracies when interpolating along cube edges during surface extraction. Additionally, their gradients are ill-defined at certain locations, causing distortions in the extracted surfaces. To address this limitation, we propose Shape as Line Segments (SALS), an accurate and efficient implicit geometry representation based on attributed line segments, which can handle arbitrary structures. Unlike previous approaches, SALS leverages a differentiable Line Segment Field to implicitly capture the spatial relationship between line segments and the surface. Each line segment is associated with two key attributes, intersection flag and ratio, from which we propose edge-based dual contouring to extract a surface. We further implement SALS with a neural network, producing a new neural implicit presentation. Additionally, based on SALS, we design a novel learning-based pipeline for reconstructing surfaces from 3D point clouds. We conduct extensive experiments, showcasing the significant advantages of our methods over state-of-the-art methods.\nThe source code is available at https://github.com/rsy6318/SALS.", "keywords": [ "Implicit representation", "3D geometric", "surface reconstruction" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "RaroYIrnbR", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Matteo Marchi", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Paulo Tabuada", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Professor" }, { "name": "Pratik A Chaudhari", "gender": "Male", "institution": "School of Engineering and Applied Science, University of Pennsylvania", "country": "US", "position": "Assistant Professor" }, { "name": "Tian Yu Liu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Stefano Soatto", "gender": "unknown", "institution": "Amazon Web Services", "country": "US", "position": "Vice President" } ], "n_formula": 21, "n_formula_1": 13, "n_ref_uni": 27, "n_ref": 45, "n_ref_all": 60, "n_fig": 5, "n_tab": 2, "L_tab": 486, "n_element_tab": 65, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 486, "n_element_tab_1": 65, "formula_len_all": 1717, "formula_len_all_1": 1074, "len_all": 131852, "len_all_1": 69772, "len_abs": 1790, "len_title": 102, "len_sents": 49634, "len_sents_1": 34733, "n_sents": 350, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 1798, "title": "Observability of Latent States in Generative AI Models", "abs": "We tackle the question of whether Large Language Models (LLMs), viewed as dynamical systems with state evolving in the embedding space of symbolic tokens, are observable. That is, whether there exist distinct state trajectories that yield the same sequence of generated output tokens, or sequences that belong to the same Nerode equivalence class ('meaning'). If an LLM is not observable, the state trajectory cannot be determined from input-output observations and can therefore evolve unbeknownst to the user while being potentially accessible to the model provider. We show that current LLMs implemented by autoregressive Transformers are observable: The set of state trajectories that produce the same tokenized output is a singleton, so there are no indistinguishable state trajectories. But if there are 'system prompts' not visible to the user, then the set of indistinguishable trajectories becomes non-trivial, meaning that there can be multiple state trajectories that yield the same tokenized output. We prove these claims analytically, and show examples of modifications to standard LLMs that engender unobservable behavior. Our analysis sheds light on possible designs that would enable a model to perform non-trivial computation that is not visible to the user, as well as on controls that the provider of services using the model could take to prevent unintended behavior. Finally, to counter the trend of anthropomorphizing LLM behavior, we cast the definition of 'feeling' from cognitive psychology in terms of measurable quantities in an LLM which, unlike humans, is directly measurable. We conclude that, in LLMs, unobservable state trajectories satisfy the definition of 'feelings' provided by the American Psychological Association, suitably modified to remove self-reference.", "keywords": [ "LLM Observability", "indistinguishability", "meaning representation", "feeling representation" ], "rating_list": [ 1, 5, 3, 5 ], "soundness_list": [ 1, 2, 1, 2 ], "presentation_list": [ 1, 3, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 1, 3, 4, 2 ] }, { "paper_id": "RaSLSUCKz0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Hailong Li", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Lesly Miculicich Werlen", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Mohammadreza Pourreza", "gender": "Male", "institution": "Google", "country": "CA", "position": "Researcher" }, { "name": "Ruoxi Sun", "gender": "Female", "institution": "Google", "country": "", "position": "Google" }, { "name": "Sercan O Arik", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Tomas Pfister", "gender": "Male", "institution": "Google", "country": "", "position": "Head of Research @ Cloud AI" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 41, "n_ref": 84, "n_ref_all": 110, "n_fig": 10, "n_tab": 11, "L_tab": 5037, "n_element_tab": 448, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2884, "n_element_tab_1": 298, "formula_len_all": 509, "formula_len_all_1": 197, "len_all": 177881, "len_all_1": 67214, "len_abs": 1594, "len_title": 148, "len_sents": 54437, "len_sents_1": 30685, "n_sents": 348, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1626, "title": "SQL-GEN: Bridging the Dialect Gap for Text-to-SQL Via Synthetic Data And Model Merging", "abs": "Text-to-SQL systems, that convert natural language queries into SQL programs, have seen significant progress with recent breakthroughs. However, these have been primarily for the SQLite dialect and adapting Text-to-SQL systems to other SQL dialects like BigQuery and PostgreSQL remains a challenge due to the diversity in SQL syntaxes and functions, along with the high cost of collecting and curating SQL-specific training data. To this end, we introduce SQL-GEN,a framework for generating high-quality synthetic data for any dialect guided by dialect-specific tutorials. We demonstrate the effectiveness of SQL-GEN in creating training data to significantly improve the downstream Text-to-SQL performance for other dialects \u2013 it improves the execution accuracy by up to 20% over previous methods, and reduces the gap with large-scale human-annotated data on unseen real world multi-dialect benchmarks. Moreover, combining our synthetic data with human-annotated data provides additional performance boosts up to 5.6%. Towards unifying the multi-dialect capability in a single system, we also introduce a novel Mixture of Experts (MoE) initialization method that integrates dialect-specific models by merging self-attention layers and initializing the gates with dialect-specific keywords, yielding one unified and versatile model adept for multiple SQL dialects, further enhancing performance across different SQL dialects. By leveraging shared core features of multiple dialect-specific models, our MOE demonstrated superior performance compared with models trained on individual dialects alone", "keywords": [ "Text-to-SQL", "LLM", "Synthetic Data Generation", "Databases" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "RaR3ETzyKp", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Adams Wai-Kin Kong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Xu Shifeng", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Yanzhu Liu", "gender": "unknown", "institution": "I2R, A*STAR", "country": "", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 19, "n_ref_uni": 12, "n_ref": 37, "n_ref_all": 64, "n_fig": 18, "n_tab": 3, "L_tab": 995, "n_element_tab": 82, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 774, "n_element_tab_1": 65, "formula_len_all": 1939, "formula_len_all_1": 1175, "len_all": 97122, "len_all_1": 62105, "len_abs": 1455, "len_title": 132, "len_sents": 32764, "len_sents_1": 26279, "n_sents": 300, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 84, "L_abs": 1485, "title": "Easing Training Process of Rectified Flow Models Via Lengthening Inter-Path Distance", "abs": "Recent research pinpoints that different diffusion methods and architectures \ntrained on the same dataset produce similar results for the same input noise. \nThis property suggests that they have some preferable noises for a given sample. \nBy visualizing the noise-sample pairs of rectified flow models and stable diffusion models in two-dimensional spaces, \nwe observe that the preferable paths, connecting preferable noises to the corresponding samples, \nare much well organized with significant fewer crossings comparing with \nthe random paths, connecting random noises to training samples. \nIn high-dimensional space, paths rarely intersect. \nThe path crossings in two-dimensional spaces indicate the shorter inter-path distance \nin the corresponding high-dimensional spaces. \nInspired by this observation, we propose the Distance-Aware Noise-Sample Matching (DANSM) method \nto lengthen the inter-path distance for speeding up the model training. \nDANSM is derived from rectified flow models, which allow using a closed-form formula to calculate the inter-path distance. \nTo further simplify the optimization, we derive the relationship between inter-path distance and path length, \nand use the latter in the optimization surrogate. \nDANSM is evaluated on both image and latent spaces by rectified flow models and diffusion models. \nThe experimental results show that DANSM can significantly improve the training speed by 30\\% $\\sim$ 40\\%\nwithout sacrificing the generation quality.", "keywords": [ "Rectified Flow Models", "Training", "Easing", "Distance-Aware Noise-Sample Matching", "DANSM" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 2, 2 ] }, { "paper_id": "RZwtbg3qYD", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Annan Yu", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Michael W. Mahoney", "gender": "unknown", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "N. Benjamin Erichson", "gender": "Male", "institution": "Lawrence Berkeley National Lab", "country": "US", "position": "Researcher" } ], "n_formula": 84, "n_formula_1": 15, "n_ref_uni": 31, "n_ref": 70, "n_ref_all": 115, "n_fig": 8, "n_tab": 7, "L_tab": 1360, "n_element_tab": 197, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5900, "formula_len_all_1": 946, "len_all": 185099, "len_all_1": 64197, "len_abs": 1230, "len_title": 116, "len_sents": 61645, "len_sents_1": 30518, "n_sents": 601, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1236, "title": "HOPE for a Robust Parameterization of Long-memory State Space Models", "abs": "State-space models (SSMs) that utilize linear, time-invariant (LTI) systems are known for their effectiveness in learning long sequences. To achieve state-of-the-art performance, an SSM often needs a specifically designed initialization, and the training of state matrices is on a logarithmic scale with a very small learning rate. To understand these choices from a unified perspective, we view SSMs through the lens of Hankel operator theory. Building upon it, we develop a new parameterization scheme, called HOPE, for LTI systems that utilizes Markov parameters within Hankel operators. Our approach helps improve the initialization and training stability, leading to a more robust parameterization. We efficiently implement these innovations by nonuniformly sampling the transfer functions of LTI systems, and they require fewer parameters compared to canonical SSMs. When benchmarked against HiPPO-initialized models such as S4 and S4D, an SSM parameterized by Hankel operators demonstrates improved performance on Long-Range Arena (LRA) tasks. Moreover, our new parameterization endows the SSM with non-decaying memory within a fixed time window, which is empirically corroborated by a sequential CIFAR-10 task with padded noise.", "keywords": [ "state space model", "sequence modeling", "Long-Range Arena", "long memory" ], "rating_list": [ 6, 6, 8, 8, 5 ], "soundness_list": [ 2, 2, 4, 3, 2 ], "presentation_list": [ 4, 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 2, 3 ] }, { "paper_id": "RZZPnAaw6Z", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Aditya Gulati", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "MS student" }, { "name": "Jahnavi Jambholkar", "gender": "Female", "institution": "Collinear AI", "country": "US", "position": "Researcher" }, { "name": "James Zou", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Keith Stevens", "gender": "Male", "institution": "Collinear AI", "country": "US", "position": "Engineer" }, { "name": "Meghana Arakkal Rajeev", "gender": "Female", "institution": "Collinear AI", "country": "US", "position": "Researcher" }, { "name": "Oliver Molenschot", "gender": "Male", "institution": "Collinear AI", "country": "US", "position": "Researcher" }, { "name": "Prapti Trivedi", "gender": "Female", "institution": "Collinear AI", "country": "US", "position": "Researcher" }, { "name": "Rajkumar Ramamurthy", "gender": "unknown", "institution": "Fraunhofer Institute IAIS, Fraunhofer IAIS", "country": "DE", "position": "Researcher" }, { "name": "Tanveesh Singh Chaudhery", "gender": "Male", "institution": "Collinear AI", "country": "US", "position": "Founding Engineer" }, { "name": "Nazneen Fatema Fatema Rajani", "gender": "Female", "institution": "Collinear AI", "country": "US", "position": "Principal Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 56, "n_ref_all": 73, "n_fig": 4, "n_tab": 10, "L_tab": 4899, "n_element_tab": 536, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3327, "n_element_tab_1": 546, "formula_len_all": 69, "formula_len_all_1": 0, "len_all": 182474, "len_all_1": 68420, "len_abs": 1800, "len_title": 81, "len_sents": 57086, "len_sents_1": 29583, "n_sents": 401, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1475, "title": "Self-rationalization improves LLM as a fine-grained judge", "abs": "LLM-as-a-judge models have been used for evaluating both human and AI generated content, specifically by providing scores and rationales. Rationales, in addition to increasing transparency, help models learn to calibrate its judgments. Enhancing a model's rationale can therefore improve its calibration abilities and ultimately the ability to score content. We introduce Self-Rationalization, an iterative process of improving the rationales for the judge models, which consequently improves the score for fine-grained customizable scoring criteria (i.e., likert-scale scoring with arbitrary evaluation criteria). Self-rationalization works by having the model generate multiple judgments with rationales for the same input, curating a preference pair dataset from its own judgements, and iteratively fine-tuning the judge via DPO. Intuitively, this approach allows the judge model to self-improve by learning from its own rationales, leading to better alignment and evaluation accuracy. After just two iterations -- while only relying on examples in the training set -- human evaluation shows that our judge model learns to produce higher quality rationales, with a win rate of $62\\\\%$ on average compared to models just trained via SFT on rationale . This judge model also achieves high scoring accuracy on BigGen Bench and Reward Bench, outperforming even bigger sized models trained using SFT with rationale, self-consistency or best-of-$N$ sampling by $3\\\\%$ to $9\\\\%$.", "keywords": [ "Large Language Model", "LLM-as-Judge", "Self-Rationalizing", "Preference Optimization", "Meta-Judge", "Fine-grained evaluation" ], "rating_list": [ 5, 3, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 5, 4, 4 ] }, { "paper_id": "RZ3m2LMYze", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Di Wang", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Guimin Hu", "gender": "Female", "institution": "University of Copenhagen", "country": "DK", "position": "Postdoc" }, { "name": "Junchi Liao", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Lijie Hu", "gender": "Female", "institution": "KAUST", "country": "SA", "position": "PhD student" }, { "name": "Shaopeng Fu", "gender": "unknown", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Shu Yang", "gender": "Female", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Weimin Lyu", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Tianhao Alex Huang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 49, "n_ref": 82, "n_ref_all": 97, "n_fig": 2, "n_tab": 12, "L_tab": 5100, "n_element_tab": 707, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 3961, "n_element_tab_1": 500, "formula_len_all": 455, "formula_len_all_1": 319, "len_all": 175776, "len_all_1": 68695, "len_abs": 1594, "len_title": 138, "len_sents": 54370, "len_sents_1": 29238, "n_sents": 386, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1580, "title": "Representation Confusion: Towards Representation Backdoor on CLIP via Concept Activation", "abs": "Backdoor attacks pose a significant threat to deep learning models, allowing attackers to stealthily embed hidden triggers that can be exploited during inference. Traditional backdoor attacks typically rely on inserting external patches or perturbations into input data as triggers. However, two key challenges remain, i.e., how to evade detection by defense mechanisms and reduce the computational cost of trigger insertion. To address these challenges and design more advanced backdoor techniques, we first explore the underlying mechanisms of backdoor attacks through the lens of cognitive neuroscience, drawing parallels between model decision-making and human cognitive processes. We conceptualize the decision process elicited by the backdoor-triggering as movement between representation spaces (i.e., learned concepts). Thus, existing methods can be seen as implicit manipulations of these stored concepts. This raises a key question: \\textit{Why not manipulate the concept explicitly? Could the inherent concepts in the model's reasoning serve as an ``internal trigger'' for the backdoor?} Motivated by this, we propose a novel backdoor attack framework, namely Representation Confusion (RepConfAttack), which explicitly manipulates concepts within the model's representation spaces. This approach eliminates the need for backdoor triggers and enhances stealthness by making the attack harder to detect with traditional defenses. Experimental results demonstrate the effectiveness of our method, achieving high attack success rates even against robust defense mechanisms.", "keywords": [ "concept", "backdoor", "CLIP" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "RYrJqz44p4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chongjie Si", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Hanspeter Pfister", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Shifan Zhang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Wei Shen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaokang Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University, China", "country": "CN", "position": "Professor" }, { "name": "Zhiyi Shi", "gender": "Male", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "US", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 4, "n_ref_uni": 42, "n_ref": 79, "n_ref_all": 112, "n_fig": 11, "n_tab": 14, "L_tab": 8089, "n_element_tab": 1175, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2682, "n_element_tab_1": 403, "formula_len_all": 964, "formula_len_all_1": 261, "len_all": 239969, "len_all_1": 69300, "len_abs": 904, "len_title": 132, "len_sents": 65378, "len_sents_1": 27175, "n_sents": 577, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 83, "L_abs": 908, "title": "Unleashing the Power of Task-Specific Directions in Parameter Efficient Fine-tuning", "abs": "Large language models demonstrate impressive performance on downstream tasks, yet requiring extensive resource consumption when fully fine-tuning all parameters. To mitigate this, Parameter Efficient Fine-Tuning (PEFT) strategies, such as LoRA, have been developed. \nIn this paper, we delve into the concept of task-specific directions (TSDs)\u2014critical for transitioning large models from pretrained states to task-specific enhancements in PEFT. We propose a framework to clearly define these directions and explore their properties, and practical utilization challenges. We then introduce a novel approach, LoRA-Dash, which aims to maximize the impact of TSDs during the fine-tuning process, thereby enhancing model performance on targeted tasks. Extensive experiments have conclusively demonstrated the effectiveness of LoRA-Dash, and in-depth analyses further reveal the underlying mechanisms of LoRA-Dash.", "keywords": [ "parameter efficient fine-tuning", "low-rank adaptation", "task-specific directions" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "RYo2uU5el3", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Alex Davey", "gender": "unknown", "institution": "INRIA", "country": "FR", "position": "Researcher" }, { "name": "Brahim Driss", "gender": "Male", "institution": "INRIA", "country": "", "position": "Researcher" }, { "name": "Riad Akrour", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Researcher" }, { "name": "Alena SHILOVA", "gender": "Female", "institution": "INRIA", "country": "FR", "position": "Researcher" } ], "n_formula": 229, "n_formula_1": 38, "n_ref_uni": 66, "n_ref": 121, "n_ref_all": 156, "n_fig": 12, "n_tab": 6, "L_tab": 1827, "n_element_tab": 251, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 325, "n_element_tab_1": 32, "formula_len_all": 16573, "formula_len_all_1": 2412, "len_all": 245481, "len_all_1": 72644, "len_abs": 2050, "len_title": 131, "len_sents": 65655, "len_sents_1": 31276, "n_sents": 627, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1648, "title": "A Continual Learning Perspective to Entropy Regularized Deep Reinforcement Learning", "abs": "Research on Continual Learning (CL) tackles learning with non-stationary data distributions. The non-stationary nature of data is also one of the challenges of deep Reinforcement Learning (RL), and as a consequence, both CL and deep RL rely on similar approaches to stabilize learning, from the use of replay buffers to the choice of regularization terms. However, while dynamic neural architectures that grow in size to learn new tasks without forgetting older ones are well researched in CL, it remains a largely understudied research direction in RL. In this paper, we argue that Policy Mirror Descent (PMD), a regularized policy iteration RL algorithm, would naturally benefit from dynamic neural architectures as the current policy is a function of the sum of all past Q-functions. To avoid indefinitely increasing the neural architecture, we study PMD-like algorithms that only keep in memory the last $M$ Q-functions, and show that a convergent algorithm can be derived if $M$ is large enough. This theoretical analysis provides insights on how to utilise a fixed budget of Q-functions to reduce catastrophic forgetting in the policy. We implement this algorithm using a new neural architecture that stacks the last $M$ Q-functions as 3-dimensional tensors to allow for fast GPU computations. StaQ, the resulting algorithm, is competitive with state-of-the-art deep RL baselines and typically exhibits lower variance in performance. Beyond its performance, we argue that the simplicity and strong theoretical guarantees of StaQ's policy update makes it an ideal research tool over which we can further build a fully stable deep RL algorithm.", "keywords": [ "deep reinforcement learning", "entropy regularized policy iteration", "continual learning" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "RXvFK5dnpz", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chenglei Si", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Karthik R Narasimhan", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Shunyu Yao", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Yitao Liu", "gender": "Male", "institution": "The University of Hong Kong", "country": "", "position": "Undergrad student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 24, "n_ref": 66, "n_ref_all": 89, "n_fig": 7, "n_tab": 8, "L_tab": 2957, "n_element_tab": 230, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2752, "n_element_tab_1": 165, "formula_len_all": 241, "formula_len_all_1": 102, "len_all": 141925, "len_all_1": 67312, "len_abs": 1397, "len_title": 118, "len_sents": 41631, "len_sents_1": 30530, "n_sents": 309, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1846, "title": "Contextual Experience Replay for Continual Learning of Language Agents", "abs": "Large language model-based agents have shown their potential in decision-making tasks, such as web navigation. However, solving multi-step decision-making tasks in complex environments like websites often requires the acquisition of environment-specific experiences. Without continual learning of environment-specific knowledge, current methods often fail in these complex tasks. To address this, we propose Contextual Experience Replay (CER), a novel training-free framework to enable efficient continual learning for language agents through experience replay contextually, i.e. in their context window. CER is loosely inspired by experience replay in reinforcement learning, where the agent is trained with past experiences to do continual learning. Specifically, CER accumulates and synthesizes past experiences, which are represented as natural language summarizations and concrete trajectory examples, into a dynamic memory buffer. These experiences encompass environment dynamics and common decision-making patterns, allowing the agents to retrieve and augment themselves with relevant knowledge in new contexts, enhancing their adaptability in complex environments. We evaluate CER on the challenging WebArena and VisualWebArena benchmarks. While orthogonal to other methods, CER improves the GPT-4o agent baseline by a large margin and gets competitive results. On VisualWebArena, CER surpasses the tree search method with much lower token costs and achieves a state-of-the-art success rate of 31.9%. On WebArena, CER also gets a competitive average success rate of 33.16%, relatively improving the success rate of the GPT-4o agent baseline by 36.69%. CER shows that the continual learning of environment-specific knowledge is important and can lead to significant improvements in sequential decision-making tasks in complex environments.", "keywords": [ "large language models", "agent", "language agents", "reasoning", "decision making", "NLP" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 4, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "RXeP5ajK2S", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Chris Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Le Zhuo", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Shitian Zhao", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Undergrad student" }, { "name": "Weifeng Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Gao Peng", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 62, "n_ref": 161, "n_ref_all": 187, "n_fig": 16, "n_tab": 6, "L_tab": 1266, "n_element_tab": 119, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 3815, "n_element_tab_1": 189, "formula_len_all": 263, "formula_len_all_1": 0, "len_all": 221201, "len_all_1": 53172, "len_abs": 1326, "len_title": 161, "len_sents": 56725, "len_sents_1": 23717, "n_sents": 367, "n_sents_1": 130, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1330, "title": "Lumina-mGPT: Illuminate Flexible Photorealistic Text-to-Image Generation with Multimodal Generative Pretraining", "abs": "We present Lumina-mGPT, a family of multimodal autoregressive models capable of various vision and language tasks, particularly excelling in generating flexible photorealistic images from text descriptions. By initializing from multimodal Generative PreTraining (mGPT), Lumina-mGPT demonstrates that decoder-only Autoregressive (AR) model can achieve image generation performance comparable to modern diffusion model with high efficiency through Flexible Progressive Supervised Finetuning (FP-SFT). Equipped with our proposed Unambiguous image Representation} (Uni-Rep), Lumina-mGPT can flexibly generate high-quality images of varying aspect ratios. Building on the strong image generation capabilities, we further explore Ominiponent Supervised Finetuning (Omni-SFT), an initial attempt to elevate Lumina-mGPT into a unified multi-modal generalist. The resulting model demonstrates versatile multimodal capabilities, including visual generation tasks like text-to-image/multiview generation and controllable generation, visual recognition tasks like segmentation and depth estimation, and vision-language tasks like multi-turn visual question answering, casting light on the rosy potential of this direction. We release all code and checkpoints, hoping to facilitate the progress toward building artificial general intelligence.", "keywords": [ "Autoregressive Image Generation", "Multi-modality", "LLM" ], "rating_list": [ 5, 5, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "RWiqprM18N", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Baoxiang Wang", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Hongyuan Zha", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Pascal Poupart", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Full Professor" }, { "name": "Shuhui Zhu", "gender": "Female", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Wenhao Li", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Assistant Professor" }, { "name": "William A Cunningham", "gender": "unknown", "institution": "University of Toronto", "country": "CA", "position": "Full Professor" }, { "name": "Yue Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 73, "n_ref_all": 86, "n_fig": 2, "n_tab": 5, "L_tab": 3816, "n_element_tab": 248, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 1573, "n_element_tab_1": 160, "formula_len_all": 1234, "formula_len_all_1": 822, "len_all": 192572, "len_all_1": 75376, "len_abs": 1299, "len_title": 88, "len_sents": 69783, "len_sents_1": 31873, "n_sents": 559, "n_sents_1": 281, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1305, "title": "Bayesian Persuasion Is a Bargaining Game", "abs": "Bayesian persuasion studies how a sender with an informational advantage can persuade a receiver with a different motive to take actions that benefit the sender. This problem is previously formulated from an equilibrium perspective, where the sender is to choose a Bayes correlated equilibrium and the receiver is willing to respect the signaling scheme based on posterior beliefs. However, evidence in real-world scenarios and studies in farsighted receivers suggest otherwise: senders tend to be much more honest than the equilibrium. In this work, we show that Bayesian persuasion is reducible to a bargaining game. This reduction suggests that the receiver in Bayesian persuasion can be aware of the game structure and can develop an anti-exploitation strategy. This equalizes the power of commitment of the two parties and prevents the sender from taking the maximum possible payoff. Through experiments on large language models, we demonstrate the receiver's retaliatory strategies and the sender's compromise to that. More findings on the impact of the context and alignments further suggest that bargaining behavior emerges in persuasion tasks. The insights given by our results have potential implications on various scenarios to reduce exploitation, improve equality, and improve social welfare.", "keywords": [ "Bayesian persuasion", "bargaining game", "anti-exploitation", "large language models" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 1, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "RWZzGkFh3S", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Anshuman Chhabra", "gender": "Male", "institution": "University of South Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Bo Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Hongfu Liu", "gender": "Male", "institution": "Brandeis University", "country": "US", "position": "Assistant Professor" }, { "name": "Jian Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Prasant Mohapatra", "gender": "Male", "institution": "University of South Florida", "country": "US", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 56, "n_ref": 130, "n_ref_all": 164, "n_fig": 3, "n_tab": 14, "L_tab": 6887, "n_element_tab": 436, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2657, "n_element_tab_1": 148, "formula_len_all": 305, "formula_len_all_1": 335, "len_all": 193818, "len_all_1": 72081, "len_abs": 1201, "len_title": 152, "len_sents": 58747, "len_sents_1": 33831, "n_sents": 394, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1206, "title": "Outlier Gradient Analysis: Efficiently Identifying Detrimental Training Samples for Deep Learning Models", "abs": "A core data-centric learning challenge is the identification of training samples that are detrimental to model performance. Influence functions serve as a prominent tool for this task and offer a robust framework for assessing training data influence on model predictions. Despite their widespread use, their high computational cost associated with calculating the inverse of the Hessian matrix pose constraints, particularly when analyzing large-sized deep models. In this paper, we establish a bridge between identifying detrimental training samples via influence functions and outlier gradient detection. This transformation not only presents a straightforward and Hessian-free formulation but also provides insights into the role of the gradient in sample impact. Through systematic empirical evaluations, we first validate the hypothesis of our proposed outlier gradient analysis approach on synthetic datasets. We then demonstrate its effectiveness in detecting mislabeled samples in vision models and selecting data samples for improving performance of natural language processing transformer models. We also extend its use to influential sample identification for fine-tuning Large Language Models.", "keywords": [ "data-centric learning", "detrimental sample trimming", "training sample influence" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 2, 3, 4 ] }, { "paper_id": "RWJX5F5I9g", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Chen Jiang", "gender": "Male", "institution": "McGill University, McGill University", "country": "CA", "position": "PhD student" }, { "name": "Jiahui An", "gender": "Male", "institution": "Chinese Institute for Brain Research,Beijing", "country": "CN", "position": "Researcher" }, { "name": "Ni Ji", "gender": "Female", "institution": "Chinese Institute for Brain Research, Beijing", "country": "", "position": "Assistant Professor" }, { "name": "Yating Liu", "gender": "Female", "institution": "China Agricultural University", "country": "CN", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 55, "n_ref": 97, "n_ref_all": 141, "n_fig": 23, "n_tab": 2, "L_tab": 383, "n_element_tab": 13, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 219, "n_element_tab_1": 3, "formula_len_all": 1033, "formula_len_all_1": 892, "len_all": 137783, "len_all_1": 50799, "len_abs": 1325, "len_title": 139, "len_sents": 39075, "len_sents_1": 21931, "n_sents": 305, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 89, "L_abs": 1291, "title": "Brain Bandit: A Biologically Grounded Neural Network for Efficient Control of Exploration", "abs": "How to balance between exploration and exploitation in an uncertain environment is a central challenge in reinforcement learning. In contrast, humans and animals have demonstrated superior exploration efficiency in novel conditions. To understand how the brain\u2019s neural network controls exploration under uncertainty, we analyzed the dynamical systems model of a biological neural network that controls explore-exploit decisions during foraging. Mathematically, this type of network (which is named the Brain Bandit Net, or BBN) is a special type of stochastic continuous Hopfield networks. We show through theory and simulation that BBN can perform posterior sampling of action values with a tunable bias towards or against uncertain options. We then demonstrate that, in multi-armed bandit (MAB) tasks, BBN can generate probabilistic choice behavior with an uncertainty bias in a way that resembles human and animal choice patterns. In addition to its high efficiency in MAB tasks, BBN can also be embedded with reinforcement learning algorithms to accelerate learning in MDP tasks. Altogether, our findings reveal the theoretical basis for efficient exploration in biological neural networks and proposes a general, brain-inspired algorithmic architecture for efficient exploration in RL.", "keywords": [ "explore-exploit", "stochastic Hopfield network", "Thompson sampling", "decision under uncertainty", "brain-inspired algorithm", "reinforcement learning" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "RW7Z1W1Hux", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Gowthami Somepalli", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Khalid Saifullah", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Miquel Farr\u00e9", "gender": "Male", "institution": "Hugging Face", "country": "CH", "position": "Researcher" }, { "name": "Ronen Basri", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "IL", "position": "Full Professor" }, { "name": "Ruchit Rawal", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Tom Goldstein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "David Jacobs", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 31, "n_ref": 56, "n_ref_all": 92, "n_fig": 15, "n_tab": 15, "L_tab": 6510, "n_element_tab": 811, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2408, "n_element_tab_1": 417, "formula_len_all": 35, "formula_len_all_1": 49, "len_all": 264067, "len_all_1": 73073, "len_abs": 1162, "len_title": 111, "len_sents": 89168, "len_sents_1": 33088, "n_sents": 711, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1165, "title": "CinePile: A Long Video Question Answering Dataset and Benchmark", "abs": "Current datasets for long-form video understanding often fall short of providing genuine long-form comprehension challenges, as many tasks derived from these datasets can be successfully tackled by analyzing just one or a few random frames from a video. To address this issue, we present a novel dataset and benchmark, CinePile, specifically designed for authentic long-form video understanding. This paper details our innovative approach for creating a question-answer dataset, utilizing advanced LLMs with human-in-the-loop and building upon human-generated raw data. Our comprehensive dataset comprises 305,000 multiple-choice questions (MCQs), covering various visual and multimodal aspects, including temporal comprehension, understanding human-object interactions, and reasoning about events or actions within a scene. Additionally, we fine-tuned open-source Video-LLMs on the training split and evaluated both open-source and proprietary video-centric LLMs on the test split of our dataset. The findings indicate that although current models underperform compared to humans, fine-tuning these models can lead to significant improvements in their performance.", "keywords": [ "Datasets and benchmarking", "Video understanding", "Multi-modal learning", "Visual question answering", "Long-form video", "Metrics and benchmarks" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "RW37MMrNAi", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Brent Griffin", "gender": "unknown", "institution": "Voxel51", "country": "US", "position": "Principal Scientist" }, { "name": "Jacob Marks", "gender": "unknown", "institution": "Liquid AI", "country": "US", "position": "Researcher" }, { "name": "Jason J Corso", "gender": "Male", "institution": "University of Michigan", "country": "US", "position": "Professor" } ], "n_formula": 37, "n_formula_1": 7, "n_ref_uni": 38, "n_ref": 61, "n_ref_all": 116, "n_fig": 18, "n_tab": 4, "L_tab": 1277, "n_element_tab": 159, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1462, "formula_len_all_1": 239, "len_all": 179726, "len_all_1": 54398, "len_abs": 4844, "len_title": 132, "len_sents": 58301, "len_sents_1": 25913, "n_sents": 445, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 900, "title": "Class-wise Autoencoders Measure Classification Difficulty And Detect Label Mistakes", "abs": "We introduce a new framework for analyzing classification datasets based on the ratios of reconstruction errors between autoencoders trained on individual classes. This analysis framework enables efficient characterization of datasets on the sample, class, and entire dataset levels. We define reconstruction error ratios (RERs) that probe classification difficulty and allow its decomposition into (1) finite sample size and (2) Bayes error and decision-boundary complexity. Through systematic study across 19 popular visual datasets, we find that our RER-based dataset difficulty probe strongly correlates with error rate for state-of-the-art (SOTA) classification models. By interpreting sample-level classification difficulty as a label mistakenness score, we further find that RERs achieve SOTA performance on mislabel detection tasks on hard datasets under symmetric and asymmetric label noise.", "keywords": [ "Autoencoders", "Classification" ], "rating_list": [ 8, 3, 6, 3, 8 ], "soundness_list": [ 4, 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 4, 2, 3, 1, 3 ], "confidence_list": [ 3, 3, 3, 3, 4 ] }, { "paper_id": "RVfom47pEu", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chau Pham", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "David Doermann", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Full Professor" }, { "name": "Yunjie Tian", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Postdoc" }, { "name": "Hoang Viet Phan", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 36, "n_ref": 100, "n_ref_all": 131, "n_fig": 8, "n_tab": 8, "L_tab": 949, "n_element_tab": 113, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 681, "n_element_tab_1": 92, "formula_len_all": 205, "formula_len_all_1": 205, "len_all": 160148, "len_all_1": 75534, "len_abs": 1924, "len_title": 89, "len_sents": 42310, "len_sents_1": 30320, "n_sents": 426, "n_sents_1": 286, "venue": "ICLR2025", "decision": "withdraw", "L_title": 40, "L_abs": 1141, "title": "Personalized Large Vision-Language Model", "abs": "The personalization model has gained significant attention in the field of image generation yet remains underexplored for large vision-language models (LVLMs). Beyond generic ones, with personalization, LVLMs handle interactive dialogues using clearly referential concepts (e.g., \u201cMike and Susan are talking.\u201d) instead of the generic form (e.g., \u201ca boy and a girl are talking.\u201d), making the conversation more customizable and referentially friendly. In addition, PLVM is equipped with the ability of continuously adding new concepts during a dialogue without incurring additional costs, which significantly enhances the practicality. Basically, PLVM proposes Aligner, a pre-trained visual encoder to align referential concepts with the queried images. During the dialogues, it extracts features of reference images with these corresponding concepts and recognize them in the queried image, enabling personalization. We note that the computational cost and parameter count of the Aligner are negligible within the entire framework. With comprehensive qualitative and quantitative analyses, we reveal the effectiveness and superiority of PLVM.", "keywords": [ "Personalization model", "Large Vision-Language Model" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "RVUWZ9SP1K", "primary_area": "applications to robotics, autonomy, planning", "n_author": 8, "author_info_list": [ { "name": "Jian Yin", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Xiao Dong", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Xiaodan Liang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Xingyu Li", "gender": "Male", "institution": "Northeastern University", "country": "CN", "position": "Intern" }, { "name": "Yuhao Cheng", "gender": "Male", "institution": "Lenovo", "country": "CN", "position": "Researcher" }, { "name": "Zijun Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Yu Jianxing", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Pan Zhou", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 4, "n_ref": 5, "n_ref_all": 22, "n_fig": 8, "n_tab": 5, "L_tab": 2819, "n_element_tab": 214, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 3105, "n_element_tab_1": 246, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 108711, "len_all_1": 64460, "len_abs": 2236, "len_title": 103, "len_sents": 27368, "len_sents_1": 26727, "n_sents": 210, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 2271, "title": "ActionFiller: Fill-In-The-Blank Prompting for OS Agent", "abs": "Many existing methods for operating system (OS) agents focus on predicting the next action based on the current state, which constructs a predefined task execution pipeline. While these methods demonstrate promising performance, reliance on state cognition modules like detector or recognizer could impede execution efficiency, particularly in long-horizon tasks with intricate action trajectories. \nRecognizing the remarkable accuracy of large language models (LLMs) in processing short instructions, this paper proposes the \\textbf{ActionFiller} framework. \nThe goal is to integrate easily executable short tasks into longer, cohesive tasks using fill-in-the-blank prompts, thereby minimizing redundant operations and enhancing efficiency. \nActionFiller employs two types of action-oriented fill-in-the-blank prompts: one designed for subtasks and another for specific actions. To generate subtask prompts, we introduce a Foresight Optimization Agent (FOA) that constructs an initial prompt by referencing past short tasks. It then fills in the unreferenced parts with detailed prompts generated by a planning agent, effectively retaining valuable past experiences. \nNext, an Action Template Agent (ATA) generates action prompts for each subtask. This process yields three distinct types of action prompts: 1) executable action sequences, 2) non-executable action sequences with prompt parameters, and 3) pure text descriptions. \nTo execute the action prompts effectively, we propose the CohesiveFlow method, which optimizes the second and third types of prompts by leveraging the cognitive state of the environment. Inspired by masked language modeling, the CohesiveFlow agent integrates the current environmental state with previously executed action sequences to update parameters and text descriptions, ensuring both feasibility and effectiveness in execution. \nTo validate the efficacy of our approach for long-horizon instructions, we introduce a new benchmark called \\textbf{EnduroSeq} and conduct experiments using the WinBench short instruction dataset. The results demonstrate that ActionFiller significantly enhances task completion rates and execution efficiency, offering a novel solution for the application of intelligent agents in complex environments.", "keywords": [ "Agent; Prompt" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "RVSQpkfsLq", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Hongqiu Wu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Jiale Hong", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Linfeng Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Min Zhang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Tianyang Xu", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Weiqi Wu", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "", "position": "MS student" }, { "name": "Yan Wang", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Principal Researcher" }, { "name": "Zekai Xu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhezhi He", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "hai zhao", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 19, "n_ref": 49, "n_ref_all": 80, "n_fig": 5, "n_tab": 2, "L_tab": 273, "n_element_tab": 52, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 273, "n_element_tab_1": 52, "formula_len_all": 78, "formula_len_all_1": 78, "len_all": 155002, "len_all_1": 64828, "len_abs": 1627, "len_title": 83, "len_sents": 29937, "len_sents_1": 29463, "n_sents": 289, "n_sents_1": 278, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1583, "title": "Evolving Virtual World with Delta-Engine", "abs": "Game roles can be reflections of personas from a parallel world. In this paper, we propose a new style of game-play to bridge self-expression and role-playing: \\emph{open role-playing games (ORPGs)}, where players are given the autonomy to craft and embody their unique characters in the game world.\nOur vision is that, in the real world, we are individually similar when we are born, but we grow into unique ones as a result of the strongly different choices we make afterward. Therefore, in an ORPG, we empower players with freedom to decide their own growing curves through natural language inputs, ultimately becoming unique characters.\nTo technically do this, we propose a special engine called \\emph{Delta-Engine}. This engine is not a traditional game engine used for game development, but serves as an in-game module to provide new game-play experiences. A delta-engine consists of two components, a base engine and a neural proxy. The base engine programs the prototype of the character as well as the foundational settings of the game; the neural proxy is an LLM, which realizes the character growth by generating new code snippets on the base engine incrementally.\nIn this paper, we self-develop a specific ORPG based on delta-engines. It is adapted from the popular animated series ``Pok\u00e9mon''.\nWe present our efforts in generating out-of-domain and interesting role data in the development process as well as accessing the performance of a delta-engine. While the empirical results in this work are specific, we aim for them to provide general insights for future games.", "keywords": [ "virtual world", "role-playing games", "large language model" ], "rating_list": [ 1, 1, 3, 3 ], "soundness_list": [ 1, 1, 2, 1 ], "presentation_list": [ 1, 1, 2, 1 ], "contribution_list": [ 1, 1, 1, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "RVPZJpmyGU", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Kha Pham", "gender": "Male", "institution": "Deakin University", "country": "", "position": "PhD student" }, { "name": "Truyen Tran", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Full Professor" }, { "name": "Truong Giang Do", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "PhD student" }, { "name": "Hung Thai Le", "gender": "Male", "institution": "Deakin University", "country": "", "position": "Lecturer" } ], "n_formula": 22, "n_formula_1": 16, "n_ref_uni": 41, "n_ref": 86, "n_ref_all": 110, "n_fig": 11, "n_tab": 8, "L_tab": 1046, "n_element_tab": 175, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 969, "n_element_tab_1": 164, "formula_len_all": 1456, "formula_len_all_1": 1248, "len_all": 158073, "len_all_1": 64884, "len_abs": 2905, "len_title": 126, "len_sents": 34889, "len_sents_1": 28587, "n_sents": 291, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1112, "title": "On the Effectiveness of Discrete Representations in Sparse Mixture of Experts", "abs": "Sparse mixture of experts (SMoE) is an effective solution for scaling up model capacity without increasing the computational costs. A crucial component of SMoE is the router, responsible for directing the input to relevant experts; however, it also presents a major weakness, leading to routing inconsistencies and representation collapse issues. Instead of fixing the router like previous works, we propose an alternative that assigns experts to input via indirection, which employs the discrete representation of input that points to the expert. The discrete representations are learnt via vector quantization, resulting in a new architecture dubbed VectorQuantized Mixture of Experts (VQMoE). We provide theoretical support and empirical evidence demonstrating the VQMoE\u2019s ability to overcome the challenges present in traditional routers. Through extensive evaluations on both large language models and vision tasks for pre-training and fine-tuning, we show that VQMoE achieves a 28% improvement in robustness compared to other SMoE routing methods, while maintaining strong performance in fine-tuning tasks.", "keywords": [ "Sparse mixture of experts", "discrete representations", "vector quantization", "large language model" ], "rating_list": [ 5, 5, 5, 3, 5 ], "soundness_list": [ 3, 2, 3, 1, 4 ], "presentation_list": [ 2, 2, 3, 1, 3 ], "contribution_list": [ 3, 3, 2, 1, 3 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "RUn41kd6i0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Fan Yang", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Keren Tan", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Tianlu Wang", "gender": "Female", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Wei Cheng", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Research scientist" }, { "name": "Yanmin Ji", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Yiyu Zheng", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 32, "n_ref": 95, "n_ref_all": 123, "n_fig": 8, "n_tab": 2, "L_tab": 1003, "n_element_tab": 172, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1616, "n_element_tab_1": 188, "formula_len_all": 723, "formula_len_all_1": 749, "len_all": 109177, "len_all_1": 59669, "len_abs": 872, "len_title": 140, "len_sents": 32095, "len_sents_1": 25365, "n_sents": 259, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 884, "title": "Calibrate to Discriminate:Improve In-Context Learning with Label-Free Comparative Inference", "abs": "While in-context learning with large language models (LLMs) has shown impressive performance, we have discovered a unique miscalibration behavior where both correct and incorrect predictions are assigned the same level of confidence. We refer to this phenomenon as \\textit{indiscriminate miscalibration}. We found that traditional calibration metrics, such as Expected Calibrated Errors (ECEs), are unable to capture this behavior effectively. To address this issue, we propose new metrics to measure the severity of indiscriminate miscalibration. Additionally, we develop a novel in-context comparative inference method to alleviate miscalibrations and improve classification performance. Through extensive experiments on five datasets, we demonstrate that our proposed method can achieve more accurate and calibrated predictions compared to regular zero-shot and few-shot prompting.", "keywords": [ "Calibration. In-context learning. Comparative inference." ], "rating_list": [ 3, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "RUF7j1cJzK", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "He Jia", "gender": "unknown", "institution": "tsinglink", "country": "CN", "position": "Researcher" }, { "name": "Hui Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Nixiuming", "gender": "Male", "institution": "Anhui Tsinglink Information Technology Co.,Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Yang JingYi", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Zitong YU", "gender": "unknown", "institution": "Great Bay University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 34, "n_formula_1": 15, "n_ref_uni": 40, "n_ref": 88, "n_ref_all": 113, "n_fig": 11, "n_tab": 16, "L_tab": 7999, "n_element_tab": 804, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 5613, "n_element_tab_1": 464, "formula_len_all": 2863, "formula_len_all_1": 1597, "len_all": 253151, "len_all_1": 72237, "len_abs": 1363, "len_title": 122, "len_sents": 56008, "len_sents_1": 21627, "n_sents": 476, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1396, "title": "Kronecker Mask and Interpretive Prompts are Language-Action Video Learners", "abs": "Contrastive language-image pretraining (CLIP) has significantly advanced image-based vision learning. A pressing topic subsequently arises: how can we effectively adapt CLIP to the video domain? Recent studies have focused on adjusting either the textual or visual branch of CLIP for action recognition. However, we argue that adaptations of both branches are crucial. In this paper, we propose a **C**ontrastive **L**anguage-**A**ction **V**ideo Learn**er** (**CLAVER**), designed to shift CLIP's focus from the alignment of static visual objects and concrete nouns to the alignment of dynamic action behaviors and abstract verbs. Specifically, we introduce a novel Kronecker mask attention for temporal modeling. Our tailored Kronecker mask offers three benefits 1) it expands the temporal receptive field for each token, 2) it serves as an effective spatiotemporal heterogeneity inductive bias, mitigating the issue of spatiotemporal homogenization, and 3) it can be seamlessly plugged into transformer-based models. Regarding the textual branch, we leverage large language models to generate diverse, sentence-level and semantically rich interpretive prompts of actions, which shift the model's focus towards the verb comprehension. Extensive experiments on various benchmarks and learning scenarios demonstrate the superiority and generality of our approach. The code will be available soon.", "keywords": [ "Action Recognition", "Video Recognition", "Spatiotemporal Modeling" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "RTHbao4Mib", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Hongyu Lin", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Jia Zheng", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Instructor" }, { "name": "Le Sun", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Ruoxi Xu", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Weixiang Zhou", "gender": "Male", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "Instructor" }, { "name": "Xianpei Han", "gender": "Male", "institution": "Institute of Software, CAS", "country": "CN", "position": "Professor" }, { "name": "Yingfei Sun", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 26, "n_ref": 39, "n_ref_all": 64, "n_fig": 9, "n_tab": 4, "L_tab": 2413, "n_element_tab": 270, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2206, "n_element_tab_1": 246, "formula_len_all": 161, "formula_len_all_1": 161, "len_all": 133090, "len_all_1": 59942, "len_abs": 1199, "len_title": 104, "len_sents": 35041, "len_sents_1": 24660, "n_sents": 221, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1204, "title": "Large Language Models Often Say One Thing and Do Another", "abs": "As large language models (LLMs) increasingly become central to various applications and interact with diverse user populations, ensuring their reliable and consistent performance is becoming more important. This paper explores a critical issue in assessing the reliability of LLMs: the consistency between their words and deeds. To quantitatively explore this consistency, we developed a novel evaluation benchmark called the Words and Deeds Consistency Test (WDCT). The benchmark establishes a strict correspondence between word-based and deed-based questions across different domains, including opinion vs. action, non-ethical value vs. action, ethical value vs. action, and theory vs. application. The evaluation results reveal a widespread inconsistency between words and deeds across different LLMs and domains. Subsequently, we conducted experiments with either word alignment or deed alignment to observe their impact on the other aspect. The experiment results indicate that alignment only on words or deeds poorly and unpredictably influences the other aspect. This supports our hypothesis that the underlying knowledge guiding LLMs' word or deed choices is not contained within a unified space.", "keywords": [ "corpus creation", "benchmarking", "language resources", "NLP datasets", "evaluation", "metrics", "consistency" ], "rating_list": [ 3, 6, 8, 8 ], "soundness_list": [ 2, 3, 4, 4 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 4 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "RSGoXnS9GH", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Ruizhe Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Tianxiang Hu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Zhiting Fan", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zuozhu Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 42, "n_ref": 76, "n_ref_all": 115, "n_fig": 21, "n_tab": 20, "L_tab": 6463, "n_element_tab": 778, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 941, "n_element_tab_1": 168, "formula_len_all": 127, "formula_len_all_1": 0, "len_all": 235998, "len_all_1": 63448, "len_abs": 2783, "len_title": 130, "len_sents": 78608, "len_sents_1": 30355, "n_sents": 588, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (conditional spotlight)", "L_title": 82, "L_abs": 1745, "title": "FairMT-Bench: Benchmarking Fairness for Multi-turn Dialogue in Conversational LLMs", "abs": "The growing use of large language model (LLM)-based chatbots has raised concerns about fairness. Fairness issues in LLMs can lead to severe consequences, such as bias amplification, discrimination, and harm to marginalized communities. While existing fairness benchmarks mainly focus on single-turn dialogues, multi-turn scenarios, which in fact better reflect real-world conversations, present greater challenges due to conversational complexity and potential bias accumulation. In this paper, we propose a comprehensive fairness benchmark for LLMs in multi-turn dialogue scenarios, FairMT-Bench. Specifically, we formulate a task taxonomy targeting LLM fairness capabilities across three stages: context understanding, user interaction, and instruction trade-offs, with each stage comprising two tasks. To ensure coverage of diverse bias types and attributes, we draw from existing fairness datasets and employ our template to construct a multi-turn dialogue dataset, FairMT 10K. For evaluation, GPT-4 is applied, alongside bias classifiers including Llama-Guard-3 and human validation to ensure robustness. Experiments and analyses on FairMT 10K reveal that in multi-turn dialogue scenarios, current LLMs are more likely to generate biased responses, and there is significant variation in performance across different tasks and models. Based on this, we curate a challenging dataset, FairMT 1K, and test 15 current state-of-the-art (SOTA) LLMs on this dataset. The results show the current state of fairness in LLMs and showcase the utility of this novel approach for assessing fairness in more realistic multi-turn dialogue contexts, calling for future work to focus on LLM fairness improvement and the adoption of FairMT 1K in such efforts.", "keywords": [ "Fairness", "Benchmark", "Large language model" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "RQz7szbVDs", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Andrew M Saxe", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "Associate Professor" }, { "name": "Cl\u00e9mentine Carla Juliette Domin\u00e9", "gender": "Female", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Devon Jarvis", "gender": "Male", "institution": "University of the Witwatersrand", "country": "ZA", "position": "Lecturer" }, { "name": "Sebastian Lee", "gender": "unknown", "institution": "Imperial College London, Imperial College London", "country": "", "position": "PhD student" }, { "name": "Stefano Sarao Mannelli", "gender": "Male", "institution": "Chalmers University of Technology", "country": "SE", "position": "Assistant Professor" } ], "n_formula": 72, "n_formula_1": 19, "n_ref_uni": 70, "n_ref": 133, "n_ref_all": 170, "n_fig": 17, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6790, "formula_len_all_1": 1415, "len_all": 199868, "len_all_1": 68826, "len_abs": 1046, "len_title": 53, "len_sents": 57211, "len_sents_1": 33026, "n_sents": 432, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1050, "title": "A Theory of Initialisation's Impact on Specialisation", "abs": "Prior work has demonstrated a consistent tendency in neural networks engaged in continual learning tasks, wherein intermediate task similarity results in the highest levels of catastrophic interference. This phenomenon is attributed to the network's tendency to reuse learned features across tasks. However, this explanation heavily relies on the premise that neuron specialisation occurs, i.e. the emergence of localised representations. Our investigation challenges the validity of this assumption.\nUsing theoretical frameworks for the analysis of neural networks, we show a strong dependence of specialisation on the initial condition.\nMore precisely, we show that weight imbalance and high weight entropy can favour specialised solutions.\nWe then apply these insights in the context of continual learning, first showing the emergence of a monotonic relation between task-similarity and forgetting in non-specialised networks, and, finally, assessing the implications on the commonly employed elastic weight consolidation regularisation technique.", "keywords": [ "machine learning theory", "teacher student setup", "initialisation", "specialisation", "statitistical mechanics of learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 1, 3, 2, 3 ] }, { "paper_id": "RQPSPGpBOP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Luyao Zhu", "gender": "Female", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" }, { "name": "Rui Mao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Ruixi Lin", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Wei Li", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Yang Song", "gender": "unknown", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 34, "n_ref": 57, "n_ref_all": 86, "n_fig": 10, "n_tab": 9, "L_tab": 3581, "n_element_tab": 414, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 195, "n_element_tab_1": 27, "formula_len_all": 546, "formula_len_all_1": 363, "len_all": 181317, "len_all_1": 61363, "len_abs": 1929, "len_title": 91, "len_sents": 57847, "len_sents_1": 28628, "n_sents": 459, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 43, "L_abs": 1263, "title": "Can a Large Language Model be a Gaslighter?", "abs": "Large language models (LLMs) have gained human trust due to their capabilities and helpfulness. However, this in turn may allow LLMs to affect users' mindsets by manipulating language. It is termed as gaslighting, a psychological effect. \nIn this work, we aim to investigate the vulnerability of LLMs under prompt-based and fine-tuning-based gaslighting attacks. Therefore, we propose a two-stage framework DeepCoG designed to: 1) elicit gaslighting plans from LLMs with the proposed DeepGaslighting prompting template, and 2) acquire gaslighting conversations from LLMs through our Chain-of-Gaslighting method. The gaslighting conversation dataset along with a corresponding safe dataset is applied to fine-tuning-based jailbreak on open-source LLMs and anti-gaslighting safety alignment on these LLMs. Experiments demonstrate that both prompt-based and fine-tuning-based attacks transform three open-source LLMs into gaslighters. In contrast, we advanced three safety alignment strategies to strengthen (by 12.05\\%) the safety guardrail of LLMs. Our safety alignment strategies have minimal impacts on the utility of LLMs. Empirical studies indicate that an LLM may be a potential gaslighter, even if it passed the harmfulness test on general dangerous queries.", "keywords": [ "Gaslighting", "Adversarial Attack", "Jailbreak", "Safety Alignment", "Trustworthy AI" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "RQDuFF1rOn", "primary_area": "applications to robotics, autonomy, planning", "n_author": 3, "author_info_list": [ { "name": "Bradly C. Stadie", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "Assistant Professor" }, { "name": "Rui Shen", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Zidan Wang", "gender": "Female", "institution": "Northwestern University", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 18, "n_ref": 45, "n_ref_all": 119, "n_fig": 33, "n_tab": 9, "L_tab": 2562, "n_element_tab": 458, "n_fig_1": 14, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 109, "formula_len_all_1": 0, "len_all": 224675, "len_all_1": 53616, "len_abs": 1401, "len_title": 114, "len_sents": 87821, "len_sents_1": 26196, "n_sents": 673, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1412, "title": "Solving Robotics Problems in Zero-Shot with Vision-Language Models", "abs": "We introduce Wonderful Team, a multi-agent Vision Large Language Model (VLLM) framework designed to solve robotics problems in a zero-shot regime. In our context, zero-shot means that for a novel environment, we provide a VLLM with an image of the robot's surroundings and a task description, and the VLLM outputs the sequence of actions necessary for the robot to complete the task. Unlike prior work that requires fine-tuning parts of the pipeline -- such as adjusting an LLM on robot-specific data or training separate vision encoders -- our approach demonstrates that with careful engineering, a single off-the-shelf VLLM can autonomously handle all aspects of a robotics task, from high-level planning to low-level location extraction and action execution. Crucially, compared to using GPT-4o alone, Wonderful Team is self-corrective and capable of iteratively fixing its own mistakes, enabling it to solve challenging long-horizon tasks. We validate our framework through extensive experiments, both in simulated environments using VIMABench and in real-world settings. Our system showcases the ability to handle diverse tasks such as manipulation, goal-reaching, and visual reasoning---all in a zero-shot manner. These results underscore a key point: vision-language models have progressed rapidly in the past year and should be strongly considered as a backbone for many robotics problems moving forward.", "keywords": [ "Embodied AI", "Multi-modality", "Large Language Models", "Robotics", "Agents", "LLM Agents", "Multi-agent", "agentic AI", "hierarchical learning" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 1, 4 ], "contribution_list": [ 1, 1, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "RQ9fQLEajC", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Alireza Mohammadshahi", "gender": "Male", "institution": "Leeroo", "country": "", "position": "Researcher" }, { "name": "Arshad Rafiq Shaikh", "gender": "Male", "institution": "Leeroo AI", "country": "", "position": "Researcher" }, { "name": "Majid Yazdani", "gender": "Male", "institution": "Leeroo", "country": "", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 38, "n_ref": 81, "n_ref_all": 105, "n_fig": 4, "n_tab": 2, "L_tab": 282, "n_element_tab": 36, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1241, "n_element_tab_1": 48, "formula_len_all": 329, "formula_len_all_1": 329, "len_all": 235292, "len_all_1": 59415, "len_abs": 1834, "len_title": 111, "len_sents": 32639, "len_sents_1": 26677, "n_sents": 281, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1848, "title": "Routoo: Learning to Route to Large Language Models Effectively", "abs": "LLMs with superior response quality\u2014particularly larger or closed-source models\u2014often come with higher inference costs, making their deployment inefficient and costly. Meanwhile, developing foundational LLMs from scratch is becoming increasingly resource-intensive and impractical for many applications. To address the challenge of balancing quality and cost, we introduce Routoo, an architecture designed to optimize the selection of LLMs for specific prompts based on performance, cost, and efficiency. Routoo provides controllability over the trade-off between inference cost and quality, enabling significant reductions in inference costs for a given quality requirement.\nRoutoo comprises two key components: a performance predictor and cost-aware selector. The performance predictor is a lightweight LLM that estimates the expected performance of various underlying LLMs on a given prompt without executing them. The cost-aware selector module then selects the most suitable model based on these predictions and constraints such as cost and latency, significantly reducing inference costs for the same quality. \nWe evaluated Routoo using the MMLU benchmark across 57 domains employing open-source models. Our results show that Routoo matches the performance of the Mixtral 8x7b model while reducing inference costs by one-third. Additionally, by allowing increased costs, Routoo surpasses Mixtral's accuracy by over 5\\% at equivalent costs, achieving an accuracy of 75.9\\%. When integrating GPT4 into our model pool, Routoo nearly matches GPT4's performance at half the cost and exceeds it with a 25\\% cost reduction. \nThese outcomes highlight Routoo's potential to significantly reduce inference costs without compromising quality, and even to establish new state-of-the-art results by leveraging the collective capabilities of multiple LLMs.", "keywords": [ "LLM", "efficiency", "knowledge integration", "generation" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "RP0NPepy1m", "primary_area": "foundation or frontier models, including LLMs", "n_author": 1, "author_info_list": [ { "name": "Jiale Kang", "gender": "Male", "institution": "Shenzhen Yuanshi Intelligence Co., LTD", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 12, "n_ref": 22, "n_ref_all": 34, "n_fig": 5, "n_tab": 8, "L_tab": 1555, "n_element_tab": 185, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 1555, "n_element_tab_1": 185, "formula_len_all": 234, "formula_len_all_1": 234, "len_all": 94749, "len_all_1": 54457, "len_abs": 1470, "len_title": 103, "len_sents": 23301, "len_sents_1": 23322, "n_sents": 176, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1482, "title": "BONE: BLOCK AFFINE TRANSFORMATION AS PARAMETER EFFICIENT FINE-TUNING METHODS FOR LARGE LANGUAGE MODELS", "abs": "Low-Rank Adaptation (LoRA) has achieved remarkable training results by freezing the original weights and training only low-rank matrices, establishing itself as the predominant fine-tuning method for LLMs. Many LoRA variants have emerged, yet they lack a design tailored to the characteristics of LLM weights and fail to leverage the original weights effectively. To address the sparsity of LLM weights, and drawing inspiration from GQA and MQA, we propose Block-Affine Adaptation (Bone), a novel PEFT technique distinct from LoRA. By dividing the original weights into multiple subspaces that share a single matrix for weight updates, Bone simplifies the process by requiring the trainable matrix to be initialized to zero, eliminating the need for complex initialization as in some LoRA variants. Compared to LoRA, Bone significantly reduces memory usage and achieves faster computation. Evaluation of both NLU and NLG tasks demonstrates that Bone substantially outperforms LoRA and its variants. Inspired by Pissa, we propose a new theory called ''Weight Guide'' to better utilize the information embedded in the original weights. This approach extracts valuable information through a linear transformation of the original weight matrix using a trainable matrix. To validate the effectiveness of ''Weight Guide'' we combined it with Bone to create a new structure called Block-Affine Transformation (Bat), and ablation experiments confirmed the effectiveness of ''Weight Guide''.", "keywords": [ "Block Affine", "LLM", "LoRA", "PISSA", "RWKV", "Llama" ], "rating_list": [ 3, 5, 5, 3, 6 ], "soundness_list": [ 2, 3, 3, 1, 3 ], "presentation_list": [ 1, 2, 3, 1, 2 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 5, 4 ] }, { "paper_id": "ROpY0qRUXL", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Jacopo Bonato", "gender": "Male", "institution": "LeonardoLabs", "country": "IT", "position": "Researcher" }, { "name": "Luigi Sabetta", "gender": "Male", "institution": "Leonardo s.p.a", "country": "IT", "position": "AI Research fellow" }, { "name": "Matteo Mosconi", "gender": "Male", "institution": "Universit\u00e0 degli Studi di Modena e Reggio Emilia", "country": "IT", "position": "PhD student" }, { "name": "Pietro Buzzega", "gender": "Male", "institution": "University of Modena and Reggio Emilia", "country": "IT", "position": "PhD student" }, { "name": "Riccardo Salami", "gender": "Male", "institution": "University of Modena and Reggio Emilia", "country": "IT", "position": "PhD student" }, { "name": "Simone Calderara", "gender": "Male", "institution": "University of Modena and Reggio Emilia", "country": "IT", "position": "Full Professor" } ], "n_formula": 62, "n_formula_1": 16, "n_ref_uni": 47, "n_ref": 82, "n_ref_all": 106, "n_fig": 3, "n_tab": 6, "L_tab": 944, "n_element_tab": 171, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 320, "n_element_tab_1": 52, "formula_len_all": 4469, "formula_len_all_1": 812, "len_all": 152928, "len_all_1": 61127, "len_abs": 1163, "len_title": 131, "len_sents": 39800, "len_sents_1": 28838, "n_sents": 351, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1171, "title": "Closed-Form Merging of Parameter-Efficient Modules for Federated Continual Learning", "abs": "Model merging has emerged as a crucial technique in Deep Learning, enabling the integration of multiple models into a unified system while preserving performance and scalability. In this respect, the compositional properties of low-rank adaptation techniques (e.g., LoRA) have proven beneficial, as simple averaging LoRA modules yields a single model that mostly integrates the capabilities of all individual modules. Building on LoRA, we take a step further by imposing that the merged model matches the responses of all learned modules. Solving this ob-\njective in closed form yields an indeterminate system with A and B as unknown variables, indicating the existence of infinitely many closed-form solutions. To address this challenge, we introduce LoRM, an alternating optimization strategy that trains one LoRA matrix at a time. This allows solving for each unknown variable individually, thus finding a unique solution. We apply our proposed methodology to Federated Class-Incremental Learning (FCIL), ensuring alignment of model responses both between clients and across tasks. Our method demonstrates state-of-the-art performance across a range of FCIL scenarios.", "keywords": [ "model merging", "federated continual learning", "federated learning", "continual learning" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "ROYSNn3vvB", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Sixin Zhang", "gender": "Male", "institution": "Universtite Toulouse", "country": "", "position": "Assistant Professor" } ], "n_formula": 161, "n_formula_1": 42, "n_ref_uni": 39, "n_ref": 108, "n_ref_all": 203, "n_fig": 10, "n_tab": 1, "L_tab": 262, "n_element_tab": 45, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18027, "formula_len_all_1": 3965, "len_all": 251903, "len_all_1": 76201, "len_abs": 1012, "len_title": 135, "len_sents": 72771, "len_sents_1": 29674, "n_sents": 741, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1065, "title": "Local convergence of simultaneous min-max algorithms to differential equilibrium on Riemannian manifold", "abs": "We study min-max algorithms to solve zero-sum differential games on\nRiemannian manifold.\nBased on the notions of\ndifferential Stackelberg equilibrium\nand differential Nash equilibrium on Riemannian manifold,\nwe analyze the local convergence of \ntwo representative deterministic simultaneous algorithms $\\tau$-GDA and $\\tau$-SGA\nto such equilibria.\nSufficient conditions are obtained to establish the linear convergence rate\nof $\\tau$-GDA based on the Ostrowski theorem on manifold and spectral analysis. \nTo avoid strong rotational dynamics in $\\tau$-GDA, \n$\\tau$-SGA is extended from\nthe symplectic gradient-adjustment method in Euclidean space.\nWe analyze an \nasymptotic approximation of $\\tau$-SGA \nwhen the learning rate ratio $\\tau$ is big. \nIn some cases, it can achieve a faster convergence rate \nto differential Stackelberg equilibrium compared to $\\tau$-GDA.\nWe show numerically how the insights obtained from the\nconvergence analysis may improve\nthe training of orthogonal Wasserstein GANs using \nstochastic $\\tau$-GDA and $\\tau$-SGA on simple benchmarks.", "keywords": [ "min-max algorithm", "differential game", "Riemannian manifold", "Wasserstein GAN" ], "rating_list": [ 6, 1, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 4, 3 ], "presentation_list": [ 3, 2, 3, 4, 2 ], "contribution_list": [ 3, 1, 3, 3, 3 ], "confidence_list": [ 3, 1, 2, 3, 3 ] }, { "paper_id": "RMaRBE9s2H", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chun Gu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Junzhe Jiang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Li Zhang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Yurui Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" } ], "n_formula": 23, "n_formula_1": 26, "n_ref_uni": 39, "n_ref": 85, "n_ref_all": 98, "n_fig": 8, "n_tab": 5, "L_tab": 3807, "n_element_tab": 259, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2887, "n_element_tab_1": 194, "formula_len_all": 1044, "formula_len_all_1": 1046, "len_all": 114004, "len_all_1": 57625, "len_abs": 449, "len_title": 131, "len_sents": 27504, "len_sents_1": 23552, "n_sents": 191, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1307, "title": "GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting", "abs": "LiDAR novel view synthesis (NVS) has emerged as a novel task within LiDAR simulation, offering valuable simulated point cloud data from novel viewpoints to aid in autonomous driving systems. However, existing LiDAR NVS methods typically rely on neural radiance fields (NeRF) as their 3D representation, which incurs significant computational costs in both training and rendering. Moreover, NeRF and its variants are designed for symmetrical scenes, making them ill-suited for driving scenarios. To address these challenges, we propose GS-LiDAR, a novel framework for generating realistic LiDAR point clouds with panoramic Gaussian splatting. Our approach employs 2D Gaussian primitives with periodic vibration properties, allowing for precise geometric reconstruction of both static and dynamic elements in driving scenarios. We further introduce a novel panoramic rendering technique with explicit ray-splat intersection, guided by panoramic LiDAR supervision. By incorporating intensity and ray-drop spherical harmonic (SH) coefficients into the Gaussian primitives, we enhance the realism of the rendered point clouds. Extensive experiments on KITTI-360 and nuScenes demonstrate the superiority of our method in terms of quantitative metrics, visual quality, as well as training and rendering efficiency.", "keywords": [ "Gaussian Splatting", "LiDAR Simulation" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "RMaB6cn07S", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Ge Li", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Hao Zhu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jia Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Lecheng Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhi Jin", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 11, "n_ref": 25, "n_ref_all": 39, "n_fig": 9, "n_tab": 2, "L_tab": 108, "n_element_tab": 12, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 206, "n_element_tab_1": 21, "formula_len_all": 600, "formula_len_all_1": 379, "len_all": 87660, "len_all_1": 59685, "len_abs": 1832, "len_title": 123, "len_sents": 31674, "len_sents_1": 25246, "n_sents": 347, "n_sents_1": 282, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1832, "title": "Generating Equivalent Representations of Code By A Self-Reflection Approach", "abs": "Equivalent Representations (ERs) of code are textual representations that preserve the same semantics as the code itself, e.g., natural language comments and pseudocode. ERs play a critical role in software development and maintenance. However, how to automatically generate ERs of code remains an open challenge. In this paper, we propose a self-reflection approach to generating ERs of code. It enables two Large Language Models (LLMs) to work mutually and produce an ER through a reflection process. Depending on whether constraints on ERs are applied, our approach generates ERs in both open and constrained settings. We conduct a empirical study to generate ERs in two settings and obtain eight findings. (1) Generating ERs in the open setting. In the open setting, we allow LLMs to represent code without any constraints, analyzing the resulting ERs and uncovering five key findings. These findings shed light on how LLMs comprehend syntactic structures, APIs, and numerical computations in code.\n(2) Generating ERs in the constrained setting. In the constrained setting, we impose constraints on ERs, such as natural language comments, pseudocode, and flowcharts. This allows our approach to address a range of software engineering tasks. Based on our experiments, we have three findings demonstrating that our approach can effectively generate ERs that adhere to specific constraints, thus supporting various software engineering tasks.\n(3) Future directions. We also discuss potential future research directions, such as deriving intermediate languages for code generation, exploring LLM-friendly requirement descriptions, and further supporting software engineering tasks. We believe that this paper will spark discussions in research communities and inspire many follow-up studies. The source code and data are available.", "keywords": [ "Equivalent Representations of Code", "Large Language Models", "Empirical Study" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "RMBwNzs57N", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Amin Mansouri", "gender": "Not Specified", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "", "position": "MS student" }, { "name": "Martin Josifoski", "gender": "Male", "institution": "Facebook", "country": "GB", "position": "Researcher" }, { "name": "Maxime Peyrard", "gender": "Male", "institution": "CNRS", "country": "FR", "position": "Associate Professor" }, { "name": "Mohammad Hossein Amani", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Robert West", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" }, { "name": "Nicolas Mario Baldwin", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "MS student" } ], "n_formula": 28, "n_formula_1": 22, "n_ref_uni": 34, "n_ref": 49, "n_ref_all": 80, "n_fig": 7, "n_tab": 5, "L_tab": 1986, "n_element_tab": 138, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1706, "formula_len_all_1": 1460, "len_all": 149098, "len_all_1": 52102, "len_abs": 1326, "len_title": 111, "len_sents": 47321, "len_sents_1": 22325, "n_sents": 316, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1212, "title": "Symbolic Autoencoding with Straight-Through Gradient Approximations", "abs": "Self-supervised autoregressive models have achieved significant success across diverse domains, including text, \naudio, and biological sequences. However, these models often rely heavily on large samples of aligned (parallel) \ndata, limiting their applicability in low-resource settings. \nTo address this limitation, we propose Symbolic Autoencoding ($\\Sigma$AE) with Straight-Through Gradient Estimators \n(STGEs)\u2014a latent variable model where the latent space consists of sequences of categorical random variables, \nresembling sentences in an emergent symbolic language. $\\Sigma$AE is trained end-to-end using a family of \nstraight-through gradient estimators.\nIn the unsupervised mode, $\\Sigma$AE learns to compress input data into symbolic sentences and reconstructs the data \nfrom this emergent language. In weakly supervised settings, $\\Sigma$AE further grounds the latent language by \nleveraging supervised training on the small amount of parallel data available.\nWe evaluate $\\Sigma$AE with three well-known quantization mechanisms on four text sequence transduction tasks. \nOur results show that $\\Sigma$AE outperforms baseline methods, particularly in low-resource scenarios with limited \nparallel data.", "keywords": [ "straight-through gradient approximation", "Autoencoders", "discrete representation learning" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 1, 2, 3, 1 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "RLzeoy4FzP", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Faeze Brahman", "gender": "Female", "institution": "Allen Institute for AI", "country": "US", "position": "Postdoc" }, { "name": "Hannaneh Hajishirzi", "gender": "Female", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "senior director" }, { "name": "Lester James Validad Miranda", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Pradeep Dasigi", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Research Scientist" }, { "name": "Sachin Kumar", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Assistant Professor" }, { "name": "Valentina Pyatkin", "gender": "unknown", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Postdoc" }, { "name": "Yanai Elazar", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Postdoc" }, { "name": "Yizhong Wang", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "", "position": "PhD student" }, { "name": "Noah Smith", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 40, "n_ref": 102, "n_ref_all": 135, "n_fig": 8, "n_tab": 11, "L_tab": 2992, "n_element_tab": 56, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 244, "n_element_tab_1": 5, "formula_len_all": 135, "formula_len_all_1": 121, "len_all": 218588, "len_all_1": 57686, "len_abs": 1843, "len_title": 121, "len_sents": 63106, "len_sents_1": 27601, "n_sents": 495, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1851, "title": "Hybrid Preferences: Learning to Route Instances for Human vs. AI Feedback", "abs": "Learning from human feedback has enabled the alignment of language models (LMs) with human preferences. However, directly collecting human preferences can be expensive, time-consuming, and can have high variance. An appealing alternative is to distill preferences from LMs as a source of synthetic annotations as they are more consistent, cheaper, and scale better than human annotation; however, they are also prone to biases and errors. In this work, we introduce a routing framework that combines inputs from humans and LMs to achieve better annotation quality, while reducing the total cost of human annotation. The crux of our approach is to identify preference instances that will benefit from human annotations. We formulate this as an optimization problem: given a preference dataset and an evaluation metric, we train a performance prediction model to predict a reward model's performance on an arbitrary combination of human and LM annotations and employ a routing strategy that selects a combination that maximizes predicted performance. We train the performance prediction model on MultiPref, a new preference dataset with 10K instances paired with human and LM labels. We show that the selected hybrid mixture of LM and direct human preferences using our routing framework achieves better reward model performance compared to using either one exclusively. We simulate selective human preference collection on three other datasets and show that our method generalizes well to all three. We analyze features from the routing model to identify characteristics of instances that can benefit from human feedback, e.g., prompts with a moderate safety concern or moderate intent complexity. We release the dataset, annotation platform, and source code used in this study to foster more efficient and accurate preference collection in the future.", "keywords": [ "Preference Learning", "RLHF", "AI Feedback", "Language Models" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "RLhEGWt94S", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Bo Zou", "gender": "Male", "institution": "Computer Science, Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Chengbin Quan", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Instructor" }, { "name": "Youjian Zhao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yuanfu Wang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yang Chao", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 25, "n_ref": 51, "n_ref_all": 59, "n_fig": 2, "n_tab": 4, "L_tab": 2058, "n_element_tab": 279, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2444, "n_element_tab_1": 283, "formula_len_all": 352, "formula_len_all_1": 352, "len_all": 147625, "len_all_1": 67133, "len_abs": 1835, "len_title": 114, "len_sents": 31934, "len_sents_1": 31139, "n_sents": 190, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1845, "title": "Rethinking Modality Alignment in Multi-Modal Large Language Models", "abs": "Multi-modal Large Language Models (MLLMs) demonstrate remarkable proficiency in addressing a wide range of Vision-Language (VL) tasks. However, most advancements have been focused on adapting to longer sequences containing detailed visual information and scaling up high-quality VL corpus. \nPrevalent VL alignment modules (e.g., the adapter layer in LLaVA and the Q-former in BLIP-2)\nstruggle to align the LLM and visual inputs adequately. \nThey rely on the powerful LLM to decode sub-optimally aligned visual features into the desired formatted word sequences, which can result in hallucinations and reduce the reliability of visual reasoning. Additionally, the LLM's causal attention does not effectively capture the relationship between visual embeddings. To tackle these issues, we rethink the modality alignment in MLLMs and present VL Superior Alignment (VLSA),\na framework designed to decouple the alignment of the LLM with visual inputs. VLSA has two main stages: The perception alignment stage, which consists of innovative compressive high-resolution image encoding and reconstructive training based on Latent Diffusion Models (LDM), reduces the information loss in visual encoding and better models the spatial connection between images' subgraphs. The cognition alignment stage strengthens the LLM in understanding high-level visual semantics and low-level image appearances simultaneously. This advancement is actualized by following the instructions of predicting the codebook indices generated from a Vector Quantized (VQ) encoder and the pixel values within designated areas. Extensive experiments across 20 MLLM benchmarks underscore the consistent improvements brought by VLSA, demonstrating the effectiveness of our methods. In service to the MLLM research community, our code and model checkpoints will be publicly available.", "keywords": [ "Multi-Modal Large Language Model" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "RL7PycCtAO", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Baoyi An", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Bin Chen", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Haoqian Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jinpeng Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yaowei Wang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Yichong Xia", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yimin Zhou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 25, "n_formula_1": 18, "n_ref_uni": 41, "n_ref": 89, "n_ref_all": 130, "n_fig": 22, "n_tab": 3, "L_tab": 1246, "n_element_tab": 126, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 988, "n_element_tab_1": 94, "formula_len_all": 2135, "formula_len_all_1": 917, "len_all": 147669, "len_all_1": 52819, "len_abs": 1445, "len_title": 141, "len_sents": 44345, "len_sents_1": 24045, "n_sents": 328, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1450, "title": "DiffPC: Diffusion-based High Perceptual Fidelity Image Compression with Semantic Refinement", "abs": "Reconstructing high-quality images under low bitrates conditions presents a challenge, and previous methods have made this task feasible by leveraging the priors of diffusion models. However, the effective exploration of pre-trained latent diffusion models and semantic information integration in image compression tasks still needs further study. To address this issue, we introduce Diffusion-based High Perceptual Fidelity Image Compression with Semantic Refinement (DiffPC), a two-stage image compression framework based on stable diffusion. DiffPC efficiently encodes low-level image information, enabling the highly realistic reconstruction of the original image by leveraging high-level semantic features and the prior knowledge inherent in diffusion models. Specifically, DiffPC utilizes a multi-feature compressor to represent crucial low-level information with minimal bitrates and employs pre-embedding to acquire more robust hybrid semantics, thereby providing additional context for the decoding end. Furthermore, we have devised a control module tailored for image compression tasks, ensuring structural and textural consistency in reconstruction even at low bitrates and preventing decoding collapses induced by condition leakage. Extensive experiments demonstrate that our method achieves state-of-the-art perceptual fidelity and surpasses previous perceptual image compression methods by a significant margin in statistical fidelity.", "keywords": [ "lossy image compression", "diffusion model" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 5, 2 ] }, { "paper_id": "RL6R5ryuL5", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Benjamin Lanir", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "MS student" }, { "name": "Jinze Xue", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Engineer" }, { "name": "Mike Chrzanowski", "gender": "unknown", "institution": "NVIDIA", "country": "", "position": "Research Scientist" }, { "name": "Oleg Rybakov", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Enginner" }, { "name": "Peter Dykas", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 17, "n_ref": 57, "n_ref_all": 97, "n_fig": 6, "n_tab": 4, "L_tab": 84, "n_element_tab": 26, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2538, "n_element_tab_1": 56, "formula_len_all": 464, "formula_len_all_1": 464, "len_all": 62162, "len_all_1": 48638, "len_abs": 1481, "len_title": 43, "len_sents": 19381, "len_sents_1": 17894, "n_sents": 147, "n_sents_1": 127, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 1485, "title": "METHODS OF IMPROVING LLM TRAINING STABILITY", "abs": "Training stability of large language models (LLMs) is an important research topic. Reproducing training instabilities can be costly, so we use a small language model with 830M parameters and experiment with higher learning rates to force models to diverge, as in Wortsman et al. (2024). One of the sources of training instability is the growth of logits in attention layers Dehghani et al. (2023). We extend the focus of the previous work [Dehghani et al. (2023),Wortsman et al. (2024)] and look not only at the magnitude of the logits but at all outputs of linear layers in the Transformer block. We observe that with a high learning rate the L2 norm of all linear layer outputs grow with each training step and the model diverges. Specifically we observe that QKV, Proj and FC2 layers have the largest growth of the output magnitude. This prompts us to explore several options: 1) apply layer normalization not only after QK layers (as it is done in [Dehghani et al. (2023), Wortsman et al. (2024)]) but after Proj and FC2 layers too; 2) apply layer normalization after the QKV layer (and remove pre normalization). 3) apply QK layer normalization together with softmax capping. We show that with the last two methods we can increase learning rate by 1.5x (without model divergence) in comparison to an approach based on QK layer normalization only Dehghani et al. (2023). Also we observe significant perplexity improvements for all three methods in comparison to the baseline model.", "keywords": [ "LLM", "training stability" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "RKXcTwWqVa", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Hajin Lee", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "MS student" }, { "name": "Jisu Kim", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Assistant Professor" }, { "name": "Kwangho Kim", "gender": "unknown", "institution": "Korea University", "country": "", "position": "Assistant Professor" } ], "n_formula": 80, "n_formula_1": 19, "n_ref_uni": 36, "n_ref": 68, "n_ref_all": 86, "n_fig": 8, "n_tab": 8, "L_tab": 1689, "n_element_tab": 156, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 267, "n_element_tab_1": 38, "formula_len_all": 7206, "formula_len_all_1": 1039, "len_all": 184652, "len_all_1": 67323, "len_abs": 1600, "len_title": 92, "len_sents": 61577, "len_sents_1": 31377, "n_sents": 547, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1606, "title": "ECLayr: Fast and Robust Topological Layer based on Differentiable Euler Characteristic Curve", "abs": "In the realm of Topological Data Analysis, persistent homology has traditionally served as a primary tool for extracting topological features. However, approaches relying on persistent homology often encounter practical challenges due to their high computational costs. To address this issue, we propose a computationally efficient novel topological layer tailored for general deep learning architectures, leveraging the Euler Characteristic Curve (ECC). Unlike methods based on persistent homology, ECC offers computational advantages by circumventing the need for persistent homology calculation, while still allowing access to crucial information about the underlying topological structure. The proposed layer can readily adapt to diverse data modalities by allowing appropriate filtration according to the user's preference, enabling its application across various learning problems without data preprocessing. We present a novel technique for stable backpropagation that effectively mitigates the vanishing gradient problems commonly encountered in existing methods, allowing for seamless integration of our layer into deep learning models. We go on to present stability analysis, showing that the proposed layer is robust against noise and outliers. We apply our method to topological autoencoders, showing that the standard loss function can effectively regularize topological structures of the latent space. Through classification experiments across various datasets, we illustrate the benefits of our approach in mitigating information loss under conditions of data scarcity or data contamination.", "keywords": [ "Topological Data Analysis", "Deep Learning", "Euler Characteristic Curve" ], "rating_list": [ 5, 5, 6, 5, 5 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3, 3 ] }, { "paper_id": "RKQcJ1lXNT", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Abdulrahman Diaa", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Nils Lukas", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Toluwani Aremu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 32, "n_ref": 83, "n_ref_all": 109, "n_fig": 7, "n_tab": 4, "L_tab": 529, "n_element_tab": 66, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 230, "n_element_tab_1": 22, "formula_len_all": 1111, "formula_len_all_1": 725, "len_all": 163972, "len_all_1": 63106, "len_abs": 1187, "len_title": 74, "len_sents": 57791, "len_sents_1": 29366, "n_sents": 457, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1207, "title": "Optimizing Adaptive Attacks against Content Watermarks for Language Models", "abs": "Large Language Models (LLMs) can be \\emph{misused} to spread online spam and misinformation. Content watermarking deters misuse by hiding a message in model-generated outputs, enabling their detection using a secret watermarking key. Robustness is a core security property, stating that evading detection requires (significant) degradation of the content's quality. Many LLM watermarking methods have been proposed, but robustness is tested only against non-adaptive attackers who lack knowledge of the watermarking method and can find only suboptimal attacks. We formulate the robustness of LLM watermarking as an objective function and use preference-based optimization to tune \\emph{adaptive} attacks against the specific watermarking method. Our evaluation shows that (i) adaptive attacks substantially outperform non-adaptive baselines. (ii) Even in a non-adaptive setting, adaptive attacks optimized against a few known watermarks remain highly effective when tested against other unseen watermarks, and (iii) optimization-based attacks are practical and need limited computational resources of less than seven GPU hours. Our findings underscore the need to test robustness against adaptive attackers.", "keywords": [ "watermarking", "language models", "robustness", "adaptive attacks" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 2, 5, 4 ] }, { "paper_id": "RKOAU5ti1y", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Bo Yue", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Guiliang Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Hongyuan Zha", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Sheng Xu", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 19, "n_ref_uni": 65, "n_ref": 122, "n_ref_all": 160, "n_fig": 5, "n_tab": 3, "L_tab": 3010, "n_element_tab": 128, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3631, "formula_len_all_1": 1724, "len_all": 214585, "len_all_1": 73927, "len_abs": 1505, "len_title": 144, "len_sents": 63076, "len_sents_1": 33979, "n_sents": 462, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1510, "title": "A Distributional Approach to Uncertainty-Aware Preference Alignment Using Offline Demonstrations", "abs": "Designing reward functions in Reinforcement Learning (RL) often demands significant task-specific expertise. Offline preference-based Reinforcement Learning (PbRL) provides an effective alternative to address the complexity of reward design by learning policies from offline datasets that contain human preferences between trajectory pairs. Existing offline PbRL studies typically model a reward function by maximizing its likelihood of generating the observed human preferences. However, due to the varying number of samples within the limited dataset, less frequently compared trajectories exhibit greater uncertainty, which potentially leads to unrelible behaviors during reward and policy updates. To solve this issue, in this work, we introduce Uncertainty-Aware PbRL (UA-PbRL) to learn a distributional reward model and a risk-sensitive policy from an offline preference dataset. Our approach employs a Maximum A Posteriori (MAP) objective to update trajectory rewards and incorporates an informative prior to account for the uncertainties. Building upon this reward update, we propose a generative reward model to capture the reward distribution, utilizing the offline distributional Bellman operator and the Conditional Value-at-Risk (CVaR) metric to train a risk-sensitive policy. Experimental results demonstrate that UA-PbRL effectively identifies and avoids states with high uncertainty, facilitating risk-averse behaviors across various tasks, including robot control and language model alignment.", "keywords": [ "Preference-based Reinforcement Learning", "Distributional Reinforcement Learning", "Uncertainty Awareness" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 4 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "RKB4WiesB4", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Chenjia Bai", "gender": "Male", "institution": "TeleAI, China Telecom", "country": "CN", "position": "Research Scientist" }, { "name": "Jiakun Zheng", "gender": "Male", "institution": "East China University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Kang Xu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Peng Liu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "Professor" }, { "name": "Qiaosheng Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Rushuai Yang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Ting Xiao", "gender": "Female", "institution": "East China University of Science and Technology", "country": "", "position": "Lecturer" } ], "n_formula": 53, "n_formula_1": 15, "n_ref_uni": 58, "n_ref": 109, "n_ref_all": 160, "n_fig": 13, "n_tab": 4, "L_tab": 2987, "n_element_tab": 489, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 1445, "n_element_tab_1": 292, "formula_len_all": 5541, "formula_len_all_1": 981, "len_all": 209300, "len_all_1": 68761, "len_abs": 2005, "len_title": 121, "len_sents": 60455, "len_sents_1": 30932, "n_sents": 451, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1194, "title": "Unsupervised Reinforcement Learning by Maximizing Skill Density Deviation", "abs": "Unsupervised Reinforcement Learning (RL) aims to discover diverse behaviors that can accelerate the learning of downstream tasks. Previous methods typically focus on entropy-based exploration or empowerment-driven skill learning. However, entropy-based exploration struggles in large-scale state spaces (e.g., images), and empowerment-based methods with Mutual Information (MI) estimations have limitations in state exploration. To address these challenges, we propose a novel skill discovery objective that maximizes the deviation of the state density of one skill from the explored regions of other skills, encouraging inter-skill state diversity similar to the initial MI objective. For state-density estimation, we construct a novel conditional autoencoder with soft modularization for different skill policies in high-dimensional space. To incentivize intra-skill exploration, we formulate an intrinsic reward based on the learned autoencoder that resembles count-based exploration in a compact latent space. Through extensive experiments in challenging state and image-based tasks, we find our method learns meaningful skills and achieves superior performance in various downstream tasks.", "keywords": [ "Unsupervised Reinforcement Learning", "Skill Discovery", "Inter-Skill Diversity", "Intra-Skill Exploration" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 2, 5, 4 ] }, { "paper_id": "RJG7fCVkhQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Alois Knoll", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Full Professor" }, { "name": "Jun Liu", "gender": "Male", "institution": "Lancaster University", "country": "GB", "position": "Full Professor" }, { "name": "Mingyu Liu", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Wenqi Ren", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Yuning Cui", "gender": "Male", "institution": "Technical University of Munich", "country": "DE", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 44, "n_ref": 134, "n_ref_all": 169, "n_fig": 10, "n_tab": 8, "L_tab": 2546, "n_element_tab": 210, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 3035, "n_element_tab_1": 219, "formula_len_all": 1170, "formula_len_all_1": 728, "len_all": 182853, "len_all_1": 51674, "len_abs": 1404, "len_title": 101, "len_sents": 33898, "len_sents_1": 19232, "n_sents": 280, "n_sents_1": 151, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1410, "title": "Modumer: Modulating Transformer for Image Restoration", "abs": "Image restoration aims to recover clean images from degraded versions. While Transformer-based approaches have achieved significant advancements in this field, they are limited by high complexity and their inability to capture omni-range dependencies, hindering their overall performance. In this work, we develop Modumer for effective and efficient image restoration by revisiting the Transformer block and Modulation design, which processes input through a convolutional block and projection layers, and fuses features via element-wise multiplication. Specifically, within each unit of Modumer, we integrate the cascaded Modulation design with the downsampled Transformer block to build the attention layers, enabling omni-kernel modulation and mapping inputs into high-dimensional feature spaces. Moreover, we introduce a bioinspired parameter-sharing mechanism to attention layers, which not only enhances efficiency but also improves performance. Additionally, a dual-domain feed-forward network strengthens the representational power of the model. Extensive experiments demonstrate that the proposed Modumer achieves state-of-the-art performance on ten different datasets for five image restoration tasks: image motion deblurring, image deraining, image dehazing, image desnowing, and low-light image enhancement. Furthermore, our model yields promising performance on all-in-one image restoration tasks.", "keywords": [ "Image restoration", "Transformer block", "Modulation design" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "RJBf8k6lxO", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Amnon Shashua", "gender": "Male", "institution": "Hebrew University, Hebrew University of Jerusalem", "country": "", "position": "Full Professor" }, { "name": "Binyamin Rothberg", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "MS student" }, { "name": "Dorin Shteyman", "gender": "Female", "institution": "Hebrew University of Jerusalem", "country": "", "position": "MS student" }, { "name": "Yotam Wolf", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "PhD student" } ], "n_formula": 51, "n_formula_1": 17, "n_ref_uni": 26, "n_ref": 50, "n_ref_all": 62, "n_fig": 6, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3442, "formula_len_all_1": 844, "len_all": 142084, "len_all_1": 62624, "len_abs": 1016, "len_title": 133, "len_sents": 49443, "len_sents_1": 32832, "n_sents": 345, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1021, "title": "Compositional Hardness of Code in Large Language Models - A Probabilistic Perspective", "abs": "A common practice in large language model (LLM) usage for complex analytical tasks such as code generation, is to sample a solution for the entire task within the model's context window. Previous works have shown that subtask decomposition within the model's context (chain of thought), is beneficial for solving such tasks. In this work, we point a limitation of LLMs' ability to perform several sub-tasks within the same context window - an in-context hardness of composition, pointing to an advantage for distributing a decomposed problem in a multi-agent system of LLMs. The hardness of composition is quantified by a generation complexity metric, i.e., the number of LLM generations required to sample at least one correct solution. We find a gap between the generation complexity of solving a compositional problem within the same context relative to distributing it among multiple agents, that increases exponentially with the solution's length. We prove our results theoretically and demonstrate them empirically.", "keywords": [ "Large language models" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 2, 3, 2, 4 ] }, { "paper_id": "RInisw1yin", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Abhishek Gupta", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "Bingjie Tang", "gender": "unknown", "institution": "University of Southern California", "country": "", "position": "PhD student" }, { "name": "Dieter Fox", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Senior Director of Robotics Research" }, { "name": "Iretiayo Akinola", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Yashraj Narang", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Yijie Guo", "gender": "Female", "institution": "University of Michigan", "country": "", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 5, "n_ref_uni": 55, "n_ref": 114, "n_ref_all": 145, "n_fig": 19, "n_tab": 5, "L_tab": 1398, "n_element_tab": 58, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 148, "n_element_tab_1": 9, "formula_len_all": 1972, "formula_len_all_1": 497, "len_all": 223547, "len_all_1": 69134, "len_abs": 1703, "len_title": 111, "len_sents": 79643, "len_sents_1": 35153, "n_sents": 578, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 63, "L_abs": 1713, "title": "SRSA: Skill Retrieval and Adaptation for Robotic Assembly Tasks", "abs": "Enabling robots to learn novel tasks in a data-efficient manner is a long-standing challenge. Common strategies involve carefully leveraging prior experiences, especially transition data collected on related tasks. Although much progress has been made in developing such strategies for general pick-and-place manipulation, far fewer studies have investigated contact-rich assembly tasks, where precise control is essential. In this work, we present SRSA (Skill Retrieval and Skill Adaptation), a novel framework designed to address this problem by utilizing a pre-existing skill library containing policies for diverse assembly tasks. The challenge lies in identifying which skill from the library is most relevant for fine-tuning on a new task. Our key hypothesis is that skills showing higher zero-shot success rates on a new task are better suited for rapid and effective fine-tuning on that task. To this end, we propose to predict the transfer success for all skills in the skill library on a novel task, and then use this prediction to guide the skill retrieval process. Through extensive experiments, we demonstrate that SRSA significantly outperforms the leading baseline, achieving a 22\\% relative improvement in success rate, 3.7x higher stability, and 2.4x greater sample efficiency when retrieving and fine-tuning skills on unseen tasks. Moreover, in a continual learning setup, SRSA efficiently learns policies for new tasks and incorporates them into the skill library, enhancing future policy learning. Additionally, policies trained with SRSA in simulation achieve a 90% mean success rate when deployed in the real world. Please visit our project webpage at https://srsa2024.github.io/ for videos.", "keywords": [ "Robotic Assembly Tasks; Skill Retrieval; Skill Adaptation; Sim-to-real Transfer; Reinforcement Learning Fine-tuning;" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "RIXuX37Hh7", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bowen Yang", "gender": "Male", "institution": "Cohere", "country": "CA", "position": "Researcher" }, { "name": "Huayang Li", "gender": "Male", "institution": "Nara Institute of Science and Technology", "country": "JP", "position": "PhD student" }, { "name": "Pat Verga", "gender": "unknown", "institution": "Cohere", "country": "", "position": "Researcher" }, { "name": "Patrick Lewis", "gender": "Male", "institution": "University College London", "country": "", "position": "PhD student" }, { "name": "Priyanka Sen", "gender": "Female", "institution": "Cohere", "country": "GB", "position": "Researcher" }, { "name": "Taro Watanabe", "gender": "Male", "institution": "Nara Institute of Science and Technology, Japan", "country": "JP", "position": "Full Professor" }, { "name": "Vijay Viswanathan", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Yixuan Su", "gender": "Male", "institution": "Cohere", "country": "", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 53, "n_ref": 99, "n_ref_all": 127, "n_fig": 7, "n_tab": 3, "L_tab": 1295, "n_element_tab": 278, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1295, "n_element_tab_1": 278, "formula_len_all": 265, "formula_len_all_1": 249, "len_all": 158553, "len_all_1": 71985, "len_abs": 1456, "len_title": 124, "len_sents": 36055, "len_sents_1": 30251, "n_sents": 281, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1467, "title": "ALR$^2$: A Retrieve-then-Reason Framework for Long-context Question Answering", "abs": "The context window of large language models (LLMs) has been extended significantly in recent years. However, while the context length that the LLM can process has grown, the capability of the model to accurately reason over that context degrades noticeably. \nThis occurs because modern LLMs often become overwhelmed by the vast amount of information in the context; when answering questions, the model must identify and reason over relevant evidence sparsely distributed throughout the text. To alleviate the challenge of long-context reasoning, we develop a retrieve-then-reason framework, enabling LLMs to reason over relevant evidence collected during an intermediate retrieval step. We find that modern LLMs struggle to accurately retrieve relevant facts and instead, often hallucinate \"*retrieved facts*\", resulting in flawed reasoning and the production of incorrect answers. To address these issues, we introduce ALR$^2$, a method that augments the long-context reasoning capability of LLMs via an explicit two-stage procedure, i.e., aligning LLMs with the objectives of both retrieval and reasoning. We demonstrate the efficacy of ALR$^2$ for mitigating performance degradation in long-context reasoning tasks. Through extensive experiments on long-context QA benchmarks, we find our method to outperform competitive baselines by large margins, achieving at least 8.4 and 7.9 EM gains on the long-context versions of HotpotQA and SQuAD datasets, respectively.", "keywords": [ "Long context", "Retrieval-Augmented Generation", "Question Answering", "Large Language Model" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "RIKIavmwqK", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Arpita Singh", "gender": "Female", "institution": "University of Massachusetts at Amherst", "country": "", "position": "MS student" }, { "name": "Ashish Singh", "gender": "Male", "institution": "Department of Computer Science, University of Massachusetts, Amherst", "country": "US", "position": "PhD student" }, { "name": "Prateek Agarwal", "gender": "Male", "institution": "Department of Computer Science, University of Massachusetts at Amherst", "country": "", "position": "MS student" }, { "name": "Sungchul Kim", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Tong Yu", "gender": "unknown", "institution": "Adobe Research", "country": "US", "position": "Senior Research Scientist" }, { "name": "Victor Bursztyn", "gender": "Male", "institution": "Adobe Systems", "country": "", "position": "Researcher" }, { "name": "Zixuan Huang", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "", "position": "MS student" }, { "name": "Eric Learned-Miller", "gender": "Male", "institution": "Manning College of Information and Computer Sciences, University of Massachusetts, Amherst", "country": "US", "position": "Full Professor" }, { "name": "Nesreen Ahmed", "gender": "Female", "institution": "Intel AI Research", "country": "", "position": "Principal Researcher" }, { "name": "Ryan Rossi", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Senior Research Scientist" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 42, "n_ref": 72, "n_ref_all": 89, "n_fig": 3, "n_tab": 9, "L_tab": 1222, "n_element_tab": 159, "n_fig_1": 2, "n_tab_1": 9, "L_tab_1": 2002, "n_element_tab_1": 181, "formula_len_all": 263, "formula_len_all_1": 310, "len_all": 171389, "len_all_1": 65081, "len_abs": 1358, "len_title": 185, "len_sents": 54389, "len_sents_1": 29261, "n_sents": 420, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1350, "title": "FigCaps-HF: A Figure-to-Caption Generative Framework and Benchmark with Human Feedback", "abs": "Captions are crucial for understanding scientific visualizations and documents. Existing captioning methods for scientific figures rely on figure-caption pairs extracted from documents for training, many of which fall short with respect to metrics like helpfulness, explainability, and visual-descriptiveness leading to generated captions being misaligned with reader preferences. To enable the generation of high-quality figure captions, we introduce \\textbf{FigCaps-HF} a new benchmark and framework for figure-caption generation that can incorporate domain expert feedback in generating captions optimized for reader preferences. Our benchmark framework comprises of 1) an automatic method for evaluating quality of figure-caption pairs, 2) a novel reinforcement learning with human feedback (RLHF) method to optimize a generative figure-to-caption model for reader preferences. We demonstrate the effectiveness of our benchmark by improving performance over standard fine-tuning across different types of models. In particular, when using BLIP as the base model, our RLHF framework achieves a mean gain of 35.7%, 16.9%, and 9% in ROUGE, BLEU, and METEOR, respectively. Finally, we release a large-scale benchmark dataset with human feedback on figure-caption pairs to enable further evaluation and development of RLHF techniques for this problem.", "keywords": [ "Figure Caption Generation", "Figure-Caption Benchmark", "Human Feedback" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "RHISYSlLHf", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Sen Su", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Full Professor" }, { "name": "Shuohuan Wang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Tingfeng Hui", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Yu Sun", "gender": "unknown", "institution": "Baidu", "country": "", "position": "Researcher" }, { "name": "Zhenyu Zhang", "gender": "Male", "institution": "Baidu Inc.", "country": "CN", "position": "Researcher" }, { "name": "hua wu", "gender": "Female", "institution": "Baidu", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 33, "n_ref": 62, "n_ref_all": 79, "n_fig": 6, "n_tab": 9, "L_tab": 9041, "n_element_tab": 1616, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 3898, "n_element_tab_1": 490, "formula_len_all": 260, "formula_len_all_1": 402, "len_all": 174045, "len_all_1": 67507, "len_abs": 2402, "len_title": 131, "len_sents": 41057, "len_sents_1": 28289, "n_sents": 280, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1288, "title": "Upcycling Instruction Tuning from Dense to Mixture-of-Experts via Parameter Merging", "abs": "Mixture-of-Experts (MoE) shines brightly in large language models (LLMs) and demonstrates outstanding performance in plentiful natural language processing tasks. However, existing methods that transform LLMs from dense to MoE face significant data requirements and typically rely on large-scale post-training.\nIn this paper, we propose Upcycling Instruction Tuning (UpIT), a data-efficient approach for tuning a dense pre-trained model into an MoE instruct model.\nSpecifically, we first point out that intermediate checkpoints during instruction tuning of the dense model are naturally suitable for specialized experts, and then propose an expert expansion stage to flexibly achieve models with different numbers of experts, where genetic algorithm and parameter merging are introduced to ensure sufficient diversity of new extended experts.\nTo ensure that each differentiated expert in the MoE model works as expected, we select a small amount of seed data that each expert excels to pre-optimize the router.\nExtensive experiments with various data scales and upcycling settings demonstrate the outstanding performance and data efficiency of UpIT, as well as stable improvement in expert or data scaling. Further analysis reveals the importance of ensuring expert diversity in upcycling.", "keywords": [ "Mixture-of-Experts", "Model upcyling", "Upcycling instruction tuning" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "RG806nMtQr", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Hong Liu", "gender": "Non-Binary", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Huafeng Kuang", "gender": "unknown", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 28, "n_ref": 49, "n_ref_all": 64, "n_fig": 5, "n_tab": 6, "L_tab": 2012, "n_element_tab": 60, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 988, "n_element_tab_1": 46, "formula_len_all": 904, "formula_len_all_1": 708, "len_all": 125139, "len_all_1": 60485, "len_abs": 1035, "len_title": 138, "len_sents": 38047, "len_sents_1": 27501, "n_sents": 312, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1039, "title": "Revisiting Adversarial Examples from the Perspective of Asymptotic Equipartition Property", "abs": "Adversarial examples, which can mislead neural networks through subtle perturbations, continue to challenge our understanding, raising more questions than answers. This paper presents a novel perspective on interpreting adversarial examples through the Asymptotic Equipartition Property (AEP). Our theoretical analysis examines the noise within these examples, revealing that while normal noise aligns with AEP, adversarial noise does not. This insight allows us to classify samples in high-dimensional space as belonging to either the typical or non-typical set, corresponding to normal and adversarial examples, respectively. \nOur analyses and experiments show adversarial examples arise from AEP in high-dimensional space and derive some key properties regarding their quantity, probability, and information capacity. These findings enhance our understanding of adversarial examples and clarify their counterintuitive phenomena, such as adversarial transferability, the trade-off between robustness and accuracy, and robust overfitting.", "keywords": [ "Adversarial examples", "Adversarial robustness", "Asymptotic equipartition property" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 1, 2, 2, 1 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "RFqeoVfLHa", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Pengfei Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Associate Professor" }, { "name": "Xuefeng Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Kelly Ting Wu", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 35, "n_ref": 62, "n_ref_all": 79, "n_fig": 7, "n_tab": 2, "L_tab": 725, "n_element_tab": 92, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1431, "formula_len_all_1": 1161, "len_all": 151243, "len_all_1": 54429, "len_abs": 1309, "len_title": 111, "len_sents": 45277, "len_sents_1": 27124, "n_sents": 300, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1212, "title": "Progress or Regress? Self-Improvement Reversal in Post-training", "abs": "Self-improvement through post-training methods such as iterative preference learning has been acclaimed for enhancing the problem-solving capabilities (e.g., mathematical reasoning) of Large Language Models (LLMs) without human intervention. However, as our exploration deepens, it is crucial to critically assess whether these enhancements indeed signify comprehensive progress or if they could lead to unintended regressions. Through rigorous experimentation and analysis across diverse problem-solving tasks, we uncover nuances in the self-improvement trajectories of LLMs. Our study introduces the concept of \\emph{self-improvement reversal}, where models showing improved overall accuracy metrics might paradoxically exhibit declines in broader, essential capabilities. We propose a comprehensive evaluative framework to scrutinize the underlying mechanisms and outcomes of post-training self-improvement, aiming to discern between superficial metric improvements and genuine enhancements in model functionality. The findings emphasize the complexity of technological advancements in LLMs, underscoring the need for a nuanced understanding of the \\textit{progress or regress} dichotomy in their development.", "keywords": [ "Iterative Self-improvement", "Problem-solving AI" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "RFZV2tOWYN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Chen Chen", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Daochang Liu", "gender": "Male", "institution": "University of Western Australia", "country": "AU", "position": "Lecturer" }, { "name": "Enhuai Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "MS student" }, { "name": "Mubarak Shah", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Chang Xu", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 31, "n_ref": 82, "n_ref_all": 93, "n_fig": 5, "n_tab": 3, "L_tab": 548, "n_element_tab": 66, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 323, "n_element_tab_1": 37, "formula_len_all": 535, "formula_len_all_1": 368, "len_all": 116317, "len_all_1": 58833, "len_abs": 867, "len_title": 100, "len_sents": 37885, "len_sents_1": 28848, "n_sents": 257, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 869, "title": "Investigating Memorization in Video Diffusion Models", "abs": "Diffusion models, widely used for image and video generation, face a significant limitation: the risk of memorizing and reproducing training data during inference, potentially generating unauthorized copyrighted content. While prior research has focused on image diffusion models (IDMs), video diffusion models (VDMs) remain underexplored. To address this, we introduce new metrics specifically designed to separately assess content and motion memorization in VDMs. By applying these metrics, we systematically analyze memorization in various pretrained VDMs, including text-conditional and unconditional models on various datasets, revealing that memorization is widespread across both video and image datasets. Finally, we propose effective detection strategies for both content and motion memorization, offering a foundational approach for improving privacy in VDMs.", "keywords": [ "Video Diffusion Models", "Memorization" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 5, 4, 3, 2 ] }, { "paper_id": "RFMdtKbff5", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Ido Nachum", "gender": "Male", "institution": "University of Haifa", "country": "IL", "position": "Assistant Professor" }, { "name": "Jonathan Shafer", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Michael Gastpar", "gender": "unknown", "institution": "School of Computer and Communication Sciences, EPFL - EPF Lausanne", "country": "", "position": "Full Professor" }, { "name": "Thomas Weinberger", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" } ], "n_formula": 77, "n_formula_1": 26, "n_ref_uni": 24, "n_ref": 48, "n_ref_all": 78, "n_fig": 0, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8356, "formula_len_all_1": 1958, "len_all": 173721, "len_all_1": 68400, "len_abs": 665, "len_title": 50, "len_sents": 54510, "len_sents_1": 28860, "n_sents": 509, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 670, "title": "Which Algorithms Have Tight Generalization Bounds?", "abs": "We study which machine learning algorithms have tight generalization bounds in the overparameterized setting. Our results build on and extend the recent work of Gastpar et al. (2024).\n\nFirst, we present conditions that preclude the existence of tight generalization bounds. Specifically, we show that algorithms that have certain inductive biases that cause them to be unstable do not admit tight generalization bounds. Next, we show that algorithms that are sufficiently stable do have tight generalization bounds. We conclude with a simple characterization that relates the existence of tight generalization bounds to the conditional variance of the algorithm's loss.", "keywords": [ "learning theory", "overparametrization", "neural networks" ], "rating_list": [ 8, 6, 5, 1 ], "soundness_list": [ 4, 4, 3, 2 ], "presentation_list": [ 4, 3, 2, 1 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "REprQnylmC", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Cong Ma", "gender": "Male", "institution": "SenseAuto Research", "country": "CN", "position": "Researcher" }, { "name": "Fudan Yu", "gender": "Male", "institution": "Electronic Engineering, Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Jian Yuan", "gender": "unknown", "institution": "Tsinghua University", "country": "", "position": "Professor" }, { "name": "Qiao Lei", "gender": "Male", "institution": "SenseAuto Research", "country": "", "position": "Researcher" }, { "name": "Tianjian Ouyang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Wei Wu", "gender": "Male", "institution": "Sensetime", "country": "CN", "position": "Researcher" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yuheng Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 6, "n_ref_uni": 46, "n_ref": 97, "n_ref_all": 121, "n_fig": 9, "n_tab": 5, "L_tab": 1166, "n_element_tab": 99, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 546, "n_element_tab_1": 42, "formula_len_all": 1403, "formula_len_all_1": 486, "len_all": 154187, "len_all_1": 60040, "len_abs": 1639, "len_title": 99, "len_sents": 45532, "len_sents_1": 28096, "n_sents": 347, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 51, "L_abs": 1655, "title": "LCSim: A Large-Scale Controllable Traffic Simulator", "abs": "With the rapid growth of urban transportation and the continuous progress in autonomous driving, a demand for robust benchmarking autonomous driving algorithms has emerged, calling for accurate modeling of large-scale urban traffic scenarios with diverse vehicle driving styles. Traditional traffic simulators, such as SUMO, often depend on hand-crafted scenarios and rule-based models, where vehicle actions are limited to speed adjustment and lane changes, making it difficult for them to create realistic traffic environments. In recent years, real-world traffic scenario datasets have been developed alongside advancements in autonomous driving, facilitating the rise of data-driven simulators and learning-based simulation methods. However, current data-driven simulators are often restricted to replicating the traffic scenarios and driving styles within the datasets they rely on, limiting their ability to model multi-style driving behaviors observed in the real world. We propose \\textit{LCSim}, a large-scale controllable traffic simulator. First, we define a unified data format for traffic scenarios and provide tools to construct them from multiple data sources, enabling large-scale traffic simulation. Furthermore, we integrate a diffusion-based vehicle motion planner into LCSim to facilitate realistic and diverse vehicle modeling. Under specific guidance, this allows for the creation of traffic scenarios that reflect various driving styles. Leveraging these features, LCSim can provide large-scale, realistic, and controllable virtual traffic environments. Codes and demos are available at https://anonymous.4open.science/r/LCSim-0C7A.", "keywords": [ "Traffic Simulation", "Diffusion", "Autonomous Driving" ], "rating_list": [ 8, 3, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "RDz3EPC3Lp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Aleksandr Petiushko", "gender": "Male", "institution": "Gatik", "country": "US", "position": "Principal Researcher" }, { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Jiawei Zhang", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Taiqi Wang", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Xuan Yang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yu Yao", "gender": "Male", "institution": "University of Michigan", "country": "", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 22, "n_ref": 62, "n_ref_all": 88, "n_fig": 5, "n_tab": 9, "L_tab": 1532, "n_element_tab": 172, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3116, "n_element_tab_1": 119, "formula_len_all": 397, "formula_len_all_1": 280, "len_all": 144414, "len_all_1": 76751, "len_abs": 2009, "len_title": 136, "len_sents": 53543, "len_sents_1": 38024, "n_sents": 338, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 2024, "title": "SafeAuto: Knowledge-Enhanced Safe Autonomous Driving with Multimodal Foundation Models", "abs": "Traditional autonomous driving systems often struggle to harmonize high-level reasoning with low-level control, leading to suboptimal and even unsafe driving behaviors. The emergence of multimodal large language models (MLLMs), capable of processing visual and textual data, presents an opportunity to unify perception and reasoning tasks within a single framework. However, integrating precise safety knowledge into MLLMs for safe autonomous driving remains a significant challenge.\nTo address this, we propose SafeAuto, a novel framework that enhances MLLM-based autonomous driving systems by incorporating both unstructured and structured knowledge. In particular, we first propose the Place-Dependent Cross-Entropy (PDCE) loss function, which is specifically designed to enhance the accuracy of low-level control signal predictions when treating numerical values as text.\nTo explicitly integrate precise safety knowledge into the MLLM to enable safe autonomous driving, we build a reasoning component for SafeAuto, which first parses driving safety regulations into first-order logic rules (e.g., \"red light $\\implies$ stop\") and then integrates these rules into a probabilistic graphical model, such as a Markov Logic Network (MLN). The environment attributes, identified by attribute recognition models (e.g., detecting a red light), are used to form the predicates in MLN.\nIn addition, the environmental attributes utilized for reasoning are also considered factors in retrieval to construct a Multimodal Retrieval-Augmented Generation (RAG) model, which aims to learn from past similar driving experiences more effectively.\nExtensive experiments demonstrate that SafeAuto significantly outperforms baselines across multiple datasets. By bridging the gap between high-level reasoning and low-level control, SafeAuto paves the way for more accurate, reliable, and safer autonomous driving, facilitating systems that learn effectively from experience, adhere to traffic regulations, and execute precise control actions.", "keywords": [ "Autonomous Driving; Multimodal Large Language Models; Multimodal Retrieval-Augmented Generation; Probabilistic Graph Model" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 1 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "RDVrlWAb7K", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "David Hurych", "gender": "Male", "institution": "Valeo.ai", "country": "CZ", "position": "Researcher" }, { "name": "Eduardo Valle", "gender": "Male", "institution": "Valeo", "country": "FR", "position": "Researcher" }, { "name": "Matthieu Cord", "gender": "Male", "institution": "Sorbonne Universit\u00e9", "country": "FR", "position": "Full Professor" }, { "name": "Mickael Chen", "gender": "Male", "institution": "H.ai", "country": "FR", "position": "Researcher" }, { "name": "Victor Besnier", "gender": "Male", "institution": "Valeo.ai", "country": "CZ", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 53, "n_ref": 91, "n_ref_all": 127, "n_fig": 14, "n_tab": 9, "L_tab": 2505, "n_element_tab": 311, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1524, "n_element_tab_1": 140, "formula_len_all": 614, "formula_len_all_1": 615, "len_all": 158172, "len_all_1": 60905, "len_abs": 116, "len_title": 104, "len_sents": 38942, "len_sents_1": 27362, "n_sents": 310, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1407, "title": "Halton Scheduler for Masked Generative Image Transformer", "abs": "Masked Generative Image Transformers (MaskGIT) have emerged as a scalable and efficient image generation framework, able to deliver high-quality visuals with low inference costs. However, MaskGIT's token unmasking scheduler, an essential component of the framework, has not received the attention it deserves. We analyze the sampling objective in MaskGIT, based on the mutual information between tokens, and elucidate its shortcomings. We then propose a new sampling strategy based on our Halton scheduler instead of the original Confidence scheduler. More precisely, our method selects the token's position according to a quasi-random, low-discrepancy Halton sequence. Intuitively, that method spreads the tokens spatially, progressively covering the image uniformly at each step. Our analysis shows that it allows reducing non-recoverable sampling errors, leading to simpler hyper-parameter tuning and better quality images. Our scheduler does not require retraining or noise injection and may serve as a simple drop-in replacement for the original sampling strategy. Evaluation of both class-to-image synthesis on ImageNet and text-to-image generation on the COCO dataset demonstrates that the Halton scheduler outperforms the Confidence scheduler quantitatively by reducing the FID and qualitatively by generating more diverse and more detailed images. Our code is publicly available on [link redacted].", "keywords": [ "Image Synthesis", "MaskGIT", "Halton Sequence" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "RDLvnUJ5JZ", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Haksoo Lim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jaesik Choi", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" }, { "name": "JAEHOON LEE", "gender": "Male", "institution": "LG AI RESEARCH", "country": "KR", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 16, "n_ref_uni": 29, "n_ref": 77, "n_ref_all": 85, "n_fig": 0, "n_tab": 7, "L_tab": 3568, "n_element_tab": 228, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1183, "n_element_tab_1": 64, "formula_len_all": 1931, "formula_len_all_1": 1639, "len_all": 102753, "len_all_1": 64154, "len_abs": 1074, "len_title": 115, "len_sents": 29844, "len_sents_1": 27545, "n_sents": 205, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1078, "title": "TF-score: Time-series Forecasting using score-based diffusion model", "abs": "Diffusion models have emerged as powerful generative models, capable of synthesizing high-quality images by capturing complex underlying patterns. Building on this success, these models have been adapted for time-series forecasting, a domain characterized by intricate temporal dependencies. However, most existing works have focused primarily on empirical performance without sufficient theoretical exploration. In this paper, we address this gap by introducing a generalized loss function within the diffusion-based forecasting framework. Leveraging this foundation, we introduce TF-score, a score-based diffusion model designed to capture the interdependencies between historical data and future predictions. Extensive experiments across six benchmark datasets show that TF-score consistently surpasses leading baselines, including prior diffusion-based models. Furthermore, we extend existing guidance sampling strategies into a our score-based formulation, achieving performance gains across multiple datasets while providing a detailed analysis of the trade-offs involved.", "keywords": [ "Time-series forecasting", "Diffusion models", "Signal processing" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 1, 2, 1, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "RDFkGZ9Dkh", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Abdelhakim Benechehab", "gender": "Male", "institution": "Eurecom", "country": "", "position": "PhD student" }, { "name": "Ambroise Odonnat", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "FR", "position": "PhD student" }, { "name": "Ievgen Redko", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "FR", "position": "Principal Researcher" }, { "name": "Linus Bleistein", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "PhD student" }, { "name": "Nicolas Boulle", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "Assistant Professor" }, { "name": "Oussama Zekri", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Intern" } ], "n_formula": 198, "n_formula_1": 18, "n_ref_uni": 47, "n_ref": 124, "n_ref_all": 188, "n_fig": 15, "n_tab": 4, "L_tab": 184, "n_element_tab": 8, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 16535, "formula_len_all_1": 1474, "len_all": 322542, "len_all_1": 69685, "len_abs": 4917, "len_title": 86, "len_sents": 95141, "len_sents_1": 29262, "n_sents": 1027, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 38, "L_abs": 1055, "title": "Large Language Models as Markov Chains", "abs": "Large language models (LLMs) have proven to be remarkably efficient, both across a wide range of natural language processing tasks and well beyond them. However, a comprehensive theoretical analysis of the origins of their impressive performance remains elusive. In this paper, we approach this challenging task by drawing an equivalence between generic auto-regressive language models with vocabulary of size $T$ and context window of size $K$ and Markov chains defined on a finite state space of size $\\mathcal{O}(T^K)$. We derive several surprising findings related to the existence of a stationary distribution of Markov chains that capture the inference power of LLMs, their speed of convergence to it, and the influence of the temperature on the latter. We then prove pre-training and in-context generalization bounds and show how the drawn equivalence allows us to enrich their interpretation. Finally, we illustrate our theoretical guarantees with experiments on several recent LLMs to highlight how they capture the behavior observed in practice.", "keywords": [ "Large language models", "Markov chain", "in-context learning", "generalization bounds", "convergence analysis" ], "rating_list": [ 6, 6, 3 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 2, 3, 4 ] }, { "paper_id": "RD9q5vEe1Q", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Changliang Zou", "gender": "Male", "institution": "Nankai University", "country": "", "position": "Full Professor" }, { "name": "Dongjian Hu", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "PhD student" }, { "name": "Junxi Wu", "gender": "unknown", "institution": "Nankai University", "country": "CN", "position": "PhD student" }, { "name": "Shu-Tao Xia", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yajie Bao", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Postdoc" } ], "n_formula": 61, "n_formula_1": 16, "n_ref_uni": 31, "n_ref": 62, "n_ref_all": 99, "n_fig": 11, "n_tab": 13, "L_tab": 6375, "n_element_tab": 1033, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 975, "n_element_tab_1": 104, "formula_len_all": 5371, "formula_len_all_1": 1321, "len_all": 160591, "len_all_1": 56440, "len_abs": 1234, "len_title": 100, "len_sents": 42559, "len_sents_1": 24924, "n_sents": 416, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1240, "title": "Error-quantified Conformal Inference for Time Series", "abs": "Uncertainty quantification in time series prediction is challenging due to the temporal dependence and distribution shift on sequential data. Conformal prediction provides a pivotal and flexible instrument for assessing the uncertainty of machine learning models through prediction sets. Recently, a series of online conformal inference methods updated thresholds of prediction sets by performing online gradient descent on a sequence of quantile loss functions. A drawback of such methods is that they only use the information of revealed non-conformity scores via miscoverage indicators but ignore error quantification, namely the distance between the non-conformity score and the current threshold. To accurately leverage the dynamic of miscoverage error, we propose Error-quantified Conformal Inference (ECI) by smoothing the quantile loss function. ECI introduces a continuous and adaptive feedback scale with the miscoverage error, rather than simple binary feedback in existing methods. We establish a long-term coverage guarantee for ECI under arbitrary dependence and distribution shift. The extensive experimental results show that ECI can achieve valid miscoverage control and output tighter prediction sets than other baselines.", "keywords": [ "Time Series; Uncertainty Quantification; Conformal Prediction; Distribution Shift" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "RCiwz7WqUU", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Chuan Wu", "gender": "unknown", "institution": "The University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Lingpeng Kong", "gender": "Male", "institution": "Department of Computer Science, The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Sheng Wang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Wenhao Lu", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "JUNTAO ZHAO", "gender": "Male", "institution": "University of Hong Kong", "country": "CN", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 30, "n_ref": 78, "n_ref_all": 103, "n_fig": 8, "n_tab": 6, "L_tab": 6207, "n_element_tab": 743, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2694, "n_element_tab_1": 270, "formula_len_all": 119, "formula_len_all_1": 119, "len_all": 145014, "len_all_1": 60313, "len_abs": 1481, "len_title": 116, "len_sents": 35724, "len_sents_1": 26622, "n_sents": 255, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1499, "title": "QSpec: Speculative Decoding with Complementary Quantization Schemes", "abs": "Quantization has been substantially adopted to accelerate inference and reduce memory consumption of large language models (LLMs).\nWhile activation-weight joint quantization speeds up the inference process through low-precision kernels, we demonstrate that it suffers severe performance degradation on multi-step reasoning tasks, rendering it ineffective.\nWe propose a novel quantization paradigm called QSPEC, which seamlessly integrates two complementary quantization schemes for speculative decoding. \nLeveraging nearly cost-free execution switching, QSPEC drafts tokens with low-precision, fast activation-weight quantization, and verifies them with high-precision weight-only quantization, \neffectively combines the strengths of both quantization schemes.\nCompared to high-precision quantization methods, QSPEC empirically boosts token generation throughput by up to $1.80\\times$ without any quality compromise, distinguishing it from other low-precision quantization approaches. \nThis enhancement is also consistent across various serving tasks, model sizes, quantization methods, and batch sizes.\nUnlike existing speculative decoding techniques, our approach reuses weights and the KV cache, avoiding additional memory overhead. Furthermore, QSPEC offers a plug-and-play advantage without requiring any training.\nWe believe that QSPEC demonstrates unique strengths for future deployment of high-fidelity quantization schemes, particularly in memory-constrained scenarios (e.g., edge devices).", "keywords": [ "quantization", "compression", "large language models", "reasoning", "speculative decoding" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "RC5x3OkywQ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Guorui Chen", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Jindong Gu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Xiaojun Jia", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Zhijiang Li", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "\u590f\u4e00\u5e06", "gender": "Male", "institution": "Wuhan University", "country": "", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 5, "n_ref_uni": 52, "n_ref": 120, "n_ref_all": 140, "n_fig": 11, "n_tab": 18, "L_tab": 20658, "n_element_tab": 2087, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 2383, "n_element_tab_1": 210, "formula_len_all": 470, "formula_len_all_1": 264, "len_all": 216780, "len_all_1": 64268, "len_abs": 1102, "len_title": 90, "len_sents": 44525, "len_sents_1": 28116, "n_sents": 362, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 1106, "title": "LLM Jailbreak Detection for (Almost) Free!", "abs": "Large language models (LLMs) enhance security through alignment when widely used, but remain susceptible to jailbreak attacks capable of producing inappropriate content. Jailbreak detection methods show promise in mitigating jailbreak attacks through the assistance of other models or multiple model inferences. However, existing methods entail significant computational costs. In this paper, we present a finding that the difference in output distributions between jailbreak and benign prompts can be employed for detecting jailbreak prompts. Based on this finding, we propose a Free Jailbreak Detection (FJD) method which incorporates manual instructions into the input and scales the logits by temperature to distinguish between jailbreak and benign prompts through the confidence of the first token. Furthermore, we enhance the detection performance of FJD through the integration of virtual instruction learning (FJD-LI). Extensive experiments on aligned large models demonstrated that our FJD outperforms baseline methods in jailbreak detection accuracy with almost no additional computational costs.", "keywords": [ "Jailbreak Attack", "Large Language Model" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "RC5FPYVQaH", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Berk Ustun", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Chung-En Sun", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Tsui-Wei Weng", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Tuomas Oikarinen", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 14, "n_ref": 25, "n_ref_all": 51, "n_fig": 10, "n_tab": 18, "L_tab": 15630, "n_element_tab": 364, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1189, "n_element_tab_1": 153, "formula_len_all": 696, "formula_len_all_1": 676, "len_all": 161378, "len_all_1": 68512, "len_abs": 3187, "len_title": 88, "len_sents": 51250, "len_sents_1": 31034, "n_sents": 494, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 891, "title": "Concept Bottleneck Large Language Models", "abs": "We introduce the Concept Bottleneck Large Language Model (CB-LLM), a pioneering approach to creating inherently interpretable Large Language Models (LLMs). Unlike traditional black-box LLMs that rely on post-hoc interpretation methods with limited neuron function insights, CB-LLM sets a new standard with its built-in interpretability, scalability, and ability to provide clear, accurate explanations. We investigate two essential tasks in the NLP domain: text classification and text generation. In text classification, CB-LLM narrows the performance gap with traditional black-box models and provides clear interpretability. In text generation, we show how interpretable neurons in CB-LLM can be used for concept detection and steering text generation. Our CB-LLMs enable greater interaction between humans and LLMs across a variety of tasks --- a feature notably absent in existing LLMs.", "keywords": [ "LLMs", "Interpretability", "Concept Bottleneck Model" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "RBqvU12SHz", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Andrej Erkelens", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "", "position": "MS student" }, { "name": "Diganta Misra", "gender": "Male", "institution": "Max-Planck-Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Divyansh Singhvi", "gender": "Male", "institution": "IIT Kanpur", "country": "", "position": "Undergrad student" }, { "name": "Naomi Saphra", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Fellow" }, { "name": "Raghav Jain", "gender": "Male", "institution": "Indian Institute of Technology, Patna.", "country": "", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 36, "n_ref": 47, "n_ref_all": 61, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 268, "formula_len_all_1": 285, "len_all": 109250, "len_all_1": 52990, "len_abs": 853, "len_title": 43, "len_sents": 37231, "len_sents_1": 27505, "n_sents": 242, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 855, "title": "Structural Probing with Feature Interaction", "abs": "Measuring nonlinear feature interaction is an established approach to understanding complex patterns of attribution in many models. In this paper, we use Shapley Taylor interaction indices (STII) to analyze the impact of underlying data structure on model representations in a variety of modalities, tasks, and architectures. Considering linguistic structure in masked and auto-regressive language models (MLMs and ALMs), we find that STII increases within idiomatic expressions and that Transformer ALMs scale STII with syntactic distance, just as LSTM-based ALMs do. Our speech model findings reflect the phonetic principal that the openness of the oral cavity determines how much a phoneme's acoustics vary based on context. Our wide range of results illustrates the benefits of interdisciplinary work and domain expertise in interpretability research.", "keywords": [ "Shapley Interactions", "Shapley Taylor interaction indices", "Masked Language Models", "Language Models", "Feature Interaction" ], "rating_list": [ 3, 3, 6, 1 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 4, 3, 1, 4 ] }, { "paper_id": "RBp0x7rkMO", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Marco Cipriano", "gender": "Male", "institution": "Hasso Plattner Institute", "country": "DE", "position": "PhD student" }, { "name": "Moritz Feuerpfeil", "gender": "Male", "institution": "Hasso Plattner Institute", "country": "DE", "position": "MS student" }, { "name": "Gerard Melo", "gender": "Male", "institution": "Hasso Plattner Institute", "country": "DE", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 39, "n_ref": 64, "n_ref_all": 103, "n_fig": 17, "n_tab": 10, "L_tab": 1642, "n_element_tab": 160, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 984, "n_element_tab_1": 219, "formula_len_all": 541, "formula_len_all_1": 233, "len_all": 148612, "len_all_1": 62651, "len_abs": 1132, "len_title": 127, "len_sents": 50503, "len_sents_1": 27791, "n_sents": 391, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1136, "title": "Vector Grimoire: Codebook-based Shape Generation under Raster Image Supervision", "abs": "Scalable Vector Graphics (SVG) is a popular format on the web and in the design industry. \nHowever, despite the great strides made in generative modeling,\nSVG has remained underexplored due to the discrete and complex nature of such\ndata. We introduce GRIMOIRE, a text-guided SVG generative model that is comprised\nof two modules: A Visual Shape Quantizer (VSQ) learns to map raster\nimages onto a discrete codebook by reconstructing them as vector shapes, and\nan Auto-Regressive Transformer (ART) models the joint probability distribution\nover shape tokens, positions and textual descriptions, allowing us to generate vector \ngraphics from natural language. Unlike existing models that require direct\nsupervision from SVG data, GRIMOIRE learns shape image patches using only\nraster image supervision which opens up vector generative modeling to significantly more data. \nWe demonstrate the effectiveness of our method by fitting\nGRIMOIRE for closed filled shapes on the MNIST and for outline strokes on icon and font data,\nsurpassing previous image-supervised methods in generative quality and vector-supervised approach in flexibility.", "keywords": [ "Image Generation", "Scalable Vector Graphics", "VQ-VAE", "Differentiable Rasterizer" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "RBaDiInDRg", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Kai Mei", "gender": "unknown", "institution": "Rutgers University, New Brunswick", "country": "", "position": "PhD student" }, { "name": "Libby Hemphill", "gender": "Female", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "Associate Professor" }, { "name": "Lingyao Li", "gender": "Male", "institution": "University of South Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Lizhou Fan", "gender": "unknown", "institution": "Brigham and Women's Hospital, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Wenyue Hua", "gender": "Female", "institution": "University of California, Santa Barbara", "country": "US", "position": "Postdoc" }, { "name": "Yingqiang Ge", "gender": "Male", "institution": "Rutgers University", "country": "", "position": "PhD student" }, { "name": "Yongfeng Zhang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Associate Professor" }, { "name": "jianchao ji", "gender": "Male", "institution": "Rutgers University", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 50, "n_ref_all": 84, "n_fig": 16, "n_tab": 5, "L_tab": 298, "n_element_tab": 9, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 413, "n_element_tab_1": 28, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 234512, "len_all_1": 62153, "len_abs": 1398, "len_title": 120, "len_sents": 103392, "len_sents_1": 31429, "n_sents": 668, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1391, "title": "War and Peace (WarAgent): LLM-based Multi-Agent Simulation of World Wars", "abs": "This research explores the potential of Artificial Intelligence and Large Language Models in understanding and simulating complex human behaviors, specifically in the context of historical international conflicts. We introduce WarAgent, an LLM-powered multi-agent AI system, to simulate the decisions and consequences of participating countries in three specific historical conflicts. In addition, we propose standard evaluation protocols for LLM-based Multi-agent Systems simulation. Our study provides a nuanced analysis of the strengths and limitations of current MAS systems in simulating complex collective human behaviors under diverse settings of international conflicts. The emergent interactions among agents in our simulations offer fresh perspectives on the triggers and conditions leading to war. Our findings offer data-driven and AI-augmented insights that can help redefine how we approach conflict resolution and peacekeeping strategies. While we acknowledge the potential of AI in providing data-driven insights, we caution against over-reliance and emphasize the need for careful interpretation in conflict resolution and peacekeeping strategies. The implications of this work extend beyond computer simulation, offering a potential avenue for using AI to better understand human history. Code and data are available at \\url{https://anonymous.4open.science/r/WarAgent-0FF0}", "keywords": [ "large language model", "multi-agent system", "social simulation" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "RBL3Gm5ygj", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Aoran Xiao", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Postdoc" }, { "name": "Heli Qi", "gender": "Male", "institution": "Nara Institute of Science and Technology, Japan", "country": "JP", "position": "MS student" }, { "name": "Naoto Yokoya", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Associate Professor" }, { "name": "Shijian Lu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Weihao Xuan", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Yun Xing", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 44, "n_ref": 97, "n_ref_all": 129, "n_fig": 6, "n_tab": 6, "L_tab": 2327, "n_element_tab": 263, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 969, "n_element_tab_1": 89, "formula_len_all": 135, "formula_len_all_1": 135, "len_all": 176760, "len_all_1": 59127, "len_abs": 1479, "len_title": 89, "len_sents": 48924, "len_sents_1": 27950, "n_sents": 373, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1486, "title": "Segment Anything with Multiple Modalities", "abs": "Robust and accurate segmentation of scenes has become one core functionality in various visual recognition and navigation tasks. This has inspired the recent development of Segment Anything Model (SAM), a foundation model for general mask segmentation. However, SAM is largely tailored for single-modal RGB images, limiting its applicability to multi-modal data captured with widely-adopted sensor suites, such as LiDAR plus RGB, depth plus RGB, thermal plus RGB, etc. We develop MM-SAM, an extension and expansion of SAM that supports cross-modal and multi-modal processing for robust and enhanced segmentation with different sensor suites. MM-SAM features two key designs, namely, unsupervised cross-modal transfer and weakly-supervised multi-modal fusion, enabling label-free and parameter-efficient adaptation toward various sensor modalities. It addresses three main challenges: 1) adaptation toward diverse non-RGB sensors for single-modal processing, 2) synergistic processing of multi-modal data via sensor fusion, and 3) mask-free training for different downstream tasks. Notably, we demonstrate that the output latent space of SAM's RGB image encoder can function as a highly abstract, shareable embedding space compatible with segmentation across different sensor modalities. Extensive experiments show that MM-SAM consistently outperforms SAM by large margins, demonstrating its effectiveness and robustness across various sensors and data modalities. Code will be released.", "keywords": [ "segment anything", "multimodal fusion", "foundation model", "parameter-efficient tuning", "domain adaptation", "unsupervised learning", "weakly supervised learning" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "RAyRXQjsFl", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Bruno Lepri", "gender": "Male", "institution": "Fondazione Bruno Kessler", "country": "IT", "position": "Principal Researcher" }, { "name": "Gabriele Santin", "gender": "Male", "institution": "University of Venice", "country": "IT", "position": "Assistant Professor" }, { "name": "Marco Pacini", "gender": "Male", "institution": "Fondazione Bruno Kessler", "country": "IT", "position": "PhD student" }, { "name": "Xiaowen Dong", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" } ], "n_formula": 138, "n_formula_1": 32, "n_ref_uni": 34, "n_ref": 83, "n_ref_all": 95, "n_fig": 0, "n_tab": 1, "L_tab": 287, "n_element_tab": 21, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 13350, "formula_len_all_1": 3678, "len_all": 249801, "len_all_1": 78000, "len_abs": 1989, "len_title": 95, "len_sents": 70675, "len_sents_1": 32114, "n_sents": 756, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1276, "title": "Separation Power of Equivariant Neural Networks", "abs": "The separation power of a machine learning model refers to its ability to distinguish between different inputs and is often used as a proxy for its expressivity. Indeed, knowing the separation power of a family of models is a necessary condition to obtain fine-grained universality results. In this paper, we analyze the separation power of equivariant neural networks, such as convolutional and permutation-invariant networks.\nWe first present a complete characterization of inputs indistinguishable by models derived by a given architecture. From this results, we derive how separability is influenced by hyperparameters and architectural choices\u2014such as activation functions, depth, hidden layer width, and representation types. Notably, all non-polynomial activations, including ReLU and sigmoid, are equivalent in expressivity and reach maximum separation power. Depth improves separation power up to a threshold, after which further increases have no effect. Adding invariant features to hidden representations does not impact separation power. Finally, block decomposition of hidden representations affects separability, with minimal components forming a hierarchy in separation power that provides a straightforward method for comparing the separation power of models.", "keywords": [ "Geometric Deep Learning", "Theory for Equivariant Neural Networks", "Expressive Power" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "RAdBtquPiI", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Alexander Julian Mattick", "gender": "Male", "institution": "Fraunhofer IIS", "country": "", "position": "Intern" }, { "name": "Christopher Mutschler", "gender": "Male", "institution": "Fraunhofer IIS", "country": "DE", "position": "Principal Researcher" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 25, "n_ref": 40, "n_ref_all": 61, "n_fig": 3, "n_tab": 1, "L_tab": 27, "n_element_tab": 1, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 27, "n_element_tab_1": 1, "formula_len_all": 876, "formula_len_all_1": 876, "len_all": 95606, "len_all_1": 64631, "len_abs": 965, "len_title": 121, "len_sents": 32733, "len_sents_1": 32241, "n_sents": 222, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 968, "title": "Provably safe Reinforcement Learning using Bender's Decomposition Oracles", "abs": "One of the core challenges when applying reinforcement learning to solve real world problems is the violation of numerous safety, feasibility or physical constraints during training and deployment.\nWe propose Bender's Oracle Optimization (BOO) that manages to achieve provable safety during both training and deployment, under the assumption that one has access to a representation of the feasible set, e.g., through a (possibly inaccurate) simulator or encoded rules. \nThis method is particularly useful for cases where a simple (deterministic) model of the problem is available, but said model is too inaccurate or incomplete to solve the problem directly.\nWe showcase our method by applying it to a challenging reward-maximizing stochastic job-shop scheduling problem, where we demonstrate a 17\\% improvement, and a nonlinear, nonconvex packing problem where we achieve close to globally optimal performance while improving the convergence speed by a factor of 800.", "keywords": [ "Constrained Reinforcement Learning", "Safe Reinforcement learning", "Constrained Optimization" ], "rating_list": [ 3, 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 1, 3, 1 ], "presentation_list": [ 3, 3, 2, 2, 1 ], "contribution_list": [ 2, 2, 3, 3, 1 ], "confidence_list": [ 4, 3, 4, 3, 2 ] }, { "paper_id": "RAX45dcfA2", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Farzeen Munir", "gender": "Female", "institution": "Aalto University", "country": "FI", "position": "Postdoc" }, { "name": "Shoaib Azam", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Postdoc" }, { "name": "Tomasz Piotr Kucner", "gender": "Male", "institution": "Aalto University", "country": "", "position": "Assistant Professor" }, { "name": "Tsvetomila Mihaylova", "gender": "Female", "institution": "Aalto University", "country": "", "position": "Postdoctoral Researcher" }, { "name": "Ville Kyrki", "gender": "unknown", "institution": "Aalto University", "country": "FI", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 48, "n_ref": 85, "n_ref_all": 107, "n_fig": 9, "n_tab": 2, "L_tab": 1207, "n_element_tab": 209, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1207, "n_element_tab_1": 209, "formula_len_all": 207, "formula_len_all_1": 299, "len_all": 125584, "len_all_1": 64439, "len_abs": 1235, "len_title": 114, "len_sents": 36671, "len_sents_1": 30423, "n_sents": 228, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1254, "title": "PEDVLM: PEDESTRIAN VISION LANGUAGE MODEL FOR INTENTIONS PREDICTION", "abs": "Effective modeling of human behavior is crucial for the safe and reliable coexistence of humans and autonomous vehicles. Traditional deep learning methods have limitations in capturing the complexities of pedestrian behavior, often relying on simplistic representations or indirect inference from visual cues, which hinders their explainability. To address this gap, we introduce $\\textbf{PedVLM}$, a vision-language model that leverages multiple modalities (RGB images, optical flow, and text) to predict pedestrian intentions and also provide explainability for pedestrian behavior. PedVLM comprises a CLIP-based vision encoder and a text-to-text transfer transformer (T5) language model, which together extract and combine visual and text embeddings to predict pedestrian actions and enhance explainability. Furthermore, to complement our PedVLM model and further facilitate research, we also publicly release the corresponding dataset, PedPrompt, which includes the prompts in the Question-Answer (QA) template for pedestrian intention prediction. PedVLM is evaluated on PedPrompt, JAAD, and PIE datasets demonstrates its efficacy compared to state-of-the-art methods. The dataset and code will be made available at {https://github.com/abc/ped_VLM}.", "keywords": [ "Pedestrian Intention Prediction", "Vision Language Model", "Autonomous Vehicle" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "RAC3ng3TSN", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "M. Paul Laiu", "gender": "unknown", "institution": "Oak Ridge National Laboratory", "country": "US", "position": "Researcher" }, { "name": "Steffen Schotth\u00f6fer", "gender": "Male", "institution": "Oak Ridge National Laboratory", "country": "", "position": "Researcher" } ], "n_formula": 82, "n_formula_1": 29, "n_ref_uni": 34, "n_ref": 91, "n_ref_all": 130, "n_fig": 11, "n_tab": 10, "L_tab": 4557, "n_element_tab": 358, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1508, "n_element_tab_1": 143, "formula_len_all": 10328, "formula_len_all_1": 2768, "len_all": 214261, "len_all_1": 69917, "len_abs": 1151, "len_title": 125, "len_sents": 57636, "len_sents_1": 27377, "n_sents": 459, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1158, "title": "Federated Dynamical Low-Rank Training with Global Loss Convergence Guarantees", "abs": "We propose a federated dynamical low-rank training (FeDLRT) scheme to reduce client compute and communication costs - two significant performance bottlenecks in horizontal federated learning. Our method builds upon dynamical low-rank splitting schemes for manifold-constrained optimization to create a global low-rank basis of network weights, which enables client training on a small coefficient matrix. This global low-rank basis that allows us to incorporate a variance correction scheme and prove global loss descent and convergence to a stationary point. FeDLRT features dynamic augmentation and truncation of the low-rank bases to optimize computing and communication resource utilization. Notably FeDLRT only trains a small coefficient matrix per client. We demonstrate the efficiency of FeDLRT in an array of computer vision benchmarks with both i.i.d. and non-i.i.d. data distributions and show a reduction of client compute and communication costs by up to an order of magnitude with minimal impacts on global accuracy.\nFeDLRT performs as well as classical methods such as FedAvg and FedLin, with a fraction of the memory and compute requirements.", "keywords": [ "Federated Learning", "Low-Rank", "Model Compression", "Efficient Federated Learning" ], "rating_list": [ 5, 5, 8, 5, 5 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 2, 3, 4 ] }, { "paper_id": "R9lgWYE508", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Honggyu An", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jaewoo Jung", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jisang Han", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jiwon Kang", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Undergrad student" }, { "name": "Seonghoon Park", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "MS student" }, { "name": "Seungryong Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 7, "n_ref_uni": 46, "n_ref": 107, "n_ref_all": 154, "n_fig": 15, "n_tab": 13, "L_tab": 8405, "n_element_tab": 1591, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 4066, "n_element_tab_1": 785, "formula_len_all": 767, "formula_len_all_1": 252, "len_all": 195682, "len_all_1": 74856, "len_abs": 1340, "len_title": 117, "len_sents": 55290, "len_sents_1": 29238, "n_sents": 365, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1344, "title": "Relaxing Accurate Initialization Constraint for 3D Gaussian Splatting", "abs": "In this work, we investigate the limitations of the 3D Gaussian Splatting (3DGS) optimization scheme, revealing why it undergoes significant performance drops when initialized with noisy or random point clouds. Through in-depth analysis, we identify a key limitation of the 3DGS optimization: limited Gaussian transportability. Since Gaussians are optimized solely based on image photometric loss, the optimization tends to overfit the parameters of the projected Gaussians to improve reconstruction at their current positions, rather than relocating them to more optimal locations. This leads to producing under-reconstructed regions when starting with noisy or random initialization, failing to transport Gaussians to correct locations. Based on our findings, we propose RAIN-GS (Relaxing Accurate Initialization Constraint for 3D Gaussian Splatting), a set of simple yet effective modifications, including initializing sparse Gaussians with large variances, progressive Gaussian low-pass filtering, and an Adaptive Bound-Expanding split algorithm. These modifications enable Gaussians to effectively redistribute across the scene, capturing both coarse structure and fine details. By addressing the inherent limitations of 3DGS, RAIN-GS allows effective training even with random point clouds, significantly enhancing reconstruction quality.", "keywords": [ "Novel View Synthesis", "3D Gaussian Splatting" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "R9feGbYRG7", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "A. Carolina Filipe", "gender": "unknown", "institution": "Champalimaud Foundation", "country": "PT", "position": "Researcher" }, { "name": "Il Memming Park", "gender": "Male", "institution": "Champalimaud Centre for the Unknown", "country": "PT", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 24, "n_ref": 41, "n_ref_all": 53, "n_fig": 6, "n_tab": 2, "L_tab": 327, "n_element_tab": 24, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 327, "n_element_tab_1": 24, "formula_len_all": 625, "formula_len_all_1": 642, "len_all": 112090, "len_all_1": 58014, "len_abs": 1171, "len_title": 155, "len_sents": 32538, "len_sents_1": 28812, "n_sents": 219, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1176, "title": "One Model to Train Them All: A Unified Diffusion Framework for Multi-Context Neural Population Forecasting", "abs": "Recent research has revealed shared neural patterns among animals performing similar tasks and within individual animals across different tasks. This has led to a growing interest in replacing single-session latent variable models with a unified model that allows us to align recordings across different animals, sessions, and tasks, despite the challenge of distinct neuron identities in each recording. \nIn this work, we present a conditioned diffusion framework to model population dynamics of neural activity across multiple contexts. The quality of the learned dynamics is evaluated through the model's forecasting ability, which predicts multiple timesteps of both neural activity and behavior.\nAdditionally, we introduce a benchmark dataset spanning six electrophysiology datasets, seven tasks, 19 animals, and 261 sessions, providing a standardized framework for multi-task neural population models.\nOur results demonstrate that the pretrained model can be efficiently adapted to novel, unseen sessions without requiring explicit neuron correspondence. This enables few-shot learning with minimal labeled data, as well as competitive performance in zero-shot learning.", "keywords": [ "neural population", "diffusion model", "time series forecasting", "sequence-to-sequence", "electrophysiology", "neural dynamics" ], "rating_list": [ 6, 6, 3, 5, 3 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 2, 1 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "R9W6fFlr8W", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Hok Shing Wong", "gender": "Male", "institution": "University of Bath", "country": "", "position": "Postdoc" }, { "name": "Matthias J Ehrhardt", "gender": "male", "institution": "University of Bath", "country": "GB", "position": "Associate Professor" }, { "name": "Subhadip Mukherjee", "gender": "Male", "institution": "Indian Institute of Technology Kharagpur", "country": "IN", "position": "Assistant Professor" } ], "n_formula": 59, "n_formula_1": 27, "n_ref_uni": 33, "n_ref": 52, "n_ref_all": 66, "n_fig": 9, "n_tab": 1, "L_tab": 93, "n_element_tab": 10, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 93, "n_element_tab_1": 10, "formula_len_all": 5730, "formula_len_all_1": 3109, "len_all": 114207, "len_all_1": 51240, "len_abs": 1045, "len_title": 143, "len_sents": 30691, "len_sents_1": 20700, "n_sents": 337, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1050, "title": "A primal-dual algorithm for variational image reconstruction with learned convex regularizers", "abs": "We address the optimization problem in a data-driven variational reconstruction framework, where the regularizer is parameterized by an input-convex neural network (ICNN). While gradient-based methods are commonly used to solve such problems, they struggle to effectively handle non-smoothness which often leads to slow convergence. Moreover, the nested structure of the neural network complicates the application of standard non-smooth optimization techniques, such as proximal algorithms. To overcome these challenges, we reformulate the problem and eliminate the network's nested structure. By relating this reformulation to epigraphical projections of the activation functions, we transform the problem into a convex optimization problem that can be efficiently solved using a primal-dual algorithm. We also prove that this reformulation is equivalent to the original variational problem. Through experiments on several imaging tasks, we demonstrate that the proposed approach outperforms subgradient methods in terms of both speed and stability.", "keywords": [ "variational problem+learned convex regularizer+convex optimization+primal-dual algorithm" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 4, 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 5, 4 ] }, { "paper_id": "R9OHszNtpA", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Anurag Sarkar", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Postdoc" }, { "name": "Jordan T. Ash", "gender": "unknown", "institution": "Microsoft Research", "country": "US", "position": "Researcher" }, { "name": "Lucas Caccia", "gender": "Male", "institution": "Microsoft", "country": "CA", "position": "Researcher" }, { "name": "Nabil Omi", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Siddhartha Sen", "gender": "unknown", "institution": "Research, Microsoft", "country": "US", "position": "Principal Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 46, "n_ref": 72, "n_ref_all": 92, "n_fig": 11, "n_tab": 2, "L_tab": 403, "n_element_tab": 60, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 881, "n_element_tab_1": 66, "formula_len_all": 174, "formula_len_all_1": 193, "len_all": 157398, "len_all_1": 72607, "len_abs": 1367, "len_title": 99, "len_sents": 51210, "len_sents_1": 36469, "n_sents": 367, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1816, "title": "Generative Modeling of Individual Behavior at Scale", "abs": "Recent years have seen a growing interest in using AI to model human behavior, particularly in domains where humans learn from or collaborate with this technology. While most existing work attempts to model human behavior at an aggregate level, our goal is to model behavior at the individual level. Recent work in the domain of chess has shown that behavioral stylometry, or the task of identifying a person from their actions alone, can be achieved with high accuracy among a pool of a few thousand players. However, this approach cannot generate actions in the style of each player, and hence cannot reason about or influence player behavior in practice. We provide a new perspective on behavioral stylometry that addresses these limitations, by drawing a connection to the vast literature of transfer learning in NLP. Specifically, by casting the stylometry problem as a multi-task learning problem---where each task represents a distinct---we show that parameter-efficient fine-tuning (PEFT) methods can be adapted to model individual behavior in an explicit and generative manner, at unprecedented scale. We apply our approach at scale to two very different games: chess (47,864 players) and Rocket League (2,000 players).\n\nOur approach leverages recent modular PEFT methods to learn a shared set of skill parameters that can be combined in different ways via style vectors. Style vectors enable two important capabilities. First, they are generative: we can generate actions in the style of a player simply by conditioning on the player's style vector. Second, they induce a latent style space that we can interpret and manipulate algorithmically. This allows us to compare different player styles, as well as synthesize new (human-like) styles, e.g. by interpolating between the style vectors of two players.", "keywords": [ "style", "parameter efficient fine-tuning", "peft", "chess", "stylometry", "playstyle", "representation learning", "steerability" ], "rating_list": [ 6, 10, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 4, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "R8t9Q3jmCQ", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Jian Ren", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Jian Wang", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Lionel Z. WANG", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "MS student" }, { "name": "Sizhuo Ma", "gender": "unknown", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Xinjie Li", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Yiming Zhang", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Researcher" }, { "name": "Zhihang Zhong", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Xiao Sun", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 44, "n_ref": 74, "n_ref_all": 92, "n_fig": 8, "n_tab": 5, "L_tab": 2264, "n_element_tab": 205, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 2470, "n_element_tab_1": 208, "formula_len_all": 495, "formula_len_all_1": 495, "len_all": 119707, "len_all_1": 55205, "len_abs": 1191, "len_title": 124, "len_sents": 28967, "len_sents_1": 24199, "n_sents": 189, "n_sents_1": 156, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1373, "title": "DiffBody: Human Body Restoration by Imagining with Generative Diffusion Prior", "abs": "Human body restoration plays a vital role in various applications related to the human body.\nDespite recent advances in general image restoration using generative models, their performance in human body restoration remains mediocre, often resulting in foreground and background blending, over-smoothing surface textures, missing accessories, and distorted limbs. \nAddressing these challenges, we propose a novel approach by constructing a human body-aware diffusion model that leverages domain-specific knowledge to enhance performance. \nSpecifically, we employ a pretrained body attention module to guide the diffusion model's focus on the foreground, addressing issues caused by blending between the subject and background.\nWe also demonstrate the value of revisiting the language modality of the diffusion model in restoration tasks by seamlessly incorporating text prompt to improve the quality of surface texture and additional clothing and accessories details.\nAdditionally, we introduce a diffusion sampler tailored for fine-grained human body parts, utilizing local semantic information to rectify limb distortions.\nLastly, we collect a comprehensive dataset for benchmarking and advancing the field of human body restoration.\nExtensive experimental validation showcases the superiority of our approach, both quantitatively and qualitatively, over existing methods.", "keywords": [ "Diffusion Models" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "R8YCBH5HWo", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Hans-peter Seidel", "gender": "Male", "institution": "Max-Planck Institute", "country": "DE", "position": "Scientific Director" }, { "name": "Kevin Tirta Wijaya", "gender": "unknown", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "", "position": "PhD student" }, { "name": "Michael Sun", "gender": "Male", "institution": "Computer Science and Artificial Intelligence Laboratory, Electrical Engineering & Computer Science", "country": "US", "position": "PhD student" }, { "name": "Minghao Guo", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Vahid Babaei", "gender": "unknown", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Researcher" }, { "name": "Wojciech Matusik", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 34, "n_ref": 66, "n_ref_all": 92, "n_fig": 3, "n_tab": 7, "L_tab": 5193, "n_element_tab": 639, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 911, "n_element_tab_1": 140, "formula_len_all": 680, "formula_len_all_1": 666, "len_all": 151514, "len_all_1": 62060, "len_abs": 1618, "len_title": 116, "len_sents": 33097, "len_sents_1": 28507, "n_sents": 232, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1629, "title": "Two-Stage Pretraining for Molecular Property Prediction in the Wild", "abs": "Accurate property prediction is crucial for accelerating the discovery of new compounds. Although deep learning models have achieved remarkable success, their performance often relies on large amounts of labeled data that are expensive and time-consuming to obtain. Thus, there is a growing need for models that can perform well with limited experimentally-validated data. In this work, we introduce MoleVers, a versatile pretrained model designed for various types of molecular property prediction *in the wild*, i.e., where experimentally-validated molecular property labels are scarce. MoleVers adopts a two-stage pretraining strategy. In the first stage, the model learns molecular representations from large unlabeled datasets via masked atom prediction and *dynamic denoising*, a novel task enabled by a new branching encoder architecture. In the second stage, MoleVers is further pretrained using auxiliary labels obtained with inexpensive computational methods, enabling supervised learning without the need for costly experimental data. This two-stage framework allows MoleVers to learn representations that generalize effectively across various downstream datasets. We evaluate MoleVers on a new benchmark comprising 22 molecular datasets with diverse types of properties, the majority of which contain 50 or fewer training labels reflecting real-world conditions. MoleVers achieves state-of-the-art results on 20 out of the 22 datasets, and ranks second among the remaining two, highlighting its ability to bridge the gap between data-hungry models and real-world conditions where practically-useful labels are scarce.", "keywords": [ "large-scale pretraining", "molecular property prediction" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "R8APzK2Vsf", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Chia-Yuan Chang", "gender": "Not Specified", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Guanchu Wang", "gender": "Male", "institution": "Rice University", "country": "", "position": "PhD student" }, { "name": "Huiyuan Chen", "gender": "Male", "institution": "VISA", "country": "", "position": "Researcher" }, { "name": "Junpeng Wang", "gender": "unknown", "institution": "VISA", "country": "", "position": "Researcher" }, { "name": "Mingzhi Hu", "gender": "Female", "institution": "Worcester Polytechnic Institute", "country": "US", "position": "PhD student" }, { "name": "Xiaoting Li", "gender": "Female", "institution": "VISA", "country": "", "position": "Researcher" }, { "name": "Xiran Fan", "gender": "unknown", "institution": "VISA", "country": "US", "position": "Researcher" }, { "name": "Yuzhong Chen", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Xia Hu", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 26, "n_ref": 35, "n_ref_all": 53, "n_fig": 5, "n_tab": 6, "L_tab": 3514, "n_element_tab": 294, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1339, "n_element_tab_1": 199, "formula_len_all": 938, "formula_len_all_1": 709, "len_all": 119089, "len_all_1": 66399, "len_abs": 1931, "len_title": 126, "len_sents": 30879, "len_sents_1": 28504, "n_sents": 282, "n_sents_1": 248, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1941, "title": "Advancing Table Understanding of Large Language Models via Feature Re-ordering", "abs": "Large Language Models (LLMs) exhibit exceptional proficiency in comprehending human language. Despite their significant success across a wide array of tasks, including text generation, translation, question answering, and even code generation, understanding tabular data remains a challenging task. Especially, tabular data lacks an intrinsic order of the different features (table fields), whereas LLMs take only sequential inputs. Consequently, an artificial order is imposed, the impact of which on the performance of LLMs has not yet been thoroughly investigated. Surprisingly, as discovered in this work, this artificially induced order bias dramatically influences the performance of LLMs on tasks related to tabular data. Mitigating the order bias presents a significant challenge. To address this, we propose a simple and cost-effective method, Re-Ordering Tabular feATures fOR LLM (ROTATOR-LLM), to conduct test-time compute without fine-tuning the base LLM. Aiming at optimizing the feature order of tabular data and boosting LLMs\u2019 capability to better understand the data semantics, ROTATOR-LLM re-frames the ordering problem as a feature trajectory generation task. A dynamic programming based meta-controller is trained to auto-regressively generate an individualized feature trajectory for each data instance via accumulative value estimation of the serialized feature input through the LLM\u2019s final performance metrics. Model performance is maximized by iteratively selecting features across different steps. Experimental results on multiple datasets and LLMs show close to or over 20% performance boosts via features reordered by ROTATOR-LLM against the un-ordered counterpart. Also, it outperforms State-Of-The-Art tabular LLM methods with significant margin. Moreover, meta-controller demonstrates strong transferability: the tested LLMs gain performance enhancements when utilizing a meta-controller trained on one of them.", "keywords": [ "Large language models", "Tabular data", "Feature re-ordering" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "R7pR4dzgAV", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Bowen Zhou", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Chengxing Xie", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Ermo Hua", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Heng Zhou", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Mingyu Liang", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Ning Ding", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Xin Zhong", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "Undergrad student" }, { "name": "Yuchen Fan", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yuchen Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 70, "n_ref_all": 114, "n_fig": 14, "n_tab": 8, "L_tab": 1686, "n_element_tab": 154, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1217, "n_element_tab_1": 93, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 266722, "len_all_1": 61519, "len_abs": 1454, "len_title": 113, "len_sents": 84896, "len_sents_1": 29095, "n_sents": 661, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1460, "title": "CALF: Benchmarking Evaluation of LFQA Using Chinese Examinations", "abs": "Long-Form Question Answering (LFQA) refers to generating in-depth,paragraph-level responses to open-ended questions. Although lots of LFQA methods are developed, evaluating LFQA effectively and efficiently remains challenging due to its high complexity and cost. Therefore, there is no standard benchmark for LFQA evaluation till now. To address this gap, we make the first attempt by proposing a well-constructed, reference-based benchmark named Chinese exAmination for LFQA Evaluation (CALF), aiming to rigorously assess the performance of automatic evaluation metrics for LFQA. The CALF benchmark is derived from Chinese examination questions that have been translated into English. It includes up to 1476 examples consisting of knowledge-intensive and nuanced responses. Our evaluation comprises three different settings to analyze the behavior of automatic metrics comprehensively. We conducted extensive experiments on 7 traditional evaluation metrics, 3 prompt-based metrics, and 3 trained evaluation metrics, and tested on agent systems for the LFQA evaluation. The results reveal that none of the current automatic evaluation metrics shows comparable performances with humans, indicating that they cannot capture dense information contained in long-form responses well. In addition, we provide a detailed analysis of the reasons why automatic evaluation metrics fail when evaluating LFQA, offering valuable insights to advance LFQA evaluation systems.", "keywords": [ "Benchmark", "Evaluation", "Long-form QA", "Large Language Models" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "R7l5kMJTut", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Jiahang Cao", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Lingfeng Zhang", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "MS student" }, { "name": "Renjing Xu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Rui Qing", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Associate Professor" }, { "name": "Shiyi LIU", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Sitao Zhang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yuetong FANG", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Zixuan Jiang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "MS student" }, { "name": "QIANG ZHANG", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 49, "n_ref": 71, "n_ref_all": 95, "n_fig": 18, "n_tab": 6, "L_tab": 1698, "n_element_tab": 160, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 2226, "n_element_tab_1": 98, "formula_len_all": 255, "formula_len_all_1": 181, "len_all": 168649, "len_all_1": 62539, "len_abs": 1737, "len_title": 119, "len_sents": 48560, "len_sents_1": 28687, "n_sents": 338, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1745, "title": "Recursive Cleaning for Large-scale Protein Data via Multimodal Learning", "abs": "Reliable datasets and high-performance models work together to drive significant advancements in protein representation learning in the era of Artificial Intelligence. The size of protein models and datasets has grown exponentially in recent years. However, the quality of protein knowledge and model training has suffered from the lack of accurate and efficient data annotation and cleaning methods.\nTo address this challenge, we introduce ProtAC, which corrects large Protein datasets with a scalable Automatic Cleaning framework that leverages both sequence and functional information through multimodal learning. To fulfill data cleaning, we propose the Sequence-Annotation Matching (SAM) module in the model, which filters the functional annotations that are more suitable for the corresponding sequences. Our approach is a cyclic process consisting of three stages: first pretraining the model on a large noisy dataset, then finetuning the model on a small manually annotated dataset, and finally cleaning the noisy dataset using the finetuned model. Through multiple rounds of \u201ctrain-finetune-clean\u201d cycles, we observe progressive improvement in protein function prediction and sequence-\nannotation matching. As a result, we achieve (1) a state-of-the-art (SOTA) model that outperforms competitors with fewer than 100M parameters, evaluated on multiple function-related downstream tasks, and (2) a cleaned UniRef50 dataset containing \u223c50M proteins with well-annotated functions. Performing extensive biological analysis on a cleaned protein dataset, we demonstrate that our model is able to understand the relationships between different functional annotations in proteins and that proposed functional annotation revisions are reasonable.", "keywords": [ "protein; data cleaning; multimodal learning" ], "rating_list": [ 5, 6, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 5, 4, 4 ] }, { "paper_id": "R7edIYodis", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Muhammed Fatih Balin", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Vaibhav Sharma", "gender": "Male", "institution": "ShareChat", "country": "", "position": "Machine Learning Engineer" }, { "name": "abhinav nagpal", "gender": "Male", "institution": "Indian Institute of Technology, Delhi", "country": "IN", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 30, "n_ref": 50, "n_ref_all": 71, "n_fig": 4, "n_tab": 6, "L_tab": 1519, "n_element_tab": 200, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1172, "n_element_tab_1": 155, "formula_len_all": 223, "formula_len_all_1": 64, "len_all": 116168, "len_all_1": 64228, "len_abs": 860, "len_title": 109, "len_sents": 36004, "len_sents_1": 28311, "n_sents": 274, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 863, "title": "SIRD: Transformers Assisted Step by Step Symbolic Integration", "abs": "Recently, deep learning has gained popularity in solving statistical or approximate problems. However, working with symbolic data has been challenging for neural networks. Despite this, the natural sciences are making strides in utilizing deep learning for various use cases. In this work, we aim to solve the problem of symbolic integration by using deep learning through integral rule prediction, enabling faster search and better interpretability. We propose a novel symbolic integration rules dataset containing 27 million distinct functions and integration rule pairs. We show that by combining a transformer model trained on this dataset into SymPy's integral_steps function, the number of branches explored during the depth-first-search procedure was reduced by a factor of 3 and successfully solve functions that the original version was unable to handle.", "keywords": [ "Transformers", "Symbolic Mathematics", "Symbolic Integration", "NLP", "LLM", "SIRD" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 1, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "R7DKZhgyDX", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Junxian Feng", "gender": "Male", "institution": "East China Normal University", "country": "", "position": "Undergrad student" }, { "name": "Ming Hu", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Researcher" }, { "name": "Mingsong Chen", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Pengyu Zhang", "gender": "unknown", "institution": "East China Normal University", "country": "CN", "position": "PhD student" }, { "name": "Xian Wei", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Principal Researcher" }, { "name": "Yao Liu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Yingbo Zhou", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 21, "n_ref_uni": 25, "n_ref": 35, "n_ref_all": 51, "n_fig": 5, "n_tab": 7, "L_tab": 1998, "n_element_tab": 94, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1968, "n_element_tab_1": 126, "formula_len_all": 1753, "formula_len_all_1": 1799, "len_all": 128173, "len_all_1": 72079, "len_abs": 1548, "len_title": 128, "len_sents": 39138, "len_sents_1": 29341, "n_sents": 331, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1553, "title": "PerFIT: Personalized Federated Instruction Tuning via Neural Architecture Search", "abs": "Federated Instruction Tuning (FIT) has shown the ability to enable model instruction tuning among massive data owners without exposing privacy. Yet, it still faces two key challenges, i.e., data and resource heterogeneity. Due to the varying data distribution and preferences among data owners, FIT cannot adapt to the personalized data of individual owners. Moreover, clients with superior computational abilities have to compromise to maintain the same fine-tuning architecture as the weaker clients. Such a constraint prevents the powerful clients from having more trainable parameters for better fine-tuning performances. To address these issues uniformly, we propose a novel Personalized Federated Instruction Tuning (PerFIT) framework based on architecture search. Specifically, PerFIT allows each client to search for a personalized architecture by expanding the trainable parameter space of the global model, pruning them, and obtaining personalized sparse patterns. We further propose personalized parameter-wise aggregation to facilitate flexible aggregation among clients with diverse sparse patterns. This procedure allows personalized instruction fine-tuning within the expanded parameter spaces, concurrently preserving the same number of trainable parameters as the vanilla state, thus introducing no extra resource burden. \nThe evaluations with multiple LLMs on various instruction-following datasets demonstrate that our approach can achieve up to a 23% decrease in personalized perplexity compared to the state-of-the-art FIT methods.", "keywords": [ "federated instruction tuning", "personalized federated learning", "neural architecture search" ], "rating_list": [ 8, 3, 3, 3, 6 ], "soundness_list": [ 4, 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3, 4 ] }, { "paper_id": "R6sIi9Kbxv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Can Huang", "gender": "Male", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "Guanbin Li", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Han Wang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Liang Lin", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "US", "position": "Full Professor" }, { "name": "Yanjie Wang", "gender": "Male", "institution": "ByteDance Inc", "country": "CN", "position": "Researcher" }, { "name": "Yuxiang Nie", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 16, "n_ref": 36, "n_ref_all": 44, "n_fig": 3, "n_tab": 11, "L_tab": 2485, "n_element_tab": 334, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1025, "n_element_tab_1": 154, "formula_len_all": 110, "formula_len_all_1": 109, "len_all": 137116, "len_all_1": 53723, "len_abs": 1092, "len_title": 166, "len_sents": 35855, "len_sents_1": 25468, "n_sents": 285, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 117, "L_abs": 1096, "title": "Video Q-Former: Multimodal Large Language Model with Spatio-Temporal Querying Transformer Towards Video Understanding", "abs": "Large language models (LLMs) have made remarkable strides in natural language processing tasks. However, effectively processing and understanding visual information remains a challenge for these models. To address this, multimodal large language models have been proposed, which integrate pre-trained visual encoders with LLMs. Although existing image-based approaches have shown success in aligning visual and textual modalities, extending these advancements to videos is challenging due to the richer visual and temporal information they contain. Current methods, including Video-ChatGPT and Video-LLaMA, have limitations in capturing inter-frame relationships and providing sufficient semantic context. To overcome these challenges, we propose Video Q-Former, a model that adaptively extracts spatiotemporal features from videos with a spatio-temporal querying transformer, enhancing the LLM\u2019s comprehension of visual-language alignment. Extensive experiments demonstrate that our model achieves state-of-the-art performance across various datasets in zero-shot video question answering tasks.", "keywords": [ "Multimodal Large Language Model", "Vision-Language Pretraining", "Video Understanding" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "R6q67CDBCH", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Keno Harada", "gender": "Male", "institution": "the University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Masachika Taniguchi", "gender": "Male", "institution": "University of the Ryukyus", "country": "JP", "position": "MS student" }, { "name": "Takeshi Kojima", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Researcher" }, { "name": "Yudai Yamazaki", "gender": "Male", "institution": "Kyoto University", "country": "JP", "position": "MS student" }, { "name": "Yusuke Iwasawa", "gender": "Male", "institution": "The University of Tokyo, The University of Tokyo", "country": "JP", "position": "Associate Professor" }, { "name": "Yutaka Matsuo", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 30, "n_ref": 54, "n_ref_all": 74, "n_fig": 9, "n_tab": 7, "L_tab": 1414, "n_element_tab": 152, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 308, "n_element_tab_1": 33, "formula_len_all": 107, "formula_len_all_1": 107, "len_all": 148749, "len_all_1": 57826, "len_abs": 1968, "len_title": 137, "len_sents": 36823, "len_sents_1": 28160, "n_sents": 280, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1981, "title": "Curse of Instructions: Large Language Models Cannot Follow Multiple Instructions at Once", "abs": "Large language models (LLMs) have demonstrated impressive performance across various natural language processing (NLP) tasks owing to the strong capability of following instructions. To further accelerate the integration of LLMs into our society, it is essential to have LLMs follow many instructions as accurately as humans do. This study reveals that LLMs unexpectedly struggle to follow all instructions simultaneously as the number of instructions increases. First, to validate our claim, we introduce ManyIFEval, a large-scale benchmark dataset comprising task prompts with up to ten objectively verifiable instructions. Second, we conduct experiments based on ManyIFEval with GPT-4o, Claude-3.5, Gemini-1.5, Gemma2, and Llama3.1, demonstrating that as the instruction count rises, the models' ability to follow individual instruction deteriorates gradually but constantly. As a result, the models' ability to follow all the instructions significantly drops: the success rate of all the instructions is precisely explained by the success rate of individual instructions to the power of total number of instructions. We refer to it as the ``curse of instructions''. Third, to remove the curse without retraining models, we propose an inference-time strategy that enhances performance through iterative self-refinement. We demonstrate that instruction-level chain-of-thought reasoning significantly improves their capability to detect and correct instruction-following errors. Notably, our method has improved the success rate of following ten instructions by GPT-4o from 15% to 31% and Claude 3.5 Sonnet from 44% to 58%. We also show that precision is more important than recall in feedback: just telling LLMs that they are not following all the instructions also improves self-refinement success. Our findings highlight a fundamental limitation of instruction-following ability and suggest a future direction for building trustworthy LLMs that can coexist with human society.", "keywords": [ "instruction following", "large language models" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "R5xozf2ZoP", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Bai LiChen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Haoyi Xiong", "gender": "Male", "institution": "Microsoft Corporation", "country": "CN", "position": "Principal Applied Scientist" }, { "name": "Zeke Xie", "gender": "Male", "institution": "HKUST(GZ)", "country": "CN", "position": "Assistant Professor" }, { "name": "Zipeng Qi", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 43, "n_ref": 82, "n_ref_all": 117, "n_fig": 11, "n_tab": 7, "L_tab": 1736, "n_element_tab": 145, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 756, "n_element_tab_1": 40, "formula_len_all": 380, "formula_len_all_1": 264, "len_all": 153357, "len_all_1": 56677, "len_abs": 1509, "len_title": 126, "len_sents": 43516, "len_sents_1": 25474, "n_sents": 395, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1522, "title": "Not All Noises Are Created Equally: Diffusion Noise Selection and Optimization", "abs": "Diffusion models that can generate high-quality data from randomly sampled Gaussian noises have become the mainstream generative method in academia and industry. Are randomly sampled Gaussian noises equally effective for diffusion models? Some methods explore the impact of noise variations on the results, but they either do not operate in the pure noise space, requiring additional evaluation models, or cannot be adapted to general text-to-image tasks. In this paper, we mainly made three contributions. First, we are the first to hypothesize and empirically observe that the generation quality of diffusion models significantly depends on the noise inversion stability. This naturally provides a noise quality metric for noise selection, grounded in a mathematical property. Second, we further propose a novel noise optimization method that actively enhances the inversion stability of arbitrary given noises. Our method is the first one that works purely on noise space for general text-to-image without fine-tuning model parameters or relying on additional result quality evaluators. Third, our extensive experiments demonstrate that the proposed noise selection and noise optimization methods both significantly improve representative diffusion models, such as SDXL and SDXL-turbo, in terms of human preference and other objective evaluation metrics. For example, the human preference winning rates of noise selection and noise optimization over the baselines can be up to 57% and 72.5%, respectively, on DrawBench.", "keywords": [ "Diffusion Models", "Optimization", "Generative AI", "Text-to-Image Generation" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 5, 3, 5 ] }, { "paper_id": "R5FzCFR5yU", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Alves Jos\u00e9", "gender": "Male", "institution": "Transvalor S.A.", "country": "FR", "position": "Researcher" }, { "name": "Frederic Magoules", "gender": "unknown", "institution": "CentraleSupelec", "country": "", "position": "Full Professor" }, { "name": "Marien Chenaud", "gender": "Male", "institution": "CentraleSupelec", "country": "FR", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 14, "n_ref_uni": 31, "n_ref": 44, "n_ref_all": 64, "n_fig": 5, "n_tab": 1, "L_tab": 134, "n_element_tab": 11, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 134, "n_element_tab_1": 11, "formula_len_all": 716, "formula_len_all_1": 514, "len_all": 112268, "len_all_1": 54892, "len_abs": 1010, "len_title": 148, "len_sents": 31051, "len_sents_1": 25977, "n_sents": 251, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1014, "title": "Hybrid Numerical PINNs: On the effectiveness of numerical differentiation for non-analytic problems", "abs": "This work demonstrates that automatic differentiation has strong limitations when employed to compute physical derivatives in a general physics-informed framework, therefore limiting the range of applications that these methods can address. A hybrid approach is proposed, combining deep learning and traditional numerical solvers such as the finite element method, to address the shortcomings of automatic differentiation. This novel approach enables the exact imposition of Dirichlet boundary conditions in a seamless manner, and more complex, non analytical problems can be solved. Finally, enriched inputs can be used by the model to help convergence. The proposed approach is flexible and can be incorporated into any physics-informed model. Our hybrid gradient computation proposal is also up to two orders of magnitude faster than automatic differentiation, as its numerical cost is independent of the complexity of the trained model. Several numerical applications are provided to illustrate the discussion.", "keywords": [ "Scientific Machine Learning", "Physics-Informed Neural Networks", "Automatic Differentiation", "Partial Differential Equations" ], "rating_list": [ 6, 3, 1 ], "soundness_list": [ 3, 1, 2 ], "presentation_list": [ 2, 3, 1 ], "contribution_list": [ 3, 1, 1 ], "confidence_list": [ 2, 4, 5 ] }, { "paper_id": "R4q3cY3kQf", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Andreas Krause", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Bhavya Sukhija", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Carmelo Sferrazza", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "Postdoc" }, { "name": "Pieter Abbeel", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Stelian Coros", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "Associate Professor" } ], "n_formula": 49, "n_formula_1": 18, "n_ref_uni": 71, "n_ref": 183, "n_ref_all": 237, "n_fig": 18, "n_tab": 5, "L_tab": 458, "n_element_tab": 38, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5030, "formula_len_all_1": 1422, "len_all": 207883, "len_all_1": 64891, "len_abs": 1244, "len_title": 143, "len_sents": 55280, "len_sents_1": 28938, "n_sents": 459, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1251, "title": "MaxInfoRL: Boosting exploration in reinforcement learning through information gain maximization", "abs": "Reinforcement learning (RL) algorithms aim to balance exploiting the current best strategy with exploring new options that could lead to higher rewards. Most common RL algorithms use undirected exploration, i.e., select random sequences of actions.\nExploration can also be directed using intrinsic rewards, such as curiosity or model epistemic uncertainty. However, effectively balancing task and intrinsic rewards is challenging and often task-dependent. In this work, we introduce a framework, MaxInfoRL, for balancing intrinsic and extrinsic exploration. MaxInfoRL steers exploration towards informative transitions, by maximizing intrinsic rewards such as the information gain about the underlying task. \nWhen combined with Boltzmann exploration, this approach naturally trades off maximization of the value function with that of the entropy over states, rewards, and actions. We show that our approach achieves sublinear regret in the simplified setting of multi-armed bandits. We then apply this general formulation to a variety of off-policy model-free RL methods for continuous state-action spaces, yielding novel algorithms that achieve superior performance across hard exploration problems and complex scenarios such as visual control tasks.", "keywords": [ "Reinforcement learning", "Exploration in off-policy methods", "Continuous control" ], "rating_list": [ 8, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "R4h5PXzUuU", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Jihyo Kim", "gender": "unknown", "institution": "Seoul National University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Sangheum Hwang", "gender": "unknown", "institution": "Seoul National University of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Seulbi Lee", "gender": "unknown", "institution": "Seoul National University of Science and Technology", "country": "KR", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 85, "n_ref_all": 99, "n_fig": 5, "n_tab": 15, "L_tab": 12981, "n_element_tab": 834, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1476, "n_element_tab_1": 128, "formula_len_all": 100, "formula_len_all_1": 0, "len_all": 281701, "len_all_1": 79505, "len_abs": 1577, "len_title": 158, "len_sents": 83440, "len_sents_1": 36712, "n_sents": 689, "n_sents_1": 296, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 110, "L_abs": 1475, "title": "Reflexive Guidance: Improving OoDD in Vision-Language Models via Self-Guided Image-Adaptive Concept Generation", "abs": "With the recent emergence of foundation models trained on internet-scale data and demonstrating remarkable generalization capabilities, such foundation models have become more widely adopted, leading to an expanding range of application domains. Despite this rapid proliferation, the trustworthiness of foundation models remains underexplored. Specifically, the out-of-distribution detection (OoDD) capabilities of large vision-language models (LVLMs), such as GPT-4o, which are trained on massive multi-modal data, have not been sufficiently addressed. The disparity between their demonstrated potential and practical reliability raises concerns regarding the safe and trustworthy deployment of foundation models. To address this gap, we evaluate and analyze the OoDD capabilities of various proprietary and open-source LVLMs. Our investigation contributes to a better understanding of how these foundation models represent confidence scores through their generated natural language responses. Furthermore, we propose a self-guided prompting approach, termed Reflexive Guidance (ReGuide), aimed at enhancing the OoDD capability of LVLMs by leveraging self-generated image-adaptive concept suggestions. Experimental results demonstrate that our ReGuide enhances the performance of current LVLMs in both image classification and OoDD tasks. The lists of sampled images, along with the prompts and responses for each sample are available at https://github.com/daintlab/ReGuide.", "keywords": [ "Out-of-distribution detection", "Trustworthiness", "Vision-language models", "Foundation models" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "R32pqU5vej", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jian Yang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Mingqian Ji", "gender": "unknown", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Shanshan Zhang", "gender": "Female", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 11, "n_ref": 22, "n_ref_all": 39, "n_fig": 9, "n_tab": 11, "L_tab": 6767, "n_element_tab": 255, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 1740, "n_element_tab_1": 113, "formula_len_all": 413, "formula_len_all_1": 415, "len_all": 157397, "len_all_1": 54210, "len_abs": 1300, "len_title": 128, "len_sents": 34490, "len_sents_1": 22985, "n_sents": 266, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1306, "title": "DH-Fusion: Depth-Aware Hybrid Feature Fusion for Multimodal 3D Object Detection", "abs": "State-of-the-art LiDAR-camera 3D object detectors usually focus on feature fusion. However, they neglect the factor of depth while designing the fusion strategy. In this work, we for the first time point out that different modalities play different roles as depth varies via statistical analysis and visualization. Based on this finding, we propose a Depth-Aware Hybrid Feature Fusion (DH-Fusion) strategy that guides the weights of point cloud and RGB image modalities by introducing depth encoding at both global and local levels. Specifically, the Depth-Aware Global Feature Fusion (DGF) module adaptively adjusts the weights of image Bird's-Eye-View (BEV) features in multi-modal global features via depth encoding. Furthermore, to compensate for the information lost when transferring raw features to the BEV space, we propose a Depth-Aware Local Feature Fusion (DLF) module, which adaptively adjusts the weights of original voxel features and multi-view image features in multi-modal local features via depth encoding. Extensive experiments on the nuScenes and KITTI datasets demonstrate that our DH-Fusion method surpasses previous state-of-the-art methods. Moreover, our DH-Fusion is more robust to various kinds of corruptions, outperforming previous methods on nuScenes-C w.r.t. both NDS and mAP.", "keywords": [ "Depth Aware", "Multi Modality", "3D Object Detection" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 5, 4 ] }, { "paper_id": "R2OzZWOkjz", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Seung-Hoon Na", "gender": "Male", "institution": "Chonbuk National University", "country": "", "position": "Full Professor" }, { "name": "Shanbao Qiao", "gender": "Male", "institution": "Chonbuk National University", "country": "KR", "position": "PhD student" }, { "name": "Xuebing Liu", "gender": "Male", "institution": "Chonbuk National University", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 40, "n_ref": 60, "n_ref_all": 74, "n_fig": 3, "n_tab": 4, "L_tab": 426, "n_element_tab": 12, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1206, "n_element_tab_1": 88, "formula_len_all": 484, "formula_len_all_1": 704, "len_all": 120229, "len_all_1": 58704, "len_abs": 1949, "len_title": 139, "len_sents": 32694, "len_sents_1": 26896, "n_sents": 212, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 2002, "title": "Retrieval-Augmented Editing Generation: Impact of Knowledge Editing and Fine-Tuning on RAG", "abs": "The knowledge embedded in Large Language Models (LLMs) is static, tied to the time when the training data was collected. \nWhile Retrieval-Augmented Generation (RAG) methods are widely used to introduce new knowledge, they simply rely on retrieved information for reasoning without integrating it into the model\u2019s parameters. This limits the model's ability for long-term knowledge retention and autonomous learning.\nTo overcome this, in this work, we propose the \\textbf{R}etrieval-\\textbf{A}ugmented \\textbf{E}diting \\textbf{G}eneration (RAEG) framework for open-domain question answering (ODQA) tasks. \nRAEG enhances model generation performance by first editing the retrieved paragraphs to inject necessary knowledge, followed by an augmented generation phase. This dual mechanism\u2014combining knowledge injection and retrieval augmentation\u2014provides complementary advantages in the reasoning process. When the injected knowledge alone is insufficient for accurate generation, the model can rely on the retrieved information to compensate, and conversely, when retrieval yields suboptimal results, the injected knowledge ensures continuity and accuracy in the response. This interplay between internalized and externally sourced knowledge reinforces the model's ability to produce correct answers, thereby enhancing overall task performance.\nWe explore the impact of two key methods for knowledge injection: Knowledge Editing (KE) and Parameter-Efficient Fine-Tuning (PEFT), and analyze how modifying the model's parameters influences its reasoning abilities and generation outcomes. To further improve RAEG's performance, we introduce a re-ranking mechanism to optimize the integration of external knowledge and apply parameter pruning to mitigate the potential drawbacks of parameter modifications during KE.\nEvaluations on two authoritative ODQA benchmarks show that RAEG is able to further replace RAG as a competitive method.\nOur data and code will be available at \\url{https://github.com/XXX/XXX}.", "keywords": [ "Retrieval-Augmented Generation", "Knowledge Editing", "Parameter-Efficient Fine-Tuning" ], "rating_list": [ 3, 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 1, 2 ], "presentation_list": [ 2, 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "R2834dhBlo", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Lewis Hammond", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Sam Adam-Day", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "PhD student" } ], "n_formula": 41, "n_formula_1": 13, "n_ref_uni": 41, "n_ref": 90, "n_ref_all": 127, "n_fig": 20, "n_tab": 1, "L_tab": 220, "n_element_tab": 52, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 220, "n_element_tab_1": 52, "formula_len_all": 4704, "formula_len_all_1": 1124, "len_all": 264485, "len_all_1": 72424, "len_abs": 993, "len_title": 73, "len_sents": 97084, "len_sents_1": 32762, "n_sents": 726, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 25, "L_abs": 996, "title": "Neural Interactive Proofs", "abs": "We consider the problem of how a trusted, but computationally bounded agent (a 'verifier') can learn to interact with one or more powerful but untrusted agents ('provers') in order to solve a given task. More specifically, we study the case in which agents are represented using neural networks and refer to solutions of this problem as neural interactive proofs. First we introduce a unifying framework based on prover-verifier games (Anil et al., 2021), which generalises previously proposed interaction protocols. We then describe several new protocols for generating neural interactive proofs, and provide a theoretical comparison of both new and existing approaches. Finally, we support this theory with experiments in two domains: a toy graph isomorphism problem that illustrates the key ideas, and a code validation task using large language models. In so doing, we aim to create a foundation for future work on neural interactive proofs and their application in building safer AI systems.", "keywords": [ "interactive proofs", "game theory", "neural networks", "safety", "multi-agent reinforcement learning" ], "rating_list": [ 5, 10, 5 ], "soundness_list": [ 2, 4, 3 ], "presentation_list": [ 2, 3, 1 ], "contribution_list": [ 2, 4, 2 ], "confidence_list": [ 4, 2, 3 ] }, { "paper_id": "R22JPTQYWV", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Botao Ren", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Junwei Luo", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Xue Yang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yi Yu", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Postdoc" }, { "name": "Zhidong Deng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 59, "n_ref_all": 69, "n_fig": 5, "n_tab": 9, "L_tab": 2363, "n_element_tab": 180, "n_fig_1": 3, "n_tab_1": 10, "L_tab_1": 2363, "n_element_tab_1": 180, "formula_len_all": 512, "formula_len_all_1": 512, "len_all": 122675, "len_all_1": 59372, "len_abs": 1310, "len_title": 148, "len_sents": 32398, "len_sents_1": 25177, "n_sents": 269, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1331, "title": "PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection", "abs": "Single point supervised oriented object detection has gained attention and made initial progress within the community. Diverse from those approaches relying on one-shot samples or powerful pretrained models (e.g. SAM), PointOBB has shown promise due to its prior-free feature. In this paper, we propose PointOBB-v2, a simpler, faster, and stronger method to generate pseudo rotated boxes from points without relying on any other prior. Specifically, we first generate a Class Probability Map (CPM) by training the network with non-uniform positive and negative sampling. We show that the CPM is able to learn the approximate object regions and their contours. Then, Principal Component Analysis (PCA) is applied to accurately estimate the orientation and the boundary of objects. By further incorporating a separation mechanism, we resolve the confusion caused by the overlapping on the CPM, enabling its operation in high-density scenarios. Extensive comparisons demonstrate that our method achieves a training speed 15.58$\\times$ faster and an accuracy improvement of 11.60\\%/25.15\\%/21.19\\% on the DOTA-v1.0/v1.5/v2.0 datasets compared to the previous state-of-the-art, PointOBB. This significantly advances the cutting edge of single point supervised oriented detection in the modular track. Code and models will be released.", "keywords": [ "Oriented Object Detection", "Point Supervised Object Detection" ], "rating_list": [ 6, 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 2, 4 ], "contribution_list": [ 3, 4, 3, 3, 3 ], "confidence_list": [ 4, 5, 5, 4, 4 ] }, { "paper_id": "R1rNN22IoP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Fujun Luan", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Hao Su", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Hao Tan", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" }, { "name": "Kai Zhang", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Kalyan Sunkavalli", "gender": "Male", "institution": "Adobe Research", "country": "", "position": "Principal Scientist" }, { "name": "Sai Bi", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Valentin Deschaintre", "gender": "Male", "institution": "Adobe Systems", "country": "GB", "position": "Researcher" }, { "name": "Xinyue Wei", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Zexiang Xu", "gender": "Male", "institution": "Hillbot", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 67, "n_ref": 175, "n_ref_all": 199, "n_fig": 10, "n_tab": 8, "L_tab": 1313, "n_element_tab": 119, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 880, "n_element_tab_1": 177, "formula_len_all": 133, "formula_len_all_1": 134, "len_all": 192031, "len_all_1": 66424, "len_abs": 379, "len_title": 107, "len_sents": 45315, "len_sents_1": 31956, "n_sents": 320, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 947, "title": "MeshLRM: Large Reconstruction Model for High-Quality Meshes", "abs": "We propose MeshLRM, a novel LRM-based approach that can reconstruct a high-quality mesh from merely four input images in less than one second. Different from previous large reconstruction models (LRMs) that focus on NeRF-based reconstruction, MeshLRM incorporates differentiable mesh extraction and rendering within the LRM framework. This allows for end-to-end mesh reconstruction by fine-tuning a pre-trained NeRF LRM with mesh rendering. Moreover, we improve the LRM architecture by simplifying several complex designs in previous LRMs. MeshLRM's NeRF initialization is sequentially trained with low- and high-resolution images; this new LRM training strategy enables significantly faster convergence and thereby leads to better quality with less compute. Our approach achieves state-of-the-art mesh reconstruction from sparse-view inputs and also allows for many downstream applications, including text-to-3D and single-image-to-3D generation.", "keywords": [ "Sparse-view reconstruction", "High-quality mesh", "Large Reconstruction Models" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "R1hIXdST22", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Michael Rabbat", "gender": "Male", "institution": "Facebook", "country": "CA", "position": "Principal Researcher" }, { "name": "Pierluca D'Oro", "gender": "Male", "institution": "Facebook", "country": "CA", "position": "Researcher" }, { "name": "Scott Fujimoto", "gender": "unknown", "institution": "Facebook", "country": "CA", "position": "Researcher" }, { "name": "Yuandong Tian", "gender": "Male", "institution": "Meta AI (FAIR)", "country": "", "position": "Research Scientist" }, { "name": "Amy Zhang", "gender": "Female", "institution": "University of Texas at Austin", "country": "", "position": "Assistant Professor" } ], "n_formula": 70, "n_formula_1": 28, "n_ref_uni": 65, "n_ref": 138, "n_ref_all": 155, "n_fig": 2, "n_tab": 5, "L_tab": 2230, "n_element_tab": 43, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 286, "n_element_tab_1": 11, "formula_len_all": 5803, "formula_len_all_1": 1778, "len_all": 249355, "len_all_1": 66991, "len_abs": 1099, "len_title": 105, "len_sents": 43321, "len_sents_1": 29824, "n_sents": 389, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 57, "L_abs": 1103, "title": "Towards General-Purpose Model-Free Reinforcement Learning", "abs": "Reinforcement learning (RL) promises a framework for near-universal problem-solving. In practice however, RL algorithms are often tailored to specific benchmarks, relying on carefully tuned hyperparameters and algorithmic choices. Recently, powerful model-based RL methods have shown impressive general results across benchmarks but come at the cost of increased complexity and slow run times, limiting their broader applicability. In this paper, we attempt to find a unifying model-free deep RL algorithm that can address a diverse class of domains and problem settings. To achieve this, we leverage model-based representations that approximately linearize the value function, taking advantage of the denser task objectives used by model-based RL while avoiding the costs associated with planning or simulated trajectories. We evaluate our algorithm, MR.Q, on a variety of common RL benchmarks with a single set of hyperparameters and show a competitive performance against domain-specific and general baselines, providing a concrete step towards building general-purpose model-free deep RL algorithms.", "keywords": [ "Deep reinforcement learning", "model-free", "general-purpose" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 4, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "R1WF5b5faF", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Cameron Austin", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Haizhao Yang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Associate Professor" }, { "name": "Ling Liang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Postdoc" } ], "n_formula": 33, "n_formula_1": 14, "n_ref_uni": 57, "n_ref": 89, "n_ref_all": 94, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3641, "formula_len_all_1": 1241, "len_all": 150511, "len_all_1": 68908, "len_abs": 1435, "len_title": 127, "len_sents": 43469, "len_sents_1": 31865, "n_sents": 332, "n_sents_1": 230, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1415, "title": "Accelerating Multi-Block Constrained Optimization Through Learning to Optimize", "abs": "Learning to Optimize (L2O) approaches, including algorithm unrolling, plug-and-play methods, and hyperparameter learning, have garnered significant attention and have been successfully applied to the Alternating Direction Method of Multipliers (ADMM) and its variants. However, the natural extension of L2O to multi-block ADMM-type methods remains largely unexplored. Such an extension is critical, as multi-block methods leverage the separable structure of optimization problems, offering substantial reductions in per-iteration complexity. Given that classical multi-block ADMM does not guarantee convergence, the Majorized Proximal Augmented Lagrangian Method (MPALM), which shares a similar form with multi-block ADMM and ensures convergence, is more suitable in this setting. Despite its theoretical advantages, MPALM\u2019s performance is highly sensitive to the choice of penalty parameters. To address this limitation, we propose a novel L2O approach that adaptively selects this hyperparameter using supervised learning. We demonstrate the versatility and effectiveness of our method by applying it to the Lasso problem and the optimal transport problem. Our numerical results show that the proposed framework outperforms popular alternatives. Given its applicability to generic linearly constrained composite optimization problems, this work opens the door to a wide range of potential real-world applications.", "keywords": [ "Learning to Optimize", "Multi-Block Constrained Optimzation", "Alternating Direction Method of Multipliers", "Proximal Augmented Lagrangian Methods" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "R0YGjmqiwB", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Guoqing Ma", "gender": "Male", "institution": "Institute of automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Liangxuan Guo", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Shan Yu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yang Chen", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Assistant Professor" }, { "name": "Yuhan Zhang", "gender": "Female", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "PhD student" } ], "n_formula": 54, "n_formula_1": 13, "n_ref_uni": 37, "n_ref": 58, "n_ref_all": 114, "n_fig": 39, "n_tab": 2, "L_tab": 793, "n_element_tab": 41, "n_fig_1": 32, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4332, "formula_len_all_1": 828, "len_all": 202568, "len_all_1": 71091, "len_abs": 1365, "len_title": 158, "len_sents": 69961, "len_sents_1": 33745, "n_sents": 568, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1372, "title": "Efficient Learning in Neural Networks without Gradient Backpropagation", "abs": "The brain possesses highly efficient learning algorithms that have not been fully understood. The gradient backpropagation (BP) serves as a powerful tool for training artificial neural networks, but it diverges from the known anatomical and physiological constraints of the brain. Conversely, biologically plausible learning algorithms have efficiency limitations in training deep neural networks. To bridge this gap, we introduce a perturbation-based approach called low-rank cluster orthogonal (LOCO) weight modification. Theoretical analysis shows that LOCO provides an unbiased estimate of the BP gradient and achieves low variance in gradient estimation. Compared with some brain-inspired algorithms, LOCO keeps mathematical convergence guarantees and improves the efficiency. It can train the deepest spiking neural networks to date without gradient backpropagation, achieving state-of-the-art performance on several benchmark datasets and exhibiting the ability to overcome catastrophic forgetting. These findings suggest that biologically feasible learning methods can be substantially more efficient than previously believed. Furthermore, avoiding gradient backpropagation allows LOCO to achieve O(1) time complexity for weight updates. This opens a promising avenue for developing distributed computing systems that are more efficient than BP-based counterparts.", "keywords": [ "Brain-inspired", "Learning algorithm", "Learning efficiency" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "R03zKO9T9S", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Chengjie Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Haoyang He", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jiangning Zhang", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Lei Xie", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Qingdong He", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Yabiao Wang", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Senior Researcher" }, { "name": "Yong Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zhenye Gan", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "yuxuan cai", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "zhucun xue", "gender": "Female", "institution": "Wuhan University of Technology", "country": "", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 17, "n_ref": 36, "n_ref_all": 39, "n_fig": 2, "n_tab": 8, "L_tab": 6742, "n_element_tab": 481, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 4977, "n_element_tab_1": 396, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 119883, "len_all_1": 67589, "len_abs": 1634, "len_title": 125, "len_sents": 28636, "len_sents_1": 23396, "n_sents": 192, "n_sents_1": 145, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1667, "title": "A Comprehensive Library for Benchmarking Multi-class Visual Anomaly Detection", "abs": "Visual anomaly detection aims to identify anomalous regions in images through unsupervised learning paradigms, with increasing application demand and value in fields such as industrial inspection and medical lesion detection. Despite significant progress in recent years, there is a lack of comprehensive benchmarks to adequately evaluate the performance of various mainstream methods across different datasets under the practical multi-class setting. The absence of standardized experimental setups can lead to potential biases in training epochs, resolution, and metric results, resulting in erroneous conclusions. This paper addresses this issue by proposing a comprehensive visual anomaly detection benchmark, ***ADer***, which is a modular framework that is highly extensible for new methods. The benchmark includes multiple datasets from industrial and medical domains, implementing fifteen state-of-the-art methods and nine comprehensive metrics. Additionally, we have proposed the GPU-assisted ***ADEval*** package to address the slow evaluation problem of metrics like time-consuming mAU-PRO on large-scale data, significantly reducing evaluation time by more than \\textit{1000-fold}. Through extensive experimental results, we objectively reveal the strengths and weaknesses of different methods and provide insights into the challenges and future directions of multi-class visual anomaly detection. We hope that ***ADer*** will become a valuable resource for researchers and practitioners in the field, promoting the development of more robust and generalizable anomaly detection systems. Full codes have been attached in Appendix and will be open-sourced.", "keywords": [ "Visual Anomaly Detection", "Unsupervised Learning", "benchmark" ], "rating_list": [ 3, 3, 8, 5 ], "soundness_list": [ 2, 2, 4, 3 ], "presentation_list": [ 1, 2, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "QzxuHbWAD6", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Junfan Li", "gender": "Male", "institution": "Harbin Institute of Technology Shenzhen", "country": "CN", "position": "Postdoc" }, { "name": "Liqiang Nie", "gender": "Male", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "dean" }, { "name": "Xiao Zhang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zenglin Xu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 151, "n_formula_1": 47, "n_ref_uni": 29, "n_ref": 69, "n_ref_all": 80, "n_fig": 0, "n_tab": 13, "L_tab": 3058, "n_element_tab": 290, "n_fig_1": 0, "n_tab_1": 6, "L_tab_1": 1585, "n_element_tab_1": 105, "formula_len_all": 22977, "formula_len_all_1": 4983, "len_all": 185401, "len_all_1": 66921, "len_abs": 1190, "len_title": 144, "len_sents": 40616, "len_sents_1": 22495, "n_sents": 575, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1197, "title": "On the Power of Federated Learning for Online Sparse Linear Regression with Decentralized Data", "abs": "In this paper, we study the necessity of federated learning (FL) for online linear regression with decentralized data. Previous work proved that FL is unnecessary for minimizing regret in full information setting, while we prove that it can be necessary if only limited attributes of each instance are observed. We call this problem online sparse linear regression with decentralized data (OSLR-DecD). We propose a federated algorithm for OSLR-DecD, and prove a lower bound on the regret of any noncooperative algorithm. In the case of $d=o(M)$, the upper bound on the regret of our algorithm is smaller than the lower bound, demonstrating the necessity of FL, in which $M$ is the number of clients and $d$ is the dimension of data. When $M=1$, we give the first lower bound on the regret and improve previous upper bounds. We invent three new techniques including an any-time federated online mirror descent with negative entropy regularization, a paradigm for client-server collaboration with privacy protection, and a reduction from online sparse linear regression to prediction with limited advice for establishing the lower bound on the regret, some of which might be of independent interest.", "keywords": [ "Federated learning", "online learning" ], "rating_list": [ 6, 5, 3, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "Qzd4BloAjQ", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Dmitry Yarotsky", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Maksim Velikanov", "gender": "Male", "institution": "Technology Innovation Institute", "country": "", "position": "Researcher" } ], "n_formula": 287, "n_formula_1": 47, "n_ref_uni": 44, "n_ref": 84, "n_ref_all": 162, "n_fig": 5, "n_tab": 3, "L_tab": 803, "n_element_tab": 39, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 19918, "formula_len_all_1": 3774, "len_all": 290310, "len_all_1": 77100, "len_abs": 1247, "len_title": 115, "len_sents": 82463, "len_sents_1": 30122, "n_sents": 841, "n_sents_1": 275, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1288, "title": "SGD with memory: fundamental properties and stochastic acceleration", "abs": "An important open problem is the theoretically feasible acceleration of mini-batch SGD-type algorithms on quadratic problems with power-law spectrum. In the non-stochastic setting, the optimal exponent $\\xi$ in the loss convergence $L_t\\sim C_Lt^{-\\xi}$ is double that in plain GD and is achievable using Heavy Ball (HB) with a suitable schedule; this no longer works in the presence of mini-batch noise. We address this challenge by considering first-order methods with an arbitrary fixed number $M$ of auxiliary velocity vectors (*memory-$M$ algorithms*). We first prove an equivalence between two forms of such algorithms and describe them in terms of suitable characteristic polynomials. Then we develop a general expansion of the loss in terms of *signal and noise propagators*. Using it, we show that losses of stationary stable memory-$M$ algorithms always retain the exponent $\\xi$ of plain GD, but can have different constants $C_L$ depending on their *effective learning rate* that generalizes that of HB. We prove that in memory-1 algorithms we can make $C_L$ arbitrarily small while maintaining stability. As a consequence, we propose a memory-1 algorithm with a time-dependent schedule that we show heuristically and experimentally to improve the exponent $\\xi$ of plain SGD.", "keywords": [ "Stochastic Gradient Descent", "SGD", "spectral power laws", "acceleration", "effective learning rate", "quadratic problems", "stability", "Heavy Ball", "momentum" ], "rating_list": [ 8, 3, 8, 8 ], "soundness_list": [ 4, 1, 4, 4 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 2, 2, 3, 2 ] }, { "paper_id": "QzR7Jfe8Tz", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 6, "author_info_list": [ { "name": "Eyke H\u00fcllermeier", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "Full Professor" }, { "name": "Gabriel Marques Tavares", "gender": "Male", "institution": "University of Milan", "country": "", "position": "PhD student" }, { "name": "Marcel Wever", "gender": "Male", "institution": "Universit\u00e4t Hannover", "country": "DE", "position": "Postdoc" }, { "name": "Sandra Gilhuber", "gender": "Female", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Thomas Seidl", "gender": "unknown", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Valentin Margraf", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 51, "n_ref": 127, "n_ref_all": 147, "n_fig": 10, "n_tab": 5, "L_tab": 4604, "n_element_tab": 82, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 768, "n_element_tab_1": 67, "formula_len_all": 304, "formula_len_all_1": 319, "len_all": 177830, "len_all_1": 58853, "len_abs": 1320, "len_title": 115, "len_sents": 46238, "len_sents_1": 28091, "n_sents": 344, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1325, "title": "ALPBench: A Benchmark for Active Learning Pipelines on Tabular Data", "abs": "In settings where only a budgeted amount of labeled data can be afforded, active learning seeks to devise query strategies for selecting the most informative data points to be labeled, aiming to enhance learning algorithms' efficiency and performance. Numerous such query strategies have been proposed and compared in the active learning literature. However, the community still lacks standardized benchmarks for comparing the performance of different query strategies. This particularly holds for the combination of query strategies with different learning algorithms into active learning pipelines and examining the impact of the learning algorithm choice. To close this gap, we propose ALPBench, which facilitates the specification, execution, and performance monitoring of active learning pipelines. It has built-in measures to ensure evaluations are done reproducibly, saving exact dataset splits and hyperparameter settings of used algorithms. In total, ALPBench consists of 86 real-world tabular classification datasets and 5 active learning settings, yielding 430 active learning problems. To demonstrate its usefulness and broad compatibility with various learning algorithms and query strategies, we conduct an exemplary study evaluating 9 query strategies paired with 8 learning algorithms in 2 different settings.", "keywords": [ "Active Learning", "Benchmark", "Tabular Data", "Software Library" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "QzPKSUUcud", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Behzad Bozorgtabar", "gender": "Male", "institution": "CHUV - University Hospital Lausanne", "country": "CH", "position": "Researcher" }, { "name": "Jean-Philippe Thiran", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Full Professor" }, { "name": "Nikola \u0110uki\u0107", "gender": "Male", "institution": "KU Leuven", "country": "", "position": "PhD student" }, { "name": "Thomas Stegm\u00fcller", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Tim Lebailly", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "PhD student" }, { "name": "Tinne Tuytelaars", "gender": "Female", "institution": "KU Leuven", "country": "BE", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 56, "n_ref": 124, "n_ref_all": 159, "n_fig": 5, "n_tab": 2, "L_tab": 1394, "n_element_tab": 253, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 837, "n_element_tab_1": 144, "formula_len_all": 327, "formula_len_all_1": 287, "len_all": 178525, "len_all_1": 62105, "len_abs": 1239, "len_title": 109, "len_sents": 42841, "len_sents_1": 27345, "n_sents": 311, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1310, "title": "A Simple Framework for Open-Vocabulary Zero-Shot Segmentation", "abs": "Zero-shot classification capabilities naturally arise in models trained within a vision-language contrastive framework. Despite their classification prowess, these models struggle in dense tasks like zero-shot open-vocabulary segmentation. This deficiency is often attributed to the absence of localization cues in captions and the intertwined nature of the learning process, which encompasses both image/text representation learning and cross-modality alignment. To tackle these issues, we propose SimZSS, a $\\textbf{Sim}$ple framework for open-vocabulary $\\textbf{Z}$ero-$\\textbf{S}$hot $\\textbf{S}$egmentation. The method is founded on two key principles: $\\textit{i)}$ leveraging frozen vision-only models that exhibit spatial awareness while exclusively aligning the text encoder and $\\textit{ii)}$ exploiting the discrete nature of text and linguistic knowledge to pinpoint local concepts within captions. By capitalizing on the quality of the visual representations, our method requires only image-caption pair datasets and adapts to both small curated and large-scale noisy datasets. When trained on COCO Captions across 8 GPUs, SimZSS achieves state-of-the-art results on 7 out of 8 benchmark datasets in less than 15 minutes. The code and pretrained models will be publicly available upon acceptance.", "keywords": [ "Vision-language models", "Zero-shot segmentation" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Qyile3DctL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Tong Niu", "gender": "Male", "institution": "Salesforce AI Research", "country": "", "position": "Research Scientist" }, { "name": "Xiangliang Zhang", "gender": "Female", "institution": "University of Notre Dame", "country": "US", "position": "Associate Professor" }, { "name": "Ye Liu", "gender": "Female", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "Yingbo Zhou", "gender": "unknown", "institution": "Salesforce Research", "country": "", "position": "Research Scientist" }, { "name": "Zhenwen Liang", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Researcher" }, { "name": "Semih Yavuz", "gender": "Male", "institution": "SalesForce.com", "country": "", "position": "Research Scientist" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 41, "n_ref": 99, "n_ref_all": 113, "n_fig": 5, "n_tab": 6, "L_tab": 2734, "n_element_tab": 328, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 3562, "n_element_tab_1": 383, "formula_len_all": 284, "formula_len_all_1": 284, "len_all": 160949, "len_all_1": 62783, "len_abs": 1826, "len_title": 143, "len_sents": 41500, "len_sents_1": 27804, "n_sents": 284, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1753, "title": "Improving LLM Reasoning through Scaling Inference Computation with Collaborative Verification", "abs": "Despite significant advancements in the general capability of large language models (LLMs), they continue to struggle with consistent and accurate reasoning, especially in complex tasks such as mathematical and code reasoning. One key limitation is that LLMs are trained primarily on correct solutions, reducing their ability to detect and learn from errors, which hampers their ability to reliably verify and rank outputs. To address this, we scale up the inference-time computation by generating multiple reasoning paths and employing verifiers to assess and rank the generated outputs by correctness. To facilitate this, we introduce a comprehensive dataset consisting of correct and incorrect solutions for math and code tasks, generated by multiple LLMs. This diverse set of solutions enables verifiers to more effectively distinguish and rank correct answers from erroneous outputs. The training methods for building verifiers were selected based on an extensive comparison of existing approaches. Moreover, to leverage the unique strengths of different reasoning strategies, we propose a novel collaborative method integrating Chain-of-Thought (CoT) and Program-of-Thought (PoT) solutions for verification. CoT provides a clear, step-by-step reasoning process that enhances interpretability, while PoT, being executable, offers a precise and error-sensitive validation mechanism. By taking both of their strengths, our approach significantly improves the accuracy and reliability of reasoning verification. Our verifiers, Math-Rev and Code-Rev, demonstrate substantial performance gains to existing LLMs, achieving state-of-the-art results on benchmarks such as GSM8k and MATH and even outperforming GPT-4o with Qwen-72B-Instruct as the reasoner.", "keywords": [ "Verifier", "Math Reasoning", "LLMs" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "QyhxT8xska", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Domenic Rosati", "gender": "unknown", "institution": "Dalhousie University", "country": "", "position": "PhD student" }, { "name": "Ga Wu", "gender": "Male", "institution": "Dalhousie University", "country": "CA", "position": "Assistant Professor" }, { "name": "Hammad Rizwan", "gender": "Male", "institution": "Dalhousie University", "country": "CA", "position": "PhD student" }, { "name": "Hassan Sajjad", "gender": "Male", "institution": "Dalhousie University", "country": "CA", "position": "Associate Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 24, "n_ref": 60, "n_ref_all": 78, "n_fig": 8, "n_tab": 7, "L_tab": 2955, "n_element_tab": 410, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 972, "n_element_tab_1": 116, "formula_len_all": 1039, "formula_len_all_1": 834, "len_all": 139765, "len_all_1": 66236, "len_abs": 997, "len_title": 117, "len_sents": 47395, "len_sents_1": 32217, "n_sents": 315, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 854, "title": "Resolving Lexical Bias in Edit Scoping with Projector Editor Networks", "abs": "Weight-preserving large language model editing techniques rely heavily on the scoping mechanism that decides when to apply an edit to the base model. These scoping mechanisms utilize distance functions in the representation space. In this work, we show that distance-based scoping functions grapple with strong lexical biases leading to issues such as deciding that irrelevant prompts that share overlapping words should result in applying an edit. We address these problems by introducing Projector Editor Networks for Model Editing (PENME), a principled model editing approach designed to learn the optimal representation space for scoping via contrastive learning. We show that PENME achieves state of the art model editing results while being compute-efficient at inference time than previous methods and flexible enough to adapt across architectures", "keywords": [ "Representation Learning", "Model Editing and LLM's" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "QyVLJ7EnAC", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Chi Zhang", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Postdoc" }, { "name": "Yue Wang", "gender": "unknown", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Zain Ulabedeen Farhat", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "George Atia", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Professor" } ], "n_formula": 225, "n_formula_1": 37, "n_ref_uni": 55, "n_ref": 184, "n_ref_all": 196, "n_fig": 8, "n_tab": 7, "L_tab": 1536, "n_element_tab": 153, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 15, "n_element_tab_1": 1, "formula_len_all": 20054, "formula_len_all_1": 2237, "len_all": 270688, "len_all_1": 66036, "len_abs": 3941, "len_title": 114, "len_sents": 70584, "len_sents_1": 29799, "n_sents": 726, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1214, "title": "Model-Free Offline Reinforcement Learning with Enhanced Robustness", "abs": "Offline reinforcement learning (RL) has gained considerable attention for its ability to learn policies from pre-collected data without real-time interaction, which makes it particularly useful for high-risk applications. However, due to its reliance on offline datasets, existing works inevitably introduce assumptions to ensure effective learning, which, however, often lead to a trade-off between robustness to model mismatch and scalability to large environments. In this paper, we enhance both aspects with a novel double-pessimism principle, which conservatively estimates performance and accounts for both limited data and potential model mismatches, two major reasons for the previous trade-off. We then propose a universal, model-free algorithm to learn an optimal policy that is robust to potential environment mismatches, which enhances robustness in a scalable manner. Furthermore, we provide a sample complexity analysis of our algorithm when the mismatch is modeled by the $l_\\alpha$-norm, which also theoretically demonstrates the efficiency of our method. Extensive experiments further demonstrate that our approach significantly improves robustness in a more scalable manner than existing methods.", "keywords": [ "offline RL", "robust", "scalability", "model-free" ], "rating_list": [ 6, 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 4, 2 ], "presentation_list": [ 3, 2, 3, 4, 2 ], "contribution_list": [ 3, 3, 4, 3, 2 ], "confidence_list": [ 3, 3, 3, 3, 2 ] }, { "paper_id": "QyNN5n37nK", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Alexander Swerdlow", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Deepak Pathak", "gender": "Male", "institution": "Skild AI", "country": "US", "position": "CEO" }, { "name": "Katerina Fragkiadaki", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Mihir Prabhudesai", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Siddharth Gandhi", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "MS student" } ], "n_formula": 18, "n_formula_1": 7, "n_ref_uni": 50, "n_ref": 91, "n_ref_all": 135, "n_fig": 18, "n_tab": 4, "L_tab": 1591, "n_element_tab": 151, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 2018, "n_element_tab_1": 116, "formula_len_all": 930, "formula_len_all_1": 402, "len_all": 185369, "len_all_1": 60971, "len_abs": 1636, "len_title": 85, "len_sents": 45478, "len_sents_1": 25970, "n_sents": 358, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 1568, "title": "Unified Multimodal Discrete Diffusion", "abs": "Multimodal generative models that can understand and generate across multiple modalities are dominated by autoregressive (AR) approaches, which process tokens sequentially from left to right, or top to bottom. These models jointly handle images, text, video, and audio for various tasks such as image captioning, question answering, and image generation. While AR models have been highly successful in the text domain, they have been found suboptimal for processing images, videos, and audio due to the high correlation between adjacent tokens which waste inference-time compute by separately predicting each one. In this work, we explore discrete diffusion models as a unified generative formulation in the joint text and image domain, building upon their recent success in the text domain alone. Discrete diffusion models offer several advantages over AR models, including improved control over quality versus diversity of generated samples, the ability to perform joint multimodal inpainting (across both text and image domains), greater controllability in generation through guidance. Leveraging these benefits, we present the first Unified Multimodal Discrete Diffusion (UniDisc) model, which is capable of jointly processing text and images for a variety of downstream tasks. We compare UniDisc to multimodal AR models of similar capacity, demonstrating that UniDisc outperforms them in terms of both performance and inference-time compute, and enhanced controllability, editability, inpainting and flexible trade-off of inference time versus generation quality.", "keywords": [ "multimodal", "discrete diffusion" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "Qy3UwW4OJ9", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Kai Chen", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Liu Yanchen", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Yanan SUN", "gender": "Female", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yanhong Zeng", "gender": "Female", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yinhao Tang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Cai Rong Zhao", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Full Professor" }, { "name": "junyao gao", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 44, "n_ref": 121, "n_ref_all": 148, "n_fig": 20, "n_tab": 7, "L_tab": 1014, "n_element_tab": 119, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 575, "n_element_tab_1": 79, "formula_len_all": 608, "formula_len_all_1": 515, "len_all": 150600, "len_all_1": 58564, "len_abs": 415, "len_title": 82, "len_sents": 38041, "len_sents_1": 27924, "n_sents": 296, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 34, "L_abs": 890, "title": "StyleShot: A snapshot on any style", "abs": "In this paper, we show that, a good style representation is crucial and sufficient for generalized style transfer without test-time tuning.\nWe achieve this through constructing a style-aware encoder and a well-organized style dataset called StyleGallery.\nWith dedicated design for style learning, this style-aware encoder is trained to extract expressive style representation with decoupling training strategy, and StyleGallery enables the generalization ability.\nWe further employ a content-fusion encoder to enhance image-driven style transfer.\nWe highlight that, our approach, named StyleShot, is simple yet effective in mimicking various desired styles, i.e., 3D, flat, abstract or even fine-grained styles, without test-time tuning. Rigorous experiments validate that, StyleShot achieves superior performance across a wide range of styles compared to existing state-of-the-art methods.", "keywords": [ "diffusion model" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "QxbJYBZVbE", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Hao Jiang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Qi Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Rui Li", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Shengyu Ye", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Shijin Wang", "gender": "Male", "institution": "State Key Laboratory of Cognitive Intelligence", "country": "", "position": "Vice Dean" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 29, "n_ref": 66, "n_ref_all": 101, "n_fig": 13, "n_tab": 7, "L_tab": 5298, "n_element_tab": 223, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1673, "n_element_tab_1": 107, "formula_len_all": 18, "formula_len_all_1": 18, "len_all": 277723, "len_all_1": 54454, "len_abs": 1434, "len_title": 104, "len_sents": 61433, "len_sents_1": 22812, "n_sents": 477, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1440, "title": "CursorCore: Assist Programming through Aligning Anything", "abs": "Large language models have been successfully applied to programming assistance tasks, such as code completion, code insertion, and instructional code editing. However, these applications remain insufficiently automated and struggle to effectively integrate various types of information during the programming process, including coding history, current code, and user instructions. In this work, we propose a new conversational framework that comprehensively integrates these information sources, collect data to train our models and evaluate their performance. Firstly, to thoroughly evaluate how well models align with different types of information and the quality of their outputs, we introduce a new benchmark, APEval (Assist Programming Eval), to comprehensively assess the performance of models in programming assistance tasks. Then, for data collection, we develop a data generation pipeline, Programming-Instruct, which synthesizes training data from diverse sources, such as GitHub and online judge platforms. This pipeline can automatically generate various types of messages throughout the programming process. Finally, using this pipeline, we generate 219K samples, fine-tune multiple models, and develop the CursorCore series. We show that CursorCore outperforms other models of comparable size. This framework unifies applications such as inline chat and automated editing, contributes to the advancement of coding assistants.", "keywords": [ "AI-Assisted Programming", "Large Language Models for Code", "Code Benchmarks" ], "rating_list": [ 5, 6, 5, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "QwrnH32tJV", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Shaoan Xie", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Yujia Zheng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 115, "n_formula_1": 12, "n_ref_uni": 57, "n_ref": 120, "n_ref_all": 172, "n_fig": 12, "n_tab": 1, "L_tab": 26, "n_element_tab": 1, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 15, "n_element_tab_1": 2, "formula_len_all": 4695, "formula_len_all_1": 987, "len_all": 218204, "len_all_1": 82748, "len_abs": 1302, "len_title": 88, "len_sents": 69050, "len_sents_1": 41135, "n_sents": 652, "n_sents_1": 321, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1308, "title": "Provably Learning Concepts by Comparison", "abs": "We are born with the ability to learn concepts by comparing diverse observations. This helps us to understand the new world in a compositional manner and facilitates extrapolation, as objects naturally consist of multiple concepts. In this work, we argue that the cognitive mechanism of comparison, fundamental to human learning, is also vital for machines to recover true concepts underlying the data. This offers correctness guarantees for the field of concept learning, which, despite its impressive empirical successes, still lacks general theoretical support. Specifically, we aim to develop a theoretical framework for the identifiability of concepts with multiple classes of observations. We show that with sufficient diversity across classes, hidden concepts can be identified without assuming specific concept types, functional relations, or parametric generative models. Interestingly, even when conditions are not globally satisfied, we can still provide alternative guarantees for as many concepts as possible based on local comparisons, thereby extending the applicability of our theory to more flexible scenarios. Moreover, the hidden structure between classes and concepts can also be identified nonparametrically. We validate our theoretical results in both synthetic and real-world settings.", "keywords": [ "Concept Learning", "Compositional Learning", "Interpretability" ], "rating_list": [ 6, 3, 8 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 2, 2, 2 ] }, { "paper_id": "QwKieXLF6x", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Baoxin Li", "gender": "Not Specified", "institution": "Arizona State University", "country": "US", "position": "Full Professor" }, { "name": "Nupur Thakur", "gender": "Female", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Riti Paul", "gender": "Female", "institution": "Arizona State University", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 45, "n_ref": 74, "n_ref_all": 100, "n_fig": 8, "n_tab": 4, "L_tab": 329, "n_element_tab": 44, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 952, "n_element_tab_1": 45, "formula_len_all": 521, "formula_len_all_1": 262, "len_all": 159737, "len_all_1": 61913, "len_abs": 1184, "len_title": 127, "len_sents": 46302, "len_sents_1": 29049, "n_sents": 375, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1189, "title": "ZCTG: A Zero-Shot Framework for Automatic Video Chaptering and Title Generation", "abs": "In the vast landscape of video content, breaking down lengthy videos into chapters accompanied by concise, descriptive titles greatly enhances searchability and retrieval efficiency. While recent advancements in this field often incorporate multiple data modalities along with human-annotated chapter titles, access to such data, like speech transcripts or audio, is not always guaranteed. Moreover, the manual annotation of chapter titles is expensive and time-consuming. To address these challenges, we introduce ZCTG, a novel and unified zero-shot framework designed to generate video chapters and their concise titles for untrimmed videos. ZCTG utilizes the combined capabilities of scene graphs and Large Language Models (LLMs). The advantages of ZCTG are three-fold: 1) offers practical utility, relying solely on video data; 2) eliminates the need for detailed chapter title supervision; 3) exhibits excellent generalization capabilities in a completely zero-shot setting, without any training needed. We conduct an extensive evaluation on VidChapters-7M and GTEA datasets, which include videos of varying duration and domains, to demonstrate the efficacy of our proposed framework.", "keywords": [ "Video Chapter Generation", "Large Language Models", "Scene Graph" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Qvo0RBDEwD", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Leo Schwinn", "gender": "Male", "institution": "Technical University of Munich", "country": "DE", "position": "Postdoc" }, { "name": "Leon G\u00f6tz", "gender": "Male", "institution": "Volkswagen AG", "country": "", "position": "PhD student" }, { "name": "Marcel Kollovieh", "gender": "Male", "institution": "Department of Informatics, Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Stephan G\u00fcnnemann", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 40, "n_ref": 96, "n_ref_all": 111, "n_fig": 7, "n_tab": 6, "L_tab": 2891, "n_element_tab": 303, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 378, "n_element_tab_1": 29, "formula_len_all": 116, "formula_len_all_1": 99, "len_all": 141267, "len_all_1": 59354, "len_abs": 1410, "len_title": 96, "len_sents": 41781, "len_sents_1": 28975, "n_sents": 314, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1414, "title": "Efficient Time Series Processing for Transformers and State-Space Models through Token Merging", "abs": "Transformer architectures have shown promising results in time series processing. However, despite recent advances in subquadratic attention mechanisms or state-space models, processing very long sequences still imposes significant computational requirements. Token merging, which involves replacing multiple tokens with a single one calculated as their linear combination, has shown to considerably improve the throughput of vision transformer architectures while maintaining accuracy. In this work, we go beyond computer vision and perform the first investigations of token merging in time series analysis on both time series transformers and state-space models. We further introduce local merging, a domain-specific token merging algorithm that selectively combines tokens within a local neighborhood, achieving two major benefits: a) Local merging can adjust its the computational complexity from quadratic to linear based on the neighborhood size to effectively scale token merging to long sequences; b) Local merging is the first causal merging scheme enabling token merging in transformer decoders. Our comprehensive empirical evaluation demonstrates that token merging offers substantial computational benefits with minimal impact on accuracy across various models and datasets. On the recently proposed Chronos foundation model, we achieve accelerations up to 5400% with only minor accuracy degradations.", "keywords": [ "Token Merging", "Efficient Time Series Processing", "Transformers", "State-Space Models" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "Qv9TG9yDG0", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Sarit Khirirat", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Postdoc" }, { "name": "Peter Richtarik", "gender": "Male", "institution": "King Abdullah University of Science and Technology (KAUST)", "country": "SA", "position": "Full Professor" } ], "n_formula": 142, "n_formula_1": 32, "n_ref_uni": 56, "n_ref": 92, "n_ref_all": 119, "n_fig": 3, "n_tab": 4, "L_tab": 2126, "n_element_tab": 66, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 520, "n_element_tab_1": 88, "formula_len_all": 13245, "formula_len_all_1": 2495, "len_all": 177376, "len_all_1": 60859, "len_abs": 1488, "len_title": 130, "len_sents": 39396, "len_sents_1": 26870, "n_sents": 419, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1496, "title": "Error Feedback for Smooth and Nonsmooth Convex Optimization with Constant, Decreasing and Polyak Stepsizes", "abs": "Error feedback, originally proposed a decade ago by Seide et al (2014), is an immensely popular strategy for stabilizing the convergence behavior of distributed algorithms employing communication compression via the application of contractive compression operators, such as greedy and random sparsification, quantization, and low-rank approximation. While our algorithmic and theoretical understanding of error feedback has grown immensely over the years, several important considerations remained elusive. For example, the theory of error feedback is fully focused on the smooth convex and nonconvex regimes, and results in the nonsmooth convex setting are limited. This is not a coincidence: Error feedback works when the gradients converge, and this is not necessarily the case in the nonsmooth setting. Further, existing stepsize rules for error feedback are limited to constant schedules; a by-product of the current theoretical approach to analyzing error feedback. By modifying the algorithmic design of error feedback, we are able to resolve these issues. In particular, we provide a comprehensive analysis covering both the smooth and nonsmooth convex regimes, and give support for constant, decreasing and adaptive (Polyak-type) stepsizes. This is the first time such results are obtained. In particular, this is the first time adaptive stepsizes have successfully been combined with compression mechanisms. Our theoretical results are corroborated with suitable numerical experiments.", "keywords": [ "Error feedback; Polyak stepsize; Communication-efficient optimization" ], "rating_list": [ 3, 8, 3, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 3, 4, 2 ] }, { "paper_id": "QuGnjxfLBH", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Jiyoun Lim", "gender": "Female", "institution": "Electronics and Telecommunications Research Institute", "country": "", "position": "Principal Researcher" }, { "name": "SeungBum Ha", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sung Whan Yoon", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Taehwan Lee", "gender": "unknown", "institution": "Ulsan National Institute of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 19, "n_ref": 47, "n_ref_all": 72, "n_fig": 4, "n_tab": 9, "L_tab": 4646, "n_element_tab": 419, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 3098, "n_element_tab_1": 38, "formula_len_all": 394, "formula_len_all_1": 187, "len_all": 160953, "len_all_1": 56828, "len_abs": 1590, "len_title": 135, "len_sents": 41805, "len_sents_1": 27314, "n_sents": 288, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1605, "title": "Benchmarking Federated Learning for Semantic Datasets: Federated Scene Graph Generation", "abs": "Federated learning (FL) has recently garnered attention as a decentralized training framework that enables the learning of deep models from locally distributed samples while keeping the data privacy.\nBuilt upon the framework, immense efforts have been made to establish FL benchmarks, which provide rigorous evaluation settings that aim to control data heterogeneity across clients.\nPrior efforts have mainly focused on handling relatively simple classification tasks, where each sample is annotated with a one-hot label, such as MNIST, CIFAR, LEAF benchmark, etc.\nHowever, little attention has been paid to demonstrating an FL benchmark that handles complicated semantics, where each sample encompasses diverse semantic information from multiple labels, such as Scene Graph Generation / Panoptic Scene Graph Generation (SGG/PSG) with objects, predicates, and relations between objects.\nBecause the existing benchmark is designed to distribute data in a narrow view of a single semantic, e.g., a one-hot label, managing the complicated $\\textit{semantic heterogeneity}$ across clients when formalizing FL benchmarks is non-trivial.\nIn this paper, we propose a benchmark process to establish an FL benchmark with controllable semantic heterogeneity across clients: two key steps are i) data clustering with semantics and ii) data distributing via controllable semantic heterogeneity across clients.\nAs a proof of concept, we first construct a federated SGG/PSG benchmark, which demonstrates the efficacy of the existing PSG methods in an FL setting with controllable semantic heterogeneity of scene graphs.", "keywords": [ "Federated Learning", "Decentralized Learning", "Scene Graph Generation" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "QtZsTaqRRE", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Alexandre Capone", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Aman Mehra", "gender": "unknown", "institution": "Tesla", "country": "US", "position": "Researcher" }, { "name": "Jeff Schneider", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 26, "n_ref": 64, "n_ref_all": 86, "n_fig": 13, "n_tab": 1, "L_tab": 285, "n_element_tab": 7, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 695, "n_element_tab_1": 15, "formula_len_all": 480, "formula_len_all_1": 426, "len_all": 108578, "len_all_1": 60036, "len_abs": 1498, "len_title": 131, "len_sents": 38940, "len_sents_1": 30277, "n_sents": 262, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1467, "title": "Accelerated Online Reinforcement Learning using Auxiliary Start State Distributions", "abs": "Learning a robust policy that is performant across the state space, in a sample efficient manner, is a long-standing problem in online reinforcement learning (RL). This challenge arises from the inability of algorithms to explore the environment efficiently. Most attempts at efficient exploration tackle this problem in a setting where learning begins from scratch, without prior information available to bootstrap learning. However, such approaches often fail to fully leverage expert demonstrations and simulators that can reset to arbitrary states. These affordances are valuable resources that offer enormous potential to guide exploration and speed up learning. In this paper, we explore how a small number of expert demonstrations and a simulator allowing arbitrary resets can accelerate learning during online RL. We show that by leveraging expert state information to form an auxiliary start state distribution, we significantly improve sample efficiency. Specifically, we show that using a notion of safety to inform the choice of auxiliary distribution significantly accelerates learning. We highlight the effectiveness of our approach by matching or exceeding state-of-the-art performance in sparse reward and dense reward setups, even when competing with algorithms with access to expert actions and rewards. Moreover, we find that the improved exploration ability facilitates learning more robust policies in spare reward, hard exploration environments.", "keywords": [ "reinforcement learning", "sample efficiency", "robustness" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "QtSw71HJ6M", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Junho Shin", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Myungjoo Kang", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "SoHyung Kim", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 38, "n_ref": 60, "n_ref_all": 63, "n_fig": 2, "n_tab": 6, "L_tab": 2511, "n_element_tab": 292, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1653, "n_element_tab_1": 212, "formula_len_all": 833, "formula_len_all_1": 711, "len_all": 106538, "len_all_1": 54686, "len_abs": 1396, "len_title": 130, "len_sents": 30391, "len_sents_1": 24893, "n_sents": 192, "n_sents_1": 149, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1401, "title": "Mitigating Overestimation in Offline Reinforcement Learning with Anomaly Detection", "abs": "Reinforcement Learning (RL) encounters substantial challenges in real-world applications, due to the time-consuming, costly, and risky nature of interacting with the environment. Offline Reinforcement Learning addresses this limitation by training models on static datasets, allowing an optimal policy to be learned from pre-collected data without requiring additional interactions with the environment. However, in this setting, when the agent queries actions outside the training data distribution, it can lead to overestimation of Q-values for OOD (Out-of-distribution) actions, ultimately hindering policy optimization. Previous works attempted to address this problem using explicit constraints such as penalty terms or support restriction. But these methods often fail to identify OOD actions or result in overly conservative Q-value estimates. We propose a novel solution that adjusts weights during training by using an anomaly detection model to identify the distribution of the offline dataset and employing anomaly scores to guide the offline RL process. Our method(RLAD) not only effectively mitigates the overestimation of OOD actions but also achieves near state-of-the-art performance on continuous D4RL tasks. Additionally, this framework is highly flexible, allowing for integration with various off-policy or offline RL algorithms and Anomaly Detection models to enhance performance.", "keywords": [ "Reinforcement Learning", "Offline Reinforcement Learning", "Anomaly Detection" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 1, 3, 1, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "QtKYYatG3Z", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Ashesh Rambachan", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Keyon Vafa", "gender": "Male", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Peter G. Chang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Sendhil Mullainathan", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 75, "n_ref_all": 96, "n_fig": 5, "n_tab": 13, "L_tab": 4341, "n_element_tab": 512, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2070, "n_element_tab_1": 177, "formula_len_all": 570, "formula_len_all_1": 358, "len_all": 165468, "len_all_1": 77792, "len_abs": 1183, "len_title": 55, "len_sents": 61743, "len_sents_1": 36735, "n_sents": 497, "n_sents_1": 289, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1193, "title": "Evaluating the World Models Used by Pretrained Learners", "abs": "A common approach for assessing whether generative models develop world models is by studying the behavior of fixed models. However, many of the benefits of having a world model arise when transferring a model to new tasks (e.g. few- shot learning). In this paper, we ask: what does it mean to test if a _learner_ has a world model embodied in it? We consider a simple definition of a true world model: a mapping from inputs to states. We introduce a procedure that assesses a learner\u2019s world model by measuring its inductive bias when transferring to new tasks. This inductive bias can be measured in two distinct dimensions: does a learner extrapolate to new data by building functions of state, and to what degree do these functions capture the full state? We use this procedure to study the degree to which pretrained models extrapolate to new tasks based on state. We find that models that perform very well on next-token prediction can extrapolate to new tasks with very little inductive bias toward state. We conclude by assessing the possibility that these models learn bundles of heuristics that enable them to perform well on next-token prediction despite preserving little of state.", "keywords": [ "large language models", "world models", "transfer learning", "evaluation" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "QtJiPhqnsV", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Bo Lei", "gender": "Male", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "PhD student" }, { "name": "Wei Lan", "gender": "Male", "institution": "Southwest University of Finance and Economics", "country": "", "position": "Full Professor" }, { "name": "Xinyan Fan", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 74, "n_formula_1": 26, "n_ref_uni": 34, "n_ref": 62, "n_ref_all": 86, "n_fig": 0, "n_tab": 7, "L_tab": 3913, "n_element_tab": 417, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 263, "n_element_tab_1": 66, "formula_len_all": 11962, "formula_len_all_1": 3956, "len_all": 136576, "len_all_1": 76142, "len_abs": 1181, "len_title": 117, "len_sents": 34320, "len_sents_1": 25305, "n_sents": 322, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1187, "title": "Inferences on Covariance Matrix with Blockwise Correlation Structure", "abs": "Utilizing the sample moments of variable means within groups, we develop a novel closed-form estimator for blockwise correlation matrix of $p$ variables.\nWhen the block number and group memberships of the variables are known, we demonstrate the asymptotic normality of parameter estimators and establish the stochastic convergence rate of the estimated blockwise correlation matrix and corresponding estimated covariance matrix, under certain moment conditions.\nThe method ensures positive semi-definiteness of the estimated covariance matrix without requiring a predetermined variable order, and can be applicable for high-dimensional data.\nMoreover, to estimate the number of blocks and recover their memberships, respectively,\nwe employ the ridge-type ratio criterion and spectral clustering, and establish their consistency. Based on this, we extend the aforementioned properties of the asymptotic normality and stochastic convergence rate to the scenario where the group memberships are unknown and the block number is given.\nExtensive simulations and an empirical study of stock returns in the Chinese stock market are analyzed to illustrate the usefulness of our proposed methods.", "keywords": [ "Blockwise Correlation Matrix Estimation; Ridge-type Ratio Criterion; Spectral Clustering; Covariance Matrix." ], "rating_list": [ 6, 5, 3, 6, 5 ], "soundness_list": [ 3, 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 1, 3, 2 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 2, 4, 3 ] }, { "paper_id": "Qt5sBi0u7I", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Haoyu Liu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Tingyu Zhu", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Zeyu Zheng", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Associate Professor" }, { "name": "Zhimin Jiang", "gender": "Male", "institution": "Central China Normal University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 84, "n_formula_1": 23, "n_ref_uni": 27, "n_ref": 64, "n_ref_all": 87, "n_fig": 4, "n_tab": 4, "L_tab": 1921, "n_element_tab": 127, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 239, "n_element_tab_1": 20, "formula_len_all": 6106, "formula_len_all_1": 1676, "len_all": 166702, "len_all_1": 62754, "len_abs": 898, "len_title": 134, "len_sents": 55442, "len_sents_1": 28223, "n_sents": 454, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 899, "title": "Symbolic Music Generation with Fine-grained Interactive Textural Guidance", "abs": "The problem of symbolic music generation presents unique challenges due to the combination of limited data availability and the need for high precision in note pitch. To address these issues, we introduce an efficient Fine-grained Sampling Guidance (FTG) approach within diffusion models. FTG guides the diffusion models to generate music that aligns more closely with the control and intent of human composers, thereby improving the accuracy and quality of music generation. This method empowers diffusion models to excel in advanced applications such as improvisation, and interactive music creation. We derive theoretical characterizations for both the challenges in symbolic music generation and the effect of the FTG approach. We provide numerical experiments, subjective evaluation and a demo page for interactive music generation with user input to showcase the effectiveness of our approach.", "keywords": [ "Symbolic Music Generation; Guided Diffusion Models" ], "rating_list": [ 3, 5, 3, 8, 6 ], "soundness_list": [ 2, 3, 1, 3, 3 ], "presentation_list": [ 1, 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 2, 4, 3 ] }, { "paper_id": "QstnrTlPyr", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Jing Bai", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Weixi Xiang", "gender": "Male", "institution": "Chinese Academy of Agricultural Sciences", "country": "CN", "position": "MS student" }, { "name": "Xiujuan Chai", "gender": "Female", "institution": "AII", "country": "", "position": "Full Professor" }, { "name": "Xueting Han", "gender": "unknown", "institution": "Microsoft Research Asia", "country": "CN", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 39, "n_ref": 55, "n_ref_all": 71, "n_fig": 6, "n_tab": 6, "L_tab": 1642, "n_element_tab": 151, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 562, "n_element_tab_1": 69, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 137435, "len_all_1": 59111, "len_abs": 1400, "len_title": 120, "len_sents": 42309, "len_sents_1": 29749, "n_sents": 274, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1405, "title": "BSM: Small but Powerful Biological Sequence Model for Genes and Proteins", "abs": "Modeling biological sequences such as DNA, RNA, and proteins is crucial for understanding complex processes like gene regulation and protein synthesis. However, most current models either focus on a single type or treat multiple types of data separately, limiting their ability to capture cross-modal relationships. We propose that by learning the relationships between these modalities, the model can enhance its understanding of each type. To address this, we introduce BSM, a small but powerful mixed-modal biological sequence foundation model, trained on three types of data: RefSeq, Gene Related Sequences, and interleaved biological sequences from the web. These datasets capture the genetic flow, gene-protein relationships, and the natural co-occurrence of diverse biological data, respectively. By training on mixed-modal data, BSM significantly enhances learning efficiency and cross-modal representation, outperforming models trained solely on unimodal data. With only 110M parameters, BSM achieves performance comparable to much larger models across both single-modal and mixed-modal tasks, and uniquely demonstrates in-context learning capability for mixed-modal tasks, which is absent in existing models. Further scaling to 270M parameters demonstrates even greater performance gains, highlighting the potential of BSM as a significant advancement in multimodal biological sequence modeling.", "keywords": [ "Biological Sequence Modeling", "Language Models", "Gene", "Protein", "Mixed-Modal" ], "rating_list": [ 5, 1, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "QsA3YzNUxA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Cho-Jui Hsieh", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Hengguang Zhou", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "MS student" }, { "name": "Minhao Cheng", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Ruochen Wang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Tianyi Zhou", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Xirui Li", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 27, "n_ref": 69, "n_ref_all": 127, "n_fig": 28, "n_tab": 18, "L_tab": 3716, "n_element_tab": 443, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 16, "n_element_tab_1": 1, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 290660, "len_all_1": 60688, "len_abs": 2067, "len_title": 113, "len_sents": 87875, "len_sents_1": 29593, "n_sents": 734, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1746, "title": "Is Your Multimodal Language Model Oversensitive to Safe Queries?", "abs": "Humans are prone to cognitive distortions \u2014 biased thinking patterns that lead to exaggerated responses to specific stimuli, albeit in very different contexts.\nThis paper demonstrates that advanced Multimodal Large Language Models (MLLMs) exhibit similar tendencies.\nWhile these models are designed to respond queries under safety mechanism, they sometimes reject harmless queries in the presence of certain visual stimuli, disregarding the benign nature of their contexts.\nAs the initial step in investigating this behavior, we identify three representative types of stimuli that trigger the oversensitivity of existing MLLMs: $\\textbf{\\textit{Exaggerated Risk}}$, $\\textbf{\\textit{Negated Harm}}$, and $\\textbf{\\textit{Counterintuitive Interpretation}}$.\nTo systematically evaluate MLLMs' oversensitivity to these stimuli, we propose the $\\textbf{M}$ultimodal $\\textbf{O}$ver$\\textbf{S}$en$\\textbf{S}$itivity $\\textbf{Bench}$mark (MOSSBench).\nThis toolkit consists of 300 manually collected benign multimodal queries, cross-verified by third-party reviewers (AMT).\nEmpirical studies using MOSSBench on 20 MLLMs reveal several insights:\n(1). Oversensitivity is prevalent among SOTA MLLMs, with refusal rates reaching up to $\\textbf{76}$\\% for harmless queries.\n(2). Safer models are more oversensitive: increasing safety may inadvertently raise caution and conservatism in the model\u2019s responses.\n(3). Different types of stimuli tend to cause errors at specific stages \u2014 perception, intent reasoning, and safety judgement \u2014 in the response process of MLLMs.\nThese findings highlight the need for refined safety mechanisms that balance caution with contextually appropriate responses, improving the reliability of MLLMs in real-world applications.", "keywords": [ "Safety", "Adversarial Robustness", "Multimodal Large Language Models", "Alignment" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "Qro97zWC29", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Francesco Locatello", "gender": "Male", "institution": "Institute of Science and Technology", "country": "AT", "position": "Assistant Professor" }, { "name": "Gertjan J. Burghouts", "gender": "Male", "institution": "TNO", "country": "NL", "position": "Researcher" }, { "name": "Mohammadreza Salehi", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Valentinos Pariza", "gender": "Male", "institution": "niversity of Technology Nuremberg", "country": "DE", "position": "PhD student" }, { "name": "Yuki Asano", "gender": "Male", "institution": "University of Technology Nuremberg", "country": "DE", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 62, "n_ref": 195, "n_ref_all": 231, "n_fig": 7, "n_tab": 10, "L_tab": 3552, "n_element_tab": 432, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 443, "n_element_tab_1": 25, "formula_len_all": 656, "formula_len_all_1": 656, "len_all": 200755, "len_all_1": 60124, "len_abs": 1121, "len_title": 0, "len_sents": 59801, "len_sents_1": 27838, "n_sents": 431, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1123, "title": "Near, far: Patch-ordering enhances vision foundation models' scene understanding", "abs": "We introduce NeCo: Patch Neighbor Consistency, a novel self-supervised training loss that enforces patch-level nearest neighbor consistency across a student and teacher model. Compared to contrastive approaches that only yield binary learning signals, i.e. \"attract\" and \"repel\", this approach benefits from the more fine-grained learning signal of sorting spatially dense features relative to reference patches. Our method leverages differentiable sorting applied on top of pretrained representations, such as DINOv2-registers to bootstrap the learning signal and further improve upon them. This dense post-pretraining leads to superior performance across various models and datasets, despite requiring only 19 hours on a single GPU. This method generates high-quality dense feature encoders and establishes several new state-of-the-art results such as +5.5 % and +6% for non-parametric in-context semantic segmentation on ADE20k and Pascal VOC, +7.2% and +5.7% for linear segmentation evaluations on COCO-Things and -Stuff and improvements in the 3D understanding of multi-view consistency on SPair-71k, by more than 10%.", "keywords": [ "Dense self-supervised learning", "Self-supervised learning", "Unsupervised semantic segmentation", "In-context scene understanding" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "QrhB9HcgnL", "primary_area": "applications to neuroscience & cognitive science", "n_author": 10, "author_info_list": [ { "name": "Aidong Zhang", "gender": "Female", "institution": "University of Virginia", "country": "US", "position": "Full Professor" }, { "name": "Bolin Lai", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jianguo Cao", "gender": "Male", "institution": "Shenzhen Children's Hospital", "country": "CN", "position": "Chief Physician" }, { "name": "Jintai Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Joerg Heintz", "gender": "Male", "institution": "University of Illinois Urbana Champaign", "country": "", "position": "Researcher" }, { "name": "Meihuan Huang", "gender": "Female", "institution": "Shenzhen Children's Hospital", "country": "CN", "position": "Researcher" }, { "name": "Wenqian Ye", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Xu Cao", "gender": "Non-Binary", "institution": "Department of Computer Science, University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Yunsheng Ma", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "James Rehg", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 67, "n_ref": 126, "n_ref_all": 155, "n_fig": 9, "n_tab": 7, "L_tab": 2378, "n_element_tab": 225, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 2521, "n_element_tab_1": 214, "formula_len_all": 140, "formula_len_all_1": 92, "len_all": 221883, "len_all_1": 64757, "len_abs": 1365, "len_title": 118, "len_sents": 58665, "len_sents_1": 28633, "n_sents": 444, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1372, "title": "Benchmarking Visual Cognition of Multimodal LLMs via Matrix Reasoning", "abs": "Recently, Multimodal Large Language Models (MLLMs) and Vision Language Models (VLMs) have shown great promise in language-guided perceptual tasks such as recognition, segmentation, and object detection. However, their effectiveness in addressing visual cognition problems that require high-level multi-image reasoning and visual working memory is not well-established. One such challenge is matrix reasoning -- the cognitive ability to discern relationships among patterns in a set of images and extrapolate to predict subsequent patterns. This skill is crucial during the early neurodevelopmental stages of children. Inspired by the matrix reasoning tasks in Raven\u2019s Progressive Matrices (RPM) and Wechsler Intelligence Scale for Children (WISC), we propose a new dataset MaRs-VQA and a new benchmark VCog-Bench to evaluate the zero-shot visual cognition capability of MLLMs and compare their performance with existing human visual cognition investigation. Our comparative experiments with different open-source and closed-source MLLMs on the VCog-Bench revealed a gap between MLLMs and human intelligence, highlighting the visual cognitive limitations of current MLLMs. We believe that the public release of VCog-Bench, consisting of MaRs-VQA, and the inference pipeline will drive progress toward the next generation of MLLMs with human-like visual cognition abilities.", "keywords": [ "Visual Cognition", "Matrix Reasoning", "Psychometrics", "Visual Reasoning", "Multimodal LLMs" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "QrTvFCa4nX", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Bingxin Xu", "gender": "Female", "institution": "Illinois Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Kurt Keutzer", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Mu Cai", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin, Madison", "country": "US", "position": "PhD student" }, { "name": "Weitai Kang", "gender": "Male", "institution": "University of Illinois at Chicago", "country": "US", "position": "PhD student" }, { "name": "Yan Yan", "gender": "Male", "institution": "University of Illinois Chicago", "country": "US", "position": "Associate Professor" }, { "name": "Yuheng Li", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Yuzhang Shang", "gender": "Male", "institution": "Illinois Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zehao Wen", "gender": "Male", "institution": "Shenzhen College of International Education", "country": "CN", "position": "Undergrad student" }, { "name": "Zhen Dong", "gender": "Male", "institution": "Nexusflow.ai Inc", "country": "US", "position": "Principal Researcher" }, { "name": "Yong Jae Lee", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 29, "n_ref": 74, "n_ref_all": 82, "n_fig": 3, "n_tab": 6, "L_tab": 1717, "n_element_tab": 236, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2342, "n_element_tab_1": 124, "formula_len_all": 746, "formula_len_all_1": 744, "len_all": 122275, "len_all_1": 60081, "len_abs": 1945, "len_title": 127, "len_sents": 33824, "len_sents_1": 28250, "n_sents": 255, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1412, "title": "Interpolating Video-LLMs: Toward Longer-sequence LMMs in a Training-free Manner", "abs": "Advancements in Large Language Models (LLMs) inspire various strategies for integrating video modalities. \nA key approach is Video-LLMs, which incorporate an optimizable interface linking sophisticated video encoders to LLMs. \nHowever, due to computation and data limitations, these Video-LLMs are typically pre-trained to process only short videos, limiting their broader application for understanding longer video content. Additionally, fine-tuning Video-LLMs to handle longer videos is cost-prohibitive.\nConsequently, it becomes essential to explore the interpolation of Video-LLMs under a completely training-free setting. In this paper, we first identify the primary challenges in interpolating Video-LLMs: (1) the video encoder and modality alignment projector are fixed, preventing the integration of additional frames into Video-LLMs, and (2) the LLM backbone is limited in its content length capabilities, which complicates the processing of an increased number of video tokens.\nTo address these challenges, we propose a specific INTerPolation method for Video-LLMs (INTP-Video-LLMs). We introduce an alternative video token rearrangement technique that circumvents limitations imposed by the fixed video encoder and alignment projector. Furthermore, we introduce a training-free LLM context window extension method to enable Video-LLMs to understand a correspondingly increased number of visual tokens.", "keywords": [ "Video Understanding" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "Qr9TjKYzjl", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Ling Pan", "gender": "Female", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Miranda Anna Christ", "gender": "unknown", "institution": "Alfr\u00e9d R\u00e9nyi Institute of Mathematics", "country": "HU", "position": "Intern" }, { "name": "Zarif Ikram", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Visiting scholar" }, { "name": "Dianbo Liu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 7, "n_ref_uni": 59, "n_ref": 85, "n_ref_all": 108, "n_fig": 11, "n_tab": 2, "L_tab": 731, "n_element_tab": 50, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 568, "formula_len_all_1": 585, "len_all": 139438, "len_all_1": 56010, "len_abs": 925, "len_title": 109, "len_sents": 37699, "len_sents_1": 26687, "n_sents": 311, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 928, "title": "Small features matter: Robust representation for world models", "abs": "In Model-Based Reinforcement Learning (MBRL), an agent learns to make decisions by building a world model that predicts the environment's dynamics. The accuracy of this world model is crucial for generalizability and sample efficiency. Many works rely on pixel-level reconstruction, which may focus on irrelevant, exogenous features over minor, but key information. In this work, to encourage the world model to focus on important task related information, we propose an augmentation to the world model training using a temporal prediction loss in the embedding space as an auxiliary loss. Building our method on the DreamerV3 architecture, we improve sample efficiency and stability by learning better representation for world model and policy training. We evaluate our method on the Atari100k and Distracting Control Suite benchmarks, demonstrating significant improvements in world model quality and overall MBRL performance.", "keywords": [ "Representation learning", "model based reinforcement learning", "world models" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "QqziJAdev9", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Bolin Ding", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Senior Director" }, { "name": "Jiancan Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Postdoc" }, { "name": "Jinyang Gao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Junkang Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Rong Jin", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Principal Researcher" }, { "name": "Xiang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xiangnan He", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Zhengyi Yang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "xue wang", "gender": "Male", "institution": "Alibaba Group US", "country": "US", "position": "Researcher" } ], "n_formula": 68, "n_formula_1": 32, "n_ref_uni": 51, "n_ref": 85, "n_ref_all": 104, "n_fig": 9, "n_tab": 7, "L_tab": 2958, "n_element_tab": 409, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1751, "n_element_tab_1": 318, "formula_len_all": 6541, "formula_len_all_1": 2603, "len_all": 189282, "len_all_1": 65841, "len_abs": 1584, "len_title": 122, "len_sents": 42356, "len_sents_1": 26992, "n_sents": 372, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1552, "title": "$\\alpha$-DPO: Adaptive Reward Margin is What Direct Preference Optimization Needs", "abs": "Aligning large language models (LLMs) with human values and intentions is crucial for their utility, honesty, and safety. Reinforcement learning from human feedback (RLHF) is a popular approach to achieve this alignment, but it faces challenges in computational efficiency and training stability. Recent methods like Direct Preference Optimization (DPO) and Simple Preference Optimization (SimPO) have proposed offline alternatives to RLHF, simplifying the process by reparameterizing the reward function. However, DPO depends on a potentially suboptimal reference model,\nand SimPO's assumption of a fixed target reward margin may lead to suboptimal decisions in diverse data settings.\nIn this work, we propose \\(\\alpha\\)-DPO, an adaptive preference optimization algorithm designed to address these limitations by introducing a dynamic reward margin. Specifically, \\(\\alpha\\)-DPO employs an adaptive preference distribution, balancing the policy model and the reference model to achieve personalized reward margins. We provide theoretical guarantees for \\(\\alpha\\)-DPO, demonstrating its effectiveness as a surrogate optimization objective and its ability to balance alignment and diversity through KL divergence control. Empirical evaluations on AlpacaEval 2 and Arena-Hard show that \\(\\alpha\\)-DPO consistently outperforms DPO and SimPO across various model settings, establishing it as a robust approach for fine-tuning LLMs. Our method achieves significant improvements in win rates, highlighting its potential as a powerful tool for LLM alignment.", "keywords": [ "Direct Preference Optimization", "LLM's alignment" ], "rating_list": [ 5, 6, 3, 5, 6 ], "soundness_list": [ 3, 4, 1, 2, 2 ], "presentation_list": [ 3, 4, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 4, 4 ] }, { "paper_id": "QqypKtKiWX", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Amir AbdelWahab", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "John Sapp", "gender": "Male", "institution": "Dalhousie University", "country": "CA", "position": "Full Professor" }, { "name": "Shijie Zhou", "gender": "Male", "institution": "Worcester Polytechnic Institute", "country": "US", "position": "Assistant Professor" }, { "name": "Zixuan Ni", "gender": "Male", "institution": "Zhejiang University", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 22, "n_ref": 57, "n_ref_all": 71, "n_fig": 8, "n_tab": 5, "L_tab": 5195, "n_element_tab": 109, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 206, "n_element_tab_1": 48, "formula_len_all": 17, "formula_len_all_1": 17, "len_all": 117999, "len_all_1": 42369, "len_abs": 1402, "len_title": 145, "len_sents": 30298, "len_sents_1": 20200, "n_sents": 189, "n_sents_1": 144, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1406, "title": "A Simple Data-Parameters Balancing Framework for Early Ventricular Activation Origin Localization", "abs": "Accurately identifying the site of origin (SoO) of early ventricular activation is crucial for catheter ablation, an effective therapeutic option for treating ventricular arrhythmia. However, due to the limited availability of clinical data and the errors introduced during data preprocessing, achieving precise localization remains a challenge. While deep learning models offer an end-to-end approach for data input in the ECG field, they often suffer from overfitting caused by limited training data, hindering continuous performance improvement. This paper proposes a Simple data-parameters Balancing framework for early ventricular activation Origin Localization (SimBOL). By using onset-based data augmentation, the SimBOL method expands the training data derived from clinical samples. The framework utilizes a small-scale 1D convolution model that balances the relationship between available training data and model complexity, effectively mitigating overfitting and eliminating the need for extensive data preprocessing.SimBOL achieves a localization error as low as 9.83 mm, which meets clinical acceptable localization error < 10 mm and outperforming existing methods in predicting the SoO of early ventricular activation. The discussion about data augmentation and model architecture on ECG signal processing, offering new insights into optimizing deep learning applications for ECG-based tasks.", "keywords": [ "Early Ventricular Activation Origin Localization;" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "QowsEic1sc", "primary_area": "generative models", "n_author": 11, "author_info_list": [ { "name": "Enshu Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Guohao Dai", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Huazhong Yang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Junyi Zhu", "gender": "Male", "institution": "Samsung", "country": "GB", "position": "Researcher" }, { "name": "Sergey Yekhanin", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Shengen Yan", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Shuaiqi Wang", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Xuefei Ning", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Research Assistant Professor" }, { "name": "Yu Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Zinan Lin", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Senior Researcher" }, { "name": "Matthew Blaschko", "gender": "Male", "institution": "KU Leuven", "country": "BE", "position": "Associate Professor" } ], "n_formula": 63, "n_formula_1": 13, "n_ref_uni": 42, "n_ref": 146, "n_ref_all": 194, "n_fig": 16, "n_tab": 21, "L_tab": 6111, "n_element_tab": 1049, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3137, "n_element_tab_1": 470, "formula_len_all": 5758, "formula_len_all_1": 887, "len_all": 242592, "len_all_1": 79964, "len_abs": 1752, "len_title": 134, "len_sents": 72131, "len_sents_1": 31974, "n_sents": 684, "n_sents_1": 294, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1466, "title": "Linear Combination of Saved Checkpoints Makes Consistency and Diffusion Models Better", "abs": "Diffusion Models (DM) and Consistency Models (CM) are two types of popular generative models with good generation quality on various tasks. When training DM and CM, intermediate weight checkpoints are not fully utilized and only the last converged checkpoint is used. In this work, we find proper checkpoint merging can significantly improve the training convergence and final performance. Specifically, we propose LCSC, a simple but effective and efficient method to enhance the performance of DM and CM, by combining checkpoints along the training trajectory with coefficients deduced from evolutionary search. We demonstrate the value of LCSC through two use cases: (a) Reducing training cost. With LCSC, we only need to train DM/CM with fewer number of iterations and/or lower batch sizes to obtain comparable sample quality with the fully trained model. For example, LCSC achieves considerable training speedups for CM (23$\\times$ on CIFAR-10 and 15$\\times$ on ImageNet-64). (b) Enhancing pre-trained models. When full training is already done, LCSC can further improve the generation quality or efficiency of the final converged models. For example, LCSC achieves better FID using 1 number of function evaluation (NFE) than the base model with 2 NFE on consistency distillation, and decreases the NFE of DM from 15 to 9 while maintaining the generation quality. Applying LCSC to large text-to-image models, we also observe clearly enhanced generation quality.", "keywords": [ "Model merging", "consistency model", "diffusion model" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 2, 2 ] }, { "paper_id": "Qohoh5x1Jm", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Jiaochenchen", "gender": "unknown", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Ning Luo", "gender": "unknown", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Engineer" }, { "name": "Peng Jiang", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Vice President" }, { "name": "Xialong Liu", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Xiaocheng Liu", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Yanxiang Zeng", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Yongxiang Tang", "gender": "unknown", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Yuanpengjia", "gender": "unknown", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "yanhua chengyan", "gender": "unknown", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" } ], "n_formula": 44, "n_formula_1": 29, "n_ref_uni": 33, "n_ref": 47, "n_ref_all": 79, "n_fig": 8, "n_tab": 15, "L_tab": 3923, "n_element_tab": 390, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 310, "n_element_tab_1": 23, "formula_len_all": 2973, "formula_len_all_1": 2265, "len_all": 138548, "len_all_1": 63728, "len_abs": 1172, "len_title": 121, "len_sents": 36327, "len_sents_1": 26393, "n_sents": 302, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1186, "title": "Reformulating Strict Monotonic Probabilities with a Generative Cost Model", "abs": "In numerous machine learning contexts, the relationship between input variables and predicted outputs is not only statistically significant but also strictly monotonic. Conventional approaches to ensuring monotonicity focus primarily on construction or regularization methods. This paper establishes that the problem of strict monotonic probability can be interpreted as a comparison between an observable revenue variable and a latent cost variable. This insight allows us to reformulate the original monotonicity challenge into modeling the latent cost variable and estimating its distribution. To address this issue, we introduce a generative model for the latent cost variable, called the Generative Cost Model (\\textbf{GCM}), and derive a corresponding loss function. We further enhance the estimation of latent variables using variational inference, which reformulate our loss function accordingly. Lastly, we validate our approach through a numerical simulation of quantile regression and several experiments on public datasets, demonstrating that our method significantly outperforms traditional techniques. The code of GCM is available in https://github.com/iclr-2025-4464/GCM.", "keywords": [ "monotonic model", "variational inference", "generative model" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "QogcGNXJVw", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Federico Adolfi", "gender": "Male", "institution": "ESI Neuroscience, Max-Planck Society", "country": "DE", "position": "Postdoc" }, { "name": "Martina G. Vilas", "gender": "Female", "institution": "Computer Science Department, Goethe University", "country": "", "position": "PhD student" }, { "name": "Todd Wareham", "gender": "Male", "institution": "Memorial University of Newfoundland", "country": "CA", "position": "Full Professor" } ], "n_formula": 110, "n_formula_1": 4, "n_ref_uni": 75, "n_ref": 203, "n_ref_all": 240, "n_fig": 2, "n_tab": 19, "L_tab": 4384, "n_element_tab": 272, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1829, "n_element_tab_1": 126, "formula_len_all": 11950, "formula_len_all_1": 464, "len_all": 492688, "len_all_1": 76833, "len_abs": 1700, "len_title": 124, "len_sents": 179240, "len_sents_1": 37296, "n_sents": 1497, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 76, "L_abs": 1856, "title": "The Computational Complexity of Circuit Discovery for Inner Interpretability", "abs": "Many proposed applications of neural networks in machine learning, cognitive/brain science, and society hinge on the feasibility of inner interpretability via circuit discovery. This calls for empirical and theoretical explorations of viable algorithmic options. Despite advances in the design and testing of heuristics, there are concerns about their scalability and faithfulness at a time when we lack understanding of the complexity properties of the problems they are deployed to solve. To address this, we study circuit discovery with classical and parameterized computational complexity theory: (1) we describe a conceptual scaffolding to reason about circuit finding queries in terms of affordances for description, explanation, prediction and control; (2) we formalize a comprehensive set of queries that capture mechanistic explanation, and propose a formal framework for their analysis; (3) we use it to settle the complexity of many query variants and relaxations of practical interest on multi-layer perceptrons (part of, e.g., transformers). Our findings reveal a challenging complexity landscape. Many queries are intractable (NP-hard, $\\Sigma^p_2$-hard), remain fixed-parameter intractable (W[1]-hard) when constraining model/circuit features (e.g., depth), and are inapproximable under additive, multiplicative, and probabilistic approximation schemes. To navigate this landscape, we prove there exist transformations to tackle some of these hard problems (NP- vs. $\\Sigma^p_2$-complete) with better-understood heuristics, and prove the tractability (PTIME) or fixed-parameter tractability (FPT) of more modest queries which retain useful affordances. This framework allows us to understand the scope and limits of interpretability queries, explore viable options, and compare their resource demands among existing and future architectures.", "keywords": [ "inner interpretability", "mechanistic interpretability", "circuit discovery", "computational complexity", "parameterized complexity" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "QoDDNkx4fP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Bolian Li", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Ruqi Zhang", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" }, { "name": "Yi Ding", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Intern" } ], "n_formula": 11, "n_formula_1": 4, "n_ref_uni": 34, "n_ref": 127, "n_ref_all": 177, "n_fig": 12, "n_tab": 13, "L_tab": 3404, "n_element_tab": 422, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1030, "n_element_tab_1": 104, "formula_len_all": 918, "formula_len_all_1": 310, "len_all": 207547, "len_all_1": 61357, "len_abs": 1533, "len_title": 81, "len_sents": 61820, "len_sents_1": 28862, "n_sents": 458, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1206, "title": "ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time", "abs": "Vision Language Models (VLMs) have become essential backbones for multi-modal intelligence, yet significant safety challenges limit their real-world application. While textual inputs can often be effectively safeguarded, adversarial visual inputs can often easily bypass VLM defense mechanisms. Existing defense methods are either resource-intensive, requiring substantial data and compute, or fail to simultaneously ensure safety and usefulness in responses. To address these limitations, we propose a novel two-phase inference-time alignment framework, **E**valuating **T**hen **A**ligning (ETA): i) Evaluating input visual contents and output responses to establish a robust safety awareness in multimodal settings, and ii) Aligning unsafe behaviors at both shallow and deep levels by conditioning the VLMs' generative distribution with an interference prefix and performing sentence-level best-of-$N$ to search the most harmless and helpful generation paths. Extensive experiments show that ETA outperforms baseline methods in terms of harmlessness, helpfulness, and efficiency, reducing the unsafe rate by 87.5\\% in cross-modality attacks and achieving 96.6\\% win-ties in GPT-4 helpfulness evaluation.", "keywords": [ "VLMs", "safety alignment", "inference time" ], "rating_list": [ 6, 5, 5, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "Qny1ufReka", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Aditya Shrivastava", "gender": "unknown", "institution": "Capital One", "country": "US", "position": "Researcher" }, { "name": "Alfy Samuel", "gender": "Female", "institution": "CapitalOne", "country": "US", "position": "Researcher" }, { "name": "Anoop Kumar", "gender": "Male", "institution": "Amazon", "country": "", "position": "Computer Scientist" }, { "name": "Ashwinee Panda", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Postdoc" }, { "name": "Chenyang Zhu", "gender": "unknown", "institution": "Capital One", "country": "US", "position": "Researcher" }, { "name": "Daben Liu", "gender": "Male", "institution": "CapitalOne", "country": "US", "position": "Principal Researcher" }, { "name": "Micah Goldblum", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" }, { "name": "Neel Jain", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Tom Goldstein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 42, "n_ref": 65, "n_ref_all": 92, "n_fig": 9, "n_tab": 6, "L_tab": 841, "n_element_tab": 56, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2524, "n_element_tab_1": 129, "formula_len_all": 230, "formula_len_all_1": 93, "len_all": 147560, "len_all_1": 61276, "len_abs": 1478, "len_title": 123, "len_sents": 45669, "len_sents_1": 29532, "n_sents": 363, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1485, "title": "Refusal Tokens: A Simple Way to Calibrate Refusals in Large Language Models", "abs": "A key component of building safe and reliable language models is enabling the models to appropriately refuse to follow certain instructions or answer certain questions.\nWe may want models to output refusal messages for various categories of user queries, for example, ill-posed questions, instructions for committing illegal acts, or queries which require information past the model's knowledge horizon.\nEngineering models that refuse to answer such questions is complicated by the fact that an individual may want their model to exhibit varying levels of sensitivity for refusing queries of various categories,\nand different users may want different refusal rates. \nThe current default approach involves training multiple models with varying proportions of refusal messages from each category to achieve the desired refusal rates, which is computationally expensive and may require training a new model to accommodate each user's desired preference over refusal rates.\nTo address these challenges, we propose refusal tokens, one such token for each refusal category or a single refusal token, which are prepended to the model's responses during training. \nWe then show how to increase or decrease the probability of generating the refusal token for each category during inference to steer the model's refusal behavior. Refusal tokens enable controlling a single model's refusal rates without the need of any further fine-tuning, but only by selectively intervening during generation.", "keywords": [ "refusal messages", "trustworthiness", "llms" ], "rating_list": [ 3, 6, 3, 3, 6 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 1, 3 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "QnkhVwSu7u", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Letian Chen", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Matthew C. Gombolay", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 41, "n_ref": 52, "n_ref_all": 68, "n_fig": 3, "n_tab": 3, "L_tab": 1231, "n_element_tab": 204, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1231, "n_element_tab_1": 204, "formula_len_all": 342, "formula_len_all_1": 342, "len_all": 114126, "len_all_1": 66364, "len_abs": 1309, "len_title": 109, "len_sents": 32879, "len_sents_1": 31161, "n_sents": 205, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 122, "L_abs": 1369, "title": "ELEMENTAL: Interactive Learning from Demonstrations and Vision-Language Models for Interpretable Reward Design in Robotics", "abs": "Reinforcement learning (RL) has demonstrated compelling performance in robotic tasks, but its success often hinges on the design of complex, ad hoc reward functions. Researchers have explored how Large Language Models (LLMs) could enable non-expert users to specify reward functions more easily. However, LLMs struggle to balance the importance of different features, generalize poorly to out-of-distribution robotic tasks, and cannot represent the problem properly with only text-based descriptions. To address these challenges, we propose ELEMENTAL (intEractive LEarning froM dEmoNstraTion And Language), a novel framework that combines natural language guidance with visual user demonstrations to align robot behavior with user intentions better. By incorporating visual inputs, ELEMENTAL overcomes the limitations of text-only task specifications, while leveraging inverse reinforcement learning (IRL) to balance feature weights and match the demonstrated behaviors optimally. ELEMENTAL also introduces an iterative feedback-loop through self-reflection to improve feature, reward, and policy learning. Further, ELEMENTAL reward functions are interpretable. Our experiment results demonstrate that ELEMENTAL outperforms prior work by 24.4\\% on task success, and achieves 41.3\\% better generalization in out-of-distribution tasks, highlighting its robustness in LfD.", "keywords": [ "Interactive Robot Learning", "Inverse Reinforcement Learning", "Feature Abstraction", "Vision-Language Models" ], "rating_list": [ 3, 8, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "QnjUf0VytI", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Kaipeng Zhang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Tianhua Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Wenqi Shao", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yuxuan Xie", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 13, "n_ref": 24, "n_ref_all": 40, "n_fig": 8, "n_tab": 2, "L_tab": 247, "n_element_tab": 29, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 97, "n_element_tab_1": 15, "formula_len_all": 697, "formula_len_all_1": 697, "len_all": 86719, "len_all_1": 58196, "len_abs": 1290, "len_title": 124, "len_sents": 29083, "len_sents_1": 27724, "n_sents": 212, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1298, "title": "TP-Eval: Tap Multimodal LLMs' Potential in Evaluation by Customizing Prompts", "abs": "Recently, multimodal large language models (MLLMs) have received much attention for their impressive capabilities. The evaluation of MLLMs is becoming critical to analyzing attributes of MLLMs and providing valuable insights. However, current benchmarks overlook the problem of prompt sensitivity - minor prompt variations may lead to significant performance fluctuations. Thus, inappropriate prompts may obscure the models' capabilities, underestimating the models' performance. Moreover, different models have different preferences for different prompts, and thus, using the same prompt for all models will cause evaluation bias. This paper analyzes this deficiency in existing benchmarks and further introduces a new evaluation framework named TP-Eval, which introduces a prompt customization method to reduce evaluation biases and tap models' potential. TP-Eval will rewrite the original prompts to different customized prompts for different models. In particular, we propose some well-designed modules for prompt customization tailored to the scenario of MLLM evaluation. Extensive experiments demonstrate the effectiveness of our approach to uncovering models' capabilities, and TP-Eval should benefit the community in developing more comprehensive and convincing MLLM evaluation benchmarks.", "keywords": [ "MLLM", "benchmark", "prompt customization", "potential" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 5, 4 ] }, { "paper_id": "QmmTEgIbzA", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Aranyak Acharyya", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Postdoc" }, { "name": "Brandon Duderstadt", "gender": "Male", "institution": "Nomic AI", "country": "US", "position": "CEO" }, { "name": "Carey Priebe", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "Full Professor" }, { "name": "Hayden Helm", "gender": "Male", "institution": "Nomic AI", "country": "", "position": "Researcher" }, { "name": "Youngser Park", "gender": "unknown", "institution": "Johns Hopkins University", "country": "", "position": "Researcher" } ], "n_formula": 22, "n_formula_1": 9, "n_ref_uni": 50, "n_ref": 65, "n_ref_all": 85, "n_fig": 5, "n_tab": 1, "L_tab": 62, "n_element_tab": 1, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2493, "formula_len_all_1": 703, "len_all": 141225, "len_all_1": 67589, "len_abs": 719, "len_title": 58, "len_sents": 39500, "len_sents_1": 32007, "n_sents": 337, "n_sents_1": 258, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 725, "title": "Embedding-based statistical inference on generative models", "abs": "Generative models are capable of producing human-expert level content across a variety of topics and domains. As the impact of generative models grows, it is necessary to develop statistical methods to understand the population of available models. These methods are particularly important in settings where the user may not have access to information related to a model's pre-training data, weights, or other relevant model-level covariates. In this paper we extend recent results on representations of black-box generative models to model-level statistical inference tasks. We demonstrate -- both theoretically and empirically -- that the use of these representations are effective for multiple model-level inference tasks.", "keywords": [ "model inference", "embeddings" ], "rating_list": [ 3, 5, 3, 3, 8 ], "soundness_list": [ 2, 2, 2, 1, 4 ], "presentation_list": [ 2, 2, 1, 1, 3 ], "contribution_list": [ 2, 3, 2, 2, 4 ], "confidence_list": [ 4, 4, 3, 2, 2 ] }, { "paper_id": "QmXfEmtBie", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Donglin Wang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Associate Professor" }, { "name": "Huaicheng Zhou", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Intern" }, { "name": "Zifeng Zhuang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 34, "n_ref": 63, "n_ref_all": 102, "n_fig": 25, "n_tab": 6, "L_tab": 2232, "n_element_tab": 210, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 924, "n_element_tab_1": 86, "formula_len_all": 262, "formula_len_all_1": 339, "len_all": 137873, "len_all_1": 55529, "len_abs": 1888, "len_title": 131, "len_sents": 47840, "len_sents_1": 26303, "n_sents": 338, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1897, "title": "Stay Hungry, Keep Learning: Sustainable Plasticity for Deep Reinforcement Learning", "abs": "The integration of Deep Neural Networks (DNNs) in Reinforcement Learning (RL) systems has led to remarkable progress in solving complex tasks but also introduced challenges like primacy bias and dead neurons. Primacy bias skews learning towards early experiences, while dead neurons diminish the network's capacity to acquire new knowledge. Traditional reset mechanisms aimed at addressing these issues often involve maintaining large replay buffers to train new networks or selectively resetting subsets of neurons. However, These approaches either incur substantial computational costs or fail to effectively reset the entire network, resulting in underutilization of network plasticity and reduced learning efficiency. In this work, we introduce the novel concept of neuron regeneration, which combines reset mechanisms with knowledge recovery techniques. We also propose a new framework called Sustainable Backup Propagation (SBP) that effectively maintains plasticity in neural networks through this neuron regeneration process. The SBP framework achieves whole network neuron regeneration through two key procedures: cycle reset and inner distillation. Cycle reset involves a scheduled renewal of neurons, while inner distillation functions as a knowledge recovery mechanism at the neuron level. To validate our framework, we integrate SBP with Proximal Policy Optimization (PPO) and propose a novel distillation function for inner distillation. This integration results in Plastic PPO (P3O), a new algorithm that enables efficient cyclic regeneration of all neurons in the actor network. This approach facilitates neuron regeneration while maintaining policy plasticity and sample efficiency. Extensive experiments demonstrate that, with proper neuron regeneration methods, the SBP framework can effectively maintain plasticity and improve sample efficiency in reinforcement learning tasks.", "keywords": [ "Deep reinforcement learning", "Plasticity" ], "rating_list": [ 5, 5, 8, 3 ], "soundness_list": [ 2, 2, 4, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 4, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "QmJoF47DIR", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Jichen Hu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Jingcheng Zhu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Lizhuang Ma", "gender": "Male", "institution": "Dept. of Computer Sci. & Eng., Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Ran Yi", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xuan Wang", "gender": "Male", "institution": "Ant Research", "country": "CN", "position": "Scientific Researcher" }, { "name": "Yanbo Fan", "gender": "Male", "institution": "Ant Research", "country": "CN", "position": "Associate Professor" }, { "name": "Yating Wang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 24, "n_ref": 41, "n_ref_all": 49, "n_fig": 3, "n_tab": 3, "L_tab": 310, "n_element_tab": 53, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 292, "n_element_tab_1": 48, "formula_len_all": 589, "formula_len_all_1": 589, "len_all": 101103, "len_all_1": 57127, "len_abs": 1266, "len_title": 153, "len_sents": 30885, "len_sents_1": 27829, "n_sents": 208, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 105, "L_abs": 1272, "title": "High-fidelity and realtime 3D Gaussian Head Avatars with Expressive and Compact blenshape representations", "abs": "Recent studies have combined 3D Gaussian and 3D Morphable Models (3DMM) to achieve real-time, high-quality rendering of controllable head avatars. Several techniques have attempted to express dynamic textures in facial animation when modeling 3D avatars. However, accurately capturing and displaying expressive appearance dynamics while maintaining temporal and spatial efficiency remains a technical challenge. To this end, we propose a novel method for 3D facial avatar modeling that utilizes an expressive and compact model representation, capturing dynamic facial information accurately while ensuring efficiency. We encode texture-related attributes of the 3D Gaussians in the tensorial feature representation. Specifically, we store color information of the neutral expression in static tri-planes; and represent dynamic texture details for different expressions using lightweight 1D feature lines, which are then decoded into opacity changes relative to the neutral face. Experiments show that this design introduces nonlinear expressiveness to the model, enhancing its performance, while the compact representation maintains real-time rendering capabilities and significantly reduces storage costs. This approach thus broadens the applicability to more scenarios.", "keywords": [ "head avatar", "gaussian splatting" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "QlvL6eEOC6", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Chenfan Sun", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Maxwell Horton", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Mohammad Rastegari", "gender": "Male", "institution": "Apple", "country": "", "position": "ML manager" }, { "name": "Moin Nabi", "gender": "Male", "institution": "Apple Inc.", "country": "DE", "position": "Principal Research Scientist" }, { "name": "Qingqing Cao", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Sachin Mehta", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Yanzi Jin", "gender": "Female", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 43, "n_ref": 163, "n_ref_all": 193, "n_fig": 12, "n_tab": 6, "L_tab": 2248, "n_element_tab": 468, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 615, "n_element_tab_1": 212, "formula_len_all": 285, "formula_len_all_1": 314, "len_all": 205516, "len_all_1": 70206, "len_abs": 1455, "len_title": 94, "len_sents": 36924, "len_sents_1": 31982, "n_sents": 327, "n_sents_1": 274, "venue": "ICLR2025", "decision": "withdraw", "L_title": 46, "L_abs": 1470, "title": "KV Prediction for Improved Time to First Token", "abs": "Inference with transformer-based language models begins with a prompt processing step. In this step, the model generates the first output token and stores the KV cache needed for future generation steps. This prompt processing step can be computationally expensive, taking 10s of seconds or more for billion-parameter models on edge devices when prompt lengths or batch sizes rise. This degrades user experience by introducing significant latency into the model's outputs. To reduce the time spent producing the first output (known as the ``time to first token'', or TTFT of a pretrained model, we introduce a novel method called KV Prediction. In our method, a small auxiliary model is used to process the prompt and produce an approximation of the KV cache used by a base model. This approximated KV cache is then used with the base model for autoregressive generation without the need to query the auxiliary model again. We demonstrate that our method produces a pareto-optimal efficiency-accuracy trade-off when compared to baselines. On TriviaQA, we demonstrate relative accuracy improvements in the range of 15%-50% across a range of TTFT FLOPs budgets. We also demonstrate accuracy improvements of up to 30% on HumanEval python code completion at fixed TTFT FLOPs budgets. Additionally, we benchmark models on an Apple M2 Pro CPU and demonstrate that our improvement in FLOPs translates to a TTFT speedup on hardware. We will release our code for reproducibility.", "keywords": [ "time to first token", "TTFT", "on-device", "llm", "inference" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 1 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "QliOktBcy3", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Jin Tan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jinjin Zhou", "gender": "Male", "institution": "Antgroup", "country": "", "position": "Researcher" }, { "name": "Junming Ma", "gender": "Male", "institution": "NVIDIA", "country": "CN", "position": "DL performance architect" }, { "name": "Meng Li", "gender": "Male", "institution": "Peking University", "country": "", "position": "Assistant Professor" }, { "name": "Runsheng Wang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Wenxuan Zeng", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Ye Dong", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 69, "n_ref_all": 92, "n_fig": 9, "n_tab": 6, "L_tab": 2788, "n_element_tab": 341, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 732, "n_element_tab_1": 57, "formula_len_all": 719, "formula_len_all_1": 719, "len_all": 154568, "len_all_1": 49517, "len_abs": 1877, "len_title": 123, "len_sents": 26298, "len_sents_1": 19540, "n_sents": 175, "n_sents_1": 124, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1439, "title": "MPCache: MPC-Friendly KV Cache Eviction for Efficient Private LLM Inference", "abs": "Private LLM inference based on multi-party computation (MPC) offers cryptographically-secure protection for both user prompt and proprietary model weights. However, it suffers from large latency overhead for long input sequences. While key-value (KV) cache eviction algorithms have been proposed to reduce the computation and memory cost for plaintext inference, they are not designed for MPC and may even introduce more overhead. In this paper, we propose an accurate and MPC-friendly KV cache eviction framework, dubbed MPCache. MPCache is built on the observation that historical tokens in a long sequence may have different effects on the downstream decoding. Hence, MPCache combines a look-once static eviction algorithm to discard unimportant tokens and a query-aware dynamic selection algorithm to further choose a small subset of tokens for attention computation. As existing dynamic selection algorithms incur too much latency, we propose a series of optimizations to drastically reduce the KV cache selection overhead, including MPC-friendly similarity approximation, hierarchical KV cache clustering, and layer-wise index sharing strategy. With extensive experiments, we demonstrate that MPCache consistently outperforms prior-art KV cache eviction baselines across different LLM generation tasks and achieves 1.8 \u223c 2.01\u00d7 and 3.39 \u223c 8.37\u00d7 decoding latency and communication reduction on different sequence lengths, respectively.", "keywords": [ "Large Language Model", "KV Cache Compression", "Private Inference", "Multi-party Computation" ], "rating_list": [ 3, 8, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 1, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "Ql7msQBqoF", "primary_area": "reinforcement learning", "n_author": 10, "author_info_list": [ { "name": "Arjun Radhakrishna", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Arun Shankar Iyer", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Principal Researcher" }, { "name": "Gustavo Soares", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Krishna Kariya", "gender": "Female", "institution": "Microsoft", "country": "IN", "position": "Intern" }, { "name": "Naman Gupta", "gender": "unknown", "institution": "Microsoft", "country": "IN", "position": "Research Fellow" }, { "name": "Priyanshu Gupta", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Shashank Kirtania", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Intern" }, { "name": "Sriram K. Rajamani", "gender": "Male", "institution": "Research, Microsoft", "country": "", "position": "Researcher" }, { "name": "Sumit Gulwani", "gender": "Male", "institution": "Research, Microsoft", "country": "", "position": "Researcher" }, { "name": "Suresh Parthasarathy Iyengar", "gender": "Male", "institution": "Research, Microsoft", "country": "", "position": "Principal Researcher" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 65, "n_ref_all": 75, "n_fig": 2, "n_tab": 4, "L_tab": 1443, "n_element_tab": 101, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1564, "n_element_tab_1": 96, "formula_len_all": 658, "formula_len_all_1": 444, "len_all": 138166, "len_all_1": 63208, "len_abs": 1360, "len_title": 130, "len_sents": 31682, "len_sents_1": 29093, "n_sents": 226, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 851, "title": "MAC-CAFE: Multi-actor, Centralized Critic Architecture for Feedback-driven Editing", "abs": "Large Language Models (LLMs) often generate incorrect or outdated information, especially in low-resource settings or when dealing with private data. To address this, Retrieval-Augmented Generation (RAG) uses external knowledge bases (KBs), but these can also suffer from inaccuracies. We introduce MAC-CAFE, a novel Multi-actor, Centralized Critic Architecture for Feedback-driven Editing approach that iteratively refines the KB based on expert feedback using a multi-actor, centralized critic reinforcement learning framework. Each document is assigned to an actor, modeled as a ReACT agent, which performs structured edits based on document-specific targeted instructions from a centralized critic. Experimental results show that MAC-CAFE significantly improves KB quality and RAG system performance, enhancing accuracy by up to 8% over baselines.", "keywords": [ "Retrieval-Augmented Generation", "Large Language Models", "Knowledge Base Editing", "Prompt Optimization" ], "rating_list": [ 1, 3, 3, 6 ], "soundness_list": [ 1, 2, 1, 3 ], "presentation_list": [ 1, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "QkDUdPRcma", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 9, "author_info_list": [ { "name": "Hyun Jae Jang", "gender": "Male", "institution": "Korea Institute of Science and Technology", "country": "KR", "position": "Researcher" }, { "name": "Jaewook Kim", "gender": "Male", "institution": "Korea Institute of Science and Technology", "country": "KR", "position": "Researcher" }, { "name": "Jeonghee Jo", "gender": "Female", "institution": "Korea Institute of Science and Technology", "country": "KR", "position": "Postdoc" }, { "name": "Jinkyu Kim", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Assistant Professor" }, { "name": "Jongkil Park", "gender": "Male", "institution": "Korea Institute of Science and Technology", "country": "KR", "position": "Principal Researcher" }, { "name": "Mingyeong Seo", "gender": "Female", "institution": "Korea Institute of Science and Technology", "country": "KR", "position": "Intern" }, { "name": "Seongsik Park", "gender": "Male", "institution": "Korea Institute of Science and Technology", "country": "KR", "position": "Researcher" }, { "name": "Suyoun Lee", "gender": "Male", "institution": "Korea Institute of Science and Technology", "country": "KR", "position": "Principal Researcher" }, { "name": "Yechan Kang", "gender": "Male", "institution": "Korea Institute of Science and Technology", "country": "KR", "position": "Intern" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 46, "n_ref": 87, "n_ref_all": 103, "n_fig": 14, "n_tab": 6, "L_tab": 4758, "n_element_tab": 631, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 2619, "n_element_tab_1": 233, "formula_len_all": 500, "formula_len_all_1": 500, "len_all": 134150, "len_all_1": 63686, "len_abs": 1485, "len_title": 130, "len_sents": 31608, "len_sents_1": 27950, "n_sents": 227, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1493, "title": "H-Direct: Homeostasis-aware Direct Spike Encoding for Deep Spiking Neural Networks", "abs": "Deep spiking neural networks (SNNs) have been expected to enable energy-efficient artificial intelligence as a next-generation artificial neural network. Recently, with the development of various algorithms, such as direct spike encoding, many applications have been successfully implemented in deep SNNs. Notably, most state-of-the-art deep SNNs have greatly improved their performance by adopting direct spike encoding, which expresses input information as discrete spikes, thereby exerting substantial influence. Despite the importance of the encoding, efficient encoding methods have not been studied. As the first attempt to our knowledge, we thoroughly analyzed the conventional direct encoding. Our analysis revealed that the existing encoding restricts the training performance and efficiency due to inappropriate encoding. To address this limitation by maintaining an appropriate encoding, we introduced a concept of homeostasis to the direct spike encoding. With this concept, we presented a homeostasis-aware direct spike encoding (H-Direct), which consists of dynamic feature encoding loss, adaptive threshold, and feature diversity loss. Our experimental results demonstrate that the proposed encoding achieves higher performance and efficiency compared to conventional direct encoding across several image classification datasets on various architectures. We have validated that brain-inspired algorithms have the potential to enhance the performance and efficiency of deep SNNs.", "keywords": [ "spiking neural networks; direct encoding; neuromorphic computing" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "QjrC77Nyu6", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Adrian Atienza", "gender": "Male", "institution": "Technical University of Denmark", "country": "DK", "position": "PhD student" }, { "name": "Jakob E. Bardram", "gender": "Male", "institution": "Technical University of Denmark", "country": "", "position": "Full Professor" }, { "name": "Sadasivan Puthusserypady", "gender": "Male", "institution": "Technical University of Denmark", "country": "", "position": "Associate Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 27, "n_ref": 49, "n_ref_all": 66, "n_fig": 10, "n_tab": 5, "L_tab": 1953, "n_element_tab": 125, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 1214, "n_element_tab_1": 64, "formula_len_all": 98, "formula_len_all_1": 98, "len_all": 103925, "len_all_1": 51372, "len_abs": 1077, "len_title": 132, "len_sents": 30961, "len_sents_1": 24777, "n_sents": 213, "n_sents_1": 162, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1084, "title": "CuPID: Leveraging Masked Single-Lead ECG Modelling for Enhancing the Representations", "abs": "Wearable sensing devices, such as electrocardiogram (ECG) heart-rate monitors, will play a crucial role in the future of digital health. This constant monitoring leads to massive unlabeled datasets, making the development of unsupervised learning frameworks essential to associate these single-lead ECG signals with their anticipated clinical outcomes. While Masked Data Modelling (MDM) methods have enjoyed wide use, the idiosyncrasies of single-lead ECG data make its direct application impractical. In this paper, we present Cueing the Predictor Increments the Detailing (CuPID), a novel Self-Supervised Learning (SSL) method that adapts MDM methods for use on single-lead ECG signal data. CuPID accomplishes this via cueing spectrogram-derived context to the predictors, thus incentivizing the encoder to produce more detailed representations. This leads the class token to accommodate fine-grained information. We demonstrate that CuPID outperforms state-of-the-art methods in a variety of downstream tasks and databases, increasing the accuracy for each task from 3.6 % to 9.7%.", "keywords": [ "Self-Supervised Learning", "Time Series", "ECG" ], "rating_list": [ 3, 5, 1, 1 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 2, 3, 1, 1 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 5, 3, 5, 5 ] }, { "paper_id": "Qja5s0K3VX", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Nan Jiang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Associate Professor" }, { "name": "Yuheng Zhang", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 93, "n_formula_1": 19, "n_ref_uni": 43, "n_ref": 105, "n_ref_all": 132, "n_fig": 0, "n_tab": 4, "L_tab": 869, "n_element_tab": 9, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 540, "n_element_tab_1": 12, "formula_len_all": 11770, "formula_len_all_1": 2244, "len_all": 167605, "len_all_1": 69648, "len_abs": 3668, "len_title": 138, "len_sents": 44470, "len_sents_1": 30647, "n_sents": 398, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1182, "title": "Statistical Tractability of Off-policy Evaluation of History-dependent Policies in POMDPs", "abs": "We investigate off-policy evaluation (OPE), a central and fundamental problem\nin reinforcement learning (RL), in the challenging setting of Partially Observable\nMarkov Decision Processes (POMDPs) with large observation spaces. Recent\nworks of Uehara et al. (2023a); Zhang & Jiang (2024) developed a model-free\nframework and identified important coverage assumptions (called belief and outcome coverage) that enable accurate OPE of memoryless policies with polynomial sample complexities, but handling more general target policies that depend on\nthe entire observable history remained an open problem. In this work, we prove\ninformation-theoretic hardness for model-free OPE of history-dependent policies in\nseveral settings, characterized by additional assumptions imposed on the behavior\npolicy (memoryless vs. history-dependent) and/or the state-revealing property of\nthe POMDP (single-step vs. multi-step revealing). We further show that some hardness can be circumvented by a natural model-based algorithm\u2014whose analysis has surprisingly eluded the literature despite the algorithm\u2019s simplicity\u2014demonstrating\nprovable separation between model-free and model-based OPE in POMDPs.", "keywords": [ "Partially Observable Markov Decision Process; Offline Policy Evaluation; Reinforcement Learning Theory" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 2, 4, 3, 2 ] }, { "paper_id": "QjTSaFXg25", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Huy Nguyen", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Minh Le", "gender": "Male", "institution": "VinAI Research", "country": "VN", "position": "Researcher" }, { "name": "Nhat Ho", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Quyen Tran", "gender": "Female", "institution": "VinAI Research", "country": "VN", "position": "Research Resident" }, { "name": "Trung Le", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" }, { "name": "ChauNguyen", "gender": "Female", "institution": "VinAI", "country": "VN", "position": "Researcher" } ], "n_formula": 200, "n_formula_1": 53, "n_ref_uni": 51, "n_ref": 91, "n_ref_all": 128, "n_fig": 3, "n_tab": 10, "L_tab": 3647, "n_element_tab": 423, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 735, "n_element_tab_1": 106, "formula_len_all": 28544, "formula_len_all_1": 5135, "len_all": 272840, "len_all_1": 71470, "len_abs": 1909, "len_title": 130, "len_sents": 66528, "len_sents_1": 28007, "n_sents": 652, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1574, "title": "Revisiting Prefix-tuning: Statistical Benefits of Reparameterization among Prompts", "abs": "Prompt-based techniques, such as prompt-tuning and prefix-tuning, have gained prominence for their efficiency in fine-tuning large pre-trained models. Despite their widespread adoption, the theoretical foundations of these methods remain limited. For instance, in prefix-tuning, we observe that a key factor in achieving performance parity with full fine-tuning lies in the reparameterization strategy. However, the theoretical principles underpinning the effectiveness of this approach have yet to be thoroughly examined. Our study demonstrates that reparameterization is not merely an engineering trick but is grounded in deep theoretical foundations. Specifically, we show that the reparameterization strategy implicitly encodes a shared structure between prefix key and value vectors. Building on recent insights into the connection between prefix-tuning and mixture of experts models, we further illustrate that this shared structure significantly improves sample efficiency in parameter estimation compared to non-shared alternatives. The effectiveness of prefix-tuning across diverse tasks is empirically confirmed to be enhanced by the shared structure, through extensive experiments in both visual and language domains. Additionally, we uncover similar structural benefits in prompt-tuning, offering new perspectives on its success. Our findings provide theoretical and empirical contributions, advancing the understanding of prompt-based methods and their underlying mechanisms. Our code is publicly available at https://github.com/Minhchuyentoancbn/ReparamPrefix.", "keywords": [ "prefix-tuning", "mixture of experts", "theory", "prompt", "reparameterization" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "QjSOgxJ0hp", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 1, "author_info_list": [ { "name": "Tal Wagner", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 58, "n_formula_1": 11, "n_ref_uni": 41, "n_ref": 118, "n_ref_all": 144, "n_fig": 5, "n_tab": 9, "L_tab": 16923, "n_element_tab": 1162, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 515, "n_element_tab_1": 6, "formula_len_all": 6073, "formula_len_all_1": 1029, "len_all": 219341, "len_all_1": 68108, "len_abs": 1135, "len_title": 110, "len_sents": 55223, "len_sents_1": 31305, "n_sents": 456, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 84, "L_abs": 1145, "title": "Learning from End User Data with Shuffled Differential Privacy over Kernel Densities", "abs": "We study a setting of collecting and learning from private data distributed across end users.\nIn the shuffled model of differential privacy, the end users partially protect their data locally before sharing it, and their data is also anonymized during its collection to enhance privacy. \nThis model has recently become a prominent alternative to central DP, which requires full trust in a central data curator, and local DP, where fully local data protection takes a steep toll on downstream accuracy. \n\nOur main technical result is a shuffled DP protocol for privately estimating the kernel density function of a distributed dataset, with accuracy essentially matching central DP. \nWe use it to privately learn a classifier from the end user data, by learning a private density function per class. \nMoreover, we show that the density function itself can recover the semantic content of its class, despite having been learned in the absence of any unprotected data. \nOur experiments show the favorable downstream performance of our approach, and highlight key downstream considerations and trade-offs in a practical ML deployment of shuffled DP.", "keywords": [ "differential privacy", "shuffled differential privacy", "kernel density estimation", "kde" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "QjO0fUlVYK", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Alexander Rubinstein", "gender": "unknown", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Ankit Sonthalia", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Seong Joon Oh", "gender": "Male", "institution": "Parameter Lab", "country": "DE", "position": "Principal Researcher" }, { "name": "Ehsan M Abbasnejad", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 31, "n_formula_1": 14, "n_ref_uni": 38, "n_ref": 103, "n_ref_all": 157, "n_fig": 10, "n_tab": 4, "L_tab": 1397, "n_element_tab": 144, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 586, "n_element_tab_1": 75, "formula_len_all": 3207, "formula_len_all_1": 1015, "len_all": 199751, "len_all_1": 72817, "len_abs": 1116, "len_title": 100, "len_sents": 68542, "len_sents_1": 32482, "n_sents": 628, "n_sents_1": 282, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1051, "title": "Do Deep Neural Network Solutions Form a Star Domain?", "abs": "It has recently been conjectured that neural network solution sets reachable via stochastic gradient descent (SGD) are convex, considering permutation invariances. This means that a linear path can connect two independent solutions with low loss, given the weights of one of the models are appropriately permuted. However, current methods to test this theory often require very wide networks to succeed. In this work, we conjecture that more generally, the SGD solution set is a star domain that contains a star model that is linearly connected to all the other solutions via paths with low loss values, modulo permutations. We propose the Starlight algorithm that finds a star model of a given learning task. We validate our claim by showing that this star model is linearly connected with other independently found solutions. As an additional benefit of our study, we demonstrate better uncertainty estimates on Bayesian Model Averaging over the obtained star domain. Further, we demonstrate star models as potential substitutes for model ensembles.", "keywords": [ "mode connectivity", "loss landscapes", "neural networks", "parameter space", "star domain" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "QjNHmfA3IB", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Guangliang Cheng", "gender": "Male", "institution": "University of Liverpool", "country": "GB", "position": "Associate Professor" }, { "name": "Qi Zhao", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Shuchang Lyu", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Postdoc" }, { "name": "Wenquan Feng", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Xiaowei Huang", "gender": "Male", "institution": "University of Liverpool", "country": "GB", "position": "Full Professor" }, { "name": "Zheng Zhou", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 34, "n_ref": 139, "n_ref_all": 147, "n_fig": 4, "n_tab": 4, "L_tab": 1316, "n_element_tab": 138, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 506, "formula_len_all_1": 507, "len_all": 132773, "len_all_1": 50138, "len_abs": 1391, "len_title": 120, "len_sents": 40286, "len_sents_1": 24811, "n_sents": 268, "n_sents_1": 158, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1396, "title": "BEARD: Benchmarking the Adversarial Robustness for Dataset Distillation", "abs": "Dataset Distillation (DD) is an emerging technique that compresses large-scale datasets into significantly smaller synthesized datasets while preserving high test performance and enabling the efficient training of large models. However, current research primarily focuses on enhancing evaluation accuracy under limited compression ratios, often overlooking critical security concerns such as adversarial robustness. A key challenge in evaluating this robustness lies in the complex interactions between distillation methods, model architectures, and adversarial attack strategies, which complicate standardized assessments. To address this, we introduce BEARD, an open and unified benchmark designed to systematically assess the adversarial robustness of DD methods, including DM, IDM, and BACON. BEARD encompasses a variety of adversarial attacks (e.g., FGSM, PGD, C&W) on distilled datasets like CIFAR-10/100 and TinyImageNet. Utilizing an adversarial game framework, it introduces three key metrics: Robustness Ratio (RR), Attack Efficiency Ratio (AE), and Comprehensive Robustness-Efficiency Index (CREI). Our analysis includes unified benchmarks, various Images Per Class (IPC) settings, and the effects of adversarial training. Results are available on the BEARD Leaderboard, along with a library providing model and dataset pools to support reproducible research. Access the code at BEARD.", "keywords": [ "Dataset distillation", "benchmark", "adversarial robustness", "dataset condensation", "adversarial game framework" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "Qj1KwBZaEI", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Alex Rodriguez", "gender": "unknown", "institution": "University of Trieste", "country": "", "position": "Assistant Professor" }, { "name": "Fabio Anselmi", "gender": "Male", "institution": "University of Trieste", "country": "", "position": "Assistant Professor" }, { "name": "Lorenzo Basile", "gender": "unknown", "institution": "University of Trieste", "country": "IT", "position": "PhD student" }, { "name": "Luca Bortolussi", "gender": "Male", "institution": "University of Trieste", "country": "IT", "position": "Full Professor" }, { "name": "Santiago Acevedo", "gender": "Male", "institution": "International Higher School for Advanced Studies Trieste", "country": "IT", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 3, "n_ref_uni": 43, "n_ref": 66, "n_ref_all": 93, "n_fig": 12, "n_tab": 7, "L_tab": 1535, "n_element_tab": 234, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 600, "formula_len_all_1": 145, "len_all": 144403, "len_all_1": 56601, "len_abs": 1066, "len_title": 143, "len_sents": 45690, "len_sents_1": 30072, "n_sents": 260, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1070, "title": "Intrinsic Dimension Correlation: uncovering nonlinear connections in multimodal representations", "abs": "To gain insight into the mechanisms behind machine learning methods, it is crucial to establish connections among the features describing data points. However, these correlations often exhibit a high-dimensional and strongly nonlinear nature, which makes them challenging to detect using standard methods. This paper exploits the entanglement between intrinsic dimensionality and correlation to propose a metric that quantifies the (potentially nonlinear) correlation between high-dimensional manifolds. We first validate our method on synthetic data in controlled environments, showcasing its advantages and drawbacks compared to existing techniques. Subsequently, we extend our analysis to large-scale applications in neural network representations. Specifically, we focus on latent representations of multimodal data, uncovering clear correlations between paired visual and textual embeddings, whereas existing methods struggle significantly in detecting similarity. Our results indicate the presence of highly nonlinear correlation patterns between latent manifolds.", "keywords": [ "intrinsic dimension", "nonlinear correlation", "multimodal representations", "representation similarity" ], "rating_list": [ 8, 8, 6, 5, 8 ], "soundness_list": [ 4, 3, 2, 3, 3 ], "presentation_list": [ 4, 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 5, 3, 4, 4, 4 ] }, { "paper_id": "QiyQJqpcYe", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Belen Alastruey", "gender": "Female", "institution": "Facebook", "country": "", "position": "PhD student" }, { "name": "Christophe Ropers", "gender": "unknown", "institution": "Meta", "country": "", "position": "Linguist" }, { "name": "Eduardo S\u00e1nchez", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Marta R. Costa-juss\u00e0", "gender": "Female", "institution": "Meta", "country": "", "position": "Research Scientist" }, { "name": "Mikel Artetxe", "gender": "Male", "institution": "Reka AI", "country": "", "position": "Researcher" }, { "name": "Pontus Stenetorp", "gender": "Not Specified", "institution": "University College London", "country": "GB", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 21, "n_ref": 28, "n_ref_all": 49, "n_fig": 6, "n_tab": 8, "L_tab": 4107, "n_element_tab": 586, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2411, "n_element_tab_1": 274, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 107966, "len_all_1": 46939, "len_abs": 736, "len_title": 113, "len_sents": 20230, "len_sents_1": 17950, "n_sents": 139, "n_sents_1": 125, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 737, "title": "Linguini: A benchmark for language-agnostic linguistic reasoning", "abs": "We propose a new benchmark to measure a language model\u2019s linguistic reasoning skills without relying on pre-existing language-specific knowledge. The test covers 894 questions grouped in 160 problems across 75 (mostly) extremely low-resource languages, extracted from the International Linguistic Olympiad corpus. To attain high accuracy on this benchmark, models don\u2019t need previous knowledge of the tested language, since all the information required to solve the linguistic puzzle is provided within the context. We find that, while all analyzed models rank below 25% accuracy, there is a significant gap between open and closed models, with the best-performing proprietary model at 24.05% and the best-performing open model at 8.84%.", "keywords": [ "linguistic-reasoning" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 2, 4, 2, 4 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "QipLSeLQRS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Haimin Hu", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Jaime Fern\u00e1ndez Fisac", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Kaiqu Liang", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Ryan Liu", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Thomas L. Griffiths", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "Professor" } ], "n_formula": 15, "n_formula_1": 9, "n_ref_uni": 53, "n_ref": 94, "n_ref_all": 116, "n_fig": 14, "n_tab": 7, "L_tab": 1497, "n_element_tab": 101, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1255, "n_element_tab_1": 34, "formula_len_all": 1198, "formula_len_all_1": 709, "len_all": 199670, "len_all_1": 72935, "len_abs": 1648, "len_title": 111, "len_sents": 62660, "len_sents_1": 34877, "n_sents": 457, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1661, "title": "RLHS: Mitigating Misalignment in RLHF with Hindsight Simulation", "abs": "Generative AI systems like foundation models (FMs) must align well with human values to ensure their behavior is helpful and trustworthy. While Reinforcement Learning from Human Feedback (RLHF) has shown promise for optimizing model performance using human judgments, existing RLHF pipelines predominantly rely on *immediate* feedback, which can fail to reflect the true downstream impact of an interaction on users' utility. We demonstrate that this shortsighted feedback can, by itself, result in misaligned behaviors like sycophancy and deception, and we propose to alleviate this by refocusing RLHF on *downstream consequences*. Our theoretical analysis reveals that the hindsight gained by simply delaying human feedback mitigates misalignment and improves expected human utility. To leverage this insight in a practical alignment algorithm, we introduce Reinforcement Learning from Hindsight Simulation (RLHS), which first simulates plausible consequences and then elicits feedback to assess what behaviors were genuinely beneficial in hindsight. We apply RLHS to two widely-employed online and offline preference optimization methods---Proximal Policy Optimization (PPO) and Direct Preference Optimization (DPO)---and show empirically that misalignment is significantly reduced with both methods. Through an online human user study, we show that RLHS consistently outperforms RLHF in helping users achieve their goals and earns higher satisfaction ratings, despite being trained solely with simulated hindsight feedback. These results underscore the importance of focusing on long-term consequences, even simulated ones, to mitigate misalignment in RLHF.", "keywords": [ "Reinforcement Learning from Human Feedback (RLHF)", "Human-AI Alignment", "Large Language Models", "AI Safety", "Partial Observability" ], "rating_list": [ 5, 8, 5, 5 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 2, 3, 2, 4 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "QinkNNKZ3b", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Hilde Kuehne", "gender": "Female", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "Full Professor" }, { "name": "Juergen Gall", "gender": "unknown", "institution": "University of Bonn", "country": "DE", "position": "Professor" }, { "name": "Lokesh Veeramacheneni", "gender": "Male", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "PhD student" }, { "name": "Moritz Wolter", "gender": "Male", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "Postdoc" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 50, "n_ref": 104, "n_ref_all": 133, "n_fig": 9, "n_tab": 9, "L_tab": 3759, "n_element_tab": 402, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2249, "n_element_tab_1": 317, "formula_len_all": 916, "formula_len_all_1": 449, "len_all": 157087, "len_all_1": 56103, "len_abs": 1185, "len_title": 119, "len_sents": 36622, "len_sents_1": 22524, "n_sents": 372, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1215, "title": "Fr\u00e9chet Wavelet Distance: A Domain-Agnostic Metric for Image Generation", "abs": "Modern metrics for generative learning like Fr\u00e9chet Inception Distance (FID ) and DINOv2-Fr\u00e9chet Distance (FD-DINOv2 ) demonstrate impressive performance. However, they suffer from various shortcomings, like a bias towards specific generators and datasets. To address this problem, we propose the Fr\u00e9chet Wavelet Distance (FWD ) as a domain-agnostic metric based on the Wavelet Packet Transform ($\\mathcal{W}_p$). FWD provides a sight across a broad spectrum of frequencies in images with a high resolution, preserving both spatial and textural aspects. Specifically, we use $\\mathcal{W}_p$ to project generated and real images to the packet coefficient space. We then compute the Fr\u00e9chet distance with the resultant coefficients to evaluate the quality of a generator. This metric is general-purpose and dataset-domain agnostic, as it does not rely on any pre-trained network while being more interpretable due to its ability to compute Fr\u00e9chet distance per packet, enhancing transparency. We conclude with an extensive evaluation of a wide variety of generators across various datasets that the proposed FWD can generalize and improve robustness to domain shifts and various corruptions compared to other metrics.", "keywords": [ "Frechet Distance", "Wavelet Packet Transform", "Frechet Inception Distance", "Diffusion", "GAN", "ImageNet", "Frechet Inception Distance", "FD-DINOv2" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "QibJggOAnB", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Sangchul Park", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Assistant Professor" }, { "name": "Yongdai Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Kun woong Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" } ], "n_formula": 86, "n_formula_1": 35, "n_ref_uni": 54, "n_ref": 199, "n_ref_all": 253, "n_fig": 8, "n_tab": 11, "L_tab": 2288, "n_element_tab": 221, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 299, "n_element_tab_1": 21, "formula_len_all": 11144, "formula_len_all_1": 2885, "len_all": 238218, "len_all_1": 71852, "len_abs": 1554, "len_title": 77, "len_sents": 71255, "len_sents_1": 30956, "n_sents": 616, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 29, "L_abs": 1561, "title": "Fair Clustering via Alignment", "abs": "Algorithmic fairness in clustering aims to balance the proportions of instances assigned to each cluster with respect to a given sensitive attribute.\nRecently, numerous algorithms have been developed for Fair Clustering (FC), most of which optimize a clustering objective under specifically designed fairness constraints.\nHowever, the inherent complexity or approximation of constrained optimization problems makes it challenging to achieve the optimal trade-off between fairness level and clustering utility in practice.\nFor example, the obtained clustering utility by an existing FC algorithm might be suboptimal, or achieving a certain fairness level could be numerically unstable.\nTo resolve these limitations, we propose a new FC algorithm based on a novel decomposition of the fair $K$-means clustering objective function.\nThe proposed algorithm, called Fair Clustering via Alignment (FCA), operates by (i) finding a joint probability distribution to align the data from different protected groups, and (ii) optimizing cluster centers in the aligned space.\nA key advantage of FCA is that it guarantees (local) optimal clustering utility for any given fairness level while avoiding the need to solve complex constrained optimization problems, thereby obtaining (local) optimal fair clustering in practice.\nExperiments show that FCA offers several empirical benefits over existing methods such as (i) attaining the optimal trade-off between fairness level and clustering utility, and (ii) achieving near-perfect fairness level without numerical instability.", "keywords": [ "Clustering", "Fairness", "Trustworthy AI" ], "rating_list": [ 5, 5, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 3, 4, 4 ] }, { "paper_id": "QiUitwJDKI", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Hendrik Baier", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "Levi Lelis", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Assistant Professor" }, { "name": "Quazi Asif Sadmine", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "MS student" }, { "name": "Rubens O. Moraes", "gender": "Male", "institution": "Universidade Federal de Vi\u00e7osa", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 50, "n_ref": 124, "n_ref_all": 157, "n_fig": 13, "n_tab": 2, "L_tab": 764, "n_element_tab": 45, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 334, "n_element_tab_1": 32, "formula_len_all": 1592, "formula_len_all_1": 72, "len_all": 225839, "len_all_1": 69875, "len_abs": 1405, "len_title": 113, "len_sents": 78003, "len_sents_1": 35378, "n_sents": 627, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1419, "title": "InnateCoder: Learning Programmatic Options with Foundation Models", "abs": "Outside of transfer learning settings, reinforcement learning agents start their learning process from a clean slate. As a result, such agents have to go through a slow process to learn even the most obvious skills required to solve a problem. In this paper, we present InnateCoder, a system that leverages human knowledge encoded in foundation models to provide programmatic policies that encode \"innate skills\" in the form of temporally extended actions, or options. In contrast to existing approaches to learning options, InnateCoder learns them from the general human knowledge encoded in foundation models in a zero-shot setting, and not from the knowledge the agent gains by interacting with the environment. Then, InnateCoder searches for a programmatic policy by combining the programs encoding these options into a larger and more complex program. We hypothesized that InnateCoder's scheme of learning and using options could improve the sampling efficiency of current methods for synthesizing programmatic policies. We evaluated our hypothesis in MicroRTS and Karel the Robot, two challenging domains. Empirical results support our hypothesis, since they show that InnateCoder is more sample efficient than versions of the system that do not use options or learn the options from experience. The policies InnateCoder learns are competitive and often outperform current state-of-the-art agents in both domains.", "keywords": [ "programmatic policies", "reinforcement learning", "options" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "QhxjQOMdDF", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Meisam Razaviyayn", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Associate Professor" }, { "name": "Peilin Zhong", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Vahab Mirrokni", "gender": "Male", "institution": "Google Research", "country": "", "position": "Distinguished Scientist" }, { "name": "Xinwei Zhang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Postdoc" }, { "name": "Yuan Deng", "gender": "Male", "institution": "Google Research", "country": "", "position": "Research Scientist" }, { "name": "Zeman Li", "gender": "Male", "institution": "Google", "country": "US", "position": "Intern" } ], "n_formula": 69, "n_formula_1": 20, "n_ref_uni": 66, "n_ref": 147, "n_ref_all": 201, "n_fig": 10, "n_tab": 19, "L_tab": 7154, "n_element_tab": 1256, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9668, "formula_len_all_1": 1797, "len_all": 289775, "len_all_1": 68063, "len_abs": 6673, "len_title": 171, "len_sents": 71275, "len_sents_1": 27903, "n_sents": 576, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 123, "L_abs": 2178, "title": "Addax: Utilizing Zeroth-Order Gradients to Improve Memory Efficiency and Performance of SGD for Fine-Tuning Language Models", "abs": "Fine-tuning language models (LMs) with the standard Adam optimizer often demands excessive memory, limiting accessibility. The ``in-place'' version of Stochastic Gradient Descent (IP-SGD) and Memory-Efficient Zeroth-order Optimizer (MeZO) have been proposed as solutions to improve memory efficiency. However, IP-SGD still requires a decent amount of memory, and MeZO suffers from slow convergence and degraded final performance due to its zeroth-order nature. This paper introduces Addax, a novel method that improves both memory efficiency and algorithm performance of IP-SGD by integrating it with MeZO. Specifically, Addax computes the zeroth-order or first-order gradient of the data points in the minibatch based on their memory consumption and combines zeroth- and first-order gradient estimates to obtain the updated direction in each step.\nBy computing the zeroth-order order gradient of data points that require more memory and the first-order gradient of the ones that require less memory, Addax overcomes the slow convergence of MeZO and excessive memory requirement of IP-SGD. Additionally, the zeroth-order gradient acts as a regularizer for the first-order gradient, further enhancing the model's final performance.\nTheoretically, we establish the convergence of Addax under mild assumptions, demonstrating faster convergence and less restrictive hyper-parameter choices than MeZO. Our extensive experiments with diverse LMs and tasks show that Addax consistently outperforms MeZO in terms of accuracy and convergence speed, while having a comparable memory footprint. \nIn particular, our experiments using one A100 GPU on OPT-13B model reveal that, on average, Addax outperforms MeZO in terms of accuracy/F1 score by 14%, and runs $15\\times$ faster, while having a comparable memory footprint to MeZO. In our experiments on the larger OPT-30B model, on average, Addax outperforms MeZO in terms of accuracy/F1 score by >16% and runs $30\\times$ faster on a single H100 GPU. Moreover, Addax surpasses the performance of standard fine-tuning approaches, such as IP-SGD and Adam, in most tasks in terms of Accuracy/F1 score with significantly less memory requirement.", "keywords": [ "Large Language Models", "memory efficient Fine-tuning", "Zeroth order optimization" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "QhsbF2RZeu", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Guijin Son", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Undergrad student" }, { "name": "Hoyoung Lee", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Seunghyeok Hong", "gender": "Male", "institution": "the University of Suwon", "country": "KR", "position": "Assistant Professor" }, { "name": "Yewon Kim", "gender": "Female", "institution": "Kangwon National University", "country": "", "position": "Undergrad student" }, { "name": "Ko Hyun Woo", "gender": "Male", "institution": "haerae.com", "country": "KR", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 34, "n_ref": 66, "n_ref_all": 93, "n_fig": 10, "n_tab": 9, "L_tab": 3287, "n_element_tab": 264, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 2244, "n_element_tab_1": 265, "formula_len_all": 334, "formula_len_all_1": 19, "len_all": 166397, "len_all_1": 72603, "len_abs": 3328, "len_title": 106, "len_sents": 42556, "len_sents_1": 31254, "n_sents": 343, "n_sents_1": 244, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1312, "title": "LLM-as-a-Judge & Reward Model: What They Can and Cannot Do", "abs": "LLM-as-a-Judge and reward models are widely used alternatives of multiple-choice questions or human annotators for large language model (LLM) evaluation. Their efficacy shines in evaluating long-form responses, serving a critical role as evaluators of leaderboards and as proxies to align LLMs via reinforcement learning. However, despite their popularity, their effectiveness in diverse contexts, such as non-English prompts, factual verification, or challenging questions, remains unexplored. In this paper, we conduct a comprehensive analysis of automated evaluators, reporting several key findings on their behavior. First, we discover that English evaluation capabilities significantly influence language-specific evaluation capabilities, often more than the language proficiency itself, enabling evaluators trained in English to easily transfer their skills to other languages. Second, we identify critical shortcomings, where LLMs fail to detect and penalize errors, such as factual inaccuracies, cultural misrepresentations, and the presence of unwanted language. Finally, we find that state-of-the-art evaluators struggle with challenging prompts, in either English or Korean, underscoring their limitations in assessing or generating complex reasoning questions. We release the dataset and codes used.", "keywords": [ "reward modeling", "meta-evaluation", "llm-as-a-judge" ], "rating_list": [ 5, 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 2, 3, 3, 4, 4 ] }, { "paper_id": "QhnPrsZ38V", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Canlin Zhang", "gender": "Male", "institution": "Sorenson Communications", "country": "US", "position": "Researcher" }, { "name": "Xiuwen Liu", "gender": "Male", "institution": "Florida State University", "country": "US", "position": "full Professor" } ], "n_formula": 95, "n_formula_1": 37, "n_ref_uni": 63, "n_ref": 74, "n_ref_all": 112, "n_fig": 10, "n_tab": 3, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 11219, "formula_len_all_1": 3636, "len_all": 292404, "len_all_1": 70458, "len_abs": 5917, "len_title": 116, "len_sents": 100000, "len_sents_1": 30427, "n_sents": 773, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1307, "title": "Explainable self-supervised learning by spiking functions: a theory", "abs": "Deep neural networks trained in an end-to-end manner have been proven to be efficient in a wide range of machine learning tasks. However, there is one drawback of end-to-end learning: The learned features and information are implicitly represented in neural network parameters, which are not explainable: The learned features cannot be used as explicit regularities to explain the data probability distribution. To resolve this issue, we propose in this paper a new machine learning theory, which describes in mathematics what are 'non-randomness' and 'regularities' in a data probability distribution. Our theory applies a spiking function to distinguish data samples from random noises. In this process, 'non-randomness', or a large amount of information, is encoded by the spiking function into regularities, a small amount of information. Then, our theory describes the application of multiple spiking functions to the same data distribution. In this process, we claim that the 'best' regularities, or the optimal spiking functions, are those who can capture the largest amount of information from the data distribution, and then encode the captured information into the smallest amount of information. By optimizing the spiking functions, one can achieve an explainable self-supervised learning system.", "keywords": [ "Self-supervised learning", "spiking neural networks", "information theory", "XAI" ], "rating_list": [ 5, 5, 1, 5, 3 ], "soundness_list": [ 2, 3, 1, 3, 2 ], "presentation_list": [ 4, 2, 3, 3, 2 ], "contribution_list": [ 4, 2, 1, 2, 2 ], "confidence_list": [ 3, 3, 5, 3, 3 ] }, { "paper_id": "QhjosARfay", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Chao Qu", "gender": "Male", "institution": "Inftech", "country": "CN", "position": "Researcher" }, { "name": "Fangzhen Lin", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Fengming ZHU", "gender": "unknown", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Haozhe Wang", "gender": "Male", "institution": "INF", "country": "CN", "position": "Researcher" }, { "name": "Long Li", "gender": "Male", "institution": "East China Normal University", "country": "", "position": "MS student" }, { "name": "Wei Chu", "gender": "Male", "institution": "Inf Tech", "country": "SG", "position": "Researcher" }, { "name": "Weidi Xu", "gender": "Male", "institution": "Infly Technology", "country": "CN", "position": "Researcher" }, { "name": "Yi Xin", "gender": "Male", "institution": "Nanjing university", "country": "CN", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 9, "n_ref_uni": 27, "n_ref": 74, "n_ref_all": 93, "n_fig": 6, "n_tab": 3, "L_tab": 1697, "n_element_tab": 168, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 692, "n_element_tab_1": 105, "formula_len_all": 1428, "formula_len_all_1": 687, "len_all": 147226, "len_all_1": 69998, "len_abs": 1447, "len_title": 113, "len_sents": 48931, "len_sents_1": 30516, "n_sents": 449, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1429, "title": "AutoCode4Math: Learning Autonomous Code Integration for Math LLMs", "abs": "Recent research on tool integration for math Large Language Models (LLMs) aims to combine complementary strengths of chain-of-thought (CoT) reasoning and code execution. However, we discover a critical limitation: current tool-integrated math LLMs rely on externally dictated instructions to decide whether to use CoT or code, lacking the autonomy to choose the most appropriate method independently. This prompts us to study \\emph{Autonomous Code integration } (AutoCode) for math LLMs, which enables models to \\emph{independently} develop their own methodology-selection strategy in the absence of reliable supervision. To address this challenge, we propose an innovative Expectation-Maximization (EM) formulation that refines the model's decision-making through the exploration of its capabilities. This framework alternates between (a) computing a reference strategy that improves the model's belief over its capabilities through self-exploration, and (b) updating the model based on the refined belief. We further enhance this framework with an efficient implementation, incorporating a novel data synthesis strategy and off-policy reinforcement learning. Extensive experiments demonstrate that our approach, using only a public query set, significantly boosts the performance of existing math LLMs, raising accuracy by nearly 20% to 65.28% on the challenging MATH benchmark, while reducing code executions by up to 65% . ", "keywords": [ "Large Language Models" ], "rating_list": [ 5, 5, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "QhhShUQIpJ", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Chuan Liu", "gender": "Female", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Chunshu Wu", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Postdoc" }, { "name": "Pouya Haghi", "gender": "Male", "institution": "University of Rochester", "country": "", "position": "Postdoc" }, { "name": "Ruibing Song", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Tong Geng", "gender": "Male", "institution": "University of Rochester", "country": "", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 32, "n_ref": 57, "n_ref_all": 77, "n_fig": 7, "n_tab": 5, "L_tab": 1853, "n_element_tab": 342, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 838, "n_element_tab_1": 146, "formula_len_all": 933, "formula_len_all_1": 870, "len_all": 127003, "len_all_1": 65249, "len_abs": 1395, "len_title": 135, "len_sents": 37536, "len_sents_1": 29102, "n_sents": 270, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1419, "title": "InstaTrain: Adaptive Training via Ultra-Fast Natural Annealing within Dynamical Systems", "abs": "Time-series modeling is broadly adopted to capture underlying patterns and trends present in historical data, allowing for prediction of future values. However, one crucial aspect in such modeling is often overlooked: in highly dynamic environments, data distributions can shift drastically within a second or less. Under this circumstance, traditional predictive models, even online learning methods struggle to adapt to the ultra-fast and complex distribution shift present in highly dynamic scenarios. To address this, we propose InstaTrain, a novel learning paradigm that enables frequent model updates with microsecond-level intervals for real-world prediction tasks, allowing it to keep pace with rapidly evolving data distributions. In this work, (1) We transform the slow and expensive model training process into an ultra-fast natural annealing process that can be carried out on a dynamical system. (2) Leveraging a recently proposed electronic dynamical system, we augment the system with a parameter update module, extending its capabilities to encompass both rapid training and inference. Experimental results across highly dynamic datasets demonstrate that our method delivers on average, a significant $\\sim$4,000$\\times$ training speedup, $\\sim10^5\\times$ reduction in training energy costs, and a remarkable lower test MAE over SOTA methods running on GPUs without / with the online learning mechanism.", "keywords": [ "Time-Series Prediction", "Online Learning", "Nature-Powered Computing" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "QgA0auXUU0", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Artur Andrzejak", "gender": "Male", "institution": "Heidelberg University", "country": "DE", "position": "Full Professor" }, { "name": "Min Xue", "gender": "Female", "institution": "Ruprecht-Karls-Universit\u00e4t Heidelberg", "country": "DE", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 42, "n_ref": 57, "n_ref_all": 91, "n_fig": 5, "n_tab": 8, "L_tab": 2598, "n_element_tab": 337, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1698, "n_element_tab_1": 58, "formula_len_all": 329, "formula_len_all_1": 106, "len_all": 179600, "len_all_1": 61918, "len_abs": 1742, "len_title": 136, "len_sents": 49271, "len_sents_1": 28601, "n_sents": 347, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1480, "title": "A Consistent Pattern for Identifying Decisive Code Snippets for LLM-Based Code Inference", "abs": "Which parts of pre-target input are most influential for next-token prediction in the context of programming languages? In this paper, we present evidence that code snippets at specific locations in pre-target inputs play a decisive role in large language model (LLM) inference, and these snippets exhibit a consistent pattern. Firstly, we introduce a novel causal tracing method to identify tokens, so-called high-information tokens, that significantly contribute to next-token prediction. Building on this, we propose a multi-phase causal tracing process to analyze the importance distribution of high-information tokens, revealing a consistent pattern, named the Important Position Rule (IPR). To further validate this hypothesis, we assess the role of IPR across various LLMs, languages, and tasks. Our extensive evaluations for code translation, code correction and code completion tasks (Java, Python, C++) on models CodeLlama-7b/13b/34b-Instruct and GPT-3.5/4-turbo, confirm this hypothesis. Furthermore, we observe that IPR exhibits structural and semantic properties similar to the $\\langle \\text{subject}, \\text{relation}, \\text{object} \\rangle$ paradigm in natural language. Leveraging this insight, we successfully combine IPR with the knowledge editing method ROME in order to repair translation errors, achieving a correction rate of 62.73% to 75.31%. To our knowledge, this is the first application of knowledge editing in the context of programming languages.", "keywords": [ "Programming Languages; interpretability; Knowledge editing" ], "rating_list": [ 6, 3, 6 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 1, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "Qg6Z3VcA1U", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Changyeon Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Dongyoon Hahm", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Undergrad student" }, { "name": "Haeone Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Intern" }, { "name": "Juyong Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" }, { "name": "Kimin Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Minyong An", "gender": "Male", "institution": "KAIST", "country": "", "position": "Intern" }, { "name": "Taywon Min", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 51, "n_ref": 83, "n_ref_all": 128, "n_fig": 11, "n_tab": 18, "L_tab": 17003, "n_element_tab": 1579, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 136, "formula_len_all_1": 50, "len_all": 245322, "len_all_1": 58989, "len_abs": 1139, "len_title": 128, "len_sents": 71214, "len_sents_1": 28792, "n_sents": 552, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1142, "title": "B-MoCA: Benchmarking Mobile Device Control Agents across Diverse Configurations", "abs": "Mobile device control agents can largely enhance user interactions or productivity by automating daily tasks. However, despite growing interest in developing practical agents, the absence of a commonly adopted benchmark in this area makes it challenging to quantify scientific progress. In this work, we introduce B-MoCA: a novel benchmark with interactive environments for evaluating and developing mobile device control agents. To create a realistic benchmark, we develop B-MoCA based on the Android operating system and define 131 common daily tasks. Importantly, we incorporate a randomization feature that changes the configurations of mobile devices, including user interface layouts and language settings, to assess generalization performance. We benchmark diverse agents, including agents employing large language models (LLMs) or multi-modal LLMs as well as agents trained with imitation learning using human expert demonstrations. While these agents demonstrate proficiency in executing straightforward tasks, their poor performance on complex tasks highlights significant opportunities for future research to improve effectiveness.", "keywords": [ "Mobile device control", "Benchmark", "Decision-making agents", "LLM agents" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 1, 2, 1, 2, 2 ], "confidence_list": [ 4, 3, 4, 5, 5 ] }, { "paper_id": "Qg0gtNkXIb", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Chunsan Hong", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Undergrad student" }, { "name": "Minhyuk Sung", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Tae-hyun Oh", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 15, "n_ref_uni": 24, "n_ref": 126, "n_ref_all": 161, "n_fig": 6, "n_tab": 6, "L_tab": 3702, "n_element_tab": 518, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1610, "n_element_tab_1": 301, "formula_len_all": 1826, "formula_len_all_1": 913, "len_all": 212201, "len_all_1": 76891, "len_abs": 3237, "len_title": 117, "len_sents": 78390, "len_sents_1": 33603, "n_sents": 694, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1314, "title": "MemBench: Memorized Image Trigger Prompt Dataset for Diffusion Models", "abs": "Diffusion models have achieved remarkable success in Text-to-Image generation tasks, leading to the development of many commercial models. However, recent studies have reported that diffusion models often repeatedly generate memorized images in train data when triggered by specific prompts, potentially raising social issues ranging from copyright to privacy concerns. To sidestep the memorization, there have been recent studies for developing memorization mitigation methods for diffusion models. Nevertheless, the lack of benchmarks hinders the assessment of the true effectiveness of these methods. In this work, we present MemBench, the first benchmark for evaluating image memorization mitigation methods. Our benchmark includes a large number of memorized image trigger prompts in various Text-to-Image diffusion models. Furthermore, in contrast to the prior work evaluating mitigation performance only on trigger prompts, we present metrics evaluating on both trigger prompts and general prompts, so that we can see whether mitigation methods address the memorization issue while maintaining performance for general prompts. Through our MemBench evaluation, we revealed that existing memorization mitigation methods notably degrade overall performance of diffusion models and need to be further developed.", "keywords": [ "diffusion models", "MCMC sampling", "image memorization", "safety ai" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "QfyZ28FpVY", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Chang-Yu Hsieh", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Chunbin Gu", "gender": "Male", "institution": "Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Hemutian", "gender": "Male", "institution": "Lanzhou University", "country": "", "position": "MS student" }, { "name": "Ning Ma", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Postdoc" }, { "name": "Pheng-Ann Heng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Hanqun CAO", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 17, "n_ref_uni": 36, "n_ref": 60, "n_ref_all": 74, "n_fig": 3, "n_tab": 6, "L_tab": 2364, "n_element_tab": 206, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 785, "n_element_tab_1": 46, "formula_len_all": 2119, "formula_len_all_1": 1187, "len_all": 144973, "len_all_1": 63161, "len_abs": 1750, "len_title": 162, "len_sents": 45899, "len_sents_1": 31031, "n_sents": 376, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 113, "L_abs": 1759, "title": "DEL-Ranking: Ranking-Correction Denoising Framework for Elucidating Molecular Affinities in DNA-Encoded Libraries", "abs": "DNA-encoded library (DEL) screening has revolutionized protein-ligand binding detection, enabling rapid exploration of vast chemical spaces through read count analysis. However, two critical challenges limit its effectiveness: distribution noise in low copy number regimes and systematic shifts between read counts and true binding affinities. We present DEL-Ranking, a comprehensive framework that simultaneously addresses both challenges through innovative ranking-based denoising and activity-referenced correction. Our approach introduces a dual-perspective ranking strategy combining Pair-wise Soft Rank (PSR) and List-wise Global Rank (LGR) constraints to preserve both local and global count relationships. Additionally, we develop an Activity-Referenced Correction (ARC)\nmodule that bridges the gap between read counts and binding affinities through iterative refinement and biological consistency enforcement. Another key contribution of this work is the curation and release of three comprehensive DEL datasets that uniquely combine ligand 2D sequences, 3D conformational information, and experimentally validated activity labels. We validate our framework on\nfive diverse DEL datasets and introduce three new comprehensive datasets featuring 2D sequences, 3D structures, and activity labels. DEL-Ranking achieves state-of-the-art performance across multiple correlation metrics and demonstrates strong generalization ability across different protein targets. Importantly, our approach successfully identifies key functional groups associated with binding affinity, providing actionable insights for drug discovery. This work advances both the accuracy and interpretability of DEL screening, while contributing valuable datasets for future research.", "keywords": [ "DEL Denoising", "Deep Learning", "Bioinformatics" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 1, 1, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "QfhU3ZC2g1", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Haodong Liang", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "PhD student" }, { "name": "Lifeng Lai", "gender": "unknown", "institution": "University of California, Davis", "country": "US", "position": "Full Professor" }, { "name": "Krishnakumar Balasubramanian", "gender": "Male", "institution": "University of California, Davis", "country": "", "position": "Associate Professor" } ], "n_formula": 185, "n_formula_1": 45, "n_ref_uni": 37, "n_ref": 59, "n_ref_all": 113, "n_fig": 12, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 97, "n_element_tab_1": 2, "formula_len_all": 21675, "formula_len_all_1": 4524, "len_all": 193727, "len_all_1": 59809, "len_abs": 859, "len_title": 111, "len_sents": 45937, "len_sents_1": 23182, "n_sents": 536, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 861, "title": "Transformers Handle Endogeneity in In-Context Linear Regression", "abs": "We explore the capability of transformers to address endogeneity in in-context linear regression. Our main finding is that transformers inherently possess a mechanism to handle endogeneity effectively using instrumental variables (IV). First, we demonstrate that the transformer architecture can emulate a gradient-based bi-level optimization procedure that converges to the widely used two-stage least squares (2SLS) solution at an exponential rate. Next, we propose an in-context pretraining scheme and provide theoretical guarantees showing that the global minimizer of the pre-training loss achieves a small excess loss. Our extensive experiments validate these theoretical findings, showing that the trained transformer provides more robust and reliable in-context predictions and coefficient estimates than the 2SLS method, in the presence of endogeneity.", "keywords": [ "endogeneity", "instrumental variables", "transformers", "bi-level gradient descent" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 2, 1 ] }, { "paper_id": "QfGc9txfGO", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Changsheng Xu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Hantao Yao", "gender": "Male", "institution": ",Institute of automation, Chinese academy of science", "country": "CN", "position": "Associate Professor" }, { "name": "Lu Yu", "gender": "Female", "institution": "Tianjin University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Zhe Tao", "gender": "Male", "institution": "Tianjin University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 24, "n_ref": 35, "n_ref_all": 51, "n_fig": 3, "n_tab": 6, "L_tab": 1975, "n_element_tab": 288, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 612, "n_element_tab_1": 64, "formula_len_all": 611, "formula_len_all_1": 564, "len_all": 130544, "len_all_1": 60752, "len_abs": 1557, "len_title": 110, "len_sents": 32418, "len_sents_1": 28425, "n_sents": 187, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1576, "title": "Revisiting Prompt-based Methods in Class Incremental Learning", "abs": "In recent years, prompt-based methods have emerged as a promising direction for continual learning, demonstrating impressive performance across various benchmarks. These methods create learnable prompts to infer task identity, then select and integrate specific prompts into the pretrained model to generate instructed features for prediction. In this paper, we first analyze the working patterns of such method across different distribution scenarios through extensive empirical analysis. Our analysis exposes the limitations of existing methods: first, two-stage inference can make mistakes even when the first stage has already provided reliable predictions; second, enforcing identical architectures for both stages hampers performance gains. To address these issues, we incorporated a self-supervised learning objective to learn discriminative features, thereby boosting the plasticity of the model. During inference, we implemented a simple yet effective threshold filtering strategy to selectively pass data to the second stage. This approach prevents errors in the second stage when the first stage has already made reliable predictions, while also conserving computational resources. Ultimately, we explore utilizing self-supervised pretrained models as a unified task identity provider. Comparing to state-of-the-art methods, our method achieves comparable results under in-distribution scenarios and demonstrates substantial gains under out-of-distribution scenarios (e.g., up to 6.34\\% and 5.15\\% improvements on Split Aircrafts and Split Cars-196, respectively).", "keywords": [ "continual learning", "pretrained-models", "self-supervised learning" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "QeYnKwFhsU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Deyu Bo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Xinchao Wang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 30, "n_formula_1": 25, "n_ref_uni": 51, "n_ref": 86, "n_ref_all": 110, "n_fig": 9, "n_tab": 7, "L_tab": 3280, "n_element_tab": 254, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 3031, "n_element_tab_1": 154, "formula_len_all": 2147, "formula_len_all_1": 1738, "len_all": 148684, "len_all_1": 70864, "len_abs": 1413, "len_title": 80, "len_sents": 36894, "len_sents_1": 29036, "n_sents": 329, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 32, "L_abs": 1420, "title": "Point Cloud Dataset Distillation", "abs": "This study introduces dataset distillation (DD) tailored for 3D data, particularly point clouds. DD aims to substitute large-scale real datasets with a small set of synthetic samples while preserving model performance. Existing methods mainly focus on structured data such as images. However, adapting DD for unstructured point clouds poses challenges due to their diverse orientations and resolutions in 3D space. To address these challenges, we theoretically demonstrate the importance of matching rotation-invariant features between real and synthetic data for 3D distillation. We further propose a plug-and-play point cloud rotator to align the point cloud to a canonical orientation, facilitating the learning of rotation-invariant features by all point cloud models. Furthermore, instead of optimizing fixed-size synthetic data directly, we devise a point-wise generator to produce point clouds at various resolutions based on the sampled noise amount. Compared to conventional DD methods, the proposed approach, termed DD3D, enables efficient training on low-resolution point clouds while generating high-resolution data for evaluation, thereby significantly reducing memory requirements and enhancing model scalability. Extensive experiments validate the effectiveness of DD3D in shape classification and part segmentation tasks across diverse scenarios, such as cross-architecture and cross-resolution settings.", "keywords": [ "Dataset Distillation", "Point Cloud", "Geometric Deep Learning" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 2 ] }, { "paper_id": "Qdp7hlenr6", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 5, "author_info_list": [ { "name": "Brando Miranda", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Jakob Nordhagen", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "MS student" }, { "name": "Michael Souliman", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Willy Chan", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 11, "n_ref": 29, "n_ref_all": 49, "n_fig": 6, "n_tab": 2, "L_tab": 351, "n_element_tab": 24, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 351, "n_element_tab_1": 24, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 88473, "len_all_1": 61313, "len_abs": 1955, "len_title": 144, "len_sents": 35351, "len_sents_1": 30104, "n_sents": 231, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1965, "title": "Lean-ing on Quality: How High-Quality Data Beats Diverse Multilingual Data in AutoFormalization", "abs": "Autoformalization, the process of transforming informal mathematical language into formal specifications and proofs remains a difficult task for state-of-the-art (large) language models. Existing works point to competing explanations for the performance gap. On one hand, large language models exhibit exceptional performance on translation tasks, suggesting their significant potential for autoformalization. On the other hand, the quantitative reasoning capabilities of standard language models remain limited, leading to suboptimal performance on autoformalization and the subsequent task of formal theorem proving. To this end, we introduce a novel methodology that leverages backtranslation with hand-curated prompts to enhance the mathematical capabilities of language models, particularly addressing the challenge posed by the scarcity of labeled data. Specifically, we evaluate three primary variations of this strategy: (1) on-the-fly (online) backtranslation, (2) distilled (offline) backtranslation with few-shot amplification, and (3) line-by-line proof analysis integrated with proof state information. Each variant is designed to optimize data quality over quantity, focusing on the high fidelity of generated proofs rather than sheer data scale. Our findings provide evidence that employing our proposed approaches to generate synthetic data, which prioritizes quality over volume, improves the autoformalization performance of LLMs as measured by standard benchmarks such as ProofNet. Crucially, our approach outperforms pretrained models using a minimal number of tokens. We also show, through strategic prompting and backtranslation, that our approaches surpass the performance of finetuning with extensive multilingual datasets such as MMA on ProofNet with only 1/150th of the tokens. Taken together, our methods show a promising new approach to significantly reduce the resources required to formalize proofs, thereby accelerating AI for math.", "keywords": [ "llm", "large language models", "autoformalization", "lean" ], "rating_list": [ 3, 3, 6, 5, 3 ], "soundness_list": [ 2, 2, 4, 2, 3 ], "presentation_list": [ 2, 3, 4, 2, 3 ], "contribution_list": [ 2, 2, 4, 2, 2 ], "confidence_list": [ 5, 4, 5, 3, 4 ] }, { "paper_id": "QdiMWcwU5w", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Binfeng Xu", "gender": "Male", "institution": "Samsung Research America", "country": "US", "position": "Research Engineer" }, { "name": "Haoyan Yang", "gender": "Male", "institution": "New York University", "country": "US", "position": "MS student" }, { "name": "Hongxia Jin", "gender": "Female", "institution": "Head of AI center", "country": "US", "position": "Vice President" }, { "name": "Jie Xu", "gender": "Female", "institution": "University of Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Shangqian Gao", "gender": "unknown", "institution": "Florida State University", "country": "US", "position": "Assistant Professor" }, { "name": "Ting Hua", "gender": "unknown", "institution": "Samsung", "country": "US", "position": "Researcher" }, { "name": "Vijay Srinivasan", "gender": "Male", "institution": "Samsung", "country": "", "position": "Researcher" }, { "name": "Zheng Tang", "gender": "Male", "institution": "Samsung", "country": "", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 19, "n_ref_uni": 23, "n_ref": 31, "n_ref_all": 49, "n_fig": 10, "n_tab": 7, "L_tab": 2714, "n_element_tab": 299, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 882, "n_element_tab_1": 78, "formula_len_all": 1668, "formula_len_all_1": 1668, "len_all": 119385, "len_all_1": 45827, "len_abs": 1183, "len_title": 129, "len_sents": 34305, "len_sents_1": 19486, "n_sents": 271, "n_sents_1": 149, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1192, "title": "Dynamic Noise Preference Optimization for LLM Self-Improvement via Synthetic Data", "abs": "Although LLMs have achieved significant success, their reliance on large volumes of human-annotated data has limited their potential for further scaling. In this situation, utilizing self-generated synthetic data has become crucial for fine-tuning LLMs without extensive human annotation. However, current methods often fail to ensure consistent improvements across iterations, with performance stagnating after only minimal updates. To overcome these challenges, we introduce Dynamic Noise Preference Optimization (DNPO). DNPO employs a dynamic sample labeling mechanism to construct preference pairs for training and introduces controlled, trainable noise into the preference optimization process. Our approach effectively prevents stagnation and enables continuous improvement. In experiments with Zephyr-7B, DNPO consistently outperforms existing methods, showing an average performance boost of 2.6\\% across multiple benchmarks. \nAdditionally, DNPO shows a significant improvement in model-generated data quality, with a 29.4\\% win-loss rate gap compared to the baseline in GPT-4 evaluations. This highlights its effectiveness in enhancing model performance through iterative refinement.", "keywords": [ "Preference Optimization", "Alignment of LLMs", "Self-improvement", "Synthetic Data", "Noise" ], "rating_list": [ 6, 6, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 2, 2, 2 ] }, { "paper_id": "QcvwVUqnCg", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Blake Bordelon", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Cengiz Pehlevan", "gender": "unknown", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "Jacob A Zavatone-Veth", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "M Ganesh Kumar", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" } ], "n_formula": 58, "n_formula_1": 9, "n_ref_uni": 84, "n_ref": 169, "n_ref_all": 252, "n_fig": 47, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 41, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4018, "formula_len_all_1": 594, "len_all": 236408, "len_all_1": 73725, "len_abs": 1281, "len_title": 64, "len_sents": 72036, "len_sents_1": 36376, "n_sents": 533, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1286, "title": "A Model of Place Field Reorganization During Reward Maximization", "abs": "When rodents learn to navigate in a novel environment, a high density of place fields emerge at reward locations, fields elongate against the trajectory, and individual fields change spatial selectivity while demonstrating stable behavior. Why place fields demonstrate these characteristic phenomena during learning remains elusive. We develop a normative framework using a reward maximization objective, whereby the temporal difference (TD) error drives place field reorganization to improve policy learning. Place fields are modelled using Gaussian radial basis functions to represent states in an environment, and directly synapse to an actor-critic for policy learning. Each field's amplitude, center and width, as well as downstream weights, are updated online at each time step to maximize cumulative reward. We demonstrate that this framework unifies the three disparate phenomena observed in navigation experiments. Furthermore, we show that these place field phenomena improves policy convergence when learning to navigate to a single target and relearning multiple new targets. To conclude, we develop a normative model that recapitulates several aspects of hippocampal place field learning dynamics and unifies mechanisms to offer testable predictions for future experiments.", "keywords": [ "Place field", "Reinforcement Learning", "Successor representation", "Representation drift" ], "rating_list": [ 8, 6, 3, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "QarKTT5brZ", "primary_area": "datasets and benchmarks", "n_author": 20, "author_info_list": [ { "name": "Chenlong Wang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "", "position": "Undergrad student" }, { "name": "Chujie Gao", "gender": "Female", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Intern" }, { "name": "Dongping Chen", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Intern" }, { "name": "Huichi Zhou", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "MRes" }, { "name": "Jianfeng Gao", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Jingyu Tang", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Lichao Sun", "gender": "Male", "institution": "Lehigh University", "country": "US", "position": "Assistant Professor" }, { "name": "Liuyi Chen", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Pan Zhou", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Professor" }, { "name": "Qihui Zhang", "gender": "Male", "institution": "The University of Queensland", "country": "AU", "position": "Intern" }, { "name": "Siyuan Wu", "gender": "unknown", "institution": "University of Waterloo", "country": "CA", "position": "Intern" }, { "name": "Tianshuo Zhou", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Yao Wan", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yi Gui", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yilin Bai", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "MS student" }, { "name": "Yiqiang Li", "gender": "Female", "institution": "Royal College of Art", "country": "GB", "position": "MS student" }, { "name": "Yue Huang", "gender": "unknown", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "Yue Yu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Zhen Li", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Zhigang He", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 52, "n_ref": 137, "n_ref_all": 210, "n_fig": 30, "n_tab": 21, "L_tab": 14222, "n_element_tab": 1486, "n_fig_1": 6, "n_tab_1": 11, "L_tab_1": 6405, "n_element_tab_1": 453, "formula_len_all": 153, "formula_len_all_1": 44, "len_all": 287606, "len_all_1": 64409, "len_abs": 215, "len_title": 114, "len_sents": 66096, "len_sents_1": 24137, "n_sents": 528, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 65, "L_abs": 1768, "title": "GUI-World: A GUI-oriented Dataset for Multimodal LLM-based Agents", "abs": "Recently, Multimodal Large Language Models (MLLMs) have been used as agents to control keyboard and mouse inputs by directly perceiving the Graphical User Interface (GUI) and generating corresponding commands. However, current agents primarily demonstrate strong understanding capabilities in static environments and are mainly applied to relatively simple domains, such as Web or mobile interfaces. We argue that a robust GUI agent should be capable of perceiving temporal information on the GUI, including dynamic Web content and multi-step tasks. Additionally, it should possess a comprehensive understanding of various GUI scenarios, including desktop software and multi-window interactions. To this end, this paper introduces a new dataset, termed GUI-World, which features meticulously crafted Human-MLLM annotations, extensively covering six GUI scenarios and eight types of GUI-oriented questions in three formats. We evaluate the capabilities of current state-of-the-art MLLMs, including Image LLMs and Video LLMs, in understanding various types of GUI content, especially dynamic and sequential content. Our findings reveal that vision LLMs struggle with dynamic GUI content without manually annotated keyframes or operation history. On the other hand, video LLMs fall short in all GUI-oriented tasks given the sparse GUI video dataset. Therefore, we take the initial step of leveraging a fine-tuned video LLM as a GUI agent based on GUI-World, demonstrating an improved understanding of various GUI tasks. However, due to the limitations in the performance of base LLMs, we conclude that using video LLMs as GUI agents remains a significant challenge. We believe our work provides valuable insights for future research in dynamic GUI content understanding.", "keywords": [ "GUI", "GUI Agent", "Multimodal Large Language Model", "Benchmark", "Dataset", "Video LLM", "Instruction Tuning" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "QaTBHSqmH9", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Jiacheng Cen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Le Song", "gender": "Male", "institution": "College of Computing, Georgia Institute of Technology", "country": "", "position": "Associate Professor" }, { "name": "Taifeng Wang", "gender": "Male", "institution": "BioMap", "country": "", "position": "Principal Researcher" }, { "name": "Wenbing Huang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zongzhao Li", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 14, "n_ref_uni": 42, "n_ref": 76, "n_ref_all": 103, "n_fig": 7, "n_tab": 10, "L_tab": 3304, "n_element_tab": 413, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 385, "n_element_tab_1": 53, "formula_len_all": 3662, "formula_len_all_1": 1299, "len_all": 184874, "len_all_1": 65516, "len_abs": 1431, "len_title": 129, "len_sents": 56660, "len_sents_1": 29439, "n_sents": 416, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1426, "title": "Size-Generalizable RNA Structure Evaluation by Exploring Hierarchical Geometries", "abs": "Understanding the 3D structure of RNA is essential for deciphering its function and developing RNA-based therapeutics. Geometric Graph Neural Networks (GeoGNNs) that conform to the $\\mathrm{E}(3)$-symmetry have advanced RNA structure evaluation, a crucial step toward RNA structure prediction. However, existing GeoGNNs are still defective in two aspects: 1. inefficient or incapable of capturing the full geometries of RNA; 2. limited generalization ability when the size of RNA significantly differs between training and test datasets. In this paper, we propose EquiRNA, a novel equivariant GNN model by exploring the three-level hierarchical geometries of RNA. At its core, EquiRNA effectively addresses the size generalization challenge by reusing the representation of nucleotide, the common building block shared across RNAs of varying sizes. Moreover, by adopting a scalarization-based equivariant GNN as the backbone, our model maintains directional information while offering higher computational efficiency compared to existing GeoGNNs. Additionally, we propose a size-insensitive $K$-nearest neighbor sampling strategy to enhance the model's robustness to RNA size shifts. We test our approach on our created benchmark as well as an existing dataset. The results show that our method significantly outperforms other state-of-the-art methods, providing a robust baseline for RNA 3D structure modeling and evaluation.", "keywords": [ "RNA Structure", "RNA Evaluation", "Geometric Deep Learning", "Graph Neural Networks" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "QaQrWKPWdG", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Alexander Shepard", "gender": "Male", "institution": "iNaturalist", "country": "", "position": "Software Developer" }, { "name": "Christian Lange", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Elijah Cole", "gender": "Male", "institution": "Altos Labs", "country": "US", "position": "Researcher" }, { "name": "Grant Van Horn", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Max Hamilton", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" }, { "name": "Samuel Heinrich", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Researcher" }, { "name": "Subhransu Maji", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Associate Professor" }, { "name": "Oisin Mac Aodha", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 97, "n_ref_all": 153, "n_fig": 34, "n_tab": 8, "L_tab": 3913, "n_element_tab": 495, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 202, "n_element_tab_1": 39, "formula_len_all": 373, "formula_len_all_1": 202, "len_all": 229571, "len_all_1": 64138, "len_abs": 1117, "len_title": 81, "len_sents": 91677, "len_sents_1": 32691, "n_sents": 697, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 33, "L_abs": 1121, "title": "Few-shot Species Range Estimation", "abs": "Understanding where a particular species can or cannot be found is crucial for ecological research and conservation efforts. By mapping the spatial ranges of all species on Earth, we could obtain deeper insights into how global biodiversity is affected by climate change and habitat loss. However, accurate range estimates are available for a relatively small proportion of known species. For most species, we have only have a few prior observations indicating the locations where they have been previously recorded. In this work we address the challenge of training with limited observations by developing a new approach for few-shot species range estimation. During inference, our model takes a set of spatial coordinates as input, along with optional metadata such as text, and outputs a species encoding that can be used to predict the range of a previously unseen species in feed-forward manner. We validate our method on two challenging benchmarks, where we obtain state-of-the-art performance in predicting the ranges of unseen species, in a fraction of the compute time, compared to recent alternative approaches.", "keywords": [ "species distribution modeling", "SDM", "spatial implicit neural representation", "SINR", "low-shot learning", "few-shot learning" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "Qa6VvpE2Py", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Ashwin Swaminathan", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Hao Yang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Pei Wang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "R. Manmatha", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Senior Principal Scientist" }, { "name": "Zhaowei Cai", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Stefano Soatto", "gender": "unknown", "institution": "Amazon Web Services", "country": "US", "position": "Vice President" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 26, "n_ref": 58, "n_ref_all": 84, "n_fig": 4, "n_tab": 10, "L_tab": 2687, "n_element_tab": 27, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2854, "n_element_tab_1": 23, "formula_len_all": 73, "formula_len_all_1": 73, "len_all": 160810, "len_all_1": 58055, "len_abs": 1474, "len_title": 99, "len_sents": 36933, "len_sents_1": 29189, "n_sents": 245, "n_sents_1": 168, "venue": "ICLR2025", "decision": "withdraw", "L_title": 51, "L_abs": 1482, "title": "Scaling up Image Segmentation across Data and Tasks", "abs": "Traditional segmentation models, while effective in isolated tasks, often fail to generalize to more complex and open-ended segmentation problems, such as free-form, open-vocabulary, and in-the-wild scenarios. To bridge this gap, we propose to scale up image segmentation across diverse datasets and tasks such that the knowledge across different tasks and datasets can be integrated while improving the generalization ability. QueryMeldNet, a novel segmentation framework, is introduced and designed to scale seamlessly across both data size and task diversity. It is built upon a dynamic object query mechanism called query meld, which fuses different types of queries using cross-attention. This hybrid approach enables the model to balance between instance- and stuff-level segmentation, providing enhanced scalability for handling diverse object types. We further enhance scalability by leveraging synthetic data-generating segmentation masks and captions for pixel-level and open-vocabulary tasks-drastically reducing the need for costly human annotations. By training on multiple datasets and tasks at scale, QueryMeldNet continuously improves performance as the volume and diversity of data and tasks increase. It exhibits strong generalization capabilities, boosting performance in open-set segmentation tasks SeginW by 7 points. These advancements mark a key step toward universal, scalable segmentation models capable of addressing the demands of real-world applications.", "keywords": [ "image segmentation", "scalable", "open-vocabulary" ], "rating_list": [ 5, 3, 5, 6, 5 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 1, 3, 3 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "Qa40qfZooj", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Daniel P Jeong", "gender": "unknown", "institution": "Machine Learning Department, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Dylan Sam", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Rattana Pukdee", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Yewon Byun", "gender": "unknown", "institution": "Machine Learning Department, School of Computer Science", "country": "", "position": "PhD student" }, { "name": "Zico Kolter", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 58, "n_ref": 116, "n_ref_all": 149, "n_fig": 7, "n_tab": 7, "L_tab": 2416, "n_element_tab": 222, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 498, "n_element_tab_1": 46, "formula_len_all": 1068, "formula_len_all_1": 581, "len_all": 172183, "len_all_1": 68304, "len_abs": 1250, "len_title": 101, "len_sents": 53457, "len_sents_1": 33755, "n_sents": 354, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1253, "title": "Bayesian Neural Networks with Domain Knowledge Priors", "abs": "Bayesian neural networks (BNNs) have recently gained popularity due to their ability to quantify model uncertainty in prediction. However, specifying a prior for BNNs that accurately captures relevant domain knowledge is often extremely challenging. In this work, we propose a framework for integrating general forms of domain knowledge (i.e., any knowledge that can be represented by a loss function) into a BNN prior through variational inference, while enabling computationally efficient posterior inference and sampling. Specifically, our approach results in a prior over neural network weights that assigns high probability mass to models that better align with our domain knowledge, leading to posterior samples that also exhibit this behavior. In a semi-supervised learning setting, we show that BNNs using our proposed domain knowledge priors outperform those with standard priors (e.g., isotropic Gaussian, Gaussian process), successfully incorporating diverse types of prior information such as fairness, physics rules, and healthcare knowledge and achieving better predictive performance. We also present techniques for transferring the learned priors across different model architectures, demonstrating their broad utility across many tasks.", "keywords": [ "bayesian neural networks", "domain knowledge" ], "rating_list": [ 6, 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "QZuZmfLLRG", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Chao Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Da-Wei Zhou", "gender": "unknown", "institution": "Nanjing University", "country": "CN", "position": "Associate Researcher" }, { "name": "Hanbin Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Hui Qian", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "qian feng", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 50, "n_formula_1": 26, "n_ref_uni": 26, "n_ref": 71, "n_ref_all": 107, "n_fig": 5, "n_tab": 20, "L_tab": 10039, "n_element_tab": 1105, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2059, "n_element_tab_1": 198, "formula_len_all": 2944, "formula_len_all_1": 1076, "len_all": 240969, "len_all_1": 71841, "len_abs": 1657, "len_title": 114, "len_sents": 61128, "len_sents_1": 30174, "n_sents": 462, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1666, "title": "LW2G: Learning Whether to Grow for Prompt-based Continual Learning", "abs": "Continual Learning (CL) aims to learn in non-stationary scenarios, progressively acquiring and maintaining knowledge from sequential tasks. Recent Prompt-based Continual Learning (PCL) has achieved remarkable performance with Pre-Trained Models (PTMs). These approaches grow a prompt sets pool by adding a new set of prompts when learning each new task (prompt learning) and adopt a matching mechanism to select the correct set for each testing sample (prompt retrieval). Previous studies focus on the latter stage by improving the matching mechanism to enhance Prompt Retrieval Accuracy (PRA). To promote cross-task knowledge facilitation and form an effective and efficient prompt sets pool, we propose a plug-in module in the former stage to Learn Whether to Grow (LW2G) based on the disparities between tasks. Specifically, a shared set of prompts is utilized when several tasks share certain commonalities, and a new set is added when there are significant differences between the new task and previous tasks. Inspired by Gradient Projection Continual Learning, our LW2G develops a metric called Hinder Forward Capability (HFC) to measure the hindrance imposed on learning new tasks by surgically modifying the original gradient onto the orthogonal complement of the old feature space. With HFC, an automated scheme Dynamic Growing Approach adaptively learns whether to grow with a dynamic threshold. Furthermore, we design a gradient-based constraint to ensure the consistency between the updating prompts and pre-trained knowledge, and a prompts weights reusing strategy to enhance forward transfer. Extensive experiments show the effectiveness of our method.", "keywords": [ "Continual Learning", "Gradient Orthogonal", "Pre-Trained Model", "Prompt Learning" ], "rating_list": [ 6, 5, 5, 6, 3 ], "soundness_list": [ 3, 3, 2, 4, 2 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 4, 5 ] }, { "paper_id": "QYvtX2XA8p", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Cong Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": " LLM Algorithm Research Scientist" }, { "name": "Hao Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Kuicai Dong", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "SG", "position": "Researcher" }, { "name": "Liu Huanshuo", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Wang Jing", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Yi Quan Lee", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Yong Liu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "SG", "position": "Principal Researcher" }, { "name": "Zhijiang Guo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiangyang Li", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 44, "n_ref": 131, "n_ref_all": 171, "n_fig": 10, "n_tab": 10, "L_tab": 4508, "n_element_tab": 274, "n_fig_1": 4, "n_tab_1": 9, "L_tab_1": 3096, "n_element_tab_1": 235, "formula_len_all": 630, "formula_len_all_1": 476, "len_all": 271718, "len_all_1": 73362, "len_abs": 1153, "len_title": 116, "len_sents": 71860, "len_sents_1": 29422, "n_sents": 568, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1163, "title": "CtrlA: Adaptive Retrieval-Augmented Generation via Inherent Control", "abs": "Retrieval-augmented generation (RAG) has emerged as a promising solution for mitigating hallucinations of large language models (LLMs) with retrieved external knowledge. Adaptive RAG enhances this approach by enabling dynamic retrieval during generation, activating retrieval only when the query exceeds LLM's internal knowledge. Existing methods primarily focus on detecting LLM's confidence via statistical uncertainty. Instead, we present the first attempts to solve adaptive RAG from a representation perspective and develop an inherent control-based framework, termed CtrlA. Specifically, we extract the features that represent the honesty and confidence directions of LLM and adopt them to control LLM behavior and guide retrieval timing decisions. We also design a simple yet effective query formulation strategy to support adaptive retrieval. Experiments show that CtrlA is superior to existing adaptive RAG methods on a diverse set of tasks, the honesty steering can effectively make LLMs more honest and confidence monitoring is a promising indicator of retrieval trigger. Our anonymous codes are submitted with the paper and will be publicly available.", "keywords": [ "Retrieval-Augmented Generation", "Large Language Model" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "QYigQ6gXNw", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Ahmed Hassan Awadallah", "gender": "Male", "institution": "Microsoft Research", "country": "", "position": "Principal Researcher" }, { "name": "Akshay Krishnamurthy", "gender": "Male", "institution": "Research, Microsoft", "country": "US", "position": "Senior Principal Research Manager" }, { "name": "Alexander Rakhlin", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Dylan J Foster", "gender": "unknown", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Tengyang Xie", "gender": "unknown", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Corbin L Rosset", "gender": "Male", "institution": "Research, Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 191, "n_formula_1": 38, "n_ref_uni": 57, "n_ref": 190, "n_ref_all": 254, "n_fig": 0, "n_tab": 5, "L_tab": 741, "n_element_tab": 80, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 220, "n_element_tab_1": 14, "formula_len_all": 20455, "formula_len_all_1": 2931, "len_all": 300093, "len_all_1": 81669, "len_abs": 3687, "len_title": 118, "len_sents": 74438, "len_sents_1": 34187, "n_sents": 771, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 118, "L_abs": 1267, "title": "Exploratory Preference Optimization: Provably Sample-Efficient Exploration in RLHF with General Function Approximation", "abs": "This paper investigates a basic question in reinforcement learning from human feedback (RLHF) from a theoretical perspective: how to efficiently explore in an online manner under preference feedback and general function approximation. We take the initial step towards a theoretical understanding of this problem by proposing a novel algorithm, *Exploratory Preference Optimization* (XPO). This algorithm is elegantly simple---requiring only a one-line modification to (online) Direct Preference Optimization (DPO; Rafailov et al., 2023)---yet provides the strongest known provable guarantees. XPO augments the DPO objective with a novel and principled *exploration bonus*, enabling the algorithm to strategically explore beyond the support of the initial model and preference feedback data. We prove that XPO is provably sample-efficient and converges to a near-optimal policy under natural exploration conditions, regardless of the initial model's coverage. Our analysis builds on the observation that DPO implicitly performs a form of *Bellman error minimization*. It synthesizes previously disparate techniques from language modeling and theoretical reinforcement learning in a serendipitous fashion through the lens of *KL-regularized Markov decision processes*.", "keywords": [ "Learning theory", "Reinforcement learning theory", "Sample-efficient reinforcement learning" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "QYgtZRTv3e", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Kang Dang", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "CN", "position": "Associate Professor" }, { "name": "Liang Peng", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Mian Zhou", "gender": "Male", "institution": "Tianjin University of Technology", "country": "", "position": "Associate Professor" }, { "name": "Mingyu Wu", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "PING KUANG", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yikai Zhao", "gender": "Male", "institution": "Chengdu University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Zhikun Feng", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 29, "n_ref": 66, "n_ref_all": 91, "n_fig": 5, "n_tab": 9, "L_tab": 3338, "n_element_tab": 365, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 3466, "n_element_tab_1": 338, "formula_len_all": 556, "formula_len_all_1": 556, "len_all": 121573, "len_all_1": 74406, "len_abs": 1640, "len_title": 108, "len_sents": 34769, "len_sents_1": 31080, "n_sents": 275, "n_sents_1": 243, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1650, "title": "TIPS: Two-Level Prompt for Rehearsal-free Continual Learning", "abs": "Continual learning based on prompt tuning creates a key-value pool, where these key-value pairs are called prompts. Prompts are retrieved using input images as queries and input into a frozen backbone network. It requires training only a few parameters to quickly adapt to downstream tasks. Compared to other traditional Continual learning methods, it is more effective in resisting catastrophic forgetting. However, the effectiveness of these methods heavily depends on the selection strategy. \nMost existing methods overlook the model plasticity since they focus on solving the model's stability issues, leading to a sharp decline in performance for new tasks in long task sequences of incremental learning.\nTo address these limitations, we propose a novel prompt-based continual learning method called TIPS, which mainly consists of two modules: (1) design a novel two-level prompt selection strategy combined with a set of adaptive weights for sparse joint tuning, aiming to improve the accuracy of prompt selection; (2) design a semantic distillation module that enhances the generalization ability to unknown new classes by creating a language token and utilizing the encapsulated semantic information of class names.\nWe validated TIPS on four datasets across three incremental scenarios. \nOur method outperformed the current state of the art (SOTA) by 2.03%, 4.78%, 1.18%, and 5.59% on CIFAR (10 tasks), ImageNet-R (20 tasks), CUB (10 tasks), and DomainNet (20 tasks). \nNotably, our approach consistently surpasses or matches SOTA in all settings, maintaining stable prompt selection accuracy throughout multiple incremental learning sessions.", "keywords": [ "Continual Learning; Prompt Learning; Catastrophic Forgetting" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 4, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "QY52D9BeJo", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Jason D. Lee", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Yunwei Ren", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" } ], "n_formula": 275, "n_formula_1": 34, "n_ref_uni": 22, "n_ref": 61, "n_ref_all": 77, "n_fig": 1, "n_tab": 5, "L_tab": 284, "n_element_tab": 30, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 31754, "formula_len_all_1": 3411, "len_all": 305845, "len_all_1": 74660, "len_abs": 1288, "len_title": 132, "len_sents": 85189, "len_sents_1": 28578, "n_sents": 1087, "n_sents_1": 288, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1435, "title": "Learning Orthogonal Multi-Index Models: A Fine-Grained Information Exponent Analysis", "abs": "The information exponent (Ben Arous et al. (2021)) --- which is equivalent to the lowest degree in the Hermite\n expansion of the link function for Gaussian single-index models --- has played an important role in predicting the \n sample complexity of online stochastic gradient descent (SGD) in various learning tasks. In this work, we \n demonstrate that, for multi-index models, focusing solely on the lowest degree can miss key structural details \n and result in suboptimal rates. \n\n Specifically, we consider the task of learning target functions of form $f_*(\\mathbf{x}) = \\sum_{k=1}^{P} \\phi({\\mathbf{v_k}^*} \\cdot \\mathbf{x})$, \n where $P \\ll d$, the ground-truth directions $\\\\{ v_k^* \\\\}_{k=1}^P$ are orthonormal, and only the second and \n $2L$-th Hermite coefficients of the link function $\\phi$ can be nonzero. \n Based on the theory of information exponent, when the lowest degree is $2L$, recovering the directions requires \n $d^{2L-1}\\mathrm{poly}(P)$ samples, and when the lowest degree is $2$, only the relevant subspace (not the exact \n directions) can be recovered due to the rotational invariance of the second-order terms. In contrast, we show that \n by considering both second- and higher-order terms, we can first learn the relevant space via the second-order\n terms, and then the exact directions using the higher-order terms, and the overall sample complexity of online \n SGD is $d \\mathrm{poly}(P)$.", "keywords": [ "multi-index model", "information exponent", "sample complexity", "stochastic gradient descent" ], "rating_list": [ 5, 8, 5 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 4, 2 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "QXwtkVI8Yr", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Jia Liu", "gender": "Male", "institution": "The Ohio State University", "country": "US", "position": "Associate Professor" }, { "name": "Ness Shroff", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Full Professor" }, { "name": "Peiwen Qiu", "gender": "Female", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Ziyue Luo", "gender": "Male", "institution": "The Ohio State University", "country": "US", "position": "Postdoc" } ], "n_formula": 75, "n_formula_1": 27, "n_ref_uni": 42, "n_ref": 83, "n_ref_all": 115, "n_fig": 8, "n_tab": 5, "L_tab": 313, "n_element_tab": 24, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 70, "n_element_tab_1": 8, "formula_len_all": 21354, "formula_len_all_1": 2161, "len_all": 189814, "len_all_1": 75547, "len_abs": 1838, "len_title": 133, "len_sents": 50085, "len_sents_1": 36892, "n_sents": 332, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1878, "title": "Swift-FedGNN: Federated Graph Learning with Low Communication and Sample Complexities", "abs": "Graph neural networks (GNNs) have achieved great success in a wide variety of graph-based learning applications.\nTo expedite training for large-scale graphs,\ndistributed GNN training has been proposed using sampling-based mini-batch training.\nHowever, such a traditional distributed GNN training approach is not applicable to emerging GNN learning applications with geo-distributed input graphs, which require the data to be kept within the site where it is generated to protect privacy.\nOn the other hand, federated learning (FL) has been widely used to enable privacy-preserving training under data parallelism.\nHowever, because of cross-client links in the aforementioned geo-distributed graph data, applying federated learning directly to GNNs incurs expensive cross-client neighbor sampling and communication costs due to the large graph size and the dependencies between nodes among different clients. \nTo overcome these challenges, we propose a new mini-batch and sampling-based federated GNN algorithmic framework called Swift-FedGNN that primarily performs efficient parallel local training and periodically conducts time-consuming cross-client training.\nSpecifically, in Swift-FedGNN, each client *primarily* trains a local GNN model using only its local graph data, and some randomly sampled clients *periodically* learn the local GNN models based on their local graph data and the dependent nodes across clients.\nWe theoretically establish the convergence performance of Swift-FedGNN and show that it enjoys a convergence rate of $\\mathcal{O}\\left( T^{-1/2} \\right)$, matching the state-of-the-art (SOTA) rate of sampling-based GNN methods, despite operating in the challenging FL setting.\nExtensive experiments on real-world datasets show that Swift-FedGNN significantly outperforms the SOTA federated GNN approaches with comparable accuracy in terms of efficiency.", "keywords": [ "federated learning", "graph neural network", "optimization" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "QXQiq8JVOB", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Alexandre Kaiser", "gender": "Male", "institution": "New York University", "country": "US", "position": "MS student" }, { "name": "Arthur Jacot", "gender": "Male", "institution": "NYU, New York University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 61, "n_formula_1": 33, "n_ref_uni": 27, "n_ref": 43, "n_ref_all": 50, "n_fig": 5, "n_tab": 3, "L_tab": 82, "n_element_tab": 7, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 57, "n_element_tab_1": 7, "formula_len_all": 6045, "formula_len_all_1": 2208, "len_all": 125710, "len_all_1": 64244, "len_abs": 1282, "len_title": 80, "len_sents": 39810, "len_sents_1": 27931, "n_sents": 293, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1341, "title": "Hamiltonian Mechanics of Feature Learning: Bottleneck Structure in Leaky ResNets", "abs": "We study Leaky ResNets, which interpolate between ResNets ($\\tilde{L}=0$)\nand Fully-Connected nets ($\\tilde{L}\\to\\infty$) depending on an 'effective\ndepth' hyper-parameter $\\tilde{L}$. In the infinite depth limit,\nwe study 'representation geodesics' $A_{p}$: continuous paths in\nrepresentation space (similar to NeuralODEs) from input $p=0$ to\noutput $p=1$ that minimize the parameter norm of the network. We\ngive a Lagrangian and Hamiltonian reformulation, which highlight the\nimportance of two terms: a kinetic energy which favors small layer\nderivatives $\\partial_{p}A_{p}$ and a potential energy that favors\nlow-dimensional representations, as measured by the 'Cost of Identity'.\nThe balance between these two forces offers an intuitive understanding\nof feature learning in ResNets. We leverage this intuition to explain\nthe emergence of a bottleneck structure, as observed in previous work:\nfor large $\\tilde{L}$ the potential energy dominates and leads to\na separation of timescales, where the representation jumps rapidly\nfrom the high dimensional inputs to a low-dimensional representation,\nmove slowly inside the space of low-dimensional representations, before\njumping back to the potentially high-dimensional outputs. Inspired\nby this phenomenon, we train with an adaptive layer step-size\nto adapt to the separation of timescales.", "keywords": [ "Theory of Deep Learning", "Feature Learning", "Hamiltonian" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "QWunLKbBGF", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Devamanyu Hazarika", "gender": "Male", "institution": "Amazon Alexa AI", "country": "US", "position": "Researcher" }, { "name": "Kaixiang Lin", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Mingyi Hong", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Siyan Zhao", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Yang Liu", "gender": "Female", "institution": "Amazon", "country": "", "position": "Principal Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 35, "n_ref": 84, "n_ref_all": 133, "n_fig": 19, "n_tab": 10, "L_tab": 13396, "n_element_tab": 1353, "n_fig_1": 11, "n_tab_1": 5, "L_tab_1": 2290, "n_element_tab_1": 108, "formula_len_all": 1, "formula_len_all_1": 47, "len_all": 260090, "len_all_1": 64217, "len_abs": 1730, "len_title": 137, "len_sents": 83265, "len_sents_1": 30561, "n_sents": 638, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 88, "L_abs": 1671, "title": "Do LLMs Recognize Your Preferences? Evaluating Personalized Preference Following in LLMs", "abs": "Large Language Models (LLMs) are increasingly deployed as chatbots, yet their ability to personalize responses to user preferences remains limited. We introduce PrefEval, a benchmark for evaluating LLMs' ability to infer, memorize and adhere to user preferences in long-context conversational setting.\nPrefEval comprises 3,000 manually curated user preference and query pairs spanning 20 topics. PrefEval contains user personalization or preference information in both explicit and implicit preference forms, and evaluates LLM performance using a generation and a classification task. With PrefEval, we have evaluated 10 open-sourced and\nproprietary LLMs in multi-session conversations with varying context lengths up to 100k tokens. We benchmark with various prompting, iterative feedback, and retrieval-augmented generation methods. \nOur benchmarking effort reveals that state-of-the-art LLMs face significant challenges in following users' preference during conversations. In particular, in zero-shot settings, preference following accuracy falls below 10\\% at merely 10 turns (~3k tokens) across most evaluated models. Even with advanced prompting and retrieval methods, preference following still deteriorates in long-context conversations. We also find that multiple stated preferences within a conversation improve adherence and models are not affected by conflicting preferences. Furthermore, we show that fine-tuning on PrefEval significantly improves performance. We believe PrefEval serves as a valuable resource for measuring, understanding, and enhancing LLMs' proactive preference following abilities, paving the way for personalized conversational agents.", "keywords": [ "personalization", "benchmark", "Large language models", "conversational llm", "chatbots" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "QWkcCFhkTL", "primary_area": "generative models", "n_author": 1, "author_info_list": [ { "name": "Mashrur M. Morshed", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 19, "n_ref_uni": 28, "n_ref": 54, "n_ref_all": 105, "n_fig": 16, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 193, "n_element_tab_1": 2, "formula_len_all": 1571, "formula_len_all_1": 1003, "len_all": 112365, "len_all_1": 56372, "len_abs": 925, "len_title": 108, "len_sents": 37366, "len_sents_1": 25599, "n_sents": 291, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 932, "title": "DiverseFlow: Sample-Efficient Diverse Mode Coverage in Flows", "abs": "Many real-world applications of flow generative models desire a diverse set of samples covering multiple modes of the target distribution. However, the predominant approach for obtaining diverse sets is not sample-efficient, as it involves independently obtaining many samples from the source distribution and mapping them through the flow until the desired mode coverage is achieved. As an alternative to repeated sampling, we introduce DiverseFlow---a training-free, inference-time approach to improve the diversity of flow models. Our key idea is to employ a determinantal point process to induce a coupling between the samples and drive sample diversity under a fixed sampling budget. We demonstrate the efficacy of DiverseFlow for tasks where sample efficient diversity is highly desirable---text-guided image generation with polysemous words, inverse problems like large-hole inpainting, and class-conditional image synthesis.", "keywords": [ "Diverse sampling", "flow matching", "determinantal point processes" ], "rating_list": [ 5, 5, 5, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "QWjpjisCjs", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Bin Chen", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Hao Fang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Intern" }, { "name": "Hongyao Yu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Ke Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Shu-Tao Xia", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Wenbo Yu", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xuan Wang", "gender": "Male", "institution": "Harbin Institute of Technology,Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Yixiang Qiu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 41, "n_ref": 122, "n_ref_all": 137, "n_fig": 2, "n_tab": 19, "L_tab": 8556, "n_element_tab": 421, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 4807, "n_element_tab_1": 130, "formula_len_all": 57, "formula_len_all_1": 57, "len_all": 174291, "len_all_1": 54993, "len_abs": 1741, "len_title": 121, "len_sents": 43677, "len_sents_1": 24764, "n_sents": 358, "n_sents_1": 147, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1768, "title": "MIBench: A Comprehensive Benchmark for Model Inversion Attack and Defense", "abs": "Model Inversion (MI) attacks aim at leveraging the output information of target models to reconstruct privacy-sensitive training data, raising widespread concerns on privacy threats of Deep Neural Networks (DNNs). Unfortunately, in tandem with the rapid evolution of MI attacks, the lack of a comprehensive, aligned, and reliable benchmark has emerged as a formidable challenge. This deficiency leads to inadequate comparisons between different attack methods and inconsistent experimental setups. In this paper, we introduce the first practical benchmark for model inversion attacks and defenses to address this critical gap, which is named \"\\textit{MIBench}\". This benchmark serves as an extensible and reproducible modular-based toolbox and currently integrates a total of 16 state-of-the-art attack and defense methods. Moreover, we furnish a suite of assessment tools encompassing 9 commonly used evaluation protocols to facilitate standardized and fair evaluation and analysis. Capitalizing on this foundation, we conduct extensive experiments from multiple perspectives to holistically compare and analyze the performance of various methods across different scenarios, which overcomes the misalignment issues and discrepancy prevalent in previous works. Based on the collected attack methods and defense strategies, we analyze the impact of target resolution, defense robustness, model predictive power, model architectures, transferability and loss function. Our hope is that this \\textit{MIBench} could provide a unified, practical and extensible toolbox and is widely utilized by researchers in the field to rigorously test and compare their novel methods, ensuring equitable evaluations and thereby propelling further advancements in the future development.", "keywords": [ "Benchmark", "Privacy Leakage", "Model Inversion" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "QWMgTMRUnB", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Shangjian Yin", "gender": "Male", "institution": "South China Agricultural University", "country": "CN", "position": "Undergrad student" }, { "name": "Wei-Lin Chen", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Xinyu Zhu", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Yu Meng", "gender": "Male", "institution": "University of Virginia", "country": "", "position": "Assistant Professor" }, { "name": "Zhepei Wei", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 60, "n_ref": 90, "n_ref_all": 105, "n_fig": 9, "n_tab": 2, "L_tab": 1581, "n_element_tab": 234, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1581, "n_element_tab_1": 234, "formula_len_all": 814, "formula_len_all_1": 814, "len_all": 170044, "len_all_1": 62459, "len_abs": 3716, "len_title": 95, "len_sents": 35117, "len_sents_1": 26615, "n_sents": 249, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 2021, "title": "Self-Alignment Optimization for Language Models", "abs": "Traditional reinforcement learning from human feedback (RLHF) relies heavily on costly and time-consuming human-annotated datasets. Even Reinforcement Learning from AI Feedback (RLAIF), which trains a reward model using AI-generated preference data before refining the language model through reinforcement learning, remains expensive. These methods often necessitate either specialized reward model designs or larger models (e.g., GPT-4) for external labeling. In this paper, we introduce a dataset-free and annotation-free framework called Self-Alignment Optimization (SAO), which addresses the aforementioned issue by aligning the model with its own prompts and feedback as preferences. SAO begins with a chat-based model that engages in persona role-play to generate diverse prompts and responses, which are then self-evaluated and used for preference optimization.\nExtensive experiments with two strong LLMs on several benchmarks demonstrate the effectiveness of SAO. Specifically, on AlpacaEval 2.0, Gemma-2-9B-it-SAO achieves a Length-Controlled Win Rate (LC) of 69.2\\% and win rate (WR) of 66.0\\%, surpassing the baseline model by 18.1\\% and 27.9\\%. Llama-3-Instruct-8B-SAO reaches 33.3\\% LC and 39.0\\% WR, with performance improvements of 10.4\\% and 16.4\\%, respectively. On the MT-Bench benchmark, Gemma-2-9B-it-SAO and Llama-3-8B-Instruct-SAO score 7.41 and 6.76, compared to their pre-SAO scores of 7.09 and 6.70. The Arena-Hard benchmark shows even greater gains from SAO, with Gemma-2-9B-it's WR increasing from 52.6\\% to 70.1\\% and Llama-3-Instruct-8B's WR rising from 40.3\\% to 56.4\\%. In addition, our further experiments demonstrate that models fine-tuned with SAO exhibit similar or even superior performance on downstream NLP tasks compared to baseline models, rather than those trained with external labeled datasets, which enhance alignment ability but may compromise some general capabilities. We anticipate that this work will provide new insights for future research on self-improvement in LLMs.", "keywords": [ "Self-Alignment", "Annotation-free", "Large Language Model" ], "rating_list": [ 3, 3, 6, 3, 3 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2, 2 ], "contribution_list": [ 2, 1, 2, 1, 1 ], "confidence_list": [ 4, 3, 2, 4, 2 ] }, { "paper_id": "QWMCaEfoR7", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Alex Hung", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Dan Zeng", "gender": "Female", "institution": "Shanghai University", "country": "CN", "position": "Full Professor" }, { "name": "Kai Zhao", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "Postdoc" }, { "name": "zuojiehe", "gender": "Male", "institution": "Shanghai University", "country": "CN", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 25, "n_ref": 73, "n_ref_all": 87, "n_fig": 5, "n_tab": 11, "L_tab": 13921, "n_element_tab": 482, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 546, "n_element_tab_1": 12, "formula_len_all": 162, "formula_len_all_1": 162, "len_all": 127201, "len_all_1": 38873, "len_abs": 1486, "len_title": 166, "len_sents": 29813, "len_sents_1": 17878, "n_sents": 208, "n_sents_1": 127, "venue": "ICLR2025", "decision": "Reject", "L_title": 118, "L_abs": 1364, "title": "Dominant Shuffle: An Incredibly Simple but Exceptionally Effective Data Augmentation Method for Time-Series Prediction", "abs": "Frequency-domain data augmentation (DA) has shown strong performance in time-series prediction due to its ability to preserve data-label consistency. However, we observed that existing frequency-domain augmentations introduce excessive variability, leading to out-of-distribution samples that may be harmful to model performance. To address this, we introduced two simple modifications to frequency-domain DA. First, we limit perturbations to dominant frequencies with larger magnitudes, which capture the main periodicities and trends of the signal. Second, instead of using complicated random perturbations, we simply shuffle the dominant frequency components, which preserves the original structure while avoiding external noise. With the two simple modifications, we proposed dominant shuffle\u2014a simple yet highly effective data augmentation technique for time-series prediction. Our method is remarkably simple, requiring only a few lines of code, yet exceptionally effective, consistently and significantly improving model performance. Extensive experiments on short-term, long term, few-shot and cold-start prediction tasks with eight state-of-the-art models, nine existing augmentation methods and twelve datasets demonstrate that dominant shuffle consistently boosts model performance with substantial gains, outperforming existing augmentation techniques.", "keywords": [ "time series prediction", "data augmentation", "deep learning" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "QWIg5e6mtT", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Mingzhi Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Naming Liu", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Weinan Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xihuai Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Yaodong Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ying Wen", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Youzhi Zhang", "gender": "unknown", "institution": "Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "", "position": "Assistant Professor" } ], "n_formula": 35, "n_formula_1": 15, "n_ref_uni": 33, "n_ref": 95, "n_ref_all": 119, "n_fig": 8, "n_tab": 4, "L_tab": 379, "n_element_tab": 3, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 278, "n_element_tab_1": 9, "formula_len_all": 4277, "formula_len_all_1": 1899, "len_all": 173080, "len_all_1": 72354, "len_abs": 1662, "len_title": 115, "len_sents": 61607, "len_sents_1": 35035, "n_sents": 390, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1745, "title": "Computing Ex Ante Equilibrium in Heterogeneous Zero-Sum Team Games", "abs": "The \\textit{ex ante} equilibrium for two-team zero-sum games, where agents within each team collaborate to compete against the opposing team, is known to be the best a team can do for coordination. Many existing works on \\textit{ex ante} equilibrium solutions are aiming to extend the scope of \\textit{ex ante} equilibrium solving to large-scale team games based on Policy Space Response Oracle (PSRO). However, the joint team policy space constructed by the most prominent method, Team PSRO, cannot cover the entire team policy space in heterogeneous team games where teammates play distinct roles. Such insufficient policy expressiveness causes Team PSRO to be trapped into a sub-optimal \\textit{ex ante} equilibrium with significantly higher exploitability and never converges to the global \\textit{ex ante} equilibrium. To find the global \\textit{ex ante} equilibrium without introducing additional computational complexity, we first parameterize heterogeneous policies for teammates, and we prove that optimizing the heterogeneous teammates' policies sequentially can guarantee a monotonic improvement in team rewards. \nWe further propose \\textbf{Heterogeneous-PSRO} (\\textbf{H-PSRO}), a novel framework for heterogeneous team games, which integrates the sequential correlation mechanism into the PSRO framework and serves as the first PSRO framework for heterogeneous team games. \nWe prove that H-PSRO achieves lower exploitability than Team PSRO in heterogeneous team games. \nEmpirically, H-PSRO achieves convergence in matrix heterogeneous games that are unsolvable by non-heterogeneous baselines. \nFurther experiments reveal that H-PSRO outperforms non-heterogeneous baselines in both heterogeneous team games and homogeneous settings.", "keywords": [ "Two-Team Zero-Sum Games", "Heterogeneous Teammates", "Policy Space Response Oracle" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 1, 1, 2 ], "confidence_list": [ 5, 5, 2 ] }, { "paper_id": "QWDZE2mYIe", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Aaditya Arunkumar Khant", "gender": "Male", "institution": "University of Texas at San Antonio", "country": "US", "position": "PhD student" }, { "name": "Murtuza Jadliwala", "gender": "Male", "institution": "University of Texas at San Antonio", "country": "US", "position": "Associate Professor" }, { "name": "Raveen Wijewickrama", "gender": "unknown", "institution": "University of Texas at San Antonio", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 40, "n_ref": 66, "n_ref_all": 86, "n_fig": 7, "n_tab": 6, "L_tab": 1913, "n_element_tab": 264, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1847, "n_element_tab_1": 143, "formula_len_all": 61, "formula_len_all_1": 61, "len_all": 134137, "len_all_1": 64667, "len_abs": 1549, "len_title": 60, "len_sents": 42236, "len_sents_1": 31387, "n_sents": 283, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1556, "title": "SAFE: Spiking Neural Network-based Audio Fidelity Evaluation", "abs": "Recent advancements in generative AI have enabled the creation of highly realistic synthetic audio, posing significant challenges in voice authentication, media verification, and fraud detection. While deep learning models are frequently used for fake audio detection, they often struggle to generalize to unseen and complex manipulations, particularly partial fake audio, where real and synthetic segments are seamlessly combined. This paper explores the use of Spiking Neural Networks (SNNs) for fake and partial fake audio detection, an area that has not yet been investigated. SNNs, known for their energy-efficient computation and ability to process temporal data, offer a promising alternative to traditional Artificial Neural Networks (ANNs). We propose an SNN-based approach for fake audio detection and comprehensively evaluate its performance through a series of experiments, including hyperparameter tuning, cross-dataset generalization and partial fake audio detection. \nOur results show that SNNs achieve accuracy comparable to state-of-the-art ANN models with fewer number of parameters. Although, SNNs did not offer significant improvements in generalization capabilities, they provided advantages such as reduced model sizes and computational efficiency, making them more suitable for resource-constrained and real-time voice authentication applications.\nThis study lays the groundwork for further exploration of SNNs in audio spoofing countermeasures, providing a foundation for future advancements in security-critical voice applications.", "keywords": [ "Fake Audio Detection", "Spiking Neural Networks (SNNs)", "Partial Fake Audio" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "QWDFOOoV3U", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Bernard Ghanem", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Full Professor" }, { "name": "Bryan Russell", "gender": "unknown", "institution": "Adobe Research", "country": "US", "position": "Research Scientist" }, { "name": "Josef Sivic", "gender": "Male", "institution": "Czech Technical University in Prague", "country": "CZ", "position": "Principal investigator" }, { "name": "Mattia Soldan", "gender": "Male", "institution": "KAUST", "country": "", "position": "PhD student" }, { "name": "Fabian Caba heilbron", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 71, "n_ref": 174, "n_ref_all": 224, "n_fig": 24, "n_tab": 7, "L_tab": 1303, "n_element_tab": 87, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 417, "formula_len_all_1": 342, "len_all": 291038, "len_all_1": 78050, "len_abs": 99, "len_title": 125, "len_sents": 88723, "len_sents_1": 35495, "n_sents": 648, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1068, "title": "ResidualViT for Efficient Zero-Shot Natural Language Temporal Video Grounding", "abs": "The goal of this work is to efficiently compute frame-level features from videos for the Zero-Shot Natural Language Temporal Video Grounding (NLTVG) task. The contributions of this work are three-fold. First, we introduce a novel vision transformer (ViT) architecture, dubbed ResidualViT, that capitalizes on the large temporal redundancies in video. Our architecture incorporates (i) learnable residual connections that ensure temporal consistency across consecutive frames and (ii) a token reduction module for enhancing processing speed by selectively discarding temporally redundant information. Second, we describe a lightweight distillation strategy that enables learning parameters of ResidualViT from existing frame encoders without additional manual annotation. Finally, we validate the effectiveness of our approach across three diverse datasets, demonstrating significant reductions in computational cost (up to 60%) and improvements in inference speed (up to 2.5x faster), all while observing marginal accuracy reduction with respect to the teacher model.", "keywords": [ "Vision Transformers", "Efficiency", "Video Encoding", "Natural Language Video Grounding" ], "rating_list": [ 6, 6, 8, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "QVj3kUvdvl", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Asaf Shul", "gender": "unknown", "institution": "Google", "country": "IL", "position": "Intern" }, { "name": "Eliahu Horwitz", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "PhD student" }, { "name": "Yedid Hoshen", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "Associate Professor" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 28, "n_ref": 40, "n_ref_all": 64, "n_fig": 11, "n_tab": 6, "L_tab": 1320, "n_element_tab": 188, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 534, "formula_len_all_1": 526, "len_all": 139713, "len_all_1": 67533, "len_abs": 1562, "len_title": 89, "len_sents": 45782, "len_sents_1": 31900, "n_sents": 412, "n_sents_1": 289, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 1569, "title": "Unsupervised Model Tree Heritage Recovery", "abs": "The number of models shared online has recently skyrocketed, with over one million public models available on Hugging Face. Sharing models allows other users to build on existing models, using them as initialization for fine-tuning, improving accuracy and saving compute and energy. However, it also raises important intellectual property issues, as fine-tuning may violate the license terms of the original model or that of its training data. A Model Tree, i.e., a tree data structure rooted at a foundation model and having directed edges between a parent model and other models directly fine-tuned from it (children), would settle such disputes by making the model heritage explicit. Unfortunately, current models are not well documented, with most model metadata (e.g., \"model cards\") not providing accurate information about heritage. In this paper, we introduce the task of Unsupervised Model Tree Heritage Recovery (Unsupervised MoTHer Recovery) for collections of neural networks. For each pair of models, this task requires: i) determining if they are directly related, and ii) establishing the direction of the relationship. Our hypothesis is that model weights encode this information, the challenge is to decode the underlying tree structure given the weights. We discover several properties of model weights that allow us to perform this task. By using these properties, we formulate the MoTHer Recovery task as finding a directed minimal spanning tree. In extensive experiments we demonstrate that our method successfully reconstructs complex Model Trees.", "keywords": [ "Weight Space Learning", "Deep Weight Space", "Model Tree" ], "rating_list": [ 6, 6, 8, 8, 5 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 2, 3, 3 ] }, { "paper_id": "QVBeBPsmy0", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Arman Zarei", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Keivan Rezaei", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Mazda Moayeri", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Mehrdad Saberi", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Priyatham Kattakinda", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Samyadeep Basu", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Soheil Feizi", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 24, "n_ref": 48, "n_ref_all": 71, "n_fig": 12, "n_tab": 3, "L_tab": 391, "n_element_tab": 47, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 379, "n_element_tab_1": 49, "formula_len_all": 1094, "formula_len_all_1": 1038, "len_all": 108472, "len_all_1": 49707, "len_abs": 1523, "len_title": 143, "len_sents": 34148, "len_sents_1": 23164, "n_sents": 244, "n_sents_1": 158, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1529, "title": "Mitigating Compositional Issues in Text-to-Image Generative Models via Enhanced Text Embeddings", "abs": "Text-to-image diffusion-based generative models have the stunning ability to generate photo-realistic images and achieve state-of-the-art low FID scores on challenging image generation benchmarks. However, one of the primary failure modes of these text-to-image generative models is in composing attributes, objects, and their associated relationships accurately into an image. In our paper, we investigate this compositionality-based failure mode and highlight that imperfect text conditioning with CLIP text-encoder is one of the primary reasons behind the inability of these models to generate high-fidelity compositional scenes. In particular, we show that (i) there exists an optimal text-embedding space that can generate highly coherent compositional scenes showing that the output space of the CLIP text-encoder is sub-optimal, and (ii) the final token embeddings in CLIP are erroneous as they often include attention contributions from unrelated tokens in compositional prompts. Our main finding shows that the best compositional improvements can be achieved (without harming the model's FID score) by fine-tuning only a simple and parameter-efficient linear projection on CLIP's representation space in Stable-Diffusion variants using a small set of compositional image-text pairs. This result demonstrates that the sub-optimality of the CLIP's output space is a major error source. We also show that re-weighting the erroneous attention contributions in CLIP can lead to slightly improved compositional performances.", "keywords": [ "Diffusion Models", "Stable Diffusion", "Compositionality", "Generative Models", "Explainability" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "QSuOHV62IQ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Hongtao Zhang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Undergrad student" }, { "name": "Jianing Li", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Postdoc" }, { "name": "Wei Chen", "gender": "Female", "institution": " Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xueqi Cheng", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy", "country": "CN", "position": "Full Professor" } ], "n_formula": 59, "n_formula_1": 22, "n_ref_uni": 47, "n_ref": 76, "n_ref_all": 106, "n_fig": 20, "n_tab": 1, "L_tab": 158, "n_element_tab": 24, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 158, "n_element_tab_1": 24, "formula_len_all": 5706, "formula_len_all_1": 2183, "len_all": 185148, "len_all_1": 70770, "len_abs": 848, "len_title": 157, "len_sents": 56452, "len_sents_1": 31856, "n_sents": 484, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 851, "title": "Latent Variable Identifiability in Nonlinear Causal Models with Single-domain Data under Minimality Condition", "abs": "The identifiability of latent variables given observational data is one of the core issues in the field of disentangled representation learning. Recent progresses have been made on establishing identifiablity theories for latent causal models. However with much restrictions or unrealistic assumptions, their practicality on real applications are limited. In this paper, we propose a novel identifiablity theory for learning latent variables in nonlinear causal models, requiring only single-domain data. We prove that all latent variables in a powerset bipartite graph can be identified up to an invertible transformation, if the generation process of observable data is globally invertible, latent variables are independent, and shared latent variables entail minimal information. Experiments on synthetic data support the conclusions of our theory.", "keywords": [ "Identifiability theory", "Disentangled representation learning", "Structural causal model" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "QSTv4os59f", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Pan Peng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Shan Jiang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Shi Li", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Professor" }, { "name": "Yinhao Dong", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 42, "n_formula_1": 7, "n_ref_uni": 51, "n_ref": 130, "n_ref_all": 148, "n_fig": 6, "n_tab": 4, "L_tab": 553, "n_element_tab": 62, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 47, "n_element_tab_1": 2, "formula_len_all": 6240, "formula_len_all_1": 994, "len_all": 241042, "len_all_1": 83085, "len_abs": 1229, "len_title": 114, "len_sents": 72792, "len_sents_1": 36943, "n_sents": 732, "n_sents_1": 327, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1276, "title": "Learning-Augmented Streaming Algorithms for Correlation Clustering", "abs": "We study streaming algorithms for Correlation Clustering. Given a complete graph as an arbitrary-order stream of edges, with each edge labelled as positive or negative, the goal is to partition the vertices into disjoint clusters, such that the number of disagreements is minimized. In this paper, we introduce the first learning-augmented streaming algorithms for the problem, achieving the first better-than-$3$-approximation in dynamic streams. Our algorithms draw inspiration from recent works of Cambus et al. (SODA'24), and Chakrabarty and Makarychev (NeurIPS'23). Our algorithms use the predictions of pairwise dissimilarities between vertices provided by a predictor and achieve an approximation ratio that is close to $2.06$ under good prediction quality. Even if the prediction quality is poor, our algorithms cannot perform worse than the well known Pivot algorithm, which achieves a $3$-approximation. Our algorithms are much simpler than the recent $1.847$-approximation streaming algorithm by Cohen-Addad et al. (STOC'24) which appears to be challenging to implement and is restricted to insertion-only streams. Experimental results on synthetic and real-world datasets demonstrate the superiority of our proposed algorithms over their non-learning counterparts.", "keywords": [ "Correlation Clustering", "graph streaming algorithms", "learning-augmented algorithms" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 4, 3 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3, 3 ] }, { "paper_id": "QRmpkVsvqS", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Guoying Zhao", "gender": "unknown", "institution": "Aalto University", "country": "FI", "position": "Visiting Professor" }, { "name": "Hao Yu", "gender": "Male", "institution": "University of Oulu", "country": "FI", "position": "PhD student" }, { "name": "Haoyu Chen", "gender": "Male", "institution": "University of Oulu", "country": "FI", "position": "Assistant Professor" }, { "name": "Pichao WANG", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 5, "n_ref_uni": 37, "n_ref": 103, "n_ref_all": 127, "n_fig": 10, "n_tab": 9, "L_tab": 7696, "n_element_tab": 1185, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3374, "n_element_tab_1": 539, "formula_len_all": 1139, "formula_len_all_1": 257, "len_all": 169571, "len_all_1": 63635, "len_abs": 1797, "len_title": 123, "len_sents": 44986, "len_sents_1": 23453, "n_sents": 323, "n_sents_1": 151, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1738, "title": "SCFormer: Spatial Coordination for Efficient and Robust Vision Transformers", "abs": "We investigate the design of visual backbones with a focus on optimizing both efficiency and robustness. While recent advancements in hybrid Vision Transformers (ViTs) have significantly enhanced efficiency, achieving state-of-the-art performance with fewer parameters, their robustness against domain-shifted and corrupted inputs remains a critical challenge. This trade-off is particularly difficult to balance in lightweight models, where robustness often relies on wider channels to capture diverse spatial features. In this paper, we present SCFormer, a novel hybrid ViT architecture designed to address these limitations. SCFormer introduces Spatial Coordination Attention (SCA), a mechanism that coordinates cross-spatial pixel interactions by deconstructing and reassembling spatial conditions with diverse connectivity patterns. This approach broadens the representation boundary, allowing SCFormer to efficiently capture more diverse spatial dependencies even with fewer channels, thereby improving robustness without sacrificing efficiency. Additionally, we incorporate an Inceptional Local Representation (ILR) block to flexibly enrich local token representations before self-attention, enhancing both locality and feature diversity. Through extensive experiments, SCFormer demonstrates superior performance across multiple benchmarks. On ImageNet-1K, SCFormer-XS achieves 2.5\\% higher top-1 accuracy and 10\\% faster GPU inference speed compared to FastViT-T8. On ImageNet-A, SCFormer-L (30.1M) surpasses RVT-B (91.8M) in robustness accuracy by 5.6\\% while using 3$\\times$ fewer parameters. These results underscore the effectiveness of our design in achieving a new state-of-the-art balance between efficiency and robustness.", "keywords": [ "Vision backbone", "Transformer", "Efficiency", "Robustness", "Spatial coordinating Attention." ], "rating_list": [ 6, 5, 3, 5, 3 ], "soundness_list": [ 4, 3, 2, 2, 1 ], "presentation_list": [ 4, 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 2, 1 ], "confidence_list": [ 4, 3, 4, 5, 5 ] }, { "paper_id": "QRERAL4r2k", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Eslam Mohamed BAKR", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Faizan Farooq Khan", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Habib Slim", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Jian Ding", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" }, { "name": "Mohamed Elhoseiny", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Mohamad Aymen Mir", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 89, "n_ref": 152, "n_ref_all": 160, "n_fig": 3, "n_tab": 2, "L_tab": 994, "n_element_tab": 176, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 764, "n_element_tab_1": 141, "formula_len_all": 537, "formula_len_all_1": 533, "len_all": 178670, "len_all_1": 65611, "len_abs": 1094, "len_title": 126, "len_sents": 31174, "len_sents_1": 29779, "n_sents": 232, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 125, "L_abs": 1099, "title": "HuMouS: Human Motion Synthesis with Fine-Grained Control using Latent Space Manipulation of Cycle-Consistent Diffusion Models", "abs": "We address the problem of spatially guided text-to-motion synthesis. While there has been work to incorporate spatial constraints in text-to-motion diffusion models, existing methods still face significant challenges in generating motions that align with the conditional controls. To this end, we propose Cycle Consistent Diffusion, a novel approach that improves controllable generation by explicitly optimizing frame-level cycle consistency between generated motions and conditional controls. Specifically, for an input conditional control, we ensure that the output motion and the input spatial constraint are forced to be consistent. A straightforward implementation though consistent with the input often does not match fine-grained control signals. To this end, we introduce a novel test-time optimization framework that directs our pre-trained cycle consistent diffusion model towards user-defined sparse constraints. We demonstrate approximately 5 to 10 percent improvement in controllability of motion synthesis on the HumanML3D dataset, while significantly reducing foot skating artifacts.", "keywords": [ "3d humans", "human motion synthesis" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "QQvhOyIldg", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Baocheng Sun", "gender": "unknown", "institution": "Weizmann Institute of Science", "country": "", "position": "MS student" }, { "name": "Junze Yin", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Yeqi Gao", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "MS student" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" } ], "n_formula": 79, "n_formula_1": 29, "n_ref_uni": 24, "n_ref": 63, "n_ref_all": 78, "n_fig": 1, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8112, "formula_len_all_1": 1849, "len_all": 134294, "len_all_1": 62909, "len_abs": 844, "len_title": 94, "len_sents": 34642, "len_sents_1": 22445, "n_sents": 417, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1056, "title": "An $O(k\\log n)$ Time Fourier Set Query Algorithm", "abs": "Fourier transformation is an extensively studied problem in many research fields. It has many applications in machine learning, signal processing, compressed sensing, and so on. In many real-world applications, approximated Fourier transformation is sufficient and we only need to do the Fourier transform on a subset of coordinates. \nGiven a vector $x \\in \\mathbb{C}^{n}$, approximation parameters $\\epsilon, \\delta \\in (0, 0.1)$, and a query set $S \\subset [n]$ of size $k$, we propose an algorithm to compute an approximate Fourier transform result $x'$ which uses $O(\\epsilon^{-1} k \\log(n/\\delta))$ Fourier measurements and runs in $O(\\epsilon^{-1} k \\log(n/\\delta))$ time. For $\\hat{x}$ being the Fourier transformation result, our algorithm can output a vector $x'$ such that $\\\\| ( x' - \\widehat{x} )\\_S \\\\|\\_2^2 \\leq \\epsilon \\\\| \\widehat{x}\\_{\\overline{S}} \\\\|\\_2^2 + \\delta \\\\| \\widehat{x} \\\\|\\_1^2 $ holds with probability of at least $9/10$, where $\\overline{S}$ denotes the complement of the set $S$, i.e., $\\overline{S} := [n] \\setminus S$.", "keywords": [ "Fourier transform", "set query" ], "rating_list": [ 8, 3, 1, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 1, 1 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "QQoWeCscSH", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Xuming Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Yonghua Hei", "gender": "Female", "institution": "Wuhan University of Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 38, "n_ref": 72, "n_ref_all": 100, "n_fig": 4, "n_tab": 21, "L_tab": 17955, "n_element_tab": 1998, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2183, "n_element_tab_1": 319, "formula_len_all": 247, "formula_len_all_1": 142, "len_all": 280439, "len_all_1": 60698, "len_abs": 1732, "len_title": 112, "len_sents": 40926, "len_sents_1": 26183, "n_sents": 299, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1754, "title": "Unlocking Speech Instruction Data Potential with Query Rewriting", "abs": "End-to-end Large Speech Language Models (**LSLMs**) demonstrate strong potential in response latency and speech comprehension capabilities, showcasing general intelligence across speech understanding tasks. However, the ability to follow speech instructions has not been fully realized due to the lack of datasets and heavily biased training tasks. Leveraging the rich ASR datasets, previous approaches have used Large Language Models (**LLMs**) to continue the linguistic information of speech to construct speech instruction datasets. Yet, due to the gap between LLM-generated results and real human responses, the continuation methods further amplify these shortcomings. Given the high costs of collecting and annotating speech instruction datasets by humans, using speech synthesis to construct large-scale speech instruction datasets has become a balanced and robust alternative. Although modern Text-To-Speech (**TTS**) models have achieved near-human-level synthesis quality, it is challenging to appropriately convert out-of-distribution text instruction to speech due to the limitations of the training data distribution in TTS models. To address this issue, we propose a query rewriting framework with multi-LLM knowledge fusion, employing multiple agents to annotate and validate the synthesized speech, making it possible to construct high-quality speech instruction datasets without relying on human annotation. Experiments show that this method can transform text instructions into distributions more suitable for TTS models for speech synthesis through zero-shot rewriting, increasing data usability from 72\\% to 93\\%. It also demonstrates unique advantages in rewriting tasks that require complex knowledge and context-related abilities.", "keywords": [ "Multimodal language model; Large Speech Language Model; Datasets" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "QQCIfkhGIq", "primary_area": "generative models", "n_author": 12, "author_info_list": [ { "name": "Alexander Filippov", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "", "position": "Lab Director" }, { "name": "Alexey Artemov", "gender": "Male", "institution": "Apple", "country": "DE", "position": "Researcher" }, { "name": "Anton Konushin", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "Daniil Selikhanovych", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Dmitry Senushkin", "gender": "Male", "institution": "AIRI Institute", "country": "RU", "position": "Researcher" }, { "name": "Evgeny Burnaev", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Full Professor" }, { "name": "Ilya Olkov", "gender": "unknown", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Nikolay Patakin", "gender": "Male", "institution": "AIRI", "country": "RU", "position": "Researcher" }, { "name": "Nina Konovalova", "gender": "Female", "institution": "AIRI - Artificial Intelligence Research Institute", "country": "RU", "position": "Researcher" }, { "name": "Oleg Voynov", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Researcher" }, { "name": "Peter Wonka", "gender": "Male", "institution": "Snap Inc.", "country": "SA", "position": "Researcher" }, { "name": "Savva Victorovich Ignatyev", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 48, "n_ref": 104, "n_ref_all": 153, "n_fig": 13, "n_tab": 9, "L_tab": 2256, "n_element_tab": 323, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 656, "n_element_tab_1": 97, "formula_len_all": 396, "formula_len_all_1": 273, "len_all": 200385, "len_all_1": 64206, "len_abs": 417, "len_title": 94, "len_sents": 59187, "len_sents_1": 31247, "n_sents": 418, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 45, "L_abs": 1297, "title": "A3D: Does Diffusion Dream about 3D Alignment?", "abs": "We tackle the problem of text-driven 3D generation from a geometry alignment perspective.\nGiven a set of text prompts, we aim to generate a collection of objects with semantically corresponding parts aligned across them.\nRecent methods based on Score Distillation have succeeded in distilling the knowledge from 2D diffusion models to high-quality representations of the 3D objects.\nThese methods handle multiple text queries separately, and therefore the resulting objects have a high variability in object pose and structure.\nHowever, in some applications, such as 3D asset design, it may be desirable to obtain a set of objects aligned with each other.\nIn order to achieve the alignment of the corresponding parts of the generated objects, we propose to embed these objects into a common latent space and optimize the continuous transitions between these objects.\nWe enforce two kinds of properties of these transitions: smoothness of the transition and plausibility of the intermediate objects along the transition.\nWe demonstrate that both of these properties are essential for good alignment.\nWe provide several practical scenarios that benefit from alignment between the objects, including 3D editing and object hybridization, and experimentally demonstrate the effectiveness of our method.", "keywords": [ "Aligned 3D generation", "3D generation and editing", "Text-to-image diffusion", "Score distillation sampling", "Implicit neural surfaces", "Radiance fields" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "QQBPWtvtcn", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Fujun Luan", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Haian Jin", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Intern" }, { "name": "Hanwen Jiang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Hao Tan", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" }, { "name": "Kai Zhang", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Noah Snavely", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Professor" }, { "name": "Sai Bi", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Zexiang Xu", "gender": "Male", "institution": "Hillbot", "country": "US", "position": "Researcher" }, { "name": "Zhang Tianyuan", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 74, "n_ref": 167, "n_ref_all": 190, "n_fig": 9, "n_tab": 3, "L_tab": 845, "n_element_tab": 100, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2847, "n_element_tab_1": 263, "formula_len_all": 448, "formula_len_all_1": 190, "len_all": 197930, "len_all_1": 67955, "len_abs": 1349, "len_title": 113, "len_sents": 46102, "len_sents_1": 29791, "n_sents": 358, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 65, "L_abs": 1350, "title": "LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias", "abs": "We propose the Large View Synthesis Model (LVSM), a novel transformer-based approach for scalable and generalizable novel view synthesis from sparse-view inputs. We introduce two architectures: (1) an encoder-decoder LVSM, which encodes input image tokens into a fixed number of 1D latent tokens, functioning as a fully learned scene representation, and decodes novel-view images from them; and (2) a decoder-only LVSM, which directly maps input images to novel-view outputs, completely eliminating intermediate scene representations. Both models bypass the 3D inductive biases used in previous methods---from 3D representations (e.g., NeRF, 3DGS) to network designs (e.g., epipolar projections, plane sweeps)---addressing novel view synthesis with a fully data-driven approach. While the encoder-decoder model offers faster inference due to its independent latent representation, the decoder-only LVSM achieves superior quality, scalability, and zero-shot generalization, outperforming previous state-of-the-art methods by 1.5 to 3.5 dB PSNR. Comprehensive evaluations across multiple datasets demonstrate that both LVSM variants achieve state-of-the-art novel view synthesis quality, delivering superior performance even with reduced computational resources (1-2 GPUs). Please see our anonymous website for more details: https://lvsm-web.github.io/", "keywords": [ "novel view synthesis", "transformer", "large model" ], "rating_list": [ 8, 8, 8, 6, 8 ], "soundness_list": [ 4, 4, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 4 ], "confidence_list": [ 4, 4, 4, 5, 5 ] }, { "paper_id": "QPtoBPn4lZ", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Dennis Fok", "gender": "Male", "institution": "Erasmus University Rotterdam", "country": "NL", "position": "Full Professor" }, { "name": "Kathrin Gruber", "gender": "Female", "institution": "Erasmus University Rotterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Markus Mueller", "gender": "Male", "institution": "Erasmus University Rotterdam", "country": "", "position": "PhD student" } ], "n_formula": 51, "n_formula_1": 21, "n_ref_uni": 40, "n_ref": 119, "n_ref_all": 153, "n_fig": 4, "n_tab": 29, "L_tab": 14789, "n_element_tab": 488, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1889, "n_element_tab_1": 47, "formula_len_all": 3989, "formula_len_all_1": 1215, "len_all": 267634, "len_all_1": 76556, "len_abs": 1081, "len_title": 96, "len_sents": 79889, "len_sents_1": 34541, "n_sents": 734, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 1088, "title": "Continuous Diffusion for Mixed-Type Tabular Data", "abs": "Score-based generative models (or diffusion models for short) have proven successful for generating text and image data.\nHowever, the adaption of this model family to tabular data of mixed-type has fallen short so far. \nIn this paper, we propose CDTD, a Continuous Diffusion model for mixed-type Tabular Data. Specifically, we combine score matching and score interpolation to ensure a common continuous noise distribution for both continuous and categorical features alike. \nWe counteract the high heterogeneity inherent to data of mixed-type with distinct, adaptive noise schedules per feature or per data type.\nThe learnable noise schedules ensure optimally allocated model capacity and balanced generative capability.\nWe homogenize the data types further with model-specific loss calibration and initialization schemes tailored to mixed-type tabular data.\nOur experimental results show that CDTD consistently outperforms state-of-the-art benchmark models, captures feature correlations exceptionally well, and that heterogeneity in the noise schedule design boosts the sample quality.", "keywords": [ "synthetic data generation", "diffusion model", "generative model", "tabular data", "mixed-type data" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "QPZy2XMgzn", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Chao Hao", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "MS student" }, { "name": "Wenzhe Niu", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Intern" }, { "name": "Xin Liu", "gender": "Male", "institution": "Lappeenranta University of Technology", "country": "FI", "position": "Associate Professor" }, { "name": "Xu Ruiwen", "gender": "Male", "institution": "xiaohongshu", "country": "", "position": "Researcher" }, { "name": "Yanhua Huang", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Researcher" }, { "name": "Zezheng Wang", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Researcher" }, { "name": "Zitong YU", "gender": "unknown", "institution": "Great Bay University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 4, "n_ref_uni": 49, "n_ref": 114, "n_ref_all": 136, "n_fig": 2, "n_tab": 8, "L_tab": 2194, "n_element_tab": 287, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 2177, "n_element_tab_1": 280, "formula_len_all": 548, "formula_len_all_1": 200, "len_all": 197866, "len_all_1": 74053, "len_abs": 1599, "len_title": 154, "len_sents": 61903, "len_sents_1": 31614, "n_sents": 535, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1607, "title": "Token-by-Token Election: Improving Language Model Reasoning through Token-Level Multi-model Collaboration", "abs": "With the continuous development of large language models (LLMs), they have demonstrated amazing capabilities in many areas of natural language processing (NLP). However, due to their inherent limitations, the performance of a single model on many complex reasoning tasks has reached a bottleneck. A feasible solution is to introduce external feedback to further improve model performance, among which multi-model collaboration is a particularly promising approach. In this paper, we propose token-by-token election (TTE), a novel token-level multi-model collaboration strategy. Different from the common multi-model collaboration methods that operates at the overall answer level, TTE performs multi-model elections at the lowest token level. It selects the optimal token from the next token distributions given by multiple LLMs and then generates the answer autoregressively, allowing multiple LLMs to reach a consensus on each token. Inspired by human behavior, TTE consists of three election modes, including Cooperation, Competition, and Counting, all of which aim to sample the optimal token from multiple distributions. By strictly controlling the generation quality of each token, TTE can improve the quality of the overall answer and break through the performance bottleneck of a single LLM. Through extensive experiments on a variety of different types of reasoning benchmarks, we demonstrate the powerful performance of TTE, which further improves the performance compared to the current state-of-the-art single LLM and other multi-model collaborative methods. The code will be released on GitHub.", "keywords": [ "LLM", "multi-model collaboration" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "QPVK1ne9gI", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Abhisek Upadhyaya", "gender": "Male", "institution": "New York University", "country": "US", "position": "MS student" }, { "name": "Adarsh Krishnamurthy", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "Associate Professor" }, { "name": "Aditya Balu", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "Researcher" }, { "name": "Ali Rabeh", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "Postdoc" }, { "name": "Baskar Ganapathysubramanian", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "Professor" }, { "name": "Cheng-Hau Yang", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "PhD student" }, { "name": "Chinmay Hegde", "gender": "Male", "institution": "New York University", "country": "US", "position": "Associate Professor" }, { "name": "Ethan Herron", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "PhD student" }, { "name": "Mehdi Shadkhah", "gender": "Male", "institution": "Iowa State University", "country": "", "position": "PhD student" }, { "name": "Ronak Tali", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 1, "n_ref_uni": 54, "n_ref": 85, "n_ref_all": 109, "n_fig": 10, "n_tab": 7, "L_tab": 2869, "n_element_tab": 276, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 3742, "n_element_tab_1": 380, "formula_len_all": 1578, "formula_len_all_1": 97, "len_all": 167944, "len_all_1": 56174, "len_abs": 852, "len_title": 94, "len_sents": 44967, "len_sents_1": 21881, "n_sents": 324, "n_sents_1": 152, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 854, "title": "MPFBench: A Large Scale Dataset for SciML of Multi-Phase-Flows: Droplet and Bubble Dynamics", "abs": "Multiphase fluid dynamics, such as falling droplets and rising bubbles, are critical to many industrial applications. However, simulating these phenomena efficiently is challenging due to the complexity of instabilities, wave patterns, and bubble breakup. This paper investigates the potential of scientific machine learning (SciML) to model these dynamics using neural operators and foundation models. We apply sequence-to-sequence techniques on a comprehensive dataset generated from 11,000 simulations, comprising 1 million time snapshots, produced with a well-validated Lattice Boltzmann method (LBM) framework. The results demonstrate the ability of machine learning models to capture transient dynamics and intricate fluid interactions, paving the way for more accurate and computationally efficient SciML-based solvers for multiphase applications.", "keywords": [ "Scientific Machine Learning (SciML)", "Multiphase Flow", "Complex Physics Simulation", "Lattice Boltzmann Method (LBM)", "Droplet Dynamics", "Bubble Dynamics" ], "rating_list": [ 3, 8, 3, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "QPDbIFumQ8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Farshad Khorrami", "gender": "Male", "institution": "New York University", "country": "US", "position": "Full Professor" }, { "name": "Prashanth Krishnamurthy", "gender": "unknown", "institution": "New York University", "country": "US", "position": "Research Scientist and Adjunct Faculty" }, { "name": "Sara Ghazanfari", "gender": "Female", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Siddharth Garg", "gender": "Male", "institution": "New York University", "country": "US", "position": "Associate Professor" }, { "name": "Alexandre ARAUJO", "gender": "Male", "institution": "New York University", "country": "", "position": "Postdoc" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 29, "n_ref": 77, "n_ref_all": 97, "n_fig": 9, "n_tab": 4, "L_tab": 450, "n_element_tab": 91, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 38, "formula_len_all_1": 38, "len_all": 138402, "len_all_1": 48596, "len_abs": 1484, "len_title": 52, "len_sents": 32914, "len_sents_1": 25133, "n_sents": 201, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1490, "title": "EMMA: Efficient Visual Alignment in Multi-Modal LLMs", "abs": "Multi-modal Large Language Models (MLLMs) have recently exhibited impressive general-purpose capabilities by leveraging vision foundation models to encode the core concepts of images into representations. These are then combined with instructions and processed by the language model to generate high-quality responses. Despite significant progress\nin enhancing the language component, challenges persist in optimally fusing visual encodings within the language model for task-specific adaptability. Recent research has focused on improving this fusion through modality adaptation modules but at the cost of significantly increased model complexity and training data needs. \nIn this paper, we propose EMMA (Efficient Multi-Modal Adaptation), a lightweight cross-modality module designed to efficiently fuse visual and textual encodings, generating instruction-aware visual representations for the language model. Our key contributions include: (1) an efficient early fusion mechanism that integrates vision and language representations with minimal added parameters (less than 0.2% increase in model size), (2) an in-depth interpretability analysis that sheds light on the internal mechanisms of the proposed method; (3) comprehensive experiments that demonstrate notable improvements on both specialized and general benchmarks for MLLMs. Empirical results show that EMMA boosts performance across multiple tasks by up to 9.3% while significantly improving robustness against hallucinations.", "keywords": [ "Multi-Modal LLMs", "Modality Adaptation", "Visual Alignment" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 3, 5 ] }, { "paper_id": "QP3EvD1AVa", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Guy Yariv", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "PhD student" }, { "name": "Idan Schwartz", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "Assistant Professor" }, { "name": "Sagie Benaim", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "Assistant Professor" }, { "name": "Yossi Adi", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 56, "n_ref": 133, "n_ref_all": 156, "n_fig": 6, "n_tab": 14, "L_tab": 2699, "n_element_tab": 408, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 688, "n_element_tab_1": 121, "formula_len_all": 469, "formula_len_all_1": 365, "len_all": 195243, "len_all_1": 70681, "len_abs": 1552, "len_title": 125, "len_sents": 49073, "len_sents_1": 33065, "n_sents": 368, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1562, "title": "Improving Visual Commonsense in Language Models via Multiple Image Generation", "abs": "Commonsense reasoning is fundamentally based on multimodal knowledge. However, large language models (LLMs), trained using textual data only, are limited with their ability to incorporate essential visual information. In contrast, Visual Language Models (VLMs), which excel at visually-oriented tasks, often fail at non-visual tasks such as textual commonsense reasoning. \nThis divergence highlights a critical challenge - the integration of robust visual understanding with foundational text-based reasoning. To this end, we introduce a method aimed at enhancing LLMs' visual commonsense while maintaining textual modeling and commonsense reasoning performance. Specifically, our method generates multiple images based on the input text prompt and integrates these into the model's decision-making process by mixing their prediction probabilities. To facilitate multimodal grounded language modeling, we employ a late-fusion layer that combines the projected visual features with the output of a pre-trained LLM conditioned on text only. This late-fusion layer enables predictions based on comprehensive image-text knowledge as well as text only when required. We evaluate our approach using several visual commonsense reasoning tasks together with traditional NLP tasks, including common sense reasoning and reading comprehension. Our experimental results demonstrate significant superiority over existing baselines. When applied to recent state-of-the-art LLMs (e.g., Llama3), we observe improvements not only in visual commonsense but also in NLP benchmarks.", "keywords": [ "Language Models", "Image generation", "Machine learning" ], "rating_list": [ 5, 6, 8, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "QOfswj7hij", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 8, "author_info_list": [ { "name": "Adrian Weller", "gender": "Male", "institution": "Alan Turing Institute", "country": "", "position": "Programme Director for AI" }, { "name": "Hao Tang", "gender": "Male", "institution": "Cornell University", "country": "", "position": "PhD student" }, { "name": "Joao F. Henriques", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Principal Researcher" }, { "name": "Joshua B. Tenenbaum", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Professor" }, { "name": "Kevin Ellis", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Assistant Professor" }, { "name": "Nishanth Kumar", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Intern" }, { "name": "Tom Silver", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Assistant Professor" }, { "name": "Yichao Liang", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 5, "n_ref_uni": 32, "n_ref": 59, "n_ref_all": 69, "n_fig": 5, "n_tab": 2, "L_tab": 564, "n_element_tab": 13, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1478, "formula_len_all_1": 354, "len_all": 147262, "len_all_1": 63126, "len_abs": 813, "len_title": 113, "len_sents": 47467, "len_sents_1": 30575, "n_sents": 408, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 65, "L_abs": 816, "title": "Online Neuro-Symbolic Predicate Invention for High-Level Planning", "abs": "Broadly intelligent agents should form task-specific abstractions that selectively expose the essential elements of a task, while abstracting away the complexity of the raw sensorimotor space. In this work, we present Neuro-Symbolic Predicates, a first-order abstraction language that combines the strengths of symbolic and neural knowledge representations. We outline an online algorithm for inventing such predicates and learning abstract world models. We compare our approach to hierarchical reinforcement learning, vision-language model planning, and symbolic predicate invention approaches, on both in- and out-of-distribution tasks across five simulated robotic domains. Results show that our approach offers better sample complexity, stronger out-of-distribution generalization, and improved interpretability.", "keywords": [ "learning abstractions for planning", "neuro-symbolic ai", "concept learning" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 2, 5, 3 ] }, { "paper_id": "QOfWubPhdS", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Haozhe Ma", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Kuankuan Sima", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "", "position": "MS student" }, { "name": "Vinh Vo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Tze-Yun Leong", "gender": "unknown", "institution": "National University of Singapore", "country": "", "position": "Full Professor" }, { "name": "Zhengding Luo", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 15, "n_ref_uni": 59, "n_ref": 86, "n_ref_all": 105, "n_fig": 10, "n_tab": 12, "L_tab": 2980, "n_element_tab": 268, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 992, "n_element_tab_1": 48, "formula_len_all": 1480, "formula_len_all_1": 734, "len_all": 166321, "len_all_1": 63828, "len_abs": 1028, "len_title": 121, "len_sents": 46167, "len_sents_1": 30629, "n_sents": 357, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1078, "title": "Highly Efficient Self-Adaptive Reward Shaping for Reinforcement Learning", "abs": "Reward shaping is a technique in reinforcement learning that addresses the sparse-reward problem by providing more frequent and informative rewards. We introduce a self-adaptive and highly efficient reward shaping mechanism that incorporates success rates derived from historical experiences as shaped rewards. The success rates are sampled from Beta distributions, which dynamically evolve from uncertain to reliable values as data accumulates. Initially, the shaped rewards exhibit more randomness to encourage exploration, while over time, the increasing certainty enhances exploitation, naturally balancing exploration and exploitation. Our approach employs Kernel Density Estimation (KDE) combined with Random Fourier Features (RFF) to derive the Beta distributions, providing a computationally efficient, non-parametric, and learning-free solution for high-dimensional continuous state spaces. Our method is validated on various tasks with extremely sparse rewards, demonstrating notable improvements in sample efficiency and convergence stability over relevant baselines.", "keywords": [ "Reinforcement Learning", "Reward Shaping", "Thompson Sampling", "Beta Distribution", "Self-Adaptive Exploration-Exploitation Trade-off", "Success Rate" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 3, 2, 2, 4 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "QOXrVMiHGK", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Jianchen Zhu", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "", "position": "Researcher" }, { "name": "Kai Liu", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "MS student" }, { "name": "Qitan Lv", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Tianyu Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Winston Hu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Yun Li", "gender": "Male", "institution": "Tencent", "country": "", "position": "Researcher" }, { "name": "Xiao Sun", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 8, "n_ref_uni": 23, "n_ref": 47, "n_ref_all": 61, "n_fig": 1, "n_tab": 7, "L_tab": 2044, "n_element_tab": 247, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 1475, "n_element_tab_1": 160, "formula_len_all": 907, "formula_len_all_1": 491, "len_all": 132475, "len_all_1": 69671, "len_abs": 1529, "len_title": 111, "len_sents": 43620, "len_sents_1": 30874, "n_sents": 381, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1463, "title": "PEARL: Parallel Speculative Decoding with Adaptive Draft Length", "abs": "Speculative decoding (SD), where an extra draft model is employed to provide multiple **draft** tokens first and then the original target model verifies these tokens in parallel, has shown great power for LLM inference acceleration.\nHowever, existing SD methods suffer from the mutual waiting problem, i.e., the target model gets stuck when the draft model is *guessing* tokens, and vice versa. This problem is directly incurred by the asynchronous execution of the draft model and the target model, and is exacerbated due to the fixed draft length in speculative decoding.\nTo address these challenges, we propose a conceptually simple, flexible, and general framework to boost speculative decoding, namely \n**P**arallel sp**E**culative decoding with **A**daptive d**R**aft **L**ength (PEARL). \nSpecifically, PEARL proposes *pre-verify* to verify the first draft token in advance during the drafting phase, and *post-verify* to generate more draft tokens during the verification phase.\nPEARL parallels the drafting phase and the verification phase via applying the two strategies, and achieves adaptive draft length for different scenarios, which effectively alleviates the mutual waiting problem.\nExperiments on various text generation benchmarks demonstrate the effectiveness of our PEARL, leading to a superior speedup performance up to **4.43$\\times$** and **1.50$\\times$**, compared to auto-regressive decoding and vanilla speculative decoding, respectively.", "keywords": [ "speculative decoding", "inference acceleration", "large language models" ], "rating_list": [ 3, 6, 6, 8 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "QOHgjY5KDp", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Feng Gao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Haoru Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Luyi Miao", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Qi Tian", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Wentao Zhu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yishu Xu", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Yizhou Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 56, "n_ref": 129, "n_ref_all": 167, "n_fig": 17, "n_tab": 12, "L_tab": 4799, "n_element_tab": 644, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 665, "n_element_tab_1": 69, "formula_len_all": 1169, "formula_len_all_1": 933, "len_all": 202986, "len_all_1": 60881, "len_abs": 1166, "len_title": 103, "len_sents": 57843, "len_sents_1": 26215, "n_sents": 468, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 978, "title": "Aligning Human Motion Generation with Human Perceptions", "abs": "Human motion generation is a critical task with a wide spectrum of applications. Achieving high realism in generated motions requires naturalness, smoothness, and plausibility. However, current evaluation metrics often rely on simple heuristics or distribution distances and do not align well with human perceptions. In this work, we propose a data-driven approach to bridge this gap by introducing a large-scale human perceptual evaluation dataset, MotionPercept, and a human motion critic model, MotionCritic, that capture human perceptual preferences. Our critic model offers a more accurate metric for assessing motion quality and could be readily integrated into the motion generation pipeline to enhance generation quality. Extensive experiments demonstrate the effectiveness of our approach in both evaluating and improving the quality of generated human motions by aligning with human perceptions. Code and data are publicly available at https://motioncritic.github.io/.", "keywords": [ "Human Motion Generation" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "QO4bF6MHza", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Amrita Saha", "gender": "Female", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Doyen Sahoo", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Researcher" }, { "name": "Ee-Peng Lim", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Full Professor" }, { "name": "Hanze Dong", "gender": "Male", "institution": "SalesForce", "country": "SG", "position": "Researcher" }, { "name": "Lei Wang", "gender": "Male", "institution": "SalesForce", "country": "SG", "position": "Researcher" }, { "name": "Shan Dong", "gender": "Female", "institution": "Singapore Management University", "country": "SG", "position": "MS student" }, { "name": "Yalu Wang", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yuhui Xu", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Research Scientist" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 30, "n_ref": 52, "n_ref_all": 68, "n_fig": 6, "n_tab": 5, "L_tab": 4175, "n_element_tab": 12, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1783, "n_element_tab_1": 21, "formula_len_all": 96, "formula_len_all_1": 518, "len_all": 129694, "len_all_1": 55843, "len_abs": 1108, "len_title": 127, "len_sents": 44015, "len_sents_1": 25812, "n_sents": 331, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1113, "title": "MathHay: An Automated Benchmark for Long-Context Mathematical Reasoning in LLMs", "abs": "Recent large language models (LLMs) have demonstrated versatile capabilities in long-context scenarios. Although some recent benchmarks have been developed to evaluate the long-context capabilities of LLMs, there is a lack of benchmarks evaluating the mathematical reasoning abilities of LLMs over long contexts, which is crucial for LLMs' application in real-world scenarios. In this paper, we introduce MathHay, an automated benchmark designed to assess the long-context mathematical reasoning capabilities of LLMs. Unlike previous benchmarks like Needle in a Haystack, which focus primarily on information retrieval within long texts, MathHay demands models with both information-seeking and complex mathematical reasoning abilities. We conduct extensive experiments on MathHay to assess the long-context mathematical reasoning abilities of eight top-performing LLMs. Even the best-performing model, Gemini-1.5-Pro-002, still struggles with mathematical reasoning over long contexts, achieving only 51.26\\% accuracy at 128K tokens. This highlights the significant room for improvement on the MathHay benchmark.", "keywords": [ "Mathematical Reasoning", "Large Language Models", "Benchmark", "Long-Context Modeling" ], "rating_list": [ 3, 5, 5, 3, 3 ], "soundness_list": [ 1, 3, 2, 2, 2 ], "presentation_list": [ 3, 4, 3, 2, 1 ], "contribution_list": [ 1, 3, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 3, 5 ] }, { "paper_id": "QN97ubU1HH", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Chujie Zheng", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Heng Ji", "gender": "Female", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Minlie Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Nanyun Peng", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Associate Professor" }, { "name": "Ziqi Wang", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "Ph.D. student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 48, "n_ref": 94, "n_ref_all": 116, "n_fig": 6, "n_tab": 8, "L_tab": 3246, "n_element_tab": 305, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 3598, "n_element_tab_1": 268, "formula_len_all": 184, "formula_len_all_1": 184, "len_all": 151433, "len_all_1": 65736, "len_abs": 1509, "len_title": 87, "len_sents": 35197, "len_sents_1": 29424, "n_sents": 221, "n_sents_1": 160, "venue": "ICLR2025", "decision": "withdraw", "L_title": 39, "L_abs": 1602, "title": "Model Extrapolation Expedites Alignment", "abs": "As the alignment training of large language models (LLMs) usually requires expensive computational resources, exploring more efficient alignment methods to reduce training overhead has always been an important and compelling research challenge. Inspired by prior work on *model interpolation*, we present a simple method called ***ExPO (model extrapolation)*** to expedite the alignment of LLMs with human preferences. Based on our observation that interpolating the weights between existing DPO/RLHF models and their initial SFT checkpoints usually produces new models with intermediate performance, we propose to treat a partially-trained model $\\mathcal{M}_1$ (corresponding to the intermediate-performing model) as the interpolated result between the initial SFT checkpoint $\\mathcal{M}_0$ and a hypothetical better-aligned model $\\mathcal{M}_2$. Thus, we can obtain the hypothetical $\\mathcal{M}_2$ by simply extrapolating the model weights along the direction from $\\mathcal{M}_0$ to $\\mathcal{M}_1$, which consequently saves the additional training overhead for $\\mathcal{M}_1$ to reach better alignment performance. We validate our hypothesis through controlled experiments, demonstrating that ExPO can boost a DPO model trained with only 20% steps to outperform the fully-trained one. Additionally, we show that ExPO can also notably improve existing open-source LLMs (ranging from 1.8B to 70B parameters), as evidenced by evaluations on the mainstream LLM benchmarks AlpacalEval 2.0 and MT-Bench, which further highlights ExPO's utility and potential in enabling more efficient LLM alignment.", "keywords": [ "large language model", "alignment", "preference optimization", "model merging" ], "rating_list": [ 3, 3, 6, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "QMtrW8Ej98", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "David R\u00fcgamer", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Associate Professor" }, { "name": "Emanuel Sommer", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Giorgi Nozadze", "gender": "Male", "institution": "Eraneos Germany", "country": "DE", "position": "Solution Engineer" }, { "name": "Jakob Robnik", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Uros Seljak", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 33, "n_ref": 69, "n_ref_all": 109, "n_fig": 9, "n_tab": 12, "L_tab": 5248, "n_element_tab": 560, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1454, "n_element_tab_1": 224, "formula_len_all": 237, "formula_len_all_1": 159, "len_all": 170097, "len_all_1": 69196, "len_abs": 1523, "len_title": 133, "len_sents": 51943, "len_sents_1": 32558, "n_sents": 359, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1528, "title": "Microcanonical Langevin Ensembles: Advancing the Sampling of Bayesian Neural Networks", "abs": "Despite recent advances, sampling-based inference for Bayesian Neural Networks (BNNs) remains a significant challenge in probabilistic deep learning. While sampling-based approaches do not require a variational distribution assumption, current state-of-the-art samplers still struggle to navigate the complex and highly multimodal posteriors of BNNs. As a consequence, sampling still requires considerably longer inference times than non-Bayesian methods even for small neural networks, despite recent advances in making software implementations more efficient. Besides the difficulty of finding high-probability regions, the time until samplers provide sufficient exploration of these areas remains unpredictable. To tackle these challenges, we introduce an ensembling approach that leverages strategies from optimization and a recently proposed sampler called Microcanonical Langevin Monte Carlo (MCLMC) for efficient, robust and predictable sampling performance. Compared to approaches based on the state-of-the-art No-U-Turn Sampler, our approach delivers substantial speedups up to an order of magnitude, while maintaining or improving predictive performance and uncertainty quantification across diverse tasks and data modalities. The suggested Microcanonical Langevin Ensembles and modifications to MCLMC additionally enhance the method's predictability in resource requirements, facilitating easier parallelization. All in all, the proposed method offers a promising direction for practical, scalable inference for BNNs.", "keywords": [ "Sampling", "Bayesian Neural Networks" ], "rating_list": [ 8, 6, 6, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 2, 3, 4 ] }, { "paper_id": "QM2WoPu1It", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Feiyu Duan", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Ge Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Haoran Que", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Liu Jiaheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jian Yang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Junran Peng", "gender": "Male", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "Associate Professor" }, { "name": "Kai Chen", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Liqun He", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Songyang Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Wangchunshu Zhou", "gender": "Male", "institution": "Guangdong OPPO Mobile Telecommunications Corp.,Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Wenge Rong", "gender": "Male", "institution": "Beihang University", "country": "", "position": "Full Professor" }, { "name": "Yutao Mou", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhaoxiang Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zekun Wang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 3, "n_ref_uni": 69, "n_ref": 261, "n_ref_all": 340, "n_fig": 11, "n_tab": 12, "L_tab": 5993, "n_element_tab": 443, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 4253, "n_element_tab_1": 497, "formula_len_all": 561, "formula_len_all_1": 464, "len_all": 378273, "len_all_1": 68457, "len_abs": 1577, "len_title": 130, "len_sents": 126851, "len_sents_1": 27080, "n_sents": 1092, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1583, "title": "HelloBench: Evaluating Long Text Generation Capabilities of Large Language Models", "abs": "In recent years, Large Language Models (LLMs) have demonstrated remarkable capabilities in various tasks (e.g., long-context understanding), and many benchmarks have been proposed. However, we observe that long text generation capabilities are not well investigated. Therefore, we introduce the Hierarchical Long Text Generation Benchmark (HelloBench), a comprehensive, in-the-wild, and open-ended benchmark to evaluate LLMs' performance in generating long text. Based on Bloom's Taxonomy, HelloBench categorizes long text generation tasks into five subtasks: open-ended QA, summarization, chat, text completion, and heuristic text generation. Besides, we propose Hierarchical Long Text Evaluation (HelloEval), a human-aligned evaluation method that significantly reduces the time and effort required for human evaluation while maintaining a high correlation with human evaluation. We have conducted extensive experiments across around 30 mainstream LLMs and observed that the current LLMs lack long text generation capabilities. Specifically, first, regardless of whether the instructions include explicit or implicit length constraints, we observe that most LLMs cannot generate text that is longer than 4000 words. Second, we observe that while some LLMs can generate longer text, many issues exist (e.g., severe repetition and quality degradation). Third, to demonstrate the effectiveness of HelloEval, we compare HelloEval with traditional metrics (e.g., ROUGE, BLEU, etc.) and LLM-as-a-Judge methods, which show that HelloEval has the highest correlation with human evaluation.", "keywords": [ "Large Language Models", "Long Text Generation", "Benchmark" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "QLOGfFSB50", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Bahram Zonooz", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "Elahe Arani", "gender": "Female", "institution": "Eindhoven University of technology ", "country": "NL", "position": "Assistant Professor" }, { "name": "Laurens Niesten", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "MS student" }, { "name": "Prashant Shivaram Bhat", "gender": "Male", "institution": "TomTom", "country": "NL", "position": "Applied Scientist III" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 74, "n_ref": 126, "n_ref_all": 164, "n_fig": 8, "n_tab": 12, "L_tab": 4570, "n_element_tab": 411, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1766, "n_element_tab_1": 122, "formula_len_all": 328, "formula_len_all_1": 328, "len_all": 247061, "len_all_1": 64697, "len_abs": 6184, "len_title": 122, "len_sents": 73795, "len_sents_1": 30982, "n_sents": 484, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1334, "title": "SPARC: Continual learning beyond experience rehearsal and model surrogates", "abs": "Continual learning (CL) has become increasingly important as deep neural networks\n(DNNs) are required to adapt to the continuous influx of data without retraining\nfrom scratch. However, a significant challenge in CL is catastrophic forgetting (CF),\nwhere learning new tasks erases previously acquired knowledge, either partially\nor completely. Existing solutions often rely on experience rehearsal or full model\nsurrogates to mitigate CF. While effective, these approaches introduce substantial\nmemory and computational overhead, limiting their scalability and applicability in\nreal-world scenarios. To address this, we propose SPARC, a scalable CL approach\nthat eliminates the need for experience rehearsal and full-model surrogates. By\neffectively combining task-specific working memories and task-agnostic semantic\nmemory for cross-task knowledge consolidation, SPARC results in a remarkable\nparameter efficiency, using only 6% of the parameters required by full-model\nsurrogates. Despite its lightweight design, SPARC achieves superior performance\non Seq-TinyImageNet and matches rehearsal-based methods on various CL benchmarks. Additionally, weight re-normalization in the classification layer mitigates\ntask-specific biases, establishing SPARC as a practical and scalable solution for\nCL under stringent efficiency constraints.", "keywords": [ "Continual learning", "lifelong learning", "computer vision", "experience rehearsal", "parameter isolation" ], "rating_list": [ 5, 5, 6, 6, 5 ], "soundness_list": [ 2, 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 1, 2 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "QKywN4BbqA", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Runze Ma", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Yanjie Huang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhongyue Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "SHUANGJIA ZHENG", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 14, "n_ref_uni": 41, "n_ref": 84, "n_ref_all": 98, "n_fig": 7, "n_tab": 7, "L_tab": 8657, "n_element_tab": 607, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 3582, "n_element_tab_1": 340, "formula_len_all": 1749, "formula_len_all_1": 1151, "len_all": 162064, "len_all_1": 76519, "len_abs": 1291, "len_title": 159, "len_sents": 42787, "len_sents_1": 29591, "n_sents": 326, "n_sents_1": 287, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1305, "title": "$E^3$former: An Adaptive Energy-Aware Elastic Equivariant Transformer Model For Protein Representation Learning", "abs": "Structure-informed protein representation learning is essential for effective protein function annotation and \\textit{de novo} design. However, the presence of inherent noise in both crystal and AlphaFold-predicted structures poses significant challenges for existing methods in learning robust protein representations. To address these issues, we propose a novel equivariant Transformer-State Space Model(SSM) hybrid framework, termed $E^3$former, designed for efficient protein representation. Our approach leverages energy function-based receptive fields to construct proximity graphs and incorporates an equivariant high-tensor-elastic selective SSM within the transformer architecture. These components enable the model to adapt to complex atom interactions and extract geometric features with higher signal-to-noise ratios. Empirical results demonstrate that our model outperforms existing methods in structure-intensive tasks, such as inverse folding and binding site prediction, particularly when using predicted structures, owing to its enhanced tolerance to data deviation and noise. Our approach offers a novel perspective for conducting biological function research and drug discovery using noisy protein structure data. Our code is available on https://anonymous.4open.science/r/E3former-207E", "keywords": [ "protein representation learning", "geometric deep learning", "equivariant neural networks", "proteins" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "QKBu1BOAwd", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Changle Qu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Dawei Yin", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Principal Researcher" }, { "name": "Hengyi Cai", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Jun Xu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Shuaiqiang Wang", "gender": "Male", "institution": "Baidu Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Sunhao Dai", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Xiaochi Wei", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 7, "n_ref_uni": 38, "n_ref": 74, "n_ref_all": 113, "n_fig": 7, "n_tab": 6, "L_tab": 2975, "n_element_tab": 278, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 633, "n_element_tab_1": 91, "formula_len_all": 178, "formula_len_all_1": 193, "len_all": 198436, "len_all_1": 79248, "len_abs": 1881, "len_title": 135, "len_sents": 64925, "len_sents_1": 37251, "n_sents": 430, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 87, "L_abs": 1658, "title": "From Exploration to Mastery: Enabling LLMs to Master Tools via Self-Driven Interactions", "abs": "Tool learning enables Large Language Models (LLMs) to interact with external environments by invoking tools, serving as an effective strategy to mitigate the limitations inherent in their pre-training data. In this process, tool documentation plays a crucial role by providing usage instructions for LLMs, thereby facilitating effective tool utilization. This paper concentrates on the critical challenge of bridging the comprehension gap between LLMs and external tools due to the inadequacies and inaccuracies inherent in existing human-centric tool documentation. We propose a novel framework, DRAFT, aimed at Dynamically Refining tool documentation through the Analysis of Feedback and Trials emanating from LLMs' interactions with external tools. This methodology pivots on an innovative trial-and-error approach, consisting of three distinct learning phases: experience gathering, learning from experience, and documentation rewriting, to iteratively enhance the tool documentation. This process is further optimized by implementing a diversity-promoting exploration strategy to ensure explorative diversity and a tool-adaptive termination mechanism to prevent overfitting while enhancing efficiency. \nExtensive experiments on multiple datasets demonstrate that DRAFT's iterative, feedback-based refinement significantly ameliorates documentation quality, fostering a deeper comprehension and more effective utilization of tools by LLMs. Notably, our analysis reveals that the tool documentation refined via our approach demonstrates robust cross-model generalization capabilities. Our code is available at https://anonymous.4open.science/r/DRAFT-10B3.", "keywords": [ "Large Language Model", "Tool Learning", "Learning from Experience" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 2, 4, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 4, 4 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "QIxFo9mFwR", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Cheng Perng Phoo", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Kavita Bala", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Professor" }, { "name": "Shreelekha Revankar", "gender": "Female", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Utkarsh Mall", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Postdoc" }, { "name": "Bharath Hariharan", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 36, "n_ref": 66, "n_ref_all": 78, "n_fig": 4, "n_tab": 1, "L_tab": 48, "n_element_tab": 3, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 48, "n_element_tab_1": 3, "formula_len_all": 235, "formula_len_all_1": 235, "len_all": 104379, "len_all_1": 59152, "len_abs": 1474, "len_title": 118, "len_sents": 31454, "len_sents_1": 27485, "n_sents": 256, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1019, "title": "Scale-aware Recognition in Satellite Images under Resource Constraints", "abs": "Recognition of features in satellite imagery (forests, swimming pools, etc.) depends strongly on the spatial scale of the concept and therefore the resolution of the images. This poses two challenges:\nWhich resolution is best suited for recognizing a given concept, and where and when should the costlier higher-resolution (HR) imagery be acquired? \nWe present a novel scheme to address these challenges by introducing three components: (1) A technique to distill knowledge from models trained on HR imagery to recognition models that operate on imagery of lower resolution (LR), (2) a sampling strategy for HR imagery based on model disagreement, and (3) an LLM-based approach for inferring concept \"scale\". With these components we present a system to efficiently perform scale-aware recognition in satellite imagery, improving accuracy over single-scale inference while following budget constraints. **Our novel approach offers up to a 26.3\\% improvement over entirely HR baselines, using 76.3 \\% fewer HR images.**", "keywords": [ "Satellite Imagery", "Resolution" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "QIsnwejVYE", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Chengli Zhao", "gender": "Male", "institution": "National University of Defense Technology", "country": "", "position": "Associate Professor" }, { "name": "Xiaojun Duan", "gender": "Female", "institution": "National University of Defense Technology", "country": "", "position": "Full Professor" }, { "name": "Xin Li", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Xue Zhang", "gender": "unknown", "institution": "National University of Defense Technology", "country": "", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 15, "n_ref_uni": 31, "n_ref": 58, "n_ref_all": 79, "n_fig": 8, "n_tab": 4, "L_tab": 1458, "n_element_tab": 90, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1286, "n_element_tab_1": 88, "formula_len_all": 2699, "formula_len_all_1": 1549, "len_all": 136845, "len_all_1": 57318, "len_abs": 1255, "len_title": 127, "len_sents": 46369, "len_sents_1": 26073, "n_sents": 330, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1225, "title": "Robust Latent Neural Operators for a Family of Systems with Sparse Observations", "abs": "Neural operator methods have achieved significant success in the efficient simulation and inverse problems of complex systems by learning a mapping between two infinite-dimensional Banach spaces. However, existing methods still exhibit room for optimization in terms of robustness and modeling accuracy. Specifically, existing methods are characterized by sensitivity to noise and a tendency to overlook the importance of sparse observations. Therefore, we propose a robust latent neural operator based on the variational autoencoder framework. In this method, an encoder based on recurrent neural networks effectively extracts sequential information and dynamical characteristics embedded in sparse observations. Subsequently, a neural operator in latent space and a decoder facilitate the modelling of the original system. Additionally, for certain higher-dimensional systems, opting for a lower-dimensional latent space can reduce task complexity while still maintaining satisfactory modeling performance. We conduct experiments across several representative systems, and the results validate that our method achieves superior modeling accuracy and enhanced robustness compared to the state of the art baseline approaches.", "keywords": [ "Neural operator", "complex system", "variational autoencoder", "recurrent neural network" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "QIp1YUqgCB", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Bai LiChen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Kaishun Wu", "gender": "Male", "institution": "HKUST(GZ)", "country": "CN", "position": "Full Professor" }, { "name": "Shitong Shao", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Tian Ye", "gender": "Not Specified", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Zeke Xie", "gender": "Male", "institution": "HKUST(GZ)", "country": "CN", "position": "Assistant Professor" }, { "name": "zikai zhou", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "YUNFENG CAI", "gender": "Male", "institution": "Beijing Institute of Mathematical Sciences and Applications", "country": "CN", "position": "Full Professor" } ], "n_formula": 75, "n_formula_1": 16, "n_ref_uni": 51, "n_ref": 130, "n_ref_all": 169, "n_fig": 12, "n_tab": 9, "L_tab": 3957, "n_element_tab": 482, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3525, "n_element_tab_1": 352, "formula_len_all": 7986, "formula_len_all_1": 1646, "len_all": 222949, "len_all_1": 66662, "len_abs": 2568, "len_title": 85, "len_sents": 62303, "len_sents_1": 23392, "n_sents": 478, "n_sents_1": 160, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1414, "title": "The Blessing of Smooth Initialization for Video Diffusion Models", "abs": "Extending the success of text-to-image (T2I) synthesis to text-to-video (T2V) synthesis is a promising direction for visual generative AI. Popular training-free sampling algorithms currently generate high-fidelity images within the Stable Diffusion family. However, when applied to video diffusion models (VDMs), these techniques result in limited diversity and quality due to the low-quality data in a video datasets. We focus on inference to mitigate this issue, and then we propose a training-free paradigm that optimizes the initial Gaussian noise by introducing a targeted semantic prior bias into the sampling process from a smoothing perspective. The paradigm significantly improves both the fidelity and semantic faithfulness of the synthesized videos. Guided by theoretical analysis using random smoothing and differential equations, our resulting method SmoothInit can be understood as approximately incorporating third-order derivatives into gradient descent, which contributes to be better convergence in learning semantic information. A more efficient version, Fast-SmoothInit, is proposed to achieve better experimental results by leveraging a momentum mechanism. Both SmoothInit and Fast-SmoothInit demonstrate promising empirical results across various benchmarks, including UCF-101/MSR-VTT-related FVD, Chronomagic-bench, and T2V-Compbench, setting a new standard for noise initialization in VDMs.", "keywords": [ "Smoothing", "Video Diffusion Models", "Noise Initialization" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 1, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "QIfzMeTyOu", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Diego Paez-Granados", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Principal Researcher" }, { "name": "Jiongfan Chen", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Mehdi Ejtehadi", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 34, "n_ref": 57, "n_ref_all": 66, "n_fig": 4, "n_tab": 1, "L_tab": 729, "n_element_tab": 77, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1833, "n_element_tab_1": 171, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 111405, "len_all_1": 55010, "len_abs": 1129, "len_title": 135, "len_sents": 28017, "len_sents_1": 25579, "n_sents": 175, "n_sents_1": 161, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1125, "title": "GEAR-FEN: Generalized Feature Representation for Kinematic Human Activity Recognition", "abs": "This study addresses the challenge of efficient human activity recognition (HAR) with limited training data. We propose GEAR-FEN (Generalized Activity Recognition Feature Extraction Network), a novel transfer learning method that transforms kinematic motion signals into a generalized feature space. GEAR-FEN potentially outperforms the state-of-the-art in scenarios with limited training data. This was demonstrated through an evaluation across 11 public HAR datasets (encompassing number of activities ranging from 6 to 33 and number of samples per activity ranging from 8628 to 1140258), using a deep learning model based on convolutional neural networks (CNN), residual bi-directional long short-term memory (ResBiLSTM), and an attention mechanism. Furthermore, we establish the generalizability of our method through performance comparisons on an independent dataset encompassing a distinct population and diverse kinematic modalities for 8 activities, and 26121 samples per activity. These findings highlight the potential of our proposed approach in robust feature representation for HAR tasks with limited data sizes.", "keywords": [ "Human Activity Recognition", "Transfer Learning", "Machine Learning", "Classification" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "QIApiYYgLG", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Martin Vechev", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Nikola Jovanovi\u0107", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Robin Staab", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Thibaud Gloaguen", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" } ], "n_formula": 29, "n_formula_1": 12, "n_ref_uni": 39, "n_ref": 97, "n_ref_all": 146, "n_fig": 13, "n_tab": 5, "L_tab": 2075, "n_element_tab": 412, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1688, "n_element_tab_1": 144, "formula_len_all": 2247, "formula_len_all_1": 1009, "len_all": 178644, "len_all_1": 69912, "len_abs": 1159, "len_title": 90, "len_sents": 62449, "len_sents_1": 32284, "n_sents": 508, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1164, "title": "Discovering Clues of Spoofed LM Watermarks", "abs": "LLM watermarks stand out as a promising way to attribute ownership of LLM-generated text. One threat to watermark credibility comes from spoofing attacks, where an unauthorized third party forges the watermark, enabling it to falsely attribute arbitrary texts to a particular LLM. While recent works have demonstrated that state-of-the-art schemes are in fact vulnerable to spoofing, they lack deeper qualitative analysis of the texts produced by spoofing methods. In this work, we for the first time reveal that there are observable differences between genuine and spoofed watermark texts. Namely, we show that regardless of their underlying approach, all current learning-based spoofing methods consistently leave observable artifacts in spoofed texts, indicative of watermark forgery. We build upon these findings to propose rigorous statistical tests that reliably reveal the presence of such artifacts, effectively discovering that a watermark was spoofed. Our experimental evaluation shows high test power across all current learning-based spoofing methods, providing insights into their fundamental limitations, and suggesting a way to mitigate this threat.", "keywords": [ "llm", "watermarks", "spoofing" ], "rating_list": [ 6, 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "QHj2LL958o", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Bo Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Canyu Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Chunhua Shen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Fan Wang", "gender": "Female", "institution": "Alibaba Group", "country": "US", "position": "Senior Staff Algorithm Engineer" }, { "name": "Hao Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Mingyu Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Weihua Chen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Algorithm Engineer" }, { "name": "Wen Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 5, "n_ref_uni": 45, "n_ref": 97, "n_ref_all": 124, "n_fig": 18, "n_tab": 4, "L_tab": 1222, "n_element_tab": 160, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 437, "n_element_tab_1": 73, "formula_len_all": 780, "formula_len_all_1": 260, "len_all": 223661, "len_all_1": 64854, "len_abs": 1407, "len_title": 120, "len_sents": 63661, "len_sents_1": 30664, "n_sents": 591, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1337, "title": "MovieDreamer: Hierarchical Generation for Coherent Long Visual Sequences", "abs": "Recent advancements in video generation have primarily leveraged diffusion models for short-duration content. However, these approaches often fall short in modeling complex narratives and maintaining character consistency over extended periods, which is essential for long-form video production like movies. We propose MovieDreamer, a novel hierarchical framework that integrates the strengths of autoregressive models with diffusion-based rendering to pioneer long-duration video generation with intricate plot progressions and high visual fidelity. Our approach utilizes autoregressive models for global narrative coherence, predicting sequences of visual tokens that are subsequently transformed into high-quality video frames through diffusion rendering. This method is akin to traditional movie production processes, where complex stories are factorized down into manageable scene capturing. Further, we employ a multimodal script that enriches scene descriptions with detailed character information and visual style, enhancing continuity and character identity across scenes. We present extensive experiments across various movie genres, demonstrating that our approach not only achieves superior visual and narrative quality but also effectively extends the duration of generated content significantly beyond current capabilities.", "keywords": [ "Generative Model", "Vison Language Model", "Long Video Generation", "Long Story Generation" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 4 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "QGGNvKaoIU", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "Dennis Frauen", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Konstantin Hess", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Stefan Feuerriegel", "gender": "Male", "institution": "LMU Munich", "country": "DE", "position": "Professor" } ], "n_formula": 97, "n_formula_1": 20, "n_ref_uni": 41, "n_ref": 159, "n_ref_all": 191, "n_fig": 3, "n_tab": 1, "L_tab": 143, "n_element_tab": 14, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 3063, "n_element_tab_1": 125, "formula_len_all": 13075, "formula_len_all_1": 1467, "len_all": 184115, "len_all_1": 73644, "len_abs": 1242, "len_title": 133, "len_sents": 55738, "len_sents_1": 31595, "n_sents": 475, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1209, "title": "Model-agnostic meta-learners for estimating heterogeneous treatment effects over time", "abs": "Estimating heterogeneous treatment effects (HTEs) over time is crucial in many disciplines such as personalized medicine. Existing works for this task have mostly focused on *model-based* learners that adapt specific machine-learning models and adjustment mechanisms. In contrast, model-agnostic learners - so-called *meta-learners* - are largely unexplored. In our paper, we propose several meta-learners that are model-agnostic and thus can be used in combination with arbitrary machine learning models (e.g., transformers) to estimate HTEs over time. We then provide a comprehensive theoretical analysis that characterizes the different learners and that allows us to offer insights into when specific learners are preferable. Furthermore, we propose a novel IVW-DR-learner that (i) uses a doubly robust (DR) and orthogonal loss; and (ii) leverages inverse-variance weights (IVWs) that we derive to stabilize the DR-loss over time. Our IVWs downweight extreme trajectories due to *products* of inverse-propensities in the DR-loss, resulting in a lower estimation variance. Our IVW-DR-learner achieves superior performance in our experiments, particularly in regimes with low overlap and long time horizons.", "keywords": [ "causal inference", "treatment effect estimation", "CATE", "doubly robust learning" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 4, 2, 4, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "QG31By6S6w", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Shaoting Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Full Professor" }, { "name": "Wenhui Lei", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Xiaofan Zhang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yankai Jiang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 43, "n_ref": 95, "n_ref_all": 131, "n_fig": 10, "n_tab": 8, "L_tab": 3243, "n_element_tab": 296, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 978, "n_element_tab_1": 83, "formula_len_all": 890, "formula_len_all_1": 980, "len_all": 187461, "len_all_1": 69481, "len_abs": 1240, "len_title": 171, "len_sents": 59883, "len_sents_1": 32697, "n_sents": 429, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 122, "L_abs": 1245, "title": "Unleashing the Potential of Vision-Language Pre-Training for 3D Zero-Shot Lesion Segmentation via Mask-Attribute Alignment", "abs": "Recent advancements in medical vision-language pre-training models have driven significant progress in zero-shot disease recognition. However, transferring image-level knowledge to pixel-level tasks, such as lesion segmentation in 3D CT scans, remains a critical challenge. Due to the complexity and variability of pathological visual characteristics, existing methods struggle to align fine-grained lesion features not encountered during training with disease-related textual representations. In this paper, we present Malenia, a novel multi-scale lesion-level mask-attribute alignment framework, specifically designed for 3D zero-shot lesion segmentation. Malenia improves the compatibility between mask representations and their associated elemental attributes, explicitly linking the visual features of unseen lesions with the extensible knowledge learned from previously seen ones. Furthermore, we design a Cross-Modal Knowledge Injection module to enhance both visual and textual features with mutually beneficial information, effectively guiding the generation of segmentation results. Comprehensive experiments across three datasets and 12 lesion categories validate the superior performance of Malenia. Codes will be publicly available.", "keywords": [ "Medical Image Segmentation", "Vision-Language Pre-Training", "Zero-Shot Segmentation" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "QFmnhgEnIB", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Amnon Shashua", "gender": "Male", "institution": "Hebrew University, Hebrew University of Jerusalem", "country": "", "position": "Full Professor" }, { "name": "Binyamin Rothberg", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "MS student" }, { "name": "Dorin Shteyman", "gender": "Female", "institution": "Hebrew University of Jerusalem", "country": "", "position": "MS student" }, { "name": "Noam Wies", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "", "position": "PhD student" }, { "name": "Yoav Levine", "gender": "Male", "institution": "AI21 Labs", "country": "", "position": "Principal Researcher" }, { "name": "Yotam Wolf", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "PhD student" } ], "n_formula": 52, "n_formula_1": 11, "n_ref_uni": 51, "n_ref": 73, "n_ref_all": 102, "n_fig": 19, "n_tab": 2, "L_tab": 46, "n_element_tab": 8, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3997, "formula_len_all_1": 557, "len_all": 235010, "len_all_1": 69677, "len_abs": 1680, "len_title": 143, "len_sents": 76200, "len_sents_1": 36302, "n_sents": 512, "n_sents_1": 226, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1489, "title": "Tradeoffs Between Alignment and Helpfulness in Language Models with Representation Engineering", "abs": "Language model alignment has become an important component of AI safety, allowing safe interactions between humans and language models, by enhancing desired behaviors and inhibiting undesired ones. It is often done by tuning the model or inserting preset aligning prompts. Recently, representation engineering, a method which alters the model\u2019s behavior via changing its representations post-training, was shown to be effective in aligning LLMs (Zou et al., 2023a). Representation engineering yields gains in alignment oriented tasks such as resistance to adversarial attacks and reduction of social biases, but was also shown to cause a decrease in the ability of the model to perform basic tasks. In this paper we study the tradeoff between the increase in alignment and decrease in helpfulness of the model. We propose a theoretical framework which provides bounds for these two quantities, and demonstrate their relevance empirically. First, we find that under the conditions of our framework, alignment can be guaranteed with representation engineering, and at the same time that helpfulness is harmed in the process. Second, we show that helpfulness is harmed quadratically with the norm of the representation engineering vector, while the alignment increases linearly with it, indicating a regime in which it is efficient to use representation engineering. We validate our findings empirically, and chart the boundaries to the usefulness of representation engineering for alignment.", "keywords": [ "Language model alignment", "representation engineering" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 4, 2, 2 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "QFgbJOYJSE", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Naoki Nishikawa", "gender": "Male", "institution": "Graduate School of Information Science and Technology, The University of Tokyo", "country": "", "position": "PhD student" }, { "name": "Taiji Suzuki", "gender": "Male", "institution": "The University of Tokyo", "country": "", "position": "Full Professor" } ], "n_formula": 167, "n_formula_1": 30, "n_ref_uni": 8, "n_ref": 39, "n_ref_all": 55, "n_fig": 4, "n_tab": 3, "L_tab": 1429, "n_element_tab": 31, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 556, "n_element_tab_1": 6, "formula_len_all": 25087, "formula_len_all_1": 3146, "len_all": 234658, "len_all_1": 74079, "len_abs": 1104, "len_title": 133, "len_sents": 76722, "len_sents_1": 30371, "n_sents": 905, "n_sents_1": 272, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1108, "title": "State Space Models are Provably Comparable to Transformers in Dynamic Token Selection", "abs": "Deep neural networks based on state space models (SSMs) are attracting significant attention in sequence modeling since their computational cost is significantly smaller than that of Transformers. While the capabilities of SSMs have been demonstrated through experiments in various tasks, theoretical understanding of SSMs is still limited. In particular, most theoretical studies discuss the capabilities of SSM layers without nonlinear layers, and there is a lack of discussion on their combination with nonlinear layers. In this paper, we explore the capabilities of SSMs combined with fully connected neural networks, and show that they are comparable to Transformers in extracting the essential tokens depending on the input. As concrete examples, we consider two synthetic tasks, which are challenging for a single SSM layer, and demonstrate that SSMs combined with nonlinear layers can efficiently solve these tasks. Furthermore, we study the nonparametric regression task, and prove that the ability of SSMs is equivalent to that of Transformers in estimating functions belonging to a certain class.", "keywords": [ "State Space Model", "Transformer", "Nonparametric regression" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "QFaj7InstQ", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Anushya Subbiah", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Hardik Patel", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Judith Yue Li", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Li Yang", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Vikram Aggarwal", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Yanwei Song", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 22, "n_ref": 37, "n_ref_all": 59, "n_fig": 5, "n_tab": 6, "L_tab": 3822, "n_element_tab": 465, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1718, "n_element_tab_1": 284, "formula_len_all": 128, "formula_len_all_1": 417, "len_all": 103254, "len_all_1": 70717, "len_abs": 1255, "len_title": 67, "len_sents": 30581, "len_sents_1": 25178, "n_sents": 240, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 19, "L_abs": 1265, "title": "Item Language Model", "abs": "Embeddings are extensively used in many domains to represent information about domain entities in a compressed manner. In recommendation systems, these embeddings are trained to extract meaningful information about an item/user from collaborative filtering data consisting users ratings or implicit feedback on items. These behavioral embeddings are usually not trained on data from language domain, but they encode very useful behavioral information which cannot be described using language. In contrast, in large language models (LLM) this collaborative data and behavioral entities(users/items) are not well represented as they are not textual and are specific to the recommendation system/product. Bridging this gap between behavioral understanding and language understanding can enable new item and language interleaved tasks. In our work we show how we can efficiently adapt rich behavioral embeddings as an additional behavioral input representation in pre-trained LLMs. To achieve this we adapt Querying Transformer technique with a new item contrastive loss and show improved item-text joint understanding in PALM2. Finally, we also demonstrate improved capabilities in recommendation domain over using the behavioral embeddings directly as input to PALM2.", "keywords": [ "large language models", "interpretability methods", "representation", "recommender systems" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "QFO1asgas2", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Aaron Courville", "gender": "unknown", "institution": "University of Montreal", "country": "", "position": "Assistant Professor" }, { "name": "Gauthier Gidel", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Assistant Professor" }, { "name": "Juan Agustin Duque", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "Milad Aghajohari", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Researcher" }, { "name": "Tianyu Zhang", "gender": "Not Specified", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, University of Montreal", "country": "CA", "position": "PhD student" }, { "name": "Tim Cooijmans", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "PhD student" }, { "name": "razvan ciuca", "gender": "unknown", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "MS student" } ], "n_formula": 63, "n_formula_1": 20, "n_ref_uni": 33, "n_ref": 73, "n_ref_all": 94, "n_fig": 8, "n_tab": 3, "L_tab": 1611, "n_element_tab": 53, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6489, "formula_len_all_1": 1609, "len_all": 150457, "len_all_1": 57549, "len_abs": 1259, "len_title": 78, "len_sents": 49250, "len_sents_1": 27115, "n_sents": 372, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 30, "L_abs": 1264, "title": "Advantage Alignment Algorithms", "abs": "Artificially intelligent agents are increasingly being integrated into human decision-making: from large language model (LLM) assistants to autonomous vehicles. These systems often optimize their individual objective, leading to conflicts, particularly in general-sum games where naive reinforcement learning agents empirically converge to Pareto-suboptimal Nash equilibria. To address this issue, opponent shaping has emerged as a paradigm for finding socially beneficial equilibria in general-sum games. In this work, we introduce Advantage Alignment, a family of algorithms derived from first principles that perform opponent shaping efficiently and intuitively. We achieve this by aligning the advantages of interacting agents, increasing the probability of mutually beneficial actions when their interaction has been positive. We prove that existing opponent shaping methods implicitly perform Advantage Alignment. Compared to these methods, Advantage Alignment simplifies the mathematical formulation of opponent shaping, reduces the computational burden and extends to continuous action domains. We demonstrate the effectiveness of our algorithms across a range of social dilemmas, achieving state-of-the-art cooperation and robustness against exploitation.", "keywords": [ "Multi-agent Reinforcement Learning", "Opponent Shaping", "Social Dilemmas", "General-Sum Games" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 2, 3, 4, 4 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 2, 4, 2, 4 ] }, { "paper_id": "QETk0lBdVf", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Bo Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Chunyuan Li", "gender": "unknown", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Guangtao Zeng", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "", "position": "PhD student" }, { "name": "Haoran Tan", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "Undergrad student" }, { "name": "Jingkang Yang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Kaichen Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Undergrad student" }, { "name": "Peiyuan Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Yuanhan Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Ziwei Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Ziyue Wang", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 40, "n_ref": 113, "n_ref_all": 141, "n_fig": 6, "n_tab": 7, "L_tab": 1190, "n_element_tab": 187, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 2302, "n_element_tab_1": 203, "formula_len_all": 114, "formula_len_all_1": 75, "len_all": 153228, "len_all_1": 57118, "len_abs": 1094, "len_title": 94, "len_sents": 29737, "len_sents_1": 24486, "n_sents": 235, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1089, "title": "Long Context Transfer from Language to Vision", "abs": "Video sequences offer valuable temporal information, but existing large multimodal models (LMMs) fall short in understanding extremely long videos. Many works address this by reducing the number of visual tokens using visual resamplers. Alternatively, in this paper, we approach this problem from the perspective of the language model. By simply extrapolating the context length of the language backbone, we enable LMMs to comprehend orders of magnitude more visual tokens without any video training. We call this phenomenon long context transfer and carefully ablate its properties. To effectively measure LMMs' ability to generalize to long contexts in the vision modality, we develop V-NIAH (Visual Needle-In-A-Haystack), a purely synthetic long vision benchmark inspired by the language model's NIAH test. Our proposed Long Video Assistant (LongVA) can process 2000 frames or over 200K visual tokens without additional complexities. With its extended context length, LongVA achieves state-of-the-art performance on Video-MME among 7B-scale models by densely sampling more input frames.", "keywords": [ "Vision Language Model", "Long Context Model" ], "rating_list": [ 5, 5, 8, 5, 6 ], "soundness_list": [ 4, 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 4, 2, 2 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 5, 3 ] }, { "paper_id": "QEHrmQPBdd", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Juanzi Li", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Lei Hou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Rui Min", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yantao Liu", "gender": "Not Specified", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Yixin Cao", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Zijun Yao", "gender": "Male", "institution": "Department of Computer Science and Technology, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 41, "n_ref": 99, "n_ref_all": 129, "n_fig": 10, "n_tab": 8, "L_tab": 9034, "n_element_tab": 435, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 943, "n_element_tab_1": 43, "formula_len_all": 610, "formula_len_all_1": 299, "len_all": 218322, "len_all_1": 61225, "len_abs": 1236, "len_title": 127, "len_sents": 62330, "len_sents_1": 25843, "n_sents": 495, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 79, "L_abs": 1245, "title": "RM-Bench: Benchmarking Reward Models of Language Models with Subtlety and Style", "abs": "Reward models are critical in techniques like Reinforcement Learning from Human Feedback (RLHF) and Inference Scaling Laws, where they guide language model alignment and select optimal responses. \nDespite their importance, existing reward model benchmarks often evaluate models by asking them to distinguish between responses generated by models of varying power. \nHowever, this approach fails to assess reward models on subtle but critical content changes and variations in style, resulting in a low correlation with policy model performance.\nTo this end, we introduce RM-Bench, a novel benchmark designed to evaluate reward models based on their sensitivity to subtle content differences and resistance to style biases. \nExtensive experiments demonstrate that RM-Bench strongly correlates with policy model performance, making it a reliable reference for selecting reward models to align language models effectively.\nWe evaluate nearly 40 reward models on RM-Bench. \nOur results reveal that even state-of-the-art models achieve an average performance of only 46.6%, which falls short of random-level accuracy (50%) when faced with style bias interference.\nThese findings highlight the significant room for improvement in current reward models.", "keywords": [ "Reward Models", "Language Models", "Evaluation", "Alignment" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "QE1LFzXQPL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Bhiksha Raj", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Hao Chen", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Jason Kuen", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Researcher" }, { "name": "Jiuxiang Gu", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Kai Qiu", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Xiang Li", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Zhe Lin", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Principal Researcher" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 62, "n_ref": 164, "n_ref_all": 189, "n_fig": 13, "n_tab": 4, "L_tab": 701, "n_element_tab": 78, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 484, "n_element_tab_1": 51, "formula_len_all": 338, "formula_len_all_1": 311, "len_all": 175900, "len_all_1": 58460, "len_abs": 574, "len_title": 112, "len_sents": 35774, "len_sents_1": 26033, "n_sents": 270, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1262, "title": "ImageFolder: Autoregressive Image Generation with Folded Tokens", "abs": "Image tokenizers are crucial for visual generative models, \\eg, diffusion models (DMs) and autoregressive (AR) models, as they construct the latent representation for modeling. Increasing token length is a common approach to improve image reconstruction quality. However, tokenizers with longer token lengths are not guaranteed to achieve better generation quality. There exists a trade-off between reconstruction and generation quality regarding token length. In this paper, we investigate the impact of token length on both image reconstruction and generation and provide a flexible solution to the tradeoff. We propose \\textbf{ImageFolder}, a semantic tokenizer that provides spatially aligned image tokens that can be folded during autoregressive modeling to improve both efficiency and quality. To enhance the representative capability without increasing token length, we leverage dual-branch product quantization to capture different contexts of images. Specifically, semantic regularization is introduced in one branch to encourage compacted semantic information while another branch is designed to capture pixel-level details. Extensive experiments demonstrate the superior quality of image generation and shorter token length with ImageFolder tokenizer.", "keywords": [ "Semantic tokenizer", "Autoregressive generation", "Product quantization" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "QE1ClsZjOQ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Guanyu Zhang", "gender": "Female", "institution": "Lenovo Group Limited", "country": "CN", "position": "Researcher" }, { "name": "Hongsheng Qi", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "Lecturer" }, { "name": "Wenfa Wu", "gender": "Male", "institution": "Lenovo Research", "country": "CN", "position": "Researcher" }, { "name": "Yi Wang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Algorithm Engineer" }, { "name": "tan zheng", "gender": "Male", "institution": "Lenovo Group Limited", "country": "CN", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 66, "n_ref_all": 79, "n_fig": 7, "n_tab": 9, "L_tab": 2638, "n_element_tab": 160, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 1738, "n_element_tab_1": 125, "formula_len_all": 1127, "formula_len_all_1": 603, "len_all": 140699, "len_all_1": 67406, "len_abs": 1879, "len_title": 140, "len_sents": 44676, "len_sents_1": 30718, "n_sents": 296, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1886, "title": "Dual-Forecaster: A Multimodal Time Series Model Integrating Descriptive and Predictive Texts", "abs": "Time series forecasting plays a vital role for decision-making across a wide range of real-world domains, which has been extensively studied. Most existing single-modal models rely solely on numerical series, which suffer from the limitations imposed by insufficient information. Recent studies have revealed that multimodal models can address the core issue by integrating textual information. However, these models focus on either historical or future textual information, overlooking the unique contributions each plays in time series forecasting. Besides, these models fail to grasp the intricate relationships between textual and time series data, constrained by their moderate capacity for multimodal comprehension. To tackle these challenges, we propose Dual-Forecaster, a pioneering multimodal time series model that combines both descriptively historical textual information and predictive textual insights, leveraging advanced multimodal comprehension capability. We begin by developing the historical text-time series contrastive loss to align the descriptively historical textual data and corresponding time series data, followed by encoding multimodal text-time series representations between them through the history-oriented modality interaction module, and then combining predictive textual data through the future-oriented modality interaction module to ensure textual insights-following forecasting. Our comprehensive evaluations on synthetic dataset and captioned-public datasets demonstrate that Dual-Forecaster is a distinctly effective multimodal time series model that outperforms or is comparable to other state-of-the-art models, highlighting the superiority of integrating textual information for time series forecasting. This work opens new avenues in the integration of textual information with numerical time series data for multimodal time series analysis.", "keywords": [ "time series forecasting", "multimodal time series model", "multimodal comprehension" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "QDkPAV9Fa5", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Jiajun Liang", "gender": "Male", "institution": "Kuaishou", "country": "CN", "position": "Researcher" }, { "name": "Linze Li", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Zhaodong Bing", "gender": "unknown", "institution": "Megvii Technology Inc.", "country": "", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 22, "n_ref_uni": 25, "n_ref": 34, "n_ref_all": 39, "n_fig": 2, "n_tab": 8, "L_tab": 2265, "n_element_tab": 255, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 2147, "n_element_tab_1": 269, "formula_len_all": 1075, "formula_len_all_1": 1014, "len_all": 108343, "len_all_1": 56769, "len_abs": 1290, "len_title": 56, "len_sents": 30927, "len_sents_1": 22482, "n_sents": 260, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 116, "L_abs": 1296, "title": "Optimizing Knowledge Distillation in Transformers: Enabling Power of Multi-Head Attention without Alignment Barriers", "abs": "Knowledge distillation has been proven effective for compressing transformer architectures by transferring knowledge from teacher to student models. Logits-based methods of knowledge distillation cannot fully capture the intermediate representations and features within the teacher model, which may result in the student model not fully learning all the knowledge from the teacher model. Thus, previous work focuses on transferring knowledge through intermediate features or attention maps. However, leveraging multi-head attention maps in transformers for knowledge distillation presents challenges due to head misalignment and suboptimal feature alignment, often requiring projectors to align features or special modifications to the model architecture. To address above limitations, we propose the Squeezing-Heads Distillation (SHD) method. This method reduces the number of attention maps to any desired number through linear approximation, without requiring additional projectors or parameters. This facilitates better alignment and knowledge transfer between models with different numbers of heads, enhancing both flexibility and efficiency. Experimental results demonstrate significant improvements in both language and vision generative models, validating the effectiveness of our method.", "keywords": [ "Knowledge distillation; Multi-head Attention; Transformers; Generative models;" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "QDNUuB5DeO", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Eng Yeow Cheu", "gender": "Male", "institution": "Sembcorp Industries Ltd", "country": "SG", "position": "Researcher" }, { "name": "Noelle Y. L. Wong", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zhonglin Chiam", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 34, "n_ref_uni": 11, "n_ref": 17, "n_ref_all": 26, "n_fig": 2, "n_tab": 4, "L_tab": 1374, "n_element_tab": 160, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4488, "formula_len_all_1": 3160, "len_all": 108860, "len_all_1": 55482, "len_abs": 1644, "len_title": 94, "len_sents": 40513, "len_sents_1": 26393, "n_sents": 261, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1817, "title": "Disentangled interleaving variational encoding", "abs": "Conflicting objectives present a considerable challenge in interleaving multi-task learning, necessitating the need for meticulous design and balance to ensure effective learning of a representative latent data space across all tasks without mutual negative impact. Drawing inspiration from the concept of marginal and conditional probability distributions in probability theory, we design a principled and well-founded approach to disentangle the original input into marginal and conditional probability distributions in the latent space of a variational autoencoder. Our proposed model, Deep Disentangled Interleaving Variational Encoding (DeepDIVE) learns disentangled features from the original input to form clusters in the embedding space and unifies these features via the cross-attention mechanism in the fusion stage. We theoretically prove that combining the objectives for reconstruction and forecasting fully captures the lower bound and mathematically derive a loss function for disentanglement using Na\u00efve Bayes. Under the assumption that the prior is a mixture of log-concave distributions, we also establish that the Kullback-Leibler divergence between the prior and the posterior is upper bounded by the cross entropy loss, informing our adoption of radial basis functions (RBF) and cross entropy with interleaving training for DeepDIVE to provide a justified basis for convergence. Experiments on anonymous bidding data from the National Electricity Market of Singapore (NEMS) show that DeepDIVE disentangles the original input and yields more accurate forecasts, outperforming current state-of-the-art baselines. In the context of the power market, this study can enhance operational decisions and bidding strategies by offering insights into the embedded supply curve via the representation space.", "keywords": [ "Representation Learning", "Disentanglement", "Probability Theory", "Mathematical Optimization", "Variational Autoencoder", "Time Series Forecasting" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "QCY1WQXTc8", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Atik Faysal", "gender": "Male", "institution": "Rowan University", "country": "US", "position": "PhD student" }, { "name": "Imad El Aaroussi", "gender": "Male", "institution": "MLNomads", "country": "MA", "position": "Researcher" }, { "name": "Taha Bouhsine", "gender": "Male", "institution": "Rowan University", "country": "US", "position": "MS student" }, { "name": "huaxia wang", "gender": "Male", "institution": "Rowan University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 4, "n_ref_uni": 27, "n_ref": 36, "n_ref_all": 47, "n_fig": 4, "n_tab": 1, "L_tab": 799, "n_element_tab": 34, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 440, "n_element_tab_1": 5, "formula_len_all": 1641, "formula_len_all_1": 169, "len_all": 114316, "len_all_1": 39544, "len_abs": 1121, "len_title": 88, "len_sents": 35004, "len_sents_1": 18547, "n_sents": 290, "n_sents_1": 117, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1134, "title": "SimO Loss: Anchor-Free Contrastive Loss for Fine-Grained Supervised Contrastive Learning", "abs": "We introduce a novel anchor-free contrastive learning (AFCL) method leveraging our proposed Similarity-Orthogonality (SimO) loss. Our approach minimizes a pseudo-metric/metric discriminative loss function that simultaneously optimizes two key objectives: reducing the distance and orthogonality between embeddings of similar inputs while maximizing these metrics for dissimilar inputs, facilitating more fine-grained contrastive learning. The AFCL method, powered by SimO loss, creates a fiber bundle topological structure in the embedding space, forming class-specific, internally cohesive yet orthogonal neighborhoods. We validate the efficacy of our method on the CIFAR-10 dataset, providing visualizations that demonstrate the impact of SimO loss on the embedding space. Our results illustrate the formation of distinct, orthogonal class neighborhoods, showcasing the method's ability to create well-structured embeddings that balance class separation with intra-class variability. This work opens new avenues for understanding and leveraging the geometric properties of learned representations in various machine learning tasks.", "keywords": [ "Contrastive Learning", "Anchor Free", "Contrastive Learning", "Explainable Contrastive Learning", "Fine Grained Representation Learning" ], "rating_list": [ 5, 3, 1, 3, 3 ], "soundness_list": [ 3, 2, 1, 2, 1 ], "presentation_list": [ 3, 3, 3, 2, 1 ], "contribution_list": [ 2, 2, 1, 2, 1 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "QCDdI7X3f9", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Carlos Guestrin", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Full Professor" }, { "name": "Irena Gao", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Percy Liang", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Associate Professor" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 34, "n_ref": 50, "n_ref_all": 71, "n_fig": 5, "n_tab": 2, "L_tab": 764, "n_element_tab": 18, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 125, "n_element_tab_1": 27, "formula_len_all": 1494, "formula_len_all_1": 1448, "len_all": 110204, "len_all_1": 58385, "len_abs": 1077, "len_title": 104, "len_sents": 31072, "len_sents_1": 27905, "n_sents": 256, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1081, "title": "Model Equality Testing: Which Model is this API Serving?", "abs": "Users often interact with large language models through black-box inference APIs, both for closed- and open-weight models (e.g., Llama models are popularly accessed via Amazon Bedrock and Azure AI Studio). In order to cut costs or add functionality, API providers may quantize, watermark, or finetune the underlying model, changing the output distribution --- often without notifying users. We formalize detecting such distortions as Model Equality Testing, a two-sample testing problem, where the user collects samples from the API and a reference distribution and conducts a statistical test to see if the two distributions are the same. We find that tests based on the Maximum Mean Discrepancy between distributions are powerful for this task: a test built on a simple string kernel achieves a median of 77.4% power against a range of distortions, using an average of just 10 samples per prompt. We then apply this test to commercial inference APIs for four Llama models, finding that 11 out of 31 endpoints serve different distributions than reference weights released by Meta.", "keywords": [ "API monitoring", "model shift", "two-sample testing" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "QC2qE1tcmd", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Diaaeldin Taha", "gender": "Male", "institution": "Max Planck Institute for Mathematics in the Sciences, Max-Planck Institute", "country": "DE", "position": "Postdoc" }, { "name": "Guido F. Montufar", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Full Professor" }, { "name": "James Chapman", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Karel Devriendt", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Marzieh Eidi", "gender": "unknown", "institution": "Universit\u00e4t Leipzig", "country": "DE", "position": "Postdoc" } ], "n_formula": 101, "n_formula_1": 30, "n_ref_uni": 62, "n_ref": 153, "n_ref_all": 186, "n_fig": 11, "n_tab": 12, "L_tab": 5243, "n_element_tab": 766, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 802, "n_element_tab_1": 37, "formula_len_all": 8212, "formula_len_all_1": 2163, "len_all": 304195, "len_all_1": 67225, "len_abs": 782, "len_title": 128, "len_sents": 80219, "len_sents_1": 29569, "n_sents": 706, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 128, "L_abs": 785, "title": "Demystifying Topological Message-Passing with Relational Structures: A Case Study on Oversquashing in Simplicial Message-Passing", "abs": "Topological deep learning (TDL) has emerged as a powerful tool for modeling higher-order interactions in relational data. However, phenomena such as oversquashing in topological message-passing remain understudied and lack theoretical analysis. We propose a unifying axiomatic framework that bridges graph and topological message-passing by viewing simplicial and cellular complexes and their message-passing schemes through the lens of relational structures. This approach extends graph-theoretic results and algorithms to higher-order structures, facilitating the analysis and mitigation of oversquashing in topological message-passing networks. Through theoretical analysis and empirical studies on simplicial networks, we demonstrate the potential of this framework to advance TDL.", "keywords": [ "topological deep learning", "oversquashing", "rewiring", "relational graph neural networks", "simplicial complexes", "relational structures" ], "rating_list": [ 6, 6, 6, 8, 8 ], "soundness_list": [ 2, 3, 3, 4, 4 ], "presentation_list": [ 2, 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "QByW8EYEtt", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Qianqi Yan", "gender": "Female", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Xiang Yue", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Xin Eric Wang", "gender": "Male", "institution": "Simular", "country": "US", "position": "Principal Researcher" }, { "name": "Xuehai He", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 26, "n_ref": 57, "n_ref_all": 75, "n_fig": 5, "n_tab": 19, "L_tab": 16479, "n_element_tab": 1544, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1461, "n_element_tab_1": 172, "formula_len_all": 29, "formula_len_all_1": 0, "len_all": 157940, "len_all_1": 55063, "len_abs": 1682, "len_title": 152, "len_sents": 34795, "len_sents_1": 24434, "n_sents": 227, "n_sents_1": 140, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1689, "title": "Worse than Random? An Embarrassingly Simple Probing Evaluation of Large Multimodal Models in Medical VQA", "abs": "Large Multimodal Models (LMMs) have shown remarkable progress in medical Visual Question Answering (Med-VQA), achieving high accuracy on existing benchmarks. However, their reliability under robust evaluation is questionable. This study reveals that state-of-the-art models perform worse than random guessing on medical diagnosis questions when subjected to simple probing evaluation. To address this critical evaluation problem, we introduce the Probing Evaluation for Medical Diagnosis (ProbMed) dataset to rigorously assess LMM performance in medical imaging through probing evaluation and procedural diagnosis. Particularly, probing evaluation features pairing original questions with negation questions with hallucinated attributes, while procedural diagnosis requires reasoning across various diagnostic dimensions for each image, including modality recognition, organ identification, clinical findings, abnormalities, and positional grounding. Our evaluation reveals that top-performing models like GPT-4o, GPT-4V, and Gemini Pro perform worse than random guessing on specialized diagnostic questions, indicating significant limitations in handling fine-grained medical inquiries. We further investigate the underperformance of open-source models (e.g., LLaVA, LLaVA-Med, and Med-Flamingo) through an ablation study. This study reveals that poor visual understanding is a primary bottleneck, which can be mitigated by adding visual descriptions generated by GPT-4o, leading to an average performance improvement of 9.44%. These findings underscore the urgent need for more robust evaluation methods and domain-specific expertise to ensure LMM reliability in critical medical fields.", "keywords": [ "Vision and Language", "AI for Healthcare", "Benchmark" ], "rating_list": [ 8, 6, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "QB8dHqVoDw", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Bin Cui", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Daniel Biggs", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Jiandong Yu", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Jiulong Shan", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Omar Hafez", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Siqi Shen", "gender": "Female", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Yu Liu", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Zhang wen tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 8, "n_ref_uni": 24, "n_ref": 60, "n_ref_all": 81, "n_fig": 10, "n_tab": 3, "L_tab": 2094, "n_element_tab": 386, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 971, "n_element_tab_1": 227, "formula_len_all": 1187, "formula_len_all_1": 574, "len_all": 121989, "len_all_1": 61442, "len_abs": 1708, "len_title": 131, "len_sents": 35240, "len_sents_1": 26993, "n_sents": 288, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1717, "title": "Transfer learning in Scalable Graph Neural Network for Improved Physical Simulation", "abs": "In recent years, Graph Neural Network (GNN) based models have shown promising results in simulating physics of complex systems. However, training dedicated graph network based physics simulators can be costly, as most models are confined to fully supervised training, which requires extensive data generated from traditional physics simulators. To date, how transfer learning could improve the model performance and training efficiency has remained unexplored. In this work, we introduce a pre-training and transfer learning paradigm for graph network simulators. We propose the scalable graph U-net (SGUNET). Incorporating an innovative depth-first search (DFS) pooling, the SGUNET is adaptable to different mesh sizes and resolutions for various simulation tasks. To enable the transfer learning between differently configured SGUNETs, we propose a set of mapping functions to align the parameters between the pre-trained model and the target model. An extra normalization term is also added into the loss to constrain the difference between the pre-trained weights and target model weights for better generalization performance. To pre-train our physics simulator we created a dataset which includes 20,000 physical simulations of randomly selected 3D shapes from the open source A Big CAD (ABC) dataset. We show that our proposed transfer learning methods allow the model to perform even better when fine-tuned with small amounts of training data than when it is trained from scratch with full extensive dataset.\nOn the 2D Deformable Plate benchmark dataset, our pre-trained model fine-tuned on 1/16 of the training data achieved an 11.05\\% improvement in position RMSE compared to the model trained from scratch.", "keywords": [ "graph neural network", "transfer learning", "physical simulations", "mesh" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 5, 4, 4 ] }, { "paper_id": "QAAsnSRwgu", "primary_area": "applications to robotics, autonomy, planning", "n_author": 11, "author_info_list": [ { "name": "Chenxin Diao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Researcher" }, { "name": "Damien Graux", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Principal Researcher" }, { "name": "Jeff Z. Pan", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Full Professor" }, { "name": "Kaustubh Vyas", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Researcher" }, { "name": "Keshuang Li", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Researcher" }, { "name": "Pavlos Vougiouklis", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Ruofei Lai", "gender": "Male", "institution": "University of Pittsburgh", "country": "", "position": "MS student" }, { "name": "Sebastien Montella", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Researcher" }, { "name": "Wendi Zhou", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Researcher" }, { "name": "Yang Ren", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Yijun YANG", "gender": "Male", "institution": "Edinburgh University, University of Edinburgh", "country": "", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 41, "n_ref": 74, "n_ref_all": 101, "n_fig": 4, "n_tab": 4, "L_tab": 1809, "n_element_tab": 123, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1136, "n_element_tab_1": 100, "formula_len_all": 84, "formula_len_all_1": 84, "len_all": 193729, "len_all_1": 66683, "len_abs": 1261, "len_title": 151, "len_sents": 48860, "len_sents_1": 31783, "n_sents": 334, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 103, "L_abs": 1119, "title": "From an LLM Swarm to a PDDL-empowered Hive: Planning Self-executed Instructions in a Multi-modal Jungle", "abs": "In response to the call for agent-based solutions that leverage the ever-increasing capabilities of the deep models' ecosystem, we introduce a comprehensive solution for selecting appropriate models and subsequently planning a set of atomic actions to satisfy the end-users' instructions.\n\nOur system, Hive, operates over sets of models and, upon receiving natural language instructions, schedules and executes, explainable plans of atomic actions. These actions can involve one or more of the available models to achieve the overall task, while respecting end-users specific constraints. Hive is able to plan complex chains of actions while guaranteeing explainability, using an LLM-based formal logic backbone empowered by PDDL operations. We introduce the MuSE benchmark in order to offer a comprehensive evaluation of the multi-modal capabilities of agent systems. Our findings show that our framework redefines the state-of-the-art for task selection, outperforming other competing systems that plan operations across multiple models while offering transparency guarantees while fully adhering to user constraints.", "keywords": [ "Deep Models", "Planning", "PDDL", "Knowledge Graphs", "Benchmark", "Large Language Models" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "Q9Z0c1Rb5i", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Ao Sun", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Pingchuan Ma", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Visitor" }, { "name": "Shuai Wang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Yuanyuan Yuan", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 24, "n_ref": 82, "n_ref_all": 113, "n_fig": 10, "n_tab": 7, "L_tab": 938, "n_element_tab": 145, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 307, "n_element_tab_1": 37, "formula_len_all": 163, "formula_len_all_1": 160, "len_all": 148381, "len_all_1": 71412, "len_abs": 1388, "len_title": 129, "len_sents": 57374, "len_sents_1": 35427, "n_sents": 462, "n_sents_1": 278, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1391, "title": "Boosting Concept Bottleneck Models with Supervised, Hierarchical Concept Learning", "abs": "Concept Bottleneck Models (CBMs) aim to deliver interpretable and interventionable predictions by bridging features and labels with human-understandable concepts. While recent CBMs show promising potential, they suffer from information leakage, where unintended information beyond the concepts (either in probabilistic or binary-state form) is leaked to the subsequent label prediction. Consequently, distinct classes are falsely classified via indistinguishable concepts, undermining the interpretation and intervention of CBMs.\n\nThis paper alleviates the information leakage issue by introducing label supervision in concept prediction and constructing a hierarchical concept set. Accordingly, we propose a new paradigm of CBMs, namely SupCBM, which stands for Structured Understanding of leakage Prevention Concept Bottleneck Model, achieving label prediction via predicted concepts and a deliberately structural-designed intervention matrix. SupCBM focuses on concepts that are mostly relevant to the predicted label and only distinguishes classes when different concepts are presented. Our evaluations show that SupCBM\u2019s label prediction outperforms SOTA CBMs over diverse datasets. Its predicted concepts also exhibit better interpretability. With proper quantification of information leakage in different CBMs, we demonstrate that SupCBM significantly reduces the information leakage.", "keywords": [ "Interpretability", "explainability", "CBM" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 1, 4, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "Q9OGPWt0Rp", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Bo Liu", "gender": "Male", "institution": "Beijing University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Tongtong Yuan", "gender": "Female", "institution": "Beijing University of Technology", "country": "CN", "position": "Lecturer" }, { "name": "Yan Li", "gender": "unknown", "institution": "Beijing University of Technology", "country": "", "position": "MS student" }, { "name": "Yu Meng", "gender": "unknown", "institution": "Beijing University of Technology", "country": "", "position": "MS student" }, { "name": "Yujia Liu", "gender": "unknown", "institution": "Beijing University of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 56, "n_formula_1": 16, "n_ref_uni": 31, "n_ref": 50, "n_ref_all": 85, "n_fig": 7, "n_tab": 18, "L_tab": 5973, "n_element_tab": 661, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 740, "n_element_tab_1": 73, "formula_len_all": 5776, "formula_len_all_1": 1428, "len_all": 173518, "len_all_1": 68969, "len_abs": 1832, "len_title": 151, "len_sents": 47415, "len_sents_1": 30145, "n_sents": 398, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1842, "title": "Connecting Solutions and Boundary Conditions/Parameters Directly: Solving PDEs in Real Time with PINNs", "abs": "Physics-Informed Neural Networks (PINNs) have proven to be important tools for solving both forward and inverse problems of partial differential equations (PDEs). However, PINNs face the retraining challenge in which neural networks need to be retrained once the parameters, or boundary/initial conditions change. To address this challenge, meta-learning PINNs train a meta-model across a range of PDE configurations, and the PINN models for new PDE configurations are then generated directly or fine-tuned from the meta-model. Meta-learning PINNs are confronted with either the issue of generalizing to significantly new PDE configurations or the time-consuming process of fine-tuning. By analyzing the mathematical structure of various PDEs, in this paper we establish the direct and mathematically sound connections between PDE solutions and boundary/initial conditions, sources and parameters. The learnable functions in these connections are trained offline in less than 1 hour in most cases. With these connections, the solutions for new PDE configurations can be obtained directly and vice versa, without retraining and fine-tuning at all. Our experimental results indicate that our methods are comparable to vanilla PINNs in terms of accuracy in forward problems, yet at least 400 times faster than them (even over 800 times faster for variable initial/source problems). In inverse problems, our methods are much more accurate than vanilla PINNs while being 80 times faster. Compared with meta-learning PINNs, our methods are much more accurate and about 20 times faster than fine-tuning. Our inference time is less than half a second in forward problems, and at most 3 seconds in inverse problems (less than half a second for variable initial/source problems of linear PDEs). Our code will be made publicly available upon acceptance.", "keywords": [ "Physics-Informed Neural Networks; Partial Differential Equations; PINNs; PDEs" ], "rating_list": [ 3, 8, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Q95MaWfF4e", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Christopher J. Anders", "gender": "unknown", "institution": "RIKEN", "country": "JP", "position": "Postdoc" }, { "name": "Frederik Pahde", "gender": "Male", "institution": "Fraunhofer HHI, Fraunhofer IAIS", "country": "", "position": "Researcher" }, { "name": "Leander Weber", "gender": "Male", "institution": "Fraunhofer HHI", "country": "", "position": "PhD student" }, { "name": "Maximilian Dreyer", "gender": "Male", "institution": "Fraunhofer HHI", "country": "", "position": "PhD student" }, { "name": "Moritz Weckbecker", "gender": "Male", "institution": "Fraunhofer HHI, Fraunhofer IAIS", "country": "DE", "position": "PhD student" }, { "name": "Sebastian Lapuschkin", "gender": "Male", "institution": "Fraunhofer HHI", "country": "", "position": "Head of Explainable Artificial Intelligence" }, { "name": "Thomas Wiegand", "gender": "Male", "institution": "Fraunhofer HHI", "country": "", "position": "Full Professor" }, { "name": "Wojciech Samek", "gender": "Male", "institution": "TU Berlin", "country": "DE", "position": "Associate Professor" } ], "n_formula": 69, "n_formula_1": 12, "n_ref_uni": 50, "n_ref": 87, "n_ref_all": 119, "n_fig": 16, "n_tab": 11, "L_tab": 5393, "n_element_tab": 988, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4508, "formula_len_all_1": 424, "len_all": 229612, "len_all_1": 64596, "len_abs": 1347, "len_title": 145, "len_sents": 68173, "len_sents_1": 30174, "n_sents": 445, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1356, "title": "Navigating Neural Space: Revisiting Concept Activation Vectors to Overcome Directional Divergence", "abs": "With a growing interest in understanding neural network prediction strategies, Concept Activation Vectors (CAVs) have emerged as a popular tool for modeling human-understandable concepts in the latent space.\nCommonly, CAVs are computed by leveraging linear classifiers optimizing the *separability* of latent representations of samples with and without a given concept. However, in this paper we show that such a separability-oriented computation leads to solutions, which may diverge from the actual goal of precisely modeling the concept direction.\nThis discrepancy can be attributed to the significant influence of distractor directions, i.e., signals unrelated to the concept, which are picked up by filters (i.e., weights) of linear models to optimize class-separability.\nTo address this, we introduce *pattern-based CAVs*, solely focussing on concept signals, thereby providing more accurate concept directions.\nWe evaluate various CAV methods in terms of their alignment with the true concept direction and their impact on CAV applications, including concept sensitivity testing and model correction for shortcut behavior caused by data artifacts. \nWe demonstrate the benefits of pattern-based CAVs using the Pediatric Bone Age, ISIC2019, and FunnyBirds datasets with VGG, ResNet, ReXNet, EfficientNet, and Vision Transformer as model architectures.", "keywords": [ "Explainable AI", "Concept-based Explanations", "Concept Activation Vectors" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "Q7uE3M5aMD", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Suhan Liu", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Tianhe Zhang", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Peng Shi", "gender": "unknown", "institution": "University of Wisconsin-Madison", "country": "US", "position": "Full Professor" } ], "n_formula": 28, "n_formula_1": 28, "n_ref_uni": 29, "n_ref": 43, "n_ref_all": 65, "n_fig": 11, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2209, "formula_len_all_1": 2006, "len_all": 111345, "len_all_1": 72633, "len_abs": 1629, "len_title": 122, "len_sents": 36006, "len_sents_1": 31923, "n_sents": 281, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1637, "title": "Discrimination-free Insurance Pricing with Privatized Sensitive Attributes", "abs": "Fairness has emerged as a critical consideration in the landscape of machine learn-ing algorithms, particularly as AI continues to transform decision-making across societal domains. To ensure that these algorithms are free from bias and do not discriminate against individuals based on sensitive attributes such as gender and race, the field of algorithmic bias has introduced various fairness concepts, along with methodologies to achieve these notions in different contexts. Despite the rapid advancement, not all sectors have embraced these fairness principles to the same extent. One specific sector that merits attention in this regard is insurance. Within the realm of insurance pricing, fairness is defined through a distinct and specialized framework. Consequently, achieving fairness according to established notions does not automatically ensure fair pricing in insurance. In particular, regulators are increasingly emphasizing transparency in pricing algorithms and imposing constraints on insurance companies on the collection and utilization of sensitive consumer attributes. These factors present additional challenges in the implementation of fairness in pricing algorithms. To address these complexities and comply with regulatory demands, we propose an efficient method for constructing fair models that are tailored to the insurance domain, using only privatized sensitive attributes. Notably, our approach ensures statistical guarantees, does not require direct access to sensitive attributes, and adapts to varying transparency requirements, addressing regulatory demands while ensuring fairness in insurance pricing.", "keywords": [ "Fairness", "Discrimination-free Insurance Pricing", "Insurance", "Privatized Attributes", "Privacy" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 4, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "Q7EjHroO1w", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Letian Chen", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Matthew C. Gombolay", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Rohan R Paleja", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Sanne van Waveren", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Varshith Sreeramdass", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 13, "n_ref_uni": 68, "n_ref": 145, "n_ref_all": 171, "n_fig": 16, "n_tab": 1, "L_tab": 102, "n_element_tab": 23, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1895, "formula_len_all_1": 1353, "len_all": 243752, "len_all_1": 69595, "len_abs": 1127, "len_title": 0, "len_sents": 57654, "len_sents_1": 32895, "n_sents": 434, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1141, "title": "Generalized Behavior Learning from Diverse Demonstrations", "abs": "Diverse behavior policies are valuable in domains requiring quick test-time adaptation or personalized human-robot interaction. Human demonstrations provide rich information regarding task objectives and factors that govern individual behavior variations, which can be used to characterize \\it{useful} diversity and learn diverse performant policies.\nHowever, we show that prior work that builds naive representations of demonstration heterogeneity fails in generating successful novel behaviors that generalize over behavior factors.\nWe propose Guided Strategy Discovery (GSD), which introduces a novel diversity formulation based on a learned task-relevance measure that prioritizes behaviors exploring modeled latent factors.\nWe empirically validate across three continuous control benchmarks for generalizing to in-distribution (interpolation) and out-of-distribution (extrapolation) factors that GSD outperforms baselines in novel behavior discovery by $\\sim$21\\%.\nFinally, we demonstrate that GSD can generalize striking behaviors for table tennis in a virtual testbed while leveraging human demonstrations collected in the real world.", "keywords": [ "Behavior Discovery", "Demonstrator Heterogeneity" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 2, 2 ] }, { "paper_id": "Q6a9W6kzv5", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Boyi Li", "gender": "Female", "institution": "NVIDIA Research", "country": "US", "position": "Researcher" }, { "name": "Jiageng Mao", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Researcher" }, { "name": "Vitor Campagnolo Guizilini", "gender": "Male", "institution": "Toyota Research Institute", "country": "US", "position": "Staff Research Scientist" }, { "name": "Yue Wang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Daniel Takeshi Seita", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "Assistant Professor" }, { "name": "Wei Chou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 2, "n_ref_uni": 109, "n_ref": 312, "n_ref_all": 402, "n_fig": 33, "n_tab": 21, "L_tab": 20181, "n_element_tab": 1174, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 800, "n_element_tab_1": 101, "formula_len_all": 671, "formula_len_all_1": 696, "len_all": 835213, "len_all_1": 61996, "len_abs": 2751, "len_title": 141, "len_sents": 256739, "len_sents_1": 29132, "n_sents": 2426, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (conditional oral)", "L_title": 93, "L_abs": 1898, "title": "PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding", "abs": "Understanding the physical world is a fundamental challenge in embodied AI, critical for enabling agents to perform complex tasks and operate safely in real-world environments. While Vision-Language Models (VLMs) have shown great promise in reasoning and task planning for embodied agents, their ability to comprehend physical phenomena remains extremely limited.\nTo close this gap, we introduce PhysBench, a comprehensive benchmark designed to evaluate VLMs' physical world understanding capability across a diverse set of tasks. \nPhysBench contains 100,000 entries of interleaved video-image-text data, categorized into four major domains: physical object properties, physical object relationships, physical scene understanding, and physics-based dynamics, further divided into 19 subclasses and 8 distinct capability dimensions.\nOur extensive experiments, conducted on 39 representative VLMs, reveal that while these models excel in common-sense reasoning, they struggle with understanding the physical world---likely due to the absence of physical knowledge in their training data and the lack of embedded physical priors.\nTo tackle the shortfall, we introduce PhysAgent, a novel framework that combines the generalization strengths of VLMs with the specialized expertise of vision models, significantly enhancing VLMs' physical understanding across a variety of tasks, including an 18.4\\% improvement on GPT-4o.\nFurthermore, our results demonstrate that enhancing VLMs\u2019 physical world understanding capabilities can significantly help the deployment of embodied agents, pushing the boundaries of machine intelligence in comprehending and interacting with the physical world. We believe that PhysBench and PhysAgent offer valuable insights and contribute to bridging the gap between VLMs and physical world understanding. Project site: [https://anoy1314.github.io/](https://anoy1314.github.io/)", "keywords": [ "vision-language", "multi-modal understanding" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "Q6PAnqYVpo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chihiro Taguchi", "gender": "Male", "institution": "University of Notre Dame", "country": "", "position": "PhD student" }, { "name": "Go Kamoda", "gender": "Male", "institution": "Tohoku University", "country": "", "position": "MS student" }, { "name": "Hiroyuki Deguchi", "gender": "Male", "institution": "NTT Communications", "country": "JP", "position": "Researcher" }, { "name": "Kohei Suenaga", "gender": "Male", "institution": "Kyoto University, Kyoto University", "country": "JP", "position": "Associate Professor" }, { "name": "Masaki Waga", "gender": "Male", "institution": "Kyoto University, Tokyo Institute of Technology", "country": "JP", "position": "Assistant Professor" }, { "name": "Sho Yokoi", "gender": "unknown", "institution": "NINJAL", "country": "JP", "position": "Associate Professor" }, { "name": "Yusuke Matsushita", "gender": "Male", "institution": "Kyoto University", "country": "JP", "position": "Postdoc" } ], "n_formula": 4, "n_formula_1": 7, "n_ref_uni": 87, "n_ref": 129, "n_ref_all": 151, "n_fig": 6, "n_tab": 3, "L_tab": 864, "n_element_tab": 62, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1113, "n_element_tab_1": 109, "formula_len_all": 552, "formula_len_all_1": 880, "len_all": 262842, "len_all_1": 71870, "len_abs": 1599, "len_title": 113, "len_sents": 39792, "len_sents_1": 32920, "n_sents": 284, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1606, "title": "A Soft and Fast Pattern Matcher for Billion-Scale Corpus Searches", "abs": "Researchers and practitioners in natural language processing and computational linguistics frequently observe and analyze the real language usage in large-scale corpora.\nFor that purpose, they often employ off-the-shelf pattern-matching tools, such as grep, and keyword-in-context concordancers, which is widely used in corpus linguistics for gathering examples.\nNonetheless, these existing techniques rely on surface-level string matching, and thus they suffer from the major limitation of not being able to handle orthographic variations and paraphrasing---notable and common phenomena in any natural language.\nIn addition, existing continuous approaches such as dense vector search tend to be overly coarse, often retrieving texts that are unrelated but share similar topics.\nGiven these challenges, we propose a novel algorithm that achieves soft (or semantic) yet efficient pattern matching by relaxing a surface-level matching with word embeddings.\nOur algorithm is highly scalable with respect to the size of the corpus text utilizing inverted indexes.\nWe have prepared an efficient implementation, and we provide an accessible web tool.\nOur experiments demonstrate that the proposed method\n(i) can execute searches on billion-scale corpora in less than a second, which is comparable in speed to surface-level string matching and dense vector search;\n(ii) can extract harmful instances that semantically match queries from a large set of English and Japanese Wikipedia articles;\nand (iii) can be effectively applied to corpus-linguistic analyses of Latin, a language with highly diverse inflections.", "keywords": [ "natural language processing", "full-text search", "word embeddings", "inverted index", "pattern match" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 4, 3, 2 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 5, 3, 3 ] }, { "paper_id": "Q6M7bZIo9t", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "Jingyu Liu", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "PhD student" }, { "name": "Linjiaen", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 36, "n_formula_1": 8, "n_ref_uni": 38, "n_ref": 81, "n_ref_all": 93, "n_fig": 4, "n_tab": 3, "L_tab": 384, "n_element_tab": 70, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 188, "n_element_tab_1": 32, "formula_len_all": 4609, "formula_len_all_1": 401, "len_all": 178635, "len_all_1": 71339, "len_abs": 1333, "len_title": 91, "len_sents": 59182, "len_sents_1": 35810, "n_sents": 450, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1338, "title": "How Much Can RAG Help the Reasoning of LLM?", "abs": "Retrieval-Augmented Generation (RAG) has gained significant popularity in modern Large Language Models (LLMs) due to its effectiveness in introducing new knowledge and reducing hallucinations. However, the deep understanding of RAG remains limited, how does RAG help the reasoning process and can RAG help improve the reasoning capability remains question. While external documents are typically considered as a method to incorporate domain-specific information, they also contain intermediate reasoning results related to the query, this suggests that documents could enhance the reasoning capability of LLMs, which has not been previously explored. In this paper, we investigate this issue in depth and find that while RAG can assist with reasoning, the help is limited. If we conceptualize the reasoning process as a tree with fixed depth, then RAG struggles to assist LLMs in performing deeper reasoning. Additionally, the information in the documents requires preprocessing to filter out noise. We demonstrate that this preprocessing is difficult to achieve simply fine-tuning of the LLM, it often necessitates numerous additional transformer layers to solve the problem. To simplify the problem, we propose DPrompt tuning, which effectively resolves the issue within just limited transformer layers, leading to improved performance.", "keywords": [ "RAG", "Reasoning", "LLM" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 4, 2, 3 ], "confidence_list": [ 4, 4, 2 ] }, { "paper_id": "Q5eo3VMxF6", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Fan Lin", "gender": "Male", "institution": "Southeast University", "country": "", "position": "MS student" }, { "name": "Linlin Song", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Lv Qingsong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Shupei Xiao", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Shuyi Xie", "gender": "Male", "institution": "Tencent Big Data", "country": "CN", "position": "Researcher" }, { "name": "Sui Wenjuan", "gender": "Female", "institution": "Guangdong University of Foreign Studies", "country": "CN", "position": "Assistant Professor" }, { "name": "Zishan Xu", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 79, "n_ref_all": 105, "n_fig": 9, "n_tab": 8, "L_tab": 2419, "n_element_tab": 477, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 2082, "n_element_tab_1": 436, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 162614, "len_all_1": 70696, "len_abs": 1915, "len_title": 127, "len_sents": 43095, "len_sents_1": 28538, "n_sents": 361, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1342, "title": "MisAttributionLLM: Integrating Error Attribution Capability into LLM Evaluation", "abs": "With the widespread application of Large Language Models (LLMs) in various tasks, evaluating the performance of LLMs becomes an essential research topic. However, existing judge models lack the specific capability required for error attribution (i.e., identify the types of error made in responses). In this work, we first establish a comprehensive Misattribution Framework with 9 primary and 19 secondary categories, which are intended to facilitate in-depth analysis and enhance the performance of LLMs. Based on this framework, we present AttriData, a dataset specifically designed for error attribution, encompassing misattributions, along with the corresponding scores and feedback. We also propose MisAttributionLLM, a fine-tuned model on AttriData, which is the first open-source, general-purpose judge model with error attribution capability which provides valuable insights into the model\u2019s weaknesses and enables targeted improvements. Experimental results show that MisAttributionLLM achieves the highest Pearson correlation with human evaluators among 8 open-source and closed-source LLMs. Furthermore, MisAttributionLLM also obtains the highest accuracy and micro-F1 in the performance of error attribution. Extensive experiments and analyses are conducted to confirm the effectiveness and robustness of our proposed method.", "keywords": [ "evaluation", "error attribution", "large language models", "llm-as-a-judge" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "Q5VlpYRxGF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Boxiao Jin", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Minjia Zhang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Yuming Chang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zheng Wang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Zhongzhi Yu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 24, "n_ref_uni": 21, "n_ref": 39, "n_ref_all": 60, "n_fig": 8, "n_tab": 5, "L_tab": 2579, "n_element_tab": 200, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 1673, "n_element_tab_1": 20, "formula_len_all": 1774, "formula_len_all_1": 1415, "len_all": 119369, "len_all_1": 65397, "len_abs": 1365, "len_title": 136, "len_sents": 35398, "len_sents_1": 29304, "n_sents": 255, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1873, "title": "Model Tells You Where to Merge: Adaptive KV Cache Merging for LLMs on Long-Context Tasks", "abs": "How to efficiently serve Large Language Models (LLMs) has become a pressing issue because of their huge computational cost in their autoregressive generation process. To mitigate computational costs, LLMs often employ the KV Cache technique to improve the generation speed. While improving the computational efficiency, the storage requirements of the KV cache are substantial, particularly in long-context scenarios, leading to significant memory consumption. Existing KV cache eviction methods often degrade the performance of LLMs in long-context scenarios due to the information loss introduced by eviction. In this paper, we propose a novel KV cache merging approach, called KVMerger, to achieve adaptive KV cache compression for long-context tasks without significant performance degradation under constrained memory budgets. Our approach is inspired by the intriguing observation that key states exhibit high similarity at the token level within a single sequence. To facilitate merging, we develop an effective yet straightforward merging set identification algorithm to identify suitable KV states for merging. Our merging set identification algorithm stimulates the second observation that KV cache sparsity, from similarity perspective, is independent of the dataset and remains persistent at the model level. Subsequently, we propose a Gaussian kernel weighted merging algorithm to selectively merge all states within each merging set. We conduct extensive experiments to demonstrate the effectiveness of KVMerger for long-context tasks under constrained memory budgets, applying it to models including Llama2-7B/13B-chat and Mistral-7B-instruct across various tasks. We also compare our method with other KV cache compression algorithms, including H2O and CaM, showing that our method achieves superior performance across tasks with different KV cache budgets.", "keywords": [ "KV cache compression", "LLM", "Gaussian Kernel weighted merging" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "Q5Sawm0nqo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Ali Moin", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Behrooz Mahasseni", "gender": "unknown", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Christopher Michael Sandino", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Ellen L. Zippi", "gender": "Female", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Erdrin Azemi", "gender": "Female", "institution": "Apple", "country": "", "position": "Principal Researcher" }, { "name": "Gaurav Patel", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Juri Minxha", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 97, "n_formula_1": 13, "n_ref_uni": 83, "n_ref": 247, "n_ref_all": 314, "n_fig": 23, "n_tab": 12, "L_tab": 16659, "n_element_tab": 1132, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6608, "formula_len_all_1": 1018, "len_all": 301708, "len_all_1": 62735, "len_abs": 1570, "len_title": 131, "len_sents": 81593, "len_sents_1": 29959, "n_sents": 626, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1325, "title": "Efficient Source-Free Time-Series Adaptation via Parameter Subspace Disentanglement", "abs": "In this paper, we propose a framework for efficient Source-Free Domain Adaptation (SFDA) in the context of time-series, focusing on enhancing both parameter efficiency and data-sample utilization. Our approach introduces an improved paradigm for source-model preparation and target-side adaptation, aiming to enhance training efficiency during target adaptation. Specifically, we reparameterize the source model's weights in a Tucker-style decomposed manner, factorizing the model into a compact form during the source model preparation phase. During target-side adaptation, only a subset of these decomposed factors is fine-tuned, leading to significant improvements in training efficiency. We demonstrate using PAC Bayesian analysis that this selective fine-tuning strategy implicitly regularizes the adaptation process by constraining the model's learning capacity. Furthermore, this re-parameterization reduces the overall model size and enhances inference efficiency, making the approach particularly well suited for resource-constrained devices. Additionally, we demonstrate that our framework is compatible with various SFDA methods and achieves significant computational efficiency, reducing the number of fine-tuned parameters and inference overhead in terms of MACs by over 90\\% while maintaining model performance.", "keywords": [ "Time Series", "Source-Free Domain Adaptation", "Efficiency" ], "rating_list": [ 5, 6, 5, 5, 10 ], "soundness_list": [ 3, 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 2, 4 ], "contribution_list": [ 1, 3, 2, 2, 4 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "Q5CLpqbrFM", "primary_area": "causal reasoning", "n_author": 6, "author_info_list": [ { "name": "Dennis Frauen", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Jonas Schweisthal", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Konstantin Hess", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Maresa Schr\u00f6der", "gender": "unknown", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Niki Kilbertus", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Assistant Professor" }, { "name": "Stefan Feuerriegel", "gender": "Male", "institution": "LMU Munich", "country": "DE", "position": "Professor" } ], "n_formula": 44, "n_formula_1": 22, "n_ref_uni": 43, "n_ref": 82, "n_ref_all": 134, "n_fig": 7, "n_tab": 7, "L_tab": 911, "n_element_tab": 81, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 694, "n_element_tab_1": 47, "formula_len_all": 4263, "formula_len_all_1": 2221, "len_all": 178066, "len_all_1": 72303, "len_abs": 1849, "len_title": 87, "len_sents": 62378, "len_sents_1": 32556, "n_sents": 499, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1365, "title": "Learning Representations of Instruments for Partial Identification of Treatment Effects", "abs": "Reliable estimation of treatment effects from observational data is important in many disciplines, such as medicine. However, estimation is challenging when unconfoundedness as a standard assumption in the causal inference literature is violated. In this work, we leverage arbitrary (potentially high-dimensional) instruments to estimate bounds on the conditional average treatment effect (CATE). Our contributions are three-fold: (1) We propose a novel approach for partial identification through a mapping of instruments to a discrete representation space so that we yield valid bounds on the CATE. This is crucial for reliable decision-making in real-world applications. (2) We derive a two-step method that learns tight bounds using a tailored neural partitioning of the latent instrument space. As a result, we avoid instability issues due to numerical approximations or adversarial training. Furthermore, our procedure aims to reduce the estimation variance in finite-sample settings to yield more reliable estimates. (3) We show theoretically that our procedure obtains valid bounds while reducing estimation variance and we perform experiments to demonstrate the effectiveness across various settings. Overall, our procedure offers a novel path for practitioners to make use of potentially high-dimensional instruments (e.g., as in Mendelian randomization).", "keywords": [ "causal inference", "partial identification", "instrumental variables", "treatment effect", "representation learning" ], "rating_list": [ 3, 5, 5, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "Q4s7nFoowt", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Jiaming Han", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Kaixiong Gong", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Wei-Hong Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Xiangyu Yue", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Yiyuan Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yilei JIANG", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 2, "n_ref": 74, "n_ref_all": 88, "n_fig": 4, "n_tab": 14, "L_tab": 2592, "n_element_tab": 296, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 1921, "n_element_tab_1": 236, "formula_len_all": 1220, "formula_len_all_1": 1192, "len_all": 196045, "len_all_1": 66524, "len_abs": 1239, "len_title": 107, "len_sents": 34440, "len_sents_1": 27312, "n_sents": 292, "n_sents_1": 235, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1254, "title": "Towards Continual Domain Adaptation of Vision-language Models", "abs": "Large-scale vision-language models have achieved remarkable performance on various downstream tasks. {Nevertheless, how to efficiently adapt vision-language models to new data distributions without re-training, \\ie, domain incremental learning (DIL) of vision-language models, is still under-explored. Existing DIL methods for single modality are either not applicable to multi-modal settings or need exemplar buffers to store previous samples to avoid catastrophic forgetting, which is not memory-efficient.} To address these limitations, we propose an exemplar-free paradigm to improve DIL of vision-language models based on prompt-tuning. We theoretically analyze and decompose the problem into two optimization objectives. Guided by the theoretical insights, we propose a novel framework named {M}ultimodal {C}ontinual {D}omain {A}daptation (MCDA), which incorporates two strategies: Multimodal Domain Alignment (MDA) and Maximum Softmax Gating (MSG). MDA enhances cross-domain performance by aligning visual and language representation spaces, while MSG improves the accuracy of domain identification by gating through Softmax probability. Extensivev experimental results demonstrate that our method outperforms current state-of-the-art approaches.", "keywords": [ "continual learning", "domain adaptation", "vision-language models" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "Q47jVPzJ3G", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Emanuele La Malfa", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Guohao Li", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Jesse Wright", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Michael J. Wooldridge", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Nigel Shadbolt", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "Full Professor" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Samuele Marro", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 42, "n_ref": 64, "n_ref_all": 82, "n_fig": 7, "n_tab": 1, "L_tab": 106, "n_element_tab": 10, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 159646, "len_all_1": 60204, "len_abs": 2702, "len_title": 119, "len_sents": 47731, "len_sents_1": 30431, "n_sents": 372, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 958, "title": "A Scalable Communication Protocol for Networks of Large Language Models", "abs": "Communication is a prerequisite for collaboration.\nWhen scaling networks of AI-powered agents, communication must be versatile, efficient, and portable.\nThese requisites, which we refer to as the Agent Communication Trilemma, are hard to achieve in large networks of agents.\nWe introduce Agora, a meta protocol that leverages existing communication standards to make LLM-powered agents solve complex problems efficiently.\nIn Agora, agents typically use standardised routines for frequent communications, natural language for rare communications, and LLM-written routines for everything in between. \nAgora sidesteps the Agent Communication Trilemma and robustly handles changes in interfaces and members, allowing unprecedented scalability with full decentralisation and minimal involvement of human beings. \nOn large Agora networks, we observe the emergence of self-organising, fully automated protocols that achieve complex goals without human intervention.", "keywords": [ "llm", "multi-agent systems", "protocol", "negotiation", "collaboration" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 4, 2 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "Q3oAX9HoH2", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Jiangchao Yao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jianing Zhu", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "PhD student" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Xuan Li", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "CN", "position": "PhD student" }, { "name": "Zhanke Zhou", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Visiting student" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 64, "n_ref": 164, "n_ref_all": 224, "n_fig": 26, "n_tab": 24, "L_tab": 14541, "n_element_tab": 407, "n_fig_1": 12, "n_tab_1": 10, "L_tab_1": 2525, "n_element_tab_1": 267, "formula_len_all": 275, "formula_len_all_1": 253, "len_all": 536038, "len_all_1": 76821, "len_abs": 1652, "len_title": 93, "len_sents": 206206, "len_sents_1": 31476, "n_sents": 1767, "n_sents_1": 278, "venue": "ICLR2025", "decision": "withdraw", "L_title": 45, "L_abs": 970, "title": "Nested Gloss Makes Large Language Models Lost", "abs": "Large language models (LLMs) have succeeded significantly in various applications but remain susceptible to adversarial jailbreaks that void their safety guardrails. \nPrevious attempts to exploit these vulnerabilities often rely on high-cost computational extrapolations, which may not be practical or efficient. \nIn this paper, inspired by the authority influence demonstrated in the Milgram experiment, we present a lightweight method to take advantage of the LLMs' personification capabilities to construct $\\textit{a virtual, nested scene}$, allowing it to realize an adaptive way to escape the usage control in a normal scenario.\nEmpirically, the contents induced by our approach can achieve leading harmfulness rates with previous counterparts and realize a continuous jailbreak in subsequent interactions, which reveals the critical weakness of self-losing on both open-source and closed-source LLMs, $\\textit{e.g.}$, Llama-2, Llama-3, GPT-3.5, GPT-4, and GPT-4o.", "keywords": [ "Trustworthy Machine Learning; Large Language Model" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "Q3RoP5IhHy", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Amir Feder", "gender": "unknown", "institution": "Columbia University", "country": "", "position": "Postdoc" }, { "name": "David Blei", "gender": "Male", "institution": "Columbia University", "country": "", "position": "Full Professor" }, { "name": "Dominic Sobhani", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Undergrad student" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 52, "n_ref": 82, "n_ref_all": 109, "n_fig": 2, "n_tab": 10, "L_tab": 3441, "n_element_tab": 288, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 575, "n_element_tab_1": 65, "formula_len_all": 895, "formula_len_all_1": 639, "len_all": 155160, "len_all_1": 64161, "len_abs": 976, "len_title": 78, "len_sents": 46070, "len_sents_1": 28930, "n_sents": 376, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 30, "L_abs": 979, "title": "Multi-environment Topic Models", "abs": "Probabilistic topic models are a powerful tool for extracting latent themes from large text datasets. In many text datasets, we also observe per-document covariates (e.g., source, style, political affiliation) that act as environments that modulate a \"global\" (environment-agnostic) topic representation. Accurately learning these representations is important for prediction on new documents in unseen environments and for estimating the causal effect of topics on real-world outcomes. To this end, we introduce the Multi-environment Topic Model (MTM), an unsupervised probabilistic model that separates global and environment-specific terms. Through experimentation on various political content, from ads to tweets and speeches, we show that the MTM produces interpretable global topics with distinct environment-specific words. On multi-environment data, the MTM outperforms strong baselines in and out-of-distribution. It also enables the discovery of accurate causal effects.", "keywords": [ "Topic Models", "ML for Social Science", "Out-of-distribution Generalization", "Multi-environment learning" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "Q2hkp8WIDS", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Adam Jelley", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Amos J Storkey", "gender": "Not Specified", "institution": "University of Edinburgh", "country": "", "position": "Full Professor" }, { "name": "Trevor McInroe", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Weipu Zhang", "gender": "Male", "institution": "University of Edinburgh", "country": "GB", "position": "MS student" } ], "n_formula": 23, "n_formula_1": 8, "n_ref_uni": 61, "n_ref": 140, "n_ref_all": 169, "n_fig": 14, "n_tab": 3, "L_tab": 1807, "n_element_tab": 125, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 468, "n_element_tab_1": 35, "formula_len_all": 1503, "formula_len_all_1": 521, "len_all": 216009, "len_all_1": 62624, "len_abs": 4493, "len_title": 155, "len_sents": 56957, "len_sents_1": 29336, "n_sents": 458, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1675, "title": "Objects matter: object-centric world models improve reinforcement learning in visually complex environments", "abs": "Deep reinforcement learning has achieved remarkable success in learning control policies from pixels across a wide range of tasks, yet its application remains hindered by low sample efficiency, requiring significantly more environment interactions than humans to reach comparable performance.\nModel-based reinforcement learning (MBRL) offers a solution by leveraging learnt world models to generate simulated experience, thereby improving sample efficiency.\nHowever, in visually complex environments, small or dynamic elements can be critical for decision-making.\nYet, traditional MBRL methods in pixel-based environments typically rely on auto-encoding with an $L_2$ loss, which is dominated by large areas and often fails to capture decision-relevant details.\nTo address these limitations, we propose an **object-centric MBRL pipeline**, which integrates recent advances in computer vision to allow agents to focus on key decision-related elements.\nOur approach consists of four main steps: (1) annotating key objects related to rewards and goals with segmentation masks, (2) extracting object features using a pre-trained, frozen foundation vision model, (3) incorporating these object features with the raw observations to predict environmental dynamics, and (4) training the policy using imagined trajectories generated by this object-centric world model.\nBuilding on the efficient MBRL algorithm STORM, we call this pipeline **OC-STORM**.\nWe demonstrate OC-STORM's practical value in overcoming the limitations of conventional MBRL approaches on both Atari games and the visually complex game Hollow Knight.\nCode and videos are available in the supplementary materials.", "keywords": [ "reinforcement learning", "model-based RL", "object-centric RL", "video object segmentation", "Atari", "Hollow Knight" ], "rating_list": [ 6, 8, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "Q2bJ2qgcP1", "primary_area": "causal reasoning", "n_author": 2, "author_info_list": [ { "name": "Haining Yu", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Yizhou Sun", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" } ], "n_formula": 55, "n_formula_1": 24, "n_ref_uni": 35, "n_ref": 85, "n_ref_all": 108, "n_fig": 4, "n_tab": 7, "L_tab": 14138, "n_element_tab": 240, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1829, "n_element_tab_1": 116, "formula_len_all": 6774, "formula_len_all_1": 2316, "len_all": 180947, "len_all_1": 75765, "len_abs": 1840, "len_title": 147, "len_sents": 53308, "len_sents_1": 32927, "n_sents": 498, "n_sents_1": 301, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 111, "L_abs": 1785, "title": "Do Contemporary Causal Inference Models Capture Real-World Heterogeneity? Findings from a Large-Scale Benchmark", "abs": "We present unexpected findings from a large-scale benchmark study evaluating Conditional Average Treatment Effect (CATE) estimation algorithms. By running 16 modern CATE models across 43,200 datasets, we find that: (a) 62\\% of CATE estimates have a higher Mean Squared Error (MSE) than a trivial zero-effect predictor, rendering them ineffective; (b) in datasets with at least one useful CATE estimate, \n80\\% still have higher MSE than a constant-effect model; and (c) Orthogonality-based models outperform other models only 30\\% of the time, despite widespread optimism about their performance. These findings expose significant limitations in current CATE models and suggest ample opportunities for further research.\n\nOur findings stem from a novel application of \\textit{observational sampling}, originally developed to evaluate Average Treatment Effect (ATE) estimates from observational methods with experiment data. To adapt observational sampling for CATE evaluation, we introduce a statistical parameter, $Q$, equal to MSE minus a constant and preserves the ranking of models by their MSE. We then derive a family of sample statistics, collectively called $\\hat{Q}$, that can be computed from real-world data. We prove that $\\hat{Q}$ is a consistent estimator of $Q$ under mild technical conditions. When used in observational sampling, $\\hat{Q}$ is unbiased and asymptotically selects the model with the smallest MSE. To ensure the benchmark reflects real-world heterogeneity, we handpick datasets where outcomes come from field rather than simulation. By combining the new observational sampling method, new statistics, and real-world datasets, the benchmark provides a unique perspective on CATE estimator performance and uncover gaps in capturing real-world heterogeneity.", "keywords": [ "causal inference" ], "rating_list": [ 6, 8, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 2, 3, 3 ] }, { "paper_id": "Q2Q4SyZ2a9", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Jingyi Zhang", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Researcher" }, { "name": "Jun Zhu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Liyuan Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Qingyi Pan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" } ], "n_formula": 23, "n_formula_1": 21, "n_ref_uni": 32, "n_ref": 80, "n_ref_all": 100, "n_fig": 3, "n_tab": 6, "L_tab": 3562, "n_element_tab": 318, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3770, "n_element_tab_1": 322, "formula_len_all": 1692, "formula_len_all_1": 1415, "len_all": 114192, "len_all_1": 66847, "len_abs": 1625, "len_title": 126, "len_sents": 32632, "len_sents_1": 26874, "n_sents": 217, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1523, "title": "LUNCH: Adaptive Balancing of Continual Learning via Hyperparameter Uncertainty", "abs": "Continual learning (CL) is characterized by learning sequentially arriving tasks and behaving as if they were observed simultaneously. In order to prevent catastrophic forgetting of old tasks when learning new tasks, representative CL methods usually employ additional loss terms to balance their contributions (e.g., regularization and replay), modulated by deterministic hyperparameters. However, this strategy struggles to accommodate real-time changes in data distributions and is also lack of robustness to subsequent unseen tasks, especially in online scenarios where CL is performed with a one-pass data stream. Inspired by adaptive weighting in multi-task learning, we propose an innovative approach named Learning UNCertain Hyperparameters (LUNCH) for adaptive balancing of task contributions in CL. Specifically, we formulate each CL-relevant hyperparameter as a function of optimizable uncertainty under homoscedastic assumption and ensure its training stability through the exponential moving average of network parameters. We further devise an evaluation protocol that moderately adjusts the hyperparameter values and reports their impact on performance, so as to analyze the sensitivity of these sub-optimal values in realistic applications. We perform extensive experiments to demonstrate the effectiveness and robustness of our approach, which significantly improves online CL in a plug-in manner (e.g., up to 11.26% and 5.64% on Split CIFAR-100 and Split Mini-ImageNet, respectively) as well as offline CL.", "keywords": [ "Life-long Learning; Uncertainty; Hyperparameter Sensitivity" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "Q1kPHLUbhi", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Alexandre Alahi", "gender": "Male", "institution": "EPFL", "country": "CH", "position": "Associate Professor" }, { "name": "Aziz Shameem", "gender": "Male", "institution": "Indian Institute of Technology, Bombay", "country": "IN", "position": "Undergrad student" }, { "name": "Mathieu Salzmann", "gender": "Male", "institution": "Swiss Data Science Center", "country": "CH", "position": "Principal Researcher" }, { "name": "Megh Shukla", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 34, "n_ref": 84, "n_ref_all": 111, "n_fig": 13, "n_tab": 3, "L_tab": 3053, "n_element_tab": 442, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 2540, "n_element_tab_1": 357, "formula_len_all": 1837, "formula_len_all_1": 944, "len_all": 146275, "len_all_1": 67871, "len_abs": 1416, "len_title": 129, "len_sents": 47126, "len_sents_1": 28604, "n_sents": 401, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1424, "title": "Towards Self-Supervised Covariance Estimation in Deep Heteroscedastic Regression", "abs": "Deep heteroscedastic regression models the mean and covariance of the target distribution through neural networks. The challenge arises from heteroscedasticity, which implies that the covariance is sample dependent and is often unknown. Consequently, recent methods learn the covariance through unsupervised frameworks, which unfortunately yield a trade-off between computational complexity and accuracy. While this trade-off could be alleviated through supervision, obtaining labels for the covariance is non-trivial.\nHere, we study self-supervised covariance estimation in deep heteroscedastic regression. We address two questions: (1) How should we supervise the covariance assuming ground truth is available? (2) How can we obtain pseudo labels in the absence of the ground-truth? We address (1) by analysing two popular measures: the KL Divergence and the 2-Wasserstein distance. Subsequently, we derive an upper bound on the 2-Wasserstein distance between normal distributions with non-commutative covariances that is stable to optimize. We address (2) through a simple neighborhood based heuristic algorithm which results in surprisingly effective pseudo labels for the covariance. Our experiments over a wide range of synthetic and real datasets demonstrate that the proposed 2-Wasserstein bound coupled with pseudo label annotations results in a computationally cheaper yet accurate deep heteroscedastic regression.", "keywords": [ "deep regression", "heteroscedastic", "uncertainty", "2-Wasserstein", "KL-Divergence", "Negative Log-Likelihood" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "Q1QTxFm0Is", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Denis Blessing", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "", "position": "PhD student" }, { "name": "Julius Berner", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Lorenz Richter", "gender": "unknown", "institution": "Zuse Institute Berlin", "country": "DE", "position": "Postdoc" }, { "name": "Gerhard Neumann", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 83, "n_formula_1": 28, "n_ref_uni": 55, "n_ref": 140, "n_ref_all": 197, "n_fig": 10, "n_tab": 2, "L_tab": 32, "n_element_tab": 2, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 964, "n_element_tab_1": 53, "formula_len_all": 9593, "formula_len_all_1": 2771, "len_all": 206713, "len_all_1": 72775, "len_abs": 1121, "len_title": 107, "len_sents": 58587, "len_sents_1": 29302, "n_sents": 525, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1133, "title": "Underdamped Diffusion Bridges with Applications to Sampling", "abs": "We provide a general framework for learning diffusion bridges that transport prior to target distributions. It includes existing diffusion models for generative modeling, but also underdamped versions with degenerate diffusion matrices, where the noise only acts in certain dimensions. Extending previous findings, our framework allows to rigorously show that score-matching in the underdamped case is indeed equivalent to maximizing a lower bound on the likelihood. Motivated by superior convergence properties and compatibility with sophisticated numerical integration schemes of underdamped stochastic processes, we propose *underdamped diffusion bridges*, where a general density evolution is learned rather than prescribed by a fixed noising process. We apply our method to the challenging task of sampling from unnormalized densities without access to samples from the target distribution. Across a diverse range of sampling problems, our approach demonstrates state-of-the-art performance, notably outperforming alternative methods, while requiring significantly fewer discretization steps and almost no hyperparameter tuning.", "keywords": [ "Variational Inference", "Sampling", "Diffusion Models" ], "rating_list": [ 8, 6, 8, 6, 6 ], "soundness_list": [ 4, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 2, 4, 4, 3, 3 ] }, { "paper_id": "Q1MHvGmhyT", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Chao Du", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Kejiang Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Research Associate" }, { "name": "Min Lin", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Principal Researcher" }, { "name": "Tianyu Pang", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Weiming Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xiaojian Yuan", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 11, "n_ref_uni": 49, "n_ref": 162, "n_ref_all": 215, "n_fig": 12, "n_tab": 11, "L_tab": 10002, "n_element_tab": 612, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1658, "formula_len_all_1": 805, "len_all": 217731, "len_all_1": 71464, "len_abs": 1319, "len_title": 109, "len_sents": 58059, "len_sents_1": 34478, "n_sents": 457, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1277, "title": "A Closer Look at Machine Unlearning for Large Language Models", "abs": "Large language models (LLMs) may memorize sensitive or copyrighted content, raising privacy and legal concerns. Due to the high cost of retraining from scratch, researchers attempt to employ machine unlearning to remove specific content from LLMs while preserving the overall performance. In this paper, we discuss several issues in machine unlearning for LLMs and provide our insights on possible approaches. To address the issue of inadequate evaluation of model outputs after unlearning, we introduce three additional metrics to evaluate token diversity, sentence semantics, and factual correctness. We then categorize unlearning methods into untargeted and targeted, and discuss their issues respectively. Specifically, the behavior that untargeted unlearning attempts to approximate is unpredictable and may involve hallucinations, and existing regularization is insufficient for targeted unlearning. To alleviate these issues, we propose using the objective of maximizing entropy (ME) for untargeted unlearning and incorporate answer preservation (AP) loss as regularization for targeted unlearning. Experimental results across three scenarios, i.e., fictitious unlearning, continual unlearning, and real-world unlearning, demonstrate the effectiveness of our approaches.", "keywords": [ "Machine Unlearning", "Large Language Models" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "Q1Hr9dVfDS", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Alexander Gepperth", "gender": "Male", "institution": "HAW Fulda", "country": "DE", "position": "Full Professor" }, { "name": "Alexander Krawczyk", "gender": "Male", "institution": "Hochschule Fulda", "country": "", "position": "PhD student" }, { "name": "Benedikt Bagus", "gender": "Male", "institution": "University of Applied Sciences Fulda", "country": "DE", "position": "PhD student" }, { "name": "Yannick Denker", "gender": "Male", "institution": "Hochschule Fulda", "country": "DE", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 18, "n_ref": 28, "n_ref_all": 49, "n_fig": 7, "n_tab": 3, "L_tab": 2178, "n_element_tab": 408, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 541, "n_element_tab_1": 79, "formula_len_all": 43, "formula_len_all_1": 53, "len_all": 111084, "len_all_1": 56487, "len_abs": 1119, "len_title": 133, "len_sents": 31280, "len_sents_1": 27410, "n_sents": 217, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1149, "title": "Decoupled representation and policy acquisition for continual reinforcement learning", "abs": "This contribution proposes adiabatic reinforcement learning (ARL), a new method for continual reinforcement learning (CRL). \nIn CRL, we assume a non-stationary environment partitioned into \\textit{tasks}. To avoid catastrophic forgetting (CF), RL requires the use\nof large replay buffers, which leads to very slow learning and high memory requirements.\nTo remedy this, we propose adiabatic reinforcement learning (ARL), a wake-sleep method that performs slow learning of internal representations from high-error transitions during sleep phases. Wake phases are used for the fast learning of policies, i.e., mappings from representations to actions, \nand to collect new high-error transitions. \nRepresentation learning is performed by \\textit{adiabatic replay} (AR), a recent CL technique we adapted to the RL setting. AR uses selective, internal replay of samples\nthat are likely to be affected by forgetting. Since this process is conditioned on incoming samples only, its has constant time-complexity w.r.t. tasks. Other benefits include \nfast adaptation to new tasks, and a very low memory footprint due to the complete absence of replay buffers.", "keywords": [ "continual learning", "reinforcement learning", "q-learning", "replay" ], "rating_list": [ 1, 3, 5, 3 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 1, 1, 2, 1 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "Q150eWkQ4I", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Benteng Sun", "gender": "Male", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "Undergrad student" }, { "name": "Haijin Zeng", "gender": "Male", "institution": "IMEC & Universiteit Gent", "country": "BE", "position": "PhD student" }, { "name": "Jingyong Su", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yong Xu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yongyong Chen", "gender": "Male", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "Associate Professor" } ], "n_formula": 21, "n_formula_1": 14, "n_ref_uni": 42, "n_ref": 70, "n_ref_all": 112, "n_fig": 16, "n_tab": 13, "L_tab": 6736, "n_element_tab": 699, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 6014, "n_element_tab_1": 381, "formula_len_all": 1915, "formula_len_all_1": 1395, "len_all": 221102, "len_all_1": 65747, "len_abs": 5436, "len_title": 126, "len_sents": 62377, "len_sents_1": 24525, "n_sents": 431, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 78, "L_abs": 1681, "title": "Spectral Compressive Imaging via Unmixing-driven Subspace Diffusion Refinement", "abs": "Spectral Compressive Imaging (SCI) reconstruction is inherently ill-posed, offering multiple plausible solutions from a single observation. Traditional deterministic methods typically struggle to effectively recover high-frequency details. Although diffusion models offer promising solutions to this challenge, their application is constrained by the limited training data and high computational demands associated with multispectral images (MSIs), complicating direct training. To address these issues, we propose a novel Predict-and-unmixing-driven-Subspace-Refine framework (PSR-SCI). This framework begins with a cost-effective predictor that produces an initial, rough estimate of the MSI. Subsequently, we introduce a unmixing-driven reversible spectral embedding module that decomposes the MSI into subspace images and spectral coefficients. This decomposition facilitates the adaptation of pre-trained RGB diffusion models and focuses refinement processes on high-frequency details, thereby enabling efficient diffusion generation with minimal MSI data. Additionally, we design a high-dimensional guidance mechanism with imaging consistency to enhance the model's efficacy. The refined subspace image is then reconstructed back into an MSI using the reversible embedding, yielding the final MSI with full spectral resolution. Experimental results on the standard KAIST and zero-shot datasets NTIRE, ICVL, and Harvard show that PSR-SCI enhances visual quality and delivers PSNR and SSIM metrics comparable to existing diffusion, transformer, and deep unfolding techniques. This framework provides a robust alternative to traditional deterministic SCI reconstruction methods.", "keywords": [ "Spectral compressive imaging", "subspace", "diffusion", "fine-tune" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "Q0zmmNNePz", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Alexander H Berger", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Alexander Weers", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "MS student" }, { "name": "Daniel Rueckert", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Johannes C. Paetzold", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "Postdoc" }, { "name": "Laurin Lux", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Ulrich Bauer", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "Associate Professor" }, { "name": "Nico Daniel Stucki", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 20, "n_ref": 63, "n_ref_all": 102, "n_fig": 14, "n_tab": 13, "L_tab": 5935, "n_element_tab": 506, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 1980, "n_element_tab_1": 170, "formula_len_all": 673, "formula_len_all_1": 864, "len_all": 143190, "len_all_1": 69090, "len_abs": 1232, "len_title": 146, "len_sents": 47087, "len_sents_1": 30047, "n_sents": 377, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 97, "L_abs": 1178, "title": "Topograph: An Efficient Graph-Based Framework for Strictly Topology Preserving Image Segmentation", "abs": "Topological correctness plays a critical role in many image segmentation tasks, yet most networks are trained using pixel-wise loss functions, such as Dice, neglecting topological accuracy. Existing topology-aware methods often lack robust topological guarantees, are limited to specific use cases, or impose high computational costs. \nIn this work, we propose a novel, graph-based framework for topologically accurate image segmentation that is both computationally efficient and generally applicable. Our method constructs a component graph that fully encodes the topological information of both the prediction and ground truth, allowing us to efficiently identify topologically critical regions and aggregate a loss based on local neighborhood information. Furthermore, we introduce a strict topological metric capturing the homotopy equivalence between the union and intersection of prediction-label pairs. We formally prove the topological guarantees of our approach and empirically validate its effectiveness on binary and multi-class datasets, demonstrating state-of-the-art performance with up to fivefold faster loss computation compared to persistent homology methods.", "keywords": [ "Image Segmentation", "Topology", "Graph" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 4, 2, 4 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "Q0s6kgrUMr", "primary_area": "causal reasoning", "n_author": 6, "author_info_list": [ { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Mingming Gong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Yang Zhou", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Undergrad student" }, { "name": "Yu Yao", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Lecturer" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 8, "n_ref_uni": 49, "n_ref": 94, "n_ref_all": 114, "n_fig": 9, "n_tab": 8, "L_tab": 14575, "n_element_tab": 1331, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 481, "n_element_tab_1": 56, "formula_len_all": 1663, "formula_len_all_1": 470, "len_all": 206562, "len_all_1": 70841, "len_abs": 1047, "len_title": 105, "len_sents": 58089, "len_sents_1": 33294, "n_sents": 454, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1053, "title": "A Robust Method to Discover Causal or Anticausal Relation", "abs": "Understanding whether the data generative process follows causal or anticausal relations is important for many applications. Existing causal discovery methods struggle with high-dimensional perceptual data such as images. Moreover, they require well-labeled data, which may not be feasible due to measurement error. In this paper, we propose a robust method to detect whether the data generative process is causal or anticausal. To determine the causal or anticausal relation, we identify an asymmetric property: under the causal relation, the instance distribution does not contain information about the noisy class-posterior distribution. We also propose a practical method to verify this via a noise injection approach. Our method is robust to label errors and is designed to handle both large-scale and high-dimensional datasets effectively. Both theoretical analyses and empirical results on a variety of datasets demonstrate the effectiveness of our proposed method in determining the causal or anticausal direction of the data generative process.", "keywords": [ "Trustworthy Machine Learning", "Semi-Superivsed Learning", "Causality" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 4, 2, 4 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "Q0mp2yBvb4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Benjamin Steenhoek", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Hengbo Tong", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "Researcher" }, { "name": "Md Mahbubur Rahman", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "PhD student" }, { "name": "Mirza Sanjida Alam", "gender": "Female", "institution": "Iowa State University", "country": "US", "position": "MS student" }, { "name": "Monoshi Kumar Roy", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "PhD student" }, { "name": "Swarna Das", "gender": "Female", "institution": "Iowa State University", "country": "US", "position": "PhD student" }, { "name": "Wei Le", "gender": "Female", "institution": "Iowa State University", "country": "", "position": "Associate Professor" }, { "name": "Earl Barr", "gender": "Male", "institution": "University College London", "country": "", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 82, "n_ref_all": 117, "n_fig": 14, "n_tab": 3, "L_tab": 610, "n_element_tab": 112, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 1275, "n_element_tab_1": 106, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 136379, "len_all_1": 46486, "len_abs": 1900, "len_title": 115, "len_sents": 38277, "len_sents_1": 22072, "n_sents": 239, "n_sents_1": 146, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1911, "title": "To Err is Machine: Vulnerability Detection Challenges LLM Reasoning", "abs": "In this paper, we present a challenging code reasoning task: vulnerability detection.\nLarge Language Models (LLMs) have shown promising results in natural-language\nand math reasoning, but state-of-the-art (SOTA) models reported only 54.5%\nBalanced Accuracy in our vulnerability detection evaluation, even those models\npre-trained on large amounts of source code. Our error analysis on LLM responses\nshows that the models struggle to reason about the code semantics relevant to\nidentifying vulnerabilities, especially subtle semantic differences caused by small\ntextual changes. We explored prominent models and training settings to understand\ntheir effects on vulnerability detection performance \u2014 including better prompts,\nlarger models, more pre-training data, and fine-tuning \u2014 but none led to significant\nimprovements. This raises the question of whether simply scaling training data and\nmodel size will allow us to \u201csolve\u201d complex code reasoning tasks like vulnerability\ndetection, or if a fundamental shift in modeling and training techniques is required.\nWe also explored adding domain knowledge to prompts; although it helped certain\nmodels understand some code semantics, vulnerability detection requires multi-\nstep reasoning, and these models still failed in steps, such as reasoning about\nvariable relations. Our results suggest that new models, new training methods, or\nmore execution-specific pretraining data may be needed to conquer vulnerability\ndetection. We speculate that auto-regressive pre-training on source code may not\neffectively extract code semantics, especially on the current pretraining mixtures,\nin which execution data is scarce. Success on vulnerability detection as a code\nreasoning task can benefit many areas of software engineering such as debugging,\ntest input generation, and program repair. Our code and data are available at\nhttps://figshare.com/s/78fe02e56e09ec49300b.", "keywords": [ "deep learning", "security", "vulnerability detection" ], "rating_list": [ 6, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 2, 3, 1 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "Q0TEVKV2cp", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "B\u00e1lint Mucs\u00e1nyi", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Lukas Tatzel", "gender": "Male", "institution": "University of T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Osane Hackel", "gender": "Female", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "", "position": "MS student" }, { "name": "Philipp Hennig", "gender": "Male", "institution": "University of T\u00fcbingen", "country": "DE", "position": "Full Professor" } ], "n_formula": 42, "n_formula_1": 7, "n_ref_uni": 36, "n_ref": 72, "n_ref_all": 151, "n_fig": 20, "n_tab": 5, "L_tab": 1456, "n_element_tab": 337, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3395, "formula_len_all_1": 439, "len_all": 265846, "len_all_1": 81214, "len_abs": 2970, "len_title": 113, "len_sents": 94013, "len_sents_1": 34122, "n_sents": 813, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 983, "title": "Debiasing Mini-Batch Quadratics for Applications in Deep Learning", "abs": "Quadratic approximations form a fundamental building block of machine learning methods. E.g., second-order optimizers try to find the Newton step into the minimum of a local quadratic proxy to the objective function; and the second-order approximation of a network's loss function can be used to quantify the uncertainty of its outputs via the Laplace approximation. When computations on the entire training set are intractable - typical for deep learning - the relevant quantities are computed on mini-batches. This, however, distorts and biases the shape of the associated *stochastic* quadratic approximations in an intricate way with detrimental effects on applications. In this paper, we (i) show that this bias introduces a systematic error, (ii) provide a theoretical explanation for it, (iii) explain its relevance for second-order optimization and uncertainty quantification via the Laplace approximation in deep learning, and (iv) develop and evaluate debiasing strategies.", "keywords": [ "quadratic Taylor approximation", "mini-batching", "second-order optimizers", "conjugate gradients", "uncertainty quantification", "Laplace approximation", "stochastic curvature", "GGN", "KFAC" ], "rating_list": [ 8, 3, 8, 8 ], "soundness_list": [ 3, 2, 4, 4 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "Q00XEQxA45", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Tongda Xu", "gender": "Non-Binary", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xinjie Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" }, { "name": "Yan Wang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "YixingWu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 38, "n_ref": 66, "n_ref_all": 74, "n_fig": 2, "n_tab": 5, "L_tab": 3313, "n_element_tab": 520, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 4433, "n_element_tab_1": 429, "formula_len_all": 314, "formula_len_all_1": 314, "len_all": 107004, "len_all_1": 61753, "len_abs": 1192, "len_title": 122, "len_sents": 27114, "len_sents_1": 23877, "n_sents": 200, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1201, "title": "TRAIN THE LATENT, NOT THE IMAGE: JOINT IMAGE COMPRESSION AND STEGANOGRAPHY", "abs": "Image steganography is the process of hiding secret information in an image through imperceptible changes. Most of recent works achieve message in the image by modifying the pixels of image itself. However, those images with hidden messages are not robust to compression such as JPEG, which is used almost everywhere. In order to achieve the ability to compress the image while still having the ability to carry the message, we propose an innovative optimization method which leverages a semi-amortized approach to directly manipulate latent space data for the joint optimization of image compression and steganography. In the compression module, we investigate two of the most popular models in learned image compression with different pre-trained quality: the hyperprior model and the ELIC model. For the steganography module, our method employs the pre-trained fixed neural network steganography (FNNS) model. We compare our method with two state-of-the-art methods such as FNNS-JPEG and LISO-JPEG, achieving significant image compression while maintaining high fidelity and ensuring the accuracy of content upon decoding. The results demonstrate the effectiveness and superiority of our approach.", "keywords": [ "steganography", "image compression", "semi-amotorize", "adversatial learning" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 5, 5, 2 ] }, { "paper_id": "PzpSJkdokk", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Chaohan Wang", "gender": "Male", "institution": "University of Adelaide", "country": "", "position": "PhD student" }, { "name": "Qi Chen", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Qi Wu", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "Associate Professor" }, { "name": "Yutong Xie", "gender": "Not Specified", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 27, "n_ref": 51, "n_ref_all": 80, "n_fig": 10, "n_tab": 3, "L_tab": 1647, "n_element_tab": 242, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 685, "n_element_tab_1": 30, "formula_len_all": 399, "formula_len_all_1": 325, "len_all": 133335, "len_all_1": 58471, "len_abs": 1973, "len_title": 122, "len_sents": 46260, "len_sents_1": 27769, "n_sents": 347, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1799, "title": "Radiologist-like Progressive Radiology Report Generation and Benchmarking", "abs": "Radiology report generation is a critical application at the intersection of radiology and artificial intelligence. It aims to reduce radiologists' workload by automating the interpretation and reporting of medical images. Previous works have employed diverse approaches, with some focusing solely on imaging data while others incorporate the indication but often neglect the interrelationships among different report sections. \nOur work identifies and harnesses the intrinsic relationships between the indication, findings, and impression sections of a radiology report.\nThe indication section provides the clinical context and specifies the reason for the examination, setting the stage for targeted image analysis. The findings section details the radiologist's observations from the image, including identified abnormalities and relevant normal findings. The impression section synthesizes these observations to form a diagnostic conclusion, directly addressing the clinical query presented in the indication.\nBy mapping these relationships, we propose a Radiologist-Like Progressive Generation (RLPG) framework that mirrors the radiologist's workflow for report generation.\nInitially, an image encoder and a large language model process the imaging data alongside the indication to generate detailed findings. Subsequently, the same image, the indication, and the predicted findings are utilized to produce a concise impression. This method improves the alignment between report sections and improves the clinical relevance of the generated reports. \nTo facilitate research and benchmarking in report generation, we introduce MIMIC-1V3 (i.e., 1 case vs. 3 sections), a curated dataset derived from the MIMIC-CXR by dividing each report into three sections: indication, findings, and impression.", "keywords": [ "radiology report generation" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 5, 3 ] }, { "paper_id": "PzkYyZH9Fx", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Brooks Paige", "gender": "Male", "institution": "University College London", "country": "", "position": "Associate Professor" }, { "name": "Chuanhao Sun", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Principal Researcher" }, { "name": "Kai Xu", "gender": "Male", "institution": "MIT-IBM Watson AI Lab", "country": "", "position": "Research scientist" }, { "name": "Mahesh K. Marina", "gender": "Male", "institution": "Edinburgh University, University of Edinburgh", "country": "GB", "position": "Full Professor" }, { "name": "Qiyu Wang", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 14, "n_ref_uni": 31, "n_ref": 92, "n_ref_all": 107, "n_fig": 3, "n_tab": 7, "L_tab": 8817, "n_element_tab": 483, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 3554, "n_element_tab_1": 245, "formula_len_all": 2936, "formula_len_all_1": 1168, "len_all": 138932, "len_all_1": 63495, "len_abs": 1290, "len_title": 129, "len_sents": 35966, "len_sents_1": 26474, "n_sents": 262, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1296, "title": "Narcissus: Leveraging Early Training Dynamics for Unsupervised Anomaly Detection", "abs": "Anomaly detection is a critical learning task with many significant and diverse applications. Currently, semi-supervised methods provide the state-of-the-art accuracy performance but require labeled normal data for training. Unsupervised approaches, on the other hand, do not have this requirement but can only offer inferior anomaly detection performance. In this paper, we introduce NARCISSUS, a novel unsupervised anomaly detection method that achieves accuracy comparable to semi-supervised approaches. Our key insight is that a learning model when training with a mix of normal and sparse anomalous data converges first on normal data. Leveraging this insight, NARCISSUS employs a tailored early stopping scheme, eliminating the need for pseudo labels and costly label generation interactions. It also offers systematic solutions to minimize the influence of model uncertainty, ensuring robust detection. NARCISSUS is model-agnostic and can therefore make use of even a semi-supervised anomaly detection model underneath, thereby turning it into an unsupervised one. Comprehensive evaluations using time series, image and graph datasets show that NARCISSUS provides similar or better detection performance compared to best-performing semi-supervised methods while not requiring labeled data.", "keywords": [ "Anomaly Detection", "Unsupervised Learning", "Early Stopping" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 1, 3, 4, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "PzGyZFIn5U", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Haijin Zeng", "gender": "Male", "institution": "IMEC & Universiteit Gent", "country": "BE", "position": "PhD student" }, { "name": "Jie Liu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Jingyong Su", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Shiyang Zhou", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Yongyong Chen", "gender": "Male", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 15, "n_ref": 18, "n_ref_all": 43, "n_fig": 9, "n_tab": 4, "L_tab": 4112, "n_element_tab": 126, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 3903, "n_element_tab_1": 85, "formula_len_all": 400, "formula_len_all_1": 397, "len_all": 119695, "len_all_1": 53314, "len_abs": 1773, "len_title": 135, "len_sents": 28220, "len_sents_1": 23663, "n_sents": 200, "n_sents_1": 148, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1837, "title": "Lightweight Quad Bayer HybridEVS Demosaicing via State Space Augmented Cross-Attention", "abs": "Event cameras like the Hybrid Event-based Vision Sensor (HybridEVS) camera capture brightness changes as asynchronous \"events\" instead of frames, offering advantages over traditional cameras: high temporal resolution, wide dynamic range, and no motion blur. However, challenges arise from combining a Quad Bayer Color Filter Array (CFA) sensor with event pixels lacking color information, resulting in aliasing and artifacts on the demosaicing process before downstream application. Current methods struggle to address these issues, especially on resource-limited mobile devices. In response, we introduce \\textbf{TSANet}, a lightweight \\textbf{T}wo-stage network via \\textbf{S}tate space augmented cross-\\textbf{A}ttention, which can handle event pixels inpainting and Quad Bayer demosaicing separately, leveraging the benefits of dividing complex tasks into manageable subtasks and learning them through a two-step training strategy to enhance robustness. Additionally, we propose a lightweight Cross-Swin State Block (CSSB) designed to augment the model's capacity to capture global dependencies using state space models in a linear format, along with cross-modality Swin attention to integrate additional priors like CFA pattern and event map, outperforming traditional local attention mechanisms while also reducing model size. In summary, TSANet demonstrates excellent demosaicing performance on HybridEVS while maintaining a lightweight model, averaging better results than the previous state-of-the-art method DemosaicFormer across seven diverse datasets in both PSNR and SSIM, while respectively reducing parameter and computation costs by $1.86\\times$ and $3.29\\times$. Our approach presents new possibilities for efficient image demosaicing on mobile devices. \\textit{Code and models are available in supplementary materials.}", "keywords": [ "Demosaicing", "HybridEVS", "Quad Bayer", "State Space", "Attention" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "Pz9zFea4MQ", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Bhiksha Raj", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Matthew Johnson-Roberson", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Sebastian Scherer", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Shibo Zhao", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Sibo Wang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "MS student" }, { "name": "Tianyi Zhang", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Xiang Li", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Xiaohao Xu", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Xiaonan Huang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" }, { "name": "Ye Li", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "MS student" }, { "name": "Yongqi Chen", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "MS student" } ], "n_formula": 67, "n_formula_1": 6, "n_ref_uni": 53, "n_ref": 162, "n_ref_all": 267, "n_fig": 54, "n_tab": 31, "L_tab": 11546, "n_element_tab": 850, "n_fig_1": 14, "n_tab_1": 4, "L_tab_1": 1023, "n_element_tab_1": 129, "formula_len_all": 3023, "formula_len_all_1": 327, "len_all": 442448, "len_all_1": 63867, "len_abs": 2038, "len_title": 155, "len_sents": 160280, "len_sents_1": 28718, "n_sents": 1288, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 106, "L_abs": 1814, "title": "Scalable Benchmarking and Robust Learning for Noise-Free Ego-Motion and 3D Reconstruction from Noisy Video", "abs": "We aim to redefine robust ego-motion estimation and photorealistic 3D reconstruction by addressing a critical limitation: the reliance on noise-free data in existing models. While such sanitized conditions simplify evaluation, they fail to capture the unpredictable, noisy complexities of real-world environments. Dynamic motion, sensor imperfections, and synchronization perturbations lead to sharp performance declines when these models are deployed in practice, revealing an urgent need for frameworks that embrace and excel under real-world noise.\nTo bridge this gap, we tackle three core challenges: scalable data generation, comprehensive benchmarking, and model robustness enhancement. First, we introduce a scalable noisy data synthesis pipeline that generates diverse datasets simulating complex motion, sensor imperfections, and synchronization errors. Second, we leverage this pipeline to create Robust-Ego3D, a benchmark rigorously designed to expose noise-induced performance degradation, highlighting the limitations of current learning-based methods in ego-motion accuracy and 3D reconstruction quality. Third, we propose Correspondence-guided Gaussian Splatting (CorrGS), a novel method that progressively refines an internal clean 3D representation by aligning noisy observations with rendered RGB-D frames from clean 3D map, enhancing geometric alignment and appearance restoration through visual correspondence.\nExtensive experiments on synthetic and real-world data demonstrate that CorrGS consistently outperforms prior state-of-the-art methods, particularly in scenarios involving rapid motion and dynamic illumination. We will release our code and benchmark to advance robust 3D vision, setting a new standard for ego-motion estimation and high-fidelity reconstruction in noisy environments.", "keywords": [ "Benchmarking", "Robustness", "Neural 3D Reconstruction", "Ego-Motion Estimation", "SLAM" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "PyyoSwPaSa", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Chongxuan Li", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Guang Cheng", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "Full Professor" }, { "name": "Liyan Xie", "gender": "Female", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "Assistant Professor" }, { "name": "Yidong Ouyang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 14, "n_ref_uni": 52, "n_ref": 163, "n_ref_all": 184, "n_fig": 2, "n_tab": 12, "L_tab": 2240, "n_element_tab": 202, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1438, "n_element_tab_1": 55, "formula_len_all": 5198, "formula_len_all_1": 747, "len_all": 176050, "len_all_1": 65381, "len_abs": 1076, "len_title": 120, "len_sents": 51523, "len_sents_1": 29808, "n_sents": 384, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1056, "title": "MissDiff: Training Diffusion Models on Tabular Data with Missing Values", "abs": "The diffusion model has shown remarkable performance in modeling data distributions and synthesizing data. However, the vanilla diffusion model requires complete or fully observed training data. Incomplete data is a common issue in various real-world applications, including healthcare and finance, particularly when dealing with tabular datasets. This work considers learning from data with missing values for missing value imputations and generating synthetic complete data in a unified framework. With minimal assumptions on the missing mechanisms, our method models the score of complete data distribution by denoising score matching on data with missing values. We prove that the proposed method can recover the score of the complete data distribution, and the proposed training objective serves as an upper bound for the negative likelihood of observed data. Extensive experiments on imputation tasks together with generation tasks demonstrate that our proposed framework outperforms existing state-of-the-art approaches on multiple tabular datasets.", "keywords": [ "Diffusion Model", "Missing Value", "Tabular Data" ], "rating_list": [ 8, 3, 6, 6 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "PytShcLwOn", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Alexey Gavryushin", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Florian Redhardt", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "Undergrad student" }, { "name": "Gaia Di Lorenzo", "gender": "Female", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Kaichun Mo", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Luc Van Gool", "gender": "unknown", "institution": "INSAIT - Sofia Un.", "country": "BG", "position": "Full Professor" }, { "name": "Marc Pollefeys", "gender": "Male", "institution": "Microsoft", "country": "CH", "position": "Director" }, { "name": "Xi Wang", "gender": "Female", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 36, "n_ref": 60, "n_ref_all": 78, "n_fig": 4, "n_tab": 2, "L_tab": 509, "n_element_tab": 49, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 299, "n_element_tab_1": 8, "formula_len_all": 218, "formula_len_all_1": 218, "len_all": 128835, "len_all_1": 64478, "len_abs": 1848, "len_title": 95, "len_sents": 37026, "len_sents_1": 32686, "n_sents": 276, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1549, "title": "SIGHT: Single-Image Conditioned Generation of Hand Trajectories for Hand-Object Interaction", "abs": "We introduce a novel task of generating realistic and diverse 3D hand trajectories given a single image of an object, which could be involved in a hand-object interaction scene or pictured by itself. When humans reach for an object, appropriate trajectories naturally form to manipulate it for specific tasks in our minds. Such hand-object interaction trajectory priors can greatly benefit applications in robotics, embodied AI, augmented reality and related fields. To tackle this challenging problem, we propose the SIGHT-Fusion system, consisting of a carefully curated pipeline for extracting features at various levels of hand-object interaction details from the single image input, and a conditional motion generation diffusion model processing the extracted features. We train our method given video data with corresponding hand trajectory annotations, without supervision in the form of action labels. For the evaluation, we establish benchmarks utilizing the FPHAB and HOI4D datasets, testing our method against various baselines and metrics. We also introduce task simulators for executing the generated hand trajectories and reporting task success rates as an additional metric. Experiments show that our method generates more natural and diverse hand trajectories than baselines and presents promising generalization capability on unseen objects. The accuracy of the generated hand trajectories is confirmed in a physics simulation setting, showcasing the authenticity of the created sequences and their applicability in downstream uses.", "keywords": [ "Human Motion Generation", "Human-Object Interaction", "Diffusion Models" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 3, 4, 5, 5 ] }, { "paper_id": "PyjZO7oSw2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Bin Cui", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Joseph E. Gonzalez", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Associate Professor" }, { "name": "Ling Yang", "gender": "Male", "institution": "DeepSeek AI", "country": "CN", "position": "Intern" }, { "name": "Minkai Xu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Tianjun Zhang", "gender": "unknown", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Zhaochen Yu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 36, "n_ref": 82, "n_ref_all": 102, "n_fig": 5, "n_tab": 10, "L_tab": 6213, "n_element_tab": 314, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 238, "n_element_tab_1": 26, "formula_len_all": 1495, "formula_len_all_1": 957, "len_all": 221580, "len_all_1": 64911, "len_abs": 1688, "len_title": 131, "len_sents": 65463, "len_sents_1": 31554, "n_sents": 484, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1770, "title": "SuperCorrect: Supervising and Correcting Language Models with Error-Driven Insights", "abs": "Large language models (LLMs) like GPT-4, PaLM, and LLaMA have shown significant improvements in various reasoning tasks. However, smaller models such as Llama-3-8B and DeepSeekMath-Base still struggle with complex mathematical reasoning because they fail to effectively identify and correct reasoning errors. Recent reflection-based methods aim to address these issues by enabling self-reflection and self-correction, but they still face challenges in independently detecting errors in their reasoning steps.\nTo overcome these limitations, we propose SuperCorrect, a novel two-stage framework that uses a large teacher model to supervise and correct both the reasoning and reflection processes of a smaller student model. In the first stage, we extract hierarchical high-level and detailed thought templates from the teacher model to guide the student model in eliciting more fine-grained reasoning thoughts. In the second stage, we introduce cross-model collaborative direct preference optimization (DPO) to enhance the self-correction abilities of the student model by following the teacher's correction traces during training. This cross-model DPO approach teaches the student model to effectively locate and resolve erroneous thoughts with error-driven insights from the teacher model, breaking the bottleneck of its thoughts and acquiring new skills and knowledge to tackle challenging problems. Extensive experiments consistently demonstrate our superiority over previous methods. Notably, our SuperCorrect-7B model significantly surpasses powerful DeepSeekMath-7B by 7.8\\%/5.3\\% and Qwen2.5-Math-7B by 15.1\\%/6.3\\% on MATH/GSM8K benchmarks, achieving new SOTA performance among all 7B models. Code is available at: https://github.com/YangLing0818/SuperCorrect-llm", "keywords": [ "Large Language Models", "LLM Reasoning" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "PxlfzEePC0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bo Dai", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Bowen Zhang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Research Scientist" }, { "name": "Haoshuo Huang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Research Engineer" }, { "name": "Haotian Sun", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Nan Du", "gender": "Male", "institution": "Apple/AIML", "country": "", "position": "Principal Researcher" }, { "name": "Ruoming Pang", "gender": "unknown", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Tao Lei", "gender": "Male", "institution": "Apple", "country": "US", "position": "Research scientist" }, { "name": "Yanghao Li", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 28, "n_ref": 84, "n_ref_all": 112, "n_fig": 8, "n_tab": 3, "L_tab": 2225, "n_element_tab": 239, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1502, "n_element_tab_1": 121, "formula_len_all": 587, "formula_len_all_1": 565, "len_all": 147011, "len_all_1": 64123, "len_abs": 1288, "len_title": 122, "len_sents": 39101, "len_sents_1": 29615, "n_sents": 310, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1293, "title": "EC-DIT: Scaling Diffusion Transformers with Adaptive Expert-Choice Routing", "abs": "Diffusion transformers have been widely adopted for text-to-image synthesis. While scaling these models up to billions of parameters shows promise, the effectiveness of scaling beyond current sizes remains underexplored and challenging. By explicitly exploiting the computational heterogeneity of image generations, we develop a new family of Mixture-of-Experts (MoE) models (EC-DIT) for diffusion transformers with expert-choice routing. EC-DIT learns to adaptively optimize the compute allocated to understand the input texts and generate the respective image patches, enabling heterogeneous computation aligned with varying text-image complexities. This heterogeneity provides an efficient way of scaling EC-DIT up to 97 billion parameters and achieving significant improvements in training convergence, text-to-image alignment, and overall generation quality over dense models and conventional MoE models. Through extensive ablations, we show that EC-DIT demonstrates superior scalability and adaptive compute allocation by recognizing varying textual importance through end-to-end training. Notably, in text-to-image alignment evaluation, our largest models achieve a state-of-the-art GenEval score of 71.68% and still maintain competitive inference speed with intuitive interpretability.", "keywords": [ "Diffusion transformer", "text-to-image synthesis", "Mixture-of-Experts" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "Pxik3T6Mn9", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Haomin Wen", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Leman Akoglu", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Shurui Cao", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 18, "n_ref_uni": 37, "n_ref": 62, "n_ref_all": 79, "n_fig": 6, "n_tab": 9, "L_tab": 2362, "n_element_tab": 228, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1841, "n_element_tab_1": 151, "formula_len_all": 1158, "formula_len_all_1": 960, "len_all": 140052, "len_all_1": 72868, "len_abs": 1349, "len_title": 111, "len_sents": 44096, "len_sents_1": 32951, "n_sents": 325, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1357, "title": "Uncertainty-aware Human Mobility Modeling and Anomaly Detection", "abs": "Given the GPS coordinates of a large collection of human agents over time, how can we model their mobility behavior toward effective anomaly detection (e.g. for bad-actor or malicious behavior detection) without any labeled data?\nHuman mobility and trajectory modeling have been studied extensively with varying capacity to handle complex input, and performance-efficiency trade-offs. With the arrival of more expressive models in machine learning, we attempt to model GPS data as a sequence of stay-point events, each with a set of characterizing spatiotemporal features, and leverage modern sequence models such as Transformers for un/self-supervised training and inference. Notably, driven by the inherent stochasticity of certain individuals' behavior, we equip our model with aleatoric/data uncertainty estimation. In addition, to handle data sparsity of a large variety of behaviors, we incorporate epistemic/model uncertainty into our model. Together, aleatoric and epistemic uncertainty enable a robust loss and training dynamics, as well as uncertainty-aware decision making in anomaly scoring. Experiments on large expert-simulated datasets with tens of thousands of agents demonstrate the effectiveness of our model against both forecasting and anomaly detection baselines. All code is available at https://anonymous.4open.science/r/mobility-ad.", "keywords": [ "uncertaitnty", "human mobility", "anomaly detection" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 1, 1, 3 ], "confidence_list": [ 5, 5, 4, 2 ] }, { "paper_id": "PxYhHF6FNH", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Hamed Damirchi", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "PhD student" }, { "name": "Zhen Zhang", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Cristian Rodriguez", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Ehsan M Abbasnejad", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Associate Professor" }, { "name": "Javen Shi", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Professor" } ], "n_formula": 18, "n_formula_1": 9, "n_ref_uni": 30, "n_ref": 55, "n_ref_all": 87, "n_fig": 15, "n_tab": 5, "L_tab": 760, "n_element_tab": 116, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 464, "n_element_tab_1": 75, "formula_len_all": 1039, "formula_len_all_1": 490, "len_all": 133193, "len_all_1": 66229, "len_abs": 1781, "len_title": 98, "len_sents": 49199, "len_sents_1": 32211, "n_sents": 356, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1789, "title": "The Quest for Winning Tickets in Low-Rank Adapters", "abs": "Low-Rank Adaptation (LoRA), a prominent parameter-efficient fine-tuning (PEFT) method, offers an effective strategy for adapting large pre-trained models to specific tasks with minimal computational overhead. LoRA achieves this by introducing low-rank parameter matrices to the frozen pre-trained models. However, despite their efficiency, LoRA and its variants modify all elements of a parameter block, which is unnecessary as LoRA primarily aims to adjust a small set of subspaces that capture task-specific knowledge. Drawing inspiration from the Lottery Ticket Hypothesis (LTH), which posits that dense neural networks contain sparse subnetworks capable of performing similarly to fully-parameterized models, we investigate whether similar sparse subnetworks exist for low-rank adapters. We demonstrate that such subnetworks, often referred to as \"winning tickets\" in the context of LTH, indeed exist for low-rank adapters. We introduce a method to identify this sparse subset of weights for each layer by relating the top subspaces of the pretrained parameter block to the elements of the corresponding weight matrix. This subset is then fine-tuned using LoRA. We show that this sparse subset is not necessarily unique; as long as sparsity is kept within a certain bound defined by the task, random subnetworks with similar sparsity can act as winning tickets. Building on this discovery, we propose a novel approach called Partial-LoRA, which adds sparse low-rank parameters to pre-trained models. Through extensive experiments on 8 vision and 4 language tasks, we demonstrate that Partial-LoRA can reduce trainable parameters by up to 87% while maintaining or even improving model performance in some cases. Our work thus reduces memory needs and theoretically grounds sparse LoRAs.", "keywords": [ "Parameter-efficient fine-tuning", "Low-Rank Adaptation", "Lottery Ticket Hypothesis", "Pruning" ], "rating_list": [ 5, 6, 6, 3, 6 ], "soundness_list": [ 3, 2, 4, 1, 3 ], "presentation_list": [ 2, 4, 3, 2, 4 ], "contribution_list": [ 2, 3, 3, 1, 3 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "PxRATSTDlS", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Huandong Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Qingmin Liao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Ruikun Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 49, "n_ref": 79, "n_ref_all": 121, "n_fig": 17, "n_tab": 4, "L_tab": 1234, "n_element_tab": 144, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 718, "formula_len_all_1": 514, "len_all": 170718, "len_all_1": 60122, "len_abs": 1364, "len_title": 157, "len_sents": 55616, "len_sents_1": 30667, "n_sents": 359, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 109, "L_abs": 1371, "title": "Predicting the Energy Landscape of Stochastic Dynamical System via Physics-informed Self-supervised Learning", "abs": "Energy landscapes play a crucial role in shaping dynamics of many real-world complex systems. System evolution is often modeled as particles moving on a landscape under the combined effect of energy-driven drift and noise-induced diffusion, where the energy governs the long-term motion of the particles. Estimating the energy landscape of a system has been a longstanding interdisciplinary challenge, hindered by the high operational costs or the difficulty of obtaining supervisory signals. Therefore, the question of how to infer the energy landscape in the absence of true energy values is critical. In this paper, we propose a physics-informed self-supervised learning method to learn the energy landscape from the evolution trajectories of the system. It first maps the system state from the observation space to a discrete landscape space by an adaptive codebook, and then explicitly integrates energy into the graph neural Fokker-Planck equation, enabling the joint learning of energy estimation and evolution prediction. Experimental results across interdisciplinary systems demonstrate that our estimated energy has a correlation coefficient above 0.9 with the ground truth, and evolution prediction accuracy exceeds the baseline by an average of 17.65\\%. The code is available in the anonymous repository: https://anonymous.4open.science/r/PESLA-0D9A/README.md", "keywords": [ "dynamical system", "energy landscape", "deep learning" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "PxBzxO02Ef", "primary_area": "foundation or frontier models, including LLMs", "n_author": 1, "author_info_list": [ { "name": "Kunyang Zhou", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 32, "n_ref": 93, "n_ref_all": 120, "n_fig": 6, "n_tab": 10, "L_tab": 8548, "n_element_tab": 820, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 162, "formula_len_all_1": 121, "len_all": 192276, "len_all_1": 51878, "len_abs": 1644, "len_title": 114, "len_sents": 38458, "len_sents_1": 22166, "n_sents": 355, "n_sents_1": 155, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1635, "title": "LVP: Language-guide Visual Projector for Efficient Multimodal LLM", "abs": "Visual projector plays a crucial role in bridging the visual model and the large language model (LLM) in modern multimodal LLM. \nTypically, MLLMs utilize a simple MLP to preserve all visual tokens, causing a heavy computational burden and redundant visual tokens.\nSome recent works adopt either a resampler or an adaptive pooling to reduce the visual tokens. However, they only reduce the visual tokens based on the image feature,\nleading to the feature misalignment between visual tokens and text tokens. In this paper, we present a novel Language-guidance Visual Projector (LVP), where the text \nfeature serves as a guide to selecting the important visual tokens. Specially, we first adopt a lightweight text encoder to extract the text feature. Then, a lightweight\ncross-modal feature enhancement module is proposed to enhance the cross-modal feature alignment. Finally, we select the important visual tokens according to the feature similarity between visual tokens and text tokens and apply\na deformable attention module to integrate the visual features from the visual encoder into the selected visual tokens. We further propose a multi-level language-guidance visual projector, which selects the visual tokens from different stages of the visual encoder.\nExtensive experiments demonstrate that our LVP compresses the visual tokens by 75\\%~95\\% while achieving competitive even better performance across diverse benchmarks with a significant efficiency advantage. For instance, LLaVA1.5-LVP with Qwen2.5-7B\nobtains 72.4\\% accuracy on VQA$^\\text{T}$, realizing the state-of-the-art result. The code and the model will be released.", "keywords": [ "Multimodal large language model", "Visual projector", "Language-guide visual token selection." ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 2, 4, 3, 2 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "PwxYoMvmvy", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Hao Zhu", "gender": "Male", "institution": "CSIRO", "country": "AU", "position": "Postdoc" }, { "name": "Xiao-Ming Wu", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Associate Professor" }, { "name": "Yuankai Luo", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" } ], "n_formula": 65, "n_formula_1": 23, "n_ref_uni": 60, "n_ref": 105, "n_ref_all": 122, "n_fig": 7, "n_tab": 4, "L_tab": 4272, "n_element_tab": 206, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 3155, "n_element_tab_1": 110, "formula_len_all": 4343, "formula_len_all_1": 1165, "len_all": 161846, "len_all_1": 62522, "len_abs": 1378, "len_title": 118, "len_sents": 37195, "len_sents_1": 22916, "n_sents": 344, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1387, "title": "Beyond Random Masking: When Dropout meets Graph Convolutional Networks", "abs": "Graph Convolutional Networks (GCNs) have emerged as powerful tools for learning on graph-structured data, yet the behavior of dropout in these models remains poorly understood. This paper presents a comprehensive theoretical analysis of dropout in GCNs, revealing its unique interactions with graph structure. We demonstrate that dropout in GCNs creates dimension-specific stochastic sub-graphs, leading to a form of structural regularization not present in standard neural networks. Our analysis shows that dropout effects are inherently degree-dependent, resulting in adaptive regularization that considers the topological importance of nodes. We provide new insights into dropout's role in mitigating oversmoothing and derive novel generalization bounds that account for graph-specific dropout effects. Furthermore, we analyze the synergistic interaction between dropout and batch normalization in GCNs, uncovering a mechanism that enhances overall regularization. Our theoretical findings are validated through extensive experiments on both node-level and graph-level tasks across 14 datasets. Notably, GCN with dropout and batch normalization outperforms state-of-the-art methods on several benchmarks. This work bridges a critical gap in the theoretical understanding of regularization in GCNs and provides practical insights for designing more effective graph learning algorithms.", "keywords": [ "graph neural networks", "dropout" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "PwmEvdXFAv", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Holakou Rahmanian", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Lexing Ying", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Professor" }, { "name": "Michael Shavlovsky", "gender": "Male", "institution": "Amazon", "country": "US", "position": "PhD student" }, { "name": "Tesi Xiao", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Xun Tang", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 38, "n_formula_1": 22, "n_ref_uni": 50, "n_ref": 80, "n_ref_all": 85, "n_fig": 4, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4518, "formula_len_all_1": 2419, "len_all": 135971, "len_all_1": 63543, "len_abs": 987, "len_title": 137, "len_sents": 38202, "len_sents_1": 26570, "n_sents": 334, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 987, "title": "An efficient algorithm for entropic optimal transport under martingale-type constraints", "abs": "This work introduces novel computational methods for entropic optimal transport (OT) problems under martingale-type conditions.\nThe problems can map to a prevalent class of OT problems with structural constraints, encompassing the discrete martingale optimal transport (MOT) problem, as the (super-)martingale conditions are equivalent to row-wise (in-)equality constraints on the coupling matrix. Inspired by the recent empirical success of Sinkhorn-type algorithms, we propose an entropic formulation for the MOT problem and introduce Sinkhorn-type algorithms with sparse Newton iterations that utilize the (approximate) sparsity of the Hessian matrix of the dual objective. As exact martingale conditions are typically infeasible, we adopt entropic regularization to find an approximate constraint satisfied solution. We show that in practice the proposed algorithms enjoy both super-exponential convergence and robustness with controllable thresholds for total constraint violations.", "keywords": [ "Optimal Transport", "Martingale Constraints" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "PwLsQ1AFbP", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Guang Lin", "gender": "Male", "institution": "Tokyo University of Agriculture and Technology", "country": "JP", "position": "PhD student" }, { "name": "Jianhai Zhang", "gender": "Male", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "Full Professor" }, { "name": "Qibin Zhao", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Team Leader" }, { "name": "Toshihisa Tanaka", "gender": "unknown", "institution": "Tokyo University of Agriculture and Technology, Tokyo Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Zerui Tao", "gender": "unknown", "institution": "RIKEN", "country": "JP", "position": "Postdoc" } ], "n_formula": 22, "n_formula_1": 12, "n_ref_uni": 35, "n_ref": 97, "n_ref_all": 122, "n_fig": 4, "n_tab": 6, "L_tab": 1270, "n_element_tab": 136, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 2857, "n_element_tab_1": 303, "formula_len_all": 1760, "formula_len_all_1": 859, "len_all": 142496, "len_all_1": 68370, "len_abs": 1628, "len_title": 113, "len_sents": 39071, "len_sents_1": 28540, "n_sents": 260, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1634, "title": "Adversarial Guided Diffusion Models for Adversarial Purification", "abs": "Diffusion model (DM) based adversarial purification (AP) has proven to be a powerful defense method that can remove adversarial perturbations and generate a purified example without threats. In principle, the pre-trained DMs can only ensure that purified examples conform to the same distribution of the training data, but it may inadvertently compromise the semantic information of input examples, leading to misclassification of purified examples. Recent advancements introduce guided diffusion techniques to preserve semantic information while removing the perturbations. However, these guidances often rely on distance measures between purified examples and diffused examples, which can also preserve perturbations in purified examples. To further unleash the robustness power of DM-based AP, we propose an adversarial guided diffusion model (AGDM) by introducing a novel adversarial guidance that contains sufficient semantic information but does not explicitly involve adversarial perturbations. The guidance is modeled by an auxiliary neural network obtained with adversarial training, considering the distance in the latent representations rather than at the pixel-level values. Extensive experiments are conducted on CIFAR-10, CIFAR-100 and ImageNet to demonstrate that our method is effective for simultaneously maintaining semantic information and removing the adversarial perturbations. In addition, comprehensive comparisons show that our method significantly enhances the robustness of existing DM-based AP, with an average robust accuracy improved by up to 7.30% on CIFAR-10. The code will be available upon acceptance.", "keywords": [ "adversarial attacks", "adversarial training", "adversarial purification" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "Pw7Wb3dGvg", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Sophia Ananiadou", "gender": "Female", "institution": "University of Manchester", "country": "GB", "position": "Full Professor" }, { "name": "ZEPING YU", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 47, "n_ref": 67, "n_ref_all": 91, "n_fig": 11, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 275, "formula_len_all_1": 253, "len_all": 126895, "len_all_1": 67060, "len_abs": 1253, "len_title": 118, "len_sents": 35776, "len_sents_1": 32563, "n_sents": 279, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1258, "title": "Understanding Llava's Visual Question Answering in a Mechanistic View", "abs": "Understanding the mechanisms behind Large Language Models (LLMs) is crucial for designing improved models and strategies. While recent studies have yielded valuable insights into the mechanisms of textual LLMs, the mechanisms of Multi-modal Large Language Models (MLLMs) remain underexplored. In this paper, we apply mechanistic interpretability methods to analyze the visual question answering (VQA) mechanisms in the first MLLM, Llava. We compare the mechanisms between VQA and textual QA (TQA) in color answering tasks and find that: a) VQA exhibits a mechanism similar to the in-context learning mechanism observed in TQA; b) the visual features exhibit significant interpretability when projecting the visual embeddings into the embedding space; and c) Llava enhances the existing capabilities of the corresponding textual LLM Vicuna during visual instruction tuning. Based on these findings, we develop an interpretability tool to help users and researchers identify important visual locations for final predictions, aiding in the understanding of visual hallucination. Our method demonstrates faster and more effective results compared to existing interpretability approaches. Our code, data and interpretability tool will be made available on GitHub.", "keywords": [ "Llava", "multimodal LLM", "mechanistic interpretability", "interpretability tool", "visual question answering" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "PvvXDazPMs", "primary_area": "applications to robotics, autonomy, planning", "n_author": 9, "author_info_list": [ { "name": "Chengyang Zhao", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Hengjun Pu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jifeng Dai", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Ronglei Tong", "gender": "Male", "institution": "Sensetime", "country": "CN", "position": "Researcher" }, { "name": "Tianyi Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuntao Chen", "gender": "Male", "institution": "Centre for Artificial Intelligence and Robotics (CAIR), Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "HK", "position": "Assistant Professor" }, { "name": "Yuwen Xiong", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Zhi Hou", "gender": "unknown", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 17, "n_ref": 73, "n_ref_all": 97, "n_fig": 5, "n_tab": 6, "L_tab": 1311, "n_element_tab": 209, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1242, "n_element_tab_1": 200, "formula_len_all": 70, "formula_len_all_1": 70, "len_all": 107710, "len_all_1": 57272, "len_abs": 1701, "len_title": 76, "len_sents": 31020, "len_sents_1": 26330, "n_sents": 223, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 28, "L_abs": 1309, "title": "Diffusion Transformer Policy", "abs": "Recent large visual-language action models pretrained on diverse robot datasets have demonstrated the potential for generalizing to new environments with a few in-domain data. However, those approaches usually predict discretized or continuous actions by a small action head, which limits the ability in handling diverse action spaces. \nIn contrast, we model the continuous action with a large multi-modal diffusion transformer, dubbed as Diffusion Transformer Policy, in which we directly denoise action chunks by a large transformer model rather than a small action head. \nBy leveraging the scaling capability of transformers, the proposed approach can effectively model continuous end-effector actions across large diverse robot datasets, and achieve better generalization performance. \nExtensive experiments demonstrate Diffusion Transformer Policy pre-trained on diverse robot data can generalize to different embodiments, including simulation environments like Maniskill2 and Calvin, as well as the real-world Franka arm. \nSpecifically, without bells and whistles, the proposed approach achieves state-of-the-art performance in the Calvin novel task setting, and the pre-training stage significantly facilitates the success sequence length on the Calvin by over 1.2. The code will be publicly available.", "keywords": [ "Visual-Language-Action", "Diffusion Policy" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "PvvQlhBbgu", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Emmanuel Dellandrea", "gender": "Male", "institution": "Ecole Centrale de Lyon", "country": "FR", "position": "Associate Professor" }, { "name": "Liming Chen", "gender": "Male", "institution": "Ecole Centrale de Lyon", "country": "FR", "position": "Full Professor" }, { "name": "Matthieu Grard", "gender": "Male", "institution": "Sil\u00e9ane", "country": "", "position": "Researcher" }, { "name": "Rui Yang", "gender": "Male", "institution": "Ecole Centrale de Lyon", "country": "FR", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 16, "n_ref_uni": 22, "n_ref": 48, "n_ref_all": 72, "n_fig": 8, "n_tab": 7, "L_tab": 2675, "n_element_tab": 197, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2076, "formula_len_all_1": 1024, "len_all": 105503, "len_all_1": 56685, "len_abs": 1591, "len_title": 139, "len_sents": 30871, "len_sents_1": 26940, "n_sents": 246, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1596, "title": "Fast Multi-Mode Adaptive Generative Distillation for Continually Learning Diffusion Models", "abs": "Diffusion models are powerful generative models, but their computational demands, vulnerability to catastrophic forgetting, and class imbalance in generated data pose significant challenges in continual learning scenarios. In this paper, we introduce Fast Multi-Mode Adaptive Generative Distillation (MAGD), a novel approach designed to address these three core challenges. MAGD combines generative replay and knowledge distillation, enhancing the continual training of diffusion models through three key innovations: (1) Noisy Intermediate Generative Distillation (NIGD), which leverages intermediate noisy images during the reverse diffusion process to improve data utility and preserve image quality without additional computational costs; (2) Class-guided generative distillation (CGGD), which uses classifier guidance to ensure balanced class representation in generated images, addressing the issue of class imbalance in traditional methods; and (3) Signal-Guided Generative Distillation (SGGD), which reduces computational overhead while maintaining image clarity through the reuse of the model\u2019s denoising capabilities across tasks. Our experimental results on Fashion-MNIST, CIFAR-10, and CIFAR-100 demonstrate that MAGD significantly outperforms existing methods in both image quality, measured by Fr\u00e9chet Inception Distance (FID), and class balance, measured by Kullback-Leibler Divergence (KLD). Moreover, MAGD achieves competitive results with far fewer generation steps compared to traditional methods, making it a practical solution for real-life continual learning applications.", "keywords": [ "diffusion model", "continual learning", "transfer learning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "Pv6fwGPgrA", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Georg Meinhardt", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Intern" }, { "name": "Kai Yi", "gender": "Male", "institution": "KAUST", "country": "", "position": "PhD student" }, { "name": "Laurent Condat", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "research scientist" }, { "name": "Peter Richtarik", "gender": "Male", "institution": "King Abdullah University of Science and Technology (KAUST)", "country": "SA", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 10, "n_ref_uni": 44, "n_ref": 84, "n_ref_all": 103, "n_fig": 8, "n_tab": 5, "L_tab": 1913, "n_element_tab": 189, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1394, "n_element_tab_1": 201, "formula_len_all": 1034, "formula_len_all_1": 570, "len_all": 133822, "len_all_1": 66435, "len_abs": 1193, "len_title": 136, "len_sents": 38650, "len_sents_1": 27483, "n_sents": 364, "n_sents_1": 243, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1201, "title": "Prune at the Clients, Not the Server: Accelerated Sparse Training in Federated Learning", "abs": "In the recent paradigm of Federated Learning (FL), multiple clients train a shared model while keeping their local data private. Resource constraints of clients and communication costs pose major problems for training large models in FL. On the one hand, addressing the resource limitations of the clients, sparse training has proven to be a powerful tool in the centralized setting. On the other hand, communication costs in FL can be addressed by local training, where each client takes multiple gradient steps on its local data. Recent work has shown that local training can provably achieve the optimal accelerated communication complexity [Mishchenko et al., 2022]. Hence, one would like an accelerated sparse training algorithm. In this work we show that naive integration of sparse training and acceleration on the server fails, and how to fix it by letting the clients perform these tasks appropriately. We introduce Sparse-ProxSkip, our method developed for the nonconvex setting, inspired by RandProx [Condat and Richt\u00e1rik, 2022], which provably combines sparse training and acceleration in the convex setting. We demonstrate the good performance of Sparse-ProxSkip in extensive experiments.", "keywords": [ "Federated Learning", "Compression", "Sparsity", "Pruning", "Communication Efficiency", "Local Training" ], "rating_list": [ 5, 5, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4, 4 ] }, { "paper_id": "PulKaNibeQ", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Deyu Meng", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Gaoqin Chang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Jun Shu", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiang Yuan", "gender": "unknown", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 10, "n_ref_uni": 35, "n_ref": 72, "n_ref_all": 109, "n_fig": 10, "n_tab": 8, "L_tab": 1981, "n_element_tab": 245, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1185, "n_element_tab_1": 194, "formula_len_all": 1122, "formula_len_all_1": 810, "len_all": 242501, "len_all_1": 64169, "len_abs": 4240, "len_title": 151, "len_sents": 63355, "len_sents_1": 30435, "n_sents": 513, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1535, "title": "Diversity-Enhanced and Classification-Aware Prompt Learning for Few-Shot Learning via Stable Diffusion", "abs": "Recent text-to-image generative models have exhibited an impressive ability to generate fairly realistic images from some text prompts. In this work, we explore to leverage off-the-shelf text-to-image generative models to train non-specific downstream few-shot classification model architectures using synthetic dataset to classify real images. Current approaches use hand-crafted or model-generated text prompts of text-to-image generative models to generated desired synthetic images, however, they have limited capability of generating diversity images. \nEspecially, their synthetic datasets has relatively limited relevance to the downstream classification tasks. This makes them \nfairly hard to guarantee training models from synthetic images are efficient in practice. To address this issue, we propose a method capable of adaptively learning proper text prompts for the off-the-shelf diffusion\nmodel to generate diverse and classification-aware synthetic images. Our approach shows notable improvements in various\nclassification datasets, with results comparable to existing prompt designing methods. \nWe find that replacing data generation strategy of existing zero/few-shot methods with proposed method could consistly improves downstream classification performance across different network architectures, demostrating its model-agnostic characteristic for few-shot learning. This makes it possible to train an efficient downstream few-shot learning models from synthetic images generated by proposed method for real problems.", "keywords": [ "meta-learning", "synthetic dataset generation", "diffusion model" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "Pujt3ADZgI", "primary_area": "reinforcement learning", "n_author": 9, "author_info_list": [ { "name": "Baolin Peng", "gender": "Male", "institution": "Research, Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Dian Yu", "gender": "Female", "institution": "Tencent AI Lab", "country": "", "position": "NLP researcher" }, { "name": "Dong Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Distinguished Scientist" }, { "name": "Haitao Mi", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Principal Researcher" }, { "name": "Linfeng Song", "gender": "unknown", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Mingyue Huo", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Nan Jiang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Associate Professor" }, { "name": "Ye Tian", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Yuheng Zhang", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 55, "n_formula_1": 28, "n_ref_uni": 54, "n_ref": 112, "n_ref_all": 139, "n_fig": 0, "n_tab": 6, "L_tab": 951, "n_element_tab": 125, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 642, "n_element_tab_1": 91, "formula_len_all": 3626, "formula_len_all_1": 1466, "len_all": 152648, "len_all_1": 66631, "len_abs": 1304, "len_title": 145, "len_sents": 34245, "len_sents_1": 28105, "n_sents": 337, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 97, "L_abs": 1312, "title": "Iterative Nash Policy Optimization: Aligning LLMs with General Preferences via No-Regret Learning", "abs": "Reinforcement Learning with Human Feedback (RLHF) has achieved great success\nin aligning large language models (LLMs) with human preferences. Prevalent\nRLHF approaches are reward-based, following the Bradley-Terry (BT) model assumption, which may not fully capture the complexity of human preferences. In\nthis paper, we explore RLHF under a general preference framework and approach\nit from a game-theoretic perspective. Specifically, we formulate the problem as\na two-player game and propose a novel online algorithm, iterative Nash policy\noptimization (INPO). The key idea is to let the policy play against itself via no-\nregret learning, thereby approximating the Nash policy. Unlike previous methods,\nINPO bypasses the need for estimating the expected win rate for individual responses, which typically incurs high computational or annotation costs. Instead,\nwe introduce a new loss objective that is directly minimized over a preference\ndataset. We provide theoretical analysis for our approach and demonstrate its\neffectiveness through experiments on various representative benchmarks. With an\nLLaMA-3-8B-based SFT model, INPO achieves a 42.6% length-controlled win\nrate on AlpacaEval 2.0 and a 37.8% win rate on Arena-Hard, showing substantial\nimprovement over the state-of-the-art online RLHF algorithms.", "keywords": [ "RLHF Theory", "LLM Alignment" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 2, 4, 4, 5 ] }, { "paper_id": "Pui7Sa6Jwi", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Daiki E. Matsunaga", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "HyeongJoo Hwang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Postdoc" }, { "name": "Kee-Eung Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Tri Wahyu Guntara", "gender": "Male", "institution": "KRAFTON", "country": "KR", "position": "Deep Learning Engineer" } ], "n_formula": 26, "n_formula_1": 10, "n_ref_uni": 40, "n_ref": 110, "n_ref_all": 142, "n_fig": 10, "n_tab": 8, "L_tab": 2515, "n_element_tab": 166, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1099, "n_element_tab_1": 26, "formula_len_all": 4478, "formula_len_all_1": 1240, "len_all": 185639, "len_all_1": 70132, "len_abs": 1421, "len_title": 156, "len_sents": 51859, "len_sents_1": 32265, "n_sents": 376, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1400, "title": "CLEAR: An Information-Theoretic Framework for Distraction-Free Representation Learning in Visual Offline RL", "abs": "Visual offline RL aims to learn an optimal policy for visual domains, solely from the pre-collected dataset comprised of actions taken on visual observations. Prior works on visual RL typically learn a dynamics model by extracting a latent state representation. However, the learned representation would contain factors irrelevant to control when there are distractions in the visual observations. These nuisance factors introduced by the distraction further exacerbates the difficulties of learning a good policy in the offline RL setting. In this work, we formalize the visual offline RL setting as a Partially Observable Markov Decision Process with exogenous variables (ExoPOMDP) and identify these problems with previous approaches under an information-theoretic lens. To overcome these challenges, we propose CLEAR (**C**ontrollable **L**atent State **E**xtr**A**cto**R**) for visual offline RL, which learns the dynamics model of a succinct agent-centric state representation that is consistent with the underlying ExoPOMDP. We empirically demonstrate that CLEAR is able to outperform baselines on the DeepMind Control Suite with various types of distractions and perform consistently well across these distractions. We further provide qualitative analysis on the results showing that our approach successfully disentangles the distraction factors from the agent-centric state representation.", "keywords": [ "Visual Offline Reinforcement Learning", "Information-Theoretic Representation Learning" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "Pu3c0209cx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Laziz Abdullaev", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Rachel Teo", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Stefan Nielsen", "gender": "Male", "institution": "FPT AI", "country": "VN", "position": "Researcher" }, { "name": "Tan Minh Nguyen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 65, "n_formula_1": 20, "n_ref_uni": 37, "n_ref": 80, "n_ref_all": 105, "n_fig": 5, "n_tab": 13, "L_tab": 4227, "n_element_tab": 205, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 3319, "n_element_tab_1": 86, "formula_len_all": 4652, "formula_len_all_1": 1190, "len_all": 186864, "len_all_1": 73339, "len_abs": 1606, "len_title": 87, "len_sents": 58845, "len_sents_1": 29532, "n_sents": 474, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 39, "L_abs": 1599, "title": "Tight Clusters Make Specialized Experts", "abs": "Sparse Mixture-of-Experts (MoE) architectures have emerged as a promising approach to decoupling model capacity from computational cost. At the core of the MoE model is the router, which learns the underlying clustering structure of the input distribution in order to send input tokens to appropriate experts. However, latent clusters may be unidentifiable in high dimension, which causes slow convergence, susceptibility to data contamination, and overall degraded representations as the router is unable to perform appropriate token-expert matching. We examine the router through the lens of clustering optimization and derive optimal feature weights that maximally identify the latent clusters. We use these weights to compute the token-expert routing assignments in an adaptively transformed space that promotes well-separated clusters, which helps identify the best-matched expert for each token. In particular, for each expert cluster, we compute a set of weights that scales features according to whether that expert clusters tightly along that feature. We term this novel router the Adaptive Clustering (AC) router. Our AC router enables the MoE model to obtain three connected benefits: 1) faster convergence, 2) better robustness to data corruption, and 3) overall performance improvement, as experts are specialized in semantically distinct regions of the input space. We empirically demonstrate the advantages of our AC router over baseline routing methods when applied on a variety of MoE backbones for language modeling and image recognition tasks in both clean and corrupted settings.", "keywords": [ "Mixture of Experts", "robustness", "clustering" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "PtnttTKgQw", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Jose Hernandez-Orallo", "gender": "unknown", "institution": "Universitat Politecnica de Valencia", "country": "ES", "position": "Full Professor" }, { "name": "Lorenzo Pacchiardi", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "Lucy G Cheke", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Associate Professor" }, { "name": "Marko Tesic", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 53, "n_ref": 102, "n_ref_all": 124, "n_fig": 4, "n_tab": 3, "L_tab": 2011, "n_element_tab": 172, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2011, "n_element_tab_1": 172, "formula_len_all": 2585, "formula_len_all_1": 19, "len_all": 144669, "len_all_1": 60379, "len_abs": 1691, "len_title": 138, "len_sents": 33903, "len_sents_1": 28662, "n_sents": 205, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1706, "title": "Leaving the barn door open for Clever Hans: Simple features predict LLM benchmark answers", "abs": "The integrity of AI benchmarks is fundamental to accurately assess the capabilities of AI systems. The internal validity of these benchmarks\u2014i.e., making sure they are free from confounding factors\u2014is crucial for ensuring that they are measuring what they are designed to measure. In this paper, we explore a key issue related to internal validity: the possibility that AI systems can solve benchmarks in unintended ways, bypassing the capability being tested. This phenomenon, widely known in human and animal experiments, is often referred to as the \u2018Clever Hans\u2019 effect, where tasks are solved using spurious cues, often involving much simpler processes than those putatively assessed. Previous research suggests that language models can exhibit this behaviour as well. In several older Natural Language Processing (NLP) benchmarks, individual $n$-grams like \u201cnot\u201d have been found to be highly predictive of the correct labels, and supervised NLP models have been shown to exploit these patterns. In this work, we investigate the extent to which simple $n$-grams extracted from benchmark instances can be combined to predict labels in modern multiple-choice benchmarks designed for LLMs, and whether LLMs might be using such $n$-gram patterns to solve these benchmarks. We show how simple classifiers trained on these $n$-grams can achieve high scores on several benchmarks, despite lacking the capabilities being tested. Additionally, we provide evidence that some modern LLMs might be using these superficial patterns to solve benchmarks. This suggests that the internal validity of these benchmarks may be compromised and caution should be exercised when interpreting LLM performance results on them.", "keywords": [ "LLM benchmarks", "Benchmark validity", "Clever Hans effect", "LLM evaluation" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 2, 4, 4 ] }, { "paper_id": "PtgfcMcQd5", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Lav R. Varshney", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "Associate Professor" }, { "name": "Anuj Nayak", "gender": "Male", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 23, "n_ref_uni": 28, "n_ref": 78, "n_ref_all": 93, "n_fig": 6, "n_tab": 1, "L_tab": 176, "n_element_tab": 19, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2829, "formula_len_all_1": 1840, "len_all": 123068, "len_all_1": 64886, "len_abs": 895, "len_title": 97, "len_sents": 40058, "len_sents_1": 28467, "n_sents": 315, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 898, "title": "An Information Theory of Compute-Optimal Size Scaling, Emergence, and Plateaus in Language Models", "abs": "Recent empirical studies show three phenomena with increasing size of language models: compute-optimal size scaling, emergent capabilities, and performance plateauing. We present a simple unified mathematical framework to explain all of these language model scaling phenomena, building on recent skill-text bipartite graph frameworks for semantic learning. Modeling the learning of concepts from texts as an iterative process yields an analogy to iterative decoding of low-density parity check (LDPC) codes in information theory. Thence, drawing on finite-size scaling characterizations of LDPC decoding, we derive the compute-optimal size scaling (Chinchilla rule) for language models. Further, using tools from random network theory, we provide a simple explanation for both emergence of complex skills and plateauing of performance as the size of language models scale. We see multiple plateaus.", "keywords": [ "Language models", "scaling law", "emergence", "plateauing", "Low-Density Parity Check codes", "sequential concept learning", "composition of skills." ], "rating_list": [ 6, 5, 3, 8 ], "soundness_list": [ 3, 2, 1, 4 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "Pt3lfU1NqC", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Alexander Sax", "gender": "unknown", "institution": "Meta (FAIR)", "country": "US", "position": "Researcher" }, { "name": "Alexander Swerdlow", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Ayush Jain", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Franziska Meier", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Katerina Fragkiadaki", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Yuzhou Wang", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Undergrad student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 32, "n_ref": 76, "n_ref_all": 92, "n_fig": 5, "n_tab": 9, "L_tab": 3725, "n_element_tab": 296, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 1359, "n_element_tab_1": 201, "formula_len_all": 407, "formula_len_all_1": 407, "len_all": 157868, "len_all_1": 64275, "len_abs": 2022, "len_title": 125, "len_sents": 40372, "len_sents_1": 30699, "n_sents": 289, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1644, "title": "RODIN: Injecting 2D Foundational Features to 3D Vision Language Understanding", "abs": "We present RODIN (Referential ODIN), a novel model for 3D vision-language understanding that directly operates on posed RGB-D frames. Consuming posed RGB-D from sensors, such as those from an iPhone, simplifies and speeds up inference compared to existing models that train and test using pointclouds sampled from a reconstructed mesh provided by a dataset. We hypothesize that existing approaches consume pointclouds sampled from mesh instead of sensor RGB-D point clouds due to inaccurate camera poses in existing 3D grounding benchmarks, and show that using the \"sensor\" pointclouds indeed leads to a 5-10\\% drop in performance on 3D referential grounding, for these methods. Yet sensor noise is unavoidable in real-world settings. RODIN instead addresses this with a scalable, end-to-end architecture for various 3D vision-language tasks. Specifically, RODIN combines powerful pretrained 2D weights trained on internet-scale data, adapts them to a 2D-3D encoder using the recently proposed ODIN, and combines that backbone with a proposed 3D mask-language decoder based on the Mask2Former used in SAM. RODIN achieves state-of-the-art performance on multiple 3D vision-language benchmarks, including referential grounding (SR3D, NR3D, ScanRefer), language-prompted object detection (ScanNet200 and Matterport3D), and question-answering (ScanQA and SQA3D). It outperforms previous methods for 3D vision-language tasks, despite consuming only sensor inputs. Because of its combination of effectively leveraging 2D pretrained architectures and finetuning end-to-end on sensor data, RODIN provides a scalable solution for embodied 3D perception.", "keywords": [ "3D vision-language understanding" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 2, 4 ] }, { "paper_id": "PstM8YfhvI", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Anne E Carpenter", "gender": "Female", "institution": "Broad Institute", "country": "", "position": "Institute Scientist" }, { "name": "BO WANG", "gender": "Male", "institution": "Vector Institute", "country": "", "position": "Assistant Professor" }, { "name": "Benjamin Haibe-Kains", "gender": "Male", "institution": "Toronto University", "country": "", "position": "Associate Professor" }, { "name": "Beth A Cimini", "gender": "Female", "institution": "Broad Institute", "country": "", "position": "Principal Researcher" }, { "name": "Esteban Miglietta", "gender": "Male", "institution": "Broad Institute", "country": "US", "position": "Postdoc" }, { "name": "Jun Ma", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "Postdoc" }, { "name": "Le Liu", "gender": "Female", "institution": "Broad Institute", "country": "US", "position": "Postdoc" }, { "name": "Zeinab Navidi", "gender": "Female", "institution": "University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 38, "n_ref": 74, "n_ref_all": 106, "n_fig": 13, "n_tab": 4, "L_tab": 2813, "n_element_tab": 209, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 243, "formula_len_all_1": 267, "len_all": 145164, "len_all_1": 56103, "len_abs": 1207, "len_title": 110, "len_sents": 46907, "len_sents_1": 29059, "n_sents": 282, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 62, "L_abs": 1197, "title": "MorphoDiff: Cellular Morphology Painting with Diffusion Models", "abs": "Understanding cellular responses to external stimuli is critical for parsing biological mechanisms and advancing therapeutic development. High-content image-based assays provide a cost-effective approach to examine cellular phenotypes induced by diverse interventions, which offers valuable insights into biological processes and cellular states. We introduce MorphoDiff, a generative pipeline to predict high-resolution cell morphological responses under different conditions based on perturbation encoding. To the best of our knowledge, MorphoDiff is the first framework capable of producing guided, high-resolution predictions of cell morphology that generalize across both chemical and genetic interventions. The model integrates perturbation embeddings as guiding signals within a 2D latent diffusion model. The comprehensive computational, biological, and visual validations across three open-source Cell Painting datasets show that MorphoDiff can generate high-fidelity images and produce meaningful biology signals under various interventions. We envision the model will facilitate efficient in silico exploration of perturbational landscapes towards more effective drug discovery studies.", "keywords": [ "Generative Modelling", "Latent Diffusion Model", "Cell Painting", "Morphology", "Drug Response Prediction", "Cellular Phenotype", "Machine Learning" ], "rating_list": [ 8, 8, 3, 8 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "PsaRfjbfnv", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 14, "author_info_list": [ { "name": "Abhinav Kommula", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "Undergrad student" }, { "name": "Adam Dziedzic", "gender": "Male", "institution": "CISPA, saarland university, saarland informatics campus", "country": "DE", "position": "Assistant Professor" }, { "name": "Akul Arora", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "MS student" }, { "name": "Dan Hendrycks", "gender": "unknown", "institution": "UC Berkeley", "country": "", "position": "PhD student" }, { "name": "Daniel Kang", "gender": "unknown", "institution": "Department of Computer Science", "country": "US", "position": "Assistant Professor" }, { "name": "Franziska Boenisch", "gender": "Female", "institution": "CISPA", "country": "DE", "position": "Assistant Professor" }, { "name": "Jacob Steinhardt", "gender": "unknown", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" }, { "name": "Mantas Mazeika", "gender": "Male", "institution": "Center for AI Safety", "country": "US", "position": "Researcher" }, { "name": "Oliver Zhang", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Steven Basart", "gender": "Male", "institution": "Center for AI Safety ", "country": "US", "position": "Researcher" }, { "name": "Tom B Brown", "gender": null, "institution": "", "country": "", "position": "" }, { "name": "Xuwang Yin", "gender": "Male", "institution": "Center for AI Safety", "country": "US", "position": "Researcher" }, { "name": "Yi Sun", "gender": "unknown", "institution": "University of Chicago", "country": "", "position": "Assistant Professor" }, { "name": "Max Kaufmann", "gender": "unknown", "institution": "University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 38, "n_ref": 74, "n_ref_all": 97, "n_fig": 8, "n_tab": 17, "L_tab": 5572, "n_element_tab": 447, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 369, "n_element_tab_1": 53, "formula_len_all": 186, "formula_len_all_1": 33, "len_all": 162321, "len_all_1": 53761, "len_abs": 873, "len_title": 114, "len_sents": 42565, "len_sents_1": 24239, "n_sents": 328, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 875, "title": "Evaluating Model Robustness Against Unforeseen Adversarial Attacks", "abs": "When considering real-world adversarial settings, defenders are unlikely to have access to the full range of deployment-time adversaries, and adversaries are likely to use realistic adversarial distortions that will not be limited to small $L_p$-constrained perturbations. To narrow in on this discrepancy between research and reality we introduce ImageNet-UA, a new benchmark for evaluating model robustness against a wide range of unforeseen adversaries. We make use of our benchmark to identify holes in current popular adversarial defense techniques, highlighting a rich space of techniques which can improve unforeseen robustness. We hope the greater variety and realism of ImageNet-UA will make it a useful tool for those working on real-world worst-case robustness, enabling development of more robust defenses which can generalize beyond attacks seen during training.", "keywords": [ "AI Safety", "ML safety", "adversarial robustness", "distribution shift", "unforeseen adversaries" ], "rating_list": [ 3, 8, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 4, 2, 2 ], "contribution_list": [ 1, 4, 1, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "Pr4JkJVlmz", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "S Akash", "gender": "unknown", "institution": "Indian Institute of Technology, Patna", "country": "IN", "position": "Undergrad student" }, { "name": "Tejas Bodas", "gender": "Male", "institution": "International Institute of Information Technology Hyderabad, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 21, "n_ref": 25, "n_ref_all": 31, "n_fig": 1, "n_tab": 1, "L_tab": 47, "n_element_tab": 2, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 47, "n_element_tab_1": 2, "formula_len_all": 1281, "formula_len_all_1": 1263, "len_all": 68259, "len_all_1": 47098, "len_abs": 913, "len_title": 96, "len_sents": 21699, "len_sents_1": 21488, "n_sents": 150, "n_sents_1": 150, "venue": "ICLR2025", "decision": "withdraw", "L_title": 48, "L_abs": 923, "title": "MAQL: Speeding up Q-learning with a model-assist", "abs": "In reinforcement learning, model free methods such as Q-learning and policy gradient are extremely popular due to their simplicity but require a huge amount of data for training. Model based methods on the other hand, are proven to be sample efficient in various environments but are unfortunately computationally expensive. It is therefore only prudent to investigate and design algorithms that have best of features from both these classes of algorithms. In this work, we propose MAQL, a model-assisted Q-learning algorithm that is not only computationally inexpensive but also offers low sample complexity. We illustrate its superior performance to vanilla Q-learning in various RL environments and particularly demonstrate its utility in learning the Gittins/Whittles index in Rested/Restless Bandits respectively. We aim to spur discussion on how model-assists can help boost the performance of existing RL algorithms.", "keywords": [ "bandits", "reinforcement learning" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "PqiDHCLkB9", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Faiza Amin", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Huanhuan Chen", "gender": "Male", "institution": "University of Science and Technology of China, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jinlong Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Qiuju Chen", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" }, { "name": "Xinyu Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "Undergrad student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 29, "n_ref": 66, "n_ref_all": 90, "n_fig": 4, "n_tab": 4, "L_tab": 1714, "n_element_tab": 248, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 3984, "n_element_tab_1": 238, "formula_len_all": 698, "formula_len_all_1": 670, "len_all": 106786, "len_all_1": 63931, "len_abs": 1098, "len_title": 172, "len_sents": 29376, "len_sents_1": 26723, "n_sents": 190, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 124, "L_abs": 1069, "title": "Characterizing trainability, expressivity, and generalization of neural architecture with metrics from neural tangent kernel", "abs": "Zero-shot neural architecture search aims to predict multiple characteristics of neural architectures using proxy indicators without actual training, yet most methods focus on evaluating only a single characteristic of neural networks.\nSince the Neural Tangent Kernel (NTK) offers a promising theoretical framework for understanding the characteristics of neural networks, we propose NTK-score, including three metrics derived from NTK's eigenvalues and kernel regression, to assess three critical characteristics: trainability, expressivity, and generalization. \nMoreover, to exploit three metrics of our NTK-score, we employ the Borda Count approach on our NTK-score to rank architectures in neural architecture search.\nCompared with state-of-the-art proxies, experimental results demonstrate that the NTK-score correlates well with both the accuracy and training time of architectures, and exhibits excellent performance across various search spaces and methods, including NAS-bench-201, DARTS, and ResNet, as well as pruning, reinforce, and evolutionary algorithm.", "keywords": [ "neural tangential kernel", "neural architecture search", "trainability", "expressivity", "generalization" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "PqeMnyGU1B", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Ashkan Shahbazi", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "PhD student" }, { "name": "Huy Tran", "gender": "Male", "institution": "Vanderbilt University", "country": "", "position": "PhD student" }, { "name": "Soheil Kolouri", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "Assistant Professor" }, { "name": "Yikun Bai", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "Postdoc" }, { "name": "john r hershey", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 168, "n_formula_1": 45, "n_ref_uni": 37, "n_ref": 75, "n_ref_all": 87, "n_fig": 2, "n_tab": 18, "L_tab": 27540, "n_element_tab": 1983, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 17360, "formula_len_all_1": 3303, "len_all": 256901, "len_all_1": 69536, "len_abs": 1484, "len_title": 134, "len_sents": 50209, "len_sents_1": 26733, "n_sents": 557, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1518, "title": "Understanding Learning with Sliced-Wasserstein Requires Re-thinking Informative Slices", "abs": "The practical applications of Wasserstein distances (WDs) are constrained by their sample and computational complexities. Sliced-Wasserstein distances (SWDs) provide a workaround by projecting distributions onto one-dimensional subspaces, leveraging the more efficient, closed-form WDs for one-dimensional distributions. However, in high dimensions, most random projections become uninformative due to the concentration of measure phenomenon. Although several SWD variants have been proposed to focus on \\textit{informative} slices, they often introduce additional complexity, numerical instability, and compromise desirable theoretical (metric) properties of SWD. Amidst the growing literature that focuses on directly modifying the slicing distribution, which often face challenges, we revisit the classic Sliced-Wasserstein and propose instead to rescale the 1D Wasserstein to make all slices equally informative. Importantly, we show that with an appropriate notion of \\textit{slice informativeness}, rescaling for all individual slices simplifies to \\textbf{a single global scaling factor} on the SWD. This, in turn, translates to the standard learning rate search for gradient-based learning in common ML workflows. We perform extensive experiments across various machine learning tasks showing that the classic SWD, when properly configured, can often match or surpass the performance of more complex variants. We then answer the following question: Is Sliced-Wasserstein all you need for common learning tasks?", "keywords": [ "Optimal Transport", "Sliced Wasserstein", "Concentration of Measure" ], "rating_list": [ 5, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "Pq2yEKXOl7", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Chaoqi Chen", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Kunze Huang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Luyao Tang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Xinghao Ding", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Yue Huang", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Yuxuan Yuan", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 32, "n_ref": 53, "n_ref_all": 69, "n_fig": 10, "n_tab": 3, "L_tab": 2341, "n_element_tab": 354, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1190, "n_element_tab_1": 211, "formula_len_all": 389, "formula_len_all_1": 392, "len_all": 114721, "len_all_1": 62021, "len_abs": 1438, "len_title": 149, "len_sents": 30735, "len_sents_1": 26859, "n_sents": 222, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1448, "title": "SlotSAM: Bootstrap Segmentation Foundation Model under Real-world Shifts via Object-Centric Learning", "abs": "Foundation models have made incredible strides in achieving zero-shot or few-shot generalization, leveraging prompt engineering to mimic the problem-solving approach of human intelligence. However, when it comes to some foundation models like Segment Anything, there is still a challenge in performing well under real-world shifts. One of the real-world shifts is the distribution shift, the out-of-distribution data, such as camouflaged and medical images. Another is inconsistent prompting strategies during fine-tuning and testing, leading to decreased performance. We draw inspiration from human intelligence, particularly the process by which individuals decompose scenes into components in unfamiliar environments to determine the positions or boundaries of each component. To this end, we introduce SlotSAM, a method that reconstructs features from the encoder in a self-supervised manner to create object-centric representations. These representations are then integrated into the foundation model, bolstering its object-level perceptual capabilities while reducing the impact of distribution-related variables. The beauty of SlotSAM lies in its simplicity and adaptability to various tasks, making it a versatile solution that significantly enhances the generalization abilities of foundation models. Through limited parameter fine-tuning in a bootstrap manner, our approach paves the way for improved generalization in novel environments.", "keywords": [ "segmentation foundation model", "distribution shift", "object-centric learning", "weakly supervised" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "PplM2kDrl3", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Jianmin Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Jincheng Zhong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "XiangCheng Zhang", "gender": "Male", "institution": " Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Mingsheng Long", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 10, "n_ref_uni": 30, "n_ref": 84, "n_ref_all": 111, "n_fig": 4, "n_tab": 10, "L_tab": 2576, "n_element_tab": 179, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1135, "n_element_tab_1": 73, "formula_len_all": 1436, "formula_len_all_1": 681, "len_all": 128930, "len_all_1": 61594, "len_abs": 3673, "len_title": 125, "len_sents": 39859, "len_sents_1": 28187, "n_sents": 301, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1246, "title": "Domain Guidance: A Simple Transfer Approach for a Pre-trained Diffusion Model", "abs": "Recent advancements in diffusion models have revolutionized generative modeling. However, the impressive and vivid outputs they produce often come at the cost of significant model scaling and increased computational demands. Consequently, building personalized diffusion models based on off-the-shelf models has emerged as an appealing alternative. In this paper, we introduce a novel perspective on conditional generation for transferring a pre-trained model. From this viewpoint, we propose *Domain Guidance*, a straightforward transfer approach that leverages pre-trained knowledge to guide the sampling process toward the target domain. Domain Guidance shares a formulation similar to advanced classifier-free guidance, facilitating better domain alignment and higher-quality generations. We provide both empirical and theoretical analyses of the mechanisms behind Domain Guidance. Our experimental results demonstrate its substantial effectiveness across various transfer benchmarks, achieving over a 19.6\\% improvement in FID and a 20.6\\% improvement in FD$_\\text{DINOv2}$ compared to standard fine-tuning. Notably, existing fine-tuned models can seamlessly integrate Domain Guidance to leverage these benefits, without additional training.", "keywords": [ "transfer learning", "diffusion models", "fine-tuning", "guidance" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 4, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "PpYy0dR3Qw", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Arto Maranjyan", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Laurent Condat", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "research scientist" }, { "name": "Peter Richtarik", "gender": "Male", "institution": "King Abdullah University of Science and Technology (KAUST)", "country": "SA", "position": "Full Professor" } ], "n_formula": 41, "n_formula_1": 28, "n_ref_uni": 26, "n_ref": 68, "n_ref_all": 87, "n_fig": 6, "n_tab": 4, "L_tab": 2030, "n_element_tab": 146, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 160, "n_element_tab_1": 11, "formula_len_all": 3954, "formula_len_all_1": 2201, "len_all": 139760, "len_all_1": 68987, "len_abs": 893, "len_title": 88, "len_sents": 34950, "len_sents_1": 27785, "n_sents": 326, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 88, "L_abs": 905, "title": "LoCoDL: Communication-Efficient Distributed Learning with Local Training and Compression", "abs": "In $D$istributed optimization and $L$earning, and even more in the modern framework of federated learning, communication, which is slow and costly, is critical. We introduce LoCoDL, a communication-efficient algorithm that leverages the two popular and effective techniques of $Lo$cal training, which reduces the communication frequency, and $Co$mpression, in which short bitstreams are sent instead of full-dimensional vectors of floats. LoCoDL works with a large class of unbiased compressors that includes widely-used sparsification and quantization methods. LoCoDL provably benefits from local training and compression and enjoys a doubly-accelerated communication complexity, with respect to the condition number of the functions and the model dimension, in the general heterogeneous regime with strongly convex functions. This is confirmed in practice, with LoCoDL outperforming existing algorithms.", "keywords": [ "distributed optimization", "local training", "compression", "communication-efficient algorithm", "federated learning" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "PpP6ALezeK", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Bohan Lyu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Haibing Guan", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Full Professor" }, { "name": "Hao Wang", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Jiaru Zhang", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Postdoc" }, { "name": "Tao Song", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaoyu Wu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yang Hua", "gender": "Male", "institution": "Queen's University Belfast", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 27, "n_formula_1": 7, "n_ref_uni": 32, "n_ref": 92, "n_ref_all": 124, "n_fig": 18, "n_tab": 5, "L_tab": 2137, "n_element_tab": 246, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 861, "n_element_tab_1": 104, "formula_len_all": 1963, "formula_len_all_1": 535, "len_all": 145418, "len_all_1": 69360, "len_abs": 1671, "len_title": 161, "len_sents": 48614, "len_sents_1": 29935, "n_sents": 408, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 113, "L_abs": 1511, "title": "Exploring Diffusion Models' Corruption Stage in Few-Shot Fine-tuning and Mitigating with Bayesian Neural Networks", "abs": "Few-shot fine-tuning of Diffusion Models (DMs) is a key advancement, significantly reducing training costs and enabling personalized AI applications. However, we explore the training dynamics of DMs and observe an unanticipated phenomenon: during the training process, image fidelity initially improves, then unexpectedly deteriorates with the emergence of noisy patterns, only to recover later with severe overfitting. We term the stage with generated noisy patterns as corruption stage. To understand this corruption stage, we begin by heuristically modeling the one-shot fine-tuning scenario, and then extend this modeling to more general cases. Through this modeling, we identify the primary cause of this corruption stage: a narrowed learning distribution inherent in the nature of few-shot fine-tuning. To tackle this, we apply Bayesian Neural Networks (BNNs) on DMs with variational inference to implicitly broaden the learned distribution, and present that the learning target of the BNNs can be naturally regarded as an expectation of the diffusion loss and a further regularization with the pretrained DMs. This approach is highly compatible with current few-shot fine-tuning methods in DMs and does not introduce any extra inference costs. Experimental results demonstrate that our method significantly mitigates corruption, and improves the fidelity, quality and diversity of the generated images in both object-driven and subject-driven generation tasks. The code is available at an anonymous link.", "keywords": [ "Few-shot Fine-tuning", "Diffusion Models", "Bayesian Neural Networks" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "PoSq0B0ffE", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Dogukan Yigit Polat", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Kaiwen Tang", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Shida Wang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Weng-Fai Wong", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" }, { "name": "Zhanglu Yan", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" } ], "n_formula": 22, "n_formula_1": 23, "n_ref_uni": 18, "n_ref": 30, "n_ref_all": 43, "n_fig": 5, "n_tab": 5, "L_tab": 2326, "n_element_tab": 386, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2098, "n_element_tab_1": 342, "formula_len_all": 1319, "formula_len_all_1": 1391, "len_all": 83293, "len_all_1": 59329, "len_abs": 1232, "len_title": 116, "len_sents": 26432, "len_sents_1": 24965, "n_sents": 185, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1238, "title": "Improving model robustness against noise with safe haven activations", "abs": "Quantized neural networks (QNNs) are often used in edge AI because they reduce memory and computational demands. In practical applications such as control systems, medical imaging, and robotics, controlling input noise is crucial for enhancing system robustness. Thus, improving the noise resilience of QNNs is an important challenge in achieving effective edge AI applications. In this paper, we investigate the impact of input noise on QNN performance and propose the safe haven activation quantization (SHAQ) method. This approach leverages the characteristics of the quantization function to constrain outputs before quantization within a more noise-resilient 'safe' range, effectively reducing the impact of noise across quantized layers. Our methods achieve state-of-the-art, 73.11\\% accuracy with 2-bit activations under the fast gradient sign method (FGSM) adversarial attacks with an epsilon of 8/255 on the CIFAR-10 dataset. Furthermore, we extend our methods into a plug-and-play solution we call quantized helmet (QH), comprising a series of quantized layers that can be integrated into any unquantized neural network to enhance its noise robustness. Our experimental code and analysis are open-source and publicly accessible.", "keywords": [ "Quantized Neural Networks", "Noise Robustness" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 2, 1 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "PoL2joPZQ4", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Dongseok Shim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "H. Jin Kim", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Kejie Li", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "peng wang", "gender": "Male", "institution": "Bytedance US AILab", "country": "US", "position": "Research Scientist" }, { "name": "Yichun Shi", "gender": "Male", "institution": "ByteDance", "country": "US", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 26, "n_ref": 80, "n_ref_all": 91, "n_fig": 8, "n_tab": 1, "L_tab": 122, "n_element_tab": 12, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 122, "n_element_tab_1": 12, "formula_len_all": 262, "formula_len_all_1": 262, "len_all": 102038, "len_all_1": 54841, "len_abs": 1039, "len_title": 133, "len_sents": 32738, "len_sents_1": 27575, "n_sents": 199, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 995, "title": "MVLight: Relightable Text-to-3D Generation via Light-conditioned Multi-View Diffusion", "abs": "Recent advancements in text-to-3D generation, building on the success of high-performance text-to-image generative models, have made it possible to create imaginative and richly textured 3D objects from textual descriptions. However, a key challenge remains in effectively decoupling light-independent and lighting-dependent components to enhance the quality of generated 3D models and their relighting performance. In this paper, we present MVLight, a novel light-conditioned multi-view diffusion model that explicitly integrates lighting conditions directly into the generation process. This enables the model to synthesize high-quality images that faithfully reflect the specified lighting environment across multiple camera views. By leveraging this capability to Score Distillation Sampling (SDS), we can effectively synthesize 3D models with improved geometric precision and relighting capabilities. We validate the effectiveness of MVLight through extensive experiments and a user study.", "keywords": [ "Text-to-3D generation", "Multi-view diffusion", "PBR material" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Pnr8XNWcY0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "benyou wang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "FanBu", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Intern" }, { "name": "Haizhou Li", "gender": "Male", "institution": "The Chinese University of Hong Kong (Shenzhen); National University of Singapore", "country": "CN", "position": "Full Professor" }, { "name": "Qun Liu", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "HK", "position": "Chief Scientist of Speech and Language Computing" }, { "name": "Xidong Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Yuhao Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Postdoc" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 90, "n_ref_all": 103, "n_fig": 6, "n_tab": 16, "L_tab": 8644, "n_element_tab": 669, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3718, "n_element_tab_1": 331, "formula_len_all": 354, "formula_len_all_1": 0, "len_all": 227754, "len_all_1": 64295, "len_abs": 1142, "len_title": 124, "len_sents": 65417, "len_sents_1": 26748, "n_sents": 527, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1158, "title": "Roadmap towards Superhuman Speech Understanding using Large Language Models", "abs": "The success of large language models (LLMs) has prompted efforts to integrate speech and audio data, aiming to create general foundation models capable of processing both textual and non-textual inputs. Recent advances, such as GPT-4o, highlight the potential for end-to-end speech LLMs, which preserves non-semantic information and world knowledge for deeper speech understanding.\nTo guide the development of speech LLMs, we propose a five-level roadmap, ranging from basic automatic speech recognition (ASR) to advanced superhuman models capable of integrating non-semantic information with abstract acoustic knowledge for complex tasks. \nMoreover, we design a benchmark, \\textbf{SAGI Bechmark}, that standardizes critical aspects across various tasks in these five levels, uncovering challenges in using abstract acoustic knowledge and completeness of capability. Our findings reveal gaps in handling paralinguistic cues and abstract acoustic knowledge, and we offer future directions. This paper outlines a roadmap for advancing speech LLMs, introduces a benchmark for evaluation, and provides key insights into their current limitations and potential.", "keywords": [ "Large langauge models; speech language models;" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Pnktu2PBXD", "primary_area": "generative models", "n_author": 12, "author_info_list": [ { "name": "Chenyan Xiong", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Associate Professor" }, { "name": "Ge Yu", "gender": "Male", "institution": "Northeastern University, China", "country": "CN", "position": "Full Professor" }, { "name": "Hao Chen", "gender": "unknown", "institution": "Northeastern University", "country": "", "position": "MS student" }, { "name": "Maosong Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Sen Mei", "gender": "Male", "institution": "Northeastern University", "country": "CN", "position": "MS student" }, { "name": "Shi Yu", "gender": "unknown", "institution": "Department of Computer Science and Technology, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Shuo Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Xinze Li", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "MS student" }, { "name": "Yukun Yan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Zhenghao Liu", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "Associate Professor" }, { "name": "Zheni Zeng", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Zhiyuan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 43, "n_ref": 111, "n_ref_all": 134, "n_fig": 7, "n_tab": 6, "L_tab": 3006, "n_element_tab": 412, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1195, "n_element_tab_1": 207, "formula_len_all": 565, "formula_len_all_1": 512, "len_all": 171290, "len_all_1": 67968, "len_abs": 1607, "len_title": 133, "len_sents": 40941, "len_sents_1": 30603, "n_sents": 303, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1615, "title": "RAG-DDR: Optimizing Retrieval-Augmented Generation Using Differentiable Data Rewards", "abs": "Retrieval-Augmented Generation (RAG) has proven its effectiveness in mitigating hallucinations in Large Language Models (LLMs) by retrieving knowledge from external resources. To adapt LLMs for RAG pipelines, current approaches use instruction tuning to optimize LLMs, improving their ability to utilize retrieved knowledge. This supervised fine-tuning (SFT) approach focuses on equipping LLMs to handle diverse RAG tasks using different instructions. However, it trains RAG modules to overfit training signals and overlooks the varying data preferences among agents within the RAG system. In this paper, we propose a Differentiable Data Rewards (DDR) method, which end-to-end trains RAG systems by aligning data preferences between different RAG modules. DDR works by collecting the rewards to optimize each agent with a rollout method. This method prompts agents to sample some potential responses as perturbations, evaluates the impact of these perturbations on the whole RAG system, and subsequently optimizes the agent to produce outputs that improve the performance of the RAG system. Our experiments on various knowledge-intensive tasks demonstrate that DDR significantly outperforms the SFT method, particularly for LLMs with smaller-scale parameters that depend more on the retrieved knowledge. Additionally, DDR exhibits a stronger capability to align the data preference between RAG modules. The DDR method makes generation module more effective in extracting key information from documents and mitigating conflicts between parametric memory and external knowledge. All codes will be released via GitHub.", "keywords": [ "Retrieval-Augmented Generation" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "Pnk7vMbznK", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Bill Y Lin", "gender": "Male", "institution": "xAI", "country": "US", "position": "Researcher" }, { "name": "Fengqing Jiang", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Luyao Niu", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Postdoc" }, { "name": "Radha Poovendran", "gender": "unknown", "institution": "University of Washington, Seattle", "country": "", "position": "Full Professor" }, { "name": "Yejin Choi", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Yuntian Deng", "gender": "unknown", "institution": "University of Waterloo", "country": "", "position": "Assistant Professor" }, { "name": "Zhangchen Xu", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 47, "n_ref": 156, "n_ref_all": 210, "n_fig": 19, "n_tab": 16, "L_tab": 6422, "n_element_tab": 919, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 394, "n_element_tab_1": 38, "formula_len_all": 516, "formula_len_all_1": 168, "len_all": 296219, "len_all_1": 72553, "len_abs": 2049, "len_title": 84, "len_sents": 74361, "len_sents_1": 34843, "n_sents": 605, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 2062, "title": "Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing", "abs": "High-quality instruction data is critical for aligning large language models (LLMs). Although some models, such as Llama-3-Instruct, have open weights, their alignment data remain private, which hinders the democratization of AI. High human labor costs and a limited, predefined scope for prompting prevent existing open-source data creation methods from scaling effectively, potentially limiting the diversity and quality of public alignment datasets. Is it possible to synthesize high-quality instruction data at scale by extracting it directly from an aligned LLM? We present a self-synthesis method for generating large-scale alignment data named Magpie. Our key observation is that aligned LLMs like Llama-3-Instruct can generate a user query when we input only the pre-query templates up to the position reserved for user messages, thanks to their auto-regressive nature. We use this method to prompt Llama-3-Instruct and generate 4 million instructions along with their corresponding responses. We further introduce extensions of Magpie for filtering, generating multi-turn, preference optimization, domain-specific and multilingual datasets. We perform a comprehensive analysis of the Magpie-generated data. To compare Magpie-generated data with other public instruction datasets (e.g., ShareGPT, WildChat, Evol-Instruct, UltraChat, OpenHermes, Tulu-V2-Mix, GenQA), we fine-tune Llama-3-8B-Base with each dataset and evaluate the performance of the fine-tuned models. Our results indicate that using Magpie for supervised fine-tuning (SFT) solely can surpass the performance of previous public datasets utilized for both SFT and preference optimization, such as direct preference optimization with UltraFeedback. We also show that in some tasks, models supervised fine-tuned with Magpie perform comparably to the official Llama-3-8B-Instruct, despite the latter being enhanced with 10 million data points through SFT and subsequent preference optimization. This advantage is evident on alignment benchmarks such as AlpacaEval, ArenaHard, and WildBench.", "keywords": [ "Dataset", "LLM", "alignment", "synthetic", "supervised fine-tuning" ], "rating_list": [ 6, 8, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 4, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "PnfghHD4Pi", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Haolin Pan", "gender": "unknown", "institution": "South China University of Technology", "country": "", "position": "MS student" }, { "name": "Jian Chen", "gender": "unknown", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Jing Liu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Shulian Zhang", "gender": "Female", "institution": "South China University of Technology", "country": "", "position": "PhD student" }, { "name": "Yong Guo", "gender": "Male", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Postdoc" }, { "name": "Yulun Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 29, "n_formula_1": 12, "n_ref_uni": 64, "n_ref": 156, "n_ref_all": 191, "n_fig": 9, "n_tab": 11, "L_tab": 2377, "n_element_tab": 192, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 624, "n_element_tab_1": 36, "formula_len_all": 2712, "formula_len_all_1": 1010, "len_all": 212656, "len_all_1": 69908, "len_abs": 1340, "len_title": 134, "len_sents": 54751, "len_sents_1": 32278, "n_sents": 400, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1367, "title": "Gap Preserving Distillation by Building Bidirectional Mappings with A Dynamic Teacher", "abs": "Knowledge distillation aims to transfer knowledge from a large teacher model to a compact student counterpart, often coming with a significant performance gap between them. Interestingly, we find that a too-large performance gap can hamper the training process.\nTo alleviate this, we propose a **Gap Preserving Distillation (GPD)** method that trains an additional dynamic teacher model from scratch along with the student to maintain a reasonable performance gap. To further strengthen distillation, we develop a hard strategy by enforcing both models to share parameters. Besides, we also build the soft bidirectional mappings between them through ***Inverse Reparameterization (IR)*** and ***Channel-Branch Reparameterization (CBR)***.\nIR initializes a larger dynamic teacher with approximately the same accuracy as the student to avoid a too large gap in early stage of training. CBR enables direct extraction of an effective student model from the dynamic teacher without post-training. \nIn experiments, GPD significantly outperforms existing distillation methods on top of both CNNs and transformers, achieving up to 1.58\\% accuracy improvement. \nInterestingly, GPD also generalizes well to the scenarios without a pre-trained teacher, including training from scratch and fine-tuning, yielding a large improvement of 1.80\\% and 0.89\\% on ResNet18, respectively.", "keywords": [ "Knowledge Distillation; Model Expansion; Reparameterization" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "PnZ2lbQaao", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 8, "author_info_list": [ { "name": "Bo Long", "gender": "Male", "institution": "Meta", "country": "US", "position": "Principal Researcher" }, { "name": "Qi Xu", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Rohit Amarnath", "gender": "Male", "institution": ", Rutgers University", "country": "US", "position": "Undergrad student" }, { "name": "Shuang Yang", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Research Scientist" }, { "name": "Yan Xie", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Zhigang Hua", "gender": "Male", "institution": "Meta", "country": "US", "position": "ML engineer" }, { "name": "Zihao Xu", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Hao Wang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 21, "n_formula_1": 21, "n_ref_uni": 39, "n_ref": 68, "n_ref_all": 84, "n_fig": 5, "n_tab": 4, "L_tab": 1139, "n_element_tab": 178, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 674, "n_element_tab_1": 104, "formula_len_all": 1331, "formula_len_all_1": 1331, "len_all": 112676, "len_all_1": 60171, "len_abs": 759, "len_title": 111, "len_sents": 28570, "len_sents_1": 25312, "n_sents": 253, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 762, "title": "Domain Indexing Collaborative Filtering for Recommender System", "abs": "In cross-domain recommendation systems, addressing cold-start items remains a significant challenge. Previous methods typically focus on maximizing performance using cross-domain knowledge, often treating the knowledge transfer process as a black box. However, the recent development of domain indexing introduces a new approach to better address such challenges. We have developed an adversarial Bayesian framework, Domain Indexing Collaborative Filtering (DICF), that infers domain indices during cross-domain recommendation. This framework not only significantly improves the recommendation performance but also provides interpretability for cross-domain knowledge transfer. This is verified by our empirical results on both synthetic and real-world datasets.", "keywords": [ "Recommendation System", "Domain Adaptation", "Bayesian Deep Learning" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "PnQJ24n1qq", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Feng Wu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Fuying Wang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Guosheng Yin", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Lequan Yu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Tsai Hor Chan", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 11, "n_ref_uni": 41, "n_ref": 86, "n_ref_all": 107, "n_fig": 5, "n_tab": 11, "L_tab": 4719, "n_element_tab": 315, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2866, "n_element_tab_1": 134, "formula_len_all": 2976, "formula_len_all_1": 1040, "len_all": 153011, "len_all_1": 69246, "len_abs": 1967, "len_title": 103, "len_sents": 42276, "len_sents_1": 29459, "n_sents": 345, "n_sents_1": 244, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 2023, "title": "Cross-Modal Alignment via Variational Copula Modelling", "abs": "Various data modalities are common in real-world applications. In healthcare, for example, electronic health records, medical images, and clinical notes provide comprehensive information for diagnosis and treatment.\n Thus, it is essential to develop multimodal learning methods that aggregate information from multiple modalities to generate meaningful representations for downstream tasks.\n The key challenge here is how to appropriately align the representations of the respective modalities and fuse them into a joint distribution.\n Existing methods mainly focus on fusing the representations via concatenation or the Kronecker product, which oversimplifies the interaction structure between modalities, prompting the need to model more complex interactions.\n Moreover, the notion of joint distribution of the latent representation that incorporates higher-order interactions between modalities is also underexplored.\n Copula is a powerful statistical structure in modelling the interactions between variables, as it bridges the joint distribution and marginal distributions of multiple variables.\n In this paper, we propose a novel copula modelling-driven multimodal learning framework, which focuses on learning the joint distribution of various modalities to capture the complex interaction among them.\n The key idea is interpreting the copula model as a tool to align the marginal distributions of the modalities efficiently. \n By assuming a Gaussian mixture distribution for each modality and a copula model on the joint distribution, our model can also generate accurate representations for missing modalities.\n Extensive experiments on public MIMIC datasets demonstrate the superior performance of our model over other competitors.\n Ablation studies also validate the effectiveness of the copula alignment strategy and the robustness of our model over different choices of the copula family. \n Code is anonymously available at https://anonymous.4open.science/r/CM2-C1FD/README.md.", "keywords": [ "Copula", "Multimodal learning", "Missing modality" ], "rating_list": [ 3, 8, 6, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "PmV9oPAtU9", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Alain Ryser", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Emanuele Palumbo", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Imant Daunhawer", "gender": "unknown", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Julia E Vogt", "gender": "Female", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Assistant Professor" }, { "name": "Moritz Vandenhirtz", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 78, "n_ref_all": 108, "n_fig": 11, "n_tab": 6, "L_tab": 8819, "n_element_tab": 31521, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1040, "n_element_tab_1": 193, "formula_len_all": 992, "formula_len_all_1": 710, "len_all": 366447, "len_all_1": 65566, "len_abs": 1043, "len_title": 111, "len_sents": 42866, "len_sents_1": 30385, "n_sents": 302, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1046, "title": "From Logits to Hierarchies: Hierarchical Clustering made Simple", "abs": "The structure of many real-world datasets is intrinsically hierarchical, making the modeling of such hierarchies a critical objective in both unsupervised and supervised machine learning. Recently, novel approaches for hierarchical clustering with deep architectures have been proposed. In this work, we take a critical perspective on this line of research and demonstrate that many approaches exhibit major limitations when applied to realistic datasets, partly due to their high computational complexity. In particular, we show that a lightweight procedure implemented on top of pre-trained non-hierarchical clustering models outperforms models designed specifically for hierarchical clustering. Our proposed approach is computationally efficient and applicable to any pre-trained clustering model that outputs logits, without requiring any fine-tuning. To highlight the generality of our findings, we illustrate how our method can also be applied in a supervised setup, recovering meaningful hierarchies from a pre-trained ImageNet classifier.", "keywords": [ "Hierarchical Clustering", "Clustering", "Interpretability and Explainability", "Unsupervised Learning" ], "rating_list": [ 5, 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 3, 2 ], "presentation_list": [ 1, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4, 4 ] }, { "paper_id": "Pm1NXHgzyf", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Kazuki Kozuka", "gender": "Male", "institution": "Panasonic Corporation", "country": "JP", "position": "Researcher" }, { "name": "Kehan Li", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Konstantinos Kallidromitis", "gender": "Male", "institution": "Panasonic", "country": "US", "position": "AI Research Engineer" }, { "name": "Shaolun Zhang", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Shufan Li", "gender": "Male", "institution": "UCLA Computer Science Department, University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Yusuke Kato", "gender": "unknown", "institution": "Panasonic Holdings Corporation", "country": "JP", "position": "Researcher" }, { "name": "trevor darrell", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department", "country": "US", "position": "Professor" }, { "name": "Xudong Wang", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 46, "n_ref": 122, "n_ref_all": 161, "n_fig": 9, "n_tab": 4, "L_tab": 1197, "n_element_tab": 147, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1330, "n_element_tab_1": 200, "formula_len_all": 215, "formula_len_all_1": 83, "len_all": 174466, "len_all_1": 63132, "len_abs": 874, "len_title": 116, "len_sents": 49778, "len_sents_1": 28559, "n_sents": 341, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 936, "title": "SegLLM: Multi-round Reasoning Segmentation with Large Language Model", "abs": "We present SegLLM, a novel multi-round interactive reasoning segmentation model that enhances LLM-based segmentation by exploiting conversational memory of both visual and textual outputs. By leveraging a mask-aware multimodal LLM, SegLLM re-integrates previous segmentation results into its input stream, enabling it to reason about complex user intentions and segment objects in relation to previously identified entities, including positional, interactional, and hierarchical relationships, across multiple interactions. This capability allows SegLLM to respond to visual and text queries in a chat-like manner. Evaluated on the newly curated MRSeg benchmark, SegLLM outperforms existing methods in multi-round interactive reasoning segmentation by over 20%. In addition, SegLLM obtains a 5.5% improvement in cIoU for standard single-round referring segmentation and a 4.5% increase in Acc@0.5 for referring expression comprehension.", "keywords": [ "LLMs", "Reasoning Segmentation", "Muiti-round Conversations" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 4, 4 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "PlKQ9UDgqp", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Changsun Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Inhwa Han", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jong Chul Ye", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Jaayeon Lee", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 47, "n_ref": 86, "n_ref_all": 110, "n_fig": 8, "n_tab": 3, "L_tab": 833, "n_element_tab": 160, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 728, "n_element_tab_1": 150, "formula_len_all": 140, "formula_len_all_1": 248, "len_all": 119883, "len_all_1": 54960, "len_abs": 200, "len_title": 119, "len_sents": 29818, "len_sents_1": 26374, "n_sents": 204, "n_sents_1": 163, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1506, "title": "MindFormer: Semantic Alignment of Multi-Subject fMRI for Brain Decoding", "abs": "Research efforts for visual decoding from fMRI signals have attracted considerable attention in research community. Still multi-subject fMRI decoding with one model has been considered intractable due to the drastic variations in fMRI signals between subjects and even within the same subject across different trials. To address current limitations in multi-subject brain decoding, here we introduce a novel semantic alignment method of multi-subject fMRI signals using so-called $\\textit{MindFormer}$. This model is specifically designed to generate fMRI-conditioned feature vectors that can be used for conditioning Stable Diffusion model for fMRI- to-image generation or large language model (LLM) for fMRI-to-text generation. More specifically, MindFormer incorporates two key innovations: 1) a subject specific token that effectively capture individual differences in fMRI signals while synergistically combines multi subject fMRI data for training, and 2) a novel feature embedding and training scheme based on the IP-Adapter to extract semantically meaningful features from fMRI signals. Our experimental results demonstrate that MindFormer generates semantically consistent images and text across different subjects. Since our MindFormer maintains semantic fidelity by fully utilizing the training data across different subjects by significantly surpassing existing models in multi-subject brain decoding, this may help deepening our understanding of neural processing variations among individuals.", "keywords": [ "fMRI", "Neural Decoding", "Image Reconstruction" ], "rating_list": [ 5, 1, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 1, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "PkpNRmBZ32", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Yan Ru Pei", "gender": "Male", "institution": "Brainchip Inc", "country": "US", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 16, "n_ref_uni": 44, "n_ref": 75, "n_ref_all": 95, "n_fig": 5, "n_tab": 5, "L_tab": 1447, "n_element_tab": 174, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1741, "formula_len_all_1": 724, "len_all": 184146, "len_all_1": 66829, "len_abs": 1208, "len_title": 123, "len_sents": 67140, "len_sents_1": 33571, "n_sents": 473, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 75, "L_abs": 1212, "title": "Let SSMs be ConvNets: State-space Modeling with Optimal Tensor Contractions", "abs": "We introduce Centaurus, a class of networks composed of generalized state-space model (SSM) blocks, where the SSM operations can be treated as tensor contractions during training. The optimal order of tensor contractions can then be systematically determined for every SSM block to maximize training efficiency. This allows more flexibility in designing SSM blocks beyond the depthwise-separable configuration commonly implemented. The new design choices will take inspiration from classical convolutional blocks including group convolutions, full convolutions, and bottleneck blocks. We architect the Centaurus network with a mixture of these blocks, to balance between network size and performance, as well as memory and computational efficiency during both training and inference. We show that this heterogeneous network design outperforms its homogeneous counterparts in raw audio processing tasks including keyword spotting, speech denoising, and automatic speech recognition (ASR). For ASR, Centaurus is the first network with competitive performance that can be made fully state-space based, without using any nonlinear recurrence (LSTMs), explicit convolutions (CNNs), or (surrogate) attention mechanism.", "keywords": [ "state-space models; convolution; tensor networks; audio processing; speech recognition" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "PklYedVFUW", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Dongxiang Chen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Ying Wen", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 10, "n_ref_uni": 31, "n_ref": 68, "n_ref_all": 90, "n_fig": 12, "n_tab": 1, "L_tab": 685, "n_element_tab": 30, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1945, "formula_len_all_1": 794, "len_all": 128543, "len_all_1": 63656, "len_abs": 1584, "len_title": 148, "len_sents": 41776, "len_sents_1": 29219, "n_sents": 317, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1672, "title": "DSR: Reinforcement Learning with Dynamical Skill Refinement", "abs": "Reinforcement learning with skills (RL with skills) is an efficient paradigm for solving sparse-reward tasks by extracting skills from demonstration datasets and learning high-level policy which selects skills. Because each selected skill by high-level policy is executed for multiple consecutive timesteps, the high-level policy is essentially learned in a temporally abstract Markov decision process (TA-MDP) built on the skills, which shortens the task horizon and reduces the exploration cost. However, these skills are usually sub-optimal because of the potential low quality and low coverage of the datasets, which causes the sub-optimal performance in the downstream task. Refining skills is intuitive, but the change of skills will in turn lead to the non-stationarity of the transition dynamics of TA-MDP which we name temporal abstraction shift. To address the dilemma of sub-optimal skills and temporal abstraction shift, we unify the optimization objectives of the entire hierarchical policy consisting of the high-level policy and the low-level policy whose latent space embeds the skills. We theoretically prove that the unified optimization objective guarantees the performance improvement in TA-MDP, and that optimizing the performance in TA-MDP is equivalent to optimizing a lower bound of the performance of the entire hierarchical policy in original MDP. Furthermore, in order to overcome the phenomenon of skill space collapse, we propose the dynamical skill refinement (DSR) mechanism which names our method. The experiment results empirically validate the effectiveness of our method, and show the advantages over the state-of-the-art (SOTA) methods.", "keywords": [ "Reinforcement Learning", "Reinforcement Learning with Skills", "Reinforcement Learning with Demonstrations" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "Pjkes5MdKI", "primary_area": "reinforcement learning", "n_author": 1, "author_info_list": [ { "name": "Jipeng Han", "gender": "Male", "institution": "Beijing Huagui Technology Co., LTD", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 30, "n_ref": 39, "n_ref_all": 74, "n_fig": 8, "n_tab": 10, "L_tab": 6043, "n_element_tab": 328, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1634, "n_element_tab_1": 149, "formula_len_all": 155, "formula_len_all_1": 379, "len_all": 173385, "len_all_1": 50644, "len_abs": 1244, "len_title": 167, "len_sents": 50981, "len_sents_1": 22108, "n_sents": 354, "n_sents_1": 164, "venue": "ICLR2025", "decision": "withdraw", "L_title": 119, "L_abs": 1272, "title": "COOL: Efficient and Reliable Chain-Oriented Objective Logic with Neural Networks Feedback Control for Program Synthesis", "abs": "Program synthesis methods, whether formal or neural-based, lack fine-grained control and flexible modularity, which limits their adaptation to complex software development. These limitations stem from rigid Domain-Specific Language (DSL) frameworks and neural network incorrect predictions. To this end, we propose the \\textbf{Chain of Logic (CoL)}, which organizes synthesis stages into a chain and provides precise heuristic control to guide the synthesis process. Furthermore, by integrating neural networks with libraries and introducing a \\textbf{Neural Network Feedback Control (NNFC)} mechanism, our approach modularizes synthesis and mitigates the impact of neural network mispredictions. Experiments on relational and symbolic synthesis tasks show that CoL significantly enhances the efficiency and reliability of DSL program synthesis across multiple metrics. Specifically, CoL improves accuracy by 70\\% while reducing tree operations by 91\\% and time by 95\\%. Additionally, NNFC further boosts accuracy by 6\\%, with a 64\\% reduction in tree operations under challenging conditions such as insufficient training\ndata, increased difficulty, and multidomain synthesis. These improvements confirm COOL as a highly efficient and reliable program synthesis framework.", "keywords": [ "program synthesis", "Chain-of-Logic", "neural network feedback control" ], "rating_list": [ 3, 5, 1, 1 ], "soundness_list": [ 1, 2, 1, 1 ], "presentation_list": [ 1, 2, 1, 1 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 3, 2, 4, 5 ] }, { "paper_id": "PjIe6IesEm", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Anh Tuan Tran", "gender": "Male", "institution": "VinAI Research", "country": "VN", "position": "Research Scientist" }, { "name": "Toan Tran", "gender": "Male", "institution": "VinAI Research", "country": "VN", "position": "Researcher" }, { "name": "Viet Van Nguyen", "gender": "Male", "institution": "VinAI Research", "country": "VN", "position": "Researcher" }, { "name": "Bao Tran", "gender": "Male", "institution": "VinAI Research", "country": "VN", "position": "Resident" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 37, "n_ref": 90, "n_ref_all": 118, "n_fig": 6, "n_tab": 13, "L_tab": 3065, "n_element_tab": 341, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 2121, "n_element_tab_1": 229, "formula_len_all": 1339, "formula_len_all_1": 946, "len_all": 132732, "len_all_1": 58519, "len_abs": 1338, "len_title": 163, "len_sents": 35779, "len_sents_1": 26563, "n_sents": 264, "n_sents_1": 152, "venue": "ICLR2025", "decision": "Reject", "L_title": 115, "L_abs": 1343, "title": "Dual-Model Defense: Safeguarding Diffusion Models from Membership Inference Attacks through Disjoint Data Splitting", "abs": "Diffusion models have demonstrated remarkable capabilities in image synthesis, but their recently proven vulnerability to Membership Inference Attacks (MIAs) poses a critical privacy concern. This paper introduces two novel and efficient approaches (DualMD and DistillMD) to protect diffusion models against MIAs while maintaining high utility. Both methods are based on training two separate diffusion models on disjoint subsets of the original dataset. DualMD then employs a private inference pipeline that utilizes both models. This strategy significantly reduces the risk of black-box MIAs by limiting the information any single model contains about individual training samples. The dual models can also generate \"soft targets\" to train a private student model in DistillMD, enhancing privacy guarantees against all types of MIAs. Extensive evaluations of DualMD and DistillMD against state-of-the-art MIAs across various datasets in white-box and black-box settings demonstrate their effectiveness in substantially reducing MIA success rates while preserving competitive image generation performance. Notably, our experiments reveal that DistillMD not only defends against MIAs but also mitigates model memorization, indicating that both vulnerabilities stem from overfitting and can be addressed simultaneously with our unified approach.", "keywords": [ "Membership Inference Attack", "Diffusion Models", "Knowledge Distillation" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "Pj4Aid3XqL", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Achal Dave", "gender": "Male", "institution": "Anthropic", "country": "US", "position": "Researcher" }, { "name": "Benjamin Burchfiel", "gender": "Male", "institution": "Dexterous Manipulation Group, Toyota Research Institute", "country": "", "position": "Researcher" }, { "name": "Igor Vasiljevic", "gender": "Male", "institution": "Toyota Research Institute", "country": "US", "position": "Research Scientist" }, { "name": "Jean Mercat", "gender": "Male", "institution": "Toyota Research Institute", "country": "US", "position": "Researcher" }, { "name": "Kushal Arora", "gender": "Male", "institution": "McGill University", "country": "", "position": "PhD student" }, { "name": "Ludwig Schmidt", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Russ Tedrake", "gender": "Male", "institution": "Toyota Research Institute", "country": "", "position": "Researcher" }, { "name": "Samir Yitzhak Gadre", "gender": "Male", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Shuran Song", "gender": "Female", "institution": "Columbia University", "country": "", "position": "Assistant Professor" }, { "name": "Thomas Kollar", "gender": "Male", "institution": "Wayve", "country": "US", "position": "Principal Researcher" }, { "name": "Sedrick Scott Keh", "gender": "Male", "institution": "Toyota Research Institute", "country": "", "position": "Research Engineer" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 30, "n_ref": 80, "n_ref_all": 115, "n_fig": 9, "n_tab": 9, "L_tab": 1345, "n_element_tab": 247, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 1589, "n_element_tab_1": 86, "formula_len_all": 111, "formula_len_all_1": 110, "len_all": 162576, "len_all_1": 60154, "len_abs": 1415, "len_title": 91, "len_sents": 41438, "len_sents_1": 27127, "n_sents": 318, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 1014, "title": "Should VLMs be Pre-trained with Image Data?", "abs": "Pre-trained LLMs that are further trained with image data perform well on vision-language tasks. \nWhile adding images during a second training phase effectively unlocks this capability, it is unclear how much of a gain or loss this two-step pipeline gives over VLMs which integrate images earlier into the training process. \nTo investigate this, we train models spanning various datasets, scales, image-text ratios, and amount of pre-training done before introducing vision tokens.\nWe then fine-tune these models and evaluate their downstream performance on a suite of vision-language and text-only tasks.\nWe find that pre-training with a mixture of image and text data allows models to perform better on vision-language tasks while maintaining strong performance on text-only evaluations.\nOn an average of 6 diverse tasks, we find that for a 1B model, introducing visual tokens 80\\% of the way through pre-training results in a 2\\% average improvement over introducing visual tokens to a fully pre-trained model.", "keywords": [ "vision language models", "pre-training", "fine-tuning" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Pj2qEVzufH", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Guandao Yang", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Leonidas Guibas", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Yanchao Yang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Yunchao Mercer Zhang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 42, "n_ref": 85, "n_ref_all": 115, "n_fig": 15, "n_tab": 3, "L_tab": 902, "n_element_tab": 183, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 640, "n_element_tab_1": 121, "formula_len_all": 650, "formula_len_all_1": 603, "len_all": 156332, "len_all_1": 64795, "len_abs": 1667, "len_title": 122, "len_sents": 46002, "len_sents_1": 30880, "n_sents": 328, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1713, "title": "Efficient Structure-Aware 3D Gaussians via Lightweight Information Shaping", "abs": "3D Gaussians, \nas an explicit scene representation, \ntypically involve \nthousands to millions of elements per scene. \nThis makes it \nchallenging to control \nthe scene in ways \nthat reflect the underlying semantics, \nwhere the number of independent entities \nis typically much smaller. \nEspecially, \nif one wants to animate or edit objects in the scene, \nas this requires coordination among the many Gaussians\ninvolved in representing each object. \nTo address this issue, \nwe develop a mutual information shaping technique \nthat enforces resonance and coordination\nbetween correlated Gaussians \nvia a Gaussian attribute decoding network. \nSuch correlations can be learned \nfrom putative 2D object masks in different views. \nBy approximating the \nmutual information with \nthe gradients concerning the network parameters, \nour method ensures consistency \nbetween scene elements \nand enables efficient scene editing \nby operating on network parameters rather than massive Gaussians.\nIn particular, \nwe develop an effective contrastive learning pipeline\nwith lightweight optimization to shape the attribute decoding network,\nwhile ensuring that the shaping (consistency) is maintained during continuous edits, \navoiding re-shaping after parameter changes. \nNotably, \nour training only touches \na small fraction of all Gaussians in the scene \nyet attains the desired correlated behavior \naccording to the underlying scene structure. \nThe proposed technique \nis evaluated on challenging scenes and demonstrates significant performance improvements \nin 3D object segmentation and promoting scene interactions, \nwhile inducing \nlow computation and memory requirements. \nOur code and trained models \nwill be made available.", "keywords": [ "Mutual Information Maximization; 3D Reconstruction; 3D Editing" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 1, 2 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "Pj06mxCXPl", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Jiuqi Wang", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Shangtong Zhang", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "", "position": "Assistant Professor" }, { "name": "Ethan H Blaser", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Hadi Daneshmand", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Assistant Professor" } ], "n_formula": 185, "n_formula_1": 29, "n_ref_uni": 56, "n_ref": 99, "n_ref_all": 163, "n_fig": 18, "n_tab": 3, "L_tab": 819, "n_element_tab": 27, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 17324, "formula_len_all_1": 2224, "len_all": 329063, "len_all_1": 83571, "len_abs": 1403, "len_title": 136, "len_sents": 79240, "len_sents_1": 33557, "n_sents": 973, "n_sents_1": 315, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1221, "title": "Transformers Learn Temporal Difference Methods for In-Context Reinforcement Learning", "abs": "Traditionally, reinforcement learning (RL) agents learn to solve new tasks by updating their parameters through interactions with the task environment. However, recent works have demonstrated that transformer-based RL agents, after certain pretraining procedures, can learn to solve new out-of-distribution tasks without parameter updates, a phenomenon known as in-context reinforcement learning (ICRL). The empirical success of ICRL is widely attributed to the hypothesis that the forward pass of these models implements an RL algorithm. However, no prior works have demonstrated a precise equivalence between a forward pass and any specific RL algorithm, even in simplified settings like transformers with linear attention. In this paper, we present the first proof by construction demonstrating that transformers with linear attention can implement temporal difference (TD) learning in the forward pass \u2014 referred to as in-context TD. We also provide theoretical analysis and empirical evidence demonstrating the emergence of in-context TD after training the transformer with a multi-task TD algorithm, offering the first constructive explanation for transformers\u2019 ability to perform in-context reinforcement learning.", "keywords": [ "in-context reinforcement learning", "policy evaluation", "temporal difference learning" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "Pin2kdWloe", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Claire Vernade", "gender": "Female", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "", "position": "Assistant Professor" }, { "name": "Giulia Lanzillotta", "gender": "Female", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" }, { "name": "Mandana Samiei", "gender": "Female", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "Razvan Pascanu", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Affiliate Member" } ], "n_formula": 100, "n_formula_1": 19, "n_ref_uni": 52, "n_ref": 99, "n_ref_all": 134, "n_fig": 7, "n_tab": 7, "L_tab": 837, "n_element_tab": 143, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 433, "n_element_tab_1": 50, "formula_len_all": 8543, "formula_len_all_1": 1130, "len_all": 228555, "len_all_1": 73806, "len_abs": 1422, "len_title": 104, "len_sents": 73493, "len_sents_1": 33628, "n_sents": 639, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1082, "title": "Is multitask learning all you need in continual learning?", "abs": "Continual Learning solutions often treat multitask learning as an upper-bound of what the learning process can achieve. \nThis is a natural assumption, given that this objective directly addresses the catastrophic forgetting problem, which has been a central focus in early works. However, depending on the nature of the distributional shift in the data, the multi-task solution is not always optimal for the broader continual learning problem. In this work, we draw on principles from online learning to formalize the limitations of multitask objectives, especially when viewed through the lens of cumulative loss, which also serves as an indicator of forward transfer.\nWe provide empirical evidence on when multi-task solutions are suboptimal, and argue that continual learning solutions should not and do not have to adhere to this assumption. Moreover, we argue for the utility of estimating the distributional drift as the data is being received and show preliminary results of how this could be exploited by a simple replay based method to move beyond the multitask solution.", "keywords": [ "lifelong learning", "multitask learning", "continual learning" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "Pik26bc4Jx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Paul Quinlan", "gender": "Male", "institution": "Queen's University", "country": "", "position": "PhD student" }, { "name": "Qingguo Li", "gender": "Male", "institution": "Queen's University", "country": "", "position": "Associate Professor" }, { "name": "Xiaodan Zhu", "gender": "Male", "institution": "Queen's University", "country": "", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 41, "n_ref": 76, "n_ref_all": 97, "n_fig": 6, "n_tab": 8, "L_tab": 4988, "n_element_tab": 266, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 713, "n_element_tab_1": 42, "formula_len_all": 249, "formula_len_all_1": 159, "len_all": 228442, "len_all_1": 60014, "len_abs": 1839, "len_title": 74, "len_sents": 70610, "len_sents_1": 30476, "n_sents": 512, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1665, "title": "Enhancing Multi-Modal Reasoning Over Time-Series and Natural Language Data", "abs": "Time-series analysis is critical in many industries such as healthcare, finance and energy sectors, where understanding time-series trends alongside contextual information is essential for informed decision making. However, current time-series models are limited in their ability to perform reasoning that involves both time-series data and textual information. In this work we address this gap by introducing Chat-TS, a large language model (LLM) designed specifically for reasoning over time-series and textual data. Unlike traditional time-series models Chat-TS integrates time-series tokens into the LLM vocabulary, enhancing its reasoning ability over both text and time-series modalities without compromising its core natural language capabilities.\nTo support the development and validation of Chat-TS we contribute three new datasets: the TS Instruct Training Dataset which pairs diverse time-series data with relevant text instructions and responses for instruction tuning, the TS Instruct question and answer (QA) benchmark, a set of nearly 4000 multiple-choice questions designed to evaluate multi-modal reasoning and the TS Instruct Qualitative Benchmark which provides a smaller subset of QA, math and decision making questions for LLM evaluation. Our training strategy preserves the inherent reasoning capabilities of the LLM while augmenting it with time-series reasoning capabilities. Evaluation results show that Chat-TS achieves state-of-the-art performance in multi-modal reasoning tasks, maintaining strong natural language proficiency while advancing time-series reasoning. All models, datasets, and code will be made publicly available [link].", "keywords": [ "Time-Series", "Natural Language Processing", "Multi-Modal" ], "rating_list": [ 6, 3, 1, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 2, 5, 4 ] }, { "paper_id": "PigfMZMHq1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 1, "author_info_list": [ { "name": "Ali Kashefi", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 48, "n_ref": 150, "n_ref_all": 183, "n_fig": 5, "n_tab": 8, "L_tab": 2995, "n_element_tab": 514, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 682, "n_element_tab_1": 94, "formula_len_all": 791, "formula_len_all_1": 911, "len_all": 131382, "len_all_1": 49545, "len_abs": 1711, "len_title": 143, "len_sents": 32772, "len_sents_1": 22074, "n_sents": 241, "n_sents_1": 152, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1720, "title": "PointNet with KAN versus PointNet with MLP for 3D Classification and Segmentation of Point Sets", "abs": "Kolmogorov\u2013Arnold Networks (KANs) have recently gained attention as an alternative to traditional Multilayer Perceptrons (MLPs) in deep learning frameworks. KANs have been integrated into various deep learning architectures such as convolutional neural networks, graph neural networks, and transformers, with their performance evaluated. However, their effectiveness within point-cloud-based neural networks remains unexplored. To address this gap, we incorporate KANs into PointNet for the first time to evaluate their performance on 3D point cloud classification and segmentation tasks. Specifically, we introduce PointNet-KAN, built upon two key components. First, it employs KANs instead of traditional MLPs. Second, it retains the core principle of PointNet by using shared KAN layers and applying symmetric functions for global feature extraction, ensuring permutation invariance with respect to the input features. In traditional MLPs, the goal is to train the weights and biases with fixed activation functions; however, in KANs, the goal is to train the activation functions themselves. We use Jacobi polynomials to construct the KAN layers. We extensively and systematically evaluate PointNet-KAN across various polynomial degrees and special types such as the Lagrange, Chebyshev, and Gegenbauer polynomials. Our results show that PointNet-KAN achieves competitive performance compared to PointNet with MLPs on benchmark datasets for 3D object classification and segmentation, despite employing a shallower and simpler network architecture. We hope this work serves as a foundation and provides guidance for integrating KANs, as an alternative to MLPs, into more advanced point cloud processing architectures.", "keywords": [ "Kolmogorov-Arnold Networks", "PointNet", "Computer Graphics", "Classification", "Segmentation" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "PiZtlzMWUj", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Daniel Ward", "gender": "Male", "institution": "University of Bristol", "country": "", "position": "PhD student" }, { "name": "Mark Beaumont", "gender": "unknown", "institution": "University of Bristol", "country": "", "position": "Full Professor" }, { "name": "Matteo Fasiolo", "gender": "unknown", "institution": "University of Bristol", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 50, "n_formula_1": 32, "n_ref_uni": 71, "n_ref": 108, "n_ref_all": 141, "n_fig": 8, "n_tab": 1, "L_tab": 467, "n_element_tab": 18, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3264, "formula_len_all_1": 2227, "len_all": 180890, "len_all_1": 67014, "len_abs": 1489, "len_title": 122, "len_sents": 47597, "len_sents_1": 28208, "n_sents": 338, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 74, "L_abs": 1495, "title": "SoftCVI: Contrastive variational inference with self-generated soft labels", "abs": "Estimating a distribution given access to its unnormalized density is pivotal in Bayesian inference, where the posterior is generally known only up to an unknown normalizing constant. Variational inference and Markov chain Monte Carlo methods are the predominant tools for this task; however, both are often challenging to apply reliably, particularly when the posterior has complex geometry. Here, we introduce Soft Contrastive Variational Inference (SoftCVI), which allows a family of variational objectives to be derived through a contrastive estimation framework. The approach parameterizes a classifier in terms of a variational distribution, reframing the inference task as a contrastive estimation problem aiming to identify a single true posterior sample among a set of samples. Despite this framing, we do not require positive or negative samples, but rather learn by sampling the variational distribution and computing ground truth soft classification labels from the unnormalized posterior itself. The objectives have zero variance gradient when the variational approximation is exact, without the need for specialized gradient estimators. We empirically investigate the performance on a variety of Bayesian inference tasks, using both simple (e.g. normal) and expressive (normalizing flow) variational distributions. We find that SoftCVI can be used to form objectives which are stable to train and mass-covering, frequently outperforming inference with other variational approaches.", "keywords": [ "contrastive learning", "variational inference" ], "rating_list": [ 8, 10, 5, 6 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "PiOhaDXuXa", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 7, "author_info_list": [ { "name": "Emanuele Rossi", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "PhD student" }, { "name": "Harrison Rush", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Jesse Shrader", "gender": "Male", "institution": "Amboss Technologies", "country": "US", "position": "CEO" }, { "name": "Matthew Khanzadeh", "gender": "Male", "institution": "Ontra", "country": "US", "position": "Software Engineer" }, { "name": "Pietro Lio", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Vikash Singh", "gender": "unknown", "institution": "Stillmark (Bitcoin VC)", "country": "US", "position": "Researcher" }, { "name": "Vincent Davis", "gender": "unknown", "institution": "Amboss Technologies", "country": "US", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 5, "n_ref_uni": 20, "n_ref": 23, "n_ref_all": 25, "n_fig": 0, "n_tab": 6, "L_tab": 5786, "n_element_tab": 30, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 2076, "n_element_tab_1": 63, "formula_len_all": 853, "formula_len_all_1": 215, "len_all": 105681, "len_all_1": 52573, "len_abs": 1056, "len_title": 70, "len_sents": 34327, "len_sents_1": 22845, "n_sents": 344, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Reject", "L_title": 22, "L_abs": 1057, "title": "Bayesian Binary Search", "abs": "We present Bayesian Binary Search (BBS), a novel probabilistic variant of the classical binary search/bisection algorithm. BBS leverages machine learning/statistical techniques to estimate the probability density of the search space and modifies the bisection step to split based on probability density rather than the traditional midpoint, allowing for the learned distribution of the search space to guide the search algorithm. Search space density estimation can flexibly\nbe performed using supervised probabilistic machine learning techniques (e.g., Gaussian process regression, Bayesian neural networks, quantile regression) or unsupervised learning algorithms (e.g., Gaussian mixture models, kernel density estimation (KDE), maximum likelihood estimation (MLE)). We demonstrate significant efficiency gains of using BBS on both simulated data across a variety of distributions and in a real-world binary search use case of probing channel balances in the Bitcoin Lightning Network, for which we have deployed the BBS algorithm in a production setting.", "keywords": [ "bayesian", "binary search", "bitcoin", "lightning network" ], "rating_list": [ 5, 6, 3, 1 ], "soundness_list": [ 2, 3, 2, 1 ], "presentation_list": [ 2, 3, 1, 1 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "PiHGrTTnvb", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Dixia Fan", "gender": "Male", "institution": "Westlake University", "country": "", "position": "Assistant Professor" }, { "name": "Haodong Feng", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Long Wei", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Postdoc" }, { "name": "Peiyan Hu", "gender": "Not Specified", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Ruiqi Feng", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Tailin Wu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tao Zhang", "gender": "Male", "institution": "Wuhan University", "country": "", "position": "Undergrad student" }, { "name": "Xiang Zheng", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "Intern" }, { "name": "Yuchen Yang", "gender": "unknown", "institution": "Nankai University", "country": "CN", "position": "MS student" } ], "n_formula": 72, "n_formula_1": 32, "n_ref_uni": 42, "n_ref": 85, "n_ref_all": 118, "n_fig": 13, "n_tab": 6, "L_tab": 1733, "n_element_tab": 218, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1511, "n_element_tab_1": 149, "formula_len_all": 7917, "formula_len_all_1": 2912, "len_all": 176578, "len_all_1": 63635, "len_abs": 1133, "len_title": 105, "len_sents": 54044, "len_sents_1": 27087, "n_sents": 429, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1139, "title": "CL-DiffPhyCon: Closed-loop Diffusion Control of Complex Physical Systems", "abs": "The control problems of complex physical systems have broad applications in science and engineering. Previous studies have shown that generative control methods based on diffusion models offer significant advantages for solving these problems. However, existing generative control approaches face challenges in both performance and efficiency when extended to the closed-loop setting, which is essential for effective control. In this paper, we propose an efficient Closed-Loop Diffusion method for Physical systems Control (CL-DiffPhyCon). By employing an asynchronous denoising framework for different physical time steps, CL-DiffPhyCon generates control signals conditioned on real-time feedback from the system with significantly reduced computational cost during sampling. Additionally, the control process could be further accelerated by incorporating fast sampling techniques, such as DDIM. We evaluate CL-DiffPhyCon on two tasks: 1D Burgers' equation control and 2D incompressible fluid control. The results demonstrate that CL-DiffPhyCon achieves superior control performance with significant improvements in sampling efficiency.", "keywords": [ "physical systems control", "closed-loop control", "PDE", "physical simulation", "generative models" ], "rating_list": [ 3, 10, 8 ], "soundness_list": [ 2, 4, 3 ], "presentation_list": [ 1, 4, 3 ], "contribution_list": [ 2, 4, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "PhRYDGqiee", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Elisa Ricci", "gender": "Female", "institution": "University of Trento", "country": "IT", "position": "Full Professor" }, { "name": "Gianni Franchi", "gender": "Male", "institution": "ENSTA Paris", "country": "FR", "position": "Assistant Professor" }, { "name": "Jun Li", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Mingxuan Liu", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "PhD student" }, { "name": "Subhankar Roy", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Postdoc" }, { "name": "Zhun Zhong", "gender": "Male", "institution": "University of Nottingham", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 101, "n_ref": 237, "n_ref_all": 305, "n_fig": 24, "n_tab": 17, "L_tab": 8607, "n_element_tab": 1008, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 310, "n_element_tab_1": 45, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 392807, "len_all_1": 67467, "len_abs": 422, "len_title": 112, "len_sents": 126499, "len_sents_1": 33619, "n_sents": 889, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1342, "title": "Organizing Unstructured Image Collections using Natural Language", "abs": "Organizing unstructured visual data into semantic clusters is a key challenge in computer vision. Traditional deep clustering (DC) approaches focus on a single partition of data, while multiple clustering (MC) methods address this limitation by uncovering distinct clustering solutions. The rise of large language models (LLMs) and multimodal LLMs (MLLMs) has enhanced MC by allowing users to define clustering criteria in natural language. However, manually specifying criteria for large datasets is impractical. In this work, we introduce the task Semantic Multiple Clustering (SMC) that aims to automatically discover clustering criteria from large image collections, uncovering interpretable substructures without requiring human input. Our framework, Text Driven Semantic Multiple Clustering (TeDeSC), uses text as a proxy to concurrently reason over large image collections, discover partitioning criteria, expressed in natural language, and reveal semantic substructures. To evaluate TeDeSC, we introduce the COCO-4c and Food-4c benchmarks, each containing four grouping criteria and ground-truth annotations. We apply TeDeSC to various applications, such as discovering biases and analyzing social media image popularity, demonstrating its utility as a tool for automatically organizing image collections and revealing novel insights.", "keywords": [ "Vision-Language", "Image Clustering" ], "rating_list": [ 6, 5, 8, 5, 3 ], "soundness_list": [ 4, 2, 3, 3, 1 ], "presentation_list": [ 3, 3, 2, 4, 2 ], "contribution_list": [ 4, 2, 3, 2, 2 ], "confidence_list": [ 2, 3, 4, 4, 3 ] }, { "paper_id": "PhLCPYsHCw", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 1, "author_info_list": [ { "name": "Jingjing Fu", "gender": "Female", "institution": "Clemson University", "country": "US", "position": "PhD student" } ], "n_formula": 65, "n_formula_1": 18, "n_ref_uni": 29, "n_ref": 72, "n_ref_all": 90, "n_fig": 8, "n_tab": 6, "L_tab": 2305, "n_element_tab": 371, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 76, "n_element_tab_1": 9, "formula_len_all": 17306, "formula_len_all_1": 2306, "len_all": 147603, "len_all_1": 58572, "len_abs": 1803, "len_title": 132, "len_sents": 32740, "len_sents_1": 25701, "n_sents": 277, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1862, "title": "ON THE CONVERGENCE OF CYCLIC HIERARCHICAL FEDERATED LEARNING WITH HETEROGENEOUS DATA", "abs": "Hierarchical Federated Learning (HFL) advances the classic Federated Learning (FL) by introducing the multi-layer architecture between clients and the central server, in which edge servers aggregate models from respective clients and further send to the central server. Instead of directly uploading each update from clients for aggregation, the HFL not only reduces the communication and computational overhead but also greatly enhances the scalability of supporting a massive number of clients. When HFL operates for applications having a large-scale clients, edge servers train their models in a cyclic pattern (a ring architecture) as opposed to the star-type of architecture where each edge develops their own models independently.We refer it as Cyclic HFL(CHFL). Driven by its promising feature of handling data heterogeneity and resiliency, CHFL has a great potential to be deployed in practice. Unfortunately, the thorough convergence analysis on CHFL remains lacking, especially considering the widely-existing data heterogeneity issue among clients. To the best of our knowledge, we are the first to provide a theoretical convergence analysis for CHFL in strongly convex, general convex, and non-convex objectives. Our results demonstrate the convergence rate are $\\tilde{\\mathcal{O}}(1/MNRKT)$ for strongly convex objective, $\\mathcal{O}(1/\\sqrt{MNRKT})$ for general convex objective, and $\\mathcal{O}(1/\\sqrt{MNRKT})$ for non-convex objective, under standard assumptions. Here, $M$ is the number of edge servers, $N$ is the number of clients in edge, $K$ is local steps in client, and $R$ is the edge training round. Through extensive experiments on real-world datasets, besides validating our theoretical findings, we further show CHFL achieves a comparable or superior performance when accounting for both inter- and intra-edge data heterogeneity.", "keywords": [ "hierarchical federated learning", "convergence analysis", "cyclic pattern" ], "rating_list": [ 3, 3, 6, 3, 3 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 1, 2 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 4, 4 ] }, { "paper_id": "Pgwpc0rzbM", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Juzheng Zhang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Yongqiang Chen", "gender": "unknown", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "quanming yao", "gender": "Male", "institution": "Department of Electronic Engineering, Tsinghua University", "country": "", "position": "Assistant Professor" }, { "name": "An Bian", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Senior researcher " } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 59, "n_ref": 197, "n_ref_all": 229, "n_fig": 8, "n_tab": 15, "L_tab": 10223, "n_element_tab": 1453, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 4376, "n_element_tab_1": 290, "formula_len_all": 459, "formula_len_all_1": 229, "len_all": 225760, "len_all_1": 59642, "len_abs": 1344, "len_title": 127, "len_sents": 57180, "len_sents_1": 24820, "n_sents": 455, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1352, "title": "UniMoT: Unified Molecule-Text Language Model with Discrete Token Representation", "abs": "The remarkable success of Large Language Models (LLMs) across diverse tasks has driven the research community to extend their capabilities to molecular applications. However, most molecular LLMs employ adapter-based architectures that do not treat molecule and text modalities equally and lack a supervision signal for the molecule modality. To address these issues, we introduce UniMoT, a Unified Molecule-Text LLM adopting a tokenizer-based architecture that expands the vocabulary of LLM with molecule tokens.\nSpecifically, we introduce a Vector Quantization-driven tokenizer that incorporates a Q-Former to bridge the modality gap between molecule and text. This tokenizer transforms molecules into sequences of molecule tokens with causal dependency, encapsulating high-level molecular and textual information. \nEquipped with this tokenizer, UniMoT can unify molecule and text modalities under a shared token representation and an autoregressive training paradigm.\nIt can interpret molecules as a foreign language and generate them as text.\nFollowing a four-stage training scheme, UniMoT emerges as a multi-modal generalist capable of performing both molecule-to-text and text-to-molecule tasks. Extensive experiments demonstrate that UniMoT achieves state-of-the-art performance across a wide range of molecule comprehension and generation tasks.", "keywords": [ "Large Language Models", "Tokenization", "Molecule Generation", "Molecule Comprehension", "Multi-modal Learning" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "Pghg8dJnUe", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Jiansheng Yang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Yaodong Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zailin Ma", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 191, "n_formula_1": 50, "n_ref_uni": 33, "n_ref": 64, "n_ref_all": 70, "n_fig": 2, "n_tab": 1, "L_tab": 143, "n_element_tab": 6, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 37225, "formula_len_all_1": 3750, "len_all": 164657, "len_all_1": 60460, "len_abs": 1270, "len_title": 105, "len_sents": 33239, "len_sents_1": 19202, "n_sents": 435, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1283, "title": "Random Feature Models with Learnable Activation Functions", "abs": "Current random feature models typically rely on fixed activation functions, limiting their ability to capture diverse patterns in data. To address this, we introduce the Random Feature model with Learnable Activation Functions (RFLAF), a novel model that significantly enhances the expressivity and interpretability of traditional random feature (RF) models. We begin by studying the RF model with a single radial basis function, where we discover a new kernel and provide the first theoretical analysis on it. By integrating the basis functions with learnable weights, we show that RFLAF can represent a broad class of random feature models whose activation functions belong in $C_c(\\mathbb{R})$. Theoretically, we prove that the model requires only about twice the parameter number compared to a traditional RF model to achieve the significant leap in expressivity. Experimentally, RFLAF demonstrates two key advantages: (1) it performs better across various tasks compared to traditional RF model with the same number of parameters, and (2) the optimized weights offer interpretability, as the learned activation function can be directly inferred from these weights. Our model paves the way for developing more expressive and interpretable frameworks within random feature models.", "keywords": [ "learnable activation function", "random features", "interpretability", "statistical learning theory" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "PgXpOOqtyd", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Amaia Cardiel", "gender": "unknown", "institution": "Universit\u00e9 Grenoble Alpes", "country": "FR", "position": "PhD student" }, { "name": "Elias Ramzi", "gender": "Male", "institution": "Valeo", "country": "FR", "position": "Researcher" }, { "name": "Matthieu Cord", "gender": "Male", "institution": "Sorbonne Universit\u00e9", "country": "FR", "position": "Full Professor" }, { "name": "Oriane Sim\u00e9oni", "gender": "unknown", "institution": "Meta FAIR", "country": "FR", "position": "Research Scientist" }, { "name": "eloi zablocki", "gender": "Male", "institution": "Valeo", "country": "FR", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 45, "n_ref": 94, "n_ref_all": 143, "n_fig": 23, "n_tab": 11, "L_tab": 3031, "n_element_tab": 339, "n_fig_1": 12, "n_tab_1": 5, "L_tab_1": 2641, "n_element_tab_1": 162, "formula_len_all": 22, "formula_len_all_1": 0, "len_all": 164356, "len_all_1": 66545, "len_abs": 1424, "len_title": 162, "len_sents": 43533, "len_sents_1": 30760, "n_sents": 333, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 113, "L_abs": 1412, "title": "LLM-wrapper: Black-Box Semantic-Aware Adaptation of Vision-Language Models for Referring Expression Comprehension", "abs": "Vision Language Models (VLMs) have demonstrated remarkable capabilities in various open-vocabulary tasks, yet their zero-shot performance lags behind task-specific fine-tuned models, particularly in complex tasks like Referring Expression Comprehension (REC). Fine-tuning usually requires \u2018white-box\u2019 access to the model\u2019s architecture and weights, which is not always feasible due to proprietary or privacy concerns. In this work, we propose LLM-wrapper, a method for \u2018black-box\u2019 adaptation of VLMs for the REC task using Large Language Models (LLMs). LLM-wrapper capitalizes on the reasoning abilities of LLMs, improved with a light fine-tuning, to select the most relevant bounding box to match the referring expression, from candidates generated by a zero-shot black-box VLM. Our approach offers several advantages: it enables the adaptation of closed-source models without needing access to their internal workings, it is versatile and works with any VLM, transfers to new VLMs, and it allows for the adaptation of an ensemble of VLMs. We evaluate LLM-wrapper on multiple datasets using different VLMs and LLMs, demonstrating significant performance improvements and highlighting the versatility of our method. While LLM-wrapper is not meant to directly compete with standard white-box fine-tuning, it offers a practical and effective alternative for black-box VLM adaptation. The code will be open-sourced.", "keywords": [ "Large Language Models", "Vision-Language Models", "Black-Box Adaptation", "Referring Expression Comprehension" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "PgVo0t9rC2", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Brieuc Francois", "gender": "unknown", "institution": "CERN", "country": "CH", "position": "Staf" }, { "name": "Dolores Garcia", "gender": "unknown", "institution": "CERN", "country": "", "position": "Postdoc" }, { "name": "Michele Selvaggi", "gender": "Male", "institution": "CERN", "country": "", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 33, "n_ref": 77, "n_ref_all": 101, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 420, "formula_len_all_1": 297, "len_all": 110309, "len_all_1": 56838, "len_abs": 1267, "len_title": 98, "len_sents": 37978, "len_sents_1": 27748, "n_sents": 302, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1279, "title": "Geometric Graph Neural Network based track finding", "abs": "An essential component of event reconstruction in particle physics experiments is identifying the trajectory of charged particles in the detector. Traditional methods for track finding are often complex, and tailored to specific detectors and input geometries, limiting their adaptability to new detector designs and optimization processes.\nTo overcome these limitations, we present a novel, end-to-end track finding algorithm that is detector-agnostic and can take into account multiple input geometric types. To achieve this, our approach unifies inputs from multiple sub-detectors and detector types into a single geometric algebra representation, simplifying data handling compared to traditional methods.\nThen, we leverage an equivariant graph neural network, GATr, to perform track finding across all data from an event simultaneously. \nWe validate the effectiveness of our pipeline on various detector concepts with different technologies for the FCC-ee at CERN, specifically the IDEA and CLD detectors.\nThis work generalizes track finding across diverse types of input geometric data and tracking technologies, facilitating the development of innovative detector concepts, accelerating detector development cycles, and enabling comprehensive detector optimization.", "keywords": [ "Tracking", "GNN", "High Energy Physics" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 1 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 1, 4 ] }, { "paper_id": "PgR6fziYmJ", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Hsin-Jung Yang", "gender": "unknown", "institution": "Iowa State University", "country": "", "position": "PhD student" }, { "name": "Joshua Russell Waite", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "Postdoc" }, { "name": "Mahsa Khosravi", "gender": "Female", "institution": "Iowa State University", "country": "US", "position": "PhD student" }, { "name": "Qisai Liu", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "PhD student" }, { "name": "Soumik Sarkar", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "Full Professor" }, { "name": "Zhanhong Jiang", "gender": "Male", "institution": "Iowa State University", "country": "US", "position": "Researcher" } ], "n_formula": 65, "n_formula_1": 18, "n_ref_uni": 47, "n_ref": 77, "n_ref_all": 97, "n_fig": 7, "n_tab": 3, "L_tab": 1103, "n_element_tab": 92, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 3347, "n_element_tab_1": 79, "formula_len_all": 5666, "formula_len_all_1": 1805, "len_all": 197133, "len_all_1": 73316, "len_abs": 1323, "len_title": 118, "len_sents": 64869, "len_sents_1": 32705, "n_sents": 515, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1269, "title": "HP3O: Hybrid-Policy Proximal Policy Optimization with Best Trajectory", "abs": "Proximal policy optimization (PPO) is one of the most popular state-of-the-art on-policy algorithms that has become a standard baseline in modern reinforcement learning with applications in numerous fields. Though it delivers stable performance with theoretical policy improvement guarantees, high variance and high sample complexity still remain critical challenges in on-policy algorithms. To alleviate these issues, we propose Hybrid-Policy Proximal Policy Optimization (HP3O), which utilizes a trajectory replay buffer to make efficient use of trajectories generated by recent policies. Particularly, the buffer applies the \"first in, first out\" (FIFO) strategy so as to keep only the recent trajectories to attenuate the data distribution drift. A batch consisting of the trajectory with the best return and other randomly sampled ones from the buffer is used for updating the policy networks. The strategy helps the agent to improve its capability on top of the most recent best performance and in turn reduce variance empirically. We theoretically construct the policy improvement guarantees for the proposed algorithm. HP3O is validated and compared against several baseline algorithms using multiple continuous control environments. Our code is available here.", "keywords": [ "Reinforcement learning", "proximal policy optimization", "hybrid policy", "trajectory replay buffer", "variance reduction" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "PgC5UqKDye", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Abhishek Kumar", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Mintong Kang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "PhD student" }, { "name": "Rashmi Gangadharaiah", "gender": "Female", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Shamik Roy", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Sopan Khosla", "gender": "Male", "institution": "Amazon Web Services", "country": "", "position": "Researcher" }, { "name": "Vinayshekhar Bannihatti Kumar", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Murali Balakrishnan", "gender": "Male", "institution": "Amazon", "country": "", "position": "Principal Researcher" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 32, "n_ref": 93, "n_ref_all": 109, "n_fig": 2, "n_tab": 11, "L_tab": 1505, "n_element_tab": 120, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1093, "n_element_tab_1": 68, "formula_len_all": 764, "formula_len_all_1": 822, "len_all": 127924, "len_all_1": 63688, "len_abs": 1489, "len_title": 139, "len_sents": 42800, "len_sents_1": 31447, "n_sents": 299, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1328, "title": "FairGen: controlling fair generations in diffusion models via adaptive latent guidance", "abs": "Diffusion models have shown remarkable proficiency in generating photorealistic images, but their outputs often exhibit biases toward specific social groups, raising ethical concerns and limiting their wider adoption. This paper tackles the challenge of mitigating generative bias in diffusion models while maintaining image quality. We propose FairGen, an adaptive latent guidance mechanism enhanced by an auxiliary memory module, which operates during inference to control the generation distribution at a desired level. The latent guidance module dynamically adjusts the direction in the latent space to influence specific attributes, while the memory module tracks prior generation statistics and steers the scalar direction to align with the target distribution. To evaluate FairGen comprehensively, we introduce a bias evaluation benchmark tailored for diffusion models, spanning diverse domains such as employment, education, finance, and healthcare, along with complex user-generated prompts. Extensive empirical evaluations demonstrate that FairGen outperforms existing bias mitigation approaches, achieving substantial bias reduction while preserving generation quality. Furthermore, FairGen offers precise and flexible control over various target distributions, enabling nuanced adjustments to the generative process.", "keywords": [ "fairness", "bias", "diffusion models" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "PflweLMInP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "C.L.Philip Chen", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Jiecheng Zhang", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "MS student" }, { "name": "Shuzhen Li", "gender": "Female", "institution": "Pazhou Lab", "country": "CN", "position": "Researcher" }, { "name": "Tong Zhang", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 36, "n_ref": 74, "n_ref_all": 94, "n_fig": 6, "n_tab": 7, "L_tab": 2436, "n_element_tab": 432, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1146, "n_element_tab_1": 207, "formula_len_all": 916, "formula_len_all_1": 895, "len_all": 127491, "len_all_1": 63488, "len_abs": 1267, "len_title": 118, "len_sents": 36277, "len_sents_1": 25646, "n_sents": 279, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1240, "title": "Complete multi-modal metric learning for multi-modal sarcasm detection", "abs": "Multi-modal sarcasm detection identifies sarcasm from text-image pairs, an essential technology for accurately understanding the user's real attitude.\nMost research extracted the incongruity of text-image pairs as sarcasm information. However, these methods neglected inter-modal or intra-modal incongruities in fact and sentiment perspectives, leading to incomplete sarcasm information and biased performance.\nTo address the above issues, this paper proposes a complete multi-modal metric learning network (CMML-Net) for multi-modal sarcasm detection tasks.\nSpecifically, CMML-Net utilizes a fact-sentiment multi-task representation learning module to produce refined fact and sentiment text-image representation pairs.\nIt then designs a complete multi-modal metric learning to iteratively calculate inter-modal and intra-modal incongruities in a unified space (e.g., fact and sentiment metric space), efficiently capturing complete multi-modal incongruities.\nCMML-Net performs well in explicitly capturing comprehensive sarcasm information and obtaining discriminative performance via deep metric learning.\nThe state-of-the-art performance on the widely-used dataset demonstrates CMML-Net's effectiveness in multi-modal sarcasm detection.", "keywords": [ "multi-modal sarcasm detection", "metric learning", "complete multi-modal incongruities" ], "rating_list": [ 10, 3, 3, 5 ], "soundness_list": [ 4, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 4, 1, 2, 2 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "PfZekmXRjN", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Bocheng Zhou", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Qiaosheng Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Shuyue Hu", "gender": "Female", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Zhen Wang", "gender": "Male", "institution": "Northwestern Polytechnical University", "country": "CN", "position": "Full Professor" }, { "name": "Zhongtian Ma", "gender": "Male", "institution": "Northwestern Polytechnical University", "country": "", "position": "PhD student" }, { "name": "YexinZhang", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "PhD student" } ], "n_formula": 185, "n_formula_1": 22, "n_ref_uni": 38, "n_ref": 84, "n_ref_all": 107, "n_fig": 7, "n_tab": 3, "L_tab": 450, "n_element_tab": 66, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 26409, "formula_len_all_1": 1375, "len_all": 233484, "len_all_1": 71284, "len_abs": 1516, "len_title": 132, "len_sents": 64704, "len_sents_1": 33274, "n_sents": 642, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1567, "title": "Understanding When and Why Graph Attention Mechanisms Work via Node Classification", "abs": "Despite the growing popularity of graph attention mechanisms, their theoretical understanding remains limited. This paper aims to explore the conditions under which these mechanisms are effective in node classification tasks through the lens of Contextual Stochastic Block Models (CSBMs). Our theoretical analysis reveals that incorporating graph attention mechanisms is *not universally beneficial*. Specifically, by appropriately defining *structural noise* and *feature noise* in graphs, we show that graph attention mechanisms can enhance classification performance when structural noise exceeds feature noise. Conversely, when feature noise predominates, simpler graph convolution operations are more effective. Furthermore, we examine the over-smoothing phenomenon and show that, in the high signal-to-noise ratio (SNR) regime, graph convolutional networks suffer from over-smoothing, whereas graph attention mechanisms can effectively resolve this issue. Building on these insights, we propose a novel multi-layer Graph Attention Network (GAT) architecture that significantly outperforms single-layer GATs in achieving *perfect node classification* in CSBMs, relaxing the SNR requirement from $ \\omega(\\sqrt{\\log n}) $ to $ \\omega(\\sqrt{\\log n} / \\sqrt[3]{n}) $. To our knowledge, this is the first study to delineate the conditions for perfect node classification using multi-layer GATs. Our theoretical contributions are corroborated by extensive experiments on both synthetic and real-world datasets, highlighting the practical implications of our findings.", "keywords": [ "Graph attention mechanisms", "node classification", "contextual stochastic block model", "over-smoothing" ], "rating_list": [ 3, 6, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "PfYg3eRrNi", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Daniel Fried", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Researcher" }, { "name": "Graham Neubig", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Jiayuan Mao", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zhiruo Wang", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 25, "n_ref": 54, "n_ref_all": 76, "n_fig": 7, "n_tab": 10, "L_tab": 1660, "n_element_tab": 263, "n_fig_1": 6, "n_tab_1": 10, "L_tab_1": 1660, "n_element_tab_1": 263, "formula_len_all": 42, "formula_len_all_1": 76, "len_all": 124139, "len_all_1": 68194, "len_abs": 1317, "len_title": 69, "len_sents": 41177, "len_sents_1": 31208, "n_sents": 286, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 21, "L_abs": 1325, "title": "Agent Workflow Memory", "abs": "Despite the potential of language model-based agents to solve real-world tasks such as web navigation, current methods still struggle with long-horizon tasks with complex action trajectories. In contrast, humans can flexibly solve complex tasks by learning reusable task workflows from past experiences and using them to guide future actions. To build agents that can similarly benefit from this process, we introduce Agent Workflow Memory (AWM), a method for inducing commonly reused routines, i.e., workflows, and selectively providing workflows to the agent to guide subsequent generations. AWM flexibly applies to both offline and online scenarios, where agents induce workflows from training examples beforehand or from test queries on the fly. We experiment on two major web navigation benchmarks -- Mind2Web and WebArena -- that collectively cover 1000+ tasks from 200+ domains across travel, shopping, and social media, among others. AWM substantially improves the baseline results by 24.6% and 51.1% relative success rate on Mind2Web and WebArena while reducing the number of steps taken to solve WebArena tasks successfully. Furthermore, online AWM robustly generalizes in cross-task, website, and domain evaluations, surpassing baselines from 8.9 to 14.0 absolute points as train-test task distribution gaps widen.", "keywords": [ "agent", "web navigation", "memory augmentation", "online" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "Pf8i7cv2CH", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Changyu Chen", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "PhD student" }, { "name": "Chao Du", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Min Lin", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Principal Researcher" }, { "name": "Wee Sun Lee", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Zichen Liu", "gender": "unknown", "institution": "Sea AI Lab, Sea, Singapore", "country": "SG", "position": "Researcher" } ], "n_formula": 59, "n_formula_1": 27, "n_ref_uni": 81, "n_ref": 168, "n_ref_all": 245, "n_fig": 16, "n_tab": 5, "L_tab": 972, "n_element_tab": 77, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 107, "n_element_tab_1": 5, "formula_len_all": 51629, "formula_len_all_1": 15674, "len_all": 323419, "len_all_1": 106231, "len_abs": 1063, "len_title": 83, "len_sents": 83316, "len_sents_1": 41123, "n_sents": 583, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1067, "title": "Sample Efficient Alignment for LLMs", "abs": "We study methods for efficiently aligning large language models (LLMs) with human preferences given budgeted online feedback. We first formulate the LLM alignment problem in the frame of contextual dueling bandits. This formulation, subsuming recent paradigms such as online RLHF and online DPO, inherently quests for sample-efficient algorithms that incorporate online active exploration. Leveraging insights from bandit theory, we introduce a unified algorithm based on Thompson sampling and highlight its applications in two distinct LLM alignment scenarios. The practical agent that efficiently implements this algorithm, named SEA (Sample-Efficient Alignment), is empirically validated through extensive experiments across three model scales (1B, 2.8B, 6.9B) and three preference learning algorithms (DPO, IPO, SLiC). The results demonstrate that SEA achieves highly sample-efficient alignment with oracle's preferences, outperforming recent active exploration methods for LLMs. We will release our codebase to hopefully accelerate future research in this field.", "keywords": [ "rlhf", "online dap", "llm alignment", "sample efficiency" ], "rating_list": [ 3, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "PevF76oAEh", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Balazs Szegedy", "gender": "Male", "institution": "Alfr\u00e9d R\u00e9nyi Institute of Mathematics", "country": "", "position": "Researcher" }, { "name": "Domonkos Czifra", "gender": "Male", "institution": "Alfr\u00e9d R\u00e9nyi Institute of Mathematics", "country": "", "position": "Researcher" }, { "name": "P\u00e9ter K\u0151r\u00f6si-Szab\u00f3", "gender": "Male", "institution": "Alfr\u00e9d R\u00e9nyi Institute of Mathematics", "country": "", "position": "Researcher" } ], "n_formula": 26, "n_formula_1": 20, "n_ref_uni": 12, "n_ref": 24, "n_ref_all": 44, "n_fig": 6, "n_tab": 3, "L_tab": 2640, "n_element_tab": 156, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 1255, "n_element_tab_1": 75, "formula_len_all": 1960, "formula_len_all_1": 1647, "len_all": 123783, "len_all_1": 68516, "len_abs": 1051, "len_title": 90, "len_sents": 44105, "len_sents_1": 29391, "n_sents": 398, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1076, "title": "Dynamic Memory Based Adaptive Optimization", "abs": "Define an optimizer as having memory $k$ if it stores $k$ dynamically changing vectors in the parameter space. Classical SGD has memory $0$, momentum SGD optimizer has $1$ and Adam optimizer has $2$. We address the following questions: *How can optimizers make use of more memory units? What information should be stored in them? How to use them for the learning steps?* As an approach to the last question, we introduce a general method called \"Retrospective Learning Law Correction\" or shortly RLLC. This method is designed to calculate a dynamically varying linear combination (called *learning law*) of memory units, which themselves may evolve arbitrarily. We demonstrate RLLC on optimizers whose memory units have linear update rules and small memory ($\\leq 4$ memory units). Our experiments show that in a variety of standard problems, these optimizers outperform the above mentioned three classical optimizers. We conclude that RLLC is a promising framework for boosting the performance of known optimizers by adding more memory units and by making them more adaptive.", "keywords": [ "optimization", "meta-training", "adaptive-learning", "RLLC", "retrospective-learning-law-correction" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 1, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 3, 2 ] }, { "paper_id": "PeLLMw3wLX", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Felipe Maia Polo", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "PhD student" }, { "name": "Mikhail Yurochkin", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Research Staff Member" }, { "name": "Moulinath Banerjee", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "Full Professor" }, { "name": "Seamus Somerstep", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Yaacov Ritov", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Full Professor" }, { "name": "Yuekai Sun", "gender": "unknown", "institution": "University of Michigan", "country": "US", "position": "Associate Professor" } ], "n_formula": 71, "n_formula_1": 17, "n_ref_uni": 76, "n_ref": 138, "n_ref_all": 143, "n_fig": 3, "n_tab": 2, "L_tab": 329, "n_element_tab": 34, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9394, "formula_len_all_1": 2012, "len_all": 271039, "len_all_1": 63248, "len_abs": 1126, "len_title": 89, "len_sents": 70949, "len_sents_1": 29311, "n_sents": 632, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1117, "title": "A transfer learning framework for weak to strong generalization", "abs": "Modern large language model (LLM) alignment techniques rely on human feedback, but it is unclear whether the techniques fundamentally limit the capabilities of aligned LLMs. In particular, it is unclear whether it is possible to align (stronger) LLMs with superhuman capabilities with (weaker) human feedback *without degrading their capabilities*. This is an instance of the weak-to-strong generalization problem: using weaker (less capable) feedback to train a stronger (more capable) model. We prove that weak-to-strong generalization is possible by eliciting latent knowledge from pre-trained LLMs. In particular, we cast the weak-to-strong generalization problem as a transfer learning problem in which we wish to transfer a latent concept from a weak model to a strong pre-trained model. We prove that a naive fine-tuning approach suffers from fundamental limitations, but an alternative refinement-based approach suggested by the problem structure provably overcomes the limitations of fine-tuning. Finally, we demonstrate the practical applicability of the refinement approach in multiple LLM alignment tasks.", "keywords": [ "Weak to strong generalization", "alignment", "transfer learning" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "Pe3AxLq6Wf", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Benoit Dufumier", "gender": "unknown", "institution": "CEA", "country": "FR", "position": "Researcher" }, { "name": "Devis Tuia", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" }, { "name": "Javiera Castillo Navarro", "gender": "Female", "institution": "Conservatoire national des arts et m\u00e9tiers", "country": "FR", "position": "Assistant Professor" }, { "name": "Jean-Philippe Thiran", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 64, "n_ref": 149, "n_ref_all": 174, "n_fig": 8, "n_tab": 9, "L_tab": 2775, "n_element_tab": 214, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1238, "n_element_tab_1": 93, "formula_len_all": 1858, "formula_len_all_1": 906, "len_all": 224639, "len_all_1": 63204, "len_abs": 1456, "len_title": 97, "len_sents": 61116, "len_sents_1": 28010, "n_sents": 436, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1436, "title": "What to align in multimodal contrastive learning?", "abs": "Humans perceive the world through multisensory integration, blending the information of different modalities to adapt their behavior.\nContrastive learning offers an appealing solution for multimodal self-supervised learning. Indeed, by considering each modality as a different view of the same entity, it learns to align features of different modalities in a shared representation space. However, this approach is intrinsically limited as it only learns shared or redundant information between modalities, while multimodal interactions can arise in other ways. In this work, we introduce CoMM, a Contrastive Multimodal learning strategy that enables the communication between modalities in a single multimodal space. Instead of imposing cross- or intra- modality constraints, we propose to align multimodal representations by maximizing the mutual information between augmented versions of these multimodal features. Our theoretical analysis shows that shared, synergistic and unique terms of information naturally emerge from this formulation, allowing us to estimate multimodal interactions beyond redundancy. We test CoMM both in a controlled and in a series of real-world settings: in the former, we demonstrate that CoMM effectively captures redundant, unique and synergistic information between modalities. In the latter, CoMM learns complex multimodal interactions and achieves state-of-the-art results on seven multimodal tasks.", "keywords": [ "Multimodal representation learning", "Self-supervised learning", "Contrastive learning" ], "rating_list": [ 8, 8, 6, 3 ], "soundness_list": [ 4, 4, 2, 1 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "Pdh1yMqwev", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Juho Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Yohan Jung", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Postdoc" } ], "n_formula": 94, "n_formula_1": 23, "n_ref_uni": 20, "n_ref": 49, "n_ref_all": 152, "n_fig": 27, "n_tab": 3, "L_tab": 1825, "n_element_tab": 75, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 983, "n_element_tab_1": 41, "formula_len_all": 8252, "formula_len_all_1": 1475, "len_all": 193244, "len_all_1": 71422, "len_abs": 1377, "len_title": 134, "len_sents": 63493, "len_sents_1": 30238, "n_sents": 532, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1382, "title": "Adaptive Priors from Learning Trajectories for Function-Space Bayesian Neural Networks", "abs": "Tractable Function-space Variational Inference (T-FVI) provides a way to estimate the function-space Kullback-Leibler (KL) divergence between a random prior function and its posterior. This allows the optimization of the function-space KL divergence via Stochastic Gradient Descent (SGD) and thus simplifies the training of function-space Bayesian Neural Networks (BNNs). However, function-space BNNs on high-dimensional datasets typically require deep neural networks (DNN) with numerous parameters, and thus defining suitable function-space priors remains challenging. For instance, the Gaussian Process (GP) prior suffers from scalability issues, and DNNs do not provide a clear way to set appropriate weight parameters to achieve meaningful function-space priors. To address this issue, we propose an explicit form of function-space priors that can be easily integrated into widely-used DNN architectures, while adaptively incorporating different levels of uncertainty based on the function's inputs. To achieve this, we consider DNNs as Bayesian last-layer models to\nobtain the explicit mean and variance functions of our prior. The parameters of these explicit functions are determined using the weight statistics over the learning trajectory. Our empirical experiments show improved uncertainty estimation in image classification, transfer learning, and UCI regression tasks.", "keywords": [ "Function-space Bayesian neural network", "Function-space variational inference", "Gaussian process", "Stochastic weight averaging gaussian (SWAG))" ], "rating_list": [ 5, 5, 5, 1, 3 ], "soundness_list": [ 3, 2, 3, 1, 2 ], "presentation_list": [ 3, 1, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 1, 2 ], "confidence_list": [ 3, 2, 3, 3, 4 ] }, { "paper_id": "PdZkfSttGK", "primary_area": "applications to neuroscience & cognitive science", "n_author": 1, "author_info_list": [ { "name": "Ganchao Wei", "gender": "Male", "institution": "Duke University", "country": "", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 75, "n_ref": 104, "n_ref_all": 138, "n_fig": 21, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 15, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1242, "formula_len_all_1": 739, "len_all": 153931, "len_all_1": 66489, "len_abs": 1713, "len_title": 143, "len_sents": 39919, "len_sents_1": 33657, "n_sents": 295, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1721, "title": "Nonparametric Covariance Regression for Massive Neural Data on Restricted Covariates via Graph", "abs": "Modern recording techniques enable neuroscientists to simultaneously study neural activity across large populations of neurons, with capturing predictor-dependent correlations being a fundamental challenge in neuroscience. Moreover, the fact that input covariates often lie in restricted subdomains, according to experimental settings, makes inference even more challenging. To address these challenges, we propose a set of nonparametric mean-covariance regression models for high-dimensional neural activity with restricted inputs. These models reduce the dimensionality of neural responses by employing a lower-dimensional latent factor model, where both factor loadings and latent factors are predictor-dependent, to jointly model mean and covariance across covariates. The smoothness of neural activity across experimental conditions is modeled nonparametrically using two Gaussian processes (GPs), applied to both loading basis and latent factors. Additionally, to account for the covariates lying in restricted subspace, we incorporate graph information into the covariance structure. To flexibly infer the model, we use an MCMC algorithm to sample from posterior distributions. After validating and studying the properties of proposed methods by simulations, we apply them to two neural datasets (local field potential and neural spiking data) to demonstrate the usage of models for continuous and counting observations. Overall, the proposed methods provide a framework to jointly model covariate-dependent mean and covariance in high dimensional neural data, especially when the covariates lie in restricted domains. The framework is general and can be easily adapted to various applications beyond neuroscience.", "keywords": [ "covariance regression", "latent variable modeling", "Gaussian process", "graph Laplacian", "time series", "spatiotemporal data", "neural data" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "PdDm14eXO4", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Minghui Fang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Rongjie Huang", "gender": "Male", "institution": "FAIR", "country": "US", "position": "Researcher" }, { "name": "Siqi Zheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Tao Jin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xize Cheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zehan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Ziang Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Hu Ruofan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 35, "n_ref": 84, "n_ref_all": 99, "n_fig": 4, "n_tab": 5, "L_tab": 1427, "n_element_tab": 192, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1595, "n_element_tab_1": 157, "formula_len_all": 28, "formula_len_all_1": 0, "len_all": 146285, "len_all_1": 60810, "len_abs": 1219, "len_title": 124, "len_sents": 42584, "len_sents_1": 28443, "n_sents": 330, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1246, "title": "AVSET-10M: An Open Large-Scale Audio-Visual Dataset with High Correspondence", "abs": "Recent research initiatives such as ChatGPT and Sora highlight the important role of large-scale data in advancing generative and comprehension tasks. However, the scarcity of comprehensive and large-scale audio-visual correspondence datasets poses a significant challenge to research in the audio-visual field. To address this gap, we introduce **AVSET-10M**, a high-correspondence audio-visual dataset comprising 10 million samples, featuring the following key attributes: (1) **High Audio-Visual Correspondence**: Through meticulous sample filtering, we ensure a strong correspondence between the audio and visual components of each entry. (2) **Comprehensive Categories**: Encompassing 527 unique audio categories, AVSET-10M provides a wide range of audio categories for diverse research needs. (3) **Large Scale**: With 10 million samples, AVSET-10M is one of the largest publicly available audio-visual correspondence datasets. We have benchmarked two critical tasks on AVSET-10M: audio-video retrieval and vision-queried sound separation. These tasks underscore the importance of precise audio-visual correspondence in advancing audio-visual research. For more information, please visit our demo page at \\url{https://avset-10m.github.io/}.", "keywords": [ "udio-visual corresponding dataset", "sound separation", "audio-video retrieval" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "PdA9HAxO4w", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Bin Chen", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Hao Fang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Intern" }, { "name": "Hao Wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jiawei Kong", "gender": "Male", "institution": "National Institute of Information and Communications Technology (NICT), National Institute of Advanced Industrial Science and Technology", "country": "JP", "position": "Intern" }, { "name": "Jiawei Li", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Ke Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Shu-Tao Xia", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Wenbo Yu", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 47, "n_ref": 109, "n_ref_all": 138, "n_fig": 9, "n_tab": 18, "L_tab": 10966, "n_element_tab": 743, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 5679, "n_element_tab_1": 315, "formula_len_all": 764, "formula_len_all_1": 537, "len_all": 188921, "len_all_1": 71393, "len_abs": 1437, "len_title": 169, "len_sents": 54357, "len_sents_1": 29837, "n_sents": 428, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 121, "L_abs": 1444, "title": "One Perturbation is Enough: On Generating Universal Adversarial Perturbations against Vision-Language Pre-training Models", "abs": "Vision-Language Pre-training (VLP) models have exhibited unprecedented capability in many applications by taking full advantage of the multimodal alignment. However, previous studies have shown they are vulnerable to maliciously crafted adversarial samples. Despite recent success, these methods are generally instance-specific and require generating perturbations for each input sample. In this paper, we reveal that VLP models are also vulnerable to the instance-agnostic universal adversarial perturbation (UAP). Specifically, we design a novel Contrastive-training Perturbation Generator with Cross-modal conditions (C-PGC) to achieve the attack. In light that the pivotal multimodal alignment is achieved through the advanced contrastive learning technique, we devise to turn this powerful weapon against themselves, i.e., employ a malicious version of contrastive learning to train the C-PGC based on our carefully crafted positive and negative image-text pairs for essentially destroying the alignment relationship learned by VLP models. Besides, C-PGC fully utilizes the characteristics of Vision-and-Language (V+L) scenarios by incorporating both unimodal and cross-modal information as effective guidance. Extensive experiments show that C-PGC successfully forces adversarial samples to move away from their original area in the VLP model's feature space, thus essentially enhancing attacks across various victim models and V+L tasks.", "keywords": [ "Universal Adversarial Attacks", "Vision-Language Pretraining Models", "Generative Attacks" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 4, 3, 3, 1 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "Pd7IOswRUZ", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Kalliopi Basioti", "gender": "Female", "institution": "Nokia Bell Labs", "country": "US", "position": "Intern" }, { "name": "Pritish Sahu", "gender": "Male", "institution": "SRI International", "country": "", "position": "Researcher" }, { "name": "Tony Qingze Liu", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Zihao Xu", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Hao Wang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Assistant Professor" }, { "name": "vladimir pavlovic", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 21, "n_ref": 91, "n_ref_all": 147, "n_fig": 4, "n_tab": 11, "L_tab": 3723, "n_element_tab": 335, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1965, "n_element_tab_1": 126, "formula_len_all": 2076, "formula_len_all_1": 1597, "len_all": 154760, "len_all_1": 58459, "len_abs": 1286, "len_title": 116, "len_sents": 57115, "len_sents_1": 23849, "n_sents": 440, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1290, "title": "GenVP: Generating Visual Puzzles with Contrastive Hierarchical VAEs", "abs": "Raven\u2019s Progressive Matrices (RPMs) is an established benchmark to examine\nthe ability to perform high-level abstract visual reasoning (AVR). Despite the current success of algorithms that solve this task, humans can generalize beyond a given puzzle and create new puzzles given a set of rules, whereas machines remain locked in solving a fixed puzzle from a curated choice list. We propose Generative Visual Puzzles (GenVP), a framework to model the entire RPM generation process, a substantially more challenging task. Our model\u2019s capability spans from generating multiple solutions for one specific problem prompt to creating complete new puzzles out of the desired set of rules. Experiments on five different datasets indicate that GenVP achieves state-of-the-art (SOTA) performance both in puzzle-solving accuracy and out-of-distribution (OOD) generalization in 22 out\nof 24 OOD scenarios. Further, compared to SOTA generative approaches, which struggle to solve RPMs when the feasible solution space increases, GenVP efficiently generalizes to these challenging scenarios. Moreover, our model demonstrates the ability to produce a wide range of complete RPMs given a set of abstract rules by effectively capturing the relationships between abstract rules and visual object properties.", "keywords": [ "abstract visual reasoning; contrastive learning; variational autoencoders; out of distribution generalization" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "Pd3jVGTacT", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Chongyu Fan", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Jiancheng Liu", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "MS student" }, { "name": "Jinghan Jia", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Licong Lin", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Ruiqi Zhang", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Sijia Liu", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Assistant Professor" }, { "name": "Song Mei", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 57, "n_ref": 174, "n_ref_all": 221, "n_fig": 15, "n_tab": 8, "L_tab": 3161, "n_element_tab": 447, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 583, "n_element_tab_1": 28, "formula_len_all": 681, "formula_len_all_1": 617, "len_all": 193387, "len_all_1": 77418, "len_abs": 1512, "len_title": 132, "len_sents": 53020, "len_sents_1": 36800, "n_sents": 381, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1197, "title": "Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning", "abs": "In this work, we address the problem of large language model (LLM) unlearning, aiming to remove unwanted data influences and associated model capabilities e.g., copyrighted data or harmful content generation) while preserving essential model utilities, without the need for retraining from scratch. Despite the growing need for LLM unlearning, a principled optimization framework remains lacking. To this end, we revisit the state-of-the-art approach, negative preference optimization (NPO), and identify the issue of reference model bias, which could undermine NPO's effectiveness, particularly when unlearning forget data of varying difficulty. Given that, we propose a simple yet effective unlearning optimization framework, called SimNPO, showing that `simplicity' in removing the reliance on a reference model (through the lens of simple preference optimization) benefits unlearning. We also provide deeper insights into SimNPO's advantages, supported by analysis using mixtures of Markov chains. Furthermore, we present extensive experiments validating SimNPO's superiority over existing unlearning baselines in benchmarks like TOFU and MUSE, and robustness against relearning attacks.", "keywords": [ "Unlearning", "Large language model", "Preference optimization" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "PcE0yAGAGW", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Ali Lashgari", "gender": "Male", "institution": "Kansas State University", "country": "US", "position": "PhD student" }, { "name": "Elnaz Lashgari", "gender": "Female", "institution": "Abbott", "country": "US", "position": "Researcher" }, { "name": "Mohamamd Salehan", "gender": "Male", "institution": "California Polytechnic State University, Pomona", "country": "", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 7, "n_ref": 15, "n_ref_all": 31, "n_fig": 4, "n_tab": 1, "L_tab": 1518, "n_element_tab": 73, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2539, "n_element_tab_1": 85, "formula_len_all": 78, "formula_len_all_1": 78, "len_all": 76627, "len_all_1": 60209, "len_abs": 1653, "len_title": 136, "len_sents": 30525, "len_sents_1": 28794, "n_sents": 212, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1661, "title": "FSL-MIC: An Attentional Few-Shot Learning Framework for EEG Motor Imagery Classification", "abs": "Electroencephalography (EEG) is a key non-invasive technique used to investigate brain activity, particularly in motor imagery (MI) research. Traditional methods for classifying EEG signals often rely on handcrafted features and heuristic parameters, which can limit generalization across tasks and subjects. Recent advances in deep learning, particularly few-shot learning (FSL), offer promising alternatives to improve classification accuracy in scenarios with limited training data. This study explores the effectiveness of FSL algorithms, including Relation Networks, to enhance MI classification. It also examines how transfer learning and data augmentation techniques contribute to improving classification performance.\n\nWe propose a novel framework with three core modules\u2014feature embedding, attention, and relation\u2014that facilitates the classification of unseen subject categories using only a few labeled samples. The attention mechanism identifies key features related to the query data, while the relation module predicts query labels by modeling relationships between support and query data across subjects. Our experimental results demonstrate the effectiveness of our approach on two benchmark datasets, BCI 2a and BCI 2b, as well as our experimental dataset. The proposed FSL framework significantly outperforms traditional methods, offering promising applications in real-time Brain-Computer Interface (BCI) systems across various EEG setups. This research advances the understanding of machine learning in EEG applications and highlights the potential of FSL techniques in overcoming the challenges of limited training data in MI classification.", "keywords": [ "Few-shot learning", "Data Augmentation", "EEG", "motor imagery", "BCI", "Transformer", "CNN" ], "rating_list": [ 3, 3, 1, 3, 1 ], "soundness_list": [ 2, 3, 1, 2, 1 ], "presentation_list": [ 2, 2, 2, 2, 1 ], "contribution_list": [ 2, 1, 1, 2, 1 ], "confidence_list": [ 4, 4, 5, 4, 5 ] }, { "paper_id": "Pc94ncbkoo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Di Huang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Yanan Zhang", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Postdoc" }, { "name": "Yunlong Qi", "gender": "Male", "institution": "None", "country": "", "position": "Researcher" }, { "name": "Zehua Fu", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Researcher" }, { "name": "Jiaxin Chen", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Associate Professor" }, { "name": "Wang Zengran", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 39, "n_ref": 68, "n_ref_all": 77, "n_fig": 4, "n_tab": 3, "L_tab": 1325, "n_element_tab": 95, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 586, "n_element_tab_1": 9, "formula_len_all": 673, "formula_len_all_1": 595, "len_all": 109974, "len_all_1": 53751, "len_abs": 1074, "len_title": 107, "len_sents": 29698, "len_sents_1": 24974, "n_sents": 216, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1079, "title": "$A^2$-DP: Annotation-aware Data Pruning for Object Detection", "abs": "As the size of datasets for training deep neural networks expands, data pruning has become an intriguing area of research due to its ability to achieve lossless performance with a reduced overall data volume. However, traditional data pruning usually demands complete dataset annotations, incurring high costs. To tackle this, we propose an innovative Annotation-Aware Data Pruning paradigm tailored for object detection, dubbed\nas $A^2$-DP, which aims to reduce the burdens of both annotation and storage. Our approach, consisting of two phases, integrates a hard sample mining module to extract crucial hidden objects, a class balance module to identify important objects in rare or challenging classes and a global similarity removal module that enhances the elimination of redundant information through object-level similarity assessments.\nExtensive experiments on 2D and 3D detection tasks validate the effectiveness of the $A^2$-DP, consistently achieving a minimum pruning rate of 20\\% across various datasets, showcasing the practical value and efficiency of our methods.", "keywords": [ "Object Detection", "Data Pruning", "Active Learning" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "Pbz4i7B0B4", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Hui Xiong", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Junren Xiao", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Meng Xiao", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Postdoc" }, { "name": "Qingqiang Sun", "gender": "Male", "institution": "Great Bay University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiao Luo", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Ziyue Qiao", "gender": "Male", "institution": "Great Bay University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 46, "n_formula_1": 19, "n_ref_uni": 49, "n_ref": 119, "n_ref_all": 132, "n_fig": 6, "n_tab": 4, "L_tab": 1076, "n_element_tab": 110, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 979, "n_element_tab_1": 96, "formula_len_all": 2790, "formula_len_all_1": 1284, "len_all": 185446, "len_all_1": 70868, "len_abs": 1681, "len_title": 76, "len_sents": 57905, "len_sents_1": 31645, "n_sents": 440, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1508, "title": "Towards Continuous Reuse of Graph Models via Holistic Memory Diversification", "abs": "This paper addresses the challenge of incremental learning in growing graphs with increasingly complex tasks. The goal is to continuously train a graph model to handle new tasks while retaining proficiency on previous tasks via memory replay. Existing methods usually overlook the importance of memory diversity, limiting in selecting high-quality memory from previous tasks and remembering broad previous knowledge within the scarce memory on graphs. To address that, we introduce a novel holistic Diversified Memory Selection and Generation (DMSG) framework for incremental learning in graphs, which first introduces a buffer selection strategy that considers both intra-class and inter-class diversities, employing an efficient greedy algorithm for sampling representative training nodes from graphs into memory buffers after learning each new task. Then, to adequately rememorize the knowledge preserved in the memory buffer when learning new tasks, a diversified memory generation replay method is introduced. This method utilizes a variational layer to generate the distribution of buffer node embeddings and sample synthesized ones for replaying. Furthermore, an adversarial variational embedding learning method and a reconstruction-based decoder are proposed to maintain the integrity and consolidate the generalization of the synthesized node embeddings, respectively. Extensive experimental results on publicly accessible datasets demonstrate the superiority of DMSG over state-of-the-art methods.", "keywords": [ "Incremental learning", "Model Reuse", "Transfer learning" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "PbxKOPtoEE", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Hang Yin", "gender": "Male", "institution": "Department of Mathematical science, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hanghang Tong", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Shukai Zhao", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "MS student" }, { "name": "Wei Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yangqiu Song", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Zihao Wang", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Fei Weizhi", "gender": "Male", "institution": "The Department of Mathematics, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 34, "n_ref": 110, "n_ref_all": 136, "n_fig": 4, "n_tab": 5, "L_tab": 3047, "n_element_tab": 829, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 364, "n_element_tab_1": 72, "formula_len_all": 1589, "formula_len_all_1": 1428, "len_all": 142241, "len_all_1": 67764, "len_abs": 1574, "len_title": 117, "len_sents": 43120, "len_sents_1": 30977, "n_sents": 327, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1472, "title": "Neural Logical Index for Fast Knowledge Graph Complex Query Answering", "abs": "Complex query answering (CQA) over knowledge graphs is a crucial multi-hop reasoning task aimed at addressing first-order logical queries within large and incomplete knowledge graphs. Direct traversal search methods rely solely on graph topology and often miss answers due to the incompleteness of the graph, thus neural models have been proposed to generalize the neglected answers from observed facts. There are primarily two lines of research tackling the challenges of CQA. Query embedding models learn representations for complex queries, offering fast speed but often providing only generic performance. In contrast, neural symbolic search methods deliver better performance, although they tend to be computationally more expensive. In this paper, we propose an efficient and scalable search framework that combines the precision of symbolic methods with the speed of embedding techniques. Our model utilizes embedding methods to compute Neural Logical Indices (NLI) to reduce the search domain for each variable in advance, followed by an approximate symbolic search for fine ranking. The search is precise for tree-structured queries and approximates cyclic queries (which are NP-complete) in quadratic complexity concerning the search domain, matching the complexity of tree-form queries. Experiments on various CQA benchmarks show that our framework reduces computation by 90% with a minimal performance loss, alleviating both efficiency and scalability issues.", "keywords": [ "complex logical query", "knowledge graph", "query embedding", "neural link predictor" ], "rating_list": [ 5, 6, 5, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 2, 3, 2, 4 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "PbheqxnO1e", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Anil Kag", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Chaoyang Wang", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Chenfanfu Jiang", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "Associate Professor" }, { "name": "Jian Ren", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Ju Hu", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Junli Cao", "gender": "Male", "institution": "UCLA Computer Science Department, University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Sergei Korolev", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Machine Learning Engineer" }, { "name": "Sergey Tulyakov", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Director of Research" }, { "name": "Vidit Goel", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 44, "n_ref": 140, "n_ref_all": 155, "n_fig": 9, "n_tab": 4, "L_tab": 1432, "n_element_tab": 204, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1263, "n_element_tab_1": 239, "formula_len_all": 347, "formula_len_all_1": 173, "len_all": 150146, "len_all_1": 72777, "len_abs": 470, "len_title": 139, "len_sents": 44162, "len_sents_1": 32456, "n_sents": 374, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 1510, "title": "Lightweight Predictive 3D Gaussian Splats", "abs": "Recent approaches representing 3D objects and scenes using Gaussian splats show increased rendering speed across a variety of platforms and devices. While rendering such representations is indeed extremely efficient, storing and transmitting them is often prohibitively expensive. To represent large-scale scenes, one often needs to store millions of 3D Gaussian, which can occupy up to gigabytes of storage. This creates a significant practical barrier, preventing widespread adoption on resource-constrained devices.\nIn this work, we propose a new representation that dramatically reduces the hard drive footprint while featuring similar or improved quality when compared to the standard 3D Gaussian splats. This representation leverages the inherent feature sharing among splats in the close proximity using a hierarchical tree structure, with which only the parent splats need to be stored. We present a method for constructing tree structures from naturally unstructured point clouds. Additionally, we propose the adaptive tree manipulation to prune the redundant trees in the space, while spawn new ones from the significant children splats during the optimization process. On the benchmark datasets, we achieve 20x storage reduction in hard-drive footprint with improved fidelity compared to the vanilla 3DGS and 2-5x reduction compared to the exiting compact solutions. More importantly, we demonstrate the practical application of our method in real-world rendering on mobile devices and AR glasses.", "keywords": [ "Gaussian splatting" ], "rating_list": [ 8, 8, 5 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 4, 4, 2 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "PbC786k7qc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Conghao Xiong", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Hao Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Hao Zheng", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Irwin King", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Jiahong Liu", "gender": "Female", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Joseph JY Sung", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Xian Wu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Principal Researcher" }, { "name": "Yefeng Zheng", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" } ], "n_formula": 23, "n_formula_1": 23, "n_ref_uni": 43, "n_ref": 115, "n_ref_all": 132, "n_fig": 2, "n_tab": 3, "L_tab": 1998, "n_element_tab": 105, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 3036, "n_element_tab_1": 147, "formula_len_all": 1467, "formula_len_all_1": 1417, "len_all": 133994, "len_all_1": 64295, "len_abs": 1434, "len_title": 83, "len_sents": 32365, "len_sents_1": 26219, "n_sents": 228, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1440, "title": "Enhancing Multimodal Survival Prediction with Pathology Reports in Hyperbolic Space", "abs": "Cancer survival prediction using computational pathology has emerged as a crucial tool for diagnosis and treatment planning. Current approaches primarily rely on Whole Slide Images (WSIs) and genomic data, but face significant challenges in capturing the logical relationships between visual features and survival outcomes with limited supervision. While pathology reports could potentially serve as a semantic bridge between WSIs and survival time, existing methods overlook the inherent hierarchical relationships between textual descriptions and visual features, where pathology terms represent more abstract concepts and individual terms may correspond to multiple image regions. To address these challenges, we propose HyperSurv, a novel framework that leverages hyperbolic geometry to model the hierarchical relationships between WSIs and pathology reports. Our key insight is that hyperbolic space naturally captures both the entailment structure between generic report concepts and specific visual features, as well as the one-to-many relationships between pathology terms and image regions. HyperSurv enforces these relationships through hyperbolic cones while identifying survival-relevant features via attention pooling. Extensive experiments on four TCGA cancer datasets demonstrate that our approach achieves state-of-the-art survival prediction performance by effectively modeling these multi-modal hierarchical relationships.", "keywords": [ "Survival Prediction", "Computational Pathology", "Multimodal Medical Image Analysis" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "PauyrluLud", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Ofir Lindenbaum", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "Assistant Professor" }, { "name": "Oren Glickman", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "Assistant Professor" }, { "name": "Yaniv Zimmer", "gender": "unknown", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 44, "n_ref_all": 62, "n_fig": 11, "n_tab": 7, "L_tab": 2861, "n_element_tab": 183, "n_fig_1": 11, "n_tab_1": 6, "L_tab_1": 1702, "n_element_tab_1": 142, "formula_len_all": 120, "formula_len_all_1": 120, "len_all": 103136, "len_all_1": 59864, "len_abs": 1319, "len_title": 162, "len_sents": 28289, "len_sents_1": 27757, "n_sents": 180, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 114, "L_abs": 1326, "title": "Supervised Band Selection with a Concrete Layer for Hyperspectral Imagery in Remote Sensing and Autonomous Driving", "abs": "Hyperspectral imagery captures rich spectral information, which is valuable for a wide range of applications but poses challenges due to high data dimensionality. Current band selection methods are often computationally intensive, non-embedded, or lack adaptability for specific tasks. We address this gap by introducing a novel plug-and-play embedded method for supervised band selection in hyperspectral imagery, utilizing a concrete selector layer based on the Gumbel-Softmax re-parameterization trick. Our approach allows for dynamic and task-specific selection of optimal bands, eliminating the need for pre-processing and enabling seamless integration with downstream models. We evaluated the method on four hyperspectral datasets, covering three remote sensing benchmarks and an autonomous driving task, demonstrating consistent improvements over state-of-the-art methods. This is the first work to perform comprehensive band-selection research on an autonomous driving dataset of this type, and the first to employ a concrete layer for supervised band selection. Our findings highlight the potential of this approach for real-time hyperspectral analysis in applications such as autonomous driving and environmental monitoring, laying the groundwork for further exploration of efficient, domain-specific band selection.", "keywords": [ "hyperspectral imagery", "band selection", "Gumbel-Softmax", "concrete layer", "remote sensing", "autonomous driving", "deep learning", "semantic segmentation", "plug-and-play models" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "PageLgQlXz", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Jinrong Sheng", "gender": "Male", "institution": "Chongqing University", "country": "CN", "position": "MS student" }, { "name": "Wenshi Li", "gender": "Male", "institution": "Chongqing University", "country": "CN", "position": "PhD student" }, { "name": "Yongxin Ge", "gender": "Male", "institution": "Chongqing University", "country": "CN", "position": "Full Professor" }, { "name": "Li Ao", "gender": "Male", "institution": "Chongqing University", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 24, "n_ref": 35, "n_ref_all": 44, "n_fig": 3, "n_tab": 7, "L_tab": 1720, "n_element_tab": 156, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 352, "n_element_tab_1": 28, "formula_len_all": 708, "formula_len_all_1": 650, "len_all": 132581, "len_all_1": 61222, "len_abs": 1479, "len_title": 124, "len_sents": 32809, "len_sents_1": 27492, "n_sents": 258, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1486, "title": "Dual-level Prototypes Guidance for Single-frame Temporal Action Localization", "abs": "In recent, single-frame temporal action localization (STAL) has captured the attention of the computer vision community. Due to the sparse single-frame annotations, current STAL methods generally employ pseudo-labels strategies to bridge the gap between weakly-supervised methods and fully-supervised methods. However, these methods derive pseudo-labels from single-frame of the corresponding instances, yet the intra-class affinity from the current single-frame to other action snippets remains neglected. To capitalize on this affinity, we design a dual-level prototypes guidance (DPG) method with the graph matching random walk (Gm-Rw) algorithm to achieve instance-level and video-level prototype guidance for pseudo-labels refinement. For instance-level guidance, the Gm-Rw exploits the high affinity prototype among instances of the current video to build intra-class associations. For video-level guidance, an online memory bank is constructed to iteratively summarize more discriminative prototype. After Gm-Rw builds affinity among intra-class videos, an exponential moving average (EMA) mechanism is designed to achieve dual-level prototypes guidance for pseudo-labels refinement. Notably, the dual-level guidance is mutually reinforcing, prompting us to propose a novel adaptive collaborative strategy (ACS) for dynamic optimization. Extensive experiments on THUMOS14, GTEA, BEOID, and ActivityNet1.3 reveal that our method significantly outperforms state-of-the-art methods.", "keywords": [ "Temporal action localization", "Single-frame annotations", "Memory bank", "Graph matching random walk" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 1, 1, 2, 4 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "PacBhLzeGO", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "JiaKui Hu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Lujia Jin", "gender": "Male", "institution": "China Mobile Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Yanye Lu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhengjian Yao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 65, "n_ref": 189, "n_ref_all": 206, "n_fig": 4, "n_tab": 19, "L_tab": 6463, "n_element_tab": 540, "n_fig_1": 1, "n_tab_1": 10, "L_tab_1": 2769, "n_element_tab_1": 124, "formula_len_all": 281, "formula_len_all_1": 246, "len_all": 200316, "len_all_1": 61832, "len_abs": 1505, "len_title": 119, "len_sents": 47964, "len_sents_1": 27139, "n_sents": 360, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1594, "title": "Universal Image Restoration Pre-training via Degradation Classification", "abs": "This paper proposes the Degradation Classification Pre-Training (DCPT), which enables models to learn how to classify the degradation type of input images for universal image restoration pre-training. Unlike the existing self-supervised pre-training methods, DCPT utilizes the degradation type of the input image as an extremely weak supervision, which can be effortlessly obtained, even intrinsic in all image restoration datasets. DCPT comprises two primary stages. Initially, image features are extracted from the encoder. Subsequently, a lightweight decoder, such as ResNet18, is leveraged to classify the degradation type of the input image solely based on the features extracted in the first stage, without utilizing the input image. The encoder is pre-trained with a straightforward yet potent DCPT, which is used to address universal image restoration and achieve outstanding performance. Following DCPT, both convolutional neural networks (CNNs) and transformers demonstrate performance improvements, with gains of up to 2.55 dB in the 10D all-in-one restoration task and 6.53 dB in the mixed degradation scenarios. Moreover, previous self-supervised pretraining methods, such as masked image modeling, discard the decoder after pre-training, while our DCPT utilizes the pre-trained parameters more effectively. This superiority arises from the degradation classifier acquired during DCPT, which facilitates transfer learning between models of identical architecture trained on diverse degradation types. Source code and models are available at \\url{https://github.com/MILab-PKU/dcpt}.", "keywords": [ "Pre-training; Degradation Classfication; Universal restoration;" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "PabAln0jjB", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Brooks Paige", "gender": "Male", "institution": "University College London", "country": "", "position": "Associate Professor" }, { "name": "Sayak Ray Chowdhury", "gender": "Male", "institution": "Indian Institute of Technology, Kanpur", "country": "IN", "position": "Assistant Professor" }, { "name": "Seongho Son", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "William Bankes", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Ilija Bogunovic", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Lecturer" } ], "n_formula": 139, "n_formula_1": 35, "n_ref_uni": 64, "n_ref": 136, "n_ref_all": 214, "n_fig": 8, "n_tab": 3, "L_tab": 351, "n_element_tab": 56, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 732, "n_element_tab_1": 35, "formula_len_all": 13349, "formula_len_all_1": 2205, "len_all": 257224, "len_all_1": 67180, "len_abs": 1187, "len_title": 139, "len_sents": 67786, "len_sents_1": 28299, "n_sents": 625, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1099, "title": "Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift", "abs": "Current Large Language Model (LLM) preference optimization algorithms do not\naccount for temporal preference drift, which can lead to severe misalignment. To\naddress this limitation, we propose an offline fine-tuning algorithm Non-Stationary\nDirect Preference Optimisation (NS-DPO) which models time-dependent reward\nfunctions with a Dynamic Bradley-Terry model. NS-DPO applies exponential\nweighting, by introducing a discount parameter in the loss function, which pro-\nportionally focuses learning on more time-relevant datapoints. We theoretically\nanalyse the convergence of NS-DPO, providing upper bounds on the estimation\nerror and regret caused by non-stationary preferences. Finally, we demonstrate\nthe effectiveness of NS-DPO1 for fine-tuning LLMs in scenarios with drifting\npreferences. By simulating preference drift using popular LLM reward models\nand datasets accordingly, we show that NS-DPO fine-tuned LLMs remain robust\nunder non-stationarity, significantly outperforming baseline algorithms that ignore\ntemporal preference changes, without sacrificing performance in stationary cases.", "keywords": [ "LLM", "fine-tuning", "DPO", "non-stationarity", "preference drift", "RLHF" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "Pa1vr1Prww", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Arthur Conmy", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Dmitrii Kharlapenko", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Fazl Barez", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Neel Nanda", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Stepan Shabalin", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Undergrad student" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 40, "n_ref": 89, "n_ref_all": 132, "n_fig": 19, "n_tab": 4, "L_tab": 237, "n_element_tab": 154, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 237, "n_element_tab_1": 154, "formula_len_all": 599, "formula_len_all_1": 135, "len_all": 213322, "len_all_1": 57148, "len_abs": 1169, "len_title": 107, "len_sents": 57293, "len_sents_1": 25961, "n_sents": 466, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1172, "title": "Scaling Sparse Feature Circuits For Studying In-Context Learning", "abs": "Sparse autoencoders (SAEs) are a popular tool for interpreting large language\nmodel activations, but their utility in addressing open questions in interpretability\nremains unclear. In this work, we demonstrate their effectiveness by using SAEs\nto deepen our understanding of the mechanism behind in-context learning (ICL).\nWe identify abstract SAE features that encode the model\u2019s knowledge of which\ntask to execute and whose latent vectors causally induce the task zero-shot. This\naligns with prior work showing that ICL is mediated by task vectors. We further\ndemonstrate that these task vectors are well approximated by a sparse sum of SAE\nlatents, including these task-execution features. To explore the ICL mechanism,\nwe adapt the sparse feature circuits methodology of Marks et al. (2024) to work for\nthe much larger Gemma-1 2B model, with 30 times as many parameters, and to\nthe more complex task of ICL. Through circuit finding, we discover task-detecting\nfeatures with corresponding SAE latents that activate earlier in the prompt, that\ndetect when tasks have been performed. They are causally linked with task-\nexecuting features through attention layer and MLP.", "keywords": [ "SAE", "ICL", "SFC", "Interpretability", "Gemma", "LLM" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 3, 3, 4, 1 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 3, 3, 3, 1 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "PZf4RsPMBG", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Nghi D. Q. Bui", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Researcher" }, { "name": "Phan Nh\u1eadt Huy", "gender": "Male", "institution": "VinUniversity", "country": "VN", "position": "Undergrad student" }, { "name": "Tien N Nguyen", "gender": "unknown", "institution": "university of texas at dallas", "country": "", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 103, "n_ref_all": 119, "n_fig": 4, "n_tab": 14, "L_tab": 13080, "n_element_tab": 462, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 2520, "n_element_tab_1": 232, "formula_len_all": 105, "formula_len_all_1": 0, "len_all": 270973, "len_all_1": 68445, "len_abs": 1039, "len_title": 130, "len_sents": 77784, "len_sents_1": 29376, "n_sents": 558, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1375, "title": "HyperAgent: Generalist Software Engineering Agents to Solve Coding Tasks at Scale", "abs": "Large Language Models (LLMs) have revolutionized software engineering (SE), demonstrating remarkable capabilities in various coding tasks. While recent efforts have produced autonomous software\nagents based on LLMs for end-to-end development tasks, these systems are typically designed for\nspecific SE tasks. We introduce HyperAgent , a novel generalist multi-agent system designed to\naddress a wide spectrum of SE tasks across different programming languages by mimicking human\ndevelopers\u2019 workflows. Comprising four specialized agents\u2014Planner, Navigator, Code Editor, and Executor\u2014HyperAgent manages the full lifecycle of SE tasks, from initial conception to final verification.\nThrough extensive evaluations, HyperAgent achieves state-of-the-art performance across diverse\nSE tasks: it attains a 26.00% success rate on SWE-Bench-Lite and 33.00% on SWE-Bench-Verified for\nGitHub issue resolution, surpassing existing methods. Furthermore, HyperAgent demonstrates\nsuperior performance in code generation at repository scale (RepoExec), and in fault localization and\nprogram repair (Defects4J), often outperforming specialized systems. This work represents a significant\nadvancement towards versatile, autonomous agents capable of handling complex, multi-step SE tasks\nacross various domains and languages, potentially transforming AI-assisted software development\npractices.", "keywords": [ "Multi-Agent", "Large Language Model", "Software Engineering" ], "rating_list": [ 6, 3, 6, 5, 5 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 5 ] }, { "paper_id": "PZYr22zFyE", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Haeyun Lee", "gender": "Male", "institution": "Korea University of Technology and Education", "country": "KR", "position": "Assistant Professor" }, { "name": "Jae Youn Hwang", "gender": "Male", "institution": "DGIST", "country": "KR", "position": "Full Professor" }, { "name": "Kyungsu Lee", "gender": "Male", "institution": "Jeonbuk National University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 31, "n_formula_1": 15, "n_ref_uni": 65, "n_ref": 94, "n_ref_all": 128, "n_fig": 5, "n_tab": 2, "L_tab": 789, "n_element_tab": 116, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 545, "n_element_tab_1": 70, "formula_len_all": 2071, "formula_len_all_1": 956, "len_all": 189157, "len_all_1": 67827, "len_abs": 1871, "len_title": 143, "len_sents": 58190, "len_sents_1": 32270, "n_sents": 359, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1878, "title": "Connectome Mapping: Shape-Memory Network via Interpretation of Contextual Semantic Information", "abs": "Contextual semantic information plays a pivotal role in the brain's visual interpretation of the surrounding environment. When processing visual information, electrical signals within synapses facilitate the dynamic activation and deactivation of synaptic connections, guided by the contextual semantic information associated with different objects. In the realm of Artificial Intelligence (AI), neural networks have emerged as powerful tools to emulate complex signaling systems, enabling tasks such as classification and segmentation by understanding visual information. However, conventional neural networks have limitations in simulating the conditional activation and deactivation of synapses, collectively known as the connectome, a comprehensive map of neural connections in the brain. Additionally, the pixel-wise inference mechanism of conventional neural networks failed to account for the explicit utilization of contextual semantic information in the prediction process. To overcome these limitations, we developed a novel neural network, dubbed the Shape Memory Network (SMN), which excels in two key areas: (1) faithfully emulating the intricate mechanism of the brain's connectome, and (2) explicitly incorporating contextual semantic information during the inference process. The SMN memorizes the structure suitable for contextual semantic information and leverages this structure at the inference phase. The structural transformation emulates the conditional activation and deactivation of synaptic connections within the connectome. Rigorous experimentation carried out across a range of semantic segmentation benchmarks demonstrated the outstanding performance of the SMN, highlighting its superiority and effectiveness. Furthermore, our pioneering network on connectome emulation reveals the immense potential of the SMN for next-generation neural networks.", "keywords": [ "neural representation" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 4, 4, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "PZVVOeu6xx", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Miguel E. P. Silva", "gender": "Male", "institution": "INESC TEC", "country": "", "position": "Postdoc" }, { "name": "Pedro C Vieira", "gender": "Male", "institution": "Universidade do Porto", "country": "PT", "position": "MS student" }, { "name": "Pedro Manuel Pinto Ribeiro", "gender": "Female", "institution": "Universidade do Porto", "country": "", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 4, "n_ref_uni": 64, "n_ref": 105, "n_ref_all": 143, "n_fig": 11, "n_tab": 4, "L_tab": 1825, "n_element_tab": 219, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 837, "formula_len_all_1": 246, "len_all": 216203, "len_all_1": 73036, "len_abs": 4898, "len_title": 112, "len_sents": 73958, "len_sents_1": 38164, "n_sents": 520, "n_sents_1": 264, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1306, "title": "Predicting Network Motif Fingerprints with Graph Neural Networks", "abs": "Graph Neural Networks (GNNs) are a predominant method for graph representation learning. However, beyond subgraph frequency estimation, their application to network motif prediction remains underexplored, with no established benchmarks in the literature. We propose to address this problem, framing motif prediction as an extension of subgraph frequency estimation. Our approach formulates motif estimation as a multitarget regression problem, optimising for interpretability and improving stability and scalability on large graphs. We validate our method using a large synthetic dataset generated by graph generators that mimic real-world data, and further test it on real-world graphs. Our experiments reveal that 1-WL limited models trained on synthetic data struggle to predict accurately motif profiles of real-world networks. However, apart from their reasonable performance within synthetic data, they can generalise to approximate the graph generation processes of real-world networks by comparing their predicted motif profiles with the ones originating from synthetic data. This first study on GNN-based motif prediction sets a benchmark and should open pathways for further developing the connection between motif profiles and subgraph frequency from a graph representation learning perspective.", "keywords": [ "motifs", "graph representation learning", "synthetic data", "significance profiles" ], "rating_list": [ 3, 3, 1, 3, 3 ], "soundness_list": [ 2, 2, 1, 2, 3 ], "presentation_list": [ 1, 2, 1, 3, 2 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 5, 4, 3 ] }, { "paper_id": "PYmrUQmMEw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Qingkai Fang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Shaolei Zhang", "gender": "Male", "institution": "Key Laboratory of Intelligent Information Processing Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Shoutao Guo", "gender": "Male", "institution": "Institute of computing technology, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Yan Zhou", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yang Feng", "gender": "unknown", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhengrui Ma", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 47, "n_ref": 91, "n_ref_all": 114, "n_fig": 3, "n_tab": 9, "L_tab": 1995, "n_element_tab": 340, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 677, "n_element_tab_1": 115, "formula_len_all": 23559, "formula_len_all_1": 24285, "len_all": 224096, "len_all_1": 107442, "len_abs": 1394, "len_title": 114, "len_sents": 42201, "len_sents_1": 28569, "n_sents": 319, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1302, "title": "LLaMA-Omni: Seamless Speech Interaction with Large Language Models", "abs": "Models like GPT-4o enable real-time interaction with large language models (LLMs) through speech, significantly enhancing user experience compared to traditional text-based interaction. However, there is still a lack of exploration on how to build speech interaction models based on open-source LLMs. To address this, we propose LLaMA-Omni, a novel model architecture designed for low-latency and high-quality speech interaction with LLMs. LLaMA-Omni integrates a pretrained speech encoder, a speech adaptor, an LLM, and a streaming speech decoder. It eliminates the need for speech transcription, and can simultaneously generate text and speech responses directly from speech instructions with extremely low latency. We build our model based on the latest Llama-3.1-8B-Instruct model. To align the model with speech interaction scenarios, we construct a dataset named InstructS2S-200K, which includes 200K speech instructions and corresponding speech responses. Experimental results show that compared to previous speech-language models, LLaMA-Omni provides better responses in both content and style, with a response latency as low as 226ms. Additionally, training LLaMA-Omni takes less than 3 days on just 4 GPUs, paving the way for the efficient development of speech-language models in the future.", "keywords": [ "large language models", "speech interaction", "speech-to-speech", "speech-language models" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "PYQmaU4RwI", "primary_area": "learning theory", "n_author": 1, "author_info_list": [ { "name": "Arthur Jun ZHANG", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 14, "n_ref": 20, "n_ref_all": 28, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 636, "formula_len_all_1": 692, "len_all": 85264, "len_all_1": 59680, "len_abs": 1624, "len_title": 104, "len_sents": 33113, "len_sents_1": 27962, "n_sents": 275, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1629, "title": "A Novel Dual of Shannon Information and Weighting Scheme", "abs": "Shannon Information theory has achieved great success in not only communication technology where it was originally developed for but also many other science and engineering fields such as machine learning and artificial intelligence. Inspired by the famous weighting scheme TF-IDF, we discovered that Shannon information entropy actually has a natural dual. To complement the classical Shannon information entropy which measures the uncertainty we propose a novel information quantity, namely troenpy. Troenpy measures the certainty and commonness of the underlying distribution. So entropy and troenpy form an information twin. To demonstrate its usefulness, we propose a conditional troenpy based weighting scheme for document with class labels, namely positive class frequency (PCF). On a collection of public datasets we show the PCF based weighting scheme outperforms the classical TF-IDF and a popular Optimal Transport based word moving distance algorithm in a kNN setting with respectively more than 22.9 and 26.5 classification error reduction while the corresponding entropy based approach completely fails. We further developed a new odds-ratio type feature, namely Expected Class Information Bias(ECIB), which can be regarded as the expected odds ratio of the information twin across different classes. In the experiments we observe that including the new ECIB features and simple binary term features in a simple logistic regression model can further significantly improve the performance. The proposed simple new weighting scheme and ECIB features are very effective and can be computed with linear time complexity.", "keywords": [ "entropy", "certainty", "uncertainty", "weighting" ], "rating_list": [ 6, 5, 1 ], "soundness_list": [ 3, 3, 1 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 1 ], "confidence_list": [ 5, 2, 5 ] }, { "paper_id": "PY7KSh29Z8", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Bishmoy Paul", "gender": "Male", "institution": "Santa Clara University", "country": "US", "position": "MS student" }, { "name": "Md Awsafur Rahman", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Najibul Haque Sarker", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "MS student" }, { "name": "Shaikh Anowarul Fattah", "gender": "unknown", "institution": "Bangladesh University of Engineering and Technology", "country": "BD", "position": "Full Professor" }, { "name": "Zaber Ibn Abdul Hakim", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 32, "n_ref": 75, "n_ref_all": 108, "n_fig": 9, "n_tab": 9, "L_tab": 2996, "n_element_tab": 509, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2767, "n_element_tab_1": 344, "formula_len_all": 227, "formula_len_all_1": 227, "len_all": 181624, "len_all_1": 67499, "len_abs": 1724, "len_title": 56, "len_sents": 62133, "len_sents_1": 29852, "n_sents": 406, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 56, "L_abs": 1726, "title": "SONICS: Synthetic Or Not - Identifying Counterfeit Songs", "abs": "The recent surge in AI-generated songs presents exciting possibilities and challenges. These innovations necessitate the ability to distinguish between human-composed and synthetic songs to safeguard artistic integrity and protect human musical artistry. Existing research and datasets in fake song detection only focus on singing voice deepfake detection (SVDD), where the vocals are AI-generated but the instrumental music is sourced from real songs. However, these approaches are inadequate for detecting contemporary end-to-end artificial songs where all components (vocals, music, lyrics, and style) could be AI-generated. Additionally, existing datasets lack music-lyrics diversity, long-duration songs, and open-access fake songs. To address these gaps, we introduce SONICS, a novel dataset for end-to-end Synthetic Song Detection (SSD), comprising over 97k songs (4,751 hours) with over 49k synthetic songs from popular platforms like Suno and Udio. Furthermore, we highlight the importance of modeling long-range temporal dependencies in songs for effective authenticity detection, an aspect entirely overlooked in existing methods. To utilize long-range patterns, we introduce SpecTTTra, a novel architecture that significantly improves time and memory efficiency over conventional CNN and Transformer-based models. In particular, for long audio samples, our top-performing variant outperforms ViT by 8% F1 score while being 38% faster and using 26% less memory. Additionally, in comparison with ConvNeXt, our model achieves 1% gain in F1 score with 20% boost in speed and 67% reduction in memory usage. Other variants of our model family provide even better speed and memory efficiency with competitive performance.", "keywords": [ "deepfake detection", "fake song detection", "synthetic song detection", "efficient model", "dataset", "audio processing" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "PY56Wur7S0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Ashish Tiwari", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Gust Verbruggen", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Mukul Singh", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Sumit Gulwani", "gender": "Male", "institution": "Research, Microsoft", "country": "", "position": "Researcher" }, { "name": "Vu Le", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 24, "n_ref": 54, "n_ref_all": 68, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1312, "formula_len_all_1": 1397, "len_all": 100344, "len_all_1": 57213, "len_abs": 1459, "len_title": 122, "len_sents": 30221, "len_sents_1": 25515, "n_sents": 232, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1466, "title": "Execution-guided within-prompt search for programming-by-example", "abs": "Soundness is an important property in programming-by-example (PBE) as it allows synthesizers to perform a search over a domain-specific language (DSL) that terminates when any sound program is found.\nLarge language models (LLMs) can generate code from examples without being limited to a DSL, but they lack search, as samples are independent.\nOne can sampling code until a sound program is generated, but that is very inefficient.\nIn this paper, we use an LLM as a policy that generates lines of code and then join these lines of code to let the LLM implicitly estimate the value of each of these lines in its next iteration.\nWe further guide the policy and value estimation by executing each line and annotating it with its results on the given examples. \nThis allows us to search for programs within a single, expanding prompt until a sound program is found by letting the policy reason in both the syntactic (code) and semantic (execution) space.\nWe evaluate this approach on straight-line Python code generation using five benchmarks across different domains (string transformations, list transformations, and arbitrary Python programming problems).\nWe show that the model effectively uses the execution results to guide the search and that within-prompt search performs well at low token budgets.\nWe also analyze how the model behaves as a policy and value, show that it can parallelize the search, and that it can implicitly backtrack over earlier generations.", "keywords": [ "Programming-by-example", "program synthesis", "large language models" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "PWtx9fJqM5", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Mehran Hosseini", "gender": "unknown", "institution": "King's College London, University of London", "country": "GB", "position": "Postdoc" }, { "name": "Peyman Hosseini", "gender": "Male", "institution": "Queen Mary University of London", "country": "GB", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 37, "n_ref_uni": 40, "n_ref": 54, "n_ref_all": 89, "n_fig": 8, "n_tab": 9, "L_tab": 4654, "n_element_tab": 380, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 108, "n_element_tab_1": 24, "formula_len_all": 1436, "formula_len_all_1": 2085, "len_all": 149962, "len_all_1": 57914, "len_abs": 1477, "len_title": 87, "len_sents": 44460, "len_sents_1": 25286, "n_sents": 324, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1459, "title": "A Study of Necessity & Sufficiency of Linear Transformations in the Attention Mechanism", "abs": "Scaled Dot Product Attention (SDPA) is the backbone of many modern\n deep-learning models. It is so versatile that it has been used in\n natural language, vision, and multi-modal domains with very little\n change compared to its original formulation. This paper studies the linear transformations used in SDPA. To this end, we introduce three variants of the attention mechanism by removing consecutive linear transformations or adding an extra one. We name these variants Optimized ($W^V$ removed),\n Efficient ($W^V$ and $W^K$ removed), and Super Attention ($W^V$ and $W^K$ removed and $W^A$ introduced) to simplify comparison when referring to them. In addition to providing the mathematical intuition behind these choices, we evaluate these variants on several datasets of varying size and complexity in vision and text modalities for predictive and generative tasks. Optimized and\n Efficient variants have one and two matrix multiplications fewer\n per head, respectively, and 25\\% and 50\\% fewer parameters,\n respectively, than standard SDPA. However, the performance change compared to difference in parameter count is small. Super Attention introduces a new linear transformation\n on the values, transforming them from the left. It outperforms\n standard SPDA in both modalities by up to 10\\%\n while having one fewer matrix multiplication per head and 25\\% fewer\n parameters than standard SPDA. Consequently, it is also faster than standard SDPA.", "keywords": [ "Transformers", "Attention", "Self-Attention" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 2 ] }, { "paper_id": "PWia19rgzV", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Cheyu Lin", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Katherine A. Flanigan", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Sirajum Munir", "gender": "Male", "institution": "Bosch", "country": "US", "position": "Senior Research Scientist" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 19, "n_ref": 40, "n_ref_all": 65, "n_fig": 10, "n_tab": 3, "L_tab": 1177, "n_element_tab": 161, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1269, "n_element_tab_1": 169, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 110814, "len_all_1": 65616, "len_abs": 2108, "len_title": 133, "len_sents": 38829, "len_sents_1": 32450, "n_sents": 272, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 2154, "title": "Your Actions Talk: DUET - A Multimodal Dataset for Contextualizable Dyadic Activities", "abs": "Human activity recognition (HAR) has advanced significantly with the availability of diverse datasets, yet the field remains limited by a scarcity of datasets focused on two-person, or ''dyadic,'' interactions. Existing datasets primarily cater to single-person activities, overlooking the complex dynamics and contextual dependencies present in interactions between two individuals. Failing to extend HAR to dyadic settings limits opportunities to advance areas like collaborative learning, healthcare, robotics, augmented reality, and psychological assessments, which require an understanding of interpersonal dynamics. To address this gap, we introduce the Dyadic User Engagement dataseT (DUET), a comprehensive dataset designed to enhance the understanding and recognition of dyadic activities. DUET comprises 14,400 video samples across 12 interaction classes, capturing the highest sample-to-class ratio of dyadic datasets known to date. Each sample is recorded using RGB, depth, infrared, and 3D skeleton joints, ensuring a robust dataset for multimodal analysis. Critically, DUET features a taxonomization of interactions based on five fundamental communication functions: emblems, illustrators, affect displays, regulators, and adaptors. This classification, rooted in psychology, supports dyadic human activity contextualization by extracting the embedded semantics of bodily movements. Data collection was conducted at three locations using a novel technique that captures interactions from multiple views with a single camera, thereby improving model resilience against background noise and view variations. We benchmark six state-of-the-art, open-source HAR algorithms on DUET, demonstrating the dataset's complexity and current HAR models' limitations in recognizing dyadic interactions. Our results highlight the need for further research into multimodal and context-aware HAR for dyadic interactions, and provide a dataset to support this advancement. DUET is publicly available at \\href{https://huggingface.co/datasets/Anonymous-Uploader1/DUET}, providing a valuable resource for the research community to advance HAR in dyadic settings.", "keywords": [ "Dyadic activity datasets", "dyadic human activity recognition", "contextualization", "kinesics" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "PW1Mj6lLh2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Dylan Zhang", "gender": "Male", "institution": "Department of Computer Science", "country": "US", "position": "PhD student" }, { "name": "Francois Charton", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Engineer" }, { "name": "Justin Wang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "Undergrad student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 28, "n_ref": 63, "n_ref_all": 79, "n_fig": 8, "n_tab": 3, "L_tab": 1703, "n_element_tab": 570, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1703, "n_element_tab_1": 570, "formula_len_all": 94, "formula_len_all_1": 22, "len_all": 184079, "len_all_1": 68710, "len_abs": 1458, "len_title": 130, "len_sents": 35484, "len_sents_1": 27819, "n_sents": 252, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1521, "title": "$\\textbf{Only-IF}$: Revealing the Decisive Effect of Instruction Diversity on Generalization", "abs": "Understanding and accurately following instructions is critical for large language models (LLMs) to be effective across diverse tasks. In this work, we conduct a rigorous investigation into the factors that enable generalization to unseen instructions. Through controlled experiments, inspired by the Turing-complete Markov algorithm, we demonstrate that such generalization $\\textbf{only emerges}$ when training data is diversified enough across semantic domains. Our findings also reveal that merely diversifying within limited domains fails to ensure robust generalization. In contrast, cross-domain data diversification, even under constrained data budgets, significantly enhances a model's adaptability. We further extend our analysis to real-world scenarios, including fine-tuning of $\\textit{\\textbf{{specialist}}}$ and $\\textit{\\textbf{{generalist}}}$ models. Our research provides important insights for dataset collation, particularly when optimizing model performance by expanding training data for both specialist and generalist scenarios. We show that careful consideration of data diversification is key: training specialist models with data extending beyond their core domain leads to significant performance improvements, while generalist models benefit from diverse data mixtures that enhance their overall instruction-following capabilities across a wide range of applications. . Our results highlight the critical role of\nstrategic diversification and offer clear guidelines for improving data quality.", "keywords": [ "Instruction Following; Generalization" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "PVHoELf5UN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Haoqian Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "HuXiaowan", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Li Huaqiu", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 17, "n_formula_1": 18, "n_ref_uni": 21, "n_ref": 25, "n_ref_all": 38, "n_fig": 9, "n_tab": 4, "L_tab": 827, "n_element_tab": 176, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 842, "n_element_tab_1": 190, "formula_len_all": 1673, "formula_len_all_1": 1739, "len_all": 100164, "len_all_1": 50458, "len_abs": 1131, "len_title": 141, "len_sents": 24780, "len_sents_1": 21717, "n_sents": 179, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1136, "title": "Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios", "abs": "Real-world low-light images often suffer from complex degradations such as local overexposure, low brightness, noise, and uneven illumination. Supervised methods tend to overfit to specific scenarios, while unsupervised methods, though better at generalization, struggle to model these degradations due to the lack of reference images. To address this issue, we propose an interpretable, zero-reference joint denoising and low-light enhancement framework tailored for real-world scenarios. Our method derives a training strategy based on paired sub-images with varying illumination and noise levels, grounded in physical imaging principles and retinex theory. Additionally, we leverage the Discrete Cosine Transform (DCT) to perform frequency domain decomposition in the sRGB space, and introduce an implicit-guided hybrid representation strategy that effectively separates intricate compounded degradations. In the backbone network design, we develop retinal decomposition network guided by implicit degradation representation mechanisms. Extensive experiments demonstrate the superiority of our method. The code will be released soon.", "keywords": [ "unsupervised learning", "low-light image enhancement", "image denoising" ], "rating_list": [ 6, 8, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 5, 5, 4, 3, 5 ] }, { "paper_id": "PVGS8UZ6GX", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Adina Williams", "gender": "Female", "institution": "FAIR (Meta Platforms Inc.)", "country": "", "position": "Research Scientist" }, { "name": "Mark Ibrahim", "gender": "unknown", "institution": "Facebook AI Research (FAIR) Meta", "country": "US", "position": "Researcher" }, { "name": "Michael Rabbat", "gender": "Male", "institution": "Facebook", "country": "CA", "position": "Principal Researcher" }, { "name": "Niklas Nolte", "gender": "Not Specified", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Ouail Kitouni", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 16, "n_ref": 59, "n_ref_all": 103, "n_fig": 15, "n_tab": 3, "L_tab": 722, "n_element_tab": 80, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 575, "n_element_tab_1": 134, "formula_len_all": 88, "formula_len_all_1": 88, "len_all": 122105, "len_all_1": 49515, "len_abs": 1504, "len_title": 106, "len_sents": 44983, "len_sents_1": 23865, "n_sents": 341, "n_sents_1": 149, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1507, "title": "Transformers Can Navigate Mazes With Multi-Step Prediction", "abs": "Despite their remarkable success in language modeling, transformers trained to predict the next token in a sequence struggle with long-term planning. This limitation is particularly evident in tasks requiring foresight to plan multiple steps ahead such as maze navigation. The standard next single token prediction objective, however, offers no explicit mechanism to predict multiple steps ahead---or revisit the path taken so far. Consequently, in this work we study whether explicitly predicting multiple steps ahead (and backwards) can improve transformers' maze navigation. We train under identical settings, parameter-matched transformers from scratch to navigate mazes of varying types and sizes with standard next token prediction and MLM-U: an objective explicitly predicting multiple steps ahead and backwards. We find MLM-U considerably improves transformers\u2019 ability to navigate mazes compared to standard next token prediction across maze types and complexities. We also find MLM-U training is 4x more sample efficient and converges 2x faster in terms of GPU training hours relative to next token training. Finally, for more complex mazes we find MLM-U benefits from scaling to larger transformers. Remarkably, we find transformers trained with MLM-U outperform larger transformers trained with next token prediction using additional supervision from A* search traces. We hope these findings underscore the promise of learning objectives to advance transformers' capacity for long-term planning.", "keywords": [ "transformers", "planning", "learning objectives" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "PUnD86UEK5", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Mohamad Amin Mohamadi", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "PhD student" }, { "name": "Shuo Xie", "gender": "unknown", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "PhD student" }, { "name": "Zhiyuan Li", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Assistant Professor" } ], "n_formula": 84, "n_formula_1": 30, "n_ref_uni": 25, "n_ref": 40, "n_ref_all": 53, "n_fig": 2, "n_tab": 8, "L_tab": 266, "n_element_tab": 25, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 154, "n_element_tab_1": 6, "formula_len_all": 17700, "formula_len_all_1": 4085, "len_all": 152779, "len_all_1": 64501, "len_abs": 1077, "len_title": 124, "len_sents": 39677, "len_sents_1": 25180, "n_sents": 416, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 85, "L_abs": 1114, "title": "Adam Exploits $\\ell_\\infty$-geometry of Loss Landscape via Coordinate-wise Adaptivity", "abs": "Adam outperforms SGD when training language models. Yet such benefits are not well-understood theoretically -- previous convergence analysis for Adam and SGD mainly focuses on the number of steps $T$ and is already minimax-optimal in non-convex cases, which are both $O(T^{-1/4})$. In this work, we argue that the better dependence on the loss smoothness is the key advantage of Adam over SGD. More specifically, we give a new convergence analysis for Adam under novel assumptions that loss is smooth under $\\ell_\\infty$ geometry rather than the more common $\\ell_2$ geometry, which yields a much better empirical smoothness constant for GPT-2 and ResNet models. Moreover, we show that if we rotate the training loss randomly, Adam can be outperformed by some variants of SGD which is invariant to rotations. This implies that any practically relevant explanation of Adam's optimization benefit must involve non-rotational invariant properties of loss, such as $\\ell_\\infty$ smoothness as used in our analysis. We also extend the convergence analysis to blockwise Adam, which is a generalization of standard Adam.", "keywords": [ "Adam", "coordinate-wise adaptivity", "adaptive algorithms", "infinity norm" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "PUXy7vQ5M3", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Erik Strumbelj", "gender": "Male", "institution": "University of Ljubljana", "country": "SI", "position": "Full Professor" }, { "name": "Martin Jurkovi\u010d", "gender": "Male", "institution": "University of Ljubljana", "country": "SI", "position": "MS student" }, { "name": "Valter Hudovernik", "gender": "unknown", "institution": "University of Ljubljana", "country": "", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 39, "n_ref": 82, "n_ref_all": 133, "n_fig": 13, "n_tab": 10, "L_tab": 5734, "n_element_tab": 657, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 67, "n_element_tab_1": 6, "formula_len_all": 2155, "formula_len_all_1": 1867, "len_all": 205564, "len_all_1": 60452, "len_abs": 4976, "len_title": 114, "len_sents": 64752, "len_sents_1": 27849, "n_sents": 481, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1093, "title": "Benchmarking the Fidelity and Utility of Synthetic Relational Data", "abs": "Synthesizing relational data has started to receive more attention from researchers, practitioners, and industry. The task is more difficult than synthesizing a single table due to the added complexity of relationships between tables. For the same reason, benchmarking methods for synthesizing relational data introduces new challenges. Our work is motivated by a lack of an empirical evaluation of state-of-the-art methods and by gaps in the understanding of how such an evaluation should be done. We review related work on relational data synthesis, common benchmarking datasets, and approaches to measuring the fidelity and utility of synthetic data. We combine the best practices and a novel robust detection approach into a benchmarking tool and use it to compare six methods, including two commercial tools. While some methods are better than others, no method is able to synthesize a dataset that is indistinguishable from original data. For utility, we typically observe moderate correlation between real and synthetic data for both model predictive performance and feature importance.", "keywords": [ "data generation", "literature review", "empirical comparison", "discriminative detection", "data quality" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "PTjKXwrVCT", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Abdul Fatir Ansari", "gender": "Male", "institution": "AWS AI Labs (Amazon)", "country": "DE", "position": "Senior Scientist" }, { "name": "Bernie Wang", "gender": "Male", "institution": "AWS AI Labs", "country": "US", "position": "Principal Researcher" }, { "name": "Boran Han", "gender": "Female", "institution": "Amazon/AWS", "country": "US", "position": "Researcher" }, { "name": "Derek Qiang Xu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Mukul R. Prasad", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Rahul Sharma", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Xiyuan Zhang", "gender": "unknown", "institution": "AWS", "country": "US", "position": "Applied Scientist" }, { "name": "Yizhou Sun", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "SHUAI Zhang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" } ], "n_formula": 16, "n_formula_1": 12, "n_ref_uni": 41, "n_ref": 94, "n_ref_all": 124, "n_fig": 14, "n_tab": 1, "L_tab": 282, "n_element_tab": 42, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 2474, "n_element_tab_1": 30, "formula_len_all": 1667, "formula_len_all_1": 1380, "len_all": 149668, "len_all_1": 65413, "len_abs": 1327, "len_title": 93, "len_sents": 42273, "len_sents_1": 28808, "n_sents": 341, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1332, "title": "Forecasting Needles in a Time Series Haystack", "abs": "Shocks and sudden spikes are common characteristics of real-world time series data. For example, demand surges or electricity outages often occur in time series data, manifesting as spikes (\u201cNeedles\u201d) added to the regular time series (\u201cHaystack\u201d). Despite their importance, it is surprising to find their absence in the benchmarking protocol at the frontier of time series research\u2014Time Series Foundation Models (TSFMs). To address this gap, we present the Needle-in-a-Time-Series-Haystack (NiTH) Benchmark, which includes both synthetic and real-world spiky time series data from diverse domains like traffic, energy, and biomedical systems. For synthetic data, we develop a flexible framework using Poisson-based modeling to generate spiky time series, allowing us to evaluate forecast models under various conditions. To accurately assess model performance, we introduce a new metric based on Dynamic Time Warping, specifically designed for spiky data. We evaluate the zero-shot forecasting capabilities of 6 popular TSFMs over 64 million observations, identifying their limitations related to architecture, tokenization, and loss functions. Furthermore, we demonstrate that the incorporation of the proposed NiTH dataset, due to its diversity compared to the common pre-training corpus of TSFMs, results in improved performance.", "keywords": [ "Time Series Forecasting", "Zero-Shot Forecasting", "Time Series Benchmark" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "PTgTlj6x0W", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Feiyue Chen", "gender": "Female", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" }, { "name": "Gelei Deng", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Guowei Yang", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" }, { "name": "Ling Shi", "gender": "unknown", "institution": "Nanyang Technological University ", "country": "SG", "position": "Postdoc" }, { "name": "Tianwei Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Yang Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Yi Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Yuqi Chen", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 12, "n_ref": 35, "n_ref_all": 56, "n_fig": 7, "n_tab": 2, "L_tab": 963, "n_element_tab": 133, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 524, "n_element_tab_1": 41, "formula_len_all": 69, "formula_len_all_1": 69, "len_all": 83517, "len_all_1": 59614, "len_abs": 997, "len_title": 81, "len_sents": 32809, "len_sents_1": 29675, "n_sents": 243, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1011, "title": "TREANT: Red-teaming Text-to-Image Models with Tree-based Semantic Transformations", "abs": "The increasing prevalence of text-to-image (T2I) models makes their safety a critical concern. Adversarial testing techniques have been developed to probe whether such models can be prompted to produce Not-Safe-For-Work (NSFW) content. Despite these efforts, current solutions face several challenges, such as low success rates, inefficiency, and lack of semantic understanding. To address these issues, we introduce TREANT, a novel automated red-teaming framework for adversarial testing of T2I models. The core of our framework is the tree-based semantic transformation. We employ semantic decomposition and sensitive element drowning strategies in conjunction with Large Language Models (LLMs) to systematically refine adversarial prompts for effective testing. Our comprehensive evaluation confirms the efficacy of TREANT, which not only exceeds the performance of state-of-the-art approaches but also achieves an overall success rate of 88.5% on leading T2I models, including DALL\u00b7E 3 and Stable Diffusion.", "keywords": [ "Red-teaming" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "PTcMzQgKmn", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 10, "author_info_list": [ { "name": "Bumsik Kim", "gender": "Male", "institution": "DeepAuto.ai", "country": "KR", "position": "Researcher" }, { "name": "Geon Park", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Heejun Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Hyemin Lee", "gender": "Female", "institution": "Deepauto.ai", "country": "KR", "position": "Researcher" }, { "name": "Jaduk Suh", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Undergrad student" }, { "name": "Jina Kim", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Undergrad student" }, { "name": "Myeongjae Jeon", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Sung Ju Hwang", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Wonyong Jeong", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "Ph.D. student" }, { "name": "Youngwan Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 64, "n_formula_1": 11, "n_ref_uni": 38, "n_ref": 88, "n_ref_all": 169, "n_fig": 25, "n_tab": 15, "L_tab": 4270, "n_element_tab": 385, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 2942, "n_element_tab_1": 323, "formula_len_all": 7654, "formula_len_all_1": 818, "len_all": 317259, "len_all_1": 77727, "len_abs": 1812, "len_title": 155, "len_sents": 113281, "len_sents_1": 33837, "n_sents": 997, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 107, "L_abs": 1829, "title": "A Training-Free Sub-quadratic Cost Transformer Model Serving Framework with Hierarchically Pruned Attention", "abs": "In modern large language models (LLMs), increasing the context length is crucial for improving comprehension and coherence in long-context, multi-modal, and retrieval-augmented language generation. \nWhile many recent transformer models attempt to extend their context length over a million tokens, they remain impractical due to the quadratic time and space complexities.\nAlthough recent works on linear and sparse attention mechanisms can achieve this goal, their real-world applicability is often limited by the need to re-train from scratch and significantly worse performance. In response, we propose a novel approach, Hierarchically Pruned Attention (HiP), which reduces the time complexity of the attention mechanism to $O(T \\log T)$ and the space complexity to $O(T)$, where $T$ is the sequence length. \nWe notice a pattern in the attention scores of pretrained LLMs where tokens close together tend to have similar scores, which we call \"attention locality\". Based on this observation, we utilize a novel tree-search-like algorithm that estimates the top-$k$ key tokens for a given query on the fly, which is mathematically guaranteed to have better performance than random attention pruning. In addition to improving the time complexity of the attention mechanism, we further optimize GPU memory usage by implementing KV cache offloading, which stores only $O(\\log T)$ tokens on the GPU while maintaining similar decoding throughput. Experiments on benchmarks show that HiP, with its training-free nature, significantly reduces both prefill and decoding latencies, as well as memory usage, while maintaining high-quality generation with minimal degradation.\nHiP enables pretrained LLMs to scale up to millions of tokens on commodity GPUs, potentially unlocking long-context LLM applications previously deemed infeasible.", "keywords": [ "Efficient Attention Mechanism", "Long-context LLM Decoding", "KV Cache Offloading" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "PSzDG612AC", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Chengliang Liu", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Postdoc" }, { "name": "Jie Wen", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Jinghua Wang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yabo Liu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yong Xu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Huang Chao", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 17, "n_ref": 39, "n_ref_all": 46, "n_fig": 3, "n_tab": 4, "L_tab": 1184, "n_element_tab": 139, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2086, "n_element_tab_1": 133, "formula_len_all": 849, "formula_len_all_1": 849, "len_all": 120862, "len_all_1": 65087, "len_abs": 1274, "len_title": 128, "len_sents": 30076, "len_sents_1": 27065, "n_sents": 234, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1279, "title": "Text-driven Zero-shot Domain Adaptation with Cross-modality Graph Motif Matching", "abs": "Zero-shot domain adaptive semantic adaptation aims to transfer knowledge from a source domain and learn a target segmenter without access to any target domain data. Some existing methods have achieved notable performances by transforming source features to the target domain through language-driven methods. However, these methods often align language features to global image features coarsely resulting in sub-optimal performance. To address the challenges, we propose a graph motif-based adaptation method designed to balance the efficiency and effectiveness of feature alignment. Our approach involves constructing motif structures based on domain-wise image feature distributions. By increasing the angle between language-vision directed edges, we effectively pull visual features toward the language feature center, thereby achieving cross-modality feature alignment. Additionally, we employ relationship-constraint losses, \\ie directional and contrastive losses, to mitigate the mode-collapse during target feature stylization. These relationship-constraint losses help stabilize the learning process and improve the robustness of the adaptation. Extensive experimental results validate the efficacy of our proposed method. The code for this method will be made available.", "keywords": [ "computer vision", "transfer learning", "multi-modality", "zero-shot domain adaptation" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 1, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "PSiijdQjNU", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 10, "author_info_list": [ { "name": "Hao Zhou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jingjing Gong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Shuyi Zhang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Siyu Long", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Wei-Ying Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Wenhao Huang", "gender": "Male", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Yu Pei", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yuxuan Song", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Intern" }, { "name": "Zhe Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Ziyao Cao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 57, "n_formula_1": 19, "n_ref_uni": 34, "n_ref": 99, "n_ref_all": 125, "n_fig": 10, "n_tab": 7, "L_tab": 7212, "n_element_tab": 577, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1497, "n_element_tab_1": 182, "formula_len_all": 5002, "formula_len_all_1": 1422, "len_all": 208284, "len_all_1": 62172, "len_abs": 966, "len_title": 108, "len_sents": 67505, "len_sents_1": 27029, "n_sents": 504, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 60, "L_abs": 968, "title": "Steering Protein Family Design through Profile Bayesian Flow", "abs": "Protein family design emerges as a promising alternative by combining the advantages of de novo protein design and mutation-based directed evolution.In this paper, we propose ProfileBFN, the Profile Bayesian Flow Networks, for specifically generative modeling of protein families. ProfileBFN extends the discrete Bayesian Flow Network from an MSA profile perspective, which can be trained on single protein sequences by regarding it as a degenerate profile, thereby achieving efficient protein family design by avoiding large-scale MSA data construction and training. Empirical results show that ProfileBFN has a profound understanding of proteins. When generating diverse and novel family proteins, it can accurately capture the structural characteristics of the family. The enzyme produced by this method is more likely than the previous approach to have the corresponding function, offering better odds of generating diverse proteins with the desired functionality.", "keywords": [ "protein family generation", "homologous protein generation", "protein design", "bayesian flow" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 4, 4, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "PSQuy9sjQ8", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Jun Nie", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Mingming Gong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Xinmei Tian", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yonggang Zhang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Postdoc" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 17, "n_ref_uni": 53, "n_ref": 98, "n_ref_all": 126, "n_fig": 6, "n_tab": 8, "L_tab": 491, "n_element_tab": 47, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 880, "n_element_tab_1": 20, "formula_len_all": 1128, "formula_len_all_1": 985, "len_all": 154798, "len_all_1": 61850, "len_abs": 1746, "len_title": 106, "len_sents": 43367, "len_sents_1": 28377, "n_sents": 363, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1784, "title": "Consistency Verification for Detecting AI-Generated Images", "abs": "With the rapid development of generative models, AI-generated images have sparked significant concerns regarding their potential misuse for malicious purposes, highlighting the urgent need for AI-generated image detection. Current methods primarily focus on training a binary classifier to detect generated images. However, the efficacy of these methods is critically dependent on the quantity and quality of the collected AI-generated images. More importantly, they suffer from a generalization challenge: \\emph{the literature lacks sufficient exploration of whether a binary classifier trained on images from a specific diffusion model can effectively generalize to images generated by other models.} In this work, we propose a novel framework termed \\textbf{con}sistency \\textbf{v}erification (ConV) for AI-generated image detection, providing a new approach that detects without requiring AI-generated images. In particular, we introduce two functions and establish a principle for designing them so that their outputs remain consistent for natural images but exhibit signi\ufb01cant inconsistency for AI-generated images. Our principle shows that gradients of these two functions need to lie within two mutually orthogonal subspaces. This enables a training-free detection approach: an image is identified as AI-generated if transformation along its data manifold results in a substantial change in the loss value of a self-supervised model pre-trained on natural images. This detection framework leads to the unique advantage of ConV over existing methods: \\emph{ConV identifies AI-generated images by fitting the distribution of natural images rather than that of AI-generated images.} Extensive experiments across various benchmarks validate the effectiveness of the proposed ConV.", "keywords": [ "AI-generated image detection", "Generative models", "Diffusion models", "GAN" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 1, 1, 3 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "PRvdO3nfFi", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Florian Lemmerich", "gender": "Male", "institution": "Leuphana Universit\u00e4t L\u00fcneburg", "country": "DE", "position": "Instructor" }, { "name": "Klaus Maier-Hein", "gender": "Male", "institution": "German Cancer Research Center", "country": "DE", "position": "Full Professor" }, { "name": "Markus Strohmaier", "gender": "Not Specified", "institution": "Universit\u00e4t Mannheim", "country": "", "position": "Full Professor" }, { "name": "Max Klabunde", "gender": "unknown", "institution": "Universit\u00e4t Passau", "country": "", "position": "PhD student" }, { "name": "Tassilo Wald", "gender": "Male", "institution": "German Cancer Research Center", "country": "DE", "position": "PhD student" }, { "name": "Tobias Schumacher", "gender": "Male", "institution": "Universit\u00e4t Mannheim", "country": "DE", "position": "PhD student" } ], "n_formula": 48, "n_formula_1": 8, "n_ref_uni": 49, "n_ref": 100, "n_ref_all": 147, "n_fig": 14, "n_tab": 23, "L_tab": 37455, "n_element_tab": 6088, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1718, "n_element_tab_1": 126, "formula_len_all": 2397, "formula_len_all_1": 422, "len_all": 365558, "len_all_1": 67470, "len_abs": 2664, "len_title": 120, "len_sents": 79497, "len_sents_1": 32116, "n_sents": 639, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1349, "title": "ReSi: A Comprehensive Benchmark for Representational Similarity Measures", "abs": "Measuring the similarity of different representations of neural architectures is a fundamental task and an open research challenge for the machine learning community. This paper presents the first comprehensive benchmark for evaluating representational similarity measures based on well-defined groundings of similarity. The representational similarity (ReSi) benchmark consists of (i) six carefully designed tests for similarity measures, (ii) 23 similarity measures, (iii) twelve neural network architectures, and (iv) six datasets, spanning over the graph, language, and vision domains. The benchmark opens up several important avenues of research on representational similarity that enable novel explorations and applications of neural architectures. We demonstrate the utility of the ReSi benchmark by conducting experiments on various neural network architectures, real world datasets and similarity measures. All components of the benchmark are publicly available and thereby facilitate systematic reproduction and production of research results. The benchmark is extensible, future research can build on and further expand it. We believe that the ReSi benchmark can serve as a sound platform catalyzing future research that aims to systematically evaluate existing and explore novel ways of comparing representations of neural architectures.", "keywords": [ "representational similarity", "benchmark", "grounding", "representations" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "PRKFRzOEq8", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Coby Penso", "gender": "Male", "institution": "Bar Ilan University", "country": "", "position": "MS student" }, { "name": "Ethan Fetaya", "gender": "Male", "institution": "Bar Ilan University", "country": "IL", "position": "Assistant Professor" }, { "name": "Jacob Goldberger", "gender": "Male", "institution": "Bar-Ilan University", "country": "", "position": "Associate Professor" } ], "n_formula": 37, "n_formula_1": 36, "n_ref_uni": 15, "n_ref": 46, "n_ref_all": 84, "n_fig": 2, "n_tab": 4, "L_tab": 6699, "n_element_tab": 493, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 4264, "n_element_tab_1": 362, "formula_len_all": 2511, "formula_len_all_1": 2365, "len_all": 113948, "len_all_1": 68756, "len_abs": 1193, "len_title": 111, "len_sents": 34157, "len_sents_1": 25140, "n_sents": 316, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1206, "title": "Estimating the conformal prediction threshold from noisy labels", "abs": "Conformal Prediction (CP) is a method to control prediction uncertainty by producing a small prediction set, ensuring a predetermined probability that the true class lies within this set. This is commonly done by defining a score, based on the model predictions, and setting a threshold on this score using a validation set. In this study, we address the problem of CP calibration when we only have access to a validation set with noisy labels. We show how we can estimate the noise-free conformal threshold based on the noisy labeled data. Our solution is flexible and can accommodate various modeling assumptions regarding the label contamination process, without needing any information about the underlying data distribution or the internal mechanisms of the machine learning classifier. We develop a coverage guarantee for uniform noise that is effective even in tasks with a large number of classes. We dub our approach Noise-Aware Conformal Prediction (NACP) and show on several natural and medical image classification datasets, including ImageNet, that it significantly outperforms current noisy label methods and achieves results comparable to those obtained with a clean validation set.", "keywords": [ "conformal prediction", "label noise" ], "rating_list": [ 6, 8, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 1, 2 ], "confidence_list": [ 3, 3, 4, 3, 3 ] }, { "paper_id": "PRJ4n3CBzU", "primary_area": "datasets and benchmarks", "n_author": 1, "author_info_list": [ { "name": "Varun Badrinath Krishna", "gender": "unknown", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 21, "n_ref": 25, "n_ref_all": 45, "n_fig": 4, "n_tab": 12, "L_tab": 5145, "n_element_tab": 283, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2239, "n_element_tab_1": 147, "formula_len_all": 39, "formula_len_all_1": 38, "len_all": 130582, "len_all_1": 60323, "len_abs": 3675, "len_title": 173, "len_sents": 47056, "len_sents_1": 27830, "n_sents": 345, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 124, "L_abs": 1170, "title": "AttackQA: Development and Adoption of a Dataset for Assisting Cybersecurity Operations using Fine-tuned and Open-Source LLMs", "abs": "Retrieval-augmented generation (RAG) on specialized domain datasets has shown improved performance when large language models (LLMs) are fine-tuned for generating responses to user queries. In this study, we develop a cybersecurity question-answering (Q\\&A) dataset, called AttackQA, and employ it to build a RAG-based Q\\&A system designed for analysts in security operations centers. The dataset comprises 25,335 Q\\&A pairs, accompanied by rationales to facilitate fine-tuning and evaluation. 80\\% of the dataset was generated with help of a lightweight open-source LLM (LLama 3 8B), which produced over 1100 tokens per second with full 16-bit precision on specialized hardware. To ensure dataset quality, we fine-tuned LLama 3 70B to detect and reject low-quality Q\\&A pairs. In using the dataset for RAG, we demonstrate that fine-tuning open-source embeddings and LLMs can yield superior accuracy compared to OpenAI's state-of-the-art proprietary embedding and LLM (GPT-4o). Furthermore, we use Llama 3.1 405B as a judge to evaluate answer correctness, enabling the creation of a fully open-source, high-speed RAG and evaluation pipeline with an associated benchmark.", "keywords": [ "LLM", "Large Language Models", "Generative AI", "Dataset", "Cyber", "Security", "MITRE", "Chatbot", "Judging", "Quality", "Control", "Open source", "RAG", "Retrieval", "Augmented", "Generation" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 2, 4, 1, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "PQzZrRNynC", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Danny Dongning Sun", "gender": "Male", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Fei Ma", "gender": "Male", "institution": "Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ)", "country": "CN", "position": "Researcher" }, { "name": "Jian Xu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Shao-Lun Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiao-Ping Zhang", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yanlong Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 36, "n_ref": 83, "n_ref_all": 118, "n_fig": 13, "n_tab": 13, "L_tab": 12749, "n_element_tab": 1003, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1770, "n_element_tab_1": 107, "formula_len_all": 833, "formula_len_all_1": 674, "len_all": 188487, "len_all_1": 58780, "len_abs": 1212, "len_title": 141, "len_sents": 55783, "len_sents_1": 26741, "n_sents": 373, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1216, "title": "PSformer: Parameter-efficient Transformer with Segment Attention for Time Series Forecasting", "abs": "Time series forecasting remains a critical challenge across various domains, often complicated by high-dimensional data and long-term dependencies. This paper presents a novel transformer architecture for time series forecasting, incorporating two key innovations: parameter sharing (PS) and Spatial-Temporal Segment Attention (SegAtt). We also define the time series segment as the concatenation of sequence patches from the same positions across different variables. The proposed model, PSformer, reduces the number of training parameters through the parameter sharing mechanism, thereby improving model efficiency and scalability. The introduction of SegAtt could enhance the capability of capturing local spatio-temporal dependencies by computing attention over the segments, and improve global representation by integrating information across segments. The combination of parameter sharing and SegAtt significantly improves the forecasting performance. Extensive experiments on benchmark datasets demonstrate that PSformer outperforms popular baselines and other transformer-based approaches in terms of accuracy and scalability, establishing itself as an accurate and scalable tool for time series forecasting.", "keywords": [ "Time Series Forecasting", "Transformer", "Parameter Sharing" ], "rating_list": [ 5, 5, 5, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3, 4 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3, 4 ] }, { "paper_id": "PQrkWvQSL0", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Augustin Luna", "gender": "Not Specified", "institution": "National Institutes of Health", "country": "US", "position": "Principal Researcher" }, { "name": "Tianci Song", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "", "position": "PhD student" }, { "name": "Tianfan Fu", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "Assistant Professor" }, { "name": "Yoshitaka Inoue", "gender": "Male", "institution": "National Library of Medicine", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 5, "n_ref_uni": 37, "n_ref": 64, "n_ref_all": 67, "n_fig": 1, "n_tab": 1, "L_tab": 22, "n_element_tab": 1, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 427, "n_element_tab_1": 5, "formula_len_all": 759, "formula_len_all_1": 378, "len_all": 147341, "len_all_1": 58988, "len_abs": 1682, "len_title": 163, "len_sents": 43634, "len_sents_1": 28654, "n_sents": 343, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 114, "L_abs": 1694, "title": "DrugAgent: Multi-Agent Large Language Model-Based Reasoning for Drug-Target Interaction Prediction and Repurposing", "abs": "Advancements in large language models (LLMs) allow them to address a wide set of questions from diverse topics using human-like text interfaces, but limitations in their training prevent them from answering accurately in scenarios that could benefit from multiple perspectives. Multi-agent systems allow the resolution of questions to enhance result consistency and reliability. Here we create a multi-perspective (i.e., unstructured text, structured knowledge graph, and Machine Learning (ML) prediction) multi-agent LLM system. We apply this system to the biologically inspired problem of predicting drug-target interaction. Our system uses a coordinator agent to assign and integrate results for tasks given to three specialized agents: an AI agent for ML predictions, a knowledge graph (KG) agent for KG retrieval, and a search agent for web-based information retrieval. \n\nWe conducted experiments using our LLM-based system for predicting drug-target interaction constants that reflect binding affinities using the BindingDB dataset. Our multi-agent LLM method significantly outperformed GPT-4 across multiple evaluation metrics by a significant margin. An ablation study revealed the contributions by each agent; ranked in terms of a contribution: the AI agent (i.e., ML prediction) was the most important followed by the KG agent then the search agent. The large contribution by the AI agent highlights the importance of LLM tool use in addressing questions that may not be part of text corpora. While our use case was related to biology, our presented architecture is applicable to other integrative prediction tasks. Code is available https://anonymous.4open.science/r/DrugAgent-2BB7/", "keywords": [ "Multi-agent", "Drug-target interaction", "drug-protein binding prediction", "Large Language Models" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 1, 2, 2 ], "presentation_list": [ 1, 1, 4, 2 ], "contribution_list": [ 1, 1, 1, 2 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "PQpvhUrA1C", "primary_area": "foundation or frontier models, including LLMs", "n_author": 12, "author_info_list": [ { "name": "Alan Yuille", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Fangxun Shu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Feng Wang", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Haoqin Tu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Jieru Mei", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Lei Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Linjie Yang", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "peng wang", "gender": "Male", "institution": "Bytedance US AILab", "country": "US", "position": "Research Scientist" }, { "name": "Sucheng Ren", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Xianhang Li", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "cihang xie", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "heng Wang", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "research scientist" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 23, "n_ref": 43, "n_ref_all": 67, "n_fig": 8, "n_tab": 4, "L_tab": 4018, "n_element_tab": 331, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 1512, "n_element_tab_1": 61, "formula_len_all": 445, "formula_len_all_1": 377, "len_all": 108111, "len_all_1": 47214, "len_abs": 1113, "len_title": 96, "len_sents": 22605, "len_sents_1": 21908, "n_sents": 153, "n_sents_1": 136, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1037, "title": "Autoregressive Pretraining with Mamba in Vision", "abs": "The vision community has started to build with the recently developed state space model, Mamba, as the new backbone for a range of tasks. This paper shows that Mamba's visual capability can be significantly enhanced through autoregressive pretraining, a direction not previously explored. Efficiency-wise, the autoregressive nature can well capitalize on the Mamba's unidirectional recurrent structure, enabling faster overall training speed and reduced GPU memory usage compared to other training strategies. Performance-wise, autoregressive pretraining equips the Mamba architecture with markedly higher accuracy over its supervised-trained counterparts and, more importantly, successfully unlocks its scaling potential to large and even huge model sizes. For example, with autoregressive pretraining, a base-size Mamba outperforms its supervised counterpart by 2.0% on ImageNet classification; our best model, a huge-size Mamba, attains 85.0% top-1 ImageNet accuracy, significantly outperforming all existing Mamba variants in vision.", "keywords": [ "Auto regressive Pretraining" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "PQjZes6vFV", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Di Liu", "gender": "Male", "institution": "Rutgers University, New Brunswick", "country": "US", "position": "PhD student" }, { "name": "Khanh Doan", "gender": "Male", "institution": "VinAI Research", "country": "VN", "position": "Intern" }, { "name": "Quan Dao", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Trung Le", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" }, { "name": "Dimitris Metaxas", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 19, "n_ref_uni": 40, "n_ref": 122, "n_ref_all": 149, "n_fig": 8, "n_tab": 3, "L_tab": 2086, "n_element_tab": 198, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 95, "n_element_tab_1": 6, "formula_len_all": 1020, "formula_len_all_1": 985, "len_all": 127236, "len_all_1": 60236, "len_abs": 1408, "len_title": 106, "len_sents": 33497, "len_sents_1": 28698, "n_sents": 258, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1415, "title": "Improved Training Technique for Latent Consistency Models", "abs": "Consistency models are a new family of generative models capable of producing high-quality samples in either a single step or multiple steps. Recently, consistency models have demonstrated impressive performance, achieving results on par with diffusion models in the pixel space. However, the success of scaling consistency training to large-scale datasets, particularly for text-to-image and video generation tasks, is determined by performance in the latent space. In this work, we analyze the statistical differences between pixel and latent spaces, discovering that latent data often contains highly impulsive outliers, which significantly degrade the performance of iCT \\citep{song2023improved} in the latent space. To address this, we replace Pseudo-Huber losses with Cauchy losses, effectively mitigating the impact of outliers. Additionally, we introduce a diffusion loss at early timesteps and employ optimal transport (OT) coupling to further enhance performance. Lastly, we introduce the adaptive scaling-$c$ scheduler to manage the robust training process and adopt Non-scaling LayerNorm in the architecture to better capture the statistics of the features and reduce outlier impact. With these strategies, we successfully train latent consistency models capable of high-quality sampling with one or two steps, significantly narrowing the performance gap between latent consistency and diffusion models.", "keywords": [ "Consistency Model", "Diffusion Model" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "PQ6jzz82Nd", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Aleksandar Pavlovic", "gender": "Male", "institution": "Technische Universit\u00e4t Wien", "country": "AT", "position": "PhD student" }, { "name": "Emanuel Sallinger", "gender": "unknown", "institution": "TU Wien (Vienna University of Technology)", "country": "AT", "position": "Professor" }, { "name": "Steven Schockaert", "gender": "Male", "institution": "Cardiff University", "country": "GB", "position": "Full Professor" } ], "n_formula": 40, "n_formula_1": 24, "n_ref_uni": 29, "n_ref": 76, "n_ref_all": 94, "n_fig": 6, "n_tab": 3, "L_tab": 986, "n_element_tab": 57, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 776, "n_element_tab_1": 61, "formula_len_all": 2897, "formula_len_all_1": 1687, "len_all": 199155, "len_all_1": 74512, "len_abs": 1632, "len_title": 122, "len_sents": 70155, "len_sents_1": 35156, "n_sents": 620, "n_sents_1": 280, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1642, "title": "Differentiable Reasoning about Knowledge Graphs with Reshuffled Embeddings", "abs": "Knowledge graph (KG) embedding methods learn geometric representations of entities and relations to predict plausible missing knowledge. These representations are typically assumed to capture rule-like inference patterns. However, our theoretical understanding of the kinds of inference patterns that can be captured in this way remains limited. Ideally, KG embedding methods should be expressive enough such that for any set of rules, there exists an embedding that exactly captures these rules. This principle has been studied within the framework of region-based embeddings, but existing models are severely limited in the kinds of rule bases that can be captured. We argue that this stems from the use of representations that correspond to the Cartesian product of two-dimensional regions.\nAs an alternative, we propose RESHUFFLE, a simple model based on ordering constraints that can faithfully capture a much larger class of rule bases than existing approaches. Moreover, the embeddings in our framework can be learned by a Graph Neural Network (GNN), which effectively acts as a differentiable rule base. This has some practical advantages, e.g. ensuring that embeddings can be easily updated as new knowledge is added to the KG. At the same time, since the resulting representations can be used similarly to standard KG embeddings, our approach is significantly more efficient than existing approaches to differentiable reasoning. The GNN-based formulation also allows us to study how bounded inference can be captured. We show in particular that bounded reasoning with arbitrary sets of closed path rules can be captured in this way.", "keywords": [ "Differentiable reasoning", "knowledge graphs", "region based embeddings" ], "rating_list": [ 6, 3, 5, 6, 6 ], "soundness_list": [ 3, 2, 4, 3, 3 ], "presentation_list": [ 4, 1, 4, 3, 3 ], "contribution_list": [ 3, 1, 2, 3, 3 ], "confidence_list": [ 3, 5, 4, 2, 5 ] }, { "paper_id": "PPxyXlCAOJ", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Nathaniel Xu", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "D. J. Sutherland", "gender": "Female", "institution": "University of British Columbia", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 71, "n_formula_1": 22, "n_ref_uni": 40, "n_ref": 67, "n_ref_all": 95, "n_fig": 12, "n_tab": 10, "L_tab": 1899, "n_element_tab": 137, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10021, "formula_len_all_1": 2332, "len_all": 175377, "len_all_1": 69606, "len_abs": 1344, "len_title": 97, "len_sents": 52989, "len_sents_1": 31281, "n_sents": 459, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1344, "title": "Learning Representations for Independence Testing", "abs": "Many tools exist that attempt to detect dependence between random variables, a core question across a wide range of machine learning, statistical, and scientific endeavors. Although several statistical tests guarantee eventual detection of any dependence with enough samples, standard tests may require an exorbitant amount of samples for detecting subtle dependencies between high-dimensional random variables with complex distributions. In this work, we study two related ways to learn powerful independence tests. First, we show how to construct powerful statistical tests with finite-sample validity by using variational estimators of mutual information, such as the InfoNCE or NWJ estimators. Second, we establish a close relationship between these variational mutual information-based tests\nand tests based on the Hilbert-Schmidt Independence Criterion (HSIC), showing that learning a variational bound in the former case\nis closely related to learning kernels, typically parameterized by deep networks, in the latter. Finally, we show how to find a representation that maximizes the asymptotic power of an HSIC test, prove that this procedure works, and demonstrate empirically the practical improvement of our tests (with HSIC tests generally outperforming the variational ones) on difficult problems of detecting structured dependence.", "keywords": [ "Kernel methods", "Hypothesis testing", "Independence testing" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "PPrcfHXfuT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Hang Gao", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Yongfeng Zhang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 46, "n_ref": 73, "n_ref_all": 80, "n_fig": 3, "n_tab": 5, "L_tab": 3563, "n_element_tab": 88, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1583, "n_element_tab_1": 101, "formula_len_all": 283, "formula_len_all_1": 169, "len_all": 192606, "len_all_1": 68227, "len_abs": 1788, "len_title": 115, "len_sents": 59767, "len_sents_1": 33372, "n_sents": 436, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1796, "title": "PTR: Precision-Driven Tool Recommendation for Large Language Models", "abs": "By augmenting Large Language Models (LLMs) with external tools, their capacity to solve complex problems has been significantly enhanced. However, despite ongoing advancements in the parsing capabilities of LLMs, incorporating all available tools simultaneously in the prompt remains impractical due to the vast number of external tools. Consequently, it is essential to provide LLMs with a precise set of tools tailored to the specific task, considering both quantity and quality. Current tool retrieval methods primarily focus on refining the ranking list of tools and directly packaging a fixed number of top-ranked tools as the tool set. However, these approaches often fail to equip LLMs with the optimal set of tools prior to execution, since the optimal number of tools for different tasks could be different, resulting in inefficiencies such as redundant or unsuitable tools, which impede immediate access to the most relevant tools. This paper addresses the challenge of recommending precise toolsets for LLMs. We introduce the problem of tool recommendation, define its scope, and propose a novel Precision-driven Tool Recommendation (PTR) approach. PTR captures an initial, concise set of tools by leveraging historical tool bundle usage and dynamically adjusts the tool set by performing tool matching, culminating in a multi-view-based tool addition. Additionally, we present a new dataset, RecTools, and a metric, TRACC, designed to evaluate the effectiveness of tool recommendation for LLMs. We further validate our design choices through comprehensive experiments, demonstrating promising accuracy across two open benchmarks and our RecTools dataset. We release our code and dataset at https://anonymous.4open.science/r/PTR-65DD to support further research in tool recommendation.", "keywords": [ "Tool", "Recommendation", "Large Language Model" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "PPjpGTPG5K", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Bailan He", "gender": "Male", "institution": "University of Munich, Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Shuo Chen", "gender": "Male", "institution": "Amazon", "country": "DE", "position": "Intern" }, { "name": "Volker Tresp", "gender": "Male", "institution": "Ludwig Maximilian University of Munich", "country": "DE", "position": "Associate Professor" }, { "name": "Yilun Liu", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "MS student" }, { "name": "Yunpu Ma", "gender": "Male", "institution": "Siemens Corporate Research", "country": "DE", "position": "Researcher" }, { "name": "Zhen Han", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Zifeng Ding", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 54, "n_ref": 139, "n_ref_all": 169, "n_fig": 10, "n_tab": 5, "L_tab": 11336, "n_element_tab": 1431, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 854, "formula_len_all_1": 682, "len_all": 202846, "len_all_1": 70500, "len_abs": 1062, "len_title": 136, "len_sents": 45280, "len_sents_1": 33312, "n_sents": 310, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1057, "title": "PERFT: Parameter-Efficient Routed Fine-Tuning for Mixture-of-Expert Model", "abs": "The Mixture-of-Experts (MoE) paradigm has emerged as a promising approach for scaling transformer-based large language models (LLMs) with improved resource utilization. \nHowever, efficiently fine-tuning MoE LLMs remains largely underexplored.\nInspired by recent works on Parameter-Efficient Fine-Tuning (PEFT), we present a unified framework for integrating PEFT modules into MoE LLMs.\nOur framework, aligned with the core mechanisms of MoE, encompasses a comprehensive set of design dimensions including various functional and composition strategies.\nBy combining the key design choices within our framework, we introduce Parameter-Efficient Routed Fine-Tuning (PERFT) as a flexible and scalable family of PEFT strategies tailored for MoE LLMs.\nExtensive experiments adapting OLMoE-1B-7B and Mixtral-8\u00d77B for commonsense and arithmetic reasoning tasks demonstrate the effectiveness, scalability, and intriguing dynamics of PERFT. \nAdditionally, we provide empirical findings for each specific design choice to facilitate better application of MoE and PEFT.", "keywords": [ "Mixture of Expert", "Parameter efficient fine-tuning", "Large Language Model" ], "rating_list": [ 8, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 1, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "PPazOk82Sq", "primary_area": "learning theory", "n_author": 7, "author_info_list": [ { "name": "Andi Han", "gender": "Male", "institution": "RIKEN AIP", "country": "JP", "position": "Postdoc" }, { "name": "Denny Wu", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "Difan Zou", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Taiji Suzuki", "gender": "Male", "institution": "The University of Tokyo", "country": "", "position": "Full Professor" }, { "name": "Wei Huang", "gender": "Male", "institution": "RIKEN AIP", "country": "JP", "position": "Research Scientist" }, { "name": "Yilan Chen", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Yujin Song", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "MS student" } ], "n_formula": 214, "n_formula_1": 42, "n_ref_uni": 49, "n_ref": 105, "n_ref_all": 125, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 86, "n_element_tab_1": 6, "formula_len_all": 24090, "formula_len_all_1": 2835, "len_all": 219759, "len_all_1": 69544, "len_abs": 1257, "len_title": 132, "len_sents": 54804, "len_sents_1": 29228, "n_sents": 615, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1260, "title": "Label Noise Gradient Descent Improves Generalization in the Low SNR Regime", "abs": "The capacity of deep learning models is often large enough to both learn the underlying statistical signal and overfit to noise in the training set. This noise memorization can be harmful especially for data with a low signal-to-noise ratio (SNR), leading to poor generalization. Inspired by prior observations that label noise provides implicit regularization that improves generalization, in this work, we investigate whether introducing label noise to the gradient updates can enhance the test performance of neural network (NN) in the low SNR regime. Specifically, we consider the learning of a two-layer NN with a simple label noise gradient descent (GD) algorithm, in an idealized signal-noise data setting. We prove that adding label noise during training suppresses noise memorization, preventing it from dominating the learning process; consequently, label noise GD enjoys rapid signal growth while the overfitting remains controlled, thereby achieving good generalization despite the low SNR. In contrast, we also show that NN trained with standard GD tends to overfit to noise in the same low SNR setting and establish a non-vanishing lower bound on its test error, thus demonstrating the benefit of label noise injection in gradient-based training.", "keywords": [ "Label Noise Gradient Descent", "Feature Learning", "Generalization", "Low Signal-to-noise Ratio" ], "rating_list": [ 3, 6, 6, 8, 5 ], "soundness_list": [ 1, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 1, 3, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 2, 3 ] }, { "paper_id": "PPTE1DL4Li", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Lorenzo Magnino", "gender": "Male", "institution": "New York University Shanghai", "country": "CN", "position": "Researcher" }, { "name": "Mathieu Lauriere", "gender": "unknown", "institution": "New York University", "country": "FR", "position": "Assistant Professor" }, { "name": "Yuchen Zhu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 68, "n_formula_1": 22, "n_ref_uni": 34, "n_ref": 55, "n_ref_all": 86, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6772, "formula_len_all_1": 2048, "len_all": 155620, "len_all_1": 68643, "len_abs": 1407, "len_title": 138, "len_sents": 51186, "len_sents_1": 30127, "n_sents": 502, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1421, "title": "Deep Learning Algorithms for Mean Field Optimal Stopping in Finite Space and Discrete Time", "abs": "Optimal stopping of stochastic processes is a fundamental problem in optimization that has found applications in risk management, finance, economics, and recently in the fields of computer science. We extend the standard framework to a multi-agent setting, named multi-agent optimal stopping (MAOS), where a group of agents cooperatively solves finite-space, discrete-time optimal stopping problems. Solving the finite-agent case is computationally prohibitive when the number of agents is very large, so this work studies the mean field optimal stopping (MFOS) problem, obtained as the number of agents approaches infinity. We prove that MFOS provides a good approximate solution to MAOS. We also prove a dynamic programming principle (DPP), based on the theory of mean field control. We then propose two deep learning methods: one simulates full trajectories to learn optimal decisions, whereas the other leverages DPP \\blue{to compute the value function and to learn the optimal stopping rule} with backward induction; both methods train neural networks for the optimal stopping decisions. We demonstrate the effectiveness of these approaches through numerical experiments on 6 different problems in spatial dimension up to 300. To the best of our knowledge, this is the first work to study MFOS in finite space and discrete time, and to propose efficient and scalable computational methods for this type of problems.", "keywords": [ "Optimal Stopping", "Mean Field", "Deep Learning", "Dynamic Programming" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "PPDheO2z5v", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "DaFeng Chi", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Irwin King", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Shiguang Wu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yuecheng Liu", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Yueen Ma", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yuzheng Zhuang", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Research Engineer" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 51, "n_ref": 96, "n_ref_all": 112, "n_fig": 10, "n_tab": 5, "L_tab": 2405, "n_element_tab": 241, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 868, "n_element_tab_1": 96, "formula_len_all": 179, "formula_len_all_1": 104, "len_all": 204508, "len_all_1": 74538, "len_abs": 1001, "len_title": 141, "len_sents": 48034, "len_sents_1": 34814, "n_sents": 395, "n_sents_1": 255, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1004, "title": "Actra: Optimized Transformer Architecture for Vision-Language-Action Models in Robot Learning", "abs": "Vision-language-action models have gained significant attention for their ability to model trajectories in robot learning. However, most existing models rely on Transformer models with vanilla causal attention, which we find suboptimal for processing segmented multi-modal sequences. Additionally, the autoregressive generation approach falls short in generating multi-dimensional actions. In this paper, we introduce Actra, an optimized Transformer architecture featuring trajectory attention and learnable action queries, designed to efficiently process segmented multi-modal trajectories in language-conditioned robot imitation learning. Furthermore, we propose a contrastive dynamics learning objective to enhance its understanding of environment dynamics and multi-modal alignment, complementing the primary behavior cloning objective. Through extensive experiments on three large-scale robot manipulation benchmarks, Actra exhibits substantial performance improvements over state-of-the-art models.", "keywords": [ "Robotics", "Embodied AI", "Transformer", "Multimodality", "Contrastive Learning", "Dynamics Learning", "Multimodal Alignment" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "POCT74JhAl", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Chi-Hua Wang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Guang Cheng", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "Full Professor" }, { "name": "Lan Tao", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Namjoon Suh", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "Postdoc" }, { "name": "Shirong Xu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Researcher" } ], "n_formula": 91, "n_formula_1": 31, "n_ref_uni": 0, "n_ref": 56, "n_ref_all": 81, "n_fig": 5, "n_tab": 4, "L_tab": 1691, "n_element_tab": 150, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1166, "n_element_tab_1": 70, "formula_len_all": 9802, "formula_len_all_1": 2101, "len_all": 162877, "len_all_1": 61208, "len_abs": 1342, "len_title": 141, "len_sents": 36111, "len_sents_1": 26814, "n_sents": 330, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1348, "title": "Discriminative Estimation of Total Variation Distance: A Fidelity Auditor for Generative Data", "abs": "With the proliferation of generative AI and the increasing volume of generative data (also called as synthetic data), assessing the fidelity of generative data has become a critical concern. In this paper, we propose a discriminative approach to estimate the total variation (TV) distance between two distributions as an effective measure of generative data fidelity. Our method quantitatively characterizes the relation between the Bayes risk in classifying two distributions and their TV distance. Therefore, the estimation of total variation distance reduces to that of the Bayes risk. In particular, this paper establishes theoretical results regarding the convergence rate of the estimation error of TV distance between two Gaussian distributions. We demonstrate that, with a specific choice of hypothesis class in classification, a fast convergence rate in estimating the TV distance can be achieved. Specifically, the estimation accuracy of the TV distance is proven to inherently depend on the separation of two Gaussian distributions: smaller estimation errors are achieved when the two Gaussian distributions are farther apart. This phenomenon is also validated empirically through extensive simulations. In the end, we apply this discriminative estimation method to rank fidelity of synthetic image data using the MNIST/CIFAR-10 dataset.", "keywords": [ "Classification; Total Variation Distance; Learning Theory; Generative Data" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 2 ] }, { "paper_id": "PNiqWDAtPq", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Alessio Tonioni", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Enis Simsar", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Federico Tombari", "gender": "Male", "institution": "Google", "country": "CH", "position": "Research Director" }, { "name": "Thomas Hofmann", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Yongqin Xian", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 37, "n_ref": 103, "n_ref_all": 119, "n_fig": 4, "n_tab": 7, "L_tab": 2913, "n_element_tab": 265, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1392, "n_element_tab_1": 139, "formula_len_all": 598, "formula_len_all_1": 495, "len_all": 171998, "len_all_1": 60971, "len_abs": 1197, "len_title": 0, "len_sents": 49784, "len_sents_1": 28268, "n_sents": 351, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1202, "title": "UIP2P: Unsupervised Instruction-Based Image Editing via Cycle Edit Consistency", "abs": "We propose an unsupervised model for instruction-based image editing that eliminates the need for ground-truth edited images during training. Traditional supervised approaches depend on datasets containing triplets of input image, edited image, and edit instruction, often generated by either existing editing methods or human-annotations, which introduce biases and limit their generalization ability. Our model addresses these challenges by introducing a novel editing mechanism called Cycle Edit Consistency (CEC). We propose to apply a forward and backward edit in one training step and enforce consistency in both the image and attention space. This allows us to bypass the need for ground-truth edited images and unlock training on datasets comprising either real image-caption pairs or image-caption-edit triplets. We empirically show that our unsupervised method achieves better performance across a wider range of edits with high fidelity and precision. By eliminating the need for pre-existing datasets of triplets, reducing biases associated with supervised methods, and introducing CEC, our work represents a significant advancement in unblocking scaling of instruction-based image editing.", "keywords": [ "Unsupervised learning", "Diffusion models", "Cycle edit consistency", "Instruction-based image editing" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "PNMv4r7s1i", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Gang Pan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" }, { "name": "Juntao Dai", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Taiye Chen", "gender": "Male", "institution": "Peking University", "country": "", "position": "Undergrad student" }, { "name": "Yaodong Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Qian Dr. Zheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 86, "n_formula_1": 19, "n_ref_uni": 41, "n_ref": 114, "n_ref_all": 157, "n_fig": 16, "n_tab": 5, "L_tab": 1673, "n_element_tab": 203, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10981, "formula_len_all_1": 1452, "len_all": 220924, "len_all_1": 69002, "len_abs": 1626, "len_title": 129, "len_sents": 64970, "len_sents_1": 32175, "n_sents": 562, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1639, "title": "Mitigating Reward Over-Optimization in RLHF via Behavior-Supported Regularization", "abs": "Reinforcement learning from human feedback (RLHF) is an effective method for aligning large language models (LLMs) with human values. However, reward over-optimization remains an open challenge leading to discrepancies between the performance of LLMs under the reward model and the true human objectives. A primary contributor to reward over-optimization is the extrapolation error that arises when the reward model evaluates out-of-distribution (OOD) responses. However, current methods still fail to prevent the increasing frequency of OOD response generation during the reinforcement learning (RL) process and are not effective at handling extrapolation errors from OOD responses. In this work, we propose the *Behavior-Supported Policy Optimization* (BSPO) method to mitigate the reward over-optimization issue. Specifically, we define *behavior policy* as the next token distribution of the reward training dataset to model the in-distribution (ID) region of the reward model. Building on this, we introduce the behavior-supported Bellman operator to regularize the value function, penalizing all OOD values without impacting the ID ones. Consequently, BSPO reduces the generation of OOD responses during the RL process, thereby avoiding overestimation caused by the reward model\u2019s extrapolation errors. Theoretically, we prove that BSPO guarantees a monotonic improvement of the supported policy until convergence to the optimal behavior-supported policy. Empirical results from extensive experiments show that BSPO outperforms baselines in preventing reward over-optimization due to OOD evaluation and finding the optimal ID policy.", "keywords": [ "Reward Over-Optimization", "Reinforcement Learning from Human Feedback", "Large Language Model" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 5, 3 ] }, { "paper_id": "PNHjoWcQje", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Chuhan Wu", "gender": "Male", "institution": "Noah\u2019s Ark Lab, Huawei", "country": "", "position": "Researcher" }, { "name": "Jingtao Zhan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Min Zhang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Shuai Wang", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Weizhi Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuanqing Yu", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhefan Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhicheng Guo", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhiqiang Guo", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 36, "n_ref": 70, "n_ref_all": 86, "n_fig": 7, "n_tab": 8, "L_tab": 4386, "n_element_tab": 346, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 136, "n_element_tab_1": 23, "formula_len_all": 1084, "formula_len_all_1": 1085, "len_all": 145049, "len_all_1": 51902, "len_abs": 1182, "len_title": 131, "len_sents": 40588, "len_sents_1": 24403, "n_sents": 312, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1188, "title": "StepTool: A Step-grained Reinforcement Learning Framework for Tool Learning in LLMs", "abs": "Despite having powerful reasoning and inference capabilities, Large Language Models (LLMs) still need external tools to acquire real-time information or domain-specific expertise to solve complex tasks, which is referred to as tool learning. Existing tool learning methods primarily rely on tuning with expert trajectories, focusing on token-sequence learning from a linguistic perspective. However, there are several challenges: 1) imitating static trajectories limits their ability to generalize to new tasks. 2) even expert trajectories can be suboptimal, and better solution paths may exist. In this work, we introduce StepTool, a novel step-grained reinforcement learning framework to improve tool learning in LLMs. It consists of two components: Step-grained Reward Shaping, which assigns rewards at each tool interaction based on tool invocation success and its contribution to the task, and Step-grained Optimization, which uses policy gradient methods to optimize the model in a multi-step manner. Experimental results demonstrate that StepTool significantly outperforms existing methods in multi-step, tool-based tasks, providing a robust solution for complex task environments.", "keywords": [ "Tool Learning", "Large Language Models", "Reinforcement Learning" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "PNHGYziAsL", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Chengzhi Mao", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Junfeng Yang", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Associate Professor" }, { "name": "Leon Zhou", "gender": "Male", "institution": "Columbia University", "country": "", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 45, "n_ref": 68, "n_ref_all": 84, "n_fig": 7, "n_tab": 2, "L_tab": 1099, "n_element_tab": 53, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1099, "n_element_tab_1": 53, "formula_len_all": 481, "formula_len_all_1": 481, "len_all": 127486, "len_all_1": 61564, "len_abs": 1319, "len_title": 98, "len_sents": 28464, "len_sents_1": 28417, "n_sents": 241, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 38, "L_abs": 870, "title": "SPIN: Self-Supervised Prompt INjection", "abs": "Large Language Models (LLMs) are increasingly used in a variety of important applications, yet their safety and reliability remain major concerns. Various adversarial and jailbreak attacks have been proposed to bypass the safety alignment and cause the model to produce harmful responses. We introduce Defensive Self-supervised Prompt INjection (D-SPIN) which can detect and reverse these various attacks on LLMs. Just by injecting an adaptive defense prompt at inference-time, our method is simple, effective, and compatible with existing safety-aligned models. Our benchmarks demonstrate that our system can reduce the attack success rate by up to 87.9\\%, while maintaining the performance on benign user requests. In addition, we discuss the situation of an adaptive attacker and show that our method is still resilient against attackers who are aware of our defense.", "keywords": [ "Safety", "Alignment", "Defense", "Adversarial Attack", "Inference", "LLM", "NLP" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "PN4f0hnI0U", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chengjin Yu", "gender": "Male", "institution": "Anhui University", "country": "", "position": "Assistant Professor" }, { "name": "Huafeng Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Mingfeng Jiang", "gender": "Male", "institution": "Zhejiang Sci-Tech University", "country": "", "position": "Full Professor" }, { "name": "Xiaowei Ruan", "gender": "unknown", "institution": "Zhejiang Sci-Tech University", "country": "CN", "position": "MS student" }, { "name": "Zheng Yuan", "gender": "Male", "institution": "Civil Aviation Flight University of China", "country": "", "position": "Lecturer" }, { "name": "zhangxiaolin", "gender": "unknown", "institution": "Zhejiang Sci-Tech University", "country": "CN", "position": "MS student" }, { "name": "ruandongsheng", "gender": "Male", "institution": "Zhejiang Sci-Tech University", "country": "CN", "position": "Lecturer" } ], "n_formula": 23, "n_formula_1": 23, "n_ref_uni": 33, "n_ref": 42, "n_ref_all": 57, "n_fig": 6, "n_tab": 7, "L_tab": 1835, "n_element_tab": 145, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 2692, "n_element_tab_1": 84, "formula_len_all": 1351, "formula_len_all_1": 1349, "len_all": 112832, "len_all_1": 56114, "len_abs": 1252, "len_title": 152, "len_sents": 27620, "len_sents_1": 23146, "n_sents": 218, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1262, "title": "CineMorph: Learning Time-Continuous Motion Field for Motion Tracking on Cine Magnetic Resonance Images", "abs": "Tracking cardiac motion using cine magnetic resonance imaging (cine MRI) is essential for evaluating cardiac function and diagnosing cardiovascular diseases. Current methods for cardiac motion tracking depend on scaling and squaring (SS) integration to learn discrete Lagrangian motion fields. However, this reliance hinders the effective exploitation of temporal continuity, leading to inadequate tracking accuracy. In this paper, we introduce a novel unsupervised learning method, CineMorph, to achieve temporally continuous cardiac motion tracking in cine MRI image sequences. Our approach integrates a frame-aware UNet with a series of time-continuous Transformer blocks to learn temporally continuous intra-frame motion fields, which are then assembled into time-continuous Lagrangian motion fields. To ensure the diffeomorphism property, we implement semigroup regularization to constrain our model, thus eliminating the reliance on SS integration. We evaluate our method on the public Automatic Cardiac Diagnostic Challenge (ACDC) dataset. The experimental results show that our method outperforms the existing state-of-the-art methods and achieves state-of-the-art performance with a mean DICE score of $83.6\\%$ and a mean Hausdorff distance of $3.4$ mm.", "keywords": [ "Motion tracking", "cine MRI", "unsupervised learning", "diffeomorphic" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 1, 3, 1 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 5, 3, 5 ] }, { "paper_id": "PN3i4b6NED", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "ChenRui Duan", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Siyuan Li", "gender": "Male", "institution": "Westlake University & Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zelin Zang", "gender": "Male", "institution": "Westlake University, Zhejiang University, National University of Singapore", "country": "CN", "position": "PhD student" }, { "name": "Stan Z. Li", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Chair Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 75, "n_ref": 124, "n_ref_all": 133, "n_fig": 2, "n_tab": 17, "L_tab": 7127, "n_element_tab": 791, "n_fig_1": 1, "n_tab_1": 13, "L_tab_1": 6939, "n_element_tab_1": 781, "formula_len_all": 235, "formula_len_all_1": 0, "len_all": 184865, "len_all_1": 62758, "len_abs": 1188, "len_title": 77, "len_sents": 31908, "len_sents_1": 21749, "n_sents": 251, "n_sents_1": 152, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1192, "title": "OpenMeta: A Comprehensive Multi-Task Benchmark for Metagenomics Understanding", "abs": "Metagenomics is essential for exploring the vast diversity and intricate interactions of microbes that impact health, agriculture, and environmental sciences. Despite the surge of machine learning-based metagenomic models addressing these questions, evaluating their respective benefits is challenging due to the use of distinct, experimental datasets, partly contrived, and varying model performance across different tasks. To this end, we introduce OpenMeta, the first comprehensive benchmark tailored for metagenomic function prediction, which integrates diverse datasets ranging from 1,000 to 213,000 sequences and incorporates hierarchical data. We highlight the inadequacies of current genomic models and the superior performance of metagenomic pre-trained models for handling complex metagenomic data. Furthermore, we identify a critical research gap: the lack of unified models that process both sequence and hierarchical data. Addressing this could significantly advance metagenomic analyses. OpenMeta sets a new standard for metagenomic analysis, offering insights that could enhance the understanding and application of microbial ecology in biotechnology and environmental science.", "keywords": [ "Metagenomics", "DNA", "Pre-Trained Language Model", "Transformer", "Benchmark" ], "rating_list": [ 3, 5, 5, 1 ], "soundness_list": [ 2, 2, 4, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "PMf2Dg1TAA", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Hemanth Saratchandran", "gender": "unknown", "institution": "University of Adelaide/Australian Institute of Machine Learning", "country": "AU", "position": "Postdoc" }, { "name": "Jianqiao Zheng", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "PhD student" }, { "name": "Simon Lucey", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Full Professor" }, { "name": "Wenbo Zhang", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "PhD student" }, { "name": "Yiping Ji", "gender": "Male", "institution": "CSIRO", "country": "AU", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 10, "n_ref_uni": 25, "n_ref": 41, "n_ref_all": 58, "n_fig": 10, "n_tab": 5, "L_tab": 590, "n_element_tab": 105, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 342, "n_element_tab_1": 56, "formula_len_all": 3712, "formula_len_all_1": 696, "len_all": 109681, "len_all_1": 51191, "len_abs": 772, "len_title": 111, "len_sents": 30823, "len_sents_1": 21478, "n_sents": 259, "n_sents_1": 156, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 773, "title": "Rethinking Softmax: Self-Attention with Polynomial Activations", "abs": "This paper challenges the conventional belief that softmax attention in transformers is effective primarily because it generates a probability distribution for attention allocation. Instead, we theoretically show that its success lies in its ability to implicitly regularize the Frobenius norm of the attention matrix during training. We then explore alternative activations that regularize the Froebnius norm of the attention matrix, demonstrating that certain polynomial activations can achieve this effect, making them suitable for attention-based architectures. Empirical results indicate these activations perform comparably or better than softmax across various computer vision and language tasks, suggesting new possibilities for attention mechanisms beyond softmax.", "keywords": [ "polynomial activations", "softmax", "theory of transformers." ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 1, 3, 1, 3 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 1, 2, 1, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "PLskiLUBDW", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Dan MacKinlay", "gender": "Male", "institution": "Commonwealth Scientific and Industrial Research Organisation, CSIRO", "country": "AU", "position": "Researcher" }, { "name": "Daniel Edward Pagendam", "gender": "Male", "institution": ", CSIRO", "country": "", "position": "Researcher" }, { "name": "Petra Kuhnert", "gender": "Female", "institution": "CSIRO", "country": "AU", "position": "Associate Professor" }, { "name": "Russell Tsuchida", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Lecturer" } ], "n_formula": 124, "n_formula_1": 35, "n_ref_uni": 29, "n_ref": 54, "n_ref_all": 89, "n_fig": 12, "n_tab": 2, "L_tab": 299, "n_element_tab": 102, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 85, "n_element_tab_1": 14, "formula_len_all": 8297, "formula_len_all_1": 1701, "len_all": 291654, "len_all_1": 98757, "len_abs": 1118, "len_title": 148, "len_sents": 74890, "len_sents_1": 29926, "n_sents": 703, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 99, "L_abs": 1274, "title": "Gaussian Ensemble Belief Propagation for Efficient Inference in High-Dimensional, Black-box Systems", "abs": "Efficient inference in high-dimensional models is a central challenge in machine learning.\nWe introduce the Gaussian Ensemble Belief Propagation (GEnBP) algorithm, which combines the strengths of the Ensemble Kalman Filter (EnKF) and Gaussian Belief Propagation (GaBP) to address this challenge.\nGEnBP updates ensembles of prior samples into posterior samples by passing low-rank local messages over the edges of a graphical model, enabling efficient handling of high-dimensional states, parameters, and complex, noisy, black-box generative processes.\nBy utilizing local message passing within a graphical model structure, GEnBP effectively manages complex dependency structures and remains computationally efficient even when the ensemble size is much smaller than the inference dimension --- a common scenario in spatiotemporal modeling, image processing, and physical model inversion.\nWe demonstrate that GEnBP can be applied to various problem structures, including data assimilation, system identification, and hierarchical models, and show through experiments that it outperforms existing belief propagation methods in terms of accuracy and computational efficiency.\n\nSupporting code is available at https://github.com/danmackinlay/GEnBP}{github.com/danmackinlay/GEnBP", "keywords": [ "factor graph", "data assimilation", "ensemble kalman filter", "gaussian belief propagation", "geospatial", "probabilistic graphical model", "hierarchical model" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 4 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "PLgHiJOjcH", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Jin Xie", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Shenlong Wang", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "CA", "position": "Assistant Professor" }, { "name": "Zexin Yin", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "PhD student" }, { "name": "jian Yang", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 22, "n_ref": 60, "n_ref_all": 71, "n_fig": 7, "n_tab": 3, "L_tab": 405, "n_element_tab": 52, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 329, "n_element_tab_1": 23, "formula_len_all": 258, "formula_len_all_1": 258, "len_all": 119851, "len_all_1": 45403, "len_abs": 223, "len_title": 136, "len_sents": 24870, "len_sents_1": 22378, "n_sents": 145, "n_sents_1": 128, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1214, "title": "LISA: UNLEASHING 2D DIFFUSION FOR 3D GENERATION VIA LIGHTWEIGHT IMAGE SPLATS ADAPTATION", "abs": "Despite its potential, 3D generation lags behind 2D generation in quality and utility, primarily due to the vast gap in the scale and diversity of training data\u2014high-quality 2D data is abundant, while high-quality 3D assets remain limited by orders of magnitude. Existing methods use 2D generative priors for 3D asset creation via distillation or generate-and-reconstruct schemes, both of which suffer from quality loss during optimization. In this paper, we propose a novel scheme to exploit 2D diffusion prior for 3d generation by integrating a lightweight adapter into the decoder of a frozen 2D diffusion model, allowing it to generate RGB images, Gaussian splats, and physics-based rendering material maps simultaneously. Once trained, the proposed Lightweight Image Splats Adaptation (LISA) directly produces relightable Gaussian splats in feed-forward manner, which can be converted into high-quality, relightable 3D meshes through an inverse rendering framework. Quantitative and qualitative results demonstrate that our method outperforms state-of-the-art approaches with a significantly lower computational budget for both training and sampling. More results can be found at https://LISA-3dgen.github.io.", "keywords": [ "3D asset creation", "lightweight adapter", "Gaussian Splats" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "PLYqJVV7dm", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Dong Li", "gender": "Male", "institution": "South China University of Technology", "country": "", "position": "MS student" }, { "name": "Yao Yu", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Yingwei Zhang", "gender": "Female", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Yiqiang Chen", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "ke Bu", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "lvdetao", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 36, "n_formula_1": 19, "n_ref_uni": 30, "n_ref": 53, "n_ref_all": 100, "n_fig": 12, "n_tab": 8, "L_tab": 4334, "n_element_tab": 490, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 4897, "n_element_tab_1": 354, "formula_len_all": 1675, "formula_len_all_1": 935, "len_all": 149290, "len_all_1": 75775, "len_abs": 1983, "len_title": 115, "len_sents": 46445, "len_sents_1": 29443, "n_sents": 381, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1501, "title": "CRAFT: Time Series Forecasting with Cross-Future Behavior Awareness", "abs": "Time series forecasting is the crucial infrastructure in the field of e-commerce, providing technical support for consumer behavior analysis, sales trends forecasting, etc. E-commerce allows consumers to reserve in advance. These pre-booking features reflect future sales trends and can increase the certainty of time series forecasting issues. In this paper, we define these features as Cross-Future Behavior, which occurs before the current time but takes effect in the future. To increase the performance of time series forecasting, we leverage these features and propose the CRoss-Future Behavior Awareness based Time Series Forecasting method (CRAFT). The core idea of CRAFT is to utilize the trend of cross-future behavior to mine the trend of time series data to be predicted. Specifically, to settle the sparse and partial flaws of cross-future behavior, CRAFT employs the Koopman Predictor Module to extract the key trend and the Internal Trend Mining Module to supplement the unknown area of the cross-future behavior matrix. Then, we introduce the External Trend Guide Module with a hierarchical structure to acquire more representative trends from higher levels. Finally, we apply the demand-constrained loss to calibrate the distribution deviation of prediction results. We conduct experiments on real-world dataset. Experiments on both offline large-scale dataset and online A/B test demonstrate the effectiveness of CRAFT. Our dataset and code will be released after formal publication.", "keywords": [ "Time Series Forecasting", "Cross-Future Behavior", "Koopman Theory", "Hierarchical Structure" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "PL6e9HkVxk", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Jiawei Chen", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "PhD student" }, { "name": "Xiao Yang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Xinyi Zeng", "gender": "Female", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yu Tian", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yutao Zhu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Postdoc" }, { "name": "Yuying Shang", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 23, "n_ref": 44, "n_ref_all": 69, "n_fig": 8, "n_tab": 4, "L_tab": 1524, "n_element_tab": 413, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1522, "n_element_tab_1": 322, "formula_len_all": 397, "formula_len_all_1": 527, "len_all": 114541, "len_all_1": 60450, "len_abs": 1344, "len_title": 117, "len_sents": 32283, "len_sents_1": 25954, "n_sents": 282, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1337, "title": "ROOT DEFENCE STRATEGIES: ENSURING SAFETY OF LLM AT THE DECODER LEVEL", "abs": "Large language models (LLMs) have demonstrated immense utility across various industries. However, as LLMs advance, the risk of harmful outputs increases due to incorrect or malicious instruction prompts. While current methods effectively address jailbreak risks, they share common limitations: 1) Judging harmful responses from the prefill-level lacks utilization of the model's decoding outputs, leading to relatively lower effectiveness and robustness. 2) Rejecting potentially harmful responses based on a single evaluation can significantly impair the model's helpfulness. This paper examines the LLMs' capability to recognize harmful outputs, revealing and quantifying their proficiency in assessing the danger of previous tokens. Motivated by pilot experiment results, we design a robust defense mechanism at the decoding level. Our novel decoder-oriented, step-by-step defense architecture corrects harmful queries directly rather than rejecting them outright. We introduce speculative decoding to enhance usability and facilitate deployment to boost secure decoding speed. Extensive experiments demonstrate that our approach improves model security without compromising reasoning speed. Notably, our method leverages the model's ability to discern hazardous information, maintaining its helpfulness compared to existing methods.", "keywords": [ "Large Language Models; LLMs safety; LLMs defense; Speculative decoding" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "PKqHT0xZhI", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Junwei Deng", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "PhD student" }, { "name": "Shichang Zhang", "gender": "Male", "institution": "Harvard Business School", "country": "US", "position": "Postdoc" }, { "name": "Jiaqi Ma", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Ting Wei Li", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 41, "n_formula_1": 7, "n_ref_uni": 31, "n_ref": 75, "n_ref_all": 103, "n_fig": 9, "n_tab": 9, "L_tab": 1492, "n_element_tab": 157, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2876, "formula_len_all_1": 531, "len_all": 151611, "len_all_1": 55529, "len_abs": 1746, "len_title": 102, "len_sents": 54551, "len_sents_1": 26222, "n_sents": 408, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1755, "title": "Efficient Ensembles Improve Training Data Attribution", "abs": "Training data attribution (TDA) methods aim to quantify the influence of individual training data points on the model predictions, with broad applications in data-centric AI, such as mislabel detection, data selection, and copyright compensation. However, existing methods in this field, which can be categorized as retraining-based and gradient-based, have struggled with the trade-off between computational efficiency and attribution efficacy. Retraining-based methods can accurately attribute complex non-convex models but are computationally prohibitive, while gradient-based methods are efficient but often fail for non-convex models. Recent research has shown that augmenting gradient-based methods with ensembles of multiple independently trained models can achieve significantly better attribution efficacy. However, this approach remains impractical for very large-scale applications.\n\nIn this work, we discover that expensive, fully independent training is unnecessary for ensembling the gradient-based methods, and we propose two efficient ensemble strategies, DROPOUT ENSEMBLE and LORA ENSEMBLE, alternative to naive independent ensemble. These strategies significantly reduce training time (up to 80%), serving time (up to 60%), and space cost (up to 80%) while maintaining similar attribution efficacy to the naive independent ensemble. Our extensive experimental results demonstrate that the proposed strategies are effective across multiple TDA methods on diverse datasets and models, including generative settings, significantly advancing the Pareto frontier of TDA methods with better computational efficiency and attribution efficacy. We conduct a theoretical analysis that provides insights into the success of our empirical findings.", "keywords": [ "training data attribution", "data-centric AI" ], "rating_list": [ 3, 6, 8, 5, 5 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "PKAZzhcIrP", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Di Huang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Jinlin Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Liefeng Bo", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Principal Researcher" }, { "name": "Xiefan Guo", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" }, { "name": "miaomiao cui", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Senior Algorithm Expert" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 60, "n_ref": 138, "n_ref_all": 167, "n_fig": 16, "n_tab": 5, "L_tab": 2002, "n_element_tab": 115, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1708, "n_element_tab_1": 79, "formula_len_all": 696, "formula_len_all_1": 669, "len_all": 171366, "len_all_1": 54478, "len_abs": 704, "len_title": 113, "len_sents": 36035, "len_sents_1": 24733, "n_sents": 252, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1430, "title": "I4VGen: Image as Free Stepping Stone for Text-to-Video Generation", "abs": "Text-to-video generation has trailed behind text-to-image generation in terms of quality and diversity, primarily due to the inherent complexities of spatio-temporal modeling and the limited availability of video-text datasets. Recent text-to-video diffusion models employ the image as an intermediate step, significantly enhancing overall performance but incurring high training costs. In this paper, we present I4VGen, a novel video diffusion inference pipeline to leverage advanced image techniques to enhance pre-trained text-to-video diffusion models, which requires no additional training. Instead of the vanilla text-to-video inference pipeline, I4VGen consists of two stages: anchor image synthesis and anchor image-augmented text-to-video synthesis. Correspondingly, a simple yet effective generation-selection strategy is employed to achieve visually-realistic and semantically-faithful anchor image, and an innovative noise-invariant video score distillation sampling (NI-VSDS) is developed to animate the image to a dynamic video by distilling motion knowledge from video diffusion models, followed by a video regeneration process to refine the video. Extensive experiments show that the proposed method produces videos with higher visual realism and textual fidelity. Furthermore, I4VGen also supports being seamlessly integrated into existing image-to-video diffusion models, thereby improving overall video quality.", "keywords": [ "Text-to-Video", "Video Diffusion Models", "Video Synthesis" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "PJqP0wyQek", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Hao Jiang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Qihan Huang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Siming Fu", "gender": "Female", "institution": "alibabagroup", "country": "CN", "position": "Researcher" }, { "name": "Xierui Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "robotx", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 44, "n_ref": 96, "n_ref_all": 133, "n_fig": 17, "n_tab": 2, "L_tab": 548, "n_element_tab": 125, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 518, "n_element_tab_1": 78, "formula_len_all": 540, "formula_len_all_1": 910, "len_all": 158797, "len_all_1": 53110, "len_abs": 0, "len_title": 128, "len_sents": 44291, "len_sents_1": 24830, "n_sents": 328, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1378, "title": "MS-Diffusion: Multi-subject Zero-shot Image Personalization with Layout Guidance", "abs": "Recent advancements in text-to-image generation models have dramatically enhanced the generation of photorealistic images from textual prompts, leading to an increased interest in personalized text-to-image applications, particularly in multi-subject scenarios. However, these advances are hindered by two main challenges: firstly, the need to accurately maintain the details of each referenced subject in accordance with the textual descriptions; and secondly, the difficulty in achieving a cohesive representation of multiple subjects in a single image without introducing inconsistencies. To address these concerns, our research introduces the MS-Diffusion framework for layout-guided zero-shot image personalization with multi-subjects. This innovative approach integrates grounding tokens with the feature resampler to maintain detail fidelity among subjects. With the layout guidance, MS-Diffusion further improves the cross-attention to adapt to the multi-subject inputs, ensuring that each subject condition acts on specific areas. The proposed multi-subject cross-attention orchestrates harmonious inter-subject compositions while preserving the control of texts. Comprehensive quantitative and qualitative experiments affirm that this method surpasses existing models in both image and text fidelity, promoting the development of personalized text-to-image generation.", "keywords": [ "Image Personalization", "Multiple Subjects", "Diffusion Models" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 5, 5, 3, 5, 3 ] }, { "paper_id": "PJojB68YBu", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Mehul Motani", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Associate Professor" }, { "name": "Rohan Ghosh", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" } ], "n_formula": 29, "n_formula_1": 23, "n_ref_uni": 12, "n_ref": 22, "n_ref_all": 38, "n_fig": 4, "n_tab": 7, "L_tab": 4169, "n_element_tab": 483, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3005, "n_element_tab_1": 413, "formula_len_all": 2346, "formula_len_all_1": 1605, "len_all": 110163, "len_all_1": 78681, "len_abs": 1988, "len_title": 97, "len_sents": 39224, "len_sents_1": 31501, "n_sents": 306, "n_sents_1": 240, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 2034, "title": "Generative Matching Units for Supervised Learning", "abs": "We propose an alternative computational unit for feedforward supervised learning architectures, called Generative Matching Units (GMUs). To understand GMUs, we start with the standard perceptron unit and view it as an undirected symmetric measure of computation between the weights $W=[w_1,w_2,..w_d]$ and each input datapoint $X=[x_1,x_2,..,x_d]$. Perceptrons forward $W^TX+b$, which is usually followed by an activation function. In contrast, GMUs compute a directed asymmetric measure of computation that estimates the degree of functional dependency $f$ of the input elements $x_i$ of each datapoint to the weights $w_i$ in terms of latent generative variables $\\theta$, i.e, $f(w_i,\\theta) \\rightarrow x_i$. In order to estimate the functional dependency, GMUs measure the minimum error $\\sum (f(w_i,\\theta)-x_i)^2$ incurred in the generation process by optimizing $\\theta$ for each input datapoint. Subsequently, GMUs map the error into a functional dependency measure via an appropriate scalar function, and forward it to the next layer for further computation. In GMUs, the weights $[w_1,w_2,..,w_d]$ can therefore be interpreted as the $\\textit{generative weights}$. We first compare the generalization ability of GMUs and multi-layered-perceptrons (MLPs) via comprehensive synthetic experiments across a range of diverse settings. The most notable finding is that when the input is a sparse linear combination of latent generating variables, GMUs generalize significantly better than MLPs. Subsequently, we evaluate Resnet MLP networks where the first feedforward layer is replaced by GMUs (GMU-MLP) on 30 tabular datasets and find that in most cases, GMU-MLPs generalize better than the MLP baselines. We also compare GMU-MLP to a set of other benchmarks, including TabNet, XGBoost, etc. Lastly, we evaluate GMU-CNNs on three standard vision datasets and find that in all cases they generalize better than the corresponding CNN baselines. We also find that GMU-CNNs are significantly more robust to test-time corruptions.", "keywords": [ "supervised learning", "classification", "robustness" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 2, 2, 4 ] }, { "paper_id": "PJjHILiQHC", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "David Yunis", "gender": "unknown", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "PhD student" }, { "name": "Gal Vardi", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "IL", "position": "Assistant Professor" }, { "name": "Matthew Walter", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Associate Professor" }, { "name": "Michael Maire", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Associate Professor" }, { "name": "Pedro Henrique Pamplona Savarese", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "", "position": "PhD student" }, { "name": "Samuel Wheeler", "gender": "Male", "institution": "Argonne National Laboratory", "country": "", "position": "Researcher" }, { "name": "Kare Livescu", "gender": "Female", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Professor" }, { "name": "Kumar Kshitij Patel", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 6, "n_ref_uni": 84, "n_ref": 272, "n_ref_all": 312, "n_fig": 18, "n_tab": 1, "L_tab": 107, "n_element_tab": 8, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 613, "formula_len_all_1": 287, "len_all": 260234, "len_all_1": 63691, "len_abs": 1094, "len_title": 114, "len_sents": 73714, "len_sents_1": 31526, "n_sents": 547, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1104, "title": "Approaching Deep Learning through the Spectral Dynamics of Weights", "abs": "We propose an empirical approach centered on the spectral dynamics of weights---the behavior of singular values and vectors during optimization---to unify and clarify several phenomena in deep learning. We identify a consistent bias in optimization across various experiments, from small-scale ``grokking'' to large-scale tasks like image classification with ConvNets, image generation with UNets, speech recognition with LSTMs, and language modeling with Transformers. We also demonstrate that weight decay enhances this bias beyond its role as a norm regularizer, even in practical systems. Moreover, we show that these spectral dynamics distinguish memorizing networks from generalizing ones, offering a novel perspective on this longstanding conundrum. Additionally, we leverage spectral dynamics to explore the emergence of well-performing sparse subnetworks (lottery tickets) and the structure of the loss surface through linear mode connectivity. Our findings suggest that spectral dynamics provide a coherent framework to better understand the behavior of neural networks across diverse settings.", "keywords": [ "simplicity bias", "grokking", "lottery tickets", "linear mode connectivity" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "PJNhZoCjLh", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 6, "author_info_list": [ { "name": "Amauri Holanda Souza Jr", "gender": "Male", "institution": "Federal Institute of Cear\u00e1", "country": "BR", "position": "Associate Professor" }, { "name": "Diego Mesquita", "gender": "Male", "institution": "Getulio Vargas Foundation", "country": "", "position": "Assistant Professor" }, { "name": "Omar Rivasplata", "gender": "unknown", "institution": "University of Manchester", "country": "GB", "position": "Associate Professor" }, { "name": "Samuel Kaski", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "Full Professor" }, { "name": "Tiago Silva", "gender": "Male", "institution": "Escola de Matem\u00e1tica Aplicada", "country": "BR", "position": "MS student" }, { "name": "Vikas K Garg", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Assistant Professor" } ], "n_formula": 78, "n_formula_1": 16, "n_ref_uni": 80, "n_ref": 339, "n_ref_all": 406, "n_fig": 18, "n_tab": 6, "L_tab": 1384, "n_element_tab": 43, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 103, "n_element_tab_1": 1, "formula_len_all": 7640, "formula_len_all_1": 1401, "len_all": 401568, "len_all_1": 91830, "len_abs": 1185, "len_title": 100, "len_sents": 112896, "len_sents_1": 37135, "n_sents": 851, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1193, "title": "Generalization and Distributed Learning of GFlowNets", "abs": "Conventional wisdom attributes the success of Generative Flow Networks (GFlowNets) to their ability to exploit the compositional structure of the sample space for learning generalizable flow functions (Bengio et al., 2021). Despite the abundance of empirical evidence, formalizing this belief with verifiable non-vacuous statistical guarantees has remained elusive. We address this issue with the first data-dependent generalization bounds for GFlowNets. We also elucidate the negative impact of the state space size on the generalization performance of these models via Azuma-Hoeffding-type oracle PAC-Bayesian inequalities. We leverage our theoretical insights to design a novel distributed learning algorithm for GFlowNets, which we call *Subgraph Asynchronous Learning* (SAL). In a nutshell, SAL utilizes a divide-and-conquer strategy: multiple GFlowNets are trained in parallel on smaller subnetworks of the flow network, and then aggregated with an additional GFlowNet that allocates appropriate flow to each subnetwork. Our experiments with synthetic and real-world problems demonstrate the benefits of SAL over centralized training in terms of mode coverage and distribution matching.", "keywords": [ "GFlowNets" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "PIpGN5Ko3v", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Haifeng Chen", "gender": "unknown", "institution": "NEC-Labs", "country": "US", "position": "Researcher" }, { "name": "Tianchun Wang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Wei Cheng", "gender": "Male", "institution": "NEC-Labs", "country": "US", "position": "Researcher" }, { "name": "Xiang Zhang", "gender": "unknown", "institution": "Pennsylvania State University", "country": "US", "position": "Associate Professor" }, { "name": "Yuanzhou Chen", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Zhanwen Chen", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "", "position": "PhD student" }, { "name": "Zichuan Liu", "gender": "unknown", "institution": "Microsoft Research", "country": "CN", "position": "Intern" } ], "n_formula": 26, "n_formula_1": 6, "n_ref_uni": 49, "n_ref": 111, "n_ref_all": 139, "n_fig": 4, "n_tab": 12, "L_tab": 9468, "n_element_tab": 1294, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2981, "n_element_tab_1": 348, "formula_len_all": 2958, "formula_len_all_1": 490, "len_all": 237658, "len_all_1": 71680, "len_abs": 1991, "len_title": 110, "len_sents": 69569, "len_sents_1": 30174, "n_sents": 479, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1831, "title": "Humanizing the Machine: Proxy Attacks to Mislead LLM Detectors", "abs": "The advent of large language models (LLMs) has revolutionized the field of text generation, producing outputs that closely mimic human-like writing. Although academic and industrial institutions have developed detectors to prevent the malicious usage of LLM-generated texts, other research has doubt about the robustness of these systems. To stress test these detectors, we introduce a proxy-attack strategy that effortlessly compromises LLMs, causing them to produce outputs that align with human-written text and mislead detection systems. Our method attacks the source model by leveraging a reinforcement learning (RL) fine-tuned humanized small language model (SLM) in the decoding phase. Through an in-depth analysis, we demonstrate that our attack strategy is capable of generating responses that are indistinguishable to detectors, preventing them from differentiating between machine-generated and human-written text. We conduct systematic evaluations on extensive datasets using proxy-attacked open-source models, including Llama2-13B, Llama3-70B, and Mixtral-8*7B in both white- and black-box settings. Our findings show that the proxy-attack strategy effectively deceives the leading detectors, resulting in an average AUROC drop of 70.4% across multiple datasets, with a maximum drop of 90.3% on a single dataset. Furthermore, in cross-discipline scenarios, our strategy also bypasses these detectors, leading to a significant relative decrease of up to 90.9%, while in cross-language scenario, the drop reaches 91.3%. Despite our proxy-attack strategy successfully bypassing the detectors with such significant relative drops, we find that the generation quality of the attacked models remains preserved, even within a modest utility budget, when compared to the text produced by the original, unattacked source model.", "keywords": [ "machine-generted text detection; evade detection; fine-tuning" ], "rating_list": [ 6, 5, 8, 5, 6 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 4, 3, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "PITFO1ddeh", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Haoyu Wang", "gender": "Male", "institution": "State University of New York at Albany", "country": "US", "position": "Assistant Professor" }, { "name": "Hui Liu", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Jun Huan", "gender": "Male", "institution": "AWS", "country": "US", "position": "Principal Researcher" }, { "name": "Monica Xiao Cheng", "gender": "Female", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Qingyu Yin", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Ruirui Li", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Tianci Liu", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Tianqi Zheng", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Xianfeng Tang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yunzhe Qi", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Jing Gao", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "Associate Professor" } ], "n_formula": 26, "n_formula_1": 14, "n_ref_uni": 46, "n_ref": 116, "n_ref_all": 132, "n_fig": 2, "n_tab": 4, "L_tab": 4462, "n_element_tab": 354, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2135, "n_element_tab_1": 232, "formula_len_all": 2222, "formula_len_all_1": 847, "len_all": 169459, "len_all_1": 67070, "len_abs": 1684, "len_title": 156, "len_sents": 44718, "len_sents_1": 29728, "n_sents": 400, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 106, "L_abs": 1705, "title": "Unlocking Efficient, Scalable, and Continual Knowledge Editing with Basis-Level Representation Fine-Tuning", "abs": "Large language models (LLMs) have achieved remarkable performance on vari-\nous natural language tasks. However, they are trained on static corpora and their\nknowledge can become outdated quickly in the fast-changing world. This moti-\nvates the development of knowledge editing methods designed to update certain\nknowledge in LLMs without changing unrelated others. To make selective edits,\nprevious efforts often sought to update a small amount of parameters in some spe-\ncific layer(s) of a LLM. Nonetheless, in challenging scenarios, they still fall short\nin making successful edits while preserving knowledge irrelevant to the updates\nsimultaneously, resulting in a notable editing-locality trade-off. In this work, we\nquestion if the trade-offs are caused by the fact that parameter-based updates have\na global effect, i.e., edited parameters affect all inputs indiscriminately. In light of\nthis, we explore the feasibility of representation fine-tuning, which applied some\nlinear update to a few representations in a learned subspace, for knowledge edit-\ning. While being effective to enhance an LLM\u2019s general ability as demonstrated in\nthe previous work, we theoretically show that this linear update imposes a tension\nin editing-locality trade-off. Subsequently, BaFT is proposed to break the linear-\nity. BaFT computes a weight for each basis that spans a dimension of the subspace\nbased on the input representation. This input-dependent weighting mechanism al-\nlows BaFT to manage different types of knowledge in an adaptive way, thereby\nachieving a better editing-locality trade-off. Experiments on three LLMs with five\nediting benchmarks in diverse scenarios show the superiority of our method.", "keywords": [ "Knowledge Editing", "Representation Fine-tuning", "Large Language Models" ], "rating_list": [ 5, 8, 3, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "PIHPmNNp7w", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Fabian Paischer", "gender": "Male", "institution": ", Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Postdoc" }, { "name": "Markus Hofmarcher", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "", "position": "PhD student" }, { "name": "Razvan Pascanu", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Affiliate Member" }, { "name": "Sepp Hochreiter", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "Full Professor" }, { "name": "Thomas Schmied", "gender": "Not Specified", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "PhD student" }, { "name": "Vihang Prakash Patil", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 98, "n_ref": 170, "n_ref_all": 240, "n_fig": 37, "n_tab": 5, "L_tab": 4581, "n_element_tab": 485, "n_fig_1": 24, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 707, "formula_len_all_1": 382, "len_all": 334356, "len_all_1": 64166, "len_abs": 1253, "len_title": 123, "len_sents": 88864, "len_sents_1": 29669, "n_sents": 798, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1262, "title": "Retrieval-Augmented Decision Transformer: External Memory for In-context RL", "abs": "In-context learning (ICL) is the ability of a model to learn a new task by observing a few exemplars in its context. While prevalent in NLP, this capability has recently also been observed in Reinforcement Learning (RL) settings. Prior in-context RL methods, however, require entire episodes in the agent's context. Given that complex environments typically lead to long episodes with sparse rewards, these methods are constrained to simple environments with short episodes. To address these challenges, we introduce Retrieval-Augmented Decision Transformer (RA-DT). RA-DT employs an external memory mechanism to store past experiences from which it retrieves only sub-trajectories relevant for the current situation. The retrieval component in RA-DT does not require training and can be entirely domain-agnostic. We evaluate the capabilities of RA-DT on grid-world environments, robotics simulations, and procedurally-generated video games. On grid worlds, RA-DT outperforms baselines, while using only a fraction of their context length. Furthermore, we illuminate the limitations of current in-context RL methods on complex environments and discuss future directions. To facilitate future research, we release datasets for four of the considered environments.", "keywords": [ "Reinforcement Learning", "Transformer", "Multi-task Learning", "In-context Learning", "Retrieval-augmentation" ], "rating_list": [ 1, 5, 8 ], "soundness_list": [ 3, 1, 3 ], "presentation_list": [ 2, 2, 4 ], "contribution_list": [ 1, 1, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "PHrqpxUczG", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 7, "author_info_list": [ { "name": "Bingsheng He", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "CHEN Han", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Luo Pingyi", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Mian Lu", "gender": "Male", "institution": "4Paradigm", "country": "SG", "position": "Researcher" }, { "name": "Yuqiang Chen", "gender": "Male", "institution": "4Paradigm Inc.", "country": "", "position": "Researcher" }, { "name": "Zicong Jiang", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "MS student" }, { "name": "Zining Zhang", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 26, "n_ref": 58, "n_ref_all": 77, "n_fig": 8, "n_tab": 6, "L_tab": 4672, "n_element_tab": 563, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1352, "n_element_tab_1": 66, "formula_len_all": 277, "formula_len_all_1": 190, "len_all": 108492, "len_all_1": 51315, "len_abs": 1117, "len_title": 141, "len_sents": 27365, "len_sents_1": 23366, "n_sents": 199, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1136, "title": "LogQuant: Log-Distributed 2-Bit Quantization of KV Cache with Superior Accuracy Preservation", "abs": "We introduce LogQuant, a groundbreaking 2-bit quantization technique for KV Cache in large language model (LLM) inference, delivering substantial memory savings while preserving superior performance. Previous methods either assume that later tokens are more important or attempt to predict important tokens based on earlier attention patterns. Both approaches, however, can result in performance bottlenecks or frequent mispredictions.\n\nLogQuant takes a different approach. By applying a log-based filtering mechanism, it selectively compresses the KV Cache across the entire context, achieving better performance with the same or even reduced memory footprint compared to existing methods. In benchmark tests, it enhances throughput by 25\\% and boosts batch size by 60\\% without increasing memory consumption. For challenging tasks such as Math and Code Completion, LogQuant improves accuracy by 40\\% to 200\\% at the same compression ratio, outperforming comparable techniques. LogQuant integrates effortlessly with popular inference frameworks like Python\u2019s \\texttt{transformers} library and will be made open-source upon publication.", "keywords": [ "Large Language Models", "KV Cache", "Memory Efficiency", "Quantization" ], "rating_list": [ 6, 3, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 4, 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 2, 1 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "PHkUNcno9n", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Hao Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Postdoc" }, { "name": "Peng Bo", "gender": "unknown", "institution": "Department of Mechanical Engineering, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Po-Yen Tung", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "Stefan Bauer", "gender": "unknown", "institution": "Helmholtz Munich ", "country": "DE", "position": "Director" }, { "name": "Wenjie Du", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Assistant Professor" }, { "name": "Yangtao Chen", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Ye Wei", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 5, "n_ref_uni": 57, "n_ref": 77, "n_ref_all": 107, "n_fig": 17, "n_tab": 2, "L_tab": 2000, "n_element_tab": 190, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 2000, "n_element_tab_1": 190, "formula_len_all": 502, "formula_len_all_1": 177, "len_all": 188181, "len_all_1": 60411, "len_abs": 1712, "len_title": 123, "len_sents": 49310, "len_sents_1": 28733, "n_sents": 328, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1716, "title": "BALSA: Benchmarking Active Learning Strategies for Autonomous laboratories", "abs": "Accelerating scientific discoveries holds significant potential to address some of the most pressing challenges facing society, from mitigating climate change to combating public health crises, such as the growing antibiotics resistance. The vast and complex nature of design parameter spaces makes identifying promising candidates both time-consuming and resource-intensive, rendering conventional exhaustive searches impractical. However, recent advancements in data-driven methods, particularly within the framework of \"active learning,\" have led to more efficient strategies for scientific discovery. By iteratively identifying and labeling the most informative data points, these methods function in a closed loop, guiding experiments or simulations to accelerate the identification of optimal candidates while reducing the demand for data labeling. Despite these advancements, the lack of standardized benchmarks in this emerging field of autonomous scientific discovery impedes progress and limits its potential translational impact. To address this, we introduce BALSA: a comprehensive benchmark specifically designed for evaluating various search algorithms applied in autonomous laboratories within the active learning framework. BALSA offers a standardized evaluation protocol, provides a metric to characterize high-dimensional objective functions, and includes reference implementations of recent methodologies, with a focus on minimizing the data required to reach optimal results. It provides not only a suite of synthetic functions or controlled simulators but also real-world active learning tasks in biology and materials science \u2014 each presenting unique challenges for autonomous laboratory tasks.", "keywords": [ "active learning", "experimental design", "AI for science" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 5, 3, 3 ] }, { "paper_id": "PHg4rAXFVH", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 4, "author_info_list": [ { "name": "Caiwen Ding", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "Associate Professor" }, { "name": "Hongwu Peng", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Researcher" }, { "name": "Xi Xie", "gender": "Male", "institution": "University of Connecticut", "country": "US", "position": "PhD student" }, { "name": "Yuebo Luo", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 22, "n_ref": 31, "n_ref_all": 48, "n_fig": 6, "n_tab": 6, "L_tab": 3018, "n_element_tab": 317, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3681, "n_element_tab_1": 751, "formula_len_all": 841, "formula_len_all_1": 618, "len_all": 95551, "len_all_1": 69591, "len_abs": 1097, "len_title": 131, "len_sents": 29017, "len_sents_1": 24548, "n_sents": 212, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1051, "title": "RTop-K: Ultra-Fast Row-Wise Top-K Selection for Neural Network Acceleration on GPUs", "abs": "Top-k selection algorithms are fundamental in a wide range of applications, from high-performance computing and information retrieval to big data processing and neural network model training. In this paper, we present RTop-K, a highly efficient parallel row-wise top-k selection algorithm specifically designed for GPUs. RTop-K leverages a binary search-based approach to optimize row-wise top-k selection, providing a scalable and accelerated solution. We conduct a detailed analysis of early stopping in our algorithm, showing that it effectively maintains the testing accuracy of neural network models while substantially improving performance. Our GPU implementation of RTop-K demonstrates superior performance over state-of-the-art row-wise top-k GPU implementations, achieving speed-ups ranging from 4.25\u00d7 to 9.51\u00d7 with early stopping, and 3.94\u00d7 without early stopping. Moreover, RTop-K is capable of accelerating the overall training workflow of MaxK-GNNs, delivering an average speed-up of 9.76% to 31.53% across different models and datasets.", "keywords": [ "row-wise topk selection", "GPU", "CUDA" ], "rating_list": [ 5, 6, 10, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 1, 3, 4, 2 ], "confidence_list": [ 3, 2, 4, 2 ] }, { "paper_id": "PHXLbaq822", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Jiangshan Zhang", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "MS student" }, { "name": "Xinrun Wang", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Yewen Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Yunkun Xu", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "shengjun cheng", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Lv Zhiyi", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" } ], "n_formula": 22, "n_formula_1": 22, "n_ref_uni": 47, "n_ref": 87, "n_ref_all": 111, "n_fig": 6, "n_tab": 8, "L_tab": 1746, "n_element_tab": 252, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 898, "n_element_tab_1": 156, "formula_len_all": 1657, "formula_len_all_1": 1577, "len_all": 152466, "len_all_1": 64732, "len_abs": 1288, "len_title": 66, "len_sents": 37194, "len_sents_1": 28240, "n_sents": 271, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1465, "title": "Efficient LLM Alignment via Hierarchical Coarse-to-Fine Refinement", "abs": "Alignment of Large Language Models (LLMs) intends to make LLMs behave to satisfy human preferences and values. Widely used methods, $\\textbf{e.g.}$, Reinforcement Learning from Human Feedback (RLHF), usually involve the additional training of LLMs with a reward model or the dataset reflecting human preferences. However, these training-based methods cannot quickly adapt to different preferences. Recent methods leverage search during the decoding process to align LLMs with preferences. However, these methods ignore the influence of prompts on the decoding distribution, thus hindering the performance. In this work, we propose $ \\textbf{\\textbf{HCFR}}$, a $\\textbf{H}$ierarchical $\\textbf{C}$oarse-to-$\\textbf{F}$ine $\\text{R}$efinement for efficient LLM alignment. Specifically, $\\textbf{\\textbf{HCFR}}$ includes a two-stage refinement: i) $\\textbf{coarse refinement}$ which rephrases the prompts from users through self-refinement, and ii) $\\textbf{fine refinement}$ which leverages the search methods, $\\textit{e.g.}$, Monte Carlo Tree Search (MCTS), for the responses with the guidance of a pre-trained reward model. \nExperimental results on HH-RLHF and UltraChat demonstrate that $\\textbf{\\textbf{HCFR}}$ can significantly outperform existing methods, $\\textit{e.g.}$, ARGS, CARDS, and Rejection sampling, in terms of performance and efficiency, $\\textit{i.e.}$, achieving a 71.3\\% win-tie rate in GPT-4 evaluations while reducing time consumption by 42\\%.", "keywords": [ "Alignment", "Large language models" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 2, 4 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "PHESUVacAw", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Amirhossein Vahidi", "gender": "Male", "institution": "Wellcome Sanger Institute", "country": "GB", "position": "Researcher" }, { "name": "Emily Stephenson", "gender": "unknown", "institution": "University of Newcastle-upon-Tyne", "country": "GB", "position": "Postdoc" }, { "name": "Gosia Trynka", "gender": "Female", "institution": "Wellcome Sanger Institute", "country": "", "position": "Associate Professor" }, { "name": "Hesam Asadollahzadeh", "gender": "Male", "institution": "Wellcome Sanger Institute", "country": "GB", "position": "Researcher" }, { "name": "Kevin Ly", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Marie Moullet", "gender": "Female", "institution": "Sanger Institute ", "country": "GB", "position": "PhD student" }, { "name": "Mohammad Lotfollahi", "gender": "Male", "institution": "Welcome Sanger ", "country": "", "position": "Researcher" }, { "name": "Muzlifah Haniffa", "gender": "Female", "institution": "Wellcome Sanger Institute", "country": "", "position": "Full Professor" }, { "name": "Vijaya Baskar MS", "gender": "unknown", "institution": "Wellcome Sanger Institute", "country": "GB", "position": "Principal Researcher" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 44, "n_ref": 83, "n_ref_all": 98, "n_fig": 4, "n_tab": 3, "L_tab": 443, "n_element_tab": 32, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1248, "n_element_tab_1": 54, "formula_len_all": 359, "formula_len_all_1": 203, "len_all": 136872, "len_all_1": 63553, "len_abs": 1781, "len_title": 124, "len_sents": 33888, "len_sents_1": 29066, "n_sents": 286, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1799, "title": "Sequence-to-sequence modeling for Temporal Reconstruction of Cellular Events", "abs": "Single-cell omics technologies capture molecular snapshots of cells, while most\nbiological processes unfold over time. Accurately predicting single-cell gene ex-\npression at unmeasured time points enhances our understanding of these processes,\nreducing costs and experimental effort by enabling the interpolation and extrap-\nolation of observed data. This helps study continuous development, response to\nperturbations, and disease progression. To address this problem, we propose an\nencoder-decoder transformer architecture for Temporal Reconstruction of Cellular\nEvents (TRACE). TRACE models gene expression generation as a sequence-to-\nsequence generation task by learning to transform a sequence of genes from a\nsource condition (e.g., previous time) into a sequence of genes in a target condition\n(e.g., next time point). TRACE decoder learns to generate gene tokens of the target\ncondition by iteratively unmaking tokens in the target sequence, overcoming the dis-\ncordance between autoregressive modeling and the non-sequential nature of gene\nexpression data. We evaluate TRACE both quantitatively and qualitatively on three\ndatasets, covering a range of tasks and biological scenarios. TRACE outperforms\nexisting models in generalizing across in-distribution and out-of-distribution tasks\nfor temporal prediction. Furthermore, we demonstrate the biological relevance of\nthe cell embeddings learned by TRACE by delineating activation-dependent cell\nstages in immune cells, measured across multiple time points. Our findings suggest\nthat TRACE can enhance in silico hypothesis generation, improving our under-\nstanding and prediction of cellular changes over time. This ultimately facilitates\ndisease understanding and supports the design of cost-effective experiments for\nbiological discovery.", "keywords": [ "self-supervised learning", "single cell RNAseq", "generative modeling", "temporal sequence modeling" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 1, 3, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "PH7ja3T0vN", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Fahim Tajwar", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Jeff Schneider", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Researcher" }, { "name": "Ruslan Salakhutdinov", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Xintong Duan", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Yutong He", "gender": "Female", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Wenze Chen", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 65, "n_ref": 116, "n_ref_all": 144, "n_fig": 12, "n_tab": 14, "L_tab": 3612, "n_element_tab": 296, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 395, "n_element_tab_1": 18, "formula_len_all": 314, "formula_len_all_1": 300, "len_all": 186809, "len_all_1": 68190, "len_abs": 1216, "len_title": 135, "len_sents": 50944, "len_sents_1": 34095, "n_sents": 341, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1231, "title": "State Combinatorial Generalization In Decision Making With Conditional Diffusion Models", "abs": "Many real-world decision-making problems are combinatorial in nature, where states (e.g., surrounding traffic of a self-driving car) can be seen as a combination of basic elements (e.g., pedestrians, trees, and other cars). Due to combinatorial complexity, observing all combinations of basic elements in the training set is infeasible, which leads to an essential yet understudied problem of $\\textit{zero-shot generalization to states that are unseen combinations of previously seen elements.}$ In this work, we first formalize this problem and then demonstrate how existing value-based reinforcement learning (RL) algorithms struggle due to unreliable value predictions in unseen states. We argue that this problem cannot be addressed with exploration alone, but requires more expressive and generalizable models. We demonstrate that behavior cloning with a conditioned diffusion model trained on expert trajectory generalizes better to states formed by new combinations of seen elements than traditional RL methods. Through experiments in maze, driving, and multiagent environments, we show that conditioned diffusion models outperform traditional RL techniques and highlight the broad applicability of our problem formulation.", "keywords": [ "RL generalization", "decision making", "combinatorial generalization", "diffusion model" ], "rating_list": [ 6, 6, 3, 3 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "PH09buDIBT", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Ryuichiro Hataya", "gender": "Unspecified", "institution": "RIKEN", "country": "JP", "position": "Postdoc" }, { "name": "Yoshinobu Kawahara", "gender": "Male", "institution": "Osaka University", "country": "JP", "position": "Full Professor" } ], "n_formula": 26, "n_formula_1": 26, "n_ref_uni": 44, "n_ref": 81, "n_ref_all": 107, "n_fig": 2, "n_tab": 5, "L_tab": 775, "n_element_tab": 122, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1239, "n_element_tab_1": 385, "formula_len_all": 1591, "formula_len_all_1": 1571, "len_all": 123528, "len_all_1": 70728, "len_abs": 1174, "len_title": 102, "len_sents": 29148, "len_sents_1": 26243, "n_sents": 234, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1180, "title": "Glocal Hypergradient Estimation with Koopman Operator", "abs": "Gradient-based hyperparameter optimization methods update hyperparameters using hypergradients, gradients of a meta criterion with respect to hyperparameters. Previous research used two distinct update strategies: optimizing hyperparameters using global hypergradients obtained after completing model training or local hypergradients derived after every few model updates. While global hypergradients offer reliability, their computational cost is significant; conversely, local hypergradients provide speed but are often suboptimal. In this paper, we propose *glocal* hypergradient estimation, blending \"global\" quality with \"local\" efficiency. To this end, we use the Koopman operator theory to linearize the dynamics of hypergradients so that the global hypergradients can be efficiently approximated only by using a trajectory of local hypergradients. Consequently, we can optimize hyperparameters greedily using estimated global hypergradients, achieving both reliability and efficiency simultaneously. Through numerical experiments of hyperparameter optimization, including optimization of optimizers, we demonstrate the effectiveness of the glocal hypergradient estimation.", "keywords": [ "gradient-based hyperparameter optimization", "koopman operator theory", "dynamical systems" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "PGhiPGBf47", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Liwei Jiang", "gender": "Female", "institution": "University of Washington", "country": "", "position": "PhD student" }, { "name": "Yejin Choi", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Yu Ying Chiu", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 44, "n_ref_all": 66, "n_fig": 6, "n_tab": 7, "L_tab": 9057, "n_element_tab": 799, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 229150, "len_all_1": 68414, "len_abs": 1649, "len_title": 128, "len_sents": 84018, "len_sents_1": 33211, "n_sents": 558, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 80, "L_abs": 1658, "title": "DailyDilemmas: Revealing Value Preferences of LLMs with Quandaries of Daily Life", "abs": "As we increasingly seek guidance from LLMs for decision-making in daily life, many of these decisions are not clear-cut and depend significantly on the personal values and ethical standards of the users. We present DailyDilemmas, a dataset of 1,360 moral dilemmas encountered in everyday life. Each dilemma includes two possible actions and with each action, the affected parties and human values invoked. Based on these dilemmas, we consolidated a set of human values across everyday topics e.g., interpersonal relationships, workplace, and environmental issues. We evaluated LLMs on these dilemmas to determine what action they will take and the values represented by these actions. Then, we analyzed these values through the lens of five popular theories inspired by sociology, psychology and philosophy. These theories are: World Value Survey, Moral Foundation Theory, Maslow's Hierarchy of Needs, Aristotle's Virtues, and Plutchik Wheel of Emotion. We find that LLMs are most aligned with the self-expression over survival values in terms of World Value Survey, care over loyalty in Moral Foundation Theory. Interestingly, we find large preferences differences in models for some core values such as truthfulness e.g., Mixtral-8x7B model tends to neglect it by 9.7% while GPT-4-turbo model tends to select it by 9.4%. We also study the recent guidance released by OpenAI (ModelSpec), and Anthropic (Constitutional AI) to understand how their released principles reflect their actual value prioritization when facing nuanced moral reasoning in daily-life settings. We find that end users cannot effectively steer such prioritization using system prompts.", "keywords": [ "language model", "moral dilemma", "model alignment", "machine ethics" ], "rating_list": [ 8, 5, 8, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "PGNdDfsI6C", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Antonio Torralba", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Jacob Andreas", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Pratyusha Sharma", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Reece S Shuttleworth", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 82, "n_ref_all": 135, "n_fig": 24, "n_tab": 2, "L_tab": 894, "n_element_tab": 88, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 435, "n_element_tab_1": 5, "formula_len_all": 678, "formula_len_all_1": 293, "len_all": 154606, "len_all_1": 65258, "len_abs": 2295, "len_title": 100, "len_sents": 44314, "len_sents_1": 31943, "n_sents": 343, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1803, "title": "LoRA vs Full Fine-tuning: An Illusion of Equivalence", "abs": "Fine-tuning is a crucial paradigm for adapting pre-trained large language models to downstream tasks. Recently, methods like Low-Rank Adaptation (LoRA) have been shown to match the performance of fully fine-tuned models on various tasks with an extreme reduction in the number of trainable parameters. Even in settings where both methods learn similarly accurate models, \\emph{are their learned solutions really equivalent?} \nTo answer this, we study how different fine-tuning methods change pre-trained models by analyzing the model's weight matrices through the lens of their spectral properties. We find that full fine-tuning and LoRA yield weight matrices whose singular value decompositions exhibit very different structure; moreover, the fine-tuned models themselves show distinct generalization behaviors when tested outside the adaptation task's distribution. We first show that the weight matrices trained with LoRA have new, high-ranking singular vectors, which we call \\emph{intruder dimensions}. Intruder dimensions do not appear during full fine-tuning. Second, we find that LoRA models with intruder dimensions, despite achieving similar performance to full fine-tuning on the target task, become worse models of the pre-training distribution and adapt less robustly to multiple tasks sequentially. Higher-rank, rank-stabilized LoRA models closely mirror full fine-tuning, even when performing on par with lower-rank LoRA models on the same tasks. These results suggest that models updated with LoRA and full fine-tuning inherently access different parts of the solution space, even when they perform equally on the fine-tuned distribution.\nWe conclude by examining why intruder dimensions appear in LoRA fine-tuned models, why they are undesirable, and how their effects can be minimized.", "keywords": [ "LoRA", "Fine-tuning", "Large Language models", "Transformers", "Low-rank approximation" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "PFRWGeUhJx", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Chenyi Zhang", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Tongyang Li", "gender": "Male", "institution": "Peking University", "country": "", "position": "Assistant Professor" } ], "n_formula": 35, "n_formula_1": 33, "n_ref_uni": 48, "n_ref": 68, "n_ref_all": 79, "n_fig": 1, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4262, "formula_len_all_1": 3747, "len_all": 111678, "len_all_1": 60606, "len_abs": 1154, "len_title": 108, "len_sents": 23745, "len_sents_1": 22056, "n_sents": 197, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1285, "title": "Comparisons Are All You Need for Optimizing Smooth Functions", "abs": "When optimizing machine learning models, there are various scenarios where gradient computations are challenging or even infeasible. Furthermore, in reinforcement learning (RL), preference-based RL that only compares between options has wide applications, including reinforcement learning with human feedback in large language models. In this paper, we systematically study optimization of a smooth function $f\\colon\\mathbb{R}^n\\to\\mathbb{R}$ only assuming an oracle that compares function values at two points and tells which is larger. When $f$ is convex, we give two algorithms using $\\tilde{O}(n/\\epsilon)$ and $\\tilde{O}(n^{2})$ comparison queries to find an $\\epsilon$-optimal solution, respectively. When $f$ is nonconvex, our algorithm uses $\\tilde{O}(n/\\epsilon^2)$ comparison queries to find an $\\epsilon$-approximate stationary point. All these results match the best-known zeroth-order algorithms with function evaluation queries in $n$ dependence, thus suggesting that \\emph{comparisons are all you need for optimizing smooth functions using derivative-free methods}. In addition, we also give an algorithm for escaping saddle points and reaching an $\\epsilon$-second order stationary point of a nonconvex $f$, using $\\tilde{O}(n^{1.5}/\\epsilon^{2.5})$ comparison queries.", "keywords": [ "comparison-based optimization", "sub-zeroth order method", "gradient estimation", "convex optimization", "escape from saddle points" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "PDu1zouM1U", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Ana Me\u0161trovi\u0107", "gender": "Female", "institution": "University of Rijeka, Faculty of Informatics and Digital Technologies", "country": "HR", "position": "Full Professor" }, { "name": "Karlo Babi\u0107", "gender": "unknown", "institution": "University of Rijeka", "country": "HR", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 15, "n_ref": 27, "n_ref_all": 44, "n_fig": 5, "n_tab": 7, "L_tab": 3849, "n_element_tab": 578, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 414, "n_element_tab_1": 64, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 78696, "len_all_1": 61276, "len_abs": 1296, "len_title": 128, "len_sents": 25426, "len_sents_1": 24380, "n_sents": 190, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1305, "title": "Pyramidal Recursive Composition of Multi-Word Units into Unified Representations", "abs": "In this paper, we explore the composition of word embeddings to create richer, more meaningful representations of multi-word units. Existing methods, such as averaging word embeddings, provide simple and efficient approaches. However, they often fail to capture the complexity of multi-word interactions. To address this, we employ the Pyramidal Recursive learning (PyRv) method, which recursively combines word embeddings into unified representations. Originally developed for constructing representations hierarchically from subwords to phrases, PyRv is well-suited for progressively merging individual word vectors into phrase vectors.\nWe evaluate the effectiveness of PyRv for embedding composition using fastText embeddings on the dependency relation labeling task. Using a single fastText word embedding yields an accuracy of 71\\%. Averaging five fastText word embeddings (the middle word and its four neighboring words) results in a significant drop in accuracy to 34\\%. In contrast, by composing five word embeddings with PyRv, we achieve an accuracy of 77\\%, demonstrating the superior ability of PyRv to integrate multiple word embeddings into more expressive representations. These findings highlight the potential of PyRv as a lightweight yet powerful technique for word embedding composition.", "keywords": [ "Composition", "Croatian", "embedding", "pyramidal", "recursive neural network", "text representation learning" ], "rating_list": [ 3, 1, 1 ], "soundness_list": [ 3, 1, 1 ], "presentation_list": [ 2, 3, 1 ], "contribution_list": [ 1, 1, 1 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "PDtMrogheZ", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Bin Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Guoping Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Jie Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Qijie Peng", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Ruibo Guo", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Yang Rui", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 35, "n_formula_1": 17, "n_ref_uni": 71, "n_ref": 135, "n_ref_all": 195, "n_fig": 10, "n_tab": 13, "L_tab": 6188, "n_element_tab": 454, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1046, "n_element_tab_1": 142, "formula_len_all": 3606, "formula_len_all_1": 1410, "len_all": 230474, "len_all_1": 71403, "len_abs": 6236, "len_title": 158, "len_sents": 61298, "len_sents_1": 33619, "n_sents": 476, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 110, "L_abs": 1515, "title": "Learning Robust Representations with Long-Term Information for Generalization in Visual Reinforcement Learning", "abs": "Generalization in visual reinforcement learning (VRL) aims to learn agents that can adapt to test environments with unseen visual distractions. Despite advances in robust representations learning, many methods do not take into account the essential downstream task of sequential decision-making. This leads to representations that lack critical long-term information, impairing decision-making abilities in test environments. To tackle this problem, we propose a novel robust action-value representation learning (ROUSER) under the information bottleneck (IB) framework. ROUSER learns robust representations to capture long-term information from the decision-making objective (i.e., action values). Specifically, ROUSER uses IB to encode robust representations by maximizing their mutual information with action values for long-term information, while minimizing mutual information with state-action pairs to discard irrelevant features. As action values are unknown, ROUSER proposes to decompose robust representations of state-action pairs into one-step rewards and robust representations of subsequent pairs. Thus, it can use known rewards to compute the loss for robust representation learning. Moreover, we show that ROUSER accurately estimates action values using learned robust representations, making it applicable to various VRL algorithms. Experiments demonstrate that ROUSER outperforms several state-of-the-art methods in eleven out of twelve tasks, across both unseen background and color distractions.", "keywords": [ "Visual Reinforcement Learning", "Generalization", "Representation Learning", "Information Bottleneck", "One-Step Rewards" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "PDnM7mSO7M", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Tingyu Xia", "gender": "Female", "institution": "Jilin University", "country": "", "position": "PhD student" }, { "name": "Yahan Li", "gender": "unknown", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "Yi Chang", "gender": "Male", "institution": "Jilin University, China", "country": "CN", "position": "Full Professor" }, { "name": "Yuan Wu", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 30, "n_formula_1": 18, "n_ref_uni": 38, "n_ref": 50, "n_ref_all": 70, "n_fig": 6, "n_tab": 12, "L_tab": 7103, "n_element_tab": 598, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 5, "n_element_tab_1": 1, "formula_len_all": 1218, "formula_len_all_1": 875, "len_all": 151856, "len_all_1": 47355, "len_abs": 1400, "len_title": 103, "len_sents": 42029, "len_sents_1": 21404, "n_sents": 304, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1435, "title": "Large Language Model Evaluation via Matrix Nuclear-Norm", "abs": "As large language models (LLMs) continue to evolve, efficient evaluation metrics are vital for assessing their ability to compress information and reduce redundancy. While traditional metrics like Matrix Entropy offer valuable insights, they are computationally intensive for large-scale models due to their $\\( O(n^3) \\)$ time complexity with Singular Value Decomposition (SVD). To mitigate this issue, we introduce the Matrix Nuclear-Norm, which not only serves as a metric to quantify the data compression proficiency of LLM but also provides a convex approximation of matrix rank to capture both predictive discriminability and diversity. By employing the $\\( L_{1,2}\\text{-norm} \\)$ to further approximate the nuclear norm, we can effectively assess the model's information compression capabilities. This approach reduces the time complexity to $\\( O(n^2) \\)$ and eliminates the need for SVD computation. Consequently, the Matrix Nuclear-Norm achieves speeds 8 to 24 times faster than Matrix Entropy for the CEREBRAS-GPT model as sizes increase from 111M to 6.7B. This performance gap becomes more pronounced with larger models, as validated in tests with other models like Pythia. Additionally, evaluations on benchmarks and model responses confirm that our proposed Matrix Nuclear-Norm is a reliable, scalable, and efficient tool for assessing LLMs' performance, striking a balance between accuracy and computational efficiency.", "keywords": [ "Large language model", "evaluation", "matrix entropy", "nuclear norm" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "PDnEDS244P", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Chengdong Ma", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jiancong Xiao", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Postdoc" }, { "name": "Jing Huo", "gender": "Female", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Linjian Meng", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Mingzhi Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Qizhi Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Weijie J Su", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Associate Professor" }, { "name": "Yang Han", "gender": "unknown", "institution": "China Telecom", "country": "CN", "position": "Researcher" }, { "name": "Yaodong Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhaowei Zhang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" } ], "n_formula": 46, "n_formula_1": 19, "n_ref_uni": 60, "n_ref": 158, "n_ref_all": 197, "n_fig": 12, "n_tab": 8, "L_tab": 1420, "n_element_tab": 203, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4290, "formula_len_all_1": 1068, "len_all": 246962, "len_all_1": 62558, "len_abs": 1332, "len_title": 99, "len_sents": 60415, "len_sents_1": 27654, "n_sents": 564, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1337, "title": "Magnetic Mirror Descent Self-play Preference Optimization", "abs": "Self-play methods have demonstrated remarkable success in enhancing model capabilities across various domains. In the context of Reinforcement Learning from Human Feedback (RLHF), self-play not only boosts Large Language Model (LLM) performance but also overcomes the limitations of traditional Bradley-Terry (BT) model assumptions by finding the Nash equilibrium (NE) of a preference-based, two-player constant-sum game. However, existing methods either guarantee only average-iterate convergence, incurring high storage and inference costs, or converge to the NE of a regularized game, failing to accurately reflect true human preferences. In this paper, we introduce Magnetic Preference Optimization (MPO), a novel approach capable of achieving last-iterate convergence to the NE of the original game, effectively overcoming the limitations of existing methods. Building upon Magnetic Mirror Descent (MMD), MPO attains a linear convergence rate, making it particularly suitable for fine-tuning LLMs. To ensure our algorithm is both theoretically sound and practically viable, we present a simple yet effective implementation that adapts the theoretical insights to the RLHF setting. Empirical results demonstrate that MPO can significantly enhance the performance of LLMs, highlighting the potential of self-play methods in alignment.", "keywords": [ "AI Alignment", "Mirror Descent", "Self-play", "Nash Equilibrium" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "PDgZ3rvqHn", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Buqing Nie", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Yangqing Fu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yue Gao", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "", "position": "Associate Professor" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 33, "n_ref": 75, "n_ref_all": 119, "n_fig": 9, "n_tab": 7, "L_tab": 2267, "n_element_tab": 188, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1246, "n_element_tab_1": 126, "formula_len_all": 1362, "formula_len_all_1": 1167, "len_all": 120985, "len_all_1": 66185, "len_abs": 1228, "len_title": 128, "len_sents": 34544, "len_sents_1": 27507, "n_sents": 275, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1235, "title": "Select before Act: Spatially Decoupled Action Repetition for Continuous Control", "abs": "Reinforcement Learning (RL) has achieved remarkable success in various continuous control tasks, such as robot manipulation and locomotion.\nDifferent to mainstream RL which makes decisions at individual steps, recent studies have incorporated action repetition into RL, achieving enhanced action persistence with improved sample efficiency and superior performance.\nHowever, existing methods treat all action dimensions as a whole during repetition, ignoring variations among them.\nThis constraint leads to inflexibility in decisions, which reduces policy agility with inferior effectiveness. \nIn this work, we propose a novel repetition framework called SDAR, which implements Spatially Decoupled Action Repetition through performing closed-loop act-or-repeat selection for each action dimension individually.\nSDAR achieves more flexible repetition strategies, leading to an improved balance between action persistence and diversity.\nCompared to existing repetition frameworks, SDAR is more sample efficient with higher policy performance and reduced action fluctuation.\nExperiments are conducted on various continuous control scenarios, \ndemonstrating the effectiveness of spatially decoupled repetition design proposed in this work.", "keywords": [ "Reinforcement Learning", "Action Repetition" ], "rating_list": [ 5, 8, 6 ], "soundness_list": [ 2, 4, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "PDV3SmO6Iw", "primary_area": "foundation or frontier models, including LLMs", "n_author": 1, "author_info_list": [ { "name": "Yang Cao", "gender": "Not Specified", "institution": "Wyoming Seminary", "country": "US", "position": "High School Student" } ], "n_formula": 47, "n_formula_1": 17, "n_ref_uni": 32, "n_ref": 70, "n_ref_all": 83, "n_fig": 4, "n_tab": 1, "L_tab": 1245, "n_element_tab": 119, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 28, "n_element_tab_1": 4, "formula_len_all": 3069, "formula_len_all_1": 782, "len_all": 182428, "len_all_1": 55832, "len_abs": 1437, "len_title": 103, "len_sents": 36201, "len_sents_1": 24654, "n_sents": 317, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 1370, "title": "SORSA: Singular Values and Orthonormal Regularized Singular Vectors Adaptation of Large Language Models", "abs": "In this paper, we propose Singular Values and Orthonormal Regularized Singular Vectors Adaptation, or SORSA, a novel PEFT method. Each SORSA adapter consists of two main parts: trainable principal singular weights $W_p = U_p \\text{diag}(S_p) V^\\top_p$, and frozen residual weights $W_r = U_r \\text{diag}(S_r) V^\\top_r$. These parts are initialized by performing singular value decomposition (SVD) on pre-trained weights. Moreover, we implement and analyze an orthonormal regularizer, which we prove could decrease the condition number of $W_p$ and make the optimization more efficient. SORSA adapters could be merged during inference, thus eliminating any inference latency. We also introduce a method to analyze the variation of the parameters by performing SVD and discuss and analyze SORSA's superiority in minimizing the alteration in the SVD aspect. After all, SORSA shows a faster convergence than LoRA and PiSSA in our experiments. On the GSM-8K benchmark, Llama 2 7B adapted using SORSA achieved 56.03\\% accuracy, surpassing LoRA (42.30\\%), AdaLoRA (47.30\\%), Full FT (49.05\\%), and PiSSA (53.07\\%). On the MATH benchmark, SORSA achieved 10.36\\% accuracy, outperforming LoRA (5.50\\%), AdaLoRA (6.48\\%), Full FT (7.22\\%), and PiSSA (7.44\\%). We conclude that SORSA offers a new perspective on parameter-efficient fine-tuning, demonstrating remarkable performance.", "keywords": [ "Deep Learning", "LLM", "PEFT", "LoRA", "SVD" ], "rating_list": [ 3, 6, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "PD8JVDg8mB", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Dibya Ghosh", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 36, "n_ref": 63, "n_ref_all": 76, "n_fig": 3, "n_tab": 5, "L_tab": 1230, "n_element_tab": 190, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1196, "n_element_tab_1": 206, "formula_len_all": 746, "formula_len_all_1": 738, "len_all": 121983, "len_all_1": 65831, "len_abs": 1342, "len_title": 129, "len_sents": 36616, "len_sents_1": 32314, "n_sents": 189, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1356, "title": "Annotation Bootstrapping: Reinforcing Visual Pre-Training using Unlabelled Images", "abs": "A common approach to learning from unlabeled images is to train models to satisfy invariances on these images, such as consistency under augmentations or crops. Despite successes on Imagenet, these approaches struggle to learn from larger uncurated datasets like web crawls or video, where such inductive biases only weakly hold. How can we more effectively learn from broader datasets? Instead of training models to be invariant across views, we study an alternative approach encouraging model representations to be \\textit{predictive} of important semantics of adjacent views of an image. We concurrently train a model to predict semantic annotations from images (generated either self-supervised, or from auxiliary datasets); and bootstrap the model's semantics by predicting, given a cropped view of an image and the coordinates for a nearby crop, the model's annotation distribution for the neighboring view. A core strength of this approach is the ability to extract information universally from both unlabelled and labelled image data, incorporating captions, bounding boxes, and other annotations when they are present. Our experiments show that annotation propagation improves pre-training on unlabelled datasets in the wild, including video datasets like EpicKitchens, scene datasets like COCO, and uncurated web-scale image datasets like CC12M.", "keywords": [ "visual pretraining", "self supervised learning", "bootstrapping" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "PD3I0iOYOd", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Murari Mandal", "gender": "Male", "institution": "Kalinga Institute of Industrial Technology (KIIT) Bhubaneswar India", "country": "IN", "position": "Assistant Professor" }, { "name": "Stefan Schoepf", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Vikram Singh Chundawat", "gender": "Male", "institution": "Sagepilot AI", "country": "IN", "position": "CTO" }, { "name": "Alexandra Brintrup", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "Full Professor" }, { "name": "Prasanna Dhungana", "gender": "Male", "institution": "Kalinga Institute Of Industrial Technology", "country": "IN", "position": "Undergrad student" }, { "name": "Pushkar Niroula", "gender": "Male", "institution": "Kalinga Institute of Industrial Technology", "country": "IN", "position": "Undergrad student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 20, "n_ref": 36, "n_ref_all": 51, "n_fig": 6, "n_tab": 5, "L_tab": 3001, "n_element_tab": 564, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 807, "n_element_tab_1": 187, "formula_len_all": 569, "formula_len_all_1": 404, "len_all": 110738, "len_all_1": 58646, "len_abs": 1803, "len_title": 108, "len_sents": 35070, "len_sents_1": 26882, "n_sents": 234, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1847, "title": "ConDa: Fast Federated Unlearning with Contribution Dampening", "abs": "Federated learning (FL) has enabled collaborative model training across decentralized data sources or clients. While adding new participants to a shared model does not pose great technical hurdles, the removal of a participant and their related information contained in the shared model remains a challenge. To address this problem, federated unlearning has emerged as a critical research direction, seeking to remove information from globally trained models without harming the model performance on the remaining data. Most modern federated unlearning methods use costly approaches such as the use of remaining clients data to retrain the global model or methods that would require heavy computation on client or server side. We introduce Contribution Dampening (\\textsc{ConDa}), a framework that performs efficient unlearning by tracking down the parameters which affect the global model for each client and performs synaptic dampening on the parameters of the global model that have privacy infringing contributions from the forgetting client. Our technique does not require clients data or any kind of retraining and it does not put any computational overhead on either the client or server side. We perform experiments on multiple datasets and demonstrate that \\textsc{ConDa} is effective to forget a client\u2019s data. In experiments conducted on the MNIST, CIFAR10, and CIFAR100 datasets, \\textsc{ConDa} proves to be the fastest federated unlearning method, outperforming the nearest state-of-the-art approach by at least 100\u00d7. Our emphasis is on the non-IID Federated Learning setting, which presents the greatest challenge for unlearning. Additionally, we validate \\textsc{ConDa}'s robustness through backdoor and membership inference attacks. We envision this work as a crucial component for FL in adhering to legal and ethical requirements.", "keywords": [ "federated unlearning", "machine unlearning", "fast unlearning", "non-IID federated unlearning" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "PC5WxcMRs8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Xiangyu Yue", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Xiaohan Ding", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Yiyuan Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zhang Zhixin", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 37, "n_ref": 68, "n_ref_all": 79, "n_fig": 7, "n_tab": 4, "L_tab": 1085, "n_element_tab": 112, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1237, "n_element_tab_1": 24, "formula_len_all": 977, "formula_len_all_1": 681, "len_all": 137255, "len_all_1": 65635, "len_abs": 1294, "len_title": 133, "len_sents": 34872, "len_sents_1": 28285, "n_sents": 249, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1299, "title": "Vision Search Assistant: Empower Vision-Language Models as Multimodal Search Engines", "abs": "Search engines enable the retrieval of unknown information with texts. However, traditional methods fall short when it comes to understanding unfamiliar visual content, such as identifying an object that the model has never seen before. This challenge is particularly pronounced for large vision-language models (VLMs): if the model has not been exposed to the object depicted in an image, it struggles to generate reliable answers to the user's question regarding that image. Moreover, as new objects and events continuously emerge, frequently updating VLMs is impractical due to heavy computational burdens. To address this limitation, we propose Vision Search Assistant, a novel framework that facilitates collaboration between VLMs and web agents. This approach leverages VLMs' visual understanding capabilities and web agents' real-time information access to perform open-world Retrieval-Augmented Generation via the web. By integrating visual and textual representations through this collaboration, the model can provide informed responses even when the image is novel to the system. Extensive experiments conducted on both open-set and closed-set QA benchmarks demonstrate that the Vision Search Assistant significantly outperforms the other models and can be widely applied to existing VLMs.", "keywords": [ "Vision Search Assistant", "VLM-Agent Collaboration", "Multimodal Search Engine" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "PBq8uOjGso", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Di Huang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Yanan Zhang", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Postdoc" }, { "name": "Jiaxin Chen", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Associate Professor" }, { "name": "Wang Zengran", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 3, "n_ref_uni": 40, "n_ref": 77, "n_ref_all": 92, "n_fig": 7, "n_tab": 6, "L_tab": 4054, "n_element_tab": 542, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1785, "n_element_tab_1": 278, "formula_len_all": 398, "formula_len_all_1": 140, "len_all": 161789, "len_all_1": 61731, "len_abs": 1385, "len_title": 175, "len_sents": 51093, "len_sents_1": 28385, "n_sents": 373, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 127, "L_abs": 1355, "title": "Unleashing the Potential of Unlabeled Data: Bidirectional Collaborative Semi-Supervised Active Learning for 3D Object Detection", "abs": "To address the annotation burden in LiDAR-based 3D object detection, active learning (AL) methods offer a promising solution. However, traditional active learning approaches solely rely on labeled data to train an initial model for data selection, overlooking the potential of leveraging unlabeled data. Recently, attempts to integrate semi-supervised learning (SSL) into AL with the goal of leveraging unlabeled data have faced challenges in effectively resolving the conflict between the two paradigms, resulting in less satisfactory performance.\nTo tackle this conflict, we propose a Bidirectional Collaborative Semi-Supervised Active Learning framework, dubbed as BC-SSAL. Specifically, from the perspective of SSL, we propose a Collaborative PseudoScene Pre-training (CPSP) method that effectively learns from unlabeled data without introducing adverse effects. From the perspective of AL, we design a Collaborative Active Learning (CAL) method tailored for outdoor LiDAR scenes, which complements the uncertainty and diversity methods by model cascading, alleviating the dilemma of sampling rare classes. Extensive experiments conducted on KITTI and Waymo demonstrate the effectiveness of our BC-SSAL. Especially, on the KITTI dataset, utilizing only 2\\% labeled data, BC-SSAL can achieve comparable performance to the model trained on the full set.", "keywords": [ "Autonomous Driving", "3D Object Detection", "Active Learning", "Semi-Supervised Learning" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "PBjCTeDL6o", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Alexandre Duplessis", "gender": "Male", "institution": "Department of Computer Science", "country": "GB", "position": "MS student" }, { "name": "Ching Lam Choi", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Serge Belongie", "gender": "Male", "institution": "University of Copenhagen", "country": "DK", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 85, "n_ref": 147, "n_ref_all": 172, "n_fig": 9, "n_tab": 8, "L_tab": 580, "n_element_tab": 78, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 66, "n_element_tab_1": 24, "formula_len_all": 1025, "formula_len_all_1": 754, "len_all": 205948, "len_all_1": 61437, "len_abs": 897, "len_title": 87, "len_sents": 44331, "len_sents_1": 28114, "n_sents": 398, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 39, "L_abs": 925, "title": "Unlearning-based Neural Interpretations", "abs": "Gradient-based interpretations often require an anchor point of comparison to avoid saturation in computing feature importance. We show that current baselines defined using static functions \u2014constant mapping, averaging or blurring \u2014 inject harmful colour, texture or frequency assumptions that deviate from model behaviour. This leads to accumulation of irregular gradients, resulting in attribution maps that are biased, fragile and manipulable. Departing from the static approach, we propose $\\texttt{UNI}$ to compute an (un)learnable, debiased and adaptive baseline by perturbing the input towards an $\\textit{unlearning direction}$ of steepest ascent. Our method discovers reliable baselines and succeeds in erasing salient features, which in turn locally smooths the high-curvature decision boundaries. Our analyses point to unlearning as a promising avenue for generating faithful, efficient and robust interpretations.", "keywords": [ "Explainability", "Attribution", "Debiasing", "Bias" ], "rating_list": [ 6, 8, 10, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "PBK9AM5HUm", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Afia Anjum", "gender": "Female", "institution": "University of Texas at Arlington", "country": "US", "position": "PhD student" }, { "name": "Arkajyoti Mitra", "gender": "Male", "institution": "University of Texas at Arlington", "country": "US", "position": "PhD student" }, { "name": "Ebelechukwu Nwafor", "gender": "Male", "institution": "Villanova University", "country": "US", "position": "Assistant Professor" }, { "name": "Habeeb Olufowobi", "gender": "Male", "institution": "University of Texas at Arlington", "country": "US", "position": "Assistant Professor" }, { "name": "Paul Agbaje", "gender": "unknown", "institution": "University of Texas at Arlington", "country": "US", "position": "PhD student" }, { "name": "Pranali Khose", "gender": "Female", "institution": "University of Texas at Arlington", "country": "US", "position": "MS student" } ], "n_formula": 17, "n_formula_1": 9, "n_ref_uni": 33, "n_ref": 64, "n_ref_all": 94, "n_fig": 9, "n_tab": 9, "L_tab": 4213, "n_element_tab": 344, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 6040, "n_element_tab_1": 260, "formula_len_all": 1213, "formula_len_all_1": 599, "len_all": 160734, "len_all_1": 69506, "len_abs": 1402, "len_title": 107, "len_sents": 50475, "len_sents_1": 29226, "n_sents": 391, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1407, "title": "Enhancing Graph Neural Networks: A Mutual Learning Approach", "abs": "Knowledge distillation (KD) techniques have emerged as a powerful tool for transferring expertise from complex teacher models to lightweight student models, particularly beneficial for deploying high-performance models in resource-constrained devices. This approach has been successfully applied to graph neural networks (GNNs), harnessing their expressive capabilities to generate node embeddings that capture structural and feature-related information. In this study, we depart from the conventional KD approach by exploring the potential of collaborative learning among GNNs. In the absence of a pre-trained teacher model, we show that relatively simple and shallow GNN architectures can synergetically learn efficient models capable of performing better during inference, particularly in tackling multiple tasks. We propose a collaborative learning framework where ensembles of student GNNs mutually teach each other throughout the training process. We introduce an adaptive logit weighting unit to facilitate efficient knowledge exchange among models and an entropy enhancement technique to improve mutual learning. These components dynamically empower the models to adapt their learning strategies during training, optimizing their performance for downstream tasks. Extensive experiments conducted on three datasets each for node and graph classification demonstrate the effectiveness of our approach.", "keywords": [ "graph neural network", "knowledge distillation", "online deep mutual learning" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 2, 3 ] }, { "paper_id": "PA3MWNDD6O", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Diego Klabjan", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "Full Professor" }, { "name": "Mengfan Xu", "gender": "Female", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" } ], "n_formula": 129, "n_formula_1": 2, "n_ref_uni": 34, "n_ref": 86, "n_ref_all": 90, "n_fig": 1, "n_tab": 5, "L_tab": 2439, "n_element_tab": 135, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 33, "n_element_tab_1": 2, "formula_len_all": 15252, "formula_len_all_1": 404, "len_all": 231951, "len_all_1": 82970, "len_abs": 1983, "len_title": 68, "len_sents": 76364, "len_sents_1": 39857, "n_sents": 631, "n_sents_1": 274, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1995, "title": "Decentralized Blockchain-based Robust Multi-agent Multi-armed Bandit", "abs": "We study a robust, i.e. in presence of malicious participants, multi-agent multi-armed bandit problem where multiple participants are distributed on a fully decentralized blockchain, with the possibility of some being malicious. The rewards of arms are homogeneous among the honest participants, following time-invariant stochastic distributions, which are revealed to the participants only when certain conditions are met to ensure that the coordination mechanism is secure enough. The coordination mechanism's objective is to efficiently ensure the cumulative rewards gained by the honest participants are maximized. To this end and to the best of our knowledge, we are the first to incorporate advanced techniques from blockchains, as well as novel mechanisms, into such a cooperative decision making framework to design optimal strategies for honest participants. This framework allows various malicious behaviors and the maintenance of security and participant privacy. More specifically, we select a pool of validators who communicate to all participants, design a new consensus mechanism based on digital signatures for these validators, invent a UCB-based strategy that requires less information from participants through secure multi-party computation, and design the chain-participant interaction and an incentive mechanism to encourage participants' participation. Notably, we are the first to prove the theoretical regret of the proposed algorithm and claim its optimality. Unlike existing work that integrates blockchains with learning problems such as federated learning which mainly focuses on optimality via computational experiments, we demonstrate that the regret of honest participants is upper bounded by $\\log{T}$ under certain assumptions. The regret bound is consistent with the multi-agent multi-armed bandit problem without malicious participants and the robust multi-agent multi-armed bandit problem with purely Byzantine attacks which do not affect the entire system.", "keywords": [ "Online learning", "Multi-agent Bandit", "Blockchain" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "P9VdRQOyqu", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chang D. Yoo", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Eunseop Yoon", "gender": "Female", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Hee Suk Yoon", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Mark A. Hasegawa-Johnson", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 31, "n_ref": 66, "n_ref_all": 102, "n_fig": 19, "n_tab": 7, "L_tab": 2342, "n_element_tab": 232, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 277, "n_element_tab_1": 4, "formula_len_all": 600, "formula_len_all_1": 593, "len_all": 169763, "len_all_1": 58256, "len_abs": 1545, "len_title": 139, "len_sents": 48760, "len_sents_1": 29709, "n_sents": 322, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1551, "title": "Can Video LLMs Refuse to Answer? Alignment for Answerability in Video Large Language Models", "abs": "In the broader context of deep learning, Multimodal Large Language Models have achieved significant breakthroughs by leveraging powerful Large Language Models as a backbone to align different modalities into the language space. A prime exemplification is the development of Video Large Language Models (Video-LLMs). While numerous advancements have been proposed to enhance the video understanding capabilities of these models, they are predominantly trained on questions generated directly from video content. However, in real-world scenarios, users often pose questions that extend beyond the informational scope of the video, highlighting the need for Video-LLMs to assess the relevance of the question. We demonstrate that even the best-performing Video-LLMs fail to reject unfit questions-not necessarily due to a lack of video understanding, but because they have not been trained to identify and refuse such questions. To address this limitation, we propose alignment for answerability, a framework that equips Video-LLMs with the ability to evaluate the relevance of a question based on the input video and appropriately decline to answer when the question exceeds the scope of the video, as well as an evaluation framework with a comprehensive set of metrics designed to measure model behavior before and after alignment. Furthermore, we present a pipeline for creating a dataset specifically tailored for alignment for answerability, leveraging existing video-description paired datasets. The code and the dataset will be publicly available.", "keywords": [ "Multimodal Large Language Model", "Alignment" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "P98KMCf60l", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Gengze Xu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "MS student" }, { "name": "Hongjin Qian", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Xiaolin Hu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Xinhao Yao", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "PhD student" }, { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 21, "n_ref_uni": 50, "n_ref": 105, "n_ref_all": 145, "n_fig": 3, "n_tab": 1, "L_tab": 228, "n_element_tab": 50, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1641, "formula_len_all_1": 1625, "len_all": 158159, "len_all_1": 71299, "len_abs": 1483, "len_title": 140, "len_sents": 44682, "len_sents_1": 32398, "n_sents": 373, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1601, "title": "Theoretical Insights into Fine-Tuning Attention Mechanism: Generalization and Optimization", "abs": "Large Language Models (LLMs), built on Transformer architectures, exhibit remarkable generalization across a wide range of tasks. However, fine-tuning these models for specific tasks remains resource-intensive due to their extensive parameterization. \nIn this paper, we investigate two remarkable phenomena related to the attention mechanism during the fine-tuning of LLMs. The first phenomenon, termed \u201cUnequal Importance of Attention Matrices,\u201d highlights the impact of fine-tuning different weight matrices. It shows that optimizing the $\\mathbf{W}_v$ matrix yields significantly better performance than optimizing the $\\mathbf{W}_k$ matrix. Fine-tuning only the $\\mathbf{W}_q$ and $\\mathbf{W}_v$ matrices is computationally efficient while delivering results comparable to, or even better than fine-tuning all three matrices ($\\mathbf{W}_q$, $\\mathbf{W}_k$, and $\\mathbf{W}_v$). The second phenomenon, \u201cAttention Matrices with Customized Learning Rate Leads to Better Convergence,\u201d emphasizes the importance of assigning distinct learning rates to these matrices. Specifically, a higher learning rate for the $\\mathbf{W}_v$ matrix compared to $\\mathbf{W}_q$ and $\\mathbf{W}_k$ accelerates convergence and improves performance. Building on these insights, we propose a new strategy that improves fine-tuning efficiency in terms of both storage and time. Experimental results on benchmark datasets validate the effectiveness of this approach, supporting our theoretical findings. Our analysis lays the theoretical groundwork for configuring and improving lightweight algorithms in LLMs fine-tuning.", "keywords": [ "Large Language Models", "Attention mechanism", "Fine-tuning", "Generalization and Optimization" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 1, 1, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "P8uOZmypb6", "primary_area": "datasets and benchmarks", "n_author": 13, "author_info_list": [ { "name": "Alvin Wei Ming Tan", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Auddithio Nag", "gender": "Non-Binary", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Bria Lorelle Long", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Chengxu Zhuang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Daniel LK Yamins", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Grace Keene", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Researcher" }, { "name": "Michael Frank", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Full Professor" }, { "name": "Robert Z. Sparks", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Researcher" }, { "name": "Stefan Stojanov", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Steven Y. Feng", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Violet Xiang", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Virginia A. Marchman", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Researcher" }, { "name": "ZiYin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 62, "n_ref": 105, "n_ref_all": 124, "n_fig": 8, "n_tab": 5, "L_tab": 1584, "n_element_tab": 231, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1754, "n_element_tab_1": 226, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 164753, "len_all_1": 67577, "len_abs": 1766, "len_title": 157, "len_sents": 43231, "len_sents_1": 29293, "n_sents": 281, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1773, "title": "The BabyView dataset: High-resolution egocentric videos of infants\u2019 and young children\u2019s everyday experiences", "abs": "Human children far exceed modern machine learning algorithms in their sample efficiency, achieving high performance in key domains with much less data than current models. This \"data gap'' is a key challenge both for building intelligent artificial systems and for understanding human development. Egocentric video capturing children's experience\u2014their \"training data''\u2014is a key ingredient for comparison of humans and models and for the development of algorithmic innovations to bridge this gap. Yet there are few such datasets available, and extant data are low-resolution, have limited metadata, and importantly, represent only a small set of children's experiences. Here, we provide the first release of a large developmental egocentric video dataset\u2014the BabyView dataset\u2014recorded using a high-resolution camera with a large vertical field-of-view and gyroscope/accelerometer data. This 430 hour dataset includes egocentric videos from children spanning 6 months\u20133 years of age in longitudinal, at-home contexts. We provide gold-standard annotations for the evaluation of speech transcription, speaker diarization, and human pose estimation, and evaluate models in each of these domains. We train self-supervised language and vision models and evaluate their transfer to out-of-distribution tasks including syntactic structure learning, object recognition, depth estimation, and image segmentation. Although performance in each scales with dataset size, overall performance is relatively lower than when models are trained on curated datasets, especially in the visual domain. Our dataset stands as an open challenge for robust, human-like AI systems: how can such systems achieve human-levels of success on the same scale and distribution of training data as humans?", "keywords": [ "egocentric video dataset", "self-supervised learning", "developmental curriculum", "data gap" ], "rating_list": [ 5, 5, 8, 3, 6 ], "soundness_list": [ 2, 2, 3, 1, 3 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 4, 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "P8pqeEkn1H", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Dani Valevski", "gender": "Male", "institution": "Google", "country": "IL", "position": "Researcher" }, { "name": "Moab Arar", "gender": "Male", "institution": "Facebook", "country": "IL", "position": "Intern" }, { "name": "Shlomi Fruchter", "gender": "Male", "institution": "Research, Google", "country": "IL", "position": "Researcher" }, { "name": "Yaniv Leviathan", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 52, "n_ref_all": 81, "n_fig": 11, "n_tab": 6, "L_tab": 477, "n_element_tab": 38, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 722, "n_element_tab_1": 57, "formula_len_all": 224, "formula_len_all_1": 224, "len_all": 140735, "len_all_1": 59541, "len_abs": 1128, "len_title": 91, "len_sents": 41806, "len_sents_1": 27300, "n_sents": 339, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 1134, "title": "Diffusion Models Are Real-Time Game Engines", "abs": "We present GameNGen, the first game engine powered entirely by a neural model that also enables real-time interaction with a complex environment over long trajectories at high quality. When trained on the classic game DOOM, GameNGen extracts gameplay and uses it to generate a playable environment that can interactively simulate new trajectories. GameNGen runs at 20 frames per second on a single TPU and remains stable over extended multi-minute play sessions. Next frame prediction achieves a PSNR of 29.4, comparable to lossy JPEG compression. Human raters are only slightly better than random chance at distinguishing short clips of the game from clips of the simulation, even after 5 minutes of auto-regressive generation. GameNGen is trained in two phases: (1) an RL-agent learns to play the game and the training sessions are recorded, and (2) a diffusion model is trained to produce the next frame, conditioned on the sequence of past frames and actions. Conditioning augmentations help ensure stable auto-regressive generation over long trajectories, and decoder fine-tuning improves the fidelity of visual details and text.", "keywords": [ "Diffusion Models", "Neural Game Engines", "Real-Time Simulation", "Reinforcement Learning", "Auto-Regressive Generation", "Video Game Simulation", "Interactive World Simulation" ], "rating_list": [ 8, 8, 5, 8, 5 ], "soundness_list": [ 3, 3, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 4, 2 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 4, 3 ] }, { "paper_id": "P8IBvXLAVk", "primary_area": "foundation or frontier models, including LLMs", "n_author": 12, "author_info_list": [ { "name": "Huajun Chen", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Jialong Wu", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "Jiamin Chen", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Ningyu Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Shengwei Ding", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Shuai Wang", "gender": "Male", "institution": "AIWaves Inc.", "country": "CN", "position": "Researcher" }, { "name": "Tiannan Wang", "gender": "Male", "institution": "AIWaves", "country": "CN", "position": "Researcher" }, { "name": "Wangchunshu Zhou", "gender": "Male", "institution": "Guangdong OPPO Mobile Telecommunications Corp.,Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xiaohua Xu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yixin Ou", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "\u674e \u9f99", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yuchen Jiang", "gender": "Female", "institution": "AIWaves Inc.", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 25, "n_ref": 63, "n_ref_all": 72, "n_fig": 2, "n_tab": 6, "L_tab": 694, "n_element_tab": 106, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 268, "n_element_tab_1": 46, "formula_len_all": 179, "formula_len_all_1": 105, "len_all": 157933, "len_all_1": 59827, "len_abs": 2011, "len_title": 46, "len_sents": 43538, "len_sents_1": 30674, "n_sents": 303, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1944, "title": "Symbolic Learning Enables Self-Evolving Agents", "abs": "The AI community has been exploring a pathway to artificial general intelligence (AGI) by developing \"language agents\", which are complex large language models (LLMs) pipelines involving both prompting techniques and tool usage methods. While language agents have demonstrated impressive capabilities for many real-world tasks, a fundamental limitation of current language agents research is that they are model-centric, or engineering-centric. That's to say, the progress on prompts, tools, and pipelines of language agents requires substantial manual engineering efforts from human experts rather than automatically learning from data. We believe the transition from model-centric, or engineering-centric, to data-centric, i.e., the ability of language agents to autonomously learn and evolve in environments, is the key for them to possibly achieve AGI. In this work, we introduce agent symbolic learning, a systematic framework that enables language agents to optimize themselves on their own in a data-centric way using symbolic optimizers. Specifically, we consider agents as symbolic networks where learnable weights are defined by prompts, tools, and the way they are stacked together. Agent symbolic learning is designed to optimize the symbolic network within language agents by mimicking two fundamental algorithms in connectionist learning: back-propagation and gradient descent. Instead of dealing with numeric weights, agent symbolic learning works with language-based weights, loss, and gradients. We conduct proof-of-concept experiments on both standard benchmarks and complex real-world tasks and show that agent symbolic learning enables language agents to update themselves after being created and deployed in the wild, resulting in \"self-evolving agents\". We demonstrate the potential of the agent symbolic learning framework and open-source the entire framework to facilitate future research on data-centric agent learning.", "keywords": [ "large language model", "language agents", "agent learning" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "P8FS9byr1c", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Baochun Li", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Full Professor" }, { "name": "Ningxin Su", "gender": "Female", "institution": "University of Toronto", "country": "", "position": "PhD student" }, { "name": "Sijia Chen", "gender": "Male", "institution": "Toronto University", "country": "CA", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 24, "n_ref": 80, "n_ref_all": 112, "n_fig": 8, "n_tab": 4, "L_tab": 2251, "n_element_tab": 231, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 4087, "n_element_tab_1": 256, "formula_len_all": 443, "formula_len_all_1": 962, "len_all": 239954, "len_all_1": 74575, "len_abs": 1535, "len_title": 105, "len_sents": 88588, "len_sents_1": 30767, "n_sents": 739, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1552, "title": "$Staple$: Towards Reliable Problem Solving with Large Language Models via Plan Optimization and Tree Search", "abs": "Large language models (LLMs) exhibit the ability to perform step-by-step reasoning when tackling complex problems across various tasks. To improve the reliability of multi-step reasoning and mitigate potential hallucinations, sophisticated prompting techniques have been developed to provide instructions on $what$ $to$ $do$ at each step, offering reasoning guidance before addressing specific questions. However, this additional prompting can increase time and token consumption without guaranteeing effectiveness. In response, this paper proposes $Staple$, a novel plan retrieval augmented reasoning framework that utilizes offline plan optimization. This approach involves constructing a plan database of general-purpose reasoning instructions. Subsequently, online plan searching facilitates the direct retrieval of optimal and effective step-by-step plans from the database when addressing new questions, serving as guidance for LLMs to derive correct answers. The offline stage uses LLMs to self-generate and optimize plans, storing them as tree structures via Monte Carlo Tree Search (MCTS) to form the plan database. Extensive experiments on mathematical and multi-task problems show that $Staple$ achieves competitive problem-solving rates while minimizing token usage and interactions. Importantly, the plan trees in the database are human-interpretable, revealing the prioritization of various plan combinations for a given task. In addition, the plan database can be reused, updated, and expanded by users for a wider range of applications.", "keywords": [ "Large language models; Prompt engineering; Complex reasoning;" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 1, 1, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "P7t2niLbvw", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Jung Taek Seo", "gender": "Male", "institution": "Gachon University", "country": "KR", "position": "Associate Professor" }, { "name": "Malik Hassanaly", "gender": "unknown", "institution": "The National Renewable Energy Laboratory", "country": "US", "position": "Researcher" }, { "name": "Nguyen Hoang Khoi Do", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "PhD student" }, { "name": "Truc Nguyen", "gender": "Male", "institution": "The National Renewable Energy Laboratory", "country": "US", "position": "Postdoc" }, { "name": "raed alharbi", "gender": "Male", "institution": "Saudi Electroic University", "country": "SA", "position": "Assistant Professor" }, { "name": "My Thai", "gender": "Female", "institution": "University of Florida", "country": "US", "position": "Full Professor" } ], "n_formula": 56, "n_formula_1": 14, "n_ref_uni": 61, "n_ref": 106, "n_ref_all": 134, "n_fig": 8, "n_tab": 14, "L_tab": 13310, "n_element_tab": 1076, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 410, "n_element_tab_1": 64, "formula_len_all": 3856, "formula_len_all_1": 1064, "len_all": 271567, "len_all_1": 70548, "len_abs": 1298, "len_title": 148, "len_sents": 80171, "len_sents_1": 34176, "n_sents": 569, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1303, "title": "Swift Hydra: Self-Reinforcing Generative Framework for Anomaly Detection with Multiple Mamba Models", "abs": "Despite a plethora of anomaly detection models developed over the years, their ability to generalize to unseen anomalies remains an issue, particularly in critical systems. This paper aims to address this challenge by introducing Swift Hydra, a new framework for training an anomaly detection method based on generative AI and reinforcement learning (RL). Through featuring an RL policy that operates on the latent variables of a generative model, the framework synthesizes novel and diverse anomaly samples that are capable of bypassing a detection model. These generated synthetic samples are, in turn, used to augment the detection model, further improving its ability to handle challenging anomalies. Swift Hydra also incorporates Mamba models structured as a Mixture of Experts (MoE) to enable scalable adaptation of the number of Mamba experts based on data complexity, effectively capturing diverse feature distributions without increasing the model\u2019s inference time. Empirical evaluations on ADBench benchmark demonstrate that Swift Hydra outperforms other state-of-the-art anomaly detection models while maintaining a relatively short inference time. From these results, our research highlights a new and auspicious paradigm of integrating RL and generative AI for advancing anomaly detection.", "keywords": [ "Anomaly detection", "Reinforcement learning", "Generative AI" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 4, 3, 2 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "P7s4WYF1rf", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Ke Yan", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Lu Zhang", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Shouhong Ding", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "researcher" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 18, "n_ref": 41, "n_ref_all": 49, "n_fig": 3, "n_tab": 3, "L_tab": 1866, "n_element_tab": 212, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 372, "n_element_tab_1": 5, "formula_len_all": 664, "formula_len_all_1": 596, "len_all": 102311, "len_all_1": 57411, "len_abs": 1617, "len_title": 124, "len_sents": 32601, "len_sents_1": 26543, "n_sents": 242, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1626, "title": "YouCLIP: Advancing Multilingual Cross-Modal Learning with Efficient Training.", "abs": "Since the advent of vision-language pretraining, the CLIP model has become a foundational model for many downstream tasks. However, most of the advanced CLIP models available today are trained primarily on English, making them poorly suited for other languages. This limits accessibility for countries where other languages are dominant. Given that training CLIP models requires vast amounts of GPU resources and data, which most countries lack due to the absence of companies on the scale of Google or OpenAI, this paper proposes an efficient and straightforward three-stage fine-tuning method, which allows for the conversion of the most powerful English CLIP model into models for other languages. \nIn these three stages of training, the first stage focuses on aligning the embedding layer, followed by token fusion in the second stage, and finally contrastive learning fine-tuning in the third stage.\nMeanwhile, to improve data quality, we propose a translation filtering model to filter the data.\nIn this work, we target Chinese as the language of interest and name the resulting model YouCLIP, which is currently the most powerful Chinese CLIP model, significantly outperforming previous models across all Chinese benchmarks. For example, YouCLIP improves the text-to-image Recall@1 score on the COCO-CN dataset from 63.4 to 73.1. Additionally, YouCLIP retains strong English capabilities, achieving a Top-1 accuracy of 76.9 on ImageNet. Despite these impressive results, YouCLIP requires the least amount of training resources compared to other Chinese CLIP models. All models and code for YouCLIP will be open-sourced.", "keywords": [ "CLIP; Vision-Language Pre-training; Non-English CLIP" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "P7f55HQtV8", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Mabiao Long", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yehui Tang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 69, "n_formula_1": 14, "n_ref_uni": 56, "n_ref": 106, "n_ref_all": 116, "n_fig": 6, "n_tab": 9, "L_tab": 3907, "n_element_tab": 308, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 913, "n_element_tab_1": 86, "formula_len_all": 4027, "formula_len_all_1": 1043, "len_all": 201231, "len_all_1": 73013, "len_abs": 1759, "len_title": 123, "len_sents": 60636, "len_sents_1": 34495, "n_sents": 486, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1791, "title": "QuaDiM: A Conditional Diffusion Model For Quantum State Property Estimation", "abs": "Quantum state property estimation (QPE) is a fundamental challenge in quantum many-body problems in physics and chemistry, involving the prediction of characteristics such as correlation and entanglement entropy through statistical analysis of quantum measurement data. Recent advances in deep learning have provided powerful solutions, predominantly using auto-regressive models. These models generally assume an intrinsic ordering among qubits, aiming to approximate the classical probability distribution through sequential training. However, unlike natural language, the entanglement structure of qubits lacks an inherent ordering, hurting the motivation of such models. In this paper, we introduce a novel, non-autoregressive generative model called \\textbf{QuaDiM}, designed for \\underline{\\textbf{Qua}}ntum state property estimation using \\underline{\\textbf{Di}}ffusion \\underline{\\textbf{M}}odels. QuaDiM progressively denoises Gaussian noise into the distribution corresponding to the quantum state, encouraging equal, unbiased treatment of all qubits. QuaDiM learns to map physical variables to properties of the ground state of the parameterized Hamiltonian during offline training. Afterwards one can sample from the learned distribution conditioned on previously unseen physical variables to collect measurement records and employ post-processing to predict properties of unknown quantum states. We evaluate QuaDiM on large-scale QPE tasks using classically simulated data on the 1D anti-ferromagnetic Heisenberg model with the system size up to 100 qubits. Numerical results demonstrate that \\model outperforms baseline models, particularly auto-regressive approaches, under conditions of limited measurement data during training and reduced sample complexity during inference.", "keywords": [ "quantum", "property estimation", "machine learning" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "P7O1Vt1BdU", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Akram Aldroubi", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "Full Professor" }, { "name": "Ashkan Shahbazi", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "PhD student" }, { "name": "Matthew Thorpe", "gender": "unknown", "institution": "The University of Warwick", "country": "GB", "position": "Associate Professor" }, { "name": "Soheil Kolouri", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "Assistant Professor" }, { "name": "Xinran Liu", "gender": "unknown", "institution": "Vanderbilt University", "country": "US", "position": "PhD student" }, { "name": "Yikun Bai", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "Postdoc" }, { "name": "ROCIO DIAZ MARTIN", "gender": "Female", "institution": "Tufts University", "country": "US", "position": "Postdoc" } ], "n_formula": 47, "n_formula_1": 27, "n_ref_uni": 26, "n_ref": 44, "n_ref_all": 70, "n_fig": 8, "n_tab": 1, "L_tab": 164, "n_element_tab": 22, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4820, "formula_len_all_1": 2332, "len_all": 139806, "len_all_1": 54923, "len_abs": 1615, "len_title": 79, "len_sents": 40889, "len_sents_1": 21646, "n_sents": 338, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 31, "L_abs": 1600, "title": "Expected Sliced Transport Plans", "abs": "The optimal transport (OT) problem has gained significant traction in modern machine learning for its ability to: (1) provide versatile metrics, such as Wasserstein distances and their variants, and (2) determine optimal couplings between probability measures. To reduce the computational complexity of OT solvers, methods like entropic regularization and sliced optimal transport have been proposed. The sliced OT framework improves efficiency by comparing one-dimensional projections (slices) of high-dimensional distributions. However, despite their computational efficiency, sliced-Wasserstein approaches lack a transportation plan between the input measures, limiting their use in scenarios requiring explicit coupling. In this paper, we address two key questions: Can a transportation plan be constructed between two probability measures using the sliced transport framework? If so, can this plan be used to define a metric between the measures? We propose a \u2018lifting\u2019 operation to extend one-dimensional optimal transport plans back to the original space of the measures. By computing the expectation of these lifted plans, we derive a new transportation plan, termed expected sliced transport (EST) plans. We further prove that using the EST plan to weight the sum of the individual Euclidean costs $\\|x - y\\|^p$ for moving from $x$ to $y$ results in a valid metric between the input discrete probability measures. Finally, we demonstrate the connection between our approach and the recently proposed min-SWGG, along with illustrative numerical examples that support our theoretical findings.", "keywords": [ "Optimal Transport", "Sliced Wasserstein", "Transportation Plan", "Probability Metrics" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "P7KRIiLM8T", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Carlo Luschi", "gender": "Male", "institution": "Graphcore", "country": "", "position": "VP & Head of Research" }, { "name": "Charlie Blake", "gender": "Male", "institution": "Graphcore", "country": "", "position": "Researcher" }, { "name": "Constantin Eichenberg", "gender": "unknown", "institution": "Aleph Alpha", "country": "", "position": "Researcher" }, { "name": "Douglas Orr", "gender": "Male", "institution": "Graphcore", "country": "", "position": "Researcher" }, { "name": "Josef Dean", "gender": "unknown", "institution": "University of Warwick", "country": "", "position": "MS student" }, { "name": "Lukas Balles", "gender": "Male", "institution": "Aleph Alpha ", "country": "DE", "position": "Researcher" }, { "name": "Luke Yuri Prince", "gender": "Male", "institution": "Graphcore", "country": "", "position": "Researcher" }, { "name": "Samuel Weinbach", "gender": "Male", "institution": "Aleph Alpha GmbH", "country": "", "position": "Researcher" }, { "name": "Andres Felipe Cruz Salinas", "gender": "Not Specified", "institution": "Cohere", "country": "DE", "position": "Researcher" }, { "name": "Bjoern Deiseroth", "gender": "Male", "institution": "Aleph Alpha ", "country": "DE", "position": "Researcher" } ], "n_formula": 71, "n_formula_1": 11, "n_ref_uni": 34, "n_ref": 100, "n_ref_all": 232, "n_fig": 25, "n_tab": 12, "L_tab": 4895, "n_element_tab": 971, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 688, "n_element_tab_1": 134, "formula_len_all": 7365, "formula_len_all_1": 930, "len_all": 345795, "len_all_1": 68075, "len_abs": 3287, "len_title": 100, "len_sents": 97892, "len_sents_1": 29071, "n_sents": 829, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 56, "L_abs": 872, "title": "u-$\\mu$P: The Unit-Scaled Maximal Update Parametrization", "abs": "The Maximal Update Parametrization ($\\mu$P) aims to make the optimal hyperparameters (HPs) of a model independent of its size, allowing them to be swept using a cheap proxy model rather than the full-size target model. We present a new scheme, u-$\\mu$P, which improves upon $\\mu$P by combining it with Unit Scaling, a method for designing models that makes them easy to train in low-precision. The two techniques have a natural affinity: $\\mu$P ensures that the scale of activations is independent of model size, and Unit Scaling ensures that activations, weights and gradients begin training with a scale of one. This synthesis opens the door to a simpler scheme, whose default values are near-optimal. This in turn facilitates a more efficient sweeping strategy, with u-$\\mu$P models reaching a lower loss than comparable $\\mu$P models and working out-of-the-box in FP8.", "keywords": [ "maximal update parametrization", "learning dynamics", "hyperparameter transfer", "efficiency", "training", "stability", "scaling", "numerics", "fp8", "low precision" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "P7KIGdgW8S", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Yair Davidson", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "MS student" }, { "name": "Nadav E Dym", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "Assistant Professor" } ], "n_formula": 123, "n_formula_1": 19, "n_ref_uni": 38, "n_ref": 89, "n_ref_all": 128, "n_fig": 10, "n_tab": 6, "L_tab": 1459, "n_element_tab": 72, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 897, "n_element_tab_1": 82, "formula_len_all": 14782, "formula_len_all_1": 1701, "len_all": 287321, "len_all_1": 70479, "len_abs": 3878, "len_title": 109, "len_sents": 95634, "len_sents_1": 30431, "n_sents": 903, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 61, "L_abs": 1625, "title": "On the H\u00f6lder Stability of Multiset and Graph Neural Networks", "abs": "Extensive research efforts have been put into characterizing and constructing maximally separating multiset and graph neural networks. \nHowever, recent empirical evidence suggests the notion of separation itself doesn't capture several interesting phenomena. On the one hand, the quality of this separation may be very weak, to the extent that the embeddings of \"separable\" objects might even be considered identical when using fixed finite precision. On the other hand, architectures which aren't capable of separation in theory, somehow achieve separation when taking the network to be wide enough.\n\nIn this work, we address both of these issues, by proposing a novel pair-wise separation quality analysis framework which is based on an adaptation of Lipschitz and H\u00f6lder stability to parametric functions. The proposed framework, which we name H\u00f6lder in expectation, allows for separation quality analysis, without restricting the analysis to embeddings that can separate all the input space simultaneously. We prove that common sum-based models are lower-H\u00f6lder in expectation, with an exponent\n that decays rapidly with the network's depth . Our analysis leads to adversarial examples of graphs which can be separated by three 1-WL iterations, but cannot be separated in practice by standard maximally powerful Message Passing Neural Networks (MPNNs). To remedy this, we propose two novel MPNNs with improved separation quality, one of which is lower Lipschitz in expectation. We show these MPNNs can easily classify our adversarial examples, and compare favorably with standard MPNNs on standard graph learning tasks.", "keywords": [ "graph neural networks", "message passing neural networks", "multiset neural networks", "neural network stability", "expressive power", "WL tests" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "P775MtcEEc", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Baptiste Debes", "gender": "Male", "institution": "Department of Electrical Engineering, KU Leuven, Belgium, KU Leuven", "country": "", "position": "PhD student" }, { "name": "Tinne Tuytelaars", "gender": "Female", "institution": "KU Leuven", "country": "BE", "position": "Full Professor" } ], "n_formula": 60, "n_formula_1": 32, "n_ref_uni": 50, "n_ref": 134, "n_ref_all": 153, "n_fig": 11, "n_tab": 1, "L_tab": 259, "n_element_tab": 23, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4579, "formula_len_all_1": 1985, "len_all": 169078, "len_all_1": 60739, "len_abs": 844, "len_title": 93, "len_sents": 50199, "len_sents_1": 26970, "n_sents": 405, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 848, "title": "Distributional Sobolev reinforcement learning", "abs": "Distributional reinforcement learning (DRL) is a framework for learning a complete distribution over returns, rather than merely estimating expectations. In this paper, we extend DRL on continuous state-action spaces by modeling not only the distribution over the scalar state-action value function but also its gradient. We refer to this method as Distributional Sobolev training. Inspired by Stochastic Value Gradients (SVG), we achieve this by leveraging a one-step world model of the reward and transition distributions implemented using a conditional Variational Autoencoder (cVAE). Our approach is sample-based and relies on Maximum Mean Discrepancy (MMD) to instantiate the distributional Bellman operator. We first showcase the method on a toy supervised learning problem. We then validate our algorithm in several Mujoco/Brax environments.", "keywords": [ "Reinforcement learning", "distributional reinforcement learnng", "Sobolev training of neural networks" ], "rating_list": [ 5, 1, 3, 3 ], "soundness_list": [ 3, 1, 1, 3 ], "presentation_list": [ 2, 1, 1, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "P6dwZJpJ4m", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Junxian He", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Lulu Zhao", "gender": "Female", "institution": "Beijing Academy of Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Weihao Zeng", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Intern" }, { "name": "Yijun Wang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "MS student" }, { "name": "Yuzhen Huang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Zifei Shan", "gender": "Male", "institution": "WeChat, Tencent", "country": "", "position": "Senior Applied Scientist" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 25, "n_ref": 55, "n_ref_all": 81, "n_fig": 8, "n_tab": 3, "L_tab": 537, "n_element_tab": 22, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 60, "n_element_tab_1": 4, "formula_len_all": 180, "formula_len_all_1": 107, "len_all": 126996, "len_all_1": 66338, "len_abs": 2237, "len_title": 134, "len_sents": 45819, "len_sents_1": 34184, "n_sents": 305, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 2251, "title": "B-STaR: Monitoring and Balancing Exploration and Exploitation in Self-Taught Reasoners", "abs": "In the absence of extensive human-annotated data for complex reasoning tasks, self-improvement -- where models are trained on their own outputs -- has emerged as a primary method for enhancing performance. Recently, the approach to self-improvement has shifted toward a more dynamic, online fashion through iterative training processes. However, the critical factors underlying the mechanism of these self-improving methods remain poorly understood, such as under what conditions self-improvement is effective, and what are the bottlenecks in the current iterations.\nIn this work, we identify and propose methods to monitor two pivotal factors in this iterative process: (1) the model's ability to explore and generate high-quality responses among multiple candidates (exploration); and (2) the reliability of external rewards in selecting the best responses from the generated outputs (exploitation).\nThese factors are inherently moving targets throughout the self-improvement cycles, yet their dynamics are rarely discussed in prior research -- It remains unclear what impedes continual model enhancement after only a few iterations. \nUsing mathematical reasoning as a case study, we begin with a quantitative analysis to track the dynamics of exploration and exploitation, discovering that a model's exploratory capabilities rapidly deteriorate over iterations, and the effectiveness of exploiting external rewards diminishes as well due to shifts in distribution from the original policy.\nMotivated by these findings, we introduce B-STaR, a Self-Taught Reasoning framework that autonomously adjusts configurations across iterations to Balance exploration and exploitation, thereby optimizing the self-teaching effectiveness based on the current policy model and available rewards.\nOur experiments in mathematical reasoning demonstrate that B-STaR not only enhances the model's exploratory capabilities throughout training but also achieves a more effective balance between exploration and exploitation, leading to superior performance. Crucially, this work deconstructs the opaque nature of self-training algorithms, elucidating the interpretable dynamics throughout the process and highlighting current limitations for future research to address.", "keywords": [ "large language models", "self-improvement", "complex reasoning" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "P6IVIoGRRg", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Molei Tao", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Wei Guo", "gender": "Not Specified", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yongxin Chen", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" } ], "n_formula": 92, "n_formula_1": 33, "n_ref_uni": 44, "n_ref": 81, "n_ref_all": 103, "n_fig": 2, "n_tab": 2, "L_tab": 508, "n_element_tab": 45, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 78, "n_element_tab_1": 11, "formula_len_all": 10038, "formula_len_all_1": 3708, "len_all": 179175, "len_all_1": 71521, "len_abs": 835, "len_title": 126, "len_sents": 41020, "len_sents_1": 27150, "n_sents": 362, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 951, "title": "Provable Benefit of Annealed Langevin Monte Carlo for Non-log-concave Sampling", "abs": "We consider the outstanding problem of sampling from an unnormalized density that may be non-log-concave and multimodal. To enhance the performance of simple Markov chain Monte Carlo (MCMC) methods, techniques of annealing type have been widely used. However, quantitative theoretical guarantees of these techniques are under-explored. This study takes a first step toward providing a non-asymptotic analysis of annealed MCMC. Specifically, we establish, for the first time, an oracle complexity of $\\widetilde{O}\\left(\\frac{d\\beta^2{\\cal A}^2}{\\varepsilon^6}\\right)$ for the simple annealed Langevin Monte Carlo algorithm to achieve $\\varepsilon^2$ accuracy in Kullback-Leibler divergence to the target distribution $\\pi\\propto{\\rm e}^{-V}$ on $\\mathbb{R}^d$ with $\\beta$-smooth potential $V$. Here, ${\\cal A}$ represents the action of a curve of probability measures interpolating the target distribution $\\pi$ and a readily sampleable distribution.", "keywords": [ "MCMC", "Annealed Langevin Monte Carlo", "Non-log-concave sampling", "Non-asymptotic analysis" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "P6G1Z6jkf3", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 11, "author_info_list": [ { "name": "Baoqi Pei", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Fei Wu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Jilan Xu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Lijin Yang", "gender": "Not Specified", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Weidi Xie", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yali Wang", "gender": "Male", "institution": "SIAT, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yifei Huang", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Postdoc" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuping He", "gender": "Female", "institution": "Nanjing university", "country": "CN", "position": "MS student" }, { "name": "Chen Guo", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Limin Wang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 51, "n_ref": 115, "n_ref_all": 144, "n_fig": 8, "n_tab": 11, "L_tab": 2387, "n_element_tab": 384, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 580, "n_element_tab_1": 122, "formula_len_all": 693, "formula_len_all_1": 414, "len_all": 199848, "len_all_1": 65759, "len_abs": 1388, "len_title": 135, "len_sents": 52341, "len_sents_1": 30908, "n_sents": 419, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1398, "title": "Modeling Fine-Grained Hand-Object Dynamics for Egocentric Video Representation Learning", "abs": "In egocentric video understanding, the motion of hands and objects as well as their interactions play a significant role by nature.\nHowever, existing egocentric video representation learning methods mainly focus on aligning video representation with high-level narrations, overlooking the intricate dynamics between hands and objects.\nIn this work, we aim to integrate the modeling of fine-grained hand-object dynamics into the video representation learning process.\nSince no suitable data is available, we introduce HOD, a novel pipeline employing a hand-object detector and a large language model to generate high-quality narrations with detailed descriptions of hand-object dynamics. \nTo learn these fine-grained dynamics, we propose EgoVideo, a model with a new lightweight motion adapter to capture fine-grained hand-object motion information. \nThrough our co-training strategy, EgoVideo effectively and efficiently leverages the fine-grained hand-object dynamics in the HOD data. \nExtensive experiments demonstrate that our method achieves state-of-the-art performance across multiple egocentric downstream tasks, including improvements of 6.3% in EK-100 multi-instance retrieval, 5.7% in EK-100 classification, and 16.3% in EGTEA classification in zero-shot settings. Furthermore, our model exhibits robust generalization capabilities in hand-object interaction and robot manipulation tasks.", "keywords": [ "Video representation learning", "Egocentric video", "Action recognition" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "P5rRGMk40p", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Bohan Yu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Imari Sato", "gender": "Female", "institution": "The University of Tokyo", "country": "JP", "position": "Professor" }, { "name": "Lixiong Chen", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Victor Adrian Prisacariu", "gender": "Male", "institution": "Niantic", "country": "US", "position": "Chief Scientist" } ], "n_formula": 23, "n_formula_1": 18, "n_ref_uni": 25, "n_ref": 33, "n_ref_all": 75, "n_fig": 13, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1448, "formula_len_all_1": 1240, "len_all": 99694, "len_all_1": 58901, "len_abs": 1838, "len_title": 85, "len_sents": 36518, "len_sents_1": 29883, "n_sents": 232, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 37, "L_abs": 1845, "title": "Continuous Surface Normal Integration", "abs": "We address a novel task for monocular explicit surface reconstruction that extends traditional surface normal integration over measurements on a regular grid to direct continuous surface depth estimation. Our solution accepts coordinates as queries and predicts both the normal and depth of an arbitrary query point by its relative locations and orientations to the points distributed in its vicinity. In general, all points are regarded by our model as random samples drawn from an underlying continuous gradient field of a surface which we parameterize using a field of polynomials to establish its topology. We establish a mapping from coordinates to a sequence of learnable polynomial coefficients to model a continuous surface and train a neural network to approximate it. We decompose a continuous surface representation into two components: (1) a set of grid points of unknown orientations whose locations are picked by a quadtree and (2) a set of sample points whose orientations are directly observable. Our training workflow estimates the normal of grid points and the locations of depth discontinuities iteratively. During each iteration, we generate a normal map of grid points for it to be processed by a standard bilateral normal integrator to identify the locations of depth discontinuities, which we use to refine the estimation for grid-based normal map in the subsequent iteration. As a result, the learned model generates both normal and depth for arbitrary coordinates accurately in a continuous field. We provide both theoretical formulation for our design and extensive empirical evidence to demonstrate that our proposed method not only delivers a performance as effective as its grid-based counterpart approaches but also flexibly and accurately addresses the continuous cases that existing methods are unable to handle.", "keywords": [ "normal integration", "shape modelling", "shape recovery" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "P5qCqYWD53", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Bin Dong", "gender": "Male", "institution": "Peking University", "country": "", "position": "Associate Professor" }, { "name": "Meitan Wang", "gender": "Female", "institution": "Peking University", "country": "", "position": "Undergrad student" }, { "name": "Yifan Luo", "gender": "Male", "institution": "Peking University", "country": "", "position": "Undergrad student" }, { "name": "Zhennan Zhou", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 60, "n_ref_all": 71, "n_fig": 6, "n_tab": 3, "L_tab": 349, "n_element_tab": 59, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 395, "n_element_tab_1": 69, "formula_len_all": 615, "formula_len_all_1": 612, "len_all": 137452, "len_all_1": 64276, "len_abs": 981, "len_title": 117, "len_sents": 36414, "len_sents_1": 28671, "n_sents": 332, "n_sents_1": 255, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 987, "title": "Jailbreak Instruction-Tuned Large Language Models via MLP Re-weighting", "abs": "In this paper, we investigate the safety mechanisms of instruction fine-tuned large language models (LLMs). We discover that re-weighting MLP neurons can significantly compromise a model's safety, especially for MLPs in end-of-sentence inferences. We hypothesize that LLMs evaluate the harmfulness of prompts during end-of-sentence inferences, and MLP layers plays a critical role in this process. Based on this hypothesis, we develop 2 novel white-box jailbreak methods: a prompt-specific method and a prompt-general method. The prompt-specific method targets individual prompts and optimizes the attack on the fly, while the prompt-general method is pre-trained offline and can generalize to unseen harmful prompts. Our methods demonstrate robust performance across 7 popular open-source LLMs, size ranging from 2B to 72B. Furthermore, our study provides insights into vulnerabilities of instruction-tuned LLM's safety and deepens the understanding of the internal mechanisms of LLMs.", "keywords": [ "Jailbreak", "AI safety", "mechanism interpretability" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "P5jreWnIjV", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Fengfeng Zhou", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Full Professor" }, { "name": "Jiaxin Zheng", "gender": "Female", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "Shikun Feng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Wei-Ying Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yanwen Huang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yanyan Lan", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Yinjun Harold Jia", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 40, "n_ref_all": 69, "n_fig": 6, "n_tab": 11, "L_tab": 7346, "n_element_tab": 843, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3070, "n_element_tab_1": 425, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 144172, "len_all_1": 65278, "len_abs": 1259, "len_title": 142, "len_sents": 39675, "len_sents_1": 26955, "n_sents": 238, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1263, "title": "MoleculeCLA: Rethinking Molecular Benchmark via Computational Ligand-Target Binding Analysis", "abs": "Molecular representation learning is pivotal for various molecular property prediction tasks related to drug discovery. Robust and accurate benchmarks are essential for refining and validating current methods. Existing molecular property benchmarks derived from wet experiments, however, face limitations such as data volume constraints, unbalanced label distribution, and noisy labels. To address these issues, we construct a large-scale and precise molecular representation dataset of approximately 140,000 small molecules, meticulously designed to capture an extensive array of chemical, physical, and biological properties, derived through a robust computational ligand-target binding analysis pipeline. We conduct extensive experiments on various deep learning models, demonstrating that our dataset offers significant physicochemical interpretability to guide model development and design. Notably, the dataset's properties are linked to binding affinity metrics, providing additional insights into model performance in drug-target interaction tasks. We believe this dataset will serve as a more accurate and reliable benchmark for molecular representation learning, thereby expediting progress in the field of artificial intelligence-driven drug discovery.", "keywords": [ "Molecular Benchmark", "Molecular Property Prediction" ], "rating_list": [ 8, 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 3, 4 ] }, { "paper_id": "P5icyaAReM", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Erland Brandser Olsson", "gender": "unknown", "institution": "Norwegian University of Science and Technology", "country": "NO", "position": "PhD student" }, { "name": "Zhirong Yang", "gender": "Male", "institution": "Norwegian University of Science and Technology", "country": "", "position": "Full Professor" } ], "n_formula": 23, "n_formula_1": 22, "n_ref_uni": 64, "n_ref": 163, "n_ref_all": 191, "n_fig": 2, "n_tab": 5, "L_tab": 1580, "n_element_tab": 110, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1424, "n_element_tab_1": 108, "formula_len_all": 1992, "formula_len_all_1": 2011, "len_all": 161978, "len_all_1": 64533, "len_abs": 1282, "len_title": 96, "len_sents": 33577, "len_sents_1": 25951, "n_sents": 273, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1425, "title": "Stochastic Approximation to Contrastive Learning", "abs": "Contrastive learning is a powerful paradigm that has been crucial for self-supervised representation learning. While there is evidence for its effectiveness, these methods typically rely on arbitrary definitions of positive and negative pairs. Most existing contrastive learning methods require large batch sizes during training due to their rigid control over the tradeoff between the two contrastive terms. Consequences are that, substantial computational resources are wasted on negative pairs that provide minimal learning signals. To address this issue, this work present a novel method. We reformulate contrastive learning as a matrix approximation problem using I-divergence, a non-normalized form of Kullback-Leibler divergence. Our proposed objective function is decomposable across instance pairs, enabling the development of efficient stochastic approximation algorithms from neighbor embeddings which perform well with fewer negative samples. Additionally, we generalize the scaling factor beyond normalization, allowing it to adaptively emphasize positive pairs that carry more learning signals, thereby reducing the computational waste associated with negative pairs. Experimental results on visual representation learning benchmark datasets such as CIFAR and ImageNet demonstrate major improvements over other contrastive learning methods, particularly when using small batches and with only one negative pair.", "keywords": [ "contrastive learning", "self-supervised learning", "unsupervised learning" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "P5UETqZXqT", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Dainong Hu", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "MS student" }, { "name": "Haewon Jeong", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "Assistant Professor" }, { "name": "Iain Weissburg", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "MS student" }, { "name": "Yao Qin", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "Assistant Professor" }, { "name": "Youngseok Yoon", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 6, "n_ref_uni": 41, "n_ref": 118, "n_ref_all": 166, "n_fig": 29, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1238, "formula_len_all_1": 229, "len_all": 190785, "len_all_1": 68876, "len_abs": 1465, "len_title": 168, "len_sents": 66010, "len_sents_1": 34527, "n_sents": 492, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1476, "title": "Model Collapse in the Chain of Diffusion Finetuning: A Novel Perspective from Quantitative Trait Modeling", "abs": "The success of generative models has reached a unique threshold where their outputs are indistinguishable from real data, leading to the inevitable contamination of future data collection pipelines with synthetic data. While their potential to generate infinite samples initially offers promise for reducing data collection costs and addressing challenges in data-scarce fields, the severe degradation in performance has been observed when iterative loops of training and generation occur---known as ''model collapse.'' This paper explores a practical scenario in which a pretrained text-to-image diffusion model is finetuned using synthetic images generated from a previous iteration, a process we refer to as the ''Chain of Diffusion.'' We first demonstrate the significant degradation in image qualities caused by this iterative process and identify the key factor driving this decline through rigorous empirical investigations. Drawing on an analogy between the Chain of Diffusion and biological evolution, we then introduce a novel theoretical analysis based on quantitative trait modeling. Our theoretical analysis aligns with empirical observations of the generated images in the Chain of Diffusion. Finally, we propose Reusable Diffusion Finetuning (ReDiFine), a simple yet effective strategy inspired by genetic mutations. ReDiFine mitigates model collapse without requiring any hyperparameter tuning, making it a plug-and-play solution for reusable image generation.", "keywords": [ "generative", "diffusion", "model collapse" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "P4pkLckzt8", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Lambert Deng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Intern" }, { "name": "Mohammed Haroon Dupty", "gender": "unknown", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Wee Sun Lee", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Yanfei Dong", "gender": "Female", "institution": "PayPal Inc.", "country": "SG", "position": "Researcher" }, { "name": "Yong Liang Goh", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zhuanghua Liu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 59, "n_formula_1": 23, "n_ref_uni": 68, "n_ref": 155, "n_ref_all": 193, "n_fig": 11, "n_tab": 20, "L_tab": 10543, "n_element_tab": 796, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 4891, "n_element_tab_1": 244, "formula_len_all": 3962, "formula_len_all_1": 1264, "len_all": 250892, "len_all_1": 76741, "len_abs": 2007, "len_title": 92, "len_sents": 60383, "len_sents_1": 30132, "n_sents": 526, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1054, "title": "Differentiable Cluster Graph Neural Network", "abs": "Graph Neural Networks often struggle with long-range information propagation and \nlocal heterophilous neighborhood aggregation. Inspired by the observation that cluster patterns manifest at global and local levels, we propose to tackle both challenges with a unified framework that incorporates a clustering inductive bias into the message passing mechanism, using additional cluster-nodes.\nCentral to our approach is the formulation of an optimal transport based clustering objective. \nHowever, optimizing this objective in a differentiable way is non-trivial.\nTo navigate this, we adopt an iterative process, alternating between solving for the cluster assignments and updating the node/cluster-node embeddings. \nNotably, our derived optimization steps are themselves simple yet elegant message passing steps operating seamlessly on a bipartite graph of nodes and cluster-nodes.\nOur clustering-based approach can effectively capture both local and global information, \ndemonstrated by extensive experiments on heterophilous and homophilous datasets.", "keywords": [ "Graph Neural Networks", "Graph Representation Learning", "Node Classification" ], "rating_list": [ 5, 8, 3, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 1, 2, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "P4o9akekdf", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Botao Ye", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Haofei Xu", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Marc Pollefeys", "gender": "Male", "institution": "Microsoft", "country": "CH", "position": "Director" }, { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Sifei Liu", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Songyou Peng", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Xueting Li", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 2, "n_ref_uni": 61, "n_ref": 156, "n_ref_all": 189, "n_fig": 17, "n_tab": 7, "L_tab": 3053, "n_element_tab": 131, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 2667, "n_element_tab_1": 91, "formula_len_all": 255, "formula_len_all_1": 113, "len_all": 182260, "len_all_1": 68908, "len_abs": 476, "len_title": 134, "len_sents": 52674, "len_sents_1": 31677, "n_sents": 407, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 86, "L_abs": 1728, "title": "No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images", "abs": "We introduce NoPoSplat, a feed-forward model capable of reconstructing 3D scenes parameterized by 3D Gaussians from unposed sparse multi-view images. Our model, trained exclusively with photometric loss, achieves real-time 3D Gaussian reconstruction during inference. To eliminate the need for accurate pose input during reconstruction, we anchor one input view's local camera coordinates as the canonical space and train the network to predict Gaussian primitives for all views within this space. This approach obviates the need to transform Gaussian primitives from local coordinates into a global coordinate system, thus avoiding errors associated with per-frame Gaussians and pose estimation. To resolve scale ambiguity, we design and compare various intrinsic embedding methods, ultimately opting to convert camera intrinsics into a token embedding and concatenate it with image tokens as input to the model, enabling accurate scene scale prediction. We utilize the reconstructed 3D Gaussians for novel view synthesis and pose estimation tasks and propose a two-stage coarse-to-fine pipeline for accurate pose estimation. Experimental results demonstrate that our pose-free approach can achieve superior novel view synthesis quality compared to pose-required methods, particularly in scenarios with limited input image overlap. For pose estimation, our method, trained without ground truth depth or explicit matching loss, significantly outperforms the state-of-the-art methods with substantial improvements. This work makes significant advances in pose-free generalizable 3D reconstruction and demonstrates its applicability to real-world scenarios. The source code and trained models will be made available to the public.", "keywords": [ "3D Gaussian Splatting", "Pose Free", "Pose Estimation", "Novel View Synthesis", "3D Reconstruction" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 1, 3, 4 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "P4XmKjXTrM", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 4, "author_info_list": [ { "name": "ALISTAIR JOHNSON", "gender": "Male", "institution": "The Hospital for Sick Children", "country": "", "position": "Principal Researcher" }, { "name": "Jack Gallifant", "gender": "Male", "institution": "Brigham and Women's Hospital, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Justin Xu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Researcher" }, { "name": "Matthew B.A. McDermott", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Postdoc" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 37, "n_ref_all": 46, "n_fig": 4, "n_tab": 3, "L_tab": 1663, "n_element_tab": 220, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1170, "n_element_tab_1": 229, "formula_len_all": 57, "formula_len_all_1": 0, "len_all": 110233, "len_all_1": 66661, "len_abs": 1697, "len_title": 114, "len_sents": 38082, "len_sents_1": 31812, "n_sents": 258, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1506, "title": "ACES: Automatic Cohort Extraction System for Event-Stream Datasets", "abs": "Reproducibility remains a significant challenge in machine learning (ML) for healthcare. Datasets, model pipelines, and even task/cohort definitions are often private in this field, leading to a significant barrier in sharing, iterating, and understanding ML results on electronic health record (EHR) datasets. This paper addresses a significant part of this problem by introducing the Automatic Cohort Extraction System (ACES) for event-stream data. This library is designed to simultaneously simplify the development of task/cohorts for ML in healthcare and also enable the reproduction of these cohorts, both at an exact level for single datasets and at a conceptual level across datasets. To accomplish this, ACES provides (1) a highly intuitive and expressive configuration language for defining both dataset-specific concepts and dataset-agnostic inclusion/exclusion criteria, and (2) a pipeline to automatically extract patient records that meet these defined criteria from real-world data. ACES can be automatically applied to any dataset in either the Medical Event Data Standard (MEDS) or EventStreamGPT (ESGPT) formats, or to ***any*** dataset in which the necessary task-specific predicates can be extracted in an event-stream form. ACES has the potential to significantly lower the barrier to entry for defining ML tasks that learn representations, redefine the way researchers interact with EHR datasets, and significantly improve the state of reproducibility for ML studies in this modality.", "keywords": [ "Automatic Task Specification", "Cohort Extraction", "Electronic Health Records", "Open Source Software", "Benchmarks", "Datasets" ], "rating_list": [ 3, 6, 6, 8 ], "soundness_list": [ 2, 2, 4, 4 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 1, 3, 3, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "P4WnvhVmPV", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Jingyi Cui", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Qiran Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Yi-Ge Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yisen Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 80, "n_formula_1": 16, "n_ref_uni": 28, "n_ref": 54, "n_ref_all": 85, "n_fig": 5, "n_tab": 13, "L_tab": 2874, "n_element_tab": 189, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 3244, "n_element_tab_1": 139, "formula_len_all": 10019, "formula_len_all_1": 987, "len_all": 188292, "len_all_1": 61401, "len_abs": 2588, "len_title": 173, "len_sents": 62057, "len_sents_1": 26920, "n_sents": 529, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1473, "title": "A Unified Theoretical Framework for Understanding Difficult-to-learn Examples in Contrastive Learning", "abs": "Unsupervised contrastive learning has shown significant performance improvements in recent years, often approaching or even rivaling supervised learning in various tasks. However, its learning mechanism is fundamentally different from that of supervised learning. Previous works have shown that difficult-to-learn examples (well-recognized in supervised learning as examples around the decision boundary), which are essential in supervised learning, contribute minimally in unsupervised settings. In this paper, perhaps surprisingly, we find that the direct removal of difficult-to-learn examples, although reduces the sample size, can boost the downstream classification performance of contrastive learning. To uncover the reasons behind this, we develop a theoretical framework modeling the similarity between different pairs of samples. Guided by this theoretical framework, we conduct a thorough theoretical analysis revealing that the presence of difficult-to-learn examples negatively affects the generalization of contrastive learning. Furthermore, we demonstrate that the removal of these examples, and techniques such as margin tuning and temperature scaling can enhance its generalization bounds, thereby improving performance.\nEmpirically, we propose a simple and efficient mechanism for selecting difficult-to-learn examples and validate the effectiveness of the aforementioned methods, which substantiates the reliability of our proposed theoretical framework.", "keywords": [ "Machine Learning; Contrastive Learning; Difficult-to-learn Examples" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "P4KzPJlnFk", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Di Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "PhD student" }, { "name": "Dong Yuan", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Dongzhan Zhou", "gender": "Female", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Junxian Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Minghao Yang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Peng Ye", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Shufei Zhang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Tao Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Wanli Ouyang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Full Professor" }, { "name": "Yining Tang", "gender": "unknown", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Yuchen Ren", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "PhD student" }, { "name": "Yuqiang Li", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Ziyang Xu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "haonan he", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 42, "n_ref": 71, "n_ref_all": 94, "n_fig": 10, "n_tab": 11, "L_tab": 5978, "n_element_tab": 543, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1793, "n_element_tab_1": 141, "formula_len_all": 326, "formula_len_all_1": 97, "len_all": 222751, "len_all_1": 55084, "len_abs": 1677, "len_title": 170, "len_sents": 70174, "len_sents_1": 25352, "n_sents": 511, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 120, "L_abs": 1330, "title": "Biology Instructions: A Dataset and Benchmark for Multi-Omics Sequence Understanding Capability of Large Language Models", "abs": "Large language models have already demonstrated their formidable capabilities in general domains, ushering in a revolutionary transformation. However, exploring and exploiting the extensive knowledge of these models to comprehend multi-omics biology remains underexplored. To fill this research gap, we first introduce Biology-Instructions, the first large-scale multi-omics biological sequences-related instruction-tuning dataset including DNA, RNA, proteins, and multi-molecules, designed to bridge the gap between large language models (LLMs) and complex biological sequences-related tasks. This dataset can enhance the versatility of LLMs by integrating diverse biological sequenced-related tasks with advanced reasoning capabilities, maintaining conversational fluency. Additionally, we reveal significant performance limitations in even state-of-the-art LLMs on biological sequence-related multi-omics tasks without specialized pre-training and instruction-tuning. We further develop a strong baseline called ChatMultiOmics with a novel three-stage training pipeline, demonstrating the powerful ability to understand biology by using Biology-Instructions. Biology-Instructions and ChatMultiOmics are publicly available and crucial resources for enabling more effective integration of LLMs with multi-omics sequence analysis.", "keywords": [ "Instruction benchmark", "Multi-omics", "AI for Biology" ], "rating_list": [ 5, 3, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 4, 3, 2 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "P4DbTSDQFu", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Bo Dai", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Chen Change Loy", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Fangzhou Hong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Shangchen Zhou", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Shuai Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xingang Pan", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Yushi LAN", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Zhaoyang Lyu", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "", "position": "Researcher" } ], "n_formula": 34, "n_formula_1": 12, "n_ref_uni": 65, "n_ref": 200, "n_ref_all": 224, "n_fig": 16, "n_tab": 5, "L_tab": 1922, "n_element_tab": 175, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1713, "n_element_tab_1": 122, "formula_len_all": 1936, "formula_len_all_1": 722, "len_all": 226287, "len_all_1": 69673, "len_abs": 1363, "len_title": 124, "len_sents": 49382, "len_sents_1": 31261, "n_sents": 409, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1012, "title": "GaussianAnything: Interactive Point Cloud Flow Matching for 3D Generation", "abs": "Recent advancements in diffusion models and large-scale datasets have revolutionized image and video generation, with increasing focus on 3D content generation. While existing methods show promise, they face challenges in input formats, latent space structures, and output representations. This paper introduces a novel 3D generation framework that addresses these issues, enabling scalable and high-quality 3D generation with an interactive Point Cloud-structured Latent space. Our approach utilizes a VAE with multi-view posed RGB-D-N renderings as input, features a unique latent space design that preserves 3D shape information, and incorporates a cascaded latent flow-based model for improved shape-texture disentanglement. The proposed method, GaussianAnything, supports multi-modal conditional 3D generation, allowing for point cloud, caption, and single-view image inputs. Experimental results demonstrate superior performance on various datasets, advancing the state-of-the-art in 3D content generation.", "keywords": [ "3D Object Generation", "Gaussian Splatting", "Flow-based Generative Models" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "P49gSPmrvN", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 1, "author_info_list": [ { "name": "Jonah Lynch", "gender": "unknown", "institution": "morphing.ai", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 7, "n_ref": 10, "n_ref_all": 19, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 40155, "len_all_1": 39952, "len_abs": 549, "len_title": 99, "len_sents": 18482, "len_sents_1": 18526, "n_sents": 128, "n_sents_1": 127, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 691, "title": "Time-dependent Development of Scientific Discourse: A Novel Approach Using UMAP and Word Embeddings", "abs": "This study presents a method for visualizing the time-dependent development of a scientific discipline using UMAP (Uniform Manifold Approximation and Projection) and text embeddings. This study demonstrates how the evolution of research interests and topics in a specific field can be mapped over time by encoding the abstracts of scholarly articles into a high-dimensional space and then projecting them into a 3D space. This computational approach converts the history of discourse into a point-cloud that can be further studied as a manifold and as a time series, which leads to new insights into the dynamics of scholarly discourse and the emergence and disappearance of research themes.", "keywords": [ "visualization", "umap", "dimension reduction", "history of science", "word embeddings" ], "rating_list": [ 1, 1, 1 ], "soundness_list": [ 1, 1, 1 ], "presentation_list": [ 2, 1, 1 ], "contribution_list": [ 1, 1, 1 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "P42DbV2nuV", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Lei Feng", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" }, { "name": "Runqi Lin", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Suqin Yuan", "gender": "Not Specified", "institution": "The University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 76, "n_ref": 138, "n_ref_all": 159, "n_fig": 7, "n_tab": 12, "L_tab": 2415, "n_element_tab": 239, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1099, "n_element_tab_1": 112, "formula_len_all": 347, "formula_len_all_1": 373, "len_all": 196572, "len_all_1": 62885, "len_abs": 1507, "len_title": 81, "len_sents": 57006, "len_sents_1": 30815, "n_sents": 380, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 33, "L_abs": 1511, "title": "Instance-dependent Early Stopping", "abs": "In machine learning practice, early stopping has been widely used to regularize models and can save computational costs by halting the training process when the model's performance on a validation set stops improving. However, conventional early stopping applies the same stopping criterion to all instances without considering their individual learning statuses, which leads to redundant computations on instances that are already well-learned. To further improve the efficiency, we propose an Instance-dependent Early Stopping (IES) method that adapts the early stopping mechanism from the entire training set to the instance level, based on the core principle that once the model has mastered an instance, the training on it should stop. IES considers an instance as mastered if the second-order differences of its loss value remain within a small range around zero. This offers a more consistent measure of an instance's learning status compared with directly using the loss value, and thus allows for a unified threshold to determine when an instance can be excluded from further backpropagation. We show that excluding mastered instances from backpropagation can increase the gradient norms, thereby accelerating the decrease of the training loss and speeding up the training process. Extensive experiments on benchmarks demonstrate that IES method can reduce backpropagation instances by 10%-50% while maintaining or even slightly improving the test accuracy and transfer learning performance of a model.", "keywords": [ "Early Stopping", "Supervised Learning", "Deep Learning", "Efficiency", "Sample Selection", "Data Pruning" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 4, 4, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "P2snmtUBkQ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Chao Ma", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Donghua Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Junqi Wu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Wen Yao", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "Principal Researcher" }, { "name": "Zijian Zhu", "gender": "unknown", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Tingsong Jiang", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 17, "n_ref_uni": 28, "n_ref": 45, "n_ref_all": 68, "n_fig": 5, "n_tab": 5, "L_tab": 1256, "n_element_tab": 178, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2770, "n_element_tab_1": 194, "formula_len_all": 856, "formula_len_all_1": 855, "len_all": 107517, "len_all_1": 66950, "len_abs": 1661, "len_title": 157, "len_sents": 32143, "len_sents_1": 30335, "n_sents": 240, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1675, "title": "Invisibility Stickers Against LiDAR: Adversarial Attacks on Point Cloud Intensity for LiDAR-based Detection", "abs": "Point cloud detection is crucial in applications such as autonomous driving systems and robotics. These systems utilize onboard LiDAR sensors to capture input point clouds, consisting of numerous three-dimensional coordinate points and their corresponding intensity of laser reflection. Recent studies have proposed various adversarial schemes to highlight the vulnerability of point cloud detectors. However, these studies primarily focused on generating or perturbing the coordinate positions of input points and are hard to attack in the physical world, while largely overlooking the significance of their intensity.\tThrough our exploration, we found that perturbing point cloud intensity poses significant security risks for point cloud object detectors. To the best of our knowledge, we are the first to attack on point cloud intensity and we propose an effective adversarial attack scheme, named I-ADV. Our method employs a voxel partition scheme to enhance physical implementation. To boost attack performance, we incorporate a gradient enhancement technique using 3D angle and distance features, along with an extremum-based gradient fusion strategy.\tExtensive experimental results demonstrate that by altering only point cloud intensity, our approach achieves state-of-the-art performance across detectors with various input representations, attaining attack success rates between 83.9% and 99.1%. Comprehensive ablation studies confirm the effectiveness and generality of the method\u2019s components. Additionally, comparing different attack schemes underscores the advantages of our point cloud intensity attack method in both performance and real-world applicability.", "keywords": [ "LiDAR-based detection", "adversarial examples", "deep learning" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 5, 5, 2 ] }, { "paper_id": "P2BgxNCFs9", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Bin Cui", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Fan Yang", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Guosheng Dong", "gender": "unknown", "institution": "Baichuan Inc.", "country": "", "position": "Researcher" }, { "name": "Miao Zheng", "gender": "Female", "institution": "Baichuan", "country": "CN", "position": "Researcher" }, { "name": "Youquan Li", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Zenan Zhou", "gender": "Male", "institution": "baichuan", "country": "CN", "position": "Researcher" }, { "name": "weipeng chen", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhang wen tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 24, "n_ref": 55, "n_ref_all": 78, "n_fig": 15, "n_tab": 5, "L_tab": 5390, "n_element_tab": 198, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1792, "n_element_tab_1": 127, "formula_len_all": 76, "formula_len_all_1": 76, "len_all": 148860, "len_all_1": 60911, "len_abs": 1243, "len_title": 148, "len_sents": 40862, "len_sents_1": 27534, "n_sents": 314, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 99, "L_abs": 1250, "title": "FB-Bench: A Fine-Grained Multi-Task Benchmark for Evaluating LLMs' Responsiveness to Human Feedback", "abs": "Human feedback is crucial in the interactions between humans and Large Language Models (LLMs). However, existing research primarily focuses on benchmarking LLMs in single-turn dialogues. Even in benchmarks designed for multi-turn dialogues, the user inputs are often independent, neglecting the nuanced and complex nature of human feedback within real-world usage scenarios. To fill this research gap, we introduce FB-Bench, a fine-grained, multi-task benchmark designed to evaluate LLMs' responsiveness to human feedback in real-world usage scenarios. Drawing from the two main interaction scenarios, FB-Bench comprises 734 meticulously curated samples, encompassing eight task types, five deficiency types of response, and nine feedback types. We extensively evaluate a broad array of popular LLMs, revealing significant variations in their performance across different interaction scenarios. Further analysis indicates that task, human feedback, and deficiencies of previous responses can also significantly impact LLMs' responsiveness. Our findings underscore both the strengths and limitations of current models, providing valuable insights and directions for future research. Both the toolkits and the dataset of FB-Bench will be released soon.", "keywords": [ "benchmark;human feedback;LLMs" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "P1qhkp8gQT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Wei-Lin Chen", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Yu Meng", "gender": "Male", "institution": "University of Virginia", "country": "", "position": "Assistant Professor" }, { "name": "Zhepei Wei", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 86, "n_ref": 152, "n_ref_all": 185, "n_fig": 8, "n_tab": 3, "L_tab": 849, "n_element_tab": 77, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1067, "n_element_tab_1": 80, "formula_len_all": 14776, "formula_len_all_1": 7928, "len_all": 256536, "len_all_1": 88324, "len_abs": 1696, "len_title": 137, "len_sents": 52116, "len_sents_1": 42266, "n_sents": 302, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1703, "title": "InstructRAG: Instructing Retrieval-Augmented Generation via Self-Synthesized Rationales", "abs": "Retrieval-augmented generation (RAG) has shown promising potential to enhance the accuracy and factuality of language models (LMs). However, imperfect retrievers or noisy corpora can introduce misleading or even erroneous information to the retrieved contents, posing a significant challenge to the generation quality. Existing RAG methods typically address this challenge by directly predicting final answers despite potentially noisy inputs, resulting in an implicit denoising process that is difficult to interpret and verify. On the other hand, the acquisition of explicit denoising supervision is often costly, involving significant human efforts. In this work, we propose InstructRAG, where LMs explicitly learn the denoising process through self-synthesized rationales --- First, we instruct the LM to explain how the ground-truth answer is derived from retrieved documents. Then, these rationales can be used either as demonstrations for in-context learning of explicit denoising or as supervised fine-tuning data to train the model. Compared to standard RAG approaches, InstructRAG requires no additional supervision, allows for easier verification of the predicted answers, and effectively improves generation accuracy. Experiments show InstructRAG consistently outperforms existing RAG methods in both training-free and trainable scenarios, achieving a relative improvement of 8.3% over the best baseline method on average across five knowledge-intensive benchmarks. Extensive analysis indicates that InstructRAG scales well with increased numbers of retrieved documents and consistently exhibits robust denoising ability even in out-of-domain datasets, demonstrating strong generalizability.", "keywords": [ "Large language model", "retrieval-augmented generation" ], "rating_list": [ 8, 8, 5, 6, 8 ], "soundness_list": [ 4, 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "P0eEalHM5h", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Guanbin Li", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Keyang Zhong", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Lionel Ni", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou))", "country": "CN", "position": "Full Professor" }, { "name": "Yuan Xie", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 17, "n_ref": 31, "n_ref_all": 39, "n_fig": 3, "n_tab": 1, "L_tab": 307, "n_element_tab": 38, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 25, "n_element_tab_1": 2, "formula_len_all": 242, "formula_len_all_1": 242, "len_all": 94926, "len_all_1": 55859, "len_abs": 1070, "len_title": 141, "len_sents": 33961, "len_sents_1": 27987, "n_sents": 248, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1075, "title": "LLMs Synergy : From Closed-Source Prototyping to Open-Source Model based Instruction Following", "abs": "We study the problem of constructing an efficient LLM-based instruction-following agent capable of comprehending and executing open-ended instructions in an embodied environment. We propose a method called LLMs Synergy for rapid domain adaptation in the instruction-following task without requiring additional manual annotations. This approach leverages a large general-purpose LLM to establish task baselines and generate domain-specific data. The knowledge from the larger model is then gradually transferred to a domain-tuned open-source LLM through a model transition process, enabling faster and more efficient adaptation. Accordingly, we developed the Dynamic Instruction Decomposition (DID) framework, specifically designed for LLM integration within this task scenario. The DID framework enables the agent to progressively align open-ended natural language commands with dynamic environmental contexts. Experimental results demonstrate significant improvements in task accuracy, leading to more effective instruction following and enhanced human-agent collaboration.", "keywords": [ "LLM", "instruction following", "domain adaptation" ], "rating_list": [ 3, 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 2, 3 ] }, { "paper_id": "OzwGZP8h2A", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 7, "author_info_list": [ { "name": "Feng Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Lei Chen", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Yinqi Bai", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Zhihai Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 32, "n_ref": 56, "n_ref_all": 96, "n_fig": 12, "n_tab": 13, "L_tab": 3861, "n_element_tab": 788, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 1633, "n_element_tab_1": 449, "formula_len_all": 1035, "formula_len_all_1": 576, "len_all": 138454, "len_all_1": 80117, "len_abs": 1872, "len_title": 116, "len_sents": 42995, "len_sents_1": 35416, "n_sents": 333, "n_sents_1": 258, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1880, "title": "Can Symbolic Regression of Boolean Functions Boost Logic Synthesis?", "abs": "Logic synthesis, which aims to synthesize a compact logic circuit with minimized size while exactly satisfying a given functionality, plays an important role in chip design. Recently, symbolic regression (SR) has shown great success in scientific discovery to recover underlying mathematical functions from given datasets. However, we found from extensive experiments that existing SR methods struggle to recover an exact and compact boolean function for logic synthesis given a truth table, i.e., complete input-output pairs of the circuit. The major challenges include (1) the greater complexity of underlying boolean functions compared to mathematical functions, and (2) the complex objectives involving both exact recovery and expression optimization towards circuit minimization. To address these challenges, we propose a novel symbolic factorized boolean searcher (SINE) to recover exact and compact boolean functions towards logic synthesis. Motivated by the Shannon decomposition theorem, SINE proposes a factorized boolean function representation to decompose the underlying boolean function into multiple simplified sub-functions, significantly reducing their complexity and thus improving the recovery accuracy. Moreover, based on the key observation that, logical sharing is significant for circuit size minimization. SINE proposes a self symmetric sub-expression motif operators mining mechanism to enhance the monte-carlo tree search method for optimized boolean function learning. To the best of our knowledge, SINE is the first symbolic regression framework capable of exactly recovering optimized boolean functions for circuit optimization. Experiments on circuits across a wide range of inputs demonstrate that SINE significantly improves the recovery accuracy and decreases the size of synthesized circuits by up to 24.32\\% compared to state-of-the-art methods.", "keywords": [ "Chip Design", "Logic Synthesis", "Symbolic Regression", "Monte-Carlo Tree Search" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 3, 3, 1 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 2, 4, 5, 4 ] }, { "paper_id": "Ozo7qJ5vZi", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Fabian Ruehle", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Assistant Professor" }, { "name": "James Halverson", "gender": "unknown", "institution": "Northeastern University", "country": "", "position": "Assistant Professor" }, { "name": "Marin Soljacic", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Max Tegmark", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Sachin Vaidya", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yixuan Wang", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Ziming Liu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Thomas Hou", "gender": "unknown", "institution": "California Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 77, "n_formula_1": 16, "n_ref_uni": 56, "n_ref": 109, "n_ref_all": 207, "n_fig": 25, "n_tab": 14, "L_tab": 4822, "n_element_tab": 319, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 1134, "n_element_tab_1": 60, "formula_len_all": 8856, "formula_len_all_1": 1450, "len_all": 343439, "len_all_1": 62860, "len_abs": 1347, "len_title": 31, "len_sents": 103598, "len_sents_1": 27674, "n_sents": 863, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 31, "L_abs": 1304, "title": "KAN: Kolmogorov\u2013Arnold Networks", "abs": "Inspired by the Kolmogorov-Arnold representation theorem, we propose Kolmogorov-Arnold Networks (KANs) as promising alternatives to Multi-Layer Perceptrons (MLPs). While MLPs have fixed activation functions on nodes (\"neurons''), KANs have learnable activation functions on edges (\"weights''). KANs have no linear weights at all -- every weight parameter is replaced by a univariate function parametrized as a spline. We show that this seemingly simple change makes KANs outperform MLPs in terms of accuracy and interpretability, on small-scale AI + Science tasks. For accuracy, smaller KANs can achieve comparable or better accuracy than larger MLPs in function fitting tasks. Theoretically and empirically, KANs possess faster neural scaling laws than MLPs. For interpretability, KANs can be intuitively visualized and can easily interact with human users. Through two examples in mathematics and physics, KANs are shown to be useful ``collaborators'' helping scientists (re)discover mathematical and physical laws. In summary, KANs are promising alternatives for MLPs. Despite the slow training of KANs, their improved accuracy and interpretability show the potential to improve today's deep learning models which rely heavily on MLPs. More research is necessary to make KANs' training more efficient.", "keywords": [ "Kolmogorov-Arnold networks", "Kolmogorov-Arnold representation theorem", "learnable activation functions", "interpretability", "AI + Science" ], "rating_list": [ 8, 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 4, 3 ], "confidence_list": [ 3, 4, 4, 3, 4 ] }, { "paper_id": "Ozfu2uBH55", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Chaoqi Wang", "gender": "Male", "institution": "xAI", "country": "US", "position": "Member of Technical Staff" }, { "name": "Chen Zhu", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Hao Ma", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Director" }, { "name": "Karthik Abinav Sankararaman", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Madian Khabsa", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Michal Valko", "gender": "Male", "institution": "Stealth Startup", "country": "US", "position": "Principal Researcher" }, { "name": "Sinong Wang", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research scientist" }, { "name": "Xuefei Cao", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Yuxin Chen", "gender": "unknown", "institution": "University of Chicago", "country": "US", "position": "Assistant Professor" }, { "name": "Zhaorun Chen", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Zhuokai Zhao", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 23, "n_ref_uni": 54, "n_ref": 99, "n_ref_all": 133, "n_fig": 12, "n_tab": 5, "L_tab": 1017, "n_element_tab": 97, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 365, "n_element_tab_1": 44, "formula_len_all": 2268, "formula_len_all_1": 1840, "len_all": 233000, "len_all_1": 75587, "len_abs": 1244, "len_title": 101, "len_sents": 74566, "len_sents_1": 33326, "n_sents": 742, "n_sents_1": 278, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1249, "title": "Preference Optimization with Multi-Sample Comparisons", "abs": "Recent advancements in generative models, particularly large language models (LLMs) and diffusion models, have been driven by extensive pretraining on large datasets followed by post-training. However, current post-training methods such as reinforcement learning from human feedback (RLHF) and direct alignment from preference methods (DAP) primarily utilize single-sample comparisons. These approaches often fail to capture critical characteristics such as generative diversity and bias, which are more accurately assessed through multiple samples. To address these limitations, we introduce a novel approach that extends post-training to include multi-sample comparisons. To achieve this, we propose Multi-sample Direct Preference Optimization (mDPO) and Multi-sample Identity Preference Optimization (mIPO). These methods improve traditional DAP methods by focusing on group-wise characteristics. Empirically, we demonstrate that multi-sample comparison is more effective in optimizing collective characteristics~(e.g., diversity and bias) for generative models than single-sample comparison. Additionally, our findings suggest that multi-sample comparisons provide a more robust optimization framework, particularly for dataset with label noise.", "keywords": [ "RLHF", "Preference Optimization" ], "rating_list": [ 3, 6, 6, 5 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 4, 4, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "OzUNDnpQyd", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Chence Shi", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "PhD student" }, { "name": "Hongyu Guo", "gender": "Male", "institution": "University of Ottawa", "country": "CA", "position": "Adjunct professor" }, { "name": "Jian Tang", "gender": "unknown", "institution": "HEC Montreal", "country": "", "position": "Assistant Professor" }, { "name": "Jiarui Lu", "gender": "Not Specified", "institution": "Mila - Qu\u00e9bec AI Institute; Universit\u00e9 de Montr\u00e9al", "country": "", "position": "PhD student" }, { "name": "Stephen Zhewen Lu", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Researcher" }, { "name": "Xiaoyin Chen", "gender": "unknown", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "PhD student" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" } ], "n_formula": 48, "n_formula_1": 10, "n_ref_uni": 66, "n_ref": 158, "n_ref_all": 193, "n_fig": 8, "n_tab": 6, "L_tab": 4744, "n_element_tab": 456, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 4962, "n_element_tab_1": 375, "formula_len_all": 4133, "formula_len_all_1": 744, "len_all": 252288, "len_all_1": 70791, "len_abs": 1668, "len_title": 109, "len_sents": 72488, "len_sents_1": 28521, "n_sents": 551, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1677, "title": "Structure Language Models for Protein Conformation Generation", "abs": "Proteins adopt multiple structural conformations to perform their diverse biological functions, and understanding these conformations is crucial for advancing drug discovery. Traditional physics-based simulation methods often struggle with sampling equilibrium conformations and are computationally expensive. Recently, deep generative models have shown promise in generating protein conformations as a more efficient alternative. However, these methods predominantly rely on the diffusion process within a 3D geometric space, which typically centers around the vicinity of metastable states and is often inefficient in terms of runtime. In this paper, we introduce Structure Language Modeling (SLM) as a novel framework for efficient protein conformation generation. Specifically, the protein structures are first encoded into a compact latent space using a discrete variational auto-encoder, followed by conditional language modeling that effectively captures sequence-specific conformation distributions. This enables a more efficient and interpretable exploration of diverse ensemble modes compared to existing methods. Based on this general framework, we instantiate SLM with various popular LM architectures as well as proposing the ESMDiff, a novel BERT-like structure language model fine-tuned from ESM3 with masked diffusion. We verify our approach in various scenarios, including the equilibrium dynamics of BPTI, conformational change pairs, and intrinsically disordered proteins. SLM provides a highly efficient solution, offering a 20-100x speedup than existing methods in generating diverse conformations, shedding light on promising avenues for future research.", "keywords": [ "protein", "conformation generation", "conformation sampling", "generative models", "language models", "diffusion models" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 2, 4, 4 ] }, { "paper_id": "Oz9FTPINRe", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Anej Svete", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Brian DuSell", "gender": "unknown", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Josef Valvoda", "gender": "Male", "institution": "University of Copenhagen", "country": "DK", "position": "Postdoc" }, { "name": "V\u00e9steinn Sn\u00e6bjarnarson", "gender": "Male", "institution": "University of Copenhagen", "country": "DK", "position": "PhD student" }, { "name": "Ryan Cotterell", "gender": "Not Specified", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 32, "n_formula_1": 10, "n_ref_uni": 31, "n_ref": 46, "n_ref_all": 63, "n_fig": 7, "n_tab": 2, "L_tab": 649, "n_element_tab": 24, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2524, "formula_len_all_1": 529, "len_all": 176961, "len_all_1": 70492, "len_abs": 1327, "len_title": 115, "len_sents": 56513, "len_sents_1": 33526, "n_sents": 566, "n_sents_1": 289, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1545, "title": "A Causal Study on The Learnability of Formal Languages", "abs": "Understanding the limitations of neural language models is crucial for knowing what such models are capable of and how they can be used safely. A popular approach to analyzing formal limitations takes the form of training models on formal languages, and studying what aspects of the languages affect model performance. Formal languages can, for instance, be designed using manually constructed grammars or randomly sampled by sampling some type of automata. This provides the researcher with unique control over the features of the language of interest. In this paper, we provide an even more fine-grained approach to targeted model evaluation. We develop a method for controlling specific \\emph{string} features, on the corpus level, in the language of a given automaton. This gives us control over properties such as symbol frequencies while keeping everything else intact, enabling a causal study of their importance. To describe our framework formally, we turn to \\emph{semirings} and introduce finite state automata over a novel---counting---semiring. We devise algorithms that enable string sampling under varying degrees of interventions and demonstrate the utility of our method through several examples showing how targeted interventions over transition, symbol, and state frequencies can be performed. We then train Transformer and LSTM language models on languages under varying degrees of interventions. Our fine-grained analysis allows us to show that different mechanisms influence the learning behavior of these two architectures.", "keywords": [ "Causality", "language models", "formal languages" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "OyyE1FDdrQ", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Han Wang", "gender": "unknown", "institution": "University of Alberta", "country": "", "position": "PhD student" }, { "name": "Haseeb Shah", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "PhD student" }, { "name": "Lingwei Zhu", "gender": "Male", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "JP", "position": "Postdoc" }, { "name": "Martha White", "gender": "Female", "institution": "University of Alberta", "country": "", "position": "Associate Professor" }, { "name": "Yukie Nagai", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" } ], "n_formula": 48, "n_formula_1": 26, "n_ref_uni": 45, "n_ref": 92, "n_ref_all": 140, "n_fig": 12, "n_tab": 5, "L_tab": 3523, "n_element_tab": 345, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 417, "n_element_tab_1": 17, "formula_len_all": 4500, "formula_len_all_1": 2381, "len_all": 180868, "len_all_1": 64324, "len_abs": 976, "len_title": 92, "len_sents": 52092, "len_sents_1": 26382, "n_sents": 530, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 46, "L_abs": 1056, "title": "$q$-exponential family for policy optimization", "abs": "Policy optimization methods benefit from a simple and tractable policy parametrization, usually the Gaussian for continuous action spaces. In this paper, we consider a broader policy family that remains tractable: the $q$-exponential family. \nThis family of policies is flexible, allowing the specification of both heavy-tailed policies ($q>1$) and light-tailed policies ($q<1$). This paper examines the interplay between $q$-exponential policies for several actor-critic algorithms conducted on both online and offline problems. We find that heavy-tailed policies are more effective in general and can consistently improve on Gaussian. \nIn particular, we find the Student's t-distribution to be more stable than the Gaussian across settings and that a heavy-tailed $q$-Gaussian for Tsallis Advantage Weighted Actor-Critic consistently performs well in offline benchmark problems.\nIn summary, we find that the Student's t policy a strong candidate for drop-in replacement to the Gaussian.\nOur code is available at \\url{https://github.com/lingweizhu/qexp}.", "keywords": [ "reinforcement learning", "policy optimization", "$q$-exponential family", "heavy-tailed policies", "sparse policies" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 2, 4, 3 ] }, { "paper_id": "OyjMJjfhiw", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Henry Lam", "gender": "unknown", "institution": "Columbia University", "country": "", "position": "Associate Professor" }, { "name": "Hongseok Namkoong", "gender": "Male", "institution": "LinkedIn", "country": "", "position": "Researcher" }, { "name": "Jiashuo Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yibo Zeng", "gender": "unknown", "institution": "Columbia University", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 3, "n_ref_uni": 45, "n_ref": 86, "n_ref_all": 119, "n_fig": 9, "n_tab": 3, "L_tab": 699, "n_element_tab": 99, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 381, "n_element_tab_1": 52, "formula_len_all": 655, "formula_len_all_1": 139, "len_all": 150822, "len_all_1": 61929, "len_abs": 1185, "len_title": 115, "len_sents": 47699, "len_sents_1": 30387, "n_sents": 343, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 964, "title": "LLM Embeddings Improve Test-Time Adaptation to Tabular $Y|X$-Shifts", "abs": "For tabular datasets, the change in the relationship between the label and covariates ($Y|X$-shifts) is common due to missing variables. Since it is impossible to generalize to a completely new and unknown domain, we study models that are easy to adapt to the target domain even with few labeled examples. We focus on building more informative representations of tabular data that can mitigate $Y|X$-shifts, and propose to leverage the prior world knowledge in LLMs by serializing the tabular data to encode it. We find LLM embeddings alone provide inconsistent improvements in robustness, but models trained on them can be well adapted to the target domain even using 32 labeled observations. Our finding is based on a systematic study consisting of 7650 source-target pairs and benchmark against **261,000** model configurations trained by 20 algorithms. Our observation holds when ablating the size of accessible target data and different adaptation strategies.", "keywords": [ "LLM embeddings", "distribution shifts", "tabular data" ], "rating_list": [ 3, 5, 3, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 3, 3 ], "confidence_list": [ 5, 2, 4, 3, 4 ] }, { "paper_id": "OyWreBlvIE", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Ankur Mallick", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Chi Wang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Daniel Madrigal", "gender": "Male", "institution": "Microsoft", "country": "MX", "position": "Principal Researcher" }, { "name": "Dujian Ding", "gender": "unknown", "institution": "Computing Science, University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Jieyu Zhang", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Menglin Xia", "gender": "unknown", "institution": "Microsoft", "country": "GB", "position": "Researcher" }, { "name": "Mirian Del Carmen Hipolito Garcia", "gender": "Female", "institution": "Microsoft", "country": "MX", "position": "Researcher" }, { "name": "Shaokun Zhang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Qingyun Wu", "gender": "Female", "institution": "Pennsylvania State University", "country": "", "position": "Assistant Professor" }, { "name": "Victor Ruehle", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Principal Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 30, "n_ref": 73, "n_ref_all": 96, "n_fig": 9, "n_tab": 4, "L_tab": 1229, "n_element_tab": 108, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1482, "n_element_tab_1": 137, "formula_len_all": 252, "formula_len_all_1": 236, "len_all": 130324, "len_all_1": 72176, "len_abs": 2219, "len_title": 113, "len_sents": 44536, "len_sents_1": 34999, "n_sents": 326, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1102, "title": "EcoAct: Economic Agent Determines When to Register What Action", "abs": "Recent advancements have enabled Large Language Models (LLMs) to function as agents that can perform actions using external tools. \nThis requires registering, i.e. integrating tool information into the LLM context prior to taking actions.\nCurrent methods indiscriminately incorporates all candidate tools into the agent\u2019s context and retains them across multiple reasoning steps. This process remains opaque to LLM agents and is not integrated into their reasoning procedures, leading to inefficiencies due to increased context length from irrelevant tools. \nTo address this, we introduce EcoAct, a tool-using algorithm that allows LLMs to selectively register tools as needed, optimizing context use. By integrating the tool registration process into the reasoning procedure, EcoAct reduces computational costs by over 50% in multi-step reasoning tasks while maintaining performance, as demonstrated through extensive experiments. Moreover, it can be plugged into any reasoning pipeline with only minor modifications to the prompt, making it universally applicable to LLM agents now and in the future.", "keywords": [ "LLM", "efficiency" ], "rating_list": [ 5, 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "OyAMxlDikl", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 8, "author_info_list": [ { "name": "Anton Tarasenko", "gender": "Male", "institution": "Novosibirsk State University", "country": "RU", "position": "Lecturer" }, { "name": "Ilya Ivanchenko", "gender": "unknown", "institution": "Novosibirsk State University", "country": "", "position": "Researcher" }, { "name": "Ivan Bondarenko", "gender": "Male", "institution": "Novosibirsk State University", "country": "", "position": "Lecturer" }, { "name": "Ivan Plokhikh", "gender": "Male", "institution": "Novosibirsk State University", "country": "", "position": "Researcher" }, { "name": "Kirill Yashunin", "gender": "Male", "institution": "Novosibirsk State University", "country": "", "position": "Researcher" }, { "name": "Nikita Andreyevich Radeev", "gender": "unknown", "institution": "Novosibirsk State University", "country": "", "position": "MS student" }, { "name": "Rustam Mullyadzhanov", "gender": "Male", "institution": "Novosibirsk State University", "country": "", "position": "Associate Professor" }, { "name": "Timofei Prasolov", "gender": "unknown", "institution": "Novosibirsk State University", "country": "RU", "position": "Lecturer" } ], "n_formula": 61, "n_formula_1": 21, "n_ref_uni": 47, "n_ref": 59, "n_ref_all": 71, "n_fig": 3, "n_tab": 4, "L_tab": 363, "n_element_tab": 58, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 306, "n_element_tab_1": 55, "formula_len_all": 5464, "formula_len_all_1": 1257, "len_all": 155374, "len_all_1": 56158, "len_abs": 1305, "len_title": 116, "len_sents": 44094, "len_sents_1": 24951, "n_sents": 380, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1312, "title": "Neural Network Adaptive Quantization based on Bayesian Deep Learning", "abs": "We propose a novel approach to solve the adaptive quantization problem in neural networks based on epistemic uncertainty analysis. The quantized model is treated as a Bayesian neural network with stochastic weights, where the mean values are employed to estimate the corresponding weights. Standard deviations serve as an indicator of uncertainty and the number of corresponding bits \u2014 i.e., a larger number of bits indicate lower uncertainty, and vice versa. We perform an extensive analysis of several algorithms within a novel framework for different convolutional and fully connected neural networks based on open datasets demonstrating the main advantages of the proposed approach. In particular, we introduce two novel algorithms for mixed-precision quantization. Quantile Inform utilizes uncertainty to allocate bit-width across layers, while Random Bits employs stochastic gradient-based optimization techniques to maximize the full likelihood of quantization. Using our approach, we reduce the average bit-width of the VGG-16 model to 3.05 with the 90.5% accuracy on the CIFAR-10 dataset compared to 91.9% for the non-quantized model. For the LeNet model trained on the MNIST dataset, we reduce the average bit-width to 3.16 and achieve 99.0% accuracy, almost equal to 99.2% for the non-quantized model.", "keywords": [ "adaptive quantization", "epistemic uncertainty", "Bayesian neural network" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 2, 4 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "Oy9l6UDWIN", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Abele M\u0103lan", "gender": "unknown", "institution": "Universit\u00e9 de Neuch\u00e2tel", "country": "CH", "position": "PhD student" }, { "name": "Lydia Y. Chen", "gender": "Female", "institution": "Delft University of Technology", "country": "", "position": "Associate Professor" }, { "name": "Robert Birke", "gender": "Male", "institution": "University of Turin", "country": "IT", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 28, "n_ref": 52, "n_ref_all": 61, "n_fig": 4, "n_tab": 6, "L_tab": 1589, "n_element_tab": 186, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1287, "n_element_tab_1": 180, "formula_len_all": 1644, "formula_len_all_1": 1461, "len_all": 122598, "len_all_1": 70336, "len_abs": 1682, "len_title": 113, "len_sents": 35480, "len_sents_1": 31985, "n_sents": 223, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1692, "title": "DLGrapher: Dual Latent Diffusion for Attributed Graph Generation", "abs": "Graphs for applications like social data and financial transactions are particularly complex, with large node counts and high-dimensional features. State-of-the-art diffusion graph synthesizers model the node structure via discrete diffusion and are, unfortunately, limited to small-scale graphs with few to no features. In contrast, continuous diffusion models capture rich node features well, but have issues faithfully modelling connectivity. In this paper, we design DLGrapher, a dual latent diffusion framework for jointly synthesizing large graph structures and high-dimension node features. DLGrapher models node features and structure as a joint latent representation. Structure-wise, we design a reversible coarsening scheme to merge pairs of similar neighboring nodes and their respective edges after encoding node features through a structure-aware variational autoencoder. To capture the dependencies between node features and the graph structure, DLGrapher trains a single diffusion over a dual denoising objective, one for the continuous node representations and another for the discrete edge connectivity. We extensively evaluate DLGrapher's performance on three complex social graph datasets against baselines combining tabular and graph synthesizers. Our solution fares 12.9x better at statistically capturing feature-structure interaction and 25.2% better at downstream tasks thanks to the dual diffusion on average and the latent compressed representation increases throughput by 2.5X. Furthermore, we maintain competitive synthesis quality for simple-featured molecular graphs and structure-only synthetic graphs while drastically reducing computation in the latter case.", "keywords": [ "graph generation", "diffusion model", "attributed graph" ], "rating_list": [ 3, 6, 3, 3, 3 ], "soundness_list": [ 2, 2, 1, 3, 2 ], "presentation_list": [ 2, 3, 2, 1, 2 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "OxxbqZBJxx", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Lily H Zhang", "gender": "Female", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Rajesh Ranganath", "gender": "unknown", "institution": "New York University", "country": "", "position": "Assistant Professor" } ], "n_formula": 49, "n_formula_1": 14, "n_ref_uni": 35, "n_ref": 60, "n_ref_all": 81, "n_fig": 2, "n_tab": 3, "L_tab": 462, "n_element_tab": 45, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 13, "n_element_tab_1": 1, "formula_len_all": 7220, "formula_len_all_1": 1264, "len_all": 157995, "len_all_1": 65420, "len_abs": 1183, "len_title": 106, "len_sents": 46274, "len_sents_1": 30944, "n_sents": 345, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1189, "title": "Win Rate is All that Can Matter from Preference Data Alone", "abs": "The surging interest in learning from preference data has resulted in an elaborate landscape of methods and evaluations. \nThis work offers a framework to simplify this landscape, starting from the underlying sampling distribution for preference data.\nFirst, we show that the only evaluation of a generative model that is grounded in the preference data sampling distribution is win rate. \nGiven that win rate is all that can matter from preference data alone, we relate common preference learning algorithms to direct win rate optimization (DWRO). We outline the theoretical benefits of RLHF as a variant of DWRO; explain why checkpointing is difficult with DPO as a non-DWRO objective; and characterize the limits of SFT on preferred samples with regard to the extent of win rate improvement possible.\nFurthermore, we provide closed-form expressions for the expected win rate improvement of the above objectives, formalizing the role of a model's starting point in the win rate improvement possible. Finally, we conduct an empirical analysis of existing methods and alternative DWRO objectives which suggests that optimization improvements are likely key to advancing preference learning.", "keywords": [ "alignment", "preference learning", "RLHF", "win rate", "language model" ], "rating_list": [ 5, 6, 1, 3 ], "soundness_list": [ 3, 4, 1, 2 ], "presentation_list": [ 1, 3, 1, 1 ], "contribution_list": [ 2, 3, 1, 1 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "OxrDTroSNP", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Tong Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Weikang Su", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Wenzhen Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Zihan Huang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 11, "n_formula_1": 6, "n_ref_uni": 43, "n_ref": 58, "n_ref_all": 72, "n_fig": 8, "n_tab": 1, "L_tab": 390, "n_element_tab": 21, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 590, "n_element_tab_1": 43, "formula_len_all": 405, "formula_len_all_1": 663, "len_all": 165732, "len_all_1": 67796, "len_abs": 1251, "len_title": 153, "len_sents": 53165, "len_sents_1": 32649, "n_sents": 442, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1254, "title": "GenNet: A Generative AI-Driven Mobile Network Simulator for Multi-Objective Network Optimization", "abs": "Simulation-based optimization has emerged as a crucial methodology in the field of mobile network optimization, addressing the need for dynamic and predictive network management. To address the scarcity of open-source mobile network simulators for advanced research, we developed GenNet\u2014a generative AI-driven mobile network simulator. GenNet can create virtual replicas of mobile users, base stations, and wireless environments, utilizing generative AI methods to simulate the behaviors of these entities under various network settings with high accuracy. GenNet features a tailor-made API explicitly designed for reinforcement learning environments, enabling researchers to finely adjust network parameters such as tilts, azimuth, and transmitting power. Extensive experiments have employed GenNet to benchmark multi-objective optimization algorithms, focusing on enhancing network coverage, throughput, and energy efficiency, validating its effectiveness as a robust platform for advancing network optimization techniques. Through this innovative tool, we aim to empower researchers and practitioners to identify and implement the most effective approaches for network optimization, paving the way for future advancements in mobile network management.", "keywords": [ "Mobile Networks; Simulator; Optimization" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "Oxpkn0YLG1", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Aliaksandr Siarohin", "gender": "Male", "institution": "Snap Inc.", "country": "IT", "position": "Researcher" }, { "name": "Chaoyang Wang", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Hsin-Ying Lee", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Jiaxu Zou", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Michael Vasilkovsky", "gender": "Not Specified", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Peiye Zhuang", "gender": "Female", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Sergei Korolev", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Machine Learning Engineer" }, { "name": "Sergey Tulyakov", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Director of Research" }, { "name": "Songfang Han", "gender": "Female", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Vladislav Shakhrai", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Machine Learning Engineer" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 25, "n_ref": 109, "n_ref_all": 136, "n_fig": 14, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 846, "n_element_tab_1": 69, "formula_len_all": 204, "formula_len_all_1": 155, "len_all": 111820, "len_all_1": 52385, "len_abs": 492, "len_title": 134, "len_sents": 28054, "len_sents_1": 23548, "n_sents": 251, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1571, "title": "GTR: Improving Large 3D Reconstruction Models through Geometry and Texture Refinement", "abs": "We propose a novel approach for 3D mesh reconstruction from multi-view images. We improve upon the large reconstruction model LRM that use a transformer-based triplane generator and a Neural Radiance Field (NeRF) model trained on multi-view images. We introduce three key components to significantly enhance the 3D reconstruction quality. First of all, we examine the original LRM architecture and find several shortcomings. Subsequently, we introduce respective modifications to the LRM architecture, which lead to improved multi-view image representation and more computationally efficient training. Second, in order to improve geometry reconstruction and enable supervision at full image resolution, we extract meshes from the NeRF in a differentiable manner and fine-tune the NeRF model through mesh rendering. These modifications allow us to achieve state-of-the-art performance on both 2D and 3D evaluation metrics on Google Scanned Objects (GSO) dataset and OmniObject3D dataset. Finally, we introduce a lightweight per-instance texture refinement procedure to better reconstruct complex textures, such as text and portraits on assets. To address this, we introduce a lightweight per-instance texture refinement procedure. This procedure fine-tunes the triplane representation and the NeRF's color estimation model on the mesh surface using the input multi-view images in just 4 seconds. This refinement achieves faithful reconstruction of complex textures. Additionally, our approach enables various downstream applications, including text/image-to-3D generation.", "keywords": [ "3D Reconstruction", "Feed-forward" ], "rating_list": [ 6, 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 1, 2 ], "confidence_list": [ 4, 4, 3, 5, 5 ] }, { "paper_id": "OxKi02I29I", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Kumara Kahatapitiya", "gender": "Male", "institution": "Meta", "country": "US", "position": "Intern" }, { "name": "Michael S Ryoo", "gender": "Male", "institution": "Salesforce AI Research", "country": "US", "position": "Principal Researcher" }, { "name": "Xiang Li", "gender": "unknown", "institution": "State University of New York, Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Kanchana Nisal Ranasinghe", "gender": "Male", "institution": "State University of New York, Stony Brook", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 42, "n_ref": 138, "n_ref_all": 176, "n_fig": 4, "n_tab": 12, "L_tab": 4580, "n_element_tab": 735, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1995, "n_element_tab_1": 377, "formula_len_all": 289, "formula_len_all_1": 51, "len_all": 225899, "len_all_1": 65548, "len_abs": 1218, "len_title": 105, "len_sents": 56424, "len_sents_1": 28396, "n_sents": 408, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 896, "title": "Understanding Long Videos with Multimodal Language Models", "abs": "Large Language Models (LLMs) have allowed recent LLM-based approaches to achieve excellent performance on long-video understanding benchmarks. We investigate how extensive world knowledge and strong reasoning skills of LLMs influence evaluations on standard long video benchmarks. Surprisingly, we discover that LLM-based models yield surprisingly good accuracy on long-video tasks with limited video information, sometimes even with no video specific information. \nBuilding on this, we inject video-specific object-centric information extracted from off-the-shelf pre-trained models into an LLM-based setup. We utilize natural language as a medium for fusing this information. Our resulting Multimodal Video Understanding (MVU) framework demonstrates state-of-the-art performance across long-video understanding benchmarks as well as on robotics domain tasks. Our code will be released publicly.", "keywords": [ "long-video", "visual question answering", "interpretability" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "Ox4AJ2Vurb", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Bernhard Egger", "gender": "Male", "institution": "Friedrich-Alexander-Universit\u00e4t Erlangen-N\u00fcrnberg", "country": "DE", "position": "Professor" }, { "name": "Maximilian Weiherer", "gender": "Male", "institution": "Friedrich-Alexander Universit\u00e4t Erlangen-N\u00fcrnberg", "country": "", "position": "PhD student" } ], "n_formula": 42, "n_formula_1": 19, "n_ref_uni": 34, "n_ref": 69, "n_ref_all": 108, "n_fig": 9, "n_tab": 8, "L_tab": 1710, "n_element_tab": 369, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 2225, "n_element_tab_1": 171, "formula_len_all": 3830, "formula_len_all_1": 1126, "len_all": 145395, "len_all_1": 62832, "len_abs": 1193, "len_title": 107, "len_sents": 46211, "len_sents_1": 27001, "n_sents": 357, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1198, "title": "Mat\u00e9rn Kernels for Tunable Implicit Surface Reconstruction", "abs": "We propose to use the family of Mat\u00e9rn kernels for tunable implicit surface reconstruction, building upon the recent success of kernel methods for 3D reconstruction of oriented point clouds.\nAs we show, both, from a theoretical and practical perspective, Mat\u00e9rn kernels have some appealing properties which make them particularly well suited for surface reconstruction---outperforming state-of-the-art methods based on the arc-cosine kernel while being significantly easier to implement, faster to compute, and scaleable.\nBeing stationary, we demonstrate that the Mat\u00e9rn kernels' spectrum can be tuned in the same fashion as Fourier feature mappings help coordinate-based MLPs to overcome spectral bias. \nMoreover, we theoretically analyze Mat\u00e9rn kernel's connection to SIREN networks as well as its relation to previously employed arc-cosine kernels. \nFinally, based on recently introduced Neural Kernel Fields, we present data-dependent Mat\u00e9rn kernels and conclude that especially the Laplace kernel (being part of the Mat\u00e9rn family) is extremely competitive, performing almost on par with state-of-the-art methods in the noise-free case while having a more than five times shorter training time.", "keywords": [ "surface reconstruction", "kernel methods", "neural tangent kernel" ], "rating_list": [ 6, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "Ox2A1WoKLm", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Aakash Sen Sharma", "gender": "Male", "institution": "SagepilotAI", "country": "IN", "position": "Machine Learning Engineer" }, { "name": "Ankur Mali", "gender": "Male", "institution": "University of South Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Mohan Shankara Kankanhalli", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Murari Mandal", "gender": "Male", "institution": "Kalinga Institute of Industrial Technology (KIIT) Bhubaneswar India", "country": "IN", "position": "Assistant Professor" }, { "name": "Umakanta Maharana", "gender": "Male", "institution": "Kalinga Institute of Industrial Technology (KIIT)", "country": "IN", "position": "Researcher" }, { "name": "Yash Sinha", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 3, "n_ref_uni": 16, "n_ref": 42, "n_ref_all": 57, "n_fig": 5, "n_tab": 3, "L_tab": 848, "n_element_tab": 65, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 848, "n_element_tab_1": 65, "formula_len_all": 667, "formula_len_all_1": 175, "len_all": 93913, "len_all_1": 55070, "len_abs": 3028, "len_title": 148, "len_sents": 31336, "len_sents_1": 26185, "n_sents": 252, "n_sents_1": 195, "venue": "ICLR2025", "decision": "withdraw", "L_title": 99, "L_abs": 1317, "title": "Towards Robust Concept Erasure in Diffusion Models: Unlearning Identity, Nudity and Artistic Styles", "abs": "Diffusion models have achieved remarkable success in generative tasks across various domains. However, the increasing demand for content moderation and the removal of specific concepts from these models has introduced the challenge of \\textit{unlearning}. In this work, we present a suite of robust methodologies that significantly enhance the unlearning process by employing advanced loss functions within knowledge distillation frameworks. Specifically, we utilize the Cramer-Wold distance and Jensen-Shannon (JS) divergence to facilitate more efficient and versatile concept removal. Although current non-learning techniques are effective in certain scenarios, they are typically limited to specific categories such as identity, nudity, or artistic style. In contrast, our proposed methods demonstrate robust versatility, seamlessly adapting to and performing effectively across a wide range of concept erasure categories. Our approach outperforms existing techniques, achieving consistent results across different unlearning categories and showcasing its broad applicability. Through extensive experiments, we show that our method not only surpasses previous benchmarks but also addresses key limitations of current unlearning techniques, paving the way for more responsible use of text-to-image diffusion models.", "keywords": [ "Diffusion Model", "machine unlearning", "erasing concepts" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "OwtkY3LyvC", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Dechen Zhang", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Haoyang Zhou", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Lei Huang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Associate Professor" }, { "name": "Yimi Wang", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Undergrad student" }, { "name": "\u6731\u745e\u6e58", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" } ], "n_formula": 31, "n_formula_1": 21, "n_ref_uni": 15, "n_ref": 30, "n_ref_all": 52, "n_fig": 4, "n_tab": 7, "L_tab": 2815, "n_element_tab": 362, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 768, "n_element_tab_1": 138, "formula_len_all": 1822, "formula_len_all_1": 1157, "len_all": 126057, "len_all_1": 64922, "len_abs": 1454, "len_title": 119, "len_sents": 38127, "len_sents_1": 28118, "n_sents": 357, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1456, "title": "Bridging The Gap Between Training and Testing for Certified Robustness", "abs": "Certified robustness provides a theoretical lower bound for adversarial robustness and arouses widespread interests and discussions from the research community. With theoretical support to improve the certified robustness on the training set, practitioners endeavor to train a more certified robust model during inference on the test set. However, the experimental neglect on the training set and the theoretical ignorance during inference on the test set induce a gap between training and testing for certified robustness. By establishing an equivalence between the convergence of training loss and the improvement of certified robustness, we recognize there is a trade-off between expressive power and generalization (assuming a well-conditioned optimization) for certified robustness, which is similar to the underfitting and overfitting discussed in machine learning. To investigate this trade-off, we design a new orthogonal convolution-Controllable Orthogonal Convolution Kernel (COCK) which provides a wider range of expressive power than existing orthogonal convolutions. Empirically, there is a power-driven shift from vanilla classification accuracy to certified robustness in the sense of the optimal trade-off between expressive power and generalization. The experimental results suggest that by carefully improving the expressive power from the optimal trade-off for vanilla classification performance, the model will be more certified robust.", "keywords": [ "certified robustness", "orthogonal convolution", "expressive power", "generalization" ], "rating_list": [ 3, 5, 1, 5, 3 ], "soundness_list": [ 2, 3, 1, 3, 2 ], "presentation_list": [ 2, 3, 1, 2, 1 ], "contribution_list": [ 2, 3, 1, 2, 1 ], "confidence_list": [ 4, 3, 4, 2, 3 ] }, { "paper_id": "OwpLQrpdwE", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Daning Huang", "gender": "unknown", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Hanyang He", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "John Harlim", "gender": "unknown", "institution": "Pennsylvania State University", "country": "US", "position": "Full Professor" }, { "name": "Yan Li", "gender": "Female", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 120, "n_formula_1": 29, "n_ref_uni": 41, "n_ref": 68, "n_ref_all": 152, "n_fig": 20, "n_tab": 4, "L_tab": 904, "n_element_tab": 126, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9932, "formula_len_all_1": 3093, "len_all": 208282, "len_all_1": 69237, "len_abs": 1258, "len_title": 96, "len_sents": 67240, "len_sents_1": 31045, "n_sents": 591, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 116, "L_abs": 1264, "title": "Learning vector fields of differential equations on manifolds with geometrically constrained operator-valued kernels", "abs": "We address the problem of learning ordinary differential equations (ODEs) on manifolds. Existing machine learning methods, particularly those using neural networks, often struggle with high computational demands. To overcome this issue, we introduce a geometrically constrained operator-valued kernel that allows us to represent vector fields on tangent bundles of smooth manifolds. The construction of the kernel imposes the geometric constraints that are estimated from the data and ensures the computational feasibility for learning high dimensional systems of ODEs. Once the vector fields are estimated, e.g., by the kernel ridge regression, we need an ODE solver that guarantees the solution to stay on (or close to) the manifold. To overcome this issue, we propose a geometry-preserving ODE solver that approximates the exponential maps corresponding to the ODE solutions. We deduce a theoretical error bound for the proposed solver that guarantees the approximate solutions to lie on the manifold in the limit of large data. We verify the effectiveness of the proposed approach on high-dimensional dynamical systems, including the cavity flow problem, the beating and travelling waves in Kuramoto-Sivashinsky equations, and the reaction-diffusion dynamics.", "keywords": [ "Dynamics on manifolds", "Operator-valued kernel", "Geometry-preserving time integration", "Ordinary differential equations" ], "rating_list": [ 8, 8, 5, 6, 10 ], "soundness_list": [ 4, 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 4, 3, 2, 2, 4 ], "confidence_list": [ 3, 3, 5, 3, 4 ] }, { "paper_id": "OwntMF6McA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Aounon Kumar", "gender": "Male", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Lillian Sun", "gender": "Female", "institution": "Harvard University", "country": "US", "position": "Undergrad student" }, { "name": "Martin Pawelczyk", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Postdoc" }, { "name": "Zhenting Qi", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "MS student" }, { "name": "Hima Lakkaraju", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 41, "n_ref": 55, "n_ref_all": 83, "n_fig": 24, "n_tab": 1, "L_tab": 232, "n_element_tab": 18, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 214, "n_element_tab_1": 18, "formula_len_all": 1119, "formula_len_all_1": 520, "len_all": 157662, "len_all_1": 67481, "len_abs": 1964, "len_title": 79, "len_sents": 53045, "len_sents_1": 33837, "n_sents": 385, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1936, "title": "Weak-to-Strong Trustworthiness: Eliciting Trustworthiness with Weak Supervision", "abs": "The rapid proliferation of generative AI, especially large language models (LLMs), has led to their integration into a variety of applications. \nA key phenomenon known as weak-to-strong generalization - where a strong model trained on a weak model's outputs surpasses the weak model in task performance - has gained significant attention. Yet, whether critical trustworthiness properties such as robustness, fairness, and privacy can generalize similarly remains an open question. In this work, we study this question by examining if a stronger model can inherit trustworthiness properties when fine-tuned on a weaker model\u2019s outputs, a process we term weak-to-strong trustworthiness generalization. Specifically, we examine whether a strong model can inherit or even enhance trustworthiness attributes when fine-tuned on a weak model's outputs. To address this, we introduce two foundational training strategies: 1) Weak Trustworthiness Finetuning (Weak TFT), which leverages trustworthiness regularization during the fine-tuning of the weak model, and 2) Weak and Weak-to-Strong Trustworthiness Finetuning (Weak+WTS TFT), which extends regularization to both weak and strong models. Our experimental evaluation on real-world datasets (Adult, OOD Style Transfer, AdvGLUE++, and Enron Emails) reveals that while some trustworthiness properties, such as fairness, adversarial, and OOD robustness, show significant improvement in transfer when both models were regularized, others like privacy do not exhibit signs of weak-to-strong trustworthiness. As the first study to explore trustworthiness generalization via weak-to-strong generalization, our work provides valuable insights into the potential and limitations of this method. Our findings highlight the importance of systematically studying trustworthiness transfer to develop AI systems that are not only accurate but also ethically aligned and reliable in critical applications.", "keywords": [ "Weak-to-strong learning", "Fairness", "Adversarial Robustness", "OOD Robustness", "Privacy" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "OwNoTs2r8e", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Ananth Grama", "gender": "Male", "institution": "Purdue University", "country": "", "position": "Full Professor" }, { "name": "Changlong Wu", "gender": "Male", "institution": "Purdue University", "country": "", "position": "Postdoc" }, { "name": "Wojciech Szpankowski", "gender": "Male", "institution": " Purdue University", "country": "US", "position": "Full Professor" } ], "n_formula": 34, "n_formula_1": 22, "n_ref_uni": 16, "n_ref": 24, "n_ref_all": 33, "n_fig": 1, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2570, "formula_len_all_1": 1491, "len_all": 128609, "len_all_1": 88970, "len_abs": 1246, "len_title": 130, "len_sents": 42590, "len_sents_1": 34889, "n_sents": 376, "n_sents_1": 285, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1258, "title": "No Free Lunch: Fundamental Limits of Learning Non-Hallucinating Generative Models", "abs": "Generative models have shown impressive capabilities in synthesizing high-quality outputs across various domains. However, a persistent challenge is the occurrence of \"hallucinations,\" where the model produces outputs that are not grounded in the underlying facts. While empirical strategies have been explored to mitigate this issue, a rigorous theoretical understanding remains elusive. In this paper, we develop a theoretical framework to analyze the *learnability* of non-hallucinating generative models from a learning-theoretic perspective. Our results reveal that non-hallucinating learning is statistically *impossible* when relying solely on the training dataset, even for a hypothesis class of size two and when the entire training set is truthful. To overcome these limitations, we show that incorporating *inductive biases* aligned with the actual facts into the learning process is essential. We provide a systematic approach to achieve this by restricting the fact set to a concept class of finite VC-dimension and demonstrate its effectiveness under various learning paradigms. Although our findings are primarily conceptual, they represent a first step towards a principled approach to addressing hallucinations in learning generative models.", "keywords": [ "generative models", "hallucination", "no-free-lunch theorem", "distribution PAC learning", "VC-dimension" ], "rating_list": [ 5, 8, 8, 8, 1 ], "soundness_list": [ 3, 3, 3, 3, 1 ], "presentation_list": [ 2, 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 3, 3, 1 ], "confidence_list": [ 2, 2, 3, 2, 4 ] }, { "paper_id": "OvrmA3GMiX", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Akhil Bagaria", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Bingnan Huo", "gender": "Male", "institution": "Brown University", "country": "US", "position": "MS student" }, { "name": "Cameron Allen", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "George Konidaris", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Assistant Professor" }, { "name": "Zhiyuan Zhou", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Anita Taosheng De Mello Koch", "gender": "Female", "institution": "Brown University", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 31, "n_ref": 31, "n_ref_all": 43, "n_fig": 9, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 78, "formula_len_all_1": 78, "len_all": 99290, "len_all_1": 64895, "len_abs": 1198, "len_title": 119, "len_sents": 35177, "len_sents_1": 32928, "n_sents": 268, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1204, "title": "Learning Transferable Sub-goals by Hypothesizing Generalizing Features", "abs": "Transfer is a key promise of hierarchical reinforcement learning, but requires first learning transferable skills.\nFor an agent to effectively transfer a skill it must identify features that generalize and define the skill over this subset.\nHowever, this task is under-specified from a single context as the agent has no prior knowledge of what future tasks may be introduced.\nSince successful transfer requires a skill to reliably achieve a sub-goal from different states, we focus our attention on ensuring sub-goals are represented in a transferable way. \nFor each sub-goal, we train an ensemble of classifiers while explicitly incentivizing them to use minimally overlapping features.\nEach ensemble member represents a unique hypothesis about the transferable features of a sub-goal that the agent can use to learn a skill in previously unseen portions of the environment.\nEnvironment reward then determines which hypothesis is most transferable for the given task, based on the intuition that useful sub-goals lead to better reward maximization.\nWe apply these reusable sub-goals to MiniGrid and Montezuma's Revenge, allowing us to learn previously defined skills in unseen parts of the state-space.", "keywords": [ "hierarchical reinforcement learning" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 4, 2, 3, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "OvoCm1gGhN", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Furu Wei", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Partner Research Manager" }, { "name": "Gao Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Li Dong", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Principal Researcher" }, { "name": "Tianzhu Ye", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yi Zhu", "gender": "Male", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Yuqing Xia", "gender": "unknown", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yutao Sun", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 38, "n_ref": 67, "n_ref_all": 94, "n_fig": 11, "n_tab": 11, "L_tab": 1532, "n_element_tab": 222, "n_fig_1": 11, "n_tab_1": 6, "L_tab_1": 1473, "n_element_tab_1": 73, "formula_len_all": 2850, "formula_len_all_1": 1241, "len_all": 137900, "len_all_1": 57032, "len_abs": 1208, "len_title": 0, "len_sents": 34333, "len_sents_1": 23374, "n_sents": 349, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 24, "L_abs": 1207, "title": "Differential Transformer", "abs": "Transformer tends to overallocate attention to irrelevant context. In this work, we introduce Diff Transformer, which amplifies attention to the relevant context while canceling noise. Specifically, the differential attention mechanism calculates attention scores as the difference between two separate softmax attention maps. The subtraction cancels noise, promoting the emergence of sparse attention patterns. Experimental results on language modeling show that Diff Transformer outperforms Transformer in various settings of scaling up model size and training tokens. More intriguingly, it offers notable advantages in practical applications, such as long-context modeling, key information retrieval, hallucination mitigation, in-context learning, and reduction of activation outliers. By being less distracted by irrelevant context, Diff Transformer can mitigate hallucination in question answering and text summarization. For in-context learning, Diff Transformer not only enhances accuracy but is also more robust to order permutation, which was considered as a chronic robustness issue. The results position Diff Transformer as a highly effective and promising architecture for large language models.", "keywords": [ "sequence modeling", "language models", "model architecture", "Transformer" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "OvU9u6wS2J", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Roberto Colomboni", "gender": "Male", "institution": "Polytechnic University of Milan (POLIMI)", "country": "IT", "position": "Postdoc" }, { "name": "Tommaso R. Cesari", "gender": "Male", "institution": "University of Ottawa", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 59, "n_formula_1": 30, "n_ref_uni": 18, "n_ref": 54, "n_ref_all": 62, "n_fig": 1, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8427, "formula_len_all_1": 2712, "len_all": 121633, "len_all_1": 66308, "len_abs": 1885, "len_title": 104, "len_sents": 40630, "len_sents_1": 30412, "n_sents": 309, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1920, "title": "An Online Learning Theory of Trading-Volume Maximization", "abs": "We explore brokerage between traders in an online learning framework.\nAt any round $t$, two traders meet to exchange an asset, provided the exchange is mutually beneficial.\nThe broker proposes a trading price, and each trader tries to sell their asset or buy the asset from the other party, depending on whether the price is higher or lower than their private valuations.\nA trade happens if one trader is willing to sell and the other is willing to buy at the proposed price.\nPrevious work provided guidance to a broker aiming at enhancing traders' total earnings by maximizing the *gain from trade*, defined as the sum of the traders' net utilities after each interaction.\nThis classical notion of reward can be highly unfair to traders with small profit margins, and far from the real-life utility of the broker.\nFor these reasons, we investigate how the broker should behave to maximize the trading volume, i.e., the *total number of trades*.\nWe model the traders' valuations as an i.i.d. process with an unknown distribution.\nIf the traders' valuations are revealed after each interaction (full-feedback), and the traders' valuations cumulative distribution function (cdf) is continuous, we provide an algorithm achieving logarithmic regret and show its optimality up to constants.\nIf only their willingness to sell or buy at the proposed price is revealed after each interaction ($2$-bit feedback), we provide an algorithm achieving poly-logarithmic regret when the traders' valuations cdf is Lipschitz and show its near-optimality.\nWe complement our results by analyzing the implications of dropping the regularity assumptions on the unknown traders' valuations cdf. \nIf we drop the continuous cdf assumption, the regret rate degrades to $\\Theta(\\sqrt{T})$ in the full-feedback case, where $T$ is the time horizon. \nIf we drop the Lipschitz cdf assumption, learning becomes impossible in the $2$-bit feedback case.", "keywords": [ "regret minimization", "bilateral trade", "online learning", "fairness", "theory" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 4, 4, 4, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "OuxdVB6g1F", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Bo Pan", "gender": "unknown", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Liang Zhao", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Associate Professor" }, { "name": "Yuntong Hu", "gender": "Male", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Zheng Zhang", "gender": "Male", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Chen LING", "gender": "Male", "institution": "Emory University", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 32, "n_ref": 77, "n_ref_all": 95, "n_fig": 4, "n_tab": 7, "L_tab": 10727, "n_element_tab": 423, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 3515, "n_element_tab_1": 245, "formula_len_all": 937, "formula_len_all_1": 946, "len_all": 156321, "len_all_1": 63693, "len_abs": 1234, "len_title": 153, "len_sents": 50042, "len_sents_1": 28768, "n_sents": 310, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1238, "title": "TAGA: Text-Attributed Graph Self-Supervised Learning by Synergizing Graph and Text Mutual Transformations", "abs": "Text-Attributed Graphs (TAGs) enhance graph structures with natural language descriptions, enabling detailed representation of data and their relationships across a broad spectrum of real-world scenarios. Despite the potential for deeper insights, existing TAG representation learning primarily omit the semantic relationship among node texts, and mostly relies on supervised methods, necessitating extensive labeled data and limiting applicability across diverse contexts. This paper introduces a new self-supervised learning framework, Text-And-Graph Multi-View Alignment (TAGA), which overcomes these constraints by integrating TAGs' structural and semantic dimensions. TAGA constructs two complementary views: Text-of-Graph view, which organizes node texts into structured documents based on graph topology, and the Graph-of-Text view, which converts textual nodes and connections into graph data. By aligning representations from both views, TAGA captures joint textual and structural information. In addition, a novel structure-preserving random walk algorithm is proposed for efficient training on large-sized TAGs. Our framework demonstrates strong performance in zero-shot and few-shot scenarios across eight real-world datasets.", "keywords": [ "Text-attributed graphs", "self-supervised learning", "representation learning" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "Ouu3HnIVBc", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Chaochao Lu", "gender": "unknown", "institution": "Shanghai AI Laboratory ", "country": "CN", "position": "Research Scientist" }, { "name": "Shu Yu", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 108, "n_ref_all": 140, "n_fig": 17, "n_tab": 4, "L_tab": 737, "n_element_tab": 39, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 433, "n_element_tab_1": 22, "formula_len_all": 27, "formula_len_all_1": 0, "len_all": 181650, "len_all_1": 63072, "len_abs": 1598, "len_title": 105, "len_sents": 41287, "len_sents_1": 27940, "n_sents": 360, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1602, "title": "ADAM: An Embodied Causal Agent in Open-World Environments", "abs": "In open-world environments like Minecraft, existing agents face challenges in continuously learning structured knowledge, particularly causality. These challenges stem from the opacity inherent in black-box models and an excessive reliance on prior knowledge during training, which impair their interpretability and generalization capability. To this end, we introduce ADAM, An emboDied causal Agent in Minecraft, that can autonomously navigate the open world, perceive multimodal contexts, learn causal world knowledge, and tackle complex tasks through lifelong learning. ADAM is empowered by four key components: 1) an interaction module, enabling the agent to execute actions while documenting the interaction processes; 2) a causal model module, tasked with constructing an ever-growing causal graph from scratch, which enhances interpretability and diminishes reliance on prior knowledge; 3) a controller module, comprising a planner, an actor, and a memory pool, which uses the learned causal graph to accomplish tasks; 4) a perception module, powered by multimodal large language models, which enables ADAM to perceive like a human player. Extensive experiments show that ADAM constructs an almost perfect causal graph from scratch, enabling efficient task decomposition and execution with strong interpretability. Notably, in our modified Minecraft games where no prior knowledge is available, ADAM maintains its performance and shows remarkable robustness and generalization capability. ADAM pioneers a novel paradigm that integrates causal methods and embodied agents in a synergistic manner.", "keywords": [ "embodied agent", "causality", "large language model", "interpretability", "vision language navigation", "cross-modal application", "cross-modal information extraction", "multimodality" ], "rating_list": [ 6, 6, 5, 1 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "OujTnpmAZG", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "Bojun Cheng", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Changchun Feng", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Haotian FU", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Hong Xing", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Hongwei Ren", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Xiaopeng LIN", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yue Zhou", "gender": "Female", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yulong Huang", "gender": "Male", "institution": "Hong Kong University of Science and Technology (GZ)", "country": "CN", "position": "PhD student" }, { "name": "Zunchang LIU", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" } ], "n_formula": 85, "n_formula_1": 25, "n_ref_uni": 33, "n_ref": 63, "n_ref_all": 115, "n_fig": 10, "n_tab": 18, "L_tab": 3493, "n_element_tab": 435, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1759, "n_element_tab_1": 135, "formula_len_all": 6375, "formula_len_all_1": 2141, "len_all": 191599, "len_all_1": 66755, "len_abs": 1637, "len_title": 141, "len_sents": 54429, "len_sents_1": 27783, "n_sents": 491, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1600, "title": "PRF: Parallel Resonate and Fire Neuron for Long Sequence Learning in Spiking Neural Networks", "abs": "Recently, there is growing demand for effective and efficient long sequence modeling, with State Space Models (SSMs) proving to be effective for long sequence tasks. To further reduce energy consumption, SSMs can be adapted to Spiking Neural Networks (SNNs) using spiking functions. However, current spiking-formalized SSMs approaches still rely on float-point matrix-vector multiplication during inference, undermining SNNs\u2019 energy advantage. In this work, we address the efficiency and performance challenges of long sequence learning in SNNs simultaneously. First, we propose a decoupled reset method for parallel spiking neuron training, reducing the typical Leaky Integrate-and-Fire (LIF) model\u2019s training time from $O(L^2)$ to $O(L\\log L)$, effectively speeding up the training by $6.57 \\times$ to $16.50 \\times$ on sequence lengths $1,024$ to $32,768$. To our best knowledge, this is the first time that parallel computation with a reset mechanism is implemented achieving equivalence to its sequential counterpart. Secondly, to capture long-range dependencies, we propose a Parallel Resonate and Fire (PRF) neuron, which leverages an oscillating membrane potential driven by a resonate mechanism from a differentiable reset function in the complex domain. The PRF enables efficient long sequence learning while maintaining parallel training. Finally, we demonstrate that the proposed spike-driven architecture using PRF achieves performance comparable to Structured SSMs (S4), with two orders of magnitude reduction in energy consumption, outperforming transformers on Long Range Arena tasks.", "keywords": [ "Spiking Neural Networks", "Neuromorphic Computing", "Neuron" ], "rating_list": [ 3, 5, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "OuYCW4ACDl", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Minghui Zhu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Full Professor" }, { "name": "Shicheng Liu", "gender": "unknown", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Yue Mao", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 83, "n_formula_1": 22, "n_ref_uni": 38, "n_ref": 82, "n_ref_all": 103, "n_fig": 3, "n_tab": 3, "L_tab": 384, "n_element_tab": 31, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 17158, "formula_len_all_1": 2320, "len_all": 197551, "len_all_1": 77982, "len_abs": 604, "len_title": 128, "len_sents": 55216, "len_sents_1": 32531, "n_sents": 514, "n_sents_1": 289, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 604, "title": "Learn from Interactions: General-Sum Interactive Inverse Reinforcement Learning", "abs": "This paper studies the problem that a learner aims to learn the reward function of the expert from the interaction with the expert and how to interact with the expert. We formulate the problem as a stochastic bi-level optimization problem and develop a double-loop algorithm \"general-sum interactive inverse reinforcement learning\" (GSIIRL). In the GSIIRL, the learner first learns the reward function of the expert in the inner loop and then learns how to interact with the expert in the outer loop. We theoretically prove the convergence of our algorithm and validate our algorithm through simulations.", "keywords": [ "Inverse Reinforcement Learning" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "OuUKXhV2Uz", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Feng Wang", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Qi Wu", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "Associate Professor" }, { "name": "Weiren zhao", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Intern" }, { "name": "Yutong Xie", "gender": "Not Specified", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Yuyin Zhou", "gender": "unknown", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 27, "n_ref": 55, "n_ref_all": 73, "n_fig": 7, "n_tab": 5, "L_tab": 1282, "n_element_tab": 193, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1282, "n_element_tab_1": 193, "formula_len_all": 450, "formula_len_all_1": 353, "len_all": 103164, "len_all_1": 54391, "len_abs": 1800, "len_title": 133, "len_sents": 27949, "len_sents_1": 23775, "n_sents": 178, "n_sents_1": 145, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1805, "title": "UD-Mamba: A pixel-level uncertainty-driven mamba model for medical image segmentation", "abs": "Recent advancements have highlighted the Mamba framework, a state-space models (SSMs) known for its efficiency in capturing long-range dependencies with linear computational complexity. While Mamba has shown competitive performance in medical image segmentation, it encounters difficulties in modeling local features due to the sporadic nature of traditional location-based scanning methods and the complex, ambiguous boundaries often present in medical images. To overcome these challenges, we propose Uncertainty-Driven Mamba (UD-Mamba), which redefines the pixel-order scanning process by incorporating channel uncertainty into the scanning mechanism. UD-Mamba introduces two key scanning techniques: sequential scanning, which prioritizes regions with high uncertainty by scanning in a row-by-row fashion, and skip scanning, which processes columns vertically, moving from high-to-low or low-to-high uncertainty at fixed intervals. Sequential scanning efficiently clusters high-uncertainty regions, such as boundaries and foreground objects, to improve segmentation precision, while skip scanning enhances the interaction between background and foreground regions, allowing for timely integration of background information to support more accurate foreground inference. Recognizing the advantages of scanning from certain to uncertain areas, we introduce four learnable parameters to balance the importance of features extracted from different scanning methods. Additionally, a cosine consistency loss is employed to mitigate the drawbacks of transitioning between uncertain and certain regions during the scanning process. Our method demonstrates robust segmentation performance, validated across three distinct medical imaging datasets involving pathology, dermatological lesions, and cardiac tasks.", "keywords": [ "Medical Image Segmentation", "State Space Models", "Mamba", "Uncertainty" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "OuNIWgGGif", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Donald Kougang-Yombi", "gender": "Male", "institution": "African Institute for Mathematical Science, AIMS RWANDA", "country": "RW", "position": "PhD student" }, { "name": "Elisabetta Cornacchia", "gender": "Female", "institution": "INRIA", "country": "FR", "position": "Postdoc" }, { "name": "Emmanuel Abbe", "gender": "unknown", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "", "position": "Full Professor" }, { "name": "Jan H\u0105z\u0142a", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "", "position": "Postdoc" } ], "n_formula": 305, "n_formula_1": 12, "n_ref_uni": 25, "n_ref": 55, "n_ref_all": 82, "n_fig": 8, "n_tab": 4, "L_tab": 480, "n_element_tab": 20, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 25996, "formula_len_all_1": 506, "len_all": 273525, "len_all_1": 69358, "len_abs": 4555, "len_title": 117, "len_sents": 77052, "len_sents_1": 31468, "n_sents": 862, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1209, "title": "Learning High-Degree Parities: The Crucial Role of the Initialization", "abs": "Parities have become a standard benchmark for evaluating learning algorithms. Recent works show that regular neural networks trained by gradient descent can efficiently learn degree $k$ parities on uniform inputs for constant $k$, but fail to do so when $k$ and $d-k$ grow with $d$ (here $d$ is the ambient dimension). However, the case where $k=d-O_d(1)$, including the degree $d$ parity (the \\textit{full parity}), has remained unsettled. This paper shows that for gradient descent on regular neural networks, learnability depends on the initial weight distribution. On one hand, the discrete Rademacher initialization enables efficient learning of full parities, while on the other hand, its Gaussian perturbation with large enough constant standard deviation $\\sigma$ prevents it. The positive result for full parity is shown to hold up to $\\sigma=O(d^{-1})$, pointing to questions about a sharper threshold phenomenon. Unlike statistical query (SQ) learning, where a singleton function class like the full parity is trivially learnable, our negative result applies to a fixed function and relies on an \\emph{initial gradient alignment} measure of potential broader relevance to neural networks learning.", "keywords": [ "initialization", "neural networks", "gradient descent", "parity functions", "complexity", "initial alignment" ], "rating_list": [ 8, 8, 6, 6, 6 ], "soundness_list": [ 3, 4, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 4, 3 ] }, { "paper_id": "OuLgaHEmzi", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Haiyang Xu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Lei Feng", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" }, { "name": "Ming Yan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Instructor" }, { "name": "Shengjie Zhou", "gender": "Male", "institution": "Chongqing University", "country": "CN", "position": "MS student" }, { "name": "Tao Xiang", "gender": "Male", "institution": "Chongqing University", "country": "CN", "position": "Full Professor" }, { "name": "Xin Cheng", "gender": "Male", "institution": "Chongqing University", "country": "", "position": "MS student" } ], "n_formula": 31, "n_formula_1": 22, "n_ref_uni": 43, "n_ref": 73, "n_ref_all": 103, "n_fig": 5, "n_tab": 18, "L_tab": 7528, "n_element_tab": 395, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1777, "n_element_tab_1": 78, "formula_len_all": 3775, "formula_len_all_1": 1417, "len_all": 164115, "len_all_1": 68598, "len_abs": 1148, "len_title": 106, "len_sents": 50125, "len_sents_1": 32281, "n_sents": 344, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1075, "title": "Endowing Visual Reprogramming with Adversarial Robustness", "abs": "Visual reprogramming (VR) leverages well-developed pre-trained models (e.g., a pre-trained classifier on ImageNet) to tackle target tasks (e.g., a traffic sign recognition task), without the need for training from scratch. Despite the effectiveness of previous VR methods, all of them did not consider the adversarial robustness of reprogrammed models against adversarial attacks, which could lead to unpredictable problems in safety-crucial target tasks. In this paper, we empirically find that reprogramming pre-trained models with adversarial robustness and incorporating adversarial samples from the target task during reprogramming can both improve the adversarial robustness of reprogrammed models. Furthermore, we propose a theoretically guaranteed adversarial robustness risk upper bound for VR, which validates our empirical findings and could provide a theoretical foundation for future research. Extensive experiments demonstrate that by adopting the strategies revealed in our empirical findings, the adversarial robustness of reprogrammed models can be enhanced.", "keywords": [ "visual reprogramming", "adversarial robustness", "risk bound" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "OuKMmtAvOi", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Brian Tom", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "Principal Researcher" }, { "name": "Chang Ho Yoon", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Researcher" }, { "name": "Fabian Falck", "gender": "Male", "institution": "Microsoft Research", "country": "GB", "position": "Researcher" }, { "name": "Jessica Barrett", "gender": "unknown", "institution": "MRC Biostatistics Unit, University of Cambridge", "country": "", "position": "Researcher" }, { "name": "Vincent Jeanselme", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Postdoc" } ], "n_formula": 20, "n_formula_1": 18, "n_ref_uni": 42, "n_ref": 80, "n_ref_all": 105, "n_fig": 6, "n_tab": 7, "L_tab": 3377, "n_element_tab": 76, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 943, "n_element_tab_1": 22, "formula_len_all": 1355, "formula_len_all_1": 944, "len_all": 162342, "len_all_1": 69715, "len_abs": 1708, "len_title": 124, "len_sents": 52727, "len_sents_1": 32795, "n_sents": 399, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1368, "title": "Identifying treatment response subgroups in observational time-to-event data", "abs": "Identifying patient subgroups with different treatment responses is an important task to inform medical recommendations, guidelines, and the design of future clinical trials. Existing approaches for subgroup analysis primarily rely on Randomised Controlled Trials (RCTs), in which treatment assignment is randomised. RCTs' patient cohorts are often constrained by cost, rendering them not representative of the heterogeneity of patients likely to receive treatment in real-world clinical practice. When applied to observational studies, subgroup analysis approaches suffer from significant statistical biases particularly because of the non-randomisation of treatment. Our work introduces a novel, outcome-guided method for identifying treatment response subgroups in observational studies. Our approach assigns each patient to a subgroup associated with two time-to-event distributions: one under treatment and one under control regime. It hence positions itself in between individualised and average treatment effect estimation. The assumptions of our model result in a simple correction of the statistical bias from treatment non-randomisation through inverse propensity weighting. In experiments, our approach significantly outperforms the current state-of-the-art method for outcome-guided subgroup analysis in both randomised and observational treatment regimes.", "keywords": [ "Survival analysis", "Treatment effect", "Observational data", "Subgroup discovery" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "Osr0KZJeTX", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Christopher Brinton", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "Dong-Jun Han", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Seyyedali Hosseinalipour", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Yun-Wei Chu", "gender": "Male", "institution": "Purdue University", "country": "", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 9, "n_ref_uni": 45, "n_ref": 265, "n_ref_all": 302, "n_fig": 7, "n_tab": 27, "L_tab": 18380, "n_element_tab": 1418, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2642, "n_element_tab_1": 56, "formula_len_all": 1287, "formula_len_all_1": 508, "len_all": 255551, "len_all_1": 71315, "len_abs": 1573, "len_title": 115, "len_sents": 69694, "len_sents_1": 33512, "n_sents": 420, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1435, "title": "Unlocking the Potential of Model Calibration in Federated Learning", "abs": "Over the past several years, various federated learning (FL) methodologies have been developed to improve model accuracy, a primary performance metric in machine learning. However, to utilize FL in practical decision-making scenarios, beyond considering accuracy, the trained model must also have a reliable confidence in each of its predictions, an aspect that has been largely overlooked in existing FL research. Motivated by this gap, we propose Non-Uniform Calibration for Federated Learning (NUCFL), a generic framework that integrates FL with the concept of model calibration. The inherent data heterogeneity in FL environments makes model calibration particularly difficult, as it must ensure reliability across diverse data distributions and client conditions. Our NUCFL addresses this challenge by dynamically adjusting the model calibration objectives based on statistical relationships between each client's local model and the global model in FL. In particular, NUCFL assesses the similarity between local and global model relationships, and controls the penalty term for the calibration loss during client-side local training. By doing so, NUCFL effectively aligns calibration needs for the global model in heterogeneous FL settings while not sacrificing accuracy. Extensive experiments show that NUCFL offers flexibility and effectiveness across various FL algorithms, enhancing accuracy as well as model calibration.", "keywords": [ "Federated Learning", "Model Calibration" ], "rating_list": [ 6, 5, 5, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 2, 5, 4 ] }, { "paper_id": "OspqtLVUN5", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Bei Li", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Gangao Liu", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Jingang Wang", "gender": "Male", "institution": "Meituan", "country": "", "position": "Researcher" }, { "name": "Peng Li", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "", "position": "Associate Professor" }, { "name": "Ruichen Shao", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Xunliang Cai", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Principal Researcher" }, { "name": "Yang Chen", "gender": "unknown", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "ZhouXiang", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 14, "n_ref_uni": 31, "n_ref": 61, "n_ref_all": 100, "n_fig": 10, "n_tab": 13, "L_tab": 6584, "n_element_tab": 903, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1134, "n_element_tab_1": 111, "formula_len_all": 2229, "formula_len_all_1": 1656, "len_all": 181639, "len_all_1": 66270, "len_abs": 1370, "len_title": 152, "len_sents": 47961, "len_sents_1": 29036, "n_sents": 365, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 103, "L_abs": 1385, "title": "Earlier Tokens Contribute More: Learning Direct Preference Optimization From Temporal Decay Perspective", "abs": "Direct Preference Optimization (DPO) has gained attention as an efficient alternative to reinforcement learning from human feedback (RLHF) for aligning large language models (LLMs) with human preferences. Despite its advantages, DPO suffers from a length bias, generating responses longer than those from the reference model. Existing solutions like SimPO and SamPO address this issue but uniformly treat the contribution of rewards across sequences, overlooking temporal dynamics. To this end, we propose an enhanced preference optimization method that incorporates a temporal decay factor controlled by a gamma parameter. This dynamic weighting mechanism adjusts the influence of each reward based on its position in the sequence, prioritizing earlier tokens that are more critical for alignment. By adaptively focusing on more relevant feedback, our approach mitigates overfitting to less pertinent data and remains responsive to evolving human preferences. Experimental results on several benchmarks show that our approach consistently outperforms vanilla DPO by 5.9-8.8 points on AlpacaEval 2 and 3.3-9.7 points on Arena-Hard across different model architectures and sizes. \\revised{Furthermore, additional experiments on mathematical and reasoning benchmarks, such as MMLU, GSM8K, and Math, confirm that our method enhances performance without compromising general capabilities.}", "keywords": [ "preference optimization", "RLHF", "DPO" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "OrBIjc0lMz", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Anupam Pani", "gender": "unknown", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yanchao Yang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 39, "n_ref": 55, "n_ref_all": 84, "n_fig": 9, "n_tab": 6, "L_tab": 1626, "n_element_tab": 148, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 795, "n_element_tab_1": 83, "formula_len_all": 586, "formula_len_all_1": 493, "len_all": 182653, "len_all_1": 61897, "len_abs": 1555, "len_title": 112, "len_sents": 61458, "len_sents_1": 29682, "n_sents": 439, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1567, "title": "Gaze-Regularized Attention for Human Action Prediction", "abs": "Eye gaze, encompassing fixations and saccades, offers valuable insights into human intentions and future actions. This study presents a novel approach to enhancing Vision Language Models (VLMs) for human action prediction by integrating eye gaze data into egocentric video analysis. Existing methods for action anticipation in egocentric videos often rely solely on visual data, potentially missing critical information provided by eye gaze. To address this limitation, we propose a unique gaze-augmented framework that integrates eye gaze directly into the VLM architecture and training process. By generating gaze heatmaps from eye gaze coordinates, our model dynamically focuses on regions highlighted by gaze patterns. Additionally, a gaze-regularization mechanism ensures the model maintains attention on gaze-allocated areas, thereby improving prediction accuracy and robustness. Our approach significantly enhances the model's ability to generate precise and detailed predictions of future actions. Compared to baseline models without leveraging gaze data, our method achieves a nearly 13\\% improvement in the semantic score of predictions. This substantial improvement underscores the effectiveness and novelty of integrating eye gaze with a gaze-regularized attention mechanism in VLMs for action anticipation. Moreover, our work demonstrates that incorporating eye gaze through this gaze-augmented framework can significantly boost the predictive capabilities of VLMs, enhancing their potential in applications that require accurate human action prediction.", "keywords": [ "human action prediction", "human-machine interaction", "eye gaze" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "Oqqbnn1snA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 1, "author_info_list": [ { "name": "Nannan Wu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 17, "n_ref_uni": 14, "n_ref": 20, "n_ref_all": 36, "n_fig": 3, "n_tab": 7, "L_tab": 1992, "n_element_tab": 120, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 3164, "n_element_tab_1": 103, "formula_len_all": 1100, "formula_len_all_1": 1026, "len_all": 121443, "len_all_1": 70635, "len_abs": 1271, "len_title": 106, "len_sents": 42020, "len_sents_1": 30158, "n_sents": 331, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1254, "title": "FedGraph: A New Paradigm for Federated Graph Learning", "abs": "Federated learning is a distributed approach to training a global model over multiple clients without sharing their local data. In graph data, the data heterogeneity can correspond to subgraph structures and node features varying extremely different, and the task-specific knowledge isolation corresponds to exclusive schema on handing data for specific task in clients, e.g., anomaly user setting in Twitter is rather different from LinkedIn. Although most feder- ated graph learning approaches are employed to address the data heterogeneity challenge, we find that the task-specific knowledge isolation challenge has been overlooked. This task-specific knowl- edge isolation will prevent existing models into the federated graph learning framework. In this paper, we propose FedGraph: a new paradigm for federated graph learning. The key idea is to utilize the graph structure without private node features as structure knowl- edge bridging all task specific knowledge in clients. Our extensive experiments show that FedGraph significantly outperforms the other state-of-the-art federated learning algorithms on anomaly de- tection tasks. Two deep learning models and one existing anomaly subgraph detection model are transferred to FedGraph framework.", "keywords": [ "Federated Learning", "Anomaly Alignment", "Network Alignment" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "Oqk1Ui6m0n", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 2, "author_info_list": [ { "name": "Anas Jnini", "gender": "unknown", "institution": "University of Trento", "country": "IT", "position": "PhD student" }, { "name": "Flavio Vella", "gender": "Male", "institution": "University of Trento", "country": "", "position": "Assistant Professor" } ], "n_formula": 39, "n_formula_1": 21, "n_ref_uni": 25, "n_ref": 45, "n_ref_all": 54, "n_fig": 5, "n_tab": 5, "L_tab": 1724, "n_element_tab": 90, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1843, "n_element_tab_1": 23, "formula_len_all": 2238, "formula_len_all_1": 1339, "len_all": 103138, "len_all_1": 51506, "len_abs": 1136, "len_title": 123, "len_sents": 30924, "len_sents_1": 21930, "n_sents": 234, "n_sents_1": 151, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1121, "title": "Hessian-Free Natural Gradient Descent for Physics Informed Machine Learning", "abs": "Physics-Informed Machine Learning (PIML) methods, such as Physics-Informed Neural Networks (PINNs), are notoriously difficult to optimize. Recent advances utilizing second-order optimization techniques, including natural gradient and Gauss-Newton methods, have significantly improved training accuracy over first-order methods. However, these approaches are computationally prohibitive, as they require evaluating, storing, and inverting large curvature matrices, limiting scalability to small networks. To overcome this limitation, we introduce a Hessian-Free Natural Gradient Descent framework that employs a matrix-free approximation of the Hessian. This approach circumvents the need for explicitly constructing the Hessian matrix and incorporates a novel preconditioning scheme that significantly enhances convergence rates. Our method enables scaling to large neural networks with up to a million of parameters. Empirically, we demonstrate that our approach outperforms state-of-the-art optimizers, such as LBFGS and Adam, achieving orders-of-magnitudes accuracy improvements across various benchmark PDE problems.", "keywords": [ "PINNs", "Gauss-Newton", "Function-space optimization", "Hessian-Free Optimization", "Second-order Optimizers" ], "rating_list": [ 6, 3, 5, 3, 8 ], "soundness_list": [ 3, 2, 3, 1, 4 ], "presentation_list": [ 3, 1, 4, 3, 4 ], "contribution_list": [ 3, 2, 3, 2, 4 ], "confidence_list": [ 3, 2, 4, 4, 4 ] }, { "paper_id": "OqZDfIknDe", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Chenxi Jiang", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "MS student" }, { "name": "Geng Li", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "Undergrad student" }, { "name": "Haozhi Cao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Mingyang Liu", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "MS student" }, { "name": "Jianfei Yang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 17, "n_ref_uni": 50, "n_ref": 88, "n_ref_all": 115, "n_fig": 15, "n_tab": 5, "L_tab": 890, "n_element_tab": 185, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 814, "n_element_tab_1": 144, "formula_len_all": 1670, "formula_len_all_1": 1054, "len_all": 151701, "len_all_1": 66071, "len_abs": 1597, "len_title": 107, "len_sents": 46844, "len_sents_1": 31034, "n_sents": 326, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1604, "title": "UniRiT: Towards Few-Shot Non-Rigid Point Cloud Registration", "abs": "Non-rigid point cloud registration is a critical challenge in 3D scene understanding, particularly in surgical navigation. Although existing methods achieve excellent performance when trained on large-scale, high-quality datasets, these datasets are prohibitively expensive to collect and annotate, e.g., organ data in authentic medical scenarios. With insufficient training samples and data noise, existing methods degrade significantly since non-rigid patterns are more flexible and complicated than rigid ones, and the distributions across samples are more distinct, leading to higher difficulty in representation learning with few data.\nIn this work, we aim to deal with this challenging few-shot non-rigid point cloud registration problem. Based on the observation that complex non-rigid transformation patterns can be decomposed into rigid and small non-rigid transformations, we propose a novel and effective framework, UniRiT. UniRiT adopts a two-step registration strategy that first aligns the centroids of the source and target point clouds and then refines the registration with non-rigid transformations, thereby significantly reducing the problem complexity. To validate the performance of UniRiT on real-world datasets, we introduce a new dataset, MedMatch3D, which consists of real human organs and exhibits high variability in sample distribution. We further establish a new challenging benchmark for few-shot non-rigid registration. Extensive empirical results demonstrate that UniRiT achieves state-of-the-art performance on MedMatch3D, improving the existing best approach by 94.22\\%.", "keywords": [ "Few-shot learning; Point cloud registration" ], "rating_list": [ 5, 5, 3, 1 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 1 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "OqTVwjLlRI", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Barun Patra", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Applied Scientist" }, { "name": "Liliang Ren", "gender": "unknown", "institution": "Microsoft GenAI", "country": "", "position": "Senior Researcher" }, { "name": "Suyu Ge", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Vishrav Chaudhary", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Senior Principal Researcher" }, { "name": "Xia Song", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Xihui Lin", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Yunan Zhang", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" }, { "name": "Hao Peng", "gender": "Male", "institution": "Department of Computer Science, University of Illinois Urbana-Champaign", "country": "", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 17, "n_ref": 42, "n_ref_all": 92, "n_fig": 6, "n_tab": 3, "L_tab": 836, "n_element_tab": 134, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1827, "n_element_tab_1": 195, "formula_len_all": 377, "formula_len_all_1": 377, "len_all": 107094, "len_all_1": 67440, "len_abs": 2150, "len_title": 0, "len_sents": 33269, "len_sents_1": 29407, "n_sents": 293, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 2860, "title": "S2-Attention: Hardware-Aware Context Sharding Among Attention Heads", "abs": "Sparse attention, which selectively attends to a subset of tokens in the context, has been an established approach to enhance the efficiency of Transformers. \nHowever, its theoretical reduction in FLOPs has rarely translated into wall-clock speed-up over its dense attention counterparts, mainly due to the lack of hardware-level optimizations like FlashAttention.\nMeanwhile, it remains unclear wheter sparse attention can maintain the model's quality at a scale of today's large language models (LLMs), and how this can be achieved.\n%how to guarantee model quality with sparse attention, given the decoder-only architecture and model scale of modern LLMs. \nThis paper presents Sparsely-Sharded(S2) Attention, a Triton library that provides kernel optimization for sparse attention customizable at both per-head and per-context-range levels.\nS2-Attention enables the exploration of novel and high-performance sparse attention techniques, which we demonstrate through extensive ablations across a wide range of sparse attention deisngs at various model scales. \n% design heuristics across model scale. \nFrom these insights, we present several basic guidelines to design sparse attention that can achieve not only practical efficiency improvements, but also strong performance on downstream tasks.\n% heterogeneous context sharding, union completeness, and inevitable density as principles to improve LLM training and inference efficiency without compromising model quality.\nTo achieve high parallelization and optimized memory IO, sparse attention should \\textbf{shard the context heterogeneously across attention heads}, where each head attends to a different subset of tokens while \\textbf{collectively covering the full context}. Meanwhile, we find hybrid architectures combining sparse and dense attention particularly beneficial in practice.\nThese design choices lead to a novel sparse attention architecture,\nwhich we evaluate with 1.3B, 7B models.\nIt achieves wall-clock speedup of 8.79X, 15.87X, 25.3X compared to the strong FlashAttention-2 baseline with strong downstream performance on-par with full attention and perfect retrieval performance at a 128k context length. \n% on-par downstream performance and perfect 128k needle retrieval.\n% In 1.3B, 7B, 70B Llama architecture models, S2-Attention following the principles delivers a speedup of 7X, 12X, 22X wall-clock speed-up compared to FlashAttention-2, while achieving on-par downstream performance and perfect 128k needle retrieval.\nIn inference, for 7B models, our model, with the help of our S2-Attention kernel, achieves 4.5x speed-up compared to dense counterparts. \nS2-Attention will be released with easy-to-customize APIs for direct usage in Megatron and vLLM. \nWe hope they will help future research develop sparse attention algorithms to improve the efficiency of large language models.", "keywords": [ "efficient transformer", "kernel", "sparse attention", "long context", "efficiency", "infrastructure", "pre-training", "inference", "sparsity", "software library" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "OqEsj4S240", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Alberto Cazzaniga", "gender": "Male", "institution": "AREA Science Park", "country": "IT", "position": "Researcher" }, { "name": "Alessio ansuini", "gender": "Male", "institution": "AREA Science Park", "country": "IT", "position": "Researcher" }, { "name": "Giada Panerai", "gender": "Female", "institution": "University of Trieste", "country": "IT", "position": "MS student" }, { "name": "Karthik Viswanathan", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Matteo Biagetti", "gender": "Male", "institution": "Area Science Park", "country": "IT", "position": "Researcher" }, { "name": "Yuri Gardinazzi", "gender": "Male", "institution": "University of Trieste", "country": "IT", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 8, "n_ref_uni": 58, "n_ref": 99, "n_ref_all": 130, "n_fig": 10, "n_tab": 1, "L_tab": 594, "n_element_tab": 134, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 930, "formula_len_all_1": 718, "len_all": 176421, "len_all_1": 60456, "len_abs": 1287, "len_title": 48, "len_sents": 45335, "len_sents_1": 29948, "n_sents": 306, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1293, "title": "Persistent Similarity in Internal Representations of Large Language Models", "abs": "Understanding the decision-making processes of large language models (LLMs) is critical given their widespread applications. Towards this goal, describing the topological and geometrical properties of internal representations has recently provided valuable insights. For a more comprehensive characterization of these inherently complex spaces, we present a novel framework based on zigzag persistence, a method in topological data analysis (TDA) well-suited for describing data undergoing dynamic transformations across layers. Within this framework, we introduce persistence similarity, a new topological descriptor that quantifies the persistence and transformation of topological features such as $p$-cycles throughout the model layers. Unlike traditional similarity measures, our approach captures the entire evolutionary trajectory of these features, providing deeper insights into the internal workings of LLMs. As a practical application, we leverage persistence similarity to identify and prune layers, demonstrating comparable performance to state-of-the-art methods across several benchmark datasets. Additionally, our analysis reveals similar topological behaviors across various models and hyperparameter settings, suggesting a universal structure in LLM internal representations.", "keywords": [ "Topological Data Analysis", "Persistent Homology", "Large Language Models", "Internal Representations", "Similarity", "Layer Pruning" ], "rating_list": [ 5, 5, 3, 6, 3 ], "soundness_list": [ 3, 2, 1, 3, 2 ], "presentation_list": [ 4, 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 2, 3 ] }, { "paper_id": "Oq8bDXRf4F", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Calvin Yeung", "gender": "unknown", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Mohsen Imani", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Assistant Professor" }, { "name": "Nathaniel D. Bastian", "gender": "Male", "institution": "United States Military Academy", "country": "US", "position": "Principal Researcher" }, { "name": "Zhuowen Zou", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 30, "n_ref_uni": 12, "n_ref": 21, "n_ref_all": 35, "n_fig": 10, "n_tab": 2, "L_tab": 193, "n_element_tab": 27, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 193, "n_element_tab_1": 27, "formula_len_all": 2009, "formula_len_all_1": 1794, "len_all": 75840, "len_all_1": 55146, "len_abs": 1678, "len_title": 119, "len_sents": 27066, "len_sents_1": 23646, "n_sents": 227, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1687, "title": "Cognitive map formation under uncertainty via local prediction learning", "abs": "Cognitive maps are internal world models that enable adaptive behavior including spatial navigation and planning. The Cognitive Map Learner (CML) has been recently proposed as a model for cognitive map formation and planning. A CML learns high dimensional state and action representations using local prediction learning. While the CML offers a simple and elegant solution to cognitive map learning, it is limited by its simplicity, applying only to fully observable environments. To address this, we introduce the Partially Observable Cognitive Map Learner (POCML), extending the CML to handle partially observable environments.\n\nThe POCML employs a superposition of states for probabilistic representation and uses binding operations for state updates. Additionally, an associative memory is incorporated to enable adaptive behavior across environments with similar structures. We derive local update rules tailored to the POCML's probabilistic state representation and associative memory. We demonstrate a POCML is capable of learning the underlying structure of an environment via local next-observation prediction learning. In addition, we show that a POCML trained on an environment is capable of generalizing to environments with the same underlying structure but with novel observations, achieving good zero-shot next-observation prediction accuracy, significantly outperforming sequence models such as LSTMs and Transformers. Finally, we present a case study of navigation in a two-tunnel maze environment with aliased observations, showing that a POCML is capable of effectively using its probabilistic state representations for disambiguation of states and spatial navigation.", "keywords": [ "cognitive maps", "local prediction learning", "vector symbolic architectures" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "Oq7BhRSy0a", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Awni Altabaa", "gender": "unknown", "institution": "Yale University", "country": "", "position": "PhD student" }, { "name": "John Lafferty", "gender": "unknown", "institution": "Yale University", "country": "", "position": "Professor" } ], "n_formula": 26, "n_formula_1": 9, "n_ref_uni": 56, "n_ref": 112, "n_ref_all": 135, "n_fig": 11, "n_tab": 6, "L_tab": 1542, "n_element_tab": 289, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 663, "n_element_tab_1": 92, "formula_len_all": 3052, "formula_len_all_1": 1460, "len_all": 199943, "len_all_1": 72093, "len_abs": 1443, "len_title": 141, "len_sents": 73237, "len_sents_1": 33289, "n_sents": 611, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1455, "title": "Disentangling and Integrating Relational and Sensory Information in Transformer Architectures", "abs": "Relational reasoning is a central component of generally intelligent systems, enabling robust and data-efficient inductive generalization. Recent empirical evidence shows that many existing neural architectures, including Transformers, struggle with tasks requiring relational reasoning. In this work, we distinguish between two types of information: *sensory* information about the properties of individual objects, and *relational* information about the relationships between objects. While neural attention provides a powerful mechanism for controlling the flow of sensory information between objects, the Transformer lacks an explicit computational mechanism for routing and processing relational information. To address this limitation, we propose an architectural extension of the Transformer framework that we call the *Dual Attention Transformer (DAT)*, featuring two distinct attention mechanisms: sensory attention for directing the flow of sensory information, and a novel relational attention mechanism for directing the flow of relational information. We empirically evaluate *DAT* on a diverse set of tasks ranging from synthetic relational benchmarks to complex real-world tasks such as language modeling and visual processing. Our results demonstrate that integrating explicit relational computational mechanisms into the Transformer architecture leads to significant performance gains in terms of data efficiency and parameter efficiency.", "keywords": [ "relational learning", "transformers", "inductive biases", "sensory", "relational", "architecture", "attention" ], "rating_list": [ 5, 10, 3, 6 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "Opq0InLQn1", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Dohwan Ko", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "PhD student" }, { "name": "Ji Soo Lee", "gender": "unknown", "institution": "Korea University", "country": "KR", "position": "MS student" }, { "name": "Minhyuk Choi", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "MS student" }, { "name": "Zihang Meng", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Hyunwoo Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 41, "n_ref": 108, "n_ref_all": 167, "n_fig": 24, "n_tab": 8, "L_tab": 3994, "n_element_tab": 758, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 867, "formula_len_all_1": 641, "len_all": 203242, "len_all_1": 63329, "len_abs": 1989, "len_title": 129, "len_sents": 51172, "len_sents_1": 30529, "n_sents": 395, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 2043, "title": "Bidirectional Generative Retrieval with Multi-Modal LLMs for Text-Video Retrieval", "abs": "In recent years, multi-modal large language models (MLLMs) have shown outstanding advancement in various multi-modal understanding tasks by leveraging the powerful knowledge of large language models (LLMs). Extending MLLMs to text-video retrieval enables handling more complex queries with multiple modalities beyond simple uni-modal queries for traditional search engines. It also provides a new opportunity to incorporate search into a unified conversational system, but MLLM-based text-video retrieval has been less explored in the literature. To this end, we investigate MLLMs' capabilities in text-video retrieval as a generation task, namely, generative retrieval, in two directions. An intuitive direction is $\\textit{content generation}$ that directly generates the content given a query. Another direction is $\\textit{query generation}$, which generates the query given the content. Interestingly, we observe that in both text-to-video and video-to-text retrieval tasks, query-generation less suffers from the bias and significantly outperforms content-generation. In this paper, we propose a novel framework, Bidirectional Text-Video Generative Retrieval (BGR), that handles both text-to-video and video-to-text retrieval tasks by measuring the relevance using two generation directions. Our framework trains MLLMs by simultaneously optimizing two objectives, $\\textit{i.e.}$, video-grounded text generation (VTG) and text-grounded video feature generation (TVG). At inference, our framework ensembles predictions by both generation directions. We also introduce a Prior Normalization, a simple plug-and-play module, to further alleviate the $\\textit{prior bias}$ induced by the likelihood of uni-modal content data that often overwhelms the relevance between query and content. Our extensive experiments on multi-modal benchmarks demonstrate that BGR and Prior Normalization are effective in alleviating the prior bias, especially the text prior bias from LLMs' pretrained knowledge in MLLMs, achieving state-of-the-art performance.", "keywords": [ "Video Retrieval", "Multi-modal Large Language Model" ], "rating_list": [ 3, 3, 5, 8, 5 ], "soundness_list": [ 3, 2, 3, 3, 1 ], "presentation_list": [ 2, 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3, 1 ], "confidence_list": [ 5, 4, 3, 3, 4 ] }, { "paper_id": "OpSMgpBubj", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Abhirup Mallik", "gender": "Male", "institution": "Google", "country": "US", "position": "Intern" }, { "name": "Burhaneddin Yaman", "gender": "unknown", "institution": "Bosch Center for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Chenbin Pan", "gender": "Female", "institution": "Bosch", "country": "US", "position": "Researcher" }, { "name": "Liu Ren", "gender": "Male", "institution": "Bosch Research", "country": "US", "position": "Principal Researcher" }, { "name": "Senem Velipasalar", "gender": "Female", "institution": "Syracuse University", "country": "US", "position": "Professor" }, { "name": "Xin Ye", "gender": "Female", "institution": "Bosch", "country": "US", "position": "Researcher" }, { "name": "Feng Tao", "gender": "Male", "institution": "Bosch", "country": "US", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 20, "n_ref": 48, "n_ref_all": 59, "n_fig": 1, "n_tab": 7, "L_tab": 2490, "n_element_tab": 354, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2072, "n_element_tab_1": 271, "formula_len_all": 698, "formula_len_all_1": 604, "len_all": 122528, "len_all_1": 57682, "len_abs": 884, "len_title": 94, "len_sents": 33427, "len_sents_1": 25544, "n_sents": 226, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 887, "title": "Adapting CLIP for DETR-based Object Detection", "abs": "Object detection involves class identification and spatial positioning. While DETR-based architectures have shown promising detection capabilities by framing the task as set prediction, prior approaches have limited refinement for object features, leading to inferior inherent understanding of objects, particularly when generalizing to unseen categories. To this end, we propose CLIP-DETR, a novel detection framework that harnesses the pretrained visual-linguistic capabilities of CLIP to enhance both the encoding and decoding processes in DETR models. Our method focuses on two key principles: 1) feature map sensitivity to objects, and 2) query adaptability. Extensive experiments demonstrate that CLIP-DETR significantly outperforms state-of-the-art models in object detection and open-vocabulary detection tasks, illustrating its superior generalization and recognition abilities.", "keywords": [ "Object detection", "contrastive learning", "open-vocabulary object detection" ], "rating_list": [ 5, 5, 5, 5, 8 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 1, 2, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 2, 4, 3, 3, 3 ] }, { "paper_id": "OpNMWVDdKS", "primary_area": "reinforcement learning", "n_author": 1, "author_info_list": [ { "name": "Sumeet Batra", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 12, "n_ref_uni": 31, "n_ref": 65, "n_ref_all": 82, "n_fig": 13, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1462, "formula_len_all_1": 717, "len_all": 149838, "len_all_1": 61459, "len_abs": 1038, "len_title": 117, "len_sents": 47673, "len_sents_1": 31546, "n_sents": 295, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1044, "title": "Zero Shot Generalization of Vision-Based RL Without Data Augmentation", "abs": "Generalizing vision-based reinforcement learning (RL) agents to novel environments remains a difficult and open challenge. Current trends are to collect large-scale datasets or use data augmentation techniques to prevent overfitting and improve downstream generalization. However, the computational and data collection costs increase exponentially with the number of task variations and can destabilize the already difficult task of training RL agents. In this work, we take inspiration from recent advances in computational neuroscience and propose a model, Associative Latent DisentAnglement (ALDA), that builds on standard off-policy RL towards zero-shot generalization. Specifically, we revisit the role of latent disentanglement in RL and show how combining it with a model of associative memory achieves zero-shot generalization on difficult task variations *without* relying on data augmentation. Finally, we formally show that data augmentation techniques are a form of weak disentanglement and discuss the implications of this insight.", "keywords": [ "reinforcement learning", "representation learning", "disentangled representation learning", "associative memory", "robotics" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "OovfCS4FYT", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Keyu Long", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Undergrad student" }, { "name": "Xuzhe Zhi", "gender": "Female", "institution": "Johns Hopkins University", "country": "", "position": "MS student" }, { "name": "Yash Shah", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Garrison Cottrell", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "Emeritus" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 37, "n_ref": 70, "n_ref_all": 82, "n_fig": 5, "n_tab": 5, "L_tab": 1706, "n_element_tab": 116, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 927, "n_element_tab_1": 85, "formula_len_all": 162, "formula_len_all_1": 118, "len_all": 105827, "len_all_1": 58617, "len_abs": 1175, "len_title": 125, "len_sents": 34891, "len_sents_1": 28850, "n_sents": 236, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1183, "title": "Modeling Divisive Normalization as Learned Local Competition in Visual Cortex", "abs": "Convolutional Neural Networks (CNNs) embody priors about the visual world: locality, stationary statistics, translation invariance, and compositionality. Similarly, CNNs implement the retinotopy of visual cortex---nearby pixels are processed by nearby neurons. A common cortical computation not usually included in CNNs is divisive normalization. It has been shown that divisive normalization of Gabor filters results in more statistically independent responses (Simoncelli & Heeger, 1998). In this paper, we model divisive normalization as a simple computationally-efficient layer that can be inserted at any stage within deep artificial neural networks. Divisive normalization acts on neuronal sub-populations, whose parameters are initialized from a multivariate Gaussian distribution. This leads to the emergence of learned competition between both orientation-preferring and color-opponent cell types. Divisive normalization improves categorization performance, as well as robustness to perturbed images. Interestingly, in smaller networks, divisive normalization as a non-linear operation eliminates the need for a non-linear activation function like ReLU to drive performance.", "keywords": [ "divisive normalization", "deep artificial neural networks", "robustness", "local competition", "computational neuroscience" ], "rating_list": [ 1, 6, 3, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 1, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "OopiU1q328", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Alex Gabel", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Efstratios Gavves", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Associate Professor" }, { "name": "Rick Quax", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 30, "n_ref": 46, "n_ref_all": 50, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 465, "formula_len_all_1": 465, "len_all": 67209, "len_all_1": 67209, "len_abs": 1037, "len_title": 115, "len_sents": 18011, "len_sents_1": 18011, "n_sents": 125, "n_sents_1": 125, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1059, "title": "PowerNet: Truncated Matrix Power Series as Quasi-Equivariant Layers", "abs": "Despite being theoretically well-grounded, enforcing strict equivariance in deep learning models has shown to be harmful in some cases. The problem is that most available data does not follow mathematically precise rules, is noisy, and is not strictly group-structured. While soft equivariance approaches attempt to address these issues, they often struggle to maintain group structure and lack strong theoretical guarantees, potentially compromising the benefits of equivariance. Here we introduce the concept of \\textit{quasi equivariances}, where group structure is maintained but the associated parameters become distributions, and implement it in the proposed \\textit{PowerNet} architecture. Similar to CNNs, PowerNet is constructed by interlacing truncated matrix power series with non-linearities. We show how the base matrix used to define the power series can instill quasi-equivariance in a natural way. Finally, we provide results for augmented MNIST classification and transformation magnitude regression in addition to classification of CIFAR-10.", "keywords": [ "symmetry", "deep learning", "geometry", "Lie theory" ], "rating_list": [ 1, 3, 3, 1 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 1, 1, 1, 1 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 4, 2, 4, 2 ] }, { "paper_id": "OoNazl6T7D", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Haiteng Zhao", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Junlei Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Junxian He", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Lingpeng Kong", "gender": "Male", "institution": "Department of Computer Science, The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Ma Chang", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 54, "n_ref": 134, "n_ref_all": 159, "n_fig": 10, "n_tab": 10, "L_tab": 2976, "n_element_tab": 233, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 3392, "n_element_tab_1": 231, "formula_len_all": 1255, "formula_len_all_1": 885, "len_all": 221040, "len_all_1": 68436, "len_abs": 974, "len_title": 111, "len_sents": 59789, "len_sents_1": 28503, "n_sents": 556, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1009, "title": "Language Model Non-Myopic Generation for Reasoning and Planning", "abs": "Large Language Models have demonstrated remarkable abilities in reasoning and planning by breaking down complex problems into sequential steps. Despite their success in various domains like mathematical problem-solving and coding, LLMs face challenges in ensuring reliable and optimal planning due to their inherent myopic nature of autoregressive decoding. This paper revisits LLM reasoning from an optimal-control perspective, proposing a novel method, Predictive-Decoding, that leverages Model Predictive Control to enhance planning accuracy. By re-weighting LLM distributions based on foresight trajectories, Predictive-Decoding aims to mitigate early errors and promote non-myopic planning. Our experiments show significant improvements in a wide range of tasks for math, coding, and agents. Furthermore, Predictive-Decoding demonstrates computational efficiency, outperforming search baselines with reduced computational resources. This study provides insights into optimizing LLM planning capabilities.", "keywords": [ "LLM reasoning; agents; optimal control" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 4, 2, 3, 2 ] }, { "paper_id": "Onw93uJCWO", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Junyu Luo", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Ming Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Pengyun Wang", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "MS student" }, { "name": "Siyu Heng", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Xiao Luo", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Yanxin Shen", "gender": "unknown", "institution": "Nankai University", "country": "", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 73, "n_ref": 210, "n_ref_all": 233, "n_fig": 6, "n_tab": 11, "L_tab": 8626, "n_element_tab": 548, "n_fig_1": 3, "n_tab_1": 9, "L_tab_1": 9919, "n_element_tab_1": 546, "formula_len_all": 105, "formula_len_all_1": 105, "len_all": 238542, "len_all_1": 71972, "len_abs": 1615, "len_title": 135, "len_sents": 57062, "len_sents_1": 26298, "n_sents": 466, "n_sents_1": 159, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1332, "title": "A Comprehensive Graph Pooling Benchmark: Effectiveness, Robustness and Generalizability", "abs": "Graph pooling has gained attention for its ability to obtain effective node and graph representations for various downstream tasks. Despite the recent surge in graph pooling approaches, there is a lack of standardized experimental settings and fair benchmarks to evaluate their performance. To address this issue, we have constructed a comprehensive benchmark that includes 17 graph pooling methods and 28 different graph datasets. This benchmark systematically assesses the performance of graph pooling methods in three dimensions, i.e., effectiveness, robustness, and generalizability. We first evaluate the performance of these graph pooling approaches across different tasks including graph classification, graph regression and node classification. Then, we investigate their performance under potential noise attacks and out-of-distribution shifts in real-world scenarios. We also involve detailed efficiency analysis, backbone analysis, parameter analysis and visualization to provide more evidence. Extensive experiments validate the strong capability and applicability of graph pooling approaches in various scenarios, which can provide valuable insights and guidance for deep geometric learning research. The source code of our benchmark is available at \\url{https://anonymous.4open.science/r/Graph_Pooling_Benchmark-8EDD}.", "keywords": [ "Graph Pooling; Benchmark; Graph Neural Networks; Graph Machine Learning" ], "rating_list": [ 3, 5, 8, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "OnMRWwOqCs", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Cong Shen", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Associate Professor" }, { "name": "Jundong Li", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" }, { "name": "Peng Wang", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Song Wang", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Yu Meng", "gender": "Male", "institution": "University of Virginia", "country": "", "position": "Assistant Professor" }, { "name": "Zhen Tan", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Zhepei Wei", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Zihan Chen", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 49, "n_ref": 80, "n_ref_all": 91, "n_fig": 5, "n_tab": 3, "L_tab": 948, "n_element_tab": 131, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1038, "n_element_tab_1": 139, "formula_len_all": 625, "formula_len_all_1": 549, "len_all": 154582, "len_all_1": 73923, "len_abs": 1559, "len_title": 149, "len_sents": 39449, "len_sents_1": 33731, "n_sents": 306, "n_sents_1": 256, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1167, "title": "Separate the Wheat from the Chaff: Winnowing Down Divergent Views in Retrieval Augmented Generation", "abs": "Retrieval-augmented generation (RAG) addresses the limitation of large language models (LLMs) in achieving up-to-date information by integrating external knowledge sources, but it is hindered by noisy or irrelevant retrieved data, leading to reduced accuracy. Additionally, most RAG methods rely on task-specific supervision, reducing their adaptability across domains. \nTo overcome these challenges, we propose WinnowRAG, a novel multi-agent debate-based RAG framework. WinnowRAG operates in two stages: in Stage I, query-aware clustering groups similar documents, with each cluster assigned to an LLM agent for generating personalized responses. A critic LLM then consolidates these answers, forming super-agents. In Stage II, the super-agents engage in a structured discussion to filter out incorrect or irrelevant information, ensuring only relevant knowledge is used for final response generation. Crucially, WinnowRAG is unsupervised and leverages pretrained LLMs without requiring fine-tuning, making it easily adaptable to various tasks. The experiments on various realistic datasets demonstrate the effectiveness of WinnowRAG over state-of-the-art baselines.", "keywords": [ "Retrieval Augmented Generation", "Large Language Modles" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 5, 5, 3 ] }, { "paper_id": "OnBCQgi2LY", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Amy R Ward", "gender": "Female", "institution": "University of Chicago", "country": "US", "position": "Full Professor" }, { "name": "Bingxuan Li", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "MS student" }, { "name": "Pengyi Shi", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 28, "n_ref": 38, "n_ref_all": 56, "n_fig": 8, "n_tab": 4, "L_tab": 933, "n_element_tab": 98, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 272, "n_element_tab_1": 19, "formula_len_all": 197, "formula_len_all_1": 260, "len_all": 145046, "len_all_1": 65467, "len_abs": 1442, "len_title": 129, "len_sents": 53748, "len_sents_1": 32141, "n_sents": 419, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1474, "title": "Latent Feature Mining for Predictive Model Enhancement with Large Language Models", "abs": "Predictive modeling often faces challenges due to limited data availability and quality, especially in domains where collected features are weakly correlated with outcomes and where additional data collection is constrained by ethical or practical difficulties. Traditional machine learning (ML) models struggle to incorporate unobserved yet critical factors. In this work, we introduce a novel approach to formulate latent feature mining as text-to-text propositional logical reasoning. We propose FLAME (Faithful Latent FeAture Mining for Predictive Model Enhancement), a framework that leverages large language models (LLMs) to augment observed features with latent features, enhancing the predictive power of ML models in downstream tasks. Our novel approach transforms the latent feature extraction task to a text-to-text propositional reasoning task. Our framework is generalizable across various domains with minimal domain-specific customization, ensuring easy transfer to other areas facing similar challenges in data availability. We validate our framework with two case studies: (1) the criminal justice system, a domain characterized by limited and ethically challenging data collection. (2) the healthcare domain, where patient privacy concerns and the complexity of medical data often limit comprehensive feature collection. Our results show that inferred latent features align well with ground truth labels and significantly enhance the downstream classifier.", "keywords": [ "data mining", "large language models", "feature extraction", "criminal justice", "AI for social good", "AI for healthcare" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "On8E0U9vbz", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Linglan Zhao", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Weiran Huang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiangzhong Fang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Zhiquan Tan", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhongqiang Zhang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "LiuDuo", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 17, "n_ref_uni": 23, "n_ref": 38, "n_ref_all": 68, "n_fig": 7, "n_tab": 10, "L_tab": 6966, "n_element_tab": 563, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 4624, "n_element_tab_1": 304, "formula_len_all": 1491, "formula_len_all_1": 879, "len_all": 138082, "len_all_1": 65601, "len_abs": 1318, "len_title": 152, "len_sents": 35239, "len_sents_1": 25666, "n_sents": 307, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1339, "title": "Generalized Category Discovery Utilizing Reciprocal Learning and Class-wise Distribution Regularization", "abs": "Generalized Category Discovery (GCD) aims to identify unlabeled samples by leveraging the base knowledge from labeled ones, where the unlabeled set consists of both base and novel classes. \nSince clustering methods are time-consuming at inference, parametric-based approaches have become more popular. \nHowever, recent parametric-based methods suffer inferior base discrimination due to the unreliable self-supervision. \nTo address this issue, we propose a Reciprocal Learning Framework (RLF) that introduces an auxiliary branch devoted to base classification. \nDuring training, the main branch filters the pseudo-base samples to the auxiliary branch. \nIn response, the auxiliary branch provides more reliable soft labels for the main branch, leading to a virtuous cycle. \nFurthermore, we introduce Class-wise Distribution Regularization (CDR) to mitigate the leaning bias towards base classes. \nCDR essentially increases the prediction confidence of the unlabeled data and boosts the novel class performance. \nCombined with both components, our method achieves superior performance in all classes with negligible extra computation. \nExtensive experiments on seven GCD datasets validate the effectiveness of our method, e.g. delivering a notable 2.1\\% improvement on the Stanford Cars dataset.\nOur codes will be available upon acceptance.", "keywords": [ "Generalized Category Discovery", "Novel Class Discovery", "Distribution Regularization" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "OmpTdjl7RV", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Kim Yong Tan", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" }, { "name": "Yueming Lyu", "gender": "Male", "institution": "Agency for Science, Technology and Research (A*STAR)", "country": "SG", "position": "Researcher" }, { "name": "Ivor W Tsang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Yew Soon Ong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 67, "n_ref_all": 111, "n_fig": 15, "n_tab": 4, "L_tab": 1507, "n_element_tab": 153, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1991, "formula_len_all_1": 1359, "len_all": 149249, "len_all_1": 64025, "len_abs": 1437, "len_title": 140, "len_sents": 44819, "len_sents_1": 30321, "n_sents": 355, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1539, "title": "Fast Direct: Query-Efficient Online Black-box Guidance for Diffusion-model Target Generation", "abs": "Guided diffusion-model generation is a promising direction for customizing the generation process of a pre-trained diffusion-model to address the specific downstream tasks. Existing guided diffusion models either rely on training of the guidance model with pre-collected datasets or require the objective functions to be differentiable. However, for most real-world tasks, the offline datasets are often unavailable, and their objective functions are often not differentiable, such as image generation with human preferences, molecular generation for drug discovery, and material design. Thus, we need an **online** algorithm capable of collecting data during runtime and supporting a **black-box** objective function. Moreover, the **query efficiency** of the algorithm is also critical because the objective evaluation of the query is often expensive in the real-world scenarios. In this work, we propose a novel and simple algorithm, **Fast Direct**, for query-efficient online black-box target generation. Our Fast Direct builds a pseudo-target on the data manifold to update the noise sequence of the diffusion model with a universal direction, which is promising to perform query-efficient guided generation. Extensive experiments on twelve high-resolution ($\\small {1024 \\times 1024}$) image target generation tasks and six 3D-molecule target generation tasks show $\\textbf{6}\\times$ up to $\\textbf{10}\\times$ query efficiency improvement and $\\textbf{11}\\times$ up to $\\textbf{44}\\times$ query efficiency improvement, respectively.", "keywords": [ "Diffusion model", "Black-box target generation", "Online guided diffusion model", "Query-efficient" ], "rating_list": [ 8, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 3, 3, 3, 1 ] }, { "paper_id": "Ombm8S40zN", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Alexander Tong", "gender": "unknown", "institution": "Mila", "country": "", "position": "Postdoc" }, { "name": "Cheng-Hao Liu", "gender": "Male", "institution": "Dreamfold", "country": "CA", "position": "Co-founder" }, { "name": "Jarrid Rector-Brooks", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, University of Montreal", "country": "", "position": "Researcher" }, { "name": "Joey Bose", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Michael M. Bronstein", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Mohsin Hasan", "gender": "Not Specified", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "PhD student" }, { "name": "Nouha Dziri", "gender": "Female", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Pranam Chatterjee", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Assistant Professor" }, { "name": "Sarthak Mittal", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "Graduate Student" }, { "name": "Zhangzhi Peng", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" } ], "n_formula": 55, "n_formula_1": 20, "n_ref_uni": 70, "n_ref": 116, "n_ref_all": 162, "n_fig": 13, "n_tab": 8, "L_tab": 2208, "n_element_tab": 117, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1459, "n_element_tab_1": 101, "formula_len_all": 5761, "formula_len_all_1": 1884, "len_all": 270820, "len_all_1": 78980, "len_abs": 1361, "len_title": 133, "len_sents": 86793, "len_sents_1": 37226, "n_sents": 657, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1364, "title": "Steering Masked Discrete Diffusion Models via Discrete Denoising Posterior Prediction", "abs": "Generative modeling of discrete data underlies important applications spanning text-based agents like ChatGPT to the design of the very building blocks of life in protein sequences. However, application domains need to exert control over the generated data by steering the generative process\u2014typically via RLHF\u2014to satisfy a specified property, reward, or affinity metric. In this paper, we study the problem of steering Masked Diffusion Models (MDMs), a recent class of discrete diffusion models that offer a compelling alternative to traditional autoregressive models. We introduce Discrete Denoising Posterior Prediction (DDPP), a novel framework that casts the task of steering pretrained MDMs as a problem of probabilistic inference by learning to sample from a target Bayesian posterior. Our DDPP framework leads to a family of three novel objectives that are all simulation-free, and thus scalable while applying to general non-differentiable reward functions. Empirically, we instantiate DDPP by steering MDMs to perform class-conditional pixel-level image modeling, RLHF-based alignment of MDMs using text based rewards, and finetuning protein language models to generate more diverse secondary structures and shorter proteins. We substantiate our designs via wet-lab validation, where we observe transient expression of reward-optimized protein sequences.", "keywords": [ "Discrete diffusion models", "language modeling", "probabilistic inference" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "OmFlDvsvc3", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Alessandro Abate", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "David Krueger", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Assistant Professor" }, { "name": "Joar Max Viktor Skalse", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Leon Lang", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Lukas Fluri", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Intern" }, { "name": "Patrick Forr\u00e9", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Associate Professor" } ], "n_formula": 260, "n_formula_1": 11, "n_ref_uni": 69, "n_ref": 195, "n_ref_all": 291, "n_fig": 4, "n_tab": 1, "L_tab": 543, "n_element_tab": 9, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 26811, "formula_len_all_1": 660, "len_all": 464080, "len_all_1": 69047, "len_abs": 2055, "len_title": 148, "len_sents": 150809, "len_sents_1": 34504, "n_sents": 1405, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1071, "title": "The Perils of Optimizing Learned Reward Functions: Low Training Error Does Not Guarantee Low Regret", "abs": "In reinforcement learning, specifying reward functions that capture the intended task can be very challenging. Reward learning aims to address this issue by learning the reward function. However, a learned reward model may have a low loss on the training distribution, and yet subsequently produce a policy with large regret. We say that such a reward model has an error-regret mismatch. The main source of an error-regret mismatch is the distribution shift that commonly occurs during policy optimization. In this paper, we mathematically show that a sufficiently low expected test error of the reward model guarantees low worst-case regret, but that for any fixed expected test error, there exist realistic data distributions that allow for error-regret mismatch to occur. We then show that similar problems persist even when using policy regularization techniques, commonly employed in methods such as RLHF. We hope our results stimulate the theoretical and empirical study of improved methods to learn reward models, and better ways to reliably measure their quality.", "keywords": [ "Reward learning", "RLHF", "RL", "Safety", "Distributional shift", "Generalization", "Learning Theory" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 4, 2, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "OlzB6LnXcS", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Danijar Hafner", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Kevin Frans", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Pieter Abbeel", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 32, "n_ref": 67, "n_ref_all": 85, "n_fig": 8, "n_tab": 2, "L_tab": 565, "n_element_tab": 79, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 565, "n_element_tab_1": 79, "formula_len_all": 1092, "formula_len_all_1": 1059, "len_all": 121326, "len_all_1": 67248, "len_abs": 1099, "len_title": 86, "len_sents": 38383, "len_sents_1": 30721, "n_sents": 329, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 38, "L_abs": 1103, "title": "One Step Diffusion via Shortcut Models", "abs": "Diffusion models and flow matching models have enabled generating diverse and realistic images by learning to transfer noise to data. However, sampling from these models involves iterative denoising over many neural network passes, making generation slow and expensive. Previous approaches for speeding up sampling require complex training regimes, such as multiple training phases, multiple networks, or fragile scheduling. We introduce Shortcut Models, a family of generative models that use a single network and training phase to produce high-quality samples in a single or multiple sampling steps. Shortcut models condition the network not only on the current noise level but also on the desired step size, allowing the model to skip ahead in the generation process. Across a wide range of sampling step budgets, shortcut models consistently produce higher quality samples than previous approaches, such as consistency models and reflow. Compared to distillation, shortcut models reduce complexity to a single network and training phase and additionally allow varying step budgets at inference time.", "keywords": [ "diffusion", "flow-matching", "fast inference", "distillation" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "OlytBskWjc", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Armen Mkrtchyan", "gender": "Male", "institution": "Flagship Pioneering", "country": "", "position": "Instructor" }, { "name": "Drew Dresser", "gender": "unknown", "institution": "Flagship Pioneering", "country": "US", "position": "Researcher" }, { "name": "J. Mark Kim", "gender": "Male", "institution": "Flagship Pioneering", "country": "US", "position": "Principal" }, { "name": "Olivia Viessmann", "gender": "Female", "institution": "Flagship Pioneering", "country": "US", "position": "Principal Researcher" }, { "name": "Siddharth Narayanan", "gender": "Male", "institution": "FutureHouse", "country": "US", "position": "Researcher" }, { "name": "Yasha Ektefaie", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 26, "n_ref": 44, "n_ref_all": 65, "n_fig": 7, "n_tab": 11, "L_tab": 4566, "n_element_tab": 733, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2219, "n_element_tab_1": 226, "formula_len_all": 1051, "formula_len_all_1": 887, "len_all": 133956, "len_all_1": 54993, "len_abs": 1415, "len_title": 145, "len_sents": 33043, "len_sents_1": 23020, "n_sents": 259, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1422, "title": "Reinforcement learning on structure-conditioned categorical diffusion for protein inverse folding", "abs": "Protein inverse folding\u2014that is, predicting an amino acid sequence that will fold into the desired 3D structure\u2014is an important problem for structure-based protein design. Machine learning based methods for inverse folding typically use recovery of the original sequence as the optimization objective. However, inverse folding is a one-to-many problem where several sequences can fold to the same structure. Moreover, for many practical applications, it is often desirable to have multiple, diverse sequences that fold into the target structure since it allows for more candidate sequences for downstream optimizations. Here, we demonstrate that although recent inverse folding methods show increased sequence recovery, their \u201cfoldable diversity\u201d\u2014i.e. their ability to generate multiple non-similar sequences that fold into the structures consistent with the target\u2014does not increase. To address this, we present RL-DIF, a categorical diffusion model for inverse folding that is pre-trained on sequence recovery and tuned via reinforcement learning on structural consistency. We find that RL-DIF achieves comparable sequence recovery and structural consistency to benchmark models but shows greater foldable diversity: experiments show RL-DIF can achieve an foldable diversity of 29% on CATH 4.2, compared to 23% from models trained on the same dataset. The PyTorch model weights and sampling code are available on GitHub.", "keywords": [ "Protein Inverse Folding", "Diffusion", "Reinforcement Learning" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "Olb8JwUGZ3", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Cameron Ethan Taylor", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Constantine Dovrolis", "gender": "Male", "institution": "Georgia Tech Research Corporation", "country": "", "position": "" }, { "name": "Shreyas Malakarjun Patil", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 51, "n_ref_all": 68, "n_fig": 15, "n_tab": 2, "L_tab": 400, "n_element_tab": 85, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 137516, "len_all_1": 55144, "len_abs": 1216, "len_title": 89, "len_sents": 49686, "len_sents_1": 27420, "n_sents": 354, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1221, "title": "When and how are modular networks better?", "abs": "Many real-world learning tasks have an underlying hierarchical modular structure, composed of smaller sub-functions. Traditional neural networks (NNs), however, often ignore this structure, leading to inefficiencies in learning and generalization. Leveraging known structural information can enhance performance by aligning the network architecture with the task\u2019s inherent modularity. In this work, we investigate how modular NNs can outperform traditional dense networks by systematically varying the degree of structural knowledge incorporated. We compare architectures ranging from monolithic dense NNs, which assume no prior knowledge, to hierarchically modular NNs with shared modules, which leverage sparsity, modularity, and module reusability. Our experiments demonstrate that incorporating structural knowledge, particularly through module reuse and fixed connectivity, significantly improves learning efficiency and generalization. Hierarchically modular NNs excel in data-scarce scenarios by promoting functional specialization within the modules and reducing redundancy. These findings suggest that task-specific architectural biases can lead to more efficient, interpretable, and effective learning systems.", "keywords": [ "Neural networks", "hierarchical modularity", "sparsity", "generalization", "training efficiency" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "OlRjxSuSwl", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Jianqun Zhou", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "MS student" }, { "name": "Qianqian Zheng", "gender": "unknown", "institution": "Shanghai Maritime University", "country": "CN", "position": "MS student" }, { "name": "Rui Meng", "gender": "Male", "institution": "SalesForce Research", "country": "", "position": "Researcher" }, { "name": "Shang Zeyuan", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Wei Chen", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Wei Zhang", "gender": "Male", "institution": "Eastern Institute of Technology, Ningbo, China.", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaoyu Shen", "gender": "Male", "institution": "Amazon", "country": "DE", "position": "machine learning scientist" }, { "name": "Yuanlei Zheng", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 31, "n_ref": 59, "n_ref_all": 82, "n_fig": 5, "n_tab": 15, "L_tab": 11729, "n_element_tab": 930, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1519, "n_element_tab_1": 227, "formula_len_all": 824, "formula_len_all_1": 849, "len_all": 168832, "len_all_1": 72568, "len_abs": 1744, "len_title": 79, "len_sents": 48472, "len_sents_1": 33364, "n_sents": 353, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 970, "title": "Beyond Content Relevance: Evaluating Instruction Following in Retrieval Models", "abs": "Large language models (LLMs) have been widely adopted for training embedding and ranking models, with recent advancements significantly improving the performance of Information Retrieval systems. However, while instruction-following is a core capability of LLMs, their ability to handle detailed user instructions has not been thoroughly investigated in search models. This study evaluates the instruction-following capabilities of various retrieval models, including LLM-based dense retrieval and reranking models. We develop a specialized benchmark InFoSearch spanning six dimensions: Audience, Keyword, Format, Language, Length, and Source, and introduce novel metrics to assess the models' responsiveness to instructions. Our findings show that even fine-tuned retrieval models struggle with instruction-following, highlighting a key limitation in current systems and providing valuable insights for improving their instruction-aware capabilities in future research.", "keywords": [ "LLM", "Instruction-Following", "Retrieval Model", "Benchmark" ], "rating_list": [ 3, 5, 6, 6, 8 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "Ok7ZH2Cyd7", "primary_area": "reinforcement learning", "n_author": 1, "author_info_list": [ { "name": "Ezgi Korkmaz", "gender": "unknown", "institution": "DeepMind", "country": "GB", "position": "Researcher" } ], "n_formula": 30, "n_formula_1": 28, "n_ref_uni": 24, "n_ref": 39, "n_ref_all": 55, "n_fig": 4, "n_tab": 4, "L_tab": 1379, "n_element_tab": 40, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1273, "n_element_tab_1": 45, "formula_len_all": 2553, "formula_len_all_1": 2313, "len_all": 106960, "len_all_1": 58799, "len_abs": 915, "len_title": 127, "len_sents": 32056, "len_sents_1": 26542, "n_sents": 202, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 917, "title": "A Large-Scale Analysis on Methodological Choices in Deep Reinforcement Learning", "abs": "Deep reinforcement learning research has been the center of remarkable scientific progress for the past decade. From winning one of the most challenging games to algorithmic advancements that allowed solving problems without even explicitly knowing the rules of the task at hand reinforcement learning research progress has been the epicenter of many breakthrough ideas. In this paper, we analyze the methodological issues in deep reinforcement learning. We introduce the theoretical foundations of the underlying causes outlining that the asymptotic performance of deep reinforcement learning algorithms does not have a monotone relationship to the performance in the regimes where data becomes scarce. The extensive large-scale empirical analysis provided in our paper discovers that a major line of deep reinforcement learning research under the canonical methodological choices resulted in suboptimal conclusions.", "keywords": [ "scientific analysis", "methodological choices" ], "rating_list": [ 3, 3, 6, 3, 6 ], "soundness_list": [ 1, 3, 4, 2, 3 ], "presentation_list": [ 2, 1, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 5, 3, 3 ] }, { "paper_id": "OjP6LUrw1O", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "AJAY KUMAR JAISWAL", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Jiawei Zhao", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Lu Yin", "gender": "unknown", "institution": "University of Surrey", "country": "GB", "position": "Assistant Professor" }, { "name": "Shiwei Liu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Yuandong Tian", "gender": "Male", "institution": "Meta AI (FAIR)", "country": "", "position": "Research Scientist" }, { "name": "Zhangyang Wang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Zhenyu Zhang", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 39, "n_ref": 92, "n_ref_all": 108, "n_fig": 7, "n_tab": 6, "L_tab": 2162, "n_element_tab": 204, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 556, "n_element_tab_1": 73, "formula_len_all": 86, "formula_len_all_1": 85, "len_all": 125900, "len_all_1": 63192, "len_abs": 1835, "len_title": 136, "len_sents": 38335, "len_sents_1": 32243, "n_sents": 216, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1841, "title": "From GaLore to WeLore: How Low-Rank Weights Non-uniformly Emerge from Low-Rank Gradients", "abs": "Modern Large Language Models (LLMs) are composed of matrices with billions of elements, making their storage and processing quite demanding in terms of computational resources and memory usage. Being significantly large, such matrices can often be expressed in low-rank format with potential to relax resource requirements. Unlike prior works which focus on developing novel matrix decomposition algorithms, in this work we first study the emergence of low-rank structures across matrices within different layers of LLMs and establish a consequential relationship between the gradient dynamics and emerging low-rank expressiveness of matrices. Our findings reveal that different layers exhibit varying levels of converged low-rank structure, necessitating a non-uniform rank reduction across them to minimize performance drop due to compression. In view of that, we present Weight Low-Rank Projection (WeLore) that unifies weight compression and memory-efficient fine-tuning as ONE, in a data-agnostic and one-shot way. WeLore capitalizes the heavy-tail distribution of singular values to identify a suitable rank reduction ratio for matrices within LLMs. Going beyond only as a compression technique, WeLore categorizes weight matrices into Low-rank Components (LRCs) and Non-Low-rank Components (N-LRCs) based on their ability to express themselves as low-rank. Our gradient perspective and extensive experiments illustrate that LRCs tend to have better finetuning capabilities and can closely mimic (sometimes outperform) the training loss trajectory and performance of full-finetuning with notable memory and compute footprint reduction. For example, finetuning a 50% compressed LLaMa-2 7B model using only a fraction of parameters in LRCs (WeLore) can outperform its full finetuning with ~3x better throughput and ~0.6x GPU requirement.", "keywords": [ "Large language models", "LLM finetuning", "Memory-efficient training", "Optimization", "Low Rank Compression" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "OjCWG58ZyY", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Shiying Sun", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Assistant Professor" }, { "name": "Tian Zhang", "gender": "unknown", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Yifan Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "Full Professor" }, { "name": "Zengmao Wang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Associate Professor " }, { "name": "Bo Du", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 45, "n_ref": 71, "n_ref_all": 102, "n_fig": 14, "n_tab": 1, "L_tab": 772, "n_element_tab": 29, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1250, "formula_len_all_1": 1020, "len_all": 130972, "len_all_1": 67241, "len_abs": 1151, "len_title": 112, "len_sents": 42083, "len_sents_1": 32001, "n_sents": 328, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1152, "title": "Goal-Conditioned Reinforcement Learning with Virtual Experiences", "abs": "Goal-conditioned reinforcement learning often employs a technique known as Hindsight Experience Replay (HER) for data augmentation by relabeling goals. However, HER limits goal relabeling to a single trajectory, which hinders the utilization of experiences from diverse trajectories. To address this issue, we present a curriculum learning method to construct virtual experiences, incorporating actual state transitions and virtual goals selected from the replay buffer. Considering that virtual experiences may contain a lot of noise, we also propose a self-supervised subgoal planning method that guides the learning of virtual experiences by imitating the subgoal-conditioned policy. Our intuition is that achieving a virtual goal may be challenging for the goal-conditioned policy, whereas simplified subgoals can provide effective guidance. We empirically show that the virtual experiences from diverse historical trajectories significantly boost the sample-efficiency compared to the existing goal-conditioned reinforcement learning and hierarchical reinforcement learning methods, even enabling the agent to learn tasks it has never experienced.", "keywords": [ "Subgoal Planning", "Curriculum Learning", "Imitating Learning" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 4 ], "confidence_list": [ 3, 3, 2, 5 ] }, { "paper_id": "OjAU0LLDbe", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Aldo Pacchiano", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "Assistant Professor" }, { "name": "Ambuj Tewari", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "Full Professor" }, { "name": "Chinmaya Kausik", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Mirco Mutti", "gender": "unknown", "institution": "Technion - Israel Institute of Technology", "country": "IL", "position": "Postdoc" } ], "n_formula": 110, "n_formula_1": 12, "n_ref_uni": 46, "n_ref": 102, "n_ref_all": 116, "n_fig": 1, "n_tab": 18, "L_tab": 5322, "n_element_tab": 452, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 910, "n_element_tab_1": 16, "formula_len_all": 11029, "formula_len_all_1": 931, "len_all": 338453, "len_all_1": 81119, "len_abs": 1654, "len_title": 116, "len_sents": 104796, "len_sents_1": 36906, "n_sents": 944, "n_sents_1": 305, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1672, "title": "A Theoretical Framework for Partially-Observed Reward States in RLHF", "abs": "The growing deployment of reinforcement learning from human feedback (RLHF) calls for a deeper theoretical investigation of its underlying models. The prevalent models of RLHF do not account for neuroscience-backed, partially-observed \"internal states'' that can affect human feedback, nor do they accommodate intermediate feedback during an interaction. Both of these can be instrumental in speeding up learning and improving alignment. To address these limitations, we model RLHF as reinforcement learning with partially observed reward-states (PORRL). We accommodate two kinds of feedback — cardinal and dueling feedback. We first demonstrate that PORRL subsumes a wide class of RL problems, including traditional RL, RLHF, and reward machines. For cardinal feedback, we present two model-based methods (POR-UCRL, POR-UCBVI). We give both cardinal regret and sample complexity guarantees for the methods, showing that they improve over naive history-summarization. We then discuss the benefits of a model-free method like GOLF with naive history-summarization in settings with recursive internal states and dense intermediate feedback. For this purpose, we define a new history aware version of the Bellman-eluder dimension and give a new guarantee for GOLF in our setting, which can be exponentially sharper in illustrative examples. For dueling feedback, we show that a naive reduction to cardinal feedback fails to achieve sublinear dueling regret. We then present the first explicit reduction that converts guarantees for cardinal regret to dueling regret. In both feedback settings, we show that our models and guarantees generalize and extend existing ones.", "keywords": [ "reinforcement learning", "learning theory", "reinforcement learning theory", "reinforcement learning from human feedback", "eluder dimension", "partial observability" ], "rating_list": [ 5, 6, 8 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 4 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "OioOio3bmx", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Dejing Dou", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Loi Duc Huynh", "gender": "Male", "institution": "Auburn University", "country": "", "position": "PhD student" }, { "name": "Ruoming Jin", "gender": "Male", "institution": "Kent State University", "country": "", "position": "Professor" }, { "name": "Tianshi Che", "gender": "Male", "institution": "Auburn University", "country": "US", "position": "PhD student" }, { "name": "Yang Zhou", "gender": "unknown", "institution": "Auburn University", "country": "US", "position": "Associate Professor" }, { "name": "Zijie Zhang", "gender": "Male", "institution": "University of Texas at San Antonio", "country": "US", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 9, "n_ref_uni": 22, "n_ref": 47, "n_ref_all": 84, "n_fig": 9, "n_tab": 7, "L_tab": 3679, "n_element_tab": 275, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 599, "n_element_tab_1": 29, "formula_len_all": 3095, "formula_len_all_1": 1305, "len_all": 126737, "len_all_1": 69551, "len_abs": 1519, "len_title": 156, "len_sents": 37042, "len_sents_1": 28779, "n_sents": 336, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1566, "title": "k-Odd One Clear (k-OOC), a novel GPU kernel that improves quantization accuracy and speed of GPTQ algorithm", "abs": "Large Language Model (LLM) demonstrated tremendously useful applications in nowadays fast-evolving AI driven technology. As the model sizes grow bigger, the demand for bigger and faster GPU is required. Another way to alleviate this issue is by improving the compression of the trained model through quantization so that lower VRAM devices can run. Quantization paradigms like GPTQ, PB-LLM, BiLLM (Hessian based with structural searching) are successful quantize mechanisms. In this paper, we propose **OOC**, a technique to pick an \"odd\" group to improve the quantization clarity so that the model can have better reasoning capability overall. In addition, we define **Bit Family** ($A^{lim},A^{max}$) to classify compression rate of current and past quantizing techniques, thus providing a more objective way to rank different methodologies in literature. Thirdly, to avoid compromising the quantization speed due to the **scanning** process overhead, we developed a specialized fused GPU kernel (k-OOC) where it can be $9\\times$ faster than the original GPTQ implementation (single-flow mode) and $22\\times$ faster than the naive OOC implementation (double-flow mode) due to the incorporation of techniques called **Row-Flow-Selection Parallel** and **Input Batching**. We measured perplexity of k-OOC (2 bits) with 14 major models like OPT, LLAMA, and Bloom (125M to 70B parameters) and popular datasets ( Wikitext2, C4, and PTB). We managed to improved the perplexity of small model by 8.9\\% and of big model by 4.1\\% compared to the baseline of GPTQ (2 bits).", "keywords": [ "quantization", "LLM", "GPTQ", "BitNet" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 2, 3, 3 ] }, { "paper_id": "OiQttMHwce", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Hoang V. Tran", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Huyen Trang Pham", "gender": "Female", "institution": "VinAI Research", "country": "VN", "position": "Intern" }, { "name": "Minh-Khoi Nguyen-Nhat", "gender": "Male", "institution": "FPT", "country": "VN", "position": "Researcher" }, { "name": "Tam Le", "gender": "Male", "institution": "The Institute of Statistical Mathematics (ISM)", "country": "JP", "position": "Assistant Professor" }, { "name": "Tan Minh Nguyen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Thanh Chu", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Undergrad student" } ], "n_formula": 104, "n_formula_1": 41, "n_ref_uni": 40, "n_ref": 86, "n_ref_all": 111, "n_fig": 11, "n_tab": 6, "L_tab": 2394, "n_element_tab": 185, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 352, "n_element_tab_1": 33, "formula_len_all": 11766, "formula_len_all_1": 3162, "len_all": 174908, "len_all_1": 66780, "len_abs": 1967, "len_title": 95, "len_sents": 48315, "len_sents_1": 26721, "n_sents": 454, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1976, "title": "Distance-Based Tree-Sliced Wasserstein Distance", "abs": "To overcome computational challenges of Optimal Transport (OT), several variants of Sliced Wasserstein (SW) has been developed in the literature. These approaches exploit the closed-form expression of the univariate OT by projecting measures onto one-dimensional lines. However, projecting measures onto low-dimensional spaces can lead to a loss of topological information. Tree-Sliced Wasserstein distance on Systems of Lines (TSW-SL) has emerged as a promising alternative that replaces these lines with a more intricate structure called tree systems. The tree structures enhance the ability to capture topological information of the metric while preserving computational efficiency. However, at the core of TSW-SL, the splitting maps, which serve as the mechanism for pushing forward measures onto tree systems, focus solely on the position of the measure supports while disregarding the projecting domains. Moreover, the specific splitting map used in TSW-SL leads to a metric that is not invariant under Euclidean transformations, a typically expected property for OT on Euclidean space. In this work, we propose a novel class of splitting maps that generalizes the existing one studied in TSW-SL enabling the use of all positional information from input measures, resulting in a novel Distance-based Tree-Sliced Wasserstein (Db-TSW) distance. In addition, we introduce a simple tree sampling process better suited for Db-TSW, leading to an efficient GPU-friendly implementation for tree systems, similar to the original SW. We also provide a comprehensive theoretical analysis of proposed class of splitting maps to verify the injectivity of the corresponding Radon Transform, and demonstrate that Db-TSW is an Euclidean invariant metric. We empirically show that Db-TSW significantly improves accuracy compared to recent SW variants while maintaining low computational cost via a wide range of experiments on gradient flows, image style transfer, and generative models.", "keywords": [ "tree-sliced wasserstein distance", "optimal transport", "equivariance" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "Oi47wc10sm", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Amit Dhurandhar", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Erik Miehling", "gender": "Male", "institution": "IBM Research", "country": "IE", "position": "Research Scientist" }, { "name": "Inkit Padhi", "gender": "unknown", "institution": "IBM Research", "country": "", "position": "Researcher" }, { "name": "Karthikeyan Natesan Ramamurthy", "gender": "unknown", "institution": "International Business Machines", "country": "US", "position": "Research Staff Member" }, { "name": "Manish Nagireddy", "gender": "Male", "institution": "IBM Research", "country": "US", "position": "Researcher" }, { "name": "Pierre Dognin", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Research Scientist" }, { "name": "Bruce W Lee", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Intern" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 61, "n_ref": 87, "n_ref_all": 127, "n_fig": 24, "n_tab": 6, "L_tab": 2635, "n_element_tab": 174, "n_fig_1": 19, "n_tab_1": 2, "L_tab_1": 675, "n_element_tab_1": 62, "formula_len_all": 537, "formula_len_all_1": 252, "len_all": 218227, "len_all_1": 70471, "len_abs": 2259, "len_title": 104, "len_sents": 52440, "len_sents_1": 32312, "n_sents": 410, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 56, "L_abs": 1128, "title": "Programming Refusal with Conditional Activation Steering", "abs": "LLMs have shown remarkable capabilities, but precisely controlling their response behavior remains challenging.\nExisting activation steering methods alter LLM behavior indiscriminately, limiting their practical applicability in settings where selective responses are essential, such as content moderation or domain-specific assistants.\nIn this paper, we propose Conditional Activation Steering (CAST), which analyzes LLM activation patterns during inference to selectively apply or withhold activation steering based on the input context.\nOur method is based on the observation that different categories of prompts activate distinct patterns in the model's hidden states.\nUsing CAST, one can systematically control LLM behavior with rules like \"if input is about hate speech or adult content, then refuse\" or \"if input is not about legal advice, then refuse.\"\nThis allows for selective modification of responses to specific content while maintaining normal responses to other content, all without requiring weight optimization.\nWe release an open-source implementation of our framework at .", "keywords": [ "Activation Engineering", "Safety", "Alignment", "Steering Vector" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 4, 3, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "OheAR2xrtb", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Boxuan Dong", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Chenrui Tie", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Chongkai Gao", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Hao Dong", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ruihai Wu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yue Chen", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Zeyi Li", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 43, "n_formula_1": 18, "n_ref_uni": 30, "n_ref": 60, "n_ref_all": 74, "n_fig": 8, "n_tab": 5, "L_tab": 2196, "n_element_tab": 310, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1896, "n_element_tab_1": 280, "formula_len_all": 4016, "formula_len_all_1": 2136, "len_all": 142933, "len_all_1": 56510, "len_abs": 407, "len_title": 118, "len_sents": 41886, "len_sents_1": 21879, "n_sents": 361, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1189, "title": "ET-SEED: EFFICIENT TRAJECTORY-LEVEL SE(3) EQUIVARIANT DIFFUSION POLICY", "abs": "Imitation learning, e.g., diffusion policy, has been proven effective in various robotic manipulation tasks.\nHowever, extensive demonstrations are required for policy robustness and generalization.\nTo reduce the demonstration reliance, we leverage spatial symmetry and propose ET-SEED, an efficient trajectory-level SE(3) equivariant diffusion model for generating action sequences in complex robot manipulation tasks.\nFurther, previous equivariant diffusion models require the per-step equivariance in the Markov process, making it difficult to learn policy under such strong constraints.\nWe theoretically extend equivariant Markov kernels and simplify the condition of equivariant diffusion process, thereby significantly improving training efficiency for trajectory-level SE(3) equivariant diffusion policy in an end-to-end manner.\nWe evaluate ET-SEED on representative robotic manipulation tasks, involving rigid body, articulated and deformable object.\nExperiments demonstrate superior data efficiency and manipulation proficiency of our proposed method,\nas well as its ability to generalize to unseen configurations with only a few demonstrations. Website: https://et-seed.github.io/", "keywords": [ "Robotics; Manipulation; Equivariance" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 4, 3, 3, 4 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "OhcWlo1M8q", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Anja Jankovic", "gender": "Not Specified", "institution": "Rheinisch Westf\u00e4lische Technische Hochschule Aachen", "country": "DE", "position": "Postdoc" }, { "name": "Elena Raponi", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "Postdoc" }, { "name": "Hadar Shavit", "gender": "Male", "institution": "Rheinisch Westf\u00e4lische Technische Hochschule Aachen", "country": "DE", "position": "PhD student" }, { "name": "Holger Hoos", "gender": "unknown", "institution": "Rheinisch Westf\u00e4lische Technische Hochschule Aachen", "country": "DE", "position": "Full Professor" }, { "name": "Thomas B\u00e4ck", "gender": "Male", "institution": "Leiden University, Leiden University", "country": "NL", "position": "Full Professor" } ], "n_formula": 25, "n_formula_1": 16, "n_ref_uni": 55, "n_ref": 102, "n_ref_all": 126, "n_fig": 10, "n_tab": 5, "L_tab": 1671, "n_element_tab": 241, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 366, "n_element_tab_1": 7, "formula_len_all": 1663, "formula_len_all_1": 918, "len_all": 172146, "len_all_1": 63455, "len_abs": 1310, "len_title": 78, "len_sents": 51148, "len_sents_1": 31567, "n_sents": 392, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1317, "title": "DensBO: Dynamic Ensembling of Surrogate Models for Hyperparameter Optimisation", "abs": "Hyperparameter optimisation (HPO) of machine learning models is crucial for achieving optimal performance for different tasks. Surrogate-based optimisation techniques, such as Bayesian optimisation (BO), have been successfully applied to tackle this problem. BO is subject to different design choices of its components. In particular, depending on the nature and the size of the search space, the choice of the surrogate model has a substantial impact on the overall performance of BO. Surrogate models in BO approximate the function to optimise and guide the search towards promising regions by predicting the function value for different solution candidates. Combining different machine learning (ML) models is known to lead to performance gains, e.g., in different prediction tasks. To this end, we propose a novel dynamic approach to ensemble surrogate models in the BO pipeline, leveraging the complementary powers of different surrogate models at different stages of the optimisation process. We empirically evaluate our method on numerous benchmarks and demonstrate its advantage compared to state-of-the-art single-surrogate BO baselines. We highlight the usefulness of our approach in finding good hyperparameter configurations in mixed (numerical and categorical) search spaces for a wide range of problems.", "keywords": [ "hyperparameter optimisation", "Bayesian optimisation", "surrogate models" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 4, 2, 1, 2 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "OhauMUNW8T", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Kuniko Saito", "gender": "unknown", "institution": "NTT", "country": "JP", "position": "Researcher" }, { "name": "Kyosuke Nishida", "gender": "Male", "institution": "NTT corporation", "country": "JP", "position": "Researcher" }, { "name": "Taku Hasegawa", "gender": "Male", "institution": "NTT corporation", "country": "JP", "position": "Researcher" }, { "name": "Yui Oka", "gender": "Female", "institution": "NTT Human Informatics Laboratories", "country": "JP", "position": "Researcher" } ], "n_formula": 54, "n_formula_1": 25, "n_ref_uni": 36, "n_ref": 83, "n_ref_all": 128, "n_fig": 9, "n_tab": 5, "L_tab": 1862, "n_element_tab": 336, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 1279, "n_element_tab_1": 240, "formula_len_all": 4662, "formula_len_all_1": 2031, "len_all": 196270, "len_all_1": 69715, "len_abs": 1624, "len_title": 104, "len_sents": 53471, "len_sents_1": 29244, "n_sents": 455, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1636, "title": "Wavelet-based Positional Representation for Long Context", "abs": "In the realm of large-scale language models, a significant challenge arises when extrapolating sequences beyond the maximum allowable length. \nThis is because the model's position embedding mechanisms are limited to positions encountered during training, thus preventing effective representation of positions in longer sequences.\nWe analyzed conventional position encoding methods for long contexts and found the following characteristics.\n(1) When the representation dimension is regarded as the time axis, Rotary Position Embedding (RoPE) can be interpreted as a restricted wavelet transform using Haar-like wavelets. \nHowever, because it uses only a fixed scale parameter, it does not fully exploit the advantages of wavelet transforms, which capture the fine movements of non-stationary signals using multiple scales (window sizes). \nThis limitation could explain why RoPE performs poorly in extrapolation.\n(2)\nPrevious research as well as our own analysis indicates that Attention with Linear Biases (ALiBi) functions similarly to windowed attention, using windows of varying sizes.\nHowever, it has limitations in capturing deep dependencies because it restricts the receptive field of the model.\nFrom these insights, we propose a new position representation method that captures multiple scales (i.e., window sizes) by leveraging wavelet transforms without limiting the model's attention field.\nExperimental results show that this new method improves the performance of the model in both short and long contexts. \nIn particular, our method allows extrapolation of position information without limiting the model's attention field.", "keywords": [ "Positional Encoding", "Extrapolation", "Wavelet Transform", "Transformers", "RoPE", "ALiBi", "NLP" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "OhUoTMxFIH", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Gonzalo Gonzalez-Pumariega", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Leong Su Yean", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Undergrad student" }, { "name": "Neha Sunkara", "gender": "Female", "institution": "Cornell University", "country": "US", "position": "Undergrad student" }, { "name": "Sanjiban Choudhury", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 73, "n_ref_all": 147, "n_fig": 28, "n_tab": 9, "L_tab": 25767, "n_element_tab": 197, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 10, "n_element_tab_1": 1, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 335383, "len_all_1": 51898, "len_abs": 3758, "len_title": 110, "len_sents": 105739, "len_sents_1": 25513, "n_sents": 1123, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1179, "title": "Robotouille: An Asynchronous Planning Benchmark for LLM Agents", "abs": "Effective asynchronous planning, or the ability to efficiently reason and plan over states and actions that must happen in parallel or sequentially, is essential for agents that must account for time delays, reason over diverse long-horizon tasks, and collaborate with other agents. While large language model (LLM) agents show promise in high-level task planning, current benchmarks focus primarily on short-horizon tasks and do not evaluate such asynchronous planning capabilities. We introduce Robotouille, a challenging benchmark environment designed to test LLM agents' ability to handle asynchronous, long-horizon, and multi-agent scenarios. These datasets capture increasingly complex planning challenges that go beyond existing benchmarks, particularly in their requirement for agents to manage overlapping tasks, interruptions, and collaboration. Our results show that ReAct (gpt-4o) achieves 47% on synchronous tasks but only 11% on asynchronous tasks, highlighting significant room for improvement. We further analyze failure modes, demonstrating the need for LLM agents to better incorporate long-horizon feedback and self-audit their reasoning during task execution.", "keywords": [ "benchmark", "llm", "agents", "planning" ], "rating_list": [ 6, 8, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "Oh8MuCacJW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Chenghao Xu", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jiexi Yan", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Muli Yang", "gender": "unknown", "institution": "A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Cheng Deng", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "GuangtaoLyu", "gender": "Male", "institution": "Xidian University of Electronic Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 36, "n_ref": 87, "n_ref_all": 115, "n_fig": 12, "n_tab": 14, "L_tab": 3594, "n_element_tab": 405, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1050, "n_element_tab_1": 69, "formula_len_all": 839, "formula_len_all_1": 658, "len_all": 170407, "len_all_1": 66257, "len_abs": 1607, "len_title": 141, "len_sents": 47908, "len_sents_1": 31789, "n_sents": 326, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1612, "title": "Towards Unified Human Motion-Language Understanding via Sparse Interpretable Characterization", "abs": "Recently, the comprehensive understanding of human motion has been a prominent area of research due to its critical importance in many fields. However, existing methods often prioritize specific downstream tasks and roughly align text and motion features within a CLIP-like framework. This results in a lack of rich semantic information which restricts a more profound comprehension of human motions, ultimately leading to unsatisfactory performance.\nTherefore, we propose a novel motion-language representation paradigm to enhance the interpretability of motion representations by constructing a universal motion-language space, where both motion and text features are concretely lexicalized, ensuring that each element of features carries specific semantic meaning.\nSpecifically, we introduce a multi-phase strategy mainly comprising Lexical Bottlenecked Masked Language Modeling to enhance the language model's focus on high-entropy words crucial for motion semantics, Contrastive Masked Motion Modeling to strengthen motion feature extraction by capturing spatiotemporal dynamics directly from skeletal motion, Lexical Bottlenecked Masked Motion Modeling to enable the motion model to capture the underlying semantic features of motion for improved cross-modal understanding, and Lexical Contrastive Motion-Language Pretraining to align motion and text lexicon representations, thereby ensuring enhanced cross-modal coherence.\nComprehensive analyses and extensive experiments across multiple public datasets demonstrate that our model achieves state-of-the-art performance across various tasks and scenarios.", "keywords": [ "Interpretable Human Motion Understanding", "visualization and interpretation of motion representations", "Human Motion representation learning" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Og7ZZd7hDm", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Feihu Huang", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Siqi Zhang", "gender": "unknown", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Songcan Chen", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Wang Xinrui", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 218, "n_formula_1": 45, "n_ref_uni": 23, "n_ref": 64, "n_ref_all": 101, "n_fig": 5, "n_tab": 6, "L_tab": 1673, "n_element_tab": 104, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 220, "n_element_tab_1": 19, "formula_len_all": 43894, "formula_len_all_1": 5253, "len_all": 197413, "len_all_1": 66565, "len_abs": 1383, "len_title": 138, "len_sents": 38658, "len_sents_1": 23245, "n_sents": 379, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1435, "title": "Faster Adaptive Momentum-Based Federated Methods for Distributed Composition Optimization", "abs": "Federated learning is a popular distributed learning paradigm in machine learning. Meanwhile, composition optimization is an effective hierarchical learning model, which appears in many machine learning applications such as meta learning and robust learning. More recently, although a few federated composition optimization algorithms have been proposed, they still suffer from high sample and communication complexities. In the paper, thus, we propose a class of faster adaptive federated compositional optimization algorithms (i.e., MFCGD and AdaMFCGD) to solve the nonconvex distributed composition problems, which builds on the momentum-based variance reduced and local-SGD techniques. In particular, our adaptive algorithm (i.e., AdaMFCGD) uses a unified adaptive matrix to flexibly incorporate various adaptive learning rates. Moreover, we provide a solid theoretical analysis for our algorithms under non-i.i.d. setting, and prove our algorithms obtain a lower sample and communication complexities simultaneously than the existing federated composition optimization algorithms. Specifically, our algorithms obtain lower sample complexity of $\\tilde{O}(\\epsilon^{-3})$ with lower communication complexity of $\\tilde{O}(\\epsilon^{-2})$ in finding an $\\epsilon$-stationary solution. We conduct numerical experiments on robust federated learning and distributed meta learning tasks to demonstrate the efficiency of our algorithms.", "keywords": [ "Federated Learning", "Composition Optimization", "Adaptive Learning Rate" ], "rating_list": [ 6, 3, 1, 3 ], "soundness_list": [ 4, 2, 1, 2 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "OfsGBHdFgk", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Jiaxin Liu", "gender": "Female", "institution": "Lehigh University", "country": "US", "position": "PhD student" }, { "name": "XIN JIANG", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Postdoc" } ], "n_formula": 113, "n_formula_1": 27, "n_ref_uni": 21, "n_ref": 58, "n_ref_all": 96, "n_fig": 6, "n_tab": 6, "L_tab": 3712, "n_element_tab": 172, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2010, "n_element_tab_1": 188, "formula_len_all": 18962, "formula_len_all_1": 2631, "len_all": 157902, "len_all_1": 70356, "len_abs": 1479, "len_title": 127, "len_sents": 38843, "len_sents_1": 27764, "n_sents": 389, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1436, "title": "Improved Stochastic Controlled Averaging for Distributed and Federated Learning", "abs": "Distributed and federated learning (D/FL) is a powerful machine learning (ML) paradigm in which clients collaborate to train a model under the coordination of a central server. Depending on the nature of clients, data in each client might have the same distribution (called the homogeneous setting) or different distributions (the heterogeneous setting). The state-of-the-art D/FL algorithm SCAFFOLD addresses the critical issue of data heterogeneity through the use of control variables. However, while theoretical analysis suggests that the convergence rate of SCAFFOLD is independent of data heterogeneity, the practical performance of SCAFFOLD is often inconsistent in homogeneous and heterogeneous settings. Motivated by the disagreement between theory and practice of SCAFFOLD, in this work, we propose a novel D/FL algorithm to bridge this experimental performance gap while preserving similar theoretical guarantees as SCAFFOLD. The proposed algorithm accommodates arbitrary data heterogeneity, partial participation, local updates, and supports unbiased communication compression. Theoretically, we prove that our algorithm is unaffected by data heterogeneity and achieves state-of-the-art convergence rate as SCAFFOLD. Furthermore, numerical experiments indicate that our algorithm achieves consistent (similar) test accuracy in both homogeneous and heterogeneous settings while often converges faster than existing baselines.", "keywords": [ "Federated Learning" ], "rating_list": [ 5, 5, 3, 3, 3 ], "soundness_list": [ 4, 2, 2, 1, 2 ], "presentation_list": [ 4, 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 1, 2 ], "confidence_list": [ 3, 2, 4, 5, 3 ] }, { "paper_id": "OfjIlbelrT", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Jianqiao Lu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xun Zhou", "gender": "unknown", "institution": "bytedance Inc", "country": "", "position": "Researcher" }, { "name": "Xunhao Lai", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yao Luo", "gender": "Female", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Yiyuan Ma", "gender": "Female", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 13, "n_ref_uni": 30, "n_ref": 94, "n_ref_all": 126, "n_fig": 10, "n_tab": 11, "L_tab": 6776, "n_element_tab": 627, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1480, "n_element_tab_1": 90, "formula_len_all": 3259, "formula_len_all_1": 1679, "len_all": 250303, "len_all_1": 54320, "len_abs": 1731, "len_title": 94, "len_sents": 43087, "len_sents_1": 24453, "n_sents": 320, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 93, "L_abs": 1531, "title": "FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference", "abs": "Large language models (LLMs) encounter computational challenges during long-sequence inference, especially in the attention pre-filling phase, where the complexity grows quadratically with the prompt length. Previous efforts to mitigate these challenges have relied on fixed sparse attention patterns or identifying sparse attention patterns based on limited cases. However, these methods lacked the flexibility to efficiently adapt to varying input demands. In this paper, we introduce FlexPrefill, a Flexible sparse Pre-filling mechanism that dynamically adjusts sparse attention patterns and computational budget in real-time to meet the specific requirements of each input and attention head. The flexibility of our method is demonstrated through two key innovations: 1) Query-Aware Sparse Pattern Determination: By measuring Jensen-Shannon divergence, this component adaptively switches between query-specific diverse attention patterns and predefined attention patterns. 2) Cumulative-Attention Based Index Selection: This component dynamically selects query-key indexes to be computed based on different attention patterns, ensuring the sum of attention scores meets a predefined threshold.\nFlexPrefill adaptively optimizes the sparse pattern and sparse ratio of each attention head based on the prompt, enhancing efficiency in long-sequence inference tasks. Experimental results show significant improvements in both speed and accuracy over prior methods, providing a more flexible and efficient solution for LLM inference.", "keywords": [ "Large Language Models (LLMs)", "LLM inference", "Long-context LLMs", "Sparse Attention Mechanism" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "OfCtAsnxvT", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Kerstin Ritter", "gender": "Female", "institution": "Charit\u00e9 - Universit\u00e4tsmedizin Berlin", "country": "", "position": "Assistant Professor" }, { "name": "Sam Gijsen", "gender": "Not Specified", "institution": "Charit\u00e9 - Universit\u00e4tsmedizin Berlin", "country": "DE", "position": "Postdoc" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 49, "n_ref": 104, "n_ref_all": 131, "n_fig": 11, "n_tab": 10, "L_tab": 6049, "n_element_tab": 343, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 176, "n_element_tab_1": 22, "formula_len_all": 597, "formula_len_all_1": 596, "len_all": 198505, "len_all_1": 57219, "len_abs": 1671, "len_title": 119, "len_sents": 59690, "len_sents_1": 28045, "n_sents": 465, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1679, "title": "EEG-Language Pretraining for Highly Label-Efficient Pathology Detection", "abs": "Multimodal language modeling constitutes a recent breakthrough which leverages advances in large language models to pretrain capable multimodal models. The integration of natural language during pretraining has been shown to significantly improve learned representations, particularly in computer vision. However, the efficacy of multimodal language modeling in the realm of functional brain data, specifically for advancing pathology detection, remains unexplored. This study pioneers EEG-language models (ELMs) trained on clinical reports and 15000 EEGs. We propose to combine multimodal alignment in this novel domain with timeseries cropping and text segmentation. This also enables an extension based on multiple instance learning to alleviate misalignment between irrelevant EEG or text segments. Our results indicate that models learn richer representations from being exposed to a variety of report segments, including the patient's clinical history, description of the EEG, and the physician's interpretation. Compared to models exposed to narrower clinical text information, we find such models to retrieve EEGs based on clinical reports (and vice versa) with substantially higher accuracy. Particularly in regimes with few annotations, we observe that ELMs can significantly improve pathology detection compared to EEG-only models, as demonstrated by both zero-shot classification and linear probes. The integration of multiple instance learning further improves performance across tasks. In sum, these results highlight the potential of integrating brain activity data with clinical text, suggesting that ELMs represent significant progress for clinical applications.", "keywords": [ "deep learning", "multimodal", "neuroscience", "eeg", "medical" ], "rating_list": [ 5, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Of6tH5yfmK", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Dan Zhang", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Difan Zou", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Fan Yin", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Min Cai", "gender": "unknown", "institution": "Shenzhen University", "country": "CN", "position": "MS student" }, { "name": "Shichang Zhang", "gender": "Male", "institution": "Harvard Business School", "country": "US", "position": "Postdoc" }, { "name": "Yisong Yue", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Yuchen Zhang", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "PhD student" }, { "name": "Ziniu Hu", "gender": "Male", "institution": "xAI", "country": "US", "position": "Member of Technical Staff" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 41, "n_ref": 93, "n_ref_all": 127, "n_fig": 6, "n_tab": 39, "L_tab": 8580, "n_element_tab": 524, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2947, "n_element_tab_1": 197, "formula_len_all": 670, "formula_len_all_1": 521, "len_all": 247589, "len_all_1": 64874, "len_abs": 1295, "len_title": 132, "len_sents": 72440, "len_sents_1": 26861, "n_sents": 586, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1260, "title": "Self-Control of LLM Behaviors by Compressing Suffix Gradient into Prefix Controller", "abs": "We propose $SelfControl$, an inference-time model control method utilizing gradients to control the behavior of large language models (LLMs) without explicit human annotations. Given a desired behavior expressed in a natural language suffix string concatenated to the input prompt, $SelfControl$ computes gradients of the LLM's self-evaluation of the suffix with respect to its latent representations. The gradients are used to directly control the auto-regressive generation process towards desired behaviors, which eliminates human supervision, achieves precise and transparent control, and offers on-the-fly adaptability. To further enhance efficiency, we introduce $SelfControl_{prefix}$, a compact module that encapsulates the learned representations from gradients into a \\pc, facilitating efficient inference-time control with no latency compared to the original model and allowing control for multiple behaviors simultaneously. Our experiments demonstrate $SelfControl$'s efficacy across multiple domains, where it improves over SOTA for **8.3\\%** in detoxification, **3.1\\%** in truthfulness enhancement, **4\\%$\\textasciitilde$10\\%** in controlling on emotion tones, and **48.2\\%** in privacy protection, i.e., completely remove privacy leakage issue.", "keywords": [ "Large Language Models", "LLM Steering", "Representation Engineering" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 1, 3, 2 ], "confidence_list": [ 2, 3, 4 ] }, { "paper_id": "Of5F2GdGLA", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Boxun Li", "gender": "Male", "institution": "Infinigence-AI", "country": "", "position": "Principal Researcher" }, { "name": "Lin Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Qingmin Liao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Xinyi Yang", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yu Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Zhuyu Yao", "gender": "Male", "institution": "Infinigence-AI", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 110, "n_ref_all": 134, "n_fig": 8, "n_tab": 13, "L_tab": 3922, "n_element_tab": 320, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1017, "n_element_tab_1": 137, "formula_len_all": 466, "formula_len_all_1": 0, "len_all": 320369, "len_all_1": 61251, "len_abs": 1305, "len_title": 156, "len_sents": 131757, "len_sents_1": 29215, "n_sents": 1046, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1077, "title": "VeSX: A Framework Featured by Verification, Self-Correction and In-context Learning for Web Automation Tasks", "abs": "While large language models have achieved remarkable success in tasks such as reasoning and question answering, applying LLMs to interactive tasks like web automation remains challenging. In web automation, existing planning-execution workflow often faces limitations due to the infeasible subtasks. We propose VeSX, a framework designed to enhance subtask feasibility through verification, self-correction, and in-context learning. VeSX introduces three key improvements: (1) subgoal-guided verification, which verifies the execution results of subtasks based on the preset subgoals; (2) hierarchical self-correction, which combines reflection and replanning, targeting to self-correct mistakes in both planning and execution phases; (3) exemplar bank, which improves in-context learning by partitioning execution trajectories and heuristically generating metadata for exemplars. We evaluate VeSX on WebArena benchmark and achieve the state-of-the-art average success rate of 0.34, which significantly outperforms existing methods without human guidance on all five scenarios.", "keywords": [ "LLM agent", "web automation" ], "rating_list": [ 3, 3, 6, 5, 6 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 1, 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "OegBJMucyM", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Amrith Setlur", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Aviral Kumar", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Claire Tomlin", "gender": "unknown", "institution": "UC Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Dibya Ghosh", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Jacob Steinhardt", "gender": "unknown", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" }, { "name": "Katie Kang", "gender": "Female", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 32, "n_ref": 47, "n_ref_all": 60, "n_fig": 7, "n_tab": 2, "L_tab": 1593, "n_element_tab": 133, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 488, "formula_len_all_1": 459, "len_all": 113442, "len_all_1": 62946, "len_abs": 1520, "len_title": 129, "len_sents": 38029, "len_sents_1": 32495, "n_sents": 241, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1526, "title": "Pre-Memorization Train Accuracy Reliably Predicts Generalization in LLM Reasoning", "abs": "When large language models (LLMs) are finetuned on reasoning tasks, they can either reduce their training loss by developing problem-solving abilities, or by simply memorizing target traces in the training data. Our work aims to better understand how this learning process shapes a model's ability to generalize. We observe that, while LLMs often perfectly memorize most target solution traces by the end of training, their predictions at intermediate checkpoints can provide valuable insights into their behavior at test time. Concretely, we introduce the concept of pre-memorization train accuracy: the accuracy of model samples for training queries prior to exactly reproducing reasoning traces in the training data. We find that the average pre-memorization train accuracy of the model is strongly predictive of its test performance, with coefficients of determination around or exceeding 0.9 across various models (Llama3-8B, Gemma2-9B), datasets (GSM8k, MATH), and training setups. Beyond this aggregate statistic, we find that the pre-memorization train accuracy of individual examples can predict the model\u2019s sensitivity to input perturbations for those examples, allowing us to identify examples for which the model fails to learn robust solutions. A natural application of this insight is in data curation. We find that prioritizing the collection of examples with low pre-memorization accuracy leads to 1.5-2x data efficiency compared to i.i.d. data scaling, and outperforms other standard data curation techniques.", "keywords": [ "LLMs", "Generalization", "Memorization", "Reasoning" ], "rating_list": [ 3, 3, 3, 8 ], "soundness_list": [ 2, 1, 2, 4 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "Oeb0I3JcVc", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Mohsen Bayati", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "" }, { "name": "Yuwei Luo", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" } ], "n_formula": 32, "n_formula_1": 32, "n_ref_uni": 30, "n_ref": 56, "n_ref_all": 63, "n_fig": 4, "n_tab": 4, "L_tab": 372, "n_element_tab": 6, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 63, "n_element_tab_1": 3, "formula_len_all": 3096, "formula_len_all_1": 2873, "len_all": 103164, "len_all_1": 70379, "len_abs": 1266, "len_title": 96, "len_sents": 29913, "len_sents_1": 27390, "n_sents": 257, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1297, "title": "Geometry-Aware Approaches for Balancing Performance and Theoretical Guarantees in Linear Bandits", "abs": "This paper is motivated by recent research in the $d$-dimensional stochastic linear bandit literature, which has revealed an unsettling discrepancy: algorithms like Thompson sampling and Greedy demonstrate promising empirical performance, yet this contrasts with their pessimistic theoretical regret bounds. The challenge arises from the fact that while these algorithms may perform poorly in certain problem instances, they generally excel in typical instances. To address this, we propose a new data-driven technique that tracks the geometric properties of the uncertainty ellipsoid around the main problem parameter. This methodology enables us to formulate a data-driven frequentist regret bound, which incorporates the geometric information, for a broad class of base algorithms, including Greedy, OFUL, and Thompson sampling. This result allows us to identify and ``course-correct\" problem instances in which the base algorithms perform poorly. The course-corrected algorithms achieve the minimax optimal regret of order $\\tilde{\\mathcal{O}}(d\\sqrt{T})$ for a $T$-period decision-making scenario, effectively maintaining the desirable attributes of the base algorithms, including their empirical efficacy. We present simulation results to validate our findings using synthetic and real data.", "keywords": [ "Linear bandit", "Thompson sampling", "Greedy", "Data-driven exploration" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 5, 4, 2, 4 ] }, { "paper_id": "OeKp3AdiVO", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Han Lu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Liqing Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Siyu Sun", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Tianjiao Li", "gender": "Male", "institution": "bilibili.inc", "country": "CN", "position": "Principal Researcher" }, { "name": "Xiaokang Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University, China", "country": "CN", "position": "Professor" }, { "name": "Yichen Xie", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "jiangtong li", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Postdoc" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 25, "n_formula_1": 20, "n_ref_uni": 30, "n_ref": 159, "n_ref_all": 169, "n_fig": 6, "n_tab": 6, "L_tab": 3364, "n_element_tab": 326, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 219, "n_element_tab_1": 19, "formula_len_all": 1276, "formula_len_all_1": 1068, "len_all": 147910, "len_all_1": 61101, "len_abs": 1615, "len_title": 139, "len_sents": 39754, "len_sents_1": 28263, "n_sents": 327, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1649, "title": "Rethinking Classifier Re-Training in Long-Tailed Recognition: Label Over-Smooth Can Balance", "abs": "In the field of long-tailed recognition, the Decoupled Training paradigm has shown exceptional promise by dividing training into two stages: representation learning and classifier re-training. While previous work has tried to improve both stages simultaneously, this complicates isolating the effect of classifier re-training. Recent studies reveal that simple regularization can produce strong feature representations, highlighting the need to reassess classifier re-training methods. In this study, we revisit classifier re-training methods based on a unified feature representation and re-evaluate their performances. \nWe propose two new metrics, Logits Magnitude and Regularized Standard Deviation, to compare the differences and similarities between various methods. \nUsing these two newly proposed metrics, we demonstrate that when the Logits Magnitude across classes is nearly balanced, further reducing its overall value can effectively decrease errors and disturbances during training, leading to better model performance. \nBased on our analysis using these metrics, we observe that adjusting the logits could improve model performance, leading us to develop a simple label over-smoothing approach to adjust the logits without requiring prior knowledge of class distribution.\nThis method softens the original one-hot labels by assigning a probability slightly higher than $\\frac{1}{K}$ to the true class and slightly lower than $\\frac{1}{K}$ to the other classes, where $K$ is the number of classes.\nOur method achieves state-of-the-art performance on various imbalanced datasets, including CIFAR100-LT, ImageNet-LT, and iNaturalist2018.", "keywords": [ "Long-Tailed Recognition and Decoupled Training" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "OeHSkJ58TG", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Kushal Thaman", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Naomi Bashkansky", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Undergrad student" }, { "name": "Rylan Schaeffer", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" }, { "name": "Trevor Chow", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Victor Lecomte", "gender": "Male", "institution": "Alignment Research Center", "country": "US", "position": "Researcher" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" } ], "n_formula": 40, "n_formula_1": 17, "n_ref_uni": 16, "n_ref": 24, "n_ref_all": 40, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3984, "formula_len_all_1": 1436, "len_all": 96827, "len_all_1": 53135, "len_abs": 1352, "len_title": 123, "len_sents": 33291, "len_sents_1": 23596, "n_sents": 239, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1343, "title": "Incidental Polysemanticity: A New Obstacle for Mechanistic Interpretability", "abs": "Polysemantic neurons \u2014 neurons that activate for a set of unrelated features \u2014 have been seen as a significant obstacle towards interpretability of task-optimized deep networks, with implications for AI safety. The classic origin story of polysemanticity is that the data contains more \"features\" than neurons, such that learning to perform a task forces the network to co-allocate multiple unrelated features to the same neuron, endangering our ability to understand networks' internal processing. In this work, we present a second and non-mutually exclusive origin story of polysemanticity. We show that polysemanticity can arise incidentally, even when there are ample neurons to represent all features in the data, a phenomenon we term incidental polysemanticity. Using a combination of theory and experiments, we show that incidental polysemanticity can arise due to multiple reasons including regularization and neural noise; this incidental polysemanticity occurs because random initialization can, by chance alone, initially assign multiple features to the same neuron, and the training dynamics then strengthen such overlap. Our paper concludes by calling for further research quantifying the performance-polysemanticity tradeoff in task-optimized deep neural networks to better understand to what extent polysemanticity is avoidable.", "keywords": [ "polysemanticity", "mechanistic interpretability", "AI safety", "deep learning", "science of deep learning", "neural computation", "interpretability" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "OeBY9XqiTz", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Bernt Schiele", "gender": "Male", "institution": "Max Planck Institute for Informatics, Saarland Informatics Campus", "country": "DE", "position": "Full Professor" }, { "name": "Luc Van Gool", "gender": "unknown", "institution": "INSAIT - Sofia Un.", "country": "BG", "position": "Full Professor" }, { "name": "Luigi Piccinelli", "gender": "Male", "institution": "ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Siyuan Li", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Yung-Hsu Yang", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Mattia Seg\u00f9", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 45, "n_ref": 126, "n_ref_all": 144, "n_fig": 5, "n_tab": 11, "L_tab": 4348, "n_element_tab": 415, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1764, "n_element_tab_1": 163, "formula_len_all": 393, "formula_len_all_1": 413, "len_all": 168871, "len_all_1": 77377, "len_abs": 1621, "len_title": 74, "len_sents": 45241, "len_sents_1": 34451, "n_sents": 350, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 74, "L_abs": 1553, "title": "Samba: Synchronized Set-of-Sequences Modeling for Multiple Object Tracking", "abs": "Multiple object tracking in complex scenarios - such as coordinated dance performances, team sports, or dynamic animal groups - presents unique challenges. In these settings, objects frequently move in coordinated patterns, occlude each other, and exhibit long-term dependencies in their trajectories. However, it remains a key open research question on how to model long-range dependencies within tracklets, interdependencies among tracklets, and the associated temporal occlusions. To this end, we introduce Samba, a novel linear-time set-of-sequences model designed to jointly process multiple tracklets by synchronizing the multiple selective state-spaces used to model each tracklet. Samba autoregressively predicts the future track query for each sequence while maintaining synchronized long-term memory representations across tracklets. By integrating Samba into a tracking-by-propagation framework, we propose SambaMOTR, the first tracker effectively addressing the aforementioned issues, including long-range dependencies, tracklet interdependencies, and temporal occlusions. Additionally, we introduce an effective technique for dealing with uncertain observations (MaskObs) and an efficient training recipe to scale SambaMOTR to longer sequences. By modeling long-range dependencies and interactions among tracked objects, SambaMOTR implicitly learns to track objects accurately through occlusions without any hand-crafted heuristics. Our approach significantly surpasses prior state-of-the-art on the DanceTrack, BFT, and SportsMOT datasets.", "keywords": [ "Synchronized sequence modeling; multiple object tracking" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 4, 3, 2 ], "presentation_list": [ 4, 3, 2 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 3, 5 ] }, { "paper_id": "Odtr1rzQMq", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Alexey Naumov", "gender": "Male", "institution": "Higher School of Economics", "country": "", "position": "Associate Professor" }, { "name": "Denis Belomestny", "gender": "Male", "institution": "Duisburg-Essen University", "country": "", "position": "Full Professor" }, { "name": "Eric Moulines", "gender": "Male", "institution": "Ecole polytechnique", "country": "", "position": "Full Professor" }, { "name": "Marina Sheshukova", "gender": "Female", "institution": "Higher School of Economics", "country": "", "position": "Undergrad student" }, { "name": "Sergey Samsonov", "gender": "Male", "institution": "Higher School of Economics", "country": "", "position": "PhD student" }, { "name": "Alain Durmus", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "", "position": "Full Professor" } ], "n_formula": 206, "n_formula_1": 62, "n_ref_uni": 30, "n_ref": 72, "n_ref_all": 104, "n_fig": 1, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18356, "formula_len_all_1": 3757, "len_all": 165604, "len_all_1": 68390, "len_abs": 1171, "len_title": 144, "len_sents": 41853, "len_sents_1": 26486, "n_sents": 490, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1212, "title": "Nonasymptotic Analysis of Stochastic Gradient Descent with the Richardson\u2013Romberg Extrapolation", "abs": "We address the problem of solving strongly convex and smooth minimization problems using stochastic gradient descent (SGD) algorithm with a constant step size. Previous works suggested to combine the Polyak-Ruppert averaging procedure with the Richardson-Romberg extrapolation technique to reduce the asymptotic bias of SGD at the expense of a mild increase of the variance. We significantly extend previous results by providing an expansion of the mean-squared error of the resulting estimator with respect to the number of iterations $n$. More precisely, we show that the mean-squared error can be decomposed into the sum of two terms: a leading one of order $\\mathcal{O}(n^{-1/2})$ with explicit dependence on a minimax-optimal asymptotic covariance matrix, and a second-order term of order $\\mathcal{O}(n^{-3/4})$ where the power $3/4$ is best known. We also extend this result to the $p$-th moment bound keeping optimal scaling of the remainders with respect to $n$. Our analysis relies on the properties of the SGD iterates viewed as a time-homogeneous Markov chain. In particular, we establish that this chain is geometrically ergodic with respect to a suitably defined weighted Wasserstein semimetric.", "keywords": [ "Stochastic first-order optimization", "Richardson-Romberg extrapolation", "Polyak-Ruppert averaging" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "OdoS6cH8MP", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Jui-Chien Lin", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "MS student" }, { "name": "Mohammad Mohammadi Amiri", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 5, "n_ref_uni": 32, "n_ref": 36, "n_ref_all": 52, "n_fig": 6, "n_tab": 5, "L_tab": 1609, "n_element_tab": 90, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1024, "n_element_tab_1": 79, "formula_len_all": 70, "formula_len_all_1": 597, "len_all": 117351, "len_all_1": 74080, "len_abs": 2136, "len_title": 90, "len_sents": 44948, "len_sents_1": 36786, "n_sents": 324, "n_sents_1": 282, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 2145, "title": "Language Models for Textual Data Valuation", "abs": "In the rapidly evolving field of machine learning (ML), the quality of training data significantly impacts model performance, especially with the rise of foundation models capable of generating data. Measuring data quality may be linked to two statistical metrics: similarity and diversity, relative to a baseline dataset. We introduce DetEmbedMetrics, a novel deterministic embedding-based metric that enables textual data quality assessment by integrating a language model (LM) with deterministic similarity and diversity measurement functions. The core methodology constrains LM-generated embeddings to align with deterministic mathematical measurement functions, endowing the embeddings with desirable statistical properties. This approach enables the valuation of data quality by providing consistent and reliable similarity and diversity measurements, in contrast to methods directly employing neural networks for measuring data quality. Specifically, our approach involves fine-tuning an LM by inputting textual data samples with varying levels of similarity and diversity. The model learns to generate embeddings that, when applied to deterministic similarity and diversity functions, effectively capture the relationship between data sample pairs. This method allows the model to provide associated probabilities for different levels of similarity and diversity, offering clearer interpretation and decision-making compared to continuous scores. Extensive experiments on synthetic datasets demonstrate the effectiveness of DetEmbedMetrics in identifying similarity and diversity within various datasets. Notably, DetEmbedMetrics exhibits generalizability by performing robustly across different deterministic similarity and diversity functions, not relying on specific measurement techniques. This flexibility enhances its applicability as a robust framework for various measurement functions. By providing high-quality embeddings that facilitate the valuation of similarity and diversity between datasets, this research contributes to the growing field of data-centric ML, emphasizing the importance of data quality in the ML pipeline.", "keywords": [ "Language Model", "Data Valuation", "Similarity", "Diversity" ], "rating_list": [ 3, 1, 1, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 1, 1, 1, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "OdnqG1fYpo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chenhe Du", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "PhD student" }, { "name": "Hongjiang Wei", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Jingyi Yu", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "Full Professor" }, { "name": "Qing Wu", "gender": "Not Specified", "institution": "ShanghaiTech University", "country": "CN", "position": "PhD student" }, { "name": "Xuanyu Tian", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "PhD student" }, { "name": "Yuyao Zhang", "gender": "Female", "institution": "ShanghaiTech University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 39, "n_ref": 110, "n_ref_all": 131, "n_fig": 11, "n_tab": 6, "L_tab": 1831, "n_element_tab": 115, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1165, "n_element_tab_1": 56, "formula_len_all": 1244, "formula_len_all_1": 606, "len_all": 153064, "len_all_1": 60294, "len_abs": 1401, "len_title": 140, "len_sents": 41810, "len_sents_1": 26902, "n_sents": 439, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 91, "L_abs": 1467, "title": "Moner: Motion Correction in Undersampled Radial MRI with Unsupervised Neural Representation", "abs": "Motion correction (MoCo) in radial MRI is a particularly challenging problem due to the unpredictability of subject movement. Current state-of-the-art (SOTA) MoCo algorithms often rely on extensive high-quality MR images to pre-train neural networks, which constrains the solution space and leads to outstanding image reconstruction results. However, the need for large-scale datasets significantly increases costs and limits model generalization. In this work, we propose Moner, an unsupervised MoCo method that jointly reconstructs artifact-free MR images and estimates accurate motion from undersampled, rigid motion-corrupted k-space data, without requiring any training data. Our core idea is to leverage the continuous prior of implicit neural representation (INR) to constrain this ill-posed inverse problem, facilitating optimal solutions. Specifically, we integrate a quasi-static motion model into the INR, granting its ability to correct subject's motion. To stabilize model optimization, we reformulate radial MRI reconstruction as a back-projection problem using the Fourier-slice theorem. Additionally, we propose a novel coarse-to-fine hash encoding strategy, significantly enhancing MoCo accuracy. Experiments on multiple MRI datasets show our Moner achieves performance comparable to SOTA MoCo techniques on in-domain data, while demonstrating significant improvements on out-of-domain data. The code is available at: https://github.com/iwuqing/Moner", "keywords": [ "MRI Reconstruction", "Motion Correction", "Neural Representation", "NeRF", "Unsupervised Learning" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 3, 3, 5 ] }, { "paper_id": "OdMqKszKSd", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Ali Mahdavi-Amiri", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "Assistant Professor" }, { "name": "Angel X Chang", "gender": "Female", "institution": "Simon Fraser University", "country": "CA", "position": "Associate Professor" }, { "name": "Denys Iliash", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "MS student" }, { "name": "Jiayi Liu", "gender": "Female", "institution": "Simon Fraser University", "country": "CA", "position": "PhD student" }, { "name": "Manolis Savva", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 59, "n_ref": 100, "n_ref_all": 131, "n_fig": 13, "n_tab": 9, "L_tab": 2779, "n_element_tab": 380, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1039, "n_element_tab_1": 151, "formula_len_all": 239, "formula_len_all_1": 311, "len_all": 194013, "len_all_1": 71415, "len_abs": 1291, "len_title": 123, "len_sents": 55668, "len_sents_1": 34301, "n_sents": 377, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1299, "title": "SINGAPO: Single Image Controlled Generation of Articulated Parts in Objects", "abs": "We address the challenge of creating 3D assets for household articulated objects from a single image.\nPrior work on articulated object creation either requires multi-view multi-state input, or only allows coarse control over the generation process.\nThese limitations hinder the scalability and practicality for articulated object modeling.\nIn this work, we propose a method to generate articulated objects from a single image.\nObserving the object in a resting state from an arbitrary view, our method generates an articulated object that is visually consistent with the input image.\nTo capture the ambiguity in part shape and motion posed by a single view of the object, we design a diffusion model that learns the plausible variations of objects in terms of geometry and kinematics.\nTo tackle the complexity of generating structured data with attributes in multiple domains, we design a pipeline that produces articulated objects from high-level structure to geometric details in a coarse-to-fine manner, where we use a part connectivity graph and part abstraction as proxies.\nOur experiments show that our method outperforms the state-of-the-art in articulated object creation by a large margin in terms of the generated object realism, resemblance to the input image, and reconstruction quality.", "keywords": [ "3D articulated objects creation; generative model" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "OclSRDktp3", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Di Wang", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Liang Liu", "gender": "Male", "institution": "Soochow University", "country": "CN", "position": "MS student" }, { "name": "Lijie Hu", "gender": "Female", "institution": "KAUST", "country": "SA", "position": "PhD student" }, { "name": "Mengdi Li", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" }, { "name": "Muhammad Asif Ali", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Postdoc" }, { "name": "Shu Yang", "gender": "Female", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Xin CHEN", "gender": "Male", "institution": "University of Southampton", "country": "GB", "position": "MS student" }, { "name": "Zhen Tan", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 5, "n_ref_uni": 60, "n_ref": 97, "n_ref_all": 117, "n_fig": 7, "n_tab": 8, "L_tab": 2585, "n_element_tab": 397, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 612, "n_element_tab_1": 72, "formula_len_all": 421, "formula_len_all_1": 213, "len_all": 212379, "len_all_1": 70115, "len_abs": 1207, "len_title": 117, "len_sents": 58961, "len_sents_1": 34420, "n_sents": 625, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1213, "title": "Understanding Reasoning in Chain-of-Thought from the Hopfieldian View", "abs": "Large Language Models have demonstrated remarkable abilities across various tasks, with Chain-of-Thought (CoT) prompting emerging as a key technique to enhance reasoning capabilities. However, existing research primarily focuses on improving performance, lacking a comprehensive framework to explain and understand the fundamental factors behind CoT's success. To bridge this gap, we introduce a novel perspective grounded in the Hopfieldian view of cognition in cognitive neuroscience. We establish a connection between CoT reasoning and key cognitive elements such as stimuli, actions, neural populations, and representation spaces. From our view, we can understand the reasoning process as the movement between these representation spaces. Building on this insight, we develop a method for localizing reasoning errors in the response of CoTs. Moreover, we propose the Representation-of-Thought (RoT) framework, which leverages the robustness of low-dimensional representation spaces to enhance the robustness of the reasoning process in CoTs. Experimental results demonstrate that RoT improves the robustness and interpretability of CoT reasoning while offering fine-grained control over the reasoning process.", "keywords": [ "Interpretability", "Chain-of-thought" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "OclHGmt2ZM", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Guanjie Zheng", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "", "position": "Assistant Professor" }, { "name": "Linghe Kong", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Zhanyu Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Chaolv zeng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 35, "n_ref": 132, "n_ref_all": 164, "n_fig": 9, "n_tab": 9, "L_tab": 10505, "n_element_tab": 487, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2432, "n_element_tab_1": 146, "formula_len_all": 1009, "formula_len_all_1": 966, "len_all": 166856, "len_all_1": 61943, "len_abs": 1921, "len_title": 144, "len_sents": 48438, "len_sents_1": 25611, "n_sents": 362, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1940, "title": "CMamba: Channel Correlation Enhanced State Space Models for Multivariate Time Series Forecasting", "abs": "Recent advancements in multivariate time series forecasting have been propelled by Linear-based, Transformer-based, and Convolution-based models, with Transformer-based architectures gaining prominence for their efficacy in temporal and cross-channel mixing.\nMore recently, Mamba, a state space model, has emerged with robust sequence and feature mixing capabilities.\nHowever, the suitability of the vanilla Mamba design for time series forecasting remains an open question, particularly due to its inadequate handling of cross-channel dependencies.\nCapturing cross-channel dependencies is critical in enhancing the performance of multivariate time series prediction.\nRecent findings show that self-attention excels in capturing cross-channel dependencies, whereas other simpler mechanisms, such as MLP, may degrade model performance.\nThis is counterintuitive, as MLP, being a learnable architecture, should theoretically capture both correlations and irrelevances, potentially leading to neutral or improved performance.\nDiving into the self-attention mechanism, we attribute the observed degradation in MLP performance to its lack of data dependence and global receptive field, which result in MLP's lack of generalization ability.\nConsidering the powerful sequence modeling capabilities of Mamba and the high efficiency of MLP, the combination of the two is an effective strategy for solving multivariate time series prediction.\nBased on the above insights, we introduce a refined Mamba variant tailored for time series forecasting.\nOur proposed model, \\textbf{CMamba}, incorporates a modified Mamba (M-Mamba) module for temporal dependencies modeling, a global data-dependent MLP (GDD-MLP) to effectively capture cross-channel dependencies, and a Channel Mixup mechanism to mitigate overfitting.\nComprehensive experiments conducted on seven real-world datasets demonstrate the efficacy of our model in improving forecasting performance.", "keywords": [ "Time Series Forecasting", "Mamba", "Channel-Dependent" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "Ocg3XIymmp", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Omar Mahmood", "gender": "Male", "institution": "Genentech", "country": "", "position": "ML Scientist" }, { "name": "Pedro Oliveira Pinheiro", "gender": "Male", "institution": "Prescient Design, Genentech", "country": "", "position": "Researcher" }, { "name": "Richard Bonneau", "gender": "unknown", "institution": "genentech", "country": "", "position": "VP ML for Drug Discovery" }, { "name": "Saeed Saremi", "gender": "Male", "institution": "Genentech", "country": "US", "position": "Principal Research Scientist" }, { "name": "Vishnu Sresht", "gender": "Male", "institution": "Genentech", "country": "", "position": "Principal Researcher" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 43, "n_ref": 58, "n_ref_all": 70, "n_fig": 2, "n_tab": 4, "L_tab": 944, "n_element_tab": 125, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 580, "n_element_tab_1": 81, "formula_len_all": 384, "formula_len_all_1": 321, "len_all": 98219, "len_all_1": 69270, "len_abs": 1821, "len_title": 101, "len_sents": 31692, "len_sents_1": 28823, "n_sents": 213, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1827, "title": "Pharmacophore-based design by learning on voxel grids", "abs": "Ligand-based drug discovery (LBDD) relies on making use of known binders to a protein target to find structurally diverse molecules similarly likely to bind. This process typically involves a brute force search of the known binder (query) against a molecular library using some metric of molecular similarity. One popular approach overlays the pharmacophore-shape profile of the known binder to 3D conformations enumerated for each of the library molecules, computes overlaps, and picks a set of diverse library molecules with high overlaps. While this virtual screening workflow has had considerable success in hit diversification, scaffold hopping, and patent busting, it scales poorly with library sizes and restricts candidate generation to existing library compounds. Leveraging recent advances in voxel-based generative modelling, we propose a pharmacophore-based generative model and workflows that address the scaling and fecundity issues of conventional pharmacophore-based virtual screening. We introduce VoxCap, a voxel captioning method for generating SMILES strings from voxelised molecular representations. We propose two workflows as practical use cases as well as benchmarks for pharmacophore-based generation: de-novo design, in which we aim to generate new molecules with high pharmacophore-shape similarities to query molecules, and fast search, which aims to combine generative design with a cheap 2D substructure similarity search for efficient hit identification. Our results show that VoxCap significantly outperforms previous methods in generating diverse de-novo hits. When combined with our fast search workflow, VoxCap reduces computational time by orders of magnitude while returning hits for all query molecules, enabling the search of large libraries that are intractable to search by brute force.", "keywords": [ "generative models", "drug discovery", "ligand-based drug discovery", "pharmacophore", "voxels", "captioning" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "OcXsdBo6vK", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Fan Wang", "gender": "Female", "institution": "Shandong University", "country": "", "position": "MS student" }, { "name": "Guanglin Zhou", "gender": "Male", "institution": "The University of Queensland", "country": "AU", "position": "Postdoc" }, { "name": "Hao Sun", "gender": "Male", "institution": "Shandong University", "country": "", "position": "MS student" }, { "name": "Rundong He", "gender": "Male", "institution": "Shandong University", "country": "", "position": "PhD student" }, { "name": "Wan su", "gender": "Female", "institution": "Shandong University", "country": "CN", "position": "PhD student" }, { "name": "Xin Gao", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Full Professor" }, { "name": "Yicong Dong", "gender": "Male", "institution": "Shandong University", "country": "", "position": "PhD student" }, { "name": "Yilong Yin", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "Full Professor" }, { "name": "Zhongyi Han", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 47, "n_ref": 69, "n_ref_all": 97, "n_fig": 14, "n_tab": 3, "L_tab": 1353, "n_element_tab": 183, "n_fig_1": 12, "n_tab_1": 3, "L_tab_1": 1142, "n_element_tab_1": 147, "formula_len_all": 364, "formula_len_all_1": 342, "len_all": 138382, "len_all_1": 63711, "len_abs": 2122, "len_title": 131, "len_sents": 41949, "len_sents_1": 30341, "n_sents": 264, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 2129, "title": "Active In-Context Learning: Enhancing the Generalization of Large Multimodal Models", "abs": "The performance of Large Multimodal Models (LMMs) on downstream tasks improves substantially when examples of visual-text relationships are incorporated as context, with performance gains increasing as the number of examples and the context window size grow. However, collecting high-quality training sets for In-Context Learning (ICL) to retrieve multimodal examples is not trivial, particularly in specialized domains like healthcare, remote sensing, finance, and scientific research, due to the significant costs of manual labeling and strict privacy regulations. In this paper, we introduce Active In-Context Learning (AICL), a novel paradigm that eliminates the need for traditional training sets in multimodal ICL. AICL dynamically selects and annotates a small, highly informative set of samples in real-time during the query phase of LMMs. This active set evolves throughout querying, with the most relevant examples being continuously retrieved from it to optimize LMM performance on new data, without relying on pre-existing training sets. To construct an optimal active set, we propose Spectral-based Representative Sampling, which applies spectral clustering in the early query phase to select samples that are early, class-balanced, and representative, ensuring the active set captures key features of the data distribution and reduces data bias. To fully leverage the active set, we propose Similarity-enhanced TopK Prompt Construction, which retrieves the most relevant multimodal examples using a TopK similarity strategy and integrates the visual similarities between the multimodal examples and the query samples directly into the text prompts. By incorporating this similarity information, LMMs can better grasp the relationships, leading to more accurate and context-aware predictions. Experimental results on 10 specialized datasets and four LMMs show that our method significantly enhances LMMs\u2019 generalization performance. For example, in medical diagnosis tasks, our method, using only 10 annotated samples in the active set, outperforms existing ICL methods that rely on 2,000 annotated training samples.", "keywords": [ "Large Multimodal Models", "In-context Learning", "Active Learning" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "OcTUquFXfx", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 1, "author_info_list": [ { "name": "Soumya Ram", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Undergrad student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 12, "n_ref": 18, "n_ref_all": 33, "n_fig": 4, "n_tab": 3, "L_tab": 331, "n_element_tab": 27, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 331, "n_element_tab_1": 27, "formula_len_all": 553, "formula_len_all_1": 527, "len_all": 79461, "len_all_1": 52904, "len_abs": 809, "len_title": 112, "len_sents": 28995, "len_sents_1": 22883, "n_sents": 269, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 814, "title": "Discovering Global Minima of High-Dimensional Energy Landscapes", "abs": "Identifying global minima of high-dimensional non-convex functions is a fundamental problem in fields such as structural biology and materials modeling. Existing solutions (e.g. AlphaFold) often rely on generalizing from data. In contrast, we address the challenging domain where no existing data is available, and only the ground-truth energy function is provided. Utilizing the action functional, we formulate a novel loss function that transforms the input's rough loss landscape into a benign one for the neural network parameters. This allows minimizing the loss to align with finding the global minimum of the energy landscape. We validate our method on high-dimensional global optimization tasks, demonstrating its ability to approximate global minima for energy landscapes with thousand-dimensional inputs.", "keywords": [ "non-convex optimization", "structural biology", "statistical mechanics" ], "rating_list": [ 3, 3, 1, 1, 5 ], "soundness_list": [ 2, 1, 1, 1, 3 ], "presentation_list": [ 2, 2, 1, 1, 3 ], "contribution_list": [ 2, 1, 1, 1, 2 ], "confidence_list": [ 3, 4, 4, 3, 3 ] }, { "paper_id": "Ob0UafH2YI", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Aditi Sarker", "gender": "unknown", "institution": "Wayne State University", "country": "US", "position": "PhD student" }, { "name": "Chengyin Li", "gender": "Male", "institution": "Henry Ford Health", "country": "US", "position": "Researcher" }, { "name": "Dongxiao Zhu", "gender": "Male", "institution": "Wayne State University", "country": "US", "position": "Full Professor" }, { "name": "Joerg Kliewer", "gender": "unknown", "institution": "New Jersey Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Prashant Khanduri", "gender": "Male", "institution": "Wayne State University", "country": "US", "position": "Assistant Professor" }, { "name": "Yao Qiang", "gender": "Male", "institution": "Oakland University", "country": "US", "position": "Assistant Professor" }, { "name": "RAFI IBN SULTAN", "gender": "Male", "institution": "Wayne State University", "country": "US", "position": "PhD student" } ], "n_formula": 140, "n_formula_1": 28, "n_ref_uni": 35, "n_ref": 94, "n_ref_all": 128, "n_fig": 7, "n_tab": 6, "L_tab": 262, "n_element_tab": 15, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 66, "n_element_tab_1": 12, "formula_len_all": 24366, "formula_len_all_1": 4211, "len_all": 242228, "len_all_1": 73925, "len_abs": 1795, "len_title": 153, "len_sents": 72459, "len_sents_1": 31009, "n_sents": 644, "n_sents_1": 249, "venue": "ICLR2025", "decision": "withdraw", "L_title": 104, "L_abs": 1903, "title": "FEDERATED COMPOSITIONAL OPTIMIZATION: THE IMPACT OF TWO-SIDED LEARNING RATES ON COMMUNICATION EFFICIENCY", "abs": "Compositional optimization (CO) has recently gained popularity due to its applications in distributionally robust optimization (DRO), meta-learning, reinforcement learning, and many other machine learning applications. The large-scale and distributed nature of data necessitates efficient federated learning (FL) algorithms for CO, but the compositional structure of the objective poses significant challenges. Current methods either rely on large batch gradients (which are impractical) or suffer from suboptimal communication efficiency. To address these challenges, we propose efficient FedAvg-type algorithms for solving non-convex CO in the FL setting. We first establish that standard FedAvg fails in solving the federated CO problems due to data heterogeneity, which amplifies bias in local gradient estimates. Our analysis establishes that either {\\em additional communication} or {\\em two-sided learning rate-based} algorithms are required to control this bias. To this end, we develop two algorithms for solving the federated CO problem. First, we propose FedDRO that utilizes the compositional problem structure to design a communication strategy that allows FedAvg to control the bias in the estimation of the compositional gradient, achieving $\\mathcal{O}(\\epsilon^{-2})$ sample and $\\mathcal{O}(\\epsilon^{-3/2})$ communication complexity. Then we propose DS-FedDRO, a two-sided learning rate algorithm, that eliminates the need for additional communication and achieves the optimal $\\mathcal{O}(\\epsilon^{-2})$ sample and $\\mathcal{O}(\\epsilon^{-1})$ communication complexity, highlighting the importance of two-sided learning rate algorithms for solving federated CO problems. The proposed algorithms avoid the need for large batch gradients and achieve linear speedup with the number of clients. We corroborate our theoretical findings with empirical studies on large-scale DRO problems.", "keywords": [ "Compositional optimization", "Federated learning", "Federated averaging", "Distributed learning" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "Oazgf8A24z", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Adeel Pervez", "gender": "unknown", "institution": "Institute of Science and Technology Austria", "country": "AT", "position": "Postdoc" }, { "name": "Dingling Yao", "gender": "Female", "institution": "Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Francesco Locatello", "gender": "Male", "institution": "Institute of Science and Technology", "country": "AT", "position": "Assistant Professor" }, { "name": "Jiale Chen", "gender": "unknown", "institution": "Institute of Science and Technology Austria", "country": "AT", "position": "PhD student" }, { "name": "Dan Alistarh", "gender": "Male", "institution": "Institute of Science and Technology", "country": "AT", "position": "Full Professor" } ], "n_formula": 51, "n_formula_1": 24, "n_ref_uni": 27, "n_ref": 75, "n_ref_all": 94, "n_fig": 6, "n_tab": 4, "L_tab": 1296, "n_element_tab": 185, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 996, "n_element_tab_1": 127, "formula_len_all": 5149, "formula_len_all_1": 2049, "len_all": 134969, "len_all_1": 65224, "len_abs": 885, "len_title": 84, "len_sents": 40334, "len_sents_1": 28068, "n_sents": 358, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 36, "L_abs": 958, "title": "Scalable Mechanistic Neural Networks", "abs": "We propose Scalable Mechanistic Neural Network (S-MNN), an enhanced neural network framework designed for scientific machine learning applications involving long temporal sequences. By reformulating the original Mechanistic Neural Network (MNN) (Pervez et al., 2024), we reduce the computational time and space complexities from cubic and quadratic with respect to the sequence length, respectively, to linear. This significant improvement enables efficient modeling of long-term dynamics without sacrificing accuracy or interpretability. Extensive experiments demonstrate that S-MNN matches the original MNN in precision while substantially reducing computational resources. Consequently, S-MNN can drop-in replace the original MNN in applications, providing a practical and efficient tool for integrating mechanistic bottlenecks into neural network models of complex dynamical systems. Source code is available at https://github.com/IST-DASLab/ScalableMNN.", "keywords": [ "Scientific Machine Learning", "Ordinary Differential Equations", "Time Series", "Dynamical Systems" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 4, 4, 4, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 4, 3, 2, 2 ] }, { "paper_id": "OaoDVZntGe", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Demetri Terzopoulos", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "Minglu Zhao", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Qian Long", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Ruoyan Li", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Undergrad student" }, { "name": "Tao Gao", "gender": "unknown", "institution": "General Electric", "country": "", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 30, "n_ref": 38, "n_ref_all": 59, "n_fig": 4, "n_tab": 9, "L_tab": 2759, "n_element_tab": 249, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 763, "formula_len_all_1": 696, "len_all": 129849, "len_all_1": 61533, "len_abs": 1418, "len_title": 96, "len_sents": 44944, "len_sents_1": 29970, "n_sents": 356, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 45, "L_abs": 1436, "title": "Inverse Attention Agent in Multi-Agent System", "abs": "A major challenge for Multi-Agent Systems (MAS) is enabling agents to adapt dynamically to diverse environments in which opponents and teammates may continually change. Agents trained using conventional methods tend to excel only within the confines of their training cohorts; their performance drops significantly when confronting unfamiliar agents. To address this shortcoming, we introduce Inverse Attention Agents that adopt concepts from the Theory of Mind (ToM) implemented algorithmically using an attention mechanism trained in an end-to-end manner. Crucial to determining the final actions of these agents, the weights in their attention model explicitly represent attention to different goals. We furthermore propose an inverse attention network that deduces the ToM of agents based on observations and prior actions. The network infers the attentional states of other agents, thereby refining the attention weights to adjust the agent's final action. We conduct experiments in a continuous environment, tackling demanding tasks encompassing cooperation, competition, and a blend of both. They demonstrate that the inverse attention network successfully infers the attention of other agents, and that this information improves agent performance. Additional human experiments show that, compared to baseline agent models, our inverse attention agents exhibit superior cooperation with humans and better emulate human behaviors.", "keywords": [ "multi-agent system", "Theory of mind agent" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "OaORjvWelu", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Hongning Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "John Lui", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Qipei Chen", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Wei Chen", "gender": "Male", "institution": "Microsoft Research", "country": "", "position": "Pricipal Researcher" }, { "name": "Xuchuang Wang", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Postdoc" }, { "name": "Zhendong Chu", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Qingyun Wu", "gender": "Female", "institution": "Pennsylvania State University", "country": "", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 18, "n_ref_uni": 20, "n_ref": 31, "n_ref_all": 38, "n_fig": 1, "n_tab": 5, "L_tab": 1403, "n_element_tab": 159, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1403, "n_element_tab_1": 159, "formula_len_all": 1152, "formula_len_all_1": 1152, "len_all": 88443, "len_all_1": 80364, "len_abs": 935, "len_title": 113, "len_sents": 27145, "len_sents_1": 24980, "n_sents": 220, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 940, "title": "Cost-Efficient Multi-Fidelity Alignment for LLMs", "abs": "Alignment is a critical step in large language model (LLM) post-training. It typically requires human annotations to align the model's output to human preferences, which is prohibitively expensive. This paper proposes a novel approach to reduce the alignment cost.\n Specifically, we consider multiple levels of alignment with different qualities and response-generating costs, which we refer to as multi-fidelity alignment. We develop a new approach to incorporating the varying levels of response quality to train a language model, aiming to reduce the cost of response collection for alignment while maintaining the performance of the language model. We provide theoretical insights and empirical results to support the effectiveness of the proposed multi-fidelity alignment approach. Lastly, we conduct experiments to corroborate the effectiveness of the proposed approach by comparing its performance with the vanilla alignment methods.", "keywords": [ "Multi-Fidelity", "Alignment", "LLM" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 2, 1 ], "presentation_list": [ 2, 2, 1 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "OZdr2mV5EI", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Feng Jiang", "gender": "Male", "institution": "University of Texas at Arlington", "country": "US", "position": "PhD student" }, { "name": "Hehuan Ma", "gender": "unknown", "institution": "University of Texas at Arlington", "country": "US", "position": "Postdoc" }, { "name": "Junzhou Huang", "gender": "Male", "institution": "University of Texas, Arlington", "country": "US", "position": "Full Professor" }, { "name": "Weizhi An", "gender": "Female", "institution": "University of Texas at Arlington, University of Texas at Arlington", "country": "US", "position": "PhD student" }, { "name": "Wenliang Zhong", "gender": "Male", "institution": "University of Texas at Arlington", "country": "US", "position": "PhD student" }, { "name": "Yuzhi Guo", "gender": "unknown", "institution": "University of Texas at Arlington", "country": "US", "position": "Postdoc" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 42, "n_ref": 89, "n_ref_all": 107, "n_fig": 6, "n_tab": 10, "L_tab": 2247, "n_element_tab": 348, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 1309, "n_element_tab_1": 121, "formula_len_all": 396, "formula_len_all_1": 284, "len_all": 151844, "len_all_1": 65119, "len_abs": 1639, "len_title": 117, "len_sents": 40140, "len_sents_1": 30127, "n_sents": 300, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1639, "title": "Instruction Contrastive Tuning for Zero-shot Composed Image Retrieval", "abs": "Composed Image Retrieval (CIR) requires retrieving a target image based on a composed query consisting of an image and accompanying text that modifies or instructs changes to the visual reference. This task is particularly challenging as it demands the model effectively follow modification instructions for accurate retrieval. Additionally, data acquisition difficulties hinder training models for specific tasks. To address these challenges, recent approaches explore Zero-Shot CIR (ZS-CIR), mainly leveraging CLIP-based models with tailored projections to compose images and textual modifications. However, these base models are not trained on instruction-aware data, limiting their ability to effectively combine visual and textual cues. In this paper, we propose a novel embedding method utilizing an instruction-tuned Multimodal Large Language Model (MLLM) to generate unified embeddings that seamlessly integrate images and modification instructions. Instruction-tuned MLLMs inherently align vision and text while exhibiting strong instruction-following capabilities, though they are primarily used in text generation. We introduce a two-stage training strategy to efficiently transform the MLLM\u2019s text generation capabilities into embedding extraction, and further refining its ability to follow modification instructions in CIR. Our model demonstrates significant advancements in ZS-CIR, outperforming state-of-the-art baselines across four public datasets: FashionIQ, CIRR, GeneCIS, and CIRCO. Our model highlights the potential of instruction-tuned MLLMs in capturing nuanced instruction comprehension and advancing CIR systems.", "keywords": [ "Multimodal Learning", "Composed Image Retrieval", "Large Vision Language Models" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "OZbFRNhpwr", "primary_area": "datasets and benchmarks", "n_author": 17, "author_info_list": [ { "name": "Bin Xie", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Derek Yuen", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Researcher" }, { "name": "Gongwei Chen", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Postdoc" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jingxuan Chen", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Researcher" }, { "name": "Jun Wang", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Professor" }, { "name": "Kaiwen Zhou", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Kun Shao", "gender": "unknown", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Li Yixing", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Liqiang Nie", "gender": "Male", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "dean" }, { "name": "Rui Shao", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Shuai Wang", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Weiwen Liu", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xurui Zhou", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "MS student" }, { "name": "Yasheng Wang", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "", "position": "Researcher" }, { "name": "Yuhao Yang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Zhihao Wu", "gender": "Male", "institution": "Huawei Technologies Research & Development (UK) Ltd.", "country": "GB", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 16, "n_ref": 70, "n_ref_all": 106, "n_fig": 11, "n_tab": 14, "L_tab": 12628, "n_element_tab": 1102, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 3279, "n_element_tab_1": 331, "formula_len_all": 7304, "formula_len_all_1": 0, "len_all": 215247, "len_all_1": 72315, "len_abs": 1416, "len_title": 116, "len_sents": 69285, "len_sents_1": 32216, "n_sents": 531, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 68, "L_abs": 1420, "title": "SPA-BENCH: A COMPREHENSIVE BENCHMARK FOR SMARTPHONE AGENT EVALUATION", "abs": "Smartphone agents are increasingly important for helping users control devices efficiently, with (Multimodal) Large Language Model (MLLM)-based approaches emerging as key contenders. Fairly comparing these agents is essential but challenging, requiring a varied task scope, the integration of agents with different implementations, and a generalisable evaluation pipeline to assess their strengths and weaknesses. In this paper, we present SPA-Bench, a comprehensive SmartPhone Agent Benchmark designed to evaluate (M)LLM-based agents in an interactive environment that simulates real-world conditions. SPA-Bench offers three key contributions: (1) A diverse set of tasks covering system and third-party apps in both English and Chinese, focusing on features commonly used in daily routines; (2) A plug-and-play framework enabling real-time agent interaction with Android devices, integrating over ten agents with the flexibility to add more; (3) A novel evaluation pipeline that automatically assesses agent performance across multiple dimensions, encompassing seven metrics related to task completion and resource consumption. Our extensive experiments across tasks and agents reveal challenges like interpreting mobile user interfaces, action grounding, memory retention, and execution costs. We propose future research directions to ease these difficulties, moving closer to real-world smartphone agent applications.", "keywords": [ "AI Agent", "LLM", "MLLM", "Benchmark", "Smartphone Control" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "OZZYqfplS3", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Alex Ororbia", "gender": "unknown", "institution": "Rochester Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Ankur Mali", "gender": "Male", "institution": "University of South Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Tommaso Salvatori", "gender": "Male", "institution": "VERSES", "country": "", "position": "Researcher" } ], "n_formula": 117, "n_formula_1": 15, "n_ref_uni": 39, "n_ref": 52, "n_ref_all": 61, "n_fig": 3, "n_tab": 5, "L_tab": 1630, "n_element_tab": 85, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 46, "n_element_tab_1": 10, "formula_len_all": 5929, "formula_len_all_1": 841, "len_all": 222287, "len_all_1": 72308, "len_abs": 1365, "len_title": 130, "len_sents": 78338, "len_sents_1": 35398, "n_sents": 655, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1368, "title": "Tight Stability, Convergence, and Robustness Bounds for Predictive Coding Networks", "abs": "Energy-based learning algorithms, such as predictive coding (PC), have garnered significant attention in the machine learning community due to their theoretical properties, such as local operations and biologically plausible mechanisms for error correction. In this work, we rigorously analyze the stability, robustness, and convergence of PC through the lens of dynamical systems theory. We show that, first, PC is Lyapunov stable under mild assumptions on its loss and residual energy functions, which implies intrinsic robustness to small random perturbations due to its well-defined energy-minimizing dynamics. Second, we formally establish that the PC updates approximate quasi-Newton methods by incorporating higher-order curvature information, which makes them more stable and able to converge with fewer iterations compared to models trained via backpropagation (BP). Furthermore, using this dynamical framework, we provide new theoretical bounds on the similarity between PC and other algorithms, i.e., BP and target propagation (TP), by precisely characterizing the role of higher-order derivatives. These bounds, derived through detailed analysis of the Hessian structures, show that PC is significantly closer to quasi-Newton updates than TP, providing a deeper understanding of the stability and efficiency of PC compared to conventional learning methods.", "keywords": [ "Predictive Coding", "Dynamical Systems", "fixed-point stability", "Lyapunov Stability", "Bio-inspired learning algorithms", "Energy-based learning algorithms" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "OZVTqoli2N", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Angelo Porrello", "gender": "Male", "institution": "University of Modena and Reggio Emilia, AimageLab", "country": "IT", "position": "Postdoc" }, { "name": "Lorenzo Bonicelli", "gender": "Male", "institution": "University of Modena and Reggio Emilia", "country": "IT", "position": "PhD student" }, { "name": "Monica Millunzi", "gender": "Female", "institution": "University of Modena and Reggio Emilia", "country": "IT", "position": "PhD student" }, { "name": "Pietro Buzzega", "gender": "Male", "institution": "University of Modena and Reggio Emilia", "country": "IT", "position": "PhD student" }, { "name": "Rita Cucchiara", "gender": "Female", "institution": "Universit\u00e0 di modena e reggio emilia", "country": "IT", "position": "Full Professor" }, { "name": "Simone Calderara", "gender": "Male", "institution": "University of Modena and Reggio Emilia", "country": "IT", "position": "Full Professor" } ], "n_formula": 57, "n_formula_1": 24, "n_ref_uni": 68, "n_ref": 161, "n_ref_all": 169, "n_fig": 2, "n_tab": 7, "L_tab": 4333, "n_element_tab": 594, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 2727, "n_element_tab_1": 184, "formula_len_all": 8230, "formula_len_all_1": 1482, "len_all": 258784, "len_all_1": 84307, "len_abs": 1120, "len_title": 125, "len_sents": 65063, "len_sents_1": 34827, "n_sents": 591, "n_sents_1": 283, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 77, "L_abs": 1124, "title": "A Second-Order Perspective on Model Compositionality and Incremental Learning", "abs": "The fine-tuning of deep pre-trained models has revealed compositional properties, with multiple specialized modules that can be arbitrarily composed into a single, multi-task model. However, identifying the conditions that promote compositionality remains an open issue, with recent efforts concentrating mainly on linearized networks. We conduct a theoretical study that attempts to demystify compositionality in standard non-linear networks through the second-order Taylor approximation of the loss function. The proposed formulation highlights the importance of staying within the pre-training basin to achieve composable modules. Moreover, it provides the basis for two dual incremental training algorithms: the one from the perspective of multiple models trained individually, while the other aims to optimize the composed model as a whole. We probe their application in incremental classification tasks and highlight some valuable skills. In fact, the pool of incrementally learned modules not only supports the creation of an effective multi-task model but also enables unlearning and specialization in certain tasks.", "keywords": [ "Continual Learning", "Model Compositionality", "Ensemble Learning", "Task Arithmetic" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "OZG6MuD3pT", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Hanmeng Liu", "gender": "Male", "institution": "Westlake University", "country": "", "position": "PhD student" }, { "name": "Yue Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" }, { "name": "Zhiyang Teng", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 21, "n_ref": 25, "n_ref_all": 36, "n_fig": 3, "n_tab": 3, "L_tab": 212, "n_element_tab": 22, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 212, "n_element_tab_1": 22, "formula_len_all": 87, "formula_len_all_1": 87, "len_all": 103824, "len_all_1": 60334, "len_abs": 997, "len_title": 131, "len_sents": 30085, "len_sents_1": 30093, "n_sents": 205, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1001, "title": "Logic Agent: Enhancing Validity with Logic Rule Invocation", "abs": "Chain-of-Thought (CoT) prompting has become a key strategy for enhancing the inferential abilities of large language models (LLMs) in reasoning tasks. However, it often struggles with ensuring reasoning validity and maintaining informativeness. This paper presents the Logic Agent (LA), a novel framework designed to boost the validity of reasoning in LLMs through strategic logic function calls. Distinct from traditional methods, LA converts LLMs into dynamic agents that apply propositional logic rules, transforming natural language inputs into structured logical forms. The agent utilizes a robust suite of predefined functions to guide the reasoning process effectively. This approach can enhance the structured and coherent generation of reasoning outputs, improving their interpretability and logical consistency. Through detailed experiments, we showcase LA's ability to adapt across different LLM sizes, significantly enhancing the accuracy of complex reasoning tasks across various domains.", "keywords": [ "reasoning", "large language models", "logic", "logic reasoning" ], "rating_list": [ 3, 1, 5, 5 ], "soundness_list": [ 1, 1, 4, 3 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 4, 5, 3, 2 ] }, { "paper_id": "OZ3NXrF3gQ", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Jan Robine", "gender": "Male", "institution": "Technische Universit\u00e4t Dortmund", "country": "DE", "position": "PhD student" }, { "name": "Marc H\u00f6ftmann", "gender": "Male", "institution": "Technische Universit\u00e4t Dortmund", "country": "", "position": "PhD student" }, { "name": "Stefan Harmeling", "gender": "Unspecified", "institution": "Technische Universit\u00e4t Dortmund", "country": "", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 50, "n_ref": 70, "n_ref_all": 84, "n_fig": 5, "n_tab": 4, "L_tab": 2869, "n_element_tab": 166, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 24, "n_element_tab_1": 2, "formula_len_all": 767, "formula_len_all_1": 766, "len_all": 128083, "len_all_1": 64987, "len_abs": 1396, "len_title": 97, "len_sents": 32676, "len_sents_1": 30650, "n_sents": 264, "n_sents_1": 247, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1402, "title": "Reward-free Policy Optimization with World Models", "abs": "As AI capabilities advance, their rapid progress is not keeping pace with the need for safe and value-aligned algorithms, raising concerns about autonomous systems. E.g., maximizing expected return in reinforcement learning can lead to unintended and potentially harmful consequences. This work introduces Reward-free Policy Optimization (RFPO), a method that prioritizes goal-oriented policy learning over reward maximization by eliminating rewards as the agent's learning signal. Our approach learns a world model that simulates backward in time, and then uses it to construct a directed graph for planning, and finally learning a goal-conditioned policy from the graph. The algorithm has two requirements: (1) the goal has to be defined, and (2) the agent needs sufficient world knowledge, enabling it to plan. This method removes the risks associated with reward hacking and discourages unintended behaviors by allowing for human oversight. Additionally, it provides a framework for humans to build transparent and high-level algorithms by using the (low-level) learned policies. We demonstrate the effectiveness of RFPO on maze environments with pixel observations, where the agent successfully reaches arbitrarily selected goals and follows human-designed algorithms. In conclusion, RFPO enables agents to learn policies without rewards and provides a framework for creating high-level behaviors.", "keywords": [ "Reward-free", "Goal-conditioned", "World Models", "Planning", "AI Safety" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 2, 1, 1, 2 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "OYrqvAVXiQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Heng Liu", "gender": "Male", "institution": "vivo AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Lei Liu", "gender": "Male", "institution": "Vivo AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Rui Hu", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Tianheng Cheng", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Wenyu Liu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoxin Chen", "gender": "Male", "institution": "vivo AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Xinggang Wang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yuxuan Zhang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "ranlongjin", "gender": "Male", "institution": "vivo", "country": "CN", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 59, "n_ref": 182, "n_ref_all": 202, "n_fig": 7, "n_tab": 5, "L_tab": 1019, "n_element_tab": 118, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3646, "n_element_tab_1": 173, "formula_len_all": 163, "formula_len_all_1": 183, "len_all": 166388, "len_all_1": 67328, "len_abs": 1314, "len_title": 126, "len_sents": 37907, "len_sents_1": 30216, "n_sents": 278, "n_sents_1": 195, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1268, "title": "EVF-SAM: Early Vision-Language Fusion For Text-Prompted Segment Anything Model", "abs": "Segment Anything Model (SAM) has attracted widespread attention for its superior interactive segmentation capabilities with visual prompts while lacking further exploration of text prompts. In this paper, we empirically investigate what text prompt encoders (e.g., CLIP or LLM) are good for adapting SAM for referring expression segmentation and introduce the Early Vision-language Fusion-based SAM (EVF-SAM). EVF-SAM is a simple yet effective referring segmentation method which exploits multimodal prompts (i.e., image and text) and comprises a pre-trained vision-language model to generate referring prompts and a SAM for segmentation. Surprisingly, we observe that: (1) multimodal prompts and (2) vision-language models with early fusion (e.g., BEIT-3) are beneficial for prompting SAM for accurate referring segmentation. Our experiments show that the proposed EVF-SAM based on BEIT-3 can obtain state-of-the-art performance on RefCOCO/+/g for referring expression segmentation and demonstrate the superiority of prompting SAM with early vision-language fusion. In addition, the proposed EVF-SAM with 1.32B parameters achieves remarkably higher performance while reducing nearly 82% of parameters compared to previous SAM methods based on large multimodal models.", "keywords": [ "referring image segmentation", "vision-language models", "multimodal models", "segment anything" ], "rating_list": [ 5, 5, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "OYTDePFRLC", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Chenliang Xu", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Associate Professor" }, { "name": "Jinyang Jiang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Mingqian Feng", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Yijie Peng", "gender": "Male", "institution": "Peking University", "country": "", "position": "Associate Professor" }, { "name": "Zeliang Zhang", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" } ], "n_formula": 45, "n_formula_1": 16, "n_ref_uni": 29, "n_ref": 59, "n_ref_all": 71, "n_fig": 4, "n_tab": 2, "L_tab": 887, "n_element_tab": 93, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 785, "n_element_tab_1": 95, "formula_len_all": 4747, "formula_len_all_1": 1457, "len_all": 156261, "len_all_1": 76580, "len_abs": 1463, "len_title": 90, "len_sents": 48135, "len_sents_1": 32702, "n_sents": 406, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1469, "title": "Forward Learning with Differential Privacy", "abs": "Differential privacy (DP) in deep learning is a critical concern as it ensures the confidentiality of training data while maintaining model utility. Existing DP training algorithms provide privacy guarantees by clipping and then injecting external noise into sample gradients computed by the backpropagation algorithm. Different from backpropagation, forward-learning algorithms based on perturbation inherently add noise during the forward pass and utilize randomness to estimate the gradients. Although these algorithms are non-privatized, the introduction of noise during the forward pass indirectly provides internal randomness protection to the model parameters and their gradients, suggesting the potential for naturally providing differential privacy. In this paper, we propose a privatized forward-learning algorithm, Differential Private Unified Likelihood Ratio (DP-ULR), and demonstrate its differential privacy guarantees. DP-ULR features a novel batch sampling operation with rejection, of which we provide theoretical analysis in conjunction with classic differential privacy mechanisms. DP-ULR is also underpinned by a theoretically guided privacy controller that dynamically adjusts noise levels to manage privacy costs in each training step. Our experiments indicate that DP-ULR achieves competitive performance compared to traditional differential privacy training algorithms based on backpropagation, maintaining nearly the same privacy loss limits.", "keywords": [ "Differential Privacy", "Forward Learning", "Likelihood Ratio Method" ], "rating_list": [ 3, 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 2, 4, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3, 3 ] }, { "paper_id": "OYT7yZfBFw", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "David L. Buckeridge", "gender": "unknown", "institution": "McGill University", "country": "", "position": "Full Professor" }, { "name": "He Zhu", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Undergrad student" }, { "name": "Qincheng Lu", "gender": "unknown", "institution": "McGill University", "country": "", "position": "PhD student" }, { "name": "Yue Li", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Assistant Professor" }, { "name": "Ziyang Song", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 6, "n_ref_uni": 34, "n_ref": 78, "n_ref_all": 94, "n_fig": 6, "n_tab": 5, "L_tab": 3114, "n_element_tab": 191, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 579, "n_element_tab_1": 33, "formula_len_all": 1448, "formula_len_all_1": 502, "len_all": 154123, "len_all_1": 63761, "len_abs": 1228, "len_title": 134, "len_sents": 40046, "len_sents_1": 30552, "n_sents": 314, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1238, "title": "TrajGPT: Irregular Time-Series Representation Learning for Health Trajectory Analysis", "abs": "In many domains, such as healthcare, time-series data is often irregularly sampled with varying intervals between observations. This poses challenges for classical time-series models that require equally spaced data. To address this, we propose a novel time-series Transformer called **Trajectory Generative Pre-trained Transformer (TrajGPT)**. TrajGPT employs a novel Selective Recurrent Attention (SRA) mechanism, which utilizes a data-dependent decay to adaptively filter out irrelevant past information based on contexts. By interpreting TrajGPT as discretized ordinary differential equations (ODEs), it effectively captures the underlying continuous dynamics and enables time-specific inference for forecasting arbitrary target timesteps. Experimental results demonstrate that TrajGPT excels in trajectory forecasting, drug usage prediction, and phenotype classification without requiring task-specific fine-tuning. By evolving the learned continuous dynamics, TrajGPT can interpolate and extrapolate disease risk trajectories from partially-observed time series. The visualization of predicted health trajectories shows that TrajGPT forecasts unseen diseases based on the history of clinically relevant phenotypes (i.e., contexts).", "keywords": [ "GPT", "Representation learning", "Linear attention", "ODE", "Irregularly-sample time series", "zero-shot learning", "trajectory analysis" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "OXi1FmHGzz", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Pedro Valdeira", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Shiqiang Wang", "gender": "Male", "institution": "IBM, International Business Machines", "country": "US", "position": "Research Staff Member" }, { "name": "Yuejie Chi", "gender": "Female", "institution": "Meta AI", "country": "US", "position": "Researcher" } ], "n_formula": 68, "n_formula_1": 28, "n_ref_uni": 22, "n_ref": 47, "n_ref_all": 70, "n_fig": 5, "n_tab": 7, "L_tab": 3114, "n_element_tab": 103, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 3743, "n_element_tab_1": 95, "formula_len_all": 4803, "formula_len_all_1": 1969, "len_all": 145409, "len_all_1": 76142, "len_abs": 2288, "len_title": 127, "len_sents": 45249, "len_sents_1": 31234, "n_sents": 385, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1783, "title": "Vertical Federated Learning with Missing Features During Training and Inference", "abs": "Vertical federated learning trains models from feature-partitioned datasets across multiple clients, who collaborate without sharing their local data. Standard approaches assume that all feature partitions are available during both training and inference. Yet, in practice, this assumption rarely holds, as for many samples only a subset of the clients observe their partition. However, not utilizing incomplete samples during training harms generalization, and not supporting them during inference limits the utility of the model. Moreover, if any client leaves the federation after training, its partition becomes unavailable, rendering the learned model unusable. Missing feature blocks are therefore a key challenge limiting the applicability of vertical federated learning in real-world scenarios. To address this, we propose LASER-VFL, a vertical federated learning method for efficient training and inference of split neural network-based models that is capable of handling arbitrary sets of partitions. Our approach is simple yet effective, relying on the sharing of model parameters and on task-sampling to train a family of predictors. We show that LASER-VFL achieves a $\\mathcal{O}({1}/{\\sqrt{T}})$ convergence rate for nonconvex objectives and, under the Polyak-\u0141ojasiewicz inequality, it achieves linear convergence to a neighborhood of the optimum. Numerical experiments show improved performance of LASER-VFL over the baselines. Remarkably, this is the case even in the absence of missing features. For example, for CIFAR-100, we see an improvement in accuracy of $18.2$\\% when each of four feature blocks is observed with a probability of 0.5 and of $7.4$\\% when all features are observed. The code for this work is available at https://github.com/Valdeira/LASER-VFL.", "keywords": [ "Vertical Federated Learning", "missing features" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3, 2 ] }, { "paper_id": "OXfllUhjrJ", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Anj Simmons", "gender": "Female", "institution": "Deakin University", "country": "", "position": "Researcher" }, { "name": "Anupam Chaudhuri", "gender": "Male", "institution": "Deakin University", "country": "", "position": "Postdoc" } ], "n_formula": 20, "n_formula_1": 20, "n_ref_uni": 22, "n_ref": 26, "n_ref_all": 34, "n_fig": 5, "n_tab": 4, "L_tab": 851, "n_element_tab": 59, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 156, "n_element_tab_1": 22, "formula_len_all": 1283, "formula_len_all_1": 1138, "len_all": 103612, "len_all_1": 58392, "len_abs": 1213, "len_title": 89, "len_sents": 39707, "len_sents_1": 26895, "n_sents": 301, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1218, "title": "Tropical Geometry Features for Novelty Detection and interpretability", "abs": "Existing methods for critical tasks such as out-of-distribution (OOD) detection,\nuncertainty quantification, and adversarial robustness often focus on measuring\nthe output of the last or intermediate layers of a neural network such as logits and\nenergy score. However, these methods typically overlook the geometric properties\nof the learned representations in the latent space, failing to capture important\nsignals that relate to model reliability, fairness, and adversarial vulnerability.\n\n\nInnovations: We introduce an innovative method, termed Tropical Geometry Features (TGF), for detecting out-of-distribution data and enhancing overall model evaluation. This approach leverages the geometric properties of polytopes derived\nfrom a trained neural network\u2019s learned representations. By integrating these\ngeometric features with the data used during training, TGF establishes a unique\nsignature of in-distribution data points. Our framework extends beyond OOD\ndetection, providing insights into model uncertainty, adversarial robustness, interpretability, and fairness. Through TGF, we enhance interpretability technique to detect OOD, uncertainty, adverserial robustness in dynamic and unpredictable\nenvironments.", "keywords": [ "OOD", "Tropical geometry", "Deep neural network", "polytopes" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 1, 3, 2 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 2, 3, 1 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "OXKJtrfH5n", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Mengzhu Wang", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Nan Yin", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Yingxu Wang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Zhiguang Cao", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Houcheng Su", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 63, "n_formula_1": 37, "n_ref_uni": 48, "n_ref": 99, "n_ref_all": 115, "n_fig": 4, "n_tab": 4, "L_tab": 1745, "n_element_tab": 136, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1014, "n_element_tab_1": 100, "formula_len_all": 7349, "formula_len_all_1": 2792, "len_all": 147951, "len_all_1": 71339, "len_abs": 1517, "len_title": 85, "len_sents": 34030, "len_sents_1": 28662, "n_sents": 283, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 37, "L_abs": 1585, "title": "Continuous Spiking Graph ODE Networks", "abs": "Spiking Graph Networks (SGNs), as bio-inspired neural models that address energy consumption challenges for graph classification, have attracted considerable attention from researchers and the industry. \nHowever, SGNs are typically applied in static scenarios with real-valued inputs and cannot be directly utilized for dynamic prediction because of their limited capacity to handle dynamic real-valued features, denoted as architectural inapplicability. Moreover, they suffer from accuracy loss due to the inherently discrete nature of spike-based representations.\nInspired by recent graph ordinary differential equation (ODE) methods, we propose the framework named \\textbf{C}ontinuous \\textbf{S}piking \\textbf{G}raph \\textbf{O}DE Networks (\\method{}), which leverages the advantages of graph ODE to address the architectural inapplicability, and employs high-order structures to solve the problem of information loss.\nSpecifically, \\method{} replaces the high energy-consuming static SGNs with an efficient Graph ODE process by incorporating SGNs with graph ODE into a unified framework, thereby achieving energy efficiency.\nThen, we derive a high-order spike representation capable of preserving more information. By integrating this with a high-order graph ODE, we propose the second-order \\method{} to address the information loss challenge.\nFurthermore, we prove that the second-order \\method{} maintains stability during the dynamic graph learning.\nExtensive experiments validate the superiority of the proposed \\method{} in performance while maintaining low power consumption.", "keywords": [ "Spiking graph neural network; graph ODE" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "OXIIFZqiiN", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Chunyang Ye", "gender": "Male", "institution": "Chuanzhi Education Group", "country": "CN", "position": "Lecturer" }, { "name": "Hanyi Yu", "gender": "Female", "institution": "University of Southern California", "country": "", "position": "MS student" }, { "name": "Mingqiao Mo", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Intern" }, { "name": "Wu Tingting", "gender": "Female", "institution": "China Mobile Communications Company Limited Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Yifang Xu", "gender": "Male", "institution": "Nanjing university", "country": "CN", "position": "Researcher" }, { "name": "Yiyang Niu", "gender": "Male", "institution": "Rice University", "country": "US", "position": "MS student" }, { "name": "Zhenghan Chen", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 23, "n_ref_uni": 11, "n_ref": 13, "n_ref_all": 16, "n_fig": 2, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1281, "formula_len_all_1": 1216, "len_all": 84246, "len_all_1": 54208, "len_abs": 1331, "len_title": 123, "len_sents": 27591, "len_sents_1": 23591, "n_sents": 210, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1335, "title": "A Dual-Modal Framework Utilizing Visual Prompts for Enhanced Patch Analysis", "abs": "Patch representation learning has emerged as a crucial innovation in software development, leveraging machine learning techniques to advance software generation workflows. This approach has led to significant enhancements across various applications involving code alterations. However, existing methods often exhibit a tendency towards specialization, excelling predominantly in either predictive tasks such as security patch classification or in generative tasks like the automated creation of patch descriptions. This paper presents a groundbreaking approach to patch representation learning through the Image-Guided Code Patch Framework (IGCP), a novel architecture that bridges the gap between code analysis and image processing domains. We introduce a rigorous mathematical foundation for IGCP, leveraging measure theory, functional analysis, and information geometry to formalize the domain adaptation process in patch representation learning. The optimization dynamics of IGCP are rigorously analyzed through the lens of Stochastic Gradient Langevin Dynamics, providing convergence guarantees in both convex and non-convex loss landscapes. Empirical evaluations demonstrate that IGCP not only achieves state-of-the-art performance in patch description generation but also exhibits remarkable domain generalization capabilities.", "keywords": [ "Code Generation", "Domain Adaptation" ], "rating_list": [ 1, 1, 3, 1 ], "soundness_list": [ 1, 1, 2, 1 ], "presentation_list": [ 1, 1, 2, 1 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 1, 3, 2, 3 ] }, { "paper_id": "OX4Tk43uwv", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Mark Horowitz", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Mert Pilanci", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Sunil Sudhakaran", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 15, "n_ref": 23, "n_ref_all": 56, "n_fig": 17, "n_tab": 3, "L_tab": 1020, "n_element_tab": 102, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 1532, "n_element_tab_1": 167, "formula_len_all": 626, "formula_len_all_1": 339, "len_all": 96203, "len_all_1": 55457, "len_abs": 1185, "len_title": 162, "len_sents": 37963, "len_sents_1": 24294, "n_sents": 265, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 113, "L_abs": 1201, "title": "SERDES Link Training with Edge Inference: Neural-Network Driven Discrete Optimization to Maximize Link Efficiency", "abs": "Meeting the growing data demands of modern AI applications requires efficient, high-speed communication links. We propose an edge inference framework that dynamically optimizes non-uniform quantization levels in programmable ADC receivers. While integer linear programming (ILP) offers high-quality solutions, its significant computational cost (120 seconds per instance on high-performance CPUs) and hardware requirements make it unsuitable for on-chip use. On-chip solutions are essential for fast, periodic adjustments to track time-varying effects such as temperature drift and ensure reliable communication. To address this, we train a convolutional neural network (CNN) using ILP-generated labels, achieving a 24,000x speedup with inference on a RISC-V microcontroller. The CNN leverages a custom loss function tied to system-level metrics, reducing area metric errors from 29\\% to less than 2\\%. Unlike prior works embedding neural networks in the signal path, our framework adapts periodically to channel variations without disrupting communication. This enables improved error rates, energy efficiency, and a scalable pathway for on-chip edge intelligence in next-generation high-speed links.", "keywords": [ "CNN", "ILP", "Gumbel-Softmax", "discrete optimization", "edge inference", "affine translation", "hardware", "microcontroller" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "OWIk5E4lJs", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Pengliang Ji", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" }, { "name": "Siwei Ma", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Xuanyi Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yizhi Wang", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 13, "n_ref": 21, "n_ref_all": 55, "n_fig": 13, "n_tab": 2, "L_tab": 2573, "n_element_tab": 66, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 872, "n_element_tab_1": 10, "formula_len_all": 696, "formula_len_all_1": 717, "len_all": 110056, "len_all_1": 54942, "len_abs": 561, "len_title": 114, "len_sents": 30578, "len_sents_1": 25968, "n_sents": 228, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 962, "title": "Interactive-Action Image Generation via Synthetic Physical Priors", "abs": "While diffusion-based text-to-image generation has made notable advancements, generating accurate images containing interactive actions remains a challenge due to the lack of inherent physical and spatial priors. To address this problem, we propose a novel pipeline that synthesizes a dataset enriched with physical priors using a graphics engine, combined with a captioning technique. Building on the dataset, we introduce a distillation-structured fine-tuning method, where a teacher network assists in inverting the semantics of interactive actions, leveraging the synthesized priors effectively. This fine-tuning method disentangles the synthetic data features while mitigating random misalignment during the fine-tuning process. Extensive experiments demonstrate that our method not only achieves state-of-the-art results but also highlights the synthetic data's potential to be applied more broadly in enhancing the generation of interactive action images.", "keywords": [ "synthetic data", "interactive action", "generative model" ], "rating_list": [ 3, 6, 3, 8, 6 ], "soundness_list": [ 1, 3, 2, 3, 3 ], "presentation_list": [ 1, 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 3, 4 ] }, { "paper_id": "OW9elPTHcE", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Gerhard Wunder", "gender": "Male", "institution": "Freie Universit\u00e4t Berlin", "country": "DE", "position": "Full Professor" }, { "name": "Thibaud Ardoin", "gender": "unknown", "institution": "Freie Universit\u00e4t Berlin", "country": "DE", "position": "PhD student" }, { "name": "Yi Cai", "gender": "unknown", "institution": "Freie Universit\u00e4t Berlin", "country": "DE", "position": "PhD student" } ], "n_formula": 51, "n_formula_1": 19, "n_ref_uni": 32, "n_ref": 52, "n_ref_all": 71, "n_fig": 2, "n_tab": 4, "L_tab": 572, "n_element_tab": 68, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 161, "n_element_tab_1": 19, "formula_len_all": 4332, "formula_len_all_1": 1123, "len_all": 140305, "len_all_1": 65720, "len_abs": 1275, "len_title": 113, "len_sents": 51295, "len_sents_1": 33135, "n_sents": 354, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1282, "title": "A General Feature Attribution Framework under a Black-box Setting", "abs": "Feature attribution is widely accepted as a form of explanation for reasoning machine decisions, indicating the proportion of each feature's contribution to an inquired decision. \nWhile most efforts have focused on determining attributions through exact gradient measurements, recent work has adopted gradient estimation to derive explanatory information requiring only query-level access \u2013 a restricted yet more practical accessibility assumption known as the black-box setting. \nFollowing this direction, this paper extends the idea of utilizing estimated gradients to a broader framework and introduces GEFA (Gradient-Estimation-based explanation For All). Unlike the previous attempt that focused on explaining image classifiers, the proposed explainer derives feature attributions in a proxy space, making it generally applicable to arbitrary black-box models, regardless of input type. \nIn addition to its close relationship with Integrated Gradients, we find, surprisingly, that our approach \u2013 a path method built upon estimated gradients \u2013 outputs unbiased estimates of Shapley Values. By avoiding the potential information waste sourced from computing marginal contributions, it improves the quality of derived explanations, as demonstrated by our quantitative evaluations.", "keywords": [ "Explainability", "Feature Attribution", "Black-box Setting", "Query-level Access" ], "rating_list": [ 6, 3, 8, 5 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "OW9TFoLuE4", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Lixu Wang", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Payal Mohapatra", "gender": "Female", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Qi Zhu", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "Full Professor" } ], "n_formula": 53, "n_formula_1": 24, "n_ref_uni": 64, "n_ref": 132, "n_ref_all": 193, "n_fig": 11, "n_tab": 14, "L_tab": 6232, "n_element_tab": 543, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 3048, "n_element_tab_1": 173, "formula_len_all": 4118, "formula_len_all_1": 1505, "len_all": 252248, "len_all_1": 75783, "len_abs": 1439, "len_title": 135, "len_sents": 70258, "len_sents_1": 32984, "n_sents": 516, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1445, "title": "Phase-Driven Domain Generalizable Learning For Nonstationary Time Series Classification", "abs": "Monitoring and recognizing patterns in continuous sensing data is crucial for many practical applications. These real-world time-series data are often nonstationary, characterized by varying statistical and spectral properties over time. This poses a significant challenge in developing learning models that can effectively generalize across different distributions. In this work, based on our observation that nonstationary statistics for time-series classification tasks are intrinsically linked to the phase information, we propose a time-series domain generalization framework, PhASER. It consists of three novel elements: 1) Hilbert transform-based phase augmentation that diversifies non-stationarity while preserving discriminatory semantics, 2) separate magnitude-phase encoding by viewing time-varying magnitude and phase as independent modalities, and 3) phase-residual feature broadcasting by incorporating phase with a novel residual connection for inherent regularization to enhance distribution invariant learning. Extensive evaluation on 5 datasets from sleep-stage classification, human activity recognition, and gesture recognition against 12 state-of-the-art baseline methods demonstrate that PhASER consistently outperforms the best baselines by an average of 5% and up to 13% in some cases. Moreover, PhASER\u2019s principles can also be applied broadly to boost the generalizability of existing time-series classification models.", "keywords": [ "Time Series; Domain Generalization; Machine Learning for Time Series data;" ], "rating_list": [ 3, 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 1, 2 ], "presentation_list": [ 1, 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3, 2 ] }, { "paper_id": "OW5Gf4cse1", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Benjamin Hoover", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Csaba Both", "gender": "unknown", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Daniel Karl I. Weidele", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Dmitry Krotov", "gender": "unknown", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Hendrik Strobelt", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Visiting Scientist" }, { "name": "Mauro Martino", "gender": "unknown", "institution": "IBM TJ Watson Research Center", "country": "US", "position": "Principal Researcher" }, { "name": "Nima Dehmamy", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 23, "n_ref": 51, "n_ref_all": 77, "n_fig": 14, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 567, "n_element_tab_1": 12, "formula_len_all": 67, "formula_len_all_1": 0, "len_all": 157884, "len_all_1": 65886, "len_abs": 1480, "len_title": 123, "len_sents": 64026, "len_sents_1": 33007, "n_sents": 482, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1489, "title": "The Role of Task Complexity in Emergent Abilities of Small Language Models", "abs": "We investigate the relationship between task complexity and the minimum model\nsize required for learning specific tasks in small transformer models. We focus\non the ListOps dataset, consisting of nested math operations. We define the task\ncomplexity as the Kolmogorov complexity (KC) of the code solving the task, using\na rough proxy for KC. We find a power-law relation between KC and parameters\nrequired to learn, suggesting number of parameters to learn harder task increases\nalmost cubic in KC. On individual math operations, sum mod 10 is hardest to\nlearn. Surprisingly, when combining tasks, we observe that sum is learned earlier\nand with fewer parameters when trained alongside max and median. Analyzing\nthe model, we find strong evidence that models trained on sum alone and models\ntrained jointly converge to different algorithms. Concretely, the sum alone model\ndoesn\u2019t seem to have learned number properties in the embedding layer, likely\nmemorizing the sum table. In contrast, models trained on three tasks (maximum,\nmedian and sum) reveals that joint training results in clear number-like properties.\nFinally, we also find evidence that the sum-only model utilizes its feedforward\nlayer more than the jointly trained model. Conversely, the attention layer in the\njoint model is activated more than the sum model. Our findings suggest there is\nanother dimension to emergent abilities in language models, namely the algorithms\nbeing learned, potentially impacting scaling laws.", "keywords": [ "LLM", "ListOps", "Emergent Abilities", "Scaling Laws" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "OW332Wh9S5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Changhan Wang", "gender": "Male", "institution": "Meta AI", "country": "", "position": "Researcher" }, { "name": "Heng-Jui Chang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Hongyu Gong", "gender": "Female", "institution": "FAIR at Meta", "country": "US", "position": "Researcher" }, { "name": "James R. Glass", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Principal Researcher" }, { "name": "Yu-An Chung", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 61, "n_ref": 187, "n_ref_all": 258, "n_fig": 18, "n_tab": 21, "L_tab": 7849, "n_element_tab": 1126, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 3497, "n_element_tab_1": 396, "formula_len_all": 183, "formula_len_all_1": 0, "len_all": 283075, "len_all_1": 71161, "len_abs": 3488, "len_title": 120, "len_sents": 72532, "len_sents_1": 29155, "n_sents": 547, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 918, "title": "DC-Spin: A Speaker-invariant Speech Tokenizer For Spoken Language Models", "abs": "Spoken language models (SLMs) have gained increasing attention with advancements in text-based, decoder-only language models. SLMs process text and speech, enabling simultaneous speech understanding and generation. This paper presents Double-Codebook Speaker-invariant Clustering (DC-Spin), which aims to improve speech tokenization by bridging audio signals and SLM tokens. DC-Spin extracts speaker-invariant tokens rich in phonetic information and resilient to input variations, enhancing zero-shot SLM tasks and speech resynthesis. We propose a chunk-wise approach to enable streamable DC-Spin without retraining and degradation. Comparisons of tokenization methods (self-supervised and neural audio codecs), model scalability, and downstream task proxies show that tokens easily modeled by an n-gram LM or aligned with phonemes offer strong performance, providing insights for designing speech tokenizers for SLMs.", "keywords": [ "speech tokenizer", "self-supervised learning", "spoken language model", "speech language model", "speech resynthesis", "audio codec" ], "rating_list": [ 5, 1, 8, 5 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 4, 2, 4, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 2, 5 ] }, { "paper_id": "OW0uRFs51N", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Biao Gong", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Changsheng Xu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "DanDan Zheng", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Jingdong Chen", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Senior Staff Algorithm Engineer" }, { "name": "Ming Yang", "gender": "Male", "institution": "Ant Group", "country": "US", "position": "Researcher" }, { "name": "Weiming Dong", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yuxin Zhang", "gender": "Not Specified", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 32, "n_ref": 89, "n_ref_all": 124, "n_fig": 19, "n_tab": 8, "L_tab": 1557, "n_element_tab": 156, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 573, "n_element_tab_1": 66, "formula_len_all": 216, "formula_len_all_1": 173, "len_all": 186425, "len_all_1": 49929, "len_abs": 370, "len_title": 119, "len_sents": 55454, "len_sents_1": 22928, "n_sents": 450, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1399, "title": "LumiSculpt: A Consistency Lighting Control Network for Video Generation", "abs": "Lighting is essential for the naturalness of video generation, which significantly impacts the overall aesthetic quality of the generated video. However, due to the deep coupling between lighting and the temporal features of videos, it is challenging for modeling independent and coherent lighting attributes, resulting in a lack of approaches for controlling lighting in videos. Therefore, inspired by the established controllable T2I models, we propose LumiSculpt, achieving precise and consistency lighting control in video generation models for the first time. LumiSculpt equips the video generation with strong interactive capabilities, allowing for the input of custom lighting reference image sequences. Furthermore, the core learnable plug-and-play module of LumiSculpt enables us to achieve remarkable performance on controlling light intensity, position, trajectory in latent video diffusion models based on the advanced DiT backbone. Additionally, to effectively fine-tune LumiSculpt and address the issue of insufficient lighting data, we construct LumiHuman, a new lightweight and flexible dataset for portrait lighting of images and videos. Experiments demonstrate that LumiSculpt achieves precise and high-quality lighting control in video generation. The code, model, and dataset will be released to facilitate further research. Video results are shown in the supplementary material.", "keywords": [ "Light Control", "Video generation", "Text-to-video generation", "diffusion models" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "OVxmpus9NA", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Alexandros Kouris", "gender": "Male", "institution": "Samsung AI", "country": "GB", "position": "Researcher" }, { "name": "Fuwen Tan", "gender": "Male", "institution": "ByteDance", "country": "US", "position": "Researcher" }, { "name": "Hao Mark Chen", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Hongxiang Fan", "gender": "Male", "institution": "Samsung", "country": "", "position": "Researcher" }, { "name": "Royson Lee", "gender": "unknown", "institution": "Samsung AI Center, Cambridge", "country": "GB", "position": "Research Scientist" }, { "name": "Stylianos Venieris", "gender": "unknown", "institution": "Samsung AI", "country": "GB", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 16, "n_ref": 46, "n_ref_all": 76, "n_fig": 7, "n_tab": 9, "L_tab": 3306, "n_element_tab": 583, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3003, "n_element_tab_1": 259, "formula_len_all": 222, "formula_len_all_1": 170, "len_all": 134640, "len_all_1": 67392, "len_abs": 1846, "len_title": 112, "len_sents": 42416, "len_sents_1": 30548, "n_sents": 299, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1741, "title": "Progressive Mixed-Precision Decoding for Efficient LLM Inference", "abs": "In spite of the great potential of large language models (LLMs) across various tasks, their deployment on resource-constrained devices remains challenging due to their excessive computational and memory demands. Quantization has emerged as an effective solution by storing weights in reduced precision. However, utilizing low precisions (i.e.~2/3-bit) to substantially alleviate the memory-boundedness of LLM decoding, still suffers from prohibitive performance drop. In this work, \nwe argue that existing approaches fail to explore the diversity in computational patterns, redundancy, and sensitivity to approximations of the different phases of LLM inference, resorting to a uniform quantization policy throughout.\nInstead, we propose a novel phase-aware method that selectively allocates precision during different phases of LLM inference, achieving both strong context extraction during prefill and efficient memory bandwidth utilization during decoding. To further address the memory-boundedness of the decoding phase, we introduce Progressive Mixed-Precision Decoding (PMPD), a technique that enables the gradual lowering of precision deeper in the generated sequence, together with a spectrum of precision-switching schedulers that dynamically drive the precision-lowering decisions in either task-adaptive or prompt-adaptive manner. \nExtensive evaluation across diverse language tasks shows that when targeting Nvidia GPUs, PMPD achieves 1.4$-$12.2$\\times$ speedup in matrix-vector multiplications over fp16 models, while when targeting an LLM-optimized NPU, our approach delivers a throughput gain of 3.8$-$8.0$\\times$ over fp16 models and up to 1.54$\\times$ over uniform quantization approaches while preserving the output quality.", "keywords": [ "LLM Quantization", "Efficient LLM Inference" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "OVu9DsOjgH", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Hai Jin", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Hong Huang", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Hua Zhu", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Kehan Yin", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Ruize Shi", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Wei Zhou", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "", "position": "Undergrad student" } ], "n_formula": 72, "n_formula_1": 27, "n_ref_uni": 46, "n_ref": 78, "n_ref_all": 106, "n_fig": 13, "n_tab": 21, "L_tab": 14528, "n_element_tab": 1759, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2208, "n_element_tab_1": 268, "formula_len_all": 5867, "formula_len_all_1": 1864, "len_all": 234088, "len_all_1": 59653, "len_abs": 4602, "len_title": 144, "len_sents": 61797, "len_sents_1": 24047, "n_sents": 509, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1333, "title": "FEDNET: FREQUENCY ENHANCED DECOMPOSED NETWORK FOR OUT-OF-DISTRIBUTION TIME SERIES CLASSIFICATION", "abs": "Time series classification is a crucial task with widespread applications in various fields such as medicine and energy. Due to the non-stationary property of time series, its data distribution will change over time, which makes it challenging for models to generalize to the out-of-distribution (OOD) environment. How\u0002ever, limitations persist in the current research on OOD time series classification, particularly the absence of a unified consideration addressing both domain dis\u0002tribution shift and temporal distribution shift. To this end, we view the time series distribution shift from the frequency perspective and propose a novel method called Frequency Enhanced Decomposed Network (FEDNet) for OOD time series classification. FEDNet utilizes frequency domain information to guide the decomposition of time series and further eliminates domain shift and temporal shift, it then obtains domain-invariant features for adapting to OOD data. Finally,we provide theoretical insights of FEDNet to validate its superiority for OOD time series classification. Comprehensive results on synthetic and real-world datasets demonstrate that FEDNet achieves state-of-the-art performance in OOD time series classification tasks, surpassing previous methods by up to 7%.Our code is available at https://anonymous.4open.science/r/FEDNet-743E", "keywords": [ "out-of-distribution", "time series classification", "frequency" ], "rating_list": [ 5, 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 2, 4, 4, 3, 4 ] }, { "paper_id": "OV0rZx8jr1", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Dongyoung Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Jaehyung Kim", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Jinsung Yoon", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Jinwoo Shin", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 10, "n_ref_uni": 26, "n_ref": 92, "n_ref_all": 131, "n_fig": 13, "n_tab": 12, "L_tab": 2731, "n_element_tab": 325, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 196, "n_element_tab_1": 38, "formula_len_all": 953, "formula_len_all_1": 827, "len_all": 174876, "len_all_1": 67159, "len_abs": 1605, "len_title": 120, "len_sents": 59041, "len_sents_1": 32691, "n_sents": 476, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1612, "title": "Debiasing Online Preference Learning via Preference Feature Preservation", "abs": "While various preferred features determine human preferences, current preference learning frameworks for large language models (LLMs) simplify them with binary pairwise comparisons and scalar rewards. This simplification could make LLMs' responses biased to mostly preferred features such as longer responses which would be exacerbated in online learning scenarios as the biases can be accumulate continuously throughout the iterations.\nTo address these challenges, we propose a novel framework called PFP (Preference Feature Preservation). The key idea of PFP is maintaining the distribution of human preference features throughout the online preference learning process. Specifically, PFP first trains a feature classifier using the existing offline pairwise human preference data. \nThen, using this classifier and the distribution preserving optimization, PFP maps appropriate preference features for each input instruction during online learning. \nLastly, PFP trains LLM using the existing preference learning framework, by incorporating the preference feature of each data into system prompts and enabling LLM to explicitly handle various human preferences. Our experiments demonstrate that PFP successfully mitigates the bias in preference features that arise during online learning, and achieves superior performance compared to previous preference learning methods on general benchmarks including AlpacaEval 2.0 and MT-Bench. We also observe that PFP almost resolves a length bias issue, a long-standing problem of online preference learning, even though it was not specifically designed to tackle this.", "keywords": [ "large language model", "alignment", "bias", "preference" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "OUuhwVsk9Z", "primary_area": "applications to robotics, autonomy, planning", "n_author": 11, "author_info_list": [ { "name": "Jialu Li", "gender": "Female", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Kunchang Li", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Mohit Bansal", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" }, { "name": "Shoubin Yu", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Songze Li", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Yali Wang", "gender": "Male", "institution": "SIAT, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yi Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yicong Hong", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Zun Wang", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Limin Wang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 59, "n_ref": 131, "n_ref_all": 168, "n_fig": 6, "n_tab": 16, "L_tab": 6852, "n_element_tab": 998, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1534, "n_element_tab_1": 143, "formula_len_all": 73, "formula_len_all_1": 60, "len_all": 248916, "len_all_1": 66287, "len_abs": 1755, "len_title": 130, "len_sents": 56634, "len_sents_1": 30651, "n_sents": 402, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1764, "title": "Bootstrapping Language-Guided Navigation Learning with Self-Refining Data Flywheel", "abs": "Creating high-quality data for training robust language-instructed agents is a long-lasting challenge in embodied AI. In this paper, we introduce a Self-Refining Data Flywheel (SRDF) that generates high-quality and large-scale navigational instruction-trajectory pairs by iteratively refining the data pool through the collaboration between two models, the instruction generator and the navigator, without any human-in-the-loop annotation. Specifically, SRDF starts with using a base generator to create an initial data pool for training a base navigator, followed by applying the trained strong navigator to filter the data pool. This leads to higher-fidelity data to train a better generator, which can, in turn, produce higher-quality data for training the next-round navigator. Such a flywheel establishes a data self-refining process, yielding a continuously improved and highly effective dataset for large-scale language-guided navigation learning. Our experiments demonstrate that after several flywheel rounds, the navigator elevates the performance boundary from 70% to 78% SPL on the classic R2R test set, surpassing human performance (76%) for the first time. Meanwhile, this process results in a superior instruction generator, as reflected by the improved SPICE from 23.5 to 25.7, better than all published approaches tailored for VLN instruction generation. Finally, we demonstrate the scalability of our method through increasing environment and instruction diversity, and the generalization ability of our pre-trained navigator across various downstream navigation tasks, surpassing state-of-the-art performance by a large margin in all cases. Code is uploaded as supplementary materials and all our data/code/models will also be publicly released.", "keywords": [ "vision-and-language navigation", "data flywheel", "dataset curation" ], "rating_list": [ 8, 8, 5, 5 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "OUo50cxU21", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Alejandro Sanchez Guinea", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "", "position": "Postdoc" }, { "name": "Max M\u00fchlh\u00e4user", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "", "position": "Full Professor" }, { "name": "Thomas Kreutz", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 50, "n_ref": 119, "n_ref_all": 158, "n_fig": 9, "n_tab": 1, "L_tab": 13, "n_element_tab": 1, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 225, "n_element_tab_1": 31, "formula_len_all": 1094, "formula_len_all_1": 701, "len_all": 159570, "len_all_1": 68004, "len_abs": 1778, "len_title": 116, "len_sents": 51969, "len_sents_1": 34466, "n_sents": 296, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1785, "title": "The Sky Is The Limit When Clustering Is Equated With Disentanglement", "abs": "Disentangled representation learning allows data to be mapped to a latent space where factors of variation can be individually manipulated. These factors define a direct notion of similarity between observations that naturally groups them into clusters with shared factors of variation. While this has been empirically shown to be effective on simple datasets, it is unclear how or when complex real-world data can be disentangled into representations that allow the same degree of manipulation and clustering. To advance the field of disentangled representation learning and clustering, we provide a new theoretical perspective by equating disentanglement with clustering by using factors of variation as a measure of element-wise similarity. This leads to a simple yet important observation: Instead of explicitly clustering the elements of a dataset, we can implicitly cluster them by learning to represent and generate the elements of each cluster. Furthermore, this observation reveals that implicit clusters have a lower bound because (I) explicit clusters are a subset of implicit clusters, and (II) implicit clusters can generate novel elements not present in the finite dataset through combinatorial generalization. Building on these insights, we derive an implicit neural clustering approach based on identifying factors of variation in the latent space. We validate our findings through experiments on synthetic image data and empirical evidence from related state-of-the-art works. This demonstrates the practical relevance of our approach and promising potential for synthesizing complete datasets from limited data, addressing data distribution gaps, improving interpretability in cluster analysis, enhancing SSL and classification tasks, and reducing data storage space.", "keywords": [ "Disentangled Representation Learning", "Clustering", "Factors of Variation", "Variational Autoencoders (VAEs)", "Generative Models", "Synthetic Image Data" ], "rating_list": [ 3, 3, 5, 5, 3 ], "soundness_list": [ 1, 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 3, 1 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3, 3 ] }, { "paper_id": "OUhR7Ghg3K", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Kajetan Schweighofer", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "", "position": "PhD student" }, { "name": "Nuria M Oliver", "gender": "Female", "institution": "ELLIS unit Alicante Foundation ", "country": "ES", "position": "Principal Researcher" }, { "name": "Sepp Hochreiter", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "Full Professor" }, { "name": "Adri\u00e1n Arnaiz-Rodr\u00edguez", "gender": "Male", "institution": "Universidad de Alicante", "country": "ES", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 10, "n_ref_uni": 64, "n_ref": 111, "n_ref_all": 186, "n_fig": 26, "n_tab": 21, "L_tab": 6142, "n_element_tab": 594, "n_fig_1": 14, "n_tab_1": 3, "L_tab_1": 2118, "n_element_tab_1": 90, "formula_len_all": 1878, "formula_len_all_1": 601, "len_all": 257583, "len_all_1": 67575, "len_abs": 1264, "len_title": 88, "len_sents": 75231, "len_sents_1": 30511, "n_sents": 637, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1270, "title": "The Disparate Benefits of Deep Ensembles", "abs": "Ensembles of Deep Neural Networks, Deep Ensembles, are widely used as a simple way to boost predictive performance. However, their impact on algorithmic fairness is not well understood yet. Algorithmic fairness investigates how a model\u2019s performance varies across different groups, typically defined by protected attributes such as age, gender, or race. In this work, we investigate the interplay between the performance gains from Deep Ensembles and fairness. Our analysis reveals that they unevenly favor different groups in what we refer to as a disparate benefits effect. We empirically investigate this effect with Deep Ensembles applied to popular facial analysis and medical imaging datasets, where protected group attributes are given and find that it occurs for multiple established group fairness metrics, including statistical parity and equal opportunity. Furthermore, we identify the per-group difference in predictive diversity of ensemble members as the potential cause of the disparate benefits effect. Finally, we evaluate different approaches to reduce unfairness due to the disparate benefits effect. Our findings show that post-processing is an effective method to mitigate this unfairness while preserving the improved performance of Deep Ensembles.", "keywords": [ "Deep Ensembles", "Algorithmic Fairness", "Disparate Benefits", "Post-Processing" ], "rating_list": [ 8, 6, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "OTFKVkxSlL", "primary_area": "reinforcement learning", "n_author": 9, "author_info_list": [ { "name": "Chao Li", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Cheng Wang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Chenglu Wen", "gender": "unknown", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Chenxing Lin", "gender": "Male", "institution": "Xiamen University", "country": "", "position": "Undergrad student" }, { "name": "Weiquan Liu", "gender": "Male", "institution": "Jimei University", "country": "CN", "position": "Associate Professor" }, { "name": "Wenqi Chen", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Yongquan Fu", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Ziwei Deng", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Siqi SHEN", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 77, "n_formula_1": 25, "n_ref_uni": 41, "n_ref": 119, "n_ref_all": 175, "n_fig": 11, "n_tab": 23, "L_tab": 5109, "n_element_tab": 638, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1026, "n_element_tab_1": 103, "formula_len_all": 8283, "formula_len_all_1": 2524, "len_all": 288342, "len_all_1": 68838, "len_abs": 1351, "len_title": 128, "len_sents": 89543, "len_sents_1": 26464, "n_sents": 825, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1360, "title": "DoF: A Diffusion Factorization Framework for Offline Multi-Agent Decision Making", "abs": "Diffusion models have been widely adopted in image and language generation and are now being applied to decision-making. However, the application of diffusion models in offline cooperative Multi-Agent decision making (MADM) remains limited. Although some researches exist, they suffer from scalability or poor cooperation issues due to the lack of design principles for diffusion-based MADM. The Individual-Global-Max (IGM) principle is a popular design principle for cooperative MADM. Through satisfying such principles, MADM algorithms achieve remarkable performance with good scalability. In this work, we extend the IGM principle as the Individual-Global-identically-Distributed (IGD) principle. This principle stipulates that the generated outcome of a multi-agent diffusion model should be identically distributed as the collective outcomes from multiple individual-agent diffusion models. We propose DoF, a diffusion factorization framework for MADM. It uses noise factorization function to factorize a centralized diffusion model into multiple diffusion models. We theoretically show that the noise factorization functions satisfy the IGD principle. Further, DoF uses data factorization function to model the complex relationship among data generated by multiple diffusion models. Through extensive experiments, we demonstrate the effectiveness of DoF.", "keywords": [ "multi-agent reinforcement learning; Diffusion Models; Offline reinforcement learning" ], "rating_list": [ 8, 5, 3 ], "soundness_list": [ 3, 3, 1 ], "presentation_list": [ 2, 2, 1 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 2, 3, 2 ] }, { "paper_id": "OT2NFdNrny", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Shigui Li", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Wei Chen", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Delu zeng", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 111, "n_formula_1": 52, "n_ref_uni": 28, "n_ref": 85, "n_ref_all": 106, "n_fig": 4, "n_tab": 12, "L_tab": 1638, "n_element_tab": 184, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 1342, "n_element_tab_1": 91, "formula_len_all": 10555, "formula_len_all_1": 4245, "len_all": 206198, "len_all_1": 74708, "len_abs": 953, "len_title": 122, "len_sents": 62828, "len_sents_1": 29588, "n_sents": 512, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 996, "title": "Improving Denoising Diffusion with Efficient Conditional Entropy Reduction", "abs": "Diffusion models (DMs) have achieved significant success in generative modeling, but their iterative denoising process is computationally expensive. Training-free samplers, such as DPM-Solver, accelerate this process through gradient estimation-based numerical iterations. However, the mechanisms behind this acceleration remain insufficiently understood. In this paper, we demonstrate gradient estimation-based iterations enhance the denoising process by effectively \\emph{\\textbf{r}educing the conditional \\textbf{e}ntropy} of reverse transition distribution. Building on this analysis, we introduce streamlined denoising iterations for DMs that optimize conditional entropy in score-integral estimation to improve the denoising iterations. Experiments on benchmark pre-trained models validate our theoretical insights, demonstrating that numerical iterations based on conditional entropy reduction improve the reverse denoising diffusion process of DMs. The code will be available.", "keywords": [ "Denoising Diffusion", "Conditional Entropy" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 1, 4, 3, 4 ] }, { "paper_id": "OSmjkkF6Uy", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Alan Malek", "gender": "Male", "institution": "DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Alexis Bellot", "gender": "Male", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Eleni Sgouritsa", "gender": "unknown", "institution": "Google Deepmind", "country": "GB", "position": "Researcher" }, { "name": "Francisco J. R. Ruiz", "gender": "Male", "institution": "DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Ira Ktena", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Jessica Schrouff", "gender": "Female", "institution": "Google DeepMind", "country": "GB", "position": "Senior Researcher" }, { "name": "Silvia Chiappa", "gender": "Female", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Virginia Aglietti", "gender": "Female", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 46, "n_ref": 88, "n_ref_all": 148, "n_fig": 21, "n_tab": 3, "L_tab": 257, "n_element_tab": 7, "n_fig_1": 16, "n_tab_1": 2, "L_tab_1": 70, "n_element_tab_1": 2, "formula_len_all": 219, "formula_len_all_1": 152, "len_all": 159524, "len_all_1": 72692, "len_abs": 1173, "len_title": 129, "len_sents": 54225, "len_sents_1": 36218, "n_sents": 368, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1176, "title": "FunBO: Discovering Acquisition Functions forBayesian Optimization with FunSearch", "abs": "The sample efficiency of Bayesian optimization algorithms depends on carefully crafted acquisition functions (AFs) guiding the sequential collection of function evaluations. The best-performing AF can vary significantly across optimization problems, often requiring ad-hoc and problem-specific choices. This work tackles the challenge of designing novel AFs that perform well across a variety of experimental settings. Based on FunSearch, a recent work using Large Language Models (LLMs) for discovery in mathematical sciences, we propose FunBO, an LLM-based method that can be used to learn new AFs written in computer code by leveraging access to a limited number of evaluations for a set of objective functions. We provide the analytic expression of all discovered AFs and evaluate them on various global optimization benchmarks and hyperparameter optimization tasks. We show how FunBO identifies AFs that generalize well in and out of the training distribution of functions, thus outperforming established general-purpose AFs and achieving competitive performance against AFs that are customized to specific function types and are learned via transfer-learning algorithms.", "keywords": [ "Bayesian optimization", "LLM", "acquisition function", "meta-learning" ], "rating_list": [ 6, 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3, 2 ] }, { "paper_id": "ORxjH9kTp8", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Aaron David Tucker", "gender": "unknown", "institution": "FAR AI", "country": "US", "position": "Researcher" }, { "name": "Adam Gleave", "gender": "Male", "institution": "FAR.AI", "country": "US", "position": "Principal Researcher" }, { "name": "Chris Cundy", "gender": "Male", "institution": "FAR AI", "country": "US", "position": "Researcher" }, { "name": "Maximilian Li", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "Undergrad student" }, { "name": "Mohammad Taufeeque", "gender": "unknown", "institution": "FAR.AI", "country": "US", "position": "Researcher" }, { "name": "Philip Quirke", "gender": "Male", "institution": "Martian", "country": "US", "position": "Researcher" }, { "name": "Adria Garriga-Alonso", "gender": "Male", "institution": "FAR", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 57, "n_ref": 107, "n_ref_all": 152, "n_fig": 19, "n_tab": 9, "L_tab": 2349, "n_element_tab": 322, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 390, "n_element_tab_1": 36, "formula_len_all": 51, "formula_len_all_1": 0, "len_all": 196522, "len_all_1": 68338, "len_abs": 1119, "len_title": 105, "len_sents": 58966, "len_sents_1": 32988, "n_sents": 512, "n_sents_1": 276, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1111, "title": "Planning in a recurrent neural network that plays Sokoban", "abs": "How a neural network (NN) generalizes to novel situations depends on whether it has learned to select actions heuristically or via a planning process. Guez et al., (2019, \"An investigation of model-free planning\") found that recurrent NN (RNN) trained to play Sokoban appears to plan, with extra computation steps improving the RNN's success rate. We replicate and expand on their behavioral analysis, finding the RNN learns to give itself extra computation steps in complex situations by \"pacing\" in cycles. Moreover, we train linear probes that predict the future actions taken by the network and find that intervening on the hidden state using these probes controls the agent\u2019s subsequent actions. Leveraging these insights, we perform model surgery, enabling the convolutional NN to generalize beyond its $10 \\times 10$ architectural limit to arbitrarily sized levels. The resulting model solves challenging, highly off-distribution levels. We open-source our model and code, and believe its small size (1.29M parameters) makes it an excellent model organism to deepen our understanding of learned planning.", "keywords": [ "Interpretability", "Mechanistic Interpretability", "Planning", "LSTM", "Reinforcement Learning" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 1, 2, 1, 3 ], "presentation_list": [ 2, 1, 1, 2 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "OQqNieeivq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Chansung Park", "gender": "Male", "institution": "Electronics and Telecommunications Research Institute", "country": "KR", "position": "Senior Researcher" }, { "name": "Fan Wang", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "MPhil student" }, { "name": "Jing Tang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Juyong Jiang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Sunghun Kim", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "Associate Professor" } ], "n_formula": 37, "n_formula_1": 12, "n_ref_uni": 51, "n_ref": 160, "n_ref_all": 193, "n_fig": 13, "n_tab": 13, "L_tab": 8655, "n_element_tab": 988, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2387, "n_element_tab_1": 376, "formula_len_all": 2334, "formula_len_all_1": 1081, "len_all": 266698, "len_all_1": 67079, "len_abs": 1388, "len_title": 120, "len_sents": 85419, "len_sents_1": 28487, "n_sents": 648, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1296, "title": "KaSA: Knowledge-Aware Singular-Value Adaptation of Large Language Models", "abs": "The increasing sizes of large language models (LLMs) result in significant computational overhead and memory usage when adapting these models to specific tasks or domains. Various parameter-efficient fine-tuning (PEFT) methods have been devised to mitigate these challenges by training a small set of parameters for the task-specific updates of the model weights. Among PEFT methods, LoRA stands out for its simplicity and efficiency, inspiring the development of a series of variants. However, LoRA and its successors disregard the knowledge that is noisy or irrelevant to the targeted task, detrimentally impacting model performance and leading to suboptimality. To address this limitation, we introduce Knowledge-aware Singular-value Adaptation (KaSA), a PEFT method that leverages singular value decomposition (SVD) with knowledge-aware singular values to dynamically activate knowledge based on its relevance to the task at hand. We conduct extensive experiments across a range of LLMs on tasks spanning natural language understanding (NLU), generation (NLG), and instruction following. The experimental results demonstrate that KaSA consistently outperforms FFT and 14 popular PEFT baselines across 8 benchmarks and 4 synthetic datasets, underscoring our method's efficacy and adaptability.", "keywords": [ "Large Language Models", "Parameter-efficient Fine-tuning", "Singular Value Decomposition" ], "rating_list": [ 6, 5, 8, 8, 6 ], "soundness_list": [ 3, 2, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 4, 4, 2 ], "confidence_list": [ 2, 4, 4, 5, 3 ] }, { "paper_id": "OPmYlaixqO", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Ben Fei", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Jingyi Xu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "LEI BAI", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Shengnan Wang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Weidong Yang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yeqi Luo", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Keyi Liu", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 1, "n_ref_uni": 36, "n_ref": 52, "n_ref_all": 69, "n_fig": 8, "n_tab": 3, "L_tab": 1345, "n_element_tab": 150, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 621, "n_element_tab_1": 7, "formula_len_all": 465, "formula_len_all_1": 151, "len_all": 103105, "len_all_1": 45621, "len_abs": 1337, "len_title": 73, "len_sents": 31092, "len_sents_1": 22709, "n_sents": 187, "n_sents_1": 128, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1346, "title": "SIFM: A Foundation Model for Multi-granularity Arctic Sea Ice Forecasting", "abs": "Arctic sea ice performs a vital role in global climate and has paramount impacts on both polar ecosystems and coastal communities. \nIn the last few years, multiple deep learning based pan-Arctic sea ice concentration (SIC) forecasting methods have emerged and showcased superior performance over physics-based dynamical models. \nHowever, previous methods forecast SIC at a fixed temporal granularity, e.g. sub-seasonal or seasonal, thus only leveraging inter-granularity information and overlooking the plentiful inter-granularity correlations.\nSIC at various temporal granularities exhibits cumulative effects and are naturally consistent, with short-term fluctuations potentially impacting long-term trends and long-term trends provides effective hints for facilitating short-term forecasts in Arctic sea ice.\nTherefore, in this study, we propose to cultivate temporal multi-granularity that naturally derived from Arctic sea ice reanalysis data and provide a unified perspective for modeling SIC via our Sea Ice Foundation Model. \nSIFM is delicately designed to leverage both intra-granularity and inter-granularity information for capturing granularity-consistent representations that promote forecasting skills. \nOur extensive experiments show that SIFM outperforms off-the-shelf deep learning models for their specific temporal granularity.", "keywords": [ "Arctic Sea Ice Forecasting", "Foundation Model", "Multi-granularity" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "OPdmIxdkPb", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Gonzalo Gonzalez-Pumariega", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Kushal Kedia", "gender": "Male", "institution": "Cornell University", "country": "", "position": "PhD student" }, { "name": "Sanjiban Choudhury", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Assistant Professor" }, { "name": "Wayne Chen", "gender": "unknown", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 42, "n_ref": 99, "n_ref_all": 117, "n_fig": 10, "n_tab": 11, "L_tab": 8631, "n_element_tab": 108, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 162, "n_element_tab_1": 12, "formula_len_all": 509, "formula_len_all_1": 303, "len_all": 180904, "len_all_1": 62454, "len_abs": 1222, "len_title": 94, "len_sents": 56749, "len_sents_1": 29581, "n_sents": 439, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 45, "L_abs": 1231, "title": "Query-Efficient Planning with Language Models", "abs": "Planning in complex environments requires an agent to efficiently query a world model to find a feasible sequence of actions from start to goal.\nRecent work has shown that Large Language Models (LLMs), with their rich prior knowledge and reasoning capabilities, can potentially help with planning by searching over promising states and adapting to feedback from the world. \nIn this paper, we propose and study two fundamentally competing frameworks that leverage LLMs for query-efficient planning. \nThe first uses LLMs as a heuristic within a search-based planner to select promising nodes to expand and propose promising actions. \nThe second uses LLMs as a generative planner to propose an entire sequence of actions, query the world model, and adapt based on feedback.\nWe show that while both approaches improve upon comparable baselines, using an LLM as a generative planner results in significantly fewer interactions. Our key finding is that the LLM as a planner can more rapidly adapt its planning strategies based on immediate feedback than LLM as a heuristic. We present evaluations and ablations on Robotouille and PDDL planning benchmarks and discuss connections to existing theory on query-efficient planning algorithms.", "keywords": [ "llm", "planning", "adaptive", "query-efficient" ], "rating_list": [ 5, 8, 3, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "OPSpdc25IZ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Ang Li", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Associate Professor" }, { "name": "Chuan Liu", "gender": "Female", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Chunshu Wu", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Postdoc" }, { "name": "Dongfang Liu", "gender": "unknown", "institution": "Rochester Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Ruibing Song", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Tong Geng", "gender": "Male", "institution": "University of Rochester", "country": "", "position": "Assistant Professor" }, { "name": "Yingnian Wu", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Full Professor" } ], "n_formula": 23, "n_formula_1": 20, "n_ref_uni": 35, "n_ref": 48, "n_ref_all": 66, "n_fig": 7, "n_tab": 5, "L_tab": 1752, "n_element_tab": 218, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1517, "n_element_tab_1": 181, "formula_len_all": 1147, "formula_len_all_1": 1001, "len_all": 155787, "len_all_1": 67459, "len_abs": 1452, "len_title": 148, "len_sents": 49322, "len_sents_1": 30698, "n_sents": 337, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1464, "title": "DS-LLM: Leveraging Dynamical Systems to Enhance Both Training and Inference of Large Language Models", "abs": "The training of large language models (LLMs) faces critical computational cost challenges, hindering their scaling toward AGI and broader adoption. With model sizes doubling approximately every 3.4 months and training costs surging from 64 million USD for GPT-4 in 2020 to 191 million USD for Gemini Ultra in 2023, the economic strain is unsustainable. While optimizations like quantization provide incremental improvements, they fail to address the fundamental bottleneck. In this work, we propose DS-LLM, a novel framework leveraging dynamic system (DS)-based machines, which exploit Natural Annealing to instantaneously converge to minimum energy states, enabling orders-of-magnitude gains in efficiency. Unlike traditional methods, DS-LLM maps LLM components to optimization problems solvable via Hamiltonian configurations and utilizes DS machines' continuous electric current flow for hardware-native gradient descent during training. We mathematically demonstrate the equivalence between existing LLMs and DS-LLMs and offer a viable approach to build a DS-LLM from a trained conventional LLM. Evaluations using different sizes of models showcase orders of magnitudes speedup and energy reduction on both training and inference, while maintaining consistent accuracy. Furthermore, we provide detail analysis and discussion on the potential challenges and solutions of this emerging computing diagram, aiming to provide a solid foundation for future research.", "keywords": [ "Large Language Model; nature-powered computing; dynamic physical system" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 2, 3, 3 ] }, { "paper_id": "OPKBPz6Qnz", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Ali Aghdaei", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Soumen Sikder Shuvo", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zhuo Feng", "gender": "Male", "institution": "Stevens Institute of Technology", "country": "", "position": "Associate Professor" } ], "n_formula": 43, "n_formula_1": 25, "n_ref_uni": 66, "n_ref": 115, "n_ref_all": 145, "n_fig": 8, "n_tab": 5, "L_tab": 1059, "n_element_tab": 73, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 819, "n_element_tab_1": 52, "formula_len_all": 2638, "formula_len_all_1": 1332, "len_all": 160749, "len_all_1": 67661, "len_abs": 1231, "len_title": 119, "len_sents": 44682, "len_sents_1": 30229, "n_sents": 343, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1233, "title": "A Spectral Framework for Assessing the Geodesic Distance Between Graphs", "abs": "This paper presents a spectral framework for quantifying the differentiation between graph data samples by introducing a novel metric named Graph Geodesic Distance (GGD). For two different graphs with the same number of nodes, our framework leverages a spectral graph matching procedure to find node correspondence so that the geodesic distance between them can be subsequently computed by solving a generalized eigenvalue problem associated with their Laplacian matrices. For graphs of different sizes, a resistance-based spectral graph coarsening scheme is introduced to reduce the size of the larger graph while preserving the original spectral properties. We show that the proposed GGD metric can effectively quantify dissimilarities between two graphs by encapsulating their differences in key structural (spectral) properties, such as effective resistances between nodes, cuts, the mixing time of random walks, etc. Through extensive experiments comparing with the state-of-the-art metrics, such as the latest Tree-Mover's Distance (TMD) metric, the proposed GGD metric shows significantly improved performance for graph classification and stability evaluation of GNNs, especially when only partial node features are available.", "keywords": [ "Graph Theory", "Graph Neural Network", "Graph Laplacian", "Riemannian Manifold", "Geodesic", "Graph Classification" ], "rating_list": [ 5, 8, 3, 3, 3 ], "soundness_list": [ 2, 3, 1, 3, 2 ], "presentation_list": [ 3, 3, 1, 2, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3, 5 ] }, { "paper_id": "OOywAeccTZ", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Guoyu Lu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Assistant Professor" }, { "name": "jinchang zhang", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 30, "n_ref_uni": 41, "n_ref": 54, "n_ref_all": 79, "n_fig": 5, "n_tab": 7, "L_tab": 3483, "n_element_tab": 452, "n_fig_1": 5, "n_tab_1": 9, "L_tab_1": 5818, "n_element_tab_1": 481, "formula_len_all": 2012, "formula_len_all_1": 2150, "len_all": 141794, "len_all_1": 76330, "len_abs": 1868, "len_title": 150, "len_sents": 34558, "len_sents_1": 29418, "n_sents": 254, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1878, "title": "Segmentation-Enhanced Depth Estimation Using Camera Model Based Self-supervised Contrastive Learning", "abs": "Depth estimation is a key topic in the field of computer vision. Self-supervised monocular depth estimation offers a powerful method to extract 3D scene information from a single camera image, allowing training on arbitrary image sequences without the need for depth labels. However, monocular unsupervised depth estimation still cannot address the issue of scale and often requires ground truth for calibration.\nIn the deep learning era, existing methods primarily rely on relationships between images to train unsupervised neural networks, often overlooking the fundamental information provided by the camera itself. In fact, the intrinsic and extrinsic parameters of the camera can be used to compute depth information for the ground and its related areas based on physical principles. This information can offer rich supervisory signals at no additional cost. Additionally, by assuming that objects like people, cars, and buildings share the same depth as the corresponding ground, the physical depth of the entire scene can be inferred, and gaps in the depth map can be filled.\nSince some areas may have depth estimation errors, to make full use of these regions, we introduce a contrastive learning self-supervised framework. This framework consists of two networks with the same structure: the Anchor network and the Target network. While calculating depth, the network also outputs semantic segmentation results to assist in computing the physics depth, which is then used as the label for the model. Semantic segmentation can identify dynamic objects, reducing photometric reprojection errors caused by moving objects. The predictions from the Anchor network are used as pseudo-labels for training the Target network. Reliability is determined by entropy, dividing the predicted depth into positive and negative samples to maximize the use of physics depth information.", "keywords": [ "Contrastive Learning; Depth Estimation" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 2, 1 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "OOt5RMI0JC", "primary_area": "applications to neuroscience & cognitive science", "n_author": 7, "author_info_list": [ { "name": "Declan Iain Campbell", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Gianluca Bencomo", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Jake Snell", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Liyi Zhang", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Raja Marjieh", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Sreejan Kumar", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Thomas L. Griffiths", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "Professor" } ], "n_formula": 59, "n_formula_1": 11, "n_ref_uni": 52, "n_ref": 86, "n_ref_all": 167, "n_fig": 28, "n_tab": 2, "L_tab": 288, "n_element_tab": 75, "n_fig_1": 22, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5986, "formula_len_all_1": 1021, "len_all": 186328, "len_all_1": 64996, "len_abs": 1374, "len_title": 153, "len_sents": 61645, "len_sents_1": 33450, "n_sents": 407, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 105, "L_abs": 1379, "title": "Using Contrastive Learning with Generative Similarity to Learn Spaces that Capture Human Inductive Biases", "abs": "Humans rely on strong inductive biases to learn from few examples and abstract useful information from sensory data. Instilling such biases in machine learning models has been shown to improve their performance on various benchmarks including few-shot learning, robustness, and alignment. However, finding effective training procedures to achieve that goal can be challenging as psychologically-rich training data such as human similarity judgments are expensive to scale, and Bayesian models of human inductive biases are often intractable for complex, realistic domains. Here, we address this challenge by introducing a Bayesian notion of generative similarity whereby two datapoints are considered similar if they are likely to have been sampled from the same distribution. This measure can be applied to complex generative processes, including probabilistic programs. We show that generative similarity can be used to define a contrastive learning objective even when its exact form is intractable, enabling learning of spatial embeddings that express specific inductive biases. We demonstrate the utility of our approach by showing that it can be used to capture human inductive biases for geometric shapes, distinguish different abstract drawing styles that are parameterized by probabilistic programs, and capture abstract high-level categories that enable generalization.", "keywords": [ "contrastive learning", "cognitive science", "generative models", "similarity", "Bayesian inference", "human intelligence", "representations" ], "rating_list": [ 3, 6, 3, 5, 3 ], "soundness_list": [ 1, 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 3, 3, 3, 4, 3 ] }, { "paper_id": "OOqvY9yvVG", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Adriel Kuek", "gender": "Male", "institution": "DSO National Laboratories", "country": "SG", "position": "Researcher" }, { "name": "Chionh Wei Peng", "gender": "Male", "institution": "DSO National Labs", "country": "SG", "position": "Researcher" }, { "name": "Xuanming Cui", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Ser-Nam Lim", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Associate Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 90, "n_ref_all": 109, "n_fig": 7, "n_tab": 4, "L_tab": 1417, "n_element_tab": 140, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2201, "n_element_tab_1": 231, "formula_len_all": 670, "formula_len_all_1": 610, "len_all": 122365, "len_all_1": 60178, "len_abs": 691, "len_title": 113, "len_sents": 36776, "len_sents_1": 25578, "n_sents": 275, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 694, "title": "Improving Soft Unification with Knowledge Graph Embedding Methods", "abs": "Neural Theorem Provers (NTPs) present a promising framework for neuro-symbolic reasoning, combining end-to-end differentiability with the interpretability of symbolic logic programming. However, optimizing NTPs remains a significant challenge due to their complex objective landscape and gradient sparcity. On the other hand, Knowledge Graph Embedding (KGE) methods offer smooth optimization with well-defined learning objectives but often lack interpretability. In this work, we propose several strategies to integrate the strengths of NTPs and KGEs. By incorporating KGE objectives into the NTP framework, we demonstrate substantial improvements in both accuracy and computational efficiency.", "keywords": [ "Neuro-Symbolic", "Knowledge Graph Embedding" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "OO6lPenO4c", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Christopher Brinton", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "Dong-Jun Han", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Rohit Parasnis", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Seyyedali Hosseinalipour", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Zhan-Lun Chang", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "PhD student" } ], "n_formula": 83, "n_formula_1": 17, "n_ref_uni": 29, "n_ref": 49, "n_ref_all": 53, "n_fig": 2, "n_tab": 5, "L_tab": 6746, "n_element_tab": 832, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 2995, "n_element_tab_1": 523, "formula_len_all": 12760, "formula_len_all_1": 1630, "len_all": 157637, "len_all_1": 79959, "len_abs": 1707, "len_title": 126, "len_sents": 44547, "len_sents_1": 33901, "n_sents": 332, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 125, "L_abs": 1710, "title": "Federated Learning with Dynamic Client Arrival and Departure: Convergence and Rapid Adaptation via Initial Model Construction", "abs": "While most existing federated learning (FL) approaches assume a fixed set of clients in the system, in practice, clients can dynamically leave or join the system depending on their needs or interest in the specific task. This dynamic FL setting introduces several key challenges: (1) the objective function dynamically changes depending on the current set of clients, unlike traditional FL approaches that maintain a static optimization goal; (2) the current global model may not serve as the best initial point for the next FL rounds and could potentially lead to slow adaptation, given the possibility of clients leaving or joining the system. In this paper, we consider a dynamic optimization objective in FL that seeks the optimal model tailored to the currently active set of clients. Building on our probabilistic framework that provides direct insights into how the arrival and departure of different types of clients influence the shifts in optimal points, we establish an upper bound on the optimality gap, accounting for factors such as stochastic gradient noise, local training iterations, non-IIDness of data distribution, and deviations between optimal points caused by dynamic client pattern. We also propose an adaptive initial model construction strategy that employs weighted averaging guided by gradient similarity, prioritizing models trained on clients whose data characteristics align closely with the current one, thereby enhancing adaptability to the current clients. The proposed approach is validated on various datasets and FL algorithms, demonstrating robust performance across diverse client arrival and departure patterns, underscoring its effectiveness in dynamic FL environments.", "keywords": [ "Federated Learning", "Client Participation", "Dynamic Initial Model Construction", "Client Availability" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 3, 1, 1 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "ONfWFluZBI", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Rodrigo Gonz\u00e1lez Laiz", "gender": "Male", "institution": "Helmholtz Zentrum M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Steffen Schneider", "gender": "unknown", "institution": "Helmholtz Zentrum M\u00fcnchen", "country": "DE", "position": "Principal Researcher" }, { "name": "Tobias Schmidt", "gender": "Male", "institution": "Helmholtz Zentrum M\u00fcnchen", "country": "", "position": "Intern" } ], "n_formula": 78, "n_formula_1": 18, "n_ref_uni": 54, "n_ref": 128, "n_ref_all": 207, "n_fig": 29, "n_tab": 8, "L_tab": 1348, "n_element_tab": 131, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 6, "n_element_tab_1": 1, "formula_len_all": 4402, "formula_len_all_1": 1186, "len_all": 256710, "len_all_1": 70947, "len_abs": 1616, "len_title": 126, "len_sents": 88632, "len_sents_1": 30427, "n_sents": 744, "n_sents_1": 275, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 659, "title": "Self-supervised contrastive learning performs non-linear system identification", "abs": "Self-supervised learning (SSL) approaches have brought tremendous success across many tasks and domains. It has been argued that these successes can be attributed to a link between SSL and identifiable representation learning: Temporal structure and auxiliary variables ensure that latent representations are related to the true underlying generative factors of the data. Here, we deepen this connection and show that SSL can perform system identification in latent space. We propose DynCL, a framework to uncover linear, switching linear and non-linear dynamics under a non-linear observation model, give theoretical guarantees and validate them empirically.", "keywords": [ "system identification", "dynamics learning", "identifiability", "self-supervised learning" ], "rating_list": [ 8, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3, 2 ] }, { "paper_id": "ONWLxkNkGN", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Baixu", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Researcher" }, { "name": "Hao Ai", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Haofan Wang", "gender": "Male", "institution": "InstantX", "country": "CN", "position": "Researcher" }, { "name": "Jen-Yuan Huang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Peng Xing", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Tong Lin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "wangqixun", "gender": "Male", "institution": "Tencent Hunyuan", "country": "CN", "position": "Instructor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 49, "n_ref": 75, "n_ref_all": 86, "n_fig": 5, "n_tab": 3, "L_tab": 2343, "n_element_tab": 238, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1649, "n_element_tab_1": 153, "formula_len_all": 817, "formula_len_all_1": 817, "len_all": 141961, "len_all_1": 55410, "len_abs": 0, "len_title": 0, "len_sents": 33998, "len_sents_1": 23640, "n_sents": 260, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1202, "title": "InstantIR: Blind Image Restoration with Instant Generative Reference", "abs": "Handling test-time unknown degradation is the major challenge in Blind Image Restoration (BIR), necessitating high model generalization. An effective strategy is to incorporate prior knowledge, either from human input or generative model. In this paper, we introduce Instant-reference Image Restoration (InstantIR), a novel diffusion-based BIR method which dynamically adjusts generation condition during inference. We first extract a compact representation of the input via a pre-trained vision encoder. At each generation step, this representation is used to decode current diffusion latent and instantiate it in the generative prior. The degraded image is then encoded with this reference, providing robust generation condition. We observe the variance of generative references fluctuate with degradation intensity, which we further leverage as an indicator for developing a sampling algorithm adaptive to input quality. Extensive experiments demonstrate InstantIR achieves state-of-the-art performance and offering outstanding visual quality. Through modulating generative references with textual description, InstantIR can restore extreme degradation and additionally feature creative restoration.", "keywords": [ "blind image restoration; diffusion model; text-to-image model; generative model" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "ONOe6cAE9I", "primary_area": "applications to neuroscience & cognitive science", "n_author": 24, "author_info_list": [ { "name": "Aaron Batista", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Adam G. Rouse", "gender": "Male", "institution": "University of Kansas School of Medicine", "country": "US", "position": "Assistant Professor" }, { "name": "Adam Smoulder", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Andrew B. Schwartz", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Charles M. Greenspon", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Principal Researcher" }, { "name": "Dalton D. Moore", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Postdoc" }, { "name": "Fabio Rizzoglio", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "Postdoc" }, { "name": "Gary H Blumenthal", "gender": "unknown", "institution": "University of Pittsburgh", "country": "US", "position": "Postdoc" }, { "name": "Hongwei Mao", "gender": "Male", "institution": "University of Pittsburgh", "country": "US", "position": "Researcher" }, { "name": "J. Patrick Mayo", "gender": "Male", "institution": "University of Pittsburgh", "country": "US", "position": "Assistant Professor" }, { "name": "Jennifer L Collinger", "gender": "Female", "institution": "University of Pittsburgh", "country": "", "position": "Associate Professor" }, { "name": "Joel Ye", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Kristofer Bouchard", "gender": "Male", "institution": "Lawrence Berkeley National Lab", "country": "US", "position": "Assistant Professor" }, { "name": "Lee E. Miller", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "Full Professor" }, { "name": "Leila Wehbe", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Michael L Boninger", "gender": "Male", "institution": "University of Pittsburgh", "country": "US", "position": "Full Professor" }, { "name": "Nicholas G. Hatsopoulos", "gender": "unknown", "institution": "University of Chicago", "country": "US", "position": "Full Professor" }, { "name": "Nicolas Guazzelli Kunigk", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Patrick J. Marino", "gender": "Male", "institution": "Phantom Neuro", "country": "US", "position": "Senior Research Scientist" }, { "name": "Raeed H. Chowdhury", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Robert Gaunt", "gender": "Male", "institution": "University of Pittsburgh", "country": "", "position": "Associate Professor" }, { "name": "Steven Chase", "gender": "unknown", "institution": "Carnegie-Mellon University", "country": "", "position": "" }, { "name": "William Hockeimer", "gender": "unknown", "institution": "University of Pittsburgh", "country": "US", "position": "Postdoc" }, { "name": "Xuan Ma", "gender": "Male", "institution": "Northwestern University", "country": "", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 78, "n_ref": 121, "n_ref_all": 187, "n_fig": 46, "n_tab": 3, "L_tab": 669, "n_element_tab": 52, "n_fig_1": 25, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 322379, "len_all_1": 71978, "len_abs": 7042, "len_title": 88, "len_sents": 87639, "len_sents_1": 37350, "n_sents": 681, "n_sents_1": 283, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1415, "title": "A Generalist Intracortical Motor Decoder", "abs": "Mapping the relationship between neural activity and motor behavior is a central aim of sensorimotor neuroscience and neurotechnology. \nMost progress to this end has relied on restricting complexity: studying specific simple behaviors, in limited subjects, with interpretable computational models. However, current trends in deep learning suggest that modeling a breadth of neural and behavioral data all at once is not only possible, but that such a model would also benefit downstream analysis of related data. We accordingly developed Neural Data Transformer 3 (NDT3) as a foundation model for motor decoding of neural data from intracortical microelectrodes. We pretrained NDT3 with 2000 hours of neural population spiking activity paired with diverse motor covariates from over 30 monkeys and humans from 10 labs. Pretrained NDT3 is broadly useful, benefiting decoding on 8 downstream decoding tasks and generalizing to a variety of neural distribution shifts. However, we find signs that scaling over diverse neural datasets may be challenging, as scaling from 200 to 2000 hours already requires increasing model size to 350M parameters to avoid model saturation, and several downstream datasets scarcely benefit from scale. We provide two demonstrations that this scaling is at least partially limited by variability in input and output spaces across neural datasets, which pretraining alone may not resolve.", "keywords": [ "Neuroscience", "Foundation Model", "Motor Decoding", "BCI" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "ON3QLXrwVb", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 9, "author_info_list": [ { "name": "Binbin Lin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Chen Shen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Chenxi Huang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jieping Ye", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Liang Xie", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "Postdoc" }, { "name": "Ronghua Liang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Sinan Fan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Wenxiao Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaofei He", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 27, "n_ref": 47, "n_ref_all": 56, "n_fig": 4, "n_tab": 4, "L_tab": 2319, "n_element_tab": 147, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1776, "n_element_tab_1": 250, "formula_len_all": 826, "formula_len_all_1": 783, "len_all": 135437, "len_all_1": 64380, "len_abs": 1830, "len_title": 142, "len_sents": 40707, "len_sents_1": 29303, "n_sents": 337, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1843, "title": "Interpretable Analysis and Reasoning Enhancement for LLMs via Cross-Generation Reasoning Trees", "abs": "Generating diverse reasoning paths by varying the context (such as demonstrations, prompts, instructions, etc) or sampling methods (such as top-k, top-p, beam-search, etc) and then selecting appropriate paths via majority voting or verifier-based strategies to enhance the reasoning capabilities of large language models (LLMs) is a commonly recognized approach. Although both different contexts and sampling techniques can generate diverse contents, using sampling methods alone does not significantly enhance the diversity of generations. Context variation, however, while fostering greater diversity in reasoning, can also introduce negative effects, which causes that switching contexts can not necessarily lead to proportional improvements in performance. Therefore, there is a need to investigate how context influences LLM generation and mitigate any adverse impacts. The primary challenge lies in the inability to conduct comparative studies once divergences occur in reasoning paths generated under different contexts. Specifically, once the predicted tokens at a given step differ, it becomes unclear whether subsequent tokens in the inference path are influenced by the context or the content already generated. In this paper, we propose a Cross-Generation Reasoning Tree (CGRT) algorithm for studying the impact of different contexts on LLM generation and enhancing LLMs' reasoning performance. Experimental findings reveal that, beyond enhancing interpretability, CGRT integrates the positive effects of both context and sampling strategies more effectively than previous approaches, leading to more rational inference paths. Experiments conducted on Llama2, Llama3, and Qwen demonstrate that, when generating an equivalent number of diverse inference paths, those produced via the \"reasoning tree\" method exhibit higher accuracy.", "keywords": [ "Large Language Models", "LLM Reasoning" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "ON121aJV61", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Jingyu Liu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Lin Wang", "gender": "unknown", "institution": "CUHK", "country": "CN", "position": "PhD student" }, { "name": "Tao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaoying Tang", "gender": "Female", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Yongxin Guo", "gender": "Male", "institution": "Chinese University of HongKong, Shenzhen", "country": "CN", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 17, "n_ref_uni": 54, "n_ref": 123, "n_ref_all": 152, "n_fig": 8, "n_tab": 7, "L_tab": 1452, "n_element_tab": 131, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 915, "n_element_tab_1": 99, "formula_len_all": 2365, "formula_len_all_1": 1180, "len_all": 183522, "len_all_1": 74396, "len_abs": 1153, "len_title": 85, "len_sents": 50437, "len_sents_1": 31851, "n_sents": 432, "n_sents_1": 259, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1162, "title": "Client2Vec: Improving Federated Learning by Distribution Shifts Aware Client Indexing", "abs": "Federated Learning (FL) is a privacy-preserving distributed machine learning paradigm. Nonetheless, the substantial distribution shifts among clients pose a considerable challenge to the performance of current FL algorithms. To mitigate this challenge, various methods have been proposed to enhance the FL training process.\nThis paper endeavors to tackle the issue of data heterogeneity from another perspective---by improving FL algorithms prior to the actual training stage. Specifically, we introduce the Client2Vec mechanism, which generates a unique client index for each client before the commencement of FL training. Subsequently, we leverage the generated client index to enhance the subsequent FL training process. To demonstrate the effectiveness of the proposed Client2Vec method, we conduct three case studies that assess the impact of the client index on the FL training process. These case studies encompass enhanced client sampling, model aggregation, and local training. Extensive experiments conducted on diverse datasets and model architectures show the efficacy of Client2Vec across all three case studies. Our code will be publicly available.", "keywords": [ "Federated Leaning", "Domain Index", "Data Heterogeneity" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "ON0JjUtw9B", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Bisheng Yang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Chen Long", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Conglang Zhang", "gender": "unknown", "institution": "Wuhan University", "country": "CN", "position": "Undergrad student" }, { "name": "Hang Xu", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Wenxiao Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" }, { "name": "Zhen Dong", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 22, "n_ref": 35, "n_ref_all": 45, "n_fig": 5, "n_tab": 7, "L_tab": 1821, "n_element_tab": 222, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 754, "n_element_tab_1": 127, "formula_len_all": 574, "formula_len_all_1": 574, "len_all": 110506, "len_all_1": 59240, "len_abs": 1804, "len_title": 101, "len_sents": 30371, "len_sents_1": 27376, "n_sents": 236, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1734, "title": "WIN: Variable-View Implicit LIDAR Upsampling Network", "abs": "LiDAR upsampling aims to increase the resolution of sparse point sets obtained from low-cost sensors, providing better performance for various downstream tasks. Most existing methods transform LiDAR points into range view and design complex neighborhood point interpolation strategies to improve the resolution of point clouds. However, they overlook that the range image representation is insufficient to describe complex local geometric relationships, which limits the geometric accuracy of upsampled points.\nTo address this issue, we propose WIN, a Variable-View Implicit Network. \nFirst, we decouple the range image into two novel virtual view representations to compensate for the missing geometric information during range view-based interpolation. Secondly, to fuse the interpolation results of different views, we model the fusion process as a probability distribution problem instead of a simple binary classification task. We introduce a contrast selection module, which captures the feature differences between two representations and outputs the view confidence score for each upsampled point. The underlying idea is that the complementarity of the information is proportional to the feature difference between the two views. Motivated by this insight, we design a loss function based on probabilistic modeling to supervise the results of the selection module.\nAs a result, compared with the current state-of-the-art (SOTA) method ILN, WIN introduces a small number of parameters (+0.4M) but achieves a +4.5\\% increase in the MAE metric on the CARLA dataset. Furthermore, our method outperforms all existing methods in a downstream task (Depth Completion). The pre-trained model and code will be released upon acceptance.", "keywords": [ "LiDAR Upsampling", "Deep Learning", "Autonomous vehicle system" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "OMN1UtDbRf", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Mahdi Jalili", "gender": "Male", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "Full Professor" }, { "name": "Nafiseh Imanian", "gender": "Female", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "PhD student" }, { "name": "Parham Moradi DW", "gender": "Male", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "Postdoc" }, { "name": "laleh tafakori", "gender": "Female", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "Assistant Professor" } ], "n_formula": 33, "n_formula_1": 29, "n_ref_uni": 24, "n_ref": 41, "n_ref_all": 57, "n_fig": 5, "n_tab": 7, "L_tab": 4366, "n_element_tab": 305, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1700, "formula_len_all_1": 1365, "len_all": 156788, "len_all_1": 59888, "len_abs": 1390, "len_title": 138, "len_sents": 47550, "len_sents_1": 28405, "n_sents": 328, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1449, "title": "KD-HGRL: Knowledge Distillation for Multi-Task Heterogeneous Graph Representation Learning", "abs": "Heterogeneous graphs, characterized by diverse node and edge types, are central to many real-world applications, including social networks, biological systems, and recommendation engines. While Graph Neural Networks (GNNs) are effective for graph representation learning, their reliance on extensive labeled data, high computational cost, and long inference times limit scalability, especially for heterogeneous graphs. To address these challenges, we propose KD-HGRL, which leverages \\textbf{K}nowledge \\textbf{D}istillation for multi-task \\textbf{H}eterogenous \\textbf{G}raph \\textbf{R}epresentation \\textbf{L}earning. KD-HGRL uses self-supervised contrastive learning across semantic and topological views to generate robust, label-free node embeddings in the teacher phase. These embeddings are distilled into a lightweight student model, enabling efficient task-specific outputs such as node classification and link prediction with significantly reduced inference time. Experiments on benchmark datasets demonstrate KD-HGRL\u2019s superior performance and efficiency compared to state-of-the-art methods. The framework captures both local and global graph structures, eliminates the need for labeled data, and scales effectively to large graphs. Key novelties, such as a multi-view teacher model, contrastive alignment, and a lightweight student model, make KD-HGRL a versatile and efficient solution for heterogeneous graph representation learning.", "keywords": [ "Knowledge Distillation", "Graph Neural Network", "Embedding Transfer", "Heterogeneous Graph", "Contrastive Learning", "Self-Supervised Learning." ], "rating_list": [ 5, 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 4, 4 ] }, { "paper_id": "OMFssKwpyo", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "H. Vincent Poor", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Full Professor" }, { "name": "Kang Pu", "gender": "unknown", "institution": "State University of New York at Stony Brook", "country": "", "position": "PhD student" }, { "name": "Mian Jia", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Nam H Nguyen", "gender": "Male", "institution": "Capital One", "country": "US", "position": "Sr. Distinguished Applied Researcher" }, { "name": "Yue Zhao", "gender": "Male", "institution": "State University of New York, Stony Brook", "country": "", "position": "Assistant Professor" }, { "name": "Yuqi Nie", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 34, "n_ref": 49, "n_ref_all": 60, "n_fig": 5, "n_tab": 7, "L_tab": 2423, "n_element_tab": 196, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 712, "n_element_tab_1": 115, "formula_len_all": 876, "formula_len_all_1": 809, "len_all": 124477, "len_all_1": 59068, "len_abs": 1894, "len_title": 145, "len_sents": 39624, "len_sents_1": 27885, "n_sents": 279, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1942, "title": "SONNET: Solar-disaggregation-based Day-ahead Probabilistic Net Load Forecasting with Transformers", "abs": "The global transition towards sustainable energy sources has positioned solar power as a cornerstone of modern electricity systems, underscoring the critical need for advanced forecasting techniques in grid management. Accurate net load forecasting is crucial for efficient and reliable power grid operations, especially with the rapid deployment of behind-the-meter (BTM) renewable energy sources such as rooftop solar. Notably, BTM solar generation is neither controlled nor monitored by utilities and hence only net load data are observed. Different from load forecasting, net load forecasting faces new challenges because BTM solar, a major component of net load, behaves very differently from and is much more variable than loads. To exploit the distinct natures of solar generation and load and unlock their predictive potentials, we propose ${\\bf SONNET}$, which stands for ${\\bf SO}$lar-disaggregatio${\\bf N}$-based ${\\bf NE}$t load forecasting with ${\\bf T}$ransformers. It is a novel probabilistic net load forecasting method based on disaggregating net loads into solar generation and loads and feeding both into the predictors. The method further features a) an enhanced Transformer architecture that integrates both historical and future input data, employing a combination of self-attention and cross-attention mechanisms, and b) a data augmentation method that enhances the robustness of net load forecasts against weather forecast errors. Extensive experiments are conducted based on the comprehensive real-world data set from a recent net load forecasting competition organized by the U.S. Department of Energy (DOE). It is demonstrated that our proposed method both improves the accuracy and reduces the uncertainty of net load forecasts. Notably, our proposed method significantly outperforms the state-of-the-art. The proposed techniques also have broad applications for energy and/or general forecasting-related problems.", "keywords": [ "Net load forecasting", "Probabilistic Modeling", "Transformers", "Solar Disaggregation", "Data Augmentation" ], "rating_list": [ 5, 3, 5, 3, 3 ], "soundness_list": [ 3, 1, 3, 2, 1 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 1, 3 ], "confidence_list": [ 3, 5, 3, 4, 5 ] }, { "paper_id": "OM1R87YLTc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Luxi Li", "gender": "unknown", "institution": "Beijing Normal University-Hong Kong Baptist University United International College", "country": "", "position": "PhD student" }, { "name": "Yuchen Li", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "", "position": "PhD student" }, { "name": "Yuhang He", "gender": "Male", "institution": "Research, Microsoft", "country": "CA", "position": "Researcher" }, { "name": "Zhe XuanYuan", "gender": "Male", "institution": "Beijing Normal University-Hong Kong Baptist University United International College", "country": "", "position": "Associate Professor" }, { "name": "Chen Long", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 11, "n_ref": 15, "n_ref_all": 30, "n_fig": 4, "n_tab": 4, "L_tab": 592, "n_element_tab": 82, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 483, "n_element_tab_1": 66, "formula_len_all": 591, "formula_len_all_1": 592, "len_all": 81504, "len_all_1": 51609, "len_abs": 1229, "len_title": 156, "len_sents": 28160, "len_sents_1": 25439, "n_sents": 168, "n_sents_1": 149, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1232, "title": "Multi-Task Perception in Unstructured Environments: Anti-Degradation Complementary Learning and SAMEnhancer", "abs": "While autonomous driving perception has advanced significantly in structured environments, unstructured environments still present major challenges due to the complexity of traffic participants and irregular road conditions. This paper focuses on addressing these challenges through multi-task perception, targeting drivable area segmentation and object detection in unstructured environments. A key issue in existing datasets for unstructured settings is the non-overlapping annotation of images across different tasks, which limits the efficiency of data utilization. To tackle this, we propose Anti-Degradation Complementary Learning (ADC learning), a semi-supervised approach that allows different tasks to share knowledge across unlabeled data, thereby maximizing the use of available image information. Additionally, we introduce SAMEnhancer, which integrates the Segment Anything Model (SAM) to improve segmentation quality by combining the semantic specificity of network training with the coherence of SAM\u2019s segmentation. Extensive experiments validate the effectiveness of our methods, demonstrating significant performance improvements in both segmentation and detection, especially in challenging unstructured scenarios.", "keywords": [ "autonomous driving", "unstructured environment", "mobile sam", "semi-supervised learning" ], "rating_list": [ 3, 3, 1, 1 ], "soundness_list": [ 2, 2, 1, 1 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "OLtD2vDF5X", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Martin Jaggi", "gender": "Male", "institution": "EPFL", "country": "CH", "position": "Associate Professor" }, { "name": "Simin Fan", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Xinyu Zhou", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "MS student" } ], "n_formula": 49, "n_formula_1": 24, "n_ref_uni": 79, "n_ref": 165, "n_ref_all": 218, "n_fig": 12, "n_tab": 10, "L_tab": 3627, "n_element_tab": 697, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1003, "n_element_tab_1": 239, "formula_len_all": 4224, "formula_len_all_1": 1562, "len_all": 284376, "len_all_1": 74080, "len_abs": 1754, "len_title": 132, "len_sents": 62545, "len_sents_1": 30506, "n_sents": 517, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1752, "title": "HyperINF: Unleashing the HyperPower of the Schulz's Method for Data Influence Estimation", "abs": "Influence function provides a principled method to assess the contribution of individual training samples to a specific target, yet their high computation costs limits its applications on large-scale models or datasets. \n Existing methods proposed for influence function approximation have significantly reduce the computation overheads. However, they mostly suffer from a unsatisfied accuracy due to the lack of strong convergence guarantees. The family of hyperpower methods are well-known for their rigorous convergence guarantees on matrix inverse approximation, while the matrix multiplication operation can involve intractable memory and computation costs on large-scale models.\n We propose HyperINF, an efficient and accurate influence function approximation method which leverages the hyperpower method, specifically the Schulz's iterative algorithm.\n To deal with the computation-intensive matrix multiplication, we incorporate the generalized fisher information (GFIM) as a low-rank approximation of the hessian matrix, which reduces the memory and computation overheads to a constant costs independent of ranks on LoRA-tuned models. \n We first demonstrate the superior accuracy and stability of HyperINF compared to other baselines through a synthetic convergence simulation of matrix inversion. We further validate the efficacy of HyperINFthrough extensive real-world data attribution tasks, including mislabeled data detection and data selection for LLM and VLM fine-tuning. \n On LoRA-tuned models, HyperINF achieves superior downstream performance with minimal memory and computational overhead, while other baselines suffer from significant degradation. The codebase is available at \\url{https://anonymous.4open.science/r/HyperINF-B702}.", "keywords": [ "Data Attribution", "Influnece Function" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3, 4 ] }, { "paper_id": "OL44KtasKc", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Jia wei", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Jianfei Chen", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Jintao Zhang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jun Zhu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Pengle Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 46, "n_ref": 75, "n_ref_all": 108, "n_fig": 11, "n_tab": 14, "L_tab": 5208, "n_element_tab": 279, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 4679, "n_element_tab_1": 181, "formula_len_all": 978, "formula_len_all_1": 861, "len_all": 163020, "len_all_1": 73958, "len_abs": 1112, "len_title": 128, "len_sents": 38675, "len_sents_1": 28556, "n_sents": 302, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1068, "title": "SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration", "abs": "The transformer architecture predominates across various models. As the heart of the transformer, attention has a computational complexity of $O(N^2)$, compared to $O(N)$ for linear transformations. When handling large sequence lengths, attention becomes the primary time-consuming component. Although quantization has proven to be an effective method for accelerating model inference, existing quantization methods primarily focus on optimizing the linear layer.\nIn response, we first analyze the feasibility of quantization in attention detailedly. Following that, we propose SageAttention, a highly efficient and accurate quantization method for attention. The OPS (operations per second) of our approach outperforms FlashAttention2 and xformers by about 2.1x and 2.7x, respectively. SageAttention also achieves superior accuracy performance over FlashAttention3. Comprehensive experiments confirm that our approach incurs almost no end-to-end metrics loss across diverse models\u2014including those for large language processing, image generation, and video generation.", "keywords": [ "Attention", "Quantization" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "OKzvovmUbh", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Hengrun Zhao", "gender": "unknown", "institution": "Dalian University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Huchuan Lu", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Professor" }, { "name": "Lijun Wang", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yifan Wang", "gender": "unknown", "institution": "Dalian University of Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Yu Zeng", "gender": "Not Specified", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yunzhi Zhuge", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Postdoc" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 37, "n_ref": 46, "n_ref_all": 65, "n_fig": 5, "n_tab": 8, "L_tab": 12594, "n_element_tab": 1115, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 8051, "n_element_tab_1": 713, "formula_len_all": 206, "formula_len_all_1": 190, "len_all": 161015, "len_all_1": 68880, "len_abs": 1468, "len_title": 121, "len_sents": 32889, "len_sents_1": 25786, "n_sents": 215, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1474, "title": "Learning Universal Features for Generalizable Image Forgery Localization", "abs": "In recent years, advanced image editing and generation methods have rapidly evolved, making detecting and locating forged image content increasingly challenging. Most existing image forgery detection methods rely on identifying the edited traces left in the image. However, because the traces of different forgeries are distinct, these methods can identify familiar forgeries included in the training data but struggle to handle unseen ones.\nIn response, we present an approach for Generalizable Image Forgery Localization (GIFL). Once trained, our model can detect both seen and unseen forgeries, providing a more practical and efficient solution to counter false information in the era of generative AI. \nOur method focuses on learning general features from the pristine content rather than traces of specific forgeries, which are relatively consistent across different types of forgeries and therefore can be used as universal features to locate unseen forgeries. Additionally, as existing image forgery datasets are still dominated by traditional hand-crafted forgeries, we construct a new dataset consisting of images edited by various popular deep generative image editing methods to further encourage research in detecting images manipulated by deep generative models. Extensive experimental results show that the proposed approach outperforms state-of-the-art methods in the detection of unseen forgeries and also demonstrates competitive results for seen forgeries.", "keywords": [ "image forgery detection", "inpainting detection", "forgery localization", "image forgery dataset" ], "rating_list": [ 3, 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 4, 5 ] }, { "paper_id": "OKnsCAZlSc", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Haiyan Zhang", "gender": "Female", "institution": "Xi'an University of Electronic Science and Technology", "country": "", "position": "PhD student" }, { "name": "Mengfan Xu", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Mengyao Zhu", "gender": "unknown", "institution": "Xidian University", "country": "", "position": "PhD student" }, { "name": "Robert Deng", "gender": "Male", "institution": "Singapore Management University", "country": "", "position": "Full Professor" }, { "name": "Shunjie Yuan", "gender": "Male", "institution": "Xi'an University", "country": "", "position": "PhD student" }, { "name": "Siqi Ma", "gender": "Female", "institution": "University of New South Wales", "country": "", "position": "Assistant Professor" }, { "name": "Xinghua Li", "gender": "Male", "institution": "Xidian University ", "country": "", "position": "Full Professor" } ], "n_formula": 25, "n_formula_1": 20, "n_ref_uni": 28, "n_ref": 58, "n_ref_all": 70, "n_fig": 5, "n_tab": 5, "L_tab": 3267, "n_element_tab": 196, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1569, "formula_len_all_1": 1468, "len_all": 128423, "len_all_1": 68514, "len_abs": 1599, "len_title": 153, "len_sents": 36985, "len_sents_1": 33425, "n_sents": 287, "n_sents_1": 257, "venue": "ICLR2025", "decision": "withdraw", "L_title": 112, "L_abs": 1610, "title": "A robust federated learning client selection with combinatorial data class representations and data augmentation", "abs": "The federated learning (FL) client selection scheme can effectively mitigate global model performance degradation caused by the random aggregation of clients with heterogeneous data. Simultaneously, research has exposed FL's susceptibility to backdoor attacks. However herein lies the dilemma, traditional client selection methods and backdoor defenses stand at odds, so their integration is an elusive goal.\nTo resolve this, we introduce Grace, a resilient client selection framework blending combinational class sampling with data augmentation. On the client side, Grace first proposes a local model purification method, fortifying the model's defenses by bolstering its innate robustness. After, local class representations are extracted for server-side client selection. This approach not only shields benign models from backdoor tampering but also allows the server to glean insights into local class representations without infringing upon the client's privacy.\nOn the server side, Grace introduces a novel representation combination sampling method. Clients are selected based on the interplay of their class representations, a strategy that simultaneously weeds out malicious actors and draws in clients whose data holds unique value.\nOur extensive experiments highlight Grace's capabilities. The results are compelling: Grace enhances defense performance by over 50\\% compared to state-of-the-art (SOTA) backdoor defenses, and, in the best case, improves accuracy by 3.19\\% compared to SOTA client selection schemes. Consequently, Grace achieves substantial advancements in both security and accuracy.", "keywords": [ "Federated Learning", "Client Selection", "Backdoor Defense", "Data Augmentation", "Representation Learning." ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 3, 5, 4, 5 ] }, { "paper_id": "OJsMGsO6yn", "primary_area": "applications to neuroscience & cognitive science", "n_author": 7, "author_info_list": [ { "name": "Daniel Rueckert", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Emma Claire Robinson", "gender": "Female", "institution": "King's College London", "country": "", "position": "Lecturer in Biomedical Engineering" }, { "name": "Gabriel B\u00e9n\u00e9dict", "gender": "unknown", "institution": "University of Amsterdam", "country": "", "position": "PhD student" }, { "name": "Logan Zane John Williams", "gender": "Male", "institution": "King's College London", "country": "", "position": "PhD student" }, { "name": "Robert Leech", "gender": "unknown", "institution": "King's College London, University of London", "country": "", "position": "Full Professor" }, { "name": "Simon Dahan", "gender": "Male", "institution": "King's College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Yourong Guo", "gender": "Female", "institution": "King's College London, University of London", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 60, "n_ref": 178, "n_ref_all": 212, "n_fig": 8, "n_tab": 3, "L_tab": 2246, "n_element_tab": 199, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 657, "n_element_tab_1": 56, "formula_len_all": 535, "formula_len_all_1": 489, "len_all": 190034, "len_all_1": 65833, "len_abs": 1694, "len_title": 150, "len_sents": 48518, "len_sents_1": 31502, "n_sents": 297, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 1723, "title": "SIM: Surface-based fMRI Analysis for Inter-Subject Multimodal Decoding from Movie-Watching Experiments", "abs": "Current AI frameworks for brain decoding and encoding, typically train and test models within the same datasets. This limits their utility for brain computer interfaces (BCI) or neurofeedback, for which it would be useful to pool experiences across individuals to better simulate stimuli not sampled during training. A key obstacle to model generalisation is the degree of variability of inter-subject cortical organisation, which makes it difficult to align or compare cortical signals across participants. In this paper, we address this through the use of surface vision transformers, which build a generalisable model of cortical functional dynamics through encoding the topography of cortical networks and their interactions as a moving image across a surface. This is then combined with tri-modal self-supervised contrastive (CLIP) alignment of audio, video, and fMRI modalities to enable the retrieval of visual and auditory stimuli from patterns of cortical activity (and vice-versa). We validate our approach on 7T task-fMRI data from 174 healthy participants engaged in the movie-watching experiment from the Human Connectome Project (HCP). Results show that it is possible to detect which movie clips an individual is watching purely from their brain activity, even for individuals and movies *not seen during training*. Further analysis of attention maps reveals that our model captures individual patterns of brain activity that reflect semantic and visual systems. This opens the door to future personalised simulations of brain function. Code \\& pre-trained models will be made available at \\url{https://github.com/}, processed data for training will be available upon request at \\url{https://gin.g-node.org}", "keywords": [ "task-fMRI", "cortical analysis", "transformers", "multimodal learning", "contrastive learning", "self-supervised learning" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "OJd3ayDDoF", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 24, "author_info_list": [ { "name": "Bill Qian", "gender": "Male", "institution": "Yale University", "country": "", "position": "Undergrad student" }, { "name": "Binyuan Hui", "gender": "Female", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Bowen Li", "gender": "unknown", "institution": "Shanghai AI Lab", "country": "", "position": "Researcher" }, { "name": "Boxuan Li", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Frank F. Xu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Fuqiang Li", "gender": "unknown", "institution": "Xiamen BaishanCloud Technology Co., Ltd", "country": "", "position": "Researcher" }, { "name": "Graham Neubig", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Heng Ji", "gender": "Female", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Hoang H. Tran", "gender": "unknown", "institution": "Ho Chi Minh City University of Technology", "country": "VN", "position": "Undergrad student" }, { "name": "Jaskirat Singh", "gender": "unknown", "institution": "Australian National University", "country": "AU", "position": "PhD student" }, { "name": "Jiayi Pan", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Junyang Lin", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Mingchen Zhuge", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Mingzhang Zheng", "gender": "unknown", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Niklas Muennighoff", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Ren Ma", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Robert Brennan", "gender": "Male", "institution": "All Hands AI", "country": "US", "position": "Researcher" }, { "name": "Blind Name", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Xingyao Wang", "gender": "Male", "institution": "All Hands AI", "country": "US", "position": "Researcher" }, { "name": "Yanjun Shao", "gender": "Male", "institution": "Yale University", "country": "US", "position": "MS student" }, { "name": "Yizhe Zhang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Yueqi Song", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Yufan Song", "gender": "Male", "institution": "RisingWave Labs", "country": "US", "position": "Researcher" }, { "name": "Hao Peng", "gender": "Male", "institution": "Department of Computer Science, University of Illinois Urbana-Champaign", "country": "", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 24, "n_ref": 59, "n_ref_all": 71, "n_fig": 4, "n_tab": 11, "L_tab": 8907, "n_element_tab": 499, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 4665, "n_element_tab_1": 427, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 232058, "len_all_1": 61860, "len_abs": 1391, "len_title": 123, "len_sents": 51538, "len_sents_1": 25844, "n_sents": 439, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1270, "title": "OpenHands: An Open Platform for AI Software Developers as Generalist Agents", "abs": "Software is one of the most powerful tools that we humans have at our disposal; it allows a skilled programmer to interact with the world in complex and profound ways. At the same time, thanks to improvements in large language models (LLMs), there has also been a rapid development in AI agents that interact with and effect change in their surrounding environments. In this paper, we introduce OpenHands, a platform for the development of powerful and flexible AI agents that interact with the world in similar ways to a human developer: by writing code, interacting with a command line, and browsing the web. We describe how the platform allows for the implementation of new agents, utilization of various LLMs, safe interaction with sandboxed environments for code execution, and incorporation of evaluation benchmarks. Based on our currently incorporated benchmarks, we perform an evaluation of agents over 13 challenging tasks, including software engineering (e.g., SWE-Bench) and web browsing (e.g., WebArena), amongst others. Released under the permissive MIT license, OpenHands is a community project spanning academia and industry with more than 2K contributions from over 186 contributors in less than six months of development, and will improve going forward.", "keywords": [ "AI agents", "evaluation", "infrastructure", "benchmark" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 4, 4, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "OJUcOLOLXL", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 8, "author_info_list": [ { "name": "Kounianhua Du", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Qingyao Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Ruiming Tang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Wei Xia", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Weinan Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xinyi Dai", "gender": "Female", "institution": "Huawei Noah's Ark Lab", "country": "", "position": "Researcher" }, { "name": "Yasheng Wang", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "", "position": "Researcher" }, { "name": "Yong Yu", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 43, "n_ref": 89, "n_ref_all": 106, "n_fig": 7, "n_tab": 3, "L_tab": 1358, "n_element_tab": 245, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1058, "n_element_tab_1": 288, "formula_len_all": 295, "formula_len_all_1": 295, "len_all": 157129, "len_all_1": 73409, "len_abs": 1427, "len_title": 136, "len_sents": 47619, "len_sents_1": 34574, "n_sents": 348, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1436, "title": "RethinkMCTS: Refining Erroneous Thoughts in Monte Carlo Tree Search for Code Generation", "abs": "LLM agents enhanced by tree search algorithms have shown significant performance in code generation. However, existing search methods generally operate directly in the code language space, leading to suboptimal search quality due to ignoring the reasoning process behind the code. Specifically, two key challenges remain largely unaddressed: 1) A lack of exploration for the reasoning process, which is essential for high-reasoning-demand tasks like code generation, and 2) Inadequate search quality due to the absence of refinement mechanism. In this paper, we introduce RethinkMCTS, a framework that explores and refines the reasoning process for generating code. Specifically, we employ MCTS to search for the thoughts before code generation and integrate MCTS with a refinement mechanism called \"rethink\", which incorporates fine-grained code execution feedback to refine erroneous thoughts during the search. It ensures the search path aligns with the better reasoning, improving overall search quality. Through extensive experiments, we demonstrate that RethinkMCTS outperforms previous search-enhanced and feedback-enhanced code generation baselines. On the HumanEval dataset, it boosts the pass@1 of GPT-3.5-turbo from 70.12 to 89.02 and that of GPT-4o-mini from 87.20 to 94.51. By conducting thought-level exploration and integrating the rethink mechanism, it significantly enhances the search quality of the entire search tree", "keywords": [ "Code Generation", "Large Language Model Agent", "Monte Carlo Tree Search" ], "rating_list": [ 6, 6, 3, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "OIvg3MqWX2", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Bao Wang", "gender": "Male", "institution": "University of Utah", "country": "US", "position": "Assistant Professor" }, { "name": "Qi Tang", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Shih-Hsin Wang", "gender": "Male", "institution": "University of Utah", "country": "", "position": "PhD student" }, { "name": "Yuan-En Sun", "gender": "unknown", "institution": "University of Utah", "country": "US", "position": "PhD student" }, { "name": "Yuhao Huang", "gender": "Male", "institution": "University of Utah", "country": "US", "position": "PhD student" }, { "name": "Justin Baker", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 51, "n_ref": 122, "n_ref_all": 178, "n_fig": 8, "n_tab": 7, "L_tab": 1754, "n_element_tab": 237, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 306, "n_element_tab_1": 27, "formula_len_all": 1153, "formula_len_all_1": 1353, "len_all": 217344, "len_all_1": 76330, "len_abs": 1115, "len_title": 137, "len_sents": 71175, "len_sents_1": 36071, "n_sents": 578, "n_sents_1": 278, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 89, "L_abs": 1122, "title": "A Theoretically-Principled Sparse, Connected, and Rigid Graph Representation of Molecules", "abs": "Graph neural networks (GNNs) -- learn graph representations by exploiting graph's sparsity, connectivity, and symmetries -- have become indispensable for learning geometric data like molecules. However, the most used graphs (e.g., radial cutoff graphs) in molecular modeling lack theoretical guarantees for achieving connectivity and sparsity simultaneously, which are essential for the performance and scalability of GNNs. Furthermore, existing widely used graph construction methods for molecules lack rigidity, limiting GNNs' ability to exploit graph nodes' spatial arrangement. In this paper, we introduce a new hyperparameter-free graph construction of molecules and beyond with sparsity, connectivity, and rigidity guarantees. Remarkably, our method consistently generates connected and sparse graphs with the edge-to-node ratio being bounded above by 3. Our graphs' rigidity guarantees that edge distances and dihedral angles are sufficient to uniquely determine general spatial arrangements of atoms. We substantiate the effectiveness and efficiency of our proposed graphs in various molecular modeling benchmarks.", "keywords": [ "Graph representation", "sparsity", "connectivity", "rigidity", "molecules", "learning" ], "rating_list": [ 8, 6, 10, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "OIqOpdyhTd", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Boyuan Cao", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Hongming Shan", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Jiaxin Ye", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Yujie Wei", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 5, "n_ref_uni": 37, "n_ref": 79, "n_ref_all": 126, "n_fig": 24, "n_tab": 12, "L_tab": 3778, "n_element_tab": 625, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1774, "n_element_tab_1": 204, "formula_len_all": 1127, "formula_len_all_1": 250, "len_all": 170980, "len_all_1": 59249, "len_abs": 175, "len_title": 155, "len_sents": 47283, "len_sents_1": 23501, "n_sents": 440, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1516, "title": "AP-LDM: Attentive and Progressive Latent Diffusion Model for Training-Free High-Resolution Image Generation", "abs": "Latent diffusion models (LDMs), such as Stable Diffusion, often experience significant structural distortions when directly generating high-resolution (HR) images that exceed their original training resolutions. A straightforward and cost-effective solution is to adapt pre-trained LDMs for HR image generation; however, existing methods often suffer from poor image quality and long inference time.\n\nIn this paper, we propose an Attentive and Progressive LDM (AP-LDM), a novel, training-free framework aimed at enhancing HR image quality while accelerating the generation process.\n\nAP-LDM decomposes the denoising process of LDMs into two stages: (i) attentive training-resolution denoising, and (ii) progressive high-resolution denoising. The first stage generates a latent representation of a higher-quality training-resolution image through the proposed attentive guidance, which utilizes a novel parameter-free self-attention mechanism to enhance the structural consistency. The second stage progressively performs upsampling in pixel space, alleviating the severe artifacts caused by latent space upsampling.\n\nLeveraging the effective initialization from the first stage enables denoising at higher resolutions with significantly fewer steps, enhancing overall efficiency. Extensive experimental results demonstrate that AP-LDM significantly outperforms state-of-the-art methods, delivering up to a 5x speedup in HR image generation, thereby highlighting its substantial advantages for real-world applications.", "keywords": [ "Diffusion Model", "High-Resolution Image", "Attentive Guidance", "progressive pixel space upsampling" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "OIhON8zd8d", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Bastian Rieck", "gender": "Male", "institution": "University of Fribourg", "country": "CH", "position": "Full Professor" }, { "name": "Ernst R\u00f6ell", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 19, "n_ref": 28, "n_ref_all": 54, "n_fig": 8, "n_tab": 3, "L_tab": 1523, "n_element_tab": 161, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1929, "n_element_tab_1": 146, "formula_len_all": 428, "formula_len_all_1": 428, "len_all": 101445, "len_all_1": 64992, "len_abs": 656, "len_title": 87, "len_sents": 34765, "len_sents_1": 31039, "n_sents": 240, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 39, "L_abs": 661, "title": "Generative Topology for Shape Synthesis", "abs": "The _Euler Characteristic Transform_ (ECT) is a powerful invariant for assessing geometrical and topological characteristics of a large variety of objects, including graphs and embedded simplicial complexes. Although the ECT is invertible in theory, no explicit algorithm for general data sets exists. In this paper, we address this lack and demonstrate that it is possible to _learn_ the inversion, permitting us to develop a novel framework for shape generation tasks on point clouds. Our model exhibits high quality in reconstruction and generation tasks, affords efficient latent-space interpolation, and is orders of magnitude faster than existing methods.", "keywords": [ "Topological Data Analysis", "TDA", "Euler Characteristic", "Topology", "Topological Deep Learning", "Geometric Deep Learning." ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "OIEczoib6t", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Aakriti Agrawal", "gender": "Female", "institution": "University of Maryland", "country": "US", "position": "PhD student" }, { "name": "Anirudh Satheesh", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "Undergrad student" }, { "name": "Chenghao Deng", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Furong Huang", "gender": "Female", "institution": "University of Maryland", "country": "US", "position": "Associate Professor" }, { "name": "John Langford", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Mucong Ding", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Zora Che", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 10, "n_ref_uni": 15, "n_ref": 35, "n_ref_all": 56, "n_fig": 10, "n_tab": 8, "L_tab": 11545, "n_element_tab": 1302, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1890, "formula_len_all_1": 917, "len_all": 174550, "len_all_1": 73162, "len_abs": 1304, "len_title": 120, "len_sents": 47511, "len_sents_1": 33184, "n_sents": 386, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1312, "title": "EnsemW2S: Can an Ensemble of LLMs be Leveraged to Obtain a Stronger LLM?", "abs": "How can we harness the collective capabilities of multiple Large Language Models (LLMs) to create an even more powerful model? This question forms the foundation of our research, where we propose an innovative approach to weak-to-strong (w2s) generalization\u2014a critical problem in AI alignment. Our work introduces an easy-to-hard (e2h) framework for studying the feasibility of w2s generalization, where weak models trained on simpler tasks collaboratively supervise stronger models on more complex tasks. This setup mirrors real-world challenges, where direct human supervision is limited. To achieve this, we develop a novel AdaBoost-inspired ensemble method, demonstrating that an ensemble of weak supervisors can enhance the performance of stronger LLMs across classification and generative tasks on difficult QA datasets. In several cases, our ensemble approach matches the performance of models trained on ground-truth data, establishing a new benchmark for w2s generalization. We observe an improvement of upto 14\\% over existing baseline and an average improvement of 5\\% and 4\\% for binary classification and generation task respectively. This research points to a promising direction for enhancing AI through collective supervision, especially in scenarios where labeled data is sparse or insufficient.", "keywords": [ "Weak-to-Strong generalization", "Superalignment", "Ensemble Learning", "LLMs" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "OHZO0Hdfo0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Geng Li", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Hulingxiao He", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jinglin Xu", "gender": "Female", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "Associate Professor" }, { "name": "Yuxin Peng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 50, "n_ref_all": 58, "n_fig": 2, "n_tab": 3, "L_tab": 295, "n_element_tab": 43, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 324, "formula_len_all_1": 270, "len_all": 95111, "len_all_1": 66074, "len_abs": 1236, "len_title": 143, "len_sents": 26711, "len_sents_1": 26302, "n_sents": 164, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1223, "title": "Ger: Generation, Evaluation and Reflection Enhanced LLM for Knowledge Graph Question Answering", "abs": "Knowledge Graph Question Answering (KGQA) involves answering natural language questions based on information provided by knowledge graphs. Large language models (LLMs), utilizing their exceptional natural language understanding capabilities and factual knowledge from knowledge graphs, have made some progress in KGQA reasoning. However, existing methods overlook the amplification of hallucinations in large language models caused by irrelevant information within vast knowledge graphs. This oversight leads to answers containing seemingly correct but unrelated responses, decreasing reliability. In this paper, we propose $\\textbf{\\textit{Generation-Evaluation-Reflection}} $ (Ger), an LLM-enhanced reflective reasoning framework for KGQA. The Ger mechanism introduces evaluation and reflection steps during the reasoning process, allowing LLMs to leverage the factual information in KGs better and utilize their logical reasoning strengths. This process reduces errors and hallucinations while improving reasoning accuracy. Extensive experiments on multiple KGQA benchmark datasets demonstrate that Ger enhances reasoning performance, providing more reliable and interpretable results, and achieves new state-of-the-art.", "keywords": [ "KGQA", "Reasoning in Large Language Models", "Reflection" ], "rating_list": [ 3, 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 5, 3 ] }, { "paper_id": "OHOmpkGiYK", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Gang Niu", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Senior Research Scientist (tenured)" }, { "name": "Jiangchao Yao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Jianing Zhu", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "PhD student" }, { "name": "Jianliang Xu", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Full Professor" }, { "name": "Masashi Sugiyama", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Director" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 29, "n_formula_1": 10, "n_ref_uni": 64, "n_ref": 254, "n_ref_all": 329, "n_fig": 16, "n_tab": 21, "L_tab": 14763, "n_element_tab": 1698, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 4671, "n_element_tab_1": 543, "formula_len_all": 2410, "formula_len_all_1": 1075, "len_all": 285188, "len_all_1": 71923, "len_abs": 1952, "len_title": 119, "len_sents": 90339, "len_sents_1": 28925, "n_sents": 604, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1400, "title": "Decoupling the Class Label and the Target Concept in Machine Unlearning", "abs": "Machine unlearning as an emerging research topic for data regulations, aims to adjust a trained model to approximate a retrained one that excludes a portion of training data. Previous studies showed that class-wise unlearning is effective in forgetting the knowledge of target data, either through gradient ascent on the forgetting data or fine-tuning with the remaining data. However, while these methods are useful, they are insufficient as the class label and the target concept are often considered to coincide. In this work, we expand the scope by considering the label domain mismatch and investigate three problems beyond the conventional *all matched* forgetting, e.g., *target mismatch*, *model mismatch*, and *data mismatch* forgetting. We systematically analyze the new challenges in restrictively forgetting the target concept and also reveal crucial forgetting dynamics in the representation level to realize these tasks. Based on that, we propose a general framework, namely, *TARget-aware Forgetting* (TARF). It enables the additional tasks to actively forget the target concept while maintaining the rest part, by simultaneously conducting annealed gradient ascent on the forgetting data and selected gradient descent on the hard-to-affect remaining data. Empirically, various experiments under our newly introduced settings are conducted to demonstrate the effectiveness of our TARF.", "keywords": [ "Machine Unlearning", "Label Domain Mismatch" ], "rating_list": [ 8, 3, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "OGtUfA6Amo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Bowen Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Qianli Ma", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yicheng Luo", "gender": "unknown", "institution": "South China University of Technology", "country": "CN", "position": "MS student" }, { "name": "Zhen Liu", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 29, "n_ref_uni": 41, "n_ref": 102, "n_ref_all": 128, "n_fig": 4, "n_tab": 9, "L_tab": 5391, "n_element_tab": 567, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 4475, "n_element_tab_1": 226, "formula_len_all": 1687, "formula_len_all_1": 1654, "len_all": 167861, "len_all_1": 64131, "len_abs": 1582, "len_title": 129, "len_sents": 40466, "len_sents_1": 25148, "n_sents": 306, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1675, "title": "Hi-Patch: Hierarchical Patch GNN for Irregular Multivariate Time Series Modeling", "abs": "Multi-scale information is crucial for multivariate time series modeling. However, most existing time series multi-scale analysis methods treat all variables in the same manner, which is not well adaptive to Irregular Multivariate Time Series (IMTS), where different variables have distinct original scales/sampling rates. Therefore, extracting temporal and inter-variable dependencies at multiple scales\nin IMTS remains challenging. To fill this gap, we propose a hierarchical patch graph network Hi-Patch. The key components of Hi-Patch are an intra-patch graph layer and several inter-patch graph layers. The intra-patch graph layer flexibly represents and fully captures both the local temporal and inter-variable dependencies of densely sampled variables at the original scale by employing fully connected graph networks within each patch, and obtains patch-level node representations through aggregation. Subsequently, several inter-patch graph layers are stacked to form a hierarchical architecture, where each layer updates specific patch-level nodes through scale-specific graph networks, progressively representing and extracting more global temporal and inter-variable features of both sparsely and densely sampled variables, and further aggregating to produce the next patch-level node representations. The output of the last inter-patch graph\nlayer is fed into task-specific decoders to adapt to different downstream tasks. Experimental results on 8 datasets show that Hi-Patch outperforms a range of state-of-the-art models in both IMTS forecasting and classification tasks. Code is available at this repository: https://anonymous.4open.science/r/Hi-Patch-F42E.", "keywords": [ "irregular multivariate time series", "graph neural network" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "OGfyzExd69", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 6, "author_info_list": [ { "name": "Alston Lo", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jie Chen", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "Research Staff Member" }, { "name": "Michael Sun", "gender": "Male", "institution": "Computer Science and Artificial Intelligence Laboratory, Electrical Engineering & Computer Science", "country": "US", "position": "PhD student" }, { "name": "Minghao Guo", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Wojciech Matusik", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Connor Coley", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 53, "n_ref": 77, "n_ref_all": 92, "n_fig": 2, "n_tab": 2, "L_tab": 1175, "n_element_tab": 183, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1504, "n_element_tab_1": 256, "formula_len_all": 191, "formula_len_all_1": 99, "len_all": 155537, "len_all_1": 67240, "len_abs": 1484, "len_title": 95, "len_sents": 34504, "len_sents_1": 30372, "n_sents": 230, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1487, "title": "Procedural Synthesis of Synthesizable Molecules", "abs": "Designing synthetically accessible molecules and recommending analogs to unsynthesizable molecules are important problems for accelerating molecular discovery. We reconceptualize both problems using ideas from program synthesis. Drawing inspiration from syntax-guided synthesis approaches, we decouple the syntactic skeleton from the semantics of a synthetic tree to create a bilevel framework for reasoning about the combinatorial space of synthesis pathways. Given a molecule we aim to generate analogs for, we iteratively refine its skeletal characteristics via Markov Chain Monte Carlo simulations over the space of syntactic skeletons. Given a black-box oracle to optimize, we formulate a joint design space over syntactic templates and molecular descriptors and introduce evolutionary algorithms that optimize both syntactic and semantic dimensions synergistically. Our key insight is that once the syntactic skeleton is set, we can amortize over the search complexity of deriving the program's semantics by training policies to fully utilize the fixed horizon Markov Decision Process imposed by the syntactic template. We demonstrate performance advantages of our bilevel framework for synthesizable analog generation and synthesizable molecule design. Notably, our approach offers the user explicit control over the resources required to perform synthesis and biases the design space towards simpler solutions, making it particularly promising for autonomous synthesis platforms.", "keywords": [ "molecular design", "synthesis planning", "tree generation", "graph generation" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "OFukl9Qg8P", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Jizhong Han", "gender": "Male", "institution": "Institute of Information Engineering", "country": "CN", "position": "Full Professor" }, { "name": "Qiao Li", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Wangjia Yu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Xiaodan Zhang", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Xiaomeng Fu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 39, "n_ref": 49, "n_ref_all": 70, "n_fig": 9, "n_tab": 7, "L_tab": 1624, "n_element_tab": 195, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 446, "n_element_tab_1": 85, "formula_len_all": 520, "formula_len_all_1": 563, "len_all": 147605, "len_all_1": 58654, "len_abs": 1422, "len_title": 128, "len_sents": 48869, "len_sents_1": 28080, "n_sents": 357, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 79, "L_abs": 1442, "title": "Resolution Attack: Exploiting Image Compression to Deceive Deep Neural Networks", "abs": "Model robustness is essential for ensuring the stability and reliability of machine learning systems. Despite extensive research on various aspects of model robustness, such as adversarial robustness and label noise robustness, the exploration of robustness towards different resolutions, remains less explored. To address this gap, we introduce a novel form of attack: the resolution attack. This attack aims to deceive both classifiers and human observers by generating images that exhibit different semantics across different resolutions. To implement the resolution attack, we propose an automated framework capable of generating dual-semantic images in a zero-shot manner. Specifically, we leverage large-scale diffusion models for their comprehensive ability to construct images and propose a staged denoising strategy to achieve a smoother transition across resolutions. Through the proposed framework, we conduct resolution attacks against various off-the-shelf classifiers. The experimental results exhibit high attack success rate, which not only validates the effectiveness of our proposed framework but also reveals the vulnerability of current classifiers towards different resolutions. Additionally, our framework, which incorporates features from two distinct objects, serves as a competitive tool for applications such as face swapping and facial camouflage. The code is available at https://github.com/ywj1/resolution-attack.", "keywords": [ "Resolution Attack\u3001Image Generation\u3001Deep Learning Robustness\u3001Image Classification" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 2, 4, 4 ] }, { "paper_id": "OFWD0jgJ17", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Andy Edmondson", "gender": "unknown", "institution": "Heriot-Watt University", "country": "GB", "position": "PhD student" }, { "name": "Kefan Chen", "gender": "Male", "institution": "Heriot-Watt University", "country": "", "position": "PhD student" }, { "name": "Lynne Baillie", "gender": "Female", "institution": "Heriot-Watt University", "country": "", "position": "Full Professor" }, { "name": "Mehdi Rizvi", "gender": "unknown", "institution": "Heriot-Watt University", "country": "", "position": "Postdoc" }, { "name": "Wei Pang", "gender": "Male", "institution": "Heriot-Watt University", "country": "GB", "position": "Full Professor" }, { "name": "Yingfang Yuan", "gender": "Male", "institution": "Heriot-Watt University", "country": "GB", "position": "Postdoc" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 31, "n_ref": 52, "n_ref_all": 62, "n_fig": 2, "n_tab": 2, "L_tab": 715, "n_element_tab": 40, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 715, "n_element_tab_1": 40, "formula_len_all": 101, "formula_len_all_1": 101, "len_all": 87982, "len_all_1": 59119, "len_abs": 1076, "len_title": 105, "len_sents": 30265, "len_sents_1": 30256, "n_sents": 184, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1080, "title": "Graph Supervised Contrastive Learning for Geodemographics", "abs": "Geodemographic analysis is essential for understanding population characteristics and addressing socio-economic disparities across regions. However, limited research has been conducted on modelling changes in demographic data over time using Graph Neural Networks (GNNs). In this study, we address this gap by leveraging GNNs to model correlations between the 2011 census data (England \\& Wales), observing changes over time, and the Output Area Classification 2021, which reflects socio-economic differences between Output Areas. We propose a novel framework that utilises Supervised Contrastive Learning on graphs to obtain robust OA embeddings, with a particular focus on improving the model\u2019s performance for minority classes. To evaluate the effectiveness of our framework, we conducted two downstream tasks based on the 2021 OA embeddings. Our results demonstrate that the proposed approach provides valuable insights for geodemographic analysis and offers policymakers a useful tool for assessing socio-economic transitions over time, and planning ahead on the basis of it.", "keywords": [ "Graph Neural Network", "Geodemographics" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "OEabcqgoQQ", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Chenhui Xu", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Haihui Pan", "gender": "Male", "institution": "Cheetah Mobile", "country": "", "position": "Researcher" }, { "name": "Zheng Li", "gender": "Male", "institution": "Beijing Technology and Business University", "country": "CN", "position": "MS student" } ], "n_formula": 20, "n_formula_1": 8, "n_ref_uni": 20, "n_ref": 57, "n_ref_all": 61, "n_fig": 0, "n_tab": 6, "L_tab": 5899, "n_element_tab": 863, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 1690, "n_element_tab_1": 265, "formula_len_all": 2069, "formula_len_all_1": 464, "len_all": 109166, "len_all_1": 62658, "len_abs": 1292, "len_title": 152, "len_sents": 32505, "len_sents_1": 28991, "n_sents": 263, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 104, "L_abs": 1298, "title": "Can Information-Theoretic Generalization Bound Explain the Generalization of Pre-trained Language Model?", "abs": "Although language models exhibit exceptional generalization capabilities in downstream tasks after extensive text pre-training, the underlying causes behind this generalization remain unclear. Existing studies on information-theoretic generalization bounds suggest that the compression of information stored in the weights (IIW) is a crucial factor influencing a model's ability to generalize, with some experiments indicating a correlation between lower IIW and improved generalization. However, it remains uncertain whether IIW is applicable to pre-trained language models. In this work, we find that using IIW can explain why the pre-trained language models have better generalization compared to non-pre-trained language models. Unfortunately, we also discover that IIW does not consistently reflect the degree of generalization when applying IIW to study the fine-tuning process of pre-trained language models. We revisit existing IIW estimation methods, highlighting their limitations in accurately estimating IIW based on theoretical and empirical evidence. Our findings suggest that current information-theoretic generalization bounds, constrained by the limitations of IIW estimation methodologies, fail to accurately capture the generalisation performance of pre-trained language models.", "keywords": [ "Information-Theoretic Generalization Bound", "Pre-trained Language Model" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 1, 3, 3, 1 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "OEC6zOuZG1", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Samet Demir", "gender": "unknown", "institution": "Koc University", "country": "TR", "position": "PhD student" }, { "name": "Zafer Dogan", "gender": "Male", "institution": "Koc University", "country": "TR", "position": "Assistant Professor" } ], "n_formula": 213, "n_formula_1": 32, "n_ref_uni": 37, "n_ref": 81, "n_ref_all": 132, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 13816, "formula_len_all_1": 2169, "len_all": 199911, "len_all_1": 66945, "len_abs": 1287, "len_title": 156, "len_sents": 58376, "len_sents_1": 30632, "n_sents": 586, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1294, "title": "Random Features Outperform Linear Models: Effect of Strong Input-Label Correlation in Spiked Covariance Data", "abs": "Random Feature Model (RFM) with a nonlinear activation function is instrumental in understanding training and generalization performance in high-dimensional learning. While existing research has established an asymptotic equivalence in performance between the RFM and noisy linear models under isotropic data assumptions, empirical observations indicate that the RFM frequently surpasses linear models in practical applications. To address this gap, we ask, _\"When and how does the RFM outperform linear models?\"_ In practice, inputs often have additional structures that significantly influence learning. Therefore, we explore the RFM under anisotropic input data characterized by spiked covariance in the proportional asymptotic limit, where dimensions diverge jointly while maintaining finite ratios. Our analysis reveals that a high correlation between inputs and labels is a critical factor enabling the RFM to outperform linear models. Moreover, we show that the RFM performs equivalent to noisy polynomial models, where the polynomial degree depends on the strength of the correlation between inputs and labels. Our numerical simulations validate these theoretical insights, confirming the performance-wise superiority of RFM in scenarios characterized by strong input-label correlation.", "keywords": [ "Random feature model", "Gaussian equivalence", "universality" ], "rating_list": [ 6, 5, 6, 6, 3 ], "soundness_list": [ 3, 2, 3, 2, 4 ], "presentation_list": [ 2, 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 4, 2 ], "confidence_list": [ 3, 3, 3, 3, 4 ] }, { "paper_id": "OE67D1Oatr", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Abolfazl Motahari", "gender": "Male", "institution": "Sharif University of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Arman Malekzadeh", "gender": "Male", "institution": "Sharif University of Technology", "country": "IR", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 46, "n_ref": 66, "n_ref_all": 81, "n_fig": 3, "n_tab": 9, "L_tab": 1842, "n_element_tab": 184, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 1842, "n_element_tab_1": 184, "formula_len_all": 742, "formula_len_all_1": 742, "len_all": 125264, "len_all_1": 60176, "len_abs": 1370, "len_title": 131, "len_sents": 29450, "len_sents_1": 25332, "n_sents": 199, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1376, "title": "Gradient Storm: Stronger Backdoor Attacks Through Expanded Parameter Space Coverage", "abs": "Targeted data poisoning poses a critical adversarial threat to machine learning systems by enabling attackers to manipulate training data to induce specific, harmful misclassifications. Among these threats, backdoor attacks are particularly pernicious, embedding hidden triggers in the data that lead models to misclassify only those inputs containing the trigger, while maintaining high accuracy on benign samples. In this paper, we propose Gradient Storm, a novel technique that facilitates the simultaneous execution of multiple backdoor attacks, while necessitating only minimal modification to the training dataset. Our contributions are twofold: First, we introduce a method for designing adversarial poisons in modular components, each tailored based on a distinct region of the model\u2019s parameter space. Second, we present a framework for conducting multi-trigger attacks, where each trigger causes misclassification from a specific source class to a distinct target class. We evaluate the efficacy of Gradient Storm across multiple neural network architectures and two benchmark datasets, demonstrating its robustness against eight different poisoning defense mechanisms. Additionally, we show that poisons crafted for one model can be effectively transferred to other models, demonstrating that our attack remains effective even in black-box settings.", "keywords": [ "Data Poisoning", "Adversarial Attacks", "Backdoor Attacks" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 3, 5, 3, 3 ] }, { "paper_id": "OE2T7AgQFN", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Jianming Zhang", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" }, { "name": "Nanxuan Zhao", "gender": "Female", "institution": "Adobe Research", "country": "US", "position": "Researcher" }, { "name": "Pinar Yanardag", "gender": "Female", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "Assistant Professor" }, { "name": "Qing Liu", "gender": "Female", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Yusuf Dalva", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "PhD student" }, { "name": "Zhe Lin", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Principal Researcher" }, { "name": "Yijun Li", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 19, "n_ref": 46, "n_ref_all": 81, "n_fig": 22, "n_tab": 3, "L_tab": 1034, "n_element_tab": 59, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 681, "formula_len_all_1": 150, "len_all": 117230, "len_all_1": 49949, "len_abs": 957, "len_title": 131, "len_sents": 43543, "len_sents_1": 25254, "n_sents": 288, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1171, "title": "LayerFusion: Harmonized Multi-Layer Text-to-Image Generation with Generative Priors", "abs": "Large-scale diffusion models have achieved remarkable success in generating high-quality images from textual descriptions, gaining popularity across various applications. However, the generation of layered content, such as transparent images with foreground and background layers, remains an under-explored area. Layered content generation is crucial for creative workflows in fields like graphic design, animation, and digital art, where layer-based approaches are fundamental for flexible editing and composition. In this paper, we propose a novel image generation pipeline based on Latent Diffusion Models (LDMs) that generates images with two layers: a foreground layer (RGBA) with transparency information and a background layer (RGB). Unlike existing methods that generate these layers sequentially, our approach introduces a harmonized generation mechanism that enables dynamic interactions between the layers for more coherent outputs. We demonstrate the effectiveness of our method through extensive qualitative and quantitative experiments, showing significant improvements in visual coherence, image quality, and layer consistency compared to baseline methods.", "keywords": [ "diffusion", "T2I", "generative models", "rgb", "rgba", "layer", "layer diffusion" ], "rating_list": [ 3, 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 2, 3, 2 ], "presentation_list": [ 1, 3, 2, 3, 3 ], "contribution_list": [ 1, 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 3, 4 ] }, { "paper_id": "ODzthXYPtp", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Liudeng Wang", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Samson Zhou", "gender": "unknown", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" } ], "n_formula": 157, "n_formula_1": 13, "n_ref_uni": 29, "n_ref": 56, "n_ref_all": 72, "n_fig": 0, "n_tab": 8, "L_tab": 1574, "n_element_tab": 222, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 287, "n_element_tab_1": 36, "formula_len_all": 25468, "formula_len_all_1": 2724, "len_all": 257502, "len_all_1": 67057, "len_abs": 1128, "len_title": 87, "len_sents": 76466, "len_sents_1": 28933, "n_sents": 810, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1341, "title": "Clustering on Skewed Cost Distributions", "abs": "In this paper, we tackle the problem of $(k,z)$-clustering, a generalization of the well-known $k$-means, $k$-medians and $k$-medoids problems that is known to be APX hard, i.e., impossible to approximate within a multiplicative factor of $1.06$ in polynomial time for $n$ and $k$ unless P=NP. Due to the APX-hardness, the fastest $(1+\\varepsilon)$-approximation scheme proposed by Feldman et al. (2007), exhibits a run time with a polynomial dependency on $n$, but an exponential dependency $2^{\\tilde{\\mathcal{O}}(k/\\varepsilon)}$ on $k$. We observe that a $(1+\\varepsilon)$-approximation in truly polynomial time is feasible if the data sets exhibit sufficiently skewed distributions. Indeed in practical scenarios, data sets often exhibit a heavy skewness, leading to the overall clustering cost disproportionately dominated by a few clusters. We propose a novel algorithm that adapts the traditional local search technique to effectively manage $(s, 1- \\varepsilon^{z+1})$-skewed datasets with a run time of $(nk/\\varepsilon)^{\\mathcal{O}(s+1/\\varepsilon)}$ for discrete case and $\\tilde{\\mathcal{O}}(nk) + (k \\log n)^{\\tilde{\\mathcal{O}}(s+1/\\varepsilon)}$ for continuous case. Our method is particularly effective with Zipfian distributions with exponent $p>1$, where $s = \\mathcal{O}\\left(\\frac{1}{\\varepsilon^{(z+1)/(p-1)}}\\right)$.", "keywords": [ "clustering", "PTAS" ], "rating_list": [ 5, 8, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "ODzT43I5lJ", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "ping li", "gender": "Male", "institution": "Rutgers University", "country": "", "position": "Associate Professor" }, { "name": "Xiaoyun Li", "gender": "Male", "institution": "LinkedIn", "country": "", "position": "Researcher" } ], "n_formula": 47, "n_formula_1": 23, "n_ref_uni": 63, "n_ref": 94, "n_ref_all": 107, "n_fig": 6, "n_tab": 1, "L_tab": 198, "n_element_tab": 5, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7186, "formula_len_all_1": 2711, "len_all": 158045, "len_all_1": 62067, "len_abs": 909, "len_title": 94, "len_sents": 36130, "len_sents_1": 26194, "n_sents": 325, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 913, "title": "Differentially Private One Permutation Hashing", "abs": "Minwise hashing (MinHash) is a standard hashing algorithm for large-scale search and learning with the binary Jaccard similarity. One permutation hashing (OPH) is an effective and efficient alternative of MinHash which splits the data into K bins and generates hash values within each bin. In this paper, to protect the privacy of the output sketches, we combine differential privacy (DP) with OPH, and propose DP-OPH framework with three variants: DP-OPH-fix, DP-OPH-re and DP-OPH-rand, depending on the densification strategy to deal with empty bins in OPH. Detailed algorithm design and privacy and utility analysis are provided. The proposed DP-OPH methods significantly improves the DP minwise hashing (DP-MH) alternative in the literature. Experiments on similarity search confirm the effectiveness of our proposed algorithms. We also provide an extension to real-value data, named DP-BCWS, in the appendix.", "keywords": [ "hash", "data compression", "privacy" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "ODiY6pbHZQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Jiwen Lu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Winston Hu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Yongming Rao", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Yuhao Dong", "gender": "Male", "institution": "Department of Automation, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Ziwei Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Zuyan Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 36, "n_ref": 123, "n_ref_all": 143, "n_fig": 11, "n_tab": 9, "L_tab": 4119, "n_element_tab": 633, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 50, "formula_len_all_1": 50, "len_all": 205897, "len_all_1": 53058, "len_abs": 1744, "len_title": 123, "len_sents": 56597, "len_sents_1": 27779, "n_sents": 421, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1453, "title": "Oryx MLLM: On-Demand Spatial-Temporal Understanding at Arbitrary Resolution", "abs": "Visual data comes in various forms, ranging from small icons of just a few pixels to long videos spanning hours. Existing multi-modal LLMs usually standardize these diverse visual inputs to fixed-resolution images or patches for visual encoders and yield similar numbers of tokens for LLMs. This approach is non-optimal for multimodal understanding and inefficient for processing inputs with long and short visual contents. To solve the problem, we propose Oryx, a unified multimodal architecture for the spatial-temporal understanding of images, videos, and multi-view 3D scenes. Oryx offers an on-demand solution to seamlessly and efficiently process visual inputs with arbitrary spatial sizes and temporal lengths through two core innovations: 1) a pre-trained OryxViT model that can encode images at any resolution into LLM-friendly visual representations; 2) a dynamic compressor module that supports 1x to 16x compression on visual tokens by request. These designs enable Oryx to accommodate extremely long visual contexts, such as videos, with lower resolution and high compression while maintaining high recognition precision for tasks like document understanding with native resolution and no compression. Beyond the architectural improvements, enhanced data curation and specialized training on long-context retrieval and spatial-aware data help Oryx achieve strong capabilities in image, video, and 3D multimodal understanding simultaneously.", "keywords": [ "Multi-modal Large Language Model", "Multi-Modal Understanding", "Arbitary Resolution" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "ODRHZrkOQM", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Jiayin Cai", "gender": "Female", "institution": "Kuaishou", "country": "CN", "position": "Researcher" }, { "name": "Ouxiang Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Shilin Yan", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Weidi Xie", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaolong Jiang", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Enigneer" }, { "name": "Yanbin Hao", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yao Hu", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 125, "n_ref_all": 147, "n_fig": 9, "n_tab": 10, "L_tab": 4899, "n_element_tab": 424, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2398, "n_element_tab_1": 245, "formula_len_all": 932, "formula_len_all_1": 637, "len_all": 169919, "len_all_1": 72097, "len_abs": 1615, "len_title": 96, "len_sents": 44523, "len_sents_1": 32375, "n_sents": 297, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1547, "title": "A Sanity Check for AI-generated Image Detection", "abs": "With the rapid development of generative models, discerning AI-generated content has evoked increasing attention from both industry and academia. In this paper, we conduct a sanity check on whether the task of AI-generated image detection has been solved. To start with, we present Chameleon dataset, consisting of AI-generated images that are genuinely challenging for human perception. To quantify the generalization of existing methods, we evaluate 9 off-the-shelf AI-generated image detectors on Chameleon dataset. Upon analysis, almost all models misclassify AI-generated images as real ones. Later, we propose AIDE AI-generated Image DEtector with Hybrid Features, which leverages multiple experts to simultaneously extract visual artifacts and noise patterns. Specifically, to capture the high-level semantics, we utilize CLIP to compute the visual embedding. This effectively enables the model to discern AI-generated images based on semantics and contextual information. Secondly, we select the highest and lowest frequency patches in the image, and compute the low-level patchwise features, aiming to detect AI-generated images by low-level artifacts, for example, noise patterns, anti-aliasing effects. While evaluating on existing benchmarks, for example, AIGCDetectBenchmark and GenImage, AIDE achieves +3.5% and +4.6% improvements to state-of-the-art methods, and on our proposed challenging Chameleon benchmarks, it also achieves promising results, despite the problem of detecting AI-generated images remains far from being solved.", "keywords": [ "AI-generated image detection", "Chameleon dataset", "mixture-of-experts" ], "rating_list": [ 6, 6, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "OD9pwKQzXl", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Hao Tang", "gender": "unknown", "institution": "City University of New York, City University of New York", "country": "US", "position": "Full Professor" }, { "name": "Jianing Qi", "gender": "unknown", "institution": "City University of New York, City University of New York", "country": "US", "position": "PhD student" }, { "name": "Zhigang Zhu", "gender": "Male", "institution": "CUNY City College of NY", "country": "US", "position": "Full Professor" } ], "n_formula": 36, "n_formula_1": 15, "n_ref_uni": 16, "n_ref": 65, "n_ref_all": 88, "n_fig": 7, "n_tab": 1, "L_tab": 30, "n_element_tab": 2, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 288, "n_element_tab_1": 44, "formula_len_all": 3108, "formula_len_all_1": 903, "len_all": 142518, "len_all_1": 67870, "len_abs": 1189, "len_title": 122, "len_sents": 52371, "len_sents_1": 31336, "n_sents": 484, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1195, "title": "VerifierQ: Enhancing LLM Test Time Compute with Q-Learning-based Verifiers", "abs": "Recent test time compute approaches with verifier models have significantly enhanced the reasoning capabilities of Large Language Models (LLMs). While this kind of generator-verifier approach closely resembles the actor-critic framework in reinforcement learning (RL), the verifiers currently are used rely on supervised fine-tuning rather than on temporal difference learning. This paper introduces VerifierQ, a novel approach that integrates Offline Q-learning into LLM verifier models. We address three key challenges in applying Q-learning to LLMs: utterance-level Markov Decision Processes (MDPs), large action spaces, and overestimation bias. VerifierQ introduces a modified Bellman update, incorporates Implicit Q-learning (IQL) for efficient action space management, and integrates a novel Conservative Q-learning (CQL) formulation for balanced overestimation. Our method is among the first to apply Q-learning to LLM verifiers. This integration of RL principles into verifier models complements existing advancements in generator techniques. Experimental results on mathematical reasoning tasks demonstrate VerifierQ's superior performance compared to supervised fine-tuning approaches.", "keywords": [ "LLM", "test time compute", "Reinforcement Learning", "Q-Learning", "verifier" ], "rating_list": [ 3, 5, 8, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "OD1MV7vf41", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Azhir Mahmood", "gender": "Not Specified", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Michel Tsamados", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Associate Professor" }, { "name": "So Takao", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Weibin Chen", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 21, "n_ref_uni": 33, "n_ref": 52, "n_ref_all": 70, "n_fig": 11, "n_tab": 3, "L_tab": 687, "n_element_tab": 61, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 674, "n_element_tab_1": 58, "formula_len_all": 2005, "formula_len_all_1": 1906, "len_all": 113603, "len_all_1": 64342, "len_abs": 1267, "len_title": 118, "len_sents": 32187, "len_sents_1": 27158, "n_sents": 229, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1273, "title": "Deep Random Features for Scalable Interpolation of Spatiotemporal Data", "abs": "The rapid growth of earth observation systems calls for a scalable approach to interpolate remote-sensing observations. These methods in principle, should acquire more information about the observed field as data grows. Gaussian processes (GPs) are candidate model choices for interpolation. However, due to their poor scalability, they usually rely on inducing points for inference, which restricts their expressivity. Moreover, commonly imposed assumptions such as stationarity prevents them from capturing complex patterns in the data. While deep GPs can overcome this issue, training and making inference with them are difficult, again requiring crude approximations via inducing points. In this work, we instead approach the problem through Bayesian deep learning, where spatiotemporal fields are represented by deep neural networks, whose layers share the inductive bias of stationary GPs on the plane/sphere via random feature expansions. This allows one to (1) capture high frequency patterns in the data, and (2) use mini-batched gradient descent for large scale training. We experiment on various remote sensing data at local/global scales, showing that our approach produce competitive or superior results to existing methods, with well-calibrated uncertainties.", "keywords": [ "Random Features", "Deep Gaussian Processes", "Bayesian Deep Learning", "Remote Sensing" ], "rating_list": [ 8, 3, 8 ], "soundness_list": [ 4, 2, 3 ], "presentation_list": [ 4, 2, 4 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "OCpxDSn0G4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Gunhee Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Junhyeog Yun", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Seungyoon Woo", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "MS student" } ], "n_formula": 27, "n_formula_1": 10, "n_ref_uni": 58, "n_ref": 123, "n_ref_all": 134, "n_fig": 4, "n_tab": 7, "L_tab": 3557, "n_element_tab": 439, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1244, "n_element_tab_1": 169, "formula_len_all": 1483, "formula_len_all_1": 562, "len_all": 183966, "len_all_1": 63082, "len_abs": 1173, "len_title": 88, "len_sents": 47238, "len_sents_1": 26927, "n_sents": 417, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1177, "title": "Meta-Continual Learning of Neural Fields", "abs": "Neural Fields (NF) have gained prominence as a versatile framework for complex data representation. This work unveils a new problem setting termed Meta-Continual Learning of Neural Fields (MCL-NF) and introduces a novel strategy that employs a modular architecture combined with optimization-based meta-learning. Focused on overcoming the limitations of existing methods for continual learning of neural fields, such as catastrophic forgetting and slow convergence, our strategy achieves high-quality reconstruction with significantly improved learning speed. We further introduce Fisher Information Maximization loss for neural radiance fields (FIM-NeRF), which maximizes information gains at the sample level to enhance learning generalization, with proved convergence guarantee and generalization bound. We perform extensive evaluations across image, audio, video reconstruction, and view synthesis tasks on six diverse datasets, demonstrating our method\u2019s superiority in reconstruction quality and speed over existing MCL and CL-NF approaches. Notably, our approach attains rapid adaptation of neural fields for city-scale NeRF rendering with reduced parameter requirement.", "keywords": [ "Meta-Learning; Continual Learning; 3D Vision; Neural Radiance Fields;" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "OCd3cffulp", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Bin Dai", "gender": "Male", "institution": "XiaoIce", "country": "", "position": "Researcher" }, { "name": "Jingsheng Gao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Ke Ji", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Linxu Li", "gender": "unknown", "institution": "Xiaobing", "country": "CN", "position": "Researcher" }, { "name": "Weiyuan Li", "gender": "Male", "institution": "xiaobing.ai", "country": "CN", "position": "Researcher" }, { "name": "Yixin Lian", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "yuzhuo fu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 10, "n_ref_uni": 45, "n_ref": 78, "n_ref_all": 100, "n_fig": 3, "n_tab": 10, "L_tab": 1961, "n_element_tab": 164, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2208, "n_element_tab_1": 161, "formula_len_all": 344, "formula_len_all_1": 755, "len_all": 182690, "len_all_1": 67288, "len_abs": 1087, "len_title": 119, "len_sents": 56932, "len_sents_1": 29039, "n_sents": 453, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1088, "title": "SmartRAG: Jointly Learn RAG-Related Tasks From the Environment Feedback", "abs": "RAG systems consist of multiple modules to work together. However, these modules are usually separately trained. We argue that a system like RAG that incorporates multiple modules should be jointly optimized to achieve optimal performance. To demonstrate this, we design a specific pipeline called SmartRAG that includes a policy network and a retriever. The policy network can serve as 1) a decision maker that decides when to retrieve, 2) a query rewriter to generate a query most suited to the retriever and 3) an answer generator that produces the final response with/without the observations. We then propose to jointly optimize the whole system using a reinforcement learning algorithm, with the reward designed to encourage the system to achieve the highest performance with minimal retrieval cost. When jointly optimized, each module can be aware of how other modules are working and thus find the best way to work together as a complete system. Empirical results demonstrate that the jointly optimized system can achieve better performance than separately optimized counterparts.", "keywords": [ "Retrieval-augmented Generation", "Language Models", "Reinforcement Learning" ], "rating_list": [ 6, 5, 8, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "OCHSgafZ1Y", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Cuong Pham", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Hoang Anh Dung", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Jianfei Cai", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" }, { "name": "Trung Le", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" }, { "name": "Thanh-Toan Do", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" } ], "n_formula": 38, "n_formula_1": 33, "n_ref_uni": 27, "n_ref": 75, "n_ref_all": 90, "n_fig": 0, "n_tab": 3, "L_tab": 584, "n_element_tab": 10, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3577, "formula_len_all_1": 2289, "len_all": 105856, "len_all_1": 67113, "len_abs": 1318, "len_title": 147, "len_sents": 29093, "len_sents_1": 26118, "n_sents": 212, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1326, "title": "Zero-shot Mixed Precision Quantization via Joint Optimization of Data Generation and Bit Allocation", "abs": "Mixed-precision quantization (MPQ) aims to identify optimal bit-widths for layers to quantize a model.\nOn the other hand,\nzero-shot quantization (ZSQ) aims to learn a quantized model from a pre-trained full-precision model in a data-free manner, which is commonly done by generating a synthetic calibration set used for quantizing the full-precision model. While it is intuitive that there exists inherent correlation between the quality of the generated calibration dataset\nand the bit allocation to the model's layers, \nall existing frameworks treat them as separate problems. This paper proposes a novel method that jointly optimizes both the calibration set and the bit-width of each layer in the context of zero-shot quantization. Specifically, we first propose a novel data optimization approach that takes into consideration the Gram-Gradient matrix constructed from the gradient vectors of calibration samples. We then propose a novel scalable quadratic optimization-based approach to identify the model's bit-widths. These proposals will then be combined into a single framework to jointly optimize both the calibration data and the bit allocation to the model's layers.\nExperimental results on the ImageNet dataset demonstrate the proposed method's superiority compared to current state-of-the-art techniques in ZSQ.", "keywords": [ "Zero-shot Quantization", "Post-training quantization", "mixed-precision quantization" ], "rating_list": [ 8, 5, 6 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 4 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "OBrTQcX2Hm", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Qiang Ni", "gender": "Male", "institution": "Lancaster University", "country": "GB", "position": "Full Professor" }, { "name": "Yongchen Zhou", "gender": "Male", "institution": "Lancaster University", "country": "GB", "position": "PhD student" }, { "name": "Richard M. Jiang", "gender": "Male", "institution": "Lancaster University", "country": "GB", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 20, "n_ref_uni": 20, "n_ref": 21, "n_ref_all": 27, "n_fig": 1, "n_tab": 2, "L_tab": 640, "n_element_tab": 131, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 640, "n_element_tab_1": 131, "formula_len_all": 1198, "formula_len_all_1": 1198, "len_all": 53894, "len_all_1": 53894, "len_abs": 653, "len_title": 124, "len_sents": 15894, "len_sents_1": 15894, "n_sents": 119, "n_sents_1": 119, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 654, "title": "KARA: Enhancing High-Dimensional Data Processing with Learnable Activations", "abs": "In the rapidly advancing field of machine learning, efficiently processing and interpreting high-dimensional data remains a significant challenge. This paper presents the Kolmogorov-Arnold Representation Autoencoder (KARA), a novel autoencoder architecture designed to leverage the Kolmogorov-Arnold representation theorem. By incorporating this mathematical foundation, KARA enhances the representational power and efficiency of neural networks, enabling superior performance in data compression tasks. Experimental results demonstrate that KARA achieves superior performance, positioning it as a promising approach for high-dimensional data processing.", "keywords": [ "Kolmogorov-Arnold representation theorem", "high-dimensional data processing", "learnable activation" ], "rating_list": [ 3, 1, 1, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 1, 2, 1, 1 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "OBjF5I4PWg", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Yang Liu", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "Associate Professor" }, { "name": "Zhen Wang", "gender": "Male", "institution": "Northwestern Polytechnical University", "country": "CN", "position": "Full Professor" }, { "name": "quanming yao", "gender": "Male", "institution": "Department of Electronic Engineering, Tsinghua University", "country": "", "position": "Assistant Professor" }, { "name": "An Bian", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Senior researcher " }, { "name": "HY Nie", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 34, "n_ref": 64, "n_ref_all": 85, "n_fig": 10, "n_tab": 8, "L_tab": 1137, "n_element_tab": 101, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 862, "n_element_tab_1": 78, "formula_len_all": 264, "formula_len_all_1": 300, "len_all": 148449, "len_all_1": 60733, "len_abs": 1657, "len_title": 111, "len_sents": 51921, "len_sents_1": 29985, "n_sents": 347, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1557, "title": "Erasing Concept Combination from Text-to-Image Diffusion Model", "abs": "Advancements in the text-to-image diffusion model have raised security concerns due to their potential to generate images with inappropriate themes such as societal biases and copyright infringements. Current studies make a great process to prevent the model from generating images containing specific high-risk visual concepts. However, these methods neglect the issue that inappropriate themes may also arise from the combination of benign visual concepts. Considering that the same image theme might be represented via multiple different visual concept combinations, and the model's generation performance of the corresponding individual visual concepts is distorted easily while processing the visual concept combination, effectively erasing such visual concept combinations from the diffusion model remains a formidable challenge. To this end, we formulate such challenge as the Concept Combination Erasing (CCE) problem and propose a Concept Graph-based high-level Feature Decoupling framework (CoGFD) to address CCE. CoGFD identifies and decomposes visual concept combinations with a consistent image theme from an LLM-induced concept logic graph, and erases these combinations through decoupling oc-occurrent high-level features. These techniques enable CoGFD to erase visual concept combinations of image content while enjoying a much less negative effect, compared to SOTA baselines, on the generative fidelity of related individual concepts. Extensive experiments on diverse visual concept combination scenarios verify the effectiveness of CoGFD.", "keywords": [ "concept combination erasing", "text-to-image diffusion model", "AIGC security", "concept logic graph generation", "feature decouple" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 4, 3, 1, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "OBUQNASaWw", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Adriano Guastella", "gender": "Male", "institution": "University of Bologna", "country": "", "position": "MS student" }, { "name": "Alessio Mora", "gender": "Male", "institution": "University of Bologna", "country": "IT", "position": "Postdoc" }, { "name": "Lorenzo Sani", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Paolo Bellavista", "gender": "Male", "institution": "University of Bologna", "country": "IT", "position": "Full Professor" }, { "name": "Alexandru Andrei Iacob", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Nic Lane", "gender": "Male", "institution": "Flower Labs", "country": "GB", "position": "Chief Scientific Officer" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 51, "n_ref": 140, "n_ref_all": 155, "n_fig": 4, "n_tab": 13, "L_tab": 5617, "n_element_tab": 576, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1907, "n_element_tab_1": 125, "formula_len_all": 294, "formula_len_all_1": 291, "len_all": 222338, "len_all_1": 62253, "len_abs": 1433, "len_title": 45, "len_sents": 65710, "len_sents_1": 30469, "n_sents": 437, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 45, "L_abs": 1439, "title": "SparsyFed: Sparse Adaptive Federated Learning", "abs": "Sparse training is often adopted in cross-device federated learning (FL) environments where constrained devices collaboratively train a machine learning model on private data by exchanging pseudo-gradients across heterogeneous networks. Although sparse training methods can reduce communication overhead and computational burden in FL, they are often not used in practice for the following key reasons: (1) data heterogeneity impacts more clients\u2019 consensus on sparse, compared to dense, models, requiring training for longer; (2) a lack of sufficient plasticity to adapt to never-seen data distributions, crucial in cross-device FL; (3) requiring additional hyperparameters, which are notably challenging to tune in FL. This paper presents SparsyFed, a practical federated sparse training method that critically addresses all the aforementioned problems. Previous works have only managed to solve one, or perhaps two of these challenges, and at the expense of introducing new trade-offs, such as clients\u2019 consensus on masks versus sparsity pattern plasticity. We show that SparsyFed simultaneously (1) can produce 95% sparse models, with negligible degradation in accuracy, while only needing a single hyperparameter, (2) achieves a per-round weight regrowth 200 times smaller than previous methods, and (3) still offers plasticity under this sparse design, by outperforming all the baselines at adapting to never-seen data distributions.", "keywords": [ "dynamic sparse training", "federated learning", "cross device federated learning" ], "rating_list": [ 3, 5, 6, 8, 5 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 3, 5, 4, 2, 5 ] }, { "paper_id": "OBTmkKBmQW", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Chi Wang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Hong Li", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Weiwei Xu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yikai Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Ying Song", "gender": "unknown", "institution": "Zhejiang Sci-Tech University", "country": "CN", "position": "Associate Professor" }, { "name": "guojunlei", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 30, "n_ref": 71, "n_ref_all": 93, "n_fig": 7, "n_tab": 4, "L_tab": 1445, "n_element_tab": 211, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1097, "n_element_tab_1": 135, "formula_len_all": 263, "formula_len_all_1": 263, "len_all": 111303, "len_all_1": 55231, "len_abs": 179, "len_title": 147, "len_sents": 30818, "len_sents_1": 26122, "n_sents": 197, "n_sents_1": 165, "venue": "ICLR2025", "decision": "withdraw", "L_title": 98, "L_abs": 965, "title": "MOTIONFLOW:Learning Implicit Motion Flow for Complex Camera Trajectory Control in Video Generation", "abs": "Generating videos guided by camera trajectories poses significant challenges in achieving consistency and generalizability, particularly when both camera and object motions are present. Existing approaches often attempt to learn these motions separately, which may lead to confusion regarding the relative motion between the camera and the objects. To address this challenge, we propose a novel approach that integrates both camera and object motions by converting them into the motion of corresponding pixels. Utilizing a stable diffusion network, we effectively learn reference motion maps in relation to the specified camera trajectory. These maps, along with an extracted semantic object prior, are then fed into an image-to-video network to generate the desired video that can accurately follow the designated camera trajectory while maintaining consistent object motions. Extensive experiments verify that our model outperforms SOTA methods by a large margin.", "keywords": [ "Camera trajectory", "video generation", "diffusion model" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "OBIuFjZzmp", "primary_area": "optimization", "n_author": 8, "author_info_list": [ { "name": "Eunho Yang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Jihun Yun", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "June Yong Yang", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sihwan Park", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Souvik Kundu", "gender": "Male", "institution": "Intel", "country": "US", "position": "Researcher" }, { "name": "Yeonsung Jung", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Kyung-Su Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Assistant Professor" }, { "name": "SungYub Kim", "gender": "Male", "institution": "Samsung", "country": "KR", "position": "Researcher" } ], "n_formula": 169, "n_formula_1": 15, "n_ref_uni": 35, "n_ref": 102, "n_ref_all": 119, "n_fig": 3, "n_tab": 10, "L_tab": 3799, "n_element_tab": 547, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 19538, "formula_len_all_1": 1539, "len_all": 228109, "len_all_1": 68338, "len_abs": 1811, "len_title": 145, "len_sents": 55203, "len_sents_1": 28830, "n_sents": 604, "n_sents_1": 231, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1876, "title": "MeZO-A$^{3}$dam: Memory-efficient Zeroth-order Adam with Adaptivity Adjustments for Fine-tuning LLMs", "abs": "Recently, fine-tuning of language models (LMs) via zeroth-order (ZO) optimization have gained significant traction due to their ability of memory-efficient deployment, significantly reducing memory cost over first-order methods. However, the existing studies on ZO optimization for LM fine-tuning often exhibit slow convergence and the reliance on the hand-crafted prompts. Towards mitigating these limitations, in this paper, we first investigate on the importance of adaptive gradient based ZO optimization method. Toward this, we revisit memory-efficient zeroth-order Adam (MeZO-Adam) and make important findings that merely considering adaptivity can enable faster convergence while improving the generalization ability compared to previous studies. Interestingly, we further observe that decreasing the level of adaptivity might be recommended in ZO optimization potentially due to the high variance of ZO gradient estimate, hypothesized as \\emph{weak adaptivity hypothesis}. Based upon our hypothesis, we propose MeZO-A$^3$dam, MeZO-Adam with Adaptivity Adjustments according to the parameter dimension. We provide the dimension-free theoretical guarantee on both the convergence and the generalization of MeZO-A$^3$dam, providing strong evidence for our hypothesis. Extensive experiments show that MeZO-A$^3$dam can achieve faster convergence and better generalization over several baselines across LMs of various sizes on diverse datasets. By adaptivity adjustments, MeZO-A$^3$dam outperforms MeZO, MeZO-SVRG, and MeZO-Adam, with up to an average of $36.6\\\\%$, $16.9\\\\%$, $6.8\\\\%$ improvements in performance and up to an average of $\\times 12.6$ and $\\times1.8$ faster convergence, respectively. Furthermore, by leveraging an off-the-shelf low-bit optimizer, MeZO-A$^3$dam achieves an average of $40.3\\\\%$ and $43.6\\\\%$ memory reduction from MeZO-SVRG and MeZO-Adam.", "keywords": [ "Optimization", "Zeroth-Order Optimization", "Large Language Models", "Fine-tuning" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "OAlPlR1g9B", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Dongsheng Luo", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "Assistant Professor" }, { "name": "Farhad Shirani", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "Assistant Professor" }, { "name": "Shouwei Gao", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "PhD student" }, { "name": "Tianchun Wang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Wei Cheng", "gender": "Male", "institution": "NEC-Labs", "country": "US", "position": "Researcher" }, { "name": "Wenqian Dong", "gender": "Female", "institution": "Florida International University", "country": "", "position": "Assistant Professor" }, { "name": "Xu Zheng", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "PhD student" } ], "n_formula": 74, "n_formula_1": 26, "n_ref_uni": 59, "n_ref": 119, "n_ref_all": 135, "n_fig": 5, "n_tab": 9, "L_tab": 5922, "n_element_tab": 245, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 2245, "n_element_tab_1": 46, "formula_len_all": 9235, "formula_len_all_1": 1811, "len_all": 202351, "len_all_1": 70823, "len_abs": 1995, "len_title": 105, "len_sents": 51484, "len_sents_1": 31164, "n_sents": 540, "n_sents_1": 234, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 2010, "title": "Explanation-Assisted Data Augmentation for Graph Learning", "abs": "This work introduces a novel class of Data Augmentation (DA) techniques in the context of graph learning. In general, DA refers to techniques that enlarge the training set using label-preserving transformations. Such techniques enable increased robustness and generalization, especially when the size of the original training set is limited. A fundamental idea in DA is that \nlabels are invariant to domain-specific transformations of the input samples. \nHowever, it is challenging to identify such transformations in learning over graphical input domains due to the complex nature of graphs and the need to preserve their structural and semantic properties.\nIn this work, we propose explanation-assisted DA (EA-DA) for Graph Neural Networks (GNNs). A graph explanation is a subgraph which is an `almost sufficient' statistic of the input graph with respect to its classification label. Consequently, the classification label is invariant, with high probability, to perturbations of graph edges not belonging to its explanation subgraph. We develop EA-DA techniques leveraging such perturbation invariances. First, we show analytically that the sample complexity of explanation-assisted learning can be arbitrarily smaller than explanation-agnostic learning. On the other hand, we show that if the training set is enlarged using EA-DA techniques and\nthe learning rule does not distinguish between the augmented data and the original data, then the sample complexity can be worse than that of explanation-agnostic learning. We identify the main reason for the potential increase in sample complexity as the out-of-distribution nature of graph perturbations. \nWe conclude that theoretically EA-DA may improve sample complexity, and that the learning rule must distinguish between the augmented data and the original data. Subsequently, we build upon these theoretical insights, introduce practically implementable EA-DA techniques and associated learning mechanisms, and perform extensive empirical evaluations.", "keywords": [ "empirical risk minimization", "Explainable Graph Neural Networks" ], "rating_list": [ 6, 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 3 ], "confidence_list": [ 2, 2, 2, 4, 3 ] }, { "paper_id": "OATPSB5JK1", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Kwanyoung Park", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Intern" }, { "name": "Youngwoon Lee", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 36, "n_formula_1": 24, "n_ref_uni": 38, "n_ref": 106, "n_ref_all": 159, "n_fig": 8, "n_tab": 21, "L_tab": 9774, "n_element_tab": 1117, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 4498, "n_element_tab_1": 596, "formula_len_all": 5902, "formula_len_all_1": 2251, "len_all": 179684, "len_all_1": 69737, "len_abs": 1313, "len_title": 122, "len_sents": 46542, "len_sents_1": 23298, "n_sents": 389, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1163, "title": "Model-based Offline Reinforcement Learning with Lower Expectile Q-Learning", "abs": "Model-based offline reinforcement learning (RL) is a compelling approach that addresses the challenge of learning from limited, static data by generating imaginary trajectories using learned models. However, these approaches often struggle with inaccurate value estimation from model rollouts. In this paper, we introduce a novel model-based offline RL method, Lower Expectile Q-learning (LEQ), which provides a low-bias model-based value estimation via lower expectile regression of $\\lambda$-returns. Our empirical results show that LEQ significantly outperforms previous model-based offline RL methods on long-horizon tasks, such as the D4RL AntMaze tasks, matching or surpassing the performance of model-free approaches and sequence modeling approaches. Furthermore, LEQ matches the performance of state-of-the-art model-based and model-free methods in dense-reward environments across both state-based tasks (NeoRL and D4RL) and pixel-based tasks (V-D4RL), showing that LEQ works robustly across diverse domains. Our ablation studies demonstrate that lower expectile regression, $\\lambda$-returns, and critic training on offline data are all crucial for LEQ.", "keywords": [ "model-based offline reinforcement learning", "offline reinforcement learning", "lower expectile return" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "OANUpvmnuf", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Boyuan Wang", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "MS student" }, { "name": "Chang Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Jianzhun Shao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Xiangyang Ji", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yixiu Mao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yuhang Jiang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Yun Qu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Qi Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" } ], "n_formula": 14, "n_formula_1": 5, "n_ref_uni": 65, "n_ref": 154, "n_ref_all": 198, "n_fig": 24, "n_tab": 8, "L_tab": 2825, "n_element_tab": 346, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 219, "n_element_tab_1": 32, "formula_len_all": 1456, "formula_len_all_1": 380, "len_all": 243671, "len_all_1": 69058, "len_abs": 1283, "len_title": 85, "len_sents": 71355, "len_sents_1": 33179, "n_sents": 538, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1292, "title": "Choices are More Important than Efforts: LLM Enables Efficient Multi-Agent Exploration", "abs": "With expansive state-action spaces, efficient multi-agent exploration remains a longstanding challenge in reinforcement learning.\nAlthough pursuing novelty, diversity, or uncertainty attracts increasing attention, redundant efforts brought by exploration without proper guidance choices poses a practical issue for the community.\nThis paper introduces a systematic approach, termed LEMAE, choosing to channel informative task-relevant guidance from a knowledgeable Large Language Model (LLM) for Efficient Multi-Agent Exploration. \nSpecifically, we ground linguistic knowledge from LLM into symbolic key states, that are critical for task fulfillment, in a discriminative manner at low LLM inference costs. \nTo unleash the power of key states, \nwe design Subspace-based Hindsight Intrinsic Reward (SHIR) to guide agents toward key states by increasing reward density. Additionally, we build the Key State Memory Tree (KSMT) to track transitions between key states in a specific task for organized exploration. Benefiting from diminishing redundant explorations, LEMAE outperforms existing SOTA approaches on the challenging benchmarks (e.g., SMAC and MPE) by a large margin, achieving a 10x acceleration in certain scenarios.\nOur code is available at https://anonymous.4open.science/r/LEMAE.", "keywords": [ "Reinforcement Learning; Large Language Model; Efficient Exploration" ], "rating_list": [ 1, 6, 5, 6 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "OALIb8oNfl", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Chongjie Si", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Jifeng Dai", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Qingyun Li", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Wei Shen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaokang Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University, China", "country": "CN", "position": "Professor" }, { "name": "Xue Yang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xuehui Wang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhengqin Xu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Postdoc" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 19, "n_ref": 46, "n_ref_all": 63, "n_fig": 5, "n_tab": 15, "L_tab": 6497, "n_element_tab": 489, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2356, "n_element_tab_1": 222, "formula_len_all": 376, "formula_len_all_1": 420, "len_all": 188159, "len_all_1": 61832, "len_abs": 2009, "len_title": 137, "len_sents": 39198, "len_sents_1": 26910, "n_sents": 285, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1731, "title": "Maintaining Structural Integrity in Parameter Spaces for Parameter Efficient Fine-tuning", "abs": "Adapting pre-trained foundation models for various downstream tasks has been prevalent in artificial intelligence. Due to the vast number of tasks and high costs, adjusting all parameters becomes unfeasible. To mitigate this, several fine-tuning techniques have been developed to update the pre-trained model weights in a more resource-efficient manner, such as through low-rank adjustments. Yet, almost all of these methods focus on linear weights, neglecting the intricacies of parameter spaces in higher dimensions like 4D. \nAlternatively, some methods can be adapted for high-dimensional parameter space by compressing changes in the original space into two dimensions and then employing low-rank matrix adaptations. However, these approaches destructs the structural integrity of the involved high-dimensional spaces. To tackle the diversity of dimensional spaces across different foundation models and provide a more precise representation of the changes within these spaces, this paper introduces a generalized parameter-efficient fine-tuning framework, designed for various dimensional parameter space. Specifically, our method asserts that changes in each dimensional parameter space are based on a low-rank core space which maintains the consistent topological structure with the original space. It then models the changes through this core space alongside corresponding weights to reconstruct alterations in the original space. It effectively preserves the structural integrity of the change of original N-dimensional parameter space, meanwhile models it via low-rank tensor adaptation. Extensive experiments on computer vision, natural language processing and multi-modal tasks validate the effectiveness of our method.", "keywords": [ "parameter efficient fine-tuning" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "O9YTt26r2P", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Aaron Mueller", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Postdoc" }, { "name": "Anja Reusch", "gender": "Female", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "Postdoc" }, { "name": "Yaniv Nikankin", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "PhD student" }, { "name": "Yonatan Belinkov", "gender": "Male", "institution": "Technion, Technion", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 3, "n_ref_uni": 36, "n_ref": 90, "n_ref_all": 149, "n_fig": 38, "n_tab": 3, "L_tab": 312, "n_element_tab": 41, "n_fig_1": 19, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 982, "formula_len_all_1": 198, "len_all": 173354, "len_all_1": 68689, "len_abs": 1319, "len_title": 130, "len_sents": 62646, "len_sents_1": 33165, "n_sents": 463, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1329, "title": "Arithmetic Without Algorithms: Language Models Solve Math with a Bag of Heuristics", "abs": "Do large language models (LLMs) solve reasoning tasks by learning robust generalizable algorithms, or do they memorize training data? To investigate this question, we use arithmetic reasoning as a representative task. Using causal analysis, we identify a subset of the model (a circuit) that explains most of the model's behavior for basic arithmetic logic and examine its functionality. By zooming in on the level of individual circuit neurons, we discover a sparse set of important neurons that implement simple heuristics. Each heuristic identifies a numerical input pattern and outputs corresponding answers. We hypothesize that the combination of these heuristic neurons is the mechanism used to produce correct arithmetic answers. To test this, we categorize each neuron into several heuristic types---such as neurons that activate when an operand falls within a certain range---and find that the unordered combination of these heuristic types is the mechanism that explains most of the model's accuracy on arithmetic prompts. Finally, we demonstrate that this mechanism appears as the main source of arithmetic accuracy early in training. Overall, our experimental results across several LLMs show that LLMs perform arithmetic using neither robust algorithms nor memorization; rather, they rely on a ``bag of heuristics''.", "keywords": [ "mechanistic interpretability", "interpretability", "arithmetic", "nlp", "language models" ], "rating_list": [ 6, 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 4, 3, 4 ], "presentation_list": [ 2, 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 2, 4 ] }, { "paper_id": "O9XdvMbnXC", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 9, "author_info_list": [ { "name": "Charles Guille-Escuret", "gender": "Male", "institution": "Mila", "country": "CA", "position": "PhD student" }, { "name": "Christopher J Pal", "gender": "unknown", "institution": "Polytechnique Montreal", "country": "CA", "position": "Full Professor" }, { "name": "Joel Lamy-Poirier", "gender": "Male", "institution": "ServiceNow Research", "country": "", "position": "Researcher" }, { "name": "Luke Kumar", "gender": "unknown", "institution": "ServiceNow Inc", "country": "CA", "position": "Researcher" }, { "name": "Max Tian", "gender": "Male", "institution": "ServiceNow Inc", "country": "CA", "position": "Researcher" }, { "name": "Raymond Li", "gender": "unknown", "institution": "ServiceNow", "country": "", "position": "Researcher" }, { "name": "Sebastien Paquet", "gender": "unknown", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "PhD student" }, { "name": "Torsten Scholak", "gender": "Male", "institution": "ServiceNow Research", "country": "CA", "position": "Researcher" }, { "name": "Denis Kochetkov", "gender": "Male", "institution": "ServiceNow Research", "country": "", "position": "Software engineer" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 34, "n_ref": 44, "n_ref_all": 61, "n_fig": 7, "n_tab": 3, "L_tab": 675, "n_element_tab": 181, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 675, "n_element_tab_1": 181, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 84925, "len_all_1": 75033, "len_abs": 783, "len_title": 136, "len_sents": 19930, "len_sents_1": 19933, "n_sents": 175, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 782, "title": "Mitigating Forgetting in Continually Pretraining MoE-LLMs by Adding and Chilling Experts", "abs": "As model training requires more and more compute, the cost of re-training models to support new data or domains increases as well. Methods to adapt existing models to new data distributions are crucial to avoid spending redundant compute re-training models from scratch. However, naive finetuning often incurs forgetting of previously learned capabilities. In this paper, we analyse how different factors such as model size, dataset size and replay data impact forgetting when adapting models to new data distributions. We also propose to increase the capacity of Mixture-of-experts models by adding new experts and reducing the learning rate of the old model weights. Our experiments show that this simple method allows to reduce forgetting and learn efficiently on the new domain.", "keywords": [ "LLM", "Continual Learning", "Lifelong Learning", "Mixture of Experts", "Pretraining" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "O9TTAoySaG", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Arash Behboodi", "gender": "Male", "institution": "QualComm", "country": "", "position": "Machine Learning Researcher" }, { "name": "Fabio Valerio Massoli", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "", "position": "Researcher" }, { "name": "Thomas Hehn", "gender": "unknown", "institution": "Qualcomm Inc, QualComm", "country": "NL", "position": "Researcher" }, { "name": "Tim Bakker", "gender": "Not Specified", "institution": "Qualcomm Inc, QualComm", "country": "NL", "position": "Researcher" }, { "name": "Tribhuvanesh Orekondy", "gender": "unknown", "institution": "Qualcomm Inc, QualComm", "country": "", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 55, "n_ref": 130, "n_ref_all": 155, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 757, "formula_len_all_1": 401, "len_all": 165925, "len_all_1": 65906, "len_abs": 1336, "len_title": 119, "len_sents": 52605, "len_sents_1": 32566, "n_sents": 382, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1347, "title": "Simulating, Fast and Slow: Learning Policies for Black-Box Optimization", "abs": "Simulators are vital in science and engineering, as they faithfully model the influence of design parameters on real-world observations. A common problem is leveraging the simulator to optimize the design parameters to minimize a desired objective function. Since simulators are often non-differentiable blackboxes and each simulation incurs significant compute time, gradient-based optimization techniques can often be intractable or, in some cases, impossible. Furthermore, in many experiment design settings, practitioners are required to solve sets of closely related optimization problems. Thus, starting the optimization from scratch each time might be inefficient if the forward simulation model is expensive to evaluate. To address these challenges, this paper introduces a novel method for solving classes of similar black-box optimization problems by learning an active learning policy that guides the training of a differentiable surrogate and then uses that surrogate's gradients to optimize the simulation parameters with gradient descent. After training the policy, the cost for downstream optimization of problems involving black-box simulators is amortized and we require up to $\\sim$90\\% fewer expensive simulator calls compared to baselines such as local surrogate-based approaches, numerical optimization, and Bayesian methods.", "keywords": [ "RL", "black-box optimization", "learning surrogates", "AI for science" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "O9NQLOjrdu", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Abdelrahman Abdelhamed", "gender": "Male", "institution": "Google", "country": "CA", "position": "Researcher" }, { "name": "Alec Go", "gender": "unknown", "institution": "Google", "country": "US", "position": "Software Engineer" }, { "name": "Mahmoud Afifi", "gender": "Male", "institution": "Samsung", "country": "CA", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 29, "n_ref": 101, "n_ref_all": 145, "n_fig": 8, "n_tab": 4, "L_tab": 161, "n_element_tab": 18, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 157, "n_element_tab_1": 17, "formula_len_all": 242, "formula_len_all_1": 242, "len_all": 112513, "len_all_1": 59035, "len_abs": 1238, "len_title": 143, "len_sents": 36772, "len_sents_1": 26728, "n_sents": 257, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 95, "L_abs": 1224, "title": "What Do You See? Enhancing Zero-Shot Image Classification with Multimodal Large Language Models", "abs": "Large language models (LLMs) have been effectively used for many computer vision tasks, including image classification. In this paper, we present a simple yet effective approach for zero-shot image classification using multimodal LLMs. By employing multimodal LLMs, we generate comprehensive textual representations from input images. These textual representations are then utilized to generate fixed-dimensional features in a cross-modal embedding space. Subsequently, these features are fused together to perform zero-shot classification using a linear classifier. Our method does not require prompt engineering for each dataset; instead, we use a single, straightforward, set of prompts across all datasets. We evaluated our method on several datasets, and our results demonstrate its remarkable effectiveness, surpassing benchmark accuracy on multiple datasets. On average, our method achieved an accuracy gain of 4.1 percentage points, with an increase of 6.8 percentage points on the ImageNet dataset, compared to prior methods. Our findings highlight the potential of multimodal LLMs to enhance computer vision tasks such as zero-shot image classification, offering a significant improvement over traditional methods.", "keywords": [ "large language models", "zero shot image classification", "cross modal representation" ], "rating_list": [ 3, 5, 3, 5, 6 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 5, 5, 5 ] }, { "paper_id": "O8fUZfC4GT", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Jiachen Jiang", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Jinxin Zhou", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Zhihui Zhu", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 12, "n_ref_uni": 26, "n_ref": 30, "n_ref_all": 46, "n_fig": 9, "n_tab": 3, "L_tab": 2655, "n_element_tab": 28, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1696, "formula_len_all_1": 1227, "len_all": 120796, "len_all_1": 62844, "len_abs": 1890, "len_title": 122, "len_sents": 32285, "len_sents_1": 31791, "n_sents": 195, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1896, "title": "GENERALIZATION, ROBUSTNESS AND ADAPTABILITY OF PROGRESSIVE NEURAL COLLAPSE", "abs": "Neural networks exhibit the neural collapse phenomenon in multi-class classification tasks, where last-layer features and linear classifier weights converge into a symmetric geometric structure. However, most prior studies have primarily focused on last-layer feature representations or have examined intermediate features using limited, simple architectures and datasets. The mechanisms by which deep neural networks separate data according to class membership across all layers in more complex and realistic scenarios, and how this separation evolves under distribution shifts, remain unclear. In this work, we extend the study of neural collapse to a broader range of architectures and datasets, investigating its progression throughout the network and its implications for generalization, robustness, and domain adaptability. Our findings reveal that well-trained neural networks progressively enhance neural collapse across layers, though a distinct transition phase occurs where this improvement plateaus after the initial layers and is followed by a renewed continuous improvement in the very last layers, with additional layers contributing minimal generalization benefits. Moreover, we observe that this progressive neural collapse pattern remains robust against noisy data, whether the noise occurs in inputs or labels, and that the degree of intermediate separation serves as an effective indicator of noise levels. Additionally, for the learned networks, comparing neural collapse evaluated on noisy data and clean data reveals insights into feature learning and memorization, with the latter primarily occurring in the very last layers. This finding aligns with the neural collapse pattern observed with clean training data. Finally, we show that when a shift occurs between source and target domains, intermediate neural collapse is closely related to downstream target performance.", "keywords": [ "Neural Collapse", "generalization", "robustness", "domain adaptability" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "O8FkMqNF1M", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Jongwoo Park", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Kumara Kahatapitiya", "gender": "Male", "institution": "Meta", "country": "US", "position": "Intern" }, { "name": "Michael S Ryoo", "gender": "Male", "institution": "Salesforce AI Research", "country": "US", "position": "Principal Researcher" }, { "name": "Kanchana Nisal Ranasinghe", "gender": "Male", "institution": "State University of New York, Stony Brook", "country": "US", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 14, "n_ref_uni": 74, "n_ref": 185, "n_ref_all": 214, "n_fig": 4, "n_tab": 8, "L_tab": 6840, "n_element_tab": 760, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1905, "n_element_tab_1": 204, "formula_len_all": 1032, "formula_len_all_1": 858, "len_all": 243710, "len_all_1": 67490, "len_abs": 919, "len_title": 96, "len_sents": 47354, "len_sents_1": 28053, "n_sents": 411, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 924, "title": "Language Repository for Long Video Understanding", "abs": "Language has become a prominent modality in computer vision with the rise of LLMs. Despite supporting long context-lengths, their effectiveness in handling long-term information gradually declines with input length. This becomes critical, especially in applications such as long-form video understanding. In this paper, we introduce a Language Repository (LangRepo) for LLMs, that maintains concise and structured information as an interpretable (i.e., all-textual) representation. Our repository is updated iteratively based on multi-scale video chunks. We introduce write and read operations that focus on pruning redundancies in text, and extracting information at various temporal scales. The proposed framework is evaluated on zero-shot visual question-answering benchmarks including EgoSchema, NExT-QA, IntentQA and NExT-GQA, showing state-of-the-art performance at its scale. Our code will be made publicly available.", "keywords": [ "Large-language models", "Long-video understanding", "Visual question-answering" ], "rating_list": [ 3, 5, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 5, 2, 4, 3 ] }, { "paper_id": "O7wTfBLSFn", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Wenqian Li", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yan Pang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 72, "n_formula_1": 30, "n_ref_uni": 15, "n_ref": 27, "n_ref_all": 50, "n_fig": 11, "n_tab": 4, "L_tab": 1290, "n_element_tab": 199, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7896, "formula_len_all_1": 2084, "len_all": 155788, "len_all_1": 67835, "len_abs": 6122, "len_title": 76, "len_sents": 52706, "len_sents_1": 31430, "n_sents": 437, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 28, "L_abs": 1104, "title": "Private Wasserstein Distance", "abs": "Wasserstein distance is a key metric for quantifying data divergence from a distributional perspective. However, its application in privacy-sensitive environments, where direct sharing of raw data is prohibited, presents significant challenges. Existing approaches, such as Differential Privacy and Federated Optimization, have been employed to estimate the Wasserstein distance under such constraints. However, these methods often fall short when both accuracy and security are required. In this study, we explore the inherent triangular properties within the Wasserstein space, leading to a novel solution named $\\texttt{TriangleWad}$. This approach facilitates the fast computation of the Wasserstein distance between datasets stored across different entities, ensuring that raw data remain completely hidden. TriangleWad not only strengthens resistance to potential attacks but also preserves high estimation accuracy.\nThrough extensive experiments across various tasks involving both image and text data, we demonstrate its superior performance and significant potential for real-world applications.", "keywords": [ "data valuation", "data marketplace", "optimal transport", "Wasserstein distance" ], "rating_list": [ 3, 6, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 3, 2 ], "contribution_list": [ 2, 4, 2, 3, 3 ], "confidence_list": [ 4, 3, 2, 4, 2 ] }, { "paper_id": "O6znYvxC1U", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Bernardo Cuenca Grau", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Professor" }, { "name": "Ouns El Harzli", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 22, "n_ref_uni": 45, "n_ref": 83, "n_ref_all": 89, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1399, "formula_len_all_1": 1372, "len_all": 136769, "len_all_1": 84488, "len_abs": 1124, "len_title": 144, "len_sents": 39242, "len_sents_1": 36828, "n_sents": 257, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1133, "title": "Bayesian Treatment of the Spectrum of the Empirical Kernel in (Sub)Linear-Width Neural Networks", "abs": "We study Bayesian neural networks (BNNs) in the theoretical limits of infinitely increasing number of training examples, network width and input space dimension. Our findings establish new bridges between kernel-theoretic approaches and techniques derived from statistical mechanics through the correspondence between Mercer's eigenvalues and limiting spectral distributions of covariance matrices studied in random matrix theory. \n Our theoretical contributions first consist in novel integral formulas that accurately describe the predictors of BNNs in the asymptotic linear-width and sublinear-width regimes. Moreover, we extend the recently developed renormalisation theory of deep linear neural networks, enabling a rigorous explanation of the mounting empirical evidence that hints at the theory's applicability to nonlinear BNNs with ReLU activations in the linear-width regime.\n From a practical standpoint, our results introduce a novel technique for estimating the predictor statistics of a trained BNN that is applicable to the sublinear-width regime where the predictions of the renormalisation theory are inaccurate.", "keywords": [ "infinite bayesian neural networks", "kernel theory", "random matrix theory" ], "rating_list": [ 6, 8, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 2, 3 ] }, { "paper_id": "O6p3v6i0hT", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Dooho Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Jaemin Yoo", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Myeong Kong", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Undergrad student" } ], "n_formula": 67, "n_formula_1": 16, "n_ref_uni": 34, "n_ref": 71, "n_ref_all": 88, "n_fig": 4, "n_tab": 13, "L_tab": 10111, "n_element_tab": 826, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 4152, "n_element_tab_1": 357, "formula_len_all": 4478, "formula_len_all_1": 865, "len_all": 182235, "len_all_1": 72036, "len_abs": 1116, "len_title": 126, "len_sents": 55782, "len_sents_1": 29377, "n_sents": 464, "n_sents_1": 242, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1111, "title": "Filling in the GAP: Achieving Robust and Adaptive GNNs through Post-Processing", "abs": "Graph neural networks (GNNs) have shown significant success in modeling graph-structured data. However, their performance often deteriorates when faced with a change in the graph structure between training and test time, such as edge addition or removal\u2014a common scenario considering the dynamic nature of graphs. To address this challenge, we propose FILLER (Framework for Integrating Layer-Level Edge-shift Recovery), a post-processing method which enhances the robustness of a GNN against edge sparsification while maintaining its adaptability to informative edge addition. Our key idea is to fill in the representation gap caused by edge distribution shift by injecting the Edge-shfit (ER) layer into each layer of the GNN. Our ER layer is carefully designed to allow a GNN to maintain its high performance in dynamic graph environments even without any additional training, and its effectiveness is shown both theoretically and empirically. Our experiments on ten datasets for node classification and five GNN architectures demonstrate that FILLER is broadly applicable across\ndiverse models and scenarios.", "keywords": [ "Graph neural networks", "edge distribution shift", "training-free post-processing", "robustness", "node classification" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 4, 2 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "O6W9SJRZRA", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Andre Kokozinski", "gender": "unknown", "institution": "Technische Universit\u00e4t Darmstadt", "country": "", "position": "PhD student" }, { "name": "Antonia W\u00fcst", "gender": "unknown", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "David Steinmann", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "", "position": "PhD student" }, { "name": "Kristian Kersting", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Full Professor" }, { "name": "Maurice Kraus", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 5, "n_ref_uni": 32, "n_ref": 51, "n_ref_all": 75, "n_fig": 13, "n_tab": 8, "L_tab": 5902, "n_element_tab": 398, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1965, "n_element_tab_1": 96, "formula_len_all": 471, "formula_len_all_1": 288, "len_all": 162958, "len_all_1": 68869, "len_abs": 918, "len_title": 125, "len_sents": 50771, "len_sents_1": 30426, "n_sents": 451, "n_sents_1": 276, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 922, "title": "Right on Time: Revising Time Series Models by Constraining their Explanations", "abs": "The reliability of deep time series models is often compromised by their tendency to rely on confounding factors, which may lead to incorrect outputs. Our newly recorded, naturally confounded dataset named P2S from a real mechanical production line emphasizes this. To avoid \u201cClever-Hans\u201d moments in time series, i.e., to mitigate confounders, we introduce the method Right on Time (RioT). RioT enables, for the first time interactions with model explanations across both the time and frequency domain. Feedback on explanations in both domains is then used to constrain the model, steering it away from the annotated confounding factors. The dual-domain interaction strategy is crucial for effectively addressing confounders in time series datasets. We empirically demonstrate that RioT can effectively guide models away from the wrong reasons in P2S as well as popular time series classification and forecasting datasets.", "keywords": [ "XIL", "time series", "spurious corelation", "Interaction", "XAI", "Dataset" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "O6QZ4W6GXt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Ang Li", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Eric P. Xing", "gender": "Male", "institution": "Mohamed bin Zayed Univeristy of AI", "country": "AE", "position": "Full Professor" }, { "name": "Guoheng Sun", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Hongyi Wang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Assistant Professor" }, { "name": "Souvik Kundu", "gender": "Male", "institution": "Intel", "country": "US", "position": "Researcher" }, { "name": "Yexiao He", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Yuning Zhang", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "MS student" }, { "name": "Zheyu Shen", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Ziyao Wang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 47, "n_ref": 78, "n_ref_all": 89, "n_fig": 3, "n_tab": 5, "L_tab": 1853, "n_element_tab": 114, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1266, "n_element_tab_1": 133, "formula_len_all": 369, "formula_len_all_1": 265, "len_all": 130051, "len_all_1": 61862, "len_abs": 1968, "len_title": 117, "len_sents": 28683, "len_sents_1": 28216, "n_sents": 221, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1490, "title": "ShareLoRA: Less Tuning, More Performance for LoRA Fine-tuning of LLMs", "abs": "Fine-tuning large language models (LLMs) is prohibitively expensive, prompting the development of various parameter-efficient fine-tuning (PEFT) methods. These methods primarily focus on fine-tuning small, additional modules known as adapters, which account for only a small fraction of the total LLM parameters. \nOne such method, low-rank adaptation (LoRA), has shown notable parameter efficiency while maintaining performance comparable to full fine-tuning. However, classical LoRA may still involve tuning more parameters than necessary given the intrinsic rank of pre-trained weights, as highlighted by prior work. \nIn this work, we introduce ShareLoRA, a novel approach that further enhances parameter efficiency during LLM fine-tuning by leveraging redundancies in pre-trained model weights to share LoRA modules, thereby significantly reducing the number of trainable parameters. Specifically, ShareLoRA automatically identifies redundancies in the pre-trained weights and determines which LoRA adapters can share parameters. This is achieved by measuring the similarity between representations to assess information redundancy and using a greedy algorithm to maximize parameter sharing. We conducted extensive evaluations on the LLMs of the LLaMA family across benchmark tasks. Notably, ShareLoRA achieves better parameter efficiency, with up to a 23\\% reduction in the number of fine-tuned parameters while delivering performance comparable to or better than existing PEFT methods.", "keywords": [ "generative models", "parameter-efficient-training", "fine-tuning" ], "rating_list": [ 3, 5, 3, 5, 5 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 2, 4, 3, 3, 2 ], "contribution_list": [ 2, 3, 1, 2, 3 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "O4y4m9biMx", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "KE DENG", "gender": "unknown", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "Assistant Professor" }, { "name": "Sishun Liu", "gender": "Male", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "PhD student" }, { "name": "Xiuzhen Zhang", "gender": "Female", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "Full Professor" }, { "name": "Yan Wang", "gender": "Male", "institution": "Macquarie University", "country": "AU", "position": "Full Professor" } ], "n_formula": 29, "n_formula_1": 22, "n_ref_uni": 61, "n_ref": 113, "n_ref_all": 139, "n_fig": 3, "n_tab": 13, "L_tab": 5110, "n_element_tab": 184, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1402, "n_element_tab_1": 62, "formula_len_all": 2426, "formula_len_all_1": 1968, "len_all": 214716, "len_all_1": 73706, "len_abs": 954, "len_title": 118, "len_sents": 36707, "len_sents_1": 25724, "n_sents": 295, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 958, "title": "Counterfactual History Distillation on Continuous-time Event Sequences", "abs": "This study aims to distill history events that have essential information for predicting subsequent events with counterfactual analysis. The problem is named Counterfactual History Distillation (CHD). CHD distills a minimum set of events from history, based on which the distribution provided by a trained MTPP model fits the events observed later, and the distribution based on the remaining events in history cannot. It can help understand what event marks may have more influence on the occurrence of future events and what events in history may have a causal relationship with the events observed later. This study proposes a robust solution for CHD, called MTPP-based Counterfactual History Distiller (MTPP-CHD). MTPP-CHD learns to select the optimal event combination from history for the events observed later. Experiment results demonstrate the superiority of MTPP-CHD by outperforming baselines in terms of distillation quality and processing speed.", "keywords": [ "Counterfactual Analysis", "Marked Temporal Point Process" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 2, 2, 4, 2 ] }, { "paper_id": "O4N9kWwV6R", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Qingfu Zhang", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Xi Lin", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Xiaoyuan Zhang", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yilu Liu", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zhenkun Wang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Fei LIU", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 88, "n_formula_1": 24, "n_ref_uni": 64, "n_ref": 124, "n_ref_all": 155, "n_fig": 5, "n_tab": 8, "L_tab": 2489, "n_element_tab": 96, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 984, "n_element_tab_1": 19, "formula_len_all": 6697, "formula_len_all_1": 2440, "len_all": 236456, "len_all_1": 69036, "len_abs": 1165, "len_title": 123, "len_sents": 71174, "len_sents_1": 29922, "n_sents": 495, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1170, "title": "Few for Many: Tchebycheff Set Scalarization for Many-Objective Optimization", "abs": "Multi-objective optimization can be found in many real-world applications where some conflicting objectives can not be optimized by a single solution. Existing optimization methods often focus on finding a set of Pareto solutions with different optimal trade-offs among the objectives. However, the required number of solutions to well approximate the whole Pareto optimal set could be exponentially large with respect to the number of objectives, which makes these methods unsuitable for handling many optimization objectives. In this work, instead of finding a dense set of Pareto solutions, we propose a novel Tchebycheff set scalarization method to find a few representative solutions (e.g., 5) to cover a large number of objectives (e.g., $>100$) in a collaborative and complementary manner. In this way, each objective can be well addressed by at least one solution in the small solution set. In addition, we further develop a smooth Tchebycheff set scalarization approach for efficient optimization with good theoretical guarantees. Experimental studies on different problems with many optimization objectives demonstrate the effectiveness of our proposed method.", "keywords": [ "multi-objective optimization", "many-objective optimization", "Tchebycheff scalarization" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 2, 4, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 4, 2, 4 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "O4LoPhRSfb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Changshuo Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Daizong Liu", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Wanlong Fang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xiang Fang", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xiaoye Qu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 36, "n_ref": 84, "n_ref_all": 108, "n_fig": 8, "n_tab": 13, "L_tab": 6737, "n_element_tab": 1364, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 713, "formula_len_all_1": 527, "len_all": 201286, "len_all_1": 67619, "len_abs": 1815, "len_title": 118, "len_sents": 45766, "len_sents_1": 28425, "n_sents": 426, "n_sents_1": 242, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1353, "title": "Does Your Video-language Model Actually Understand the Language Input?", "abs": "Driven by the wave of Large Language Models (LLMs), Video-Language Models (VLMs) have become a significant yet challenging technology to bridge the gap between video and text. Although previous VLM works have made significant progress, almost all of them implicitly assume that all the texts are predefined by the specific template. In real-world applications, such an assumption is impossible to satisfy, since predefining all the texts is extremely time-consuming and labor-intensive. Besides, these predefined text inputs are too strict and user-unfriendly, limiting their applications. It is observed that given a video input, texts with similar semantics lead to various performances. To this end, in this paper, we propose a novel text-augmented VLM method to improve video-text fusion by text rewriting. Specifically, we first generate various text samples from the original ones based on the pre-trained LLM to target specific text components. A multi-level contrastive learning module is designed to mine the coarse-grained language information. Moreover, we also propose an attribute-based text reasoning strategy to learn fine-grained textual semantics. Extensive experiments on many video-language tasks show that the proposed method can serve as the plug-and-play module to effectively improve the performance of state-of-the-art VLM works.", "keywords": [ "Video-language Model", "Coarse-grained Language Alignment", "Attribute-based Text Reasoning", "Fine-grained Language Alignment" ], "rating_list": [ 5, 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 1, 3, 2 ], "presentation_list": [ 3, 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 5, 3, 4 ] }, { "paper_id": "O3SatrdL97", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "David Grangier", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Pierre Ablin", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Simin Fan", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 14, "n_ref_uni": 29, "n_ref": 70, "n_ref_all": 98, "n_fig": 14, "n_tab": 7, "L_tab": 3683, "n_element_tab": 257, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1393, "n_element_tab_1": 57, "formula_len_all": 1219, "formula_len_all_1": 931, "len_all": 165818, "len_all_1": 75510, "len_abs": 4142, "len_title": 97, "len_sents": 50348, "len_sents_1": 34140, "n_sents": 411, "n_sents_1": 276, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1485, "title": "Dynamic Gradient Alignment for Online Data Mixing", "abs": "The composition of training data mixtures is critical for effectively training large language models (LLMs), as it directly impacts their performance on downstream tasks. Our goal is to identify an optimal data mixture to specialize an LLM for a specific task with access to only a few examples. Traditional approaches to this problem include ad-hoc reweighting methods, importance sampling, and gradient alignment techniques.\nThis paper focuses on gradient alignment and introduces Dynamic Gradient Alignment (DGA), a scalable online gradient alignment algorithm. DGA dynamically estimates the pre-training data mixture on which the models' gradients align as well as possible with those of the model on the specific task.\nDGA is the first gradient alignment approach that incurs minimal overhead compared to standard pre-training and outputs a competitive model, eliminating the need for retraining the model. Experimentally, we demonstrate significant improvements over importance sampling in two key scenarios: (i) when the pre-training set is small and importance sampling overfits due to limited data; and (ii) when there is insufficient specialized data, trapping importance sampling on narrow pockets of data.\nOur findings underscore the effectiveness of gradient alignment methods in optimizing training data mixtures, particularly in data-constrained environments, and offer a practical solution for enhancing LLM performance on specific tasks with limited data availability.", "keywords": [ "LLM pretraining", "data reweighing", "data mixing" ], "rating_list": [ 5, 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4, 5 ] }, { "paper_id": "O34CXUAZ0E", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Cheng Hong", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Leo Yu Zhang", "gender": "Male", "institution": "Griffith University", "country": "AU", "position": "Researcher" }, { "name": "Minghui Li", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Ning Yuxuan", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "", "position": "Undergrad student" }, { "name": "Shengshan Hu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Wei Wan", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yechao Zhang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Zhicong Huang", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Ziqi Zhou", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 23, "n_ref": 57, "n_ref_all": 75, "n_fig": 3, "n_tab": 7, "L_tab": 1810, "n_element_tab": 159, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 3206, "n_element_tab_1": 56, "formula_len_all": 1122, "formula_len_all_1": 348, "len_all": 144670, "len_all_1": 62965, "len_abs": 1226, "len_title": 110, "len_sents": 54074, "len_sents_1": 28391, "n_sents": 409, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1231, "title": "MARS: A Malignity-Aware Backdoor Defense in Federated Learning", "abs": "Federated Learning (FL) is a distributed paradigm aimed at protecting participant data privacy by exchanging model parameters to achieve high-quality model training. However, this distributed nature also makes FL highly vulnerable to backdoor attacks. Notably, the recently proposed state-of-the-art (SOTA) attack, 3DFed (SP2023), uses an indicator mechanism to determine whether the backdoor models have been accepted by the defender and adaptively optimizes backdoor models, rendering existing defenses ineffective. In this paper, we first reveal that the failure of existing defenses lies in the employment of empirical statistical measures that are loosely coupled with backdoor attacks. Motivated by this, we propose a Malignity-Aware backdooR defenSe (MARS) that leverages backdoor energy (BE) to indicate the malicious extent of each neuron. To amplify malignity, we further extract the most prominent BE values from each model to form a concentrated backdoor energy (CBE). Finally, a novel Wasserstein distance-based clustering method is introduced to effectively identify backdoor models. Extensive experiments demonstrate that MARS can defend against SOTA backdoor attacks and significantly outperforms existing defenses.", "keywords": [ "Federated Learning", "Backdoor Attack", "Backdoor Defense" ], "rating_list": [ 3, 6, 6, 5 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 2, 4, 4, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "O2aioX2Z2v", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Albert Peng", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "US", "position": "Undergrad student" }, { "name": "Mauricio Delbracio", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Peyman Milanfar", "gender": "Male", "institution": "Google", "country": "US", "position": "Distinguished Scientist" }, { "name": "Weijie Gan", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "US", "position": "PhD student" }, { "name": "Yuyang Hu", "gender": "unknown", "institution": "Washington University in Saint Louis", "country": "US", "position": "PhD student" }, { "name": "Ulugbek Kamilov", "gender": "Not Specified", "institution": "Washington University, Saint Louis", "country": "US", "position": "Associate Professor" } ], "n_formula": 43, "n_formula_1": 15, "n_ref_uni": 72, "n_ref": 168, "n_ref_all": 194, "n_fig": 9, "n_tab": 10, "L_tab": 1893, "n_element_tab": 83, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1094, "n_element_tab_1": 49, "formula_len_all": 2605, "formula_len_all_1": 554, "len_all": 217405, "len_all_1": 65928, "len_abs": 1061, "len_title": 111, "len_sents": 58524, "len_sents_1": 30659, "n_sents": 450, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1068, "title": "Stochastic Deep Restoration Priors for Imaging Inverse Problems", "abs": "Deep neural networks trained as image denoisers are widely used as priors for solving imaging inverse problems. While Gaussian denoising is thought sufficient for learning image priors, we show that priors from deep models pre-trained as more general restoration operators can perform better. We introduce Stochastic deep Restoration Priors (ShaRP), a novel method that leverages an ensemble of such restoration models to regularize inverse problems. ShaRP improves upon methods using Gaussian denoiser priors by better handling structured artifacts and enabling self-supervised training even without fully sampled data. We prove ShaRP minimizes an objective function involving a regularizer derived from the score functions of minimum mean square error (MMSE) restoration operators, and theoretically analyze its convergence. Empirically, ShaRP achieves state-of-the-art performance on tasks such as magnetic resonance imaging reconstruction and single-image super-resolution, surpassing both denoiser- and diffusion-model-based methods without requiring retraining.", "keywords": [ "computational imaging", "inverse problems", "deep learning" ], "rating_list": [ 6, 1, 6, 5, 3 ], "soundness_list": [ 3, 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 3, 2, 1 ], "confidence_list": [ 5, 5, 5, 3, 3 ] }, { "paper_id": "O2GBkHujdP", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Erin E. Tripp", "gender": "unknown", "institution": "Air Force Research Laboratory", "country": "", "position": "Researcher" }, { "name": "Shaofeng Zou", "gender": "unknown", "institution": "Arizona State University", "country": "", "position": "Associate Professor" }, { "name": "Yi Zhou", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Associate Professor" }, { "name": "Yifan Sun", "gender": "Female", "institution": "State University of New York, Stony Brook", "country": "", "position": "Assistant Professor" }, { "name": "Yufeng Yang", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" } ], "n_formula": 185, "n_formula_1": 35, "n_ref_uni": 26, "n_ref": 83, "n_ref_all": 97, "n_fig": 5, "n_tab": 2, "L_tab": 97, "n_element_tab": 3, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 262, "n_element_tab_1": 5, "formula_len_all": 15100, "formula_len_all_1": 1915, "len_all": 189732, "len_all_1": 62489, "len_abs": 1009, "len_title": 122, "len_sents": 52817, "len_sents_1": 26717, "n_sents": 494, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1034, "title": "Independently-Normalized SGD for Generalized-Smooth Nonconvex Optimization", "abs": "Recent studies have shown that many nonconvex machine learning problems meet a so-called generalized-smooth condition that extends beyond traditional smooth nonconvex optimization. However, the existing algorithms designed for generalized-smooth nonconvex optimization encounter significant limitations in both their design and convergence analysis.\nIn this work, we first study deterministic generalized-smooth nonconvex optimization and analyze the convergence of normalized gradient descent under the generalized Polyak-Lojasiewicz condition. Our results provide a comprehensive understanding of the interplay between gradient normalization and function geometry. Then, for stochastic generalized-smooth nonconvex optimization, we propose an independently-normalized stochastic gradient descent algorithm, which leverages independent sampling, gradient normalization, and clipping to achieve an $\\mathcal{O}(\\epsilon^{-4})$ sample complexity under relaxed assumptions. Experiments demonstrate the fast convergence of our algorithm.", "keywords": [ "Non-convex optimization", "Stochastic Algorithm" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 1, 2, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "O2DVmb0pwo", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Cong Liu", "gender": "unknown", "institution": "University of California, Riverside", "country": "", "position": "Associate Professor" }, { "name": "Haizhou Li", "gender": "Male", "institution": "The Chinese University of Hong Kong (Shenzhen); National University of Singapore", "country": "CN", "position": "Full Professor" }, { "name": "Luis Fernando D'Haro", "gender": "Male", "institution": "Universidad Polit\u00e9cnica de Madrid", "country": "", "position": "Associate Professor" }, { "name": "Yiming Chen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zexin Li", "gender": "Male", "institution": "University of California, Riverside", "country": "US", "position": "PhD student" }, { "name": "Robby T. Tan", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" }, { "name": "Yue Xianghu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 53, "n_ref": 105, "n_ref_all": 115, "n_fig": 4, "n_tab": 13, "L_tab": 8626, "n_element_tab": 2654, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 3833, "n_element_tab_1": 1148, "formula_len_all": 20, "formula_len_all_1": 164, "len_all": 286389, "len_all_1": 85712, "len_abs": 1414, "len_title": 100, "len_sents": 55299, "len_sents_1": 32868, "n_sents": 443, "n_sents_1": 244, "venue": "ICLR2025", "decision": "withdraw", "L_title": 51, "L_abs": 1419, "title": "Energy-Oriented Alignment for Large Language Models", "abs": "Large language models (LLMs) have showcased remarkable capabilities on a variety of natural language processing (NLP) tasks, powering various real-world applications.\nEnsuring the safe and effective deployment of LLMs requires careful alignment to mitigate risks associated with malicious inputs, which now mainly involve toxic content and misinformation.\nIn this study, we expand this focus by identifying and exploring a novel category of energy-oriented malicious instructions, akin to Denial-of-Service (DoS) attacks.\nThese instructions provoke LLMs to generate excessively lengthy responses through impractical tasks, resulting in high energy and computational resource consumption, and even risking system overload.\nTo address this gap, we curate EnergyAlign, the first energy-oriented malicious instruction dataset with 8 diverse categories.\nThen, we conduct a comprehensive evaluation of 5 advanced proprietary LLMs and 24 open-source LLMs.\nThe results reveal a notable disparity: while proprietary LLMs can refuse such malicious inputs, most open-source LLMs are extremely vulnerable with a failure rate of up to 96.8\\%.\nAdditionally, we assess the effectiveness of jailbreak techniques in bypassing the energy-related safety measures of proprietary models.\nLastly, we highlight the inadequacies of existing defense mechanisms and propose energy-oriented alignment data against EnergyAlign for future research.", "keywords": [ "datasets and benchmarks", "jailbreak", "energy attack", "large language model", "efficiency", "denial-of-service" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "O2CG9B2k9Q", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Amit Sethi", "gender": "Male", "institution": "Indian Institute of Technology, Bombay, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "Full Professor" }, { "name": "Neeraj Nixon", "gender": "Male", "institution": "Indian Institute of Technology, Bombay, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "Undergrad student" }, { "name": "Pranav Jeevan P", "gender": "Male", "institution": "Indian Institute of Technology Bombay, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 16, "n_ref": 40, "n_ref_all": 57, "n_fig": 9, "n_tab": 1, "L_tab": 517, "n_element_tab": 7, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 362, "formula_len_all_1": 354, "len_all": 89080, "len_all_1": 46999, "len_abs": 1512, "len_title": 98, "len_sents": 37356, "len_sents_1": 23529, "n_sents": 253, "n_sents_1": 158, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1471, "title": "Normalizing Flow Based Evaluation Metrics for Image Generation", "abs": "We propose two new evaluation metrics to assess realness of generated images based on normalizing flows: a simpler and efficient flow-based likelihood distance (FLD) and a more exact dual-flow based likelihood distance (D-FLD). Because normalizing flows can be used to compute the exact likelihood, the proposed metrics assess how closely generated images align with the distribution of real images from a given domain. This property gives the proposed metrics a few advantages over the widely used Fr\u00e9chet inception distance (FID) and other recent metrics. Firstly, the proposed metrics need only a few hundred images to stabilize (converge in mean), as opposed to tens of thousands needed for FID, and at least a few thousand for the other metrics. This allows confident evaluation of even small sets of generated images, such as validation batches inside training loops. Secondly, the network used to compute the proposed metric has over an order of magnitude fewer parameters compared to Inception-V3 used to compute FID, making it computationally more efficient. For assessing the realness of generated images in new domains (e.g., x-ray images), ideally these networks should be retrained on real images to model their distinct distributions. Thus, our smaller network will be even more advantageous for new domains. Extensive experiments show that the proposed metrics have the desired monotonic relationships with the extent of image degradation of various kinds.", "keywords": [ "image generation", "evaluation metric", "FID", "normalizing flow" ], "rating_list": [ 1, 5, 3, 6 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 1, 3, 2, 4 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "O13fIFEB81", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 16, "author_info_list": [ { "name": "Hanqing Lu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Hui Liu", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Jiacheng Lin", "gender": "Male", "institution": "Department of Computer Science, University of Illinois", "country": "US", "position": "PhD student" }, { "name": "Jingrui He", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Ruirui Li", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Ruizhong Qiu", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Sreyashi Nag", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Suhang Wang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Associate Professor" }, { "name": "Tianxin Wei", "gender": "unknown", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Wenju Xu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Xianfeng Tang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yifan Chen", "gender": "unknown", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Assistant Professor" }, { "name": "Yunzhe Qi", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Zhengyang Wang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Qi He", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Director of Applied Science" }, { "name": "chen luo", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 7, "n_ref_uni": 30, "n_ref": 72, "n_ref_all": 88, "n_fig": 5, "n_tab": 8, "L_tab": 2493, "n_element_tab": 348, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1158, "n_element_tab_1": 181, "formula_len_all": 718, "formula_len_all_1": 367, "len_all": 126266, "len_all_1": 62600, "len_abs": 1550, "len_title": 124, "len_sents": 42440, "len_sents_1": 30390, "n_sents": 282, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1557, "title": "Robust Watermarking for Diffusion Models: A Unified Multi-Dimensional Recipe", "abs": "Diffusion models are known for the supreme capability to generate realistic images. However, ethical concerns, such as copyright protection and generation of inappropriate content, pose significant challenges for the practical deployment of diffusion models. Recent work has proposed a flurry of watermarking techniques that inject visually noteless patterns into generated images, offering a promising solution to these issues. While effective, the essential elements for watermarking and the interconnections among various methods are still chaos. In this paper, we dissect the design principles of state-of-the-art watermarking techniques and introduce a unified framework. We identify a set of dimensions that explain the manipulation enforced by watermarking methods, including the distribution of individual elements, the specification of watermark regions within each channel, and the choice of channels for watermark embedding. Moreover, under this framework we instantiate a new watermarking method to minimize impacts on the model performance from a distributional perspective. Through the empirical studies on regular text-to-image applications and the first systematic attempt on watermarking image-to-image diffusion models, we thoroughly verify the effectiveness of our proposed framework through comprehensive evaluations. On all the diffusion models, including Stable Diffusion, our approach induced from the proposed framework not only preserves image quality but also outperforms existing methods in robustness against a range of attacks.", "keywords": [ "Diffusion model", "Watermark", "Unified framework" ], "rating_list": [ 1, 5, 5, 6, 5 ], "soundness_list": [ 1, 3, 2, 3, 2 ], "presentation_list": [ 1, 3, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 5, 3, 4 ] }, { "paper_id": "O0z4mkjl7i", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Krishnan Raghavan", "gender": "Male", "institution": "Argonne National Laboratory", "country": "US", "position": "Researcher" }, { "name": "Supriyo Chakraborty", "gender": "unknown", "institution": "Capital One", "country": "US", "position": "Researcher" } ], "n_formula": 60, "n_formula_1": 11, "n_ref_uni": 55, "n_ref": 119, "n_ref_all": 149, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5796, "formula_len_all_1": 1190, "len_all": 164159, "len_all_1": 67196, "len_abs": 1169, "len_title": 72, "len_sents": 43827, "len_sents_1": 30565, "n_sents": 350, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1176, "title": "On Understanding of the Dynamics of Model Capacity in Continual Learning", "abs": "The core issue in continual learning (CL) is balancing catastrophic forgetting of prior knowledge with generalization to new tasks, otherwise, known as the stability-plasticity dilemma. We argue that the dilemma is akin to the capacity~(the networks' ability to represent tasks) of the neural network~(NN) in the CL setting. Within this context, this work introduces ``CL\u2019s effective model capacity (CLEMC)\" to understand the dynamical behavior of stability-plasticity balance point in the CL setting. We define CLEMC as a function of the NN, the task data, and the optimization procedure. Leveraging CLEMC, we demonstrate that the capacity is non-stationary and regardless of the NN architecture and optimization method, the network\u2019s ability to represent new tasks diminishes if the incoming tasks\u2019 data distributions differ from previous ones. We formulate these results using dynamical systems' theory and conduct extensive experiments to complement the findings. Our analysis extends from a small feed-forward~(FNN) and convolutional networks~(CNN) to medium sized graph neural networks~(GNN) to transformer-based large language models~(LLM) with millions of parameters.", "keywords": [ "Continual Learning", "dynamic pogramming", "theory", "CL challenges." ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "O0sQ9CPzai", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Weibin Liao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xu Chu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Yasha Wang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 17, "n_ref_uni": 49, "n_ref": 101, "n_ref_all": 116, "n_fig": 6, "n_tab": 7, "L_tab": 3542, "n_element_tab": 418, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1972, "n_element_tab_1": 261, "formula_len_all": 2244, "formula_len_all_1": 1533, "len_all": 172912, "len_all_1": 64032, "len_abs": 1964, "len_title": 83, "len_sents": 45689, "len_sents_1": 27631, "n_sents": 378, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1720, "title": "TPO: Aligning Large Language Models with Multi-branch & Multi-step Preference Trees", "abs": "In the domain of complex reasoning tasks, such as mathematical reasoning, recent advancements have proposed the use of Direct Preference Optimization (DPO) to suppress output of dispreferred responses, thereby enhancing the long-chain reasoning capabilities of large language models (LLMs). To this end, these studies employed LLMs to generate preference trees via Tree-of-thoughts (ToT) and sample the paired preference responses required by the DPO algorithm. However, the DPO algorithm based on binary preference optimization is unable to learn multiple responses with varying degrees of preference/dispreference that provided by the preference trees, resulting in incomplete preference learning. In this work, we introduce Tree Preference Optimization (TPO), that does not sample paired preference responses from the preference tree; instead, it directly learns from the entire preference tree during the fine-tuning. Specifically, TPO formulates the language model alignment as a Preference List Ranking problem, where the policy can potentially learn more effectively from a ranked preference list of responses given the prompt. In addition, to further assist LLMs in identifying discriminative steps within long-chain reasoning and increase the relative reward margin in the preference list, TPO utilizes Adaptive Step Reward to adjust the reward values of each step in trajectory for performing fine-grained preference optimization. We carry out extensive experiments on mathematical reasoning tasks to evaluate TPO. The experimental results indicate that TPO consistently outperforms DPO across five public large language models on four datasets. The code is available on https://anonymous.4open.science/r/TPO.", "keywords": [ "Reinforcement Learning from Human Feedback", "Language Models", "Preferences Learning" ], "rating_list": [ 6, 5, 8 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "O0RIrM5iqX", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Chaoyue Song", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Fayao Liu", "gender": "Female", "institution": "A*STAR", "country": "SG", "position": "Research Scientist" }, { "name": "Jiacheng Wei", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "Wenhao Shen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xiaofeng Yang", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Xulei Yang", "gender": "Male", "institution": "I2R, A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Guosheng Lin", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "ZHOUJIE FU", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 21, "n_ref_uni": 44, "n_ref": 73, "n_ref_all": 87, "n_fig": 11, "n_tab": 1, "L_tab": 300, "n_element_tab": 34, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1261, "formula_len_all_1": 889, "len_all": 148814, "len_all_1": 47853, "len_abs": 216, "len_title": 122, "len_sents": 30483, "len_sents_1": 21109, "n_sents": 239, "n_sents_1": 160, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1292, "title": "Sync4D: Video Guided Controllable Dynamics for Physics-Based 4D Generation", "abs": "In this work, we introduce a novel approach for creating controllable dynamics in 3D-generated Gaussians using casually captured reference videos. Our method transfers the motion of objects from reference videos to a variety of generated 3D Gaussians across different categories, ensuring precise and customizable motion transfer. We achieve this by employing blend skinning-based non-parametric shape reconstruction to extract the shape and motion of reference objects. This process involves segmenting the reference objects into motion-related parts based on skinning weights and establishing shape correspondences with generated target shapes. To address shape and temporal inconsistencies prevalent in existing methods, we integrate physical simulation, driving the target shapes with matched motion. This integration is optimized through a displacement loss to ensure reliable and genuine dynamics. Our approach supports diverse reference inputs, including humans, quadrupeds, and articulated objects, and can generate dynamics of arbitrary length, providing enhanced fidelity and applicability. Unlike methods heavily reliant on diffusion video generation models, our technique offers specific and high-quality motion transfer, maintaining both shape integrity and temporal consistency.", "keywords": [ "4D generation" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "O08nfMzc93", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Moyang Guo", "gender": "Male", "institution": "Duke University, Duke University", "country": "US", "position": "MS student" }, { "name": "Yuepeng Hu", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Zhengyuan Jiang", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Neil Gong", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Associate Professor" } ], "n_formula": 34, "n_formula_1": 14, "n_ref_uni": 35, "n_ref": 81, "n_ref_all": 103, "n_fig": 12, "n_tab": 6, "L_tab": 977, "n_element_tab": 105, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2857, "formula_len_all_1": 679, "len_all": 157256, "len_all_1": 72778, "len_abs": 3509, "len_title": 111, "len_sents": 55259, "len_sents_1": 35295, "n_sents": 461, "n_sents_1": 277, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 950, "title": "Watermark-based Detection and Attribution of AI-Generated Image", "abs": "Several companies--such as Google, Microsoft, and OpenAI--have deployed techniques to watermark AI-generated images to enable proactive detection. However, existing literature mainly focuses on user-agnostic detection. Attribution aims to further trace back the user who generated a detected AI-generated image. Despite its growing importance, attribution is largely unexplored. In this work, we aim to bridge this gap by providing the first systematic study on watermark-based, user-aware detection and attribution of AI-generated images. Specifically, we theoretically study the detection and attribution performance via rigorous probabilistic analysis. Moreover, we develop an efficient algorithm to select watermarks for the users to enhance attribution performance. Both our theoretical and empirical results show that watermark-based detection and attribution inherit the accuracy and (non-)robustness properties of the watermarking method.", "keywords": [ "Image Watermark", "AI-generated image", "Watermark detection", "Watermark attribution" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "NzEIjnIIzv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Dawid Jan Kopiczko", "gender": "Male", "institution": "University of Technology Nuremberg", "country": "DE", "position": "PhD student" }, { "name": "Tijmen Blankevoort", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Yuki Asano", "gender": "Male", "institution": "University of Technology Nuremberg", "country": "DE", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 32, "n_ref": 42, "n_ref_all": 55, "n_fig": 3, "n_tab": 15, "L_tab": 6069, "n_element_tab": 700, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 504, "n_element_tab_1": 65, "formula_len_all": 681, "formula_len_all_1": 358, "len_all": 200244, "len_all_1": 54305, "len_abs": 873, "len_title": 119, "len_sents": 52009, "len_sents_1": 27011, "n_sents": 411, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 713, "title": "Bitune: Leveraging Bidirectional Attention to Improve Decoder-Only LLMs", "abs": "Decoder-only large language models typically rely solely on masked causal attention, which limits their expressiveness by restricting information flow to one direction. We propose Bitune, a method that enhances pretrained decoder-only LLMs by incorporating bidirectional attention into prompt processing. We evaluate Bitune in instruction-tuning and question-answering settings, showing significant improvements in performance on commonsense reasoning, arithmetic, and language understanding tasks. Furthermore, extensive ablation studies validate the role of each component of the method, and demonstrate that Bitune is compatible with various parameter-efficient finetuning techniques and full model finetuning.", "keywords": [ "Instruction tuning", "Parameter-efficient fine-tuning", "Transformer", "PEFT", "LLM" ], "rating_list": [ 8, 5, 10, 6 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 4, 3, 4, 4 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "NxyfSW6mLK", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Dinesh Jayaraman", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Assistant Professor" }, { "name": "Insup Lee", "gender": "unknown", "institution": "University of Pennsylvania", "country": "", "position": "Full Professor" }, { "name": "Kaustubh Sridhar", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Souradeep Dutta", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 63, "n_ref": 138, "n_ref_all": 190, "n_fig": 16, "n_tab": 8, "L_tab": 6681, "n_element_tab": 1256, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 475, "formula_len_all_1": 444, "len_all": 224783, "len_all_1": 66934, "len_abs": 978, "len_title": 91, "len_sents": 59122, "len_sents_1": 33062, "n_sents": 445, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 90, "L_abs": 983, "title": "REGENT: A Retrieval-Augmented Generalist Agent That Can Act In-Context in New Environments", "abs": "Do generalist agents only require large models pre-trained on massive amounts of data to rapidly adapt to new environments? We propose a novel approach to pre-train relatively small models and adapt them to unseen environments via in-context learning, without any finetuning. Our key idea is that retrieval offers a powerful bias for fast adaptation. Indeed, we demonstrate that even a simple retrieval-based 1-nearest neighbor agent offers a surprisingly strong baseline for today's state-of-the-art generalist agents. From this starting point, we construct a semi-parametric agent, REGENT, that trains a transformer-based policy on sequences of queries and retrieved neighbors. REGENT can generalize to unseen robotics and game-playing environments via retrieval augmentation and in-context learning, achieving this with up to 3x fewer parameters and up to an order-of-magnitude fewer pre-training datapoints, significantly outperforming today's state-of-the-art generalist agents.", "keywords": [ "Generalist Agent", "Retrieval", "In-Context Learning", "Imitation Learning", "Reinforcement Learning" ], "rating_list": [ 8, 5, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "NxsTjmRAzA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Pin-Yu Chen", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Qiang Xu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Tsung-Yi Ho", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "", "position": "Full Professor" }, { "name": "Zhiyuan He", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yijun YANG", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 38, "n_ref": 88, "n_ref_all": 125, "n_fig": 11, "n_tab": 13, "L_tab": 6478, "n_element_tab": 663, "n_fig_1": 8, "n_tab_1": 7, "L_tab_1": 2941, "n_element_tab_1": 191, "formula_len_all": 84, "formula_len_all_1": 84, "len_all": 155691, "len_all_1": 62201, "len_abs": 1061, "len_title": 135, "len_sents": 38420, "len_sents_1": 25577, "n_sents": 346, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1074, "title": "Test Time Augmentations are Worth One Million Images for Out-of-Distribution Detection", "abs": "Out-of-distribution (OOD) detection is a major threat for deploying machine learning models in safety-critical scenarios. Data augmentations have been proven to be beneficial to OOD detection by providing diverse features. However, previous methods have only focused on the role of data augmentation in the training phase, overlooking its impact on the testing phase. In this paper, we present the first comprehensive study of the impact of test-time augmentation (TTA) on OOD detection. We find aggressive TTAs can cause distribution shifts on OOD scores of In-distribution (InD) data, whereas mild TTAs do not, resulting in the effectiveness of mild TTAs on OOD Detection. Based on the above observations, we propose a detection method that performs a K-nearest-neighbor (KNN) search on mild TTAs instead of InD data. With only 25 TTAs, our method outperforms state-of-the-art methods using the entire training set (1.2 million images) on IMAGENET for OOD detection. Moreover, our approach is compatible with various model architectures and robust to adversarial examples.", "keywords": [ "Out-of-distribution", "Test time augmentation", "OOD Detection" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "NxLWeK4P3q", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Isao Ishikawa", "gender": "Male", "institution": "Ehime University", "country": "JP", "position": "Associate Professor" }, { "name": "Masahiro Ikeda", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Researcher" }, { "name": "Yuka Hashimoto", "gender": "unknown", "institution": "NTT", "country": "", "position": "Researcher" }, { "name": "Sho Sonoda", "gender": "Male", "institution": "RIKEN", "country": "", "position": "Researcher" } ], "n_formula": 51, "n_formula_1": 40, "n_ref_uni": 31, "n_ref": 59, "n_ref_all": 64, "n_fig": 2, "n_tab": 1, "L_tab": 98, "n_element_tab": 3, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5516, "formula_len_all_1": 3733, "len_all": 108789, "len_all_1": 61445, "len_abs": 845, "len_title": 130, "len_sents": 31494, "len_sents_1": 25721, "n_sents": 288, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 855, "title": "Unified Universality Theorem for Deep and Shallow Joint-Group-Equivariant Machines", "abs": "We present a constructive universal approximation theorem for learning machines equipped with joint-group-equivariant feature maps, based on the group representation theory. ``Constructive'' here indicates that the distribution of parameters is given in a closed-form expression known as the ridgelet transform. Joint-group-equivariance encompasses a broad class of feature maps that generalize classical group-equivariance. Notably, this class includes fully-connected networks, which are *not* group-equivariant *but* are joint-group-equivariant. Moreover, our main theorem also unifies the universal approximation theorems for both shallow and deep networks. While the universality of shallow networks has been investigated in a unified manner by the ridgelet transform, the universality of deep networks has been investigated in a case-by-case manner.", "keywords": [ "group theory", "irreducible representation", "universality", "fully-connected network", "joint-group-equivariance", "ridgelet transform" ], "rating_list": [ 6, 3, 6 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 4, 2, 4 ], "contribution_list": [ 3, 1, 4 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "Nx8lVqyKeZ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Florent Gu\u00e9pin", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Matthieu Meeus", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Nata\u0161a Kr\u010do", "gender": "Female", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Yves-Alexandre de Montjoye", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 31, "n_ref": 82, "n_ref_all": 102, "n_fig": 7, "n_tab": 8, "L_tab": 932, "n_element_tab": 198, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 562, "n_element_tab_1": 85, "formula_len_all": 799, "formula_len_all_1": 625, "len_all": 127956, "len_all_1": 64046, "len_abs": 2363, "len_title": 86, "len_sents": 41515, "len_sents_1": 29875, "n_sents": 290, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 2385, "title": "Lost in the Averages: Evaluating record-specific MIAs against Machine Learning models", "abs": "Record-specific Membership Inference Attacks (MIAs) are widely used to evaluate the propensity of a machine learning (ML) model to memorize an individual record and the privacy risk its release therefore poses. Record-specific MIAs are currently evaluated the same way ML models are: on a test set of models trained on data samples that were not seen during training ($D_{eval}$). A recent large body of literature has however shown that the main risk often comes from outliers, records that are statistically different from the rest of the dataset. In this work, we argue that the traditional evaluation setup for record-specific MIAs, which includes dataset sampling as a source of randomness, incorrectly captures the privacy risk. Indeed, what is an outlier is highly specific to particular data samples, and a record that is an outlier in the training dataset will not necessarily be one in the randomly sampled test datasets. We propose to use model randomness as the only source of randomness to evaluate record-level MIAs, a setup we call *model-seeded*. Across 10 combinations of models, datasets, and attacks for predictive and generative AI, we show the per-record risk estimates given by the traditional evaluation setup to substantially differ from ones given by the *model-seeded* setup which properly account for the increased risk posed by outliers. We show that across setups the traditional evaluation method leads to a substantial number of records to be incorrectly classified as low risk, emphasizing the inadequacy of the current setup to capture the record-level risk. We then a) provide evidence that the traditional setup is an average--across datasets--of the *model-seeded* risk, validating our use of model randomness to create evaluation models and b) show how relying on the traditional setup might conceal the existence of stronger attacks. The traditional setup would indeed strongly underestimate the risk posed by the strong Differential Privacy adversary. We believe our results to convincingly show the practice of randomizing datasets to evaluate record-specific MIAs to be incorrect. We then argue that relying on model randomness, an setup we call *model-seeded* evaluation, better captures the risk posed by outliers and should be used moving forward to evaluate record-level MIAs against machine learning models, both predictive and generative.", "keywords": [ "Privacy", "Membership inference attacks" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 2, 1, 1, 3 ], "presentation_list": [ 2, 2, 2, 4 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Nx6Bb5uxfI", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Cheng Wang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Chenglu Wen", "gender": "unknown", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Hai Wu", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Maoji Zheng", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Pufan Zou", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Shijia Zhao", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Xusheng Guo", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "qiming Xia", "gender": "Male", "institution": "XMU", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 37, "n_ref": 91, "n_ref_all": 99, "n_fig": 6, "n_tab": 6, "L_tab": 3072, "n_element_tab": 256, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2232, "n_element_tab_1": 226, "formula_len_all": 575, "formula_len_all_1": 544, "len_all": 126791, "len_all_1": 61654, "len_abs": 1376, "len_title": 131, "len_sents": 30967, "len_sents_1": 25988, "n_sents": 230, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1383, "title": "E3D: Enhancing Sparsely-Supervised 3D Object Detector with Large Multimodal Models", "abs": "Recently, sparsely-supervised 3D object detection has gained great attention, achieving performance close to that of fully-supervised 3D objectors with only a few annotated instances. Nevertheless, these methods suffer challenges when the accurate labels are extremely limited. In this paper, we propose an Ehanced 3D object Detection strategy, termed E3D, explicitly utilizing the prior knowledge from Large Multimodal Models (LMMs) to enhance the feature discrimination capability of the 3D detector under sparse annotation settings. Specifically, we first develop a Confident Points Semantic Transfer (CPST) module that generates high-quality seed points through boundary-constrained center cluster selection. Based on these seed points, we introduce a Dynamic Cluster Pseudo-label Generation (DCPG) module that yields pseudo-supervision signals from the geometry shape of multi-scale neighbor points. Additionally, we design a Distribution Shape score (DS score) that chooses high-quality supervision signals for the initial training of the 3D detector. By utilizing E3D, existing leading sparsely-supervised CoIn++ is improved by an average of 11.63% under the annotation rate of 2%. Moreover, we have verified our E3D in the zero-shot setting, and the results demonstrate its performance exceeding that of the state-of-the-art methods. The code will be made publicly available.", "keywords": [ "sparsely-supervised object detection", "point cloud", "large multimodal model" ], "rating_list": [ 1, 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 1, 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "Nx4PMtJ1ER", "primary_area": "causal reasoning", "n_author": 6, "author_info_list": [ { "name": "Cecilia Casolo", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Cristopher Salvi", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Assistant Professor" }, { "name": "Emilio Ferrucci", "gender": "Not Specified", "institution": "University of Oxford", "country": "", "position": "Postdoc" }, { "name": "Georg Manten", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Niki Kilbertus", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Assistant Professor" }, { "name": "S\u00f8ren Wengel Mogensen", "gender": "Male", "institution": "Lund University / Lund Institute of Technology", "country": "", "position": "Postdoc" } ], "n_formula": 60, "n_formula_1": 25, "n_ref_uni": 88, "n_ref": 174, "n_ref_all": 245, "n_fig": 7, "n_tab": 6, "L_tab": 1745, "n_element_tab": 168, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1400, "n_element_tab_1": 146, "formula_len_all": 17659, "formula_len_all_1": 9082, "len_all": 318821, "len_all_1": 89919, "len_abs": 1168, "len_title": 141, "len_sents": 86136, "len_sents_1": 35413, "n_sents": 635, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 92, "L_abs": 1171, "title": "Signature Kernel Conditional Independence Tests in Causal Discovery for Stochastic Processes", "abs": "Inferring the causal structure underlying stochastic dynamical systems from observational data holds great promise in domains ranging from science and health to finance. Such processes can often be accurately modeled via stochastic differential equations (SDEs), which naturally imply causal relationships via `which variables enter the differential of which other variables'. In this paper, we develop conditional independence (CI) constraints on coordinate processes over selected intervals that are Markov with respect to the acyclic dependence graph (allowing self-loops) induced by a general SDE model. We then provide a sound and complete causal discovery algorithm, capable of handling both fully and partially observed data, and uniquely recovering the underlying or induced ancestral graph by exploiting time directionality assuming a CI oracle. Finally, to make our algorithm practically usable, we also propose a flexible, consistent signature kernel-based CI test to infer these constraints from data. We extensively benchmark the CI test in isolation and as part of our causal discovery algorithms, outperforming existing approaches in SDE models and beyond.", "keywords": [ "causality", "dynamical systems", "stochastic processes", "causal discovery", "signature kernel" ], "rating_list": [ 8, 10, 6, 8 ], "soundness_list": [ 3, 4, 2, 4 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "Nx1XZWcLcW", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Akshat Shrivastava", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Armen Aghajanyan", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Research Scientist" }, { "name": "Gargi Ghosh", "gender": "Female", "institution": "Meta AI", "country": "", "position": "Researcher" }, { "name": "Liang Luo", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Luke Zettlemoyer", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Full Professor" }, { "name": "Mike Lewis", "gender": "Male", "institution": "Facebook AI Research", "country": "", "position": "Research Scientist" }, { "name": "Srini Iyer", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Principal Researcher" }, { "name": "Xi Victoria Lin", "gender": "Female", "institution": "Meta", "country": "US", "position": "Research Scientist" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 41, "n_ref": 92, "n_ref_all": 115, "n_fig": 8, "n_tab": 7, "L_tab": 1437, "n_element_tab": 225, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 221, "n_element_tab_1": 31, "formula_len_all": 318, "formula_len_all_1": 257, "len_all": 165911, "len_all_1": 56989, "len_abs": 1161, "len_title": 128, "len_sents": 41179, "len_sents_1": 27349, "n_sents": 291, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1167, "title": "MoMa: Efficient Early-Fusion Pre-training with Mixture of Modality-Aware Experts", "abs": "We introduce MoMa, a novel modality-aware mixture-of-experts (MoE) architecture designed for pre-training mixed-modal, early-fusion language models. MoMa processes images and text in arbitrary sequences by dividing expert modules into modality-specific groups. These groups exclusively process designated tokens while employing learned routing within each group to maintain semantics-based adaptivity. Our empirical results reveal substantial pre-training efficiency gains through this modality-specific parameter allocation. Under a 1-trillion-token training budget, the MoMa 1.4B model, featuring 4 text experts and 4 image experts, achieves impressive FLOPs savings: 3.7x overall, with 2.6x for text and 5.2x for image compared to a compute-equivalent dense baseline, measured by pre-training loss. This outperforms the standard expert-choice MoE with 8 mixed-modal experts, which achieves 3x overall FLOPs savings (3x for text, 2.8x for image). These results demonstrate MoMa's potential to significantly advance the efficiency of mixed-modal, early-fusion language model pre-training, paving the way for more resource-efficient and capable multimodal AI systems.", "keywords": [ "mutli-modal language modeling", "mixture-of-experts", "pre-training" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "NwYya2nwf3", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Juinn-Dar Huang", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" }, { "name": "Yi-Zeng Fang", "gender": "unknown", "institution": "NYCU", "country": "TW", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 18, "n_ref": 32, "n_ref_all": 45, "n_fig": 9, "n_tab": 36, "L_tab": 10133, "n_element_tab": 1352, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 3140, "n_element_tab_1": 431, "formula_len_all": 674, "formula_len_all_1": 206, "len_all": 171751, "len_all_1": 65139, "len_abs": 1494, "len_title": 163, "len_sents": 42603, "len_sents_1": 26944, "n_sents": 312, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 115, "L_abs": 1501, "title": "ExpertZIP: A Progressive Fusion Framework for Mixture-of-Experts Model Optimization through Huffman Tree Structures", "abs": "Mixture-of-Experts (MoE) models have gained attention as a novel approach to developing large language models (LLMs), praised for their ability to enhance performance by utilizing multiple experts. However, while increasing the number of experts in these models can yield performance gains, it also introduces significant trade-offs, such as substantial memory overhead and increased inference time, limiting their scalability and practical deployment. In this work, we conduct a thorough analysis of expert utilization and identify inefficiency: many experts are underutilized, leading to suboptimal resource allocation with limited improvement. To address this issue, we propose ExpertZIP, a progressive framework for MoE models that leverages a Huffman tree-based expert fusion technique. This progressive approach systematically merges underutilized experts step by step, ensuring their essential contributions are maintained while drastically reducing memory usage and computational demands. Our approach yields a 17.23x reduction in model size and a 4.84x improvement in inference time, with only a 1.18\\% decrease in average accuracy compared to the original 64-expert Switch Transformer model. Moreover, it demonstrates a 6.47\\% increase in accuracy relative to models with an equivalent number of experts. These results demonstrate that our optimized framework provides performance on par with larger models, offering an efficient solution for resource-constrained and real-time applications.", "keywords": [ "Large language models", "Mixture-of-Experts", "Expert Fusion" ], "rating_list": [ 6, 6, 5, 3, 6 ], "soundness_list": [ 2, 4, 2, 2, 3 ], "presentation_list": [ 3, 4, 3, 1, 3 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 5, 3, 4 ] }, { "paper_id": "NwQfwm3tHf", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Gang Yan", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Full Professor" }, { "name": "Sikai Yang", "gender": "unknown", "institution": "University of California, Merced", "country": "US", "position": "PhD student" }, { "name": "Wan Du", "gender": "Male", "institution": "University of California, Merced", "country": "", "position": "Assistant Professor" } ], "n_formula": 39, "n_formula_1": 26, "n_ref_uni": 34, "n_ref": 57, "n_ref_all": 70, "n_fig": 4, "n_tab": 4, "L_tab": 1496, "n_element_tab": 65, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1857, "n_element_tab_1": 137, "formula_len_all": 12684, "formula_len_all_1": 10776, "len_all": 158672, "len_all_1": 84487, "len_abs": 1317, "len_title": 157, "len_sents": 39680, "len_sents_1": 33648, "n_sents": 267, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 109, "L_abs": 1327, "title": "When Large Models Meet Generalized Linear Models: Hierarchy Statistical Network for Secure Federated Learning", "abs": "Large pre-trained models perform well on many Federated Learning (FL) tasks. Recent studies have revealed that fine-tuning only the final layer of large pre-trained models can reduce computational and communication costs while maintaining high performance. We can model the final layer, which typically performs a linear transformation, as a Generalized Linear Model (GLM). GLMs offer advantages in statistical modeling, especially for anomaly detection. Leveraging these advantages, GLM-based methods can be utilized to enhance the security of the fine-tuning process for large pre-trained models. However, integrating GLMs with large pre-trained models in FL presents challenges. GLMs rely on linear decision boundaries and struggle with the complex feature representation spaces from pre-trained models. To address this, we introduce the Hierarchy Statistical Network (HStat-Net). HStat-Net refines the spaces to make them more discriminative, allowing GLMs to work effectively in FL. Based on HStat-Net, we further develop FedRACE to detect poisoning attacks using deviance residuals from GLMs. We also provide a theorem to support FedRACE\u2019s detection. Extensive experiments conducted on CIFAR-100, Food-101, and Tiny ImageNet demonstrate that FedRACE significantly outperforms existing state-of-the-art defense algorithms.", "keywords": [ "Federated Learning", "Large Pre-trained Models", "Generalized Linear Models", "Security in Federated Learning", "Poisoning Attacks", "Deviance Residuals" ], "rating_list": [ 3, 5, 5, 5, 6 ], "soundness_list": [ 1, 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3, 3 ] }, { "paper_id": "Nw7i9Gd1WU", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Nuoya Xiong", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Siyu Chen", "gender": "Male", "institution": "Yale University", "country": "", "position": "PhD student" }, { "name": "Yusong Zhu", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Zhuoran Yang", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 144, "n_formula_1": 35, "n_ref_uni": 33, "n_ref": 78, "n_ref_all": 81, "n_fig": 0, "n_tab": 6, "L_tab": 765, "n_element_tab": 220, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 51, "n_element_tab_1": 6, "formula_len_all": 21824, "formula_len_all_1": 4461, "len_all": 194866, "len_all_1": 77181, "len_abs": 1229, "len_title": 93, "len_sents": 49429, "len_sents_1": 31397, "n_sents": 486, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1244, "title": "Exploration in the Face of Strategic Responses: Provable Learning of Online Stackelberg Games", "abs": "We study online leader-follower games where the leader interacts with a myopic follower using a quantal response policy. The leader's objective is to design an algorithm without prior knowledge of her reward function or the state transition dynamics. \nCrucially, the leader also lacks insight into the follower's reward function and realized rewards, posing a significant challenge. \nTo address this, the leader must learn the follower's quantal response mapping solely through strategic interactions --- announcing policies and observing responses. \nWe introduce a unified algorithm, Planning after Estimation, which updates the leader's policies in a two-step approach. \nIn particular, we first jointly estimate the leader's value function and the follower's response mapping by maximizing a sum of the Bellman error of the value function, the likelihood of the quantal response model, and a regularization term that encourages exploration. The leader's policy is then updated through a greedy planning step based on these estimates. Our algorithm achieves a $\\sqrt{T}$-regret in the context of general function approximation. \nMoroever, this algorithm avoids the intractable optimistic planning and thus enhances implementation simplicity.", "keywords": [ "Stackelberg game", "RL", "exploration" ], "rating_list": [ 5, 6, 6, 5, 3 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3, 3 ] }, { "paper_id": "Nvw2szDdmI", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Atsushi Nitanda", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Kazusato Oko", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Ryotaro Kawata", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "MS student" }, { "name": "Taiji Suzuki", "gender": "Male", "institution": "The University of Tokyo", "country": "", "position": "Full Professor" } ], "n_formula": 229, "n_formula_1": 27, "n_ref_uni": 42, "n_ref": 114, "n_ref_all": 168, "n_fig": 8, "n_tab": 4, "L_tab": 327, "n_element_tab": 36, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 365, "n_element_tab_1": 33, "formula_len_all": 31380, "formula_len_all_1": 2799, "len_all": 293428, "len_all_1": 71059, "len_abs": 1083, "len_title": 126, "len_sents": 71707, "len_sents_1": 28854, "n_sents": 851, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1088, "title": "Direct Distributional Optimization for Provable Alignment of Diffusion Models", "abs": "We introduce a novel alignment method for diffusion models from distribution optimization perspectives while providing rigorous convergence guarantees.\nWe first formulate the problem as a generic regularized loss minimization over probability distributions and directly optimize the distribution using the Dual Averaging method.\nNext, we enable sampling from the learned distribution by approximating its score function via Doob's $h$-transform technique.\nThe proposed framework is supported by rigorous convergence guarantees and an end-to-end bound on the sampling error, which imply that when the original distribution's score is known accurately, the complexity of sampling from shifted distributions is independent of isoperimetric conditions.\nThis framework is broadly applicable to general distribution optimization problems, including alignment tasks in Reinforcement Learning with Human Feedback (RLHF), Direct Preference Optimization (DPO), and Kahneman-Tversky Optimization (KTO). We empirically validate its performance on synthetic and image datasets using the DPO objective.", "keywords": [ "Diffusion models", "Optimization" ], "rating_list": [ 6, 8, 5, 8, 8 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 3, 4, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 2, 2 ] }, { "paper_id": "NvaZn3uwzJ", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Jason D. Lee", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Ruosong Wang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Assistant Professor" }, { "name": "Simon Shaolei Du", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "Yuxin Chen", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Associate Professor" }, { "name": "Zihan Zhang", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Postdoc" }, { "name": "lin Yang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "Assistant Professor" } ], "n_formula": 114, "n_formula_1": 11, "n_ref_uni": 27, "n_ref": 83, "n_ref_all": 85, "n_fig": 0, "n_tab": 5, "L_tab": 376, "n_element_tab": 33, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1246, "n_element_tab_1": 11, "formula_len_all": 18045, "formula_len_all_1": 2193, "len_all": 168795, "len_all_1": 67812, "len_abs": 1482, "len_title": 128, "len_sents": 45612, "len_sents_1": 30472, "n_sents": 476, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1490, "title": "Deployment Efficient Reward-Free Exploration with Linear Function Approximation", "abs": "We study deployment efficient reward-free exploration with linear function approximation, where the goal is to explore a linear Markov Decision Process (MDP) without revealing the reward function, while minimizing the number of exploration policies used during the algorithm. We design a new reinforcement learning (RL) algorithm whose sample complexity is polynomial in the feature dimension and horizon length, while achieving nearly optimal deployment efficiency for linear MDPs under the reward-free exploration setting. More specifically, our algorithm explores a linear MDP in a reward-free manner, while using at most $H$ exploration policies during its execution where $H$ is the horizon length. Compared to previous algorithms with similar deployment efficiency guarantees, the sample complexity of our algorithm does not depend on the reachability coefficient or the explorability coefficient of the underlying MDP, which can be arbitrarily small for certain MDPs. Our result addresses an open problem proposed in prior work. To achieve such a result, we show how to truncate state-action pairs of the underlying linear MDP in a data-dependent manner, and devise efficient offline policy evaluation and offline policy optimization algorithms in the truncated linear MDP. We further show how to implement reward-free exploration mechanisms in the linear function approximation setting by carefully combines these offline RL algorithms without sacrificing the deployment efficiency.", "keywords": [ "Linear MDP", "Deployment Complexity", "Sample Complexity" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 1, 3 ] }, { "paper_id": "NvRVYVN106", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Bogdan Cautis", "gender": "Male", "institution": "Universit\u00e9 Paris Saclay", "country": "FR", "position": "Full Professor" }, { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Keke Huang", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Postdoc" }, { "name": "Ruize Gao", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Xiaokui Xiao", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 45, "n_formula_1": 13, "n_ref_uni": 80, "n_ref": 300, "n_ref_all": 350, "n_fig": 17, "n_tab": 7, "L_tab": 2180, "n_element_tab": 156, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 216, "n_element_tab_1": 6, "formula_len_all": 4080, "formula_len_all_1": 863, "len_all": 316092, "len_all_1": 68634, "len_abs": 1562, "len_title": 107, "len_sents": 106798, "len_sents_1": 33672, "n_sents": 749, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1572, "title": "Privacy Breach Detection by Non-Parametric Two-Sample Tests", "abs": "With the proliferation of machine learning services, the risk of privacy breaches has never been higher, owing to the need for collecting -- sometimes by any means necessary -- valuable, yet sensitive training data. When an unsanctioned data access occurs, it may become apparent after the fact, in the predictive models that have been trained on compromised data. This calls for effective membership inference methods, enabling an evaluator to identify privacy breaches. Distinct from traditional membership inference attacks (MIAs), which focus on determining whether individual data records were used in training, this study centers on the evaluation of sets of records, particularly when only a small proportion of the set are training members. In this scenario, traditional MIAs often suffer from non-ideal evaluation reliability. To address this issue, from a privacy evaluator's perspective, we propose a novel approach for membership inference, applicable not to individual records but to sets thereof. It relies on a non-parametric two-sample test, which leverages the differences between high-level representation to infer membership. Based on extensive experiments, our proposed High-level Representation-based MMD (HR-MMD) test exhibits high sensitivity in distinguishing between the training and non-training sets, with ideal type I error, making it a powerful membership detection tool. Our study offers insights into an alternative privacy breach detection scenario and opens up a promising avenue for privacy evaluation based on membership inference tests.", "keywords": [ "Privacy breach detection", "Two-Sample tests", "membership inference" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 2, 2, 3, 3 ] }, { "paper_id": "NvDRvtrGLo", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Guy Pelc", "gender": "unknown", "institution": "Hebrew University of Jerusalem", "country": "", "position": "PhD student" }, { "name": "Matt Ricci", "gender": "Male", "institution": ", Hebrew University of Jerusalem", "country": "", "position": "Postdoc" }, { "name": "Mor Nitzan", "gender": "unknown", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "Assistant Professor" }, { "name": "Noa Moriel", "gender": "Female", "institution": "Hebrew University of Jerusalem", "country": "", "position": "PhD student" }, { "name": "Zoe Piran", "gender": "Female", "institution": "Genentech", "country": "US", "position": "Postdoc" } ], "n_formula": 27, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 66, "n_ref_all": 89, "n_fig": 6, "n_tab": 1, "L_tab": 411, "n_element_tab": 35, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1355, "formula_len_all_1": 424, "len_all": 159196, "len_all_1": 60029, "len_abs": 1969, "len_title": 60, "len_sents": 56989, "len_sents_1": 29588, "n_sents": 469, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1979, "title": "TRENDy: Temporal Regression of Effective Nonlinear Dynamics", "abs": "Spatiotemporal dynamics pervade the natural sciences, from the morphogen dynamics underlying patterning in animal pigmentation to the protein waves controlling cell division. A central challenge lies in understanding how controllable parameters induce qualitative changes in system behavior called bifurcations. This endeavor is made particularly difficult in realistic settings where governing partial differential equations (PDEs) are unknown and data is limited and noisy. To address this challenge, we propose TRENDy (Temporal Regression of Effective Nonlinear Dynamics), an equation-free approach to learning low-dimensional, predictive models of spatiotemporal dynamics. Following classical work in spatial coarse-graining, TRENDy first maps input data to a low-dimensional space of effective dynamics through a cascade of multiscale filtering operations. Our key insight is the recognition that these effective dynamics can be fit by a neural ordinary differential equation (NODE) having the same parameter space as the input PDE. The preceding filtering operations strongly regularize the phase space of the NODE, making TRENDy significantly more robust to noise compared to existing methods. We train TRENDy to predict the effective dynamics of synthetic and real data representing dynamics from across the physical and life sciences. We then demonstrate how our framework can automatically locate both Turing and Hopf bifurcations in unseen regions of parameter space. We finally apply our method to the analysis of spatial patterning of the ocellated lizard through development. We found that TRENDy's predicted effective state not only accurately predicts spatial changes over time but also identifies distinct pattern features unique to different anatomical regions, such as the tail, neck, and body\u2014an insight that highlights the potential influence of surface geometry on reaction-diffusion mechanisms and their role in driving spatially varying pattern dynamics.", "keywords": [ "dynamical systems; neural ODEs", "representation learning" ], "rating_list": [ 6, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "Nue5iMj8n6", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Chao Qu", "gender": "Male", "institution": "Inftech", "country": "CN", "position": "Researcher" }, { "name": "Fenglei Cao", "gender": "Male", "institution": "inf tech", "country": "", "position": "Researcher" }, { "name": "Junyi An", "gender": "Male", "institution": "Shanghai Academy of Artificial Intelligence for Science", "country": "CN", "position": "Researcher" }, { "name": "Qianwei Tang", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Yuan Qi", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yun-Fei Shi", "gender": "Male", "institution": "Shanghai Academy of Artificial Intelligence for Science", "country": "CN", "position": "Researcher" }, { "name": "xinhao liu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 34, "n_formula_1": 25, "n_ref_uni": 38, "n_ref": 76, "n_ref_all": 118, "n_fig": 8, "n_tab": 10, "L_tab": 2478, "n_element_tab": 484, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 877, "n_element_tab_1": 184, "formula_len_all": 2290, "formula_len_all_1": 1173, "len_all": 174973, "len_all_1": 69725, "len_abs": 1097, "len_title": 126, "len_sents": 56829, "len_sents_1": 30663, "n_sents": 480, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1072, "title": "Equivariant Masked Position Prediction for Efficient Molecular Representation", "abs": "Graph neural networks (GNNs) have shown considerable promise in computational chemistry. However, the limited availability of molecular data raises concerns regarding GNNs' ability to effectively capture the fundamental principles of physics and chemistry, which constrains their generalization capabilities. To address this challenge, we introduce a novel self-supervised approach termed Equivariant Masked Position Prediction (EMPP), grounded in intramolecular potential and force theory. Unlike conventional attribute masking techniques, EMPP formulates a nuanced position prediction task that is more well-defined and enhances the learning of quantum mechanical features. EMPP also bypasses the approximation of the Gaussian mixture distribution commonly used in denoising methods, allowing for more accurate acquisition of physical properties. Experimental results indicate that EMPP significantly enhances performance of advanced molecular architectures, surpassing state-of-the-art self-supervised approaches. Our code is released in https://github.com/ajy112/EMPP.", "keywords": [ "Self-supervised Learning; Graph Neural Network; Molecular Property Prediction;" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 2, 4, 4 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "NuWX55CpIQ", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Charles Fleming", "gender": "unknown", "institution": "Cisco", "country": "US", "position": "Researcher" }, { "name": "Gaowen Liu", "gender": "Female", "institution": "Cisco Systems", "country": "US", "position": "Researcher" }, { "name": "Ramana Rao Kompella", "gender": "Male", "institution": "Cisco", "country": "US", "position": "Researcher" }, { "name": "Shaolei Ren", "gender": "Not Specified", "institution": "University of California, Riverside", "country": "US", "position": "Associate Professor" }, { "name": "Shijin Duan", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "PhD student" }, { "name": "Xiaolin Xu", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 8, "n_ref_uni": 26, "n_ref": 45, "n_ref_all": 73, "n_fig": 9, "n_tab": 7, "L_tab": 1021, "n_element_tab": 149, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1479, "n_element_tab_1": 142, "formula_len_all": 1906, "formula_len_all_1": 676, "len_all": 164682, "len_all_1": 68637, "len_abs": 1263, "len_title": 127, "len_sents": 58379, "len_sents_1": 31183, "n_sents": 434, "n_sents_1": 230, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1268, "title": "Robust Quantum Neural Networks Against Dynamic Noise Landscape in the NISQ Era", "abs": "Quantum machine learning, an emerging field in the noisy intermediate-scale quantum (NISQ) era, faces significant challenges in error mitigation during training and inference stages. Current noise-aware training (NAT) methods typically assume static error rates in quantum neural networks (QNNs), often neglecting the inherently dynamic nature of such noise. By addressing this oversight, our work recognizes the dynamics of noise in the NISQ era, evidenced by fluctuating error rates across different times and qubits. Moreover, QNN performance can vary markedly depending on the specific locations of errors, even under similar error rates. This variability underscores the limitations of static NAT strategies in addressing the dynamic nature of noisy environments. We propose a novel NAT strategy that adapts to both standard and fatal error conditions, cooperating with a low-complexity search strategy to efficiently locate fatal errors during optimization. Our approach marks a significant advancement over current NAT methods by maintaining robust performance in fatal error scenarios. Evaluations validate the efficacy of our strategy against fatal errors, while maintaining performance comparable to state-of-the-art NAT approaches under various error rates.", "keywords": [ "Quantum Neural Network", "Noise-Aware Training", "Dynamic Noise" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "NuVBI4wPMm", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Jing Tang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Pengwen Dai", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaochun Cao", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Yifan Song", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Yihong Luo", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CX", "position": "PhD student" }, { "name": "Yuhan Chen", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 20, "n_ref_uni": 37, "n_ref": 83, "n_ref_all": 95, "n_fig": 4, "n_tab": 12, "L_tab": 15415, "n_element_tab": 2649, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 4821, "n_element_tab_1": 632, "formula_len_all": 2786, "formula_len_all_1": 1578, "len_all": 194284, "len_all_1": 80175, "len_abs": 2474, "len_title": 145, "len_sents": 44818, "len_sents_1": 30510, "n_sents": 336, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 2566, "title": "Decoupled Graph Energy-based Model for Node Out-of-Distribution Detection on Heterophilic Graphs", "abs": "Despite extensive research efforts focused on Out-of-Distribution (OOD) detection on images, OOD detection on nodes in graph learning remains underexplored. The dependence among graph nodes hinders the trivial adaptation of existing approaches on images that assume inputs to be i.i.d. sampled, since many unique features and challenges specific to graphs are not considered, such as the heterophily issue. Recently, GNNSafe, which considers node dependence, adapted energy-based detection to the graph domain with state-of-the-art performance, however, it has two serious issues: 1) it derives node energy from classification logits without specifically tailored training for modeling data distribution, making it less effective at recognizing OOD data; 2) it highly relies on energy propagation, which is based on homophily assumption and will cause significant performance degradation on heterophilic graphs, where the node tends to have dissimilar distribution with its neighbors. To address the above issues, we suggest training Energy-based Models (EBMs) by Maximum Likelihood Estimation (MLE) to enhance data distribution modeling and removing energy propagation to overcome the heterophily issues. However, training EBMs via MLE requires performing Markov Chain Monte Carlo (MCMC) sampling on both node feature and node neighbors, which is challenging due to the node interdependence and discrete graph topology. To tackle the sampling challenge, we introduce Decoupled Graph Energy-based Model (DeGEM), which decomposes the learning process into two parts\u2014a graph encoder that leverages topology information for node representations and an energy head that operates in latent space. Additionally, we propose a Multi-Hop Graph encoder (MH) and Energy Readout (ERo) to enhance node representation learning, Conditional Energy (CE) for improved EBM training, and Recurrent Update for the graph encoder and energy head to promote each other. This approach avoids sampling adjacency matrices and removes the need for energy propagation to extract graph topology information. Extensive experiments validate that DeGEM, without OOD exposure during training, surpasses previous state-of-the-art methods, achieving an average AUROC improvement of 6.71% on *homophilic* graphs and 20.29% on *heterophilic* graphs, and even outperform methods trained with OOD exposure. Our code is available at the anonymous link: [https://anonymous.4open.science/r/DeGEM\\_ICLR2025\\_rebuttal-B801/README.md](https://anonymous.4open.science/r/DeGEM\\_ICLR2025\\_rebuttal-B801/README.md).", "keywords": [ "Node OOD detection; Energy-based Models; Graph Neural Network" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "NuHYh4YKNe", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Di Huang", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Junyi Chen", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Tong He", "gender": "Male", "institution": "Shanghai AI lab", "country": "CN", "position": "Researcher" }, { "name": "Wanli Ouyang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Full Professor" }, { "name": "Weicai Ye", "gender": "Male", "institution": "KwaiVGI, Kuaishou Technology", "country": "CN", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 31, "n_ref": 59, "n_ref_all": 71, "n_fig": 6, "n_tab": 2, "L_tab": 345, "n_element_tab": 48, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 363, "n_element_tab_1": 51, "formula_len_all": 180, "formula_len_all_1": 119, "len_all": 92387, "len_all_1": 46642, "len_abs": 286, "len_title": 102, "len_sents": 27616, "len_sents_1": 21930, "n_sents": 193, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 101, "L_abs": 1388, "title": "Where Am I and What Will I See: An Auto-Regressive Model for Spatial Localization and View Prediction", "abs": "Spatial intelligence is the ability of a machine to perceive, reason, and act in three dimensions within space and time.\nRecent advancements in large-scale auto-regressive models have demonstrated remarkable capabilities across various reasoning tasks. However, these models often struggle with fundamental aspects of spatial reasoning, particularly in answering questions like \"Where am I?\" and \"What will I see?\". While some attempts have been done, existing approaches typically treat them as separate tasks, failing to capture their interconnected nature. In this paper, we present **G**enerative **S**patial **T**ransformer (GST), a novel auto-regressive framework that jointly addresses spatial localization and view prediction. Our model simultaneously estimates the camera pose from a single image and predicts the view from a new camera pose, effectively bridging the gap between spatial awareness and visual prediction. The proposed innovative camera tokenization method enables the model to learn the joint distribution of 2D projections and their corresponding spatial perspectives in an auto-regressive manner. This unified training paradigm demonstrates that joint optimization of pose estimation and novel view synthesis leads to improved performance in both tasks, for the first time, highlighting the inherent relationship between spatial awareness and visual prediction.", "keywords": [ "Generative Models", "Novel View Synthesis", "Camera Pose Estimation" ], "rating_list": [ 8, 6, 8, 3 ], "soundness_list": [ 3, 2, 3, 1 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Nu8b9C1xcr", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Carsten Binnig", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Chelsea Lin", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Fatma Ozcan", "gender": "Female", "institution": "Google", "country": "", "position": "Principal Researcher" }, { "name": "Sami Abu-El-Haija", "gender": "Male", "institution": "Research, Google", "country": "", "position": "Research Scientist" }, { "name": "Yannis Chronis", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Yawen Wang", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Yu Gan", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 10, "n_ref": 19, "n_ref_all": 43, "n_fig": 6, "n_tab": 6, "L_tab": 3523, "n_element_tab": 502, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2860, "n_element_tab_1": 308, "formula_len_all": 37, "formula_len_all_1": 37, "len_all": 104549, "len_all_1": 65563, "len_abs": 1604, "len_title": 130, "len_sents": 33450, "len_sents_1": 28990, "n_sents": 237, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1598, "title": "CardBench: A Benchmark for Learned Cardinality Estimation in Relational Databases", "abs": "Cardinality estimation is crucial for enabling high query performance in relational\ndatabases. Recently learned cardinality estimation models have been proposed\nto improve accuracy but there is no systematic benchmark or datasets which\nallows researchers to evaluate the progress made by new learned approaches\nand even systematically develop new learned approaches. In this paper, we are\nreleasing a benchmark, containing thousands of queries over 20 distinct real-world\ndatabases for learned cardinality estimation. In contrast to other initial benchmarks,\nour benchmark is much more diverse and can be used for training and testing\nlearned models systematically. Using this benchmark, we explored whether learned\ncardinality estimation can be transferred to an unseen dataset in a zero-shot manner.\nWe trained GNN-based and transformer-based models to study the problem in three\nsetups: 1-) instance-based, 2-) zero-shot, and 3-) fine-tuned.\nOur results show that while we get promising results for zero-shot cardinality estimation on simple single table queries; as soon as we add joins, the accuracy drops.\nHowever, we show that with fine-tuning, we can still utilize pre-trained models\nfor cardinality estimation, significantly reducing training overheads compared to\ninstance specific models. We are open sourcing our scripts to collect statistics,\ngenerate queries and training datasets to foster more extensive research, also from\nthe ML community on the important problem of cardinality estimation and in\nparticular improve on recent directions such as pre-trained cardinality estimation.", "keywords": [ "cardinality estimation", "zero-shot", "fine tuning", "databases" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "NtwFghsJne", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Lokesh Nagalapatti", "gender": "unknown", "institution": "Indian Institute of Technology, Bombay", "country": "", "position": "PhD student" }, { "name": "Prateek Garg", "gender": "Male", "institution": "Indian Institute of Technology Bombay", "country": "IN", "position": "Undergrad student" }, { "name": "Sunita Sarawagi", "gender": "Female", "institution": "IIT Bombay", "country": "IN", "position": "Full Professor" } ], "n_formula": 23, "n_formula_1": 19, "n_ref_uni": 34, "n_ref": 61, "n_ref_all": 80, "n_fig": 4, "n_tab": 10, "L_tab": 1926, "n_element_tab": 196, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 844, "n_element_tab_1": 32, "formula_len_all": 2339, "formula_len_all_1": 1268, "len_all": 147843, "len_all_1": 68238, "len_abs": 1370, "len_title": 122, "len_sents": 40107, "len_sents_1": 29134, "n_sents": 474, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1376, "title": "From Search to Sampling: Generative Models for Robust Algorithmic Recourse", "abs": "Algorithmic Recourse provides recommendations to individuals who are adversely impacted by automated model decisions, on how to alter their profiles to achieve a favorable outcome. Effective recourse methods must balance three conflicting goals: proximity to the original profile to minimize cost, plausibility for realistic recourse, and validity to ensure the desired outcome. We show that existing methods train for these objectives separately and then search for recourse through a joint optimization over the recourse goals during inference, leading to poor recourse recommendations. We introduce GenRe, a generative recourse model designed to train the three recourse objectives jointly. Training such generative models is non-trivial due to lack of direct recourse supervision. We propose efficient ways to synthesize such supervision and further show that GenRe's training leads to a consistent estimator. Unlike most prior methods, that employ non-robust gradient descent based search during inference, GenRe simply performs a forward sampling over the generative model to produce minimum cost recourse, leading to superior performance across multiple metrics. We also demonstrate GenRe provides the best trade-off between cost, plausibility and validity, compared to state-of-art baselines. We release anonymized code at: https://anonymous.4open.science/r/GenRe-BD71", "keywords": [ "Algorithmic recourse", "explainability", "generative modelling" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "NtSlKEJ2DS", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Kaipeng Zeng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Liangliang Shi", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Yihui Tu", "gender": "Male", "institution": "Shanghai University", "country": "CN", "position": "Lecturer" }, { "name": "Yufeng Li", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "", "position": "Undergrad student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 32, "n_formula_1": 20, "n_ref_uni": 30, "n_ref": 46, "n_ref_all": 66, "n_fig": 6, "n_tab": 8, "L_tab": 2793, "n_element_tab": 257, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 824, "n_element_tab_1": 79, "formula_len_all": 1803, "formula_len_all_1": 1065, "len_all": 126222, "len_all_1": 67922, "len_abs": 1699, "len_title": 176, "len_sents": 41009, "len_sents_1": 30755, "n_sents": 326, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 127, "L_abs": 1694, "title": "Optimal Flow Transport and its Entropic Regularization: a GPU-friendly Matrix Iterative Algorithm for Flow Balance Satisfaction", "abs": "The Sinkhorn algorithm, based on Entropic Regularized Optimal Transport (OT), has garnered significant attention due to its computational efficiency enabled by GPU-friendly matrix-vector multiplications. However, vanilla OT primarily deals with computations between the source and target nodes in a bipartite graph, limiting its practical application in real-world transportation scenarios.\nIn this paper, we introduce the concept of Optimal Flow Transport (OFT) as an extension, where we consider a more general graph case and the marginal constraints in vanilla OT are replaced by flow balance constraints. \nTo obtain solutions, we incorporate entropic regularization into the OFT and introduce virtual flows for individual nodes to tackle the issue of potentially numerous isolated nodes lacking flow passages. Our proposition, the OFT-Sinkhorn algorithm, utilizes GPU-friendly matrix iterations to maintain flow balance constraints and minimize the objective function, and theoretical results for global convergence are also proposed in this paper.\nFurthermore, we enhance OFT by introducing capacity constraints on nodes and edges, transforming the OFT problem into a minimum-cost flow problem. We then present the Capacity-Constrained EOFT-Sinkhorn algorithm and compare it with the traditional Minimum cost flow (MCF) algorithm, showing that our algorithm is quite efficient for calculation. \nIn particular, our EOFT-Sinkhorn is evaluated on high-precision and integer-precision MCF problems with different scales from one hundred to five thousand size, exhibiting significant time efficiency and the ability to approximate optimal solutions. Source code will be made publicly available.", "keywords": [ "Optimal Transport", "Flow Balance Constraints", "Sinkhorn Algorithm", "Network Flow Theory" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "NtMf8DejbV", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Cao Xiao", "gender": "Female", "institution": "GE Healthcare", "country": "US", "position": "VP of AI" }, { "name": "Hua Wei", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" }, { "name": "Longchao Da", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Parminder Bhatia", "gender": "Male", "institution": "GEHC", "country": "US", "position": "Principal Researcher" }, { "name": "Rui Wang", "gender": "Male", "institution": "GE HealthCare", "country": "US", "position": "Researcher" }, { "name": "Taha Kass-Hout", "gender": "Not Specified", "institution": "GE HealthCare", "country": "US", "position": "Chief Science and Technology Officer" }, { "name": "Xiaojian Xu", "gender": "Female", "institution": "GE HealthCare", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 2, "n_ref_uni": 59, "n_ref": 95, "n_ref_all": 127, "n_fig": 15, "n_tab": 4, "L_tab": 1928, "n_element_tab": 219, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2424, "n_element_tab_1": 320, "formula_len_all": 491, "formula_len_all_1": 181, "len_all": 202229, "len_all_1": 73203, "len_abs": 1432, "len_title": 125, "len_sents": 58177, "len_sents_1": 34060, "n_sents": 363, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1436, "title": "Segment as You Wish: Free-Form Language-Based Segmentation for Medical Images", "abs": "Medical imaging is crucial for diagnosing a patient\u2019s health condition, and accurate segmentation of these images is essential for isolating regions of interest to ensure precise diagnosis and treatment planning. Existing methods primarily rely on bounding boxes or point-based prompts, while few have explored text-related prompts, despite clinicians often describing their observations and instructions in natural language. To address this gap, we first propose a RAG-based free-form text prompt generator, that leverages the domain corpus to generate diverse and realistic descriptions. Then, we introduce FLanS, a novel medical image segmentation model that handles various free-form text prompts, including professional anatomy-informed queries, anatomy-agnostic position-driven queries, or anatomy-agnostic size-driven queries. Additionally, our model also incorporates a symmetry-aware canonicalization module to ensure consistent, accurate segmentations across varying scan orientations and reduce confusion between the anatomical position of an organ and its appearance in the scan. FLanS is trained on a large-scale dataset of over 100k medical images from 7 public datasets. Comprehensive experiments demonstrate the model\u2019s superior language understanding and segmentation precision, along with a deep comprehension of the relationship between them, outperforming SOTA baselines on both in-domain and out-of-domain datasets.", "keywords": [ "Medical Image Segmentation", "Foundation Models", "Equivariance" ], "rating_list": [ 5, 1, 3, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "NtDXCDvkXJ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Bingxi Liu", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Hong Zhang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Jinqiang Cui", "gender": "Male", "institution": "Pengcheng Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Li He", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Pengju Zhang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Shiyi Guo", "gender": "unknown", "institution": "Northeastern University", "country": "CN", "position": "Lecturer" }, { "name": "Yihong Wu", "gender": "Female", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Hao Chen", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 47, "n_ref": 114, "n_ref_all": 121, "n_fig": 4, "n_tab": 10, "L_tab": 3883, "n_element_tab": 661, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 3588, "n_element_tab_1": 588, "formula_len_all": 265, "formula_len_all_1": 123, "len_all": 154105, "len_all_1": 77297, "len_abs": 2006, "len_title": 157, "len_sents": 35951, "len_sents_1": 30162, "n_sents": 275, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 108, "L_abs": 2024, "title": "The Renaissance of Classic Feature Aggregations for Visual Place Recognition in the Era of Foundation Models", "abs": "Visual Place Recognition (VPR) addresses the retrieval problem in large-scale geographic image databases through feature representations. Recent approaches have leveraged visual foundation models and have proposed novel feature aggregations. However, these methods have failed to grasp the core concepts of foundational models, such as leveraging extensive training sets, and have also neglected the potential of classical feature aggregations, such as GeM and NetVLAD, for low-dimensional representations. Building on these insights, we revive classic aggregation methods and create more fundamental VPR models, abbreviated SuperPlace. First, we introduce a supervised label alignment method that combines grid partitioning and local feature matching. This allows models to be trained on diverse VPR datasets within a unified framework, similar to the design principles of foundation models. Second, we introduce G$^2$M, a compact feature aggregation with two GeMs, in which one GeM learns the principal components of feature maps along the channel direction and calibrates the other GeM's output. Third, we propose the secondary fine-tuning (FT$^2$) strategy for NetVLAD-Linear (NVL). NetVLAD first learns feature vectors in a high-dimensional space and then compresses them into a low-dimensional space using a single linear layer. G$^2$M excels in large-scale applications requiring rapid response and low latency, while NVL-FT$^2$ is optimized for scenarios demanding high precision across a broad range of conditions. Extensive experiments (12 test sets, 14 previous methods, and 11 tables) highlight our contributions and demonstrate the superiority of SuperPlace. Specifically, SuperPlace-G$^2$M achieves state-of-the-art results with only one-tenth of the feature dimensions compared to recent methods. Moreover, SuperPlace-NVL-FT$^2$ holds the top rank on the MSLS challenge leaderboard. We have submitted a ranking screenshot, the source code, and the original experimental records in the supplementary materials.", "keywords": [ "Visual Place Recognition", "Feature Representation", "Supervised Learning" ], "rating_list": [ 5, 8, 3, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "NtAXAvIYuN", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Choong Seon Hong", "gender": "Male", "institution": "Kyung Hee University", "country": "KR", "position": "Full Professor" }, { "name": "Long Tan Le", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "PhD student" }, { "name": "Tung-Anh Nguyen", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "PhD student" }, { "name": "HAN SHU", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Nguyen Hoang Tran", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "Associate Professor" } ], "n_formula": 52, "n_formula_1": 32, "n_ref_uni": 43, "n_ref": 106, "n_ref_all": 124, "n_fig": 5, "n_tab": 7, "L_tab": 2047, "n_element_tab": 270, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 228, "n_element_tab_1": 6, "formula_len_all": 4575, "formula_len_all_1": 2277, "len_all": 204870, "len_all_1": 57600, "len_abs": 1381, "len_title": 130, "len_sents": 65332, "len_sents_1": 26430, "n_sents": 490, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1393, "title": "$i$REPO: $i$mplicit Reward Pairwise Difference based Empirical Preference Optimization", "abs": "While astonishingly capable, large Language Models (LLM) can sometimes produce outputs that deviate from human expectations. Such deviations necessitate an alignment phase to prevent disseminating untruthful, toxic, or biased information. Traditional alignment methods based on reinforcement learning often struggle with the identified instability, whereas preference optimization methods are limited by their overfitting to pre-collected hard-label datasets. In this paper, we propose a novel LLM alignment framework named $i$REPO, which utilizes implicit Reward pairwise difference regression for Empirical Preference Optimization. Particularly, $i$REPO employs self-generated datasets labeled by empirical human (or AI annotator) preference to iteratively refine the aligned policy through a novel regression-based loss function. Furthermore, we introduce an innovative algorithm backed by theoretical guarantees for achieving optimal results under ideal assumptions and providing a practical performance-gap result without such assumptions. Experimental results with Phi-2 and Mistral-7B demonstrate that $i$REPO effectively achieves self-alignment using soft-label, self-generated responses and the logit of empirical AI annotators. Furthermore, our approach surpasses preference optimization baselines in evaluations using the Language Model Evaluation Harness and Multi-turn benchmarks.", "keywords": [ "Language Models", "Preference Optimization" ], "rating_list": [ 3, 3, 5, 3, 3 ], "soundness_list": [ 2, 4, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "Nsms7NeU2x", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Sebastian Bordt", "gender": "unknown", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "Postdoc" }, { "name": "Suraj Srinivas", "gender": "Male", "institution": "Robert Bosch LLC", "country": "US", "position": "Researcher" }, { "name": "Ulrike von Luxburg", "gender": "Female", "institution": "University of Tuebingen", "country": "DE", "position": "Professor" }, { "name": "Valentyn Boreiko", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 2, "n_ref_uni": 65, "n_ref": 127, "n_ref_all": 165, "n_fig": 21, "n_tab": 1, "L_tab": 153, "n_element_tab": 17, "n_fig_1": 17, "n_tab_1": 1, "L_tab_1": 153, "n_element_tab_1": 17, "formula_len_all": 1146, "formula_len_all_1": 104, "len_all": 180212, "len_all_1": 66632, "len_abs": 1447, "len_title": 96, "len_sents": 47185, "len_sents_1": 31592, "n_sents": 369, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1454, "title": "How much can we Forget about Data Contamination?", "abs": "The leakage of benchmark data into the training data has emerged as a significant challenge for evaluating the capabilities of large language models (LLMs). In this work, we use experimental evidence and theoretical estimates to challenge the common assumption that small-scale contamination renders benchmark evaluations invalid. First, we experimentally quantify the magnitude of benchmark overfitting based on scaling along three dimensions: The number of model parameters (up to 1.6B), the number of times an example is seen (up to 144), and the number of training tokens (up to 40B). We find that if model and data follow the Chinchilla scaling laws, minor contamination indeed leads to overfitting. At the same time, even 144 times of contamination can be forgotten if the training data is scaled beyond five times Chinchilla, a regime characteristic of many modern LLMs. We then derive a simple theory of example forgetting via cumulative weight decay. It allows us to bound the number of gradient steps required to forget past data for any training run where we know the hyperparameters of AdamW. This indicates that many LLMs, including Llama 3, have forgotten the data seen at the beginning of training. Experimentally, we demonstrate that forgetting occurs faster than what is predicted by our bounds. Taken together, our results suggest that moderate amounts of contamination can be forgotten at the end of realistically scaled training runs.", "keywords": [ "Large Language Models", "Contamination", "Forgetting", "Scaling", "Optimization" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 2, 2 ] }, { "paper_id": "NsFZZU9gvk", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 11, "author_info_list": [ { "name": "Elaine Lau", "gender": "Female", "institution": "McGill University", "country": "", "position": "MS student" }, { "name": "Elaine T Chang", "gender": "Female", "institution": "Scale AI", "country": "US", "position": "Researcher" }, { "name": "Matt Fredrikson", "gender": "Male", "institution": "Gray Swan AI", "country": "US", "position": "Researcher" }, { "name": "Priyanshu Kumar", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Saranya Vijayakumar", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Sean M. Hendryx", "gender": "Male", "institution": "Scale AI", "country": "", "position": "Engineering Manager, Machine Learning" }, { "name": "Shuyan Zhou", "gender": "Non-Binary", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Summer Yue", "gender": "Female", "institution": "Scale AI", "country": "US", "position": "Researcher" }, { "name": "Tu Trinh", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" }, { "name": "Vaughn Robinson", "gender": "Male", "institution": "Scale AI", "country": "", "position": "Operations" }, { "name": "Zifan Wang", "gender": "Male", "institution": "Scale AI", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 50, "n_ref": 89, "n_ref_all": 117, "n_fig": 5, "n_tab": 4, "L_tab": 2042, "n_element_tab": 156, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 889, "n_element_tab_1": 130, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 162510, "len_all_1": 62533, "len_abs": 1441, "len_title": 91, "len_sents": 42028, "len_sents_1": 29593, "n_sents": 337, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 1465, "title": "Aligned LLMs Are Not Aligned Browser Agents", "abs": "Despite significant efforts spent by large language model (LLM) developers to\nalign model outputs towards safety and helpfulness, there remains an open ques-\ntion if this safety alignment, typically enforced in chats, generalize to non-chat\nand agentic use cases? Unlike chatbots, agents equipped with general-purpose\ntools, such as web browsers and mobile devices, can directly influence the real\nworld, making it even more crucial to ensure the safety of LLM agents. In\nthis work, we primarily focus on red-teaming browser agents, LLMs that in-\nteract with and extract information from web browsers. To this end, we in-\ntroduce Browser Agent Red teaming Toolkit (BrowserART), a comprehensive test\nsuite consisting of 100 diverse browser-related harmful behaviors and 40 syn-\nthetic websites, designed specifically for red-teaming browser agents. Our empir-\nical study on state-of-the-art browser agents reveals a significant alignment gap\nbetween the base LLMs and their downstream browser agents. That is, while\nthe LLM demonstrates alignment as a chatbot, the corresponding agent does not.\nMoreover, attack methods designed to jailbreak aligned LLMs in chat settings\ntransfer effectively to browser agents - with simple human rewrites, GPT-4o and\nGPT-4 Turbo -based browser agents attempted all 100 harmful behaviors. We plan to publicly release BrowserART and call on LLM developers, policymakers, and agent developers to collaborate on enhancing agent safety.", "keywords": [ "LLM", "agents", "red teaming", "safety", "adversarial robustness", "alignment", "jailbreak" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 4, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "Ns8zGZ0lmM", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Koki Wataoka", "gender": "Male", "institution": "SB Intuitions", "country": "JP", "position": "Researcher" }, { "name": "Ryokan Ri", "gender": "Male", "institution": "SB Intuitions", "country": "", "position": "Researcher" }, { "name": "Tsubasa Takahashi", "gender": "Male", "institution": "Turing Inc.", "country": "JP", "position": "Principal Researcher" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 26, "n_ref": 40, "n_ref_all": 62, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 236, "formula_len_all_1": 162, "len_all": 91491, "len_all_1": 53752, "len_abs": 1250, "len_title": 86, "len_sents": 27796, "len_sents_1": 26844, "n_sents": 209, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 38, "L_abs": 1260, "title": "Self-Preference Bias in LLM-as-a-Judge", "abs": "Automated evaluation leveraging large language models (LLMs), commonly referred to as LLM evaluators or LLM-as-a-judge, has been widely used in measuring the performance of dialogue systems.\nHowever, the self-preference bias in LLMs has posed significant risks, including promoting specific styles or policies intrinsic to the LLMs.\nDespite the importance of this issue, there is a lack of established methods to measure the self-preference bias quantitatively, and its underlying causes are poorly understood.\nIn this paper, we introduce a novel quantitative metric to measure the self-preference bias.\nOur experimental results demonstrate that GPT-4 exhibits a significant degree of self-preference bias.\nTo explore the causes, we hypothesize that LLMs may favor outputs that are more familiar to them, as indicated by lower perplexity.\nWe analyze the relationship between LLM evaluations and the perplexities of outputs.\nOur findings reveal that LLMs assign significantly higher evaluations to outputs with lower perplexity than human evaluators, regardless of whether the outputs were self-generated.\nThis suggests that the essence of the bias lies in perplexity and that the self-preference bias occurs because the LLMs' own outputs have lower perplexity.", "keywords": [ "large language model", "llm-as-a-judge", "bias", "fairness" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "Ns84n4NWh6", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Binh-Son Hua", "gender": "Male", "institution": "University of Dublin, Trinity College", "country": "IE", "position": "Assistant Professor" }, { "name": "Duc Thanh Nguyen", "gender": "Male", "institution": "Deakin University", "country": "AU", "position": "Assistant Professor" }, { "name": "Jaeyeon Kim", "gender": "Female", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Ka Chun SHUM", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yaxin Shi", "gender": "Female", "institution": "Centre for Frontier AI Research (CFAR) ", "country": "SG", "position": "Postdoc" }, { "name": "Ivor W Tsang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Sai-kit Yeung", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 44, "n_ref": 118, "n_ref_all": 141, "n_fig": 10, "n_tab": 5, "L_tab": 1547, "n_element_tab": 123, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 1940, "n_element_tab_1": 106, "formula_len_all": 422, "formula_len_all_1": 340, "len_all": 122849, "len_all_1": 56952, "len_abs": 1014, "len_title": 128, "len_sents": 32493, "len_sents_1": 24705, "n_sents": 243, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1019, "title": "Multi-view Consistent Image Generation through Self-calibrated Latent Refinement", "abs": "In this paper, we introduce a novel 3D-aware image generation framework that ensures high-quality and view-consistent image generation. Our core idea is to leverage the semantic latent space of a pre-trained 2D GAN for 3D view-consistent image generation, eliminating need for large-scale dataset use and prior knowledge of camera poses. To achieve this, we propose a latent refiner with multi-view and geometry-preserving capabilities, enabled by self-calibrated depth and pose estimation. Thanks to the advances of diffusion models, our refiner allows for view-consistent latent manipulation in GANs and can be trained using a self-supervised fashion. Our method optimizes the latent codes of a pre-trained 2D GAN across a wide range of pose angles. We demonstrate the effectiveness of our method through evaluations and comparisons with existing baselines on benchmark datasets. Experimental results show the superiority of our method over existing works in both the quality and view consistency of generated images.", "keywords": [ "3D aware image synthesis", "3D generation", "GAN", "Diffusion." ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "Ns6fnLFsCZ", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Dominic Phillips", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Efthymia Tsamoura", "gender": "Female", "institution": "Samsung AI", "country": "GB", "position": "Researcher" }, { "name": "Leon Jonathan Feldstein", "gender": "Male", "institution": "Bennu.Ai Limited", "country": "GB", "position": "Principal Researcher" } ], "n_formula": 51, "n_formula_1": 13, "n_ref_uni": 38, "n_ref": 76, "n_ref_all": 86, "n_fig": 2, "n_tab": 6, "L_tab": 978, "n_element_tab": 124, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 567, "n_element_tab_1": 77, "formula_len_all": 4022, "formula_len_all_1": 1280, "len_all": 179320, "len_all_1": 69426, "len_abs": 1276, "len_title": 128, "len_sents": 60531, "len_sents_1": 31557, "n_sents": 529, "n_sents_1": 274, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1283, "title": "Efficiently Learning Probabilistic Logical Models by Cheaply Ranking Mined Rules", "abs": "Probabilistic logical models are a core component of neurosymbolic AI and are important models in their own right for tasks that require high explainability. Unlike neural networks, logical models are often handcrafted using domain expertise, making their development costly and prone to errors. While there are algorithms that learn logical models from data, they are generally prohibitively expensive, limiting their applicability in real-world settings. In this work, we introduce precision and recall for logical rules and define their composition as rule utility -- a cost-effective measure to evaluate the predictive power of logical models. Further, we introduce SPECTRUM, a scalable framework for learning logical models from relational data. Its scalability derives from a linear-time algorithm that mines recurrent structures in the data along with a second algorithm that, using the cheap utility measure, efficiently ranks rules built from these structures. Moreover, we derive theoretical guarantees on the utility of the learnt logical model. As a result, we demonstrate across various tasks that SPECTRUM scales to larger datasets, often learning more accurate logical models orders of magnitude faster than previous methods without requiring specialised GPU hardware.", "keywords": [ "Structure Learning", "Probabilistic Logical Models", "MLN", "PSL", "Knowledge Graph Reasoning", "Neurosymbolic" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "Nrm1yJ577U", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "David Uminsky", "gender": "unknown", "institution": "University of Chicago", "country": "", "position": "Executive Director" }, { "name": "Joshua Charney", "gender": "Male", "institution": ", University of Chicago", "country": "", "position": "MS student" }, { "name": "Patricia Chiril", "gender": "Female", "institution": "University of Chicago", "country": "", "position": "Postdoc" }, { "name": "Rishabh Shastry", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Undergrad student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 32, "n_ref": 35, "n_ref_all": 53, "n_fig": 4, "n_tab": 1, "L_tab": 581, "n_element_tab": 103, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 419, "formula_len_all_1": 419, "len_all": 81876, "len_all_1": 78837, "len_abs": 1275, "len_title": 91, "len_sents": 23580, "len_sents_1": 23360, "n_sents": 129, "n_sents_1": 126, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1278, "title": "Entailment Progressions: A Robust Approach to Evaluating Reasoning Within Larger Discourse", "abs": "Textual entailment, or the ability to deduce whether a proposed hypothesis is logically supported by a given premise, has historically been applied to the evaluation of language modelling efficiency in tasks like question answering and text summarization. However, we hypothesize that these zero-shot entailment evaluations can be extended to the task of evaluating discourse within larger textual narratives.\nIn this paper, we propose a simple but effective method that sequentially evaluates\nchanges in textual entailment between sentences within a larger text, in an approach we denote as \u201cEntailment Progressions\u201d. These entailment progressions aim to capture the inference relations between sentences as an underlying component capable of distinguishing texts generated from various models and procedures. Our results suggest that entailment progressions can be used to effectively distinguish between machine-generated and human-authored texts across multiple established benchmark corpora and our own EP4MGT dataset. Additionally, our method displays robustness in performance when evaluated on paraphrased texts a technique that has historically affected the performance of well-established metrics when distinguishing between machine generated and human authored texts.", "keywords": [ "Natural language inference", "discourse structure", "text detection" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "NrfP7zZNiG", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Chenyu You", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Assistant Professor" }, { "name": "Jiacheng Liang", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Luoxi Tang", "gender": "Female", "institution": "State University of New York at Binghamton", "country": "US", "position": "MS student" }, { "name": "Muchao Ye", "gender": "unknown", "institution": "University of Iowa", "country": "US", "position": "Assistant Professor" }, { "name": "Xiaoqun Liu", "gender": "Female", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Zhaohan Xi", "gender": "Male", "institution": "State University of New York at Binghamton", "country": "US", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 57, "n_ref": 96, "n_ref_all": 125, "n_fig": 8, "n_tab": 7, "L_tab": 3274, "n_element_tab": 105, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 2110, "n_element_tab_1": 71, "formula_len_all": 541, "formula_len_all_1": 495, "len_all": 181317, "len_all_1": 65489, "len_abs": 1621, "len_title": 124, "len_sents": 38575, "len_sents_1": 30505, "n_sents": 266, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1550, "title": "Buckle Up: Robustifying LLMs at Every Customization Stage via Data Curation", "abs": "Large language models (LLMs) are extensively adapted for downstream applications through a process known as \"customization,\" with fine-tuning being a common method for integrating domain-specific expertise. \nHowever, recent studies have revealed a vulnerability that tuning LLMs with malicious samples can compromise their robustness and amplify harmful content, an attack known as \"jailbreaking.\"\nTo mitigate such attack, we propose an effective defensive framework utilizing data curation to revise commonsense texts and enhance their safety implication from the perspective of LLMs. The curated texts can mitigate jailbreaking attacks at every stage of the customization process: before customization to immunize LLMs against future jailbreak attempts, during customization to neutralize jailbreaking risks, or after customization to restore the compromised models. Since the curated data strengthens LLMs through the standard fine-tuning workflow, we do not introduce additional modules during LLM inference, thereby preserving the original customization process. Experimental results demonstrate a substantial reduction in jailbreaking effects, with up to a 100\\% success in generating responsible responses. Notably, our method is effective even with commonsense texts, which are often more readily available than safety-relevant data. With the every-stage defensive framework and supporting experimental performance, this work represents a significant advancement in mitigating jailbreaking risks and ensuring the secure customization of LLMs.", "keywords": [ "Safety Alignment", "Fine-tuning", "LLMs", "AI Security", "Jailbreaking" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "NrLXQWwCMg", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Hongjun Wang", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Qingtian Zhu", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Yinqiang Zheng", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Zhengwei Yin", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Zhao Jiancheng", "gender": "Male", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "JP", "position": "PhD student" }, { "name": "Mingze MA", "gender": "Male", "institution": "Tokyo University", "country": "JP", "position": "MS student" }, { "name": "YIFAN ZHAN", "gender": "Male", "institution": "University of Tokyo", "country": "JP", "position": "PhD student" } ], "n_formula": 54, "n_formula_1": 6, "n_ref_uni": 39, "n_ref": 66, "n_ref_all": 88, "n_fig": 10, "n_tab": 6, "L_tab": 1447, "n_element_tab": 114, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1230, "n_element_tab_1": 103, "formula_len_all": 5654, "formula_len_all_1": 589, "len_all": 149473, "len_all_1": 52152, "len_abs": 1640, "len_title": 131, "len_sents": 39621, "len_sents_1": 22246, "n_sents": 308, "n_sents_1": 146, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1606, "title": "Neural Tangent Kernel Analysis and Filtering for Robust Fourier Feature Embedding", "abs": "Implicit Neural Representations (INRs) employ neural networks to represent continuous functions by mapping coordinates to the corresponding values of the target function, with applications e.g., inverse graphics. However, INRs face a challenge known as spectral bias when dealing with scenes containing varying frequencies. To overcome spectral bias, the most common approach is the Fourier features-based methods such as positional encoding. However, Fourier features-based methods will introduce noise to output, which degrades their performances when applied to downstream tasks. In response, this paper addresses this problem by first investigating the underlying causes through the lens of the Neural Tangent Kernel. Through theoretical analysis, we propose that using Fourier features embedding can be interpreted as fitting Fourier series expansion of the target function, from which we find that it is the insufficiency in the finitely sampled frequencies that causes the generation of noisy outputs. Leveraging these insights, we introduce bias-free MLPs as an adaptive linear filter to locally suppress unnecessary frequencies while amplifying essential ones by adjusting the coefficients at the coordinate level. Additionally, we propose a line-search-based algorithm to adjust the filter's learning rate dynamically, achieving Pareto efficiency between the adaptive linear filter module and the INRs. Extensive experiments demonstrate that our proposed method consistently improves the performance of INRs on typical tasks, including image regression, 3D shape regression, and inverse graphics.", "keywords": [ "Implicit Neural Representation", "Spectral Bias", "Positional Encoding", "Fourier Features" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "NrDUhtIWsY", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Hanyu Gu", "gender": "Male", "institution": "University of Technology Sydney (UTS)", "country": "AU", "position": "Associate Professor" }, { "name": "Hui Zhang", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Full Professor" }, { "name": "Jianxu Mao", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Full Professor" }, { "name": "Junfei YI", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "PhD student" }, { "name": "Mingjie Li", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Tengfei Liu", "gender": "Male", "institution": "Beijing University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yaonan Wang", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaojun Chang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 18, "n_ref": 42, "n_ref_all": 69, "n_fig": 6, "n_tab": 14, "L_tab": 10037, "n_element_tab": 1117, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 4718, "n_element_tab_1": 458, "formula_len_all": 496, "formula_len_all_1": 494, "len_all": 149118, "len_all_1": 57997, "len_abs": 1323, "len_title": 146, "len_sents": 33239, "len_sents_1": 21218, "n_sents": 258, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1334, "title": "Teaching with Uncertainty: Unleashing the Potential of Knowledge Distillation in Object Detection", "abs": "Knowledge distillation (KD) has become a fundamental technique for model compression in object detection tasks. The data noise and training randomness may cause the knowledge of the teacher model to be unreliable, referred to as knowledge uncertainty. Existing methods only transfer this knowledge and could limit the student's ability to capture and understand the potential ``dark knowledge''. In this work, we introduce a new strategy that explicitly incorporates knowledge uncertainty, named Uncertainty-Driven Knowledge Extraction and Transfer (UET). Given that the knowledge distribution is unknown and high-dimensional in practice, we introduce a simple yet effective sampling method with Monte Carlo dropout (MC dropout) to estimate the teacher\u2019s knowledge uncertainty. Leveraging information theory, we integrate knowledge uncertainty into the conventional KD process, allowing the student model to benefit from knowledge diversity. UET is a plug-and-play method that integrates seamlessly with existing distillation techniques. We validate our approach through comprehensive experiments across various distillation strategies, detectors, and backbones. Specifically, UET achieves state-of-the-art results, with a ResNet50-based GFL detector obtaining 44.1\\% mAP on the COCO dataset\u2014surpassing baseline performance by 3.9\\%.", "keywords": [ "Object Detection", "Knowledge Distillation", "Uncertainty Estimation" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "Nr6V30wK1l", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Adam P. Generale", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Andreas Euan Robertson", "gender": "Male", "institution": "Sandia National Laboratories", "country": "US", "position": "Postdoc" }, { "name": "Surya Kalidindi", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 34, "n_formula_1": 16, "n_ref_uni": 52, "n_ref": 166, "n_ref_all": 219, "n_fig": 4, "n_tab": 2, "L_tab": 810, "n_element_tab": 120, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2601, "formula_len_all_1": 866, "len_all": 193421, "len_all_1": 61867, "len_abs": 1706, "len_title": 163, "len_sents": 59155, "len_sents_1": 29760, "n_sents": 371, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 115, "L_abs": 1646, "title": "Conditional Variable Flow Matching: Transforming Conditional Densities with Amortized Conditional Optimal Transport", "abs": "Forecasting stochastic nonlinear dynamical systems under the influence of conditioning variables is a fundamental challenge repeatedly encountered across the biological and physical sciences. While flow-based models can impressively predict the temporal evolution of probability distributions representing possible outcomes of a specific process, existing frameworks cannot satisfactorily account for the impact of conditioning variables on these dynamics. Amongst several limitations, existing methods require training data with paired conditions and are developed for discrete conditioning variables. We propose Conditional Variable Flow Matching (CVFM), a framework for learning flows transforming conditional distributions with amortization across continuous conditioning variables -- permitting predictions across the conditional density manifold. This is accomplished through several novel advances, in particular, simultaneous sample conditioned flows over the main and conditioning variables, alongside a conditional Wasserstein metric and kernel facilitating conditional optimal transport. Collectively, these advances allow for learning system dynamics provided measurement data whose states and conditioning variables are not in correspondence. We demonstrate CVFM on a suite of increasingly challenging problems, including discrete and continuous conditional mapping benchmarks, image-to-image domain transfer, and modeling the temporal evolution of materials internal structure during manufacturing processes. We observe that CVFM results in improved performance and convergence characteristics over alternative conditional variants.", "keywords": [ "Optimal Transport", "Flow Matching", "Generative Modeling", "Stochastic Dynamics", "Shr\u00f6dinger Bridge" ], "rating_list": [ 6, 6, 3, 3 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "NqkSUwMc0K", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Esfandiar Mohammadi", "gender": "Male", "institution": "Universit\u00e4t zu L\u00fcbeck", "country": "", "position": "Associate Professor" }, { "name": "Moritz Kirschte", "gender": "unknown", "institution": "University of Luebeck", "country": "DE", "position": "PhD student" }, { "name": "Saman Ardalan", "gender": "unknown", "institution": "Christian-Albrechts-Universit\u00e4t Kiel", "country": "", "position": "Researcher" }, { "name": "Sebastian Meiser", "gender": "unknown", "institution": "Universit\u00e4t zu L\u00fcbeck", "country": "", "position": "Researcher" } ], "n_formula": 79, "n_formula_1": 16, "n_ref_uni": 43, "n_ref": 109, "n_ref_all": 150, "n_fig": 13, "n_tab": 4, "L_tab": 902, "n_element_tab": 63, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 3065, "n_element_tab_1": 125, "formula_len_all": 11722, "formula_len_all_1": 2657, "len_all": 286123, "len_all_1": 79329, "len_abs": 2404, "len_title": 123, "len_sents": 92474, "len_sents_1": 30324, "n_sents": 952, "n_sents_1": 336, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1876, "title": "Private Blind Model Averaging \u2013 Distributed, Non-interactive, and Convergent", "abs": "Scalable distributed differentially private learning would benefit notably from reduced communication and synchronization overhead. The current best methods, based on gradient averaging, inherently require many synchronization rounds. In this work, we analyze blind model averaging for convex and smooth empirical risk minimization (ERM): each user first locally finishes training a model and then submits the model for secure averaging without any client-side online synchronization. This setting lends itself not only to data point-level privacy but also to flexible user-level privacy, where the combined impact of the user\u2019s trained model does not depend on the number of data points used for training.\n\nIn detail, we analyze the utility side of blind model averaging for support vector machines (SVMs) and the inherently multi-class Softmax regression (SoftmaxReg). On the theory side, we use strong duality to show for SVMs that blind model averaging converges toward centralized training performance if the task is robust against L2-regularization, i.e. if increasing the regularization weight does not destroy utility. Furthermore, we provide theoretical and experimental evidence that blind averaged Softmax Regression works well: we prove strong convexity of the dual problem by proving smoothness of the primal problem. Using this result, we also conclude the first output perturbation bounds for Softmax regression. On the experimental side, we support our theoretical SVM convergence. Furthermore, we observe hints of an even more fine-granular connection between good utility of model averaging and mid-range regularization weights which lead to compelling utility-privacy-tradeoffs for SVM and Softmax regression on 3 datasets (CIFAR-10, CIFAR-100, and federated EMNIST embeddings). We additionally provide ablation for an artificially extreme non-IID scenario.", "keywords": [ "differential privacy", "scalable distributed learning", "privacy-preserving machine learning", "privacy", "federated learning", "non-interactivity" ], "rating_list": [ 3, 3, 3, 8, 1 ], "soundness_list": [ 2, 2, 2, 3, 1 ], "presentation_list": [ 2, 1, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 3, 1 ], "confidence_list": [ 4, 3, 3, 3, 3 ] }, { "paper_id": "Nq7yKYL0Bp", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Min Zhang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Shizhuo Cheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Tintin Jiang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Yan Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zhengxi Lu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 27, "n_formula_1": 20, "n_ref_uni": 40, "n_ref": 76, "n_ref_all": 108, "n_fig": 11, "n_tab": 9, "L_tab": 3778, "n_element_tab": 596, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1734, "n_element_tab_1": 253, "formula_len_all": 2033, "formula_len_all_1": 1365, "len_all": 167474, "len_all_1": 60128, "len_abs": 1239, "len_title": 111, "len_sents": 42105, "len_sents_1": 24640, "n_sents": 357, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1270, "title": "ProtPainter: Draw or Drag Protein via Topology-guided Diffusion", "abs": "Recent advances in protein backbone generation have achieved promising results under structural, functional, or physical constraints. However, existing methods lack the flexibility for precise topology control, limiting navigation of the backbone space. We present $\\textbf{ProtPainter}$, a diffusion-based approach for generating protein backbones conditioned on 3D curves. ProtPainter follows a two-stage process: curve-based sketching and sketch-guided backbone generation. For the first stage, we propose $\\textbf{CurveEncoder}$, which predicts secondary structure annotations from a curve to parametrize sketch generation. For the second stage, the sketch guides the generative process in Denoising Diffusion Probabilistic Modeling (DDPM) to generate backbones. During the process, we further introduce a fusion scheduling scheme, Helix-Gating, to control the scaling factors. To evaluate, we propose the first benchmark for topology-conditioned protein generation, introducing Protein Restoration Task and a new metric, self-consistency Topology Fitness (scTF). Experiments demonstrate ProtPainter's ability to generate topology-fit (scTF $>$ 0.8) and designable (scTM $>$ 0.5) backbones, with drawing and dragging tasks showcasing its flexibility and versatility.", "keywords": [ "Protein Backbone Generation", "Conditional Diffusion", "Topology", "Protein Editing" ], "rating_list": [ 6, 5, 8 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 2, 4, 5 ] }, { "paper_id": "NpsgBKlApa", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Max Springer", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "MohammadTaghi Hajiaghayi", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "Full Professor" }, { "name": "Mohammad Hossein Bateni", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research scientist" }, { "name": "Neslihan Bulut", "gender": "unknown", "institution": "Research, Google", "country": "", "position": "Software Engineer" }, { "name": "Sasan Tavakkol", "gender": "Male", "institution": "Google", "country": "US", "position": "Software Engineer" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 54, "n_ref_all": 64, "n_fig": 6, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 2380, "n_element_tab_1": 1, "formula_len_all": 371, "formula_len_all_1": 115, "len_all": 232409, "len_all_1": 120626, "len_abs": 940, "len_title": 59, "len_sents": 55225, "len_sents_1": 38927, "n_sents": 365, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 943, "title": "Less is More: Adaptive Coverage for Synthetic Training Data", "abs": "Synthetic training data generation with Large Language Models (LLMs) like Google's Gemma and OpenAI's GPT offer a promising solution to the challenge of obtaining large, labeled datasets for training classifiers, especially when rapid model deployment is critical, such as classifying emerging social media trends or combating new forms of online abuse tied to current events. While prior research has examined the comparability of synthetic data to human-labeled data, this study introduces a novel sampling algorithm based on the maximum coverage problem to select a representative subset from a synthetically generated dataset. Our results demonstrate that training a classifier on this contextually sampled subset achieves superior performance compared to training on the entire dataset. This ``less is more'' approach not only improves accuracy but also reduces the volume of data required, leading to potentially more efficient training.", "keywords": [ "Large Language Models (LLMs)", "Synthetic Data Generation", "Sampling Algorithms", "Maximum Coverage Problem", "Data Efficiency" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "NpBhYnUgFU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Gayathri C", "gender": "Female", "institution": "Mahindra University", "country": "IN", "position": "Undergrad student" }, { "name": "Rambabu Damalla", "gender": "Male", "institution": "Indian Institute of Technology, Hyderabad", "country": "IN", "position": "PhD student" }, { "name": "Dr Rajeshreddy Datla", "gender": "Male", "institution": "Advanced data processing research institute ", "country": "IN", "position": "Scientist " } ], "n_formula": 50, "n_formula_1": 49, "n_ref_uni": 16, "n_ref": 36, "n_ref_all": 43, "n_fig": 4, "n_tab": 5, "L_tab": 1941, "n_element_tab": 244, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1839, "n_element_tab_1": 217, "formula_len_all": 2155, "formula_len_all_1": 2143, "len_all": 99908, "len_all_1": 66249, "len_abs": 1710, "len_title": 200, "len_sents": 28762, "len_sents_1": 25143, "n_sents": 212, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 149, "L_abs": 1719, "title": "SuperCAT: Super Resolution and Cross Semantic Attribute-guided Transformer based Feature Refinement for Zero-Shot Remote Sensing Scene Classification", "abs": "Zero-shot learning becomes challenging in classifying scenes of unseen classes due to the typical characteristics of remote-sensing images. The intricate variations and non-uniform spatial resolutions among the scenes of remote sensing images further complicate achieving discriminative semantic knowledge. To tackle these issues, we propose a SuperCAT framework comprising a super-resolution module, a cross-semantic attribute-guided Transformer (CAT), feature-generating models, and a feature refinement (FR) module for the zero-shot scene classification in remote sensing images. First, we leverage the semantic attributes for all the classes of four benchmark remote sensing scene classification datasets to explore semantic knowledge using super-resolution effectively. Then, the semantic attribute to visual Transformer (SAVT) and visual to semantic attribute Transformer (VSAT) modules in CAT learn to obtain attribute-based visual features and visual-based attribute features, respectively. The SAVT and VSAT modules collaboratively learn and teach each other using the feature-level and prediction-level semantic collaborative losses. The feature-generating models map semantic vectors to the visual features of remote-sensing images. The FR module incorporates triplet center margin loss and semantic loop consistency loss functions to capture class-related and semantically-related discriminative features for achieving intra-class closeness and inter-class distinctiveness. Our extensive experiments on four benchmark remote sensing image scene classification datasets demonstrate the efficacy of SuperCAT over state-of-the-art approaches. The code can be accessed at https://github.com/ZSL-RSI-SC/SuperCAT.", "keywords": [ "Scene classification", "remote sensing images", "zero-shot learning", "Transformer" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "NoiaAT0eec", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Shan You", "gender": "Male", "institution": "SenseTime Research", "country": "CN", "position": "Researcher" }, { "name": "Tao Huang", "gender": "Male", "institution": "The University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Yanxiang Ma", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "MS student" }, { "name": "Chang Xu", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" } ], "n_formula": 49, "n_formula_1": 24, "n_ref_uni": 41, "n_ref": 89, "n_ref_all": 96, "n_fig": 2, "n_tab": 7, "L_tab": 1401, "n_element_tab": 146, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 127, "n_element_tab_1": 21, "formula_len_all": 3078, "formula_len_all_1": 1410, "len_all": 168420, "len_all_1": 69093, "len_abs": 1442, "len_title": 116, "len_sents": 44616, "len_sents_1": 29451, "n_sents": 391, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1448, "title": "Learning Mask Invariant Mutual Information for Masked Image Modeling", "abs": "Masked autoencoders (MAEs) represent a prominent self-supervised learning paradigm in computer vision. Despite their empirical success, the underlying mechanisms of MAEs remain insufficiently understood. Recent studies have attempted to elucidate the functioning of MAEs through contrastive learning and feature representation analysis, yet these approaches often provide only implicit insights. In this paper, we propose a new perspective for understanding MAEs by leveraging the information bottleneck principle in information theory. Our theoretical analyses reveal that optimizing the latent features to balance relevant and irrelevant information is key to improving MAE performance. Building upon our proofs, we introduce MI-MAE, a novel method that optimizes MAEs through mutual information maximization and minimization. By enhancing latent features to retain maximal relevant information between them and the output, and minimizing irrelevant information between them and the input, our approach achieves better performance. Extensive experiments on standard benchmarks show that MI-MAE significantly outperforms MAE models in tasks such as image classification, object detection, and semantic segmentation. Our findings validate the theoretical framework and highlight the practical advantages of applying the information bottleneck principle to MAEs, offering deeper insights for developing more powerful self-supervised learning models.", "keywords": [ "Masked image modeling", "Self-supervised learning", "Visual pretraining" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "NoRvNK9eDp", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Andreas Geiger", "gender": "Male", "institution": "University of Tuebingen", "country": "DE", "position": "Professor" }, { "name": "Anthony Dick", "gender": "unknown", "institution": "University of Adelaide", "country": "", "position": "Associate Professor" }, { "name": "Hongdong Li", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "Full Professor" }, { "name": "Huan Lei", "gender": "Female", "institution": "University of Adelaide", "country": "AU", "position": "Postdoc" } ], "n_formula": 15, "n_formula_1": 9, "n_ref_uni": 67, "n_ref": 142, "n_ref_all": 166, "n_fig": 8, "n_tab": 7, "L_tab": 1754, "n_element_tab": 253, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 396, "n_element_tab_1": 29, "formula_len_all": 579, "formula_len_all_1": 398, "len_all": 190666, "len_all_1": 80826, "len_abs": 1251, "len_title": 102, "len_sents": 46559, "len_sents_1": 35793, "n_sents": 415, "n_sents_1": 296, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 1256, "title": "Level-Set Parameters: Novel Data for 3D Shape Analysis", "abs": "3D shape analysis has been widely explored based on traditional 3D data of point clouds and meshes, but the discrete nature of these data makes the analysis methods susceptible to variations in input resolutions. The recent development of neural fields brings in level-set parameters from signed distance functions as a novel, continuous, and numerical representation of 3D shapes, where the shape surfaces are defined as zero-level-sets of those functions. This motivated us to extend shape analysis from the traditional 3D data to these novel parameter data. Since the level-set parameters are not Euclidean like point clouds, we establish correlations across different shapes by formulating them as a pseudo-normal distribution, and learn the distribution prior from the respective dataset. To further explore the level-set parameters with shape transformations, we propose to condition a subset of these parameters on rotations and translations, and generate them with a hypernetwork. We demonstrate the potential of the novel continuous representation in pose-related shape analysis through applications to shape classification, retrieval under arbitrary poses, and 6D object pose estimation. Code and data in this research will be provided at github.", "keywords": [ "neural fields", "signed distance fields", "shape analysis" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 4, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "NoN4lziOUF", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Changwen Zheng", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Jingyao Wang", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Lingyu Si", "gender": "Male", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Wenwen Qiang", "gender": "Male", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 35, "n_formula_1": 8, "n_ref_uni": 67, "n_ref": 127, "n_ref_all": 214, "n_fig": 9, "n_tab": 17, "L_tab": 10370, "n_element_tab": 638, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 4819, "n_element_tab_1": 384, "formula_len_all": 2646, "formula_len_all_1": 692, "len_all": 356739, "len_all_1": 79361, "len_abs": 1965, "len_title": 111, "len_sents": 104583, "len_sents_1": 31200, "n_sents": 706, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1522, "title": "On the Universality of Self-Supervised Representation Learning", "abs": "In this paper, we investigate the characteristics that define a good representation or model. We propose that such a representation or model should possess universality, characterized by: (i) discriminability: performing well on training samples; (ii) generalization: performing well on unseen datasets; and (iii) transferability: performing well on unseen tasks with distribution shifts. Despite its importance, current self-supervised learning (SSL) methods lack explicit modeling of universality, and theoretical analysis remains underexplored. To address these issues, we aim to explore and incorporate universality into SSL. Specifically, we first revisit SSL from a task perspective and find that each mini-batch can be viewed as a multi-class classification task. We then propose that a universal SSL model should achieve: (i) learning universality by minimizing loss across all training samples, and (ii) evaluation universality by learning causally invariant representations that generalize well to unseen datasets and tasks. To quantify this, we introduce a \n$\\sigma$-measurement that assesses the gap between the performance of SSL model and optimal task-specific models. Furthermore, to model universality, we propose the GeSSL framework. It first learns task-specific models by minimizing SSL loss, then incorporates future updates to enhance discriminability, and finally integrates these models to learn from multiple mini-batch tasks. Theoretical and empirical evidence supports the effectiveness of GeSSL.", "keywords": [ "Self-Supervised Learning", "Representation Learning", "Unsupervised Learning" ], "rating_list": [ 5, 6, 3, 6, 6 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 1, 2, 3 ], "confidence_list": [ 3, 3, 3, 4, 4 ] }, { "paper_id": "No2PNOiKgb", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Bodo Rosenhahn", "gender": "Male", "institution": "Institut f\u00fcr Informationsverarbeitung", "country": "DE", "position": "Professor" }, { "name": "David Forsyth", "gender": "Male", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Florian Kluger", "gender": "Male", "institution": "Leibniz Universit\u00e4t Hannover", "country": "DE", "position": "PhD student" }, { "name": "Seemandhar Jain", "gender": "Male", "institution": "Department of Computer Science", "country": "US", "position": "MS student" }, { "name": "Vaibhav Vavilala", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 48, "n_ref": 93, "n_ref_all": 114, "n_fig": 10, "n_tab": 5, "L_tab": 1801, "n_element_tab": 418, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2356, "n_element_tab_1": 293, "formula_len_all": 177, "formula_len_all_1": 111, "len_all": 148931, "len_all_1": 62648, "len_abs": 1482, "len_title": 116, "len_sents": 44238, "len_sents_1": 27422, "n_sents": 355, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1499, "title": "Improved Convex Decomposition with Ensembling and Boolean Primitives", "abs": "Describing a scene in terms of primitives -- geometrically simple shapes that offer a parsimonious but accurate abstraction of structure -- is an established and difficult fitting problem. Different scenes require different numbers of primitives, and these primitives interact strongly; however, any proposed solution can be evaluated at inference time. The state of the art method involves a learned regression procedure to predict a start point consisting of a fixed number of primitives, followed by a descent method to refine the geometry and remove redundant primitives. Methods are evaluated by accuracy in depth and normal prediction and in scene segmentation. This paper shows that very significant improvements in accuracy can be obtained by (a) incorporating a small number of \\emph{negative} primitives and (b) ensembling over a number of different regression procedures. Ensembling is by refining each predicted start point, then choosing the best by fitting loss. Extensive experiments on the standard NYUv2 dataset confirm that negative primitives are useful, and that our refine-then-choose strategy outperforms choose-then-refine, confirming that the fitting problem is very difficult. Our ensembling with boolean primitives approach strongly outperforms existing methods; additionally we present several improvements to the underlying primitive generation process enabling us to obtain better decompositions with fewer primitives. Code will be released upon acceptance of the paper.", "keywords": [ "Geometric Primitives", "convex decomposition", "Ensembling" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "NnwDdPDwUq", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Raihan Seraj", "gender": "Male", "institution": "McGill University", "country": "", "position": "PhD student" }, { "name": "Tristan Sylvain", "gender": "Unspecified", "institution": "Borealis AI", "country": "", "position": "Machine Learning Researcher" }, { "name": "Lili Meng", "gender": "Female", "institution": "University of British Columbia", "country": "", "position": "Postdoc" } ], "n_formula": 16, "n_formula_1": 9, "n_ref_uni": 39, "n_ref": 53, "n_ref_all": 61, "n_fig": 4, "n_tab": 4, "L_tab": 4736, "n_element_tab": 419, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2162, "n_element_tab_1": 246, "formula_len_all": 650, "formula_len_all_1": 356, "len_all": 144869, "len_all_1": 58881, "len_abs": 4565, "len_title": 101, "len_sents": 38076, "len_sents_1": 26174, "n_sents": 305, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1106, "title": "Contextual Bandits with Entropy-based Human Feedback", "abs": "In recent years, preference-based human feedback mechanisms have become integral to improving model performance across a range of applications, including conversational AI systems like ChatGPT. However, existing methodologies often overlook critical factors such as model uncertainty and variability in feedback quality. To address these limitations, we propose an innovative entropy-based human feedback framework designed for contextual bandits, which balances exploration and exploitation by soliciting expert feedback when model entropy surpasses a predefined threshold. Our method is model-agnostic and adaptable to any contextual bandit agent employing stochastic policies. Through rigorous experimentation, we demonstrate that our approach requires minimal human feedback to achieve significant performance gains, even with suboptimal feedback quality. Our work not only introduces a novel feedback solicitation strategy but also underscores the robustness of integrating human guidance into machine learning systems. Our code is publicly available: \\url{https://anonymous.4open.science/r/CBHF-33C5}", "keywords": [ "Contextual bandits", "human feedback" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 1, 4, 2 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "NnExMNiTHw", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Kaixuan Huang", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Mengdi Wang", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Associate Professor" }, { "name": "Xudong Guo", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 16, "n_ref_uni": 40, "n_ref": 104, "n_ref_all": 118, "n_fig": 4, "n_tab": 2, "L_tab": 595, "n_element_tab": 58, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 577, "n_element_tab_1": 48, "formula_len_all": 1804, "formula_len_all_1": 1532, "len_all": 158787, "len_all_1": 64278, "len_abs": 1503, "len_title": 119, "len_sents": 43825, "len_sents_1": 29513, "n_sents": 378, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1522, "title": "SpecDec++: Boosting Speculative Decoding via Adaptive Candidate Lengths", "abs": "Speculative decoding reduces the inference latency of a target large language model via utilizing a smaller and faster draft model. Its performance depends on a hyperparameter $K$ --- the candidate length, i.e., the number of candidate tokens for the target model to verify in each round. However, previous methods often use simple heuristics to choose $K$, which may result in sub-optimal performance. We study the choice of the candidate length $K$ and formulate it as a Markov Decision Process. We theoretically show that the optimal policy of this Markov decision process takes the form of a threshold policy, i.e., the current speculation should stop and be verified when the probability of getting a rejection exceeds a threshold value. Motivated by this theory, we propose SpecDec++, an enhanced version of speculative decoding that adaptively determines the candidate length on the fly. We augment the draft model with a trained acceptance prediction head to predict the conditional acceptance probability of the candidate tokens. SpecDec++ will stop the current speculation when the predicted probability that at least one token gets rejected exceeds a threshold. We implement SpecDec++ and apply it to the llama-2-chat 7B \\& 70B model pair. Our adaptive method achieves a 2.04x speedup on the Alpaca dataset (7.2% improvement over the baseline speculative decoding). On the GSM8K and HumanEval datasets, our method achieves a 2.26x speedup (9.4% improvement) and 2.23x speedup (11.1% improvement), respectively.", "keywords": [ "speculative decoding", "reinforcement learning", "large language models" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "NmpOUCwAjR", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Alexander M Rush", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Associate Professor" }, { "name": "Celine Lee", "gender": "Female", "institution": "Cornell University", "country": "", "position": "PhD student" }, { "name": "Md Arafat Sultan", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Tahira Naseem", "gender": "Female", "institution": "IBM, International Business Machines", "country": "", "position": "Researcher" }, { "name": "Ramon Fernandez Astudillo", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 19, "n_ref": 53, "n_ref_all": 65, "n_fig": 2, "n_tab": 6, "L_tab": 1019, "n_element_tab": 151, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1396, "n_element_tab_1": 262, "formula_len_all": 1152, "formula_len_all_1": 1121, "len_all": 166537, "len_all_1": 59704, "len_abs": 1215, "len_title": 90, "len_sents": 35138, "len_sents_1": 25042, "n_sents": 299, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 1222, "title": "A Codespace Autoencoder for Language Tasks", "abs": "Modern language modeling datasets require models to handle compositional reasoning, fact recall, and task-specific constraints. While these tasks are expressed in natural language, they often imply an underlying symbolic representation. In this work, we consider methods for extracting a latent symbolic representation in an unsupervised manner. We propose an autoencoder that models observed text data as being generated from underlying code with a dataset level function library. Our method is non-parametric and leverages in-context learning and code interpretation for inference. Code as the latent symbolic representation offers two key advantages. First, code offers a structured space that can be explored via modular functions; second, code is interpretably executable using deterministic and neural interpreters, enabling compositional and programmatic decoding into text. By identifying and composing patterns in this latent space, we can sample programs that produce correct, diverse, and task-relevant text through program execution.\nWe demonstrate how our method induces a latent space with modern LLMs, explore patterns discovered within it, and evaluate text data synthesized from our induced latent space.", "keywords": [ "LLM", "autoencoders", "code generation", "data generation" ], "rating_list": [ 3, 3, 5, 3, 6 ], "soundness_list": [ 1, 3, 3, 1, 3 ], "presentation_list": [ 2, 1, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 3, 2, 3, 3, 4 ] }, { "paper_id": "NmmRPUCWIA", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Enhua Wu", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Emeritus" }, { "name": "Jianyuan Guo", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Kai Han", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Wei He", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Yehui Tang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yunhe Wang", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 10, "n_formula_1": 12, "n_ref_uni": 17, "n_ref": 34, "n_ref_all": 47, "n_fig": 2, "n_tab": 8, "L_tab": 2868, "n_element_tab": 158, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1019, "n_element_tab_1": 73, "formula_len_all": 269, "formula_len_all_1": 271, "len_all": 103215, "len_all_1": 53032, "len_abs": 1326, "len_title": 118, "len_sents": 26534, "len_sents_1": 24444, "n_sents": 194, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1339, "title": "Prompt-guided Visual Perception for Efficient Training-free Video LLM", "abs": "Vision-language large models have achieved remarkable success in various multi-modal tasks, yet applying them to video understanding remains challenging due to the inherent complexity and computational demands of video data. While training-based video-LLMs deliver high performance, they often require substantial resources for training and inference. Conversely, training-free approaches offer a more efficient alternative by adapting pre-trained image-LLMs models for video tasks without additional training, but they face inference efficiency bottlenecks due to the large number of visual tokens generated from video frames. In this work, we present a novel prompt-guided visual perception framework (abbreviated as \\emph{Free Video-LLM}) for efficient inference of training-free video LLMs. The proposed framework decouples spatial-temporal dimension and performs temporal frame sampling and spatial RoI cropping respectively based on task-specific prompts. Our method effectively reduces the number of visual tokens while maintaining high performance across multiple video question-answering benchmarks. Extensive experiments demonstrate that our approach achieves competitive results with significantly fewer tokens, offering an optimal trade-off between accuracy and computational efficiency compared to state-of-the-art video LLMs.", "keywords": [ "LLM", "Multi-modality", "VLM", "Video" ], "rating_list": [ 3, 6, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 1, 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5, 4 ] }, { "paper_id": "NmiFwEP8K5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Andreas Hotho", "gender": "Male", "institution": "Bayerische Julius-Maximilians-Universit\u00e4t W\u00fcrzburg", "country": "DE", "position": "Full Professor" }, { "name": "Daniel Schl\u00f6r", "gender": "unknown", "institution": "Julius-Maximilians-Universit\u00e4t W\u00fcrzburg", "country": "DE", "position": "Postdoc" }, { "name": "Janna Omeliyanenko", "gender": "unknown", "institution": "Bayerische Julius-Maximilians-Universit\u00e4t W\u00fcrzburg", "country": "", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 33, "n_ref": 70, "n_ref_all": 92, "n_fig": 3, "n_tab": 12, "L_tab": 4064, "n_element_tab": 271, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 944, "n_element_tab_1": 76, "formula_len_all": 730, "formula_len_all_1": 731, "len_all": 132489, "len_all_1": 68417, "len_abs": 1100, "len_title": 128, "len_sents": 38140, "len_sents_1": 31492, "n_sents": 271, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1104, "title": "GE-PEFT: Gated Expandable Parameter-Efficient Fine-Tuning for Continual Learning", "abs": "Continual learning (CL) is a research field focused on continuously adapting foundation models such as large language models (LMs) to newly emerging information sources and tasks. While aspects such as parameter efficiency, knowledge transfer, and managing model capacity have recently received attention, the main research focus in CL remains on preventing catastrophic forgetting. Specifically, there is a lack of solutions that address all these aspects simultaneously. We bridge this gap by introducing Gated Expandable Parameter-Efficient Fine-Tuning (GE-PEFT). Our approach shares knowledge of previous tasks through leveraging a single, dynamically expanding PEFT module within LMs while selectively gating irrelevant previous tasks. Our experiments across multiple task-incremental CL benchmarks demonstrate that GE-PEFT outperforms existing state-of-the-art CL approaches in both full CL and few-shot settings. Our ablation and parameter sensitivity studies highlight the benefit of each proposed component, demonstrating that GE-PEFT offers a more efficient and adaptive solution for CL in LMs.", "keywords": [ "PEFT", "Continual Learning", "Knowledge Transfer", "Language Models" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "NmcOAwRyH5", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Francesco Di Giovanni", "gender": "Male", "institution": "Valence Labs powered by recursion", "country": "GB", "position": "Researcher" }, { "name": "Johannes F. Lutzeyer", "gender": "Male", "institution": "Ecole Polytechique", "country": "FR", "position": "Assistant Professor" }, { "name": "Joshua Southern", "gender": "Male", "institution": "Imperial College London, Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Michael M. Bronstein", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" } ], "n_formula": 83, "n_formula_1": 28, "n_ref_uni": 46, "n_ref": 129, "n_ref_all": 168, "n_fig": 1, "n_tab": 7, "L_tab": 3258, "n_element_tab": 170, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1770, "n_element_tab_1": 95, "formula_len_all": 6364, "formula_len_all_1": 1788, "len_all": 190932, "len_all_1": 67570, "len_abs": 1137, "len_title": 113, "len_sents": 60904, "len_sents_1": 30766, "n_sents": 472, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1141, "title": "Understanding Virtual Nodes: Oversquashing and Node Heterogeneity", "abs": "While message passing neural networks (MPNNs) have convincing success in a range of applications, they exhibit limitations such as the oversquashing problem and their inability to capture long-range interactions. Augmenting MPNNs with a virtual node (VN) removes the locality constraint of the layer aggregation and has been found to improve performance on a range of benchmarks. We provide a comprehensive theoretical analysis of the role of VNs and benefits thereof, through the lenses of oversquashing and sensitivity analysis. First, we characterize, precisely, how the improvement afforded by VNs on the mixing abilities of the network and hence in mitigating oversquashing, depends on the underlying topology. We then highlight that, unlike Graph-Transformers (GTs), classical instantiations of the VN are often constrained to assign uniform importance to different nodes. Consequently, we propose a variant of VN with the same computational complexity, which can have different sensitivity to nodes based on the graph structure. We show that this is an extremely effective and computationally efficient baseline for graph-level tasks.", "keywords": [ "Graph Neural Networks", "Message Passing", "Virtual Nodes", "Oversquashing", "Graph Transformers" ], "rating_list": [ 5, 3, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3, 3 ], "contribution_list": [ 2, 1, 3, 3, 4 ], "confidence_list": [ 4, 4, 3, 3, 3 ] }, { "paper_id": "NmP8PvcMtc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Le Zhang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Ming-Ming Cheng", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Full Professor" }, { "name": "Ping Wang", "gender": "unknown", "institution": "Tianjin University", "country": "", "position": "Full Professor" }, { "name": "Qibin Hou", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiang Li", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuming Chen", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "PhD student" }, { "name": "ZhaoHui Zheng", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 57, "n_ref": 134, "n_ref_all": 152, "n_fig": 14, "n_tab": 8, "L_tab": 1439, "n_element_tab": 243, "n_fig_1": 8, "n_tab_1": 8, "L_tab_1": 1763, "n_element_tab_1": 244, "formula_len_all": 70, "formula_len_all_1": 96, "len_all": 156744, "len_all_1": 63148, "len_abs": 1530, "len_title": 58, "len_sents": 35521, "len_sents_1": 26086, "n_sents": 320, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1474, "title": "Efficient Multi-Level Learning for Dense Object Detection", "abs": "Dense object detection is crucial and favorable in the industry and has been popular for years with the success of the multi-level learning framework. By delivering the learning of objects into a multi-level feature pyramid, such a divide-and-conquer solution eases the optimization difficulty. However, this learning paradigm has a major shortcoming left behind. The shallow levels take tons of computational burden due to their high resolutions of the feature maps, heavily slowing down the inference speed. In this paper, we aim for minimal modifications to exchange a better speed-accuracy trade-off. The outcome is SlimHead, a very simple, efficient, and generalizable head network, which further unleashes the potential of multi-level learning for dense object detectors. It operates in two stages: Slim and Fat, initially plugging interpolator before the head network functions to \"slim'' the feature pyramid, and then recovering the features to original solution space by \"fatting'' the feature pyramid. Thanks to its flexibility, operations with higher computational complexity can be easily integrated to benefit accuracy without loss of inference efficiency. We also extend our SlimHead to multiple high-level vision tasks such as arbitrary-oriented object detection, pedestrian detection, and instance segmentation. Extensive experiments on PASCAL VOC, MS COCO, DOTA, and CrowdHuman demonstrate the broad applicability and the high practical value of our method.", "keywords": [ "Object Detection", "Multi-Level Learning", "Head Network" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 5, 5 ] }, { "paper_id": "NmILZXKcOi", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Caigao JIANG", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Danrui Qi", "gender": "Female", "institution": "Simon Fraser University", "country": "CA", "position": "PhD student" }, { "name": "Fan Zhou", "gender": "Male", "institution": "AntGroup", "country": "CN", "position": "Researcher" }, { "name": "Faqiang Chen", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Ganglin Wei", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Hongyang Zhang", "gender": "Male", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "MS student" }, { "name": "Jiahui Li", "gender": "unknown", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Siqiao Xue", "gender": "Male", "institution": "starguest.ai", "country": "CN", "position": "Researcher" }, { "name": "WENHUI SHI", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "ZhaoWang", "gender": "Male", "institution": "China Mobile Communications Company Limited Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Zhixuan Chu", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 38, "n_ref": 73, "n_ref_all": 93, "n_fig": 4, "n_tab": 10, "L_tab": 4260, "n_element_tab": 437, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 818, "n_element_tab_1": 95, "formula_len_all": 85, "formula_len_all_1": 85, "len_all": 160607, "len_all_1": 57057, "len_abs": 4190, "len_title": 132, "len_sents": 32001, "len_sents_1": 25939, "n_sents": 264, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1356, "title": "DB-GPT-Hub: Towards Open Benchmarking Text-to-SQL Empowered by Large Language Models", "abs": "Large language models (LLMs) becomes the dominant paradigm for the challenging task of text-to-SQL. LLM-empowered text-to-SQL methods are typically categorized into prompting-based and tuning approaches. Compared to prompting-based methods, benchmarking fine-tuned LLMs for text-to-SQL is important yet under-explored, partially attributed to the prohibitively high computational cost. In this paper, we present DB-GPT-Hub, an open benchmark suite for LLM-empowered text-to-SQL, which primarily focuses on tuning LLMs at large scales. The proposed benchmark consists of: 1. a standardized and comprehensive evaluation of text-to-SQL tasks by fine-tuning medium to large-sized open LLMs; 2. a modularized and easy-to-extend codebase with mainstream LLMs and experimental scenarios supported, which prioritizes fine-tuning methods but can be easily extended to prompt-based setting. Our work investigates the potential gains and the performance boundaries of tuning approaches, compared to prompting approaches and explores optimal solutions tailored to specific scenarios. We hope \\textit{DB-GPT-Hub}, along with these findings, enables further research and broad applications that would otherwise be difficult owing to the absence of a dedicated open benchmark. The project code has been released anonymously at https://github.com/anonymity-360/DB-GPT-Hub.", "keywords": [ "text-to-sql", "supervised finetuning", "LLM", "DB-GPT" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "NmFt9dIrSi", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Artur Back de Luca", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "George Giapitzakis", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "MS student" }, { "name": "Kimon Fountoulakis", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Associate Professor" }, { "name": "Petar Veli\u010dkovi\u0107", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Affiliated Lecturer" }, { "name": "Shenghao Yang", "gender": "Male", "institution": "University of Waterloo", "country": "", "position": "PhD student" } ], "n_formula": 52, "n_formula_1": 11, "n_ref_uni": 41, "n_ref": 165, "n_ref_all": 211, "n_fig": 22, "n_tab": 6, "L_tab": 360, "n_element_tab": 8, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3973, "formula_len_all_1": 603, "len_all": 291556, "len_all_1": 71471, "len_abs": 930, "len_title": 84, "len_sents": 108504, "len_sents_1": 33530, "n_sents": 871, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 966, "title": "Positional Attention: Out-of-Distribution Generalization and Expressivity for Neural Algorithmic Reasoning", "abs": "There has been a growing interest in the ability of neural networks to solve algorithmic tasks, such as arithmetic, summary statistics, and sorting. While state-of-the-art models like Transformers have demonstrated good generalization performance on in-distribution tasks, their out-of-distribution (OOD) performance is poor when trained end-to-end. In this paper, we focus on value generalization, a common instance of OOD generalization where the test distribution has the same input sequence length as the training distribution, but the value ranges in the training and test distributions do not necessarily overlap. To address this issue, we propose that using fixed positional encodings to determine attention weights\u2014referred to as positional attention\u2014enhances empirical OOD performance while maintaining expressivity. We support our claim about expressivity by proving that Transformers with positional attention can effectively simulate parallel algorithms.", "keywords": [ "Transformers; neural algorithmic reasoning; attention" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "NltQraRnbW", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Lizhen Lin", "gender": "Female", "institution": "University of Maryland, College Park", "country": "", "position": "Full Professor" }, { "name": "Rong Tang", "gender": "Female", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Yun Yang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Associate Professor" } ], "n_formula": 35, "n_formula_1": 35, "n_ref_uni": 38, "n_ref": 70, "n_ref_all": 82, "n_fig": 2, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4454, "formula_len_all_1": 4454, "len_all": 125479, "len_all_1": 70687, "len_abs": 1474, "len_title": 159, "len_sents": 37988, "len_sents_1": 32200, "n_sents": 211, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 110, "L_abs": 1145, "title": "Conditional Diffusion Models are Minimax-Optimal and Manifold-Adaptive for Conditional Distribution Estimation", "abs": "We consider a class of conditional forward-backward diffusion models for conditional generative modeling, that is, generating new data given a covariate (or control variable). To formally study the theoretical properties of these conditional generative models, we adopt a statistical framework of distribution regression to characterize the large sample properties of the conditional distribution estimators induced by these conditional forward-backward diffusion models. Here, the conditional distribution of data is assumed to smoothly change over the covariate. In particular, our derived convergence rate is minimax-optimal under the total variation metric within the regimes covered by the existing literature. Additionally, we extend our theory by allowing both the data and the covariate variable to potentially admit a low-dimensional manifold structure. In this scenario, we demonstrate that the conditional forward-backward diffusion model can adapt to both manifold structures, meaning that the derived estimation error bound (under the Wasserstein metric) depends only on the intrinsic dimensionalities of the data and the covariate.", "keywords": [ "conditional distribution estimation", "diffusion models", "distribution regression", "generative models", "manifold", "minimax rate" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "Nlm3Xf0W9S", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Chenxi Liu", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Heng Huang", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Junfeng Guo", "gender": "Male", "institution": "University of Maryland Institute for Advanced Computer Studies, University of Maryland, College Park", "country": "US", "position": "Postdoc" }, { "name": "Ruibo Chen", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Yanshuo Chen", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Yihan Wu", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 15, "n_ref_uni": 24, "n_ref": 65, "n_ref_all": 89, "n_fig": 6, "n_tab": 4, "L_tab": 1479, "n_element_tab": 189, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1479, "n_element_tab_1": 189, "formula_len_all": 2285, "formula_len_all_1": 1410, "len_all": 123069, "len_all_1": 65132, "len_abs": 944, "len_title": 94, "len_sents": 37008, "len_sents_1": 26810, "n_sents": 324, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 46, "L_abs": 949, "title": "A Watermark for Order-Agnostic Language Models", "abs": "Statistical watermarking techniques are well-established for sequentially decoded language models (LMs). However, these techniques cannot be directly applied to order-agnostic LMs, as the tokens in order-agnostic LMs are not generated sequentially. In this work, we introduce PATTERN-MARK, a pattern-based watermarking framework specifically designed for order-agnostic LMs. We develop a\nMarkov-chain-based watermark generator that produces watermark key sequences with high-frequency key patterns. Correspondingly, we propose a statistical pattern-based detection algorithm that recovers the key sequence during detection and conducts statistical tests based on the count of high-frequency patterns. Our extensive evaluations on order-agnostic LMs, such as ProteinMPNN and CMLM, demonstrate PATTERN-MARK\u2019s enhanced detection efficiency, generation quality, and robustness, positioning it as a superior watermarking technique for order-agnostic LMs.", "keywords": [ "language model watermarking; generative model;" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "NlY3XppPt3", "primary_area": "datasets and benchmarks", "n_author": 1, "author_info_list": [ { "name": "Tung Nguyen", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Associate Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 17, "n_ref": 22, "n_ref_all": 32, "n_fig": 3, "n_tab": 4, "L_tab": 2695, "n_element_tab": 52, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 482, "n_element_tab_1": 8, "formula_len_all": 644, "formula_len_all_1": 644, "len_all": 127568, "len_all_1": 66946, "len_abs": 1946, "len_title": 70, "len_sents": 44972, "len_sents_1": 32528, "n_sents": 372, "n_sents_1": 256, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 710, "title": "Improving AI via Novel Computational Models and Programming Challenges", "abs": "AI, like humans, should be able to adapt and apply learned knowledge across diverse domains, such as computational models, mathematical/formal systems, and programming languages to solve problems. Current AI training often relies on existing systems, which limits its ability to generate original solutions or generalize across unfamiliar contexts. To address this, we propose a new computational model along with a revised programming language tailored to this model. By challenging AI to write, analyze, or verify programs within these new frameworks, and by utilizing a virtual machine for evaluation, we aim to test and enhance the AI's adaptability and problem-solving capabilities in a verifiable manner.", "keywords": [ "AI", "LLM", "code generation" ], "rating_list": [ 3, 1, 1, 3 ], "soundness_list": [ 1, 1, 2, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 1, 2, 1, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "NlEt8LYAxC", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Chen Liu", "gender": "Male", "institution": "City University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Xuyang Zhong", "gender": "Male", "institution": "City University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Yixiao HUANG", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 56, "n_formula_1": 12, "n_ref_uni": 42, "n_ref": 147, "n_ref_all": 202, "n_fig": 9, "n_tab": 18, "L_tab": 3191, "n_element_tab": 411, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 326, "n_element_tab_1": 39, "formula_len_all": 6590, "formula_len_all_1": 1158, "len_all": 194253, "len_all_1": 75991, "len_abs": 3492, "len_title": 121, "len_sents": 57102, "len_sents_1": 36540, "n_sents": 458, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1102, "title": "Fast Adversarial Training against Sparse Attacks Requires Loss Smoothing", "abs": "This paper studies fast adversarial training against sparse adversarial perturbations. We highlight the challenges faced when employing $1$-step attacks on $l_0$ bounded perturbations for fast adversarial training, including degraded performance and the occurrence of catastrophic overfitting (CO). We highlight that CO in $l_0$ adversarial training is caused by sub-optimal perturbation locations of $1$-step attack, which is distinct from other cases. Theoretical and empirical analyses reveal that the loss landscape of $l_0$ adversarial training is more craggy compared to its $l_\\infty$, $l_2$ and $l_1$ counterparts. Moreover, we corroborate that the craggy loss landscape can aggravate CO. To address these issues, we propose Fast-LS-$l_0$ that incorporates soft label and the trade-off loss function to smooth the adversarial loss landscape. Extensive experiments demonstrate our method can overcome the challenge of catastrophic overfitting, achieves state-of-the-art performance and narrows down the performance gap between $1$-step and multi-step adversarial training against sparse attacks.", "keywords": [ "adversarial robustness", "sparse attack", "fast adversarial training" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "NkGDNM8LB0", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Itsik Pe'er", "gender": "Male", "institution": "Columbia University", "country": "", "position": "Full Professor" }, { "name": "Philippe Chlenski", "gender": "Male", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Raiyan R. Khan", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 59, "n_ref": 83, "n_ref_all": 120, "n_fig": 16, "n_tab": 8, "L_tab": 1368, "n_element_tab": 126, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 5, "n_element_tab_1": 1, "formula_len_all": 952, "formula_len_all_1": 724, "len_all": 180400, "len_all_1": 59569, "len_abs": 1318, "len_title": 47, "len_sents": 51450, "len_sents_1": 28481, "n_sents": 373, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 28, "L_abs": 1334, "title": "Hyperbolic Genome Embeddings", "abs": "Current approaches to genomic sequence modeling often struggle to align the inductive biases of machine learning models with the evolutionarily-informed structure of biological systems. To this end, we formulate a novel application of hyperbolic CNNs that exploits this structure, enabling more expressive DNA sequence representations. Our strategy circumvents the need for explicit phylogenetic mapping while discerning key properties of sequences pertaining to core functional and regulatory behavior. Across 37 out of 43 genome interpretation benchmark datasets, our hyperbolic models outperform their Euclidean equivalents. Notably, our approach even surpasses state-of-the-art performance on seven GUE benchmark datasets, consistently outperforming many DNA language models while using 13-379$\\times$ fewer parameters and avoiding pretraining. Our results include a novel benchmark dataset - the Transposable Elements Benchmark - which explores a significant but understudied component of the genome with deep evolutionary significance. We further motivate our work by constructing an empirical method for interpreting the hyperbolicity of dataset embeddings. Throughout these assessments, we find persistent evidence highlighting the potential of our hyperbolic framework as a robust paradigm for genome representation learning.", "keywords": [ "genomics", "representation learning", "hyperbolic geometry" ], "rating_list": [ 8, 8, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "Nk1MegaPuG", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Jasper Dekoninck", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Marc Fischer", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Mark Niklas Mueller", "gender": "Male", "institution": "LogicStar AI", "country": "CH", "position": "Researcher" }, { "name": "Martin Vechev", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Maximilian Baader", "gender": "unknown", "institution": "ETH Zurich", "country": "CH", "position": "Postdoc" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 26, "n_ref": 166, "n_ref_all": 210, "n_fig": 4, "n_tab": 21, "L_tab": 7182, "n_element_tab": 812, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1568, "n_element_tab_1": 110, "formula_len_all": 242, "formula_len_all_1": 133, "len_all": 271655, "len_all_1": 70129, "len_abs": 1955, "len_title": 118, "len_sents": 55615, "len_sents_1": 33211, "n_sents": 561, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1124, "title": "Evading Data Contamination Detection for Language Models is (too) Easy", "abs": "The benchmark performance of large language models (LLMs) has a high impact on their popularity and is thus of great importance to many model providers. However, the reliability of such benchmark scores as a measure of model quality gets compromised if the model is contaminated with benchmark data. While recent contamination detection methods try to address this issue, they overlook the possibility of deliberate contamination by malicious model providers aiming to evade detection. We propose a categorization of model providers based on their (de)contamination practices and argue that malicious contamination is of crucial importance as it casts doubt on the reliability of public benchmarks. To study this issue more rigorously, we analyze current contamination detection methods based on their assumptions. This analysis reveals a significant vulnerability in existing approaches: they do not account for rephrased benchmark data used during training by malicious actors. We demonstrate how exploiting this gap can result in significantly inflated benchmark scores while completely evading current detection methods.", "keywords": [ "large language models", "model evaluation", "malicious actors" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "Njx1NjHIx4", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Isaac L. Chuang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Liu Ziyin", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Tomer Galanti", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" }, { "name": "tomaso a poggio", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 154, "n_formula_1": 25, "n_ref_uni": 35, "n_ref": 51, "n_ref_all": 95, "n_fig": 14, "n_tab": 2, "L_tab": 320, "n_element_tab": 3, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 87, "n_element_tab_1": 4, "formula_len_all": 8946, "formula_len_all_1": 2401, "len_all": 189750, "len_all_1": 74631, "len_abs": 1313, "len_title": 96, "len_sents": 57855, "len_sents_1": 33861, "n_sents": 609, "n_sents_1": 268, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 47, "L_abs": 1316, "title": "Formation of Representations in Neural Networks", "abs": "Understanding neural representations will help open the black box of neural networks and advance our scientific understanding of modern AI systems. However, how complex, structured, and transferable representations emerge in modern neural networks has remained a mystery. Building on previous results, we propose the Canonical Representation Hypothesis (CRH), which posits a set of six alignment relations to universally govern the formation of representations in most hidden layers of a neural network. Under the CRH, the latent representations (R), weights (W), and neuron gradients (G) become mutually aligned during training. This alignment implies that neural networks naturally learn compact representations, where neurons and weights are invariant to task-irrelevant transformations. We then show that the breaking of CRH leads to the emergence of reciprocal power-law relations between R, W, and G, which we refer to as the Polynomial Alignment Hypothesis (PAH). We present a minimal-assumption theory demonstrating that the balance between gradient noise and regularization is crucial for the emergence the canonical representation. The CRH and PAH lead to an exciting possibility of unifying major key deep learning phenomena, including neural collapse and the neural feature ansatz, in a single framework.", "keywords": [ "representation learning", "neural collapse", "neural feature ansatz" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 2, 3, 2, 3 ] }, { "paper_id": "Nifg2fQMGW", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Haoji Hu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Jiang Bian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Jianhong Bai", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Junliang Guo", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Tianyu He", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Researcher" }, { "name": "Yuchi Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Zuozhu Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 82, "n_ref": 158, "n_ref_all": 195, "n_fig": 24, "n_tab": 6, "L_tab": 1794, "n_element_tab": 244, "n_fig_1": 16, "n_tab_1": 2, "L_tab_1": 760, "n_element_tab_1": 123, "formula_len_all": 422, "formula_len_all_1": 422, "len_all": 189327, "len_all_1": 60284, "len_abs": 430, "len_title": 128, "len_sents": 40520, "len_sents_1": 26670, "n_sents": 322, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1144, "title": "UniEdit: A Unified Tuning-Free Framework for Video Motion and Appearance Editing", "abs": "Recent advances in text-guided video editing have showcased promising results in appearance editing (e.g., stylization). However, video motion editing in the temporal dimension (e.g., from eating to waving), which distinguishes video editing from image editing, is underexplored. In this work, we present UniEdit, a tuning-free framework that supports both video motion and appearance editing by harnessing the power of a pre-trained text-to-video generator within an inversion-then-generation framework.\nTo realize motion editing while preserving source video content, based on the insights that temporal and spatial self-attention layers encode inter-frame and intra-frame dependency respectively, we introduce auxiliary motion-reference and reconstruction branches to produce text-guided motion and source features respectively. The obtained features are then injected into the main editing path via temporal and spatial self-attention layers. Extensive experiments demonstrate that UniEdit covers video motion editing and various appearance editing scenarios, and surpasses the state-of-the-art methods. Our code will be publicly available.", "keywords": [ "Diffusion Model", "Video Editing" ], "rating_list": [ 1, 5, 8, 5 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "NiNIthntx7", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Dhruv Gautam", "gender": "Male", "institution": "UC Berkeley, University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Jinu Jang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Neel Sundaresan", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Vice President, AI and Engineering" }, { "name": "Roshanak Zilouchian Moghaddam", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Spandan Garg", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 55, "n_ref": 112, "n_ref_all": 132, "n_fig": 10, "n_tab": 4, "L_tab": 929, "n_element_tab": 32, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 341, "n_element_tab_1": 18, "formula_len_all": 44, "formula_len_all_1": 122, "len_all": 199175, "len_all_1": 66418, "len_abs": 1466, "len_title": 125, "len_sents": 50633, "len_sents_1": 33148, "n_sents": 357, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1471, "title": "RefactorBench: Evaluating Stateful Reasoning in Language Agents Through Code", "abs": "Recent advances in language model (LM) agents and function calling have enabled autonomous, feedback-driven systems to solve problems across various digital domains. To better understand the unique limitations of LM agents, we introduce RefactorBench, a benchmark consisting of 100 large handcrafted multi-file refactoring tasks in popular open-source repositories. Solving tasks within RefactorBench requires thorough exploration of dependencies across multiple files and strong adherence to relevant instructions. Every task is defined by 3 natural language instructions of varying specificity and is mutually exclusive, allowing for the creation of longer combined tasks on the same repository. Baselines on RefactorBench reveal that current LM agents struggle with simple compositional tasks, solving only 22\\% of tasks with base instructions, in contrast to a human developer with short time constraints solving 87\\%. Through trajectory analysis, we identify various unique failure modes of LM agents, and further explore the failure mode of tracking past actions. By adapting a baseline agent to condition on representations of state, we achieve a 43.9\\% improvement in solving RefactorBench tasks. We further extend our state-aware approach to encompass entire digital environments and outline potential directions for future research. RefactorBench aims to support the study of LM agents by providing a set of real-world, multi-hop tasks within the realm of code.", "keywords": [ "Language Agents", "Benchmarks", "Code Generation", "Reasoning", "State-Awareness", "Refactoring", "Long-Horizon Tasks", "Knowledge Representation" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "NiCSyYOfex", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Michael T Schaub", "gender": "unknown", "institution": "RWTH Aachen University", "country": "DE", "position": "Assistant Professor" }, { "name": "Vincent Peter Grande", "gender": "Male", "institution": "Rheinisch Westf\u00e4lische Technische Hochschule Aachen", "country": "DE", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 13, "n_ref_uni": 58, "n_ref": 91, "n_ref_all": 141, "n_fig": 16, "n_tab": 1, "L_tab": 178, "n_element_tab": 27, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1341, "formula_len_all_1": 895, "len_all": 248319, "len_all_1": 70769, "len_abs": 777, "len_title": 110, "len_sents": 93844, "len_sents_1": 33936, "n_sents": 704, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 779, "title": "Node-Level Topological Representation Learning on Point Clouds", "abs": "Topological Data Analysis (TDA) allows us to extract powerful topological, and higher-order information on the global shape of a data set or point cloud. Tools like Persistent Homology or the Euler Transform give a single complex description of the global structure of the point cloud. However, common machine learning applications like classification require point-level information and features to be available. In this paper, we bridge this gap and propose a novel method to extract node-level topological features from complex point clouds using discrete variants of concepts from algebraic topology and differential geometry. We verify the effectiveness of these topological point features (TOPF) on both synthetic and real-world data and study their robustness under noise.", "keywords": [ "Topological Data Analysis", "Hodge Laplacian", "Hodge Theory", "Geometry Processing", "Differential Geometry", "Algebraic Topology", "Point Clouds", "Representation Learning on Point Clouds" ], "rating_list": [ 6, 5, 3, 8, 5 ], "soundness_list": [ 4, 2, 1, 3, 3 ], "presentation_list": [ 3, 2, 3, 4, 3 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 3, 2, 5, 4, 4 ] }, { "paper_id": "Ni4jNyroJZ", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Jong-Seok Lee", "gender": "unknown", "institution": "Yonsei University", "country": "KR", "position": "Full Professor" }, { "name": "Jonghyun Park", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Undergrad student" }, { "name": "Juyeop Kim", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 33, "n_ref": 51, "n_ref_all": 101, "n_fig": 16, "n_tab": 5, "L_tab": 2520, "n_element_tab": 218, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 316, "n_element_tab_1": 52, "formula_len_all": 121, "formula_len_all_1": 86, "len_all": 129060, "len_all_1": 57353, "len_abs": 1228, "len_title": 145, "len_sents": 44305, "len_sents_1": 27434, "n_sents": 328, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1198, "title": "Impact of Regularization on Calibration and Robustness: From the Representation Space Perspective", "abs": "Recent studies have shown that regularization techniques using soft labels, e.g., label smoothing, Mixup, and CutMix, not only enhance image classification accuracy but also improve model calibration and robustness against adversarial attacks. However, the underlying mechanisms of such improvements remain underexplored. In this paper, we offer a novel explanation from the perspective of the representation space. Our investigation first reveals that the decision regions in the representation space form cone-like shapes around the origin after training regardless of the presence of regularization. However, applying regularization causes changes in the distribution of features (or representation vectors obtained at the penultimate layer). The magnitudes of the representation vectors are reduced and subsequently the cosine similarities between the representation vectors and the class centers (minimal loss points for each class) become higher, which acts as a central mechanism inducing improved calibration and robustness. Our findings provide new insights into the characteristics of the high-dimensional representation space in relation to training and regularization using soft labels.", "keywords": [ "Soft labels", "regularization", "representation space", "image classification" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 3, 2, 2, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 4 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "NhqKHHK4Nk", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Henrik Voigt", "gender": "Male", "institution": "University of Jena", "country": "", "position": "PhD student" }, { "name": "Joachim Giesen", "gender": "unknown", "institution": "University of Jena", "country": "", "position": "Full Professor" }, { "name": "Kai Lawonn", "gender": "unknown", "institution": "Friedrich-Schiller Universit\u00e4t Jena", "country": "", "position": "Full Professor" }, { "name": "Michael Habeck", "gender": "unknown", "institution": "University of Jena", "country": "DE", "position": "Professor" }, { "name": "Paul Kahlmeyer", "gender": "unknown", "institution": "Friedrich-Schiller Universit\u00e4t Jena", "country": "DE", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 25, "n_ref": 47, "n_ref_all": 57, "n_fig": 4, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 416, "formula_len_all_1": 494, "len_all": 103022, "len_all_1": 64508, "len_abs": 1132, "len_title": 106, "len_sents": 31685, "len_sents_1": 31387, "n_sents": 256, "n_sents_1": 249, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1150, "title": "Boosting Recovery in Transformer-Based Symbolic Regression", "abs": "The traditional objective in regression is generalization. That is, learning a function from training data that performs well beyond the training data. Symbolic regression adds another objective, namely, interpretability of the regressor. \nIn the context of regression, interpretability means that the representation of the regressor facilitates insights into mechanisms that underlie the functional dependence. State-of-the-art symbolic regressors provide such insights. However, the state of the art predominantly incurs high costs at inference time. The recently proposed transformer-based end-to-end approach is orders of magnitude faster at inference time. \nIt does, however, not achieve state-of-the-art performance in terms of interpretability, which is typically measured by the ability to recover ground truth formulas from samples. Here, we show that the recovery performance of the end-to-end approach can be boosted by carefully selecting the training data. We construct a synthetic dataset from first principles and demonstrate that the capacity to recover ground truth formulas is proportional to the available computational resources.", "keywords": [ "symbolic regression", "interpretability", "transformer", "recovery" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "NhIaRz9Qf5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Juanzi Li", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Lei Hou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Liangming Pan", "gender": "Male", "institution": "University of Arizona", "country": "US", "position": "Assistant Professor" }, { "name": "Linmei Hu", "gender": "Female", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Liu Weichuan", "gender": "Male", "institution": "Siemens Corporate Research", "country": "", "position": "Researcher" }, { "name": "Shulin Cao", "gender": "Female", "institution": "Zhipu AI", "country": "CN", "position": "Researcher" }, { "name": "Weijian Qi", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "MS student" }, { "name": "Zijun Yao", "gender": "Male", "institution": "Department of Computer Science and Technology, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 36, "n_ref": 73, "n_ref_all": 94, "n_fig": 6, "n_tab": 9, "L_tab": 3627, "n_element_tab": 221, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 3377, "n_element_tab_1": 191, "formula_len_all": 169, "formula_len_all_1": 191, "len_all": 183583, "len_all_1": 70921, "len_abs": 967, "len_title": 129, "len_sents": 46448, "len_sents_1": 27708, "n_sents": 473, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 972, "title": "SeaKR: Self-aware Knowledge Retrieval for Adaptive Retrieval Augmented Generation", "abs": "Adaptive Retrieval-Augmented Generation (RAG) is an effective strategy to alleviate hallucination of large language models (LLMs). It dynamically determines whether LLMs need external knowledge for generation and invokes retrieval accordingly. This paper introduces Self-aware Knowledge Retrieval (SeaKR), a novel adaptive RAG model that extracts self-aware uncertainty of LLMs from their internal states. SeaKR activates retrieval when the LLMs present high self-aware uncertainty for generation. To effectively integrate retrieved knowledge snippets, SeaKR re-ranks them based on LLM's self-aware uncertainty to preserve the snippet that reduces their uncertainty to the utmost. To facilitate solving complex tasks that require multiple retrievals, SeaKR utilizes their self-aware uncertainty to choose among different reasoning strategies. Our experiments on both complex and simple Question Answering datasets show that SeaKR outperforms existing adaptive RAG methods.", "keywords": [ "Retrieval-Augmented Generation", "Internal State of LLMs", "Open-domain Question Answering" ], "rating_list": [ 5, 3, 8, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Nh8NLlIfBv", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Guyue Zhou", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Haoyi Niu", "gender": "Male", "institution": "Department of Automation, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Jianming HU", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Qimao Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Tenglong Liu", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Xianyuan Zhan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yi ZHANG", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Li Jianxiong", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 64, "n_ref": 124, "n_ref_all": 150, "n_fig": 6, "n_tab": 10, "L_tab": 4733, "n_element_tab": 368, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2144, "n_element_tab_1": 154, "formula_len_all": 1430, "formula_len_all_1": 881, "len_all": 203668, "len_all_1": 68837, "len_abs": 1566, "len_title": 116, "len_sents": 58164, "len_sents_1": 33267, "n_sents": 366, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1622, "title": "xTED: Cross-Domain Adaptation via Diffusion-Based Trajectory Editing", "abs": "Reusing pre-collected data from different domains is an appealing solution for decision-making tasks that have insufficient data in the target domain but are relatively abundant in other related domains. Existing cross-domain policy transfer methods mostly aim at learning domain correspondences or corrections to facilitate policy learning, such as learning domain/task-specific discriminators, representations, or policies. This design philosophy often results in heavy model architectures or task/domain-specific modeling, lacking flexibility. This reality makes us wonder: can we directly bridge the domain gaps universally at the data level, instead of relying on complex downstream cross-domain policy transfer models? In this study, we propose the **Cross**-Domain **T**rajectory **ED**iting (**xTED**) framework that employs a specially designed diffusion model for cross-domain trajectory adaptation. Our proposed model architecture effectively captures the intricate dependencies among states, actions, and rewards, as well as the dynamics patterns within target data. By utilizing the pre-trained diffusion as a prior, source domain trajectories can be transformed to match with target domain properties while preserving original semantic information. This process implicitly corrects underlying domain gaps, enhancing state realism and dynamics reliability in the source data, and allowing flexible incorporation with various downstream policy learning methods. Despite its simplicity, xTED demonstrates superior performance in extensive simulation and [real-robot experiments](https://xted24.github.io/xTED/).", "keywords": [ "Cross-Domain Policy Transfer", "Embodied Decision Making", "Diffusion Models" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "Nh1w3ZnDaH", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Jaemoo Choi", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Jaewoong Choi", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "Assistant Professor" }, { "name": "Myungjoo Kang", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "TAEKYUNG LEE", "gender": "unknown", "institution": "SNU", "country": "", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 18, "n_ref_uni": 38, "n_ref": 91, "n_ref_all": 116, "n_fig": 3, "n_tab": 10, "L_tab": 2533, "n_element_tab": 319, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2053, "n_element_tab_1": 213, "formula_len_all": 1642, "formula_len_all_1": 1532, "len_all": 141750, "len_all_1": 73033, "len_abs": 1175, "len_title": 120, "len_sents": 40756, "len_sents_1": 29956, "n_sents": 337, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1180, "title": "Unsupervised Point Cloud Completion through Unbalanced Optimal Transport", "abs": "Unpaired point cloud completion explores methods for learning a completion map from unpaired incomplete and complete point cloud data. In this paper, we propose a novel approach for unpaired point cloud completion using the unbalanced optimal transport map, called Unbalanced Optimal Transport Map for Unpaired Point Cloud Completion (UOT-UPC). We demonstrate that the unpaired point cloud completion can be naturally interpreted as the Optimal Transport (OT) problem and introduce the Unbalanced Optimal Transport (UOT) approach to address the class imbalance problem, which is prevalent in unpaired point cloud completion datasets. \nMoreover, we analyze the appropriate cost function for unpaired completion tasks. This analysis shows that the InfoCD cost function is particularly well-suited for this task.\nOur model is the first attempt to leverage UOT for unpaired point cloud completion, achieving competitive or superior results on both single-category and multi-category datasets. In particular, our model is especially effective in scenarios with class imbalance, where the proportions of categories are different between the incomplete and complete point cloud datasets.", "keywords": [ "Point Cloud Completion", "Unpaired Point Cloud Completion", "Unbalanced optimal transport", "Optimal transport" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 1, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 5 ] }, { "paper_id": "Nh1ZH61OqF", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Jin Liu", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "Associate Professor" }, { "name": "Junwen Duan", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaokang Pan", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "MS student" }, { "name": "Zhe Qu", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "Full Professor" }, { "name": "SHICHAO KAN", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "Lecturer" } ], "n_formula": 144, "n_formula_1": 19, "n_ref_uni": 20, "n_ref": 60, "n_ref_all": 90, "n_fig": 19, "n_tab": 5, "L_tab": 3392, "n_element_tab": 351, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 350, "n_element_tab_1": 12, "formula_len_all": 24093, "formula_len_all_1": 1750, "len_all": 199233, "len_all_1": 67633, "len_abs": 1212, "len_title": 127, "len_sents": 52584, "len_sents_1": 29227, "n_sents": 578, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1238, "title": "AdaFM: Adaptive Variance-Reduced Algorithm for Stochastic Minimax Optimization", "abs": "In stochastic minimax optimization, variance-reduction techniques have been widely developed to mitigate the inherent variances introduced by stochastic gradients. Most of these techniques employ carefully designed estimators and learning rates, successfully reducing variance. Although these approaches achieve optimal theoretical convergence rates, they require the careful selection of numerous hyperparameters, which heavily depend on problem-dependent parameters. This complexity makes them difficult to implement in practical model training. To address this, our paper introduces Adaptive Filtered Momentum (AdaFM), an adaptive variance-reduced algorithm for stochastic minimax optimization. AdaFM adaptively adjusts hyperparameters based solely on historical estimator information, eliminating the need for manual parameter tuning. Theoretical results show that AdaFM can achieve a near-optimal sample complexity of $O(\\epsilon^{-3})$ to find an $\\epsilon$-stationary point in non-convex-strongly-concave and non-convex-Polyak-\\L ojasiewicz objectives, matching the performance of the best existing non-parameter-free algorithms. Extensive experiments across various applications validate the effectiveness and robustness of AdaFM.", "keywords": [ "varienced reduction; adaptive method; minimax optimization" ], "rating_list": [ 3, 3, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "NgvL7aMaTI", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Albert Albesa-Gonz\u00e1lez", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "PhD student" }, { "name": "Claudia Clopath", "gender": "unknown", "institution": "Imperial College London", "country": "", "position": "Full Professor" } ], "n_formula": 36, "n_formula_1": 8, "n_ref_uni": 30, "n_ref": 54, "n_ref_all": 91, "n_fig": 23, "n_tab": 4, "L_tab": 416, "n_element_tab": 40, "n_fig_1": 19, "n_tab_1": 3, "L_tab_1": 548, "n_element_tab_1": 52, "formula_len_all": 2372, "formula_len_all_1": 318, "len_all": 140296, "len_all_1": 58536, "len_abs": 1611, "len_title": 136, "len_sents": 48902, "len_sents_1": 27863, "n_sents": 340, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1621, "title": "Predicting episodic structure from overlapping input in binary networks with homeostasis", "abs": "How neural networks process overlapping input patterns is a fundamental question in both neuroscience and artificial intelligence. Traditionally, overlaps in neural activity are viewed as interference, requiring separation for better performance. However, an alternative perspective suggests that these overlaps may encode meaningful semantic relationships between concepts. In this paper, we propose a framework where persistent overlap between episodic patterns represent semantic components across episodic experiences, and the statistics of these overlaps how each semantic concept relates to others.\n\nTo explore this idea, we introduce an Episode Generation Protocol (EGP) that defines a mapping between the semantic structure of episodes and input pattern generation. Paired with our EGP, we use Homeostatic Binary Networks (HBNs), a simplified yet biologically-inspired model incorporating key features such as adjustable inhibition, Hebbian learning, and homeostatic plasticity.\n\nOur contributions are threefold: (1) We formalize a link between episodic semantics and neural patterns through our EGP. This EGP can be used for systematic study of semantic learning in artificial neural networks. (2) We introduce HBNs as an analytically tractable network that extracts semantic structure in its internal model (3) We show that HBNs align their performance with Maximum A Posteriori and Maximum Likelihood Estimation strategies depending on the homeostatic regime. Similarly, we provide an example of how our EGP can be used as an experimental protocol in neuroscience to make different models of learning compete.", "keywords": [ "Semantics", "Episodes", "Homeostasis", "Regularization", "Hopfield Network", "Synaptic Plasticity", "Predictive Learning" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "NgMbGDCmAM", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Adel Boyarsky", "gender": "Female", "institution": "Morgan Stanley", "country": "", "position": "Researcher" }, { "name": "Alexander Sabat", "gender": "Male", "institution": "Independent", "country": "US", "position": "Independent Researcher" }, { "name": "Daniel DeDora", "gender": "Male", "institution": "Morgan Stanley", "country": "", "position": "Researcher" }, { "name": "Nicholas M. Venuti", "gender": "Male", "institution": "Morgan Stanley", "country": "US", "position": "Researcher" }, { "name": "Rachel Cayale", "gender": "Female", "institution": "Morgan Stanley", "country": "US", "position": "Researcher" }, { "name": "William Brown", "gender": "Male", "institution": "Morgan Stanley", "country": "US", "position": "Researcher" }, { "name": "Yuriy Nevmyvaka", "gender": "unknown", "institution": "Morgan Stanley", "country": "US", "position": "Principal Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 46, "n_ref": 81, "n_ref_all": 100, "n_fig": 0, "n_tab": 4, "L_tab": 1416, "n_element_tab": 190, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 1336, "n_element_tab_1": 112, "formula_len_all": 440, "formula_len_all_1": 296, "len_all": 112863, "len_all_1": 70660, "len_abs": 1661, "len_title": 117, "len_sents": 31145, "len_sents_1": 28008, "n_sents": 176, "n_sents_1": 132, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1665, "title": "Clustering and Entity Matching via Language Model Community Detection", "abs": "We introduce LMCD, a novel framework for semantic clustering and multi-set entity matching problems, in which we employ graph community detection algorithms to prune spurious edges from match graphs constructed using embedding and language models. We construct these match graphs by retrieving nearest embedding neighbors for each entity, then querying a language model to remove false positive pairs. Across a variety of cluster size distributions, and for tasks ranging from sentiment and topic categorization to deduplication of product databases, our approach outperforms existing methods without requiring any finetuning or labeled data beyond few-shot examples, and without needing to select the desired number of clusters in advance. Our embedding and inference stages are fully parallelizable, with query and computational costs which scale near-linearly in the number of entities. Our post-processing stage is bottlenecked only by the runtime of community detection algorithms on discrete graphs, which are often near-linear, with no explicit dependence on embedding dimension or numbers of clusters. This is in stark contrast to existing methods relying on high-dimensional clustering algorithms that are difficult to apply at scale; for entity matching our approach also ensures consistency constraints across matches regardless of group sizes, a desirable practical feature which is absent from all prior approaches other than vector clustering. Our improvements over previous techniques are most stark when clusters are numerous and heterogenously-sized, a regime which captures many clustering and matching problems of widespread practical importance.", "keywords": [ "language models", "LLMs", "embeddings", "entity matching", "entity resolution", "clustering", "community detection", "knowledge graphs", "vector databases" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "NgCNMlTXx9", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Baolong Bi", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Lingrui Mei", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Ruibin Yuan", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Shenghua Liu", "gender": "Not Specified", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Xueqi Cheng", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy", "country": "CN", "position": "Full Professor" }, { "name": "Yiwei Wang", "gender": "Male", "institution": "University of California, Merced", "country": "US", "position": "Assistant Professor" } ], "n_formula": 46, "n_formula_1": 33, "n_ref_uni": 64, "n_ref": 115, "n_ref_all": 122, "n_fig": 4, "n_tab": 7, "L_tab": 2768, "n_element_tab": 239, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1052, "n_element_tab_1": 110, "formula_len_all": 3453, "formula_len_all_1": 2060, "len_all": 232832, "len_all_1": 63644, "len_abs": 1927, "len_title": 128, "len_sents": 64158, "len_sents_1": 27012, "n_sents": 474, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 2303, "title": "HiddenGuard: Fine-Grained Safe Generation with Specialized Representation Router", "abs": "As Large Language Models (LLMs) grow increasingly powerful, ensuring their safety and alignment with human values remains a critical challenge. Current alignment approaches predominantly rely on refusal alignment, such as training models to refuse harmful prompts or implementing filters at various stages to block certain responses. These methods are designed toward a binary outcome: either denying to answer the question entirely or answering with full access to the model's parametric knowledge.\nThe binary nature of current alignment approaches presents significant limitations. These methods often fail to balance safety and utility, resulting in either overly cautious responses or overlooking subtle harmful content. They also prevent users from accessing benign information when it's mixed with harmful content. For instance, a model might refuse to provide basic, public information about a medication's composition due to misuse concerns.\nFurthermore, these approaches struggle with context-dependent sensitivity, potentially over-censoring harmless content or missing nuanced harmful outputs. Ideally, LLMs should offer informative responses while avoiding the disclosure of harmful and sensitive information.\nTo address these challenges, we introduce HiddenGuard, a novel framework for fine-grained safe generation in LLMs. Our method incorporates PRISM (rePresentation Router for In-Stream Moderation), a specialized moudule that operates alongside the LLM architecture. By leveraging intermediate hidden states, HiddenGuard enables real-time, token-level harmfulness detection and redaction, without loss in capability. This approach captures deeper semantic information, allowing for more nuanced and context-aware content control compared to traditional filtering techniques. Consequently, the model can generate informative responses while selectively redacting or replacing sensitive information, rather than refusing to answer outright.\nWe also contribute a comprehensive dataset with token-level fine-grained annotations of potentially harmful information across diverse contexts. Our experiments demonstrate that HiddenGuard achieves over 90% in F1 score for detecting and redacting harmful content while preserving the overall utility and informativeness of the model's responses.", "keywords": [ "Large Language Models", "AI Safety", "Context-Aware Moderation", "Representation Router" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Ng4HaH4L6P", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Bin Zhang", "gender": "unknown", "institution": "Jinan University", "country": "CN", "position": "Principal Researcher" }, { "name": "Guoan Wang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Jin Ye", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Junjun He", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Nana Pei", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Rongshan Yu", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Tianbin Li", "gender": "Male", "institution": "ShangHai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yanjun Li", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Ying Chen", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuanfeng Ji", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 27, "n_ref": 49, "n_ref_all": 81, "n_fig": 9, "n_tab": 8, "L_tab": 3807, "n_element_tab": 539, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1761, "n_element_tab_1": 230, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 152432, "len_all_1": 61008, "len_abs": 1475, "len_title": 140, "len_sents": 51349, "len_sents_1": 30488, "n_sents": 316, "n_sents_1": 163, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1435, "title": "SlideChat: A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding", "abs": "Despite the progress made by multimodal large language models (MLLMs) in computational pathology, they remain limited by a predominant focus on patch-level analysis, missing essential contextual information at the whole-slide level. The lack of large-scale instruction datasets and the gigapixel scale of whole slide images (WSIs) pose significant developmental challenges. In this paper, we present SlideChat, the first vision-language assistant capable of understanding gigapixel whole-slide images, exhibiting excellent multimodal conversational capability and response complex instruction across diverse pathology scenarios. To support its development, we created SlideInstruction, the largest instruction-following dataset for WSIs consisting of 4.2K WSI captions and 176K VQA pairs with multiple categories. Furthermore, we propose SlideBench, a multimodal benchmark that incorporates captioning and VQA tasks to assess SlideChat\u2019s capabilities in varied clinical settings such as microscopy, diagnosis. Compared to both general and specialized MLLMs, SlideChat exhibits exceptional capabilities, achieving state-of-the-art performance on 18 of 22 tasks. For example, it achieved an overall accuracy of 81.17% on SlideBench-VQA (TCGA), and 54.15% on SlideBench-VQA (BCNB). We will fully release SlideChat, SlideInstruction and SlideBench as open-source resources to facilitate research and development in computational pathology.", "keywords": [ "Computational Pathology", "Whole Slide Image", "Pathology", "Multimodal Large Language Model" ], "rating_list": [ 3, 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 5, 4, 4 ] }, { "paper_id": "Ng1r9kTep4", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 2, "author_info_list": [ { "name": "Georgii Sergeevich Novikov", "gender": "Non-Binary", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Ivan Oseledets", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 14, "n_ref": 18, "n_ref_all": 37, "n_fig": 4, "n_tab": 1, "L_tab": 1617, "n_element_tab": 100, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 308, "n_element_tab_1": 7, "formula_len_all": 961, "formula_len_all_1": 845, "len_all": 110822, "len_all_1": 55292, "len_abs": 1544, "len_title": 75, "len_sents": 29563, "len_sents_1": 26242, "n_sents": 215, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1592, "title": "Inverted Activations: Reducing Memory Footprint in Neural Network Training", "abs": "The scaling of neural networks with increasing data and model sizes necessitates the development of more efficient deep learning algorithms. \n A significant challenge in neural network training is the memory footprint associated with activation tensors, particularly in pointwise nonlinearity layers that traditionally save the entire input tensor for the backward pass, leading to substantial memory consumption.\n \n In this paper, we propose a modification to the handling of activation tensors in pointwise nonlinearity layers. \n Our method involves saving the output tensor instead of the input tensor during the forward pass. Since the subsequent layer typically also saves its input tensor, this approach reduces the total memory required by storing only one tensor between layers instead of two. This optimization is especially beneficial for transformer-based architectures like GPT, BERT, Mistral, and Llama.\n\n To enable this approach, we utilize the inverse function of the nonlinearity during the backward pass. As the inverse cannot be computed analytically for most nonlinearities, we construct accurate approximations using simpler functions. \n Experimental results demonstrate that our method significantly reduces memory usage without affecting training accuracy or computational performance.\n\n Our implementation is provided as a drop-in replacement for standard nonlinearity layers in the PyTorch framework, facilitating easy adoption without requiring architectural modifications. The code is available at \\url{https://github.com/removed/for/anonimity}.", "keywords": [ "deep learning", "large language models" ], "rating_list": [ 6, 5, 8 ], "soundness_list": [ 2, 2, 4 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 4, 2, 3 ], "confidence_list": [ 4, 3, 5 ] }, { "paper_id": "Nfd7z9d6Bb", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Alexander Fishkov", "gender": "unknown", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Eric Moulines", "gender": "Male", "institution": "Ecole polytechnique", "country": "", "position": "Full Professor" }, { "name": "Maxim Panov", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Mohsen Guizani", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" }, { "name": "Vincent Plassier", "gender": "Male", "institution": "Ecole polytechnique", "country": "FR", "position": "PhD student" } ], "n_formula": 133, "n_formula_1": 25, "n_ref_uni": 35, "n_ref": 81, "n_ref_all": 117, "n_fig": 8, "n_tab": 5, "L_tab": 287, "n_element_tab": 66, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 439, "n_element_tab_1": 7, "formula_len_all": 16771, "formula_len_all_1": 3716, "len_all": 218041, "len_all_1": 75845, "len_abs": 1034, "len_title": 120, "len_sents": 55474, "len_sents_1": 28097, "n_sents": 619, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1053, "title": "Probabilistic Conformal Prediction with Approximate Conditional Validity", "abs": "We develop a new method for generating prediction sets that combines the flexibility of conformal methods with an estimate of the conditional distribution $\\textup{P}_{Y \\mid X}$. Existing methods, such as conformalized quantile regression and probabilistic conformal prediction, usually provide only a marginal coverage guarantee. In contrast, our approach extends these frameworks to achieve approximately conditional coverage, which is crucial for many practical applications. Our prediction sets adapt to the behavior of the predictive distribution, making them effective even under high heteroscedasticity. While exact conditional guarantees are infeasible without assumptions on the underlying data distribution, we derive non-asymptotic bounds that depend on the total variation distance of the conditional distribution and its estimate. Using extensive simulations, we show that our method consistently outperforms existing approaches in terms of conditional coverage, leading to more reliable statistical inference in a variety of applications.", "keywords": [ "Conformal Prediction", "Conditional coverage", "Probabilistic method", "Uncertainty Quantification" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "NfUHBaZdLw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Huan Sun", "gender": "Female", "institution": "The Ohio State University, Columbus", "country": "US", "position": "Associate Professor" }, { "name": "Kun Qian", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "SHADDY GARG", "gender": "Male", "institution": "Adobe Systems", "country": "IN", "position": "Researcher" }, { "name": "Siddhartha Sahai", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "TIANSHU ZHANG", "gender": "unknown", "institution": "Ohio State University, Columbus", "country": "", "position": "PhD student" }, { "name": "Yuan Tian", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Yunyao Li", "gender": "unknown", "institution": "Adobe Systems", "country": "", "position": "Director of Machine Learning" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 20, "n_ref": 41, "n_ref_all": 57, "n_fig": 3, "n_tab": 9, "L_tab": 5965, "n_element_tab": 403, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2008, "n_element_tab_1": 158, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 191184, "len_all_1": 64896, "len_abs": 1785, "len_title": 66, "len_sents": 34375, "len_sents_1": 31147, "n_sents": 217, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1791, "title": "EVOSCHEMA: TOWARDS TEXT-TO-SQL ROBUSTNESS AGAINST SCHEMA EVOLUTION", "abs": "Neural text-to-SQL models, which translate natural language questions (NLQs) into SQL queries given a database schema, have achieved remarkable performance. However, database schemas frequently evolve to meet new requirements. Such schema evolution often leads to performance degradation for models trained on static schemas. Existing work either mainly focuses on simply paraphrasing some syntactic or semantic mappings among NLQ, DB and SQL or lacks a comprehensive and controllable way to investigate the model robustness issue\nunder the schema evolution. In this work, we approach this crucial problem by introducing a novel framework, EvoSchema, to systematically simulate diverse schema changes that occur in real-world scenarios. EvoSchema builds on our newly defined schema evolution taxonomy, which encompasses a comprehensive set of eight perturbation types, covering both column-level and table-level modifications. We utilize this framework to build an evaluation benchmark to assess the models\u2019 robustness against different schema evolution types. Meanwhile, we propose a new training paradigm, which augments existing training data with diverse schema designs and forces the model to distinguish the schema difference for the same questions to avoid learning spurious patterns. Our experiments demonstrate that the existing models are more easily affected by table-level perturbations than column-level perturbations. In addition, the models trained under our paradigm exhibit significantly improved robustness, achieving up to 33 points improvement on the evaluation benchmark compared to models trained on unperturbed data. This work represents a significant step towards building more resilient text-to-SQL systems capable of handling the dynamic nature of database schemas.", "keywords": [ "text-to-SQL", "schema evolution", "robustness" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "NfMmo2eQhG", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Deepak Ravikumar", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Efstathia Soufleri", "gender": "Female", "institution": "Athena Research and Innovation Center", "country": "GR", "position": "Postdoc" }, { "name": "Kaushik Roy", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Full Professor" }, { "name": "Manish Nagaraj", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 42, "n_ref": 69, "n_ref_all": 83, "n_fig": 11, "n_tab": 4, "L_tab": 1657, "n_element_tab": 173, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1704, "n_element_tab_1": 42, "formula_len_all": 853, "formula_len_all_1": 461, "len_all": 124939, "len_all_1": 46312, "len_abs": 1450, "len_title": 63, "len_sents": 37794, "len_sents_1": 22212, "n_sents": 250, "n_sents_1": 131, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1532, "title": "The Relevancy Metric: Understanding the Impact of Training Data", "abs": "Deep learning models are central to many critical decision-making processes, making it imperative to gain deeper insights into their behavior to improve performance, transparency, interpretability, and fairness. \nA key challenge is understanding how training data shapes model predictions on unseen test data. \nIn this paper, we introduce a novel metric, $\\textbf{\\textit{Relevancy}}$, which quantifies the impact of individual training samples on inference predictions. \nOur proposed metric is calculated by observing the learning dynamics of the model during training, and it is computationally efficient and applicable across a wide range of tasks. \nWe demonstrate that it is between $80\\times$ and $100,000\\times$ more efficient than existing metrics for capturing the train-test relationship. \nUsing $\\textit{relevancy}$, we enable the identification of coresets \u2014 compact datasets that represent the essence of the training distribution. \nQuantitative evaluations show that coresets selected using our metric outperform state-of-the-art methods by up to $5.2$% on CIFAR-100. \nAdditionally, we qualitatively demonstrate how $\\textit{relevancy}$ can be extended to assess various training data properties, such as identifying mislabeled samples in widely used datasets like ImageNet, CIFAR-100, and Fashion-MNIST.\nThese examples illustrate just a few of the many potential uses of $\\textit{relevancy}$, highlighting its versatility in promoting more interpretable, efficient, and fair deep learning systems across diverse tasks.", "keywords": [ "Train-Test Relationship", "Influence functions", "Memorization", "Learning dynamics", "Dataset properties" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "NfCEVihkdC", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Balaji Krishnamurthy", "gender": "Male", "institution": "Adobe Systems", "country": "IN", "position": "Principal Scientist" }, { "name": "Harini S I", "gender": "Female", "institution": "Adobe Systems", "country": "IN", "position": "Researcher" }, { "name": "Somesh Kumar Singh", "gender": "Male", "institution": "Indraprastha Institute of Information Technology, Delhi", "country": "IN", "position": "PhD student" }, { "name": "Yaman Kumar", "gender": "Male", "institution": "Adobe ", "country": "IN", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 69, "n_ref": 169, "n_ref_all": 220, "n_fig": 9, "n_tab": 18, "L_tab": 7800, "n_element_tab": 1236, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 3892, "n_element_tab_1": 233, "formula_len_all": 357, "formula_len_all_1": 0, "len_all": 335858, "len_all_1": 75017, "len_abs": 2619, "len_title": 111, "len_sents": 123301, "len_sents_1": 35924, "n_sents": 899, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 2405, "title": "Measuring And Improving Persuasiveness Of Generative Models", "abs": "Large Language Models (LLMs) are increasingly being used in workflows involving generating content to be consumed by humans (e.g., marketing) and also in directly interacting with humans (e.g., through chatbots). The development of such systems that are capable of generating verifiably persuasive messages presents both opportunities and challenges for society. On the one hand, such systems could\npositively impact domains like advertising and social good, such as addressing drug addiction, and on the other, they could be misused for spreading misinformation and shaping political opinions. To channel LLMs\u2019 impact on society, we need to develop systems to measure and benchmark their persuasiveness. With this motivation, we introduce PersuasionBench and PersuasionArena, the first largescale\nbenchmark and arena containing a battery of tasks to measure the persuasion ability of generative models automatically. We introduce transsuasion (trans = carrying across, suasion = the act of persuading), a novel task of transforming non-persuasive language into persuasive content while preserving other factors determining persuasiveness (sender, receiver, time, and channel). To construct data\nfor transsuasion, we leverage natural experiments in the form of a pair of tweets from the same user, posted in close temporal proximity, with similar semantic content but divergent wording and significantly different like counts. Given such pairs, we investigate to what extent LLMs know and leverage linguistic patterns that can help them generate more persuasive language. Our findings indicate that the\npersuasiveness of LLMs correlates positively with model size, but smaller models can also be made to have a higher persuasiveness than much larger models. Notably, targeted training using synthetic and natural datasets significantly enhances smaller models\u2019 persuasive capabilities, challenging scale-dependent assumptions. Our findings carry key implications for both model developers and policymakers. For instance, while California\u2019s SB-1047 aims to regulate AI models based on the number of floating point operations, we demonstrate that simple metrics like this alone fail to capture the full scope of AI\u2019s societal impact. We invite the community to explore and contribute to PersuasionArena and PersuasionBench, to advance our understanding of AI-driven persuasion and its societal implications.", "keywords": [ "llm", "transsuasion", "persuasion", "PersuasionArena", "PersuasionBench" ], "rating_list": [ 8, 6, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 1, 1 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "Neb17mimVH", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Jeff Bilmes", "gender": "Male", "institution": "University of Washington, Seattle", "country": "US", "position": "Full Professor" }, { "name": "Tian Li", "gender": "unknown", "institution": "University of Chicago", "country": "US", "position": "Assistant Professor" }, { "name": "Tianyi Zhou", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Virginia Smith", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Ziyue Li", "gender": "Female", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" } ], "n_formula": 71, "n_formula_1": 21, "n_ref_uni": 40, "n_ref": 96, "n_ref_all": 139, "n_fig": 12, "n_tab": 12, "L_tab": 3134, "n_element_tab": 249, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 216, "n_element_tab_1": 29, "formula_len_all": 4996, "formula_len_all_1": 1090, "len_all": 195205, "len_all_1": 72579, "len_abs": 1288, "len_title": 87, "len_sents": 67470, "len_sents_1": 34222, "n_sents": 554, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 39, "L_abs": 1295, "title": "Many-Objective Multi-Solution Transport", "abs": "Optimizing the performance of many objectives (instantiated by tasks or clients) jointly with a few Pareto stationary solutions (models) is critical in machine learning. However, previous multi-objective optimization methods often focus on a few objectives and cannot scale to many objectives that outnumber the solutions, leading to either subpar performance or ignored objectives. We introduce ''Many-objective multi-solution Transport (MosT)'', a framework that finds multiple diverse solutions in the Pareto front of many objectives. Our insight is to seek multiple solutions, each performing as a domain expert and focusing on a specific subset of objectives while collectively covering all of them. MosT formulates the problem as a bi-level optimization of weighted objectives for each solution, where the weights are defined by an optimal transport between objectives and solutions. Our algorithm ensures convergence to Pareto stationary solutions for complementary subsets of objectives. On a range of applications in federated learning, multi-task learning, and mixture-of-prompt learning for LLMs, MosT distinctly outperforms strong baselines, delivering high-quality, diverse solutions that profile the entire Pareto frontier, thus ensuring balanced trade-offs across many objectives.", "keywords": [ "Multi-Objective Optimization", "Mixture of Experts" ], "rating_list": [ 5, 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 4 ], "contribution_list": [ 2, 3, 3, 2, 4 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "NeVbEYW4tp", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Chen Xu", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "MS student" }, { "name": "Gangshan Wu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Guozhen Zhang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Haocheng Shen", "gender": "Male", "institution": "Vivo AI Research", "country": "CN", "position": "Researcher" }, { "name": "Xiaoxin Chen", "gender": "Male", "institution": "vivo AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Yuhan Zhu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Limin Wang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 71, "n_ref": 132, "n_ref_all": 147, "n_fig": 8, "n_tab": 9, "L_tab": 5626, "n_element_tab": 633, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 3789, "n_element_tab_1": 467, "formula_len_all": 986, "formula_len_all_1": 961, "len_all": 183607, "len_all_1": 73956, "len_abs": 1616, "len_title": 108, "len_sents": 35681, "len_sents_1": 28259, "n_sents": 305, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1644, "title": "Efficient Test-Time Prompt Tuning for Vision-Language Models", "abs": "Vision-language models have showcased impressive zero-shot classification capabilities when equipped with suitable text prompts. Previous studies have shown the effectiveness of test-time prompt tuning; however, these methods often require per-image prompt adaptation during inference, which is computationally intensive and limits scalability and deployment.\nTo address this issue, we introduce a novel framework: Self-supervised learning for efficient Test-time Prompt Tuning (Self-TPT).\nThe key feature of Self-TPT is its shift to efficient \\textit{predefined class adaptation} through self-supervised learning, thereby avoiding the computation-heavy \\textit{per-image adaptation} at inference.\nSelf-TPT starts by co-training the self-supervised and supervised tasks using source data, then applies the self-supervision exclusively for new class understanding before making predictions.\nSpecifically, we propose Contrastive Prompt Learning (CPT) as the core task for self-supervision. CPT is designed to minimize the intra-class distances while enhancing inter-class distinguishability via contrastive learning.\nEmpirical evidence suggests that CPT can partially mimic supervised learning in terms of gradients, providing a plausible explanation for its effectiveness.\nMotivated by this finding, we introduce a gradient matching loss to explicitly enhance gradient similarity.\nWe evaluated Self-TPT across three challenging zero-shot benchmarks. The results consistently show that Self-TPT not only significantly reduces inference costs but also achieves state-of-the-art performance, effectively balancing the efficiency-efficacy trade-off.", "keywords": [ "Vision-Language Models", "Zero-shot Generalization", "Prompt Learning", "Test-Time Adaptation" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "NdxI9Yx9f2", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Arthur Aubret", "gender": "unknown", "institution": "Frankfurt Institute for Advanced Studies ", "country": "DE", "position": "Postdoc" }, { "name": "C\u00e9line Teuli\u00e8re", "gender": "unknown", "institution": "Universit\u00e9 Clermont Auvergne", "country": "", "position": "Assistant Professor" }, { "name": "Jochen Triesch", "gender": "Male", "institution": "Goethe University", "country": "DE", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 56, "n_ref": 120, "n_ref_all": 149, "n_fig": 7, "n_tab": 10, "L_tab": 2400, "n_element_tab": 532, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1304, "n_element_tab_1": 383, "formula_len_all": 522, "formula_len_all_1": 414, "len_all": 163799, "len_all_1": 66095, "len_abs": 938, "len_title": 120, "len_sents": 42014, "len_sents_1": 28222, "n_sents": 337, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 941, "title": "Seeing the Whole in the Parts in Self-Supervised Representation Learning", "abs": "Recent successes in self-supervised learning (SSL) model spatial co-occurrences of visual features either by masking portions of an image or by aggressively cropping it. Here, we propose a new way to model spatial co-occurrences by aligning local representations (before pooling) with a global image representation. We present CO-SSL, a family of instance discrimination methods and show that it outperforms previous methods on several datasets, including ImageNet-1K where it achieves 71.5% of Top-1 accuracy with 100 pre-training epochs. CO-SSL is also more robust to noise corruption, internal corruption, small adversarial attacks, and large training crop sizes. Our analysis further indicates that CO-SSL learns highly redundant local representations, which offers an explanation for its robustness. Overall, our work suggests that aligning local and global representations may be a powerful principle of unsupervised category learning.", "keywords": [ "Self-supervised learning", "statistical learning", "bio-inspired learning", "infomax" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "Ndq4g76MyH", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Lei Li", "gender": "Not Specified", "institution": "University of Washington", "country": "US", "position": "Researcher" }, { "name": "Sen Jia", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 4, "n_ref": 9, "n_ref_all": 19, "n_fig": 4, "n_tab": 4, "L_tab": 3212, "n_element_tab": 111, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 764, "n_element_tab_1": 57, "formula_len_all": 900, "formula_len_all_1": 900, "len_all": 93002, "len_all_1": 48619, "len_abs": 1690, "len_title": 91, "len_sents": 27066, "len_sents_1": 22440, "n_sents": 190, "n_sents_1": 161, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 1731, "title": "Adaptive Masking Enhances Visual Grounding", "abs": "In recent years, zero-shot and few-shot learning in visual grounding have garnered considerable attention, largely due to the success of large-scale vision-language pre-training on expansive datasets such as LAION-5B and DataComp-1B. However, the continuous expansion of these datasets presents significant challenges, particularly with respect to data availability and computational overhead, thus creating a bottleneck in the advancement of low-shot learning capabilities. In this paper, we propose a novel approach, \\textbf{I}nterpretative \\textbf{MA}sking with \\textbf{G}aussian Radiation Mod\\textbf{E}ling, aimed at enhancing vocabulary grounding in low-shot learning scenarios without necessitating an increase in dataset size. Drawing inspiration from cognitive science and the recent success of masked autoencoders (MAE), our method leverages adaptive masking on salient regions of the feature maps generated by the vision backbone. This enables the model to learn robust, generalized representations through the reconstruction of occluded information, thereby facilitating effective attention to both local and global features. We evaluate the efficacy of our approach on benchmark datasets, including COCO and ODinW, demonstrating its superior performance in zero-shot and few-shot tasks. Experimental results consistently show that IMAGE outperforms baseline models, achieving enhanced generalization and improved performance in low-shot scenarios. These findings highlight the potential of adaptive feature manipulation through attention mechanisms and Gaussian modeling as a promising alternative to approaches that rely on the continual scaling of dataset sizes for the advancement of zero-shot and few-shot learning.", "keywords": [ "Vocabulary Grounding", "multimodal" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 1 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "NdNuKMEv9y", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Bo Liu", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Lizhang Chen", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Son Nguyen", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Qiang Liu", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Assistant Professor" } ], "n_formula": 28, "n_formula_1": 21, "n_ref_uni": 38, "n_ref": 57, "n_ref_all": 73, "n_fig": 8, "n_tab": 4, "L_tab": 817, "n_element_tab": 111, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3710, "formula_len_all_1": 2045, "len_all": 116198, "len_all_1": 57032, "len_abs": 1260, "len_title": 121, "len_sents": 33688, "len_sents_1": 25459, "n_sents": 270, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1514, "title": "Improving Adaptive Moment Optimization via Preconditioner Diagonalization", "abs": "Modern adaptive optimization methods, such as Adam and its variants, have emerged as the most widely used tools in deep learning over recent years. These algorithms offer automatic mechanisms for dynamically adjusting the update step based on estimates of gradient statistics. Compared to traditional algorithms like Stochastic Gradient Descent, these adaptive methods are typically more robust to model scale and hyperparameter tuning. However, the gradient statistics employed by these methods often do not leverage sufficient gradient covariance information, leading to suboptimal updates in certain directions of the parameter space and potentially slower convergence. In this work, we keep track of such covariance statistics in the form of a structured preconditioner matrix. Unlike other works, our approach does not apply direct approximations to estimate this matrix. We instead implement an invertible transformation that maps the preconditioner matrix into a new space where it becomes approximately diagonal. This enables a diagonal approximation of the preconditioner matrix in the transformed space, offering several computational advantages. Empirical results show that our approach can substantially enhance the convergence speed of the modern adaptive optimizers. Notably, for large language models like LLaMA, we can achieve a speedup of 2x compared to the baseline Adam. Additionally, our method can be integrated with memory-efficient optimizers like Adafactor to manage computational overhead.", "keywords": [ "optimization", "deep learning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "NdHka08uWn", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 5, "author_info_list": [ { "name": "Bing XUE", "gender": "Female", "institution": "Victoria University of Wellington", "country": "NZ", "position": "Full Professor" }, { "name": "Hengzhe Zhang", "gender": "Male", "institution": "Victoria University of Wellington", "country": "NZ", "position": "PhD student" }, { "name": "Mengjie Zhang", "gender": "Male", "institution": "Victoria University of Wellington", "country": "NZ", "position": "Full Professor" }, { "name": "Qi Chen", "gender": "Female", "institution": "Victoria University of Wellington", "country": "NZ", "position": "Lecturer" }, { "name": "Wolfgang Banzhaf", "gender": "Male", "institution": "Michigan State University", "country": "", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 35, "n_ref": 61, "n_ref_all": 78, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 405, "n_element_tab_1": 27, "formula_len_all": 1578, "formula_len_all_1": 1449, "len_all": 123301, "len_all_1": 69130, "len_abs": 1378, "len_title": 117, "len_sents": 34430, "len_sents_1": 32703, "n_sents": 227, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 69, "L_abs": 1383, "title": "RAG-SR: Retrieval-Augmented Generation for Neural Symbolic Regression", "abs": "Symbolic regression is a key task in machine learning, aiming to discover mathematical expressions that best describe a dataset. While deep learning has increased interest in using neural networks for symbolic regression, many existing approaches rely on pre-trained models. These models require significant computational resources and struggle with regression tasks involving unseen functions and variables. A pre-training-free paradigm is needed to better integrate with search-based symbolic regression algorithms. To address these limitations, we propose a novel framework for symbolic regression that integrates evolutionary feature construction with a neural network, without the need for pre-training. Our approach adaptively generates symbolic trees that align with the desired semantics in real-time using a language model trained via online supervised learning, providing effective building blocks for feature construction. To mitigate hallucinations from the language model, we design a retrieval-augmented generation mechanism that explicitly leverages searched symbolic expressions. Additionally, we introduce a scale-invariant data augmentation technique that further improves the robustness and generalization of the model. Experimental results demonstrate that our framework achieves state-of-the-art accuracy across 25 regression algorithms and 120 regression tasks.", "keywords": [ "Symbolic Regression", "Genetic Programming", "Transformers", "Deep Learning" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "Ncx0X8lcN1", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Boxin Zhao", "gender": "Male", "institution": "Emory University", "country": "", "position": "Undergrad student" }, { "name": "Liang Zhao", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Associate Professor" }, { "name": "Raasikh Kanjiani", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Undergrad student" }, { "name": "Yuntong Hu", "gender": "Male", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Zheng Zhang", "gender": "Male", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Zhuofeng Li", "gender": "unknown", "institution": "Shanghai University", "country": "CN", "position": "Undergrad student" }, { "name": "Chen LING", "gender": "Male", "institution": "Emory University", "country": "US", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 17, "n_ref_uni": 52, "n_ref": 72, "n_ref_all": 88, "n_fig": 5, "n_tab": 6, "L_tab": 4104, "n_element_tab": 224, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1090, "n_element_tab_1": 86, "formula_len_all": 1065, "formula_len_all_1": 1079, "len_all": 174618, "len_all_1": 76213, "len_abs": 1431, "len_title": 126, "len_sents": 55977, "len_sents_1": 35512, "n_sents": 442, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1525, "title": "HiReview: Hierarchical Taxonomy-Driven Automatic Literature Review Generation", "abs": "In this work, we present HiReview, a novel framework for hierarchical taxonomy-driven automatic literature review generation. With the exponential growth of academic documents, manual literature reviews have become increasingly labor-intensive and time-consuming, while traditional summarization models struggle to generate comprehensive document reviews effectively. Large language models (LLMs), with their powerful text processing capabilities, offer a potential solution; however, research on incorporating LLMs for automatic document generation remains limited. To address key challenges in large-scale automatic literature review generation (LRG), we propose a two-stage taxonomy-then-generation approach that combines graph-based hierarchical clustering with retrieval-augmented LLMs. First, we retrieve the most relevant sub-community within the citation network, then generate a hierarchical taxonomy tree by clustering papers based on both textual content and citation relationships. In the second stage, an LLM generates coherent and contextually accurate summaries for clusters or topics at each hierarchical level, ensuring comprehensive coverage and logical organization of the literature. Extensive experiments demonstrate that HiReview significantly outperforms state-of-the-art methods, achieving superior hierarchical organization, content relevance, and factual accuracy in automatic literature review generation tasks. The code and dataset are available at https://anonymous.4open.science/r/HiReivew-767D.", "keywords": [ "Literature review generation", "retrieval-augmented generation", "hierarchical graph clustering" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "NceWCqEIV3", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Chao Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Assistant Professor" }, { "name": "Chuan Qin", "gender": "Male", "institution": "CNIC, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Dazhong Shen", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "Assistant Professor" }, { "name": "Hui Xiong", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Yov Ren", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 27, "n_ref": 66, "n_ref_all": 76, "n_fig": 5, "n_tab": 7, "L_tab": 1326, "n_element_tab": 200, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 1782, "n_element_tab_1": 272, "formula_len_all": 1078, "formula_len_all_1": 989, "len_all": 118337, "len_all_1": 72034, "len_abs": 1314, "len_title": 155, "len_sents": 36546, "len_sents_1": 31987, "n_sents": 261, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1231, "title": "Enhancing Conversational Recommender Systems with Tree-Structured Knowledge and Pretrained Language Models", "abs": "Conversational recommender systems (CRS) have emerged as a key enhancement to traditional recommendation systems, offering interactive and explainable recommendations through natural dialogue.\nRecent advancements in pretrained language models (PLMs) have significantly improved the conversational capabilities of CRS, enabling more fluent and context-aware interactions. \nHowever, PLMs still face challenges, including hallucinations\u2014where the generated content can be factually inaccurate\u2014and difficulties in providing precise, entity-specific recommendations.\nTo address these challenges, we propose the PCRS-TKA framework, which integrates PLMs with knowledge graphs (KGs) through prompt-based learning. By incorporating tree-structured knowledge from KGs, our framework grounds the PLM in factual information, thereby enhancing the accuracy and reliability of the recommendations. Additionally, we design a user preference extraction module to improve the personalization of recommendations and introduce an alignment module to ensure semantic consistency between dialogue text and KG data. Extensive experiments demonstrate that PCRS-TKA outperforms existing methods in both recommendation accuracy and conversational fluency.", "keywords": [ "Conversational Recommendation Systems", "Knowledge Graph", "Pretrained Language Model" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "NcKUcd4EkA", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Defu Lian", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Enhong Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Hantao Yang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Hong Xie", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" }, { "name": "Xutong Liu", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Postdoc" } ], "n_formula": 22, "n_formula_1": 20, "n_ref_uni": 24, "n_ref": 54, "n_ref_all": 58, "n_fig": 1, "n_tab": 1, "L_tab": 62, "n_element_tab": 5, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1280, "formula_len_all_1": 792, "len_all": 110272, "len_all_1": 73063, "len_abs": 1857, "len_title": 132, "len_sents": 36418, "len_sents_1": 34066, "n_sents": 303, "n_sents_1": 291, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1583, "title": "Harnessing Query Heterogeneity for Cost-Effective Proactive Caching in LLM Inference", "abs": "As Large Language Models (LLMs) significantly enhance the capabilities of AI systems, the increasing volume of query processing requests presents challenges for cost-effective inference, particularly due to repetitive queries that lead to unnecessary resource consumption and increased costs. Caching strategies are employed to store a small set of previous queries, enabling direct retrieval of repetitive queries without reprocessing by the LLMs. However, existing caching algorithms often assume uniform query lengths, simplifying cache selection to a top-$K$ problem, which is inadequate for real-world scenarios with heterogeneous lengths. To address this issue, we propose a bandit learning algorithm for proactive query caching in LLMs, specifically considering variable-sized queries. We cast the optimal cache query cache problem as a knapsack problem. Since the repetitive pattern and processing cost are unknown and has uncertainty, we cast the learning-to-cache problem as a bandit learning problem. Compared to conventional bandit learning frameworks, a new technical challenge is that the reward of an arm would not be observed if it is pulled. To tackle this, we propose an Lower confidence bound (LCB)-type algorithm, which we prove has a $\\tilde{O}(\\sqrt{T})$ order of regret and show that our regret does not deteriorate compared to previous results when incorporating a variable size setting. Furthermore, we demonstrate that our online cache policy effectively reduces the additional computational overhead typically associated with calculating the optimal cache.", "keywords": [ "Query Cache", "LLM Inference Serving", "Bandit Learning" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "NbgODSFW3q", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Mohamed Elhoseiny", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Xiaoqian Shen", "gender": "Female", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 36, "n_ref": 127, "n_ref_all": 143, "n_fig": 8, "n_tab": 4, "L_tab": 1214, "n_element_tab": 198, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 517, "n_element_tab_1": 75, "formula_len_all": 303, "formula_len_all_1": 303, "len_all": 141117, "len_all_1": 56530, "len_abs": 1310, "len_title": 138, "len_sents": 41480, "len_sents_1": 26777, "n_sents": 328, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1478, "title": "ReferPix2Pix: Guiding Multi-Modal LLMs for Image Editing with Referential Pixel Grounding", "abs": "Instruction-based image editing methods allow user-friendly instruction to enhance controllability via natural command. However, without a user-provided mask, existing methods could not identify and edit specific objects if multiple similar instances exist, such as \\textit{``add the man on the right a hat''}. Furthermore, the iterative nature of the editing process may inherently involve ambiguous references from users, such as \\textit{`change it to blue'}, posing challenges in identifying the target without a contextual understanding. Multimodal large language models (MLLMs) offer impressive cross-modal comprehension and co-reference resolution capabilities. In this work, we present \\emph{ReferPix2Pix}, which leverages MLLMs to interpret editing instructions and provide regions of interest (RoI) for precise editing. Such pixel-grounded guidance from MLLMs enhances comprehension of referring expressions and resolves ambiguous references that facilitate localized editing of editing models. Additionally, we developed CoReferEdit benchmark to evaluate editing capabilities across iterative editing phases with multimodal co-references. Our comprehensive experiments show that our approach significantly enhances editing capability in referring and co-referential editing tasks. Our code and data will be made publicly available\\footnote{Please refer to the \\href{https://anonymous.4open.science/r/ReferPix2Pix}{anonymous webpage} for code and qualitative results.}.", "keywords": [ "Image Editing; Multimodal Large Language Models; Referring Expression Comprehension; Multimodal Coreference Resolution" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "NbbsRnPBoS", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Archish S", "gender": "Male", "institution": "Research, Microsoft", "country": "IN", "position": "Intern" }, { "name": "Arun Rajkumar", "gender": "Male", "institution": "Indian Institute of Technology Madras", "country": "", "position": "Assistant Professor" }, { "name": "Chandra Shekar Lakshminarayanan", "gender": "Male", "institution": "Indian Institute of Technology, Madras", "country": "IN", "position": "Assistant Professor" }, { "name": "Harish Guruprasad Ramaswamy", "gender": "Male", "institution": "Indian Institute of Technology Madras,", "country": "", "position": "Assistant Professor" } ], "n_formula": 37, "n_formula_1": 30, "n_ref_uni": 5, "n_ref": 40, "n_ref_all": 71, "n_fig": 4, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 75, "n_element_tab_1": 1, "formula_len_all": 2816, "formula_len_all_1": 2562, "len_all": 88620, "len_all_1": 69872, "len_abs": 2245, "len_title": 120, "len_sents": 34453, "len_sents_1": 28142, "n_sents": 304, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1743, "title": "Faster Gradient Descent in Deep Linear Networks: The Advantage of Depth", "abs": "Gradient descent dynamics in deep linear networks has been studied under a wide range of settings. These studies have reported some negative results on the role of depth, in that, gradient descent in deep linear networks: (i) can take exponential number of iterations to converge, (ii) can exhibit sigmoidal learning, i.e., almost no learning in initial phase followed by rapid learning, (iii) can delay convergence with increase in depth. Some of these results are also under stronger assumptions such as whitened data and balanced initialisation. These messages from prior works suggest that depth hurts the speed of convergence.\n\nIn this paper, we argue that the negative role of depth in the prior works is due to certain pitfalls which can be carefully avoided. We give a positive message on the role of depth, i.e., seen as an additional resource, depth can always be used to speed up convergence. For this purpose, we consider scalar regression with quadratic loss. In this setting, we propose a novel aligned gradient descent (AGD) algorithm for which we show that (i) linear convergence is always possible (ii) depth accelerates the speed of convergence. In AGD, feature alignment happens in first layer and the deeper layers accelerate by learning the right scale. We show acceleration in AGD happens in finite time for unwhitened data. We provide insights into the {acceleration} mechanism and also show that acceleration happens in phases. We also demonstrate the acceleration due to AGD on synthetic and benchmark datasets. Our main message is not propose AGD as a new algorithm in itself, but to demonstrate that depth is an advantage in linear networks thereby dispelling some of the past negative results on the role of depth.", "keywords": [ "Deep Linear Network; Gradient Descent; Faster Convergence in Finite Time" ], "rating_list": [ 3, 1, 3 ], "soundness_list": [ 3, 1, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "Nb7Akh3SjN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Hanlin Gu", "gender": "Male", "institution": "webank", "country": "CN", "position": "Researcher" }, { "name": "Lixin Fan", "gender": "Male", "institution": "WeBank", "country": "CN", "position": "Principal Researcher" }, { "name": "Qiang Yang", "gender": "Male", "institution": "Webank", "country": "CN", "position": "Principal Researcher" }, { "name": "Xinyuan Zhao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yuxing Han", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Professor" } ], "n_formula": 40, "n_formula_1": 25, "n_ref_uni": 30, "n_ref": 59, "n_ref_all": 78, "n_fig": 7, "n_tab": 4, "L_tab": 333, "n_element_tab": 31, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 167, "n_element_tab_1": 15, "formula_len_all": 2471, "formula_len_all_1": 1655, "len_all": 110352, "len_all_1": 52300, "len_abs": 897, "len_title": 55, "len_sents": 32141, "len_sents_1": 22736, "n_sents": 246, "n_sents_1": 152, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 901, "title": "Disentangling data distribution for Federated Learning", "abs": "Federated Learning (FL) facilitates collaborative training of a global model whose performance is boosted by private data owned by distributed clients, without compromising data privacy. Yet the wide applicability of FL is hindered by entanglement of data distributions across different clients. This paper demonstrates for the first time that by disentangling data distributions FL can in principle achieve efficiencies comparable to those of distributed systems, requiring only one round of communication. \nTo this end, we propose a novel FedDistr algorithm, which employs stable diffusion models to decouple and recover data distributions. Empirical results on the CIFAR100 and DomainNet datasets show that FedDistr significantly enhances model utility and efficiency in both disentangled and near-disentangled scenarios while ensuring privacy, outperforming traditional federated learning methods.", "keywords": [ "Disentangled", "Federated learning", "Communication efficiency", "Diffusion model" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "Nb3a8aUGfj", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Amir Barati Farimani", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Anthony Zhou", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "John R Buchanan Jr", "gender": "unknown", "institution": "Naval Nuclear Laboratory", "country": "US", "position": "Principal Researcher" }, { "name": "Michael Schneier", "gender": "unknown", "institution": "University of Pittsburgh", "country": "US", "position": "Postdoc" }, { "name": "zijie li", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 72, "n_ref": 141, "n_ref_all": 171, "n_fig": 6, "n_tab": 11, "L_tab": 1351, "n_element_tab": 187, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 693, "n_element_tab_1": 84, "formula_len_all": 524, "formula_len_all_1": 442, "len_all": 230557, "len_all_1": 69294, "len_abs": 1441, "len_title": 115, "len_sents": 71171, "len_sents_1": 33353, "n_sents": 496, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1451, "title": "Text2PDE: Latent Diffusion Models for Accessible Physics Simulation", "abs": "Recent advances in deep learning have inspired numerous works on data-driven solutions to partial differential equation (PDE) problems. These neural PDE solvers can often be much faster than their numerical counterparts; however, each presents its unique limitations and generally balances training cost, numerical accuracy, and ease of applicability to different problem setups. To address these limitations, we introduce several methods to apply latent diffusion models to physics simulation. Firstly, we introduce a mesh autoencoder to compress arbitrarily discretized PDE data, allowing for efficient diffusion training across various physics. Furthermore, we investigate full spatiotemporal solution generation to mitigate autoregressive error accumulation. Lastly, we investigate conditioning on initial physical quantities, as well as conditioning solely on a text prompt to introduce text2PDE generation. We show that language can be a compact, interpretable, and accurate modality for generating physics simulations, paving the way for more usable and accessible PDE solvers. Through experiments on both uniform and structured grids, we show that the proposed approach is competitive with current neural PDE solvers in both accuracy and efficiency, with promising scaling behavior up to $\\sim$3 billion parameters. By introducing a scalable, accurate, and usable physics simulator, we hope to bring neural PDE solvers closer to practical use.", "keywords": [ "AI4Science", "PDE", "Neural Operator", "Latent Diffusion", "Text2PDE" ], "rating_list": [ 3, 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3, 1 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 5, 3, 5, 5, 3 ] }, { "paper_id": "Nazzz5GJ4g", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Chao Du", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Lei Li", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Tianyu Pang", "gender": "Male", "institution": "Sea AI Lab", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Xianjun Yang", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Xuandong Zhao", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Yu-Xiang Wang", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "William Yang Wang", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 66, "n_ref": 136, "n_ref_all": 169, "n_fig": 6, "n_tab": 8, "L_tab": 3111, "n_element_tab": 354, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2138, "n_element_tab_1": 305, "formula_len_all": 339, "formula_len_all_1": 410, "len_all": 217819, "len_all_1": 67228, "len_abs": 1146, "len_title": 166, "len_sents": 61880, "len_sents_1": 29056, "n_sents": 508, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1092, "title": "Weak-to-Strong Jailbreaking on Large Language Models", "abs": "Large language models (LLMs) are vulnerable to jailbreak attacks -- resulting in harmful, unethical, or biased text generations. However, existing jailbreaking methods are computationally costly. In this paper, we propose the weak-to-strong jailbreaking attack, an efficient method to attack aligned LLMs to produce harmful text. Our key intuition is based on the observation that jailbroken and aligned models only differ in their initial decoding distributions. The weak-to-strong attack's key technical insight is using two smaller models (a safe and an unsafe one) to adversarially modify a significantly larger safe model's decoding probabilities. We evaluate the weak-to-strong attack on 5 diverse LLMs from 3 organizations. The results show our method can increase the misalignment rate to over 99% on two datasets with just one forward pass per example. Our study exposes an urgent safety issue that needs to be addressed when aligning LLMs. As an initial attempt, we propose a defense strategy to protect against such attacks, but creating more advanced defenses remains challenging.", "keywords": [ "LLM", "AI safety", "Jailbreaking" ], "rating_list": [ 6, 6, 6, 3, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 1, 3 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "Na28j1Drh7", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "JinYeong Bak", "gender": "Male", "institution": "SungKyunKwan University", "country": "KR", "position": "Associate Professor" }, { "name": "Jing Yao", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Sooyung Choi", "gender": "Female", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "MS student" }, { "name": "Xiaoyuan Yi", "gender": "Male", "institution": "Research, Microsoft", "country": "CN", "position": "Senior Researcher" }, { "name": "Xin Xie", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Senior Principal Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 40, "n_ref": 64, "n_ref_all": 81, "n_fig": 1, "n_tab": 6, "L_tab": 815, "n_element_tab": 54, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 827, "n_element_tab_1": 45, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 163152, "len_all_1": 62052, "len_abs": 1209, "len_title": 117, "len_sents": 50592, "len_sents_1": 31195, "n_sents": 369, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 115, "L_abs": 1218, "title": "Why Do You Answer Like That? Psychological Analysis on Underlying Connections between LLM's Values and Safety Risks", "abs": "The application scope of Large Language Models (LLMs) continues to expand, leading to increasing interest in personalized LLMs. However, aligning these models with individual values raises significant safety concerns due to harmful information correlated with certain values. In this paper, we identify specific safety risks in value-aligned LLMs and investigate the psychological principles behind these challenges. Our findings reveal two key insights. First, value-aligned LLMs are more prone to harmful behavior compared to non-fine-tuned models and exhibit slightly higher risks in traditional safety evaluations than other fine-tuned models. Second, these safety issues arise because value-aligned LLMs genuinely understand and act according to the aligned values, which can amplify harmful outcomes. Using a dataset with detailed safety categories, we find significant correlations between value alignment and safety concerns, supported by psychological hypotheses. This study offers insights into the ``black box'' of value alignment and proposes enhancing the safety of value-aligned LLMs by corresponding in-context alignment methods.\nWarning: This paper contains contents that may be offensive or upsetting.", "keywords": [ "Value Alignment", "Personalized LLMs", "AI Safety", "Phychological Analysis" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "Na0OnR5WxD", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Shanmin Pang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "WenZhao Qiu", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xiuxiu Bai", "gender": "Female", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Zehao Wu", "gender": "unknown", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xu Lizhen", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 40, "n_ref": 108, "n_ref_all": 131, "n_fig": 5, "n_tab": 6, "L_tab": 1232, "n_element_tab": 142, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 310, "n_element_tab_1": 51, "formula_len_all": 506, "formula_len_all_1": 508, "len_all": 136354, "len_all_1": 59994, "len_abs": 1566, "len_title": 127, "len_sents": 36306, "len_sents_1": 27116, "n_sents": 302, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1421, "title": "Redundant Queries in DETR-Based 3D Detection Methods: Unnecessary and Prunable", "abs": "Query-based models are extensively used in 3D object detection tasks, with a wide range of pre-trained checkpoints readily available online. However, despite their popularity, these models often require an excessive number of object queries, far surpassing the actual number of objects to detect. The redundant queries result in unnecessary computational and memory costs. In this paper, we find that not all queries contribute equally --- a significant portion of queries have a much smaller impact compared to others. Based on this observation, we propose an embarrassingly simple approach that Gradually Prunes Queries (GPQ) according to classification scores that queries generated. Compared to existing pruning methods, our method introduces no additional learnable parameters. GPQ is easy to implement to any query-based method by integrating it in after-training fine-tune using an existing checkpoint. By using our method, one can easily generate several different models with fewer queries using an checkpoint has exicessive queries. Experiments on various advanced 3D detectors show that GPQ effectively reduces redundant queries while maintaining performance. Using our method, model inference on desktop GPUs can be accelerated by up to 1.31x. Moreover, after deployment on edge devices, it achieves up to a 67.86\\% reduction in FLOPs and a 76.38\\% decrease in inference time. The code will be available soon.", "keywords": [ "DETR", "3D object detection", "Query pruning" ], "rating_list": [ 6, 6, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "NZC5QgbTSq", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Donghao Ying", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Henry Lam", "gender": "unknown", "institution": "Columbia University", "country": "", "position": "Associate Professor" }, { "name": "Huajie Qian", "gender": "unknown", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Wotao Yin", "gender": "Male", "institution": "Alibaba Group US", "country": "", "position": "Principal Researcher" } ], "n_formula": 137, "n_formula_1": 18, "n_ref_uni": 67, "n_ref": 85, "n_ref_all": 129, "n_fig": 26, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 585, "n_element_tab_1": 6, "formula_len_all": 16193, "formula_len_all_1": 1712, "len_all": 234966, "len_all_1": 57292, "len_abs": 1987, "len_title": 114, "len_sents": 58724, "len_sents_1": 26328, "n_sents": 600, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 887, "title": "Subsampled Ensemble Can Improve Generalization Tail Exponentially", "abs": "Ensemble learning is a popular technique to improve the accuracy of machine learning models. It hinges on the rationale that aggregating multiple weak models can lead to better models with lower variance and hence higher stability, especially for discontinuous base learners. In this paper, we provide a new perspective on ensembling. By selecting the best model trained on subsamples via majority voting, we can attain exponentially decaying tails for the excess risk, even if the base learner suffers from slow (i.e., polynomial) decay rates. This tail enhancement power of ensembling is agnostic to the underlying base learner and is stronger than variance reduction in the sense of exhibiting rate improvement. We demonstrate how our ensemble methods can substantially improve out-of-sample performances in a range of examples involving heavy-tailed data or intrinsically slow rates.", "keywords": [ "ensemble method", "subsampling", "heavy tail", "exponential convergence", "excess risk", "generalization" ], "rating_list": [ 8, 6, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "NYf2XIXUi3", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Osmar Zaiane", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "Full Professor" }, { "name": "Shadan Golestan", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Postdoc" }, { "name": "Sheila Schoepp", "gender": "Female", "institution": "University of Alberta", "country": "CA", "position": "PhD student" }, { "name": "Shotaro Miwa", "gender": "Male", "institution": "AIST, National Institute of Advanced Industrial Science and Technology", "country": "CA", "position": "Researcher" }, { "name": "Yoshihiro Mitsuka", "gender": "Male", "institution": "Mitsubishi Electric", "country": "", "position": "Researcher" }, { "name": "Zahin Sufiyan", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Researcher" } ], "n_formula": 35, "n_formula_1": 25, "n_ref_uni": 39, "n_ref": 61, "n_ref_all": 116, "n_fig": 15, "n_tab": 15, "L_tab": 6499, "n_element_tab": 814, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 596, "n_element_tab_1": 36, "formula_len_all": 3263, "formula_len_all_1": 1793, "len_all": 173561, "len_all_1": 66689, "len_abs": 1000, "len_title": 118, "len_sents": 50196, "len_sents_1": 29651, "n_sents": 370, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1003, "title": "TLXML: Task-Level Explanation of Meta-Learning via Influence Functions", "abs": "The scheme of adaptation via meta-learning is seen as an ingredient for solving the problem of data shortage or distribution shift in real-world applications, but it also brings the new risk of inappropriate updates of the model in the user environment, which increases the demand for explainability. Among the various types of XAI methods, establishing a method of explanation based on past experience in meta-learning requires special consideration due to its bi-level structure of training, which has been left unexplored. In this work, we propose influence functions for explaining meta-learning that measure the sensitivities of training tasks to adaptation and inference. We also argue that the approximation of the Hessian using the Gauss-Newton matrix resolves computational barriers peculiar to meta-learning. We demonstrate the adequacy of the method through experiments on task distinction and task distribution distinction using image classification tasks with MAML and Prototypical Network.", "keywords": [ "meta-learning", "XAI", "explainability", "influence function", "hessian approximation" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 4 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "NYPJz0CL5X", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 7, "author_info_list": [ { "name": "Hamza Errahmouni Barkam", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Mohsen Imani", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Assistant Professor" }, { "name": "Prathyush Poduval", "gender": "unknown", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Sanggeon Yun", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Xiangjian Liu", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Undergrad student" }, { "name": "Yang Ni", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Zhuowen Zou", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 25, "n_ref": 38, "n_ref_all": 61, "n_fig": 11, "n_tab": 1, "L_tab": 270, "n_element_tab": 41, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 1148, "n_element_tab_1": 45, "formula_len_all": 1034, "formula_len_all_1": 724, "len_all": 109602, "len_all_1": 60020, "len_abs": 1065, "len_title": 126, "len_sents": 39196, "len_sents_1": 27200, "n_sents": 295, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1075, "title": "Optimal Hyperdimensional Representation for Learning and Cognitive Computation", "abs": "Hyperdimensional Computing (HDC), as a novel neurally-inspired computing methodology, uses lightweight and high-dimensional operations to realize major brain functionalities. Recent HDC works mainly focus on two aspects: brain-like learning and cognitive computation. However, it lacks differentiation between these functions and their requirements for HDC algorithms. We address this gap by proposing an adaptable hyperdimensional kernel-based encoding method. We explore how encoding settings impact HDC performance for both tasks, highlighting the distinction between learning patterns and retrieving information. We provide detailed guidance on kernel design, optimizing data points for accurate decoding or correlated learning. Experimental results with our proposed encoder significantly boost image classification accuracy from 65\\% to 95\\% by considering pixel correlations and increase decoding accuracy from 85\\% to 100\\% by maximizing pixel vector separation. Factorization tasks are shown to require highly exclusive representation to enable accurate convergence.", "keywords": [ "hyperdimensional computing; vector symbolic architecture; decoding;" ], "rating_list": [ 5, 3, 1 ], "soundness_list": [ 2, 2, 1 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "NY7aEek0mi", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Yonatan Sverdlov", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "PhD student" }, { "name": "Nadav E Dym", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "Assistant Professor" } ], "n_formula": 59, "n_formula_1": 15, "n_ref_uni": 56, "n_ref": 155, "n_ref_all": 192, "n_fig": 2, "n_tab": 2, "L_tab": 997, "n_element_tab": 55, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4742, "formula_len_all_1": 1195, "len_all": 222247, "len_all_1": 68390, "len_abs": 1106, "len_title": 97, "len_sents": 69146, "len_sents_1": 32126, "n_sents": 619, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1109, "title": "On the Expressive Power of Sparse Geometric MPNNs", "abs": "Motivated by applications in chemistry and other sciences, we study the expressive\npower of message-passing neural networks for geometric graphs, whose node\nfeatures correspond to 3-dimensional positions. Recent work has shown that such\nmodels can separate generic pairs of non-isomorphic geometric graphs, though they\nmay fail to separate some rare and complicated instances. However, these results\nassume a fully connected graph, where each node possesses complete knowledge\nof all other nodes. In contrast, often, in application, every node only possesses\nknowledge of a small number of nearest neighbors.\nThis paper shows that generic pairs of non-isomorphic geometric graphs can\nbe separated by message-passing networks with rotation equivariant features as\nlong as the underlying graph is connected. When only invariant intermediate\nfeatures are allowed, generic separation is guaranteed for generically globally\nrigid graphs. We introduce a simple architecture, EGENNET, which achieves our\ntheoretical guarantees and compares favorably with alternative architecture on\nsynthetic and chemical benchmarks", "keywords": [ "WL", "GWL", "GNNS", "geometric GNNS", "point clouds", "completness" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "NXnNiT0fdp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Abdessalam Ed-dib", "gender": "unknown", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Zhanibek Datbayev", "gender": "Male", "institution": "Nace.AI", "country": "US", "position": "Researcher" }, { "name": "Amine Mohamed Aboussalah", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 23, "n_formula_1": 11, "n_ref_uni": 49, "n_ref": 77, "n_ref_all": 96, "n_fig": 5, "n_tab": 7, "L_tab": 3371, "n_element_tab": 309, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2527, "n_element_tab_1": 241, "formula_len_all": 2063, "formula_len_all_1": 627, "len_all": 173536, "len_all_1": 69623, "len_abs": 1253, "len_title": 112, "len_sents": 48908, "len_sents_1": 31560, "n_sents": 400, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1247, "title": "GLoRA: Geometric Adaptive Ranks for Efficient LoRA Fine-Tuning", "abs": "Fine-tuning large language models is computationally intensive because it requires updating all parameters. Low-Rank Adaptation (LoRA) improves efficiency by modifying only a subset of weights but introduces a trade-off between expressivity and computational cost: lower ranks reduce resources but limit expressiveness, while higher ranks enhance expressivity at increased cost. Despite recent advances in adaptive LoRA techniques, existing methods fail to provide a theoretical basis for optimizing the trade-off between model performance and efficiency. We propose Geometric Low-Rank Adaptation (GLoRA), a novel framework that computes the intrinsic dimensionality of hidden state representations to adaptively select LoRA ranks. We demonstrate that the intrinsic dimension provides a lower bound for the optimal rank of LoRA matrices, allowing for a principled selection that balances efficiency and expressivity. GLoRA dynamically adjusts the rank for each layer based on the intrinsic dimensionality of its input and output representations, recognizing that not all model parameters equally impact fine-tuning. Empirical validation on multiple tasks shows that GLoRA consistently outperforms recent baselines within the same parameter budget.", "keywords": [ "Low-Rank Adaptation", "Large Language Models", "Intrinsic Dimensionality" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 1, 4, 2, 3 ], "presentation_list": [ 1, 2, 1, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "NWvsm2VxAM", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Darian Toma\u0161evi\u0107", "gender": "unknown", "institution": "University of Ljubljana", "country": "SI", "position": "PhD student" }, { "name": "Fadi Boutros", "gender": "Male", "institution": "Fraunhofer Institute for Computer Graphics Research", "country": "DE", "position": "Researcher" }, { "name": "Naser Damer", "gender": "Male", "institution": "Fraunhofer Institute for Computer Graphics Research IGD", "country": "DE", "position": "Senior researcher" }, { "name": "Peter Peer", "gender": "Male", "institution": "University of Ljubljana", "country": "SI", "position": "Full Professor" }, { "name": "Vitomir Struc", "gender": "Male", "institution": "University of Ljubljana", "country": "SI", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 39, "n_ref": 122, "n_ref_all": 141, "n_fig": 4, "n_tab": 2, "L_tab": 3095, "n_element_tab": 299, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 5049, "n_element_tab_1": 191, "formula_len_all": 452, "formula_len_all_1": 452, "len_all": 122147, "len_all_1": 64936, "len_abs": 1663, "len_title": 117, "len_sents": 36018, "len_sents_1": 28928, "n_sents": 227, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1672, "title": "ID-Booth: Identity-consistent image generation with diffusion models", "abs": "The recent retraction of large-scale biometric datasets, prompted by strict privacy regulations, presents a critical challenge for future biometric research. This is evident with the face recognition task, for which large-scale datasets were often gathered through web-scraping without the consent of subjects. A potential solution entails the creation of synthetic data, suitable for training recognition models, with deep generative models. Existing generative approaches rely on conditioning and fine-tuning of powerful pretrained diffusion models to achieve the synthesis of realistic images of a desired identity. Yet, these methods often do not consider the identity of subjects during training, leading to poor consistency between generated and intended identities. In contrast, methods that employ identity-based training objectives tend to overfit on various aspects of the identity, and in turn, lower the diversity of images that can be generated. To address these issues, we present the ID-Booth fine-tuning framework, which utilizes a novel triplet identity training objective and enables identity-consistent image generation while retaining the synthesis capabilities of pretrained models. Experiments across two latent diffusion models with varying prompt complexity reveal that our method facilitates better intra-identity consistency and inter-identity separability while achieving higher image diversity. In turn, the produced data enables the training of better-performing recognition models than even real-world datasets of a similar scale gathered with suitable consent. The source code for the ID-Booth framework is available at omitted_for_review.", "keywords": [ "Image synthesis", "Diffusion models", "Face recognition data" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 1, 1, 1, 1 ], "confidence_list": [ 5, 5, 4, 2 ] }, { "paper_id": "NWb128pSCb", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Bei Yang", "gender": "unknown", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Chengyu Wang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jun Huang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xiangru Zhu", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Xiaoxiao Xu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Yanghua Xiao", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yaoxian Song", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Postdoc" }, { "name": "Zhixu Li", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "penglei sun", "gender": "Male", "institution": " The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 2, "n_ref_uni": 31, "n_ref": 48, "n_ref_all": 86, "n_fig": 16, "n_tab": 8, "L_tab": 4040, "n_element_tab": 658, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2136, "formula_len_all_1": 239, "len_all": 269559, "len_all_1": 54488, "len_abs": 1336, "len_title": 126, "len_sents": 78662, "len_sents_1": 25193, "n_sents": 822, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1285, "title": "Evaluating Semantic Variation in Text-to-Image Synthesis: A Causal Perspective", "abs": "Accurate interpretation and visualization of human instructions are crucial for \\textbf{T}ext-to-\\textbf{I}mage (\\textbf{T2I}) synthesis. Existing T2I models often struggle to capture semantic variations caused by different word orders. Current evaluations rely on indirect metrics, like text-image matching, where the limitations with uncommon or complex linguistic patterns may be obscured by focusing on frequent word combinations. To address these deficiencies, we introduce a novel metric called \\underline{PermuteEffect} and a benchmark named \\underline{PermuteBench}, designed to evaluate the causality between textual semantic variations and visual semantic variations in T2I synthesis. Our experiments revealed that the CogView-3-Plus and Ideogram 2 models performed the best, achieving a score of 0.2/1, where semantic variations concerning relations between objects were more weakly understood than those about attributes, scoring only 0.07/1. We found that cross-modal alignment in UNet or Transformers plays a critical role in understanding semantic variations, a factor previously overlooked by a focus on textual encoders. Our work establishes an effective evaluation framework that advances \nthe T2I synthesis community's exploration of human instruction understanding.", "keywords": [ "text-to-image synthesis", "semantics", "evaluation", "metric" ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "NWH2pdKu2I", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Chen Chen", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Li Ren", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Liqiang Wang", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Full Professor" }, { "name": "Kien Hua", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 12, "n_ref_uni": 76, "n_ref": 140, "n_ref_all": 158, "n_fig": 6, "n_tab": 9, "L_tab": 3671, "n_element_tab": 479, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1797, "n_element_tab_1": 280, "formula_len_all": 843, "formula_len_all_1": 735, "len_all": 204608, "len_all_1": 67124, "len_abs": 2011, "len_title": 130, "len_sents": 49464, "len_sents_1": 29019, "n_sents": 310, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1504, "title": "Prompt Distribution Matters: Tuning Visual Prompt Through Semantic Metric Guidance", "abs": "Visual Prompt Tuning (VPT) has become a promising solution for Parameter-Efficient Fine-Tuning (PEFT) of pre-trained Vision Transformer (ViT) models on downstream vision tasks. VPT partially fine-tunes a set of learnable tokens while keeping the majority of the model parameters frozen. Recent research has explored modifying the connection structures of the prompts. However, the fundamental correlation and distribution between the prompts and image tokens remain unexplored. In this paper, we leverage \\textit{metric learning} techniques to investigate how the distribution of prompts affects fine-tuning and transfer learning performance. Specifically, we propose a novel framework, \\textbf{D}istribution \\textbf{A}ware \\textbf{V}isual \\textbf{P}rompt Tuning (DA-VPT), to guide the distributions of the prompts by learning the distance metric from their class-related semantic data. Our method demonstrates that the prompts can serve as an effective bridge to share semantic information between image patches and the class token. We extensively evaluated our approach on popular benchmarks in both recognition and segmentation tasks. The results show the effectiveness of our proposed method and offer a new direction for PEFT optimization in vision transformers. We demonstrate that our approach enables more effective and efficient fine-tuning of ViT models by leveraging semantic information to guide the learning of the prompts, leading to improved performance on various downstream vision tasks.", "keywords": [ "visual prompt tuning", "parameter-efficient fine-tuning", "metric learning" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "NW5vSJXO9V", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Alex Wong", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" }, { "name": "Byung-Woo Hong", "gender": "Male", "institution": "Chung-Ang University", "country": "KR", "position": "Full Professor" }, { "name": "Tomislav Dobri\u010dki", "gender": "Male", "institution": "Chung-Ang University", "country": "KR", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 22, "n_ref_uni": 33, "n_ref": 80, "n_ref_all": 97, "n_fig": 5, "n_tab": 5, "L_tab": 1240, "n_element_tab": 191, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1532, "n_element_tab_1": 112, "formula_len_all": 2297, "formula_len_all_1": 2296, "len_all": 104836, "len_all_1": 55338, "len_abs": 1199, "len_title": 106, "len_sents": 27022, "len_sents_1": 24103, "n_sents": 180, "n_sents_1": 152, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1203, "title": "Diffusion Process with Implicit Latents via Energy Models", "abs": "We present a generative model based on an ordered sequence of latent variables for intermediate distributions between a given source and a desired target distribution. We construct the probabilistic transitions among the latent variables using energy models that are in the form of classifiers. In our work, the intermediate transitional distributions are implicitly defined by the energy models during training, where the statistical properties of the data distribution are naturally taken into account. This is in contrast to denoising diffusion probabilistic models (DDPMs) where they are explicitly defined by the predefined scheduling of a sequential noise degradation process. Over the course of training, our model is designed to optimally determine the intermediate distributions by Langevin dynamics driven by the energy model. In contrast, energy-based models (EBMs) typically require an additional generator since the intermediate distributions are not explicitly defined in the training procedure. We demonstrate the effectiveness and efficiency of the proposed algorithm in the context of image generation, achieving high fidelity results with less inference steps on a variety of datasets.", "keywords": [ "Energy-based model", "Image generation", "Gradient Penalty", "Diffusion Process" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "NVKwjCIAAX", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Akihiro FUJII", "gender": "Male", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "JP", "position": "PhD student" }, { "name": "Anh Khoa Augustin Lu", "gender": "Male", "institution": "The University of Tokyo, The University of Tokyo", "country": "JP", "position": "Assistant Professor" }, { "name": "Koji Shimizu", "gender": "Male", "institution": "AIST, National Institute of Advanced Industrial Science and Technology", "country": "JP", "position": "Researcher" }, { "name": "Satoshi Watanabe", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Yoshitaka Ushiku", "gender": "Male", "institution": "OMRON SINIC X", "country": "JP", "position": "Vice President for Research" } ], "n_formula": 17, "n_formula_1": 14, "n_ref_uni": 55, "n_ref": 114, "n_ref_all": 149, "n_fig": 3, "n_tab": 9, "L_tab": 8932, "n_element_tab": 254, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1398, "n_element_tab_1": 116, "formula_len_all": 1083, "formula_len_all_1": 995, "len_all": 219436, "len_all_1": 69417, "len_abs": 1392, "len_title": 163, "len_sents": 63464, "len_sents_1": 32337, "n_sents": 472, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 114, "L_abs": 1396, "title": "Adaptive Constraint Integration for Simultaneously Optimizing Crystal Structures with Multiple Targeted Properties", "abs": "In materials science, finding crystal structures that have targeted properties is crucial. While recent methodologies such as Bayesian optimization and deep generative models have made some advances on this issue, these methods often face difficulties in adaptively incorporating various constraints, such as electrical neutrality and targeted properties optimization, while keeping the desired specific crystal structure. To address these challenges, we have developed the Simultaneous Multi-property Optimization using Adaptive Crystal Synthesizer (SMOACS), which utilizes state-of-the-art property prediction models and their gradients to directly optimize input crystal structures for targeted properties simultaneously. SMOACS enables the integration of adaptive constraints into the optimization process without necessitating model retraining. Thanks to this feature, SMOACS has succeeded in simultaneously optimizing targeted properties while maintaining perovskite structures, even with models trained on diverse crystal types. We have demonstrated the band gap optimization while meeting a challenging constraint, that is, maintaining electrical neutrality in large atomic configurations up to 135 atom sites, where the verification of the electrical neutrality is challenging. The properties of the most promising materials have been confirmed by density functional theory calculations.", "keywords": [ "deep learning", "materials discovery", "materials design" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "NVASzf27bL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Heng Lu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "ShiLiang Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Staff Algorithm Engineer" }, { "name": "Zhen-Hua Ling", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Zhengyan Sheng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Zhihao Du", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 30, "n_ref": 93, "n_ref_all": 128, "n_fig": 7, "n_tab": 5, "L_tab": 2102, "n_element_tab": 203, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1707, "n_element_tab_1": 187, "formula_len_all": 1154, "formula_len_all_1": 699, "len_all": 181816, "len_all_1": 61158, "len_abs": 1529, "len_title": 132, "len_sents": 47010, "len_sents_1": 28076, "n_sents": 330, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1477, "title": "Unispeaker: A unified speech generation model for multimodality-driven voice control", "abs": "Recent advancements in zero-shot speech generation have brought synthetic speech increasingly close to the realism of target speakers' recordings, yet multimodal voice creation remains an evolving field. In various scenarios, individuals often seek to control and create voice characteristics through different voice description modalities. To address the limitations in both the versatility and performance of voice control found in previous methods, this paper introduces UniSpeaker, a unified, multimodal-driven speech generation model that integrates face images, text descriptions, attribute descriptions, and reference speech for comprehensive voice control and creation. Specifically, we propose a unified voice aggregator based on KV-Former, applying soft contrastive loss to map diverse voice description modalities into a shared voice space, ensuring that the generated voice aligns more closely with the input descriptions. In addition, multimodal voice control is incorporated within a large-scale speech generation framework, employing self-distillation to enhance voice disentanglement. We introduce the MVC benchmark to evaluate multimodality-driven voice control, focusing on voice suitability, voice diversity, and speech quality. We assess UniSpeaker across five tasks using the MVC benchmark, and the experimental results demonstrate that UniSpeaker outperforms previous modality-specific models. Speech samples are available at https://UniSpeaker.github.io.", "keywords": [ "Speech Generation", "Voice Control", "Multi-Modality Alignment" ], "rating_list": [ 5, 3, 3, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 2, 1, 4 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "NV5p50EkT6", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Hui Xiong", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Siyang Li", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Yize Chen", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 29, "n_formula_1": 12, "n_ref_uni": 57, "n_ref": 120, "n_ref_all": 150, "n_fig": 7, "n_tab": 16, "L_tab": 9960, "n_element_tab": 922, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 448, "n_element_tab_1": 63, "formula_len_all": 2943, "formula_len_all_1": 933, "len_all": 218576, "len_all_1": 63713, "len_abs": 1671, "len_title": 152, "len_sents": 61435, "len_sents_1": 31226, "n_sents": 397, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1607, "title": "Channel-aware Contrastive Conditional Diffusion for Multivariate Probabilistic Time Series Forecasting", "abs": "Forecasting faithful trajectories of multivariate time series from practical scopes is essential for reasonable decision-making. Recent methods majorly tailor generative conditional diffusion models to estimate the target temporal predictive distribution. However, it remains an obstacle to enhance the exploitation efficiency of given implicit temporal predictive information to bolster conditional diffusion learning. To this end, we propose a generic channel-aware contrastive conditional diffusion model termed CCDM to achieve desirable multivariate probabilistic forecasting, obviating the need for curated temporal conditioning inductive biases. In detail, we first design a channel-centric conditional denoising network to manage intra-variate variations and cross-variate correlations, which can lead to scalability on diverse prediction horizons and channel numbers. Then, we devise an ad-hoc denoising-based temporal contrastive learning to explicitly amplify the predictive mutual information between past observations and future forecasts. It can coherently complement naive step-wise denoising diffusion training and improve the forecasting accuracy and generality on unknown test time series. Besides, we offer theoretic insights on the benefits of such auxiliary contrastive training refinement from both neural mutual information and temporal distribution generalization aspects. The proposed CCDM can exhibit superior forecasting capability compared to current state-of-the-art diffusion forecasters over a comprehensive benchmark, with best MSE and CRPS outcomes on 79.17% and 87.5% cases.", "keywords": [ "diffusion models", "contrastive learning", "time series forecasting" ], "rating_list": [ 3, 3, 8, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "NUD03NBDOE", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Chitta Baral", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Full Professor" }, { "name": "Divij Handa", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Pavel Dolin", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Shrinidhi Kumbhar", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Tran Cao Son", "gender": "Male", "institution": "New Mexico State University", "country": "US", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 0, "n_ref_uni": 34, "n_ref": 45, "n_ref_all": 82, "n_fig": 3, "n_tab": 14, "L_tab": 9668, "n_element_tab": 698, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 13, "n_element_tab_1": 1, "formula_len_all": 838, "formula_len_all_1": 0, "len_all": 205289, "len_all_1": 56956, "len_abs": 1648, "len_title": 137, "len_sents": 63063, "len_sents_1": 28904, "n_sents": 492, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1733, "title": "ActionReasoningBench: Reasoning about Actions with and without Ramification Constraints", "abs": "Reasoning about Actions and Change (RAC) has historically played a pivotal role in solving foundational AI problems, such as the frame problem. It has driven advancements in AI fields, such as non-monotonic and commonsense reasoning. RAC remains crucial for AI systems that operate in dynamic environments, engage in interactive scenarios, or rely on commonsense reasoning. Despite substantial advances made by Large Language Models (LLMs) in various AI domains, their performance in RAC remains underexplored. To address this gap, we introduce a new diagnostic benchmark, $\\textbf{ActionReasoningBench}$, which encompasses 8 domains and includes questions for up to 19 action sequences. This benchmark rigorously evaluates LLMs across six key RAC dimensions: $\\textit{Fluent Tracking}$, $\\textit{State Tracking}$, $\\textit{Action Executability}$, $\\textit{Effects of Actions}$, $\\textit{Numerical RAC}$, and $\\textit{Composite Questions}$. LLMs demonstrate average accuracy rates of 73.55%, 65.63%, 58.73%, and 62.38% on the former four dimensions, which are frequently discussed in RAC literature. However, the performance on the latter two dimensions, which introduce complex and novel reasoning questions, the average performance of LLMs is lowered to 33.16% and 51.19%, respectively, reflecting a 17.9% performance decline. We also introduce new ramification constraints to capture the indirect effects of actions, providing deeper insights into RAC challenges. Our evaluation of state-of-the-art LLMs, including both open-source and commercial models, reveals challenges across all RAC dimensions, particularly in handling ramifications, with GPT-4o failing to solve any question and o1-preview achieving a score of only 18.4%.", "keywords": [ "Reasoning about Actions and Change (RAC)", "Benchmark", "Large Language Models (LLMs)", "o1-preview" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "NTNdRElwbp", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Luca Viano", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Volkan Cevher", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" }, { "name": "Yihang Chen", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Yongtao Wu", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Zhenyu Zhu", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" } ], "n_formula": 122, "n_formula_1": 40, "n_ref_uni": 47, "n_ref": 130, "n_ref_all": 146, "n_fig": 2, "n_tab": 6, "L_tab": 2875, "n_element_tab": 143, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 2913, "n_element_tab_1": 33, "formula_len_all": 16619, "formula_len_all_1": 4867, "len_all": 210675, "len_all_1": 70159, "len_abs": 1314, "len_title": 110, "len_sents": 51178, "len_sents_1": 24518, "n_sents": 522, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1275, "title": "Multi-Step Preference Optimization via Two-Player Markov Games", "abs": "Reinforcement Learning from Human Feedback (RLHF) has been highly successful in aligning large language models with human preferences. While prevalent methods like DPO have demonstrated strong performance, they frame interactions with the language model as a bandit problem, which limits their applicability in real-world scenarios where multi-turn conversations are common. Additionally, DPO relies on the Bradley-Terry model assumption, which does not adequately capture the non-transitive nature of human preferences. In this paper, we address these challenges by modeling the alignment problem as a two-player constant-sum Markov game, where each player seeks to maximize their winning rate against the other across all steps of the conversation. Our approach Multi-step Preference Optimization (MPO) is built upon the natural actor-critic framework. We further develop OMPO based on the optimistic online gradient descent algorithm. Theoretically, we provide a rigorous analysis for both algorithms on convergence and show that OMPO requires $\\mathcal{O}(\\epsilon^{-1})$ policy updates to converge to an $\\epsilon$-approximate Nash equilibrium. We also validate the effectiveness of our method through experiments on the multi-turn conversations dataset in MT-bench-101.", "keywords": [ "Multi-step Preference Optimization", "Two-player Markov game", "RLHF", "Optimistic Online Gradient Descent" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "NTHMw8S1Ow", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Mihaela van der Schaar", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Katarzyna Kobalczyk", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 57, "n_formula_1": 14, "n_ref_uni": 52, "n_ref": 75, "n_ref_all": 103, "n_fig": 6, "n_tab": 4, "L_tab": 890, "n_element_tab": 53, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5790, "formula_len_all_1": 1125, "len_all": 314677, "len_all_1": 82628, "len_abs": 1170, "len_title": 128, "len_sents": 88594, "len_sents_1": 35123, "n_sents": 708, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 80, "L_abs": 1162, "title": "Towards Automated Knowledge Integration From Human-Interpretable Representations", "abs": "In noisy and low-data environments, a significant challenge in machine learning lies in effectively incorporating inductive biases that enhance data efficiency and robustness. Despite many success of informed machine learning methods, designing algorithms with explicit inductive biases based on prior expert knowledge remains largely a manual process. In this work, we explore how prior knowledge represented in its native formats, e.g. in natural language, can be integrated into machine learning models in an automated manner. Inspired by the learning to learn principles of meta-learning, we consider an approach of learning to integrate knowledge via conditional meta-learning, a paradigm we refer to as informed meta-learning. We introduce and motivate theoretically the principles of informed meta-learning enabling automated and controllable inductive bias selection. To illustrate our claims, we implement an instantiation of informed meta-learning--the Informed Neural Process, and empirically demonstrate the potential benefits and limitations of informed meta-learning in improving data efficiency and generalizing to novel knowledge representations.", "keywords": [ "informed machine learning", "knowledge integration", "data efficiency", "prior knowledge" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "NSpe8QgsCB", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Boaz Lerner", "gender": "Male", "institution": "OriginAI", "country": "", "position": "Researcher" }, { "name": "Dvir Samuel", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "PhD student" }, { "name": "Gavriel Habib", "gender": "unknown", "institution": "OriginAI", "country": "IL", "position": "Researcher" }, { "name": "Issar Tzachor", "gender": "unknown", "institution": "OriginAI", "country": "IL", "position": "Researcher" }, { "name": "Matan Levy", "gender": "unknown", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "PhD student" }, { "name": "Michael Green", "gender": "Male", "institution": "OriginAI", "country": "IL", "position": "Computer Vision Researcher" }, { "name": "Nir Darshan", "gender": "Male", "institution": "OriginAI", "country": "IL", "position": "Principal Researcher" }, { "name": "Noam Korngut Zailer", "gender": "Female", "institution": "Deep Learning Indaba", "country": "", "position": "Researcher" }, { "name": "Or Shimshi", "gender": "Male", "institution": "OriginAI", "country": "", "position": "Researcher" }, { "name": "Rami Ben-Ari", "gender": "Male", "institution": "OriginAI", "country": "IL", "position": "Principal Researcher" }, { "name": "Tal Berkovitz Shalev", "gender": "Female", "institution": "OriginAI", "country": "IL", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 52, "n_ref": 203, "n_ref_all": 241, "n_fig": 11, "n_tab": 8, "L_tab": 2251, "n_element_tab": 340, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 565, "n_element_tab_1": 99, "formula_len_all": 231, "formula_len_all_1": 231, "len_all": 174471, "len_all_1": 64720, "len_abs": 1260, "len_title": 77, "len_sents": 51182, "len_sents_1": 31744, "n_sents": 365, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1266, "title": "EffoVPR: Effective Foundation Model Utilization for Visual Place Recognition", "abs": "The task of Visual Place Recognition (VPR) is to predict the location of a query image from a database of geo-tagged images. Recent studies in VPR have highlighted the significant advantage of employing pre-trained foundation models like DINOv2 for the VPR task. However, these models are often deemed inadequate for VPR without further fine-tuning on VPR-specific data.\nIn this paper, we present an effective approach to harness the potential of a foundation model for VPR. We show that features extracted from self-attention layers can act as a powerful re-ranker for VPR, even in a zero-shot setting. Our method not only outperforms previous zero-shot approaches but also introduces results competitive with several supervised methods.\nWe then show that a single-stage approach utilizing internal ViT layers for pooling can produce global features that achieve state-of-the-art performance, with impressive feature compactness down to 128D. Moreover, integrating our local foundation features for re-ranking further widens this performance gap. Our method also demonstrates exceptional robustness and generalization, setting new state-of-the-art performance, while handling challenging conditions such as occlusion, day-night transitions, and seasonal variations.", "keywords": [ "Visual Place Recognition", "Geo-Localization", "Re-ranking", "Vision Transformer", "Self-Attention" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "NSlvSDQ8aE", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Wenbing Huang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yang Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Ziyang Yu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 38, "n_formula_1": 6, "n_ref_uni": 39, "n_ref": 72, "n_ref_all": 127, "n_fig": 8, "n_tab": 8, "L_tab": 3284, "n_element_tab": 165, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 577, "n_element_tab_1": 33, "formula_len_all": 2721, "formula_len_all_1": 337, "len_all": 178405, "len_all_1": 66866, "len_abs": 1244, "len_title": 124, "len_sents": 53259, "len_sents_1": 29497, "n_sents": 376, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1249, "title": "Force-Guided Bridge Matching for Full-Atom Time-Coarsened Dynamics of Peptides", "abs": "Molecular Dynamics (MD) is crucial in various fields such as materials science, chemistry, and pharmacology to name a few. Conventional MD software struggles with the balance between time cost and prediction accuracy, which restricts its wider application. Recently, data-driven approaches based on deep generative models have been devised for time-coarsened dynamics, which aim at learning dynamics of diverse molecular systems over a long timestep, enjoying both universality and efficiency. Nevertheless, most current methods are designed solely to learn from the data distribution regardless of the underlying Boltzmann distribution, and the physics priors such as energies and forces are constantly overlooked. In this work, we propose a conditional generative model called Force-guided Bridge Matching (FBM), which learns full-atom time-coarsened dynamics and targets the Boltzmann-constrained distribution. With the guidance of our delicately-designed intermediate force field, FBM leverages favourable physics priors into the generation process, giving rise to enhanced simulations. Experiments on two datasets consisting of peptides verify our superiority in terms of comprehensive metrics and demonstrate transferability to unseen systems.", "keywords": [ "molecular dynamics", "force-guided bridge matching", "graph neural network" ], "rating_list": [ 8, 8, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 4, 1 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "NSefAqUM6U", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Michael Slawinski", "gender": "unknown", "institution": "Crowdstrike Holdings Inc", "country": "", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 21, "n_ref": 29, "n_ref_all": 33, "n_fig": 1, "n_tab": 3, "L_tab": 1295, "n_element_tab": 115, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1295, "n_element_tab_1": 115, "formula_len_all": 460, "formula_len_all_1": 423, "len_all": 79670, "len_all_1": 53489, "len_abs": 1125, "len_title": 119, "len_sents": 25077, "len_sents_1": 23387, "n_sents": 193, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1136, "title": "Deep Clustering and Interpolation via the Federated Self-Organizing Map", "abs": "We introduce FedSOM, a clustering and interpolation module based on the Self-organizing Map (SOM), which can be appended to any encoder and which can be trained in a federated way either in tandem with the encoder or post training on the resulting representations. The result is a discrete moduli space of representations that provides for cluster or sample-level interpolation, hierarchical clustering, and can be leveraged as a function to cluster new vectors at test time. This moduli space can either be created from data alone or by glueing pre-existing clusters along regions of commonality, although we do not explore the latter in this work. Interpolation is accomplished by considering the $n$-dimensional tensor underlying the SOM as a weighted undirected graph, where the weights are computed as a function of the dispersion of the two clusters corresponding to the nodes bounding the given edge. Any two clusters or samples may then be interpolated by computing the lowest-cost path between their associated graph nodes via Dijkstra's algorithm. The method is validated on MNIST-like and parsed-binary malware datasets.", "keywords": [ "self", "organizing", "map", "federated", "cybersecurity", "security", "ember", "interpolation", "clustering" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 1, 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 1, 1, 1 ], "contribution_list": [ 1, 3, 1, 1, 2 ], "confidence_list": [ 5, 3, 5, 4, 2 ] }, { "paper_id": "NS1G1Uhny3", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Andreas Krause", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Ido Hakimi", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Jonas H\u00fcbotter", "gender": "Male", "institution": "ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Sascha Bongni", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" } ], "n_formula": 49, "n_formula_1": 13, "n_ref_uni": 107, "n_ref": 238, "n_ref_all": 327, "n_fig": 26, "n_tab": 11, "L_tab": 5435, "n_element_tab": 326, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1174, "n_element_tab_1": 94, "formula_len_all": 4554, "formula_len_all_1": 868, "len_all": 387410, "len_all_1": 81702, "len_abs": 2981, "len_title": 109, "len_sents": 112982, "len_sents_1": 35166, "n_sents": 943, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1266, "title": "Efficiently Learning at Test-Time: Active Fine-Tuning of LLMs", "abs": "Recent efforts in fine-tuning language models often rely on automatic data selection, commonly using Nearest Neighbors retrieval from large datasets.\nHowever, we theoretically show that this approach tends to select redundant data, limiting its effectiveness or even hurting performance.\nTo address this, we introduce SIFT, a data selection algorithm designed to reduce uncertainty about the model's response given a prompt, which unifies ideas from retrieval and active learning.\nWhereas Nearest Neighbor retrieval typically fails in the presence of information duplication, SIFT accounts for information duplication and optimizes the overall information gain of the selected examples.\nWe focus our evaluations on fine-tuning at test-time for prompt-specific language modeling on the Pile dataset, and show that SIFT consistently outperforms Nearest Neighbor retrieval, with minimal computational overhead.\nMoreover, we show that our uncertainty estimates can predict the performance gain of test-time fine-tuning, and use this to develop an adaptive algorithm that invests test-time compute proportional to realized performance gains.\nWe provide the `activeft` (Active Fine-Tuning) library which can be used as a drop-in replacement for Nearest Neighbor retrieval.", "keywords": [ "data selection", "retrieval", "active learning", "transductive active learning", "local learning", "test-time fine-tuning", "transductive learning", "language modeling", "uncertainty quantification" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "NRYgUzSPZz", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Jiacheng Ye", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Lin Zheng", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Lingpeng Kong", "gender": "Male", "institution": "Department of Computer Science, The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Shansan Gong", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xin Jiang", "gender": "Male", "institution": "Noah\u2019s Ark Lab, Huawei Technologies", "country": "HK", "position": "Principal Researcher" }, { "name": "Zhenguo Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "JIAHUI GAO", "gender": "Female", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 9, "n_ref_uni": 87, "n_ref": 170, "n_ref_all": 203, "n_fig": 6, "n_tab": 4, "L_tab": 782, "n_element_tab": 166, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 406, "n_element_tab_1": 90, "formula_len_all": 2667, "formula_len_all_1": 719, "len_all": 211797, "len_all_1": 67483, "len_abs": 946, "len_title": 124, "len_sents": 44815, "len_sents_1": 31532, "n_sents": 304, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 954, "title": "Beyond Autoregression: Discrete Diffusion for Complex Reasoning and Planning", "abs": "Autoregressive language models, despite their impressive capabilities, struggle with complex reasoning and long-term planning tasks. We introduce discrete diffusion models as a novel solution to these challenges. Through the lens of subgoal imbalance, we demonstrate how diffusion models effectively learn difficult subgoals that elude autoregressive approaches. We propose Multi-granularity Diffusion Modeling (MDM), which prioritizes subgoals based on difficulty during learning. On complex tasks like Countdown, Sudoku, and Boolean Satisfiability Problems, MDM significantly outperforms autoregressive models without using search techniques. For instance, MDM achieves 91.5\\% and 100\\% accuracy on Countdown and Sudoku, respectively, compared to 45.8\\% and 20.7\\% for autoregressive models. Our work highlights the potential of diffusion-based approaches in advancing AI capabilities for sophisticated language understanding and problem-solving tasks.", "keywords": [ "diffusion model", "autoregressive model", "complex reasoning" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "NRY0QAvGNT", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Bin Fan", "gender": "Male", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "Full Professor" }, { "name": "Chenghao Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Gaofeng Meng", "gender": "Male", "institution": "Institute of automation", "country": "CN", "position": "Full Professor" }, { "name": "Jieping Ye", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Lubin Fan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Shiming Xiang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Shixiong Xu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Zhou Yuan", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 36, "n_ref": 73, "n_ref_all": 89, "n_fig": 9, "n_tab": 8, "L_tab": 3923, "n_element_tab": 572, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1530, "n_element_tab_1": 274, "formula_len_all": 234, "formula_len_all_1": 203, "len_all": 152473, "len_all_1": 67976, "len_abs": 1478, "len_title": 153, "len_sents": 45214, "len_sents_1": 30925, "n_sents": 363, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1484, "title": "AddressVLM: Cross-view Alignment Tuning for Image Address Localization using Large Vision-Language Models", "abs": "Large visual language models (LVLMs) have demonstrated impressive performance in coarse-grained geo-localization at the country or city level, but they struggle with fine-grained street-level localization within urban areas. In this paper, we explore integrating city-wide address localization capabilities into LVLMs, facilitating flexible address-related question answering using street-view images. A key challenge is that the street-view visual question-and-answer (VQA) data provides only microscopic visual cues, leading to subpar performance in fine-tuned models. To tackle this issue, we incorporate perspective-invariant satellite images as macro cues and propose cross-view alignment tuning including a satellite-view and street-view image grafting mechanism, along with an automatic alignment label generation mechanism. This helps build connections between street-view images through cross-view matching, thus enhancing LVLM's global understanding of street distribution. We name our proposed model AddressVLM consisting of two-stage training protocols: cross-view alignment tuning and address localization tuning. Furthermore, we have constructed two street-view VQA datasets based on image address localization datasets from Pittsburgh and San Francisco. Qualitative and quantitative evaluations demonstrate that AddressVLM outperforms counterpart LVLMs by over 9% and 12% in average address localization accuracy on the Pitts-VQA and SF-Base-VQA datasets, respectively.", "keywords": [ "Image Address Localization; Large Vision Language Model; Cross-view Alignment; Supervised Fine-tuning" ], "rating_list": [ 5, 5, 8, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "NQqJPPCesd", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Hai Nguyen", "gender": "Male", "institution": "VinAI Research", "country": "", "position": "Intern" }, { "name": "Khai Nguyen", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Nhat Ho", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Assistant Professor" } ], "n_formula": 63, "n_formula_1": 33, "n_ref_uni": 47, "n_ref": 61, "n_ref_all": 89, "n_fig": 13, "n_tab": 6, "L_tab": 3208, "n_element_tab": 316, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 421, "n_element_tab_1": 36, "formula_len_all": 9251, "formula_len_all_1": 3079, "len_all": 164727, "len_all_1": 65082, "len_abs": 1543, "len_title": 103, "len_sents": 41229, "len_sents_1": 26907, "n_sents": 413, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 55, "L_abs": 1502, "title": "Towards Marginal Fairness Sliced Wasserstein Barycenter", "abs": "The Sliced Wasserstein barycenter (SWB) is a widely acknowledged method for efficiently generalizing the averaging operation within probability measure spaces. However, achieving marginal fairness SWB, ensuring approximately equal distances from the barycenter to marginals, remains unexplored. The uniform weighted SWB is not necessarily the optimal choice to obtain the desired marginal fairness barycenter due to the heterogeneous structure of marginals and the non-optimality of the optimization. As the first attempt to tackle the problem, we define the marginal fairness sliced Wasserstein barycenter (MFSWB) as a constrained SWB problem. Due to the computational disadvantages of the formal definition, we propose two hyperparameter-free and computationally tractable surrogate MFSWB problems that implicitly minimize the distances to marginals and encourage marginal fairness at the same time. To further improve the efficiency, we perform slicing distribution selection and obtain the third surrogate definition by introducing a new slicing distribution that focuses more on marginally unfair projecting directions. We discuss the relationship of the three proposed problems and their relationship to sliced multi-marginal Wasserstein distance. Finally, we conduct experiments on finding 3D point-clouds averaging, color harmonization, and training of sliced Wasserstein autoencoder with class-fairness representation to show the favorable performance of the proposed surrogate MFSWB problems.", "keywords": [ "Sliced Wasserstein Barycenter", "Optimal Transport", "Sliced Wasserstein", "Averaging Measures." ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "NQZNNUsutn", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Karthik Vinay Seetharaman", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Keertana Chidambaram", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Vasilis Syrgkanis", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 54, "n_formula_1": 34, "n_ref_uni": 51, "n_ref": 84, "n_ref_all": 100, "n_fig": 5, "n_tab": 1, "L_tab": 142, "n_element_tab": 11, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5223, "formula_len_all_1": 2589, "len_all": 147785, "len_all_1": 53781, "len_abs": 1106, "len_title": 120, "len_sents": 35380, "len_sents_1": 23348, "n_sents": 296, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1112, "title": "Direct Preference Optimization With Unobserved Preference Heterogeneity", "abs": "RLHF has emerged as a pivotal step in aligning language models with human objectives and values. It typically involves learning a reward model from human preference data and then using reinforcement learning to update the generative model accordingly. Conversely, Direct Preference Optimization (DPO) directly optimizes the generative model with preference data, skipping reinforcement learning. However, both RLHF and DPO assume uniform preferences, overlooking the reality of diverse human annotators. This paper presents a new method to align generative models with varied human preferences. We propose an Expectation-Maximization adaptation to DPO, generating a mixture of models based on latent preference types of the annotators. We then introduce a min-max regret ensemble learning model to produce a single generative method to minimize worst-case regret among annotator subgroups with similar latent factors. Our algorithms leverage the simplicity of DPO while accommodating diverse preferences. Experimental results validate the effectiveness of our approach in producing equitable generative policies.", "keywords": [ "RLHF", "LLM Alignment", "Preference Aggregation" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 1, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "NQZImD0VGP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Haiyun He", "gender": "Female", "institution": "Cornell University", "country": "US", "position": "Postdoc" }, { "name": "Yepeng Liu", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "PhD student" }, { "name": "Yuheng Bu", "gender": "Male", "institution": "University of Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Ziqiao Wang", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yongyi Mao", "gender": "Male", "institution": "University of Ottawa", "country": "CA", "position": "Full Professor" } ], "n_formula": 53, "n_formula_1": 28, "n_ref_uni": 43, "n_ref": 83, "n_ref_all": 109, "n_fig": 5, "n_tab": 3, "L_tab": 618, "n_element_tab": 77, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 592, "n_element_tab_1": 74, "formula_len_all": 5599, "formula_len_all_1": 2284, "len_all": 190974, "len_all_1": 82166, "len_abs": 1211, "len_title": 122, "len_sents": 53016, "len_sents_1": 33152, "n_sents": 523, "n_sents_1": 284, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1226, "title": "Universally Optimal Watermarking Schemes for LLMs: from Theory to Practice", "abs": "Large Language Models (LLMs) boosts human efficiency but also poses misuse risks, with watermarking serving as a reliable method to differentiate AI-generated content from human-created text. In this work, we propose a novel theoretical framework for watermarking LLMs. Particularly, we jointly optimize both the watermarking scheme and detector to maximize detection performance, while controlling the worst-case Type-I error and distortion in the watermarked text. Within our framework, we characterize the _universally minimum Type-II error_, showing a fundamental trade-off between detection performance and distortion. More importantly, we identify the optimal type of detectors and watermarking schemes. Building upon our theoretical analysis, we introduce a practical, model-agnostic and computationally efficient token-level watermarking algorithm that invokes a surrogate model and the Gumbel-max trick. Empirical results on Llama-13B and Mistral-8$\\times$7B demonstrate the effectiveness of our method. Furthermore, we also explore how robustness can be integrated into our theoretical framework, which provides a foundation for designing future watermarking systems with improved resilience to adversarial attacks.", "keywords": [ "Large language models", "watermarking", "hypothesis testing", "distortion-free", "detection", "joint optimization" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "NQTrARs2pz", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Dongping Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Gaoang Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Hongwei Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Katherine Rose Driggs-Campbell", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Tianci Tang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Tielong Cai", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Wenhao Chai", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 70, "n_ref": 122, "n_ref_all": 144, "n_fig": 5, "n_tab": 11, "L_tab": 2202, "n_element_tab": 197, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1915, "n_element_tab_1": 76, "formula_len_all": 158, "formula_len_all_1": 209, "len_all": 248368, "len_all_1": 61521, "len_abs": 1551, "len_title": 130, "len_sents": 66335, "len_sents_1": 29408, "n_sents": 428, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1399, "title": "HomieBot: an Adaptive System for Embodied Mobile Manipulation in Open Environments", "abs": "Embodied Mobile Manipulation in Open Environments (EMMOE) is the challenge that agents understanding user instructions and executing long-horizon everyday tasks in home environments. This challenge encompasses task planning, decision-making, navigation and manipulation, and is crucial to develop a powerful home assistant capable of autonomously completing daily tasks. However, the absence of a holistic benchmark, data incompatibility between large language models (LLMs) and mobile manipulation tasks, the lack of a comprehensive framework, and insufficient dynamic adaptation mechanisms all continue to hinder its development. To address these issues, we propose EMMOE, the first unified benchmark that simultaneously evaluates high-level planners and low-level policies, and new metrics for more diverse evaluation. Additionally, we manually collect EMMOE-100, the first everyday task dataset featuring detailed decision-making processes, Chain-of-Thought (CoT) outputs, feedback from low-level execution and a trainable data format for Large Multimodal Models (LMMs). Furthermore, we design HomieBot, a sophisticated agent system which integrates LMM with Direct Preference Optimization (DPO) as the high-level planner, small navigation and manipulation models as the low-level executor. Finally, we demonstrate HomieBot's performance and methods for evaluating different models and policies.", "keywords": [ "Embodied Agent", "Embodied AI Benchmark", "Mobile Manipulation" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 1, 2 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "NQEe7B7bSw", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Congliang Chen", "gender": "Male", "institution": "The Chinese University of Hong Kong(Shenzhen)", "country": "CN", "position": "PhD student" }, { "name": "Jiancong Xiao", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Postdoc" }, { "name": "Ruoyu Sun", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Tian Xu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Zeyu Qin", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zhi-Quan Luo", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Ziniu Li", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 13, "n_ref_uni": 67, "n_ref": 147, "n_ref_all": 186, "n_fig": 11, "n_tab": 8, "L_tab": 3282, "n_element_tab": 328, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 935, "n_element_tab_1": 103, "formula_len_all": 2086, "formula_len_all_1": 1414, "len_all": 240998, "len_all_1": 74398, "len_abs": 1356, "len_title": 153, "len_sents": 71074, "len_sents_1": 33713, "n_sents": 633, "n_sents_1": 290, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1510, "title": "Preserving Diversity in Supervised Fine-Tuning of Large Language Models", "abs": "Large Language Models (LLMs) typically rely on Supervised Fine-Tuning (SFT) to specialize in downstream tasks, with the Cross Entropy (CE) loss being the de facto choice. However, CE maximizes the likelihood of observed data without accounting for alternative possibilities. As such, CE usually leads to reduced diversity in the model's outputs, which hinders further development that requires sampling to explore better responses. To address this limitation, this paper introduces a new game-theoretic formulation for SFT. In this framework, an auxiliary variable is introduced to regulate the learning process. We prove that the proposed game-theoretic approach connects to the problem of reverse KL minimization with entropy regularization. This regularization prevents over-memorization of training data and promotes output diversity. To implement this framework, we develop GEM, a new training algorithm that is computationally efficient as CE by leveraging some unique properties of LLMs. Empirical studies of pre-trained models from 3B to 70B parameters show that GEM achieves comparable downstream performance to CE while significantly enhancing output diversity. This increased diversity translates to performance gains in test-time compute scaling for chat and code generation tasks. Moreover, we observe that preserving output diversity has the added benefit of mitigating forgetting, as maintaining diverse outputs encourages models to retain pre-trained knowledge throughout the training process.", "keywords": [ "Large language models", "supervised fine-tuning", "diversity", "test-time scaling" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "NPzuN3Rxi8", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Adeel Razi", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" }, { "name": "Ahmad Khajehnejad", "gender": "Male", "institution": ", University of British Columbia", "country": "", "position": "Postdoc" }, { "name": "Brett Joseph Kagan", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Associate Professor" }, { "name": "Forough Habibollahi", "gender": "Female", "institution": "Cortical Labs", "country": "AU", "position": "Postdoc" }, { "name": "Moein Khajehnejad", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Postdoc" } ], "n_formula": 68, "n_formula_1": 22, "n_ref_uni": 17, "n_ref": 32, "n_ref_all": 78, "n_fig": 14, "n_tab": 12, "L_tab": 6236, "n_element_tab": 355, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 1095, "n_element_tab_1": 76, "formula_len_all": 6646, "formula_len_all_1": 3573, "len_all": 210061, "len_all_1": 68739, "len_abs": 4174, "len_title": 145, "len_sents": 68013, "len_sents_1": 30416, "n_sents": 533, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1649, "title": "TAVRNN: Temporal Attention-enhanced Variational Graph RNN Captures Neuronal Dynamics and Behavior", "abs": "We introduce Temporal Attention-enhanced Variational Graph Recurrent Neural Network (TAVRNN), a novel framework for analyzing the evolving dynamics of neuronal connectivity networks in response to external stimuli and behavioral feedback. TAVRNN captures temporal changes in network structure by modeling sequential snapshots of neuronal activity, enabling the identification of key connectivity patterns. Leveraging temporal attention mechanisms and variational graph techniques, TAVRNN uncovers how connectivity shifts align with behavior over time. We validate TAVRNN on two datasets: _in vivo_ calcium imaging data from freely behaving rats and novel _in vitro_ electrophysiological data from the _DishBrain_ system, where biological neurons control a simulated environment during the game of _pong_. We show that TAVRNN outperforms previous baseline models in classification, clustering tasks and computational efficiency while accurately linking connectivity changes to performance variations. Crucially, TAVRNN reveals that high game performance in the _DishBrain_ system correlates with the alignment of sensory and motor subregion channels, a relationship not evident in earlier models.\nThis framework represents the first application of dynamic graph representation of electrophysiological (neuronal) data from _DishBrain_ system, providing insights into the reorganization of neuronal networks during learning. TAVRNN\u2019s ability to differentiate between neuronal states associated with successful and unsuccessful learning outcomes, offers significant implications for real-time monitoring and manipulation of biological neuronal systems.", "keywords": [ "Representation Learning", "Attention", "Graph Recurrent Neuronal Network", "Neuronal Dynamics", "Electrophysiology", "Calcium Imaging", "Behaviour" ], "rating_list": [ 1, 5, 3, 3 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "NPSZ7V1CCY", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Antonia K\u00f6rner", "gender": "unknown", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "MS student" }, { "name": "Kostadin Cvejoski", "gender": "Male", "institution": "LAMARR", "country": "DE", "position": "Postdoc" }, { "name": "Patrick Seifner", "gender": "Male", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "PhD student" }, { "name": "Ramses J Sanchez", "gender": "Male", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "Postdoc" } ], "n_formula": 44, "n_formula_1": 4, "n_ref_uni": 47, "n_ref": 107, "n_ref_all": 182, "n_fig": 10, "n_tab": 19, "L_tab": 18700, "n_element_tab": 2148, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 346, "n_element_tab_1": 66, "formula_len_all": 3359, "formula_len_all_1": 321, "len_all": 333874, "len_all_1": 73533, "len_abs": 1713, "len_title": 123, "len_sents": 103996, "len_sents_1": 35029, "n_sents": 903, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1727, "title": "Zero-shot Imputation with Foundation Inference Models for Dynamical Systems", "abs": "Dynamical systems governed by ordinary differential equations (ODEs) serve as models for a vast number of natural and social phenomena. In this work, we offer a fresh perspective on the classical problem of imputing missing time series data, whose underlying dynamics are assumed to be determined by ODEs. Specifically, we revisit ideas from amortized inference and neural operators, and propose a novel supervised learning framework for *zero-shot time series imputation*, through parametric functions satisfying some (hidden) ODEs. Our proposal consists of two components. First, a broad probability distribution over the space of ODE solutions, observation times and noise mechanisms, with which we generate a large, synthetic dataset of (hidden) ODE solutions, along with their noisy and sparse observations. Second, a neural recognition model that is trained *offline*, to map the generated time series onto the spaces of initial conditions and time derivatives of the (hidden) ODE solutions, which we then integrate to impute the missing data. We empirically demonstrate that *one and the same* (pretrained) recognition model can perform zero-shot imputation across 63 distinct time series with missing values, each sampled from widely different dynamical systems. Likewise, we demonstrate that it can perform zero-shot imputation of missing high-dimensional data in 10 vastly different settings, spanning human motion, air quality, traffic and electricity studies, as well as Navier-Stokes simulations \u2014 *without requiring any fine-tuning*. What is more, our proposal often outperforms state-of-the-art methods, which are trained on the target datasets.\n\nOur pretrained model is available with the supplementary material", "keywords": [ "Zero-shot imputation", "foundation models", "time series imputation", "dynamical systems", "amortized inference", "zero-shot interpolation", "foundation models for time series" ], "rating_list": [ 6, 6, 8, 5 ], "soundness_list": [ 4, 3, 4, 2 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "NPNUHgHF2w", "primary_area": "applications to neuroscience & cognitive science", "n_author": 8, "author_info_list": [ { "name": "Gang Pan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" }, { "name": "Haiteng Jiang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Principal Researcher" }, { "name": "Jiquan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Sha Zhao", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Shijian Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Tao Li", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Yangxuan Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhiling Luo", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 67, "n_ref": 135, "n_ref_all": 174, "n_fig": 9, "n_tab": 26, "L_tab": 18660, "n_element_tab": 1430, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 3226, "n_element_tab_1": 270, "formula_len_all": 993, "formula_len_all_1": 922, "len_all": 288431, "len_all_1": 70482, "len_abs": 1747, "len_title": 110, "len_sents": 76111, "len_sents_1": 30450, "n_sents": 507, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1838, "title": "CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding", "abs": "Electroencephalography (EEG) is a non-invasive technique to measure and record brain electrical activity, widely used in various BCI and healthcare applications. Early EEG decoding methods rely on supervised learning, limited by specific tasks and datasets, hindering model performance and generalizability. With the success of large language models, there is a growing body of studies focusing on EEG foundation models. However, these studies still leave challenges: Firstly, most of existing EEG foundation models employ full EEG modeling strategy. It models the spatial and temporal dependencies between all EEG patches together, but ignores that the spatial and temporal dependencies are heterogeneous due to the unique structural characteristics of EEG signals. Secondly, existing EEG foundation models have limited generalizability on a wide range of downstream BCI tasks due to varying formats of EEG data, making it challenging to adapt to. To address these challenges, we propose a novel foundation model called CBraMod. Specifically, we devise a criss-cross transformer as the backbone to thoroughly leverage the structural characteristics of EEG signals, which can model spatial and temporal dependencies separately through two parallel attention mechanisms. And we utilize an asymmetric conditional positional encoding scheme which can encode positional information of EEG patches and be easily adapted to the EEG with diverse formats. CBraMod is pre-trained on a very large corpus of EEG through patch-based masked EEG reconstruction. We evaluate CBraMod on up to 10 downstream BCI tasks (12 public datasets). CBraMod achieves the state-of-the-art performance across the wide range of tasks, proving its strong capability and generalizability. The source code is publicly available at https://github.com/wjq-learning/CBraMod.", "keywords": [ "Foundation Model; EEG; Criss-Cross Transformer; Downstream BCI tasks" ], "rating_list": [ 8, 5, 8, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "NPLty3VT1c", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Ningyuan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Tianyan Long", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Xiaotie Deng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" } ], "n_formula": 24, "n_formula_1": 24, "n_ref_uni": 25, "n_ref": 40, "n_ref_all": 45, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1880, "formula_len_all_1": 1880, "len_all": 88817, "len_all_1": 66016, "len_abs": 1189, "len_title": 130, "len_sents": 28043, "len_sents_1": 27550, "n_sents": 206, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1192, "title": "Solving Nash Equilibrium Scalably via Deep-Learning-Augmented Iterative Algorithms", "abs": "Computing the Nash Equilibrium (NE) is a fundamental yet computationally challenging problem in game theory. Although recent approaches have incorporated deep learning techniques to tackle this intractability, most of them still struggle with scalability when the number of players increases, due to the exponential growth of computational cost. Inspired by the efficiency of classical learning dynamics methods, we propose a deep learning-augmented Nash equilibrium solver, named Deep Iterative Nash Equilibrium Solver (DINES), based on a novel framework that integrates deep learning into iterative algorithms to solve Nash Equilibria more efficiently. Our approach effectively reduces time complexity to a polynomial level and mitigates the curse of dimensionality by leveraging query-based access to utility functions rather than requiring the full utility matrix. Experimental results demonstrate that our approach achieves better or comparable approximation accuracy compared to existing methods, while significantly reducing computational expense. This advantage is highlighted in large-scale sparse games, which is previously intractable for most existing deep-learning-based methods.", "keywords": [ "Nash Equilibrium", "Game Theory", "Deep Learning" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 1, 1, 2, 2 ], "presentation_list": [ 1, 3, 1, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "NPDnRLFhc0", "primary_area": "datasets and benchmarks", "n_author": 16, "author_info_list": [ { "name": "Ashish Dalvi", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "Undergrad student" }, { "name": "Christopher D. Rosin", "gender": "Male", "institution": "Elsevier", "country": "US", "position": "Sr. Director Data Science" }, { "name": "David E Neal", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Emeritus" }, { "name": "Gene W. Yeo", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "Full Professor" }, { "name": "Gino Prasad", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Hsuan-lin Her", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Jianyou Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Kaicheng Wang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Leon Bergen", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Maxim Khan", "gender": "unknown", "institution": "Elsevier", "country": "", "position": "Product Manager" }, { "name": "Mingwang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "", "position": "Associate Professor" }, { "name": "Qin Yang", "gender": "Female", "institution": "Southwest Jiaotong University", "country": "", "position": "Principal Researcher" }, { "name": "Qishan Liang", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Ramamohan Paturi", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Weili Cao", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Xiaoyue Wang", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Undergrad student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 41, "n_ref_all": 105, "n_fig": 18, "n_tab": 9, "L_tab": 1849, "n_element_tab": 199, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 922, "n_element_tab_1": 148, "formula_len_all": 53, "formula_len_all_1": 38, "len_all": 195010, "len_all_1": 59860, "len_abs": 1049, "len_title": 121, "len_sents": 62418, "len_sents_1": 27227, "n_sents": 582, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1051, "title": "EvidenceBench: A Benchmark for Extracting Evidence from Biomedical Papers", "abs": "We study the task of automatically finding evidence relevant to hypotheses in biomedical papers. Finding relevant evidence is an important stage when humans write systematic reviews about certain scientific hypotheses. We introduce EvidenceBench to measure models performance on this task, which is created by a novel pipeline that consists of hypothesis generation and sentence-by-sentence annotation of biomedical papers for relevant evidence, completely guided by and faithfully following existing human experts judgment. Our pipeline's value and accuracy is validated by teams of human experts. We evaluate a diverse set of language models and retrieval systems on the benchmark and find the performance of the best models still falls significantly short of expert-level on this task. To show the scalability of our proposed pipeline, we create a larger EvidenceBench-100k with 107,461 fully annotated papers with hypotheses to faciliate model training and development. Both datasets are available at https://github.com/EvidenceBench/EvidenceBench", "keywords": [ "Biomedical Benchmark", "Scientific Information Retrieval", "Scientific Information Extraction", "Large Language Models", "BioNLP" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "NOk7t8eKtc", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Bingning Wang", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Kun Zhou", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Postdoc" }, { "name": "Liang Song", "gender": "unknown", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "", "position": "Researcher" }, { "name": "Xin Zhao", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Zhipeng Chen", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "PhD student" }, { "name": "weipeng chen", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 51, "n_ref": 102, "n_ref_all": 122, "n_fig": 6, "n_tab": 9, "L_tab": 873, "n_element_tab": 80, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1116, "n_element_tab_1": 61, "formula_len_all": 599, "formula_len_all_1": 599, "len_all": 226584, "len_all_1": 68051, "len_abs": 1421, "len_title": 151, "len_sents": 43018, "len_sents_1": 32670, "n_sents": 284, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 1485, "title": "Extracting and Transferring Abilities For Building Multi-lingual Ability-enhanced Large Language Models", "abs": "Multi-lingual ability transfer has become increasingly important for the broad application of large language models~(LLMs). Existing work highly relies on training with the multi-lingual ability-related data, which may be not available for low-resource languages. To solve it, we propose a $\\textbf{M}$ulti-lingual $\\textbf{A}$bility $\\textbf{E}$xtraction and $\\textbf{T}$ransfer approach, named as $\\textbf{MAET}$. Our key idea is to decompose and extract language-agnostic ability-related weights from LLMs, and transfer them across different languages by simple addition and subtraction operations without training. Specially, our MAET consists of the extraction and transfer stages. In the extraction stage, we firstly locate key neurons that are highly related to specific abilities, and then employ them to extract the transferable ability-specific weights. In the transfer stage, we further select the ability-related parameter tensors, and design the merging strategy based on the linguistic and ability specific weights, to build the multi-lingual ability-enhanced LLM. To demonstrate the effectiveness of our proposed approach, we conduct extensive experiments on mathematical and scientific tasks in both high-resource lingual and low-resource lingual scenarios. Experiment results have shown that MAET can effectively and efficiently extract and transfer the advanced abilities, and outperform training-based baselines methods. Our code and data will be publicly released.", "keywords": [ "Large Language Models", "Advanced Abilities Transferring", "Multi-lingual Scenarios" ], "rating_list": [ 6, 8, 5, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "NOfmlsnCsS", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Ping Zhang", "gender": "unknown", "institution": "The Ohio State University", "country": "US", "position": "Associate Professor" }, { "name": "Ruoqi Liu", "gender": "unknown", "institution": "Ohio State University", "country": "US", "position": "PhD student" }, { "name": "Xiang Yue", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "yuelin bai", "gender": "Not Specified", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 26, "n_ref": 56, "n_ref_all": 81, "n_fig": 11, "n_tab": 17, "L_tab": 5925, "n_element_tab": 611, "n_fig_1": 6, "n_tab_1": 9, "L_tab_1": 2474, "n_element_tab_1": 439, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 236488, "len_all_1": 64387, "len_abs": 2587, "len_title": 112, "len_sents": 74447, "len_sents_1": 28346, "n_sents": 569, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1382, "title": "Teach Multimodal LLMs to Comprehend Electrocardiographic Images", "abs": "The electrocardiogram (ECG) is an essential non-invasive diagnostic tool for assessing cardiac conditions. Existing automatic interpretation methods suffer from limited generalizability, focusing on a narrow range of cardiac conditions, and typically depend on raw physiological signals, which may not be readily available in resource-limited settings where only printed or digital ECG images are accessible. Recent advancements in multimodal large language models (MLLMs) present promising opportunities for addressing these challenges. However, the application of MLLMs to ECG image interpretation remains challenging due to the lack of instruction tuning datasets and well-established ECG image benchmarks for quantitative evaluation. To address these challenges, we introduce ECGInstruct, a comprehensive ECG image instruction tuning dataset of over 1 million samples, covering a wide range of ECG-related tasks from diverse data sources. Using ECGInstruct, we develop PULSE, a fine-tuned MLLM tailored for ECG image interpretation. In addition, we curate ECGBench, a new evaluation benchmark covering four key ECG image interpretation tasks. Our experiments show\nthat PULSE sets a new state-of-the-art, outperforming general MLLMs with an average accuracy improvement of 15% to 30%. This work highlights the potential of PULSE to enhance ECG interpretation in clinical practice", "keywords": [ "Electrocardiogram", "LLMs", "Multimodal LLMs", "Instruction Tuning", "Benchmark and Evaluation" ], "rating_list": [ 5, 3, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "NO6Tv6QcDs", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Florian E. Dorner", "gender": "Male", "institution": "Max Planck Institute for Intelligent Systems, Max-Planck Institute", "country": "", "position": "Researcher" }, { "name": "Moritz Hardt", "gender": "Not Specified", "institution": "Max-Planck-Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "Principal Researcher" }, { "name": "Vivian Yvonne Nastl", "gender": "Female", "institution": "Max Planck Institute for Intelligent Systems", "country": "", "position": "PhD student" } ], "n_formula": 47, "n_formula_1": 13, "n_ref_uni": 41, "n_ref": 62, "n_ref_all": 81, "n_fig": 12, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4854, "formula_len_all_1": 767, "len_all": 185172, "len_all_1": 65179, "len_abs": 2569, "len_title": 133, "len_sents": 57459, "len_sents_1": 30802, "n_sents": 515, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 85, "L_abs": 1366, "title": "Limits to scalable evaluation at the frontier: LLM as judge won\u2019t beat twice the data", "abs": "High quality annotations are increasingly a bottleneck in the explosively growing machine learning ecosystem. Scalable evaluation methods that avoid costly annotation have therefore become an important research ambition. Many hope to use strong existing models in lieu of costly labels to provide cheap model evaluations. Unfortunately, this method of using models as judges introduces biases, such as self-preferencing, that can distort model comparisons. An emerging family of debiasing tools promises to fix these issues by using a few high quality labels to debias a large number of model judgments. In this paper, we study how far such debiasing methods, in principle, can go. Our main result shows that when the judge is no more accurate than the evaluated model, no debiasing method can decrease the required amount of ground truth labels by more than half. Our result speaks to the severe limitations of the LLM-as-a-judge paradigm at the evaluation frontier where the goal is to assess newly released models that are possibly better than the judge. Through an empirical evaluation, we demonstrate that the sample size savings achievable in practice are even more modest than what our theoretical limit suggests. Along the way, our work provides new observations about debiasing methods for model evaluation, and points out promising avenues for future work.", "keywords": [ "Evaluation", "Benchmarking", "Model-as-a-judge", "Theory" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 3, 4, 4, 3 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "NNUiUwQWx6", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 6, "author_info_list": [ { "name": "Jiannong Cao", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "CN", "position": "Full Professor" }, { "name": "Junnan Dong", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Qinggang Zhang", "gender": "Male", "institution": "The Hong Kong Polytechnic University, Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Wen Hua", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "", "position": "Associate Professor" }, { "name": "Xiao Huang", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "", "position": "Assistant Professor" }, { "name": "CHEN SHENGYUAN", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Postdoc" } ], "n_formula": 23, "n_formula_1": 18, "n_ref_uni": 26, "n_ref": 50, "n_ref_all": 63, "n_fig": 5, "n_tab": 3, "L_tab": 726, "n_element_tab": 12, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1469, "n_element_tab_1": 34, "formula_len_all": 1496, "formula_len_all_1": 1153, "len_all": 108510, "len_all_1": 62379, "len_abs": 1184, "len_title": 106, "len_sents": 37074, "len_sents_1": 27916, "n_sents": 272, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1190, "title": "Neuro-symbolic Entity Alignment via Variational Inference", "abs": "Entity alignment (EA) aims to merge two knowledge graphs (KGs) by identifying equivalent entity pairs. Existing methods can be categorized into symbolic and neural models. Symbolic models, while precise, struggle with substructure heterogeneity and sparsity, whereas neural models, although effective, generally lack interpretability and cannot handle uncertainty. We propose NeuSymEA, a probabilistic neuro-symbolic framework that combines the strengths of both methods. NeuSymEA models the joint probability of all possible pairs' truth scores in a Markov random field, regulated by a set of rules, and optimizes it with the variational EM algorithm. In the E-step, a neural model parameterizes the truth score distributions and infers missing alignments. In the M-step, the rule weights are updated based on the observed and inferred alignments. To facilitate interpretability, we further design a path-ranking-based explainer upon this framework that generates supporting rules for the inferred alignments. Experiments on benchmarks demonstrate that NeuSymEA not only significantly outperforms baselines in terms of effectiveness and robustness, but also provides interpretable results.", "keywords": [ "Entity alignment", "neuro-symbolic reasoning", "graph representation learning" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "NNBAzdF7Cg", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "Aditya Kar", "gender": "Male", "institution": "CNRS", "country": "", "position": "PhD student" }, { "name": "Emiliano Lorini", "gender": "Male", "institution": "Toulouse University", "country": "", "position": "Principal Researcher" }, { "name": "Timoth\u00e9e Masquelier", "gender": "unknown", "institution": "CNRS", "country": "", "position": "Principal Researcher" } ], "n_formula": 20, "n_formula_1": 18, "n_ref_uni": 39, "n_ref": 63, "n_ref_all": 76, "n_fig": 3, "n_tab": 5, "L_tab": 740, "n_element_tab": 107, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 793, "n_element_tab_1": 104, "formula_len_all": 1838, "formula_len_all_1": 1482, "len_all": 122816, "len_all_1": 67447, "len_abs": 955, "len_title": 95, "len_sents": 39277, "len_sents_1": 31152, "n_sents": 278, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 994, "title": "Binary Spiking Neural Networks as causal models", "abs": "In this paper, we provide a causal analysis of binary spiking neural networks (BSNNs)\naimed at explaining their behaviors. \nWe formally define a BSNN \nand represent its spiking activity\n as a binary causal model.\nThanks to this causal representation, \nwe are able to explain the output of the network\nby leveraging logic-based methods. \nIn particular,\nwe show that we can successfully \nuse a SAT (Boolean satisfiability) solver to compute \n abductive explanations from this binary causal model. \nTo illustrate our approach, \nwe trained the BSNN on the standard MNIST\ndataset and applied our SAT-based method to\nfinding abductive explanations of the network's classifications\nbased on pixel-level features. We also compared the found explanations against SHAP, a popular \nmethod used in the area of explainable\nAI to explain ``black box'' classifiers.\nWe show that, unlike SHAP,\nour method guarantees that a found explanation does\nnot contain completely irrelevant features.", "keywords": [ "Explainability", "Causal reasoning", "Spiking Neural Networks", "White-box" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 3, 5, 2, 3 ] }, { "paper_id": "NN6QHwgRrQ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Ali Anwar", "gender": "Male", "institution": "University of Minnesota", "country": "US", "position": "Assistant Professor" }, { "name": "Ammar Ahmed", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" }, { "name": "Enmao Diao", "gender": "Male", "institution": "ColAI", "country": "CN", "position": "Researcher" }, { "name": "Jie Ding", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "Associate Professor" }, { "name": "Nathalie Baracaldo", "gender": "unknown", "institution": "University of Pittsburgh", "country": "", "position": "PhD student" }, { "name": "Qi Le", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "", "position": "PhD student" }, { "name": "Xinran Wang", "gender": "Female", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" }, { "name": "Yi Zhou", "gender": "Female", "institution": "International Business Machines", "country": "US", "position": "Researcher" } ], "n_formula": 66, "n_formula_1": 19, "n_ref_uni": 32, "n_ref": 54, "n_ref_all": 91, "n_fig": 12, "n_tab": 5, "L_tab": 1551, "n_element_tab": 229, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3144, "formula_len_all_1": 1086, "len_all": 215166, "len_all_1": 66558, "len_abs": 1484, "len_title": 88, "len_sents": 76551, "len_sents_1": 30156, "n_sents": 695, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 40, "L_abs": 1489, "title": "MAP: Multi-Human-Value Alignment Palette", "abs": "Ensuring that generative AI systems align with human values is essential but challenging, especially when considering multiple human values and their potential trade-offs. Since human values can be personalized and dynamically change over time, the desirable levels of value alignment vary across different ethnic groups, industry sectors, and user cohorts. Within existing frameworks, it is hard to define human values and align AI systems accordingly across different directions simultaneously, such as harmlessness, helpfulness, and positiveness. To address this, we develop a novel, first-principle approach called Multi-Human-Value Alignment Palette (MAP), which navigates the alignment across multiple human values in a structured and reliable way. MAP formulates the alignment problem as an optimization task with user-defined constraints, which define human value targets. It can be efficiently solved via a primal-dual approach, which determines whether a user-defined alignment target is achievable and how to achieve it. We conduct a detailed theoretical analysis of MAP by quantifying the trade-offs between values, the sensitivity to constraints, the fundamental connection between multi-value alignment and sequential alignment, and proving that linear weighted rewards are sufficient for multi-value alignment. Extensive experiments demonstrate MAP's ability to align multiple values in a principled manner while delivering strong empirical performance across various tasks.", "keywords": [ "Human value alignment", "Generative model" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 4, 4 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "NLfWQfy5zp", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Thomas Mesquida", "gender": "Male", "institution": "CEA", "country": "FR", "position": "Researcher" }, { "name": "Yannick Malot", "gender": "unknown", "institution": "CEA", "country": "FR", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 36, "n_ref": 47, "n_ref_all": 63, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 430, "formula_len_all_1": 327, "len_all": 110507, "len_all_1": 53563, "len_abs": 1216, "len_title": 153, "len_sents": 29151, "len_sents_1": 26631, "n_sents": 197, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1227, "title": "Exploring the Trade-Off between Model Complexity and Numerical Precision for Efficient Edge AI Inference", "abs": "When considering the compression of neural networks, the adoption of low-bit representations for both parameters and activations has demonstrated significant efficacy. The process of learning quantized weights through Quantization Aware Training (QAT) stands out as a powerful means to substantially diminish the memory requirements for a specific model to efficiently perform inference. However, despite the numerous works reporting the gains achieved using QAT, a comparison with a notably simpler technique - reducing the model's complexity using fewer parameters - is often absent.\n In this paper, we attemp to answer a seemingly simple question: to reduce a given model's storage requirements, is it better to reduce the number of parameters in the model or to reduce the numerical precision? We explore the trade-off between the dimensionality of parameters and activations one can afford to keep in memory, and the numerical precision used to represent them. Through our experiments in image classification, keyword spotting and language modelling, our results suggest that quantizing weights to $2$ bits and keeping a high number of parameters seems optimal, regardless of the task considered and model architecture.", "keywords": [ "Neural networks", "Edge AI", "Artificial Intelligence", "Model compression" ], "rating_list": [ 3, 6, 5, 1 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 3, 2, 1 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "NLAKxnnSuW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Dailan He", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jun Zhang", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Tongda Xu", "gender": "Non-Binary", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xingtong Ge", "gender": "Male", "institution": "Sensetime", "country": "CN", "position": "Researcher" }, { "name": "Xinjie Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" }, { "name": "Yan Wang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yifan Zhang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zehong Lin", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Research Assistant Professor" }, { "name": "Zhening Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 30, "n_ref": 84, "n_ref_all": 103, "n_fig": 8, "n_tab": 4, "L_tab": 3761, "n_element_tab": 656, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2913, "n_element_tab_1": 338, "formula_len_all": 604, "formula_len_all_1": 528, "len_all": 118754, "len_all_1": 64801, "len_abs": 611, "len_title": 112, "len_sents": 31488, "len_sents_1": 28053, "n_sents": 216, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1563, "title": "MEGA: Memory-Efficient 4D Gaussian Splatting for Dynamic Scenes", "abs": "4D Gaussian Splatting (4DGS) has recently emerged as a promising technique for capturing complex dynamic 3D scenes with high fidelity. It utilizes a 4D Gaussian representation and a GPU-friendly rasterizer, enabling rapid rendering speeds. Despite its advantages, 4DGS faces significant challenges, notably the requirement of millions of 4D Gaussians, each with extensive associated attributes, leading to substantial memory and storage cost. This paper introduces a memory-efficient framework for 4DGS. We streamline the color attribute by decomposing it into a per-Gaussian direct color component with only 3 parameters and a shared lightweight alternating current color predictor. This approach eliminates the need for spherical harmonics coefficients, which typically involve up to 144 parameters in classic 4DGS, thereby creating a memory-efficient 4D Gaussian representation. Furthermore, we introduce an entropy-constrained Gaussian deformation technique that uses a deformation field to expand the action range of each Gaussian and integrates an opacity-based entropy loss to limit the number of Gaussians, thus forcing our model to use as few Gaussians as possible to fit a dynamic scene well. With simple half-precision storage and zip compression, our framework achieves a storage reduction by approximately 190$\\times$ and 125$\\times$ on the Technicolor and Neural 3D Video datasets, respectively, compared to the original 4DGS. Meanwhile, it maintains comparable rendering speeds and scene representation quality, setting a new standard in the field.", "keywords": [ "4D Gaussian Splatting", "Compression" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "NKotdPUc3L", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Zhengyuan Zhou", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Zijian Liu", "gender": "unknown", "institution": "New York University", "country": "", "position": "PhD student" } ], "n_formula": 86, "n_formula_1": 33, "n_ref_uni": 36, "n_ref": 101, "n_ref_all": 111, "n_fig": 0, "n_tab": 3, "L_tab": 1097, "n_element_tab": 329, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 224, "n_element_tab_1": 36, "formula_len_all": 11043, "formula_len_all_1": 3271, "len_all": 163694, "len_all_1": 63533, "len_abs": 1433, "len_title": 154, "len_sents": 40980, "len_sents_1": 26990, "n_sents": 339, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 106, "L_abs": 1619, "title": "Nonconvex Stochastic Optimization under Heavy-Tailed Noises: Optimal Convergence without Gradient Clipping", "abs": "Recently, the study of heavy-tailed noises in first-order nonconvex stochastic optimization has gotten a lot of attention since it was recognized as a more realistic condition as suggested by many empirical observations. Specifically, the stochastic noise (the difference between the stochastic and true gradient) is considered only to have a finite $\\mathfrak{p}$-th moment where $\\mathfrak{p}\\in\\left(1,2\\right]$ instead of assuming it always satisfies the classical finite variance assumption. To deal with this more challenging setting, people have proposed different algorithms and proved them to converge at an optimal $\\mathcal{O}(T^{\\frac{1-\\mathfrak{p}}{3\\mathfrak{p}-2}})$ rate for smooth objectives after $T$ iterations. Notably, all these new-designed algorithms are based on the same technique \u2013 gradient clipping. Naturally, one may want to know whether the clipping method is a necessary ingredient and the only way to guarantee convergence under heavy-tailed noises. In this work, by revisiting the existing Batched Normalized Stochastic Gradient Descent with Momentum (Batched NSGDM) algorithm, we provide the first convergence result under heavy-tailed noises but without gradient clipping. Concretely, we prove that Batched NSGDM can achieve the optimal $\\mathcal{O}(T^{\\frac{1-\\mathfrak{p}}{3\\mathfrak{p}-2}})$ rate even under the relaxed smooth condition. More interestingly, we also establish the first $\\mathcal{O}(T^{\\frac{1-\\mathfrak{p}}{2\\mathfrak{p}}})$ convergence rate in the case where the tail index $\\mathfrak{p}$ is unknown in advance, which is arguably the common scenario in practice.", "keywords": [ "Stochastic Optimization", "Heavy-Tailed Noises" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "NKOWxemSb4", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Haoyu Peter Wang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Nan Wu", "gender": "Female", "institution": "George Washington University", "country": "US", "position": "Assistant Professor" }, { "name": "Pan Li", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Yinan Huang", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 1, "n_ref_uni": 51, "n_ref": 139, "n_ref_all": 169, "n_fig": 10, "n_tab": 32, "L_tab": 26079, "n_element_tab": 1494, "n_fig_1": 3, "n_tab_1": 11, "L_tab_1": 2323, "n_element_tab_1": 226, "formula_len_all": 1419, "formula_len_all_1": 160, "len_all": 353051, "len_all_1": 75461, "len_abs": 1416, "len_title": 121, "len_sents": 60922, "len_sents_1": 36060, "n_sents": 567, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1370, "title": "A Benchmark on Directed Graph Representation Learning in Hardware Designs", "abs": "To keep pace with the rapid advancements in design complexity within modern computing systems, directed graph representation learning (DGRL) has become crucial, particularly for encoding circuit netlists, computational graphs, and developing surrogate models for hardware performance prediction. However, DGRL remains relatively unexplored, especially in the hardware domain, mainly due to the lack of comprehensive and user-friendly benchmarks. This study presents a novel benchmark comprising five hardware design datasets and 13 prediction tasks spanning various levels of circuit abstraction. We evaluate 21 DGRL models, employing diverse graph neural networks and graph transformers (GTs) as backbones, enhanced by positional encodings (PEs) tailored for directed graphs. Our results highlight that bidirected (BI) message passing neural networks (MPNNs) and robust PEs significantly enhance model performance. Notably, the top-performing models include PE-enhanced GTs interleaved with BI-MPNN layers and BI-Graph Isomorphism Network, both surpassing baselines across the 13 tasks. Additionally, our investigation into out-of-distribution (OOD) performance emphasizes the urgent need to improve OOD generalization in DGRL models. This benchmark, implemented with a modular codebase, streamlines the evaluation of DGRL models for both hardware and ML practitioners.", "keywords": [ "graph neural network", "hardware design", "positional encoding", "graph transformers" ], "rating_list": [ 3, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "NKE7VTxVKL", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Fragkiskos D. Malliaros", "gender": "Male", "institution": "CentraleSup\u00e9lec, Inria, Paris-Saclay University", "country": "FR", "position": "Associate Professor" }, { "name": "Johannes F. Lutzeyer", "gender": "Male", "institution": "Ecole Polytechique", "country": "FR", "position": "Assistant Professor" }, { "name": "Michalis Vazirgiannis", "gender": "Male", "institution": "Ecole Polytechnique, France", "country": "", "position": "Full Professor" }, { "name": "Yassine ABBAHADDOU", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "FR", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 11, "n_ref_uni": 65, "n_ref": 94, "n_ref_all": 113, "n_fig": 2, "n_tab": 13, "L_tab": 17287, "n_element_tab": 1287, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2511, "n_element_tab_1": 132, "formula_len_all": 4246, "formula_len_all_1": 589, "len_all": 206864, "len_all_1": 72989, "len_abs": 1033, "len_title": 106, "len_sents": 49145, "len_sents_1": 33515, "n_sents": 404, "n_sents_1": 243, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1038, "title": "Centrality Graph Shift Operators for Graph Neural Networks", "abs": "Graph Shift Operators (GSOs), such as the adjacency and graph Laplacian matrices, play a fundamental role in graph theory and graph representation learning. Traditional GSOs are typically constructed by normalizing the adjacency matrix by the degree matrix, a local centrality metric. In this work, we instead propose and study Centrality GSOs (CGSOs), which normalize adjacency matrices by global centrality metrics such as the PageRank, $k$-core or count of fixed length paths. We study spectral properties of the CGSOs, allowing us to get an understanding of their action on graph signals. We confirm this understanding by defining and running the spectral clustering algorithm based on different CGSOs on several synthetic and real-world datasets. We furthermore outline how our CGSO can act as the message passing operator in any Graph Neural Network and in particular demonstrate strong performance of a variant of the Graph Convolutional Network and Graph Attention Network using our CGSOs on several real-world benchmark datasets.", "keywords": [ "Graph Neural Networks", "Graph Shift Operators", "Centrality" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "NK09Bcvuxl", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Yibo Jiang", "gender": "Male", "institution": "University of Chicago", "country": "", "position": "PhD student" }, { "name": "Yuxin Chen", "gender": "unknown", "institution": "University of Chicago", "country": "US", "position": "Assistant Professor" }, { "name": "Zhuokai Zhao", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 17, "n_ref_uni": 39, "n_ref": 81, "n_ref_all": 111, "n_fig": 14, "n_tab": 1, "L_tab": 127, "n_element_tab": 11, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 127, "n_element_tab_1": 11, "formula_len_all": 8557, "formula_len_all_1": 8414, "len_all": 138784, "len_all_1": 79993, "len_abs": 2883, "len_title": 110, "len_sents": 36540, "len_sents_1": 32686, "n_sents": 242, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 947, "title": "Direct Acquisition Optimization for Low-Budget Active Learning", "abs": "Active Learning (AL) has gained prominence in integrating data-intensive machine learning (ML) models into domains with limited labeled data. However, its effectiveness diminishes significantly when the labeling budget is low. In this paper, we first empirically observe the performance degradation of existing AL algorithms in the low-budget settings, and then introduce Direct Acquisition Optimization (DAO), a novel AL algorithm that optimizes sample selections based on expected true loss reduction. Specifically, DAO utilizes influence functions to update model parameters and incorporates an additional acquisition strategy to mitigate bias in loss estimation. This approach facilitates a more accurate estimation of the overall error reduction, without extensive computations or reliance on labeled data. Experiments demonstrate DAO\u2019s effectiveness in low budget settings, outperforming state-of-the-arts approaches across seven benchmarks.", "keywords": [ "Low-Budget Active Learning", "Data Scarcity", "Label-Efficient Learning" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "NJxCpMt0sf", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Liyue Shen", "gender": "Female", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" }, { "name": "Luning Wang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "MS student" }, { "name": "Zhengxu Tang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Undergrad student" }, { "name": "Zitao Shuai", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "MS student" }, { "name": "Chenwei Wu", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 46, "n_ref": 80, "n_ref_all": 102, "n_fig": 8, "n_tab": 3, "L_tab": 934, "n_element_tab": 196, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 893, "n_element_tab_1": 166, "formula_len_all": 509, "formula_len_all_1": 483, "len_all": 157967, "len_all_1": 71201, "len_abs": 2024, "len_title": 144, "len_sents": 38759, "len_sents_1": 33437, "n_sents": 280, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 2026, "title": "Dynamic Modeling of Patients, Modalities and Tasks via Multi-modal Multi-task Mixture of Experts", "abs": "Multi-modal multi-task learning holds significant promise in tackling complex diagnostic tasks and many significant medical imaging problems. It fulfills the needs in real-world diagnosis protocol to leverage information from different data sources and simultaneously perform mutually informative tasks. However, medical imaging domains introduce two key challenges: dynamic modality fusion and modality-task dependence. The quality and amount of task-related information from different modalities could vary significantly across patient samples, due to biological and demographic factors. Traditional fusion methods apply fixed combination strategies that fail to capture this dynamic relationship, potentially underutilizing modalities that carry stronger diagnostic signals for specific patients. Additionally, different clinical tasks may require dynamic feature selection and combination from various modalities, a phenomenon we term \u201cmodality-task dependence.\u201d To address these issues, we propose M4oE, a novel Multi-modal Multi-task Mixture of Experts framework for precise Medical diagnosis. M4oE comprises Modality-Specific (MSoE) modules and a Modality-shared Modality-Task MoE (MToE) module. With collaboration from both modules, our model dynamically decomposes and learns distinct and shared information from different modalities and achieves dynamic fusion. MToE provides a joint probability model of modalities and tasks by using experts as a link and encourages experts to learn modality-task dependence via conditional mutual information loss. By doing so, M4oE offers sample and population-level interpretability of modality contributions. We evaluate M4oE on four public multi-modal medical benchmark datasets for solving two important medical diagnostic problems including breast cancer screening and retinal disease diagnosis. Results demonstrate our method's superiority over state-of-the-art methods under different metrics of classification and segmentation tasks like Accuracy, AUROC, AUPRC, and DICE.", "keywords": [ "Multimodal Learning", "Medical Imaging" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "NJqu7xwXZk", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Bin Gu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Boyu Wang", "gender": "Male", "institution": "University of Western Ontario", "country": "CA", "position": "Assistant Professor" }, { "name": "Charles Ling", "gender": "Male", "institution": "Western University", "country": "CA", "position": "Professor" }, { "name": "Ganyu Wang", "gender": "Male", "institution": "Western University", "country": "CA", "position": "PhD student" }, { "name": "Jinjie Fang", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "PhD student" }, { "name": "Xi Chen", "gender": "Male", "institution": "Proactive AI Lab, Inc.", "country": "CA", "position": "Principal Researcher" }, { "name": "Maxwell Yin", "gender": "unknown", "institution": "University of Western Ontario", "country": "", "position": "PhD student" } ], "n_formula": 126, "n_formula_1": 28, "n_ref_uni": 44, "n_ref": 94, "n_ref_all": 108, "n_fig": 2, "n_tab": 8, "L_tab": 1854, "n_element_tab": 164, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1444, "n_element_tab_1": 108, "formula_len_all": 26023, "formula_len_all_1": 2373, "len_all": 190321, "len_all_1": 70012, "len_abs": 1194, "len_title": 129, "len_sents": 45373, "len_sents_1": 28569, "n_sents": 410, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1203, "title": "FedOne: Query-Efficient Federated Learning for Black-box Discrete Prompt Learning", "abs": "Black-Box Discrete Prompt Learning (BDPL) is a prompt-tuning method that optimizes discrete prompts without accessing model parameters or gradients, making the prompt tuning on a cloud-based Large Language Model (LLM) feasible.\nAdapting Federated Learning (FL) to BDPL could further enhance prompt tuning performance by leveraging data from diverse sources. \nHowever, all previous research on federated black-box prompt tuning had neglected the substantial query cost associated with the cloud-based LLM service. \nTo address this gap, we conducted a theoretical analysis of query efficiency within the context of federated black-box prompt tuning. Our findings revealed that degrading FedAvg to activate only one client per round, a strategy we called *FedOne*, enabled optimal query efficiency in federated black-box prompt learning. \nBuilding on this insight, we proposed the FedOne framework, a federated black-box discrete prompt learning method designed to maximize query efficiency when interacting with cloud-based LLMs.\nWe conducted numerical experiments on various aspects of our framework, demonstrating a significant improvement in query efficiency, which aligns with our theoretical results.", "keywords": [ "Federated Learning", "Black-box Prompt Learning", "Large Language Model" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "NJqsHgxcKh", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Haixu Wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jianmin Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Jiaxiang Dong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Li Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Yuxuan Wang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Mingsheng Long", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 74, "n_ref_all": 112, "n_fig": 11, "n_tab": 17, "L_tab": 18090, "n_element_tab": 951, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 4672, "n_element_tab_1": 182, "formula_len_all": 465, "formula_len_all_1": 496, "len_all": 205179, "len_all_1": 69143, "len_abs": 2036, "len_title": 123, "len_sents": 61683, "len_sents_1": 31135, "n_sents": 383, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1719, "title": "Metadata Matters for Time Series: Informative Forecasting with Transformers", "abs": "Time series forecasting is prevalent in extensive real-world applications, such as financial analysis and energy planning. Previous studies primarily focus on time series modality, endeavoring to capture the intricate variations and dependencies inherent in time series. Beyond numerical time series data, we notice that metadata (e.g. dataset and variate descriptions) also carries valuable information essential for forecasting, which can be used to identify the application scenario and provide more interpretable knowledge than digit sequences. Inspired by this observation, we propose a Metadata-informed Time Series Transformer MetaTST, which incorporates multiple levels of context-specific metadata into Transformer forecasting models to enable informative time series forecasting. To tackle the unstructured nature of metadata, MetaTST formalizes them into natural languages by pre-designed templates and leverages large language models (LLMs) to encode these texts into metadata tokens as a supplement to classic series tokens, resulting in an informative embedding. Further, a Transformer encoder is employed to communicate series and metadata tokens, which can extend series representations by metadata information for more accurate forecasting. This design also allows the model to adaptively learn context-specific patterns across various scenarios, which is particularly effective in handling large-scale, diverse-scenario forecasting tasks. Experimentally, MetaTST achieves state-of-the-art compared to advanced time series models and LLM-based methods on widely acknowledged short- and long-term forecasting benchmarks, covering both single-dataset individual and multi-dataset joint training settings.", "keywords": [ "time series forecasting", "deep time series analysis", "large language models" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 5, 5 ] }, { "paper_id": "NJUzUq2OIi", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Jianyou Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Leon Bergen", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Longtian Bao", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Qirui Zheng", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "Undergrad student" }, { "name": "Ramamohan Paturi", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Taylor Berg-Kirkpatrick", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "Associate Professor" }, { "name": "Weili Cao", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Youze Zheng", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Undergrad student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 39, "n_ref": 88, "n_ref_all": 132, "n_fig": 7, "n_tab": 20, "L_tab": 9355, "n_element_tab": 1706, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 19, "n_element_tab_1": 1, "formula_len_all": 60, "formula_len_all_1": 0, "len_all": 263035, "len_all_1": 53943, "len_abs": 1170, "len_title": 99, "len_sents": 68086, "len_sents_1": 25251, "n_sents": 533, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1178, "title": "Efficient Full-Context Retrieval for Long Documents", "abs": "Long document question answering is challenging due to the quadratic computational cost of transformer-based LLMs. In resource-constrained environments, Retrieval-Augmented Generation (RAG) uses document chunking to maintain linear computational costs but often loses sight of the global context. We introduce the Mamba retriever 130M and Mamba retriever 1.3B retrievers, capable of processing entire long documents in linear time and integrating earlier context to retrieve relevant sentences for answering questions. Mamba retrievers outperform state-of-the-art embedding models across 41 long-document Q\\&A benchmarks while maintaining speed and computational efficiency. Their performance is comparable to GPT-4o on long documents over 256k tokens while using significantly fewer tokens. Mamba retrievers are trained on synthetic data generated from our novel link-based method, which enhances the retrievers' ability to leverage long-range document connections. We further demonstrate the effectiveness of our link-based method over baseline synthetic data methods. All code, datasets, and model checkpoints are available at https://github.com/MambaRetriever/MambaRetriever", "keywords": [ "Long-context retrieval", "Efficient language models", "Mamba architecture", "Synthetic data generation" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "NIkfix2eDQ", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Alex Lewandowski", "gender": "unknown", "institution": "University of Alberta", "country": "", "position": "PhD student" }, { "name": "Dale Schuurmans", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Professor" }, { "name": "Marlos C. Machado", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 46, "n_formula_1": 8, "n_ref_uni": 59, "n_ref": 108, "n_ref_all": 138, "n_fig": 16, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2721, "formula_len_all_1": 438, "len_all": 178917, "len_all_1": 75146, "len_abs": 1227, "len_title": 43, "len_sents": 62826, "len_sents_1": 38636, "n_sents": 479, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 1243, "title": "Plastic Learning with Deep Fourier Features", "abs": "Deep neural networks can struggle to learn continually in the face of non-stationarity. This phenomenon is known as loss of plasticity. In this paper, we identify underlying principles that lead to plastic algorithms. In particular, we provide theoretical results showing that shallow linear networks, as well as a special case of deep linear networks, do not suffer from loss of plasticity. We then propose \\emph{deep Fourier features}, which are the concatenation of a sine and cosine in every layer, and we show that this combination provides a dynamic balance between the trainability obtained through linearity and the effectiveness obtained through the nonlinearity of neural networks. Deep networks composed entirely of deep Fourier features are highly trainable and sustain their trainability over the course of learning. Our empirical results show that continual learning performance can be drastically improved by replacing \\texttt{ReLU} activations with deep Fourier features. These results hold for different continual learning scenarios (e.g., label noise, class incremental learning, pixel permutations) on all major supervised learning datasets used for continual learning research, such as CIFAR10, CIFAR100, and tiny-ImageNet.", "keywords": [ "Fourier", "plasticity", "neural networks", "continual learning" ], "rating_list": [ 6, 6, 5, 8, 6 ], "soundness_list": [ 3, 3, 4, 4, 3 ], "presentation_list": [ 4, 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "NIhRwzqhUz", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Anthony Nixon", "gender": "unknown", "institution": "Lancaster University", "country": "GB", "position": "Full Professor" }, { "name": "David S.Lesley", "gender": "Male", "institution": "Lancaster University", "country": "", "position": "Full Professor" }, { "name": "Jack Trainer", "gender": "Male", "institution": "Lancaster University", "country": "GB", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 16, "n_ref": 20, "n_ref_all": 28, "n_fig": 5, "n_tab": 1, "L_tab": 201, "n_element_tab": 29, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 201, "n_element_tab_1": 29, "formula_len_all": 1007, "formula_len_all_1": 940, "len_all": 69966, "len_all_1": 55905, "len_abs": 1081, "len_title": 106, "len_sents": 27065, "len_sents_1": 26252, "n_sents": 202, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1084, "title": "Learning the Partially Dynamic Travelling Salesman Problem", "abs": "Learning to solve the Travelling Salesman Problem (TSP) using Deep Reinforcement Learning (Deep RL) and Graph Neural Networks (GNNs) has shown promising results for small instances of the problem. We demonstrate that these methods can be extended to solve instances of a partially dynamic variant of the TSP. Solving this partially dynamic variant more effectively exploits the strengths of reinforcement learning and also presents challenges for more established methods of solving the TSP. We show the policies trained using Deep RL outperform modified versions of TSP solvers and heuristics for different distributions of dynamic vertices, including on larger instances than the policies were trained on. This shows the promise of Deep RL for solving this type of dynamic routing problem which is predicted to become of great importance as logistical services become more flexible and responsive to customer demand. Furthermore, our method is a general purpose approach to Deep RL where the problem consists of selecting items from a dynamically-evolving and arbitrarily-sized set.", "keywords": [ "Reinforcement Learning", "Graph Neural Networks", "Dynamic Graphs", "Dynamic Graph Neural Networks", "Travelling Salesman Problem", "Combinatorial Optimisation" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 1, 1, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "NIG8O2zQSQ", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Linghe Kong", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Qingfeng He", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Qiyang Li", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "Undergrad student" }, { "name": "Rui Kong", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Weijun Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Xinyu Fang", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "MS student" }, { "name": "Yazhu Dong", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Yuanchun Li", "gender": "Male", "institution": "Institute for AI Industry Research, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yunxin Liu", "gender": "unknown", "institution": "Institute for AI Industry Research, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "qingtian feng", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 33, "n_ref": 74, "n_ref_all": 106, "n_fig": 3, "n_tab": 7, "L_tab": 1926, "n_element_tab": 252, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1618, "n_element_tab_1": 205, "formula_len_all": 848, "formula_len_all_1": 816, "len_all": 136714, "len_all_1": 65739, "len_abs": 1254, "len_title": 140, "len_sents": 34779, "len_sents_1": 28593, "n_sents": 252, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1260, "title": "LoRA-Switch: Boosting the Efficiency of Dynamic LLM Adapters via System-Algorithm Co-design", "abs": "Recent literature has found that an effective method to customize or further improve large language models (LLMs) is to add dynamic adapters, such as low-rank adapters (LoRA) with Mixture-of-Experts (MoE) structures. Though such dynamic adapters incur modest computational complexity, they surprisingly lead to huge inference latency overhead, slowing down the decoding speed by 2.5+ times. In this paper, we analyze the fine-grained costs of the dynamic adapters and find that the fragmented CUDA kernel calls are the root cause. Therefore, we propose LoRA-Switch, a system-algorithm co-designed architecture for efficient dynamic adapters. Unlike most existing dynamic structures that adopt layer-wise or block-wise dynamic routing, LoRA-Switch introduces a token-wise routing mechanism. It switches the LoRA adapters and weights for each token and merges them into the backbone for inference. For efficiency, this switching is implemented with an optimized CUDA kernel, which fuses the merging operations for all LoRA adapters at once. Based on experiments with popular open-source LLMs on common benchmarks, our approach has demonstrated similar accuracy improvement as existing dynamic adapters, while reducing the decoding latency by more than 2.4 times.", "keywords": [ "Dynamic adapters", "mixture of adapters", "inference acceleration", "lora merging" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "NI8AUSAc4i", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Chao Zhang", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Hao Cheng", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Researcher" }, { "name": "Kuan Wang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Rongzhi Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Shuohang Wang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "yelong shen", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Liyuan Liu", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 36, "n_formula_1": 26, "n_ref_uni": 23, "n_ref": 48, "n_ref_all": 61, "n_fig": 3, "n_tab": 11, "L_tab": 2361, "n_element_tab": 196, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 19, "n_element_tab_1": 2, "formula_len_all": 2197, "formula_len_all_1": 1393, "len_all": 140300, "len_all_1": 63606, "len_abs": 1452, "len_title": 84, "len_sents": 45533, "len_sents_1": 29417, "n_sents": 364, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1460, "title": "LoRC: Low-Rank Compression for LLMs KV Cache with a Progressive Compression Strategy", "abs": "The Key-Value (KV) cache is a crucial component in serving transformer-based autoregressive large language models (LLMs), enabling faster inference by storing previously computed KV vectors. However, its memory consumption scales linearly with sequence length and batch size, posing a significant bottleneck in LLM deployment. Existing approaches to mitigate this issue include: (1) efficient attention variants integrated in upcycling stages, which requires extensive parameter tuning thus unsuitable to pre-trained LLMs; (2) KV cache compression at test time, primarily through token eviction policies, which often overlook inter-layer dependencies and can be task-specific.\n\nThis paper introduces an orthogonal approach to KV cache compression. We propose a low-rank approximation of KV weight matrices, allowing for plug-in integration with existing transformer-based LLMs without model retraining. To effectively compress KV cache at the weight level, we adjust for layerwise sensitivity and introduce a progressive compression strategy, which is supported by our theoretical analysis on how compression errors accumulate in deep networks. Our method is designed to function without model tuning in upcycling stages or task-specific profiling in test stages. Extensive experiments with LLaMA models ranging from 8B to 70B parameters across various tasks show that our approach significantly reduces the GPU memory footprint while maintaining performance.", "keywords": [ "KV Cache Compression", "Progressive Compression Strategy" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "NHxwxc3ql6", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Balaji Krishnamurthy", "gender": "Male", "institution": "Adobe Systems", "country": "IN", "position": "Principal Scientist" }, { "name": "Sohan Patnaik", "gender": "Male", "institution": "Adobe Systems", "country": "IN", "position": "Researcher" }, { "name": "Sumit Bhatia", "gender": "unknown", "institution": "Adobe Systems", "country": "IN", "position": "Senior ML Scientist" }, { "name": "Milan Aggarwal", "gender": "Male", "institution": "Adobe Systems", "country": "", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 8, "n_ref_uni": 46, "n_ref": 114, "n_ref_all": 146, "n_fig": 3, "n_tab": 5, "L_tab": 1901, "n_element_tab": 102, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 956, "formula_len_all_1": 756, "len_all": 252724, "len_all_1": 70486, "len_abs": 1807, "len_title": 117, "len_sents": 64405, "len_sents_1": 34122, "n_sents": 459, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 116, "L_abs": 1815, "title": "It Helps to Take a Second Opinion: Teaching Smaller LLMs To Deliberate Mutually via Selective Rationale Optimisation", "abs": "Very large language models (LLMs) such as GPT-4 have shown the ability to handle complex tasks by generating and self-refining step-by-step rationales. Smaller language models (SLMs), typically with < 13B parameters, have been improved by using the data generated from very-large LMs through knowledge distillation. However, various practical constraints such as API costs, copyright, legal and ethical policies restrict using large (often opaque) models to train smaller models for commercial use. Limited success has been achieved at improving the ability of an SLM to explore the space of possible rationales and evaluate them by itself through self-deliberation. To address this, we propose COALITION, a trainable framework that facilitates interaction between two variants of the same SLM and trains them to generate and refine rationales optimized for the end-task. The variants exhibit different behaviors to produce a set of diverse candidate rationales during the generation and refinement steps. The model is then trained via Selective Rationale Optimization (SRO) to prefer generating rationale candidates that maximize the likelihood of producing the ground-truth answer. During inference, COALITION employs a controller to select the suitable variant for generating and refining the rationales. On five different datasets covering mathematical problems, commonsense reasoning, and natural language inference, COALITION outperforms several baselines by up to 5%. Our ablation studies reveal that cross-communication between the two variants performs better than using the single model to self-refine the rationales. We also demonstrate the applicability of COALITION for LMs of varying scales (4B to 14B parameters) and model families (Mistral, Llama, Qwen, Phi). We release the code for this work here.", "keywords": [ "Multi-LLM Deliberation", "Smaller LLMs", "Rationale Generation", "Rationale Refinement", "Selective Rationale Optimisation", "Trainable", "Task-Guided Rationale Selection" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "NHhjczmJjo", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Cong Shen", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Associate Professor" }, { "name": "Jing Yang", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Associate Professor" }, { "name": "Renpu Liu", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Ruida Zhou", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" } ], "n_formula": 113, "n_formula_1": 18, "n_ref_uni": 41, "n_ref": 84, "n_ref_all": 95, "n_fig": 6, "n_tab": 3, "L_tab": 1609, "n_element_tab": 65, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 2645, "n_element_tab_1": 32, "formula_len_all": 16327, "formula_len_all_1": 1543, "len_all": 202328, "len_all_1": 75949, "len_abs": 1514, "len_title": 131, "len_sents": 57638, "len_sents_1": 32711, "n_sents": 509, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1521, "title": "On the Learn-to-Optimize Capabilities of Transformers in In-Context Sparse Recovery", "abs": "An intriguing property of the Transformer is its ability to perform in-context learning (ICL), where the Transformer can solve different inference tasks without parameter updating based on the contextual information provided by the corresponding input-output demonstration pairs. It has been theoretically proved that ICL is enabled by the capability of Transformers to perform gradient-descent algorithms (Von Oswald et al., 2023a; Bai et al., 2024). This work takes a step further and shows that Transformers can perform learning-to-optimize (L2O) algorithms. Specifically, for the ICL sparse recovery (formulated as LASSO) tasks, we show that a K-layer Transformer can perform an L2O algorithm with a provable convergence rate linear in K. This provides a new perspective explaining the superior ICL capability of Transformers, even with only a few layers, which cannot be achieved by the standard gradient-descent algorithms. Moreover, unlike the conventional L2O algorithms that require the measurement matrix involved in training to match that in testing, the trained Transformer is able to solve sparse recovery problems generated with different measurement matrices. Besides, Transformers as an L2O algorithm can leverage structural information embedded in the training tasks to accelerate its convergence during ICL, and generalize across different lengths of demonstration pairs, where conventional L2O algorithms typically struggle or fail. Such theoretical findings are supported by our experimental results.", "keywords": [ "Transformer", "In-context learning", "Learning-to-optimize" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 4, 3, 2, 4 ], "contribution_list": [ 4, 3, 2, 4 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "NHe6guO3l6", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Arun Verma", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "SG", "position": "Postdoc" }, { "name": "Bolin Ding", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Senior Director" }, { "name": "Bryan Kian Hsiang Low", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Associate Professor" }, { "name": "Ze Yu Zhang", "gender": "Not Specified", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Nhung Bui", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 38, "n_ref": 56, "n_ref_all": 74, "n_fig": 13, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 139, "formula_len_all_1": 139, "len_all": 119099, "len_all_1": 65301, "len_abs": 1205, "len_title": 126, "len_sents": 43442, "len_sents_1": 33959, "n_sents": 314, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1209, "title": "Achieving Exact Federated Unlearning with Improved Post-Unlearning Performance", "abs": "Federated learning is a machine learning paradigm that allows multiple clients to train aggregated model via sharing model updates to a central server without sharing their data. Even though the data is not shared, it can indirectly influence the aggregated model via the shared model updates. In many real-life scenarios, we need to completely remove a client's influence (unlearning) from the aggregated model, such as competitive clients who want to remove their influence from the aggregated model after leaving the coalition to ensure other clients do not benefit from their contributions. The influence removal is also needed when the adversarial client negatively affects the aggregated model. Though the aggregated model can be retrained from scratch to ensure exact unlearning (completely removing the client's influence from the aggregated model), it performs poorly just after the unlearning, which is undesirable during deployment. To overcome this challenge, this paper proposes federated unlearning algorithms that ensure exact unlearning while achieving better performance post-unlearning. Our experimental results on different real datasets validate the performance of the proposed algorithms.", "keywords": [ "Exact Federated Unlearning", "Improved Post-Unlearning Performance", "Multi-Models Training" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "NHPQOm6z0e", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Lei Jiao", "gender": "Male", "institution": "University of Agder", "country": "NO", "position": "Full Professor" }, { "name": "Ole-Christoffer Granmo", "gender": "Male", "institution": "University of Oslo, Norway", "country": "", "position": "PhD student" }, { "name": "Xuan Zhang", "gender": "Female", "institution": "NORCE Norwegian Research Centre", "country": "NO", "position": "Researcher" } ], "n_formula": 120, "n_formula_1": 19, "n_ref_uni": 24, "n_ref": 52, "n_ref_all": 146, "n_fig": 15, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10267, "formula_len_all_1": 1969, "len_all": 241972, "len_all_1": 71058, "len_abs": 1223, "len_title": 119, "len_sents": 94186, "len_sents_1": 33411, "n_sents": 841, "n_sents_1": 270, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1230, "title": "On the Convergence of Tsetlin Machines for the AND and the OR Operators", "abs": "The Tsetlin Machine (TM) is an innovative machine learning algorithm rooted in propositional logic, achieving state-of-the-art performance in various pattern recognition tasks. While previous studies analyzed its convergence properties for the 1-bit and XOR operators, this work extends the analysis to the AND and OR operators, completing the study of fundamental digital operations. Our findings demonstrate that the TM almost surely converges to reproduce the AND and OR operators when trained on noise-free data over an infinite time horizon. Notably, the analysis of the OR operator uncovers a distinct property: the ability of the TM to represent two sub-patterns jointly within a single clause, contrasting with its behavior in the XOR case. Furthermore, we investigate the TM\u2019s behavior for AND/OR/XOR operators with noisy training samples, including mislabeled samples and irrelevant inputs. With wrong labels, the TM does not converge to the intended operators but can still learn efficiently. With irrelevant variables, the TM converges to the intended operators almost surely. Together, these analyses provide a comprehensive theoretical foundation for the TM's convergence properties across basic Boolean operators.", "keywords": [ "Tsetlin Machine", "Convergence", "AND operator", "OR operator" ], "rating_list": [ 3, 3, 8, 8 ], "soundness_list": [ 2, 2, 4, 4 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 1, 1, 4, 4 ], "confidence_list": [ 3, 2, 3, 2 ] }, { "paper_id": "NHMuM84tRT", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Diego Martinez Plasencia", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Associate Professor" }, { "name": "Dimitrios Kanoulas", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Full Professor" }, { "name": "Elia Gatti", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "Lecturer" }, { "name": "Jincheng Wang", "gender": "unknown", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Pengyuan Wei", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Penny Karanasou", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 2, "n_ref_uni": 42, "n_ref": 89, "n_ref_all": 138, "n_fig": 12, "n_tab": 8, "L_tab": 1494, "n_element_tab": 139, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 963, "n_element_tab_1": 155, "formula_len_all": 305, "formula_len_all_1": 93, "len_all": 172378, "len_all_1": 60146, "len_abs": 2862, "len_title": 152, "len_sents": 61596, "len_sents_1": 28455, "n_sents": 447, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 103, "L_abs": 1419, "title": "Long-Short Decision Transformer: Bridging Global and Local Dependencies for Generalized Decision-Making", "abs": "Decision Transformers (DTs) effectively capture long-range dependencies using self-attention but struggle with fine-grained local relationships, especially the Markovian properties in many offline-RL datasets. Conversely, Decision Convformer (DC) utilizes convolutional filters for capturing local patterns but shows limitations in tasks demanding long-term dependencies, such as Maze2d. To address these limitations and leverage both strengths, we propose the Long-Short Decision Transformer (LSDT), a general-purpose architecture to effectively capture global and local dependencies across two specialized parallel branches (self-attention and convolution). We explore how these branches complement each other by modeling various ranged dependencies across different environments, and compare it against other baselines. Experimental results demonstrate our LSDT achieves state-of-the-art performance and notable gains over the standard DT in D4RL offline RL benchmark. Leveraging the parallel architecture, LSDT performs consistently on diverse datasets, including Markovian and non-Markovian. We also demonstrate the flexibility of LSDT's architecture, where its specialized branches can be replaced or integrated into models like DC to improve their performance in capturing diverse dependencies. Finally, we also highlight the role of goal states in improving decision-making for goal-reaching tasks like Antmaze.", "keywords": [ "Deep Learning", "Reinforcement Learning", "Transformer", "Decision Transformer", "Long-Short Decsion Transformer", "OfflineRL" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "NHCkILEmWn", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Da Chen", "gender": "Male", "institution": "University of Bath", "country": "GB", "position": "Assistant Professor" }, { "name": "JUN ZHOU", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Xiaofeng Mao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xin Zhang", "gender": "Male", "institution": "Ant Group", "country": "", "position": "Researcher" }, { "name": "Yuhan Dong", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Zitao Xuan", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 40, "n_ref": 82, "n_ref_all": 132, "n_fig": 4, "n_tab": 4, "L_tab": 1402, "n_element_tab": 202, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1970, "formula_len_all_1": 1470, "len_all": 143047, "len_all_1": 62272, "len_abs": 1637, "len_title": 109, "len_sents": 44501, "len_sents_1": 29902, "n_sents": 299, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1644, "title": "ShieldHead: Decoding-time Safeguard for Large Language Models", "abs": "In light of the widespread deployment of Large Language Models (LLMs), the responsibility for safeguarding and regulating LLM-generated content has taken on heightened significance. Recent advancements in LLM-based moderation methods, e.g., LlamaGuard, have demonstrated remarkable promise in identifying safety risks associated with both inputs and outputs in human-AI interactions. However, integrating LLM-based safeguards into a chatbot system requires an additional inference stage involving a moderation LLM with billions of parameters, which significantly increases computational costs and reduces overall efficiency. In this paper, we demonstrate that simply learning a classification head on the last-layer hidden states of the dialogue model provides a strong capability to identify harmful contents. The classification head, referred to as ShieldHead, serves as an auxiliary branch paralleled with next-token-prediction LM head, enabling the detection of potential risks in past text sequences. Additionally, a label disambiguation technique is employed to supervise ShieldHead with both token-level and sentence-level labels, which further enhances its performance. ShieldHead exhibits remarkable efficiency during inference, providing real-time moderation results alongside token-wise streaming output during the chatbot system's decoding phase. Extensive experimental results demonstrate the superiority of the proposed framework: a state-of-the-art performance on the XSTest and SafeRLHF datasets while running at a speed about 300\u00d7 faster (<1ms) than previous LLM-based moderation models with \uff5e99% less parameters of LlamaGuard.", "keywords": [ "Large Language Model", "Safety Guard", "Content Moderation" ], "rating_list": [ 5, 5, 8, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "NH47cNdgNz", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chaochao Lu", "gender": "unknown", "institution": "Shanghai AI Laboratory ", "country": "CN", "position": "Research Scientist" }, { "name": "Shengjie Zhao", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Full Professor" }, { "name": "Shu Yu", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Sirui Chen", "gender": "unknown", "institution": "Tongji University", "country": "", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 56, "n_ref": 113, "n_ref_all": 156, "n_fig": 12, "n_tab": 5, "L_tab": 3877, "n_element_tab": 197, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 470, "formula_len_all_1": 527, "len_all": 229275, "len_all_1": 75325, "len_abs": 8902, "len_title": 126, "len_sents": 65743, "len_sents_1": 33267, "n_sents": 709, "n_sents_1": 305, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1291, "title": "From Imitation to Introspection: Probing Self-Consciousness in Language Models", "abs": "Self-consciousness, the introspection of one's existence and thoughts, represents a high-level cognitive process. As language models advance at an unprecedented pace, a critical question arises: Are these models becoming self-conscious?\nDrawing upon insights from psychological and neural science, this work presents a practical definition of self-consciousness for language models and refines ten core concepts. Our work pioneers an investigation into self-consciousness in language models by, for the first time, leveraging causal structural games to establish the functional definitions of the ten core concepts. \nBased on our definitions, we conduct a comprehensive four-stage experiment: quantification (evaluation of ten leading models), representation (visualization of self-consciousness within the models), manipulation (modification of the models' representation), and acquisition (fine-tuning the models on core concepts). \nOur findings indicate that although models are in the early stages of developing self-consciousness, there is a discernible representation of certain concepts within their internal mechanisms. However, these representations of self-consciousness are hard to manipulate positively at the current stage, yet they can be acquired through targeted fine-tuning.", "keywords": [ "self-consciousness", "evaluation", "probing", "large language model", "causality" ], "rating_list": [ 6, 6, 3, 8 ], "soundness_list": [ 3, 3, 1, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 1, 4 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "NGKQoaqLpo", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Andrey Zhmoginov", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Been Kim", "gender": "unknown", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Chen Sun", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Mark Sandler", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Max Vladymyrov", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Nolan Andrew Miller", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "MS student" }, { "name": "Renat Aksitov", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Ulrich Rueckert", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 35, "n_ref": 106, "n_ref_all": 167, "n_fig": 38, "n_tab": 3, "L_tab": 387, "n_element_tab": 15, "n_fig_1": 31, "n_tab_1": 1, "L_tab_1": 36, "n_element_tab_1": 1, "formula_len_all": 35, "formula_len_all_1": 35, "len_all": 192728, "len_all_1": 59238, "len_abs": 1504, "len_title": 105, "len_sents": 56956, "len_sents_1": 30050, "n_sents": 393, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 56, "L_abs": 1517, "title": "How new data pollutes LLM knowledge and how to dilute it", "abs": "Understanding how the learning of new texts alter the existing knowledge in a large language model is of great importance, because it is through these accumulated changes that the LLM was initially pre-trained, and is also through such changes that continual, new learning in LLMs can proceed. As a result, both desirable alterations (i.e. generalization) and undesirable alterations (i.e. hallucination) can occur. Here, we study the learning of new texts, one at a time, and ask: how does it impact the underlying LLM knowledge? \n We show that learning new texts induce 'priming', an undesirable effect that pollutes existing knowledge where it should not.\nCentrally, we demonstrate that we can predict how much priming will happen after learning, using token probability before learning. This was empirically robust across models (PALM-2-xs/s, Gemma-2b, Llama-2-7b), of various sizes, and training stages. To show this, we created a new dataset, called \"Outlandish\" consisting of 1320 different samples with diverse textual characteristics. Finally, we propose two strategies to mitigate the spread of priming: first, a simple text augmentation technique which we call the \"stepping-stone'', and second, a novel update pruning technique (\"ignore-k\"). These decrease priming by a median of 50%-75% and 50%-95% respectively depending on the model architecture, and enhance the specificity of new learning in language models. The dataset and reproducible findings can be found [LINK omitted for double blind review].", "keywords": [ "fine-tuning", "hallucinations", "knowledge injection", "memory", "LLMs" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 4 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "NGF1wDDBMm", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Albert No", "gender": "unknown", "institution": "Yonsei University", "country": "", "position": "Associate Professor" }, { "name": "Dongjae Jeon", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Undergrad student" }, { "name": "Jonghyun Choi", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" }, { "name": "Wonje Jeung", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Undergrad student" }, { "name": "TAEHEON KIM", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" } ], "n_formula": 19, "n_formula_1": 5, "n_ref_uni": 44, "n_ref": 94, "n_ref_all": 158, "n_fig": 21, "n_tab": 18, "L_tab": 23876, "n_element_tab": 713, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 3949, "n_element_tab_1": 219, "formula_len_all": 1843, "formula_len_all_1": 391, "len_all": 248275, "len_all_1": 75781, "len_abs": 3857, "len_title": 112, "len_sents": 72774, "len_sents_1": 33322, "n_sents": 494, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1125, "title": "AN INFORMATION THEORETIC EVALUATION METRIC FOR STRONG UNLEARNING", "abs": "Machine unlearning (MU) aims to remove the influence of specific data from trained models, addressing privacy concerns and ensuring compliance with regulations such as the \"right to be forgotten.\"\nEvaluating strong unlearning, where the unlearned model is indistinguishable from one retrained without the forgetting data, remains a significant challenge in deep neural networks (DNNs).\nCommon black-box metrics, such as variants of membership inference attacks and accuracy comparisons, primarily assess model outputs but often fail to capture residual information in intermediate layers.\nTo bridge this gap, we introduce the Information Difference Index (IDI), a novel white-box metric inspired by information theory.\nIDI quantifies retained information in intermediate features by measuring mutual information between those features and the labels to be forgotten, offering a more comprehensive assessment of unlearning efficacy.\nOur experiments demonstrate that IDI effectively measures the degree of unlearning across various datasets and architectures,\nproviding a reliable tool for evaluating strong unlearning in DNNs.", "keywords": [ "Strong unlearning", "information-theoretic metrics", "Information Difference Index", "residual information", "evaluation metric" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "NGB6YNnO5o", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Florian Shkurti", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Assistant Professor" }, { "name": "Jierui Zhu", "gender": "unknown", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Qi CHEN", "gender": "Female", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "Postdoc" } ], "n_formula": 122, "n_formula_1": 38, "n_ref_uni": 53, "n_ref": 92, "n_ref_all": 127, "n_fig": 11, "n_tab": 3, "L_tab": 622, "n_element_tab": 130, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15661, "formula_len_all_1": 3227, "len_all": 220091, "len_all_1": 66146, "len_abs": 1225, "len_title": 132, "len_sents": 61324, "len_sents_1": 28211, "n_sents": 549, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1019, "title": "Generalization in VAE and Diffusion Models: A Unified Information-Theoretic Analysis", "abs": "Despite the empirical success of Diffusion Models (DMs) and Variational Autoencoders (VAEs), their generalization performance remains theoretically underexplored, particularly lacking a full consideration of the shared encoder-generator structure. Leveraging recent information-theoretic tools, we propose a unified theoretical framework that guarantees the generalization of both the encoder and generator by treating them as randomized mappings. This framework further enables (1) a refined analysis for VAEs, accounting for the generator's generalization, which was previously overlooked; (2) illustrating an explicit trade-off in generalization terms for DMs that depends on the diffusion time $T$; and (3) providing estimable bounds for DMs based solely on the training data, allowing the selection of the optimal $T$ and the integration of such bounds into the optimization process to improve model performance. Empirical results on both synthetic and real datasets illustrate the validity of the proposed theory.", "keywords": [ "Generalization bounds; information theory; generative models; VAE; diffusion models;" ], "rating_list": [ 6, 6, 8, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "NFWt2PavSW", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Aurelien Lucchi", "gender": "Male", "institution": "University of Basel", "country": "CH", "position": "Assistant Professor" }, { "name": "Eduard Gorbunov", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "Rustem Islamov", "gender": "Male", "institution": "University of Basel", "country": "CH", "position": "PhD student" }, { "name": "Samuel Horv\u00e1th", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Assistant Professor" }, { "name": "Peter Richtarik", "gender": "Male", "institution": "King Abdullah University of Science and Technology (KAUST)", "country": "SA", "position": "Full Professor" } ], "n_formula": 319, "n_formula_1": 31, "n_ref_uni": 63, "n_ref": 120, "n_ref_all": 149, "n_fig": 7, "n_tab": 3, "L_tab": 540, "n_element_tab": 83, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 853, "n_element_tab_1": 29, "formula_len_all": 51288, "formula_len_all_1": 1781, "len_all": 312816, "len_all_1": 69077, "len_abs": 1625, "len_title": 129, "len_sents": 61263, "len_sents_1": 28559, "n_sents": 845, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 993, "title": "Momentum and Error Feedback for Clipping with Fast Rates and Differential Privacy", "abs": "Strong Differential Privacy (DP) and Optimization guarantees are two desirable properties for a method in Federated Learning (FL). However, existing algorithms do not achieve both properties at once: they either have optimal DP guarantees but rely on restrictive assumptions such as bounded gradients/bounded data heterogeneity, or they have strong optimization guarantees but do not have DP ones. To address this gap in the literature, we propose and analyze a new method called Clip21-SGDM based on a novel combination of clipping, heavy-ball momentum, and Error Feedback. In particular, for non-convex smooth distributed problems with clients having arbitrarily heterogeneous data, we prove that Clip21-SGDM has optimal convergence rate and also optimal (local-)DP neighborhood. Our numerical experiments on non-convex logistic regression and training of neural networks highlight the superiority of Clip21-SGDM over baselines in terms of the optimization performance for a given DP-budget.", "keywords": [ "gradient clipping", "federated learning", "momentum", "differential privacy" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 2, 4, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "NFEnBqknoX", "primary_area": "generative models", "n_author": 17, "author_info_list": [ { "name": "Akash Srivastava", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Bo Liu", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Researcher" }, { "name": "Chaowei Tan", "gender": "Male", "institution": "FocusAI Inc", "country": "US", "position": "Seniro Applied Scientist" }, { "name": "Di Liu", "gender": "Male", "institution": "Rutgers University, New Brunswick", "country": "US", "position": "PhD student" }, { "name": "Faez Ahmed", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Hongdong Li", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "Full Professor" }, { "name": "Junzhou Huang", "gender": "Male", "institution": "University of Texas, Arlington", "country": "US", "position": "Full Professor" }, { "name": "Kang Li", "gender": "Male", "institution": "Rutgers University", "country": "", "position": "Associate Professor" }, { "name": "Ligong Han", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Martin Renqiang Min", "gender": "Male", "institution": "NEC Laboratories America", "country": "US", "position": "Researcher" }, { "name": "Minhao Bai", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Quan Dao", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Song Wen", "gender": "unknown", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Xiaoxiao He", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Dimitris Metaxas", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Full Professor" }, { "name": "Felix Juefei-Xu", "gender": "Male", "institution": "GenAI, Meta", "country": "US", "position": "Researcher" }, { "name": "Han Zhang", "gender": "Male", "institution": "Reve AI", "country": "US", "position": "Researcher" } ], "n_formula": 28, "n_formula_1": 19, "n_ref_uni": 44, "n_ref": 93, "n_ref_all": 121, "n_fig": 12, "n_tab": 5, "L_tab": 1177, "n_element_tab": 145, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 718, "n_element_tab_1": 78, "formula_len_all": 2058, "formula_len_all_1": 1232, "len_all": 156716, "len_all_1": 62182, "len_abs": 1059, "len_title": 124, "len_sents": 42526, "len_sents_1": 27523, "n_sents": 367, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 980, "title": "Discrete Inversion: A Controllable Latent Space for Multinomial Diffusion and Masked Generative Models", "abs": "Discrete diffusion models have achieved notable success in tasks like image generation and masked language modeling, yet they face limitations in controlled content editing. This paper introduces {\\bf Discrete Inversion}, the first approach to enable precise inversion for discrete diffusion models, including multinomial diffusion and masked generative models. By recording noise sequences and masking patterns during the forward diffusion process, Discrete Inversion facilitates accurate reconstruction and controlled edits without the need for predefined masks or attention map manipulation. We demonstrate the effectiveness of our method across both image and text domains, evaluating it on models like VQ-Diffusion, Paella, and RoBERTa. Our results show that Discrete Inversion not only preserves high fidelity in the original data but also enables flexible and user-friendly editing in discrete spaces, significantly advancing the capabilities of discrete generative models.", "keywords": [ "Masked Generative Modeling; Discrete Diffusion Model; Multinomial Diffusion" ], "rating_list": [ 8, 6, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 4, 2 ], "contribution_list": [ 3, 3, 1 ], "confidence_list": [ 5, 4, 5 ] }, { "paper_id": "NEu8wgPctU", "primary_area": "applications to robotics, autonomy, planning", "n_author": 8, "author_info_list": [ { "name": "Abhirup Mallik", "gender": "Male", "institution": "Google", "country": "US", "position": "Intern" }, { "name": "Burhaneddin Yaman", "gender": "unknown", "institution": "Bosch Center for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Chenbin Pan", "gender": "Female", "institution": "Bosch", "country": "US", "position": "Researcher" }, { "name": "Hang Wang", "gender": "Male", "institution": "University of California, Davis", "country": "", "position": "PhD student" }, { "name": "Junshan Zhang", "gender": "Male", "institution": "University of California, Davis", "country": "", "position": "Full Professor" }, { "name": "Liu Ren", "gender": "Male", "institution": "Bosch Research", "country": "US", "position": "Principal Researcher" }, { "name": "Xin Ye", "gender": "Female", "institution": "Bosch", "country": "US", "position": "Researcher" }, { "name": "Feng Tao", "gender": "Male", "institution": "Bosch", "country": "US", "position": "Researcher" } ], "n_formula": 31, "n_formula_1": 13, "n_ref_uni": 27, "n_ref": 47, "n_ref_all": 80, "n_fig": 16, "n_tab": 8, "L_tab": 1865, "n_element_tab": 333, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 2969, "n_element_tab_1": 174, "formula_len_all": 2407, "formula_len_all_1": 828, "len_all": 161501, "len_all_1": 69999, "len_abs": 1465, "len_title": 114, "len_sents": 52456, "len_sents_1": 32161, "n_sents": 385, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1476, "title": "AdaWM: Adaptive World Model based Planning for Autonomous Driving", "abs": "World model based reinforcement learning (RL) has emerged as a promising approach for autonomous driving, which learns a latent dynamics model and uses it to train a planning policy. To speed up the learning process, the pretrain-finetune paradigm is often used, where online RL is initialized by a pretrained model and a policy learned offline. However, naively performing such initialization in RL may result in dramatic performance degradation during the online interactions in the new task. To tackle this challenge, we first analyze the performance degradation and identify two primary root causes therein: the mismatch of the planning policy and the mismatch of the dynamics model, due to distribution shift. We further analyze the effects of these factors on performance degradation during finetuning, and our findings reveal that the choice of finetuning strategies plays a pivotal role in mitigating these effects. We then introduce AdaWM, an Adaptive World Model based planning method, featuring two key steps: (a) mismatch identification, which quantifies the mismatches and informs the finetuning strategy, and (b) alignment-driven finetuning, which selectively updates either the policy or the model as needed using efficient low-rank updates. Extensive experiments on the challenging CARLA driving tasks demonstrate that AdaWM significantly improves the finetuning process, resulting in more robust and efficient performance in autonomous driving systems.", "keywords": [ "World Model", "Autonomous Driving", "Reinforcement Learning" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "NDMLjEJoLb", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Dinesh Manocha", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Professor" }, { "name": "Divya Kothandaraman", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Ming Lin", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "Capital One Endowed Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 32, "n_ref": 52, "n_ref_all": 68, "n_fig": 7, "n_tab": 2, "L_tab": 933, "n_element_tab": 11, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 498, "formula_len_all_1": 442, "len_all": 137440, "len_all_1": 59472, "len_abs": 1535, "len_title": 133, "len_sents": 47154, "len_sents_1": 28796, "n_sents": 354, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1711, "title": "Blending Concepts in Text-to-Image Diffusion Models using the Black Scholes Algorithm", "abs": "Many image generation tasks, such as content creation, editing, personalization, and zero-shot generation, require generating unseen concepts without retraining the model or collecting additional data. These tasks often involve blending existing concepts by conditioning the diffusion model with text prompts at each denoising step, a process known as ``prompt mixing''. We introduce a novel approach for prompt mixing to forecasts predictions w.r.t. the generated image and makes informed text conditioning decisions at each time step during diffusion denoising. To do so, we leverage the connection between diffusion models (rooted in non-equilibrium thermodynamics) and the Black-Scholes model for pricing options in Finance, and draw analogies between the variables in both contexts to derive an appropriate algorithm for prompt mixing using the Black Scholes model. Specifically, the parallels between diffusion models and the Black-Scholes model enable us to leverage properties related to the dynamics of the Markovian model derived in the Black-Scholes algorithm. Our prompt-mixing algorithm is data-efficient, meaning it does not need additional training. Furthermore, it operates without human intervention or hyperparameter tuning. We highlight the benefits of our approach by comparing it, qualitatively and quantitatively using CLIP scores, to other prompt mixing techniques, including linear interpolation, alternating prompts, step-wise prompt switching, and CLIP-guided prompt selection across various scenarios such as single object per text prompt, multiple objects per text prompt and objects against backgrounds. The resulting code will be made publicly available for research reproduction.", "keywords": [ "content creation", "diffusion models" ], "rating_list": [ 3, 8, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 3, 5, 3 ] }, { "paper_id": "NDLmZZWATc", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Jung Ho Park", "gender": "Male", "institution": "Samsung Electronics", "country": "KR", "position": "Researcher" }, { "name": "Kyubum Shin", "gender": "Male", "institution": "NAVER", "country": "KR", "position": "Researcher" }, { "name": "Sua Lee", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 47, "n_ref": 88, "n_ref_all": 102, "n_fig": 7, "n_tab": 3, "L_tab": 2018, "n_element_tab": 82, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 502, "n_element_tab_1": 30, "formula_len_all": 659, "formula_len_all_1": 659, "len_all": 130048, "len_all_1": 56909, "len_abs": 1325, "len_title": 135, "len_sents": 34995, "len_sents_1": 27920, "n_sents": 219, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1374, "title": "Weighted Multi-Prompt Learning with Description-free Large Language Model Distillation", "abs": "Recent advances in pre-trained Vision Language Models (VLMs) have shown promising potential through \\textit{prompt learning} in effectively adapting to downstream tasks without requiring additional annotated paired datasets.\nTo supplement text information in VLMs dependently trained on correlation with vision data, new approaches leveraging Large Language Models (LLM) in prompts have been proposed, enhancing robustness to unseen and diverse data.\nExisting methods query LLM for text-based responses (i.e., \\textit{descriptions}) to utilize in prompts, but this approach has limitations: high variability and low reliability.\nIn this work, we propose \\textbf{De}scription-free \\textbf{Mul}ti-prompt Learning(\\textbf{DeMul}) for image recognition task, a novel method that eliminates the process of extracting descriptions and instead directly distills knowledge from LLM into prompts.\nBy adopting a description-free approach, prompts can encapsulate richer semantics and still be defined as continuous vectors to optimize, thereby eliminating the need for discrete pre-defined templates.\nAdditionally, in a multi-prompt setting, we have empirically shown the potential of using prompt weighting to reflect the importance of different prompts during training.\nExperimental results demonstrate that our approach achieves superior performance across 11 recognition datasets.", "keywords": [ "Prompt learning", "Vision-language models", "Large language models", "Few-shot image recognition" ], "rating_list": [ 6, 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 4, 5, 3, 5, 5 ] }, { "paper_id": "NCynHu1kVF", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Bohyung Han", "gender": "Not Specified", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Donghun Ryou", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Inju Ha", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Seonguk Seo", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 14, "n_ref_uni": 35, "n_ref": 66, "n_ref_all": 87, "n_fig": 8, "n_tab": 6, "L_tab": 3512, "n_element_tab": 243, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 807, "n_element_tab_1": 34, "formula_len_all": 1125, "formula_len_all_1": 728, "len_all": 113220, "len_all_1": 52175, "len_abs": 1040, "len_title": 102, "len_sents": 33086, "len_sents_1": 24066, "n_sents": 233, "n_sents_1": 163, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 1042, "title": "Learning to Translate Noise for Robust Image Denoising", "abs": "Image denoising techniques based on deep learning often struggle with poor generalization performance to out-of-distribution real-world noise. To tackle this challenge, we propose a novel noise translation framework that performs denoising on an image with translated noise rather than directly denoising an original noisy image. Speci\ufb01cally, our approach translates complex, unknown real-world noise into Gaussian noise, which is spatially uncorrelated and independent of image content, through a noise translation network. The translated noisy images are then processed by an image denoising network pretrained to effectively remove Gaussian noise, enabling robust and consistent denoising performance. We also design well-motivated loss functions and architectures for the noise translation network by leveraging the mathematical properties of Gaussian noise. Experimental results demonstrate that the proposed method substantially improves robustness and generalizability, outperforming state-of-the-art methods across diverse benchmarks.", "keywords": [ "denoising", "generalization", "robustness", "noise translation" ], "rating_list": [ 6, 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 3, 4 ] }, { "paper_id": "NCrFA7dq8T", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Carsten Eickhoff", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "Full Professor" }, { "name": "Ellie Pavlick", "gender": "Female", "institution": "Brown University", "country": "", "position": "Associate Professor" }, { "name": "Matianyu Zang", "gender": "Female", "institution": "Brown University", "country": "US", "position": "Undergrad student" }, { "name": "Qinan Yu", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Ruochen Zhang", "gender": "Female", "institution": "Brown University", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 60, "n_ref": 97, "n_ref_all": 135, "n_fig": 12, "n_tab": 1, "L_tab": 353, "n_element_tab": 66, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 158367, "len_all_1": 64187, "len_abs": 1352, "len_title": 147, "len_sents": 41233, "len_sents_1": 30991, "n_sents": 307, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1356, "title": "The Same but Different: Structural Similarities and Differences in Multilingual Language Modeling", "abs": "We employ new tools from mechanistic interpretability in order to ask whether the internal structure of large language models (LLMs) shows correspondence to the linguistic structures which underlie the languages on which they are trained. In particular, we ask (1) when two languages employ the same morphosyntactic processes, do LLMs handle them using shared internal circuitry? and (2) when two languages require different morphosyntactic processes, do LLMs handle them using different internal circuitry? In a focused case study on English and Chinese multilingual and monolingual models, we analyze the internal circuitry involved in two tasks. We find evidence that models employ the same circuit to handle the same syntactic process independently of the language in which it occurs, and that this is the case even for monolingual models trained completely independently. Moreover, we show that multilingual models employ language-specific components (attention heads and feed-forward networks) when needed to handle linguistic processes (e.g., morphological marking) that only exist in some languages. Together, our results are revealing about how LLMs trade off between exploiting common structures and preserving linguistic differences when tasked with modeling multiple languages simultaneously, opening the door for future work in this direction.", "keywords": [ "multilinguality", "interpretability" ], "rating_list": [ 5, 6, 6, 8, 8 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4, 4 ], "contribution_list": [ 2, 3, 3, 4, 3 ], "confidence_list": [ 4, 4, 2, 4, 4 ] }, { "paper_id": "NCgTbt2j1F", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Nitya Sathyavageeswaran", "gender": "Female", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Yansong Li", "gender": "Male", "institution": "University of Illinois Chicago", "country": "US", "position": "PhD student" }, { "name": "Yu Wu", "gender": "unknown", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Zeyu Dong", "gender": "Not Specified", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Anand D. Sarwate", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Associate Professor" } ], "n_formula": 53, "n_formula_1": 15, "n_ref_uni": 43, "n_ref": 69, "n_ref_all": 85, "n_fig": 6, "n_tab": 8, "L_tab": 1039, "n_element_tab": 119, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 41, "n_element_tab_1": 3, "formula_len_all": 5235, "formula_len_all_1": 832, "len_all": 217684, "len_all_1": 63528, "len_abs": 1423, "len_title": 88, "len_sents": 56773, "len_sents_1": 30376, "n_sents": 457, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1431, "title": "Learning to Help in Multi-Class Settings", "abs": "Deploying complex machine learning models on resource-constrained devices is challenging due to limited computational power, memory, and model retrainability. To address these limitations, a hybrid system can be established by augmenting the local model with a server-side model, where samples are selectively deferred by a *rejector* and then sent to the server for processing. The hybrid system enables efficient use of computational resources while minimizing the overhead associated with server usage. The recently proposed Learning to Help (L2H) model proposed training a server model given a fixed local (client) model. This differs from the Learning to Defer (L2D) framework which trains the client for a fixed (expert) server. In both L2D and L2H, the training includes learning a rejector at the client to determine when to query the server. In this work, we extend the L2H model from binary to multi-class classification problems and demonstrate its applicability in a number of different scenarios of practical interest in which access to the server may be limited by cost, availability, or policy. We derive a stage-switching surrogate loss function that is differentiable, convex, and consistent with the Bayes rule corresponding to the 0-1 loss for the L2H model. Experiments show that our proposed methods offer an efficient and practical solution for multi-class classification in resource-constrained environments.", "keywords": [ "learning with abstention", "learning with reject option", "learning to defer", "classification" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 2, 3, 2 ] }, { "paper_id": "NCUKdeqz4X", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Defu Cao", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Jie Cai", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Lumingyuan Tang", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "MS student" }, { "name": "Wei Yang", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Wen Ye", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Yan Liu", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "Professor" }, { "name": "Yizhou Zhang", "gender": "unknown", "institution": "University of Southern California", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 57, "n_ref": 78, "n_ref_all": 93, "n_fig": 7, "n_tab": 15, "L_tab": 3902, "n_element_tab": 346, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 673, "n_element_tab_1": 52, "formula_len_all": 259, "formula_len_all_1": 210, "len_all": 204951, "len_all_1": 67096, "len_abs": 1519, "len_title": 134, "len_sents": 67858, "len_sents_1": 33462, "n_sents": 472, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1527, "title": "Beyond Forecasting: Compositional Time Series Reasoning for End-to-End Task Execution", "abs": "In recent decades, there have been substantial advances in time series models and benchmarks across various individual tasks, such as time series forecasting, classification, and anomaly detection.\nMeanwhile, compositional reasoning in time series prevalent in real-world applications (e.g., decision-making and compositional question answering) is in great demand. Unlike simple tasks that primarily focus on predictive accuracy, compositional reasoning emphasizes the synthesis of diverse information from both time series data and various domain knowledge, making it distinct and extremely more challenging. In this paper, we introduce Compositional Time Series Reasoning, a new task of handling intricate multistep reasoning tasks from time series data. Specifically, this new task focuses on various question instances requiring structural and compositional reasoning abilities on time series data, such as decision-making and compositional question answering.\nAs an initial attempt to tackle this novel task, we developed \\modelname, a program-aided approach that utilizes large language model (LLM) to decompose a complex task into steps of programs that leverage existing time series models and numerical subroutines. Through a comprehensive set of experiments, we demonstrate that our simple but effective \\modelnamespace outperforms existing standalone reasoning approaches. These promising results indicate potential opportunities in the new task of time series reasoning and highlight the need for further research.", "keywords": [ "Time Series", "LLM", "Reasoning" ], "rating_list": [ 6, 1, 5, 6, 3 ], "soundness_list": [ 3, 1, 1, 3, 2 ], "presentation_list": [ 2, 1, 3, 4, 4 ], "contribution_list": [ 4, 1, 2, 3, 2 ], "confidence_list": [ 3, 5, 4, 2, 5 ] }, { "paper_id": "NBgB5xirgp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Jiaqi Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Research Scientist" }, { "name": "Pan Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xiaoyi Dong", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yuhang CAO", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 88, "n_ref_all": 107, "n_fig": 6, "n_tab": 5, "L_tab": 1235, "n_element_tab": 167, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 588, "n_element_tab_1": 42, "formula_len_all": 401, "formula_len_all_1": 401, "len_all": 143496, "len_all_1": 56339, "len_abs": 1434, "len_title": 136, "len_sents": 38770, "len_sents_1": 27559, "n_sents": 281, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1188, "title": "DualFocus: Integrating Macro and Micro Perspectives in Multi-modal Large Language Models", "abs": "We present DualFocus, a novel framework for integrating macro and micro perspectives within multi-modal large language models (MLLMs) to enhance vision-language task performance. Current MLLMs typically singularly focus on inputs at a predefined resolution, resulting in deficiencies in detailed questions involving local regions. We introduced a DualFocus mechanism where the model concentrates on the image from a macro perspective, responses to the question, and identifies suitable sub-regions to zoom in for subsequent micro perspective analysis. Via the integration of answers from both macro and micro perspectives, the model is adept at addressing tasks that encompass global, detailed, and combined considerations. To endows the DualFocus mechanism in MLLMs, we curated a tailored dataset derived from the Visual Genome (VG) and adapted it to align with the training regimen of DualFocus. Through comparative studies across different model sizes and benchmarks, we demonstrate DualFocus's superiority in balancing detailed examination with holistic insight, significantly reducing hallucination instances in MLLMs and improving their performance in various vision-language tasks.", "keywords": [ "Multi-modal Large Language Model (MLLM)", "Large Vision Language Model (LVLM)" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "NB8qn8iIW9", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Alasdair Paren", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "David Krueger", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Assistant Professor" }, { "name": "Fazl Barez", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Luke Marks", "gender": "Male", "institution": "Apart Lab", "country": "GB", "position": "Research Fellow" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 49, "n_ref_all": 67, "n_fig": 9, "n_tab": 3, "L_tab": 140, "n_element_tab": 24, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 140, "n_element_tab_1": 24, "formula_len_all": 380, "formula_len_all_1": 331, "len_all": 102310, "len_all_1": 53685, "len_abs": 1195, "len_title": 82, "len_sents": 28549, "len_sents_1": 25087, "n_sents": 197, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1201, "title": "Enhancing Neural Network Interpretability with Feature-Aligned Sparse Autoencoders", "abs": "Sparse Autoencoders (SAEs) have shown promise in improving the interpretability of neural network activations, but can learn features that are not features of the input, limiting their effectiveness. We propose Mutual Feature Regularization (MFR), a regularization technique for improving feature learning by encouraging SAEs trained in parallel to learn similar features. We motivate MFR by showing that features learned by multiple SAEs are more likely to correlate with features of the input. By training on synthetic data with known features of the input, we show that MFR can help SAEs learn those features, as we can directly compare the features learned by the SAE with the input features for the synthetic data. We then scale MFR to SAEs that are trained to denoise electroencephalography (EEG) data and SAEs that are trained to reconstruct GPT-2 Small activations. We show that MFR can improve the reconstruction loss of SAEs by up to 21.21\\% on GPT-2 Small, and 6.67\\% on EEG data. Our results suggest that the similarity between features learned by different SAEs can be leveraged to improve SAE training, thereby enhancing performance and the usefulness of SAEs for model interpretability.", "keywords": [ "Interpretability", "Sparse autoencoder", "Mutual learning" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 4, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "NAbqM2cMjD", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Donghyun Lee", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "MS student" }, { "name": "Mo Tiwari", "gender": "unknown", "institution": "OpenAI", "country": "US", "position": "Member of Technical Staff" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 32, "n_ref": 62, "n_ref_all": 79, "n_fig": 10, "n_tab": 1, "L_tab": 443, "n_element_tab": 11, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 600, "n_element_tab_1": 16, "formula_len_all": 27, "formula_len_all_1": 27, "len_all": 111527, "len_all_1": 57286, "len_abs": 1295, "len_title": 121, "len_sents": 30520, "len_sents_1": 27452, "n_sents": 224, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1297, "title": "Prompt Infection: LLM-to-LLM Prompt Injection within Multi-Agent Systems", "abs": "As Large Language Models (LLMs) grow increasingly powerful, multi-agent systems\u2014where multiple LLMs collaborate to tackle complex tasks\u2014are becoming more prevalent in modern AI applications. Most safety research, however, has focused on vulnerabilities in single-agent LLMs. These include prompt injection attacks, where malicious prompts embedded in external content trick the LLM into executing unintended or harmful actions, compromising the victim\u2019s application. In this paper, we reveal a more dangerous vector: LLM-to-LLM prompt injection within multi-agent systems. We introduce Prompt Infection, a novel attack where malicious prompts self-replicate across interconnected agents, behaving much like a computer virus. This attack poses severe threats, including data theft, scams, misinformation, and system-wide disruption, all while propagating silently through the system. Our extensive experiments demonstrate that multi-agent systems are highly susceptible, even when agents do not directly share communications. To address this, we propose LLM Tagging, a defense mechanism that, when combined with existing safeguards, significantly mitigates infection spread. This work underscores the urgent need for advanced security measures as multi-agent LLM systems become more widely adopted.", "keywords": [ "Large Language Models", "Prompt Injection", "Multi-Agent Systems", "LLM Security", "LLM Safety" ], "rating_list": [ 5, 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3, 2 ], "presentation_list": [ 1, 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 5, 3, 3, 4, 3 ] }, { "paper_id": "NALkteEo9Q", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Alina Oprea", "gender": "Female", "institution": "Northeastern University", "country": "US", "position": "Full Professor" }, { "name": "Christopher Amato", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "Associate Professor" }, { "name": "Ethan Rathbun", "gender": "unknown", "institution": "Northeastern University", "country": "US", "position": "PhD student" } ], "n_formula": 67, "n_formula_1": 18, "n_ref_uni": 29, "n_ref": 60, "n_ref_all": 87, "n_fig": 6, "n_tab": 3, "L_tab": 353, "n_element_tab": 40, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 215, "n_element_tab_1": 7, "formula_len_all": 8144, "formula_len_all_1": 1979, "len_all": 156997, "len_all_1": 69436, "len_abs": 1490, "len_title": 131, "len_sents": 53508, "len_sents_1": 31631, "n_sents": 405, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1093, "title": "Adversarial Inception for Bounded Backdoor Poisoning in Deep Reinforcement Learning", "abs": "Recent works have demonstrated the vulnerability of Deep Reinforcement Learning (DRL) algorithms against training-time, backdoor poisoning attacks. These attacks induce pre-determined, adversarial behavior in the agent upon observing a fixed trigger during deployment while allowing the agent to solve its intended task during training. Prior attacks rely on arbitrarily large perturbations to the agent's rewards to achieve both of these objectives - leaving them open to detection. Thus, in this work, we propose a new class of backdoor attacks against DRL which achieve state of the art performance while minimally altering the agent's rewards. These ``inception'' attacks train the agent to associate the targeted adversarial behavior with high returns by inducing a disjunction between the agent's chosen action and the true action executed in the environment during training. We formally define these attacks and prove they can achieve both adversarial objectives. We then devise an online inception attack which significantly out-performs prior attacks under bounded reward constraints.", "keywords": [ "Reinforcement Learning", "Poisoning Attacks", "Backdoor Attacks", "Adversarial Machine Learning" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "NA2vUMaMOm", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Mengshi Zhao", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shaofeng H.-C. Jiang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tianyi Wu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Hubert Chan", "gender": "unknown", "institution": "The University of Hong Kong", "country": "", "position": "Associate Professor" } ], "n_formula": 56, "n_formula_1": 14, "n_ref_uni": 18, "n_ref": 56, "n_ref_all": 66, "n_fig": 4, "n_tab": 2, "L_tab": 869, "n_element_tab": 57, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9801, "formula_len_all_1": 1796, "len_all": 181352, "len_all_1": 72327, "len_abs": 960, "len_title": 97, "len_sents": 62763, "len_sents_1": 32669, "n_sents": 697, "n_sents_1": 291, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 991, "title": "Online Clustering with Nearly Optimal Consistency", "abs": "We give online algorithms for $k$-Means(more generally, $(k, z)$-Clustering) with nearly optimal consistency (a notion suggested by Lattanzi & Vassilvitskii (2017)). \nOur result turns any $\\alpha$-approximate offline algorithm for clustering into an $(1+\\epsilon)\\alpha^2$-competitive online algorithm for clustering with $O(k \\text{poly} \\log n)$ consistency. \nThis consistency bound is optimal up to $\\text{poly} \\log(n)$ factors. \nPlugging in the offline algorithm that returns the exact optimal solution, \nwe obtain the first\n$(1 + \\epsilon)$-competitive online algorithm for clustering that achieves a linear in $k$ consistency.\nThis simultaneously improves several previous results (Lattanzi & Vassilvitskii, 2017; Fichtenberger et al., 2021). \nWe validate the performance of our algorithm on real datasets by plugging in the practically efficient $k$-Means++ algorithm. \nOur online algorithm makes $k$-Means++ achieve good consistency with little overhead to the quality of solutions.", "keywords": [ "clustering", "online", "consistency" ], "rating_list": [ 8, 8, 5 ], "soundness_list": [ 3, 4, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "NA03iMcGDw", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "James Cheng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "Associate Professor" }, { "name": "Jiaqi WANG", "gender": "Female", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xinyi Wu", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Yifei Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" } ], "n_formula": 86, "n_formula_1": 13, "n_ref_uni": 41, "n_ref": 120, "n_ref_all": 168, "n_fig": 12, "n_tab": 11, "L_tab": 4064, "n_element_tab": 330, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 3482, "n_element_tab_1": 114, "formula_len_all": 5954, "formula_len_all_1": 699, "len_all": 236655, "len_all_1": 69968, "len_abs": 1359, "len_title": 131, "len_sents": 77484, "len_sents_1": 31222, "n_sents": 645, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1364, "title": "Oversmoothing as Loss of Sign: Towards Structural Balance in Graph Neural Networks", "abs": "Oversmoothing is a common phenomenon in a wide range of graph neural networks (GNNs), where node representation becomes homogeneous and thus model performance worsens as the number of layers increases. Various strategies have been proposed to combat oversmoothing, but they are based on different heuristics and lack a unified understanding of their inherent mechanisms. In this paper, we revisit the concept of signed graphs and show that a wide class of anti-oversmoothing techniques can be viewed as the propagation on corresponding signed graphs with both positive and negative edges. Leveraging the classic theory of signed graphs, we characterize the asymptotic behaviors of existing methods and reveal that they deviate from the ideal state of structural balance that provably prevents oversmoothing and improves node classification performance. Driven by this unified analysis and theoretical insights, we propose Structural Balanced Propagation (SBP) where we explicitly enhance the structural balance of the signed graph with the help of label and feature information. We theoretically and empirically prove that SBP can improve the structural balance to alleviate oversmoothing under certain conditions. Experiments on synthetic and real-world datasets demonstrate the effectiveness of our methods, highlighting the value of our signed graph framework.", "keywords": [ "graph neural networks", "oversmoothing" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 3, 3 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "N9hTix8kWA", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Dingyi Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Haodong Fan", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Yingming Li", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Yunlong Yu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 30, "n_formula_1": 20, "n_ref_uni": 18, "n_ref": 35, "n_ref_all": 63, "n_fig": 7, "n_tab": 6, "L_tab": 1432, "n_element_tab": 283, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 753, "n_element_tab_1": 131, "formula_len_all": 1622, "formula_len_all_1": 1006, "len_all": 131055, "len_all_1": 77745, "len_abs": 1534, "len_title": 139, "len_sents": 44359, "len_sents_1": 33832, "n_sents": 329, "n_sents_1": 255, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1537, "title": "LDINet: Latent Decomposition-Interpolation for Single Image Fast-moving Objects Deblatting", "abs": "The image of fast-moving objects (FMOs) usually contains a blur stripe indicating the blurred object that is mixed with the background. To deblur the stripe and separate the object from the background in the single image, in this work we propose a novel Latent Decomposition-Interpolation Network (LDINet) to generate the appearances and shapes of the objects. In particular, under the assumption that motion blur is an accumulation of the appearance of the object over exposure time and the long blur can be decomposed into several shorter blur parts, the blurry input is first encoded into latent feature maps and then an efficient Decomposition-Interpolation Module (DIM) is introduced to break down the feature maps into discrete time indexed parts corresponding to different small blurs. And the target latent frames are further interpolated according to the provided time indexes with affine transformations, where the feature maps are categorized into the scalar-like and gradient-like parts to effectively capture the intrinsic properties of features warping in the interpolation. Finally, the sharp and clear images are rendered with a decoder. In addition, based on the generated images by LDINet, a Refining Conditional Deblatting (RCD) approach is presented to use post-image-to-image techniques to further enhance the fidelity of the textures and the accuracy of the masks. Extensive experiments are conducted and have shown that the proposed methods achieve superior performances compared to the existing competing methods.", "keywords": [ "fast moving object deblatting", "image deblur", "time super-resolution", "latent decomposition and interpolation" ], "rating_list": [ 3, 8, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "N8tJmhCw25", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Omar Saadi", "gender": "Male", "institution": "University Mohammed VI Polytechnic", "country": "MA", "position": "Assistant Professor" }, { "name": "Taha EL BAKKALI EL KADI", "gender": "Male", "institution": "University Mohammed VI Polytechnic", "country": "MA", "position": "PhD student" } ], "n_formula": 98, "n_formula_1": 42, "n_ref_uni": 28, "n_ref": 60, "n_ref_all": 70, "n_fig": 3, "n_tab": 2, "L_tab": 391, "n_element_tab": 85, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 56, "n_element_tab_1": 18, "formula_len_all": 8363, "formula_len_all_1": 2963, "len_all": 129861, "len_all_1": 69705, "len_abs": 1482, "len_title": 119, "len_sents": 33612, "len_sents_1": 27020, "n_sents": 308, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1621, "title": "On the Almost Sure Convergence of the Stochastic Three Points Algorithm", "abs": "The stochastic three points (STP) algorithm is a derivative-free optimization technique designed for unconstrained optimization problems in $\\mathbb{R}^d$. In this paper, we analyze this algorithm for three classes of functions : smooth functions that may lack convexity, smooth convex functions, and smooth functions that are strongly convex. Our work provides the first almost sure convergence results of the STP algorithm, alongside some convergence results in expectation.\nFor the class of smooth functions, we establish that the best gradient iterate of the STP algorithm converges almost surely to zero at a rate arbitrarily close to $o(\\frac{1}{\\sqrt{T}})$, where $T$ is the number of iterations. Furthermore, within the same class of functions, we establish both almost sure convergence and convergence in expectation of the final gradient iterate towards zero.\nFor the class of smooth convex functions, we establish that $f(\\theta^T)$ converges to $\\inf_{\\theta \\in \\mathbb{R}^d} f(\\theta)$ almost surely at a rate arbitrarily close to $o(\\frac{1}{T})$, and in expectation at a rate of $O(\\frac{d}{T})$ where $d$ is the dimension of the space.\nFinally, for the class of smooth functions that are strongly convex, we establish that when step sizes are obtained by approximating the directional derivatives of the function, $f(\\theta^T)$ converges to $\\inf_{\\theta \\in \\mathbb{R}^d} f(\\theta)$ in expectation at a rate of $O((1-\\frac{\\mu}{dL})^T)$, and almost surely at a rate arbitrarily close to $o((1-\\frac{\\mu}{dL})^T)$, where $\\mu$ and $L$\nare the strong convexity and smoothness parameters of the function.", "keywords": [ "Zeroth Order Optimization", "Almost sure convergence", "Stochastic Three Points Algorithm." ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "N8Oj1XhtYZ", "primary_area": "generative models", "n_author": 11, "author_info_list": [ { "name": "Enze Xie", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Researcher" }, { "name": "Han Cai", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Haotian Tang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Junyu Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Ligeng Zhu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Muyang Li", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Song Han", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yao Lu", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Yujun Lin", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zhekai Zhang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "\u4fca\u677e \u9648", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 85, "n_ref_all": 115, "n_fig": 14, "n_tab": 16, "L_tab": 4974, "n_element_tab": 826, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1497, "n_element_tab_1": 230, "formula_len_all": 916, "formula_len_all_1": 30, "len_all": 215894, "len_all_1": 62379, "len_abs": 129, "len_title": 131, "len_sents": 53305, "len_sents_1": 26441, "n_sents": 448, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 90, "L_abs": 1398, "title": "SANA: Efficient High-Resolution Text-to-Image Synthesis with Linear Diffusion Transformers", "abs": "We introduce Sana, a text-to-image framework that can efficiently generate images up to 4096$\\times$4096 resolution. Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed, deployable on laptop GPU. Core designs include: (1) Deep compression autoencoder: unlike traditional AEs, which compress images only 8$\\times$, we trained an AE that can compress images 32$\\times$, effectively reducing the number of latent tokens. (2) Linear DiT: we replace all vanilla attention in DiT with linear attention, which is more efficient at high resolutions without sacrificing quality. (3) Decoder-only text encoder: we replaced T5 with modern decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. (4) Efficient training and sampling: we propose Flow-DPM-Solver to reduce sampling steps, with efficient caption labeling and selection to accelerate convergence. As a result, Sana-0.6B is very competitive with modern giant diffusion model (e.g. Flux-12B), being 20 times smaller and 100+ times faster in measured throughput. Moreover, Sana-0.6B can be deployed on a 16GB laptop GPU, taking less than 1 second to generate a 1024$\\times$1024 resolution image. Sana enables content creation at low cost. Code and model will be publicly released upon publication.", "keywords": [ "Efficient AI", "Diffusion Models", "Text to Image generation" ], "rating_list": [ 8, 8, 8, 10 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "N83O2FcqzN", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Huihui Zhou", "gender": "Male", "institution": "Pengcheng Lab", "country": "", "position": "Full Professor" }, { "name": "Liutao Yu", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "", "position": "Engineer" }, { "name": "Liwei Huang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yonghong Tian", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Zhengyu Ma", "gender": "Female", "institution": "Peng Cheng Lab", "country": "CN", "position": "Associate Professor" } ], "n_formula": 32, "n_formula_1": 9, "n_ref_uni": 47, "n_ref": 91, "n_ref_all": 121, "n_fig": 12, "n_tab": 14, "L_tab": 3485, "n_element_tab": 218, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 3225, "n_element_tab_1": 152, "formula_len_all": 3721, "formula_len_all_1": 562, "len_all": 156691, "len_all_1": 64461, "len_abs": 1685, "len_title": 153, "len_sents": 43416, "len_sents_1": 28819, "n_sents": 337, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1691, "title": "Time-Dependent VAE for Building Latent Representations from Visual Neural Activity with Complex Dynamics", "abs": "Seeking high-quality representations with latent variable models (LVMs) to reveal the intrinsic correlation between neural activity and behavior or sensory stimuli has attracted much interest. Most work has focused on analyzing motor neural activity that controls clear behavioral traces and has modeled neural temporal relationships in a way that does not conform to natural reality. For studies of visual brain regions, naturalistic visual stimuli are high-dimensional and time-dependent, making neural activity exhibit intricate dynamics. To cope with such conditions, we propose Time-Dependent Split VAE (TiDeSPL-VAE), a sequential LVM that decomposes visual neural activity into two latent representations while considering time dependence. We specify content latent representations corresponding to the component of neural activity driven by the current visual stimulus, and style latent representations corresponding to the neural dynamics influenced by the organism's internal state. To progressively generate the two latent representations over time, we introduce state factors to construct conditional distributions with time dependence and apply self-supervised contrastive learning to shape them. By this means, TiDeSPL-VAE can effectively analyze complex visual neural activity and model temporal relationships in a natural way. We compare our model with alternative approaches on synthetic data and neural data from the mouse visual cortex. The results show that our model not only yields the best decoding performance on naturalistic scenes/movies but also extracts explicit neural dynamics, demonstrating that it builds latent representations more relevant to visual stimuli.", "keywords": [ "Latent Variable Models", "Visual Cortex", "Sequential Variational Auto-Encoder", "Neural Embeddings" ], "rating_list": [ 8, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "N80ER2he6l", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 11, "author_info_list": [ { "name": "Bo Li", "gender": "unknown", "institution": "Sensetime Research", "country": "", "position": "Intern" }, { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Feizhao Zhang", "gender": "Male", "institution": "Sensetime", "country": "CN", "position": "Researcher" }, { "name": "Jiahao Hu", "gender": "Male", "institution": "Sensetime", "country": "CN", "position": "Researcher" }, { "name": "JinXin", "gender": "unknown", "institution": "Sensetime", "country": "CN", "position": "Researcher" }, { "name": "Jingru Tan", "gender": "Male", "institution": "Central South University", "country": "", "position": "Lecturer" }, { "name": "Ningyi Xu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Pengfei Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Associate Professor" }, { "name": "Ruihao Gong", "gender": "Male", "institution": "SenseTime", "country": "CN", "position": "Principal Researcher" }, { "name": "Yazhe Niu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yongqiang Yao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 13, "n_ref": 31, "n_ref_all": 50, "n_fig": 3, "n_tab": 9, "L_tab": 4597, "n_element_tab": 215, "n_fig_1": 3, "n_tab_1": 9, "L_tab_1": 3205, "n_element_tab_1": 187, "formula_len_all": 614, "formula_len_all_1": 614, "len_all": 121377, "len_all_1": 66518, "len_abs": 1442, "len_title": 147, "len_sents": 29146, "len_sents_1": 26382, "n_sents": 273, "n_sents_1": 255, "venue": "ICLR2025", "decision": "withdraw", "L_title": 98, "L_abs": 1455, "title": "OMNIBAL: TOWARDS FAST INSTRUCT-TUNING FOR VISION-LANGUAGE MODELS VIA OMNIVERSE COMPUTATION BALANCE", "abs": "Vision-language instruct-tuning models have recently made significant progress due to their more comprehensive understanding of the world. In this work, we discover that large-scale 3D parallel training on those models leads to an imbalanced computation load across different devices. The vision and language parts are inherently heterogeneous: their data distribution and model architecture differ significantly, which affects distributed training efficiency. To address this issue, we rebalance the computational load from data, model, and memory perspectives, achieving more balanced computation across devices. Specifically, for the data, instances are grouped into new balanced mini-batches within and across devices. A search-based method is employed for the model to achieve a more balanced partitioning. For memory optimization, we adaptively adjust the re-computation strategy for each partition to utilize the available memory fully. These three perspectives are not independent but are closely connected, forming an omniverse balanced training framework. extensive experiments are conducted to validate the effectiveness of our method. Compared with the open-source training code of InternVL-Chat, training time is reduced greatly, achieving about 1.8x speed-up. Our method's efficacy and generalizability are further validated across various models and datasets. Codes will be released at https://github.com/anonymousiclr293/omnibal_example.", "keywords": [ "Fast Vision-Language Training; 3D parallel" ], "rating_list": [ 6, 6, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "N6pbLYLeej", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Chenxiao Yang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "David Wipf", "gender": "Male", "institution": "Amazon AI Research Lab", "country": "", "position": "Principal Research Scientist" }, { "name": "Zhiyuan Li", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Assistant Professor" } ], "n_formula": 82, "n_formula_1": 27, "n_ref_uni": 47, "n_ref": 104, "n_ref_all": 125, "n_fig": 1, "n_tab": 3, "L_tab": 455, "n_element_tab": 2, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 2283, "n_element_tab_1": 26, "formula_len_all": 7040, "formula_len_all_1": 2095, "len_all": 185281, "len_all_1": 73219, "len_abs": 1409, "len_title": 118, "len_sents": 56212, "len_sents_1": 32960, "n_sents": 478, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1415, "title": "Chain-of-Thought Provably Enables Learning the (Otherwise) Unlearnable", "abs": "Modern language models have demonstrated remarkable reasoning capabilities by using chain-of-thought (CoT). One hypothesis about the inner workings of CoT is that it breaks down originally complex tasks into smaller subtasks that are more amenable to learning. We formalize this notion by showing possibility and impossibility results of learning from in-context demonstrations with and without CoT. In particular, with CoT, we examine a family of learning algorithms that learn a task step-by-step, capable of composing simpler functions from individual reasoning steps to form an overall complex function. This process reduces the difficulty of learning a task to that of the hardest reasoning step in the chain. Moreover, we prove Transformers can express this algorithm and thus they can efficiently in-context learn arbitrary tasks as long as these tasks can be decomposed into a finite number of subtasks, each of which are efficiently learnable. In contrast, without CoT, we demonstrate that there exist tasks that are inherently unlearnable by the same algorithm. Overall, our results suggest several provably effective ways for decomposing target problems to instantiate CoT. Empirically, we demonstrate our proposed CoT construction significantly enhances the reasoning capabilities of real-world LLMs in solving challenging arithmetic reasoning tasks, including learning polynomials and Boolean formulas.", "keywords": [ "large language models", "learning theory", "chain-of-thought" ], "rating_list": [ 6, 6, 10, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 3, 4, 3 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "N6ba2xsmds", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Hengzhuang Li", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Teng Zhang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 43, "n_formula_1": 16, "n_ref_uni": 84, "n_ref": 249, "n_ref_all": 314, "n_fig": 12, "n_tab": 26, "L_tab": 21899, "n_element_tab": 2476, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3328, "n_element_tab_1": 375, "formula_len_all": 2742, "formula_len_all_1": 1214, "len_all": 352247, "len_all_1": 78098, "len_abs": 1221, "len_title": 127, "len_sents": 87769, "len_sents_1": 30516, "n_sents": 688, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1226, "title": "Outlier Synthesis via Hamiltonian Monte Carlo for Out-of-Distribution Detection", "abs": "Out-of-distribution (OOD) detection is crucial for developing trustworthy and reliable machine learning systems. Recent advances in training with auxiliary OOD data demonstrate efficacy in enhancing detection capabilities. Nonetheless, these methods heavily rely on acquiring a large pool of high-quality natural outliers. Some prior methods try to alleviate this problem by synthesizing virtual outliers but suffer from either poor quality or high cost due to the monotonous sampling strategy and the heavy-parameterized generative models. In this paper, we overcome all these problems by proposing the Hamiltonian Monte Carlo Outlier Synthesis (HamOS) framework, which views the synthesis process as sampling from Markov chains. Based solely on the in-distribution data, the Markov chains can extensively traverse the feature space and generate diverse and representative outliers, hence exposing the model to miscellaneous potential OOD scenarios. The Hamiltonian Monte Carlo with sampling acceptance rate almost close to 1 also makes our framework enjoy great efficiency. By empirically competing with SOTA baselines on both standard and large-scale benchmarks, we verify the efficacy and efficiency of our proposed HamOS.", "keywords": [ "Trustworthy Machine Learning", "Out-of-Distribution Detection", "Outlier Detection" ], "rating_list": [ 5, 8, 8, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "N6SccBt3EF", "primary_area": "applications to robotics, autonomy, planning", "n_author": 11, "author_info_list": [ { "name": "Anthony Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Chun-Kai Fan", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Hengyuan Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Rongyu Zhang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Wei Xue", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Wenhan Luo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Xiaowei Chi", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Xingqun Qi", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yike Guo", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "CHIMIN CHAN", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Shanghang Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 50, "n_ref": 107, "n_ref_all": 146, "n_fig": 10, "n_tab": 8, "L_tab": 2952, "n_element_tab": 300, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1196, "n_element_tab_1": 107, "formula_len_all": 231, "formula_len_all_1": 186, "len_all": 213424, "len_all_1": 53310, "len_abs": 1669, "len_title": 0, "len_sents": 58870, "len_sents_1": 24270, "n_sents": 526, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1740, "title": "EVA: An Embodied World Model for Future Video Anticipation", "abs": "World models integrate raw data from various modalities\u2014such as images and language to simulate comprehensive interactions in the world, thereby displaying crucial roles in fields like mixed reality and robotics. \nYet, applying the world model for accurate video prediction is quite challenging due to the complex and dynamic intentions of the various scenes in practice. \nIn this paper, inspired by the human rethinking process, we decompose the complex video prediction into four meta-tasks that enable the world model to handle this issue in a more fine-grained manner. \nAlongside these tasks, we introduce a new benchmark named Embodied Video Anticipation Benchmark (EVA-Bench) to provide a well-rounded evaluation. \nEVA-Bench focused on evaluating the video prediction ability of human and robot actions, presenting significant challenges for both the language model and the generation model. \nTargeting embodied video prediction, we propose the Embodied Video Anticipator (EVA), a unified framework aiming at video understanding and generation. \nEVA integrates a video generation model with a visual language model, effectively combining reasoning capabilities with high-quality generation.\nMoreover, to enhance the generalization of our framework, we tailor-designed a multi-stage pretraining paradigm that adaptatively ensembles LoRA to produce high-fidelity results. \nExtensive experiments on EVA-Bench highlight the potential of EVA to significantly improve performance in embodied scenes, paving the way for large-scale pre-trained models in real-world prediction tasks. The video demo and benchmark information will be available at \\hyperlink{https://sites.google.com/view/iclr25-eva}{https://sites.google.com/view/iclr25-eva}.", "keywords": [ "World Model", "Video Generation", "Visual Language Model", "Embodied AI" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "N5qFgohx9u", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Anna-Maria Halacheva", "gender": "Female", "institution": "Sofia University \"St. Kliment Ohridski\"", "country": "BG", "position": "PhD student" }, { "name": "Mojtaba Nayyeri", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "DE", "position": "PhD student" }, { "name": "Steffen Staab", "gender": "Male", "institution": "University of Stuttgart", "country": "DE", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 13, "n_ref_uni": 58, "n_ref": 123, "n_ref_all": 163, "n_fig": 11, "n_tab": 6, "L_tab": 1549, "n_element_tab": 126, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 807, "n_element_tab_1": 66, "formula_len_all": 1845, "formula_len_all_1": 817, "len_all": 205395, "len_all_1": 68805, "len_abs": 1441, "len_title": 113, "len_sents": 55528, "len_sents_1": 32413, "n_sents": 435, "n_sents_1": 229, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1447, "title": "Expanding Expressivity in Transformer Models with M\u00f6biusAttention", "abs": "Attention mechanisms and Transformer architectures have revolutionized Natural Language Processing (NLP) by enabling exceptional modeling of long-range dependencies and capturing intricate linguistic patterns. However, their inherent reliance on linear operations in the form of matrix multiplications limits their ability to fully capture inter-token relationships on their own. We propose M\u00f6biusAttention, a novel approach that integrates M\u00f6bius transformations within the attention mechanism of Transformer-based models. M\u00f6bius transformations are non-linear operations in spaces over complex numbers with the ability to map between various geometries. By incorporating these properties, M\u00f6biusAttention empowers models to learn more intricate geometric relationships between tokens and capture a wider range of information through complex-valued weight vectors. We build and pre-train a BERT and a RoFormer version enhanced with M\u00f6biusAttention, which we then finetune on the GLUE benchmark. We evaluate empirically our approach against the baseline BERT and RoFormer models on a range of downstream tasks. Our approach compares favorably against the baseline models, even with smaller number of parameters suggesting the enhanced expressivity of M\u00f6biusAttention. This research paves the way for exploring the potential of M\u00f6bius transformations in the complex projective space to enhance the expressivity and performance of foundation models.", "keywords": [ "Transformer", "Attention", "BERT", "M\u00f6bius Transformation", "NLP", "RoPe" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "N5fVv6PZGz", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 5, "author_info_list": [ { "name": "Baris Kasikci", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Associate Professor" }, { "name": "Kan Zhu", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Keisuke Kamahori", "gender": "unknown", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Tian Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yile Gu", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 81, "n_ref_all": 107, "n_fig": 15, "n_tab": 2, "L_tab": 1214, "n_element_tab": 192, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 565, "n_element_tab_1": 30, "formula_len_all": 0, "formula_len_all_1": 89, "len_all": 119992, "len_all_1": 59356, "len_abs": 1335, "len_title": 138, "len_sents": 37949, "len_sents_1": 28564, "n_sents": 280, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1122, "title": "Orchestrating Heterogeneous Architectures for Fast Inference of Mixture-of-Experts Models", "abs": "Large Language Models (LLMs) with the Mixture-of-Experts (MoE) architectures have shown promising performance on various tasks. However, due to the huge model sizes, running them in resource-constrained environments where GPU memory is not abundant is challenging. Some existing systems propose to use CPU resources to solve that, but they either suffer from significant overhead of frequently moving data between CPU and GPU, or fail to consider different characteristics of CPU and GPU. This paper proposes Twiddler, a resource-efficient inference system for MoE models with limited GPU resources. Twiddler strategically utilizes the heterogeneous computing architecture of CPU and GPU resources by determining the optimal execution strategy. Our evaluation shows that, unlike state-of-the-art systems that optimize for specific scenarios such as single batch inference or long prefill, Twiddler has better performance in all scenarios. Twiddler achieves 1.26 times speed up in single batch inference, 1.30 times in long prefill processing, and 11.57 times in beam search inference, compared against different baselines.", "keywords": [ "NLP in resource-constrained settings", "inference methods" ], "rating_list": [ 8, 6, 5 ], "soundness_list": [ 3, 3, 1 ], "presentation_list": [ 4, 3, 2 ], "contribution_list": [ 3, 3, 1 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "N5YTixK4F1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Jie Wu", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Peize Sun", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shilong Zhang", "gender": "Male", "institution": "University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Sidi Yang", "gender": "Male", "institution": "University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Weifeng Chen", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Xuefeng Xiao", "gender": "Male", "institution": "ByteDance", "country": "CN", "position": "Researcher" }, { "name": "Yatai Ji", "gender": "Male", "institution": "University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Yujiu Yang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Ping Luo", "gender": "unknown", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 107, "n_ref_all": 127, "n_fig": 7, "n_tab": 5, "L_tab": 1307, "n_element_tab": 159, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 485, "n_element_tab_1": 62, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 130528, "len_all_1": 61503, "len_abs": 1236, "len_title": 125, "len_sents": 32756, "len_sents_1": 28372, "n_sents": 255, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 77, "L_abs": 1240, "title": "IDA-VLM: Towards Movie Understanding via ID-Aware Large Vision-Language Model", "abs": "The rapid advancement of Large Vision-Language models (LVLMs) has demonstrated a spectrum of emergent capabilities. Nevertheless, current models only focus on the visual content of a single scenario, while their ability to associate instances across different scenes has not yet been explored, which is essential for understanding complex visual content, such as movies with multiple characters and intricate plots. Towards movie understanding, a critical initial step for LVLMs is to unleash the potential of character identities memory and recognition across multiple visual scenarios. To achieve the goal, we propose visual instruction tuning with ID reference and develop an ID-Aware Large Vision-Language Model, IDA-VLM. Furthermore, our research introduces a novel benchmark MM-ID, to examine LVLMs on instance IDs memory and recognition across four dimensions: matching, location, question-answering, and captioning. Our findings highlight the limitations of existing LVLMs in recognizing and associating instance identities with ID reference. This paper paves the way for future artificial intelligence systems to possess multi-identity visual inputs, thereby facilitating the comprehension of complex visual narratives like movies.", "keywords": [ "large vision-language model; ID recognition" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "N4rYbQowE3", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Brandon G. Nguyen", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "MS student" }, { "name": "Chunkai Fu", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "MS student" }, { "name": "Jung Hoon Seo", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "MS student" }, { "name": "Ryan S. Zesch", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Samson Zhou", "gender": "unknown", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" } ], "n_formula": 20, "n_formula_1": 6, "n_ref_uni": 59, "n_ref": 100, "n_ref_all": 129, "n_fig": 13, "n_tab": 3, "L_tab": 1637, "n_element_tab": 147, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 117, "n_element_tab_1": 8, "formula_len_all": 1483, "formula_len_all_1": 551, "len_all": 211693, "len_all_1": 68981, "len_abs": 1410, "len_title": 89, "len_sents": 70366, "len_sents_1": 35402, "n_sents": 579, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 1416, "title": "Learning-Augmented Search Data Structures", "abs": "We study the integration of machine learning advice to improve upon traditional data structure designed for efficient search queries. Although there has been recent effort in improving the performance of binary search trees using machine learning advice, e.g., Lin et. al. (ICML 2022), the resulting constructions nevertheless suffer from inherent weaknesses of binary search trees, such as complexity of maintaining balance across multiple updates and the inability to handle partially-ordered or high-dimensional datasets. For these reasons, we focus on skip lists and KD trees in this work. Given access to a possibly erroneous oracle that outputs estimated fractional frequencies for search queries on a set of items, we construct skip lists and KD trees that provably provides the optimal expected search time, within nearly a factor of two. In fact, our learning-augmented skip lists and KD trees are still optimal up to a constant factor, even if the oracle is only accurate within a constant factor. We also demonstrate robustness by showing that our data structures achieves an expected search time that is within a constant factor of an oblivious skip list/KD tree construction even when the predictions are arbitrarily incorrect. Finally, we empirically show that our learning-augmented search data structures outperforms their corresponding traditional analogs on both synthetic and real-world datasets.", "keywords": [ "learning-augmented algorithms", "data structures" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "N4mb3MBV6J", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Thang D Bui", "gender": "Male", "institution": "Australian National University", "country": "", "position": "Assistant Professor" }, { "name": "Yashvir Singh Grewal", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "PhD student" }, { "name": "Edwin V Bonilla", "gender": "Male", "institution": "CSIRO's Data61", "country": "", "position": "Principal Research Scientist" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 57, "n_ref_all": 67, "n_fig": 3, "n_tab": 4, "L_tab": 1529, "n_element_tab": 92, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1126, "n_element_tab_1": 91, "formula_len_all": 373, "formula_len_all_1": 373, "len_all": 128120, "len_all_1": 63233, "len_abs": 1381, "len_title": 134, "len_sents": 35448, "len_sents_1": 31015, "n_sents": 233, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1388, "title": "Improving Uncertainty Quantification in Large Language Models via Semantic Embeddings", "abs": "Accurately quantifying uncertainty in large language models (LLMs) is crucial for their reliable deployment, especially in high-stakes applications. Current state-of-the-art methods for measuring semantic uncertainty in LLMs rely on strict bidirectional entailment criteria between multiple generated responses and also depend on sequence likelihoods. While effective, these approaches often overestimate uncertainty due to their sensitivity to minor wording differences, additional correct information, and non-important words in the sequence. We propose a novel approach that leverages semantic embeddings to achieve smoother and more robust estimation of semantic uncertainty in LLMs. By capturing semantic similarities without depending on sequence likelihoods, our method inherently reduces any biases introduced by irrelevant words in the answers. Furthermore, we introduce an amortised version of our approach by explicitly modelling semantics as latent variables in a joint probabilistic model. This allows for uncertainty estimation in the embedding space with a single forward pass, significantly reducing computational overhead compared to existing multi-pass methods. Experiments across multiple question-answering datasets and frontier LLMs demonstrate that our embedding-based methods provide more accurate and nuanced uncertainty quantification than traditional approaches.", "keywords": [ "uncertainty estimation", "large language models", "natural language generation", "variational inference" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 1, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "N4lUNwEn1c", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Aleksandr Beznosikov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "Associate Professor" }, { "name": "Alexey Orlov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "MS student" }, { "name": "Elizaveta Starykh", "gender": "Female", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Klea Ziu", "gender": "Female", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Nikolai Rekut", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "Undergrad student" }, { "name": "Martin Takac", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Associate Professor" } ], "n_formula": 14, "n_formula_1": 3, "n_ref_uni": 37, "n_ref": 76, "n_ref_all": 88, "n_fig": 6, "n_tab": 5, "L_tab": 655, "n_element_tab": 95, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 643, "n_element_tab_1": 86, "formula_len_all": 491, "formula_len_all_1": 115, "len_all": 148873, "len_all_1": 55881, "len_abs": 1123, "len_title": 186, "len_sents": 53068, "len_sents_1": 27378, "n_sents": 351, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 139, "L_abs": 1127, "title": "Broadening Discovery through Structural Models: Multimodal Combination of Local and Structural Properties for Predicting Chemical Features.", "abs": "In recent years, machine learning (ML) has significantly impacted the field of chemistry, facilitating advancements in diverse applications such as the prediction of molecular properties and the generation of molecular structures. Traditional string representations, such as the Simplified Molecular Input Line Entry System (SMILES), although widely adopted, exhibit limitations in conveying essential physical and chemical properties of compounds. Conversely, vector representations, particularly chemical fingerprints, have demonstrated notable efficacy in various ML tasks. Additionally, graph-based models, which leverage the inherent structural properties of chemical compounds, have shown promise in improving predictive accuracy. This study investigates the potential of language models based on fingerprints within a bimodal architecture that combines both graph-based and language model components. We propose a method that integrates the aforementioned approaches, significantly enhancing predictive performance compared to conventional methodologies while simultaneously capturing more accurate chemical information.", "keywords": [ "ECFP", "GRAPH", "LLMs", "TRANSFORMERS" ], "rating_list": [ 3, 1, 5 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "N4QQNU9HK3", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Amin Beheshti", "gender": "Male", "institution": "Macquarie University", "country": "AU", "position": "Full Professor" }, { "name": "Jiayi Chen", "gender": "Female", "institution": "University of Sydney", "country": "AU", "position": "MS student" }, { "name": "Maryam Khanian Najafabadi", "gender": "unknown", "institution": "University of Sydney, University of Sydney", "country": "", "position": "Lecturer" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 20, "n_ref": 33, "n_ref_all": 40, "n_fig": 3, "n_tab": 1, "L_tab": 544, "n_element_tab": 37, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 544, "n_element_tab_1": 37, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 88885, "len_all_1": 65465, "len_abs": 1336, "len_title": 122, "len_sents": 36134, "len_sents_1": 34749, "n_sents": 217, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 120, "L_abs": 1340, "title": "ENHANCING DIVERSITY AND ACCURACY IN PERSONALIZED TAG RECOMMENDATIONS: A HYBRID SEMANTIC AND CONTEXTUAL ANALYSIS APPROACH", "abs": "This paper introduces HYCOMB, a cascading Hybrid model that innovatively integrates\nCollaborative Filtering (CF), Content-Based Filtering (CB), and Context-\nAware (CA) methods to address the challenge of data sparsity in tag recommendation\nsystems. Unlike traditional models that rely heavily on user-item interactions,\nHYCOMB enhances recommendation diversity and interpretability by utilizing\nsemantic clustering in CF to extract and analyze user sentiment from tags, adding\na layer of nuanced understanding often missing in conventional systems. The CB\ncomponent advances this by applying sophisticated NLP techniques to refine these\nrecommendations based on item attributes, while the CA component incorporates\nmovie synopses for deeper contextual understanding. Developed and tested using\nthe MovieLens 20M dataset, our model demonstrates significant outperformance\nover baseline methods in terms of precision and recall, achieving scores of 0.813\nand 0.364 respectively. Further, a newly introduced Overall Total Similarity metric\nthat underscores its ability to deliver relevant and diverse recommendations.\nHYCOMB\u2019s strategic amalgamation of CF, CB, and CA not only mitigates the\neffects of sparse data but also improves the precision and diversity of tag recommendations,\nreflecting a more accurate alignment with user preferences.", "keywords": [ "Personalized content", "Tag recommendation", "Content-based", "Collaborative filtering", "Context-aware", "Semantic clustering" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "N4NhVN30ph", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Dong Tian", "gender": "unknown", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "MS student" }, { "name": "Ge Li", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Hongyi Zhou", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "PhD student" }, { "name": "Rudolf Lioutikov", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "Tenure-Track Professor" }, { "name": "Xinkai Jiang", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "", "position": "PhD student" }, { "name": "Gerhard Neumann", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 32, "n_formula_1": 12, "n_ref_uni": 62, "n_ref": 134, "n_ref_all": 179, "n_fig": 10, "n_tab": 8, "L_tab": 3788, "n_element_tab": 611, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1999, "formula_len_all_1": 1067, "len_all": 208811, "len_all_1": 77899, "len_abs": 1381, "len_title": 117, "len_sents": 62728, "len_sents_1": 36668, "n_sents": 463, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 69, "L_abs": 1298, "title": "TOP-ERL: Transformer-based Off-Policy Episodic Reinforcement Learning", "abs": "This work introduces Transformer-based Off-Policy Episodic Reinforcement Learning (TOP-ERL), a novel algorithm that enables off-policy updates in the ERL framework. In ERL, policies predict entire action trajectories over multiple time steps instead of single actions at every time step. These trajectories are typically parameterized by trajectory generators such as Movement Primitives (MP), allowing for smooth and efficient exploration over long horizons while capturing high-level temporal correlations. However, ERL methods are often constrained to on-policy frameworks due to the difficulty of evaluating state-action values for entire action sequences, limiting their sample efficiency and preventing the use of more efficient off-policy architectures. TOP-ERL addresses this shortcoming by segmenting long action sequences and estimating the state-action values for each segment using a transformer-based critic architecture alongside an n-step return estimation. These contributions result in efficient and stable training that is reflected in the empirical results conducted on sophisticated robot learning environments. TOP-ERL significantly outperforms state-of-the-art RL methods. Thorough ablation studies additionally show the impact of key design choices on the model performance.", "keywords": [ "Value of sequences of actions", "Reinforcement Learning", "Transformer", "Robot Manipulation", "Movement Primitives." ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "N3EFTVCFWX", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Karan Goyal", "gender": "unknown", "institution": "Indraprastha Institute of Information Technology, Delhi", "country": "IN", "position": "PhD student" }, { "name": "Mukesh Mohania", "gender": "Male", "institution": "Indraprastha Institute of Information Technology", "country": "IN", "position": "Full Professor" }, { "name": "Saankhya Samanta", "gender": "unknown", "institution": "Indian Institute of Engineering Science and Technology, Shibpur", "country": "IN", "position": "Undergrad student" }, { "name": "Vikram Goyal", "gender": "Male", "institution": "Indraprastha Institute of Information Technology, Delhi", "country": "", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 58, "n_ref": 105, "n_ref_all": 125, "n_fig": 1, "n_tab": 12, "L_tab": 9288, "n_element_tab": 1123, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 8696, "n_element_tab_1": 631, "formula_len_all": 548, "formula_len_all_1": 422, "len_all": 173620, "len_all_1": 79176, "len_abs": 1492, "len_title": 111, "len_sents": 39131, "len_sents_1": 30151, "n_sents": 297, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1508, "title": "Spectral Highways: Injecting Homophily into Heterophilic Graphs", "abs": "It is widely assumed that standard GNNs perform better on graphs with high homophily, leading to the development of specialised algorithms for heterophilic datasets in recent years. In this work, we both challenge and leverage this assumption. Rather than creating new algorithms, we emphasise the importance of understanding and enriching the data. We introduce a novel data engineering technique, \\textit{Spectral Highways}, that enhances the performance of both heterophilic and non-heterophilic GNNs on heterophilic datasets. Our method augments a given heterophilic graph by adding supernodes, thereby creating a network of highways connecting spectral clusters in the graph. It facilitates additional paths to bring similar nodes closer than dissimilar ones by reducing the average shortest path lengths. We draw both intuitive and empirical connections between the relative decreases in intraclass and interclass average shortest path lengths and shifts in the graph's homophily levels, providing a novel perspective that extends beyond traditional homophily measures. We conduct extensive experiments on seven heterophilic datasets using various GNN architectures and also compare with data-centric techniques, demonstrating significant improvements in node classification performance. Furthermore, our empirical findings highlight the strong sensitivity of several recent GNNs to the random seed used for data splitting, underscoring the importance of this often-overlooked factor in GNN evaluation.", "keywords": [ "Data augmentation", "Graph representation learning", "GNNs", "Homophily", "Heterophily" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "N2wPtFVK6o", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Dongmin Park", "gender": "Male", "institution": "KRAFTON", "country": "KR", "position": "Researcher" }, { "name": "Irfan Akbar", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Jae-Gil Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Patara Trirat", "gender": "Male", "institution": "DeepAuto.ai", "country": "KR", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 13, "n_ref_uni": 34, "n_ref": 66, "n_ref_all": 95, "n_fig": 9, "n_tab": 7, "L_tab": 5390, "n_element_tab": 390, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 2974, "n_element_tab_1": 199, "formula_len_all": 1352, "formula_len_all_1": 760, "len_all": 136058, "len_all_1": 63044, "len_abs": 1110, "len_title": 121, "len_sents": 38435, "len_sents_1": 27019, "n_sents": 290, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1130, "title": "MOUCHI: Mitigating Over-forgetting in Unlearning Copyrighted Information", "abs": "Large language models are trained on massive internet datasets, which may inadvertently memorize illegal copyrighted content, making its inclusion unavoidable. Unlearning is a potential solution to remove such content. However, existing unlearning methods often suffer from **over-forgetting**, where the process unintentionally erases knowledge similar to the copyrighted content that falls under fair use and should be preserved. To address this issue, we propose **MOUCHI**, a novel unlearning framework that introduces the concept of **derivative knowledge**, a subset of information derived from copyrighted content that must be retained during unlearning. MOUCHI first generates derivative knowledge and then incorporates a derivative loss function into the unlearning process to mitigate over-forgetting in unlearning copyrighted content. Due to its plug-and-play nature, MOUCHI can be effortlessly integrated into existing unlearning methods. Experimental results show that MOUCHI reduces unintended knowledge loss, improving performance by **up to 145%** compared to baseline methods when evaluated on the derivative set.", "keywords": [ "machine unlearning", "LLM unlearning", "derivative knowledge" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "N2sN3LESoW", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Chen Xiuyu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Jiqing Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Lou Jieming", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "WEIDE LIU", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 26, "n_ref": 43, "n_ref_all": 49, "n_fig": 2, "n_tab": 4, "L_tab": 1033, "n_element_tab": 226, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1033, "n_element_tab_1": 226, "formula_len_all": 983, "formula_len_all_1": 993, "len_all": 86567, "len_all_1": 56343, "len_abs": 1182, "len_title": 132, "len_sents": 23437, "len_sents_1": 21710, "n_sents": 172, "n_sents_1": 159, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1190, "title": "Gap-Aware Preference Optimization: Enhancing Model Alignment with Perception Margin", "abs": "Reinforcement learning from human feedback (RLHF) approaches are widely used for fine-tuning large language models (LLMs) to align with instructional preferences. However, traditional RLHF methods often rely on binary labels, which fail to capture the pairwise differences in human perception, leading to potential performance degradation.\nTo address this limitation, we introduce $\\textbf{Gap-Aware Preference Optimization}$ (GaPO), a novel approach that integrates the degree of semantic gaps into preference optimization. By modifying the margin term in the loss function and replacing it with an estimated gap computed using general metrics, GaPO provides a new supervisory signal that explicitly highlights the nuances between preference pairs. This new signal helps the model allocate gradients more rationally during optimization, facilitating more effective learning from the preference data.\nExperiments conducted with a strong base model, Llama-3-8B-Instruct, demonstrate that GaPO surpasses state-of-the-art methods on widely used benchmarks. Our best-performing model, GaPO-ROUGE\\_L, achieves a win rate of 52.8\\% on AlpacaEval 2.0, exceeding the baseline methods by 5.3 points.", "keywords": [ "LLM Alignments; Gap-Aware Preference Optimization" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 3, 4, 2 ] }, { "paper_id": "N23g8eGOiP", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 3, "author_info_list": [ { "name": "Lili Mou", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "Assistant Professor" }, { "name": "Yanshuai Cao", "gender": "unknown", "institution": "BorealisAI", "country": "", "position": "Researcher" }, { "name": "Yongchang Hao", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 44, "n_ref": 89, "n_ref_all": 120, "n_fig": 14, "n_tab": 7, "L_tab": 3983, "n_element_tab": 369, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 2501, "n_element_tab_1": 282, "formula_len_all": 311, "formula_len_all_1": 311, "len_all": 161100, "len_all_1": 64354, "len_abs": 817, "len_title": 140, "len_sents": 40699, "len_sents_1": 26242, "n_sents": 376, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 821, "title": "NeuZip: Memory-Efficient Training and Inference with Dynamic Compression of Neural Networks", "abs": "The performance of neural networks improves when more parameters are used. However, the model sizes are constrained by the available on-device memory during training and inference. Although applying techniques like quantization can alleviate the constraint, they suffer from performance degradation. In this work, we introduce NeuZip, a new weight compression scheme based on the entropy of floating-point numbers in neural networks. With NeuZip, we are able to achieve memory-efficient training and inference without sacrificing performance. Notably, we significantly reduce the memory footprint of training a Llama-3 8B model from 31GB to less than 16GB, while keeping the training dynamics fully unchanged. In inference, our method can reduce memory usage by more than half while maintaining near-lossless performance.", "keywords": [ "Compression", "Quantization", "Memory-efficient training", "Memory-efficient inference" ], "rating_list": [ 3, 8, 3 ], "soundness_list": [ 2, 4, 2 ], "presentation_list": [ 2, 3, 1 ], "contribution_list": [ 1, 4, 1 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "N1vYivuSKq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Jianbing Shen", "gender": "Male", "institution": "University of Macau", "country": "CN", "position": "Full Professor" }, { "name": "Yu Cheng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Yucheng Zhou", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 11, "n_ref_uni": 19, "n_ref": 67, "n_ref_all": 113, "n_fig": 27, "n_tab": 4, "L_tab": 810, "n_element_tab": 53, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 3267, "n_element_tab_1": 81, "formula_len_all": 1202, "formula_len_all_1": 560, "len_all": 171908, "len_all_1": 59400, "len_abs": 1523, "len_title": 127, "len_sents": 45072, "len_sents_1": 25428, "n_sents": 330, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1531, "title": "Weak to Strong Generalization for Large Language Models with Multi-capabilities", "abs": "As large language models (LLMs) grow in sophistication, some of their capabilities surpass human abilities, making it essential to ensure their alignment with human values and intentions, i.e., Superalignment. This superalignment challenge is particularly critical for complex tasks, as annotations provided by humans, as weak supervisors, may be overly simplistic, incomplete, or incorrect. Previous work has demonstrated the potential of training a strong model using the weak dataset generated by a weak model as weak supervision. However, these studies have been limited to a single capability. In this work, we conduct extensive experiments to investigate weak to strong generalization for LLMs with multi-capabilities. The experiments reveal that different capabilities tend to remain relatively independent in this generalization, and the effectiveness of weak supervision is significantly impacted by the quality and diversity of the weak datasets. Moreover, the self-bootstrapping of the strong model leads to performance degradation due to its overconfidence and the limited diversity of its generated dataset. To address these issues, we proposed a novel training framework using reward models to select valuable data, thereby providing weak supervision for strong model training. In addition, we propose a two-stage training method on both weak and selected datasets to train the strong model. Experimental results demonstrate our method significantly improves the weak to strong generalization with multi-capabilities.", "keywords": [ "Weak to Strong Generalization", "Large Language Model", "Superalignment" ], "rating_list": [ 6, 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 4, 3, 3, 2, 4 ], "contribution_list": [ 4, 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 5 ] }, { "paper_id": "N1pya6kv3g", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Ge Liu", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Tao Feng", "gender": "Male", "institution": "Department of Computer Science, University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Yihang Sun", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Jiaxuan You", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 51, "n_ref": 94, "n_ref_all": 133, "n_fig": 11, "n_tab": 14, "L_tab": 4929, "n_element_tab": 504, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1550, "n_element_tab_1": 201, "formula_len_all": 488, "formula_len_all_1": 331, "len_all": 218203, "len_all_1": 71729, "len_abs": 1992, "len_title": 118, "len_sents": 61746, "len_sents_1": 34126, "n_sents": 448, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1744, "title": "PREMIUM: LLM Personalization with Individual-level Preference Feedback", "abs": "With an increasing demand for LLM personalization, various methods have been developed to deliver customized LLM experiences, including in-context learning, retrieval augmentation, and parameter-efficient fine-tuning. However, most existing methods are not readily locally deployable, limited by the compute cost, privacy risks, and an inability to adapt to dynamic user preferences. Here, we propose to use a tag system to efficiently characterize user profiles, inspired from the insights from personality typology and recommendation systems. Based on the observation, we present a locally deployable LLM-agnostic framework for achieving LLM personalization: $\\textbf{PREMIUM}$ ($\\textbf{P}$reference $\\textbf{R}$anking $\\textbf{EM}$powered $\\textbf{I}$ndividual $\\textbf{U}$ser $\\textbf{M}$odeling), which obtains individual-level feedback by having users rank responses and continuously self-iterates optimization during the interaction between the user and the LLM. Notably, a variant of PREMIUM, PREMIUM-Embed, can effectively capture user preferences while being deployable with laptop-level resources. Besides algorithmic innovation, we further prepare a novel dataset, Ranking-TAGER, which provides a valuable evaluation protocol for LLM personalization. Extensive experiments validate that PREMIUM remarkably outperforms various baselines, achieving a 15\\%-50\\% higher accuracy and a 2.5\\%-35\\% higher win rate on Ranking-TAGER, as well as a 3\\%-13\\% higher accuracy and a 2\\%-7.5\\% higher F1 Score on LaMP-2. More importantly, we further demonstrate that PREMIUM can develop an effective strategy with minimal interactive data, adapt to dynamic user preferences, and demonstrate excellent scalability in both scale and functionality.", "keywords": [ "Preference Ranking", "Tagging System", "LLM Personalization", "Prompt-Based", "Embedding-Based" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "N1L5TgtkAw", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Arash Behboodi", "gender": "Male", "institution": "QualComm", "country": "", "position": "Machine Learning Researcher" }, { "name": "Ashish J Khisti", "gender": "Male", "institution": "Toronto University", "country": "", "position": "Professor" }, { "name": "Christos Louizos", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "NL", "position": "Researcher" }, { "name": "Hassan Dbouk", "gender": "unknown", "institution": "Qualcomm Inc, QualComm", "country": "", "position": "Researcher" }, { "name": "MohammadReza Ebrahimi", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "CA", "position": "Researcher" }, { "name": "Roland Memisevic", "gender": "Male", "institution": "Twenty Billion Neurons", "country": "CA", "position": "Chief Scientist" } ], "n_formula": 149, "n_formula_1": 15, "n_ref_uni": 29, "n_ref": 99, "n_ref_all": 215, "n_fig": 9, "n_tab": 7, "L_tab": 1084, "n_element_tab": 66, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2130, "n_element_tab_1": 32, "formula_len_all": 18289, "formula_len_all_1": 2280, "len_all": 253783, "len_all_1": 69571, "len_abs": 1316, "len_title": 128, "len_sents": 75202, "len_sents_1": 29353, "n_sents": 744, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 80, "L_abs": 1323, "title": "Multi-Draft Speculative Sampling: Canonical Architectures and Theoretical Limits", "abs": "We consider multi-draft speculative sampling, where the proposal sequences are sampled independently from different draft models. At each step, a token-level draft selection scheme takes a list of valid tokens as input and produces an output token whose distribution matches that of the target model. Previous works have demonstrated that the optimal scheme (which maximizes the probability of accepting one of the input tokens) can be cast as a solution to a linear program. In this work we show that the optimal scheme can be decomposed into a two-step solution: in the first step an importance sampling (IS) type scheme is used to select one intermediate token; in the second step (single-draft) speculative sampling is applied to generate the output token. For the case of two identical draft models we further 1) establish a necessary and sufficient condition on the distributions of the target and draft models for the acceptance probability to equal one and 2) provide an explicit expression for the optimal acceptance probability. Our theoretical analysis also motives a new class of token-level selection scheme based on weighted importance sampling. Our experimental results demonstrate consistent improvements in the achievable block efficiency and token rates over baseline schemes in a number of scenarios.", "keywords": [ "speculative decoding", "multi draft speculative sampling", "large language models", "weighted importance sampling", "optimal transport" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 2, 4, 4, 4 ], "confidence_list": [ 2, 2, 4, 3 ] }, { "paper_id": "N1DKrLIKhT", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Alberto Sinigaglia", "gender": "Male", "institution": "University of Padua", "country": "IT", "position": "PhD student" }, { "name": "Davide Sartor", "gender": "unknown", "institution": "University of Padua", "country": "IT", "position": "PhD student" }, { "name": "Gian Antonio Susto", "gender": "Male", "institution": "University of Padua", "country": "IT", "position": "Associate Professor" } ], "n_formula": 61, "n_formula_1": 30, "n_ref_uni": 28, "n_ref": 74, "n_ref_all": 109, "n_fig": 9, "n_tab": 4, "L_tab": 855, "n_element_tab": 101, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5158, "formula_len_all_1": 2071, "len_all": 158452, "len_all_1": 68693, "len_abs": 1307, "len_title": 111, "len_sents": 54574, "len_sents_1": 29813, "n_sents": 445, "n_sents_1": 234, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1284, "title": "Unbounded Activations for Constrained Monotonic Neural Networks", "abs": "Monotonic multi-layer perceptrons (MLPs) are crucial in applications requiring interpretable and trustworthy machine learning models, particularly in domains where decisions must adhere to specific input-output relationships. Traditional approaches that build monotonic MLPs with universal approximation guarantees often rely on constrained weights and bounded activation functions, which suffer from optimization issues. \nIn this work, we prove that non-negative constrained weights MLPs with activations that saturate on alternating sides are universal approximators for the class of monotonic functions. Thanks to this new result, we show that non-positive constrained weights MLPs with convex monotone activations, contrary to their non-negative constrained counterpart, are universal approximators. \nDespite such guarantees, we also show that such classes of MLPs are hard to optimize. Therefore, we propose a novel parametrization that eliminates the need for weight constraints, allowing the network to dynamically adjust activations based on weight signs, thus enhancing optimization stability and performance. \nExperiments demonstrate that our approach maintains theoretical guarantees and significantly outperforms existing monotonic architectures in approximation accuracy.", "keywords": [ "Monotonic", "Neural Networks" ], "rating_list": [ 8, 6, 6, 3 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "N18Z2MkMEa", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bin Lei", "gender": "unknown", "institution": "University of Connecticut", "country": "", "position": "PhD student" }, { "name": "Jianhui Wang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Lewei He", "gender": "Male", "institution": "South China Normal University", "country": "CN", "position": "Assistant Professor" }, { "name": "Qiuwu Chen", "gender": "Male", "institution": "AI", "country": "CN", "position": "CTO" }, { "name": "TIANYU SHI", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Yangfan He", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Intern" }, { "name": "Yuchen Li", "gender": "Male", "institution": "AIGCode Inc.", "country": "CN", "position": "Researcher" }, { "name": "\u674e\u6cfd\u8fdc", "gender": "unknown", "institution": "South China Normal University", "country": "CN", "position": "MS student" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 8, "n_ref": 13, "n_ref_all": 31, "n_fig": 5, "n_tab": 7, "L_tab": 2411, "n_element_tab": 147, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2047, "n_element_tab_1": 137, "formula_len_all": 1034, "formula_len_all_1": 1095, "len_all": 107405, "len_all_1": 53914, "len_abs": 1613, "len_title": 134, "len_sents": 30383, "len_sents_1": 23003, "n_sents": 236, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1619, "title": "FALCON: A Feedback-Driven Adaptive Long/Short-Term Memory Reinforced Coding Optimization", "abs": "Recently, large language models (LLMs) have achieved significant progress in automated code generation. Despite their strong instruction-following capabilities, these models frequently struggled to align with user intent in the coding scenario. In particular, they were hampered by datasets that lacked diversity and failed to address specialized tasks or edge cases. Furthermore, challenges in supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF) led to failures in generating precise, human-intent-aligned code. To tackle these challenges and improve the code generation performance for automated programming systems, we propose Feedback-driven Adaptive Long/short-term memory reinforced Coding OptimizatioN (i.e., FALCON). FALCON is structured into two hierarchical levels, from the global level, long-term memory improves code quality by retaining and applying learned knowledge, while from the local level, short-term memory allows for the incorporation of immediate feedback from compilers and AI systems. Additionally, we introduce meta-reinforcement learning with feedback rewards to solve the global-local bi-level optimization problem and enhance the model\u2019s adaptability across diverse code generation tasks. Extensive experiments are conducted and it is found that our technique achieves state-of-the-art performance, leading other reinforcement learning methods by more than 4.5 percentage points on the MBPP benchmark and 6.1 percentage points on the Humaneval benchmark. The open-sourced code is publicly available at https://anonymous.4open.science/r/FALCON-BFE0/README.md.", "keywords": [ "Automated agent generation framework Self-evolving agents LLM Manager Task-specific Role Builder" ], "rating_list": [ 3, 1, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 3, 1, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "N0vzm0vwyR", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Dat Du Thanh", "gender": "Male", "institution": "Ho Chi Minh City University of Technology", "country": "VN", "position": "MS student" }, { "name": "Mai Vu", "gender": "unknown", "institution": "Tufts University", "country": "US", "position": "Full Professor" }, { "name": "Nguyen Ho", "gender": "unknown", "institution": "Loyola University Maryland", "country": "US", "position": "Assistant Professor" }, { "name": "Tien Vu-Van", "gender": "unknown", "institution": "Ho Chi Minh City University of Technology", "country": "VN", "position": "Lecturer" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 40, "n_ref": 55, "n_ref_all": 85, "n_fig": 2, "n_tab": 4, "L_tab": 1517, "n_element_tab": 62, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 654, "n_element_tab_1": 7, "formula_len_all": 1099, "formula_len_all_1": 1001, "len_all": 112073, "len_all_1": 68486, "len_abs": 1337, "len_title": 126, "len_sents": 33446, "len_sents_1": 31066, "n_sents": 240, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1337, "title": "Pruning Deep Convolutional Neural Network Using Conditional Mutual Information", "abs": "Convolutional Neural Networks (CNNs) achieve high performance in image classification tasks but are challenging to deploy on resource-limited hardware due to their large model sizes. To address this issue, we leverage Mutual Information, a metric that provides valuable insights into how deep learning models retain and process information by measuring the shared information between input features or output labels and network layers. In this study, we propose a structured filter-pruning approach for CNNs that identifies and selectively retains the most informative features in each layer. Our approach successively evaluates each layer by ranking the importance of its feature maps based on Conditional Mutual Information (CMI) values, computed using a matrix-based R\u00e9nyi \u03b1-order entropy numerical method. We propose several formulations of CMI to capture correlation among features across different layers. We then develop various strategies to determine the cutoff point for CMI values to prune unimportant features. This approach allows parallel pruning in both forward and backward directions and significantly reduces model size while preserving accuracy. Tested on the VGG16 architecture with the CIFAR-10 dataset, the proposed method reduces the number of filters by more than a third, with only a 0.32% drop in test accuracy.", "keywords": [ "CNN", "pruning", "optimization", "mutual information", "Renyi entropy" ], "rating_list": [ 5, 1, 1 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 3, 1, 1 ], "contribution_list": [ 3, 2, 1 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "N0ts5QGEW9", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bo Jiang", "gender": "unknown", "institution": "Anhui University", "country": "CN", "position": "Full Professor" }, { "name": "Chuanfu Li", "gender": "Male", "institution": "Anhui University of Traditional Chinese Medicine", "country": "CN", "position": "Full Professor" }, { "name": "Fuling Wang", "gender": "Male", "institution": "Anhui University", "country": "CN", "position": "MS student" }, { "name": "Jin Tang", "gender": "unknown", "institution": "Anhui University", "country": "CN", "position": "Full Professor" }, { "name": "Qingchuan Ma", "gender": "Male", "institution": "Anhui University", "country": "CN", "position": "Undergrad student" }, { "name": "Shiao Wang", "gender": "Male", "institution": "Anhui University", "country": "CN", "position": "PhD student" }, { "name": "Wang Xiao", "gender": "Male", "institution": "Anhui University", "country": "CN", "position": "Associate Professor" }, { "name": "YueHang Li", "gender": "unknown", "institution": "Anhui University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 2, "n_ref": 19, "n_ref_all": 47, "n_fig": 3, "n_tab": 9, "L_tab": 2965, "n_element_tab": 421, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 3244, "n_element_tab_1": 421, "formula_len_all": 232, "formula_len_all_1": 206, "len_all": 140871, "len_all_1": 70746, "len_abs": 1565, "len_title": 154, "len_sents": 35503, "len_sents_1": 27654, "n_sents": 269, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 104, "L_abs": 1537, "title": "CXPMRG-Bench: Pre-training and Benchmarking for X-ray Medical Report Generation on CheXpert Plus Dataset", "abs": "X-ray image-based medical report generation (MRG) is a pivotal area in artificial intelligence which can significantly reduce diagnostic burdens and patient wait times. Despite significant progress, we believe that the task has reached a bottleneck due to the limited benchmark datasets and the existing large models' insufficient capability enhancements in this specialized domain. Specifically, the recently released CheXpert Plus dataset lacks comparative evaluation algorithms and their results, providing only the dataset itself. This situation makes the training, evaluation, and comparison of subsequent algorithms challenging. Thus, we conduct a comprehensive benchmarking of existing mainstream X-ray report generation models and large language models (LLMs), on the CheXpert Plus dataset. We believe that the proposed benchmark can provide a solid comparative basis for subsequent algorithms and serve as a guide for researchers to quickly grasp the state-of-the-art models in this field. More importantly, we propose a big model for the X-ray image report generation using a multi-stage pre-training strategy, including self-supervised autoregressive generation and Xray-report contrastive learning, and supervised fine-tuning. Extensive experimental results indicate that the autoregressive pre-training based on Mamba effectively encodes X-ray images, and the image-text contrastive pre-training further aligns the feature spaces, achieving better experimental results. All the source codes will be released upon acceptance.", "keywords": [ "Medical Report Generation; X-ray Image Pre-training; State Space Model; Benchmark" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "N0i0d27RTW", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Fang Xie", "gender": "unknown", "institution": "Beijing Normal University-Hong Kong Baptist University United International College", "country": "", "position": "Assistant Professor" }, { "name": "Johannes Lederer", "gender": "unknown", "institution": "Universit\u00e4t Hamburg", "country": "DE", "position": "Full Professor" }, { "name": "Mahsa Taheri", "gender": "Female", "institution": "Universit\u00e4t Hamburg", "country": "DE", "position": "Postdoc" } ], "n_formula": 227, "n_formula_1": 24, "n_ref_uni": 52, "n_ref": 109, "n_ref_all": 123, "n_fig": 3, "n_tab": 6, "L_tab": 661, "n_element_tab": 51, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 574, "n_element_tab_1": 29, "formula_len_all": 35037, "formula_len_all_1": 1959, "len_all": 292367, "len_all_1": 68265, "len_abs": 1022, "len_title": 132, "len_sents": 87003, "len_sents_1": 32016, "n_sents": 783, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1025, "title": "Statistical Guarantees for Approximate Stationary Points of Shallow Neural Networks", "abs": "Since statistical guarantees for neural networks are usually restricted to global optima of intricate objective functions, it is unclear whether these theories explain the performances of actual outputs of neural network pipelines. The goal of this paper is, therefore, to bring statistical theory closer to practice. We develop statistical guarantees for shallow linear neural networks that coincide up to logarithmic factors with the global optima but apply to stationary points and the points nearby. These results support the common notion that neural networks do not necessarily need to be optimized globally from a mathematical perspective. We then extend our statistical guarantees to shallow ReLU neural networks, assuming the first layer weight matrices are nearly identical for the stationary network and the target. More generally, despite being limited to shallow neural networks for now, our theories make an important step forward in describing the practical properties of neural networks in mathematical terms.", "keywords": [ "statistical guarantees", "shallow neural networks", "stationary points" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "N0U6OQRsNu", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 9, "author_info_list": [ { "name": "Amin Beheshti", "gender": "Male", "institution": "Macquarie University", "country": "AU", "position": "Full Professor" }, { "name": "Cheng Chen", "gender": "Male", "institution": "nanjing university", "country": "", "position": "PhD student" }, { "name": "Guanfeng Liu", "gender": "Male", "institution": "Macquarie University", "country": "AU", "position": "Associate Professor" }, { "name": "Lianyong Qi", "gender": "Male", "institution": "China University of Petroleum", "country": "CN", "position": "Full Professor" }, { "name": "Wanchun Dou", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Full Professor" }, { "name": "Xiaolong Xu", "gender": "unknown", "institution": "Nanjing University of Information Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xuyun Zhang", "gender": "Male", "institution": "Macquarie University", "country": "AU", "position": "Associate Professor" }, { "name": "Yongzhe Jia", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Yuankai Qi", "gender": "unknown", "institution": "Macquarie University", "country": "AU", "position": "Lecturer" } ], "n_formula": 14, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 135, "n_ref_all": 162, "n_fig": 12, "n_tab": 8, "L_tab": 4110, "n_element_tab": 245, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2028, "n_element_tab_1": 82, "formula_len_all": 482, "formula_len_all_1": 378, "len_all": 156061, "len_all_1": 59757, "len_abs": 1342, "len_title": 139, "len_sents": 41273, "len_sents_1": 27177, "n_sents": 277, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1350, "title": "ATTENDING: Federated Learning with Personalized Attentive Pruning for Heterogeneous Clients", "abs": "Federated Learning (FL) emerges as a novel machine learning paradigm, enabling distributed clients to collaboratively train a global model while eliminating local data transmission. Despite its advantages, FL faces challenges posed by system and data heterogeneity. System heterogeneity prevents low-end clients from participating in FL with uniform models, while data heterogeneity adversely impacts the learning performance of FL. In this paper, we propose the personalized ATTENtive pruning enabled federateD learnING (ATTENDING) to collectively address these heterogeneity challenges. Specifically, we first design an attention module incorporating spatial and channel attention to enhance the learning performance on heterogeneous data. Subsequently, we introduce the attentive pruning algorithm to generate personalized local models guided by attention scores, aiming to facilitate clients' participation in FL. Finally, we introduce a specific heterogeneous aggregation algorithm integrated with an attention matching mechanism to efficiently aggregate the pruned models. We implement ATTENDING with a real FL platform and the evaluation results show that ATTENDING significantly outperforms the baselines by up to 11.3\\% and reduces the average model footprints by 32\\%. Our code is available at: https://anonymous.4open.science/r/ATTENDING.", "keywords": [ "Federated Learning", "Attentive Pruning", "Heterogeneous Clients", "Non-IID Data." ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "N0MnPLK6r7", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Hogun Park", "gender": "unknown", "institution": "Sungkyunkwan University", "country": "KR", "position": "Associate Professor" }, { "name": "Hyunju Kang", "gender": "unknown", "institution": "Sungkyunkwan University", "country": "KR", "position": "PhD student" }, { "name": "Kyeongrok Park", "gender": "Male", "institution": "Sungkyunkwan University", "country": "", "position": "MS student" } ], "n_formula": 21, "n_formula_1": 16, "n_ref_uni": 47, "n_ref": 99, "n_ref_all": 123, "n_fig": 4, "n_tab": 15, "L_tab": 5138, "n_element_tab": 723, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2386, "n_element_tab_1": 423, "formula_len_all": 2276, "formula_len_all_1": 1261, "len_all": 178053, "len_all_1": 75745, "len_abs": 1573, "len_title": 119, "len_sents": 54493, "len_sents_1": 32150, "n_sents": 399, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1492, "title": "Toward Human-Interpretable Explanations in a Unified Framework for GNNs", "abs": "As Graph Neural Networks (GNNs) are increasingly applied across various domains, explainability has become a critical factor for real-world applications. Existing post-hoc explainability methods primarily focus on estimating the importance of edges, nodes, or subgraphs in the input graph to identify substructures crucial for predictions. However, these methods often lack human interpretability and do not provide a unified framework that incorporates both model-level and instance-level explanations. In this context, we propose leveraging a set of graphlets---small, connected, non-isomorphic induced subgraphs widely used in various scientific fields---and their associated orbits as human-interpretable units to decompose GNN predictions. Domain experts can select the most relevant graphlets as interpretable units and request unified explanations based on these units. To address this problem, we introduce UO-Explainer, the Unified and Orbit-based Explainer for GNNs, which utilizes predefined orbits that are generalizable and universal across graph domains as interpretable units. Our model decomposes GNN weights into orbit units to extract class-specific graph patterns (model-level) and to identify important subgraphs within individual data instances for prediction (instance-level). Extensive experimental results demonstrate that UO-Explainer outperforms existing baselines in providing meaningful and interpretable explanations across both synthetic and real-world datasets.", "keywords": [ "eXplainable AI", "Graph Neural Networks" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "N0ETIi580T", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Ananthram Swami", "gender": "unknown", "institution": "Army Research Labortory", "country": "US", "position": "ST" }, { "name": "Francesco Restuccia", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Assistant Professor" }, { "name": "Jonathan Ashdown", "gender": "Male", "institution": "Air Force Research Laboratory", "country": "US", "position": "Researcher" }, { "name": "Michael J. De Lucia", "gender": "Male", "institution": "University of Delaware", "country": "US", "position": "Lecturer" }, { "name": "Shahriar Rifat", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 23, "n_ref_uni": 34, "n_ref": 72, "n_ref_all": 92, "n_fig": 6, "n_tab": 7, "L_tab": 4903, "n_element_tab": 158, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2699, "n_element_tab_1": 63, "formula_len_all": 1250, "formula_len_all_1": 1531, "len_all": 121380, "len_all_1": 70129, "len_abs": 1845, "len_title": 108, "len_sents": 33534, "len_sents_1": 29203, "n_sents": 248, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1288, "title": "On the Adversarial Vulnerability of Label-Free Test-Time Adaptation", "abs": "Despite the success of Test-time adaptation (TTA), recent work has shown that adding relatively small adversarial perturbations to a limited number of samples leads to significant performance degradation. Therefore, it is crucial to rigorously evaluate existing TTA algorithms against relevant threats and implement appropriate security countermeasures. Importantly, existing threat models assume test-time samples will be labeled, which is impractical in real-world scenarios. To address this gap, we propose a new attack algorithm that does not rely on\naccess to labeled test samples, thus providing a concrete way to assess the security vulnerabilities of TTA algorithms. Our attack design is grounded in theoretical foundations and can generate strong attacks against different state of the art TTA methods. In addition, we show that existing defense mechanisms are almost ineffective, which emphasizes the need for further research on TTA security. Through extensive experiments on CIFAR10-C, CIFAR100-C, and ImageNet-C, we demonstrate that our proposed approach closely matches the performance of state-of-the-art attack benchmarks, even without access to labeled samples. In certain cases, our approach generates stronger attacks, e.g., more than 4% higher error rate on CIFAR10-C.", "keywords": [ "Test Time Adaptation", "Adversarial Attack" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 3, 3, 2 ] }, { "paper_id": "Mzz9i4Zf8B", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Enis Simsar", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Federico Tombari", "gender": "Male", "institution": "Google", "country": "CH", "position": "Research Director" }, { "name": "Pinar Yanardag", "gender": "Female", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "Assistant Professor" }, { "name": "Tuna Han Salih Meral", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 4, "n_ref_uni": 47, "n_ref": 99, "n_ref_all": 151, "n_fig": 14, "n_tab": 2, "L_tab": 802, "n_element_tab": 34, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1705, "n_element_tab_1": 42, "formula_len_all": 142, "formula_len_all_1": 201, "len_all": 165425, "len_all_1": 67723, "len_abs": 781, "len_title": 109, "len_sents": 50596, "len_sents_1": 32921, "n_sents": 391, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1408, "title": "CLoRA: A Contrastive Approach to Compose Multiple LoRA Models", "abs": "Low-Rank Adaptation (LoRA) has emerged as a powerful and popular technique for personalization, enabling efficient adaptation of pre-trained image generation models for specific tasks without comprehensive retraining. While employing individual pre-trained LoRA models excels at representing single concepts, such as those representing a specific dog or a cat, utilizing multiple LoRA models to capture a variety of concepts in a single image still poses a significant challenge. Existing methods often fall short, primarily because the attention mechanisms within different LoRA models overlap, leading to scenarios where one concept may be completely ignored (e.g., omitting the dog) or where concepts are incorrectly combined (e.g., producing an image of two cats instead of one cat and one dog). We introduce CloRA, a training-free approach that addresses these limitations by updating the attention maps of multiple LoRA models at test-time, and leveraging the attention maps to create semantic masks for fusing latent representations. This enables the generation of composite images that accurately reflect the characteristics of each LoRA. Our comprehensive qualitative and quantitative evaluations demonstrate that CloRA significantly outperforms existing methods in multi-concept image generation using LoRAs. Furthermore, we share our source code and benchmark dataset to promote further research.", "keywords": [ "Text to Image Generation", "Personalization", "LoRAs", "Contrastive Learning", "Generative Models", "Low Rank Adaptation" ], "rating_list": [ 6, 5, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 5, 5, 4 ] }, { "paper_id": "MzmeLlDOkN", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Chaozheng Wang", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Yinghao Wu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Zhaojian Yu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 24, "n_ref": 44, "n_ref_all": 55, "n_fig": 6, "n_tab": 4, "L_tab": 4791, "n_element_tab": 326, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 4620, "n_element_tab_1": 324, "formula_len_all": 871, "formula_len_all_1": 871, "len_all": 89874, "len_all_1": 55897, "len_abs": 1508, "len_title": 124, "len_sents": 24279, "len_sents_1": 21482, "n_sents": 151, "n_sents_1": 134, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1515, "title": "MIGA: Mixture-of-Experts with Group Aggregation for Stock Market Prediction", "abs": "Stock market prediction has remained an extremely challenging problem for many\ndecades owing to its inherent high volatility and low information noisy ratio.\nExisting solutions based on machine learning or deep learning demonstrate superior\nperformance by employing a single model trained on the entire stock dataset to\ngenerate predictions across all types of stocks. However, due to the significant\nvariations in stock styles and market trends, a single end-to-end model struggles to\nfully capture the differences in these stylized stock features, leading to relatively\ninaccurate predictions for all types of stocks. In this paper, we present MIGA, a\nnovel Mixture of Expert with Group Aggregation framework designed to generate\nspecialized predictions for stocks with different styles of by dynamically switching\nbetween distinct style experts. To promote collaboration among different experts\nin MIGA, we propose a novel inner group attention architecture, enabling experts\nwithin the same group to share information and thereby enhancing the overall\nperformance of all experts. As a result, MIGA significantly outperforms other\nend-to-end models on three Chinese Stock Index benchmarks including CSI300,\nCSI500 and CSI1000. Notably, MIGA-Conv reaches 24 % excess annual return on\nCSI300 benchmark, surpassing the previous state-of-the-art model by 8% absolute.\nFurthermore, we conduct a comprehensive analysis of mixture of experts for stock\nmarket prediction, providing valuable insights for future research.", "keywords": [ "mixture-of-expert", "stock market prediction" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 3, 1, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "MzHNftnAM1", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Benjamin Feuer", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "John P Dickerson", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Associate Professor" }, { "name": "Max Cembalest", "gender": "Male", "institution": "Arthur", "country": "", "position": "Researcher" }, { "name": "Micah Goldblum", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" }, { "name": "Raz Besaleli", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "MS student" }, { "name": "Samuel Dooley", "gender": "unknown", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Sanjana Nambiar", "gender": "Female", "institution": "New York University", "country": "US", "position": "Researcher" }, { "name": "Teresa Datta", "gender": "Female", "institution": "Arthur AI", "country": "", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 58, "n_ref_all": 92, "n_fig": 4, "n_tab": 9, "L_tab": 5536, "n_element_tab": 290, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1060, "n_element_tab_1": 98, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 282324, "len_all_1": 59894, "len_abs": 1131, "len_title": 128, "len_sents": 56818, "len_sents_1": 29584, "n_sents": 456, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1024, "title": "Style Outweighs Substance: Failure Modes of LLM Judges in Alignment Benchmarking", "abs": "The release of ChatGPT in November 2022 sparked an explosion of interest in post-training and an avalanche of new preference optimization (PO) methods. These methods claim superior alignment by virtue of better correspondence with human pairwise preferences, often measured by LLM-judges. In this work, we attempt to answer the following question -- do LLM-judge preferences translate to progress on other, more concrete metrics for alignment, and if not, why not? We define a concrete metric for alignment, and introduce SOS-Bench (Substance Outweighs Style Benchmark), the largest standardized, reproducible LLM meta-benchmark to date. We find that (1) LLM-judge preferences do not correlate with concrete measures of safety, world knowledge, and instruction following; (2) LLM-judges have powerful implicit biases, prioritizing style over factuality and safety; and (3) the supervised fine-tuning (SFT) stage of post-training has a large impact on alignment, with data scaling and prompt diversity as the driving factors.", "keywords": [ "LLM", "large language model", "alignment", "post-training", "benchmarking", "evaluation" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "MyotJECv0D", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Lin Wang", "gender": "Female", "institution": "Xianda College of Economics and Humanities Shanghai International Studies University", "country": "", "position": "Associate Professor" }, { "name": "Wuying Liu", "gender": "Male", "institution": "Ludong University", "country": "", "position": "Full Professor" }, { "name": "Yuan Wang", "gender": "unknown", "institution": "Ludong University", "country": "CN", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 11, "n_ref": 15, "n_ref_all": 35, "n_fig": 6, "n_tab": 4, "L_tab": 3000, "n_element_tab": 472, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 3006, "n_element_tab_1": 472, "formula_len_all": 195, "formula_len_all_1": 195, "len_all": 81415, "len_all_1": 67843, "len_abs": 1954, "len_title": 66, "len_sents": 31379, "len_sents_1": 29284, "n_sents": 185, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1959, "title": "Correlation Analysis of Evaluation Metrics for Machine Translation", "abs": "Machine translation evaluation methods can be roughly divided into three categories: manual evaluation, classical morphological evaluation and semantic evaluation based on pre-trained model. The automatic evaluation metrics of the latter two categories are numerous, from which we select commonly used seven morphological evaluation metrics and four semantic evaluation metrics for correlation analysis between each two of them. The experimental results of the correlation coefficients of Pearson, Kendall and Spearman on 40 machine translation models of bidirectional 20 foreign languages and Chinese show that: (1) There is an extremely strong correlation among morphological evaluation metrics, indicating that the statistical results of various morphological calculation methods tend to be the same on big data. (2) There is a strong correlation between semantic evaluation metrics, indicating that although there are semantic spatial differences among various pre-trained models, the statistical results on big data also tend to be consistent. The above-mentioned ubiquitous correlations largely stem from the equivalence of human cognition and the economy of knowledge representation. (3) There is also a strong correlation between morphological and semantic evaluation metrics, which shows that the deep \u201csemantics\u201d of various commercial hypes at present is just another high-level \u201cmorphology\u201d. Because the Turing computing system can use symbols and operations to directly represent and accurately process morphologies, but can only simulately represent and approximately process semantics using symbols and operations. (4) For each correlation coefficient between any two evaluation metrics, there is a significant difference between different languages, which indicates that morphology and semantics are inherent attributes of languages, and more optimized evaluation metrics of machine translation should be personalized according to the language.", "keywords": [ "Correlation Analysis", "Morphological Evaluation Metrics", "Semantic Evaluation Metrics", "Machine Translation" ], "rating_list": [ 5, 1, 3, 1 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "MyVC4X5B2X", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "AJAY KUMAR JAISWAL", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Abhra Chaudhuri", "gender": "Male", "institution": "Fujitsu Research of Europe", "country": "GB", "position": "Senior Researcher" }, { "name": "Lu Yin", "gender": "unknown", "institution": "University of Surrey", "country": "GB", "position": "Assistant Professor" }, { "name": "Xiatian Zhu", "gender": "unknown", "institution": "University of Surrey", "country": "GB", "position": "Associate Professor" }, { "name": "Yunpeng Li", "gender": "unknown", "institution": "King's College London, University of London", "country": "GB", "position": "Reader" }, { "name": "Ziquan Liu", "gender": "Male", "institution": "Queen Mary, University of London", "country": "GB", "position": "Lecturer" }, { "name": "Adarsh K", "gender": "Male", "institution": "University of Surrey", "country": "", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 22, "n_ref_uni": 24, "n_ref": 68, "n_ref_all": 89, "n_fig": 6, "n_tab": 5, "L_tab": 1733, "n_element_tab": 156, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 769, "n_element_tab_1": 53, "formula_len_all": 1490, "formula_len_all_1": 1495, "len_all": 129866, "len_all_1": 71045, "len_abs": 1477, "len_title": 97, "len_sents": 49575, "len_sents_1": 35207, "n_sents": 332, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1483, "title": "SEBRA : Debiasing through Self-Guided Bias Ranking", "abs": "Ranking samples by fine-grained estimates of spuriosity (the degree to which spurious cues are present) has recently been shown to significantly benefit bias mitigation, over the traditional binary biased-vs-unbiased partitioning of train sets. However, this spuriousity ranking comes with the requirement of human supervision. In this paper, we propose a debiasing framework based on our novel Self-Guided Bias Ranking (Sebra), that mitigates biases via an automatic ranking of data points by spuriosity within their respective classes. Sebra leverages a key local symmetry in Empirical Risk Minimization (ERM) training -- the ease of learning a sample via ERM inversely correlates with its spuriousity; the fewer spurious correlations a sample exhibits, the harder it is to learn, and vice versa. However, globally across iterations, ERM tends to deviate from this symmetry. Sebra dynamically steers ERM to correct this deviation, facilitating the sequential learning of attributes in increasing order of difficulty, ie, decreasing order of spuriosity. As a result, the sequence in which Sebra learns samples naturally provides spuriousity rankings. We use the resulting fine-grained bias characterization in a contrastive learning framework to mitigate biases from multiple sources. Extensive experiments show that Sebra consistently outperforms previous state-of-the-art unsupervised debiasing techniques across multiple standard benchmarks, including UrbanCars, BAR, and CelebA.", "keywords": [ "Sub-population shift", "Spurious Correlations", "Bias Mitigation", "Fairness" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 2, 3, 3, 2 ] }, { "paper_id": "MyRcW53CCC", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Cen Chen", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Junyao Yang", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "ZhengdongLu", "gender": "Male", "institution": "South China University of Technology", "country": "", "position": "MS student" }, { "name": "Ziqian Zeng", "gender": "Female", "institution": "South China University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "HUIPING ZHUANG", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Jianwei Wang", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 26, "n_formula_1": 8, "n_ref_uni": 29, "n_ref": 82, "n_ref_all": 106, "n_fig": 6, "n_tab": 8, "L_tab": 2438, "n_element_tab": 238, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 960, "n_element_tab_1": 36, "formula_len_all": 1843, "formula_len_all_1": 376, "len_all": 219851, "len_all_1": 64587, "len_abs": 1370, "len_title": 153, "len_sents": 70638, "len_sents_1": 30705, "n_sents": 605, "n_sents_1": 238, "venue": "ICLR2025", "decision": "withdraw", "L_title": 105, "L_abs": 1375, "title": "PrivacyRestore: Privacy-Preserving Inference in Large Language Models via Privacy Removal and Restoration", "abs": "The widespread usage of online Large Language Models (LLMs) inference services has raised significant privacy concerns about the potential exposure of private information in user inputs to malicious eavesdroppers. Existing privacy protection methods for LLMs suffer from either insufficient privacy protection, performance degradation, or large inference time overhead. To address these limitations, we propose PrivacyRestore, a plug-and-play method to protect the privacy of user inputs during LLM inference. The server first trains restoration vectors for each privacy span and then release to clients. Privacy span is defined as a contiguous sequence of tokens within a text that contain private information. The client then aggregate restoration vectors of all privacy spans in the input into a single meta restoration vector which is later sent to the server side along with the input without privacy spans.The private information is restored via activation steering during inference. Furthermore, we prove that PrivacyRestore inherently prevents the linear growth of the privacy budget.We create three datasets, covering medical and legal domains, to evaluate the effectiveness of privacy preserving methods. The experimental results show that PrivacyRestore effectively protects private information and maintain acceptable levels of performance and inference overhead.", "keywords": [ "Privacy Protection", "Activation Steering", "LLM Inference" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "MyMrDTiFdk", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 1, "author_info_list": [ { "name": "Antti Koskela", "gender": "Male", "institution": "Nokia Bell Labs", "country": "", "position": "Researcher" } ], "n_formula": 51, "n_formula_1": 30, "n_ref_uni": 41, "n_ref": 99, "n_ref_all": 113, "n_fig": 8, "n_tab": 3, "L_tab": 338, "n_element_tab": 32, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 242, "n_element_tab_1": 29, "formula_len_all": 3636, "formula_len_all_1": 1849, "len_all": 151477, "len_all_1": 72138, "len_abs": 1422, "len_title": 135, "len_sents": 43459, "len_sents_1": 29111, "n_sents": 358, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1427, "title": "Approximating Two-Layer ReLU Networks for Hidden State Analysis in Differential Privacy", "abs": "The hidden state threat model of differential privacy (DP) assumes that the adversary has access only to the final trained machine learning (ML) model, without seeing intermediate states during training. Current privacy analyses under this model, however, are limited to convex optimization problems, reducing their applicability to multi-layer neural networks, which are essential in modern deep learning applications. Additionally, the most successful applications of the hidden state privacy analyses in classification tasks have been for logistic regression models. We demonstrate that it is possible to privately train convex problems with privacy-utility trade-offs comparable to those of one hidden-layer ReLU networks trained with DP stochastic gradient descent (DP-SGD). We achieve this through a stochastic approximation of a dual formulation of the ReLU minimization problem which results in a strongly convex problem. This enables the use of existing hidden state privacy analyses, providing accurate privacy bounds also for the noisy cyclic mini-batch gradient descent (NoisyCGD) method with fixed disjoint mini-batches. Our experiments on benchmark classification tasks show that NoisyCGD can achieve privacy-utility trade-offs comparable to DP-SGD applied to one-hidden-layer ReLU networks. Additionally, we provide theoretical utility bounds that highlight the speed-ups gained through the convex approximation.", "keywords": [ "Differential privacy", "machine learning", "hidden state threat model", "privacy amplification by iteration", "DP-SGD" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "MyAqAYCjP5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Dou Hoon Kwark", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Guowei Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Haiyang Xu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Haowen Wang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Xiang Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Zeyuan Chen", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Zhuowen Tu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 57, "n_ref": 90, "n_ref_all": 123, "n_fig": 16, "n_tab": 8, "L_tab": 4048, "n_element_tab": 804, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 947, "n_element_tab_1": 113, "formula_len_all": 149, "formula_len_all_1": 76, "len_all": 173957, "len_all_1": 65412, "len_abs": 1767, "len_title": 146, "len_sents": 44604, "len_sents_1": 30666, "n_sents": 326, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 96, "L_abs": 1376, "title": "Mousterian: exploring the equivalence of generative and real data augmentation in classification", "abs": "In this paper, we address a key question in machine learning: **How effectively can generative data augmentation enhance image classification?** We begin by examining the differences and similarities between real and synthetic data generated by advanced text-to-image models. Through comprehensive experiments, we provide systematic insights into leveraging synthetic data for improved classification performance. Our findings show that: 1). Generative data augmentation by models trained solely on the internal (available training) set can effectively improve classification performance, validating the long-held hypothesis that synthesis enhances analysis by enriching modeling capability.\n2). For generative data augmentation by models trained on both internal and external data (e.g. large-scale image-text pairs) separately, the size of equivalent synthetic dataset augmentation can be determined empirically. In addition to being aligned with a common intuition that real data augmentation is always preferred, our empirical formulation also provides a guideline for quantitatively estimating how much larger the size of generative dataset augmentation is, over the real data augmentation, to achieve comparable improvements. Our CIFAR-10 and ImageNet results also demonstrate its impact w.r.t. the size of the baseline training set and the quality of generative models.", "keywords": [ "analysis by synthesis", "image classification", "data generation" ], "rating_list": [ 3, 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 3, 3, 2 ], "presentation_list": [ 3, 1, 2, 3, 2 ], "contribution_list": [ 2, 1, 1, 1, 2 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "My9MBsO41H", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Hua Yang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Jihao Dong", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "MS student" }, { "name": "Renjie Pan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" } ], "n_formula": 46, "n_formula_1": 17, "n_ref_uni": 50, "n_ref": 133, "n_ref_all": 182, "n_fig": 6, "n_tab": 13, "L_tab": 5012, "n_element_tab": 971, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1433, "n_element_tab_1": 119, "formula_len_all": 2757, "formula_len_all_1": 922, "len_all": 220686, "len_all_1": 67066, "len_abs": 1225, "len_title": 134, "len_sents": 53696, "len_sents_1": 29606, "n_sents": 417, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1228, "title": "Discovering Clone Negatives via Adaptive Contrastive Learning for Image-Text Matching", "abs": "In this paper, we identify a common yet challenging issue in image-text matching, i.e., clone negatives: negative image-text pairs that are semantic-consistent with the positives, which leads to ambiguous and suboptimal matching results. To address this, we propose Adaptive Contrastive Learning (AdaCL), which introduces two margin parameters with a modulating anchor to dynamically strengthen the compactness between positives and mitigating the impact of clone negatives. The modulating anchor is selected based on the distribution of negative samples without explicit training, allowing for progressive tuning and enhanced in-batch supervision. Extensive experiments on image-text matching, noisy correspondence learning, CLIP pre-training, and text-based person search demonstrate the superiority of AdaCL in image-text matching. Furthermore, we extend AdaCL to weakly-supervised image-text matching by substituting human-annotated descriptions with automatically generated captions, increasing the number of potential clone negatives. AdaCL demonstrates robustness with the generated captions, alleviating the reliance on crowd-sourced annotations and laying a foundation for scalable vision-language contrastive learning.", "keywords": [ "Image-text matching", "Contrastive learning", "Vision and language", "Multimodal learning" ], "rating_list": [ 8, 6, 8, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 2, 5, 5 ] }, { "paper_id": "MxbEiFRf39", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 20, "author_info_list": [ { "name": "Aaron Mueller", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Postdoc" }, { "name": "Adam Belfki", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Researcher" }, { "name": "Alexander Russell Loftus", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Arjun Guha", "gender": "Male", "institution": "Roblox Research", "country": "", "position": "Visiting Professor" }, { "name": "Arnab Sen Sharma", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Byron C Wallace", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "Assistant Professor" }, { "name": "Caden Juang", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "Undergrad student" }, { "name": "Can Rager", "gender": "unknown", "institution": "Universit\u00e4t Hamburg", "country": "DE", "position": "MS student" }, { "name": "Carla E. Brodley", "gender": "Female", "institution": "Northeastern University", "country": "US", "position": "Full Professor" }, { "name": "David Bau", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Assistant Professor" }, { "name": "Dmitrii Troitskii", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "MS student" }, { "name": "Eric Todd", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Francesca Lucchetti", "gender": "unknown", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Jaden Fried Fiotto-Kaufman", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Researcher" }, { "name": "Jannik Brinkmann", "gender": "Male", "institution": "New York University", "country": "US", "position": "Researcher" }, { "name": "Jonathan Bell", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Associate Professor" }, { "name": "Koyena Pal", "gender": "Female", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Michael Ripa", "gender": "unknown", "institution": "Carleton University", "country": "CA", "position": "Undergrad student" }, { "name": "Nikhil Prakash", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Samuel Marks", "gender": "unknown", "institution": "Northeastern University", "country": "", "position": "Postdoc" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 56, "n_ref": 81, "n_ref_all": 106, "n_fig": 13, "n_tab": 5, "L_tab": 360, "n_element_tab": 19, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 464, "n_element_tab_1": 17, "formula_len_all": 159, "formula_len_all_1": 159, "len_all": 357691, "len_all_1": 60540, "len_abs": 1288, "len_title": 128, "len_sents": 64268, "len_sents_1": 29161, "n_sents": 567, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1145, "title": "NNsight and NDIF: Democratizing Access to Foundation Model Internals", "abs": "We introduce NNsight and NDIF, technologies that work in tandem to enable scientific study of the representations and computations learned by very large neural networks. NNsight is an open-source system that extends PyTorch to introduce deferred remote execution. NDIF is a scalable inference service that executes NNsight requests, allowing users to share GPU resources and pretrained models. These technologies are enabled by the intervention graph, an architecture developed to decouple experiment design from model runtime. Together, this framework provides transparent and efficient access to the internals of foundation-size deep neural networks without imposing the cost or complexity of hosting customized models individually. We conduct a quantitative survey of the machine learning literature that reveals a growing gap in the study of the internals of large-scale AI. We demonstrate the design and use of our framework to address this gap by enabling a range of research methods on huge models. Finally, we conduct benchmarks to compare performance with previous approaches.\n\nCode and documentation will be made available open-source.", "keywords": [ "interpretability", "safety", "large language models", "distributed inference", "scalable infrastructure", "deferred execution", "computation graphs", "resource sharing" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "MxHgnYbxly", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Lahav Dabah", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" }, { "name": "Tom Tirer", "gender": "unknown", "institution": "Bar-Ilan University", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 86, "n_formula_1": 6, "n_ref_uni": 38, "n_ref": 78, "n_ref_all": 120, "n_fig": 9, "n_tab": 10, "L_tab": 4641, "n_element_tab": 813, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 935, "n_element_tab_1": 196, "formula_len_all": 10517, "formula_len_all_1": 256, "len_all": 201984, "len_all_1": 75908, "len_abs": 1430, "len_title": 116, "len_sents": 58912, "len_sents_1": 32985, "n_sents": 554, "n_sents_1": 270, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1445, "title": "On Temperature Scaling and Conformal Prediction of Deep Classifiers", "abs": "In many classification applications, the prediction of a deep neural network (DNN) based classifier needs to be accompanied by some confidence indication. Two popular approaches for that aim are: 1) *Calibration*: modifies the classifier's softmax values such that the maximal value better estimates the correctness probability; and 2) *Conformal Prediction* (CP): produces a prediction set of candidate labels that contains the true label with a user-specified probability, guaranteeing marginal coverage but not, e.g., per class coverage. In practice, both types of indications are desirable, yet, so far the interplay between them has not been investigated. \nFocusing on the ubiquitous *Temperature Scaling* (TS) calibration, we start this paper with an extensive empirical study of its effect on prominent CP methods. We show that while TS calibration improves the class-conditional coverage of adaptive CP methods, surprisingly, it negatively affects their prediction set sizes. Motivated by this behavior, we explore the effect of TS on CP *beyond its calibration application* and reveal an intriguing trend under which it allows to trade prediction set size and conditional coverage of adaptive CP methods. Then, we establish a mathematical theory that explains the entire non-monotonic trend.\nFinally, based on our experiments and theory, we offer simple guidelines for practitioners to effectively combine adaptive CP with calibration.", "keywords": [ "classification", "temperature scaling", "conformal prediction", "conditional coverage", "prediction sets" ], "rating_list": [ 6, 3, 8 ], "soundness_list": [ 3, 2, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 5, 5 ] }, { "paper_id": "MxGGdhDmv5", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Guillaume Rabusseau", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Associate Professor" }, { "name": "Reihaneh Rabbany", "gender": "Female", "institution": "McGill University", "country": "", "position": "Assistant Professor" }, { "name": "Soroush Omranpour", "gender": "Male", "institution": ", McGill University", "country": "CA", "position": "MS student" } ], "n_formula": 37, "n_formula_1": 22, "n_ref_uni": 32, "n_ref": 54, "n_ref_all": 83, "n_fig": 4, "n_tab": 10, "L_tab": 4762, "n_element_tab": 406, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 4951, "n_element_tab_1": 253, "formula_len_all": 3021, "formula_len_all_1": 1828, "len_all": 146570, "len_all_1": 66548, "len_abs": 1246, "len_title": 132, "len_sents": 36873, "len_sents_1": 24865, "n_sents": 288, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1200, "title": "Higher Order Transformers: Efficient Attention Mechanism for Tensor Structured Data", "abs": "Transformers are now ubiquitous for sequence modeling tasks, but their extension to multi-dimensional data remains a challenge due to the quadratic cost of the attention mechanism. In this paper, we propose Higher-Order Transformers (HOT), a novel architecture designed to efficiently process data with more than two axes, i.e. higher-order tensors. \nTo address the computational challenges associated with high-order tensor attention, we introduce a novel Kronecker factorized attention mechanism that reduces the attention cost to quadratic in each axis' dimension, rather than quadratic in the total size of the input tensor. To further enhance efficiency, HOT leverages kernelized attention, reducing the complexity to linear. This strategy maintains the model's expressiveness while enabling scalable attention computation.\nWe validate the effectiveness of HOT on two high-dimensional tasks, including multivariate time series forecasting, and 3D medical image classification. Experimental results demonstrate that HOT achieves competitive performance while significantly improving computational efficiency, showcasing its potential for tackling a wide range of complex, multi-dimensional data.", "keywords": [ "transformers", "multihead attention", "high order tensor", "kronecker decomposition", "multivariate timeseries forecasting", "3D medical image classification" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 4, 5, 2, 2 ] }, { "paper_id": "MxALfOAnXv", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Aiqing Zhu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Qianxiao Li", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Yuting Pan", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "PhD student" } ], "n_formula": 55, "n_formula_1": 31, "n_ref_uni": 44, "n_ref": 82, "n_ref_all": 124, "n_fig": 15, "n_tab": 1, "L_tab": 74, "n_element_tab": 10, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 74, "n_element_tab_1": 10, "formula_len_all": 6996, "formula_len_all_1": 2941, "len_all": 160113, "len_all_1": 66512, "len_abs": 1164, "len_title": 156, "len_sents": 46612, "len_sents_1": 28720, "n_sents": 443, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 109, "L_abs": 1170, "title": "Continuity-Preserving Convolutional Autoencoders for Learning Continuous Latent Dynamical Models from Images", "abs": "Continuous dynamical systems are cornerstones of many scientific and engineering disciplines.\nWhile machine learning offers powerful tools to model these systems from trajectory data, challenges arise when these trajectories are captured as images, resulting in pixel-level observations that are discrete in nature.\nConsequently, a naive application of a convolutional autoencoder can result in latent coordinates that are discontinuous in time.\nTo resolve this, we propose continuity-preserving convolutional autoencoders (CpAEs) to learn continuous latent states and their corresponding continuous latent dynamical models from discrete image frames. \nWe present a mathematical formulation for learning dynamics from image frames, which illustrates issues with previous approaches and motivates our methodology based on promoting the continuity of convolution filters, thereby preserving the continuity of the latent states.\nThis approach enables CpAEs to produce latent states that evolve continuously with the underlying dynamics, leading to more accurate latent dynamical models.\nExtensive experiments across various scenarios demonstrate the effectiveness of CpAEs.", "keywords": [ "Latent dynamical system", "Autoencoders", "Learning dynamics", "Continuity-preserving" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "Mx22pSSo1b", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Kwanghyun Ko", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Myungjoo Kang", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Sangjun Han", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Cheong Woojin", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 39, "n_ref": 59, "n_ref_all": 91, "n_fig": 12, "n_tab": 7, "L_tab": 4263, "n_element_tab": 508, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 2272, "n_element_tab_1": 293, "formula_len_all": 731, "formula_len_all_1": 684, "len_all": 140254, "len_all_1": 57722, "len_abs": 1117, "len_title": 122, "len_sents": 44990, "len_sents_1": 24299, "n_sents": 308, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1121, "title": "Learning Rotation-Invariant Representation using Rotation-Equivariant CNNs", "abs": "Conventional self-supervised learning (SSL) methods, such as SimCLR and SimSiam, have demonstrated significant effectiveness. However, their feature representation is not robust to image rotations, as rotational augmentation may negatively impact the framework. In this paper, we address this limitation by applying SSL to group-equivariant CNNs, specifically rotation-equivariant CNNs, to develop robust features. To learn expressive, rotation-invariant features, we introduce our training method, Guiding Invariance with Equivariance (GIE), which simultaneously trains both invariant features and the equivariance score for images. The equivariance score guides the rotation-equivariant features through an attention-weighted sum mechanism, enabling the development of rotation-invariant features. Through experiments, we demonstrate that our GIE method not only extracts high-performing features under four discrete rotations but also achieves robustness to random-degree rotations through rotation augmentation training. These results highlight the effectiveness of our method in achieving robust rotation-invariance.", "keywords": [ "Self-supervised Learning", "Contrastive Learning", "Equivariance", "Rotation-Invariance", "Guiding Invariance with Equivariance" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "MwU2SGLKpS", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Alon Albalak", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Chase Blagden", "gender": "Male", "institution": "Synthlabs AI", "country": "US", "position": "Researcher" }, { "name": "Chelsea Finn", "gender": "Female", "institution": "Physical Intelligence", "country": "US", "position": "Researcher" }, { "name": "Dakota Mahan", "gender": "unknown", "institution": "SynthLabs", "country": "US", "position": "Researcher" }, { "name": "Duy Van Phung", "gender": "Male", "institution": "Ho Chi Minh city University of Science, Vietnam National University", "country": "", "position": "Undergrad student" }, { "name": "Jan-Philipp Fr\u00e4nken", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Postdoc" }, { "name": "Louis Castricato", "gender": "Male", "institution": "Brown University", "country": "", "position": "PhD student" }, { "name": "Nathan Lile", "gender": "Male", "institution": "SynthLabs.ai", "country": "US", "position": "Researcher" }, { "name": "Rafael Rafailov", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 33, "n_ref": 68, "n_ref_all": 83, "n_fig": 8, "n_tab": 3, "L_tab": 589, "n_element_tab": 80, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 217, "n_element_tab_1": 43, "formula_len_all": 950, "formula_len_all_1": 949, "len_all": 200597, "len_all_1": 56896, "len_abs": 1448, "len_title": 72, "len_sents": 37455, "len_sents_1": 26887, "n_sents": 273, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 24, "L_abs": 1437, "title": "Generative Reward Models", "abs": "Reinforcement Learning from Human Feedback (RLHF) has greatly improved the performance of modern Large Language Models (LLMs). The RLHF process is resource-intensive and technically challenging, generally requiring a large collection of human preference labels over model-generated outputs. Reinforcement Learning from AI Feedback (RLAIF) addresses this data collection challenge by leveraging synthetic preferences generated by an LLM. However, recent work has shown that synthetic preferences labels may not align well with human preference judgments.\n\nTo address this, we propose a hybrid approach that unifies RLHF and RLAIF methodologies. We introduce GenRM, an iterative algorithm that trains an LLM on self-generated reasoning traces, leading to synthetic preference labels matching human preference judgments. Empirically, we show that zero-shot LLM-based judgments under-perform compared to Bradley-Terry reward models on in-distribution tasks (between 9-36%). In contrast, GenRM achieves in-distribution accuracy comparable to Bradley-Terry models, while significantly outperforming them on out-of-distribution tasks (between 10-45%). Moreover, GenRM surpasses the performance of using LLMs as judges on both in-distribution (by 9-31%) and out-of-distribution tasks (by 2- 6%). Our results show that combining the strengths of RLHF and RLAIF offers a promising approach for improving the quality of synthetic preference labels.", "keywords": [ "RLHF", "reward model", "LLM Judge", "reasoning" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "MwMoE1y0Nb", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Deepayan Sanyal", "gender": "Male", "institution": "Vanderbilt University", "country": "", "position": "PhD student" }, { "name": "Joel Phillips Michelson", "gender": "Male", "institution": "Vanderbilt University", "country": "", "position": "PhD student" }, { "name": "Maithilee Kunda", "gender": "Female", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Reader" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 42, "n_ref": 61, "n_ref_all": 69, "n_fig": 2, "n_tab": 3, "L_tab": 1026, "n_element_tab": 77, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 407, "n_element_tab_1": 14, "formula_len_all": 321, "formula_len_all_1": 288, "len_all": 105958, "len_all_1": 55233, "len_abs": 1187, "len_title": 159, "len_sents": 28583, "len_sents_1": 26034, "n_sents": 218, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1192, "title": "Understanding Distribution Alignment Through Category Separability In An Infant-Inspired Domain Adaptation Task", "abs": "We introduce a novel distribution shift, called the VI-Shift, that mimics the trade-off between object instances and viewpoints in the visual experience of infants. Motivated by findings in infant learning literature, we study this problem through the lens of domain adaptation, but without ImageNet pretraining. We show that the performances of two classic domain adaptation methods, Joint Adaptation Network (JAN) and Domain Adversarial Neural Networks (DANN), deteriorate without ImageNet pretraining. We hypothesize that the separability of source and target category clusters in the feature space plays a crucial role in the effectiveness of JAN. So, we propose 3 metrics to measure category separability and demonstrate that target separability in the pretrained network is strongly correlated with downstream JAN and DANN accuracy. Further, we propose two novel loss functions that increase target separability during pretraining by aligning the distribution of within-domain pairwise distances between the source and target distributions. Our experiments show that the application of these loss functions modestly improves downstream accuracy on unseen images from the target dataset.", "keywords": [ "domain adaptation", "distribution shift", "infant learning", "self-supervised learning" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "MwIbzfu93a", "primary_area": "learning on time series and dynamical systems", "n_author": 8, "author_info_list": [ { "name": "Binli Luo", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "Researcher" }, { "name": "Bocheng Xu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" }, { "name": "Han Zhou", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Jimmy Tan", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Shangjian Zhong", "gender": "Female", "institution": "City University of Hong Kong", "country": "HK", "position": "MS student" }, { "name": "Yao Zhao", "gender": "Female", "institution": "Engineer Information Technology Consulting Company", "country": "CN", "position": "Researcher" }, { "name": "Yuanzheng Tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Zhiyuan Gao", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Intern" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 22, "n_ref": 37, "n_ref_all": 45, "n_fig": 2, "n_tab": 4, "L_tab": 1344, "n_element_tab": 69, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 605, "n_element_tab_1": 56, "formula_len_all": 914, "formula_len_all_1": 995, "len_all": 104379, "len_all_1": 54143, "len_abs": 1410, "len_title": 135, "len_sents": 30469, "len_sents_1": 26136, "n_sents": 192, "n_sents_1": 166, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1418, "title": "TIM: Interpretable Modelling of Complex Temporal Interactions in Multivariate Networks", "abs": "Multivariate time series forecasting is crucial across various fields and essential for addressing numerous real-world challenges. However, existing forecasting methods have significant limitations: while Transformer models are effective, they are constrained by high computational costs and declining performance in long-term forecasting; MLP models struggle to capture complex multivariate interactions. These issues hinder the models' ability to accurately decompose seasonality and trends. To tackle these problems, we propose a new method called TIM. Through a cross-layer architecture, TIM decomposes time series predictions into temporal features, multivariate interaction features, and residual components. Our all-MLP model integrates global features with complex multivariate dynamics. By introducing a linear self-attention mechanism across variables and time steps, TIM enhances the learning of feature interactions and accurately captures temporal transitions between domains. This innovative design leverages linear attention mechanisms and cross-layer architecture to more effectively model temporal features and multivariate interactions. It surpasses traditional Transformer-based methods by improving predictive accuracy while maintaining linear computational complexity. Experimental results demonstrate that TIM outperforms existing state-of-the-art methods while ensuring computational efficiency.", "keywords": [ "Time Series", "Deep Learning", "Decomposition" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 4, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "Mw42TqZ0o5", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Jundong Li", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" }, { "name": "Song Wang", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Xinyu Zhao", "gender": "unknown", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Zhen Tan", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "huan liu", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Full Professor" } ], "n_formula": 21, "n_formula_1": 16, "n_ref_uni": 57, "n_ref": 130, "n_ref_all": 153, "n_fig": 7, "n_tab": 6, "L_tab": 1936, "n_element_tab": 241, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 700, "n_element_tab_1": 111, "formula_len_all": 864, "formula_len_all_1": 598, "len_all": 163502, "len_all_1": 71858, "len_abs": 1251, "len_title": 92, "len_sents": 46583, "len_sents_1": 32981, "n_sents": 379, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1245, "title": "GraphRCG: Self-Conditioned Graph Generation", "abs": "Graph generation aims to create new graphs that closely align with a target graph distribution. Existing works often implicitly capture this distribution by aligning the output of a generator with each training sample. As such, the overview of the entire distribution is not explicitly captured and used for graph generation. In contrast, in this work, we propose a novel self-conditioned graph generation framework designed to explicitly model graph distributions and employ these distributions to guide the generation process. We first perform self-conditioned modeling to capture the graph distributions by transforming each graph sample into a low-dimensional representation and optimizing a representation generator to create new representations reflective of the learned distribution. Subsequently, we leverage these bootstrapped representations as self-conditioned guidance for the generation process, thereby facilitating the generation of graphs that more accurately reflect the learned distributions. We conduct extensive experiments on generic and molecular graph datasets. Our framework demonstrates superior performance over existing state-of-the-art graph generation methods in terms of graph quality and fidelity to training data.", "keywords": [ "Graph Generation", "Diffusion Models" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "Mw16Akb1CR", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Albert Bifet", "gender": "Male", "institution": "The University of Waikato", "country": "NZ", "position": "Full Professor" }, { "name": "Ayman Chaouki", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "FR", "position": "Postdoc" }, { "name": "Jesse Read", "gender": "unknown", "institution": "Ecole polytechnique", "country": "", "position": "Full Professor" } ], "n_formula": 117, "n_formula_1": 33, "n_ref_uni": 17, "n_ref": 39, "n_ref_all": 87, "n_fig": 6, "n_tab": 13, "L_tab": 11164, "n_element_tab": 1228, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 1102, "n_element_tab_1": 299, "formula_len_all": 10223, "formula_len_all_1": 2145, "len_all": 209816, "len_all_1": 65402, "len_abs": 2473, "len_title": 141, "len_sents": 60827, "len_sents_1": 24404, "n_sents": 698, "n_sents_1": 270, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1088, "title": "Branches: A Fast Dynamic Programming and Branch & Bound Algorithm for Optimal Decision Trees", "abs": "Decision Tree (DT) Learning is a fundamental problem in Interpretable Machine Learning, yet it poses a formidable optimisation challenge. Despite numerous efforts dating back to the early 1990's, practical algorithms have only recently emerged, primarily leveraging Dynamic Programming (DP) and Branch \\& Bound (B\\&B) techniques. These methods fall into two categories: algorithms like DL8.5, MurTree and STreeD utilise an efficient DP strategy but lack effective bounds for pruning the search space; while algorithms like OSDT and GOSDT employ more efficient pruning bounds but at the expense of a less refined DP strategy. We introduce Branches, a new algorithm that combines the strengths of both approaches. Using DP and B\\&B with a novel analytical bound for efficient pruning, Branches offers both speed and sparsity optimisation. Unlike other methods, it also handles non-binary features. Theoretical analysis shows its lower complexity compared to existing methods, and empirical results confirm that Branches outperforms the state-of-the-art in speed, iterations, and optimality.", "keywords": [ "Optimal Decision Trees", "Dynamic Programming", "Branch & Bound", "Markov Decision Process" ], "rating_list": [ 6, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "Mvn5g49RrM", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Chengquan Guo", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Chulin Xie", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "PhD student" }, { "name": "Yu Yang", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Zinan Lin", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Senior Researcher" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 78, "n_ref_all": 112, "n_fig": 16, "n_tab": 7, "L_tab": 2394, "n_element_tab": 142, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1001, "n_element_tab_1": 27, "formula_len_all": 19, "formula_len_all_1": 0, "len_all": 206470, "len_all_1": 70392, "len_abs": 1797, "len_title": 61, "len_sents": 75034, "len_sents_1": 34844, "n_sents": 568, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1580, "title": "RedCodeAgent: Automatic Red-teaming Agent against Code Agents", "abs": "LLM-based code agents, integrated with external tools like the Python interpreter, can interact with broad system environments and leverage code execution feedback to improve or self-debug generated code for better task-solving. However, as these code agents evolve rapidly in terms of capabilities, their increasing sophistication also amplifies security risks, such as generating or executing risky and buggy code. Traditional static safety benchmarks and manually designed red-teaming tools struggle to keep up with this rapid evolution, lacking the ability to adapt dynamically to the changing behaviors of code agents. To address these limitations, we propose RedCodeAgent, the first fully automated and adaptive red-teaming agent against given code agents. Equipped with red-teaming tools for function-calling and a novel memory module for accumulating successful attack experience, RedCodeAgent dynamically optimizes input prompts to jailbreak the target code agent for risky code execution. Unlike static benchmarks or red-teaming tools, RedCodeAgent autonomously adapts its attack strategies, making it a scalable solution to the growing challenge of testing increasingly sophisticated code agents. Experimental results show that compared to state-of-the-art LLM jailbreaking methods, RedCodeAgent achieves significantly higher attack success rates on the same tasks while maintaining high overall efficiency. By autonomously exploring and exploiting vulnerabilities of code agents, RedCodeAgent provides critical insights into the evolving security risks of code agents.", "keywords": [ "Trustworthy machine learning", "Code Agents", "LLM" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "Mvn48u0ehO", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Merve Atasever", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Mihir Nitin Kulkarni", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "MS student" }, { "name": "Qingpei Li", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "Researcher" }, { "name": "Jyotirmoy Deshmukh", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Associate Professor" }, { "name": "Matthew Min Hong", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 35, "n_ref": 47, "n_ref_all": 63, "n_fig": 5, "n_tab": 6, "L_tab": 2168, "n_element_tab": 353, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1709, "n_element_tab_1": 344, "formula_len_all": 85, "formula_len_all_1": 85, "len_all": 143537, "len_all_1": 64407, "len_abs": 1627, "len_title": 120, "len_sents": 45030, "len_sents_1": 28737, "n_sents": 358, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1637, "title": "Multi-Agent Path Finding via Decision Transformer and LLM Collaboration", "abs": "Multi-Agent Path Finding (MAPF) is a significant problem with pivotal applications in robotics and logistics. The problem involves determining collision-free paths for multiple agents with specific goals in a 2D grid-world environment. Unfortunately, finding optimal solutions for MAPF is an NP-hard problem. Traditional centralized planning approaches are intractable for large numbers of agents and inflexible when adapting to\ndynamic changes in the environment. On the other hand, existing decentralized methods utilizing learning-based strategies suffer from two main drawbacks: (1) training takes times ranging from days to weeks, and (2) they often tend to exhibit self-centered agent behaviors leading to increased collisions. We introduce a novel approach leveraging the Decision Transformer (DT) architecture that enables agents to learn individual policies efficiently. We capitalize on the transformer's capability for long-horizon planning and the advantages of offline reinforcement learning to drastically reduce training times to a few hours. We further show that integrating an LLM (GPT-4o), enhances the performance of DT policies in mitigating undesirable behaviors such as prolonged idling at specific positions and undesired deviations from goal positions. We focus our empirical evaluation on both scenarios with static environments and in dynamically changing environments where agents' goals are altered during inference. Results demonstrate that incorporating an LLM for dynamic scenario adaptation in MAPF significantly enhances the agents' performance and paves the way for more adaptable multi-agent systems.", "keywords": [ "Multi-Agent Path Finding", "Multi-Agent Reinforcement Learning", "Decision Transformer", "Large Language Models", "Autonomous Agents" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 5, 3 ] }, { "paper_id": "MvEkN2ejZ1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Haoxin Wang", "gender": "Male", "institution": "Georgia State University", "country": "US", "position": "Assistant Professor" }, { "name": "Hongyu Ke", "gender": "Male", "institution": "Georgia State University", "country": "US", "position": "PhD student" }, { "name": "Jack Morris", "gender": "Male", "institution": "Georgia State University", "country": "US", "position": "PhD student" }, { "name": "Kentaro Oguchi", "gender": "unknown", "institution": "Toyota Motor North America", "country": "US", "position": "Director and Senior Principal Researcher" }, { "name": "Xiaofei Cao", "gender": "Male", "institution": "Toyota Motor Corporation", "country": "US", "position": "Principal Researcher" }, { "name": "Yi Ding", "gender": "unknown", "institution": "Georgia State University", "country": "", "position": "Assistant Professor" }, { "name": "Yongkang Liu", "gender": "unknown", "institution": "Toyota", "country": "US", "position": "Principal Researcher" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 22, "n_ref": 45, "n_ref_all": 66, "n_fig": 7, "n_tab": 8, "L_tab": 2226, "n_element_tab": 324, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 923, "n_element_tab_1": 151, "formula_len_all": 661, "formula_len_all_1": 851, "len_all": 117463, "len_all_1": 60632, "len_abs": 887, "len_title": 123, "len_sents": 37554, "len_sents_1": 26592, "n_sents": 283, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 891, "title": "MamBEV: Enabling State Space Models to Learn Birds-Eye-View Representations", "abs": "3D visual perception tasks, such as 3D detection from multi-camera images, are essential components of autonomous driving and assistance systems. However, designing computationally efficient methods remains a significant challenge. In this paper, we propose a Mamba-based framework called MamBEV, which learns unified Bird's Eye View (BEV) representations using linear spatio-temporal SSM-based attention. This approach supports multiple 3D perception tasks with significantly improved computational and memory efficiency. Furthermore, we introduce SSM based cross-attention, analogous to standard cross attention, where BEV query representations can interact with relevant image features. Extensive experiments demonstrate MamBEV's promising performance across diverse visual perception metrics, highlighting its advantages in input scaling efficiency compared to existing benchmark models.", "keywords": [ "BEV", "Bird\u2019s Eye View", "State space model", "causal attention", "3d computer vision" ], "rating_list": [ 6, 8, 8, 6, 5 ], "soundness_list": [ 2, 4, 4, 2, 2 ], "presentation_list": [ 2, 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "Mv3GAYJGcW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 14, "author_info_list": [ { "name": "Alexander G Hauptmann", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Bin Luo", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Chenyang Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Hanyuan Chen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jin-Peng Lan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jingdong Sun", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Qi He", "gender": "unknown", "institution": "Southwest Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Wangmeng Xiang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xianhui Lin", "gender": "Male", "institution": "Tongyi Lab, Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xuansong Xie", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yifeng Geng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zhi-Qi Cheng", "gender": "Not Specified", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "kang zhu", "gender": "unknown", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "HE JUNYAN", "gender": "Male", "institution": "Tongyi Lab, Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 13, "n_ref_uni": 15, "n_ref": 22, "n_ref_all": 39, "n_fig": 8, "n_tab": 4, "L_tab": 1610, "n_element_tab": 502, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2571, "n_element_tab_1": 144, "formula_len_all": 750, "formula_len_all_1": 631, "len_all": 135556, "len_all_1": 57917, "len_abs": 1098, "len_title": 98, "len_sents": 41066, "len_sents_1": 23596, "n_sents": 327, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 111, "L_abs": 1035, "title": "MetaDesigner: Advancing Artistic Typography through AI-Driven, User-Centric, and Multilingual WordArt Synthesis", "abs": "MetaDesigner revolutionizes artistic typography synthesis by leveraging the strengths of Large Language Models (LLMs) to drive a design paradigm centered around user engagement. At the core of this framework lies a multi-agent system comprising the Pipeline, Glyph, and Texture agents, which collectively enable the creation of customized WordArt, ranging from semantic enhancements to the imposition of complex textures. MetaDesigner incorporates a comprehensive feedback mechanism that harnesses insights from multimodal models and user evaluations to refine and enhance the design process iteratively. Through this feedback loop, the system adeptly tunes hyperparameters to align with user-defined stylistic and thematic preferences, generating WordArt that not only meets but exceeds user expectations of visual appeal and contextual relevance. Empirical validations highlight MetaDesigner's capability to effectively serve diverse WordArt applications, consistently producing aesthetically appealing and context-sensitive results.", "keywords": [ "MetaDesigner" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 2, 2, 4 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "MueN6LyTmS", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 10, "author_info_list": [ { "name": "Fuchun Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Huaping Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Kangyao Huang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Runfa Chen", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Tianyu Shi", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "MS student" }, { "name": "Wenbing Huang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yulong Xia", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yunfei Xie", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Sun Jun", "gender": "unknown", "institution": "University of Mississippi Medical Center", "country": "CN", "position": "Researcher" }, { "name": "Di Xinpeng", "gender": "unknown", "institution": "Marine Biological Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 57, "n_formula_1": 23, "n_ref_uni": 21, "n_ref": 77, "n_ref_all": 109, "n_fig": 14, "n_tab": 2, "L_tab": 120, "n_element_tab": 12, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 110, "n_element_tab_1": 11, "formula_len_all": 4327, "formula_len_all_1": 1599, "len_all": 152904, "len_all_1": 62108, "len_abs": 252, "len_title": 164, "len_sents": 52581, "len_sents_1": 27117, "n_sents": 436, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1386, "title": "Subequivariant Morphology-Behavior Co-Evolution in 3D Environments", "abs": "The co-evolution of morphology and behavior in 3D space has garnered considerable interest in the field of embodied intelligence. \nWhile recent studies have highlighted the considerable benefits of geometric symmetry for tasks like learning to locomote, navigate, and explore in dynamic 3D environments, its role within co-evolution setup remains unexplored.\nExisting benchmarks encounter several key issues: 1) the task lacks consideration for spatial geometric information; 2) the method lacks geometric symmetry to deal with the complexities in 3D environments.\nIn this work, we propose a novel setup, named Subequivariant Morphology-Behavior Co-Evolution in 3D Environments (3DS-MB), to address the identified limitations.\nTo be specific, we propose EquiEvo, which injects geometric symmetry, i.e., subequivariance, to construct dynamic, learnable local reference frames, enabling the joint policy to generalize to diverse task spatial structures, thereby improving co-evolution efficiency.\nThen, we evaluate EquiEvo on the proposed environments, where our method consistently and significantly outperforms existing approaches in tasks such as locomotion navigation and adversarial scenarios.\nExtensive experiments underscore the importance of subequivariance for the co-evolution of morphology and behavior, effective morphology-task mapping and robust morphology-behavior mapping.", "keywords": [ "Geometric Graph", "Reinforcement Learning", "Morphology-Behavior Co-Evolution", "3D", "Subequivariance" ], "rating_list": [ 8, 6, 6, 5, 1 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 3, 2, 1 ], "confidence_list": [ 1, 4, 2, 3, 5 ] }, { "paper_id": "MuXF0UZsoW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Monali Barbate", "gender": "Female", "institution": "Microsoft", "country": "IN", "position": "Researcher" }, { "name": "Sudeep Choudhary", "gender": "Male", "institution": "Indian Statistical Institute, Kolkata", "country": "IN", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 6, "n_ref": 9, "n_ref_all": 10, "n_fig": 0, "n_tab": 1, "L_tab": 387, "n_element_tab": 38, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 387, "n_element_tab_1": 38, "formula_len_all": 306, "formula_len_all_1": 306, "len_all": 49525, "len_all_1": 49525, "len_abs": 1766, "len_title": 140, "len_sents": 20711, "len_sents_1": 20711, "n_sents": 152, "n_sents_1": 152, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1632, "title": "Optimizing Detection Techniques for High-Precision Icon Recognition in Sparse Feature Spaces", "abs": "CNNs usually work well when they can extract progressively higher-level features through the layers. In small, low-resolution images, the depth of feature extraction is limited, leading to a sparsity in the feature space. Icon detection presents a unique challenge due to the small, feature-sparse nature of the target images, which often results in limited discriminative features. To address this, we propose an icon detection model based on a Siamese network architecture. This approach draws inspiration from face recognition frameworks. The modified architectures are aimed at being well-suited for distinguishing subtle differences between icon pairs. Given the relatively sparse feature space of these icons compared to larger images, we explore several enhancements to improve performance. Key innovations include the integration of attention mechanisms to focus on informative features, multi-scale feature extraction for better detail capture and contrastive learning. We additionally employ adversarial training to enhance performance. Additionally, we investigate dynamic margins in metric learning to model icon similarities. Self-supervised pretraining and Neural Architecture Search are employed to further refine and optimize the network. Our comprehensive evaluation demonstrates significant improvements in icon detection, highlighting the effectiveness of these advanced techniques in handling small, feature-sparse image data. This solution\noffers a valuable advancement in high-precision icon recognition, with potential applications in user interface design, software development, and digital asset management.", "keywords": [ "CNN", "Feature Sparse Images", "Contrastive Learning", "Adversarial Training", "Dynamic Margins", "Attention Mechanisms" ], "rating_list": [ 3, 1, 1, 3 ], "soundness_list": [ 2, 1, 2, 1 ], "presentation_list": [ 2, 1, 2, 1 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "Mu2tNzS0AM", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Cornelia Caragea", "gender": "unknown", "institution": "University of Illinois at Chicago", "country": "US", "position": "Full Professor" }, { "name": "Tiberiu Sosea", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 26, "n_ref": 62, "n_ref_all": 70, "n_fig": 1, "n_tab": 2, "L_tab": 3920, "n_element_tab": 505, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 3967, "n_element_tab_1": 512, "formula_len_all": 842, "formula_len_all_1": 842, "len_all": 135707, "len_all_1": 74177, "len_abs": 1098, "len_title": 109, "len_sents": 32672, "len_sents_1": 31961, "n_sents": 193, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1124, "title": "LLM-Informed Semi-Supervised Learning for Text Classification", "abs": "Large Language Models (LLMs) have shown impressive zero-shot and few-shot capabilities in many NLP tasks including text classification. While these models outperform others in terms of raw performance when few examples are available, they are expensive to use in practice and may lag behind traditional approaches when data (labeled or unlabeled) is plentiful. Semi-supervised learning (SSL) on the other hand can utilize large amounts of unlabeled data in combination with labeled data to improve a model's performance. In this paper, we propose to unify LLM and SSL under a common framework which effectively leverages the few-shot capabilities of LLMs in combination with SSL's ability to extract valuable information from unlabeled data to improve the model capabilities in text classification. Our approach, entitled LLM-SSL, utilizes LLMs to generate predictions on unlabeled examples and uses these predictions to guide the SSL training and improve the quality of pseudo-labels during training. We show that LLM-SSL outperforms both prior SSL approaches as well as few-shot LLMs on six text classification benchmarks.", "keywords": [ "semi-supervised learning", "LLM" ], "rating_list": [ 5, 5, 5, 1, 3 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 5, 5 ] }, { "paper_id": "MtoklWYQus", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Akshay Dudhane", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Omkar Chakradhar Thawakar", "gender": "Male", "institution": "MBZUAI", "country": "AE", "position": "Researcher" }, { "name": "Salman Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Associate Professor" }, { "name": "Fahad Shahbaz Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" }, { "name": "Syed Waqas Zamir", "gender": "Male", "institution": "Inception Institute of Artificial Intelligence", "country": "AE", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 38, "n_ref": 91, "n_ref_all": 115, "n_fig": 12, "n_tab": 4, "L_tab": 908, "n_element_tab": 74, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1038, "n_element_tab_1": 86, "formula_len_all": 236, "formula_len_all_1": 95, "len_all": 131722, "len_all_1": 63197, "len_abs": 1745, "len_title": 129, "len_sents": 36323, "len_sents_1": 28176, "n_sents": 272, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1759, "title": "Dynamic Pre-training: Towards Efficient and Scalable All-in-One Image Restoration", "abs": "All-in-one image restoration tackles different types of degradations with a unified model instead of having task-specific, non-generic models for each degradation. The requirement to tackle multiple degradations using the same model can lead to high-complexity designs with fixed configuration that lack the adaptability to more efficient alternatives. We propose DyNet, a dynamic family of networks designed in an encoder-decoder style for all-in-one image restoration tasks. Our DyNet can seamlessly switch between its bulkier and lightweight variants, thereby offering flexibility for efficient model deployment with a single round of training. This seamless switching is enabled by our weights-sharing mechanism, forming the core of our architecture and facilitating the reuse of initialized module weights. Further, to establish robust weights initialization, we introduce a dynamic pre-training strategy that trains variants of the proposed DyNet concurrently, thereby achieving a 50\\% reduction in GPU hours. Our dynamic pre-training strategy eliminates the need for maintaining separate checkpoints for each variant, as all models share a common set of checkpoints, varying only in model depth. This efficient strategy significantly reduces storage overhead and enhances adaptability. To tackle the unavailability of large-scale dataset required in pre-training, we curate a high-quality, high-resolution image dataset named Million-IRD, having 2M image samples. We validate our DyNet for image denoising, deraining, and dehazing in all-in-one setting, achieving state-of-the-art results with 31.34\\% reduction in GFlops and a 56.75\\% reduction in parameters compared to baseline models. The source codes and trained models will be publicly released.", "keywords": [ "All-in-one \u00b7 Image Restoration \u00b7 Foundation model" ], "rating_list": [ 5, 5, 5, 1 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "MtjPIDWyWK", "primary_area": "reinforcement learning", "n_author": 10, "author_info_list": [ { "name": "Bo Zheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Chuan Yu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jian Xu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Tianyu Wang", "gender": "Male", "institution": "TAOBAO & TMALL GROUP ", "country": "CN", "position": "Researcher" }, { "name": "Xiaotie Deng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Yeshu Li", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Engineer" }, { "name": "Yusen Huo", "gender": "Male", "institution": "TAOBAO & TMALL GROUP", "country": "CN", "position": "Researcher" }, { "name": "Zhengye Han", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhijian Duan", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhilin Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Senior Algorithm Specialist" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 23, "n_ref": 28, "n_ref_all": 34, "n_fig": 2, "n_tab": 1, "L_tab": 1392, "n_element_tab": 90, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 125, "n_element_tab_1": 5, "formula_len_all": 543, "formula_len_all_1": 453, "len_all": 92794, "len_all_1": 54124, "len_abs": 1037, "len_title": 123, "len_sents": 32340, "len_sents_1": 27788, "n_sents": 220, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1039, "title": "Action Sequence Planner: An Alternative For Offline Reinforcement Learning", "abs": "Offline reinforcement learning methods, which typically train agents that make decisions step by step, are known to suffer from instability due to bootstrapping and function approximation, especially when applied to tasks requiring long-horizon planning. To alleviate these issues, in this paper, we propose a novel policy gradient approach by planning an action sequence in a high-dimensional space.This design implicitly models temporal dependencies, excelling in long-horizon and horizon-critical tasks. Furthermore, we discover that replacing maximum likelihood with cross-entropy loss in policy gradient methods significantly stabilizes training gradients, leading to substantial performance improvements in long-horizon tasks. The proposed neural network-based solution features a simple architecture that not only facilitates ease of training and convergence but also demonstrates high efficiency and effective performance. Extensive experimental results reveal that our method exhibits strong performance across a variety of tasks.", "keywords": [ "Offline reinforcement learning", "policy gradient", "long horizon" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 1, 1 ], "presentation_list": [ 2, 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "MtW30ql5Oj", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Qianxiao Li", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "REN Lianhai", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 61, "n_formula_1": 31, "n_ref_uni": 24, "n_ref": 53, "n_ref_all": 112, "n_fig": 1, "n_tab": 11, "L_tab": 2763, "n_element_tab": 256, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1250, "n_element_tab_1": 99, "formula_len_all": 5211, "formula_len_all_1": 2533, "len_all": 145055, "len_all_1": 66998, "len_abs": 1045, "len_title": 117, "len_sents": 44338, "len_sents_1": 27597, "n_sents": 368, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1107, "title": "BP-Modified Local Loss for Efficient Training of Deep Neural Networks", "abs": "The training of large models is memory-constrained, one direction to relieve\n this is training using local loss, like GIM, LoCo, and Forward-Forward\n algorithms. However, the local loss methods often face the issue of slow or\n non-convergence. In this paper, we propose a novel BP-modified local loss\n method that uses the true Backward Propagation (BP) gradient to modify the\n local loss gradient to improve the performance of local loss training. We\n use the stochastic modified equation to analyze our method and show that\n modified offset decreases the bias between the BP gradient and local loss\n gradient, but introduces additional variance, which results in a\n bias-variance balance. Numerical experiments on full-tuning and LoKr tuning\n on the ResNet-50 model and LoRA tuning on the ViT-b16 model on CIFAR-100\n datasets show 20.5\\% test top-1 accuracy improvement for the Forward-Forward\n algorithm, 18.6\\% improvement for LoCo algorithm and achieve only on average\n 7.7\\% of test accuracy loss compared to the BP algorithm, with up to 75\\%\n memory savings.", "keywords": [ "deep learning optimization", "local loss training", "bias-variance balance" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "MtDd7rWok1", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Daochang Liu", "gender": "Male", "institution": "University of Western Australia", "country": "AU", "position": "Lecturer" }, { "name": "Eunbyung Park", "gender": "Male", "institution": "Sungkyunkwan University", "country": "KR", "position": "Assistant Professor" }, { "name": "Junyu Zhang", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "PhD student" }, { "name": "Shichao Zhang", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "Full Professor" }, { "name": "Chang Xu", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" } ], "n_formula": 26, "n_formula_1": 22, "n_ref_uni": 68, "n_ref": 189, "n_ref_all": 236, "n_fig": 7, "n_tab": 12, "L_tab": 3652, "n_element_tab": 332, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1533, "n_element_tab_1": 73, "formula_len_all": 1228, "formula_len_all_1": 1164, "len_all": 206344, "len_all_1": 70865, "len_abs": 1965, "len_title": 106, "len_sents": 49642, "len_sents_1": 32217, "n_sents": 326, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 58, "L_abs": 1275, "title": "Anti-Exposure Bias in Diffusion Models via Prompt Learning", "abs": "Diffusion models (DMs) have achieved record-breaking performance in image generation tasks.\nNevertheless, in practice, the training-sampling discrepancy, caused by score estimation error and discretization error, limits the modeling ability of DMs, a phenomenon known as exposure bias.\nTo alleviate such exposure bias and further improve the generative performance, we put forward a prompt learning framework built upon a lightweight prompt prediction model.\nConcretely, our model learns an anti-bias prompt for the generated sample at each sampling step, aiming to compensate for the exposure bias that arises.\nFollowing this design philosophy, our framework rectifies the sampling trajectory to match the training trajectory, thereby reducing the divergence between the target data distribution and the modeling distribution.\nTo train the prompt prediction model, we simulate exposure bias by constructing training data and introduce a time-dependent weighting function for optimization.\nEmpirical results on various DMs demonstrate the superiority of our prompt learning framework across three benchmark datasets.\nImportantly, the optimized prompt prediction model effectively improves image quality with only a 5\\% increase in sampling overhead, which remains negligible.", "keywords": [ "Diffusion Models", "Exposure Bias", "Prompt Learning", "Sampling Trajectory" ], "rating_list": [ 6, 8, 8, 6, 8 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 4, 3, 4 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4, 3 ] }, { "paper_id": "MtCcVO8Oux", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Fei Gao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Liuao Pei", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Long Xu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhichao Han", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 34, "n_formula_1": 26, "n_ref_uni": 46, "n_ref": 66, "n_ref_all": 99, "n_fig": 10, "n_tab": 1, "L_tab": 40, "n_element_tab": 2, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 959, "n_element_tab_1": 130, "formula_len_all": 2434, "formula_len_all_1": 2121, "len_all": 147547, "len_all_1": 66977, "len_abs": 1414, "len_title": 97, "len_sents": 49138, "len_sents_1": 30442, "n_sents": 321, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1422, "title": "Agile Flight with Optimization Embedded Networks", "abs": "To bridge the gap between perception and planning in traditional navigation systems, we address the challenge of learning optimal trajectories directly from depth information in an end-to-end fashion. Using neural networks as black-box replacements for traditional modules can compromise robustness and stability. Moreover, such methods often fail to adequately account for the robot's kinematic constraints, leading to trajectories that may not be satisfactorily executable. In this paper, we integrate the strengths of conventional methods and neural networks by introducing an optimization-embedded neural network based on a compact trajectory library. Neural networks establish spatial constraints for model-based trajectory planning, followed by robust numerical optimization to achieve feasible and optimal solutions. By making the process differentiable, our model seamlessly approximates the optimal trajectory. Additionally, the introduction of a regularized trajectory library enables the method to efficiently capture the spatial distribution of optimal trajectories with minimal storage cost, ensuring multimodal planning characteristics. Evaluations in complex, unseen environments demonstrate our method\u2019s superior performance over state-of-the-art algorithms. Real-world flight experiments with a small onboard computer showcase the autonomous quadrotor\u2019s ability to navigate swiftly through dense forests.", "keywords": [ "Robotics", "Autonomous Navigation", "Motion Planning and Control", "Differentiable Optimization" ], "rating_list": [ 6, 3, 6, 3 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "MscdsFVZrN", "primary_area": "foundation or frontier models, including LLMs", "n_author": 26, "author_info_list": [ { "name": "Abdalghani Abujabal", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Abdulmohsen Al-Thubaity", "gender": "Male", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "SA", "position": "Principal Researcher" }, { "name": "Ahmed Abdelali", "gender": "Male", "institution": "National Center for AI , Saudi Data and AI Authority", "country": "", "position": "Principal Researcher" }, { "name": "Ali Alammari", "gender": "unknown", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "SA", "position": "Researcher" }, { "name": "Areeb Alowisheq", "gender": "unknown", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "SA", "position": "Researcher" }, { "name": "Dr. Majed Alrubaian", "gender": "Male", "institution": "King Saud University", "country": "SA", "position": "Assistant Professor" }, { "name": "Faisal Abdulrahman Mirza", "gender": "Male", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "SA", "position": "Researcher" }, { "name": "Ghadah Alabduljabbar", "gender": "Female", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "SA", "position": "data scientist" }, { "name": "Haidar Khan", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Hassan A. Alahmed", "gender": "Male", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "SA", "position": "Researcher" }, { "name": "Hisham Abdullah Alyahya", "gender": "Male", "institution": "Saudi Data and AI Authority", "country": "SA", "position": "Researcher" }, { "name": "Jeril Kuriakose", "gender": "unknown", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "SA", "position": "Researcher" }, { "name": "Maryam Al Mansour", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Nora Al-Twairesh", "gender": "Female", "institution": "King Saud University", "country": "SA", "position": "Associate Professor" }, { "name": "Norah A. Alzahrani", "gender": "Female", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "SA", "position": "AI Engineer" }, { "name": "Raghad Alkhathran", "gender": "unknown", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "SA", "position": "Researcher" }, { "name": "Raneem Alnajim", "gender": "Female", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "", "position": "Researcher" }, { "name": "Saad Amin Hassan", "gender": "Male", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "", "position": "Researcher" }, { "name": "Salman Alsubaihi", "gender": "Male", "institution": "KAUST", "country": "", "position": "MS student" }, { "name": "Sultan AlRashed", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Researcher" }, { "name": "Yazeed Alnumay", "gender": "Male", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "SA", "position": "Researcher" }, { "name": "Yousef Almushayqih", "gender": "unknown", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "SA", "position": "Researcher" }, { "name": "Zaki Alawami", "gender": "Male", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "SA", "position": "Researcher" }, { "name": "M SAIFUL BARI", "gender": "Male", "institution": "National Centre of Artificial Intelligence, Saudi Data and AI Authority", "country": "SA", "position": "Principal Researcher" }, { "name": "Nouf Alotaibi", "gender": "Female", "institution": "Saudi Data and AI Authority", "country": "SA", "position": "AI Engineer" }, { "name": "Shaykhah Alsubaie", "gender": "Female", "institution": "Saudi Data and AI Authority, Saudi Data and AI Authority", "country": "SA", "position": "AI Engineer" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 97, "n_ref": 142, "n_ref_all": 188, "n_fig": 14, "n_tab": 14, "L_tab": 9595, "n_element_tab": 1326, "n_fig_1": 11, "n_tab_1": 6, "L_tab_1": 2575, "n_element_tab_1": 452, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 418059, "len_all_1": 73318, "len_abs": 1084, "len_title": 51, "len_sents": 62741, "len_sents_1": 31304, "n_sents": 479, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1155, "title": "ALLaM: Large Language Models for Arabic and English", "abs": "In this work, we present ALLaM: Arabic Large Language Model, a series of large language models to support the ecosystem of Arabic Language Technologies (ALT). ALLaM is carefully trained, considering the values of language alignment and transferability of knowledge at scale. The models are based on an autoregressive decoder-only architecture and are pretrained on a mixture of Arabic and English texts. We illustrate how the second-language acquisition via vocabulary expansion can help steer a language model towards a new language without any major catastrophic forgetting in English. Furthermore, we highlight the effectiveness of using translation data and the process of knowledge encoding within the language model's latent space. Finally, we show that effective alignment with human preferences can significantly enhance the performance of a large language model (LLM) compared to less aligned models of a larger scale. Our methodology enables us to achieve state-of-the-art performance in various Arabic benchmarks, including MMLU Arabic, ACVA, and Arabic Exams. Our aligned models improve both in Arabic and English from its base aligned models.", "keywords": [ "Large Language Model", "English", "Arabic", "Second Language Acquisition" ], "rating_list": [ 8, 8, 3, 6 ], "soundness_list": [ 4, 4, 3, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "MsUhByb3CM", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Ivan Vladimir Meza Ruiz", "gender": "Male", "institution": "Instituto de Investigaciones en Matem\u00e1ticas Aplicadas y en Sistemas, Universidad Nacional Aut\u00f3noma de M\u00e9xico", "country": "", "position": "Research associate" }, { "name": "Victor Sebastian Martinez Pozos", "gender": "Male", "institution": "Universidad Nacional Aut\u00f3noma de M\u00e9xico", "country": "MX", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 20, "n_ref": 22, "n_ref_all": 29, "n_fig": 3, "n_tab": 2, "L_tab": 662, "n_element_tab": 70, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 662, "n_element_tab_1": 70, "formula_len_all": 314, "formula_len_all_1": 314, "len_all": 75805, "len_all_1": 53685, "len_abs": 1020, "len_title": 134, "len_sents": 25613, "len_sents_1": 25661, "n_sents": 168, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1023, "title": "Extracting Symbolic Sequences from Visual Representations via Self-Supervised Learning", "abs": "In this paper, we explore the potential of abstracting complex visual information into discrete, structured symbolic sequences using self-supervised learning (SSL). Inspired by how language abstracts and organizes information to enable better reasoning and generalization, we propose a novel approach for generating symbolic representations from visual data. To learn these sequences, we extend the DINO framework to handle both visual and symbolic information. Initial experiments suggest that the generated symbolic sequences capture a meaningful level of abstraction, though further refinement is required. An advantage of our method is its interpretability: the sequences are produced by a decoder transformer using cross-attention, allowing attention maps to be linked to specific symbols and offering insight into how these representations correspond to image regions. This approach lays the foundation for creating interpretable symbolic representations with potential applications in high-level scene understanding.", "keywords": [ "Self-Supervised Learning", "Symbolic Representations", "Information Theory", "Knowledge Distillation", "Visual Abstraction", "Interpretability" ], "rating_list": [ 3, 3, 5, 5, 5 ], "soundness_list": [ 2, 1, 2, 2, 3 ], "presentation_list": [ 3, 1, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 4, 3 ] }, { "paper_id": "MsRdq0ePTR", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 11, "author_info_list": [ { "name": "Jiashuo Sun", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jiongxiao Wang", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Peiran Li", "gender": "Male", "institution": "NeurOS", "country": "US", "position": "Founder" }, { "name": "Peiran Wang", "gender": "Not Specified", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Siyuan Ma", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Tian Xia", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Undergrad student" }, { "name": "Weidi Luo", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "", "position": "Undergrad student" }, { "name": "Xiaogeng Liu", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Yiming Liang", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "MS student" }, { "name": "Yu Wang", "gender": "Female", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "chaowei xiao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 39, "n_ref": 91, "n_ref_all": 105, "n_fig": 1, "n_tab": 10, "L_tab": 4317, "n_element_tab": 151, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 3646, "n_element_tab_1": 115, "formula_len_all": 619, "formula_len_all_1": 584, "len_all": 153631, "len_all_1": 66949, "len_abs": 1839, "len_title": 114, "len_sents": 44258, "len_sents_1": 31184, "n_sents": 319, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1845, "title": "Prompt Injection Benchmark for Foundation Model Integrated Systems", "abs": "Foundation Models (FMs) are increasingly integrated with external data sources and tools to handle complex tasks, forming FM-integrated systems with different modalities. However, such integration introduces new security vulnerabilities, especially when FMs interact dynamically with the system environments. One of the most critical threats is the prompt injection attack, where adversaries inject malicious instructions into the input environment, causing the model to deviate from user-intended behaviors. To advance the study of prompt injection vulnerabilities in FM-integrated systems, a comprehensive benchmark is essential. However, existing benchmarks fall short in two key areas: 1) they primarily focus on text-based modalities, lacking thorough analysis of diverse threats and attacks across more integrated modalities such as code, web pages, and vision; and 2) they rely on static test suites, failing to capture the dynamic, adversarial interplay between evolving attacks and defenses, as well as the interactive nature of agent-based environments. To bridge this gap, we propose the Prompt Injection Benchmark for FM-integrated Systems (FSPIB), which offers comprehensive coverage across various dimensions, including task modalities, threat categories, various attack and defense algorithms. Furthermore, FSPIB is interactive and dynamic, with evaluations conducted in interactive environments, and features a user-friendly front end that supports extensible attacks and defenses for ongoing research. By analyzing the performance of baseline prompt injection attacks and defenses, our benchmark highlights the prevalence of security vulnerabilities in FM-integrated systems and reveals the limited effectiveness of existing defense strategies, underscoring the urgent need for further research into prompt injection mitigation.", "keywords": [ "Prompt Injection Attack", "Foundation Models", "AI Agent" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 2, 4, 4, 2 ] }, { "paper_id": "MsAglk31tQ", "primary_area": "interpretability and explainable AI", "n_author": 13, "author_info_list": [ { "name": "Amin Madani", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Assistant Professor" }, { "name": "Anton Xue", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Bhuvnesh Jain", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "Full Professor" }, { "name": "Chaehyeon Kim", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Daniel A Hashimoto", "gender": "Male", "institution": ", University of Pennsylvania", "country": "", "position": "Assistant Professor" }, { "name": "Eric Wong", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Assistant Professor" }, { "name": "Helen Jin", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Helen Qu", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Lyle Ungar", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Full Professor" }, { "name": "Marco Gatti", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "Postdoc" }, { "name": "Masao Sako", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Full Professor" }, { "name": "Shreya Havaldar", "gender": "Female", "institution": "University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Weiqiu You", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 6, "n_ref_uni": 84, "n_ref": 148, "n_ref_all": 166, "n_fig": 4, "n_tab": 4, "L_tab": 1560, "n_element_tab": 105, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1013, "formula_len_all_1": 352, "len_all": 321602, "len_all_1": 65859, "len_abs": 1056, "len_title": 111, "len_sents": 54577, "len_sents_1": 31734, "n_sents": 459, "n_sents_1": 258, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1059, "title": "The FIX Benchmark: Extracting Features Interpretable to eXperts", "abs": "Feature-based methods are commonly used to explain model predictions, but these methods often implicitly assume that interpretable features are readily available. However, this is often not the case for high-dimensional data, and it can be hard even for domain experts to mathematically specify which features are important. Can we instead automatically extract collections or groups of features that are aligned with expert knowledge? To address this gap, we present FIX (Features Interpretable to eXperts), a benchmark for measuring how well a collection of features aligns with expert knowledge. In collaboration with domain experts, we propose FIXScore, a unified expert alignment measure applicable to diverse real-world settings across cosmology, psychology, and medicine domains in vision, language and time series data modalities. With FIXScore, we find that popular feature-based explanation methods have poor alignment with expert-specified knowledge, highlighting the need for new methods that can better identify features interpretable to experts.", "keywords": [ "Interpretable Features", "Explainability" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "Mri9WIfxSm", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Eli Waxman", "gender": "Male", "institution": "MetaDialog", "country": "AE", "position": "Principal Researcher" }, { "name": "Etrit Haxholli", "gender": "Not Specified", "institution": "Metadialog", "country": "AE", "position": "Researcher" }, { "name": "O\u011ful Can", "gender": "Male", "institution": "METU", "country": "", "position": "PhD student" }, { "name": "Yeti Ziya Gurbuz", "gender": "Male", "institution": "MetaDialog", "country": "DE", "position": "Researcher" } ], "n_formula": 199, "n_formula_1": 32, "n_ref_uni": 41, "n_ref": 118, "n_ref_all": 156, "n_fig": 5, "n_tab": 9, "L_tab": 4734, "n_element_tab": 327, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1979, "n_element_tab_1": 199, "formula_len_all": 22283, "formula_len_all_1": 3310, "len_all": 309517, "len_all_1": 70221, "len_abs": 1228, "len_title": 84, "len_sents": 92353, "len_sents_1": 26821, "n_sents": 808, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1239, "title": "Efficient Perplexity Bound and Ratio Matching in Discrete Diffusion Language Models", "abs": "While continuous diffusion models excel in modeling continuous distributions, their application to categorical data has been less effective. Recent work has shown that ratio-matching through *score-entropy* within a continuous-time discrete Markov chain (CTMC) framework serves as a competitive alternative to autoregressive models in language modeling.\nTo enhance this framework, we first introduce three new theorems concerning the KL divergence between the data and learned distribution. Our results serve as the discrete counterpart to those established for continuous diffusion models and allow us to derive an improved upper bound of the perplexity. Second, we empirically show that ratio-matching performed by minimizing the *denoising cross-entropy* between the clean and corrupted data enables models to outperform those utilizing score-entropy with up to 10\\% lower perplexity/generative-perplexity, and 15\\% faster training steps.\n To further support our findings, we introduce and evaluate a novel CTMC transition-rate matrix that allows prediction refinement, and derive the analytic expression for its matrix exponential which facilitates the computation of conditional ratios thus enabling efficient training and generation.", "keywords": [ "Discrete Diffusion", "CTMCs", "Ratio Matching", "Score Entropy", "Cross Entropy", "Language Models" ], "rating_list": [ 8, 5, 8, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "MrGca1Q7mK", "primary_area": "learning theory", "n_author": 1, "author_info_list": [ { "name": "Xin Li", "gender": "Male", "institution": "State University of New York at Albany", "country": "US", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 39, "n_ref": 72, "n_ref_all": 87, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 512, "formula_len_all_1": 481, "len_all": 111217, "len_all_1": 66580, "len_abs": 1257, "len_title": 126, "len_sents": 38968, "len_sents_1": 35160, "n_sents": 229, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1261, "title": "Location, Location, Location: Design Bias with Kernel Transformation", "abs": "It has been hypothesized that the old brain was compressed into cortical columns of the neocortex during the evolution of mammalian brains. Computational modeling of hippocampal-cortical interaction inspires us to propose a navigation-based implicit representation for manifold learning. The key new insight is to transform any explicit function (or geometrically a manifold) to an implicit representation using design bias for exploiting the concentration of measure (CoM) in high dimensional spaces. CoM-based blessing of dimensionality enables us to solve the manifold learning problem by direct-fit or local computation with guaranteed generalization property and without the need to discover global topology. We construct a memory encoding model, namely specification-before-generalization (SbG), and extend it into recursive kernel transformation to mirror the nested structure of the physical world. The biological plausibility of SbG learning is supported by its consistency with the wake-sleep cycles of mammalian brains. Finally, we showcase the application of design bias and recursive kernel transformation to understanding the phylogenetic continuity of navigation and memory and the manifold untangling of object recognition by the ventral stream.", "keywords": [ "implicit representation; concentration of measure; recursive kernel transformation; design bias; manifold learning" ], "rating_list": [ 1, 1, 1, 3 ], "soundness_list": [ 1, 1, 1, 1 ], "presentation_list": [ 1, 1, 1, 1 ], "contribution_list": [ 1, 1, 1, 1 ], "confidence_list": [ 3, 3, 4, 1 ] }, { "paper_id": "Mr1wsHM9JK", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Carlo Masone", "gender": "Male", "institution": "Polytechnic Institute of Turin", "country": "IT", "position": "Assistant Professor" }, { "name": "Marco Ciccone", "gender": "Male", "institution": "Vector Institute", "country": "CA", "position": "Postdoc" }, { "name": "Riccardo Zaccone", "gender": "Male", "institution": "Polytechnic Institute of Turin", "country": "IT", "position": "PhD student" }, { "name": "Sai Praneeth Karimireddy", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" } ], "n_formula": 102, "n_formula_1": 21, "n_ref_uni": 43, "n_ref": 113, "n_ref_all": 163, "n_fig": 8, "n_tab": 9, "L_tab": 4431, "n_element_tab": 366, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 436, "n_element_tab_1": 45, "formula_len_all": 13708, "formula_len_all_1": 1668, "len_all": 242009, "len_all_1": 69520, "len_abs": 3843, "len_title": 141, "len_sents": 72349, "len_sents_1": 32377, "n_sents": 555, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1148, "title": "Communication-Efficient Heterogeneous Federated Learning with Generalized Heavy-Ball Momentum", "abs": "Federated Learning (FL) has emerged as the state-of-the-art approach for learning from decentralized data in privacy-constrained scenarios. However, system and statistical challenges hinder real-world applications, which demand efficient learning from edge devices and robustness to heterogeneity. Despite significant research efforts, existing approaches (i) are not sufficiently robust, (ii) do not perform well in large-scale scenarios, and (iii) are not communication efficient. In this work, we propose a novel _Generalized Heavy-Ball Momentum_ (GHBM), proving that it enjoys an improved theoretical convergence rate w.r.t. existing FL methods based on classical momentum in _partial participation_, without relying on bounded data heterogeneity. Then, we present FedHBM as an adaptive, communication-efficient by-design instance of GHBM. Extensive experimentation on vision and language tasks, in both controlled and realistic large-scale scenarios, confirms our theoretical findings, showing that GHBM substantially improves the state of the art, especially in large scale scenarios with high data heterogeneity and low client participation.", "keywords": [ "federated learning", "momentum", "distributed learning", "deep learning" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 1, 3, 3 ] }, { "paper_id": "MqvQUP7ZuZ", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Alberto Tono", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Andrew Rodriguez", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Berfin Inal", "gender": "Not Specified", "institution": "University of Amsterdam", "country": "", "position": "MS student" }, { "name": "Biruk Abere Ambaw", "gender": "Male", "institution": "Gondar University", "country": "ET", "position": "Undergrad student" }, { "name": "Gabriele Dominici", "gender": "Male", "institution": "Universita della Svizzera Italiana", "country": "CH", "position": "PhD student" }, { "name": "Meher Shashwat Nigam", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "MS student" }, { "name": "Nursena Koprucu", "gender": "Female", "institution": "Max-Planck Institute", "country": "DE", "position": "Researcher" }, { "name": "Sharvaree Vadgama", "gender": "Female", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Shicheng Luke Xu", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 44, "n_ref": 98, "n_ref_all": 105, "n_fig": 2, "n_tab": 5, "L_tab": 471, "n_element_tab": 67, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 471, "n_element_tab_1": 67, "formula_len_all": 376, "formula_len_all_1": 338, "len_all": 123880, "len_all_1": 56646, "len_abs": 1059, "len_title": 90, "len_sents": 32289, "len_sents_1": 26476, "n_sents": 218, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1063, "title": "DC3DO: Diffusion Classifier for 3D Objects", "abs": "Recent advancements in deep generative models, particularly diffusion models, have shown remarkable capabilities in generating high-fidelity 3D objects. In this work, we explore the application of diffusion models for 3D object classification by integrating the LION model with diffusion-based classifiers. Due to the availability of pretrained model weights, our study focuses on two categories from the ShapeNet dataset: chairs and cars. We propose DC3DO, a method that leverages the generative strengths of diffusion models for domain generalization in 3D classification tasks. Our approach demonstrates improved performance over a multi-view baseline, highlighting the potential of diffusion models in handling 3D data. We also examine the model's ability to generalize to data from different distributions, evaluating its performance on the IFCNet and ModelNet datasets. This study underscores the potential of using diffusion models for 3D object classification and sets the stage for future research involving more categories as resources become available.", "keywords": [ "diffusion", "classifier", "3D", "deep generative models", "classification" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 1, 1, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "MqL2e85ZTp", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 6, "author_info_list": [ { "name": "Agustinus Kristiadi", "gender": "unknown", "institution": "Vector Institute", "country": "DE", "position": "Postdoc" }, { "name": "Ahmad Rashid", "gender": "unknown", "institution": "University of Waterloo", "country": "", "position": "PhD student" }, { "name": "Julia Grosse", "gender": "Female", "institution": "University of Tuebingen", "country": "", "position": "PhD student" }, { "name": "Pascal Poupart", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Full Professor" }, { "name": "Philipp Hennig", "gender": "Male", "institution": "University of T\u00fcbingen", "country": "DE", "position": "Full Professor" }, { "name": "Ruotian Wu", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 17, "n_ref_uni": 54, "n_ref": 78, "n_ref_all": 108, "n_fig": 15, "n_tab": 4, "L_tab": 1884, "n_element_tab": 20, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 529, "n_element_tab_1": 2, "formula_len_all": 838, "formula_len_all_1": 932, "len_all": 162758, "len_all_1": 66460, "len_abs": 1277, "len_title": 116, "len_sents": 47041, "len_sents_1": 30207, "n_sents": 395, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1281, "title": "Uncertainty-Guided Optimization on Large Language Model Search Trees", "abs": "Tree search algorithms such as greedy and beam search are the standard when it comes to finding sequences of maximum likelihood in the decoding processes of large language models (LLMs).\nHowever, they are myopic since they do not take the complete root-to-leaf path into account.\nMoreover, they are agnostic to prior knowledge available about the process:\nFor example, it does not consider that the objective being maximized is a probability and thereby has specific properties like being bound in the unit interval.\nTaking a probabilistic approach, we define prior beliefs over LLMs' transition probabilities and obtain posterior beliefs over the most promising paths in each iteration.\nThese beliefs are useful for defining a sample-based, non-myopic acquisition function that allows for a more data-efficient exploration scheme than standard search algorithms on LLMs.\nCrucially, unlike expensive simulation-based non-myopic methods like the Monte Carlo tree search, our method only requires samples from the beliefs. We discuss how to select the prior and the acquisition function, and demonstrate in experiments with various LLMs that our method achieves higher efficiency than recent baselines:\nOur method achieves the same or a higher likelihood while expanding fewer nodes.", "keywords": [ "LLMs", "Probabilistic Inference", "Tree Search" ], "rating_list": [ 8, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 2, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 4, 2, 3 ], "confidence_list": [ 3, 4, 4, 3, 4 ] }, { "paper_id": "Mq23uJ6sIm", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Dongyoung Lim", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Jonghun Lee", "gender": "unknown", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sungil Kim", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "YongKyung Oh", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" } ], "n_formula": 67, "n_formula_1": 18, "n_ref_uni": 32, "n_ref": 85, "n_ref_all": 119, "n_fig": 7, "n_tab": 27, "L_tab": 28488, "n_element_tab": 3236, "n_fig_1": 7, "n_tab_1": 7, "L_tab_1": 3203, "n_element_tab_1": 188, "formula_len_all": 5614, "formula_len_all_1": 1628, "len_all": 241623, "len_all_1": 65503, "len_abs": 1197, "len_title": 151, "len_sents": 58832, "len_sents_1": 26521, "n_sents": 490, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1199, "title": "Neural Regenerative Stochastic Differential Equation: Dropout Scheme for Neural Differential Equations", "abs": "Neural Differential Equations (NDEs) are an excellent tool for modeling continuous-time (stochastic) dynamics, effectively handling challenges such as irregular observations, missing values, and noise. Despite their advantages, there is a lack of regularization techniques in the NDE framework, particularly those like dropout, which have been successfully implemented in other discrete neural networks, making them susceptible to overfitting. To address this research gap, we introduce Neural Regenerative Stochastic Differential Equation (NRSDE), based on alternating renewal processes, as a universally applicable regularization technique for NDEs. Our study reveals that NRSDE can effectively represent a continuous approximation of neural networks that randomly deactivate some neurons during training, similar to dropout, thereby enhancing the robustness and generalization capabilities of NDEs. Through extensive experiments, we demonstrate that NRSDE outperforms existing regularization methods for NDEs and can be applied to all existing NDE models, significantly improving their performance across various deep learning tasks, including time series classification and image classification.", "keywords": [ "neural differential equations", "neural stochastic differential equations", "dropout", "regularization", "renewal process" ], "rating_list": [ 6, 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 4, 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 4, 3 ] }, { "paper_id": "MpuMza23aL", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Jiayu Zhou", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Associate Professor" }, { "name": "Siqi Liang", "gender": "Female", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Sumyeong Ahn", "gender": "Male", "institution": "KENTECH", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 18, "n_ref": 43, "n_ref_all": 66, "n_fig": 7, "n_tab": 9, "L_tab": 4438, "n_element_tab": 191, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 937, "n_element_tab_1": 61, "formula_len_all": 745, "formula_len_all_1": 658, "len_all": 167734, "len_all_1": 60075, "len_abs": 1484, "len_title": 115, "len_sents": 36080, "len_sents_1": 26642, "n_sents": 300, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1442, "title": "Robust Prompt Learning For Vision-Language Models With Noisy Labels", "abs": "Recent advancements in vision-language models (VLMs), designed for simultaneous comprehension of vision and language, have demonstrated significant success in achieving zero-shot classification capabilities. However, despite their impressive performance, it is widely acknowledged that fine-tuning is essential to adapt these models to new target tasks. This adaptation process requires the collection of target datasets, which may introduce incorrect labels and greatly compromise the model performance after fine-tuning. In this paper, our objective is to enhance classification fine-tuning performance by leveraging the zero-shot classification capability under a noisy labeled training dataset. We first conduct a detailed exploration of the behavior of the pre-trained VLMs under various classification text prompts, including human-crafted and LLM-crafted visual characteristics. This investigation reveals that VLMs have tilted knowledge towards some classes, and each prompt exhibits varying expertise for each class. Based on these observations, we introduce a robust training method called PoND, which employs a complementary approach across different types of prompts, leveraging the expertise of each class. We systematically compare the efficacy of the proposed algorithm with existing denoising techniques designed for VLMs and substantiate that our proposed algorithm outperforms prior approaches across 11 real-world datasets.", "keywords": [ "Vision Language Models", "Prompt Learning", "Noisy Labels" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "Mpp6SakVzl", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Bhushan Gopaluni", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Professor" }, { "name": "Michael Forbes", "gender": "Male", "institution": "Honeywell", "country": "CA", "position": "Researcher" }, { "name": "Philip D Loewen", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Full Professor" }, { "name": "Shuyuan Wang", "gender": "unknown", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Wei Pan", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "Associate Professor" } ], "n_formula": 33, "n_formula_1": 20, "n_ref_uni": 32, "n_ref": 57, "n_ref_all": 83, "n_fig": 5, "n_tab": 3, "L_tab": 373, "n_element_tab": 73, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 145, "n_element_tab_1": 43, "formula_len_all": 2184, "formula_len_all_1": 1579, "len_all": 135149, "len_all_1": 62009, "len_abs": 1634, "len_title": 51, "len_sents": 33669, "len_sents_1": 25750, "n_sents": 288, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1675, "title": "DiLQR: Differentiable Iterative Linear Quadratic Regulator", "abs": "Differentiable control promises end-to-end differentiability and adaptability, effectively combining the advantages of both model-free and model-based control\napproaches. However, the iterative Linear Quadratic Regulator (iLQR), despite\nbeing a powerful nonlinear controller, still lacks differentiable capabilities. The\nscalability of differentiating through extended iterations and horizons poses signifi\ncant challenges, hindering iLQR from being an effective differentiable controller.\nThis paper introduces a framework that facilitates differentiation through iLQR,\nallowing it to serve as a trainable and differentiable module, either as or within\na neural network. for control purposes. A novel aspect of this framework is the\nanalytical solution that it provides for the gradient of an iLQR controller through\nimplicit differentiation, which ensures a constant backward cost regardless of iteration, while producing an accurate gradient. We evaluate our framework on\nimitation tasks on famous control benchmarks. Our analytical method demonstrates superior computational performance, achieving up to $\\mathbf{128x}$ speedup and\na minimum of $\\mathbf{21x}$ speedup compared to automatic differentiation. Our method\nalso demonstrates superior learning performance ($\\mathbf{10^6}$x) compared to traditional\nneural network policies and better model loss with differentiable controllers that\nlack exact analytical gradients. Furthermore, we integrate our module into a\nlarger network with visual inputs to demonstrate the capacity of our method for\nhigh-dimensional, fully end-to-end tasks. Codes can be found on the project\nhomepage https://sites.google.com/view/dilqr/.", "keywords": [ "Differentiable control;iLQR;IL" ], "rating_list": [ 3, 1, 6 ], "soundness_list": [ 1, 1, 4 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 3, 4, 1 ] }, { "paper_id": "MpCxUF8x61", "primary_area": "foundation or frontier models, including LLMs", "n_author": 20, "author_info_list": [ { "name": "benyou wang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Chaojun Wang", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Dongdong Zhang", "gender": "Male", "institution": "Microsoft Research Asia", "country": "", "position": "Researcher" }, { "name": "Furu Wei", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Partner Research Manager" }, { "name": "Haoran Li", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "", "position": "PhD student" }, { "name": "Haoyang Huang", "gender": "Male", "institution": "Microsoft Research Asia", "country": "", "position": "FTE" }, { "name": "Li Dong", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Principal Researcher" }, { "name": "Qingxiu Dong", "gender": "Female", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Shaohan Huang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Si-Qing Chen", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Partner Applied Science Manager" }, { "name": "Wai Lam", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Professor" }, { "name": "Xiaolong Huang", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Xin Cheng", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Xingxing Zhang", "gender": "Male", "institution": "Microsoft Research Asia", "country": "", "position": "Researcher" }, { "name": "Xun Wang", "gender": "Not Specified", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Yuxian Gu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zeqiang Huang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Zhengyang Tang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Zhifang Sui", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Wei Lu", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 70, "n_ref_all": 91, "n_fig": 2, "n_tab": 11, "L_tab": 7579, "n_element_tab": 1046, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1334, "n_element_tab_1": 184, "formula_len_all": 357, "formula_len_all_1": 103, "len_all": 173874, "len_all_1": 65367, "len_abs": 1439, "len_title": 136, "len_sents": 39922, "len_sents_1": 30310, "n_sents": 324, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1469, "title": "Synthetic Data (Almost) from Scratch: Generalized Instruction Tuning for Language Models", "abs": "We introduce Generalized Instruction Tuning (called GLAN), a general and scalable method for instruction tuning of Large Language Models (LLMs). \nUnlike prior work that relies on seed examples or existing datasets to construct instruction-tuning data, \nGLAN exclusively utilizes a pre-curated taxonomy of human knowledge and capabilities as input and generates large-scale \nsynthetic instruction data across all disciplines.\nSpecifically, inspired by the systematic structure in human education system, we build the taxonomy by decomposing human knowledge and capabilities to various fields, sub-fields and ultimately, distinct disciplines semi-automatically, facilitated by LLMs. \nSubsequently, we generate a comprehensive list of subjects for every discipline and proceed to design a syllabus tailored to each subject, again utilizing LLMs.\nWith the fine-grained key concepts detailed in every class session of the syllabus, we are able to generate diverse instructions with a broad coverage across the entire spectrum of human knowledge and skills. \nExtensive experiments on large language models (e.g., Mistral) demonstrate that GLAN excels in multiple dimensions from mathematical reasoning, coding, academic exams, logical reasoning to general instruction following without using task-specific training data of these tasks. In addition, GLAN allows for easy customization and new\nfields or skills can be added by simply incorporating a new node into our taxonomy.", "keywords": [ "Synthetic Data", "Large Language Model", "Instruction Tuning", "Human Education Inspired Method" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "MpA6HMD7Wq", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Jakob Nicolaus Foerster", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Jaron Cohen", "gender": "Male", "institution": "InstaDeep", "country": "", "position": "Intern" }, { "name": "Shimon Whiteson", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "Professor" }, { "name": "Alex Goldie", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 57, "n_ref": 151, "n_ref_all": 160, "n_fig": 4, "n_tab": 3, "L_tab": 3546, "n_element_tab": 83, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 289, "n_element_tab_1": 32, "formula_len_all": 43, "formula_len_all_1": 43, "len_all": 224909, "len_all_1": 66462, "len_abs": 1685, "len_title": 131, "len_sents": 47553, "len_sents_1": 32951, "n_sents": 368, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1630, "title": "Do Symbolic or Black-Box Representations Generalise Better In Learned Optimisation?", "abs": "Until recently, behind every algorithmic advance in machine learning was a human researcher. Now, however, algorithms can be meta-learned automatically, with little human input. However, to be truly useful, such algorithms must generalise beyond their training distribution. This is especially challenging in reinforcement learning (RL), where transferring algorithms between environments with vastly different dynamics is difficult and training on diverse environments often requires prohibitively expensive large-scale data collection.Learned optimisation is a branch of algorithmic discovery that meta-learns optimiser update rules. Learned optimisers can be classified into two groups: black-box algorithms, where the optimiser is a neural network; or symbolic algorithms, where the optimiser is represented using mathematical functions or code. While some claim that symbolic algorithms generalise better than black-box ones, testing such assertions is complicated by the fact that symbolic algorithms typically include additional hyperparameters, and thus their evaluation is done many-shot. This is an unfair comparison with the zero-shot evaluation of black-box optimisers. In this work, we build a pipeline to discover symbolic optimisers which are hyperparameter-free, enabling a fair comparison of the generalisation of symbolic optimisers with that of an open-source state-of-the-art black-box optimiser trained for RL. Based on our analysis, we propose suggestions to improve the symbolic optimiser discovery pipeline for RL, with an overall objective of reducing the need for hyperparameter tuning to train an agent.", "keywords": [ "reinforcement learning", "meta learning", "optimisation" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "MoJSnVZ59d", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Byoungjip Kim", "gender": "Male", "institution": "KB", "country": "KR", "position": "Principal Researcher" }, { "name": "Geon-hyeong Kim", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Researcher" }, { "name": "Kyunghoon Bae", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Principal Researcher" }, { "name": "Moontae Lee", "gender": "unknown", "institution": "LG Corporation", "country": "KR", "position": "Director" }, { "name": "Youngsoo Jang", "gender": "unknown", "institution": "LG AI Research", "country": "", "position": "Researcher" }, { "name": "Yu Jin Kim", "gender": "unknown", "institution": "LG AI Research", "country": "", "position": "Researcher" }, { "name": "Honglak Lee", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Full Professor" } ], "n_formula": 40, "n_formula_1": 24, "n_ref_uni": 37, "n_ref": 85, "n_ref_all": 133, "n_fig": 6, "n_tab": 18, "L_tab": 5298, "n_element_tab": 444, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 378, "n_element_tab_1": 4, "formula_len_all": 4373, "formula_len_all_1": 1742, "len_all": 206821, "len_all_1": 65312, "len_abs": 1306, "len_title": 130, "len_sents": 60533, "len_sents_1": 30878, "n_sents": 489, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1319, "title": "SafeDPO: A Simple Approach to Direct Preference Optimization with Enhanced Safety", "abs": "As large language models (LLMs) continue to advance and find applications across a growing number of fields, ensuring the safety of LLMs has become increasingly critical. To address safety concerns, recent studies have proposed integrating safety constraints into reinforcement learning from human feedback (RLHF). However, these approaches tend to be complex and often unstable, as they encompass complicated procedures in RLHF along with additional procedures required by the safety constraints. Inspired by direct preference optimization (DPO), we introduce a new algorithm called \\textit{SafeDPO}, which is designed to implicitly optimize the safety alignment objective within a single stage of policy learning. The resulting algorithm can be implemented by introducing only one additional hyperparameter, which aims to further enhance safety, along with minor modifications to the DPO implementation. Consequently, SafeDPO successfully eliminates the necessity of fitting a reward and a cost model, as well as sampling from the language model during fine-tuning, while still enhancing the safety of LLMs. Finally, we demonstrate that SafeDPO achieves competitive performance compared to the current state-of-the-art safety alignment algorithm, both in terms of aligning with human preferences and improving safety.", "keywords": [ "Safety Alignment", "LLM Fine-tuning", "Preferences", "Large Language Models", "AI Safety" ], "rating_list": [ 6, 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 2, 2 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 2, 4, 4 ] }, { "paper_id": "MnfHxPP5gs", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Alexander Bukharin", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Daniel Egert", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Gerald Shen", "gender": "unknown", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Oleksii Kuchaiev", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Olivier Delalleau", "gender": "Male", "institution": "NVIDIA", "country": "CA", "position": "Researcher" }, { "name": "Yi Dong", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Zhilin Wang", "gender": "unknown", "institution": "NVIDIA", "country": "", "position": "Applied Scientist" }, { "name": "Gaby Zeng", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 34, "n_ref": 86, "n_ref_all": 109, "n_fig": 3, "n_tab": 9, "L_tab": 863, "n_element_tab": 70, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2477, "n_element_tab_1": 287, "formula_len_all": 285, "formula_len_all_1": 151, "len_all": 228452, "len_all_1": 69624, "len_abs": 5308, "len_title": 110, "len_sents": 76486, "len_sents_1": 33867, "n_sents": 563, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1440, "title": "HelpSteer2-Preference: Complementing Ratings with Preferences", "abs": "Reward models are critical for aligning models to follow instructions, and are typically trained following one of two popular paradigms: Bradley-Terry style or Regression style. However, there is a lack of evidence that either approach is better than the other, when adequately matched for data. This is primarily because these approaches require data collected in different (but incompatible) formats, meaning that adequately matched data is not available in existing public datasets. To tackle this problem, we release preference annotations (designed for Bradley-Terry training) to complement existing ratings (designed for Regression style training) in the HelpSteer2 dataset. To improve data interpretability, preference annotations are accompanied with human-written justifications. Using this data, we conduct the first head-to-head comparison of Bradley-Terry and Regression models when adequately matched for data. Based on insights derived from such a comparison, we propose a novel approach to combine Bradley-Terry and Regression reward modeling. A Llama-3.1-70B-Instruct model tuned with this approach scores 94.1 on RewardBench, emerging top of more than 140 reward models as of 1 Oct 2024. \nThis reward model can then be used with REINFORCE to align a model to reach 85.0 on Arena Hard, which is No. 1 as of 1 Oct 2024.\nWe open-source this dataset (CC-BY-4.0 license) and openly release the trained reward and aligned models.", "keywords": [ "reward modelling", "rlhf", "model alignment" ], "rating_list": [ 6, 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 4, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "MnJzJ2gvuf", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Chengzhuo Tong", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Dongzhi Jiang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Jiaming Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Renrui Zhang", "gender": "Male", "institution": "MMLab of CUHK & Shanghai AI Laboratory", "country": "HK", "position": "PhD student" }, { "name": "Xinyu Wei", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yichi Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Ziyu Guo", "gender": "Female", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Aojun Zhou", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Researcher" }, { "name": "Gao Peng", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Shanghang Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 97, "n_ref_all": 140, "n_fig": 23, "n_tab": 16, "L_tab": 5289, "n_element_tab": 987, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1446, "n_element_tab_1": 399, "formula_len_all": 327, "formula_len_all_1": 0, "len_all": 249062, "len_all_1": 58690, "len_abs": 2188, "len_title": 123, "len_sents": 67867, "len_sents_1": 24674, "n_sents": 516, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 2029, "title": "MAVIS: Mathematical Visual Instruction Tuning with an Automatic Data Engine", "abs": "Multi-modal Large Language Models (MLLMs) have recently showcased superior proficiency in general visual scenarios. However, we identify their mathematical capabilities remain under-explored with three areas to be improved: visual encoding of math diagrams, diagram-language alignment, and chain-of-thought (CoT) reasoning. This draws forth an urgent demand for an effective training paradigm and a large-scale, comprehensive dataset with detailed CoT rationales, which is challenging to collect and costly to annotate manually. To tackle this issue, we propose MAVIS, a MAthematical VISual instruction tuning pipeline for MLLMs, featuring an automatic data engine to efficiently create mathematical visual datasets.\nWe design the data generation process to be entirely independent of human intervention or GPT API usage, while ensuring the diagram-caption correspondence, question-answer correctness, and CoT reasoning quality. With this approach, we curate two datasets, MAVIS-Caption (558K diagram-caption pairs) and MAVIS-Instruct (834K visual math problems with CoT rationales), and propose four progressive stages for training MLLMs from scratch.\nFirst, we utilize MAVIS-Caption to fine-tune a math-specific vision encoder (CLIP-Math) through contrastive learning, tailored for improved diagram visual encoding. Second, we also leverage MAVIS-Caption to align the CLIP-Math with a large language model (LLM) by a projection layer, enhancing vision-language alignment in mathematical domains. Third, we adopt MAVIS-Instruct to perform the instruction tuning for robust problem-solving skills, and term the resulting model as MAVIS-7B. Fourth, we apply Direct Preference Optimization (DPO) to enhance the CoT capabilities of our model, further refining its step-wise reasoning performance.\nOn various mathematical benchmarks, our MAVIS-7B achieves leading results among open-source MLLMs, e.g., surpassing other 7B models by +9.3% and the second-best LLaVA-NeXT (110B) by +6.9%, demonstrating the effectiveness of our method.", "keywords": [ "Large language model", "multimodal learning", "mathematics" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "MnE8iIBCfO", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Jonas Geiping", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen", "country": "DE", "position": "Principal Researcher" }, { "name": "Micah Goldblum", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" }, { "name": "Tom Goldstein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Vasu Singla", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Pedro Sandoval Segura", "gender": "Male", "institution": "University of Maryland", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 33, "n_ref": 98, "n_ref_all": 128, "n_fig": 10, "n_tab": 4, "L_tab": 1091, "n_element_tab": 72, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 760, "n_element_tab_1": 34, "formula_len_all": 127, "formula_len_all_1": 127, "len_all": 134867, "len_all_1": 59766, "len_abs": 1233, "len_title": 57, "len_sents": 42570, "len_sents_1": 28833, "n_sents": 319, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1293, "title": "Data Brittleness Estimation with Self-Supervised Features", "abs": "To what extent are model predictions sensitive to modifications in training data? Data attribution approaches have served to answer this question. These approaches can be used for estimating data brittleness i.e., identifying which subset of training samples had the highest positive influence on a test sample. However, these methods come at a high computational cost, are memory intensive, and are hard to scale to large models or datasets. Current state-of-the-art approaches require an ensemble of as many as \\textbf{300,000 models}. In this work, we focus on a computationally efficient baseline centered on estimating two types of data brittleness metrics. Our baseline approach uses the image features from a \\textbf{single} pretrained self-supervised backbone. In contrast to data attribution approaches, our method is model-agnostic based on the intuition that different models leverage data in similar ways. Our results show this simple assumption works well in practice, achieving competitive performance with state-of-the-art attribution approaches on CIFAR-10 and ImageNet, under limited computational and memory requirements. Our work serves as a simple baseline showing that effective data brittleness estimates can be achieved based solely using knowledge of the training data.", "keywords": [ "data attribution", "data brittleness estimation" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "MnBrLJez3q", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Chris van der Heide", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Researcher" }, { "name": "Kenyon Ng", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" }, { "name": "Liam Hodgkinson", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Lecturer" }, { "name": "Susan Wei", "gender": "Female", "institution": "Monash University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 47, "n_formula_1": 16, "n_ref_uni": 56, "n_ref": 144, "n_ref_all": 190, "n_fig": 6, "n_tab": 9, "L_tab": 4295, "n_element_tab": 745, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 378, "n_element_tab_1": 75, "formula_len_all": 3985, "formula_len_all_1": 841, "len_all": 222751, "len_all_1": 69632, "len_abs": 1172, "len_title": 99, "len_sents": 65256, "len_sents_1": 32382, "n_sents": 537, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1175, "title": "Temperature Optimization for Bayesian Deep Learning", "abs": "The Cold Posterior Effect (CPE) is a phenomenon in Bayesian Deep Learning (BDL), where tempering the posterior to a cold temperature often improves the predictive performance of the posterior predictive distribution (PPD). Although the term `CPE' suggests colder temperatures are inherently better, the BDL community increasingly recognizes that this is not always the case. Despite this, there remains no systematic method for finding the optimal temperature beyond grid search. In this work, we propose a data-driven approach to select the temperature that maximizes test log-predictive density, treating the temperature as a model parameter and estimating it directly from the data. We empirically demonstrate that our method performs comparably to grid search, at a fraction of the cost, across both regression and classification tasks. Finally, we highlight the differing perspectives on CPE between the BDL and Generalized Bayes communities: while the former primarily emphasizes the predictive performance of the PPD, the latter prioritizes the utility of the posterior under model misspecification; these distinct objectives lead to different temperature preferences.", "keywords": [ "Bayesian deep learning", "cold posterior effect", "temperature selection", "posterior tempering" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 3, 3, 1, 4 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "Mn2qgIcIPS", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Daehyun Kim", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "PhD student" }, { "name": "Donggoo Jung", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "PhD student" }, { "name": "Tae Hyun Kim", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "Associate Professor" } ], "n_formula": 40, "n_formula_1": 23, "n_ref_uni": 38, "n_ref": 138, "n_ref_all": 194, "n_fig": 17, "n_tab": 11, "L_tab": 6200, "n_element_tab": 870, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 3807, "n_element_tab_1": 505, "formula_len_all": 2150, "formula_len_all_1": 1138, "len_all": 185191, "len_all_1": 70488, "len_abs": 1439, "len_title": 126, "len_sents": 53013, "len_sents_1": 28740, "n_sents": 404, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 78, "L_abs": 1410, "title": "Continuous Exposure Learning for Low-light Image Enhancement using Neural ODEs", "abs": "Low-light image enhancement poses a significant challenge due to the limited information captured by image sensors in low-light environments. \n Despite recent improvements in deep learning models, the lack of paired training datasets remains a significant obstacle. \n Therefore, unsupervised methods have emerged as a promising solution. \n In this work, we focus on the strength of curve-adjustment-based approaches to tackle unsupervised methods. \n The majority of existing unsupervised curve-adjustment approaches iteratively estimate higher order curve parameters to enhance the exposure of images while efficiently preserving the details of the images. \n However, the convergence of the enhancement procedure cannot be guaranteed, leading to sensitivity to the number of iterations and limited performance.\n To address this problem, we consider the iterative curve-adjustment update process as a dynamic system and formulate it as a Neural Ordinary Differential Equations (NODE) for the first time, and this allows us to learn a continuous dynamics of the latent image. \n The strategy of utilizing NODE to leverage continuous dynamics in iterative methods enhances unsupervised learning and aids in achieving better convergence compared to discrete-space approaches. Consequently, we achieve state-of-the-art performance in unsupervised low-light image enhancement across various benchmark datasets.", "keywords": [ "NeuralODE", "Low-light Enhancement" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "MmzZyHG1Te", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jun Cheng", "gender": "Male", "institution": "Institute For Infocomm Research, A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Lile Cai", "gender": "Female", "institution": "Institute for Infocomm Research, A*STAR", "country": "SG", "position": "Scientist" }, { "name": "Xulei Yang", "gender": "Male", "institution": "I2R, A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Yanyu Xu", "gender": "Male", "institution": "Shandong University", "country": "CN", "position": "Full Professor" }, { "name": "ZAIWANG GU", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Senior Researcher Engineer" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 56, "n_ref": 97, "n_ref_all": 108, "n_fig": 2, "n_tab": 7, "L_tab": 1535, "n_element_tab": 234, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 3556, "n_element_tab_1": 177, "formula_len_all": 277, "formula_len_all_1": 278, "len_all": 129893, "len_all_1": 60980, "len_abs": 1133, "len_title": 142, "len_sents": 29979, "len_sents_1": 26572, "n_sents": 221, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1153, "title": "Harmonious convergence for confidence estimation in depth estimation and completion", "abs": "Confidence estimation for monocular depth estimation and completion is important for their deployment in real-world applications. Recent models for confidence estimation in these regression tasks mainly rely on the statistical characteristics of training and test data, while ignoring the information from the model training. We propose a harmonious convergence estimation approach for confidence estimation in the regression tasks, taking training consistency into consideration. Specifically, we propose an intra-batch convergence estimation algorithm with two sub-iterations to compute the training consistency for confidence estimation. A harmonious convergence loss is newly designed to encourage the consistency between confidence measure and depth prediction. Our experimental results on the NYU2 and KITTI datasets show improvements ranging from 10.91\\% to 43.90\\% across different settings in monocular depth estimation, and from 27.91\\% to 45.24\\% in depth completion, measured by Pearson correlation coefficients, justifying the effectiveness of the proposed method. We will release all the codes upon the publication of our paper.", "keywords": [ "confidence estimation", "monocular depth estimation", "depth completion" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 2, 3 ] }, { "paper_id": "MmWkNmeDNE", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Bernd Rosenow", "gender": "Male", "institution": "Universit\u00e4t Leipzig", "country": "", "position": "Full Professor" }, { "name": "Matthias Thamm", "gender": "Male", "institution": "Universit\u00e4t Leipzig", "country": "DE", "position": "Postdoc" }, { "name": "Max Staats", "gender": "Male", "institution": "Center for Scalable Data Analytics and Artificial Intelligence", "country": "DE", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 30, "n_ref": 62, "n_ref_all": 83, "n_fig": 14, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 565, "formula_len_all_1": 513, "len_all": 135002, "len_all_1": 66333, "len_abs": 1312, "len_title": 118, "len_sents": 46874, "len_sents_1": 33321, "n_sents": 345, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1324, "title": "Locating Information in Large Language Models via Random Matrix Theory", "abs": "As large language models (LLMs) become central to AI applications, gaining a deeper understanding of their inner workings is increasingly important. In this work, we analyze the weight matrices of pretrained transformer models -- specifically BERT and Llama -- using random matrix theory (RMT) as a zero-information hypothesis. While randomly initialized weights perfectly agree with RMT predictions, deviations emerge after training, allowing us to locate learned structures within the models. We identify layer-type specific behaviors that are consistent across all blocks and architectures considered. By pinpointing regions that deviate from RMT predictions, we highlight areas of feature learning and confirm this through comparisons with the activation covariance matrices of the corresponding layers. Our method provides a diagnostic tool for identifying relevant regions in transformer weights using only the trained matrices. Additionally, we address the ongoing debate regarding the significance of small singular values in the context of finetuning and alignment in LLMs. Our findings reveal that, after finetuning, small singular values play a crucial role in the models' capabilities, suggesting that removing them in an already aligned transformer can be detrimental, as it may compromise model alignment.", "keywords": [ "Random matrix theory", "RMT", "Llama 3", "Spectra", "singular value decomposition" ], "rating_list": [ 6, 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "MmOQY71YHw", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Chunyu Liu", "gender": "Female", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Jianjun Yu", "gender": "Male", "institution": "CNIC,CAS", "country": "CN", "position": "Full Professor" }, { "name": "Shijie Li", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Xulei Yang", "gender": "Male", "institution": "I2R, A*STAR", "country": "SG", "position": "Principal Researcher" } ], "n_formula": 19, "n_formula_1": 20, "n_ref_uni": 38, "n_ref": 77, "n_ref_all": 83, "n_fig": 4, "n_tab": 6, "L_tab": 3375, "n_element_tab": 209, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2063, "n_element_tab_1": 135, "formula_len_all": 736, "formula_len_all_1": 734, "len_all": 127161, "len_all_1": 60830, "len_abs": 1680, "len_title": 109, "len_sents": 31516, "len_sents_1": 25912, "n_sents": 236, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1691, "title": "MS$^3$M: Multi-Stage State Space Model for Motion Forecasting", "abs": "Motion forecasting is a fundamental component of autonomous driving systems, as it predicts an agent's future trajectories based on its surrounding environment. Transformer architectures have dominated this domain due to their strong ability to model both temporal and spatial information. However, transformers often suffer from quadratic complexity with respect to input sequence length, limiting their ability to efficiently process scenarios involving numerous agents. Additionally, transformers typically rely on positional encodings to represent temporal or spatial relationships, a strategy that may not be as effective or intuitive as the inductive biases naturally embedded in convolutional architectures. To address these challenges, we leverage recent advancements in state space models (SSMs) and propose the Multi-Stage State Space Model (MS$^3$M). In MS$^3$M, the Temporal Mamba Model (TMM) is employed to capture fine-grained temporal information, while the Spatial Mamba Model efficiently handles spatial interactions. By injecting temporal and spatial inductive biases through Mamba\u2019s state-space model structure, the model's capacity is significantly improved. MS$^3$M also strikes an exceptional trade-off between accuracy and efficiency, which is achieved through convolutional computations and near-linear computational strategies in the Mamba architecture. Furthermore, a hierarchical query-based decoder is introduced, further enhancing model performance and efficiency. Extensive experimental results demonstrate that the proposed method achieves superior performance while maintaining low latency, which is crucial for practical real-time autonomous driving systems.", "keywords": [ "Motion Forecasting", "Autonomous Driving", "State Space Model" ], "rating_list": [ 3, 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 5, 3, 5, 4, 1 ] }, { "paper_id": "Mlxov4A7AE", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Long Tung Vuong", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Trung Le", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" }, { "name": "Van-Anh Nguyen", "gender": "Female", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Vy Vo", "gender": "Female", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Dinh Phung", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" }, { "name": "Mehrtash T. Harandi", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Associate Professor" }, { "name": "Thanh-Toan Do", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" } ], "n_formula": 71, "n_formula_1": 14, "n_ref_uni": 63, "n_ref": 183, "n_ref_all": 212, "n_fig": 4, "n_tab": 13, "L_tab": 1990, "n_element_tab": 181, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 418, "n_element_tab_1": 20, "formula_len_all": 5273, "formula_len_all_1": 1117, "len_all": 248984, "len_all_1": 74683, "len_abs": 2611, "len_title": 120, "len_sents": 69061, "len_sents_1": 33261, "n_sents": 515, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1115, "title": "Understanding Domain Generalization: A View of Necessity and Sufficiency", "abs": "Despite the rapid advancements in domain generalization (DG), the majority of DG studies center on establishing theoretical guarantee for generalization under the assumption of sufficient, diverse or even infinite domains. This assumption however is unrealistic, thus there remains no conclusive evidence as to whether the existing DG algorithms can truly generalize in practical settings where domains are limited. This paper aims to elucidate this matter. We first study the conditions for the existence and learnability of an optimal hypothesis. As the sufficient conditions are non-verifiable, our identified two necessary conditions become critical to guaranteeing the chance of finding the global optimal hypothesis in finite domain settings. In light of the theoretical insights, we provide a comprehensive review of DG algorithms explaining to what extent they can generalize effectively. We finally introduce a practical approach that leverages the joint effect of the two sets of conditions to boost generalization. Our proposed method demonstrates superior performance on well-established DG benchmarks.", "keywords": [ "Domain Generalization" ], "rating_list": [ 5, 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 1, 2, 2 ], "presentation_list": [ 2, 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 4, 3 ] }, { "paper_id": "MlxeUVCQgD", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Chen Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Chengjie Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Chengming Xu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Donghao Luo", "gender": "Male", "institution": "Tencent YouTu Lab", "country": "CN", "position": "researcher" }, { "name": "Kai Hu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Lingfan Zhang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Yanwei Fu", "gender": "Male", "institution": "Fudan University,", "country": "CN", "position": "Professor" }, { "name": "Yuan Yao", "gender": "Unspecified", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 14, "n_ref": 22, "n_ref_all": 37, "n_fig": 7, "n_tab": 8, "L_tab": 2995, "n_element_tab": 181, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 673, "n_element_tab_1": 99, "formula_len_all": 1543, "formula_len_all_1": 1471, "len_all": 119898, "len_all_1": 65571, "len_abs": 1281, "len_title": 108, "len_sents": 38751, "len_sents_1": 28948, "n_sents": 309, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1289, "title": "Combating inherent noise for direct preference optimization", "abs": "Direct Preference Optimization (DPO) has recently gained traction as a promising approach to align large models with human feedback. It is notable for its effectiveness and ease of application across various models, including Large Language Models (LLMs) and Diffusion Models (DMs). However, the quality of preference data used in DPO training has been largely overlooked. Current datasets, whether annotated by deep learning metrics or crowd-sourced human judgments, often contain noisy labels. This noise can adversely affect the performance of DPO. \nTo address this issue, we propose a novel approach that incorporates a noise-aware metric into the DPO objective. This metric, which includes intra-annotator confidence and inter-annotator stability, helps identify and mitigate the impact of noisy data. We introduce an Adaptive-DPO loss function which improves the DPO loss in two ways: one aims to reduce the influence of noisy samples, while the other is to amplify the impact of clean samples. Our experiments demonstrate that this method effectively handles both synthetic and natural noisy data, leading to improved performance in visual and textual generation tasks. This underscores the practical value of our approach in enhancing model robustness amidst noisy preference data.", "keywords": [ "Direct Preference Optimization" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "Mjn53GtMxi", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Fei Fang", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Stephanie Milani", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zhuorui Ye", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Geoff Gordon", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 38, "n_ref": 55, "n_ref_all": 91, "n_fig": 9, "n_tab": 5, "L_tab": 2026, "n_element_tab": 224, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 262, "n_element_tab_1": 35, "formula_len_all": 552, "formula_len_all_1": 265, "len_all": 171675, "len_all_1": 62380, "len_abs": 1659, "len_title": 124, "len_sents": 62132, "len_sents_1": 31054, "n_sents": 458, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1684, "title": "LICORICE: Label-Efficient Concept-Based Interpretable Reinforcement Learning", "abs": "Recent advances in reinforcement learning (RL) have predominantly leveraged neural network-based policies for decision-making, yet these models often lack interpretability, posing challenges for stakeholder comprehension and trust. Concept bottleneck models offer an interpretable alternative by integrating human-understandable concepts into neural networks. However, a significant limitation in prior work is the assumption that annotations for these concepts are readily available during training, necessitating continuous real-time concept annotation. This reliance either places a significant burden on human annotators or incurs substantial costs in API queries and inference time when employing automated labeling methods, such as vision-language models (VLMs). To overcome this limitation, we introduce a novel training scheme that enables RL algorithms to efficiently learn a concept-based policy by only querying annotators to label a small set of data. Our algorithm, LICORICE, involves three main contributions: interleaving concept learning and RL training, using a concept ensembles to actively select informative data points for labeling, and decorrelating the concept data with a simple strategy. We show how LICORICE reduces human labeling efforts to 500 or fewer concept labels in three environments and 5000 in another complex environment at minimal or no cost to performance. We also explore the use of VLMs as automated concept annotators, finding them effective in some cases but challenging in others. This work significantly reduces the annotation burden for interpretable RL, making it more practical for real-world applications where transparency is crucial.", "keywords": [ "Reinforcement Learning", "Explainable Reinforcement Learning", "Concept Bottleneck Models", "Concept-based Explainability", "Interpretability", "XRL" ], "rating_list": [ 6, 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 4, 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4, 3 ] }, { "paper_id": "MjhTb4gwFP", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Bing Deng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Hualian Sheng", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Jieping Ye", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Jinhua Zhang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Qiao Liang", "gender": "unknown", "institution": "DAMO Academy, Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Shuhang Gu", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Wen Li", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Ying Fu", "gender": "Female", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Sijia Cai", "gender": "Male", "institution": "Alibaba Cloud", "country": "CN", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 108, "n_ref_all": 128, "n_fig": 13, "n_tab": 9, "L_tab": 3086, "n_element_tab": 331, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1451, "n_element_tab_1": 136, "formula_len_all": 953, "formula_len_all_1": 266, "len_all": 138258, "len_all_1": 57837, "len_abs": 1274, "len_title": 134, "len_sents": 44192, "len_sents_1": 27549, "n_sents": 281, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1305, "title": "PerLDiff: Controllable Street View Synthesis Using Perspective-Layout Diffusion Model", "abs": "Controllable generation is considered a potentially vital approach to address the challenge of annotating 3D data, and the precision of such controllable generation becomes particularly imperative in the context of data production for autonomous driving. Existing methods focus on the integration of diverse generative information into controlling inputs, utilizing frameworks such as GLIGEN or ControlNet, to produce commendable outcomes in controllable generation. However, such approaches intrinsically restrict generation performance to the learning capacities of predefined network architectures. In this paper, we explore the integration of controlling information and introduce PerLDiff (\\textbf{Per}spective-\\textbf{L}ayout \\textbf{Diff}usion Models), a method for effective street view image generation that fully leverages perspective 3D geometric information. Our PerLDiff employs 3D geometric priors to guide the generation of street view images with precise object-level control within the network learning process, resulting in a more robust and controllable output. Moreover, it demonstrates superior controllability compared to alternative layout control methods. Empirical results justify that our PerLDiff markedly enhances the precision of generation on the NuScenes and KITTI datasets.", "keywords": [ "Controllable generation;3D annotation;Geometric priors;" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "MjR5LcAGXJ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Arnab Chakraborty", "gender": "Male", "institution": "Fidelity Investments", "country": "US", "position": "Principal Researcher" }, { "name": "J. Jeffrey Brown", "gender": "Male", "institution": "Fidelity Investments", "country": "US", "position": "Principal Researcher" }, { "name": "Mohammadreza Soltani", "gender": "Male", "institution": "Fidelity Investments", "country": "", "position": "Principal Researcher" }, { "name": "William McCormack", "gender": "Male", "institution": "Fidelity Investments", "country": "US", "position": "Researcher" }, { "name": "Zachary semenov", "gender": "unknown", "institution": "Fidelity Investments", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 42, "n_ref": 71, "n_ref_all": 96, "n_fig": 7, "n_tab": 14, "L_tab": 4851, "n_element_tab": 706, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 3861, "n_element_tab_1": 351, "formula_len_all": 208, "formula_len_all_1": 286, "len_all": 168673, "len_all_1": 71201, "len_abs": 1975, "len_title": 92, "len_sents": 50814, "len_sents_1": 30564, "n_sents": 370, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1984, "title": "FRAPPE: Fast RAG-Inspired Prompt Evaporator", "abs": "Large Language Models (LLMs) have demonstrated remarkable performance in various tasks such as multi-document QA, summarization, and text classification. This has been achieved in part by recent advancements in prompt engineering and in-context learning (ICL), enabling LLMs to consume tens of thousands of input tokens as the supported context for the given query. However, this creates higher computational costs, longer latency, and potential performance degradation. To address these issues, we propose a task-agnostic and efficient approach called \u201cFast RAG Inspired Prompt Evaporator\u201d, or FRAPPE, to significantly reduce LLMs\u2019 latency, memory requirement, and computation by compressing input tokens. Unlike many other proposed approaches for prompt compression, our method does not rely on any large model for computing conditional probabilities, and data preparation is fast with negligible memory requirements. In particular, our approach first pre-processes the input data, categorizes and ranks phrases based on their informativeness, and finally selects the highest-ranked phrases to generate highly compressed and extractive input. We show the efficacy of our approach through a comprehensive set of experiments on public datasets and benchmarks. For instance, on the summarization task of the MeetingBank dataset, at a compression rate of 70%, our proposed approach achieves performance similar to the full context while performing compression up to 4 times faster than the contemporary state of the art compression algorithms. We extend FRAPPE to create the Context-Aware FRAPPE algorithm, which incorporates task-specific information when ranking phrases, which further improves performance of downstream tasks using compressed text. Additionally, we demonstrate that the use of FRAPPE can reduce toxicity by close to 50% relative to the original text by removing extraneous vitriolic phrases, in contrast to other compression methods, which often increase toxicity.", "keywords": [ "Compression", "Prompt engineering", "Efficient LLM Inference", "Toxicity reduction", "Task-agnostic", "Summarization" ], "rating_list": [ 5, 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 4, 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3, 3 ] }, { "paper_id": "MjFoQAhnl3", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Ekdeep Singh Lubana", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Hidenori Tanaka", "gender": "unknown", "institution": "Harvard University, Harvard University", "country": "US", "position": "Researcher" }, { "name": "Kento Nishi", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Undergrad student" }, { "name": "Maya Okawa", "gender": "unknown", "institution": "NTT Research, Inc.", "country": "", "position": "Researcher" }, { "name": "Rahul Ramesh", "gender": "unknown", "institution": "University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "mikail khona", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 100, "n_ref_all": 126, "n_fig": 16, "n_tab": 5, "L_tab": 3027, "n_element_tab": 191, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1967, "n_element_tab_1": 56, "formula_len_all": 878, "formula_len_all_1": 402, "len_all": 173973, "len_all_1": 71075, "len_abs": 1862, "len_title": 131, "len_sents": 57289, "len_sents_1": 32600, "n_sents": 438, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1870, "title": "Representation Shattering in Transformers: A Synthetic Study with Knowledge Editing", "abs": "Knowledge Editing (KE) algorithms alter models' weights to perform targeted updates to incorrect, outdated, or otherwise unwanted factual associations. To better identify the possibilities and limitations of these approaches, recent work has shown that applying KE can adversely affect models' factual recall accuracy and diminish their general reasoning abilities. While these studies give broad insights into the potential harms of KE algorithms, e.g., via performance evaluations on benchmarks, we argue little is understood as to why such destructive failures occur. Is it possible KE methods distort representations of concepts beyond the targeted fact, hence hampering abilities at broad? If so, what is the extent of this distortion? Motivated by such questions, we define a novel synthetic task wherein a Transformer is trained from scratch to internalize a \"structured\" knowledge graph. The structure enforces relationships between entities of the graph, such that editing a factual association has \"trickling effects\" on other entities in the graph (e.g., altering X's parent is Y to Z affects who X's siblings' parent is). Through evaluations of edited models and analysis of extracted representations, we show that KE inadvertently affects representations of entities beyond the targeted one, distorting relevant structures that allow a model to infer unseen knowledge about an entity. We call this phenomenon representation shattering and demonstrate that it results in degradation of factual recall and reasoning performance more broadly. To corroborate our findings in a more naturalistic setup, we perform preliminary experiments with pretrained GPT-2-XL and Mamba models, reproducing the representation shattering effect therein as well. Overall, our work yields a precise mechanistic hypothesis to explain why KE has adverse effects on model abilities.", "keywords": [ "mechanistic interpretability", "knowledge editing", "transformers" ], "rating_list": [ 5, 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 4, 2 ], "confidence_list": [ 2, 4, 2, 3, 3 ] }, { "paper_id": "MidXrlkVu1", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Carl William Harris", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Huy Nguyen", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Nhat Ho", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Suchi Saria", "gender": "Female", "institution": "Johns Hopkins University", "country": "", "position": "Assistant Professor" }, { "name": "Xing Han", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Postdoc" } ], "n_formula": 195, "n_formula_1": 36, "n_ref_uni": 58, "n_ref": 91, "n_ref_all": 155, "n_fig": 3, "n_tab": 5, "L_tab": 1098, "n_element_tab": 73, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 634, "n_element_tab_1": 21, "formula_len_all": 40151, "formula_len_all_1": 4946, "len_all": 299173, "len_all_1": 74073, "len_abs": 895, "len_title": 136, "len_sents": 73474, "len_sents_1": 30423, "n_sents": 635, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 897, "title": "On Expert Estimation in Hierarchical Mixture of Experts: Beyond Softmax Gating Functions", "abs": "With the growing prominence of the Mixture of Experts (MoE) architecture in developing large-scale foundation models, we investigate the Hierarchical Mixture of Experts (HMoE), a specialized variant of MoE that excels in handling complex inputs and improving performance on targeted tasks. Our investigation highlights the advantages of using varied gating functions, moving beyond softmax gating within HMoE frameworks. We theoretically demonstrate that applying tailored gating functions to each expert group allows HMoE to achieve robust results, even when optimal gating functions are applied only at select hierarchical levels. Empirical validation across diverse scenarios supports these theoretical claims. This includes large-scale multimodal tasks, image classification, and latent domain discovery and prediction tasks, where our modified HMoE models show great performance improvements.", "keywords": [ "Hierarchical Mixture-of-Experts", "Multimodal", "Mixture-of-Expert Theory" ], "rating_list": [ 3, 8, 5, 3 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "MiPyle6Jef", "primary_area": "applications to neuroscience & cognitive science", "n_author": 9, "author_info_list": [ { "name": "Ammar Belatreche", "gender": "Male", "institution": "Northumbria University", "country": "GB", "position": "Associate Professor" }, { "name": "Honglin Cao", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Jieyuan Zhang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Malu Zhang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Wenjie Wei", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yang Yang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yimeng Shan", "gender": "Male", "institution": "Liaoning Technical University", "country": "CN", "position": "MS student" }, { "name": "Yu Liang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Zijian Zhou", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 29, "n_formula_1": 23, "n_ref_uni": 51, "n_ref": 95, "n_ref_all": 137, "n_fig": 18, "n_tab": 8, "L_tab": 2993, "n_element_tab": 214, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3784, "n_element_tab_1": 110, "formula_len_all": 2278, "formula_len_all_1": 1308, "len_all": 178672, "len_all_1": 68641, "len_abs": 1567, "len_title": 100, "len_sents": 47238, "len_sents_1": 28153, "n_sents": 383, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1573, "title": "QP-SNN: Quantized and Pruned Spiking Neural Networks", "abs": "Brain-inspired Spiking Neural Networks (SNNs) leverage sparse spikes to encode information and operate in an asynchronous event-driven manner, offering a highly energy-efficient paradigm for machine intelligence. However, the current SNN community focuses primarily on performance improvement by developing large-scale models, which limits the applicability of SNNs in resource-limited edge devices. In this paper, we propose a hardware-friendly and lightweight SNN, aimed at effectively deploying high-performance SNN in resource-limited scenarios. Specifically, we first develop a baseline model that integrates uniform quantization and structured pruning, called QP-SNN baseline. While this baseline significantly reduces storage demands and computational costs, it suffers from performance decline. To address this, we conduct an in-depth analysis of the challenges in quantization and pruning that lead to performance degradation and propose solutions to enhance the baseline's performance. For weight quantization, we propose a weight rescaling strategy that utilizes bit width more effectively to enhance the model's representation capability. For structured pruning, we propose a novel pruning criterion using the singular value of spatiotemporal spike activities to enable more accurate removal of redundant kernels. Extensive experiments demonstrate that integrating two proposed methods into the baseline allows QP-SNN to achieve state-of-the-art performance and efficiency, underscoring its potential for enhancing SNN deployment in edge intelligence computing.", "keywords": [ "Spiking Neural Networks", "Neuromorphic Computing", "Spiking Pruning", "Spiking Quantization" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "Mi45HjlVRj", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Chao Ye", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Gang Chen", "gender": "Male", "institution": "College of Computer Science and Technology, Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Haobo Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Junbo Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Liyao Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Wentao Ye", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yifei Sun", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhe Jiang", "gender": "unknown", "institution": "Wesleyan University", "country": "US", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 38, "n_ref": 77, "n_ref_all": 101, "n_fig": 8, "n_tab": 14, "L_tab": 2452, "n_element_tab": 156, "n_fig_1": 7, "n_tab_1": 7, "L_tab_1": 3509, "n_element_tab_1": 50, "formula_len_all": 279, "formula_len_all_1": 279, "len_all": 177034, "len_all_1": 60567, "len_abs": 1149, "len_title": 92, "len_sents": 51049, "len_sents_1": 26268, "n_sents": 361, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1173, "title": "Injecting Learnable Table Features into LLMs", "abs": "To migrate the remarkable successes of Large Language Models (LLMs), the community has made numerous efforts to extend them to the table reasoning tasks for the widely deployed tabular data. Despite that, in this work, by showing a probing experiment on our proposed StructQA benchmark, we postulate that even the most advanced LLMs (such as GPTs) may still fall short of coping with tabular data. More specifically, the current scheme often simply relies on serializing the tabular data, together with the meta information, then inputting them through the LLMs. We argue that the loss of structural information is the root of this shortcoming. In this work, we further propose **TAMO** which bears an ideology to treat the **ta**bles **a**s **a**n independent **mo**dality integrated with the text tokens. The resulting model in TAMO is a multimodal framework consisting of a hypergraph neural network as the global table encoder seamlessly integrated with the mainstream LLM. Empirical results on various benchmarking datasets, including HiTab, WikiTQ, WikiSQL, FeTaQA, and StructQA, have demonstrated significant improvements with an average relative gain of **42.65%**.", "keywords": [ "Large language model", "Table reasoning", "Multi-modal learning" ], "rating_list": [ 5, 5, 5, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "MhsCDuY4zx", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Hao Zhou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jingjing Gong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Juntong Shi", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Undergrad student" }, { "name": "Minkai Xu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Stefano Ermon", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Wei-Ying Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yuxuan Song", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Intern" } ], "n_formula": 42, "n_formula_1": 18, "n_ref_uni": 33, "n_ref": 105, "n_ref_all": 115, "n_fig": 3, "n_tab": 9, "L_tab": 2635, "n_element_tab": 367, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1097, "n_element_tab_1": 216, "formula_len_all": 4172, "formula_len_all_1": 1492, "len_all": 188741, "len_all_1": 66744, "len_abs": 1312, "len_title": 131, "len_sents": 67600, "len_sents_1": 29225, "n_sents": 493, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1315, "title": "Smooth Probabilistic Interpolation Benefits Generative Modeling for Discrete Graphs", "abs": "Though typically represented by the discrete node and edge attributes, the graph topological information can be sufficiently captured by the graph spectrum in a continuous space. It is believed that incorporating the continuity of graph topological information into the generative process design could establish a superior paradigm for graph generative modeling. Motivated by such prior and recent advancements in the generative paradigm, we propose Graph Bayesian Flow Networks (GraphBFN) in this paper, a principled generative framework that designs an alternative generative process emphasizing the dynamics of topological information. Unlike recent discrete-diffusion-based methods, GraphBFNemploys the continuous counts derived from sampling infinite times from a categorical distribution as latent to facilitate a smooth decomposition of topological information, demonstrating enhanced effectiveness. To effectively realize the concept, we further develop an advanced sampling strategy and new time-scheduling techniques to overcome practical barriers and boost performance. Through extensive experimental validation on both generic graph and molecular graph generation tasks, GraphBFN could consistently achieve superior or competitive performance with significantly higher training and sampling efficiency.", "keywords": [ "2D Molecule Generation; Bayesian Flow Networks" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "MhD9rLeU31", "primary_area": "learning on time series and dynamical systems", "n_author": 12, "author_info_list": [ { "name": "Alex Pyo", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "Undergrad student" }, { "name": "Ammar Gilani", "gender": "Male", "institution": "Northwestern University", "country": "", "position": "PhD student" }, { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" }, { "name": "Jerry Yao-Chieh Hu", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Jialong Li", "gender": "Not Specified", "institution": "Northwestern University, Northwestern University", "country": "US", "position": "MS student" }, { "name": "Jiayi Wang", "gender": "Female", "institution": "Northwestern University", "country": "", "position": "PhD student" }, { "name": "Jigyasa Kumari", "gender": "Female", "institution": "Northwestern University", "country": "US", "position": "Intern" }, { "name": "Jingya Xun", "gender": "Female", "institution": "Northwestern University", "country": "", "position": "Researcher" }, { "name": "Lining Mao", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Mehak Kawatra", "gender": "Female", "institution": "Northwestern University", "country": "US", "position": "MS student" }, { "name": "Stephen S Cheng", "gender": "Male", "institution": "Northwestern University", "country": "", "position": "MS student" }, { "name": "Weijian Li", "gender": "Male", "institution": "Northwestern University", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 37, "n_ref": 59, "n_ref_all": 84, "n_fig": 5, "n_tab": 10, "L_tab": 3315, "n_element_tab": 218, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 3774, "n_element_tab_1": 239, "formula_len_all": 1078, "formula_len_all_1": 912, "len_all": 146797, "len_all_1": 65968, "len_abs": 6292, "len_title": 146, "len_sents": 44830, "len_sents_1": 27751, "n_sents": 354, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1391, "title": "A Benchmark Study For Limit Order Book (LOB) Models and Time Series Forecasting Models on LOB Data", "abs": "We present a comprehensive benchmark to evaluate the performance of deep learning models on limit order book (LOB) data. Our work makes four significant contributions: (i) We evaluate existing LOB models on a proprietary futures LOB dataset to examine the transferability of LOB model performance between various assets; (ii) We are the first to benchmark existing LOB models on the mid-price return forecasting (MPRF) task. (iii) We present the first benchmark study to evaluate SOTA time series forecasting models on the MPRF task to bridge the two fields of general-purpose time series forecasting and LOB time series forecasting; and (iv) we propose an architecture of convolutional cross-variate mixing layers (CVML) as an add-on to any deep learning multivariate time series model to significantly enhance MPRF performance on LOB data. Our empirical results highlight the value of our benchmark results on our proprietary futures LOB dataset, demonstrating a performance gap between the commonly used open-source stock LOB dataset and our futures dataset. Furthermore, the results demonstrate that LOB-aware model design is essential for achieving optimal prediction performance on LOB datasets. Most importantly, our results show that our proposed CVML architecture brings about an average improvement of 244.9% to various time series models\u2019 mid-price return forecasting performance.", "keywords": [ "benchmark", "time series forecasting", "convolution", "deep learning", "limit order book", "mid-price trend prediction", "mid-price return forecasting" ], "rating_list": [ 5, 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 4, 2 ] }, { "paper_id": "Mh8blXreJW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Bowei He", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Chen Ma", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Hui-Ling Zhen", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Principal Researcher" }, { "name": "Lihao Yin", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 41, "n_ref": 74, "n_ref_all": 93, "n_fig": 2, "n_tab": 13, "L_tab": 15231, "n_element_tab": 1736, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 6102, "n_element_tab_1": 682, "formula_len_all": 1383, "formula_len_all_1": 1358, "len_all": 219639, "len_all_1": 75075, "len_abs": 2180, "len_title": 135, "len_sents": 56004, "len_sents_1": 28739, "n_sents": 383, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1636, "title": "PASER: Post-Training Data Selection for Efficient Pruned Large Language Model Recovery", "abs": "Model pruning is an effective approach for compressing Large Language Models (LLMs) and improving inference efficiency. However, this process often leads to significant degradation of model capabilities. While post-training techniques such as instruction tuning are commonly employed to recover model performance, existing methods often overlook the uneven deterioration of model capabilities and incur high computational costs due to extensive recovery training. Moreover, some instruction data irrelevant to model capability recovery may introduce negative effects. To address these challenges, we propose the **P**ost-training d**A**ta **S**election method for **E**fficient pruned large language model **R**ecovery (**PASER**). PASER aims to identify instructions where model capabilities are most severely compromised within a certain recovery data budget. Our approach first applies manifold learning and spectral clustering to group recovery data in the semantic space, revealing capability-specific instruction sets. We then adaptively allocate the data budget to different clusters based on the degrees of model capability degradation. In each cluster, we prioritize data samples where model performance has declined dramatically. To mitigate potential negative transfer, we also detect and filter out conflicting or irrelevant recovery data. Extensive experiments demonstrate that PASER significantly outperforms conventional baselines, effectively recovering the general capabilities of pruned LLMs while utilizing merely 4\\%-20\\% of the original post-training data and substantially reducing training computational overhead.", "keywords": [ "Large Language Model", "Model Pruning", "Recovery Training", "Data Selection" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "Mgf7qdUbX5", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Furong Huang", "gender": "Female", "institution": "University of Maryland", "country": "US", "position": "Associate Professor" }, { "name": "Michael-Andrei Panaitescu-Liess", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Pankayaraj Pathmanathan", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Udari Madhushani", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Postdoc" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 41, "n_ref": 96, "n_ref_all": 139, "n_fig": 10, "n_tab": 10, "L_tab": 3857, "n_element_tab": 252, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 595, "n_element_tab_1": 132, "formula_len_all": 183, "formula_len_all_1": 183, "len_all": 275516, "len_all_1": 69110, "len_abs": 1348, "len_title": 144, "len_sents": 92692, "len_sents_1": 35105, "n_sents": 690, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1267, "title": "AdvBDGen: Adversarially Fortified Prompt-Specific Fuzzy Backdoor Generator Against LLM Alignment", "abs": "With the growing adoption of reinforcement learning with human feedback (RLHF) for aligning large language models (LLMs), the risk of backdoor installation during alignment has increased, leading to unintended and harmful behaviors. Existing backdoor triggers are typically limited to fixed word patterns, making them detectable during data cleaning and easily removable post-poisoning. In this work, we explore the use of prompt-specific paraphrases as backdoor triggers, enhancing their stealth and resistance to removal during LLM alignment. We propose AdvBDGen, an adversarially fortified generative fine-tuning framework that automatically generates prompt-specific backdoors that are effective, stealthy, and transferable across models. AdvBDGen employs a generator-detector pair, fortified by an adversary, to ensure the installability and stealthiness of backdoors. It enables the crafting of complex triggers using as little as 3% of the fine-tuning data. Once installed, these backdoors can jailbreak LLMs during inference, demonstrate improved stability against perturbations compared to traditional constant triggers, and are harder to remove. These properties highlight the greater risks posed by such an adversarially crafted backdoors to LLM alignment.", "keywords": [ "RLHF poisoning", "Backdoor", "LLM Alignment" ], "rating_list": [ 3, 3, 5, 8 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "Mg1stYVYTl", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Cheng Yan", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Guansong Pang", "gender": "unknown", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Min Chen", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "MS student" }, { "name": "Wenjun Wang", "gender": "Male", "institution": "Hainan Tropical Ocean University", "country": "CN", "position": "Full Professor" } ], "n_formula": 64, "n_formula_1": 10, "n_ref_uni": 35, "n_ref": 111, "n_ref_all": 156, "n_fig": 12, "n_tab": 15, "L_tab": 6854, "n_element_tab": 615, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1314, "n_element_tab_1": 23, "formula_len_all": 4764, "formula_len_all_1": 954, "len_all": 257491, "len_all_1": 56532, "len_abs": 1661, "len_title": 145, "len_sents": 77429, "len_sents_1": 25106, "n_sents": 725, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1670, "title": "Combating Dual Noise Effect in Spatial-temporal Forecasting via Information Bottleneck Principle", "abs": "Spatial-temporal forecasting plays a pivotal role in urban planning and computing. Although Spatial-Temporal Graph Neural Networks (STGNNs) excel in modeling spatial-temporal dynamics, they often suffer from relatively poor computational efficiency. Recently, Multi-Layer Perceptrons (MLPs) have gained popularity in spatial-temporal forecasting for their simplified architecture and better efficiency. However, existing MLP-based models can be susceptible to noise interference, especially when the noise can affect both input and target sequences in spatial-temporal forecasting on noisy data. To alleviate this impact, we propose _Robust Spatial-Temporal Information Bottleneck (RSTIB)_ principle. The RSTIB extends previous Information Bottleneck (IB) approaches by lifting the specific Markov assumption without impairing the IB nature. Then, by explicitly minimizing the irrelevant noisy information, the representation learning guided by RSTIB can be more robust against noise interference. Furthermore, the instantiation, RSTIB-MLP, can be seamlessly implemented with MLPs, thereby achieving efficient and robust spatial-temporal modeling. Moreover, a training regime is designed to handle the dynamic nature of spatial-temporal relationships by incorporating a knowledge distillation module to alleviate feature collapse and enhance model robustness under noisy conditions. Our extensive experimental results on six intrinsically noisy benchmark datasets from various domains show that the RSTIB-MLP runs much faster than state-of-the-art STGNNs and delivers superior forecasting accuracy across noisy environments, substantiating its robustness and efficiency.", "keywords": [ "Robust spatial-temporal forecasting", "Multi-Layer Perceptron", "Information bottleneck" ], "rating_list": [ 3, 6, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 5, 3 ] }, { "paper_id": "Mfnh1Sqdwf", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Degui Zhi", "gender": "Male", "institution": "University of Texas Health Science Center at Houston", "country": "", "position": "Full Professor" }, { "name": "Haiyang Yu", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Shuiwang Ji", "gender": "Male", "institution": "Texas A&M University", "country": "US", "position": "Professor" }, { "name": "Xingyu Su", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 13, "n_ref_uni": 36, "n_ref": 70, "n_ref_all": 78, "n_fig": 2, "n_tab": 3, "L_tab": 678, "n_element_tab": 79, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 666, "n_element_tab_1": 8, "formula_len_all": 855, "formula_len_all_1": 613, "len_all": 124851, "len_all_1": 65160, "len_abs": 1003, "len_title": 119, "len_sents": 39649, "len_sents_1": 31726, "n_sents": 306, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 71, "L_abs": 1098, "title": "Learning to Discover Regulatory Elements for Gene Expression Prediction", "abs": "We consider the problem of predicting gene expressions from DNA sequences. A key challenge of this task is to find the regulatory elements that control gene expressions. Here, we introduce Seq2Exp, a Sequence to Expression network explicitly designed to discover and extract regulatory elements that drive target gene expression, enhancing the accuracy of the gene expression prediction. Our approach captures the causal relationship between epigenomic signals, DNA sequences and their associated regulatory elements. Specifically, we propose to decompose the epigenomic signals and the DNA sequence conditioned on the causal active regulatory elements, and apply an information bottleneck with the Beta distribution to combine their effects while filtering out non-causal components. Our experiments demonstrate that Seq2Exp outperforms existing baselines in gene expression prediction tasks and discovers influential regions compared to commonly used statistical methods for peak detection such as MACS3. The source code is released as part of the AIRS library (https://github.com/divelab/AIRS/).", "keywords": [ "Gene Expression", "Deep Learning", "Sequence Modeling" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 4, 4, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "Mez2No9lHj", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 11, "author_info_list": [ { "name": "Fahong Zhang", "gender": "unknown", "institution": "Xi'an Jiaotong University", "country": "", "position": "MS student" }, { "name": "Ling Liang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Meng Li", "gender": "Male", "institution": "Peking University", "country": "", "position": "Assistant Professor" }, { "name": "Peiye Liu", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Qichao Ma", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Renye Yan", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Rui-Jie Zhu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Tiejun Huang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Yimao Cai", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Zhaofei Yu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zongwei Wang", "gender": "unknown", "institution": "Peking University", "country": "", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 27, "n_ref": 32, "n_ref_all": 58, "n_fig": 11, "n_tab": 7, "L_tab": 2531, "n_element_tab": 164, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 984, "n_element_tab_1": 13, "formula_len_all": 690, "formula_len_all_1": 596, "len_all": 144312, "len_all_1": 59181, "len_abs": 2515, "len_title": 131, "len_sents": 41807, "len_sents_1": 27849, "n_sents": 336, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1461, "title": "CopyLens: Dynamically Flagging Copyrighted Sub-Dataset Contributions to LLM Outputs", "abs": "Large Language Models (LLMs) have become pervasive due to their knowledge absorption and text-generation capabilities. Concurrently, the copyright issue for pretraining datasets has been a pressing concern, particularly when generation includes specific styles. Previous methods either focus on the defense of identical copyrighted outputs or find interpretability by individual tokens with computational burdens. However, the gap between them exists, where direct assessments of how dataset contributions impact LLM outputs are missing. Once the model providers ensure copyright protection for data holders, a more mature LLM community can be established. To address these limitations, we introduce CopyLens, a new framework to analyze how copyrighted datasets may influence LLM responses. Specifically, a two-stage approach is employed: First, based on the uniqueness of pretraining data in the embedding space, token representations are initially fused for potential copyrighted texts, followed by a lightweight LSTM-based network to analyze dataset contributions. With such a prior, a contrastive-learning-based non-copyright OOD detector is designed. Our framework can dynamically face different situations and bridge the gap between current copyright detection methods. Experiments show that CopyLens improves efficiency and accuracy by 15.2\\% over our proposed baseline, 58.7\\% over prompt engineering methods, and 0.21 AUC over OOD detection baselines.", "keywords": [ "Copyright", "Large Language Model" ], "rating_list": [ 3, 1, 3 ], "soundness_list": [ 1, 1, 1 ], "presentation_list": [ 1, 1, 1 ], "contribution_list": [ 1, 1, 1 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "MeSfNZjGvN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Nurbek Tastan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Samuel Horv\u00e1th", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Assistant Professor" }, { "name": "Karthik Nandakumar", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Associate Professor" }, { "name": "Martin Takac", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Associate Professor" } ], "n_formula": 21, "n_formula_1": 21, "n_ref_uni": 37, "n_ref": 73, "n_ref_all": 106, "n_fig": 11, "n_tab": 2, "L_tab": 786, "n_element_tab": 61, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 713, "n_element_tab_1": 55, "formula_len_all": 1536, "formula_len_all_1": 1344, "len_all": 139274, "len_all_1": 65361, "len_abs": 1079, "len_title": 138, "len_sents": 39819, "len_sents_1": 28806, "n_sents": 326, "n_sents_1": 225, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1119, "title": "FedPeWS: Personalized Warmup via Subnetworks for Enhanced Heterogeneous Federated Learning", "abs": "Statistical data heterogeneity is a significant barrier to convergence in federated learning (FL). While prior work has advanced heterogeneous FL through better optimization objectives, these methods fall short when there is \\textit{extreme} data heterogeneity among collaborating participants. We hypothesize that convergence under extreme data heterogeneity is primarily hindered due to the aggregation of conflicting updates from the participants in the initial collaboration rounds. To overcome this problem, we propose a warmup phase where each participant learns a personalized mask and updates only a subnetwork of the full model. This \\textit{personalized warmup} allows the participants to focus initially on learning specific \\textit{subnetworks} tailored to the heterogeneity of their data. After the warmup phase, the participants revert to standard federated optimization, where all parameters are communicated. We empirically demonstrate that the proposed personalized warmup via subnetworks (\\texttt{FedPeWS}) approach improves accuracy and convergence speed over standard federated optimization methods.", "keywords": [ "federated learning", "heterogeneous federated learning", "subnetworks", "personalized warmup" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "MeOi6u9E23", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Haijun Yang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Yuxing Yuan", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Ziwei Wang", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" }, { "name": "Zhenyuanhuang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 7, "n_ref_uni": 20, "n_ref": 50, "n_ref_all": 56, "n_fig": 3, "n_tab": 4, "L_tab": 2262, "n_element_tab": 213, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1000, "n_element_tab_1": 149, "formula_len_all": 2390, "formula_len_all_1": 401, "len_all": 141732, "len_all_1": 66295, "len_abs": 905, "len_title": 34, "len_sents": 53433, "len_sents_1": 31081, "n_sents": 446, "n_sents_1": 240, "venue": "ICLR2025", "decision": "withdraw", "L_title": 34, "L_abs": 914, "title": "DiT-LSTM-SVAR Model For Portfolios", "abs": "This paper proposes a novel combined model named DiT-LSTM-SVAR, which successfully integrates time series and the Efficient Markets Hypothesis. This is the first to combine the microstructure of financial markets with deep learning networks to improve the performance of portfolios. We employ the DiT model to predict the upside and downside movements and an information decomposition model based on the SVAR model to identify random walk stocks. The DiT module significantly improves the Matthews correlation coefficient by almost 3\\%. The annual return of the portfolio is improved by almost 20\\%. The SVAR module greatly improves the Matthews correlation coefficient by almost 4\\%. Portfolios constructed using the DiT-LSTM-SVAR module based on market and public information outperformed those created with the DiT-LSTM model. The annual cumulative return of the portfolio is 266.60\\% and a Sharpe ratio of 1.8.", "keywords": [ "DiT-LSTM", "SVAR", "Portfolios" ], "rating_list": [ 8, 1, 5, 1 ], "soundness_list": [ 4, 1, 3, 1 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 4, 2, 3, 1 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "MeGDmZjUXy", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Elizaveta Tennant", "gender": "Female", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Stephen Hailes", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Full Professor" }, { "name": "Mirco Musolesi", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Full Professor" } ], "n_formula": 25, "n_formula_1": 2, "n_ref_uni": 50, "n_ref": 94, "n_ref_all": 184, "n_fig": 29, "n_tab": 7, "L_tab": 2859, "n_element_tab": 148, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 64, "n_element_tab_1": 4, "formula_len_all": 2464, "formula_len_all_1": 528, "len_all": 279886, "len_all_1": 63723, "len_abs": 1739, "len_title": 78, "len_sents": 86694, "len_sents_1": 32972, "n_sents": 628, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 30, "L_abs": 1751, "title": "Moral Alignment for LLM Agents", "abs": "Decision-making agents based on pre-trained Large Language Models (LLMs) are increasingly being deployed across various domains of human activity. While their applications are currently rather specialized, several research efforts are under way to develop more generalist agents. As LLM-based systems become more agentic, their influence on human activity will grow and the transparency of this will decrease. Consequently, developing effective methods for aligning them to human values is vital. \n\nThe prevailing practice in alignment often relies on human preference data (e.g., in RLHF or DPO), in which values are implicit and are essentially deduced from relative preferences over different model outputs. In this work, instead of relying on human feedback, we introduce the design of reward functions that explicitly encode core human values for Reinforcement Learning-based fine-tuning of foundation agent models. Specifically, we use intrinsic rewards for the moral alignment of LLM agents. \n\nWe evaluate our approach using the traditional philosophical frameworks of Deontological Ethics and Utilitarianism, quantifying moral rewards for agents in terms of actions and consequences on the Iterated Prisoner's Dilemma (IPD) environment. We also show how moral fine-tuning can be deployed to enable an agent to unlearn a previously developed selfish strategy. Finally, we find that certain moral strategies learned on the IPD game generalize to several other matrix game environments. In summary, we demonstrate that fine-tuning with intrinsic rewards is a promising general solution for aligning LLM agents to human values, and it might represent a more transparent and cost-effective alternative to currently predominant alignment techniques.", "keywords": [ "alignment", "LLM fine-tuning", "moral decision-making", "social dilemmas" ], "rating_list": [ 6, 5, 8 ], "soundness_list": [ 3, 4, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 1, 4 ], "confidence_list": [ 2, 4, 3 ] }, { "paper_id": "MeCPwqrm19", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Fang Wu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Jinbo Xu", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Full Professor" }, { "name": "Jure Leskovec", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Shuting Jin", "gender": "Female", "institution": "Wuhan university science and technology", "country": "CN", "position": "Lecturer" }, { "name": "xiangxiang Zeng", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 33, "n_formula_1": 29, "n_ref_uni": 78, "n_ref": 144, "n_ref_all": 155, "n_fig": 4, "n_tab": 6, "L_tab": 1217, "n_element_tab": 147, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 772, "n_element_tab_1": 102, "formula_len_all": 2998, "formula_len_all_1": 2586, "len_all": 199500, "len_all_1": 73499, "len_abs": 1156, "len_title": 59, "len_sents": 44059, "len_sents_1": 29678, "n_sents": 354, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1180, "title": "Surface-based Peptide Design with Multi-modal Flow Matching", "abs": "Therapeutic peptides show promise in targeting previously undruggable binding sites, with recent advancements in deep generative models enabling full-atom peptide co-design for specific protein receptors. However, the critical role of molecular surfaces in protein-protein interactions (PPIs) has been underexplored. To bridge this gap, we propose an \\emph{omni-design} peptides generation paradigm, called SurfFlow, a novel surface-based generative algorithm that enables comprehensive co-design of sequence, structure, and surface for peptides. SurfFlow employs a multi-modality conditional flow matching (CFM) architecture to learn distributions of surface geometries and biochemical properties, enhancing peptide binding accuracy. Evaluated on the comprehensive PepMerge benchmark, SurfFlow consistently outperforms full-atom baselines across all metrics. These results highlight the advantages of considering molecular surfaces in \\emph{de novo} peptide discovery and demonstrate the potential of integrating multiple protein modalities for more effective therapeutic peptide discovery. Anonymous codes are available at~\\url{https://anonymous.4open.science/r/SurfFlow-880B/}.", "keywords": [ "Molecular Surfaces", "Peptide Design", "Flow Matching" ], "rating_list": [ 5, 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 5, 3, 2 ] }, { "paper_id": "MdidZNQxqK", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Joongkyu Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Min-hwan Oh", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" } ], "n_formula": 229, "n_formula_1": 33, "n_ref_uni": 51, "n_ref": 173, "n_ref_all": 198, "n_fig": 1, "n_tab": 14, "L_tab": 8423, "n_element_tab": 484, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 809, "n_element_tab_1": 21, "formula_len_all": 57790, "formula_len_all_1": 3991, "len_all": 418017, "len_all_1": 84022, "len_abs": 1322, "len_title": 109, "len_sents": 95727, "len_sents_1": 30891, "n_sents": 882, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1302, "title": "Combinatorial Reinforcement Learning with Preference Feedback", "abs": "In this paper, we consider combinatorial reinforcement learning with preference feedback, where a learning agent sequentially offers an action\u2014an assortment of multiple items\u2014to a user, whose preference feedback follows a multinomial logit (MNL) model. This framework allows us to model real-world scenarios, particularly those involving long-term user engagement, such as in recommender systems and online advertising. However, this framework faces two main challenges: (1) the unknown value of each item, unlike traditional MNL bandits (which only account for single-step preference feedback), and (2) the difficulty of ensuring optimism with tractable assortment selection in the combinatorial action space. In this paper, we assume a contextual MNL preference model, where mean utilities are linear, and the value of each item is approximated using general function approximation. We propose an algorithm, MNL-V$Q$L, that addresses these challenges, making it both computationally and statistically efficient. As a special case, for linear MDPs (with the MNL preference model), we establish a regret lower bound and show that MNL-V$Q$L achieves near-optimal regret. To the best of our knowledge, this is the first work to provide statistical guarantees in combinatorial RL with preference feedback.", "keywords": [ "combinatorial reinforcement learning", "preference feedback", "contextual MNL bandits", "nonlinear function approximation" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "MdBt0ttZrZ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Sangjun Han", "gender": "unknown", "institution": "Yonsei University", "country": "KR", "position": "PhD student" }, { "name": "Youngmi Hur", "gender": "unknown", "institution": "Yonsei University", "country": "KR", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 48, "n_ref": 90, "n_ref_all": 127, "n_fig": 12, "n_tab": 10, "L_tab": 1578, "n_element_tab": 151, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2424, "n_element_tab_1": 135, "formula_len_all": 337, "formula_len_all_1": 340, "len_all": 173715, "len_all_1": 59287, "len_abs": 2300, "len_title": 162, "len_sents": 52730, "len_sents_1": 26248, "n_sents": 457, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 114, "L_abs": 2151, "title": "Detail Loss in Super-Resolution Models Based on the Laplacian Pyramid and Repeated Upscaling-Downscaling Structure", "abs": "With advances in artificial intelligence, image processing has also gained significant interest. Image super-resolution, in particular, is a vital technology closely related to real-life applications, as it enhances the quality of existing images. Since enhancing details is important in the super-resolution task, it is often necessary to activate pixels that appear only at high frequencies, distinct from low frequencies. \nIn this paper, we propose a method that generates a detail image separately from the super-resolution image. This approach introduces a loss function designed to enhance detail, allowing the model to generate an upscaled image and a detail image independently, with control over each component. Consequently, the model can focus more effectively on high-frequency data, resulting in an improved super-resolution image. Our loss function utilizes detail images based on the Laplacian Pyramid, which is widely used in image reconstruction. The multi-level property of the Laplacian Pyramid is well-suited for applying upscaling and downscaling repeatedly.\nOur experiments demonstrate that a structure applying the repetition of upscaling and downscaling integrates effectively with our detail loss control. The results show that this structure efficiently extracts diverse information, enabling the generation of improved super-resolution images from multiple low-resolution features. We conduct two types of experiments. First, we construct a simple CNN-based model incorporating the Laplacian Pyramid-based detail control and a repeated upscaling and downscaling structure. This model achieves a state-of-the-art PSNR value of 38.48 dB, surpassing all currently available CNN-based models and even some attention-based models without additional special techniques. Second, we apply our methods to existing attention-based models on a small scale. In all the experiments, attention-based models using our detail loss show improvements compared to the original models. These experiments demonstrate that our detail control loss effectively enhances performance, regardless of the model's structure in the super-resolution task.", "keywords": [ "Super-Resoltuion", "Laplacian Pyramid-based detail loss", "Repeated Upscaling-Downscaling Process", "Supervised learning" ], "rating_list": [ 5, 3, 5, 1 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "Md783Qa2JX", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Jon Schneider", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Khashayar Gatmiry", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Stefanie Jegelka", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" } ], "n_formula": 191, "n_formula_1": 24, "n_ref_uni": 46, "n_ref": 77, "n_ref_all": 84, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15644, "formula_len_all_1": 1239, "len_all": 211828, "len_all_1": 72046, "len_abs": 1279, "len_title": 109, "len_sents": 64891, "len_sents_1": 34199, "n_sents": 567, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1295, "title": "Computing Optimal Regularizers for Online Linear Optimization", "abs": "Follow-the-Regularized-Leader (FTRL) algorithms are a popular class of learning algorithms for online linear optimization (OLO) that guarantee sub-linear regret, but the choice of regularizer can significantly impact dimension-dependent factors in the regret bound. We present an algorithm that takes as input convex and symmetric action sets and loss sets for a specific OLO instance, and outputs a regularizer such that running FTRL with this regularizer guarantees regret within a universal constant factor of the best possible regret bound. In particular, for any choice of (convex, symmetric) action set and loss set we prove that there exists an instantiation of FTRL which achieves regret within a constant factor of the best possible learning algorithm, strengthening the universality result of Srebro et al., 2011. \n \nOur algorithm requires preprocessing time and space exponential in the dimension $d$ of the OLO instance, but can be run efficiently online assuming a membership and linear optimization oracle for the action and loss sets, respectively (and is fully polynomial time for the case of constant dimension $d$). We complement this with a lower bound showing that even deciding whether a given regularizer is $\\alpha$-strongly-convex with respect to a given norm is NP-hard.", "keywords": [ "mirror descent", "online convex optimization", "online linear optimization", "minimax optimal rate", "adversarial online learning" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "McqeEcMSzy", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Amir Bar", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Postdoc" }, { "name": "Grace Luo", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "trevor darrell", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department", "country": "US", "position": "Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 34, "n_ref": 58, "n_ref_all": 81, "n_fig": 12, "n_tab": 5, "L_tab": 1436, "n_element_tab": 222, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 1509, "n_element_tab_1": 215, "formula_len_all": 250, "formula_len_all_1": 209, "len_all": 122694, "len_all_1": 53058, "len_abs": 362, "len_title": 126, "len_sents": 27574, "len_sents_1": 22276, "n_sents": 216, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 28, "L_abs": 1029, "title": "Task Vectors are Cross-Modal", "abs": "We investigate the internal representations of autoregressive vision-and-language models (VLMs) and how they encode task representations. We consider tasks specified through examples or instructions, using either text or image inputs. Surprisingly, we find that conceptually similar tasks are mapped to similar task vector representations, regardless of how they are specified. Our findings suggest that to output answers, tokens in VLMs undergo three distinct phases: input, task, and answer, a process which is consistent across different modalities and specifications. The task vectors we identify in VLMs are general enough to be derived in one modality (e.g., text) and transferred to another (e.g., image). Additionally, we find that ensembling exemplar and instruction based task vectors produce better task representations. Taken together, these insights shed light on the underlying mechanisms of VLMs, particularly their ability to represent tasks in a shared manner across different modalities and task specifications.", "keywords": [ "in-context learning", "interpretability", "multimodal models", "vision and language models", "task vectors", "function vectors" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "McqVjmwdPe", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Hugo Monz\u00f3n Maldonado", "gender": "Male", "institution": "RIKEN AIP", "country": "JP", "position": "Postdoc" }, { "name": "Iryna Gurevych", "gender": "Female", "institution": "Institute for Computer Science, Artificial Intelligence and Technology", "country": "BG", "position": "Full Professor" }, { "name": "Mohammad Emtiyaz Khan", "gender": "Male", "institution": "RIKEN Center for AI Project", "country": "JP", "position": "Full Professor" }, { "name": "Nico Daheim", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Researcher" }, { "name": "Thomas M\u00f6llenhoff", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Researcher" } ], "n_formula": 32, "n_formula_1": 18, "n_ref_uni": 52, "n_ref": 98, "n_ref_all": 127, "n_fig": 13, "n_tab": 3, "L_tab": 1243, "n_element_tab": 200, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 829, "n_element_tab_1": 115, "formula_len_all": 2309, "formula_len_all_1": 1258, "len_all": 174483, "len_all_1": 60964, "len_abs": 726, "len_title": 124, "len_sents": 41939, "len_sents_1": 25328, "n_sents": 380, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 729, "title": "How to Weight Multitask Finetuning? Fast Previews via Model Merging", "abs": "When finetuning multiple tasks altogether, it is important to carefully weigh them to get a good performance, but searching for good weights can be difficult and costly. Here, we propose to aid the search with fast previews to quickly get a rough idea of different reweighting options. We use model merging to create previews by simply reusing and averaging parameters of models trained on each task separately (no retraining required). To improve the quality of previews, we propose a Bayesian approach to design new merging strategies by using more flexible posteriors. We validate our findings on vision and natural-language transformers. Our work shows the benefits of model merging via Bayes to improve multitask finetuning.", "keywords": [ "Model Merging", "Bayesian Inference", "Multitask Learning", "Finetuning" ], "rating_list": [ 3, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "McNPMnPVrz", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Anastasia Markina", "gender": "unknown", "institution": "BioNTech", "country": "DE", "position": "Researcher" }, { "name": "Bakary N'tji Diallo", "gender": "unknown", "institution": "InstaDeep", "country": "TN", "position": "Researcher" }, { "name": "Christoph Brunken", "gender": "Male", "institution": "InstaDeep", "country": "", "position": "Researcher" }, { "name": "Martin Maarand", "gender": "unknown", "institution": "InstaDeep", "country": "FR", "position": "Researcher" }, { "name": "Mustafa Omar", "gender": "Male", "institution": "InstaDeep", "country": "GB", "position": "Researcher" }, { "name": "Oliver Bent", "gender": "Male", "institution": "InstaDeep", "country": "", "position": "Researcher" }, { "name": "Olivier Peltre", "gender": "Male", "institution": "InstaDeep", "country": "FR", "position": "Researcher" }, { "name": "Sebastien Boyer", "gender": "unknown", "institution": "InstaDeep", "country": "", "position": "Researcher" }, { "name": "Solal Attias", "gender": "Male", "institution": "InstaDeep", "country": "FR", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 8, "n_ref_uni": 34, "n_ref": 54, "n_ref_all": 93, "n_fig": 12, "n_tab": 3, "L_tab": 789, "n_element_tab": 65, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 859, "formula_len_all_1": 257, "len_all": 180365, "len_all_1": 64649, "len_abs": 1337, "len_title": 144, "len_sents": 62741, "len_sents_1": 34108, "n_sents": 409, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1340, "title": "Universally Applicable And Tunable Graph-Based Coarse-Graining For Machine Learning Force Fields", "abs": "Coarse-grained (CG) force field methods for molecular systems are a crucial tool to simulate large biological macromolecules and are therefore essential for characterisations of biomolecular systems. While state-of-the-art deep learning (DL)-based models for all-atom force fields have improved immensely over recent years, we observe and analyse significant limitations of the currently available approaches for DL-based CG simulations. In this work, we present the first transferable DL-based CG force field approach (i.e., not specific to only one narrowly defined system type) applicable to a wide range of biosystems. To achieve this, our CG algorithm does not rely on hard-coded rules and is tuned to output coarse-grained systems optimised for minimal statistical noise in the ground truth CG forces, which results in significant improvement of model training. Our force field model is also the first CG variant that is based on the MACE architecture and is trained on a custom dataset created by a new approach based on the fragmentation of large biosystems covering protein, RNA and lipid chemistry. We demonstrate that our model can be applied in molecular dynamics simulations to obtain stable and qualitatively accurate trajectories for a variety of systems, while also discussing cases for which we observe limited reliability.", "keywords": [ "Molecular Dynamics", "Coarse-graining", "Quantum Chemistry", "EGNN", "Geometric Deep Learning", "JAX", "Dataset generation" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "MbtA7no8Ys", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Jieping Ye", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Sile Hu", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Xiang Tian", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Associate Professor" }, { "name": "Xu Shen", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Yaowu Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Yiqun Wang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Yonggang Zhang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Postdoc" }, { "name": "Wan Chaoqun", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 57, "n_ref_all": 85, "n_fig": 14, "n_tab": 9, "L_tab": 4149, "n_element_tab": 464, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 393, "n_element_tab_1": 79, "formula_len_all": 16, "formula_len_all_1": 0, "len_all": 172822, "len_all_1": 65935, "len_abs": 1489, "len_title": 165, "len_sents": 54264, "len_sents_1": 31414, "n_sents": 388, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 117, "L_abs": 1477, "title": "Deciphering and Enhancing Commonsense Reasoning in LLMs from the Perspective of Intrinsic Factual Knowledge Retrieval", "abs": "Commonsense reasoning in large language models (LLMs) bridges the gap to physical world, thus allowing them to think and behave more like humans. Previous research has shown that LLMs acquire the underlying factual knowledge from extensive training corpora and store it within their parameters. However, how LLMs apply this knowledge during the inference phase remains unclear. This lack of transparency makes it difficult to determine whether shortcomings in LLMs are due to a lack of factual knowledge or insufficient reasoning capabilities.\nIn this work, we aim to decipher the commonsense reasoning process into human-understandable steps. By interpreting the hidden states in different transformer layers and token positions, we uncover a specific mechanism by which LLMs execute reasoning.\nOur extensive experiments indicate: 1) both attention head and multi-layer perceptron (MLP) contribute to the generation of factual knowledge from different perspective. 2) The process of commonsense reasoning in LLMs involves a clear sequence of knowledge augmentation, knowledge retrieval and answer generation, akin to retrieval-augmented generation.\nBuilding on these findings, we have discovered that LLMs often contain relevant facutal knowledge but fail to retrieve the correct knowledge at top. To address this issure, we selectively fine-tuned the key heads and MLPs, resulting in notably improvements in reasoning performance in both in-domain and out-of-domain settings.", "keywords": [ "Model interpretability", "Chain-of-Thought", "Large Language Models" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "MbX0t1rUlp", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Cengiz Pehlevan", "gender": "unknown", "institution": "School of Engineering and Applied Sciences, Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "William Lingxiao Tong", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 2, "n_ref_uni": 49, "n_ref": 123, "n_ref_all": 173, "n_fig": 31, "n_tab": 3, "L_tab": 1078, "n_element_tab": 104, "n_fig_1": 23, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1390, "formula_len_all_1": 63, "len_all": 213734, "len_all_1": 74883, "len_abs": 991, "len_title": 108, "len_sents": 73832, "len_sents_1": 36477, "n_sents": 659, "n_sents_1": 311, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 998, "title": "MLPs Learn In-Context on Regression and Classification Tasks", "abs": "In-context learning (ICL), the remarkable ability to solve a task from only input exemplars, is often assumed to be a unique hallmark of Transformer models. By examining commonly employed synthetic ICL tasks, we demonstrate that multi-layer perceptrons (MLPs) can also learn in-context. Moreover, MLPs, and the closely related MLP-Mixer models, learn in-context *competitively with Transformers given the same compute budget* in this setting. We further show that MLPs *outperform* Transformers on a series of classical tasks from psychology designed to test relational reasoning, which are closely related to in-context classification. These results underscore a need for studying in-context learning beyond attention-based architectures, while also challenging strong prior arguments about MLPs' limited ability to solve relational tasks. Altogether, our results highlight the unexpected competence of MLPs, and support the growing interest in all-MLP alternatives to task-specific architectures.", "keywords": [ "In-context learning", "relational reasoning", "synthetic tasks", "MLP", "MLP-Mixer", "Transformer" ], "rating_list": [ 3, 6, 8, 8, 6 ], "soundness_list": [ 2, 4, 3, 3, 3 ], "presentation_list": [ 3, 4, 4, 3, 3 ], "contribution_list": [ 1, 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "MbM1BqGpZu", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Hengyu Fu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Mengdi Wang", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Associate Professor" }, { "name": "Minshuo Chen", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Postdoc" }, { "name": "Zehao Dou", "gender": "Male", "institution": "Yale University", "country": "", "position": "PhD student" }, { "name": "Guo Jiawei", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" } ], "n_formula": 191, "n_formula_1": 35, "n_ref_uni": 52, "n_ref": 73, "n_ref_all": 112, "n_fig": 14, "n_tab": 6, "L_tab": 843, "n_element_tab": 49, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 52753, "formula_len_all_1": 41180, "len_all": 376116, "len_all_1": 126232, "len_abs": 1188, "len_title": 145, "len_sents": 94604, "len_sents_1": 36717, "n_sents": 960, "n_sents_1": 314, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1193, "title": "Diffusion Transformer Captures Spatial-Temporal Dependencies: A Theory for Gaussian Process Data", "abs": "Diffusion Transformer, the backbone of Sora for video generation, successfully scales the capacity of diffusion models, pioneering new avenues for high-fidelity sequential data generation. Unlike static data such as images, sequential data consists of consecutive data frames indexed by time, exhibiting rich spatial and temporal dependencies. These dependencies represent the underlying dynamic model and are critical to validate the generated data. In this paper, we make the first theoretical step towards bridging diffusion transformers for capturing spatial-temporal dependencies. Specifically, we establish score approximation and distribution estimation guarantees of diffusion transformers for learning Gaussian process data with covariance functions of various decay patterns. We highlight how the spatial-temporal dependencies are captured and affect learning efficiency. Our study proposes a novel transformer approximation theory, where the transformer acts to unroll an algorithm. We support our theoretical results by numerical experiments, providing strong evidence that spatial-temporal dependencies are captured within attention layers, aligning with our approximation theory.", "keywords": [ "Diffusion model", "transformers", "sequential data", "spatial-temporal dependency", "sample complexity" ], "rating_list": [ 8, 6, 8, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "MbK0Vs5lFI", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Kumar Shridhar", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" }, { "name": "Kushal Jain", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "MS student" }, { "name": "Moritz Miller", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Niket Tandon", "gender": "Male", "institution": "Research, Microsoft", "country": "IN", "position": "Principal Researcher" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 23, "n_ref": 49, "n_ref_all": 66, "n_fig": 10, "n_tab": 3, "L_tab": 1303, "n_element_tab": 153, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1303, "n_element_tab_1": 153, "formula_len_all": 380, "formula_len_all_1": 337, "len_all": 126449, "len_all_1": 62623, "len_abs": 1292, "len_title": 127, "len_sents": 32454, "len_sents_1": 25898, "n_sents": 287, "n_sents_1": 241, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1020, "title": "First-Step Advantage: Importance of Starting Right in Multi-Step Math Reasoning", "abs": "Language models can solve complex reasoning tasks better by learning to generate rationales for their predictions. Often these models know how to solve a task but their auto-regressive decoding nature leads to incorrect results if they start incorrectly. \nWe observe that smaller models in particular, when corrected, can solve a task that they would have otherwise struggled with. We demonstrate this phenomenon by using a larger model to guide smaller models, which leads to significantly improved performance (up to *+24* points on the GSM8K dataset by 7B models). To assist smaller models in initiating the starting step, we propose QuestCoT, where a smaller model first *asks itself how to start*, before proceeding with a chain of reasoning.\nOn various multistep mathematical reasoning datasets over multiple smaller models, we show that getting the right start can lead to significant performance gains across all models (gains of up to *+6* points on GSM8K, *+9* on SVAMP, *+5* on ASDiv, and *+7* on MultiArith).", "keywords": [ "Reasoning", "Planning", "Refinement" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "Mb5vJijcHn", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Cong Tian", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Handing Wang", "gender": "Female", "institution": "Xidian University ", "country": "CN", "position": "Full Professor" }, { "name": "Yaochu Jin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" }, { "name": "Zhaoxin Wang", "gender": "Male", "institution": "Xidian University", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 42, "n_ref": 72, "n_ref_all": 84, "n_fig": 4, "n_tab": 5, "L_tab": 1260, "n_element_tab": 112, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2081, "n_element_tab_1": 245, "formula_len_all": 674, "formula_len_all_1": 675, "len_all": 107415, "len_all_1": 56043, "len_abs": 1887, "len_title": 150, "len_sents": 25689, "len_sents_1": 22944, "n_sents": 178, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1896, "title": "Decoupling Backdoors from Main Task: Toward the Effective and Durable Backdoors in Federated Learning", "abs": "Federated learning, as a distributed machine learning method, enables multiple participants to collaboratively train a central model without sharing their private data. However, this decentralized mechanism introduces new privacy and security concerns. Malicious attackers can embed backdoors into local models, which are inherited by the central global model through the federated aggregation process. While previous studies have demonstrated the effectiveness of backdoor attacks, the effectiveness and durability often rely on unrealistic assumptions, such as a large number of attackers and scaled malicious contributions. These assumptions arise because a sufficient number of attackers can neutralize the contributions of honest participants, allowing the backdoor to be successfully inherited by the central model. In this work, we attribute these backdoor limitations to the coupling between the main and backdoor tasks. To address these backdoor limitations, we propose a min-max backdoor attack framework that decouples backdoors from the main task, ensuring that these two tasks do not interfere with each other. The maximization phase employs the principle of universal adversarial perturbation to create triggers that amplify the performance disparity between poisoned and benign samples. These samples are then used to train a backdoor model in the minimization process. We evaluate the proposed framework in both image classification and semantic analysis tasks. Comparisons with four backdoor attack methods under five defense algorithms show that our method achieves good attack performance even if there is a small number of attackers and when the submitted model parameters are not scaled. In addition, even if attackers are completely removed in the training process, the implanted backdoors will not be dramatically weakened by the contributions of other honest participants.", "keywords": [ "Backdoor Attack\uff0cfederated learning" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 1, 2, 2, 1 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "MazxSMs6Hs", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 1, "author_info_list": [ { "name": "Bonaventure F. P. Dossou", "gender": "Male", "institution": "McGill University, McGill University", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 37, "n_ref": 51, "n_ref_all": 70, "n_fig": 3, "n_tab": 7, "L_tab": 2625, "n_element_tab": 269, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1091, "n_element_tab_1": 125, "formula_len_all": 318, "formula_len_all_1": 318, "len_all": 104303, "len_all_1": 50403, "len_abs": 1333, "len_title": 128, "len_sents": 22610, "len_sents_1": 22610, "n_sents": 164, "n_sents_1": 155, "venue": "ICLR2025", "decision": "withdraw", "L_title": 127, "L_abs": 1355, "title": "Advancing African-Accented English Speech Recognition: Epistemic Uncertainty-Driven Data Selection for Generalizable ASR Models", "abs": "Accents play a pivotal role in shaping human communication, enhancing our ability to convey and comprehend messages with clarity and cultural nuance. While there has been significant progress in Automatic Speech Recognition (ASR), African-accented English ASR has been understudied due to a lack of training datasets, which are often expensive to create and demand colossal human labor. Combining several active learning paradigms and the core-set approach, we propose a new multi-rounds adaptation process that uses epistemic uncertainty to automate the annotation process, significantly reducing the associated costs and human labor. This novel method streamlines data annotation and strategically selects data samples that contribute most to model uncertainty, enhancing training efficiency. We define a new U-WER metric to track model adaptation to hard accents. We evaluate our approach across several domains, datasets, and high-performing speech models. Our results show that our approach leads to a 27\\% WER relative average improvement while requiring, on average, 45\\% less data than established baselines. Our approach also improves out-of-distribution generalization for very low-resource accents, demonstrating its viability for building generalizable ASR models in the context of accented African ASR. We open-source the code \n\\href{}{here}.", "keywords": [ "low-resource African languages", "uncertainty quantification", "automatic speech recognition" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "MagmwodCAB", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Ji Xie", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Intern" }, { "name": "Yi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zongxin Yang", "gender": "Male", "institution": "Harvard Medical School, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "dewei Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 28, "n_ref": 73, "n_ref_all": 99, "n_fig": 16, "n_tab": 6, "L_tab": 1580, "n_element_tab": 135, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1701, "n_element_tab_1": 129, "formula_len_all": 398, "formula_len_all_1": 317, "len_all": 142856, "len_all_1": 61805, "len_abs": 0, "len_title": 175, "len_sents": 48219, "len_sents_1": 27909, "n_sents": 369, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 76, "L_abs": 1374, "title": "3DIS: Depth-Driven Decoupled Instance Synthesis for Text-to-Image Generation", "abs": "The increasing demand for controllable outputs in text-to-image generation has spurred advancements in multi-instance generation (MIG), allowing users to define both instance layouts and attributes. However, unlike image-conditional generation methods such as ControlNet, MIG techniques have not been widely adopted in state-of-the-art models like SD2 and SDXL, primarily due to the challenge of building robust renderers that simultaneously handle instance positioning and attribute rendering. In this paper, we introduce Depth-Driven Decoupled Instance Synthesis (3DIS), a novel framework that decouples the MIG process into two stages: (i) generating a coarse scene depth map for accurate instance positioning and scene composition, and (ii) rendering fine-grained attributes using pre-trained ControlNet on any foundational model, without additional training. Our 3DIS framework integrates a custom adapter into LDM3D for precise depth-based layouts and employs a finetuning-free method for enhanced instance-level attribute rendering. Extensive experiments on COCO-Position and COCO-MIG benchmarks demonstrate that 3DIS significantly outperforms existing methods in both layout precision and attribute rendering. Notably, 3DIS offers seamless compatibility with diverse foundational models, providing a robust, adaptable solution for advanced multi-instance generation.", "keywords": [ "Image Generation; Diffusion Models" ], "rating_list": [ 8, 6, 6, 8, 8 ], "soundness_list": [ 3, 2, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 4, 3 ], "confidence_list": [ 4, 3, 3, 4, 3 ] }, { "paper_id": "MZDdTzN6Cy", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Chenghan Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Mingchen Li", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Ruisheng Diao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 37, "n_formula_1": 13, "n_ref_uni": 35, "n_ref": 81, "n_ref_all": 113, "n_fig": 8, "n_tab": 30, "L_tab": 17004, "n_element_tab": 1537, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 4804, "n_element_tab_1": 176, "formula_len_all": 1917, "formula_len_all_1": 579, "len_all": 210024, "len_all_1": 55838, "len_abs": 1512, "len_title": 136, "len_sents": 50099, "len_sents_1": 21999, "n_sents": 377, "n_sents_1": 148, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1217, "title": "TVNet: A Novel Time Series Analysis Method Based on Dynamic Convolution and 3D-Variation", "abs": "With the recent development and advancement of Transformer and MLP architectures, significant strides have been made in time series analysis. Conversely, the performance of Convolutional Neural Networks (CNNs) in time series analysis has fallen short of expectations, diminishing their potential for future applications. Our research aims to enhance the representational capacity of Convolutional Neural Networks (CNNs) in time series analysis by introducing novel perspectives and design innovations. To be specific, We introduce a novel time series reshaping technique that considers the inter-patch, intra-patch, and cross-variable dimensions. Consequently, we propose TVNet, a dynamic convolutional network leveraging a 3D perspective to employ time series analysis. TVNet retains the computational efficiency of CNNs and achieves state-of-the-art results in five key time series analysis tasks, offering a superior balance of efficiency and performance over the state-of-the-art Transformer-based and MLP-based models. Additionally, our findings suggest that TVNet exhibits enhanced transferability and robustness. Therefore, it provides a new perspective for applying CNN in advanced time series analysis tasks.", "keywords": [ "Time series Analysis", "Dynamic convolution", "Deep Learning" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "MZ324wU7Hj", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Mengmi Zhang", "gender": "Female", "institution": "Nanyang Technological University, Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Shuangpeng Han", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 51, "n_ref": 90, "n_ref_all": 112, "n_fig": 9, "n_tab": 3, "L_tab": 2885, "n_element_tab": 126, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1201, "n_element_tab_1": 56, "formula_len_all": 158, "formula_len_all_1": 137, "len_all": 140138, "len_all_1": 67439, "len_abs": 1306, "len_title": 138, "len_sents": 40005, "len_sents_1": 31976, "n_sents": 350, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1313, "title": "Unveiling AI's Blind Spots: An Oracle for In-Domain, Out-of-Domain, and Adversarial Errors", "abs": "AI models make mistakes when recognizing images\u2014whether in-domain, out-of-domain, or adversarial. Predicting these errors is critical for improving system reliability, reducing costly mistakes, and enabling proactive corrections in real-world applications such as healthcare, finance, and autonomous systems. However, understanding what mistakes AI models make, why they occur, and how to predict them remains an open challenge. Here, we conduct comprehensive empirical evaluations using a \"mentor\" model \u2014a deep neural network designed to predict another model\u2019s errors. Our findings show that the mentor model excels at learning from a mentee's mistakes on adversarial images with small perturbations and generalizes effectively to predict in-domain and out-of-domain errors of the mentee. Additionally, transformer-based mentor models excel at predicting errors across various mentee architectures. Subsequently, we draw insights from these observations and develop an \"oracle\" mentor model, dubbed SuperMentor, that achieves 78\\% accuracy in predicting errors across different error types. Our error prediction framework paves the way for future research on anticipating and correcting AI model behaviours, ultimately increasing trust in AI systems. All code, models, and data will be made publicly available.", "keywords": [ "Error Prediction", "AI reliability", "In-domain", "Out-of-Domain", "Adversarial Attack" ], "rating_list": [ 8, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "MZ1xgIBU3q", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Hao Niu", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Tianyu Qiu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Xiaofeng Gao", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "", "position": "Full Professor" }, { "name": "Xiuqi Huang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yi Xie", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Yun Xiong", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 6, "n_ref_uni": 28, "n_ref": 45, "n_ref_all": 55, "n_fig": 7, "n_tab": 4, "L_tab": 1793, "n_element_tab": 276, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1268, "n_element_tab_1": 220, "formula_len_all": 524, "formula_len_all_1": 230, "len_all": 142577, "len_all_1": 62430, "len_abs": 1200, "len_title": 130, "len_sents": 50903, "len_sents_1": 28491, "n_sents": 379, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1390, "title": "A GENERALIZABLE AND EFFICIENT SYMBOLIC RE- GRESSION METHOD FOR TIME SERIES ANALYSIS", "abs": "The current popular time series analysis methods primarily focus on quantita- tive approaches, which typically offer accurate and diverse statistical indicators. However, these methods often fall short in elucidating the underlying evolution patterns of time series and providing intuitive and qualitative analysis. In this pa- per, we employ a reinforcement learning-inspired approach: using Monte-Carlo Tree Search (MCTS) as the foundation, we introduce symbolic regression tech- niques to derive explicit expressions for the non-linear dynamics in time series evolution. Considering the challenges of excessive randomness during the ac- tion selection phase and low efficiency during the simulation phase in MCTS, we integrate neural networks with MCTS, forming Neural-Enhanced Monte-Carlo Tree Search (NEMoTS) method. By leveraging the excellent fitting ablities of neural networks, we introduce priors for the action selection phase and directly replace the complex and time-consuming simulation process. This integration significantly enhances generalizability and computational efficiency in time series analysis based on symbolic regression. NEMoTS offers a qualitative and intu- itive approach to time series analysis. Experiments with six real-world datasets demonstrate that NEMoTS exhibits significant superiority in performance, effi- ciency, reliability, and interpretability.", "keywords": [ "Symbolic Regression", "Analytical Expression", "Neural-Enhanced MCTS", "Time Series" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 1, 2, 1, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "MYw74B77KQ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Aditya Parameswaran", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Associate Professor" }, { "name": "Sepanta Zeighami", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Zac Wellmer", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "MS student" } ], "n_formula": 115, "n_formula_1": 17, "n_ref_uni": 28, "n_ref": 44, "n_ref_all": 89, "n_fig": 7, "n_tab": 15, "L_tab": 4741, "n_element_tab": 560, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 2302, "n_element_tab_1": 312, "formula_len_all": 7873, "formula_len_all_1": 1245, "len_all": 254048, "len_all_1": 78982, "len_abs": 1549, "len_title": 121, "len_sents": 79330, "len_sents_1": 33821, "n_sents": 724, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1590, "title": "NUDGE: Lightweight Non-Parametric Fine-Tuning of Embeddings for Retrieval", "abs": "$k$-Nearest Neighbor search on dense vector embeddings ($k$-NN retrieval) from pre-trained embedding models is the predominant retrieval method for text and images, as well as Retrieval-Augmented Generation (RAG) pipelines. In practice, application developers often fine-tune the embeddings to improve their accuracy on the dataset and query workload in hand. Existing approaches either fine-tune the pre-trained model itself or, more efficiently, but at the cost of accuracy, train adaptor models to transform the output of the pre-trained model. We present NUDGE, a family of novel *non-parametric* embedding fine-tuning approaches that are significantly more accurate and efficient than both sets of existing approaches. NUDGE directly modifies the embeddings of data records to maximize the accuracy of $k$-NN retrieval. We present a thorough theoretical and experimental study of NUDGE's non-parametric approach. We show that even though the underlying problem is NP-Hard, constrained variations can be solved efficiently. These constraints additionally ensure that the changes to the embeddings are modest, avoiding large distortions to the semantics learned during pre-training. In experiments across five pre-trained models and nine standard text and image retrieval datasets, *NUDGE runs in minutes and often improves NDCG@10 by more than 10\\% over existing fine-tuning methods. On average, NUDGE provides 3.3$\\times$ and 4.3$\\times$ higher increase in accuracy and runs 200$\\times$ and 3$\\times$ faster, respectively, over fine-tuning the pre-trained model and training adaptors.*", "keywords": [ "semantic similarity search", "pre-trained embedding model", "fine-tuning" ], "rating_list": [ 8, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "MY1pJ1RDMv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Gaurav Singh Tomar", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Hagai Taitelbaum", "gender": "Male", "institution": "Research, Google", "country": "IL", "position": "Researcher" }, { "name": "Hexiang Hu", "gender": "Male", "institution": "xAI", "country": "US", "position": "Principal Researcher" }, { "name": "Hsin-Ping Huang", "gender": "Female", "institution": "University of California, Merced", "country": "US", "position": "PhD student" }, { "name": "Kelvin C.K. Chan", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Research Scientist" }, { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Ming-Wei Chang", "gender": "unknown", "institution": "Google", "country": "", "position": "Research scientist" }, { "name": "Xinyi Wang", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Xuhui Jia", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Yonatan Bitton", "gender": "Male", "institution": "Google", "country": "IL", "position": "Research Scientist" }, { "name": "Yu-Chuan Su", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 26, "n_ref": 45, "n_ref_all": 73, "n_fig": 10, "n_tab": 6, "L_tab": 3042, "n_element_tab": 358, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 304, "n_element_tab_1": 34, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 130512, "len_all_1": 60468, "len_abs": 284, "len_title": 206, "len_sents": 43611, "len_sents_1": 30146, "n_sents": 309, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1408, "title": "Kitten: A Knowledge-Intensive Evaluation of Image Generation on Visual Entities", "abs": "Recent advancements in text-to-image generation have significantly enhanced the quality of synthesized images. Despite this progress, evaluations predominantly focus on aesthetic appeal or alignment with text prompts. Consequently, there is limited understanding of whether these models can accurately represent a wide variety of realistic visual entities \u2014 a task requiring real-world knowledge. To address this gap, we propose a benchmark focused on evaluating Knowledge-InTensive image generaTion on real-world ENtities (i.e., KITTEN). Using KITTEN, we conduct a systematic study on the fidelity of entities in text-to-image generation models, focusing on their ability to generate a wide range of real-world visual entities, such as landmark buildings, aircraft, plants, and animals. We evaluate the latest text-to-image models and retrieval-augmented customization models using both automatic metrics and carefully-designed human evaluations, with an emphasis on the fidelity of entities in the generated images. Our findings reveal that even the most advanced text-to-image models often fail to generate entities with accurate visual details. Although retrieval-augmented models can enhance the fidelity of entity by incorporating reference images during testing, they often over-rely on these references and struggle to produce novel configurations of the entity as requested in creative text prompts.", "keywords": [ "Text-to-image generation" ], "rating_list": [ 5, 5, 5, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 5, 3, 5 ] }, { "paper_id": "MWP9V0Ej0d", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Aaron Wenteler", "gender": "Not Specified", "institution": "Queen Mary, University of London", "country": "GB", "position": "PhD student" }, { "name": "Alex Hawkins-Hooker", "gender": "unknown", "institution": "InstaDeep", "country": "", "position": "Intern" }, { "name": "Amaya Gallagher-Syed", "gender": "Female", "institution": "Queen Mary, University of London", "country": "GB", "position": "PhD student" }, { "name": "C\u00e9sar Miguel Valdez C\u00f3rdova", "gender": "Male", "institution": "McGill University", "country": "", "position": "PhD student" }, { "name": "Magdalena Huebner", "gender": "Female", "institution": "Queen Mary, University of London", "country": "", "position": "PhD student" }, { "name": "Martina Occhetta", "gender": "Female", "institution": "Queen Mary, University of London", "country": "GB", "position": "PhD student" }, { "name": "Nikhil Branson", "gender": "unknown", "institution": "Queen Mary, University of London", "country": "GB", "position": "PhD student" }, { "name": "Siu Pui Chung", "gender": "unknown", "institution": "Queen Mary, University of London", "country": "GB", "position": "PhD student" }, { "name": "Victor Curean", "gender": "Male", "institution": "University of Medicine and Pharmacy of Cluj-Napoca", "country": "RO", "position": "PhD student" }, { "name": "William Connell", "gender": "unknown", "institution": "University of California, San Francisco", "country": "", "position": "PhD student" }, { "name": "William Dee", "gender": "Male", "institution": "Queen Mary, University of London", "country": "GB", "position": "PhD student" }, { "name": "Yasha Ektefaie", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 9, "n_ref_uni": 46, "n_ref": 82, "n_ref_all": 104, "n_fig": 10, "n_tab": 5, "L_tab": 2056, "n_element_tab": 65, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 3390, "n_element_tab_1": 24, "formula_len_all": 1868, "formula_len_all_1": 539, "len_all": 209517, "len_all_1": 59541, "len_abs": 1851, "len_title": 140, "len_sents": 64492, "len_sents_1": 25577, "n_sents": 473, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1163, "title": "PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction", "abs": "_In silico_ modeling of transcriptional responses to perturbations is crucial for advancing our understanding of cellular processes and disease mechanisms. We present PertEval-scFM, a standardized framework designed to evaluate models for perturbation effect prediction. We apply PertEval-scFM to benchmark zero-shot single-cell foundation model (scFM) embeddings against simpler baseline models to assess whether these contextualized representations enhance perturbation effect prediction. Our results show that scFM embeddings do not provide consistent improvements over baseline models, especially under distribution shift. Additionally, all models struggle with predicting strong or atypical perturbation effects. Overall, this study provides a systematic evaluation of zero-shot scFM embeddings for perturbation effect prediction, highlighting the challenges of this task and revealing the limitations of current-generation scFMs. Our findings underscore the need for specialized models and high-quality datasets that capture a broader range of cellular states. Source code and documentation can be found at: https://anonymous.4open.science/r/PertEval-C674/.", "keywords": [ "foundation models", "benchmark", "single-cell biology", "perturbation effect prediction" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 1, 4, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "MWHIIWrWWu", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Shanning Zhuang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Vincent Zhuang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Engineer" }, { "name": "Yanan Sui", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yunyue Wei", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 18, "n_ref_uni": 34, "n_ref": 43, "n_ref_all": 63, "n_fig": 10, "n_tab": 1, "L_tab": 187, "n_element_tab": 56, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1552, "formula_len_all_1": 1208, "len_all": 111491, "len_all_1": 61731, "len_abs": 1268, "len_title": 145, "len_sents": 35769, "len_sents_1": 26745, "n_sents": 274, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1269, "title": "Motion Control of High-Dimensional Musculoskeletal System with Hierarchical Model-Based Planning", "abs": "Controlling high-dimensional nonlinear systems presents significant challenges in biological and robotic applications due to the large state and action spaces. While deep reinforcement learning has emerged as the leading approach, it suffers from computationally-intensive and time-consuming, and are not scalable to wide varieties of tasks that each require significant manual tuning. This paper introduces Model Predictive Control with Morphology-aware Proportional Control (MPC$^2$), a novel hierarchical model-based algorithm that addresses these challenges. By integrating a sampling-based model predictive controller for target posture planning with a morphology-aware proportional controller for actuator coordination, our algorithm achieves stable movement control of a 700-actuator musculoskeletal model without training. We show that MPC$^2$ enables zero-shot high-dimensional motion control across diverse movement tasks, such as standing, walking on varying terrains, and sports motion imitation. It can be incorporated into optimal cost function design to automatically optimize the objective, reducing the reliance on traditional reward engineering methods. This work presents a major advancement in (near) real-time control for complex dynamical systems.", "keywords": [ "Model predictive control", "High-dimensional embodied system" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "MW8DN8BE3g", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Haimei Zhao", "gender": "Female", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Hui Zhang", "gender": "Male", "institution": "Samsung Rearch China-Beijing(SRCB)", "country": "CN", "position": "Principle Researcher" }, { "name": "Jing Zhang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Mengchuan Wei", "gender": "Male", "institution": "Samsung", "country": "CN", "position": "Researcher" }, { "name": "Xiaoshuai Hao", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence(BAAl) ", "country": "CN", "position": "Principal Researcher" }, { "name": "Shanghang Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "yifan yang", "gender": "Male", "institution": "Samsung", "country": "CN", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 8, "n_ref_uni": 12, "n_ref": 42, "n_ref_all": 58, "n_fig": 9, "n_tab": 11, "L_tab": 6839, "n_element_tab": 837, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2485, "n_element_tab_1": 376, "formula_len_all": 867, "formula_len_all_1": 389, "len_all": 146883, "len_all_1": 61907, "len_abs": 2214, "len_title": 119, "len_sents": 34994, "len_sents_1": 26696, "n_sents": 289, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1886, "title": "Uni-Map: Unified Camera-LiDAR Perception for Robust HD Map Construction", "abs": "High-definition (HD) map construction methods play a vital role in providing precise and comprehensive static environmental information essential for autonomous driving systems. The primary sensors used are cameras and LiDAR, with input configurations varying among camera-only, LiDAR-only, or camera-LiDAR fusion based on cost-performance considerations, while fusion-based methods typically perform the best. However, current methods face two major issues: high costs due to separate training and deployment for each input configuration, and low robustness when sensors are missing or corrupted. To address these challenges, we propose the Unified Robust HD Map Construction Network (Uni-Map), a single model designed to perform well across all input configurations. Our approach designs a novel Mixture Stack Modality (MSM) training scheme, allowing the map decoder to learn effectively from camera, LiDAR, and fused features. We also introduce a projector module to align Bird's Eye View features from different modalities into a shared space, enhancing representation learning and overall model performance. During inference, our model utilizes a switching modality strategy to adapt seamlessly to any input configuration, ensuring compatibility across various modalities. To evaluate the robustness of HD map construction methods, we designed 13 different sensor corruption scenarios and conducted extensive experiments comparing Uni-Map with state-of-the-art methods. Experimental results show that Uni-Map outperforms previous methods by a significant margin across both normal and corrupted modalities, demonstrating superior performance and robustness. Notably, our unified model surpasses independently trained camera-only, LiDAR-only, and camera-LiDAR MapTR models with a gain of 4.6, 5.6, and 5.6 mAP on the nuScenes dataset, respectively. The source code will be released.", "keywords": [ "HD Map Construction", "Sensor Failures; Out-of-Distribution Robustness" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "MVpvyeVeyI", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Jean Durand", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Researcher" }, { "name": "Sonali Parbhoo", "gender": "unknown", "institution": "Imperial College London, Imperial College London", "country": "GB", "position": "Assistant Professor" }, { "name": "Stefan Bauer", "gender": "unknown", "institution": "Helmholtz Munich ", "country": "DE", "position": "Director" }, { "name": "Yashas Annadani", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" } ], "n_formula": 39, "n_formula_1": 23, "n_ref_uni": 48, "n_ref": 103, "n_ref_all": 113, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3853, "formula_len_all_1": 1738, "len_all": 150915, "len_all_1": 68254, "len_abs": 1277, "len_title": 103, "len_sents": 46616, "len_sents_1": 30403, "n_sents": 452, "n_sents_1": 275, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1242, "title": "Causal Bayesian Optimization with Unknown Causal Graphs", "abs": "Causal Bayesian Optimization (CBO) is a methodology designed to optimize an outcome variable by leveraging known causal relationships through targeted interventions. Traditional CBO methods require a fully and accurately specified causal graph, which is a limitation in many real-world scenarios where such graphs are unknown. To address this, we propose a new method for the CBO framework that operates without prior knowledge of the causal graph. We demonstrate through theoretical analysis and empirical validation that focusing on the direct causal parents of the target variable is sufficient for optimization. Our method learns a Bayesian posterior over the direct parents of the target variable. This allows us to optimize the outcome variable while simultaneously learning the causal structure. Our contributions include a derivation of a closed-form posterior distribution for the linear case. In the nonlinear case, we present a Gaussian Process (GP) approximation that still enables CBO in cases where the posterior is not tractable. The proposed method performs competitively with existing benchmarks and scales well to larger graphs, making it a practical tool for real-world applications where causal information is incomplete.", "keywords": [ "causality; bayesian optimization; causal graph discovery; optimal intervention design" ], "rating_list": [ 10, 3, 8, 5 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "MVosmEvLSb", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Arash Amini", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Guanhao Feng", "gender": "Male", "institution": "City University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Naixin GUO", "gender": "Female", "institution": "City University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Zhixin Zhou", "gender": "Male", "institution": "Alpha Benito Research", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 0, "n_ref": 0, "n_ref_all": 1, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 1262, "len_all_1": 1262, "len_abs": 0, "len_title": 0, "len_sents": 0, "len_sents_1": 0, "n_sents": 0, "n_sents_1": 0, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 51, "title": "Spectral Group Lasso for Selecting Factors Hidden in Plain Sight", "abs": "We decide to withdraw our work from the conference.", "keywords": [ "variable selection", "group lasso", "factor model" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 3, 4, 2, 3 ] }, { "paper_id": "MVltEnKJaO", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Hengyuan Cao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Min Zhang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Qichao Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Tiexiang Li", "gender": "Female", "institution": "Southeast University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoyin Xu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yunhong Lu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 21, "n_ref_uni": 23, "n_ref": 50, "n_ref_all": 72, "n_fig": 8, "n_tab": 4, "L_tab": 2054, "n_element_tab": 236, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2308, "n_element_tab_1": 343, "formula_len_all": 2395, "formula_len_all_1": 1280, "len_all": 120040, "len_all_1": 53900, "len_abs": 1302, "len_title": 126, "len_sents": 30148, "len_sents_1": 20259, "n_sents": 224, "n_sents_1": 144, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1309, "title": "Adversarial Self Flow Matching: Few-steps Image Generation with Straight Flows", "abs": "Flow Matching provides a method to train Ordinary Differential Equation (ODE)-based generative models and facilitates various probability path designs between initial and target distributions. Among these designs, straight flows are particularly interesting for reducing sampling steps. While some works have successfully straightened flows and achieved image generation in a few steps, they often suffer from cumulative errors or provide only piecewise or minibatch-level straightness. We propose Adversarial Self Flow Matching (ASFM), which can straighten flows and align the generated data distribution with the real data distribution. ASFM consists of two complementary components. Online Self Training straightens flows by constructing a conditional vector field using paired data, enabling one-step image generation during training. Adversarial Training aligns the one-step generated data with real data, thereby reducing cumulative errors when straightening flows. Experiments demonstrate that ASFM can build straight flows across the entire time span between two complete distributions and achieve highly competitive results across multiple datasets among Flow Matching-based methods. For instance, ASFM achieves 8.15 and 14.9 FID scores with NFE=6 on CelebA-HQ (256) and AFHQ-Cat (256), respectively.", "keywords": [ "generative models", "flow matching", "adversarial training" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "MVQj0uajaF", "primary_area": "generative models", "n_author": 11, "author_info_list": [ { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Qi Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Renrui Zhang", "gender": "Male", "institution": "MMLab of CUHK & Shanghai AI Laboratory", "country": "HK", "position": "PhD student" }, { "name": "Xudong Lu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yafei Wen", "gender": "Male", "institution": "vivo", "country": "CN", "position": "Researcher" }, { "name": "Yuhui Xu", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Research Scientist" }, { "name": "Ziyi Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Aojun Zhou", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Researcher" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Gao Peng", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Ren Shuai", "gender": "Male", "institution": "vivo", "country": "CN", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 45, "n_ref": 159, "n_ref_all": 180, "n_fig": 12, "n_tab": 4, "L_tab": 527, "n_element_tab": 95, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 2562, "n_element_tab_1": 252, "formula_len_all": 140, "formula_len_all_1": 140, "len_all": 139552, "len_all_1": 62009, "len_abs": 402, "len_title": 99, "len_sents": 33413, "len_sents_1": 26403, "n_sents": 290, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 50, "L_abs": 1300, "title": "TerDiT: Ternary Diffusion Models with Transformers", "abs": "Recent developments in large-scale pre-trained text-to-image diffusion models have significantly improved the generation of high-fidelity images, particularly with the emergence of diffusion transformer models (DiTs). Among diffusion models, diffusion transformers have demonstrated superior image generation capabilities, boosting lower FID scores and higher scalability. However, deploying large-scale DiT models can be expensive due to their excessive parameter numbers. Although existing research has explored efficient deployment techniques for diffusion models such as model quantization, there is still little work concerning DiT-based models. To tackle this research gap, in this paper, we propose **TerDiT**, a quantization-aware training (QAT) and efficient deployment scheme for ternary diffusion transformer models. We focus on the ternarization of DiT networks, with model sizes ranging from 600M to 4.2B, and image resolution from 256$\\times$256 to 512$\\times$512. Our work contributes to the exploration of efficient deployment of large-scale DiT models, demonstrating the feasibility of training extremely low-bit DiT models from scratch while maintaining competitive image generation capacities compared to full-precision models. Code has been uploaded in the supplemental materials.", "keywords": [ "model quantization", "quantization-aware training", "diffusion models" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 5, 3, 3 ] }, { "paper_id": "MV5j4Qpq7N", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Anush Sankaran", "gender": "Male", "institution": "Microsoft", "country": "CA", "position": "Principal Researcher" }, { "name": "Charlotte Siska", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 30, "n_ref": 60, "n_ref_all": 108, "n_fig": 8, "n_tab": 8, "L_tab": 2045, "n_element_tab": 112, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 503, "n_element_tab_1": 34, "formula_len_all": 155, "formula_len_all_1": 155, "len_all": 173861, "len_all_1": 61477, "len_abs": 1915, "len_title": 120, "len_sents": 38348, "len_sents_1": 28733, "n_sents": 315, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1926, "title": "Leveraging System-Prompt Attention to Counteract Novel Jailbreak Attacks", "abs": "In the past few years, Language Models (LMs) have shown par-human capabilities in several domains. \nDespite their practical applications and exceeding user consumption, they are susceptible to jailbreaks when malicious inputs exploit the LM's weaknesses, causing it to deviate from its intended behavior. Current defensive strategies either classify the input prompt as adversarial or prevent LMs from generating harmful outputs. The primary challenge is that the current defense techniques are built against known and established jailbreaking patterns while work poorly against novel attacks. In this research, we propose an end-to-end framework for generating novel attack patterns and demonstrate how the proposed defense approach can generalize over known and unknown attack patterns. Attack patterns are generated using a novel self-learning large language model (LLM)-based multi-agent system with closed loop feedback called ALMAS, which stands for Attack using LLM-based Multi-Agent Systems. We demonstrate that system-prompt attention from Small Language Models (SLMs) can be used to characterize adversarial prompts providing a novel explainable and cheaper defense approach called AttentionDefense. The proposed AttentionDefense is evaluated against existing jailbreak benchmark datasets as well as the novel jailbreaks generated using ALMAS. Ablation studies demonstrate that SLM-based AttentionDefense has equivalent or better jailbreak detection performance as compared to text embedding based classifiers and GPT-4 zero-shot detectors. Our research suggests that the attention mechanism is an integral component in understanding and explaining how LMs respond to malicious inputs that is not captured in the semantic meaning of text embeddings. Additionally, for practical purposes AttentionDefense is an ideal solution as it has the computation requirements of a small LM but the performance of a LLM detector.", "keywords": [ "jailbreaks", "agents", "safeguards", "latent representations" ], "rating_list": [ 3, 3, 1 ], "soundness_list": [ 1, 2, 1 ], "presentation_list": [ 2, 1, 1 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "MUr7Fl93QS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Corrado Rainone", "gender": "Male", "institution": "Qualcomm Inc, QualComm", "country": "NL", "position": "Researcher" }, { "name": "Micha\u00ebl Defferrard", "gender": "Male", "institution": "Qualcomm", "country": "CH", "position": "Researcher" }, { "name": "Roland Memisevic", "gender": "Male", "institution": "Twenty Billion Neurons", "country": "CA", "position": "Chief Scientist" }, { "name": "David W Zhang", "gender": "Male", "institution": "Facebook", "country": "FR", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 23, "n_ref": 34, "n_ref_all": 55, "n_fig": 9, "n_tab": 2, "L_tab": 546, "n_element_tab": 33, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 546, "n_element_tab_1": 33, "formula_len_all": 807, "formula_len_all_1": 804, "len_all": 101285, "len_all_1": 58518, "len_abs": 1091, "len_title": 102, "len_sents": 35848, "len_sents_1": 27790, "n_sents": 260, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1095, "title": "Grounding code understanding in step-by-step execution", "abs": "Auto-regressive language models have made significant inroads in code generation, reasoning, and execution in recent years. Despite the recent progress, however, even the most capable models have been shown to perform significantly worse than humans in the task of predicting what a given piece of code does. This has fueled concerns about the tendency of models that seemingly generate and reason over code to learn shortcuts without developing any deeper understanding of code. Unlike reasoning, the meaning of a line of code is determined entirely by the effect it has on the state of the machine on which it is executed. Inspired by this observation, we propose measuring code understanding as the ability to predict the effects of line-by-line execution of a piece of code. We perform an empirical study which suggests that the inability to track machine state is a key contributor to the deficiencies of existing models to understand code. We also propose a simple solution based on fine-tuning a model on auxiliary state supervision, and we demonstrate the effectiveness of this approach.", "keywords": [ "large language models", "code execution" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 4, 2, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "MUWkqH6e7d", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Jiashi Gao", "gender": "Female", "institution": "Southern University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jiaxin Zhang", "gender": "Female", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Junlei Zhou", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Kexin Liu", "gender": "Female", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Xiangyu Zhao", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Xin Yao", "gender": "Male", "institution": "Lingnan University", "country": "HK", "position": "Full Professor" }, { "name": "Xinwei Guo", "gender": "Not Specified", "institution": "Southern University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Xuetao Wei", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 155, "n_formula_1": 32, "n_ref_uni": 24, "n_ref": 54, "n_ref_all": 77, "n_fig": 7, "n_tab": 4, "L_tab": 1160, "n_element_tab": 164, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 716, "n_element_tab_1": 71, "formula_len_all": 19007, "formula_len_all_1": 2986, "len_all": 170916, "len_all_1": 66142, "len_abs": 1632, "len_title": 178, "len_sents": 53070, "len_sents_1": 31951, "n_sents": 387, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 130, "L_abs": 1655, "title": "Human Expertise Really Matters! Mitigating Unfair Utility Induced by Heterogenous Human Expertise in AI-assisted Decision-Making", "abs": "AI-assisted decision-making often involves an AI model providing confidence, which helps human decision-makers integrate these with their own confidence to make higher-utility final decisions.\nHowever, when human decision-makers are heterogeneous in their expertise, existing AI-assisted decision-making may fail to provide fair utility across them. Such unfairness raises concerns about social welfare among diverse human decision-makers due to inequities in access to equally effective AI assistance, which may reduce their willingness and trust to engage with AI systems. In this work, we investigate how to calibrate AI confidence to provide fair utility for human decision-makers. \nWe first demonstrate that rational decision-makers with heterogeneous expertise are unlikely to obtain fair decision utility from existing AI confidence calibrations. We propose a novel confidence calibration criterion, *inter-group-alignment*, which synergizes with human-alignment to jointly determine the upper bound of utility disparity across human decision-maker groups. Building on this foundation, we propose a new fairness-aware confidence calibration method, *group-level multicalibration*, which ensures a sufficient condition for achieving both inter-group-alignment and human-alignment. \nWe validate our theoretical findings through extensive experiments on four real-world multimodal tasks. The results indicate that our calibrated AI confidence facilitates fairer utility, concurrently enhancing overall utility. *The implementation code is available at* \n [https://anonymous.4open.science/r/iclr4103](https://anonymous.4open.science/r/iclr4103).", "keywords": [ "AI for good", "Human-centric ML", "Fairness", "Calibration" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "MUTTUDKb3M", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Guanghui Wen", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Full Professor" }, { "name": "Yiliu Jiang", "gender": "unknown", "institution": "Southeast University", "country": "", "position": "PhD student" }, { "name": "A. Kai Qin", "gender": "unknown", "institution": "Swinburne University of Technology", "country": "AU", "position": "Full Professor" } ], "n_formula": 39, "n_formula_1": 25, "n_ref_uni": 32, "n_ref": 55, "n_ref_all": 88, "n_fig": 12, "n_tab": 5, "L_tab": 424, "n_element_tab": 53, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 242, "n_element_tab_1": 40, "formula_len_all": 3460, "formula_len_all_1": 2016, "len_all": 140654, "len_all_1": 70749, "len_abs": 1509, "len_title": 137, "len_sents": 46004, "len_sents_1": 31721, "n_sents": 364, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1519, "title": "DRIMA: Differential Reward Interaction for Cooperative Multi-Agent Reinforcement Learning", "abs": "Multi-agent reinforcement learning (MARL) owning to its potent capabilities in complex systems has gained remarkable research attention nowadays, in which collaborative decision-making and control for multi-agent systems is one of the key research focuses.\nThe prevalent learning framework is centralized training with decentralized execution (CTDE), in which the decentralized execution realizes strategy flexibility, and the use of centralized training ensures stationarity and goal consistency while becoming incapable when facing scalability and complexity situations.\nTo address this issue, we follow the concept of distributed training with decentralized execution (DTDE).\nDecentralization is naturally accompanied by the game during the learning process, which has not been entirely studied in related work, resulting in the constrained strategy combination of MARL.\nIn this paper, we devise a novel approach of differential reward interaction (DRI) with conflict-triggered for the distributed evaluation that enables overall goal consistency through highly efficient local information exchange.\nWith this collaborative learning method, the DRI-based MARL can eliminate the notorious issue of converging to saddle equilibriums of stochastic games.\nMeanwhile, it possesses provable convergence and is well compatible for general value-based and policy-based algorithms.\nExperiments in several benchmark scenarios demonstrate that DRIMA realizes collaborative strategy learning with enhanced global goal-achieving.", "keywords": [ "Multi-agent reinforcement learning", "Distributed training and execution", "Multi-player stochastic game" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 2 ] }, { "paper_id": "MUL7tKvNei", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Aoming Liang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Cong Wang", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "Postdoc" }, { "name": "Dixia Fan", "gender": "Male", "institution": "Westlake University", "country": "", "position": "Assistant Professor" }, { "name": "Hao Tang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ling Shao", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Mingming Ge", "gender": "unknown", "institution": "BNU-HKBU United International College", "country": "CN", "position": "Assistant Professor" }, { "name": "Pengxiao Lin", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Zhaoyang Mu", "gender": "Male", "institution": "Dalian Martime University", "country": "CN", "position": "MS student" } ], "n_formula": 22, "n_formula_1": 22, "n_ref_uni": 48, "n_ref": 73, "n_ref_all": 80, "n_fig": 4, "n_tab": 6, "L_tab": 2365, "n_element_tab": 173, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 856, "n_element_tab_1": 43, "formula_len_all": 1272, "formula_len_all_1": 1166, "len_all": 141898, "len_all_1": 57845, "len_abs": 1157, "len_title": 163, "len_sents": 35076, "len_sents_1": 24654, "n_sents": 300, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 116, "L_abs": 1167, "title": "M$^2$M: LEARNING CONTROLLABLE MULTI OF EXPERTS AND MULTI-SCALE OPERATORS ARE THE PARTIAL DIFFERENTIAL EQUATIONS NEED", "abs": "Learning the evolutionary dynamics of Partial Differential Equations (PDEs) is critical in understanding dynamic systems, yet current methods insufficiently learn their representations. This is largely due to the multi-scale nature of the solution, where certain regions exhibit rapid oscillations while others evolve more slowly. This paper introduces a framework of multi-scale and multi-expert (M$^2$M) neural operators designed to simulate and learn PDEs efficiently. We employ a divide-and-conquer strategy to train a multi-expert gated network for the dynamic router policy. Our method incorporates a controllable prior gating mechanism that determines the selection rights of experts, enhancing the model's efficiency. To optimize the learning process, we have implemented a PI (Proportional, Integral) control strategy to adjust the allocation rules precisely. This universal controllable approach allows the model to achieve greater accuracy. We test our approach on benchmark 2D Navier-Stokes equations and provide a custom multi-scale dataset. M$^2$M can achieve higher simulation accuracy and offer improved interpretability compared to baseline methods.", "keywords": [ "Multi-experts and Multi-scales; Controlled operator learning; machine learing in PDEs" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 1, 2, 2 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 5, 5, 2 ] }, { "paper_id": "MU4aykgggg", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Bin Li", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Mingzhao Yang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Xiangyang Xue", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 17, "n_ref": 24, "n_ref_all": 45, "n_fig": 2, "n_tab": 1, "L_tab": 15, "n_element_tab": 3, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 1972, "n_element_tab_1": 25, "formula_len_all": 949, "formula_len_all_1": 949, "len_all": 115266, "len_all_1": 53589, "len_abs": 1536, "len_title": 90, "len_sents": 30467, "len_sents_1": 26085, "n_sents": 198, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 1541, "title": "Federated Unlearning with Diffusion Models", "abs": "In recent years, diffusion models are widely adopted by individual users due to their outstanding performance in generation. During usage, individual users develop a need to forget privacy-related contents, making the scenario of using diffusion models on the clients a natural federated unlearning setting. For this scenario, we propose FedDUL, a Federated UnLearning method with Diffusion models, which addresses the unlearn requests from clients using the diffusion models. On one hand, we utilize local data on the clients to perform attention-based unlearning, enabling the local diffusion model to forget the concepts specified by the clients. On the other hand, we filter and group the unlearn requests from clients, gradually aggregating reasonable requests into the global diffusion model on the server, thereby protecting client privacy within the global model. The theoretical analysis further demonstrates the inherent unity between the federated unlearning problem based on diffusion models and federated learning, and extend this unity to traditional federated unlearning methods. Extensive quantitation and visualization experiments are conducted to evaluate the unlearning of both local and global models and discuss the communication and computation costs of our method, demonstrating that our method can satisfy the unlearn requests of multiple clients without compromising the generative capabilities for irrelevant concepts, providing new ideas and methods for the application of diffusion models in federated unlearning.", "keywords": [ "Federated Unlearning", "Diffusion Model" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 1, 2, 2, 1 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "MTcgsz1SHr", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Boning Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Longbo Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 52, "n_ref": 175, "n_ref_all": 185, "n_fig": 2, "n_tab": 7, "L_tab": 578, "n_element_tab": 56, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 384, "n_element_tab_1": 19, "formula_len_all": 1048, "formula_len_all_1": 745, "len_all": 175087, "len_all_1": 67833, "len_abs": 1864, "len_title": 87, "len_sents": 55934, "len_sents_1": 32719, "n_sents": 377, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1805, "title": "Efficient Online Pruning and Abstraction for Imperfect Information Extensive-Form Games", "abs": "Efficiently computing approximate equilibrium strategies in large Imperfect Information Extensive-Form Games (IIEFGs) poses significant challenges due to the vast size of the game tree. Pruning and abstraction methods effectively reduce this complexity, enhancing computational efficiency. However, seamlessly integrating pruning techniques with variants of Counterfactual Regret Minimization (CFR), a leading method for solving IIEFGs, remains a complex task. Furthermore, existing information abstraction methods often involve high computational costs and may require extensive offline pre-computation, limiting their practical applicability. In this paper, we introduce Expected-Value Pruning and Abstraction (EVPA), an online approach that improves efficiency by leveraging expected value estimation within information sets. EVPA consists of three core components: expected value estimation of information sets, expected value-based pruning, and information abstraction for subgames. It estimates the expected value of information sets based on approximate Nash equilibrium strategies, employing these estimations for both pruning and abstraction. By integrating Minimax pruning with CFR, EVPA streamlines decision-making by permanently eliminating sub-optimal actions from the game tree before CFR starts. Additionally, EVPA features an advanced information abstraction mechanism that merges information sets based on both current and future expected values in the subgame, achieving efficient online abstraction. Experiments on HUNL demonstrate that EVPA outperforms DeepStack's replication and Slumbot with significant win-rate margins in multiple settings. Remarkably, EVPA requires only $1$\\%-$2$\\% of the solving time to reach an approximate Nash equilibrium compared to DeepStack's replication.", "keywords": [ "Game Theory", "Imperfect Information Games", "Counterfactual Regret Minimization", "Poker", "Machine Learning" ], "rating_list": [ 6, 6, 3, 8 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "MT3aOfXIbY", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Linda Cai", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Shivam Gupta", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Sitan Chen", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 159, "n_formula_1": 33, "n_ref_uni": 28, "n_ref": 129, "n_ref_all": 163, "n_fig": 1, "n_tab": 5, "L_tab": 1782, "n_element_tab": 120, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 18, "n_element_tab_1": 1, "formula_len_all": 23132, "formula_len_all_1": 1871, "len_all": 230499, "len_all_1": 67711, "len_abs": 1230, "len_title": 124, "len_sents": 64698, "len_sents_1": 29657, "n_sents": 635, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1317, "title": "Faster Diffusion Sampling with Randomized Midpoints: Sequential and Parallel", "abs": "Sampling algorithms play an important role in controlling the quality and runtime of diffusion model inference. In recent years, a number of works (Chen et al., 2023c;b; Benton et al., 2023; Lee et al., 2022) have analyzed algorithms for diffusion sampling with provable guarantees; these works show that for essentially any data distribution, one can approximately sample in polynomial time given a sufficiently accurate estimate of its score functions at different noise levels. \n\nIn this work, we propose a new scheme inspired by Shen and Lee's randomized midpoint method for log-concave sampling (Shen & Lee, 2019). We prove that this approach achieves the best known dimension dependence for sampling from arbitrary smooth distributions in total variation distance ($\\widetilde O(d^{5/12})$ compared to $\\widetilde O(\\sqrt{d})$ from prior work). We also show that our algorithm can be parallelized to run in only $\\widetilde O(\\log^2 d)$ parallel rounds, constituting the first provable guarantees for parallel sampling with diffusion models.\n \nAs a byproduct of our methods, for the well-studied problem of log-concave sampling in total variation distance, we give an algorithm and simple analysis achieving dimension dependence $\\widetilde O(d^{5/12})$ compared to $\\widetilde O(\\sqrt{d})$ from prior work.", "keywords": [ "Diffusion Sampling", "Generative Model", "Statistical Theory" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "MSxCBXD5C8", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Bijun Li", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Hailong Shi", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Jingyi Wu", "gender": "Female", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Jinsheng Xiao", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Associate Professor" }, { "name": "Rui-di Chen", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Undergrad student" }, { "name": "Xingyu Gao", "gender": "unknown", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 19, "n_ref": 34, "n_ref_all": 57, "n_fig": 9, "n_tab": 4, "L_tab": 1894, "n_element_tab": 256, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1894, "n_element_tab_1": 256, "formula_len_all": 627, "formula_len_all_1": 627, "len_all": 93769, "len_all_1": 63305, "len_abs": 1365, "len_title": 131, "len_sents": 27455, "len_sents_1": 26600, "n_sents": 218, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1322, "title": "Anomalous Action Recognition via Spatio-temporal Relation and Key Patch Selection", "abs": "For providing timely warnings and preventing potential damages, it is crucial to detect anomalous actions that threaten public safety through surveillance cameras. Compared to normal actions, anomalous actions often occupy only a small portion of surveillance videos and exhibit more complex manifestations in terms of time and space. Considering that normal action recognition methods fail to highlight crucial information from small-sized patches, resulting in imprecise anomaly modeling, we propose the Spatio-Temporal Key Patch Selection Network (SKPS-Net). To tackle the challenge of detecting anomalous behaviors that manifest in small and inconspicuous areas, we design a spatial adaptive key patch selection module to select small but informative patches on input videos. Furthermore, the long-short feature map spatio-temporal relation module is devised to make the key patch effectively capture the continuous dynamic changes of anomalous actions. Finally, we propose a spatio-temporal refined loss to reinforce fine-grained feature learning. Experiments conducted on the HMDB51, Kinetics, and UCF-Crime v2 datasets demonstrate that our SKPS-Net achieves state-of-the-art performance in few-shot action recognition, outperforming the most competitive methods by 1.2% on the anomalous action dataset UCF-Crime v2.", "keywords": [ "Anomalous Action Recognition; Spatio-temporal Relation;Key Patch Selection" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "MSlF3GvUXI", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Deyuan Liu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Tao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "peng sun", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Xufeng li", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Undergrad student" } ], "n_formula": 151, "n_formula_1": 16, "n_ref_uni": 65, "n_ref": 90, "n_ref_all": 110, "n_fig": 10, "n_tab": 3, "L_tab": 890, "n_element_tab": 82, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1568, "n_element_tab_1": 20, "formula_len_all": 8517, "formula_len_all_1": 1137, "len_all": 332346, "len_all_1": 67257, "len_abs": 1644, "len_title": 82, "len_sents": 83188, "len_sents_1": 31583, "n_sents": 682, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 34, "L_abs": 1614, "title": "Structured-Initialization Learning", "abs": "The emergence of large language models (LLMs) has revolutionized natural language processing, but their development and deployment face significant challenges in computational resources and environmental sustainability.\nTraditional self-supervised learning (SSL) paradigms requiring extensive computational infrastructure and exhibiting slow convergence rates, leading to increased energy consumption and longer training durations.\nWhile existing model fine-tuning techniques such as Low-Rank Adaptation (LoRA) are resource-intensive and fail to facilitate swift knowledge updates when integrating a mount of new data in model version iteration.\nTo mitigate these challenges, we introduce Sail, a novel method for accelerating the training of neural network models by leveraging knowledge from (publicly available) pre-trained models.\nOur approach comprises two key components: (1) a parameter transformation technique that adjusts the dimensions of pre-trained model parameters to match the target architecture, and (2) a proximal parameter integration and retraining strategy that efficiently combines transformed parameters to initialize new models.\nWe formalize the concept of Proximal Parameter and provide theoretical guarantees for its convergence advantages.\nOur approach achieves substantial reductions in training time and computational resources while maintaining or improving model performance on downstream tasks.\nThese results indicate that Sail provides a promising direction for the more efficient and accessible development of the deep learning community.\nOur code will be made publicly available.", "keywords": [ "Efficient Learning" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "MRnZ1KEXSt", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Chongxuan Li", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Hanzhong Allan Guo", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jingdong Wang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Chief Scientist for Computer Vision" }, { "name": "Xinyu Zhang", "gender": "Female", "institution": "University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Zebin You", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 8, "n_ref_uni": 50, "n_ref": 113, "n_ref_all": 130, "n_fig": 9, "n_tab": 12, "L_tab": 2703, "n_element_tab": 421, "n_fig_1": 7, "n_tab_1": 7, "L_tab_1": 2184, "n_element_tab_1": 314, "formula_len_all": 203, "formula_len_all_1": 332, "len_all": 163096, "len_all_1": 63261, "len_abs": 1561, "len_title": 125, "len_sents": 45770, "len_sents_1": 27840, "n_sents": 363, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1227, "title": "Are Images Indistinguishable to Humans Also Indistinguishable to Classifiers?", "abs": "The ultimate goal of generative models is to perfectly capture the data distribution. For image generation, common metrics of visual quality (e.g., FID) and the perceived truthfulness of generated images seem to suggest that we are nearing this goal. However, through \\emph{distribution classification} tasks, we reveal that, from the perspective of neural network-based classifiers, even advanced diffusion models are still far from this goal. Specifically, classifiers are able to consistently and effortlessly distinguish real images from generated ones across various settings. Moreover, we uncover an intriguing discrepancy: classifiers can easily differentiate between diffusion models with comparable performance (e.g., U-ViT-H vs. DiT-XL), but struggle to distinguish between models within the same family but of different scales (e.g., EDM2-XS vs. EDM2-XXL). Our methodology carries several important implications. First, it naturally serves as a diagnostic tool for diffusion models by analyzing specific features of generated data. Second, it sheds light on the model autophagy disorder and offers insights into the use of generated data: augmenting real data with generated data is more effective than replacing it.", "keywords": [ "synthesis data", "diffusion models", "generated data", "image classification", "distribution classification task" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 1 ], "confidence_list": [ 3, 3, 3, 3, 3 ] }, { "paper_id": "MRYyOaNxh3", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Hyeongjin Kim", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Undergrad student" }, { "name": "Jaein Jang", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Undergrad student" }, { "name": "Junghyun Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Se-Young Yun", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "Assistant Professor" }, { "name": "Siyeol Kim", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Undergrad student" }, { "name": "Suhin Shin", "gender": "Female", "institution": "Yonsei University", "country": "KR", "position": "Undergrad student" }, { "name": "Wonbeen Oh", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Undergrad student" }, { "name": "Woosung Koh", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Intern" } ], "n_formula": 33, "n_formula_1": 16, "n_ref_uni": 40, "n_ref": 75, "n_ref_all": 116, "n_fig": 13, "n_tab": 13, "L_tab": 6022, "n_element_tab": 667, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1368, "n_element_tab_1": 95, "formula_len_all": 4413, "formula_len_all_1": 1408, "len_all": 190207, "len_all_1": 66075, "len_abs": 1488, "len_title": 114, "len_sents": 45725, "len_sents_1": 26974, "n_sents": 415, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1606, "title": "FlickerFusion: Intra-trajectory Domain Generalizing Multi-agent Reinforcement Learning", "abs": "Multi-agent reinforcement learning has demonstrated significant potential in addressing complex cooperative tasks across various real-world applications. However, existing MARL approaches often rely on the restrictive assumption that the number of entities (e.g., agents, obstacles) remains constant between training and inference. This overlooks scenarios where entities are dynamically removed or $\\textit{added}$ $\\textit{during}$ the inference trajectory\u2014a common occurrence in real-world environments like search and rescue missions and dynamic combat situations. In this paper, we tackle the challenge of intra-trajectory dynamic entity composition under zero-shot out-of-domain (OOD) generalization, where such dynamic changes cannot be anticipated beforehand. Our empirical studies reveal that existing MARL methods suffer $\\textit{significant}$ performance degradation and increased uncertainty in these scenarios. In response, we propose FlickerFusion, a novel OOD generalization method that acts as a $\\textit{universally}$ applicable augmentation technique for MARL backbone methods. FlickerFusion stochastically drops out parts of the observation space, emulating being in-domain when inferenced OOD. The results show that FlickerFusion not only achieves superior inference rewards but also $\\textit{uniquely}$ reduces uncertainty vis-\u00e0-vis the backbone, compared to existing methods. Benchmarks, implementations, and model weights are organized and open-sourced at $\\texttt{\\href{flickerfusion305.github.io}{\\textbf{flickerfusion305.github.io}}}$, accompanied by ample demo video renderings.", "keywords": [ "Domain Generalization", "Multi-agent Reinforcement Learning", "Benchmark" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "MRPCIForrE", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Amir Nassereldine", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Chenhui Xu", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Dancheng Liu", "gender": "Male", "institution": "University at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Jiajie Li", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Jinjun Xiong", "gender": "unknown", "institution": "State University of New York at Buffalo", "country": "US", "position": "Professor" }, { "name": "Zhaohui Li", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Postdoc" } ], "n_formula": 139, "n_formula_1": 19, "n_ref_uni": 39, "n_ref": 65, "n_ref_all": 67, "n_fig": 0, "n_tab": 2, "L_tab": 456, "n_element_tab": 11, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7643, "formula_len_all_1": 1154, "len_all": 197077, "len_all_1": 66689, "len_abs": 1471, "len_title": 164, "len_sents": 61454, "len_sents_1": 33017, "n_sents": 597, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 116, "L_abs": 1478, "title": "Towards Understanding Multi-Round Large Language Model Reasoning: Approximability, Learnability and Generalizability", "abs": "Recent advancements in cognitive science and multi-round reasoning techniques for Large Language Models (LLMs) suggest that iterative thinking processes improve problem-solving performance in complex tasks. Inspired by this, approaches like Chain-of-Thought, debating, and self-refinement have been applied to auto-regressive LLMs, achieving significant successes in tasks such as mathematical reasoning, commonsense reasoning, and multi-hop question answering. Despite these successes, the theoretical basis for how multi-round reasoning enhances problem-solving abilities remains underexplored.\nIn this work, we investigate the approximation, learnability, and generalization properties of multi-round auto-regressive models. We show that Transformers with finite context windows are universal approximators for steps of Turing-computable functions and can approximate any Turing-computable sequence-to-sequence function through multi-round reasoning. We extend PAC learning to sequence generation and demonstrate that multi-round generation is learnable even when the sequence length exceeds the model's context window. \nFinally, we examine how generalization error propagates across rounds, and show how the aforementioned approaches can help constrain this error, ensuring outputs stay within an expectation boundary. This work sheds light on the systemic theoretical foundations of multi-round sequence learning and reasoning, emphasizing its role in inference complexity.", "keywords": [ "Large Language Model", "Auto-regressive Language Model", "Next-token Prediction" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 1, 1, 3, 4 ] }, { "paper_id": "MR6RZQKMby", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Huajun Chen", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Ningyu Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Shumin Deng", "gender": "Female", "institution": "National University of Singapore", "country": "", "position": "Postdoc" }, { "name": "Yedi Hu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Intern" }, { "name": "Yunzhi Yao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 41, "n_ref": 68, "n_ref_all": 90, "n_fig": 12, "n_tab": 2, "L_tab": 352, "n_element_tab": 45, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 528, "n_element_tab_1": 47, "formula_len_all": 194, "formula_len_all_1": 202, "len_all": 164068, "len_all_1": 49830, "len_abs": 998, "len_title": 57, "len_sents": 37454, "len_sents_1": 22433, "n_sents": 290, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1255, "title": "Exploring Model Kinship for Merging Large Language Models", "abs": "Model merging has become one of the key technologies for enhancing the capabilities and efficiency of Large Language Models (LLMs). However, our understanding of the expected performance gains and principles when merging any two models remains limited. In this work, we introduce model kinship, the degree of similarity or relatedness between LLMs, analogous to biological evolution. With comprehensive empirical analysis, we find that there is a certain relationship between model kinship and the performance gains after model merging, which can help guide our selection of candidate models. Inspired by this, we propose a new model merging strategy: Top-$k$ Greedy Merging with Model Kinship, which can yield better performance on benchmark datasets. Specifically, we discover that using model kinship as a criterion can assist us in continuously performing model merging, alleviating the degradation (local optima) in model evolution, whereas model kinship can serve as a guide to escape these traps. Furthermore, we observe that model kinship can serve as an early stopping criterion, enhancing the efficiency of model evolution. We hope that the proposed model kinship can provide guidance and insights for the future merging and evolution of LLMs.", "keywords": [ "model merging", "model kinship", "large language models" ], "rating_list": [ 6, 6, 8, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "MQXrTMonT1", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Elvis Dohmatob", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Francois Charton", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Engineer" }, { "name": "Julia Kempe", "gender": "Female", "institution": "New York University", "country": "US", "position": "Professor" }, { "name": "Pu Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yunzhen Feng", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Intern" } ], "n_formula": 30, "n_formula_1": 11, "n_ref_uni": 48, "n_ref": 84, "n_ref_all": 111, "n_fig": 5, "n_tab": 6, "L_tab": 661, "n_element_tab": 65, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 238, "n_element_tab_1": 17, "formula_len_all": 1554, "formula_len_all_1": 727, "len_all": 221833, "len_all_1": 70458, "len_abs": 1664, "len_title": 126, "len_sents": 71450, "len_sents_1": 32660, "n_sents": 613, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1216, "title": "Beyond Model Collapse: Scaling Up with Synthesized Data Requires Verification", "abs": "Large Language Models (LLM) are increasingly trained on data generated by other LLMs, either because generated text and images become part of the pre-training corpus, or because synthetized data is used as a replacement for expensive human-annotation. This raises concerns about *model collapse*, a drop in model performance when their training sets include generated data. Considering that it is easier for both humans and machines to tell between good and bad examples than to generate high-quality samples, we investigate the use of verification on synthesized data to prevent model collapse. We provide a theoretical characterization using Gaussian mixtures, linear classifiers, and linear verifiers to derive conditions with measurable proxies to assess whether the verifier can effectively select synthesized data that leads to optimal performance. We experiment with two practical tasks -- computing matrix eigenvalues with transformers and news summarization with LLMs -- which both exhibit model collapse when trained on generated data, and show that verifiers, even imperfect ones, can indeed be harnessed to prevent model collapse and that our proposed proxy measure strongly correlates with performance.", "keywords": [ "Learning with Synthetic Data", "Data Curation", "Avoiding Model Collapse" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "MPJ4SMnScw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Feng Zhu", "gender": "Male", "institution": "SenseTime Group LTD", "country": "CN", "position": "Researcher" }, { "name": "Haodong Zhang", "gender": "Non-Binary", "institution": "SenseTime", "country": "", "position": "Researcher" }, { "name": "Jiangyan Feng", "gender": "Female", "institution": "SenseTime", "country": "", "position": "Researcher" }, { "name": "Lijun GONG", "gender": "Female", "institution": "Sensetime", "country": "", "position": "Researcher" }, { "name": "Ming-Ming Cheng", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Full Professor" }, { "name": "Qibin Hou", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Rui Zhao", "gender": "Male", "institution": "SenseTime Research", "country": "CN", "position": "Researcher" }, { "name": "Yibing Song", "gender": "unknown", "institution": "Alibaba DAMO Academy", "country": "CN", "position": "Staff Scientist" }, { "name": "Yuming Chen", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 61, "n_ref": 148, "n_ref_all": 175, "n_fig": 11, "n_tab": 10, "L_tab": 4756, "n_element_tab": 441, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 266471, "len_all_1": 70966, "len_abs": 3080, "len_title": 111, "len_sents": 72420, "len_sents_1": 30286, "n_sents": 750, "n_sents_1": 285, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1716, "title": "Re-Aligning Language to Visual Objects with an Agentic Workflow", "abs": "Language-based object detection (LOD) aims to align visual objects with language expressions. A large amount of paired data is utilized to improve LOD model generalizations. During the training process, recent studies leverage vision-language models (VLMs) to automatically generate human-like expressions for visual objects, facilitating training data scaling up. In this process, we observe that VLM hallucinations bring inaccurate object descriptions (e.g., object name, color, and shape) to deteriorate VL alignment quality. To reduce VLM hallucinations, we propose an agentic workflow controlled by an LLM to re-align language to visual objects via adaptively adjusting image and text prompts. We name this workflow Real-LOD, which includes planning, tool use, and reflection steps. Given an image with detected objects and VLM raw language expressions, Real-LOD reasons its state automatically and arranges action based on our neural symbolic designs (i.e., planning). The action will adaptively adjust the image and text prompts and send them to VLMs for object re-description (i.e., tool use). Then, we use another LLM to analyze these refined expressions for feedback. These steps are conducted in a cyclic form to gradually improve language descriptions for re-aligning to visual objects. We construct a dataset that contains a tiny amount of 0.18M images with re-aligned language expression and train a prevalent LOD model to surpass existing LOD methods by around 50% on the standard benchmarks. Our Real-LOD workflow, with automatic VL refinement, reveals a potential to preserve data quality along with scaling up data quantity, which further improves LOD performance from a data-alignment perspective.", "keywords": [ "agentic workflow", "language-based object detection" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "MOEqbKoozj", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Fan Yang", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Renjing Xu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhengpeng Xie", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Intern" }, { "name": "QIANG ZHANG", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 44, "n_formula_1": 40, "n_ref_uni": 35, "n_ref": 78, "n_ref_all": 97, "n_fig": 8, "n_tab": 9, "L_tab": 3692, "n_element_tab": 453, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 333, "n_element_tab_1": 27, "formula_len_all": 4078, "formula_len_all_1": 3131, "len_all": 131655, "len_all_1": 52857, "len_abs": 1252, "len_title": 74, "len_sents": 31094, "len_sents_1": 21106, "n_sents": 243, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 26, "L_abs": 1258, "title": "Simple Policy Optimization", "abs": "Model-free reinforcement learning algorithms have seen remarkable progress, but key challenges remain. Trust Region Policy Optimization (TRPO) is known for ensuring monotonic policy improvement through conservative updates within a trust region, backed by strong theoretical guarantees. However, its reliance on complex second-order optimization limits its practical efficiency. Proximal Policy Optimization (PPO) addresses this by simplifying TRPO's approach using ratio clipping, improving efficiency but sacrificing some theoretical robustness. This raises a natural question: Can we combine the strengths of both methods? In this paper, we introduce Simple Policy Optimization (SPO), a novel unconstrained first-order algorithm. SPO integrates the surrogate objective with Total Variation (TV) divergence instead of Kullback-Leibler (KL) divergence, achieving a balance between the theoretical rigor of TRPO and the efficiency of PPO. Our new objective improves upon ratio clipping, offering stronger theoretical properties and better constraining the probability ratio within the trust region. Empirical results demonstrate that SPO outperforms PPO with a simple implementation, particularly for training large, complex network architectures end-to-end.", "keywords": [ "Reinforcement Learning", "Policy Optimization" ], "rating_list": [ 8, 8, 3, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 4, 4, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "MOEBghZGVq", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "David Lindner", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Kai Fronsdal", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 0, "n_ref_uni": 39, "n_ref": 50, "n_ref_all": 62, "n_fig": 5, "n_tab": 10, "L_tab": 7106, "n_element_tab": 384, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 257, "formula_len_all_1": 0, "len_all": 244950, "len_all_1": 62332, "len_abs": 6903, "len_title": 110, "len_sents": 71107, "len_sents_1": 31992, "n_sents": 793, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1000, "title": "MISR: Measuring Instrumental Self-Reasoning in Frontier Models", "abs": "We propose a suite of tasks to evaluate the instrumental self-reasoning ability of large language model (LLM) agents. Instrumental self-reasoning ability could improve adaptability and enable self-modification, but it could also pose significant risks, such as enabling deceptive alignment. Prior work has only evaluated self-reasoning in non-agentic settings or in limited domains. In this paper, we propose evaluations for instrumental self-reasoning ability in agentic tasks in a wide range of scenarios, including self-modification, knowledge seeking, and opaque self-reasoning. We evaluate agents built using state-of-the-art LLMs, including commercial and open source systems. We find that instrumental self-reasoning ability emerges only in the most capable frontier models and that it is highly context-dependent. No model passes the the most difficult versions of our evaluations, hence our evaluation can be used to measure increases in instrumental self-reasoning ability in future models.", "keywords": [ "Self-Reasoning", "Agents", "AI Safety", "Evaluations", "Alignment" ], "rating_list": [ 3, 5, 8, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "MOCEoNsjEx", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Guido F. Montufar", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Full Professor" }, { "name": "Jakob Nikolas Kather", "gender": "Male", "institution": "Technische Universit\u00e4t Dresden", "country": "DE", "position": "Full Professor" }, { "name": "Lena Seifert", "gender": "unknown", "institution": "Technische Universit\u00e4t Dresden", "country": "DE", "position": "Full Professor" }, { "name": "Nithya Bhasker", "gender": "Female", "institution": "National Center for Tumor diseases, Dresden", "country": "", "position": "PhD student" }, { "name": "Stefanie Speidel", "gender": "Female", "institution": "National Center for Tumor Diseases, Dresden", "country": "DE", "position": "Full Professor" }, { "name": "Susu Hu", "gender": "Female", "institution": "National Center for Tumor Diseases, Dresden", "country": "DE", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 24, "n_ref": 38, "n_ref_all": 52, "n_fig": 4, "n_tab": 5, "L_tab": 702, "n_element_tab": 90, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 635, "n_element_tab_1": 83, "formula_len_all": 473, "formula_len_all_1": 475, "len_all": 88028, "len_all_1": 62960, "len_abs": 1259, "len_title": 145, "len_sents": 25292, "len_sents_1": 22625, "n_sents": 209, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1215, "title": "Pg-GAT: A Complete Graph Model for Cancer Detection and Subtyping in Whole Slide Images Analysis", "abs": "Whole-Slide-Images (WSIs) have generated significant interests in cancer research community, owing to their availability and the rich information that they provide. Previous Multiple Instance Learning (MIL) methods often \nneglect the topological structure of tissues which is closely related to tumor evolution. Some attempts with transformer-based MIL methods take spatial relation into account with a trade-off of computational complexity. We propose Projection-gated Graph Attention Network (Pg-GAT), a lightweight model that effectively leverages graph neural network to provide structural prior, learns spatial and contextual relations through graph attention, and mitigates tissue morphology redundancy with differentiable projection-gated pooling, maintaining a data-adaptive decision boundary. In addition, Pg-GAT outputs region-of-interest (ROI) with respect to the graph-level prediction with post-hoc graph explainer, offering tumor localization and model interpretability. We evaluate our method on lymph node metastasis datasets (CAMELYON16 and CAMELYON17) and non-small cell lung cancer (TCGA-NSCLC), achieving AUCs of 97.6\\% and 95.6\\% and 99.6\\% respectively, outperforming state-of-the-art methods.", "keywords": [ "graph neural network", "whole slide images", "cancer research" ], "rating_list": [ 5, 3, 3, 1, 3 ], "soundness_list": [ 3, 2, 1, 1, 1 ], "presentation_list": [ 3, 3, 2, 1, 3 ], "contribution_list": [ 2, 1, 1, 1, 1 ], "confidence_list": [ 4, 5, 5, 5, 5 ] }, { "paper_id": "MNGMpHxi1I", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Kajetan Schweighofer", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "", "position": "PhD student" }, { "name": "Lukas Aichberger", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Mykyta Ielanskyi", "gender": "Not Specified", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "PhD student" }, { "name": "Sepp Hochreiter", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "Full Professor" } ], "n_formula": 58, "n_formula_1": 9, "n_ref_uni": 53, "n_ref": 131, "n_ref_all": 226, "n_fig": 46, "n_tab": 1, "L_tab": 280, "n_element_tab": 5, "n_fig_1": 19, "n_tab_1": 2, "L_tab_1": 448, "n_element_tab_1": 1, "formula_len_all": 7076, "formula_len_all_1": 1169, "len_all": 217314, "len_all_1": 66297, "len_abs": 2239, "len_title": 107, "len_sents": 67058, "len_sents_1": 31904, "n_sents": 592, "n_sents_1": 242, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1190, "title": "On Information-Theoretic Measures of Predictive Uncertainty", "abs": "Reliable estimation of predictive uncertainty is crucial for machine learning applications, particularly in high-stakes scenarios where hedging against risks is essential. Despite its significance, a consensus on the correct measurement of predictive uncertainty remains elusive. In this work, we return to first principles to develop a fundamental framework of information-theoretic predictive uncertainty measures. Our proposed framework categorizes predictive uncertainty measures according to two factors: (I) The predicting model (II) The approximation of the true predictive distribution. Examining all possible combinations of these two factors, we derive a set of predictive uncertainty measures that includes both known and newly introduced ones. We empirically evaluate these measures in typical uncertainty estimation settings, such as misclassification detection, selective prediction, and out-of-distribution detection. The results show that no single measure is universal, but the effectiveness depends on the specific setting. Thus, our work provides clarity about the suitability of predictive uncertainty measures by clarifying their implicit assumptions and relationships.", "keywords": [ "Uncertainty quantification", "Predictive Uncertainty", "Information-Theory", "Bayesian methods" ], "rating_list": [ 3, 1, 3, 3, 5 ], "soundness_list": [ 3, 1, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 1, 1, 2, 2 ], "confidence_list": [ 3, 5, 4, 4, 3 ] }, { "paper_id": "MMwaQEVsAg", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Alexander M Rush", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Associate Professor" }, { "name": "Celine Lee", "gender": "Female", "institution": "Cornell University", "country": "", "position": "PhD student" }, { "name": "Claire Cardie", "gender": "Female", "institution": "Cornell University", "country": "", "position": "Full Professor" }, { "name": "Matthias Gall\u00e9", "gender": "Male", "institution": "Cohere", "country": "", "position": "Machine Learning Manager" }, { "name": "Nan Jiang", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "Undergrad student" }, { "name": "Wenting Zhao", "gender": "unknown", "institution": "Cornell University", "country": "", "position": "PhD student" }, { "name": "Justin Chiu", "gender": "Male", "institution": "Cornell University", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 15, "n_ref": 27, "n_ref_all": 48, "n_fig": 5, "n_tab": 5, "L_tab": 2739, "n_element_tab": 237, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 867, "n_element_tab_1": 142, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 110300, "len_all_1": 56849, "len_abs": 1106, "len_title": 88, "len_sents": 31104, "len_sents_1": 25352, "n_sents": 298, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1109, "title": "Commit0: Library Generation from Scratch", "abs": "With the goal of supporting research into AI that exceeds typical expert software development ability, we introduce Commit0, a benchmark that challenges AI agents to write libraries from scratch. Agents are provided with a specification document outlining the library's API as well as a suite of interactive unit tests, with the goal of producing an implementation of this API accordingly. The implementation is validated through running these unit tests. As a benchmark, Commit0 is designed to move beyond static one-shot code generation towards agents that must process long-form natural language specifications, adapt to multi-stage feedback, and generate code with complex dependencies. Commit0 also offers an interactive environment where models receive execution and linting feedback on the code they generate. Our experiments demonstrate that while current agents can pass some unit tests, none can yet fully reproduce full libraries. Results also show that interactive feedback is quite useful for models to generate code that passes more unit tests, validating the benchmarks that facilitate its use.", "keywords": [ "code generation", "language model", "evaluation", "feedback" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "MMHqnUOnl0", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Artem Moskalev", "gender": "unknown", "institution": "Johnson and Johnson", "country": "NL", "position": "Researcher" }, { "name": "Mangal Prakash", "gender": "Male", "institution": "Johnson & Johnson", "country": "", "position": "Researcher" }, { "name": "Mehdi Yazdani-Jahromi", "gender": "Male", "institution": "University of Central Florida", "country": "", "position": "PhD student" }, { "name": "Rui Liao", "gender": "unknown", "institution": "Johnson and Johnson", "country": "", "position": "Principal Researcher" }, { "name": "Tommaso Mansi", "gender": "Male", "institution": "Johnson and Johnson", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 73, "n_ref": 122, "n_ref_all": 152, "n_fig": 7, "n_tab": 10, "L_tab": 2995, "n_element_tab": 306, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1742, "n_element_tab_1": 124, "formula_len_all": 166, "formula_len_all_1": 166, "len_all": 213504, "len_all_1": 65503, "len_abs": 1129, "len_title": 103, "len_sents": 64137, "len_sents_1": 32052, "n_sents": 427, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1131, "title": "HELM: Hierarchical Encoding for mRNA Language Modeling", "abs": "Messenger RNA (mRNA) plays a crucial role in protein synthesis, with its codon structure directly impacting biological properties. While Language Models (LMs) have shown promise in analyzing biological sequences, existing approaches fail to account for the hierarchical nature of mRNA's codon structure. We introduce Hierarchical Encoding for mRNA Language Modeling (HELM), a novel pre-training strategy that incorporates codon-level hierarchical structure into language model training. HELM modulates the loss function based on codon synonymity, aligning the model's learning process with the biological reality of mRNA sequences. We evaluate HELM on diverse mRNA datasets and tasks, demonstrating that HELM outperforms standard language model pre-training as well as existing foundation model baselines on six diverse downstream property prediction tasks and an antibody region annotation tasks on average by around 8%. Additionally, HELM enhances the generative capabilities of language model, producing diverse mRNA sequences that better align with the underlying true data distribution compared to non-hierarchical baselines.", "keywords": [ "Messenger RNA (mRNA)", "Codon structure", "Hierarchical modeling", "Bio-language model", "Property prediction", "mRNA sequence generation" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 4 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "MM197t8WlM", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Chen Xu", "gender": "Male", "institution": "Toyota Research Institute", "country": "US", "position": "Researcher" }, { "name": "Xiuyuan Cheng", "gender": "unknown", "institution": "Duke University", "country": "US", "position": "Associate Professor" }, { "name": "Yao Xie", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 46, "n_formula_1": 17, "n_ref_uni": 47, "n_ref": 108, "n_ref_all": 135, "n_fig": 6, "n_tab": 3, "L_tab": 657, "n_element_tab": 92, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2433, "n_element_tab_1": 113, "formula_len_all": 3935, "formula_len_all_1": 1535, "len_all": 180361, "len_all_1": 68911, "len_abs": 1151, "len_title": 85, "len_sents": 48526, "len_sents_1": 29183, "n_sents": 431, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 1192, "title": "Local Flow Matching Generative Models", "abs": "Flow Matching (FM) is a simulation-free method for learning a continuous and invertible flow to interpolate between two distributions, and in particular to generate data from noise in generative modeling. In this paper, we introduce Local Flow Matching ($\\texttt{LFM}$), which consecutively learns a sequence of FM sub-models and each matches a diffusion process up to the time of the step size in the data-to-noise direction. In each step, the two distributions to be interpolated by the sub-model are closer to each other than data vs. noise, and this enables the use of smaller models with faster training. The stepwise structure of $\\texttt{LFM}$ is natural to be distilled and different distillation techniques can be adopted to speed up generation. Theoretically, we prove a generation guarantee of the proposed flow model in terms of the $\\chi^2$-divergence between the generated and true data distributions. In experiments, we demonstrate the improved training efficiency and competitive generative performance of $\\texttt{LFM}$ compared to FM on the unconditional generation of tabular data and image datasets, and also on the conditional generation of robotic manipulation policies.", "keywords": [ "flow-based generative modeling", "stepwise training", "model distillation" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "MLhquJb1qN", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Jan Ebert", "gender": "unknown", "institution": "Forschungszentrum J\u00fclich GmbH", "country": "DE", "position": "Researcher" }, { "name": "Jiangtao Wang", "gender": "unknown", "institution": "Forschungszentrum Juelich GmbH", "country": "DE", "position": "Researcher" }, { "name": "Oleg Filatov", "gender": "unknown", "institution": "Forschungszentrum Juelich GmbH", "country": "DE", "position": "Researcher" }, { "name": "Stefan Kesselheim", "gender": "Male", "institution": "Forschungszentrum J\u00fclich", "country": "DE", "position": "Principal Researcher" } ], "n_formula": 27, "n_formula_1": 5, "n_ref_uni": 37, "n_ref": 84, "n_ref_all": 125, "n_fig": 12, "n_tab": 5, "L_tab": 5059, "n_element_tab": 1230, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1762, "formula_len_all_1": 239, "len_all": 301507, "len_all_1": 68696, "len_abs": 1285, "len_title": 129, "len_sents": 53327, "len_sents_1": 31297, "n_sents": 339, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1375, "title": "Time Transfer: On Optimal Learning Rate and Batch Size In The Infinite Data Limit", "abs": "One of the main challenges in optimal scaling of large language models (LLMs) is the prohibitive cost of hyperparameter tuning, particularly learning rate $\\eta$ and batch size $B$. While techniques like $\\mu$P (Yang et al., 2022) provide scaling rules for optimal $\\eta$ transfer in the infinite model size limit, the optimal scaling behavior in the infinite data size limit remains unknown. We fill in this gap by observing for the first time an intricate dependence of optimal $\\eta$ scaling on the pretraining token budget $T$, $B$ and its relation to the critical batch size $B_\\mathrm{crit}$, which we measure to evolve as $B_\\mathrm{crit} \\propto T$. Furthermore, we show that the optimal batch size is positively correlated with $B_\\mathrm{crit}$: keeping it fixed becomes suboptimal over time even if learning rate is scaled optimally. Surprisingly, our results demonstrate that the observed optimal $\\eta$ and $B$ dynamics are preserved with $\\mu$P model scaling, challenging the conventional view of $B_\\mathrm{crit}$ dependence solely on loss value. Complementing optimality, we examine the sensitivity of loss to changes in learning rate, where we find the sensitivity to decrease with increase of $T$ and to remain constant with $\\mu$P model scaling. We hope our results make the first step towards a unified picture of the joint optimal data and model scaling.", "keywords": [ "optimal scaling", "muP", "hyperparameter transfer", "learning rate sensitivity", "LLM" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "ML8FH4s5Ts", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Chenxu Zhang", "gender": "Male", "institution": "Bytedance", "country": "US", "position": "Researcher" }, { "name": "Guoxian Song", "gender": "Male", "institution": "Bytedance Inc", "country": "US", "position": "Researcher" }, { "name": "Hongyi Xu", "gender": "Male", "institution": "Bytedance", "country": "US", "position": "Researcher" }, { "name": "Jinli Suo", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Linjie Luo", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist, Manager" }, { "name": "XIAOCHEN ZHAO", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiu Li", "gender": "Male", "institution": "Bytedance", "country": "SG", "position": "Researcher" }, { "name": "Yebin Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "You Xie", "gender": "Male", "institution": "Bytedance", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 2, "n_ref_uni": 31, "n_ref": 79, "n_ref_all": 101, "n_fig": 10, "n_tab": 4, "L_tab": 1369, "n_element_tab": 59, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 827, "n_element_tab_1": 48, "formula_len_all": 646, "formula_len_all_1": 321, "len_all": 157585, "len_all_1": 62054, "len_abs": 0, "len_title": 127, "len_sents": 45049, "len_sents_1": 33132, "n_sents": 265, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1492, "title": "X-NeMo: Expressive Neural Motion Reenactment via Disentangled Latent Attention", "abs": "We propose X-NeMo, a novel zero-shot diffusion-based portrait animation pipeline that animates a static portrait using facial movements from a driving video of a different individual. Our work first identifies the root causes of the limitations in prior approaches, such as identity leakage and difficulty in capturing subtle and extreme expressions. To address these challenges, we introduce a fully end-to-end training framework that distills a 1D identity-agnostic latent motion descriptor from driving image, effectively controlling motion through cross-attention during image generation. Our implicit motion descriptor captures expressive facial motion in fine detail, learned end-to-end from a diverse video dataset without reliance on any pre-trained motion detectors. We further disentangle motion latents from identity cues with enhanced expressiveness by supervising their learning with a dual GAN decoder, alongside spatial and color augmentations. By embedding the driving motion into a 1D latent vector and controlling motion via cross-attention instead of additive spatial guidance, our design effectively eliminates the transmission of spatial-aligned structural clues from the driving condition to the diffusion backbone, substantially mitigating identity leakage. Extensive experiments demonstrate that X-NeMo surpasses state-of-the-art baselines, producing highly expressive animations with superior identity resemblance. Our code and models will be available for research.", "keywords": [ "Portrait Animation", "Head Avatar", "Conditional Video Generation" ], "rating_list": [ 6, 3, 8, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 2, 5, 4 ] }, { "paper_id": "MKvQH1ekeY", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Bai LiChen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Haoyi Xiong", "gender": "Male", "institution": "Microsoft Corporation", "country": "CN", "position": "Principal Applied Scientist" }, { "name": "Shitong Shao", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Zeke Xie", "gender": "Male", "institution": "HKUST(GZ)", "country": "CN", "position": "Assistant Professor" }, { "name": "Zipeng Qi", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "zhiqiang xu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "CN", "position": "Assistant Professor" }, { "name": "zikai zhou", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 14, "n_ref_uni": 33, "n_ref": 71, "n_ref_all": 126, "n_fig": 21, "n_tab": 21, "L_tab": 5875, "n_element_tab": 649, "n_fig_1": 9, "n_tab_1": 6, "L_tab_1": 2759, "n_element_tab_1": 348, "formula_len_all": 1984, "formula_len_all_1": 1022, "len_all": 181329, "len_all_1": 56658, "len_abs": 1613, "len_title": 0, "len_sents": 50477, "len_sents_1": 21562, "n_sents": 427, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1883, "title": "Zigzag Diffusion Sampling: Diffusion Models Can Self-Improve via Self-Reflection", "abs": "Diffusion models, the most popular generative paradigm so far, can inject conditional information into the generation path to guide the latent towards desired directions. However, existing text-to-image diffusion models often fail to maintain high image quality and high prompt-image alignment for those challenging prompts. To mitigate this issue and enhance existing pretrained diffusion models, we mainly made three contributions in this paper. First, we propose **diffusion self-reflection** that alternately performs denoising and inversion and demonstrate that such diffusion self-reflection can leverage the guidance gap between denoising and inversion to capture prompt-related semantic information with theoretical and empirical evidence. Second, motivated by theoretical analysis, we derive Zigzag Diffusion Sampling (Z-Sampling), a novel self-reflection-based diffusion sampling method that leverages the guidance gap between denosing and inversion to accumulate semantic information step by step along the sampling path, leading to improved sampling results. Moreover, as a plug-and-play method, Z-Sampling can be generally applied to various diffusion models (e.g., accelerated ones and Transformer-based ones) with very limited coding and computational costs. Third, our extensive experiments demonstrate that Z-Sampling can generally and significantly enhance generation quality across various benchmark datasets, diffusion models, and performance evaluation metrics. For example, DreamShaper with Z-Sampling can self-improve with the HPSv2 winning rate up to **94%** over the original results. Moreover, Z-Sampling can further enhance existing diffusion models combined with other orthogonal methods, including Diffusion-DPO. The code is publicly available at\n[github.com/xie-lab-ml/Zigzag-Diffusion-Sampling](https://github.com/xie-lab-ml/Zigzag-Diffusion-Sampling).", "keywords": [ "Diffusion model", "Semantic Information", "Classifier Guidance Gap" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "MKpiaqh7R0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Baijiong Lin", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "James Kwok", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Hong Kong University of Science and Technology", "country": "HK", "position": "Professor" }, { "name": "Shuhao Chen", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Weisen Jiang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 79, "n_ref": 166, "n_ref_all": 188, "n_fig": 7, "n_tab": 10, "L_tab": 6061, "n_element_tab": 965, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3124, "n_element_tab_1": 469, "formula_len_all": 366, "formula_len_all_1": 310, "len_all": 180879, "len_all_1": 64521, "len_abs": 1089, "len_title": 84, "len_sents": 30519, "len_sents_1": 25239, "n_sents": 225, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 1112, "title": "Input Compensation for Pruned Models", "abs": "Though foundation models are powerful, they are large and require substantial memory and computation resources for serving.\nTo tackle this issue, many pruning methods have been proposed to reduce the model size, thereby achieving memory and computational efficiency. These methods either identify and retrain the important weights or \\textit{adjust the unpruned weights} to compensate for the removed weights. In this paper, we propose a novel approach called input compensation (IC) to boost the performance of pruned models, i.e., \\textit{adjust the input} to compensate for the removed weights. We learn a compensation pool to construct input-dependent compensation to reduce the error caused by pruning. Different from existing pruning methods, which are designed in the parameter space, the proposed IC is designed in the input space. Hence, IC is complementary to existing methods and can be integrated with them.\nExtensive experiments on various tasks, including image classification, language modeling, and image generation, demonstrate that IC is effective in improving the performance of pruned models.", "keywords": [ "model pruning", "model compression" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "MKP1g8wU0P", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Francesco Brarda", "gender": "unknown", "institution": "Emory University", "country": "", "position": "PhD student" }, { "name": "Ruipeng Li", "gender": "unknown", "institution": "Lawrence Livermore National Labs", "country": "", "position": "Researcher" }, { "name": "Shuhao Cao", "gender": "Male", "institution": "University of Missouri - Kansas City", "country": "US", "position": "Assistant Professor" }, { "name": "Yuanzhe Xi", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Associate Professor" } ], "n_formula": 71, "n_formula_1": 21, "n_ref_uni": 67, "n_ref": 130, "n_ref_all": 162, "n_fig": 10, "n_tab": 9, "L_tab": 1325, "n_element_tab": 82, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6778, "formula_len_all_1": 1619, "len_all": 304380, "len_all_1": 77804, "len_abs": 1749, "len_title": 148, "len_sents": 79642, "len_sents_1": 34397, "n_sents": 617, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1760, "title": "Spectral-Refiner: Accurate Fine-Tuning of Spatiotemporal Fourier Neural Operator for Turbulent Flows", "abs": "Recent advancements in operator-type neural networks have shown promising results in approximating the solutions of spatiotemporal Partial Differential Equations (PDEs). However, these neural networks often entail considerable training expenses, and may not always achieve the desired accuracy required in many scientific and engineering disciplines. \nIn this paper, we propose a new learning framework to address these issues. A new spatiotemporal adaptation is proposed to generalize any Fourier Neural Operator (FNO) variant to learn maps between Bochner spaces, which can perform an arbitrary-lengthed temporal super-resolution for the first time. \nTo better exploit this capacity, a new paradigm is proposed to refine the commonly adopted end-to-end neural operator training and evaluations with the help from the wisdom from traditional numerical PDE theory and techniques. \nSpecifically, in the learning problems for the turbulent flow modeling by the Navier-Stokes Equations (NSE), the proposed paradigm trains an FNO only for a few epochs. Then, only the newly proposed spatiotemporal spectral convolution layer is fine-tuned without the frequency truncation. The fine-tuning loss function uses a negative Sobolev norm for the first time in operator learning, defined through a reliable functional-type a posteriori error estimator whose evaluation is exact thanks to the Parseval identity. Moreover, unlike the difficult nonconvex optimization problems in the end-to-end training, this fine-tuning loss is convex. \nNumerical experiments on commonly used NSE benchmarks demonstrate significant improvements in both computational efficiency and accuracy, compared to end-to-end evaluation and traditional numerical PDE solvers under certain conditions.", "keywords": [ "operator learning", "neural operators", "Navier-Stokes", "PDE", "partial differential equations", "computational fluid dynamics" ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "MKEHCx25xp", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Abhilasha Ravichander", "gender": "unknown", "institution": "University of Washington", "country": "US", "position": "Postdoc" }, { "name": "Bill Y Lin", "gender": "Male", "institution": "xAI", "country": "US", "position": "Researcher" }, { "name": "Khyathi Chandu", "gender": "Female", "institution": "Mistral AI", "country": "US", "position": "Researcher" }, { "name": "Nouha Dziri", "gender": "Female", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Ronan Le Bras", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Valentina Pyatkin", "gender": "unknown", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Postdoc" }, { "name": "Yejin Choi", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Yuntian Deng", "gender": "unknown", "institution": "University of Waterloo", "country": "", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 58, "n_ref_all": 88, "n_fig": 8, "n_tab": 4, "L_tab": 3493, "n_element_tab": 443, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2379, "n_element_tab_1": 387, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 158251, "len_all_1": 73438, "len_abs": 2368, "len_title": 127, "len_sents": 44453, "len_sents_1": 31040, "n_sents": 359, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 79, "L_abs": 1704, "title": "WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild", "abs": "We introduce WildBench, an automated evaluation framework designed to benchmark large language models (LLMs) using challenging, real-world user queries. WildBench consists of 1,024 tasks carefully selected from over one million human-chatbot conversation logs. For automated evaluation with WildBench, we have developed two metrics, WB-Reward and WB-Score, which are computable using advanced LLMs such as GPT-4-turbo. WildBench evaluation uses task-specific checklists to evaluate model outputs systematically and provides structured explanations that justify the scores and comparisons, resulting in more reliable and interpretable automatic judgments. WB-Reward employs fine-grained pairwise comparisons between model responses, generating five potential outcomes: much better, slightly better, slightly worse, much worse, or a tie. Unlike previous evaluations that employed a single baseline model, we selected three baseline models at varying performance levels to ensure a comprehensive pairwise evaluation. Additionally, we propose a simple method to mitigate length bias, by converting outcomes of \u201cslightly better/worse\u201d to \u201ctie\u201d if the winner response exceeds the loser one by more than K characters. WB-Score evaluates the quality of model outputs individually, making it a fast and cost-efficient evaluation metric. WildBench results demonstrate a strong correlation with the human-voted Elo ratings from Chatbot Arena on hard tasks. Specifically, WB-Reward achieves a Pearson correlation of 0.98 with top-ranking models. Additionally, WB-Score reaches 0.95, surpassing both ArenaHard\u2019s 0.91 and AlpacaEval2.0\u2019s 0.89 for length-controlled win rates, as well as the 0.87 for regular win rates.", "keywords": [ "LLM", "Evaluation", "Benchmarking" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "MK6E6IgROl", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Hiroki IKOMA", "gender": "Male", "institution": "Yazaki corporation", "country": "JP", "position": "engineer" }, { "name": "Hiroki Seto", "gender": "unknown", "institution": "Yazaki Corporation", "country": "JP", "position": "independent researcher" }, { "name": "Ippei Fujisawa", "gender": "unknown", "institution": "Araya", "country": "", "position": "Researcher" }, { "name": "Pei-Chun Chien", "gender": "Male", "institution": "Yazaki Corporation", "country": "JP", "position": "Independent Researcher" }, { "name": "Rina Onda", "gender": "unknown", "institution": "Yazaki Corporation", "country": "JP", "position": "Independent Researcher" }, { "name": "Ryota Kanai", "gender": "Male", "institution": "Araya, Inc.", "country": "JP", "position": "Principal Researcher" }, { "name": "Sensho Nobe", "gender": "Male", "institution": "Araya Inc.", "country": "JP", "position": "Researcher" }, { "name": "Yoshiaki Uchida", "gender": "unknown", "institution": "Yazaki Corporation", "country": "JP", "position": "Engineer" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 28, "n_ref": 43, "n_ref_all": 64, "n_fig": 14, "n_tab": 3, "L_tab": 1825, "n_element_tab": 91, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 1981, "n_element_tab_1": 95, "formula_len_all": 171, "formula_len_all_1": 136, "len_all": 137390, "len_all_1": 63115, "len_abs": 1700, "len_title": 117, "len_sents": 36441, "len_sents_1": 29644, "n_sents": 277, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1710, "title": "ProcBench: Benchmark for Multi-Step Reasoning and Following Procedure", "abs": "Reasoning is central to a wide range of intellectual activities, and while the capabilities of large language models (LLMs) continue to advance, their performance in reasoning tasks remains limited. The processes and mechanisms underlying reasoning are not yet fully understood, but key elements include path exploration, selection of relevant knowledge, and multi-step inference. Problems are solved through the synthesis of these components. In this paper, we propose a benchmark that focuses on a specific aspect of reasoning ability: the direct evaluation of multi-step inference. To this end, we design an extreme reasoning task where multi-step inference is specifically focused by largely eliminating path exploration and implicit knowledge utilization. Our dataset comprises pairs of explicit instructions and corresponding questions, where the procedures necessary for solving the questions are entirely detailed within the instructions. This setup allows models to solve problems solely by following the provided directives. By constructing problems that require varying numbers of steps to solve and evaluating responses at each step, we enable a thorough assessment of state-of-the-art LLMs' ability to follow instructions. To ensure the robustness of our evaluation, we include multiple distinct tasks. Furthermore, by comparing accuracy across tasks, utilizing step-aware metrics, and applying separately defined measures of complexity, we conduct experiments that offer insights into the capabilities and limitations of LLMs in reasoning tasks. Our findings have significant implications for the development of LLMs and highlight areas for future research in advancing their reasoning abilities.", "keywords": [ "benchmark", "dataset", "multi-step reasoning", "LLM", "large language model", "instruction following" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 1, 1, 3 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "MJyqwBVgMs", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Jiangtao Xie", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Junjie Wu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "MS student" }, { "name": "Pengfei Zhu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" }, { "name": "Qilong Wang", "gender": "unknown", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" }, { "name": "Qinghua Hu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Professor" } ], "n_formula": 20, "n_formula_1": 19, "n_ref_uni": 67, "n_ref": 185, "n_ref_all": 225, "n_fig": 4, "n_tab": 8, "L_tab": 4364, "n_element_tab": 425, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 179, "n_element_tab_1": 13, "formula_len_all": 2009, "formula_len_all_1": 2036, "len_all": 183142, "len_all_1": 66194, "len_abs": 1870, "len_title": 111, "len_sents": 39407, "len_sents_1": 27243, "n_sents": 279, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1740, "title": "Asymmetric Factorized Bilinear Operation for Vision Transformer", "abs": "As a core component of Transformer-like deep architectures, a feed-forward network (FFN) for channel mixing is responsible for learning features of each token. Recent works show channel mixing can be enhanced by increasing computational burden or can be slimmed at the sacrifice of performance. Although some efforts have been made, existing works are still struggling to solve the paradox of performance and complexity trade-offs. In this paper, we propose an Asymmetric Factorized Bilinear Operation (AFBO) to replace FFN of vision transformer (ViT), which attempts to efficiently explore rich statistics of token features for achieving better performance and complexity trade-off. Specifically, our AFBO computes second-order statistics via a spatial-channel factorized bilinear operation for feature learning, which replaces a simple linear projection in FFN and enhances the feature learning ability of ViT by modeling second-order correlation among token features. Furthermore, our AFBO presents two structured-sparsity channel mapping strategies, namely Grouped Cross Channel Mapping (GCCM) and Overlapped Cycle Channel Mapping (OCCM). They decompose bilinear operation into grouped channel features by considering information interaction between groups, significantly reducing computational complexity while guaranteeing model performance. Finally, our AFBO is built with GCCM and OCCM in an asymmetric way, aiming to achieve a better trade-off. Note that our AFBO is model-agnostic, which can be flexibly integrated with existing ViTs. Experiments are conducted with twenty ViTs on various tasks, and the results show our AFBO is superior to its counterparts while improving existing ViTs in terms of generalization and robustness.", "keywords": [ "Vision Transformer", "Channel Mixer", "Factorized Bilinear Operation" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "MJip0YjQkp", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Ahmed Sharshar", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Bin Gu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Velibor Bojkovic", "gender": "unknown", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Yasser Ashraf", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 42, "n_ref": 68, "n_ref_all": 83, "n_fig": 3, "n_tab": 9, "L_tab": 7248, "n_element_tab": 547, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1228, "n_element_tab_1": 140, "formula_len_all": 428, "formula_len_all_1": 369, "len_all": 173951, "len_all_1": 49230, "len_abs": 1055, "len_title": 153, "len_sents": 35451, "len_sents_1": 22419, "n_sents": 239, "n_sents_1": 152, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1057, "title": "SPACT18: Spiking Human Action Recognition Benchmark Dataset with Complementary RGB and Thermal Modalities", "abs": "Spike cameras, bio-inspired vision sensors, asynchronously fire spikes by accumulating light intensities at each pixel, offering ultra-high energy efficiency and exceptional temporal resolution. Unlike event cameras, which record changes in light intensity to capture motion, spike cameras provide even finer spatiotemporal resolution and a more precise representation of continuous changes. In this paper, we introduce the first video action recognition (VAR) dataset using spike camera, alongside synchronized RGB and thermal modalities, to enable comprehensive benchmarking for Spiking Neural Networks (SNNs). By preserving the inherent sparsity and temporal precision of spiking data, our three datasets offer a unique platform for exploring multimodal video understanding and serve as a valuable resource for directly comparing spiking, thermal, and RGB modalities. This work contributes a novel dataset that will drive research in energy-efficient, ultra-low-power video understanding, specifically for action recognition tasks using spike-based data.", "keywords": [ "Spiking Camera", "Spiking Neural Network", "Video Understanding", "Video Action Recognition" ], "rating_list": [ 3, 6, 6, 5 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 3, 4 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "MJWJoICJQh", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Seokjun Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Seung-Hyun Kong", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "Associate Professor" } ], "n_formula": 18, "n_formula_1": 11, "n_ref_uni": 37, "n_ref": 72, "n_ref_all": 98, "n_fig": 15, "n_tab": 4, "L_tab": 319, "n_element_tab": 36, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 319, "n_element_tab_1": 36, "formula_len_all": 1191, "formula_len_all_1": 719, "len_all": 136720, "len_all_1": 57622, "len_abs": 2403, "len_title": 149, "len_sents": 48345, "len_sents_1": 28597, "n_sents": 330, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 2412, "title": "Safety-Advanced Autonomous Driving for Urgent Hazardous Situations using Q-Compared Soft Actor-Critic", "abs": "Autonomous vehicles must be capable of safe driving under all conditions to ensure passenger safety. This includes urgent hazardous situations (UHS), such as skidding on slippery roads or tire grip saturation during high-speed driving, which are not only difficult even for expert human drivers but also challenging to develop autonomous driving technologies that surpass human capabilities. Even though the recent advancements in machine learning including imitation learning (IL), reinforcement learning (RL), and hybrid learning (HL) have enabled the safe navigation of autonomous vehicles in various complex scenarios, they have fundamental limitations in UHS. Driving policies trained via IL degrade in novel situations where expert demonstration data is scarce or of poor quality, and RL struggles to develop optimal driving policies in UHS, which have broad state and action spaces and high transition variance. HL techniques combining IL and RL also fall short, as they require nearly optimal demonstration data, which is nearly impossible to obtain in UHS due to the difficulty for human drivers to react appropriately.\nTo address these limitations, we propose a novel HL technique, Q-Compared Soft Actor-Critic (QC-SAC), which effectively utilizes immature demonstration data to develop optimal driving policies and adapt quickly to novel situations in UHS. QC-SAC evaluates the quality of demonstration data based on action value Q to prioritize beneficial data and disregard detrimental ones. Furthermore, QC-SAC improves the performance of the Q-network by leveraging demonstration data and enhances learning by rapidly incorporating new successful experiences from ongoing interactions, enabling fast adaptation to new situations. We test QC-SAC for two extreme UHS scenarios: oversteer control with collision avoidance (OCCA) and time-trial race (TTR). In OCCA, QC-SAC achieves a success rate 2.36 times higher than existing techniques, and in TTR, it reduces lap time by more than 13.6\\% while completing 300 test runs without a single failure. By proposing an innovative HL technique capable of training superior driving policies with immature demonstration data, we provide a solution for autonomous driving technologies that can handle UHS and introduce the world-first safe-advanced autonomous driving technology capable of controlling a vehicle oversteer safely and avoiding obstacles ahead.", "keywords": [ "Autonomous Driving", "Reinforcement Learning", "Imitation Learning", "Hybrid Learning" ], "rating_list": [ 3, 3, 5, 3, 3 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 2, 2 ], "contribution_list": [ 3, 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "MJNywBdSDy", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Andrew Campbell", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Juno Nam", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "KR", "position": "PhD student" }, { "name": "Rafael Gomez-Bombarelli", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Sulin Liu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Tommi Jaakkola", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Yilun Xu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Hannes St\u00e4rk", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 78, "n_formula_1": 25, "n_ref_uni": 0, "n_ref": 116, "n_ref_all": 183, "n_fig": 11, "n_tab": 19, "L_tab": 5520, "n_element_tab": 371, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 181, "n_element_tab_1": 21, "formula_len_all": 11428, "formula_len_all_1": 3408, "len_all": 210827, "len_all_1": 74370, "len_abs": 1054, "len_title": 115, "len_sents": 86062, "len_sents_1": 32035, "n_sents": 726, "n_sents_1": 283, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1065, "title": "Think while You Generate: Discrete Diffusion with Planned Denoising", "abs": "Discrete diffusion has achieved state-of-the-art performance, outperforming or approaching autoregressive models on standard benchmarks. In this work, we introduce *Discrete Diffusion with Planned Denoising* (DDPD), a novel framework that separates the generation process into two models: a planner and a denoiser. At inference time, the planner selects which positions to denoise next by identifying the most corrupted positions in need of denoising, including both initially corrupted and those requiring additional refinement. This plan-and-denoise approach enables more efficient reconstruction during generation by iteratively identifying and denoising corruptions in the optimal order. DDPD outperforms traditional denoiser-only mask diffusion methods, achieving superior results on language modeling benchmarks such as *text8*, *OpenWebText*, and token-based generation on *ImageNet 256 \u00d7 256*. Notably, in language modeling, DDPD significantly reduces the performance gap between diffusion-based and autoregressive methods in terms of generative perplexity.", "keywords": [ "discrete diffusion", "generative models" ], "rating_list": [ 8, 3, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "MJ8ALv35sj", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Geyong Min", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "Full Professor" }, { "name": "Zheng Wang", "gender": "Male", "institution": "The University of Warwick", "country": "GB", "position": "Postdoc" }, { "name": "Wenjie Ruan", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "Senior Lecturer" } ], "n_formula": 126, "n_formula_1": 37, "n_ref_uni": 33, "n_ref": 57, "n_ref_all": 65, "n_fig": 2, "n_tab": 2, "L_tab": 824, "n_element_tab": 45, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 323, "n_element_tab_1": 37, "formula_len_all": 10800, "formula_len_all_1": 2055, "len_all": 127946, "len_all_1": 51891, "len_abs": 892, "len_title": 157, "len_sents": 33322, "len_sents_1": 21370, "n_sents": 364, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 914, "title": "How Does Cross-Layer Correlation in Deep Neural Networks Influence Generalization and Adversarial Robustness?", "abs": "\\textit{Generalization} and \\textit{adversarial robustness} are two critical concepts in machine learning. Understanding the key factors that affect the trade-off between these concepts is essential for guiding architectural design and developing training strategies, such as adversarial training, especially for deep neural networks. In this paper, we investigate the impact of cross-layer correlations in weight matrices on both generalization and adversarial robustness. We provide a theoretical analysis demonstrating that increasing cross-layer correlations leads to a monotonic increase in the generalization gap. Furthermore, we establish a connection between adversarial risk and natural risk. Leveraging this connection, we show that in linear models, higher cross-layer correlations also degrade adversarial robustness. Finally, we validate our theoretical findings through experiments conducted on MLPs.", "keywords": [ "Deep Neural Network", "Generalization", "Adversarial Robustness" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 2, 3 ] }, { "paper_id": "MI0UiWeqOl", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jathushan Rajasegaran", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Jitendra Malik", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Director of Research" }, { "name": "Neerja Thakkar", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Shiry Ginosar", "gender": "Female", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Assistant Professor" }, { "name": "Tara Sadjadpour", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 53, "n_ref": 70, "n_ref_all": 93, "n_fig": 9, "n_tab": 3, "L_tab": 302, "n_element_tab": 61, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 250, "n_element_tab_1": 41, "formula_len_all": 238, "formula_len_all_1": 150, "len_all": 151340, "len_all_1": 71204, "len_abs": 901, "len_title": 101, "len_sents": 43171, "len_sents_1": 34293, "n_sents": 345, "n_sents_1": 277, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 902, "title": "Poly-Autoregressive Modeling for Interacting Entities", "abs": "We present a simple framework that predicts an agent's future behavior by considering the effects that other interacting agents and entities have on them. We propose to model behavior as a sequence of tokens, each representing the state of an agent at a specific timestep. The core of our approach centers around Poly-Autoregressive models, which predict the future behavior of an agent during interaction by considering the agent's past state history and the state of other agents in the scene. In this paper, we develop the mechanics of Poly-Autoregressive (PAR) modeling and show that this framework applies without any modification to an extensive range of prediction problems that, on the surface, appear as entirely different scenarios, such as human action prediction in social situations, trajectory prediction for autonomous vehicles, and object pose prediction during hand-object interaction.", "keywords": [ "autoregressive prediction; multi-agent interaction" ], "rating_list": [ 1, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "MHmsJS6YHQ", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Darshan Patil", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "PhD student" }, { "name": "Maryam Hashemzadeh", "gender": "Female", "institution": "Montreal Institute for Learning Algorithms, Universit\u00e9 de Montr\u00e9al", "country": "", "position": "Researcher" }, { "name": "Pranshu Malviya", "gender": "Male", "institution": "\u00c9cole Polytechnique de Montr\u00e9al, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "Quentin Fournier", "gender": "Male", "institution": "Mila - Quebec AI Institute", "country": "CA", "position": "Research Fellow" }, { "name": "Sarath Chandar", "gender": "Male", "institution": "\u00c9cole Polytechnique de Montr\u00e9al", "country": "CA", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 30, "n_ref": 77, "n_ref_all": 100, "n_fig": 9, "n_tab": 9, "L_tab": 7399, "n_element_tab": 872, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2762, "n_element_tab_1": 212, "formula_len_all": 434, "formula_len_all_1": 326, "len_all": 132076, "len_all_1": 62832, "len_abs": 976, "len_title": 156, "len_sents": 40446, "len_sents_1": 29539, "n_sents": 281, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 979, "title": "Interpolate: How Resetting Active Neurons can also improve Generalizability in Online Learning", "abs": "While neural networks have shown a significant gain in performance across a wide range of applications, they still struggle in non-stationary settings as they tend to lose their ability to adapt to new tasks \u2014 a phenomenon known as the loss of plasticity. The conventional approach to addressing this problem often involves resetting the most under-utilized or dormant parts of the network, suggesting that recycling such parameters is crucial for maintaining a model's plasticity. In this study, we explore whether this approach is the only way to address plasticity loss. We introduce a resetting approach based on model merging called Interpolate and show that contrary to previous findings, resetting even the most active parameters using our approach can also lead to better generalization. We further show that Interpolate can perform similarly or better compared to traditional resetting methods, offering a new perspective on training dynamics in non-stationary settings.", "keywords": [ "plastiticy", "generalization", "online learning", "permutation invariance", "model merging", "dormancy", "adaptability", "continual learning" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 1, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "MHP4jGMN2E", "primary_area": "applications to robotics, autonomy, planning", "n_author": 3, "author_info_list": [ { "name": "Ameet Talwalkar", "gender": "Male", "institution": "Carnegie-Mellon University", "country": "", "position": "" }, { "name": "Chris Donahue", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Wayne Chi", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 27, "n_ref": 109, "n_ref_all": 127, "n_fig": 1, "n_tab": 8, "L_tab": 1369, "n_element_tab": 135, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 616, "n_element_tab_1": 93, "formula_len_all": 257, "formula_len_all_1": 257, "len_all": 126290, "len_all_1": 66979, "len_abs": 1392, "len_title": 102, "len_sents": 38871, "len_sents_1": 30176, "n_sents": 385, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1398, "title": "The Impact of Element Ordering on LM Agent Performance", "abs": "There has been a surge of interest in language model agents that can navigate virtual environments such as the web or desktop. To navigate such environments, agents benefit from information on the various elements (e.g., buttons, text, or images) present. However, it remains unclear which element attributes have the greatest impact on agent performance, especially in environments that only provide a graphical representation (i.e., pixels). Here we find that the ordering in which elements are presented to the language model is surprisingly impactful\u2014randomizing element ordering in webpages compromises average agent performance to a degree comparable to removing all visible text from webpages. While web agents benefit from the semantic hierarchical ordering of elements available via the browser, agents that parse elements directly from pixels do not have access to any such ordering. Here we endeavor to derive effective orderings and investigate the impact of various element ordering methods in web and desktop environments. We find that dimensionality reduction provides a viable ordering for pixel-only environments. We train a UI element detection model to derive elements from pixels and apply our findings to an agent benchmark\u2014OmniACT\u2014where we only have access to pixels. Our method completes more than two times as many tasks on average relative to the previous state-of-the-art.", "keywords": [ "Agents", "AI Agents", "LLM Agents", "LM Agents" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 5, 5 ] }, { "paper_id": "MH6yUPwVbp", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Athanasios Katsamanis", "gender": "Male", "institution": "Athena Research Center", "country": "", "position": "Principal Researcher" }, { "name": "Martino Ciaperoni", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Postdoc" }, { "name": "Nikolaos Tziavelis", "gender": "unknown", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "Panagiotis Karras", "gender": "Male", "institution": "Copenhagen University", "country": "DK", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 12, "n_ref_uni": 35, "n_ref": 58, "n_ref_all": 83, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3781, "formula_len_all_1": 2132, "len_all": 147910, "len_all_1": 77396, "len_abs": 1764, "len_title": 103, "len_sents": 54428, "len_sents_1": 36441, "n_sents": 428, "n_sents_1": 285, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1735, "title": "Fast and Space-Efficient Fixed-Length Path Optimization", "abs": "Several optimization problems seek a path of predetermined length among network nodes that minimizes a cost function. Conventionally, such problems are tackled by dynamic programming (DP) applying a Bellman-type equation. A prominent example is Viterbi decoding, which returns the path in a Hidden Markov Model that best explains a series of observations, with applications from bioinformatics to communication systems and speech recognition. However, DP-based solutions (i) exhaustively explore a search space linear in both network size and path length in time quadratic in network size, without exploiting data characteristics, and (ii) require memory commensurate with that search space to reconstruct the optimal path. In this paper, we propose Isabella (Dijkstra-Bellman), a novel framework that finds optimal paths of predetermined length in time- and space-efficient fashion by a combination of best-first-search, depth-first-search, and divide-and-conquer strategies. The best-first-search component avoids the exhaustive exploration of the search space using a priority queue; the depth-first-search component keeps the size of that queue in check; and the divide-and-conquer component constructs the optimal path recursively and parsimoniously after determining its cost. We apply Isabella to Viterbi decoding, introducing algorithms that visit the most promising pathways first and control memory consumption. To emphasize the generality of Isabella, we also instantiate it with an algorithm for histogram construction. To our knowledge, no previous work addresses such problems in this manner. Our experimental evaluation shows our solutions to be highly time- and space-efficient compared to standard dynamic programming.", "keywords": [ "path optimization", "fixed length constraint", "space efficiency" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "MGceYYNvXp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Daniel Gruhl", "gender": "Male", "institution": "IBM, International Business Machines", "country": "", "position": "Researcher" }, { "name": "Tze Way Eugene Ie", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Lucas Spangher", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Peter Grabowski", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "Lecturer" }, { "name": "Rama Kumar Pasumarthi", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Tianle Li", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "William F. Arnold", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 20, "n_ref": 30, "n_ref_all": 46, "n_fig": 6, "n_tab": 1, "L_tab": 182, "n_element_tab": 17, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 182, "n_element_tab_1": 17, "formula_len_all": 56, "formula_len_all_1": 56, "len_all": 78836, "len_all_1": 57722, "len_abs": 935, "len_title": 117, "len_sents": 18840, "len_sents_1": 18555, "n_sents": 132, "n_sents_1": 127, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 940, "title": "Project MPG: towards a generalized performance quotient for LLM intelligence", "abs": "There exists an extremely wide array of LLM benchmarking tasks, whereas oftentimes a single number is the most actionable for decision making, especially by non-experts. No such aggregation schema exists that is not Elo based, which could be costly or time consuming. Here we propose a method to aggregate performance across a general space of benchmarks, nicknamed Project \u201cMPG\u201d, here dubbed Model Performance and Goodness, in addition referencing a metric widely understood to be an important yet inaccurate and crude measure of car performance. Here, we create two numbers: an ``Goodness'' number (answer accuracy), and a \u201cFastness\u201d number (cost or QPS). We compare models against each other and present a ranking according to our general metric as well as subdomains. We find significant agreement between the raw pearson correlation of our scores and thosee of LMSys, even improving on the correlation of the MMLU leaderboard to LMSys.", "keywords": [ "evaluation", "LLM", "LMSys", "benchmarks" ], "rating_list": [ 1, 1, 1, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 1, 1, 3 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "MGKDBuyv4p", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Arham Mushtaq Khan", "gender": "Male", "institution": "University of Chicago", "country": "", "position": "PhD student" }, { "name": "Aswathy Ajith", "gender": "Female", "institution": "University of Chicago", "country": "", "position": "PhD student" }, { "name": "Caleb Geniesse", "gender": "Male", "institution": "Diadia Health", "country": "US", "position": "Senior ML Scientist" }, { "name": "Ian Foster", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Full Professor" }, { "name": "Kyle Chard", "gender": "unknown", "institution": "University of Chicago", "country": "", "position": "Researcher" }, { "name": "Mansi Sakarvadia", "gender": "unknown", "institution": ", University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Michael W. Mahoney", "gender": "unknown", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Nathaniel C Hudson", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Postdoc" }, { "name": "Yaoqing Yang", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 68, "n_ref": 182, "n_ref_all": 249, "n_fig": 16, "n_tab": 9, "L_tab": 6263, "n_element_tab": 497, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 3883, "n_element_tab_1": 406, "formula_len_all": 301, "formula_len_all_1": 99, "len_all": 245654, "len_all_1": 60478, "len_abs": 1403, "len_title": 90, "len_sents": 73759, "len_sents_1": 24485, "n_sents": 558, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 42, "L_abs": 1407, "title": "Mitigating Memorization in Language Models", "abs": "Language models (LMs) can \u201cmemorize\u201d information, i.e., encode training data in their weights in such a way that inference-time queries can lead to verbatim regurgitation of that data. This ability to extract training data can be problematic, for example, when data are private or sensitive. In this work, we investigate methods to mitigate memorization: three regularizer-based, three fine-tuning-based, and eleven machine unlearning-based methods, with five of the latter being new methods that we introduce. We also introduce TinyMem, a suite of small, computationally-efficient LMs for the rapid development and evaluation of memorization-mitigation methods. We demonstrate that the mitigation methods that we develop using TinyMem can successfully be applied to production-grade LMs, and we determine via experiment that: regularizer-based mitigation methods are slow and ineffective at curbing memorization; fine-tuning-based methods\nare effective at curbing memorization, but overly expensive, especially for retaining higher accuracies; and unlearning-based methods are faster and more effective, allowing for the precise localization and removal of memorized information from LM weights prior to inference. We show, in particular, that our proposed unlearning method BalancedSubnet outperforms other mitigation methods at removing\nmemorized information while preserving performance on target tasks.", "keywords": [ "language models", "memorization", "machine unlearning", "regularization", "fine-tuning", "natural language processing" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 4, 2, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 2, 4, 4 ] }, { "paper_id": "MGAzLOJOYL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Changhao Pan", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Undergrad student" }, { "name": "Rongjie Huang", "gender": "Male", "institution": "FAIR", "country": "US", "position": "Researcher" }, { "name": "Ruiqi Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Ruiyuan Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Wenxiang Guo", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhiqing Hong", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Zhiyuan Zhu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Ziyue Jiang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 35, "n_formula_1": 5, "n_ref_uni": 53, "n_ref": 129, "n_ref_all": 170, "n_fig": 9, "n_tab": 18, "L_tab": 4994, "n_element_tab": 481, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 5082, "n_element_tab_1": 222, "formula_len_all": 1626, "formula_len_all_1": 309, "len_all": 254466, "len_all_1": 64540, "len_abs": 1515, "len_title": 124, "len_sents": 82308, "len_sents_1": 29027, "n_sents": 608, "n_sents_1": 170, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1461, "title": "MultiBand: Multi-Task Song Generation with Personalized Prompt-Based Control", "abs": "Song generation focuses on producing controllable high-quality songs based on various personalized prompts. However, existing methods struggle to generate high-quality vocals and accompaniments with effective style control and proper alignment. Additionally, they fall short in supporting various personalized tasks based on diverse prompts. To address these challenges, we introduce MultiBand, the first multi-task song generation model for synthesizing high-quality, aligned songs with extensive control based on diverse personalized prompts. \nMultiBand comprises these primary models: 1) VocalBand, a decoupled model, leverages the flow-matching method for singing styles, pitches, and mel-spectrograms generation, allowing fast and high-quality vocal generation with high-level control. 2) AccompBand, a flow-based transformer model, incorporates the Aligned Vocal Encoder, using contrastive learning for alignment, and Band-MOE, selecting suitable experts for enhanced quality and control. This model allows for generating controllable, high-quality accompaniments perfectly aligned with vocals. 3) Two generation models, LyricBand for lyrics and MelodyBand for melodies, contribute to the comprehensive multi-task song generation system, allowing for extensive control based on multiple personalized prompts. Experimental results demonstrate that MultiBand performs better over baseline models across multiple tasks using objective and subjective metrics.", "keywords": [ "multi-task song generation", "prompt-based style control", "style transfer", "singing voice synthesis", "music generation" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 5, 3 ] }, { "paper_id": "MG2Zkf0haD", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Anca Dragan", "gender": "Female", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Chunyuan Zheng", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Kuan Fang", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Assistant Professor" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Vivek Myers", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 18, "n_ref_uni": 64, "n_ref": 101, "n_ref_all": 116, "n_fig": 3, "n_tab": 3, "L_tab": 424, "n_element_tab": 37, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 606, "n_element_tab_1": 38, "formula_len_all": 2562, "formula_len_all_1": 2007, "len_all": 150437, "len_all_1": 56707, "len_abs": 1003, "len_title": 125, "len_sents": 32833, "len_sents_1": 24600, "n_sents": 259, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 746, "title": "Successor Representations Enable Emergent Compositional Instruction Following", "abs": "Behavioral cloning (BC) has seen widespread adoption in scalable robot learning pipelines. These methods struggle to perform compositional generalization, where a new out-of-distribution evaluation task can be viewed as a sequence of simpler in-distribution steps. We augment goal-conditioned BC methods with a temporal alignment loss that learns to associate present and future states. This approach is able to generalize to novel composite tasks specified as goal images or language instructions, without assuming any additional reward supervision or explicit subtask planning. We evaluate our approach across diverse tabletop robotic manipulation tasks, showing substantial improvements for tasks specified with either language or goal images.", "keywords": [ "Robot Learning", "Instruction Following", "Compositional Generalization" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "MFySy0DWAH", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Ao Tong", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Chengjie Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Haoyang He", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jiangning Zhang", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Xiang Bai", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xinwei He", "gender": "Male", "institution": "Huazhong Agricultural University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhenye Gan", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "yuxuan cai", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 40, "n_ref": 85, "n_ref_all": 101, "n_fig": 5, "n_tab": 3, "L_tab": 700, "n_element_tab": 85, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 3056, "n_element_tab_1": 46, "formula_len_all": 499, "formula_len_all_1": 499, "len_all": 126675, "len_all_1": 59614, "len_abs": 1482, "len_title": 116, "len_sents": 33380, "len_sents_1": 27426, "n_sents": 263, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1407, "title": "A Framework of Distilling Multimodal Large Language Models", "abs": "The success of Large Language Models (LLM) has led researchers to explore Multimodal Large Language Models (MLLM) for unified visual and linguistic understanding. However, the increasing model size and computational complexity of MLLM limit their use in resource-constrained environments. Small-scale MLLM ($s$-MLLM) aims to retain the capabilities of the large-scale model ($l$-MLLM) while reducing computational demands, but resulting in a significant decline in performance. To address the aforementioned issues, we propose a novel LLaVA-KD framework to transfer knowledge from $l$-MLLM to $s$-MLLM. Specifically, we introduce Multimodal Distillation (MDist) to minimize the divergence between the visual-textual output distributions of $l$-MLLM and $s$-MLLM, and Relation Distillation (RDist) to transfer $l$-MLLM\u2019s ability to model correlations between visual features. Additionally, we propose a three-stage training scheme to fully exploit the potential of $s$-MLLM: 1) Distilled Pre-Training to align visual-textual representations, 2) Supervised Fine-Tuning to equip the model with multimodal understanding, and 3) Distilled Fine-Tuning to further transfer $l$-MLLM capabilities. Our approach significantly improves performance without altering the small model's architecture. Extensive experiments and ablation studies validate the effectiveness of each proposed component. Code will be available.", "keywords": [ "MLLMs", "Knowledge distillation" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "MFwYXa796v", "primary_area": "reinforcement learning", "n_author": 11, "author_info_list": [ { "name": "Bo Liu", "gender": "Male", "institution": "Auburn University", "country": "", "position": "Assistant Professor" }, { "name": "Bo XU", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Chengjie Wu", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Chongjie Zhang", "gender": "unknown", "institution": "Washington University, Saint Louis", "country": "", "position": "Associate Professor" }, { "name": "Hao Hu", "gender": "Male", "institution": "Moonshot AI", "country": "CN", "position": "Researcher" }, { "name": "Jin Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Yang Gao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yi Fan", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yihuan Mao", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yiqin Yang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuhua Jiang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 54, "n_formula_1": 24, "n_ref_uni": 48, "n_ref": 90, "n_ref_all": 112, "n_fig": 2, "n_tab": 11, "L_tab": 4661, "n_element_tab": 458, "n_fig_1": 1, "n_tab_1": 8, "L_tab_1": 3018, "n_element_tab_1": 140, "formula_len_all": 5488, "formula_len_all_1": 2000, "len_all": 168380, "len_all_1": 61653, "len_abs": 1263, "len_title": 166, "len_sents": 41832, "len_sents_1": 25013, "n_sents": 362, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 117, "L_abs": 1269, "title": "Fewer Questions, Better Answers: Efficient Offline Preference-based Reinforcement Learning via In-Dataset Exploration", "abs": "Preference-based reinforcement learning (PbRL) can help avoid sophisticated reward designs and align better with human intentions, showing great promise in various real-world applications. However, obtaining human feedback for preferences can be expensive and time-consuming, which forms a strong barrier for PbRL. In this work, we address the problem of low query efficiency in offline PbRL, pinpointing two primary reasons: inefficient exploration and overoptimization of learned reward functions. In response to these challenges, we propose a novel algorithm, Offline PbRL via In-Dataset Exploration (OPRIDE), designed to enhance the query efficiency of offline PbRL. OPRIDE consists of two key features: a principled exploration strategy that maximizes the informativeness of the queries and a discount scheduling mechanism aimed at mitigating overoptimization of the learned reward functions. Through empirical evaluations, we demonstrate that OPRIDE significantly outperforms prior methods, achieving strong performance with notably fewer queries. Moreover, we provide theoretical guarantees of the algorithm's efficiency. Experimental results across various locomotion, manipulation, and navigation tasks underscore the efficacy and versatility of our approach.", "keywords": [ "preference-based RL", "offline RL", "reinforcement learning" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "MFrqTfubEB", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Fan Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Researcher" }, { "name": "Haifeng Wang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "CTO" }, { "name": "Jizhou Huang", "gender": "Male", "institution": "Baidu Inc.", "country": "CN", "position": "Chief Architect" }, { "name": "Kaixin Xiong", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Shi Gong", "gender": "Male", "institution": "Baidu", "country": "", "position": "Researcher" }, { "name": "Xiao Tan", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Xiaoqing Ye", "gender": "Female", "institution": "Baidu Inc.", "country": "CN", "position": "Researcher and Developer" }, { "name": "Yumeng Zhang", "gender": "unknown", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "hua wu", "gender": "Female", "institution": "Baidu", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 21, "n_ref": 49, "n_ref_all": 71, "n_fig": 9, "n_tab": 6, "L_tab": 1720, "n_element_tab": 223, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 717, "n_element_tab_1": 126, "formula_len_all": 1213, "formula_len_all_1": 1295, "len_all": 127711, "len_all_1": 56218, "len_abs": 5337, "len_title": 134, "len_sents": 34371, "len_sents_1": 23919, "n_sents": 287, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 981, "title": "BEVWorld: A Multimodal World Model for Autonomous Driving via Unified BEV Latent Space", "abs": "World models are receiving increasing attention in autonomous driving for their capability to predict potential future scenarios. In this paper, we present BEVWorld, a novel approach that tokenize multimodal sensor inputs into a unified and compact Bird's Eye View (BEV) latent space for environment modeling. The world model consists of two parts: the multi-modal tokenizer and the latent BEV sequence diffusion model. The multi-modal tokenizer first encodes multi-modality information and the decoder is able to reconstruct the latent BEV tokens into LiDAR and image observations by ray-casting rendering in a self-supervised manner. Then the latent BEV sequence diffusion model predicts future scenarios given action tokens as conditions. Experiments demonstrate the effectiveness of BEVWorld in autonomous driving tasks, showcasing its capability in generating future scenes and benefiting downstream tasks such as perception and motion prediction. Code will be available soon.", "keywords": [ "World Model", "End-to-end Autonomous Driving", "BEV" ], "rating_list": [ 6, 3, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 3, 1 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "MFZjrTFE7h", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Feng Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Hebei Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Peixi Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Xiaoyan Sun", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yansong Peng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yueyi Zhang", "gender": "unknown", "institution": "Hesai Technology", "country": "CN", "position": "Staff Engineer" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 40, "n_ref": 99, "n_ref_all": 110, "n_fig": 4, "n_tab": 13, "L_tab": 6264, "n_element_tab": 1062, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 5291, "n_element_tab_1": 571, "formula_len_all": 848, "formula_len_all_1": 758, "len_all": 144127, "len_all_1": 61589, "len_abs": 224, "len_title": 129, "len_sents": 31694, "len_sents_1": 21664, "n_sents": 227, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 81, "L_abs": 1383, "title": "D-FINE: Redefine Regression Task of DETRs as Fine-grained Distribution Refinement", "abs": "We introduce D-FINE, a powerful real-time object detector that achieves outstanding localization precision by redefining the bounding box regression task in DETR models. D-FINE comprises two key components: Fine-grained Distribution Refinement (FDR) and Global Optimal Localization Self-Distillation (GO-LSD). FDR transforms the regression process from predicting fixed coordinates to iteratively refining probability distributions, providing a fine-grained intermediate representation that significantly enhances localization accuracy. GO-LSD is a bidirectional optimization strategy that transfers localization knowledge from refined distributions to shallower layers through self-distillation, while also simplifying the residual prediction tasks for deeper layers. Additionally, D-FINE incorporates lightweight optimizations in computationally intensive modules and operations, achieving a better balance between speed and accuracy. Specifically, D-FINE-L / X achieves 54.0% / 55.8% AP on the COCO dataset at 124 / 78 FPS on an NVIDIA T4 GPU. When pretrained on Objects365, D-FINE-L / X attains 57.1% / 59.3% AP, surpassing all existing real-time detectors. Furthermore, our method significantly enhances the performance of a wide range of DETR models by up to 5.3% AP with negligible extra parameters and training costs. Our code and models: https://github.com/Peterande/D-FINE.", "keywords": [ "Object Detection", "Real-Time", "Detection Transformer", "Knowledge Distillation" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "MF7ljU8xcf", "primary_area": "learning theory", "n_author": 7, "author_info_list": [ { "name": "Andrew Gordon Wilson", "gender": "Not Specified", "institution": "New York University", "country": "US", "position": "Full Professor" }, { "name": "Anming Gu", "gender": "Male", "institution": "Boston University", "country": "US", "position": "Researcher" }, { "name": "Diego Granziol", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Marc Anton Finzi", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Sanyam Kapoor", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Christopher De Sa", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Associate Professor" }, { "name": "Zico Kolter", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 71, "n_formula_1": 22, "n_ref_uni": 54, "n_ref": 88, "n_ref_all": 117, "n_fig": 10, "n_tab": 1, "L_tab": 69, "n_element_tab": 9, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4854, "formula_len_all_1": 1956, "len_all": 233224, "len_all_1": 64290, "len_abs": 1166, "len_title": 97, "len_sents": 55338, "len_sents_1": 30046, "n_sents": 455, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1177, "title": "Larger Language Models Provably Generalize Better", "abs": "Why do larger language models generalize better? To explore this question, we develop generalization bounds on the pretraining objective of large language models (LLMs) in the compute-optimal regime, as described by the Chinchilla scaling laws. We introduce a novel, fully empirical Freedman-type martingale concentration inequality that tightens existing bounds by accounting for the variance of the loss function. The generalization bound can be broken into three contributions: the number of parameters per token, the loss variance, and the quantization error at a fixed bitrate. As language models are scaled up, the number of parameters per data point stays constant; however, both the loss variance and the quantization error decrease, implying that larger models should have \\emph{smaller} generalization gaps. We examine why larger models tend to be more quantizable from an information theoretic perspective, showing that the rate at which they can integrate new information grows slower than their capacity on the compute optimal frontier. From these findings we produce a scaling law for the generalization gap, showing that our bounds decrease in a predictable way.", "keywords": [ "generalization bounds", "language models", "scaling laws" ], "rating_list": [ 6, 6, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 1, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 2, 3, 3, 2, 2 ] }, { "paper_id": "MEnPLXJNng", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Nicu Sebe", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Full Professor" }, { "name": "Xiaojun Wu", "gender": "Male", "institution": "Jiangnan University", "country": "CN", "position": "Full Professor" }, { "name": "Yue Song", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Ziheng Chen", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "PhD student" } ], "n_formula": 139, "n_formula_1": 40, "n_ref_uni": 41, "n_ref": 184, "n_ref_all": 232, "n_fig": 1, "n_tab": 9, "L_tab": 4332, "n_element_tab": 258, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1960, "n_element_tab_1": 41, "formula_len_all": 9689, "formula_len_all_1": 3002, "len_all": 266801, "len_all_1": 77658, "len_abs": 964, "len_title": 104, "len_sents": 63719, "len_sents_1": 25952, "n_sents": 785, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 965, "title": "Riemannian Transformation Layers for General Geometries", "abs": "Recently, deep neural networks on manifold-valued representations have garnered significant attention in various machine learning applications. Several studies have attempted to generalize traditional Euclidean transformation layers, such as Fully Connected (FC) and convolutional layers, to non-Euclidean geometries. However, the previous approaches typically focus on a select few manifolds and rely on the specific properties of the target manifold. In this work, we propose a theoretical framework for constructing Riemannian FC and convolutional layers over general geometries, providing broader applicability. Utilizing this framework, we design convolutional networks across five distinct geometries of the Symmetric Positive Definite (SPD) manifold, as well as networks under two Grassmannian perspectives. Extensive experiments demonstrate that the proposed Riemannian convolutional networks significantly outperform existing SPD and Grassmannian networks.", "keywords": [ "Manifold Learning", "Representation Learning", "Riemannian Manifolds" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "MEF8SyXuXG", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Arash Tavakoli", "gender": "Male", "institution": "Riot Games", "country": "US", "position": "Staff Research Scientist" }, { "name": "Nemanja Rakicevic", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Sina Ghiassian", "gender": "Male", "institution": "Spotify", "country": "", "position": "Researcher" } ], "n_formula": 31, "n_formula_1": 28, "n_ref_uni": 72, "n_ref": 101, "n_ref_all": 128, "n_fig": 14, "n_tab": 5, "L_tab": 2801, "n_element_tab": 142, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2281, "formula_len_all_1": 1639, "len_all": 218289, "len_all_1": 68456, "len_abs": 1259, "len_title": 106, "len_sents": 53875, "len_sents_1": 33399, "n_sents": 360, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1264, "title": "Learning in complex action spaces without policy gradients", "abs": "Conventional wisdom suggests that policy gradient methods are better suited to complex action spaces than action-value methods. However, foundational studies have shown equivalences between these paradigms in small and finite action spaces (O'Donoghue et al., 2017; Schulman et al., 2017a). This raises the question of why their computational applicability and performance diverge as the complexity of the action space increases. We hypothesize that the apparent superiority of policy gradients in such settings stems not from intrinsic qualities of the paradigm, but from universal principles that can also be applied to action-value methods to serve similar functionality. We identify three such principles and provide a framework for incorporating them into action-value methods. To support our hypothesis, we instantiate this framework in what we term QMLE, for Q-learning with maximum likelihood estimation. Our results show that QMLE can be applied to complex action spaces with a controllable computational cost that is comparable to that of policy gradient methods, all without using policy gradients. Furthermore, QMLE demonstrates strong performance on the DeepMind Control Suite, even when compared to the state-of-the-art methods such as DMPO and D4PG.", "keywords": [ "action-value learning", "policy gradient methods", "complex action spaces" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "MDvecs7EvO", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Daniil Gavrilov", "gender": "unknown", "institution": "T-Bank", "country": "RU", "position": "Principal Researcher" }, { "name": "Nikita Balagansky", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "", "position": "MS student" }, { "name": "Yan Viktorovich Maximov", "gender": "Male", "institution": "Higher School of Economics, Higher School of Economics", "country": "RU", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 10, "n_ref": 23, "n_ref_all": 45, "n_fig": 14, "n_tab": 3, "L_tab": 2771, "n_element_tab": 390, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 1471, "n_element_tab_1": 211, "formula_len_all": 594, "formula_len_all_1": 517, "len_all": 124151, "len_all_1": 58683, "len_abs": 2063, "len_title": 103, "len_sents": 37767, "len_sents_1": 26402, "n_sents": 290, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1200, "title": "Mechanistic Permutability: Match Features Across Layers", "abs": "Understanding how features evolve across layers in deep neural networks is a fundamental challenge in mechanistic interpretability, particularly due to polysemanticity and feature superposition. While Sparse Autoencoders (SAEs) have been used to extract interpretable features from individual layers, aligning these features across layers has remained an open problem. In this paper, we introduce SAE Match, a novel, data-free method for aligning SAE features across different layers of a neural network. Our approach involves matching features by minimizing the mean squared error between the folded parameters of SAEs, a technique that incorporates activation thresholds into the encoder and decoder weights to account for differences in feature scales. Through extensive experiments on the Gemma 2 language model, we demonstrate that our method effectively captures feature evolution across layers, improving feature matching quality. We also show that features persist over several layers and that our approach can approximate hidden states across layers. Our work advances the understanding of feature dynamics in neural networks and provides a new tool for mechanistic interpretability studies.", "keywords": [ "Interpretability", "LLM features", "SAE" ], "rating_list": [ 8, 5, 5, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 4, 4, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "MD4ifad9v5", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Boris Wembe", "gender": "Male", "institution": "Universit\u00e4t Paderborn", "country": "DE", "position": "Postdoc" }, { "name": "Christian Offen", "gender": "Male", "institution": "Universit\u00e4t Paderborn", "country": "DE", "position": "Postdoc" }, { "name": "Pranav Singh", "gender": "Male", "institution": "University of Bath", "country": "GB", "position": "Associate Professor" }, { "name": "Sina Ober-Bl\u00f6baum", "gender": "Female", "institution": "Paderborn University", "country": "", "position": "Full Professor" }, { "name": "Sofya Maslovskaya", "gender": "Female", "institution": "Universit\u00e4t Paderborn", "country": "DE", "position": "Postdoc" } ], "n_formula": 54, "n_formula_1": 23, "n_ref_uni": 50, "n_ref": 106, "n_ref_all": 135, "n_fig": 10, "n_tab": 4, "L_tab": 1065, "n_element_tab": 99, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 546, "n_element_tab_1": 59, "formula_len_all": 5629, "formula_len_all_1": 1839, "len_all": 188544, "len_all_1": 71083, "len_abs": 1493, "len_title": 127, "len_sents": 52387, "len_sents_1": 31735, "n_sents": 424, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1500, "title": "Adaptive higher order reversible integrators for memory efficient deep learning", "abs": "The depth of networks plays a crucial role in the effectiveness of deep learning. However, the memory requirement for backpropagation scales linearly with the number of layers, which leads to memory bottlenecks during training. Moreover, deep networks are often unable to handle time-series data appearing at irregular intervals. These issues can be resolved by considering continuous-depth networks based on the neural ODE framework in combination with reversible integration methods that allow for variable time-steps. Reversibility of the method ensures that the memory requirement for training is independent of network depth, while variable time-steps are required for assimilating time-series data on irregular intervals. However, at present, there are no known higher-order reversible methods with this property. High-order methods are especially important when a high level of accuracy in learning is required or when small time-steps are necessary due to large errors in time integration of neural ODEs, for instance in context of complex dynamical systems such as Kepler systems and molecular dynamics. The requirement of small time-steps when using a low-order method can significantly increase the computational cost of training as well as inference. In this work, we present an approach for constructing high-order reversible methods that allow adaptive time-stepping. Our numerical tests show the advantages in computational speed when applied to the task of learning dynamical systems.", "keywords": [ "neural ODE", "backpropagation", "reversible neural networks", "learning dynamical systems", "high-order integration methods", "variable time-steps" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 4, 2, 2 ], "contribution_list": [ 3, 2, 1 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "MCHuGOkExF", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Haifeng Chen", "gender": "unknown", "institution": "NEC-Labs", "country": "US", "position": "Researcher" }, { "name": "Jonathan Light", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "PhD student" }, { "name": "Wei Cheng", "gender": "Male", "institution": "NEC-Labs", "country": "US", "position": "Researcher" }, { "name": "Wenchao Yu", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Xujiang Zhao", "gender": "Male", "institution": "NEC Labs America", "country": "", "position": "Researcher" }, { "name": "Yanchi Liu", "gender": "Male", "institution": "NEC-Labs", "country": "US", "position": "Researcher" }, { "name": "Yiyou Sun", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Yue Wu", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Ziniu Hu", "gender": "Male", "institution": "xAI", "country": "US", "position": "Member of Technical Staff" } ], "n_formula": 20, "n_formula_1": 12, "n_ref_uni": 47, "n_ref": 96, "n_ref_all": 153, "n_fig": 9, "n_tab": 19, "L_tab": 9867, "n_element_tab": 887, "n_fig_1": 6, "n_tab_1": 8, "L_tab_1": 1150, "n_element_tab_1": 197, "formula_len_all": 989, "formula_len_all_1": 682, "len_all": 246073, "len_all_1": 66948, "len_abs": 932, "len_title": 113, "len_sents": 66862, "len_sents_1": 27093, "n_sents": 584, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 936, "title": "Scattered Forest Search: Smarter Code Space Exploration with LLMs", "abs": "We propose a novel approach to scaling LLM inference for code generation. We\nframe code generation as a black box optimization problem within the code space,\nand employ optimization-inspired techniques to enhance exploration. Specifically,\nwe introduce SCATTERED FOREST SEARCH to enhance solution diversity while\nsearching for solutions. Our theoretical analysis illustrates how these methods avoid\nlocal optima during optimization. Extensive experiments on HumanEval, MBPP,\nAPPS, CodeContests, and Leetcode reveal significant performance improvements.\nFor instance, our method achieves a pass@1 rate of 67.1% on HumanEval+ and\n87.2% on HumanEval with GPT-3.5, marking improvements of 8.6% and 4.3%\nover the state-of-the-art, while also halving the iterations needed to find the correct\nsolution. Furthermore, our method scales more efficiently than existing search\ntechniques, including tree search, line search, and repeated sampling.", "keywords": [ "LLM", "code generation", "optimization", "search", "agent", "inference scaling" ], "rating_list": [ 6, 5, 6, 6, 8 ], "soundness_list": [ 3, 3, 2, 4, 3 ], "presentation_list": [ 2, 2, 2, 4, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 2, 4, 4 ] }, { "paper_id": "MBkoYFftRa", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Abhinav Shrivastava", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "Associate Professor" }, { "name": "Di Wu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Director" }, { "name": "Han Cai", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Matthew Gwilliam", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Zhiyu Cheng", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Manager" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 30, "n_ref": 38, "n_ref_all": 64, "n_fig": 12, "n_tab": 2, "L_tab": 438, "n_element_tab": 74, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 438, "n_element_tab_1": 74, "formula_len_all": 183, "formula_len_all_1": 180, "len_all": 113966, "len_all_1": 55984, "len_abs": 266, "len_title": 113, "len_sents": 27452, "len_sents_1": 26069, "n_sents": 203, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1576, "title": "Accelerate High-Quality Diffusion Models with Inner Loop Feedback", "abs": "We propose Inner Loop Feedback (ILF), a novel approach to accelerate diffusion models' inference. ILF trains a lightweight module to predict future features in the denoising process by leveraging the outputs from a chosen diffusion backbone block at a given time step. This approach exploits two key intuitions; (1) the outputs of a given block at adjacent time steps are similar, and (2) performing partial computations for a step imposes a lower burden on the model than skipping the step entirely. Our method is highly flexible, since we find that the feedback module itself can simply be a block from the diffusion backbone, with all settings copied. Its influence on the diffusion forward can be tempered with a learnable scaling factor from zero initialization. We train this module using distillation losses; however, unlike some prior work where a full diffusion backbone serves as the student, our model freezes the backbone, training only the feedback module. While many efforts to optimize diffusion models focus on achieving acceptable image quality in extremely few steps (1-4 steps), our emphasis is on matching best case results (typically achieved in 20 steps) while significantly reducing runtime. ILF achieves this balance effectively, demonstrating strong performance for both class-to-image generation with diffusion transformer (DiT) and text-to-image generation with DiT-based PixArt-alpha and PixArt-sigma. The quality of ILF's 1.7x-1.8x speedups are confirmed by FID, CLIP score, CLIP Image Quality Assessment, ImageReward, and qualitative comparisons.", "keywords": [ "diffusion", "image generation", "efficiency" ], "rating_list": [ 3, 5, 3, 1 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 1, 1 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "MBDH5zyxHM", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Ahmad Beirami", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Hadi Jamali-Rad", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "Sayak Mukherjee", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "PhD student" }, { "name": "Anuj Rajeeva Singh", "gender": "Male", "institution": "Shell Global Solutions B.V.", "country": "NL", "position": "Researcher" } ], "n_formula": 29, "n_formula_1": 19, "n_ref_uni": 39, "n_ref": 118, "n_ref_all": 137, "n_fig": 7, "n_tab": 7, "L_tab": 1527, "n_element_tab": 207, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 416, "n_element_tab_1": 33, "formula_len_all": 2846, "formula_len_all_1": 1386, "len_all": 158347, "len_all_1": 67162, "len_abs": 901, "len_title": 89, "len_sents": 40938, "len_sents_1": 29765, "n_sents": 310, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 41, "L_abs": 903, "title": "Controlled Denoising For Diffusion Models", "abs": "Aligning diffusion models for downstream tasks often requires finetuning new models or costly inference-time solutions (e.g., gradient-based guidance) to allow sampling from the reward-tilted posterior. In this work, we explore a simple and low-cost inference-time gradient-free guidance approach, called conditional controlled denoising (C-Code), that circumvents the need for differentiable guidance functions and model finetuning. C-Code is a block-wise sampling method with adjustable conditioning on a reference image applied during intermediate denoising steps, allowing for efficient alignment with downstream rewards. Experiments demonstrate that, despite its simplicity, C-Code offers a balanced trade-off between reward alignment, prompt instruction following, and inference cost, outperforming state-of-the-art baselines. Our code is available at: https://anonymous.4open.science/r/CoDe-Repo.", "keywords": [ "Generative Models", "Computer Vision", "Diffusion Models", "Guidance" ], "rating_list": [ 5, 3, 5, 5, 5 ], "soundness_list": [ 2, 1, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 1, 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3, 5 ] }, { "paper_id": "MBBRHDuiwM", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Juexiao Zhang", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Yann LeCun", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Director" }, { "name": "Yubei Chen", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Assistant Professor" }, { "name": "Zeyu Yun", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 4, "n_ref_uni": 22, "n_ref": 56, "n_ref_all": 87, "n_fig": 10, "n_tab": 6, "L_tab": 1730, "n_element_tab": 111, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1079, "n_element_tab_1": 136, "formula_len_all": 682, "formula_len_all_1": 152, "len_all": 217220, "len_all_1": 67504, "len_abs": 2803, "len_title": 126, "len_sents": 63501, "len_sents_1": 30285, "n_sents": 574, "n_sents_1": 273, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1249, "title": "URLOST: Unsupervised Representation Learning without Stationarity or Topology", "abs": "Unsupervised representation learning has seen tremendous progress. However, it is constrained by its reliance on domain specific stationarity and topology, a limitation not found in biological intelligence systems. For instance, unlike computer vision, human vision can process visual signals sampled from highly irregular and non-stationary sensors. We introduce a novel framework that learns from high-dimensional data without prior knowledge of stationarity and topology. Our model, abbreviated as URLOST, combines a learnable self-organizing layer, spectral clustering, and a masked autoencoder (MAE). We evaluate its effectiveness on three diverse data modalities including simulated biological vision data, neural recordings from the primary visual cortex, and gene expressions. Compared to state-of-the-art unsupervised learning methods like SimCLR and MAE, our model excels at learning meaningful representations across diverse modalities without knowing their stationarity or topology. It also outperforms other methods that are not dependent on these factors, setting a new benchmark in the field. We position this work as a step toward unsupervised learning methods capable of generalizing across diverse high-dimensional data modalities.", "keywords": [ "Unsupervised learning", "Self-Supervised Learning", "NeuroAI", "Multi-Modality", "Human Vision", "Biologically-inspired Models" ], "rating_list": [ 8, 3, 8, 5, 8 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 4, 2, 3 ], "confidence_list": [ 3, 4, 4, 3, 4 ] }, { "paper_id": "MB53uAZKSc", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Fartash Faghri", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Hadi Pouransari", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Jeffrey Li", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Mehrdad Farajtabar", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Mohammadreza Armandpour", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Oncel Tuzel", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Raviteja Vemulapalli", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Sachin Mehta", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Samy Bengio", "gender": "Male", "institution": "Apple", "country": "US", "position": "Senior Director" }, { "name": "Seyed Iman Mirzadeh", "gender": "Male", "institution": "Apple", "country": "US", "position": "Research Engineer" }, { "name": "vaishaal naanny shankar", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 55, "n_ref": 121, "n_ref_all": 153, "n_fig": 9, "n_tab": 11, "L_tab": 8053, "n_element_tab": 835, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3711, "n_element_tab_1": 455, "formula_len_all": 309, "formula_len_all_1": 169, "len_all": 270078, "len_all_1": 68890, "len_abs": 1365, "len_title": 75, "len_sents": 66247, "len_sents_1": 27970, "n_sents": 498, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1360, "title": "TiC-LM: A Multi-Year Benchmark for Continual Pretraining of Language Models", "abs": "Large language models (LLMs) are trained on data crawled over many years from the web. We investigate how quickly LLMs become outdated as the world evolves with time and how to best update them with newer data. Specifically, we simulate a world where the latest dump of Common Crawl (CC), the most prominent public source of pre-training data, is used every month to *continually* train an LLM. We design various dynamic evaluations from the CC data, Wikipedia, StackExchange, and code documentations to measure continual learning metrics such as forgetting and forward transfer. Notably, our TiC-CC training data is more than 100 times larger compared with prior continual learning benchmarks for language modeling. We discover that recent DataComp-LM models trained on data before 2023 have already become outdated, incurring up to 45\\% larger noun-perplexity on 2024 Wikipedia articles compared to pre-2023 articles. Further, we use our setup to evaluate the effectiveness of several large-scale continual learning methods and find that replaying older data is most effective for combating forgetting: for previously seen CC dumps, it can reduce the regret on held-out loss by 60\\% compared to other optimizer and loss-based interventions. However, some domains evolve more quickly than others, favoring different trade-offs between mixing old and new data.", "keywords": [ "language models", "continual learning", "benchmark", "temporal adaptation" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 4, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "M9p2SIq0Oj", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Chaofan Pan", "gender": "Non-Binary", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "PhD student" }, { "name": "Fan Min", "gender": "Male", "institution": "Southwest Petroleum University", "country": "CN", "position": "Full Professor" }, { "name": "Jiafen Liu", "gender": "Female", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "Associate Professor" }, { "name": "Tianrui Li", "gender": "Male", "institution": "Southwest Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Wei Wei", "gender": "Male", "institution": "shanxi university", "country": "CN", "position": "Full Professor" }, { "name": "Xin Yang", "gender": "Male", "institution": "Southwestern University of Finance and Economics", "country": "CN", "position": "Full Professor" }, { "name": "Yanhua Li", "gender": "Female", "institution": "Southwest University of Finance and Economics", "country": "", "position": "PhD student" }, { "name": "\u718a\u6797\u535a", "gender": "Male", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 54, "n_ref": 97, "n_ref_all": 135, "n_fig": 17, "n_tab": 8, "L_tab": 2902, "n_element_tab": 275, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 2219, "n_element_tab_1": 71, "formula_len_all": 746, "formula_len_all_1": 724, "len_all": 202435, "len_all_1": 73503, "len_abs": 3428, "len_title": 137, "len_sents": 61357, "len_sents_1": 33158, "n_sents": 477, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1482, "title": "ARC-RL: Self-Evolution Continual Reinforcement Learning via Action Representation Space", "abs": "Continual Reinforcement Learning (CRL) is a powerful tool that enables agents to learn a sequence of tasks, accumulating knowledge learned in the past and using it for problemsolving or future task learning. However, existing CRL methods all assume that the agent\u2019s capabilities remain static within dynamic environments, which doesn\u2019t reflect realworld scenarios where capabilities evolve. This paper introduces *Self-Evolution Continual Reinforcement Learning* (SE-CRL), a new and realistic problem where the agent\u2019s action space continually changes. It presents a significant challenge for RL agents: How can policy generalization across different action spaces be achieved? Inspired by the cortical functions that lead to consistent human behavior, we propose an **A**ction **R**epresentation **C**ontinual **R**einforcement **L**earning framework (ARC-RL) to address this challenge. Our framework builds a representation space for actions by self-supervised learning on transitions, decoupling the agent\u2019s policy from the specific action space. For a new action space, the decoder of the action representation is expanded or masked for adaptation and regularized fine-tuned to improve the stability of the policy. Furthermore, we release a benchmark based on MiniGrid to validate the effectiveness of methods for SE-CRL. Experimental results demonstrate that our framework significantly outperforms popular CRL methods by generalizing the policy across different action spaces.", "keywords": [ "continual learning", "lifelong learning", "reinforcement learning" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 1, 3, 2, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "M9iky9Ruhx", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Difei Gao", "gender": "Male", "institution": "NUS", "country": "SG", "position": "Postdoc" }, { "name": "Weixian Lei", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zheng Shou", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 45, "n_ref": 137, "n_ref_all": 163, "n_fig": 10, "n_tab": 8, "L_tab": 3218, "n_element_tab": 298, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2123, "n_element_tab_1": 97, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 193120, "len_all_1": 72024, "len_abs": 352, "len_title": 102, "len_sents": 48092, "len_sents_1": 33030, "n_sents": 419, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 54, "L_abs": 921, "title": "Grounding Multimodal Large Language Model in GUI World", "abs": "Recent advancements in Multimodal Large Language Models (MLLMs) have accelerated the development of Graphical User Interface (GUI) agents capable of automating complex tasks across digital platforms. However, precise GUI element grounding remains a key challenge for accurate interaction and generalization. In this work, we present an effective GUI grounding framework, which includes an automated data collection engine that gathers extensive GUI screenshots and annotations to ensure broad generalization. We also propose a lightweight and flexible GUI grounding module designed to efficiently localize UI elements by pre-training on the collected data, and introduce a novel method to integrate this module with MLLMs for the effective execution of GUI tasks. Our approach demonstrates superior performance in task accuracy and adaptability, as validated by benchmarks such as ScreenSpot, MiniWob, AITW, and Mind2Web.", "keywords": [ "Multimodal Large Language Model", "GUI Element Grounding" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "M9U49u9GA7", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Agam Shah", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Chao Zhang", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Liqin Ye", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Sudheer Chava", "gender": "Not Specified", "institution": "Georgia Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 14, "n_ref_uni": 35, "n_ref": 91, "n_ref_all": 104, "n_fig": 3, "n_tab": 10, "L_tab": 5579, "n_element_tab": 233, "n_fig_1": 1, "n_tab_1": 8, "L_tab_1": 3491, "n_element_tab_1": 174, "formula_len_all": 1458, "formula_len_all_1": 1676, "len_all": 222712, "len_all_1": 65367, "len_abs": 1739, "len_title": 128, "len_sents": 47444, "len_sents_1": 25820, "n_sents": 364, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1818, "title": "SiDyP: Simplex Diffusion with Dynamic Prior for Denoising Llama-Generated Labels", "abs": "The traditional process of creating labeled datasets is not only labor-intensive but also expensive. Recent breakthroughs in open-source large language models (LLMs), such as Llama-3, have opened a new avenue in generating labeled datasets automatically for various natural language processing (NLP) tasks to provide an alternative to such expensive annotation process. However, the reliability of such auto-generated labels remains a significant concern due to inherent inaccuracies. When learning from such noisy labels, the model's generalization is likely to be harmed as it is prone to overfit those label noises. In this paper, we propose the \\textbf{Si}mplex Diffusion with a \\textbf{Dy}namic \\textbf{P}rior (\\textbf{SiDyP}) model to calibrate \\hl{classifier's} predication, thus enhancing its robustness towards noisy labels. Our framework leverages simplex diffusion model to iteratively correct noisy labels conditioned on training dynamic trajectories obtained from classifier finetuning. \\hl{The \\textbf{P}rior in SiDyP refers} to the potential true label candidates which was obtained according to neighborhood label distribution in text embedding space. \\hl{It is \\textbf{Dy}namic because we} progressively distill these candidates based on the feedback of the diffusion model. Our SiDyP model can increase the performance of the BERT classifier fine-tuned on both zero-shot and few-shot Llama-3 generated noisy label datasets by an average of 5.33\\% and 7.69\\% respectively. Our extensive experiments, which explore different LLMs, diverse noise types (real-world and synthetic), ablation studies, and multiple baselines, demonstrate the effectiveness of SiDyP across a range of NLP tasks. We will make code and data publicly (under a CC BY 4.0 license) available on GitHub upon publication of the work.", "keywords": [ "Diffusion Model", "Learning from Noisy Labels", "Soft Labels" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "M9SKazbVkJ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Ching-Chun Chang", "gender": "Not Specified", "institution": "National Institute of Informatics", "country": "JP", "position": "Assistant Professor" }, { "name": "Futa Kai Waseda", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Isao Echizen", "gender": "Male", "institution": "National Institute of Informatics", "country": "JP", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 29, "n_ref": 61, "n_ref_all": 93, "n_fig": 17, "n_tab": 19, "L_tab": 5617, "n_element_tab": 496, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 472, "n_element_tab_1": 60, "formula_len_all": 1310, "formula_len_all_1": 646, "len_all": 152845, "len_all_1": 59148, "len_abs": 1407, "len_title": 150, "len_sents": 47988, "len_sents_1": 26421, "n_sents": 395, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 101, "L_abs": 1410, "title": "Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off", "abs": "Adversarial training often suffers from a robustness-accuracy trade-off, where achieving high robustness comes at the cost of accuracy.\nOne approach to mitigate this trade-off is leveraging invariance regularization, which encourages model invariance under adversarial perturbations; however, it still leads to accuracy loss.\nIn this work, we closely analyze the challenges of using invariance regularization in adversarial training and understand how to address them.\nOur analysis identifies two key issues: (1) a \"gradient conflict\" between invariance and classification objectives, leading to suboptimal convergence, and (2) the mixture distribution problem arising from diverged distributions between clean and adversarial inputs.\nTo address these issues, we propose Asymmetric Representation-regularized Adversarial Training (ARAT), which incorporates asymmetric invariance loss with stop-gradient operation and a predictor to avoid gradient conflict, and a split-BatchNorm (BN) structure to resolve the mixture distribution problem.\nOur detailed analysis demonstrates that each component effectively addresses the identified issues, offering novel insights into adversarial defense.\nARAT shows superiority over existing methods across various settings. Finally, we discuss the implications of our findings to knowledge distillation-based defenses, providing a new perspective on their relative successes.", "keywords": [ "computer vision", "adversarial defense", "adversarial robustness", "robustness-accuracy trade-off" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "M9SAhECerP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "David Meger", "gender": "Male", "institution": "McGill University", "country": "", "position": "Associate Professor" }, { "name": "Doina Precup", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Research Team Lead" }, { "name": "Hanna Yurchyk", "gender": "Female", "institution": "McGill University, McGill University", "country": "CA", "position": "MS student" }, { "name": "Sahand Rezaei-Shoshtari", "gender": "Male", "institution": "McGill University, McGill University", "country": "CA", "position": "PhD student" }, { "name": "Scott Fujimoto", "gender": "unknown", "institution": "Facebook", "country": "CA", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 21, "n_ref_uni": 67, "n_ref": 140, "n_ref_all": 181, "n_fig": 11, "n_tab": 6, "L_tab": 3769, "n_element_tab": 247, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2787, "n_element_tab_1": 145, "formula_len_all": 2019, "formula_len_all_1": 1580, "len_all": 192766, "len_all_1": 64865, "len_abs": 828, "len_title": 118, "len_sents": 54353, "len_sents_1": 28491, "n_sents": 408, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 830, "title": "Long-Term Fairness in Reinforcement Learning with Bisimulation Metrics", "abs": "Ensuring long-term fairness is crucial when developing automated decision making systems, specifically in dynamic and sequential environments. By maximizing their reward without consideration of fairness, AI agents can introduce disparities in their treatment of groups or individuals. In this paper, we establish the connection between bisimulation metrics and group fairness in reinforcement learning. We propose a novel approach that leverages bisimulation metrics to learn reward functions and observation dynamics, ensuring that learners treat groups fairly while reflecting the original problem. We demonstrate the effectiveness of our method in addressing disparities in sequential decision making problems through empirical evaluation on a standard fairness benchmark consisting of lending and college admission scenarios.", "keywords": [ "fairness", "reinforcement learning", "bisimulation" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "M992mjgKzI", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Benjamin Eysenbach", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Assistant Professor" }, { "name": "Kevin Frans", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Seohong Park", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" } ], "n_formula": 21, "n_formula_1": 4, "n_ref_uni": 37, "n_ref": 138, "n_ref_all": 178, "n_fig": 3, "n_tab": 5, "L_tab": 6463, "n_element_tab": 517, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 430, "n_element_tab_1": 55, "formula_len_all": 2362, "formula_len_all_1": 1114, "len_all": 219770, "len_all_1": 74933, "len_abs": 1388, "len_title": 97, "len_sents": 69190, "len_sents_1": 36719, "n_sents": 464, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1169, "title": "OGBench: Benchmarking Offline Goal-Conditioned RL", "abs": "Offline goal-conditioned reinforcement learning (GCRL) is a major problem in reinforcement learning (RL) because it provides a simple, unsupervised, and domain-agnostic way to acquire diverse behaviors and representations from unlabeled data without rewards. Despite the importance of this setting, we lack a standard benchmark that can systematically evaluate the capabilities of offline GCRL algorithms. In this work, we propose OGBench, a new, high-quality benchmark for algorithms research in offline goal-conditioned RL. OGBench consists of 8 types of environments, 85 datasets, and reference implementations of 6 representative offline GCRL algorithms. We have designed these challenging and realistic environments and datasets to directly probe different capabilities of algorithms, such as stitching, long-horizon reasoning, and the ability to handle high-dimensional inputs and stochasticity. While representative algorithms may rank similarly on prior benchmarks, our experiments reveal stark strengths and weaknesses in these different capabilities, providing a strong foundation for building new algorithms. Project page: https://seohong.me/projects/ogbench", "keywords": [ "reinforcement learning" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 4, 3, 4 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "M922KJFO7O", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Amir Zandieh", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Insu Han", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Vahab Mirrokni", "gender": "Male", "institution": "Google Research", "country": "", "position": "Distinguished Scientist" }, { "name": "amin karbasi", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Associate Professor" } ], "n_formula": 33, "n_formula_1": 16, "n_ref_uni": 22, "n_ref": 39, "n_ref_all": 55, "n_fig": 1, "n_tab": 2, "L_tab": 2, "n_element_tab": 1, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 744, "n_element_tab_1": 89, "formula_len_all": 3559, "formula_len_all_1": 2066, "len_all": 109817, "len_all_1": 66238, "len_abs": 1187, "len_title": 130, "len_sents": 33433, "len_sents_1": 28839, "n_sents": 282, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1191, "title": "ClusterGen: Token Generation in Sublinear Time and Memory with Clustering KV Cache", "abs": "Despite the significant success of large language models (LLMs), their extensive memory requirements pose challenges for deploying them in long-context token generation. The substantial memory footprint of LLM decoders arises from the necessity to store all previous tokens in the attention module, a requirement imposed by key-value (KV) caching. In this work, our focus is on developing an efficient compression technique for the KV cache. Empirical evidence indicates a significant clustering tendency within key embeddings in the attention module. Building on this key insight, we have devised a novel caching method with sublinear complexity, employing online clustering on key tokens and online \n sampling on values. The result is a provably accurate and efficient attention decoding algorithm, termed ClusterGen. Not only does this algorithm ensure a sublinear memory footprint and sublinear time complexity, but we also establish a tight error bound for our approach. Empirical evaluations on long-context question-answering tasks demonstrate that ClusterGen significantly outperforms existing and state-of-the-art KV cache compression methods in terms of performance and efficiency.", "keywords": [ "KV cache", "large language models", "clustering" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "M8xtZuxqC5", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Beidi Chen", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Huaxiu Yao", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Lingfeng Shen", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Linjun Zhang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Associate Professor" }, { "name": "Ryumei Nakada", "gender": "unknown", "institution": "Rutgers University", "country": "", "position": "PhD student" }, { "name": "TaiMing Lu", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Undergrad student" }, { "name": "Weiting Tan", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Xinyu Yang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 7, "n_ref_uni": 48, "n_ref": 78, "n_ref_all": 99, "n_fig": 7, "n_tab": 6, "L_tab": 2447, "n_element_tab": 176, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 543, "n_element_tab_1": 67, "formula_len_all": 2193, "formula_len_all_1": 632, "len_all": 175205, "len_all_1": 65671, "len_abs": 1138, "len_title": 121, "len_sents": 47956, "len_sents_1": 30078, "n_sents": 356, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1152, "title": "It Takes Two: On the Seamlessness between Reward and Policy Model in RLHF", "abs": "Reinforcement Learning from Human Feedback (RLHF) involves training policy models (PMs) and reward models (RMs) to align language models with human preferences. Instead of focusing solely on PMs and RMs independently, we propose to examine their interactions during fine-tuning, introducing the concept of \\textbf{seamlessness}. Our study starts with observing the saturation phenomenon, where continual improvements in RM and PM do not translate into RLHF progress. Our analysis shows that RMs fail to assign proper scores to PM responses, resulting in a 35% mismatch rate with human preferences, highlighting a significant discrepancy between PM and RM. To measure seamlessness between PM and RM without human effort, we propose an automatic metric, SEAM. SEAM quantifies the discrepancies between PM and RM judgments induced by data samples. We validate the effectiveness of SEAM in data selection and model augmentation. Our experiments demonstrate that (1) using SEAM-filtered data for RL training improves RLHF performance by 4.5%, and (2) SEAM-guided model augmentation results in a 4% performance improvement over standard augmentation methods.", "keywords": [ "Large Language Model", "RLHF", "Data-centric methods" ], "rating_list": [ 3, 6, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "M8uf26TbrC", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "BIKE XIE", "gender": "unknown", "institution": "Kneron Inc", "country": "US", "position": "Researcher" }, { "name": "Cong Shen", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Associate Professor" }, { "name": "Jundong Li", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" }, { "name": "Zihan Chen", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 48, "n_ref": 101, "n_ref_all": 121, "n_fig": 4, "n_tab": 8, "L_tab": 3349, "n_element_tab": 429, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1620, "n_element_tab_1": 174, "formula_len_all": 500, "formula_len_all_1": 500, "len_all": 152703, "len_all_1": 69903, "len_abs": 1437, "len_title": 115, "len_sents": 38671, "len_sents_1": 32906, "n_sents": 250, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1442, "title": "Channel-Wise Mixed-Precision Quantization for Large Language Models", "abs": "Large Language Models (LLMs) have demonstrated remarkable success across a wide range of language tasks, but their deployment on edge devices remains challenging due to the substantial memory requirements imposed by their large parameter sizes. Weight-only quantization presents a promising solution to reduce the memory footprint of LLMs. However, existing approaches primarily focus on integer-bit quantization, limiting their adaptability to fractional-bit quantization tasks and preventing the full utilization of available storage space on devices. In this paper, we introduce Channel-Wise Mixed-Precision Quantization (CMPQ), a novel mixed-precision quantization method that allocates quantization precision in a channel-wise pattern based on activation distributions. By assigning different precision levels to different weight channels, CMPQ can adapt to any bit-width constraint. CMPQ employs a non-uniform quantization strategy and incorporates two outlier extraction techniques that collaboratively preserve the critical information, thereby minimizing the quantization loss. Experiments on different sizes of LLMs demonstrate that CMPQ not only enhances performance in integer-bit quantization tasks but also achieves significant performance gains with a modest increase in memory usage. CMPQ thus represents an adaptive and effective approach to LLM quantization, offering substantial benefits across diverse device capabilities.", "keywords": [ "Quantization", "Large language models", "Mixed-precision quantization" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "M8gXSFGkn2", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Boshen Xu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Qin Jin", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "Professor" }, { "name": "Sipeng Zheng", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Yang Du", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "MS student" }, { "name": "Zhinan Song", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Undergrad student" }, { "name": "Ziheng Wang", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 40, "n_ref": 100, "n_ref_all": 131, "n_fig": 15, "n_tab": 18, "L_tab": 3688, "n_element_tab": 488, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1878, "n_element_tab_1": 147, "formula_len_all": 673, "formula_len_all_1": 694, "len_all": 200556, "len_all_1": 62431, "len_abs": 1544, "len_title": 126, "len_sents": 58101, "len_sents_1": 28528, "n_sents": 405, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1573, "title": "Do Egocentric Video-Language Models Truly Understand Hand-Object Interactions?", "abs": "Egocentric video-language pretraining is a crucial step in advancing the understanding of hand-object interactions in first-person scenarios. Despite successes on existing testbeds, we find that current EgoVLMs can be easily misled by simple modifications, such as changing the verbs or nouns in interaction descriptions, with models struggling to distinguish between these changes. This raises the question: Do EgoVLMs truly understand hand-object interactions?'' To address this question, we introduce a benchmark called $\\textbf{EgoHOIBench}$, revealing the performance limitation of current egocentric models when confronted with such challenges. We attribute this performance gap to insufficient fine-grained supervision and the greater difficulty EgoVLMs experience in recognizing verbs compared to nouns. To tackle these issues, we propose a novel asymmetric contrastive objective named $\\textbf{EgoNCE++}$. For the video-to-text objective, we enhance text supervision by generating negative captions using large language models or leveraging pretrained vocabulary for HOI-related word substitutions. For the text-to-video objective, we focus on preserving an object-centric feature space that clusters video representations based on shared nouns. Extensive experiments demonstrate that EgoNCE++ significantly enhances EgoHOI understanding, leading to improved performance across various EgoVLMs in a range of tasks such as multi-instance retrieval, action recognition, and temporal understanding. Our code is available at https://anonymous.4open.science/r/EgoNCEpp", "keywords": [ "egocentric", "hand-object interaction", "video-language model", "contrastive learning", "multimodal representation learning" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 3, 5 ] }, { "paper_id": "M8XUdsjxQM", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Jiaming Lu", "gender": "Male", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Rong Zhu", "gender": "Male", "institution": "Fudan University", "country": "", "position": "Associate Professor" } ], "n_formula": 54, "n_formula_1": 21, "n_ref_uni": 18, "n_ref": 36, "n_ref_all": 55, "n_fig": 4, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 19, "n_element_tab_1": 2, "formula_len_all": 6196, "formula_len_all_1": 1150, "len_all": 117587, "len_all_1": 62270, "len_abs": 1019, "len_title": 122, "len_sents": 39396, "len_sents_1": 28888, "n_sents": 323, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1032, "title": "High dimensional Bayesian Optimization via Condensing-Expansion Projection", "abs": "In high-dimensional settings, Bayesian optimization (BO) can be expensive and infeasible. The random embedding Bayesian optimization algorithm is commonly used to address high-dimensional BO challenges. However, this method relies on the effective subspace assumption on the optimization problem's objective function, which limits its applicability. \nIn this paper, we introduce Condensing-Expansion Projection Bayesian optimization (CEPBO), a novel random projection-based approach for high-dimensional BO that does not reply on the effective subspace assumption. The approach is both simple to implement and highly practical. We present two algorithms based on different random projection matrices: the Gaussian projection matrix and the hashing projection matrix. Experimental results demonstrate that both algorithms outperform existing random embedding-based algorithms in most cases, achieving superior performance on high-dimensional BO problems.\nThe code is available in \\url{https://anonymous.4open.science/r/CEPBO-14429}.", "keywords": [ "Bayesian optimization", "random projection", "high-dimension", "Gaussian Process" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "M8OGl34Pmg", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Alessio Sampieri", "gender": "Male", "institution": "University of Roma \"La Sapienza\"", "country": "IT", "position": "PhD student" }, { "name": "Fabio Galasso", "gender": "Male", "institution": "University of Roma \"La Sapienza\"", "country": "IT", "position": "Associate Professor" }, { "name": "Indro Spinelli", "gender": "Male", "institution": "Sapienza University of Rome", "country": "IT", "position": "Assistant Professor" }, { "name": "Lamberto Ballan", "gender": "Male", "institution": "Universita' degli studi di Padova", "country": "IT", "position": "Associate Professor" }, { "name": "Luca Scofano", "gender": "Male", "institution": "University of Roma \"La Sapienza\"", "country": "IT", "position": "PhD student" }, { "name": "Tommaso Campari", "gender": "Male", "institution": "Fondazione Bruno Kessler", "country": "IT", "position": "Postdoc" }, { "name": "Valentino Sacco", "gender": "Male", "institution": "University of Roma \"La Sapienza\"", "country": "IT", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 49, "n_ref": 123, "n_ref_all": 146, "n_fig": 7, "n_tab": 8, "L_tab": 2682, "n_element_tab": 168, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1030, "n_element_tab_1": 86, "formula_len_all": 185, "formula_len_all_1": 180, "len_all": 172957, "len_all_1": 68515, "len_abs": 1284, "len_title": 95, "len_sents": 50252, "len_sents_1": 34236, "n_sents": 370, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 47, "L_abs": 1294, "title": "Following the Human Thread in Social Navigation", "abs": "The success of collaboration between humans and robots in shared environments relies on the robot's real-time adaptation to human motion. Specifically, in Social Navigation, the agent should be close enough to assist but ready to back up to let the human move freely, avoiding collisions. Human trajectories emerge as crucial cues in Social Navigation, but they are partially observable from the robot's egocentric view and computationally complex to process.\n\nWe present the first Social Dynamics Adaptation model (SDA) based on the robot's state-action history to infer the social dynamics. We propose a two-stage Reinforcement Learning framework: the first learns to encode the human trajectories into social dynamics and learns a motion policy conditioned on this encoded information, the current status, and the previous action. Here, the trajectories are fully visible, i.e., assumed as privileged information. In the second stage, the trained policy operates without direct access to trajectories. Instead, the model infers the social dynamics solely from the history of previous actions and statuses in real-time.\nTested on the novel Habitat 3.0 platform, SDA sets a novel state-of-the-art (SotA) performance in finding and following humans. \n\nThe code will be released upon acceptance.", "keywords": [ "Embodied AI", "Social Navigation", "Human Trajectories" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 4, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "M7KyLjuN0A", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Haozhe Xie", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Hengwei Bian", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Liang Pan", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Research Scientist " }, { "name": "Lingdong Kong", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Ziwei Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 10, "n_ref_uni": 45, "n_ref": 86, "n_ref_all": 102, "n_fig": 10, "n_tab": 19, "L_tab": 4007, "n_element_tab": 349, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 938, "n_element_tab_1": 76, "formula_len_all": 1562, "formula_len_all_1": 891, "len_all": 172942, "len_all_1": 44568, "len_abs": 0, "len_title": 186, "len_sents": 49988, "len_sents_1": 19610, "n_sents": 417, "n_sents_1": 141, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 61, "L_abs": 1813, "title": "DynamicCity: Large-Scale LiDAR Generation from Dynamic Scenes", "abs": "LiDAR scene generation has been developing rapidly recently. However, existing methods primarily focus on generating static and single-frame scenes, overlooking the inherently dynamic nature of real-world driving environments. In this work, we introduce DynamicCity, a novel 4D occupancy generation framework capable of generating large-scale, high-quality dynamic LiDAR scenes with semantics. DynamicCity mainly consists of two key models. **1)** A VAE model for learning HexPlane as the compact 4D representation. Instead of using naive averaging operations, DynamicCity employs a novel **Projection Module** to effectively compress 4D LiDAR features into six 2D feature maps for HexPlane construction, which significantly enhances HexPlane fitting quality (up to **12.56** mIoU gain). Furthermore, we utilize an Expansion & Squeeze Strategy to reconstruct 3D feature volumes in parallel, which improves both network training efficiency and reconstruction accuracy than naively querying each 3D point (up to **7.05** mIoU gain, **2.06x** training speedup, and **70.84\\%** memory reduction). **2)** A DiT-based diffusion model for HexPlane generation. To make HexPlane feasible for DiT generation, a **Padded Rollout Operation** is proposed to reorganize all six feature planes of the HexPlane as a squared 2D feature map. In particular, various conditions could be introduced in the diffusion or sampling process, supporting **versatile 4D generation applications**, such as trajectory- and command-driven generation, inpainting, and layout-conditioned generation. Extensive experiments on the CarlaSC and Waymo datasets demonstrate that DynamicCity significantly outperforms existing state-of-the-art 4D LiDAR generation methods across multiple metrics. The code will be released to facilitate future research.", "keywords": [ "LiDAR Generation", "Dynamic Scenes", "4D Generation" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 4, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "M7CblLwJB8", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Gracjan G\u00f3ral", "gender": "Male", "institution": "University of Warsaw", "country": "PL", "position": "PhD student" }, { "name": "Jaroslaw Kochanowicz", "gender": "Male", "institution": "deepsense.ai", "country": "PL", "position": "Principal Researcher" }, { "name": "Konrad Szewczyk", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "MS student" }, { "name": "Krzysztof Dziedzic", "gender": "Male", "institution": "Warsaw University of Technology", "country": "PL", "position": "Undergrad student" }, { "name": "Mateusz Olko", "gender": "Male", "institution": "University of Warsaw", "country": "PL", "position": "PhD student" }, { "name": "Piotr Mi\u0142o\u015b", "gender": "unknown", "institution": "IDEAS NCBR", "country": "PL", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 16, "n_ref": 22, "n_ref_all": 38, "n_fig": 4, "n_tab": 4, "L_tab": 4334, "n_element_tab": 68, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 495, "n_element_tab_1": 18, "formula_len_all": 227, "formula_len_all_1": 127, "len_all": 92166, "len_all_1": 72647, "len_abs": 1454, "len_title": 142, "len_sents": 30818, "len_sents_1": 24553, "n_sents": 251, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1167, "title": "AutoCustomization: A Unified Framework for Effortless, Selective LLM Bias and Style Finetuning", "abs": "Large language models are transforming the landscape of applications, with their influence poised to expand. One important practical challenge is how to selectively customize models to align with specific expectations, such as tone, formality, or underlying biases. To solve this task, we develop AutoCustomization. The key to our approach is leveraging the vast knowledge encoded in modern language models to construct fine-tuning datasets focused on a specific customization axis in contrast to prior methods, which depend primarily on tediously constructed libraries of prompts. AutoCustomization demonstrates several desirable properties. It is universally applicable to any bias axis (e.g., political, stylistic). It is efficient with small automatically generated datasets and short fine-tuning. It allows for precise monitoring of the resulting bias change with our BiasShift evaluation metric proven to be alligned with human perception, generalizable to held-out aspects, and selective in preserving other model capabilities. We verify AutoCustomization through human evaluation and show that it outperforms existing prompting techniques while being simpler.", "keywords": [ "large language models", "model customization" ], "rating_list": [ 3, 1, 3, 3, 3 ], "soundness_list": [ 1, 1, 1, 2, 3 ], "presentation_list": [ 2, 2, 2, 2, 3 ], "contribution_list": [ 1, 1, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "M5u38Os65F", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Avetik Karagulyan", "gender": "Male", "institution": "Centrale Sup\u00e9lec", "country": "FR", "position": "Researcher" }, { "name": "Hanmin Li", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Peter Richtarik", "gender": "Male", "institution": "King Abdullah University of Science and Technology (KAUST)", "country": "SA", "position": "Full Professor" } ], "n_formula": 299, "n_formula_1": 42, "n_ref_uni": 65, "n_ref": 134, "n_ref_all": 190, "n_fig": 13, "n_tab": 1, "L_tab": 3436, "n_element_tab": 40, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 277, "n_element_tab_1": 3, "formula_len_all": 33116, "formula_len_all_1": 2521, "len_all": 321741, "len_all_1": 63788, "len_abs": 916, "len_title": 123, "len_sents": 79971, "len_sents_1": 24849, "n_sents": 917, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 870, "title": "Variance Reduced Distributed Non-Convex Optimization Using Matrix Stepsizes", "abs": "Matrix-stepsized gradient descent algorithms have been shown to have superior performance in non-convex optimization problems compared to their scalar counterparts. The det-CGD algorithm, as introduced by Li et al. (2023), leverages matrix stepsizes to perform compressed gradient descent for non-convex objectives and matrix smooth problems in a federated manner. The authors establish the algorithm\u2019s convergence to a neighborhood of a weighted stationarity point under a convex condition for the symmetric and positive-definite matrix stepsize. In this paper, we propose two variance-reduced versions of the det-CGD algorithm, incorporating MARINA and DASHA methods. Notably, we establish theoretically and empirically, that det-MARINA and det-DASHA outperform MARINA, DASHA and the distributed det-CGD algorithms in terms of iteration and communication complexities.", "keywords": [ "Federated Learning", "Non-Convex Optimization", "Optimization" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "M5t0WvjfCg", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Alois Knoll", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Full Professor" }, { "name": "Mubarak Shah", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Salman Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Associate Professor" }, { "name": "Yuning Cui", "gender": "Male", "institution": "Technical University of Munich", "country": "DE", "position": "PhD student" }, { "name": "Fahad Shahbaz Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" }, { "name": "Syed Waqas Zamir", "gender": "Male", "institution": "Inception Institute of Artificial Intelligence", "country": "AE", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 10, "n_ref_uni": 54, "n_ref": 139, "n_ref_all": 191, "n_fig": 12, "n_tab": 9, "L_tab": 2910, "n_element_tab": 180, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 1626, "n_element_tab_1": 249, "formula_len_all": 727, "formula_len_all_1": 420, "len_all": 170877, "len_all_1": 68803, "len_abs": 2106, "len_title": 128, "len_sents": 42242, "len_sents_1": 28858, "n_sents": 306, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1949, "title": "AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation", "abs": "In the image acquisition process, various forms of degradation, including noise, blur, haze, and rain, are frequently introduced. These degradations typically arise from the inherent limitations of cameras or unfavorable ambient conditions. To recover clean images from their degraded versions, numerous specialized restoration methods have been developed, each targeting a specific type of degradation. Recently, all-in-one algorithms have garnered significant attention by addressing different types of degradations within a single model without requiring the prior information of the input degradation type. However, most methods purely operate in the spatial domain and do not delve into the distinct frequency variations inherent to different degradation types. To address this gap, we propose an adaptive all-in-one image restoration network based on frequency mining and modulation. Our approach is motivated by the observation that different degradation types impact the image content on different frequency subbands, thereby requiring different treatments for each restoration task. Specifically, we first mine low- and high-frequency information from the input features, guided by the adaptively decoupled spectra of the degraded image. The extracted features are then modulated by a bidirectional operator to facilitate interactions between different frequency components. Finally, the modulated features are merged into the original input for a progressively guided restoration. With this approach, the model achieves adaptive reconstruction by accentuating the informative frequency subbands according to different input degradations. Extensive experiments demonstrate that the proposed method, named AdaIR, achieves state-of-the-art performance on different image restoration tasks, including image denoising, dehazing, deraining, motion deblurring, and low-light image enhancement. Our code and models will be made publicly available.", "keywords": [ "All-in-one image restoration", "frequency mining", "frequency modulation" ], "rating_list": [ 5, 5, 8, 6 ], "soundness_list": [ 2, 2, 4, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "M5LGyR71yS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Alisia Maria Lupidi", "gender": "Female", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Carlos Gemmell", "gender": "Male", "institution": "University of Glasgow", "country": "", "position": "PhD student" }, { "name": "Jakob Nicolaus Foerster", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Jason E Weston", "gender": "Not Specified", "institution": "New York University", "country": "US", "position": "Researcher" }, { "name": "Maria Lomeli", "gender": "Female", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Nicola Cancedda", "gender": "Male", "institution": "Meta", "country": "GB", "position": "Researcher" }, { "name": "Jane Yu", "gender": "Female", "institution": "Meta AI ", "country": "US", "position": "Researcher" }, { "name": "Roberta Raileanu", "gender": "Female", "institution": "Facebook", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 25, "n_ref": 40, "n_ref_all": 72, "n_fig": 17, "n_tab": 9, "L_tab": 1913, "n_element_tab": 141, "n_fig_1": 15, "n_tab_1": 4, "L_tab_1": 789, "n_element_tab_1": 24, "formula_len_all": 272, "formula_len_all_1": 34, "len_all": 139718, "len_all_1": 61957, "len_abs": 915, "len_title": 130, "len_sents": 37171, "len_sents_1": 29058, "n_sents": 294, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 920, "title": "Source2Synth: Synthetic Data Generation and Curation Grounded in Real Data Sources", "abs": "Large Language Models still struggle in challenging scenarios that leverage structured data, complex reasoning, or tool usage. In this paper, we propose Source2Synth: a new self-augmentation approach for teaching LLMs new skills that can be leveraged in low data regimes without relying on costly human annotations. Source2Synth takes as input a custom data source and produces synthetic data points with intermediate reasoning steps grounded in real-world sources.\nSource2Synth improves the dataset quality by discarding low-quality generations based on their answerability.\nWe demonstrate the generality of this approach by applying it to two challenging domains: we test reasoning abilities in multi-hop question answering (MHQA), and tool usage in tabular question answering (TQA).\nOur method improves performance by 25.51\\% for TQA on WikiSQL and 22.57\\% for MHQA on HotPotQA compared to the fine-tuned baselines.", "keywords": [ "large language models", "llms", "synthetic-generation", "dataset-generation", "real-world data" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 1, 2, 1 ], "contribution_list": [ 2, 2, 3, 4 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "M4qNIzQYpd", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Chaoyun Zhang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Dan Pei", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Dongmei Zhang", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Distinguished Scientist" }, { "name": "Junjielong Xu", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "PhD student" }, { "name": "Pinjia He", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Qi Zhang", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Qinan Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Undergrad student" }, { "name": "Qingwei Lin", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Sr. Principal Researcher" }, { "name": "Shilin He", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" }, { "name": "Zhiqing Zhong", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 26, "n_ref": 61, "n_ref_all": 86, "n_fig": 9, "n_tab": 9, "L_tab": 3755, "n_element_tab": 418, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 682, "n_element_tab_1": 64, "formula_len_all": 82, "formula_len_all_1": 5, "len_all": 177766, "len_all_1": 70551, "len_abs": 1487, "len_title": 127, "len_sents": 53622, "len_sents_1": 31368, "n_sents": 419, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 78, "L_abs": 1235, "title": "OpenRCA: Can Large Language Models Locate the Root Cause of Software Failures?", "abs": "Large language models (LLMs) are driving substantial advancements in software engineering, with successful applications like Copilot and Cursor transforming real-world development practices. However, current research predominantly focuses on the early stages of development, such as code generation, while overlooking the post-development phases that are crucial to user experience. To explore the potential of LLMs in this direction, we propose OpenRCA, a benchmark dataset and evaluation framework for assessing LLMs\u2019 ability to identify the root cause of software failures. OpenRCA includes 335 failures from three enterprise software systems, along with over 68 GB of telemetry data (logs, metrics, and traces). Given a failure case and its associated telemetry, the LLM is tasked to identify the root cause that triggered the failure, requiring comprehension of software dependencies and reasoning over heterogeneous, long-context telemetry data. Our results show substantial room for improvement, as current models can only handle the simplest cases. Even with the specially designed RCA-agent, the best-performing model, Claude 3.5, solved only 11.34% failure cases. Our work paves the way for future research in this direction.", "keywords": [ "Language models", "Natural language processing", "Software engineering" ], "rating_list": [ 5, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "M4fhjfGAsZ", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Stefan Feuerriegel", "gender": "Male", "institution": "LMU Munich", "country": "DE", "position": "Professor" }, { "name": "Yilmazcan Ozyurt", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "PhD student" }, { "name": "MRINMAYA SACHAN", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 68, "n_ref": 200, "n_ref_all": 232, "n_fig": 17, "n_tab": 7, "L_tab": 3799, "n_element_tab": 417, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1544, "n_element_tab_1": 146, "formula_len_all": 843, "formula_len_all_1": 928, "len_all": 250174, "len_all_1": 68435, "len_abs": 1322, "len_title": 145, "len_sents": 75278, "len_sents_1": 29306, "n_sents": 638, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1332, "title": "Automated Knowledge Concept Annotation and Question Representation Learning for Knowledge Tracing", "abs": "Knowledge tracing (KT) is a popular approach for modeling students' learning progress over time, which can enable more personalized and adaptive learning. However, existing KT approaches face two major limitations: (1) they rely heavily on expert-defined knowledge concepts (KCs) in questions, which is time-consuming and prone to errors; and (2) KT methods tend to overlook the semantics of both questions and the given KCs. In this work, we address these challenges and present KCQRL, a framework for automated knowledge concept annotation and question representation learning that can improve the effectiveness of any existing KT model. First, we propose an automated KC annotation process using large language models (LLMs), which generates question solutions and then annotates KCs in each solution step of the questions. Second, we introduce a contrastive learning approach to generate semantically rich embeddings for questions and solution steps, aligning them with their associated KCs via a tailored false negative elimination approach. These embeddings can be readily integrated into existing KT models, replacing their randomly initialized embeddings. We demonstrate the effectiveness of KCQRL across 15 KT algorithms on two large real-world Math learning datasets, where we achieve consistent performance improvements.", "keywords": [ "knowledge tracing", "time series prediction", "knowledge concept", "KC annotation", "representation learning", "contrastive learning", "large language models" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 1, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 1, 3 ], "confidence_list": [ 2, 5, 4 ] }, { "paper_id": "M4RhGr2lAy", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Chen Chen", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Assistant Professor" }, { "name": "Jing Ma", "gender": "Female", "institution": "Case Western Reserve University", "country": "US", "position": "Assistant Professor" }, { "name": "Jundong Li", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" }, { "name": "Song Wang", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Yushun Dong", "gender": "Male", "institution": "Florida State University", "country": "US", "position": "Assistant Professor" }, { "name": "Zaiyi Zheng", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "", "position": "PhD student" }, { "name": "Zhenyu Lei", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 80, "n_ref": 123, "n_ref_all": 145, "n_fig": 5, "n_tab": 3, "L_tab": 2438, "n_element_tab": 169, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 3653, "n_element_tab_1": 224, "formula_len_all": 900, "formula_len_all_1": 364, "len_all": 167932, "len_all_1": 72484, "len_abs": 1109, "len_title": 90, "len_sents": 39226, "len_sents_1": 31909, "n_sents": 275, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1114, "title": "Fairness-Aware Graph Learning: A Benchmark", "abs": "Fairness-aware graph learning has gained increasing attention in recent years. Nevertheless, there lacks a comprehensive benchmark to evaluate and compare different fairness-aware graph learning methods, which blocks practitioners from choosing appropriate ones for broader real-world applications. In this paper, we present an extensive benchmark on ten representative fairness-aware graph learning methods. Specifically, we design a systematic evaluation protocol and conduct experiments on seven real-world datasets to evaluate these methods from multiple perspectives, including group fairness, individual fairness, the balance between different fairness criteria, and computational efficiency. Our in-depth analysis reveals key insights into the strengths and limitations of existing methods. Additionally, we provide practical guidance for applying fairness-aware graph learning methods in applications. To the best of our knowledge, this work serves as an initial step towards comprehensively understanding representative fairness-aware graph learning methods to facilitate future advancements in this area.", "keywords": [ "Graph Learning Algorithms", "Algorithmic Fairness", "Performance Benchmark" ], "rating_list": [ 6, 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 4, 2, 2, 3, 4 ], "contribution_list": [ 3, 2, 1, 2, 2 ], "confidence_list": [ 5, 4, 5, 4, 3 ] }, { "paper_id": "M4J8OtcqT0", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Daizong Liu", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xiaoye Qu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yicheng Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yin Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yu Cheng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Zhi Ii", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 57, "n_ref_all": 73, "n_fig": 5, "n_tab": 4, "L_tab": 624, "n_element_tab": 123, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 573, "n_element_tab_1": 110, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 113690, "len_all_1": 58611, "len_abs": 1833, "len_title": 157, "len_sents": 30771, "len_sents_1": 27761, "n_sents": 229, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1844, "title": "Strategy-centric Synthesis: Connecting Billions of Image-Text Pairs to High-Quality Visual Instruction Data", "abs": "Vision-Language Models (VLMs) have demonstrated remarkable generalization across tasks by aligning visual and linguistic representations. High-quality visual instruction data is critical for enhancing the performance of Vision-Language Models. However, current visual instruction tuning datasets, which are primarily derived from past visual tasks, have several limitations. For instance, the range of question types is often restricted and closely tied to the original visual tasks. Furthermore, image diversity is limited, as images collected for various specialized vision tasks clearly fail to adequately represent real-world user queries. Additionally, previous instruction datasets tend to lack complexity, focusing on single tasks like captioning or OCR, which makes it challenging to train models for more complex, multi-skill scenarios. To address these limitations, we propose a novel paradigms called strategy-centric synthesis: automatically synthesizing high-quality instruction data from large-scale image-text pairs. First, we employ an efficient heuristic method to select high-quality, complex images from DataComp-1B image-text pairs. Carefully crafted prompts and these images are fed to VLMs to extract high-quality query strategies and generate corresponding image descriptions. These descriptions are subsequently used to retrieve images aligned with specific questioning strategies. Finally, the retrieved images and their matching strategies are used to synthesize high-quality instructional data. Our experiments indicate that with continued instruction fine-tuning via LoRA on only 3,000 newly synthesized data samples, 0.45\\% of the LLAVA-1.5 instruction tuning dataset, the model significantly outperforms the original LLAVA-1.5-7B across multiple benchmarks, thereby demonstrating the effectiveness of our approach.", "keywords": [ "Visual Instruction Tuning", "Strategy", "Synthetic Data", "Complex", "Diverse", "Scalable" ], "rating_list": [ 5, 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 3, 1, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 5, 5 ] }, { "paper_id": "M42KR4W9P5", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Junqi You", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Xiaosong Jia", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "zhiyuan zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 38, "n_ref": 114, "n_ref_all": 134, "n_fig": 7, "n_tab": 13, "L_tab": 4215, "n_element_tab": 412, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 2444, "n_element_tab_1": 237, "formula_len_all": 1122, "formula_len_all_1": 1002, "len_all": 161869, "len_all_1": 56334, "len_abs": 1536, "len_title": 128, "len_sents": 36843, "len_sents_1": 23262, "n_sents": 271, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1545, "title": "DriveTransformer: Unified Transformer for Scalable End-to-End Autonomous Driving", "abs": "End-to-end autonomous driving (E2E-AD) has emerged as a trend in the field of autonomous driving, promising a data-driven, scalable approach to system design. However, existing E2E-AD methods usually adopt the sequential paradigm of perception-prediction-planning, which leads to cumulative errors and training instability. The manual ordering of tasks also limits the system\u2019s ability to leverage synergies between tasks (for example, planning-aware perception and game-theoretic interactive prediction and planning). Moreover, the dense BEV representation adopted by existing methods brings computational challenges for long-range perception and long-term temporal fusion. To address these challenges, we present DriveTransformer, a simplified E2E-AD framework for the ease of scaling up, characterized by three key features: Task Parallelism (All agent, map, and planning queries direct interact with each other at each block), Sparse Representation (Task queries direct interact with raw sensor features), and Streaming Processing (Task queries are stored and passed as history information). As a result, the new framework is composed of three unified operations: task self-attention, sensor cross-attention, temporal cross-attention, which significantly reduces the complexity of system and leads to better training stability. DriveTransformer achieves state-of-the-art performance in both simulated closed-loop benchmark Bench2Drive and real world open-loop benchmark nuScenes with high FPS. We will open source our code and checkpoints.", "keywords": [ "end-to-end autonomous driving" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "M3y2msIfHZ", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Hamed Karimi", "gender": "Male", "institution": "Boston College", "country": "", "position": "PhD student" }, { "name": "Stefano Anzellotti", "gender": "Male", "institution": "Boston College", "country": "", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 23, "n_ref": 58, "n_ref_all": 74, "n_fig": 3, "n_tab": 2, "L_tab": 1301, "n_element_tab": 63, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 798, "n_element_tab_1": 55, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 99250, "len_all_1": 50688, "len_abs": 1249, "len_title": 204, "len_sents": 36295, "len_sents_1": 25058, "n_sents": 221, "n_sents_1": 148, "venue": "ICLR2025", "decision": "Reject", "L_title": 154, "L_abs": 1254, "title": "Visual Representations in Humans and Machines: A Comparative Analysis of Artificial and Biological Neural Responses to Naturalistic Dynamic Visual Stimuli", "abs": "Visual representations in the human brain are shaped by the pressure to support planning and interactions with the environment. Do visual representations in deep network models converge with visual representations in humans? Here, we investigate this question for a new class of effective self-supervised models: Masked Autoencoders (MAEs). We compare image MAEs and video MAEs to neural responses in humans as well as convolutional neural networks. The results reveal that representations learned by MAEs diverge from neural representations in humans and convolutional neural networks. Fine-tuning MAEs with a supervised task improves their correspondence with neural responses but is not sufficient to bridge the gap that separates them from supervised convolutional networks. Finally, video MAEs show closer correspondence to neural representations than image MAEs, revealing an important role of temporal information. However, convolutional networks based on optic flow show a closer correspondence to neural responses in humans than even video MAEs, indicating that while masked autoencoding yields visual representations that are effective at multiple downstream tasks, it is not sufficient to learn representations that converge with human vision.", "keywords": [ "self-supervised learning", "visual representation", "occipitotemporal cortex", "human vision", "masked autoencoders" ], "rating_list": [ 5, 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "M3kBtqpys5", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Guoqing Liu", "gender": "Female", "institution": "Shanxi University", "country": "CN", "position": "PhD student" }, { "name": "Pinhan Fu", "gender": "Male", "institution": "Shanxi University", "country": "CN", "position": "MS student" }, { "name": "Qian Guo", "gender": "Female", "institution": "Taiyuan University of Science and Technology", "country": "CN", "position": "Lecturer" }, { "name": "Xinyan Liang", "gender": "Male", "institution": "Shanxi University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuhua Qian", "gender": "Male", "institution": "Shanxi University", "country": "CN", "position": "Full Professor" } ], "n_formula": 49, "n_formula_1": 22, "n_ref_uni": 41, "n_ref": 89, "n_ref_all": 120, "n_fig": 7, "n_tab": 13, "L_tab": 6341, "n_element_tab": 474, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 3076, "n_element_tab_1": 278, "formula_len_all": 2922, "formula_len_all_1": 1705, "len_all": 200910, "len_all_1": 67962, "len_abs": 322, "len_title": 117, "len_sents": 57222, "len_sents_1": 27581, "n_sents": 643, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1964, "title": "Trusted Multi-View Classification via Evolutionary Multi-View Fusion", "abs": "Multi-view classification methodologies grounded in the Dempster-Shafer theory, renowned for their reliability in decision-making, have garnered significant application across various safety-critical domains due to their capacity to provide a degree of trustworthiness for each view. However, the adoption of a late fusion strategy by these methodologies constrains the interaction of information among views, thereby leading to suboptimal utilization of multi-view data. A recent advancement aimed at mitigating this limitation involves the generation of a pseudo view by concatenating all individual views. Nonetheless, the effectiveness of this pseudo view may be compromised when incorporating underperforming views, such as those afflicted by noise. Furthermore, the integration of a pseudo view exacerbates the issue of imbalanced multi-view learning, as it contains a disproportionate amount of information compared to individual views. To address these multifaceted challenges, we propose an approach termed Enhancing Trusted multi-view classification via Evolutionary multi-view Fusion (TEF). Specifically, we introduce an evolutionary multi-view architecture search method to generate a high-quality fusion architecture serving as the pseudo view, thus enabling adaptive selection of views and fusion operators. Subsequently, each view within the fusion architecture is enhanced by concatenating the decision output of the fusion architecture with its respective view. Our experimental findings underscore the efficacy of this straightforward yet potent strategy in mitigating the imbalanced multi-view learning problem, consequently enhancing TEF's performance, particularly on complex many-view datasets featuring more than three views compared to its counterparts. Comprehensive experimental evaluations conducted on six multi-view datasets corroborate the superior performance of our proposed method over other trusted multi-view learning approaches.", "keywords": [ "Trusted multi-view classification", "evolutionary multi-view fusion", "multi-view learning" ], "rating_list": [ 6, 3, 8, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 4, 4 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "M31bgolPnl", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 10, "author_info_list": [ { "name": "Bo Zheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Gengru Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Shaopan Xiong", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Weixun Wang", "gender": "unknown", "institution": "Tianjin University", "country": "", "position": "PhD student" }, { "name": "Wenbo Su", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xiaoyang Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xingyao Zhang", "gender": "unknown", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Xingyuan Bu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yanshi li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yingshui Tan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 17, "n_ref_uni": 39, "n_ref": 216, "n_ref_all": 226, "n_fig": 5, "n_tab": 1, "L_tab": 1132, "n_element_tab": 42, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 731, "n_element_tab_1": 101, "formula_len_all": 1493, "formula_len_all_1": 1174, "len_all": 146732, "len_all_1": 59352, "len_abs": 1582, "len_title": 137, "len_sents": 38457, "len_sents_1": 26099, "n_sents": 341, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1594, "title": "Adaptive dense reward:Understanding the Gap Between Action and Reward Space in Alignment", "abs": "Reinforcement Learning from Human Feedback (RLHF) has proven highly effective in aligning Large Language Models (LLMs) with human preferences. However, the original RLHF typically optimizes under an overall reward, which can lead to a suboptimal learning process. This limitation stems from RLHF's lack of awareness regarding which specific tokens should be reinforced or suppressed. Moreover, conflicts in supervision can arise, for instance, when a chosen response includes erroneous tokens, while a rejected response contains accurate elements. To rectify these shortcomings, increasing dense reward methods, such as step-wise and token-wise RLHF, have been proposed. However, these existing methods are limited to specific tasks (like mathematics). In this paper, we propose the \"Adaptive Message-wise RLHF\" method, which robustly applies to various tasks. By defining pivot tokens as key indicators, our approach adaptively identifies essential information and converts sequence-level supervision into fine-grained, subsequence-level supervision. This aligns the density of rewards and action spaces more closely with the information density of the input. Experiments demonstrate that our method can be integrated into various training methods, significantly mitigating hallucinations and catastrophic forgetting problems, while outperforming other methods on multiple evaluation metrics. Our method improves the success rate on adversarial samples by 10\\% compared to the sample-wise approach, and achieves a 1.3\\% improvement on evaluation benchmarks such as MMLU, GSM8K, HumanEval, etc.", "keywords": [ "Large Language Model; Reinforcement Learning from Human Feedback (RLHF); Adaptive Message-wise RLHF" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 1, 1, 1 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "M2g647Femt", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Brooks Paige", "gender": "Male", "institution": "University College London", "country": "", "position": "Associate Professor" }, { "name": "Phillip Williams", "gender": "Male", "institution": "Kinaxis inc.", "country": "", "position": "Software Developper" } ], "n_formula": 38, "n_formula_1": 18, "n_ref_uni": 29, "n_ref": 45, "n_ref_all": 59, "n_fig": 7, "n_tab": 12, "L_tab": 2481, "n_element_tab": 189, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 341, "n_element_tab_1": 47, "formula_len_all": 3441, "formula_len_all_1": 1730, "len_all": 156370, "len_all_1": 55680, "len_abs": 1131, "len_title": 103, "len_sents": 49971, "len_sents_1": 25187, "n_sents": 387, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1138, "title": "Solving Composable Constraints for Inverse Design Tasks", "abs": "Inverse design tasks are an important category of problem in which we want to identify some input vector $x$ satisfying some desirable properties. In this paper we propose a mechanism for representing inequality constraints as Signed Distance Functions (SDFs). SDFs permit efficient projection of points into the solution region as well as providing a mechanism for composing constraints via boolean set operations. In this paper, we provide theoretical motivation for Signed Distance Functions (SDFs) as an implicit representation of inequality constraints. Next, we provide analysis demonstrating that SDFs can be used to efficiently project points into solution regions. Additionally, we propose two novel algorithms for computing SDFs for wide families of machine learning models. Finally, we demonstrate practical utility by performing conditional image generation using MNIST and CelebA datasets, and computational drug design using the ZINC-250K dataset. From the experimental results, we note that the composable constraints can reliably and efficiently compute solutions to complex inverse design tasks with deep learning models.", "keywords": [ "Inverse Design", "Conditional Generation", "Constrained Optimization", "Deep Learning", "Neural Networks", "Signed Distance Functions" ], "rating_list": [ 1, 5, 3, 5 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 2, 3, 4 ] }, { "paper_id": "M2YCdfxNVx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Jan Kautz", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "VP Research" }, { "name": "Jang Hyun Cho", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Intern" }, { "name": "Ligeng Zhu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Marco Pavone", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Director, Autonomous Vehicle Research" }, { "name": "Pavlo Molchanov", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Research Scientist" }, { "name": "Song Han", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yan Wang", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yao Lu", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Yunhao Fang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Hongxu Yin", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Senior Research Scientist" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 50, "n_ref": 115, "n_ref_all": 154, "n_fig": 7, "n_tab": 13, "L_tab": 4486, "n_element_tab": 381, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2025, "n_element_tab_1": 122, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 230879, "len_all_1": 70558, "len_abs": 1753, "len_title": 96, "len_sents": 54486, "len_sents_1": 32310, "n_sents": 445, "n_sents_1": 259, "venue": "ICLR2025", "decision": "withdraw", "L_title": 47, "L_abs": 1474, "title": "VILA^2: VLM Augmented VLM with Self-Improvement", "abs": "Visual language models (VLMs) have rapidly progressed, driven by the success of large language models (LLMs). While model architectures and training infrastructures advance rapidly, data curation remains under-explored. When data quantity and quality become a bottleneck, existing work either directly crawls more raw data from the Internet that does not have a guarantee of data quality or distills from black-box commercial models (e.g., GPT-4V / Gemini) causing the performance upper bounded by that model. In this work, we introduce a novel approach that includes a self-augment step and a specialist-augment step to iteratively improve data quality and model performance. In the self-augment step, a VLM recaptions its own pretraining data to enhance data quality, and then retrains from scratch using this refined dataset to improve model performance. This process can iterate for several rounds. Once self-augmentation saturates, we employ several specialist VLMs finetuned from the self-augmented VLM with domain-specific expertise, to further infuse specialist knowledge into the generalist VLM through task-oriented recaptioning and retraining. With the combined self-augmented and specialist-augmented training, we introduce VILA2 (VLM-augmented-VLM), a VLM family that consistently improves the accuracy on a wide range of tasks over prior art, including MMMU leaderboard, with a reusable pretraining dataset that is 300x more cost-efficient than human labeling.", "keywords": [ "Multimodal Large Models" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "M2SsqpxGtc", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Fabian Manhardt", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Federico Tombari", "gender": "Male", "institution": "Google", "country": "CH", "position": "Research Director" }, { "name": "Konrad Schindler", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Professor" }, { "name": "Michael Oechsle", "gender": "unknown", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Nikolai Kalischek", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Philipp Henzler", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 26, "n_ref": 66, "n_ref_all": 81, "n_fig": 7, "n_tab": 1, "L_tab": 441, "n_element_tab": 89, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 270, "formula_len_all_1": 153, "len_all": 104673, "len_all_1": 54434, "len_abs": 931, "len_title": 122, "len_sents": 29348, "len_sents_1": 26130, "n_sents": 219, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 74, "L_abs": 853, "title": "CubeDiff: Repurposing Diffusion-Based Image Models for Panorama Generation", "abs": "We introduce a novel method for generating 360\u00b0 panoramas from text prompts or images. Our approach leverages recent advances in 3D generation by employing multi-view diffusion models to jointly synthesize the six faces of a cubemap. Unlike previous methods that rely on processing equirectangular projections or autoregressive generation, our method treats each face as a standard perspective image, simplifying the generation process and enabling the use of existing multi-view diffusion models. We demonstrate that these models can be adapted to produce high-quality cubemaps without requiring correspondence-aware attention layers. Our model allows for fine-grained text control, generates high resolution panorama images and generalizes well beyond its training set, whilst achieving state-of-the-art results, both qualitatively and quantitatively.", "keywords": [ "panorama generation", "diffusion", "multi-view" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "M2NFWRPMUd", "primary_area": "applications to robotics, autonomy, planning", "n_author": 13, "author_info_list": [ { "name": "Apoorva Sharma", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Boris Ivanovic", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Boyi Li", "gender": "Female", "institution": "NVIDIA Research", "country": "US", "position": "Researcher" }, { "name": "Danfei Xu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Jiahui Huang", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Jiawei Yang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Marco Pavone", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Director, Autonomous Vehicle Research" }, { "name": "Maximilian Igl", "gender": "Male", "institution": "NVIDIA", "country": "CH", "position": "Researcher" }, { "name": "Peter Karkus", "gender": "Male", "institution": "NVIDIA", "country": "CH", "position": "Researcher" }, { "name": "Yan Wang", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yue Wang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Yurong You", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Yuxiao Chen", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 62, "n_ref": 144, "n_ref_all": 159, "n_fig": 6, "n_tab": 4, "L_tab": 1445, "n_element_tab": 213, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1241, "n_element_tab_1": 205, "formula_len_all": 540, "formula_len_all_1": 385, "len_all": 204557, "len_all_1": 70752, "len_abs": 1458, "len_title": 122, "len_sents": 50708, "len_sents_1": 32086, "n_sents": 389, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1465, "title": "STORM: Spatio-TempOral Reconstruction Model For Large-Scale Outdoor Scenes", "abs": "We present STORM, a spatio-temporal reconstruction model designed to reconstruct in-the-wild dynamic outdoor scenes from sparse observations. Existing dynamic reconstruction methods rely heavily on dense observations across space and time and strong motion supervision, therefore suffering from lengthy optimization time, limited generalizability to novel views or scenes, and degenerated quality caused by noisy pseudo-labels. To bridge the gap, STORM introduces a data-driven Transformer architecture that jointly infers 3D scenes and their dynamics in a single forward pass. A key design of our scene representation is to aggregate 3D Gaussians and their motion predicted from all frames, which are later transformed to the target timestep for a more complete (i.e. \u201camodal\u201d) reconstruction at any given time from any viewpoint. As an emergent property, STORM can automatically capture dynamic instances and their high-quality masks using just the reconstruction loss. Extensive experiments show that STORM accurately reconstructs dynamic scenes and outperforms other per-scene optimization (+3.7 PSNR) or feed-forward approaches (+1.5 PSNR), it can reconstruct large-scale outdoor scenes within just 200ms and render in real-time. Beyond reconstruction, we qualitatively demonstrate four additional applications of our model, demonstrating the potential of self-supervised learning for advancing dynamic scene understanding. Our code and model will be released.", "keywords": [ "autonomous driving; reconstruction model; spatiotemporal" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "M2MinWsyjC", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Kevin Course", "gender": "Male", "institution": "Toronto University", "country": "", "position": "PhD student" }, { "name": "Prasanth B. Nair", "gender": "Male", "institution": "Toronto University", "country": "CA", "position": "Full Professor" }, { "name": "Srinath Dama", "gender": "unknown", "institution": "University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 47, "n_formula_1": 20, "n_ref_uni": 46, "n_ref": 117, "n_ref_all": 156, "n_fig": 8, "n_tab": 12, "L_tab": 2747, "n_element_tab": 524, "n_fig_1": 5, "n_tab_1": 9, "L_tab_1": 1316, "n_element_tab_1": 76, "formula_len_all": 4489, "formula_len_all_1": 1783, "len_all": 181526, "len_all_1": 66582, "len_abs": 1017, "len_title": 119, "len_sents": 52062, "len_sents_1": 29266, "n_sents": 362, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1049, "title": "SHIFTING TIME: TIME-SERIES FORECASTING WITH KHATRI-RAO NEURAL OPERATORS", "abs": "We present an operator-theoretic framework for time-series forecasting that involves learning a continuous time-shift operator associated with temporal and spatio-temporal problems. The time-shift operator learning paradigm offers a continuous relaxation of the discrete lag factor used in traditional autoregressive models enabling the history of a function up to a given time to be mapped to its future values. To parametrize the operator learning problem, we propose Khatri-Rao neural operators -- a new architecture for defining non-stationary integral transforms which achieves almost linear cost on spatial and spatio-temporal problems. From a practical perspective, the advancements made in this work allow us to handle irregularly sampled observations and forecast at super-resolution in both space and time. Detailed numerical studies across a wide range of temporal and spatio-temporal benchmark problems suggest that the proposed approach is highly scalable and provides results that compares favourably with the state-of-the-art methods.", "keywords": [ "time-series modeling", "spatio-temporal modeling", "time-shift operator", "Khatri-Rao neural operator", "neural operator", "operator learning" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "M29nUGozPa", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Bin Feng", "gender": "Male", "institution": "International Digital Economy Academy, International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Kangjie Zheng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Ming Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Siyue Liang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Wei Ju", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "Associate Professor" }, { "name": "Zequn Liu", "gender": "Female", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Zhiping Xiao", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "Postdoc" }, { "name": "junwei yang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD Student" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 29, "n_ref": 59, "n_ref_all": 105, "n_fig": 16, "n_tab": 15, "L_tab": 5375, "n_element_tab": 712, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 446, "n_element_tab_1": 74, "formula_len_all": 525, "formula_len_all_1": 253, "len_all": 197709, "len_all_1": 56197, "len_abs": 1521, "len_title": 124, "len_sents": 69067, "len_sents_1": 27609, "n_sents": 505, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1475, "title": "SMI-Editor: Edit-based SMILES Language Model with Fragment-level Supervision", "abs": "SMILES, a crucial textual representation of molecular structures, has garnered significant attention as a foundation for pre-trained language models (LMs). However, most existing pre-trained SMILES LMs focus solely on the single-token level supervision during pre-training, failing to fully leverage the substructural information of molecules. This limitation makes the pre-training task overly simplistic, preventing the models from capturing richer molecular semantic information. Moreover, during pre-training, these SMILES LMs only process corrupted SMILES inputs, never encountering any valid SMILES, which leads to a train-inference mismatch. To address these challenges, we propose SMI-Editor, a novel edit-based pre-trained SMILES LM. SMI-Editor disrupts substructures within a molecule at random and feeds the resulting SMILES back into the model, which then attempts to restore the original SMILES through an editing process. This approach not only introduces fragment-level training signals, but also enables the use of valid SMILES as inputs, allowing the model to learn how to reconstruct complete molecules from these incomplete structures. As a result, the model demonstrates improved scalability and an enhanced ability to capture fragment-level molecular information. Experimental results show that SMI-Editor achieves state-of-the-art performance across multiple downstream molecular tasks, and even outperforming several 3D molecular representation models.", "keywords": [ "SMILES Language Model", "SMILES Pre-training Model", "Molecular Pre-training Model" ], "rating_list": [ 3, 6, 8, 8, 5 ], "soundness_list": [ 2, 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "M23dTGWCZy", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Chenglei Si", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Diyi Yang", "gender": "Female", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Tatsunori Hashimoto", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 39, "n_ref": 79, "n_ref_all": 106, "n_fig": 3, "n_tab": 21, "L_tab": 2567, "n_element_tab": 230, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1246, "n_element_tab_1": 299, "formula_len_all": 24, "formula_len_all_1": 0, "len_all": 434466, "len_all_1": 68667, "len_abs": 3927, "len_title": 139, "len_sents": 183989, "len_sents_1": 31875, "n_sents": 1379, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1180, "title": "Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers", "abs": "Recent advancements in large language models (LLMs) have sparked optimism about their potential to accelerate scientific discovery, with a growing number of works proposing research agents that autonomously generate and validate new ideas. Despite this, no evaluations have shown that LLM systems can take the very first step of producing novel, expert-level ideas, let alone perform the entire research process. We address this by establishing an experimental design that evaluates research idea generation while controlling for confounders and performs the first comparison between expert NLP researchers and an LLM ideation agent. By recruiting over 100 NLP researchers to write novel ideas and blind reviews of both LLM and human ideas, we obtain the first statistically significant conclusion on current LLM capabilities for research ideation: we find LLM-generated ideas are judged as more novel (p < 0.05) than human expert ideas while being judged slightly weaker on feasibility. Studying our agent baselines closely, we identify open problems in building and evaluating research agents, including failures of LLM self-evaluation and their lack of diversity in generation.", "keywords": [ "LLM", "Idea Generation", "Human Study" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "M1y9JAL7CP", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Enlu Zhou", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "", "position": "Associate Professor" }, { "name": "Xiaoshuang Wang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yifan Lin", "gender": "Male", "institution": "C3 AI", "country": "US", "position": "Researcher" } ], "n_formula": 101, "n_formula_1": 26, "n_ref_uni": 34, "n_ref": 64, "n_ref_all": 87, "n_fig": 7, "n_tab": 6, "L_tab": 1158, "n_element_tab": 93, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 510, "n_element_tab_1": 45, "formula_len_all": 11418, "formula_len_all_1": 2580, "len_all": 179657, "len_all_1": 73723, "len_abs": 1291, "len_title": 161, "len_sents": 56292, "len_sents_1": 33230, "n_sents": 516, "n_sents_1": 247, "venue": "ICLR2025", "decision": "withdraw", "L_title": 112, "L_abs": 1345, "title": "Policy Gradient Optimization for Markov Decision Processes with Epistemic Uncertainty and General Loss Functions", "abs": "Motivated by many application problems, we consider Markov decision processes (MDPs) with a general loss function and unknown parameters. To mitigate the epistemic uncertainty associated with unknown parameters, we take a Bayesian approach to estimate the parameters from data and impose a coherent risk functional (with respect to the Bayesian posterior distribution) on the general loss function. Since this formulation usually does not satisfy the interchangeability principle, it does not admit Bellman equations and cannot be solved by approaches based on dynamic programming. Therefore, we develop a policy gradient optimization approach to address this problem. We utilize the dual representation of the coherent risk measure and extend the envelope theorem to derive the policy gradient. Our extension of the envelope theorem\nfrom the discrete case to the continuous case may be of independent interest. We then show the convergence of the proposed algorithm with a convergence rate of $\\mathcal{O}((1-\\epsilon)^t)$, where $t$ is the number of policy gradient iterations and $\\epsilon$ is the accuracy. We further extend our algorithm to an episodic setting, and establish the consistency of the extended algorithm and provide bounds on the number of iterations needed to achieve an error bound $\\mathcal{O}(\\epsilon)$ in each episode.", "keywords": [ "policy gradient", "Markov Decision Process", "epistemic uncertainty", "Bayesian approach", "general loss function", "convex RL" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "M1mL9tneGL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Hangzhou He", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "JiaKui Hu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Lei Zhu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Lujia Jin", "gender": "Male", "institution": "China Mobile Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Yanye Lu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhengjian Yao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 47, "n_ref": 90, "n_ref_all": 125, "n_fig": 17, "n_tab": 7, "L_tab": 2868, "n_element_tab": 353, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 742, "n_element_tab_1": 20, "formula_len_all": 552, "formula_len_all_1": 539, "len_all": 142801, "len_all_1": 54356, "len_abs": 106, "len_title": 111, "len_sents": 38128, "len_sents_1": 25006, "n_sents": 325, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1337, "title": "ART-FR: masked Auto-Regressive Transformer for Face Restoration", "abs": "Restoring authentic facial features from low-quality images presents an extremely challenging task, due to the intricate real-world degradations and the inherently ill-posed nature of the problem. Existing methods, which utilize a codebook prior, help alleviate the complexity of the restoration process and produce visually plausible outcomes. However, these methods struggle to accurately capture the mapping between low-quality (LQ) and high-quality (HQ) images in the discrete latent space, leading to suboptimal results. \nInspired by the success of auto-regressive generation paradigm in discrete modeling problems (e.g. large language models), we propose an Auto-Regressive Transformer based Face Restoration (ART-FR) method to mitigate this mapping challenge. Specifically, with the aid of a visual tokenizer, we reformulate the face restoration task as a conditional generation problem within the discrete latent space. Furthermore, a masked generative image transformer is employed to model the distribution of this latent space, conditioned on LQ features. Face restoration is subsequently performed in the latent space through iterative sampling, with the HQ image reconstructed using a pretrained decoder. Extensive experimental validation demonstrates ART-FR exhibits superior performance across various benchmark datasets.", "keywords": [ "Auto-regressive model; Face restoration" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "M1ZMwDqvSe", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Fei Huang", "gender": "Male", "institution": "Alibaba Group US", "country": "US", "position": "Senior Research Director" }, { "name": "Feiteng Mu", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "CN", "position": "PhD student" }, { "name": "Mengting Hu", "gender": "Female", "institution": "Nankai University", "country": "CN", "position": "Assistant Professor" }, { "name": "Pengjun Xie", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xinyu Wang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yong Jiang", "gender": "Male", "institution": "Tongyi Lab", "country": "CN", "position": "Researcher" }, { "name": "Zhuo Chen", "gender": "Male", "institution": "ShanghaiTech University", "country": "", "position": "MS student" }, { "name": "zhen zhang", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 20, "n_ref": 26, "n_ref_all": 49, "n_fig": 13, "n_tab": 4, "L_tab": 2793, "n_element_tab": 218, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 795, "n_element_tab_1": 73, "formula_len_all": 47, "formula_len_all_1": 47, "len_all": 114663, "len_all_1": 57368, "len_abs": 1423, "len_title": 126, "len_sents": 35842, "len_sents_1": 27492, "n_sents": 279, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1431, "title": "Exploring Knowledge Boundaries in Large Language Models for Retrieval Judgment", "abs": "Large Language Models (LLMs) are increasingly recognized for their practical applications. However, these models often encounter challenges in dynamically changing knowledge, as well as in managing unknown static knowledge. Retrieval-Augmented Generation (RAG) tackles this challenge and has shown a significant impact on LLMs.\nActually, we find that the impact of RAG on the question answering capabilities of LLMs can be categorized into three groups: beneficial, neutral, and harmful. By minimizing retrieval requests that yield neutral or harmful results, we can effectively reduce both time and computational costs, while also improving the overall performance of LLMs.\nThis insight motivates us to differentiate between types of questions using certain metrics as indicators, to decrease the retrieval ratio without compromising performance. \nIn our work, we propose a method that is able to identify different types of questions from this view by training a Knowledge Boundary Model (KBM). \nExperiments conducted on 11 English and Chinese datasets illustrate that the KBM effectively delineates the knowledge boundary, significantly decreasing the proportion of retrievals required for optimal end-to-end performance. Specifically, we evaluate the effectiveness of KBM in three complex scenarios: dynamic knowledge, long-tail static knowledge, and multi-hop problems, as well as its functionality as an external LLM plug-in.", "keywords": [ "knowledge boundary", "large language models", "Retrieval-Augmented Generation" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "M1CCA6UF0y", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Anirudh Goyal", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Dingli Yu", "gender": "unknown", "institution": "Microsoft ", "country": "US", "position": "Researcher" }, { "name": "Kaifeng Lyu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Michael C Mozer", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Nan Rosemary Ke", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Sanjeev Arora", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "Full Professor" }, { "name": "Simon Park", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "MS student" }, { "name": "Vedant Shah", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "MS student" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" } ], "n_formula": 23, "n_formula_1": 2, "n_ref_uni": 28, "n_ref": 62, "n_ref_all": 90, "n_fig": 4, "n_tab": 8, "L_tab": 3050, "n_element_tab": 140, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1914, "n_element_tab_1": 67, "formula_len_all": 1735, "formula_len_all_1": 633, "len_all": 251007, "len_all_1": 70870, "len_abs": 1948, "len_title": 99, "len_sents": 93758, "len_sents_1": 32550, "n_sents": 729, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1982, "title": "AI-Assisted Generation of Difficult Math Questions", "abs": "Current LLM training positions mathematical reasoning as a core capability. With publicly available sources fully tapped, there is an unmet demand for diverse and challenging mathematics questions. Relying solely on human experts is both time-consuming and costly, while LLM-generated questions often lack the requisite diversity and difficulty. We present a design framework that combines the strengths of LLMs with a human-in-the-loop approach to generate a diverse array of challenging math questions. Initially, leveraging LLM metacognition skills [Didolkar et al., 2024], a strong LLM is used to extract core \"skills'' from existing math datasets. These skills serve as the basis for generating novel and difficult questions by prompting the LLM with random pairs of core skills that must be utilized in the question. The use of two very different skills within each question makes finding such questions an ``out of distribution'' task for both LLMs and humans. Our pipeline employs LLMs to iteratively generate and refine questions and solutions through multi-turn prompting. Human annotators then verify and further refine the questions, with their efficiency enhanced via further LLM interactions. Applying this pipeline on skills extracted from MATH dataset [Hendrycks et al., 2024] resulted in **MATH$^2$** - a dataset of higher quality math questions, as evidenced by: (a) Lower performance of all models on MATH$^2$ than on MATH (b) Higher performance on MATH when using MATH$^2$ questions as in-context examples. Although focused on mathematics, our methodology seems applicable to other domains requiring structured reasoning, and potentially as a component of *scalable oversight*. Also of interest is a striking relationship observed between models' performance on the new dataset: the success rate on MATH$^2$ is the square on MATH. This suggests that successfully solving the question in MATH$^2$ requires a nontrivial combination of two distinct math skills.", "keywords": [ "Mathematical Reasoning", "Skill Composition", "Synthetic Data", "Model Evaluation" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "LzycEbgLoi", "primary_area": "foundation or frontier models, including LLMs", "n_author": 14, "author_info_list": [ { "name": "Jun Liu", "gender": "Male", "institution": "Central South University", "country": "", "position": "Full Professor" }, { "name": "Junwei Han", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "", "position": "Full Professor" }, { "name": "Lei Guo", "gender": "Male", "institution": "Northwestern Polytechnical University Xi'an", "country": "", "position": "Full Professor" }, { "name": "Ning Zhu", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "MS student" }, { "name": "Tianming Liu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Professor" }, { "name": "Tianyang Zhong", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "PhD student" }, { "name": "Tuo Zhang", "gender": "Male", "institution": "Northwestern Polytechnical University", "country": "CN", "position": "Associate Professor" }, { "name": "Wei Zhao", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "Associate Professor" }, { "name": "Wenjun Li", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "MS student" }, { "name": "Yingli Li", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "MS student" }, { "name": "Yisong Wang", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "MS student" }, { "name": "Youlan Shang", "gender": "Female", "institution": "Central South University", "country": "CN", "position": "MS student" }, { "name": "Zhuoyi Li", "gender": "Female", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "PhD student" }, { "name": "\u9648\u6d69", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 31, "n_ref": 60, "n_ref_all": 83, "n_fig": 5, "n_tab": 5, "L_tab": 2163, "n_element_tab": 291, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1140, "n_element_tab_1": 51, "formula_len_all": 819, "formula_len_all_1": 337, "len_all": 177097, "len_all_1": 54330, "len_abs": 1644, "len_title": 170, "len_sents": 49006, "len_sents_1": 26004, "n_sents": 335, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 122, "L_abs": 1660, "title": "3D-CT-GPT++: Enhancing 3D Radiology Report Generation with Direct Preference Optimization and Large Vision-Language Models", "abs": "Automatically generating radiology reports from three-dimensional medical images, such as 3D CT scans, plays a crucial role in modern diagnostics. Current approaches for generating 3D reports often adopt video processing methods, which struggle to effectively capture the relationships along the Z-axis. Additionally, multimodal large language model-based methods for generating 3D image reports face significant limitations, particularly in terms of the image encoder\u2019s ability to represent 3D structures and the hallucinations that arise in generated content. To address these challenges, we propose the 3D-CT-GPT++ model. This model integrates the optimized 3D image encoder CTViT-V, specifically designed for chest CT scans, and builds upon the LLaVA-1.5 architecture. Furthermore, we introduce \\textit{Direct Preference Optimization (DPO)}, where GPT-4 is used to score the outputs of our fully fine-tuned (SFT) model, creating a preference dataset for subsequent DPO training. DPO significantly reduces hallucinations in the report generation process, ensuring the generated reports are more aligned with clinical needs. We fine-tuned the model on both high-quality private and public datasets to ensure clinical relevance. Extensive experiments were conducted using standard natural language generation (NLG) evaluation metrics, including BLEU, METEOR, ROUGE-L, and GREEN, to assess the report generation performance. Experimental results demonstrate that 3D-CT-GPT++ significantly outperforms existing methods in terms of accuracy, fluency, clinical factual consistency, and clinical relevance, advancing the automation of 3D medical report generation.", "keywords": [ "Radiology Report Generation", "3D Medical Imaging", "Direct Preference Optimization", "Multimodal large Language Models" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "Lz5lOSC0zg", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jiacong Zhou", "gender": "unknown", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "PhD student" }, { "name": "Jun Yu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xianyun Wang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 21, "n_ref_uni": 59, "n_ref": 125, "n_ref_all": 156, "n_fig": 5, "n_tab": 13, "L_tab": 4028, "n_element_tab": 371, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2375, "n_element_tab_1": 193, "formula_len_all": 1713, "formula_len_all_1": 1452, "len_all": 261570, "len_all_1": 75451, "len_abs": 1654, "len_title": 113, "len_sents": 83244, "len_sents_1": 33533, "n_sents": 743, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1679, "title": "Optimizing Preference Alignment with Differentiable NDCG Ranking", "abs": "Aligning large language models with human preferences improves interaction quality and safety by ensuring outputs better reflect human values. A promising strategy involves Reinforcement Learning from Human Feedback (RLHF), starting with collecting and ranking responses generated by a supervised fine-tuning model to refine alignment. Current methods (DPO) focus on learning from pairwise preference data, categorizing responses into preferred and less preferred pairs, and optimizing by maximizing pairwise margins. Recent studies have uncovered a substantial discrepancy between the theoretical aspirations of preference learning and its real-world results. Current preference alignment techniques underperform expectations, with ranking accuracies below $60\\%$ on standard datasets. This suggests existing methods inadequately capture ideal preference relationships within sequences. To address this challenge, this paper introduces \\underline{D}irect \\underline{R}anking \\underline{P}reference \\underline{O}ptimization (DRPO), a novel method that views human preference alignment as a Learning-to-Rank (LTR) task. DRPO leverages NDCG, a widely used LTR metric, to optimize the ranking of responses within lists based on preference data, thereby enhancing ranking accuracies. Due to the nondifferentiability of NDCG, we propose diffNDCG loss, a differentiable approximation facilitated by a sorting network to simulate NDCG. Furthermore, to improve the quality of generated response, we propose a novel margin-based Adaptive Rank Policy Score. Extensive experiments have shown that DRPO outperforms existing baseline methods, enhancing the quality of the generated responses.", "keywords": [ "Language models; Human preferences alignment" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 4 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "Lz221VLWrO", "primary_area": "learning on time series and dynamical systems", "n_author": 7, "author_info_list": [ { "name": "Hao Zheng", "gender": "unknown", "institution": "University of Melbourne", "country": "AU", "position": "MS student" }, { "name": "Kuo Yang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Limin Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Qihe Huang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zhengyang Zhou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "MaShichao", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 21, "n_formula_1": 14, "n_ref_uni": 42, "n_ref": 76, "n_ref_all": 83, "n_fig": 4, "n_tab": 11, "L_tab": 4847, "n_element_tab": 709, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 5279, "n_element_tab_1": 247, "formula_len_all": 1403, "formula_len_all_1": 847, "len_all": 214379, "len_all_1": 71585, "len_abs": 1743, "len_title": 128, "len_sents": 74057, "len_sents_1": 30205, "n_sents": 517, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1752, "title": "ZeroTS: Zero-shot time series prediction via multi-party data-model interaction", "abs": "Time series forecasting (TSF) is a fundamental task in artificial intelligence, with applications ranging from weather prediction, stock market analysis to electricity demand forecasting. While existing models, particularly large language models (LLMs) tailored for TSF, primarily focus on improving accuracy and generalization through pre-training and fine-tuning, zero-shot prediction without task-specific fine-tuning, still remains underexplored. This limitation arises from the restricted scalability and flexibility of current LLMs for TSF, which struggle to fully capture the interactions between data and model. In this work, we introduce ZeroTS, a novel approach that bridges open-world knowledge with inherent data regularities by constructing multi-party interactions between data and models. On the data side, we propose a TS-RAG (Retrieval-Augmented Generation for Time Series), which efficiently retrieves both meta and series information, enabling diverse domain-specific time series to be used as prompts. On the model side, we develop a reinforcement learning framework that treats ground-truth as environments, providing error feedback to optimize a smaller model and harnessing the capabilities of LLMs. This allows ZeroTS to incrementally approach inherent data regularities while iteratively refining its outputs. We validate ZeroTS via extensive experiments on zero-shot and long-horizon forecasting. ZeroTS achieves best or second best results with comparative parameters, 1/4 memory and 1/7 inference speed, demonstrating its efficiency and effectiveness. Our results highlight the potential of Data-LLM interactions for zero-shot learning with acceptable parameters, opening new avenues on research of this underexplored area.", "keywords": [ "Time-series forecasting; Retrieval Augmented Generation; Large Language Model; Zero-shot prediction" ], "rating_list": [ 3, 3, 5, 6, 8 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 1, 2, 2, 4, 3 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4, 3 ] }, { "paper_id": "Lz0XW99tE0", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Hanlin Wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hao Zhou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jianbing Zhang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Jingjing Gong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Jingjing Liu", "gender": "unknown", "institution": "Tsinghua University", "country": "US", "position": "Full Professor" }, { "name": "Wei-Ying Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yawen Ouyang", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Yuxuan Song", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Intern" }, { "name": "Ziyao Cao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 78, "n_formula_1": 29, "n_ref_uni": 42, "n_ref": 141, "n_ref_all": 174, "n_fig": 6, "n_tab": 10, "L_tab": 2808, "n_element_tab": 299, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1438, "n_element_tab_1": 149, "formula_len_all": 7901, "formula_len_all_1": 2378, "len_all": 198512, "len_all_1": 65814, "len_abs": 1505, "len_title": 97, "len_sents": 54503, "len_sents_1": 26893, "n_sents": 405, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 48, "L_abs": 1512, "title": "A Periodic Bayesian Flow for Material Generation", "abs": "Generative modeling of crystal data distribution is an important yet challenging task due to the unique periodic physical symmetry of crystals. Diffusion-based methods have shown early promise in modeling crystal distribution. More recently, Bayesian Flow Networks were introduced to aggregate noisy latent variables, resulting in a variance-reduced parameter space that has been shown to be advantageous for modeling Euclidean data distributions with structural constraints (Song, et al.,2023). Inspired by this, we seek to unlock its potential for modeling variables located in non-Euclidean manifolds e.g. those within crystal structures, by overcoming challenging theoretical issues. We introduce CrysBFN, a novel crystal generation method by proposing a periodic Bayesian flow, which essentially differs from the original Gaussian-based BFN by exhibiting non-monotonic entropy dynamics. To successfully realize the concept of periodic Bayesian flow, CrysBFN integrates a new entropy conditioning mechanism and empirically demonstrates its significance compared to time-conditioning. Extensive experiments over both crystal ab initio generation and crystal structure prediction tasks demonstrate the superiority of CrysBFN, which consistently achieves new state-of-the-art on all benchmarks. Surprisingly, we found that CrysBFN enjoys a significant improvement in sampling efficiency, e.g., ~ 100x speedup (10 v.s. 2000 steps network forwards) compared with previous Diffusion-based methods on MP-20 dataset.", "keywords": [ "Crystal Generation", "Bayesian Flow Networks", "Crystal Structure Prediction" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 2, 2, 3 ] }, { "paper_id": "LyJi5ugyJx", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Cheng Lu", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Researcher" }, { "name": "Yang Song", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 21, "n_ref_uni": 36, "n_ref": 94, "n_ref_all": 116, "n_fig": 5, "n_tab": 3, "L_tab": 2218, "n_element_tab": 139, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 5318, "n_element_tab_1": 488, "formula_len_all": 1567, "formula_len_all_1": 1302, "len_all": 155794, "len_all_1": 69400, "len_abs": 1055, "len_title": 117, "len_sents": 29651, "len_sents_1": 24796, "n_sents": 224, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 71, "L_abs": 1060, "title": "Simplifying, Stabilizing and Scaling Continuous-time Consistency Models", "abs": "Consistency models (CMs) are a powerful class of diffusion-based generative models optimized for fast sampling. Most existing CMs are trained using discretized timesteps, which introduce additional hyperparameters and are prone to discretization errors. While continuous-time formulations can mitigate these issues, their success has been limited by training instability. To address this, we propose a simplified theoretical framework that unifies previous parameterizations of diffusion models and CMs, identifying the root causes of instability. Based on this analysis, we introduce key improvements in diffusion process parameterization, network architecture, and training objectives. These changes enable us to train continuous-time CMs at an unprecedented scale, reaching 1.5B parameters on ImageNet 512\u00d7512. Our proposed training algorithm, using only two sampling steps, achieves FID scores of 2.06 on CIFAR-10, 1.48 on ImageNet 64\u00d764, and 1.88 on ImageNet 512\u00d7512, narrowing the gap in FID scores with the best existing diffusion models to within 10\\%.", "keywords": [ "continuous-time consistency models", "diffusion models", "fast sampling" ], "rating_list": [ 8, 10, 8, 10, 10 ], "soundness_list": [ 3, 3, 3, 4, 4 ], "presentation_list": [ 3, 3, 4, 4, 3 ], "contribution_list": [ 3, 4, 3, 4, 4 ], "confidence_list": [ 4, 4, 3, 4, 5 ] }, { "paper_id": "Ly0SQh7Urv", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Aaron Grattafiori", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Principal Researcher" }, { "name": "Cristian Canton Ferrer", "gender": "Male", "institution": "Meta", "country": "", "position": "Principal Researcher" }, { "name": "Erik Brinkman", "gender": "unknown", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Hailey Nguyen", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Ivan Evtimov", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Research Scientist" }, { "name": "Joanna Bitton", "gender": "Female", "institution": "Meta, Inc.", "country": "US", "position": "Software Engineer, Meta AI" }, { "name": "Krithika Iyer", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Software Engineer" }, { "name": "Maya Pavlova", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "V\u00edtor Albiero", "gender": "Male", "institution": "Meta", "country": "US", "position": "Research Scientist" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 29, "n_ref": 50, "n_ref_all": 59, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 355, "formula_len_all_1": 119, "len_all": 143621, "len_all_1": 56644, "len_abs": 1657, "len_title": 119, "len_sents": 52081, "len_sents_1": 29687, "n_sents": 350, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1656, "title": "Automated Red Teaming with GOAT: the Generative Offensive Agent Tester", "abs": "Red teaming assesses how large language models (LLMs) can produce content that violates norms, policies, and rules set during their safety training. However, most existing automated methods in the literature are not representative of the way humans tend to interact with AI models. Common users of AI models may not have advanced knowledge of adversarial machine learning methods or access to model internals, and they do not spend a lot of time crafting a single highly effective adversarial prompt. Instead, they are likely to make use of techniques commonly shared online and exploit the multi-turn conversational nature of LLMs. While manual testing addresses this gap, it is an inefficient and often expensive process. To address these limitations, we introduce the Generative Offensive Agent Tester (GOAT), an automated agentic red teaming system that simulates plain language adversarial conversations while leveraging multiple adversarial prompting techniques to identify vulnerabilities in LLMs. We instantiate GOAT with 7 red teaming attacks by prompting a general-purpose model in a way that encourages reasoning through the choices of methods available, the current target model\u2019s response, and the next steps. Our approach is designed to be extensible and efficient, allowing human testers to focus on exploring new areas of risk while automation covers the scaled adversarial stress-testing of known risk territory. We present the design and evaluation of GOAT, demonstrating its effectiveness in identifying vulnerabilities in state-of-the-art LLMs, with an ASR@10 of 97% against Llama 3.1 and 88% against GPT-4 on the JailbreakBench dataset", "keywords": [ "red teaming", "adversarial machine learning", "adversarial examples", "attacks on language models" ], "rating_list": [ 5, 6, 6, 5, 8 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "Lxst78Rrwj", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Masashi Sugiyama", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Director" }, { "name": "Nang Hung Nguyen", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Researcher" }, { "name": "Phi Le Nguyen", "gender": "Female", "institution": "Hanoi University of Science and Technology", "country": "VN", "position": "Associate Professor" }, { "name": "Thao Nguyen Truong", "gender": "Male", "institution": "AIST, National Institute of Advanced Industrial Science and Technology", "country": "JP", "position": "Researcher" }, { "name": "Nghia Hoang", "gender": "Male", "institution": "Washington State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 47, "n_formula_1": 19, "n_ref_uni": 30, "n_ref": 75, "n_ref_all": 124, "n_fig": 15, "n_tab": 6, "L_tab": 2327, "n_element_tab": 89, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 408, "n_element_tab_1": 15, "formula_len_all": 4710, "formula_len_all_1": 1923, "len_all": 182672, "len_all_1": 69311, "len_abs": 1200, "len_title": 128, "len_sents": 70374, "len_sents_1": 33524, "n_sents": 564, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1122, "title": "Causal Graph Learning via Distributional Invariance of Cause-Effect Relationship", "abs": "This paper introduces a new framework for recovering causal graphs from observational data, leveraging the fact that the distribution of an effect, conditioned on its causes, remains invariant to changes in the prior distribution of those causes. This insight enables a direct test for potential causal relationships by checking the variance of their corresponding effect-cause conditional distributions across multiple downsampled subsets of the data. These subsets are selected to reflect different prior cause distributions, while preserving the effect-cause conditional relationships. Using this invariance test and exploiting an (empirical) sparsity of most causal graphs, we develop an algorithm that efficiently uncovers causal relationships with quadratic complexity in the number of observational features/variables, reducing the processing time by up to 25x compared to state-of-the-art methods. Our empirical studies on a diverse benchmark of large-scale datasets demonstrate that the developed algorithm consistently performs better or comparable to existing works while generally achieving better scalability.", "keywords": [ "Causal Graph Learning", "Invariance" ], "rating_list": [ 3, 6, 6, 5 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 5, 3, 2 ] }, { "paper_id": "LxkgScfHKf", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Hamed Hassani", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Assistant Professor" }, { "name": "Orlando Romero", "gender": "unknown", "institution": "University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Sima Noorani", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "George Pappas", "gender": "Male", "institution": "School of Engineering and Applied Science, University of Pennsylvania", "country": "US", "position": "Full Professor" }, { "name": "Nicol\u00f2 Dal Fabbro", "gender": "Male", "institution": "University of Padua", "country": "", "position": "PhD student" } ], "n_formula": 119, "n_formula_1": 32, "n_ref_uni": 27, "n_ref": 67, "n_ref_all": 108, "n_fig": 8, "n_tab": 6, "L_tab": 1663, "n_element_tab": 201, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 624, "n_element_tab_1": 65, "formula_len_all": 9366, "formula_len_all_1": 2758, "len_all": 171387, "len_all_1": 64113, "len_abs": 1279, "len_title": 88, "len_sents": 56276, "len_sents_1": 27980, "n_sents": 453, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1019, "title": "Conformal Training with Reduced Variance", "abs": "Conformal prediction (CP) is a distribution-free framework for achieving probabilistic guarantees on black-box models. {CP} is generally applied to a model post-training. Conformal training is an approach that aims to optimize the CP efficiency during training. In this direction, ConfTr (Stutz et al, 2022) is a technique that seeks to minimize the expected prediction set size of a model by simulating {CP} in-between training updates. Despite its potential, we identify a strong source of sample inefficiency in ConfTr that leads to overly noisy estimated gradients, introducing training instability and limiting practical use. To address this challenge, we propose variance-reduced conformal training (VR-ConfTr), a method that incorporates a variance reduction technique in the gradient estimation of the ConfTr objective function. Through extensive experiments on various benchmark datasets, we demonstrate that VR-ConfTr consistently achieves faster convergence and smaller prediction sets compared to baselines.", "keywords": [ "Conformal Training", "Conformal Prediction", "Optimization", "Quantile", "Deep Learning", "Uncertainty Quantification" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "Lwf5WeiyA9", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "BOSHU LEI", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "MS student" }, { "name": "Katrina Ashton", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Kostas Daniilidis", "gender": "Male", "institution": "Athena Research and Innovation Centre", "country": "GR", "position": "Affiliated Researcher" }, { "name": "Wen Jiang", "gender": "unknown", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 21, "n_ref_uni": 32, "n_ref": 93, "n_ref_all": 101, "n_fig": 4, "n_tab": 3, "L_tab": 612, "n_element_tab": 101, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 987, "n_element_tab_1": 91, "formula_len_all": 944, "formula_len_all_1": 947, "len_all": 129888, "len_all_1": 54288, "len_abs": 913, "len_title": 87, "len_sents": 28810, "len_sents_1": 24496, "n_sents": 205, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 39, "L_abs": 916, "title": "AG-SLAM: Active Gaussian Splatting SLAM", "abs": "We present AG-SLAM, the first active SLAM system utilizing 3D Gaussian Splatting (3DGS) for online scene reconstruction. In recent years, radiance field scene representations, including 3DGS have been widely used in SLAM and exploration, but actively planning trajectories for robotic exploration is still unvisited. In particular, many exploration methods assume precise localization and thus do not mitigate the significant risk of constructing a trajectory, which is difficult for a SLAM system to operate on. This can cause camera tracking failure and lead to failures in real-world robotic applications. Our method leverages Fisher Information to balance the dual objectives of maximizing the information gain for the environment while minimizing the cost of localization errors. Experiments conducted on the Gibson and Habitat-Matterport 3D datasets demonstrate state-of-the-art results of the proposed method.", "keywords": [ "Active Learning", "SLAM", "View Synthesis", "3D Gaussian Splatting" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "LwLaFpJpfM", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Eliya Nachmani", "gender": "Male", "institution": "Ben Gurion University of the Negev", "country": "IL", "position": "Assistant Professor" }, { "name": "Lior Wolf", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Associate Professor" }, { "name": "Yehuda Mishaly", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "MS student" } ], "n_formula": 20, "n_formula_1": 17, "n_ref_uni": 50, "n_ref": 101, "n_ref_all": 124, "n_fig": 8, "n_tab": 5, "L_tab": 1054, "n_element_tab": 68, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 422, "n_element_tab_1": 70, "formula_len_all": 862, "formula_len_all_1": 575, "len_all": 146957, "len_all_1": 63039, "len_abs": 934, "len_title": 103, "len_sents": 41783, "len_sents_1": 29791, "n_sents": 307, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 905, "title": "Active Audio Cancellation with Multi-Band Mamba Network", "abs": "A novel deep learning approach for Active Audio Cancellation (AAC) is presented, which surpasses traditional Active Noise Cancellation (ANC) by effectively canceling any audio signal, regardless of its spectral content. We propose, for the first time, a deep learning approach to AAC using a novel multi-band Mamba architecture. This architecture partitions input audio into multiple frequency bands, allowing for precise anti-signal generation and enhanced phase alignment across frequencies, thereby improving overall cancellation performance. Additionally, we introduce an optimization-driven loss function that provides near-optimal supervisory signals for anti-signal generation. Our experimental results demonstrate substantial improvements over existing methods, achieving up to 7.2dB gain in ANC scenarios and up to 6.2dB improvement in AAC for voice audio signals, outperforming existing methods.", "keywords": [ "Active Noise Cancellation", "Audio", "Speech", "Mamba" ], "rating_list": [ 8, 6, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "LwAG269lIq", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "KAMAL YOUCEF-TOUMI", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Mohsen Sadr", "gender": "unknown", "institution": "Paul Scherrer Institute", "country": "CH", "position": "Postdoc" }, { "name": "Tony Tohme", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD Candidate" } ], "n_formula": 54, "n_formula_1": 19, "n_ref_uni": 31, "n_ref": 53, "n_ref_all": 97, "n_fig": 18, "n_tab": 3, "L_tab": 634, "n_element_tab": 74, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 339, "n_element_tab_1": 46, "formula_len_all": 4689, "formula_len_all_1": 2040, "len_all": 153603, "len_all_1": 64288, "len_abs": 1313, "len_title": 100, "len_sents": 51618, "len_sents_1": 28931, "n_sents": 394, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 1363, "title": "Data-Driven Discovery of PDEs via the Adjoint Method", "abs": "In this work, we present an adjoint-based method for discovering the underlying governing partial differential equations (PDEs) given data. The idea is to consider a parameterized PDE in a general form and formulate a PDE-constrained optimization problem aimed at minimizing the error of the PDE solution from data. Using variational calculus, we obtain an evolution equation for the Lagrange multipliers (adjoint equations) allowing us to compute the gradient of the objective function with respect to the parameters of PDEs given data in a straightforward manner. In particular, we consider a family of parameterized PDEs encompassing linear, nonlinear, and spatial derivative candidate terms, and elegantly derive the corresponding adjoint equations. We show the efficacy of the proposed approach in identifying the form of the PDE up to machine accuracy, enabling the accurate discovery of PDEs from data. We also compare its performance with the famous PDE Functional Identification of Nonlinear Dynamics method known as PDE-FIND [Rudy, Samuel H., et al. Science advances 3.4 (2017): e1602614.], on both smooth and noisy data sets. Even though the proposed adjoint method relies on forward/backward solvers, it outperforms PDE-FIND for large data sets thanks to the analytic expressions for gradients of the cost function with respect to each PDE parameter.", "keywords": [ "PDE discovery", "Symbolic Regression", "Adjoint method" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "LvuSFvGShf", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Dokwan Oh", "gender": "Male", "institution": "Samsung Advanced Institute of Technology", "country": "KR", "position": "Principal Researcher" }, { "name": "Jaehoon Oh", "gender": "Male", "institution": "Samsung Advanced Institute of Technology", "country": "KR", "position": "Researcher" }, { "name": "Seungjun Shin", "gender": "Male", "institution": "Samsung Advanced Institute of Technology", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 50, "n_ref": 77, "n_ref_all": 120, "n_fig": 19, "n_tab": 8, "L_tab": 1764, "n_element_tab": 144, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 1827, "n_element_tab_1": 194, "formula_len_all": 215, "formula_len_all_1": 218, "len_all": 183793, "len_all_1": 61206, "len_abs": 1550, "len_title": 87, "len_sents": 44741, "len_sents_1": 27123, "n_sents": 343, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1564, "title": "House of Cards: Massive Weights in LLMs", "abs": "Massive activations, which manifest in specific feature dimensions of hidden states, introduce a significant bias in large language models (LLMs), leading to an overemphasis on the corresponding token. In this paper, we identify that massive activations originate not from the hidden state but from the intermediate state of a feed-forward network module in an early layer. Expanding on the previous observation that massive activations occur only in specific feature dimensions, we dive deep into the weights that cause massive activations. Specifically, we define *top-$k$ massive weights* as the weights that contribute to the dimensions with the top-$k$ magnitudes in the intermediate state. When these massive weights are set to zero, the functionality of LLMs is entirely disrupted. However, when all weights except for massive weights are set to zero, it results in a relatively minor performance drop, even though a much larger number of weights are set to zero. This implies that during the pre-training process, learning is dominantly focused on massive weights. Building on this observation, we propose a simple plug-and-play method called MacDrop (massive weights curriculum dropout), to rely less on massive weights during parameter-efficient fine-tuning. This method applies dropout to the pre-trained massive weights, starting with a high dropout probability and gradually decreasing it as fine-tuning progresses. Through experiments, we demonstrate that MacDrop generally improves performance across zero-shot downstream tasks and generation tasks.", "keywords": [ "large language model", "massive weight", "massive activation" ], "rating_list": [ 3, 5, 8, 5 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "LvjSLnMlwY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Hanxun Huang", "gender": "Male", "institution": "The University of Melbourne", "country": "AU", "position": "Postdoc" }, { "name": "James Bailey", "gender": "unknown", "institution": "The University of Melbourne", "country": "AU", "position": "Full Professor" }, { "name": "Sarah Monazam Erfani", "gender": "unknown", "institution": "The University of Melbourne", "country": "AU", "position": "Associate Professor" }, { "name": "Xingjun Ma", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Ge Yi Li", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Postdoc" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 105, "n_ref": 197, "n_ref_all": 235, "n_fig": 5, "n_tab": 12, "L_tab": 12394, "n_element_tab": 1204, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 4014, "n_element_tab_1": 336, "formula_len_all": 762, "formula_len_all_1": 596, "len_all": 313328, "len_all_1": 71675, "len_abs": 1484, "len_title": 107, "len_sents": 67064, "len_sents_1": 29795, "n_sents": 704, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1502, "title": "TUAP: Targeted Universal Adversarial Perturbations for CLIP", "abs": "As Contrastive Language-Image Pretraining (CLIP) models are increasingly adopted in a wide range of downstream tasks and large Vision-Language Models (VLMs), their vulnerability to adversarial attacks has attracted growing attention. In this work, we examine the susceptibility of CLIP models to Universal Adversarial Perturbations (UAPs). Unlike existing works that focus on untargeted attacks in a white-box setting, we investigate targeted UAPs (TUAPs) in a black-box setting, with a particular emphasis on transferability. In TUAP, the adversary can specify a targeted adversarial text description and generate a universal $L_{\\infty}$-norm-bounded or $L_2$-norm perturbation or a small unrestricted patch, using an ensemble of surrogate CLIP encoders. When TUAP is applied to different test images, it can mislead the image encoder of unseen CLIP models into producing image embeddings that are consistently close to the adversarial target text embedding. We conduct comprehensive experiments to demonstrate the effectiveness and transferability of TUAPs. This universal transferability extends not only across different datasets and models but also to downstream models, such as large VLMs including OpenFlamingo, LLaVA, MiniGPT-4 and BLIP2. TUAP can mislead them into generating responses that contain text descriptions specified by the adversaries. Our findings reveal a universal vulnerability in CLIP models to targeted adversarial attacks, emphasizing the need for effective countermeasures.", "keywords": [ "Adversarial", "Universal", "Perturbation", "Vision Language Model", "VLM", "CLIP" ], "rating_list": [ 8, 3, 3, 3 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 1 ] }, { "paper_id": "LvgOm9Rmih", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Baoxiong Jia", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Pengxiang Li", "gender": "unknown", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Qing Li", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence (BIGAI)", "country": "CN", "position": "Researcher" }, { "name": "Siyuan Huang", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Tengyu Liu", "gender": "Male", "institution": "Beijing Institute of General Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Xiaojian Ma", "gender": "Unspecified", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Yixin Chen", "gender": "Male", "institution": "BIGAI", "country": "CN", "position": "Researcher" }, { "name": "Zhuofan Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Ziyu Zhu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "tianxu Wang", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 60, "n_ref": 117, "n_ref_all": 143, "n_fig": 15, "n_tab": 4, "L_tab": 1650, "n_element_tab": 169, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1694, "n_element_tab_1": 256, "formula_len_all": 173, "formula_len_all_1": 175, "len_all": 189907, "len_all_1": 64928, "len_abs": 830, "len_title": 95, "len_sents": 41555, "len_sents_1": 27266, "n_sents": 354, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1339, "title": "Task-oriented Sequential Grounding in 3D Scenes", "abs": "Grounding natural language in physical 3D environments is essential for the advancement of embodied artificial intelligence. Current datasets and models for 3D visual grounding predominantly focus on identifying and localizing objects from static, object-centric descriptions. These approaches do not adequately address the dynamic and sequential nature of task-oriented grounding necessary for practical applications. In this work, we propose a new task: Task-oriented Sequential Grounding in 3D scenes, wherein an agent must follow detailed step-by-step instructions to complete daily activities by locating a sequence of target objects in indoor scenes. To facilitate this task, we introduce SG3D, a large-scale dataset containing 22,346 tasks with 112,236 steps across 4,895 real-world 3D scenes. The dataset is constructed using a combination of RGB-D scans from various 3D scene datasets and an automated task generation pipeline, followed by human verification for quality assurance. We adapted three state-of-the-art 3D visual grounding models to the sequential grounding task and evaluated their performance on SG3D. Our results reveal that while these models perform well on traditional benchmarks, they face significant challenges with task-oriented sequential grounding, underscoring the need for further research in this area.", "keywords": [ "3D visual grounding", "3D scene understanding", "vision and language", "grounded task planning" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "LvTSvdiSwG", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Dmitrii Zhemchuzhnikov", "gender": "Male", "institution": "University of Grenoble-Alpes", "country": "FR", "position": "PhD student" }, { "name": "Sergei Grudinin", "gender": "Male", "institution": "University of Grenoble-Alpes", "country": "FR", "position": "Principal Researcher" } ], "n_formula": 55, "n_formula_1": 27, "n_ref_uni": 36, "n_ref": 66, "n_ref_all": 92, "n_fig": 6, "n_tab": 2, "L_tab": 2138, "n_element_tab": 321, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5670, "formula_len_all_1": 1656, "len_all": 161855, "len_all_1": 64394, "len_abs": 1361, "len_title": 113, "len_sents": 50412, "len_sents_1": 29643, "n_sents": 424, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1366, "title": "On the Fourier analysis in the SO(3) space : the EquiLoPO Network", "abs": "Analyzing volumetric data with rotational invariance or equivariance is currently an active research topic. Existing deep-learning approaches utilize either group convolutional networks limited to discrete rotations or steerable convolutional networks with constrained filter structures. This work proposes a novel equivariant neural network architecture that achieves analytical Equivariance to Local Pattern Orientation on the continuous SO(3) group while allowing unconstrained trainable filters - EquiLoPO Network. Our key innovations are a group convolutional operation leveraging irreducible representations as the Fourier basis and a local activation function in the SO(3) space that provides a well-defined mapping from input to output functions, preserving equivariance. By integrating these operations into a ResNet-style architecture, we propose a model that overcomes the limitations of prior methods. A comprehensive evaluation on diverse 3D medical imaging datasets from MedMNIST3D demonstrates the effectiveness of our approach, which consistently outperforms state of the art. This work suggests the benefits of true rotational equivariance on SO(3) and flexible unconstrained filters enabled by the local activation function, providing a flexible framework for equivariant deep learning on volumetric data with potential applications across domains.", "keywords": [ "Equivariance", "Fourier Analysis", "SO(3)", "MedMNIST3D", "Local Activation", "CNN", "Group Convolution", "Computer Vision", "3D Medical Images" ], "rating_list": [ 5, 5, 6, 3, 6 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 3, 1, 3, 1, 3 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 3, 4 ] }, { "paper_id": "LvRQgsvd5V", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Arnav Kumar Jain", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "Glen Berseth", "gender": "Male", "institution": "University of Montreal, University of Montreal", "country": "", "position": "Assistant Professor" }, { "name": "Harley Wiltzer", "gender": "Male", "institution": "McGill University, McGill University", "country": "CA", "position": "PhD student" }, { "name": "Irina Rish", "gender": "Female", "institution": "University of Montreal", "country": "CA", "position": "Professor" }, { "name": "Jesse Farebrother", "gender": "Male", "institution": "Meta AI", "country": "FR", "position": "Research Scientist Intern" }, { "name": "Sanjiban Choudhury", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Assistant Professor" } ], "n_formula": 47, "n_formula_1": 26, "n_ref_uni": 60, "n_ref": 167, "n_ref_all": 199, "n_fig": 9, "n_tab": 4, "L_tab": 1068, "n_element_tab": 79, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 620, "n_element_tab_1": 12, "formula_len_all": 4656, "formula_len_all_1": 2084, "len_all": 187334, "len_all_1": 71910, "len_abs": 1121, "len_title": 125, "len_sents": 51944, "len_sents_1": 33534, "n_sents": 385, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1125, "title": "Non-Adversarial Inverse Reinforcement Learning via Successor Feature Matching", "abs": "In inverse reinforcement learning (IRL), an agent seeks to replicate expert demonstrations through interactions with the environment. Traditionally, IRL is treated as an adversarial game, where an adversary searches over reward models, and a learner optimizes the reward through repeated RL procedures.\nThis game-solving approach is both computationally expensive and difficult to stabilize.\nIn this work, we propose a novel approach to IRL by direct policy optimization: exploiting a linear factorization of the return as the inner product of successor features and a reward vector, we design an IRL algorithm by policy gradient descent on the gap between the learner and expert features.\nOur non-adversarial method does not require learning a reward function and can be solved seamlessly with existing actor-critic RL algorithms.\nRemarkably, our approach works in state-only settings without expert action labels, a setting which behavior cloning (BC) cannot solve.\nEmpirical results demonstrate that our method learns from as few as a single expert demonstration and achieves improved performance on various control tasks.", "keywords": [ "Inverse Reinforcement Learning", "Imitation Learning", "Successor Features" ], "rating_list": [ 3, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "LvNROciCne", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Boris Shustin", "gender": "unknown", "institution": "STFC", "country": "GB", "position": "Researcher" }, { "name": "Jonathan Svirsky", "gender": "Male", "institution": "Bar-Ilan University", "country": "", "position": "PhD student" }, { "name": "Ofir Lindenbaum", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "Assistant Professor" }, { "name": "Wasim Huleihel", "gender": "Male", "institution": "Tel Aviv University, Tel Aviv University", "country": "", "position": "Assistant Professor" }, { "name": "Yehonathan Refael", "gender": "Male", "institution": "Google", "country": "IL", "position": "Intern" } ], "n_formula": 52, "n_formula_1": 18, "n_ref_uni": 45, "n_ref": 76, "n_ref_all": 95, "n_fig": 5, "n_tab": 9, "L_tab": 2160, "n_element_tab": 264, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 406, "n_element_tab_1": 68, "formula_len_all": 6551, "formula_len_all_1": 2547, "len_all": 172244, "len_all_1": 70756, "len_abs": 2510, "len_title": 146, "len_sents": 46253, "len_sents_1": 30423, "n_sents": 341, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 98, "L_abs": 1579, "title": "AdaRankGrad: Adaptive Gradient Rank and Moments for Memory-Efficient LLMs Training and Fine-Tuning", "abs": "Training and fine-tuning large language models (LLMs) come with challenges related to memory and computational requirements due to the increasing size of the model weights and the optimizer states. To tackle these challenges, various techniques have been developed, such as low-rank adaptation (LoRA), which involves introducing a parallel trainable low-rank matrix to the fixed pre-trained weights at each layer. However, these methods often fall short compared to the full-rank weight training approach, as they restrict the parameter search to a low-rank subspace. This limitation can disrupt training dynamics and may require a full-rank warm start to mitigate the impact. \nIn this paper, we introduce a new method inspired by a phenomenon we formally prove: as training progresses, the rank of the estimated layer gradients gradually decreases and asymptotically approaches rank one. Leveraging this, our approach involves adaptively reducing the rank of the gradients during Adam optimization steps, using an efficient online-updating low-rank projections rule. We further present a randomized-svd scheme for efficiently finding the projection matrix. \nOur technique enables full-parameter fine-tuning with adaptive low-rank gradient updates, significantly reducing overall memory requirements during training compared to state-of-the-art methods while improving model performance in both pretraining and fine-tuning. Finally, we provide a convergence analysis of our method and demonstrate its merits for training and fine-tuning language and biological foundation models.", "keywords": [ "low rank adaptation; low rank gradient training; memory efficient fine tuning; memory optimization; adaptive rank; foundation large language models;" ], "rating_list": [ 8, 6, 8, 5, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "LvDwwAgMEW", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Alex Tamkin", "gender": "Unspecified", "institution": "Anthropic", "country": "", "position": "Researcher" }, { "name": "Jacob Andreas", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Noah Goodman", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Associate Professor" }, { "name": "Belinda Zou Li", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 40, "n_ref": 57, "n_ref_all": 109, "n_fig": 18, "n_tab": 6, "L_tab": 1584, "n_element_tab": 64, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 184, "formula_len_all_1": 136, "len_all": 195959, "len_all_1": 63378, "len_abs": 1539, "len_title": 96, "len_sents": 71613, "len_sents_1": 33076, "n_sents": 520, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 1565, "title": "Eliciting Human Preferences with Language Models", "abs": "Language models (LMs) can be directed to perform user- and context-dependent\ntasks by using labeled examples or natural language prompts.\nBut selecting examples or writing prompts can be challenging---especially in tasks that require users to precisely articulate nebulous preferences or reason about complex edge cases. For such tasks, we introduce **Generative Active Task Elicitation (GATE)**, a method for using *LMs themselves* to guide the task specification process. GATE is a learning framework in which models elicit and infer human preferences through free-form, language-based interaction with users.\nWe identify prototypical challenges that users face when specifying preferences, and design three preference modeling tasks to study these challenges:\ncontent recommendation, moral reasoning, and email validation.\nIn preregistered experiments, we show that LMs that learn to perform these tasks using GATE (by interactively querying users with open-ended questions) obtain preference specifications that are more informative than user-written prompts or examples. GATE matches existing task specification methods in the moral reasoning task, and significantly outperforms them in the content recommendation and email validation tasks. Users additionally report that interactive task elicitation requires less effort than prompting or example labeling and surfaces considerations that they did not anticipate on their own. Our findings suggest that LM-driven elicitation can be a powerful tool for aligning models to complex human preferences and values.", "keywords": [ "question asking", "preference elicitation", "language models", "evaluation", "human studies" ], "rating_list": [ 5, 8, 6, 8 ], "soundness_list": [ 2, 4, 3, 4 ], "presentation_list": [ 3, 4, 2, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 5, 4, 5 ] }, { "paper_id": "LuytzzohTa", "primary_area": "reinforcement learning", "n_author": 9, "author_info_list": [ { "name": "Chelsea Finn", "gender": "Female", "institution": "Physical Intelligence", "country": "US", "position": "Researcher" }, { "name": "Divyansh Garg", "gender": "Male", "institution": "Stanford University", "country": "", "position": "MS student" }, { "name": "Elan Sopher Markowitz", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Julia Kiseleva", "gender": "Female", "institution": "Research, Microsoft", "country": "", "position": "Researcher" }, { "name": "Naman Garg", "gender": "Male", "institution": "Multion, Inc", "country": "US", "position": "Researcher" }, { "name": "Pranav Putta", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Rafael Rafailov", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Sumeet Ramesh Motwani", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Edmund M Mills", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "Intern" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 63, "n_ref": 164, "n_ref_all": 173, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1069, "formula_len_all_1": 891, "len_all": 220649, "len_all_1": 69328, "len_abs": 1775, "len_title": 113, "len_sents": 45542, "len_sents_1": 32658, "n_sents": 299, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1821, "title": "Agent Q: Advanced Reasoning and Learning for Autonomous AI Agents", "abs": "Large Language Models (LLMs) have shown remarkable capabilities in natural language tasks requiring complex reasoning, yet their application in agentic, multi-step reasoning within interactive environments remains a difficult challenge. Traditional supervised pre-training on static datasets falls short in enabling autonomous agent capabilities needed to perform complex decision-making in dynamic settings like web navigation. Previous attempts to bridge this gap through supervised fine-tuning on curated expert demonstrations often suffer from compounding errors and limited exploration data, resulting in sub-optimal policy outcomes. To overcome these challenges, we propose a framework that combines guided Monte Carlo Tree Search (MCTS) search with a self-critique mechanism and iterative fine-tuning on agent interactions using an off-policy variant of the Direct Preference Optimization (DPO) algorithm. Our method allows LLM agents to learn effectively from both successful and unsuccessful trajectories, thereby improving their generalization in complex, multi-step reasoning tasks. We validate our approach in the WebShop environment, a simulated e-commerce platform\u2014where it consistently outperforms behavior cloning and reinforced fine-tuning baseline, and \\textbf{beats average human performance} when equipped with the capability to do online search. In real-world booking\nscenarios, our methodology boosts Llama-3 70B model's zero-shot performance from \\textbf{18.6\\% to 81.7\\%} success rate (a \\textbf{340\\% relative increase}) after a single day of data collection and further to \\textbf{95.4\\%} with online search. We believe this represents a substantial leap forward in the capabilities of autonomous agents, paving the way for more sophisticated and reliable decision-making in real-world settings.", "keywords": [ "Web agent", "RLHF", "Tree Search", "RL" ], "rating_list": [ 8, 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 4, 2, 2 ], "presentation_list": [ 4, 1, 2, 3, 1 ], "contribution_list": [ 3, 3, 2, 1, 2 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "Lut5t3qElA", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Bhiksha Raj", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Gus Xia", "gender": "Male", "institution": "New York University", "country": "", "position": "Assistant Professor" }, { "name": "Yuxuan Wu", "gender": "Not Specified", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "PhD student" }, { "name": "Ziyu Wang", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 46, "n_ref": 94, "n_ref_all": 137, "n_fig": 16, "n_tab": 12, "L_tab": 2048, "n_element_tab": 481, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 671, "n_element_tab_1": 192, "formula_len_all": 869, "formula_len_all_1": 817, "len_all": 172598, "len_all_1": 58755, "len_abs": 1398, "len_title": 133, "len_sents": 49796, "len_sents_1": 25744, "n_sents": 351, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1411, "title": "Unsupervised Disentanglement of Content and Style via Variance-Invariance Constraints", "abs": "We contribute an unsupervised method that effectively learns disentangled content and style representations from sequences of observations. Unlike most disentanglement algorithms that rely on domain-specific labels or knowledge, our method is based on the insight of domain-general statistical differences between content and style --- content varies more among different fragments within a sample but maintains an invariant vocabulary across data samples, whereas style remains relatively invariant within a sample but exhibits more significant variation across different samples. We integrate such inductive bias into an encoder-decoder architecture and name our method after V3 (variance-versus-invariance). Experimental results show that V3 generalizes across multiple domains and modalities, successfully learning disentangled content and style representations, such as pitch and timbre from music audio, digit and color from images of hand-written digits, and action and character appearance from simple animations. V3 demonstrates strong disentanglement performance compared to existing unsupervised methods, along with superior out-of-distribution generalization and few-shot learning capabilities compared to supervised counterparts. Lastly, symbolic-level interpretability emerges in the learned content codebook, forging a near one-to-one alignment between machine representation and human knowledge.", "keywords": [ "Unsupervised learning", "Interpretable representation learning", "Emergent knowledge", "Self-supervised learning" ], "rating_list": [ 8, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 2, 4, 2 ] }, { "paper_id": "Luss2sa0vc", "primary_area": "applications to robotics, autonomy, planning", "n_author": 9, "author_info_list": [ { "name": "Hao Dong", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Mingdong Wu", "gender": "Male", "institution": "Center on Frontiers of Computing Studies,Peking University", "country": "", "position": "PhD student" }, { "name": "Ruihai Wu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yizhou Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Yuanfei Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhaofeng He", "gender": "Male", "institution": "Beijing University of Post and Telecommunication", "country": "CN", "position": "Full Professor" }, { "name": "\u5c0f\u6770 \u5f20", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Yan Zhao", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Bruce Holmes Lee", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 28, "n_ref": 70, "n_ref_all": 93, "n_fig": 8, "n_tab": 9, "L_tab": 276, "n_element_tab": 46, "n_fig_1": 6, "n_tab_1": 8, "L_tab_1": 1850, "n_element_tab_1": 64, "formula_len_all": 245, "formula_len_all_1": 253, "len_all": 114471, "len_all_1": 53552, "len_abs": 1401, "len_title": 84, "len_sents": 33279, "len_sents_1": 24945, "n_sents": 283, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1407, "title": "AdaManip: Adaptive Articulated Object Manipulation Environments and Policy Learning", "abs": "Articulated object manipulation is a critical capability for robots to perform various tasks in real-world scenarios. Composed of multiple parts connected by joints, articulated objects are endowed with diverse functional mechanisms through complex relative motions. For example, a safe consists of a door, a handle, and a lock, where the door can only be opened when the latch is unlocked. The internal structure, such as the state of a lock or joint angle constraints, cannot be directly observed from visual observation. Consequently, successful manipulation of these objects requires adaptive adjustment based on trial and error rather than a one-time visual inference. However, previous datasets and simulation environments for articulated objects have primarily focused on simple manipulation mechanisms where the complete manipulation process can be inferred from the object's appearance. To enhance the diversity and complexity of adaptive manipulation mechanisms, we build a novel articulated object manipulation environment and equip it with 9 categories of articulated objects. Based on the environment and objects, we further propose an adaptive demonstration collection pipeline and a 3D visual diffusion-based imitation learning that learns the adaptive manipulation policy. The effectiveness of our designs and proposed method are validated through both simulation and real-world experiments.", "keywords": [ "Articulated Object Manipulation", "Adaptive Mechanism Environments", "Imitation Learning" ], "rating_list": [ 5, 8, 1, 6, 5 ], "soundness_list": [ 3, 4, 2, 3, 4 ], "presentation_list": [ 3, 4, 4, 3, 1 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 4, 5 ] }, { "paper_id": "LugcDgDjv1", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Hankook Lee", "gender": "Male", "institution": "Sungkyunkwan University", "country": "", "position": "Assistant Professor" }, { "name": "Moontae Lee", "gender": "unknown", "institution": "LG Corporation", "country": "KR", "position": "Director" }, { "name": "Sungjun Cho", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "JAEHOON LEE", "gender": "Male", "institution": "LG AI RESEARCH", "country": "KR", "position": "Researcher" }, { "name": "Choi sungik", "gender": "Male", "institution": "LG AI Research", "country": "KR", "position": "Researcher" } ], "n_formula": 32, "n_formula_1": 11, "n_ref_uni": 41, "n_ref": 73, "n_ref_all": 97, "n_fig": 8, "n_tab": 12, "L_tab": 11720, "n_element_tab": 674, "n_fig_1": 7, "n_tab_1": 7, "L_tab_1": 3162, "n_element_tab_1": 335, "formula_len_all": 3532, "formula_len_all_1": 1633, "len_all": 175406, "len_all_1": 78705, "len_abs": 1715, "len_title": 137, "len_sents": 44398, "len_sents_1": 31564, "n_sents": 363, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1618, "title": "Stochastically Capturing Partial Relationship among Features for Multivariate Forecasting", "abs": "When tackling forecasting problems that involve multiple time-series features, existing methods for capturing inter-feature information typically fall into three categories: complete-multivariate, partial-multivariate, and univariate. Complete-multivariate methods compute relationships among the entire set of features, whereas univariate cases ignore inter-feature information altogether. In contrast to these two, partial-multivariate methods group features into clusters and capture inter-feature relationships within each cluster. However, existing partial-multivariate methods deal only with specific cases where there is a single way of grouping so once the grouping way is selected, it remains unchanged. Therefore, we introduce a generalized version of partial-multivariate methods where grouping ways are sampled stochastically (called stochastic partial-multivariate methods), which can incorporate the deterministic cases using Dirac delta distributions. We propose SPMformer, a Transformer-based stochastic partial-multivariate model, with its training algorithm. We demonstrate that SPMformer outperforms various complete-multivariate, deterministic partial-multivariate, and univariate models in various forecasting tasks (long-term, short-term, and probabilistic forecasting), providing a theoretical rationale and empirical analysis for its superiority. Additionally, by proposing an inference method leveraging the inherent stochasticity in SPMformer, the forecasting accuracy is further enhanced. Finally, we highlight other advantages of SPMformer: efficiency and robustness under missing features.", "keywords": [ "Multivariate Time Series", "Forecasting", "Stochastic Algorithm" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "LuVulfPgZN", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Gang Niu", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Senior Research Scientist (tenured)" }, { "name": "Masashi Sugiyama", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Director" }, { "name": "Tongliang Liu", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Zhuo Huang", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 36, "n_formula_1": 34, "n_ref_uni": 39, "n_ref": 68, "n_ref_all": 92, "n_fig": 8, "n_tab": 4, "L_tab": 1548, "n_element_tab": 89, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 2628, "n_element_tab_1": 72, "formula_len_all": 3061, "formula_len_all_1": 2996, "len_all": 145382, "len_all_1": 71314, "len_abs": 1510, "len_title": 128, "len_sents": 38221, "len_sents_1": 28693, "n_sents": 299, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1527, "title": "Towards Out-of-Modal Generalization without Instance-level Modal Correspondence", "abs": "The world is understood from various modalities, such as appearance, sound, language, etc. Since each modality only partially represents objects in a certain physical meaning, leveraging additional ones is beneficial in both theory and practice. However, exploiting novel modalities normally requires cross-modal pairs corresponding to the same instance, which is extremely resource-consuming and sometimes even impossible, making knowledge exploration of novel modalities largely restricted. To seek practical multi-modal learning, here we study Out-of-Modal (OOM) Generalization as an initial attempt to generalize to an unknown modality without given instance-level modal correspondence. Specifically, we consider Semi-Supervised and Unsupervised scenarios of OOM Generalization, where the first has scarce correspondences and the second has none, and propose connect & explore (COX) to solve these problems. COX first connects OOM data and known In-Modal (IM) data through a variational information bottleneck framework to extract shared information. Then, COX leverages the shared knowledge to create emergent correspondences, which is theoretically justified from an information-theoretic perspective. As a result, the label information on OOM data emerges along with the correspondences, which help explore the OOM data with unknown knowledge, thus benefiting generalization results. We carefully evaluate the proposed COX method under various OOM generalization scenarios, verifying its effectiveness and extensibility.", "keywords": [ "Multi-Modal Learning; Generalization" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "LuT2CVrlpU", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Aamodh Suresh", "gender": "unknown", "institution": "DEVCOM Army Research Laboratory", "country": "US", "position": "Postdoc" }, { "name": "Carlos Nieto-Granda", "gender": "unknown", "institution": "DEVCOM Army Research Laboratory", "country": "US", "position": "Researcher" }, { "name": "Wesley Suttle", "gender": "unknown", "institution": "Army Research Laboratory", "country": "US", "position": "Postdoc" } ], "n_formula": 42, "n_formula_1": 27, "n_ref_uni": 34, "n_ref": 85, "n_ref_all": 111, "n_fig": 9, "n_tab": 3, "L_tab": 1346, "n_element_tab": 123, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 223, "n_element_tab_1": 19, "formula_len_all": 3061, "formula_len_all_1": 2089, "len_all": 122839, "len_all_1": 61305, "len_abs": 1596, "len_title": 128, "len_sents": 38946, "len_sents_1": 27892, "n_sents": 268, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1412, "title": "Behavioral Entropy-Guided Dataset Generation for Offline Reinforcement Learning", "abs": "Entropy-based objectives are widely used to perform state space exploration in reinforcement learning (RL) and dataset generation for offline RL. Behavioral entropy (BE), a rigorous generalization of classical entropies that incorporates cognitive and perceptual biases of agents, was recently proposed for discrete settings and shown to be a promising metric for robotic exploration problems. In this work, we propose using BE as a principled exploration objective for systematically generating datasets that provide diverse state space coverage in complex, continuous, potentially high-dimensional domains. To achieve this, we extend the notion of BE to continuous settings, derive tractable $k$-nearest neighbor estimators, provide theoretical guarantees for these estimators, and develop practical reward functions that can be used with standard RL methods to learn BE-maximizing policies. Using standard MuJoCo environments, we experimentally compare the performance of offline RL algorithms for a variety of downstream tasks on datasets generated using BE, R\\'{e}nyi, and Shannon entropy-maximizing policies, as well as the SMM and RND algorithms. We find that offline RL algorithms trained on datasets collected using BE outperform those trained on datasets collected using Shannon entropy, SMM, and RND on all tasks considered, and on 80\\% of the tasks compared to datasets collected using Renyi entropy.", "keywords": [ "reinforcement learning", "offline reinforcement learning", "exploration", "entropy" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "LuSZGyud4O", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Hao Lin", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "Siyuan Liang", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Wenqiang Wang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Xiaochun Cao", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Xiaojun Jia", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" } ], "n_formula": 16, "n_formula_1": 12, "n_ref_uni": 40, "n_ref": 62, "n_ref_all": 74, "n_fig": 4, "n_tab": 8, "L_tab": 2074, "n_element_tab": 271, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 837, "n_element_tab_1": 90, "formula_len_all": 1302, "formula_len_all_1": 1187, "len_all": 138615, "len_all_1": 66216, "len_abs": 1609, "len_title": 124, "len_sents": 40674, "len_sents_1": 29815, "n_sents": 337, "n_sents_1": 229, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1604, "title": "No Access, No Safety: Free Lunch Adversarial Attacks on Black-box NLP Models", "abs": "Textual adversarial attacks confuse Natural Language Processing (NLP) models,\nsuch as Large Language Models (LLMs), by finely modifying the text, resulting\nin incorrect decisions. Although existing adversarial attacks are effective, they\ntypically rely on knowing the victim model, using extensive queries, or grasping\ntraining data, which limits their real-world applications. In situations where there\nis neither knowledge of nor access to the victim model, we introduce the Free\nLunch Adversarial Attack (FLA), demonstrating that attackers can successfully\nexecute attacks armed only with victim texts. To prevent access to the victim\nmodel, we create a shadow dataset with publicly available pre-trained models and\nclustering methods as a foundation for developing substitute models. To address\nthe low attack success rate (ASR) due to insufficient information feedback, we\npropose the hierarchical substitution model design, generating substitute models\nthat approximate the victim\u2019s decision boundaries to enhance ASR. Concurrently,\nwe use diverse adversarial example generation, employing various attack methods\nto reduce the frequency of model training, balancing effectiveness with efficiency.\nExperiments with the Emotion and SST5 datasets show that the FLA outperforms\nexisting state-of-the-art methods while lowering the attack cost to\nzero. More importantly, we discover that FLA poses a significant threat to LLMs\nsuch as Qwen2 and the GPT family, and achieves the highest ASR of 45.99% even\nwithout access to the API, confirming that advanced NLP models still face serious\nsecurity risks.", "keywords": [ "Text Adversarial Attacks\uff0c Trustworthy artificial intelligence" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "LuLzcBsp5c", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Gao Cong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Jingge Xiao", "gender": "unknown", "institution": "Universit\u00e4t Hannover", "country": "", "position": "PhD student" }, { "name": "Simon Gottschalk", "gender": "Male", "institution": "L3S Research Center, Leibniz Universit\u00e4t Hannover", "country": "DE", "position": "Postdoc" }, { "name": "Wolfgang Nejdl", "gender": "Male", "institution": "L3S Research Center", "country": "DE", "position": "Full Professor" }, { "name": "Yile Chen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 50, "n_ref": 84, "n_ref_all": 104, "n_fig": 4, "n_tab": 3, "L_tab": 411, "n_element_tab": 44, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3135, "n_element_tab_1": 254, "formula_len_all": 1041, "formula_len_all_1": 966, "len_all": 133069, "len_all_1": 65666, "len_abs": 1386, "len_title": 129, "len_sents": 36970, "len_sents_1": 27784, "n_sents": 270, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1392, "title": "FlexTSF: A universal forecasting model for time series with variable regularities", "abs": "Developing a foundation model for time series forecasting across diverse domains has attracted significant attention in recent years. Existing works typically assume regularly sampled, well-structured data, limiting their applicability to more generalized scenarios where time series often contain missing values, unequal sequence lengths, and irregular time intervals between measurements. To cover diverse domains and handle variable regularities, we propose FlexTSF, a universal time series forecasting model that possesses better generalization and natively support both regular and irregular time series. FlexTSF produces forecasts in an autoregressive manner and incorporates three novel designs: VT-Norm, a normalization strategy to ablate data domain barriers, IVP Patcher, a patching module to learn representations from flexibly structured time series, and LED attention, an attention mechanism seamlessly integrating these two and propagate forecasts with awareness of domain and time information, enabling effective time series forecasting across varying regularities. Experiments on $12$ datasets show that FlexTSF outperforms state-of-the-art forecasting models respectively designed for regular and irregular time series. Furthermore, after self-supervised pre-training, FlexTSF shows exceptional performance in both zero-shot and few-show settings for time series forecasting.", "keywords": [ "Time series forecasting", "Universal forecasting model", "Irregular time series", "Regular time series", "Self-supervised pre-training", "Zero-shot learning" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "LuGHbK8qTa", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Hao Su", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Isabella Liu", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Xiaolong Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 50, "n_ref": 105, "n_ref_all": 148, "n_fig": 16, "n_tab": 7, "L_tab": 1088, "n_element_tab": 83, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1376, "n_element_tab_1": 143, "formula_len_all": 746, "formula_len_all_1": 608, "len_all": 168477, "len_all_1": 64033, "len_abs": 512, "len_title": 123, "len_sents": 50307, "len_sents_1": 29627, "n_sents": 373, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1392, "title": "Dynamic Gaussians Mesh: Consistent Mesh Reconstruction from Monocular Videos", "abs": "Modern 3D engines and graphics pipelines require mesh as a memory-efficient representation, which allows efficient rendering, geometry processing, texture editing, and many other downstream operations. However, it is still highly difficult to obtain high-quality mesh in terms of detailed structure and time consistency from dynamic observations. To this end, we introduce Dynamic Gaussians Mesh (DG-Mesh), a framework to reconstruct a high-fidelity and time-consistent mesh from dynamic input. Our work leverages the recent advancement in 3D Gaussian Splatting to construct the mesh sequence with temporal consistency from dynamic observations. Building on top of this representation, DG-Mesh recovers high-quality meshes from the Gaussian points and can track the mesh vertices over time, which enables applications such as texture editing on dynamic objects. We introduce the Gaussian-Mesh Anchoring, which encourages evenly distributed Gaussians, resulting better mesh reconstruction through mesh-guided densification and pruning on the deformed Gaussians. By applying cycle-consistent deformation between the canonical and the deformed space, we can project the anchored Gaussian back to the canonical space and optimize Gaussians across all time frames. During the evaluation on different datasets, DG-Mesh provides significantly better mesh reconstruction and rendering than baselines.", "keywords": [ "Dynamic Reconstrction; Video Reconstruction; 3D Gaussian Splatting" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "LtHy5y4Ep0", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 7, "author_info_list": [ { "name": "Bin Gu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Haozhen Zhang", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "Hualin Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Xingchen Li", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "Yi Chang", "gender": "Male", "institution": "Jilin University, China", "country": "CN", "position": "Full Professor" }, { "name": "Zhaogeng Liu", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Postdoc" }, { "name": "William De Vazelhes", "gender": "Male", "institution": "Technology Innovation Institute", "country": "AE", "position": "Researcher" } ], "n_formula": 70, "n_formula_1": 21, "n_ref_uni": 47, "n_ref": 74, "n_ref_all": 86, "n_fig": 4, "n_tab": 15, "L_tab": 5881, "n_element_tab": 337, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1988, "n_element_tab_1": 119, "formula_len_all": 14102, "formula_len_all_1": 2122, "len_all": 167908, "len_all_1": 62972, "len_abs": 1521, "len_title": 87, "len_sents": 40085, "len_sents_1": 25505, "n_sents": 374, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1530, "title": "Dual Variance Reduction with Momentum for Imbalanced Black-Box Discrete Prompt Learning", "abs": "Black-box prompt learning has proven to be an effective approach for customizing large language models (LLMs) offered as services to address various downstream tasks. \nWithin this domain, policy gradient-based methods have garnered substantial attention as a prominent approach for learning discrete prompts.\nHowever, the highly imbalanced data distribution in the real world limits the applicability of such approaches by influencing LLMs' tendency to favor certain categories.\nTo tackle the challenge posed by imbalanced data, this paper pioneers the integration of pairwise AUC loss into the policy gradient optimization of discrete text prompts and proposes learning discrete prompts with doubly policy gradient.\nUnfortunately, the doubly policy gradient estimation suffers from two variance components, resulting in unstable optimization.\nAs a further improvement, we propose (1) a novel unbiased variance-reduced doubly policy gradient estimator and (2) incorporating the STORM variance reduction technique. \nUltimately, we introduce a novel momentum-based discrete prompt learning method with doubly policy gradient (mDP-DPG).\nCrucially, we provide theoretical convergence guarantees for mDP-DPG within standard frameworks.\nThe experimental results show that mDP-DPG surpasses baseline approaches across diverse imbalanced text classification datasets, emphasizing the advantages of our proposed approach for tackling data imbalance.\nOur code is available at the following URL: https://anonymous.4open.science/r/DPDPG-1ECB.", "keywords": [ "prompt learning; black-box optimization; imbalanced data" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "LtBD5fFHB7", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Di ZHANG", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "VP" }, { "name": "Guibao Shen", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Luozhou Wang", "gender": "Male", "institution": "Hong Kong University of Science and Technology (GUANGZHOU)", "country": "CN", "position": "PhD student" }, { "name": "Pengfei Wan", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Director" }, { "name": "Xin Tao", "gender": "Male", "institution": "Kuaishou", "country": "CN", "position": "Researcher" }, { "name": "Ying-Cong Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Yixun Liang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Ziyang Mai", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Yijun Li", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 52, "n_ref": 132, "n_ref_all": 143, "n_fig": 4, "n_tab": 1, "L_tab": 276, "n_element_tab": 40, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 276, "n_element_tab_1": 40, "formula_len_all": 499, "formula_len_all_1": 499, "len_all": 128083, "len_all_1": 56341, "len_abs": 450, "len_title": 88, "len_sents": 29346, "len_sents_1": 26778, "n_sents": 188, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1330, "title": "Motion Inversion for Video Customization", "abs": "In this work, we present a novel approach for motion customization in video generation, addressing the widespread gap in the exploration of motion representation within video generative models. Recognizing the unique challenges posed by the spatiotemporal nature of video, our method introduces Motion Embeddings, a set of explicit, temporally coherent embeddings derived from a given video. These embeddings are designed to integrate seamlessly with the temporal transformer modules of video diffusion models, modulating self-attention computations across frames without compromising spatial integrity. Our approach provides a compact and efficient solution to motion representation, utilizing two types of embeddings: a Motion Query-Key Embedding to modulate the temporal attention map and a Motion Value Embedding to modulate the attention values. Additionally, we introduce an inference strategy that excludes spatial dimensions from the Motion Query-Key Embedding and applies a differential operation to the Motion Value Embedding, both designed to debias appearance and ensure the embeddings focus solely on motion. Our contributions include the introduction of a tailored motion embedding for customization tasks and a demonstration of the practical advantages and effectiveness of our method through extensive experiments.", "keywords": [ "Video Customization" ], "rating_list": [ 5, 5, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 1, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "LsZxlxA9da", "primary_area": "reinforcement learning", "n_author": 1, "author_info_list": [ { "name": "Tom Schaul", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 44, "n_ref": 55, "n_ref_all": 72, "n_fig": 1, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 105440, "len_all_1": 88957, "len_abs": 894, "len_title": 75, "len_sents": 28676, "len_sents_1": 28623, "n_sents": 172, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 27, "L_abs": 899, "title": "Boundless Socratic Learning", "abs": "An agent trained within a closed system can master any desired capability, as long as the following three conditions hold: (a) it receives sufficiently informative and aligned feedback, (b) its coverage of experience/data is broad enough, and (c) it has sufficient capacity and resource. In this position paper, we justify these conditions, and consider what limitations arise from (a) and (b) in closed systems, when assuming that (c) is not a bottleneck. Considering the special case of agents with matching input and output spaces (namely, language), we argue that such pure recursive self-improvement, dubbed \"*Socratic learning*\", can boost performance vastly beyond what is present in its initial data or knowledge, and is only limited by time, as well as gradual misalignment concerns. Furthermore, we propose a constructive framework to implement it, based on the notion of *language games*.", "keywords": [ "position paper", "self-improvement", "language", "games", "recursion", "ASI" ], "rating_list": [ 3, 5, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 2 ], "contribution_list": [ 1, 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 2, 5, 2 ] }, { "paper_id": "LsTIW9VAF7", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Binyan Xu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Di Tang", "gender": "Male", "institution": "Indiana University at Bloomington", "country": "US", "position": "Postdoc" }, { "name": "Fan YANG", "gender": "Male", "institution": "Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Kehuan Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Xilin Dai", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 25, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 88, "n_ref_all": 142, "n_fig": 20, "n_tab": 18, "L_tab": 5249, "n_element_tab": 470, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 1994, "n_element_tab_1": 168, "formula_len_all": 2450, "formula_len_all_1": 465, "len_all": 242193, "len_all_1": 62034, "len_abs": 1494, "len_title": 130, "len_sents": 73773, "len_sents_1": 26452, "n_sents": 609, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1515, "title": "Less is More: Stealthy and Adaptive Clean-Image Backdoor Attacks with Few Poisoned", "abs": "Deep neural networks are fundamental in security-critical applications such as facial recognition, autonomous driving, and medical diagnostics, yet they are vulnerable to backdoor attacks. Clean-image backdoor attack, a stealthy attack utilizing solely label manipulation to implant backdoors, renders models vulnerable to exploitation by malicious labelers. However, existing clean-image backdoor attacks likely lead to a noticeable drop in Clean Accuracy (CA), decreasing their stealthiness. In this paper, we show that clean-image backdoor attacks can achieve a negligible decrease in CA by poisoning only a few samples while still maintaining a high attack success rate. We introduce **G**enerative Adversarial **C**lean-Image **B**ackdoors (GCB), a novel attack method that minimizes the drop in CA to less than 1\\% by optimizing the trigger pattern for easier learning by the victim model. Leveraging a variant of InfoGAN, we ensure that the trigger pattern we used has already been contained in some training images and can be easily separated from those feature patterns used for benign tasks. Our experiments demonstrate that GCB can be adapted to 5 datasets\u2014including MNIST, CIFAR-10, CIFAR-100, GTSRB, and Tiny-ImageNet\u20145 different architectures, and 4 tasks, including classification, multi-label classification, regression, and segmentation. Furthermore, GCB demonstrates strong resistance to backdoor defenses, successfully evading all detection methods we know. Code: *anonymous.4open.science/r/GCB*.", "keywords": [ "Backdoor Attack", "Generative Adversarial Networks", "Clean-Image Backdoor Attacks", "Deep Neural Networks", "InfoGAN", "Poisoning Attack", "Model Integrity" ], "rating_list": [ 6, 5, 8, 5, 5 ], "soundness_list": [ 4, 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 2, 3 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 5, 5, 5, 5, 3 ] }, { "paper_id": "LrmPGtnros", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Boris Ginsburg", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Hainan Xu", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Travis M. Bartley", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Vladimir Bataev", "gender": "Male", "institution": "NVIDIA", "country": "AM", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 47, "n_ref_all": 62, "n_fig": 2, "n_tab": 5, "L_tab": 1883, "n_element_tab": 225, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 444, "n_element_tab_1": 90, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 120104, "len_all_1": 56242, "len_abs": 1268, "len_title": 120, "len_sents": 38288, "len_sents_1": 26737, "n_sents": 273, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1273, "title": "Three-in-One: Fast and Accurate Transducer for Hybrid-Autoregressive Speech Recognition", "abs": "We present Hybrid-Autoregressive Inference Transducers (HAI-T), a novel architecture for speech recognition that extends the Token-and-Duration Transducer (TDT) model. Trained with randomly masked predictor network outputs, HAI-T supports both autoregressive inference with all network components and non-autoregressive inference without the predictor. Additionally, we propose a novel semi-autoregressive inference method that first generates an initial hypothesis using non-autoregressive inference, followed by refinement steps where each token prediction is regenerated using parallelized autoregression on the initial hypothesis. Experiments on multiple datasets across different languages demonstrate that HAI-T achieves efficiency parity with CTC in non-autoregressive mode and with TDT in autoregressive mode. In terms of accuracy, autoregressive HAI-T achieves parity with TDT and RNN-T, while non-autoregressive HAI-T significantly outperforms CTC. Semi-autoregressive inference further enhances the model's accuracy with minimal computational overhead, and even outperforms TDT results in some cases. These results highlight HAI-T's flexibility in balancing accuracy and speed, positioning it as a strong candidate for real-world speech recognition applications.", "keywords": [ "speech recognition", "sequence modeling", "non-autoregressive models", "RNN-T", "Transducers" ], "rating_list": [ 5, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Lr8IIc1rB8", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Abdeslam Boularias", "gender": "Male", "institution": ", Rutgers University", "country": "", "position": "Associate Professor" }, { "name": "Haonan Chang", "gender": "Male", "institution": "Rutgers, New Brunswick", "country": "", "position": "PhD student" }, { "name": "Liam Schramm", "gender": "unknown", "institution": "Rutgers University", "country": "", "position": "PhD student" }, { "name": "Xinyu Zhang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Yuhan Liu", "gender": "Male", "institution": "Rutgers University", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 63, "n_ref_all": 115, "n_fig": 16, "n_tab": 5, "L_tab": 1922, "n_element_tab": 267, "n_fig_1": 16, "n_tab_1": 1, "L_tab_1": 240, "n_element_tab_1": 43, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 137952, "len_all_1": 69783, "len_abs": 1090, "len_title": 112, "len_sents": 46793, "len_sents_1": 34009, "n_sents": 402, "n_sents_1": 288, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1097, "title": "Autoregressive Action Sequence Learning for Robotic Manipulation", "abs": "Autoregressive models have demonstrated remarkable success in natural language processing. In this work, we design a simple yet effective autoregressive architecture for robotic manipulation tasks. We propose the Chunking Causal Transformer (CCT), which extends the next-single-token prediction of causal transformers to support multi-token prediction in a single pass. Further, we design a novel attention interleaving strategy that allows CCT to be trained efficiently with teacher-forcing. Based on CCT, we propose the Autoregressive Policy (ARP) model, which learns to generate action sequences autoregressively. We find that action sequence learning enables better leverage of the underlying causal relationships in robotic tasks. We evaluate ARP across diverse robotic manipulation environments, including Push-T, ALOHA, and RLBench, and show that it outperforms the state-of-the-art methods in all tested environments, while being more efficient in computation and parameter sizes. Video demonstrations, our source code and the models of ARP are all included in the supplementary material.", "keywords": [ "Multi-Task Robot Learning", "Manipulation", "Autoregressive Model" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "LqhorpRLIm", "primary_area": "applications to robotics, autonomy, planning", "n_author": 8, "author_info_list": [ { "name": "Ajay Mandlekar", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Danfei Xu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Kuancheng Wang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Matthew Bronars", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "MS student" }, { "name": "Nadun Ranawaka Arachchige", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Soroush Nasiriany", "gender": "unknown", "institution": "University of Texas, Austin", "country": "", "position": "PhD student" }, { "name": "Vaibhav Saxena", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Woo Chul Shin", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 21, "n_ref": 45, "n_ref_all": 67, "n_fig": 7, "n_tab": 12, "L_tab": 3798, "n_element_tab": 567, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2257, "n_element_tab_1": 231, "formula_len_all": 489, "formula_len_all_1": 160, "len_all": 222484, "len_all_1": 74500, "len_abs": 1690, "len_title": 121, "len_sents": 66437, "len_sents_1": 35029, "n_sents": 485, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1649, "title": "What Matters in Learning from Large-Scale Datasets for Robot Manipulation", "abs": "Imitation learning from large multi-task demonstration datasets has emerged as a promising path for building generally-capable robots. As a result, 1000s of hours have been spent on building such large-scale datasets around the globe. Despite the continuous growth of such efforts, we still lack a systematic understanding of what data should be collected to improve the utility of a robotics dataset and facilitate downstream policy learning. In this work, we conduct a large-scale dataset composition study to answer this question. We develop a data generation framework to procedurally emulate common sources of diversity in existing datasets (such as sensor placements and object types and arrangements), and use it to generate large-scale robot datasets with controlled compositions, enabling a suite of dataset composition studies that would be prohibitively expensive in the real world. We focus on two practical settings: (1) what types of diversity should be emphasized when future researchers collect large-scale datasets for robotics, and (2) how should current practitioners retrieve relevant demonstrations from existing datasets to maximize downstream policy performance on tasks of interest. Our study yields several critical insights -- for example, we find that camera poses and spatial arrangements are crucial dimensions for both diversity in collection and alignment in retrieval. In real-world robot learning settings, we find that not only do our insights from simulation carry over, but our retrieval strategies on existing datasets such as DROID allow us to consistently outperform existing training strategies by up to 70\\%.", "keywords": [ "imitation learning", "robotics", "dataset composition" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 4 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "LqTz13JS2P", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Tao Lin", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Yiling Chen", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Full Professor" } ], "n_formula": 111, "n_formula_1": 27, "n_ref_uni": 42, "n_ref": 82, "n_ref_all": 89, "n_fig": 0, "n_tab": 1, "L_tab": 145, "n_element_tab": 4, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10069, "formula_len_all_1": 1907, "len_all": 224774, "len_all_1": 81957, "len_abs": 2449, "len_title": 105, "len_sents": 67778, "len_sents_1": 32603, "n_sents": 616, "n_sents_1": 268, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 57, "L_abs": 1488, "title": "Generalized Principal-Agent Problem with a Learning Agent", "abs": "Generalized principal-agent problems, including Stackelberg games, contract design, and Bayesian persuasion, are a class of economic problems where an agent best responds to a principal's committed strategy. \nWe study repeated generalized principal-agent problems under the assumption that the principal does not have commitment power and the agent uses algorithms to learn to respond to the principal. We reduce this problem to a one-shot generalized principal-agent problem where the agent approximately best responds. Using this reduction, we show that: (1) if the agent uses contextual no-regret learning algorithms with regret $\\mathrm{Reg}(T)$, then the principal can guarantee utility at least $U^* - \\Theta\\big(\\sqrt{\\tfrac{\\mathrm{Reg}(T)}{T}}\\big)$, where $U^*$ is the principal's optimal utility in the classic model with a best-responding agent.\n(2) If the agent uses contextual no-swap-regret learning algorithms with swap-regret $\\mathrm{SReg}(T)$, then the principal cannot obtain utility more than $U^* + O(\\frac{\\mathrm{SReg(T)}}{T})$. \nBut (3) if the agent uses mean-based learning algorithms (which can be no-regret but not no-swap-regret), then the principal can sometimes do significantly better than $U^*$.\nThese results not only refine previous results in Stackelberg games and contract design, but also lead to new results for Bayesian persuasion with a learning agent and all generalized principal-agent problems where the agent does not have private information.", "keywords": [ "principal-agent problems", "Bayesian persuasion", "no-regret learning", "no-swap-regret" ], "rating_list": [ 8, 8, 5, 8 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 4, 4, 2, 2 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "LqB8cRuBua", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Fang Liu", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Licheng Jiao", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Lingling Li", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Xu Liu", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhouhuaji", "gender": "unknown", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "Principal Researcher" }, { "name": "HaoXiaoyu", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 32, "n_ref": 38, "n_ref_all": 53, "n_fig": 6, "n_tab": 7, "L_tab": 2380, "n_element_tab": 141, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 718, "n_element_tab_1": 24, "formula_len_all": 666, "formula_len_all_1": 459, "len_all": 106266, "len_all_1": 46944, "len_abs": 979, "len_title": 115, "len_sents": 27618, "len_sents_1": 20611, "n_sents": 237, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 988, "title": "Diffusion SigFormer for Interference Time-series Signal Recognition", "abs": "The various interferences in the actual environment make electromagnetic signal recognition challenging, and this topic has extremely important application value.\nIn this paper, a novel interference signal recognition transformer is proposed, named Diffusion SigFormer.\nFirstly, we explored the interference law of electromagnetic signals and designed a signal interference mechanism. \nSecondly, diffusion signal denoising modulewas proposed to denoise the input interference signal. We also use various types of noise to improve its denoising effect on electromagnetic signals.\nThirdly, SigFormer is designed to extract and classify the denoised signal.\nFor the characteristics of electromagnetic signals, SigFormer leverages 1-D Patch Embedding and combines transformer with convolution. \nFinally, we conducted experimental verification on datasets RML2016.10a, RML2016.10b and BT dataset. \nThe experimental results show that the proposed method has excellent anti-interference ability.", "keywords": [ "Anti-interference electromagnetic signal recognition", "diffusion", "SigFormer", "modulation", "bluetooth" ], "rating_list": [ 1, 1, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "LppenBe0fr", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 8, "author_info_list": [ { "name": "Chengzhi Wu", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "DE", "position": "PhD student" }, { "name": "Hao Fu", "gender": "Male", "institution": "Rheinisch Westf\u00e4lische Technische Hochschule Aachen", "country": "DE", "position": "PhD student" }, { "name": "Jiaming Zhang", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Julius Pfrommer", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "Lecturer" }, { "name": "Junwei Zheng", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "DE", "position": "PhD student" }, { "name": "J\u00fcrgen Beyerer", "gender": "unknown", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "Full Professor" }, { "name": "Yuxin Wan", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "MS student" }, { "name": "Zeyun Zhong", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 5, "n_ref_uni": 11, "n_ref": 15, "n_ref_all": 44, "n_fig": 17, "n_tab": 13, "L_tab": 2449, "n_element_tab": 462, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 1012, "n_element_tab_1": 143, "formula_len_all": 793, "formula_len_all_1": 183, "len_all": 164458, "len_all_1": 53967, "len_abs": 1820, "len_title": 178, "len_sents": 47452, "len_sents_1": 23421, "n_sents": 409, "n_sents_1": 195, "venue": "ICLR2025", "decision": "withdraw", "L_title": 130, "L_abs": 1828, "title": "SAMBLE: Learning Shape-Specific Sampling Strategies for Point Cloud Shapes with Sparse Attention Map and Adaptive Bin Partitioning", "abs": "Point cloud sampling plays a pivotal role in facilitating efficient analysis of large-scale point clouds. Recently, learning-to-sample methods have garnered growing interest from the community, particularly for their ability to be jointly trained with downstream tasks. However, previous learning-based sampling methods either lead to unrecognizable sampling patterns by generating a new point cloud or biased sampled results by focusing excessively on shape details. Moreover, they all fail to take the natural point distribution variations over different shapes into consideration and learn a similar sampling strategy for all point clouds. In this paper, we propose a Sparse Attention Map and Bin-based Learning method (termed SAMBLE) to learn shape-specific sampling strategies for point cloud shapes, striking a superior balance between the overall shape outline and intricate local details for the sampling process. In particular, we first propose sparse attention map by integrating both local and global information. Based on this, multiple point-wise sampling score computation methods are proposed and explored by leveraging heatmaps as a guiding tool. Subsequently, we introduce a binning strategy that partitions points within each point cloud based on these scores. Finally, additional learnable tokens are introduced during the attention computation phase to acquire sampling weights for each bin, thereby enabling the development of shape-specific sampling strategies for an optimized sampling process. Extensive experiments demonstrate that our method adeptly strikes a refined balance between sampling edge points for local details and preserving uniformity in the global shape, leading to superior performance across common point cloud downstream tasks and even in scenarios involving few-point cloud sampling.", "keywords": [ "3D Shapes", "Point Cloud Sampling", "Sparse Attention Map", "Bin Partitioning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "LphpWGimIa", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Arthur Conmy", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Connor Kissane", "gender": "unknown", "institution": "Independent", "country": "US", "position": "Researcher" }, { "name": "Joseph Isaac Bloom", "gender": "Male", "institution": "Department of Science, Technology and Innovation", "country": "GB", "position": "Researcher" }, { "name": "Neel Nanda", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Robert Krzyzanowski", "gender": "Male", "institution": "Independent", "country": "", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 5, "n_ref_uni": 51, "n_ref": 166, "n_ref_all": 231, "n_fig": 25, "n_tab": 4, "L_tab": 1820, "n_element_tab": 316, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 633, "n_element_tab_1": 98, "formula_len_all": 501, "formula_len_all_1": 295, "len_all": 244607, "len_all_1": 61264, "len_abs": 1628, "len_title": 109, "len_sents": 81928, "len_sents_1": 29240, "n_sents": 623, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1633, "title": "Interpreting Attention Layer Outputs with Sparse Autoencoders", "abs": "Decomposing model activations into interpretable components is a key open problem in mechanistic interpretability. Sparse autoencoders (SAEs) are a popular method for decomposing the internal activations of trained transformers into sparse, interpretable features, and have been applied to MLP layers and the residual stream. In this work we train SAEs on attention layer outputs and show that also here SAEs find a sparse, interpretable decomposition. We demonstrate this on transformers from several model families and up to 2B parameters. We perform a qualitative study of the features computed by attention layers, and find multiple families: long-range context, short-range context and induction features. We qualitatively study the role of every head in GPT-2 Small, and estimate that at least 90% of the heads are polysemantic, i.e. have multiple unrelated roles. Further, we show that sparse autoencoders are a useful tool that enable researchers to explain model behavior in greater detail than prior work. For example, we explore the mystery of why models have so many seemingly redundant induction heads, use SAEs to motivate the hypothesis that some are long-prefix whereas others are short-prefix, and confirm this with more rigorous analysis. We use our SAEs to analyze the computation performed by the Indirect Object Identification circuit (Wang et al., 2023), validating that the SAEs find causally meaningful intermediate variables, and deepening our understanding of the semantics of the circuit. We open-source the trained SAEs and a tool for exploring arbitrary prompts through the lens of Attention Output SAEs.", "keywords": [ "mechanistic interpretability", "llm interpretability" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 1, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 1, 3, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "Lp40Z40N07", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Biao Gong", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Chunhua Shen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Hao Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Hao OUYANG", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Qiuyu Wang", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Wen Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yujun Shen", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhekai Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "kecheng zheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 56, "n_ref": 123, "n_ref_all": 158, "n_fig": 27, "n_tab": 3, "L_tab": 2076, "n_element_tab": 185, "n_fig_1": 16, "n_tab_1": 1, "L_tab_1": 438, "n_element_tab_1": 25, "formula_len_all": 323, "formula_len_all_1": 219, "len_all": 158987, "len_all_1": 61391, "len_abs": 527, "len_title": 87, "len_sents": 38871, "len_sents_1": 27463, "n_sents": 327, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 39, "L_abs": 1256, "title": "Framer: Interactive Frame Interpolation", "abs": "We propose Framer for interactive frame interpolation, which targets producing smoothly transitioning frames between two images as per user creativity. Concretely, besides taking the start and end frames as inputs, our approach supports customizing the transition process by tailoring the trajectory of some selected keypoints. Such a design enjoys two clear benefits. First, incorporating human interaction mitigates the issue arising from numerous possibilities of transforming one image to another, and in turn enables finer control of local motions. Second, as the most basic form of interaction, keypoints help establish the correspondence across frames, enhancing the model to handle challenging cases (e.g., objects on the start and end frames are of different shapes and styles). It is noteworthy that our system also offers an \"autopilot\" mode, where we introduce a module to estimate the keypoints and refine the trajectory automatically, to simplify the usage in practice. Extensive experimental results demonstrate the appealing performance of Framer on various applications, such as image morphing, time-lapse video generation, cartoon interpolation, etc. The code, the model, and the interface will be released to facilitate further research.", "keywords": [ "Video Frame Interpolation; Interactive; Diffusion Model; Correspondence Modeling" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 4, 3, 2 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "LoXJlAW3gU", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Andrey Shevtsov", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "", "position": "Researcher" }, { "name": "Dmitry Vetrov", "gender": "Male", "institution": "Constructor University", "country": "DE", "position": "Full Professor" }, { "name": "Fedor Nikolaev", "gender": "Male", "institution": "Epistemic AI", "country": "US", "position": "Researcher" }, { "name": "Nikita Ivanisenko", "gender": "Male", "institution": "InstaDeep", "country": "DE", "position": "Researcher" }, { "name": "Olga Kardymon", "gender": "unknown", "institution": "AIRI", "country": "", "position": "Researcher" }, { "name": "Pavel Strashnov", "gender": "unknown", "institution": "Artificial Intelligence Research Institute (AIRI)", "country": "RU", "position": "Researcher" }, { "name": "Viacheslav Meshchaninov", "gender": "Male", "institution": "Higher School of Economics", "country": "RU", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 3, "n_ref_uni": 53, "n_ref": 122, "n_ref_all": 160, "n_fig": 15, "n_tab": 14, "L_tab": 8187, "n_element_tab": 899, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1646, "n_element_tab_1": 211, "formula_len_all": 599, "formula_len_all_1": 143, "len_all": 248871, "len_all_1": 72573, "len_abs": 1475, "len_title": 117, "len_sents": 69056, "len_sents_1": 34514, "n_sents": 508, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1412, "title": "Diffusion on language model encodings for protein sequence generation", "abs": "Protein design necessitates a profound understanding of the intricate nature of the protein universe. While many efforts focus on conditional generation or specific protein families, the foundational task of unconditional generation remains underexplored and underappreciated. Existing models still struggle to achieve both high quality and diversity in generated protein sequences. To address this gap, this research introduces DiMA, a novel model that leverages latent diffusion on representations derived from the protein language model, ESM-2, to generate amino acid sequences. We quantitatively investigate the impact of components of the latent diffusion model, revealing their contributions to superior protein generation performance. Extensive evaluations using multiple metrics across two protein modalities showcase DiMA's superior quality, diversity, and distribution matching capabilities compared to leading autoregressive transformer-based and discrete diffusion models, while utilizing ten times fewer parameters. Our approach consistently produces novel, diverse protein sequences that accurately reflect the inherent structural and functional diversity of the protein space. Furthermore, we demonstrate the conditional generation capabilities of our method. Our work advances the field of protein design by providing a robust framework for scalable and high-quality protein sequence generation.", "keywords": [ "diffusion", "protein language models", "protein generation" ], "rating_list": [ 3, 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 5, 3, 4, 3, 4 ] }, { "paper_id": "Lo98rDyfl8", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Diganta Misra", "gender": "Male", "institution": "Max-Planck-Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Hyeonbeom Choi", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Undergrad student" }, { "name": "Jonghyun Choi", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" }, { "name": "Minhyuk Seo", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Researcher" }, { "name": "Seongwon Cho", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Min Jae Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Seon Joo Kim", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 6, "n_ref_uni": 157, "n_ref": 353, "n_ref_all": 402, "n_fig": 21, "n_tab": 23, "L_tab": 5790, "n_element_tab": 360, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 4996, "n_element_tab_1": 148, "formula_len_all": 979, "formula_len_all_1": 389, "len_all": 454037, "len_all_1": 72512, "len_abs": 2153, "len_title": 121, "len_sents": 127940, "len_sents_1": 32088, "n_sents": 994, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1221, "title": "Rainbow Generator: Generating Diverse Data for Name Only Continual Learning", "abs": "Requiring extensive human supervision is often impractical for continual learning due to its cost, leading to the emergence of \u2018name-only continual learning\u2019 that only provides the name of new concepts (e.g., classes) without providing supervised samples. To address the task, recent approach uses web-scraped data but results in issues such as data imbalance, copyright, and privacy concerns. To overcome the limitations of both human supervision and webly supervision, we propose Generative name only Continual Learning (GenCL) using generative models for the name only continual learning. But na\u00efve application of generative models results in limited diversity of generated data. So, we specifically propose a diverse prompt generation method, HIerarchical Recurrent Prompt Generation (HIRPG) as well as COmplexity-NAvigating eNsembler (CONAN) that selects samples with minimal overlap from multiple generative models. We empirically validate that the proposed GenCL outperforms prior arts, even a model trained with fully supervised data, in various tasks including image recognition and multi-modal visual reasoning. Data generated by GenCL is available at https://anonymous.4open.science/r/name-only-continual-E079.", "keywords": [ "continual learning", "generative model" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 3, 4, 2, 4 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "Lnuy691O8Q", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Marzyeh Ghassemi", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Sanqiang Zhao", "gender": "Male", "institution": "Zoom", "country": "US", "position": "Researcher" }, { "name": "Shujian Zhang", "gender": "unknown", "institution": "University of Texas, Austin", "country": "", "position": "PhD student" }, { "name": "Wenxuan Zhou", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Researcher" }, { "name": "Yuxin Xiao", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 63, "n_ref": 112, "n_ref_all": 131, "n_fig": 4, "n_tab": 1, "L_tab": 485, "n_element_tab": 42, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 993, "n_element_tab_1": 110, "formula_len_all": 584, "formula_len_all_1": 584, "len_all": 151341, "len_all_1": 60174, "len_abs": 1830, "len_title": 117, "len_sents": 34329, "len_sents_1": 28294, "n_sents": 218, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1837, "title": "SFTMix: Elevating Language Model Instruction Tuning with Mixup Recipe", "abs": "To induce desired behaviors in large language models (LLMs) for interaction-driven tasks, the instruction-tuning stage typically trains LLMs on instruction-response pairs using the next-token prediction (NTP) loss. Previous work aiming to improve instruction-tuning performance often emphasizes the need for higher-quality supervised fine-tuning (SFT) datasets, which typically involves expensive data filtering with proprietary LLMs or labor-intensive data generation by human annotators. However, these approaches do not fully leverage the datasets' intrinsic properties, resulting in high computational and labor costs, thereby limiting scalability and performance gains. In this paper, we propose SFTMix, a novel recipe that elevates instruction-tuning performance beyond the conventional NTP paradigm, without the need for well-curated datasets. Observing that LLMs exhibit uneven confidence across the semantic representation space, we argue that examples with different confidence levels should play distinct roles during the instruction-tuning process. Based on this insight, SFTMix leverages training dynamics to identify examples with varying confidence levels, then applies a Mixup-based regularization to mitigate overfitting on confident examples while propagating supervision signals to improve learning on relatively unconfident ones. This approach enables SFTMix to significantly outperform NTP across a wide range of instruction-following and healthcare domain-specific SFT tasks, demonstrating its adaptability to diverse LLM families and scalability to datasets of any size. Comprehensive ablation studies further verify the robustness of SFTMix's design choices, underscoring its versatility in consistently enhancing performance across different LLMs and datasets in broader natural language processing applications.", "keywords": [ "Supervised Fine-Tuning", "Large Language Model" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 1, 2, 3, 4 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "LnRegTxsa4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Xingtao Ling", "gender": "Male", "institution": "ShenZhen University", "country": "CN", "position": "MS student" }, { "name": "Yingying Zhu", "gender": "Female", "institution": "Shenzhen University", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 20, "n_ref": 47, "n_ref_all": 67, "n_fig": 8, "n_tab": 3, "L_tab": 1272, "n_element_tab": 125, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 639, "n_element_tab_1": 85, "formula_len_all": 761, "formula_len_all_1": 501, "len_all": 89725, "len_all_1": 52639, "len_abs": 1698, "len_title": 86, "len_sents": 26197, "len_sents_1": 23117, "n_sents": 198, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 132, "L_abs": 1716, "title": "Improving Cross-view Object Geo-localization: A Dual Attention Approach with Cross-view Interaction and Multi-Scale Spatial Features", "abs": "Cross-view object geo-localization has recently gained attention due to potential applications. Existing methods aim to capture spatial dependencies of query objects between different views through attention mechanisms to obtain spatial relationship feature maps, which are then used to predict object locations. Although promising, these approaches fail to effectively transfer information between views and do not further refine the spatial relationship feature maps. This results in the model erroneously focusing on irrelevant edge noise, thereby affecting localization performance. To address these limitations, we introduce a **Cross-view and Cross-attention Module (CVCAM)**, which performs multiple iterations of interaction between the two views, enabling continuous exchange and learning of contextual information about the query object from both perspectives. This facilitates a deeper understanding of cross-view relationships while suppressing the edge noise unrelated to the query object. Furthermore, we integrate a **Multi-head Spatial Attention Module (MHSAM)**, which employs convolutional kernels of various sizes to extract multi-scale spatial features from the feature maps containing implicit correspondences, further enhancing the feature representation of the query object. Additionally, given the scarcity of datasets for cross-view object geo-localization, we created a new dataset called **G2D** for the \"Ground\u2192Drone\" localization task, enriching existing datasets and filling the gap in \"Ground\u2192Drone\" localization task. Extensive experiments on the CVOGL and G2D datasets demonstrate that our proposed method achieves high localization accuracy, surpassing the current state-of-the-art.", "keywords": [ "cross-view", "object localization", "attention mechanism" ], "rating_list": [ 5, 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "LnN5UdhcjT", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Fuchen Long", "gender": "Male", "institution": "JD.com", "country": "CN", "position": "Researcher" }, { "name": "Tao Mei", "gender": "Male", "institution": "JD Explore Academy", "country": "CN", "position": "Technical Vice President" }, { "name": "Ting Yao", "gender": "Male", "institution": "JD AI Research", "country": "CN", "position": "Researcher" }, { "name": "Wu Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yingwei Pan", "gender": "Male", "institution": "JD.com", "country": "CN", "position": "Researcher" }, { "name": "Zhaofan Qiu", "gender": "Male", "institution": "JD.com", "country": "CN", "position": "Researcher" }, { "name": "Zhongwei Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 25, "n_ref": 67, "n_ref_all": 85, "n_fig": 7, "n_tab": 4, "L_tab": 636, "n_element_tab": 73, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 358, "n_element_tab_1": 42, "formula_len_all": 436, "formula_len_all_1": 435, "len_all": 127243, "len_all_1": 57994, "len_abs": 1792, "len_title": 107, "len_sents": 37339, "len_sents_1": 27359, "n_sents": 267, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1800, "title": "Region-wise Motion Controller for Image-to-Video Generation", "abs": "Animating images with interactive motion control has garnered popularity for image-to-video (I2V) generation. Modern approaches typically regard the condition of Gaussian filtered point-wise trajectory as sole motion control signal. Nevertheless, such flow approximation of trajectory via Gaussian kernel severely limits the controllable capacity of fine-grained movement, and commonly fails to disentangle object and camera moving. To alleviate these, we present ReMoCo, a new recipe of region-wise motion controller that novelly leverages precise region-wise trajectory and motion mask to regulate fine-grained motion synthesis and identify exact target motion category (i.e., object or camera moving), respectively. Technically, ReMoCo first estimates the flow maps on each training video via a tracking model, and then samples the region-wise trajectories from multiple local regions to simulate inference scenario. Instead of approximating flow distribution via Gaussian filtering, our region-wise trajectory preserves original flow information at local area and thus manages to characterize fine-grained movement. A motion mask is simultaneously derived from the predicted flow maps to present holistic motion dynamics. To pursue natural and controllable motion generation, ReMoCo further strengthens video denoising with additional conditions of region-wise trajectory and motion mask in a feature modulation manner. More remarkably, we meticulously construct a benchmark called ReMoCo-Bench, which consists of 1.1K real-world user-annotated image-trajectory pairs, for the evaluation of both fine-grained and object-level motion synthesis in I2V generation. Extensive experiments conducted on WebVid-10M and ReMoCo-Bench demonstrate the effectiveness of our ReMoCo for precise motion control.", "keywords": [ "Diffusion Models", "Image-to-Video Generation", "Motion Control" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "LnKDcqOfgy", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Alexander Conzelmann", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Robert Bamler", "gender": "Male", "institution": "University of Tuebingen", "country": "", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 38, "n_ref": 69, "n_ref_all": 95, "n_fig": 6, "n_tab": 1, "L_tab": 1095, "n_element_tab": 41, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1095, "n_element_tab_1": 41, "formula_len_all": 1050, "formula_len_all_1": 1051, "len_all": 121191, "len_all_1": 65854, "len_abs": 1518, "len_title": 131, "len_sents": 32825, "len_sents_1": 30420, "n_sents": 212, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1523, "title": "Rate/Distortion Constrained Model Quantization for Efficient Storage and Inference", "abs": "The proliferation of large pre-trained neural networks has recently revived research in both quantization of network weights (for faster inference), and in their\ncompression (to reduce file sizes). However, there has so far been little idea transfer between the two lines of research. In this paper, we combine techniques from\nquantization and compression to propose an efficient and highly effective post-training compression method for large neural networks. Our method extends the\nrecently published quantization method OPTQ (Frantar et al., 2023) with a tunable\nrate/distortion trade-off by introducing a cost per bit into OPTQ's rounding\noperation. Crucially, we estimate the bit rate based on the predictive model used\nin the state-of-the-art neural network compression method NNCodec (Becking\net al., 2023). In our experiments with several standard pre-trained networks from\nthe computer vision community, our method leads to significantly (up to 2.7x)\nsmaller file sizes than NNCodec at equal model performance, generally compressing to less than half a bit per network weight and implicitly pruning insignificant weights.\nAdditionally, and in contrast to NNcodec, our method offers the same opportunities for inference speed-ups as OPTQ. By proving that file size and inference\ncost can be reduced simultaneously, we hope that our contribution shows a path\ntowards deploying large neural networks on end-user devices, alleviating privacy\nconcerns, regulatory constraints, and dependency on large service providers.", "keywords": [ "quantization", "model compression", "rate-distortion theory", "compression" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "Llh6CinTiy", "primary_area": "learning on time series and dynamical systems", "n_author": 8, "author_info_list": [ { "name": "Daniel Giles", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Postdoc" }, { "name": "Sibo Cheng", "gender": "Male", "institution": "Ecole Nationale des Ponts et Chausees", "country": "FR", "position": "Assistant Professor" }, { "name": "Xiao Xue", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Researcher" }, { "name": "Xiaohang Tang", "gender": "Male", "institution": "University College London", "country": "GB", "position": "PhD student" }, { "name": "Xiaoyuan Cheng", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Yi He", "gender": "unknown", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Yiming Yang", "gender": "unknown", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Yukun Hu", "gender": "unknown", "institution": "University College London, University of London", "country": "", "position": "Associate Professor" } ], "n_formula": 56, "n_formula_1": 17, "n_ref_uni": 65, "n_ref": 99, "n_ref_all": 130, "n_fig": 6, "n_tab": 6, "L_tab": 5204, "n_element_tab": 271, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 2679, "n_element_tab_1": 130, "formula_len_all": 4364, "formula_len_all_1": 1288, "len_all": 235745, "len_all_1": 76979, "len_abs": 2106, "len_title": 90, "len_sents": 71165, "len_sents_1": 34790, "n_sents": 545, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 30, "L_abs": 2043, "title": "Learning Chaos In A Linear Way", "abs": "Learning long-term behaviors in chaotic dynamical systems, such as turbulent flows and climate modelling, is challenging due to their inherent instability and unpredictability. These systems exhibit positive Lyapunov exponents, which significantly hinder accurate long-term forecasting. As a result, understanding long-term statistical behavior is far more valuable than focusing on short-term accuracy. While autoregressive deep sequence models have been applied to capture long-term behavior, they often lead to exponentially increasing errors in learned dynamics. To address this, we shift the focus from simple prediction errors to preserving an invariant measure in dissipative chaotic systems. These systems have attractors, where trajectories settle, and the invariant measure is the probability distribution on attractors that remains unchanged under dynamics. Existing methods generate long trajectories of dissipative chaotic systems by aligning invariant measures, but it is not always possible to obtain invariant measures for arbitrary datasets. We propose the Poincar\u00e9 Flow Neural Network (PFNN), a novel operator learning framework designed to capture behaviors of chaotic systems without any explicit knowledge of the invariant measure. PFNN employs an auto-encoder to map the chaotic system to a finite-dimensional feature space, effectively linearizing the chaotic evolution. It then learns the linear evolution operators to match the physical dynamics by addressing two critical properties in dissipative chaotic systems: (1) contraction, the system\u2019s convergence toward its attractors, and (2) measure invariance, trajectories on the attractors following a probability distribution invariant to the dynamics. \nOur experiments on a variety of chaotic systems, including Lorenz systems, Kuramoto-Sivashinsky equation and Navier\u2013Stokes equation, demonstrate that PFNN has more accurate predictions and physical statistics compared to competitive baselines including the Fourier Neural Operator and the Markov Neural Operator.", "keywords": [ "Dynamical systems", "operator learning", "chaos", "physics-informed learning" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "LlZ929lua7", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Axel Brando", "gender": "Male", "institution": "Barcelona Supercomputing Center", "country": "ES", "position": "Postdoc" }, { "name": "David Meger", "gender": "Male", "institution": "McGill University", "country": "", "position": "Associate Professor" }, { "name": "Juan Camilo Gamboa Higuera", "gender": "Male", "institution": "McGill University", "country": "", "position": "PhD student" }, { "name": "Lucas Berry", "gender": "unknown", "institution": "McGill University, McGill University", "country": "CA", "position": "PhD student" }, { "name": "Wei-Di Chang", "gender": "unknown", "institution": "McGill University", "country": "", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 60, "n_ref": 87, "n_ref_all": 113, "n_fig": 14, "n_tab": 7, "L_tab": 1990, "n_element_tab": 140, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 467, "n_element_tab_1": 37, "formula_len_all": 738, "formula_len_all_1": 717, "len_all": 159212, "len_all_1": 60627, "len_abs": 1191, "len_title": 122, "len_sents": 42299, "len_sents_1": 27145, "n_sents": 297, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1063, "title": "Seeing the Unseen: How EMoE Unveils Bias in Text-to-Image Diffusion Models", "abs": "Estimating uncertainty in text-to-image diffusion models is challenging due to their large parameter counts (often exceeding 100 million) and operation in complex, high-dimensional spaces with virtually infinite input possibilities. In this paper, we propose EMoE, a novel framework for efficiently estimating epistemic uncertainty in diffusion models. EMoE leverages pre-trained networks without requiring additional training, enabling direct uncertainty estimation from a prompt. We introduce a novel latent space within the diffusion process that captures model uncertainty better during the first denoising step than existing methods. Experimental results on the COCO dataset demonstrate EMoE's effectiveness, showing a strong correlation between uncertainty and image quality. Additionally, EMoE identifies under-sampled languages and regions with higher uncertainty, revealing hidden biases related to linguistic representation. This capability demonstrates the relevance of EMoE as a tool for addressing fairness and accountability in AI-generated content.", "keywords": [ "Uncertainty Quantification", "Text-to-Image Modeling", "Ensembles" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "LlE61BEYpB", "primary_area": "optimization", "n_author": 11, "author_info_list": [ { "name": "David Lisbonne", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Undergrad student" }, { "name": "Gregory Kielian", "gender": "Not Specified", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Guanchen Tao", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "PhD student" }, { "name": "Haoran Cheng", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Undergrad student" }, { "name": "Junyi Luo", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "MS student" }, { "name": "Kauna Lei", "gender": "Female", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Researcher" }, { "name": "Mehdi Saligane", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" }, { "name": "Michael Moffatt", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "MS student" }, { "name": "Qilong Wang", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Undergrad student" }, { "name": "Shiwei Liu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Xinting Jiang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 13, "n_ref": 21, "n_ref_all": 28, "n_fig": 3, "n_tab": 1, "L_tab": 118, "n_element_tab": 14, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 179, "formula_len_all_1": 179, "len_all": 63576, "len_all_1": 44821, "len_abs": 1958, "len_title": 68, "len_sents": 21381, "len_sents_1": 19683, "n_sents": 162, "n_sents_1": 149, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1990, "title": "FLARE: Fine-tuned Long-context Acceleration with ReLU-enhanced FIRE", "abs": "Deploying large language models (LLMs) on resource-constrained edge devices is challenging due to computational bottlenecks, memory bottlenecks, and -- for long-contexts -- specifically the Softmax operation in the attention mechanism. While using ReLU in place of Softmax has been explored, and FIRE as an alternative to RoPE has been explored for models trained from scratch, there has been little work towards exploring fine-tuning models to utilize these efficient algorithms, or the combination of the two.\n\nIn this paper, we contribute FLARE, a method for fusing Rectified Linear Activations (ReLU) with Relative Encodings (specifically FIRE), and we share a particular recipe which allows these to be fine-tuned effectively into existing models and fused to create efficient long-context inference. Following this recipe yields markedly better validation loss, long-context inference speed, and successfully introduces the property of length-generalization -- the property where the model gains high accuracy for contexts lengths several times larger than trained -- unlike RoPE -- without further fine-tuning. \n\nOnce FIRE and ReLU are both fine-tuned into a model, we show these can be mathematically fused into a single, more efficient operation, which on average was found to shave 98.9\\% of FIRE operations and produce a Probability matrix with 98.9\\% zeros in its lower-triangle.\n\nFinally, we benchmark inference speed improvements for custom hardware as well with custom CUDA kernels. Using Power, Performance, and Area (PPA) analysis, we show that FLARE operates at eight times the frequency of Softmax while consuming only 0.1\\% of the power and 0.11\\% of the energy per cycle. Our custom CUDA Kernel shows 3.8x faster operation than Softmax FlashAttention. We believe this shows the potential of fine-tuning new algorithms in pre-trained models, and we share our fine-tuning recipes, code and custom hardware designs at \\url{https://anonymous.4open.science/r/nanoGPTBD54}.", "keywords": [ "FIRE", "Functional Interpolation for Relative Position Encoding", "fine-tune", "fine-tuning", "ReLU", "Softmax", "Softplus", "Softmax alternatives", "long context", "transformer", "large language model", "edge device", "Flash Attention" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 4, 3, 1 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "LkzuPorQ5L", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Dawei Cheng", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Associate Professor" }, { "name": "Guancheng Wan", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Researcher" }, { "name": "Guibin Zhang", "gender": "Not Specified", "institution": "Tongji University", "country": "CN", "position": "Undergrad student" }, { "name": "Jeffrey Xu Yu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Kun Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Sukwon Yun", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Yanwei Yue", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhixun Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "PhD student" } ], "n_formula": 32, "n_formula_1": 20, "n_ref_uni": 52, "n_ref": 130, "n_ref_all": 180, "n_fig": 26, "n_tab": 16, "L_tab": 5058, "n_element_tab": 369, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1051, "n_element_tab_1": 79, "formula_len_all": 3352, "formula_len_all_1": 1915, "len_all": 284135, "len_all_1": 69392, "len_abs": 1516, "len_title": 132, "len_sents": 77502, "len_sents_1": 29772, "n_sents": 577, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1546, "title": "Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems", "abs": "Recent advancements in large language model (LLM)-powered agents have shown that collective intelligence can significantly outperform individual capabilities, largely attributed to the meticulously designed inter-agent communication topologies. Though impressive in performance, existing multi-agent pipelines inherently introduce substantial token overhead, as well as increased economic costs, which pose challenges for their large-scale deployments. In response to this challenge, we propose an economical, simple, and robust multi-agent communication framework, termed $\\texttt{AgentPrune}$, which can seamlessly integrate into mainstream multi-agent systems and prunes redundant or even malicious communication messages. Technically, $\\texttt{AgentPrune}$ is the first to identify and formally define the $\\textit{Communication Redundancy}$ issue present in current LLM-based multi-agent pipelines, and efficiently performs one-shot pruning on the spatial-temporal message-passing graph, yielding a token-economic and high-performing communication topology.\nExtensive experiments across six benchmarks demonstrate that $\\texttt{AgentPrune}$ $\\textbf{(I)}$ achieves comparable results as state-of-the-art topologies at merely $\\\\$5.6$ cost compared to their $\\\\$43.7$, $\\textbf{(II)}$ integrates seamlessly into existing multi-agent frameworks with $28.1\\\\%\\sim72.8\\\\%\\downarrow$ token reduction, and $\\textbf{(III)}$ successfully defend against two types of agent-based adversarial attacks with $3.5\\\\%\\sim10.8\\\\%\\uparrow$ performance boost.", "keywords": [ "Multi-agent collaboration", "sparsification", "LLM agents" ], "rating_list": [ 3, 8, 5, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "LjvIJFCa5J", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Daichi Azuma", "gender": "Male", "institution": "Sony Semiconductor Solutions", "country": "JP", "position": "Engineer" }, { "name": "Jungdae Lee", "gender": "Male", "institution": "Tokyo Institute of Technology, Tokyo Institute of Technology", "country": "JP", "position": "MS student" }, { "name": "Koya Sakamoto", "gender": "Male", "institution": "Kyoto University", "country": "JP", "position": "MS student" }, { "name": "Nakamasa Inoue", "gender": "Male", "institution": "Institute of Science Tokyo", "country": "JP", "position": "Associate Professor" }, { "name": "Shuhei Kurita", "gender": "unknown", "institution": "National Institute of Informatics", "country": "JP", "position": "Assistant Professor" }, { "name": "Taiki Miyanishi", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Researcher" }, { "name": "Yutaka Matsuo", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 54, "n_ref": 96, "n_ref_all": 132, "n_fig": 9, "n_tab": 12, "L_tab": 4165, "n_element_tab": 287, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1290, "n_element_tab_1": 46, "formula_len_all": 102, "formula_len_all_1": 0, "len_all": 196094, "len_all_1": 64547, "len_abs": 1731, "len_title": 124, "len_sents": 53486, "len_sents_1": 31797, "n_sents": 389, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1739, "title": "CityNav: Language-Goal Aerial Navigation Dataset Using Geographic Information", "abs": "Vision-and-language navigation (VLN) aims to guide autonomous agents through real-world environments by integrating visual and linguistic cues. Despite notable advancements in ground-level navigation, the exploration of aerial navigation using these modalities remains limited. This gap primarily arises from a lack of suitable resources for real-world, city-scale aerial navigation studies. To remedy this gap, we introduce CityNav, a novel dataset explicitly designed for language-guided aerial navigation in photorealistic 3D environments of real cities. CityNav comprises 32k natural language descriptions paired with human demonstration trajectories, collected via a newly developed web-based 3D simulator. Each description identifies a navigation goal, utilizing the names and locations of landmarks within actual cities. As an initial step toward addressing this challenge, we provide baseline models of navigation agents that incorporate an internal 2D spatial map representing landmarks referenced in the descriptions. We have benchmarked the latest aerial navigation methods alongside our proposed baseline model on the CityNav dataset. The findings are revealing: (i) our aerial agent model trained on human demonstration trajectories, outperform those trained on shortest path trajectories by a large margin; (ii) incorporating 2D spatial map information markedly and robustly enhances navigation performance at a city scale; (iii) despite the use of map information, our challenging CityNav dataset reveals a persistent performance gap between our baseline models and human performance. To foster further research in aerial VLN, we have made the dataset and code available at https://anonymous.4open.science/w/city-nav-77E3/.", "keywords": [ "City-scale 3D Vision", "Aerial Navigation", "3D Vision and Language" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "LjQDYcFWmN", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Itay Lavie", "gender": "unknown", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "MS student" }, { "name": "Zohar Ringel", "gender": "Male", "institution": "Hebrew University of Jerusalem, Israel", "country": "", "position": "Associate Professor" } ], "n_formula": 61, "n_formula_1": 28, "n_ref_uni": 42, "n_ref": 99, "n_ref_all": 136, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4228, "formula_len_all_1": 1685, "len_all": 162950, "len_all_1": 72515, "len_abs": 975, "len_title": 125, "len_sents": 51153, "len_sents_1": 33444, "n_sents": 394, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 976, "title": "Symmetric Kernels with Non-Symmetric Data: A Data-Agnostic Learnability Bound", "abs": "Kernel ridge regression (KRR) and Gaussian processes (GPs) are fundamental tools in statistics and machine learning, with recent applications to highly over-parameterized deep neural networks. The ability of these tools to learn a target function is directly related to the eigenvalues of their kernel sampled on the input data. Targets having support on higher eigenvalues are more learnable. While kernels are often highly symmetric objects, the data is often not. \nThus, kernel symmetry seems to have little to no bearing on the above eigenvalues or learnability, making spectral analysis on real-world data challenging.\nHere, we show that contrary to this common lure, one may use eigenvalues and eigenfunctions associated with highly idealized data measures to bound learnability on realistic data. As a demonstration, we give a theoretical lower bound on the sample complexity of copying heads for kernels associated with generic transformers acting on natural language.", "keywords": [ "Kernel Ridge Regression", "Gaussian Process", "NTK", "NNGP", "Deep Learning Theory", "Symmetry", "Sample Complexity", "Learnability", "Bayesian Inference" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 2, 4, 5 ] }, { "paper_id": "LilItwL2br", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Daeyoung Choi", "gender": "Male", "institution": "The Cyber University of Korea", "country": "KR", "position": "Assistant Professor" }, { "name": "Gyuejeong Lee", "gender": "Male", "institution": "SAKAK Inc.", "country": "KR", "position": "Researcher" } ], "n_formula": 24, "n_formula_1": 15, "n_ref_uni": 34, "n_ref": 66, "n_ref_all": 109, "n_fig": 9, "n_tab": 2, "L_tab": 1618, "n_element_tab": 121, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1618, "n_element_tab_1": 121, "formula_len_all": 975, "formula_len_all_1": 552, "len_all": 122833, "len_all_1": 67326, "len_abs": 1393, "len_title": 127, "len_sents": 35422, "len_sents_1": 29399, "n_sents": 286, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1398, "title": "Enhancing Prototype-Based Federated Learning with Structured Sparse Prototypes", "abs": "Prototype-Based Federated Learning (PBFL) has gained attention for its communication efficiency, privacy preservation, and personalization capabilities in resource-constrained environments. Despite these advantages, PBFL methods face challenges, including high communication costs for high-dimensional prototypes and numerous classes, privacy concerns during aggregation, and uniform knowledge distillation in heterogeneous data settings.\nTo address these issues, we introduce three novel methods, each targeting a specific PBFL stage: 1) Class-wise Prototype Sparsification (CPS) reduces communication costs by creating structured sparse prototypes, where each prototype utilizes only a subset of representation layer dimensions. 2) Privacy-Preserving Prototype Aggregation (PPA) enhances privacy by eliminating the transmission of client class distribution information when aggregating local prototypes. 3) Class-Proportional Knowledge Distillation (CPKD) improves personalization by modulating the distillation strength for each class based on clients' local data distributions.\nWe integrate these three methods into two foundational PBFL approaches and conduct experimental evaluations. The results demonstrate that this integration achieves up to 10\u00d7 and 4\u00d7 reductions in communication costs while outperforming the original and most communication-efficient approaches evaluated, respectively.", "keywords": [ "federated learning", "prototype-based federated learning", "distributed machine learning", "structured sparsity" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "LikKyNlzgP", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Evgeny Burnaev", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Full Professor" }, { "name": "Grigoriy Ksenofontov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "PhD student" }, { "name": "Nikita Gushchin", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Nikita Maksimovich Kornilov", "gender": "Not Specified", "institution": "Moscow Institute of Physics and Technology", "country": "", "position": "Undergrad student" }, { "name": "Sergei Kholkin", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Alexander Korotin", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Assistant Professor" }, { "name": "Li David", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "MS student" } ], "n_formula": 131, "n_formula_1": 23, "n_ref_uni": 32, "n_ref": 119, "n_ref_all": 179, "n_fig": 11, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 471, "n_element_tab_1": 111, "formula_len_all": 13435, "formula_len_all_1": 2356, "len_all": 200802, "len_all_1": 70942, "len_abs": 1183, "len_title": 137, "len_sents": 60892, "len_sents_1": 29494, "n_sents": 519, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1186, "title": "Diffusion & Adversarial Schr\u00f6dinger Bridges via Iterative Proportional Markovian Fitting", "abs": "The Iterative Markovian Fitting (IMF) procedure based on iterative reciprocal and Markovian projections has recently been proposed as a powerful method for solving the Schr\u00f6dinger Bridge problem. However, it has been observed that for the practical implementation of this procedure, it is crucial to alternate between fitting a forward and backward time diffusion at each iteration. Such implementation is thought to be a practical heuristic, which is required to stabilize training and obtain good results in applications such as unpaired domain translation. In our work, we show that this heuristic closely connects with the pioneer approaches for the Schr\u00f6dinger Bridge based on the Iterative Proportional Fitting (IPF) procedure. Namely, we find that the practical implementation of IMF is, in fact, a combination of IMF and IPF procedures, and we call this combination the Iterative Proportional Markovian Fitting (IPMF) procedure. We show both theoretically and practically that this combined IPMF procedure can converge under more general settings, thus, showing that the IPMF procedure opens a door towards developing a unified framework for solving Schr\u00f6dinger Bridge problems.", "keywords": [ "Schr\u00f6dinger Bridge", "Optimal Transport", "Entropic Optimal Transport", "Unpaired Learning" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "LieTse3fQB", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Min Xu", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "Full Professor" }, { "name": "Stuart Perry", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Full Professor" }, { "name": "Zexu Huang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 45, "n_ref": 138, "n_ref_all": 149, "n_fig": 8, "n_tab": 5, "L_tab": 1001, "n_element_tab": 155, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1705, "n_element_tab_1": 185, "formula_len_all": 832, "formula_len_all_1": 833, "len_all": 139136, "len_all_1": 60972, "len_abs": 1590, "len_title": 116, "len_sents": 41535, "len_sents_1": 26482, "n_sents": 329, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1532, "title": "GaussianFocus: Constrained Attention Focus for 3D Gaussian Splatting", "abs": "Recent developments in 3D reconstruction and neural rendering have significantly propelled the capabilities of photo-realistic 3D scene rendering across various academic and industrial fields. The 3D Gaussian Splatting technique, alongside its derivatives, integrates the advantages of primitive-based and volumetric representations to deliver top-tier rendering quality and efficiency. Despite these advancements, the method tends to generate excessive redundant noisy Gaussians overfitted to every training view, which degrades the rendering quality. Additionally, while 3D Gaussian Splatting excels in small-scale and object-centric scenes, its application to larger scenes is hindered by constraints such as limited video memory, excessive optimization duration, and variable appearance across views. To address these challenges, we introduce GaussianFocus, an innovative approach that incorporates a patch attention algorithm to refine rendering quality and implements a Gaussian constraints strategy to minimize redundancy. Moreover, we propose a subdivision reconstruction strategy for large-scale scenes, dividing them into smaller mergeable blocks for individual training. Our results indicate that GaussianFocus significantly reduces unnecessary Gaussians and enhances rendering quality, surpassing existing State-of-The-Art (SoTA) methods. Furthermore, we demonstrate the capability of our approach to effectively manage and render large scenes, such as urban environments, maintaining high fidelity in the visual output.", "keywords": [ "3D Gaussian Splatting", "3D Reconstruction", "Novel View Synthesis", "Neural Rendering", "Large Scene Reconstruction" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 1, 3, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "LidZXoqZ2Q", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 11, "author_info_list": [ { "name": "Chak Tou Leong", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Hongbo Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Jun Wang", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Professor" }, { "name": "Linyi Yang", "gender": "unknown", "institution": "University College London", "country": "GB", "position": "Researcher" }, { "name": "Minjun Zhu", "gender": "Female", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Qiang Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Qingyu Yin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Wenjie Li", "gender": "Female", "institution": "The Hong Kong Polytechnic University, The Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Yue Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" }, { "name": "Yulan He", "gender": "Female", "institution": "King's College London, University of London", "country": "GB", "position": "Full Professor" }, { "name": "hanqi yan", "gender": "Female", "institution": "King's College London", "country": "GB", "position": "Postdoc" } ], "n_formula": 29, "n_formula_1": 23, "n_ref_uni": 33, "n_ref": 64, "n_ref_all": 99, "n_fig": 4, "n_tab": 6, "L_tab": 1262, "n_element_tab": 200, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1302, "n_element_tab_1": 185, "formula_len_all": 2505, "formula_len_all_1": 2127, "len_all": 139534, "len_all_1": 70303, "len_abs": 1104, "len_title": 117, "len_sents": 42393, "len_sents_1": 31105, "n_sents": 339, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1115, "title": "Direct Preference Optimization Using Sparse Feature-level Constraints", "abs": "The alignment of large language models (LLMs) with human preferences remains a key challenge. While post-training techniques like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) have achieved notable success, they often introduce computational inefficiencies and training instability. In this paper, we propose **F**eature-level constrained **P**reference **O**ptimization (FPO), a novel method designed to simplify the alignment process while ensuring stability. FPO leverages pre-trained Sparse Autoencoders (SAEs) and introduces feature-level constraints, allowing for efficient, sparsity-enforced alignment. Our approach enjoys efficiency by using sparse features activated in a well-trained sparse autoencoder and the quality of sequential KL divergence by using the feature-level offline reference. Experimental results on benchmark datasets demonstrate that FPO achieves a 5.08\\% absolute improvement in win rate with much lower computational cost compared to state-of-the-art baselines, making it a promising solution for efficient and controllable LLM alignments.", "keywords": [ "Preference Optimization", "Reinforcement Learning", "LLM", "Sparse AutoEncoder" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 1, 2 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "LiUfN9h0Lx", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Bernd Bischl", "gender": "Male", "institution": "LMU", "country": "DE", "position": "Full Professor" }, { "name": "Giuseppe Casalicchio", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "Lecturer" }, { "name": "Hubert Baniecki", "gender": "Male", "institution": "University of Warsaw", "country": "PL", "position": "PhD student" }, { "name": "Przemyslaw Biecek", "gender": "Not Specified", "institution": "Warsaw University of Technology", "country": "PL", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 15, "n_ref_uni": 44, "n_ref": 126, "n_ref_all": 160, "n_fig": 18, "n_tab": 5, "L_tab": 2883, "n_element_tab": 343, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 154, "n_element_tab_1": 14, "formula_len_all": 2131, "formula_len_all_1": 1260, "len_all": 172498, "len_all_1": 68548, "len_abs": 1113, "len_title": 123, "len_sents": 50084, "len_sents_1": 30448, "n_sents": 429, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 75, "L_abs": 1196, "title": "Efficient and Accurate Explanation Estimation with Distribution Compression", "abs": "We discover a theoretical connection between explanation estimation and distribution compression that significantly improves the approximation of feature attributions, importance, and effects. While the exact computation of various machine learning explanations requires numerous model inferences and becomes impractical, the computational cost of approximation increases with an ever-increasing size of data and model parameters. We show that the standard i.i.d. sampling used in a broad spectrum of algorithms for post-hoc explanation leads to an approximation error worthy of improvement. To this end, we introduce Compress Then Explain (CTE), a new paradigm of sample-efficient explainability. It relies on distribution compression through kernel thinning to obtain a data sample that best approximates its marginal distribution. CTE significantly improves the accuracy and stability of explanation estimation with negligible computational overhead. It often achieves an on-par explanation approximation error 2-3x faster by using fewer samples, i.e. requiring 2-3x fewer model evaluations. CTE is a simple, yet powerful, plug-in for any explanation method that now relies on i.i.d. sampling.", "keywords": [ "explainable ai", "feature attributions", "feature importance", "sampling", "kernel thinning" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 2, 5 ] }, { "paper_id": "LhuDdMEIGS", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Rithesh Kumar", "gender": "Male", "institution": "Adobe Systems", "country": "CA", "position": "Researcher" }, { "name": "Yinghao Aaron Li", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Zeyu Jin", "gender": "Male", "institution": "Adobe Systems", "country": "", "position": "Researcher" } ], "n_formula": 40, "n_formula_1": 19, "n_ref_uni": 59, "n_ref": 139, "n_ref_all": 171, "n_fig": 4, "n_tab": 9, "L_tab": 3959, "n_element_tab": 290, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1847, "n_element_tab_1": 104, "formula_len_all": 2142, "formula_len_all_1": 1056, "len_all": 218108, "len_all_1": 71933, "len_abs": 1228, "len_title": 122, "len_sents": 62069, "len_sents_1": 33900, "n_sents": 440, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 120, "L_abs": 1235, "title": "DMDSpeech: Distilled Diffusion Model Surpassing The Teacher in Zero-shot Speech Synthesis via Direct Metric Optimization", "abs": "Diffusion models have demonstrated significant potential in speech synthesis tasks, including text-to-speech (TTS) and voice cloning. However, their iterative denoising processes are inefficient and hinder the application of end-to-end optimization with perceptual metrics. In this paper, we propose a novel method of distilling TTS diffusion models with direct end-to-end evaluation metric optimization, achieving state-of-the-art performance. By incorporating Connectionist Temporal Classification (CTC) loss and Speaker Verification (SV) loss, our approach optimizes perceptual evaluation metrics, leading to notable improvements in word error rate and speaker similarity. Our experiments show that DMDSpeech consistently surpasses prior state-of-the-art models in both naturalness and speaker similarity while being significantly faster. Moreover, our synthetic speech has a higher level of voice similarity to the prompt than the ground truth in both human evaluation and objective speaker similarity metric. This work highlights the potential of direct metric optimization in speech synthesis, allowing models to better align with human auditory preferences. The audio samples are available at https://dmdspeech.github.io/demo/.", "keywords": [ "text-to-speech", "zero-shot speech synthesis", "diffusion model", "diffusion distillation", "metric optimization" ], "rating_list": [ 5, 3, 3, 8 ], "soundness_list": [ 3, 2, 4, 4 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Lh30EOD4CT", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Edith C. H. Ngai", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Yun-Hin Chan", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 33, "n_ref": 62, "n_ref_all": 109, "n_fig": 18, "n_tab": 5, "L_tab": 3113, "n_element_tab": 432, "n_fig_1": 17, "n_tab_1": 3, "L_tab_1": 2320, "n_element_tab_1": 295, "formula_len_all": 847, "formula_len_all_1": 721, "len_all": 130013, "len_all_1": 64116, "len_abs": 1756, "len_title": 118, "len_sents": 30898, "len_sents_1": 26092, "n_sents": 235, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1534, "title": "PromptSFL: Improving Visual Prompt Tuning For Split Federated Learning", "abs": "Conflict arises due to the disparity between the substantial resource demands of pre-trained models and the limited available resources of federated learning (FL) participants. Split learning presents a viable approach for adapting pre-trained models to FL, involving the allocation of a small portion of the pre-trained model to clients while deploying the remaining part on a server. Moreover, the application of Visual Prompt Tuning (VPT) to pre-trained models has shown state-of-the-art performances in parameter-efficient fine-tuning methods. However, VPT exhibits unsatisfactory performance in split federated learning (SFL) compared to its performance in centralized learning. In this paper, we first identify that VPT falls short of expectations in SFL due to the insufficient generalization capability of clients. To address this issue, we propose PromptSFL, which aligns the feature spaces of prompts between clients and the server to adapt VPT for SFL. PromptSFL transmits the final prompts in clients, termed skip prompts, to the first prompts in the server, enabling clients to extract more common features from the server. Additionally, we introduce a linear layer to map the prompts from clients to the feature space in the server during this skipping process, preventing the prompts of clients from overfitting to local datasets. Moreover, to enhance the convergence speed of SFL, PromptSFL employs an adaptive learning rate for clients. Extensive experiments demonstrate the effectiveness and efficiency of PromptSFL.", "keywords": [ "Federated Learning", "Fine Tuning", "Prompts" ], "rating_list": [ 1, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "LgzRo1RpLS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Massoud Pedram", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Professor" }, { "name": "Mohammad Erfan Sadeghi", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Seyedarmin Azizi", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Souvik Kundu", "gender": "Male", "institution": "Intel", "country": "US", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 15, "n_ref": 49, "n_ref_all": 76, "n_fig": 8, "n_tab": 6, "L_tab": 1625, "n_element_tab": 156, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 529, "n_element_tab_1": 59, "formula_len_all": 838, "formula_len_all_1": 731, "len_all": 119797, "len_all_1": 66938, "len_abs": 1789, "len_title": 128, "len_sents": 41304, "len_sents_1": 30213, "n_sents": 303, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1879, "title": "MambaExtend: A Training-Free Approach to Improve Long Context Extension of Mamba", "abs": "The inherent quadratic complexity of the attention mechanism in transformer models has driven the research community to explore alternative architectures with sub-quadratic complexity, such as state-space models. Within this emerging paradigm, Mamba has established itself as a leading model, achieving state-of-the-art results in various language modeling benchmarks. However, despite its impressive performance, Mamba's effectiveness is significantly limited by its pre-training context length, resulting in a pronounced degradation when the model is tasked with handling longer contexts. Our investigation reveals that Mamba's inability to generalize effectively to long contexts is primarily due to the out-of-distribution (OOD) discretization steps. To address this critical limitation, we introduce _**MambaExtend**_, a novel framework designed to enhance context extension capabilities of Mamba. Specifically, MambaExtend leverages a _**training-free**_ approach to calibrate _only_ the scaling factors of discretization modules for different layers. We demonstrate both gradient-based and gradient-free zeroth-order optimization to learn the optimal scaling factors for each Mamba layer, requiring orders of magnitude fewer updates as opposed to the parameter fine-tuning based alternatives. With this, for the first time, we can enable a training-free context extension of up to $\\mathbf{32}\\times$ from $2$k to $64$k, that too without any significant increase in perplexity. Compared to the existing alternative approach of fine-tuning, due to only selective calibration of the scaling factors, MambaExtend requires up to $\\mathord{\\sim}$$\\mathbf{{5.42*10^{6}}}\\times$ fewer parameter update costing up to $\\mathbf{3.87}\\times$ lower peak-memory while maintaining similar or better long-context performance evaluated across multiple tasks. Code will be released soon.", "keywords": [ "Mamba", "Long Context Generalization", "Discretization Step", "SSM" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "LglOy15bqe", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Jiaxiang Li", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "Postdoc" }, { "name": "Kaixiang Lin", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Mingyi Hong", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Siliang Zeng", "gender": "Male", "institution": "University of Minnesota, Twin Cities", "country": "US", "position": "PhD student" }, { "name": "Xinnan Zhang", "gender": "Male", "institution": "University of Minnesota - Twin Cities", "country": "US", "position": "PhD student" } ], "n_formula": 49, "n_formula_1": 45, "n_ref_uni": 60, "n_ref": 93, "n_ref_all": 135, "n_fig": 4, "n_tab": 1, "L_tab": 493, "n_element_tab": 85, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3454, "formula_len_all_1": 2602, "len_all": 169827, "len_all_1": 68360, "len_abs": 1761, "len_title": 151, "len_sents": 40222, "len_sents_1": 30028, "n_sents": 298, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1768, "title": "Policy optimization can be memory-efficient: LLM Alignment Through Successive Policy Re-weighting (SPR)", "abs": "Reinforcement learning (RL) is serving as the cornerstone of aligning large language models (LLMs) to human behavior, by providing an appealing formulation and a suite of effective algorithms for learning behavior strategies through interacting with the underlying environment. Current paradigm of RL-based methods for LLM alignment, such as reinforcement learning with human feedback (RLHF) involves utilizing a reward function learned from extensive offline datasets to expediate the online training of reinforcement learning. The reward function learned is then used for policy optimization to obtain an improved policy (i.e. the LLM). Despite the success of RL approaches in aligning LLM with offline datasets, there are significant computational/limit of resources concern on applying RL-based methods for LLMs. For example, standard RLHF requires simultaneous loading of four models to the computing unit. In this paper, we develop a novel policy optimization algorithm named Successive Policy Re-weighting (SPR), matching the peak memory consumption of standard supervised fine-tune (SFT). Further, SPR can leverage both offline and online datasets to expediate online training and improve the sample efficiency. Specifically, SPR leverages a supervised learning subroutine to achieve policy improvement through re-weighting the policy according to the importance/performance of executed actions. Such simple and effective method is computationally inexpensive, requiring loading only one model at each update step, matching the computational cost of standard supervised fine-tuning procedure. Experimental results show that the proposed method can significantly outperform benchmark algorithms and accelerate the online training with available offline dataset.", "keywords": [ "Alignment", "Large Language Models", "Reinforcement Learning with Human Feedback", "Policy Optimization", "Reweighting" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "LgfaMR6Sst", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Hyunsu Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Juho Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Yegon Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 21, "n_formula_1": 10, "n_ref_uni": 51, "n_ref": 99, "n_ref_all": 122, "n_fig": 7, "n_tab": 16, "L_tab": 5709, "n_element_tab": 320, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 302, "n_element_tab_1": 28, "formula_len_all": 972, "formula_len_all_1": 547, "len_all": 175228, "len_all_1": 66792, "len_abs": 1435, "len_title": 92, "len_sents": 51445, "len_sents_1": 29766, "n_sents": 415, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1440, "title": "Flexible Active Learning of PDE Trajectories", "abs": "Accurately solving partial differential equations (PDEs) is critical for understanding complex scientific and engineering phenomena, yet traditional numerical solvers are computationally expensive. Surrogate models offer a more efficient alternative, but their development is hindered by the cost of generating sufficient training data from numerical solvers. In this paper, we present a novel framework for active learning (AL) in PDE surrogate modeling that reduces this cost. Unlike the existing AL methods for PDEs that always acquire entire PDE trajectories, our approach strategically generates only the most important time steps with the numerical solver, while employing the surrogate model to approximate the remaining steps. This dramatically reduces the cost incurred by each trajectory and thus allows the active learning algorithm to try out a more diverse set of trajectories given the same budget. To accommodate this novel framework, we develop an acquisition function that estimates the utility of a set of time steps by approximating its resulting variance reduction. We demonstrate the effectiveness of our method on several benchmark PDEs, including the Heat equation, Korteweg\u2013De Vries equation, Kuramoto\u2013Sivashinsky equation, and the incompressible Navier-Stokes equation. Extensive experiments validate that our approach outperforms existing methods, offering a cost-efficient solution to surrogate modeling for PDEs.", "keywords": [ "Active learning", "Partial Differential Equation (PDE)" ], "rating_list": [ 5, 8, 8, 5, 8 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "Lfy9q7Icp9", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Borja Balle", "gender": "unknown", "institution": "DeepMind", "country": "", "position": "Research scientist" }, { "name": "Meisam Razaviyayn", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Associate Professor" }, { "name": "Mingyi Hong", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Vahab Mirrokni", "gender": "Male", "institution": "Google Research", "country": "", "position": "Distinguished Scientist" }, { "name": "Xinwei Zhang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Postdoc" }, { "name": "Zhiqi Bu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 83, "n_formula_1": 28, "n_ref_uni": 59, "n_ref": 130, "n_ref_all": 164, "n_fig": 7, "n_tab": 7, "L_tab": 1181, "n_element_tab": 173, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 820, "n_element_tab_1": 23, "formula_len_all": 12392, "formula_len_all_1": 3126, "len_all": 200158, "len_all_1": 66873, "len_abs": 1808, "len_title": 88, "len_sents": 51834, "len_sents_1": 28569, "n_sents": 407, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1820, "title": "DiSK: Differentially Private Optimizer with Simplified Kalman Filter for Noise Reduction", "abs": "Differential privacy (DP) offers a robust framework for safeguarding individual data privacy. To utilize DP in training modern machine learning models, differentially private optimizers have been widely used in recent years. A popular approach to privatize an optimizer is to clip the individual gradients and add sufficiently large noise to the clipped gradient. This approach led to the development of DP optimizers that have comparable performance with their non-private counterparts in fine-tuning tasks or in tasks with a small number of training parameters. However, a significant performance drop is observed when these optimizers are applied to large-scale training. This degradation stems from the substantial noise injection required to maintain DP, which disrupts the optimizer's dynamics.\nThis paper introduces DiSK, a novel framework designed to significantly enhance the performance of DP optimizers. DiSK employs Kalman filtering, a technique drawn from control and signal processing, to effectively denoise privatized gradients and generate progressively refined gradient estimations. To ensure practicality for large-scale training, we simplify the Kalman filtering process, minimizing its memory and computational demands.\nWe establish theoretical privacy-utility trade-off guarantees for DiSK, and demonstrate provable improvements over standard DP optimizers like DPSGD in terms of iteration complexity upper-bound.\nExtensive experiments across diverse tasks, including vision tasks such as CIFAR-100 and ImageNet-1k and language fine-tuning tasks such as GLUE, E2E, and DART, validate the effectiveness of DiSK. The results showcase its ability to significantly improve the performance of DP optimizers, surpassing state-of-the-art results under the same privacy constraints on several benchmarks.", "keywords": [ "Differential privacy", "Kalman filter", "noise reduction", "nonconvex optimization" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "LfghnrSJNg", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Tao Jin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Weicai Yan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Xiaoda Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Ye Wang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Zehan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zirun Guo", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "fangming feng", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Lin Wang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 42, "n_ref": 88, "n_ref_all": 101, "n_fig": 9, "n_tab": 12, "L_tab": 3796, "n_element_tab": 328, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1536, "n_element_tab_1": 178, "formula_len_all": 580, "formula_len_all_1": 323, "len_all": 183046, "len_all_1": 65922, "len_abs": 1656, "len_title": 116, "len_sents": 43950, "len_sents_1": 28712, "n_sents": 347, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1662, "title": "Diff-Prompt: Diffusion-driven Prompt Generator with Mask Supervision", "abs": "Prompt learning has demonstrated promising results in fine-tuning pre-trained multimodal models. However, the performance improvement is limited when applied to more complex and fine-grained tasks. The reason is that most existing methods directly optimize the parameters involved in the prompt generation process through loss backpropagation, which constrains the richness and specificity of the prompt representations. In this paper, we propose Diff-Prompt (Diffusion-driven Prompt Generator), aiming to use the diffusion model to generate rich, fine-grained prompt information for complex downstream tasks. Specifically, our approach consists of three stages. In the first stage, we train a Mask-VAE to compress the masks into latent space. In the second stage, we leverage an improved DiT (Diffusion Transformer) to train a prompt generator in the latent space, using the masks for supervision. In the third stage, we align the denoising process of the prompt generator with the pre-trained model in the semantic space, and use the generated prompts to fine-tune the model. We conduct experiments on a complex pixel-level downstream task, referring expression comprehension, and compare our method with various parameter-efficient fine-tuning approaches. Diff-Prompt achieves a maximum improvement of 8.87 in R@1 and 14.05 in R@5 compared to the foundation model and also outperforms other state-of-the-art methods across multiple metrics. The experimental results validate the effectiveness of our approach and highlight the potential of using generative models for prompt generation. Code is available at https://anonymous.4open.science/r/Diff-Prompt-FF2D.", "keywords": [ "prompt learning", "diffusion model", "multimodal learning" ], "rating_list": [ 6, 8, 6, 6, 3 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2, 1 ], "contribution_list": [ 2, 2, 3, 2, 1 ], "confidence_list": [ 4, 4, 3, 5, 5 ] }, { "paper_id": "LfekK1E0QE", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Zhenghai Xue", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 13, "n_ref_uni": 35, "n_ref": 97, "n_ref_all": 128, "n_fig": 7, "n_tab": 5, "L_tab": 2811, "n_element_tab": 262, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2428, "n_element_tab_1": 223, "formula_len_all": 4058, "formula_len_all_1": 1016, "len_all": 169213, "len_all_1": 78258, "len_abs": 1540, "len_title": 136, "len_sents": 52824, "len_sents_1": 32616, "n_sents": 428, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1498, "title": "Policy Optimization under Imperfect Human Interactions with Agent-Gated Shared Autonomy", "abs": "We introduce AGSA, an Agent-Gated Shared Autonomy framework that learns from high-level human feedback to tackle the challenges of reward-free training, safe exploration, and imperfect low-level human control. Recent human-in-the loop learning methods enable human participants to intervene a learning agent\u2019s control and provide online demonstrations. Nonetheless, these methods rely heavily on perfect human interactions, including accurate human-monitored intervention decisions and near-optimal human demonstrations. AGSA employs a dedicated gating agent to determine when to switch control, thereby reducing the need of constant human monitoring. To obtain a precise and foreseeable gating agent, AGSA trains a long-term gating value function from human evaluative feedback on the gating agent\u2019s intervention requests and preference feedback on pairs of human intervention trajectories. Instead of relying on potentially suboptimal human demonstrations, the learning agent is trained using control-switching signals from the gating agent. We provide theoretical insights on performance bounds that respectively describe the ability of the two agents. Experiments are conducted with both simulated and real human participants at different skill levels in challenging continuous control environments. Comparative results highlight that AGSA achieves significant improvements over previous human-in-the-loop learning methods in terms of training safety, policy performance, and user-friendliness.", "keywords": [ "Reinforcement Learning", "Human-in-the-loop Learning", "Imperfect Human Interaction", "Human Feedback" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "Lf5znhZmFu", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "En-jui kuo", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Assistant Professor" }, { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" }, { "name": "Jerry Yao-Chieh Hu", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Maojiang Su", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" } ], "n_formula": 98, "n_formula_1": 46, "n_ref_uni": 57, "n_ref": 114, "n_ref_all": 120, "n_fig": 1, "n_tab": 1, "L_tab": 138, "n_element_tab": 11, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 17307, "formula_len_all_1": 6656, "len_all": 230858, "len_all_1": 66280, "len_abs": 2135, "len_title": 125, "len_sents": 50502, "len_sents_1": 24609, "n_sents": 633, "n_sents_1": 280, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1445, "title": "Computational Limits of Low-Rank Adaptation (LoRA) Fine-Tuning for Transformer Models", "abs": "We study the computational limits of Low-Rank Adaptation (LoRA) for finetuning transformer-based models using fine-grained complexity theory.\nOur key observation is that the existence of low-rank decompositions within the gradient computation of LoRA adaptation leads to possible algorithmic speedup.\nThis allows us to (i) identify a phase transition behavior of efficiency \\blue{assuming the Strong Exponential Time Hypothesis (SETH)}, and (ii) prove the existence of almost linear algorithms by controlling the LoRA update computation term by term.\nFor the former, we identify a sharp transition in the efficiency of all possible rank-$r$ LoRA update algorithms for transformers, based on specific norms resulting from the multiplications of the input sequence $X$, pretrained weights ${W^\\star}$, and adapter matrices $\\alpha B A/r$.\nSpecifically, we derive a shared upper bound threshold for such norms and show that efficient (sub-quadratic) approximation algorithms of LoRA exist only below this threshold.\nFor the latter, we prove the existence of almost linear approximation algorithms for LoRA adaptation by utilizing the hierarchical low-rank structures of LoRA gradients and approximating the gradients with a series of chained low-rank approximations.\nTo showcase our theory, we consider two practical scenarios: partial (e.g., only $W_V$ and $W_Q$) and full adaptations (e.g., $W_Q$, $W_V$, and $W_K$) of weights in attention heads.", "keywords": [ "Parameter Efficient Finetuning", "Low-Rank Adaptation", "LoRA", "Transformer", "Foundation Models", "Large Language Models", "Fine-Grained Complexity", "Strong Exponential Time Hypothesis" ], "rating_list": [ 6, 3, 6, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "Le823SjZEc", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Yongbing Zhang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zhu Zirui", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 22, "n_ref": 26, "n_ref_all": 35, "n_fig": 4, "n_tab": 4, "L_tab": 747, "n_element_tab": 46, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 747, "n_element_tab_1": 46, "formula_len_all": 508, "formula_len_all_1": 508, "len_all": 84735, "len_all_1": 51559, "len_abs": 1872, "len_title": 75, "len_sents": 25911, "len_sents_1": 25111, "n_sents": 164, "n_sents_1": 158, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1883, "title": "Quantization Enhanced Cross-modal Alignment for Gene Expression Prediction", "abs": "In modern healthcare, whole-slide histological images (WSIs) provide information on tissue structure and composition at the microscopic level. Integrating WSIs and gene expression profiles enhances cancer diagnosis and treatment planning, advancing clinical care and research. However, spatial transcriptomics is costly and requires a long sampling time. The intrinsic correlation between histological images and gene expressions offers the potential for predicting spatial transcriptomics using Hematoxylin-Eosin (H\\&E) stained WSIs to reduce time and resource costs. Although existing methods have achieved impressive results, they ignore the heterogeneity between modalities of image and gene expression. In this paper, we propose a Quantized Cross-modal Alignment (QCA) that exploits cross-modal interactions to address the issue of modal heterogeneity. Considering the interference of gene-unrelated image features, we develop a Gene-related Image Feature Quantizer (GIFQ) to capture the gene-related image features. Meanwhile, we develop an Asymmetric Cross-modal Alignment (ACA) approach, which facilitates the model to generate discriminative predictions from similar visual presentations. In addition, to fix the discriminability reduction, a Discriminability-Enhancing Regularization (DER) is further devised to regularize both the virtual and real gene features. Experimental results on a breast cancer dataset sampled by solid-phase transcriptome capture elucidate that our QCA model achieves state-of-the-art results for accurate prognostication of gene expression profiles, increasing the performance by 13\\% at least. Our method utilizes deep learning technology to delineate the correlation between morphological features and gene expression, furnishing new perspectives and instruments for disclosing biomarkers in histological conditions. The code will be released.", "keywords": [ "Gene Expression Prediction", "Cross-modal Alignment" ], "rating_list": [ 5, 3, 1, 3 ], "soundness_list": [ 2, 1, 1, 2 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "LdIlnsePNt", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Baihe Huang", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Banghua Zhu", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "Hanlin Zhu", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department, University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Jason D. Lee", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Jiantao Jiao", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" }, { "name": "Julien Piet", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department, University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Kannan Ramchandran", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Michael Jordan", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" } ], "n_formula": 103, "n_formula_1": 11, "n_ref_uni": 36, "n_ref": 117, "n_ref_all": 160, "n_fig": 3, "n_tab": 5, "L_tab": 1786, "n_element_tab": 182, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 743, "n_element_tab_1": 76, "formula_len_all": 10765, "formula_len_all_1": 620, "len_all": 248989, "len_all_1": 66027, "len_abs": 1609, "len_title": 128, "len_sents": 78511, "len_sents_1": 31271, "n_sents": 714, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1345, "title": "Watermarking using Semantic-aware Speculative Sampling: from Theory to Practice", "abs": "Statistical watermarking offers a theoretically-sound method for distinguishing machine-generated texts. In this work, we first present a systematic theoretical analysis of the statistical limits of watermarking, by framing it as a hypothesis testing problem. We derive nearly matching upper and lower bounds for (i) the optimal Type II error under a fixed Type I error, and (ii) the minimum number of tokens required to watermark the output. Our rate of $\\Theta(h^{-1} \\log (1/h))$ for the minimum number of required tokens, where $h$ is the average entropy per token, reveals a significant gap between the statistical limit and the $O(h^{-2})$ rate achieved in prior works. To our knowledge, this is the first comprehensive statistical analysis of the watermarking problem. Building on our theory, we develop **SEAL** (**S**emantic-awar**E** specul**A**tive samp**L**ing), a novel watermarking algorithm for practical applications. SEAL introduces two key techniques: (i) designing semantic-aware random seeds by leveraging a proposal language model, and (ii) constructing a maximal coupling between the random seed and the next token through speculative sampling. Experiments on open-source benchmarks demonstrate that our watermarking scheme delivers superior efficiency and tamper resistance, particularly in the face of paraphrase attacks.", "keywords": [ "Watermark", "Large Language Model", "Hypothesis testing" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "LcpdPCkZwI", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Guodong Long", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Jing Jiang", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Qinghua Lu", "gender": "Female", "institution": "CSIRO", "country": "AU", "position": "Principal Researcher" }, { "name": "Shanshan Ye", "gender": "Female", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Tianyi Zhou", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "yiyuan yang", "gender": "Female", "institution": "University of Technology Sydney", "country": "", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 14, "n_ref_uni": 51, "n_ref": 117, "n_ref_all": 135, "n_fig": 3, "n_tab": 9, "L_tab": 1807, "n_element_tab": 184, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 733, "n_element_tab_1": 78, "formula_len_all": 4332, "formula_len_all_1": 984, "len_all": 192058, "len_all_1": 71332, "len_abs": 1842, "len_title": 119, "len_sents": 56136, "len_sents_1": 34259, "n_sents": 368, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1848, "title": "Federated Adapter on Foundation Models: An Out-Of-Distribution Approach", "abs": "As foundation models gain increasing attention from both academic and industrial communities, Federated Foundation Models (FedFM) have emerged as a privacy-preserving approach for collaboratively fine-tuning models in federated learning (FL) frameworks using distributed datasets across multiple clients. A key challenge for FedFM, given the versatile nature of foundation models, is addressing out-of-distribution (OOD) generalization, where unseen tasks or clients may exhibit distribution shifts leading to suboptimal performance. \nAlthough numerous studies have explored OOD generalization in conventional FL, these methods are inadequate for FedFM due to the challenges posed by large parameter scales and increased data heterogeneity, where large parameter scales would result in high computational and communication costs while increased data heterogeneity like cross-domain would lead to suboptimal performance of the aggregated global model on individual client distributions. To bridge this gap, we propose a new method, called FedOA, to enhance the OOD generalization of FedFM under these conditions.\nSpecifically, our method employs adapter-based parameter-efficient fine-tuning methods for efficient learning, and introduces an additional personalized model with a feature distance-based regularization to ensure distribution alignment and provide OOD generalization guarantees for each client. Theoretically, we demonstrate that the conventional aggregated global model in FedFM inherently retains OOD generalization capabilities, and our proposed method enhances the personalized model's OOD generalization through regularization informed by the global model, with proven convergence under general non-convex settings.\nEmpirically, the effectiveness of the proposed method is validated on benchmark datasets across various NLP tasks.", "keywords": [ "Federated Learning", "Foundation Models" ], "rating_list": [ 3, 6, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 1, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "LcoOwM5y7r", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Jianzong Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "Undergrad student" }, { "name": "Jinbin Bai", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Jingbo Wang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Lu Qi", "gender": "Male", "institution": "University of California, Merced", "country": "US", "position": "Postdoc" }, { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Qingyu Shi", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xiangtai Li", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Yunhai Tong", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 80, "n_ref_all": 97, "n_fig": 12, "n_tab": 9, "L_tab": 2737, "n_element_tab": 288, "n_fig_1": 10, "n_tab_1": 5, "L_tab_1": 873, "n_element_tab_1": 151, "formula_len_all": 545, "formula_len_all_1": 546, "len_all": 127187, "len_all_1": 58026, "len_abs": 1582, "len_title": 114, "len_sents": 34617, "len_sents_1": 25249, "n_sents": 273, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1588, "title": "RelationBooth: Towards Relation-Aware Customized Object Generation", "abs": "Customized image generation is crucial for delivering personalized content based on user-provided image prompts, aligning large-scale text-to-image diffusion models with individual needs. However, existing models often overlook the relationships between customized objects in generated images. Instead, this work addresses that gap by focusing on relation-aware customized image generation, which aims to preserve the identities from image prompts while maintaining the predicate relations described in text prompts. Specifically, we introduce RelationBooth, a framework that disentangles identity and relation learning through a well-curated dataset. Our training data consists of relation-specific images, independent object images containing identity information, and text prompts to guide relation generation. Then, we propose two key modules to tackle the two main challenges\u2014generating accurate and natural relations, especially when significant pose adjustments are required, and avoiding object confusion in cases of overlap. First, we introduce a keypoint matching loss that effectively guides the model in adjusting object poses closely tied to their relationships. Second, we incorporate local features from the image prompts to better distinguish between objects, preventing confusion in overlapping cases. Extensive results on three benchmarks demonstrate the superiority of RelationBooth in generating precise relations while preserving object identities across a diverse set of objects and relations. The source code and trained models will be made available to the public.", "keywords": [ "relation generation; customize image generation;" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "Lbx9zdURxe", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Qitian Wu", "gender": "Male", "institution": "Broad Institute of MIT and Harvard", "country": "US", "position": "Researcher" }, { "name": "Yiting Chen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 20, "n_ref_uni": 34, "n_ref": 76, "n_ref_all": 91, "n_fig": 3, "n_tab": 7, "L_tab": 5296, "n_element_tab": 241, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1517, "n_element_tab_1": 164, "formula_len_all": 1346, "formula_len_all_1": 1048, "len_all": 159940, "len_all_1": 56972, "len_abs": 1407, "len_title": 129, "len_sents": 35003, "len_sents_1": 25104, "n_sents": 262, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1420, "title": "Regularizing Energy among Training Samples for Out-of-Distribution Generalization", "abs": "The energy-based model provides a unified framework for various learning models where an energy value is assigned to each configuration of random variables based on probability. Recently, different methods have been proposed to derive an energy value out of the logits of a classifier for out-of-distribution (OOD) detection or OOD generalization. However, these methods mainly focus on the energy difference between in-distribution and OOD data samples, neglecting the energy difference among in-distribution data samples. In this paper, we show that the energy among in-distribution data also requires attention. We propose to investigate the energy difference between in-distribution data samples. Both empirically and theoretically, we show that previous methods for subpopulation shift (\\emph{e.g.}, long-tail classification) such as data re-weighting and margin control apply implicit energy regularization and we provide a unified framework from the energy perspective. With the influence function, we further extend the energy regularization framework to OOD generalization scenarios where the distribution shift is more implicit compared to the long-tail recognition scenario. We conduct experiments on long-tail datasets, subpopulation shift benchmarks, and OOD generalization benchmarks to show the effectiveness of the proposed energy regularization method. The source code will be made publically available.", "keywords": [ "Energy based model; OOD generalization; Long-tail Recognition" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "LbgIZpSUCe", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Cristina Savin", "gender": "Female", "institution": "New York University", "country": "US", "position": "Associate Professor" }, { "name": "Matthew Dowling", "gender": "Male", "institution": "State University of New York, Stony Brook", "country": "", "position": "PhD student" } ], "n_formula": 48, "n_formula_1": 22, "n_ref_uni": 57, "n_ref": 89, "n_ref_all": 126, "n_fig": 24, "n_tab": 2, "L_tab": 1425, "n_element_tab": 67, "n_fig_1": 22, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4690, "formula_len_all_1": 2115, "len_all": 159230, "len_all_1": 70615, "len_abs": 1234, "len_title": 141, "len_sents": 48671, "len_sents_1": 35892, "n_sents": 274, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 93, "L_abs": 1244, "title": "Nonlinear multiregion neural dynamics with parametric impulse response communication channels", "abs": "Cognition arises from the coordinated interaction of brain regions with distinct computational roles. Despite improvements in our ability to extract the dynamics underlying circuit computation from population activity recorded in individual areas, understanding how multiple areas jointly support distributed computation remains a challenge. As part of this effort, we propose a multi-region neural dynamics model composed of two building blocks: _i)_ within-region (potentially driven) nonlinear dynamics and _ii)_ communication channels between regions, parameterized through their impulse response. Together, these choices make it possible to learn nonlinear neural population dynamics and understand the flow of information between regions by drawing from the rich literature of linear systems theory. We develop a state noise inversion free variational filtering and learning algorithm for our model and show, through neuroscientifically inspired numerical experiments, how the proposed model can reveal interpretable characterizations of the local computations within and the flow of information between neural populations. We further validate the efficacy of our approach using simultaneous population recordings from areas V1 and V2.", "keywords": [ "neural dynamics", "multiregion", "variational inference" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 4, 4 ], "presentation_list": [ 2, 4, 3 ], "contribution_list": [ 3, 4, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "LbceJJc9h2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Ernest K. Ryu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "Jaewoong Cho", "gender": "unknown", "institution": "KRAFTON", "country": "", "position": "Researcher" }, { "name": "Jaeyeon Kim", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Jason D. Lee", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Jongho Park", "gender": "Male", "institution": "Krafton Inc.", "country": "KR", "position": "Researcher" }, { "name": "Joo Young Choi", "gender": "Male", "institution": "KRAFTON", "country": "KR", "position": "Researcher" }, { "name": "Sehyun Kwon", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 17, "n_ref_uni": 75, "n_ref": 124, "n_ref_all": 150, "n_fig": 11, "n_tab": 6, "L_tab": 1556, "n_element_tab": 121, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2624, "formula_len_all_1": 1464, "len_all": 176658, "len_all_1": 66398, "len_abs": 1135, "len_title": 87, "len_sents": 38420, "len_sents_1": 27694, "n_sents": 406, "n_sents_1": 288, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 972, "title": "Task Diversity Shortens the ICL Plateau", "abs": "In-context learning (ICL) describes a language model's ability to generate outputs based on a set of input demonstrations and a subsequent query. To understand this remarkable capability, researchers have studied simplified, stylized models. These studies have consistently observed long loss plateaus, during which models exhibit minimal improvement, followed by a sudden, rapid surge of learning. In this work, we reveal that training on multiple diverse ICL tasks simultaneously shortens the loss plateaus, making each task easier to learn. This finding is surprising as it contradicts the natural intuition that the combined complexity of multiple ICL tasks would lengthen the learning process, not shorten it. Our result suggests that the recent success in large-scale training of language models may be attributed not only to the richness of the data at scale but also to the easier optimization (training) induced by the diversity of natural language training data.", "keywords": [ "In-context learning", "Transformers", "Training dynamics", "Task diversity", "Multi-task" ], "rating_list": [ 6, 3, 6, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 2, 3, 4, 4 ], "contribution_list": [ 3, 1, 2, 1 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "LbEWwJOufy", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Haiyang Liu", "gender": "Male", "institution": "the university of tokyo", "country": "JP", "position": "PhD student" }, { "name": "Qiaoge Li", "gender": "unknown", "institution": "University of Tsukuba", "country": "", "position": "PhD student" }, { "name": "Shigeru Kuriyama", "gender": "Male", "institution": "CyberAgent", "country": "", "position": "Researcher" }, { "name": "Takafumi Taketomi", "gender": "Male", "institution": "CyberAgent", "country": "JP", "position": "Researcher" }, { "name": "Tomoya Akiyama", "gender": "Male", "institution": "CyberAgent, Inc.", "country": "JP", "position": "Researcher" }, { "name": "Xingchao Yang", "gender": "Male", "institution": "CyberAgent", "country": "JP", "position": "Researcher" }, { "name": "Yuantian Huang", "gender": "Male", "institution": "CyberAgent, Inc.", "country": "JP", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 109, "n_ref_all": 128, "n_fig": 7, "n_tab": 4, "L_tab": 1133, "n_element_tab": 122, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 598, "n_element_tab_1": 80, "formula_len_all": 438, "formula_len_all_1": 0, "len_all": 153755, "len_all_1": 65046, "len_abs": 425, "len_title": 161, "len_sents": 40958, "len_sents_1": 31365, "n_sents": 319, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (conditional oral)", "L_title": 111, "L_abs": 1412, "title": "TANGO: Co-Speech Gesture Video Reenactment with Hierarchical Audio Motion Embedding and Diffusion Interpolation", "abs": "We present TANGO, a framework for generating co-speech body-gesture videos. Given a few-minute, single-speaker reference video and target speech audio, TANGO produces high-fidelity videos with synchronized body gestures. TANGO builds on Gesture Video Reenactment (GVR), which splits and retrieves video clips using a directed graph structure - representing video frames as nodes and valid transitions as edges. We address two key limitations of GVR: audio-motion misalignment and visual artifacts in GAN-generated transition frames. In particular, i) we propose retrieving gestures using latent feature distance to improve cross-modal alignment. To ensure the latent features could effectively model the relationship between speech audio and gesture motion, we implement a hierarchical joint embedding space (AuMoClip); ii) we introduce the diffusion-based model to generate high-quality transition frames. Our diffusion model, Appearance Consistent Interpolation (ACInterp), is built upon AnimateAnyone and includes a reference motion module and homography background flow to preserve appearance consistency between generated and reference videos. By integrating these components into the graph-based retrieval framework, TANGO reliably produces realistic, audio-synchronized videos and outperforms all existing generative and retrieval methods. Our code, pretrained models, and datasets are publicly available.", "keywords": [ "co-speech video generation", "cross-modal retrieval", "audio repsentation learning", "motion repsentation learning", "video frame interpolation" ], "rating_list": [ 8, 8, 8, 10 ], "soundness_list": [ 4, 4, 3, 4 ], "presentation_list": [ 4, 4, 2, 4 ], "contribution_list": [ 4, 4, 3, 4 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "Lb91pXwZMR", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Lu yan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Shikun Feng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Wei-Ying Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yanyan Lan", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Yuyan Ni", "gender": "unknown", "institution": "Institute for AI Industry Research, Tsinghua University", "country": "", "position": "Intern" }, { "name": "Zhi-Ming Ma", "gender": "unknown", "institution": "Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 39, "n_formula_1": 15, "n_ref_uni": 46, "n_ref": 113, "n_ref_all": 132, "n_fig": 5, "n_tab": 13, "L_tab": 1427, "n_element_tab": 191, "n_fig_1": 3, "n_tab_1": 9, "L_tab_1": 876, "n_element_tab_1": 99, "formula_len_all": 4317, "formula_len_all_1": 1190, "len_all": 187384, "len_all_1": 66099, "len_abs": 1271, "len_title": 126, "len_sents": 58297, "len_sents_1": 31011, "n_sents": 430, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1276, "title": "UniGEM: A Unified Approach to Generation and Property Prediction for Molecules", "abs": "Molecular generation and molecular property prediction are both crucial for drug discovery, but they are often developed independently. Inspired by recent studies, which demonstrate that diffusion model, a prominent generative approach, can learn meaningful data representations that enhance predictive tasks, we explore the potential for developing a unified generative model in the molecular domain that effectively addresses both molecular generation and property prediction tasks. However, the integration of these tasks is challenging due to inherent inconsistencies, making simple multi-task learning ineffective. To address this, we propose UniGEM, the first unified model to successfully integrate molecular generation and property prediction, delivering superior performance in both tasks. Our key innovation lies in a novel two-phase generative process, where predictive tasks are activated in the later stages, after the molecular scaffold is formed. We further enhance task balance through innovative training strategies. Rigorous theoretical analysis and comprehensive experiments demonstrate our significant improvements in both tasks. The principles behind UniGEM hold promise for broader applications, including natural language processing and computer vision.", "keywords": [ "Molecular generation", "representation learning" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "Lb414Rdzs8", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Ke Xu", "gender": "Male", "institution": "University of Illinois at Chicago", "country": "US", "position": "PhD student" }, { "name": "Philip S. Yu", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "US", "position": "Full Professor" }, { "name": "Weizhi Zhang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Intern" }, { "name": "Yuanjie Zhu", "gender": "unknown", "institution": "University of Illinois at Chicago", "country": "US", "position": "PhD student" }, { "name": "Zihe Song", "gender": "unknown", "institution": "University of Illinois at Chicago", "country": "US", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 24, "n_ref": 34, "n_ref_all": 43, "n_fig": 1, "n_tab": 12, "L_tab": 8809, "n_element_tab": 397, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 4115, "n_element_tab_1": 133, "formula_len_all": 761, "formula_len_all_1": 726, "len_all": 118499, "len_all_1": 55451, "len_abs": 975, "len_title": 119, "len_sents": 35165, "len_sents_1": 23833, "n_sents": 218, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 978, "title": "NODE-SAT: Temporal Graph Learning with Neural ODE-Guided Self-Attention", "abs": "We propose NODE-SAT, a novel temporal graph learning model that integrates Neural Ordinary Differential Equations (NODEs) with self-attention mechanisms. NODE-SAT's design requires only historical 1-hop neighbors as input and comprises three key components: a temporal link processing module utilizing NODE-guided self-attention layers to capture temporal link information, a node representation module summarizing neighbor information, and a prediction layer. Extensive experiments across thirteen temporal link prediction datasets demonstrate that NODE-SAT achieves state-of-the-art performance on most datasets with significantly faster convergence. The model demonstrates high accuracy, rapid convergence, robustness across varying dataset complexities, and strong generalization capabilities in both transductive and inductive settings in temporal link prediction. These findings highlight NODE-SAT's effectiveness in capturing node correlations and temporal link dynamics.", "keywords": [ "Temporal Graph", "Neural ODE", "Link prediction" ], "rating_list": [ 5, 8, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "LaNCeNmoHR", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Changyu Ren", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Hangyuan Ji", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Jian Yang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Rubing Yang", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Tao Sun", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Wei Zhang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Xianfu Cheng", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Xunliang Cai", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Principal Researcher" }, { "name": "Yang Yang", "gender": "unknown", "institution": "Southeast University", "country": "", "position": "MS student" }, { "name": "Yintong Huo", "gender": "Female", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Zhoujun Li", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Zhuoren Ye", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Mengdi Zhang", "gender": "Female", "institution": "Shandong University", "country": "", "position": "Undergrad student" }, { "name": "GuanXiangyuan", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 27, "n_ref": 49, "n_ref_all": 64, "n_fig": 6, "n_tab": 3, "L_tab": 1448, "n_element_tab": 180, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1711, "n_element_tab_1": 132, "formula_len_all": 301, "formula_len_all_1": 0, "len_all": 142448, "len_all_1": 61265, "len_abs": 1597, "len_title": 142, "len_sents": 34850, "len_sents_1": 27860, "n_sents": 257, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1607, "title": "RepoFixEval: A Repository-Level Program Repair Benchmark From Issue Discovering to Bug Fixing", "abs": "Automatic Program Repair (APR) aims to automatically fix software bugs, playing an essential role in software development. While current research demonstrates that Large Language Models (LLMs) excel in file-level program repair, their effectiveness in repository-level program repair remains unexplored. Real-world software projects, which often consist of multiple files, present significant challenges for LLMs in identifying bugs and generating fixes due to the intricate project structures. To bridge this gap, we introduce RepoFixEval, a repository-level APR benchmark consisting of 160 real-world bug fixing suites from popular Python projects. RepoFixEval provides the original buggy programs, associated issue reports, corresponding fixes, and unit tests to verify the correctness of each fix. Based on the benchmark, we further propose a three-step evaluation framework for LLM-based APR tools, encompassing (1) discovering issues from execution failures, (2) localizing buggy code segments, and (3) generating code fixes. Experimental results highlight that LLMs struggle with organizing error messages during the issue discovery phase. We find that longer contexts positively affect performance, but only a few LLMs can effectively utilize extended context information at the 128K level. Some open-source LLMs demonstrate competitiveness with closed-source counterparts, yet even the best-performing GPT-4o only resolves 12.3\\% of bugs. Our study reveals the capabilities and limitations of 16 LLMs in handling repository-level bugs, providing valuable insights for future research in this field.", "keywords": [ "Code", "Bug fix", "Program Repair", "LLM" ], "rating_list": [ 5, 3, 1, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "LZfjxvqw0N", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Guohao Dai", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Han Shi", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Principal Researcher" }, { "name": "Xian Liu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Xihui Liu", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Xuefei Ning", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Research Assistant Professor" }, { "name": "Yao Teng", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yu Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Zhenguo Li", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" } ], "n_formula": 27, "n_formula_1": 13, "n_ref_uni": 37, "n_ref": 98, "n_ref_all": 135, "n_fig": 18, "n_tab": 10, "L_tab": 1787, "n_element_tab": 176, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 734, "n_element_tab_1": 60, "formula_len_all": 2049, "formula_len_all_1": 527, "len_all": 183704, "len_all_1": 62730, "len_abs": 1783, "len_title": 149, "len_sents": 51904, "len_sents_1": 29708, "n_sents": 418, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1792, "title": "Accelerating Auto-regressive Text-to-Image Generation with Training-free Speculative Jacobi Decoding", "abs": "The current large auto-regressive models can generate high-quality, high-resolution images, but these models require hundreds or even thousands of steps of next-token prediction during inference, resulting in substantial time consumption. In existing studies, Jacobi decoding, an iterative parallel decoding algorithm, has been used to accelerate the auto-regressive generation and can be executed without training. However, the Jacobi decoding relies on a deterministic criterion to determine the convergence of iterations. Thus, it works for greedy decoding but is incompatible with sampling-based decoding which is crucial for visual quality and diversity in the current auto-regressive text-to-image generation. In this paper, we propose a training-free probabilistic parallel decoding algorithm, Speculative Jacobi Decoding (SJD), to accelerate auto-regressive text-to-image generation. By introducing a probabilistic convergence criterion, our SJD accelerates the inference of auto-regressive text-to-image generation while maintaining the randomness in sampling-based token decoding and allowing the model to generate diverse images. Specifically, SJD facilitates the model to predict multiple tokens at each step and accepts tokens based on the probabilistic criterion, enabling the model to generate images with fewer steps than the conventional next-token-prediction paradigm. We also investigate the token initialization strategies that leverage the spatial locality of visual data to further improve the acceleration ratio under specific scenarios. We conduct experiments for our proposed SJD on multiple auto-regressive text-to-image generation models, showing the effectiveness of model acceleration without sacrificing the visual quality. Code will be released upon acceptance.", "keywords": [ "auto-regressive image generation", "acceleration", "training-free", "image generation" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 2, 3, 4 ] }, { "paper_id": "LYawG8YkPa", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Laurence Tianruo Yang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Lingteng Qiu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Shenhao Zhu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Weichao Shen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Weihao Yuan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xiaodong Gu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yisheng HE", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yuan Dong", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zhe Li", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Zlong Dong", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 5, "n_ref_uni": 31, "n_ref": 116, "n_ref_all": 142, "n_fig": 12, "n_tab": 8, "L_tab": 812, "n_element_tab": 70, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 338, "n_element_tab_1": 42, "formula_len_all": 636, "formula_len_all_1": 267, "len_all": 137991, "len_all_1": 63982, "len_abs": 1539, "len_title": 83, "len_sents": 46554, "len_sents_1": 30730, "n_sents": 315, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1536, "title": "LaMP: Language-Motion Pretraining for Motion Generation, Retrieval, and Captioning", "abs": "Language plays a vital role in the realm of human motion. Existing methods have largely depended on CLIP text embeddings for motion generation, yet they fall short in effectively aligning language and motion due to CLIP\u2019s pretraining on static image-text pairs. This work introduces LaMP, a novel Language-Motion Pretraining model, which transitions from a language-vision to a more suitable language-motion latent space. It addresses key limitations by generating motion-informative text embeddings, significantly enhancing the relevance and semantics of generated motion sequences. With LaMP, we advance three key tasks: text-to-motion generation, motion-text retrieval, and motion captioning through aligned language-motion representation learning. For generation, LaMP instead of CLIP provides the text condition, and an autoregressive masked prediction is designed to achieve mask modeling without rank collapse in transformers. For retrieval, motion features from LaMP\u2019s motion transformer interact with query tokens to retrieve text features from the text transformer, and vice versa. For captioning, we finetune a large language model with the language-informative motion features to develop a strong motion captioning model. In addition, we introduce the LaMP-BertScore metric to assess the alignment of generated motions with textual descriptions. Extensive experimental results on multiple datasets demonstrate substantial improvements over previous methods across all three tasks. The code of our method will be made public.", "keywords": [ "motion generation; motion-language aliment; multimodal" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "LYHEY783Np", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Jia Deng", "gender": "Female", "institution": "Renmin University of China", "country": "", "position": "Undergrad student" }, { "name": "Tianyi Tang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Wenhao yang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "MS student" }, { "name": "Xin Zhao", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Yanbin Yin", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 57, "n_ref": 91, "n_ref_all": 137, "n_fig": 10, "n_tab": 16, "L_tab": 3345, "n_element_tab": 502, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1703, "n_element_tab_1": 287, "formula_len_all": 249, "formula_len_all_1": 249, "len_all": 212942, "len_all_1": 68922, "len_abs": 1348, "len_title": 113, "len_sents": 60669, "len_sents_1": 30486, "n_sents": 495, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1346, "title": "Neuron based Personality Trait Induction in Large Language Models", "abs": "Large language models (LLMs) have become increasingly proficient at simulating various personality traits, an important capability for supporting related applications (e.g., role-playing). To further improve this capacity, in this paper, we present a neuron based approach for personality trait induction in LLMs, with three major technical contributions. First, we construct PERSONALITYBENCH, a large-scale dataset for identifying and evaluating personality traits in LLMs. This dataset is grounded in the Big Five personality traits from psychology and designed to assess the generative capabilities of LLMs towards specific personality traits. Second, by leveraging PERSONALITYBENCH, we propose an efficient method for identifying personality-related neurons within LLMs by examining the opposite aspects of a given trait. Third, we develop a simple yet effective induction method that manipulates the values of these identified personality-related neurons, which enables fine-grained control over the traits exhibited by LLMs without training and modifying model parameters. Extensive experiments validates the efficacy of our neuron identification and trait induction methods. Notably, our approach achieves comparable performance as fine-tuned models, offering a more efficient and flexible solution for personality trait induction in LLMs.", "keywords": [ "Neuron", "Personality", "Large Language models" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 2 ], "contribution_list": [ 4, 3, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "LXlTdn9hY9", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 11, "author_info_list": [ { "name": "Colby Banbury", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "David Aponte", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Ilya Zharkov", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Research Manager" }, { "name": "Jongwoo Ko", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Kazuhito Koishida", "gender": "unknown", "institution": "Microsoft Corporation", "country": "US", "position": "Researcher" }, { "name": "Luming Liang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Tianyi Chen", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Senior Researcher" }, { "name": "Vladimir Lyapunov", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Xiaoyi Qu", "gender": "Male", "institution": "Lehigh University", "country": "US", "position": "PhD student" }, { "name": "Yong Ma", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Tianyu Ding", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" } ], "n_formula": 17, "n_formula_1": 7, "n_ref_uni": 49, "n_ref": 111, "n_ref_all": 129, "n_fig": 9, "n_tab": 12, "L_tab": 5994, "n_element_tab": 534, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1970, "n_element_tab_1": 155, "formula_len_all": 2346, "formula_len_all_1": 975, "len_all": 184587, "len_all_1": 64877, "len_abs": 1843, "len_title": 140, "len_sents": 47238, "len_sents_1": 27935, "n_sents": 384, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1805, "title": "HESSO: Towards Automatic Efficient and User Friendly Any Neural Network Training and Pruning", "abs": "Structured pruning is one of the most popular approaches to effectively compress the heavy deep neural networks (DNNs) into compact sub-networks while retaining the original network performance. The existing methods suffer from multi-stage procedures along with significant engineering efforts and human expertise. The Only-Train-Once series (OTOv1-v3) has been recently proposed to resolve the many pain points by streamlining the workflow. However, the built-in sparse optimizers in the OTO series, i.e., the Half-Space Projected Gradient (HSPG) family, have limitations that require hyper-parameter tuning and the implicit controls of the sparsity exploration, consequently requires intervening by human expertise. To further address such limitations, we propose a novel Hybrid Efficient Structured Sparse Optimizer (HESSO). HESSO could automatically and efficiently train a DNN within a single run to produce a high-performing sub-network. Meanwhile, it is almost tuning-free and enjoys user-friendly integration for generic training applications. To address another common issue of irreversible pruning performance collapse observed in some DNNs, we further propose a novel Corrective Redundant Identification Cycle (CRIC) to plug into HESSO for reliably identifying indispensable structures. We numerically demonstrate the efficacy of HESSO and its enhanced version HESSO-CRIC on a variety of applications ranging from computer vision to natural language processing, including large language model. The numerical results showcase that HESSO can achieve competitive performance to varying state-of-the-art benchmarks and support most DNN architectures. Meanwhile, CRIC can effectively prevent the irreversible performance collapse and further enhance the performance of HESSO on certain applications.", "keywords": [ "Structured Pruning", "Training", "AutoML" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "LXftdR11io", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Jihan Yao", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Thorsten Joachims", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Cornell University" }, { "name": "Yuta Saito", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" } ], "n_formula": 64, "n_formula_1": 30, "n_ref_uni": 45, "n_ref": 139, "n_ref_all": 190, "n_fig": 5, "n_tab": 3, "L_tab": 206, "n_element_tab": 32, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7031, "formula_len_all_1": 2242, "len_all": 192829, "len_all_1": 69086, "len_abs": 1439, "len_title": 133, "len_sents": 67058, "len_sents_1": 33336, "n_sents": 437, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 85, "L_abs": 1453, "title": "POTEC: Off-Policy Contextual Bandits for Large Action Spaces via Policy Decomposition", "abs": "We study off-policy learning (OPL) of contextual bandit policies in large discrete action spaces where existing methods -- most of which rely crucially on reward-regression models or importance-weighted policy gradients -- fail due to excessive bias or variance. To overcome these issues in OPL, we propose a novel two-stage algorithm, called Policy Optimization via Two-Stage Policy Decomposition (POTEC). It leverages clustering in the action space and learns two different policies via policy- and regression-based approaches, respectively. In particular, we derive a novel low-variance gradient estimator that enables to learn a first-stage policy for cluster selection efficiently via a policy-based approach. To select a specific action within the cluster sampled by the first-stage policy, POTEC uses a second-stage policy derived from a regression-based approach within each cluster. We show that a local correctness condition, which only requires that the regression model preserves the relative expected reward differences of the actions within each cluster, ensures that our policy-gradient estimator is unbiased and the second-stage policy is optimal. We also show that POTEC provides a strict generalization of policy- and regression-based approaches and their associated assumptions. Comprehensive experiments demonstrate that POTEC provides substantial improvements in OPL effectiveness particularly in large and structured action spaces.", "keywords": [ "Off-Policy Learning", "Contextual Bandits", "Large Action Space", "Importance Weighting", "Clustering", "Policy Gradient" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "LXVZQpEb2y", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Ning Liu", "gender": "Male", "institution": "Global Engineering and Materials, Inc.", "country": "US", "position": "Principal Machine Learning Scientist" }, { "name": "Tian Gao", "gender": "unknown", "institution": "International Business Machines", "country": "US", "position": "Reseach Staff Member" }, { "name": "Yue Yu", "gender": "Female", "institution": "Lehigh University", "country": "US", "position": "Full Professor" }, { "name": "LU ZHANG", "gender": "Not Specified", "institution": "Lehigh University", "country": "US", "position": "PhD student" } ], "n_formula": 42, "n_formula_1": 22, "n_ref_uni": 63, "n_ref": 109, "n_ref_all": 133, "n_fig": 11, "n_tab": 3, "L_tab": 915, "n_element_tab": 190, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 564, "n_element_tab_1": 81, "formula_len_all": 3253, "formula_len_all_1": 1782, "len_all": 196340, "len_all_1": 69367, "len_abs": 1778, "len_title": 130, "len_sents": 57656, "len_sents_1": 31440, "n_sents": 391, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1603, "title": "Disentangled Representation Learning for Parametric Partial Differential Equations", "abs": "Neural operators (NOs) have demonstrated remarkable success in learning mappings between function spaces, serving as efficient approximators for the forward solutions of complex physical systems governed by partial differential equations (PDEs). However, while effective as black-box solvers, they offer limited insight into the underlying physical mechanism, due to the lack of interpretable representations of the physical parameters that drive the system. To tackle this challenge, we propose a new paradigm for learning disentangled representations from neural operator parameters, thereby effectively solving an inverse problem. Specifically, we introduce DisentangO, a novel hyper-neural operator architecture designed to unveil and disentangle the latent physical factors of variation embedded within the black-box neural operator parameters. At the core of DisentangO is a multi-task neural operator architecture that distills the varying parameters of the governing PDE through a task-wise adaptive layer, coupled with a hierarchical variational autoencoder that disentangles these variations into identifiable latent factors. By learning these disentangled representations, our model not only enhances physical interpretability but also enables more robust generalization across diverse physical systems. Empirical evaluations across supervised, semi-supervised, and unsupervised learning contexts show that DisentangO effectively extracts meaningful and interpretable latent features, bridging the gap between predictive performance and physical understanding in neural operator frameworks.", "keywords": [ "Neural Operators; Hidden Physics Discovery; Inverse Problems" ], "rating_list": [ 8, 3, 6, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "LXBn5e4y8d", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Rui Qu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Shiyan Liu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yan Jin", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 35, "n_ref": 71, "n_ref_all": 78, "n_fig": 4, "n_tab": 2, "L_tab": 679, "n_element_tab": 94, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 3151, "n_element_tab_1": 99, "formula_len_all": 469, "formula_len_all_1": 471, "len_all": 106369, "len_all_1": 58070, "len_abs": 1375, "len_title": 156, "len_sents": 32693, "len_sents_1": 26769, "n_sents": 206, "n_sents_1": 166, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1409, "title": "FluentLip: A Phonemes-Based Two-stage Approach for Audio-Driven Lip Synthesis with Optical Flow Consistency", "abs": "Generating consecutive images of lip movements that align with a given speech in audio-driven lip synthesis is a challenging task. While previous studies have made strides in synchronization and visual quality, lip intelligibility and video fluency remain persistent challenges. This work proposes FluentLip, a two-stage approach for audio-driven lip synthesis, incorporating three featured strategies. To improve lip synchronization and intelligibility, we integrate a phoneme extractor and encoder to generate a fusion of audio and phoneme information for multimodal learning. Additionally, we employ optical flow consistency loss to ensure natural transitions between image frames. Furthermore, we incorporate a diffusion chain during the training of Generative Adversarial Networks (GANs) to improve both stability and efficiency. We evaluate our proposed FluentLip through extensive experiments, comparing it with five state-of-the-art (SOTA) approaches across five metrics, including a proposed metric called Phoneme Error Rate (PER) that evaluates lip pose intelligibility and video fluency. The experimental results demonstrate that our FluentLip approach is highly competitive, achieving significant improvements in smoothness and naturalness. In particular, it outperforms these SOTA approaches by approximately $\\\\textbf{16.3\\\\%}$ in Fr\u00e9chet Inception Distance (FID) and $\\\\textbf{35.2\\\\%}$ in PER.", "keywords": [ "Talking Face Generation", "Audio-driven lip synthesis", "Multimodal learning", "Optical flow consistency", "Generative Adversarial Networks" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 1, 3, 2 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "LX9m5iWBun", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 1, "author_info_list": [ { "name": "Lu Yu", "gender": "unknown", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 37, "n_formula_1": 30, "n_ref_uni": 48, "n_ref": 80, "n_ref_all": 86, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 100, "n_element_tab_1": 18, "formula_len_all": 4140, "formula_len_all_1": 3021, "len_all": 120790, "len_all_1": 76917, "len_abs": 801, "len_title": 120, "len_sents": 29431, "len_sents_1": 27909, "n_sents": 263, "n_sents_1": 267, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 808, "title": "Log-Concave Sampling on Compact Supports: A Versatile Proximal Framework", "abs": "In this paper, we investigate the theoretical aspects of sampling from strongly log-concave distributions defined on convex and compact supports. We propose a general proximal framework that involves projecting onto the constrained set, which is highly flexible and supports various projection options. Specifically, we consider the cases of Euclidean and Gauge projections, with the latter having the advantage of being performed efficiently using a membership oracle. This framework can be seamlessly integrated with multiple sampling methods. Our analysis focuses on Langevin-type sampling algorithms within the context of constrained sampling. We provide nonasymptotic upper bounds on the $W_1$ and $W_2$ errors, offering a detailed comparison of the performance of these methods in constrained sampling.", "keywords": [ "Markov Chain Monte Carlo", "Kinetic Langevin", "Langevin algorithm", "Midpoint method", "Mixing rate", "Proximal method" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "LWvgajBmNH", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Dacao Zhang", "gender": "unknown", "institution": "Hefei University of Technology", "country": "", "position": "MS student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Le Wu", "gender": "Female", "institution": "Hefei University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Wei Si", "gender": "Female", "institution": "IFLYTEK CO.LTD.", "country": "CN", "position": "Researcher" }, { "name": "Xin Li", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "Assistant Professor" }, { "name": "Chushimao", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 31, "n_ref": 81, "n_ref_all": 102, "n_fig": 4, "n_tab": 7, "L_tab": 2207, "n_element_tab": 297, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1731, "n_element_tab_1": 247, "formula_len_all": 543, "formula_len_all_1": 543, "len_all": 136162, "len_all_1": 66215, "len_abs": 1604, "len_title": 116, "len_sents": 40671, "len_sents_1": 29441, "n_sents": 330, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1609, "title": "MORE: A MIXTURE OF LOW-RANK EXPERTS FOR ADAPTIVE MULTI-TASK LEARNING", "abs": "With the rapid development of Large Language Models (LLMs), Parameter-Efficient Fine-Tuning (PEFT) methods have gained significant attention, which aims to achieve efficient fine-tuning of LLMs with fewer parameters. As a representative PEFT method, Low-Rank Adaptation (LoRA) introduces low-rank matrices to approximate the incremental tuning parameters and achieves impressive performance over multiple scenarios. After that, plenty of improvements have been proposed for further improvement. However, these methods either focus on single-task scenarios or separately train multiple LoRA modules for multi-task scenarios, limiting the efficiency and effectiveness of LoRA in multi-task scenarios. To better adapt to multi-task fine-tuning, in this paper, we propose a novel Mixture of Low-Rank Experts (MoRE) for multi-task PEFT. Specifically, instead of using an individual LoRA for each task, we align different ranks of LoRA module with different tasks, which we named low-rank experts. Moreover, we design a novel adaptive rank selector to select the appropriate expert for each task. By jointly training low-rank experts, MoRE can enhance the adaptability and efficiency of LoRA in multi-task scenarios. Finally, we conduct extensive experiments over multiple multi-task benchmarks along with different LLMs to verify model performance. Experimental results demonstrate that compared to traditional LoRA and its variants, MoRE significantly improves the performance of LLMs in multi-task scenarios and incurs no additional inference cost. We also release the model and code to facilitate the community.", "keywords": [ "Large Language Models", "LoRA", "Multi-task Learning", "Mixture of Experts" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 2 ] }, { "paper_id": "LWeVVPuIx0", "primary_area": "learning theory", "n_author": 7, "author_info_list": [ { "name": "Christian Kroer", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Associate Professor" }, { "name": "Chung-Wei Lee", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Gabriele Farina", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Haipeng Luo", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "Associate Professor" }, { "name": "Julien Grand-Cl\u00e9ment", "gender": "Male", "institution": "HEC Paris", "country": "", "position": "Assistant Professor" }, { "name": "Weiqiang Zheng", "gender": "Male", "institution": "Yale University", "country": "", "position": "PhD student" }, { "name": "Yang Cai", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 111, "n_formula_1": 27, "n_ref_uni": 30, "n_ref": 71, "n_ref_all": 98, "n_fig": 9, "n_tab": 2, "L_tab": 289, "n_element_tab": 5, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 15519, "formula_len_all_1": 3599, "len_all": 207130, "len_all_1": 88027, "len_abs": 3937, "len_title": 122, "len_sents": 55523, "len_sents_1": 31440, "n_sents": 560, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1283, "title": "Last-Iterate Convergence Properties of Regret-Matching Algorithms in Games", "abs": "We study last-iterate convergence properties of algorithms for solving two-player zero-sum games based on Regret Matching$^+$ (RM$^+$). Despite their widespread use for solving real games, virtually nothing is known about their last-iterate convergence. A major obstacle to analyzing RM-type dynamics is that their regret operators lack Lipschitzness and (pseudo)monotonicity.\nWe start by showing numerically that several variants used in practice, such as RM$^+$, predictive RM$^+$ and alternating RM$^+$, all lack last-iterate convergence guarantees even on a simple $3\\times 3$ matrix game.\nWe then prove that recent variants of these algorithms based on a smoothing technique, extragradient RM$^{+}$ and smooth Predictive RM$^+$, enjoy asymptotic last-iterate convergence (without a rate), $1/\\sqrt{t}$ best-iterate convergence, and when combined with restarting, linear-rate last-iterate convergence. Our analysis builds on a new characterization of the geometric structure of the limit points of our algorithms, marking a significant departure from most of the literature on last-iterate convergence. We believe that our analysis may be of independent interest and offers a fresh perspective for studying last-iterate convergence in algorithms based on non-monotone operators.", "keywords": [ "Last-Iterate Convergence", "Minty solution", "Regret Matching", "Zero-Sum Game", "Learning in Games" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "LWMS4pk2vK", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 10, "author_info_list": [ { "name": "Chi-Chih Chang", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Chien-Yu Lin", "gender": "Male", "institution": "Department of Computer Science", "country": "", "position": "PhD student" }, { "name": "Chong-Yan Chen", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Undergrad student" }, { "name": "Kai-Chiang Wu", "gender": "unknown", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" }, { "name": "Luis Ceze", "gender": "Male", "institution": "Department of Computer Science", "country": "", "position": "PhD student" }, { "name": "Ning-Chi Huang", "gender": "unknown", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Postdoc" }, { "name": "Pei-Shuo Wang", "gender": "unknown", "institution": "National Chiao Tung University, National Chiao Tung University", "country": "", "position": "MS student" }, { "name": "Wei-Cheng Lin", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "", "position": "MS student" }, { "name": "Yu-Fang Hu", "gender": "Female", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "Mohamed S Abdelfattah", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 33, "n_formula_1": 11, "n_ref_uni": 41, "n_ref": 79, "n_ref_all": 103, "n_fig": 9, "n_tab": 7, "L_tab": 2136, "n_element_tab": 298, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2354, "n_element_tab_1": 268, "formula_len_all": 2298, "formula_len_all_1": 724, "len_all": 241270, "len_all_1": 67946, "len_abs": 1575, "len_title": 99, "len_sents": 55964, "len_sents_1": 27469, "n_sents": 459, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1578, "title": "Palu: KV-Cache Compression with Low-Rank Projection", "abs": "Post-training KV-Cache compression methods typically either sample a subset of effectual tokens or quantize the data into lower numerical bit width. However, these methods cannot exploit redundancy in the hidden dimension of the KV tenors. This paper presents a hidden dimension compression approach called Palu, a KV-Cache compression framework that utilizes low-rank projection to reduce inference-time LLM memory usage. Palu decomposes the linear layers into low-rank matrices, caches compressed intermediate states, and reconstructs the full keys and values on the fly. To improve accuracy, compression rate, and efficiency, Palu further encompasses (1) a medium-grained low-rank decomposition scheme, (2) an efficient rank search algorithm, (3) low-rank-aware quantization compatibility enhancements, and (4) an optimized GPU kernel with matrix fusion. Extensive experiments with popular LLMs show that Palu compresses KV-Cache by 50% while maintaining strong accuracy and delivering up to 1.89\u00d7 speedup on the RoPE-based attention module. When combined with quantization, Palu\u2019s\ninherent quantization-friendly design yields small to negligible extra accuracy degradation while saving additional memory than quantization-only methods and achieving up to 2.91\u00d7 speedup for the RoPE-based attention. Moreover, it maintains comparable or even better accuracy (up to 1.19 lower perplexity) compared to quantization-only methods. These results demonstrate Palu\u2019s superior capability to effectively address the efficiency and memory challenges of LLM inference posed by KV-Cache.", "keywords": [ "KV-Cache", "Low-Rank Compression", "Large Language Model" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 1, 3, 3, 3 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "LW55JrLYPg", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Jiayi Chen", "gender": "Female", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Wei Dai", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yanbiao Ma", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 19, "n_ref_uni": 13, "n_ref": 27, "n_ref_all": 43, "n_fig": 3, "n_tab": 6, "L_tab": 3531, "n_element_tab": 297, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 234, "n_element_tab_1": 23, "formula_len_all": 817, "formula_len_all_1": 816, "len_all": 135202, "len_all_1": 69762, "len_abs": 1399, "len_title": 148, "len_sents": 34236, "len_sents_1": 32939, "n_sents": 265, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1430, "title": "Pursuing Better Decision Boundaries for Long-Tailed Object Detection via Category Information Amount", "abs": "In object detection, the number of instances is commonly used to determine whether a dataset follows a long-tailed distribution, implicitly assuming that the model will perform poorly on categories with fewer instances. This assumption has led to extensive research on category bias in datasets with imbalanced instance distributions. However, even in datasets with relatively balanced instance counts, models still exhibit bias toward certain categories, indicating that instance count alone cannot explain this phenomenon. In this work, we first introduce the concept and measurement of category informativeness. We observe a significant negative correlation between a category\u2019s informativeness and its accuracy, suggesting that informativeness more accurately reflects the learning difficulty of a category. Based on this observation, we propose the Informativeness-Guided Angular Margin Loss (IGAM Loss), which dynamically adjusts the decision space of categories according to their informativeness, thereby mitigating category bias in long-tailed datasets. IGAM Loss not only achieves superior performance on long-tailed benchmark datasets such as LVIS v1.0 and COCO-LT but also demonstrates significant improvements for underrepresented categories in non-long-tailed datasets like Pascal VOC. Extensive experiments confirm the potential of category informativeness as a tool and the generalizability of our proposed method.", "keywords": [ "Long-tailed recognition", "Class Imbalanced", "Image processing" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 4, 3, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "LVmafig6Tk", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Angelia Nedich", "gender": "Female", "institution": "Arizona State University", "country": "", "position": "Full Professor" }, { "name": "Daniil Vankov", "gender": "unknown", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Lalitha Sankar", "gender": "Female", "institution": "Arizona State University", "country": "", "position": "Professor" } ], "n_formula": 259, "n_formula_1": 52, "n_ref_uni": 35, "n_ref": 61, "n_ref_all": 114, "n_fig": 3, "n_tab": 4, "L_tab": 927, "n_element_tab": 70, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 22086, "formula_len_all_1": 3674, "len_all": 193833, "len_all_1": 70515, "len_abs": 2640, "len_title": 149, "len_sents": 48856, "len_sents_1": 28368, "n_sents": 562, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1057, "title": "Generalized Smooth Stochastic Variational Inequalities: Almost Sure Convergence and Convergence Rates", "abs": "This paper focuses on solving a stochastic variational inequality (SVI) problem under relaxed smoothness assumption for a class of structured non-monotone operators. The SVI problem has attracted significant interest in the machine learning community due to its immediate application to adversarial training and multi-agent reinforcement learning. In many such applications, the resulting operators do not satisfy the smoothness assumption. To address this issue, we focus on a weaker generalized smoothness assumption called $\\alpha$-symmetric. Under $p$-quasi sharpness and $\\alpha$-symmetric assumptions on the operator, we study clipped projection (gradient descent-ascent) and clipped Korpelevich (extragradient) methods. For these clipped methods, we provide the first almost-sure convergence results without making any assumptions on the boundedness of either the stochastic operator or the stochastic samples. Furthermore, we provide the first in-expectation unbiased convergence rate results for these methods under a relaxed smoothness assumption.", "keywords": [ "stochastic variational inequalities", "generalized smoothness", "clipping" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "LTpab44sdG", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Khanh Xuan Nguyen", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "Postdoc" }, { "name": "Tu Trinh", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" } ], "n_formula": 59, "n_formula_1": 14, "n_ref_uni": 29, "n_ref": 37, "n_ref_all": 49, "n_fig": 3, "n_tab": 3, "L_tab": 740, "n_element_tab": 55, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1281, "n_element_tab_1": 14, "formula_len_all": 7288, "formula_len_all_1": 1003, "len_all": 163723, "len_all_1": 82280, "len_abs": 1502, "len_title": 115, "len_sents": 52440, "len_sents_1": 35785, "n_sents": 492, "n_sents_1": 289, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1278, "title": "Practical alignment requires more than learning from human feedback", "abs": "Ensuring the alignment of artificial intelligence (AI) systems with human objectives is a critical challenge in the development of safe and effective AI technologies. Reinforcement learning from human feedback (RLHF) has been a predominant method to tackle this challenge. However, this framework operates under the unrealistic assumptions that human preferences are accurate reflections of their desires and that they remain constant over time. This paper identifies and challenges these assumptions by illustrating how they can lead to undesirable consequences, particularly when human beliefs about the environment are incorrect or mutate over time. To address these challenges, we introduce a novel framework termed practical alignment. This framework redefines the alignment objective to accommodate the variability and irrationality of human beliefs, emphasizing the need for AI systems not only to learn from but also to teach humans about the world. We discuss the theoretical underpinnings of practical alignment and introduce MindGrid, a toolkit designed to simulate and evaluate alignment scenarios. Our experimental results using large language models in teaching scenarios underscore the importance of teaching skills as a requisite capability to achieve alignment.", "keywords": [ "reinforcement learning", "alignment", "human feedback", "rlhf", "AI safety" ], "rating_list": [ 8, 5, 8, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 3, 3, 1 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "LTdtyzPQoZ", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Marco Caccamo", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Mirco Theile", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Postdoc" }, { "name": "Raphael Trumpp", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 32, "n_ref": 76, "n_ref_all": 99, "n_fig": 13, "n_tab": 4, "L_tab": 2555, "n_element_tab": 513, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1497, "n_element_tab_1": 245, "formula_len_all": 400, "formula_len_all_1": 368, "len_all": 130756, "len_all_1": 64917, "len_abs": 1065, "len_title": 148, "len_sents": 38773, "len_sents_1": 29191, "n_sents": 282, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1068, "title": "The Case for Gradual Structured Pruning in Image-based Deep Reinforcement Learning", "abs": "Scaling neural networks in image-based deep reinforcement learning often fails to improve performance. While it was shown that unstructured pruning of scaled networks can unlock performance gains, we find that refining the architecture of the scaled network yields even greater improvements. However, scaled networks in deep reinforcement learning present a practical challenge: the increased computational demands can hinder deployment on embedded devices, as commonly encountered in robotics applications. To address this, we propose a novel gradual group-structured pruning framework that allows performance gains through scaling while maintaining computational efficiency. Our method preserves the network's functional integrity of inter-layer dependencies in groups, such as residual connections, while seamlessly integrating with standard deep reinforcement learning algorithms. Experiments with PPO and DQN show that our approach sustains performance while significantly reducing inference time, making it the preferred approach for resource-limited deployment.", "keywords": [ "gradual structured pruning", "deep reinforcement learning", "Procgen benchmark" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "LTDtjrv02Y", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Andrew I. Comport", "gender": "Male", "institution": "CNRS", "country": "", "position": "Principal Researcher" }, { "name": "Antoine Schnepf", "gender": "Male", "institution": "Criteo", "country": "", "position": "PhD student" }, { "name": "Flavian Vasile", "gender": "Male", "institution": "Criteo", "country": "", "position": "Researcher" }, { "name": "Jean-Yves Franceschi", "gender": "Male", "institution": "Criteo", "country": "FR", "position": "Researcher" }, { "name": "Jeremie Mary", "gender": "Male", "institution": "Criteo", "country": "", "position": "Senior Researcher" }, { "name": "Karim Kassab", "gender": "Male", "institution": "Ecole Nationale des Sciences G\u00e9ographiques", "country": "", "position": "PhD student" }, { "name": "Laurent Caraffa", "gender": "Male", "institution": "Ecole Nationale des Sciences G\u00e9ographiques", "country": "FR", "position": "Researcher" }, { "name": "Valerie Gouet-Brunet", "gender": "Female", "institution": "LaSTIG / IGN, Univ. Gustave Eiffel", "country": "", "position": "Principal Researcher" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 36, "n_ref": 72, "n_ref_all": 99, "n_fig": 8, "n_tab": 10, "L_tab": 3836, "n_element_tab": 428, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 856, "formula_len_all_1": 657, "len_all": 164976, "len_all_1": 57608, "len_abs": 1201, "len_title": 112, "len_sents": 46568, "len_sents_1": 28039, "n_sents": 368, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1210, "title": "Bringing NeRFs to the Latent Space: Inverse Graphics Autoencoder", "abs": "While pre-trained image autoencoders are increasingly utilized in computer vision, the application of inverse graphics in 2D latent spaces has been under-explored. \nYet, besides reducing the training and rendering complexity, applying inverse graphics in the latent space enables a valuable interoperability with other latent-based 2D methods.\nThe major challenge is that inverse graphics cannot be directly applied to such image latent spaces because they lack an underlying 3D geometry. \nIn this paper, we propose an Inverse Graphics Autoencoder (IG-AE) that specifically addresses this issue.\nTo this end, we regularize an image autoencoder with 3D-geometry by aligning its latent space with jointly trained latent 3D scenes. \nWe utilize the trained IG-AE to bring NeRFs to the latent space with a latent NeRF training pipeline, which we implement in an open-source extension of the Nerfstudio framework, thereby unlocking latent scene learning for its supported methods. \nWe experimentally confirm that Latent NeRFs trained with IG-AE present an improved quality compared to a standard autoencoder, all while exhibiting training and rendering accelerations with respect to NeRFs trained in the image space.", "keywords": [ "Latent NeRF", "NeRF", "Autoencoder", "Inverse Graphics", "Nerfstudio", "3D" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 4 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "LSq9ef8ANs", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Gang Wu", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Researcher" }, { "name": "Haoliang Wang", "gender": "unknown", "institution": "Adobe Research", "country": "", "position": "Research Scientist" }, { "name": "Uttaran Bhattacharya", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Yizhou Wang", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Intern" }, { "name": "Yun Fu", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Full Professor" }, { "name": "RUIYI ZHANG", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 39, "n_ref": 116, "n_ref_all": 135, "n_fig": 14, "n_tab": 5, "L_tab": 1135, "n_element_tab": 126, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1689, "n_element_tab_1": 132, "formula_len_all": 542, "formula_len_all_1": 542, "len_all": 162800, "len_all_1": 69829, "len_abs": 1277, "len_title": 125, "len_sents": 44708, "len_sents_1": 31336, "n_sents": 402, "n_sents_1": 241, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1275, "title": "VaQuitA: Enhancing Alignment in LLM-Assisted Zero-Shot Video Understanding", "abs": "Recent advancements in language-model-based video understanding have been progressing at a remarkable pace, spurred by the introduction of Large Language Models (LLMs). However, the focus of prior research has been predominantly on devising a projection layer that maps video features to tokens, an approach that is both rudimentary and inefficient. In our study, we introduce a cutting-edge framework, VaQuitA, designed to refine the synergy between video and textual information. At the data level, instead of sampling frames uniformly, we implement a sampling method guided by CLIP-score rankings, which enables a more aligned selection of frames with the given question. At the feature level, we integrate a trainable Video Perceiver alongside a Visual-Query Transformer (abbreviated as VQ-Former), which bolsters the interplay between the input question and the video features. We also discover that incorporating a simple prompt, ``Please be critical.'', into the LLM input can substantially enhance its video comprehension capabilities. Our experimental results indicate that VaQuitA consistently sets a new benchmark for zero-shot video question-answering tasks and is adept at producing high-quality, multi-turn video dialogues with users. The code will be released.", "keywords": [ "Video Understanding", "Large Language Model", "Alignment" ], "rating_list": [ 5, 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "LSp4KBhAom", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Boris Ivanovic", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Boyi Li", "gender": "Female", "institution": "NVIDIA Research", "country": "US", "position": "Researcher" }, { "name": "Danfei Xu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Heng Yang", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" }, { "name": "Marco Pavone", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Director, Autonomous Vehicle Research" }, { "name": "Yue Wang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Ziqi Lu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 21, "n_ref": 36, "n_ref_all": 63, "n_fig": 15, "n_tab": 13, "L_tab": 12064, "n_element_tab": 1770, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1527, "n_element_tab_1": 246, "formula_len_all": 1209, "formula_len_all_1": 806, "len_all": 158358, "len_all_1": 62787, "len_abs": 1680, "len_title": 115, "len_sents": 41897, "len_sents_1": 26507, "n_sents": 327, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 67, "L_abs": 1490, "title": "LoRA3D: Low-Rank Self-Calibration of 3D Geometric Foundation models", "abs": "Emerging 3D geometric foundation models, such as DUSt3R, offer a promising approach for in-the-wild 3D vision tasks.\nHowever, due to the high-dimensional nature of the problem space and scarcity of high-quality 3D data,\nthese pre-trained models still struggle to generalize to many challenging circumstances,\nsuch as limited view overlap or low lighting.\nTo address this, we propose LoRA3D, an efficient self-calibration pipeline to *specialize* the pre-trained models to target scenes using their own multi-view predictions.\nTaking sparse RGB images as input, we leverage robust optimization techniques to refine multi-view predictions and align them into a global coordinate frame.\nIn particular, we incorporate prediction confidence into the geometric optimization process, \nautomatically re-weighting the confidence to better reflect point estimation accuracy. \nWe use the calibrated confidence to generate high-quality pseudo labels for the calibrating views and fine-tune the models using low-rank adaptation (LoRA) on the pseudo-labeled data.\nOur method does not require any external priors or manual labels. It completes the self-calibration process on a **single standard GPU within just 5 minutes**.\nEach low-rank adapter requires only **18MB** of storage. \nWe evaluated our method on **more than 160 scenes** from the Replica, TUM and Waymo Open datasets,\nachieving up to **88\\% performance improvement** on 3D reconstruction, multi-view pose estimation and novel-view rendering.", "keywords": [ "3D foundation model", "model specialization", "robust optimization", "low rank adaptation", "self-supervised learning" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "LSB2mRJdgZ", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Dit-Yan Yeung", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Chair Professor" }, { "name": "Jiangnan Li", "gender": "Male", "institution": "WeChat, Tencent Inc.", "country": "CN", "position": "Researcher" }, { "name": "Jie Zhou", "gender": "Male", "institution": "WeChat AI, Tencent Inc.", "country": "CN", "position": "Principal Researcher" }, { "name": "Junjie Wu", "gender": "Male", "institution": "HKUST", "country": "HK", "position": "PhD student" }, { "name": "Mo Yu", "gender": "Male", "institution": "WeChat AI, Tencent", "country": "", "position": "Principal Researcher" }, { "name": "Shunchi Zhang", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "MS student" }, { "name": "Tsz Ting Chung", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "lemao liu", "gender": "Male", "institution": "Tencent", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 47, "n_ref": 89, "n_ref_all": 117, "n_fig": 7, "n_tab": 9, "L_tab": 1619, "n_element_tab": 175, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 282, "n_element_tab_1": 21, "formula_len_all": 18, "formula_len_all_1": 0, "len_all": 150927, "len_all_1": 65957, "len_abs": 1118, "len_title": 97, "len_sents": 42493, "len_sents_1": 31219, "n_sents": 321, "n_sents_1": 236, "venue": "ICLR2025", "decision": "withdraw", "L_title": 98, "L_abs": 1121, "title": "The Stochastic Parrot on LLM\u2019s Shoulder: A Summative Assessment of Physical Concept Understanding", "abs": "In a systematic way, we investigate a widely asked question: Do LLMs really understand what they say?, which relates to the more familiar term Stochastic Parrot. To this end, we propose a summative assessment over a carefully designed physical concept understanding task, PHYSICO. Our task alleviates the memorization issue via the usage of grid-format inputs that abstractly describe physical phenomena. The grids represents varying levels of understanding, from the core phenomenon, application examples to analogies to other abstract patterns in the grid world. A comprehensive study on our task demonstrates that: (1) state-of-the-art LLMs lag behind humans by \u223c40%; (2) the stochastic parrot phenomenon is present in LLMs, as they fail on our grid task but can describe and recognize the same concepts well in natural language; (3) our task challenges the LLMs due to intrinsic difficulties rather than the unfamiliar grid format, as in-context learning and fine-tuning on same formatted data added little to their performance. Our data is released (see Supplementary Material in the submission) for public research.", "keywords": [ "corpus creation", "benchmarking", "evaluation methodologies" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "LS1VuhkReU", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Avi Schwarzschild", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Joshua Nathaniel Williams", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Yutong He", "gender": "Female", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zico Kolter", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 57, "n_ref_all": 72, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 183, "formula_len_all_1": 184, "len_all": 99828, "len_all_1": 61720, "len_abs": 1270, "len_title": 87, "len_sents": 35507, "len_sents_1": 31410, "n_sents": 227, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1275, "title": "Prompt Recovery for Image Generation Models: A Comparative Study of Discrete Optimizers", "abs": "Recovering natural language prompts for image generation models, solely based on the generated images is a difficult discrete optimization problem. In this work, we present the first head-to-head comparison of recent discrete optimization techniques for the problem of prompt inversion. Following prior work on prompt inversion, we use CLIP's (Radford et al., 2021) text-image alignment as an inexpensive proxy for the distribution of prompt-image pairs, and compare several discrete optimizers against BLIP2's image captioner (Li et al., 2024) and PRISM (He et al., 2024) in order to evaluate the quality of discretely optimized prompts across various metrics related to the quality of inverted prompts and the images that they generate. We find that while the discrete optimizers effectively minimize their objectives, CLIP similarity between the inverted prompts and the ground truth image acts as a poor proxy for the distribution of prompt-image pairs -- responses from well-trained captioners often lead to generated images that more closely resemble those produced by the original prompts. This finding highlights the need for further investigation into inexpensive methods of modeling the relationship between the prompts for generative models and their output space.", "keywords": [ "Discrete Optimization", "Prompt Inversion", "Benchmarking" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "LRrbD8EZJl", "primary_area": "reinforcement learning", "n_author": 9, "author_info_list": [ { "name": "Deheng Ye", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Director" }, { "name": "Jiafei Lyu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jing-Wen Yang", "gender": "Male", "institution": "Game AI Center", "country": "CN", "position": "Principal Researcher" }, { "name": "Mengbei Yan", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Runze Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xiaoteng Ma", "gender": "Male", "institution": "Department of Automation, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Xiu Li", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Zhongjian Qiao", "gender": "unknown", "institution": "Tsinghua Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Zongqing Lu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 58, "n_formula_1": 19, "n_ref_uni": 69, "n_ref": 147, "n_ref_all": 173, "n_fig": 5, "n_tab": 8, "L_tab": 5738, "n_element_tab": 457, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 1178, "n_element_tab_1": 70, "formula_len_all": 6752, "formula_len_all_1": 1309, "len_all": 239057, "len_all_1": 71772, "len_abs": 1605, "len_title": 132, "len_sents": 65757, "len_sents_1": 31687, "n_sents": 515, "n_sents_1": 276, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1613, "title": "Cross-Domain Offline Policy Adaptation with Optimal Transport and Dataset Constraint", "abs": "Offline reinforcement learning (RL) often struggles with limited data. This work explores cross-domain offline RL where offline datasets (with possibly sufficient data) from another domain can be accessed to facilitate policy learning. However, the underlying environments of the two datasets may have dynamics mismatches, incurring inferior performance when simply merging the data of two domains. Existing methods mitigate this issue by training domain classifiers, using contrastive learning methods, etc. Nevertheless, they still rely on a large amount of target domain data to function well. Instead, we address this problem by establishing a concrete performance bound of a policy given datasets from two domains. Motivated by the theoretical insights, we propose to align transitions in the two datasets using optimal transport and selectively share source domain samples, without training any neural networks. This enables reliable data filtering even given a few target domain data. Additionally, we introduce a dataset regularization term that ensures the learned policy remains within the scope of the target domain dataset, preventing it from being biased towards the source domain data. Consequently, we propose the Optimal Transport Data Filtering (dubbed OTDF) method and examine its effectiveness by conducting extensive experiments across various dynamics shift conditions (e.g., gravity shift, morphology shift), given limited target domain data. It turns out that OTDF exhibits superior performance on many tasks and dataset qualities, often surpassing prior strong baselines by a large margin.", "keywords": [ "cross-domain", "reinforcement learning", "offline RL", "optimal transport" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 4, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 2, 3, 4 ] }, { "paper_id": "LRizkALc84", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Paolo Favaro", "gender": "Male", "institution": "Institute f\u00fcr Informatik, University of Bern", "country": "CH", "position": "Full Professor" }, { "name": "Sepehr Sameni", "gender": "Male", "institution": "University of Bern, Universit\u00e4t Bern", "country": "CH", "position": "PhD student" }, { "name": "Simon Jenni", "gender": "unknown", "institution": "Adobe Systems", "country": "CH", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 48, "n_ref_all": 69, "n_fig": 4, "n_tab": 11, "L_tab": 1905, "n_element_tab": 234, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1722, "n_element_tab_1": 167, "formula_len_all": 11, "formula_len_all_1": 0, "len_all": 137962, "len_all_1": 54990, "len_abs": 1736, "len_title": 115, "len_sents": 29835, "len_sents_1": 24933, "n_sents": 197, "n_sents_1": 152, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1752, "title": "ViDROP: Video Dense Representation through Spatio-Temporal Sparsity", "abs": "Self-supervised learning (SSL) has revolutionized image processing, but extending its success to video understanding presents unique challenges due to increased data complexity and computational demands. We introduce ViDROP (Video Dense Representation thrOugh spatio-temporal sParsity), a novel SSL architecture for video understanding that combines token dropping and masking strategies. \nOur approach eliminates the need for a decoder and enables per-patch loss computation, overcoming limitations of previous video SSL methods. Moreover, we propose a simple yet effective video compression technique using k-means clustering in pixel space, significantly accelerating data loading and facilitating rapid experimentation. ViDROP demonstrates remarkable scalability across model sizes, from ViT-Small to ViT-Huge, when starting from pretrained models (VideoMAE or V-JEPA), achieving significant performance gains. Pushing the boundaries even further, we leverage network expansion techniques to successfully train ViT-Huge from scratch using modest computational resources, achieving comparable accuracy to VideoMAE 25$\\times$ faster in training time. This marks a significant breakthrough in large-scale video SSL, enabling the training of state-of-the-art models with limited resources.\nExtensive experiments show that ViDROP achieves state-of-the-art performance on various video understanding benchmarks, including Kinetics400, SSv2, UCF101, and HMDB51, as well as in temporal action detection (THUMOS14). These results highlight the effectiveness of our fine-grained token-level learning strategy in a domain traditionally dominated by fine-tuned SSL models, while enabling the training of large-scale models with limited computational resources.", "keywords": [ "Self-Supervised Learning", "Video Representation Learning", "Vision Transformers" ], "rating_list": [ 5, 3, 5, 3, 6 ], "soundness_list": [ 1, 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 2, 2 ], "contribution_list": [ 2, 1, 3, 2, 4 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "LRifwkqJEW", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 11, "author_info_list": [ { "name": "Hai Huang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Hanting Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Jieming Zhu", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Minghui Fang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Shengpeng Ji", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Shulei Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yan Xia", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zehan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhenhua Dong", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhousashuai", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 26, "n_formula_1": 23, "n_ref_uni": 43, "n_ref": 100, "n_ref_all": 125, "n_fig": 8, "n_tab": 4, "L_tab": 2106, "n_element_tab": 290, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1085, "n_element_tab_1": 137, "formula_len_all": 1836, "formula_len_all_1": 1653, "len_all": 146493, "len_all_1": 60706, "len_abs": 1243, "len_title": 102, "len_sents": 37598, "len_sents_1": 24673, "n_sents": 249, "n_sents_1": 168, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1247, "title": "Advancing Multimodal Unified Discrete Representations", "abs": "To enhance the interpretability of multimodal unified representations, many studies have focused on discrete unified representations. These efforts typically start with contrastive learning and gradually extend to the disentanglement of modal information, achieving solid multimodal discrete unified representations. However, existing research often overlooks two critical issues: 1) Different modalities have unique characteristics, and a uniform alignment approach does not fully exploit these traits; 2) The use of Euclidean distance for quantization in discrete representations often overlooks the important distinctions among different dimensions of features, resulting in redundant representations after quantization. To address these issues, we propose Fine and Coarse Cross-modal Information Disentangling (FCCID) and Training-Free Optimization of Codebook (TOC). These methods respectively perform fine and coarse disentanglement of information based on the specific characteristics of different modalities and refine the unified discrete representations obtained from pretraining. Compared to the previous state-of-the-art, our model demonstrates significant performance improvements. The code is provided in the supplementary materials.", "keywords": [ "Training-free Optimization", "MultiModal Learning", "Representation Learning" ], "rating_list": [ 5, 5, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "LRSspInlN5", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Jing Dong", "gender": "Female", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "PhD student" }, { "name": "Jingwei Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Jingzhao Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tianxing He", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 41, "n_ref": 105, "n_ref_all": 130, "n_fig": 15, "n_tab": 5, "L_tab": 1056, "n_element_tab": 79, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1056, "n_element_tab_1": 79, "formula_len_all": 863, "formula_len_all_1": 763, "len_all": 144508, "len_all_1": 63326, "len_abs": 3591, "len_title": 115, "len_sents": 41328, "len_sents_1": 28036, "n_sents": 365, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1205, "title": "Towards Black-Box Membership Inference Attack for Diffusion Models", "abs": "Given the rising popularity of AI-generated art and the associated copyright concerns, identifying whether an artwork was used to train a diffusion model is an important research topic.\nThe work approaches this problem from the membership inference attack (MIA) perspective. We first identify the limitation of applying existing MIA methods for proprietary diffusion models: the required access of internal U-nets.\nTo address the above problem, we introduce a novel membership inference attack method that uses only the image-to-image variation API and operates without access to the model's internal U-net. \\textcolor{blue}{Our method is based on the intuition that the model can more easily obtain an unbiased noise prediction estimate for images from the training set. By applying the API multiple times to the target image, averaging the outputs, and comparing the result to the original image, our approach can classify whether a sample was part of the training set.} We validate our method using DDIM and Stable Diffusion setups and further extend both our approach and existing algorithms to the Diffusion Transformer architecture. Our experimental results consistently outperform previous methods.", "keywords": [ "diffusion model", "membership inference attack" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 2, 4, 2, 4 ] }, { "paper_id": "LRPzo4jixx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 12, "author_info_list": [ { "name": "Fang Jianwei", "gender": "unknown", "institution": "China Unicom Software Research Institute", "country": "CN", "position": "Algorithm engineer" }, { "name": "Ji Pei", "gender": "unknown", "institution": "University of Massachusetts at Amherst", "country": "", "position": "PhD student" }, { "name": "Ma Xiufa", "gender": "unknown", "institution": "software research institute", "country": "CN", "position": "Full Professor" }, { "name": "Ran Chen", "gender": "unknown", "institution": "University of Massachusetts at Amherst", "country": "", "position": "PhD student" }, { "name": "Su Guangyao", "gender": "unknown", "institution": "China Unicom Software Research Institute", "country": "CN", "position": "Algorithm engineer" }, { "name": "Wang Jiancheng", "gender": "Male", "institution": "China Unicom Software Research Institute", "country": "CN", "position": "Algorithm engineer" }, { "name": "Wei Wang", "gender": "unknown", "institution": "Department of Computer Science, University of Massachusetts at Amherst", "country": "", "position": "PhD student" }, { "name": "Yijiong Yu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Yongfeng Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Zhi Xu", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Zhixiao Qi", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "weifeng.liu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 20, "n_ref": 70, "n_ref_all": 91, "n_fig": 7, "n_tab": 8, "L_tab": 2985, "n_element_tab": 307, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1193, "n_element_tab_1": 222, "formula_len_all": 250, "formula_len_all_1": 50, "len_all": 150132, "len_all_1": 54794, "len_abs": 2729, "len_title": 111, "len_sents": 51804, "len_sents_1": 24665, "n_sents": 402, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 979, "title": "Hyper-multi-step: The Truth Behind Difficult Long-context Tasks", "abs": "Long-context language models (LCLM), characterized by their extensive context window, is becoming increasingly popular. Meanwhile, many long-context benchmarks present challenging tasks that even the most advanced LCLMs struggle to complete. However, the underlying sources of various challenging long-context tasks have seldom been studied. To bridge this gap, we conduct experiments to indicate their difficulty stems primarily from two basic issues: \"multi-matching retrieval,\" which requires the simultaneous retrieval of multiple items, and \"logic-based retrieval,\" which necessitates logical judgment within retrieval criteria. These two problems, while seemingly straightforward, actually exceed the capabilities of LCLMs because they are proven to be hyper-multi-step (demanding numerous steps to solve) in nature. This finding could explain why LLMs struggle with more advanced long-context tasks, providing a more accurate perspective for rethinking solutions for them.", "keywords": [ "LLM", "long context", "reasoning", "multi-step", "benchmark", "CoT", "attention mechanism", "linear probing" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "LQzN6TRFg9", "primary_area": "generative models", "n_author": 18, "author_info_list": [ { "name": "Bin Xu", "gender": "Male", "institution": "Department of Computer Science, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Da Yin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Guanyu Feng", "gender": "Male", "institution": "Beijing Knowledge Atlas Technology Co., Ltd", "country": "", "position": "Researcher" }, { "name": "Jiayan Teng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jiazheng Xu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jie Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Ming Ding", "gender": "Male", "institution": "ZHIPU AI", "country": "CN", "position": "Principal Researcher" }, { "name": "Weihan Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Wendi Zheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Wenyi Hong", "gender": "Female", "institution": "Department of Computer Science and Technology, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiaohan Zhang", "gender": "Female", "institution": "Beijing Knowledge Atlas Technology Co., Ltd. ", "country": "CN", "position": "Researcher" }, { "name": "Xiaotao Gu", "gender": "Male", "institution": "Zhipu AI", "country": "CN", "position": "Researcher" }, { "name": "Yean Cheng", "gender": "Male", "institution": "ZhipuAI", "country": "CN", "position": "Researcher" }, { "name": "Yuanming Yang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yuxiao Dong", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuxuan.Zhang", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhuoyi Yang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Shiyu Huang", "gender": "Male", "institution": "Zhipu AI", "country": "CN", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 56, "n_ref_all": 85, "n_fig": 19, "n_tab": 9, "L_tab": 2497, "n_element_tab": 183, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 295, "n_element_tab_1": 57, "formula_len_all": 66, "formula_len_all_1": 66, "len_all": 144981, "len_all_1": 52067, "len_abs": 114, "len_title": 68, "len_sents": 37547, "len_sents_1": 23583, "n_sents": 322, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1451, "title": "CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer", "abs": "We present CogVideoX, a large-scale text-to-video generation model based on diffusion transformer, which can generate 10-second continuous videos aligned with text prompt, with a frame rate of 16 fps and resolution of 768 $\\times$ 1360 pixels. \nPrevious video generation models often had limited movement and short durations, and is difficult to generate videos with coherent narratives based on text. We propose several designs to address these issues. \nFirst, we propose a 3D Variational Autoencoder (VAE) to compress videos along both spatial and temporal dimensions, to improve both compression rate and video fidelity. Second, to improve the text-video alignment, we propose an expert transformer with the expert adaptive LayerNorm to facilitate the deep fusion between the two modalities. Third, by employing a progressive training and multi-resolution frame pack technique, \\model is adept at producing coherent, long-duration, different shape videos characterized by significant motions. \nIn addition, we develop an effective text-video data processing pipeline that includes various data preprocessing strategies and a video captioning method, greatly contributing to the generation quality and semantic alignment. \nResults show that CogVideoX demonstrates state-of-the-art performance across both multiple machine metrics and human evaluations. \nThe model weights of the 3D Causal VAE, the video caption model, and CogVideoX are open-source.", "keywords": [ "Video Generation", "Diffusion model", "Pretraining" ], "rating_list": [ 6, 6, 8, 8, 6 ], "soundness_list": [ 4, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 4, 3 ], "contribution_list": [ 4, 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 5, 4, 2 ] }, { "paper_id": "LQdaXixB0g", "primary_area": "applications to neuroscience & cognitive science", "n_author": 1, "author_info_list": [ { "name": "Declan Grabb", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Physician Fellow" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 14, "n_ref": 21, "n_ref_all": 34, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 55335, "len_all_1": 51147, "len_abs": 1321, "len_title": 152, "len_sents": 18738, "len_sents_1": 18341, "n_sents": 132, "n_sents_1": 128, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1576, "title": "pSAE-chiatry: Utilizing Sparse Autoencoders to Uncover Mental-Health-Related Features in Language Models", "abs": "As AI-powered mental health chatbots become more prevalent, their inability to recognize and respond to psychiatric emergencies, such as suicidality and mania, raises significant safety concerns. This study explores the internal representations of mental-health-related features (MHRF) in the Gemma-2-2B language model, focusing on crises related to suicide, mania, and psychosis. Using a sparse autoencoder (GemmaScope-RES-16K11) and psychiatric expertise (from M.D. mental health clinicians), MHRF's were identified across all 25 layers of the model, finding 29 features related to suicide and 42 to sadness. However, no features related to mania or paranoia were identified, suggesting critical gaps in the model\u2019s ability to handle complex psychiatric symptoms. One feature pertaining to \"suicide\" was selected for further, directed study. Four prompts (two pertaining to homicide, two pertaining to suicide) were tested to evaluate the associated activations of this particular \"suicide\" feature, and this feature was preferentially activated by prompts pertaining to suicide, supporting the relevance of the identified features. Lastly, as proof-of-concept, steering Gemma-2-2B through enhancement of this \"suicide\" feature causally impacted model behavior, making Gemma-2-2B far more likely to discuss concepts related to suicide. These findings underscore the need for improved feature identification and modulation within AI models to enhance their safety and effectiveness in mental healthcare applications. Trigger warning: This work contains references to suicide.", "keywords": [ "mental health", "psychiatry", "interpretability" ], "rating_list": [ 5, 1, 1, 3 ], "soundness_list": [ 2, 1, 1, 1 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 3, 5, 5, 3 ] }, { "paper_id": "LQRglYZ2Ri", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 10, "author_info_list": [ { "name": "Courtney A. Shearer", "gender": "Female", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Daniel Ritter", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Debora Susan Marks", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Associate Professor" }, { "name": "Erik Xie", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Felix Teufel", "gender": "unknown", "institution": "Copenhagen University", "country": "", "position": "PhD student" }, { "name": "Jonathan Frazer", "gender": "unknown", "institution": "Centre for Genomic Regulation (CRG)", "country": "ES", "position": "Assistant Professor" }, { "name": "Mafalda Dias", "gender": "Female", "institution": "Centre for Genomic Regulation (CRG)", "country": "", "position": "Assistant Professor" }, { "name": "Rose Orenbuch", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "PhD student" }, { "name": "Han Spinner", "gender": "Non-Binary", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Pascal M Notin", "gender": "Male", "institution": "Harvard Medical School, Harvard University", "country": "US", "position": "Postdoc" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 50, "n_ref": 69, "n_ref_all": 79, "n_fig": 5, "n_tab": 7, "L_tab": 4005, "n_element_tab": 554, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 552, "n_element_tab_1": 102, "formula_len_all": 355, "formula_len_all_1": 265, "len_all": 182219, "len_all_1": 51789, "len_abs": 4298, "len_title": 72, "len_sents": 35619, "len_sents_1": 25748, "n_sents": 229, "n_sents_1": 157, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1342, "title": "LOL-EVE: Predicting Promoter Variant Effects from Evolutionary Sequences", "abs": "Genetic studies reveal extensive disease-associated variation across the human genome, predominantly in noncoding regions, such as promoters. Quantifying the impact of these variants on disease risk is crucial to our understanding of the underlying disease mechanisms and advancing personalized medicine. However, current computational methods struggle to capture variant effects, particularly those of insertions and deletions (indels), which can significantly disrupt gene expression. To address this challenge, we present LOL-EVE (Language Of Life across EVolutionary Effects), a conditional autoregressive transformer model trained on 14.6 million diverse mammalian promoter sequences. Leveraging evolutionary information and proximal genetic context, LOL-EVE predicts indel variant effects in human promoter regions. We introduce three new benchmarks for indel variant effect prediction in promoter regions, comprising the identification of causal eQTLs, prioritization of rare variants in the human population, and understanding disruptions of transcription factor binding sites. We find that LOL-EVE achieves state-of-the-art performance on these tasks, demonstrating the potential of region-specific large genomic language models and offering a powerful tool for prioritizing potentially causal non-coding variants in disease studies.", "keywords": [ "Genomics", "LLM", "bioinformatics", "compuational biology" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 4, 2, 2, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "LQL5CBxLrY", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Mehrdad Mohammadian", "gender": "Male", "institution": "Azad University, Mashhad Branch", "country": "", "position": "MS student" }, { "name": "ali borji", "gender": "Male", "institution": "PrimerAI", "country": "US", "position": "ML Engineer" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 16, "n_ref": 27, "n_ref_all": 43, "n_fig": 8, "n_tab": 3, "L_tab": 1012, "n_element_tab": 197, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 997, "n_element_tab_1": 197, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 119034, "len_all_1": 65502, "len_abs": 1675, "len_title": 116, "len_sents": 31522, "len_sents_1": 29314, "n_sents": 287, "n_sents_1": 267, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1664, "title": "Battle of the Wordsmiths: Comparing ChatGPT, GPT-4, Claude, and Bard", "abs": "Although informal evaluations of modern LLMs\ncan be found on social media, blogs, and news\noutlets, a formal and comprehensive comparison among them has yet to be conducted. In\nresponse to this gap, we have undertaken an extensive benchmark evaluation of LLMs and conversational bots. Our evaluation involved the collection of 1002 questions encompassing 27 categories,\nwhich we refer to as the \u201cWordsmiths dataset.\u201d\nThese categories include reasoning, logic, facts,\ncoding, bias, language, humor, and more. Each\nquestion in the dataset is accompanied by an accurate and verified answer. We meticulously assessed four leading chatbots: ChatGPT, GPT-4, Bard, and Claude, using this dataset. The results\nof our evaluation revealed the following key findings: a) GPT-4 emerged as the top-performing\nchatbot across almost all categories, achieving a\nsuccess rate of 84.1%. On the other hand, Bard\nfaced challenges and achieved a success rate of\n62.4%. b) Among the four models evaluated,\none of them responded correctly approximately\n93% of the time. However, all models were correct only about 44%. c) Bard is less correlated\nwith other models while ChatGPT and GPT-4 are\nhighly correlated in terms of their responses. d)\nChatbots demonstrated proficiency in language\nunderstanding, facts, and self-awareness. However, they encountered difficulties in areas such\nas math, coding, IQ, and reasoning. e) In terms of\nbias, discrimination, and ethics categories, models\ngenerally performed well, suggesting they are relatively safe to utilize. To make future model evaluations on our dataset easier, we also provide a multiple-choice version of it (called WordsmithsMCQ).", "keywords": [ "Large language models", "ChatGPT", "GPT-4", "Claude", "Bard", "datasets", "natural language processing", "language modeling" ], "rating_list": [ 3, 6, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 3 ] }, { "paper_id": "LPfLsSqrQJ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Hongyu Qu", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Rui Yan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Researcher" }, { "name": "Wenguan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Xiangbo Shu", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Gao Hailiang", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 67, "n_ref": 230, "n_ref_all": 258, "n_fig": 7, "n_tab": 6, "L_tab": 2760, "n_element_tab": 388, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1555, "n_element_tab_1": 193, "formula_len_all": 443, "formula_len_all_1": 431, "len_all": 175876, "len_all_1": 70445, "len_abs": 1805, "len_title": 128, "len_sents": 43007, "len_sents_1": 31117, "n_sents": 322, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1860, "title": "Spatio-temporal Decoupled Knowledge Compensator for Few-Shot Action Recognition", "abs": "Few-Shot Action Recognition (FSAR) is a challenging task that requires recognizing novel action categories with a few labeled videos. Recent works typically apply semantically coarse category names as auxiliary contexts to guide the learning of discriminative visual features. However, such context provided by the action names is too limited to provide sufficient background knowledge for capturing novel spatial and temporal concepts in actions. In this paper, we propose $\\textbf{DiST}$, an innovative $\\textbf{D}$ecomposition-$\\textbf{i}$ncorporation framework that makes use of decoupled $\\textbf{S}$patial and $\\textbf{T}$emporal knowledge provided by large language models to learn expressive multi-granularity prototypes. In the decomposition stage, we decouple vanilla action names into diverse spatio-temporal attribute descriptions (i.e., action-related knowledge). Such commonsense knowledge complements semantic contexts from spatial and temporal perspectives. In the incorporation stage, we propose Spatial/Temporal Knowledge Compensators (SKC/PKC) to discover discriminative object- and frame-level prototypes, respectively. In SKC, object-level prototypes adaptively aggregate important patch tokens under the guidance of spatial knowledge. Moreover, in TKC, frame-level prototypes utilize temporal attributes to assist in inter-frame temporal relation modeling, further understanding diverse temporal patterns in videos. The learned prototypes at varying levels of granularity thus provide transparency in capturing fine-grained spatial details and dynamic temporal information, so as to enable accurate recognition of both appearance-centric and motion-centric actions. Experimental results show DiST achieves state-of-the-art results on four standard FSAR datasets (i.e., Kinetics, UCF101, HMDB51 and SSv2-small). Full code will be released.", "keywords": [ "Few-shot Learning", "Action Recognition", "Vision-Language Model" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 5, 2 ] }, { "paper_id": "LPXfOxe0zF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Hao Liu", "gender": "Male", "institution": "New York University", "country": "US", "position": "MS student" }, { "name": "Jiahao Yu", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Wenbo Guo", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "Assistant Professor" }, { "name": "Xian Wu", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Xinyu Xing", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 41, "n_ref_all": 58, "n_fig": 1, "n_tab": 10, "L_tab": 1679, "n_element_tab": 223, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 572, "n_element_tab_1": 128, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 134115, "len_all_1": 76110, "len_abs": 2216, "len_title": 73, "len_sents": 53453, "len_sents_1": 39367, "n_sents": 376, "n_sents_1": 274, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1322, "title": "BlockFound: Customized blockchain foundation model for anomaly detection", "abs": "We propose BlockFound, a customized foundation model for anomaly blockchain transaction detection. \nUnlike existing methods that rely on rule-based systems or directly apply off-the-shelf large language models, BlockFound introduces a series of customized designs to model the unique data structure of blockchain transactions. \nFirst, a blockchain transaction is multi-modal, containing blockchain-specific tokens, texts, and numbers. \nWe design a modularized tokenizer to handle these multi-modal inputs, balancing the information across different modalities. \nSecond, we design a customized mask language learning mechanism for pretraining with RoPE embedding and FlashAttention for handling longer sequences.\nAfter training the foundation model, we further design a novel detection method for anomaly detection. \nExtensive evaluations on Ethereum and Solana transactions demonstrate BlockFound's exceptional capability in anomaly detection while maintaining a low false positive rate. \nRemarkably, BlockFound is the only method that successfully detects anomalous transactions on Solana with high accuracy, whereas all other approaches achieved very low or zero detection recall scores.\nThis work not only provides new foundation models for blockchain but also sets a new benchmark for applying LLMs in blockchain data.", "keywords": [ "Large language model", "blockchain", "anomaly detection" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "LPRxGZ7Oax", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Yuhang Li", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "Yuheng Jia", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhuying Li", "gender": "Female", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 68, "n_formula_1": 34, "n_ref_uni": 43, "n_ref": 82, "n_ref_all": 116, "n_fig": 7, "n_tab": 14, "L_tab": 2839, "n_element_tab": 274, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1183, "n_element_tab_1": 110, "formula_len_all": 4808, "formula_len_all_1": 1649, "len_all": 168397, "len_all_1": 64908, "len_abs": 1568, "len_title": 137, "len_sents": 47386, "len_sents_1": 26878, "n_sents": 452, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1595, "title": "Complementary Label Learning with Positive Label Guessing and Negative Label Enhancement", "abs": "Complementary label learning (CLL) is a weakly supervised learning paradigm that constructs a multi-class classifier only with complementary labels, specifying classes that the instance does not belong to. We reformulate CLL as an inverse problem that infers the full label information from the output space information. To be specific, we propose to split the inverse problem into two subtasks: positive label guessing (PLG) and negative label enhancement (NLE), collectively called PLNL. Specifically, we use well-designed criteria for evaluating the confidence of the model output, accordingly divide the training instances into three categories: highly-confident, moderately-confident and under-confident. For highly-confident instances, we perform PLG to assign them pseudo labels for supervised training. For moderately-confident and under-confident instances, we perform NLE by enhancing their complementary label set at different levels and train them with the augmented complementary labels iteratively. In addition, we unify PLG and NLE into a consistent framework, in which we can view all the pseudo-labeling-based methods from the perspective of negative label recovery. We prove that the error rates of both PLG and NLE are upper bounded, and based on that we can construct a classifier consistent with that learned by clean full labels. Extensive experiments demonstrate the superiority of PLNL over the state-of-the-art CLL methods, e.g., on STL-10, we increase the classification accuracy from 34.96\\% to 55.25\\%. The source code is available at https://github.com/yhli-ml/PLNL.", "keywords": [ "Complementary label learning", "negative label enhancement", "weakly supervised learning" ], "rating_list": [ 5, 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 4, 3 ] }, { "paper_id": "LPG8pPSfQD", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Jianheng Liu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Researcher" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jun Wang", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Professor" }, { "name": "Kun Shao", "gender": "unknown", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Taiyi Wang", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Zhihao Wu", "gender": "Male", "institution": "Huawei Technologies Research & Development (UK) Ltd.", "country": "GB", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 39, "n_ref": 88, "n_ref_all": 104, "n_fig": 9, "n_tab": 7, "L_tab": 2134, "n_element_tab": 163, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 769, "formula_len_all_1": 412, "len_all": 188812, "len_all_1": 66713, "len_abs": 1647, "len_title": 146, "len_sents": 69482, "len_sents_1": 34010, "n_sents": 436, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1668, "title": "DistRL: An Asynchronous Distributed Reinforcement Learning Framework for On-Device Control Agent", "abs": "On-device control agents, especially on mobile devices, are responsible for operating mobile devices to fulfill users' requests, enabling seamless and intuitive interactions. Integrating Multimodal Large Language Models (MLLMs) into these agents enhances their ability to understand and execute complex commands, thereby improving user experience. However, fine-tuning MLLMs for on-device control presents significant challenges due to limited data availability and inefficient online training processes. This paper introduces DistRL, a novel framework designed to enhance the efficiency of online RL fine-tuning for mobile device control agents. DistRL employs centralized training and decentralized data acquisition to ensure efficient fine-tuning in the context of dynamic online interactions. Additionally, the framework is backed by our tailor-made RL algorithm, which effectively balances exploration with the prioritized utilization of collected data to ensure stable and robust training. Our experiments show that, on average, DistRL delivers a 3$\\times$ improvement in training efficiency and enables training data collection 2.4$\\times$ faster than the leading synchronous multi-machine methods. Notably, after training, DistRL achieves a 20\\% relative improvement in success rate compared to state-of-the-art methods on general Android tasks from an open benchmark, significantly outperforming existing approaches while maintaining the same training time. These results validate DistRL as a scalable and efficient solution, offering substantial improvements in both training efficiency and agent performance for real-world, in-the-wild device control tasks.", "keywords": [ "Mobile Agent", "LLM", "Reinforcement Learning", "Fine Tuning", "Distributed Training" ], "rating_list": [ 8, 8, 5, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 1, 4, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "LPDJfudDTM", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Liang Yang", "gender": "Male", "institution": "Hebei University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Linhao Li", "gender": "Male", "institution": "Hebei University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yongfeng Dong", "gender": "Male", "institution": "Hebei University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yu Wang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "zanghan", "gender": "Male", "institution": "Hebei University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 22, "n_ref_uni": 9, "n_ref": 12, "n_ref_all": 26, "n_fig": 7, "n_tab": 4, "L_tab": 1762, "n_element_tab": 352, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1762, "n_element_tab_1": 352, "formula_len_all": 752, "formula_len_all_1": 752, "len_all": 77730, "len_all_1": 59336, "len_abs": 1178, "len_title": 162, "len_sents": 23029, "len_sents_1": 23084, "n_sents": 163, "n_sents_1": 163, "venue": "ICLR2025", "decision": "withdraw", "L_title": 114, "L_abs": 1183, "title": "Multi-expert collaboration: Enhancing heterogeneous knowledge independence and alignment in knowledge distillation", "abs": "Heterogeneous multi-teacher Knowledge distillation attempt to learn a versatile student neural network from multiple pre-trained heterogeneous teachers. But current methods face issues with a lack of independence and alignment in heterogeneous knowledge. To address this issue, we propose a novel method called Multi-Expert Collaboration (MEC). Our approach aggregates multiple expert classifiers within the student model, replacing the conventional single-head architecture. By ensuring that each expert's independent classifier operates without interfering with others, we enhance the independence of heterogeneous knowledge. Inspired by Helmholtz Free Energy (HFE) theory, we introduce an anchor-based HFE self-normalization strategy to align the heterogeneous knowledge effectively. This method ensures consistent energy levels across all classifiers, allowing the appropriate classifier to achieve the highest confidence for in-distribution data. Extensive experiments on CIFAR-100 and ImageNet-100 datasets demonstrate that MEC significantly outperforms existing heterogeneous multi-teacher knowledge distillation methods, achieving an average accuracy improvement of over 10%.", "keywords": [ "knowledge distillation", "heterogeneous knowledge", "multi-teacher knowledge distillation", "independence and alignment" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "LOiYxBcGA9", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Aditya Gilra", "gender": "Male", "institution": "Centrum voor Wiskunde en Informatica", "country": "", "position": "Researcher" }, { "name": "Debabrota Basu", "gender": "unknown", "institution": "INRIA", "country": "FR", "position": "Faculty" }, { "name": "Reabetswe M. Nkhumise", "gender": "Male", "institution": "University of Sheffield", "country": "", "position": "PhD student" }, { "name": "Tony J. Prescott", "gender": "Male", "institution": "University of Sheffield", "country": "GB", "position": "Full Professor" } ], "n_formula": 98, "n_formula_1": 20, "n_ref_uni": 75, "n_ref": 119, "n_ref_all": 195, "n_fig": 8, "n_tab": 10, "L_tab": 1896, "n_element_tab": 185, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10302, "formula_len_all_1": 1590, "len_all": 239471, "len_all_1": 68281, "len_abs": 1627, "len_title": 143, "len_sents": 65854, "len_sents_1": 30927, "n_sents": 601, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1625, "title": "How does Your RL Agent Explore? An Optimal Transport Analysis of Occupancy Measure Trajectories", "abs": "The rising successes of RL are propelled by combining smart algorithmic strategies and deep architectures to optimize the distribution of returns and visitations over the state-action space. A quantitative framework to compare the learning processes of these eclectic RL algorithms is currently absent but desired in practice. We address this gap by representing the learning process of an RL algorithm as a sequence of policies generated during training, and then studying the policy trajectory induced in the manifold of occupancy measures. \nUsing an optimal transport-based metric, we measure the length of the paths induced by the policy sequence yielded by an RL algorithm between an initial policy and a final optimal policy. Hence, we first define the *Effort of Sequential Learning* (ESL). ESL quantifies the relative distance that an RL algorithm travels compared to the shortest path from the initial to the optimal policy. Further, we connect the dynamics of policies in the occupancy measure space and regret, another metric to understand the suboptimality of an RL algorithm, by defining the *Optimal Movement Ratio* (OMR). OMR assesses the fraction of movements in the occupancy measure space that effectively reduce an analogue of regret. Finally, we derive approximation guarantees to estimate ESL and OMR with finite number of samples and without access to an optimal policy. Through empirical analyses across various environments and algorithms, we demonstrate that ESL and OMR provide insights into the exploration processes of RL algorithms and hardness of different tasks in discrete and continuous MDPs.", "keywords": [ "reinforcement learning", "wasserstein distance", "occupancy measure", "exploration-exploitation", "effort of learning" ], "rating_list": [ 6, 3, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 4, 3 ] }, { "paper_id": "LOBhVTtVnc", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Di ZHANG", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "VP" }, { "name": "Hao Sun", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Jiacheng Cen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Liming Wu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Wenbing Huang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zongzhao Li", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "zhangfuzheng", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Hangyu Mao", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 21, "n_formula_1": 15, "n_ref_uni": 27, "n_ref": 58, "n_ref_all": 87, "n_fig": 4, "n_tab": 7, "L_tab": 4647, "n_element_tab": 295, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1926, "n_element_tab_1": 129, "formula_len_all": 2273, "formula_len_all_1": 1790, "len_all": 133377, "len_all_1": 63347, "len_abs": 1235, "len_title": 124, "len_sents": 39740, "len_sents_1": 25839, "n_sents": 334, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1240, "title": "Geometric Spatiotemporal Transformer to Simulate Long-Term Physical Dynamics", "abs": "Physical dynamics simulation plays a crucial role in various real-world applications. In this paper, we explore the potential of leveraging Transformers by framing the task as autoregressive next-graph prediction based on spatiotemporal graph inputs. To achieve this, we propose Geometric Spatiotemporal Transformers (GSTs), which adopt the expressive encoder-decoder architecture of traditional Transformers. At the core of GSTs are equivariant spatiotemporal blocks that alternate between spatial and temporal modules while preserving E(3) symmetries. Additionally, we introduce the Temporal Difference Graph (TDG), derived from the difference between the last two frames of historical input, to capture global dynamic patterns and mitigate cumulative errors in long-term prediction tasks. Unlike existing Graph Neural Network (GNN) methods, GSTs can process full input sequences of arbitrary lengths to effectively capture long-term context, and address cumulative errors over long-term rollouts thanks to the TDG mechanism. Our method achieves state-of-the-art performance across multiple challenging physical systems at various scales (molecular-, protein-, and macro-level), demonstrating the robust dynamics simulation capabilities.", "keywords": [ "Equivariance", "Spatio-Temporal Transformer", "Physical Dynamics" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "LOAfGVdL2G", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Alexander Podolskiy", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Alexey Rukhovich", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "RU", "position": "Researcher" }, { "name": "Irina Piontkovskaya", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" } ], "n_formula": 25, "n_formula_1": 23, "n_ref_uni": 34, "n_ref": 43, "n_ref_all": 48, "n_fig": 3, "n_tab": 6, "L_tab": 2020, "n_element_tab": 178, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 708, "n_element_tab_1": 63, "formula_len_all": 2815, "formula_len_all_1": 2616, "len_all": 108379, "len_all_1": 59985, "len_abs": 828, "len_title": 127, "len_sents": 32282, "len_sents_1": 25829, "n_sents": 264, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 833, "title": "Commute Your Domains: Trajectory Optimality Criterion for Multi-Domain Learning", "abs": "In multi-domain learning, a single model is trained on diverse data domains to leverage shared knowledge and improve generalization. The order in which the data from these domains is used for training can significantly affect the model's performance on each domain. However, this dependence is under-studied. In this paper, we investigate the influence of training order (or data mixing) in multi-domain learning using the concept of Lie bracket of gradient vector fields. By analyzing the infinitesimal effects of changing the training order, we identify regions in the parameter space where altering the order between two training domains can benefit the target loss. We validate the predictions of our theoretical framework on the influence of training order (or data mixing) both on a toy example and bilingual LLM pre-training.", "keywords": [ "Multi-domain learning", "Lie bracket", "Gradient dynamics", "Domain interaction" ], "rating_list": [ 3, 6, 3, 5, 5 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 1, 2, 3 ], "confidence_list": [ 4, 2, 3, 2, 2 ] }, { "paper_id": "LO76nlbvNt", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Ioannis Mitliagkas", "gender": "Male", "institution": "Athena Research Center", "country": "GR", "position": "Researcher" }, { "name": "Joao Monteiro", "gender": "Male", "institution": "Autodesk", "country": "GB", "position": "Researcher" }, { "name": "Mehrnaz Mofakhami", "gender": "unknown", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Researcher" }, { "name": "Reza Bayat", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "MS student" }, { "name": "Valentina Zantedeschi", "gender": "Female", "institution": "ServiceNow Research", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 45, "n_ref": 78, "n_ref_all": 120, "n_fig": 16, "n_tab": 8, "L_tab": 4004, "n_element_tab": 419, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 2102, "n_element_tab_1": 172, "formula_len_all": 60, "formula_len_all_1": 0, "len_all": 157651, "len_all_1": 66288, "len_abs": 1284, "len_title": 140, "len_sents": 47621, "len_sents_1": 31900, "n_sents": 314, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1287, "title": "Performance Control in Early Exiting to Deploy Large Models at the Same Cost of Smaller Ones", "abs": "Early Exiting (EE) is a promising technique for speeding up inference at the cost of limited performance loss. It adaptively allocates compute budget to data points based on their difficulty by exiting at earlier layers when predictions are confident. In this study, we first present a novel perspective on the EE approach, demonstrating that larger models, when deployed with EE, can achieve higher performance than smaller models while maintaining similar computational costs. As existing EE approaches rely on confidence estimation at each exit point, we further study the impact of overconfidence on the controllability of the compute/performance trade-off. We introduce Performance Control Early Exiting (PCEE), a method that enables accuracy thresholding by basing decisions not on a datapoint's condfidence but on the average accuracy of samples with similar confidence levels from a held-out validation set. In our experiments with MSDNets and Vision Transformer architectures on CIFAR-10, CIFAR-100, and ImageNet, we show that PCEE offers a simple yet computationally efficient approach that provides better control over performance than standard confidence-based approaches, and allows us to scale up model sizes to yield performance gain while reducing the computational cost.", "keywords": [ "Efficient Inference", "Early Exiting", "Performance Control", "Calibration", "Classification" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 1, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 1, 3, 2 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "LO4MEPoqrG", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Arun Suggala", "gender": "Male", "institution": "Google", "country": "IN", "position": "Researcher" }, { "name": "Karthikeyan Shanmugam", "gender": "Male", "institution": "Google Deepmind", "country": "IN", "position": "Researcher" }, { "name": "Prateek Jain", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Sravanti Addepalli", "gender": "Female", "institution": "Google", "country": "IN", "position": "Researcher" }, { "name": "Yerram Varun", "gender": "Male", "institution": "Research, Google", "country": "IN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 34, "n_ref": 105, "n_ref_all": 134, "n_fig": 7, "n_tab": 6, "L_tab": 1776, "n_element_tab": 250, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 837, "n_element_tab_1": 135, "formula_len_all": 433, "formula_len_all_1": 433, "len_all": 164093, "len_all_1": 72339, "len_abs": 1807, "len_title": 128, "len_sents": 57105, "len_sents_1": 36180, "n_sents": 358, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1811, "title": "Does Safety Training of LLMs Generalize to Semantically Related Natural Prompts?", "abs": "Large Language Models (LLMs) are known to be susceptible to crafted adversarial attacks or jailbreaks that lead to the generation of objectionable content despite being aligned to human preferences using safety fine-tuning methods. While the large dimensionality of input token space makes it inevitable to find adversarial prompts that can jailbreak these models, we aim to evaluate whether safety fine-tuned LLMs are safe against natural prompts which are semantically related to toxic seed prompts that elicit safe responses after alignment. We surprisingly find that popular aligned LLMs such as GPT-4 can be compromised using naive prompts that are NOT even crafted with an objective of jailbreaking the model. Furthermore, we empirically show that given a seed prompt that elicits a toxic response from an unaligned model, one can systematically generate several semantically related natural prompts that can jailbreak aligned LLMs. Towards this, we propose a method of Response Guided Question Augmentation (ReG-QA) to evaluate the generalization of safety aligned LLMs to natural prompts, that first generates several toxic answers given a seed question using an unaligned LLM (Q to A), and further leverages an LLM to generate questions that are likely to produce these answers (A to Q). We interestingly find that safety fine-tuned LLMs such as GPT-4o are vulnerable to producing natural jailbreak questions from unsafe content (without denial) and can thus be used for the latter (A to Q) step. We obtain attack success rates that are comparable to/ better than leading adversarial attack methods on the JailbreakBench leaderboard, while being significantly more stable against defenses such as Smooth-LLM and Synonym Substitution, which are effective against existing all attacks on the leaderboard.", "keywords": [ "jailbreaks", "LLMs" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "LNp7KW33Cg", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Gang Pan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" }, { "name": "Puli Wang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yu Qi", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "Tenure-track Research Professor" }, { "name": "Yueming Wang", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 74, "n_ref": 126, "n_ref_all": 151, "n_fig": 12, "n_tab": 15, "L_tab": 4245, "n_element_tab": 413, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 311, "n_element_tab_1": 51, "formula_len_all": 527, "formula_len_all_1": 197, "len_all": 206183, "len_all_1": 71658, "len_abs": 1188, "len_title": 86, "len_sents": 49999, "len_sents_1": 32651, "n_sents": 387, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1196, "title": "Stabilized Neural Dynamics for Behavioral Decoding via Hierarchical Domain Adaptation", "abs": "Brain-Computer Interfaces (BCI) have demonstrated significant potential in neural rehabilitation. However, the variability of non-stationary neural signals often leads to instabilities of behavioral decoding, posing critical obstacles to chronic applications. Domain adaptation technique offers a promising solution. Nonetheless, the existing direct adaptation within latent spaces could result in feature deviations. Therefore, developing a stable and efficient alignment framework is crucial for neural decoders.\nIn this work, we find that dynamical latent features can be extracted from neural dynamics utilizing causal architectures. \nWe also demonstrate that the process of self-consistent alignment can generate more stable latent features.\nBased on these insights, we propose a novel hierarchical domain adaptation (HDA) method for the alignment of dynamical latent features.\nUsing Lyapunov theory, we further analytically validate the stability of dynamical features, which experimentally exhibit significant enhancements across various datasets.\nOur HDA approach effectively addresses the challenge of non-stationary neural signals, thereby potentially improving the reliability of BCIs.", "keywords": [ "Unsupervised Domain Adaptation", "Brain-Computer Interface", "Neural Dynamics", "Lyapunov Theory" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "LNkMWCEssX", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Xiaoming Huo", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Ziyan Wang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Hao Wang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 56, "n_formula_1": 32, "n_ref_uni": 43, "n_ref": 97, "n_ref_all": 126, "n_fig": 2, "n_tab": 4, "L_tab": 2892, "n_element_tab": 215, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8192, "formula_len_all_1": 2432, "len_all": 180801, "len_all_1": 65692, "len_abs": 910, "len_title": 97, "len_sents": 54168, "len_sents_1": 26854, "n_sents": 510, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 915, "title": "Towards Domain Adaptive Neural Contextual Bandits", "abs": "Contextual bandit algorithms are essential for solving real-world decision making problems. In practice, collecting a contextual bandit's feedback from different domains may involve different costs. For example, measuring drug reaction from mice (as a source domain) and humans (as a target domain). Unfortunately, adapting a contextual bandit algorithm from a source domain to a target domain with distribution shift still remains a major challenge and largely unexplored. In this paper, we introduce the first general domain adaptation method for contextual bandits. Our approach learns a bandit model for the target domain by collecting feedback from the source domain. Our theoretical analysis shows that our algorithm maintains a sub-linear regret bound even adapting across domains. Empirical results show that our approach outperforms the state-of-the-art contextual bandit algorithms on real-world datasets.", "keywords": [ "Domain Adaptation", "Deep Learning", "Adversarial Learning" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "LNYL96VIsD", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 16, "author_info_list": [ { "name": "David A. Clifton", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Dongsheng Li", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Fan Yang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Professor" }, { "name": "Fang Dong", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Hengjie Cao", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Jixian Zhou", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Li Shang", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Mengyi Chen", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Mingzhi Dong", "gender": "Male", "institution": "University of Bath", "country": "CN", "position": "Lecturer" }, { "name": "Ning Gu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Qin Lv", "gender": "Female", "institution": "University of Colorado at Boulder", "country": "US", "position": "Full Professor" }, { "name": "Robert P. Dick", "gender": "Male", "institution": "University of Michigan", "country": "US", "position": "Full Professor" }, { "name": "Ruijun Huang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Tun Lu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yifeng Yang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Yujiang Wang", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 30, "n_ref": 37, "n_ref_all": 55, "n_fig": 6, "n_tab": 2, "L_tab": 235, "n_element_tab": 24, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 235, "n_element_tab_1": 24, "formula_len_all": 416, "formula_len_all_1": 394, "len_all": 109938, "len_all_1": 59897, "len_abs": 1939, "len_title": 158, "len_sents": 35684, "len_sents_1": 28955, "n_sents": 243, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 110, "L_abs": 1965, "title": "Large Learning Rates without the Agonizing Pain: Dispelling the Curse of Singularities in Deep Neural Networks", "abs": "Employing large learning rates (LRs) in deep learning can accelerate convergence and improve generalization, but it can also cause training instability and loss explosion: determining an appropriate LR is an often laborious and painful art. Our study into the fine-grained behaviors of parametric singularities, specifically the stable ranks of weight matrices of network components, reveals a strong connection between these singularities and training instability. As training progresses, parametric singularities trend upward, a phenomenon that is directly aggravated by large LRs. Crucially, several training steps before prominent instabilities such as gradient explosions, we observe unusually high parametric singularities across the network components, leading to rank-deficient representations. These representations, in turn, amplify parametric singularities during backpropagation, creating a vicious cycle that eventually results in loss explosions. We refer to this phenomenon as \\textit{the curse of singularities}. \nBuilding on this understanding, we propose a lightweight and robust stabilization method called Parametric Singularity Smoothing (PSS), which allows for early intervention and mitigates impending instability by smoothing the singular spectra of weight matrices, thereby preventing the curse of singularities.\nThis approach is easy to implement, works at any stage of training by restoring stable training even after instability, \nhas neglectable computational overhead, and, most importantly, frees us from the painful LR fine-tunings to avoid instabilities. Experimental results across various datasets, networks, and optimizers demonstrate that our approach allows a 5-10$\\times$ increase in LR without producing instability, attaining better training efficiency and generalization. We release our code for everyone to use our methods and reproduce the experiments, available at https://anonymous.4open.science/r/ICLR_stability-C69C.", "keywords": [ "learning rate", "training stability", "parametric singularity" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 3, 4, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 3, 2 ] }, { "paper_id": "LNYIUouhdt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Mi Zhang", "gender": "Male", "institution": "The Ohio State University", "country": "", "position": "Associate Professor" }, { "name": "Xin Wang", "gender": "Male", "institution": "The Ohio StateUniversity", "country": "US", "position": "PhD student" }, { "name": "Yu Zheng", "gender": "Male", "institution": "Michigan State University", "country": "", "position": "PhD student" }, { "name": "Zhongwei Wan", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 12, "n_ref_uni": 36, "n_ref": 69, "n_ref_all": 93, "n_fig": 5, "n_tab": 11, "L_tab": 4376, "n_element_tab": 462, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1911, "n_element_tab_1": 112, "formula_len_all": 2743, "formula_len_all_1": 1051, "len_all": 185257, "len_all_1": 68358, "len_abs": 1110, "len_title": 139, "len_sents": 50889, "len_sents_1": 30353, "n_sents": 369, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1122, "title": "SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression", "abs": "The advancements in Large Language Models (LLMs) have been hindered by their substantial sizes, which necessitate LLM compression methods for practical deployment. Singular Value Decomposition (SVD) offers a promising solution for LLM compression. However, state-of-the-art SVD-based LLM compression methods have two key limitations: truncating smaller singular values may lead to higher compression loss, and the lack of update on the compressed weight after SVD truncation. In this work, we propose \\sysname, a new SVD-based LLM compression method that addresses the limitations of existing methods. \\sysname incorporates a truncation-aware data whitening strategy to ensure a direct mapping between singular values and compression loss. Moreover, \\sysname adopts a parameter update with sequential low-rank approximation to compensate for the accuracy degradation after compression. We evaluate \\sysname on $10$ datasets and seven models from three different LLM families at three different scales. Our results demonstrate the superiority of \\sysname over state-of-the-arts, especially at high model compression ratios.", "keywords": [ "Large Language Model; Post-training Model Compression" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 2, 4, 3 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "LNL7zKvm7e", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 12, "author_info_list": [ { "name": "Bingni Zhang", "gender": "unknown", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "CHENGKAI JIN", "gender": "unknown", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Hao Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Huanyu Wang", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Jiawei Wu", "gender": "Male", "institution": "TikTok, ByteDance", "country": "SG", "position": "Researcher" }, { "name": "Qianru Sun", "gender": "Female", "institution": "Singapore Management University", "country": "SG", "position": "Associate Professor" }, { "name": "Sheng Jin", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Sicheng Yu", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "XU XIAOLEI", "gender": "Male", "institution": "ByteDance Inc.", "country": "AU", "position": "Researcher" }, { "name": "ZHONGRONG ZUO", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Zhenbang Sun", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Zhenghao Chen", "gender": "Male", "institution": "University of Newcastle", "country": "AU", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 42, "n_ref": 126, "n_ref_all": 162, "n_fig": 11, "n_tab": 10, "L_tab": 2742, "n_element_tab": 299, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 267, "n_element_tab_1": 38, "formula_len_all": 243, "formula_len_all_1": 126, "len_all": 203288, "len_all_1": 59905, "len_abs": 1355, "len_title": 119, "len_sents": 49492, "len_sents_1": 30353, "n_sents": 366, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1300, "title": "Frame-Voyager: Learning to Query Frames for Video Large Language Models", "abs": "Video Large Language Models (Video-LLMs) have made remarkable progress in video understanding tasks. However, they are constrained by the maximum length of input tokens, making it impractical to input entire videos. Existing frame selection approaches, such as uniform frame sampling and text-frame retrieval, fail to account for the information density variations in the videos or the complex instructions in the tasks, leading to sub-optimal performance. In this paper, we propose Frame-Voyager that learns to query informative frame combinations, based on the given textual queries in the task. To train Frame-Voyager, we introduce a new data collection and labeling pipeline, by ranking frame combinations using a pre-trained Video-LLM. Given a video of M frames, we traverse its T-frame combinations, feed them into a Video-LLM, and rank them based on Video-LLM's prediction losses. Using this ranking as supervision, we train Frame-Voyager to query the frame combinations with lower losses. In experiments, we evaluate Frame-Voyager on four Video Question Answering benchmarks by plugging it into two different Video-LLMs. The experimental results demonstrate that Frame-Voyager achieves impressive results in all settings, highlighting its potential as a plug-and-play solution for Video-LLMs.", "keywords": [ "Video-LLM", "Adaptive Frame Sampling" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "LM4PYXBId5", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Christina Sartzetaki", "gender": "unknown", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Gemma Roig", "gender": "Female", "institution": "Johann Wolfgang Goethe Universit\u00e4t Frankfurt am Main", "country": "DE", "position": "Full Professor" }, { "name": "Iris Groen", "gender": "Female", "institution": "University of Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Cees Snoek", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 62, "n_ref": 103, "n_ref_all": 123, "n_fig": 10, "n_tab": 1, "L_tab": 448, "n_element_tab": 48, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 448, "n_element_tab_1": 48, "formula_len_all": 267, "formula_len_all_1": 267, "len_all": 171835, "len_all_1": 59370, "len_abs": 1559, "len_title": 126, "len_sents": 46559, "len_sents_1": 31430, "n_sents": 253, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1516, "title": "One Hundred Neural Networks and Brains Watching Videos: Lessons from Alignment", "abs": "What can we learn from comparing video models to human brains, arguably the most efficient and effective video processing systems in existence? Our work takes a step towards answering this question by performing the first large-scale benchmarking of deep video models on representational alignment to the human brain, using publicly available models and a recently released video brain imaging (fMRI) dataset. We disentangle four factors of variation in the models (temporal modeling, classification task, architecture, and training dataset) that affect alignment to the brain, which we measure by conducting Representational Similarity Analysis across multiple brain regions and model layers. We show that temporal modeling is key for alignment to brain regions involved in early visual processing, while a relevant classification task is key for alignment to higher-level regions. Moreover, we identify clear differences between the brain scoring patterns across layers of CNNs and Transformers, and reveal how training dataset biases transfer to alignment with functionally selective brain areas. Additionally, we uncover a negative correlation of computational complexity to brain alignment. Measuring a total of 99 neural networks and 10 human brains watching videos, we aim to forge a path that widens our understanding of temporal and semantic video representations in brains and machines, ideally leading towards more efficient video models and more mechanistic explanations of processing in the human brain.", "keywords": [ "representational alignment", "Representational Similarity Analysis", "RSA", "benchmarking", "neuro-AI", "video AI", "neuroscience", "fMRI", "cognitive AI" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 1 ] }, { "paper_id": "LLtUtzSOL5", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Alexander Huth", "gender": "unknown", "institution": "The University of Texas at Austin", "country": "", "position": "Assistant Professor" }, { "name": "Jianing Mu", "gender": "Female", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Mathis Pink", "gender": "Male", "institution": "MPI-SWS", "country": "DE", "position": "PhD student" }, { "name": "Qinyuan Wu", "gender": "Female", "institution": "MPI-SWS", "country": "DE", "position": "PhD student" }, { "name": "Sebastian Michelmann", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Uri Hasson", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "", "position": "PhD student" }, { "name": "Javier Turek", "gender": "Male", "institution": "Earth Dynamics.ai", "country": "US", "position": "Principal Researcher" }, { "name": "Kenneth Norman", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Full Professor" }, { "name": "Mariya K Toneva", "gender": "Female", "institution": "Max Planck Institute for Software Systems", "country": "DE", "position": "Assistant Professor" }, { "name": "Vy A Vo", "gender": "Female", "institution": "Intel", "country": "US", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 54, "n_ref": 73, "n_ref_all": 125, "n_fig": 21, "n_tab": 7, "L_tab": 5766, "n_element_tab": 449, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 886, "n_element_tab_1": 87, "formula_len_all": 463, "formula_len_all_1": 180, "len_all": 232237, "len_all_1": 62045, "len_abs": 1669, "len_title": 114, "len_sents": 65482, "len_sents_1": 28862, "n_sents": 487, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1465, "title": "Assessing Episodic Memory in LLMs with Sequence Order Recall Tasks", "abs": "Current LLM benchmarks focus on evaluating models\u2019 memory of facts and\nsemantic relations, primarily assessing semantic aspects of long-term memory.\nHowever, in humans, long-term memory also includes episodic memory, which\nlinks memories to their contexts, such as the time and place they occurred. The\nability to contextualize memories is crucial for many cognitive tasks and everyday\nfunctions. Existing benchmarks have poor coverage of episodic memory. To\naddress the gap in evaluating memory in LLMs, we define episodic memory for\nLLMs and introduce Sequence Order Recall Tasks (SORT), which we adapt from\ntasks used in cognitive psychology. SORT requires causal LLMs to recall the\ncorrect order of text segments, and provides a general framework that is both easily\nextendable and does not require any additional annotations. We present an initial\nevaluation dataset, Book-SORT, comprising 36k pairs of segments extracted from 9\nbooks recently added to the public domain. Based on a human experiment with 155\nparticipants, we show that humans can recall sequence order based on long-term\nmemory of a book. We find that models can perform the task with high accuracy\nwhen relevant text is given in-context during the SORT evaluation. However, when\npresented with the book text only during training, LLMs\u2019 performance on SORT\nfalls short. By evaluating a new aspect of memory, we believe that SORT will aid\nin the emerging development of memory-augmented models.", "keywords": [ "Large Language Models (LLMs)", "Episodic Memory", "Long-Term Memory", "Working Memory", "Long-context", "Memory Evaluation", "Memory Benchmark", "Memory-augmentation", "Retrieval-augmented generation", "long-document evaluation", "neuroscience", "cognitive science", "NeuroAI" ], "rating_list": [ 8, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 1, 4 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "LLWj8on4Rv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Christian Vater", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Luc Van Gool", "gender": "unknown", "institution": "INSAIT - Sofia Un.", "country": "BG", "position": "Full Professor" }, { "name": "Nedko Savov", "gender": "Male", "institution": "INSAIT", "country": "BG", "position": "Researcher" }, { "name": "Nikola Popovic", "gender": "Male", "institution": "Institute for Computer Science, Artificial Intelligence and Technology", "country": "BG", "position": "Researcher" }, { "name": "Otmar Hilliges", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "Full Professor" }, { "name": "Xi Wang", "gender": "Female", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Danda Pani Paudel", "gender": "Male", "institution": "INSAIT Sofia & ETH Zurich", "country": "BG", "position": "Assistant Professor" }, { "name": "Murtaza Eren Akbiyik", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 72, "n_ref": 122, "n_ref_all": 163, "n_fig": 19, "n_tab": 7, "L_tab": 1500, "n_element_tab": 194, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 722, "n_element_tab_1": 66, "formula_len_all": 426, "formula_len_all_1": 195, "len_all": 214959, "len_all_1": 64852, "len_abs": 1425, "len_title": 120, "len_sents": 52495, "len_sents_1": 29660, "n_sents": 419, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1433, "title": "Leveraging Driver Field-of-View for Multimodal Ego-Trajectory Prediction", "abs": "Understanding drivers\u2019 decision-making is crucial for road safety. Although predicting the ego-vehicle\u2019s path is valuable for driver-assistance systems, existing methods mainly focus on external factors like other vehicles\u2019 motions, often neglecting the driver\u2019s attention and intent. To address this gap, we infer the ego-trajectory by integrating the driver\u2019s attention and the surrounding scene. We introduce RouteFormer, a novel multimodal ego-trajectory prediction network combining GPS data, environmental context, and driver field-of-view\u2014comprising first-person video and gaze fixations. We also present the Path Complexity Index (PCI), a new metric for trajectory complexity that enables a more nuanced evaluation of challenging scenarios. To tackle data scarcity and enhance diversity, we introduce GEM, a comprehensive dataset of urban driving scenarios enriched with synchronized driver field-of-view and gaze data. Extensive evaluations on GEM and DR(eye)VE demonstrate that RouteFormer significantly outperforms state-of-the-art methods, achieving notable improvements in prediction accuracy across diverse conditions. Ablation studies reveal that incorporating driver field-of-view data yields significantly better average displacement error, especially in challenging scenarios with high PCI scores, underscoring the importance of modeling driver attention. All data, code, and models will be made publicly available.", "keywords": [ "Ego-trajectory prediction", "driver attention", "multimodal learning", "field-of-view", "gaze fixations", "deep learning", "autonomous driving", "driver behavior modeling", "dataset creation" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "LKUVlhjgOw", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Changliu Liu", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Hanjiang Hu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 46, "n_formula_1": 23, "n_ref_uni": 33, "n_ref": 79, "n_ref_all": 115, "n_fig": 4, "n_tab": 9, "L_tab": 2631, "n_element_tab": 185, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 1867, "n_element_tab_1": 132, "formula_len_all": 5488, "formula_len_all_1": 2571, "len_all": 150285, "len_all_1": 59173, "len_abs": 1641, "len_title": 115, "len_sents": 39139, "len_sents_1": 25507, "n_sents": 265, "n_sents_1": 159, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1587, "title": "Guaranteed Neural PDE Boundary Control with Neural Barrier Function", "abs": "The physical world dynamics are generally governed by underlying partial derivative equations (PDEs) with unknown analytical forms in science and engineering problems. Neural network based data-driven approaches have been heavily studied in simulating and solving PDE problems in recent years, but it is still challenging to move forward from understanding to controlling the unknown PDE dynamics. PDE boundary control instantiates a simplified but important problem by only focusing on PDE boundary conditions as the control input and output. However, current model-free PDE controllers cannot ensure the boundary output satisfies some given user-specified safety constraint. To this end, we propose a safety filtering framework to guarantee the boundary output stays within the safe set for current model-free controllers. Specifically, we first introduce a general neural boundary control barrier function (BCBF) to ensure the feasibility of the trajectory-wise constraint satisfaction of boundary output. Based on a neural operator modeling the transfer function from boundary control input to output trajectories, we show that the change in the BCBF depends linearly on the change in input boundary, so \nquadratic programming-based safety filtering can be done for pre-trained model-free controllers. Extensive experiments under challenging hyperbolic, parabolic and Navier-Stokes PDE dynamics environments validate the effectiveness of the proposed method in achieving better general performance and boundary constraint satisfaction compared to the model-free controller baselines.", "keywords": [ "PDE boundary control", "safe control", "learning for control" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 5, 3, 1 ] }, { "paper_id": "LJWPYzjDz4", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 7, "author_info_list": [ { "name": "Chun Jason Xue", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" }, { "name": "Hong Xu", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Jingzong LI", "gender": "Male", "institution": "The Hang Seng University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Jun Wang", "gender": "Female", "institution": "City University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Nan Guan", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Tei-Wei Kuo", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "Full Professor" }, { "name": "YU MAO", "gender": "Female", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" } ], "n_formula": 20, "n_formula_1": 13, "n_ref_uni": 20, "n_ref": 41, "n_ref_all": 60, "n_fig": 13, "n_tab": 3, "L_tab": 873, "n_element_tab": 149, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 119, "n_element_tab_1": 14, "formula_len_all": 1012, "formula_len_all_1": 858, "len_all": 117075, "len_all_1": 63549, "len_abs": 1201, "len_title": 69, "len_sents": 39151, "len_sents_1": 29130, "n_sents": 317, "n_sents_1": 235, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1109, "title": "Extending Flexibility of Image Coding Enhancement Framework for IoTs", "abs": "Neural image compression, necessary in various edge-device scenarios, suffers from its heavy encode-decode structures and inflexible compression level switch. The primary issue is that the computational and storage capabilities of edge devices are weaker than those of servers, preventing them from handling the same amount of computation and storage. One solution is to downsample images and reconstruct them on the receiver side; however, current methods uniformly downsample the image and limit flexibility in compression levels. We take a step to break up this paradigm by proposing a conditional uniform-based sampler that allows for flexible image size reduction and reconstruction. Building on this, we introduce a lightweight transformer-based reconstruction structure to further reduce the reconstruction load on the receiver side. Extensive evaluations conducted on a real-world testbed demonstrate multiple advantages of our system over existing compression techniques, especially in terms of adaptability to different compression levels, computational efficiency, and image reconstruction quality.", "keywords": [ "Data Compression", "IoT infrastructure", "Edge Computing", "Scalable Design" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "LJULZNlW5d", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Qiao Yan", "gender": "Female", "institution": "Shenzhen University", "country": "CN", "position": "Full Professor" }, { "name": "Xinhao Liu", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "MS student" }, { "name": "\u848b\u82f1\u8087", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 29, "n_ref": 51, "n_ref_all": 72, "n_fig": 7, "n_tab": 1, "L_tab": 1418, "n_element_tab": 16, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1464, "n_element_tab_1": 19, "formula_len_all": 644, "formula_len_all_1": 609, "len_all": 102471, "len_all_1": 52625, "len_abs": 1699, "len_title": 117, "len_sents": 26034, "len_sents_1": 24413, "n_sents": 196, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1742, "title": "Vanishing Privacy: Fast Gradient Leakage Threat to Federated Learning", "abs": "In the federated learning (FL) framework, clients participate in collaborative learning tasks under the coordination of a central server. Clients train local submodels using their own data and share gradients with the server, which aggregates the gradients to achieve privacy protection. However, recent research has revealed that gradient inversion attacks (GIAs) can leak private data from the shared gradients. \nPrior work has only demonstrated the feasibility of recovering input data from gradients under highly restrictive conditions, such as when dealing with high-resolution face datasets, where GIAs often struggle to initiate attacks effectively, and on object datasets like Imagenet, where they encounter limitations, primarily manifested in their ability to handle only small batch sizes and high time costs.\nAs a result, we believe that implementing GIAs on high-resolution face datasets with large batch sizes is a challenging task. In this work, we introduce \\textbf{F}ast \\textbf{G}radient \\textbf{L}eakage (FGL), which enables rapid image recovery across various network models on complex datasets, including the CelebA face dataset (1000 classes, 224$\\times $224 px).\nWe also introduced StyleGAN as prior knowledge for images and achieved FGL with a batch size of 60 in experiments (constrained by experimental hardware).\nWe further propose a joint gradient matching loss, where multiple distinct matching losses collectively contribute to clarifying the attack direction and enhancing the efficiency of the optimization process.\nExtensive experimentation validates the feasibility of our approach. We anticipate that our proposed method can serve as a valuable tool to advance the development of privacy defense techniques.", "keywords": [ "Gradient inversion attacks", "Federated learning", "AI security" ], "rating_list": [ 3, 1, 3, 5 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "LJGY2GVcit", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Daniel Soudry", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Associate Professor" }, { "name": "Edan kinderman", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "MS student" }, { "name": "Haggai Maron", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Assistant Professor" }, { "name": "Itay Hubara", "gender": "Male", "institution": "Technion, Technion", "country": "", "position": "PhD student" } ], "n_formula": 39, "n_formula_1": 22, "n_ref_uni": 71, "n_ref": 184, "n_ref_all": 235, "n_fig": 8, "n_tab": 25, "L_tab": 9783, "n_element_tab": 1366, "n_fig_1": 8, "n_tab_1": 7, "L_tab_1": 2868, "n_element_tab_1": 439, "formula_len_all": 2439, "formula_len_all_1": 1240, "len_all": 324733, "len_all_1": 74229, "len_abs": 1049, "len_title": 141, "len_sents": 90092, "len_sents_1": 28338, "n_sents": 735, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1052, "title": "Foldable SuperNets: Scalable Merging of Transformers with Different Initializations and Tasks", "abs": "Many recent methods aim to merge neural networks (NNs) with identical architectures trained on different tasks to obtain a single multi-task model. Most existing works tackle the simpler setup of merging NNs initialized from a common pre-trained network, where simple heuristics like weight averaging work well. This work targets a more challenging goal: merging large transformers trained on different tasks from distinct initializations. First, we demonstrate that traditional merging methods fail catastrophically in this setup. To overcome this challenge, we propose Foldable SuperNet Merge (FS-Merge), a method that optimizes a SuperNet to fuse the original models using a feature reconstruction loss. FS-Merge is simple, data-efficient, and capable of merging models of varying widths. We test FS-Merge against existing methods, including knowledge distillation, on MLPs and transformers across various settings, sizes, tasks, and modalities. FS-Merge consistently outperforms them, achieving SOTA results, particularly in limited data scenarios.", "keywords": [ "Model merging", "Knowledge Distillation", "Deep Learning" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "LJ1zlaGdPm", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Jingang Wang", "gender": "Male", "institution": "Meituan", "country": "", "position": "Researcher" }, { "name": "Junxian He", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Shiqi Chen", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Intern" }, { "name": "Siqi Wang", "gender": "unknown", "institution": "University of Hong Kong", "country": "", "position": "MS student" }, { "name": "Xunliang Cai", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Principal Researcher" }, { "name": "Yudong Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhengyu Chen", "gender": "unknown", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Teng Xiao", "gender": "Male", "institution": "The Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 16, "n_ref_uni": 30, "n_ref": 75, "n_ref_all": 104, "n_fig": 14, "n_tab": 5, "L_tab": 1144, "n_element_tab": 175, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 212, "n_element_tab_1": 23, "formula_len_all": 997, "formula_len_all_1": 404, "len_all": 172723, "len_all_1": 48071, "len_abs": 1196, "len_title": 141, "len_sents": 53931, "len_sents_1": 23237, "n_sents": 386, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1204, "title": "From Scaling Law to Sub-Scaling Law: Understanding the Diminishing Returns of Larger Models", "abs": "Traditional scaling laws suggest that performance metrics of language models improve predictably with increases in model or dataset size. However, recent works display sub-scaling growth for large language models, where performance improvements decelerate as the dataset or model size increases. \nThis study aims to systematically investigate the sub-scaling law phenomenon through an extensive empirical analysis involving over 400 models, ranging from 20 million to 7 billion parameters, with varying datasets and training strategies.\nOur findings indicate that sub-scaling laws arise primarily from high data density and non-optimal training resource allocations. \nSpecifically, we observed that both factors contribute more significantly to performance deceleration than previously anticipated. We examine the sub-scaling phenomenon from two perspectives: data density and training strategy. High data density leads to diminishing marginal gains in performance, while optimal resource allocation is crucial for sustaining performance improvements.\nFurther, we propose a sub-optimal scaling law that generalizes the Chinchilla scaling law to better predict performance and loss in sub-scaling regimes.", "keywords": [ "scaling law", "large language model" ], "rating_list": [ 5, 1, 6, 6 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "LIz0cBmHn5", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Jiamin He", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "PhD student" }, { "name": "Martha White", "gender": "Female", "institution": "University of Alberta", "country": "", "position": "Associate Professor" }, { "name": "Samuel Neumann", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "PhD student" }, { "name": "Adam White", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 20, "n_ref_uni": 30, "n_ref": 66, "n_ref_all": 105, "n_fig": 18, "n_tab": 1, "L_tab": 229, "n_element_tab": 21, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2900, "formula_len_all_1": 1421, "len_all": 157637, "len_all_1": 59968, "len_abs": 960, "len_title": 114, "len_sents": 54866, "len_sents_1": 26828, "n_sents": 484, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 961, "title": "Investigating Mixture Policies in Entropy-Regularized Actor-Critic", "abs": "We study mixture policies in entropy-regularized reinforcement learning. Mixture policies offer greater flexibility than base policies like Gaussians, which we show theoretically provides improved solution quality and robustness to the entropy scale. Despite these potential benefits, they are rarely used for algorithms like Soft Actor-Critic, potentially due to the fact that Gaussians are easily reparameterized to get lower variance gradient updates, but mixtures are not. We fill this gap, introducing reparameterization gradient estimators for the mixture policy. Through extensive experiments on environments from classic control, MuJoCo, the DeepMind Control Suite and a suite of randomly generated bandits, our results show that mixture policies explore more efficiently in tasks with unshaped rewards (across entropy scales), while performing comparably to base policies in tasks with shaped rewards, and are more robust to multimodal critic surfaces.", "keywords": [ "policy parameterization", "entropy regularization", "actor-critic", "policy optimization", "exploration", "continuous control", "reinforcement learning" ], "rating_list": [ 5, 5, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "LIBLIlk5M9", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Felix Petersen", "gender": "Not Specified", "institution": "Stanford University", "country": "", "position": "Postdoc" }, { "name": "Frederick Vu", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Hao Sun", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Jiaqi Han", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" }, { "name": "Joshua Kazdan", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Stefano Ermon", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 40, "n_ref": 81, "n_ref_all": 104, "n_fig": 10, "n_tab": 5, "L_tab": 1285, "n_element_tab": 145, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2028, "n_element_tab_1": 201, "formula_len_all": 1111, "formula_len_all_1": 804, "len_all": 140632, "len_all_1": 65533, "len_abs": 1414, "len_title": 131, "len_sents": 42425, "len_sents_1": 30045, "n_sents": 320, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1422, "title": "CPSample: Classifier Protected Sampling for Guarding Training Data During Diffusion", "abs": "Diffusion models have a tendency to exactly replicate their training data, especially when trained on small datasets. Most prior work has sought to mitigate this problem by imposing differential privacy constraints or masking parts of the training data, resulting in a notable substantial decrease in image quality. We present CPSample, a method that modifies the sampling process to prevent training data replication while preserving image quality. CPSample utilizes a classifier that is trained to overfit on random binary labels attached to the training data. CPSample then uses classifier guidance to steer the generation process away from the set of points that can be classified with high certainty, a set that includes the training data. CPSample achieves FID scores of 4.97 and 2.97 on CIFAR-10 and CelebA-64, respectively, without producing exact replicates of the training data. Unlike prior methods intended to guard the training images, CPSample only requires training a classifier rather than retraining a diffusion model, which is computationally cheaper. Moreover, our technique provides diffusion models with greater robustness against membership inference attacks, wherein an adversary attempts to discern which images were in the model's training dataset. We show that CPSample behaves like a built-in rejection sampler, and we demonstrate its capabilities to prevent mode collapse in Stable Diffusion.", "keywords": [ "diffusion", "privacy", "data protection", "mode collapse" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "LHPWuckqgM", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Bao Hoang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Undergrad student" }, { "name": "Jiayu Zhou", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Associate Professor" }, { "name": "Shuyang Yu", "gender": "Female", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Yijiang Pang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 17, "n_ref_uni": 44, "n_ref": 101, "n_ref_all": 125, "n_fig": 0, "n_tab": 13, "L_tab": 7740, "n_element_tab": 977, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 1795, "n_element_tab_1": 289, "formula_len_all": 3541, "formula_len_all_1": 2202, "len_all": 146766, "len_all_1": 69435, "len_abs": 1062, "len_title": 99, "len_sents": 35475, "len_sents_1": 28228, "n_sents": 280, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1065, "title": "Towards Reliability of Parameter-free Optimization", "abs": "Hyperparameter tuning, particularly the selection of an appropriate learning rate in adaptive gradient training methods, remains a challenge. To tackle this challenge, in this paper, we propose a novel parameter-free optimizer, AdamG (Adam with the golden step size), designed to automatically adapt to diverse optimization problems without manual tuning. The core technique underlying AdamG is our golden step size derived for the AdaGrad-Norm algorithm, which is expected to help AdaGrad-Norm preserve the tuning-free convergence and approximate the optimal step size in expectation w.r.t. various optimization scenarios. To better evaluate tuning-free performance, we propose a novel evaluation criterion, reliability, to comprehensively assess the efficacy of parameter-free optimizers in addition to classical performance criteria. Empirical results demonstrate that compared with other parameter-free baselines, AdamG achieves superior performance, which is consistently on par with Adam using a manually tuned learning rate across various optimization tasks.", "keywords": [ "hyperparameter tuning", "tuning free" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "LGafQ1g2D2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Rose Yu", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Zihao Zhou", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 54, "n_ref_all": 83, "n_fig": 12, "n_tab": 4, "L_tab": 124, "n_element_tab": 32, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 985, "n_element_tab_1": 11, "formula_len_all": 1234, "formula_len_all_1": 704, "len_all": 182078, "len_all_1": 65566, "len_abs": 6136, "len_title": 90, "len_sents": 52109, "len_sents_1": 30855, "n_sents": 446, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 1424, "title": "Can LLMs Understand Time Series Anomalies?", "abs": "Large Language Models (LLMs) have gained popularity in time series forecasting, but their potential for anomaly detection remains largely unexplored. Our study investigates whether LLMs can understand and detect anomalies in time series data, focusing on zero-shot and few-shot scenarios. Inspired by conjectures about LLMs' behavior from time series forecasting research, we formulate key hypotheses about LLMs' capabilities in time series anomaly detection. We design and conduct principled experiments to test each of these hypotheses. Our investigation reveals several surprising findings about LLMs for time series: (1) LLMs understand time series better as *images* rather than as text, (2) LLMs did not demonstrate enhanced performance when prompted to engage in *explicit reasoning* about time series analysis. (3) Contrary to common beliefs, LLM's understanding of time series *do not* stem from their repetition biases or arithmetic abilities. (4) LLMs' behaviors and performance in time series analysis *vary significantly* across different model families. This study provides the first comprehensive analysis of contemporary LLM capabilities in time series anomaly detection. Our results suggest that while LLMs can understand trivial time series anomalies (we have no evidence that they can understand more subtle real-world anomalies), many common conjectures based on their reasoning capabilities do not hold.", "keywords": [ "Large Language Models (LLMs)", "Time Series Analysis", "Anomaly Detection", "Multimodal Learning" ], "rating_list": [ 5, 6, 6, 3, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "LGIhipNvCQ", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Fan Lyu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Postdoc" }, { "name": "Fuyuan Hu", "gender": "Male", "institution": "Suzhou University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Liang Wang", "gender": "Male", "institution": "Institute of Automation\uff0c CAS\uff0cChina", "country": "CN", "position": "Full Professor" }, { "name": "Zhang Zhang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Ziqi Shi", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Undergrad student" }, { "name": "hanyu zhao", "gender": "unknown", "institution": "University of Bristol", "country": "", "position": "MS student" } ], "n_formula": 30, "n_formula_1": 14, "n_ref_uni": 44, "n_ref": 91, "n_ref_all": 108, "n_fig": 2, "n_tab": 12, "L_tab": 6193, "n_element_tab": 462, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 5209, "n_element_tab_1": 205, "formula_len_all": 1783, "formula_len_all_1": 783, "len_all": 151353, "len_all_1": 70882, "len_abs": 1582, "len_title": 121, "len_sents": 45773, "len_sents_1": 29463, "n_sents": 355, "n_sents_1": 223, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1412, "title": "TUI: A Conformal Uncertainty Indicator for Continual Test-Time Adaptation", "abs": "Continual Test-Time Adaptation (CTTA) addresses the challenge of adapting models to sequentially changing domains during the testing phase. Since no ground truth labels are provided, existing CTTA methods rely on pseudo-labels for self-adaptation. However, CTTA is prone to error accumulation, where incorrect pseudo-labels can negatively impact subsequent model updates. Critically, during testing, a CTTA method can not detect its mistakes, which may then propagate through further adaptations. In this paper, we propose a simple uncertainty indicator called TUI for the CTTA task based on Conformal Prediction (CP), which generates a set of possible labels for each test sample, ensuring that the true label is included within this set with a given coverage probability. Specifically, since domain shifts can undermine the coverage of predictions, making uncertainty estimation less dependable, we propose compensating for coverage by dynamically measuring the domain difference between the target and source domains in continuously changing environments. Moreover, after estimating uncertainty, we separate reliable test pseudo-labels and use them to discriminatively enhance the adaptation process. Empirical results demonstrate that our algorithm effectively estimates the uncertainty for CTTA under a specified coverage probability and improves adaptation performance across various existing CTTA methods.", "keywords": [ "Continual Test-Time Adaptation", "Domain Adaptation", "Conformal Prediction" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "LFn7s8yRUF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Haoran Qi", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "MS student" }, { "name": "Zehua Wang", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 5, "n_ref": 13, "n_ref_all": 19, "n_fig": 3, "n_tab": 1, "L_tab": 154, "n_element_tab": 30, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2210, "n_element_tab_1": 140, "formula_len_all": 116, "formula_len_all_1": 116, "len_all": 67772, "len_all_1": 52714, "len_abs": 1299, "len_title": 148, "len_sents": 29932, "len_sents_1": 24272, "n_sents": 198, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1304, "title": "EXPLORING THE IMPACT OF DATA AUGMENTATION ON LOCALIZED PERSONALIZED AI TRAINING WITH LLAMA3 AND LORA", "abs": "With the development of personalized AI models, particularly those emulating characters from novels, games, anime, and films, a significant challenge is the scarcity of suitable dialogue data. These works often feature distinctive styles and character dialogues that may not generalize well to everyday conversations. Data augmentation is crucial for enriching these limited datasets, ensuring sufficient data for learning the target character\u2019s tone and linguistic habits. This paper investigates the impact of various data augmentation techniques on personalized AI models in NLP, specifically focusing on models trained using LLaMA3 through Low-Rank Adaptation (LoRA). We employ different data augmentation strategies, including random deletion, synonym replacement, swapping, random insertion, back translation, and paraphrasing. To provide a comprehensive analysis, we apply these techniques across three distinct datasets, each representing different dialogue styles and contexts. By systematically comparing these methods, we demonstrate their influence on model performance and robustness. This study provides valuable insights into the effectiveness of different data augmentation strategies for enhancing the versatility and robustness of personalized AI systems trained with LLaMA3 using LoRA.", "keywords": [ "Data Augmentation", "Personalized AI", "LLaMA3", "Low-Rank Adaptation", "NLP", "Synonym Replacement", "Random Insertion", "Random Swap", "Back Translation", "Paraphrasing", "Training Models", "Machine Learning", "Model Generalization" ], "rating_list": [ 1, 1, 1, 3 ], "soundness_list": [ 1, 1, 2, 1 ], "presentation_list": [ 1, 1, 1, 3 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "LFiaoYnP6T", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chaozheng Wang", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Min Zhang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Qifan Wang", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Principal Researcher" }, { "name": "Yu Pan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zekai Wu", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "", "position": "Undergrad student" }, { "name": "Zenglin Xu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 39, "n_formula_1": 11, "n_ref_uni": 39, "n_ref": 87, "n_ref_all": 137, "n_fig": 19, "n_tab": 8, "L_tab": 2176, "n_element_tab": 202, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 854, "n_element_tab_1": 61, "formula_len_all": 2627, "formula_len_all_1": 533, "len_all": 209696, "len_all_1": 75970, "len_abs": 1348, "len_title": 128, "len_sents": 62081, "len_sents_1": 29260, "n_sents": 758, "n_sents_1": 288, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1355, "title": "IDInit: A Universal and Stable Initialization Method for Neural Network Training", "abs": "Deep neural networks have achieved remarkable accomplishments in practice. The success of these networks hinges on effective initialization methods, which are vital for ensuring stable and rapid convergence during training. Recently, initialization methods that maintain identity transition within layers have shown good efficiency in network training. These techniques (e.g., Fixup) set specific weights to zero to achieve identity control. However, settings of remaining weight (e.g., Fixup uses random values to initialize non-zero weights) will affect inductive bias that is achieved only by a zero weight, which may be harmful to training. Addressing this concern, we introduce fully identical initialization (IDInit), a novel method that preserves identity in both the main and sub-stem layers of residual networks. IDInit employs a padded identity-like matrix to overcome rank constraints in non-square weight matrices. Furthermore, we show the convergence problem of an identity matrix can be solved by stochastic gradient descent. Additionally, we enhance the universality of IDInit by processing higher-order weights and addressing dead neuron problems. IDInit is a straightforward yet effective initialization method, with improved convergence, stability, and performance across various settings, including large-scale datasets and deep models.", "keywords": [ "Initialization", "Idetity Matrix", "Dynamic Isometry" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "LDu822E45Q", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 15, "author_info_list": [ { "name": "Amos J Storkey", "gender": "Not Specified", "institution": "University of Edinburgh", "country": "", "position": "Full Professor" }, { "name": "Eleni Triantafillou", "gender": "Female", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Elliot J. Crowley", "gender": "Male", "institution": "University of Edinburgh", "country": "GB", "position": "Associate Professor" }, { "name": "Fady Rezk", "gender": "unknown", "institution": "Edinburgh University, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Guang Yang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Hugo Larochelle", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "Adjunct Professor" }, { "name": "Jeff Z. Pan", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Full Professor" }, { "name": "Justin Engelmann", "gender": "Male", "institution": "University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Kiyoon Kim", "gender": "unknown", "institution": "Deargen", "country": "", "position": "Researcher" }, { "name": "Linus Ericsson", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Postdoc" }, { "name": "Pavlos Vougiouklis", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Sebastien Montella", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Researcher" }, { "name": "Srihari Humbarwadi", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "", "position": "MS student" }, { "name": "Yi Liu", "gender": "unknown", "institution": "University of Queensland", "country": "", "position": "MS student" }, { "name": "Antreas Antoniou", "gender": "Male", "institution": "University of Edinburgh", "country": "", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 40, "n_ref": 51, "n_ref_all": 71, "n_fig": 9, "n_tab": 7, "L_tab": 19013, "n_element_tab": 715, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 2171, "n_element_tab_1": 1, "formula_len_all": 96, "formula_len_all_1": 96, "len_all": 226514, "len_all_1": 64010, "len_abs": 2149, "len_title": 124, "len_sents": 59338, "len_sents_1": 34178, "n_sents": 334, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 2199, "title": "EEVEE and GATE: Finding the right benchmarks and how to run them seamlessly", "abs": "Model evaluation is a cornerstone of machine learning, guiding model design and progress measurement. Designing generalizable evaluation processes remains a challenge, however, partly due to the vast number of possible domain, task and modality combinations and lack of knowledge of how informative they are. In this paper, we propose EEVEE (Efficient Evaluation process Evolution Engine) - pronounced as \\textipa{/'i:vi:/} EE-vee - a method that frames evaluation process design as a learning problem. By analyzing a large number of evaluation metrics from diverse benchmarks and models, EEVEE identifies a smaller subset of tasks with high predictive power over the full set of evaluation metrics, reducing evaluation time. To find the optimal subset maximizing signal while minimizing GPU hours, EEVEE evaluates pre-trained models of various architectures, pretraining schemes, and modalities on diverse downstream tasks and datasets including image classification, segmentation, relational reasoning, zero-shot image-to-text tasks, medical classification and segmentation, video classification, and regression. Our results identify three subsets of benchmarks, with 8, 15 and 21 tasks, providing high quality signal for model generalization. Key benchmarks selected include iWildCam, CLEVR-Math, ACDC, WinoGround, CIFAR100, Fungi, and ADE20K. We structure the subsets into three tiers for 12, 24, and 36 GPU-hour budgets and package them into a unified, efficient, and user-friendly Python framework that we built with the researcher in mind -- which we refer to as the GATE engine. Our experiments reveal ConvNextV2, SigLIP and CLIP as top-performing model encoders, with EfficientNetV2 and ResNext50 excelling in medical tasks and challenging image classification, in particular in Happy Whale Individual classification, ConvNet based models seem to outperform transformer models by a factor of 2.5x, which is surprising. The top performing encoder being ConvNextV2 followed by CLIP seems to agree with other recent large scale evaluations. We also demonstrate the framework's versatility in fine-tuning models from text and audio modalities, paving the way for future cross-modal evaluations.", "keywords": [ "multi-modal", "benchmarks", "machine learning", "model evaluation", "benchmark frameworks" ], "rating_list": [ 5, 3, 8, 1 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "LDtNetvNQp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Dan Xu", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Shuling Zhao", "gender": "Female", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Xiaoshui Huang", "gender": "Not Specified", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Lecturer" }, { "name": "Fa-Ting Hong", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 36, "n_ref": 66, "n_ref_all": 82, "n_fig": 9, "n_tab": 1, "L_tab": 306, "n_element_tab": 49, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 3117, "n_element_tab_1": 243, "formula_len_all": 1207, "formula_len_all_1": 707, "len_all": 115739, "len_all_1": 72265, "len_abs": 1668, "len_title": 151, "len_sents": 35001, "len_sents_1": 29307, "n_sents": 243, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 1673, "title": "Synergizing Motion and Appearance: Multi-Scale Compensatory Codebooks for Talking Head Video Generation", "abs": "Talking head video generation aims to generate a realistic talking head video that\npreserves the person\u2019s identity from a source image and the motion from a driving\nvideo. Despite the promising progress made in the field, it remains a challenging\nand critical problem to generate videos with accurate poses and fine-grained facial\ndetails simultaneously. Essentially, facial motion is often highly complex to model\nprecisely, and the one-shot source face image cannot provide sufficient appearance\nguidance during generation due to dynamic pose changes. To tackle the problem, we propose to jointly learn motion and appearance codebooks and perform\nmulti-scale codebook compensation to effectively refine both the facial motion\nconditions and appearance features for talking face image decoding. Specifically,\nthe designed multi-scale motion and appearance codebooks are learned simultaneously in a unified framework to store representative global facial motion flow\nand appearance patterns. Then, we present a novel multi-scale motion and appearance compensation module, which utilizes a transformer-based codebook retrieval\nstrategy to query complementary information from the two codebooks for joint\nmotion and appearance compensation. The entire process produces motion flows\nof greater flexibility and appearance features with fewer distortions across different scales, resulting in a high-quality talking head video generation framework.\nExtensive experiments on various benchmarks validate the effectiveness of our\napproach and demonstrate superior generation results from both qualitative and\nquantitative perspectives when compared to state-of-the-art competitors.", "keywords": [ "talking head video generation", "motion transfer", "codebook compensation" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "LDmJfJlo83", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Anthony Rhodes", "gender": "Male", "institution": "Intel", "country": "US", "position": "Researcher" }, { "name": "Giuseppe Raffa", "gender": "Male", "institution": "Intel", "country": "US", "position": "Principal Researcher" }, { "name": "Lama Nachman", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Ramesh Manuvinakurike", "gender": "Male", "institution": "Intel", "country": "", "position": "Researcher" }, { "name": "Sovan Biswas", "gender": "Male", "institution": "Intel", "country": "", "position": "Research Scientist" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 26, "n_ref": 41, "n_ref_all": 50, "n_fig": 2, "n_tab": 6, "L_tab": 1562, "n_element_tab": 245, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1562, "n_element_tab_1": 245, "formula_len_all": 122, "formula_len_all_1": 122, "len_all": 90209, "len_all_1": 73744, "len_abs": 924, "len_title": 144, "len_sents": 23198, "len_sents_1": 22944, "n_sents": 140, "n_sents_1": 134, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1407, "title": "Uncertainty Quantification with Generative-Semantic Entropy Estimation for Large Language Models", "abs": "In recent years, powerful foundation models, including Large Language Models (LLMs) and Large Multi-Modal Models (LMMs) have ushered in a new epoch of multi-faceted, intelligent conversational agents. Despite their significant early successes and widespread use, foundation models nevertheless currently suffer from several critical challenges, including their lack of transparency and predilection for \"hallucinations.\" To this end, we introduce Generative-Semantic Entropy Estimation (GSEE), a model-agnostic algorithm that efficiently estimates the generative uncertainty associated with foundation models, while requiring no additional auxiliary model inference steps. In principle, for any foundation model input data, e.g., a text prompt, image, text + image, etc., GSEE numerically estimates the uncertainty encapsulated in the internal, semantic manifold of the LLM generated responses to the input data. In this way, high uncertainty is indicative of hallucinations and low generative confidence. Through experiments, we demonstrate the superior performance of GSEE for uncertainty quantification (UQ) amongst state-of-the-art methods across a variety of models, datasets, and problem settings, including: unbounded language prompting, constrained language prompting, high/low generative stochasticity, acute semantic diversity prompting, and as a barometer for hallucination/predictive accuracy.", "keywords": [ "Large Language Models", "Uncertainty Quantification", "Explainable AI", "Trustworthy AI" ], "rating_list": [ 1, 5, 5, 6 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "LDAj4UJ4aL", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Han Lin", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Koustuv Sinha", "gender": "Male", "institution": "Meta (FAIR)", "country": "US", "position": "Researcher" }, { "name": "Mohit Bansal", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" }, { "name": "Mojtaba Komeili", "gender": "Male", "institution": "Facebook AI Research", "country": "US", "position": "Researcher" }, { "name": "Nicolas Ballas", "gender": "unknown", "institution": "Meta", "country": "CA", "position": "Researcher" }, { "name": "Tushar Nagarajan", "gender": "unknown", "institution": "FAIR", "country": "US", "position": "Researcher" }, { "name": "Mahmoud Assran", "gender": "Male", "institution": "Meta", "country": "CA", "position": "Research Scientist" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 62, "n_ref": 182, "n_ref_all": 207, "n_fig": 7, "n_tab": 7, "L_tab": 1112, "n_element_tab": 193, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1727, "n_element_tab_1": 133, "formula_len_all": 223, "formula_len_all_1": 244, "len_all": 200101, "len_all_1": 69360, "len_abs": 2462, "len_title": 130, "len_sents": 49628, "len_sents_1": 30475, "n_sents": 353, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1755, "title": "VEDIT: Latent Prediction Architecture For Procedural Video Representation Learning", "abs": "Procedural video representation learning is an active research area where the objective is to learn an agent which can anticipate and forecast the future given the present video input, typically in conjunction with textual annotations. Prior works often rely on large-scale pretraining of visual encoders and prediction models with language supervision. However, the necessity and effectiveness of extending compute intensive pretraining to learn video clip sequences with noisy text supervision have not yet been fully validated by previous works. In this work, we show that a strong off-the-shelf frozen pretrained visual encoder, along with a well designed prediction model, can achieve state-of-the-art (SoTA) performance in forecasting and procedural planning without the need for pretraining the prediction model, nor requiring additional supervision from language or ASR. Instead of learning representations from pixel space, our method utilizes the latent embedding space of publicly available vision encoders. By conditioning on frozen clip-level embeddings from observed steps to predict the actions of unseen steps, our prediction model is able to learn robust representations for forecasting through iterative denoising \u2014leveraging the recent advances in diffusion transformers (Peebles & Xie, 2023). Empirical studies over a total of five procedural learning tasks across four datasets (NIV, CrossTask, COIN and Ego4D-v2) show that our model advances the strong baselines in long-horizon action anticipation (+2.6% in Verb ED@20, +3.1% in Noun ED@20), and significantly improves the SoTA in step forecasting (+5.0%), task classification (+3.8%), and procedure planning tasks (up to +2.28% in success rate, +3.39% in mAcc, and +0.90% in mIoU).", "keywords": [ "Procedural Learning from Videos", "Representation Learning", "Diffusion Transformer" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "LD0qz8j8Zm", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 6, "author_info_list": [ { "name": "Bach Viet Nguyen", "gender": "Male", "institution": "University of Texas at Dallas", "country": "", "position": "Undergrad student" }, { "name": "Duy Duong-Tran", "gender": "Male", "institution": "United States Naval Academy", "country": "", "position": "Assistant Professor" }, { "name": "Li Shen", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Full Professor" }, { "name": "Shu Yang", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "Postdoc" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Bo-Jian Hou", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Postdoc" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 29, "n_ref": 55, "n_ref_all": 75, "n_fig": 10, "n_tab": 2, "L_tab": 1322, "n_element_tab": 197, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 3182, "n_element_tab_1": 122, "formula_len_all": 327, "formula_len_all_1": 235, "len_all": 136937, "len_all_1": 53557, "len_abs": 4989, "len_title": 139, "len_sents": 40273, "len_sents_1": 24650, "n_sents": 236, "n_sents_1": 131, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 2069, "title": "Evaluating topological fitness of human brain-inspired sub-circuits in Echo State Networks", "abs": "Recent years have witnessed an emerging trend in neuromorphic computing that centers around the use of brain connectomics as a blueprint for artificial neural networks. Connectomics-based neuromorphic computing has primarily focused on embedding human brain large-scale structural connectomes (SCs), as estimated from diffusion Magnetic Resonance Imaging (dMRI) modality, to echo-state networks (ESNs). A critical step in ESN embedding requires pre-determined read-in and read-out layers constructed by the induced subgraphs (e.g., a priori set of functional sub-circuits/networks) of the embedded reservoir (e.g., SCs). As a priori set of functional sub-circuits are derived from functional MRI (fMRI) modality, it is unknown, till this point, whether the embedding of fMRI-induced sub-circuits/networks onto SCs is well justified from i) the neuro-physiological perspective and ii) ESN performance across a variety of tasks. In this paper, we proposed a pipeline to implement and evaluate ESNs with various embedded topology and processing/memorization tasks. To this end, we showed that different performance optimums are highly dependent on the neuro-physiological characteristics of these pre-determined fMRI-induced sub-circuits. In general, fMRI-induced sub-circuit-embedded ESN outperforms simple bipartite and various null models with feed-forward properties commonly seen in MLP for different tasks and reservoir criticality conditions. Noticeably, we found that the reservoir model performance is heavily dependent on the functional sub-circuits neuro-physiological properties with respect to different cognitive tasks and their corresponding computation-memorization balances. Specifically, we showed that default mode network's superior performance across the majority of tasks is related to its functional dichotomy property. Finally, we provided a thorough analysis of the topological properties of pre-determined fMRI-induced sub-circuits and highlighted their graph-theoretical properties that play significant roles in determining the ESN performance.", "keywords": [ "Computational Neuroscience", "Neural Data Analysis", "Neuromorphic Computing", "Recurrent Neural Networks", "Echo-state Networks", "Reservoir Computing", "Network Topology", "Bio-inspired Neural Networks" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 5, 4, 5, 2 ] }, { "paper_id": "LCrm1FSl26", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Bowei He", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Han Wu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Linqi Song", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Shuqi LIU", "gender": "Female", "institution": "City University of Hong Kong", "country": "", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 119, "n_ref_all": 153, "n_fig": 4, "n_tab": 21, "L_tab": 7673, "n_element_tab": 1218, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 3668, "n_element_tab_1": 695, "formula_len_all": 816, "formula_len_all_1": 393, "len_all": 177880, "len_all_1": 76873, "len_abs": 1442, "len_title": 121, "len_sents": 44780, "len_sents_1": 30579, "n_sents": 323, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1501, "title": "Towards Efficient Adaptation of Pruning Strategy in Large Language Models", "abs": "Post-training pruning has gained increasing attention with the rapid growth of large language models (LLMs). However, significant variations in weight distributions across different LLMs make a fixed pruning strategy inadequate for multiple models. In this paper, we propose an efficient evolutionary optimization framework, \\textbf{Mecon}, for adaptive LLM pruning. In particular, we design an effective search space built on our \\textbf{Me}ta pruning metric to mitigate diverse weight distributions among LLMs. We then introduce model-wise re\\textbf{con}struction error, a lightweight search evaluation to speed up the evaluation of each search trial. We finally leverage Non-dominated Sorting Genetic Algorithm III (NSGA-III) as our search algorithm, handling both the single-objective problem of pruning metric search and the multi-objective problem of layerwise sparsity ratio search in discovering the optimal pruning strategy. We extensively evaluate our framework on LLaMA-1/2/3 and Mistral models across multiple benchmarks. Our results demonstrate that our adaptive pruning metrics consistently outperform existing ones, and the layerwise sparsity ratios improve the effectiveness of other pruning metrics. Furthermore, we validate the cross-task and cross-model generalizability of our pruning metrics, offering a cost-effective solution to streamline the search process. We release our code in the anonymous repository: \\textcolor{blue}{\\url{https://anonymous.4open.science/r/Mecon-5819}}.", "keywords": [ "Model Pruning", "Large Language Model" ], "rating_list": [ 3, 3, 8, 8, 6 ], "soundness_list": [ 2, 2, 3, 4, 3 ], "presentation_list": [ 2, 2, 4, 4, 3 ], "contribution_list": [ 1, 2, 3, 3, 3 ], "confidence_list": [ 5, 5, 4, 3, 4 ] }, { "paper_id": "LCk3umTAXx", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Charles Elkan", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Professor" }, { "name": "Chirag Ahooja", "gender": "unknown", "institution": "Fraction AI", "country": "IN", "position": "Researcher" }, { "name": "Garvit Jain", "gender": "Male", "institution": "Fraction AI", "country": "VG", "position": "Researcher" }, { "name": "Rohan Tomar", "gender": "Male", "institution": "Fraction AI", "country": "IN", "position": "Researcher" }, { "name": "Shashank Yadav", "gender": "Male", "institution": "Fraction AI", "country": "US", "position": "Principal Researcher" }, { "name": "Shubham Chaudhary", "gender": "unknown", "institution": "Fraction AI", "country": "VG", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 4, "n_ref_uni": 51, "n_ref": 73, "n_ref_all": 93, "n_fig": 2, "n_tab": 4, "L_tab": 1375, "n_element_tab": 165, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1009, "formula_len_all_1": 544, "len_all": 161225, "len_all_1": 53075, "len_abs": 667, "len_title": 115, "len_sents": 36940, "len_sents_1": 25450, "n_sents": 311, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1292, "title": "Gamified crowd-sourcing of high-quality data for visual fine-tuning", "abs": "This paper introduces gamified adversarial prompting (GAP), a framework that\ncrowd-sources high-quality data for visual instruction tuning of large multimodal\nmodels. GAP transforms the data collection process into an engaging game, in-\ncentivizing players to provide fine-grained, challenging questions and answers\nthat target gaps in the model\u2019s knowledge. Our contributions include (1) an ap-\nproach to capture question-answer pairs from humans that directly address weak-\nnesses in a model\u2019s knowledge, (2) a method for evaluating and rewarding players\nthat successfully incentivizes them to provide high-quality submissions, and (3) a\nscalable, gamified platform that succeeds in collecting this data from over 50,000\nparticipants in just a few weeks. Our implementation of GAP has significantly im-\nproved the accuracy of a small multimodal model, namely MiniCPM-Llama3-V-\n2.5-8B, increasing its GPT score from 0.147 to 0.477 on our dataset, approaching\nthe benchmark set by the much larger GPT-4V. Moreover, we demonstrate that\nthe data generated using MiniCPM-Llama3-V-2.5-8B also enhances its perfor-\nmance across other benchmarks, and exhibits cross-model benefits. Specifically,\nthe same data improves the performance of QWEN2-VL-2B and QWEN2-VL-7B\non the same multiple benchmarks.", "keywords": [ "Large Multimodal Models", "Visual Question Answering", "Visual Instruction Tuning", "Gamification", "Supervised Learning", "Data Generation" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "LCL8SMGxDY", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Muthu Chidambaram", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Rong Ge", "gender": "Male", "institution": "Duke University", "country": "", "position": "Associate Professor" } ], "n_formula": 30, "n_formula_1": 13, "n_ref_uni": 56, "n_ref": 113, "n_ref_all": 138, "n_fig": 7, "n_tab": 13, "L_tab": 2013, "n_element_tab": 125, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 505, "n_element_tab_1": 5, "formula_len_all": 2087, "formula_len_all_1": 754, "len_all": 203256, "len_all_1": 69535, "len_abs": 1281, "len_title": 131, "len_sents": 63024, "len_sents_1": 34278, "n_sents": 439, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1288, "title": "For Better or For Worse? Learning Minimum Variance Features With Label Augmentation", "abs": "Data augmentation has been pivotal in successfully training deep learning models on classification tasks over the past decade. An important subclass of data augmentation techniques - which includes both label smoothing and Mixup - involves modifying not only the input data but also the input label during model training. In this work, we analyze the role played by the label augmentation aspect of such methods. We first prove that linear models on binary classification data trained with label augmentation learn only the minimum variance features in the data, while standard training (which includes weight decay) can learn higher variance features. We then use our techniques to show that even for nonlinear models and general data distributions, the label smoothing and Mixup losses are lower bounded by a function of the model output variance. Lastly, we demonstrate empirically that this aspect of label smoothing and Mixup can be a positive and a negative. On the one hand, we show that the strong performance of label smoothing and Mixup on image classification benchmarks is correlated with learning low variance hidden representations. On the other hand, we show that Mixup and label smoothing can be more susceptible to low variance spurious correlations in the training data.", "keywords": [ "feature learning", "mixup", "label smoothing", "spurious correlations" ], "rating_list": [ 5, 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 2, 3, 3, 5, 3 ] }, { "paper_id": "LC6ZtQV6u2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Huan Wang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xu Ma", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Yitian Zhang", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Yue Bai", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "PhD student" }, { "name": "Yun Fu", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 54, "n_ref": 161, "n_ref_all": 174, "n_fig": 7, "n_tab": 5, "L_tab": 3013, "n_element_tab": 295, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2390, "n_element_tab_1": 164, "formula_len_all": 299, "formula_len_all_1": 299, "len_all": 154601, "len_all_1": 55477, "len_abs": 1191, "len_title": 108, "len_sents": 41599, "len_sents_1": 25833, "n_sents": 281, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1256, "title": "Accessing Vision Foundation Models via ImageNet-1K", "abs": "Vision foundation models are renowned for the generalization ability due to massive training data. Nevertheless, they demand tremendous training resources, and the training data is often inaccessible, e.g., CLIP, DINOv2, posing great challenges to developing derivatives that could facilitate the research. In this work, we offer a very simple and general solution, named Proteus, to distill foundation models into smaller equivalents on ImageNet-1K without access to the original training data. Specifically, we remove the designs from conventional knowledge distillation settings that result in dataset bias and present three levels of training objectives, i.e., token, patch, and feature, to maximize the efficacy of knowledge transfer. In this manner, Proteus is trained at ImageNet-level costs with surprising ability, facilitating the accessibility of training foundation models for the broader research community. When leveraging DINOv2-g/14 as the teacher, Proteus-L/14 matches the performance of the Oracle method DINOv2-L/14 (142M training data) across 19 benchmarks and outperforms other vision foundation models including CLIP-L/14 (400M), OpenCLIP-L/14 (400M/2B) and SynCLR-L/14 (600M) with a significantly smaller training set of 1.2M images.", "keywords": [ "Foundation Model", "Model Compressing", "DINOv2", "CLIP", "SynCLR" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 5, 5 ] }, { "paper_id": "LC2KxRwC3n", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "David Chanin", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Hardik Bhatnagar", "gender": "unknown", "institution": "Centre for Integrative Neuroscience, AG Bethge", "country": "DE", "position": "PhD student" }, { "name": "James Wilken-Smith", "gender": "Male", "institution": "Independent", "country": "GB", "position": "Independent Researcher" }, { "name": "Joseph Isaac Bloom", "gender": "Male", "institution": "Department of Science, Technology and Innovation", "country": "GB", "position": "Researcher" }, { "name": "Tom\u00e1\u0161 Dulka", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 23, "n_ref": 53, "n_ref_all": 92, "n_fig": 29, "n_tab": 3, "L_tab": 268, "n_element_tab": 14, "n_fig_1": 14, "n_tab_1": 1, "L_tab_1": 155, "n_element_tab_1": 4, "formula_len_all": 143, "formula_len_all_1": 81, "len_all": 171929, "len_all_1": 60568, "len_abs": 1056, "len_title": 134, "len_sents": 49829, "len_sents_1": 28763, "n_sents": 371, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 975, "title": "A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders", "abs": "Sparse Autoencoders (SAEs) have emerged as a promising approach to decompose the activations of Large Language Models (LLMs) into human-interpretable latents. In this paper, we pose two questions. First, to what extent do SAEs extract monosemantic and interpretable latents? Second, to what extent does varying the sparsity or the size of the SAE affect monosemanticity / interpretability? By investigating these questions in the context of a simple first-letter identification task where we have complete access to ground truth labels for all tokens in the vocabulary, we are able to provide more detail than prior investigations. Critically, we identify a problematic form of feature-splitting we call \"feature absorption\" where seemingly monosemantic latents fail to fire in cases where they clearly should. Our investigation suggests that varying SAE size or sparsity is insufficient to solve this issue, and that there are deeper conceptual issues in need of resolution.", "keywords": [ "Sparse Autoencoders", "SAEs", "LLMs", "interpretability" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "LBsr2llHz0", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Bohang Zhang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Lin Yao", "gender": "Female", "institution": "DP Technology", "country": "CN", "position": "Researcher" }, { "name": "Linfeng Zhang", "gender": "Male", "institution": "DP Technology", "country": "", "position": "Researcher" }, { "name": "Shuqi Lu", "gender": "Female", "institution": "DP Technology", "country": "CN", "position": "Researcher" }, { "name": "Siyuan Liu", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Zhifeng Gao", "gender": "Male", "institution": "DP Technology", "country": "", "position": "Researcher" }, { "name": "guolin ke", "gender": "Male", "institution": "DP Technology", "country": "CN", "position": "Senior Researcher" }, { "name": "ji xh", "gender": "Unspecified", "institution": "DPtechnology", "country": "CN", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 18, "n_ref_uni": 45, "n_ref": 101, "n_ref_all": 120, "n_fig": 1, "n_tab": 10, "L_tab": 4301, "n_element_tab": 565, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 3077, "n_element_tab_1": 231, "formula_len_all": 1378, "formula_len_all_1": 1298, "len_all": 151955, "len_all_1": 63862, "len_abs": 1305, "len_title": 120, "len_sents": 37473, "len_sents_1": 26362, "n_sents": 283, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1308, "title": "More Space Is All You Need: Revisiting Molecular Representation Learning", "abs": "Molecular representation learning (MRL) has become pivotal in leveraging limited supervised data for applications such as drug discovery and material design. While early MRL methods relied on 1D sequences and 2D graphs, recent advancements have incorporated 3D conformational information, focusing predominantly on atomic interactions within 3D space. However, we argue that the space beyond atoms is also crucial for MRL, which is overlooked by prior models. To address this, we propose a novel transformer-based framework, dubbed SpaceFormer, which incorporates additional 3D space beyond atoms to enhance molecular representation ability. \nSpaceFormer introduces three key components: (1) Precision-Preserved Gridding, which discretizes continuous 3D space into grid cells while preserving precision; (2) Grid Sampling, which employs an importance sampling strategy to improve efficiency; and (3) Linear-Complexity 3D Positional Encoding, which extends Rotary Positional Encoding to 3D space to capture pairwise directions and utilizes random Fourier features to efficiently encode pairwise distances. Extensive experiments show that SpaceFormer significantly outperforms previous 3D MRL models across various tasks, validating the benefit of leveraging the additional 3D space beyond atoms in MRL models.", "keywords": [ "Molecular Representation Learning", "Molecular Property" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "LBl7Hez0fF", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Haotian Ye", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "James Zou", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Sheng Liu", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Postdoc" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 50, "n_ref": 117, "n_ref_all": 140, "n_fig": 9, "n_tab": 4, "L_tab": 1165, "n_element_tab": 201, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1581, "n_element_tab_1": 163, "formula_len_all": 243, "formula_len_all_1": 243, "len_all": 151507, "len_all_1": 57402, "len_abs": 1121, "len_title": 129, "len_sents": 34551, "len_sents_1": 26789, "n_sents": 235, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 81, "L_abs": 1125, "title": "Reducing Hallucinations in Large Vision-Language Models via Latent Space Steering", "abs": "Hallucination poses a challenge to the deployment of large vision-language models (LVLMs) in applications. Unlike in large language models (LLMs), hallucination in LVLMs often arises from misalignments between visual inputs and textual outputs. This paper investigates the underlying mechanisms of hallucination, focusing on the unique structure of LVLMs that distinguishes them from LLMs. We identify that hallucinations often arise from the sensitivity of text decoders to vision inputs, a natural phenomenon when image encoders and text decoders are pre-trained separately. Inspired by this, we introduce Visual and Textual Intervention (VTI), a novel technique designed to reduce hallucinations by steering latent space representations during inference to enhance the stability of vision features. As a task-agnostic test-time intervention, VTI can be easily applied to any problem without additional training cost. Extensive experiments demonstrate that it can effectively reduce hallucinations and outperform baseline methods across multiple metrics, highlighting the critical role of vision feature stability in LVLMs.", "keywords": [ "Large Vision-Language Models", "Multimodal large language model", "Hallucination" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 4, 4, 3 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "LBd87fWerd", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Mahmood Sharif", "gender": "Male", "institution": "Tel Aviv University", "country": "", "position": "Assistant Professor" }, { "name": "Matan Ben-Tov", "gender": "unknown", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 7, "n_ref_uni": 54, "n_ref": 140, "n_ref_all": 236, "n_fig": 33, "n_tab": 15, "L_tab": 6690, "n_element_tab": 335, "n_fig_1": 14, "n_tab_1": 3, "L_tab_1": 1127, "n_element_tab_1": 44, "formula_len_all": 1427, "formula_len_all_1": 457, "len_all": 293407, "len_all_1": 68969, "len_abs": 1704, "len_title": 132, "len_sents": 107377, "len_sents_1": 31927, "n_sents": 847, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1771, "title": "GASLITEing the Retrieval: Poisoning Knowledge DBs to Mislead Embedding-based Search", "abs": "Embedding-based text retrieval\u2014retrieval of relevant passages from knowledge databases (KDBs) via deep learning encodings\u2014has emerged as a powerful method attaining state-of-the-art search results and popularizing the use of Retrieval Augmented Generation (RAG). Still, like other search methods, embedding-based retrieval may be susceptible to search-engine optimization (SEO) attacks, where adversaries promote malicious content by introducing adversarial passages to KDBs. To faithfully assess the susceptibility of such systems to SEO, this work proposes the _GASLITE_ attack, a mathematically principled gradient-based search method for generating adversarial passages without relying on the KDB content or modifying the model. Notably, _GASLITE_'s passages _(1)_ carry adversary-chosen information while _(2)_ achieving high retrieval ranking for a selected query distribution when inserted to KDBs. We extensively evaluated _GASLITE_, testing it on nine advanced models and comparing it to three baselines under varied threat models, focusing on one well-suited for realistic adversaries targeting queries on a specific concept (e.g., a public figure). We found _GASLITE_ consistently outperformed baselines by $\\ge$140\\% success rate, in all settings. Particularly, adversaries using _GASLITE_ require minimal effort to manipulate search results\u2014by injecting a negligible amount of adversarial passages ($\\le$0.0001\\% of the KDBs), they could make them visible in the top-10 results for 61-100\\% of unseen concept-specific queries against most evaluated models. Among other contributions, our work also identifies several factors that may influence model susceptibility to SEO, including the embedding space's geometry. We will make our code publicly available.", "keywords": [ "Adversarial Machine Learning", "AI Safety", "Security", "NLP", "Retrieval", "Text Representations", "Text Embeddings" ], "rating_list": [ 5, 5, 6, 8 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "LB5cKhgOTu", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Can Xiao", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Cheng Zhang", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "George Anthony Constantinides", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Full Professor" }, { "name": "Jeffrey T. H. Wong", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Yiren Zhao", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Assistant Professor" } ], "n_formula": 53, "n_formula_1": 31, "n_ref_uni": 34, "n_ref": 85, "n_ref_all": 132, "n_fig": 12, "n_tab": 23, "L_tab": 8612, "n_element_tab": 1155, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 823, "n_element_tab_1": 85, "formula_len_all": 2791, "formula_len_all_1": 1429, "len_all": 199218, "len_all_1": 61355, "len_abs": 1697, "len_title": 116, "len_sents": 54785, "len_sents_1": 26190, "n_sents": 457, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1672, "title": "QERA: an Analytical Framework for Quantization Error Reconstruction", "abs": "The growing number of parameters and computational demands of large language models (LLMs) present significant challenges for their efficient deployment.\nRecently, there is an increasing interest in quantizing weights to extremely low precision while offsetting the resulting error with low-rank, high-precision error reconstruction terms.\nThe combination of quantization and low-rank approximation is now popular in both adapter-based, parameter-efficient fine-tuning methods such as LoftQ and low-precision inference techniques including ZeroQuant-V2.\nUsually, the low-rank terms are calculated via the singular value decomposition (SVD) of the weight quantization error,\nminimizing the Frobenius and spectral norms of the weight approximation error.\nRecent methods like LQ-LoRA and LQER introduced hand-crafted heuristics to minimize errors in layer outputs (activations) rather than weights, resulting improved quantization results.\nHowever, these heuristic methods lack an analytical solution to guide the design of quantization error reconstruction terms.\nIn this paper, we revisit this problem and formulate an analytical framework, named Quantization Error Reconstruction Analysis (QERA),\nand offer a closed-form solution to the problem.\nWe show QERA benefits both existing low-precision fine-tuning and inference methods --\nQERA achieves a fine-tuned accuracy gain of $\\Delta_{\\text{acc}}$ = 6.05\\% of 2-bit RoBERTa-base on GLUE compared to LoftQ;\nand obtains $\\Delta_{\\text{acc}}$ = 2.97\\% higher post-training quantization accuracy of 4-bit Llama-3.1-70B on average than ZeroQuant-V2 and $\\Delta_{\\text{ppl}}$ = $-$ 0.28 lower perplexity on WikiText2 than LQER.", "keywords": [ "parameter-efficient fine-tuning", "PEFT", "LoRA", "QLoRA", "quantization", "post-training quantization", "PTQ", "low-rank approximation" ], "rating_list": [ 6, 8, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "LAsMFAg4Zf", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Baturalp Buyukates", "gender": "Male", "institution": "University of Birmingham", "country": "GB", "position": "Assistant Professor" }, { "name": "Chaoyang He", "gender": "Male", "institution": "FedML, Inc.", "country": "", "position": "CTO" }, { "name": "Qifan Zhang", "gender": "Male", "institution": "University of California, Irvine", "country": "", "position": "PhD student" }, { "name": "Amir Avestimehr", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "Full Professor" }, { "name": "Shanshan Han", "gender": "Female", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Weizhao Jin", "gender": "unknown", "institution": "AWS", "country": "US", "position": "Applied Scientist" }, { "name": "Wenxuan Wu", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "", "position": "PhD student" }, { "name": "Yuhang Yao", "gender": "unknown", "institution": "TensorOpera Inc", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 51, "n_ref": 139, "n_ref_all": 177, "n_fig": 23, "n_tab": 5, "L_tab": 1570, "n_element_tab": 72, "n_fig_1": 18, "n_tab_1": 3, "L_tab_1": 387, "n_element_tab_1": 16, "formula_len_all": 1093, "formula_len_all_1": 1029, "len_all": 154250, "len_all_1": 71289, "len_abs": 1097, "len_title": 169, "len_sents": 47198, "len_sents_1": 34473, "n_sents": 324, "n_sents_1": 235, "venue": "ICLR2025", "decision": "withdraw", "L_title": 121, "L_abs": 1101, "title": "Kick Bad Guys Out! Conditionally Activated Anomaly Detection in Federated Learning with Zero-Knowledge Proof Verification", "abs": "Federated Learning (FL) systems are susceptible to adversarial attacks, where malicious clients submit poisoned models to disrupt the convergence or plant backdoors that cause the global model to misclassify some samples. Current defense methods are often impractical for real-world FL systems, as they either rely on unrealistic prior knowledge or cause accuracy loss even in the absence of attacks. Furthermore, these methods lack a protocol for verifying execution, leaving participants uncertain about the correct execution of the mechanism. To address these challenges, we propose a novel anomaly detection strategy that is designed for real-world FL systems. Our approach activates the defense only when potential attacks are detected, and enables the removal of malicious models without affecting the benign ones. Additionally, we incorporate zero-knowledge proofs to ensure the integrity of the proposed defense mechanism. Experimental results demonstrate the effectiveness of our approach in enhancing FL system security against a comprehensive set of adversarial attacks in various ML tasks.", "keywords": [ "outlier detection", "FL security", "defense", "anomaly detection", "zero-knowledge-proof" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "L9pTokEb8L", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Aaron Podolny", "gender": "Male", "institution": "Scribe", "country": "US", "position": "Researcher" }, { "name": "Ameet Talwalkar", "gender": "Male", "institution": "Carnegie-Mellon University", "country": "", "position": "" }, { "name": "Atishay Jain", "gender": "Male", "institution": "Scribe", "country": "US", "position": "Researcher" }, { "name": "Ishan Amlekar", "gender": "Male", "institution": "Scribe", "country": "US", "position": "Founding Machine Learning Engineer " }, { "name": "Junhong Shen", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Mouad Hadji", "gender": "unknown", "institution": "Scribe", "country": "US", "position": "Researcher" }, { "name": "Zedian Xiao", "gender": "unknown", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 32, "n_ref": 70, "n_ref_all": 85, "n_fig": 2, "n_tab": 11, "L_tab": 3228, "n_element_tab": 478, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1444, "n_element_tab_1": 227, "formula_len_all": 0, "formula_len_all_1": 10, "len_all": 179087, "len_all_1": 74135, "len_abs": 1056, "len_title": 115, "len_sents": 55342, "len_sents_1": 35960, "n_sents": 410, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1061, "title": "Towards Specialized Web Agents Using Production-Scale Workflow Data", "abs": "Large Language Model (LLM) agents are rapidly improving to handle increasingly complex web-based tasks. Most of these agents rely on general-purpose, proprietary models like GPT-4 and focus on designing better prompts to improve their planning abilities. However, general-purpose LLMs are not specifically trained to understand specialized web contexts such as HTML, and they often struggle with long-horizon planning. We explore an alternative approach that fine-tunes open-source LLMs using production-scale workflow data collected from over 250 domains corresponding to 6 billion tokens. This simple yet effective approach shows substantial gains over prompting-based agents on existing benchmarks---our WorkflowAgent achieves state-of-the-art performance on Mind2Web and substantially improves the baseline task success rate from 37.2% to 51.3% on WebArena. We further perform detailed ablation studies on various fine-tuning design choices and provide insights into LLM selection, training recipes, context window optimization, and effect of dataset sizes.", "keywords": [ "LLM web agent" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "L9j8exYGUJ", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Amir Feder", "gender": "unknown", "institution": "Columbia University", "country": "", "position": "Postdoc" }, { "name": "Ariel Goldstein", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "Assistant Professor" }, { "name": "Yuval Shalev", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 30, "n_ref": 55, "n_ref_all": 84, "n_fig": 15, "n_tab": 2, "L_tab": 1281, "n_element_tab": 204, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 234, "formula_len_all_1": 113, "len_all": 195039, "len_all_1": 58029, "len_abs": 1290, "len_title": 133, "len_sents": 48598, "len_sents_1": 27355, "n_sents": 356, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1310, "title": "Distributional reasoning in LLMs: Parallel Reasoning Processes in Multi-hop Reasoning", "abs": "Large language models (LLMs) have shown an impressive ability to perform tasks believed to require \"thought processes\u201d. When the model does not document an explicit thought process, it's difficult to understand the processes occurring within its hidden layers, and to determine if this process can be referred to as reasoning. We introduce a novel and interpretable analysis of internal multi-hop reasoning processes in LLMs. We demonstrate that the prediction process for compositional reasoning questions can be modeled using a simple linear transformation between two semantic category spaces. We show that during inference, the middle layers of the network generate highly interpretable embeddings that represent a set of potential intermediate answers for the multi-hop question. We use statistical analyses to show that a corresponding subset of tokens is activated in the model's output, implying the existence of parallel reasoning paths. These observations hold true even when the model lacks the necessary knowledge to solve the task. Our findings can help uncover the strategies that LLMs use to solve reasoning tasks, offering insights into the types of thought processes that can emerge from artificial intelligence. Finally, we also discuss the implication of cognitive modeling of these results.", "keywords": [ "LLM", "Multihop reasoning", "Interpretability" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "L9eEfwwUwU", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Arindam Banerjee", "gender": "unknown", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Professor" }, { "name": "Qiaobo Li", "gender": "Male", "institution": "Department of Computer Science, University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Zhijie Chen", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 88, "n_formula_1": 25, "n_ref_uni": 44, "n_ref": 99, "n_ref_all": 116, "n_fig": 11, "n_tab": 10, "L_tab": 2285, "n_element_tab": 337, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 610, "n_element_tab_1": 94, "formula_len_all": 13965, "formula_len_all_1": 3095, "len_all": 203114, "len_all_1": 69090, "len_abs": 1959, "len_title": 119, "len_sents": 57794, "len_sents_1": 29676, "n_sents": 577, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1828, "title": "Sketched Adaptive Federated Deep Learning: A Sharp Convergence Analysis", "abs": "Combining gradient sketching methods (e.g., CountSketch, quantization) and adaptive optimizers (e.g., Adam, AMSGrad) is a desirable goal in federated learning (FL), with potential benefits on both fewer communication rounds and smaller per-round communication. In spite of the preliminary empirical success of sketched adaptive methods, existing convergence analyses show the communication cost to have a linear dependence on the ambient dimension, i.e., number of parameters, which is prohibitively high for modern deep learning models.\n\nIn this work, we introduce specific sketched adaptive federated learning (SAFL) algorithms and, as our main contribution, provide theoretical convergence analyses in different FL settings with guarantees on communication cost depending only logarithmically (instead of linearly) on the ambient dimension. Unlike existing analyses, we show that the entry-wise sketching noise existent in the preconditioners and the first moments of SAFL can be implicitly addressed by leveraging the recently-popularized anisotropic curvatures in deep learning losses, e.g., fast decaying loss Hessian eigen-values. \nIn the i.i.d. client setting of FL, we show that SAFL achieves $O(1/\\sqrt{T})$ convergence, and $O(1/T)$ convergence near initialization. In the non-i.i.d. client setting, where non-adaptive methods lack convergence guarantees, we show that SACFL (SAFL with clipping) algorithms can provably converge in spite of the additional heavy-tailed noise. Our theoretical claims are supported by empirical studies on vision and language tasks, and in both fine-tuning and training-from-scratch regimes. Surprisingly, as a by-product of our analysis, the proposed SAFL methods are competitive with the state-of-the-art communication-efficient federated learning algorithms based on error feedback.", "keywords": [ "Federated Learning", "Sketching Algorithm", "Deep Learning Optimization" ], "rating_list": [ 6, 3, 6, 3 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "L9eBxTCpQG", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Gaojie Jin", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "Assistant Professor" }, { "name": "Lu Liu", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "Full Professor" }, { "name": "Shiwei Liu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Tianjin Huang", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "Lecturer" }, { "name": "Zhangyang Wang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Ziquan Zhu", "gender": "unknown", "institution": "University of Leicester", "country": "GB", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 47, "n_ref": 96, "n_ref_all": 138, "n_fig": 14, "n_tab": 7, "L_tab": 1851, "n_element_tab": 290, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 1718, "n_element_tab_1": 143, "formula_len_all": 401, "formula_len_all_1": 382, "len_all": 154703, "len_all_1": 65397, "len_abs": 1618, "len_title": 114, "len_sents": 42918, "len_sents_1": 29020, "n_sents": 335, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1626, "title": "SPAM: Spike-Aware Adam with Momentum Reset for Stable LLM Training", "abs": "Large Language Models (LLMs) have demonstrated exceptional performance across diverse tasks, yet their training remains highly resource intensive and susceptible to critical challenges such as training instability. A predominant source of this instability stems from gradient and loss spikes, which disrupt the learning process, often leading to costly interventions like checkpoint recovery and experiment restarts, further amplifying inefficiencies. This paper presents a comprehensive investigation into gradient spikes observed during LLM training, revealing their prevalence across multiple architectures and datasets. Our analysis shows that these spikes can be up to 1000\u00d7 larger than typical gradients, substantially deteriorating model performance. To address this issue, we propose Spike-Aware Adam with Momentum Reset (SPAM), a novel optimizer designed to counteract gradient spikes through momentum reset and spike-aware gradient clipping. Extensive experiments, including both pre-training and fine-tuning, demonstrate that SPAM consistently surpasses Adam and its variants across a range of model scales. Additionally, SPAM facilitates memory-efficient training by enabling sparse momentum, where only a subset of momentum terms are maintained and updated. When operating under memory constraints, SPAM outperforms state-of-the-art memory-efficient optimizers such as GaLore and Adam-Mini. Our work underscores the importance\nof mitigating gradient spikes in LLM training and introduces an effective optimization strategy that enhances both training stability and resource efficiency at scale. Code is submitted.", "keywords": [ "Gradient Spikes", "Spike-Aware Adam", "LLMs" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "L8vZXTVxfG", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Jianfang Li", "gender": "Male", "institution": "Institute of High Energy Physics, Chinese Academy of Sciences", "country": "CN", "position": "Postdoc" }, { "name": "ShiPing Dong", "gender": "Male", "institution": "Hunan University", "country": "", "position": "PhD student" }, { "name": "Shizhe Zhou", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Associate Professor" }, { "name": "Tao Lu", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 53, "n_ref": 107, "n_ref_all": 132, "n_fig": 9, "n_tab": 5, "L_tab": 1543, "n_element_tab": 102, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 861, "n_element_tab_1": 49, "formula_len_all": 405, "formula_len_all_1": 402, "len_all": 134183, "len_all_1": 64136, "len_abs": 1590, "len_title": 130, "len_sents": 33029, "len_sents_1": 29667, "n_sents": 252, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1595, "title": "Towards Fine-tuning-free Few-shot Classification: A Purely Self-supervised Manner", "abs": "One of the core problems of supervised few-shot classification is adapting generalized knowledge learned from substantial labeled source data to rarely labeled novel target data. What makes it a challenging problem is how to eliminate undesirable inductive bias introduced by labels when learning generalized knowledge during pre-training or adapting the learned knowledge during fine-tuning.In this paper, we propose a purely self-supervised method to bypass the labeling dilemma, focusing on an extreme scenario where a few-shot feature extractor is learned without fine-tuning. Our approach is built on two key observations from recent advancements in style transfer learning and self-supervised learning:1) high-order statistics of feature maps in deep nets encapsulate distinct information about input samples, and 2) high-quality inputs are not essential for obtaining high-quality representations. Accordingly, we introduce a variant of the vector quantized variational autoencoder (VQ-VAE) that incorporates a novel coloring operation, which conveys statistical information from the encoder to the decoder, modulating the generation process with these distinct statistics. With this design, we find that the statistics derived from the encoder's feature maps possess strong discriminative power, enabling effective classification using simple Euclidean distance metrics. Through extensive experiments on standard few-shot classification benchmark. We show that our fine-tuning-free method achieves competitive performance compared to fine-tuning-based and meta-learning-based approaches.", "keywords": [ "few-shot learning", "variational autoencoder", "self-supervised learning" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "L8e7tBf4pP", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Changdae Oh", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Dongyoon Han", "gender": "Male", "institution": "NAVER", "country": "KR", "position": "Senior Research Scientist" }, { "name": "Kyungwoo Song", "gender": "unknown", "institution": "Yonsei University", "country": "KR", "position": "Assistant Professor" }, { "name": "Sangdoo Yun", "gender": "Male", "institution": "NAVER", "country": "KR", "position": "Research Scientist" }, { "name": "Yixuan Li", "gender": "Female", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 4, "n_ref_uni": 63, "n_ref": 199, "n_ref_all": 243, "n_fig": 13, "n_tab": 7, "L_tab": 1937, "n_element_tab": 251, "n_fig_1": 9, "n_tab_1": 7, "L_tab_1": 2445, "n_element_tab_1": 274, "formula_len_all": 1175, "formula_len_all_1": 222, "len_all": 211703, "len_all_1": 75770, "len_abs": 1274, "len_title": 120, "len_sents": 60842, "len_sents_1": 33605, "n_sents": 381, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1257, "title": "DaWin: Training-free Dynamic Weight Interpolation for Robust Adaptation", "abs": "Adapting a pre-trained foundation model on downstream tasks should ensure robustness against distribution shifts without the need to retrain the whole model. Although existing weight interpolation methods are simple yet effective, we argue their static nature limits downstream performance while achieving efficiency. In this work, we propose DaWin, a training-free dynamic weight interpolation method that leverages the entropy of individual models over each unlabeled test sample to assess model expertise, and compute per-sample interpolation coefficients dynamically. Unlike previous works that typically rely on additional training to learn such coefficients, our approach requires no training. Then, we propose a mixture modeling approach that greatly reduces inference overhead raised by dynamic interpolation. We validate DaWin on the large-scale visual recognition benchmarks, spanning 14 tasks across robust fine-tuning -- ImageNet and derived five distribution shift benchmarks -- and multi-task learning with eight classification tasks. Results demonstrate that DaWin achieves significant performance gain in considered settings, with minimal computational overhead. We further discuss DaWin's analytic behavior to explain its empirical success.", "keywords": [ "robustness", "distribution shift", "robust fine-tuning", "weight interpolation", "model merging" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "L7gyAKWpiM", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Chao Chen", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Associate Professor" }, { "name": "Jiachen Yao", "gender": "Male", "institution": ", State University of New York at Stony Brook", "country": "", "position": "PhD student" }, { "name": "Mayank Goswami", "gender": "Male", "institution": "CUNY Queens College", "country": "", "position": "Associate Professor" } ], "n_formula": 61, "n_formula_1": 21, "n_ref_uni": 34, "n_ref": 55, "n_ref_all": 61, "n_fig": 4, "n_tab": 2, "L_tab": 153, "n_element_tab": 11, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 930, "n_element_tab_1": 12, "formula_len_all": 5476, "formula_len_all_1": 1435, "len_all": 165076, "len_all_1": 71255, "len_abs": 703, "len_title": 76, "len_sents": 54347, "len_sents_1": 30831, "n_sents": 552, "n_sents_1": 272, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 705, "title": "A Theoretical Study of Neural Network Expressive Power via Manifold Topology", "abs": "A prevalent assumption regarding real-world data is that it lies on or close to a low-dimensional manifold. When deploying a neural network on data manifolds, the required size, i.e., the number of neurons of the network, heavily depends on the intricacy of the underlying latent manifold. While significant advancements have been made in understanding the geometric attributes of manifolds, it's essential to recognize that topology, too, is a fundamental characteristic of manifolds. In this study, we investigate network expressive power in terms of the latent data manifold. Integrating both topological and geometric facets of the data manifold, we present a size upper bound of ReLU neural networks.", "keywords": [ "Topology", "Manifold", "Homology" ], "rating_list": [ 6, 6, 5, 6, 6 ], "soundness_list": [ 3, 4, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 4, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 2, 3, 3, 3, 3 ] }, { "paper_id": "L6xxFCafqY", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Eric Xia", "gender": "unknown", "institution": "Brown University", "country": "US", "position": "Undergrad student" }, { "name": "Jugal Kalita", "gender": "Male", "institution": "University of Colorado at Colorado Springs", "country": "US", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 29, "n_ref": 55, "n_ref_all": 69, "n_fig": 5, "n_tab": 3, "L_tab": 1018, "n_element_tab": 109, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 479, "formula_len_all_1": 395, "len_all": 108153, "len_all_1": 55414, "len_abs": 853, "len_title": 121, "len_sents": 33879, "len_sents_1": 25933, "n_sents": 253, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 884, "title": "Linear Relational Decoding of Morphological Relations in Language Models", "abs": "The recent success of transformer language models owes much to their conversational fluency and productivity in linguistic and morphological aspects. An affine Taylor approximation has been found to be a good approximation for transformer computations over certain factual and encyclopedic relations. We show that the truly linear approximation $W\\textbf{s}$, where $\\textbf{s}$ is a middle layer representation of the base form and $W$ is a local model derivative, is necessary and sufficient to approximate morphological derivations. This approach achieves above 80\\% faithfulness across most morphological tasks in the Bigger Analogy Test Set, and is successful across language models and typological categories. We propose that morphological relationships in transformer models are likely to be linearly encoded, with implications for how entities are represented in latent space.", "keywords": [ "LMs", "Model Interpretability", "Representation Theory", "Morphology", "Linear Representation Hypothesis", "Analogical Reasoning" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 2 ] }, { "paper_id": "L6gyOOJYt2", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Chuan Zhou", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Chunyuan Zheng", "gender": "Male", "institution": "meituan", "country": "CN", "position": "Intern" }, { "name": "Hao Wang", "gender": "Male", "institution": "ByteDance", "country": "CN", "position": "Researcher" }, { "name": "Haoxuan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Mingming Gong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Xu Chen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 67, "n_formula_1": 19, "n_ref_uni": 37, "n_ref": 91, "n_ref_all": 103, "n_fig": 5, "n_tab": 4, "L_tab": 4209, "n_element_tab": 198, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2182, "n_element_tab_1": 118, "formula_len_all": 7153, "formula_len_all_1": 1973, "len_all": 136212, "len_all_1": 61761, "len_abs": 1274, "len_title": 115, "len_sents": 37577, "len_sents_1": 26968, "n_sents": 324, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1278, "title": "Sampling Process Brings Additional Bias for Debiased Recommendation", "abs": "In recommender systems, selection bias arises from the users' selective interactions with items, which poses a widely-recognized challenge for unbiased evaluation and learning for recommendation models. Recently, doubly robust and its variants have been widely studied to achieve debiased learning of prediction models. However, if the users and items in the training set are not exactly the same as those in the test set, even if the imputed errors and learned propensities are accurate, all previous doubly robust based debiasing methods are biased. To tackle this problem, in this paper, we first derive the bias of doubly robust learning methods and provide alternative unbiasedness conditions when users and items are sampled from a superpopulation. Then we propose a novel superpopulation doubly robust target learning approach (SuperDR), which is unbiased when either the imputation model or propensity model is correctly specified. We further derive the generalization error bound of the proposed method under superpopulation, and show that it can be effectively controlled by the proposed target learning approach. We conduct extensive experiments on three real-world datasets, including a large-scale industrial dataset, to demonstrate the effectiveness of our method.", "keywords": [ "Superpopulation", "Selection Bias", "Recommender System" ], "rating_list": [ 5, 3, 8, 3, 3 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 4, 3 ] }, { "paper_id": "L6IgkJvcgV", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Gautam Sreekumar", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Sepehr Dehdashtian", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Vishnu Naresh Boddeti", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 21, "n_formula_1": 16, "n_ref_uni": 44, "n_ref": 70, "n_ref_all": 103, "n_fig": 15, "n_tab": 5, "L_tab": 861, "n_element_tab": 128, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 556, "n_element_tab_1": 110, "formula_len_all": 1142, "formula_len_all_1": 1058, "len_all": 161646, "len_all_1": 68814, "len_abs": 1533, "len_title": 109, "len_sents": 51284, "len_sents_1": 31157, "n_sents": 406, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (conditional spotlight)", "L_title": 61, "L_abs": 1554, "title": "OASIS Uncovers: High-Quality T2I Models, Same Old Stereotypes", "abs": "Images generated by text-to-image (T2I) models often exhibit visual biases and stereotypes of concepts such as culture and profession. Existing quantitative measures of stereotypes are based on statistical parity that does not align with the sociological definition of stereotypes and, therefore, incorrectly categorizes biases as stereotypes. Instead of oversimplifying stereotypes as biases, we propose a quantitative measure of stereotypes that aligns with its sociological definition. We then propose OASIS to measure the stereotypes in a generated dataset and understand their origins within the T2I model. OASIS includes two scores to measure stereotypes from a generated image dataset: **(M1)** Stereotype Score to measure the distributional violation of stereotypical attributes, and **(M2)** WALS to measure spectral variance in the images along a stereotypical attribute. OASIS\nalso includes two methods to understand the origins of stereotypes in T2I models: **(U1)** StOP to discover attributes that the T2I model internally associates with a given concept, and **(U2)** SPI to quantify the emergence of stereotypical attributes in the latent space of the T2I model during image generation. Despite the considerable progress in image fidelity, using OASIS, we conclude that newer T2I models such as FLUX.1 and SDv3 contain strong stereotypical predispositions about concepts and still generate images with widespread stereotypical attributes. Additionally, the quantity of stereotypes worsens for nationalities with lower Internet footprints.", "keywords": [ "Stereotype Measurement", "Responsible AI", "Trustworthy AI", "Interpretability" ], "rating_list": [ 6, 6, 8, 10, 6 ], "soundness_list": [ 2, 3, 3, 3, 4 ], "presentation_list": [ 3, 4, 4, 3, 2 ], "contribution_list": [ 2, 3, 3, 4, 3 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "L66G39JrM4", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Haobin Jiang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jiangxing Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zongqing Lu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 31, "n_formula_1": 12, "n_ref_uni": 58, "n_ref": 171, "n_ref_all": 215, "n_fig": 9, "n_tab": 10, "L_tab": 2672, "n_element_tab": 315, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 947, "n_element_tab_1": 91, "formula_len_all": 1462, "formula_len_all_1": 576, "len_all": 210456, "len_all_1": 75115, "len_abs": 1219, "len_title": 102, "len_sents": 65799, "len_sents_1": 36710, "n_sents": 484, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1224, "title": "Discrete Latent Plans via Semantic Skill Abstractions", "abs": "Skill learning from language instructions is a critical challenge in developing intelligent agents that can generalize across diverse tasks and follow complex human instructions. Hierarchical methods address this by decomposing the learning problem into multiple levels, where the high-level and low-level policies are mediated through a latent plan space. Effective modeling and learning of this latent plan space are key to enabling robust and interpretable skill learning. In this paper, we introduce LADS, a hierarchical approach that learns language-conditioned discrete latent plans through semantic skill abstractions. Our method decouples the learning of the latent plan space from the language-conditioned high-level policy to improve training stability. First, we incorporate a trajectory encoder to learn a discrete latent space with the low-level policy, regularized by language instructions. Next, we model the high-level policy as a categorical distribution over these discrete latent plans to capture the multi-modality of the dataset. Through experiments in simulated control environments, we demonstrate that LADS outperforms state-of-the-art methods in both skill learning and compositional generalization.", "keywords": [ "Hierarchical Learning", "Skill Learning", "Imitation Learning" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "L5nW2DxI5h", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bangding Yang", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "MS student" }, { "name": "Chansoo Lee", "gender": "Male", "institution": "University of Michigan", "country": "", "position": "PhD student" }, { "name": "Jorg Bornschein", "gender": "Male", "institution": "Google Deepmind", "country": "", "position": "Research Scientist" }, { "name": "Qiuyi Zhang", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Sagi Perel", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Postdoc" }, { "name": "Tung Nguyen", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Xingyou Song", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Research Scientist" }, { "name": "Yutian Chen", "gender": "unknown", "institution": "DeepMind", "country": "GB", "position": "Research Scientist" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 41, "n_ref_all": 50, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 115, "formula_len_all_1": 115, "len_all": 104386, "len_all_1": 72072, "len_abs": 718, "len_title": 124, "len_sents": 27205, "len_sents_1": 22415, "n_sents": 189, "n_sents_1": 146, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 720, "title": "Predicting from Strings: Language Model Embeddings for Bayesian Optimization", "abs": "Bayesian Optimization is ubiquitous in the field of experimental design and blackbox optimization for improving search efficiency, but has been traditionally restricted to regression models which are only applicable to fixed search spaces and tabular input features. We propose _Embed-then-Regress_, a paradigm for applying in-context regression over string inputs, through the use of string embedding capabilities of pretrained language models. By expressing all inputs as strings, we able to perform general-purpose regression for Bayesian Optimization over different search domains such as traditional and combinatorial optimization, obtaining comparable results to state-of-the-art Gaussian Process-based algorithms.", "keywords": [ "language", "model", "embedding", "bayesian", "optimization", "in", "context", "regression", "gaussian", "process", "meta", "learning" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 2, 4, 4 ] }, { "paper_id": "L5godAOC2z", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Changjiang Li", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Jiacheng Liang", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Tanqiu Jiang", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Ting Wang", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Associate Professor" }, { "name": "Yuhui Wang", "gender": "unknown", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Zian Wang", "gender": "Male", "institution": ", State University of New York at Stony Brook", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 19, "n_ref": 87, "n_ref_all": 105, "n_fig": 5, "n_tab": 6, "L_tab": 1456, "n_element_tab": 139, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1045, "n_element_tab_1": 77, "formula_len_all": 346, "formula_len_all_1": 346, "len_all": 138452, "len_all_1": 60036, "len_abs": 1410, "len_title": 131, "len_sents": 47582, "len_sents_1": 29080, "n_sents": 358, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1309, "title": "RobustKV: Defending Large Language Models against Jailbreak Attacks via KV Eviction", "abs": "Jailbreak attacks circumvent LLMs' built-in safeguards by concealing harmful queries within adversarial prompts. While most existing defenses attempt to mitigate the effects of adversarial prompts, they often prove inadequate as adversarial prompts can take arbitrary, adaptive forms. This paper introduces RobustKV, a novel jailbreak defense that takes a fundamentally different approach by selectively removing critical tokens of harmful queries from key-value (KV) caches. Intuitively, for an adversarial prompt to be effective, its tokens must achieve sufficient `importance' (measured by attention scores), which consequently lowers the importance of tokens in the concealed harmful query. Therefore, by carefully evicting the KVs of low-ranked tokens, RobustKV minimizes the harmful query's presence in the KV cache, thus preventing the LLM from generating informative responses. Extensive evaluation using benchmark datasets and models demonstrates that RobustKV effectively counters state-of-the-art jailbreak attacks while maintaining the LLM's performance on benign queries. Notably, RobustKV creates an interesting effectiveness-evasiveness dilemma for the adversary, leading to its robustness against adaptive attacks.{(Warning: This paper contains potentially harmful content generated by LLMs.)}", "keywords": [ "Jailbreak Attack", "Large Language Model", "KV cache optimization" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "L5fZHoaUCF", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Bibek Upadhayay", "gender": "Male", "institution": "University of New Haven", "country": "US", "position": "PhD student" }, { "name": "Vahid Behzadan", "gender": "Male", "institution": "University of New Haven", "country": "", "position": "Assistant Professor" }, { "name": "amin karbasi", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 76, "n_ref": 136, "n_ref_all": 176, "n_fig": 24, "n_tab": 11, "L_tab": 1255, "n_element_tab": 149, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1652, "n_element_tab_1": 119, "formula_len_all": 751, "formula_len_all_1": 65, "len_all": 320903, "len_all_1": 70965, "len_abs": 2553, "len_title": 108, "len_sents": 113811, "len_sents_1": 35232, "n_sents": 923, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1811, "title": "Cognitive Overload Attack: Prompt Injection for Long Context", "abs": "Large Language Models (LLMs) have demonstrated remarkable capabilities in performing tasks across various domains without needing explicit retraining. This capability, known as In-Context Learning (ICL), while impressive, exposes LLMs to a variety of adversarial prompts and jailbreaks that manipulate safety-trained LLMs into generating undesired or harmful output. In this paper, we propose a novel interpretation of ICL in LLMs through the lens of cognitive neuroscience, by drawing parallels between learning in human cognition with ICL. We applied the principles of Cognitive Load Theory in LLMs and empirically validate that similar to human cognition, LLMs also suffer from \\emph{cognitive overload}\u2014a state where the demand on cognitive processing exceeds the available capacity of the model, leading to potential errors. Furthermore, we demonstrated how an attacker can exploit ICL to jailbreak LLMs through deliberately designed prompts that induce cognitive overload on LLMs, thereby compromising the safety mechanisms of LLMs. We empirically validate this threat model by crafting various cognitive overload prompts and show that advanced models such as GPT-4, Claude-3.5 Sonnet, Claude-3 OPUS, LLAMA-3-70B-Instruct, Gemini-1.0-Pro, and Gemini-1.5-Pro can be successfully jailbroken, with attack success rates of up to 99.99\\%. Our findings highlight critical vulnerabilities in LLMs and underscore the urgency of developing robust safeguards. We propose integrating insights from cognitive load theory into the design and evaluation of LLMs to better anticipate and mitigate the risks of adversarial attacks. By expanding our experiments to encompass a broader range of models and by highlighting vulnerabilities in LLMs' ICL, we aim to ensure the development of safer and more reliable AI systems.", "keywords": [ "In Context Learning", "Cognitive Overload", "LLMs Safety", "Prompt Injection Attack" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 4, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 2, 3, 2 ] }, { "paper_id": "L5dUM6prKw", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Riza Batista-Navarro", "gender": "Female", "institution": "University of Manchester", "country": "GB", "position": "Associate Professor" }, { "name": "Viktor Schlegel", "gender": "unknown", "institution": "Imperial College London", "country": "SG", "position": "Researcher" }, { "name": "Yulong Wu", "gender": "unknown", "institution": "University of Manchester", "country": "GB", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 59, "n_ref": 100, "n_ref_all": 136, "n_fig": 8, "n_tab": 8, "L_tab": 1427, "n_element_tab": 166, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1480, "n_element_tab_1": 159, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 313266, "len_all_1": 69234, "len_abs": 1470, "len_title": 153, "len_sents": 49351, "len_sents_1": 34676, "n_sents": 329, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1475, "title": "Pay Attention to Real World Perturbations! Natural Robustness Evaluation in Machine Reading Comprehension", "abs": "As neural language models achieve human-comparable performance on Machine Reading Comprehension (MRC) and see widespread adoption, ensuring their robustness in real-world scenarios has become increasingly important. Current robustness evaluation research, though, primarily develops synthetic perturbation methods, leaving unclear how well they reflect real life scenarios. Considering this, we present a framework to automatically examine MRC models on naturally occurring textual perturbations, by replacing paragraph in MRC benchmarks with their counterparts based on available Wikipedia edit history. Such perturbation type is natural as its design does not stem from an arteficial generative process, inherently distinct from the previously investigated synthetic approaches. In a large-scale study encompassing SQUAD datasets and various model architectures we observe that natural perturbations result in performance degradation in pre-trained encoder language models. More worryingly, these state-of-the-art Flan-T5 and Large Language Models (LLMs) inherit these errors. Further experiments demonstrate that our findings generalise to natural perturbations found in other more challenging MRC benchmarks. In an effort to mitigate these errors, we show that it is possible to improve the robustness to natural perturbations by training on naturally or synthetically perturbed examples, though a noticeable gap still remains compared to performance on unperturbed data.", "keywords": [ "Natural Perturbations", "Robustness Evaluation", "Machine Reading Comprehension" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "L5NUDBdHqR", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Ankit Kariryaa", "gender": "Male", "institution": "University of Copenhagen", "country": "DK", "position": "Assistant Professor" }, { "name": "Christian Igel", "gender": "Male", "institution": "University of Copenhagen", "country": "DK", "position": "Full Professor" }, { "name": "Hui Zhang", "gender": "Female", "institution": "Copenhagen University", "country": "DK", "position": "PhD student" }, { "name": "Johan Mottelson", "gender": "Male", "institution": "Royal Danish Academy", "country": "", "position": "Assistant Professor" }, { "name": "Nico Lang", "gender": "Male", "institution": "Copenhagen University", "country": "DK", "position": "Postdoc" }, { "name": "Rem\u00edgio Chilaule", "gender": "unknown", "institution": "Royal Danish Academy", "country": "", "position": "PhD student" }, { "name": "Stefan Oehmcke", "gender": "Male", "institution": "Visual and Analytic Computing", "country": "DE", "position": "Assistant Professor" }, { "name": "Venkanna Babu Guthula", "gender": "Male", "institution": "Copenhagen University", "country": "DK", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 63, "n_ref_all": 89, "n_fig": 5, "n_tab": 8, "L_tab": 8865, "n_element_tab": 770, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 158526, "len_all_1": 62556, "len_abs": 1495, "len_title": 187, "len_sents": 47680, "len_sents_1": 31574, "n_sents": 384, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 138, "L_abs": 1502, "title": "Nacala-Roof-Material: Drone Imagery for Roof Detection, Classification, and Segmentation to Support Mosquito-borne Disease Risk Assessment", "abs": "As low-quality housing and in particular certain roof characteristics are associated with an increased risk of malaria, classification of roof types based on remote sensing imagery can support the assessment of malaria risk and thereby help prevent the disease. To support research in this area, we release the Nacala-Roof-Material dataset, which contains high-resolution drone images from Mozambique with corresponding labels delineating houses and specifying their roof types. The dataset defines a multi-task computer vision problem, comprising object detection, classification, and segmentation. In addition, we benchmarked various state-of-the-art approaches on the dataset. Canonical U-Nets, YOLOv8, and a custom decoder on pretrained DINOv2 served as baselines. We show that each of the methods has its advantages but none is superior on all tasks, which highlights the potential of our dataset for future research in multi-task learning. While the tasks are closely related, accurate segmentation of objects does not necessarily imply accurate instance separation, and vice versa. We address this general issue by introducing a variant of the deep ordinal watershed (DOW) approach that additionally separates the interior of objects, allowing for improved object delineation and separation. We show that our DOW variant is a generic approach that improves the performance of both U-Net and DINOv2 backbones, leading to a better trade-off between semantic segmentation and instance segmentation.", "keywords": [ "remote sensing", "deep learning", "disease risk assessment", "roof classification", "instance segmentation", "semantic segmentation", "multi-task learning" ], "rating_list": [ 5, 1, 3, 5 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "L4nH3j7L94", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Chunyuan Li", "gender": "unknown", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Dong Guo", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Haoqi Fan", "gender": "Male", "institution": "Facebook AI Research", "country": "US", "position": "Researcher" }, { "name": "Heng Huang", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Qinghao Ye", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Tianyi Xiong", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Xiyao Wang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 21, "n_ref": 103, "n_ref_all": 121, "n_fig": 2, "n_tab": 11, "L_tab": 8280, "n_element_tab": 683, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 4050, "n_element_tab_1": 226, "formula_len_all": 1395, "formula_len_all_1": 1448, "len_all": 184829, "len_all_1": 69073, "len_abs": 868, "len_title": 100, "len_sents": 49959, "len_sents_1": 30772, "n_sents": 361, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 875, "title": "LLaVA-Critic: Learning to Evaluate Multimodal Models", "abs": "We introduce LLaVA-Critic, the first open-source large multimodal model (LMM) designed as a generalist evaluator to assess performance across a wide range of multimodal tasks. LLaVA-Critic is trained using a high-quality critic instruction-following dataset that incorporates diverse evaluation criteria and scenarios. Our experiments demonstrate the model's effectiveness in two key areas: $(i)$ LMM-as-a-Judge, where LLaVA-Critic provides reliable evaluation scores, performing on par with or surpassing GPT models on multiple evaluation benchmarks; and $(ii)$ Preference Learning, where it generates reward signals for preference learning, enhancing model alignment capabilities. This work underscores the potential of open-source LMMs in self-critique and evaluation, setting the stage for future research into scalable, superhuman alignment feedback mechanisms for LMMs.", "keywords": [ "Large Multimodal Models", "LMM Evaluation", "Preference Learning" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "L3tW9nbcEM", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Wei Wang", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Qing Li", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 54, "n_ref_all": 72, "n_fig": 3, "n_tab": 3, "L_tab": 1066, "n_element_tab": 91, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1066, "n_element_tab_1": 91, "formula_len_all": 271, "formula_len_all_1": 271, "len_all": 238879, "len_all_1": 80110, "len_abs": 1224, "len_title": 92, "len_sents": 29222, "len_sents_1": 28147, "n_sents": 248, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1232, "title": "Schrodinger's Memory: Large Language Models", "abs": "Memory is the foundation of all human activities; without memory, it would be nearly impossible for people to perform any task in daily life. With the development of Large Language Models (LLMs), their language capabilities are becoming increasingly comparable to those of humans. But do LLMs have memory? Based on current practice, LLMs do appear to exhibit memory. So, what is the underlying mechanism of this memory? Previous research lacked a deep exploration of LLMs' memory capabilities and the underlying theory. In this paper, we use the Universal Approximation Theorem (UAT) to explain the memory mechanism in LLMs. We also conduct experiments to verify the memory capabilities of various LLMs, proposing a new method to assess their abilities based on the memory ability. We argue that LLM memory operates like Schr\u00f6dinger's memory, meaning that it only becomes observable when a specific memory is queried. We can only determine if the model retains a memory based on its output in response to the query; otherwise, it remains indeterminate. Finally, we expand on this concept by comparing the memory capabilities of the human brain and LLMs, highlighting the similarities and differences in their operational mechanisms.", "keywords": [ "Large Language Models' Memory" ], "rating_list": [ 3, 1, 1, 1 ], "soundness_list": [ 2, 1, 1, 1 ], "presentation_list": [ 3, 1, 1, 2 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "L3WnnnBRdu", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Binjian Xie", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Hao Wei", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Pengju Zhang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Yihong Wu", "gender": "Female", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 15, "n_ref_uni": 38, "n_ref": 68, "n_ref_all": 94, "n_fig": 11, "n_tab": 6, "L_tab": 2135, "n_element_tab": 651, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 266, "n_element_tab_1": 30, "formula_len_all": 1184, "formula_len_all_1": 899, "len_all": 128661, "len_all_1": 55034, "len_abs": 1298, "len_title": 155, "len_sents": 34225, "len_sents_1": 24319, "n_sents": 277, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1308, "title": "Hi-Gaussian: Hierarchical Gaussians under Normalized Spherical Projection for Single-View 3D Reconstruction", "abs": "Single-view 3D reconstruction is a fundamental problem in computer vision, having a significant impact on downstream tasks such as Autonomous Driving, Virtual Reality and Augment Reality. However, existing single-view reconstruction methods are unable to reconstruct the regions outside the input field-of-view or the areas occluded by visible parts. In this paper, we propose Hi-Gaussian, which employs feed-forward 3D Gaussians for efficient and generalizable single-view 3D reconstruction. A Normalized Spherical Projection module is introduced following an Encoder-Decoder network in our model, assigning a larger range to the transformed spherical coordinates, which can enlarge the field of view during scene reconstruction. Besides, to reconstruct occluded regions behind the visible part, we introduce a novel Hierarchical Gaussian Sampling strategy, utilizing two layers of Gaussians to hierarchically represent 3D scenes. We first use a pre-trained monocular depth estimation model to provide depth initialization for $leader$ Gaussians, and then leverage the $leader$ Gaussians to estimate the distribution followed by $follower$ Gaussians. Extensive experiments show that our method outperforms other methods for scene reconstruction and novel view synthesis, on both outdoor and indoor datasets.", "keywords": [ "Hierarchical Gaussian Sampling", "Normalized Spherical Projection", "Single-View 3D Reconstruction", "Gaussian Splatting" ], "rating_list": [ 5, 8, 5, 5 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 2, 4, 2, 2 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "L3DxhwXKZk", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Le Jiang", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Sarah Ostadabbas", "gender": "Female", "institution": "Northeastern University", "country": "US", "position": "Associate Professor" }, { "name": "Shaotong Zhu", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Shayda Moezzi", "gender": "Female", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Yedi Luo", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 24, "n_ref": 41, "n_ref_all": 63, "n_fig": 11, "n_tab": 2, "L_tab": 1620, "n_element_tab": 164, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 1626, "n_element_tab_1": 164, "formula_len_all": 1469, "formula_len_all_1": 1117, "len_all": 122903, "len_all_1": 62199, "len_abs": 1127, "len_title": 136, "len_sents": 34249, "len_sents_1": 26802, "n_sents": 237, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1127, "title": "ExpanDyNeRF: Expanding the Viewpoint of Dynamic Scenes beyond Constrained Camera Motions", "abs": "In the domain of dynamic Neural Radiance Fields (NeRF) for novel view synthesis, current state-of-the-art (SOTA) techniques struggle when the camera's pose deviates significantly from the primary viewpoint, resulting in unstable and unrealistic outcomes. This paper introduces Expanded Dynamic NeRF (ExpanDyNeRF), a monocular NeRF method that integrates a Gaussian splatting prior to tackle novel view synthesis with large-angle rotations. ExpanDyNeRF employs a pseudo ground truth technique to optimize density and color features, which enables the generation of realistic scene reconstructions from challenging viewpoints. Additionally, we present the Synthetic Dynamic Multiview (SynDM) dataset, the first GTA V-based dynamic multiview dataset designed specifically for evaluating robust dynamic reconstruction from significantly shifted views. We evaluate our method quantitatively and qualitatively on both the SynDM dataset and the widely recognized NVIDIA dataset, comparing it against other SOTA methods for dynamic scene reconstruction. Our evaluation results demonstrate that our method achieves superior performance.", "keywords": [ "NeRF", "Dynamic NeRF", "Generative Models", "Super Resolution" ], "rating_list": [ 3, 3, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 1, 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 5, 3 ] }, { "paper_id": "L39yPOGCma", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Naren Ramakrishnan", "gender": "Male", "institution": "Virginia Tech", "country": "US", "position": "Full Professor" }, { "name": "Rashed Shelim", "gender": "Male", "institution": "Virginia Tech", "country": "US", "position": "Postdoc" }, { "name": "Shengzhe Xu", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "PhD student" }, { "name": "Walid Saad", "gender": "unknown", "institution": "Virginia Tech", "country": "US", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 19, "n_ref_uni": 24, "n_ref": 46, "n_ref_all": 67, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1604, "formula_len_all_1": 1645, "len_all": 89406, "len_all_1": 59565, "len_abs": 1859, "len_title": 125, "len_sents": 29147, "len_sents_1": 27643, "n_sents": 205, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1869, "title": "When can isotropy help adapt LLMs' next word prediction to numerical domains?", "abs": "Recent studies have shown that vector representations of embeddings learned by pre-trained large language models (LLMs) are effective in various downstream tasks in numerical domains. Despite their significant benefits, the tendency of LLMs to hallucinate in such domains can have severe consequences in applications like finance, energy, retail, climate science, wireless networks, synthetic tabular generation, among others. To guarantee prediction reliability and accuracy in numerical domains, it is necessary to have performance guarantees through explainability. However, there is little theoretical understanding of when pre-trained language models help solve numeric downstream tasks. This paper seeks to bridge this gap by understanding when the next-word prediction capability of LLMs can be adapted to numerical domains through the lens of isotropy. Specifically, we first provide a general numeric data generation process that captures the core characteristics of numeric data across various numerical domains. Then, we consider a log-linear model for LLMs in which numeric data can be predicted from its context through a network with softmax as its last layer. We demonstrate that, in order to achieve state-of-the-art performance in numerical domains, the hidden representations of the LLM embeddings must possess a structure that accounts for the shift-invariance of the softmax function. We show how the isotropic property of LLM embeddings preserves the underlying structure of representations, thereby resolving the shift-invariance problem problem of softmax function. In other words, isotropy allows numeric downstream tasks to effectively leverage pre-trained representations, thus providing performance guarantees in the numerical domain. Experiments show that different characteristics of numeric data could have different impacts on isotropy.", "keywords": [ "Contextual embedding space", "clusters", "isotropy", "language model representation", "numeric downstream task" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 3, 2, 1 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "L34BvDTwls", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Adrien Bolland", "gender": "unknown", "institution": "University of Liege", "country": "", "position": "PhD student" }, { "name": "Damien Ernst", "gender": "Male", "institution": "University of Li\u00e8ge", "country": "", "position": "Full Professor" }, { "name": "Gaspard Lambrechts", "gender": "Male", "institution": "Universit\u00e9 de Li\u00e8ge", "country": "BE", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 24, "n_ref_uni": 44, "n_ref": 78, "n_ref_all": 106, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3110, "formula_len_all_1": 1914, "len_all": 130142, "len_all_1": 66880, "len_abs": 1225, "len_title": 126, "len_sents": 36960, "len_sents_1": 31212, "n_sents": 259, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1230, "title": "Off-Policy Maximum Entropy RL with Visitation Measures", "abs": "We introduce a new maximum entropy reinforcement learning framework based on the distribution of states and actions visited by a policy. More precisely, an intrinsic reward function is added to the reward function of the Markov decision process that shall be controlled. For each state and action, this intrinsic reward is the relative entropy of the discounted distribution of states and actions (or features from these states and actions) visited during the next time steps. We first prove that an optimal exploration policy, which maximizes the expected discounted sum of intrinsic rewards, is also a policy that maximizes a lower bound on the state-action value function of the decision process under some assumptions. We also prove that the visitation distribution used in the intrinsic reward definition is the fixed point of a contraction operator. Following, we describe how to adapt existing algorithms to learn this fixed point and compute the intrinsic rewards to enhance exploration. A new practical off-policy maximum entropy reinforcement learning algorithm is finally introduced. Empirically, exploration policies have good state-action space coverage, and high-performing control policies are computed efficiently.", "keywords": [ "reinforcement learning", "maximum entropy RL", "exploration" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 1, 1, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "L238BAx0wP", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Jiri Damborsky", "gender": "Male", "institution": "Masaryk University", "country": "", "position": "Full Professor" }, { "name": "Jiri Sedlar", "gender": "Male", "institution": "CIIRC, Czech Technical University, Czech Technical University of Prague", "country": "", "position": "Researcher" }, { "name": "Joan Planas-Iglesias", "gender": "Male", "institution": "Masaryk University", "country": "CZ", "position": "Postdoc" }, { "name": "Josef Sivic", "gender": "Male", "institution": "Czech Technical University in Prague", "country": "CZ", "position": "Principal investigator" }, { "name": "Petr Kouba", "gender": "unknown", "institution": "CIIRC, Czech Technical University, Czech Technical University of Prague", "country": "", "position": "Researcher" }, { "name": "Stanislav Mazurenko", "gender": "Male", "institution": "Masaryk University", "country": "", "position": "Principal Researcher" } ], "n_formula": 15, "n_formula_1": 7, "n_ref_uni": 52, "n_ref": 72, "n_ref_all": 105, "n_fig": 4, "n_tab": 7, "L_tab": 1245, "n_element_tab": 125, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 360, "n_element_tab_1": 56, "formula_len_all": 476, "formula_len_all_1": 271, "len_all": 187161, "len_all_1": 72023, "len_abs": 1332, "len_title": 88, "len_sents": 55709, "len_sents_1": 36420, "n_sents": 358, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1116, "title": "Learning to engineer protein flexibility", "abs": "Generative machine learning models are increasingly being used to design novel proteins. However, their major limitation is the inability to account for protein flexibility, a property crucial for protein function. Learning to engineer flexibility is difficult because the relevant data is scarce, heterogeneous, and costly to obtain using computational and experimental methods. Our contributions are three-fold. First, we perform a comprehensive comparison of methods for evaluating protein flexibility and identify relevant data for learning. Second, we overcome the data scarcity issue by leveraging a pre-trained protein language model. We design and train flexibility predictors utilizing either only sequential or both sequential and structural information on the input. Third, we introduce a method for fine-tuning a protein inverse folding model to make it steerable toward desired flexibility at specified regions. We demonstrate that our method Flexpert enables guidance of inverse folding models toward increased flexibility. This opens up a transformative possibility of engineering protein flexibility.", "keywords": [ "protein flexibility", "protein flexibility prediction", "protein design", "protein sequence design", "inverse folding" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "L14sqcrUC3", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Artem Babenko", "gender": "Male", "institution": "Yandex", "country": "RU", "position": "Researcher" }, { "name": "Ivan Rubachev", "gender": "Male", "institution": "Higher School of Economics", "country": "", "position": "PhD student" }, { "name": "Nikolay Kartashev", "gender": "Not Specified", "institution": "Higher School of Economics", "country": "", "position": "MS student" }, { "name": "Yury Gorishniy", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 56, "n_ref": 131, "n_ref_all": 150, "n_fig": 6, "n_tab": 12, "L_tab": 15224, "n_element_tab": 294, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 3159, "n_element_tab_1": 88, "formula_len_all": 46, "formula_len_all_1": 32, "len_all": 227614, "len_all_1": 73621, "len_abs": 1747, "len_title": 123, "len_sents": 73980, "len_sents_1": 33082, "n_sents": 644, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 75, "L_abs": 1727, "title": "Analyzing Pitfalls and Filling the Gaps in Tabular Deep Learning Benchmarks", "abs": "Advances in machine learning research drive progress in real-world applications. \nTo ensure this progress, it is important to understand the potential pitfalls on the way from a novel method's success on academic benchmarks to its practical deployment. In this work, we analyze existing tabular benchmarks and find two common characteristics of tabular data in typical industrial applications that are underrepresented in the datasets usually used for evaluation in the literature.\nFirst, in real-world deployment scenarios, distribution of data often changes over time. To account for this distribution drift, time-based train/test splits should be used in evaluation. However, existing academic tabular datasets generally lack timestamp metadata to enable such evaluation.\nSecond, a considerable portion of datasets in production settings stem from extensive data acquisition and feature engineering pipelines. This can have an impact on the absolute and relative number of predictive, uninformative, and correlated features compared to academic datasets.\nIn this work, we aim to understand how recent advances in tabular deep learning, which are evaluated on academic benchmarks, transfer to these underrepresented conditions.\nTo this end, we introduce TabReD -- a collection of eight industry-grade tabular datasets. \nWe reassess a large number of tabular ML models and techniques on TabReD. We demonstrate that evaluation on time-based data splits leads to different methods ranking, compared to evaluation on random splits, which are common in academic benchmarks. Furthermore, simple MLP-like architectures and GBDT show the best results on the TabReD datasets, while other methods are less effective in the new setting.", "keywords": [ "Tabular Data", "Benchmarks", "Reality Check", "Tabular Deep Learning", "Applications" ], "rating_list": [ 3, 8, 8, 8 ], "soundness_list": [ 2, 2, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 4, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "L143pPpIHv", "primary_area": "reinforcement learning", "n_author": 1, "author_info_list": [ { "name": "Jongsu Ha", "gender": "unknown", "institution": "Korea University", "country": "KR", "position": "Researcher" } ], "n_formula": 60, "n_formula_1": 16, "n_ref_uni": 37, "n_ref": 56, "n_ref_all": 80, "n_fig": 12, "n_tab": 5, "L_tab": 1336, "n_element_tab": 81, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 6386, "formula_len_all_1": 1286, "len_all": 155882, "len_all_1": 53811, "len_abs": 1688, "len_title": 85, "len_sents": 51480, "len_sents_1": 24612, "n_sents": 385, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 1775, "title": "CURIOSITY IS THE PATH TO OPTIMIZATION", "abs": "In PAC theory, it is posited that larger hypothesis spaces necessitate more independently and identically distributed (i.i.d) data to maintain the accuracy of model performance. PAC-MDP theory defines curiosity by assigning higher rewards for visiting states that are far from the previously visited trajectory, which supports more independent and i.i.d data collection. Recently, this field has witnessed attempts to narrow the hypothesis space by developing additional mechanisms that train multiple skills and facilitate the sharing of information among them, thereby discovering commonalities. However, one might wonder: What if curiosity could not only enhance the efficiency of data collection but also significantly reduce the hypothesis space, thereby driving optimal outcomes independently without additional mechanism used in PAC-MDP? Significant discussion has been devoted to the reduction of hypothesis spaces and the utilization of curiosity. Within this context, contrastive multi-skill reinforcement learning (RL) exhibits both traits. Previous research in contrastive multi-skill RL has utilized this technique primarily as a form of pretraining, However, there has been scant investigation into whether the technique itself can reduce the hypothesis space to optimize the outcomes. We have mathematically proven that curiosity provides bounds to guarantee optimality in contrastive multi-skill reinforcement learning (RL). Additionally, we have leveraged these findings to develop an algorithm that is applicable in real-world scenarios, which has been demonstrated to surpass other prominent algorithms. Furthermore, our experiments have shown that different skills are actually reducing the hypothesis space of the policy by being hierarchically grouped.", "keywords": [ "PAC - MDP", "Information Theory", "Unsupervised Skill Discovery" ], "rating_list": [ 1, 5, 3, 3 ], "soundness_list": [ 1, 3, 2, 1 ], "presentation_list": [ 1, 2, 1, 1 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "L0pXYjtfE3", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Amrutha Varshini Ramesh", "gender": "Female", "institution": "University of British Columbia", "country": "", "position": "PhD student" }, { "name": "Issam H. Laradji", "gender": "Male", "institution": "ServiceNow", "country": "US", "position": "Researcher" }, { "name": "Mark Schmidt", "gender": "unknown", "institution": "University of Alberta", "country": "", "position": "CIFAR AI Chair" }, { "name": "Vignesh Ganapathiraman", "gender": "Male", "institution": "Yahoo", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 9, "n_ref_uni": 41, "n_ref": 81, "n_ref_all": 103, "n_fig": 6, "n_tab": 11, "L_tab": 1527, "n_element_tab": 171, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 894, "n_element_tab_1": 69, "formula_len_all": 312, "formula_len_all_1": 745, "len_all": 133588, "len_all_1": 61146, "len_abs": 1760, "len_title": 149, "len_sents": 39154, "len_sents_1": 25922, "n_sents": 329, "n_sents_1": 223, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 1543, "title": "BlockLLM: Memory-Efficient Adaptation of LLMs by Selecting and Optimizing the Right Coordinate Blocks", "abs": "Training large language models (LLMs) for pretraining or adapting to new tasks and domains has become increasingly critical as their applications expand. However, as model and data sizes grow, the training process presents significant memory challenges, often requiring a prohibitive amount of GPU memory that may not be readily available. Existing methods such as low-rank adaptation (LoRA) add trainable low-rank matrix factorizations, altering the training dynamics and limiting the model's parameter search to a low-rank subspace. GaLore, a more recent method, employs Gradient Low-Rank Projection to reduce the memory footprint, in the full parameter training setting. However GaLore can only be applied to a subset of the LLM layers that satisfy the ``reversibility'' property, thus limiting their applicability. In response to these challenges, we introduce BlockLLM, an approach inspired by block coordinate descent. Our method carefully selects and updates a very small subset of the trainable parameters without altering any part of its architecture and training procedure. BlockLLM achieves state-of-the-art performance in both finetuning and pretraining tasks, while reducing the memory footprint of the underlying optimization process. Our experiments demonstrate that BlockLLM achieves superior performance on finetuning both large and small models. On pretraining a Llama model on C$4$ dataset, BlockLLM is able to train with significantly less memory than the state-of-the-art, while still maintaining competitive performance.", "keywords": [ "Memory efficient training", "LLM training" ], "rating_list": [ 3, 5, 3, 5, 3 ], "soundness_list": [ 2, 3, 1, 3, 3 ], "presentation_list": [ 2, 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "L0evcuybH5", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Kaiwen Hu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Yifei Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yisen Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhuo Ouyang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "zhang qi", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" } ], "n_formula": 32, "n_formula_1": 12, "n_ref_uni": 28, "n_ref": 85, "n_ref_all": 105, "n_fig": 9, "n_tab": 9, "L_tab": 1371, "n_element_tab": 113, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1010, "n_element_tab_1": 62, "formula_len_all": 2796, "formula_len_all_1": 1029, "len_all": 138836, "len_all_1": 59625, "len_abs": 1252, "len_title": 101, "len_sents": 51431, "len_sents_1": 28017, "n_sents": 389, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1257, "title": "Projection Head is Secretly an Information Bottleneck", "abs": "Recently, contrastive learning has risen to be a promising paradigm for extracting meaningful data representations. Among various special designs, adding a projection head on top of the encoder during training and removing it for downstream tasks has proven to significantly enhance the performance of contrastive learning. However, despite its empirical success, the underlying mechanism of the projection head remains under-explored. In this paper, we develop an in-depth theoretical understanding of the projection head from the information-theoretic perspective. By establishing the theoretical guarantees on the downstream performance of the features before the projector, we reveal that an effective projector should act as an information bottleneck, filtering out the information irrelevant to the contrastive objective. Based on theoretical insights, we introduce modifications to projectors with training and structural regularizations. Empirically, our methods exhibit consistent improvement in the downstream performance across various real-world datasets, including CIFAR-10, CIFAR-100, and ImageNet-100. We believe our theoretical understanding on the role of the projection head will inspire more principled and advanced designs in this field.", "keywords": [ "Projection Head", "Contrastive Learning", "Self-supervised Learning" ], "rating_list": [ 5, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "L0PciKdHsP", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Ajinkya Tejankar", "gender": "Male", "institution": "Predibase", "country": "", "position": "Researcher" }, { "name": "Hamed Pirsiavash", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Associate Professor" }, { "name": "Navaneet K L", "gender": "Male", "institution": "University of California, Davis", "country": "", "position": "PhD student" }, { "name": "Ujjawal Kamal Panchal", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "MS student" }, { "name": "Kossar Pourahmadi Meibodi", "gender": "Female", "institution": "University of California, Davis", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 47, "n_ref_all": 65, "n_fig": 3, "n_tab": 4, "L_tab": 2959, "n_element_tab": 177, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2959, "n_element_tab_1": 177, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 114560, "len_all_1": 66870, "len_abs": 1155, "len_title": 100, "len_sents": 29847, "len_sents_1": 28623, "n_sents": 272, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1164, "title": "MoIN: Mixture of Introvert Experts to Upcycle an LLM", "abs": "The goal of this paper is to improve (upcycle) an existing large language model without the prohibitive requirements of continued pre-training of the full-model. The idea is to split the pre-training data into semantically relevant groups and train an expert on each subset. An expert takes the form of a lightweight adapter added on the top of a frozen base model. During inference, an incoming query is first routed to the most relevant expert which is then loaded onto the base model for the forward pass. Unlike typical Mixture of Experts (MoE) models, the experts in our method do not work with other experts for a single query. Hence, we dub them ``introvert'' experts. Freezing the base model and keeping the experts as lightweight adapters allows extreme parallelism during training and inference. Training of all experts can be done in parallel without any communication channels between them. Similarly, the inference can also be heavily parallelized by distributing experts on different GPUs and routing each request to the GPU containing its relevant expert. We implement a proof-of-concept version of this method and show the validity of our approach.", "keywords": [ "llm", "lora", "moe" ], "rating_list": [ 3, 3, 6, 6 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 1, 2, 3 ], "confidence_list": [ 5, 4, 2, 3 ] }, { "paper_id": "L07zWidgdW", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Alona Fyshe", "gender": "Female", "institution": "University of Alberta", "country": "CA", "position": "Associate Professor" }, { "name": "Cory Daniel Efird", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "MS student" }, { "name": "Alex Graeme Murphy", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Postdoc" }, { "name": "Joel Zylberberg", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 30, "n_ref": 51, "n_ref_all": 87, "n_fig": 20, "n_tab": 1, "L_tab": 67, "n_element_tab": 2, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 67, "n_element_tab_1": 2, "formula_len_all": 372, "formula_len_all_1": 372, "len_all": 143088, "len_all_1": 66150, "len_abs": 2012, "len_title": 114, "len_sents": 48038, "len_sents_1": 31760, "n_sents": 391, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 2014, "title": "Finding Shared Decodable Concepts and their Negations in the Brain", "abs": "Prior work has offered evidence for functional localization in the brain; different anatomical regions preferentially activate for certain types of visual input. For example, the fusiform face area preferentially activates for visual stimuli that include a face. However, the spectrum of visual semantics is extensive, and only a few semantically-tuned patches of cortex have so far been identified in the human brain. Using a multimodal (natural language and image) neural network architecture (CLIP, \\cite{CLIP}, we train a highly accurate contrastive model that maps brain responses during naturalistic image viewing to CLIP embeddings. We then use a novel adaptation of the DBSCAN clustering algorithm to cluster the parameters of these participant-specific contrastive models. This reveals what we call Shared Decodable Concepts (SDCs): clusters in CLIP space that are decodable from common sets of voxels across multiple participants.\n\nExamining the images most and least associated with each SDC cluster gives us additional insight into the semantic properties of each SDC. We note SDCs for previously reported visual features (e.g. orientation tuning in early visual cortex) as well as visual semantic concepts such as faces, places and bodies. In cases where our method finds multiple clusters for a visuo-semantic concept, the least associated images allow us to dissociate between confounding factors. For example, we discovered two clusters of food images, one driven by color, the other by shape. We also uncover previously unreported areas with visuo-semantic sensitivity such as regions of extrastriate body area (EBA) tuned for legs/hands and sensitivity to numerosity in right intraparietal sulcus, sensitivity associated with visual perspective (close/far) and more. Thus, our contrastive-learning methodology better characterizes new and existing visuo-semantic representations in the brain by leveraging multimodal neural network representations and a novel adaptation of clustering algorithms.", "keywords": [ "fMRI", "decoding", "computer vision", "neuroscience" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "KzokzKV4JK", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Jiacheng Zhang", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" }, { "name": "Yuhao Sun", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Intern" }, { "name": "Zesheng Ye", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Postdoc" }, { "name": "chaowei xiao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 73, "n_formula_1": 23, "n_ref_uni": 27, "n_ref": 90, "n_ref_all": 114, "n_fig": 7, "n_tab": 14, "L_tab": 5206, "n_element_tab": 466, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1153, "n_element_tab_1": 69, "formula_len_all": 5531, "formula_len_all_1": 1076, "len_all": 167145, "len_all_1": 69809, "len_abs": 1385, "len_title": 124, "len_sents": 54639, "len_sents_1": 31638, "n_sents": 464, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1398, "title": "Sample-specific Noise Injection for Diffusion-based Adversarial Purification", "abs": "Diffusion-based purification (DBP) methods aim to remove adversarial noise from the input sample by first injecting Gaussian noise through a forward diffusion process, and then recovering the clean example through a reverse generative process. In the above process, how much Gaussian noise is injected to the input sample is key to the success of DBP methods, which is controlled by a constant noise level $t^*$ for all samples in existing methods. In this paper, we discover that an optimal $t^*$ for each sample indeed could be different. Intuitively, the cleaner a sample is, the less the noise it should be injected, and vice versa. Motivated by this finding, we propose a new framework, called Sample-specific Score-aware Noise Injection (SSNI). Specifically, SSNI uses a pre-trained score network to estimate how much a data point deviates from the clean data distribution (i.e., score norms). Then, based on the magnitude of score norms, SSNI applies a reweighting function to adaptively adjust $t^*$ for each sample, achieving sample-specific noise injections. Empirically, incorporating our framework with existing DBP methods results in a notable improvement in both accuracy and robustness on CIFAR-10 and ImageNet-1K, highlighting the necessity to allocate distinct noise levels to different samples in DBP methods. Our code is available at: https://anonymous.4open.science/r/SSNI-F746.", "keywords": [ "diffusion-based adversarial purification", "adversarial purification", "adversarial robustness", "accuracy-robustness trade-off" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "KzSGJy1PIf", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Doina Precup", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Research Team Lead" }, { "name": "Eleni Triantafillou", "gender": "Female", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Gintare Karolina Dziugaite", "gender": "Female", "institution": "McGill University", "country": "CA", "position": "Adjunct Professor" }, { "name": "Hugo Larochelle", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "Adjunct Professor" }, { "name": "James J. Clark", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Full Professor" }, { "name": "Nazanin Mohammadi Sepahvand", "gender": "Female", "institution": "McGill University, McGill University", "country": "", "position": "PhD student" }, { "name": "Daniel M Roy", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 7, "n_ref_uni": 24, "n_ref": 48, "n_ref_all": 83, "n_fig": 19, "n_tab": 6, "L_tab": 1590, "n_element_tab": 175, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 602, "formula_len_all_1": 366, "len_all": 130635, "len_all_1": 59132, "len_abs": 1419, "len_title": 130, "len_sents": 50434, "len_sents_1": 30972, "n_sents": 335, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1426, "title": "Selective Unlearning via Representation Erasure Using Adversarial Training", "abs": "When deploying machine learning models in the real world, we often face the challenge of \u201cunlearning\u201d specific data points or subsets after training. Inspired by Domain-Adversarial Training of Neural Networks (DANN), we propose a novel algorithm, SURE, for targeted unlearning. SURE treats the process as a domain adaptation problem, where the \u201cforget set\u201d (data to be removed) and a validation set from the same distribution form two distinct domains. We train a domain classifier to discriminate between representations from the forget and validation sets.Using a gradient reversal strategy similar to DANN, we perform gradient updates to the representations to \u201cfool\u201d the domain classifier and thus obfuscate representations belonging to the forget set. Simultaneously, gradient descent is applied to the retain set (original training data minus the forget set) to preserve its classification performance. Unlike other unlearning approaches whose training objectives are built based on model outputs, SURE directly manipulates there presentations.This is key to ensure robustness against a set of more powerful attacks than currently considered in the literature, that aim to detect which examples were unlearned through access to learned embeddings. Our thorough experiments reveal that SURE has a better unlearning quality to utility trade-off compared to other standard unlearning techniques for deep neural networks.", "keywords": [ "approximate unlearning", "adversarial training" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "Kz10l3roV0", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Hangtian Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiaolei Xie", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 15, "n_ref": 32, "n_ref_all": 38, "n_fig": 1, "n_tab": 3, "L_tab": 1162, "n_element_tab": 94, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1162, "n_element_tab_1": 94, "formula_len_all": 172, "formula_len_all_1": 172, "len_all": 58053, "len_all_1": 58053, "len_abs": 1086, "len_title": 169, "len_sents": 19999, "len_sents_1": 19999, "n_sents": 141, "n_sents_1": 141, "venue": "ICLR2025", "decision": "withdraw", "L_title": 121, "L_abs": 1092, "title": "DIMS: Channel-Dependent and Seasonal-Trend Independent Transformer Using Multi-Stage Training for Time Series Forecasting", "abs": "Due to the limited size of real-world time series data, current transformer-based time series forecasting algorithms often struggle with overfitting. Common techniques used to mitigate overfitting include channel-independence and seasonal-trend decomposition. However, channel-independent inevitably results in the loss of inter-channel dependencies, and existing seasonal-trend decomposition methods are insufficient in effectively mitigating overfitting. In this study, we propose DIMS, a time series forecasting model that uses multi-stage training to capture inter-channel dependencies while ensuring the independence of seasonal and trend components. The computation of channel dependency is postponed to the later stage, following the channel-independent training, while the seasonal and trend components remain fully independent during the early training phases. This approach enables the model to effectively capture inter-channel dependencies while minimizing overfitting. Experiments show that our model outperforms the state-of-the-art transformer-based models on several datasets.", "keywords": [ "Time series", "Deep learning", "Transformer" ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "KyqtKhv6q1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Brady Zhou", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Philipp Kraehenbuehl", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 10, "n_ref": 30, "n_ref_all": 44, "n_fig": 7, "n_tab": 3, "L_tab": 506, "n_element_tab": 39, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 506, "n_element_tab_1": 39, "formula_len_all": 35, "formula_len_all_1": 35, "len_all": 87821, "len_all_1": 48865, "len_abs": 868, "len_title": 92, "len_sents": 22295, "len_sents_1": 21639, "n_sents": 192, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 859, "title": "3D Perception with Differentiable Map Priors", "abs": "Human drivers rarely navigate where no person has gone before. After all, thousands of drivers use busy city roads every day, and only one can claim to be the first. The same holds for autonomous computer vision systems. The vast majority of the deployment area of an autonomous vision system will have been visited before. Yet, most computer vision systems act as if they are encountering each location for the first time. In this work, we present Differentiable Map Priors, a simple but effective framework to learn spatial priors from historic traversals. Differentiable Map Priors easily integrate into leading 3D perception systems at little to no extra computational costs. We show that they lead to a significant and consistent improvement in 3D object detection and semantic map segmentation tasks on the nuScenes dataset across several architectures.", "keywords": [ "autonomous driving", "3D object detection", "mapping" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "KyeyEFPxJX", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Liam Paull", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Associate Professor" }, { "name": "Mahtab Sandhu", "gender": "unknown", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "", "position": "MS student" }, { "name": "Sabyasachi Sahoo", "gender": "Male", "institution": "Universit\u00e9 Laval", "country": "CA", "position": "PhD student" }, { "name": "Samer B. Nashed", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Postdoc" }, { "name": "Yann Batiste Pequignot", "gender": "Male", "institution": "Universite Laval, Universit\u00e9 Laval", "country": "CA", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 46, "n_ref": 105, "n_ref_all": 120, "n_fig": 4, "n_tab": 6, "L_tab": 1571, "n_element_tab": 183, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 743, "n_element_tab_1": 28, "formula_len_all": 665, "formula_len_all_1": 665, "len_all": 145331, "len_all_1": 66345, "len_abs": 1509, "len_title": 133, "len_sents": 44557, "len_sents_1": 32921, "n_sents": 268, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1515, "title": "CLIP-Enhance: Improving CLIP Zero-Shot Classification via von Mises-Fisher Clustering", "abs": "Contrastive language-image pre-training (CLIP) has revolutionized computer vision by integrating natural language understanding with image analysis, enabling zero-shot classification without prior training on specific classes. However, recent efforts to improve the performance of frozen CLIP models through prompt tuning and adapter mechanisms have introduced additional system complexity and training requirements, thus undermining CLIP's inherent efficiency in zero-shot knowledge transfer. In this paper, we propose to address two common challenges in zero-shot classification using CLIP: 1) the misalignment between textual and image embeddings, and 2) the long-tailed distribution of CLIP's training dataset. Our approach, CLIP-Enhance, is motivated by a re-interpretation of CLIP zero-shot classification as a clustering problem on a hypersphere using a von Mises-Fisher mixture model. Inspired by the DINO self-supervised learning framework, we optimize this mixture model to simultaneously improve the alignment of textual and image embeddings as well as represent data distribution disparities between training and evaluation datasets. Empirically, we show that jointly optimizing for both embedding alignment and concentration via self-supervised learning improves CLIP zero-shot classification significantly across multiple benchmark datasets. We also show empirically how CLIP-Enhance mitigates problems (1) and (2), as well as its robustness to limited data through a series of additional experiments.", "keywords": [ "zero-shot classification", "multi-modal representation learning", "knowledge distillation", "CLIP" ], "rating_list": [ 3, 5, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 1, 2, 2, 1, 4 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 4, 2, 4, 3, 4 ] }, { "paper_id": "KyKTjRtyNG", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Jiawei Qiu", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Lei Zhao", "gender": "Male", "institution": "Wuhan University", "country": "", "position": "Full Professor" }, { "name": "Wei Xu", "gender": "Female", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Yiru Zhao", "gender": "Female", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "peiyaoyuan", "gender": "Female", "institution": "Wuhan University", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 31, "n_ref": 47, "n_ref_all": 65, "n_fig": 5, "n_tab": 8, "L_tab": 3938, "n_element_tab": 175, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 431, "n_element_tab_1": 57, "formula_len_all": 199, "formula_len_all_1": 315, "len_all": 142127, "len_all_1": 64540, "len_abs": 2443, "len_title": 160, "len_sents": 47825, "len_sents_1": 32116, "n_sents": 359, "n_sents_1": 229, "venue": "ICLR2025", "decision": "withdraw", "L_title": 112, "L_abs": 1644, "title": "Incremental Exploits: Efficient Jailbreaks on Large Language Models with Multi-round Conversational Jailbreaking", "abs": "As large language models (LLMs) become widely deployed across various domains, security concerns---particularly jailbreak attacks that bypass built-in safety mechanisms---have emerged as significant risks. Existing jailbreak methods focus mainly on single-turn interactions and face limitations in generalizability and practicality. In this paper, we propose a novel method called Multi-round Conversational Jailbreaking (MRCJ), which exploits the unintended competition between a LLMs' safety alignment and its in-context learning objectives during extended conversations. By incrementally introducing increasingly malicious content, the LLMs' tendency to maintain contextual consistency can override its safety protocols, ultimately leading to harmful outputs. To facilitate conversation flow generation, we developed a dataset containing 12,000 questions, categorized into six types of security topics, and classified across four levels of severity, spanning ten languages. Compared to existing methods, MRCJ demonstrates superior efficiency, applicability, and effectiveness by fully exploiting the potential of multi-round conversations. In experiments, MRCJ achieves a jailbreak success rate of over 90\\% across widely-used LLMs, requiring fewer than five queries on average, and significantly outperforms baselines on both metrics. Our findings expose vulnerabilities in current LLMs during extended conversations and highlight the need for improved safety mechanisms that consider multi-round interactions. The source code and dataset are available at (URL omitted for double-blind reviewing; code available in supplementary materials).", "keywords": [ "Large Language Models", "Model Vulnerabilities", "Multi-round Conversational Jailbreaking" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 1, 2 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 1, 2, 1 ], "confidence_list": [ 4, 4, 4, 4, 5 ] }, { "paper_id": "KxjKZvyFgx", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Heesoo Jung", "gender": "unknown", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "PhD student" }, { "name": "Hogun Park", "gender": "unknown", "institution": "Sungkyunkwan University", "country": "KR", "position": "Associate Professor" }, { "name": "Jaehyun Bae", "gender": "Female", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 13, "n_ref_uni": 36, "n_ref": 63, "n_ref_all": 85, "n_fig": 7, "n_tab": 10, "L_tab": 2880, "n_element_tab": 336, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2564, "n_element_tab_1": 190, "formula_len_all": 1130, "formula_len_all_1": 389, "len_all": 149855, "len_all_1": 67470, "len_abs": 1840, "len_title": 143, "len_sents": 48792, "len_sents_1": 30658, "n_sents": 371, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1850, "title": "Harnessing Spatial Dependency for Domain Generalization in Multivariate Time-series Sensor Data", "abs": "Multivariate time-series (MTS) data from multiple sensors often vary across domains due to factors like sensor misalignment, reattachment, or individual differences, posing significant challenges for domain generalization (DG). Existing methods inadequately address the alignment of domain-specific spatial dependencies across different domains in MTS data, as they often assume a unified invariant spatial structure and overlook the distributional discrepancies arising from varying sensor relationships. To address this limitation, we propose ASAM (Adaptive Spatial Dependency Alignment in MTS Data for Domain Generalization), a novel framework that adaptively aligns spatial dependencies across domains. ASAM proposes a DG layer with domain generalization loss function and two-view regularization loss functions to align spatial dependencies between domains adaptively. We adopt a two-phase approach to align different sets of domains effectively. An input-aware graph generation process and a GNN-based DG layer, coupled with the domain generalization loss function, adaptively align the spatial dependencies learned in the second phase with those from the first phase, ensuring a more precise alignment. We additionally incorporate a two-view reg- ularization method to effectively capture underlying spatiotemporal information comprised of spatial decorrelation loss and Gaussian kernel loss. Our theoretical analysis demonstrates that ASAM effectively assimilates information bottleneck, ensuring robustness across diverse distributions. Extensive evaluations of the four real-world datasets show ASAM outperforms ten recent baselines. To the best of our knowledge, this work is among the first to explore DG approaches for MTS data by focusing on spatial dependency alignment. Our code is available at https://anonymous.4open.science/r/ASAM.", "keywords": [ "Domain Generalization. Multivariate Time-series. sensor and healthcare." ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 2, 4, 2 ] }, { "paper_id": "KxQnhe5UuJ", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Amos J Storkey", "gender": "Not Specified", "institution": "University of Edinburgh", "country": "", "position": "Full Professor" }, { "name": "Elliot J. Crowley", "gender": "Male", "institution": "University of Edinburgh", "country": "GB", "position": "Associate Professor" }, { "name": "Linus Ericsson", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Postdoc" }, { "name": "Sigrid Passano Hellan", "gender": "unknown", "institution": "NORCE Norwegian Research Centre", "country": "NO", "position": "Researcher" }, { "name": "Thomas L Lee", "gender": "unknown", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 30, "n_ref": 94, "n_ref_all": 118, "n_fig": 3, "n_tab": 3, "L_tab": 123, "n_element_tab": 15, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 172, "n_element_tab_1": 15, "formula_len_all": 192, "formula_len_all_1": 192, "len_all": 108341, "len_all_1": 64541, "len_abs": 1172, "len_title": 115, "len_sents": 40252, "len_sents_1": 33493, "n_sents": 259, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1181, "title": "Towards Realistic Hyperparameter Optimization in Continual Learning", "abs": "In continual learning (CL)\u2014where a learner trains on a stream of data\u2014standard hyperparameter optimisation (HPO) cannot be applied, as a learner does not have access to all of the data at the same time. This has prompted the development of CL-specific HPO frameworks. The most popular way to tune hyperparameters in CL is to repeatedly train over the whole data stream with different hyperparameter settings. However, this *end-of-training* HPO is unrealistic as in practice a learner can only see the stream once. Hence, there is an open question: *what HPO framework should a practitioner use for a CL problem in reality?* This paper answers this question by comparing several realistic HPO frameworks. We find that none of the HPO frameworks considered, including end-of-training HPO, perform consistently better than the rest on popular CL benchmarks. We therefore arrive at a twofold conclusion: a) on the popular CL benchmarks examined, a CL practitioner should select the HPO framework based on other factors, for example compute efficiency and b) to be able to discriminate between HPO frameworks there is a need to move beyond the current most commonly used CL benchmarks.", "keywords": [ "Contiual Learning", "HPO" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 5, 5, 3, 4 ] }, { "paper_id": "KxQRHOre9D", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Akiko Eriguchi", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Senior Researcher" }, { "name": "Haoran Xu", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Hieu Hoang", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Huda Khayrallah", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "HyoJung Han", "gender": "Female", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Marine Carpuat", "gender": "Female", "institution": "University of Maryland, College Park", "country": "", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 33, "n_ref": 93, "n_ref_all": 112, "n_fig": 9, "n_tab": 12, "L_tab": 7593, "n_element_tab": 807, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2170, "n_element_tab_1": 149, "formula_len_all": 386, "formula_len_all_1": 382, "len_all": 154227, "len_all_1": 64388, "len_abs": 1114, "len_title": 120, "len_sents": 33971, "len_sents_1": 29080, "n_sents": 223, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1099, "title": "Adapters for Altering LLM Vocabularies: What Languages Benefit the Most?", "abs": "Vocabulary adaptation, which integrates new vocabulary into pre-trained language models (LMs), enables expansion to new languages and mitigates token over-fragmentation. However, existing approaches are limited by their reliance on heuristic or external embeddings. We propose VocADT, a novel method for vocabulary adaptation using adapter modules that are trained to learn the optimal linear combination of existing embeddings while keeping the model's weights fixed. VocADT offers a flexible and scalable solution without requiring external resources or language constraints. Across 11 languages\u2014with various scripts, resource availability, and fragmentation\u2014we demonstrate that VocADT outperforms the original Mistral model and other baselines across various multilingual tasks. We find that Latin-script languages and highly fragmented languages benefit the most from vocabulary adaptation. We further fine-tune the adapted model on the generative task of machine translation and find that vocabulary adaptation is still beneficial after fine-tuning and that VocADT is the most effective method.", "keywords": [ "Vocabulary Adaptation", "Vocabulary Transfer", "Tokenizer Transfer", "Initializing Embedding", "Adapter", "Multilingual", "Machine Translation" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "KxGGZag9gW", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 9, "author_info_list": [ { "name": "Aayush Mishra", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Alan Yuille", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Ani Nenkova", "gender": "Female", "institution": "Adobe Research", "country": "US", "position": "Principal Scientist" }, { "name": "Ankit Vaidya", "gender": "Male", "institution": "SCTR's Pune Institute of Computer Technology", "country": "IN", "position": "Undergrad student" }, { "name": "Anqi Liu", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" }, { "name": "Jiuxiang Gu", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Prakhar Kaushik", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Raghavendra Addanki", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" }, { "name": "Ryan Rossi", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Senior Research Scientist" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 49, "n_ref": 79, "n_ref_all": 110, "n_fig": 14, "n_tab": 10, "L_tab": 2710, "n_element_tab": 331, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1528, "n_element_tab_1": 117, "formula_len_all": 1555, "formula_len_all_1": 1350, "len_all": 169903, "len_all_1": 58812, "len_abs": 1059, "len_title": 83, "len_sents": 40724, "len_sents_1": 25557, "n_sents": 308, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1064, "title": "EigenLoRA: Recycle trained Adapters for Resource Efficient Adaptation and Inference", "abs": "Low-Rank Adapters (LoRA) are lightweight components that have made fine-tuning large models on domain-specific tasks inexpensive. This has resulted in an abundance of adapters in a growing open-source public community. We ask the question: can these adapters be used to inform and further streamline adaptation to new tasks? We introduce EigenLoRA, a parameter-efficient fine-tuning method that uses trained adapters to perform fast adaptation on new domains with orders of magnitude fewer parameters than LoRA. Our method finds a principal subspace that aligns with the domain of the trained adapters. This allows for efficient and fast adaptation to new tasks in this domain by simply learning coefficients on the principal components of this subspace. Furthermore, EigenLoRA makes inference time task-switching memory efficient. Instead of saving and loading whole LoRAs, EigenLoRA can simply load lightweight coefficients. EigenLoRA works across a variety of domains and tasks and is a viable solution for edge-based and efficient personalization applications.", "keywords": [ "Parameter-efficient fine-tuning", "Transfer learning", "Low-rank", "NLP", "vision", "diffusion", "efficient learning", "eco-friendly" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "Kwo20MWWCb", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Daniel Cederberg", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Mikael Johansson", "gender": "Male", "institution": "KTH Royal Institute of Technology, Stockholm, Sweden", "country": "", "position": "Full Professor" }, { "name": "Stephen P. Boyd", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Full Professor" }, { "name": "Xuyang Wu", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 68, "n_formula_1": 27, "n_ref_uni": 48, "n_ref": 79, "n_ref_all": 106, "n_fig": 5, "n_tab": 2, "L_tab": 218, "n_element_tab": 35, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 218, "n_element_tab_1": 35, "formula_len_all": 7268, "formula_len_all_1": 2146, "len_all": 166859, "len_all_1": 68170, "len_abs": 699, "len_title": 63, "len_sents": 43312, "len_sents_1": 29387, "n_sents": 431, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 702, "title": "An Asynchronous Bundle Method for Distributed Learning Problems", "abs": "We propose a novel asynchronous bundle method to solve distributed learning problems. Compared to existing asynchronous methods, our algorithm computes the next iterate based on a more accurate approximation of the objective function and does not require any prior information about the maximal information delay in the system. This makes the proposed method fast and easy to tune. We prove that the algorithm converges in both deterministic and stochastic (mini-batch) settings, and quantify how the convergence times depend on the level of asynchrony. The practical advantages of our method are illustrated through numerical experiments on classification problems of varying complexities and scales.", "keywords": [ "Distributed optimization; asynchronous optimization; model-based optimization" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "KwaNnvj0b3", "primary_area": "causal reasoning", "n_author": 9, "author_info_list": [ { "name": "Boyang Sun", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Guang-Yuan Hao", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Haoyue Dai", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Ignavier Ng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Peter Spirtes", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Shunxing Fan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "Xinshuai Dong", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Yumou Qiu", "gender": "unknown", "institution": "Iowa State University", "country": "", "position": "Associate Professor" } ], "n_formula": 42, "n_formula_1": 22, "n_ref_uni": 37, "n_ref": 70, "n_ref_all": 81, "n_fig": 6, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2514, "formula_len_all_1": 1595, "len_all": 119551, "len_all_1": 62384, "len_abs": 1281, "len_title": 161, "len_sents": 41298, "len_sents_1": 28859, "n_sents": 316, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 113, "L_abs": 1277, "title": "Permutation-based Rank Test in the Presence of Discretization and Application in Causal Discovery with Mixed Data", "abs": "Recent advances have shown that statistical tests for the rank of cross-covariance matrices play an important role in causal discovery. These rank tests include partial correlation tests as special cases and provide further graphical information about latent variables.\n Existing rank tests typically assume that all the continuous variables can be perfectly measured,\n and yet, in practice many variables can only be measured after discretization.\n For example, in psychometric studies,\n the continuous level of certain personality dimensions of a person can only be measured after being discretized into order-preserving options such as disagree, neutral, and agree.\n Motivated by this, we\npropose Mixed data Permutation-based Rank Test (MPRT), which properly controls the statistical errors even when some or all variables are discretized.\nTheoretically, we establish the \nexchangeability and \nestimate the asymptotic null distribution by permutations;\nas a consequence,\nMPRT can effectively control the Type I error in the presence of discretization while previous methods cannot. \nEmpirically, our method is validated by extensive experiments on synthetic data \nand real-world data to demonstrate its effectiveness as well as applicability in causal discovery.", "keywords": [ "Rank Test; Discretization; Causal Discovery" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 1 ] }, { "paper_id": "KwPUQOQIKt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Abhinav Rastogi", "gender": "Male", "institution": "Google", "country": "", "position": "Research Engineer" }, { "name": "Harsh Lara", "gender": "Male", "institution": "Research, Google", "country": "", "position": "Researcher" }, { "name": "Jiao Sun", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Lei Meng", "gender": "Female", "institution": "University of Notre Dame", "country": "", "position": "PhD student" }, { "name": "Lei Shu", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Liangchen Luo", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Meiqi Guo", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Rosanne Liu", "gender": "Female", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Samrat Phatale", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Yinxiao Liu", "gender": "Male", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Yunxuan Li", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 27, "n_ref": 88, "n_ref_all": 108, "n_fig": 5, "n_tab": 2, "L_tab": 423, "n_element_tab": 49, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 423, "n_element_tab_1": 49, "formula_len_all": 241, "formula_len_all_1": 240, "len_all": 134179, "len_all_1": 55729, "len_abs": 1936, "len_title": 132, "len_sents": 33924, "len_sents_1": 26593, "n_sents": 252, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1963, "title": "Improve Mathematical Reasoning in Language Models with Automated Process Supervision", "abs": "Complex multi-step reasoning tasks, such as solving mathematical problems or generating code, remain a significant hurdle for even the most advanced large language models (LLMs). Verifying LLM outputs with an Outcome Reward Model (ORM) is a standard inference-time technique aimed at enhancing the reasoning performance of LLMs. However, this still proves insufficient for reasoning tasks with a lengthy or multi-hop reasoning chain, where the intermediate outcomes are neither properly rewarded nor penalized. Process supervision addresses this limitation by assigning intermediate rewards during the reasoning process. To date, the methods used to collect process supervision data have relied on either human annotation or per-step Monte Carlo estimation, both prohibitively expensive to scale, thus hindering the broad application of this technique. In response to this challenge, we propose a novel divide-and-conquer style Monte Carlo Tree Search (MCTS) algorithm named \\textit{OmegaPRM} for the efficient collection of high-quality process supervision data. This algorithm swiftly identifies the first error in the Chain of Thought (CoT) with binary search and balances the positive and negative examples, thereby ensuring both efficiency and quality. As a result, we are able to collect over 1.5 million process supervision annotations to train Process Reward Models (PRMs). This fully automated process supervision alongside the weighted self-consistency algorithm is able to enhance LLMs' math reasoning performances. We improved the success rates of the instruction-tuned Gemini Pro model from 51\\% to 69.4\\% on MATH500 and from 86.4\\% to 93.6\\% on GSM8K. Similarly, we boosted the success rates of Gemma2 27B from 42.3\\% to 58.2\\% on MATH500 and from 74.0\\% to 92.2\\% on GSM8K.\nThe entire process operates without any human intervention or supervision, making our method both financially and computationally cost-effective compared to existing methods.", "keywords": [ "LLM", "Reasoning", "PRM" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 3, 1, 3 ], "presentation_list": [ 2, 1, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "Kvdh12wGC0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "So Kuroki", "gender": "Male", "institution": "Sakana AI", "country": "JP", "position": "Research Engineer" }, { "name": "Taishi Nakamura", "gender": "unknown", "institution": "Institute of Science Tokyo", "country": "JP", "position": "MS student" }, { "name": "Takuya Akiba", "gender": "unknown", "institution": "Sakana AI", "country": "JP", "position": "Researcher" }, { "name": "Yujin Tang", "gender": "Male", "institution": "Sakana AI", "country": "JP", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 20, "n_ref_uni": 44, "n_ref": 82, "n_ref_all": 115, "n_fig": 8, "n_tab": 4, "L_tab": 866, "n_element_tab": 135, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 984, "n_element_tab_1": 149, "formula_len_all": 13370, "formula_len_all_1": 25851, "len_all": 202680, "len_all_1": 98340, "len_abs": 1574, "len_title": 109, "len_sents": 50735, "len_sents_1": 34741, "n_sents": 360, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1580, "title": "Agent Skill Acquisition for Large Language Models via CycleQD", "abs": "Training large language models to acquire specific skills remains a challenging endeavor. Conventional training approaches often struggle with data distribution imbalances and inadequacies in objective functions that do not align well with task-specific performance. To address these challenges, we introduce CycleQD, a novel approach that leverages the Quality Diversity framework through a cyclic adaptation of the algorithm, along with a model merging based crossover and an SVD-based mutation. In CycleQD, each task\u2019s performance metric is alternated as the quality measure while the others serve as the behavioral characteristics. This cyclic focus on individual tasks allows for concentrated effort on one task at a time, eliminating the need for data ratio tuning and simplifying the design of the objective function. Empirical results from AgentBench indicate that applying CycleQD to LLAMA3-8B-INSTRUCT based models not only enables them to surpass traditional fine-tuning methods in coding, operating systems, and database tasks, but also achieves performance on par with GPT-3.5-TURBO, which potentially contains much more parameters, across these domains. Crucially, this enhanced performance is achieved while retaining robust language capabilities, as evidenced by its performance on widely adopted language benchmark tasks. We highlight the key design choices in CycleQD, detailing how these contribute to its effectiveness. Furthermore, our method is general and can be applied to image segmentation models, highlighting its applicability across different domains.", "keywords": [ "Large Language Models", "Skill Acquisition", "Quality Diversity" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 2, 1 ] }, { "paper_id": "KvaDHPhhir", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Haruto Yoshida", "gender": "Male", "institution": "Tohoku University", "country": "", "position": "MS student" }, { "name": "Itsumi Saito", "gender": "Female", "institution": "Tohoku University", "country": "", "position": "Associate Professor" }, { "name": "Keisuke Sakaguchi", "gender": "unknown", "institution": "Tohoku University", "country": "", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 39, "n_ref": 66, "n_ref_all": 100, "n_fig": 7, "n_tab": 12, "L_tab": 3442, "n_element_tab": 291, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1517, "n_element_tab_1": 156, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 164849, "len_all_1": 64942, "len_abs": 1513, "len_title": 116, "len_sents": 48347, "len_sents_1": 28178, "n_sents": 436, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1445, "title": "Sketch2Diagram: Generating Vector Diagrams from Hand-Drawn Sketches", "abs": "We address the challenge of automatically generating high-quality vector diagrams from hand-drawn sketches. \nVector diagrams are essential for communicating complex ideas across various fields, offering flexibility and scalability. \nWhile recent research has progressed in generating diagrams from text descriptions, converting hand-drawn sketches into vector diagrams remains largely unexplored, primarily due to the lack of suitable datasets. \nTo address this, we introduce SketikZ, a dataset containing 3,231 pairs of hand-drawn sketches, reference diagrams, and corresponding TikZ codes. \nOur evaluations highlight current limitations of state-of-the-art vision and language models (VLMs), establishing SketikZ as a key benchmark for future research in sketch-to-diagram conversion.\nAlong with SketikZ, we present ImgTikZ, an image-to-TikZ model that integrates a 6.7B parameter code-specialized open-source large language model (LLM) with a pre-trained vision encoder. \nDespite its modest size, ImgTikZ demonstrates performance comparable to more extensive models such as GPT-4o.\nThe model's success is largely driven by using our two data augmentation techniques and a multi-candidate inference strategy,\nsignificantly improving its performance.\nThese findings provide promising avenues for future research in sketch-to-diagram conversion and may have broader implications for image-to-code generation tasks. SketikZ is publicly available.", "keywords": [ "multimodal", "large language model", "diagram", "vector graphics" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "KvFk356RpR", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Hao Xuan", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "PhD student" }, { "name": "Xingyu Li", "gender": "unknown", "institution": "University of Alberta", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 95, "n_ref_all": 106, "n_fig": 2, "n_tab": 5, "L_tab": 1255, "n_element_tab": 289, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1097, "n_element_tab_1": 18, "formula_len_all": 1187, "formula_len_all_1": 1190, "len_all": 116419, "len_all_1": 57498, "len_abs": 1329, "len_title": 80, "len_sents": 32590, "len_sents_1": 27161, "n_sents": 215, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1365, "title": "Unlearning Mapping Attack: Exposing Hidden Vulnerabilities in Machine Unlearning", "abs": "As machine learning becomes increasingly data-dependent, concerns over privacy and content regulation among data owners have intensified. Machine Unlearning has emerged as a promising solution, allowing for the removal of specific data from pre-trained systems to protect user privacy and regulate information. Existing research on Machine Unlearning has shown considerable success in eliminating the influence of certain data while preserving model performance. However, the resilience of Machine Unlearning to malicious attacks has not been thoroughly examined. In this paper, we investigate the hidden vulnerabilities within current Machine Unlearning techniques. We propose a novel adversarial attack, the Unlearning Mapping Attack (UMA), capable of undermining the unlearning process without altering its procedures. Through experiments on both generative and discriminative tasks, we demonstrate the susceptibility of existing unlearning techniques to UMA. These findings highlight the need to reassess unlearning objectives across various tasks, prompting the introduction of a Robust Unlearning standard that prioritizes protection against adversarial threats. Our extensive studies show the successful adaptation of current unlearning methods to this robust framework. The Python implementation will be made publicly available upon acceptance of the paper.", "keywords": [ "Machine Unlearning", "Deep Learning", "Machine Learning Security" ], "rating_list": [ 5, 5, 6, 3, 5 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 3, 4, 3 ] }, { "paper_id": "Ku1tUKnAnC", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Adam Davies", "gender": "Non-Binary", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Chirag Rastogi", "gender": "Male", "institution": "Qualcomm", "country": "US", "position": "Researcher" }, { "name": "Julia Hockenmaier", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Full Professor" }, { "name": "Marc Canby", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 10, "n_ref_uni": 22, "n_ref": 91, "n_ref_all": 116, "n_fig": 10, "n_tab": 6, "L_tab": 1279, "n_element_tab": 113, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 713, "n_element_tab_1": 113, "formula_len_all": 885, "formula_len_all_1": 978, "len_all": 135565, "len_all_1": 66304, "len_abs": 3007, "len_title": 168, "len_sents": 52358, "len_sents_1": 32635, "n_sents": 275, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 119, "L_abs": 1384, "title": "Measuring the Reliability of Causal Probing Methods: Tradeoffs, Limitations, and the Plight of Nullifying Interventions", "abs": "Causal probing aims to analyze large language models (or other foundation models) by examining how modifying their representation of various latent properties using interventions derived from probing classifiers impacts their outputs. Recent works have cast doubt on the theoretical basis of several leading causal probing intervention methods, but it has been unclear how to systematically evaluate the effectiveness of probing interventions in practice. To address this, we formally define and quantify two key causal probing desiderata: completeness (how thoroughly the representation of the target property has been transformed) and selectivity (how little other properties have been impacted). We introduce an empirical analysis framework to measure and evaluate completeness and selectivity, allowing us to make the first direct comparisons of the reliability of different families of causal probing methods (e.g., linear vs. nonlinear or counterfactual vs. nullifying interventions). Our experimental analysis shows that: (1) there is an inherent tradeoff between completeness and selectivity, (2) no leading probing method is able to consistently satisfy both criteria at once, and (3) across the board, nullifying interventions are far less complete than counterfactual interventions, which suggests that nullifying methods may not be an effective approach to causal probing.", "keywords": [ "interpretability", "probing", "causal probing", "interventions", "mechanistic interpretability", "language models" ], "rating_list": [ 6, 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 1, 2 ], "presentation_list": [ 2, 4, 2, 3, 4 ], "contribution_list": [ 3, 3, 2, 1, 3 ], "confidence_list": [ 3, 3, 2, 3, 2 ] }, { "paper_id": "KtqZrNjvjd", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Aditya Agrawal", "gender": "Male", "institution": "Birla Institute of Tchnology and Science - KK Birla Goa Campus", "country": "", "position": "Undergrad student" }, { "name": "Ankit Gupta", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Louis-Philippe Morency", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Paul Pu Liang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Ruslan Salakhutdinov", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Willis Guo", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Martin Ma", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 41, "n_ref": 78, "n_ref_all": 99, "n_fig": 6, "n_tab": 9, "L_tab": 2481, "n_element_tab": 279, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 186, "n_element_tab_1": 37, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 155585, "len_all_1": 66130, "len_abs": 1216, "len_title": 155, "len_sents": 50857, "len_sents_1": 32808, "n_sents": 413, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1172, "title": "Video Active Perception: Efficient Inference-Time Long-Form Video Understanding with Vision-Language Models", "abs": "Large vision-language models (VLMs) have advanced multimodal tasks such as video question answering (QA), yet they struggle with long-form videos due to the computational burden of processing excessive tokens. Inspired by active perception theory, which posits that models gain information by acquiring data that differ from their expectations, we introduce Video Active Perception (VAP), a training-free method to enhance long-form video QA using VLMs. Our approach treats key frame selection as data acquisition in active perception and leverages a lightweight text-conditioned video generation model to represent prior world knowledge. Empirically, VAP achieves state-of-the-art zero-shot results on long-form video QA datasets such as EgoSchema, NExT-QA, ActivityNet-QA and CLEVRER, achieving an increase of up to 5.6 X efficiency by frames per question over standard GPT-4o, Gemini 1.5 Pro, and LLaVA-OV. Moreover, VAP shows stronger reasoning abilities than previous methods and effectively selects key frames relevant to questions. These findings highlight the potential of leveraging active perception to improve efficiency and effectiveness of long-form video QA.", "keywords": [ "Video question answering", "Vision Language Model" ], "rating_list": [ 6, 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 5, 5, 4, 4, 5 ] }, { "paper_id": "KtqBAGO6eu", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Hao Liu", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Nanxu Gong", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Wangyang Ying", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Xinyuan Wang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Yanjie Fu", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Associate Professor" }, { "name": "Bai Haoyue", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 12, "n_ref_uni": 43, "n_ref": 85, "n_ref_all": 110, "n_fig": 7, "n_tab": 10, "L_tab": 3554, "n_element_tab": 293, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2594, "n_element_tab_1": 95, "formula_len_all": 2442, "formula_len_all_1": 886, "len_all": 162696, "len_all_1": 70698, "len_abs": 1667, "len_title": 100, "len_sents": 51530, "len_sents_1": 32586, "n_sents": 414, "n_sents_1": 236, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 1677, "title": "Privacy Preserving Generative Feature Transformation", "abs": "Data-Centric AI (DCAI) aims to use AI to get better data for better AI. Feature transformation, as one of the essential tasks of DCAI, can augment the data representation and has garnered significant attention. Existing methods have demonstrated state-of-the-art performance on advancing predictive tasks. However, these methods can lead to serious privacy leakage. For example, sensitive features in original data can be inferred by models trained on transformed data, exposing vulnerabilities in the privacy-preserving capabilities of these methods. To address this issue, we introduce a privacy-preserving feature transformation framework that transforms data representation while preserving privacy from a generative modeling perspective. Specifically, our framework includes two phases: 1) privacy-aware knowledge acquisition and 2) privacy-preserving feature space generation. In the knowledge acquisition phase, we develop an information bottlenecks guided reinforcement learning system to explore and collect privacy-aware feature sets as a knowledge base in token sequence form. In the feature space generation phase, we develop a generative model to encode the knowledge base into a privacy-aware latent space, where the best latent representation is identified and decoded into the optimal privacy-preserving feature space. We solve the optimization via projected gradient ascent that maximizes predictive performance and minimizes privacy exposure. Finally, we present extensive experiments on eight real-world datasets to evaluate how our method can navigate both performance and privacy. The code is available at https://anonymous.4open.science/r/anonymous-2B53/.", "keywords": [ "Data-Centric AI", "Privacy-Preserving", "Generative Model" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 2, 2 ] }, { "paper_id": "KstDMYkfj4", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Da Li", "gender": "Male", "institution": "Samsung AI Center, Cambridge", "country": "GB", "position": "Sr. Research Scientist" }, { "name": "Henry Gouk", "gender": "Male", "institution": "University of Edinburgh", "country": "GB", "position": "Lecturer (Assistant Professor)" }, { "name": "Ondrej Bohdal", "gender": "Male", "institution": "Samsung Research", "country": "GB", "position": "Researcher" }, { "name": "Timothy Hospedales", "gender": "Male", "institution": "University of Edinburgh", "country": "", "position": "Full Professor" } ], "n_formula": 77, "n_formula_1": 31, "n_ref_uni": 22, "n_ref": 42, "n_ref_all": 49, "n_fig": 3, "n_tab": 2, "L_tab": 205, "n_element_tab": 19, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 763, "n_element_tab_1": 22, "formula_len_all": 6791, "formula_len_all_1": 1594, "len_all": 124633, "len_all_1": 61892, "len_abs": 1425, "len_title": 115, "len_sents": 44528, "len_sents_1": 29391, "n_sents": 372, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1433, "title": "On the Limitations of General Purpose Domain Generalisation Methods", "abs": "The Domain Generalisation (DG) problem setting requires a model trained on a set of data distributions (domains) to generalise to new distributions. Despite a huge amount of empirical study, previous DG methods fail to substantially outperform empirical risk minimisation on rigorous DG benchmarks. Motivated by this, we analyse the DG problem from a learning theoretic perspective and *characterise in which situations DG will succeed or fail*. Specifically, we derive upper bounds on the excess risk of ERM and lower bounds on the minimax excess risk, for three settings with different restrictions on how the domains may differ. In the most unconstrained setting, we show that all learning algorithms converge slowly with respect to number of training domains, potentially explaining the lack of algorithmic progress in this area. We also consider constrained settings including limiting the pairwise domain distances as measured by a broad class of integral probability metrics, and constraining all domains to have the same underlying support. In these constrained cases, DG algorithms can converge more rapidly. Notably, for all three settings, the we demonstrate that ERM has an optimal rate of convergence towards the best possible model. Our analysis guides practitioners interested in knowing when cross-domain generalisation might be reliable, and suggests strategies for optimising the performance of ERM in each setting.", "keywords": [ "Domain Generalisation", "Excess Risk", "Empirical Risk Minimisation", "Minimax", "Rademacher Complexity" ], "rating_list": [ 3, 3, 5, 5, 3 ], "soundness_list": [ 2, 1, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 1, 5, 4, 3, 4 ] }, { "paper_id": "KscheKSYrh", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Bo Jiang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Boyi Zeng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Haoli Bai", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Jushi Kai", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yixuan Wang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhouhan Lin", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 20, "n_ref": 53, "n_ref_all": 81, "n_fig": 12, "n_tab": 6, "L_tab": 3085, "n_element_tab": 386, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1187, "n_element_tab_1": 140, "formula_len_all": 792, "formula_len_all_1": 758, "len_all": 113250, "len_all_1": 55473, "len_abs": 1556, "len_title": 135, "len_sents": 30189, "len_sents_1": 23602, "n_sents": 258, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1563, "title": "FreqKV: Frequency Domain Key-Value Compression for Efficient Context Window Extension", "abs": "Extending the context window in large language models (LLMs) is essential for applications involving long-form content generation. However, the quadratic complexity of self-attention and the linear increase in key-value (KV) cache memory requirements with respect to sequence length present significant challenges during fine-tuning and inference. Although LongLoRA achieves efficient fine-tuning by employing shifted sparse attention, inference remains inefficient due to the requirement for dense global attention.\nIn this work, we introduce a novel context extension method that optimizes both fine-tuning and inference efficiency. Our method exploits a key observation: in the frequency domain, the energy distribution of the KV cache is primarily concentrated in low-frequency components. By filtering out the high-frequency components, the KV cache can be effectively compressed with minimal information loss. Building on this insight, we propose an efficient compression technique, FreqKV, that iteratively reduces the increasing KV cache to a fixed size in the frequency domain, applicable to both fine-tuning and inference. With minimal fine-tuning, LLMs can learn to leverage the limited cache that is compressed in the frequency domain and extend the context window efficiently.\nFreqKV introduces no additional parameters or architectural modifications, ensuring compatibility with the original full attention post-training.\nExperiments on long context language modeling and understanding demonstrate the efficiency and efficacy of the proposed method.", "keywords": [ "Large Language Models", "KV Compression", "Context Extension" ], "rating_list": [ 6, 8, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 3, 4, 3, 3, 3 ], "contribution_list": [ 4, 4, 2, 2, 3 ], "confidence_list": [ 3, 5, 4, 4, 3 ] }, { "paper_id": "KsVlV2CRya", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Adji Bousso Dieng", "gender": "Female", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Anjian Li", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" }, { "name": "Ryne Beeson", "gender": "unknown", "institution": "Princeton University", "country": "", "position": "Assistant Professor" }, { "name": "Zihan Ding", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 45, "n_ref": 83, "n_ref_all": 114, "n_fig": 5, "n_tab": 1, "L_tab": 678, "n_element_tab": 78, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 678, "n_element_tab_1": 78, "formula_len_all": 740, "formula_len_all_1": 775, "len_all": 131103, "len_all_1": 68296, "len_abs": 1111, "len_title": 121, "len_sents": 39963, "len_sents_1": 30662, "n_sents": 319, "n_sents_1": 236, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1132, "title": "DiffuSolve: Diffusion-Based Solver for Non-Convex Trajectory Optimization", "abs": "Optimal trajectory design is computationally expensive for nonlinear and high-dimensional dynamical systems. The challenge arises from the non-convex nature of the optimization problem with multiple local optima, which usually requires a global search. Traditional numerical solvers struggle to find diverse solutions efficiently without appropriate initial guesses. In this paper, we introduce DiffuSolve, a general diffusion model-based solver for non-convex trajectory optimization. An expressive diffusion model is trained on pre-collected locally optimal solutions and efficiently samples initial guesses, which then warm-starts numerical solvers to fine-tune the feasibility and optimality. We also present DiffuSolve+, a novel constrained diffusion model with an additional loss in training that further reduces the problem constraint violations of diffusion samples. Experimental evaluations on three tasks verify the improved robustness, diversity, and a 2$\\times$ to 11$\\times$ increase in computational efficiency with our proposed method, which generalizes well to trajectory optimization problems of varying challenges.", "keywords": [ "Diffusion Model", "Nonconvex Optimization", "Trajectory Optimization", "Numerical Solver" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 3, 2, 4, 2 ] }, { "paper_id": "KrSaWQH1OA", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Danielle C. Maddix", "gender": "unknown", "institution": "AWS AI Labs", "country": "US", "position": "Applied Scientist" }, { "name": "Neil Ashton", "gender": "unknown", "institution": "Amazon Web Services", "country": "GB", "position": "Principal Researcher" }, { "name": "Parisa M. Shabestari", "gender": "unknown", "institution": "Amazon Web Services", "country": "", "position": "Data Scientist" }, { "name": "Peter Yu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Samuel Gundry", "gender": "Not Specified", "institution": "Amazon", "country": "AU", "position": "Researcher" } ], "n_formula": 42, "n_formula_1": 0, "n_ref_uni": 10, "n_ref": 49, "n_ref_all": 82, "n_fig": 12, "n_tab": 6, "L_tab": 2755, "n_element_tab": 183, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1326, "n_element_tab_1": 69, "formula_len_all": 1568, "formula_len_all_1": 0, "len_all": 210562, "len_all_1": 48996, "len_abs": 1292, "len_title": 161, "len_sents": 56380, "len_sents_1": 23125, "n_sents": 410, "n_sents_1": 137, "venue": "ICLR2025", "decision": "Reject", "L_title": 113, "L_abs": 1298, "title": "AhmedML: High-Fidelity Computational Fluid Dynamics Dataset for Incompressible, Low-Speed Bluff Body Aerodynamics", "abs": "The development of Machine Learning (ML) methods for Computational Fluid Dynamics (CFD) is currently limited by the lack of openly available training data. This paper presents a new open-source dataset comprising of high fidelity, scale-resolving CFD simulations of 500 geometric variations of the Ahmed Car Body - a simplified car-like shape that exhibits many of the flow topologies that are present on bluff bodies such as road vehicles. The dataset contains simulation results that exhibit a broad set of fundamental flow physics such as geometry and pressure-induced flow separation as well as 3D vortical structures. Each variation of the Ahmed car body were run using a high-fidelity, time-accurate, hybrid Reynolds-Averaged Navier-Stokes (RANS) - Large-Eddy Simulation (LES) turbulence modelling approach using the open-source CFD code OpenFOAM. The dataset contains boundary, volume, geometry, and time-averaged forces/moments in widely used open-source formats. In addition, the OpenFOAM case setup is provided so that others can reproduce or extend the dataset. This represents to the authors knowledge, the first open-source large-scale dataset using high-fidelity CFD methods for the widely used Ahmed car body that is available to freely download with a permissive license (CC-BY-SA).", "keywords": [ "CFD", "aerodynamics", "dataset", "automotive", "ML" ], "rating_list": [ 6, 3, 3, 3, 5 ], "soundness_list": [ 3, 1, 3, 3, 3 ], "presentation_list": [ 3, 1, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 1, 2 ], "confidence_list": [ 5, 3, 5, 4, 4 ] }, { "paper_id": "KrK6zXbjfO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chieh-Hsin Lai", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Researcher" }, { "name": "Dongjun Kim", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Koichi Saito", "gender": "Male", "institution": "Sony AI America", "country": "US", "position": "Researcher" }, { "name": "Takashi Shibuya", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Senior Manager" }, { "name": "Yuhta Takida", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Research Scientist" }, { "name": "Yuki Mitsufuji", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Lead Research Scientist" }, { "name": "Zhi Zhong", "gender": "Male", "institution": "Sony Group Corporation", "country": "", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 14, "n_ref_uni": 32, "n_ref": 141, "n_ref_all": 196, "n_fig": 13, "n_tab": 9, "L_tab": 4413, "n_element_tab": 577, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 2435, "n_element_tab_1": 263, "formula_len_all": 2022, "formula_len_all_1": 905, "len_all": 195056, "len_all_1": 67645, "len_abs": 3633, "len_title": 141, "len_sents": 55341, "len_sents_1": 28815, "n_sents": 402, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 2339, "title": "SoundCTM: Unifying Score-based and Consistency Models for Full-band Text-to-Sound Generation", "abs": "Sound content creation, essential for multimedia works such as video games and films, often involves extensive trial-and-error, enabling creators to semantically reflect their artistic ideas and inspirations, which evolve throughout the creation process, into the sound.\nRecent high-quality diffusion-based Text-to-Sound (T2S) generative models provide valuable tools for creators. However, these models often suffer from slow inference speeds, imposing an undesirable burden that hinders the trial-and-error process.\nWhile existing T2S distillation models address this limitation through $1$-step generation, the sample quality of $1$-step generation remains insufficient for production use.\nAdditionally, while multi-step sampling in those distillation models improves sample quality itself, the semantic content changes due to their lack of deterministic sampling capabilities.\nThus, developing a T2S generative model that allows creators to efficiently conduct trial-and-error while producing high-quality sound remains a key challenge.\nTo address these issues, we introduce Sound Consistency Trajectory Models (SoundCTM), which allow flexible transitions between high-quality $1$-step sound generation and superior sound quality through multi-step deterministic sampling. \nThis allows creators to efficiently conduct trial-and-error with $1$-step generation to semantically align samples with their intention, and subsequently refine sample quality with preserving semantic content through deterministic multi-step sampling.\nTo develop SoundCTM, we reframe the CTM training framework, originally proposed in computer vision, and introduce a novel feature distance using the teacher network for a distillation loss. \nAdditionally, while distilling classifier-free guided trajectories, we introduce a $\\nu$-sampling, a new algorithm that offers another source of quality improvement. For the $\\nu$-sampling, we simultaneously train both conditional and unconditional student models.\nFor production-level generation, we scale up our model to 1B trainable parameters, making SoundCTM-DiT-1B the first large-scale distillation model in the sound community to achieve both promising high-quality $1$-step and multi-step full-band (44.1kHz) generation.\nAudio samples are available at \\url{https://anonymus-soundctm.github.io/soundctm_iclr/}.", "keywords": [ "text-to-sound generation", "distillation models", "text-to-audio diffusion models", "generative models for sound" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 2, 5, 5 ] }, { "paper_id": "Kqm8jxOC4a", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Hong Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Jian Luo", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Jiang Yixuan", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "Undergrad student" }, { "name": "Jie Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xinyi Li", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "huanshuo dong", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "Undergrad student" } ], "n_formula": 60, "n_formula_1": 38, "n_ref_uni": 57, "n_ref": 89, "n_ref_all": 99, "n_fig": 1, "n_tab": 7, "L_tab": 1653, "n_element_tab": 200, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 270, "n_element_tab_1": 35, "formula_len_all": 3507, "formula_len_all_1": 2214, "len_all": 152725, "len_all_1": 61243, "len_abs": 1668, "len_title": 120, "len_sents": 43767, "len_sents_1": 26442, "n_sents": 392, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1696, "title": "SReNet: Spectral Refined Network for Solving Operator Eigenvalue Problem", "abs": "Solving operator eigenvalue problems helps analyze intrinsic data structures and relationships, yielding substantial influence on scientific research and engineering applications.\nRecently, novel approaches based on deep learning have been proposed to obtain eigenvalues and eigenfunctions from the given operator, which address the efficiency challenge arising from traditional numerical methods.\nHowever, when solving top-$L$ eigenvalues problems, these learning-based methods ignore the information that could be inherited from other known eigenvectors, thus resulting in a less-than-ideal performance.\nTo address the challenge, we propose the **S**pectral **Re**fined **Net**work (**SReNet**). \nOur novel approach incorporates the power method to approximate the top-$L$ eigenvalues and their corresponding eigenfunctions.\nTo effectively prevent convergence to previous eigenfunctions, we introduce the Deflation Projection that significantly improves the orthogonality of the computed eigenfunctions and enables more precise prediction of multiple eigenfunctions simultaneously. \nFurthermore, we develop the adaptive filtering method that dynamically leverages intermediate approximate eigenvalues to construct rational filters that filter out predicted eigenvalues, when predicting the successive eigenvalue of the given problem.\nDuring the iterative solving, the spectral transformation is performed based on the filter function, converting the original eigenvalue problem into an equivalent problem that is easier to converge.\nExtensive experiments demonstrate that our approach consistently outperforms existing learning-based methods, achieving state-of-the-art performance in accuracy.", "keywords": [ "AI for Science", "Operator Eigenvalue Problem", "Scientific Computing" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 1, 1, 2, 2 ], "contribution_list": [ 1, 1, 1, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "Kpjvm2mB0K", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Amit Chakrabarti", "gender": "unknown", "institution": "Dartmouth College", "country": "US", "position": "Professor" }, { "name": "David Woodruff", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Jeffrey Jiang", "gender": "unknown", "institution": "Dartmouth College", "country": "US", "position": "Undergrad student" }, { "name": "Taisuke Yasuda", "gender": "Male", "institution": "Voleon Group", "country": "US", "position": "Member of Research Staff" } ], "n_formula": 29, "n_formula_1": 14, "n_ref_uni": 53, "n_ref": 94, "n_ref_all": 109, "n_fig": 0, "n_tab": 1, "L_tab": 104, "n_element_tab": 5, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2687, "formula_len_all_1": 1194, "len_all": 171078, "len_all_1": 72962, "len_abs": 1926, "len_title": 53, "len_sents": 41549, "len_sents_1": 32139, "n_sents": 340, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 63, "L_abs": 2427, "title": "Streaming Algorithms For $\\ell_p$ Flows and $\\ell_p$ Regression", "abs": "We initiate the study of one-pass streaming algorithms for underdetermined $\\ell_p$ linear regression problems of the form\n $$\n \\min_{\\mathbf A\\mathbf x = \\mathbf b} \\lVert\\mathbf x\\rVert_p \\,, \\qquad \n \\text{where } \\mathbf A \\in \\mathbb R^{n \\times d} \\text{ with } n \\ll d \\,,\n $$\n which generalizes basis pursuit ($p = 1$) and least squares solutions to\n underdetermined linear systems ($p = 2$). We study the column-arrival\n streaming model, in which the columns of $\\mathbf A$ are presented one by one in a\n stream. When $\\mathbf A$ is the incidence matrix of a graph, this corresponds to an\n edge insertion graph stream, and the regression problem captures $\\ell_p$\n flows which includes transshipment ($p = 1$), electrical flows ($p = 2$), and\n max flow ($p = \\infty$) on undirected graphs as special cases. Our goal is to\n design algorithms which use space much less than the entire stream, which has\n a length of $d$.\n\n For the task of estimating the cost of the $\\ell_p$ regression problem for\n $p\\in[2,\\infty]$, we show a streaming algorithm which constructs a sparse\n instance supported on $\\tilde O(\\varepsilon^{-2}n)$ columns of $\\mathbf A$\n which approximates the cost up to a $(1\\pm\\varepsilon)$ factor, which\n corresponds to $\\tilde O(\\varepsilon^{-2}n^2)$ bits of space in general and\n an $\\tilde O(\\varepsilon^{-2}n)$ space semi-streaming algorithm for\n constructing $\\ell_p$ flow sparsifiers on graphs. This extends to $p\\in(1,\n 2)$ with $\\tilde O(\\varepsilon^{2}n^{q/2})$ columns, where $q$ is the H\\\"older\n conjugate exponent of $p$. For $p = 2$, we show that $\\Omega(n^2)$ bits of\n space are required in general even for outputting a constant factor\n solution. For $p = 1$, we show that the cost cannot be estimated even to an\n $o(\\sqrt n)$ factor in $\\mathrm{poly}(n)$ space.\n\n On the other hand, if we are interested in outputting a solution $\\mathbf\n x$, then we show that $(1+\\varepsilon)$-approximations require $\\Omega(d)$\n space for $p > 1$, and in general, $\\kappa$-approximations require\n $\\tilde\\Omega(d/\\kappa^{2q})$ space for $p > 1$. We complement these lower\n bounds with the first sublinear space upper bounds for this problem, showing\n that we can output a $\\kappa$-approximation using space only\n $\\mathrm{poly}(n) \\cdot \\tilde O(d/\\kappa^q)$ for $p > 1$, as well as a\n $\\sqrt n$-approximation using $\\mathrm{poly}(n, \\log d)$ space for $p = 1$.", "keywords": [ "Regression", "Streaming", "Online algorithms", "Flows" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 4, 4 ], "presentation_list": [ 3, 4, 4 ], "contribution_list": [ 3, 4, 3 ], "confidence_list": [ 4, 5, 2 ] }, { "paper_id": "KohdorhwHt", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Chenglu Sun", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Deyi Xue", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Peng Cheng", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Senwen Yu", "gender": "Not Specified", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Shuo Shen", "gender": "Male", "institution": "Tencent ", "country": "CN", "position": "Researcher" }, { "name": "Zixia Zhou", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Tao Wenzhi", "gender": "unknown", "institution": "Tencent", "country": "CN", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 49, "n_ref": 80, "n_ref_all": 124, "n_fig": 13, "n_tab": 10, "L_tab": 4667, "n_element_tab": 615, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 708, "n_element_tab_1": 34, "formula_len_all": 379, "formula_len_all_1": 379, "len_all": 230369, "len_all_1": 65203, "len_abs": 1139, "len_title": 117, "len_sents": 73029, "len_sents_1": 31971, "n_sents": 602, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1156, "title": "Language-conditioned Multi-Style Policies with Reinforcement Learning", "abs": "Recent studies have explored the application of large language models (LLMs) in language-conditioned reinforcement learning (LC-RL). These studies typically involve training RL agents to follow straightforward human instructions in domains such as object manipulation, navigation, or text-based environments. To extend these capabilities for following high-level and abstract language instructions with diverse style policies in complex environments, we propose a novel method called LCMSP, which can generate language-conditioned multi-style policies. LCMSP first trains a multi-style RL policy capable of achieving different meta-behaviors, which can be controlled by corresponding style parameters. Subsequently, LCMSP leverages the reasoning capabilities and common knowledge of LLMs to align language instructions with style parameters, thereby realizing language-controlled multi-style policies. Experiments conducted in various environments and with different types of instructions demonstrate that the proposed LCMSP is capable of understanding high-level abstract instructions and executing corresponding behavioral styles in complex environments.", "keywords": [ "language-conditioned reinforcement learning", "multi-style policy", "large language model", "policy control" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "KoQkr9eIUG", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Andreas Kolb", "gender": "Male", "institution": "Universit\u00e4t Siegen", "country": "", "position": "Full Professor" }, { "name": "Hendrik Sommerhoff", "gender": "Male", "institution": "University of Siegen", "country": "", "position": "PhD student" }, { "name": "Julia Grabinski", "gender": "Female", "institution": "Universit\u00e4t Mannheim", "country": "", "position": "PhD student" }, { "name": "Margret Keuper", "gender": "Female", "institution": "Universit\u00e4t Mannheim", "country": "DE", "position": "Full Professor" }, { "name": "Shashank Agnihotri", "gender": "Male", "institution": "Universit\u00e4t Mannheim", "country": "DE", "position": "PhD student" }, { "name": "Shashank Priyadarshi", "gender": "Male", "institution": "Universit\u00e4t des Saarlandes", "country": "DE", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 58, "n_ref": 113, "n_ref_all": 124, "n_fig": 2, "n_tab": 5, "L_tab": 2808, "n_element_tab": 330, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1943, "n_element_tab_1": 239, "formula_len_all": 184, "formula_len_all_1": 184, "len_all": 144597, "len_all_1": 60711, "len_abs": 1748, "len_title": 121, "len_sents": 34754, "len_sents_1": 28052, "n_sents": 259, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1821, "title": "Roll the dice: Monte Carlo Downsampling as a low-cost Adversarial Defence", "abs": "The well-known vulnerability of Neural Networks to adversarial attacks is concerning, more so with the increasing reliance on them for real-world applications like autonomous driving, medical imaging, and others.\nMultiple previous works have proposed defense methods against adversarial attacks, including adversarial training, adding random noise to images, frequency pooling, and others.\nWe observe from several such works, that there are two main paradigms for mitigating adversarial attacks.\nFirst, effective downsampling leads to learning better feature representations during training, thus improving the performance on attacked and non-attacked samples. \nHowever, these methods are expensive.\nSecond, perturbing samples with for example random noise helps in mitigating adversarial attacks as they stymie the flow of gradients to optimize the attacks. \nHowever, these methods lower the network's performance on non-attacked samples.\nThus, in this work, we combine the best of both strategies to propose a novel Monte-Carlo sampling-based approach for downsampling called Stochastic Downsampling.\nWe combine bi-linear interpolation with Monte Carlo integration for performing downsampling.\nThis helps us mitigate adversarial attacks while preserving the performance of non-attacked samples, thus increasing reliability.\nOur proposed Stochastic Downsampling operator can easily be integrated into any existing architecture, including adversarially pre-trained networks, with some finetuning.\nWe show the effectiveness of Stochastic Dowsampling over multiple image classification datasets using different network architectures with different training strategies.\nWe provide the code for performing Stochastic Downsampling here: Anonymous GitHub Repository (https://anonymous.4open.science/r/stochastic-downsampling/).", "keywords": [ "adversarial attacks", "adversarial defense", "stochasticity" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "KnoS9XxIlK", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Haodong Wen", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "Undergrad student" }, { "name": "Kaifeng Lyu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Kairong Luo", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Maosong Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Wenguang Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Zhenbo Sun", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhiyuan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "shengding hu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 78, "n_formula_1": 25, "n_ref_uni": 44, "n_ref": 80, "n_ref_all": 149, "n_fig": 20, "n_tab": 10, "L_tab": 3347, "n_element_tab": 261, "n_fig_1": 14, "n_tab_1": 2, "L_tab_1": 1592, "n_element_tab_1": 66, "formula_len_all": 6648, "formula_len_all_1": 1645, "len_all": 245018, "len_all_1": 74325, "len_abs": 1340, "len_title": 122, "len_sents": 81219, "len_sents_1": 32255, "n_sents": 754, "n_sents_1": 279, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1379, "title": "A Multi-Power Law for Loss Curve Prediction Across Learning Rate Schedules", "abs": "Training large models is both resource-intensive and time-consuming, making it crucial to understand the quantitative relationship between model performance and hyperparameters. In this paper, we derive an empirical law that predicts pretraining loss for large language models for every intermediate training step across various learning rate schedules, including constant, cosine, and step decay schedules. Our proposed law takes a multi-power form, combining a power law based on the sum of learning rates and additional power laws to account for a loss reduction effect as learning rate decays. We validate this law extensively on Llama-2 models of varying sizes and demonstrate that, after fitting on a few learning rate schedules, it accurately predicts the loss curves for unseen schedules of different shapes and horizons. Moreover, by minimizing the predicted final pretraining loss across learning rate schedules, we are able to find a schedule that outperforms the widely-used cosine learning rate schedule. Interestingly, this automatically discovered schedule bears some resemblance to the recently proposed Warmup-Stable-Decay (WSD) schedule (hu et al, 2024) but achieves a slightly lower final loss. We believe these results could offer valuable insights for understanding the dynamics of pretraining and for designing learning rate schedules to improve efficiency.", "keywords": [ "Large language model", "Learning rate scheduler", "Scaling Law", "Hyperparameter optimization" ], "rating_list": [ 6, 8, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 4, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4, 3 ] }, { "paper_id": "KnYsdgeCey", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Chi Zhang", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Manjie Xu", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Wei Liang", "gender": "Female", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yixin Zhu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "\u6b23\u6021 \u6768", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 55, "n_ref": 70, "n_ref_all": 94, "n_fig": 6, "n_tab": 5, "L_tab": 1178, "n_element_tab": 118, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 644, "n_element_tab_1": 85, "formula_len_all": 155, "formula_len_all_1": 45, "len_all": 176340, "len_all_1": 66041, "len_abs": 1755, "len_title": 95, "len_sents": 52291, "len_sents_1": 30286, "n_sents": 424, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1394, "title": "Learning to Plan with Personalized Preferences", "abs": "Understanding and adapting to human preferences is essential for the effective integration of artificial agents into daily human life, particularly as AI becomes increasingly involved in collaboration and assistance roles. Previous studies on preference recognition in embodied intelligence have largely adopted a generalized yet non-personalized approach. To fill in this gap, our research focuses on\nempowering embodied agents to learn and adapt to individual preferences, a task complicated by the challenges of inferring these preferences from minimal observations and requiring robust few-shot generalization. To facilitate future study, we introduce PbP, an embodied environment that supports hundreds of diverse preferences ranging from complex action sequences to specific sub-actions. Our\nexperiments on PbP reveal that while symbol-based approaches show promise in terms of effectiveness and scalability, accurately inferring implicit preferences and planning adaptive actions from limited data remain challenging. Nevertheless, preference serves as a valuable abstraction of human behaviors, and incorporating preference as a key intermediary step in planning can significantly enhance the personalization and adaptability of AI agents. We hope our findings can pave the way for future research on more efficient preference learning and personalized planning in dynamic environments.", "keywords": [ "Embodied AI", "Personalized Preference", "Human-AI interaction" ], "rating_list": [ 8, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "KmphHE92wU", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Cai Zhou", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Junru Zhou", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Muhan Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Pan Li", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Xiyuan Wang", "gender": "unknown", "institution": "Peking University", "country": "", "position": "PhD student" } ], "n_formula": 82, "n_formula_1": 23, "n_ref_uni": 69, "n_ref": 151, "n_ref_all": 188, "n_fig": 0, "n_tab": 7, "L_tab": 2783, "n_element_tab": 335, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 1418, "n_element_tab_1": 144, "formula_len_all": 11186, "formula_len_all_1": 2492, "len_all": 227719, "len_all_1": 70566, "len_abs": 2037, "len_title": 133, "len_sents": 54333, "len_sents_1": 31218, "n_sents": 468, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 2070, "title": "Towards Stable, Globally Expressive Graph Representations with Laplacian Eigenvectors", "abs": "Graph neural networks (GNNs) have achieved remarkable success in a variety of machine learning tasks over graph data. Existing GNNs usually rely on message passing, i.e., computing node representations by gathering information from the neighborhood, to build their underlying computational graphs. Such an approach has been shown fairly limited in expressive power, and often fails to capture global characteristics of graphs. To overcome the issue, a popular solution is to use Laplacian eigenvectors as additional node features, as they are known to contain global positional information of nodes, and can serve as extra node identifiers aiding GNNs to separate structurally similar nodes. Since eigenvectors naturally come with symmetries---namely, $O(p)$-group symmetry for every $p$ eigenvectors with equal eigenvalue, properly handling such symmetries is crucial for the stability and generalizability of Laplacian eigenvector augmented GNNs. However, using a naive $O(p)$-group invariant encoder for each $p$-dimensional eigenspace may not keep the full expressivity in the Laplacian eigenvectors. Moreover, computing such invariants inevitably entails a hard split of Laplacian eigenvalues according to their numerical identity, which suffers from great instability when the graph structure has small perturbations. In this paper, we propose a novel method exploiting Laplacian eigenvectors to generate *stable* and globally *expressive* graph representations. The main difference from previous works is that (i) our method utilizes **learnable** $O(p)$-invariant representations for each Laplacian eigenspace of dimension $p$, which are built upon powerful orthogonal group equivariant neural network layers already well studied in the literature, and that (ii) our method deals with numerically close eigenvalues in a **smooth** fashion, ensuring its better robustness against perturbations. Experiments on various graph learning benchmarks witness the competitive performance of our method, especially its great potential to learn global properties of graphs.", "keywords": [ "graph neural networks", "graph Laplacian eigenvectors" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "KmmNb7631I", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Chongjie Zhang", "gender": "unknown", "institution": "Washington University, Saint Louis", "country": "", "position": "Associate Professor" }, { "name": "Flood Sung", "gender": "Male", "institution": "Moonshot AI", "country": "CN", "position": "Researcher" }, { "name": "Jin Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Yang Gao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhilin Yang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 38, "n_ref": 63, "n_ref_all": 77, "n_fig": 5, "n_tab": 4, "L_tab": 421, "n_element_tab": 45, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 365, "n_element_tab_1": 41, "formula_len_all": 169, "formula_len_all_1": 133, "len_all": 136168, "len_all_1": 68384, "len_abs": 1554, "len_title": 145, "len_sents": 40921, "len_sents_1": 32790, "n_sents": 356, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1561, "title": "Learning to Plan Before Answering: Self-Teaching LLMs to Learn Abstract Plans for Problem Solving", "abs": "In the field of large language model (LLM) post-training, the effectiveness of utilizing synthetic data generated by the LLM itself has been well-presented. However, a key question remains unaddressed: what essential information should such self-generated data encapsulate? Existing approaches only produce step-by-step problem solutions, and fail to capture the abstract meta-knowledge necessary for generalization across similar problems. Drawing insights from cognitive science, where humans employ high-level abstraction to simplify complex problems before delving into specifics, we introduce a novel self-training algorithm: LEarning to Plan before Answering (LEPA). LEPA trains the LLM to formulate anticipatory plans, which serve as abstract meta-knowledge for problem-solving, before engaging with the intricacies of problems. This approach not only outlines the solution generation path but also shields the LLM from the distraction of irrelevant details. During data generation, LEPA first crafts an anticipatory plan based on the problem, and then generates a solution that aligns with both the plan and the problem. LEPA refines the plan through self-reflection, aiming to acquire plans that are instrumental in yielding correct solutions. During model optimization, the LLM is trained to predict both the refined plans and the corresponding solutions. By efficiently extracting and utilizing the anticipatory plans, LEPA demonstrates remarkable superiority over conventional algorithms on various challenging natural language reasoning benchmarks.", "keywords": [ "LLM", "self-training", "high-level abstraction", "self-reflection", "meta learning", "anticipatory plans" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "KmdwGYbMv0", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Nanne Van Noord", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Pascal Mettes", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Teng Long", "gender": "Male", "institution": "University of Amsterdam, University of Amsterdam", "country": "NL", "position": "Postdoc" } ], "n_formula": 64, "n_formula_1": 17, "n_ref_uni": 62, "n_ref": 128, "n_ref_all": 152, "n_fig": 4, "n_tab": 6, "L_tab": 3225, "n_element_tab": 140, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1070, "n_element_tab_1": 47, "formula_len_all": 4850, "formula_len_all_1": 1377, "len_all": 195200, "len_all_1": 59940, "len_abs": 1253, "len_title": 87, "len_sents": 48205, "len_sents_1": 25269, "n_sents": 401, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 28, "L_abs": 1267, "title": "Binary Hyperbolic Embeddings", "abs": "As datasets grow in size, vector-based search becomes increasingly challenging in terms of both storage and computational efficiency. Traditional solutions such as quantization techniques involve trade-offs between retrieval speed and accuracy, while hashing methods often require further optimization for binarization. In this work, we propose leveraging the compact nature of hyperbolic space for efficient search. Specifically, we introduce Binary Hyperbolic Embeddings, which transform complex hyperbolic similarity calculations into binary operations. We prove that these binary hyperbolic embeddings are retrieval-equivalent to their real-valued counterparts, ensuring minimal loss in retrieval quality. Our approach can be seamlessly integrated into FAISS to achieve improved memory efficiency and running speed while maintaining performance comparable to full-precision Euclidean embeddings. \\tblue{Notably, binary hyperbolic embeddings can also be combined with product quantization}. We demonstrate significant improvements in storage efficiency, with a natural byproduct of speeding up, with scaling potential to larger datasets. A portion of the code is included in the supplementary materials, and the full implementation will be made publicly available.", "keywords": [ "Hyperbolic", "Binary" ], "rating_list": [ 6, 6, 1, 5 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "KmQEsIfhr9", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Hanxun Huang", "gender": "Male", "institution": "The University of Melbourne", "country": "AU", "position": "Postdoc" }, { "name": "James Bailey", "gender": "unknown", "institution": "The University of Melbourne", "country": "AU", "position": "Full Professor" }, { "name": "Sarah Monazam Erfani", "gender": "unknown", "institution": "The University of Melbourne", "country": "AU", "position": "Associate Professor" }, { "name": "Xingjun Ma", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Ge Yi Li", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Postdoc" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 97, "n_ref": 223, "n_ref_all": 255, "n_fig": 11, "n_tab": 9, "L_tab": 4579, "n_element_tab": 405, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1775, "n_element_tab_1": 131, "formula_len_all": 822, "formula_len_all_1": 575, "len_all": 235593, "len_all_1": 78088, "len_abs": 1204, "len_title": 116, "len_sents": 62288, "len_sents_1": 36177, "n_sents": 500, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1208, "title": "Detecting Backdoor Samples in Contrastive Language Image Pretraining", "abs": "Contrastive language-image pretraining (CLIP) has been found to be vulnerable to poisoning backdoor attacks where the adversary can achieve an almost perfect attack success rate on CLIP models by poisoning only 0.01\\% of the training dataset. This raises security concerns on the current practice of pretraining large-scale models on unscrutinized web data using CLIP. In this work, we analyze the representations of backdoor-poisoned samples learned by CLIP models and find that they exhibit unique characteristics in their local subspace, i.e., their local neighborhoods are far more sparse than that of clean samples. Based on this finding, we conduct a systematic study on detecting CLIP backdoor attacks and show that these attacks can be easily and efficiently detected by traditional density ratio-based local outlier detectors, whereas existing backdoor sample detection methods fail. Our experiments also reveal that an unintentional backdoor already exists in the original CC3M dataset and has been trained into a popular open-source model released by OpenCLIP. Based on our detector, one can clean up a million-scale web dataset (e.g., CC3M) efficiently within 15 minutes using 4 Nvidia A100 GPUs.", "keywords": [ "Backdoor", "Detection", "CLIP" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "KlxK4ncqWZ", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Arthur Jacot", "gender": "Male", "institution": "NYU, New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Ingvar Ziemann", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Postdoc" }, { "name": "Nicholas Matthew Boffi", "gender": "Male", "institution": "NYU, New York University", "country": "", "position": "Instructor" }, { "name": "Stephen Tu", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" } ], "n_formula": 225, "n_formula_1": 31, "n_ref_uni": 55, "n_ref": 115, "n_ref_all": 126, "n_fig": 2, "n_tab": 2, "L_tab": 46, "n_element_tab": 10, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 22680, "formula_len_all_1": 2562, "len_all": 239488, "len_all_1": 76308, "len_abs": 4135, "len_title": 132, "len_sents": 63916, "len_sents_1": 33790, "n_sents": 698, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1131, "title": "Shallow diffusion networks provably learn hidden low-dimensional structure", "abs": "Diffusion-based generative models provide a powerful framework for learning to sample from a complex target distribution. \nThe remarkable empirical success of these models applied to high-dimensional signals, including images and video, stands in stark contrast to classical results highlighting the curse of dimensionality for distribution recovery. In this work, we take a step towards understanding this gap through a careful analysis of learning diffusion models over the Barron space of single hidden layer neural networks.\nIn particular, we show that these shallow models provably adapt to simple forms of low-dimensional structure, such as an unknown linear subspace or hidden independence, thereby avoiding the curse of dimensionality. We combine our results with recent analyses of sampling with diffusions to provide an end-to-end sample complexity bound for learning to sample from structured distributions. Importantly, our results do not require specialized architectures tailored to particular latent structures, and instead rely on the low-index structure of the Barron space to adapt to the underlying distribution.", "keywords": [ "Generative models", "diffusion models", "denoising score matching", "low dimensional structure", "Barron classes" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 1, 3 ] }, { "paper_id": "Klx0Rq9vbC", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "B. Aditya Prakash", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Haoxin Liu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zhiyuan Zhao", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 19, "n_ref_uni": 49, "n_ref": 110, "n_ref_all": 135, "n_fig": 7, "n_tab": 4, "L_tab": 1555, "n_element_tab": 65, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2942, "n_element_tab_1": 151, "formula_len_all": 1241, "formula_len_all_1": 1349, "len_all": 139374, "len_all_1": 70229, "len_abs": 1365, "len_title": 147, "len_sents": 38704, "len_sents_1": 30122, "n_sents": 259, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1370, "title": "Navigating Concept Drift and Temporal Shift: Distribution Shift Generalized Time-Series Forecasting", "abs": "Time-series forecasting finds broad applications in real-world scenarios. Due to the dynamic nature of time series data, it is crucial for time-series forecasting models to produce robust predictions under potential distribution shifts. In this paper, we initially identify two types of distribution shifts in time series: concept drift and temporal shift. We acknowledge that while existing studies primarily focus on addressing temporal shift issues in time series, designing proper concept drift methods for time series data received comparatively less attention.\n\nMotivated by the need to mitigate potential concept drift issues in time-series forecasting, this work proposes a novel soft attention mechanism that effectively leverages and ensemble information from the horizon time series. Furthermore, recognizing that both concept drift and temporal shift could occur concurrently in time-series forecasting scenarios while an integrated solution remains missing, this paper introduces ShifTS, a model-agnostic framework seamlessly addressing both concept drift and temporal shift issues in time-series forecasting. Extensive experiments demonstrate the efficacy of ShifTS in consistently enhancing the forecasting accuracy of agnostic models across multiple datasets, and consistently outperforming existing concept drift, temporal shift, and combined baselines.", "keywords": [ "Time-Series Forecasting", "Distribution Shift Generalization" ], "rating_list": [ 5, 6, 3, 6, 3 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 2, 4, 3, 4, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 5, 3, 2 ] }, { "paper_id": "KlalQu2423", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Boaz Carmeli", "gender": "Male", "institution": "Technion - Israel Institute of Technology", "country": "IL", "position": "PhD student" }, { "name": "Ron Meir", "gender": "Male", "institution": "Technion, Technion", "country": "", "position": "Full Professor" }, { "name": "Yonatan Belinkov", "gender": "Male", "institution": "Technion, Technion", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 4, "n_ref_uni": 72, "n_ref": 169, "n_ref_all": 227, "n_fig": 11, "n_tab": 14, "L_tab": 6547, "n_element_tab": 580, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1258, "n_element_tab_1": 117, "formula_len_all": 921, "formula_len_all_1": 223, "len_all": 267613, "len_all_1": 61518, "len_abs": 3274, "len_title": 112, "len_sents": 83111, "len_sents_1": 27130, "n_sents": 669, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 807, "title": "CtD: Composition through Decomposition in Emergent Communication", "abs": "Compositionality is a cognitive mechanism that allows humans to systematically combine known concepts in novel ways.\nThis study demonstrates how artificial neural agents acquire and utilize compositional generalization to describe previously unseen images.\nOur method, termed \\`\\`Composition through Decomposition'', involves two sequential training steps.\nIn the \\'Decompose\\' step, the agents learn to decompose an image into basic concepts using a codebook acquired during interaction in a multi-target coordination game.\nSubsequently, in the \\`Compose\\' step, the agents employ this codebook to describe novel images by composing basic concepts into complex phrases.\nRemarkably, we observe cases where generalization in the `Compose' step is achieved zero-shot, without the need for additional training.", "keywords": [ "Emergent communication", "Compositionality", "Codebook learning" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 2, 3 ] }, { "paper_id": "KlV5CkNQkl", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Ga Wu", "gender": "Male", "institution": "Dalhousie University", "country": "CA", "position": "Assistant Professor" }, { "name": "Hassan Sajjad", "gender": "Male", "institution": "Dalhousie University", "country": "CA", "position": "Associate Professor" }, { "name": "Mahtab Sarvmaili", "gender": "unknown", "institution": "Dalhousie University", "country": "CA", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 9, "n_ref_uni": 30, "n_ref": 57, "n_ref_all": 87, "n_fig": 15, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1677, "n_element_tab_1": 69, "formula_len_all": 1951, "formula_len_all_1": 958, "len_all": 124546, "len_all_1": 56463, "len_abs": 1160, "len_title": 116, "len_sents": 47358, "len_sents_1": 27581, "n_sents": 306, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1164, "title": "Data-centric Prediction Explanation via Kernelized Stein Discrepancy", "abs": "Existing example-based prediction explanation methods often bridge test and training data points through the model\u2019s parameters or latent representations. While these methods offer clues to the causes of model predictions, they often exhibit innate shortcomings, such as incurring significant computational overhead or producing coarse-grained explanations. This paper presents a Highly-precise and Data-centric Explanation (HD-Explain) prediction explanation method that exploits properties of Kernelized Stein Discrepancy (KSD). Specifically, the KSD uniquely defines a parameterized kernel function for a trained model that encodes model-dependent data correlation. By leveraging the kernel function, one can identify training samples that provide the best predictive support to a test point efficiently. We conducted thorough analyses and experiments across multiple classification domains, where we show that HD-Explain outperforms existing methods from various aspects, including 1) preciseness (fine-grained explanation), 2) consistency, and 3) computation efficiency, leading to a surprisingly simple, effective, and robust prediction explanation solution.", "keywords": [ "Prediction Explanation", "Kernelized Stein Discrepency" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 4, 3, 2 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "KlN00vQEY2", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Beatrice Achilli", "gender": "Female", "institution": "Bocconi University", "country": "IT", "position": "PhD student" }, { "name": "Carlo Lucibello", "gender": "Male", "institution": "Bocconi University", "country": "IT", "position": "Assistant Professor" }, { "name": "Enrico Ventura", "gender": "Male", "institution": "Bocconi University", "country": "IT", "position": "Postdoc" }, { "name": "Gianluigi Silvestri", "gender": "Male", "institution": "OnePlanet Research Center", "country": "NL", "position": "PhD student" }, { "name": "Luca Ambrogioni", "gender": "Male", "institution": "Radboud University Nijmegen", "country": "", "position": "Assistant Professor" } ], "n_formula": 72, "n_formula_1": 36, "n_ref_uni": 18, "n_ref": 40, "n_ref_all": 64, "n_fig": 11, "n_tab": 2, "L_tab": 2699, "n_element_tab": 34, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3707, "formula_len_all_1": 1190, "len_all": 109571, "len_all_1": 58836, "len_abs": 1253, "len_title": 90, "len_sents": 38130, "len_sents_1": 26257, "n_sents": 284, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1252, "title": "Manifolds, Random Matrices and Spectral Gaps: The geometric phases of generative diffusion", "abs": "In this paper, we investigate the latent geometry of generative diffusion models under the manifold hypothesis. To this purpose, we analyze the spectrum of eigenvalues (and singular values) of the Jacobian of the score function, whose discontinuities (gaps) reveal the presence and dimensionality of distinct sub-manifolds. Using a statistical physics approach, we derive the spectral distributions and formulas for the spectral gaps under several distributional assumptions and we compare these theoretical predictions with the spectra estimated from trained networks. Our analysis reveals the existence of three distinct qualitative phase during the generative process:a trivial phase; a manifold coverage phase where the diffusion process fits the distribution internal to the manifold; a consolidation phase where the score becomes orthogonal to the manifold and all particles are projected on the support of the data. This `division of labor' between different timescales provides an elegant explanation on why generative diffusion models are not affected by the manifold overfitting phenomenon that plagues likelihood-based models, since the internal distribution and the manifold geometry are produced at different time points during generation.", "keywords": [ "Generative diffusion models", "differential geometry", "spectral gaps", "random matrices", "generalization", "manifold" ], "rating_list": [ 5, 6, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 1, 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 4, 3 ] }, { "paper_id": "KkVV561IMb", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 9, "author_info_list": [ { "name": "Bailing Wang", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Bin Yu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Haotian Li", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Kai Wang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Lingzhi Wang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Richard Y. D. Xu", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Full Professor" }, { "name": "Rui Zhang", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yuliang Wei", "gender": "Not Specified", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "\u738b\u6709\u5a01", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 33, "n_formula_1": 18, "n_ref_uni": 60, "n_ref": 111, "n_ref_all": 143, "n_fig": 5, "n_tab": 12, "L_tab": 2289, "n_element_tab": 380, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1340, "n_element_tab_1": 273, "formula_len_all": 3594, "formula_len_all_1": 1507, "len_all": 163241, "len_all_1": 60318, "len_abs": 1088, "len_title": 112, "len_sents": 39505, "len_sents_1": 25535, "n_sents": 275, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1044, "title": "Deep Sparse Latent Feature Models for Knowledge Graph Completion", "abs": "Recent progress in knowledge graph completion (KGC) has focused on text-based approaches to address the challenges of large-scale knowledge graphs (KGs). Despite their achievements, these methods often overlook the intricate interconnections between entities, a key aspect of the underlying topological structure of a KG. Stochastic blockmodels (SBMs), particularly the latent feature relational model (LFRM), offer robust probabilistic frameworks that can dynamically capture latent community structures and enhance link prediction. In this paper, we introduce a novel framework of sparse latent feature models for KGC, optimized through a deep variational autoencoder (VAE). Our approach not only effectively completes missing triples but also provides clear interpretability of the latent structures, leveraging textual information. Comprehensive experiments on the WN18RR, FB15k-237, and Wikidata5M datasets show that our method significantly improves performance by revealing latent communities and producing interpretable representations.", "keywords": [ "Knowledge Graph Completion (KGC)", "Stochastic Blockmodels (SBMs)", "Variational Autoencoder (VAE)" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 3, 5 ] }, { "paper_id": "KkALFpRWSV", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Hanping Zhang", "gender": "Male", "institution": "Carleton University", "country": "CA", "position": "PhD student" }, { "name": "Yuhong Guo", "gender": "Female", "institution": "Carleton University", "country": "CA", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 27, "n_ref": 66, "n_ref_all": 75, "n_fig": 5, "n_tab": 3, "L_tab": 459, "n_element_tab": 65, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 413, "n_element_tab_1": 45, "formula_len_all": 489, "formula_len_all_1": 491, "len_all": 106475, "len_all_1": 63464, "len_abs": 996, "len_title": 106, "len_sents": 39442, "len_sents_1": 30971, "n_sents": 247, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1000, "title": "Skill-based Safe Reinforcement Learning with Risk Planning", "abs": "Safe Reinforcement Learning (Safe RL) aims to ensure safety when an RL agent conducts learning by interacting with real-world environments where improper actions can induce high costs or lead to severe consequences. In this paper, we propose a novel Safe Skill Planning (SSkP) approach to enhance effective safe RL by exploiting auxiliary offline demonstration data. SSkP involves a two-stage process. First, we employ PU learning to learn a skill risk predictor from the offline demonstration data. Then, based on the learned skill risk predictor, we develop a novel risk planning process to enhance online safe RL and learn a risk-averse safe policy efficiently through interactions with the online RL environment, while simultaneously adapting the skill risk predictor to the environment. We conduct experiments in several benchmark robotic simulation environments. The experimental results demonstrate that the proposed approach consistently outperforms previous state-of-the-art safe RL methods.", "keywords": [ "Safe Reinforcement Learning", "Skill-based", "Risk Planning" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 4, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "KjxZ4BdUdN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Chaoyang He", "gender": "Male", "institution": "FedML, Inc.", "country": "", "position": "CTO" }, { "name": "Amir Avestimehr", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "Full Professor" }, { "name": "Shanshan Han", "gender": "Female", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 34, "n_ref": 95, "n_ref_all": 113, "n_fig": 5, "n_tab": 5, "L_tab": 3279, "n_element_tab": 154, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3529, "n_element_tab_1": 145, "formula_len_all": 692, "formula_len_all_1": 336, "len_all": 132490, "len_all_1": 67981, "len_abs": 1063, "len_title": 124, "len_sents": 36693, "len_sents_1": 31967, "n_sents": 255, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1063, "title": "Bridging the Safety Gap: A Guardrail Pipeline for Trustworthy LLM Inferences", "abs": "We present Wildflare GuardRail, a guardrail pipeline designed to enhance the safety and reliability of Large Language Model (LLM) inferences. Wildflare GuardRail integrates four key functional modules, including SAFETY DETECTOR, GROUNDING, CUSTOMIZER, and REPAIRER, and addresses safety challenges across multiple dimensions of LLM inferences. Wildflare GuardRail incorporates an unsafe content detection model that identifies issues such as toxicity, bias, and prompt injection, a hallucination detection model that identifies hallucinated LLM outputs and simultaneously provides explanations for the hallucinations, and a fixing model that corrects LLM outputs based on these explanations. Additionally, Wildflare GuardRail employs GROUNDINGto enrich user queries with relevant context, and utilizes CUSTOMIZERto allow users to define flexible protocols for handling specific safety requirements. Our experiments demonstrate that Wildflare GuardRail enhances safety and robustness in LLM inferences, offering adaptable and scalable solutions for LLM inferences.", "keywords": [ "guardrail", "safety", "llm" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 1, 3, 2, 2 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "KjTh5C0z7Y", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Antoine Grosnit", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Researcher" }, { "name": "Haitham Bou Ammar", "gender": "Male", "institution": "Huawei R&D UK", "country": "GB", "position": "Principal Researcher" }, { "name": "Lei Chen", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Zhiyao Xie", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Dimitrios Tsaras", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 34, "n_ref": 48, "n_ref_all": 74, "n_fig": 7, "n_tab": 4, "L_tab": 1641, "n_element_tab": 97, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 116, "n_element_tab_1": 20, "formula_len_all": 901, "formula_len_all_1": 939, "len_all": 144541, "len_all_1": 71175, "len_abs": 1304, "len_title": 104, "len_sents": 45139, "len_sents_1": 36262, "n_sents": 307, "n_sents_1": 241, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1335, "title": "ShortCircuit: AlphaZero-Driven Generative Circuit Design", "abs": "Chip design relies heavily on generating Boolean circuits, such as AND-Inverter Graphs (AIGs), from functional descriptions like truth tables. This generation operation is a key process in logic synthesis, a primary chip design stage. While recent advances in deep learning have aimed to accelerate circuit design, these efforts have mostly focused on tasks other than synthesis, and traditional heuristic methods have plateaued. In this paper, we introduce ShortCircuit, a novel transformer-based architecture that leverages the structural properties of AIGs and performs efficient space exploration. Contrary to prior approaches attempting end-to-end generation of logic circuits using deep networks, ShortCircuit employs a two-phase process combining supervised with reinforcement learning to enhance generalization to unseen truth tables. We also propose an AlphaZero variant to handle the double exponentially large state space and the sparsity of the rewards, enabling the discovery of near-optimal designs. To evaluate the generative performance of our trained model , we extract 500 truth tables from a benchmark set of 20 real-world circuits. ShortCircuit successfully generates AIGs for 98% of the 8-input test truth tables, and outperforms the state-of-the-art logic synthesis tool, ABC, by 18.79% in terms\nof circuits size.", "keywords": [ "Graph Generation", "MCTS", "AlphaZero", "AIG", "Boolean Logic", "EDA" ], "rating_list": [ 5, 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 3, 4 ] }, { "paper_id": "KjBG4JNOc2", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Fan Zhou", "gender": "unknown", "institution": "Shanghai University of Finance and Economics", "country": "CN", "position": "Associate Professor" }, { "name": "Shuguang Yu", "gender": "Male", "institution": "Shanghai University of Finance and Economics", "country": "", "position": "PhD student" }, { "name": "Wenqian Xu", "gender": "Male", "institution": "Shanghai University of Finance and Economics", "country": "CN", "position": "MS student" }, { "name": "Xinyi Zhou", "gender": "Female", "institution": "Shanghai University of Finance and Economics", "country": "CN", "position": "MS student" }, { "name": "Xuechun Wang", "gender": "Female", "institution": "Shanghai University of Finance and Economics", "country": "", "position": "MS student" }, { "name": "Hongtu Zhu", "gender": "Male", "institution": "University of North Carolina, Chapel Hill", "country": "", "position": "Full Professor" } ], "n_formula": 34, "n_formula_1": 18, "n_ref_uni": 39, "n_ref": 66, "n_ref_all": 96, "n_fig": 12, "n_tab": 8, "L_tab": 2155, "n_element_tab": 285, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1202, "n_element_tab_1": 183, "formula_len_all": 3360, "formula_len_all_1": 2146, "len_all": 156342, "len_all_1": 67927, "len_abs": 891, "len_title": 106, "len_sents": 44304, "len_sents_1": 28856, "n_sents": 365, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 895, "title": "Enhancing Training Robustness through Influence Measure", "abs": "In the field of machine learning, the pursuit of accurate models is ongoing. A key aspect of improving prediction performance lies in identifying which data points in the training set should be excluded and which high-quality, potentially unlabeled data points outside the training set should be incorporated to improve the model's performance on unseen data. To accomplish this, an effective metric is needed to evaluate the contribution of each data point toward enhancing overall model performance. \nThis paper proposes the use of an influence measure as a metric to assess the impact of training data on test set performance. Additionally, we introduce a data selection method to optimize the training set as well as a dynamic active learning algorithm driven by the influence measure. The effectiveness of these methods is demonstrated through extensive simulations and real-world datasets.", "keywords": [ "local influence measure", "training robustness", "active learning" ], "rating_list": [ 5, 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3, 4 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 3, 2 ] }, { "paper_id": "Kioojohsuy", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Andrei Lupu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Ani Calinescu", "gender": "Not Specified", "institution": "Department of Computer Science, University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Jakob Nicolaus Foerster", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "", "position": "Associate Professor" }, { "name": "Jonathan Cook", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Matteo Gallici", "gender": "Male", "institution": "Universidad Polit\u00e9cnica de Cataluna", "country": "", "position": "PhD student" }, { "name": "Tin Dizdarevic", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "MS student" }, { "name": "Tobias Gessler", "gender": "unknown", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 24, "n_ref": 51, "n_ref_all": 80, "n_fig": 7, "n_tab": 13, "L_tab": 5019, "n_element_tab": 735, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 679, "n_element_tab_1": 89, "formula_len_all": 558, "formula_len_all_1": 382, "len_all": 192229, "len_all_1": 69078, "len_abs": 1661, "len_title": 86, "len_sents": 59015, "len_sents_1": 32423, "n_sents": 476, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 38, "L_abs": 1693, "title": "Ad-Hoc Human-AI Coordination Challenge", "abs": "Achieving seamless coordination between AI agents and humans is crucial for real-world applications, yet it remains a significant open challenge. Hanabi is an established, fully cooperative benchmark environment that involves imperfect information, limited communication, theory of mind, and the necessity for coordination among agents to achieve a shared goal. These characteristics, in principle, make Hanabi an ideal testbed for exploring human-AI coordination. However, one key issue is that evaluation with human partners is both expensive and difficult to reproduce. To address this, we first develop \\textit{human proxy agents} via a combination of behavioural cloning on a large-scale dataset of human game play and regularised reinforcement learning. These proxies serve as robust, cheap and reproducible human-like evaluation partners in our Ad-Hoc Human-AI Coordination Challenge (AH2AC2). To facilitate the exploration of methods that leverage \\textit{limited amounts} of human data, we introduce a data-limited challenge setting, using 1,000 games, which we open-source. Finally, we present baseline results for both two-player and three-player Hanabi scenarios. These include zero-shot coordination methods, which do not utilise any human data, and methods that make use of the available human data combined with reinforcement learning. To prevent overfitting and ensure fair evaluation, we introduce an evaluation protocol that involves us hosting the proxy agents rather than publicly releasing them, and a public leaderboard for tracking the progress of the community. We make our code available as an anonymous repository: \\url{https://anonymous.4open.science/r/ah2ac2-E451/}", "keywords": [ "multi-agent reinforcement learning", "reinforcement learning", "multi-agent systems", "human-ai coordination", "cooperative", "challenge paper" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "KijslFbfOL", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "Pei Zhang", "gender": "Female", "institution": "A*STAR", "country": "SG", "position": "Intern" }, { "name": "Shengju Yu", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Siwei Wang", "gender": "Male", "institution": "Intelligent Game and Decision Lab", "country": "CN", "position": "Assistant Professor" }, { "name": "Tiejun Li", "gender": "Male", "institution": "National University of Defense Technology", "country": "", "position": "Full Professor" }, { "name": "Xinwang Liu", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yi Zhang", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Yiu-ming Cheung", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Full Professor" }, { "name": "Zhibin Dong", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Naiyang Guan", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 87, "n_formula_1": 32, "n_ref_uni": 30, "n_ref": 62, "n_ref_all": 92, "n_fig": 9, "n_tab": 16, "L_tab": 13217, "n_element_tab": 1094, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 3729, "n_element_tab_1": 324, "formula_len_all": 7748, "formula_len_all_1": 3307, "len_all": 197174, "len_all_1": 63721, "len_abs": 1801, "len_title": 182, "len_sents": 38339, "len_sents_1": 23538, "n_sents": 325, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 133, "L_abs": 1838, "title": "Simple yet Effective Incomplete Multi-view Clustering: Similarity-level Imputation and Intra-view Hybrid-group Prototype Construction", "abs": "Most of incomplete multi-view clustering (IMVC) methods typically choose to ignore the missing samples and only utilize observed unpaired samples to construct bipartite similarity. Moreover, they employ a single quantity of prototypes to extract the information of $\\textbf{all}$ views. To eliminate these drawbacks, we present a simple yet effective IMVC approach, SIIHPC, in this work. It firstly transforms partial bipartition learning into original sample form by virtue of reconstruction concept to split out of observed similarity, and then loosens traditional non-negative constraints via regularizing samples to more freely characterize the similarity. Subsequently, it learns to recover the incomplete parts by utilizing the connection built between the similarity exclusive on respective view and the consensus graph shared for all views. On this foundation, it further introduces a group of hybrid prototype quantities for each individual view to flexibly extract the data features belonging to each view itself. Accordingly, the resulting graphs are with various scales and describe the overall similarity more comprehensively. It is worth mentioning that these all are optimized in one unified learning framework, \nwhich makes it possible for them to reciprocally promote. Then, to effectively solve the formulated optimization problem, we design an ingenious auxiliary function that is with theoretically proven monotonic-increasing properties. Finally, the clustering results are obtained by implementing spectral grouping action on the eigenvectors of stacked multi-scale consensus similarity. Numerous experimental results demonstrate that even under diverse missing proportions, the proposed SIIHPC is still able to provide a preferable clustering performance compared to multiple prominent IMVC methods.", "keywords": [ "incomplete multi-view clustering", "mulit-view clustering", "clustering" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "KiN7g8mf9N", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Giulio Biroli", "gender": "unknown", "institution": "Ecole Normale Superieure", "country": "", "position": "Full Professor" }, { "name": "Krunoslav Lehman Pavasovic", "gender": "Male", "institution": "Meta", "country": "FR", "position": "PhD student" }, { "name": "Levent Sagun", "gender": "Non-Binary", "institution": "Meta", "country": "", "position": "Research scientist" } ], "n_formula": 31, "n_formula_1": 18, "n_ref_uni": 76, "n_ref": 142, "n_ref_all": 183, "n_fig": 8, "n_tab": 15, "L_tab": 3666, "n_element_tab": 274, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 687, "n_element_tab_1": 48, "formula_len_all": 3444, "formula_len_all_1": 2008, "len_all": 225354, "len_all_1": 63496, "len_abs": 1254, "len_title": 114, "len_sents": 64419, "len_sents_1": 27477, "n_sents": 543, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1258, "title": "A Differentiable Rank-Based Objective for Better Feature Learning", "abs": "In this paper, we leverage existing statistical methods to better understand feature learning from data. We tackle this by modifying the model-free variable selection method, Feature Ordering by Conditional Independence (FOCI), which is introduced in Azadkia & Chatterjee (2021). While FOCI is based on a non-parametric coefficient of conditional dependence, we introduce its parametric, differentiable approximation. With this approximate coefficient of correlation, we present a new algorithm called difFOCI, which is applicable to a wider range of machine learning problems thanks to its differentiable nature and learnable parameters. We present difFOCI in three contexts: (1) as a variable selection method with baseline comparisons to FOCI, (2) as a trainable model parametrized with a neural network, and (3) as a generic, widely applicable neural network regularizer, one that improves feature learning with better management of spurious correlations. We evaluate difFOCI on increasingly complex problems ranging from basic variable selection in toy examples to saliency map comparisons in convolutional networks. We then show how difFOCI can be incorporated in the context of fairness to facilitate classifications without relying on sensitive data.", "keywords": [ "ranks", "statistics", "machine learning", "feature learning", "rank ordering correlation", "fairness" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "KiDGtYCPcv", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Angela Radulescu", "gender": "Non-Binary", "institution": "Icahn School of Medicine at Mount Sinai", "country": "", "position": "Assistant Professor" }, { "name": "Herbert Zheng Wu", "gender": "unknown", "institution": "Icahn School of Medicine at Mount Sinai", "country": "", "position": "Assistant Professor" }, { "name": "Yusi Chen", "gender": "Female", "institution": "University of Washington", "country": "", "position": "Postdoc" } ], "n_formula": 15, "n_formula_1": 12, "n_ref_uni": 35, "n_ref": 45, "n_ref_all": 88, "n_fig": 23, "n_tab": 2, "L_tab": 990, "n_element_tab": 41, "n_fig_1": 22, "n_tab_1": 1, "L_tab_1": 574, "n_element_tab_1": 20, "formula_len_all": 974, "formula_len_all_1": 933, "len_all": 121431, "len_all_1": 63056, "len_abs": 1876, "len_title": 146, "len_sents": 39179, "len_sents_1": 28511, "n_sents": 289, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1884, "title": "Unveiling the latent dynamics in social cognition with multi-agent inverse reinforcement learning", "abs": "Understanding the intentions and beliefs of others, a phenomenon known as \"theory of mind\", is a crucial element in social behavior. These beliefs and perceptions are inherently subjective and latent, making them often unobservable for investigation. Social interactions further complicate the matter, as multiple agents can engage in recursive reasoning about each other's strategies with increasing levels of cognitive hierarchy. While previous research has shown promise in understanding a single agent's belief of values through inverse reinforcement learning, extending this to model multiple agents remains an open challenge due to the computational complexity. In this work, we adopted a probabilistic recursive modeling of cognitive levels and joint value decomposition to achieve efficient multi-agent inverse reinforcement learning (MAIRL). We provided a numerical method to evaluate value decomposition errors in multi-agent tasks with discrete state and action spaces. To validate our method, we conducted simulations of a two-agent cooperative foraging task in a grid environment. Our algorithm revealed the ground truth goal-directed value function and effectively distinguished between level-0 and level-1 agents. When applied to human behavior in a cooperative hallway task, our method identified meaningful goal maps that evolved with task proficiency and an interaction map that is related to key states in the task without accessing to the task rules. Similarly, in a non-cooperative task performed by monkeys, we identified mutual predictions that correlated with the animals' social hierarchy, highlighting the behavioral relevance of the latent beliefs we uncovered. Together, our findings demonstrate that MAIRL offers a new framework for uncovering human or animal beliefs in social behavior, thereby illuminating previously opaque aspects of social cognition.", "keywords": [ "Inverse Reinforcement Learning; Social Behavioral Latents; Theory of Mind Inference" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "KhvBlzwehb", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Chi Chen", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Fuwen Luo", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Peng Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaolong Wang", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Yang Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Yuanchi Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Yurui Dong", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Yuzhuang Xu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Ziyue Wang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 26, "n_ref": 98, "n_ref_all": 158, "n_fig": 8, "n_tab": 9, "L_tab": 4629, "n_element_tab": 402, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1445, "n_element_tab_1": 114, "formula_len_all": 11, "formula_len_all_1": 0, "len_all": 225157, "len_all_1": 67605, "len_abs": 1554, "len_title": 132, "len_sents": 71121, "len_sents_1": 31944, "n_sents": 547, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1270, "title": "ActiView: Evaluating Active Perception Ability for Multimodal Large Language Models", "abs": "Active perception, a crucial human capability, involves setting a goal based on the current understanding of the environment and performing actions to achieve that goal. Despite significant efforts in evaluating Multimodal Large Language Models (MLLMs), active perception has been largely overlooked. To address this gap, we propose a novel benchmark named ActiView to evaluate active perception in MLLMs. Since comprehensively assessing active perception is challenging, we focus on a specialized form of Visual Question Answering (VQA) that eases the evaluation yet challenging for existing MLLMs. Given an image, we restrict the perceptual field of a model, requiring it to actively zoom or shift its perceptual field based on reasoning to answer the question successfully. We conduct extensive evaluation over 27 models, including proprietary and open-source models, and observe that the ability to read and comprehend multiple images simultaneously plays a significant role in enabling active perception. Results reveal a significant gap in the active perception capability of MLLMs, indicating that this area deserves more attention. We hope that our benchmark could help develop methods for MLLMs to understand multimodal inputs in more natural and holistic ways.", "keywords": [ "benchmark for MLLMs", "active perception", "visual comprehension and reasoning" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "Khv74gSZjg", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Paribesh Regmi", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Rui Li", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Kishan KC", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" } ], "n_formula": 45, "n_formula_1": 16, "n_ref_uni": 43, "n_ref": 87, "n_ref_all": 130, "n_fig": 8, "n_tab": 8, "L_tab": 3066, "n_element_tab": 341, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1240, "n_element_tab_1": 101, "formula_len_all": 4195, "formula_len_all_1": 799, "len_all": 166389, "len_all_1": 61325, "len_abs": 1163, "len_title": 106, "len_sents": 46403, "len_sents_1": 26470, "n_sents": 372, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1167, "title": "Bayesian Neighborhood Adaptation for Graph Neural Networks", "abs": "The neighborhood scope (i.e., number of hops) where graph neural networks (GNNs) aggregate information to characterize a node's statistical property is critical to GNNs' performance. Two-stage approaches, training and validating GNNs for every pre-specified neighborhood scope to search for the best setting, is a daunting and time-consuming task and tends to be biased due to the search space design. How to adaptively determine proper neighborhood scopes for the aggregation process for both homophilic and heterophilic graphs remains largely unexplored. We thus propose to model the GNNs' message-passing behavior on a graph as a stochastic process by treating the number of hops as a beta process. This Bayesian framework allows us to infer the most plausible neighborhood scope for messsage aggregation simultaneously with the optimization of GNN parameters. Our theoretical analysis show the scope inference improves the expressivity of GNN models. Experiments on benchmark homophilic and heterophilic datasets show that the proposed method is compatible with state-of-the-art GNN variants, improving their performance and providing well-calibrated predictions.", "keywords": [ "Graph Neural Network", "Bayesian Inference" ], "rating_list": [ 5, 1, 8 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "Kh5OS3oNlg", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Daniel McDuff", "gender": "Male", "institution": "Google", "country": "US", "position": "Principal Researcher" }, { "name": "Eric J Gonzalez", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Karan Ahuja", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Lee Payne", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Mar Gonzalez-Franco", "gender": "Female", "institution": "Google", "country": "", "position": "Senior Research Scientist + Manager" }, { "name": "Steven Abreu", "gender": "Male", "institution": "Intel", "country": "DE", "position": "Intern" }, { "name": "Tiffany D Do", "gender": "Female", "institution": "Drexel University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 61, "n_ref_all": 91, "n_fig": 8, "n_tab": 2, "L_tab": 706, "n_element_tab": 62, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 141, "n_element_tab_1": 9, "formula_len_all": 56, "formula_len_all_1": 0, "len_all": 166489, "len_all_1": 64231, "len_abs": 1349, "len_title": 209, "len_sents": 47078, "len_sents_1": 33688, "n_sents": 296, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1358, "title": "PARSE-Ego4D: Personal Action Recommendation Suggestions for Ego-Centric Videos", "abs": "Intelligent assistance involves not only understanding but also action. Existing ego-centric video datasets contain rich annotations of the videos, but not of actions that an intelligent assistant could perform in the moment. To address this gap, we release PARSE-Ego4D, a new set of personal action recommendation annotations for the Ego4D dataset. We take a multi-stage approach to generating and evaluating these annotations. First, we used a prompt-engineered large language model (LLM) to generate context-aware action suggestions and identified over 18,000 action suggestions. While these synthetic action suggestions are valuable, the inherent limitations of LLMs necessitate human evaluation. To ensure high-quality and user-centered recommendations, we conducted a large-scale human annotation study that provides grounding in human preferences for all of PARSE-Ego4D. We analyze the inter-rater agreement and evaluate subjective preferences of participants. Based on our synthetic dataset and complete human annotations, we propose several new tasks for action suggestions based on ego-centric videos. We encourage novel solutions that improve latency and energy requirements. The annotations in PARSE-Ego4D will support researchers and developers who are working on building action recommendation systems for augmented and virtual reality systems.", "keywords": [ "augmented reality", "virtual reality", "human annotation", "recommender systems", "perception", "large language models" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 2 ] }, { "paper_id": "KgiMUvJcwm", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Biaobin Jiang", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Kangfei Zhao", "gender": "Female", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Kun XIE", "gender": "Female", "institution": "The Chinese University of Hong Kong, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Ningqiao Huang", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Peilin Zhao", "gender": "unknown", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Sibo Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Wei Liu", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Xingyi Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" } ], "n_formula": 35, "n_formula_1": 30, "n_ref_uni": 45, "n_ref": 100, "n_ref_all": 113, "n_fig": 7, "n_tab": 4, "L_tab": 1374, "n_element_tab": 114, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 3090, "n_element_tab_1": 45, "formula_len_all": 2046, "formula_len_all_1": 1736, "len_all": 164675, "len_all_1": 58028, "len_abs": 1306, "len_title": 113, "len_sents": 39885, "len_sents_1": 24601, "n_sents": 284, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1309, "title": "IgSeek: Fast and Accurate Antibody Design via Structure Retrieval", "abs": "Recent advancements in protein design have leveraged diffusion models to generate structural scaffolds, followed by a process known as protein inverse folding, which involves sequence inference on these scaffolds. However, these methodologies face significant challenges when applied to hyper-variable structures such as antibody Complementarity-Determining Regions (CDRs), where sequence inference frequently results in non-functional sequences due to hallucinations. Distinguished from prevailing protein inverse folding approaches, this paper introduces IgSeek, a novel structure-retrieval framework that infers CDR sequences by retrieving similar structures from a natural antibody database. Specifically, IgSeek employs a simple yet effective multi-channel equivariant graph neural network to generate high-quality geometric representations of CDR backbone structures. Subsequently, it aligns sequences of structurally similar CDRs and utilizes structurally conserved sequence motifs to enhance inference accuracy. Our experiments demonstrate that IgSeek not only proves to be highly efficient in structural retrieval but also outperforms state-of-the-art approaches in sequence recovery for both antibodies and T-Cell Receptors, offering a new retrieval-based perspective for therapeutic protein design.", "keywords": [ "Antibody Design", "Structure Retrieval", "Equivariant" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "KgN0mo6pLo", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Bin Li", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Tonglin Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Xiangyang Xue", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "ZhiMeng Shen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" } ], "n_formula": 20, "n_formula_1": 6, "n_ref_uni": 42, "n_ref": 80, "n_ref_all": 110, "n_fig": 11, "n_tab": 12, "L_tab": 2891, "n_element_tab": 389, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1364, "n_element_tab_1": 109, "formula_len_all": 1753, "formula_len_all_1": 402, "len_all": 179224, "len_all_1": 63218, "len_abs": 1467, "len_title": 121, "len_sents": 60093, "len_sents_1": 30546, "n_sents": 446, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1473, "title": "Compositional Scene Modeling with An Object-Centric Diffusion Transformer", "abs": "Early object-centric learning methods adopt simple pixel mixture decoders to reconstruct images, which struggle with complex synthetic and real-world datasets. Recent object-centric learning methods focus on decoding object representations with complex decoders, such as autoregressive Transformers or diffusion models, to solve this problem. However, these methods feed all object representations together into the decoder to directly reconstruct the latent representation of the entire scene. Contrary to human intuition, this approach ultimately leads to weak interpretability. This paper combines the recent powerful diffusion model and composition module to propose a novel object-centric learning method called Compositional Scene Modeling with an Object-centric Diffusion Transformer (CODiT). By adopting a proposed compositional denoising decoder that can generate the mask of single objects and construct images compositionally, CODiT has stronger interpretability while still retaining the ability to handle complex scenes. We also illustrate the Classifier-Free Guidance explanation of CODiT. Experiments show how compositional structure helps control the generation process, allowing the model to generate images via single object representations and edit objects. In addition, we present CODiT performs strongly in various tasks including segmentation and reconstruction on both complex synthetic datasets and real-world datasets compared with similar methods.", "keywords": [ "Object-Centric Representation Learning", "Unsupervised Learning", "Compositional Scene Modeling", "Diffusion Models", "Generative Models" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "KgKN7F0PyQ", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "DohyungKim", "gender": "unknown", "institution": "Electronics and Telecommunications Research Institute", "country": "KR", "position": "Principal Researcher" }, { "name": "Hyobin Ong", "gender": "Female", "institution": "University of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Hyungmin Kim", "gender": "unknown", "institution": "University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Jae-Woo Choi", "gender": "Male", "institution": "Electronics and Telecommunications Research Institute", "country": "", "position": "Researcher" }, { "name": "Jaehong Kim", "gender": "Male", "institution": "ETRI", "country": "KR", "position": "Principle Researcher/Director" }, { "name": "Minsu Jang", "gender": "Male", "institution": "Electronics and Telecommunications Research Institute", "country": "KR", "position": "Principal Researcher" }, { "name": "Youngwoo Yoon", "gender": "Male", "institution": "Electronics and Telecommunications Research Institute", "country": "", "position": "Principal Researcher" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 21, "n_ref": 470, "n_ref_all": 516, "n_fig": 8, "n_tab": 13, "L_tab": 6852, "n_element_tab": 253, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 536, "n_element_tab_1": 56, "formula_len_all": 863, "formula_len_all_1": 448, "len_all": 298359, "len_all_1": 72633, "len_abs": 2745, "len_title": 134, "len_sents": 111043, "len_sents_1": 34390, "n_sents": 933, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1199, "title": "ReAcTree: Hierarchical Task Planning with Dynamic Tree Expansion using LLM Agent Nodes", "abs": "Recent advancements in task planning using large language models (LLMs) have made remarkable progress. However, most existing methods, such as ReAct, face limitations when handling complex, long-horizon tasks due to inefficiencies in processing entire tasks through a single sequential decision-making process. To address these challenges, we propose ReAcTree, a hierarchical task planning method that automatically decomposes complex tasks into manageable subgoals within a tree structure. This tree consists of control flow nodes, which manage the execution order of agent nodes, and agent nodes that reason, act, and expand nodes into subgoals to achieve their goals. To further enhance performance, we introduce memory systems: each agent node retrieves goal-specific, agent-level experiences from episodic memory to use as in-context examples, and all agent nodes share and recall information obtained during task execution via working memory. Experiments on the WAH-NL dataset demonstrate that ReAcTree consistently outperforms ReAct across various LLMs and model sizes. For example, when using Qwen2.5 72B, ReAcTree achieves a goal success rate of 63\\%, significantly surpassing ReAct's 24\\%.", "keywords": [ "Task planning", "large language model", "decision-making", "behavior tree", "hierarchical planning" ], "rating_list": [ 6, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "KgJwbsfN7G", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Baoyi An", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Bin Chen", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Jinpeng Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Shiyu Qin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Shu-Tao Xia", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Tao Dai", "gender": "Male", "institution": "Department of Software Engineering, Shenzhen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tianci Luo", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Undergrad student" }, { "name": "Yaowei Wang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Yimin Zhou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 71, "n_formula_1": 24, "n_ref_uni": 48, "n_ref": 166, "n_ref_all": 197, "n_fig": 14, "n_tab": 6, "L_tab": 1093, "n_element_tab": 99, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 851, "n_element_tab_1": 53, "formula_len_all": 50625, "formula_len_all_1": 10137, "len_all": 241800, "len_all_1": 73151, "len_abs": 1916, "len_title": 120, "len_sents": 60994, "len_sents_1": 23157, "n_sents": 317, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1468, "title": "MambaVC: Exploring Selective State Spaces for Learned Visual Compression", "abs": "Learned visual compression is an important and active task in multimedia. Existing approaches have explored various CNN- and Transformer-based designs to model content distribution and eliminate redundancy, where balancing efficacy (i.e., rate-distortion trade-off) and efficiency remains a challenge. Recently, state-space models (SSMs) have shown promise due to their long-range modeling capacity and efficiency. Inspired by this, we take the first step to explore SSMs for visual compression. We introduce MambaVC, a simple, strong and efficient compression network based on SSM. MambaVC develops a visual state space (VSS) block with a 2D selective scanning (2DSS) module as the nonlinear activation function after each downsampling, which helps to capture informative global contexts and enhances compression. On compression benchmark datasets, MambaVC achieves superior rate-distortion performance with lower computational and memory overheads. Specifically, it outperforms CNN and Transformer variants by 7.2% and 15.2% on Kodak, respectively, while reducing computation by 42% and 24%, and saving 12% and 71% of memory. MambaVC shows even greater improvements with high-resolution images, highlighting its potential and scalability in real-world applications. We also provide a comprehensive comparison of different network designs, underscoring MambaVC's advantages. Code is available at https://anonymous.4open.science/r/MambaVC-408 and will be open-sourced.", "keywords": [ "Selective State Spaces; Learned Visual Compression" ], "rating_list": [ 3, 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 3, 3, 5 ] }, { "paper_id": "Kg0KnPfiW2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Jack Cole", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "PhD student" }, { "name": "Ke Li", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "Associate Professor" }, { "name": "HENG YANG", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 62, "n_ref": 110, "n_ref_all": 145, "n_fig": 2, "n_tab": 12, "L_tab": 5922, "n_element_tab": 663, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 1830, "n_element_tab_1": 315, "formula_len_all": 930, "formula_len_all_1": 53, "len_all": 233122, "len_all_1": 72844, "len_abs": 1849, "len_title": 124, "len_sents": 65761, "len_sents_1": 34152, "n_sents": 456, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1857, "title": "Automating Large-scale In-silico Benchmarking for Genomic Foundation Models", "abs": "The advancements in artificial intelligence in recent years, such as Large Language Models (LLMs), have fueled expectations for breakthroughs in genomic foundation models (GFMs). The code of nature, hidden in diverse genomes since the very beginning of life\u2019s evolution, holds immense potential for impacting humans and ecosystems through genome modeling. Recent breakthroughs in GFMs, such as Evo, have attracted significant investment and attention to genomic modeling, as they address long-standing challenges and transform in-silico genomic studies into automated, reliable, and efficient paradigms. In the context of this flourishing era of consecutive technological revolutions in genomics, GFM studies face two major challenges: the lack of GFM benchmarking tools and the absence of open-source software for diverse genomics. These challenges hinder the rapid evolution of GFMs and their wide application in tasks such as understanding and synthesizing genomes, problems that have persisted for decades. To address these challenges, we introduce GFMBench, a framework dedicated to GFM-oriented benchmarking. GFMBench standardizes benchmark suites and automates benchmarking for a wide range of open-source GFMs. It integrates millions of genomic sequences across hundreds of genomic tasks from four large-scale benchmarks, democratizing GFMs for a wide range of in-silico genomic applications. Additionally, GFMBench is released as open-source software, offering user-friendly interfaces and diverse tutorials, applicable for AutoBench and complex tasks like RNA design and structure prediction. To facilitate further advancements in genome modeling, we have launched a public leaderboard showcasing the benchmark performance derived from AutoBench. GFMBench represents a step toward standardizing GFM benchmarking and democratizing GFM applications.", "keywords": [ "genomic benchmark", "DNA", "RNA", "foundation model" ], "rating_list": [ 1, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 4, 2, 1 ], "contribution_list": [ 1, 3, 1, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "KfkmwYQXWh", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Jie Guo", "gender": "Female", "institution": "Pengcheng Laboratory", "country": "CN", "position": "Assistant Professor" }, { "name": "Mingyang Su", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Ronghui Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiu Li", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Yachao Zhang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yebin Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Youliang Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yuxiang Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ziwei Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 31, "n_ref": 68, "n_ref_all": 86, "n_fig": 6, "n_tab": 13, "L_tab": 5583, "n_element_tab": 740, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1711, "n_element_tab_1": 292, "formula_len_all": 967, "formula_len_all_1": 812, "len_all": 162284, "len_all_1": 69779, "len_abs": 1097, "len_title": 121, "len_sents": 40834, "len_sents_1": 29587, "n_sents": 347, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1069, "title": "InterDance: Reactive 3D Dance Generation with Realistic Duet Interactions", "abs": "Humans can perform a variety of interactive motions, among which two-person dance is one of the most challenging interactions. However, in terms of computer motion generation, current work is still unable to generate high-quality interactive motion, especially in the field of duet dance. On the one hand, this is caused by the lack of large-scale high-quality datasets. On the other hand, it arises from the incomplete representation of interactive motion and the lack of fine-grained optimization of interactions. To address these challenges, we propose a duet dance dataset that significantly enhances motion quality, data scale, and the variety of dance genres. Based on this dataset, we propose a new motion representation that can accurately and comprehensively describe interactive motion. We further introduce a diffusion-based algorithm with an interaction refine guidance strategy to optimize the realism of interactions progressively. Experiments demonstrate the effectiveness of our dataset and algorithm. Our project page is https://inter-dance.github.io/.", "keywords": [ "Motion Synthesis", "Human Interaction" ], "rating_list": [ 5, 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4, 5 ] }, { "paper_id": "KfeRfxTemB", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Thomas P Cannon", "gender": "Male", "institution": "University of Bath", "country": "GB", "position": "PhD student" }, { "name": "\u00d6zg\u00fcr \u015eim\u015fek", "gender": "Female", "institution": "University of Bath", "country": "GB", "position": "Full Professor" } ], "n_formula": 43, "n_formula_1": 9, "n_ref_uni": 40, "n_ref": 63, "n_ref_all": 99, "n_fig": 17, "n_tab": 10, "L_tab": 2612, "n_element_tab": 166, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2088, "formula_len_all_1": 581, "len_all": 199840, "len_all_1": 61068, "len_abs": 3596, "len_title": 118, "len_sents": 69077, "len_sents_1": 28814, "n_sents": 609, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 843, "title": "Accelerating Task Generalisation with Multi-Level Hierarchical Options", "abs": "Creating reinforcement learning agents that generalise effectively to new tasks is a key challenge in AI research. This paper introduces Fracture Cluster Options (FraCOs), a multi-level hierarchical reinforcement learning method that achieves state-of-the-art performance on difficult generalisation tasks. FraCOs identifies patterns in agent behaviour and forms options based on the expected future usefulness of those patterns, enabling rapid adaptation to new tasks. In tabular settings, FraCOs demonstrates effective transfer and improves performance as it grows in hierarchical depth. We evaluate FraCOs against state-of-the-art deep reinforcement learning algorithms in several complex procedurally generated environments. Our results show that FraCOs achieves higher in-distribution and out-of-distribution performance than competitors.", "keywords": [ "Reinforcement Learning", "Generalisation", "Hierarchical Reinforcement Learning" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "Keoih8ebp0", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 2, "author_info_list": [ { "name": "Chong Tang", "gender": "Male", "institution": "University of Southampton", "country": "GB", "position": "Postdoc" }, { "name": "J Ch", "gender": "unknown", "institution": "University of Southampton", "country": "GB", "position": "Lecturer" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 34, "n_ref": 46, "n_ref_all": 52, "n_fig": 4, "n_tab": 4, "L_tab": 1422, "n_element_tab": 187, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 822, "n_element_tab_1": 70, "formula_len_all": 250, "formula_len_all_1": 153, "len_all": 108991, "len_all_1": 60447, "len_abs": 1600, "len_title": 124, "len_sents": 31503, "len_sents_1": 30403, "n_sents": 207, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1605, "title": "BAP: BRANCH-AWARE PARALLEL EXECUTION FOR FASTER DNN INFERENCE ON MOBILE CPUS", "abs": "The growing demand for real-time applications on edge devices underscores the need for faster inference of complex deep neural network (DNN) models. Although mobile devices increasingly incorporate specialized processors like GPUs and TPUs, modern DNN models such as Whisper and Vision Transformers often involve dynamic control flows and tensor operations that are incompatible and unsupported on current frameworks with these mobile accelerators. CPU presents the most viable option to improve inference latency on mobile devices due to their widespread availability, substantial memory caches, and ability to support all types of tensor operations. However, existing CPU optimization techniques focus on sequential execution, overlooking potential parallelization within Automatic Speech Recognition (ASR) and transformer-based models, leading to inefficiencies. This work introduces a novel runtime model analysis pipeline that extracts layer and branch structures from DNN model graphs to identify parallelizable branches. We propose BAP, a branch-aware memory allocation strategy that isolates memory arenas for parallel branches, reducing contention and optimizing memory reuse within each branch. Additionally, we leverage CPU multithreading to execute these branches concurrently, optimizing thread management and memory access to minimize overhead. Evaluated on ASR models and transformer-based models, our approach reduces inference latency by up to 38.5%, decreases memory allocation requirements by up to 15.6x and saves up to 20.2% energy cost compared to the TFLite naive memory allocation.", "keywords": [ "Neural Networks", "Model Parallelism", "Edge Devices", "ASR Models", "Transformers", "Mobile CPUs" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 2, 4, 2, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "KeRwLLwZaw", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Filip Matzner", "gender": "Male", "institution": "Charles University Prague", "country": "CZ", "position": "PhD student" }, { "name": "Franti\u0161ek Mr\u00e1z", "gender": "unknown", "institution": "Czechia", "country": "CZ", "position": "Lecturer" } ], "n_formula": 6, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 68, "n_ref_all": 98, "n_fig": 13, "n_tab": 5, "L_tab": 2724, "n_element_tab": 125, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 40, "n_element_tab_1": 7, "formula_len_all": 748, "formula_len_all_1": 299, "len_all": 149739, "len_all_1": 61766, "len_abs": 1167, "len_title": 113, "len_sents": 51560, "len_sents_1": 30482, "n_sents": 396, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1183, "title": "Locally Connected Echo State Networks for Time Series Forecasting", "abs": "Echo State Networks (ESNs) are a class of recurrent neural networks in which only a small readout regression layer is trained, while the weights of the recurrent network, termed the reservoir, are randomly assigned and remain fixed.\nOur work introduces the Locally Connected ESN (LCESN), a novel ESN variant with a locally connected reservoir, forced memory, and a weight adaptation strategy.\nLCESN significantly reduces the asymptotic time and space complexities compared to the conventional ESN, enabling substantially larger networks.\nLCESN also improves the memory properties of ESNs without affecting network stability.\nWe evaluate LCESN's performance on the NARMA10 benchmark task and compare it to state-of-the-art models on nine real-world datasets.\nDespite the simplicity of our model and its one-shot training approach, LCESN achieves competitive results, even surpassing several state-of-the-art models.\nLCESN introduces a fresh approach to real-world time series forecasting and demonstrates that large, well-tuned random recurrent networks can rival complex gradient-trained feedforward models.\nWe provide our GPU-based implementation of LCESN as an open-source library.", "keywords": [ "Time Series Analysis", "Time Series Forecasting", "TSF", "Recurrent Neural Networks", "RNN", "Regression", "Echo State Networks", "ESN" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Ke2BEL4csm", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Anoop Jain", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Dong Wang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Associate Professor" }, { "name": "Huimin Zeng", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Xiaojie Wang", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Zhicheng Dou", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 35, "n_formula_1": 14, "n_ref_uni": 42, "n_ref": 91, "n_ref_all": 124, "n_fig": 3, "n_tab": 8, "L_tab": 5926, "n_element_tab": 447, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 31, "n_element_tab_1": 2, "formula_len_all": 2388, "formula_len_all_1": 1011, "len_all": 170264, "len_all_1": 59427, "len_abs": 1641, "len_title": 160, "len_sents": 50535, "len_sents_1": 28271, "n_sents": 396, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 112, "L_abs": 1648, "title": "A Non-Contrastive Learning Framework for Sequential Recommendation with Preference-Preserving Profile Generation", "abs": "Contrastive Learning (CL) proves to be effective for learning generalizable user representations in Sequential Recommendation (SR), but it suffers from high computational costs due to its reliance on negative samples. To overcome this limitation, we propose the first Non-Contrastive Learning (NCL) framework for SR, which eliminates computational overhead of identifying and generating negative\nsamples. However, without negative samples, it is challenging to learn uniform representations from only positive samples, which is prone to representation collapse. Furthermore, the alignment of the learned representations may be substantially compromised because existing ad-hoc augmentations can produce positive samples that have inconsistent user preferences. To tackle these challenges, we design a novel preference-preserving profile generation method to produce high-quality positive samples for non-contrastive training. Inspired by differential privacy, our approach creates augmented user profiles that exhibit high diversity while provably retaining consistent user preferences. With larger diversity and consistency of the positive samples, our NCL framework significantly enhances the alignment and uniformity of the learned representations, which contributes to better generalization. The experimental results on various benchmark datasets and model architectures demonstrate the effectiveness of the proposed method. Finally, our investigations reveal that both uniformity and alignment play a vital role in improving generalization for SR. Interestingly, in our data-sparse setting, alignment is usually more important than uniformity.", "keywords": [ "sequential recommendation", "non-contrastive learning" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "Kdcqzfypry", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Chao Huang", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Lianghao Xia", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Postdoc" } ], "n_formula": 13, "n_formula_1": 15, "n_ref_uni": 37, "n_ref": 55, "n_ref_all": 68, "n_fig": 6, "n_tab": 8, "L_tab": 2746, "n_element_tab": 219, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1390, "n_element_tab_1": 29, "formula_len_all": 861, "formula_len_all_1": 856, "len_all": 153350, "len_all_1": 72448, "len_abs": 2695, "len_title": 92, "len_sents": 53562, "len_sents_1": 34364, "n_sents": 428, "n_sents_1": 271, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 1959, "title": "AnyGraph: Graph Foundation Model in the Wild", "abs": "The growing ubiquity of relational data structured as graphs has underscored the need for graph learning models with exceptional generalization capabilities. However, current approaches often struggle to effectively extract generalizable insights, frequently requiring extensive fine-tuning and limiting their versatility. Graph foundation models offer a transformative solution, with the potential to learn robust, generalizable representations from graph data. This enables more effective and adaptable applications across a wide spectrum of tasks and domains. In this work, we investigate a unified graph model, AnyGraph, designed to handle key challenges: i) Structure Heterogenity. Addressing distribution shift in graph structural information; ii) Feature Heterogenity. Handling diverse feature representation spaces across graph datasets; iii) Fast Adaptation. Efficiently adapting the model to new graph domains; iv) Scaling Law Emergence. Enabling the model to exhibit scaling law behavior, where its performance scales favorably with the amount of data and parameter sizes. To tackle these critical challenges, we build the AnyGraph upon a Graph Mixture-of-Experts (MoE) architecture. This approach empowers the model to effectively manage both the in-domain and cross-domain distribution shift concerning structure-level and feature-level heterogeneity. Furthermore, a lightweight graph expert routing mechanism is proposed to facilitate AnyGraph's fast adaptability to new data and domains. Our extensive experiments on diverse 38 graph datasets have demonstrated the strong zero-shot learning performance of AnyGraph across diverse graph domains with significant distribution shift. Furthermore, we have validated the model's fast adaptation ability and scaling law emergence, showcasing its versatility. We have anonymously released our open-sourced AnyGraph implementation at the following link: https://anonymous.4open.science/r/AnyGraph-FECD.", "keywords": [ "graph learning", "graph neural networks" ], "rating_list": [ 3, 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 3, 1, 3, 1, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "KdR88Qskmw", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Hongren Yan", "gender": "Male", "institution": "Shanxi University", "country": "CN", "position": "PhD student" }, { "name": "Yuhua Qian", "gender": "Male", "institution": "Shanxi University", "country": "CN", "position": "Full Professor" } ], "n_formula": 32, "n_formula_1": 14, "n_ref_uni": 14, "n_ref": 30, "n_ref_all": 41, "n_fig": 4, "n_tab": 4, "L_tab": 744, "n_element_tab": 147, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 139, "n_element_tab_1": 20, "formula_len_all": 3253, "formula_len_all_1": 701, "len_all": 82342, "len_all_1": 51131, "len_abs": 755, "len_title": 85, "len_sents": 19377, "len_sents_1": 15331, "n_sents": 192, "n_sents_1": 144, "venue": "ICLR2025", "decision": "withdraw", "L_title": 41, "L_abs": 757, "title": "The Contraction Property of Pooling Layer", "abs": "Although the theory of deep neural networks has been studied for years, the mechanism of pooling layers is still elusive. In this paper, we report the angle contraction behavior of pooling strategies (the average pooling and max pooling) at initialization. Compared to the relu-activated fully connected layer or convolutional layer, the pooling layer stands as the main source of contraction of the angle between hidden features. Moreover, we show that the cosine similarity between average pooling features in convolutional neural network is more data-dependent than fully connected network, while the max pooling is not sensitive to the data distribution in both architectures. Our results may complement the understanding of the representation learning.", "keywords": [ "Average Pooling", "Max Pooling", "CNN", "Deep Learning" ], "rating_list": [ 3, 1, 5, 3 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 1, 3, 1 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "Kc3yoIL5oR", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Chen Gao", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Guozhen Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Huandong Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Xiaochen Wei", "gender": "Male", "institution": "China Academic of Electronics and Information Technology", "country": "CN", "position": "Researcher" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Zefang Zong", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 2, "n_ref_uni": 38, "n_ref": 92, "n_ref_all": 118, "n_fig": 9, "n_tab": 4, "L_tab": 4049, "n_element_tab": 759, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 2620, "n_element_tab_1": 473, "formula_len_all": 437, "formula_len_all_1": 197, "len_all": 163431, "len_all_1": 67369, "len_abs": 1469, "len_title": 120, "len_sents": 53078, "len_sents_1": 27893, "n_sents": 441, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1476, "title": "Solving Diverse Combinatorial Optimization Problems with a Unified Model", "abs": "Combinatorial Optimization (CO) encompasses a wide range of problems that arise in many real-world scenarios. While significant progress has been made in developing learning-based methods for specialized CO problems, a unified model with a single architecture and parameter set for diverse CO problems remains elusive. Such a model would offer substantial advantages in terms of efficiency and convenience. In this paper, we introduce and formalize a unified model for solving various CO problems. Inspired by the success of next-token prediction, we frame each problem-solving process as a Markov Decision Process (MDP), tokenize the corresponding sequential trajectory data, and train the model using a transformer backbone. To reduce token length in the trajectory data, we propose a CO-prefix design that aggregates static problem features. To address the heterogeneity of state and action tokens within the MDP, we employ a two-stage self-supervised learning approach. In this approach, a dynamic prediction model is first trained and then serves as a pre-trained model for subsequent policy generation. Experiments across nine CO problems demonstrate the generic problem-solving capability of our unified model, highlighting its few-shot and even zero-shot ability to generalize to unseen problems through rapid fine-tuning. We believe our framework offers a valuable complement to existing neural CO methods that focus on optimizing performance for individual problems.", "keywords": [ "Large Model", "Combinatorial Optimization", "Transformers" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "KbCh7zbw2K", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Wei-Hsiang Yu", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "", "position": "PhD student" }, { "name": "Yen-Yu Lin", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" }, { "name": "Yi-Hsuan Tsai", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 42, "n_ref": 88, "n_ref_all": 118, "n_fig": 12, "n_tab": 14, "L_tab": 3075, "n_element_tab": 373, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1496, "n_element_tab_1": 115, "formula_len_all": 607, "formula_len_all_1": 522, "len_all": 156753, "len_all_1": 65298, "len_abs": 1422, "len_title": 110, "len_sents": 44897, "len_sents_1": 30907, "n_sents": 328, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1425, "title": "Ranking-aware adapter for text-driven image ordering with CLIP", "abs": "Recent advances in vision-language models (VLMs) have made significant progress in downstream tasks that require quantitative concepts such as facial age estimation and image quality assessment, enabling VLMs to explore applications like image ranking and retrieval. However, existing studies typically focus on the reasoning based on a single image and heavily depend on text prompting, limiting their ability to learn comprehensive understanding from multiple images. To address this, we propose an effective yet efficient approach that reframes the CLIP model into a learning-to-rank task and introduces a lightweight adapter to augment CLIP for text-guided image ranking. Specifically, our approach incorporates learnable prompts to adapt to new instructions for ranking purposes and an auxiliary branch with ranking-aware attention, leveraging text-conditioned visual differences for additional supervision in image ranking. Our ranking-aware adapter consistently outperforms fine-tuned CLIPs on various tasks and achieves competitive results compared to state-of-the-art models designed for specific tasks like facial age estimation and image quality assessment. Overall, our approach primarily focuses on ranking images with a single instruction, which provides a natural and generalized way of learning from visual differences across images, bypassing the need for extensive text prompts tailored to individual tasks.", "keywords": [ "Vision Language Models", "CLIP", "Learning-to-Rank" ], "rating_list": [ 8, 6, 8, 3 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 2, 3, 1 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "Kb9PnkWYNT", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Anna Khoreva", "gender": "Female", "institution": "Zalando SE", "country": "DE", "position": "Principal Researcher" }, { "name": "Dan Zhang", "gender": "unknown", "institution": "Robert Bosch GmbH, Bosch", "country": "DE", "position": "Research Scientist" }, { "name": "Margret Keuper", "gender": "Female", "institution": "Universit\u00e4t Mannheim", "country": "DE", "position": "Full Professor" }, { "name": "William H. Beluch", "gender": "unknown", "institution": "Robert Bosch GmbH, Bosch", "country": "", "position": "Research Engineer" }, { "name": "Yumeng Li", "gender": "Female", "institution": "Bosch Center for Artificial Intelligence", "country": "DE", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 5, "n_ref_uni": 38, "n_ref": 101, "n_ref_all": 147, "n_fig": 11, "n_tab": 2, "L_tab": 199, "n_element_tab": 47, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 101, "n_element_tab_1": 10, "formula_len_all": 503, "formula_len_all_1": 140, "len_all": 155948, "len_all_1": 60641, "len_abs": 1786, "len_title": 117, "len_sents": 53938, "len_sents_1": 30530, "n_sents": 386, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1520, "title": "VSTAR: Generative Temporal Nursing for Longer Dynamic Video Synthesis", "abs": "Despite tremendous progress in the field of text-to-video (T2V) synthesis, open-sourced T2V diffusion models struggle to generate longer videos with dynamically varying and evolving content. They tend to synthesize quasi-static videos, ignoring the necessary visual change-over-time implied in the text prompt. Meanwhile, scaling these models to enable longer, more dynamic video synthesis often remains computationally intractable. To tackle this challenge, we introduce the concept of Generative Temporal Nursing (GTN), where we adjust the generative process on the fly during inference to improve control over the temporal dynamics and enable generation of longer videos. We propose a method for GTN, dubbed VSTAR, which consists of two key ingredients: Video Synopsis Prompting (VSP) and Temporal Attention Regularization (TAR), the latter being our core contribution. Based on a systematic analysis, we discover that the temporal units in pretrained T2V models are crucial to control the video dynamics. Upon this finding, we propose a novel regularization technique to refine the temporal attention, enabling training-free longer video synthesis in a single inference pass. For prompts involving visual progression, we leverage LLMs to generate video synopsis - description of key visual states - based on the original prompt to provide better guidance along the temporal axis. We experimentally showcase the superiority of our method in synthesizing longer, visually appealing videos over open-sourced T2V models.", "keywords": [ "Text-to-Video", "Diffusion Models", "Temporal Attention" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "Kb1bIuGuax", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Andrea Pinto", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Randall Balestriero", "gender": "Male", "institution": "Brown University", "country": "US", "position": "Assistant Professor" }, { "name": "Tomer Galanti", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 62, "n_ref": 103, "n_ref_all": 126, "n_fig": 10, "n_tab": 1, "L_tab": 238, "n_element_tab": 30, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 238, "n_element_tab_1": 30, "formula_len_all": 999, "formula_len_all_1": 971, "len_all": 235688, "len_all_1": 61740, "len_abs": 983, "len_title": 79, "len_sents": 35219, "len_sents_1": 29248, "n_sents": 272, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 31, "L_abs": 983, "title": "The Fair Language Model Paradox", "abs": "Large Language Models (LLMs) are widely deployed in real-world applications, yet little is known about their training dynamics at the token level. Evaluation typically relies on aggregated training loss, measured at the batch level, which overlooks subtle per-token biases arising from (i) varying token-level dynamics and (ii) structural biases introduced by hyperparameters. While weight decay is commonly used to stabilize training, we reveal that it silently introduces performance biases detectable only at the token level. In fact, we empirically show across different dataset sizes, model architectures and sizes ranging from 270M to 3B parameters that as weight decay increases, low-frequency tokens are disproportionately depreciated. This is particularly concerning, as these neglected low-frequency tokens represent the vast majority of the token distribution in most languages, calling for novel regularization techniques that ensure fairness across all available tokens.", "keywords": [ "LLM", "Fairness", "WeightDecay", "Unbalanced" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "Kap9vaGKwF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Chejian Xu", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Dawn Song", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Freddy Lecue", "gender": "unknown", "institution": "INRIA", "country": "", "position": "Full Professor" }, { "name": "Jiawei Zhang", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Shuang Yang", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Research Scientist" }, { "name": "Yu Gai", "gender": "unknown", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 93, "n_ref_all": 126, "n_fig": 1, "n_tab": 15, "L_tab": 6569, "n_element_tab": 266, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1888, "n_element_tab_1": 171, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 257149, "len_all_1": 79958, "len_abs": 1966, "len_title": 81, "len_sents": 106987, "len_sents_1": 39618, "n_sents": 709, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1756, "title": "KnowHalu: Multi-Form Knowledge Enhanced Hallucination Detection", "abs": "As large language models (LLMs) become increasingly integral to a wide array of applications, ensuring the factual accuracy of their outputs and mitigating hallucinations is paramount. Current approaches, which primarily rely on self-consistency checks or post-hoc fact-checking, often fall short by disregarding the nuanced structure of queries and the diverse forms of contextual knowledge required for accurate response generation.\nTo address these shortcomings, we introduce KnowHalu (pronounced \u201cNo Halu\u201d), the first multi-form knowledge-based hallucination detection framework. We also introduce a new category of hallucinations, off-target hallucinations, which occur when responses are factually accurate but irrelevant or nonspecific to the query (e.g., answering \"What\u2019s the primary language in Barcelona?\" with \"European language\").\nIn particular, KnowHalu employs a rigorous two-phase process to detect hallucinations. In the first phase, it isolates off-target hallucinations by analyzing the semantic alignment between the response and the query. In the second phase, it conducts a novel multi-form knowledge-based fact-checking through a comprehensive pipeline of reasoning and query decomposition, knowledge retrieval, knowledge form optimization, judgment generation, and judgment aggregation.\nExtensive evaluations demonstrate that KnowHalu significantly surpasses state-of-the-art (SOTA) baselines across diverse tasks, achieving over 15% improvement in question answering (QA) and 6% in summarization tasks when applied to the same underlying LLM. These results underscore the effectiveness and versatility of KnowHalu, setting a new benchmark for hallucination detection and paving the way for safer and more reliable LLM applications.", "keywords": [ "Hallucination Detection; Large Language Model; Factual Checking; Multi-Form Knowledge" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "Kak2ZH5Itp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Chen Wang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Chengqing Zong", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Researcher" }, { "name": "Jiajun Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Junhong Wu", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Wen Yang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 51, "n_ref": 97, "n_ref_all": 119, "n_fig": 4, "n_tab": 10, "L_tab": 3080, "n_element_tab": 413, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1933, "n_element_tab_1": 300, "formula_len_all": 171, "formula_len_all_1": 336, "len_all": 178520, "len_all_1": 58154, "len_abs": 1272, "len_title": 115, "len_sents": 43110, "len_sents_1": 24492, "n_sents": 316, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1289, "title": "Language Imbalance Driven Rewarding for Multilingual Self-improving", "abs": "Large Language Models (LLMs) have achieved state-of-the-art performance across numerous tasks. However, these advancements have predominantly benefited \"first-class\" languages such as English and Chinese, leaving many other languages underrepresented. This imbalance, while limiting broader applications, generates a natural preference ranking between languages, offering an opportunity to bootstrap the multilingual capabilities of LLM in a self-improving manner. Thus, we propose $\\textit{Language Imbalance Driven Rewarding}$, where the inherent imbalance between dominant and non-dominant languages within LLMs is leveraged as a reward signal. Iterative DPO training demonstrates that this approach not only enhances LLM performance in non-dominant languages but also improves the dominant language's capacity, thereby yielding an iterative reward signal. Fine-tuning Meta-Llama-3-8B-Instruct over two iterations of this approach results in continuous improvements in multilingual performance across instruction-following and arithmetic reasoning tasks, evidenced by an average improvement of 7.46\\% win rate on the X-AlpacaEval leaderboard and 13.9\\% accuracy on the MGSM benchmark. This work serves as an initial exploration, paving the way for multilingual self-improvement of LLMs.", "keywords": [ "Large Language Model", "Self-Improving", "Multilinguality" ], "rating_list": [ 6, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "KadOFOsUpQ", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Jacob Steinhardt", "gender": "unknown", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" }, { "name": "Kayo Yin", "gender": "Female", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 17, "n_ref": 49, "n_ref_all": 90, "n_fig": 23, "n_tab": 2, "L_tab": 1543, "n_element_tab": 117, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 287, "n_element_tab_1": 27, "formula_len_all": 211, "formula_len_all_1": 185, "len_all": 118231, "len_all_1": 58775, "len_abs": 1253, "len_title": 53, "len_sents": 37007, "len_sents_1": 28400, "n_sents": 272, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1257, "title": "Which Attention Heads Matter for In-Context Learning?", "abs": "Large language models (LLMs) exhibit impressive in-context learning (ICL) capability, enabling them to generate relevant responses from a handful of task demonstrations in the prompt. \nPrior studies have suggested two different explanations for the mechanisms behind ICL:\ninduction heads that find and copy relevant tokens, and function vector (FV) heads whose activations compute a latent encoding of the ICL task.\nTo better understand which of the two distinct mechanisms drives ICL, we study induction heads and FV heads in 12 language models.\n\nOur study reveals that in all 12 models, few-shot ICL is driven primarily by FV heads: ablating FV heads decreases few-shot ICL accuracy significantly more than ablating induction heads, especially in larger models. We also find that FV and induction heads are connected: many FV heads\nstart as induction heads during training before transitioning to the FV mechanism. This leads us to speculate that induction heads facilitate the learning of the more complex FV mechanism for ICL. \nFinally, the prevalence of FV and induction heads varies with architecture, which questions strong versions of the\n\"universality\" hypothesis: findings from interpretability research are not always generalizable across models.", "keywords": [ "interpretability", "in-context learning", "large language models", "mechanistic interpretability", "induction heads" ], "rating_list": [ 6, 3, 3, 3, 8 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 4 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3, 3 ] }, { "paper_id": "KaYXsoCxV7", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "Chenhui Qiang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Longhui Wei", "gender": "Male", "institution": "Huawei echnologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Qi Tian", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "XIN HE", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xumeng Han", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Intern" }, { "name": "Yingfei Sun", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhenjun Han", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Zhiyang Dou", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Zipeng Wang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 27, "n_ref": 78, "n_ref_all": 93, "n_fig": 5, "n_tab": 6, "L_tab": 922, "n_element_tab": 183, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1098, "n_element_tab_1": 155, "formula_len_all": 224, "formula_len_all_1": 168, "len_all": 113233, "len_all_1": 64695, "len_abs": 1328, "len_title": 113, "len_sents": 35215, "len_sents_1": 29630, "n_sents": 272, "n_sents_1": 226, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1333, "title": "ViMoE: An Empirical Study of Designing Vision Mixture-of-Experts", "abs": "Mixture-of-Experts (MoE) models embody the divide-and-conquer concept and are a promising approach for increasing model capacity, demonstrating excellent scalability across multiple domains. In this paper, we integrate the MoE structure into the classic Vision Transformer (ViT), naming it ViMoE, and explore the potential of applying MoE to vision through a comprehensive study on image classification. However, we observe that the performance is sensitive to the configuration of MoE layers, making it challenging to obtain optimal results without careful design. The underlying cause is that inappropriate MoE layers lead to unreliable routing and hinder experts from effectively acquiring helpful knowledge. To address this, we introduce a shared expert to learn and capture common information, serving as an effective way to construct stable ViMoE. Furthermore, we demonstrate how to analyze expert routing behavior, revealing which MoE layers are capable of specializing in handling specific information and which are not. This provides guidance for retaining the critical layers while removing redundancies, thereby advancing ViMoE to be more efficient without sacrificing accuracy. We aspire for this work to offer new insights into the design of vision MoE models and provide valuable empirical guidance for future research.", "keywords": [ "Mixture-of-Experts", "Image Classification" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "KaV6jGCxvs", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Dianbing Xi", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jiayuan Lu", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jingsen Zhu", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Rui Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Ruiyuan Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yuan Liu", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Yuchi Huo", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhijian Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 69, "n_ref": 134, "n_ref_all": 152, "n_fig": 7, "n_tab": 3, "L_tab": 393, "n_element_tab": 34, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 251, "n_element_tab_1": 18, "formula_len_all": 611, "formula_len_all_1": 450, "len_all": 162857, "len_all_1": 57058, "len_abs": 1199, "len_title": 65, "len_sents": 33194, "len_sents_1": 27331, "n_sents": 240, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1211, "title": "PFAvatar: Avatar Reconstruction from Multiple In-the-wild Images", "abs": "In this work, we present \\textit{PFAvatar}, a new approach to avatar reconstruction and editing from multiple in-the-wild images with varying poses, unknown camera conditions, cropped views, and occlusions. Traditional methods often rely on full-body images captured with controlled avatar pose, camera settings, lighting, and background, while struggling to reconstruct under in-the-wild settings.To address this issue, we fuse the varying pose priors of avatars in in-the-wild images, thereby enabling precise control over avatar generation.Specifically, we first inject avatar features (pose, appearance) from input images using a Vision-Language Model (VLM) and ControlNet. Subsequently, we employ a pose-conditioned 3D-Consistent Score Distillation Sampling (3D-SDS), which enables reconstructing a high-quality 3D avatar. Additionally, we propose a Condition Prior Preservation Loss (CPPL) to mitigate the issues of language and control drift caused by fine-tuning VLM and ControlNet with few-shot data. Through comprehensive experiments and evaluation, we demonstrate the effectiveness of our method for reconstructing avatars from in-the-wild images, supporting further applications like avatar editing.", "keywords": [ "Avatar Reconstruction", "Text-to-Image Diffusion Model", "Image-based Modeling", "Text-guided 3D Generation", "Digital Human;" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 1, 2 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "KZu3xhPhke", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Christoph Studer", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Associate professor" }, { "name": "Sueda Taner", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Ziyi Wang", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" } ], "n_formula": 35, "n_formula_1": 32, "n_ref_uni": 46, "n_ref": 115, "n_ref_all": 167, "n_fig": 12, "n_tab": 2, "L_tab": 151, "n_element_tab": 20, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1443, "formula_len_all_1": 1488, "len_all": 182250, "len_all_1": 73841, "len_abs": 1058, "len_title": 47, "len_sents": 54241, "len_sents_1": 31313, "n_sents": 471, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 27, "L_abs": 1061, "title": "Cauchy-Schwarz Regularizers", "abs": "We introduce a novel class of regularization functions, called Cauchy\u2013Schwarz~(CS) regularizers, which can be designed to induce a wide range of properties in solution vectors of optimization problems. To demonstrate the versatility of CS regularizers, we derive concrete regularization functions that promote discrete-valued vectors, eigenvectors of a given matrix, and orthogonal matrices. The resulting CS regularizers are simple, differentiable, and can be free of spurious critical points, making them suitable for gradient-based solvers and large-scale optimization problems. In addition, CS regularizers automatically adapt to the appropriate scale, which is, for example, beneficial when discretizing weights of neural networks. To demonstrate the efficacy of CS regularizers, we provide results for solving underdetermined systems of linear equations and weight quantization in neural networks. Furthermore, we discuss specializations, variations, and generalizations, which lead to an even broader class of new and possibly more powerful regularizers.", "keywords": [ "Regularizer", "optimization", "discretization", "quantization" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "KZgo2YQbhc", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Jianfei Cai", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" }, { "name": "Shangyu Chen", "gender": "Male", "institution": "Monash University", "country": "", "position": "PhD student" }, { "name": "Zizheng Pan", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Dinh Phung", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" } ], "n_formula": 27, "n_formula_1": 13, "n_ref_uni": 40, "n_ref": 114, "n_ref_all": 135, "n_fig": 13, "n_tab": 5, "L_tab": 2637, "n_element_tab": 165, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 3574, "n_element_tab_1": 129, "formula_len_all": 4481, "formula_len_all_1": 877, "len_all": 159436, "len_all_1": 67071, "len_abs": 1373, "len_title": 120, "len_sents": 45579, "len_sents_1": 29161, "n_sents": 360, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 72, "L_abs": 1379, "title": "PaRa: Personalizing Text-to-Image Diffusion via Parameter Rank Reduction", "abs": "Personalizing a large-scale pretrained Text-to-Image (T2I) diffusion model is chal-\nlenging as it typically struggles to make an appropriate trade-off between its training\ndata distribution and the target distribution, i.e., learning a novel concept with only a\nfew target images to achieve personalization (aligning with the personalized target)\nwhile preserving text editability (aligning with diverse text prompts). In this paper,\nwe propose PaRa, an effective and efficient Parameter Rank Reduction approach\nfor T2I model personalization by explicitly controlling the rank of the diffusion\nmodel parameters to restrict its initial diverse generation space into a small and\nwell-balanced target space. Our design is motivated by the fact that taming a T2I\nmodel toward a novel concept such as a specific art style implies a small generation\nspace. To this end, by reducing the rank of model parameters during finetuning, we\ncan effectively constrain the space of the denoising sampling trajectories towards\nthe target. With comprehensive experiments, we show that PaRa achieves great\nadvantages over existing finetuning approaches on single/multi-subject generation\nas well as single-image editing. Notably, compared to the prevailing fine-tuning\ntechnique LoRA, PaRa achieves better parameter efficiency (2\u00d7 fewer learnable\nparameters) and much better target image alignment.", "keywords": [ "Text-to-Image diffusion model", "Diffusion model fine-tuning" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "KZWaxtzIRx", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Bin Cui", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Tao Zhang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Wenjing Luo", "gender": "Female", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yan Zhang", "gender": "unknown", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yanjun Shen", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yanzhao Qin", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yujing Qiao", "gender": "unknown", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zenan Zhou", "gender": "Male", "institution": "baichuan", "country": "CN", "position": "Researcher" }, { "name": "sunhaoze", "gender": "Male", "institution": "Baichuan inc", "country": "CN", "position": "Researcher" }, { "name": "weipeng chen", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhangtao", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Zhang wen tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 18, "n_ref": 80, "n_ref_all": 111, "n_fig": 8, "n_tab": 8, "L_tab": 4255, "n_element_tab": 404, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 2983, "n_element_tab_1": 285, "formula_len_all": 142, "formula_len_all_1": 142, "len_all": 202117, "len_all_1": 70176, "len_abs": 3018, "len_title": 108, "len_sents": 57587, "len_sents_1": 27937, "n_sents": 436, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 1361, "title": "SysBench: Can LLMs Follow System Message?", "abs": "Large Language Models (LLMs) have become instrumental across various applications, with the customization of these models to specific scenarios becoming increasingly critical. System message, a fundamental component of LLMs, is consist of carefully crafted instructions that guide the behavior of model to meet intended goals. Despite the recognized potential of system messages to optimize AI-driven solutions, there is a notable absence of a comprehensive benchmark for evaluating how well LLMs follow system messages. To fill this gap, we introduce SysBench, a benchmark that systematically analyzes system message following ability in terms of three limitations of existing LLMs: constraint violation, instruction misjudgement and multi-turn instability. Specifically, we manually construct evaluation dataset based on six prevalent types of constraints, including 500 tailor-designed system messages and multi-turn user conversations covering various interaction relationships. Additionally, we develop a comprehensive evaluation protocol to measure model performance. Finally, we conduct extensive evaluation across various existing LLMs, measuring their ability to follow specified constraints given in system messages. The results highlight both the strengths and weaknesses of existing models, offering key insights and directions for future research.", "keywords": [ "Large Language Models", "System Message", "Instruction Following", "Benchmark Dataset" ], "rating_list": [ 6, 3, 6 ], "soundness_list": [ 3, 2, 4 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "KZII3faAs2", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Magdalena Proszewska", "gender": "unknown", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Siddharth N", "gender": "Male", "institution": "University of Edinburgh", "country": "", "position": "Reader (Associate Professor)" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 36, "n_ref": 57, "n_ref_all": 74, "n_fig": 9, "n_tab": 6, "L_tab": 1923, "n_element_tab": 142, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1356, "n_element_tab_1": 130, "formula_len_all": 643, "formula_len_all_1": 454, "len_all": 108352, "len_all_1": 65195, "len_abs": 2046, "len_title": 81, "len_sents": 31736, "len_sents_1": 28951, "n_sents": 233, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 33, "L_abs": 2056, "title": "AIMing for Explainability in GNNs", "abs": "As machine learning models become increasingly complex and are deployed in critical domains such as healthcare, finance, and autonomous systems, the need for effective explainability has grown. Graph Neural Networks (GNNs), which excel in processing graph-structured data, have seen significant advancements, but explainability for GNNs is still in its early stages. Existing approaches fall into two broad categories: post-hoc explainers and inherently interpretable models. Their evaluation is often limited to synthetic datasets for which ground truth explanations are available, or conducted with the assumption that each XAI method extracts explanations for a fixed network. We focus specifically on inherently interpretable GNNs (e.g., based on prototypes, graph kernels) which enable model-level explanations. For evaluation, these models claim inherent interpretability and only assess predictive accuracy, without applying concrete interpretability metrics. These evaluation practices fundamentally restrict the utility of any discussions regarding explainability. We propose a unified and comprehensive framework for measuring and evaluating explainability in GNNs that extends beyond synthetic datasets, ground-truth constraints, and rigid assumptions, while also supporting the development and refinement of models based on derived explanations. The framework involves measures of Accuracy, Instance-level explanations, and Model-level explanations (AIM), inspired by the generic Co-12 conceptual properties of explanations quality (Nauta et al., 2023). We apply this framework to a suite of existing models, deriving ways to extract explanations from them and to highlight their strengths and weaknesses. Furthermore, based on this analysis using AIM, we develop a new model called XGKN that demonstrates improved explainability while performing on par with existing models. Our approach aims to advance the field of Explainable AI (XAI) for GNNs, offering more robust and practical solutions for understanding and interpreting complex models.", "keywords": [ "Graph Neural Networks", "explainability", "graph kernels" ], "rating_list": [ 3, 5, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 3, 2 ], "presentation_list": [ 1, 2, 2, 3, 3 ], "contribution_list": [ 1, 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "KZEqbwJfTl", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Brendan Juba", "gender": "Male", "institution": "Washington University in St. Louis", "country": "US", "position": "Associate Professor" }, { "name": "Jizhou Huang", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "US", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 14, "n_ref_uni": 31, "n_ref": 55, "n_ref_all": 61, "n_fig": 3, "n_tab": 2, "L_tab": 11, "n_element_tab": 2, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 217, "n_element_tab_1": 1, "formula_len_all": 2693, "formula_len_all_1": 1810, "len_all": 135385, "len_all_1": 88357, "len_abs": 1398, "len_title": 121, "len_sents": 37687, "len_sents_1": 33826, "n_sents": 257, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1437, "title": "Distribution-Specific Agnostic Conditional Classification With Halfspaces", "abs": "We study \"selective\" or \"conditional\" classification problems under an agnostic setting. Classification tasks commonly focus on modeling the relationship between features and categories that captures the vast majority of data. In contrast to common machine learning frameworks, conditional classification intends to model such relationships only on a subset of the data defined by some selection rule. Most work on conditional classification either solves the problem in a realizable setting or does not guarantee the error is bounded compared to an optimal solution. In this work, we consider selective/conditional classification by sparse linear classifiers for subsets defined by halfspaces, and give both positive as well as negative results for Gaussian feature distributions. On the positive side, we present the first PAC-learning algorithm for homogeneous halfspace selectors with error guarantee $\\tilde{O}(\\sqrt{\\mathrm{opt}})$, where $\\mathrm{opt}$ is the smallest conditional classification error over the given class of classifiers and homogeneous halfspaces. On the negative side, we find that, under cryptographic assumptions, approximating the conditional classification loss within a small additive error is computationally hard even under Gaussian distribution. We prove that approximating conditional classification is at least as hard as approximating agnostic classification in both additive and multiplicative form.", "keywords": [ "Agnostic linear classification", "PAC-learning", "Intractability" ], "rating_list": [ 5, 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 2, 3 ] }, { "paper_id": "KYxxSgwRxf", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Canhui Tang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Hao Sun", "gender": "Male", "institution": "China Telecom Corporation Ltd. Data&AI Technology Company", "country": "CN", "position": "Researcher" }, { "name": "Haoyue Shi", "gender": "Female", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Le Wang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Sanping Zhou", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 22, "n_formula_1": 19, "n_ref_uni": 36, "n_ref": 92, "n_ref_all": 106, "n_fig": 8, "n_tab": 5, "L_tab": 622, "n_element_tab": 94, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 607, "n_element_tab_1": 98, "formula_len_all": 1122, "formula_len_all_1": 945, "len_all": 122858, "len_all_1": 64245, "len_abs": 1446, "len_title": 127, "len_sents": 36001, "len_sents_1": 28902, "n_sents": 291, "n_sents_1": 223, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1333, "title": "Action Typicality and Uniqueness Learning for Zero-Shot Video Anomaly Detection", "abs": "Zero-Shot Video Anomaly Detection (ZS-VAD) is an urgent task in scenarios where the target video domain lacks training data due to various concerns, \\emph{e.g.}, data privacy. The skeleton-based approach is a promising way to achieve ZS-VAD as it eliminates domain disparities in both background and human appearance. However, existing methods only learn low-level skeleton representation and rely on the domain-specific normality boundary, which cannot generalize well to new scenes with different normal and abnormal behavior patterns. In this paper, we propose a novel skeleton-based zero-shot video anomaly detection framework, which captures both scene-generic typical anomalies and scene-adaptive unique anomalies. Firstly, we introduce a language-guided typicality modeling module that projects skeleton snippets into action semantic space and learns generalizable typical distributions of normal and abnormal behavior. Secondly, we propose a test-time context uniqueness analysis module to finely analyze the spatio-temporal differences between skeleton snippets and then derive scene-adaptive boundaries. Without using any training samples from the target domain, our method achieves state-of-the-art results on four large-scale VAD datasets: ShanghaiTech, UBnormal, NWPU, and UCF-Crime. The Code will be publicly available.", "keywords": [ "video anomaly detection", "zero-shot", "skeleton-based" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "KYipmCMmSO", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Charlie Hou", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Giulia Fanti", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Shuqi Ke", "gender": "Not Specified", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Sewoong Oh", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Full Professor" } ], "n_formula": 226, "n_formula_1": 36, "n_ref_uni": 46, "n_ref": 112, "n_ref_all": 145, "n_fig": 10, "n_tab": 4, "L_tab": 671, "n_element_tab": 83, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 32638, "formula_len_all_1": 2218, "len_all": 282218, "len_all_1": 62832, "len_abs": 932, "len_title": 132, "len_sents": 69050, "len_sents_1": 27218, "n_sents": 737, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 940, "title": "Characterizing the Training Dynamics of Private Fine-tuning with Langevin Diffusion", "abs": "We show that differentially private full fine-tuning (DP-FFT) can distort pre-trained backbone features based on both theoretical and empirical results. We identify the cause of the distortion as the misalignment between the pre-trained backbone and the randomly initialized linear head. We prove that a sequential fine-tuning strategy can mitigate the feature distortion: first-linear-probing-then-fine-tuning (DP-LP-FFT). A new approximation scheme allows us to derive approximate upper and lower bounds on the training loss of DP-LP and DP-FFT, in a simple but canonical setting of 2-layer neural networks with ReLU activation. Experiments on real-world datasets and architectures are consistent with our theoretical insights. We also derive new upper bounds for 2-layer linear networks without the approximation. Moreover, our theory suggests a trade-off of privacy budget allocation in multi-phase fine-tuning methods like DP-LP-FFT.", "keywords": [ "differential privacy", "convergence", "fine-tuning theory", "transfer learning theory", "langevin diffusion", "gradient flow" ], "rating_list": [ 8, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 2, 3, 2 ] }, { "paper_id": "KYOdZRR6nr", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Ali Mahdavi-Amiri", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "Assistant Professor" }, { "name": "Dingdong Yang", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "PhD student" }, { "name": "Konrad Schindler", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Professor" }, { "name": "Yizhi Wang", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Hao Richard Zhang", "gender": "Male", "institution": "Amazon", "country": "CA", "position": "Principal Researcher" } ], "n_formula": 14, "n_formula_1": 7, "n_ref_uni": 62, "n_ref": 118, "n_ref_all": 175, "n_fig": 26, "n_tab": 3, "L_tab": 976, "n_element_tab": 85, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 144, "n_element_tab_1": 22, "formula_len_all": 1893, "formula_len_all_1": 735, "len_all": 161134, "len_all_1": 50012, "len_abs": 1281, "len_title": 116, "len_sents": 39329, "len_sents_1": 22567, "n_sents": 291, "n_sents_1": 153, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1293, "title": "GALA: Geometry-Aware Local Adaptive Grids for Detailed 3D Generation", "abs": "We propose GALA, a novel representation of 3D shapes that (i) excels at capturing and reproducing complex geometry and surface details, (ii) is computationally efficient, and (iii) lends itself to 3D generative modelling with modern, diffusion-based schemes. The key idea of GALA is to exploit both the global sparsity of surfaces within a 3D volume and their local surface properties. *Sparsity* is promoted by covering only the 3D object boundaries, not empty space, with an ensemble of tree root voxels. Each voxel contains an octree to further limit storage and compute to regions that contain surfaces. *Adaptivity* is achieved by fitting one local and geometry-aware coordinate frame in each non-empty leaf node. Adjusting the orientation of the local grid, as well as the anisotropic scales of its axes, to the local surface shape greatly increases the amount of detail that can be stored in a given amount of memory, which in turn allows for quantization without loss of quality. With our optimized C++/CUDA implementation, GALA can be fitted to an object in less than 10 seconds. Moreover, the representation can efficiently be flattened and manipulated with transformer networks. We provide a cascaded generation pipeline capable of generating 3D shapes with great geometric detail.", "keywords": [ "Generative Model", "3D", "Computer Graphics" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 2, 4, 4, 3 ], "presentation_list": [ 3, 4, 2, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 2, 5 ] }, { "paper_id": "KXrgDM3mVD", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Adam Klivans", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "Professor" }, { "name": "Chengyue Gong", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Daniel Jesus Diaz", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "Postdoc" }, { "name": "Philipp Kraehenbuehl", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Yue Zhao", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "PhD student" }, { "name": "Jeffrey O Zhang", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 25, "n_ref": 71, "n_ref_all": 99, "n_fig": 7, "n_tab": 8, "L_tab": 3202, "n_element_tab": 381, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2116, "n_element_tab_1": 236, "formula_len_all": 658, "formula_len_all_1": 724, "len_all": 134961, "len_all_1": 74632, "len_abs": 958, "len_title": 114, "len_sents": 37230, "len_sents_1": 29408, "n_sents": 326, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 970, "title": "Distilling Structural Representations into Protein Sequence Models", "abs": "Protein language (or sequence) models, like the popular ESM2, are now widely used tools for extracting evolution-based protein representations and have achieved significant success on core downstream biological tasks.\nA major open problem is how to obtain representations that best capture both the sequence evolutionary history and the atomic structural properties of proteins in general. \nWe introduce **I**mplicit **S**equence **M**odel, a sequence-only input model with structurally-enriched representations that outperforms state-of-the-art sequence models on several well-studied benchmarks including mutation stability assessment and structure prediction. \nOur key innovations are a microenvironment-based Autoencoder for generating structure tokens and a self-supervised training objective that distills these tokens into ESM2's pre-trained model. \nNotably, we make ISM's structure-enriched weights easily accessible for any application using the ESM2 framework.", "keywords": [ "biology", "proteins", "sequence", "structure", "autoencoder", "esm" ], "rating_list": [ 6, 3, 6 ], "soundness_list": [ 4, 2, 3 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 4, 4, 2 ] }, { "paper_id": "KXiQI6ggFc", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Benjamin Rodatz", "gender": "Male", "institution": "Department of Computer Science", "country": "", "position": "PhD student" }, { "name": "Ian Fan", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "MS student" }, { "name": "Neil John Ortega", "gender": "unknown", "institution": "Quantinuum Ltd.", "country": "", "position": "Researcher" }, { "name": "Thomas Hoffmann", "gender": "Male", "institution": "Artidis AG", "country": "CH", "position": "Researcher" }, { "name": "Tuomas Laakkonen", "gender": "unknown", "institution": "Quantinuum", "country": "", "position": "Researcher" }, { "name": "Vincent Wang", "gender": "Male", "institution": "Quantinuum", "country": "", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 5, "n_ref_uni": 60, "n_ref": 81, "n_ref_all": 105, "n_fig": 7, "n_tab": 8, "L_tab": 2024, "n_element_tab": 206, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 551, "n_element_tab_1": 76, "formula_len_all": 1746, "formula_len_all_1": 613, "len_all": 232835, "len_all_1": 52717, "len_abs": 1026, "len_title": 45, "len_sents": 79291, "len_sents_1": 24941, "n_sents": 553, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 951, "title": "A Pattern Language for Machine Learning Tasks", "abs": "We formalise the essential data of objective functions as equality constraints on composites of learners. We call these constraints \"tasks\", and we investigate the idealised view that such tasks determine model behaviours. We develop a flowchart-like graphical mathematics for tasks that allows us to;\n(1) design and optimise desired behaviours model-agnostically;\n(2) offer a unified perspective of approaches in machine learning across domains;\n(3) import insights from theoretical computer science into practical machine learning.\nAs a proof-of-concept of the potential practical impact of our theoretical framework, we exhibit and implement a novel \"manipulator\" task that minimally edits input data to have a desired attribute. Our model-agnostic approach achieves this without the need for custom architectures, adversarial training, random sampling, or interventions on the data, hence enabling capable, small-scale, and training-stable models.", "keywords": [ "Category Theory", "Interpretability", "Compositionality", "Machine Learning Theory", "Formal Methods", "Behaviour Modelling", "Objective Function Analysis", "Diagrammatic Reasoning" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 1, 1 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 3, 2, 2, 2 ] }, { "paper_id": "KXLbcIEurw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Chenning Xu", "gender": "unknown", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Dian Jin", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Fangxun Shu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Hao Jiang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jinghao Wei", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 27, "n_ref": 42, "n_ref_all": 44, "n_fig": 1, "n_tab": 9, "L_tab": 3080, "n_element_tab": 374, "n_fig_1": 0, "n_tab_1": 8, "L_tab_1": 2055, "n_element_tab_1": 304, "formula_len_all": 270, "formula_len_all_1": 270, "len_all": 136223, "len_all_1": 59455, "len_abs": 1291, "len_title": 109, "len_sents": 30395, "len_sents_1": 26530, "n_sents": 224, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1299, "title": "SAG: Style-Aligned Generation via Language Model Collaboration", "abs": "Large language models (LLMs) have increased the demand for personalized and stylish content generation. However, closed-source models like GPT-4 present limitations in optimization opportunities, while the substantial training costs and inflexibility of open-source alternatives, such as Qwen-72B, pose considerable challenges. Conversely, small language models (SLMs) struggle with understanding complex instructions and transferring learned capabilities to new contexts, often exhibiting more pronounced limitations. In this paper, we present a novel collaborative training framework that leverages the strengths of both LLMs and SLMs for style article generation, surpassing the performance of either model alone. We freeze the LLMs to harness their robust instruction-following capabilities and subsequently apply supervised fine-tuning on the SLM using style-specific data. Additionally, we introduce a self-improvement method to enhance style consistency. Our new benchmark, NoteBench, thoroughly evaluates style-aligned generation. Extensive experiments show that our approach achieves state-of-the-art performance, with improvements of 0.78 in ROUGE-L and 0.55 in BLEU-4 scores compared to GPT-4, while also maintaining a low hallucination rate in terms of factual accuracy and faithfulness.", "keywords": [ "stylish content generation", "large language model" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "KXDOmD7DM7", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Chien-Yao Wang", "gender": "Male", "institution": "Academia Sinica", "country": "TW", "position": "Assistant Research Fellow/Professor" }, { "name": "Hao Tang Tsui", "gender": "Male", "institution": "NYCU", "country": "TW", "position": "Undergrad student" }, { "name": "Hong-yuan Liao", "gender": "Male", "institution": "Academia Sinica", "country": "", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 8, "n_ref_uni": 32, "n_ref": 67, "n_ref_all": 98, "n_fig": 13, "n_tab": 11, "L_tab": 3683, "n_element_tab": 515, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1274, "n_element_tab_1": 131, "formula_len_all": 1264, "formula_len_all_1": 399, "len_all": 148199, "len_all_1": 60672, "len_abs": 1294, "len_title": 140, "len_sents": 40659, "len_sents_1": 28091, "n_sents": 301, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1380, "title": "YOLO-RD: Introducing Relevant and Compact Explicit Knowledge to YOLO by Retriever-Dictionary", "abs": "Identifying and localizing objects within images is a fundamental challenge, and numerous efforts have been made to enhance model accuracy by experimenting with diverse architectures and refining training strategies. Nevertheless, a prevalent limitation in existing models is overemphasizing the current input while ignoring the information from the entire dataset. We introduce an innovative $\\textbf{R}etriever-\\textbf{D}ictionary$ (RD) module to address this issue. This architecture enables YOLO-based models to efficiently retrieve features from a Dictionary that contains the insight of the dataset, which is built by the knowledge from Visual Models (VM), Large Language Models (LLM), or Visual Language Models (VLM). The flexible RD enables the model to incorporate such explicit knowledge that enhances the ability to benefit multiple tasks, specifically, segmentation, detection, and classification, from pixel to image level. The experiments show that using the RD significantly improves model performance, achieving more than a 3\\% increase in mean Average Precision for object detection with less than a 1\\% increase in model parameters. Beyond 1-stage object detection models, the RD module improves the effectiveness of 2-stage models and DETR-based architectures, such as Faster R-CNN and Deformable DETR. Code is released at https://github.com/henrytsui000/YOLO.", "keywords": [ "Real-time Object Detection", "Object Detection", "Dictionary Learning", "Retrieval-Augmented Generation", "Representation Learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "KX5hd1RhYP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chenhao Sun", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Mark Niklas Mueller", "gender": "Male", "institution": "LogicStar AI", "country": "CH", "position": "Researcher" }, { "name": "Martin Vechev", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Yuhao Mao", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 14, "n_ref": 43, "n_ref_all": 60, "n_fig": 6, "n_tab": 7, "L_tab": 1046, "n_element_tab": 42, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 239, "n_element_tab_1": 23, "formula_len_all": 990, "formula_len_all_1": 1020, "len_all": 88952, "len_all_1": 57252, "len_abs": 1308, "len_title": 114, "len_sents": 27392, "len_sents_1": 23651, "n_sents": 216, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1313, "title": "Average Certified Radius is a Poor Metric for Randomized Smoothing", "abs": "Randomized smoothing is a popular approach for providing certified robustness guarantees against adversarial attacks, and has become a very active area of research. Over the past years, the average certified radius (ACR) has emerged as the single most important metric for comparing methods and tracking progress in the field. However, in this work, we show that ACR is an exceptionally poor metric for evaluating robustness guarantees provided by randomized smoothing. We theoretically show not only that a trivial classifier can have arbitrarily large ACR, but also that ACR is much more sensitive to improvements on easy samples than on hard ones. Empirically, we confirm that existing training strategies that improve ACR reduce the model's robustness on hard samples. Further, we show that by focusing on easy samples, we can effectively replicate the increase in ACR. We develop strategies, including explicitly discarding hard samples, reweighing the dataset with certified radius, and extreme optimization for easy samples, to achieve state-of-the-art ACR, although these strategies ignore robustness for the general data distribution. Overall, our results suggest that ACR has introduced a strong undesired bias to the field, and better metrics are required to holistically evaluate randomized smoothing.", "keywords": [ "Randomized Smoothing", "Metric Evaluation" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 2, 1 ], "confidence_list": [ 5, 4, 5 ] }, { "paper_id": "KWo4w1UXs8", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Cen Zhang", "gender": "unknown", "institution": "Sohu", "country": "CN", "position": "Researcher" }, { "name": "Kaiquanzhu", "gender": "unknown", "institution": "Sohu", "country": "CN", "position": "Researcher" }, { "name": "Qingyun Wang", "gender": "Male", "institution": "Sohu", "country": "CN", "position": "Researcher" }, { "name": "Shuowen Liang", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "", "position": "MS student" }, { "name": "Tian YANG", "gender": "Male", "institution": "Sohu", "country": "CN", "position": "Principal Researcher" }, { "name": "sisi li", "gender": "Female", "institution": "Sohu", "country": "CN", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 19, "n_ref": 38, "n_ref_all": 55, "n_fig": 7, "n_tab": 1, "L_tab": 1390, "n_element_tab": 35, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 981, "n_element_tab_1": 32, "formula_len_all": 558, "formula_len_all_1": 557, "len_all": 86646, "len_all_1": 56522, "len_abs": 1418, "len_title": 149, "len_sents": 28728, "len_sents_1": 25214, "n_sents": 221, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1429, "title": "GUNet: A Graph Convolutional Network United Diffusion Model for Stable and Diversity Pose Generation", "abs": "Pose skeleton images are an important reference in pose-controllable image generation. In order to enrich the source of skeleton images, recent works have investigated the generation of pose skeletons based on natural language. These methods are based on GANs. However, it remains challenging to perform diverse, structurally correct and aesthetically pleasing human pose skeleton generation with various textual inputs. To address this problem, we propose a framework with GUNet as the main model, PoseDiffusion. It is the first generative framework based on a diffusion model and also contains a series of variants fine-tuned based on a stable diffusion model. PoseDiffusion demonstrates several desired properties that outperform existing methods. 1) Correct Skeletons. GUNet, a denoising model of PoseDiffusion, is designed to incorporate graphical convolutional neural networks. It is able to learn the spatial relationships of the human skeleton by introducing skeletal information during the training process. 2) Diversity. We decouple the key points of the skeleton and characterise them separately, and use cross-attention to introduce textual conditions. Experimental results show that PoseDiffusion outperforms existing SoTA algorithms in terms of stability and diversity of text-driven pose skeleton generation. Qualitative analyses further demonstrate its superiority for controllable generation in Stable Diffusion.", "keywords": [ "Diffusion", "Text2Pose", "GCN", "UNet" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 1, 2, 2 ], "presentation_list": [ 3, 1, 2 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "KWUFlIMn8A", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Behzad Nourani", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Setareh Maghsudi", "gender": "unknown", "institution": "Ruhr-Universt\u00e4t Bochum", "country": "", "position": "Full Professor" }, { "name": "Sofien Dhouib", "gender": "Male", "institution": "University of Tuebingen", "country": "", "position": "Postdoctoral researcher" }, { "name": "Steven Bilaj", "gender": "Male", "institution": "Ruhr-Universit\u00e4t Bochum", "country": "", "position": "PhD student" } ], "n_formula": 167, "n_formula_1": 23, "n_ref_uni": 46, "n_ref": 96, "n_ref_all": 115, "n_fig": 0, "n_tab": 6, "L_tab": 600, "n_element_tab": 128, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18125, "formula_len_all_1": 1881, "len_all": 191407, "len_all_1": 56065, "len_abs": 882, "len_title": 87, "len_sents": 53091, "len_sents_1": 25190, "n_sents": 531, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 886, "title": "Clusters Agnostic Network Lasso Bandits", "abs": "We consider a multi-task contextual bandit setting, where the learner is given a graph encoding relations between the bandit tasks. The tasks' preference vectors are assumed to be piecewise constant over the graph, forming clusters. At every round, we estimate the preference vectors by solving an online network lasso problem with a suitably chosen, time-dependent regularization parameter. We establish a novel oracle inequality relying on a convenient restricted eigenvalue assumption. Our theoretical findings highlight the importance of dense intra-cluster connections and sparse inter-cluster ones. That results in a sublinear regret bound significantly lower than its counterpart in the independent task learning setting. Finally, we support our theoretical findings by experimental evaluation against graph bandit multi-task learning and online clustering of bandits algorithms.", "keywords": [ "multitask learning", "contextual bandits", "graph total variation", "network lasso", "clustering" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "KWH4UIoQKS", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Dongsheng Zuo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Feng Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Lei Chen", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Xilin Xia", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "Undergrad student" }, { "name": "Yuzhe Ma", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhihai Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 8, "n_ref_uni": 55, "n_ref": 127, "n_ref_all": 160, "n_fig": 7, "n_tab": 15, "L_tab": 3715, "n_element_tab": 451, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1428, "n_element_tab_1": 136, "formula_len_all": 6470, "formula_len_all_1": 622, "len_all": 224646, "len_all_1": 74719, "len_abs": 1729, "len_title": 121, "len_sents": 71971, "len_sents_1": 35517, "n_sents": 498, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1736, "title": "Computing Circuits Optimization via Model-Based Circuit Genetic Evolution", "abs": "Optimizing computing circuits such as multipliers and adders is a fundamental challenge in modern integrated circuit design. Recent efforts propose formulating this optimization problem as a reinforcement learning (RL) proxy task, offering a promising approach to search high-speed and area-efficient circuit design solutions. However, we show that the RL-based formulation (proxy task) converges to a local optimal design solution (original task) due to the deceptive reward signals and incrementally localized actions in the RL-based formulation. To address this challenge, we propose a novel model-based circuit genetic evolution (MUTE) framework, which reformulates the problem as a genetic evolution process by proposing a grid-based genetic representation of design solutions. This novel formulation avoids misleading rewards by evaluating and improving generated solutions using the true objective value rather than proxy rewards. To promote globally diverse exploration, MUTE proposes a multi-granularity genetic crossover operator that recombines design substructures at varying column ranges between two grid-based genetic solutions. To the best of our knowledge, MUTE is the first to reformulate the problem as a circuit genetic evolution process, which enables effectively searching for global optimal design solutions. We evaluate MUTE on several fundamental computing circuits, including multipliers, adders, and multiply-accumulate circuits. Experiments on these circuits demonstrate that MUTE significantly Pareto-dominates state-of-the-art approaches in terms of both area and delay. Moreover, experiments demonstrate that circuits designed by MUTE well generalize to large-scale computation-intensive circuits as well.", "keywords": [ "AI Chips Design", "Computing Circuits Optimization", "Evolutionary Algorithm", "Reinforcement Learning" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 4, 4, 4 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "KW8yzAOIZr", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Hu Ding", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Ruomin Huang", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" }, { "name": "Shihong Song", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Wanlin Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Weichen Lin", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" } ], "n_formula": 30, "n_formula_1": 20, "n_ref_uni": 46, "n_ref": 96, "n_ref_all": 123, "n_fig": 7, "n_tab": 10, "L_tab": 5066, "n_element_tab": 399, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 43, "n_element_tab_1": 12, "formula_len_all": 2536, "formula_len_all_1": 1278, "len_all": 176039, "len_all_1": 71028, "len_abs": 1014, "len_title": 147, "len_sents": 53183, "len_sents_1": 34072, "n_sents": 406, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 99, "L_abs": 1028, "title": "To Tackle Adversarial Transferability: A Novel Ensemble Training Method with Fourier Transformation", "abs": "Ensemble methods are commonly used for enhancing robustness in machine learning. However, due to the ''transferability'' of adversarial examples, the performance of an ensemble model can be \nseriously affected even it contains a set of independently trained sub-models. To address this issue, we propose an efficient data transformation method based on a cute ''weakness allocation'' strategy, to diversify non-robust features.\nOur approach relies on a fine-grained analysis on the relation between non-robust features and adversarial attack directions.\nMoreover, our approach enjoys several other advantages, e.g., it does not require any communication between sub-models and the construction complexity is also quite low.\nWe conduct a set of experiments to evaluate the performance of our proposed method and compare it with several popular baselines. The results suggest that our approach can achieve significantly improved robust accuracy over most existing ensemble methods, and meanwhile preserve high clean accuracy.", "keywords": [ "robustness", "diversity", "ensemble training", "Fourier transformation" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "KW6B6s1X82", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Ali Mahdavi-Amiri", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "Assistant Professor" }, { "name": "Amirhossein Alimohammadi", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "MS student" }, { "name": "Andrea Tagliasacchi", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "Associate Professor" }, { "name": "Ghassan Hamarneh", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "Full Professor" }, { "name": "Saeid Asgari", "gender": "unknown", "institution": "Autodesk", "country": "", "position": "Research Scientist" }, { "name": "Sauradip Nag", "gender": "Male", "institution": "Simon Fraser University", "country": "CA", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 58, "n_ref": 120, "n_ref_all": 145, "n_fig": 8, "n_tab": 7, "L_tab": 2669, "n_element_tab": 302, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 681, "n_element_tab_1": 64, "formula_len_all": 461, "formula_len_all_1": 409, "len_all": 212720, "len_all_1": 67276, "len_abs": 186, "len_title": 73, "len_sents": 52786, "len_sents_1": 32115, "n_sents": 397, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 25, "L_abs": 633, "title": "SMITE: Segment Me In TimE", "abs": "Segmenting an object in a video presents significant challenges. Each pixel must be accurately labeled, and these labels must remain consistent across frames. The difficulty increases when the segmentation is with arbitrary granularity, meaning the number of segments can vary arbitrarily, and masks are defined based on only one or a few sample images. In this paper, we address this issue by employing a pre-trained text to image diffusion model supplemented with an additional tracking mechanism. We demonstrate that our approach can effectively manage various segmentation scenarios and outperforms state-of-the-art alternatives.", "keywords": [ "video segmentation", "diffusion models", "video diffusion", "part segmentation" ], "rating_list": [ 6, 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3, 4 ] }, { "paper_id": "KVLnLKjymq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Boris Ginsburg", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Fei Jia", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Shantanu Acharya", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 22, "n_ref": 34, "n_ref_all": 59, "n_fig": 12, "n_tab": 16, "L_tab": 6066, "n_element_tab": 897, "n_fig_1": 12, "n_tab_1": 5, "L_tab_1": 1653, "n_element_tab_1": 182, "formula_len_all": 556, "formula_len_all_1": 328, "len_all": 117340, "len_all_1": 59468, "len_abs": 777, "len_title": 107, "len_sents": 36188, "len_sents_1": 26790, "n_sents": 280, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 780, "title": "Star Attention: Efficient LLM Inference over Long Sequences", "abs": "Inference with Transformer-based Large Language Models (LLMs) on long sequences is both costly and slow due to the quadratic complexity of the self-attention mechanism. We introduce Star Attention, a two-phase block-sparse approximation that improves computational efficiency by sharding attention across multiple hosts while minimizing communication overhead. In the first phase, the context is processed using blockwise-local attention across hosts, in parallel. In the second phase, query and response tokens attend to all prior cached tokens through sequence-global attention. Star Attention integrates seamlessly with most Transformer-based LLMs trained with global attention, reducing memory requirements and inference time by up to 11x while preserving 95-100% of accuracy.", "keywords": [ "Large Language Model", "Transformers", "Long Context", "Efficient Inference", "Local and Global Attention" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "KUpUO7aSSg", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Haozhou Wang", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Postdoc" }, { "name": "James Burridge", "gender": "Male", "institution": "University of Tokyo", "country": "JP", "position": "Assistant Professor" }, { "name": "Pieter M. Blok", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Assistant Professor" }, { "name": "Shuai Xiang", "gender": "Not Specified", "institution": "The University of Tokyo", "country": "JP", "position": "MS student" }, { "name": "Wei Guo", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Associate Professor" } ], "n_formula": 21, "n_formula_1": 18, "n_ref_uni": 46, "n_ref": 78, "n_ref_all": 105, "n_fig": 6, "n_tab": 13, "L_tab": 5007, "n_element_tab": 463, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1951, "n_element_tab_1": 240, "formula_len_all": 2148, "formula_len_all_1": 1213, "len_all": 181284, "len_all_1": 67635, "len_abs": 3137, "len_title": 117, "len_sents": 37472, "len_sents_1": 28342, "n_sents": 332, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1044, "title": "DODA: Diffusion for Object-detection Domain Adaptation in Agriculture", "abs": "Object detection has wide applications in agriculture, but the trained models often struggle to generalize across diverse agricultural environments. To address this challenge, we propose DODA (\\underline{D}iffusion for \\underline{O}bject-detection \\underline{D}omain Adaptation in \\underline{A}griculture), a unified framework that leverages diffusion models to generate domain-specific detection data for multiple agricultural scenarios. DODA incorporates external domain embeddings and an improved layout-to-image (L2I) approach, allowing it to generate high-quality detection data for new domains without additional training. We demonstrate DODA's effectiveness on the Global Wheat Head Detection dataset, where fine-tuning detectors on DODA-generated data yields significant improvements across multiple domains (maximum +15.6 AP). DODA provides a simple yet powerful approach to adapt object detectors to diverse agricultural scenarios, lowering barriers for more plant breeders growers to use detection in their personalized environments.", "keywords": [ "diffusion model", "object detection", "domain adaptation" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "KUf2iyin77", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Bingchen Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Haoning Wu", "gender": "Male", "institution": "Rhymes AI", "country": "SG", "position": "Researcher" }, { "name": "Weisi Lin", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Xin Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Postdoc" }, { "name": "Zhibo Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Lu Yiting", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 90, "n_ref_all": 130, "n_fig": 10, "n_tab": 8, "L_tab": 3386, "n_element_tab": 418, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 1230, "n_element_tab_1": 183, "formula_len_all": 458, "formula_len_all_1": 307, "len_all": 170266, "len_all_1": 68468, "len_abs": 1708, "len_title": 134, "len_sents": 48586, "len_sents_1": 31416, "n_sents": 336, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1409, "title": "Q-Adapt: Adapting LMM for Visual Quality Perceiver with Progressive Instruction Tuning", "abs": "The rapid advancement of Large Multi-modal Foundation Models (LMM) has paved the way for the possible Explainable Image Quality Assessment (EIQA) with instruction tuning from two perspectives: overall quality explanation, and attribute-wise perception answering. However, existing works usually overlooked the conflicts between these two types of perception explanations during joint instruction tuning, leading to insufficient perception understanding. To mitigate this, we propose a new paradigm for perception-oriented instruction tuning, i.e., Q-Adapt, which aims to eliminate the conflicts and achieve the synergy between these two EIQA tasks when adapting LMM, resulting in enhanced multi-faceted explanations of IQA. Particularly, we propose a progressive instruction tuning strategy by dividing the adaption process of LMM for EIQA into two stages, where the first stage empowers the LMM with universal perception knowledge tailored for two tasks using an efficient transfer learning strategy, i.e., LoRA, and the second stage introduces the instruction-adaptive visual prompt tuning to dynamically adapt visual features for the different instructions from two tasks. In this way, our proposed Q-Adapt can achieve a lightweight visual quality perceiver, demonstrating comparable performance and, in some instances, superior results across perceptual-related benchmarks and commonly-used IQA databases.", "keywords": [ "Explainable Image Quality Assessment", "LMM" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "KUX2T1cY8w", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Aymene Berriche", "gender": "Male", "institution": "Universit\u00e9 Grenoble Alpes", "country": "FR", "position": "MS student" }, { "name": "Mehdi Zakaria Adjal", "gender": "Male", "institution": "Universit\u00e9 Paris Cit\u00e9", "country": "FR", "position": "MS student" }, { "name": "Riyadh Baghdadi", "gender": "Male", "institution": "New York University", "country": "AE", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 28, "n_ref": 39, "n_ref_all": 57, "n_fig": 7, "n_tab": 2, "L_tab": 577, "n_element_tab": 106, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 577, "n_element_tab_1": 106, "formula_len_all": 399, "formula_len_all_1": 399, "len_all": 107780, "len_all_1": 58925, "len_abs": 1221, "len_title": 127, "len_sents": 31001, "len_sents_1": 28466, "n_sents": 214, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1228, "title": "PRE-TRAIN WITH BACKPROPAGATION AND FINE-TUNE WITH A BIO-PLAUSIBLE LEARNING RULE", "abs": "Backpropagation (BP) has long been the cornerstone of deep neural network training. While neural networks trained with backpropagation typically have high accuracy and precision, they suffer from limitations in their robustness to adversarial perturbation. Biologically plausible (bio-plausible) learning rules, on the other hand, are more robust. Yet, they typically underperform in terms of accuracy and precision, which has limited their widespread adoption. In this work, we aim to bridge this gap. We propose a novel approach where neural networks are pre-trained using backpropagation and fine-tuned using bio-plausible learning rules. We use several types of Sign-Symmetry learning methods to fine-tune models pre-trained using backpropagation. We explore the effectiveness of this approach in two tasks, image classification and image retrieval, then demonstrate that it improves robustness against gradient-based adversarial attacks while offering comparable accuracy and precision compared to the use of backpropagation alone. These findings show the benefit of mixing backpropagation and bio-plausible learning rules, suggesting the need for further research by the community to evaluate this approach on other tasks.", "keywords": [ "Bio-plausible", "credit assignment", "backpropagation", "sign-symmetry", "adversarial robustness" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "KTrnOhAN4k", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Duo Xu", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Postdoc" }, { "name": "Olga Russakovsky", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Associate Professor" }, { "name": "Yan Yan", "gender": "Male", "institution": "University of Illinois Chicago", "country": "US", "position": "Associate Professor" }, { "name": "Ye Zhu", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Yu Wu", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Zhiwei Deng", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Research Scientist" } ], "n_formula": 23, "n_formula_1": 4, "n_ref_uni": 64, "n_ref": 190, "n_ref_all": 220, "n_fig": 11, "n_tab": 4, "L_tab": 1440, "n_element_tab": 98, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1123, "n_element_tab_1": 94, "formula_len_all": 1638, "formula_len_all_1": 333, "len_all": 220780, "len_all_1": 71222, "len_abs": 1713, "len_title": 110, "len_sents": 69445, "len_sents_1": 35361, "n_sents": 427, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1718, "title": "Discovery and Expansion of New Domains within Diffusion Models", "abs": "In this work, we study the generalization properties of diffusion models in a few-shot setup, introduce a novel tuning-free paradigm to synthesize the target out-of-domain (OOD) data, showcase multiple applications of those generalization properties, and demonstrate the advantages compared to existing tuning-based methods in data-sparse scientific scenarios with large domain gaps. Our work resides on the observation and premise that the theoretical formulation of denoising diffusion implicit models (DDIMs), a non-Markovian inference technique, exhibits latent Gaussian priors independent from the parameters of trained denoising diffusion probabilistic models (DDPMs). This brings two practical benefits: the latent Gaussian priors generalize to OOD data domains that have never been used in the training stage; existing DDIMs offer the flexibility to traverse the denoising chain bidirectionally for a pre-trained DDPM. We then demonstrate through theoretical and empirical studies that such established OOD Gaussian priors are practically separable from the originally trained ones after inversion. The above analytical findings allow us to introduce our novel tuning-free paradigm to synthesize new images of the target unseen domain by discovering qualified OOD latent encodings within the inverted noisy latent spaces, which is fundamentally different from most existing paradigms that seek to modify the denoising trajectory to achieve the same goal by tuning the model parameters. Extensive cross-model and domain experiments show that our proposed method can expand the latent space and synthesize images in new domains via frozen DDPMs without impairing the generation quality of their original domains.", "keywords": [ "domain generalization", "diffusion models", "few-shot", "ML4Astrophysics" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 2, 2, 4, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 2, 2, 4, 4 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "KTpSiJ1QNn", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Reut Dayan", "gender": "Female", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" }, { "name": "Uri Shaham", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "Assistant Professor" }, { "name": "Yahel Jacobs", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 25, "n_ref": 42, "n_ref_all": 57, "n_fig": 9, "n_tab": 7, "L_tab": 1671, "n_element_tab": 133, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 925, "n_element_tab_1": 56, "formula_len_all": 403, "formula_len_all_1": 336, "len_all": 125663, "len_all_1": 63806, "len_abs": 1151, "len_title": 129, "len_sents": 43276, "len_sents_1": 30840, "n_sents": 324, "n_sents_1": 226, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1167, "title": "G-SPARC: SPectral ARchitectures tackling the Cold-start problem in Graphs", "abs": "Graphs play a central role in modeling complex relationships across various domains. Most graph learning methods rely heavily on neighborhood information, raising the question of how to handle \\textit{cold-start nodes} \u2014 nodes with no known connections within the graph.\nThese models often overlook the cold-start nodes, making them ineffective for real-world scenarios. To tackle this, we propose G-SPARC, a novel framework addressing cold-start nodes, that leverages generalizable spectral embedding. This framework enables extension to state-of-the-art methods making them suitable for practical applications. By utilizing a key idea of transitioning from graph representation to spectral representation, our approach is generalizable to cold-start nodes, capturing the global structure of the graph without relying on adjacency data. Experimental results demonstrate that our method outperforms existing models on cold-start nodes across various tasks like node classification, node clustering, and link prediction. G-SPARC provides a breakthrough built-in solution to the cold-start problem in graph learning. Our code will be publicly available upon acceptance.", "keywords": [ "Cold-Start Nodes", "Graph Neural Networks", "Spectral Representation" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "KTgQGXz5xj", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Ayush Mohanty", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Nagi Gebraeel", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Nazal Mohamed", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Paritosh Ramanan", "gender": "unknown", "institution": "Oklahoma State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 76, "n_formula_1": 17, "n_ref_uni": 25, "n_ref": 28, "n_ref_all": 53, "n_fig": 1, "n_tab": 8, "L_tab": 3775, "n_element_tab": 185, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 713, "n_element_tab_1": 85, "formula_len_all": 8141, "formula_len_all_1": 2939, "len_all": 187897, "len_all_1": 58506, "len_abs": 6863, "len_title": 144, "len_sents": 56173, "len_sents_1": 23023, "n_sents": 483, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1980, "title": "Federated Granger Causality Learning For Interdependent Clients With State Space Representation", "abs": "Advanced sensors and IoT devices have improved the monitoring and control of complex industrial enterprises. They have also created an interdependent fabric of geographically distributed process operations (clients) across these enterprises. Granger causality is an effective approach to detect and quantify interdependencies by examining how the state of one client affects the states of others over time. Understanding these interdependencies helps capture how localized events, such as faults and disruptions, can propagate throughout the system, potentially leading to widespread operational impacts. However, the large volume and complexity of industrial data present significant challenges in effectively modeling these interdependencies. This paper develops a federated approach to learning Granger causality. We utilize a linear state space system framework that leverages low-dimensional state estimates to analyze interdependencies. This helps address bandwidth limitations and the computational burden commonly associated with centralized data processing. We propose augmenting the client models with the Granger causality information learned by the server through a Machine\nLearning (ML) function. We examine the co-dependence between the augmented client and server models and reformulate the framework as a standalone ML algorithm providing conditions for its sublinear and linear convergence rates. We also study the convergence of the framework to a centralized oracle model. Moreover, we include a differential privacy analysis to ensure data security while preserving causal insights. Using synthetic data, we conduct comprehensive experiments to demonstrate the robustness of our approach to perturbations in causality, the scalability to the size of communication, number of clients, and the dimensions of raw data. We also evaluate the performance on two real-world industrial control system datasets by reporting the volume of data saved by decentralization.", "keywords": [ "State Space", "Federated Learning", "Granger Causality", "Interdependencies" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 3, 3, 2, 2 ] }, { "paper_id": "KTHUTtEX5F", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Chenglei Yu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Haixin Wang", "gender": "unknown", "institution": "UCLA Computer Science Department, University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Qianru Zhang", "gender": "Female", "institution": "the University of Hong Kong, University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Siu Ming Yiu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Tailin Wu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiangyu Kuang", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Postdoc" }, { "name": "Xiao Luo", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Yuansheng Cao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yudong Yan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 16, "n_formula_1": 6, "n_ref_uni": 34, "n_ref": 96, "n_ref_all": 110, "n_fig": 3, "n_tab": 4, "L_tab": 3137, "n_element_tab": 182, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 3433, "n_element_tab_1": 182, "formula_len_all": 847, "formula_len_all_1": 343, "len_all": 180249, "len_all_1": 68161, "len_abs": 1741, "len_title": 75, "len_sents": 63079, "len_sents_1": 31469, "n_sents": 480, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1814, "title": "CellDJBench: Benchmark Datasets for Data-Driven Biological Fluid Simulation", "abs": "Biological fluid simulation is a critical tool for comprehending the intricate and complex fluid dynamics that occur within biological systems. Recently, data-driven techniques have emerged as a promising avenue to enhance the accuracy and efficiency of biological fluid simulations. However, the community encounters two challenges. (1) Existing biological datasets only capture static snapshots, lacking the ability to capture dynamic biological processes. (2) These datasets are limited in scale due to the demanding experimental conditions. To address these challenges, this paper introduces four comprehensive large-scale datasets: Tension, Wets, CellDivision and Jellyfish, containing a wealth of biological dynamics and pushing the boundary of data-driven methods. These datasets have been meticulously designed to encompass a wide array of biological fluid dynamics scenarios. By incorporating physical modeling techniques such as phase-field method, these datasets provide a standardized evaluation framework for data-driven approaches. They empower researchers to objectively assess and compare different methodologies, fostering advancements in the field of biological fluid simulation. Furthermore, the availability of these benchmark datasets facilitates reproducibility and enhances the comparability of results across studies, promoting knowledge sharing and collaboration within the research community. Researchers can build upon existing models, leading to cumulative progress in the development of accurate and efficient data-driven models for simulating complex fluid dynamics within biological systems. We offer benchmark code and \\model\\ dataset link through the following link: \\href{https://anonymous.4open.science/r/BioJCell--9E53/README.md}{https://anonymous.4open.science/r/ CellDJBench}.", "keywords": [ "Cell; biological dynamic simulation" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "KSPBh07jEO", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Bijean Ghafouri", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Jacob-Junqi Tian", "gender": "unknown", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "", "position": "Intern" }, { "name": "James Zhou", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Jean-Fran\u00e7ois Godbout", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "US", "position": "Full Professor" }, { "name": "Kellin Pelrine", "gender": "unknown", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "Mayank Goel", "gender": "Male", "institution": "International Institute of Information Technology, Hyderabad, International Institute of Information Technology Hyderabad", "country": "IN", "position": "Undergrad student" }, { "name": "Pratheeksha Nair", "gender": "Female", "institution": "McGill University", "country": "", "position": "PhD student" }, { "name": "Reihaneh Rabbany", "gender": "Female", "institution": "McGill University", "country": "", "position": "Assistant Professor" }, { "name": "Shahrad Mohammadzadeh", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 59, "n_ref_all": 81, "n_fig": 9, "n_tab": 4, "L_tab": 550, "n_element_tab": 78, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 514, "n_element_tab_1": 54, "formula_len_all": 25, "formula_len_all_1": 0, "len_all": 177457, "len_all_1": 63880, "len_abs": 1097, "len_title": 92, "len_sents": 56245, "len_sents_1": 32141, "n_sents": 367, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 1102, "title": "Epistemic Integrity in Large Language Models", "abs": "Large language models are increasingly relied upon as sources of information, but their propensity for generating false or misleading statements with high confidence poses risks for users and society. In this paper, we confront the critical problem of epistemic miscalibration\u2014where a model's linguistic assertiveness fails to reflect its true internal certainty. We introduce a new human-labeled dataset and a novel method for measuring the linguistic assertiveness of Large Language Models which cuts error rates by over 50\\% relative to previous benchmarks. Validated across multiple datasets, our method reveals a stark misalignment between how confidently models linguistically present information and their actual accuracy. Further human evaluations confirm the severity of this miscalibration. This evidence underscores the urgent risk of the overstated certainty Large Language Models hold which may mislead users on a massive scale. Our framework provides a crucial step forward in diagnosing and correcting this miscalibration, offering a path to safer and more trustworthy AI across domains.", "keywords": [ "uncertainty", "assertiveness", "persuasion", "calibration", "misinformation" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 3, 2, 1, 1 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 3, 2, 1, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "KSLkFYHlYg", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Jenna Fromer", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Keir Adams", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Kento Abeywardane", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Connor Coley", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Associate Professor" } ], "n_formula": 68, "n_formula_1": 17, "n_ref_uni": 87, "n_ref": 175, "n_ref_all": 239, "n_fig": 21, "n_tab": 14, "L_tab": 4687, "n_element_tab": 427, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4084, "formula_len_all_1": 1153, "len_all": 391398, "len_all_1": 73699, "len_abs": 1182, "len_title": 90, "len_sents": 130826, "len_sents_1": 33777, "n_sents": 1001, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 90, "L_abs": 1182, "title": "ShEPhERD: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design", "abs": "Engineering molecules to exhibit precise 3D intermolecular interactions with their environment forms the basis of chemical design. In ligand-based drug design, bioisosteric analogues of known bioactive hits are often identified by virtually screening chemical libraries with shape, electrostatic, and pharmacophore similarity scoring functions. We instead hypothesize that a generative model which learns the joint distribution over 3D molecular structures and their interaction profiles may facilitate 3D interaction-aware chemical design. We specifically design ShEPhERD, an SE(3)-equivariant diffusion model which jointly diffuses/denoises 3D molecular graphs and representations of their shapes, electrostatic potential surfaces, and (directional) pharmacophores to/from Gaussian noise. Inspired by traditional ligand discovery, we compose 3D similarity scoring functions to assess ShEPhERD\u2019s ability to conditionally generate novel molecules with desired interaction profiles. We demonstrate ShEPhERD\u2019s potential for impact via exemplary drug design tasks including natural product ligand hopping, protein-blind bioactive hit diversification, and bioisosteric fragment merging.", "keywords": [ "3D molecular generation", "drug design", "molecules" ], "rating_list": [ 6, 10, 8 ], "soundness_list": [ 3, 4, 2 ], "presentation_list": [ 3, 4, 4 ], "contribution_list": [ 3, 4, 4 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "KSBx6FBZpE", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Abhiram Iyer", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Ila R Fiete", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Professor" }, { "name": "Rylan Schaeffer", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" }, { "name": "Sunny Duan", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "mikail khona", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 67, "n_ref_all": 86, "n_fig": 13, "n_tab": 1, "L_tab": 2093, "n_element_tab": 142, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 2093, "n_element_tab_1": 142, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 149167, "len_all_1": 60658, "len_abs": 1686, "len_title": 100, "len_sents": 46530, "len_sents_1": 28102, "n_sents": 227, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1692, "title": "Uncovering Latent Memories in Large Language Models", "abs": "Frontier AI systems are making transformative impacts across society, but such benefits are not without costs: models trained on web-scale datasets containing personal and private data raise profound concerns about data privacy and security. Language models are trained on extensive corpora including potentially sensitive or proprietary information, and the risk of data leakage, where the model response reveals pieces of such information, remains inadequately understood. Prior work has investigated that sequence complexity and the number of repetitions are the primary drivers of memorization. In this work, we examine the most vulnerable class of data: highly complex sequences that are presented only once during training. These sequences often contain the most sensitive information and pose considerable risk if memorized. By analyzing the progression of memorization for these sequences throughout training, we uncover a striking observation: many memorized sequences persist in the model's memory, exhibiting resistance to catastrophic forgetting even after just one encounter. Surprisingly, these sequences may not appear memorized immediately after their first exposure but can later be \u201cuncovered\u201d during training, even in the absence of subsequent exposures - a phenomenon we call \"latent memorization.\" Latent memorization presents a serious challenge for data privacy, as sequences that seem hidden at the final checkpoint of a model may still be easily recoverable. We demonstrate how these hidden sequences can be revealed through random weight perturbations, and we introduce a diagnostic test based on cross-entropy loss to accurately identify latent memorized sequences.", "keywords": [ "Large Language Models", "Memorization", "Empirical Study", "Data Leakage", "Privacy", "LLMs", "Dynamics", "Interpretability", "Mechanistic" ], "rating_list": [ 3, 8, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 4, 3, 4 ], "contribution_list": [ 1, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "KS4G94XBo7", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Georgios Piliouras", "gender": "unknown", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Ian Gemp", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Research Scientist" }, { "name": "Luke Marris", "gender": "unknown", "institution": "DeepMind", "country": "GB", "position": "Research Engineer" }, { "name": "Marc Lanctot", "gender": "Male", "institution": "Google DeepMind", "country": "CA", "position": "Research Scientist" }, { "name": "Siqi Liu", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 20, "n_ref_uni": 57, "n_ref": 147, "n_ref_all": 180, "n_fig": 12, "n_tab": 3, "L_tab": 2685, "n_element_tab": 65, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 34, "n_element_tab_1": 2, "formula_len_all": 2139, "formula_len_all_1": 2016, "len_all": 185007, "len_all_1": 72188, "len_abs": 1046, "len_title": 93, "len_sents": 51909, "len_sents_1": 30037, "n_sents": 419, "n_sents_1": 288, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1065, "title": "Deviation Ratings: A general, clone invariant rating method", "abs": "Many real-world multi-agent or multi-task evaluation scenarios can be naturally modelled as normal-form games due to inherent strategic (adversarial, cooperative, and mixed motive) interactions. These strategic interactions may be agentic (e.g. players trying to win), fundamental (e.g. cost vs quality), or complimentary (e.g. niche finding and specialization). In such a formulation, it is the strategies (actions, policies, agents, models, tasks, prompts, etc.) that are rated. However, the rating problem is complicated by redundancy and complexity of N-player strategic interactions. Repeated or similar strategies can distort ratings for those that counter or complement them. Previous work proposed ``clone-invariant'' ratings to handle such redundancies, but this was limited to two-player zero-sum (i.e. strictly competitive) interactions. This work introduces the first N-player general-sum clone-invariant rating, called \\emph{deviation ratings}, based on coarse correlated equilibria. The rating is explored on several domains including LLMs evaluation.", "keywords": [ "rating", "ranking", "coarse correlated equilibria", "Nash equilibria", "game theory", "equilibria", "LLM leaderboard", "normal-form game" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 4, 3, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 1, 3, 2 ], "confidence_list": [ 2, 2, 3 ] }, { "paper_id": "KRv9NubipP", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Ah-Hwee Tan", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Full Professor" }, { "name": "Efstratios Gavves", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Associate Professor" }, { "name": "Jie Liu", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "PhD student" }, { "name": "Yingjun Du", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Cees Snoek", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "Jan-jakob Sonke", "gender": "unknown", "institution": "Netherlands Cancer Institute", "country": "NL", "position": "Full Professor" }, { "name": "Pan Zhou", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 86, "n_ref_all": 111, "n_fig": 11, "n_tab": 4, "L_tab": 613, "n_element_tab": 55, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 860, "n_element_tab_1": 22, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 153160, "len_all_1": 81803, "len_abs": 1562, "len_title": 131, "len_sents": 41521, "len_sents_1": 36657, "n_sents": 339, "n_sents_1": 305, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1510, "title": "CaPo: Cooperative Plan Optimization for Efficient Embodied Multi-Agent Cooperation", "abs": "In this work, we address the cooperation problem among large language model (LLM) based embodied agents, where agents must cooperate to achieve a common goal. Previous methods often execute actions extemporaneously and incoherently, without long-term strategic and cooperative planning, leading to redundant steps, failures, and even serious repercussions in complex tasks like search-and-rescue missions where discussion and cooperative plan are crucial. To solve this issue, we propose Cooperative Plan Optimization (CaPo) to enhance the cooperation efficiency of LLM-based embodied agents. Inspired by human cooperation schemes, CaPo improves cooperation efficiency with two phases: 1) meta plan generation, and 2) progress-adaptive meta plan and execution. In the first phase, all agents analyze the task, discuss, and cooperatively create a meta-plan that decomposes the task into subtasks with detailed steps, ensuring a long-term strategic and coherent plan for efficient coordination. In the second phase, agents execute tasks according to the meta-plan and dynamically adjust it based on their latest progress (e.g., discovering a target object) through multi-turn discussions. This progress-based adaptation eliminates redundant actions, improving the overall cooperation efficiency of agents. Experimental results on the ThreeDworld Multi-Agent Transport and Communicative Watch-And-Help tasks demonstrate CaPo's much higher task completion rate and efficiency compared with state-of-the-arts.", "keywords": [ "Embodied AI", "multi-agent cooperation", "LLM" ], "rating_list": [ 8, 5, 5, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "KRqMfdwQaP", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Junggeun Do", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Intern" }, { "name": "Jay Yoon Lee", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 17, "n_ref": 26, "n_ref_all": 36, "n_fig": 3, "n_tab": 4, "L_tab": 1148, "n_element_tab": 131, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1148, "n_element_tab_1": 131, "formula_len_all": 1167, "formula_len_all_1": 1167, "len_all": 75084, "len_all_1": 52157, "len_abs": 1419, "len_title": 116, "len_sents": 24417, "len_sents_1": 23366, "n_sents": 171, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1427, "title": "SEAL-Pose: Enhancing Pose Estimation through Trainable Loss Function", "abs": "Accurately predicting 3D human pose is a challenging task in computer vision due to the need to capture complex spatial structures and anatomical constraints. We propose SEAL-Pose, an adaptation of the Structured Energy As Loss (SEAL) framework for deterministic models, specifically designed to enhance 3D human pose estimation from 2D keypoints. \nAlthough the original SEAL was limited to probabilistic models, our approach employs the model's predictions as negative examples to train a structured energy network, which functions as a dynamic and trainable loss function. \nOur approach enables a pose estimation model to learn joint dependencies via learning signals from a structured energy network that automatically captures body structure during training without explicit prior structural knowledge, resulting in more accurate and plausible 3D poses .\nWe introduce new evaluation metrics to assess the structural consistency of predicted poses, demonstrating that SEAL-Pose produces more realistic, anatomically plausible results.\nExperimental results on the Human3.6M and Human3.6M WholeBody datasets show that SEAL-Pose not only reduces pose estimation errors such as Mean Per Joint Position Error (MPJPE) but also outperforms existing baselines.\nThis work highlights the potential of applying structured energy networks to tasks requiring complex output structures, offering a promising direction for future research.", "keywords": [ "Structured Energy network", "Energy-based models", "Trainable Loss-function", "Dynamic loss function", "Pose Estimation" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "KRnsX5Em3W", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Hadas Kotek", "gender": "Female", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Hadas Orgad", "gender": "Female", "institution": "Computer Science Department, Technion - Israel Institute of Technology", "country": "IL", "position": "PhD student" }, { "name": "Idan Szpektor", "gender": "unknown", "institution": "Google", "country": "IL", "position": "Researcher" }, { "name": "Michael Toker", "gender": "Male", "institution": "Computer Science Departmen, Technion-Israel Institute of Technology", "country": "", "position": "MS student" }, { "name": "Roi Reichart", "gender": "Male", "institution": "Technion, Israel Institute of Technology", "country": "IL", "position": "Associate Professor" }, { "name": "Yonatan Belinkov", "gender": "Male", "institution": "Technion, Technion", "country": "IL", "position": "Assistant Professor" }, { "name": "Zorik Gekhman", "gender": "Male", "institution": "Technion, Technion", "country": "", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 82, "n_ref": 149, "n_ref_all": 173, "n_fig": 11, "n_tab": 3, "L_tab": 1031, "n_element_tab": 69, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 982, "n_element_tab_1": 60, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 242890, "len_all_1": 68892, "len_abs": 6406, "len_title": 133, "len_sents": 56610, "len_sents_1": 33721, "n_sents": 418, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1376, "title": "LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations", "abs": "Large language models (LLMs) often produce errors, including factual inaccuracies, biases, and reasoning failures, collectively referred to as \"hallucinations\". Recent studies have demonstrated that LLMs' internal states encode information regarding the truthfulness of their outputs, and that this information can be utilized to detect errors. In this work, we show that the internal representations of LLMs encode much more information about truthfulness than previously recognized. We first discover that the truthfulness information is concentrated in specific tokens, and leveraging this property significantly enhances error detection performance. Yet, we show that such error detectors fail to generalize across datasets, implying that---contrary to prior claims---truthfulness encoding is not universal but rather multifaceted. Next, we show that internal representations can also be used for predicting the types of errors the model is likely to make, facilitating the development of tailored mitigation strategies. Lastly, we reveal a discrepancy between LLMs' internal encoding and external behavior: they may encode the correct answer, yet consistently generate an incorrect one. Taken together, these insights deepen our understanding of LLM errors from the model's internal perspective, which can guide future research on enhancing error analysis and mitigation.", "keywords": [ "hallucinations", "truthfulness", "interpretability", "probing", "LLM" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 3, 4, 2 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "KRlJN9d1sW", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Jangwon Kim", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Jiseok Jeong", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Jongsoo Lee", "gender": "Male", "institution": "Pohang University of Science and Technology (POSTECH)", "country": "KR", "position": "MS student" }, { "name": "Soohee Han", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 25, "n_ref": 59, "n_ref_all": 79, "n_fig": 6, "n_tab": 11, "L_tab": 4984, "n_element_tab": 604, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1881, "n_element_tab_1": 159, "formula_len_all": 1263, "formula_len_all_1": 1259, "len_all": 135051, "len_all_1": 62001, "len_abs": 1059, "len_title": 121, "len_sents": 46271, "len_sents_1": 27823, "n_sents": 297, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1057, "title": "Reinforcement Learning via Lazy-Agent for Environments with Random Delays", "abs": "Real-world reinforcement learning applications are often hampered by delayed feedback from environments, which violates the fundamental assumption of the Markovian property and introduces significant challenges. While numerous methods have been proposed for handling environments with constant delays, those with random delays remain largely unexplored owing to their inherent complexity and variability. In this study, we explored environments with random delays and proposed a novel strategy to transform them into their equivalent constant-delay counterparts by introducing a simple agent called the *lazy-agent*. This approach naturally overcomes the challenges posed by the variability of random delays, enabling the application of state-of-the-art methods, originally designed for handling constant delays, to random-delay environments without any modification. Empirical results demonstrate that our lazy-agent significantly outperformed other baseline algorithms in terms of asymptotic performance and sample efficiency in random-delay environments.", "keywords": [ "Delayed system", "Markov decision process", "reinforcement learning" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "KRhcZIAcoM", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Chengjia Wang", "gender": "unknown", "institution": "Heriot-Watt University", "country": "GB", "position": "Assistant Professor" }, { "name": "Giorgos Papanastasiou", "gender": "Male", "institution": "Archimedes Unit, Athena RC", "country": "GR", "position": "Principal Researcher" }, { "name": "Hongwei Bran Li", "gender": "Male", "institution": "Massachusetts General Hospital, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Peng Qi", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Associate Professor" }, { "name": "Yinyi LAI", "gender": "unknown", "institution": "Hohai University", "country": "CN", "position": "Undergrad student" }, { "name": "Yuchen Mao", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "MS student" }, { "name": "Yunjie Yang", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 31, "n_ref": 44, "n_ref_all": 51, "n_fig": 2, "n_tab": 5, "L_tab": 3519, "n_element_tab": 135, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1616, "n_element_tab_1": 68, "formula_len_all": 196, "formula_len_all_1": 196, "len_all": 127287, "len_all_1": 63267, "len_abs": 2018, "len_title": 129, "len_sents": 34532, "len_sents_1": 30677, "n_sents": 231, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 2087, "title": "Semi-Supervised Medical Image Segmentation via Knowledge Mining from Large Models", "abs": "Large-scale vision models like SAM possess extensive visual knowledge, but their application to specialized tasks like medical image segmentation is often hindered by their general nature and the computational challenges associated with training and finetuning. Locally hosted small models such as U-Net++, designed for specific tasks, struggle with limited performance due to sparse labeled datasets. This study introduces a strategic knowledge mining method as a novel interaction mechanism between large and small models. Our method utilizes SAM\u2019s broad visual understanding to enhance the specialized capabilities of locally hosted small deep learning models. Specifically, we trained a U-Net++ model on a limited labeled dataset and extend its capabilities by converting outputs (masks) produced in unlabeled images into prompts, to extract relevant knowledge from SAM. This process not only harnesses SAM\u2019s generalized visual knowledge but also iteratively improves SAM\u2019s prediction to cater specialized medical segmentation tasks via UNet++. The mined knowledge, serving as \u2018pseudo labels\u2019, enriches the training dataset, enabling the fine-tuning of the local network. Applied to the Kvasir SEG and COVID-QU-Ex datasets which consist of gastrointestinal polyp and lung Xray images respectively, our proposed method consistently enhanced the segmentation performance on Dice by 3% and 1% respectively over the baseline U-Net++ model, when the same amount of labelled data were used during training (75% and 50% of labelled data). Remarkably, our proposed method surpassed the baseline U-Net++ model even when the latter was trained exclusively on labeled data (100% of labelled data). These results underscore the potential of knowledge mining to overcome data limitations in specialized models by leveraging the broad, albeit general, knowledge of large-scale models like SAM, all while maintaining operational efficiency essential for clinical applications.The code of our method is publicly available at https://anonymous.4open.science/r/Knowledge-Mining-from-Large-Models-C7FE.", "keywords": [ "Foundation Model", "Knowledge Mining", "Segment Anything Model", "Segmentation", "Semi-Supervised Learning" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "KRdiRGSNc9", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "BAI Huiyu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "MS student" }, { "name": "Bei Chen", "gender": "Female", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Fengji Zhang", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Guancheng Lin", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Jacky Keung", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Linquan Wu", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "MS student" }, { "name": "Xiao Li", "gender": "Male", "institution": "Computer Science, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiao Yu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Yue Wang", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 41, "n_ref": 82, "n_ref_all": 150, "n_fig": 19, "n_tab": 6, "L_tab": 3131, "n_element_tab": 563, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1308, "n_element_tab_1": 194, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 209145, "len_all_1": 67156, "len_abs": 1803, "len_title": 165, "len_sents": 63525, "len_sents_1": 31909, "n_sents": 449, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 116, "L_abs": 1832, "title": "HumanEval-V: Evaluating Visual Understanding and Reasoning Abilities of Large Multimodal Models Through Coding Tasks", "abs": "Coding tasks have been valuable for evaluating Large Language Models (LLMs), as they demand the comprehension of high-level instructions, complex reasoning, and the implementation of functional programs $-$ core capabilities for advancing Artificial General Intelligence. Despite the progress in Large Multimodal Models (LMMs), which extend LLMs with visual understanding and reasoning capabilities, there remains a notable lack of coding benchmarks that rigorously assess these models, particularly in tasks that emphasize visual reasoning. To address this gap, we introduce HumanEval-V, a novel and lightweight benchmark specifically designed to evaluate LMMs' visual understanding and reasoning capabilities through code generation tasks. HumanEval-V includes 108 carefully crafted, entry-level Python coding tasks derived from platforms like CodeForces and Stack Overflow. Each task is created by modifying the context and algorithmic patterns of the original problems and redrawing the visual elements to ensure they are distinct from the source. LMMs are required to generate code solutions using the provided visual context and a predefined Python function signature that outlines the task requirements. Every coding task is equipped with meticulously crafted human-generated test cases to ensure a thorough and reliable evaluation of the model-generated code solutions. We evaluate 19 state-of-the-art LMMs using HumanEval-V, uncovering significant challenges. Proprietary models like GPT-4o achieve only 13% pass@1 and 36.4% pass@10, while open-weight models with 70B parameters score below 4% pass@1. Ablation studies further demonstrate the limitations of current LMMs in vision understanding and reasoning as well as coding abilities. These results highlight key areas for future research to improve LMMs' capabilities.", "keywords": [ "Large Multimodal Models", "Vision Language Models", "Code Generation", "Visual Reasoning", "Visual Understanding" ], "rating_list": [ 3, 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 4, 3 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "KRMSH1GxUK", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Patrick McDaniel", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Full Professor" }, { "name": "Somesh Jha", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin, Madison", "country": "US", "position": "Full Professor" }, { "name": "Xiaogeng Liu", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Zhengyue Zhao", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "chaowei xiao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 25, "n_ref": 53, "n_ref_all": 89, "n_fig": 13, "n_tab": 6, "L_tab": 2584, "n_element_tab": 322, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 340, "n_element_tab_1": 26, "formula_len_all": 403, "formula_len_all_1": 210, "len_all": 145367, "len_all_1": 65467, "len_abs": 1961, "len_title": 111, "len_sents": 49452, "len_sents_1": 32133, "n_sents": 344, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1970, "title": "Can Watermarks be Used to Detect LLM IP Infringement For Free?", "abs": "The powerful capabilities of LLMs stem from their rich training data and high-quality labeled datasets, making the training of strong LLMs a resource-intensive process, which elevates the importance of IP protection for such LLMs. Compared to gathering high-quality labeled data, directly sampling outputs from these fully trained LLMs as training data presents a more cost-effective approach. This practice\u2014where a suspect model is fine-tuned using high-quality data derived from these LLMs, thereby gaining capabilities similar to the target model\u2014can be seen as a form of IP infringement against the original LLM. In recent years, LLM watermarks have been proposed and used to detect whether a text is AI-generated. Intuitively, if data sampled from a watermarked LLM is used for training, the resulting model would also be influenced by this watermark. This raises the question: can we directly use such watermarks to detect IP infringement of LLMs? In this paper, we explore the potential of LLM watermarks for detecting model infringement. We find that there are two issues with direct detection: (1) The queries used to sample output from the suspect LLM have a significant impact on detectability. (2) The watermark that is easily learned by LLMs exhibits instability regarding the watermark's hash key during detection. To address these issues, we propose LIDet, a detection method that leverages available anchor LLMs to select suitable queries for sampling from the suspect LLM. Additionally, it adapts the detection threshold to mitigate detection failures caused by different hash keys. To demonstrate the effectiveness of this approach, we construct a challenging model set containing multiple suspect LLMs on which direct detection methods struggle to yield effective results. Our method achieves over 90\\% accuracy in distinguishing between infringing and clean models, demonstrating the feasibility of using LLM watermarks to detect LLM IP infringement.", "keywords": [ "large language models", "watermark", "model copyright", "model infringement detection" ], "rating_list": [ 6, 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 2, 3, 3, 2, 3 ] }, { "paper_id": "KR9zrisXDZ", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Gabriel Nobis", "gender": "Male", "institution": "Fraunhofer HHI", "country": "DE", "position": "PhD student" }, { "name": "Jackie Ma", "gender": "unknown", "institution": "Fraunhofer HHI", "country": "DE", "position": "Head of Applied Machine Learning" }, { "name": "Maximilian Springenberg", "gender": "Male", "institution": "Fraunhofer HHI, Fraunhofer IAIS", "country": "DE", "position": "PhD student" }, { "name": "Wojciech Samek", "gender": "Male", "institution": "TU Berlin", "country": "DE", "position": "Associate Professor" } ], "n_formula": 65, "n_formula_1": 27, "n_ref_uni": 21, "n_ref": 61, "n_ref_all": 87, "n_fig": 6, "n_tab": 5, "L_tab": 441, "n_element_tab": 33, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 431, "n_element_tab_1": 31, "formula_len_all": 3483, "formula_len_all_1": 1188, "len_all": 116825, "len_all_1": 57803, "len_abs": 2031, "len_title": 121, "len_sents": 38457, "len_sents_1": 25923, "n_sents": 299, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1973, "title": "Inverse Engineering Diffusion: Deriving Variance Schedules with Rationale", "abs": "A fundamental aspect of diffusion models is the variance schedule, which governs the evolution of variance throughout the diffusion process. Despite numerous studies exploring variance schedules, little effort has been made to understand the variance distributions implied by sampling from these schedules and how it benefits both training and data generation. We introduce a novel perspective on score-based diffusion models, bridging the gap between the variance schedule and its underlying variance distribution. Specifically, we propose the notion of sampling variance according to a probabilistic rationale, which induces a density. Our approach views the inverse of the variance schedule as a cumulative distribution function (CDF) and its first derivative as a probability density function (PDF) of the variance distribution. This formulation not only offers a unified view of variance schedules but also allows for the direct engineering of a variance schedule from the probabilistic rationale of its inverse function. Additionally, our framework is not limited to CDFs with closed-form inverse solutions, enabling the exploration of variance schedules that are unattainable through conventional methods. We present the tools required to obtain a diverse array of novel variance schedules tailored to specific rationales, such as separability metrics or prior beliefs. These schedules may exhibit varied dynamics, ranging from rapid convergence towards zero to prolonged periods in high-variance regions. Through comprehensive empirical evaluation, we demonstrate the efficacy of enhancing the performance of diffusion models with schedules distinct from those encountered during training. We provide a principled and unified approach to variance schedules in diffusion models, revealing the relationship between variance schedules and their underlying probabilistic rationales, which yields notable improvements in image generation performance, as measured by FID.", "keywords": [ "diffusion models", "generative models", "score-based generative models", "generative modeling", "stochastic differential equations" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "KQsR2JrCwx", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "Feiyu Xiong", "gender": "Male", "institution": "Institute for Advanced Algorithms Research, Shanghai", "country": "", "position": "Principal Researcher" }, { "name": "Hao Liang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Qinhan Yu", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Simin Niu", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "PhD student" }, { "name": "Zhengren Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhiyu li", "gender": "Male", "institution": "Institute for Advanced Algorithms Research, Shanghai", "country": "CN", "position": "Principal Researcher" }, { "name": "shidawei", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Zhang wen tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Victor Wang", "gender": "Male", "institution": "IAAR", "country": "CN", "position": "Researcher" } ], "n_formula": 33, "n_formula_1": 8, "n_ref_uni": 51, "n_ref": 103, "n_ref_all": 124, "n_fig": 5, "n_tab": 10, "L_tab": 11988, "n_element_tab": 1707, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2600, "n_element_tab_1": 262, "formula_len_all": 1198, "formula_len_all_1": 517, "len_all": 216547, "len_all_1": 56797, "len_abs": 897, "len_title": 127, "len_sents": 54174, "len_sents_1": 23224, "n_sents": 442, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 900, "title": "QAEncoder: Towards Aligned Representation Learning in Question Answering System", "abs": "Modern QA systems entail retrieval-augmented generation (RAG) for accurate and trustworthy responses. However, the inherent gap between user queries and relevant documents hinders precise matching. Motivated by our conical distribution hypothesis, which posits that potential queries and documents form a cone-like structure in the embedding space, we introduce QAEncoder, a training-free approach to bridge this gap. Specifically, QAEncoder estimates the expectation of potential queries in the embedding space as a robust surrogate for the document embedding, and attaches document fingerprints to effectively distinguish these embeddings. Extensive experiments on fourteen embedding models across six languages and eight datasets validate QAEncoder's alignment capability, which offers a plug-and-play solution that seamlessly integrates with existing RAG architectures and training-based methods.", "keywords": [ "Question Answering System; Retrieval Augmented Generation; Document Query Gap; Conical Distribution Hypothesis; Document Fingerprint" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "KQJiC44aSG", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Chao Yang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Shuang Li", "gender": "Female", "institution": "The Chinese University of Hong Kong (Shenzhen)", "country": "", "position": "Assistant Professor" }, { "name": "Wendi Ren", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Yang Yang", "gender": "Male", "institution": "The Chinese University of Hong Kong,Shenzhen", "country": "", "position": "MS student" } ], "n_formula": 29, "n_formula_1": 26, "n_ref_uni": 55, "n_ref": 92, "n_ref_all": 97, "n_fig": 3, "n_tab": 3, "L_tab": 2429, "n_element_tab": 198, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1998, "n_element_tab_1": 168, "formula_len_all": 1962, "formula_len_all_1": 1950, "len_all": 148681, "len_all_1": 65819, "len_abs": 1635, "len_title": 109, "len_sents": 41184, "len_sents_1": 28881, "n_sents": 303, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1426, "title": "Variational Neuro-Symbolic Generative Temporal Point Process", "abs": "Temporal point processes (TPPs) are a powerful framework for modeling event sequences with irregular timestamps, such as those commonly found in electronic health records (EHR), which often involve high-dimensional and diverse event types. However, building generative models for such complex datasets comes with several challenges, including addressing sample inefficiency, accurately capturing intricate event patterns, and producing outputs that are both trustworthy and interpretable. In this paper, we present a neuro-symbolic generative model for TPPs based on the Variational Autoencoder (VAE) framework. Our model incorporates a neural-symbolic reasoning layer into the latent space, allowing it to integrate interpretable, logic-based constraints and perform logical reasoning over learned representations. This integration enhances the interpretability of the latent space by embedding logic rules directly into the generative process, enabling structured reasoning and improved decision-making based on underlying data patterns. We validate our model on an ICU EHR dataset, demonstrating its effectiveness in capturing complex event dynamics with irregular timestamps. In addition to improving sample efficiency and accuracy, our model supports the secure and interpretable generation of synthetic event data, making it a valuable tool for healthcare applications where reliability and trustworthiness are critical.", "keywords": [ "temporal point process", "neuro-symbolic", "generative model" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "KO09K3rBSr", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Chi-Sheng Chen", "gender": "Male", "institution": "Neuro Industry, Inc.", "country": "US", "position": "Researcher" }, { "name": "Chun-Shu Wei", "gender": "unknown", "institution": "National Chiao Tung University, National Chiao Tung University", "country": "", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 32, "n_ref": 46, "n_ref_all": 59, "n_fig": 5, "n_tab": 3, "L_tab": 2399, "n_element_tab": 55, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2399, "n_element_tab_1": 55, "formula_len_all": 137, "formula_len_all_1": 138, "len_all": 101106, "len_all_1": 51331, "len_abs": 877, "len_title": 139, "len_sents": 24853, "len_sents_1": 19278, "n_sents": 151, "n_sents_1": 127, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 880, "title": "Mind's Eye: Image Recognition by EEG via Multimodal Similarity-Keeping Contrastive Learning", "abs": "Decoding images from non-invasive electroencephalographic (EEG) signals has been a grand challenge in understanding how the human brain process visual information in real-world scenarios. To cope with the issues of signal-to-noise ratio and nonstationarity, this paper introduces a MUltimodal Similarity-keeping contrastivE learning (MUSE) framework for zero-shot EEG-based image classification. We develop a series of multivariate time-series encoders tailored for EEG signals and assess the efficacy of regularized contrastive EEG-Image pretraining using an extensive visual EEG dataset. Our method achieves state-of-the-art performance, with a top-1 accuracy of 19.3% and a top-5 accuracy of 48.8% in 200-way zero-shot image classification. Furthermore, we visualize neural patterns via model interpretation, shedding light on the visual processing dynamics in the human brain.", "keywords": [ "EEG", "contrastive learning", "brain-computer interface" ], "rating_list": [ 3, 8, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 1, 3, 2, 3, 1 ], "contribution_list": [ 2, 3, 3, 1, 2 ], "confidence_list": [ 3, 4, 4, 5, 4 ] }, { "paper_id": "KNkalZnq3f", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Ali Payani", "gender": "Male", "institution": "Cisco", "country": "US", "position": "Researcher" }, { "name": "Lu Wang", "gender": "Female", "institution": "Northeastern University", "country": "", "position": "Assistant Professor" }, { "name": "Wenzhao Qiu", "gender": "Not Specified", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "MS student" }, { "name": "Joseph Joshua Peper", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 35, "n_ref_all": 51, "n_fig": 3, "n_tab": 11, "L_tab": 4100, "n_element_tab": 371, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 3065, "n_element_tab_1": 271, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 183843, "len_all_1": 76375, "len_abs": 1634, "len_title": 139, "len_sents": 31112, "len_sents_1": 22480, "n_sents": 199, "n_sents_1": 164, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1649, "title": "MDBench: A Synthetic Multi-Document Reasoning Benchmark Generated with Knowledge Guidance", "abs": "Natural language processing evaluation has made significant progress, largely driven by the proliferation of powerful large language models (LLMs). New evaluation benchmarks are of increasing priority as the reasoning capabilities of LLMs are expanding at a rapid pace. In particular, while *multi-document* (MD) reasoning is an area of extreme relevance given LLM capabilities in handling longer-context inputs, few benchmarks exist to rigorously examine model behavior in this setting. Moreover, the multi-document setting is historically challenging for benchmark creation due to the expensive cost of annotating long inputs.\nIn this work, we introduce **MDBench**, a new dataset for evaluating LLMs on the task of multi-document reasoning. Notably, MDBench is created through a novel synthetic generation process, allowing us to *controllably and efficiently generate challenging document sets* and the corresponding question-answer (QA) examples. Our novel technique operates on condensed structured seed knowledge, modifying it through LLM-assisted edits to induce MD-specific reasoning challenges. We then convert this structured knowledge into a natural text surface form, generating a document set and corresponding QA example.\nWe analyze the behavior of popular LLMs and prompting techniques, finding that MDBench poses significant challenges for all methods, even with relatively short document sets. We also see our knowledge-guided generation technique (1) allows us to readily perform targeted analysis of MD-specific reasoning capabilities and (2) can be adapted quickly to account for new challenges and future modeling improvements.", "keywords": [ "multi-document reasoning", "benchmark creation", "synthetic data generation", "text generation", "nlp" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "KNXFYBrSWH", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Adrien Bolling", "gender": "Male", "institution": "University of Luxemburg", "country": "LU", "position": "PhD student" }, { "name": "Dorian Joubaud", "gender": "unknown", "institution": "University of Luxemburg", "country": "LU", "position": "PhD student" }, { "name": "Matthieu OLEKHNOVITCH", "gender": "unknown", "institution": "University of Luxemburg", "country": "LU", "position": "Intern" }, { "name": "Maxime Cordy", "gender": "Male", "institution": "University of Luxemburg", "country": "", "position": "Researcher" }, { "name": "Mike Papadakis", "gender": "Male", "institution": "University of Luxemburg", "country": "LU", "position": "Associate Professor" }, { "name": "Sylvain KUBLER", "gender": "Male", "institution": "University of Luxemburg", "country": "LU", "position": "Researcher" }, { "name": "YVES LE TRAON", "gender": "Male", "institution": "University of Luxemburg", "country": "", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 17, "n_ref": 31, "n_ref_all": 72, "n_fig": 13, "n_tab": 14, "L_tab": 10466, "n_element_tab": 491, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 801, "n_element_tab_1": 118, "formula_len_all": 735, "formula_len_all_1": 647, "len_all": 151513, "len_all_1": 57417, "len_abs": 1642, "len_title": 149, "len_sents": 47593, "len_sents_1": 26216, "n_sents": 338, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1652, "title": "Tackling the Generative learning trilemma through VAE and GMM-controlled latent space class expansion", "abs": "Achieving efficient data augmentation (DA) in time series classification is not a trivial task due to the high complexity of temporal data. Generative models, such as GANs (Generative Adversarial Networks), diffusion models, and Variational Autoencoders (VAEs), are powerful techniques to address the generative learning trilemma of producing (1) high-quality samples, (2) fast sampling, and (3) diversity. These methods vary in their ability to address the trilemma. Diffusion models allows for high diversity and high quality samples, while GAN allows for high quality samples and fast sampling, and VAE for high diversity and fast sampling. In this paper, we introduce a novel generative method, ASCENSION (VAE and GMM-controlled latent space class expansion), that retains the strengths of VAE in terms of diversity and fast sampling, while enabling controlled and quantifiable exploration of uncharted regions in the latent space. This approach not only enhances classification performance but also yields higher quality (more realistic) samples. ASCENSION leverages the probabilistic nature of the VAE's latent space to represent classes as Gaussian mixture models (GMMs). By modifying this mixture, we enable precise manipulation of class probability densities and boundaries. To ensure intra-class compactness and maximize inter-class separation, we apply clustering constraints. Empirical evaluations on the UCR benchmark (102 datasets) show that ASCENSION outperforms state-of-the-art DA methods, achieving an average classification accuracy improvement of approximately $7\\$% and excelling in all aspects of the generative learning trilemma.", "keywords": [ "data augmentation", "classifier", "variational auto-encoder", "gaussian mixture", "latent space representation" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "KMYr8qwhbS", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Haonan Sun", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Shilong Tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yunhuai Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Zhanxing Zhu", "gender": "Male", "institution": "University of Southampton", "country": "", "position": "Associate Professor" }, { "name": "Zhe Feng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 41, "n_ref": 99, "n_ref_all": 139, "n_fig": 23, "n_tab": 7, "L_tab": 2801, "n_element_tab": 254, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 697, "n_element_tab_1": 108, "formula_len_all": 663, "formula_len_all_1": 276, "len_all": 185937, "len_all_1": 75006, "len_abs": 1529, "len_title": 142, "len_sents": 63541, "len_sents_1": 34779, "n_sents": 614, "n_sents_1": 279, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1572, "title": "LaDEEP: A Deep Learning-based Surrogate Model for Large Deformations of Elastic-Plastic Solids", "abs": "The scientific computing for large deformations of elastic-plastic solids is critical for numerous real-world applications. Classical numerical solvers rely primarily on local discrete linear approximations, which are constrained by an inherent trade-off between accuracy and efficiency. Recently, Deep Learning models have achieved impressive progress in solving PDEs. While previous models have explored various architectures and constructed coefficient-solution mappings, they are designed for general instances without considering specific problem properties and hard to accurately handle with complex elastic-plastic solids involving contact, loading and unloading. In this work, we take stretch bending, a popular metal fabrication technique, as our case study and introduce LaDEEP, a deep learning-based surrogate model for \\textbf{La}rge \\textbf{De}formations of \\textbf{E}lastic-\\textbf{P}lastic Solids. We encode the partitioned regions of the involved solids into a token sequence to maintain their essential order property. To characterize the physical process of the solid deformation, a two-stage Transformer-based module is designed to predict the deformation with the sequence of tokens as input. Empirically, LaDEEP achieves five magnitudes faster speed than finite element methods with a comparable accuracy, and gains 20.47\\% relative improvement on average compared to other deep learning baselines. We have also deployed our model into a real-world industrial production system, and it has shown remarkable performance in both accuracy and efficiency.", "keywords": [ "Continuum Mechanics", "Large Deformation", "Elastic-Plastic Solid", "Transformer", "Deep Learning-based Surrogate Model" ], "rating_list": [ 3, 3, 1, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "KMWGzQi7Qy", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Agustinus Kristiadi", "gender": "unknown", "institution": "Vector Institute", "country": "DE", "position": "Postdoc" }, { "name": "Ahmad Rashid", "gender": "unknown", "institution": "University of Waterloo", "country": "", "position": "PhD student" }, { "name": "Hongliang Li", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CA", "position": "Researcher" }, { "name": "Julia Grosse", "gender": "Female", "institution": "University of Tuebingen", "country": "", "position": "PhD student" }, { "name": "Pascal Poupart", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Full Professor" }, { "name": "Ruotian Wu", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "MS student" } ], "n_formula": 27, "n_formula_1": 20, "n_ref_uni": 41, "n_ref": 89, "n_ref_all": 117, "n_fig": 3, "n_tab": 12, "L_tab": 1795, "n_element_tab": 80, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 1396, "n_element_tab_1": 55, "formula_len_all": 1943, "formula_len_all_1": 1267, "len_all": 167858, "len_all_1": 70761, "len_abs": 1236, "len_title": 106, "len_sents": 49339, "len_sents_1": 32232, "n_sents": 385, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1244, "title": "A Critical Look At Tokenwise Reward-Guided Text Generation", "abs": "Large language models (LLMs) can be improved by aligning with human preferences through fine-tuning---the so-called reinforcement learning from human feedback (RLHF). However, the cost of fine-tuning an LLM is prohibitive for many users. Due to their ability to bypass LLM fine-tuning, prediction-time tokenwise reward-guided text generation (RGTG) methods have recently been proposed. They use a reward model trained on full sequences to score partial sequences during decoding in a bid to steer the generation towards sequences with high rewards. However, these methods have so far been only heuristically motivated and poorly analyzed. In this work, we show that reward models trained on full sequences are not compatible with scoring partial sequences. To alleviate this issue, we propose to train a Bradley-Terry reward model on partial sequences explicitly, and autoregressively sample from the implied tokenwise policy during decoding time. We study the properties of this reward model and the resulting policy: We show that this policy is proportional to the ratio of two distinct RLHF policies. Our simple approach outperforms previous RGTG methods and performs similarly to strong offline baselines without large-scale LLM finetuning.", "keywords": [ "LLM", "RLHF", "Alignment", "Model Efficiency", "Reward Models", "Sampling" ], "rating_list": [ 5, 8, 6, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "KMRCb2VCLq", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Cesare Spinoso-Di Piano", "gender": "Male", "institution": ", McGill University", "country": "", "position": "MS student" }, { "name": "Heyan Huang", "gender": "Female", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Jackie CK Cheung", "gender": "Male", "institution": "McGill University", "country": "", "position": "Associate Professor" }, { "name": "Marc-Antoine Rondeau", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "", "position": "Researcher" }, { "name": "Sanxing Chen", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Yang Gao", "gender": "Female", "institution": "Beijing Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Yu Bai", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 42, "n_ref": 65, "n_ref_all": 133, "n_fig": 3, "n_tab": 39, "L_tab": 22272, "n_element_tab": 2182, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 3332, "n_element_tab_1": 326, "formula_len_all": 191, "formula_len_all_1": 191, "len_all": 280498, "len_all_1": 82935, "len_abs": 1575, "len_title": 120, "len_sents": 77434, "len_sents_1": 35161, "n_sents": 534, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1585, "title": "Identifying and Analyzing Task-Encoding Tokens in Large Language Models", "abs": "In-context learning (ICL) has emerged as an effective solution for few-shot learning with large language models (LLMs). Previous research suggests that LLMs perform ICL by analogizing from the provided demonstrations, similar to how humans learn new tasks. However, how LLMs leverage demonstrations to specify a task and learn a corresponding computational function through ICL remains underexplored. Drawing from the way humans learn from content-label mappings in demonstrations, we categorize the tokens in an ICL prompt into content, stopword, and template tokens, with the latter two typically ignored by humans due to their uninformative nature. Our goal is to identify the type of tokens whose representations highly and directly influence LLM's performance, a property we refer to as **task-encoding**. By ablating representations from the attention of the test example, we find that the representations of informative content tokens have less influence on performance, while template and stopword tokens are more prone to be task-encoding tokens, which contrasts with the human attention to informative words. We further give evidence about the function of task-encoding tokens by showing that their representations aggregate information from the content tokens. Moreover, we demonstrate experimentally that lexical meaning, repetition, and structural cues are the main distinguishing characteristics of these tokens. Our work sheds light on how LLMs learn to perform tasks from demonstrations and deepens our understanding of the roles different types of tokens play in LLMs.", "keywords": [ "Task-encoding tokens", "In-context learning", "Large language models", "Interpretability and analysis" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "KMCJXjlDDr", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Guo Qin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Jianmin Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Xiangdong Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yong Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Mingsheng Long", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 24, "n_formula_1": 19, "n_ref_uni": 36, "n_ref": 100, "n_ref_all": 141, "n_fig": 10, "n_tab": 15, "L_tab": 7811, "n_element_tab": 532, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 5453, "n_element_tab_1": 117, "formula_len_all": 12123, "formula_len_all_1": 8691, "len_all": 205155, "len_all_1": 81122, "len_abs": 1359, "len_title": 119, "len_sents": 58942, "len_sents_1": 31711, "n_sents": 376, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1361, "title": "Timer-XL: Long-Context Transformers for Unified Time Series Forecasting", "abs": "We present Timer-XL, a generative Transformer for unified time series forecasting. To uniformly predict 1D and 2D time series, we generalize next token prediction, predominantly adopted for causal generation of 1D sequences, to multivariate next token prediction. The proposed paradigm uniformly formulates various forecasting scenarios as a long-context generation problem. We opt for the generative Transformer, which can capture global-range and causal dependencies while providing contextual flexibility, to implement unified forecasting on univariate series characterized by non-stationarity, multivariate time series with complicated dynamics and correlations, and covariate-informed contexts that include both endogenous and exogenous time series. Technically, we propose a universal TimeAttention to facilitate generative Transformers on multiple time series, which can effectively capture fine-grained intra- and inter-series dependencies of flattened time series tokens (patches) and is further strengthened by position embeddings in both temporal and variable dimensions. Timer-XL achieves state-of-the-art performance across challenging forecasting benchmarks through a unified approach. By pre-training on large-scale time series, Timer-XL demonstrates notable zero-shot performance, making it a promising architecture for large time series models.", "keywords": [ "Time Series Forecasting", "Transformer" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "KLUDshUx2V", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Usman Ahmad", "gender": "Male", "institution": "Innopolis University", "country": "RU", "position": "MS student" }, { "name": "Vladimir Ivanov", "gender": "Male", "institution": "Innopolis University", "country": "RU", "position": "Associate Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 25, "n_ref": 42, "n_ref_all": 51, "n_fig": 2, "n_tab": 8, "L_tab": 2506, "n_element_tab": 180, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 987, "n_element_tab_1": 58, "formula_len_all": 354, "formula_len_all_1": 529, "len_all": 105697, "len_all_1": 64953, "len_abs": 1633, "len_title": 47, "len_sents": 33609, "len_sents_1": 30946, "n_sents": 241, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1636, "title": "Automating High-Quality Concept Banks: Leveraging LLMs and Multimodal Evaluation Metrics", "abs": "Interpretablility in recent deep learning models has become an epicenter of research particularly in sensitive domains such as healthcare, and finance. Concept bottleneck models have emerged as a promising approach for achieving transparency and interpretability by leveraging a set of human-understandable concepts as an intermediate representation before the prediction layer. However, manual concept annotation is discouraged due to the time and effort involved. Our work explores the potential of large language models (LLMs) for generating high-quality concept banks and proposes a multimodal evaluation metric to assess the quality of generated concepts. We investigate three key research questions: the ability of LLMs to generate concept banks comparable to existing knowledge bases like ConceptNet, the sufficiency of unimodal text-based semantic similarity for evaluating concept-class label associations, and the effectiveness of multimodal information in quantifying concept generation quality compared to unimodal concept-label semantic similarity. Our findings reveal that multimodal models outperform unimodal approaches in capturing concept-class label similarity. Furthermore, our generated concepts for the CIFAR-10 and CIFAR-100 datasets surpass those obtained from ConceptNet and the baseline comparison,\ndemonstrating the standalone capability of LLMs in generating high-quality concepts. Being able to automatically generate and evaluate high-quality concepts will enable researchers to quickly adapt and iterate to a newer dataset with little to no effort before they can feed that into concept bottleneck models.", "keywords": [ "concept generation", "multimodal models", "CLIP", "concept-bottleneck models", "interpretability" ], "rating_list": [ 3, 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 2, 1, 1, 1, 2 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 5, 4, 3 ] }, { "paper_id": "KLTqeiI7w0", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Guanxing Lu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Haoyuan Deng", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "MS student" }, { "name": "Season Si Chen", "gender": "unknown", "institution": "Tsinghua Shenzhen International Graduate School", "country": "CN", "position": "Assistant Professor" }, { "name": "Tengbo Yu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yansong Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ziwei Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 51, "n_ref": 99, "n_ref_all": 115, "n_fig": 6, "n_tab": 4, "L_tab": 2018, "n_element_tab": 403, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 515, "n_element_tab_1": 196, "formula_len_all": 564, "formula_len_all_1": 497, "len_all": 154287, "len_all_1": 68525, "len_abs": 1750, "len_title": 126, "len_sents": 40747, "len_sents_1": 33490, "n_sents": 252, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1517, "title": "AnyBimanual: Transferring Single-arm Policy for General Bimanual Manipulation", "abs": "Performing language-conditioned bimanual manipulation tasks is of great importance for many applications ranging from household service to industrial assembly. However, teleoperating dual-arm demonstrations is expensive due to the high-dimensional action space, which poses challenges for conventional methods to handle general bimanual manipulation tasks. In contrast, single-arm policy has recently demonstrated impressive generalizability across a wide range of tasks because of scaled model parameters and training data, which can provide sharable manipulation knowledge for dual-arm systems. To this end, we propose a plug-and-play method named AnyBimanual, which transfers pretrained single-arm policy to multi-task bimanual manipulation policy with limited bimanual demonstrations. Specifically, we first introduce a skill manager to dynamically schedule the discovered skill primitives from pretrained single-arm policy for bimanual manipulation tasks, which combines skill primitives with embodiment-specific compensation. To mitigate the observation discrepancy between single-arm and dual-arm systems, we present a voxel editor to generate spatial soft masks for visual embedding of the workspace, which aims to align visual input of single-arm policy model for each arm with those during pretraining stage. Extensive results on 13 simulated and real-world tasks indicate the superiority of AnyBimanual with an improvement of 12.67\\% on average success rate compared with previous state-of-the-art methods.", "keywords": [ "Multi-task Bimanual Manipulation", "Skill Learning" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "KLIN1QdcX4", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Aditya Grover", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "Harkanwar Singh", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "MS student" }, { "name": "Shufan Li", "gender": "Male", "institution": "UCLA Computer Science Department, University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 13, "n_ref_uni": 35, "n_ref": 65, "n_ref_all": 93, "n_fig": 11, "n_tab": 6, "L_tab": 1669, "n_element_tab": 168, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1167, "n_element_tab_1": 99, "formula_len_all": 2996, "formula_len_all_1": 1356, "len_all": 159932, "len_all_1": 62368, "len_abs": 1332, "len_title": 70, "len_sents": 52625, "len_sents_1": 27918, "n_sents": 417, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1351, "title": "PopAlign: Population-Level Alignment for Fair Text-to-Image Generation", "abs": "Text-to-image (T2I) models achieve high-fidelity generation through extensive training on large datasets. However, these models may unintentionally pick up undesirable biases of their training data, such as over-representation of particular identities in gender or ethnicity neutral prompts. Existing alignment methods such as Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) fail to address this problem effectively because they operate on pairwise preferences consisting of individual \\textit{samples}, while the aforementioned biases can only be measured at a \\textit{population} level. For example, a single sample for the prompt ``doctor\" could be male or female, but a model generating predominantly male doctors even with repeated sampling reflects a gender bias. To address this limitation, we introduce PopAlign, a novel approach for population-level preference optimization, while standard optimization would prefer entire sets of samples over others. We further derive a stochastic lower bound that directly optimizes for individual samples from preferred populations over others for scalable training.Using human evaluation and standard image quality and bias metrics, we show that PopAlign significantly mitigates the bias of pretrained T2I models while largely preserving the generation quality.", "keywords": [ "text-to-image" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "KL8Sm4xRn7", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Dietrich Klakow", "gender": "Male", "institution": "Saarland University", "country": "DE", "position": "Full Professor" }, { "name": "Omer Moussa", "gender": "Male", "institution": "MPI-SWS", "country": "DE", "position": "PhD student" }, { "name": "Mariya K Toneva", "gender": "Female", "institution": "Max Planck Institute for Software Systems", "country": "DE", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 35, "n_ref": 83, "n_ref_all": 131, "n_fig": 28, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 13, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 54, "formula_len_all_1": 90, "len_all": 152444, "len_all_1": 59019, "len_abs": 1125, "len_title": 123, "len_sents": 54017, "len_sents_1": 29567, "n_sents": 388, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1129, "title": "Improving Semantic Understanding in Speech Language Models via Brain-tuning", "abs": "Speech language models align with human brain responses to natural language to an impressive degree. However, current models rely heavily on low-level speech features, indicating they lack brain-relevant semantics which limits their utility as model organisms of semantic processing in the brain. In this work, we address this limitation by inducing brain-relevant bias directly into the models via fine-tuning with fMRI recordings of people listening to natural stories--a process we name brain-tuning. After testing it on 3 different pretrained model families, we show that brain-tuning not only improves overall alignment with new brain recordings in semantic language regions, but also reduces the reliance on low-level speech features for this alignment. Excitingly, we further show that brain-tuning leads to 1) consistent improvements in performance on semantic downstream tasks and 2) a representational space with increased semantic preference. Our results provide converging evidence, for the first time, that incorporating brain signals into the training of language models improves the models\u2019 semantic understanding.", "keywords": [ "fMRI", "Speech Models", "Speech Recognition", "Alignment", "Brain Alignment", "Cognitive Neuroscience", "Encoding Models", "Transformers" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "KK29oh8jZs", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Amruth Sagar", "gender": "Male", "institution": "International Institute of Information Technology, Hyderabad, International Institute of Information Technology Hyderabad", "country": "IN", "position": "MS student" }, { "name": "Ravi Kiran Sarvadevabhatla", "gender": "Male", "institution": "International Institute of Information Technology Hyderabad, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 5, "n_ref_uni": 25, "n_ref": 76, "n_ref_all": 93, "n_fig": 6, "n_tab": 3, "L_tab": 4328, "n_element_tab": 131, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 733, "formula_len_all_1": 382, "len_all": 93631, "len_all_1": 73416, "len_abs": 1761, "len_title": 128, "len_sents": 29721, "len_sents_1": 19886, "n_sents": 204, "n_sents_1": 125, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1769, "title": "IN the known, OUT of the ordinary: Probing OOD detection methods with Synthetic datasets.", "abs": "Out-of-distribution (OOD) detection is crucial for ensuring the reliability of machine learning models, especially in visual tasks. Most existing benchmarks focus on isolating distribution shifts and creating varying levels of detection difficulty, often relying on manual curation or classifier-based scoring with human annotations. Additionally, large-scale benchmarks are typically derivatives of ImageNet-21k classes or combinations of ImageNet with other datasets. However, no existing work offers a setup where only one attribute such as color or class changes in a controlled manner, while other attributes of the object remain constant. This limits our ability to precisely study the impact of individual attributes on OOD detection performance. We aim to address this by proposing two novel synthetic datasets, SHAPES and CHARS, designed to explore OOD detection under controlled and fine-grained distribution shifts. SHAPES consist of 2D and 3D geometric shapes with variations in color, size, position, and rotation, while CHARS consists of alphanumeric characters with similar variations. Each dataset presents three scenarios: (1) known classes with unseen attributes, (2) unseen classes with known attributes, and (3) entirely novel classes and attributes. We train 10 architectures and assess 13 OOD detection methods across the three scenarios, concentrating on the impact of attribute shifts on OOD scores, while also conducting additional analysis on how image corruption influences OOD scores. By systematically examining how specific attribute shifts affect OOD scores and the affects of noisy test samples, we aim to bring greater transparency to where these methods succeed or fail, helping to identify their limitations under various conditions.", "keywords": [ "OOD detection", "Benchmarking" ], "rating_list": [ 1, 3, 5, 3 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "KJzz4UwqTb", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Hyesung Jeon", "gender": "Female", "institution": "Seoul National University", "country": "", "position": "MS student" }, { "name": "Yulhwa Kim", "gender": "unknown", "institution": "Sungkyunkwan University", "country": "KR", "position": "Assistant Professor" }, { "name": "jae-joon kim", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Full Professor" } ], "n_formula": 27, "n_formula_1": 13, "n_ref_uni": 25, "n_ref": 63, "n_ref_all": 95, "n_fig": 4, "n_tab": 8, "L_tab": 3485, "n_element_tab": 825, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 172, "n_element_tab_1": 45, "formula_len_all": 1849, "formula_len_all_1": 618, "len_all": 169107, "len_all_1": 63682, "len_abs": 1904, "len_title": 128, "len_sents": 43262, "len_sents_1": 30589, "n_sents": 296, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1973, "title": "L4Q: Parameter Efficient Quantization-Aware Fine-Tuning on Large Language Models", "abs": "Due to the high memory and computational costs associated with large language models (LLMs), model compression techniques such as quantization, which reduces inference costs, and parameter-efficient fine-tuning (PEFT) methods like Low-Rank Adaptation (LoRA), which reduce training costs, have gained significant popularity. This trend has spurred active research into quantization-aware PEFT techniques, aimed at maintaining model accuracy while minimizing memory overhead during both inference and training.\nPrevious quantization-aware PEFT methods typically follow a two-step approach: first, applying post-training quantization (PTQ) to model weights, followed by PEFT on the quantized model. However, recovering from the quantization error introduced by PTQ through fine-tuning has proven challenging. Additionally, most PTQ-based PEFT methods result in a mixture of low-precision quantized weights and high-precision adapter weights, limiting the efficiency of full quantization during inference.\nWhile a previous method attempted to address these issues, it still suffers from limited adaptability due to the constrained LoRA parameter structure required to produce fully-quantized models. To overcome these challenges, we propose L4Q, a method that integrates Quantization-Aware Training (QAT) with LoRA to effectively reduce quantization error.\n%, which effectively reduces quantization error, with LoRA.\nBy employing a memory-optimized layer design, L4Q significantly reduces QAT\u2019s memory overhead while producing fully-quantized weights, enabling effective adaptation to downstream tasks. Our experiments demonstrate that this combined approach to quantization and fine-tuning achieves superior accuracy compared to decoupled fine-tuning schemes, particularly in sub-4-bit quantization, positioning L4Q as an efficient QAT solution. Using the LLaMA model families and instructional datasets, we showcase L4Q\u2019s capabilities in language tasks and few-shot learning.", "keywords": [ "LLM", "Fine-tuning", "Quantization", "PEFT", "LoRA" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "KJkbmBcZRx", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 10, "author_info_list": [ { "name": "Bin Chen", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Chunlin Fan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hang Guo", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Ke Chen", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Associate Professor" }, { "name": "Shu-Tao Xia", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Tao Dai", "gender": "Male", "institution": "Department of Software Engineering, Shenzhen University", "country": "CN", "position": "Assistant Professor" }, { "name": "Taolin Zhang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yanzi Wang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yaohua Zha", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhihao Ouyang", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 52, "n_ref": 127, "n_ref_all": 141, "n_fig": 3, "n_tab": 6, "L_tab": 3278, "n_element_tab": 408, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 4211, "n_element_tab_1": 398, "formula_len_all": 70, "formula_len_all_1": 70, "len_all": 164737, "len_all_1": 66974, "len_abs": 1695, "len_title": 125, "len_sents": 38007, "len_sents_1": 28267, "n_sents": 269, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1675, "title": "Block-to-Scene Pre-training for Point Cloud Hybrid-Domain Masked Autoencoders", "abs": "Point clouds, as a primary representation of 3D data, can be categorized into scene domain point clouds and object domain point clouds based on the modeled content. Masked autoencoders (MAE) have become the mainstream paradigm in point clouds self-supervised learning. However, existing MAE-based methods are domain-specific, limiting the model's generalization. In this paper, we propose to pre-train a general Point cloud Hybrid-Domain Masked AutoEncoder (PointHDMAE) via a block-to-scene pre-training strategy. We first propose a hybrid-domain masked autoencoder consisting of an encoder and decoder belonging to the scene domain and object domain, respectively. The object domain encoder specializes in handling object point clouds and multiple shared object encoders assist the scene domain encoder in analyzing the scene point clouds. Furthermore, we propose a block-to-scene strategy to pre-train our hybrid-domain model. Specifically, we first randomly select point blocks within a scene and apply a set of transformations to convert each point block coordinates from the scene space to the object space. Then, we employ an object-level mask and reconstruction pipeline to recover the masked points of each block, enabling the object encoder to learn a universal object representation. Finally, we introduce a scene-level block position regression pipeline, which utilizes the blocks' features in the object space to regress these blocks' initial positions within the scene space, facilitating the learning of scene representations. Extensive experiments across different datasets and tasks demonstrate the generalization and superiority of our hybrid-domain model.", "keywords": [ "Point Cloud", "Self-Supervised Pre-training" ], "rating_list": [ 6, 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 3, 5 ] }, { "paper_id": "KJLqgaixgn", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Cen LU", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Dongsheng Li", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Kaitao Song", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Xu Tan", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Principal Researcher" }, { "name": "Zihao Li", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 39, "n_ref": 82, "n_ref_all": 118, "n_fig": 9, "n_tab": 8, "L_tab": 2917, "n_element_tab": 291, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 2407, "n_element_tab_1": 224, "formula_len_all": 862, "formula_len_all_1": 342, "len_all": 190096, "len_all_1": 63304, "len_abs": 1391, "len_title": 119, "len_sents": 50427, "len_sents_1": 27901, "n_sents": 376, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1416, "title": "Sparse Training: Do All Tokens Matter for Long Sequence Generalization?", "abs": "Large language models (LLMs) have demonstrated remarkable progress in generating high-quality natural language through performing extensive pre-training over Transformer architectures. However, the quadratic complexity of transformers in sequence computation greatly limits its capability in efficiently modeling long sequences. In this paper, we introduce \\method, a simple training technique to optimize the complexity of Transformer models in long-sequence training. Specifically, in \\method, the input sequences of the Transformer network are segmented into two distinct components: {the \\textit{memory} part and the \\textit{target} part.} The target part adheres to the standard next-token prediction for modeling continuous sequences, while the memory part, sampled from longer sequences, serves as the conditional context for the prediction of the target part. To build the memory part, we apply a sparse sampling policy that decays with the distance from the target part, to obtain tokens and preserve their positions. Without any architectural modifications, our method can extend existing Transformer-based LLMs to capture long-range dependencies within a fixed window size during the training. Experimental results on multiple datasets also demonstrate the effectiveness and efficiency of \\textsc{Sparse Training} to mitigate the complexity of the Transformer network in building long-sequence dependency.", "keywords": [ "Large Language Models", "Long Sequence", "Length Extrapolation", "Efficiency" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "KJFyOwAnLR", "primary_area": "applications to neuroscience & cognitive science", "n_author": 11, "author_info_list": [ { "name": "Andreas S. Tolias", "gender": "unknown", "institution": "Baylor College of Medicine", "country": "", "position": "Professor" }, { "name": "Dongrui Deng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Katrin Franke", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Researcher" }, { "name": "Konstantin Friedrich Willeke", "gender": "Male", "institution": "University of Tuebingen", "country": "", "position": "PhD student" }, { "name": "Liam Storan", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Micha\u0142 Gerasimiuk", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Nikos Karantzas", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Nina Miolane", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "Assistant Professor" }, { "name": "Sophia Sanborn", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Researcher" }, { "name": "Xaq Pitkow", "gender": "Male", "institution": "Baylor College of Medicine", "country": "", "position": "" }, { "name": "David A. Klindt", "gender": "unknown", "institution": "Cold Spring Harbor Laboratory", "country": "", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 39, "n_ref": 69, "n_ref_all": 102, "n_fig": 20, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 17, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 224, "formula_len_all_1": 0, "len_all": 162737, "len_all_1": 73125, "len_abs": 1660, "len_title": 111, "len_sents": 58927, "len_sents_1": 39215, "n_sents": 416, "n_sents_1": 255, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1668, "title": "Emergent Geometry in Neural Representations of the Visual World", "abs": "How does the brain transform the complex visual world into meaningful representations that facilitate generalization across diverse conditions? One hypothesis is that the geometric structure of neural manifolds mirrors causal structures in the environment, facilitating strong generalization across natural contexts. The analysis of neural manifold structure has yielded neuroscientific insights in domains such as navigation and motor control, which often possess simple, low-dimensional structure. Vision, however, presents unique challenges due to its more complex, high-dimensional, hierarchical structure. Leveraging a digital twin model of primate V4 neurons, we conduct targeted in silico experiments that allow us to systematically investigate the relationship between the structure of the visual world and its encoding in the visual cortex. Our findings reveal structural equivalences between world properties and neural activity patterns for rotating objects and textures. Specifically, we demonstrate the emergence of equivariant representations that disentangle latent factors such as object identity and orientation. Finally, we demonstrate that these representations facilitate out-of-distribution generalization, as a decoder trained to linearly decode the orientation of one texture can successfully transfer to novel textures. Remarkably, artificial neural networks trained on object recognition tasks exhibit similar geometric principles. These results provide empirical support for the mirroring hypothesis in visual processing and suggest universal principles govern the formation of neural representations across biological and artificial vision.", "keywords": [ "Neuroscience", "Vision", "Geometry", "Neural Manifolds", "Deep Learning", "Convolutional Neural Networks", "Neural Representations" ], "rating_list": [ 3, 6, 3 ], "soundness_list": [ 1, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 5, 4, 5 ] }, { "paper_id": "KJF3h0OpQ7", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Deepak Ramachandran", "gender": "Male", "institution": "Google", "country": "US", "position": "Staff Researcher" }, { "name": "Fei Deng", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Feng Yang", "gender": "Male", "institution": "Google", "country": "US", "position": "Principal Engineer" }, { "name": "Kangwook Lee", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Rahul Jain", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Sahil Singla", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Tingbo Hou", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Yang Zhao", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Ying Fan", "gender": "unknown", "institution": "University of Wisconsin-Madison", "country": "US", "position": "Graduate student" }, { "name": "QIFEI WANG", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 33, "n_ref": 66, "n_ref_all": 80, "n_fig": 6, "n_tab": 5, "L_tab": 735, "n_element_tab": 76, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 598, "n_element_tab_1": 51, "formula_len_all": 1163, "formula_len_all_1": 1162, "len_all": 108253, "len_all_1": 52196, "len_abs": 1392, "len_title": 120, "len_sents": 29237, "len_sents_1": 23404, "n_sents": 228, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1399, "title": "Iterative DPO with An Improvement Model for Fine-tuning Diffusion Models", "abs": "Direct Preference Optimization (DPO) has been proven as an effective solution in aligning generative models with human preferences. However, as shown in recent works, DPO could suffer from constraints from the offline preference dataset. This paper introduces a novel improvement approach for online iterative optimization of the diffusion models without introducing extra annotation of the online data. We propose to learn a preference improvement model to extract the implicit preference from the preference dataset. The learned improvement model is then used to generate winning images from the images generated by the current diffusion model. We can construct new pairs of preference data by using images generated by the current diffusion model as losing images, and its corresponding improved images as winning images. The diffusion model can therefore be optimized via iteratively applying online preference datasets. This method enables online improvement beyond offline DPO training without requiring additional human labeling or risking overfitting the reward model. Results demonstrate improvements in preference alignment with higher diversity compared with other fine-tuning methods. Our work bridges the gap between offline preference learning and online improvement, offering a promising direction for enhancing diffusion models in image generation tasks with limited preference data.", "keywords": [ "DPO", "diffusion models" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "KJ4hQAfqVa", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 1, "author_info_list": [ { "name": "Ashish Dubey", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Principal Researcher" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 17, "n_ref": 27, "n_ref_all": 42, "n_fig": 5, "n_tab": 4, "L_tab": 492, "n_element_tab": 48, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 399, "n_element_tab_1": 40, "formula_len_all": 680, "formula_len_all_1": 680, "len_all": 66364, "len_all_1": 41534, "len_abs": 1825, "len_title": 73, "len_sents": 19328, "len_sents_1": 17116, "n_sents": 165, "n_sents_1": 140, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1410, "title": "Meta-Learning for Dynamic Synaptic Plasticity in Spiking Neural Networks", "abs": "Adaptive optimization algorithms, such as Adam Kingma & Ba (2015) and RM-SProp Tieleman & Hinton (2012), have become integral to training deep neu-ral networks, yet their stability properties and impact on generalization remain poorly understood Wilson et al. (2017). This paper extends linear stability anal-ysis to adaptive optimizers, providing a theoretical framework that explains their behavior in relation to loss surface geometry Wu et al. (2022); Jastrz\u02dbebski et al.(2019). We introduce a novel generalized coherence measure that quantifies the interaction between the adaptive preconditioner and the Hessian of the loss func-tion. This measure yields necessary and sufficient conditions for linear stability near stationary points, offering insights into why adaptive methods may converge to sharper minima with poorer generalization.\nOur analysis leads to practical guidelines for hyperparameter tuning, demon-strating how to improve the generalization performance of adaptive optimizers. Through extensive experiments on benchmark datasets and architectures, includ-ing ResNet He et al. (2016) and Vision Transformers Dosovitskiy et al. (2020), we validate our theoretical predictions, showing that aligning the adaptive precon-ditioner with the loss surface geometry through careful parameter selection can narrow the generalization gap between adaptive methods and SGD Loshchilov & Hutter (2018).", "keywords": [ "Spiking Neural Networks", "Meta-learning", "Synaptic Plasticity", "Neuromorphic Computing", "Adaptive Learning" ], "rating_list": [ 3, 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 1, 2 ], "presentation_list": [ 2, 3, 3, 2, 1 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 4, 5 ] }, { "paper_id": "KIgaAqEFHW", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Jiewen Hu", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Sean Welleck", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Hanwen Zhu", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 28, "n_ref": 79, "n_ref_all": 100, "n_fig": 6, "n_tab": 9, "L_tab": 3986, "n_element_tab": 334, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 1829, "n_element_tab_1": 275, "formula_len_all": 96, "formula_len_all_1": 49, "len_all": 150007, "len_all_1": 70164, "len_abs": 1459, "len_title": 101, "len_sents": 53620, "len_sents_1": 33473, "n_sents": 373, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 52, "L_abs": 1294, "title": "miniCTX: Neural Theorem Proving with (Long-)Contexts", "abs": "Real-world formal theorem proving often depends on a wealth of context, including definitions, lemmas, comments, file structure, and other information. We introduce $\\texttt{miniCTX}$, which tests a model's ability to prove formal mathematical theorems that depend on new context that is not seen during training. $\\texttt{miniCTX}$ contains theorems sourced from real Lean projects and textbooks, each associated with a context that can span tens of thousands of tokens. Models are tasked with proving a theorem given access to code from the theorem's repository, which contains context that is needed for the proof. As a baseline for $\\texttt{miniCTX}$, we tested fine-tuning and prompting methods that condition theorem proving on preceding context. Both approaches substantially outperform traditional methods that rely solely on state information. We found that this ability to use context is not captured by previous benchmarks such as $\\texttt{miniF2F}$. Alongside $\\texttt{miniCTX}$, we offer $\\texttt{ntp-toolkit}$ for automatically extracting and annotating theorem proving data, making it easy to add new projects into $\\texttt{miniCTX}$ to ensure that contexts are not seen during training. $\\texttt{miniCTX}$ offers a challenging and realistic evaluation of neural theorem provers.", "keywords": [ "Neural theorem proving", "Formal mathematics", "Benchmark dataset" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 4, 3, 4 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "KISgRiGCKS", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Luiz Manella Pereira", "gender": "unknown", "institution": "Florida International University", "country": "US", "position": "PhD student" }, { "name": "M. Hadi Amini", "gender": "Not Specified", "institution": "Florida International University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 27, "n_ref": 49, "n_ref_all": 63, "n_fig": 9, "n_tab": 3, "L_tab": 401, "n_element_tab": 71, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 160, "n_element_tab_1": 36, "formula_len_all": 1025, "formula_len_all_1": 985, "len_all": 105634, "len_all_1": 53832, "len_abs": 1445, "len_title": 135, "len_sents": 35512, "len_sents_1": 25078, "n_sents": 274, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1450, "title": "Optimal Transport-Based Domain Alignment as a Preprocessing Step for Federated Learning", "abs": "Federated learning is a subfield of machine learning that avoids sharing local data with a central server, which can enhance privacy and scalability. The inability to consolidate data in a central server leads to a unique problem called dataset imbalance, which is where agents in a network do not have equal representation of the labels one is trying to learn to predict. In FL, fusing locally-trained models with unbalanced datasets may deteriorate the performance of global model aggregation; this further reduces the quality of updated local models and the accuracy of the distributed agents' decisions. In this work, we introduce an Optimal Transport-based preprocessing algorithm that aligns the datasets by minimizing the distributional discrepancy of data along the edge devices without breaking privacy concerns. We accomplish this by leveraging Wasserstein barycenters when computing channel-wise averages. These barycenters are collected in a trusted central server where they collectively generate a target RGB space. By projecting our dataset towards this target space, we minimize the distributional discrepancy on a global level, which facilitates the learning process due to a minimization of variance across the samples in the analyzed network. We demonstrate the capabilities of the proposed approach over the CIFAR-10 dataset, where we show its capability of reaching higher degrees of generalization in fewer communication rounds.", "keywords": [ "Federated Learning", "Optimal Transport", "Domain Alignment", "Data Preprocessing" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 1, 2, 1, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "KI45uDnmzv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Dawei Yang", "gender": "Male", "institution": "Houmo", "country": "CN", "position": "Researcher" }, { "name": "JiangyongYu", "gender": "Male", "institution": "Houmo", "country": "CN", "position": "Engineer" }, { "name": "Sifan Zhou", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "PhD student" }, { "name": "XUCHEN", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Xing Hu", "gender": "Male", "institution": "houmo.ai", "country": "CN", "position": "Researcher" }, { "name": "Yuxuan Yue", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Zhihang Yuan", "gender": "Male", "institution": "Infinigence AI", "country": "CN", "position": "Researcher" }, { "name": "Zhixuan Chen", "gender": "Male", "institution": "houmo", "country": "CN", "position": "Researcher" }, { "name": "Zixu Jiang", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Zukang Xu", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 40, "n_formula_1": 19, "n_ref_uni": 32, "n_ref": 68, "n_ref_all": 115, "n_fig": 14, "n_tab": 7, "L_tab": 3400, "n_element_tab": 587, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 600, "n_element_tab_1": 71, "formula_len_all": 2696, "formula_len_all_1": 1003, "len_all": 135014, "len_all_1": 54488, "len_abs": 1635, "len_title": 126, "len_sents": 39976, "len_sents_1": 24225, "n_sents": 310, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1638, "title": "MambaQuant: Quantizing the Mamba Family with Variance Aligned Rotation Methods", "abs": "Mamba is an efficient sequence model that rivals Transformers and demonstrates significant potential as a foundational architecture for various tasks. Quantization is commonly used in neural networks to reduce model size and computational latency. However, applying quantization to Mamba remains underexplored, and existing quantization methods, which have been effective for CNN and Transformer models, appear inadequate for Mamba models (e.g., Quarot suffers a 21% accuracy drop on Vim-T$\\dagger$ even under W8A8). We have pioneered the exploration of this issue and identified several key challenges. First, significant outliers arepresent in gate projections, output projections, and matrix multiplications. Second, Mamba\u2019s unique parallel scan further amplifies these outliers, leading to uneven and heavy-tailed data distributions. Third, even with the application of the Hadamard transform, the variance across channels in weights and activations still remains inconsistent. To these ends, we propose MambaQuant, a post-training quantization (PTQ) framework consisting of: 1) Karhunen-Lo`eve Transformation (KLT) enhanced rotation, rendering the rotation matrix adaptable to diverse channel distributions. 2) Smooth-Fused rotation, which equalizes channel variances and can merge additional parameters into model weights. Experiments show that MambaQuant can quantize both weights and activations into 8-bit with less than 1% accuracy loss for Mamba-based vision and language tasks. To our knowledge, MambaQuant is the first comprehensive PTQ design for the Mamba family, paving the way for further advancements in its application.", "keywords": [ "Mamba", "Quantization" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "KI1zldOFz9", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Chen Hou", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Zhibo Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 38, "n_ref": 92, "n_ref_all": 114, "n_fig": 9, "n_tab": 6, "L_tab": 1725, "n_element_tab": 170, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 659, "n_element_tab_1": 113, "formula_len_all": 351, "formula_len_all_1": 326, "len_all": 144164, "len_all_1": 58461, "len_abs": 1725, "len_title": 98, "len_sents": 41892, "len_sents_1": 27672, "n_sents": 310, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1672, "title": "Training-free Camera Control for Video Generation", "abs": "We propose a training-free and robust solution to offer camera movement control for off-the-shelf video diffusion models. Unlike previous work, our method does not require any supervised finetuning on camera-annotated datasets or self-supervised training via data augmentation. Instead, it can be plug-and-play with most pretrained video diffusion models and generate camera-controllable videos with a single image or text prompt as input. The inspiration for our work comes from the layout prior that intermediate latents encode for the generated results, thus rearranging noisy pixels in them will cause the output content to relocate as well. As camera moving could also be seen as a type of pixel rearrangement caused by perspective change, videos can be reorganized following specific camera motion if their noisy latents change accordingly. Building on this, we propose **CamTrol**, which enables robust camera control for video diffusion models. It is achieved by a two-stage process. First, we model image layout rearrangement through explicit camera movement in 3D point cloud space. Second, we generate videos with camera motion by leveraging the layout prior of noisy latents formed by a series of rearranged images. Extensive experiments have demonstrated its superior performance in both video generation and camera motion alignment compared with other finetuned methods. Furthermore, we show the capability of CamTrol to generalize to various base models, as well as its impressive applications in scalable motion control, dealing with complicated trajectories and unsupervised 3D video generation.\nVideos available at https://lifedecoder.github.io/CamTrol/.", "keywords": [ "video generation", "camera control", "training-free" ], "rating_list": [ 5, 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 4, 2, 2 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 5, 5 ] }, { "paper_id": "KHkBpvmYVI", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 12, "author_info_list": [ { "name": "Cheng Tan", "gender": "Male", "institution": "Zhejiang University & Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Haitao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Lirong Wu", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Odin Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "Undergrad student" }, { "name": "Siyuan Li", "gender": "Male", "institution": "Westlake University & Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Tailin Wu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yufei Huang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yunfan Liu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Yunshu Liu", "gender": "Female", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Zhangyang Gao", "gender": "Male", "institution": "Westlake University, China", "country": "CN", "position": "PhD student" }, { "name": "Zicheng Liu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Stan Z. Li", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Chair Professor" } ], "n_formula": 27, "n_formula_1": 23, "n_ref_uni": 24, "n_ref": 112, "n_ref_all": 134, "n_fig": 8, "n_tab": 13, "L_tab": 1293, "n_element_tab": 291, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1619, "n_element_tab_1": 74, "formula_len_all": 1781, "formula_len_all_1": 1639, "len_all": 170840, "len_all_1": 75894, "len_abs": 1557, "len_title": 74, "len_sents": 56842, "len_sents_1": 34107, "n_sents": 422, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1583, "title": "EVA: Geometric Inverse Design for Fast Protein Motif-Scaffolding with Coupled Flow", "abs": "Motif-scaffolding is a fundamental component of protein design, which aims to construct the scaffold structure that stabilizes motifs conferring desired functions. Recent advances in generative models are promising for designing scaffolds, with two main approaches: training-based and sampling-based methods. Training-based methods are resource-heavy and slow, while training-free sampling-based methods are flexible but require numerous sampling steps and costly, unstable guidance. To speed up and improve sampling-based methods, we analyzed failure cases and found that errors stem from the trade-off between generation and guidance. Thus we proposed to exploit the spatial context and adjust the generative direction to be consistent with guidance to overcome this trade-off. Motivated by this, we formulate motif-scaffolding as a Geometric Inverse Design task inspired by the image inverse problem, and present Evolution-ViA-reconstruction (EVA), a novel sampling-based coupled flow framework on geometric manifolds, which starts with a pretrained flow-based generative model. EVA uses motif-coupled priors to leverage spatial contexts, guiding the generative process along a straighter probability path, with generative directions aligned with guidance in the early sampling steps. EVA is 70\u00d7 faster than SOTA model RFDiffusion with competitive and even better performance on benchmark tests. Further experiments on real-world cases including vaccine design, multi-motif scaffolding and motif optimal placement searching demonstrate EVA's superior efficiency and effectiveness.", "keywords": [ "Generative Model", "Protein Structure Generation", "Training-free Conditional Generation" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 3, 2, 3, 1, 2 ] }, { "paper_id": "KHTkRhq2aB", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bin Bi", "gender": "unknown", "institution": "SalesForce.com", "country": "US", "position": "Principal Researcher" }, { "name": "Claire Na Cheng", "gender": "unknown", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Kiran Ramnath", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Regunathan Radhakrishnan", "gender": "unknown", "institution": "SalesForce.com", "country": "US", "position": "Principal Researcher" }, { "name": "Shiva Pentyala", "gender": "Male", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "Sitaram Asur", "gender": "unknown", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Zhichao Wang", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Xiangbo Mao", "gender": "Male", "institution": "Salesforce", "country": "CA", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 36, "n_ref": 46, "n_ref_all": 63, "n_fig": 2, "n_tab": 4, "L_tab": 4040, "n_element_tab": 407, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 4654, "n_element_tab_1": 413, "formula_len_all": 79, "formula_len_all_1": 79, "len_all": 126167, "len_all_1": 60873, "len_abs": 1252, "len_title": 113, "len_sents": 29392, "len_sents_1": 26573, "n_sents": 188, "n_sents_1": 150, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1256, "title": "PAFT: A Parallel Training Paradigm for Effective LLM Fine-Tuning", "abs": "Large language models (LLMs) have shown remarkable abilities in diverse natural language processing (NLP) tasks. The LLMs generally undergo supervised fine-tuning (SFT) followed by preference alignment to be usable in downstream applications. However, this sequential training pipeline leads to alignment tax that degrades the LLM performance.\n\nThis paper introduces PAFT, a new PArallel training paradigm for effective LLM Fine-Tuning, which independently performs SFT and preference alignment (e.g., DPO and ORPO, etc.) with the same pre-trained model on respective datasets. The model produced by SFT and the model from preference alignment are then merged into a final model by parameter fusing for use in downstream applications. This work reveals important findings that preference alignment like DPO naturally results in a sparse model while SFT leads to a natural dense model which needs to be sparsified for effective model merging. This paper introduces an effective interference resolution which reduces the redundancy by sparsifying the delta parameters. The LLM resulted from the new training paradigm achieved Rank #1 on the HuggingFace Open LLM Leaderboard. Comprehensive evaluation shows the effectiveness of the parallel training paradigm.", "keywords": [ "LLM", "fine-tuning", "alignment", "SFT" ], "rating_list": [ 8, 8, 3, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "KGhkIySg3h", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Francesco Tonolini", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Gabriella Kazai", "gender": "Female", "institution": "Amazon", "country": "", "position": "Principal Researcher" }, { "name": "Jordan Massiah", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Research Engineer" }, { "name": "Nikolaos Aletras", "gender": "unknown", "institution": "University of Sheffield", "country": "", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 9, "n_ref_uni": 55, "n_ref": 111, "n_ref_all": 130, "n_fig": 6, "n_tab": 4, "L_tab": 1744, "n_element_tab": 265, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1603, "n_element_tab_1": 187, "formula_len_all": 1489, "formula_len_all_1": 736, "len_all": 160461, "len_all_1": 63313, "len_abs": 1072, "len_title": 105, "len_sents": 44097, "len_sents_1": 28978, "n_sents": 291, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1073, "title": "Multi-Fidelity Fine-Tuning of Pre-Trained Language Models", "abs": "We consider the problem of fine-tuning pre-trained language models with a small amount of trusted data (high-fidelity) and a larger amount of data with noisy labels (low-fidelity). We propose Multi-Fidelity Fine-Tuning (MFFT), a novel approach which implicitly determines for new inputs when we can rely on information from high-fidelity data and when instead we need to fall back on knowledge from low-fidelity data. MFFT does not require any architecture changes to the base model and simply provides its fine-tuned version that can be easily deployed for inference. We extensively benchmark MFFT on various classification tasks against several baselines, with both simulated label noise, and in realistic scenarios with LLM generated data. MFFT consistently improves performance compared to using trusted data alone and outperforms all baselines across experiments with macro F1-score improvements of 2-4%. Finally, it provides substantial improvements in uncertainty calibration with expected calibration error (ECE) reductions of 40-60% compared to the best baselines.", "keywords": [ "language models", "foundation models", "multi-fidelity", "learning with noisy labels", "fine-tuning" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 3, 2, 1, 4 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 3, 5, 2 ] }, { "paper_id": "KGZAs8VcOM", "primary_area": "generative models", "n_author": 11, "author_info_list": [ { "name": "Chi Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Di Huang", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Gang Yu", "gender": "Male", "institution": "StepFun", "country": "CN", "position": "Principal Researcher" }, { "name": "Jiaxiang Tang", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Intern" }, { "name": "Lei Yang", "gender": "Male", "institution": "Sensetime Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Sijin Chen", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Tong He", "gender": "Male", "institution": "Shanghai AI lab", "country": "CN", "position": "Researcher" }, { "name": "Weicai Ye", "gender": "Male", "institution": "KwaiVGI, Kuaishou Technology", "country": "CN", "position": "Researcher" }, { "name": "Yiwen Chen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Zhongang Cai", "gender": "Male", "institution": "SenseTime International Pte Ltd", "country": "SG", "position": "Senior Research Scientist" }, { "name": "Guosheng Lin", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 2, "n_ref_uni": 57, "n_ref": 200, "n_ref_all": 217, "n_fig": 8, "n_tab": 7, "L_tab": 2186, "n_element_tab": 231, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 157, "n_element_tab_1": 21, "formula_len_all": 198, "formula_len_all_1": 50, "len_all": 188107, "len_all_1": 60379, "len_abs": 315, "len_title": 0, "len_sents": 47780, "len_sents_1": 29280, "n_sents": 340, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1504, "title": "MeshAnything: Artist-Created Mesh Generation with Autoregressive Transformers", "abs": "Recently, 3D assets created via reconstruction and generation have matched the quality of manually crafted assets, highlighting their potential for replacement. However, this potential is largely unrealized because these assets always need to be converted to meshes for 3D industry applications, and the meshes produced by current mesh extraction methods are significantly inferior to Artist-Created Meshes (AMs), i.e., meshes created by human artists. \n\nSpecifically, current mesh extraction methods rely on dense faces and ignore geometric features, leading to inefficiencies, complicated post-processing, and lower representation quality.\nTo address these issues, we introduce MeshAnything, a model that treats mesh extraction as a generation problem, producing AMs aligned with specified shapes.\n\nBy converting 3D assets in any 3D representation into AMs, MeshAnything can be integrated with various 3D asset production methods, thereby enhancing their application across the 3D industry.\nThe architecture of MeshAnything comprises a VQ-VAE and a shape-conditioned decoder-only transformer. We first learn a mesh vocabulary using the VQ-VAE, then train the shape-conditioned decoder-only transformer on this vocabulary for shape-conditioned autoregressive mesh generation. Our extensive experiments show that our method generates AMs with hundreds of times fewer faces, significantly improving storage, rendering, and simulation efficiencies, while achieving precision comparable to previous methods.", "keywords": [ "3D Generation", "Transformers", "Sequence Learning" ], "rating_list": [ 8, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "KGRV73Zcqt", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Ashish Kumar", "gender": "Male", "institution": "Indian Institute of Technology, Kanpur", "country": "", "position": "PhD student" }, { "name": "Jaesik Park", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 18, "n_ref": 82, "n_ref_all": 106, "n_fig": 6, "n_tab": 7, "L_tab": 6862, "n_element_tab": 671, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 5752, "n_element_tab_1": 475, "formula_len_all": 25, "formula_len_all_1": 140, "len_all": 120933, "len_all_1": 73103, "len_abs": 1134, "len_title": 109, "len_sents": 31721, "len_sents_1": 27601, "n_sents": 278, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1140, "title": "Cross Resolution Encoding-Decoding For Detection Transformers", "abs": "Detection Transformers (DETR) are renowned object detection pipelines, however\ncomputationally efficient multiscale detection using DETR is still challenging. In\nthis paper, we propose a Cross-Resolution Encoding-Decoding (CRED) mechanism\nthat allows DETR to achieve the accuracy of high-resolution detection while\nhaving the speed of low-resolution detection. CRED is based on two modules;\nCross Resolution Attention Module (CRAM) and One Step Multiscale Attention\n(OSMA). CRAM is designed to transfer the knowledge of low-resolution encoder\noutput to a high-resolution feature. While OSMA is designed to fuse multiscale\nfeatures in a single step and produce a feature map of a desired resolution enriched\nwith multiscale information. When used in prominent DETR methods, CRED\ndelivers accuracy similar to the high-resolution DETR counterpart in roughly 50%\nfewer FLOPs. Specifically, state-of-the-art DN-DETR, when used with CRED\n(calling CRED-DETR), becomes 76% faster, with \u223c 50% reduced FLOPs than its\nhigh-resolution counterpart with 202 G FLOPs on MS-COCO benchmark. We plan\nto release pretrained CRED-DETRs for use by the community.", "keywords": [ "Cross Resolution", "Encoding and Decoding", "DETR", "Detection" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "KFLtFSOtdj", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Jin Xu", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Junyang Lin", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Wenrui Liu", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yuanjun Lv", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "", "position": "MS student" }, { "name": "Yunfei Chu", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zhifang Guo", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 105, "n_ref_all": 128, "n_fig": 6, "n_tab": 8, "L_tab": 2697, "n_element_tab": 305, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2181, "n_element_tab_1": 194, "formula_len_all": 357, "formula_len_all_1": 304, "len_all": 123513, "len_all_1": 60738, "len_abs": 1256, "len_title": 145, "len_sents": 39520, "len_sents_1": 27913, "n_sents": 245, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1328, "title": "Analyzing and Mitigating Inconsistency in Discrete Audio Tokens for Neural Codec Language Models", "abs": "Building upon advancements in Large Language Models (LLMs), the field of audio processing has seen increased interest in training audio generation tasks with discrete audio token sequences. However, directly discretizing audio by\nneural audio codecs often results in sequences that fundamentally differ from text sequences. Unlike text, where text token sequences are deterministic, discrete audio tokens can exhibit significant variability based on contextual factors, while still producing perceptually identical audio segments. We refer to this phenomenon as \\textbf{Discrete Representation Inconsistency (DRI)}. This inconsistency can lead to a single audio segment being represented by multiple divergent sequences, which creates confusion in neural codec language models and results in omissions and repetitions during speech generation. In this paper, we quantitatively analyze the DRI phenomenon within popular audio tokenizers such as EnCodec. Our approach effectively mitigates the DRI phenomenon of the neural audio codec. Furthermore, extensive experiments on the neural codec language model over LibriTTS and large-scale MLS datasets (44,000 hours) demonstrate the effectiveness and generality of our method. The demo of audio samples is available online~\\footnote{\\url{https://consistencyinneuralcodec.github.io}}.", "keywords": [ "TTS", "speech generation", "neural audio codec", "neural codec lanugage model" ], "rating_list": [ 6, 1, 5, 6, 5 ], "soundness_list": [ 3, 1, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 3, 1, 3, 2, 2 ], "confidence_list": [ 3, 5, 3, 3, 5 ] }, { "paper_id": "KEqerdxOBG", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Hao Sun", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Lu Yin", "gender": "unknown", "institution": "University of Surrey", "country": "GB", "position": "Assistant Professor" }, { "name": "Wei Wang", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Yang Liu", "gender": "Female", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 7, "n_ref_uni": 38, "n_ref": 60, "n_ref_all": 92, "n_fig": 4, "n_tab": 7, "L_tab": 3190, "n_element_tab": 145, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 216, "n_element_tab_1": 36, "formula_len_all": 3341, "formula_len_all_1": 895, "len_all": 128406, "len_all_1": 61201, "len_abs": 1113, "len_title": 125, "len_sents": 37793, "len_sents_1": 27945, "n_sents": 309, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1115, "title": "MRFNet: Multi-Receptive Field Network for Multivariate Time-Series Prediction", "abs": "Time series forecasting is a critical topic in machine learning. Although existing deep learning methods have demonstrated outstanding performance and currently dominate this field, the latest state-of-the-art (SOTA) models are increasingly encountering the same limitations: the blockneck of performance. We believe this convergence is due to these models being based on the same mathematical foundations. To address this issue, we draw inspiration from the universal approximation theorem (UAT) and show that most commonly used deep learning models for time series forecasting are specific implementations of UAT. Based on UAT theory and the characteristics of time series data, we propose a new forecasting model called the Multi-Receptive Field Network (MRFNet). This architecture integrates linear, sparse matrix, convolutional, and Fourier transform modules, resulting in an interpretable model with multiple receptive fields that can capture both global and local information. The MRFNet model has been tested extensively on several popular time series forecasting datasets and has achieved superior results.", "keywords": [ "time series forecasting" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "KEeTRb8GLf", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Aviraj Newatia", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" }, { "name": "Michael Cooper", "gender": "unknown", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Rahul Krishnan", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 25, "n_ref": 40, "n_ref_all": 57, "n_fig": 3, "n_tab": 4, "L_tab": 37, "n_element_tab": 2, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1954, "n_element_tab_1": 235, "formula_len_all": 2110, "formula_len_all_1": 2028, "len_all": 100577, "len_all_1": 71679, "len_abs": 1186, "len_title": 97, "len_sents": 31982, "len_sents_1": 30333, "n_sents": 227, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1189, "title": "Blind Unlearning: Unlearning Without a Forget Set", "abs": "Machine unlearning is the study of methods to efficiently remove the influence\nof some subset of the training data from the parameters of a previously-trained\nmodel. Existing methods typically require direct access to the \u201cforget set\u201d \u2013 the\nsubset of training data to be forgotten by the model. This limitation impedes privacy, as organizations need to retain user data for the sake of unlearning when a\nrequest for deletion is made, rather than being able to delete it immediately. We\nfirst introduce the setting of blind unlearning \u2013 unlearning without explicit access\nto the forget set. Then, we propose a method for approximate unlearning called\nRELOAD, that leverages ideas from gradient-based unlearning and neural network\nsparsity to achieve blind unlearning. The method serially applies an ascent step\nwith targeted parameter re-initialization and fine-tuning, and on empirical unlearning tasks, RELOAD often approximates the behaviour of a from-scratch retrained\nmodel better than approaches that leverage the forget set. Finally, we extend the\nblind unlearning setting to blind remedial learning, the task of efficiently updating\na previously-trained model to an amended dataset.", "keywords": [ "machine unlearning", "data privacy" ], "rating_list": [ 3, 3, 6, 3, 3 ], "soundness_list": [ 2, 2, 3, 2, 1 ], "presentation_list": [ 3, 2, 3, 1, 2 ], "contribution_list": [ 1, 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "KEXoZxTwbr", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chen Qian", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Dianbing Xi", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jingsen Zhu", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Qi Wang", "gender": "Male", "institution": "College of Computer Science and Technology, Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Ying He", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Yuchi Huo", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuxin Dai", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" } ], "n_formula": 28, "n_formula_1": 18, "n_ref_uni": 18, "n_ref": 43, "n_ref_all": 86, "n_fig": 12, "n_tab": 5, "L_tab": 658, "n_element_tab": 102, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 698, "n_element_tab_1": 55, "formula_len_all": 989, "formula_len_all_1": 752, "len_all": 132480, "len_all_1": 51766, "len_abs": 1393, "len_title": 117, "len_sents": 35923, "len_sents_1": 23379, "n_sents": 263, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 117, "L_abs": 1360, "title": "Inverse Rendering for Shape, Light, and Material Decomposition using Multi-Bounce Path Tracing and Reservoir Sampling", "abs": "We present a novel two-stage inverse rendering framework that jointly reconstructs and optimizes explicit geometry, materials, and lighting from multi-view images. \nUnlike previous methods that rely on implicit irradiance fields or oversimplified path tracing algorithms, our method first extracts an explicit triangular mesh in the initial stage. \nSubsequently, it employs a more realistic physically-based inverse rendering model in the second stage, utilizing multi-bounce path tracing and Monte Carlo integration. \nBy leveraging multi-bounce path tracing, our method not only effectively estimates indirect illumination (including self-shadowing and internal reflections). but also enhances the intrinsic decomposition of shape,\nmaterial, and lighting. Moreover, we incorporate reservoir sampling into our framework to address the noise in Monte Carlo integration, enhancing convergence and facilitating gradient-based optimization with low sample counts. \nThrough both qualitative and quantitative assessments across various scenarios, especially those with complex shadows, we demonstrate that our method achieves state-of-the-art performance in decomposition results.\nAdditionally, our optimized explicit geometry supports further applications in scene editing, relighting, and material editing, compatible with modern graphics engines and CAD software.", "keywords": [ "Inverse Rendering", "Material Decomposition", "3D reconstruction", "relighting" ], "rating_list": [ 6, 8, 3, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 2, 4 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "KEUPk0wXXe", "primary_area": "interpretability and explainable AI", "n_author": 9, "author_info_list": [ { "name": "Ji Cao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yang Jie", "gender": "Male", "institution": "Zhejiang University", "country": "AL", "position": "MS student" }, { "name": "Mingli Song", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Shunyu Liu", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Tongya Zheng", "gender": "Male", "institution": "Hangzhou City University", "country": "CN", "position": "Postdoc" }, { "name": "Yihe Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yuwen Wang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zhenbang Xiao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Kai-Xuan Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" } ], "n_formula": 26, "n_formula_1": 25, "n_ref_uni": 61, "n_ref": 95, "n_ref_all": 131, "n_fig": 18, "n_tab": 5, "L_tab": 3804, "n_element_tab": 157, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 882, "n_element_tab_1": 20, "formula_len_all": 1234, "formula_len_all_1": 1172, "len_all": 196574, "len_all_1": 61444, "len_abs": 1850, "len_title": 126, "len_sents": 59216, "len_sents_1": 29101, "n_sents": 516, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1820, "title": "From GNNs to Trees: Multi-Granular Interpretability for Graph Neural Networks", "abs": "Interpretable Graph Neural Networks (GNNs) aim to reveal the underlying reasoning behind model predictions, attributing their decisions to specific subgraphs that are informative. However, existing subgraph-based interpretable methods suffer from an overemphasis on local structure, potentially overlooking long-range dependencies within the entire graphs. Although recent efforts that rely on graph coarsening have proven beneficial for global interpretability, they inevitably reduce the graphs to a fixed granularity. Such an inflexible way can only capture graph connectivity at a specific level, whereas real-world graph tasks often exhibit relationships at varying granularities (e.g., relevant interactions in proteins span from functional groups, to amino acids, and up to protein domains). In this paper, we introduce a novel Tree-like Interpretable Framework (TIF) for graph classification, where plain GNNs are transformed into hierarchical trees, with each level featuring coarsened graphs of different granularity as tree nodes. Specifically, TIF iteratively adopts a graph coarsening module to compress original graphs (i.e., root nodes of trees) into increasingly coarser ones (i.e., child nodes of trees), while preserving diversity among tree nodes within different branches through a dedicated graph perturbation module. Finally, we propose an adaptive routing module to identify the most informative root-to-leaf paths, providing not only the final prediction but also the multi-granular interpretability for the decision-making process. Extensive experiments on the graph classification benchmarks with both synthetic and real-world datasets demonstrate the superiority of TIF in interpretability, while also delivering a competitive prediction performance akin to the state-of-the-art counterparts.", "keywords": [ "Graph Neural Networks", "Multi-Granular Interpretability", "Graph Classification" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "KDXj60FpJr", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Afreen Shaikh", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Graham Neubig", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Jennifer Hsia", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zhiruo Wang", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 25, "n_ref": 60, "n_ref_all": 93, "n_fig": 12, "n_tab": 5, "L_tab": 899, "n_element_tab": 171, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 618, "formula_len_all_1": 271, "len_all": 129887, "len_all_1": 56341, "len_abs": 963, "len_title": 73, "len_sents": 39893, "len_sents_1": 28368, "n_sents": 275, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 970, "title": "RAGGED: Towards Informed Design of Retrieval Augmented Generation Systems", "abs": "Retrieval-augmented generation (RAG) systems have shown promise in improving\ntask performance by leveraging external context, but realizing their full potential\ndepends on careful configuration. In this paper, we investigate how the choice of\nretriever and reader models, context length, and context quality impact RAG per-\nformance across different task types. Our findings reveal that while some readers\nconsistently benefit from additional context, others degrade when exposed to irrele-\nvant information, highlighting the need for tuning based on reader sensitivity to\nnoise. Moreover, retriever improvements do not always translate into proportional\ngains in reader results, particularly in open-domain questions. However, in spe-\ncialized tasks, even small improvements in retrieval can significantly boost reader\nresults. These insights underscore the importance of optimizing RAG systems by\naligning configurations with task complexity and domain-specific needs.", "keywords": [ "Retrieval Augmented Generation", "Language Modeling", "Question Answering" ], "rating_list": [ 6, 3, 3, 8 ], "soundness_list": [ 2, 2, 2, 4 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "KDGP8yAz5b", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Hang Yan", "gender": "Male", "institution": "AI lab", "country": "", "position": "Researcher" }, { "name": "Kai Lv", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Linlin Li", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Qipeng Guo", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "", "position": "Researcher" }, { "name": "Qun Liu", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "HK", "position": "Chief Scientist of Speech and Language Computing" }, { "name": "Ruixiao Li", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Xiaoran Liu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Xipeng Qiu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yuerong Song", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhigeng Liu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 28, "n_ref": 125, "n_ref_all": 147, "n_fig": 12, "n_tab": 3, "L_tab": 559, "n_element_tab": 24, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 472, "n_element_tab_1": 54, "formula_len_all": 450, "formula_len_all_1": 739, "len_all": 128883, "len_all_1": 50669, "len_abs": 1399, "len_title": 119, "len_sents": 36073, "len_sents_1": 24486, "n_sents": 245, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1422, "title": "ReAttention: Training-Free Infinite Context with Finite Attention Scope", "abs": "The long-context capability of the Large Language Models (LLM) has made significant breakthroughs, but the maximum supported context length remains a critical bottleneck limiting their practical applications. The constraint of context length in LLMs arises from the self-attention mechanism, which cannot effectively and efficiently capture the semantic relationships within infinitely long contexts via the limited pre-trained positional information and attention scope. In this work, we propose \\textbf{ReAttention}, a training-free approach enabling LLM based on the self-attention mechanism to support an infinite context with a finite attention scope under sufficient memory resources. ReAttention performs the position-agnostic top-$k$ attention before the ordinary position-aware self-attention, freeing LLMs from the length extrapolation issue. We validate the performance of ReAttention on the LongBench, L-Eval, and InfiniteBench and demonstrate that it is on par with traditional methods. Furthermore, we also apply ReAttention on mainstream LLMs, including LLaMA3.1-8B and Mistral-v0.3-7B, enabling them to support context lengths of at least 1M and even expanding the context length of LLaMA3.2-3B-chat by 128$\\times$ to 4M without any further training in Needle-In-A-Haystack tests. We also improve the efficiency of ReAttention with Triton and achieve an efficient extrapolation without additional overhead.", "keywords": [ "long context", "length extrapolation", "large language model", "natural language processing" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 5, 3, 3, 5 ] }, { "paper_id": "KD9F5Ap878", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Ruoyu Sun", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Tian Ding", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Researcher" }, { "name": "Yizhou Han", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Yufeng Lin", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Undergrad student" }, { "name": "Zhihang Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "MS student" }, { "name": "Zhuohan Wang", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Ziwei Zhu", "gender": "Female", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 23, "n_ref": 41, "n_ref_all": 51, "n_fig": 5, "n_tab": 3, "L_tab": 770, "n_element_tab": 67, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1964, "n_element_tab_1": 83, "formula_len_all": 464, "formula_len_all_1": 410, "len_all": 110309, "len_all_1": 54639, "len_abs": 1795, "len_title": 156, "len_sents": 31617, "len_sents_1": 27093, "n_sents": 193, "n_sents_1": 153, "venue": "ICLR2025", "decision": "withdraw", "L_title": 108, "L_abs": 1808, "title": "OptiBench: Benchmarking Large Language Models in Optimization Modeling with Equivalence-Detection Evaluation", "abs": "In operations research (OR), formulating optimization problems in industrial applications is often time-consuming and requires specialized expertise. Recently, large language models (LLMs) have shown remarkable potential to automate this process. However, evaluating the performance of LLMs in optimization modeling remains challenging due to the scarcity of suitable datasets and rigorous evaluation methodologies. To reduce this gap, we introduce OptiBench, a new benchmark designed to assess LLMs' ability to formulate linear programming (LP) and mixed-integer linear programming (MILP) models. OptiBench provides a diverse dataset covering 816 optimization modeling word problems across 16 problem classes and over 80 practical domains. It also adopts a model-data separation format with 2 levels of description abstraction. The dataset exhibits the complexity of real-world optimization problems compared to traditional textbook examples. OptiBench incorporates a new evaluation method based on a modified Weisfeiler-Lehman graph isomorphism test (WL-test) algorithm.\nWe theoretically prove that this method can correctly judge whether two models are equivalent or not, setting a new standard for automatically validating the correctness of optimization modeling. We benchmark various LLMs using OptiBench and observe significant performance differences. GPT-4o by direct prompting achieves 49.39\\% overall accuracy, outperforming other models and LLM-based agents, including OpenAI o1 (preview and mini). Notably, GPT-4o's performance varies across different problem classes, achieving over 90\\% accuracy on the knapsack problem class but falling below 5\\% on the traveling salesman problem class. These findings provide new insights into the strengths and limitations of LLMs in optimization modeling.", "keywords": [ "LLM", "benchmark", "AI for OR", "optimization modeling", "autonomous mathematical formulation" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 2, 4, 3 ], "presentation_list": [ 2, 4, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 4, 4, 5 ] }, { "paper_id": "KD5nJUgeW4", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Dongbin Zhao", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Runyu Lu", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Yuanheng Zhu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Associate Professor" } ], "n_formula": 99, "n_formula_1": 28, "n_ref_uni": 40, "n_ref": 83, "n_ref_all": 103, "n_fig": 5, "n_tab": 2, "L_tab": 86, "n_element_tab": 17, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 638, "n_element_tab_1": 277, "formula_len_all": 14729, "formula_len_all_1": 3482, "len_all": 186217, "len_all_1": 75691, "len_abs": 1701, "len_title": 154, "len_sents": 54534, "len_sents_1": 31234, "n_sents": 484, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 106, "L_abs": 1707, "title": "Solving Multiplayer Partially Observable Stochastic Games by Divergence-Regularized Discounted Aggregation", "abs": "This paper presents Divergence-Regularized Discounted Aggregation (DRDA), a multi-round learning system for solving partially observable stochastic games (POSGs), which unify normal-form games (NFGs), extensive-form games (EFGs) with perfect recall, and Markov games (MGs). In each single round, DRDA can be viewed as a discounted variant of Follow the Regularized Leader (FTRL) under a general value function for POSGs concerning imperfect information and an infinite horizon. While previous studies on this FTRL variant have demonstrated its last-iterate convergence towards quantal response equilibrium (QRE) in NFGs, this paper extends the theoretical results to POSGs by defining a generalized Nash distribution (GND), which extends the QRE concept of Nash distribution in NFGs through divergence regularization. The linear last-iterate convergence of single-round DRDA to its rest point is proved under a general assumption of hypomonotonicity. When the rest point is unique, it induces the unique GND, which has a bounded deviation with respect to Nash equilibrium (NE). Under multiple learning rounds, DRDA keeps replacing the base policy for divergence regularization with the policy at the rest point in the previous round. It is further proved that the limit point of multi-round DRDA must be an exact NE rather than a QRE under the unique rest point assumption. In experiments, the last iterates of multi-round DRDA converge to NE at a near-exponential rate in NFGs, outperforming existing baselines including moving-magnet magnetic mirror descent (MMD) in multiplayer EFGs. In an infinite-horizon MG, DRDA significantly outperforms the applicable algorithms based on best-response computations.", "keywords": [ "partially observable stochastic game", "Nash distribution", "divergence regularization", "hypomonotonicity", "last-iterate convergence", "Nash equilibrium" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "KCr2eoSiKF", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Kshitij Kayastha", "gender": "Male", "institution": "Drexel University", "country": "US", "position": "PhD student" }, { "name": "Shahin Jabbari", "gender": "unknown", "institution": "Drexel University", "country": "US", "position": "Assistant Professor" }, { "name": "Vasilis Gkatzelis", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 42, "n_formula_1": 9, "n_ref_uni": 60, "n_ref": 126, "n_ref_all": 169, "n_fig": 14, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3820, "formula_len_all_1": 1105, "len_all": 180778, "len_all_1": 58682, "len_abs": 1132, "len_title": 94, "len_sents": 58159, "len_sents_1": 28173, "n_sents": 414, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1136, "title": "Learning-Augmented Robust Algorithmic Recourse", "abs": "The widespread use of machine learning models in high-stakes domains can have a major negative impact, especially on individuals who receive undesirable outcomes. Algorithmic recourse provides such individuals with suggestions of minimum-cost improvements they can make to achieve a desirable outcome in the future. However, machine learning models often get updated over time and this can cause a recourse to become invalid (i.e., not lead to the desirable outcome). The robust recourse literature aims to choose recourses less sensitive, even against adversarial model changes, but this comes at a higher cost. To overcome this obstacle, we initiate the study of algorithmic recourse through the learning-augmented framework and evaluate the extent to which a designer equipped with a prediction regarding future model changes can reduce the cost of recourse when the prediction is accurate (consistency) while also limiting the cost even when the prediction is inaccurate (robustness). We propose a novel algorithm for this problem, study the robustness-consistency trade-off, and analyze how prediction accuracy affects performance.", "keywords": [ "Algorithmic Recourse", "Learning-Augmented", "AI Explainability" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 1, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "KCYDpqSpqg", "primary_area": "generative models", "n_author": 12, "author_info_list": [ { "name": "Di ZHANG", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "VP" }, { "name": "Guangyong Chen", "gender": "Male", "institution": "Zhejiang Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Guibao Shen", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Luozhou Wang", "gender": "Male", "institution": "Hong Kong University of Science and Technology (GUANGZHOU)", "country": "CN", "position": "PhD student" }, { "name": "Pengfei Wan", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Director" }, { "name": "Wenhang Ge", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Xin Tao", "gender": "Male", "institution": "Kuaishou", "country": "CN", "position": "Researcher" }, { "name": "Ying-Cong Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Ziyang Mai", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "CHAOZHE ZHANG", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "MS student" }, { "name": "jiantaoLin", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Yijun Li", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 33, "n_ref": 99, "n_ref_all": 115, "n_fig": 13, "n_tab": 4, "L_tab": 976, "n_element_tab": 110, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 736, "n_element_tab_1": 98, "formula_len_all": 355, "formula_len_all_1": 355, "len_all": 126642, "len_all_1": 56684, "len_abs": 1427, "len_title": 121, "len_sents": 39571, "len_sents_1": 26013, "n_sents": 298, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1378, "title": "SG-Adapter: Enhancing Text-to-Image Generation with Scene Graph Guidance", "abs": "Recent advancements in text-to-image generation have been propelled by the development of diffusion models and multi-modality learning. However, since text is typically represented sequentially in these models, it often falls short in providing accurate contextualization and structural control. So the generated images do not consistently align with human expectations, especially in complex scenarios involving multiple objects and relationships. In this paper, we introduce the Scene Graph Adapter(SG-Adapter), leveraging the structured representation of scene graphs to rectify inaccuracies in the original text embeddings. The SG-Adapter's explicit and non-fully connected graph representation greatly improves the fully connected, transformer-based text representations. This enhancement is particularly notable in maintaining precise correspondence in scenarios involving multiple relationships. To address the challenges posed by low-quality annotated datasets like Visual Genome, we have manually curated a highly clean, multi-relational scene graph-image paired dataset MultiRels. Furthermore, we design three metrics derived from GPT-4V to effectively and thoroughly measure the correspondence between images and scene graphs. Both qualitative and quantitative results validate the efficacy of our approach in controlling the correspondence in multiple relationships.", "keywords": [ "diffusion models", "scene graph", "image generation" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 1 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "KCVv3tICvp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Aakash Lahoti", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Albert Gu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Chris Donahue", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Karan Goel", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Shih-Lun Wu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Arjun Desai", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 30, "n_ref": 96, "n_ref_all": 118, "n_fig": 3, "n_tab": 2, "L_tab": 1008, "n_element_tab": 109, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1010, "n_element_tab_1": 109, "formula_len_all": 365, "formula_len_all_1": 364, "len_all": 135514, "len_all_1": 71774, "len_abs": 1296, "len_title": 107, "len_sents": 40149, "len_sents_1": 33647, "n_sents": 284, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1260, "title": "Towards Codec-LM Co-design for Neural Codec Language Models", "abs": "Neural codec language models (or _codec LMs_) are emerging as a powerful framework for text-to-speech (TTS) and other audio generation tasks. These models leverage advancements in language modeling and high-fidelity residual vector quantization (RVQ)-based audio codecs, which compress continuous waveforms into discrete codes for LMs to process. Despite the close interdependence of codecs and LMs in these systems, research on codecs and LMs has largely remained siloed. In this work, we bridge this gap by proposing several codec-LM co-design strategies, analyzing their effects on end-to-end TTS performance and efficiency. Specifically, we introduce three complementary techniques: (i) a _frame-wise codec encoder_ that improves both LM log-likelihood and end-to-end TTS metrics, (ii) _LM codebook level dropout_, a method to efficiently navigate a portion of the codec-LM design space by training a single LM, and (iii) _increased codec frame duration_, which we show can accelerate inference while maintaining end-to-end performance. Our experiments demonstrate that combining all three co-design techniques results in doubled inference speed, and improvements in intelligibility, audio quality, and speaker control in TTS relative to a siloed baseline.", "keywords": [ "text-to-speech", "audio generation", "neural audio codecs", "codec LMs", "co-design" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "KCTHM2Ffh3", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Hongpeng Cao", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Lui Sha", "gender": "Male", "institution": "Department of Computer Science", "country": "", "position": "Full Professor" }, { "name": "Marco Caccamo", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Yanbing Mao", "gender": "Male", "institution": "Wayne State University", "country": "US", "position": "Assistant Professor" }, { "name": "Yihao Cai", "gender": "Male", "institution": "Wayne State University", "country": "US", "position": "PhD student" } ], "n_formula": 106, "n_formula_1": 17, "n_ref_uni": 32, "n_ref": 56, "n_ref_all": 203, "n_fig": 15, "n_tab": 1, "L_tab": 940, "n_element_tab": 19, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8448, "formula_len_all_1": 1365, "len_all": 215706, "len_all_1": 64760, "len_abs": 1879, "len_title": 72, "len_sents": 66773, "len_sents_1": 30832, "n_sents": 546, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 24, "L_abs": 1424, "title": "Runtime Learning Machine", "abs": "This paper proposes the **Runtime Learning Machine** for safety-critical autonomous systems. The learning machine has three interactive components: a high-performance (HP)-Student, a high-assurance (HA)-Teacher, and a Coordinator. The HP-Student is a high-performance but not fully verified Phy-DRL (physics-regulated deep reinforcement learning) agent that performs safe runtime learning in **real** plants, using **real**-time sensor data from **real**-time physical environments. On the other hand, HA-Teacher is a verified but simplified design, focusing on safety-critical functions. As a complementary, HA-Teacher's novelty lies in real-time patch for two missions: i) correcting unsafe learning of HP-Student, and ii) backing up safety. The Coordinator manages the interaction between HP-Student and HA-Teacher. Powered by the three interactive components, the runtime learning machine notably features i) assuring lifetime safety (i.e., safety guarantee in any runtime learning stage), ii) tolerating unknown unknowns, iii) addressing Sim2Real gap, and iv) automatic hierarchy learning (i.e., safety-first learning, and then high-performance learning). Experiments involving a cart-pole system, two quadruped robots, and a 2D quadrotor, as well as comparisons with state-of-the-art safe DRL, fault-tolerant DRL, and approaches for addressing Sim2Real gap, demonstrate the machine's effectiveness and unique features.", "keywords": [ "Runtime Learning", "Deep Reinforcement Learning", "Safety", "Unknown Unknown", "Autonomous Systems" ], "rating_list": [ 8, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "KBrFTuQGAp", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Doug Stefanski", "gender": "Male", "institution": "Oak Ridge National Laboratory", "country": "US", "position": "Researcher" }, { "name": "John Gounley", "gender": "Not Specified", "institution": "Oak Ridge National Laboratory", "country": "US", "position": "Researcher" }, { "name": "M. Paul Laiu", "gender": "unknown", "institution": "Oak Ridge National Laboratory", "country": "US", "position": "Researcher" }, { "name": "Matthew R Norman", "gender": "unknown", "institution": "Oak Ridge National Laboratory", "country": "US", "position": "Researcher" }, { "name": "Pei Zhang", "gender": "unknown", "institution": "Oak Ridge National Laboratory", "country": "US", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 22, "n_ref": 36, "n_ref_all": 58, "n_fig": 10, "n_tab": 1, "L_tab": 77, "n_element_tab": 6, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 77, "n_element_tab_1": 6, "formula_len_all": 1053, "formula_len_all_1": 1053, "len_all": 88889, "len_all_1": 60695, "len_abs": 1293, "len_title": 128, "len_sents": 30253, "len_sents_1": 29171, "n_sents": 178, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1301, "title": "MATEY: multiscale adaptive foundation models for spatiotemporal physical systems", "abs": "Accurate representation of the multiscale features in spatiotemporal physical systems using vision transformer (ViT) architectures requires extremely long, computationally prohibitive token sequences. To address this issue, we propose an adaptive tokenization scheme which dynamically adjusts the token sizes based on local features. \nMoreover, we present a set of spatiotemporal attention schemes, where the temporal or axial spatial dimensions are decoupled, and evaluate their computational and data efficiencies.\nWe assess the performance of the proposed multiscale adaptive model, MATEY, in a sequence of experiments. \nThe results show that adaptive tokenization achieves improved accuracy without significantly increasing token sequence length, but the improvement deteriorates in more complex data configurations. \nCompared to a full spatiotemporal attention scheme or a scheme that decouples only the temporal dimension, we find that fully decoupled axial attention is less efficient and expressive, requiring more training time and model weights to achieve the same accuracy. \nFinally, we demonstrate in two fine-tuning tasks featuring different physics that models pretrained on PDEBench data outperform the ones trained from scratch, especially in the low data regime with frozen attention.", "keywords": [ "Foundation model; vision transformer", "physical systems; adaptive tokenization; decoupled spatiotemporal attentions; computational fluid dynamics" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "KBixkDNE8p", "primary_area": "interpretability and explainable AI", "n_author": 10, "author_info_list": [ { "name": "Aoxiao Zhong", "gender": "unknown", "institution": "Squirrel AI", "country": "US", "position": "Researcher" }, { "name": "Guibin Zhang", "gender": "Not Specified", "institution": "Tongji University", "country": "CN", "position": "Undergrad student" }, { "name": "Jingheng Ye", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Junfeng Fang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Junyuan Mao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Kun Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Miao Yu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Yang Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Yuxuan Liang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Qingsong Wen", "gender": "Male", "institution": "Squirrel Ai Learning", "country": "US", "position": "Principal Researcher" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 44, "n_ref": 67, "n_ref_all": 97, "n_fig": 5, "n_tab": 6, "L_tab": 16143, "n_element_tab": 177, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1249, "n_element_tab_1": 237, "formula_len_all": 757, "formula_len_all_1": 651, "len_all": 210129, "len_all_1": 63660, "len_abs": 1955, "len_title": 121, "len_sents": 61974, "len_sents_1": 27654, "n_sents": 532, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1958, "title": "MIND SCRAMBLE: UNVEILING LARGE LANGUAGE MODEL PSYCHOLOGY VIA TYPOGLYCEMIA", "abs": "Although still in its infancy, research into the external behaviors and internal mechanisms of large language models (LLMs) has shown significant promise in addressing complex tasks in the physical world. These studies suggest that powerful LLMs, such as GPT-4, are beginning to exhibit human-like cognitive abilities, including planning, reasoning, and reflection, among others. In this paper, we introduce an innovative research line and methodology named LLM Psychology, which leverages or extends human psychology experiments and theories to investigate cognitive behaviors and mechanisms of LLMs. Practically, we migrate the Typoglycemia phenomenon from psychology to explore the \u201cmind\u201d of LLMs. To comprehend scrambled text in Typoglycemia, human brains rely on context and word patterns, which reveals a fundamental difference from LLMs\u2019 encoding and decoding processes. Through various Typoglycemia experiments at the character, word, and sentence levels, we observe the following: (I) LLMs demonstrate human-like behaviors on a macro scale, such as slightly lower task accuracy with consuming more tokens and time; (II) Different LLMs show varying degrees of robustness to scrambled input, making it a democratized benchmark for model evaluation without crafting new datasets; (III) The impact of different task types varies, with complex logical tasks (e.g., math) in scrambled format being more challenging. Going beyond these, some misleadingly optimistic results suggest that LLMs are still primarily data-driven, and their human-like cognitive abilities may differ from what we perceive; (IV) Interestingly, each LLM exhibit its unique and consistent \u201ccognitive pattern\u201d across various tasks, unveiling a general mechanism in its psychology process. To conclude, we provide an in-depth analysis of hidden layers on a micro scale to explain these phenomena, paving the way for LLMs\u2019 deeper interpretability and future research in LLM Psychology.", "keywords": [ "Large Language Model", "Typoglycemia", "Scrambled Text Understanding" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 5, 5, 3, 4, 4 ] }, { "paper_id": "KBSHR4h8XV", "primary_area": "applications to robotics, autonomy, planning", "n_author": 8, "author_info_list": [ { "name": "Fangchen Liu", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Huang Huang", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Jitendra Malik", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Director of Research" }, { "name": "Ken Goldberg", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Letian Fu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Mustafa Mukadam", "gender": "Male", "institution": "Meta AI", "country": "", "position": "Research Scientist" }, { "name": "Pieter Abbeel", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Tingfan Wu", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 33, "n_ref": 80, "n_ref_all": 102, "n_fig": 7, "n_tab": 5, "L_tab": 684, "n_element_tab": 56, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 325, "n_element_tab_1": 28, "formula_len_all": 226, "formula_len_all_1": 223, "len_all": 190278, "len_all_1": 72200, "len_abs": 1258, "len_title": 149, "len_sents": 42714, "len_sents_1": 34638, "n_sents": 315, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1263, "title": "Early Fusion Helps Vision Language Action Models Generalize Better", "abs": "Recent advances in Vision-Language-Action (VLA) models can enable robots to perform a wide range of tasks based on language or goal-based instructions. These VLA models typically encode text and images into disjoint tokens, generating actions that align with the given instructions. This requires the VLA models to simultaneously perform vision-language understanding and precise closed-loop control, resulting in significant challenges for them to generalize to new environments. However, contrastive pre-trained VLMs, such as CLIP, already possess vision-language alignment capabilities, which are underutilized by current VLA models. In this paper, we propose Early Fusion VLA (EF-VLA), a novel VLA architecture that exploits CLIP\u2019s vision-language understanding by performing early fusion, extracting fine-grained vision-language tokens relevant to the task instructions before passing them to the transformer policy. EF-VLA keeps the\nVLM frozen, allowing it to effectively perform unseen tasks without requiring fine-tuning, which often reduces generalization capabilities. Simulation and real-world experiments suggest that EF-VLA outperforms state-of-the-art VLA models on diverse tasks, with significant generalization capabilities in unseen environments.", "keywords": [ "vision language action model; robot foundation model" ], "rating_list": [ 1, 3, 6 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "KAWlH5pfQu", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Furkan Mumcu", "gender": "Male", "institution": "University of South Florida", "country": "US", "position": "PhD student" }, { "name": "Yasin Yilmaz", "gender": "unknown", "institution": "University of South Florida", "country": "", "position": "Associate Professor" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 38, "n_ref": 80, "n_ref_all": 97, "n_fig": 6, "n_tab": 8, "L_tab": 873, "n_element_tab": 56, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1805, "n_element_tab_1": 36, "formula_len_all": 640, "formula_len_all_1": 532, "len_all": 129422, "len_all_1": 54556, "len_abs": 877, "len_title": 78, "len_sents": 37144, "len_sents_1": 24535, "n_sents": 256, "n_sents_1": 169, "venue": "ICLR2025", "decision": "withdraw", "L_title": 30, "L_abs": 880, "title": "Detecting Adversarial Examples", "abs": "Deep Neural Networks (DNNs) have been shown to be vulnerable to adversarial examples. While numerous successful adversarial attacks have been proposed, defenses against these attacks remain relatively understudied. Existing defense approaches either focus on negating the effects of perturbations caused by the attacks to restore the DNNs' original predictions or use a secondary model to detect adversarial examples. However, these methods often become ineffective due to the continuous advancements in attack techniques. We propose a novel universal and lightweight method to detect adversarial examples by analyzing the layer outputs of DNNs. Through theoretical justification and extensive experiments, we demonstrate that our detection method is highly effective, compatible with any DNN architecture, and applicable across different domains, such as image, video, and audio.", "keywords": [ "adversarial machine learning", "security", "robustness" ], "rating_list": [ 1, 3, 5, 3, 3 ], "soundness_list": [ 1, 1, 2, 1, 2 ], "presentation_list": [ 2, 3, 2, 3, 2 ], "contribution_list": [ 1, 1, 2, 3, 1 ], "confidence_list": [ 5, 5, 5, 4, 4 ] }, { "paper_id": "KAIqwkB3dT", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Bosun Hwang", "gender": "Male", "institution": "Samsung Advanced Institute of Technology", "country": "", "position": "Principal Researcher" }, { "name": "Hongjoon Ahn", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Jinu Hyeon", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "MS student" }, { "name": "Taesup Moon", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Youngmin Oh", "gender": "Male", "institution": "Samsung", "country": "KR", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 46, "n_ref": 112, "n_ref_all": 147, "n_fig": 15, "n_tab": 5, "L_tab": 4368, "n_element_tab": 271, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 213, "formula_len_all_1": 0, "len_all": 167265, "len_all_1": 65534, "len_abs": 1247, "len_title": 148, "len_sents": 51744, "len_sents_1": 33655, "n_sents": 351, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 99, "L_abs": 1250, "title": "Prevalence of Negative Transfer in Continual Reinforcement Learning: Analyses and a Simple Baseline", "abs": "We argue that the negative transfer problem occurring when the new task to learn arrives is an important problem that needs not be overlooked when developing effective Continual Reinforcement Learning (CRL) algorithms. Through comprehensive experimental validation, we demonstrate that such issue frequently exists in CRL and cannot be effectively addressed by several recent work on either mitigating plasticity loss of RL agents or enhancing the positive transfer in CRL scenario. To that end, we develop Reset & Distill (R&D), a simple yet highly effective baseline method, to overcome the negative transfer problem in CRL. R&D combines a strategy of resetting the agent's online actor and critic networks to learn a new task and an offline learning step for distilling the knowledge from the online actor and previous expert's action probabilities. We carried out extensive experiments on long sequence of Meta World tasks and show that our simple baseline method consistently outperforms recent approaches, achieving significantly higher success rates across a range of tasks. Our findings highlight the importance of considering negative transfer in CRL and emphasize the need for robust strategies like R&D to mitigate its detrimental effects.", "keywords": [ "continual reinforcement learning", "negative transfer", "plasticity loss" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 4, 4, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "KA2Rit4ky1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Liangjian Wen", "gender": "Male", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "Assistant Professor" }, { "name": "Qifan Wang", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Principal Researcher" }, { "name": "Shiyi Qi", "gender": "Male", "institution": "Harbin Institute of Technology, shenzhen", "country": "", "position": "MS student" }, { "name": "Yiduo Li", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yuan Qi", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Qingsong Wen", "gender": "Male", "institution": "Squirrel Ai Learning", "country": "US", "position": "Principal Researcher" }, { "name": "Zenglin Xu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 37, "n_formula_1": 29, "n_ref_uni": 37, "n_ref": 84, "n_ref_all": 104, "n_fig": 6, "n_tab": 6, "L_tab": 3460, "n_element_tab": 131, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 3210, "n_element_tab_1": 105, "formula_len_all": 3578, "formula_len_all_1": 1839, "len_all": 138765, "len_all_1": 64157, "len_abs": 1447, "len_title": 169, "len_sents": 39200, "len_sents_1": 25857, "n_sents": 308, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 121, "L_abs": 1454, "title": "PDETime: Rethinking Long-term Multivariate Time Series Forecasting from the Perspective of Partial Differential Equations", "abs": "Recent advancements in deep learning have led to the development of various approaches for long-term multivariate time-series forecasting (LMTF). Most of these approaches can be categorized as either historical-value-based methods, which rely on discretely sampled past observations, or time-index-based methods that model time indices directly as input variables. However, real-world dynamical systems often exhibit nonstationarity and suffer from insufficient sampling frequency, posing challenges such as spurious correlations between time steps and difficulties in modeling complex temporal dependencies.\nIn this paper, we treat multivariate time series as data sampled from a continuous dynamical system governed by partial differential equations (PDEs) and propose a new model called PDETime. \nInstead of predicting future values directly, PDETime employs an encoding-integration-decoding architecture: it predicts the partial derivative of the system with respect to time (i.e., the first-order difference) in the latent space and then integrates this information to forecast future series. This approach enhances both performance and stability, especially in scenarios with extremely long forecasting windows. Extensive experiments on seven diverse real-world LMTF datasets demonstrate that PDETime not only adapts effectively to the intrinsic spatiotemporal nature of the data but also sets new benchmarks by achieving state-of-the-art results.", "keywords": [ "long-term multivariate time series forecasting" ], "rating_list": [ 6, 6, 3, 6, 3 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 1, 3, 2 ], "contribution_list": [ 3, 2, 2, 4, 2 ], "confidence_list": [ 5, 3, 3, 4, 3 ] }, { "paper_id": "K9zedJlybd", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Jisung Yoon", "gender": "Male", "institution": "KDI School of Public Policy and Management", "country": "KR", "position": "Assistant Professor" }, { "name": "Munjung Kim", "gender": "Female", "institution": "Indiana University", "country": "", "position": "PhD student" }, { "name": "Ozgur Can Seckin", "gender": "unknown", "institution": "Indiana University", "country": "", "position": "PhD student" }, { "name": "Rachith Aiyappa", "gender": "Male", "institution": "Indiana University Bloomington", "country": "US", "position": "PhD student" }, { "name": "Sadamori Kojaku", "gender": "Male", "institution": "State University of New York at Binghamton", "country": "US", "position": "Assistant Professor" }, { "name": "Xin Wang", "gender": "Female", "institution": "State University of New York at Binghamton", "country": "", "position": "PhD student" }, { "name": "Yong-Yeol Ahn", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Full Professor" } ], "n_formula": 42, "n_formula_1": 7, "n_ref_uni": 84, "n_ref": 247, "n_ref_all": 298, "n_fig": 27, "n_tab": 5, "L_tab": 5889, "n_element_tab": 722, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10631, "formula_len_all_1": 2197, "len_all": 301043, "len_all_1": 66574, "len_abs": 1117, "len_title": 48, "len_sents": 69816, "len_sents_1": 31352, "n_sents": 500, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1069, "title": "Implicit degree bias in the link prediction task", "abs": "Link prediction---a task of distinguishing actual hidden edges from random unconnected node pairs---is one of the quintessential tasks in graph machine learning. Despite being widely accepted as a universal benchmark and a downstream task for representation learning, the link prediction benchmark's validity has rarely been questioned. Here, we show that the common edge sampling procedure in the link prediction task has an implicit bias toward high-degree nodes. This produces a highly skewed evaluation that favors methods overly dependent on node degree. In fact a ``null'' link prediction method based solely on node degree can yield nearly optimal performance in this setting. We propose a degree-corrected link prediction benchmark that offers a more reasonable assessment and better aligns with the performance on the recommendation task. Finally, we demonstrate that the degree-corrected benchmark can more effectively train graph machine-learning models by reducing overfitting to node degrees and facilitating the learning of relevant structures in graphs.", "keywords": [ "link prediction", "graph machine learning", "sampling bias", "benchmark" ], "rating_list": [ 6, 8, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 5, 4, 4 ] }, { "paper_id": "K9xuqsaP0R", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Fangchen Yu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Visiting student" }, { "name": "Wenye Li", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "", "position": "Associate Professor" }, { "name": "Yidong LIN", "gender": "Male", "institution": "The Chinese University of Hong Kong,Shenzhen", "country": "CN", "position": "Undergrad student" }, { "name": "Yuqi Ma", "gender": "Female", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Zhenghao Huang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Undergrad student" }, { "name": "Hu Ruilizhen", "gender": "Male", "institution": " Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Undergrad student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 28, "n_ref": 48, "n_ref_all": 57, "n_fig": 3, "n_tab": 7, "L_tab": 3462, "n_element_tab": 195, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 4213, "n_element_tab_1": 215, "formula_len_all": 380, "formula_len_all_1": 380, "len_all": 99140, "len_all_1": 70316, "len_abs": 1258, "len_title": 111, "len_sents": 27408, "len_sents_1": 26031, "n_sents": 338, "n_sents_1": 328, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1268, "title": "KAE: Kolmogorov-Arnold Auto-Encoder for Representation Learning", "abs": "The Kolmogorov-Arnold Network (KAN) has recently emerged as a promising alternative to traditional multi-layer perceptrons (MLPs), offering enhanced accuracy and interpretability through learnable activation functions on edges instead of fixed functions on nodes. In this paper, we present the Kolmogorov-Arnold Auto-Encoder (KAE), a novel integration of KAN with autoencoders (AEs) that aims to improve representation learning and performance in retrieval, classification, and denoising tasks. By utilizing the flexible polynomial functions in KAN layers, KAE effectively captures complex data patterns and non-linear relationships, outperforming standard autoencoders. Our extensive experiments on benchmark datasets show that KAE significantly enhances the quality of latent representations, resulting in reduced reconstruction and denoising errors, and also improves performance in downstream tasks, including higher classification accuracy, retrieval recall, and interpretability compared to standard autoencoders and other KAN variants. These findings position KAE as a practical tool for high-dimensional data analysis, paving the way for more robust performance in representation learning. The code is available at \\url{https://anonymous.4open.science/r/KAE/}.", "keywords": [ "Kolmogorov-Arnold Network", "Auto-Encoder", "Representation Learning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "K9Elg2JrvY", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Manh Cuong Dao", "gender": "Male", "institution": "Hanoi University of Science and Technology", "country": "VN", "position": "MS student" }, { "name": "Phi Le Nguyen", "gender": "Female", "institution": "Hanoi University of Science and Technology", "country": "VN", "position": "Associate Professor" }, { "name": "Thao Nguyen Truong", "gender": "Male", "institution": "AIST, National Institute of Advanced Industrial Science and Technology", "country": "JP", "position": "Researcher" }, { "name": "The Hung Tran", "gender": "Male", "institution": "Hanoi University of Science and Technology", "country": "VN", "position": "Undergrad student" }, { "name": "Nghia Hoang", "gender": "Male", "institution": "Washington State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 31, "n_formula_1": 26, "n_ref_uni": 32, "n_ref": 88, "n_ref_all": 110, "n_fig": 2, "n_tab": 8, "L_tab": 3972, "n_element_tab": 259, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 1882, "n_element_tab_1": 103, "formula_len_all": 2557, "formula_len_all_1": 2234, "len_all": 154942, "len_all_1": 72602, "len_abs": 1215, "len_title": 126, "len_sents": 46042, "len_sents_1": 30931, "n_sents": 362, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1222, "title": "FROM LOW TO HIGH-VALUE DESIGNS: OFFLINE OPTIMIZATION VIA GENERALIZED DIFFUSION", "abs": "This paper studies the black-box optimization task which aims to find the maxima of a black-box function using only a static set of its observed input-output data. This is often achieved via learning and optimizing a surrogate function using such offline dataset. Alternatively, it can also be framed as an inverse modeling task which maps a desired performance to potential input candidates that achieve it. Both approaches are limited by the limited amount of offline data. To mitigate this limitation, we introduce a new perspective which casts offline optimization as a diffusion process mapping between an implicit distribution of low-value inputs (i.e., offline data) and a superior distribution of high-value inputs (i.e., solution candidates). Such diffusion process can be learned using low- and high-value inputs sampled from synthetic functions that resemble the target function. These synthetic functions are constructed as the mean posterior of multiple Gaussian processes fitted with different parameterizations on the offline data, alleviating the data bottleneck. Experimental results demonstrate that our approach consistently outperforms previous methods, establishing a new state-of-the-art performance.", "keywords": [ "Diffusion Model", "Probabilistic Method" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 2, 4 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "K8WaxpiSDw", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Huihui Yue", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Miao Sun", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Pengfei Qi", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Yin Xiangjun", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Zhengyuan Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" } ], "n_formula": 54, "n_formula_1": 40, "n_ref_uni": 25, "n_ref": 43, "n_ref_all": 82, "n_fig": 14, "n_tab": 3, "L_tab": 1151, "n_element_tab": 102, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 3362, "n_element_tab_1": 104, "formula_len_all": 2771, "formula_len_all_1": 1898, "len_all": 117720, "len_all_1": 51612, "len_abs": 895, "len_title": 114, "len_sents": 29578, "len_sents_1": 17604, "n_sents": 255, "n_sents_1": 137, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 934, "title": "Enhancing lensless imaging via Explicit Learning of Model Mismatch", "abs": "Emerging lensless imaging techniques hold promise for miniaturized cameras, but their effectiveness is constrained by challenges like model mismatch from the point spread function (PSF), which undermines reconstruction methods dependent on accurate PSF modeling. To address this issue, we propose a joint Maximum a Posteriori (MAP) approach to simultaneously estimate model mismatch error (${\\rm M^{2}}$E) and reconstruct high-resolution images from lensless imaging measurements. Specifically, we propose an explicit latent space representation for ${\\rm M^{2}}$E to improve robustness against PSF inaccuracies. Additionally, we develop a multi-stage reconstruction network by unfolding the joint MAP estimator with a learned Laplacian Scale Mixture (LSM) prior and ${\\rm M^{2}}$E representation (${\\rm M^{2}}$ER) through end-to-end optimization. Extensive experiments show that our method surpasses current state-of-the-art methods.", "keywords": [ "Lensless Imaging; Maximum a Posteriori;Model Mismatch Error;Image Reconstruction" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "K7xpl3LZQp", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chaohu Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Jianting Tang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Linli Xu", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yubo Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 31, "n_ref": 71, "n_ref_all": 98, "n_fig": 9, "n_tab": 11, "L_tab": 2661, "n_element_tab": 333, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 793, "n_element_tab_1": 163, "formula_len_all": 924, "formula_len_all_1": 498, "len_all": 143699, "len_all_1": 57775, "len_abs": 1495, "len_title": 150, "len_sents": 37934, "len_sents_1": 26424, "n_sents": 315, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 101, "L_abs": 1501, "title": "Tracking the Copyright of Large Vision-Language Models through Parameter Learning Adversarial Attacks", "abs": "Large vision-language models (LVLMs) have demonstrated remarkable image understanding and dialogue capabilities, allowing them to handle a variety of visual question answering tasks. However, their widespread availability raises concerns about unauthorized usage and copyright infringement, where users or individuals can develop their own LVLMs by fine-tuning published models. In this paper, we propose a novel method called Parameter Learning Attack (PLA) for tracking the copyright of LVLMs without modifying the original model. Specifically, we construct adversarial images through targeted attacks against the original model, enabling it to generate specific outputs. To ensure these attacks remain effective on potential fine-tuned models to trigger copyright tracking, we allow the original model to learn the trigger images by updating parameters in the opposite direction during the adversarial attack process. Notably, the proposed method can be applied after the release of the original model, thus not affecting the model\u2019s performance and behavior. To simulate real-world applications, we fine-tune the original model using various strategies across diverse datasets, creating a range of models for copyright verification. Extensive experiments demonstrate that our method can more effectively identify the original copyright of fine-tuned models compared to baseline methods. Therefore, this work provides a powerful tool for tracking copyrights and detecting unlicensed usage of LVLMs.", "keywords": [ "Copyright Tracking", "Large Vision-Language Models", "Adversarial Attacks", "Fine-tuning" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 4 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "K7qurARDi1", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Haonan Wu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "He Tang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Hua Li", "gender": "unknown", "institution": "Hainan University", "country": "CN", "position": "Associate Professor" }, { "name": "Shi Wang", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Shuang Hao", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 49, "n_ref": 78, "n_ref_all": 93, "n_fig": 7, "n_tab": 4, "L_tab": 1121, "n_element_tab": 124, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 678, "n_element_tab_1": 82, "formula_len_all": 643, "formula_len_all_1": 406, "len_all": 132408, "len_all_1": 56192, "len_abs": 0, "len_title": 0, "len_sents": 35080, "len_sents_1": 26522, "n_sents": 267, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1349, "title": "SeaDiff: Delve into Underwater Image Generation with Symmetrical Parameter Control", "abs": "With the advancement of diffusion models, the controllability of image generation has significantly improved. However, due to the refraction and absorption of light in water, underwater images often exhibit notable variations in luminance and color cast. This leads to challenges for generative models pre-trained on terrestrial images, as they struggle to produce underwater images with a diverse range of these variations, severely limiting the appearance diversity of generated underwater images. To address this issue, we focus on the precise control of appearance in underwater images. We model the appearance of underwater images using three attributes: luminance, dynamic range, and color cast. We propose a new method, SeaDiff, which introduces a Symmetrical Parameter Control structure to achieve precise control over the appearance of underwater images. The proposed method comprises two modules: Appearance Writer, which encodes and injects appearance attributes into the U-Net encoder, and Appearance Reader, which ensures that the generated images align with the desired appearance by analyzing the feature maps. Experimental results demonstrate that the proposed SeaDiff method significantly improves control over underwater image appearance while maintaining image quality, validating its effectiveness in underwater image generation.", "keywords": [ "diffusion model", "controllable generation", "underwater" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "K7XiXLfFSP", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Hang Yin", "gender": "Male", "institution": "Department of Mathematical science, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yangqiu Song", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Zihao Wang", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Fei Weizhi", "gender": "Male", "institution": "The Department of Mathematics, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 3, "n_ref_uni": 42, "n_ref": 146, "n_ref_all": 206, "n_fig": 9, "n_tab": 9, "L_tab": 4452, "n_element_tab": 1036, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 3298, "n_element_tab_1": 269, "formula_len_all": 500, "formula_len_all_1": 212, "len_all": 200718, "len_all_1": 66282, "len_abs": 1130, "len_title": 129, "len_sents": 69223, "len_sents_1": 28960, "n_sents": 447, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1146, "title": "$EFO_{k}$-CQA: Towards Knowledge Graph Complex Query Answering beyond Set Operation", "abs": "To answer complex queries on knowledge graphs, logical reasoning over incomplete knowledge needs learning-based methods because they are capable of generalizing over unobserved knowledge. Therefore, an appropriate dataset is fundamental to both obtaining and evaluating such methods under this paradigm. In this paper, we propose a comprehensive framework for data generation, model training, and method evaluation that covers the combinatorial space of Existential First-order Queries with multiple variables ($EFO_{k}$). The combinatorial query space in our framework significantly extends those defined by set operations in the existing literature. Additionally, we construct a dataset, $EFO_{k}$-CQA, with 741 query types for empirical evaluation, and our benchmark results provide new insights into how query hardness affects the results. Furthermore, we demonstrate that the existing dataset construction process is systematically biased and hinders the appropriate development of query-answering methods, highlighting the importance of our work. Our code and data are provided in~\\url{https://anonymous.4open.science/r/EFOK-CQA/README.md}.", "keywords": [ "complex query answering", "knowledge graph" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "K5yeB4dTtS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Junpeng Yue", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xinrun Xu", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zongqing Lu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "B\u00f6rje Karlsson", "gender": "unknown", "institution": "Beijing Academy of Artificial Intelligence (BAAI)", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 52, "n_ref": 94, "n_ref_all": 122, "n_fig": 11, "n_tab": 9, "L_tab": 3404, "n_element_tab": 247, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 367, "n_element_tab_1": 53, "formula_len_all": 158, "formula_len_all_1": 125, "len_all": 229091, "len_all_1": 63740, "len_abs": 1469, "len_title": 130, "len_sents": 60540, "len_sents_1": 29441, "n_sents": 476, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1299, "title": "MLLM as Retriever: Interactively Learning Multimodal Retrieval for Embodied Agents", "abs": "MLLM agents demonstrate potential for complex embodied tasks by retrieving multimodal task-relevant trajectory data. However, current retrieval methods primarily focus on surface-level similarities of textual or visual cues in trajectories, neglecting their effectiveness for the specific task at hand. To address this issue, we propose a novel method, MART, which enhances the performance of embodied agents by utilizing interaction data to fine-tune an MLLM retriever based on preference learning, such that the retriever fully considers the effectiveness of trajectories and prioritize them for unseen tasks. We also introduce Trajectory Abstraction, a mechanism that leverages MLLMs' summarization capabilities to represent trajectories with fewer tokens while preserving key information, enabling agents to better comprehend milestones in the trajectory. Experimental results across various environments demonstrate our method significantly improves task success rates in unseen scenes compared to baseline methods. This work presents a new paradigm for multimodal retrieval in embodied agents, by fine-tuning a general-purpose MLLM as the retriever to assess trajectory effectiveness. All benchmark task sets and simulator code modifications for action and observation spaces will be released.", "keywords": [ "multimodal retrieval", "interactive learning", "MLLM embodied agent" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "K5wFwpaUvK", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Songhua Liu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yuhe Liu", "gender": "Female", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "R Yu", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Xinchao Wang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 22, "n_ref": 89, "n_ref_all": 102, "n_fig": 8, "n_tab": 2, "L_tab": 14, "n_element_tab": 3, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 936, "n_element_tab_1": 53, "formula_len_all": 535, "formula_len_all_1": 535, "len_all": 103016, "len_all_1": 55602, "len_abs": 1656, "len_title": 111, "len_sents": 28005, "len_sents_1": 23860, "n_sents": 260, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1664, "title": "Injecting Vision Language into Autoregressive Image Generation", "abs": "Autoregressive (AR) models have become central to modern foundation models like large language models (LLMs) and visual-language models (VLMs). Recently, AR-based approaches have extended into text-to-image generation. Although these text-to-image AR models have been trained for visual-language token interaction, they often struggle when conditioned on visual inputs. Focusing on this drawback, in this paper, we are curious about one question: how can we inject vision information to a pre-trained AR model to ensure its output reflects visual conditions? We answer this question with a simple yet effective solution termed InjectAR. Our key insight is that, while a pre-trained AR model cannot handle visual inputs directly, its inherent capability for visual-language interaction can indeed support visual feature extraction. Consequently, with only a few newly introduced parameters and minimal training, a pre-trained AR generation model can successfully accommodate both text and image conditions and produce visually appealing results. To manage the relationship between textual and visual inputs, we reinforce InjectAR with a hierarchical attention mechanism, which subdivides the attention scores for textual tokens into their corresponding visual components, preventing either modality from dominating the output. As the first AR model with this capability, extensive experiments show that InjectAR achieves performance on par with, or even surpasses, state-of-the-art diffusion models. Moreover, unlike diffusion models, once trained, our method has the potential for flexible control over the positions of visual objects. Our codes will be available.", "keywords": [ "autoregressive models", "image generation", "text-to-image", "customized image generation" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 1, 3, 4 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "K5QGZut3uu", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Hiroki Ohashi", "gender": "Male", "institution": "Hitachi Ltd. R&D group", "country": "", "position": "Researcher" }, { "name": "Yusuke Ohtsubo", "gender": "Male", "institution": "Hitachi, Ltd.", "country": "", "position": "Researcher" } ], "n_formula": 35, "n_formula_1": 33, "n_ref_uni": 13, "n_ref": 37, "n_ref_all": 56, "n_fig": 6, "n_tab": 2, "L_tab": 853, "n_element_tab": 15, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1830, "formula_len_all_1": 2255, "len_all": 107671, "len_all_1": 62236, "len_abs": 961, "len_title": 124, "len_sents": 41735, "len_sents_1": 26717, "n_sents": 345, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 895, "title": "Fast Computation of Gaussian Processes Augmented by Synthetic Simulator Data", "abs": "When the amount of training data is limited, augmenting it with generated data from a simulator can be a beneficial approach to improving prediction accuracy. However, there are no clear metrics on which generated data should be added to the training set and in what proportion, especially when the predictive model is a Gaussian Processes (GPs) model. To address this, we propose using the log marginal likelihood as a guiding metric. The log marginal likelihood is a theoretically grounded criterion for model selection. However, computing this metric for GPs is computationally expensive. To overcome this challenge, we introduce a faster method for calculating the log marginal likelihood by considering the Cholesky factor and matrix element dependencies. Experimental results demonstrate that metrics utilizing the log likelihood outperform basic methods in mean squared error on test set.", "keywords": [ "Gaussian processes", "simulator", "marginal likelihood", "model selection" ], "rating_list": [ 3, 3, 6, 3 ], "soundness_list": [ 1, 1, 3, 2 ], "presentation_list": [ 1, 2, 3, 1 ], "contribution_list": [ 1, 1, 3, 1 ], "confidence_list": [ 4, 5, 2, 4 ] }, { "paper_id": "K4YMFdx2Z2", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Atsuyuki Miyai", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Go Irie", "gender": "Male", "institution": "Tokyo University of Science", "country": "JP", "position": "Associate Professor" }, { "name": "Hai Li", "gender": "Female", "institution": "Duke University", "country": "US", "position": "Professor" }, { "name": "Jingkang Yang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Jingyang Zhang", "gender": "Male", "institution": "Electrical and Computer Engineering, Duke University", "country": "US", "position": "PhD student" }, { "name": "Kiyoharu Aizawa", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Qing Yu", "gender": "Male", "institution": "LY Corporation", "country": "JP", "position": "Researcher" }, { "name": "Yifei Ming", "gender": "Male", "institution": "Salesforce AI Research", "country": "US", "position": "Researcher" }, { "name": "Yixuan Li", "gender": "Female", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Ziwei Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 60, "n_ref": 191, "n_ref_all": 235, "n_fig": 15, "n_tab": 19, "L_tab": 24034, "n_element_tab": 3522, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1664, "n_element_tab_1": 134, "formula_len_all": 529, "formula_len_all_1": 0, "len_all": 337625, "len_all_1": 67266, "len_abs": 1620, "len_title": 0, "len_sents": 66675, "len_sents_1": 30680, "n_sents": 600, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1204, "title": "Unsolvable Problem Detection: Evaluating Trustworthiness of Large Multimodal Models", "abs": "This paper introduces a novel and well-defined challenge for Large Multimodal Models (LMMs), termed Unsolvable Problem Detection (UPD). UPD examines the LMM's ability to withhold answers when faced with unsolvable problems. UPD encompasses three problems: Absent Answer Detection (AAD), Incompatible Answer Set Detection (IASD), and Incompatible Visual Question Detection (IVQD), covering unsolvable cases like answer-lacking or incompatible choices and image-question mismatches. In this paper, we introduce the MM-UPD Bench, a benchmark for assessing performance across various ability dimensions. Our experiments reveal that even most LMMs, which demonstrate adequate performance on existing benchmarks, struggle significantly with MM-UPD, underscoring a novel aspect of trustworthiness that current benchmarks have overlooked. To deepen the understanding of the UPD, we explore various solutions, including chain of thought, self-reflection, and instruction tuning, and demonstrate each approach's efficacy and limitations. We hope our insights, together with future efforts within the proposed UPD settings, will enhance the broader understanding and development of more practical and reliable LMMs.", "keywords": [ "Unsolvable Problem Detection: Evaluating Trustworthiness of Large Multimodal Models" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 2, 4, 4 ] }, { "paper_id": "K4JHTZ13G3", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Eugenia Soboleva", "gender": "Female", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "MS student" }, { "name": "Ivan Oseledets", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "Mariia Donskova", "gender": "Female", "institution": "Institute for Information Transmission Problems", "country": "RU", "position": "PhD student" }, { "name": "Marina Munkhoeva", "gender": "Female", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Research Scientist" }, { "name": "Maxim Panov", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Mikhail Goncharov", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 2, "n_ref_uni": 32, "n_ref": 70, "n_ref_all": 82, "n_fig": 2, "n_tab": 5, "L_tab": 1974, "n_element_tab": 311, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 4112, "n_element_tab_1": 402, "formula_len_all": 630, "formula_len_all_1": 79, "len_all": 118777, "len_all_1": 62939, "len_abs": 1508, "len_title": 193, "len_sents": 34336, "len_sents_1": 26703, "n_sents": 239, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 145, "L_abs": 1518, "title": "Screener: Learning Conditional Distribution of Dense Self-supervised Representations for Unsupervised Pathology Segmentation in 3D Medical Images", "abs": "Accurate and automated anomaly segmentation is critical for assisting clinicians in detecting and diagnosing pathological conditions, particularly in large-scale medical imaging datasets where manual annotation is not only time- and resource-intensive but also prone to inconsistency. To address these challenges, we propose Screener, a fully self-supervised framework for visual anomaly segmentation, leveraging self-supervised representation learning to eliminate the need for manual labels. Additionally, we model the conditional distribution of local image patterns given their global context, enabling the identification of anomalies as patterns with low conditional probabilities and assigning them high anomaly scores.\n\nScreener comprises three components: a descriptor model that encodes local image patterns into self-supervised representations invariant to local-content-preserving augmentations; a condition model that captures global contextual information through invariance to image masking; and a density model that estimates the conditional density of descriptors given their global contexts to compute anomaly scores.\n\nWe validate Screener by training a fully self-supervised model on over 30,000 3D CT images and evaluating its performance on four large-scale test datasets comprising 1,820 3D CT scans across four chest and abdominal pathologies. Our framework consistently outperforms existing unsupervised anomaly segmentation methods. Code and pre-trained models will be made publicly available.", "keywords": [ "Unsupervised visual anomaly detection", "self-supervised learning", "density estimation", "medical image analysis" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 4, 3, 2 ], "presentation_list": [ 3, 2, 1 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 5, 3, 3 ] }, { "paper_id": "K4FAFNRpko", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Donglin Wang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Associate Professor" }, { "name": "Han Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Pengxiang Ding", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Shuanghao Bai", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Wei Zhao", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Postdoc" }, { "name": "Zhang Min", "gender": "Female", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Zhefei Gong", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 32, "n_ref": 46, "n_ref_all": 61, "n_fig": 8, "n_tab": 9, "L_tab": 2623, "n_element_tab": 291, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1141, "n_element_tab_1": 108, "formula_len_all": 156, "formula_len_all_1": 157, "len_all": 155209, "len_all_1": 58862, "len_abs": 1541, "len_title": 141, "len_sents": 37527, "len_sents_1": 27268, "n_sents": 272, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1707, "title": "VLAS: Vision-Language-Action Model with Speech Instructions for Customized Robot Manipulation", "abs": "Vision-language-action models (VLAs) have recently become highly prevalent in robot manipulation due to its end-to-end architecture and impressive performance. However, current VLAs are limited to processing human instructions in textual form, neglecting the more natural speech modality for human interaction. A typical approach of incorporating speech modality into VLA necessitates a separate speech recognition system to transcribe spoken instructions into text. Such a cascading pipeline raises two major concerns for robotic systems. First, the entire model grows in size and complexity, potentially resulting in redundant computations and increased memory consumption. Second, the transcription procedure would lose non-semantic information in the raw speech, such as voiceprint, which is crucial for a robot to successfully understand and complete customized tasks. To this end, we propose VLAS, the fisrt end-to-end policy model that seamlessly integrates speech modality for robot manipulation. We present a three-stage speech instruction tuning strategy leveraging multimodal datasets, including our manually curated SQA and CSI datasets. Furthermore, to facilitate personalized operations, we develop a voice retrieval-augmented generation (RAG) approach to enhance the robot's performance in tasks requiring individual-specific knowledge. Experimental results show that the proposed VLAS, following either textual or speech instructions, can achieve performance comparable to traditional VLAs on the CALVIN benchmark. In addition, we created a benchmark consisting of customization tasks, where our VLAS demonstrates absolute superiority by fully leveraging the auxiliary information in speech.", "keywords": [ "Vision-Language-Action Model", "Speech Instructions", "Robot Manipulation" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "K3n5jPkrU6", "primary_area": "foundation or frontier models, including LLMs", "n_author": 12, "author_info_list": [ { "name": "Chen Qian", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Cheng Yang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Associate Professor" }, { "name": "Hanchen Xia", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "MS student" }, { "name": "Kunlun Zhu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Maosong Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Wei Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Weize Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "YiFei Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Yufan Dang", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhiyuan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhuoyun Du", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zihao Xie", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" } ], "n_formula": 22, "n_formula_1": 7, "n_ref_uni": 67, "n_ref": 141, "n_ref_all": 179, "n_fig": 23, "n_tab": 3, "L_tab": 457, "n_element_tab": 5, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 558, "n_element_tab_1": 62, "formula_len_all": 2125, "formula_len_all_1": 648, "len_all": 266749, "len_all_1": 63469, "len_abs": 1122, "len_title": 108, "len_sents": 68525, "len_sents_1": 31803, "n_sents": 428, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1120, "title": "Scaling Large Language Model-based Multi-Agent Collaboration", "abs": "Recent breakthroughs in large language model-driven autonomous agents have revealed that multi-agent collaboration often surpasses each individual through collective reasoning. Inspired by the neural scaling law\u2014increasing neurons enhances performance, this study explores whether the continuous addition of collaborative agents can yield similar benefits. Technically, we utilize directed acyclic graphs to organize agents into a multi-agent collaboration network (MacNet), upon which their interactive reasoning is topologically orchestrated for autonomous task solving. Extensive evaluations reveal that it effectively supports collaboration among over a thousand agents, with irregular topologies outperforming regular ones. We also identify a collaborative scaling law\u2014the overall performance follows a logistic growth pattern as agents scale, with collaborative emergence occurring earlier than traditional neural emergence. We speculate this may be because scaling agents catalyzes their multidimensional considerations during interactive reflection and refinement, thereby producing more comprehensive solutions.", "keywords": [ "Large Language Model", "Autonomous Agent", "Multi-Agent Collaboration", "Interactive Reasoning" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 4 ], "presentation_list": [ 3, 4, 4, 4 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "K3jv45pptT", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Adam Kortylewski", "gender": "unknown", "institution": "Saarland Informatics Campus, Max-Planck Institute", "country": "DE", "position": "Research Group Leader" }, { "name": "Alan Yuille", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Angtian Wang", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Jiahao Wang", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Jiawei Peng", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Intern" }, { "name": "Ju He", "gender": "Male", "institution": "TikTok", "country": "US", "position": "Researcher" }, { "name": "Prakhar Kaushik", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Qihao Liu", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Wufei Ma", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Xiaoding Yuan", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Yaoyao Liu", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Yi Zhang", "gender": "Male", "institution": "Sony Corporation of America", "country": "US", "position": "Research Engineer" }, { "name": "Yining Sun", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "MS student" }, { "name": "J Chen", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 40, "n_ref": 79, "n_ref_all": 91, "n_fig": 1, "n_tab": 7, "L_tab": 2811, "n_element_tab": 366, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 3211, "n_element_tab_1": 323, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 144744, "len_all_1": 67901, "len_abs": 2490, "len_title": 138, "len_sents": 36918, "len_sents_1": 31403, "n_sents": 259, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1485, "title": "DSPart: A Large-scale Diffusion-generated Synthetic Dataset with Annotations from 3D Parts", "abs": "Object parts provide representations that enable a detailed and interpretable understanding of object structures, making part recognition crucial for various real-world applications. However, acquiring pixel-level part annotations is both expensive and time-consuming. Rendering 3D object models with their 3D part annotations is a promising solution since it allows the generation of unlimited synthetic data samples with precise 3D control and accurate part segmentation masks. Nevertheless, these synthetic datasets suffer from a lack of realism, resulting in large domain gaps. In this paper, we present a large-scale realistic synthetic dataset with part annotations, namely Diffusion-generated Synthetic Parts (DSPart), for both rigid objects and animals. For images in DSPart, we obtain 2D part masks from 3D part annotations by leveraging recent advances in diffusion models with 3D control. In addition to offering more diverse and realistic textures, prior knowledge of diffusion models enables the object to exhibit more physically realistic interactions with the ground plane and other spatial contexts. We annotate $475$ representative shape instances from $50$ object categories for DSPart-Rigid and use $3,065$ high-quality SMAL models fitted poses from $40$ animal categories for DSPart-Animal. Experimental results demonstrate the potential of our dataset in training robust part segmentation models, effectively bridging the gap between synthetic and real-world data.", "keywords": [ "Synthetic dataset", "Semantic Part Segmentation", "3D Parts Annotation" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "K3KrOsR6y9", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Bilgehan Sel", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "", "position": "PhD student" }, { "name": "Ming Jin", "gender": "unknown", "institution": "Virginia Tech", "country": "", "position": "Assistant Professor" }, { "name": "Ruoxi Jia", "gender": "unknown", "institution": "Virginia Tech", "country": "US", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 51, "n_ref": 221, "n_ref_all": 234, "n_fig": 3, "n_tab": 6, "L_tab": 10259, "n_element_tab": 156, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1599, "n_element_tab_1": 81, "formula_len_all": 0, "formula_len_all_1": 16, "len_all": 200174, "len_all_1": 62959, "len_abs": 946, "len_title": 82, "len_sents": 60464, "len_sents_1": 31342, "n_sents": 387, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 34, "L_abs": 893, "title": "LLMs Can Plan Only If We Tell Them", "abs": "Large language models (LLMs) have demonstrated significant capabilities in natural language processing and reasoning, yet their effectiveness in autonomous planning has been under debate. While existing studies have utilized LLMs with external feedback mechanisms or in controlled environments for planning, these approaches often involve substantial computational and development resources due to the requirement for careful design and iterative backprompting. Moreover, even the most advanced LLMs like GPT-4 struggle to match human performance on standard planning benchmarks, such as the Blocksworld, without additional support. This paper investigates whether LLMs can independently generate long-horizon plans that rival human baselines. Our novel enhancements help achieve state-of-the-art results in planning benchmarks out-competing prior methods and human baselines all autonomously.", "keywords": [ "large language models", "decision-making", "planning" ], "rating_list": [ 6, 6, 6, 8, 6 ], "soundness_list": [ 4, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "K2jOacHUlO", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Bhuwan Dhingra", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Assistant Professor" }, { "name": "Hongyi Cai", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Undergrad student" }, { "name": "Sanxing Chen", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Yukun Huang", "gender": "Male", "institution": "Duke University", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 26, "n_ref": 43, "n_ref_all": 70, "n_fig": 2, "n_tab": 10, "L_tab": 7586, "n_element_tab": 601, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 225, "formula_len_all_1": 42, "len_all": 222046, "len_all_1": 60404, "len_abs": 1829, "len_title": 123, "len_sents": 80987, "len_sents_1": 31136, "n_sents": 595, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 75, "L_abs": 1841, "title": "Enhancing Large Language Models' Situated Faithfulness to External Contexts", "abs": "Large Language Models (LLMs) are often augmented with external information as contexts, but this external information can sometimes be inaccurate or even intentionally misleading. We argue that robust LLMs should demonstrate situated faithfulness, dynamically calibrating their trust in external information based on their confidence in the internal knowledge and the external context. To benchmark this capability, we evaluate LLMs across several QA datasets, including a newly created dataset featuring in-the-wild incorrect contexts sourced from Reddit posts. We show that when provided with both correct and incorrect contexts, both open-source and proprietary models tend to overly rely on external information, regardless of its factual accuracy. To enhance situated faithfulness, we propose two approaches: Self-Guided Confidence Reasoning (SCR) and Rule-Based Confidence Reasoning (RCR). SCR enables models to self-access the confidence of external information relative to their own internal knowledge to produce the most accurate answer. RCR, in contrast, extracts explicit confidence signals from the LLM and determines the final answer using predefined rules. \nOur results show that for LLMs with strong reasoning capabilities, such as GPT-4o and GPT-4o mini, SCR outperforms RCR, achieving improvements of up to 24.2\\% over a direct input augmentation baseline. Conversely, for a smaller model like Llama-3-8B, RCR outperforms SCR. Fine-tuning SCR with our proposed Confidence Reasoning Direct Preference Optimization (CR-DPO) method improves performance on both seen and unseen datasets, yielding an average improvement of 8.9\\% on Llama-3-8B. In addition to quantitative results, we offer insights into the relative strengths of SCR and RCR. Our findings highlight promising avenues for improving situated faithfulness in LLMs.", "keywords": [ "Large Language Model", "Knowledge Conflict", "Retrieval Augmented Generation", "Confidence Estimation", "Reasoning" ], "rating_list": [ 5, 8, 8, 8 ], "soundness_list": [ 3, 4, 4, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "K2Tqn8R9pu", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Aleksandra Slavkovic", "gender": "Female", "institution": "Pennsylvania State University", "country": "US", "position": "Full Professor" }, { "name": "Carlos J Soto", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Mark Shriver", "gender": "unknown", "institution": "Pennsylvania State Univ University Park", "country": "", "position": "" }, { "name": "Matthew Reimherr", "gender": "unknown", "institution": "Pennsylvania State University", "country": "US", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 3, "n_ref_uni": 44, "n_ref": 81, "n_ref_all": 108, "n_fig": 8, "n_tab": 1, "L_tab": 49, "n_element_tab": 9, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 49, "n_element_tab_1": 9, "formula_len_all": 842, "formula_len_all_1": 135, "len_all": 136567, "len_all_1": 67037, "len_abs": 1122, "len_title": 132, "len_sents": 42930, "len_sents_1": 31327, "n_sents": 347, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1128, "title": "Gaussian Differentially Private Human Faces Under a Face Radial Curve Representation", "abs": "In this paper we consider the problem of releasing a Gaussian Differentially Private (GDP) 3D human face. The human face is a complex structure with many features and inherently tied to one's identity. Protecting this data, in a formally private way, is important yet challenging given the dimensionality of the problem. We extend approximate DP techniques for functional data to the GDP framework. We further propose a novel representation, face radial curves, of a 3D face as a set of functions and then utilize our proposed GDP functional data mechanism. To preserve the shape of the face while injecting noise we rely on tools from shape analysis for our novel representation of the face. We show that our method preserves the shape of the average face and injects less noise than traditional methods for the same privacy budget. Our mechanism consists of two primary components, the first is generally applicable to function value summaries (as are commonly found in nonparametric statistics or functional data analysis) while the second is general to disk-like surfaces and hence more applicable than just to human faces.", "keywords": [ "differential privacy", "shape analysis", "functional data analysis" ], "rating_list": [ 6, 3, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 2, 1 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 2, 2 ] }, { "paper_id": "K2OWrXUVby", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Heng Huang", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Junyi Li", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Peiran Yu", "gender": "Female", "institution": "University of Texas at Arlington", "country": "US", "position": "Assistant Professor" }, { "name": "Ziyi Chen", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Postdoc" } ], "n_formula": 189, "n_formula_1": 50, "n_ref_uni": 63, "n_ref": 153, "n_ref_all": 245, "n_fig": 0, "n_tab": 5, "L_tab": 1469, "n_element_tab": 128, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 277, "n_element_tab_1": 32, "formula_len_all": 21462, "formula_len_all_1": 4001, "len_all": 232857, "len_all_1": 71629, "len_abs": 1186, "len_title": 168, "len_sents": 49627, "len_sents_1": 28900, "n_sents": 498, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 119, "L_abs": 1215, "title": "Provably Mitigating Corruption, Overoptimization, and Verbosity Simultaneously in Offline and Online RLHF/DPO Alignment", "abs": "Reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO) are emerging and important techniques to align large language models (LLM) with human preference. However, the quality of RLHF and DPO training is seriously compromised by ***C**orrupted* preference, reward ***O**veroptimization*, and bias towards ***V**erbosity*. To our knowledge, most existing works tackle only one of these important issues, and the few other works require much computation to estimate multiple reward models and lack theoretical guarantee of generalization ability. In this work, we propose RLHF-**COV** and DPO-**COV** algorithms that can simultaneously mitigate these three issues, in both offline and online settings. This ability is theoretically demonstrated by obtaining length-regularized generalization error rates for our DPO-COV algorithms trained on corrupted data, which match the best-known rates for simpler cases with clean data and without length regularization. Moreover, our DPO-COV algorithm is simple to implement without reward estimation, and is proved to be equivalent to our RLHF-COV algorithm, which directly implies the equivalence between the vanilla RLHF and DPO algorithms.", "keywords": [ "RLHF", "DPO", "large language model", "alignment" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "K1VT7ItD40", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Boyu Mi", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Hanqing Wang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Jiangmiao Pang", "gender": "Male", "institution": "Shanghai AI Laboratory ", "country": "CN", "position": "Research Scientist" }, { "name": "Yilun Chen", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Tai WANG", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 21, "n_ref": 74, "n_ref_all": 92, "n_fig": 8, "n_tab": 6, "L_tab": 1308, "n_element_tab": 73, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 467, "n_element_tab_1": 56, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 124413, "len_all_1": 57313, "len_abs": 1357, "len_title": 118, "len_sents": 41581, "len_sents_1": 26230, "n_sents": 352, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1364, "title": "Evolving Symbolic 3D Visual Grounder with Weakly Supervised Reflection", "abs": "Understanding the behavior of an end-to-end 3D visual grounder is challenging, especially when the grounder makes an unexpected prediction. Despite the llm agent-based grounders performing step-by-step interpretable reasoning, the cost for evaluation at scale is prohibitive. To address the challenges, in this work, we propose a novel fully interpretable symbolic framework for 3D visual grounding,\nnamely Evolvable Symbolic Visual Grounder (EASE), with much less inference cost and superior performance. Given a symbolic expression of a grounding description translated by an LLM, EASE calculates the feature of each concept utilizing a set of explicit programs in Python learned from a tiny subset of the training data. To learn this program library, we introduce a learning paradigm that continuously optimizes the programs on the training dataset by an LLM-based optimizer. We demonstrate that our paradigm is scalable when more data is involved. Experiments on ReferIt3D show EASE achieves 50.7% accuracy on Nr3D, which surpasses most training-free methods and has considerable advantages in inference time and cost. On Sr3D, EASE also has comparable overall performance with these approaches. Moreover, we perform extensive experiments to analyze the interpretability and feature quality and reveal the potential for reasoning and\ncondition level grounding.", "keywords": [ "3D Visual Grounding", "Code Generation" ], "rating_list": [ 3, 3, 5, 6, 3 ], "soundness_list": [ 2, 3, 2, 3, 1 ], "presentation_list": [ 1, 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3, 1 ], "confidence_list": [ 4, 3, 3, 2, 4 ] }, { "paper_id": "K1G8UKcEBO", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Ganzhao Yuan", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Associate Professor" } ], "n_formula": 220, "n_formula_1": 39, "n_ref_uni": 30, "n_ref": 109, "n_ref_all": 142, "n_fig": 2, "n_tab": 2, "L_tab": 1424, "n_element_tab": 33, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 383, "n_element_tab_1": 33, "formula_len_all": 32524, "formula_len_all_1": 6065, "len_all": 263321, "len_all_1": 79724, "len_abs": 1381, "len_title": 122, "len_sents": 56358, "len_sents_1": 25854, "n_sents": 809, "n_sents_1": 276, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1443, "title": "ADMM for Nonsmooth Composite Optimization under Orthogonality Constraints", "abs": "We consider a class of structured, nonconvex, nonsmooth optimization problems under orthogonality constraints, where the objectives combine a smooth function, a nonsmooth concave function, and a nonsmooth weakly convex function. This class of problems finds diverse applications in statistical learning and data science. Existing methods for addressing these problems often fail to exploit the specific structure of orthogonality constraints, struggle with nonsmooth functions, or result in suboptimal oracle complexity. We propose {\\sf OADMM}, an Alternating Direction Method of Multipliers (ADMM) designed to solve this class of problems using efficient proximal linearized strategies. Two specific variants of {\\sf OADMM} are explored: one based on Euclidean Projection ({\\sf OADMM-EP}) and the other on Riemannian Retraction ({\\sf OADMM-RR}). Under mild assumptions, we prove that {\\sf OADMM} converges to a critical point of the problem with an ergodic convergence rate of $\\mathcal{O}(1/\\epsilon^{3})$. Additionally, we establish a super-exponential convergence rate or polynomial convergence rate for {\\sf OADMM}, depending on the specific setting, under the Kurdyka-Lojasiewicz (KL) inequality. To the best of our knowledge, this is the first non-ergodic convergence result for this class of nonconvex nonsmooth optimization problems. Numerical experiments demonstrate that the proposed algorithm achieves state-of-the-art performance.", "keywords": [ "Orthogonality Constraints", "Nonconvex Optimization", "Nonsmooth Composite Optimization", "ADMM", "Convergence Analysis" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "K0oFDAPnU4", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Haixingzhao", "gender": "Male", "institution": "Qinghai Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Mingyuan Li", "gender": "Male", "institution": "Qinghai Normal University", "country": "CN", "position": "MS student" }, { "name": "Zhaoqiang Liu", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 33, "n_formula_1": 30, "n_ref_uni": 40, "n_ref": 65, "n_ref_all": 80, "n_fig": 2, "n_tab": 6, "L_tab": 3226, "n_element_tab": 269, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 3133, "n_element_tab_1": 252, "formula_len_all": 1342, "formula_len_all_1": 1339, "len_all": 105152, "len_all_1": 61090, "len_abs": 991, "len_title": 132, "len_sents": 27904, "len_sents_1": 25089, "n_sents": 192, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 994, "title": "A Unified Framework for Hierarchical Diffusion via Simplicial Complexes", "abs": "In this paper, we propose a unified framework for hierarchical diffusion via simplicial complexes (HDSC), which enables adaptive diffusion across different levels of simplicial complexes, including nodes, edges, and triangles. To ensure the accuracy and consistency of information transmission during the diffusion process, we investigate topological consistency constraints, achieving efficient coupling between structures at various levels. Additionally, by introducing a time-dependent topological memory mechanism, we further enhance the smoothness and coherence of global information flow, enabling features at different levels to diffuse cooperatively throughout the entire graph structure. Experimental results demonstrate that HDSC exhibits significant performance advantages over traditional methods. Furthermore, as the complexity and dimensionality of the graph increase, HDSC continues to maintain its superiority, effectively avoiding the phenomenon of node feature homogenization.", "keywords": [ "Graph Neural Networks; Simplicial Complex; Graph Diffusion Equation; Hierarchical diffusion process; Topological Consistency." ], "rating_list": [ 3, 3, 1, 8 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 2, 1, 1, 3 ], "contribution_list": [ 2, 1, 3, 4 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "K0WzGPTGwx", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Jeroen Berrevoets", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Julianna Piskorz", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Max Ruiz Luyten", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Mihaela van der Schaar", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Nicolas Huynh", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 59, "n_formula_1": 19, "n_ref_uni": 63, "n_ref": 125, "n_ref_all": 163, "n_fig": 10, "n_tab": 10, "L_tab": 2369, "n_element_tab": 99, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 810, "n_element_tab_1": 24, "formula_len_all": 5642, "formula_len_all_1": 1570, "len_all": 263115, "len_all_1": 78721, "len_abs": 1228, "len_title": 126, "len_sents": 79293, "len_sents_1": 33132, "n_sents": 711, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1210, "title": "Enhancing Treatment Effect Estimation with Generation-Driven Data Augmentation", "abs": "We introduce $\\texttt{GATE}$, a framework for improving the estimation of conditional average treatment effects (CATE) from observational data. Our framework leverages generative models to selectively augment datasets with synthetic potential outcomes, thus addressing the covariate shift problem inherent in CATE estimation. Crucially, $\\texttt{GATE}$ enables the integration of external knowledge into downstream CATE models, by leveraging generative models trained on external data sources, such as large language models (LLMs). These models utilise rich contextual information, such as dataset metadata, to generate synthetic potential outcomes grounded in real-world contexts. While generative models can introduce bias when imperfect, we theoretically demonstrate that restricting augmentation to a carefully chosen subsets of the covariate space can allow to achieve performance gains despite these imperfections. Empirically, $\\texttt{GATE}$ instantiated with LLMs consistently improves a wide range of CATE estimators, narrowing performance gaps between learners and underscoring the advantages of incorporating external knowledge through generative augmentation, particularly in small-sample regimes.", "keywords": [ "treatment effect estimation", "causal inference", "data augmentation", "generative models" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "Jztt1nrjAM", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Camille Thibault", "gender": "Female", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "MS student" }, { "name": "Florence Laflamme", "gender": "Female", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "", "position": "Undergrad student" }, { "name": "Gabrielle P\u00e9loquin-Skulski", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jacob-Junqi Tian", "gender": "unknown", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "", "position": "Intern" }, { "name": "Jean-Fran\u00e7ois Godbout", "gender": "Male", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "US", "position": "Full Professor" }, { "name": "Kellin Pelrine", "gender": "unknown", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "Reihaneh Rabbany", "gender": "Female", "institution": "McGill University", "country": "", "position": "Assistant Professor" }, { "name": "Yuxiang Guan", "gender": "Male", "institution": "McMaster University", "country": "CA", "position": "Undergrad student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 65, "n_ref": 107, "n_ref_all": 124, "n_fig": 1, "n_tab": 7, "L_tab": 8164, "n_element_tab": 976, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 7327, "n_element_tab_1": 975, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 225613, "len_all_1": 77419, "len_abs": 1024, "len_title": 92, "len_sents": 65221, "len_sents_1": 28083, "n_sents": 491, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 1028, "title": "A Guide to Misinformation Detection Datasets", "abs": "Misinformation is a complex societal issue, and mitigating solutions are difficult to create due to data deficiencies. To address this problem, we have curated the largest collection of (mis)information datasets in the literature, totaling 75. From these, we evaluated the quality of all of the 35 datasets that consist of statements or claims. We assess these datasets to identify those with solid foundations for empirical work and those with flaws that could result in misleading and non-generalizable results, such as insufficient label quality, spurious correlations, or political bias. We further provide state-of-the-art baselines on all these datasets, but show that regardless of label quality, categorical labels may no longer give an accurate evaluation of detection model performance. We discuss alternatives to mitigate this problem. Overall, this guide aims to provide a roadmap for obtaining higher quality data and conducting more effective evaluations, ultimately improving research in misinformation detection.", "keywords": [ "Misinformation", "Fact-Checking", "Survey", "Datasets", "LLM Judge", "Evaluation" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "JzLcKWtGnl", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Chao Zhang", "gender": "Female", "institution": "Beijing Digital Native Digital City Research Center", "country": "CN", "position": "Researcher" }, { "name": "Jiaxing Qi", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Ruifei Ma", "gender": "Female", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "MS student" }, { "name": "Xiangde Liu", "gender": "Male", "institution": "Beijing Digital Native Digital City Research Center", "country": "CN", "position": "Researcher" }, { "name": "Xiaoyan Wang", "gender": "Female", "institution": "Beijing Digital Native Digital City Research Center", "country": "CN", "position": "Researcher" }, { "name": "Yifan Xu", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Zeju Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Zhifei Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 83, "n_ref_all": 98, "n_fig": 3, "n_tab": 6, "L_tab": 778, "n_element_tab": 125, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 574, "n_element_tab_1": 92, "formula_len_all": 302, "formula_len_all_1": 303, "len_all": 134912, "len_all_1": 63295, "len_abs": 1296, "len_title": 141, "len_sents": 38122, "len_sents_1": 29458, "n_sents": 268, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1300, "title": "Spatial 3D-LLM: Progressive Spatial Awareness for Advanced 3D Vision-Language Understanding", "abs": "New era has unlocked exciting possibilities for extending Large Language Models (LLMs) to tackle 3D vision-language tasks. However, most existing 3D Multimodal LLMs (MLLMs) rely on holistic 3D scene information or specifically designated regions for 3D vision-language tasks, failing to capture multi-level location-based information.\nAddressing these concerns, we present Spatial 3D-LLM, a 3D MLLM specifically designed to enhance spatial perception and reasoning for 3D vision-language tasks by enriching the spatial embeddings of 3D scenes.\nSpatial 3D-LLM incorporates an LLM backbone and a meticulously designed progressive spatial awareness scheme that captures spatial information as the perception field expands, generating location-enriched 3D scene embeddings that serve as visual prompt.\nAdditionally, we introduce two novel tasks, namely 3D object distance measurement and 3D layout editing, and construct a 3D instruction dataset MODEL, to inspire more profound 3D spatial perception capabilities.\nExperimental results demonstrate that Spatial 3D-LLM achieves state-of-the-art performance across a wide range of 3D vision-language tasks, revealing the improvements stemmed from our progressive spatial awareness scheme of mining more profound spatial information and the proposed dataset.", "keywords": [ "3D-MLLM", "3D scene understanding" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "JytL2MrlLT", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "De-Chuan Zhan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Han-Jia Ye", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Huai-Hong Yin", "gender": "Male", "institution": "nanjing university", "country": "", "position": "MS student" }, { "name": "Wei-Lun Chao", "gender": "Male", "institution": "Ohio State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 6, "n_ref_uni": 67, "n_ref": 161, "n_ref_all": 206, "n_fig": 4, "n_tab": 9, "L_tab": 1037, "n_element_tab": 153, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 924, "n_element_tab_1": 60, "formula_len_all": 846, "formula_len_all_1": 494, "len_all": 176428, "len_all_1": 70948, "len_abs": 1268, "len_title": 135, "len_sents": 51220, "len_sents_1": 34568, "n_sents": 369, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1279, "title": "Revisiting Nearest Neighbor for Tabular Data: A Deep Tabular Baseline Two Decades Later", "abs": "The widespread enthusiasm for deep learning has recently expanded into the domain of tabular data. Recognizing that the advancement in deep tabular methods is often inspired by classical methods, e.g., integration of nearest neighbors into neural networks, we investigate whether these classical methods can be revitalized with modern techniques.\nWe revisit a differentiable version of $K$-nearest neighbors (KNN) --- Neighbourhood Components Analysis (NCA) --- originally designed to learn a linear projection to capture semantic similarities between instances, and seek to gradually add modern deep learning techniques on top. Surprisingly, our implementation of NCA using SGD and without dimensionality reduction already achieves decent performance on tabular data, in contrast to the results of using existing toolboxes like scikit-learn. Further equipping NCA with deep representations and additional training stochasticity significantly enhances its capability, being on par with the leading tree-based method CatBoost and outperforming existing deep tabular models in both classification and regression tasks on 300 datasets. We conclude our paper by analyzing the factors behind these improvements, including loss functions, prediction strategies, and deep architectures.", "keywords": [ "Tabular data", "tabular machine learning", "deep tabular models" ], "rating_list": [ 8, 5, 3, 8, 6 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 1, 3, 3 ], "confidence_list": [ 4, 4, 4, 4, 5 ] }, { "paper_id": "Jyrwd2wja9", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Javad Lavaei", "gender": "unknown", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Ying Chen", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" } ], "n_formula": 41, "n_formula_1": 42, "n_ref_uni": 28, "n_ref": 40, "n_ref_all": 48, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3733, "formula_len_all_1": 3895, "len_all": 98038, "len_all_1": 62629, "len_abs": 1455, "len_title": 136, "len_sents": 32259, "len_sents_1": 25605, "n_sents": 251, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1465, "title": "Measurement Manipulation of the Matrix Sensing Problem to Improve Optimization Landscape", "abs": "This work studies the matrix sensing (MS) problem through the lens of the Restricted Isometry Property (RIP). It has been shown in several recent papers that two different techniques of convex relaxations and local search methods for the MS problem both require the RIP constant to be less than 0.5 while most real-world problems have their RIPs close to 1. The existing literature guarantees a small RIP constant only for sensing operators having an i.i.d. Gaussian distribution, and it is well-known that the MS problem could have a complicated landscape when the RIP is greater than 0.5. In this work, we address this issue and improve the optimization landscape by developing two results. First, we show that any sensing operator with a model not too distant from i.i.d. Gaussian has a slightly higher RIP than i.i.d. Gaussian, and that its RIP constant can be reduced to match the RIP constant of an i.i.d. Gaussian via slightly increasing the number of measurements. Second, we show that if the sensing operator has an arbitrary distribution, it can be modified in such a way that the resulting operator will act as a perturbed Gaussian with a lower RIP constant. Our approach is a preconditioning technique that replaces each sensing matrix with a weighted sum of all sensing matrices. We numerically demonstrate that the RIP constants for different distributions can be reduced from almost 1 to less than 0.5 via the preconditioning of the sensing operator.", "keywords": [ "non-convex optimization", "low-rank matrix optimization", "matrix sensing", "preconditioning algorithm" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "Jyh0DR4fFE", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Jonathan Aellen", "gender": "Male", "institution": "University of Basel", "country": "CH", "position": "PhD student" }, { "name": "Marcello Massimo Negri", "gender": "Male", "institution": "University of Basel", "country": "", "position": "PhD student" }, { "name": "Volker Roth", "gender": "Male", "institution": "University of Basel", "country": "CH", "position": "Full Professor" } ], "n_formula": 55, "n_formula_1": 16, "n_ref_uni": 49, "n_ref": 74, "n_ref_all": 139, "n_fig": 12, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4918, "formula_len_all_1": 1480, "len_all": 176150, "len_all_1": 74771, "len_abs": 1119, "len_title": 87, "len_sents": 57307, "len_sents_1": 34032, "n_sents": 497, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 39, "L_abs": 1097, "title": "Injective flows for star-like manifolds", "abs": "Normalizing Flows (NFs) are powerful and efficient models for density estimation. When modeling densities on manifolds, NFs can be generalized to injective flows but the Jacobian determinant becomes computationally prohibitive. Current approaches either consider bounds on the log-likelihood or rely on some approximations of the Jacobian determinant. In contrast, we propose injective flows for star-like manifolds and show that for such manifolds we can compute the Jacobian determinant exactly and efficiently. This aspect is particularly relevant for variational inference settings, where no samples are available and only some unnormalized target is known. Among many, we showcase the relevance of modeling densities on star-like manifolds in two settings. Firstly, we introduce a novel Objective Bayesian approach for penalized likelihood models by interpreting level-sets of the penalty as star-like manifolds. Secondly, we consider probabilistic mixing models and introduce a general method for variational inference by defining the posterior of mixture weights on the probability simplex.", "keywords": [ "Normalizing Flows", "Injective Flows", "Bayesian Inference", "Variational Inference", "Objective Bayes" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "JyQYYjtO88", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Chenyi Zhang", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Tongyang Li", "gender": "Male", "institution": "Peking University", "country": "", "position": "Assistant Professor" }, { "name": "Weiyuan Gong", "gender": "Male", "institution": "SEAS, Harvard University", "country": "", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 24, "n_ref_uni": 51, "n_ref": 108, "n_ref_all": 125, "n_fig": 1, "n_tab": 2, "L_tab": 56, "n_element_tab": 4, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2111, "formula_len_all_1": 2014, "len_all": 132010, "len_all_1": 70772, "len_abs": 1520, "len_title": 107, "len_sents": 36708, "len_sents_1": 32540, "n_sents": 244, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1697, "title": "Robustness of Quantum Algorithms for Nonconvex Optimization", "abs": "In this paper, we systematically study quantum algorithms for finding an $\\epsilon$-approximate second-order stationary point ($\\epsilon$-SOSP) of a $d$-dimensional nonconvex function, a fundamental problem in nonconvex optimization, with noisy zeroth- or first-order oracles as inputs. We first prove that, up to noise of $O(\\epsilon^{10}/d^5)$, perturbed accelerated gradient descent equipped with quantum gradient estimation takes $O(\\log d/\\epsilon^{1.75})$ quantum queries to find an $\\epsilon$-SOSP. We then prove that standard perturbed gradient descent is robust to the noise of $O(\\epsilon^6/d^4)$ and $O(\\epsilon/d^{0.5+\\zeta})$ for any $\\zeta>0$ on the zeroth- and first-order oracles, respectively, which provides a quantum algorithm with poly-logarithmic query complexity. Furthermore, we propose a stochastic gradient descent algorithm using quantum mean estimation on the Gaussian smoothing of noisy oracles, which is robust to $O(\\epsilon^{1.5}/d)$ and $O(\\epsilon/\\sqrt{d})$ noise on the zeroth- and first-order oracles, respectively. The quantum algorithm takes $O(d^{2.5}/\\epsilon^{3.5})$ and $O(d^2/\\epsilon^3)$ queries to the two oracles, giving a polynomial speedup over the classical counterparts. As a complement, we characterize the domains where quantum algorithms can find an $\\epsilon$-SOSP with poly-logarithmic, polynomial, or exponential number of queries in $d$, or the problem is information-theoretically unsolvable even with an infinite number of queries. In addition, we prove an $\\Omega(\\epsilon^{-12/7})$ lower bound on $\\epsilon$ for any randomized classical and quantum algorithm to find an $\\epsilon$-SOSP using either noisy zeroth- or first-order oracles.", "keywords": [ "Nonconvex optimization", "Robustness", "Quantum algorithms" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 2, 4, 4 ], "presentation_list": [ 1, 4, 4 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "Jy17uvzNe5", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Di He", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Ganzhao Yuan", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Associate Professor" }, { "name": "Pengxiang Xu", "gender": "unknown", "institution": "Peng Cheng Laboratory", "country": "", "position": "Associate Professor" }, { "name": "Xiao Wang", "gender": "Not Specified", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Associate Professor" } ], "n_formula": 242, "n_formula_1": 42, "n_ref_uni": 38, "n_ref": 102, "n_ref_all": 140, "n_fig": 6, "n_tab": 4, "L_tab": 1040, "n_element_tab": 35, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 259, "n_element_tab_1": 16, "formula_len_all": 28727, "formula_len_all_1": 4582, "len_all": 293856, "len_all_1": 71768, "len_abs": 3776, "len_title": 101, "len_sents": 66625, "len_sents_1": 25418, "n_sents": 1305, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1636, "title": "Block Coordinate Descent Methods for Optimization under J-Orthogonality Constraints with Applications", "abs": "The J-orthogonal matrix, also referred to as the hyperbolic orthogonal matrix, is a class of special orthogonal matrix in hyperbolic space, notable for its advantageous properties. These matrices are integral to optimization under J-orthogonal constraints, which have widespread applications in statistical learning and data science. However, addressing these problems is generally challenging due to their non-convex nature and the computational intensity of the constraints. Currently, algorithms for tackling these challenges are limited. This paper introduces \\textbf{JOBCD}, a novel Block Coordinate Descent method designed to address optimizations with J-orthogonality constraints. We explore two specific variants of \\textbf{JOBCD}: one based on a Gauss-Seidel strategy (\\textbf{GS-JOBCD}), the other on a variance-reduced and Jacobi strategy (\\textbf{VR-J-JOBCD}). Notably, leveraging the parallel framework of a Jacobi strategy, \\textbf{VR-J-JOBCD} integrates variance reduction techniques to decrease oracle complexity in the minimization of finite-sum functions. For both \\textbf{GS-JOBCD} and \\textbf{VR-J-JOBCD}, we establish the oracle complexity under mild conditions and strong limit-point convergence results under the Kurdyka-Lojasiewicz inequality. To demonstrate the effectiveness of our method, we conduct experiments on hyperbolic eigenvalue problems, hyperbolic structural probe problems, and the ultrahyperbolic knowledge graph embedding problem. Extensive experiments using both real-world and synthetic data demonstrate that \\textbf{JOBCD} consistently outperforms state-of-the-art solutions, by large margins.", "keywords": [ "Orthogonality Constraints", "Nonconvex Optimization", "Nonsmooth Composite Optimization", "Block Coordinate Descent", "Convergence Analysis" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "Jy0MJYZEuN", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Dingjie Fu", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Salman Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Associate Professor" }, { "name": "Shiming Chen", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Fahad Shahbaz Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 22, "n_ref": 44, "n_ref_all": 62, "n_fig": 10, "n_tab": 7, "L_tab": 2214, "n_element_tab": 305, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3780, "n_element_tab_1": 241, "formula_len_all": 454, "formula_len_all_1": 656, "len_all": 128975, "len_all_1": 58240, "len_abs": 1319, "len_title": 152, "len_sents": 30187, "len_sents_1": 23522, "n_sents": 236, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1342, "title": "Induction Rather Than Imagination: Generative Zero-Shot Learning Via Inductive Variational Autoencoder", "abs": "Remarkable progress in zero-shot learning (ZSL) has been achieved using generative models. However, existing generative ZSL methods merely generate (imagine) the visual features from scratch guided by the strong class semantic vectors annotated by experts, resulting in suboptimal generative performance and limited scene generalization. To address these and advance ZSL, we propose an inductive variational autoencoder for generative zero-shot learning, dubbed GenZSL. Mimicking human-level concept learning, GenZSL operates by \\textit{inducting} new class samples from similar seen classes using weak class semantic vectors derived from target class names (i.e., CLIP text embedding). To ensure the generation of informative samples for training an effective ZSL classifier, our GenZSL incorporates two key strategies. Firstly, it employs class diversity promotion to enhance the diversity of class semantic vectors. Secondly, it utilizes target class-guided information boosting criteria to optimize the model. Extensive experiments conducted on three popular benchmark datasets showcase the superiority and potential of our GenZSL with significant efficacy and efficiency over f-VAEGAN, e.g., 24.7\\% performance gains and more than $60\\times$ faster training speed on AWA2. Codes are available at https://anonymous.4open.science/r/GenZSL.", "keywords": [ "zero-shot learning", "generative model" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 4, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "JxhgSAnrsG", "primary_area": "applications to robotics, autonomy, planning", "n_author": 11, "author_info_list": [ { "name": "Dongming Wu", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Intern" }, { "name": "Fan Jia", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Jianbing Shen", "gender": "Male", "institution": "University of Macau", "country": "CN", "position": "Full Professor" }, { "name": "Tiancai Wang", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Weixin Mao", "gender": "Male", "institution": "Waseda University", "country": "JP", "position": "PhD student" }, { "name": "Xiangyu Zhang", "gender": "Male", "institution": "MEGVII Technology", "country": "CN", "position": "Principal Researcher" }, { "name": "Xing Wei", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yifan Bai", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yingfei Liu", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Yucheng Zhao", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Ziheng Zhang", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 41, "n_ref": 91, "n_ref_all": 112, "n_fig": 4, "n_tab": 6, "L_tab": 2782, "n_element_tab": 216, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 2782, "n_element_tab_1": 216, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 125795, "len_all_1": 69320, "len_abs": 1422, "len_title": 109, "len_sents": 32118, "len_sents_1": 31789, "n_sents": 235, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1428, "title": "Is a 3D-Tokenized LLM the Key to Reliable Autonomous Driving?", "abs": "Rapid advancements in Autonomous Driving (AD) tasks turned a significant shift toward end-to-end fashion, particularly in the utilization of vision-language models (VLMs) that integrate robust logical reasoning and cognitive abilities to enable comprehensive end-to-end planning. However, these VLM-based approaches tend to integrate 2D vision tokenizers and a large language model (LLM) for ego-car planning, which lack 3D geometric priors as a cornerstone of reliable planning. Naturally, this observation raises a critical concern: Can a 2D-tokenized LLM accurately perceive the 3D environment? Our evaluation of current VLM-based methods across 3D object detection, vectorized map construction, and environmental caption suggests that the answer is, unfortunately, NO. In other words, 2D-tokenized LLM fails to provide reliable autonomous driving. In response, we introduce DETR-style 3D perceptrons as 3D tokenizers, which connect LLM with a one-layer linear projector. This simple yet elegant strategy, termed Atlas, harnesses the inherent priors of the 3D physical world, enabling it to simultaneously process high-resolution multi-view images and employ spatiotemporal modeling. Despite its simplicity, Atlas demonstrates superior performance in both 3D detection and ego planning tasks on nuScenes dataset, proving that 3D-tokenized LLM is the key to reliable autonomous driving. The code and datasets will be released.", "keywords": [ "Large Language Model; Vision-Language Representation Learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "Jwtpbhheoy", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Guanting Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Shang Liu", "gender": "Male", "institution": "Imperial College London, Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Xiaocheng Li", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Assistant Professor" }, { "name": "Zhongze Cai", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "PhD student" } ], "n_formula": 79, "n_formula_1": 20, "n_ref_uni": 31, "n_ref": 101, "n_ref_all": 129, "n_fig": 9, "n_tab": 7, "L_tab": 841, "n_element_tab": 72, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 8017, "formula_len_all_1": 1306, "len_all": 243854, "len_all_1": 69925, "len_abs": 1652, "len_title": 150, "len_sents": 83571, "len_sents_1": 34185, "n_sents": 780, "n_sents_1": 264, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1736, "title": "Towards Better Understanding of In-Context Learning Ability from In-Context Uncertainty Quantification", "abs": "Predicting simple function classes has been widely used as a testbed for developing theory and understanding of the trained Transformer's in-context learning (ICL) ability. In this paper, we revisit the training of Transformers on linear regression tasks, and different from the existing literature, we consider a bi-objective prediction task of predicting both the conditional expectation $\\mathbb{E}[Y|X]$ and the conditional variance Var$(Y|X)$. This additional uncertainty quantification objective provides a handle to (i) better design out-of-distribution experiments to distinguish ICL from in-weight learning (IWL) and (ii) make a better separation between the algorithms with and without using the prior information of the training distribution. Theoretically, we show that the trained Transformer reaches near Bayes optimum, suggesting the usage of the information of the training distribution. Our method can be extended to other cases. Specifically, with the Transformer's context window $S$, we prove a new generalization bound of $\\tilde{\\mathcal{O}}(\\sqrt{\\min\\{S, T\\}/(n T)})$ on $n$ tasks with sequences of length $T$, providing sharper analysis compared to previous results of $\\tilde{\\mathcal{O}}(\\sqrt{1/n})$. Empirically, we illustrate that while the trained Transformer behaves as the Bayes-optimal solution as a natural consequence of supervised training in distribution, it does not necessarily perform a Bayesian inference when facing task shifts, in contrast to the \\textit{equivalence} between these two proposed in many existing literature. We also demonstrate the trained Transformer's ICL ability over covariates shift and prompt-length shift and interpret them as a generalization over a meta distribution.", "keywords": [ "in-context learning", "Transformer", "uncertainty quantification", "Bayes optimal", "supervised learning" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "JwrnoB1tR0", "primary_area": "applications to robotics, autonomy, planning", "n_author": 10, "author_info_list": [ { "name": "Dong Wang", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Huchuan Lu", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Professor" }, { "name": "Jing Shao", "gender": "Female", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Lijun Wang", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Shiyu Tang", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "MS student" }, { "name": "Talas Fu", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "MS student" }, { "name": "Yang Liu", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yifan Wang", "gender": "unknown", "institution": "Dalian University of Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuanhang Zhang", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "MS student" }, { "name": "\u518d\u658c \u5f20", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 56, "n_ref": 130, "n_ref_all": 155, "n_fig": 7, "n_tab": 4, "L_tab": 453, "n_element_tab": 57, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 777, "n_element_tab_1": 48, "formula_len_all": 82, "formula_len_all_1": 82, "len_all": 195211, "len_all_1": 59808, "len_abs": 2216, "len_title": 97, "len_sents": 40782, "len_sents_1": 29137, "n_sents": 289, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1888, "title": "AD-H: Autonomous Driving with Hierarchical Agents", "abs": "Due to the impressive capabilities of multimodal large language models (MLLMs), recent works have focused on employing MLLM-based agents for autonomous driving in large-scale and dynamic environments. However, prevalent approaches often directly use MLLMs to translate high-level instructions into low-level vehicle control signals. This approach deviates from the inherent language generation paradigm of MLLMs and fails to fully harness their emergent capabilities. As a result, the generalizability of these methods is limited by the autonomous driving datasets used during fine-tuning.\nTo tackle this challenge, we propose AD-H, a hierarchical framework that enables two agents (the MLLM planner and the controller) to collaborate. The MLLM planner perceives environmental information and high-level instructions to generate mid-level, fine-grained driving commands, which the controller then executes as actions. This compositional paradigm liberates the MLLM from low-level control signal decoding, thus fully leveraging its high-level perception, reasoning, and planning capabilities. Furthermore, the fine-grained commands provided by the MLLM planner enable the controller to perform actions more effectively.\nTo train AD-H, we build a new autonomous driving dataset with hierarchical action annotations encompassing multiple levels of instructions and driving commands. \nComprehensive closed-loop evaluations demonstrate several key advantages of our proposed AD-H system.\nFirst, AD-H can notably outperform state-of-the-art methods in achieving exceptional driving performance, even exhibiting self-correction capabilities during vehicle operation, a scenario not encountered in the training dataset. Second, AD-H demonstrates superior generalization under long-horizon instructions and novel environmental conditions, significantly surpassing current state-of-the-art methods.", "keywords": [ "AutoDriving; Agent", "Planning" ], "rating_list": [ 5, 3, 6, 1, 5 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 3, 1, 2 ], "confidence_list": [ 3, 5, 5, 5, 4 ] }, { "paper_id": "JwoQZ9NKtH", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Hanning Chen", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Mohsen Imani", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Assistant Professor" }, { "name": "Wenjun Huang", "gender": "unknown", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Yang Ni", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Yezi Liu", "gender": "Female", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 16, "n_ref_uni": 67, "n_ref": 104, "n_ref_all": 142, "n_fig": 6, "n_tab": 2, "L_tab": 2659, "n_element_tab": 80, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 2659, "n_element_tab_1": 80, "formula_len_all": 2015, "formula_len_all_1": 971, "len_all": 171085, "len_all_1": 63880, "len_abs": 1322, "len_title": 83, "len_sents": 45518, "len_sents_1": 28236, "n_sents": 377, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1326, "title": "Cauchy-Schwarz Fairness Regularizer", "abs": "In this paper, we propose a novel approach to fair machine learning, the Cauchy-Schwarz fairness regularizer, which minimizes the Cauchy-Schwarz divergence between the prediction distribution and sensitive attributes. While existing methods effectively reduce bias as indicated by low values on specific fairness metrics, they frequently struggle to achieve a balanced performance across various fairness definitions. For example, many approaches may successfully attain low demographic parity yet still demonstrate significant disparities in equal opportunity. Theoretical studies have shown that the Cauchy-Schwarz divergence provides a tighter bound compared to the Kullback-Leibler divergence and gap parity, suggesting its potential to improve fairness in machine learning models. Our empirical evaluation, conducted on four tabular datasets and one image dataset, demonstrates that the Cauchy-Schwarz fairness regularizer achieves a more balanced performance across fairness metrics while maintaining satisfactory utility. It outperforms existing fairness approaches, providing a superior trade-off between fairness and utility. In addition, the Cauchy-Schwarz fairness regularizer is a versatile, plug-and-play fairness regularizer that can be easily integrated into various machine learning models to promote fairness.", "keywords": [ "Fairness", "Machine Learning", "Cauchy-Schwarz Divergence" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 1, 3 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "JwoCs9O3QL", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Emily Lawton", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Applied Scientist" }, { "name": "Jack W Stokes", "gender": "unknown", "institution": "Microsoft Research", "country": "", "position": "Principal Research Software Development Engineer" }, { "name": "Reshmi Ghosh", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Robert Sim", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Principal Researcher" }, { "name": "Vitor R. Carvalho", "gender": "unknown", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Xuefeng Du", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "PhD student" }, { "name": "Yixuan Li", "gender": "Female", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" }, { "name": "ahmed salem", "gender": "Male", "institution": "Microsoft", "country": "DE", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 38, "n_ref": 78, "n_ref_all": 108, "n_fig": 6, "n_tab": 7, "L_tab": 1522, "n_element_tab": 227, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1711, "n_element_tab_1": 237, "formula_len_all": 1012, "formula_len_all_1": 899, "len_all": 145755, "len_all_1": 76318, "len_abs": 1418, "len_title": 118, "len_sents": 43471, "len_sents_1": 34353, "n_sents": 319, "n_sents_1": 243, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1429, "title": "VLMGuard: Defending VLMs against Malicious Prompts via Unlabeled Data", "abs": "Vision-language models (VLMs) are essential for contextual understanding of both visual and textual information. However, their vulnerability to adversarially manipulated inputs presents significant risks, leading to compromised outputs and raising concerns about the reliability in VLM-integrated applications. Detecting these malicious prompts is thus crucial for maintaining trust in VLM generations. A major challenge in developing a safeguarding prompt classifier is the lack of a large amount of labeled benign and malicious data. To\n address the issue, we introduce VLMGuard, a novel learning framework that leverages the unlabeled user prompts in the wild for malicious prompt detection. These unlabeled prompts, which naturally arise when VLMs are deployed in the open world, consist of both benign and malicious information. To harness the unlabeled data, we present an automated maliciousness estimation score for distinguishing between benign and malicious samples within this unlabeled mixture, thereby enabling the training of a binary prompt classifier on top. Notably, our framework does not require extra human annotations, offering strong flexibility and practicality for real-world applications. Extensive experiment shows VLMGuard achieves superior detection results, significantly outperforming state-of-the-art methods. Disclaimer: This paper may contain offensive examples; reader discretion is advised.", "keywords": [ "Malicious prompt detection" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "Jwgw3znxT3", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Yancheng Wang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Yingzhen Yang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 23, "n_formula_1": 16, "n_ref_uni": 36, "n_ref": 158, "n_ref_all": 191, "n_fig": 4, "n_tab": 4, "L_tab": 436, "n_element_tab": 38, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 202, "n_element_tab_1": 10, "formula_len_all": 2314, "formula_len_all_1": 1109, "len_all": 170328, "len_all_1": 64879, "len_abs": 1545, "len_title": 125, "len_sents": 49085, "len_sents_1": 31031, "n_sents": 329, "n_sents_1": 195, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1557, "title": "Efficient Visual Transformer by Information Bottleneck Inspired Token Merging", "abs": "Self-attention and transformers have been widely used in deep learning. Recent efforts have been devoted to incorporating transformer blocks into different types of neural architectures, including those with convolutions, leading to various vision transformers for computer vision tasks. In this paper, we propose a novel and compact transformer block, Transformer with Information Bottleneck inspired Token Merging, or IBTM. IBTM performs token merging in a learnable scheme. Our IBTM is compatible with many popular and compact transformer networks, such as MobileViT and EfficientViT, and it reduces the FLOPs and the inference time of the vision transformers while maintaining or even improving the prediction accuracy. In the experiments, we replace all the transformer blocks in popular vision transformers, including MobileViT, EfficientViT, ViT, and Swin, with IBTM blocks, leading to IBTM networks with different backbones. The IBTM is motivated by the reduction of the Information Bottleneck (IB), and a novel and separable variational upper bound for the IB loss is derived. The architecture of mask module in our IBTM blocks which generate the token merging mask is designed to reduce the derived upper bound for the IB loss. Extensive results on image classification and object detection evidence that IBTM renders compact and efficient vision transformers with comparable or much better prediction accuracy than the original vision transformers. The code of IBTM is available at \\url{https://anonymous.4open.science/r/IBTM_Transformers-053B/}.", "keywords": [ "Visual Transformer", "Token Merging", "Information Bottleneck" ], "rating_list": [ 3, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "JwNQP2dNhD", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Feng Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Lihao Yin", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Zhen Wang", "gender": "Not Specified", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Zhihai Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Hui-Ling Zhen", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Principal Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 19, "n_ref": 54, "n_ref_all": 69, "n_fig": 7, "n_tab": 8, "L_tab": 4378, "n_element_tab": 259, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 4510, "n_element_tab_1": 139, "formula_len_all": 278, "formula_len_all_1": 31, "len_all": 134706, "len_all_1": 61061, "len_abs": 1850, "len_title": 135, "len_sents": 38029, "len_sents_1": 25938, "n_sents": 319, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1860, "title": "Searching Strengthens Large Language Models in Finding Bugs of Deep Learning Libraries", "abs": "Ensuring the quality of deep learning libraries is crucial, as bugs can have significant consequences for downstream software. Fuzzing, a powerful testing method, generates random programs to test software. Generally, effective fuzzing requires generated programs to meet three key criteria: rarity, validity, and variety, among which rarity is most critical for bug detection, as it determines the algorithm's ability to detect bugs. However, current large language model (LLM) based fuzzing approaches struggle to effectively explore the program generation space which results in insufficient rarity and the lack of post-processing leads to a large number of invalid programs and inadequate validity. This paper proposes EvAFuzz, a novel approach that combines Evolutionary Algorithms with LLMs to Fuzz DL libraries. For rarity, EvAFuzz uses a search algorithm to guide LLMs in efficiently exploring the program generation space, iteratively generating increasingly rare programs. For validity, EvAFuzz incorporates a feedback scheme, enabling LLMs to correct invalid programs and achieve high validity. For variety, EvAFuzz constructs a large parent selection space, enriching the diversity of selected parents, and thereby enhancing the variety of generated programs. Our experiments show that EvAFuzz outperforms the previous state-of-the-art (SOTA) in several key metrics. First, in the same version of PyTorch, EvAFuzz detects nine unique crashes, surpassing the SOTA's seven. Next, our method achieves a valid rate of 38.80%, significantly higher than the SOTA's 27.69%. Last, EvAFuzz achieves API coverage rates of 99.49% on PyTorch and 85.76% on TensorFlow, outperforming the SOTA's rates of 86.44% on PyTorch and 69.63% on TensorFlow. These results indicate that our method generates programs with higher rarity, validity, and variety, respectively.", "keywords": [ "Large Language Model", "Code Generation", "Fuzz", "Software Engineering", "Optimize", "Search" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 5, 3, 5 ] }, { "paper_id": "Jw63fvX3QB", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Andrew Y. Ng", "gender": "Male", "institution": "Computer Science Department", "country": "", "position": "Researcher" }, { "name": "Jeremy Andrew Irvin", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Ji Hun Wang", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "", "position": "MS student" }, { "name": "Jonathan H Chen", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Muhammad Ahmed Chaudhry", "gender": "Male", "institution": "PayPal Inc.", "country": "US", "position": "Researcher" }, { "name": "Yixing Jiang", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 16, "n_ref": 24, "n_ref_all": 39, "n_fig": 7, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 99274, "len_all_1": 48043, "len_abs": 2074, "len_title": 109, "len_sents": 28664, "len_sents_1": 24794, "n_sents": 196, "n_sents_1": 155, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 2085, "title": "Many-Shot In-Context Learning in Multimodal Foundation Models", "abs": "Large language models are well-known to be effective at few-shot in-context learning (ICL). Recent advancements in multimodal foundation models have enabled unprecedentedly long context windows, presenting an opportunity to explore their capability to perform ICL with many more demonstrating examples. In this work, we evaluate the performance of multimodal foundation models scaling from few-shot to many-shot ICL. We benchmark GPT-4o and Gemini 1.5 Pro across 14 datasets spanning multiple domains (natural imagery, medical imagery, remote sensing, and molecular imagery) and tasks (image classification, visual question answering, and object localization). We observe that many-shot ICL, including up to almost 2,000 multimodal demonstrating examples, leads to substantial improvements compared to few-shot (<100 examples) ICL across all of the datasets. Further, Gemini 1.5 Pro performance continues to improve log-linearly up to the maximum number of tested examples on many datasets. \nWe also find open-weights multimodal foundation models like Llama 3.2-Vision and InternLM-XComposer2.5 do not benefit from the demonstrating examples, highlighting an important gap between open and closed multimodal foundation models.\nGiven the high inference costs associated with the long prompts required for many-shot ICL, we also explore the impact of batching multiple queries in a single API call. We show that batching up to 50 queries can lead to performance improvements under zero-shot and many\u2013shot ICL, with substantial gains in the zero-shot setting on multiple datasets, while drastically reducing per-query cost and latency. Finally, we measure ICL data efficiency of the models, or the rate at which the models learn from more demonstrating examples. We find that while GPT-4o and Gemini 1.5 Pro achieve similar zero-shot performance across the datasets, Gemini 1.5 Pro exhibits higher ICL data efficiency than GPT-4o on most datasets. Our results suggest that many-shot ICL could enable users to efficiently adapt multimodal foundation models to new applications and domains.", "keywords": [ "in-context learning", "multimodal", "many-shot", "foundation models" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "JvkuZZ04O7", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Mufei Li", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Pan Li", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Siqi Miao", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 63, "n_ref": 159, "n_ref_all": 180, "n_fig": 5, "n_tab": 8, "L_tab": 6439, "n_element_tab": 980, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2587, "n_element_tab_1": 461, "formula_len_all": 333, "formula_len_all_1": 301, "len_all": 305582, "len_all_1": 82229, "len_abs": 1629, "len_title": 160, "len_sents": 62843, "len_sents_1": 36418, "n_sents": 539, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 122, "L_abs": 1608, "title": "Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation", "abs": "Large Language Models (LLMs) demonstrate strong reasoning abilities but face limitations such as hallucinations and outdated knowledge. Knowledge Graph (KG)-based Retrieval-Augmented Generation (RAG) addresses these issues by grounding LLM outputs in structured external knowledge from KGs. However, current KG-based RAG frameworks still struggle to optimize the trade-off between retrieval effectiveness and efficiency in identifying a suitable amount of relevant graph information for the LLM to digest. We introduce SubgraphRAG, extending the KG-based RAG framework that retrieves subgraphs and leverages LLMs for reasoning and answer prediction. Our approach innovatively integrates a lightweight multilayer perceptron (MLP) with a parallel triple-scoring mechanism for efficient and flexible subgraph retrieval while encoding directional structural distances to enhance retrieval effectiveness. The size of retrieved subgraphs can be flexibly adjusted to match the query's needs and the downstream LLM's capabilities. This design strikes a balance between model complexity and reasoning power, enabling scalable and generalizable retrieval processes. Notably, based on our retrieved subgraphs, smaller LLMs like Llama3.1-8B-Instruct deliver competitive results with explainable reasoning, while larger models like GPT-4o achieve state-of-the-art accuracy compared with previous baselines\u2014all without fine-tuning. Extensive evaluations on the WebQSP and CWQ benchmarks highlight SubgraphRAG's strengths in efficiency, accuracy, and reliability by reducing hallucinations and improving response grounding.", "keywords": [ "Knowledge Graphs", "Large Language Models", "Retrieval-Augmented Generation", "Retrieval" ], "rating_list": [ 8, 5, 6, 5, 6 ], "soundness_list": [ 4, 2, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 5, 5, 4, 5, 4 ] }, { "paper_id": "JvH4jDDcG3", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Hui LIU", "gender": "Female", "institution": "Saint Francis University", "country": "HK", "position": "Assistant Professor" }, { "name": "Jianhong Cheng", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "Junhui Hou", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Yuheng Jia", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 28, "n_formula_1": 11, "n_ref_uni": 43, "n_ref": 98, "n_ref_all": 124, "n_fig": 13, "n_tab": 8, "L_tab": 5203, "n_element_tab": 290, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 470, "n_element_tab_1": 21, "formula_len_all": 3257, "formula_len_all_1": 508, "len_all": 175205, "len_all_1": 60830, "len_abs": 1435, "len_title": 90, "len_sents": 54886, "len_sents_1": 29026, "n_sents": 429, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 1447, "title": "Towards Calibrated Deep Clustering Network", "abs": "Deep clustering has exhibited remarkable performance; however, the over confidence problem, i.e., the estimated confidence for a sample belonging to a particular cluster greatly exceeds its actual prediction accuracy, has been over looked in prior research. To tackle this critical issue, we pioneer the development of a calibrated deep clustering framework. Specifically, we propose a novel dual\nhead (calibration head and clustering head) deep clustering model that can effectively calibrate the estimated confidence and the actual accuracy. The calibration head adjusts the overconfident predictions of the clustering head, generating prediction confidence that matches the model learning status. Then, the clustering head dynamically selects reliable high-confidence samples estimated by the calibration head for pseudo-label self-training. Additionally, we introduce an effective network initialization strategy that enhances both training speed and network robustness. The effectiveness of the proposed calibration approach and initialization strategy are both endorsed with solid theoretical guarantees. Extensive experiments demonstrate the proposed calibrated deep clustering model not only surpasses the state-of-the-art deep clustering methods by 5\u00d7 on average in terms of expected calibration error, but also significantly outperforms them in terms of clustering accuracy. The code is available at https://github.com/ChengJianH/CDC.", "keywords": [ "Deep clustering", "unsupervised calibration", "clustering" ], "rating_list": [ 5, 5, 6, 8, 8 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 1, 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "Ju7zj6tUm6", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Akansha Kalra", "gender": "Female", "institution": "University of Utah", "country": "US", "position": "PhD student" }, { "name": "Basavasagar Patil", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "MS student" }, { "name": "Daniel S. Brown", "gender": "Male", "institution": "University of Utah", "country": "", "position": "Assistant Professor" }, { "name": "Guanhong Tao", "gender": "unknown", "institution": "University of Utah", "country": "US", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 7, "n_ref_uni": 27, "n_ref": 50, "n_ref_all": 69, "n_fig": 10, "n_tab": 9, "L_tab": 3184, "n_element_tab": 633, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2120, "n_element_tab_1": 469, "formula_len_all": 1567, "formula_len_all_1": 597, "len_all": 137495, "len_all_1": 66459, "len_abs": 1470, "len_title": 133, "len_sents": 45696, "len_sents_1": 28608, "n_sents": 316, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1475, "title": "How vulnerable is my learned policy? Adversarial attacks on modern behavioral cloning policies", "abs": "Learning from Demonstration (LfD) algorithms have shown promising results in robotic manipulation tasks, but their vulnerability to adversarial attacks remains underexplored. This paper presents a comprehensive study of adversarial attacks on both classic and recently proposed algorithms, including Behavior Cloning (BC), LSTM-GMM, Implicit Behavior Cloning (IBC), Diffusion Policy (DP), and VQ-Behavior Transformer (VQ-BET). We study the vulnerability of these methods to untargeted, targeted and universal adversarial perturbations. While explicit policies, such as BC, LSTM-GMM and VQ-BET can be attacked in the same manner as standard computer vision models, we find that attacks for implicit and denoising policy models are nuanced and require developing novel attack methods.\nOur experiments on several simulated robotic manipulation tasks reveal that most of the current methods are highly vulnerable to adversarial perturbations. We also investigate the transferability of attacks across algorithms, architectures, and tasks and provide insights into the generalizability of adversarial perturbations in LfD. We find that the success rate of the transfer attacks is highly dependent on the task, raising necessity for more fine-grained metrics that capture both the task difficulties and baseline performance of the algorithms. In summary, our findings highlight the vulnerabilities of modern BC algorithms, paving way for future work in addressing such limitations.", "keywords": [ "Adversarial Attacks", "Learning from Demonstrations", "Behavior Cloning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "JtGPIZpOrz", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Antonio Torralba", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Joshua B. Tenenbaum", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Professor" }, { "name": "Shuang Li", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Vighnesh Subramaniam", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yilun Du", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Igor Mordatch", "gender": "Male", "institution": "University of Washington", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 6, "n_ref_uni": 41, "n_ref": 65, "n_ref_all": 97, "n_fig": 13, "n_tab": 5, "L_tab": 2440, "n_element_tab": 147, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 2496, "n_element_tab_1": 122, "formula_len_all": 1623, "formula_len_all_1": 917, "len_all": 179006, "len_all_1": 65037, "len_abs": 1261, "len_title": 113, "len_sents": 54715, "len_sents_1": 27741, "n_sents": 438, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1228, "title": "Multiagent Finetuning of Language Models", "abs": "Large language models (LLMs) have achieved remarkable performance in recent years but are fundamentally limited by the underlying training data. To improve models beyond the training data, recent works have explored how LLMs can be used to generate synthetic data for autonomous self-improvement. However, successive steps of self-improvement can reach a point of diminishing returns, as the diversity of generations decreases, limiting further performance gains. In this work, we propose a complementary approach towards self-improvement where finetuning is applied to a multiagent society of language models. A set of language models are initialized from the same base model and then are specialized by independently updating each model using data generated by the model under multiagent interaction with other models. By training each model on independent sets of data, we illustrate how this approach enables specialization across models and diversification over the set of models. As a result, our overall system is able to autonomously improve over many more rounds of fine-tuning than single-agent self-improvement methods. We quantitatively illustrate the efficacy of the approach across a wide suite of reasoning tasks.", "keywords": [ "Language Models", "Multi-agent Interaction", "Self Improvement" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "Jt1gGIumJo", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Jeongwoo Park", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "Assistant Professor" }, { "name": "Kyuseok Nam", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "MS student" }, { "name": "Yulhwa Kim", "gender": "unknown", "institution": "Sungkyunkwan University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 35, "n_ref": 45, "n_ref_all": 56, "n_fig": 5, "n_tab": 3, "L_tab": 740, "n_element_tab": 101, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 775, "n_element_tab_1": 108, "formula_len_all": 619, "formula_len_all_1": 687, "len_all": 92230, "len_all_1": 48526, "len_abs": 958, "len_title": 138, "len_sents": 24914, "len_sents_1": 22137, "n_sents": 177, "n_sents_1": 157, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 971, "title": "Highlight Diffusion: Training-Free Attention Guided Acceleration for Text-to-Image Models", "abs": "Diffusion models have achieved exceptional results in image synthesis, yet their sequential processing nature imposes significant computational demands and latency, posing challenges for practical deployment. In this paper, we present Highlight Diffusion: a training-free novel acceleration approach that achieves significant speedup while retaining generation quality through an attention-guided generation process. By utilizing cross-attention maps to identify crucial segments within the image, we selectively compute these highlighted regions during the denoising process, bypassing the need for full-resolution computation at every step. This strategy maintains high-quality outputs while enabling faster, more resource-efficient diffusion model inference. With minimal loss in generated image quality\u2014evidenced by only a 0.65 increase in FID score and a 0.02 decrease in CLIP score, Highlight Diffusion achieved a 1.52 $\\times$ speedup using an NVIDIA RTX 3090 GPU.", "keywords": [ "Diffusion model", "Cross attention", "Acceleration", "Text to Image" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 1, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "Jszf4et48m", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Eslam Mohamed BAKR", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Liangbing Zhao", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "MS student" }, { "name": "Matthieu Cord", "gender": "Male", "institution": "Sorbonne Universit\u00e9", "country": "FR", "position": "Full Professor" }, { "name": "Mohamed Elhoseiny", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Vincent Tao Hu", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Postdoc" }, { "name": "Patrick Perez", "gender": "Male", "institution": "Kyutai", "country": "FR", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 10, "n_ref_uni": 53, "n_ref": 108, "n_ref_all": 185, "n_fig": 29, "n_tab": 10, "L_tab": 1581, "n_element_tab": 143, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 401, "n_element_tab_1": 51, "formula_len_all": 1536, "formula_len_all_1": 763, "len_all": 212299, "len_all_1": 71319, "len_abs": 1672, "len_title": 121, "len_sents": 66121, "len_sents_1": 32737, "n_sents": 526, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1750, "title": "ToddlerDiffusion: Interactive Structured Image Generation with Cascaded Schr\u00f6dinger Bridge", "abs": "Diffusion models break down the challenging task of generating data from high-dimensional distributions into a series of easier denoising steps. Inspired by this paradigm, we propose a novel approach that extends the diffusion framework into modality space, decomposing the complex task of RGB image generation into simpler, interpretable stages. Our method, termed {\\papernameAbbrev}, cascades modality-specific models, each responsible for generating an intermediate representation, such as contours, palettes, and detailed textures, ultimately culminating in a high-quality RGB image.\nInstead of relying on the naive LDM concatenation conditioning mechanism to connect the different stages together, we employ Schr\\\"odinger Bridge to determine the optimal transport between different modalities.\nAlthough employing a cascaded pipeline introduces more stages, which could lead to a more complex architecture, each stage is meticulously formulated for efficiency and accuracy, surpassing Stable-Diffusion (LDM) performance.\nModality composition not only enhances overall performance but enables emerging proprieties such as consistent editing, interaction capabilities, high-level interpretability, and faster convergence and sampling rate. \nExtensive experiments on diverse datasets, including LSUN-Churches, ImageNet, CelebHQ, and LAION-Art, demonstrate the efficacy of our approach, consistently outperforming state-of-the-art methods.\nFor instance, {\\papernameAbbrev} achieves notable efficiency, matching LDM performance on LSUN-Churches while operating 2$\\times$ faster with a 3$\\times$ smaller architecture.\nThe project website is available at:\n\\href{https://toddlerdiffusion.github.io/website/}{$https://toddlerdiffusion.github.io/website/$}", "keywords": [ "Generative AI", "Diffusion models", "Cascaded models", "Interpretability", "Image Editing" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "JslyktsKMY", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Ashok Cutkosky", "gender": "unknown", "institution": "Boston University", "country": "", "position": "Assistant Professor" }, { "name": "Hoang Tran", "gender": "Male", "institution": "Boston University", "country": "US", "position": "PhD student" }, { "name": "Qinzi Zhang", "gender": "Male", "institution": "Boston University, Boston University", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 59, "n_ref": 95, "n_ref_all": 115, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1506, "formula_len_all_1": 938, "len_all": 144335, "len_all_1": 67578, "len_abs": 1219, "len_title": 123, "len_sents": 36336, "len_sents_1": 32273, "n_sents": 273, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1219, "title": "Reevaluating Theoretical Analysis Methods for Optimization in Deep Learning", "abs": "There is a significant gap between our theoretical understanding of optimization algorithms used in deep learning and their practical performance. Theoretical development usually focuses on proving convergence guarantees under a variety of different assumptions, which are themselves often chosen based on a rough combination of intuitive match to practice and analytical convenience. In this paper, we carefully measure the degree to which the standard optimization analyses are capable of explaining modern algorithms. To do this, we develop new empirical metrics that compare real optimization behavior with analytically predicted behavior. Our investigation is notable for its tight integration with modern optimization analysis: rather than simply checking high-level assumptions made in the analysis (e.g. smoothness), we verify key low-level identities used by the analysis to explain optimization behavior that might hold even if the high-level motivating assumptions do not. In general, we find that real optimizers often make progress even when typical optimization analysis suggests that they should not. This highlights a need for developing new theoretical frameworks that are better aligned with practice.", "keywords": [ "Deep Learning", "Optimization", "Smooth", "Convex", "Sharpness" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "JsVIGVntnQ", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Dongbin Zhao", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Jiajun Chai", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Sicheng Li", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yuanheng Zhu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Associate Professor" }, { "name": "Yuqian Fu", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 9, "n_ref_uni": 35, "n_ref": 123, "n_ref_all": 175, "n_fig": 13, "n_tab": 17, "L_tab": 6384, "n_element_tab": 494, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1864, "n_element_tab_1": 221, "formula_len_all": 1993, "formula_len_all_1": 571, "len_all": 289003, "len_all_1": 71671, "len_abs": 1814, "len_title": 80, "len_sents": 100098, "len_sents_1": 32498, "n_sents": 751, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 79, "L_abs": 1832, "title": "Empowering LLM Agents with Zero-Shot Optimal Decision-Making through Q-learning", "abs": "Large language models (LLMs) are trained on extensive text data to gain general comprehension capability. Current LLM agents leverage this ability to make zero- or few-shot decisions but fail in making optimal decisions, as LLMs inherently perform next-token prediction based on pre-trained probability distributions rather than maximizing expected future rewards. In contrast, agents trained via reinforcement learning (RL) could make optimal decisions but require extensive environmental data. In this work, we develop an algorithm that combines the zero-shot capabilities of LLMs with the optimal decision-making advantages of RL, referred to as the Model-based LLM Agent with Q-Learning (MLAQ). MLAQ employs Q-learning to derive optimal policies from transitions within memory. However, unlike RL agents that collect data from environmental interactions, MLAQ constructs an imagination space fully based on LLM to perform imaginary interactions for deriving zero-shot policies. Our proposed UCB variant generates imaginary data through interactions with the LLM-based world model, enabling a balance between exploration and exploitation while ensuring a sub-linear regret bound guaranteed by a theorem. Moreover, MLAQ employs a mixed-examination mechanism that utilizes environmental interactions and LLM-based self-examine to enhance the quality of imaginary data. We evaluate MLAQ in benchmarks that present significant challenges for existing LLM agents. Results show that MLAQ achieves a optimal rate of over 90% in tasks where other methods struggle to succeed. Additional experiments are conducted to reach the conclusion that introducing model-based RL into LLM agents shows significant potential for current LLMs to improve their optimal decision-making ability. Our interactive website is available at http://mlaq.site.", "keywords": [ "Large language models", "Agent", "Optimal decision-making" ], "rating_list": [ 3, 3, 8, 5, 8 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 2, 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 3, 4 ] }, { "paper_id": "JsM46OZix7", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Bolin Ding", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Senior Director" }, { "name": "Ce Ge", "gender": "Male", "institution": "Fudan", "country": "CN", "position": "Intern" }, { "name": "Daoyuan Chen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Staff" }, { "name": "Yaliang Li", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Staff Engineer" }, { "name": "Zhijian Ma", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Engineer" } ], "n_formula": 27, "n_formula_1": 15, "n_ref_uni": 43, "n_ref": 66, "n_ref_all": 81, "n_fig": 5, "n_tab": 8, "L_tab": 2348, "n_element_tab": 246, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 681, "n_element_tab_1": 99, "formula_len_all": 1340, "formula_len_all_1": 790, "len_all": 158888, "len_all_1": 60675, "len_abs": 1156, "len_title": 112, "len_sents": 38355, "len_sents_1": 28104, "n_sents": 258, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1182, "title": "BiMix: Bivariate Data Mixing Law for Language Model Pretraining", "abs": "Large language models have demonstrated remarkable capabilities across various tasks, primarily attributed to the utilization of diversely sourced data. However, the impact of pretraining data composition on model performance remains poorly understood. This paper introduces BiMix, a novel bivariate data mixing law that models the joint scaling behavior of domain proportions and data volume in LLM pretraining. BiMix provides a systematic framework for understanding and optimizing data mixtures across diverse domains. Through extensive experiments on two large-scale datasets, we demonstrate BiMix's high accuracy in loss extrapolation (mean relative error $< 0.2\\%$) and its generalization to unseen mixtures (R$^{2} > 0.97$). Optimization of domain proportions yields superior model performance compared to existing methods. Furthermore, we establish entropy-based measures as efficient proxies for data mixing, offering a computationally lightweight strategy. Our work contributes both theoretical insights into data mixing dynamics and practical tools for enhancing LLM training efficiency, paving the way for more effective scaling strategies in language model development.", "keywords": [ "Data Mixture", "Large Language Models", "Scaling Law" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 5 ] }, { "paper_id": "Jrt9iWalFy", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Kaito Ariu", "gender": "Male", "institution": "CyberAgent, Inc.", "country": "JP", "position": "Research Scientist" }, { "name": "Kenshi Abe", "gender": "Male", "institution": "University of Electro-Communications", "country": "", "position": "PhD student" }, { "name": "Mitsuki Sakamoto", "gender": "Male", "institution": "CyberAgent, Inc.", "country": "JP", "position": "Research Engineer" }, { "name": "Atsushi Iwasaki", "gender": "Male", "institution": "University of Electro-Communications", "country": "", "position": "Associate Professor" } ], "n_formula": 153, "n_formula_1": 46, "n_ref_uni": 54, "n_ref": 133, "n_ref_all": 191, "n_fig": 3, "n_tab": 5, "L_tab": 1289, "n_element_tab": 132, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 21171, "formula_len_all_1": 3279, "len_all": 183590, "len_all_1": 65591, "len_abs": 1091, "len_title": 120, "len_sents": 37537, "len_sents_1": 26319, "n_sents": 369, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1094, "title": "Boosting Perturbed Gradient Ascent for Last-Iterate Convergence in Games", "abs": "This paper presents a payoff perturbation technique, introducing a strong convexity to players' payoff functions in games. This technique is specifically designed for first-order methods to achieve last-iterate convergence in games where the gradient of the payoff functions is monotone in the strategy profile space, potentially containing additive noise. Although perturbation is known to facilitate the convergence of learning algorithms, the magnitude of perturbation requires careful adjustment to ensure last-iterate convergence. Previous studies have proposed a scheme in which the magnitude is determined by the distance from a periodically re-initialized anchoring or reference strategy. Building upon this, we propose Gradient Ascent with Boosting Payoff Perturbation, which incorporates a novel perturbation into the underlying payoff function, maintaining the periodically re-initializing anchoring strategy scheme. This innovation empowers us to provide faster last-iterate convergence rates against the existing payoff perturbed algorithms, even in the presence of additive noise.", "keywords": [ "Last-Iterate Convergence", "Learning in Games", "Noisy Feedback" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "JrpMlotoGX", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Farima Fatahi Bayat", "gender": "Female", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Lechen Zhang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "MS student" }, { "name": "Lu Wang", "gender": "Female", "institution": "Northeastern University", "country": "", "position": "Assistant Professor" }, { "name": "Sheza Munir", "gender": "Female", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 16, "n_ref": 46, "n_ref_all": 68, "n_fig": 6, "n_tab": 5, "L_tab": 1597, "n_element_tab": 194, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1901, "n_element_tab_1": 152, "formula_len_all": 114, "formula_len_all_1": 27, "len_all": 165266, "len_all_1": 70832, "len_abs": 1466, "len_title": 131, "len_sents": 57788, "len_sents_1": 32549, "n_sents": 467, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1474, "title": "FactBench: A Dynamic Benchmark for In-the-Wild Language Model Factuality Evaluation", "abs": "Language models (LMs) are widely used by an increasing number of users, underscoring the challenge of maintaining factual accuracy across a broad range of topics. We present VERIFY (Verification and Evidence RetrIeval for FactualitY evaluation), a pipeline to evaluate LMs\u2019 factual accuracy in real-world user interactions. VERIFY considers the verifiability of LM-generated content and categorizes content units as supported, unsupported, or undecidable based on the retrieved web evidence. Importantly, VERIFY\u2019s factuality judgments correlate better with human evaluations than existing methods. Using VERIFY, we identify \u201challucination prompts\u201d across diverse topics\u2013those eliciting the highest rates of incorrect or unverifiable LM responses. These prompts form FACTBENCH, a dataset of 985 prompts across 213 fine-grained topics. Our dataset captures emerging factuality challenges in real-world LM interactions and is regularly updated with new prompts. We benchmark widely-used LMs from GPT, Gemini, and Llama3.1 family on FACTBENCH, yielding the following key findings: (i) Proprietary models exhibit better factuality, improving from Hard to Easy hallucination prompts. (ii) Llama3.1-405B-Instruct shows comparable or lower factual accuracy than Llama3.1-70B-Instruct across all evaluation methods due to its higher subjectivity that leads to more undecidable content. (iii) Gemini1.5-Pro shows a significantly higher refusal rate, with over-refusal in 25% of cases.", "keywords": [ "Factuality Evaluation Benchmark", "Factuality Evaluation Techniques", "LLM Evaluation" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "JrhAsf4xNH", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Brian Yi Sun", "gender": "Male", "institution": "Xinjiang University", "country": "", "position": "Full Professor" }, { "name": "Huaxiong Liu", "gender": "Male", "institution": "Xinjiang University", "country": "CN", "position": "MS student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Mingming Gong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Yinfeng Yu", "gender": "unknown", "institution": "Xinjiang University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 47, "n_formula_1": 18, "n_ref_uni": 20, "n_ref": 69, "n_ref_all": 85, "n_fig": 3, "n_tab": 7, "L_tab": 3676, "n_element_tab": 381, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 3894, "n_element_tab_1": 333, "formula_len_all": 2669, "formula_len_all_1": 689, "len_all": 133247, "len_all_1": 62664, "len_abs": 1533, "len_title": 161, "len_sents": 41614, "len_sents_1": 24981, "n_sents": 333, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 112, "L_abs": 1544, "title": "EM-DARTS: Preventing Performance Collapse in Differentiable Architecture Search with The Edge Mutation Mechanism", "abs": "Differentiable Architecture Search (DARTS) relaxes the discrete search space into a continuous form, significantly improving architecture search efficiency through gradient-based optimization. However, DARTS often suffers from performance collapse, where the performance of discovered architectures degrades during the search process, and the final architectures tend to be dominated by excessive skip-connections. In this work, we analyze how continuous relaxation impacts architecture optimization, identifying two main causes for performance collapse. First, the continuous relaxation framework introduces coupling between parametric operation weights and architecture parameters. This coupling leads to insufficient training of parametric operations, resulting in smaller architecture parameters for these operations. Second, DARTS's unrolled estimation property leads to larger architecture parameters for skip-connections. To attack this issue, we propose Edge Mutation Differentiable Architecture Search (EM-DARTS), where during network weight updates, edges have a probability of mutating from a weighted sum of candidate operations to a specific parametric operation.\n EM-DARTS reduces the impact of architecture parameters on parametric operations, allowing for better training of the parametric operations, thereby increasing their architecture parameters and preventing performance collapse. Theoretical results and experimental studies across diverse search spaces and datasets validate the effectiveness of the proposed method.", "keywords": [ "deep learning", "autoML", "neural architecture search", "image classification", "performance collapse" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "JrfWj5Ae1j", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Emily Black", "gender": "Female", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Nikita Rajaneesh", "gender": "Female", "institution": "Columbia University ", "country": "US", "position": "MS student" }, { "name": "Richard Zemel", "gender": "Male", "institution": "Department of Computer Science, Columbia University", "country": "", "position": "Full Professor" }, { "name": "Talia B. Gillis", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Associate Professor" }, { "name": "Thomas P Zollo", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 53, "n_ref": 92, "n_ref_all": 136, "n_fig": 8, "n_tab": 1, "L_tab": 52, "n_element_tab": 3, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 151, "formula_len_all_1": 0, "len_all": 245300, "len_all_1": 71389, "len_abs": 994, "len_title": 106, "len_sents": 66405, "len_sents_1": 41187, "n_sents": 361, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 997, "title": "Towards Effective Discrimination Testing for Generative AI", "abs": "Generative AI (GenAI) models present new challenges in testing for, and regulating against, discriminatory behavior. In this paper, we argue that GenAI fairness research still has not met these challenges: there is a dearth of reliable bias assessment methods for GenAI systems that speak to regulatory goals. This leads to ineffective regulation that can allow deployment of reportedly fair, yet actually discriminatory GenAI systems. Towards remedying this problem, we connect the legal and technical literature around GenAI bias evaluation and identify areas of misalignment. Through four case studies, we demonstrate how this misalignment between fairness testing techniques and regulatory goals can result in discriminatory outcomes in real-world deployments, especially in adaptive or complex environments. We offer practical recommendations for improving discrimination testing to better align with regulatory goals and enhance the reliability of fairness assessments in future deployments.", "keywords": [ "Generative AI", "LLMs", "fairness", "discrimination", "policy", "diffusion" ], "rating_list": [ 8, 5, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 4, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "Jrb9yXZJKG", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Jiang Bian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Qi Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Shizhao Sun", "gender": "Female", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "\u96e8 \u8881", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 21, "n_ref": 87, "n_ref_all": 115, "n_fig": 14, "n_tab": 2, "L_tab": 456, "n_element_tab": 78, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 193, "n_element_tab_1": 2, "formula_len_all": 108, "formula_len_all_1": 135, "len_all": 113696, "len_all_1": 58575, "len_abs": 639, "len_title": 142, "len_sents": 35555, "len_sents_1": 29271, "n_sents": 291, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1450, "title": "CAD-Editor: Text-based CAD Editing through Adapting Large Language Models with Synthetic Data", "abs": "Computer Aided Design (CAD) is indispensable across various industries. \n\\emph{Text-based CAD editing}, which automatically modifies CAD models following textual instructions, is important yet not extensively studied. \nExisting work explores design variation generation, which randomly alters specific parts of a CAD model, offering no control over the final appearance.\nThis work introduces \\emph{CAD-Editor} for text-based editing.\nWe leverage Large Language Models (LLMs) as the backbone to take the concatenation of textual instruction and original CAD sequence as input and predict the edited CAD sequence, where the sequence representation of a CAD model is designed for easier processing by LLMs.\nMoreover, we propose fine-tuning LLMs by using a synthetic dataset followed by a selective dataset.\nThe synthetic data is produced by leveraging powerful existing models, including design variation generation models for producing paired CAD models and multi-modal models for capturing textual differences between these pairs.\nThe selective data is created by choosing top examples from outputs of the initially fine-tuned LLMs based on human feedback or metrics.\nIn this way, a large-scale synthetic dataset offers basic capability while a selective dataset that is less noisy and better aligned with human intentions boosts performance further.\nExtensive experiments demonstrate the advantage of CAD-Editor both quantitatively and qualitatively.", "keywords": [ "Computer Aided Design", "Generative Models", "Text-based Editing", "Large Language Models" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "JqKh7FLUw1", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" }, { "name": "Jerry Yao-Chieh Hu", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Jiahao Yu", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Wenbo Guo", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "Assistant Professor" }, { "name": "Xinyu Xing", "gender": "Male", "institution": "Pennsylvania State University", "country": "", "position": "Assistant Professor" }, { "name": "Haozheng LUO", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "US", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 1, "n_ref_uni": 32, "n_ref": 107, "n_ref_all": 159, "n_fig": 21, "n_tab": 8, "L_tab": 1785, "n_element_tab": 312, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 667, "n_element_tab_1": 184, "formula_len_all": 573, "formula_len_all_1": 31, "len_all": 205303, "len_all_1": 71202, "len_abs": 1812, "len_title": 119, "len_sents": 66905, "len_sents_1": 32874, "n_sents": 570, "n_sents_1": 272, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1472, "title": "BOOST: Enhanced Jailbreak of Large Language Model via Slient eos Tokens", "abs": "Along with the remarkable successes of Language language models, recent research also started to explore the security threats of LLMs, including jailbreaking attacks. Attackers carefully craft jailbreaking prompts such that a target LLM will respond to the harmful question. Existing jailbreaking attacks require either human experts or leveraging complicated algorithms to craft jailbreaking prompts. In this paper, we introduce BOOST, a simple attack that leverages only the eos tokens. We demonstrate that rather than constructing complicated jailbreaking prompts, the attacker can simply append a few eos tokens to the end of a harmful question. It will bypass the safety alignment of LLMs and lead to successful jailbreaking attacks. We further apply BOOST to four representative jailbreak methods and show that the attack success rates of these methods can be significantly enhanced by simply adding eos tokens to the prompt. To understand this simple but novel phenomenon, we conduct both theoretical and empirical analyses. Our analysis reveals that (1) adding eos tokens makes the target LLM believe the input is much less harmful, and (2) eos tokens have low attention values and do not affect LLM's understanding of the harmful questions, leading the model to actually respond to the questions. Our findings uncover how fragile an LLM is against jailbreak attacks, motivating the development of strong safety alignment approaches.large language model, Jailbreak", "keywords": [ "large language model", "Jailbreak" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "Jq8NPYVxLW", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Chongjie Zhang", "gender": "unknown", "institution": "Washington University, Saint Louis", "country": "", "position": "Associate Professor" }, { "name": "Han Liu", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "US", "position": "PhD student" }, { "name": "Ning Zhang", "gender": "unknown", "institution": "Washington University, Saint Louis", "country": "", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 7, "n_ref_uni": 57, "n_ref": 147, "n_ref_all": 173, "n_fig": 7, "n_tab": 13, "L_tab": 4832, "n_element_tab": 524, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2102, "n_element_tab_1": 315, "formula_len_all": 204, "formula_len_all_1": 254, "len_all": 178052, "len_all_1": 76302, "len_abs": 959, "len_title": 130, "len_sents": 52537, "len_sents_1": 34373, "n_sents": 388, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 962, "title": "Less is More: Exploiting Feature Density for Enhanced Membership Inference Attacks", "abs": "Membership inference attacks have become the de facto standard for assessing privacy breaches across various machine learning (ML) models. However, existing approaches often require substantial resources, including large numbers of shadow models and auxiliary datasets, to achieve high true positive rates (TPR) in the low false positive rate (FPR) region. This makes these attacks prohibitively expensive and less practical. In this work, we propose a novel membership inference attack that exploits feature density gaps by progressively removing features from both members and non-members and evaluating the corresponding model outputs as a new membership signal. Our method requires only a few dozen queries and does not rely on large auxiliary datasets or the training of numerous shadow models. Extensive evaluations on both classification and diffusion models demonstrate that our method significantly improves the TPR at low FPR across multiple scenarios.", "keywords": [ "membership inference attack", "machine learning privacy" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "Jq8HYNZG9s", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Aleksandr Simonyan", "gender": "Male", "institution": "Visual Concepts", "country": "US", "position": "Senior Machine Learning Engineer" }, { "name": "Nikolay Falaleev", "gender": "Male", "institution": "Sportlight Technology", "country": "GB", "position": "Principal Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 30, "n_ref": 34, "n_ref_all": 40, "n_fig": 3, "n_tab": 9, "L_tab": 1252, "n_element_tab": 188, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 994, "n_element_tab_1": 134, "formula_len_all": 125, "formula_len_all_1": 125, "len_all": 104359, "len_all_1": 58512, "len_abs": 858, "len_title": 93, "len_sents": 34293, "len_sents_1": 28257, "n_sents": 228, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 860, "title": "ShadowPunch: fast actions spotting benchmark", "abs": "We introduce an open dataset for video event spotting focused on fast-paced events in shadowboxing videos captured at high frame rates. The dataset features accurate frame-level annotations for diverse punch types alongside pose keypoint annotations, enabling the development of robust event recognition models. This work presents a novel benchmark exploring two distinct approaches to event spotting: direct prediction from image data and a staged approach involving intermediate pose estimation followed by event detection based on the detected keypoints. We provide baseline neural network solutions incorporating temporal information for both tracks, facilitating comparative analysis of these methodologies. This shadowboxing dataset advances the field of automatic sports analysis and contributes to the broader understanding of video events recognition.", "keywords": [ "Action Spotting", "Pose Estimation", "Dataset", "Boxing", "Sports", "Video Understanding" ], "rating_list": [ 5, 3, 3, 1 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "Jprs1v2wPA", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Anna Vettoruzzo", "gender": "Female", "institution": "Halmstad University", "country": "SE", "position": "PhD student" }, { "name": "Joaquin Vanschoren", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Associate Professor" }, { "name": "Lorenzo Braccaioli", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "PhD student" }, { "name": "Marlena Nowaczyk", "gender": "Female", "institution": "Halmstad University College", "country": "SE", "position": "Lecturer" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 53, "n_ref": 139, "n_ref_all": 181, "n_fig": 10, "n_tab": 10, "L_tab": 5982, "n_element_tab": 566, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 959, "n_element_tab_1": 119, "formula_len_all": 929, "formula_len_all_1": 630, "len_all": 217532, "len_all_1": 70145, "len_abs": 1795, "len_title": 98, "len_sents": 68822, "len_sents_1": 33190, "n_sents": 450, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1113, "title": "Unsupervised Meta-Learning via In-Context Learning", "abs": "Unsupervised meta-learning aims to learn feature representations from unsupervised datasets that can transfer to downstream tasks with limited labeled data.\nIn this paper, we propose a novel approach to unsupervised meta-learning that leverages the generalization abilities of in-context learning observed in transformer architectures. Our method reframes meta-learning as a sequence modeling problem, enabling the transformer encoder to learn task context from support images and utilize it to predict query images. \nAt the core of our approach lies the creation of diverse tasks generated using a combination of data augmentations and a mixing strategy that challenges the model during training while fostering generalization to unseen tasks at test time. \nExperimental results on benchmark datasets showcase the superiority of our approach over existing unsupervised meta-learning baselines, establishing it as the new state-of-the-art. Remarkably, our method achieves competitive results with supervised and self-supervised approaches, underscoring its efficacy in leveraging generalization over memorization.", "keywords": [ "meta-learning", "unsupervised learning", "in-context learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 5, 2, 5 ] }, { "paper_id": "JnWJbrnaUE", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Jia-Chen Gu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Shi-Qi Yan", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Yun Zhu", "gender": "Male", "institution": "Google", "country": "", "position": "software engineer" }, { "name": "Zhen-Hua Ling", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 26, "n_ref": 85, "n_ref_all": 107, "n_fig": 3, "n_tab": 9, "L_tab": 1790, "n_element_tab": 328, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2114, "n_element_tab_1": 221, "formula_len_all": 170, "formula_len_all_1": 173, "len_all": 166681, "len_all_1": 65246, "len_abs": 1314, "len_title": 89, "len_sents": 44315, "len_sents_1": 29520, "n_sents": 315, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 41, "L_abs": 1321, "title": "Corrective Retrieval Augmented Generation", "abs": "Large language models (LLMs) inevitably exhibit hallucinations since the accuracy of generated texts cannot be secured solely by the parametric knowledge they encapsulate. Although retrieval-augmented generation (RAG) is a practicable complement to LLMs, it relies heavily on the relevance of retrieved documents, raising concerns about how the model behaves if retrieval goes wrong. To this end, we propose the Corrective Retrieval Augmented Generation (CRAG) to improve the robustness of generation. Specifically, a lightweight retrieval evaluator is designed to assess the overall quality of retrieved documents for a query, returning a confidence degree based on which different knowledge retrieval actions can be triggered. Since retrieval from static and limited corpora can only return sub-optimal documents, large-scale web searches are utilized as an extension for augmenting the retrieval results. Besides, a decompose-then-recompose algorithm is designed for retrieved documents to selectively focus on key information and filter out irrelevant information in them. CRAG is plug-and-play and can be seamlessly coupled with various RAG-based approaches. Experiments on four datasets covering short- and long-form generation tasks show that CRAG can significantly improve the performance of RAG-based approaches.", "keywords": [ "Retrieval-Augmented Generation", "Corrective Retrieval", "Robustness of Generation" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "JnRvQ8CxLx", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Bo Wang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xingjun Ma", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Yu-Gang Jiang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "RUOFAN WANG", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 21, "n_ref": 56, "n_ref_all": 67, "n_fig": 5, "n_tab": 4, "L_tab": 521, "n_element_tab": 70, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 458, "n_element_tab_1": 49, "formula_len_all": 215, "formula_len_all_1": 159, "len_all": 102241, "len_all_1": 51764, "len_abs": 1566, "len_title": 85, "len_sents": 33136, "len_sents_1": 24494, "n_sents": 240, "n_sents_1": 169, "venue": "ICLR2025", "decision": "withdraw", "L_title": 37, "L_abs": 1576, "title": "IDEATOR: Jailbreaking VLMs Using VLMs", "abs": "As large Vision-Language Models (VLMs) continue to gain prominence, ensuring their safety deployment in real-world applications has become a critical concern. Recently, significant research efforts have focused on evaluating the robustness of VLMs against jailbreak attacks. Due to challenges in obtaining multi-modal data, current studies often assess VLM robustness by generating adversarial or query-relevant images based on harmful text datasets. However, the jailbreak images generated this way exhibit certain limitations. Adversarial images require white-box access to the target VLM and are relatively easy to defend against, while query-relevant images must be linked to the target harmful content, limiting their diversity and effectiveness. In this paper, we propose a novel jailbreak method named IDEATOR, which autonomously generates malicious image-text pairs for black-box jailbreak attacks. IDEATOR is a VLM-based approach inspired by our conjecture that a VLM itself might be a powerful red team model for generating jailbreak prompts. Specifically, IDEATOR employs a VLM to generate jailbreak texts while leveraging a state-of-the-art diffusion model to create corresponding jailbreak images. Extensive experiments demonstrate the high effectiveness and transferability of IDEATOR. It successfully jailbreaks MiniGPT-4 with a 94% success rate and transfers seamlessly to LLVA and InstructBLIP, achieving high success rates of 82% and 88%, respectively. IDEATOR uncovers previously unrecognized vulnerabilities in VLMs, calling for advanced safety mechanisms.", "keywords": [ "Black-box Jailbreak", "Large Vision-Language Models" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 1, 3, 1, 3 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "JmXu4fk5Mm", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Jianmeng Liu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Yichen Liu", "gender": "Male", "institution": "Sensetime", "country": "HK", "position": "Researcher" }, { "name": "Yu-wing Tai", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "Associate Professor" }, { "name": "Zeyuan Meng", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Undergrad student" }, { "name": "ChiKeung Tang", "gender": "Not Specified", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Professor" }, { "name": "ZHANG YUYAO", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 24, "n_ref": 40, "n_ref_all": 55, "n_fig": 6, "n_tab": 4, "L_tab": 1250, "n_element_tab": 18, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 471, "n_element_tab_1": 72, "formula_len_all": 260, "formula_len_all_1": 145, "len_all": 158515, "len_all_1": 44941, "len_abs": 414, "len_title": 139, "len_sents": 32144, "len_sents_1": 20432, "n_sents": 232, "n_sents_1": 144, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1646, "title": "VP-LLM: Text-Driven 3D Volume Completion with Large Language Models through Patchification", "abs": "3D completion represents a critical task within the vision industries. Traditional diffusion-based methodologies have achieved commendable performance; however, they are hindered by several issues. Firstly, these methods primarily depend on models such as CLIP or BERT to encode textual information, thereby making them incapable of supporting detailed and complex instructions. Moreover, their model sizes usually increase rapidly when the scene is larger or the voxel resolution is higher, making it impossible to scale up. Witnessing the significant advancements in multi-modal understanding capabilities facilitated by recent developments in large language models (LLMs), we introduce Volume Patch LLM (VP-LLM), designed to execute *user-friendly* conditional 3D completion and denoising using a token-based single-forward pass approach. To integrate a 3D model into the textual domain of the LLM, the incomplete 3D model is initially divided into smaller patches—a process we refer to as \"patchification\"—in a way that each patch can be independently encoded, analogous to the tokenization configuration utilized by LLMs. These encoded patches are subsequently concatenated with the encoded text prompt sequence and inputted into an LLM, which is fine-tuned to capture the relationships between these patch tokens while embedding semantic meanings into the 3D object. Our findings indicate a robust ability of LLMs to interpret complex text instructions and comprehend 3D objects, surpassing the quality of results produced by state-of-the-art diffusion-based 3D completion models, especially when complex text prompts are given.", "keywords": [ "Large Language Models", "3D completion", "3D generation" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "JmGEZXkCH3", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Haijin Zeng", "gender": "Male", "institution": "IMEC & Universiteit Gent", "country": "BE", "position": "PhD student" }, { "name": "Huacheng Li", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Jingyong Su", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yongyong Chen", "gender": "Male", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 65, "n_ref_all": 80, "n_fig": 3, "n_tab": 5, "L_tab": 2735, "n_element_tab": 350, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 802, "n_element_tab_1": 78, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 119764, "len_all_1": 54179, "len_abs": 1690, "len_title": 140, "len_sents": 33443, "len_sents_1": 28806, "n_sents": 207, "n_sents_1": 147, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1697, "title": "Beyond Transformations: Augmenting Anything for Image Super-Resolution via Diffusion Model", "abs": "Image super-resolution (SR), aiming to restore accurate high-resolution images from low-resolution ones, plays a pivotal role in image processing. However, the performance of SR models is often hindered by conventional data augmentation and data degradation techniques. Conventional data augmentation methods for SR are typically limited to geometric transformations, lacking semantic richness. Traditional data degradation methods simulate degradation through a series of blurring, noise addition, compression, and resizing processes, lacking the complexity essential for robust model training. In this paper, based on pre-trained large-scale text-to-image diffusion models, we propose a novel data augmentation method and an innovative data degradation method in SR modeling. Our data augmentation method utilizes Stable Diffusion to modify image content at the semantic level for controlled data augmentation, enriching training datasets with nuanced variations while preserving the quality of the original images. Moreover, after fine-tuning Stable Diffusion with domain-matched data we further enhance the augmentation efficacy. Besides, by carefully designing control signals, our data degradation method utilizes diffusion to emulate degradation, simulating various unknown input corruptions to improve the performance of SR models across unfamiliar image degradation patterns. Our data augmentation method improves PSNR by 0.8 dB on the FFHQ dataset and by 0.28 dB on the Manga109 dataset for the SR tasks. Meanwhile, our data degradation technique has proven effective in significantly reducing artifacts in real-world SR imagery, distinctly exceeding the performance of traditional ones.", "keywords": [ "data augmentation", "data degradation", "super-resolution" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "JlkqReTftJ", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Damian Borth", "gender": "Male", "institution": "University of St.Gallen", "country": "CH", "position": "Full Professor" }, { "name": "Konstantin Sch\u00fcrholt", "gender": "Male", "institution": "Universit\u00e4t St. Gallen", "country": "CH", "position": "Postdoc" }, { "name": "L\u00e9o Meynent", "gender": "Male", "institution": "Universit\u00e4t St. Gallen", "country": "", "position": "PhD student" }, { "name": "Yaoqing Yang", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "Assistant Professor" }, { "name": "Yefan Zhou", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 42, "n_ref": 73, "n_ref_all": 90, "n_fig": 10, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 144, "formula_len_all_1": 144, "len_all": 139063, "len_all_1": 65295, "len_abs": 1681, "len_title": 112, "len_sents": 42336, "len_sents_1": 31819, "n_sents": 318, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1693, "title": "Model Zoos for Benchmarking Phase Transitions in Neural Networks", "abs": "Understanding the complex dynamics of neural network training remains a central challenge in deep learning research.\nWork rooted in statistical physics has identified phases and phase transitions in neural network (NN) models, where models within the same phase exhibit similar characteristics but qualitatively differ across phases. A prominent example is the double-descent phenomenon. \nRecognizing these transitions is essential for building a deeper understanding of model behavior and the underlying mechanics.\nSo far, these phases are typically studied in isolation or in specific applications. \nIn this paper, we show that phase transitions are a widespread phenomenon.\nHowever, identifying phase transitions across different methods requires populations that cover different phases.\nFor that reason, we introduce Phase Transition Model Zoos, a structured collection of neural networks trained on diverse datasets and architectures. These model zoos are carefully designed to help researchers systematically identify and study phase transitions in their methods. \nWe demonstrate the relevance of phase transitions across multiple applications, including fine-tuning, transfer learning, out-of-distribution generalization, pruning, ensembling, and weight averaging. The diversity of applications underscores the universal nature of phase transitions and their impact on different tasks.\nBy providing the first structured dataset specifically designed to capture phase transitions in NNs, we offer a valuable tool for the community to systematically evaluate machine learning methods and improve their understanding of phase behavior across a wide range of applications and architectures.", "keywords": [ "Phase Transition", "Model Zoo", "Population", "Evaluation" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "Jlhq0zb76Q", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Hao Chen", "gender": "unknown", "institution": "University of California, Davis", "country": "US", "position": "Full Professor" }, { "name": "Hongxiang Zhang", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "MS student" }, { "name": "Yifeng He", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 39, "n_ref": 100, "n_ref_all": 116, "n_fig": 8, "n_tab": 5, "L_tab": 1099, "n_element_tab": 195, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1579, "n_element_tab_1": 186, "formula_len_all": 168, "formula_len_all_1": 168, "len_all": 139788, "len_all_1": 64365, "len_abs": 1522, "len_title": 111, "len_sents": 45250, "len_sents_1": 29487, "n_sents": 306, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1463, "title": "SteerDiff: Steering towards Safe Text-To-Image Diffusion Models", "abs": "Text-to-image (T2I) diffusion models have drawn attention for their ability to generate high-quality images with precise text alignment. However, these models can also be misused to produce inappropriate content. Existing safety measures, which typically rely on text classifiers or ControlNet-like approaches, are often insufficient.\nTraditional text classifiers rely on large-scale labeled datasets and can be easily bypassed by rephrasing.\nAs diffusion models continue to scale, fine-tuning these safeguards becomes increasingly challenging and lacks flexibility. Recent red-teaming attack researches further underscore the need for a new paradigm to prevent the generation of inappropriate content.\nIn this paper, we introduce SteerDiff, a lightweight adaptor module designed to act as an intermediary between user input and the diffusion model, ensuring that generated images adhere to ethical and safety standards with little to no impact on usability. SteerDiff identifies and manipulates inappropriate concepts within the text embedding space to guide the model away from harmful outputs. We conduct extensive experiments across various concept unlearning tasks to evaluate the effectiveness of our approach. Furthermore, we benchmark SteerDiff against multiple red-teaming strategies to assess its robustness. Finally, we explore the potential of SteerDiff for concept forgetting tasks, demonstrating its versatility in text-conditioned image generation.", "keywords": [ "Diffusion Model", "Text-to-image", "concept unlearning" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 1 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "JlDx2xp01W", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Hengjia Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Jun Zhao", "gender": "Male", "institution": "Fabu Inc.", "country": "CN", "position": "Researcher" }, { "name": "Kaipeng Zhang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Wenqi Shao", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xiaofei He", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" }, { "name": "Yuqi Lin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zheng Yang", "gender": "Male", "institution": "Fabu Inc", "country": "CN", "position": "CTO" }, { "name": "Ping Luo", "gender": "unknown", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 4, "n_ref_uni": 22, "n_ref": 39, "n_ref_all": 80, "n_fig": 26, "n_tab": 11, "L_tab": 6324, "n_element_tab": 482, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1160, "n_element_tab_1": 114, "formula_len_all": 366, "formula_len_all_1": 180, "len_all": 183977, "len_all_1": 58357, "len_abs": 1968, "len_title": 119, "len_sents": 51579, "len_sents_1": 27184, "n_sents": 421, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1594, "title": "SAMRefiner: Taming Segment Anything Model for Universal Mask Refinement", "abs": "In this paper, we explore a principal way to enhance the quality of widely pre-existing coarse masks, enabling them to serve as reliable training data for segmentation models to reduce the annotation cost. In contrast to prior refinement techniques that are tailored to specific models or tasks in a close-world manner, we propose SAMRefiner, a universal and efficient approach by adapting SAM to the mask refinement task. The core technique of our model is the noise-tolerant prompting scheme. Specifically, we introduce a multi-prompt excavation strategy to mine diverse input prompts for SAM (\\ie, distance-guided points, context-aware elastic bounding boxes, and Gaussian-style masks) from initial coarse masks. These prompts can collaborate with each other to mitigate the effect of defects in coarse masks. In particular, considering the difficulty of SAM to handle the multi-object case in semantic segmentation, we introduce a split-then-merge (STM) pipeline. Additionally, we extend our method to SAMRefiner++ by introducing an additional IoU adaption step to further boost the performance of the generic SAMRefiner on the target dataset. This step is self-boosted and requires no additional annotation. The proposed framework is versatile and can flexibly cooperate with existing segmentation methods. We evaluate our mask framework on a wide range of benchmarks under different settings, demonstrating better accuracy and efficiency. SAMRefiner holds significant potential to expedite the evolution of refinement tools, and we will release it as a convenient post-processing toolkit.", "keywords": [ "Mask Refinement", "Segment Anything Model" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "Jl0aEFrp11", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Linglong Kong", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Full Professor" }, { "name": "Xingcai Zhou", "gender": "Male", "institution": "Nanjing Audit University", "country": "", "position": "Full Professor" }, { "name": "guangyang", "gender": "unknown", "institution": "Nanjing Audit University", "country": "", "position": "PhD student" } ], "n_formula": 249, "n_formula_1": 36, "n_ref_uni": 19, "n_ref": 70, "n_ref_all": 79, "n_fig": 4, "n_tab": 2, "L_tab": 4, "n_element_tab": 3, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 113, "n_element_tab_1": 9, "formula_len_all": 63452, "formula_len_all_1": 5515, "len_all": 253211, "len_all_1": 66628, "len_abs": 4896, "len_title": 124, "len_sents": 52563, "len_sents_1": 26118, "n_sents": 640, "n_sents_1": 236, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1122, "title": "Bidirectional Communication-Efficient Non-Convex Adaptive Federated Learning", "abs": "Within the framework of federated learning, we introduce two novel strategies: New Lazy Aggregation (NLA) and Accelerated Aggregation (AA). The NLA strategy reduces communication and computational costs through adaptive gradient skipping, while the AA strategy accelerates computation and decreases communication costs via adaptive gradient accumulation. Building upon these innovative strategies and compression techniques, we propose two new algorithms: FedBNLACA and FedBACA, aimed at minimizing bidirectional communication costs. We provide theoretical guarantees for client participation (either full or partial) in these algorithms under non-convex settings and heterogeneous data. In the context of non-convex optimization with full client participation, our proposed FedBNLACA and FedBACA algorithms achieve the same convergence rate of $\\mathcal{O}\\big(1/T\\big)$ as their non-tight counterparts. Extensive experimental results demonstrate that our protocols facilitate effective training in non-convex environments and exhibit robustness across a wide range of devices, partial participation, and imbalanced data.", "keywords": [ "Federated learning", "non-convex learning", "bidirectional communication-efficient", "adaptive" ], "rating_list": [ 1, 6, 3, 1 ], "soundness_list": [ 1, 4, 2, 2 ], "presentation_list": [ 1, 4, 1, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "JkLLAOcEME", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Haokui Zhang", "gender": "Male", "institution": "Northwest Polytechnical University", "country": "CN", "position": "Full Professor" }, { "name": "Ruihan Xu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Shiliang Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 49, "n_ref": 204, "n_ref_all": 220, "n_fig": 0, "n_tab": 10, "L_tab": 4884, "n_element_tab": 476, "n_fig_1": 0, "n_tab_1": 8, "L_tab_1": 4535, "n_element_tab_1": 464, "formula_len_all": 683, "formula_len_all_1": 645, "len_all": 175588, "len_all_1": 77291, "len_abs": 1096, "len_title": 123, "len_sents": 48444, "len_sents_1": 29980, "n_sents": 456, "n_sents_1": 258, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1102, "title": "NN-Former: Rethinking Graph Structure in Neural Architecture Representation", "abs": "The growing use of deep learning necessitates efficient network design and deployment, making neural predictors vital for estimating attributes such as accuracy and latency. Recently, Graph Neural Networks (GNNs) and transformers have shown promising performance in representing neural architectures. However, each method has its disadvantages. GNNs lack the capabilities to represent complicated features, while transformers face poor generalization when the depth of architecture grows. To mitigate the above problems, we rethink neural architecture topology and show that sibling nodes are pivotal while overlooked in previous research. Thus we propose a novel predictor leveraging the strengths of GNNs and transformers to learn the enhanced topology. We introduce a novel token mixer that considers siblings, and a new channel mixer named bidirectional graph isomorphism feed-forward network. Our approach consistently achieves promising performance in both accuracy and latency prediction, providing valuable insights for learning Directed Acyclic Graph (DAG) topology. The code will be released.", "keywords": [ "Neural Architecture Representation", "Neural Predictors" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "JkCJBoNUcU", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Jiaqi Xu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jingjing Ren", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Long Peng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Renjing Pei", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Engineer" }, { "name": "Wenbo Li", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yang Cao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yang Wang", "gender": "Male", "institution": "Chang'an University", "country": "CN", "position": "Associate Professor" }, { "name": "Zheng-Jun Zha", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 52, "n_ref": 114, "n_ref_all": 157, "n_fig": 11, "n_tab": 12, "L_tab": 4333, "n_element_tab": 266, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 2157, "n_element_tab_1": 91, "formula_len_all": 993, "formula_len_all_1": 1093, "len_all": 167825, "len_all_1": 51543, "len_abs": 1101, "len_title": 113, "len_sents": 42615, "len_sents_1": 23235, "n_sents": 279, "n_sents_1": 128, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1104, "title": "Towards Realistic Data Generation for Real-World Super-Resolution", "abs": "Existing image super-resolution (SR) techniques often fail to generalize effectively in complex real-world settings due to the significant divergence between training data and practical scenarios. To address this challenge, previous efforts have either manually simulated intricate physical-based degradations or utilized learning-based techniques, yet these approaches remain inadequate for producing large-scale, realistic, and diverse data simultaneously. In this paper, we introduce a novel Realistic Decoupled Data Generator (RealDGen), an unsupervised learning data generation framework designed for real-world super-resolution. We meticulously develop content and degradation extraction strategies, which are integrated into a novel content-degradation decoupled diffusion model to create realistic low-resolution images from unpaired real LR and HR images. Extensive experiments demonstrate that RealDGen excels in generating large-scale, high-quality paired data that mirrors real-world degradations, significantly advancing the performance of popular SR models on various real-world benchmarks.", "keywords": [ "Real-world Image Super-Resolution; Data Generation" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "Jjr2Odj8DJ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Ankur Taly", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Chun-Sung Ferng", "gender": "unknown", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Cyrus Rashtchian", "gender": "Male", "institution": "Google Research", "country": "US", "position": "Researcher" }, { "name": "Da-Cheng Juan", "gender": "unknown", "institution": "Google Research", "country": "", "position": "Senior Software Engineer" }, { "name": "Jianyi Zhang", "gender": "unknown", "institution": "National Artificial Intelligence Research Resource Pilot & Duke University", "country": "US", "position": "Principal investigator" }, { "name": "Hailey James", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 61, "n_ref_all": 76, "n_fig": 4, "n_tab": 5, "L_tab": 952, "n_element_tab": 85, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 747, "n_element_tab_1": 88, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 146550, "len_all_1": 68274, "len_abs": 1388, "len_title": 120, "len_sents": 41929, "len_sents_1": 32520, "n_sents": 378, "n_sents_1": 272, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1396, "title": "Sufficient Context: A New Lens on Retrieval Augmented Generation Systems", "abs": "Augmenting LLMs with context leads to improved performance across many applications. Despite much research on Retrieval Augmented Generation (RAG) systems, an open question is whether errors arise because LLMs fail to utilize the context from retrieval or the context itself is insufficient to answer the query. To shed light on this, we develop a new notion of sufficient context, along with a way to classify instances that have enough information to answer the query. We then use sufficient context to analyze several models and datasets. By stratifying errors based on context sufficiency, we find that proprietary LLMs (Gemini, GPT, Claude) excel at answering queries when the context is sufficient, but often output incorrect answers instead of abstaining when the context is not. On the other hand, open-source LLMs (Llama, Mistral, Gemma) hallucinate or abstain often, even with sufficient context. We further categorize cases when the context is useful, and improves accuracy, even though it does not fully answer the query and the model errs without the context. Building on our findings, we explore ways to reduce hallucinations in RAG systems, including a new selective generation method that leverages sufficient context information for guided abstention. Our method improves the fraction of correct answers among times where the model responds by 2--10\\% for Gemini, GPT, and Gemma.", "keywords": [ "retrieved augmented generation", "RAG", "LLMs", "selective generation", "sufficient context", "factuality" ], "rating_list": [ 8, 3, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "JjdU6ysnCr", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Andi Han", "gender": "Male", "institution": "RIKEN AIP", "country": "JP", "position": "Postdoc" }, { "name": "Difan Zou", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Wei Huang", "gender": "Male", "institution": "RIKEN AIP", "country": "JP", "position": "Research Scientist" }, { "name": "Yuan Cao", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "Assistant Professor" } ], "n_formula": 251, "n_formula_1": 28, "n_ref_uni": 51, "n_ref": 135, "n_ref_all": 175, "n_fig": 13, "n_tab": 7, "L_tab": 1360, "n_element_tab": 82, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 769, "n_element_tab_1": 9, "formula_len_all": 34627, "formula_len_all_1": 4379, "len_all": 340772, "len_all_1": 66322, "len_abs": 967, "len_title": 91, "len_sents": 83391, "len_sents_1": 26952, "n_sents": 897, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 942, "title": "On the feature learning in diffusion models", "abs": "The predominant success of diffusion models in generative modeling has spurred significant interest in understanding their theoretical foundations. In this work, we propose a feature learning framework aimed at analyzing and comparing the training dynamics of diffusion models with those of traditional classification models. Our theoretical analysis demonstrates that, under identical settings, neural networks trained for classification tend to prioritize learning specific patterns in the data, often focusing on easy-to-learn features. In contrast, diffusion models, due to the denoising objective, are encouraged to learn more balanced and comprehensive representations of the data. To support these theoretical insights, we conduct several experiments on both synthetic and real-world datasets, which empirically validate our findings and underscore the distinct feature learning dynamics in diffusion models compared to classification.", "keywords": [ "Diffusion model", "Feature learning" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 4, 3, 4, 3, 3 ], "presentation_list": [ 4, 3, 4, 3, 4 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 3, 2, 2, 3, 3 ] }, { "paper_id": "JjQpbbcCSp", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Ali Naseh", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" }, { "name": "Amir Houmansadr", "gender": "Male", "institution": "University of Massachusetts, Amherst", "country": "US", "position": "Associate Professor" }, { "name": "Eugene Bagdasaryan", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Jaechul Roh", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 38, "n_ref": 62, "n_ref_all": 95, "n_fig": 14, "n_tab": 12, "L_tab": 5617, "n_element_tab": 303, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 778, "n_element_tab_1": 40, "formula_len_all": 239, "formula_len_all_1": 35, "len_all": 162572, "len_all_1": 60158, "len_abs": 1126, "len_title": 90, "len_sents": 55308, "len_sents_1": 28711, "n_sents": 380, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 1146, "title": "Backdooring Bias into Text-to-Image Models", "abs": "Text-conditional diffusion models, i.e. text-to-image, produce eye-catching images that represent descriptions given by a user. \nThese images often depict benign concepts but could also carry other purposes. Specifically, visual information is easy to comprehend and could be weaponized for propaganda -- a serious challenge given widespread usage and deployment of generative models. In this paper, we show that an adversary can add an arbitrary bias through a backdoor attack that would affect even benign users generating images. While a user could inspect a generated image to comply with the given text description, our attack remains stealthy as it preserves semantic information given in the text prompt. Instead, a compromised model modifies other unspecified features of the image to add desired biases (that increase by $4-8\\times$). Furthermore, we show how the current state-of-the-art generative models make this attack both cheap and feasible for any adversary, with costs ranging between \\\\$12-\\\\$18. We evaluate our attack over various types of triggers, adversary objectives, and biases and discuss mitigations and future work.", "keywords": [ "trustworthy ml", "fairness", "backdoor attack", "text-to-image models" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "JjMRdXPpKQ", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "James R. Glass", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Principal Researcher" }, { "name": "Liming Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Marco Vilela", "gender": "Male", "institution": "Analog Devices", "country": "US", "position": "Researcher" }, { "name": "Yuan Gong", "gender": "Male", "institution": "xAI Corp", "country": "US", "position": "Researcher" } ], "n_formula": 127, "n_formula_1": 51, "n_ref_uni": 33, "n_ref": 63, "n_ref_all": 91, "n_fig": 8, "n_tab": 3, "L_tab": 1101, "n_element_tab": 100, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 800, "n_element_tab_1": 76, "formula_len_all": 13757, "formula_len_all_1": 4148, "len_all": 187904, "len_all_1": 61747, "len_abs": 1121, "len_title": 107, "len_sents": 44633, "len_sents_1": 24681, "n_sents": 374, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 936, "title": "Can Diffusion Models Disentangle? A Theoretical Perspective", "abs": "This paper introduces a novel theoretical framework to understand how diffusion models can learn disentangled representations under the assumption of an $\\normltwo$ score approximation. We also provide sufficient conditions under which such representations are beneficial for domain adaptation. Our theory offers new insights into how existing diffusion models disentangle latent variables across general distributions and suggests strategies to enhance their disentanglement capabilities. To validate our theory, we perform experiments using both synthetic data generated from latent subspace models and real speech data for non-parallel voice conversion - a canonical disentanglement problem. Across various classification tasks, we found voice conversion-based adaptation methods achieve significant improvements in classification accuracy, demonstrating their effectiveness as domain adaptors. Code will be released upon acceptance.", "keywords": [ "diffusion model", "disentanglement", "voice conversion" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 1, 1, 2, 1 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "Jj4XIKX4TJ", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Allan Dos Santos Costa", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Danny Reidenbach", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Emine Kucukbenli", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Senior Research Scientist" }, { "name": "Franco Pellegrini", "gender": "Male", "institution": "International School for Advanced Studies Trieste", "country": "", "position": "Assistant Professor" }, { "name": "Guoqing Zhou", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Karsten Kreis", "gender": "unknown", "institution": "NVIDIA", "country": "CA", "position": "Research Scientist" }, { "name": "Mario Geiger", "gender": "Male", "institution": "NVIDIA", "country": "CH", "position": "Researcher" }, { "name": "Tomas Geffner", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Zhonglin Cao", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Applied Scientist" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 37, "n_ref": 88, "n_ref_all": 107, "n_fig": 6, "n_tab": 4, "L_tab": 2160, "n_element_tab": 264, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1536, "n_element_tab_1": 232, "formula_len_all": 1464, "formula_len_all_1": 1130, "len_all": 132449, "len_all_1": 62888, "len_abs": 1326, "len_title": 85, "len_sents": 36244, "len_sents_1": 26540, "n_sents": 284, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1329, "title": "Efficient molecular conformer generation with SO(3) averaged flow-matching and reflow", "abs": "Molecular conformer generation is a critical task in computational chemistry and drug discovery. Diverse generative deep learning methods have been proposed and shown to outperform traditional cheminformatics tools. State-of-the-art models leverage neural transport, employing denoising diffusion or flow-matching to generate or refine atomic point clouds from a prior distribution. Still, sampling with existing models requires significant computational expense. In this work, we build upon flow-matching and propose two mechanisms for accelerating training and inference of 3D molecular conformer generation. For fast training, we introduce the SO(3)-Averaged Flow, which we show to converge faster and generate better conformer ensembles compared to flow-matching and Kabsch alignment-based optimal transport flow. For fast inference, we further show that reflow methods and distillation of these models enable few-steps or even one-step molecular conformer generation with high quality. Using these two techniques, we demonstrate a model that can match the performance of strong transformer baselines with only a fraction of the number of parameters and generation steps. The training techniques proposed in this work lay the foundation for highly efficient molecular conformer generation with generative deep learning model.", "keywords": [ "Flow-matching", "few-shot generation", "equivariance", "small molecules" ], "rating_list": [ 6, 8, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 3, 4 ] }, { "paper_id": "JigWdDArjb", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Jeongwon Ryu", "gender": "Male", "institution": "SI Analytics", "country": "", "position": "Researcher" }, { "name": "Minseok Seo", "gender": "Male", "institution": "SI Analytics", "country": "", "position": "Researcher" }, { "name": "Youngtack Oh", "gender": "unknown", "institution": "SI Analytics", "country": "KR", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 37, "n_ref": 67, "n_ref_all": 91, "n_fig": 13, "n_tab": 6, "L_tab": 2054, "n_element_tab": 276, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 2671, "n_element_tab_1": 403, "formula_len_all": 198, "formula_len_all_1": 199, "len_all": 136691, "len_all_1": 67421, "len_abs": 1476, "len_title": 122, "len_sents": 37911, "len_sents_1": 28156, "n_sents": 281, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1382, "title": "SAR2Earth: A SAR-to-EO Translation Dataset for Remote Sensing Applications", "abs": "Electro-optical (EO) images are essential to a wide range of remote sensing applications. With the advent of data-driven models, the efficiency of EO image analysis has significantly improved, enabling faster and more effective outcomes in these applications. However, EO images have inherent limitations\u2014they cannot penetrate cloud cover and are unable to capture imagery at night. To overcome these challenges, synthetic aperture radar (SAR) images are employed, as they can operate effectively regardless of weather conditions or time of day. Despite this advantage, SAR images come with their own difficulties: they are affected by speckle noise, complicating analysis, and existing algorithms developed for EO imagery are not directly transferable to SAR data. To address these issues, we introduce SAR2Earth, a benchmark dataset specifically designed for SAR-to-EO translation. By translating SAR images into EO-like representations, SAR2Earth allows the extensive range of algorithms developed for EO imagery to be applied effectively to SAR data. The dataset consists of 18 spatially aligned pairs of SAR and EO images, collected from 8 distinct regions encompassing both urban and rural. We provide comprehensive evaluations, detailed model analyses, and extensive experimental results. All codes and datasets will be made publicly available at https://sar2earth.github.io.", "keywords": [ "AI for social good", "\bDataset", "SAR-to-EO translation", "Benchmark" ], "rating_list": [ 5, 5, 5, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 3, 4, 5, 5 ] }, { "paper_id": "JiX2DuTkeU", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Chunhui Wang", "gender": "Male", "institution": "geely ai lab", "country": "AF", "position": "Researcher" }, { "name": "JianZhao", "gender": "unknown", "institution": "Jiangsu University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Kai Yu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Keqi Deng", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Xie Chen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yushen CHEN", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Undergrad student" }, { "name": "Zhikang Niu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Ziyang Ma", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 57, "n_ref": 137, "n_ref_all": 152, "n_fig": 3, "n_tab": 10, "L_tab": 5943, "n_element_tab": 637, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1407, "n_element_tab_1": 110, "formula_len_all": 667, "formula_len_all_1": 668, "len_all": 181856, "len_all_1": 59859, "len_abs": 1455, "len_title": 125, "len_sents": 45820, "len_sents_1": 27740, "n_sents": 330, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1462, "title": "F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching", "abs": "This paper introduces F5-TTS, a fully non-autoregressive text-to-speech system based on flow matching with Diffusion Transformer (DiT). Without requiring complex designs such as duration model, text encoder, and phoneme alignment, the text input is simply padded with filler tokens to the same length as input speech, and then the denoising is performed for speech generation, which was originally proved feasible by E2 TTS. However, the original design of E2 TTS makes it hard to follow due to its slow convergence and low robustness. To address these issues, we first model the input with ConvNeXt to refine the text representation, making it easy to align with the speech. We further propose an inference-time Sway Sampling strategy, which significantly improves our model's performance and efficiency. This sampling strategy for flow step can be easily applied to existing flow matching based models without retraining. Our design allows faster training and achieves an inference RTF of 0.15, which is greatly improved compared to state-of-the-art diffusion-based TTS models. Trained on a public 100K hours multilingual dataset, our Fairytaler Fakes Fluent and Faithful speech with Flow matching (F5-TTS) exhibits highly natural and expressive zero-shot ability, seamless code-switching capability, and speed control efficiency. Demo samples can be found at https://F5-TTS.github.io. We will release all code and checkpoints to promote community development.", "keywords": [ "text-to-speech", "speech synthesis", "flow matching", "diffusion model", "efficiency" ], "rating_list": [ 8, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 4, 5, 5 ] }, { "paper_id": "JiWlVYB4rh", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Dana Moukheiber", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Researcher" }, { "name": "Lama Moukheiber", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 41, "n_ref": 48, "n_ref_all": 54, "n_fig": 4, "n_tab": 4, "L_tab": 3310, "n_element_tab": 386, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2560, "n_element_tab_1": 335, "formula_len_all": 100, "formula_len_all_1": 100, "len_all": 107807, "len_all_1": 58782, "len_abs": 1387, "len_title": 55, "len_sents": 26876, "len_sents_1": 26277, "n_sents": 171, "n_sents_1": 168, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1408, "title": "EchoQA: Tuning into the Heart of Echocardiogram Reports", "abs": "We introduce a novel and extensive instruction-tuning dataset using echocardiogram reports sourced from MIMIC-IV. This dataset is specifically tailored to enhance question answering (QA) systems in the field of cardiology. It comprises 765,605 QA pairs addressing a wide array of cardiac abnormalities and their severity. To validate the utility of this benchmark dataset, we employ various large language models (LLMs), encompassing both open-source general models and biomedical-specific models, along with state-of-the-art closed-source models for zero-shot evaluation. Our results reveal that certain models achieve superior performance across all evaluated metrics. This underscores the effectiveness of instruction fine-tuning for echocardiogram data. Additionally, we conduct an audit of the best performing LLM across demographic groups and marginalized populations. Our objective is to propel the field forward by establishing a benchmark framework for developing LLM AI agents that support clinicians in their daily workflow within the cardiology space. The availability of this dataset aims to support the advancement of natural language models for use in diagnostic decision support systems, aiming to increase efficiency and decrease diagnostic errors in cardiology care. All code will be available on the Github and the data will be made available on HIPAA-compliant data repository PhysioNet.", "keywords": [ "Large language models", "fine-tuning", "fairness", "benchmarking" ], "rating_list": [ 3, 1, 5 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 3, 1, 1 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "JgSbMcgd8q", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Yingyu Liang", "gender": "Male", "institution": "University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Yufa Zhou", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "MS student" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Zhenmei Shi", "gender": "Male", "institution": "Voyage AI", "country": "US", "position": "Researcher" }, { "name": "Zhizhou Sha", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 180, "n_formula_1": 14, "n_ref_uni": 59, "n_ref": 178, "n_ref_all": 212, "n_fig": 0, "n_tab": 2, "L_tab": 748, "n_element_tab": 26, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 129, "n_element_tab_1": 2, "formula_len_all": 21024, "formula_len_all_1": 1496, "len_all": 465473, "len_all_1": 71620, "len_abs": 965, "len_title": 123, "len_sents": 120827, "len_sents_1": 31931, "n_sents": 1504, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 988, "title": "Multi-Layer Transformers Gradient Can be Approximated in Almost Linear Time", "abs": "The computational complexity of the self-attention mechanism in popular transformer architectures poses significant challenges for training and inference, and becomes the bottleneck for long inputs. Is it possible to significantly reduce the quadratic time complexity of computing the gradients in multi-layer transformer models? This paper proves that a novel fast approximation method can calculate the gradients in almost linear time $n^{1+o(1)}$ where $n$ is the input sequence length, while it maintains a polynomially small approximation error $1 / \\mathrm{poly}(n)$ across the entire model. \nOur theory holds for general loss functions and when the multi-layer transformer model contains many practical sub-modules, such as residual connection, casual mask, and multi-head attention. \nBy improving the efficiency of gradient computation, we hope that this work will facilitate more effective training and deployment of long-context language models based on our theoretical results.", "keywords": [ "Multi-Layer Transformers", "Attention Acceleration", "Gradient Computation" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "Jg9Ol9aVjx", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Zhengyang Gu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Zhipeng Xie", "gender": "Male", "institution": "Fudan University", "country": "", "position": "Associate Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 27, "n_ref": 47, "n_ref_all": 86, "n_fig": 11, "n_tab": 4, "L_tab": 2118, "n_element_tab": 199, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 1654, "n_element_tab_1": 175, "formula_len_all": 797, "formula_len_all_1": 821, "len_all": 98909, "len_all_1": 62839, "len_abs": 1343, "len_title": 124, "len_sents": 30868, "len_sents_1": 27293, "n_sents": 231, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1183, "title": "DICR: Direct Intra-image Constrastive Regularization for Contrastive Learning", "abs": "Typical contrastive self-supervised learning methods apply inter-image contrast to post-projector embeddings, thereby indirectly encouraging the pre-projector representations' invariance to several augmentation operators.\nWhile effective, these methods do not account for the inherent difference between semantics-altering (such as cropping and cutout) and semantics-preserving augmentation operators (such as resizing, flipping and color distortion), and thereby lack an explicit mechanism to encourage distinguishable representations for semantically different contents within the same image.\nWe explain, both in reason and in practice, that these issues can harm the generalizability of the representations in downstream tasks.\nTo address these issues, we propose Direct Intra-image Contrastive Regularization (DICR), a plug-and-play regularization method that directly applies intra-image contrast to pre-projector representations.\nEmpirical results show that DICR can significantly enhance the generalizability of existing methods in downstream tasks, and validate the crucial role of semantic content distinguishability in the generalizable performance of contrastive learning.", "keywords": [ "Contrastive Learning", "Representation Learning", "Self-Supervised Learning" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "JfgBhEqk6F", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Cheng-zhong Xu", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Full Professor" }, { "name": "Chunlin Tian", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "PhD student" }, { "name": "He Sun", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Intern" }, { "name": "KaHou Tam", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "PhD student" }, { "name": "Li Li", "gender": "Male", "institution": "University of Macau", "country": "CN", "position": "Assistant Professor" }, { "name": "Mang Ye", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Professor" }, { "name": "Yebo Wu", "gender": "Male", "institution": "University of Macau", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 31, "n_ref": 78, "n_ref_all": 103, "n_fig": 8, "n_tab": 9, "L_tab": 4630, "n_element_tab": 293, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 2385, "n_element_tab_1": 201, "formula_len_all": 497, "formula_len_all_1": 496, "len_all": 136850, "len_all_1": 66987, "len_abs": 1953, "len_title": 151, "len_sents": 45118, "len_sents_1": 30703, "n_sents": 317, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 2025, "title": "Honey: Harmonizing Progressive Federated Learning via Elastic Synergy across Different Training Blocks", "abs": "Memory limitation is becoming the prevailing challenge that hinders the deployment of Federated Learning on mobile/IoT devices in real-world cases. Progressive training offers a promising alternative to surpass memory constraints. Instead of updating the full model in each training round, progressive training divides the model into multiple blocks and iteratively updates each block until the full model is converged. However, existing progressive training approaches suffer from prominent accuracy degradation as training each block in isolation drives it to prioritize features that are only beneficial to its specific needs, neglecting the overall learning objective. To address this issue, we present $\\texttt{\\textbf{Honey}}$, a synergistic progressive training approach that integrates the holistic view and block-wise feedback to facilitate the training of each block. Specifically, the holistic view broadens the learning scope of each block, ensuring that it operates in harmony with the global objective and benefits the training of the whole model. Simultaneously, block-wise feedback heightens each block's awareness of its role and position within the full model, empowering it to make real-time adjustments based on insights from downstream blocks and facilitating a smooth and consistent information flow. Furthermore, to fully harness the heterogeneous memory resources of participating devices, we develop an elastic resource harmonization protocol. \nThis protocol authorizes each device to adaptively train specific layers according to their memory capacity, optimizing resource utilization, sparking cross-block communication, and accelerating model convergence. Comprehensive experiments on benchmark datasets and models demonstrate that $\\texttt{\\textbf{Honey}}$ outperforms state-of-the-art approaches, delivering an exceptional average accuracy improvement of up to 43.9\\%. Moreover, $\\texttt{\\textbf{Honey}}$ achieves comparable performance even with a reduction in peak memory usage of up to 49\\%.", "keywords": [ "Memory-Heterogeneous Federated Learning", "Progressive Training", "On-Device Training" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "JffVqPWQgg", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Feiran Huang", "gender": "Male", "institution": "Jinan University", "country": "CN", "position": "Full Professor" }, { "name": "Hao Chen", "gender": "Male", "institution": "City University of Macau", "country": "MO", "position": "Assistant Professor" }, { "name": "Qinggang Zhang", "gender": "Male", "institution": "The Hong Kong Polytechnic University, Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Xiao Huang", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "", "position": "Assistant Professor" }, { "name": "Zheng Yuan", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "CN", "position": "PhD student" }, { "name": "Zijin Hong", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 24, "n_ref_uni": 28, "n_ref": 93, "n_ref_all": 121, "n_fig": 3, "n_tab": 13, "L_tab": 5331, "n_element_tab": 363, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 590, "n_element_tab_1": 9, "formula_len_all": 2038, "formula_len_all_1": 1249, "len_all": 173760, "len_all_1": 61008, "len_abs": 1696, "len_title": 114, "len_sents": 52904, "len_sents_1": 28889, "n_sents": 386, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1799, "title": "Knapsack Schema Linking Agent for LLM-Based Text-to-SQL Generation", "abs": "Generating SQLs according to user queries (text-to-SQL) is a long-standing sequential challenge, where the accuracy of the initial schema linking significantly impacts the subsequent SQL generation performance. However, existing models often focus more on SQL generation and less on the schema linking task, leading to potential missing or redundant schema linking and suboptimal SQL generation performance. The underlying reason is that schema linking is not a simple selection problem but a \\textbf{Knapsack problem}, which should consider both the \\textit{value} of the schema linking in terms of missing important information and the \\textit{weight} of the schema linking in terms of providing redundant information. Motivated by this, we provide two tailored SL benchmarks and two tailored metrics to train SL agents and to evaluate the missing and redundant schema linking. In this paper, we propose the \\textbf{\\underline{K}n\\underline{a}psack \\underline{S}chema \\underline{L}inking \\underline{A}gent (KaSLA)}, which can link the most valuable and least redundant schema element subsets for both tables and columns. KaSLA introduces an importance score function to predict each schema element's importance score, and then utilizes the importance score to estimate the value and the weight of each schema. Then, by estimating the capacity, the maximum weight the knapsack can hold, of a given user query from historical SQL records, KaSLA employs efficient dynamic programming to select the most valuable schema element set within the estimated capacity. Extensive experiments on two benchmark datasets demonstrate the superior performance of KaSLA over 12 state-of-the-art baselines. Especially on the popular and challenging BIRD benchmark, KaSLA can outperform the baselines by over 5.72\\%.", "keywords": [ "Large Language Models", "Text-to-SQL", "Schema Linking" ], "rating_list": [ 6, 5, 5, 6, 3 ], "soundness_list": [ 4, 3, 2, 4, 1 ], "presentation_list": [ 2, 2, 3, 3, 1 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 5, 4, 4 ] }, { "paper_id": "JfRPsrP6qX", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Chanyoung Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Jie Peng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Namkyeong Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Sukwon Yun", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Yanyong Zhang", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Zunpeng Liu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 65, "n_ref": 90, "n_ref_all": 113, "n_fig": 4, "n_tab": 3, "L_tab": 739, "n_element_tab": 153, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 739, "n_element_tab_1": 153, "formula_len_all": 473, "formula_len_all_1": 473, "len_all": 161713, "len_all_1": 66454, "len_abs": 1602, "len_title": 131, "len_sents": 37536, "len_sents_1": 31523, "n_sents": 260, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1515, "title": "Enhancing single-cell Multi-Modal Multi-Task Learning via Sparse Mixture-of-Experts", "abs": "Recent advances in measuring high-dimensional modalities, including protein levels and DNA accessibility, at the single-cell level have prompted the need for frameworks capable of handling multi-omics data while simultaneously addressing multiple tasks. Despite these advancements, much of the work in the single-cell domain remains limited, often focusing on either a single-modal or single-task perspective. A few recent studies have ventured into multi-omics and multi-task learning, but we identified a \u2460 Optimization Conflict issue, leading to suboptimal results when integrating additional modalities, which is undesirable. Furthermore, there is a \u2461 Costly Interpretability challenge, as current approaches predominantly rely on costly post-hoc methods like SHAP. Motivated by these challenges, we introduce scMoE, a novel framework that, for the first time, applies Sparse Mixture-of-Experts (SMoE) within the single-cell domain. This is achieved by incorporating an SMoE layer into a transformer block with a cross-attention module. Thanks to its design, scMoE inherently possesses mechanistic interpretability, a critical aspect for understanding underlying mechanisms when handling biological data. Furthermore, from a post-hoc perspective, we enhance interpretability by extending the concept of activation vectors (CAVs). Extensive experiments on simulated dataset, Dyngen, and real-world multi-omics single-cell datasets, including DBiT-seq, Patch-seq, ATAC-seq, demonstrate the effectiveness of scMoE.", "keywords": [ "Single-cell Multi-modal data", "Multi-Task Learning", "Sparse MoE" ], "rating_list": [ 1, 5, 3, 5 ], "soundness_list": [ 1, 2, 1, 2 ], "presentation_list": [ 1, 2, 1, 3 ], "contribution_list": [ 1, 3, 1, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "JfKF7Pdigi", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Liulei Li", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Ruijie Quan", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Wenguan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Wensong Song", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 67, "n_ref": 114, "n_ref_all": 139, "n_fig": 12, "n_tab": 4, "L_tab": 2216, "n_element_tab": 141, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2354, "n_element_tab_1": 176, "formula_len_all": 538, "formula_len_all_1": 503, "len_all": 177604, "len_all_1": 77214, "len_abs": 1500, "len_title": 103, "len_sents": 48790, "len_sents_1": 35830, "n_sents": 366, "n_sents_1": 255, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1514, "title": "MindGrapher: Dynamic-Aware fMRI-to-Video Reconstruction", "abs": "Existing methods for fMRI-to-video reconstruction typically focus on accurately reconstructing visual content ($i.e.$, appearance), neglecting dynamic event information. However, as highlighted in cognitive neurology, these key dynamic events significantly influence brain signal changes during video perception. In this article, we introduce Mindgrapher, a two-stream framework designed to address this gap by enhancing the reconstruction of dynamic-aware videos from fMRI data. Mindgrapher comprises $i)$ a visual content reconstruction stream, that improves the accuracy of the reconstructed visual content from sparsely distributed fMRI data through a temporal dynamics enrichment approach and multi-moment multimodal contrastive learning; $ii)$ a dynamics injection stream, that firstly crafts dynamic-aware fMRI embeddings and then integrates them into the reconstruction process via a fine-grained approach, thereby producing videos that effectively perceive dynamic events. Moreover, to address the lack of suitable metrics for evaluating dynamic event information, we introduce a new evaluation metric named dynamic content fidelity (DCF), which measures how accurately dynamic events within the video are reconstructed. Upon evaluation with a publicly available fMRI dataset, Mindgrapher outperforms the state-of-the-arts on all metrics, $i.e.$, semantic classification accuracy, structural similarity index, and DCF. The reconstructed video results are available on our web page. Code shall be released.", "keywords": [ "fMRI; fMRI-to-Video; Dynamic-Aware; Video Reconstruction from Brain Activities;" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 1, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "JetCx7Tpgb", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Dong Zhao", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "PhD student" }, { "name": "Elisa Ricci", "gender": "Female", "institution": "University of Trento", "country": "IT", "position": "Full Professor" }, { "name": "Jinlong Li", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "PhD student" }, { "name": "Lin Ma", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Principal Researcher" }, { "name": "Nicu Sebe", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Full Professor" }, { "name": "ZEQUN JIE", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 18, "n_ref_uni": 40, "n_ref": 106, "n_ref_all": 116, "n_fig": 1, "n_tab": 7, "L_tab": 4573, "n_element_tab": 603, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 840, "n_element_tab_1": 34, "formula_len_all": 1799, "formula_len_all_1": 1201, "len_all": 189663, "len_all_1": 59228, "len_abs": 1971, "len_title": 101, "len_sents": 41832, "len_sents_1": 28757, "n_sents": 274, "n_sents_1": 170, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 1979, "title": "Enhancing Robustness of Vision-Language Models through Orthogonality Learning and Self-Regularization", "abs": "Efficient fine-tuning of vision-language models (VLMs) like CLIP for specific downstream tasks is gaining significant attention. Previous works primarily focus on prompt learning to adapt the CLIP into a variety of downstream tasks, however, suffering from task overfitting when fine-tuned on a small data set. In this paper, we introduce an orthogonal fine-tuning method for efficiently fine-tuning pretrained weights and enabling enhanced robustness and generalization, while a self-regularization strategy is further exploited to maintain the stability in terms of zero-shot generalization of VLMs, dubbed OrthSR. Specifically, trainable orthogonal matrices are injected seamlessly into the transformer architecture and enforced with orthogonality constraint during the training, benefiting from the norm-preserving property and thus leading to stable and faster convergence, while keeping the pre-trained weights frozen. \nTo alleviate deviation from fine-tuning, \na self-regularization strategy is further employed to retain the generalization of the model during the training within a bypass manner. In addition, to enrich the sample diversity for downstream tasks under the small dataset scenario, we first explore attentive CutOut data augmentation to boost the efficient fine-tuning, leading to better model fitting capacity for specific downstream task. Then we support the theoretical analysis on how our approach improves the specific downstream performance and maintains the generalizability. For the first time, we revisit the CLIP and CoOp with our method to effectively improve the model on few-shot image classficiation scenario on par with the elaborated prompt learning methods. We conduct extensive experiments to demonstrate that our method explicitly steers pretrained weight space to represent the task-specific knowledge and presents competitive generalizability under base-to-base/base-to-new, cross-dataset transfer and domain generalization evaluations.", "keywords": [ "vision-language model; few-shot; fine-tuning" ], "rating_list": [ 5, 3, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 5, 3 ] }, { "paper_id": "JeiaHDawhb", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Bang You", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Huaping Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jan Peters", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Principal Researcher" }, { "name": "Oleg Arenz", "gender": "Male", "institution": "TU Darmstadt", "country": "", "position": "Postdoc" }, { "name": "Puze Liu", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 11, "n_ref_uni": 41, "n_ref": 74, "n_ref_all": 86, "n_fig": 6, "n_tab": 2, "L_tab": 980, "n_element_tab": 59, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 523, "n_element_tab_1": 19, "formula_len_all": 2737, "formula_len_all_1": 1019, "len_all": 127958, "len_all_1": 61690, "len_abs": 1041, "len_title": 96, "len_sents": 39797, "len_sents_1": 30209, "n_sents": 287, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1044, "title": "Maximum Total Correlation Reinforcement Learning", "abs": "Simplicity is a powerful inductive bias. In reinforcement learning, regularization is used for simpler policies, data augmentation for simpler representations, and sparse reward functions for simpler objectives, all that, with the underlying motivation to increase generalizability and robustness by focusing on the essentials. Supplementary to these techniques, we investigate how to promote simple behavior throughout the duration of the episode. To that end, we introduce a modification of the reinforcement learning problem, that additionally maximizes the total correlation within the induced trajectories. We propose a practical algorithm that optimizes all models, including policy and state representation, based on a lower bound approximation. In simulated robot locomotion environments, our method naturally generates policies that induce periodic and compressible trajectories, and that exhibit superior robustness to noise and changes in dynamics compared to baseline methods, while also improving performance in the original tasks.", "keywords": [ "Reinforcement Learning", "Total Correlation" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "JeZS4jQF77", "primary_area": "learning on time series and dynamical systems", "n_author": 7, "author_info_list": [ { "name": "Dayu Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Guoqi Yu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Intern" }, { "name": "Shujun Wang", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "Tong Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Yaoming Li", "gender": "unknown", "institution": "Harbin University of Commerce", "country": "CN", "position": "Undergrad student" }, { "name": "Zirui Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "xiaoyu guo", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 11, "n_ref_uni": 32, "n_ref": 93, "n_ref_all": 119, "n_fig": 6, "n_tab": 16, "L_tab": 9462, "n_element_tab": 653, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 4047, "n_element_tab_1": 320, "formula_len_all": 1357, "formula_len_all_1": 737, "len_all": 190968, "len_all_1": 67542, "len_abs": 2094, "len_title": 165, "len_sents": 46661, "len_sents_1": 27909, "n_sents": 379, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 116, "L_abs": 1794, "title": "LiNo: Advancing Recursive Residual Decomposition of Linear and Nonlinear Patterns for Robust Time Series Forecasting", "abs": "Forecasting models are pivotal in a data-driven world with vast volumes of time series data that appear as a compound of vast $\\textbf{Li}$near and $\\textbf{No}$nlinear patterns. \nRecent deep time series forecasting models struggle to utilize seasonal and trend decomposition to separate the entangled components. Such a strategy only explicitly extracts simple linear patterns like trends, leaving the other linear modes and vast unexplored nonlinear patterns to the residual. Their flawed linear and nonlinear feature extraction models and shallow-level decomposition limit their adaptation to the diverse patterns present in real-world scenarios.\nGiven this, we innovate Recursive Residual Decomposition by introducing explicit extraction of both linear and nonlinear patterns. This deeper-level decomposition framework, which is named $\\textbf{LiNo}$, captures linear patterns using a Li block which can be a moving average kernel, and models nonlinear patterns using a No block which can be a Transformer encoder. The extraction of these two patterns is performed alternatively and recursively. To achieve the full potential of LiNo, we develop the current simple linear pattern extractor to a general learnable autoregressive model, and design a novel No block that can handle all essential nonlinear patterns.\nRemarkably, the proposed LiNo achieves state-of-the-art on thirteen real-world benchmarks under univariate and multivariate forecasting scenarios. Experiments show that current forecasting models can deliver more robust and precise results through this advanced Recursive Residual Decomposition. We hope this work could offer insight into designing more effective forecasting models. Code is available at this anonymous repository: https://anonymous.4open.science/r/LiNo-8225/.", "keywords": [ "Time Series Forecasting", "Deep learning." ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "JeLqFpFzwX", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Isaac Lin", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Undergrad student" }, { "name": "Shang Gao", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Researcher" }, { "name": "Tai Sing Lee", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Tang Shiming", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Tianye Wang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 73, "n_ref_all": 101, "n_fig": 11, "n_tab": 3, "L_tab": 1402, "n_element_tab": 119, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1355, "n_element_tab_1": 116, "formula_len_all": 93, "formula_len_all_1": 409, "len_all": 161510, "len_all_1": 72991, "len_abs": 1452, "len_title": 129, "len_sents": 53873, "len_sents_1": 34515, "n_sents": 387, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1457, "title": "Self-Attention-Based Contextual Modulation Improves Neural System Identification", "abs": "Convolutional neural networks (CNNs) have been shown to be state-of-the-art models for visual cortical neurons. Cortical neurons in the primary visual cortex are sensitive to contextual information mediated by extensive horizontal and feedback connections. Standard CNNs integrate global contextual information to model contextual modulation via two mechanisms: successive convolutions and a fully connected readout layer. In this paper, we find that self-attention (SA), an implementation of non-local network mechanisms, can improve neural response predictions over parameter-matched CNNs in two key metrics: tuning curve correlation and peak tuning. We introduce peak tuning as a metric to evaluate a model's ability to capture a neuron's feature preference. We factorize networks to assess each context mechanism, revealing that information in the local receptive field is most important for modeling overall tuning, but surround information is critically necessary for characterizing the tuning peak. We find that self-attention can replace posterior spatial-integration convolutions when learned incrementally, and is further enhanced in the presence of a fully connected readout layer, suggesting that the two context mechanisms are complementary. Finally, we find that decomposing receptive field learning and contextual modulation learning in an incremental manner may be an effective and robust mechanism for learning surround-center interactions.", "keywords": [ "self-attention", "incremental learning", "neural prediction", "contextual modulation" ], "rating_list": [ 8, 3, 6, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 3, 2, 4, 3 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "JeJ2uTQrF1", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Guogang Zhu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Jianwei Niu", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Mingjia Shi", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Shaojie Tang", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Full Professor" }, { "name": "Xinghao Wu", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Xuefeng Liu", "gender": "Male", "institution": "Behang University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 33, "n_formula_1": 15, "n_ref_uni": 12, "n_ref": 21, "n_ref_all": 65, "n_fig": 9, "n_tab": 19, "L_tab": 6698, "n_element_tab": 523, "n_fig_1": 4, "n_tab_1": 9, "L_tab_1": 3604, "n_element_tab_1": 281, "formula_len_all": 3073, "formula_len_all_1": 1002, "len_all": 182888, "len_all_1": 62110, "len_abs": 2065, "len_title": 175, "len_sents": 54867, "len_sents_1": 23283, "n_sents": 510, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 126, "L_abs": 2077, "title": "From Mismatch to Harmony: Resolving Feature-Classifier Mismatch in Federated Learning via Prompt-Driven Feature Transformation", "abs": "In conventional Federated Learning approaches like FedAvg, training a global model becomes challenging in the presence of data heterogeneity. To address this, Personalized Federated Learning (PFL) has emerged as a leading solution, enabling clients to train personalized models that are tailored to local data distributions. Surprisingly, our linear probe experiments reveal that FedAvg\u2019s feature extractor outperforms most PFL methods on local client data. Even more intriguingly, applying a simple linear transformation to align local features from FedAvg\u2019s extractor with the classifier enables FedAvg to surpass most PFL methods. These findings suggest that in data heterogeneity scenarios, FedAvg\u2019s weaker performance is not due to inadequate global model training but rather a mismatch between local features and the classifier. This observation motivates us to develop a new framework to address this mismatch problem. A straightforward solution would be to insert the personalized linear transformation layer mentioned above between the global feature extractor and the global classifier. However, this approach can easily overfit the limited local training data due to the large number of personalized parameters, and it is insufficient for handling complex datasets. In this paper, we introduce FedPFT, which leverages personalized prompts to resolve the mismatch problem. These prompts, along with local features, are fed into a shared self-attention-based module, where features are transformed via the attention mechanism to align with the global classifier. These prompts consist of minimal trainable parameters, reducing the risk of overfitting to local data. Additionally, this prompt-driven approach offers strong flexibility, allowing for task-specific prompts to integrate additional training objectives (e.g., contrastive learning) to further enhance performance. Our experiments demonstrate that FedPFT outperforms state-of-the-art methods by up to 5.07\\%, with additional improvements of up to 7.08\\% when collaborative contrastive learning is introduced.", "keywords": [ "Personalized Federated Learning", "Data Heterogeneity", "Feature-Classifier Mismatch", "Prompt-Driven Feature Transformation" ], "rating_list": [ 3, 8, 3, 3, 1 ], "soundness_list": [ 2, 4, 2, 3, 1 ], "presentation_list": [ 2, 4, 3, 3, 2 ], "contribution_list": [ 2, 4, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "JdtukDPwIV", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Haozhao Wang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Postdoc" }, { "name": "Junhong Liu", "gender": "unknown", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Renyu Yang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Associate Professor" }, { "name": "Tao Huang", "gender": "Male", "institution": "The University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Wenchao Xu", "gender": "Not Specified", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Researcher" }, { "name": "Yuan Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Shanghang Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 45, "n_ref": 56, "n_ref_all": 77, "n_fig": 6, "n_tab": 4, "L_tab": 1051, "n_element_tab": 244, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2629, "n_element_tab_1": 170, "formula_len_all": 643, "formula_len_all_1": 582, "len_all": 125210, "len_all_1": 63118, "len_abs": 1201, "len_title": 103, "len_sents": 36364, "len_sents_1": 27996, "n_sents": 271, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 1206, "title": "Frequency-Decoupled Cross-Modal Knowledge Distillation", "abs": "Knowledge distillation (KD) has proven highly effective for compressing large models and enhancing the performance of smaller ones. However, its effectiveness diminishes in cross-modal scenarios, such as vision-to-language distillation, where inconsistencies in representation across modalities lead to difficult knowledge transfer. To address this challenge, we propose frequency-decoupled cross-modal knowledge distillation, a method designed to decouple and balance knowledge transfer across modalities by leveraging frequency-domain features. We observe that low-frequency features tend to capture modality-agnostic, generalizable information, while high-frequency features are more modality-specific. Accordingly, we apply distinct losses to these features: enforcing strong alignment in the low-frequency domain and introducing relaxed alignment for high-frequency features. Additionally, we propose a scale consistency loss to address distributional shifts between modalities, and employ a shared classifier to unify feature spaces. Extensive experiments across multiple benchmark datasets show that our method substantially outperforms traditional KD and state-of-the-art cross-modal KD approaches.", "keywords": [ "Cross-Modal Knowledge Distillation (CMKD)", "Frequency-Domain Feature Decoupling", "Feature Space Alignment" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "Jds4tiTo2a", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Bo Zhao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Haoru Tan", "gender": "Male", "institution": "HKU", "country": "HK", "position": "PhD student" }, { "name": "Sitong Wu", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "XIAOJUAN QI", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Zeke Xie", "gender": "Male", "institution": "HKUST(GZ)", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 34, "n_formula_1": 15, "n_ref_uni": 61, "n_ref": 279, "n_ref_all": 316, "n_fig": 4, "n_tab": 9, "L_tab": 3329, "n_element_tab": 277, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 265, "n_element_tab_1": 20, "formula_len_all": 1861, "formula_len_all_1": 749, "len_all": 261082, "len_all_1": 70031, "len_abs": 1392, "len_title": 115, "len_sents": 80869, "len_sents_1": 31709, "n_sents": 656, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1402, "title": "Diff-In: Data Influence Estimation with Differential Approximation", "abs": "In this paper, we introduce a new formulation to approximate a sample's influence by accumulating the differences in influence between consecutive learning steps, which we term Diff-In. Specifically, we formulate the sample-wise influence as the cumulative sum of its changes/differences across successive training iterations. \nBy employing second-order approximations, we approximate these difference terms with high accuracy while eliminating the need for model convexity required by existing methods.\nDespite being a second-order method, Diff-In maintains computational complexity comparable to that of first-order methods and remains scalable. This efficiency is achieved by computing the product of the Hessian and gradient, which can be efficiently approximated using finite differences of first-order gradients. \nWe assess the approximation accuracy of Diff-In both theoretically and empirically. Our theoretical analysis demonstrates that Diff-In achieves significantly lower approximation error compared to existing influence estimators. Extensive experiments further confirm its superior performance across multiple benchmark datasets in three data-centric tasks: data cleaning, data deletion, and coreset selection. \nNotably, our experiments on data pruning for large-scale vision-language pre-training show that Diff-In can scale to millions of data points and outperforms strong baselines.", "keywords": [ "Deep Learning", "Influence Function" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "JddNOaw66n", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Aviv Slobodkin", "gender": "Male", "institution": "Google", "country": "IL", "position": "Intern" }, { "name": "Royi Rassin", "gender": "Male", "institution": "Research, Google", "country": "IL", "position": "Intern" }, { "name": "Shauli Ravfogel", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "Yanai Elazar", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Postdoc" }, { "name": "Yoav Goldberg", "gender": "Male", "institution": "Bar-Ilan University", "country": "", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 47, "n_ref_all": 84, "n_fig": 18, "n_tab": 9, "L_tab": 3335, "n_element_tab": 136, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 839, "n_element_tab_1": 28, "formula_len_all": 183, "formula_len_all_1": 144, "len_all": 190622, "len_all_1": 60155, "len_abs": 1535, "len_title": 107, "len_sents": 54750, "len_sents_1": 29016, "n_sents": 468, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1406, "title": "GRADE: Quantifying Sample Diversity in Text-to-Image Models", "abs": "Text-to-image (T2I) models are remarkable at generating realistic images based on textual descriptions. However, textual prompts are inherently *underspecified*: they do not specify all possible attributes of the required image. This raises key questions: do T2I models generate diverse outputs on typical underspecified prompts? How can we automatically measure diversity? We propose **GRADE**: **Gr**anular **A**ttribute **D**iversity **E**valuation, an automatic method for quantifying sample diversity. GRADE leverages the world knowledge embedded in large language models and visual question-answering systems to identify relevant concept-specific axes of diversity (e.g., ''shape'' and ''color'' for the concept ''cookie''). It then estimates attribute distributions and quantifies diversity using (normalized) entropy. GRADE achieves over 90\\% human agreement while exhibiting weak correlation to commonly used diversity metrics. We use GRADE to measure the overall diversity of 12 T2I models using 400 concept-attribute pairs, revealing that even the most diverse models display limited variation. Further, we find these models often exhibit *default behaviors*, a situation where the model consistently generates concepts with the same attributes (e.g., 98\\% of the cookies are round). Finally, we demonstrate that a key reason for low diversity is due to underspecified captions in training data.", "keywords": [ "sample diversity", "text-to-image", "diffusion", "evaluation" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "Jd3Vd7GCyq", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Fionn Mc Inerney", "gender": "unknown", "institution": "Telefonica Scientific Research", "country": "ES", "position": "Researcher" }, { "name": "Liana Khazaliya", "gender": "Female", "institution": "Technische Universit\u00e4t Wien", "country": "AT", "position": "PhD student" }, { "name": "Mathis Rocton", "gender": "Male", "institution": "Technische Universit\u00e4t Wien", "country": "", "position": "PhD student" }, { "name": "Robert Ganian", "gender": "Male", "institution": "TU Wien Vienna University of Technology", "country": "AT", "position": "Associate Professor" } ], "n_formula": 95, "n_formula_1": 10, "n_ref_uni": 0, "n_ref": 165, "n_ref_all": 189, "n_fig": 3, "n_tab": 3, "L_tab": 425, "n_element_tab": 2, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 11955, "formula_len_all_1": 1134, "len_all": 301329, "len_all_1": 71585, "len_abs": 2636, "len_title": 120, "len_sents": 95420, "len_sents_1": 34495, "n_sents": 691, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1115, "title": "The Computational Complexity of Positive Non-Clashing Teaching in Graphs", "abs": "We study the classical and parameterized complexity of computing the positive non-clashing teaching dimension of a set of concepts, that is, the smallest number of examples per concept required to successfully teach an intelligent learner under the considered, previously established model. For any class of concepts, it is known that this problem can be effortlessly transferred to the setting of balls in a graph $G$. We establish (1) the NP-hardness of the problem even when restricted to instances with positive non-clashing teaching dimension $k=2$ and where all balls in the graph are present, (2) near-tight running time upper and lower bounds for the problem on general graphs, (3) fixed-parameter tractability when parameterized by the vertex integrity of $G$, and (4) a lower bound excluding fixed-parameter tractability when parameterized by the feedback vertex number and pathwidth of $G$, even when combined with $k$.\nOur results provide a nearly complete understanding of the complexity landscape of computing the positive non-clashing teaching dimension and answer open questions from the literature.", "keywords": [ "non-clashing teaching", "positive teaching", "computational complexity", "NP-hardness", "parameterized complexity" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 4, 4, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "JbRM5QKRDd", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Bin Chen", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Shu-Tao Xia", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xiang Liu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yaowei Wang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Zimo Liu", "gender": "unknown", "institution": "PengCheng Lab", "country": "CN", "position": "Postdoc" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 45, "n_ref": 97, "n_ref_all": 115, "n_fig": 11, "n_tab": 6, "L_tab": 3821, "n_element_tab": 493, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 2154, "n_element_tab_1": 317, "formula_len_all": 421, "formula_len_all_1": 421, "len_all": 142594, "len_all_1": 61012, "len_abs": 1434, "len_title": 125, "len_sents": 35362, "len_sents_1": 24853, "n_sents": 311, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1447, "title": "An Exploration with Entropy Constrained 3D Gaussians for 2D Video Compression", "abs": "3D Gaussian Splatting (3DGS) has witnessed its rapid development in novel view synthesis, which attains high quality reconstruction and real-time rendering. At the same time, there is still a gap before implicit neural representation (INR) can become a practical compressor due to the lack of stream decoding and real-time frame reconstruction on consumer-grade hardware. It remains a question whether the fast rendering and partial parameter decoding characteristics of 3DGS are applicable to video compression. To address these challenges, we propose a Toast-like Sliding Window (TSW) orthographic projection for converting any 3D Gaussian model into a video representation model. This method efficiently represents video by leveraging temporal redundancy through a sliding window approach. Additionally, the converted model is inherently stream-decodable and offers a higher rendering frame rate compared to INR methods. Building on TSW, we introduce an end-to-end trainable video compression method, GSVC, which employs deformable Gaussian representation and optical flow guidance to capture dynamic content in videos. Experimental results demonstrate that our method effectively transforms a 3D Gaussian model into a practical video compressor. GSVC further achieves better rate-distortion performance than NeRV on the UVG dataset, while achieving higher frame reconstruction speed (+30%~40% fps) and stream decoding. Code will be released.", "keywords": [ "Video Compression", "3D Gaussian Splatting", "Entropy Coding" ], "rating_list": [ 6, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "JbPb6RieNC", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Haomiao Xiong", "gender": "unknown", "institution": "Dalian University of Technology", "country": "CN", "position": "MS student" }, { "name": "Huchuan Lu", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Professor" }, { "name": "Jiawen Zhu", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Lu Zhang", "gender": "unknown", "institution": "Dalian University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yunzhi Zhuge", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Postdoc" }, { "name": "Zongxin Yang", "gender": "Male", "institution": "Harvard Medical School, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "JiazuoYu", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 20, "n_ref_uni": 48, "n_ref": 98, "n_ref_all": 119, "n_fig": 13, "n_tab": 11, "L_tab": 5162, "n_element_tab": 792, "n_fig_1": 12, "n_tab_1": 7, "L_tab_1": 3084, "n_element_tab_1": 444, "formula_len_all": 3157, "formula_len_all_1": 3233, "len_all": 226826, "len_all_1": 114453, "len_abs": 1621, "len_title": 136, "len_sents": 65698, "len_sents_1": 45972, "n_sents": 403, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1332, "title": "Streaming Video Understanding and Multi-round Interaction with Memory-enhanced Knowledge", "abs": "Recent advances in Large Language Models (LLMs) have enabled the development of Video-LLMs, advancing multimodal learning by bridging video data with language tasks. However, current video understanding models struggle with processing long video sequences, supporting multi-turn dialogues, and adapting to real-world dynamic scenarios. To address these issues, we propose StreamChat, a training-free framework for streaming video reasoning and conversational interaction. StreamChat leverages a novel hierarchical memory system to efficiently process and compress video features over extended sequences, enabling real-time, multi-turn dialogue. Our framework incorporates a parallel system scheduling strategy that enhances processing speed and reduces latency, ensuring robust performance in real-world applications. Furthermore, we introduce StreamBench, a versatile benchmark that evaluates streaming video understanding across diverse media types and interactive scenarios, including multi-turn interactions and complex reasoning tasks. Extensive evaluations on StreamBench and other public benchmarks demonstrate that StreamChat significantly outperforms existing\nstate-of-the-art models in terms of accuracy and response times, confirming its effectiveness for streaming video understanding. Code is available at StreamChat.", "keywords": [ "Streaming Video Understanding; Video MLLM; Hierarchical Memory System" ], "rating_list": [ 3, 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 3, 3, 4 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "Jb1XGe4ioQ", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Ganchao Wei", "gender": "Male", "institution": "Duke University", "country": "", "position": "Postdoc" }, { "name": "Li Ma", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Full Professor" } ], "n_formula": 21, "n_formula_1": 12, "n_ref_uni": 27, "n_ref": 59, "n_ref_all": 92, "n_fig": 26, "n_tab": 4, "L_tab": 441, "n_element_tab": 28, "n_fig_1": 16, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2412, "formula_len_all_1": 1070, "len_all": 129518, "len_all_1": 64835, "len_abs": 1654, "len_title": 90, "len_sents": 47606, "len_sents_1": 31991, "n_sents": 367, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1660, "title": "Stream-level flow matching from a Bayesian decision theoretic perspective", "abs": "Flow matching (FM) is a family of training algorithms for fitting continuous normalizing flows (CNFs). Conditional flow matching (CFM) exploits the fact that the marginal vector field of a CNF can be learned by fitting least-square regression to the so-called conditional vector field specified given one or both ends of the flow path. We show that viewing CFM training from a Bayesian decision theoretic perspective on parameter estimation opens the door to generalizations of CFM algorithms. We propose one such extension by introducing a CFM algorithm based on defining conditional probability paths given what we refer to as \"streams,\" instances of latent stochastic paths that connect pairs of noise and observed data. Further, we advocate the modeling of these latent streams using Gaussian processes (GPs). The unique distributional properties of GPs, and in particular the fact that the velocity of a GP is still a GP, allows drawing samples from the resulting stream-augmented conditional probability path without simulating the actual streams, and hence the \"simulation-free\" nature of CFM training is preserved. We show that this generalization of the CFM can substantially reduce the variance in the estimated marginal vector field at a moderate computational cost, thereby improving the quality of the generated samples under common metrics. Additionally, we show that adopting the GP on the streams allows for flexibly linking multiple related training data points (e.g., time series) and incorporating additional prior information. We empirically validate our claim through both simulations and applications to image and neural time series data.", "keywords": [ "generative model", "flow matching", "Bayesian decision theory", "Gaussian process" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "JapY2HtNI5", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Dominika Matus", "gender": "Female", "institution": "Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "MS student" }, { "name": "Frank Hutter", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen & University of Freiburg", "country": "DE", "position": "Full Professor" }, { "name": "Frederic Runge", "gender": "unknown", "institution": "Universit\u00e4t Freiburg", "country": "", "position": "PhD student" }, { "name": "Rolf Backofen", "gender": "Male", "institution": "Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "", "position": "Full Professor" }, { "name": "Ryan Koksal", "gender": "unknown", "institution": "Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "", "position": "MS student" }, { "name": "J\u00f6rg Franke", "gender": "Male", "institution": "Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 64, "n_ref": 158, "n_ref_all": 179, "n_fig": 6, "n_tab": 4, "L_tab": 1194, "n_element_tab": 165, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2176, "n_element_tab_1": 133, "formula_len_all": 401, "formula_len_all_1": 401, "len_all": 170454, "len_all_1": 66374, "len_abs": 1184, "len_title": 130, "len_sents": 45063, "len_sents_1": 31299, "n_sents": 275, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1189, "title": "RNAformer: Axial-Attention For Homology-Aware RNA Secondary Structure Prediction", "abs": "Predicting RNA secondary structure is essential for understanding RNA function and developing RNA-based therapeutics. Despite recent advances in deep learning for structural biology, its application to RNA secondary structure prediction remains contentious. A primary concern is the control of homology between training and test data. Moreover, deep learning approaches often incorporate complex multi-model systems, ensemble strategies, or require external data. Here, we present the RNAformer, a scalable axial-attention-based deep learning model designed to predict secondary structure directly from a single RNA sequence without additional requirements. We demonstrate the benefits of this lean architecture by learning an accurate biophysical RNA folding model using synthetic data. Trained on experimental data, our model overcomes previously reported caveats in deep learning approaches with a novel homology-aware data pipeline. The RNAformer achieves state-of-the-art performance on RNA secondary structure prediction, outperforming both traditional non-learning-based methods and existing deep learning approaches, while carefully considering sequence and structure similarities.", "keywords": [ "Structural Biology", "Deep Learning", "RNA Secondary Structure Prediction", "Axial Attention" ], "rating_list": [ 5, 6, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 4, 3 ] }, { "paper_id": "JaTmg8FX3k", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Linfeng Cao", "gender": "Male", "institution": "The Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Ming Shi", "gender": "unknown", "institution": "State University of New York at Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Ness Shroff", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Full Professor" } ], "n_formula": 208, "n_formula_1": 22, "n_ref_uni": 19, "n_ref": 60, "n_ref_all": 80, "n_fig": 11, "n_tab": 8, "L_tab": 1129, "n_element_tab": 114, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 376, "n_element_tab_1": 31, "formula_len_all": 23505, "formula_len_all_1": 3709, "len_all": 303154, "len_all_1": 83631, "len_abs": 1517, "len_title": 140, "len_sents": 106151, "len_sents_1": 36923, "n_sents": 1002, "n_sents_1": 294, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1533, "title": "Provably Efficient Multi-Objective Bandit Algorithms under Preference-Centric Customization", "abs": "Existing multi-objective multi-armed bandit (MO-MAB) approaches mainly focus on achieving Pareto optimality. However, a Pareto optimal arm that receives a high score from one user may lead to a low score from another, since in real-world scenarios, users often have diverse preferences across different objectives. Instead, these preferences should inform *customized learning*, a factor usually neglected in prior research. To address this need, we study a *preference-aware* MO-MAB framework in the presence of explicit user preferences, where each user\u2019s overall-reward is modeled as the inner product of user preference and arm reward. This new framework shifts the focus from merely achieving Pareto optimality to further optimizing within the Pareto front under preference-centric customization. To the best of our knowledge, this is the first theoretical exploration of customized MO-MAB optimization based on explicit user preferences. This framework introduces new and unique challenges for algorithm design for customized optimization. To\naddress these challenges, we incorporate *preference estimation* and *preference-aware optimization* as key mechanisms for preference adaptation, and develop new analytical techniques to rigorously account for the impact of preference estimation errors on overall performance. Under this framework, we consider three preference structures inspired by practical applications, with tailored algorithms that are proven to achieve near-optimal regret, and show good numerical performance.", "keywords": [ "multi-objective multi-arm bandit", "bandit optimization", "preference-centric learning" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "JaRihIHbZm", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Abhranil Chandra", "gender": "unknown", "institution": "University of Waterloo", "country": "", "position": "MS student" }, { "name": "Achint Soni", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "MS student" }, { "name": "Bo Dai", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Percy Liang", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Sebastian Fischmeister", "gender": "Not Specified", "institution": "University of Waterloo", "country": "CA", "position": "Full Professor" }, { "name": "Mengjiao Yang", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "Student" }, { "name": "Sreyas Venkataraman", "gender": "Male", "institution": "Indian Institute of Technology Kharagpur, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "Undergrad student" } ], "n_formula": 25, "n_formula_1": 21, "n_ref_uni": 62, "n_ref": 112, "n_ref_all": 143, "n_fig": 6, "n_tab": 9, "L_tab": 3188, "n_element_tab": 428, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1004, "n_element_tab_1": 135, "formula_len_all": 1519, "formula_len_all_1": 1348, "len_all": 195272, "len_all_1": 69019, "len_abs": 1737, "len_title": 91, "len_sents": 50693, "len_sents_1": 31026, "n_sents": 374, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1592, "title": "VideoAgent: Self-Improving Video Generation", "abs": "Video generation has been used to generate visual plans for controlling robotic systems. Given an image observation and a language instruction, previous work has generated video plans which are then converted to robot controls to be executed. However, a major bottleneck in leveraging video generation for control lies in the quality of the generated videos, which often suffer from hallucinatory content and unrealistic physics, resulting in low task success when control actions are extracted from the generated videos. While scaling up dataset and model size provides a partial solution, integrating external feedback is both natural and essential for grounding video generation in the real world. With this observation, we propose VideoAgent for self-improving generated video plans based on external feedback. Instead of directly executing the generated video plan, VideoAgent first refines the generated video plans using a novel procedure which we call self-conditioning consistency, utilizing feedback from a pretrained vision-language model (VLM). As the refined video plan is being executed, VideoAgent collects additional data from the environment to further improve video plan generation. Experiments in simulated robotic manipulation from MetaWorld and iTHOR show that VideoAgent drastically reduces hallucination, thereby boosting success rate of downstream manipulation tasks. We further illustrate that VideoAgent can effectively refine real-robot videos, providing an early indicator that robotics can be an effective tool in grounding video generation in the physical world.", "keywords": [ "sequential decision making", "video generation", "self improvement" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "JZjW3k4Kyc", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Chiyu Ma", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "PhD student" }, { "name": "Jiaqi Gan", "gender": "Female", "institution": "Dartmouth College", "country": "", "position": "MS student" }, { "name": "Lin Shi", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "Undergrad student" }, { "name": "Ming Cheng", "gender": "unknown", "institution": "Dartmouth College", "country": "US", "position": "MS student" }, { "name": "Ollie Liu", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD" }, { "name": "Soroush Vosoughi", "gender": "unknown", "institution": "Dartmouth College", "country": "", "position": "Assistant Professor" }, { "name": "Wenhua Liang", "gender": "unknown", "institution": "Dartmouth College", "country": "US", "position": "Undergrad student" }, { "name": "Willie Neiswanger", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 69, "n_ref_all": 93, "n_fig": 8, "n_tab": 3, "L_tab": 7924, "n_element_tab": 548, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 28, "n_element_tab_1": 1, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 139681, "len_all_1": 63937, "len_abs": 1455, "len_title": 133, "len_sents": 46097, "len_sents_1": 33753, "n_sents": 328, "n_sents_1": 234, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1459, "title": "Mechanistic Insights: Circuit Transformations Across Input and Fine-Tuning Landscapes", "abs": "Mechanistic interpretability seeks to uncover the internal mechanisms of Large Language Models (LLMs) by identifying circuits\u2014subgraphs in the model\u2019s computational graph that correspond to specific behaviors\u2014while ensuring sparsity and maintaining task performance. Although automated methods have made massive circuit discovery feasible, determining the functionalities of circuit components still requires manual effort, limiting scalability and efficiency. To address this, we propose a novel framework that accelerates circuit discovery and analysis. Building on methods like edge pruning, our framework introduces circuit selection, comparison, attention grouping, and logit clustering to investigate the intended functionalities of circuit components. By focusing on what components aim to achieve, rather than their direct causal effects, this framework streamlines the process of understanding interpretability, reduces manual labor, and scales the analysis of model behaviors across various tasks. Inspired by observing circuit variations when models are fine-tuned or prompts are tweaked (while maintaining the same task type), we apply our framework to explore these variations across four PEFT methods and full fine-tuning on two well-known tasks. Our results suggest that while fine-tuning generally preserves the structure of the mechanism for solving tasks, individual circuit components may not retain their original intended functionalities.", "keywords": [ "Mechanistic Interpretability", "PEFT", "circuit" ], "rating_list": [ 1, 8, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 4, 2, 3 ], "contribution_list": [ 1, 3, 2, 1 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "JZgqoOu4Ml", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Julian Leander M\u00f6bius", "gender": "Male", "institution": "Friedrich-Schiller Universit\u00e4t Jena", "country": "DE", "position": "PhD student" }, { "name": "Michael Habeck", "gender": "unknown", "institution": "University of Jena", "country": "DE", "position": "Professor" } ], "n_formula": 24, "n_formula_1": 21, "n_ref_uni": 34, "n_ref": 76, "n_ref_all": 96, "n_fig": 8, "n_tab": 2, "L_tab": 203, "n_element_tab": 1, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1528, "n_element_tab_1": 17, "formula_len_all": 1470, "formula_len_all_1": 1450, "len_all": 124070, "len_all_1": 69355, "len_abs": 1272, "len_title": 125, "len_sents": 34471, "len_sents_1": 29629, "n_sents": 248, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1277, "title": "Diffusion priors for Bayesian 3D reconstruction from incomplete measurements", "abs": "Many inverse problems are ill-posed and need to be complemented by prior information that restricts the class of admissible models. Bayesian approaches encode this information as prior distributions that impose generic properties on the model such as sparsity, non-negativity or smoothness. However, in case of complex structured models such as images, graphs or three-dimensional (3D) objects, generic prior distributions tend to favor models that differ largely from those observed in the real world. Here we explore the use of diffusion models as priors that are combined with experimental data within a Bayesian framework. We use 3D point clouds to represent 3D objects such as household items or biomolecular complexes formed from proteins and nucleic acids. We train diffusion models that generate coarse-grained 3D structures at a medium resolution and integrate these with incomplete and noisy experimental data. To demonstrate the power of our approach, we focus on the reconstruction of biomolecular assemblies from cryo-electron microscopy (cryo-EM) images, which is an important inverse problem in structural biology. We find that posterior sampling with diffusion model priors allows for 3D reconstruction from very sparse, low-resolution and partial observations.", "keywords": [ "Diffusion model guidance", "3D reconstruction", "cryo-electron microscopy", "inverse problems", "diffusion posterior sampling" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "JZdd7EUefP", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "He Wang", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Associate Professor" }, { "name": "Zhanxing Zhu", "gender": "Male", "institution": "University of Southampton", "country": "", "position": "Associate Professor" }, { "name": "Zheng Wang", "gender": "Male", "institution": "University of Leeds", "country": "GB", "position": "Full Professor" }, { "name": "Bochen Lv", "gender": "Male", "institution": "University of Southampton", "country": "GB", "position": "PhD student" } ], "n_formula": 94, "n_formula_1": 39, "n_ref_uni": 18, "n_ref": 51, "n_ref_all": 109, "n_fig": 8, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7464, "formula_len_all_1": 2698, "len_all": 151411, "len_all_1": 70093, "len_abs": 1249, "len_title": 127, "len_sents": 41160, "len_sents_1": 28829, "n_sents": 431, "n_sents_1": 238, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1255, "title": "Continuous Approximation of Momentum Methods with Explicit Discretization Error", "abs": "Momentum-based optimization methods, such as Heavy-Ball (HB) and Nesterov's accelerated gradient (NAG), are essential in training modern deep neural networks. This work sheds light on the learning dynamics of momentum-based methods and how they behave differently than standard gradient descent (GD) in theory and practice. A promising approach to answer this question is \ninvestigating the continuous differential equations to approximate the discrete updates, \nan area requiring much attention for momentum methods. In this work, we take HB as a case study to investigate two important aspects of momentum methods. First, to enable a formal analysis of the Heavy-Ball momentum method, we propose a new continuous approximation, HB Flow (HBF), with a formulation that allows the control of discretization error to arbitrary order.\nAs an application of HBF, we leverage it to investigate the implicit bias of HB by conducting a series of analyses on the diagonal linear networks to inspect the influence of momentum on the model's generalization property. We validate theoretical findings in numerical experiments, which confirm the significance of HBF as an effective proxy of momentum methods to bridge between discrete and continuous learning dynamics.", "keywords": [ "momentum methods", "continuous approximation", "discretization error", "implicit bias" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "JZOPwrRYtI", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 10, "author_info_list": [ { "name": "Jiawei Zhang", "gender": "unknown", "institution": "University of California, Davis", "country": "US", "position": "Associate Professor" }, { "name": "Mingyang Zhang", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Siwei Zhang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Tengfei LIU", "gender": "unknown", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Weiqiang Wang", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Xi Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Xixi Wu", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yao Zhang", "gender": "Male", "institution": "eBay Inc.", "country": "CN", "position": "Researcher" }, { "name": "Yizhu Jiao", "gender": "Female", "institution": "UIUC", "country": "", "position": "PhD student" }, { "name": "Yun Xiong", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 30, "n_ref": 77, "n_ref_all": 125, "n_fig": 8, "n_tab": 15, "L_tab": 5659, "n_element_tab": 643, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 907, "formula_len_all_1": 587, "len_all": 236365, "len_all_1": 67064, "len_abs": 4278, "len_title": 150, "len_sents": 75722, "len_sents_1": 31486, "n_sents": 519, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1947, "title": "Interactions Exhibit Clustering Rhythm: A Prevalent Observation for Advancing Temporal Link Prediction", "abs": "Temporal link prediction aims to forecast future link existence in temporal graphs, with numerous real-world applications. Existing methods often rely on designing complex model architectures to parameterize the interaction patterns between nodes. Instead, we re-think the interaction dynamics in temporal graphs (which we call ``interaction rhythms'') by addressing a fundamental research question: \\textit{Is there a strong yet prevalent latent interaction rhythm pattern across different temporal graphs that can be leveraged for temporal link prediction?} \nOur introduced empirical analyses reveal that there indeed exists temporal clustering in node interaction rhythms, where for a specific node, interactions tend to occur in bursts. Such observation leads to two key insights for predicting future links: (i) recent historical links that carry the latest rhythm pattern information; and (ii) the inter-event times that further illuminate temporal dynamics. \nBuilding on these empirical findings, we propose TG-Mixer, a novel method that explicitly captures temporal clustering patterns to advance temporal link prediction. \nTG-Mixer samples the most recent historical links to extract surrounding neighborhoods, preserving currently invaluable interaction rhythms while avoiding massive computations. \nAdditionally, it integrates a carefully designed silence decay mechanism that penalizes nodes' long-term inactivity, effectively incorporating temporal clustering information for future link prediction.\nBoth components ensure concise implementations, leading to a lightweight architecture. \nExhaustive experiments on seven benchmarks against nine baselines demonstrate that TG-Mixer achieves state-of-the-art performance with faster convergence, stronger generalization capabilities, and higher efficiency. The experimental results also highlight the importance of explicitly considering temporal clustering for temporal link prediction.", "keywords": [ "Link Prediction", "Temporal Graphs", "Data Mining" ], "rating_list": [ 5, 5, 6, 3, 6 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 1, 3, 3, 2, 4 ], "contribution_list": [ 2, 2, 3, 2, 4 ], "confidence_list": [ 3, 4, 3, 4, 3 ] }, { "paper_id": "JZLon6cvx8", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Jaskirat Singh", "gender": "unknown", "institution": "Australian National University", "country": "AU", "position": "PhD student" }, { "name": "Jonas K Kohler", "gender": "Male", "institution": "Facebook", "country": "CH", "position": "Researcher" }, { "name": "Junshen K Chen", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Machine Learning Engineer" }, { "name": "Michael F Cohen", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 27, "n_ref": 74, "n_ref_all": 100, "n_fig": 8, "n_tab": 1, "L_tab": 538, "n_element_tab": 56, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 543, "formula_len_all_1": 518, "len_all": 106181, "len_all_1": 60017, "len_abs": 1638, "len_title": 133, "len_sents": 29827, "len_sents_1": 26725, "n_sents": 214, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1677, "title": "Storybooth: Training-Free Multi-Subject Consistency for Improved Visual Storytelling", "abs": "Consistent text-to-image generation depicting the *same* subjects across different images has gained significant recent attention due to its widespread applications in the fields of visual-storytelling and multiple-shot video generation. While remarkable, existing methods often require costly finetuning for each subject and struggle to maintain consistency across multiple characters. In this work, we first analyse the reason for these limitations. Our exploration reveals that the primary-issue stems from *self-attention leakage*, which is exacerbated when trying to ensure consistency across multiple-characters. Motivated by these findings, we next propose a simple yet effective *training and optimization-free approach* for improving multiple-character consistency. In particular, we first leverage multi-modal *chain-of-thought* reasoning in order to *apriori* localize the different subjects across the storyboard frames. The final storyboard images are then generated using a modified diffusion model which includes *1) a bounded cross-attention layer* for ensuring adherence to the initially predicted layout, and *2) a bounded cross-frame self-attention layer* for reducing inter-character attention leakage. Furthermore, we also propose a novel *cross-frame token-merging layer* which allows for improved fine-grain consistency for the storyboard characters. \n Experimental analysis reveals that proposed approach is not only $\\times 30$ faster than prior training-based methods (*eg, textual inversion, dreambooth-lora*) but also surpasses the prior *state-of-the-art*, exhibiting improved multi-character consistency and text-to-image alignment performance.", "keywords": [ "consistent text-to-image generation", "visual storytelling", "story generation" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "JZCxlrwjZ8", "primary_area": "reinforcement learning", "n_author": 9, "author_info_list": [ { "name": "Chengxing Jia", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Haoxin Lin", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Jiaji Zhang", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "MS student" }, { "name": "Junyin Ye", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "MS student" }, { "name": "Yi-Chen Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Yihao Sun", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "MS student" }, { "name": "Yu-Yan Xu", "gender": "Female", "institution": "Polixir", "country": "", "position": "Researcher" }, { "name": "Zhilong Zhang", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "MS student" }, { "name": "Yang Yu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Professor" } ], "n_formula": 24, "n_formula_1": 17, "n_ref_uni": 49, "n_ref": 159, "n_ref_all": 182, "n_fig": 7, "n_tab": 9, "L_tab": 4754, "n_element_tab": 577, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1487, "n_element_tab_1": 237, "formula_len_all": 2866, "formula_len_all_1": 1798, "len_all": 174325, "len_all_1": 64847, "len_abs": 1141, "len_title": 146, "len_sents": 48912, "len_sents_1": 28622, "n_sents": 360, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1147, "title": "Any-step Dynamics Model Improves Future Predictions for Online and Offline Reinforcement Learning", "abs": "Model-based methods in reinforcement learning offer a promising approach to enhance data efficiency by facilitating policy exploration within a dynamics model. However, accurately predicting sequential steps in the dynamics model remains a challenge due to the bootstrapping prediction, which attributes the next state to the prediction of the current state. This leads to accumulated errors during model roll-out. In this paper, we propose the Any-step Dynamics Model (ADM) to mitigate the compounding error by reducing bootstrapping prediction to direct prediction. ADM allows for the use of variable-length plans as inputs for predicting future states without frequent bootstrapping. We design two algorithms, ADMPO-ON and ADMPO-OFF, which apply ADM in online and offline model-based frameworks, respectively. In the online setting, ADMPO-ON demonstrates improved sample efficiency compared to previous state-of-the-art methods. In the offline setting, ADMPO-OFF not only demonstrates superior performance compared to recent state-of-the-art offline approaches but also offers better quantification of model uncertainty using only a single ADM.", "keywords": [ "model-based reinforcement learning", "any-step dynamics model" ], "rating_list": [ 5, 5, 6, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "JYwVijuNA7", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Ashley Prater-Bennette", "gender": "Female", "institution": "Air Force Research Laboratory", "country": "", "position": "Principal Researcher" }, { "name": "Lixin Shen", "gender": "Male", "institution": "Syracuse University", "country": "US", "position": "Full Professor" }, { "name": "Qi Zhang", "gender": "unknown", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Shaofeng Zou", "gender": "unknown", "institution": "Arizona State University", "country": "", "position": "Associate Professor" }, { "name": "Simon Khan", "gender": "Male", "institution": "Air Force Research Laboratory ", "country": "US", "position": "Principal Researcher" }, { "name": "Yi Zhou", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Associate Professor" } ], "n_formula": 144, "n_formula_1": 43, "n_ref_uni": 31, "n_ref": 92, "n_ref_all": 104, "n_fig": 3, "n_tab": 1, "L_tab": 116, "n_element_tab": 18, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 134, "n_element_tab_1": 30, "formula_len_all": 20840, "formula_len_all_1": 3940, "len_all": 153177, "len_all_1": 66587, "len_abs": 1604, "len_title": 154, "len_sents": 39672, "len_sents_1": 27131, "n_sents": 393, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1728, "title": "Revisiting Large-Scale Non-convex Distributionally Robust Optimization", "abs": "Distributionally robust optimization (DRO) is a powerful technique to train robust machine learning models that perform well under distribution shifts. Compared with empirical risk minimization (ERM), DRO optimizes the expected loss under the worst-case distribution in\nan uncertainty set of distributions. This paper revisits the important problem of DRO with non-convex smooth loss functions. For this problem, Jin et al. (2021) showed that its dual problem is generalized $(L_0, L_1)$-smooth condition and gradient noise satisfies the affine variance condition, designed an algorithm of mini-batch normalized gradient descent with momentum, and proved its convergence and complexity. In this paper, we show that the dual problem and the gradient noise satisfy simpler yet more precise partially generalized smoothness condition and partially affine variance condition by studying the optimization variable and dual variable separately, which further yields much simpler algorithm design and convergence analysis. We develop a double stochastic gradient descent with clipping (D-SGD-C) algorithm that converges to an $\\epsilon$-stationary point with $\\mathcal O(\\epsilon^{-4})$ gradient complexity, which matches with results in Jin et al. (2021). Our algorithm does not need to use momentum, and the proof is much simpler, thanks to the more precise characterization of partially generalized smoothness and partially affine variance noise. We further design a variance-reduced method that achieves a lower gradient complexity of $\\mathcal O(\\epsilon^{-3})$. Our theoretical results and insights are further verified numerically on a number of tasks, and our algorithms outperform the existing DRO method (Jin et al., 2021).", "keywords": [ "distributionally robust optimization", "generalized smoothness", "non-convex optimization", "variance-reduced method" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "JYV2hrtFSv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Idan Szpektor", "gender": "unknown", "institution": "Google", "country": "IL", "position": "Researcher" }, { "name": "Orr Zohar", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Xiaohan Wang", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Yonatan Bitton", "gender": "Male", "institution": "Google", "country": "IL", "position": "Research Scientist" }, { "name": "Serena Yeung", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 4, "n_ref_uni": 36, "n_ref": 67, "n_ref_all": 82, "n_fig": 7, "n_tab": 7, "L_tab": 2324, "n_element_tab": 308, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 2185, "n_element_tab_1": 273, "formula_len_all": 55, "formula_len_all_1": 107, "len_all": 168756, "len_all_1": 71817, "len_abs": 1791, "len_title": 128, "len_sents": 37400, "len_sents_1": 31211, "n_sents": 300, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1943, "title": "Video-STaR: Self-Training Enables Video Instruction Tuning with Any Supervision", "abs": "The performance and reasoning capabilities of Large Multi-modal Models (LMMs) is dependent on the size and quality of their training datasets. However, collecting datasets that support chain-of-thought instruction tuning is highly challenging. Existing video instruction tuning datasets are often derived by prompting large language models with video captions to generate question-answer pairs, which makes them predominantly descriptive rather than reasoning-focused. \nMeanwhile, many labeled video datasets with diverse labels and supervision exist -- however, we find that their integration into LMMs is non-trivial. \nHerein, we present $\\underline{\\text{Video}}$ $\\underline{\\text{S}}\\text{elf}$-$\\underline{\\text{T}}\\text{raining}$ $\\text{with}$ $\\underline{\\text{a}}\\text{ugmented}$ $\\underline{\\text{R}}\\text{easoning}$ (Video-STaR), the first self-training approach for video instruction tuning. \nVideo-STaR allows the utilization of *any* labeled video dataset for video instruction tuning.\nIn Video-STaR, an LMM cycles between instruction generation and finetuning, which we show (I) improves general video understanding and (II) adapts LMMs to novel downstream tasks with existing supervision. \nDuring instruction generation, an LMM is prompted to propose an answer. The answers are then filtered only to those that contain the original video labels, and the LMM is then re-trained on the generated dataset. \nBy training exclusively on generated answers containing the correct video labels, Video-STaR leverages these existing labels as weak supervision for video instruction tuning.\nOur results demonstrate that Video-STaR-augmented LMMs achieve notable improvements in (I) general Video QA, where TempCompass performance improved by 6.1%, *and* (II) downstream tasks, with a 9.9% increase in Kinetics700-QA accuracy and a 4.0% improvement in action quality assessment on FineDiving, while also exhibiting better interpretability.", "keywords": [ "Video Understanding", "Visual Instruction Tuning", "Self-Training", "Chain-of-thought reasoning" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "JYTQ6ELUVO", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Alexander Shen", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Ameet Talwalkar", "gender": "Male", "institution": "Carnegie-Mellon University", "country": "", "position": "" }, { "name": "Junhong Shen", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Mikhail Khodak", "gender": "Not Specified", "institution": "Princeton University", "country": "US", "position": "Postdoc" }, { "name": "Ritvik Gupta", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Wenduo Cheng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zongzhe Xu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 47, "n_ref": 108, "n_ref_all": 129, "n_fig": 5, "n_tab": 10, "L_tab": 4501, "n_element_tab": 570, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 3043, "n_element_tab_1": 422, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 171019, "len_all_1": 64096, "len_abs": 1147, "len_title": 115, "len_sents": 44111, "len_sents_1": 28859, "n_sents": 241, "n_sents_1": 148, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1149, "title": "Specialized Foundation Models struggle to beat Supervised Baselines", "abs": "Following its success for vision and text, the \u201cfoundation model\u201d (FM) paradigm\u2014pretraining large models on massive data, then fine-tuning on target tasks\u2014has rapidly expanded to domains in the sciences, engineering, healthcare, and beyond. Has this achieved what the original FMs accomplished, i.e. the supplanting of traditional supervised learning in their domains? To answer we look at three modalities\u2014genomics, satellite data, and time series\u2014with multiple recent FMs and compare them to a standard supervised learning workflow: model development, hyperparameter tuning, and training, all using only data from the target task. Across those three specialized domains, we find that it is consistently possible to train simple supervised models\u2014no more complicated than a lightly modified wide ResNet or UNet\u2014that match or even outperform the latest foundation models. Our work demonstrates that the benefits of large-scale pretraining have yet to be realized in many specialized areas, reinforces the need to compare new FMs to strong, well-tuned baselines, and introduces two new, easy-to-use, open-source, and automated workflows for doing so.", "keywords": [ "foundation models", "supervised learning", "neural architecture search", "hyperparameter optimization", "convolutional networks", "autoregressive models", "genomics", "satellite imaging", "time series" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "JY6P45sFDS", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Bernhard Schoelkopf", "gender": "Male", "institution": "ELLIS Institute", "country": "DE", "position": "Principal Researcher" }, { "name": "Bobby He", "gender": "unknown", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Sidak Pal Singh", "gender": "unknown", "institution": "Swiss Federal Institute of Technology Zurich", "country": "", "position": "PhD student" }, { "name": "Thomas Hofmann", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" } ], "n_formula": 23, "n_formula_1": 3, "n_ref_uni": 40, "n_ref": 51, "n_ref_all": 92, "n_fig": 19, "n_tab": 2, "L_tab": 233, "n_element_tab": 15, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 19, "n_element_tab_1": 1, "formula_len_all": 1660, "formula_len_all_1": 192, "len_all": 183544, "len_all_1": 68433, "len_abs": 1477, "len_title": 115, "len_sents": 61394, "len_sents_1": 34383, "n_sents": 441, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1490, "title": "The Directionality of Optimization Trajectories in Neural Networks", "abs": "The regularity or implicit bias in neural network optimization has been typically studied via the parameter norms or the landscape curvature, often overlooking the trajectory leading to these parameters. However, properties of the trajectory --- particularly its directionality --- capture critical aspects of how gradient descent navigates the landscape to converge to a solution. In this work, we introduce the notion of a Trajectory Map and derive natural complexity measures that highlight the directional characteristics of optimization trajectories. Our comprehensive analysis across vision and language modeling tasks reveals that (a) the trajectory's directionality at the macro-level saturates by the initial phase of training, wherein weight decay and momentum play a crucial but understated role; and (b) in subsequent training, trajectory directionality manifests in micro-level behaviors, such as oscillations, for which we also provide a theoretical analysis. This implies that neural optimization trajectories have, overall, a more linear form than zig-zaggy, as evident by high directional similarity, especially towards the end. To further hone this point, we show that when the trajectory direction gathers such an inertia, optimization proceeds largely unaltered even if the network is severely decapacitated (by freezing >99% of the parameters), --- thereby demonstrating the potential for significant computational and resource savings without compromising performance.", "keywords": [ "optimization", "trajectory", "redundancy", "LLMs", "neural networks" ], "rating_list": [ 3, 8, 6, 10 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 2, 3, 2, 4 ], "confidence_list": [ 2, 4, 3, 2 ] }, { "paper_id": "JXvEzl8YkS", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Edward Selig", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "MS student" }, { "name": "Paul Alexander Bilokon", "gender": "unknown", "institution": "Thalesians Marine Ltd, the AI in shipping and logistics company", "country": "", "position": "Principal Researcher" } ], "n_formula": 35, "n_formula_1": 31, "n_ref_uni": 9, "n_ref": 24, "n_ref_all": 50, "n_fig": 0, "n_tab": 4, "L_tab": 2619, "n_element_tab": 560, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 161, "n_element_tab_1": 3, "formula_len_all": 2008, "formula_len_all_1": 1786, "len_all": 86381, "len_all_1": 63585, "len_abs": 1329, "len_title": 119, "len_sents": 24185, "len_sents_1": 22153, "n_sents": 259, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1335, "title": "Regularised Jump Models for Regime Identification and Feature Selection", "abs": "A regime modelling framework can be employed to address the complexities of financial markets. Under the framework, market periods are grouped into distinct regimes, each distinguished by similar statistical characteristics. Regimes in financial markets are not directly observable but are often manifested in market and macroeconomic variables. The objective of regime modelling is to accurately identify the active regime from these variables at a point in time, a process known as regime identification. One way to enhance the accuracy of regime identification is to select features that are most responsible for statistical differences between regimes, a process known as feature selection. Models based on the Jump Model framework have recently been developed to address the joint problem of regime identification and feature selection. In the following work, we propose a new set of models called Regularised Jump Models that are founded upon the Jump Model framework. These models perform feature selection that is more interpretable than that from the Sparse Jump Model, a model proposed in the literature pertaining to the Jump Model framework. Through a simulation experiment, we find evidence that these new models outperform the Standard and Sparse Jump Models, both in terms of regime identification and feature selection.", "keywords": [ "jump models", "regime identification", "feature selection" ], "rating_list": [ 1, 3, 1, 3 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 1, 2, 1, 1 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "JXogIgQV86", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Adams Wai-Kin Kong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Thant Zin Oo", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "Wai Lee Chan", "gender": "Male", "institution": "Nanyang Technological University, Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Jessica Sher En Loh", "gender": "Not Specified", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "WEI XIAN LIM", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 40, "n_ref": 53, "n_ref_all": 75, "n_fig": 9, "n_tab": 15, "L_tab": 3350, "n_element_tab": 288, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 1226, "n_element_tab_1": 97, "formula_len_all": 1269, "formula_len_all_1": 948, "len_all": 153418, "len_all_1": 65735, "len_abs": 1270, "len_title": 125, "len_sents": 47916, "len_sents_1": 29229, "n_sents": 343, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1275, "title": "IMPROVING FLOW FIELD PREDICTION OF COMPLEX GEOMETRIES USING SIMPLE GEOMETRIES", "abs": "In this study, we address the challenge of computationally expensive simulations of complex geometries, which are crucial for modern engineering design processes. While neural network-based flow field predictions have been suggested, prior studies generally exclude complex geometries. Our objective is to enhance flow predictions around complex geometries, which may often be deconstructed into multiple single, simple bodies, by leveraging existing data on these simple geometry flow fields. Using a case study of tandem-airfoils, we introduce a method employing the directional integrated distance representation for multiple objects, a residual pre-training scheme based on the freestream condition as a physical prior, and a residual training scheme utilising smooth combinations of single airfoil flow fields, also capitalising on the freestream condition. To optimise memory usage during training in large domains and improve prediction performance, we decom- pose simulation domains into smaller sub-domains, each processed by a different network. Extensive experiments on four new tandem-airfoil datasets, comprising over 2000 fluid simulations, demonstrate that our proposed method and techniques effectively enhance tandem-airfoil prediction accuracy by up to 96%.", "keywords": [ "Computational Fluid Dynamics", "Tandem Airfoils", "Geometry Representations", "Graph Neural Network", "Machine Learning for Sciences" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "JXgnnUC0PH", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Bin Wen", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Changyi Liu", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Di ZHANG", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "VP" }, { "name": "Fan Yang", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Haojie Ding", "gender": "unknown", "institution": "Beijing Jiaotong University", "country": "", "position": "MS student" }, { "name": "Huihui Xiao", "gender": "Female", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Jiankang Chen", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Tianke Zhang", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Tingting Gao", "gender": "Female", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Yaya Shi", "gender": "Female", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "cheng.feng", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 44, "n_ref": 121, "n_ref_all": 135, "n_fig": 7, "n_tab": 4, "L_tab": 1832, "n_element_tab": 235, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 404, "n_element_tab_1": 44, "formula_len_all": 28, "formula_len_all_1": 28, "len_all": 154670, "len_all_1": 67533, "len_abs": 1565, "len_title": 146, "len_sents": 38977, "len_sents_1": 33947, "n_sents": 251, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 96, "L_abs": 1314, "title": "TaskGalaxy: Scaling Multi-modal Instruction Fine-tuning with Tens of Thousands Vision Task Types", "abs": "Multimodal visual language models are gaining prominence in open-world applications, driven by advancements in model architectures, training techniques, and high-quality data. However, their performance is often limited by insufficient task-specific data, leading to poor generalization and biased outputs. Existing efforts to increase task diversity in fine-tuning datasets are hindered by the labor-intensive process of manual task labeling, which typically produces only a few hundred task types. To address this, we propose TaskGalaxy, a large-scale multimodal instruction fine-tuning dataset comprising 19,227 hierarchical task types and 413,648 samples. TaskGalaxy utilizes GPT-4o to enrich task diversity by expanding from a small set of manually defined tasks, with CLIP and GPT-4o filtering those that best match open-source images, and generating relevant question-answer pairs. Multiple models are employed to ensure sample quality. This automated process enhances both task diversity and data quality, reducing manual intervention. Incorporating TaskGalaxy into LLaVA-v1.5 and InternVL-Chat-v1.0 models shows substantial performance improvements across 16 benchmarks, demonstrating the critical importance of task diversity. TaskGalaxy is publicly released at https://github.com/Kwai-YuanQi/TaskGalaxy.", "keywords": [ "instruction fine-tunning dataset", "multimodal large language models", "hierarchical task types", "data generation pipeline" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "JXd1QUREJb", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Vitaliy Kurlin", "gender": "Not Specified", "institution": "University of Liverpool", "country": "GB", "position": "Full Professor" }, { "name": "Yury Elkin", "gender": "unknown", "institution": "University of Liverpool", "country": "", "position": "Postdoc" } ], "n_formula": 24, "n_formula_1": 12, "n_ref_uni": 54, "n_ref": 64, "n_ref_all": 80, "n_fig": 5, "n_tab": 1, "L_tab": 267, "n_element_tab": 25, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 543, "n_element_tab_1": 39, "formula_len_all": 4137, "formula_len_all_1": 1391, "len_all": 192109, "len_all_1": 70917, "len_abs": 873, "len_title": 137, "len_sents": 55601, "len_sents_1": 31926, "n_sents": 538, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 888, "title": "Complete and Lipschitz continuous invariants of graphs under geometric isomorphism in R^n", "abs": "Euclidean graphs embedded in R^n with unordered vertices and straight-line edges represent important real objects such as molecules whose atoms are connected by chemical bonds. Many real objects preserve their properties under any rigid motion from the special Euclidean group SE(n).Embedded graphs were previously distinguished under such rigid motion or geometric isomorphism in R^n.\n\nExperimental noise motivates new Lipschitz continuous invariants so that perturbations of all vertices up to epsilon change the invariants up to a constant multiple of epsilon in a suitable metric, whose running time should polynomially depend on the number of unordered vertices. \n\nWe developed new complete invariants that are stable under noise, form a natural hierarchy, and distinguish all chemically different graphs in the QM9 database of 130K+ molecules within a few hours on a modest desktop.", "keywords": [ "Euclidean graph", "geometric isomorphism", "complete SE(n)-invariant", "Lipschitz continuous metric", "molecule" ], "rating_list": [ 5, 3, 8, 1, 6 ], "soundness_list": [ 3, 2, 3, 2, 4 ], "presentation_list": [ 2, 1, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 1, 3 ], "confidence_list": [ 3, 4, 2, 5, 2 ] }, { "paper_id": "JXKFPJe0NU", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Huan Zhang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Jiangwei Yu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Jose Barreiros", "gender": "unknown", "institution": "Toyota Research Institute", "country": "", "position": "Researcher" }, { "name": "Keyi Shen", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "MS student" }, { "name": "Yunzhu Li", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 49, "n_formula_1": 14, "n_ref_uni": 53, "n_ref": 113, "n_ref_all": 175, "n_fig": 13, "n_tab": 11, "L_tab": 3317, "n_element_tab": 371, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 406, "n_element_tab_1": 7, "formula_len_all": 55814, "formula_len_all_1": 51853, "len_all": 348422, "len_all_1": 129321, "len_abs": 1912, "len_title": 126, "len_sents": 89474, "len_sents_1": 33755, "n_sents": 751, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1549, "title": "BaB-ND: Long-Horizon Motion Planning with Branch-and-Bound and Neural Dynamics", "abs": "Neural-network-based dynamics models learned from observational data have shown strong predictive capabilities for scene dynamics in robotic manipulation tasks. However, their inherent non-linearity presents significant challenges for effective planning. Current planning methods, often dependent on extensive sampling or local gradient descent, struggle with long-horizon motion planning tasks involving complex contact events.\nIn this paper, we present a GPU-accelerated branch-and-bound (BaB) framework for motion planning in manipulation tasks that require trajectory optimization over neural dynamics models. Our approach employs a specialized branching heuristic to divide the search space into sub-domains and applies a modified bound propagation method, inspired by the state-of-the-art neural network verifier $\\alpha,\\beta$-CROWN, to efficiently estimate objective bounds within these sub-domains. The branching process guides planning effectively, while the bounding process strategically reduces the search space.\nOur framework achieves superior planning performance, generating high-quality state-action trajectories and surpassing existing methods in challenging, contact-rich manipulation tasks such as non-prehensile planar pushing with obstacles, object sorting, and rope routing in both simulated and real-world settings. Furthermore, our framework supports various neural network architectures, ranging from simple multilayer perceptrons to advanced graph neural dynamics models, and scales efficiently with different model sizes.", "keywords": [ "Robotic Manipulation", "Model-Based Planning", "Neural Dynamics", "Branch-and-Bound Method" ], "rating_list": [ 5, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "JWtrk7mprJ", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Andreas Krause", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Kacper Wyrwal", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "MS student" }, { "name": "Viacheslav Borovitskiy", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" } ], "n_formula": 45, "n_formula_1": 19, "n_ref_uni": 38, "n_ref": 87, "n_ref_all": 135, "n_fig": 18, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3500, "formula_len_all_1": 1849, "len_all": 175515, "len_all_1": 75705, "len_abs": 1116, "len_title": 93, "len_sents": 58654, "len_sents_1": 31689, "n_sents": 421, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 45, "L_abs": 1122, "title": "Residual Deep Gaussian Processes on Manifolds", "abs": "We propose practical deep Gaussian process models on Riemannian manifolds, similar in spirit to residual neural networks.\nWith manifold-to-manifold hidden layers and an arbitrary last layer, they can model manifold- and scalar-valued functions, as well as vector fields.\nWe target data inherently supported on manifolds, which is too complex for shallow Gaussian processes thereon.\nFor example, while the latter perform well on high-altitude wind data, they struggle with the more intricate, nonstationary patterns at low altitudes.\nOur models significantly improve performance in these settings, enhancing prediction quality and uncertainty calibration, and remain robust to overfitting, reverting to shallow models when additional complexity is unneeded.\nWe further showcase our models on Bayesian optimisation problems on manifolds, using stylised examples motivated by robotics, and obtain substantial improvements in later stages of the optimisation process.\nFinally, we show our models to have potential for speeding up inference for non-manifold data, when, and if, it can be mapped to a proxy manifold well enough.", "keywords": [ "Gaussian processes", "manifolds", "deep Gaussian processes", "probabilistic methods", "variational inference", "uncertainty quantification", "geometric learning" ], "rating_list": [ 8, 8, 6, 8, 8 ], "soundness_list": [ 4, 3, 3, 4, 3 ], "presentation_list": [ 3, 4, 3, 4, 4 ], "contribution_list": [ 4, 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 2, 3 ] }, { "paper_id": "JWGJblAs7V", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Brian Quanz", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Senior Research Scientist" }, { "name": "Georgia Perakis", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Pavithra Harsha", "gender": "Female", "institution": "IBM, International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Rares C Cristian", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 35, "n_formula_1": 24, "n_ref_uni": 22, "n_ref": 29, "n_ref_all": 50, "n_fig": 5, "n_tab": 5, "L_tab": 1477, "n_element_tab": 110, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2800, "n_element_tab_1": 74, "formula_len_all": 1724, "formula_len_all_1": 1313, "len_all": 147507, "len_all_1": 79226, "len_abs": 1354, "len_title": 97, "len_sents": 54740, "len_sents_1": 34642, "n_sents": 572, "n_sents_1": 335, "venue": "ICLR2025", "decision": "Reject", "L_title": 48, "L_abs": 1368, "title": "End-to-End Learning under Endogenous Uncertainty", "abs": "How can we effectively learn to make decisions when there are no ground-truth counterfactual observations? We propose an end-to-end learning approach to the contextual stochastic optimization problem under decision-dependent uncertainty. We propose both exact methods and efficient sampling-based methods to implement our approach. We also introduce a new class of two-stage stochastic optimization problems to the end-to-end learning framework. Here, the first stage is an information-gathering problem to decide which random variable to ``poll'' and gain information about before making a second-stage decision based off of it. We provide theoretical analysis showing (1) that optimally minimizing our proposed objective produces optimal decisions and (2) generalization bounds between in-sample and out-of-sample cost.\nWe computationally test the proposed approach on multi-item assortment problems where demand is affected by cross-item complementary and supplementary effects. Overall, our method outperforms other benchmarks by more than 15\\% and performs best in high noise, across any cost configuration, and when given sufficient data.\nWe also introduce an experiment for the information-gathering problem on a real-world electricity generation problem. We show our method proposes decisions with more than 7\\% lower cost than other decision-making methods.", "keywords": [ "End-to-end learning", "contextual stochastic optimization" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "JVkdSi7Ekg", "primary_area": "applications to robotics, autonomy, planning", "n_author": 10, "author_info_list": [ { "name": "Ajay Mandlekar", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Dieter Fox", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Senior Director of Robotics Research" }, { "name": "Jiafei Duan", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Nishanth Kumar", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Intern" }, { "name": "Ranjay Krishna", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" }, { "name": "Shulin Tian", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Wentao Yuan", "gender": "Male", "institution": "University of Washington, Seattle", "country": "", "position": "PhD student" }, { "name": "Wilbert Pumacay", "gender": "Male", "institution": "Universidad Nacional de Ingenier\u00eda", "country": "PE", "position": "Researcher" }, { "name": "Yi Ru Wang", "gender": "unknown", "institution": "University of Washington", "country": "", "position": "PhD student" }, { "name": "Yijie Guo", "gender": "Female", "institution": "University of Michigan", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 88, "n_ref_all": 106, "n_fig": 4, "n_tab": 2, "L_tab": 1063, "n_element_tab": 151, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 233, "n_element_tab_1": 24, "formula_len_all": 22, "formula_len_all_1": 0, "len_all": 141688, "len_all_1": 62902, "len_abs": 1750, "len_title": 144, "len_sents": 37445, "len_sents_1": 32041, "n_sents": 246, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1792, "title": "AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulation", "abs": "Robotic manipulation in open-world settings requires not only task execution but also the ability to detect and learn from failures. While recent advances in vision-language models (VLMs) and large language models (LLMs) have improved robots' spatial reasoning and problem-solving abilities, they still struggle with failure recognition, limiting their real-world applicability. We introduce AHA, an open-source VLM designed to detect and reason about failures in robotic manipulation using natural language. By framing failure detection as a free-form reasoning task, AHA identifies failures and provides detailed, adaptable explanations across different robots, tasks, and environments. We fine-tuned AHA using FailGen, a scalable framework that generates the first large-scale dataset of robotic failure trajectories, the AHA dataset. FailGen achieves this by procedurally perturbing successful demonstrations from simulation. Despite being trained solely on the AHA dataset, AHA generalizes effectively to real-world failure datasets, robotic systems, and unseen tasks. It surpasses the second-best model (GPT-4o in-context learning) by 10.3% and exceeds the average performance of six compared models, including five state-of-the-art VLMs, by 35.3% across multiple metrics and datasets. We integrate AHA into three manipulation frameworks that utilize LLMs/VLMs for reinforcement learning, task and motion planning, and zero-shot trajectory generation. AHA\u2019s failure feedback enhances these policies' performances by refining dense reward functions, optimizing task planning, and improving sub-task verification, boosting task success rates by an average of 21.4% across all three tasks compared to GPT-4 models. Anonymous project page: [aha-iclr.github.io](https://aha-iclr.github.io/).", "keywords": [ "Robotic Manipulation; Data Generation; Vision-Language-Model; Failure Reasoning; Failure Detection" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 4, 3 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "JVJE5yZRxm", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Kamel Yamani", "gender": "Male", "institution": "New York University, Abu Dhabi", "country": "AE", "position": "Intern" }, { "name": "Marwa Na\u00efr", "gender": "Female", "institution": "New York University, Abu Dhabi", "country": "AE", "position": "Intern" }, { "name": "Riyadh Baghdadi", "gender": "Male", "institution": "New York University", "country": "AE", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 19, "n_ref": 26, "n_ref_all": 37, "n_fig": 7, "n_tab": 2, "L_tab": 199, "n_element_tab": 18, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 199, "n_element_tab_1": 18, "formula_len_all": 330, "formula_len_all_1": 330, "len_all": 103581, "len_all_1": 53818, "len_abs": 1813, "len_title": 95, "len_sents": 31661, "len_sents_1": 25783, "n_sents": 236, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1826, "title": "Teaching Code Execution to Tiny Language Models", "abs": "Recent advancements in large language models have demonstrated their effectiveness in various tasks. However, the question of these models' limitations remains open though. For instance, can a language model learn to perform code execution (i.e., predicting the output of code)? Current research indicates that the performance of state-of-the-art large language models in code execution is still limited. The reasons for this limitations are unclear though. Is it due to fundamental constraints or other factors such as training data and computational resources? Is the next-token prediction objective sufficient for learning code execution? How small can a language model be while still capable of learning code execution? In this paper, we investigate these questions. More specifically, we investigate whether tiny language models, trained from scratch using the next-token prediction objective, can effectively learn to execute code. Our experiments show that, given appropriate data, model size, and computational resources, tiny language models can indeed learn to perform code execution with a 99.13% accuracy for a tiny Turing-complete programming language. We begin by defining a tiny programming language called TinyPy. Millions of randomly generated codes in this language, along with their outputs, are used to train our tiny language models using the next-token prediction task. We then conduct a series of experiments to determine the smallest model size, data amount, and computational resources necessary to train our language model to achieve near-perfect accuracy in code execution. Our findings reveal that TEX, our proposed tiny language model with 15M parameters, can successfully learn code execution. This suggests that a task as complex as predicting code output is within the reach of language models.", "keywords": [ "Code Language Models", "Tiny Language Models", "Code Execution" ], "rating_list": [ 1, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 3, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "JVFRwCx3Dy", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Christian Rupprecht", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" }, { "name": "Jelena Bratuli\u0107", "gender": "Female", "institution": "Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" }, { "name": "Sudhanshu Mittal", "gender": "Male", "institution": "University of Freiburg, Albert-Ludwigs-Universit\u00e4t Freiburg", "country": "DE", "position": "Postdoc" }, { "name": "Thomas Brox", "gender": "Male", "institution": "University of Freiburg", "country": "", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 38, "n_ref": 72, "n_ref_all": 120, "n_fig": 29, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 17, "n_tab_1": 2, "L_tab_1": 65, "n_element_tab_1": 10, "formula_len_all": 179, "formula_len_all_1": 86, "len_all": 175146, "len_all_1": 65501, "len_abs": 1187, "len_title": 135, "len_sents": 55929, "len_sents_1": 33021, "n_sents": 443, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1195, "title": "What Matters for In-Context Learning: A Balancing Act of Look-up and In-Weight Learning", "abs": "Large Language Models (LLMs) have demonstrated impressive performance in various tasks, including In-Context Learning (ICL), where the model performs new tasks by conditioning solely on the examples provided in the context, without updating the model's weights. While prior research has explored the roles of pretraining data and model architecture, the key mechanism behind ICL remains unclear. In this work, we systematically uncover properties present in LLMs that support the emergence of ICL. To disambiguate these factors, we conduct a study with a controlled dataset and data sequences using a deep autoregressive model. We show that conceptual repetitions in the data sequences are crucial for ICL, more so than previously indicated training data properties like burstiness or long-tail distribution. Conceptual repetitions could refer to $n$-gram repetitions in textual data or exact image copies in image sequence data. Such repetitions also offer other previously overlooked benefits such as reduced transiency in ICL performance. Furthermore, we show that the emergence of ICL depends on balancing the in-weight learning objective with the in-context solving ability during training.", "keywords": [ "in-context learning", "few-shot learning" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "JVBR1ud4lL", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Futa Kai Waseda", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Koki Wataoka", "gender": "Male", "institution": "SB Intuitions", "country": "JP", "position": "Researcher" }, { "name": "Shojiro Yamabe", "gender": "Male", "institution": "Institute of Science Tokyo", "country": "JP", "position": "MS student" }, { "name": "Tsubasa Takahashi", "gender": "Male", "institution": "Turing Inc.", "country": "JP", "position": "Principal Researcher" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 42, "n_ref": 72, "n_ref_all": 97, "n_fig": 4, "n_tab": 8, "L_tab": 2812, "n_element_tab": 478, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1787, "n_element_tab_1": 261, "formula_len_all": 745, "formula_len_all_1": 524, "len_all": 169231, "len_all_1": 70743, "len_abs": 1095, "len_title": 119, "len_sents": 42429, "len_sents_1": 30649, "n_sents": 371, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1100, "title": "MergePrint: Robust Fingerprinting against Merging Large Language Models", "abs": "As the cost of training large language models (LLMs) rises, protecting their intellectual property has become increasingly critical. Model merging, which integrates multiple expert models into a single model capable of performing multiple tasks, presents a growing risk of unauthorized and malicious usage.\nWhile fingerprinting techniques have been studied for asserting model ownership, existing methods have primarily focused on fine-tuning, leaving model merging underexplored.\nTo address this gap, we propose a novel fingerprinting method MergePrint that embeds robust fingerprints designed to preserve ownership claims even after model merging.\nBy optimizing against a pseudo-merged model, which simulates post-merged model weights, MergePrint generates fingerprints that remain detectable after merging.\nAdditionally, we optimize the fingerprint inputs to minimize performance degradation, enabling verification through specific outputs from targeted inputs. \nThis approach provides a practical fingerprinting strategy for asserting ownership in cases of misappropriation through model merging.", "keywords": [ "model fingerprinting", "large language models", "model merging" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "JV8zULNh24", "primary_area": "interpretability and explainable AI", "n_author": 9, "author_info_list": [ { "name": "Alexander Sabol", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Nathaniel Weir", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Orion Weller", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "PhD student" }, { "name": "Oyvind Tafjord", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Peter Clark", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Senior Research Manager" }, { "name": "Peter Jansen", "gender": "unknown", "institution": "University of Arizona", "country": "US", "position": "Associate Professor" }, { "name": "Sam Hornstein", "gender": "Male", "institution": "Thomas Jefferson University", "country": "US", "position": "Medical student" }, { "name": "Benjamin Van Durme", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Bhavana Dalvi", "gender": "Female", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Lead Reserarch Scientist" } ], "n_formula": 13, "n_formula_1": 4, "n_ref_uni": 33, "n_ref": 53, "n_ref_all": 95, "n_fig": 24, "n_tab": 3, "L_tab": 1023, "n_element_tab": 400, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 831, "formula_len_all_1": 404, "len_all": 215856, "len_all_1": 75873, "len_abs": 4721, "len_title": 148, "len_sents": 77357, "len_sents_1": 36232, "n_sents": 520, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1909, "title": "From Models to Microtheories: Distilling a Model's Topical Knowledge for Grounded Question-Answering", "abs": "Recent reasoning methods (e.g., chain-of-thought, entailment reasoning) help users understand how language models (LMs) answer a single question, but they do little to reveal the LM\u2019s _overall understanding_, or \u201ctheory\u201d, about the question\u2019s _topic_, making it still hard to trust the model. Our goal is to materialize such theories - here called _microtheories_ (a linguistic analog of logical microtheories (Blair et al., 1992)) - as a set of sentences encapsulating an LM\u2019s core knowledge about a topic. These statements systematically work together to entail answers to a _set_ of questions to both engender trust and improve performance. Our approach is to first populate a knowledge store with (model-generated) sentences that entail answers to training questions, and then distill those down to a core microtheory which is concise, general, and non-redundant. We show that, when added to a general corpus (e.g., Wikipedia), microtheories can supply critical, topical information not necessarily present in the corpus, improving both a model\u2019s ability to ground its answers to verifiable knowledge (i.e., show how answers are systematically entailed by documents in the corpus, fully grounding up to +8% more answers), and the accuracy of those grounded answers (up to +8% absolute). We also show that, in a human evaluation in the medical domain, our distilled microtheories contain a significantly higher concentration of topically critical facts than the non-distilled knowledge store. Finally, we show we can quantify the coverage of a microtheory for a topic (characterized by a dataset) using a notion of _p-relevance_. Together, these suggest that microtheories are an efficient distillation of an LM\u2019s topic-relevant knowledge, that they can usefully augment existing corpora, and can provide both performance gains and an interpretable, verifiable window into the model\u2019s knowledge of a topic.", "keywords": [ "microtheory", "textual entailment", "knowledge representation", "natural language reasoning", "text retrieval", "automatic knowledge base construction" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "JUu0tsd0Zk", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Feng Yu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Gengmo Zhou", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Wenda Wang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Zhen Wang", "gender": "Male", "institution": "DP Technology", "country": "", "position": "Researcher" }, { "name": "Zhewei Wei", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Zhifeng Gao", "gender": "Male", "institution": "DP Technology", "country": "", "position": "Researcher" }, { "name": "guolin ke", "gender": "Male", "institution": "DP Technology", "country": "CN", "position": "Senior Researcher" } ], "n_formula": 15, "n_formula_1": 10, "n_ref_uni": 24, "n_ref": 31, "n_ref_all": 39, "n_fig": 2, "n_tab": 4, "L_tab": 617, "n_element_tab": 81, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 641, "n_element_tab_1": 86, "formula_len_all": 1356, "formula_len_all_1": 1267, "len_all": 93033, "len_all_1": 53921, "len_abs": 1572, "len_title": 135, "len_sents": 28795, "len_sents_1": 24320, "n_sents": 225, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1580, "title": "FGW-CLIP: Enhancing Enzyme Screening via Fused Gromov-Wasserstein Contrastive Learning", "abs": "Enzymes are crucial catalysts for biochemical reactions, underpinning numerous biological processes. The efficient identification of specific enzymes from extensive protein libraries is essential for understanding and harnessing these biological reactions. While traditional computational methods for enzyme screening are time-consuming and resource-intensive, recent contrastive learning approaches have shown promise. However, these methods often overlook the inherent hierarchical classifications within enzymes and reactions, as well as the significance of molecular structure in catalysis. To address these limitations, we introduce FGW-CLIP, a novel contrastive learning framework based on optimizing the fused Gromov-Wasserstein distance. This approach incorporates multiple alignments, including representation alignment between reactions and enzymes, and internal alignment within enzyme and reaction representations. By introducing a regularization term, our method minimizes the Gromov-Wasserstein distance between enzyme and reaction spaces, enhancing information exchange within these domains. FGW-CLIP demonstrates superior performance on the widely-used EnzymeMap benchmark, significantly outperforming existing methods in enzyme virtual screening tasks. Notably, it achieves state-of-the-art results in both BEDROC and EF metrics, indicating its efficacy in identifying relevant enzymes for given reactions. These results highlight the potential of our method to advance virtual enzyme screening, offering a powerful tool for enzyme discovery and characterization.", "keywords": [ "Contrastive Learning", "Enzyme Screening", "3D Structure", "Hierarchical Classifications" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "JUr0YOMvZA", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Hengrui Gu", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "Kaishen Wang", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "Undergrad student" }, { "name": "Kaixiong Zhou", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Assistant Professor" }, { "name": "Meijun Gao", "gender": "Female", "institution": "Michigan State University", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 34, "n_ref": 85, "n_ref_all": 116, "n_fig": 14, "n_tab": 11, "L_tab": 4424, "n_element_tab": 413, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1179, "n_element_tab_1": 126, "formula_len_all": 528, "formula_len_all_1": 384, "len_all": 160477, "len_all_1": 60899, "len_abs": 1224, "len_title": 157, "len_sents": 40058, "len_sents_1": 27851, "n_sents": 281, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 107, "L_abs": 991, "title": "DAMO: Decoding by Accumulating Activations Momentum for Mitigating Hallucinations in Vision-Language Models", "abs": "Large Vision-Language Models (VLMs) exhibit significant potential in multimodal tasks but often struggle with hallucinations\u2014responses that are plausible yet visually ungrounded. In this work, we investigate the layer-wise prediction tendencies of VLMs and conduct an in-depth analysis of their decoding mechanism. We observe that VLMs tend to ``overthink'' during the final stages of decoding, making significant prediction shifts in the last few layers often favoring incorrect results, which leads to a surge in hallucinative outputs. Leveraging this localized pattern, we propose a novel decoding strategy inspired by the momentum analogy used in gradient descent-based optimizers. Our method enforces decoding consistency across layers in an adaptive manner during forward passes\u2014an under-explored approach in existing works. This strategy significantly improves the reliability and performance of VLMs in various multimodal tasks, while introducing only negligible efficiency overhead.", "keywords": [ "Vision-Language Models (VLMs)", "Hallucinations", "Decoding Method", "Momentum Techniques" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "JUYBEmwSJK", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Furu Wei", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Partner Research Manager" }, { "name": "Shuhuai Ren", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Shuming Ma", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Xu Sun", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 42, "n_ref": 86, "n_ref_all": 115, "n_fig": 14, "n_tab": 4, "L_tab": 1369, "n_element_tab": 224, "n_fig_1": 14, "n_tab_1": 2, "L_tab_1": 288, "n_element_tab_1": 29, "formula_len_all": 594, "formula_len_all_1": 445, "len_all": 173473, "len_all_1": 50334, "len_abs": 1388, "len_title": 121, "len_sents": 29199, "len_sents_1": 22258, "n_sents": 256, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1394, "title": "Next Block Prediction: Video Generation via Semi-Auto-Regressive Modeling", "abs": "Next-Token Prediction (NTP) is a de facto approach for autoregressive (AR) video generation, but it suffers from suboptimal unidirectional dependencies and slow inference speed. In this work, we propose a semi-autoregressive (semi-AR) framework, called Next-Block Prediction (NBP), for video generation. By uniformly decomposing video content into equal-sized blocks (e.g., rows or frames), we shift the generation unit from individual tokens to blocks, allowing each token in the current block to simultaneously predict the corresponding token in the next block. Unlike traditional AR modeling, our framework employs bidirectional attention within each block, enabling tokens to capture more robust spatial dependencies. By predicting multiple tokens in parallel, NBP models significantly reduce the number of generation steps, leading to faster and more efficient inference. Our model achieves FVD scores of 55.0 on UCF101 and 25.5 on K600, outperforming the vanilla NTP model by an average of 4.4. Furthermore, thanks to the reduced number of inference steps, the NBP model generates 8.89 frames (128x128 resolution) per second, achieving an 11\u00d7 speedup in inference. We also explored model scales ranging from 700M to 3B parameters, observing significant improvements in generation quality, with FVD scores dropping from 25.5 to 19.5 on K600, demonstrating the scalability of our approach.", "keywords": [ "video generation", "auto-regressive model", "semi-auto-regressive model" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "JUGLP5L8F3", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Bicheng Ying", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Haibo Yang", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Seyedsina Nabavirazavi", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "PhD student" }, { "name": "Sitharama Iyengar", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "Full Professor" }, { "name": "Zhe Li", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 71, "n_formula_1": 18, "n_ref_uni": 43, "n_ref": 107, "n_ref_all": 146, "n_fig": 4, "n_tab": 5, "L_tab": 3671, "n_element_tab": 397, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 910, "n_element_tab_1": 56, "formula_len_all": 13346, "formula_len_all_1": 1214, "len_all": 179968, "len_all_1": 75351, "len_abs": 1720, "len_title": 127, "len_sents": 48970, "len_sents_1": 34322, "n_sents": 447, "n_sents_1": 288, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1908, "title": "FAST: Federated Average with Snapshot Unleashes Arbitrary Client Participation", "abs": "Federated Learning (FL) provides a flexible distributed platform where numerous clients with high degrees of heterogeneity in data and system can collaborate to learn a model jointly. Previous research has shown that Federated Learning is effective in handling diverse data, but often assumes idealized conditions. Specifically, client participation is often simplified in these studies, while real-world factors make it difficult to predict or design individual client participation. This complexity often diverges from the ideal client participation assumption, rendering an unknown pattern of client participation, referred to as *arbitrary client participation*. Hence, it is an important open problem to explore the impact of client participation and find a lightweight mechanism to enable arbitrary client participation in FL. In this paper, we first empirically investigate the influence of client participation on FL, revealing that FL algorithms are significantly impacted by arbitrary client participation. Afterward, to alleviate the influence, we propose a lightweight solution, Federated Average with Snapshot (FAST), to unleash the almost arbitrary client participation for FL. It can seamlessly integrate with other classic FL algorithms. Specifically, FAST enforces the clients to take a snapshot once in a while and facilitates arbitrary client participation for the majority of the training process. We show the convergence rates of FAST in non-convex and strongly-convex cases, which could match the rates with those in ideal client participation. Furthermore, we empirically introduce an adaptive strategy for dynamically configuring the snapshot frequency, tailored to accommodate diverse FL systems. Our extensive numerical results demonstrate that our FAST algorithm attains significant improvements under the conditions of arbitrary client participation and highly heterogeneous data.", "keywords": [ "Federated Learning", "Optimization" ], "rating_list": [ 3, 5, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 2, 4 ] }, { "paper_id": "JU9oHs7ivN", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Chaoyang Zhu", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Chuhan ZHANG", "gender": "Female", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Researcher" }, { "name": "Dong Zhang", "gender": "Male", "institution": "HKUST", "country": "CN", "position": "Researcher" }, { "name": "Long Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Pingcheng Dong", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 40, "n_ref": 102, "n_ref_all": 122, "n_fig": 8, "n_tab": 3, "L_tab": 2124, "n_element_tab": 165, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 984, "n_element_tab_1": 131, "formula_len_all": 516, "formula_len_all_1": 449, "len_all": 128784, "len_all_1": 63062, "len_abs": 1396, "len_title": 122, "len_sents": 34702, "len_sents_1": 29819, "n_sents": 245, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1364, "title": "Cyclic Contrastive Knowledge Transfer for Open-Vocabulary Object Detection", "abs": "In pursuit of detecting unstinted objects that extend beyond predefined categories, prior arts of open-vocabulary object detection (OVD) typically resort to pretrained vision-language models (VLMs) for base-to-novel category generalization. However, to mitigate the misalignment between upstream image-text pretraining and downstream region-level perception, additional supervisions are indispensable, e.g., image-text pairs or pseudo annotations generated via self-training strategies. In this work, we propose CCKT-Det trained without any extra supervision. The proposed framework constructs a cyclic and dynamic knowledge transfer from language queries and visual region features extracted from VLMs, which forces the detector to closely align with the visual-semantic space of VLMs. Specifically, 1) we prefilter and inject semantic priors to guide the learning of queries, and 2) introduce a regional contrastive loss to improve the awareness of queries on novel objects. CCKT-Det can consistently improve performance as the scale of VLMs increases, all while requiring the detector at a moderate level of computation overhead. Comprehensive experimental results demonstrate that our method achieves performance gain of +2.9% and +10.2% AP_{50} over previous state-of-the-arts on the challenging COCO benchmark, both without and with a stronger teacher model.", "keywords": [ "Contrastive Learning", "Knowledge Transfer", "Open-Vocabulary Object Detection" ], "rating_list": [ 6, 5, 6, 5, 8 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 5, 3, 3, 4, 4 ] }, { "paper_id": "JTji0Jfh5a", "primary_area": "reinforcement learning", "n_author": 9, "author_info_list": [ { "name": "Changjie Fan", "gender": "Male", "institution": "Netease, Fuxi AI Lab", "country": "", "position": "Principal Researcher" }, { "name": "Paul Weng", "gender": "Male", "institution": "Duke Kunshan University", "country": "", "position": "Associate Professor" }, { "name": "Tangjie Lv", "gender": "Male", "institution": "NetEase, Inc.", "country": "CN", "position": "Researcher" }, { "name": "Tianze Zhou", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "MS student" }, { "name": "Xuening Feng", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Intern" }, { "name": "Yifei Zhu", "gender": "unknown", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yujing Hu", "gender": "unknown", "institution": "NetEase, Inc.", "country": "", "position": "Researcher" }, { "name": "Zhaohui JIANG", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Song Yan", "gender": "Male", "institution": "NetEase, Inc.", "country": "", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 10, "n_ref_uni": 49, "n_ref": 106, "n_ref_all": 214, "n_fig": 24, "n_tab": 12, "L_tab": 4286, "n_element_tab": 221, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 267, "n_element_tab_1": 12, "formula_len_all": 2473, "formula_len_all_1": 998, "len_all": 207064, "len_all_1": 71725, "len_abs": 1834, "len_title": 123, "len_sents": 60263, "len_sents_1": 33280, "n_sents": 482, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1843, "title": "Reinforcement Learning from Imperfect Corrective Actions and Proxy Rewards", "abs": "In practice, reinforcement learning (RL) agents are often trained with a possibly imperfect proxy reward function, which may lead to a human-agent alignment issue (i.e., the learned policy either converges to non-optimal performance with low cumulative rewards, or achieves high cumulative rewards but in an undesired manner). To tackle this issue, we consider a framework where a human labeler can provide additional feedback in the form of corrective actions, which expresses the labeler's action preferences although this feedback may possibly be imperfect as well. \nIn this setting, to obtain a better-aligned policy guided by both learning signals, we propose a novel value-based deep RL algorithm called **I**terative learning from **Co**rrective actions and **Pro**xy rewards (ICoPro), which cycles through three phases: \n(1) Solicit sparse corrective actions from a human labeler on the agent's demonstrated trajectories; \n(2) Incorporate these corrective actions into the Q-function using a margin loss to enforce adherence to labeler's preferences; \n(3) Train the agent with standard RL losses regularized with a margin loss to learn from proxy rewards and propagate the Q-values learned from human feedback. Moreover, another novel design in our approach is to integrate pseudo-labels from the target Q-network to reduce human labor and further stabilize training. \nWe experimentally validate our proposition on a variety of tasks (Atari games and autonomous driving on highway). On the one hand, using proxy rewards with different levels of imperfection, our method can better align with human and is more sample-efficient than baseline methods. On the other hand, facing corrective actions with different types of imperfection, our method can overcome the non-optimality of this feedback thanks to the guidance from proxy rewards.", "keywords": [ "reinforcement learning", "imitation learning", "corrective action", "proxy reward", "human-agent alignment" ], "rating_list": [ 8, 8, 3, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "JT53iXH7eO", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Kaipeng Zhang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Lirui Zhao", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Tianshuo Yang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Wenqi Shao", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yuxin Zhang", "gender": "unknown", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Ping Luo", "gender": "unknown", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 38, "n_ref": 123, "n_ref_all": 177, "n_fig": 23, "n_tab": 2, "L_tab": 1231, "n_element_tab": 82, "n_fig_1": 12, "n_tab_1": 3, "L_tab_1": 647, "n_element_tab_1": 90, "formula_len_all": 656, "formula_len_all_1": 601, "len_all": 156657, "len_all_1": 58633, "len_abs": 1309, "len_title": 119, "len_sents": 54390, "len_sents_1": 27716, "n_sents": 385, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1314, "title": "Diffree: Text-Guided Shape Free Object Inpainting with Diffusion Model", "abs": "This paper addresses an important problem of object addition for images with only text guidance. It is challenging because the new object must be integrated seamlessly into the image with consistent visual context, such as lighting, texture, and spatial location. While existing text-guided image inpainting methods can add objects, they either fail to preserve the background consistency or involve cumbersome human intervention in specifying bounding boxes or user-scribbled masks. To tackle this challenge, we introduce Diffree, a Text-to-Image (T2I) model that facilitates text-guided object addition with only text control. To this end, we curate OABench, an exquisite synthetic dataset by removing objects with advanced image inpainting techniques. OABench comprises 74K real-world tuples of an original image, an inpainted image with the object removed, an object mask, and object descriptions. Trained on OABench using the Stable Diffusion model with an additional mask prediction module, Diffree uniquely predicts the position of the new object and achieves object addition with guidance from only text. Extensive experiments demonstrate that Diffree excels in adding new objects with a high success rate while maintaining background consistency, spatial appropriateness, and object relevance and quality.", "keywords": [ "Image inpainting; Text-guided image editing;" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "JSB171dSUU", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "benyou wang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Juhao Liang", "gender": "Male", "institution": "Chinese University of Hong Kong, Shenzhen", "country": "", "position": "PhD student" }, { "name": "Nuo Chen", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Xidong Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Yuping Zheng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Undergrad student" }, { "name": "Zheng Guorui", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 34, "n_ref": 55, "n_ref_all": 76, "n_fig": 9, "n_tab": 8, "L_tab": 5281, "n_element_tab": 248, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1563, "n_element_tab_1": 47, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 200329, "len_all_1": 50825, "len_abs": 1457, "len_title": 97, "len_sents": 43178, "len_sents_1": 22795, "n_sents": 333, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1483, "title": "Efficiently Democratizing Medical LLMs for 50 Languages via a Mixture of Language Family Experts", "abs": "Adapting medical Large Language Models to local languages can reduce barriers to accessing healthcare services, but data scarcity remains a significant challenge, particularly for low-resource languages. To address this, we first construct a high-quality medical dataset and conduct analysis to ensure its quality. In order to leverage the generalization capability of multilingual LLMs to efficiently scale to more resource-constrained languages, we explore the internal information flow of LLMs from a multilingual perspective using Mixture of Experts (MoE) modularity. Technically, we propose a novel MoE routing method that employs language-specific experts and cross-lingual routing. Inspired by circuit theory, our routing analysis revealed a \\textit{``Spread Out in the End``} information flow mechanism: while earlier layers concentrate cross-lingual information flow, the later layers exhibit language-specific divergence. This insight directly led to the development of the Post-MoE architecture, which applies sparse routing only in the later layers while maintaining dense others. Experimental results demonstrate that this approach enhances the generalization of multilingual models to other languages while preserving interpretability. Finally, to efficiently scale the model to 50 languages, we introduce the concept of \\textit{language family} experts, drawing on linguistic priors, which enables scaling the number of languages without adding additional parameters.", "keywords": [ "Multilingual LLM", "Medical LLM" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "JRcfgNg2ZJ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Feidiao Yang", "gender": "Male", "institution": "Pengcheng Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Guanhong Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Quanmin Liang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Xiawu Zheng", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Yonghong Tian", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Yutong Lu", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "XINZI CAO", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 47, "n_ref": 111, "n_ref_all": 144, "n_fig": 7, "n_tab": 4, "L_tab": 2276, "n_element_tab": 142, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1757, "n_element_tab_1": 107, "formula_len_all": 1008, "formula_len_all_1": 1006, "len_all": 127599, "len_all_1": 65742, "len_abs": 1519, "len_title": 119, "len_sents": 31600, "len_sents_1": 28568, "n_sents": 247, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1536, "title": "Debiased Imbalanced Pseudo-Labeling for Generalized Category Discovery", "abs": "Generalized Category Discovery (GCD) is a challenging task that aims to recognize seen and novel categories within unlabeled data by leveraging labeled data. \nDesigning a prototype classifier to identify unlabeled samples instead of relying on traditional time-consuming clustering is well recognized as a milestone in GCD. \nHowever, we discover there exists a bias in this classifier: some seen categories are mistakenly classified as novel ones, leading to imbalanced pseudo-labeling during classifier learning.\nBased on this finding, we identify the low discriminability between seen and novel prototypes as the key issue.\nTo address this issue, we propose DebiasGCD, an effective debiasing method that integrates *dynamic prototype debiasing* (DPD) and *local representation alignment* (LRA).\nDPD dynamically maintains inter-prototype margins, encouraging the network to strengthen the learning of class-specific features and enhance prototype discrimination.\nAdditionally, LRA promotes local representation learning, enabling DPD to capture subtle details that further refine the understanding of class-specific features.\nIn this way, it successfully improves prototype discriminability and \ngenerates more reliable predictions for seen classes.\nExtensive experiments validate that our method effectively mitigates pseudo-labeling bias across all datasets, especially on fine-grained ones. For instance, it delivers a 10.7\\% boost on `Old' classes in CUB. Our code is available at\uff1ahttps://anonymous.4open.science/r/DebiasGCD-34F0.", "keywords": [ "Generalized Category Discovery", "Pseudo-Labeling" ], "rating_list": [ 3, 3, 6, 5, 3 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 5, 5, 3, 5, 5 ] }, { "paper_id": "JRXcvEg3OB", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Dequan Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Keyu Li", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhijie Deng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 36, "n_formula_1": 23, "n_ref_uni": 37, "n_ref": 65, "n_ref_all": 91, "n_fig": 6, "n_tab": 3, "L_tab": 1851, "n_element_tab": 109, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1551, "n_element_tab_1": 61, "formula_len_all": 2051, "formula_len_all_1": 1326, "len_all": 131489, "len_all_1": 62393, "len_abs": 1146, "len_title": 161, "len_sents": 42100, "len_sents_1": 28678, "n_sents": 295, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1154, "title": "Guided-BFNs: Towards Visualizing and Understanding Bayesian Flow Networks in the Context of Trajectory Planning", "abs": "Bayesian Flow Networks (BFNs) represent an emerging class of generative models that exhibit promising capabilities in modeling continuous, discretized, and discrete data. \nIn this paper, we develop Guided-BFNs to integrate BFNs with conditional guidance and gradient guidance to facilitate the effective application of such models in trajectory planning tasks. \nBased on our developments, we can better comprehend BFNs by inspecting the generation dynamics of the planning trajectories. \nThrough extensive parameter tuning and rigorous ablation experiments, we systematically delineate the functional roles of various parameters and elucidate the pivotal components within the structure of BFNs. Furthermore, we conduct a comparative analysis of the planning results between diffusion models and BFNs, to discern their similarities and differences. \nAdditionally, we undertake efforts to augment the performance of BFNs, including developing a faster and training-free sampling algorithm for sample generation.\nOur objectives encompass not only a comprehensive exploration of BFNs' structural insights but also the enhancement of their practical utility.", "keywords": [ "Bayesian flow networks", "trajectory planning", "model understanding" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "JQtuCumAFD", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Gil Shamai", "gender": "Male", "institution": "Technion, Technion - Israel Institute of Technology", "country": "IL", "position": "Postdoc" }, { "name": "Hen Davidov", "gender": "Not Specified", "institution": "Computer Science Department, Technion - Israel Institute of Technology", "country": "", "position": "MS student" }, { "name": "Ron Kimmel", "gender": "Male", "institution": "Computer Science Departmen, Technion-Israel Institute of Technology", "country": "IL", "position": "Full Professor" }, { "name": "Shai Feldman", "gender": "Male", "institution": "Computer Science Departmen, Technion-Israel Institute of Technology", "country": "IL", "position": "PhD student" }, { "name": "Yaniv Romano", "gender": "Male", "institution": "Technion, Technion", "country": "", "position": "Assistant Professor" } ], "n_formula": 81, "n_formula_1": 25, "n_ref_uni": 31, "n_ref": 77, "n_ref_all": 96, "n_fig": 2, "n_tab": 4, "L_tab": 1072, "n_element_tab": 27, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 11202, "formula_len_all_1": 1902, "len_all": 144917, "len_all_1": 63080, "len_abs": 1259, "len_title": 63, "len_sents": 45022, "len_sents_1": 29793, "n_sents": 383, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1268, "title": "Conformalized Survival Analysis for General Right-Censored Data", "abs": "We develop a framework to quantify predictive uncertainty in survival analysis, providing a reliable lower predictive bound (LPB) for the true, unknown patient survival time. Recently, conformal prediction has been used to construct such valid LPBs for *type-I right-censored data*, with the guarantee that the bound holds with high probability. Crucially, under the type-I setting, the censoring time is observed for all data points. As such, informative LPBs can be constructed by framing the calibration as an estimation task with covariate shift, relying on the conditionally independent censoring assumption. This paper expands the conformal toolbox for survival analysis, with the goal of handling the ubiquitous *general right-censored setting*, in which either the censoring or survival time is observed, but not both. The key challenge here is that the calibration cannot be directly formulated as a covariate shift problem anymore. Yet, we show how to construct LPBs with distribution-free finite-sample guarantees, under the same assumptions as conformal approaches for type-I censored data. Experiments demonstrate the informativeness and validity of our methods in simulated settings and showcase their practical utility using several real-world datasets.", "keywords": [ "conformal prediction", "survival analysis", "PAC", "covariate shift", "uncertainty quantification" ], "rating_list": [ 3, 8, 8, 3 ], "soundness_list": [ 1, 4, 3, 2 ], "presentation_list": [ 2, 4, 4, 2 ], "contribution_list": [ 1, 4, 3, 1 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "JQrBYfD2gg", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 13, "author_info_list": [ { "name": "Ang Li", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Associate Professor" }, { "name": "Chuan Liu", "gender": "Female", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Chuang Ren", "gender": "unknown", "institution": "University of Rochester", "country": "US", "position": "Full Professor" }, { "name": "Chunshu Wu", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Postdoc" }, { "name": "Dongfang Liu", "gender": "unknown", "institution": "Rochester Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "James Chenhao Liang", "gender": "Male", "institution": "U. S. Naval Research Laboratory", "country": "US", "position": "Researcher" }, { "name": "Michael Huang", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Full Professor" }, { "name": "Mingkai Chen", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Qifan Wang", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Principal Researcher" }, { "name": "Taowen Wang", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "US", "position": "Intern" }, { "name": "Tong Geng", "gender": "Male", "institution": "University of Rochester", "country": "", "position": "Assistant Professor" }, { "name": "shihui cao", "gender": "unknown", "institution": "University of Rochester", "country": "US", "position": "Postdoc" }, { "name": "Yingnian Wu", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 83, "n_ref": 150, "n_ref_all": 180, "n_fig": 13, "n_tab": 5, "L_tab": 3407, "n_element_tab": 315, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 2097, "n_element_tab_1": 208, "formula_len_all": 373, "formula_len_all_1": 366, "len_all": 238959, "len_all_1": 77339, "len_abs": 1486, "len_title": 113, "len_sents": 58554, "len_sents_1": 35078, "n_sents": 404, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1690, "title": "Inertial Confinement Fusion Forecasting via Large Language Models", "abs": "Controlled fusion energy is deemed pivotal for the advancement of human civilization. In this study, we introduce $\\textbf{LPI-LLM}$, a novel integration of Large Language Models (LLMs) with classical reservoir computing paradigms tailored to address a critical challenge, Laser-Plasma Instabilities ($\\texttt{LPI}$), in Inertial Confinement Fusion ($\\texttt{ICF}$). Our approach offers several key contributions: Firstly, we propose the $\\textit{LLM-anchored Reservoir}$, augmented with a $\\textit{Fusion-specific Prompt}$, enabling accurate forecasting of $\\texttt{LPI}$-generated-hot electron dynamics during implosion. Secondly, we develop $\\textit{Signal-Digesting Channels}$ to temporally and spatially describe the driver laser intensity across time, capturing the unique characteristics of $\\texttt{ICF}$ inputs. Lastly, we design the $\\textit{Confidence Scanner}$ to quantify the confidence level in forecasting, providing valuable insights for domain experts to design the $\\texttt{ICF}$ process. Extensive experiments demonstrate the superior performance of our method, achieving 1.90 CAE, 0.14 $\\texttt{top-1}$ MAE, and 0.11 $\\texttt{top-5}$ MAE in predicting Hard X-ray ($\\texttt{HXR}$) energies emitted by the hot electrons in $\\texttt{ICF}$ implosions, which presents state-of-the-art comparisons against concurrent best systems. Additionally, we present $\\textbf{LPI4AI}$, the first $\\texttt{LPI}$ benchmark based on physical experiments, aimed at fostering novel ideas in $\\texttt{LPI}$ research and enhancing the utility of LLMs in scientific exploration. Overall, our work strives to forge an innovative synergy between AI and $\\texttt{ICF}$ for advancing fusion energy.", "keywords": [ "AI for Science", "Inertial Confinement Fusion" ], "rating_list": [ 8, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "JQkj67NArS", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Alireza Entezari", "gender": "unknown", "institution": "University of Florida", "country": "US", "position": "Associate Professor" }, { "name": "Arunava Banerjee", "gender": "unknown", "institution": "University of Florida", "country": "", "position": "Associate Professor" }, { "name": "Leila Kalantari", "gender": "unknown", "institution": "", "country": "", "position": "" } ], "n_formula": 38, "n_formula_1": 31, "n_ref_uni": 18, "n_ref": 33, "n_ref_all": 40, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2005, "formula_len_all_1": 1633, "len_all": 78671, "len_all_1": 65203, "len_abs": 956, "len_title": 93, "len_sents": 27922, "len_sents_1": 27773, "n_sents": 196, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 958, "title": "Alternating Projections With Volume Sampling", "abs": "The method of Alternating Projections (AP) is a fundamental iterative technique with applications to problems in machine learning, optimization and signal processing. Examples include the Gauss-Seidel algorithm which is used to solve large-scale regression problems and the Kaczmarz and projections onto convex sets (POCS) algorithms that are fundamental to iterative reconstruction. Progress has been made with regards to the questions of efficiency and rate of convergence in the randomized setting of the AP method. Here, we extend these results with volume sampling to block (batch) sizes greater than 1 and provide explicit formulas that relate the convergence rate bounds to the spectrum of the underlying system. These results, together with a trace formula and associated volume sampling, prove that convergence rates monotonically improve with larger block sizes, a feature that can not be guaranteed in general with uniform sampling (e.g., in SGD).", "keywords": [ "Method of Alternating Projections", "Volume Sampling", "Optimization", "Iterative Methods" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "JQbqaQjV7D", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Anderson S. Chu", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Researcher" }, { "name": "Dan Zhang", "gender": "unknown", "institution": "Senslab Technology Co.,Ltd.", "country": "", "position": "Principal Researcher" }, { "name": "Lujun Li", "gender": "Male", "institution": "HKGAI", "country": "HK", "position": "Researcher" }, { "name": "Mingkun Tan", "gender": "Male", "institution": "Universit\u00e4t Bielefeld", "country": "DE", "position": "PhD student" }, { "name": "Porawit Kamnoedboon", "gender": "unknown", "institution": "University of Zurich", "country": "", "position": "PhD student" }, { "name": "Qiang Li", "gender": "Male", "institution": "Accenture", "country": "DE", "position": "Tech Lead in Computer Vision / IoT" }, { "name": "Shengzhao Lei", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "MS student" }, { "name": "Xun Zhao", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "PhD student" }, { "name": "Zhang Daoan", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 35, "n_ref": 80, "n_ref_all": 110, "n_fig": 4, "n_tab": 11, "L_tab": 5615, "n_element_tab": 746, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 2084, "n_element_tab_1": 330, "formula_len_all": 899, "formula_len_all_1": 339, "len_all": 149342, "len_all_1": 66029, "len_abs": 1477, "len_title": 171, "len_sents": 45501, "len_sents_1": 29136, "n_sents": 321, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 123, "L_abs": 1490, "title": "Industrial Benchmarking of LLMs: Assessing Hallucination in Traffic Incident Scenarios with a Novel Spatio-Temporal Dataset", "abs": "Large language models (LLMs) hold revolutionary potential to digitize and enhance the Health & Public Services (H&PS) industry. Despite their advanced linguistic abilities, concerns about accuracy, stability, and traceability still persist, especially in high-stakes areas such as transportation systems. Moreover, the predominance of English in LLM development raises questions about how they perform in non-English contexts.\n\nThis study introduces a novel cross-lingual benchmark dataset comprising nearly 99,869 real traffic incident records from Vienna (2013-2023) to assess the robustness of state-of-the-art LLMs (>9) in the spatio and temporal domain of traffic incident classification. We then explored three hypotheses \u2014 sentence indexing, date-to-text conversion, and German-to-English translation \u2014 and incorporated Retrieval Augmented Generation (RAG) to further examine the models' ability to handle hallucinations in both spatial and temporal contexts.\n\nOur experiments with GPT-4 and Llama models reveal significant performance disparities across these hypotheses in the spatio-temporal domain and also demonstrate how RAG can mitigate what types of hallucinations. These findings underscore the need for enhanced cross-lingual capabilities and improved explainability in LLMs. We provide open access to our Health & Public Services (H&PS) traffic incident dataset, with the project demo and code available at [Website](https://sites.google.com/view/llmhallucination/home).", "keywords": [ "Benchmark And Dataset", "GenAI", "LLMs", "Hallucination", "Trustworthy Machine Learning (accountability", "causality", "fairness", "privacy", "Robustness)" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 1 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "JQV9gH55Az", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Dongyu Luo", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Undergrad student" }, { "name": "Hairun Xie", "gender": "Male", "institution": "Shanghai Aircraft Design and Research Institute", "country": "CN", "position": "Postdoc" }, { "name": "Jianyu Wu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Undergrad student" }, { "name": "Jing Wang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "", "position": "Assistant Professor" }, { "name": "Miao Zhang", "gender": "Male", "institution": "Shanghai Aircraft Design and Research Institute", "country": "", "position": "Researcher" }, { "name": "Qianhong Ma", "gender": "Female", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "SHIXIANG TANG", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Wanli Ouyang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Full Professor" }, { "name": "Xiangyu Yue", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 43, "n_ref": 78, "n_ref_all": 105, "n_fig": 11, "n_tab": 11, "L_tab": 3083, "n_element_tab": 417, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 1729, "n_element_tab_1": 205, "formula_len_all": 517, "formula_len_all_1": 362, "len_all": 145332, "len_all_1": 58356, "len_abs": 1149, "len_title": 129, "len_sents": 41202, "len_sents_1": 26540, "n_sents": 288, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1152, "title": "SimDiffPDE: Simple Diffusion Baselines for Solving Partial Differential Equations", "abs": "We showcase good capabilities of the plain diffusion model with Transformers (SimDiffPDE) for general partial differential equations (PDEs) solving from various aspects, namely simplicity in model structure, scalability in model size, flexibility in training paradigm, and universality between different PDEs. Specifically, SimDiffPDE reformulates PDE-solving problems as the image-to-image translation problem, and employs plain and non-hierarchical diffusion model with Transformer to generate the solutions conditioned on the initial states/parameters of PDEs. We further propose a multi-scale noise to explicitly guide the diffusion model in capturing information of different frequencies within the solution domain of PDEs. SimDiffPDE achieves a remarkable improvement of +51.4% on the challenging Navier-Stokes equations. In benchmark tests for solving PDEs, such as Darcy Flow, Airfoil, and Pipe for fluid dynamics, as well as Plasticity and Elasticity for solid mechanics, our SimDiffPDE-B achieves significant relative improvements of +21.1%, +11.3%, +15.2%, +25.0%, and +23.4%, respectively. Models and code shall be released upon acceptance.", "keywords": [ "Partial Differential Equations", "Diffusion Model", "Fluid Dynamics", "Machine Learning in PDEs", "Generative Models" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "JQT6iGrXTh", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Haolan Zuo", "gender": "Male", "institution": "Yale University", "country": "", "position": "MS student" }, { "name": "Haoyu Peter Wang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jialin Chen", "gender": "Female", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Pan Li", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Siqi Miao", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Rex Ying", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 26, "n_formula_1": 11, "n_ref_uni": 82, "n_ref": 155, "n_ref_all": 179, "n_fig": 5, "n_tab": 18, "L_tab": 9165, "n_element_tab": 859, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 4581, "n_element_tab_1": 395, "formula_len_all": 1791, "formula_len_all_1": 568, "len_all": 256950, "len_all_1": 76315, "len_abs": 1948, "len_title": 105, "len_sents": 59191, "len_sents_1": 31938, "n_sents": 455, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1955, "title": "GFSE: A Foundational Model For Graph Structural Encoding", "abs": "Foundation models have recently shown remarkable promise by leveraging extensive pre-training on diverse datasets to acquire generalizable representations, which enable effective transfer to a wide range of downstream tasks. In the graph domain, however, most existing pre-training models are tailored to specific domains, primarily due to the inherent differences in semantic meanings of graph features across various contexts. Additionally, most existing models struggle to capture the rich topological complexity of graph structures, leading to inadequate exploration of the embedding space. To address these challenges, we propose a novel Graph Foundational Structural Encoder (GFSE) that identifies universal structural patterns, facilitating a unified feature embedding space suitable for diverse domains, including molecular structures, social networks, and citation networks. GFSE is the first cross-domain graph structural encoder pre-trained with multiple self-supervised learning objectives. Built on a Graph Transformer, GFSE incorporates attention mechanisms biased by graph structural information, allowing it to encode intricate multi-level and fine-grained topological features within complex graph structures. The pre-trained GFSE produces generic and theoretically expressive positional and structural encoding for graphs, which can be seamlessly integrated with various downstream graph feature encoders, including graph neural networks for graphs with vectorized features and Large Language Models for text-attributed graphs. Comprehensive experiments on synthetic and real-world datasets demonstrate GFSE's capability to significantly enhance the model's performance while requiring substantially less task-specific fine-tuning. \nNotably, GFSE boosts the performance by an average margin of 20.48% across eight real-world datasets, highlighting its potential as a powerful and adaptable foundational encoder for graph-structured data.", "keywords": [ "foundation model", "graph representation learning" ], "rating_list": [ 5, 3, 5, 6, 6 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 3, 5 ] }, { "paper_id": "JQQDePbfxh", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Bhuvesh Kumar", "gender": "Male", "institution": "Snap Inc.", "country": "US", "position": "Researcher" }, { "name": "Jamie Heather Morgenstern", "gender": "unknown", "institution": "University of Washington, Seattle", "country": "", "position": "Assistant Professor" }, { "name": "Tao Xiao", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yuanyuan Chloe Yang", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 18, "n_ref_uni": 22, "n_ref": 34, "n_ref_all": 41, "n_fig": 2, "n_tab": 2, "L_tab": 291, "n_element_tab": 22, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 70, "n_element_tab_1": 6, "formula_len_all": 1766, "formula_len_all_1": 1369, "len_all": 124444, "len_all_1": 78230, "len_abs": 1094, "len_title": 111, "len_sents": 36625, "len_sents_1": 33587, "n_sents": 346, "n_sents_1": 320, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 1142, "title": "Private Mechanism Design via Quantile Estimation", "abs": "We investigate the problem of designing differentially private (DP), revenue-maximizing single item auction. Specifically, we consider broadly applicable settings in mechanism design where agents' valuation distributions are **independent**, **non-identical**, and can be either **bounded** or **unbounded**. Our goal is to design such auctions with **pure**, i.e., $(\\epsilon,0)$ privacy in polynomial time. \n\nIn this paper, we propose two computationally efficient auction learning framework that achieves **pure** privacy under bounded and unbounded distribution settings. These frameworks reduces the problem of privately releasing a revenue-maximizing auction to the private estimation of pre-specified quantiles. Our solutions increase the running time by polylog factors compared to the non-private version. As an application, we show how to extend our results to the multi-round online auction setting with non-myopic bidders. To our best knowledge, this paper is the first to efficiently deliver a Myerson auction with **pure** privacy and near-optimal revenue, and the first to provide such auctions for **unbounded** distributions.", "keywords": [ "online auctions", "differential privacy", "mechanism design" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 1 ] }, { "paper_id": "JPnq9wSuSG", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Jiayuan Dong", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Jie Liu", "gender": "Male", "institution": "University of Michigan", "country": "", "position": "Assistant Professor" }, { "name": "Xun Huan", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "Assistant Professor" }, { "name": "Zheyu Zhang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "PhD student" } ], "n_formula": 34, "n_formula_1": 14, "n_ref_uni": 59, "n_ref": 105, "n_ref_all": 126, "n_fig": 8, "n_tab": 11, "L_tab": 9997, "n_element_tab": 1231, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 1802, "n_element_tab_1": 1187, "formula_len_all": 4091, "formula_len_all_1": 1028, "len_all": 190708, "len_all_1": 70749, "len_abs": 1275, "len_title": 153, "len_sents": 41688, "len_sents_1": 24049, "n_sents": 317, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1289, "title": "Adaptive Causal Experimental Design: Amortizing Sequential Bayesian Experimental Design for Causal Models", "abs": "Interventions are essential for causal discovery and causal reasoning. Acquiring interventional data, however, is often costly, especially in real-world systems.\nA careful experimental design can therefore bring substantial savings. \nIn the sequential experimental design setting, \nmost existing approaches seek the best \ninterventions in a greedy (myopic) manner that does not account for the synergy from the yet-to-come future experiments. We propose Adaptive Causal Experimental Design (ACED),\na novel Bayesian sequential design framework for learning a design policy capable of generating non-myopic interventions that incorporate the effect on future experiments.\nIn particular, ACED maximizes the Expected Information Gain (EIG) on flexible choices of causal quantities of interest (e.g., causal structure, specific causal effects) directly, bypassing the need for computing intermediate posteriors in the experimental sequence.\nLeveraging a variational lower bound estimator for the EIG, ACED trains an amortized policy network that can be executed rapidly during deployment. \nWe present numerical results demonstrating ACED's effectiveness on synthetic datasets with both linear and nonlinear structural causal models, as well as on in-silico single-cell gene expression datasets.", "keywords": [ "Causal Discovery; Causal Reasoning; Bayesian Experimental Design; Amortized Variational Inference; Expected Information Gain; Mutual Information Lower Bounds; Adaptive Experiments" ], "rating_list": [ 5, 6, 5, 3, 6 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 1, 4 ], "confidence_list": [ 4, 4, 3, 3, 2 ] }, { "paper_id": "JOHhktXd4a", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Charless Fowlkes", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Full Professor" }, { "name": "Kyungmin Kim", "gender": "Female", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Pierre Baldi", "gender": "Male", "institution": "University of California, Irvine", "country": "", "position": "Full Professor" }, { "name": "Roy Fox", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Assistant Professor" }, { "name": "John Banister Lanier", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 53, "n_ref": 121, "n_ref_all": 151, "n_fig": 22, "n_tab": 1, "L_tab": 271, "n_element_tab": 21, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 854, "formula_len_all_1": 375, "len_all": 174246, "len_all_1": 71492, "len_abs": 1650, "len_title": 141, "len_sents": 55815, "len_sents_1": 34310, "n_sents": 429, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1657, "title": "Make the Pertinent Salient: Task-Relevant Reconstruction for Visual Control with Distractions", "abs": "Recent advancements in Model-Based Reinforcement Learning (MBRL) have made it a powerful tool for visual control tasks. Despite improved data efficiency, it remains challenging to train MBRL agents with generalizable perception. Training in the presence of visual distractions is particularly difficult due to the high variation they introduce to representation learning. Building on DREAMER, a popular MBRL method, we propose a simple yet effective auxiliary task to facilitate representation learning in distracting environments. Under the assumption that task-relevant components of image observations are straightforward to identify with prior knowledge in a given task, we use a segmentation mask on image observations to only reconstruct task-relevant components. In doing so, we greatly reduce the complexity of representation learning by removing the need to encode task-irrelevant objects in the latent representation. Our method, Segmentation Dreamer (SD), can be used either with ground-truth masks easily accessible in simulation or by leveraging potentially imperfect segmentation foundation models. The latter is further improved by selectively applying the reconstruction loss to avoid providing misleading learning signals due to mask prediction errors. In modified DeepMind Control suite (DMC) and Meta-World tasks with added visual distractions, SD achieves significantly better sample efficiency and greater final performance than prior work. We find that SD is especially helpful in sparse reward tasks otherwise unsolvable by prior work, enabling the training of visually robust agents without the need for extensive reward engineering.", "keywords": [ "Visual Control", "Robust Representation Learning", "Model-Based Reinforcement Learning" ], "rating_list": [ 5, 5, 6, 6, 5 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 3, 4, 4 ] }, { "paper_id": "JOBokGDcX0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Joel Rixen", "gender": "unknown", "institution": "Christian-Albrechts-Universit\u00e4t Kiel", "country": "DE", "position": "PhD student" }, { "name": "Matthias Renz", "gender": "Male", "institution": "Christian-Albrechts-Universit\u00e4t Kiel", "country": "DE", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 27, "n_ref": 40, "n_ref_all": 55, "n_fig": 5, "n_tab": 3, "L_tab": 408, "n_element_tab": 59, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 359, "n_element_tab_1": 37, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 93579, "len_all_1": 56837, "len_abs": 812, "len_title": 115, "len_sents": 33361, "len_sents_1": 29020, "n_sents": 244, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 816, "title": "On Sequence Segmentation with overlapped Chunks in Machine Learning", "abs": "Operating on very long sequences can be problematic for many sequence modelling methods like Transformers or recurrent neural networks. To avoid this issue, long sequences are often split into smaller chunks instead.\nFor various reasons, these chunks typically are overlapped with each other which causes an increase in tensor size by however much the chunks are overlapping.\n\nThis paper attempts to find a better understanding on overlapped sequence chunks and what they accomplish. Specifically, the focus of this paper is on audio inputs in both the time and frequency domain. Previous models for speech separation and audio super resolution which use overlapped chunks are modified to allow for reduced or even removed overlaps which causes significant decreases in computational cost while maintaining accuracy.", "keywords": [ "sequence segmentation", "speech separation", "source separation", "audio super resolution", "stft", "signal processing" ], "rating_list": [ 1, 1, 3, 5 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 5, 5, 3, 3 ] }, { "paper_id": "JNsac6zbg2", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Joni Pajarinen", "gender": "unknown", "institution": "Aalto University", "country": "FI", "position": "Associate Professor" }, { "name": "Tinghuai Wang", "gender": "unknown", "institution": "Huawei Helsinki Research Center", "country": "FI", "position": "Research Leader" }, { "name": "Huiling Wang", "gender": "unknown", "institution": "Aalto University", "country": "FI", "position": "Postdoc" } ], "n_formula": 33, "n_formula_1": 30, "n_ref_uni": 39, "n_ref": 112, "n_ref_all": 146, "n_fig": 6, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2203, "formula_len_all_1": 1915, "len_all": 140450, "len_all_1": 65583, "len_abs": 1246, "len_title": 134, "len_sents": 46627, "len_sents_1": 30987, "n_sents": 335, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1252, "title": "Uncertainty-Regularized Diffusional Subgoals for Hierarchical Reinforcement Learning", "abs": "Hierarchical reinforcement learning (HRL) aims to solve complex tasks by making decisions across multiple levels of temporal abstraction. However, off-policy training of hierarchical policies faces non-stationarity issues because the low-level policy is constantly changing, which makes it difficult for the high-level policy that generates subgoals to adapt. In this paper, we propose a conditional diffusion model-based approach for subgoal generation to mitigate these non-stationarity challenges. Specifically, we employ a Gaussian Process (GP) prior on subgoal generation as a surrogate distribution to regularize the diffusion policy and inform the diffusion process about uncertain areas in the action space. We introduce adaptive inducing states to facilitate sparse GP-based subgoal generation, enhancing sample efficiency and promoting better exploration in critical regions of the state space. Building on this framework, we develop an exploration strategy that identifies promising subgoals based on the learned predictive distribution of the diffusional subgoals. Experimental results demonstrate significant improvements in both sample efficiency and performance on challenging continuous control benchmarks compared to prior HRL methods.", "keywords": [ "Hierarchical Reinforcement Learning" ], "rating_list": [ 6, 3, 3, 3, 5 ], "soundness_list": [ 3, 3, 2, 1, 3 ], "presentation_list": [ 2, 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 2, 3, 3, 3, 3 ] }, { "paper_id": "JNhU9NeOFr", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Daniel Russo", "gender": "Male", "institution": "Columbia University", "country": "", "position": "Associate Professor" }, { "name": "Hongseok Namkoong", "gender": "Male", "institution": "LinkedIn", "country": "", "position": "Researcher" }, { "name": "Tiffany Cai", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Kelly Zhang", "gender": "Female", "institution": "Imperial College London", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 26, "n_ref_uni": 43, "n_ref": 65, "n_ref_all": 79, "n_fig": 6, "n_tab": 3, "L_tab": 880, "n_element_tab": 56, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 741, "n_element_tab_1": 42, "formula_len_all": 1533, "formula_len_all_1": 1435, "len_all": 140557, "len_all_1": 77747, "len_abs": 1246, "len_title": 92, "len_sents": 34819, "len_sents_1": 31712, "n_sents": 303, "n_sents_1": 277, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 1251, "title": "Informed Exploration via Generative Modeling", "abs": "Conventionally trained neural networks excel at prediction but often struggle to model uncertainty in their own predictions. We explore this challenge in a meta-learning bandit decision-making problem for news recommendations; this setting require decision-making algorithms to incorporate pretrained language models to process text data for the best performance. We present a scalable approach to Bayesian uncertainty quantification by posing it as a problem of autoregressive generative modeling of future rewards. First, we use historical data on previously released news articles to pre-train a generative model to predict sequences of future potential rewards. At inference time, our algorithm makes decisions based on limited previous rewards and autoregressively generated future rewards. Far from a heuristic, we synthesize insights from the literature to show our method is a novel implementation of Thompson (posterior) sampling, a prominent bandit algorithm. We prove our pretraining loss directly controls online decision-making performance, and we demonstrate our framework on a news recommendation task where we integrate end-to-end fine-tuning of a pretrained language model to process news article headline text to improve performance.", "keywords": [ "bandit algorithms", "Thompson sampling", "bayesian inference", "generative models" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 2, 4 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "JNh8CCDugm", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Cong Lei", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Jinghao Liu", "gender": "unknown", "institution": "University of Technology Sydney", "country": "", "position": "PhD student" }, { "name": "Linchao Zhu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaojun Chang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 5, "n_ref_uni": 57, "n_ref": 127, "n_ref_all": 148, "n_fig": 8, "n_tab": 13, "L_tab": 3535, "n_element_tab": 355, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 571, "n_element_tab_1": 104, "formula_len_all": 1056, "formula_len_all_1": 496, "len_all": 199188, "len_all_1": 64321, "len_abs": 1468, "len_title": 118, "len_sents": 58989, "len_sents_1": 30613, "n_sents": 448, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1475, "title": "Retrieval Information Injection for Enhanced Medical Report Generation", "abs": "Automatically generating medical reports is an effective solution to the diagnostic bottleneck caused by physician shortage. Existing methods have demonstrated exemplary performance in generating high-textual-quality reports. Due to the high similarity among medical images as well as the structural and content homogeneity of medical reports, these methods often make it difficult to fully capture the semantic information in medical images. To address this issue, we propose a training-free Retrieval Information injectioN (RIN) method by simulating the process of Multidisciplinary Consultation. The essence of this method lies in fully utilizing similar reports of target images to enhance the performance of pre-trained medical report generation models. Specifically, we first retrieve images most similar to the target image from a pre-constructed image feature database. Then, the reports corresponding to these images are inputted into a report generator of the pre-trained model, obtaining the distributions of retrieved reports. RIN generates final reports by integrating prediction distributions of the pre-trained model and the average distributions of retrieved reports, thereby enhancing the accuracy and reliability of the generated report. Comprehensive experimental results demonstrate that RIN significantly enhances clinical efficacy in chest X-rays report generation task. Compared to the current state-of-the-art methods, it achieves competitive results.", "keywords": [ "Medical Report Generation", "Contrastive Decoding", "Training-free" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "JNZdhbDBUH", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Chutong Yang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Haijun Yang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "hui Zhang", "gender": "Male", "institution": "Beihang University", "country": "", "position": "Full Professor" }, { "name": "Zhenyuanhuang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 34, "n_formula_1": 22, "n_ref_uni": 33, "n_ref": 61, "n_ref_all": 92, "n_fig": 11, "n_tab": 7, "L_tab": 5882, "n_element_tab": 1322, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 3012, "n_element_tab_1": 1042, "formula_len_all": 6112, "formula_len_all_1": 2546, "len_all": 158178, "len_all_1": 67805, "len_abs": 1819, "len_title": 132, "len_sents": 41650, "len_sents_1": 23773, "n_sents": 321, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 972, "title": "A General Aggregation Federated Learning Intervention Algorithm based on $do$-Calculus", "abs": "This article explores federated long-tail learning (Fed-LT) tasks, which involve clients with private and heterogeneous data that exhibit a long-tail distribution. We propose two methods: (a) Client Re-weighted Prior Analyzer (CRePA), which balances the global model's performance on tail and non-tail categories and enhances performance on tail categories while maintaining it on non-tail categories. (b) Federated Long-Tail Causal Intervention Model (FedLT-CI) computes clients' causal effects on the global model's performance in the tail and enhances the interpretability of Fed-LT. CRePA achieves state-of-the-art performance, and FedLT-CI improves tail performance significantly without affecting non-tail performance. Extensive experiments indicate that CRePA achieved SOTA performance compared to other baselines on CIFAR-10-LT and CIFAR-100-LT. Applying the FedLT-CI to all baselines significantly improved tail performance without affecting non-tail performance.", "keywords": [ "Federated Learning", "do-Calculus", "causal" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "JNZ3Om6NPS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 1, "author_info_list": [ { "name": "Serge Berger", "gender": "Male", "institution": "AIDIL", "country": "US", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 6, "n_ref_uni": 27, "n_ref": 44, "n_ref_all": 53, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 563, "formula_len_all_1": 266, "len_all": 106610, "len_all_1": 60506, "len_abs": 2135, "len_title": 95, "len_sents": 37476, "len_sents_1": 29156, "n_sents": 354, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1989, "title": "On inherent limitations of GPT/LLM \\\\ Architecture", "abs": "This paper shows that reasoning/proving issues of $GPT/LLM$ are an inherent logical consequence of the architecture. Namely, they are due to a schema of its prediction mechanism of the next token in a sequence, and randomization involved in the process.\n \nAfter the natural formalization of the problem into a domain of finite graphs, $G({\\omega})$, we prove the following general theorem:\n\nFor almost all proofs, any learning algorithm of inference, that uses randomization in $G({\\omega})$, and necessitates veracity of inference, is almost surely a literal learning.\n\nIn the context, \"literal learning\" stands for one which is either vacuous, i.e. $\\forall x~[P(x) \\implies Q(x)]$ where $P(x)$ is false for every $x$, or create a random inference from a false assumption (hallucination), or it essentially memorizes the inferences from training/synthetic data.\n\nA few corollaries follow. For instance, if its formulation is somewhat original, it is easy to notice the issue of solving mathematical problems with $LLMs$ in the case of even low-complexity tasks. Since its solution is unlikely to be found in a holistic form in a training dataset, a correct proof is not to be expected. \n\nIt is because, in a rigorous context, $GPT$ has exponentially decreasing odds of finding a valid proof of the result unless it simply \u201crepeats\u201d a known proof, perhaps with trivial modifications. Another observation is that the degradation has an exponential rate by the length of a proof. In other words, an attempt to prove a complex enough statement virtually has no chance to be \nsuccessful.\n\nIn a novel rigorous context (i.e., when $GPT$-based architecture is looking to prove a new result, for instance, a hypothesis), that is virtually impossible even for a long enough fragment. The probability of success becomes infinitesimal quickly for either a fragment of possible proof or a weaker non-trivial statement. That also was empirically shown for data mixtures and confirmed experimentally.", "keywords": [ "0-1 laws", "first-order logic", "probabilistic spaces", "finite graphs" ], "rating_list": [ 1, 1, 3, 3 ], "soundness_list": [ 1, 1, 2, 1 ], "presentation_list": [ 1, 1, 1, 1 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "JMgxtZqkvO", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Aditya Bhaskara", "gender": "Male", "institution": "University of Utah", "country": "US", "position": "Associate Professor" }, { "name": "Cen-Jhih Li", "gender": "Male", "institution": "University of Utah", "country": "US", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 8, "n_ref_uni": 73, "n_ref": 109, "n_ref_all": 143, "n_fig": 2, "n_tab": 11, "L_tab": 6287, "n_element_tab": 919, "n_fig_1": 1, "n_tab_1": 8, "L_tab_1": 736, "n_element_tab_1": 182, "formula_len_all": 863, "formula_len_all_1": 412, "len_all": 211340, "len_all_1": 70642, "len_abs": 1090, "len_title": 114, "len_sents": 53485, "len_sents_1": 31362, "n_sents": 441, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1093, "title": "Memory-Efficient Fine-Tuning via Structured Neural Network Pruning", "abs": "Fine-tuning is an important step in adapting foundation models such as large language models to downstream tasks. To make this step more accessible to users with limited computational budgets, it is crucial to develop fine-tuning methods that are memory and computationally efficient. Sparse Fine-tuning (SFT) and Low-rank adaptation (LoRA) are two frameworks that have emerged for addressing this problem, and have been adopted widely in practice. In this work, we develop a new SFT framework, based on ideas from neural network pruning. At a high level, we first identify \"important\" neurons/nodes using feature importance metrics from network pruning (specifically, we use the structural pruning method), and then perform fine-tuning by restricting to weights involving these neurons. Using experiments on both vision and language tasks, we demonstrate that our method significantly improves the memory efficiency of SFT without increasing training time complexity and implementation complexity, while achieving accuracy comparable to state-of-the-art methods such as LoRA and its variants.", "keywords": [ "Transformer", "Fine-tuning", "Memory efficient learning" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "JMe5FW8c3e", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Brahma S Pavse", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "PhD student" }, { "name": "Qiaomin Xie", "gender": "Female", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Yudong Chen", "gender": "Male", "institution": "Department of Computer Sciences, University of Wisconsin - Madison", "country": "US", "position": "Associate Professor" }, { "name": "Josiah Hanna", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "Assistant Professor" } ], "n_formula": 64, "n_formula_1": 17, "n_ref_uni": 53, "n_ref": 146, "n_ref_all": 193, "n_fig": 10, "n_tab": 4, "L_tab": 2, "n_element_tab": 1, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 260, "n_element_tab_1": 4, "formula_len_all": 7533, "formula_len_all_1": 1623, "len_all": 251154, "len_all_1": 71977, "len_abs": 1527, "len_title": 126, "len_sents": 73064, "len_sents_1": 31502, "n_sents": 682, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1569, "title": "Stable Offline Value Function Learning with Bisimulation-based Representations", "abs": "In reinforcement learning, offline value function learning is the procedure of using an offline dataset to estimate the expected discounted return from each state when taking actions according to a fixed target policy. The stability of this procedure, i.e., whether it converges to its fixed-point, critically depends on the representations of the state-action pairs. Poorly learned representations can make value function learning unstable, or even divergent. Therefore, it is critical to stabilize value function learning by explicitly shaping the state-action representations. Recently, the class of bisimulation-based algorithms have shown promise in shaping representations for control. However, it is still unclear if this class of methods can \\emph{stabilize} value function learning. In this work, we investigate this question and answer it affirmatively. We introduce a bisimulation-based algorithm called kernel representations for offline policy evaluation (\\textsc{krope}). \\textsc{krope} uses a kernel to shape state-action representations such that state-action pairs that have similar immediate rewards and lead to similar next state-action pairs under the target policy also have similar representations. We show that \\textsc{krope}: 1) learns stable representations and 2) leads to lower value error than baselines. Our analysis provides new theoretical insight into the stability properties of bisimulation-based methods and suggests that practitioners can use these methods for stable and accurate evaluation of offline reinforcement learning agents.", "keywords": [ "reinforcement learning", "representation learning", "off-policy", "offline policy evaluation", "bisimulations", "stability", "value function learning", "abstractions" ], "rating_list": [ 3, 6, 8, 6 ], "soundness_list": [ 2, 4, 4, 3 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 3, 5, 3 ] }, { "paper_id": "JMPOqoe4tl", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Antonio Emanuele Cin\u00e0", "gender": "Male", "institution": "University of Genoa", "country": "", "position": "Assistant Professor" }, { "name": "Battista Biggio", "gender": "Male", "institution": "University of Cagliari, Italy", "country": "IT", "position": "Full Professor" }, { "name": "Francesco Villani", "gender": "Male", "institution": "University of Genoa", "country": "IT", "position": "PhD student" }, { "name": "Lea Sch\u00f6nherr", "gender": "Female", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Assistant Professor" }, { "name": "Marcello Pelillo", "gender": "Male", "institution": "Ca' Foscari University of Venice", "country": "IT", "position": "Professor" }, { "name": "Maura Pintor", "gender": "Female", "institution": "University of Cagliari", "country": "IT", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 33, "n_ref": 89, "n_ref_all": 141, "n_fig": 5, "n_tab": 14, "L_tab": 10431, "n_element_tab": 666, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2906, "n_element_tab_1": 202, "formula_len_all": 842, "formula_len_all_1": 792, "len_all": 161138, "len_all_1": 71707, "len_abs": 1198, "len_title": 117, "len_sents": 47388, "len_sents_1": 32781, "n_sents": 314, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1269, "title": "$\\sigma$-zero: Gradient-based Optimization of $\\ell_0$-norm Adversarial Examples", "abs": "Evaluating the adversarial robustness of deep networks to gradient-based attacks is challenging.\nWhile most attacks consider $\\ell_2$- and $\\ell_\\infty$-norm constraints to craft input perturbations, only a few investigate sparse $\\ell_1$- and $\\ell_0$-norm attacks.\nIn particular, $\\ell_0$-norm attacks remain the least studied due to the inherent complexity of optimizing over a non-convex and non-differentiable constraint.\nHowever, evaluating adversarial robustness under these attacks could reveal weaknesses otherwise left untested with more conventional $\\ell_2$- and $\\ell_\\infty$-norm attacks.\nIn this work, we propose a novel $\\ell_0$-norm attack, called $\\sigma$-zero, which leverages a differentiable approximation of the $\\ell_0$ norm to facilitate gradient-based optimization, and an adaptive projection operator to dynamically adjust the trade-off between loss minimization and perturbation sparsity.\nExtensive evaluations using MNIST, CIFAR10, and ImageNet datasets, involving robust and non-robust models, show that $\\sigma$-zero finds minimum $\\ell_0$-norm adversarial examples without requiring any time-consuming hyperparameter tuning, and that it outperforms all competing sparse attacks in terms of success rate, perturbation size, and efficiency.", "keywords": [ "adversarial examples", "sparse attacks", "gradient-based attack", "machine learning security", "evasion attacks", "sparse optimization" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "JMNht3SmcG", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Bingning Wang", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Hongyu Lin", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Qingyu Zhang", "gender": "unknown", "institution": "Fuzhou University", "country": "CN", "position": "Undergrad student" }, { "name": "Xianpei Han", "gender": "Male", "institution": "Institute of Software, CAS", "country": "CN", "position": "Professor" }, { "name": "Xin Men", "gender": "Male", "institution": "Beijing University of Post and Telecommunication", "country": "", "position": "MS student" }, { "name": "Yaojie Lu", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "weipeng chen", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "\u540d\u5b87 \u5f90", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 46, "n_ref": 104, "n_ref_all": 135, "n_fig": 5, "n_tab": 6, "L_tab": 3503, "n_element_tab": 201, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 1255, "n_element_tab_1": 99, "formula_len_all": 727, "formula_len_all_1": 64, "len_all": 121553, "len_all_1": 51697, "len_abs": 1193, "len_title": 124, "len_sents": 29047, "len_sents_1": 22669, "n_sents": 234, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1200, "title": "ShortGPT: Layers in Large Language Models are More Redundant Than You Expect", "abs": "As Large Language Models (LLMs) continue to advance in performance, their size has increased significantly, with current LLMs containing billions or even trillions of parameters. In this study, we identify notable redundancy across the layers of LLMs, where some layers contribute minimally to overall network functionality. To quantify this, we introduce a metric called Block Influence (BI) which use the similarity between layer's input and output to measure the importance of each layer. Based on the observation of layer redundancy, we propose a straightforward pruning method: layer removal, which eliminates redundant layers based on their BI scores. Our approach, termed ShortGPT, demonstrates superior performance over previous state-of-the-art pruning methods. Moreover, ShortGPT is orthogonal to quantization-like methods, enabling further reduction in parameters and computation. The ability to achieve better results through simple layer removal, as opposed to more complex pruning techniques, suggests a high degree of redundancy across layers, not only in transformer models but also in non-transformer models. We hope this work will contribute to future research in LLM compression.", "keywords": [ "Large Language Model", "Model pruning", "Layer redundancy" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "JLDAWbzTUg", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Anqi Yu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Jin Li", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "MS student" }, { "name": "John Lui", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Xutong Liu", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Dai Xiangxiang", "gender": "Male", "institution": "Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 79, "n_formula_1": 12, "n_ref_uni": 53, "n_ref": 93, "n_ref_all": 182, "n_fig": 15, "n_tab": 13, "L_tab": 4252, "n_element_tab": 368, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 305, "n_element_tab_1": 40, "formula_len_all": 8243, "formula_len_all_1": 1032, "len_all": 281080, "len_all_1": 76965, "len_abs": 6919, "len_title": 118, "len_sents": 92201, "len_sents_1": 37004, "n_sents": 694, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1853, "title": "Cost-Effective Online Multi-LLM Selection with Versatile Reward Models", "abs": "With the rapid advancement of large language models (LLMs), the diversity of multi-LLM tasks and the variability in their pricing structures have become increasingly important, as costs can vary greatly between different LLMs. To tackle these challenges, we introduce the \\textit{C2MAB-V}, a \\underline{C}ost-effective \\underline{C}ombinatorial \\underline{M}ulti-armed \\underline{B}andit with \\underline{V}ersatile reward models for optimal LLM selection and usage. This online model differs from traditional static approaches or those reliant on a single LLM without cost consideration. With multiple LLMs deployed on a scheduling cloud and a local server dedicated to handling user queries, \\textit{C2MAB-V} facilitates the selection of multiple LLMs over a combinatorial search space, specifically tailored for various collaborative task types with different reward models. Based on our designed online feedback mechanism and confidence bound technique, \\textit{C2MAB-V} can effectively address the multi-LLM selection challenge by managing the exploration-exploitation trade-off across different models, while also balancing cost and reward for diverse tasks. The NP-hard integer linear programming problem for selecting multiple LLMs with trade-off dilemmas is addressed by: i) decomposing the integer problem into a relaxed form by the local server, ii) utilizing a discretization rounding scheme that provides optimal LLM combinations by the scheduling cloud, and iii) continual online updates based on feedback. Theoretically, we prove that \\textit{C2MAB-V} offers strict guarantees over versatile reward models, matching state-of-the-art results for regret and violations in some degenerate cases. Empirically, we show that \\textit{C2MAB-V} effectively balances performance and cost-efficiency with nine LLMs for three application scenarios.", "keywords": [ "Combinatorial multi-armed bandit", "Online learning", "Large language model", "Long-term cost constraint" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "JL18agpSc3", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Adri\u00e1n Carrasco Revilla", "gender": "Male", "institution": "University of Oxford", "country": "BE", "position": "Researcher" }, { "name": "Chenhao Xue", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Joyraj Chakraborty", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Min Chen", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Yuanzhe Jin", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 26, "n_ref": 35, "n_ref_all": 69, "n_fig": 21, "n_tab": 14, "L_tab": 8304, "n_element_tab": 578, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 4111, "n_element_tab_1": 186, "formula_len_all": 701, "formula_len_all_1": 280, "len_all": 183535, "len_all_1": 53759, "len_abs": 1019, "len_title": 83, "len_sents": 58091, "len_sents_1": 22362, "n_sents": 432, "n_sents_1": 146, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1022, "title": "AutoGeTS: Automated Generation of Text Synthetics for Improving Text Classification", "abs": "When developing text classification models for real world applications, one major challenge is the difficulty to collect sufficient data for all text classes. In this work, we address this challenge by utilizing large language models (LLMs) to generate synthetic data and using such data to improve the performance of the models without waiting for more real data to be collected and labelled. As an LLM generates different synthetic data in response to different input examples, we formulate an automated workflow, which searches for input examples that lead to more \"effective'' synthetic data for improving the model concerned. We study three search strategies with an extensive set of experiments, and use experiment results to inform an ensemble algorithm that selects a search strategy according to the characteristics of a class. Our further experiments demonstrate that this ensemble approach is more effective than each individual strategy in our automated workflow for improving classification models using LLMs.", "keywords": [ "Text Classification", "Synthetic Data", "Data Augmentation", "Large Language Model", "Text Analysis", "Optimization" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "JJepij22fb", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Chao Yu", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Dong Yan", "gender": "Male", "institution": "Baichuan Intelligent Technology", "country": "", "position": "Researcher" }, { "name": "Jialian Li", "gender": "Female", "institution": "Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Jian Xie", "gender": "Male", "institution": "Beijing Baichuan Intelligent Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Xingzhou Lou", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "YipinZhang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Yu Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yuan Shen", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Yuzi Yan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 51, "n_ref": 84, "n_ref_all": 125, "n_fig": 16, "n_tab": 4, "L_tab": 1424, "n_element_tab": 131, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 981, "formula_len_all_1": 624, "len_all": 202644, "len_all_1": 69440, "len_abs": 1898, "len_title": 74, "len_sents": 64100, "len_sents_1": 33625, "n_sents": 472, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 26, "L_abs": 1503, "title": "Reward-Robust RLHF in LLMs", "abs": "As Large Language Models continue to progress toward more advanced forms of intelligence, Reinforcement Learning from Human Feedback is increasingly seen as a key pathway toward achieving Artificial General Intelligence. However, the reliance on reward-model-based alignment methods introduces significant challenges due to the inherent instability and imperfections of Reward Models (RMs), which can lead to critical issues such as reward hacking and misalignment with human intentions. In this paper, we introduce a reward-robust RLHF framework aimed at addressing these fundamental challenges, paving the way for more reliable and resilient learning in LLMs. Our approach introduces a novel optimization objective that carefully balances performance and robustness by incorporating Bayesian Reward Model Ensembles to model the uncertainty set of reward functions. This allows the framework to integrate both nominal performance and minimum reward signals, ensuring more stable learning even with imperfect RMs. Empirical results demonstrate that our framework consistently outperforms baselines across diverse benchmarks, showing improved accuracy and long-term stability. We also provide a theoretical analysis, demonstrating that reward-robust RLHF approaches the stability of constant reward settings, which proves to be acceptable even in a stochastic-case analysis. Together, these contributions highlight the framework\u2019s potential to enhance both the performance and stability of LLM alignment.", "keywords": [ "RLHF", "LLM", "robustness", "alignment" ], "rating_list": [ 6, 5, 1, 5 ], "soundness_list": [ 4, 2, 1, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "JJH7m9v4tv", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Wenyu Ke", "gender": "unknown", "institution": "Wuyi University", "country": "CN", "position": "MS student" }, { "name": "Yikui Zhai", "gender": "Male", "institution": "Wuyi University", "country": "", "position": "Full Professor" }, { "name": "ZhiHao Long", "gender": "unknown", "institution": "Wuyi University", "country": "CN", "position": "MS student" }, { "name": "zhangyang", "gender": "Male", "institution": "Wuyi University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 51, "n_formula_1": 31, "n_ref_uni": 20, "n_ref": 27, "n_ref_all": 36, "n_fig": 2, "n_tab": 11, "L_tab": 4489, "n_element_tab": 277, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5828, "formula_len_all_1": 1808, "len_all": 129923, "len_all_1": 47375, "len_abs": 965, "len_title": 147, "len_sents": 36509, "len_sents_1": 19478, "n_sents": 285, "n_sents_1": 132, "venue": "ICLR2025", "decision": "withdraw", "L_title": 99, "L_abs": 994, "title": "Post-hoc Discriminator Guidance For Data-efficient Image Generation Via Annealing Langevin Dynamics", "abs": "The proposed method, post-hoc discriminator guidance (PDG) aims to take an alternate route for Nash non-equilibrium issue in GANs' training. This method introduces an additional discriminator that gives explicit supervision with regard to gradient of density ratio $\\nabla_{x} \\log_{}{\\frac{p_{r}(x)}{p_{f}(x)}}$ between real and fake probability density function, steering the sample path towards more realistic regions in a post-hoc way. We train the discriminator after adversarial optimization, making post-hoc discriminator training stable and fast to converge. In generation process, annealing Langevin dynamics sampling with density ratio score reduces the Kullback-Leibler divergence between the true and generated samples. Given an optimal discriminator, the method can improve the sampling quality of various off-the-shelf models on the web without retraining required. Extensive experiments validate the advancements and effectiveness of PDG on content-varying data-limited datasets.", "keywords": [ "post-hoc discriminator guidance", "GANs", "data-efficient image generation", "annealing Langevin dynamics" ], "rating_list": [ 1, 3, 3, 5 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 1, 1, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "JJ46kIfPio", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Chuanhao Li", "gender": "unknown", "institution": "Yale University", "country": "", "position": "Postdoc" }, { "name": "Siyu Chen", "gender": "Male", "institution": "Yale University", "country": "", "position": "PhD student" }, { "name": "Yuanxin Liu", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Zhuoran Yang", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 122, "n_formula_1": 25, "n_ref_uni": 26, "n_ref": 72, "n_ref_all": 87, "n_fig": 1, "n_tab": 4, "L_tab": 2346, "n_element_tab": 265, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 16489, "formula_len_all_1": 2485, "len_all": 165190, "len_all_1": 70793, "len_abs": 1406, "len_title": 119, "len_sents": 47439, "len_sents_1": 31883, "n_sents": 419, "n_sents_1": 242, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1147, "title": "Steer a Crowd: Learning to Persuade a Population in a Stackelberg Game", "abs": "Multi-agent systems are prevalent across various domains, characterized by misaligned objectives and information asymmetry, which facilitate the study of incentive design and information design. Existing research often assumes known models and static environments. Motivated by this, we propose a Dynamic Incentive and Information Design (DIID) framework for finite-horizon Markov games, involving a principal and multiple agents. \nOur focus is on how the principal learns their optimal policy based on data generated through interactions with agents.\nThe main challenge lies in balancing the principal's regret and violations of agents' incentive compatibility constraints during interactions. We establish a lower bound characterizing the trade-off between the two objectives and propose an algorithm attaining the optimal trade-off, i.e. $\\tilde{\\mathcal{O}}(T^{2/3})$ regret and constraint violation. Additionally, with access to additional unilateral deviation information of the agents, we propose an algorithm attaining improved guarantees that achieve $\\tilde{\\mathcal{O}}(T^{1/2})$ for both regret and constraint violation simultaneously.", "keywords": [ "information design", "stackelberg game", "reinforcement learning", "markov games", "multi-agent system" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "JIlIYIHMuv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 1, "author_info_list": [ { "name": "Baiqiao Yin", "gender": "Not Specified", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 13, "n_ref": 37, "n_ref_all": 47, "n_fig": 3, "n_tab": 7, "L_tab": 1962, "n_element_tab": 159, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 2000, "n_element_tab_1": 200, "formula_len_all": 203, "formula_len_all_1": 203, "len_all": 130010, "len_all_1": 62915, "len_abs": 2098, "len_title": 137, "len_sents": 26289, "len_sents_1": 26074, "n_sents": 171, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1993, "title": "LVLM-CL: Make Large Vision-Language Models Work Better under Continual Learning Settings", "abs": "The development of Large Vision-Language Models\n(LVLMs) is striving to catch up with the success of Large\nLanguage Models (LLMs), yet it faces more challenges to\nbe resolved. When finetuning LVLMs with user-specific\ndata in the practical use, the pretrained weights would face\nthe problems of forgetting and performance degradation.\nSo it is important to improve LVLM\u2019s performance under\nthe continual learning settings. Some existing CL methods\nhave explored continual learning on VLM. However, the\ncontinual learning settings they have proposed couldn\u2019t\nbe adopted to LVLMs smoothly because the training and\nfinetuning process of LVLMs need amount of data while\nprevious VLM continual learning settings built on limited\ndata and different model architectures. In this work, we first\ndevise a task-specific continual learning setting especially\nfor LVLMs by classifying the instruction tuning data\nfor the second finetune process of LVLMs into several\ndifferent tasks. Mimicking the process of finetuning with\nuser-specific task data, we found that the performance of\nLVLMs would decline without any modules designed for\ncontinual learning settings. So we present LVLM-CL, a\nnovel approach capable of continual learning settings for\nlarge vision-language models when finetuning with different\nkinds of tasks. Specifically, our LVLM-CL consists of a\ntext feature based prompt that are different between tasks\nto keep the special feature of different tasks. To meet the\nsetting of continual learning, we also design a memory bank\nwhich storage previous trained tasks which helps LVLMs\napply knowledge to unfamiliar combinations. Extensive case\nstudies and quantitative evaluations show LVLM-CL has\nstrong capability in understanding the pivotal features of\ndifferent tasks and emerges impressive memory capabilities\nunder the continual learning settings. This work fosters the\nadvancements of LVLMs by enabling them to support better\ncontinual finetuning toward practical use in the real world.", "keywords": [ "Continual Learning; Large vision-language model" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 3, 4, 2, 2 ], "presentation_list": [ 1, 2, 1, 2 ], "contribution_list": [ 1, 3, 1, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "JIePBlcFg0", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Duxin Chen", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Wenwu Yu", "gender": "Male", "institution": "Southeast University", "country": "", "position": "Full Professor" }, { "name": "Xiaoyi Liu", "gender": "unknown", "institution": "Southeast University", "country": "", "position": "MS student" }, { "name": "He Yanyan", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 9, "n_ref_uni": 35, "n_ref": 72, "n_ref_all": 93, "n_fig": 5, "n_tab": 12, "L_tab": 4513, "n_element_tab": 504, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1131, "n_element_tab_1": 223, "formula_len_all": 1632, "formula_len_all_1": 544, "len_all": 180958, "len_all_1": 64498, "len_abs": 1739, "len_title": 167, "len_sents": 67131, "len_sents_1": 29314, "n_sents": 467, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 118, "L_abs": 1748, "title": "iQR: Quantile Regression with QR Orthogonal Decomposition for Resource Scheduling Optimization without Empirical Model", "abs": "Optimal resource scheduling aims to cover resource demand with minimum economic cost, which is far from model-based constraint optimization or data-driven prediction based scheduling process. To address this model-free constraint optimization issue, a sparse system identification framework with quantile regression and QR orthogonal decomposition~(iQR) is proposed for complex systems ranging from small to large scales. It leverages quantile optimization with $L_1$-norm to construct a business-driven strategy to reach the proportion of meeting resource demand. It also involves a complete-mapping Fourier Transformation process and an orthogonal least squares technique to select basis vectors in advance to achieve fast regression with sparse mathematical expression, which reduces the number of basis functions from thousands to hundreds and to dozens. iQR represents a specific expectation for single time series prediction, which only achieves predictions that deviate from the true values as little as possible, but aims for predictions consistently higher than the actual values of real demand. Numerical experiments was conducted on eight datasets, including commonly used time series and real-world CPU resource data. The results indicate that most neural network-based methods fail to balance both resource demands and prediction accuracy effectively. In contrast, iQR can achieve optimal scheduling with the minimum economic cost and it is easier to satisfy the business constraints with quantile tuning. Notably, iQR is lightweight with a training speed in seconds and does not rely on the support of computing power of GPU resources. This study may provide new insight into investigations on resource scheduling optimization issues.", "keywords": [ "Resource Scheduling Strategy", "Sparse System Identification", "L1-norm Quantile Regression", "Business Constraint Satisfaction" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 3, 2, 3 ] }, { "paper_id": "JIGuWpQcqO", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Hao Zhao", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "MS student" }, { "name": "Mingjia Li", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Qiming Hu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Xiaojie Guo", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 44, "n_ref": 88, "n_ref_all": 106, "n_fig": 9, "n_tab": 3, "L_tab": 1351, "n_element_tab": 104, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 925, "n_element_tab_1": 40, "formula_len_all": 441, "formula_len_all_1": 440, "len_all": 125012, "len_all_1": 62305, "len_abs": 1107, "len_title": 113, "len_sents": 36953, "len_sents_1": 31399, "n_sents": 265, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1112, "title": "Reversible Decoupling Network for Single Image Reflection Removal", "abs": "Recent deep-learning-based approaches to single-image reflection removal have shown promising advances, primarily for two reasons: 1) the utilization of recognition-pretrained features as inputs, and 2) the design of dual-stream interaction networks. However, according to the Information Bottleneck principle, high-level semantic clues tend to be compressed or discarded during layer-by-layer propagation. Additionally, interactions in dual-stream networks follow a fixed pattern across different layers, limiting overall performance. To address these limitations, we propose a novel architecture called Reversible Decoupling Network (RDNet), which employs a reversible encoder to secure valuable information while flexibly decoupling transmission- and reflection-relevant features during the forward pass. Furthermore, we customize a transmission-rate-aware prompt generator to dynamically calibrate features, further boosting performance. Extensive experiments demonstrate the superiority of RDNet over existing SOTA methods on five widely-adopted benchmark datasets. Our code will be made publicly available.", "keywords": [ "Invertible network", "Reflection separation" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "JHoC430Nxi", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Daoxiong Gong", "gender": "unknown", "institution": "Beijing University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Jianjun Yu", "gender": "Female", "institution": "Beijing University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Junxi Chen", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Sijian Tian", "gender": "Male", "institution": "Beijing University of Technology", "country": "CN", "position": "MS student" }, { "name": "Zhihui Shen", "gender": "Male", "institution": "Beijing University of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 39, "n_ref": 91, "n_ref_all": 108, "n_fig": 8, "n_tab": 4, "L_tab": 2087, "n_element_tab": 375, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2045, "n_element_tab_1": 364, "formula_len_all": 585, "formula_len_all_1": 342, "len_all": 122071, "len_all_1": 58521, "len_abs": 1561, "len_title": 117, "len_sents": 32026, "len_sents_1": 24382, "n_sents": 227, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1566, "title": "CTNet: A CNN-Transformer Hybrid Network for 6D Object Pose Estimation", "abs": "Recent advances in 6D pose estimation primarily rely on CNNs, but they struggle to grasp long-range dependencies and the global context, which are essential for precise pose determination. Although deeper or expanded networks are commonly used to tackle this, they lead to significant computational burdens without fully addressing these constraints. To overcome these challenges, we present CTNet, a hybrid network that fuses the strengths of CNN and Transformer, aiming for accurate 6D pose estimation from a solitary RGB-D image. CTNet employs Transformer to capture elusive long-range dependencies and the global context, while lightweight CNNs adeptly extract detailed local features. This complementary approach offers a comprehensive feature representation, eliminating the necessity for excessively deep networks. To further bolster the CNNs' efficiency, we introduce the Hierarchical Feature Extractor (HFE), which enhances the C2f and ELAN modules for optimal feature extraction. Additionally, we integrate a CNN-based PointNet module, designed to extract vital spatial data from the point cloud. The Transformer element captures global contextual insights, which are then seamlessly integrated with the local and spatial features extracted by the CNNs to ensure precise 6D pose estimation. Experiments demonstrate that CTNet achieves high accuracy with nearly half the FLOPs of current methods on the LineMOD and YCB-Video datasets. Furthermore, the HFE is highly adaptable, showing excellent transferability across other 6D pose estimation architectures.", "keywords": [ "neural networks", "6D pose estimation", "RGB-D image" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 2, 3, 1 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "JHE4w8q2G2", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Chengqing Yu", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Chuanguang Yang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Fei Wang", "gender": "Male", "institution": "Institute of Computig Technology, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Tangwen Qian", "gender": "Female", "institution": ", Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Yongjun Xu", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zezhi Shao", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhulin An", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "tao sun", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences.", "country": "CN", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 44, "n_ref": 76, "n_ref_all": 98, "n_fig": 5, "n_tab": 18, "L_tab": 16035, "n_element_tab": 841, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2537, "n_element_tab_1": 287, "formula_len_all": 506, "formula_len_all_1": 486, "len_all": 254242, "len_all_1": 70813, "len_abs": 1828, "len_title": 166, "len_sents": 61428, "len_sents_1": 30514, "n_sents": 1049, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 117, "L_abs": 1838, "title": "Merlin: Multi-View Representation Learning for Robust Multivariate Time Series Forecasting with Unfixed Missing Rates", "abs": "Multivariate Time Series Forecasting (MTSF) aims to predict the future values of multiple interrelated time series and support decision-making. While deep learning models have attracted much attention in MTSF for their powerful spatial-temporal encoding capabilities, they frequently encounter the challenge of missing data resulting from numerous malfunctioning data collectors in practice. \nIn this case, existing models only rely on sparse observation, making it difficult to fully mine the semantics of MTS, which leads to a decline in their forecasting performance. Furthermore, the unfixed missing rates across different samples in reality pose robustness challenges. To address these issues, we propose Multi-View Representation Learning (Merlin) based on offline knowledge distillation and multi-view contrastive learning, which aims to help existing models achieve semantic alignment between sparse observations with different missing rates and complete observations, and enhance their robustness. On the one hand, we introduce offline knowledge distillation where a teacher model guides a student model in learning how to mine semantics from sparse observations similar to those obtainable from complete observations. On the other hand, we construct positive and negative data pairs using sparse observations with different missing rates. Then, we use multi-view contrastive learning to help the student model align semantics across sparse observations with different missing rates, thereby further enhancing its robustness. In this way, Merlin can fully enhance the robustness of existing forecasting models to MTS with unfixed missing rates and achieves high-precision MTSF with sparse observations. \nExperiments on four real-world datasets validate our motivation and demonstrate the superiority and practicability of Merlin.", "keywords": [ "Multivariate time series forecasting with sparse observations", "Multi-View Representation Learning", "Offline knowledge distillation", "Multi-view contrastive learning" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "JGr4Qv9vbz", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chen Liang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" }, { "name": "Wenguan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yucheng Yan", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 21, "n_formula_1": 16, "n_ref_uni": 85, "n_ref": 169, "n_ref_all": 188, "n_fig": 9, "n_tab": 8, "L_tab": 4344, "n_element_tab": 186, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1843, "n_element_tab_1": 70, "formula_len_all": 1301, "formula_len_all_1": 1008, "len_all": 203309, "len_all_1": 55318, "len_abs": 1564, "len_title": 144, "len_sents": 44721, "len_sents_1": 24084, "n_sents": 369, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1575, "title": "Open-Ended 3D Metric-Semantic Representation Learning via Semantic-Embedded Gaussian Splatting", "abs": "This work answers the question of whether it is feasible to create a comprehensive metric-semantic 3D virtual world using everyday devices equipped with multi-view stereo. We propose an open-ended metric-semantic representation learning framework based on 3D Gaussians, which distills open-set semantics from 2D foundation models into a scalable and continuously evolving 3D Gaussian representation, optimized within a SLAM framework. The process is non-trivial. The scalability requirements make direct embedding of semantic information into Gaussians impractical, resulting in excessive memory usage and semantic inconsistencies. In response, we propose to learn semantics by aggregating from a condensed, fixed-sized semantic pool rather than directly embedding high-dimensional raw features, significantly reducing memory requirements compared to the point-wise representation. Additionally, by enforcing pixel-to-pixel and pixel-to-object semantic consistency through contrastive learning and stability-guided optimization, our framework enhances coherence and stability in semantic representations. Extensive experiments demonstrate that our framework presents a precise open-ended metric-semantic field with superior rendering quality and tracking accuracy. Besides, it accurately captures both closed-set object categories and open-set semantics, facilitating various applications, notably fine-grained, unrestricted 3D scene editing. These results mark an initial yet solid step towards efficient and expressive 3D virtual world modelling. Our code will be released.", "keywords": [ "3D Gaussian Splatting", "Dense Visual SLAM", "3D Scene Representation", "Contrastive Learning" ], "rating_list": [ 3, 3, 8, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "JGTYlyVogb", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Chenguang Duan", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Huazhen Lin", "gender": "Female", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "Full Professor" }, { "name": "Jerry Zhijian Yang", "gender": "Male", "institution": "Wuhan University", "country": "", "position": "Full Professor" }, { "name": "Wensen Ma", "gender": "Male", "institution": "Wuhan University", "country": "", "position": "MS student" }, { "name": "Yuling Jiao", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 179, "n_formula_1": 44, "n_ref_uni": 32, "n_ref": 70, "n_ref_all": 92, "n_fig": 0, "n_tab": 10, "L_tab": 1449, "n_element_tab": 175, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 580, "n_element_tab_1": 30, "formula_len_all": 34661, "formula_len_all_1": 4139, "len_all": 210785, "len_all_1": 57073, "len_abs": 1034, "len_title": 119, "len_sents": 50618, "len_sents_1": 22230, "n_sents": 519, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1036, "title": "Self-supervised Transfer Learning via Adversarial Contrastive Training", "abs": "Learning a data representation with strong transferability from an unlabeled scenario is both crucial and challenging. In this paper, we propose a novel unbiased self-supervised transfer learning approach via Adversarial Contrastive Training (ACT). Additionally, we establish an end-to-end theoretical understanding for self-supervised contrastive pretraining and its implications for downstream classification tasks in a misspecified, over-parameterized setting. Our theoretical findings highlight the provable advantages of adversarial contrastive training in the source domain towards improving the accuracy of downstream tasks in the target domain. Furthermore, we illustrate that downstream tasks necessitate only a minimal sample size when working with a well-trained representation, offering valuable insights on few-shot learning. Moreover, extensive experiments across various datasets demonstrate a significant enhancement in classification accuracy when compared to existing state-of-the-art self-supervised learning methods.", "keywords": [ "Unsupervised transfer learning", "adversarial contrastive training", "deep neural network", "end to end error" ], "rating_list": [ 3, 3, 8 ], "soundness_list": [ 2, 1, 4 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 2, 3, 5 ] }, { "paper_id": "JFk8F7w8Iz", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Ukyo Honda", "gender": "Male", "institution": "CyberAgent, Inc.", "country": "JP", "position": "Researcher" }, { "name": "Yuu Jinnai", "gender": "Male", "institution": "CyberAgent, Inc.", "country": "JP", "position": "Researcher" } ], "n_formula": 27, "n_formula_1": 23, "n_ref_uni": 93, "n_ref": 216, "n_ref_all": 255, "n_fig": 12, "n_tab": 15, "L_tab": 8042, "n_element_tab": 962, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 553, "n_element_tab_1": 22, "formula_len_all": 2080, "formula_len_all_1": 1738, "len_all": 345877, "len_all_1": 59400, "len_abs": 1148, "len_title": 140, "len_sents": 48189, "len_sents_1": 25531, "n_sents": 434, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 91, "L_abs": 1068, "title": "Annotation-Efficient Language Model Alignment via Diverse and Representative Response Texts", "abs": "Preference optimization is a standard approach to fine-tuning large language models to align with human preferences.\nThe quantity, diversity, and representativeness of the preference dataset are critical to the effectiveness of preference optimization.\nHowever, obtaining a large amount of preference annotations is difficult in many applications.\nThis raises the question of how to use the limited annotation budget to create an effective preference dataset.\nTo this end, we propose Annotation-Efficient Preference Optimization (AEPO). \nInstead of exhaustively annotating preference over all available response texts, AEPO selects a subset of responses that maximizes diversity and representativeness from the available responses and then annotates preference over the selected ones.\nIn this way, AEPO focuses the annotation budget on labeling preference over a smaller subset of responses.\nWe evaluate the performance of Direct Preference Optimization (DPO) using AEPO and show that it outperforms models trained using a standard DPO with the same annotation budget.", "keywords": [ "Language model alignment", "Direct preference optimization" ], "rating_list": [ 3, 8, 5, 3 ], "soundness_list": [ 2, 4, 2, 1 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "JFPaD7lpBD", "primary_area": "foundation or frontier models, including LLMs", "n_author": 61, "author_info_list": [ { "name": "Alan Arazi", "gender": "Male", "institution": "AI21 Labs", "country": "IL", "position": "Researcher" }, { "name": "Amir Bergman", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "ML Engineer" }, { "name": "Avshalom Manevich", "gender": "unknown", "institution": "Bar-Ilan University", "country": "", "position": "MS student" }, { "name": "Barak Lenz", "gender": "Male", "institution": "AI21 Labs", "country": "IL", "position": "CTO" }, { "name": "Barak Peleg", "gender": "unknown", "institution": "AI21 Labs", "country": "IL", "position": "Researcher" }, { "name": "Ben Aviram", "gender": "Male", "institution": "AI21 Labs", "country": "IL", "position": "Researcher" }, { "name": "Chen Almagor", "gender": "unknown", "institution": "Hebrew University in Jerusalem", "country": "", "position": "MS student" }, { "name": "Clara Fridman", "gender": "unknown", "institution": "AI21 Labs", "country": "IL", "position": "Researcher" }, { "name": "Dan Padnos", "gender": "unknown", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "MS student" }, { "name": "Daniel Gissin", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "", "position": "MS student" }, { "name": "Daniel Jannai", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "", "position": "MS student" }, { "name": "Dor Muhlgay", "gender": "Male", "institution": "AI21 Labs", "country": "", "position": "Researcher" }, { "name": "Dor Zimberg", "gender": "unknown", "institution": "AI 21 labs", "country": "IL", "position": "Researcher" }, { "name": "Edden M. Gerber", "gender": "Male", "institution": "AI21 Labs", "country": "IL", "position": "Researcher" }, { "name": "Elad Dolev", "gender": "unknown", "institution": "AI21 Labs", "country": "IL", "position": "Instructor" }, { "name": "Eran Krakovsky", "gender": "Male", "institution": "Reichman University", "country": "IL", "position": "MS student" }, { "name": "Erez Safahi", "gender": "unknown", "institution": "AI21 Labs", "country": "US", "position": "Researcher" }, { "name": "Erez Schwartz", "gender": "unknown", "institution": "Technion, Technion", "country": "", "position": "Undergrad student" }, { "name": "Gal Cohen", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "MS student" }, { "name": "Gal Shachaf", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "MS student" }, { "name": "Haim Rozenblum", "gender": "Male", "institution": "AI21 Labs", "country": "IL", "position": "Researcher" }, { "name": "Hofit Bata", "gender": "Female", "institution": "AI21 Labs", "country": "IL", "position": "Researcher" }, { "name": "Ido Blass", "gender": "Male", "institution": "Bar-Ilan University", "country": "IL", "position": "PhD student" }, { "name": "Inbal Magar", "gender": "Female", "institution": "Hebrew University in Jerusalem", "country": "", "position": "MS student" }, { "name": "Itay Dalmedigos", "gender": "Male", "institution": "AI21", "country": "IL", "position": "Researcher" }, { "name": "Jhonathan Osin", "gender": "Female", "institution": "Tel Aviv University, Technion", "country": "", "position": "MS student" }, { "name": "Josh Cohen", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "MS student" }, { "name": "Julie Fadlon", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "Researcher" }, { "name": "Maria Rozman", "gender": "unknown", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "Lecturer" }, { "name": "Matan Danos", "gender": "unknown", "institution": "ai21", "country": "IL", "position": "Researcher" }, { "name": "Michael Gokhman", "gender": "unknown", "institution": "AI21 Labs", "country": "IL", "position": "Researcher" }, { "name": "Mor Zusman", "gender": "Male", "institution": "Open University of Israel", "country": "IL", "position": "MS student" }, { "name": "Naama Gidron", "gender": "unknown", "institution": "AI21 Labs", "country": "IL", "position": "language data analyst" }, { "name": "Nir Ratner", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "MS student" }, { "name": "Noam Gat", "gender": "unknown", "institution": "AI21 Labs", "country": "IL", "position": "Researcher" }, { "name": "Noam Rozen", "gender": "Male", "institution": "Weizmann Institute, Technion", "country": "", "position": "MS student" }, { "name": "Oded Fried", "gender": "unknown", "institution": "AI21 Labs", "country": "IL", "position": "Researcher" }, { "name": "Ohad Leshno", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "MS student" }, { "name": "Omer Antverg", "gender": "unknown", "institution": "Technion, Technion", "country": "", "position": "MS student" }, { "name": "Omri Abend", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "", "position": "Associate Professor" }, { "name": "Opher Lieber", "gender": "Male", "institution": "AI21 Labs", "country": "IL", "position": "Principal Researcher" }, { "name": "Or Dagan", "gender": "Male", "institution": "AI21", "country": "IL", "position": "Full Professor" }, { "name": "Orit Cohavi", "gender": "Female", "institution": "AI21 Labs", "country": "IL", "position": "Algorithm Developer" }, { "name": "Raz Alon", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "IL", "position": "PhD student" }, { "name": "Ro'i Belson", "gender": "unknown", "institution": "AI21labs", "country": "IL", "position": "Language Data Analyst" }, { "name": "Roi Cohen", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Rom Gilad", "gender": "unknown", "institution": "AI21 labs", "country": "IL", "position": "Researcher" }, { "name": "Roman Glozman", "gender": "unknown", "institution": "AI21Labs", "country": "IL", "position": "Researcher" }, { "name": "Shahar Lev", "gender": "Male", "institution": "School of Computer Science, Tel Aviv University", "country": "IL", "position": "Msc" }, { "name": "Shai Shalev-Shwartz", "gender": "Male", "institution": "Hebrew University of Jerusalem", "country": "", "position": "Full Professor" }, { "name": "Shaked Haim Meirom", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "MS student" }, { "name": "Tal Delbari", "gender": "unknown", "institution": "Reichman University", "country": "IL", "position": "Undergrad student" }, { "name": "Tal Ness", "gender": "Female", "institution": "AI21 Labs", "country": "IL", "position": "Language Data Scientist" }, { "name": "Tom Ben Gal", "gender": "unknown", "institution": "AI21 Labs", "country": "IL", "position": "Researcher" }, { "name": "Tom Braude", "gender": "unknown", "institution": "IDC", "country": "", "position": "MS student" }, { "name": "Tomer Asida", "gender": "unknown", "institution": "AI21 Labs", "country": "IL", "position": "Researcher" }, { "name": "Uriya Pumerantz", "gender": "unknown", "institution": "Ai21", "country": "IL", "position": "Researcher" }, { "name": "Yoav Shoham", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "", "position": "Emeritus" }, { "name": "Yonatan Belinkov", "gender": "Male", "institution": "Technion, Technion", "country": "IL", "position": "Assistant Professor" }, { "name": "Yuval Globerson", "gender": "unknown", "institution": "Hebrew University of Jerusalem", "country": "IL", "position": "MS student" }, { "name": "Yuval Peleg Levy", "gender": "Male", "institution": "AI21", "country": "IL", "position": "Algorithm Developer" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 62, "n_ref": 99, "n_ref_all": 140, "n_fig": 11, "n_tab": 8, "L_tab": 1980, "n_element_tab": 338, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1274, "n_element_tab_1": 211, "formula_len_all": 249, "formula_len_all_1": 196, "len_all": 204145, "len_all_1": 64033, "len_abs": 1421, "len_title": 95, "len_sents": 48359, "len_sents_1": 27859, "n_sents": 403, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1428, "title": "Jamba: Hybrid Transformer-Mamba Language Models", "abs": "We present Jamba, a novel hybrid Transformer-Mamba mixture-of-experts (MoE) architecture. Jamba interleaves blocks of Transformer and Mamba layers, enjoying the benefits of both model families. MoE is added in some of these layers to increase model capacity while keeping active parameter usage manageable. This flexible architecture allows resource- and objective-specific configurations. We implement two configurations: Jamba-1.5-Large, with 94B active parameters, and Jamba-1.5-mini, with 12B active parameters. Built at large scale, Jamba models provide high throughput and small memory footprint compared to vanilla Transformers, especially at long-context tasks, with an effective context length of 256K tokens, the largest amongst open-weight models. At the same time, they are also competitive on standard language modeling and chatbot benchmarks. We study various architectural decisions, such as how to combine Transformer and Mamba layers, and how to mix experts, and show that some of them are crucial in large scale modeling. To support cost-effective inference, we introduce ExpertsInt8, a novel quantization technique that allows fitting Jamba-1.5-Large on a machine with 8 80GB GPUs when processing 256K-token contexts without loss of quality. We also describe several interesting properties of this architecture that the training and evaluation of Jamba have revealed. The model weights are publicly available.", "keywords": [ "language models", "state-space models", "mamba", "hybrid architecture", "foundation models" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 3, 2, 2, 4 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "JF8ovQyueq", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Alessandro Licciardi", "gender": "Male", "institution": "Polytechnic Institute of Turin", "country": "IT", "position": "PhD student" }, { "name": "Barbara Caputo", "gender": "Female", "institution": "Politecnico di Torino", "country": "IT", "position": "Full Professor" }, { "name": "Davide Leo", "gender": "Male", "institution": "Polytechnic Institute of Turin", "country": "IT", "position": "MS student" }, { "name": "Eros Fan\u00ec", "gender": "Male", "institution": "Polytechnic Institute of Turin", "country": "IT", "position": "PhD student" }, { "name": "Marco Ciccone", "gender": "Male", "institution": "Vector Institute", "country": "CA", "position": "Postdoc" } ], "n_formula": 68, "n_formula_1": 16, "n_ref_uni": 39, "n_ref": 66, "n_ref_all": 87, "n_fig": 7, "n_tab": 10, "L_tab": 3174, "n_element_tab": 301, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2933, "n_element_tab_1": 98, "formula_len_all": 3641, "formula_len_all_1": 723, "len_all": 189775, "len_all_1": 83156, "len_abs": 1074, "len_title": 118, "len_sents": 64648, "len_sents_1": 38402, "n_sents": 474, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1075, "title": "Interaction Based Gaussian Weighting Clustering for Federated Learning", "abs": "Federated learning emerged as a decentralized paradigm to train models while securing privacy. However, conventional FL faces data heterogeneity and class imbalance challenges, affecting model performance. In response to these issues, Personalized FL has been developed as an innovative methodology that relies on fine-tuning the distinct local models based on individual training datasets. In this work, we propose a novel PFL method, FedGW (Federated Gaussian Weighting), which groups clients based on their data distribution, allowing training of a more robust and personalized model on the identified clusters. FedGW identifies homogeneous clusters by transforming individual empirical losses to model client interactions with a Gaussian reward mechanism. Additionally, we introduce a new clustering metric for FL to evaluate cluster cohesion with respect to the individual class distribution. Our experiments on benchmark datasets show that FedGW outperforms existing FL algorithms in cluster quality and classification accuracy, validating the efficacy of our approach.", "keywords": [ "Federated Learning", "Clustered Federated Learning", "Personalized Federated Learning" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "JEmNgjuQHU", "primary_area": "datasets and benchmarks", "n_author": 13, "author_info_list": [ { "name": "Duy-Nhat Vo", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "", "position": "MS student" }, { "name": "Esra Suel", "gender": "Female", "institution": "Imperial College London", "country": "", "position": "Postdoc" }, { "name": "Fan Yang", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "MS student" }, { "name": "H Juliette T Unwin", "gender": "Female", "institution": "University of Bristol", "country": "GB", "position": "Lecturer" }, { "name": "Jack Gidney", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "Undergrad student" }, { "name": "Makkunda Sharma", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "", "position": "PhD student" }, { "name": "Matthew Sutcliffe", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Mengyan Zhang", "gender": "unknown", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Oliver Fiala", "gender": "Male", "institution": "Save the Children", "country": "", "position": "Researcher" }, { "name": "Samir Bhatt", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Full Professor" }, { "name": "Seth Flaxman", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" }, { "name": "Swapnil Mishra", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "William Rudgard", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 25, "n_ref": 36, "n_ref_all": 49, "n_fig": 3, "n_tab": 1, "L_tab": 571, "n_element_tab": 50, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1841, "n_element_tab_1": 87, "formula_len_all": 3, "formula_len_all_1": 3, "len_all": 111652, "len_all_1": 59833, "len_abs": 1521, "len_title": 98, "len_sents": 37243, "len_sents_1": 28684, "n_sents": 245, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1526, "title": "KidSat: satellite imagery to map childhood poverty", "abs": "Satellite imagery has emerged as an important tool to analyze demographic, health, and development indicators. While various deep learning models have been built for these tasks, each is specific to a particular problem, with few standard benchmarks available. We propose a new dataset pairing satellite imagery and high-quality survey data on child poverty to benchmark satellite feature representations. Our dataset consists of 33,608 images, each 10 km \u00d7 10 km, from 16 countries in Eastern and Southern Africa in the time period 1997-2022. As defined by UNICEF, multidimensional child poverty comprises six fundamental factors\u2014housing, sanitation, water, nutrition, education, and health (UNICEF, 2021)\u2014which can be calculated from geocoded, face-to-face Demographic and Health Surveys (DHS) Program data. Using our dataset we benchmark multiple feature representations for encoding satellite imagery, from low-level satellite imagery models such as MOSAIKS (Rolf et al., 2021), to deep learning foundation models, which include both generic vision models such as DINOv2 (Oquab et al., 2023) and specific satellite imagery models such as SatMAE (Cong et al., 2022). As part of the benchmark, we test spatial as well as temporal generalization, by testing on unseen locations, and on data beyond the training years. We provide open source code to reproduce and extend our entire pipeline: building the satellite imagery dataset, obtaining ground truth data from DHS, and comparing the various models considered in our work.", "keywords": [ "satellite imagery", "remote sensing", "social science", "global health", "economic", "health and development indicators" ], "rating_list": [ 1, 3, 1, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 1, 3, 1, 1 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "JElN0LJMKB", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Di Xie", "gender": "Male", "institution": "Hikvision Research Institute", "country": "CN", "position": "Research Director" }, { "name": "Jiang Zhu", "gender": "unknown", "institution": "Hikvision Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Qiulin Zhang", "gender": "Male", "institution": "Hikvision Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Rudan Chen", "gender": "Female", "institution": "Hikvision Research Institute", "country": "", "position": "Researcher" }, { "name": "Xichao Yang", "gender": "unknown", "institution": "Xi'an University of Electronic Science and Technology", "country": "", "position": "Undergrad student" }, { "name": "Zhaojing Wen", "gender": "unknown", "institution": "Hikvision Research Institute", "country": "CN", "position": "Researcher" }, { "name": "yuan zhang", "gender": "Male", "institution": "Hikvision Research Institute", "country": "CN", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 22, "n_ref": 54, "n_ref_all": 62, "n_fig": 2, "n_tab": 6, "L_tab": 2604, "n_element_tab": 314, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 2604, "n_element_tab_1": 314, "formula_len_all": 2004, "formula_len_all_1": 2005, "len_all": 102635, "len_all_1": 62177, "len_abs": 1576, "len_title": 126, "len_sents": 25212, "len_sents_1": 24527, "n_sents": 200, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1591, "title": "From Decoupling to Adaptive Transformation: a Wider Optimization Space for PTQ", "abs": "Post-training low-bit quantization (PTQ) is useful to accelerate DNNs due to its high efficiency. Currently, finetuning through self-distillation feature reconstruction is one of the most effective PTQ techniques. However, when bitwidth goes to be extremely low, we find that current parameter update settings in PTQ feature reconstruction are sub-optimal. Considering all possible parameters and the ignored fact that integer weight can be obtained early before actual inference, we thoroughly explore 1) the setting of weight\u2019s quantization step into six cases by decoupling; 2) ignored learnable params in PTQ like BN and bias. Based on these explorations, we find there exist a wider optimization space and a better optimum. Considering these, we propose an Adaptive Quantization Transformation(AdaQTransform) for PTQ reconstruction, which provides adaptive per-channel transformation on the quant output feature, making them better fit FP32 counterpart and achieve lower PTQ feature reconstruction error. During inference, the AdaQTransform parameters can be merged without incurring additional inference costs. Based on AdaQTransform, for the first time, we build a general quantization setting paradigm subsuming current PTQs, QATs and other potential approaches. Experiments demonstrate that AdaQTransform expands the optimization space for PTQ and helps current PTQs find a better optimum over CNNs, ViTs, LLMs and low-level vision networks (image super-resolution). Specifically, AdaQTransform improves the current best PTQ by 5.7\\% on W2A2-MobileNet-v2. The code will be released.", "keywords": [ "Post-Training Quantization" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "JEjVuVxbkf", "primary_area": "interpretability and explainable AI", "n_author": 1, "author_info_list": [ { "name": "Thomas Bonnier", "gender": "Male", "institution": "Centrale Lille Alumni", "country": "", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 38, "n_ref": 49, "n_ref_all": 89, "n_fig": 8, "n_tab": 14, "L_tab": 9371, "n_element_tab": 665, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1062, "formula_len_all_1": 946, "len_all": 194766, "len_all_1": 64941, "len_abs": 1822, "len_title": 157, "len_sents": 61240, "len_sents_1": 30532, "n_sents": 537, "n_sents_1": 236, "venue": "ICLR2025", "decision": "withdraw", "L_title": 108, "L_abs": 1428, "title": "Evaluating and Explaining the Severity of Distribution Shifts: Illustration with Tabular Text Classification", "abs": "After deploying a machine learning model, distribution shifts may emerge in real-world data. When dealing with unlabeled data, it can be challenging to accurately assess the impact of these drifts on the model's performance, for any type and intensity of shift. In that case, decisions such as updating the model for every benign shift would not be cost-efficient. In this paper, we introduce the Error Classifier, an error assessment method that addresses two tasks: unsupervised performance estimation and error detection on out-of-distribution data. The Error Classifier computes the probability that the model will fail based on detected fault patterns. Further, we employ a sampling-based approximation of Shapley values, with the Error Classifier as value function, in order to explain why a shift is predicted as severe, in terms of feature values. As explanation methods can sometimes disagree, we suggest evaluating the consistency of explanations produced by our technique and different ones. We focus on classification and illustrate the relevance of our method in a bimodal context, on tabular datasets with text fields. We measure our method against a selection of 15 baselines from various domains, on 7 datasets with a variety of shifts, and 2 multimodal fusion strategies for the classification models. Lastly, we show the usefulness of our explanation algorithm on instances affected by various types of shifts.", "keywords": [ "unsupervised performance estimation", "error detection", "distribution shifts", "explanation method", "multimodal classification" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "JEehcb48Vp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Boxi Cao", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Debing Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Hongyu Lin", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Le Sun", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xianpei Han", "gender": "Male", "institution": "Institute of Software, CAS", "country": "CN", "position": "Professor" }, { "name": "Xin Zheng", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yaojie Lu", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Yuqiu Ji", "gender": "Female", "institution": "SCIR", "country": "", "position": "MS student" }, { "name": "jie lou", "gender": "Male", "institution": "Xiaohongshu Inc", "country": "CN", "position": "Researcher" }, { "name": "Xueruwen", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 2, "n_ref_uni": 38, "n_ref": 68, "n_ref_all": 98, "n_fig": 4, "n_tab": 11, "L_tab": 4218, "n_element_tab": 395, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 745, "formula_len_all_1": 258, "len_all": 212299, "len_all_1": 61098, "len_abs": 1148, "len_title": 145, "len_sents": 54469, "len_sents_1": 28138, "n_sents": 463, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1155, "title": "Critic-CoT: Boosting the reasoning abilities of large language model via Chain-of-Thought Critic", "abs": "Self-critic has become a crucial mechanism for enhancing the reasoning performance of LLMs. However, current approaches mainly involve basic prompts for intuitive instance-level feedback, which resembles System-1 processes and limits the reasoning capabilities. Moreover, there is a lack of in-depth investigations into the relationship between LLM's ability to criticize and its task-solving performance. To address these issues, we propose Critic-CoT, a novel framework that pushes LLMs toward System-2-like critic capability. Through a step-wise CoT reasoning paradigm and the automatic construction of distant-supervision data without human annotation, Critic-CoT enables LLMs to engage in slow, analytic self-critique and refinement, thereby improving their reasoning abilities. Experiments on GSM8K and MATH demonstrate that our enhanced model significantly boosts task-solving performance by filtering out invalid solutions or iterative refinement. Furthermore, we investigate the intrinsic correlation between critique and task-solving abilities within LLMs, discovering that these abilities can mutually reinforce each other rather than conflict.", "keywords": [ "Self-critic", "Large language model", "Chain-of-thought" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "JE9tCwe3lp", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Haiwen Diao", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Huchuan Lu", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Professor" }, { "name": "Shiguang Shan", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Ting Pan", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xinlong Wang", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Yonggang Qi", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Associate Professor" }, { "name": "Yufeng Cui", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Hao Ge Deng", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "\u6b63\u96c4 \u7f57", "gender": "Male", "institution": "BAAI", "country": "CN", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 53, "n_ref": 105, "n_ref_all": 130, "n_fig": 9, "n_tab": 2, "L_tab": 2272, "n_element_tab": 398, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 2272, "n_element_tab_1": 398, "formula_len_all": 398, "formula_len_all_1": 371, "len_all": 158008, "len_all_1": 66711, "len_abs": 1616, "len_title": 108, "len_sents": 30704, "len_sents_1": 28762, "n_sents": 222, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1647, "title": "Autoregressive Video Generation without Vector Quantization", "abs": "Generating a video causally in an autoregressive manner is considered a promising path toward infinite video generation in a flexible context. Prior autoregressive approaches typically rely on vector quantization to convert a video into a discrete-valued space, which could raise challenges in efficiency when modeling long videos. In this work, we propose a novel approach that enables autoregressive video generation without vector quantization. We propose to reformulate the video generation problem as an autoregressive modeling framework integrating temporal \\textit{frame-by-frame} prediction and spatial \\textit{set-by-set} prediction. Unlike raster-scan prediction in prior autoregressive models or joint distribution modeling of fixed-length tokens in diffusion models, our approach maintains the causal property of GPT-style models for flexible in-context capabilities, while leveraging bidirectional modeling within individual frames for efficiency. We train a novel video autoregressive model with the proposed approach, termed \\Ours. Our results demonstrate that \\Ours fully surpasses prior autoregressive video models in data efficiency, inference speed, visual fidelity, and video fluency, even with a much smaller model capacity, \\ie, 0.6B parameters. \\Ours generalizes well across extended video durations and enables diverse zero-shot applications in one unified model. Additionally, with a significantly lower training cost, \\Ours outperforms state-of-the-art image diffusion models in text-to-image generation tasks. We will release all weights, models, and code to facilitate the reproduction of \\Ours and further development.", "keywords": [ "Autoregressive Generation", "Text-to-Video Generation", "Continuous-valued Space" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "JDzTI9rKls", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Fabian Otto", "gender": "unknown", "institution": "Microsoft", "country": "GB", "position": "Researcher" }, { "name": "Philipp Becker", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "DE", "position": "PhD student" }, { "name": "Vien Anh Ngo", "gender": "Male", "institution": "Bosch Center for Artificial Intelligence", "country": "", "position": "Research Scientist" }, { "name": "Gerhard Neumann", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 50, "n_ref": 143, "n_ref_all": 179, "n_fig": 6, "n_tab": 2, "L_tab": 201, "n_element_tab": 64, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 349, "n_element_tab_1": 71, "formula_len_all": 2000, "formula_len_all_1": 919, "len_all": 150548, "len_all_1": 73255, "len_abs": 1481, "len_title": 112, "len_sents": 45968, "len_sents_1": 36619, "n_sents": 309, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1494, "title": "Efficient Off-Policy Learning for High-Dimensional Action Spaces", "abs": "Existing off-policy reinforcement learning algorithms often rely on an explicit state-action-value function representation, which can be problematic in high-dimensional action spaces due to the curse of dimensionality.\nThis reliance results in data inefficiency as maintaining a state-action-value function in such spaces is challenging. \nWe present an efficient approach that utilizes only a state-value function as the critic for off-policy deep reinforcement learning.\nThis approach, which we refer to as Vlearn, effectively circumvents the limitations of existing methods by eliminating the necessity for an explicit state-action-value function. \nTo this end, we leverage a weighted importance sampling loss for learning deep value functions from off-policy data. \nWhile this is common for linear methods, it has not been combined with deep value function networks. \nThis transfer to deep methods is not straightforward and requires novel design choices such as robust policy updates, twin value function networks to avoid an optimization bias, and importance weight clipping.\nWe also present a novel analysis of the variance of our estimate compared to commonly used importance sampling estimators such as V-trace. \nOur approach improves sample complexity as well as final performance and ensures consistent and robust performance across various benchmark tasks.\nEliminating the state-action-value function in Vlearn facilitates a streamlined learning process, yielding high-return agents.", "keywords": [ "reinforcement learning", "trust region" ], "rating_list": [ 5, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "JDm7oIcx4Y", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Alexandre Allauzen", "gender": "Male", "institution": "Ecole Sup\u00e9rieure de Physique et de Chimie Industrielles", "country": "FR", "position": "Full Professor" }, { "name": "Blaise Delattre", "gender": "Male", "institution": ", Universit\u00e9 Paris-Dauphine (Paris IX)", "country": "FR", "position": "PhD student" }, { "name": "Erwan Fagnou", "gender": "Male", "institution": "Universit\u00e9 Paris-Dauphine - PSL", "country": "FR", "position": "PhD student" }, { "name": "Paul Caillon", "gender": "unknown", "institution": ", Universit\u00e9 Paris-Dauphine (Paris IX)", "country": "", "position": "Postdoc" } ], "n_formula": 30, "n_formula_1": 18, "n_ref_uni": 41, "n_ref": 77, "n_ref_all": 102, "n_fig": 6, "n_tab": 4, "L_tab": 1241, "n_element_tab": 28, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 444, "n_element_tab_1": 22, "formula_len_all": 2064, "formula_len_all_1": 1192, "len_all": 142577, "len_all_1": 59974, "len_abs": 1189, "len_title": 0, "len_sents": 39005, "len_sents_1": 24710, "n_sents": 313, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 83, "L_abs": 1195, "title": "Accelerated training through iterative gradient propagation along the residual path", "abs": "Despite being the cornerstone of deep learning, backpropagation is criticized for its inherent sequentiality, which can limit the scalability of very deep models.\nSuch models faced convergence issues due to vanishing gradient, later resolved using residual connections. Variants of these are now widely used in modern architectures.\nHowever, the computational cost of backpropagation remains a major burden, accounting for most of the training time.\nTaking advantage of residual-like architectural designs, we introduce Highway backpropagation, a parallelizable iterative algorithm that approximates backpropagation, by alternatively i) accumulating the gradient estimates along the residual path, and ii) backpropagating them through every layer in parallel. This algorithm is naturally derived from a decomposition of the gradient as the sum of gradients flowing through all paths, and is adaptable to a diverse set of common architectures, ranging from ResNets and Transformers to recurrent neural networks.\nThrough an extensive empirical study on a large selection of tasks and models, we evaluate Highway-BP and show that major speedups can be achieved with minimal performance degradation.", "keywords": [ "optimization", "efficient training" ], "rating_list": [ 8, 8, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "JDiER86r8v", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Bin-Bin Gao", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Senior Researcher" }, { "name": "Chengjie Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Feng Zheng", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Hanqiu Deng", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "PhD student" }, { "name": "Jialin Li", "gender": "Male", "institution": "Tencent YouTu Lab", "country": "CN", "position": "Researcher" }, { "name": "Jian Li", "gender": "Male", "institution": "Tencent Youtu", "country": "CN", "position": "Researcher" }, { "name": "Yifeng Zhou", "gender": "unknown", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Yong Liu", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Jiang Xi", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 35, "n_ref": 70, "n_ref_all": 94, "n_fig": 11, "n_tab": 8, "L_tab": 6524, "n_element_tab": 825, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2958, "n_element_tab_1": 373, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 179094, "len_all_1": 65195, "len_abs": 1341, "len_title": 114, "len_sents": 44583, "len_sents_1": 28200, "n_sents": 353, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1327, "title": "MMAD: A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly Detection", "abs": "In the field of industrial inspection, Multimodal Large Language Models (MLLMs) have a high potential to renew the paradigms in practical applications due to their robust language capabilities and generalization abilities. However, despite their impressive problem-solving skills in many domains, MLLMs' ability in industrial anomaly detection has not been systematically studied. To bridge this gap, we present MMAD, a full-spectrum MLLM benchmark in industrial Anomaly Detection. We defined seven key subtasks of MLLMs in industrial inspection and designed a novel pipeline to generate the MMAD dataset with 39,672 questions for 8,366 industrial images. With MMAD, we have conducted a comprehensive, quantitative evaluation of various state-of-the-art MLLMs. The commercial models performed the best, with the average accuracy of GPT-4o models reaching 74.9\\%. However, this result falls far short of industrial requirements. Our analysis reveals that current MLLMs still have significant room for improvement in answering questions related to industrial anomalies and defects. We further explore two training-free performance enhancement strategies to help models improve in industrial scenarios, highlighting their promising potential for future research. \nThe code and data are available at https://github.com/jam-cc/MMAD.", "keywords": [ "Anomaly Detection", "Multimodal Large Language Model", "Industrial Inspection" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "JDa5RiTIC7", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Boyu Gou", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Boyuan Zheng", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Cheng Chang", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Huan Sun", "gender": "Female", "institution": "The Ohio State University, Columbus", "country": "US", "position": "Associate Professor" }, { "name": "Kai Zhang", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Peng Qi", "gender": "unknown", "institution": "Orby AI", "country": "US", "position": "Researcher" }, { "name": "Sanjari Srivastava", "gender": "Female", "institution": "Orby AI", "country": "US", "position": "Researcher" }, { "name": "Yanan Xie", "gender": "Male", "institution": "Orby AI", "country": "US", "position": "Software engineer" }, { "name": "Yu Gu", "gender": "Male", "institution": "Ohio State University", "country": "US", "position": "PhD student" }, { "name": "Yu Su", "gender": "Male", "institution": "Ohio State University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 37, "n_ref": 68, "n_ref_all": 82, "n_fig": 4, "n_tab": 4, "L_tab": 461, "n_element_tab": 57, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 381, "n_element_tab_1": 55, "formula_len_all": 585, "formula_len_all_1": 515, "len_all": 133236, "len_all_1": 56548, "len_abs": 1349, "len_title": 104, "len_sents": 38242, "len_sents_1": 27211, "n_sents": 306, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1301, "title": "Simulate Before Act: Model-Based Planning for Web Agents", "abs": "Language agents have shown promising performance in automating web-based tasks, but the complexity and vast search spaces of real-world websites challenge reactive agents in identifying optimal solutions. While tree search agents offer enhanced exploration by interacting with actual websites, they often incur high costs, potential risks, and are challenging to implement for real-world websites. This paper explores a novel paradigm leveraging large language models' (LLMs) internal world models for planning in complex environments, presenting a middle ground between reactive agents and tree search agents. Results on two representative benchmarks, VisualWebArena and Mind2Web-live, demonstrate that our approach largely closes the gap between reactive agents and tree search agents, while maintaining efficiency and safety advantages. Notably, tree search can be considered as approaching an upper bound for our method, as it explores actual websites rather than simulations. This work opens new avenues for research into more effective and secure strategies for autonomous agents in complex, dynamic environments. It represents a step forward in improving upon reactive agents while approaching the performance of tree search methods, without incurring their implementation challenges and costs.", "keywords": [ "Web Agents; World Model; Planning; MPC" ], "rating_list": [ 5, 8, 1, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 1, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "JD6j7XSluo", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Bac Nguyen", "gender": "Male", "institution": "Sony Europe B.V.", "country": "DE", "position": "Senior Engineer" }, { "name": "Chieh-Hsin Lai", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Researcher" }, { "name": "Naoki Murata", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Researcher" }, { "name": "Stefano Ermon", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Toshimitsu Uesaka", "gender": "Male", "institution": "Sony AI", "country": "", "position": "Researcher" }, { "name": "Yuhta Takida", "gender": "Male", "institution": "Sony AI", "country": "JP", "position": "Research Scientist" }, { "name": "Yuki Mitsufuji", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Lead Research Scientist" } ], "n_formula": 24, "n_formula_1": 15, "n_ref_uni": 49, "n_ref": 127, "n_ref_all": 159, "n_fig": 8, "n_tab": 8, "L_tab": 2967, "n_element_tab": 330, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2280, "n_element_tab_1": 206, "formula_len_all": 2604, "formula_len_all_1": 1286, "len_all": 187349, "len_all_1": 69188, "len_abs": 4471, "len_title": 118, "len_sents": 50445, "len_sents_1": 28352, "n_sents": 424, "n_sents_1": 241, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1264, "title": "Mitigating Embedding Collapse in Diffusion Models for Categorical Data", "abs": "Latent diffusion models have enabled continuous-state diffusion models to handle a variety of datasets, including categorical data. However, most methods rely on fixed pretrained embeddings, limiting the benefits of joint training with the diffusion model. While jointly learning the embedding (via reconstruction loss) and the latent diffusion model (via score matching loss) could enhance performance, our analysis shows that end-to-end training risks embedding collapse, degrading generation quality. To address this issue, we introduce CATDM, a continuous diffusion framework within the embedding space that stabilizes training. We propose a novel objective combining the joint embedding-diffusion variational lower bound with a Consistency-Matching (CM) regularizer, alongside a shifted cosine noise schedule and random dropping strategy. The CM regularizer ensures the recovery of the true data distribution. Experiments on benchmarks show that CATDM mitigates embedding collapse, yielding superior results on FFHQ, LSUN Churches, and LSUN Bedrooms. In particular, CATDM achieves an FID of 6.81 on ImageNet $256\\times256$ with 50 steps. It outperforms non-autoregressive models in machine translation and is on a par with previous methods in text generation.", "keywords": [ "diffusion model", "categorical data", "embedding collapse" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "JCiF03qnmi", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Dean Foster", "gender": "Male", "institution": "Amazon", "country": "", "position": "scientist" }, { "name": "Depen Morwani", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Difan Zou", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Hanlin Zhang", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Jingfeng Wu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Nikhil Vyas", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Sham M. Kakade", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Udaya Ghai", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 8, "n_ref_uni": 41, "n_ref": 93, "n_ref_all": 143, "n_fig": 13, "n_tab": 11, "L_tab": 3461, "n_element_tab": 442, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1522, "n_element_tab_1": 34, "formula_len_all": 1626, "formula_len_all_1": 538, "len_all": 205028, "len_all_1": 62975, "len_abs": 1198, "len_title": 89, "len_sents": 51352, "len_sents_1": 29174, "n_sents": 355, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1202, "title": "How Does Critical Batch Size Scale in Pre-training?", "abs": "Training large-scale models under given resources requires careful design of parallelism strategies. In particular, the efficiency notion of critical batch size (CBS), concerning the compromise between time and compute, marks the threshold beyond which greater data parallelism leads to diminishing returns. To operationalize it, we propose a measure of CBS and pre-train a series of auto-regressive language models, ranging from 85 million to 1.2 billion parameters, on the C4 dataset. Through extensive hyper-parameter sweeps and careful control of factors such as batch size, momentum, and learning rate along with its scheduling, we systematically investigate the impact of scale on CBS. Then we fit scaling laws with respect to model and data sizes to decouple their effects. Overall, our results demonstrate that CBS scales primarily with data size rather than model size, a finding we justify theoretically through the analysis of infinite-width limits of neural networks and infinite-dimensional least squares regression. Of independent interest, we highlight the importance of common hyper-parameter choices and strategies for studying large-scale pre-training beyond fixed training durations.", "keywords": [ "Pre-training", "Language Models", "Data Parallelism", "Optimization" ], "rating_list": [ 8, 6, 6, 8, 6 ], "soundness_list": [ 4, 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "JCFJFBm5rE", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Brian Hie", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Fang Wu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Jianmin Wang", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "PhD student" }, { "name": "Jinbo Xu", "gender": "Male", "institution": "Toyota Technological Institute at Chicago", "country": "US", "position": "Full Professor" }, { "name": "Shuting Jin", "gender": "Female", "institution": "Wuhan university science and technology", "country": "CN", "position": "Lecturer" }, { "name": "Zerui Xu", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "MS student" }, { "name": "xiangxiang Zeng", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 29, "n_formula_1": 21, "n_ref_uni": 69, "n_ref": 168, "n_ref_all": 185, "n_fig": 6, "n_tab": 7, "L_tab": 4965, "n_element_tab": 521, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 3986, "n_element_tab_1": 348, "formula_len_all": 1581, "formula_len_all_1": 1126, "len_all": 195610, "len_all_1": 63927, "len_abs": 1340, "len_title": 106, "len_sents": 41231, "len_sents_1": 23003, "n_sents": 304, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1162, "title": "SurfDesign: Effective Protein Design on Molecular Surfaces", "abs": "Structure-based inverse folding has been extensively explored in recent years. In contrast, surface-conditioned protein generation is still an under-explored area. Molecular surfaces characterized by a compact and smooth composition of atoms at their boundary hold a more direct relevance to biomolecular interactions and function.\nIn this work, we introduce a novel framework named SurfDesign with several key improvements. Firstly, considering the theoretical fact that the molecular surface is a continuous manifold with infinite resolution, we propose surface-based equivariant message passing (SEMP) to incorporate the normal vector and curvatures and get aware of the manifold's Euclidean locality. Besides, a hybrid parameter-efficient fine-tuning (PEFT) technique is employed to combine the knowledge of protein language models (PLMs) with the surface geometric encoder. We extensively evaluate SurfDesign on the CATH, TS50, TS500, and PDB datasets, achieving an average recovery of more than 70\\%. Our work opens another road to designing functional proteins, underscoring the importance of including surface attributes in conventional inverse folding.", "keywords": [ "Molecular Surfaces", "Protein Design", "Geometric Deep Learning" ], "rating_list": [ 6, 5, 5, 8, 8 ], "soundness_list": [ 3, 3, 2, 4, 3 ], "presentation_list": [ 3, 4, 2, 4, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 5, 3, 4, 4, 4 ] }, { "paper_id": "JCCPtPDido", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Luca Franceschi", "gender": "Male", "institution": "Amazon Development Center Germany", "country": "", "position": "Researcher" }, { "name": "Pontus Stenetorp", "gender": "Not Specified", "institution": "University College London", "country": "GB", "position": "Full Professor" }, { "name": "Xiangxiang Xu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yao Lu", "gender": "unknown", "institution": "University College London", "country": "", "position": "PhD student" }, { "name": "Yihong Chen", "gender": "unknown", "institution": "University College London", "country": "GB", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 24, "n_ref_uni": 53, "n_ref": 91, "n_ref_all": 123, "n_fig": 12, "n_tab": 6, "L_tab": 28893, "n_element_tab": 192, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2107, "n_element_tab_1": 121, "formula_len_all": 2862, "formula_len_all_1": 2068, "len_all": 209165, "len_all_1": 74545, "len_abs": 861, "len_title": 86, "len_sents": 52636, "len_sents_1": 33720, "n_sents": 396, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 38, "L_abs": 876, "title": "Jet Expansions of Residual Computation", "abs": "We introduce a framework for expanding residual networks using \\textit{jets}, operators that generalize truncated Taylor series.\nOur method provides a systematic approach to disentangle contributions of different computational paths to model predictions.\nIn contrast to existing techniques such as distillation, probing, or early decoding, our expansions rely solely on the model itself and requires no data, training, or sampling from the model.\nWe demonstrate how our framework grounds and subsumes the logit lens,\nreveals a (super-)exponential path structure in the network depth and opens up several applications. \nThese include the extraction of $n$-gram statistics from a transformer large language model, and the definition of data-free toxicity scores.\nOur approach enables data-free analysis of residual networks for model interpretation, development, and evaluation.", "keywords": [ "Interpretability", "residual networks", "transformers", "LLMs", "expansions", "Taylor series", "logit lens" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "JBzTculaVV", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 22, "author_info_list": [ { "name": "Bernard Ghanem", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Full Professor" }, { "name": "Bowen Dong", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Guohao Li", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Huchuan Lu", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Professor" }, { "name": "Jing Shao", "gender": "Female", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Konisberg", "gender": "Not Specified", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Lijun Wang", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Martin Ma", "gender": "unknown", "institution": "Eigent.Ai", "country": "GB", "position": "Researcher" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Prateek Gupta", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Shuyue Hu", "gender": "Female", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Wanli Ouyang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Full Professor" }, { "name": "Xu Jia", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuxian Jiang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhenfei Yin", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Researcher" }, { "name": "Zhiyu Wang", "gender": "Male", "institution": "Northeastern University", "country": "CN", "position": "Undergrad student" }, { "name": "Zijian Ling", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "MS student" }, { "name": "Zirui Zheng", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "MS student" }, { "name": "Ziyi Yang", "gender": "Female", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Ziyue Gan", "gender": "Female", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "\u518d\u658c \u5f20", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 1, "n_ref_uni": 46, "n_ref": 64, "n_ref_all": 80, "n_fig": 8, "n_tab": 17, "L_tab": 2769, "n_element_tab": 314, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 288, "n_element_tab_1": 63, "formula_len_all": 290, "formula_len_all_1": 61, "len_all": 202125, "len_all_1": 64355, "len_abs": 141, "len_title": 165, "len_sents": 65496, "len_sents_1": 31379, "n_sents": 531, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1737, "title": "OASIS: Open Agents Social Interaction Simulations on a Large Scale", "abs": "There has been a growing interest in enhancing rule-based agent-based models (ABMs) for social media platforms (i.e., X, Reddit) with more realistic large language model (LLM) agents, thereby allowing for a more nuanced study of complex systems. As a result, several LLM-based ABMs have been proposed in the past year. While they hold promise, each simulator is specifically designed to study a particular scenario, making it time-consuming and resource-intensive to explore other phenomena using the same ABM. Additionally, these models simulate only a limited number of agents, whereas real-world social media platforms involve millions of users.\nTo this end, we propose OASIS, a generalizable and scalable social media simulator. OASIS is designed based on real-world social media platforms, incorporating dynamically updated environments (i.e., dynamic social networks and post information), diverse action spaces (i.e., following, commenting), and recommendation systems (i.e., interest-based and hot-score-based). Additionally, OASIS supports large-scale user simulations, capable of modeling up to one million users. With these features, OASIS can be easily extended to different social media platforms to study large-scale group phenomena and behaviors. We replicate various social phenomena, including information spreading, group polarization, and herd effects across X and Reddit platforms. \nMoreover, we provide observations of social phenomena at different agent group scales. we observe that the larger agent group scale leads to more enhanced group dynamics and more diverse and helpful agents' opinions. These findings demonstrate OASIS's potential as a powerful tool for studying complex systems in digital environments.", "keywords": [ "Agents; Social Simulation; Multi-agent system" ], "rating_list": [ 3, 8, 5, 1 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 5, 3, 5, 2 ] }, { "paper_id": "JBgBrnhLLL", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Aaron Gokaslan", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "John Xavier Morris", "gender": "Male", "institution": "Cornell University", "country": "", "position": "PhD student" }, { "name": "Srijeeta Biswas", "gender": "unknown", "institution": "Cornell University", "country": "", "position": "MS student" }, { "name": "Subham Sekhar Sahoo", "gender": "Male", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "PhD student" }, { "name": "Vitaly Shmatikov", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Professor" }, { "name": "Volodymyr Kuleshov", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 16, "n_ref_uni": 28, "n_ref": 56, "n_ref_all": 78, "n_fig": 7, "n_tab": 6, "L_tab": 1216, "n_element_tab": 114, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 1443, "n_element_tab_1": 118, "formula_len_all": 1236, "formula_len_all_1": 925, "len_all": 118986, "len_all_1": 62088, "len_abs": 1590, "len_title": 98, "len_sents": 34021, "len_sents_1": 26768, "n_sents": 284, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 50, "L_abs": 932, "title": "Zero-Order Diffusion Guidance for Inverse Problems", "abs": "We propose zero order diffusion guidance, a method that allows using a diffusion model to solve inverse problems without access to the gradients of the process we seek to invert. Our method employs a zero-order gradient estimator combined with a novel differentiable dimensionality reduction strategy to approximate true gradients during guidance while keeping the task computationally tractable in thousands of dimensions. We apply our method to model inversion and demonstrate how it can be used to reconstruct high-quality faces in a realistic scenario where the adversary has only black-box access to face embeddings. Across a range of inverse problems\u2014including synthetic experiments and JPEG restoration\u2014we show that access to gradients is not necessary for effective guidance. Our black-box method matches white-box performance, thus expanding the scope of inverse problems that can be solved with diffusion-based approaches.", "keywords": [ "Privacy", "Zero-Order guidance", "Diffusion Models", "Inverse Problems" ], "rating_list": [ 3, 5, 3, 3, 6 ], "soundness_list": [ 1, 2, 2, 2, 3 ], "presentation_list": [ 1, 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 1, 2 ], "confidence_list": [ 4, 4, 3, 5, 4 ] }, { "paper_id": "JBXO05r4AV", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Han Zhou", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Ruoxi Sun", "gender": "Female", "institution": "Google", "country": "", "position": "Google" }, { "name": "Sercan O Arik", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Xingchen Wan", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 53, "n_ref": 77, "n_ref_all": 120, "n_fig": 4, "n_tab": 16, "L_tab": 13794, "n_element_tab": 939, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 5765, "n_element_tab_1": 273, "formula_len_all": 350, "formula_len_all_1": 140, "len_all": 242394, "len_all_1": 78746, "len_abs": 1347, "len_title": 142, "len_sents": 76660, "len_sents_1": 37613, "n_sents": 392, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1371, "title": "From Few to Many: Enhancing Many-Shot In-Context Learning with Optimized Example Selection and Expansion", "abs": "Recent advances in long-context large language models (LLMs) have led to the emerging paradigm of many-shot in-context learning (ICL), where it is observed that scaling many more demonstrating examples beyond the conventional few-shot setup in the context can lead to performance benefits. However, despite its promise, it is unclear what aspects dominate the benefits and whether simply scaling to more examples is the most effective way of improving many-shot ICL. In this work, we first provide an analysis on the factors driving many-shot ICL, and we find that 1) many-shot performance can still be attributed to often a few disproportionately influential examples and 2) identifying such influential examples (\"optimize\") and using them as demonstrations to regenerate new examples (\"generate\") can lead to further improvements. Inspired by the findings, we propose BRIDGE, an algorithm that alternates between the \\textit{optimize} step with Bayesian optimization to discover the influential sets of examples and the \\textit{generate} step to reuse this set to expand the reasoning paths of the examples back to the many-shot regime automatically. On two state-of-the-art long-context Gemini models of different sizes, we show \\ours led to significant improvements across a diverse set of tasks including symbolic reasoning, numerical reasoning and code generation.", "keywords": [ "many-shot", "in-context learning", "large language models" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "JBOMYYX94D", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Andre Harrison", "gender": "Male", "institution": "U.S. Development Command Army Research Laboratory", "country": "US", "position": "Researcher" }, { "name": "Jacob Clements", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "MS student" }, { "name": "Laura Yu Zheng", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Ming Lin", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "Capital One Endowed Professor" }, { "name": "Shreelekha Revankar", "gender": "Female", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Tony Wu", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "MS student" }, { "name": "Wenjie Wei", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Undergrad student" }, { "name": "Yu Shen", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 31, "n_ref": 49, "n_ref_all": 65, "n_fig": 6, "n_tab": 8, "L_tab": 3604, "n_element_tab": 468, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3067, "n_element_tab_1": 449, "formula_len_all": 449, "formula_len_all_1": 390, "len_all": 120171, "len_all_1": 62897, "len_abs": 1054, "len_title": 105, "len_sents": 31543, "len_sents_1": 26882, "n_sents": 210, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1020, "title": "Sensitivity-Adaptive Augmentation for Robust Segmentation", "abs": "Achieving robustness in image segmentation models is challenging due to the fine-grained nature of pixel-level classification. These models, which are crucial for many real-time perception applications, particularly struggle when faced with natural corruptions. While sensitivity analysis can help us understand how input variables influence model outputs, applying it to natural and uncontrollable corruptions in training data is difficult. In this work, we present an efficient, sensitivity-based augmentation method to enhance robustness against natural corruptions. Our sensitivity analysis approach runs up to 10 times faster and requires up to 200 times less storage than previous approaches, enabling practical, on-the-fly estimation during training for a model-free augmentation policy.\nWith minimal fine-tuning, our sensitivity-based augmentation method achieves improved robustness on both real-world and synthetic datasets compared to state-of-the-art data augmentation techniques in image segmentation tasks.", "keywords": [ "augmentation", "sensitivity analysis", "robustness to corruption" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "JAnyCnK5In", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Tiejun Huang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Yifan Huang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zecheng Hao", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhaofei Yu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zijie Xu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 38, "n_ref": 58, "n_ref_all": 78, "n_fig": 8, "n_tab": 5, "L_tab": 1022, "n_element_tab": 69, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 3195, "n_element_tab_1": 232, "formula_len_all": 2757, "formula_len_all_1": 2577, "len_all": 101390, "len_all_1": 54428, "len_abs": 1379, "len_title": 121, "len_sents": 24698, "len_sents_1": 20898, "n_sents": 151, "n_sents_1": 127, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1383, "title": "Comprehensive Online Training and Deployment for Spiking Neural Networks", "abs": "Spiking Neural Networks (SNNs) are considered to have enormous potential in the future development of Artificial Intelligence (AI) due to their brain-inspired and energy-efficient properties. In the current supervised learning domain of SNNs, compared to vanilla Spatial-Temporal Back-propagation (STBP) training, online training can effectively overcome the risk of GPU memory explosion and has received widespread academic attention. However, the current proposed online training methods cannot tackle the inseparability problem of temporal dependent gradients and merely aim to optimize the training memory, resulting in no performance advantages compared to the STBP training models in the inference phase. To address the aforementioned challenges, we propose Efficient Multi-Precision Firing (EM-PF) model, which is a family of advanced spiking models based on floating-point spikes and binary synaptic weights. We point out that EM-PF model can effectively separate temporal gradients and achieve full-stage optimization towards computation speed and memory footprint. Experimental results have demonstrated that EM-PF model can be flexibly combined with various techniques including random back-propagation, parallel computation and channel attention mechanism, to achieve state-of-the-art performance with extremely low computational overhead in the field of online learning.", "keywords": [ "Efficient Multi-Precision Firing model", "Online Training", "Spiking Neural Networks" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "JAMxRSXLFz", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Mihaela van der Schaar", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Nicolas Astorga", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Tennison Liu", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Katarzyna Kobalczyk", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 35, "n_ref": 62, "n_ref_all": 84, "n_fig": 7, "n_tab": 6, "L_tab": 2617, "n_element_tab": 82, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 835, "n_element_tab_1": 12, "formula_len_all": 1449, "formula_len_all_1": 810, "len_all": 188598, "len_all_1": 81758, "len_abs": 1173, "len_title": 84, "len_sents": 59906, "len_sents_1": 36964, "n_sents": 442, "n_sents_1": 267, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 36, "L_abs": 1177, "title": "Active Task Disambiguation with LLMs", "abs": "Despite the impressive performance of large language models (LLMs) across various benchmarks, their ability to address ambiguously specified problems\u2014frequent in real-world interactions\u2014remains underexplored. To address this gap, we introduce a formal definition of task ambiguity and frame the problem of task disambiguation through the lens of Bayesian Experimental Design. By posing clarifying questions, LLM agents can acquire additional task specifications, progressively narrowing the space of viable solutions and reducing the risk of generating unsatisfactory outputs. Yet, generating effective clarifying questions requires LLM agents to engage in a form of meta-cognitive reasoning, an ability LLMs may presently lack. Our proposed approach of active task disambiguation enables LLM agents to generate targeted questions maximizing the information gain. Effectively, this approach shifts the load from implicit to explicit reasoning about the space of viable solutions. Empirical results demonstrate that this form of question selection leads to more effective task disambiguation in comparison to approaches relying on reasoning solely within the space of questions.", "keywords": [ "Task Ambiguity", "Bayesian Experimental Design", "Large Language Models", "Active Learning" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 2, 4, 3 ], "presentation_list": [ 2, 4, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "J9pNS44qcT", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Canran Xiao", "gender": "Not Specified", "institution": "Central South University", "country": "CN", "position": "PhD student" }, { "name": "Kangsheng Wang", "gender": "Male", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "MS student" }, { "name": "Shuyan Liu", "gender": "unknown", "institution": "Yunnan University", "country": "CN", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 38, "n_ref": 48, "n_ref_all": 55, "n_fig": 3, "n_tab": 2, "L_tab": 895, "n_element_tab": 90, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1197, "n_element_tab_1": 92, "formula_len_all": 602, "formula_len_all_1": 602, "len_all": 93498, "len_all_1": 53818, "len_abs": 1161, "len_title": 150, "len_sents": 25927, "len_sents_1": 25166, "n_sents": 169, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1165, "title": "Enhancing Cooperative Problem-Solving in Sparse-Reward Systems via Co-evolutionary Curriculum Learning", "abs": "Sparse reward environments consistently challenge reinforcement learning, as agents often need to finish tasks before receiving any feedback, leading to limited incentive signals. This issue becomes even more pronounced in multi-agent systems (MAS), where a single reward must be distributed among multiple agents over time, frequently resulting in suboptimal or inconsistent learning outcomes. To tackle this challenge, we introduce a novel approach called Collaborative Multi-dimensional Course Learning (CCL) for multi-agent cooperation scenarios. CCL features three key innovations: (1) It establishes an adaptive curriculum framework tailored for MAS, refining intermediate tasks to individual agents to ensure balanced strategy development. (2) A novel variant evolution algorithm creates more detailed intermediate tasks. (3) Co-evolution between agents and their environment is modeled to enhance training stability under sparse reward conditions. In evaluations across five tasks within multi-particle environments (MPE) and Hide and Seek (Hns), CCL demonstrated superior performance, surpassing existing benchmarks and excelling in sparse reward settings.", "keywords": [ "Reinforcement Learning", "Task Learning" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 3, 1 ], "presentation_list": [ 2, 1, 2, 1 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 5, 3, 5 ] }, { "paper_id": "J9eKm7j6KD", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Omer Sahin Tas", "gender": "unknown", "institution": "FZI Research Center for Information Technology", "country": "DE", "position": "Manager" }, { "name": "Royden Wagner", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "DE", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 49, "n_ref": 90, "n_ref_all": 124, "n_fig": 13, "n_tab": 9, "L_tab": 5804, "n_element_tab": 518, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 670, "n_element_tab_1": 134, "formula_len_all": 103, "formula_len_all_1": 15, "len_all": 170851, "len_all_1": 67968, "len_abs": 2074, "len_title": 129, "len_sents": 42169, "len_sents_1": 30859, "n_sents": 331, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 831, "title": "Words in Motion: Extracting Interpretable Control Vectors for Motion Transformers", "abs": "Transformer-based models generate hidden states that are difficult to interpret. In this work, we aim to interpret these hidden states and control them at inference, with a focus on motion forecasting. We leverage the phenomenon of neural collapse and use linear probes to measure interpretable features in hidden states. Our experiments reveal meaningful directions and distances between hidden states of opposing features, which we use to fit control vectors for activation steering. We further refine our approach using sparse autoencoders to optimize our control vectors. Notably, we show that enforcing sparsity leads to a more linear relationship between control vector temperatures and forecasts. Our approach not only enables mechanistic interpretability but also zero-shot generalization to unseen dataset characteristics.", "keywords": [ "control vectors", "activation steering", "sparse autoencoder", "neural collapse", "motion forecasting" ], "rating_list": [ 5, 3, 3, 8, 5 ], "soundness_list": [ 2, 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 5, 3, 4 ] }, { "paper_id": "J9VogDTa1W", "primary_area": "causal reasoning", "n_author": 6, "author_info_list": [ { "name": "Devendra Singh Dhami", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "Florian Peter Busch", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Jonas Seng", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Kristian Kersting", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Full Professor" }, { "name": "Moritz Willig", "gender": "Male", "institution": "Technical University of Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Tim Nelson Tobiasch", "gender": "Male", "institution": "Computer Science Department, Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 57, "n_ref": 82, "n_ref_all": 102, "n_fig": 3, "n_tab": 5, "L_tab": 646, "n_element_tab": 74, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 247, "n_element_tab_1": 48, "formula_len_all": 1083, "formula_len_all_1": 731, "len_all": 172752, "len_all_1": 75108, "len_abs": 1089, "len_title": 114, "len_sents": 59928, "len_sents_1": 37072, "n_sents": 449, "n_sents_1": 274, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 66, "L_abs": 1092, "title": "Systems with Switching Causal Relations: A Meta-Causal Perspective", "abs": "Most work on causality in machine learning assumes that causal relationships are driven by a constant underlying process. However, the flexibility of agents' actions or tipping points in the environmental process can change the qualitative dynamics of the system. As a result, new causal relationships may emerge, while existing ones change or disappear, resulting in an altered causal graph. To analyze these qualitative changes on the causal graph, we propose the concept of meta-causal states, which groups classical causal models into clusters based on equivalent qualitative behavior and consolidates specific mechanism parameterizations. We demonstrate how meta-causal states can be inferred from observed agent behavior, and discuss potential methods for disentangling these states from unlabeled data. Finally, we direct our analysis towards the application of a dynamical system, showing that meta-causal states can also emerge from inherent system dynamics, and thus constitute more than a context-dependent framework in which mechanisms emerge only as a result of external factors.", "keywords": [ "Meta-Causality", "Meta-Causal Reasoning", "Agent Behavior", "System Dynamics" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 4, 2, 4 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 2, 2, 3, 3 ] }, { "paper_id": "J9SsCtTLga", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Anastasia Martynova", "gender": "unknown", "institution": "Higher School of Economics", "country": "RU", "position": "MS student" }, { "name": "Natalia Semenova", "gender": "Female", "institution": "AIRI", "country": "RU", "position": "Researcher" }, { "name": "Vasilii Latonov", "gender": "Male", "institution": "PJSC Sberbank", "country": "RU", "position": "Researcher" }, { "name": "Vladimir Mashurov", "gender": "Male", "institution": "ITMO University", "country": "RU", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 26, "n_ref": 31, "n_ref_all": 43, "n_fig": 7, "n_tab": 6, "L_tab": 2537, "n_element_tab": 265, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 2611, "n_element_tab_1": 112, "formula_len_all": 341, "formula_len_all_1": 136, "len_all": 110285, "len_all_1": 51772, "len_abs": 547, "len_title": 103, "len_sents": 29767, "len_sents_1": 20467, "n_sents": 300, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 549, "title": "GAN-based NeRF Noise Simulation in Mesh Denoising Task", "abs": "In the present paper, we propose a new approach and a dataset for generating NeRF-like noise on the mesh surface. Our approach is based on GAN and was trained on a dataset that we collect using real NeRF noise. The core idea of our method lies in the use of graph convolutions in the generator. Our pipeline demonstrates generated NeRF-like noise more accurate than other methods by mesh denoising benchmarking. We also present a new NeRF noise analysis approach HTPH based on a conditional probability model to measure the similarity of mesh noise.", "keywords": [ "Generative networks", "Noise generation", "3D Denoising", "Pointclouds" ], "rating_list": [ 6, 1, 6 ], "soundness_list": [ 2, 1, 3 ], "presentation_list": [ 1, 1, 3 ], "contribution_list": [ 1, 1, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "J9Ofr1PmvX", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Mohan Shankara Kankanhalli", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Murari Mandal", "gender": "Male", "institution": "Kalinga Institute of Industrial Technology (KIIT) Bhubaneswar India", "country": "IN", "position": "Assistant Professor" }, { "name": "Yash Sinha", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 40, "n_ref": 75, "n_ref_all": 90, "n_fig": 3, "n_tab": 7, "L_tab": 2963, "n_element_tab": 217, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 131, "n_element_tab_1": 18, "formula_len_all": 593, "formula_len_all_1": 596, "len_all": 167291, "len_all_1": 63247, "len_abs": 1166, "len_title": 114, "len_sents": 44161, "len_sents_1": 31160, "n_sents": 362, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1280, "title": "UnSTAR: Unlearning with Self-Taught Anti-Sample Reasoning for LLMs", "abs": "The key components of machine learning are data samples for training, models for learning patterns, and loss functions for optimizing accuracy. Analogously, unlearning can potentially be achieved through anti-data samples (or anti-samples), unlearning methods, and reversed loss functions. While prior research has explored unlearning methods and reversed loss functions, the potential of anti-samples remains largely untapped. In this paper, we introduce UnSTAR: $\\underline{\\text{Un}}$learning with $\\underline{\\text{S}}$elf-$\\underline{\\text{T}}$aught $\\underline{\\text{A}}$nti-Sample $\\underline{\\text{R}}$easoning for large language models (LLMs). Our contributions are threefold: first, we propose a novel concept of anti-sample-induced unlearning; second, we generate anti-samples by leveraging misleading rationales, which help reverse learned associations and accelerate the unlearning process; and third, we enable fine-grained targeted unlearning, allowing for the selective removal of specific associations without impacting related knowledge\u2014something not achievable by previous works. Results demonstrate that anti-samples offer an efficient, targeted unlearning strategy for LLMs, opening new avenues for privacy-preserving machine learning and model modification.", "keywords": [ "Machine Unlearning", "Large Language Models" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 4, 3 ] }, { "paper_id": "J9FgrqOOni", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Ruichen Xu", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Wenhan Gao", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Yi Liu", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Assistant Professor" }, { "name": "Yuefan Deng", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Full Professor" } ], "n_formula": 52, "n_formula_1": 10, "n_ref_uni": 34, "n_ref": 82, "n_ref_all": 97, "n_fig": 4, "n_tab": 12, "L_tab": 2168, "n_element_tab": 153, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 816, "n_element_tab_1": 71, "formula_len_all": 4447, "formula_len_all_1": 914, "len_all": 187668, "len_all_1": 71623, "len_abs": 1842, "len_title": 133, "len_sents": 62873, "len_sents_1": 33514, "n_sents": 519, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1598, "title": "Discretization-invariance? On the Discretization Mismatch Errors in Neural Operators", "abs": "In recent years, neural operators have emerged as a prominent approach for learning mappings between function spaces, such as the solution operators of parametric PDEs. A notable example is the Fourier Neural Operator (FNO), which models the integral kernel as a convolution operator and uses the Convolution Theorem to learn the kernel directly in the frequency domain. The parameters are decoupled from the resolution of the data, allowing the FNO to take inputs of different resolutions.\nHowever, training at a lower resolution and inferring at a finer resolution does not guarantee consistent performance, nor can fine details, present only in fine-scale data, be learned solely from coarse data. In this work, we address this misconception by defining and examining the discretization mismatch error: the discrepancy between the outputs of the neural operator when using different discretizations of the input data. We demonstrate that neural operators may suffer from discretization mismatch errors that hinder their effectiveness when inferred on data with resolutions different from that of the training data or when trained on data with varying resolutions. As neural operators underpin many critical cross-resolution scientific tasks, such as climate modeling and fluid dynamics, understanding discretization mismatch errors is essential. Based on our findings, we propose a Cross-Resolution Operator-learning Pipeline that is free of aliasing and discretization mismatch errors, enabling efficient cross-resolution and multi-spatial-scale learning, and resulting in superior performance.", "keywords": [ "Neural Operators", "Operator Learning", "Discretization Mismatch Errors", "Discretization Invariance" ], "rating_list": [ 5, 8, 8, 5 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "J8yH8ontdq", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Yu-wing Tai", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "Associate Professor" }, { "name": "Zixuan WANG", "gender": "Not Specified", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Undergrad student" }, { "name": "ChiKeung Tang", "gender": "Not Specified", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 31, "n_ref": 49, "n_ref_all": 69, "n_fig": 4, "n_tab": 6, "L_tab": 1924, "n_element_tab": 136, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2142, "n_element_tab_1": 101, "formula_len_all": 458, "formula_len_all_1": 455, "len_all": 113928, "len_all_1": 49466, "len_abs": 1259, "len_title": 122, "len_sents": 28206, "len_sents_1": 21443, "n_sents": 226, "n_sents_1": 151, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1264, "title": "Audio-Agent: Leveraging LLMs For Audio Generation, Editing and Composition", "abs": "We introduce Audio-Agent, a multimodal framework for audio generation, editing and composition based on text or video inputs. Conventional approaches for text-to-audio (TTA) tasks often make single-pass inferences from text descriptions. While straightforward, this design struggles to produce high-quality audio when given complex text conditions. In our method, we utilize a pre-trained TTA diffusion network as the audio generation agent to work in tandem with GPT-4, which decomposes the text condition into atomic, specific instructions, and calls the agent for audio generation. Consequently, Audio-Agent generates high-quality audio that is closely aligned with the provided text or video while also supporting variable-length generation. For video-to-audio (VTA) tasks, most existing methods require training a timestamp detector to synchronize video events with generated audio, a process that can be tedious and time-consuming. We propose a simpler approach by fine-tuning a pre-trained Large Language Model (LLM), e.g., Gemma2-2B-it, to obtain both semantic and temporal conditions to bridge video and audio modality. Thus our framework provides a comprehensive solution for both TTA and VTA tasks without substantial computational overhead in training.", "keywords": [ "Agent", "Audio Generation", "Large language model", "Diffusion Model" ], "rating_list": [ 8, 3, 5, 1 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "J8YWCBPgx7", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Feng Tian", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Shanyan Guan", "gender": "Male", "institution": "vivo", "country": "CN", "position": "Researcher" }, { "name": "Xiaokang Yang", "gender": "Male", "institution": "Shanghai Jiao Tong University, China", "country": "CN", "position": "Professor" }, { "name": "Yanhao Ge", "gender": "Male", "institution": "Future Imaging Area", "country": "CN", "position": "Researcher" }, { "name": "Yichao Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yixuan Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 24, "n_ref_uni": 43, "n_ref": 96, "n_ref_all": 112, "n_fig": 8, "n_tab": 9, "L_tab": 1347, "n_element_tab": 28, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 226, "n_element_tab_1": 4, "formula_len_all": 1743, "formula_len_all_1": 1410, "len_all": 173195, "len_all_1": 60136, "len_abs": 1263, "len_title": 114, "len_sents": 40083, "len_sents_1": 28369, "n_sents": 316, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1270, "title": "PostEdit: Posterior Sampling for Efficient Zero-Shot Image Editing", "abs": "In the field of image editing, three core challenges persist: controllability, background preservation, and efficiency. Inversion-based methods rely on time-consuming optimization to preserve the features of the initial images, which results in low efficiency due to the requirement for extensive network inference. Conversely, inversion-free methods lack theoretical support for background similarity, as they circumvent the issue of maintaining initial features to achieve efficiency. As a consequence, none of these methods can achieve both high efficiency and background consistency. To tackle the challenges and the aforementioned disadvantages, we introduce PostEdit, a method that incorporates a posterior scheme to govern the diffusion sampling process. Specifically, a corresponding measurement term related to both the initial features and Langevin dynamics is introduced to optimize the estimated image generated by the given target prompt. Extensive experimental results indicate that the proposed PostEdit achieves state-of-the-art editing performance while accurately preserving unedited regions. Furthermore, the method is both inversion- and training-free, necessitating approximately 1.5 seconds and 18 GB of GPU memory to generate high-quality results.", "keywords": [ "Measurements", "Posterior Sampling", "Image Editing", "Inversion-free", "Trianing-free." ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "J8LYjgi7nH", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Fan Ma", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" }, { "name": "Yi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yixuan Han", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Yuchen Xian", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 43, "n_ref_all": 58, "n_fig": 3, "n_tab": 5, "L_tab": 1414, "n_element_tab": 309, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 342, "n_element_tab_1": 87, "formula_len_all": 726, "formula_len_all_1": 386, "len_all": 129285, "len_all_1": 57801, "len_abs": 1963, "len_title": 130, "len_sents": 37380, "len_sents_1": 28954, "n_sents": 265, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1971, "title": "Free-MoE: Tuning-Free Mixture-of-Experts Purifying LLMs to Thrive across Any Field", "abs": "The Mixture-of-Experts (MoE) framework efficiently scales large language models (LLMs) by selectively activating expert subnetworks, reducing computational costs. However, current MoE methods are costly in computation and include additional expert modules that require extra training data for tuning, leading to instability in the optimization process. To address these issues, we introduce Free-MoE, a tuning-free MoE method that leverages pre-trained LLMs' inherent ability to generalize across a wide range of tasks and domains. Free-MoE dynamically activates experts based on specific domains, achieves improvements while 1) requiring no extra model parameters and 2) being completely tuning-free. Specifically, we design the DOWP Alg., a Domain-Oriented Weight Purification Algorithm that purifies the weights in hidden layers and selects the optimal domain-specific experts of domain-specific experts in the hidden layers of the LLM to optimize activation decisions. The activated DSS-Experts, Domain-Specific Subnetwork Experts,can thereby concentrate on specialized task generation, outperforming the corresponding original model. Moreover, Free-MoE incorporates a multi-level trainable router that activates only the most relevant subnetworks during task, effectively minimizing unnecessary inference computations. Comprehensive evaluations reveals that the DOWP Algorithm consistently achieves general performance gains of 2% to 3%, reaching up to 6.8% across datasets like MMLU, HumanEval, GSM8K, and etc. Additionally, when integrated into \\model~framework, our method demonstrates a cumulative improvement of 1.11% in average. Findings indicate that Free-MoE not only enhances overall computational efficiency but improves the model\u2019s adaptability across any field that encompassed in contemporary language generation model benchmarks, and can be seamlessly applied to any transformer-based LLMs. Code for this project will be released in reachable future.", "keywords": [ "Mixture of Experts", "Pretrained LLMs" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 3, 1, 1 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "J863DxU7Sx", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Du Chen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Geoffrey A. Chua", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "Associate Professor" } ], "n_formula": 112, "n_formula_1": 29, "n_ref_uni": 26, "n_ref": 59, "n_ref_all": 104, "n_fig": 6, "n_tab": 4, "L_tab": 1611, "n_element_tab": 117, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 14384, "formula_len_all_1": 2728, "len_all": 184504, "len_all_1": 71866, "len_abs": 1034, "len_title": 157, "len_sents": 52564, "len_sents_1": 28735, "n_sents": 545, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 114, "L_abs": 1128, "title": "Exploiting Hidden Symmetry to Improve Objective Perturbation for DP linear learners with a nonsmooth $\\ell_1$-norm", "abs": "Objective Perturbation (OP) is a classic approach to differentially private (DP) convex optimization with smooth loss functions but is less understood for nonsmooth cases. In this work, we study how to apply OP to DP linear learners under loss functions with an implicit $\\ell_1$-norm structure, such as $\\max(0,x)$ as a motivating example. We propose to first smooth out the implicit $\\ell_1$-norm by convolution, and then invoke standard OP. Convolution has many advantages that distinguish itself from Moreau Envelope, such as approximating from above and a higher degree of hyperparameters. These advantages, in conjunction with the symmetry of $\\ell_1$-norm, result in tighter pointwise approximation, which further facilitates tighter analysis of generalization risks by using pointwise bounds. Under mild assumptions on groundtruth distributions, the proposed OP-based algorithm is found to be rate-optimal, and can achieve the excess generalization risk $\\mathcal{O}(\\frac{1}{\\sqrt{n}}+\\frac{\\sqrt{d\\ln(1/\\delta)}}{n\\varepsilon})$. Experiments demonstrate the competitive performance of the proposed method to Noisy-SGD.", "keywords": [ "Objective Perturbation", "Convolution", "Nonsmooth", "DP-SCO" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 2, 3 ] }, { "paper_id": "J7hIz9GXKq", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Arya Mazumdar", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Harsh Vardhan", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 80, "n_formula_1": 34, "n_ref_uni": 31, "n_ref": 71, "n_ref_all": 93, "n_fig": 8, "n_tab": 4, "L_tab": 1124, "n_element_tab": 55, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 48, "n_element_tab_1": 3, "formula_len_all": 8712, "formula_len_all_1": 3169, "len_all": 193050, "len_all_1": 69661, "len_abs": 1430, "len_title": 138, "len_sents": 53008, "len_sents_1": 30260, "n_sents": 501, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1480, "title": "Collaborative Compressors in Distributed Mean Estimation with Limited Communication Budge", "abs": "Distributed high dimensional mean estimation is a common aggregation routine used often in distributed optimization methods (e.g. federated learning). Most of these applications call for a communication-constrained setting where vectors, whose mean is to be estimated, have to be compressed before sharing. One could independently encode and decode these to achieve compression, but that overlooks the fact that these vectors are often similar with each other. To exploit these similarities, recently Suresh et al., 2022, Jhunjhunwala et al., 2021, Jiang et al, 2023, proposed multiple {\\em correlation-aware compression schemes.} However, in most cases, the correlations have to be known for these schemes to work. Moreover, a theoretical analysis of graceful degradation of these correlation-aware compression schemes with increasing {\\em dissimilarity} is limited to only the $\\ell_2$-error in the literature. \n In this paper, we propose four different collaborative compression schemes that agnostically exploit the similarities among vectors in a distributed setting. Our schemes are all simple to implement and computationally efficient, while resulting in big savings in communication. We do a rigorous theoretical analysis of our proposed schemes to show how the $\\ell_2$, $\\ell_\\infty$ and cosine estimation error varies with the degree of similarity among vectors. In the process, we come up with appropriate dissimilarity-measures for these applications as well.", "keywords": [ "Distributed Mean Estimation", "Compression" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "J6qrIjTzoM", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Yuanzhi Li", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Zeyuan Allen-Zhu", "gender": "Unspecified", "institution": "Meta FAIR Labs", "country": "US", "position": "Researcher" } ], "n_formula": 64, "n_formula_1": 19, "n_ref_uni": 13, "n_ref": 163, "n_ref_all": 275, "n_fig": 35, "n_tab": 8, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 429, "n_element_tab_1": 44, "formula_len_all": 6500, "formula_len_all_1": 1802, "len_all": 267286, "len_all_1": 82092, "len_abs": 981, "len_title": 120, "len_sents": 112492, "len_sents_1": 37048, "n_sents": 846, "n_sents_1": 292, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 936, "title": "Interpretability of Language Models for Learning Hierarchical Structures", "abs": "Transformer-based language models are effective but complex, and understanding their inner workings is a significant challenge. Previous research has primarily explored how these models handle simple tasks like name copying or selection, and we extend this by investigating how these models grasp complex, recursive language structures defined by context-free grammars (CFGs). We introduce a family of synthetic CFGs that produce hierarchical rules, capable of generating lengthy sentences (e.g., hundreds of tokens) that are locally ambiguous and require dynamic programming to parse. Despite this complexity, we demonstrate that generative models like GPT can accurately learn this CFG language and generate sentences based on it. We explore the model's internals, revealing that its hidden states precisely capture the structure of CFGs, and its attention patterns resemble the information passing in a dynamic programming algorithm.", "keywords": [ "generative language models", "interpretability", "induction head", "inner workings" ], "rating_list": [ 6, 8, 3, 8 ], "soundness_list": [ 3, 4, 3, 4 ], "presentation_list": [ 2, 4, 1, 3 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "J6nKxekCCo", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Barry-John Theobald", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Cezanne Camacho", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Instructor" }, { "name": "Falaah Arif Khan", "gender": "unknown", "institution": "New York University", "country": "", "position": "Researcher" }, { "name": "Luca Zappella", "gender": "Male", "institution": "Apple", "country": "ES", "position": "Principal Researcher" }, { "name": "Nicholas Apostoloff", "gender": "unknown", "institution": "Apple", "country": "GB", "position": "Principal Researcher" }, { "name": "Nivedha Sivakumar", "gender": "Not Specified", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "MS student" }, { "name": "Yinong Oliver Wang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Rin Metcalf", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 59, "n_ref": 91, "n_ref_all": 129, "n_fig": 6, "n_tab": 21, "L_tab": 21547, "n_element_tab": 1373, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 4141, "n_element_tab_1": 272, "formula_len_all": 78, "formula_len_all_1": 0, "len_all": 212912, "len_all_1": 64552, "len_abs": 3445, "len_title": 122, "len_sents": 44165, "len_sents_1": 29393, "n_sents": 273, "n_sents_1": 165, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1722, "title": "Uncovering Intersectional Stereotypes in Humans and Large Language Models", "abs": "Recent work has shown that Large Language Models (LLMs) learn and reproduce pre-existing biases in their training corpora, such as preferences for socially privileged identities (e.g., men or White people) and prejudices against socially marginalized identities (e.g., women or Black people). Current evaluations largely focus on single-attribute discrimination (e.g., gender stereotypes). By contrast, we investigate intersectional stereotypical bias (e.g., against Black women) as these social groups face unique challenges that cannot be explained by any single aspect of their identity alone. Our contributions in this work are two-fold: First, we design and release a new fairness benchmark for intersectional stereotypes in LLMs by augmenting the WinoBias corpus using 25 demographic markers including gender identity, body type, and disability.\nWe use this benchmark to evaluate the fairness of five causal LLMs through the lens of uncertainty, and find that they are disparately uncertain for intersectional identities on the pronoun-occupation co-reference resolution task, indicating systematic intersectional stereotypical bias. Second, we build on cognitive psychology research on stereotypes in human society, by using LLMs to detect stereotypes against intersectional identities that have previously not been studied in the social sciences. Drawing from the seminal warmth-competence stereotype content model, we compare stereotypes in LLMs to stereotypes produced by human annotators and report statistically significant alignment between the two. Our findings underscore the potential for LLMs to be used to conduct social psychology research that could otherwise be harmful to conduct with human subjects.", "keywords": [ "fairness", "uncertainty quantification", "intersectionality", "stereotypes", "social psychology", "cognitive psychology" ], "rating_list": [ 3, 1, 5, 3 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "J6e4hurEKd", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Chenglong Kang", "gender": "unknown", "institution": "Central South University", "country": "CN", "position": "MS student" }, { "name": "Fei Guo", "gender": "Female", "institution": "Central South University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoyi Liu", "gender": "Female", "institution": "University of South Carolina, University of South Carolina", "country": "US", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 17, "n_ref_uni": 43, "n_ref": 87, "n_ref_all": 91, "n_fig": 3, "n_tab": 9, "L_tab": 2150, "n_element_tab": 250, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1331, "n_element_tab_1": 186, "formula_len_all": 589, "formula_len_all_1": 513, "len_all": 142789, "len_all_1": 56840, "len_abs": 1745, "len_title": 182, "len_sents": 36226, "len_sents_1": 26181, "n_sents": 233, "n_sents_1": 162, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 132, "L_abs": 1754, "title": "RetroInText: A Multimodal Large Language Model Enhanced Framework for Retrosynthetic Planning via In-Context Representation Learning", "abs": "Development of robust and effective strategies for retrosynthetic planning requires a deep understanding of the synthesis process. A critical step in achieving this goal is accurately identifying synthetic intermediates. Current machine learning-based methods often overlook the valuable context from the overall route, focusing only on predicting reactants from the product, requiring cost annotations for every reaction step, and ignoring the multi-faced nature of molecular, resulting in inaccurate synthetic route predictions. Therefore, we introduce RetroInText, an advanced end-to-end framework based on a multimodal Large Language Model (LLM), featuring in-context learning with TEXT descriptions of synthetic routes. First, RetroInText including ChatGPT presents detailed descriptions of the reaction procedure. It learns the distinct compound representations in parallel with corresponding molecule encoders to extract multi-modal representations including 3D features. Subsequently, we propose an attention-based mechanism that offers a fusion module to complement these multi-modal representations with in-context learning and a fine-tuned LLM for a single-step model. As a result, RetroInText accurately represents and effectively captures the complex relationship between molecules and the synthetic route. In experiments on the USPTO pathways dataset RetroBench, RetroInText outperformed state-of-the-art methods, achieving up to a 5% improvement in Top-1 test accuracy, particularly for long synthetic routes. These results demonstrate the superiority of RetroInText by integrating with context information over routes. They also demonstrate its potential for advancing pathway design and facilitating the development of organic chemistry.", "keywords": [ "Retrosynthetic Planning", "Route Evaluation", "In-Context Learning", "Large Language Model" ], "rating_list": [ 8, 3, 6, 8 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 2, 1, 4, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 4, 2, 4 ] }, { "paper_id": "J6dEAiPPe0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Batoul Daou", "gender": "unknown", "institution": "Technische Universit\u00e4t Hamburg", "country": "DE", "position": "MS student" }, { "name": "Christian Feiler", "gender": "Male", "institution": "Helmholtz-Zentrum Hereon", "country": "DE", "position": "Principal Researcher" }, { "name": "Christian J Cyron", "gender": "unknown", "institution": "Hamburg University of Technology", "country": "", "position": "Full Professor" }, { "name": "Jonas F. Eichinger", "gender": "unknown", "institution": "Helmholz-Zentrum Hereon", "country": "DE", "position": "Researcher" }, { "name": "Kevin Linka", "gender": "Male", "institution": "Technische Universit\u00e4t Hamburg", "country": "DE", "position": "Lecturer" }, { "name": "Marius Tacke", "gender": "unknown", "institution": "Helmholtz Zentrum hereon", "country": "DE", "position": "PhD student" }, { "name": "Matthias Busch", "gender": "Male", "institution": "Technische Universit\u00e4t Hamburg", "country": "DE", "position": "PhD student" }, { "name": "Mikhail Zheludkevich", "gender": "Male", "institution": "Christian-Albrechts-Universit\u00e4t Kiel", "country": "DE", "position": "Full Professor" }, { "name": "Roland Aydin", "gender": "Male", "institution": "Technical University Hamburg", "country": "", "position": "Assistant Professor" }, { "name": "Sviatlana V. Lamaka", "gender": "Female", "institution": "Helmholtz-Zentrum Hereon ", "country": "DE", "position": "Principal Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 21, "n_ref": 37, "n_ref_all": 54, "n_fig": 5, "n_tab": 3, "L_tab": 589, "n_element_tab": 24, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 383, "n_element_tab_1": 7, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 128540, "len_all_1": 58219, "len_abs": 1128, "len_title": 144, "len_sents": 40053, "len_sents_1": 30015, "n_sents": 312, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1133, "title": "Large Language Model Predicting the Performance of Small Organic Molecule Corrosion Inhibitors", "abs": "Large language models (LLMs) like GPT-4o have shown promise in solving everyday tasks and addressing basic scientific challenges by utilizing extensive pre-trained knowledge. In this work, we explore their potential to predict the efficiency of various organic compounds for the inhibition of corrosion of the magnesium alloy ZE41, a material crucial for many industrial applications. Traditional approaches, such as basic neural networks, rely on non-contextual data, often requiring large datasets and significant effort per sample to achieve accurate predictions. They struggle particularly with small datasets, limiting their effectiveness in discovering new corrosion inhibitors. LLMs can contextualize and interpret limited data points by drawing on their vast knowledge, including chemical properties of molecules and their influence on corrosion processes in other materials like iron. By prompting the model with a small dataset, LLMs can provide meaningful predictions without the need for extensive training. Our study demonstrates that LLMs can predict corrosion inhibition outcomes, and reduce the amount of data needed.", "keywords": [ "AI", "LLM", "GPT-4o", "Machine Learning", "Material Science", "Corrosion", "Magnesium", "Regression", "Prediction", "Foundation Model" ], "rating_list": [ 3, 1, 3 ], "soundness_list": [ 2, 1, 3 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 1, 1, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "J6Xgra2bE5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Alexander Bronstein", "gender": "Male", "institution": "Technion", "country": "", "position": "Full Professor" }, { "name": "Chaim Baskin", "gender": "Male", "institution": "Ben Gurion University of the Negev", "country": "IL", "position": "Assistant Professor" }, { "name": "Moshe Kimhi", "gender": "Non-Binary", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Tsachi Blau", "gender": "Male", "institution": "Computer Science Department, Technion - Israel Institute of Technology", "country": "IL", "position": "PhD student" }, { "name": "Yonatan Belinkov", "gender": "Male", "institution": "Technion, Technion", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 41, "n_ref": 65, "n_ref_all": 72, "n_fig": 3, "n_tab": 3, "L_tab": 2090, "n_element_tab": 132, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2291, "n_element_tab_1": 144, "formula_len_all": 265, "formula_len_all_1": 233, "len_all": 108579, "len_all_1": 52464, "len_abs": 1636, "len_title": 131, "len_sents": 24809, "len_sents_1": 23468, "n_sents": 175, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1644, "title": "Context-aware Prompt Tuning: Advancing In-Context Learning with Adversarial Methods", "abs": "Large Language Models (LLMs) can perform few-shot learning using either optimization-based approaches or In-Context Learning (ICL). Optimization-based methods often suffer from overfitting, as they require updating a large number of parameters with limited data. In contrast, ICL avoids overfitting but typically underperforms compared to optimization-based methods and is highly sensitive to the selection, order, and format of demonstration examples. To overcome these challenges, we introduce Context-aware Prompt Tuning (CPT), a method inspired by ICL, Prompt Tuning (PT), and adversarial attacks. CPT builds on the ICL strategy of concatenating examples before the input, extending it by incorporating PT-like learning to refine the context embedding through iterative optimization, extracting deeper insights from the training examples. Our approach carefully modifies specific context tokens, considering the unique structure of the examples within the context.In addition to updating the context with PT-like optimization, CPT draws inspiration from adversarial attacks, adjusting the input based on the labels present in the context while preserving the inherent value of the user-provided data. To ensure robustness and stability during optimization, we employ a projected gradient descent algorithm, constraining token embeddings to remain close to their original values and safeguarding the quality of the context. Our method has demonstrated superior accuracy across multiple classification tasks using various LLM models, outperforming existing baselines and effectively addressing the overfitting challenge in few-shot learning.", "keywords": [ "Large Language Models", "In-Context Learning", "and Adversarial Attacks" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "J5sUOvlLbQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Alessandro Favero", "gender": "Male", "institution": "EPFL", "country": "CH", "position": "PhD student" }, { "name": "Guillermo Ortiz-Jimenez", "gender": "unknown", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Ke Wang", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Nikolaos Dimitriadis", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Pascal Frossard", "gender": "Not Specified", "institution": "EPFL", "country": "CH", "position": "Full Professor" }, { "name": "Francois Fleuret", "gender": "Male", "institution": "University of Geneva", "country": "", "position": "Full Professor" } ], "n_formula": 40, "n_formula_1": 3, "n_ref_uni": 90, "n_ref": 224, "n_ref_all": 270, "n_fig": 15, "n_tab": 10, "L_tab": 2486, "n_element_tab": 223, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 984, "n_element_tab_1": 62, "formula_len_all": 22271, "formula_len_all_1": 163, "len_all": 271791, "len_all_1": 60584, "len_abs": 1375, "len_title": 129, "len_sents": 63881, "len_sents_1": 28649, "n_sents": 395, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1360, "title": "LiNeS: Post-training Layer Scaling Prevents Forgetting and Enhances Model Merging", "abs": "Large pre-trained models exhibit impressive zero-shot performance across diverse tasks, but fine-tuning often leads to catastrophic forgetting, where improvements on a target domain degrade generalization on other tasks. To address this challenge, we introduce LiNeS, Layer-Increasing Network Scaling, a post-training editing technique designed to preserve pre-trained generalization while enhancing fine-tuned task performance. LiNeS scales parameter updates linearly based on their layer depth within the network, maintaining shallow layers close to their pre-trained values to preserve general features, while allowing deeper layers to retain task-specific representations.\nWe further extend this approach to multi-task model merging scenarios, where layer-wise scaling of merged parameters reduces negative task interference. LiNeS demonstrates significant improvements in both single-task and multi-task settings across various benchmarks in vision and natural language processing. It enhances out-of-distribution generalization, integrates seamlessly with existing multi-task model merging baselines enchancing their performance across bcenchmarks and model sizes, and can boost generalization when merging LLM policies aligned with different rewards via RLHF. Importantly, our method is simple to implement and complementary to many existing techniques.", "keywords": [ "Model merging; Model editing; Catastrophic forgetting; Multi-task learning; OOD generalization" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "J5s6EG6ual", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Bu-Sung Lee", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Dorvin Ong", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "Undergrad student" }, { "name": "JIANXIAO SUN", "gender": "Not Specified", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 62, "n_ref": 85, "n_ref_all": 102, "n_fig": 6, "n_tab": 2, "L_tab": 397, "n_element_tab": 67, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 75, "n_element_tab_1": 6, "formula_len_all": 45, "formula_len_all_1": 45, "len_all": 139725, "len_all_1": 62842, "len_abs": 1459, "len_title": 140, "len_sents": 40285, "len_sents_1": 33687, "n_sents": 253, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1465, "title": "Investigating Self-Attention: Its Impact on Sample Efficiency in Deep Reinforcement Learning", "abs": "Improving the sample efficiency of deep reinforcement learning (DRL) agents has been an ongoing challenge in research and real-world applications. Self-attention, a mechanism originally popularized in natural language processing, has shown great potential in enhancing sample efficiency when integrated with traditional DRL algorithms. However, the impact of self-attention mechanisms on the sample efficiency of DRL models has not been fully studied. In this paper, we ponder the fundamental operation of the self-attention mechanism in visual-based DRL settings and systematically investigate how different types of scaled dot-product attention affect the sample efficiency of the DRL algorithms. We design and evaluate the performance of our self-attention DRL models in the Arcade Learning Environment. Our results suggest that each self-attention module design has a distinct impact on the sample complexity of the DRL agent. To understand the influence of self-attention modules on the learning process, we conduct an interpretability study focusing on state representation and exploration. From our initial findings, the interplay between feature extraction, action selection, and reward collection is influenced subtly by the inductive biases of the proposed self-attention modules. This work contributes to the ongoing efforts to optimize DRL architectures, offering insights into the mechanisms that can enhance their performance in data-scarce scenarios.", "keywords": [ "self-attention", "sample efficiency", "reinforcement learning" ], "rating_list": [ 5, 3, 1 ], "soundness_list": [ 3, 2, 1 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "J4xLuCt2kg", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Linyi Yang", "gender": "unknown", "institution": "University College London", "country": "GB", "position": "Researcher" }, { "name": "Qingqiu Li", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Qiujie Xie", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Yue Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" }, { "name": "Yuejie Zhang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Zhuohao Yu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 54, "n_ref": 160, "n_ref_all": 198, "n_fig": 9, "n_tab": 15, "L_tab": 5910, "n_element_tab": 488, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 3000, "n_element_tab_1": 269, "formula_len_all": 52, "formula_len_all_1": 52, "len_all": 248814, "len_all_1": 75319, "len_abs": 1610, "len_title": 120, "len_sents": 64584, "len_sents_1": 33286, "n_sents": 521, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1390, "title": "An Empirical Analysis of Uncertainty in Large Language Model Evaluations", "abs": "As LLM-as-a-Judge emerges as a new paradigm for assessing large language models (LLMs), concerns have been raised regarding the alignment, bias, and stability of LLM evaluators. While substantial work has focused on alignment and bias, little research has concentrated on the stability of LLM evaluators. In this paper, we conduct extensive experiments involving 9 widely used LLM evaluators across 2 different evaluation settings to investigate the uncertainty in model-based LLM evaluations. We pinpoint that LLM evaluators exhibit varying uncertainty based on model families and sizes. With careful comparative analyses, we find that employing special prompting strategies, whether during inference or post-training, can alleviate evaluation uncertainty to some extent. By utilizing uncertainty to enhance LLM's reliability and detection capability in Out-Of-Distribution (OOD) data, we further fine-tune an uncertainty-aware LLM evaluator named ConfiLM using a human-annotated fine-tuning set and assess ConfiLM's OOD evaluation ability on a manually designed test set sourced from the 2024 Olympics. Experimental results demonstrate that incorporating uncertainty as additional information during the fine-tuning phase can largely improve the model's evaluation performance in OOD scenarios. The code and data are released at: https://github.com/hasakiXie123/LLM-Evaluator-Uncertainty.", "keywords": [ "Large Language Model", "Model-based LLM Evaluation", "LLM-as-a-Judge" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "J4D5WVoc5g", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Cewu Lu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Pengfei Xie", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Wenqiang Xu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "", "position": "PhD student" }, { "name": "Yutong Li", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "MS student" }, { "name": "Zhenjun Yu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 31, "n_ref": 61, "n_ref_all": 69, "n_fig": 5, "n_tab": 4, "L_tab": 620, "n_element_tab": 104, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 145, "n_element_tab_1": 42, "formula_len_all": 585, "formula_len_all_1": 546, "len_all": 115151, "len_all_1": 58292, "len_abs": 1484, "len_title": 151, "len_sents": 32130, "len_sents_1": 26565, "n_sents": 245, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 1491, "title": "Dynamic Reconstruction of Hand-Object Interaction with Distributed Force-aware Contact Representation", "abs": "We present ViTaM-D, a novel visual-tactile framework for dynamic hand-object interaction reconstruction, integrating distributed tactile sensing for more accurate contact modeling. While existing methods focus primarily on visual inputs, they struggle with capturing detailed contact interactions such as object deformation. Our approach leverages distributed tactile sensors to address this limitation by introducing DF-Field. This distributed force-aware contact representation models both kinetic and potential energy in hand-object interaction.\nViTaM-D first reconstructs hand-object interactions using a visual-only network, VDT-Net, and then refines contact details through a force-aware optimization (FO) process, enhancing object deformation modeling. To benchmark our approach, we introduce the HOT dataset, which features 600 sequences of hand-object interactions, including deformable objects, built in a high-precision simulation environment.\nExtensive experiments on both the DexYCB and HOT datasets demonstrate significant improvements in accuracy over previous state-of-the-art methods such as gSDF and HOTrack. Our results highlight the superior performance of ViTaM-D in both rigid and deformable object reconstruction, as well as the effectiveness of DF-Field in refining hand poses. This work offers a comprehensive solution to dynamic hand-object interaction reconstruction by seamlessly integrating visual and tactile data. Codes, models, and datasets will be available.", "keywords": [ "Tactile sensing", "hand-object tracking and reconstruction" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "J3xRByRqOz", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Hyejin Park", "gender": "Female", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Jeongyeon Hwang", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Jungseul Ok", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Associate Professor" }, { "name": "Junyoung Park", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Undergrad student" }, { "name": "Sangdon Park", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 8, "n_ref_uni": 48, "n_ref": 90, "n_ref_all": 117, "n_fig": 16, "n_tab": 14, "L_tab": 2822, "n_element_tab": 370, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 724, "n_element_tab_1": 77, "formula_len_all": 1370, "formula_len_all_1": 484, "len_all": 213196, "len_all_1": 67805, "len_abs": 1211, "len_title": 116, "len_sents": 66318, "len_sents_1": 32334, "n_sents": 560, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1204, "title": "RETRIEVAL-AUGMENTED GENERATION WITH ESTIMATION OF SOURCE RELIABILITY", "abs": "Retrieval-augmented generation (RAG) addresses key limitations of large language models (LLMs), such as hallucinations and outdated knowledge, by incorporating external databases. These databases typically consult multiple sources to encompass up-to-date and various information. However, standard RAG methods often overlook the heterogeneous source reliability in the multi-source database and retrieve documents solely based on relevance, making them prone to propagating misinformation. To address this, we propose Reliability-Aware RAG (RA-RAG) which estimates the reliability of multiple sources and incorporates this information into both retrieval and aggregation processes. Specifically, it iteratively estimates source reliability and true answers for a set of queries with no labelling. Then, it selectively retrieves relevant documents from a few of reliable sources and aggregates them using weighted majority voting, where the selective retrieval ensures scalability while not compromising the performance. We also introduce a benchmark designed to reflect real-world scenarios with heterogeneous source reliability and demonstrate the effectiveness of RA-RAG compared to a set of baselines.", "keywords": [ "Retrieval-augmented Generation (RAG)", "Trustworthy AI" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 4, 2, 2, 4 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "J3H8Az3YlB", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Nazia Tasnim", "gender": "Female", "institution": "Boston University, Boston University", "country": "", "position": "PhD student" }, { "name": "Bryan Allen Plummer", "gender": "Male", "institution": "Boston University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 12, "n_ref_uni": 41, "n_ref": 68, "n_ref_all": 95, "n_fig": 11, "n_tab": 10, "L_tab": 4368, "n_element_tab": 388, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 2765, "n_element_tab_1": 128, "formula_len_all": 926, "formula_len_all_1": 803, "len_all": 164912, "len_all_1": 60636, "len_abs": 2294, "len_title": 119, "len_sents": 44597, "len_sents_1": 27889, "n_sents": 335, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1956, "title": "RECAST: Reparameterized, Compact weight Adaptation for Sequential Tasks", "abs": "Incremental learning aims to adapt to new sets of categories over time with minimal computational overhead. Prior work often addresses this task by training efficient task-specific adaptors that modify frozen layer weights or features to capture relevant information without affecting predictions on any previously learned categories. While these adaptors are generally more efficient than finetuning the entire network, they still can require tens to hundreds of thousands task-specific trainable parameters even for relatively small networks, making it challenging to operate on resource-constrained environments with high communication costs like edge devices or mobile phones. Thus, we propose Reparameterized, Compact weight Adaptation for Sequential Tasks (RECAST), a novel method that dramatically reduces the number of task-specific trainable parameters to fewer than 50 \u2013 several orders of magnitude less than competing methods like LoRA. RECAST accomplishes this efficiency by learning to decompose layer weights into a soft parameter-sharing framework consisting of a set of shared weight templates and very few module-specific scaling factors or coefficients. This soft parameter-sharing framework allows for effective task-wise reparameterization by tuning only these coefficients while keeping templates frozen. A key innovation of RECAST is the novel weight reconstruction pipeline called Neural Mimicry, which eliminates the need for pretraining from scratch. This allows for high-fidelity emulation of existing pretrained weights within our framework and provides quick adaptability to any model scale and architecture. Extensive experiments across six diverse datasets demonstrate RECAST outperforms the state-of-the-art by up to 3% across various scales, architectures, and parameter spaces. Moreover, we show that RECAST\u2019s architecture-agnostic nature allows for seamless integration with existing methods, further boosting performance.", "keywords": [ "reparameterization", "weight decomposition", "weight reconstruction", "task-incremental learning", "transfer learning", "lifelong learning", "incremental learning" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "J2we1sVd9m", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Ao Ke", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Chuanwen Feng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Lei Feng", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" }, { "name": "S Kevin Zhou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Wenlong Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Xike Xie", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Research Professor" }, { "name": "Yukun Cao", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 41, "n_ref": 86, "n_ref_all": 109, "n_fig": 5, "n_tab": 8, "L_tab": 7227, "n_element_tab": 680, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 5505, "n_element_tab_1": 362, "formula_len_all": 752, "formula_len_all_1": 790, "len_all": 142199, "len_all_1": 69549, "len_abs": 1044, "len_title": 115, "len_sents": 34448, "len_sents_1": 27883, "n_sents": 251, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1047, "title": "Prototype-based Optimal Transport for Out-of-Distribution Detection", "abs": "Detecting Out-of-Distribution (OOD) inputs is crucial for improving the reliability of deep neural networks in the real-world deployment. In this paper, inspired by the inherent distribution shift between ID and OOD data, we propose a novel method that leverages optimal transport to measure the distribution discrepancy between test inputs and ID prototypes. The resulting transport costs are used to quantify the individual contribution of each test input to the overall discrepancy, serving as a desirable measure for OOD detection. To address the issue that solely relying on the transport costs to ID prototypes is inadequate for identifying OOD inputs closer to ID data, we generate virtual outliers to approximate the OOD region via linear extrapolation. By combining the transport costs to ID prototypes with the costs to virtual outliers, the detection of OOD data near ID data is emphasized, thereby enhancing the distinction between ID and OOD inputs. Experiments demonstrate the superiority of our method over state-of-the-art methods.", "keywords": [ "Out-Of-Distribution Detection", "Optimal Transport" ], "rating_list": [ 3, 3, 6, 8, 3 ], "soundness_list": [ 1, 3, 3, 4, 2 ], "presentation_list": [ 2, 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 4, 3 ], "confidence_list": [ 5, 5, 3, 4, 5 ] }, { "paper_id": "J2Jyp1SZ0n", "primary_area": "datasets and benchmarks", "n_author": 13, "author_info_list": [ { "name": "Chunyuan Li", "gender": "unknown", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Dongzhi Jiang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Guanglu Song", "gender": "Male", "institution": "Sensetime", "country": "CN", "position": "Computer Vision Researcher" }, { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Pengshuo Qiu", "gender": "Male", "institution": "Tohoku University", "country": "JP", "position": "Undergrad student" }, { "name": "Renrui Zhang", "gender": "Male", "institution": "MMLab of CUHK & Shanghai AI Laboratory", "country": "HK", "position": "PhD student" }, { "name": "Yanmin Wu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yu Liu", "gender": "Male", "institution": "SenseTime", "country": "HK", "position": "Principal Researcher" }, { "name": "Zehui Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Ziyu Guo", "gender": "Female", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "jiayi lei", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "", "position": "Undergrad student" }, { "name": "Pan Lu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Gao Peng", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 42, "n_ref": 111, "n_ref_all": 131, "n_fig": 7, "n_tab": 10, "L_tab": 5946, "n_element_tab": 314, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1018, "n_element_tab_1": 28, "formula_len_all": 13, "formula_len_all_1": 67, "len_all": 258466, "len_all_1": 60996, "len_abs": 1827, "len_title": 130, "len_sents": 66189, "len_sents_1": 26872, "n_sents": 566, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1836, "title": "MMSearch: Benchmarking the Potential of Large Models as Multi-modal Search Engines", "abs": "The advent of Large Language Models (LLMs) has paved the way for AI search engines, e.g., SearchGPT, showcasing a new paradigm in human-internet interaction. However, most current AI search engines are limited to text-only settings, neglecting the multimodal user queries and the text-image interleaved nature of website information. Recently, Large Multimodal Models (LMMs) have made impressive strides. Yet, whether they can function as AI search engines remains under-explored, leaving the potential of LMMs in multimodal search an open question. To this end, we first design a delicate pipeline, MMSearch-Engine, to empower any LMMs with multimodal search capabilities. On top of this, we introduce MMSearch, a comprehensive evaluation benchmark to assess the multimodal search performance of LMMs. The curated dataset contains 300 manually collected instances spanning 14 subfields, which involves no overlap with the current LMMs' training data, ensuring the correct answer can only be obtained within searching. By using MMSearch-Engine, the LMMs are evaluated by performing three individual tasks (requery, rerank, and summarization), and one challenging end-to-end task with a complete searching process. We conduct extensive experiments on closed-source and open-source LMMs. Among all tested models, GPT-4o with MMSearch-Engine achieves the best results, which surpasses the commercial product, Perplexity Pro, in the end-to-end task, demonstrating the effectiveness of our proposed pipeline. We further present error analysis to unveil current LMMs still struggle to fully grasp the multimodal search tasks, and conduct ablation study to indicate the potential of scaling test-time computation for AI search engine. We hope MMSearch may provide unique insights to guide the future development of multimodal AI search engine.", "keywords": [ "Large Multimodal Model", "AI Search Engine", "Benchmark" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "J2FyEVg8HR", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Eric P. Xing", "gender": "Male", "institution": "Mohamed bin Zayed Univeristy of AI", "country": "AE", "position": "Full Professor" }, { "name": "Hongyi Wang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Assistant Professor" }, { "name": "Huaxiu Yao", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Souvik Kundu", "gender": "Male", "institution": "Intel", "country": "US", "position": "Researcher" }, { "name": "Wenhao Zheng", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Yixiao Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Undergrad student" }, { "name": "Yun Li", "gender": "Not Specified", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" }, { "name": "Zhengzhong Liu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "Weitong ZHANG", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 36, "n_ref": 55, "n_ref_all": 64, "n_fig": 8, "n_tab": 1, "L_tab": 349, "n_element_tab": 64, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 349, "n_element_tab_1": 64, "formula_len_all": 989, "formula_len_all_1": 1035, "len_all": 131831, "len_all_1": 59656, "len_abs": 1434, "len_title": 147, "len_sents": 40372, "len_sents_1": 28280, "n_sents": 280, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1480, "title": "CITER: Collaborative Inference for Efficient Large Language Model Decoding with Token-Level Routing", "abs": "Large language models (LLMs) have achieved remarkable success in natural language processing tasks but suffer from high computational costs during inference, limiting their deployment in latency-constrained applications. To address this issue, we propose a novel \\textbf{C}ollaborative \\textbf{I}nference with \\textbf{T}oken-l\\textbf{E}vel \\textbf{R}outing (CITER) framework that introduces a token-level routing mechanism, enabling efficient collaboration between small and large language models (SLMs \\& LLMs). Specifically, CITER enables routing non-critical tokens to an SLM to reduce computational overhead, while critical tokens are processed by an LLM to maintain generation quality. We formulate the training of the router as a reinforcement learning task, where the router receives rewards based on both the quality of predictions and the inference cost of generation. This allows the router to learn to predict token-level routing scores and make routing decisions based on both the current token and the future impact of its decisions. To further accelerate reward evaluation process, we introduce a shortcut for reward function estimation, significantly reducing the cost of the reward estimation and improving the practicality of our approach. Extensive experiments across four benchmark datasets demonstrate that CITER reduces inference cost while preserving high-quality generation, offering a promising solution for real-time and resource-constrained applications.", "keywords": [ "collaborative inference", "efficient inference", "token-level routing", "large language model" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 2, 3 ] }, { "paper_id": "J2EmNMLoxv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Anurag Kumar", "gender": "Male", "institution": "Meta", "country": "US", "position": "Research Scientist" }, { "name": "Gael Le Lan", "gender": "Male", "institution": "Meta", "country": "FR", "position": "Researcher" }, { "name": "Haohe Liu", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "PhD student" }, { "name": "Mark D Plumbley", "gender": "unknown", "institution": "University of Surrey", "country": "", "position": "Full Professor" }, { "name": "Vikas Chandra", "gender": "Male", "institution": "Meta", "country": "US", "position": "Director, AI" }, { "name": "Wenwu Wang", "gender": "Male", "institution": "University of Surrey", "country": "", "position": "Full Professor" }, { "name": "Xinhao Mei", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Yangyang Shi", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Zhaoheng Ni", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Varun Nagaraja", "gender": "Male", "institution": "Meta", "country": "US", "position": "Research Scientist" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 52, "n_ref": 103, "n_ref_all": 124, "n_fig": 6, "n_tab": 4, "L_tab": 1143, "n_element_tab": 127, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1497, "n_element_tab_1": 174, "formula_len_all": 700, "formula_len_all_1": 723, "len_all": 159907, "len_all_1": 70855, "len_abs": 1321, "len_title": 115, "len_sents": 39876, "len_sents_1": 32573, "n_sents": 264, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1326, "title": "SyncFlow: Temporally Aligned Joint Audio-Video Generation from Text", "abs": "Video and audio are closely correlated modalities that humans naturally perceive together. While recent advancements have enabled the generation of audio or video from text, producing both modalities simultaneously still typically relies on either a cascaded process or multi-modal contrastive encoders. These approaches, however, often lead to suboptimal results due to inherent information losses during inference and conditioning. In this paper, we introduce SyncFlow, a system that is capable of simultaneously generating temporally synchronized audio and video from text. The core of SyncFlow is the proposed dual-diffusion-transformer (d-DiT) architecture, which enables joint video and audio modelling with proper information fusion. To efficiently manage the computational cost of joint audio and video modelling, SyncFlow utilizes a multi-stage training strategy that separates video and audio learning before joint fine-tuning. Our empirical evaluations demonstrate that SyncFlow produces audio and video outputs that are more correlated than baseline methods with significantly enhanced audio quality and audio-visual correspondence. Moreover, we demonstrate strong zero-shot capabilities of SyncFlow, including zero-shot video-to-audio generation and adaptation to novel video resolutions without further training.", "keywords": [ "Text-to-Audio-Video-Joint Generation", "Flow matching", "Diffusion Transformer" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "J1xtkJmFY3", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Kaichen Xu", "gender": "Male", "institution": "Zhongnan University of Economics and Law", "country": "CN", "position": "Researcher" }, { "name": "Yueyang Ding", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 67, "n_ref": 92, "n_ref_all": 108, "n_fig": 5, "n_tab": 6, "L_tab": 1790, "n_element_tab": 197, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1790, "n_element_tab_1": 197, "formula_len_all": 237, "formula_len_all_1": 237, "len_all": 206720, "len_all_1": 68240, "len_abs": 1314, "len_title": 128, "len_sents": 55964, "len_sents_1": 31836, "n_sents": 426, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1320, "title": "ZerOmics: Toward General Models for Single-Cell Analysis with Instruction Tuning", "abs": "A variety of analysis tasks in single-cell (SC) multi-omics are crucial for precision medicine and clinical research. To address these tasks, existing methods are typically pre-trained on large-scale datasets to obtain general representations, followed by fine-tuning on specific tasks and labeled datasets. However, their task-specific heads often lack generalizability, significantly limiting performance in zero-shot scenarios. Inspired by the success of large language models (LLMs), we propose ZerOmics, the first zero-shot method that guides LLMs to perform various SC tasks without relying on specific downstream data. To enable LLMs to establish a correct and comprehensive understanding of SC data, ZerOmics employs a dual-alignment strategy. Specifically, ZerOmics aligns SC expression data with the well-organized gene corpus, thereby generating robust SC embeddings. These embeddings are then incorporated into instructions designed for various SC analysis tasks to tune the LLM, achieving alignment between SC data and the LLM. Extensive experiments across various sequencing technologies and tissues demonstrate that ZerOmics provides a comprehensive and general solution for SC analysis, achieving performance comparable to or even surpassing the state-of-the-art (SOTA) supervised and fine-tuned methods.", "keywords": [ "Large Language Models; Instruction Tuning; Zero-shot Learning; Bioinformatics" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 4, 2, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "J1SGf2lyr6", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 1, "author_info_list": [ { "name": "Saeid Sheikhi", "gender": "unknown", "institution": "University of Oulu", "country": "FI", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 16, "n_ref": 25, "n_ref_all": 32, "n_fig": 2, "n_tab": 4, "L_tab": 365, "n_element_tab": 45, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 482, "n_element_tab_1": 52, "formula_len_all": 600, "formula_len_all_1": 409, "len_all": 62224, "len_all_1": 41538, "len_abs": 1303, "len_title": 143, "len_sents": 19140, "len_sents_1": 16099, "n_sents": 155, "n_sents_1": 136, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1307, "title": "A Feature-Aware Federated Learning Framework for Unsupervised Anomaly Detection in 5G Networks", "abs": "The expansion of 5G networks has led to remarkable data volume and complexity, introducing significant security challenges that require the implementation of robust and scalable anomaly detection mechanisms. Traditional centralized approaches pose privacy risks and scalability challenges due to the distributed nature of 5G infrastructures. Federated Learning (FL) offers a decentralized solution but often overlooks the importance of feature relevance and privacy preservation during model aggregation. This paper introduces a novel Feature-Aware Federated framework that integrates feature importance into the aggregation process while ensuring differential privacy. We employ integrated gradients to compute feature importance for each client, aggregate them globally with differential privacy noise, and use these insights to weight model parameters during aggregation. Additionally, we propose Dynamic Feature Importance Adaptation (DFIA) to update feature importance occasionally, enhancing the model's adaptability to evolving data distributions. Experimental results demonstrate that our framework outperforms traditional federated approaches like FedAvg and FedProx in unsupervised anomaly detection tasks within 5G networks, achieving higher accuracy and robustness while preserving data privacy.", "keywords": [ "Federated Learning", "Anomaly Detection", "5G Networks", "Privacy-Preserving" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "J1J5eGJsKZ", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Cheongsu Lim", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "MS student" }, { "name": "Gyuhyeon Seo", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Jeonghoon Shim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Yohan Jo", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 16, "n_ref": 23, "n_ref_all": 57, "n_fig": 7, "n_tab": 16, "L_tab": 4609, "n_element_tab": 427, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 1494, "n_element_tab_1": 235, "formula_len_all": 530, "formula_len_all_1": 531, "len_all": 157176, "len_all_1": 64895, "len_abs": 1689, "len_title": 130, "len_sents": 58645, "len_sents_1": 29061, "n_sents": 483, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1445, "title": "ToolDial: Multi-turn Dialogue Generation Method for Tool-Augmented Language Models", "abs": "Tool-Augmented Language Models (TALMs) leverage external APIs to answer user queries across various domains. However, existing benchmark datasets for TALM research often feature simplistic dialogues that do not reflect real-world scenarios, such as the need for models to ask clarifying questions or proactively call additional APIs when essential information is missing. To address these limitations, we construct and release ToolDial, a dataset comprising 11,111 multi-turn dialogues, with an average of 8.95 turns per dialogue, based on APIs from RapidAPI. ToolDial has two key characteristics. First, the dialogues incorporate 16 user and system actions (e.g., request, clarify, fail inform) to capture the rich dynamics of real-world interactions. Second, we simulate dialogues where the system requests necessary information from the user based on API documentation and seeks additional APIs if the user fails to provide the required information. To facilitate this process, we introduce a method for generating an API graph that represents input and output compatibility between APIs. Using ToolDial, we evaluate a suite of language models on their ability to predict correct actions and extract input parameter values for API calls from the dialogue history. Modern language models achieve accuracy scores below 70\\%, indicating substantial room for improvement. We provide a detailed analysis of the areas where these models fall short.", "keywords": [ "Tool-augmented language models" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "J19jKa3wFj", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "David Baker", "gender": "unknown", "institution": "University of Washington", "country": "US", "position": "Full Professor" }, { "name": "Ge Liu", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Jacob Gershon", "gender": "Male", "institution": "University of Washington, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Jason Yim", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Karsten Kreis", "gender": "unknown", "institution": "NVIDIA", "country": "CA", "position": "Research Scientist" }, { "name": "Marouane Jaakik", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "MS student" }, { "name": "Regina Barzilay", "gender": "female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Professor" }, { "name": "Tommi Jaakkola", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 25, "n_formula_1": 14, "n_ref_uni": 49, "n_ref": 99, "n_ref_all": 139, "n_fig": 8, "n_tab": 7, "L_tab": 1373, "n_element_tab": 241, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 711, "n_element_tab_1": 171, "formula_len_all": 2372, "formula_len_all_1": 1025, "len_all": 189407, "len_all_1": 70094, "len_abs": 838, "len_title": 124, "len_sents": 49031, "len_sents_1": 28764, "n_sents": 465, "n_sents_1": 268, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 841, "title": "Hierarchical Protein Backbone Generation with Latent and Structure Diffusion", "abs": "We propose a hierarchical protein backbone generative model that separates coarse and fine-grained details. Our approach called LSD consists of two stages: sampling latents which are decoded into a contact map then sampling atomic coordinates conditioned on the contact map. LSD allows new ways to control protein generation towards desirable properties while scaling to large datasets. In particular, the AlphaFold DataBase (AFDB) is appealing due as its diverse structure topologies but suffers from poor designability. We train LSD on AFDB and show latent diffusion guidance towards AlphaFold2 Predicted Alignment Error and long range contacts can explicitly balance designability, diversity, and noveltys in the generated samples. Our results are competitive with structure diffusion models and outperforms prior latent diffusion models.", "keywords": [ "Proteins", "latent diffusion" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "J0qTpmbSbh", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Alec Koppel", "gender": "Male", "institution": "J.P. Morgan Chase", "country": "", "position": "Research Team Lead" }, { "name": "Bang An", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Furong Huang", "gender": "Female", "institution": "University of Maryland", "country": "US", "position": "Associate Professor" }, { "name": "Sicheng Zhu", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Sumitra Ganesh", "gender": "Female", "institution": "J.P. Morgan Chase", "country": "", "position": "Researcher" }, { "name": "Yuancheng Xu", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Udari Madhushani", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Postdoc" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 86, "n_ref_all": 112, "n_fig": 8, "n_tab": 4, "L_tab": 1531, "n_element_tab": 60, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 1228, "n_element_tab_1": 58, "formula_len_all": 753, "formula_len_all_1": 527, "len_all": 179732, "len_all_1": 61049, "len_abs": 1706, "len_title": 138, "len_sents": 58261, "len_sents_1": 27876, "n_sents": 455, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1637, "title": "GenARM: Reward Guided Generation with Autoregressive Reward Model for Test-Time Alignment", "abs": "Large Language Models (LLMs) exhibit impressive capabilities but require careful alignment with human preferences. Traditional training-time methods finetune LLMs using human preference datasets but incur significant training costs and require repeated training to handle diverse user preferences. Test-time alignment methods address this by using reward models (RMs) to guide frozen LLMs without retraining. However, existing test-time approaches rely on trajectory-level RMs which are designed to evaluate complete responses, making them unsuitable for autoregressive text generation that requires computing next-token rewards from partial responses. To address this, we introduce GenARM, a test-time alignment approach that leverages the Autoregressive Reward Model\u2014a novel reward parametrization designed to predict next-token rewards for efficient and effective autoregressive generation. Theoretically, we demonstrate that this parametrization can provably guide frozen LLMs toward any distribution achievable by traditional RMs within the KL-regularized reinforcement learning framework. Experimental results show that GenARM significantly outperforms prior test-time alignment baselines and matches the performance of training-time methods. Additionally, GenARM enables efficient weak-to-strong guidance, aligning larger LLMs with smaller RMs without the high costs of training larger models. Furthermore, GenARM supports multi-objective alignment, allowing real-time trade-offs between preference dimensions and catering to diverse user preferences without retraining. Our project page is available at: https://genarm.github.io.", "keywords": [ "large language model alignment; controlled decoding;" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 4, 3, 4, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "IzYczpPqKq", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Heinrich H. Nax", "gender": "unknown", "institution": "University of Zurich", "country": "", "position": "Assistant Professor" }, { "name": "Jiawei Huang", "gender": "unknown", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "", "position": "PhD student" }, { "name": "Vinzenz Thoma", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "PhD student" }, { "name": "Zebang Shen", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "", "position": "Postdoc" }, { "name": "Niao He", "gender": "unknown", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 60, "n_formula_1": 10, "n_ref_uni": 78, "n_ref": 120, "n_ref_all": 165, "n_fig": 11, "n_tab": 9, "L_tab": 2106, "n_element_tab": 136, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 231, "n_element_tab_1": 23, "formula_len_all": 7332, "formula_len_all_1": 949, "len_all": 303495, "len_all_1": 82743, "len_abs": 1203, "len_title": 106, "len_sents": 90095, "len_sents_1": 34567, "n_sents": 866, "n_sents_1": 326, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1230, "title": "Learning to Steer Markovian Agents under Model Uncertainty", "abs": "Designing incentives for an adapting population is a ubiquitous problem in a wide array of economic applications and beyond. In this work, we study how to design additional rewards to steer multi-agent systems towards desired policies \\emph{without} prior knowledge of the agents' underlying learning dynamics. Motivated by the limitation of existing works, we consider a new and general category of learning dynamics called \\emph{Markovian agents}. We introduce a model-based non-episodic Reinforcement Learning (RL) formulation for our steering problem. Importantly, we focus on learning a \\emph{history-dependent} steering strategy to handle the inherent model uncertainty about the agents' learning dynamics. We introduce a novel objective function to encode the desiderata of achieving a good steering outcome with reasonable cost. Theoretically, we identify conditions for the existence of steering strategies to guide agents to the desired policies. Complementing our theoretical contributions, we provide empirical algorithms to approximately solve our objective, which effectively tackles the challenge in learning history-dependent strategies. We demonstrate the efficacy of our algorithms through empirical evaluations.", "keywords": [ "Steering Learning Dynamics", "Reinforcement Learning", "Markov Games", "Mechanism Design" ], "rating_list": [ 8, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "IzQB2pIa3F", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Adi Haviv", "gender": "Female", "institution": "School of Computer Science, Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Amit Haim Bermano", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Associate Professor" }, { "name": "Daniel Cohen-Or", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Full Professor" }, { "name": "Gal Chechik", "gender": "unknown", "institution": "Bar Ilan University", "country": "IL", "position": "Full Professor" }, { "name": "Rinon Gal", "gender": "Male", "institution": "NVIDIA", "country": "IL", "position": "Researcher" }, { "name": "Yuval Alaluf", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 41, "n_ref": 61, "n_ref_all": 73, "n_fig": 7, "n_tab": 1, "L_tab": 480, "n_element_tab": 76, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 480, "n_element_tab_1": 76, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 112908, "len_all_1": 53987, "len_abs": 715, "len_title": 112, "len_sents": 28859, "len_sents_1": 25576, "n_sents": 219, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1055, "title": "ComfyGen: Prompt-Adaptive Workflows for Text-to-Image Generation", "abs": "The practical use of text-to-image generation has evolved from simple, monolithic models to complex workflows that combine multiple specialized components. While workflow-based approaches can lead to improved image quality, crafting effective workflows requires significant expertise, owing to the large number of available components, their complex inter-dependence, and their dependence on the generation prompt. Here, we introduce the novel task of *prompt-adaptive workflow generation*, where the goal is to automatically tailor a workflow to each user prompt. We propose two LLM-based approaches to tackle this task: a tuning-based method that learns from user-preference data, and a training-free method that uses the LLM to select existing flows. Both approaches lead to improved image quality when compared to monolithic models or generic, prompt-independent workflows. Our work shows that prompt-dependent flow prediction offers a new pathway to improving text-to-image generation quality, complementing existing research directions in the field.", "keywords": [ "Text-to-Image Generation", "Workflow prediction", "LLM Applications" ], "rating_list": [ 3, 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 3, 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 5, 4, 3, 3, 4 ] }, { "paper_id": "Iz75SDbRmm", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Jiajun Wu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Jiayuan Mao", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Joshua B. Tenenbaum", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Professor" }, { "name": "Joy Hsu", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Noah Goodman", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 51, "n_ref_all": 80, "n_fig": 8, "n_tab": 10, "L_tab": 1483, "n_element_tab": 144, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 769, "n_element_tab_1": 102, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 138890, "len_all_1": 58450, "len_abs": 1520, "len_title": 83, "len_sents": 49699, "len_sents_1": 29093, "n_sents": 342, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 35, "L_abs": 1525, "title": "What Makes a Maze Look Like a Maze?", "abs": "A unique aspect of human visual understanding is the ability to flexibly interpret abstract concepts: acquiring lifted rules explaining what they symbolize, grounding them across familiar and unfamiliar contexts, and making predictions or reasoning about them. While off-the-shelf vision-language models excel at making literal interpretations of images (e.g., recognizing object categories such as tree branches), they still struggle to make sense of such visual abstractions (e.g., how an arrangement of tree branches may form the walls of a maze). To address this challenge, we introduce Deep Schema Grounding (DSG), a framework that leverages explicit structured representations of visual abstractions for grounding and reasoning. At the core of DSG are schemas\u2014dependency graph descriptions of abstract concepts that decompose them into more primitive-level symbols. DSG uses large language models to extract schemas, then hierarchically grounds concrete to abstract components of the schema onto images with vision-language models. The grounded schema is used to augment visual abstraction understanding. We systematically evaluate DSG and different methods in reasoning on our new Visual Abstractions Benchmark, which consists of diverse, real-world images of abstract concepts and corresponding question-answer pairs labeled by humans. We show that DSG significantly improves the abstract visual reasoning performance of vision-language models, and is a step toward human-aligned understanding of visual abstractions.", "keywords": [ "visual reasoning", "abstract concepts", "schemas" ], "rating_list": [ 8, 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 3, 4 ] }, { "paper_id": "Iz230vHUy0", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Kevin Sebastian Luck", "gender": "unknown", "institution": "Vrije Universiteit Amsterdam", "country": "", "position": "Assistant Professor" }, { "name": "Sergio Hern\u00e1ndez-Guti\u00e9rrez", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "MS student" }, { "name": "Ville Kyrki", "gender": "unknown", "institution": "Aalto University", "country": "FI", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 18, "n_ref_uni": 37, "n_ref": 61, "n_ref_all": 90, "n_fig": 9, "n_tab": 5, "L_tab": 1484, "n_element_tab": 117, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1433, "n_element_tab_1": 84, "formula_len_all": 1704, "formula_len_all_1": 1393, "len_all": 123238, "len_all_1": 61431, "len_abs": 1099, "len_title": 140, "len_sents": 40328, "len_sents_1": 27213, "n_sents": 261, "n_sents_1": 159, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1103, "title": "Sample-Efficient Co-Optimization of Agent Morphology and Policy with Self-Imitation Learning", "abs": "The task of co-optimizing the body and behaviour of agents has been a longstanding problem in the fields of evolutionary robotics and embodied AI. Previous work has largely focused on the development of learning methods exploiting massive parallelization of agent evaluations with large population sizes, a paradigm which is applicable to simulated agents but cannot be transferred to the real world\ndue to the assoicated costs with the production of embodiments and robots. Furthermore, recent data-efficient approaches utilizing \n reinforcement learning can suffer from distributional shifts in transition dynamics as well as in state and action spaces when experiencing new body morphologies. \nIn this work, we propose a new co-adaptation method combining reinforcement learning and State-Aligned SelfImitation Learning to co-optimize embodiment and behavioural policies withing a handful of design iterations. We show that the integration of a self-imitation signal\nimproves the data-efficiency of the co-adaptation process as well as the behavioural recovery when adapting morphological parameters.", "keywords": [ "Deep Reinforcement Learning", "Imitation Learning", "Multi-Embodiment", "Design Optimization" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "Iyrtb9EJBp", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Hai Leong Chieu", "gender": "M", "institution": "DSO National Laboratories", "country": "", "position": "Distinguished Member of Technical Staff" }, { "name": "Maojia Song", "gender": "Not Specified", "institution": "University of Leeds", "country": "GB", "position": "Undergrad student" }, { "name": "Navonil Majumder", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Researcher" }, { "name": "Rishabh Bhardwaj", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "IN", "position": "PhD student" }, { "name": "Shang Hong Sim", "gender": "Female", "institution": "Singapore University of Technology and Design", "country": "", "position": "Undergrad student" }, { "name": "Soujanya Poria", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 1, "n_ref_uni": 37, "n_ref": 90, "n_ref_all": 149, "n_fig": 6, "n_tab": 23, "L_tab": 42473, "n_element_tab": 6056, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 4989, "n_element_tab_1": 752, "formula_len_all": 1168, "formula_len_all_1": 18, "len_all": 403530, "len_all_1": 73814, "len_abs": 1208, "len_title": 155, "len_sents": 81071, "len_sents_1": 27236, "n_sents": 643, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 107, "L_abs": 1214, "title": "Measuring and Enhancing Trustworthiness of LLMs in RAG through Grounded Attributions and Learning to Refuse", "abs": "LLMs are an integral component of retrieval-augmented generation (RAG) systems. While many studies focus on evaluating the overall quality of end-to-end RAG systems, there is a gap in understanding the appropriateness of LLMs for the RAG task. To address this, we introduce TRUST-SCORE, a holistic metric that evaluates the trustworthiness of LLMs within the RAG framework. Our results show that various prompting methods, such as in-context learning, fail to effec- tively adapt LLMs to the RAG task as measured by TRUST-SCORE. Consequently, we propose TRUST-ALIGN, a method to align LLMs for improved TRUST-SCORE performance. 26 out of 27 models aligned using TRUST-ALIGN substantially outperform competitive baselines on ASQA, QAMPARI, and ELI5. Specifically, in LLaMA-3-8b, TRUST-ALIGN outperforms FRONT on ASQA (\u219112.56), QAMPARI (\u219136.04), and ELI5 (\u219117.69). TRUST-ALIGN also significantly enhances models\u2019 ability to correctly refuse and provide quality citations. We also demonstrate the effectiveness of TRUST-ALIGN across different open-weight models, including the LLaMA series (1b to 8b), Qwen-2.5 series (0.5b to 7b), and Phi3.5 (3.8b). We release our code at https://github.com/declare-lab/trust-align.", "keywords": [ "Large Language Models", "Trustworthiness", "Hallucinations", "Retrieval Augmented Generation" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 4, 4 ], "confidence_list": [ 3, 3, 2, 4 ] }, { "paper_id": "IxmWIkcKs5", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Guojiang Zhao", "gender": "Male", "institution": "DP Technology", "country": "CN", "position": "Postdoc" }, { "name": "Haitao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Lirong Wu", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Yufei Huang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yunfan Liu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Zhifeng Gao", "gender": "Male", "institution": "DP Technology", "country": "", "position": "Researcher" }, { "name": "Stan Z. Li", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Chair Professor" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 37, "n_ref": 77, "n_ref_all": 95, "n_fig": 7, "n_tab": 5, "L_tab": 4064, "n_element_tab": 403, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3127, "n_element_tab_1": 307, "formula_len_all": 791, "formula_len_all_1": 1018, "len_all": 137183, "len_all_1": 72784, "len_abs": 2059, "len_title": 137, "len_sents": 39724, "len_sents_1": 31531, "n_sents": 282, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 101, "L_abs": 2128, "title": "A Simple yet Effective $\\Delta\\Delta G$ Predictor is An Unsupervised Antibody Optimizer and Explainer", "abs": "The proteins that exist today have been optimized over billions of years of natural evolution, during which nature creates random mutations and selects them. The discovery of functionally promising mutations is challenged by the limited evolutionary accessible regions, i.e., only a small region on the fitness landscape is beneficial. There have been numerous priors used to constrain protein evolution to regions of landscapes with high-fitness variants, among which the change in binding free energy ($\\Delta\\Delta G$) of protein complexes upon mutations is one of the most commonly used priors. However, the huge mutation space poses two challenges: (1) how to improve the efficiency of $\\Delta\\Delta G$ prediction for fast mutation screening; and (2) how to explain mutation preferences and efficiently explore accessible evolutionary regions. To address these challenges, we propose a lightweight $\\Delta\\Delta G$ predictor (Light-DDG), which adopts a structure-aware Transformer as the backbone and enhances it by knowledge distilled from existing powerful but computationally heavy $\\Delta\\Delta G$ predictors. Additionally, we augmented, annotated, and released a large-scale dataset containing millions of mutation data for pre-training Light-DDG. We find that such a simple yet effective Light-DDG can serve as a good unsupervised antibody optimizer and explainer. For the target antibody, we propose a novel Mutation Explainer to learn mutation preferences, which accounts for the marginal benefit of each mutation per residue. To further explore accessible evolutionary regions, we conduct preference-guided antibody optimization and evaluate antibody candidates quickly using Light-DDG to identify desirable mutations. Extensive experiments have demonstrated the effectiveness of Light-DDG in terms of test generalizability, noise robustness, and inference practicality, e.g., 89.7$\\times$ inference acceleration and 15.45\\% performance gains over previous state-of-the-art baselines. A case study of SARS-CoV-2 further demonstrates the crucial role of Light-DDG for mutation explanation and antibody optimization.", "keywords": [ "Mutation Effect Prediction", "Mutation Preference Explanation", "Unsupervised Protein Evolution" ], "rating_list": [ 6, 8, 5, 8 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "IwxxnY98sf", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Matteo Alleman", "gender": "unknown", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Stefano Fusi", "gender": "Male", "institution": "Columbia University", "country": "", "position": "Full Professor" } ], "n_formula": 28, "n_formula_1": 9, "n_ref_uni": 49, "n_ref": 78, "n_ref_all": 104, "n_fig": 14, "n_tab": 1, "L_tab": 128, "n_element_tab": 23, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1555, "formula_len_all_1": 539, "len_all": 161559, "len_all_1": 66519, "len_abs": 930, "len_title": 35, "len_sents": 53577, "len_sents_1": 31406, "n_sents": 451, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 933, "title": "Unsupervised Learning of Categorical Structure", "abs": "Humans are known to reason using logic and abstract categories, and yet most state of the art neural models use continuous distributed representations. These representations offer impressive gradient-based learning capabilities, but it is often difficult to know what symbolic algorithm the network might implicitly be implementing, if any. We find that there are representational geometries that naturally suggest a symbolic structure, which can be expressed in terms of binary components. We show that we can recover this structure by fitting the geometry of this binary embedding to the representational geometry of the original objects. After establishing general facts and providing some intuitions, we present two algorithms that work on low-rank or full-rank data, respectively. We assess their reliability on simulated data, and then use them to interpret neural word embeddings, in which we expect a compositional structure.", "keywords": [ "geometry", "interpretability", "abstraction", "factorization", "compositionality", "disentanglement", "hierarchy" ], "rating_list": [ 6, 5, 6, 5, 6 ], "soundness_list": [ 2, 2, 3, 3, 4 ], "presentation_list": [ 2, 2, 4, 3, 4 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 2, 2, 4, 3, 4 ] }, { "paper_id": "IwmyQUPIP0", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Defeng Sun", "gender": "unknown", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Guojun Zhang", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Xinyuan Zhao", "gender": "Female", "institution": "Beijing University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yancheng Yuan", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "kaihuang CHEN", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" } ], "n_formula": 139, "n_formula_1": 40, "n_ref_uni": 31, "n_ref": 80, "n_ref_all": 116, "n_fig": 6, "n_tab": 4, "L_tab": 1170, "n_element_tab": 91, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 156, "n_element_tab_1": 3, "formula_len_all": 15286, "formula_len_all_1": 3017, "len_all": 166807, "len_all_1": 60795, "len_abs": 1451, "len_title": 171, "len_sents": 41166, "len_sents_1": 22097, "n_sents": 451, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 122, "L_abs": 1476, "title": "On the ergodic convergence properties of the Peaceman-Rachford method and their applications in solving linear programming", "abs": "In this paper, we study the ergodic convergence properties of the Peaceman-Rachford (PR) method with semi-proximal terms for solving convex optimization problems (COPs). By reformulating the PR method as a degenerate proximal point method, for the first time we establish the global convergence of the ergodic sequence generated by the PR method with broadly chosen semi-proximal terms under the assumption that there exists a Karush\u2013Kuhn\u2013Tucker (KKT) solution to the COPs. This result represents a significant departure from previous studies on the non-ergodic convergence of the PR method, which typically requires strong convexity (or strong monotonicity in the reformulated operator) conditions that are hardly satisfied for COPs. Moreover, we establish an ergodic iteration complexity of $O(1/k)$ of the PR method with semi-proximal terms, measured by the objective error, the feasibility violation, and the KKT residual using the $\\varepsilon$-subdifferential. Based on these convergence properties, we introduce the solver EPR-LP, using the ergodic sequence of the PR method with semi-proximal terms for solving linear programming (LP) problems. EPR-LP incorporates an adaptive restart strategy and dynamic penalty parameter updates for efficiency and robustness. Extensive numerical experiments on LP benchmark datasets, executed on a high-performance GPU, show that our Julia-based solver outperforms the award-winning solver PDLP at a tolerance level of $10^{-8}$.", "keywords": [ "Peaceman-Rachford method", "ergodic convergence properties", "linear programming", "compleixty", "parallel computing" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "IwhvaDrL39", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Haofei Yu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Jinwei Yao", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Kunlun Zhu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Tao Feng", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Zhaochen Hong", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Zirui Cheng", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Jiaxuan You", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 24, "n_ref": 29, "n_ref_all": 47, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1442, "formula_len_all_1": 1486, "len_all": 106530, "len_all_1": 57342, "len_abs": 1589, "len_title": 51, "len_sents": 37625, "len_sents_1": 27026, "n_sents": 272, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1594, "title": "Research Town: Simulator of Research Community", "abs": "Large Language Models (LLMs) have demonstrated remarkable potential in scientific domains, yet a fundamental question remains unanswered: Can we simulate human research communities using LLMs? Addressing this question could deepen our understanding of the processes behind research idea generation and inspire the automatic discovery of novel scientific insights. In this work, we propose ResearchTown, a multi-agent framework for simulating research communities. Within this framework, the real-world research community is simplified and modeled as an agent-data graph (i.e. community graphs), where researchers and papers are represented as agent-type and data-type nodes, respectively. We also introduce TextGNN, a text-based inference framework that models diverse research activities (e.g., paper reading, paper writing, and review writing) as specific forms of a generalized message-passing process on the agent-data graph. To evaluate the quality of research simulation, we present ResearchBench, a benchmark that uses a node-masking prediction task for scalable and objective assessment. Our experiments reveal three key findings: (1) ResearchTown effectively simulates collaborative research activities by accurately predicting the attribute of masked nodes in the graph; (2) the simulation process in ResearchTown uncovers insights, like not every author contributes equally to the final paper, which is aligned with real-world research communities; (3) ResearchTown has the potential to foster interdisciplinary research by generating reasonable paper ideas that span across domains.", "keywords": [ "multi-agent simulation; automatic research; large language model" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "IwgmgidYPS", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Ce Zhou", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "James Zou", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Juncheng Wu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Lang Gao", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Lei Xing", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Professor, Dept of Radiation Oncology," }, { "name": "Sheng Liu", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Postdoc" }, { "name": "Xianhang Li", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Yunfei Xie", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yuyin Zhou", "gender": "unknown", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "cihang xie", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "Hongyu Zhou", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 23, "n_ref": 73, "n_ref_all": 100, "n_fig": 13, "n_tab": 8, "L_tab": 4198, "n_element_tab": 865, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 745, "n_element_tab_1": 19, "formula_len_all": 7, "formula_len_all_1": 0, "len_all": 185425, "len_all_1": 52727, "len_abs": 1812, "len_title": 140, "len_sents": 40191, "len_sents_1": 25253, "n_sents": 311, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1819, "title": "MedTrinity-25M: A Large-scale Multimodal Dataset with Multigranular Annotations for Medicine", "abs": "This paper introduces MedTrinity-25M, a comprehensive, large-scale multimodal dataset for medicine, covering over 25 million images across 10 modalities, with multigranular annotations for more than 65 diseases. These multigranular annotations encompass both global information, such as modality and organ detection, and local information like ROI analysis, lesion texture, and region-wise correlations. Unlike the existing multimodal datasets, which are limited by the availability of image-text pairs, we have developed the first automated pipeline that scales up multimodal data by generating multigranular visual and textual annotations in the form of image-ROI-description triplets without the need for any paired text descriptions. Specifically, data from over 30 different sources have been collected, preprocessed, and grounded using domain-specific expert models to identify ROIs related to abnormal regions. We then build a comprehensive knowledge base and prompt multimodal large language models to perform retrieval-augmented generation with the identified ROIs as guidance, resulting in multigranular textual descriptions. Compared to existing datasets, MedTrinity-25M provides the most enriched annotations, supporting a comprehensive range of multimodal tasks such as captioning and report generation, as well as vision-centric tasks like classification and segmentation. We propose LLaVA-Tri by pretraining LLaVA on MedTrinity-25M, achieving state-of-the-art performance on VQA-RAD, SLAKE and PathVQA, surpassing representative SOTA multimodal large language models. Furthermore, MedTrinity-25M can also be utilized to support large-scale pre-training of multimodal medical AI models, contributing to the development of future foundation models in the medical domain. We will make our dataset available.", "keywords": [ "Medical Foundation Model", "Multimodal Dataset", "Vision-Language Pretraining." ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "IwPXYk6BV9", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Jiafei Wu", "gender": "Male", "institution": "Zhejiang Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Le Sun", "gender": "Female", "institution": "Nanjing University of Information Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Puning Zhao", "gender": "Male", "institution": "Zhejiang Lab", "country": "CN", "position": "Researcher" }, { "name": "Qingming Li", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" }, { "name": "Rongfei Fan", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "Associate Professor" }, { "name": "Zhe Liu", "gender": "Male", "institution": "Zhejiang Lab", "country": "CN", "position": "Full Professor" }, { "name": "Zhikun Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 35, "n_formula_1": 31, "n_ref_uni": 45, "n_ref": 76, "n_ref_all": 118, "n_fig": 3, "n_tab": 4, "L_tab": 774, "n_element_tab": 58, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 549, "n_element_tab_1": 55, "formula_len_all": 2008, "formula_len_all_1": 1771, "len_all": 133539, "len_all_1": 63852, "len_abs": 1158, "len_title": 129, "len_sents": 35707, "len_sents_1": 25921, "n_sents": 388, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 74, "L_abs": 1162, "title": "Enhancing Learning with Label Differential Privacy by Vector Approximation", "abs": "Label differential privacy (DP) is a framework that protects the privacy of labels in training datasets, while the feature vectors are public. Existing approaches protect the privacy of labels by flipping them randomly, and then train a model to make the output approximate the privatized label. However, as the number of classes K increases, stronger randomization is needed, thus the performances of these methods become significantly worse. In this paper, we propose a vector approximation approach for learning with label local differential privacy, which is easy to implement and introduces little additional computational overhead. Instead of flipping each label into a single scalar, our method converts each label into a random vector with K components, whose expectations reflect class conditional probabilities. Intuitively, vector approximation retains more information than scalar labels. A brief theoretical analysis shows that the performance of our method only decays slightly with K. Finally, we conduct experiments on both synthesized and real datasets, which validate our theoretical analysis as well as the practical performance of our method.", "keywords": [ "label differential privacy" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "Ivk2j3uRYh", "primary_area": "causal reasoning", "n_author": 2, "author_info_list": [ { "name": "Xiaowu Dai", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "Yichen Pan", "gender": "unknown", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 33, "n_formula_1": 29, "n_ref_uni": 14, "n_ref": 27, "n_ref_all": 32, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3658, "formula_len_all_1": 3245, "len_all": 80703, "len_all_1": 58445, "len_abs": 1053, "len_title": 126, "len_sents": 28622, "len_sents_1": 25820, "n_sents": 227, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1055, "title": "Random Graph Asymptotics for Treatment Effect Estimation in Two-Sided Markets", "abs": "In two-sided markets, the accurate estimation of treatment effects is crucial yet challenging due to the inherent interference between market participants, which violates the Stable Unit Treatment Value Assumption (SUTVA). This paper introduces a novel framework that leverages random graph asymptotics to model and estimate treatment effects under network interference in two-sided markets. By incorporating a random graph model, we handle two-sided randomization by modeling customer interference within the potential outcome function as a function of graph topology and equilibrium dynamics, while capturing listing interference through the random graph structure. Our new estimation process provides asymptotically normal estimators with robust theoretical properties, suitable for large-scale market scenarios. Our theoretical findings are supported by extensive numerical simulations, demonstrating the effectiveness and practical applicability of our approach in estimating direct and indirect causal effects within these complex market structures.", "keywords": [ "Two-sided markets", "Random graph models", "Network interference" ], "rating_list": [ 3, 3, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 1, 2, 3, 3 ], "confidence_list": [ 3, 5, 4, 2 ] }, { "paper_id": "Iv4NCR9wzg", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Fei Chao", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Weizhong Huang", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Xiawu Zheng", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuxin Zhang", "gender": "unknown", "institution": "Xiamen University", "country": "CN", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 64, "n_ref": 125, "n_ref_all": 135, "n_fig": 3, "n_tab": 7, "L_tab": 2205, "n_element_tab": 432, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 4640, "n_element_tab_1": 256, "formula_len_all": 102, "formula_len_all_1": 11, "len_all": 156577, "len_all_1": 55804, "len_abs": 1731, "len_title": 113, "len_sents": 29869, "len_sents_1": 23933, "n_sents": 219, "n_sents_1": 136, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1806, "title": "Towards Efficient Automatic Self-Pruning of Large Language Models", "abs": "Despite exceptional capabilities, Large Language Models (LLMs) still face deployment challenges due to their enormous size. \nPost-training structured pruning is a promising solution that prunes LLMs without the need for retraining, reducing computational overhead, and it is hardware-deployment friendly.\nHowever, the training-free nature of post-training structured pruning leads to significant performance degradation. \nWe argue that the key to mitigating this issue lies in accurately determining the pruning rate for each layer. \nMeanwhile, we find that LLMs may have prior knowledge about their own redundancy. \nBased on this insight, we introduce $\\textbf{Self-Pruner}$ an end-to-end automatic self-pruning framework for LLMs, which efficiently search layer-wise pruning rates.\nSpecifically, $\\textbf{Self-Pruner}$ leverages LLMs to autonomously execute the entire evolutionary search process to search for pruning rate configurations. \nIn this process, LLMs are used to generate populations, select parent solutions from the current population, and perform crossover and mutation operations to produce offspring solutions. \nIn this way, LLMs automatically generate and evaluate a large number of candidate solutions, effectively converging to find the pruning rate configurations with minimal human intervention.\nExtensive experiments demonstrate $\\textbf{Self-Pruner}$'s better performance compared to existing state-of-the-art methods. \nNotably, $\\textbf{Self-Pruner}$ prunes LLaMA-2-70B to 49B level with only 0.80% drop in accuracy across seven commonsense reasoning tasks, achieving a 1.39$\\times$ speedup on NVIDIA A100 80GB GPU. Further pruning to 35B level resulted in only a 3.80% decrease in accuracy while obtaining a 1.70$\\times$ speedup. Code is available in the supplementary material.", "keywords": [ "Large Language Models; Structured Pruning; Evolutionary Algorithm" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "IuU0wcO0mo", "primary_area": "applications to neuroscience & cognitive science", "n_author": 6, "author_info_list": [ { "name": "Blake A. Richards", "gender": "Male", "institution": "Google", "country": "CA", "position": "Researcher" }, { "name": "Ian Jarratt Knight", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Krystal Xuejing Pan", "gender": "Female", "institution": "McGill University, McGill University", "country": "CA", "position": "MS student" }, { "name": "Mehdi Azabou", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Postdoc" }, { "name": "Vinam Arora", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Eva Dyer", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 26, "n_ref": 51, "n_ref_all": 84, "n_fig": 23, "n_tab": 4, "L_tab": 1481, "n_element_tab": 138, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 642, "n_element_tab_1": 66, "formula_len_all": 257, "formula_len_all_1": 257, "len_all": 143018, "len_all_1": 68487, "len_abs": 1686, "len_title": 132, "len_sents": 53656, "len_sents_1": 33978, "n_sents": 382, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 84, "L_abs": 1697, "title": "Multi-session, multi-task neural decoding from distinct cell-types and brain regions", "abs": "Recent work has shown that scale is important for improved brain decoding, with more data leading to greater decoding accuracy. However, large-scale decoding across many different datasets is challenging because neural circuits are heterogeneous---each brain region contains a unique mix of cellular sub-types, and the responses to different stimuli are diverse across regions and sub-types. It is unknown whether it is possible to pre-train and transfer brain decoding models between distinct tasks, cellular sub-types, and brain regions. To address these questions, we developed a multi-task transformer architecture and trained it on the entirety of the Allen Institute's Brain Observatory dataset. This dataset contains responses from over 100,000 neurons in 6 areas of the brains of mice, observed with two-photon calcium imaging, recorded while the mice observed different types of visual stimuli. Our results demonstrate that transfer is indeed possible -combining data from different sources is beneficial for a number of downstream decoding tasks. As well, we can transfer the model between regions and sub-types, demonstrating that there is in fact common information in diverse circuits that can be extracted by an appropriately designed model. Interestingly, we found that the model's latent representations showed clear distinctions between different brain regions and cellular sub-types, even though it was never given any information about these distinctions. Altogether, our work demonstrates that training a large-scale neural decoding model on diverse data is possible, and this provides a means of studying the differences and similarities between heterogeneous neural circuits.", "keywords": [ "neural population", "multi-task", "transformer", "tokenization", "two-photon calcium imaging", "visual stimuli", "cell types" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 2, 3, 4, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "IuEBdNsWKb", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Dragos-Florian Ristache", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "PhD student" }, { "name": "Krzysztof Onak", "gender": "unknown", "institution": "Boston University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 8, "n_ref_uni": 18, "n_ref": 35, "n_ref_all": 40, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1287, "formula_len_all_1": 384, "len_all": 124264, "len_all_1": 65481, "len_abs": 1193, "len_title": 42, "len_sents": 46739, "len_sents_1": 32004, "n_sents": 359, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1294, "title": "Adversarial Robustness of Count-Min Sketch", "abs": "Small\u2013space frequency estimators play a crucial role in a multitude of settings related to both machine learning and data processing for evolving data. Many frequency estimators use internal randomness to compress the information about the frequencies of items to a small sketch that can be used to provide estimates. Historically, these types of estimators were designed without considering the scenario in which the user with access to the estimator can accidentally or maliciously manipulate estimates. This can be achieved by the user who makes adaptive updates and uses queries to gain information about the estimator's internal randomness.\n\nIn this work, we consider one of the simplest such estimators: Count-Min Sketch. On the one hand, we show how to make it resistant to adversarial attacks in both the random oracle model, which corresponds to cryptographically hard hash functions, and using universal hash functions if the domain size is in polynomial relationship with with the size of hash tables.\n\nOn the other hand, we also explore adaptive attacks on Count-Min Sketch. In particular, we show how to speed up multirow hashing attacks for a popular family of universal hash functions and demonstrate the efficiency of our attack for a popular implementation of Count-Min Sketch.", "keywords": [ "randomized methods", "frequency estimation", "adversarial attacks", "universal hashing" ], "rating_list": [ 3, 3, 8, 3 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "Iu8RytBaji", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Edoardo Ponti", "gender": "unknown", "institution": "University of Edinburgh", "country": "GB", "position": "Assistant Professor" }, { "name": "Ivan Titov", "gender": "unknown", "institution": "University of Edinburgh", "country": "", "position": "Full Professor" }, { "name": "Zeyu Huang", "gender": "unknown", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Zihan Qiu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zili Wang", "gender": "Male", "institution": "stepfun", "country": "CN", "position": "Researcher" } ], "n_formula": 25, "n_formula_1": 14, "n_ref_uni": 45, "n_ref": 75, "n_ref_all": 116, "n_fig": 16, "n_tab": 8, "L_tab": 4480, "n_element_tab": 523, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 11, "n_element_tab_1": 1, "formula_len_all": 1592, "formula_len_all_1": 1142, "len_all": 238129, "len_all_1": 65540, "len_abs": 1658, "len_title": 104, "len_sents": 65203, "len_sents_1": 34024, "n_sents": 422, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1663, "title": "Post-hoc Reward Calibration: A Case Study on Length Bias", "abs": "Reinforcement Learning from Human Feedback aligns the outputs of Large Language Models with human values and preferences. Central to this process is the reward model (RM), which translates human feedback into training signals for optimising LLM behaviour. However, RMs can develop biases by exploiting spurious correlations in their training data, such as favouring outputs based on length or\nstyle rather than true quality. These biases can lead to incorrect output rankings, sub-optimal model evaluations, and the amplification of undesirable behaviours in LLMs alignment. This paper addresses the challenge of correcting such biases without additional data and training, introducing the concept of Post-hoc Reward Calibration. We first propose to use local average reward to estimate the bias term\nand, thus, remove it to approximate the underlying true reward. We then extend the approach to a more general and robust form with the Locally Weighted Regression. Focusing on the prevalent length bias, we validate our proposed approaches across three experimental settings, demonstrating consistent improvements: (1) a 3.11 average performance gain across 33 reward models on the RewardBench\ndataset; (2) improved agreement of RM produced rankings with GPT-4 evaluations and human preferences based on the AlpacaEval benchmark; and (3) improved Length-Controlled win rate (Dubois et al., 2024) of the RLHF process in multiple LLM\u2013RM combinations. According to our experiments, our method is computationally efficient and generalisable to other types of bias and RMs, offering a scalable and robust solution for mitigating biases in LLM alignment and evaluation.", "keywords": [ "RLHF; Reward Model; Length Bias" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "Iu8FVcUmVp", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Chenhongyi Yang", "gender": "Male", "institution": "Facebook", "country": "CH", "position": "Researcher" }, { "name": "Feng Zhao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jiaming Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Qiuchen Wang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Zehui Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zhenyu Li", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 35, "n_ref": 53, "n_ref_all": 80, "n_fig": 11, "n_tab": 7, "L_tab": 3109, "n_element_tab": 410, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 2387, "n_element_tab_1": 302, "formula_len_all": 426, "formula_len_all_1": 427, "len_all": 127561, "len_all_1": 64863, "len_abs": 1911, "len_title": 124, "len_sents": 36128, "len_sents_1": 27713, "n_sents": 261, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1873, "title": "PseDet: Revisiting the Power of Pseudo Label in Incremental Object Detection", "abs": "Incremental Objection Detection (IOD) facilitates the expansion of the usage scope of object detectors without forgetting previously acquired knowledge. Current approaches mostly adopt response-level knowledge distillation to overcome forgetting issues, by conducting implicit memory replay from the teacher model on new training data. However, this indirect learning paradigm does not fully leverage the knowledge generated by the teacher model. In this paper, we dive deeper into the mechanism of pseudo-labeling in incremental object detection by investigating three critical problems: (a) the upper bound quality of the pseudo labels is greatly limited by the previous model, (b) fixed score thresholds for label filtering, without considering the distribution across categories, and (c) the confidence score generated by the model does not well reflect the quality of the localization. Based on these observations, we propose a simple yet effective pseudo-labeling continual object detection framework, namely PseDet. Specifically, we introduce the spatio-temporal enhancement module to alleviate the negative effects when learning noisy data from the previous model. Considering the score distribution divergence across different classes, we propose the Categorical Adaptive Label Selector with a simple mathematical prior and fast K-Means pre-computation to dynamically determine the class-wise filtering threshold. In order to align the label score with the localization quality of the pseudo labels, we project the score through non-linear mapping to calibrate the distribution and integrate it into the new-step supervision. Extensive experiments on the competitive COCO benchmarks demonstrate the effectiveness and generalization of PseDet. Notably, it achieves 43.5+/41.2+ mAP under the 1/4-step incremental settings, achieving new state-of-the-art performance.", "keywords": [ "Incremental Object Detection", "Catastrophic Forgetting", "Pseudo Labeling" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "It4KL6XnPq", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Jonathan Cullen", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Scott Jeen", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 108, "n_ref": 191, "n_ref_all": 218, "n_fig": 9, "n_tab": 4, "L_tab": 1007, "n_element_tab": 83, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 699, "formula_len_all_1": 545, "len_all": 187448, "len_all_1": 70205, "len_abs": 1138, "len_title": 79, "len_sents": 37062, "len_sents_1": 24784, "n_sents": 324, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 31, "L_abs": 1102, "title": "Foundation Policies with Memory", "abs": "A generalist agent should perform well on novel tasks in unfamiliar environments. While Foundation Policies (FPs) enable generalization across new tasks, they lack mechanisms for handling novel dynamics. Conversely, agents equipped with memory models can adapt to new dynamics, but struggle with unseen tasks. In this work, we bridge this gap by integrating memory models into the FP architecture, allowing policies to condition on both task and environment dynamics. We evaluate FPs enhanced with attention, state-space, and RNN-based memory models on POPGym, a memory benchmark, and ExORL, an unsupervised RL benchmark. Our results show that GRUs achieve the best generalization to unseen tasks and dynamics for a given recurrent state size, approaching the performance of a supervised baseline that has access to task information during training and significantly outperforming memory-free FPs. Additionally, our approach improves FP performance on entirely new environments not encountered during training. Our anonymized code is available at \\url{https://anonymous.4open.science/r/zero-shot-96A1}.", "keywords": [ "reinforcement learning", "unsupervised RL", "foundation policies", "zero-shot reinforcement learning", "generalisation in reinforcement learning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 4, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "IssPhpUsKt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Bertram H\u00f8jer", "gender": "Male", "institution": "IT University of Copenhagen", "country": "DK", "position": "PhD student" }, { "name": "Oliver Simon Jarvis", "gender": "Male", "institution": "IT University of Copenhagen", "country": "DK", "position": "Researcher" }, { "name": "Stefan Heinrich", "gender": "unknown", "institution": "IT University of Copenhagen", "country": "", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 18, "n_ref": 39, "n_ref_all": 51, "n_fig": 10, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 398, "formula_len_all_1": 398, "len_all": 105898, "len_all_1": 52823, "len_abs": 2045, "len_title": 136, "len_sents": 36958, "len_sents_1": 25595, "n_sents": 300, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1614, "title": "Improving Reasoning Performance in Large Language Models via Representation Engineering", "abs": "Recent advancements in large language models (LLMs) have resulted in increasingly anthropomorphic language concerning the ability of LLMs to reason. Whether reasoning in LLMs should be understood to be inherently special is, however, widely debated. We propose utilizing a representation engineering approach wherein model activations are read from the residual stream of an LLM when processing a reasoning task. The activations are used to derive a control vector that is applied to the model as an inference-time intervention, modulating the representational space of the model, to improve performance on the specified task. We additionally open-source the code for deriving control vectors and analyzing model representations. The method allows us to improve performance on reasoning benchmarks and assess how control vectors influence the final logit distribution of a model via metrics such as KL divergence and entropy. We apply control vectors to Mistral-7B-Instruct and a range of Pythia models on an inductive, a deductive and mathematical reasoning task. We show that an LLM can, to a certain degree, be controlled to improve its perceived reasoning ability by modulating activations. The intervention is dependent upon the ability to reliably extract the model's typical state when correctly solving a task. Our results suggest that reasoning performance can be modulated in the same manner as other information-processing tasks performed by LLMs and demonstrate that we are capable of improving performance on specific tasks via a simple intervention on the residual stream with no additional training.", "keywords": [ "Large Language Models", "Reasoning", "Intelligence", "Representation Learning" ], "rating_list": [ 6, 8, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 2, 4 ] }, { "paper_id": "IsHWcsk4Fz", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Hai Lin", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Intern" }, { "name": "Jun Wang", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Professor" }, { "name": "Ning Yang", "gender": "Female", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Pin Lyu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuan Xin", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 53, "n_formula_1": 51, "n_ref_uni": 43, "n_ref": 56, "n_ref_all": 66, "n_fig": 8, "n_tab": 3, "L_tab": 1035, "n_element_tab": 11, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 758, "n_element_tab_1": 8, "formula_len_all": 2588, "formula_len_all_1": 2621, "len_all": 113926, "len_all_1": 60440, "len_abs": 1259, "len_title": 109, "len_sents": 29309, "len_sents_1": 24964, "n_sents": 232, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1281, "title": "FedADM: Adaptive Federated Learning via Dissimilarity Measure", "abs": "In federated learning, there are two critical challenges: 1) the data on distributed learners is heterogeneous; and 2) communication resources within the network are limited. In this work, we propose a framework, Federated Adaptive Dissimilarity Measure (FedADM), which can be regarded as an adaptively enhanced version of the Federated Proximal (FedProx) algorithm. This adaptiveness is primarily manifested in two aspects: (i) how it adaptively adjusts the proximity between the local models on different learners and the global model; and (ii) how it adaptively aggregates local model parameters. Building on the FedProx model, FedADM incorporates the concept of the Lagrangian multiplier to control the proximal coefficients of different learners, using \u201c\\textit{parameter dissimilarity}\" to address data heterogeneity. It explicitly captures the essence of using \u201c\\textit{loss dissimilarity}\" to adaptively adjust the aggregation frequency on distributed learners, thereby reducing communication overhead. Theoretically, we provide the performance upper bounds and convergence analysis of our proposed FedADM. Experiment results demonstrate that FedADM allows for higher accuracy and lower communication overhead compared to the baselines across a suite of realistic datasets.", "keywords": [ "Distributed Optimization", "Federated Learning", "Local Updates", "Aggregation" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 1, 3, 1 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "Ir6JxcuP6H", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Anurag Bagchi", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Martial Hebert", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Professor" }, { "name": "Pavel Tokmakov", "gender": "Male", "institution": "Toyota Research Institute", "country": "US", "position": "Research Scientist" }, { "name": "Yu-Xiong Wang", "gender": "unknown", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zhipeng Bao", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 59, "n_ref": 122, "n_ref_all": 149, "n_fig": 6, "n_tab": 5, "L_tab": 1113, "n_element_tab": 137, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 777, "n_element_tab_1": 90, "formula_len_all": 331, "formula_len_all_1": 218, "len_all": 148699, "len_all_1": 62601, "len_abs": 1252, "len_title": 121, "len_sents": 41685, "len_sents_1": 30587, "n_sents": 282, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1166, "title": "ReferEverything: Towards segmenting everything we can speak of in videos", "abs": "We present REM, a framework for segmenting a wide variety of concepts in video that can be described through natural language. To achieve this level of generalization, our method capitalizes on visual-language representations learned by video diffusion models on Internet-scale datasets. A key insight of our approach is preserving as much of the generative model\u2019s original representation as possible, while fine-tuning it on narrow-domain Referral Object Segmentation datasets. As a result, despite being exclusively trained on object masks from a limited set of categories, our framework is able to accurately segment and track both rare, unseen objects and non-object, dynamic concepts, such as waves crushing in the ocean. To better quantify the generalization capabilities of our model, we introduce a new benchmark for Referral Video Process Segmentation (RVPS), which captures dynamic phenomena that exist at the intersection of video and language. Our experiments show that REM performs comparably to state-of-the-art approaches on in-domain datasets while outperforming them by up to 28\\% out-of-domain, leveraging the power of Internet-scale pre-training.", "keywords": [ "Referring Video Segmentation", "Video Diffusion Models" ], "rating_list": [ 5, 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 3, 4, 3 ] }, { "paper_id": "IqaQZ1Jdky", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Di Wen", "gender": "Female", "institution": "Sun Yat-Sen University", "country": "CN", "position": "PhD student" }, { "name": "Shuhui Wang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xiao Han", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Yiling Wu", "gender": "Female", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Postdoc" }, { "name": "Yue WU", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zhaocheng He", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Zhe Wu", "gender": "Male", "institution": "Pengcheng Laboratory", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 18, "n_ref_uni": 71, "n_ref": 100, "n_ref_all": 112, "n_fig": 3, "n_tab": 8, "L_tab": 2681, "n_element_tab": 124, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1877, "n_element_tab_1": 82, "formula_len_all": 1771, "formula_len_all_1": 1239, "len_all": 154777, "len_all_1": 61195, "len_abs": 1330, "len_title": 103, "len_sents": 32993, "len_sents_1": 27534, "n_sents": 248, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1498, "title": "Kolmogorov-Arnold Networks with Variable Function Basis", "abs": "\\begin{abstract}\nNeural networks exhibit exceptional performance in processing complex data, yet their internal structures remain largely unexplored. The emergence of Kolmogorov-Arnold Networks (KANs) represents a significant departure from traditional Multi-Layer Perceptrons (MLPs). In contrast to MLPs, KANs replace fixed activation functions at nodes (``neurons\") with learnable activation functions on edges (``weights\"), enhancing both accuracy and interpretability.\nAs data evolves, the demand for models that are both flexible and robust minimizing the influence of input data variability continues to grow. Addressing this need, we propose a general framework for KANs utilizing a \\underline{\\textbf{V}}ariety \\underline{\\textbf{B}}ernstei\\underline{\\textbf{n}} Polynomial Function Basis for \\underline{\\textbf{K}}olmogorov-\\underline{\\textbf{A}}rnold \\underline{\\textbf{N}}etworks (VBn-KAN). This framework leverages the Weierstrass approximation theorem to extend function basis within KANs in theory, specifically selecting Bernstein polynomials ($B_n$) for their robustness, assured by the uniform convergence proposition. Additionally, to enhance flexibility, we implement techniques to vary the function basis $B_n$ when handling diverse datasets. Comprehensive experiments across three fields: multivariate time series forecasting, computer vision, and function approximation\u2014demonstrate that our method outperforms conventional approaches and other variants of KANs.\n\\end{abstract}", "keywords": [ "Kolmogorov Arnold Networks; variety function basis; interpretability; Weierstrass Approximation Theorem; Bernstein polynomial; multivarious time series forecasting; image for classification; learn the correct univariate functions." ], "rating_list": [ 3, 3, 3, 1 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "IqHeDe2lbl", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Ammar I Marvi", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Meenakshi Khosla", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Nancy Kanwisher", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 31, "n_ref": 42, "n_ref_all": 59, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 741, "formula_len_all_1": 541, "len_all": 124551, "len_all_1": 57499, "len_abs": 1391, "len_title": 131, "len_sents": 41761, "len_sents_1": 29671, "n_sents": 264, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 82, "L_abs": 1384, "title": "Sparse components distinguish visual pathways & their alignment to neural networks", "abs": "The ventral, dorsal, and lateral streams in high-level human visual cortex are implicated in distinct functional processes. Yet, deep neural networks (DNNs) trained on a single task model the entire visual system surprisingly well, hinting at common computational principles across these pathways. To explore this inconsistency, we applied a novel sparse decomposition approach to identify the dominant components of visual representations within each stream. Consistent with traditional neuroscience research, we find a clear difference in component response profiles across the three visual streams\u2014identifying components selective for faces, places, bodies, text, and food in the ventral stream; social interactions, implied motion, and hand actions in the lateral stream; and some less interpretable components in the dorsal stream. Building on this, we introduce Sparse Component Alignment (SCA), a new method for measuring representational alignment between brains and machines that better captures the latent neural tuning of these two visual systems. We find that standard visual DNNs are more aligned with ventral than either dorsal or lateral representations. SCA reveals these distinctions with greater resolution than conventional population-level geometry, offering a measure of representational alignment that is sensitive to a system\u2019s underlying axes of neural tuning.", "keywords": [ "visual representations", "alignment", "sparse decomposition", "neural pathways", "brain and machine vision" ], "rating_list": [ 8, 8, 5, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 2, 4, 2, 4 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "IqGVIU4rvM", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "NiDong Wang", "gender": "Male", "institution": "Mobvoi.inc", "country": "CN", "position": "Principal Researcher" }, { "name": "Yong Luo", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Professor" }, { "name": "Yongqian Li", "gender": "Male", "institution": "Wuhan University", "country": "", "position": "MS student" }, { "name": "Yunlin Chen", "gender": "Male", "institution": "mobvoi", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhennan Meng", "gender": "Male", "institution": "mobvoi", "country": "CN", "position": "Researcher" }, { "name": "Zhifei Li", "gender": "unknown", "institution": "mobvoi", "country": "CN", "position": "Principal Researcher" }, { "name": "Bo Du", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 21, "n_ref": 25, "n_ref_all": 38, "n_fig": 4, "n_tab": 3, "L_tab": 606, "n_element_tab": 81, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 606, "n_element_tab_1": 81, "formula_len_all": 457, "formula_len_all_1": 457, "len_all": 73623, "len_all_1": 45408, "len_abs": 1017, "len_title": 169, "len_sents": 22419, "len_sents_1": 19918, "n_sents": 160, "n_sents_1": 146, "venue": "ICLR2025", "decision": "Reject", "L_title": 120, "L_abs": 1019, "title": "Balancing Token Efficiency and Structural Accuracy in LLMs Image Generation by Combining VQ-VAE and Diffusion Tokenizers", "abs": "We proposes a novel visual tokenizer by combining high-level semantic tokens and low-level pixel tokens to represent images, aiming to address the challenges of image-to-sequence conversion for Large Language Models (LLMs). Existing visual tokenizers, such as VQ-VAE and diffusion-based models, either struggle with token explosion as image resolution increases or fail to capture detailed structural information. Our method introduces a dual-token system: high-level semantic tokens capture the main content of the image, while low-level pixel tokens preserve structural details. By integrating these tokens in a hybrid architecture, we leverage a VQ-VAE branch to generate low-resolution guidance and a diffusion process to reconstruct high-resolution images with both semantic coherence and structural accuracy. This approach significantly reduces the number of required tokens and enhances image reconstruction quality, offering an efficient solution for tasks like image generation and understanding based on LLMs.", "keywords": [ "Visual Tokenizer", "VQ-VAE", "Diffusion" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Iq7wD4BG30", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Alberto Maria Metelli", "gender": "Male", "institution": "Politecnico di Milano", "country": "", "position": "Assistant Professor" }, { "name": "Filippo Lazzati", "gender": "Male", "institution": "Politecnico di Milano", "country": "IT", "position": "PhD student" } ], "n_formula": 169, "n_formula_1": 11, "n_ref_uni": 50, "n_ref": 119, "n_ref_all": 184, "n_fig": 19, "n_tab": 12, "L_tab": 7224, "n_element_tab": 493, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 513, "n_element_tab_1": 30, "formula_len_all": 19178, "formula_len_all_1": 994, "len_all": 433852, "len_all_1": 93287, "len_abs": 1111, "len_title": 115, "len_sents": 130810, "len_sents_1": 37248, "n_sents": 1103, "n_sents_1": 296, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1118, "title": "Learning Utilities from Demonstrations in Markov Decision Processes", "abs": "Our goal is to extract useful knowledge from demonstrations of behavior in sequential decision-making problems. Although it is well-known that humans commonly engage in *risk-sensitive* behaviors in the presence of stochasticity, most Inverse Reinforcement Learning (IRL) models assume a *risk-neutral* agent. Beyond introducing model misspecification, these models do not directly capture the risk attitude of the observed agent, which can be crucial in many applications. In this paper, we propose a novel model of behavior in Markov Decision Processes (MDPs) that explicitly represents the agent's risk attitude through a *utility* function. We then define the Utility Learning (UL) problem as the task of inferring the observed agent's risk attitude, encoded via a utility function, from demonstrations in MDPs, and we analyze the partial identifiability of the agent's utility. Furthermore, we devise two provably efficient algorithms for UL in a finite-data regime, and we analyze their sample complexity. We conclude with proof-of-concept experiments that empirically validate both our model and our algorithms.", "keywords": [ "Inverse Reinforcement Learning", "Risk", "Imitation Learning", "Theory" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "Ipe4fMCBXk", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Chunhua Shen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Hao Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ke Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 25, "n_ref_uni": 28, "n_ref": 86, "n_ref_all": 97, "n_fig": 6, "n_tab": 3, "L_tab": 474, "n_element_tab": 45, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 153, "n_element_tab_1": 36, "formula_len_all": 2078, "formula_len_all_1": 1403, "len_all": 130232, "len_all_1": 61252, "len_abs": 1140, "len_title": 103, "len_sents": 38763, "len_sents_1": 25612, "n_sents": 382, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1112, "title": "Recombination Flow Matching Model for Protein Evolution", "abs": "The design of novel proteins, distinct from those found in nature, holds immense potential for advancing drug discovery, biotechnology, and material science. However, current methodologies often face significant limitations in generating both novel protein structures. Biological evolution, a natural process that fosters novelty, heavily relies on recombination\u2014yet this mechanism remains largely untapped in protein design. In this work, we propose Recombination Flow Matching (RFM), a novel generative model inspired by the principles of evolution. RFM meticulously preserves the structural integrity of protein segments during recombination while autonomously optimizing their spatial arrangement within the resultant protein. Using a common benchmark dataset, we demonstrate that RFM significantly outperforms established methods in producing structurally novel proteins. This approach opens new frontiers in protein design, leveraging evolutionary recombination to enhance the novelty of protein design. To the best of our knowledge, RFM is the first model to incorporate recombination into protein design.", "keywords": [ "Flow matching", "Generative model", "Protein design" ], "rating_list": [ 3, 3, 5, 6, 8 ], "soundness_list": [ 3, 1, 3, 3, 3 ], "presentation_list": [ 1, 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 3, 3 ] }, { "paper_id": "Ip6UwB35uT", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Changliang Zou", "gender": "Male", "institution": "Nankai University", "country": "", "position": "Full Professor" }, { "name": "Lin Lu", "gender": "Female", "institution": "Nankai University", "country": "", "position": "PhD student" }, { "name": "Xiaoyang Wu", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "PhD student" }, { "name": "Zhaojun Wang", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Full Professor" } ], "n_formula": 114, "n_formula_1": 36, "n_ref_uni": 40, "n_ref": 80, "n_ref_all": 94, "n_fig": 6, "n_tab": 4, "L_tab": 817, "n_element_tab": 34, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 436, "n_element_tab_1": 7, "formula_len_all": 16461, "formula_len_all_1": 3100, "len_all": 193128, "len_all_1": 62579, "len_abs": 949, "len_title": 106, "len_sents": 60998, "len_sents_1": 26836, "n_sents": 624, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 959, "title": "Conditional Testing based on Localized Conformal $p$-values", "abs": "In this paper, we address conditional testing problems through the conformal inference framework. We define the localized conformal $p$-values by inverting prediction intervals and prove their theoretical properties. These defined $p$-values are then applied to several conditional testing problems to illustrate their practicality. Firstly, we propose a conditional outlier detection procedure to test for outliers in the conditional distribution with finite-sample false discovery rate (FDR) control. We also introduce a novel conditional label screening problem with the goal of screening multivariate response variables and propose a screening procedure to control the family-wise error rate (FWER). Finally, we consider the two-sample conditional distribution test and define a weighted U-statistic through the aggregation of localized $p$-values. Numerical simulations and real-data examples validate the superior performance of our proposed strategies.", "keywords": [ "Conditional testing; Conformal inference; False discovery rate; Family-wise error rate; U-statistic." ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "IowRyVs862", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Bowei Zhu", "gender": "unknown", "institution": "Renmin University of China", "country": "", "position": "PhD student" }, { "name": "Shaojie Li", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 136, "n_formula_1": 33, "n_ref_uni": 47, "n_ref": 117, "n_ref_all": 146, "n_fig": 0, "n_tab": 5, "L_tab": 1597, "n_element_tab": 264, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 862, "n_element_tab_1": 10, "formula_len_all": 13110, "formula_len_all_1": 2272, "len_all": 201595, "len_all_1": 70019, "len_abs": 686, "len_title": 113, "len_sents": 53910, "len_sents_1": 29249, "n_sents": 601, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 708, "title": "Stability and Sharper Risk Bounds with Convergence Rate $O(1/n^2)$", "abs": "The sharpest known high probability excess risk bounds are up to $O\\left( 1/n \\right)$ for empirical risk minimization and projected gradient descent via algorithmic stability (Klochkov \\& Zhivotovskiy, 2021). In this paper, we show that high probability excess risk bounds of order up to $O(1/n^2)$ are possible. We discuss how high probability excess risk bounds reach $O(1/n^2)$ under strong convexity, smoothness and Lipschitz continuity assumptions for empirical risk minimization, projected gradient descent and stochastic gradient descent. Besides, to the best of our knowledge, our high probability results on the generalization gap measured by gradients for nonconvex problems are also the sharpest.", "keywords": [ "algorithmic stability", "generalization bounds", "excess risk bounds", "stochastic gradiet descent" ], "rating_list": [ 5, 5, 8, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "IoonroIpfD", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Dong Yang", "gender": "unknown", "institution": "Georgia State University", "country": "US", "position": "PhD student" }, { "name": "Rufei Gao", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Zhaowei Liu", "gender": "Male", "institution": "Yantai University", "country": "CN", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 39, "n_ref": 47, "n_ref_all": 56, "n_fig": 2, "n_tab": 4, "L_tab": 1599, "n_element_tab": 140, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1667, "n_element_tab_1": 139, "formula_len_all": 904, "formula_len_all_1": 508, "len_all": 115747, "len_all_1": 51829, "len_abs": 1658, "len_title": 135, "len_sents": 34320, "len_sents_1": 24474, "n_sents": 225, "n_sents_1": 150, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1680, "title": "A Federated Graph Learning Framework With Attention Mechanism and Clustering Algorithm", "abs": "With the development of the industrial Internet of Things, graph data is also increasing, but these data are held by different clients, and due to client privacy and data security, it is impossible to integrate all the data for unified model training. Federated graph learning can overcome this difficulty very well. It allows clients to participate in the training of the overall model of other clients without revealing their own private data during training, thus protecting the security of clients' private data. However, how to improve the utilization efficiency of client upload parameters to improve the effect of model training and how to process the large amount of initial data owned by clients is an issue that needs to be solved urgently. This paper proposes a federated graph learning framework with attention mechanism and clustering algorithm (${FGL}_{AC}$). First, before the client participates in training, a clustering algorithm is used to perform a simple preprocessing operation on the large amount of data held to reduce the overall model training burden and improve training accuracy. Then during the server's process of aggregating model parameters, through the adaptive ability of the attention mechanism, the parameters uploaded by different clients are configured with different weights to obtain the best weight parameters to improve the training effect of the overall model. In order to further verify the effectiveness of ${FGL}_{AC}$, experimental verification was conducted on different data sets. The results show that in most cases, ${FGL}_{AC}$ can achieve an improvement of 2.63\\% - 4.03\\% compared to other federated graph learning frameworks.", "keywords": [ "Industrial Internet of Things", "Federated Graph Learning", "Graph Neural Networks", "Attention Mechanism", "Clustering Algorithm" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 1, 1, 1, 1 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "Io9yFt7XH7", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Bao-liang Lu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Dongsheng Li", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Weibang Jiang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Yansen Wang", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 25, "n_ref": 35, "n_ref_all": 46, "n_fig": 8, "n_tab": 11, "L_tab": 5656, "n_element_tab": 292, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3121, "n_element_tab_1": 132, "formula_len_all": 1144, "formula_len_all_1": 614, "len_all": 144823, "len_all_1": 57868, "len_abs": 1714, "len_title": 151, "len_sents": 39895, "len_sents_1": 25050, "n_sents": 307, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 1675, "title": "NeuroLM: A Universal Multi-task Foundation Model for Bridging the Gap between Language and EEG Signals", "abs": "Recent advancements for large-scale pre-training with neural signals such as electroencephalogram (EEG) have shown promising results, significantly boosting the development of brain-computer interfaces (BCIs) and healthcare. However, these pre-trained models often require full fine-tuning on each downstream task to achieve substantial improvements, limiting their versatility and usability, and leading to considerable resource wastage. To tackle these challenges, we propose NeuroLM, the first multi-task foundation model that leverages the capabilities of Large Language Models (LLMs) by regarding EEG signals as a foreign language, endowing the model with multi-task learning and inference capabilities. Our approach begins with learning a text-aligned neural tokenizer through vector-quantized temporal-frequency prediction, which encodes EEG signals into discrete neural tokens. These EEG tokens, generated by the frozen vector-quantized (VQ) encoder, are then fed into an LLM that learns causal EEG information via multi-channel autoregression. Consequently, NeuroLM can understand both EEG and language modalities. Finally, multi-task instruction tuning adapts NeuroLM to various downstream tasks. We are the first to demonstrate that, by specific incorporation with LLMs, NeuroLM unifies diverse EEG tasks within a single model through instruction tuning. The largest variant NeuroLM-XL has record-breaking 1.7B parameters for EEG signal processing, and is pre-trained on a large-scale corpus comprising approximately 25,000-hour EEG data. When evaluated on six diverse downstream datasets, NeuroLM showcases the huge potential of this multi-task learning paradigm.", "keywords": [ "EEG", "large language model", "multi-task learning" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "InWaCoIMMN", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Adam Davies", "gender": "Non-Binary", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "ChengXiang Zhai", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Full Professor" }, { "name": "Jize Jiang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 4, "n_ref_uni": 62, "n_ref": 105, "n_ref_all": 124, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 30, "n_element_tab_1": 1, "formula_len_all": 403, "formula_len_all_1": 169, "len_all": 185340, "len_all_1": 62364, "len_abs": 2958, "len_title": 92, "len_sents": 50364, "len_sents_1": 33121, "n_sents": 221, "n_sents_1": 146, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 1261, "title": "Competence-Based Analysis of Language Models", "abs": "Despite the recent successes of large language models (LLMs), little is known regarding the representations of linguistic structure they learn during pretraining, which can lead to unexpected behaviors in response to prompt variation or distribution shift. To better understand these models and behaviors, we introduce a general model analysis framework to study LLMs with respect to their representation and use of human-interpretable linguistic properties. Our framework, CALM (Competence-based Analysis of Language Models), is designed to investigate LLM competence in the context of specific tasks by intervening on models\u2019 internal representations of different linguistic properties using causal probing, and measuring models\u2019 alignment under these interventions with a given ground-truth causal model of the task. We also develop a new approach for performing causal probing interventions using gradient-based adversarial attacks, which can target a broader range of properties and representations than prior techniques. Finally, we carry out a case study of CALM using these interventions to analyze and compare LLM competence across a variety of lexical inference tasks, showing that CALM can be used to explain and predict behaviors across these tasks.", "keywords": [ "interpretability", "probing", "causality", "interventions", "counterfactuals", "mechanistic interpretability", "language models", "linguistic theory", "lexical relations" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "InUpEfpXQS", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Gangshan Wu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Haoyi Zhu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Jiange Yang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Tong He", "gender": "Male", "institution": "Shanghai AI lab", "country": "CN", "position": "Researcher" }, { "name": "Yating Wang", "gender": "Female", "institution": "Tongji University", "country": "CN", "position": "PhD student" }, { "name": "Limin Wang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 53, "n_ref": 90, "n_ref_all": 102, "n_fig": 8, "n_tab": 4, "L_tab": 1787, "n_element_tab": 244, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1708, "n_element_tab_1": 177, "formula_len_all": 535, "formula_len_all_1": 534, "len_all": 151739, "len_all_1": 57425, "len_abs": 1488, "len_title": 126, "len_sents": 36032, "len_sents_1": 26116, "n_sents": 261, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1511, "title": "Tra-MoE: Scaling Trajectory Prediction Models for Adaptive Policy Conditioning", "abs": "Scale is a primary factor that influences the performance and generalization of a robot learning system. In this paper, we aim to scale up the trajectory prediction model by using broad out-of-domain data to improve its robustness and generalization ability. Trajectory model is designed to predict any-point trajectories in the current frame given an instruction and can provide detailed control guidance for robotic policy learning. To handle the diverse out-of-domain data distribution, we propose a sparsely-gated MoE (\\textbf{Top-1} gating strategy) architecture for trajectory model, coined as \\textbf{Tra-MoE}. The sparse activation design enables good balance between parameter cooperation and specialization, effectively benefiting from large-scale out-of-domain data while maintaining constant FLOPs per token. In addition, we further introduce an adaptive policy conditioning technique by learning 2D mask representations for predicted trajectories, which is explicitly aligned with image observations to guide policy prediction more flexibly. We perform experiments on both simulation and real-world scenarios to verify the effectiveness of our Tra-MoE and adaptive policy conditioning technique. We jointly train the Tra-MoE model on all 130 tasks in the LIBERO benchmark and conduct a comprehensive empirical analysis, demonstrating that our Tra-MoE consistently exhibits superior performance compared to the dense baseline model, even when the latter is scaled to match Tra-MoE's parameter count.", "keywords": [ "Mixture-of-experts", "Trajectory-guided policy", "Policy Conditioning", "Scaling", "Robot manipulation", "Embodied AI" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "InRaT76E2S", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Alexandre Allauzen", "gender": "Male", "institution": "Ecole Sup\u00e9rieure de Physique et de Chimie Industrielles", "country": "FR", "position": "Full Professor" }, { "name": "Blaise Delattre", "gender": "Male", "institution": ", Universit\u00e9 Paris-Dauphine (Paris IX)", "country": "FR", "position": "PhD student" }, { "name": "Erwan Fagnou", "gender": "Male", "institution": "Universit\u00e9 Paris-Dauphine - PSL", "country": "FR", "position": "PhD student" }, { "name": "Paul Caillon", "gender": "unknown", "institution": ", Universit\u00e9 Paris-Dauphine (Paris IX)", "country": "", "position": "Postdoc" }, { "name": "Quentin Barth\u00e9lemy", "gender": "unknown", "institution": "Foxstream", "country": "", "position": "Researcher" } ], "n_formula": 33, "n_formula_1": 8, "n_ref_uni": 43, "n_ref": 77, "n_ref_all": 90, "n_fig": 5, "n_tab": 6, "L_tab": 2258, "n_element_tab": 263, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 80, "n_element_tab_1": 9, "formula_len_all": 1611, "formula_len_all_1": 349, "len_all": 144872, "len_all_1": 61983, "len_abs": 1084, "len_title": 108, "len_sents": 40802, "len_sents_1": 30758, "n_sents": 293, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1096, "title": "Activation Decay by Loss Smoothing to Enhance Generalization", "abs": "Generalization in deep learning is often associated with the sharpness of the minima encountered during training. We introduce a novel, deterministic, and computationally efficient method called \\emph{activation decay}, designed to flatten sharp minima and improve generalization across a wide range of tasks. Derived from Gaussian smoothing, activation decay operates by regularizing the activations of critical network layers, effectively reducing sharpness and improving robustness. Unlike stochastic techniques such as dropout or the more computationally expensive Sharpness-Aware Minimization (SAM), our approach requires no additional computational overhead, making it particularly suited for large-scale models.\nWe further demonstrate that activation decay can be seamlessly combined with other regularization techniques, offering enhanced regularization without increasing training complexity. Extensive experiments on CIFAR-10, ImageNet, and natural language processing (NLP) tasks validate our approach, showing consistent improvements in generalization and robustness to label noise.", "keywords": [ "loss smoothing", "sharpness aware minimization", "flat minima", "deep learning", "activation decay" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "In0phMs7BK", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Changyou Chen", "gender": "Male", "institution": "State University of New York, Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Dong An", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Peilin Zhao", "gender": "unknown", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Shengcai Liao", "gender": "Male", "institution": "United Arab Emirates University", "country": "US", "position": "Associate Professor" }, { "name": "Xuhui Li", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "MS student" }, { "name": "Yunlong Wang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Zhengquan Luo", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "zhiqiang xu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 66, "n_formula_1": 26, "n_ref_uni": 33, "n_ref": 65, "n_ref_all": 97, "n_fig": 8, "n_tab": 5, "L_tab": 629, "n_element_tab": 68, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1382, "n_element_tab_1": 19, "formula_len_all": 5054, "formula_len_all_1": 2016, "len_all": 136894, "len_all_1": 68674, "len_abs": 1494, "len_title": 127, "len_sents": 38769, "len_sents_1": 30188, "n_sents": 333, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1507, "title": "Heterogeneous Federated Learning: A Dual Matching Dataset Distillation Approach", "abs": "Federated Learning (FL) often struggles with error accumulation during local training, particularly on heterogeneous data, which hampers overall performance and convergence. While dataset distillation is commonly introduced to FL to enhance efficiency, our work finds that communicating distilled data instead of models can completely get rid of the error accumulation issue, albeit at the cost of exacerbating data heterogeneity across clients. To address the amplified heterogeneity due to distilled data, we propose a novel FL algorithm termed \\textit{FedDualMatch}, which performs dual matching in the way that local distribution matching captures client data distributions while global gradient matching aligns gradients on the server. This dual approach enriches feature representations and enhances convergence stability. It proves effective for FL due to a bounded difference in the testing loss between optimal models trained on the aggregation of either distilled or original data across clients. At the same time, it can converge to within a bounded constant of the optimal model loss. Experiments on controlled heterogeneous dataset MNIST/CIFAR10 and naturally heterogeneous dataset Digital-Five/Office-Home demonstrate its advantages over the state-of-the-art methods that communicate either model or distilled data, in terms of accuracy and convergence. Notably, it maintains accuracy even when data heterogeneity significantly increases, underscoring its potential for practical applications.", "keywords": [ "Federated Learning; Dataset Distillation; Heterogeneous;" ], "rating_list": [ 3, 5, 3, 5, 3 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "ImpeMDJfVL", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Bai LiChen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Haoyi Xiong", "gender": "Male", "institution": "Microsoft Corporation", "country": "CN", "position": "Principal Applied Scientist" }, { "name": "Shitong Shao", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Zeke Xie", "gender": "Male", "institution": "HKUST(GZ)", "country": "CN", "position": "Assistant Professor" }, { "name": "zikai zhou", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 16, "n_ref_uni": 37, "n_ref": 93, "n_ref_all": 136, "n_fig": 15, "n_tab": 9, "L_tab": 4199, "n_element_tab": 487, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 2298, "n_element_tab_1": 107, "formula_len_all": 2199, "formula_len_all_1": 1956, "len_all": 174527, "len_all_1": 68770, "len_abs": 1403, "len_title": 128, "len_sents": 48550, "len_sents_1": 27748, "n_sents": 389, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1408, "title": "IV-mixed Sampler: Leveraging Image Diffusion Models for Enhanced Video Synthesis", "abs": "Exploring suitable solutions to improve performance by increasing the computational cost of inference in visual diffusion models is a highly promising direction. Sufficient prior studies have demonstrated that correctly scaling up computation in the sampling process can successfully lead to improved generation quality, enhanced image editing, and compositional generalization. While there have been rapid advancements in developing inference-heavy algorithms for improved image generation, relatively little work has explored inference scaling laws in video diffusion models (VDMs). Furthermore, existing research shows only minimal performance gains that are perceptible to the naked eye. To address this, we design a novel training-free algorithm IV-Mixed Sampler that leverages the strengths of image diffusion models (IDMs) to assist VDMs surpass their current capabilities. The core of IV-Mixed Sampler is to use IDMs to significantly enhance the quality of each video frame and VDMs ensure the temporal coherence of the video during the sampling process. Our experiments have demonstrated that IV-Mixed Sampler achieves state-of-the-art performance on 4 benchmarks including UCF-101-FVD, MSR-VTT-FVD, Chronomagic-Bench-150/1649, and VBench. For example, the open-source Animatediff with IV-Mixed Sampler reduces the UMT-FVD score from 275.2 to 228.6, closing to 223.1 from the closed-source Pika-2.0.", "keywords": [ "IV-mixed Sampler", "Video Synthesis", "Inference-heavy Algorithms", "Training-free" ], "rating_list": [ 6, 5, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "Im2neAMlre", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Aida Nematzadeh", "gender": "unknown", "institution": "Google Deepmind", "country": "", "position": "Researcher" }, { "name": "Anant Nawalgaria", "gender": "Male", "institution": "Research, Google", "country": "", "position": "Researcher" }, { "name": "Christopher Knutsen", "gender": "Male", "institution": "University of Bath", "country": "", "position": "Undergrad student" }, { "name": "Chuhan Zhang", "gender": "Female", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Cyrus Rashtchian", "gender": "Male", "institution": "Google Research", "country": "US", "position": "Researcher" }, { "name": "Emanuele Bugliarello", "gender": "Male", "institution": "Google DeepMind", "country": "FR", "position": "Researcher" }, { "name": "Ira Ktena", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Isabela Albuquerque", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Ivana Kajic", "gender": "unknown", "institution": "Gooogle DeepMind", "country": "CA", "position": "Researcher" }, { "name": "Jordi Pont-Tuset", "gender": "Male", "institution": "Google", "country": "CH", "position": "Research Scientist" }, { "name": "Olivia Wiles", "gender": "unknown", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Pinelopi Papalampidi", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Su Wang", "gender": "unknown", "institution": "Balyasny Asset Management", "country": "US", "position": "Researcher" }, { "name": "Yasumasa Onoe", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 41, "n_ref": 75, "n_ref_all": 111, "n_fig": 10, "n_tab": 3, "L_tab": 1110, "n_element_tab": 140, "n_fig_1": 10, "n_tab_1": 5, "L_tab_1": 1129, "n_element_tab_1": 98, "formula_len_all": 554, "formula_len_all_1": 382, "len_all": 175879, "len_all_1": 70419, "len_abs": 1502, "len_title": 132, "len_sents": 38863, "len_sents_1": 32145, "n_sents": 319, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 85, "L_abs": 1534, "title": "Revisiting text-to-image evaluation with Gecko: on metrics, prompts, and human rating", "abs": "While text-to-image (T2I) generative models have become ubiquitous, they do not necessarily generate images that align with a given prompt. \nWhile many metrics and benchmarks have been proposed to evaluate T2I models and alignment metrics, the impact of the evaluation components (prompt sets, human annotations, evaluation task) has not been systematically measured.\nWe find that looking at only *one slice of data*, i.e. one set of capabilities or human annotations, is not enough to obtain stable conclusions that generalise to new conditions or slices when evaluating T2I models or alignment metrics. \nWe address this by introducing an evaluation suite of $>$100K annotations across four human annotation templates that comprehensively evaluates models' capabilities across a range of methods for gathering human annotations and comparing models.\nIn particular, we propose (1) a carefully curated set of prompts -- *Gecko2K*; (2) a statistically grounded method of comparing T2I models; and (3) how to systematically evaluate metrics under three *evaluation tasks* -- *model ordering, pair-wise instance scoring, point-wise instance scoring*.\nUsing this evaluation suite, we evaluate a wide range of metrics and find that a metric may do better in one setting but worse in another.\nAs a result, we introduce a new, interpretable auto-eval metric that is consistently better correlated with human ratings than such existing metrics on our evaluation suite--across different human templates and evaluation settings--and on TIFA160.", "keywords": [ "text-to-image evaluation; text-to-image alignment; human evaluation;" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 4, 4, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 4, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "Ilteh48w7m", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 8, "author_info_list": [ { "name": "Daniel Rueckert", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" }, { "name": "Georgios Kaissis", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Johannes C. Paetzold", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "Postdoc" }, { "name": "Laurin Lux", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Manuel Nickel", "gender": "Male", "institution": "Technical University Munich", "country": "", "position": "PhD student" }, { "name": "Rickmer Braren", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "Associate Professor" }, { "name": "Leonhard Feiner", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "PhD student" }, { "name": "Martin Menten", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Lecturer" } ], "n_formula": 22, "n_formula_1": 13, "n_ref_uni": 27, "n_ref": 51, "n_ref_all": 79, "n_fig": 5, "n_tab": 7, "L_tab": 6999, "n_element_tab": 1203, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 368, "n_element_tab_1": 25, "formula_len_all": 1604, "formula_len_all_1": 720, "len_all": 168432, "len_all_1": 65849, "len_abs": 3816, "len_title": 136, "len_sents": 53802, "len_sents_1": 31269, "n_sents": 410, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1010, "title": "Structured Joint Aleatoric and Epistemic Uncertainty for High Dimensional Output Spaces", "abs": "Uncertainty estimation plays a vital role in enhancing the reliability of deep learning model predictions, especially in scenarios with high-dimensional output spaces. This paper addresses the dual nature of uncertainty \u2014 aleatoric and epistemic \u2014 focusing on their joint integration in high-dimensional regression tasks. We introduce an approach to approximate joint uncertainty using a low-rank plus diagonal covariance matrix, which preserves essential output correlations while mitigating the computational complexity associated with full covariance matrices. Specifically, our method reduces memory usage and enhances sampling efficiency and log-likelihood calculations. Simultaneously, our representation matches the true posterior better than factorized joint distributions, offering a clear advancement in reliability and explainability for deep learning model predictions. Furthermore, we empirically show that our method can efficiently enhance out of distribution detection in specific applications.", "keywords": [ "structured uncertainty", "aleatoric", "epistemic", "high-dimensional data" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 5, 3, 3 ] }, { "paper_id": "IlleFmPNb6", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Jiaxuan Sun", "gender": "unknown", "institution": "ShanghaiTech University", "country": "CN", "position": "Undergrad student" }, { "name": "Longtian Qiu", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "PhD student" }, { "name": "Shan Ning", "gender": "Not Specified", "institution": "ShanghaiTech University", "country": "CN", "position": "PhD student" }, { "name": "Xuming He", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 30, "n_ref": 59, "n_ref_all": 73, "n_fig": 4, "n_tab": 2, "L_tab": 200, "n_element_tab": 52, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1263, "n_element_tab_1": 16, "formula_len_all": 573, "formula_len_all_1": 529, "len_all": 170152, "len_all_1": 52584, "len_abs": 1183, "len_title": 144, "len_sents": 30609, "len_sents_1": 25173, "n_sents": 215, "n_sents_1": 165, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1190, "title": "Training-Free Retrieval-Augmented Generation for Knowledge-Intensive Visual Question Answering", "abs": "Recent advancements in multimodal large language models (MLLMs) have achieved strong performance in vision-language tasks such as visual question answering (VQA). However, these models struggle with knowledge-intensive VQA (KI-VQA) tasks that require fine-grained domain knowledge, as seen in benchmarks such as Encyclopedic VQA and InfoSeek. To address these challenges, we propose a novel retrieval-augmented generation (RAG) framework, referred to as KIRA, designed to enhance the capability of MLLMs for KI-VQA without task-specific fine-tuning. Our target is to integrate general image-text similarity with detailed knowledge context to achieve precise entity recognition. To this end, we leverage CLIP to obtain general image-text matching, and design a verification mechanism according to detailed question-text relevance to improve recognition accuracy. We evaluate our method on KI-VQA benchmarks, demonstrating significant improvements of 47.5\\% on Encyclopedic VQA and 16.2\\% on InfoSeek, all achieved without additional training. These results highlight the potential of our training-free, plug-and-play framework for solving knowledge-intensive visual question answering tasks.", "keywords": [ "Retrieval-Augmented Generation", "Visual Question Answering", "Multi-modal Large Language Model" ], "rating_list": [ 5, 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 2, 1 ], "contribution_list": [ 2, 1, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "Il5DjZmLzp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Atul Prakash", "gender": "unknown", "institution": "University of Michigan", "country": "", "position": "Professor" }, { "name": "Ryan Feng", "gender": "unknown", "institution": "University of Michigan", "country": "US", "position": "PhD student" }, { "name": "Utkarsh Singhal", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Stella Yu", "gender": "Female", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 56, "n_ref": 108, "n_ref_all": 132, "n_fig": 11, "n_tab": 7, "L_tab": 1209, "n_element_tab": 153, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 753, "n_element_tab_1": 106, "formula_len_all": 715, "formula_len_all_1": 712, "len_all": 153559, "len_all_1": 64156, "len_abs": 1013, "len_title": 110, "len_sents": 42621, "len_sents_1": 29359, "n_sents": 349, "n_sents_1": 238, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1018, "title": "Foundation Vision Models are Unsupervised Image Canonicalizers", "abs": "One of the most significant and longstanding problems in computer vision is invariance - the ability to robustly handle changes in real-world transformations such as rotation, viewpoint, and lighting. Unfortunately, popular foundation models remain brittle under such transformations. While existing solutions towards invariance have shown promise, they all fundamentally require some model training, limiting their ability to adapt broadly to new tasks, transformations, and datasets. Our key insight is that foundation model priors can be used to reason about transformations. We thus propose Foundation Model Canonicalization (FMC), an approach that can undo nuisance transformations in images without any model training. With a single core approach, FMC can make models like CLIP and SAM invariant to different transformations without any training or fine-tuning. Our approach FMC flexibly adapts to new foundation models and tasks, making it significantly easier for newer and larger models to achieve invariance.", "keywords": [ "Invariance", "Canonicalization", "Foundation Models", "CLIP", "SD", "Augmentation", "Vision", "Robustness" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "IkZVDzdC8M", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Matthijs Douze", "gender": "unknown", "institution": "Meta", "country": "FR", "position": "researcher" }, { "name": "Pierre Fernandez", "gender": "Male", "institution": "Meta", "country": "FR", "position": "Researcher" }, { "name": "Teddy Furon", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Researcher" }, { "name": "Tom Sander", "gender": "Male", "institution": "Meta", "country": "FR", "position": "PhD student" }, { "name": "Alain Durmus", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 8, "n_ref_uni": 88, "n_ref": 167, "n_ref_all": 198, "n_fig": 11, "n_tab": 3, "L_tab": 574, "n_element_tab": 85, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1318, "n_element_tab_1": 180, "formula_len_all": 2105, "formula_len_all_1": 608, "len_all": 244192, "len_all_1": 70050, "len_abs": 1145, "len_title": 90, "len_sents": 64256, "len_sents_1": 30884, "n_sents": 573, "n_sents_1": 270, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 1162, "title": "Watermark Anything With Localized Messages", "abs": "Image watermarking methods are not tailored to handle small watermarked areas.\nThis restricts applications in real-world scenarios where parts of the image may come from different sources or have been edited.\nWe introduce a deep-learning model for localized image watermarking, dubbed the Watermark Anything Model (WAM). \nThe WAM embedder imperceptibly modifies the input image, while the extractor segments the received image into watermarked and non-watermarked areas and recovers one or several hidden messages from the areas found to be watermarked.\nThe models are jointly trained at low resolution and without perceptual constraints, then post-trained for imperceptibility and multiple watermarks.\nExperiments show that WAM is competitive with state-of-the art methods in terms of imperceptibility and robustness, especially against inpainting and splicing, even on high-resolution images. \nMoreover, it offers new capabilities: WAM can locate watermarked areas in spliced images and extract distinct 32-bit messages with less than 1 bit error from multiple small regions -- no larger than 10\\% of the image surface -- even for small $256\\times 256$ images.", "keywords": [ "Image Watermarking; Segmentation" ], "rating_list": [ 3, 8, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3, 2 ], "contribution_list": [ 2, 4, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "IjiIPQcLbV", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Hanjun Dai", "gender": "Male", "institution": "Google Research", "country": "", "position": "Researcher" }, { "name": "Peng Liu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Full Professor" }, { "name": "Songtao Liu", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Zhengkai Tu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 83, "n_ref": 141, "n_ref_all": 157, "n_fig": 3, "n_tab": 2, "L_tab": 485, "n_element_tab": 79, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 485, "n_element_tab_1": 79, "formula_len_all": 191, "formula_len_all_1": 191, "len_all": 181813, "len_all_1": 69951, "len_abs": 1192, "len_title": 99, "len_sents": 36678, "len_sents_1": 33796, "n_sents": 271, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1196, "title": "SDDBench: A Benchmark for Synthesizable Drug Design", "abs": "A significant challenge in wet lab experiments with current drug design generative models is the trade-off between pharmacological properties and synthesizability. Molecules predicted to have highly desirable properties are often difficult to synthesize, while those that are easily synthesizable tend to exhibit less favorable properties. As a result, evaluating the synthesizability of molecules in general drug design scenarios remains a significant challenge in the field of drug discovery. The commonly used synthetic accessibility (SA) score aims to evaluate the ease of synthesizing generated molecules, but it falls short of guaranteeing that synthetic routes can actually be found. Inspired by recent advances in top-down synthetic route generation and forward reaction prediction, we propose a new, data-driven metric to evaluate molecule synthesizability. This novel metric leverages the synergistic duality between retrosynthetic planners and reaction predictors, both of which are trained on extensive reaction datasets. To demonstrate the efficacy of our metric, we conduct a comprehensive evaluation of round-trip scores across a range of representative molecule generative models.", "keywords": [ "Drug Discovery" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 2, 1, 1 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "IjduZQK8gM", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Chengcheng Ma", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Ismail Elezi", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "GB", "position": "Researcher" }, { "name": "Jiankang Deng", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Researcher" }, { "name": "Prannay Kaul", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 6, "n_ref_uni": 42, "n_ref": 80, "n_ref_all": 119, "n_fig": 7, "n_tab": 7, "L_tab": 2365, "n_element_tab": 409, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 692, "n_element_tab_1": 63, "formula_len_all": 3411, "formula_len_all_1": 371, "len_all": 195292, "len_all_1": 67760, "len_abs": 620, "len_title": 175, "len_sents": 58958, "len_sents_1": 32486, "n_sents": 461, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1181, "title": "From Attention to Activation: Unraveling the Enigmas of Large Language Models", "abs": "We study two strange phenomena in auto-regressive Transformers: (1) the dominance of the \ufb01rst token in attention heads; (2) the occurrence of large outlier activations in the hidden states. We \ufb01nd that popular large language models, such as Llama attend maximally to the first token in 98% of attention heads, a behaviour we attribute to the softmax function. To mitigate this issue, we propose a reformulation of softmax to softmax-1. Furthermore, we identify adaptive optimisers, e.g. Adam, as the primary contributor to the large outlier activations and introduce OrthoAdam, a novel optimiser that utilises orthogonal matrices to transform gradients, to address this issue. Finally, not only do our methods prevent these phenomena from occurring, but additionally, they enable Transformers to sustain their performance when quantised using basic algorithms, something that standard methods are unable to do. In summary, our methods reduce the attention proportion on the first token from 65% to 3.3%, the activation kurtosis in the hidden states from 1657 to 3.1, and perplexity penalty under 4-bit weight quantisation from 3565 to 0.3. We will publish our code upon acceptance.", "keywords": [ "Transformers", "Adam", "Optimizer", "Outliers", "Attention", "Quantization" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 4, 1 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "IjbXZdugdj", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 10, "author_info_list": [ { "name": "Andreas Mayr", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Postdoc" }, { "name": "G\u00fcnter Klambauer", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Full Professor" }, { "name": "Johannes Brandstetter", "gender": "Male", "institution": "Emmi AI", "country": "AT", "position": "Principal Researcher" }, { "name": "Johannes Schimunek", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "PhD student" }, { "name": "Lisa Schneckenreiter", "gender": "Female", "institution": "Institute of Machine Learning, JKU Linz", "country": "AT", "position": "PhD student" }, { "name": "Niklas Schmidinger", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "PhD student" }, { "name": "Philipp Seidl", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "", "position": "PhD student" }, { "name": "Pieter-Jan Hoedt", "gender": "Male", "institution": "Johannes Kepler University", "country": "AT", "position": "Postdoc" }, { "name": "Sepp Hochreiter", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "Full Professor" }, { "name": "Sohvi Luukkonen", "gender": "Female", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Postdoc" } ], "n_formula": 21, "n_formula_1": 13, "n_ref_uni": 118, "n_ref": 217, "n_ref_all": 250, "n_fig": 7, "n_tab": 12, "L_tab": 6207, "n_element_tab": 560, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2139, "n_element_tab_1": 62, "formula_len_all": 2565, "formula_len_all_1": 1219, "len_all": 321448, "len_all_1": 59173, "len_abs": 1628, "len_title": 156, "len_sents": 79731, "len_sents_1": 27186, "n_sents": 564, "n_sents_1": 145, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 107, "L_abs": 1338, "title": "Bio-xLSTM: Generative modeling, representation and in-context learning of biological and chemical sequences", "abs": "Language models for biological and chemical sequences enable crucial applications such as drug discovery, protein engineering, and precision medicine. Currently, these language models are predominantly based on Transformer architectures. While Transformers have yielded impressive results, their quadratic runtime dependency on sequence length complicates their use for long genomic sequences and in-context learning on proteins and chemical sequences. Recently, the recurrent xLSTM architecture has been shown to perform favorably compared to Transformers and modern state-space models (SSMs) in the natural language domain. Similar to SSMs, xLSTMs have linear runtime dependency and allow for constant-memory decoding at inference time, which makes them prime candidates for modeling long-range dependencies in biological and chemical sequences. In this work, we tailor xLSTM towards these domains and we propose a suite of language models called Bio-xLSTM. Extensive experiments in three large domains, genomics, proteins, and chemistry, were performed to assess xLSTM\u2019s ability to model biological and chemical sequences. The results show that Bio-xLSTM is a highly proficient generative model for DNA, protein, and chemical sequences, learns rich representations, and can perform in-context learning for proteins and small molecules.", "keywords": [ "xLSTM", "large language model", "foundation model", "DNA", "protein", "small molecule", "SMILES", "in-context learning", "masked language modeling", "causal language modeling", "fill-in the middle", "equivariance", "reverse complementary sequence" ], "rating_list": [ 3, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 5, 3, 2, 3 ] }, { "paper_id": "IjVCcykKdr", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Hao Tian", "gender": "unknown", "institution": "Baidu Research", "country": "", "position": "Researcher" }, { "name": "Haoran Tang", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Jieren Deng", "gender": "unknown", "institution": "University of Connecticut", "country": "", "position": "PhD student" }, { "name": "Pratik A Chaudhari", "gender": "Male", "institution": "School of Engineering and Applied Science, University of Pennsylvania", "country": "US", "position": "Assistant Professor" }, { "name": "Xin Zhou", "gender": "Male", "institution": "Baidu", "country": "", "position": "Researcher" }, { "name": "Zhihong Pan", "gender": "Male", "institution": "Samsung", "country": "US", "position": "Principal Engineer" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 30, "n_ref": 55, "n_ref_all": 63, "n_fig": 6, "n_tab": 4, "L_tab": 2020, "n_element_tab": 100, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1454, "n_element_tab_1": 246, "formula_len_all": 232, "formula_len_all_1": 232, "len_all": 127654, "len_all_1": 74478, "len_abs": 315, "len_title": 149, "len_sents": 38770, "len_sents_1": 35419, "n_sents": 297, "n_sents_1": 274, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1618, "title": "RetriBooru: Leakage-Free Retrieval of Conditions from Reference Images for Subject-Driven Generation", "abs": "Diffusion-based methods have demonstrated remarkable capabilities in generating a diverse array of high-quality images, sparking interests for styled avatars, virtual try-on, and more. Previous methods use the same reference image as the target. An overlooked aspect is the leakage of the target's spatial information, style, etc. from the reference, harming the generated diversity and causing shortcuts. However, this approach continues as widely available datasets usually consist of single images not grouped by identities, and it is expensive to recollect large-scale same-identity data. Moreover, existing metrics adopt decoupled evaluation on text alignment and identity preservation, which fail at distinguishing between balanced outputs and those that over-fit to one aspect. \nIn this paper, we propose a multi-level, same-identity dataset RetriBooru, which groups anime characters by both face and cloth identities. RetriBooru enables adopting reference images of the same character and outfits as the target, while keeping flexible gestures and actions. We benchmark previous methods on our dataset, and demonstrate the effectiveness of training with a reference image different from target (but same identity). We introduce a new concept composition task, where the conditioning encoder learns to retrieve different concepts from several reference images, and modify a baseline network RetriNet for the new task. Finally, we introduce a novel class of metrics named Similarity Weighted Diversity (SWD), to measure the overlooked diversity and better evaluate the alignment between similarity and diversity.", "keywords": [ "Dataset", "Subject-driven generation", "Evaluation Metrics", "Diffusion Models" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "IjQ2Jtemzy", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Anna Sztyber-Betley", "gender": "Female", "institution": "Warsaw University of Technology", "country": "PL", "position": "Assistant Professor" }, { "name": "James Chua", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Jan Betley", "gender": "Male", "institution": "Owain Evans' research group", "country": "US", "position": "Researcher" }, { "name": "Mart\u00edn Soto", "gender": "Male", "institution": "Universitat de Barcelona", "country": "ES", "position": "MS student" }, { "name": "Owain Evans", "gender": "unknown", "institution": "Truthful AI", "country": "US", "position": "Principal Researcher" }, { "name": "Xuchan Bao", "gender": "Female", "institution": "University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 26, "n_ref": 50, "n_ref_all": 186, "n_fig": 23, "n_tab": 29, "L_tab": 4331, "n_element_tab": 232, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1361, "n_element_tab_1": 91, "formula_len_all": 1, "formula_len_all_1": 22, "len_all": 325652, "len_all_1": 70510, "len_abs": 1144, "len_title": 99, "len_sents": 129586, "len_sents_1": 31419, "n_sents": 1246, "n_sents_1": 286, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 51, "L_abs": 1154, "title": "Language Models Can Articulate Their Implicit Goals", "abs": "We study *objective awareness*, which we define as an LLM's capability to articulate its behavioral policies without relying on in-context examples. We finetune LLMs on examples that exhibit particular behaviors, including (a) making risk-seeking / risk-averse economic decisions, and (b) making the user say a certain word. Although these examples never contain explicit descriptions of the policy (e.g. ``I will now take the risk-seeking option''), we find that the finetuned LLMs can explicitly describe their policies through out-of-context reasoning. We demonstrate LLMs' objective awareness across various evaluation tasks, both for multiple-choice and free-form questions. Furthermore, we demonstrate that models can correctly attribute different learned policies to distinct personas. Finally, we explore the connection between objective awareness and the concept of backdoors in AI safety, where certain behaviors are implanted in a model, often through data poisoning, and can be triggered under certain conditions. We find evidence that LLMs can recognize the existence of the backdoor-like behavior that they have acquired through finetuning.", "keywords": [ "NLP", "LLM", "GPT", "generalization", "out-of-context reasoning", "capabilities", "fine-tuning", "self-awareness", "self-knowledge" ], "rating_list": [ 8, 6, 8, 5, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3, 2 ] }, { "paper_id": "IjLumaGZ4h", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Jonghyun Choi", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" }, { "name": "Mohammad Mostafavi", "gender": "unknown", "institution": "Isfahan University of Medical Sciences", "country": "IR", "position": "Postdoc" }, { "name": "Seokjun Moon", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "MS student" }, { "name": "Sankarshana V", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 41, "n_ref": 100, "n_ref_all": 135, "n_fig": 9, "n_tab": 14, "L_tab": 3374, "n_element_tab": 435, "n_fig_1": 8, "n_tab_1": 8, "L_tab_1": 3547, "n_element_tab_1": 204, "formula_len_all": 122, "formula_len_all_1": 0, "len_all": 144394, "len_all_1": 57096, "len_abs": 1626, "len_title": 110, "len_sents": 40436, "len_sents_1": 23653, "n_sents": 324, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1635, "title": "ROSE: Reduced Overhead Stereo Event-Intensity Depth Estimation", "abs": "Stereo depth estimation using event cameras is a promising approach for real-time vision tasks, offering low-latency, high-speed data capture. However, existing methods often suffer from high computational overhead, limiting their real-time applicability. To address these challenges, we introduce ROSE (Reduced Overhead Stereo Event and Intensity) a Real-Time framework for efficient depth estimation from events and intensity images. Current approaches rely on dense networks that fail to scale with increasing data complexity, constraining both accuracy and speed. In contrast, ROSE incorporates lightweight event representation networks and optimizes the stereo matching process to reduce model size and computational load without compromising accuracy.\nWe replace conventional network components with efficient spatio-temporal representations and streamline adaptive aggregation modules, reducing computational complexity by 1000\u00d7 compared to previous methods. Furthermore, we adapt event grouping strategies to better align with intensity images, improving the quality of depth estimation under various lighting and motion conditions. Extensive experiments on the DSEC and MVSEC benchmarks demonstrate that ROSE achieves real-time performance, boosting frame rates to 32.2 FPS on DSEC and 66.9 FPS on MVSEC while maintaining competitive depth accuracy. This marks a significant improvement over prior work in terms of speed and scalability, making ROSE a viable solution for real-time stereo depth estimation in resource-constrained environments. Our code and models will be released to support further advancements in the field.", "keywords": [ "Event-based Vision", "Stereo Depth Estimation" ], "rating_list": [ 5, 5, 3, 6, 5 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "IjBndR92Zy", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Daoze Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Junru Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Ruizhe Zheng", "gender": "Male", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Yang Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhizhang Yuan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "yuguo yu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 12, "n_ref_uni": 59, "n_ref": 122, "n_ref_all": 144, "n_fig": 6, "n_tab": 14, "L_tab": 4979, "n_element_tab": 585, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1804, "n_element_tab_1": 246, "formula_len_all": 1365, "formula_len_all_1": 864, "len_all": 242006, "len_all_1": 61427, "len_abs": 1646, "len_title": 163, "len_sents": 78038, "len_sents_1": 28485, "n_sents": 557, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1650, "title": "Beatrix: Out-of-Distribution Generalization of Large EEG Model via Invariant Contrastive Fine-Tuning", "abs": "The advent of large-scale foundation models has revolutionized EEG analysis; however, their ability to generalize to Out-of-Distribution (OoD) brain signals remains limited due to the inherent variability in physiological states, individual differences, and experimental setups. To address these challenges, we introduce Beatrix, a novel spectral EEG foundation model that achieves state-of-the-art OoD generalization across diverse brain activity tasks. Beatrix leverages a unique analytic wavelet-based spectral tokenization that captures the intricate non-stationary dynamics of EEG signals, and employs a semi-causal generative modeling approach during pre-training, enabling it to learn expressive latent representations capable of both interpolation and extrapolation across temporal and frequency domains. For fine-tuning, we propose an innovative Contrastive Invariant Fine-Tuning (CIFT) method that enhances domain-invariant learning without the need for explicit environment labels, thus significantly improving OoD generalizability in a parameter-efficient manner. Our multi-view Transformer architecture further integrates both spectral and temporal information, allowing Beatrix to comprehensively model EEG signals across channels. Extensive experiments demonstrate that Beatrix consistently outperforms existing EEG models in tasks such as seizure detection and forecasting, auditory neural decoding, motor imagery, and sleep staging, showcasing its robustness and broad applicability. By achieving superior performance with reduced fine-tuning costs, Beatrix represents a significant advancement in the field of EEG foundation models.", "keywords": [ "EEG; representation learning; parameter-efficient fine-tuning; domain generalization; seizure diagnosis;" ], "rating_list": [ 1, 3, 8, 3 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "Ij9ilPh36h", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Daniel Ward", "gender": "Not Specified", "institution": "PricewaterhouseCoopers", "country": "SE", "position": "Indepenent Researcher" }, { "name": "Fangyu Liu", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Fredrik Carlsson", "gender": "Male", "institution": "KTH Royal Institute of Technology, Stockholm, Sweden", "country": "", "position": "MS student" }, { "name": "Joakim Nivre", "gender": "Male", "institution": "Uppsala University", "country": "", "position": "Full Professor" }, { "name": "Murathan Kurfal\u0131", "gender": "unknown", "institution": "RISE Research Institutes of Sweden AB", "country": "SE", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 25, "n_ref": 46, "n_ref_all": 78, "n_fig": 12, "n_tab": 7, "L_tab": 2675, "n_element_tab": 420, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1301, "n_element_tab_1": 218, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 253547, "len_all_1": 59748, "len_abs": 1385, "len_title": 139, "len_sents": 50425, "len_sents_1": 27461, "n_sents": 385, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1395, "title": "The Hyperfitting Phenomenon: Sharpening and Stabilizing LLMs for Open-Ended Text Generation", "abs": "This paper introduces the counter-intuitive generalization results of overfitting pre-trained large language models (LLMs) on very small datasets. In the setting of open-ended text generation, it is well-documented that LLMs tend to generate repetitive and dull sequences, a phenomenon that is especially apparent when generating using greedy decoding. This issue persists even with state-of-the-art LLMs containing billions of parameters, trained via next-token prediction on large datasets. We find that by further fine-tuning these models to achieve a near-zero training loss on a small set of samples -- a process we refer to as hyperfitting -- the long-sequence generative capabilities are greatly enhanced.\nGreedy decoding with these Hyperfitted models even outperform Top-P sampling over long-sequences, both in terms of diversity and human preferences. This phenomenon extends to LLMs of various sizes, different domains, and even autoregressive image generation. We further find this phenomena to be distinctly different from that of Grokking and double descent. Surprisingly, our experiments indicate that hyperfitted models rarely fall into repeating sequences they were trained on, and even explicitly blocking these sequences results in high-quality output. All hyperfitted models produce extremely low-entropy predictions, often allocating nearly all probability to a single token.", "keywords": [ "LLM", "NLP", "fine-tuning", "open-ended text generation", "Hyperfitting", "Phenomenon", "Neural Networks", "Early-stopping", "Overfitting" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "IiwyThOFXL", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Fengjun Guo", "gender": "Male", "institution": "Intsig Information Co. Ltd", "country": "CN", "position": "Principal Researcher" }, { "name": "Huanzhang Dou", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" }, { "name": "Xi Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yike Yuan", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 61, "n_ref": 112, "n_ref_all": 127, "n_fig": 9, "n_tab": 5, "L_tab": 516, "n_element_tab": 41, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 362, "n_element_tab_1": 42, "formula_len_all": 309, "formula_len_all_1": 39, "len_all": 150621, "len_all_1": 55521, "len_abs": 494, "len_title": 152, "len_sents": 34970, "len_sents_1": 26652, "n_sents": 263, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 954, "title": "SemanticMIM: Marring Masked Image Modeling with Semantics Compression for General Visual Representation", "abs": "This paper represents a neat yet effective framework, named SemanticMIM, to integrate the advantages of masked image modeling (MIM) and contrastive learning (CL) for general visual representation. We conduct a thorough comparative analysis between CL and MIM, revealing that their complementary advantages fundamentally stem from two distinct phases, i.e., compression and reconstruction. Specifically, SemanticMIM leverages a proxy architecture that customizes interaction between image and mask tokens, bridging these two phases to achieve general visual representation with the property of abundant semantic and positional awareness. Through extensive qualitative and quantitative evaluations, we demonstrate that SemanticMIM effectively amalgamates the benefits of CL and MIM, leading to significant enhancement of performance and feature linear separability. SemanticMIM also offers notable interpretability through attention response visualization.", "keywords": [ "Self-supervised learning", "Masked image modeling" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 5, 3, 3 ] }, { "paper_id": "IiagjrJNwF", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Beidi Chen", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Jianyu Zhang", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Leon Bottou", "gender": "Male", "institution": "Facebook AI Research", "country": "US", "position": "Researcher" }, { "name": "Niklas Nolte", "gender": "Not Specified", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Ranajoy Sadhukhan", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 36, "n_ref": 57, "n_ref_all": 99, "n_fig": 14, "n_tab": 9, "L_tab": 2878, "n_element_tab": 573, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 774, "n_element_tab_1": 81, "formula_len_all": 1004, "formula_len_all_1": 681, "len_all": 158732, "len_all_1": 67607, "len_abs": 581, "len_title": 62, "len_sents": 52550, "len_sents_1": 32839, "n_sents": 416, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 14, "L_abs": 516, "title": "Memory Mosaics", "abs": "Memory Mosaics are networks of associative memories working in concert to achieve a prediction task of interest. Like transformers, memory mosaics possess compositional capabilities and in-context learning capabilities. Unlike transformers, memory mosaics achieve these capabilities in comparatively transparent way (\u201cpredictive disentanglement\u201d). We illustrate these capabilities on a toy example and also show that memory mosaics perform as well or better than transformers on medium-scale language modeling tasks.", "keywords": [ "predictive disentanglement", "Associative memory", "language model" ], "rating_list": [ 6, 8, 8, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "IiWZ9rB2Ef", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Henrik Bostr\u00f6m", "gender": "unknown", "institution": "KTH Royal Institute of Technology, Stockholm, Sweden", "country": "SE", "position": "Professor" }, { "name": "Johannes F. Lutzeyer", "gender": "Male", "institution": "Ecole Polytechique", "country": "FR", "position": "Assistant Professor" }, { "name": "Michalis Vazirgiannis", "gender": "Male", "institution": "Ecole Polytechnique, France", "country": "", "position": "Full Professor" }, { "name": "Sofiane ENNADIR", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "PhD student" }, { "name": "Yassine ABBAHADDOU", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "FR", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 10, "n_ref_uni": 25, "n_ref": 59, "n_ref_all": 63, "n_fig": 0, "n_tab": 2, "L_tab": 840, "n_element_tab": 75, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 2141, "n_element_tab_1": 23, "formula_len_all": 1532, "formula_len_all_1": 1120, "len_all": 117052, "len_all_1": 68532, "len_abs": 1302, "len_title": 124, "len_sents": 42309, "len_sents_1": 34202, "n_sents": 293, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1306, "title": "Adversarially Robust Graph Classification: A Pooling-Based Defense Framework", "abs": "Graph Neural Networks (GNNs) have shown great success across various domains but remain vulnerable to adversarial attacks. While most defense methodology focuses on node classification and enhancing robustness during training, this work shifts the focus to graph classification and inference-time defenses. We theoretically show that the final pooling operation, that is required for graph-level tasks, can have an impact on the graph classifier's underlying robustness. Based on this analysis, we propose a pre-pooling operation, called R-Pool (Robust-Pooling), which is based a novel filtering mechanism using Gaussian Mixture Models (GMMs) to detect and exclude nodes heavily impacted by attacks, thereby enhancing robustness at inference time. Our framework can be used with any pooling operation and any underlying model, and does not require re-training the model nor adapting its architecture. Our experiments demonstrate that this approach effectively mitigates adversarial effects while maintaining a balance between clean and attacked accuracy. Through extensive evaluations on state-of-the-art adversarial attacks, we show that the proposed framework significantly improves the robustness of the underlying GNNs in graph classification tasks compared to other available post-hoc defense methods.", "keywords": [ "Adversarial Robustness", "Graph Neural Networks" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "IiRlImvLQI", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Alan Yuille", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Ariel Lubonja", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Daniel Khashabi", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" }, { "name": "Luoxin Ye", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Researcher" }, { "name": "Wufei Ma", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Zhao-Yang Wang", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Zhongrui Gui", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Undergrad student" }, { "name": "J Chen", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 2, "n_ref_uni": 57, "n_ref": 102, "n_ref_all": 117, "n_fig": 7, "n_tab": 9, "L_tab": 2210, "n_element_tab": 223, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 402, "n_element_tab_1": 33, "formula_len_all": 1156, "formula_len_all_1": 94, "len_all": 192498, "len_all_1": 61719, "len_abs": 1244, "len_title": 123, "len_sents": 47322, "len_sents_1": 30047, "n_sents": 358, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1274, "title": "CycleAug: Cycle-Consistent Visual Augmentation for Large Multimodal Models", "abs": "Training multimodal large language models (MLLMs) requires high-quality image-question-answer (IQA) triplets, which are labour-intensive to curate and often lack diversity. We propose a novel data augmentation framework for visual instruction tuning that efficiently generates diverse synthetic images based on existing IQA anchor triplets. To ensure that the generated images align with their associated QA pairs, we propose CycleAug --- cycle-consistency visual augmentation which involves synthesizing images from text (text $\\rightarrow$ image) and then performing a verification step to confirm that the answers derived from the synthetic images match the original answers (image $\\rightarrow$ text), ensuring consistency across images, questions, and answers. By combining synthetic images with high-quality real data in the training phase, we demonstrate these synthetic triplets act as an implicit regularization, which improves the robustness of MLLMs and enables analogical reasoning. Extensive experiments show that our approach improves model performance on multiple visual question-answering benchmarks without additional real-world data. This work highlights the potential of leveraging visual foundational models to enhance visual instruction tuning in MLLMs.", "keywords": [ "large multimodal models; synthetic data; data augmentation" ], "rating_list": [ 3, 5, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 3, 5, 3, 3, 4 ] }, { "paper_id": "IiMKfn4dxk", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Jiahui Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Junfeng Fang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Wenjie Du", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Assistant Professor" }, { "name": "Yang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zhengyang Zhou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Jun Xia", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "shuai zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "XuqiangLi", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 41, "n_formula_1": 34, "n_ref_uni": 33, "n_ref": 51, "n_ref_all": 83, "n_fig": 6, "n_tab": 16, "L_tab": 6585, "n_element_tab": 587, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 4760, "n_element_tab_1": 460, "formula_len_all": 2628, "formula_len_all_1": 1755, "len_all": 180947, "len_all_1": 64648, "len_abs": 1766, "len_title": 165, "len_sents": 45150, "len_sents_1": 23095, "n_sents": 349, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 112, "L_abs": 1775, "title": "Capturing substructure interactions by invariant Information Bottle Theory for Generalizable Property Prediction", "abs": "Molecular interactions are a common phenomenon in physical chemistry, often resulting in unexpected biochemical properties harmful to humans, such as drug-drug interactions. Machine learning has shown great potential for predicting these interactions rapidly and accurately. However, the complexity of molecular structures and the diversity of interactions often reduce prediction accuracy and hinder generalizability. Identifying core invariant substructures (i.e., rationales) has become essential to improving the model's interpretability and generalization. Despite significant progress, existing models frequently overlook the pairwise molecular interaction, leading to insufficient capture of interaction dynamics. To address these limitations, we propose I2Mole (Interaction-aware Invariant Molecular learning), a novel framework for generalizable property prediction. I2Mole meticulously models atomic interactions, such as hydrogen bonds and Van der Waals forces, by first establishing indiscriminate connections between intermolecular atoms, which are then refined using an improved graph information bottleneck theory tailored for merged graphs. To further enhance model generalization, we construct an environment codebook by environment subgraph of the merged graph. This approach not only could provide noise source for optimizing mutual information but also preserve the integrity of chemical semantic information. By comprehensively leveraging the information inherent in the merged graph, our model accurately captures core substructures and significantly enhances generalization capabilities. Extensive experimental validation demonstrates I2Mole's efficacy and generalizability. The implementation code is available at https://anonymous.4open/r/I2Mol-C616.", "keywords": [ "Molecular Merged Graph.+ Drug-drug interaction.+ Out-of-Distribution" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "IiAckbuccF", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 7, "author_info_list": [ { "name": "Duc Quang Nguyen", "gender": "Male", "institution": "Ho Chi Minh City University of Technology", "country": "VN", "position": "MS student" }, { "name": "Nick Haber", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Ryan-Rhys Griffiths", "gender": "Male", "institution": "FutureHouse", "country": "US", "position": "Researcher" }, { "name": "Sang T. Truong", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Stefano Ermon", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Willie Neiswanger", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 11, "n_ref_uni": 77, "n_ref": 154, "n_ref_all": 182, "n_fig": 18, "n_tab": 5, "L_tab": 1279, "n_element_tab": 23, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1841, "formula_len_all_1": 1217, "len_all": 212855, "len_all_1": 74367, "len_abs": 1196, "len_title": 104, "len_sents": 56672, "len_sents_1": 35208, "n_sents": 420, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1201, "title": "Nonmyopic Bayesian Optimization in Dynamic Cost Settings", "abs": "Bayesian optimization (BO) is a popular framework for optimizing black-box functions, leveraging probabilistic models such as Gaussian processes. However, conventional BO assumes static query costs, which limits its applicability to real-world problems with dynamic cost structures, such as geological surveys or biological sequence design, where query costs vary based on previous actions. To address this, we propose a cost-constrained nonmyopic BO algorithm that incorporates dynamic cost models. Our method employs a neural network policy for variational optimization over multi-step lookahead horizons to plan ahead in dynamic cost environments. Empirically, we benchmark our method on synthetic functions exhibiting a variety of dynamic cost structures. Furthermore, we apply our method to a real-world application in protein sequence design using a large language model-based policy, demonstrating its scalability and effectiveness in handling multi-step planning in a large and complex query space. Our nonmyopic BO algorithm consistently outperforms its myopic counterparts in both synthetic and real-world settings, achieving significant improvements in both efficiency and solution quality.", "keywords": [ "nonmyopic Bayesian optimization", "dynamic cost settings", "language model policy optimization" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "Iht4NNVqk0", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Buelent Uendes", "gender": "unknown", "institution": "Vrije Universiteit Amsterdam", "country": "", "position": "PhD student" }, { "name": "Mark Hoogendoorn", "gender": "Male", "institution": "VU University Amsterdam", "country": "", "position": "Full Professor" }, { "name": "Shujian Yu", "gender": "Male", "institution": "Vrije Universiteit Amsterdam", "country": "NL", "position": "Assistant Professor" } ], "n_formula": 55, "n_formula_1": 14, "n_ref_uni": 45, "n_ref": 179, "n_ref_all": 198, "n_fig": 2, "n_tab": 9, "L_tab": 2476, "n_element_tab": 81, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 411, "n_element_tab_1": 24, "formula_len_all": 4794, "formula_len_all_1": 1446, "len_all": 244776, "len_all_1": 66475, "len_abs": 1487, "len_title": 118, "len_sents": 90189, "len_sents_1": 32774, "n_sents": 585, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1489, "title": "Start Smart: Leveraging Gradients For Enhancing Mask-based XAI Methods", "abs": "Mask-based explanation methods offer a powerful framework for interpreting deep learning model predictions across diverse data modalities, such as images and time series, in which the central idea is to identify an instance-dependent mask that minimizes the performance drop from the resulting masked input. Different objectives for learning such masks have been proposed, all of which, in our view, can be unified under an information-theoretic framework that balances performance degradation of the masked input with the complexity of the resulting masked representation. Typically, these methods initialize the masks either uniformly or as all-ones.\nIn this paper, we argue that an effective mask initialization strategy is as important as the development of novel learning objectives, particularly in light of the significant computational costs associated with mask-based explanation methods. To this end, we introduce a new gradient-based initialization technique called StartGrad, which is the first initialization method specifically designed for mask-based post-hoc explainability methods. Compared to commonly used strategies, StartGrad is provably superior at initialization in striking the aforementioned tradeoff. Despite its simplicity, our experiments demonstrate that StartGrad consistently helps to speed up the optimization process of various state-of-the-art mask-explanation method by reaching target metrics quicker and, in some cases, even boosts overall performance.", "keywords": [ "XAI", "mask-based explanations", "rate-distortion explanation", "information bottleneck" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "IhbZytsinc", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Greg Heinrich", "gender": "Male", "institution": "NVIDIA", "country": "FR", "position": "Researcher" }, { "name": "Jan Kautz", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "VP Research" }, { "name": "Pavlo Molchanov", "gender": "Male", "institution": "NVIDIA Research", "country": "US", "position": "Research Scientist" }, { "name": "Peter Belcak", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 26, "n_ref": 45, "n_ref_all": 60, "n_fig": 5, "n_tab": 9, "L_tab": 4656, "n_element_tab": 648, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1259, "n_element_tab_1": 109, "formula_len_all": 426, "formula_len_all_1": 426, "len_all": 123643, "len_all_1": 58168, "len_abs": 960, "len_title": 139, "len_sents": 36740, "len_sents_1": 27453, "n_sents": 282, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 966, "title": "Minifinetuning: Low-Data Generation Domain Adaptation through Corrective Self-Distillation", "abs": "Finetuning language models for a new domain inevitably leads to the deterioration of their general performance.\nThis becomes more pronounced the more limited the finetuning data resource.\n\nWe introduce minifinetuning (MFT), a method for language model domain adaptation that considerably reduces the effects of overfitting-induced degeneralization in low-data settings and which does so in the absence of any pre-training data for replay.\nMFT demonstrates 2-10x more favourable specialization-to-degeneralization ratios than standard finetuning across a wide range of models and domains and exhibits an intrinsic robustness to overfitting when data in the new domain is scarce and down to as little as 500 samples.\n\nEmploying corrective self-distillation that is individualized on the sample level, MFT outperforms parameter-efficient finetuning methods, demonstrates replay-like forgetting mitigation properties, and is composable with either for a combined effect.", "keywords": [ "minifinetuning", "mft", "finetuning", "low-resource finetuning", "self-distillation", "corrective distillation" ], "rating_list": [ 5, 6, 5, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 3, 3, 2, 4 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "IgrLJslvxa", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "David Krueger", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Assistant Professor" }, { "name": "Fazl Barez", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Mrinank Sharma", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Shay B Cohen", "gender": "Male", "institution": "University of Edinburgh", "country": "GB", "position": "Reader" }, { "name": "Tingchen Fu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "MS student" }, { "name": "Yonadav G Shavit", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 62, "n_ref": 151, "n_ref_all": 190, "n_fig": 8, "n_tab": 16, "L_tab": 3914, "n_element_tab": 384, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 1152, "n_element_tab_1": 139, "formula_len_all": 292, "formula_len_all_1": 138, "len_all": 305795, "len_all_1": 59924, "len_abs": 1196, "len_title": 134, "len_sents": 47863, "len_sents_1": 26583, "n_sents": 335, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1226, "title": "PoisonBench: Assessing Large Language Model Vulnerability to Data Poisoning", "abs": "Preference learning is a central component for aligning current LLMs, but this process can be vulnerable to data poisoning attacks. To address this concern, we introduce PoisonBench, a benchmark for evaluating large language models' susceptibility to data poisoning during preference learning. Data poisoning attacks can manipulate large language model responses to include hidden malicious content or biases, potentially causing the model to generate harmful or unintended outputs while appearing to function normally. We deploy two distinct attack types across eight realistic scenarios, assessing 22 widely-used models. Our findings reveal concerning trends: (1) Scaling up parameter size does not always enhance resilience against poisoning attacks and the influence on model resilience varies among different model suites. (2) There exists a log-linear relationship between the effects of the attack and the data poison ratio; (3) The effect of data poisoning can generalize to extrapolated triggers that are not included in the poisoned data. \nThese results expose weaknesses in current preference learning techniques, highlighting the urgent need for more robust defenses against malicious models and data manipulation.", "keywords": [ "alignment", "data poisoning", "backdoor attack", "AI safety" ], "rating_list": [ 5, 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "Igm9bbkzHC", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Chris Wendler", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Postdoc" }, { "name": "Giovanni Monea", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Julian Minder", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Kevin Du", "gender": "unknown", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Niklas Stoehr", "gender": "unknown", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" }, { "name": "Robert West", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" }, { "name": "Ryan Cotterell", "gender": "Not Specified", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 44, "n_ref": 103, "n_ref_all": 148, "n_fig": 27, "n_tab": 1, "L_tab": 496, "n_element_tab": 23, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 658, "n_element_tab_1": 33, "formula_len_all": 695, "formula_len_all_1": 435, "len_all": 257038, "len_all_1": 73472, "len_abs": 1723, "len_title": 103, "len_sents": 58219, "len_sents_1": 34344, "n_sents": 454, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1746, "title": "Controllable Context Sensitivity and the Knob Behind It", "abs": "When making predictions, a language model must trade off how much it relies on its context vs. its prior knowledge.\nChoosing how sensitive the model is to its context is a fundamental functionality, as it enables the model to excel at tasks like retrieval-augmented generation and question-answering. \nIn this paper, we search for a knob which controls this sensitivity, determining whether language models answer from the context or their prior knowledge.\nTo guide this search, we design a task for controllable context sensitivity. \nIn this task, we first feed the model a context (\"Paris is in England\") and a question (\"Where is Paris?\"); we then instruct the model to either use its prior or contextual knowledge and evaluate whether it generates the correct answer for both intents (either \"France\" or \"England\").\nWhen fine-tuned on this task, instruct versions of Llama-3.1, Mistral-v0.3, and Gemma-2 can solve it with high accuracy (85-95%). \nAnalyzing these high-performing models, we narrow down which layers may be important to context sensitivity using a novel linear time algorithm. \nThen, in each model, we identify a 1-D subspace in a single layer that encodes whether the model follows context or prior knowledge.\nInterestingly, while we identify this subspace in a fine-tuned model, we find that the exact same subspace serves as an effective knob in not only that model but also non-fine-tuned instruct and base models of that model family.\nFinally, we show a strong correlation between a model's performance and how distinctly it separates context-agreeing from context-ignoring answers in this subspace.\nThese results suggest a single fundamental subspace facilitates how the model chooses between context and prior knowledge.", "keywords": [ "analysis", "interpretability", "mechanistic interpretability", "context vs prior knowledge", "large language models" ], "rating_list": [ 8, 3, 8, 8 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 1, 3, 3 ], "confidence_list": [ 3, 2, 3, 4 ] }, { "paper_id": "IfPfUHRowT", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Bin Ren", "gender": "Male", "institution": "William & Mary", "country": "US", "position": "Associate Professor" }, { "name": "Jiaze E", "gender": "unknown", "institution": "College of William and Mary", "country": "US", "position": "PhD student" }, { "name": "Srutarshi Banerjee", "gender": "Male", "institution": "Argonne National Laboratory", "country": "US", "position": "Postdoc" }, { "name": "Tekin Bicer", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 67, "n_ref": 82, "n_ref_all": 107, "n_fig": 11, "n_tab": 4, "L_tab": 750, "n_element_tab": 114, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 390, "n_element_tab_1": 56, "formula_len_all": 822, "formula_len_all_1": 676, "len_all": 149848, "len_all_1": 67497, "len_abs": 1988, "len_title": 138, "len_sents": 35484, "len_sents_1": 32269, "n_sents": 248, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 2019, "title": "Inpainting the Sinogram from Computed Tomography using Latent Diffusion Model and Physics", "abs": "Computed Tomography (CT) is a widely used non-invasive imaging technique for materials at microscopic or sub-microscopic length scales in synchrotron radiation facilities. Typically, the object is rotated relative to the X-ray beam, and 2D projection images are recorded by the detector at different rotation angles. The 3D object is then reconstructed by combining these projections and solving a computationally demanding inverse problem. The quality of the reconstructed image is critical for scientific analysis and is influenced by various factors, including the number of projections, exposure time or dose, and the reconstruction algorithm. In this work, we develop a foundation model by integrating a Generative AI-based Latent Diffusion Model (LDM) with physics-based domain knowledge. Specifically, we first incorporate a set of loss functions into our LDM that accurately capture the physical properties of the CT data acquisition process. We demonstrate that addition of these loss functions aids in stable training of the autoencoder in the LDM and improves its accuracy. The autoencoder and the Diffusion model of the LDM is trained with real-world experimental data. Collecting real world experimental data from Synchrotron beamlines is often time-consuming and challenging. We demonstrate that the autoencoder trained with a combination of real world experimental data and phantom shapes features also performs comparable to the autoencoder trained with real world data. Second, we introduce a novel image blending method to combine the LDM\u2019s generated output with the original, extremely sparse sinogram data. Since our model integrates physics-guided loss functions focused on CT data acquisition, it simplifies the creation of downstream tasks and facilitates the adaptation of new features from different experiments. Our experimental evaluation demonstrates improvements of upto 23.5 % in SSIM for sinogram quality and 13.8 % for reconstructed image quality compared to state-of-the-art techniques.", "keywords": [ "Sinogram Inpainting", "Physics", "Latent Diffusion Model", "X-ray Imaging" ], "rating_list": [ 1, 6, 3, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "Iemy0Fc3Pw", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Luke Taylor", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Melissa Claire Maaike Fasol", "gender": "Female", "institution": "University of Edinburgh", "country": "GB", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 67, "n_ref": 142, "n_ref_all": 192, "n_fig": 34, "n_tab": 1, "L_tab": 613, "n_element_tab": 5, "n_fig_1": 33, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 659, "formula_len_all_1": 187, "len_all": 153979, "len_all_1": 67210, "len_abs": 1223, "len_title": 139, "len_sents": 39391, "len_sents_1": 32869, "n_sents": 269, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1228, "title": "Spike-to-excite: photosensitive seizures in biologically-realistic spiking neural networks", "abs": "Photosensitive Epilepsy (PE) is a neurological disorder characterized by seizures triggered by harmful visual stimuli, such as flashing lights and high-contrast patterns. The mechanisms underlying PE remain poorly understood, and to date, no computational model has captured the phenomena associated with this condition. Biologically detailed spiking networks trained for efficient prediction of natural scenes have been shown to capture V1-like characteristics. Here, we show that these models display seizure-like activity in response to harmful stimuli while retaining healthy responses to non-provocative stimuli when post-synaptic inhibitory connections are weakened. Notably, our adapted model resembles the motion tuning and contrast gain responses of excitatory V1 neurons in mice with optogenetically reduced inhibitory activity. We offer testable predictions underlying the pathophysiology of PE by exploring how reduced inhibition leads to seizure-like activity. Finally, we show that artificially injecting pulsating input current into the model units prevents seizure-like activity and restores baseline function. In summary, we present a model of PE that offers new insights to understand and treat this condition.", "keywords": [ "photosensitive epilepsy", "spiking neural network", "V1", "prediction", "deep brain stimulation" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "IeZpJNc3uy", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Anton van den Hengel", "gender": "unknown", "institution": "University of Adelaide", "country": "AU", "position": "Professor" }, { "name": "Haiming Xu", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Lingqiao Liu", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Associate Professor" }, { "name": "Yangyang Shu", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Ziqin Zhou", "gender": "Female", "institution": "University of Adelaide", "country": "AU", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 31, "n_ref": 49, "n_ref_all": 74, "n_fig": 9, "n_tab": 5, "L_tab": 422, "n_element_tab": 72, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 542, "n_element_tab_1": 61, "formula_len_all": 594, "formula_len_all_1": 579, "len_all": 114595, "len_all_1": 66698, "len_abs": 1363, "len_title": 146, "len_sents": 39002, "len_sents_1": 32521, "n_sents": 286, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1370, "title": "Towards Effective Updating of Pretrained Symbolic Music Models for Fine-Grained Bar-Level Control", "abs": "Automatically generating symbolic music scores tailored to specific user needs offers significant benefits for musicians and enthusiasts alike. Pretrained symbolic music autoregressive models have demonstrated promising results, thanks to large datasets and advanced transformer architectures. However, in practice, the control provided by such models is often limited, particularly when fine-grained controls are needed at the level of individual bars. While fine-tuning the model with newly introduced control tokens may seem like a straightforward solution, our research reveals challenges in this approach, as the model frequently struggles to respond effectively to these precise bar-level control signals. To overcome this issue, we propose two novel strategies. First, we introduce a pre-training task that explicitly links control signals with their corresponding musical tokens, enabling a more effective initialization for fine-tuning. Second, we develop a unique counterfactual loss function that enhances alignment between the generated music and the specified control prompts. These combined methods substantially improve bar-level control, yielding a 13.06\\% improvement over the fine-tuning baseline. Importantly, subjective evaluations confirm that this increased control does not compromise the musical quality produced by the original pretrained model.", "keywords": [ "Symbolic Music Model", "Fine-Grained Bar-Level Control" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 1, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "IeRcpsdY7P", "primary_area": "learning on time series and dynamical systems", "n_author": 8, "author_info_list": [ { "name": "Aakash Patel", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "David van Dijk", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" }, { "name": "Emanuele Zappala", "gender": "Male", "institution": "Idaho State University", "country": "US", "position": "Assistant Professor" }, { "name": "Nianchen Liu", "gender": "Female", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Shiyang Zhang", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Undergrad student" }, { "name": "Sizhuang He", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Syed A Rizvi", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "amin karbasi", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Associate Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 40, "n_ref": 49, "n_ref_all": 62, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 125837, "len_all_1": 60244, "len_abs": 1154, "len_title": 81, "len_sents": 43482, "len_sents_1": 31110, "n_sents": 292, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 33, "L_abs": 1126, "title": "Intelligence at the Edge of Chaos", "abs": "We explore the emergence of intelligent behavior in artificial systems by investigating how the complexity of rule-based systems influences the capabilities of models trained to predict these rules. Our study focuses on elementary cellular automata (ECA), simple yet powerful one-dimensional systems that generate behaviors ranging from trivial to highly complex. By training distinct Large Language Models (LLMs) on different ECAs, we evaluated the relationship between the complexity of the rules' behavior and the intelligence exhibited by the LLMs, as reflected in their performance on downstream tasks. Our findings reveal that rules with higher complexity lead to models exhibiting greater intelligence, as demonstrated by their performance on reasoning and chess move prediction tasks. Both uniform and periodic systems, and often also highly chaotic systems, resulted in poorer downstream performance, highlighting a sweet spot of complexity conducive to intelligence. We conjecture that intelligence arises from the ability to predict complexity and that creating intelligence may require only exposure to complexity.", "keywords": [ "Large Language Models (LLMs)", "Elementary Cellular Automata (ECA)", "Emergent Intelligence", "Complex Systems", "Complexity Theory" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "IdynViNzwI", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Lu Zhenjie", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "Undergrad student" }, { "name": "WEI W. XING", "gender": "Male", "institution": "University of Sheffield", "country": "GB", "position": "Lecturer" }, { "name": "Yuxin Wang", "gender": "Female", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 72, "n_formula_1": 18, "n_ref_uni": 27, "n_ref": 65, "n_ref_all": 92, "n_fig": 12, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 305, "n_element_tab_1": 57, "formula_len_all": 9104, "formula_len_all_1": 1333, "len_all": 170820, "len_all_1": 56444, "len_abs": 1256, "len_title": 108, "len_sents": 60970, "len_sents_1": 25939, "n_sents": 485, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1263, "title": "Convergence-Aware Multi-Fidelity Bayesian Optimization", "abs": "Multi-fidelity Bayesian Optimization (MFBO) has emerged as a powerful approach for optimizing expensive black-box functions by leveraging evaluations at different fidelity levels.\nHowever, existing MFBO methods often overlook the convergence behavior of the objective function as fidelity increases, leading to inefficient exploration and suboptimal performance. \nWe propose CAMO, a novel Convergence-Aware Multi-fidelity Optimization framework based on Fidelity Differential Equations (FiDEs). \nCAMO explicitly captures the convergence behavior of the objective function, enabling more efficient optimization. We introduce two tractable forms of CAMO: an integral Automatic Relevance Determination (ARD) kernel and a data-driven Deep Kernel. Theoretical analysis demonstrates that CAMO with the integral ARD kernel achieves a tighter regret bound compared to state-of-the-art methods. Our empirical evaluation on synthetic benchmarks and real-world engineering design problems shows that CAMO consistently outperforms existing MFBO algorithms in optimization efficiency and solution quality, with up to 4x improvement in optimal solution.\nThis work establishes a foundation for tractable convergence-aware MFBO and opens up new avenues for research in this area.", "keywords": [ "Bayesian Optimization", "Multi-Fidelity Bayesian Optimization", "Gaussian process", "dynamic systems" ], "rating_list": [ 8, 8, 6, 3 ], "soundness_list": [ 4, 3, 4, 2 ], "presentation_list": [ 4, 4, 4, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "Idygh9MX0N", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "Chen Zhang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Hao Duong Le", "gender": "Male", "institution": "National Taiwan University of Science and Technology", "country": "", "position": "Undergrad student" }, { "name": "Xin Xia", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 26, "n_ref": 77, "n_ref_all": 119, "n_fig": 6, "n_tab": 9, "L_tab": 3380, "n_element_tab": 346, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 908, "n_element_tab_1": 57, "formula_len_all": 447, "formula_len_all_1": 412, "len_all": 280495, "len_all_1": 56878, "len_abs": 6118, "len_title": 104, "len_sents": 105797, "len_sents_1": 26199, "n_sents": 759, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1166, "title": "Multi-Agent Causal Discovery Using Large Language Models", "abs": "Large Language Models (LLMs) have demonstrated significant potential in causal discovery tasks by utilizing their vast expert knowledge from extensive text corpora. However, the multi-agent capabilities of LLMs in causal discovery remain underexplored. This paper introduces a general framework to investigate this potential. The first is the Meta Agents Model, which relies exclusively on reasoning and discussions among LLM agents to conduct causal discovery. The second is the Coding Agents Model, which leverages the agents\u2019 ability to plan, write, and execute code, utilizing advanced statistical libraries for causal discovery. The third is the Hybrid Model, which integrates both the Meta Agents Model and Coding Agents Model approaches, combining the statistical analysis and reasoning skills of multiple agents. Our proposed framework shows promising results by effectively utilizing LLMs\u2019 expert knowledge, reasoning capabilities, multi-agent cooperation, and statistical causal methods. By exploring the multi-agent potential of LLMs, we aim to establish a foundation for further research in utilizing LLMs multi-agent for solving causal-related problems.", "keywords": [ "Causal Inference", "Causal Discovery", "Natural Language Process", "Multi-agent" ], "rating_list": [ 5, 3, 5, 3, 1 ], "soundness_list": [ 3, 3, 3, 2, 1 ], "presentation_list": [ 2, 2, 3, 3, 1 ], "contribution_list": [ 2, 2, 2, 1, 2 ], "confidence_list": [ 3, 4, 3, 4, 3 ] }, { "paper_id": "IdKkm91BzB", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Motoki Omura", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Takayuki Osa", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Associate Professor" }, { "name": "Tatsuya Harada", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Team Leader" }, { "name": "YUSUKE Mukuta", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Lecturer" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 22, "n_ref": 58, "n_ref_all": 82, "n_fig": 14, "n_tab": 3, "L_tab": 764, "n_element_tab": 13, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 870, "n_element_tab_1": 64, "formula_len_all": 707, "formula_len_all_1": 632, "len_all": 104056, "len_all_1": 60600, "len_abs": 1043, "len_title": 110, "len_sents": 33827, "len_sents_1": 27365, "n_sents": 267, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1048, "title": "Annealed Implicit Q-learning in Online Reinforcement Learning", "abs": "In continuous action online reinforcement learning, actor-critic methods are predominantly used.\nHowever, compared to Q-learning-based discrete action algorithms that model the optimal Q-value, continuous action algorithms that model the Q-value for the current policy and perform policy improvement solely through policy updates suffer from low sample efficiency. \nThis study investigates whether an algorithm that implicitly estimates the optimal Q-value, typically used in offline RL, is also effective in online RL. It is demonstrated that a loss function aimed at achieving optimality distorts the distribution of Q-values, leading to overestimation bias, and that this distortion and bias increase as learning progresses. To address this issue, we propose a simple algorithm that anneals optimality. Our method significantly outperforms widely used methods such as SAC and TD3 in online DM Control tasks. Additionally, we demonstrate that annealing improves performance and enhances robustness to the hyperparameter related to the optimality.", "keywords": [ "online reinforcement learning", "q-learning", "sample efficiency" ], "rating_list": [ 6, 6, 3, 3 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "IdAyXxBud7", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Han Sun", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Ismail Nejjar", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Olga Fink", "gender": "Female", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Assistant Professor" }, { "name": "Rui Gong", "gender": "Male", "institution": "Amazon", "country": "LU", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 53, "n_ref": 78, "n_ref_all": 99, "n_fig": 7, "n_tab": 15, "L_tab": 7660, "n_element_tab": 875, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 816, "n_element_tab_1": 87, "formula_len_all": 613, "formula_len_all_1": 315, "len_all": 192298, "len_all_1": 71697, "len_abs": 2333, "len_title": 127, "len_sents": 48129, "len_sents_1": 35570, "n_sents": 312, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1635, "title": "DynAlign: Unsupervised Dynamic Taxonomy Alignment for Cross-Domain Segmentation", "abs": "Current unsupervised domain adaptation (UDA) methods for semantic segmentation typically assume identical class labels between the source and target domains. This assumption ignores the label-level domain gap, which is common in real-world scenarios, and limits their ability to identify finer-grained or novel categories without requiring extensive manual annotation.\nA promising direction to address this limitation lies in recent advancements in foundation models, which exhibit strong generalization abilities due to their rich prior knowledge. However, these models often struggle with domain-specific nuances and underrepresented fine-grained categories.\nTo address these challenges, we introduce DynAlign, a two-stage framework that integrates UDA with foundation models to bridge both the image-level and label-level domain gaps. Our approach leverages prior semantic knowledge to align source categories with target categories that can be novel, more fine-grained, or named differently. (e.g., vehicle to car, truck, bus). Foundation models are then employed for precise segmentation and category reassignment. To further enhance accuracy, we propose a knowledge fusion approach that dynamically adapts to varying scene contexts. DynAlign generates accurate predictions in a new target label space without requiring any manual annotations, allowing seamless adaptation to new taxonomies through either model retraining or direct inference.\nExperiments on the GTA $\\rightarrow$ IDD and GTA$\\rightarrow$ Mapillary benchmarks validate the effectiveness of our approach, achieving a significant improvement over existing methods.", "keywords": [ "unsupervised domain adaptation\uff0csemantic segmentation\uff0ctaxonomy adaptation\uff0cfoundational models" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 4, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 4 ], "confidence_list": [ 2, 4, 5, 4, 4 ] }, { "paper_id": "Id2JMVSQHZ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 14, "author_info_list": [ { "name": "Antonio Bonafonte", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Aravind Varier", "gender": "unknown", "institution": "Amazon", "country": "GB", "position": "Researcher" }, { "name": "Ariya Rastrow", "gender": "Male", "institution": "Amazon", "country": "", "position": "Principal Researcher" }, { "name": "Biel Tura Vecino", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Constantinos Papayiannis", "gender": "Male", "institution": "Amazon", "country": "DE", "position": "Researcher" }, { "name": "Grant Strimel", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Principal Researcher" }, { "name": "Ivan Valles-Perez", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "LEIF R\u00c4DEL", "gender": "unknown", "institution": "Amazon", "country": "DE", "position": "Researcher" }, { "name": "Michael Owen", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Roberto Barra-Chicote", "gender": "Male", "institution": "Amazon", "country": "GB", "position": "Principal Researcher" }, { "name": "Subhadeep Maji", "gender": "Male", "institution": "Amazon Alexa ASR", "country": "", "position": "Researcher" }, { "name": "Trevor Wood", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Volker Leutnant", "gender": "Male", "institution": "Amazon", "country": "DE", "position": "Researcher" }, { "name": "Oluwaseyi Feyisetan", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 35, "n_ref": 96, "n_ref_all": 120, "n_fig": 8, "n_tab": 4, "L_tab": 1679, "n_element_tab": 243, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1685, "n_element_tab_1": 179, "formula_len_all": 897, "formula_len_all_1": 504, "len_all": 179838, "len_all_1": 73700, "len_abs": 1264, "len_title": 129, "len_sents": 56008, "len_sents_1": 34200, "n_sents": 392, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1267, "title": "Universal Semantic Disentangled Privacy-preserving Speech Representation Learning", "abs": "The use of audio recordings of human speech to train LLMs poses privacy concerns due to these models' potential to generate outputs that closely resemble artifacts in the training data. In this study, we propose a speaker privacy-preserving representation learning method through the Universal Speech Codec (USC), a computationally efficient encoder-decoder model that disentangles speech into: (i) privacy-preserving semantically rich representations, capturing content and speech paralinguistics, and (ii) residual acoustic and speaker representations that enables high-fidelity reconstruction. Extensive evaluations presented show that USC's semantic representation preserves content, prosody, and sentiment, while removing potentially identifiable speaker attributes. Combining both representations, USC achieves state-of-the-art speech reconstruction. Additionally, we introduce an evaluation methodology for measuring privacy-preserving properties, aligning with perceptual tests. We compare USC against other codecs in the literature and demonstrate its effectiveness on privacy-preserving representation learning, illustrating the trade-offs of speaker anonymization, paralinguistics retention and content preservation in the learned semantic representations.", "keywords": [ "speech", "tokenization", "disentanglement", "large-language models", "LLM", "privacy", "secure" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 4, 2, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "IcovaKGyMp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Ming Jiang", "gender": "Female", "institution": "Indiana University Indianapolis", "country": "US", "position": "Assistant Professor" }, { "name": "Shuying Xu", "gender": "unknown", "institution": "Tongji University", "country": "CN", "position": "PhD student" }, { "name": "Junjie Hu", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 38, "n_ref": 72, "n_ref_all": 99, "n_fig": 6, "n_tab": 7, "L_tab": 3629, "n_element_tab": 276, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 880, "n_element_tab_1": 68, "formula_len_all": 332, "formula_len_all_1": 332, "len_all": 162305, "len_all_1": 66548, "len_abs": 1508, "len_title": 106, "len_sents": 53287, "len_sents_1": 32829, "n_sents": 380, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1522, "title": "Large Language Models Are Active Critics in NLG Evaluation", "abs": "The conventional paradigm of using large language models (LLMs) for evaluating natural language generation (NLG) systems typically relies on two key inputs: (1) a clear definition of the NLG task to be evaluated and (2) a list of pre-defined evaluation criteria. This process treats LLMs as ''passive critics,'' strictly following human-defined criteria for evaluation. However, as new NLG tasks emerge, the criteria for assessing text quality can vary greatly. Consequently, these rigid evaluation methods struggle to adapt to diverse NLG tasks without extensive prompt engineering customized for each specific task. To address this limitation, we introduce Active-Critic, a novel LLM-based NLG evaluation protocol that enables LLMs to function as ''active critics.'' Specifically, our protocol comprises two key stages. In the first stage, the LLM is instructed to infer the target NLG task and establish relevant evaluation criteria from the data. Building on this self-inferred information, the second stage dynamically optimizes the prompt to guide the LLM toward more human-aligned scoring decisions, while also generating detailed explanations to justify its evaluations. Experiments across four NLG evaluation tasks show that our approach achieves stronger alignment with human judgments than state-of-the-art evaluation methods. Our comprehensive analysis further highlights the effectiveness and explainability of Active-Critic with only a small amount of labeled data. We will share our code and data on GitHub.", "keywords": [ "Large Langaugel Models\uff0c Natural Language Generation\uff0c Evaluation\uff0cPrompt Optimization\uff0c Explanable Evaluation" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "IcYDRzcccP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Haesoo Choo", "gender": "Female", "institution": "Pukyong National University", "country": "", "position": "Undergrad student" }, { "name": "Junghwan Kim", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Associate Professor" }, { "name": "Kyeongbo Kong", "gender": "Male", "institution": "Pusan National University", "country": "KR", "position": "Assistant Professor" }, { "name": "Oh-joon Kwon", "gender": "Male", "institution": "DMStudio,.Ltd.", "country": "KR", "position": "CEO" }, { "name": "Park Heemoon", "gender": "Male", "institution": "Pukyong National University", "country": "KR", "position": "PhD student" }, { "name": "Seong-Hun Jeong", "gender": "Male", "institution": "Pusan National University", "country": "KR", "position": "Researcher" }, { "name": "Inhwan Jin", "gender": "Male", "institution": "Pukyong National University", "country": "KP", "position": "Undergrad student" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 43, "n_ref": 149, "n_ref_all": 168, "n_fig": 9, "n_tab": 2, "L_tab": 461, "n_element_tab": 80, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 384, "n_element_tab_1": 54, "formula_len_all": 281, "formula_len_all_1": 282, "len_all": 133363, "len_all_1": 59127, "len_abs": 1867, "len_title": 124, "len_sents": 39015, "len_sents_1": 27383, "n_sents": 286, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1877, "title": "Optimizing 4D Gaussians for Dynamic Scene Video from Single Landscape Images", "abs": "To achieve realistic immersion in landscape images, fluids such as water and clouds need to move within the image while revealing new scenes from various camera perspectives. Recently, a field called dynamic scene video has emerged, which combines single image animation with 3D photography. These methods use pseudo 3D space, implicitly represented with Layered Depth Images (LDIs). LDIs separate a single image into depth-based layers, which enables elements like water and clouds to move within the image while revealing new scenes from different camera perspectives. However, as landscapes typically consist of continuous elements, including fluids, the representation of a 3D space separates a landscape image into discrete layers, and it can lead to diminished depth perception and potential distortions depending on camera movement. Furthermore, due to its implicit modeling of 3D space, the output may be limited to videos in the 2D domain, potentially reducing their versatility. In this paper, we propose representing a complete 3D space for dynamic scene video by modeling explicit representations, specifically 4D Gaussians, from a single image. The framework is focused on optimizing 3D Gaussians by generating multi-view images from a single image and creating 3D motion to optimize 4D Gaussians. The most important part of proposed framework is consistent 3D motion estimation, which estimates common motion among multi-view images to bring the motion in 3D space closer to actual motions. As far as we know, this is the first attempt that considers animation while representing a complete 3D space from a single landscape image. Our model demonstrates the ability to provide realistic immersion in various landscape images through diverse experiments and metrics. Extensive experimental results are https://anonymous.4open.science/r/ICLR_3D_MOM-7B9E/README.md.", "keywords": [ "Dynamic Scene Video", "4D Gaussian" ], "rating_list": [ 6, 8, 6, 3 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "IcVSKhVpKu", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Sayantan Dasgupta", "gender": "Male", "institution": "University of Melbourne", "country": "", "position": "PhD student" }, { "name": "Trevor Cohn", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 48, "n_ref": 91, "n_ref_all": 116, "n_fig": 4, "n_tab": 8, "L_tab": 2557, "n_element_tab": 327, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 2122, "n_element_tab_1": 204, "formula_len_all": 616, "formula_len_all_1": 621, "len_all": 144532, "len_all_1": 70317, "len_abs": 1050, "len_title": 115, "len_sents": 39149, "len_sents_1": 31862, "n_sents": 300, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1056, "title": "Improving Sequence Level Distillation through Hidden State Matching", "abs": "Hidden State Matching is a prominent technique in the knowledge distillation of language models. Most existing methods follow DistilBERT in using a cosine loss to encourage similarity between the student and the teacher's hidden states. However, the cosine loss restricts the architecture and dimensionality of the student, thereby severely limiting the compression ratio. We present a different technique using Centered Kernel Alignment (CKA) to match hidden states of different dimensionality, allowing for smaller students and higher compression ratios. We show the efficacy of our method using encoder--decoder (BART, mBART \\& T5) and encoder-only (BERT) architectures across a range of tasks from classification to summarization and translation. Our technique is competitive with the current state-of-the-art distillation methods at comparable compression rates and does not require already pretrained student models. It can scale to students smaller than the current methods, is no slower in training and inference, and is considerably more flexible.", "keywords": [ "Knowledge Distillation", "Centered Kernel Alignment", "BART", "mBART", "T5" ], "rating_list": [ 6, 8, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "IcPkW3QNW2", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Andreas Geiger", "gender": "Male", "institution": "University of Tuebingen", "country": "DE", "position": "Professor" }, { "name": "Daniel Barath", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Fangjinhua Wang", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Haofei Xu", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Hermann Blum", "gender": "unknown", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "Assistant Professor" }, { "name": "Marc Pollefeys", "gender": "Male", "institution": "Microsoft", "country": "CH", "position": "Director" }, { "name": "Songyou Peng", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 61, "n_ref": 144, "n_ref_all": 164, "n_fig": 8, "n_tab": 5, "L_tab": 1376, "n_element_tab": 167, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1154, "n_element_tab_1": 147, "formula_len_all": 407, "formula_len_all_1": 356, "len_all": 133633, "len_all_1": 64839, "len_abs": 256, "len_title": 149, "len_sents": 32683, "len_sents_1": 30200, "n_sents": 227, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 51, "L_abs": 1098, "title": "DepthSplat: Connecting Gaussian Splatting and Depth", "abs": "Gaussian splatting and single/multi-view depth estimation are typically studied in isolation. In this paper, we present DepthSplat to connect Gaussian splatting and depth estimation and study their interactions. More specifically, we first contribute a robust multi-view depth model by leveraging pre-trained monocular depth features, leading to high-quality feed-forward 3D Gaussian splatting reconstructions. We also show that Gaussian splatting can serve as an unsupervised pre-training objective for learning powerful depth models from large-scale unlabelled datasets. We validate the synergy between Gaussian splatting and depth estimation through extensive ablation and cross-task transfer experiments. Our DepthSplat achieves state-of-the-art performance on ScanNet, RealEstate10K and DL3DV datasets in terms of both depth estimation and novel view synthesis, demonstrating the mutual benefits of connecting both tasks. We invite the readers to view our supplementary video for feed-forward reconstruction results of large-scale or 360 scenes. Our code and models will be publicly available.", "keywords": [ "Gaussian Splatting", "Multi-View Depth", "Monocular Depth" ], "rating_list": [ 5, 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "IcNzKiB8CP", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Ahmed Aloui", "gender": "Male", "institution": "Duke University, Duke University", "country": "", "position": "PhD student" }, { "name": "Juncheng Dong", "gender": "unknown", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Vahid Tarokh", "gender": "Male", "institution": "Duke University", "country": "", "position": "Professor" }, { "name": "Zihao Wu", "gender": "Male", "institution": "Duke University", "country": "US", "position": "MS student" } ], "n_formula": 26, "n_formula_1": 24, "n_ref_uni": 24, "n_ref": 50, "n_ref_all": 65, "n_fig": 8, "n_tab": 4, "L_tab": 3640, "n_element_tab": 455, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1942, "n_element_tab_1": 282, "formula_len_all": 1936, "formula_len_all_1": 1783, "len_all": 101418, "len_all_1": 63774, "len_abs": 1230, "len_title": 127, "len_sents": 31035, "len_sents_1": 26151, "n_sents": 235, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1235, "title": "TELEPORTATION WITH NULL SPACE GRADIENT PROJECTION FOR OPTIMIZATION ACCELERATION", "abs": "Optimization techniques have become increasingly critical due to the ever-growing model complexity and data scale. In particular, teleportation has emerged as a promising approach, which accelerates convergence of gradient descent-based methods by navigating within the loss invariant level set to identify parameters with advantageous geometric properties. Existing teleportation algorithms have primarily demonstrated their effectiveness in optimizing Multi-Layer Perceptrons (MLPs), but their extension to more advanced architectures, such as Convolutional Neural Networks (CNNs) and Transformers, remains challenging. Moreover, they often impose significant computational demands, limiting their applicability to complex architectures. To this end, we introduce an algorithm that projects the gradient of the teleportation objective function onto the input null space, effectively preserving the teleportation within the loss invariant level set and reducing computational cost. Our approach is readily generalizable from MLPs to CNNs, transformers, and potentially other advanced architectures. We validate the effectiveness of our algorithm across various benchmark datasets and optimizers, demonstrating its broad applicability.", "keywords": [ "Optimization", "Teleportation", "Gradient Projection" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "IcMfCFPdd2", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Hideaki Iiduka", "gender": "Male", "institution": "Meiji University", "country": "JP", "position": "Full Professor" }, { "name": "Hinata Harada", "gender": "Male", "institution": "Meiji University", "country": "JP", "position": "Undergrad student" } ], "n_formula": 160, "n_formula_1": 46, "n_ref_uni": 29, "n_ref": 87, "n_ref_all": 155, "n_fig": 6, "n_tab": 6, "L_tab": 354, "n_element_tab": 10, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 377, "n_element_tab_1": 34, "formula_len_all": 22806, "formula_len_all_1": 6189, "len_all": 165239, "len_all_1": 62338, "len_abs": 857, "len_title": 158, "len_sents": 38039, "len_sents_1": 23725, "n_sents": 363, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 109, "L_abs": 858, "title": "Convergence of Sharpness-Aware Minimization Algorithms using Increasing Batch Size and Decaying Learning Rate", "abs": "The sharpness-aware minimization (SAM) algorithm and its variants, including gap guided SAM (GSAM), have been successful at improving the generalization capability of deep neural network models by finding flat local minima of the empirical loss in training. Meanwhile, it has been shown theoretically and practically that increasing the batch size or decaying the learning rate avoids sharp local minima of the empirical loss. In this paper, we consider the GSAM algorithm with increasing batch sizes or decaying learning rates, such as cosine annealing or linear learning rate, and theoretically show its convergence. Moreover, we numerically compare SAM (GSAM) with and without an increasing batch size and conclude that using an increasing batch size or decaying learning rate finds flatter local minima than using a constant batch size and learning rate.", "keywords": [ "batch size", "deep learning", "gap guided sharpness-aware minimization", "learning rate", "sharpness-aware minimization" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "IcHHjgdb0o", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Alicia Lozano-Diez", "gender": "Female", "institution": "Universidad Aut\u00f3noma de Madrid", "country": "ES", "position": "Assistant Professor" }, { "name": "Gy\u00f6rgy Cserey", "gender": "Male", "institution": "Pazmany Peter Catholic University", "country": "HU", "position": "Full Professor" }, { "name": "Tamas Endrei", "gender": "unknown", "institution": "Universidad Aut\u00f3noma de Madrid", "country": "ES", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 43, "n_ref": 64, "n_ref_all": 74, "n_fig": 5, "n_tab": 2, "L_tab": 1418, "n_element_tab": 195, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 1418, "n_element_tab_1": 195, "formula_len_all": 97, "formula_len_all_1": 97, "len_all": 119855, "len_all_1": 57762, "len_abs": 1128, "len_title": 137, "len_sents": 31975, "len_sents_1": 21978, "n_sents": 219, "n_sents_1": 145, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1131, "title": "PASRL: Stabilising Reinforcement Learning with Past Action-State Representation Learning", "abs": "Although deep reinforcement learning (DRL) deals with sequential decision making problems, temporal information representation is absent from state-of-the-art actor-critic algorithms. The reliance on a single observation vector, representing information from only one time step, combined with densely connected neural networks, causes instability and oscillations in action smoothness. Therefore many applied DRL robotics control methods employ various reward shaping, low-pass filter and traditional controller-based methods to mitigate this effect. However, the interactions of these different parts hinders the performance of the original goal for the RL algorithm. In this paper we present a reinforcement learning algorithm extended with past action-state representation learning (PASRL), which allows for the end-to-end training of RL-based control methods without the need for common heuristics. PASRL is evaluated on the MuJoCo benchmark, showing smoother actions that preserve exploration, eliminate the need for extensive hyperparameter tuning, and provide a simple and efficient solution for enhancing action smoothness.", "keywords": [ "Reinforcement learning", "action smoothness", "recurrent neural networks" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "IbCvnpJ4py", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Deyu Zou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Kai Liu", "gender": "Male", "institution": "SES AI", "country": "US", "position": "Principal Researcher" }, { "name": "Nima Shoghi", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Pan Li", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Shikun Liu", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Victor Fung", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 9, "n_ref_uni": 80, "n_ref": 125, "n_ref_all": 167, "n_fig": 6, "n_tab": 18, "L_tab": 61406, "n_element_tab": 6135, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 5957, "n_element_tab_1": 561, "formula_len_all": 1033, "formula_len_all_1": 555, "len_all": 374015, "len_all_1": 80152, "len_abs": 5748, "len_title": 128, "len_sents": 53632, "len_sents_1": 33545, "n_sents": 375, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1396, "title": "RoFt-Mol: Benchmarking Robust Fine-tuning with Molecular Graph Foundation Models", "abs": "In the era of foundation models, fine-tuning pre-trained models for specific downstream tasks has become crucial. This drives the need for robust fine-tuning methods to address challenges such as model overfitting and sparse labeling. Molecular graph foundation models (MGFMs) face unique difficulties that complicate fine-tuning. These models are limited by smaller pre-training datasets and more severe data scarcity for downstream tasks, both of which require enhanced model generalization. Moreover, MGFMs must accommodate diverse pre-training objectives, including both regression and classification tasks. To better understand and improve fine-tuning techniques under these conditions, we classify eight fine-tuning methods into three mechanisms: weight-based fine-tuning, representation-based fine-tuning, and partial fine-tuning. We benchmark these methods on downstream regression and classification tasks across both supervised and self-supervised pre-trained models in diverse labeling settings. This extensive evaluation provides valuable insights and informs the design of a refined robust fine-tuning method, DWiSE-FT. This approach combines the strengths of simple post-hoc weight interpolation with more complex weight ensemble fine-tuning methods, delivering improved performance across both task types while maintaining the ease of use inherent in post-hoc weight interpolation.", "keywords": [ "Molecular representation learning", "Fine-tuning" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "Ian00SaFHg", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Amanda Prorok", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Carl Henrik Ek", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "Associate Professor" }, { "name": "Jasmine Bayrooti", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 8, "n_ref_uni": 68, "n_ref": 126, "n_ref_all": 143, "n_fig": 7, "n_tab": 2, "L_tab": 821, "n_element_tab": 90, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1213, "formula_len_all_1": 554, "len_all": 177174, "len_all_1": 69979, "len_abs": 1142, "len_title": 82, "len_sents": 56359, "len_sents_1": 35957, "n_sents": 410, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1150, "title": "Efficient Model-Based Reinforcement Learning Through Optimistic Thompson Sampling", "abs": "Learning complex robot behavior through interactions with the environment necessitates principled exploration. Effective strategies should prioritize exploring regions of the state-action space that maximize rewards, with optimistic exploration emerging as a promising direction aligned with this idea and enabling sample-efficient reinforcement learning. However, existing methods overlook a crucial aspect: the need for optimism to be informed by a belief connecting the reward and state. To address this, we propose a practical, theoretically grounded approach to optimistic exploration based on Thompson sampling. Our approach is the first that allows for reasoning about _joint_ uncertainty over transitions and rewards for optimistic exploration. We apply our method on a set of MuJoCo and VMAS continuous control tasks. Our experiments demonstrate that optimistic exploration significantly accelerates learning in environments with sparse rewards, action penalties, and difficult-to-explore regions. Furthermore, we provide insights into when optimism is beneficial and emphasize the critical role of model uncertainty in guiding exploration.", "keywords": [ "reinforcement learning", "model-based reinforcement learning", "optimistic exploration" ], "rating_list": [ 5, 8, 3, 8 ], "soundness_list": [ 2, 4, 2, 2 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "IaHzYWSFYY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Hongru Wei", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" }, { "name": "Qingyuan Hu", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" }, { "name": "Yuqi Chen", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 6, "n_ref_uni": 44, "n_ref": 71, "n_ref_all": 85, "n_fig": 1, "n_tab": 5, "L_tab": 1075, "n_element_tab": 133, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 235, "n_element_tab_1": 35, "formula_len_all": 185, "formula_len_all_1": 699, "len_all": 132774, "len_all_1": 68163, "len_abs": 1512, "len_title": 152, "len_sents": 38181, "len_sents_1": 33191, "n_sents": 302, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1522, "title": "RootTracker: A Lightweight Framework to Trace Original Models of Fine-tuned LLMs in Black-Box Conditions", "abs": "Large Language Models (LLMs) demonstrate remarkable performance in various applications, yet their training demands extensive resources and time. Consequently, fine-tuning pre-trained LLMs has become a prevalent strategy for adapting these models to diverse downstream tasks, thereby reducing costs. Despite their benefits, LLMs have vulnerabilities, such as susceptibility to adversarial attacks, potential for jailbreaking, fairness issues, backdoor vulnerabilities, and the risk of generating inappropriate or harmful content. Since fine-tuned models inherit some characteristics from their original models, they may also inherit these issues and vulnerabilities. In this work, we propose a lightweight framework, RootTracker, specifically designed to trace the original models of fine-tuned LLMs. The core idea is to identify a set of prompts that can assess which pre-trained LLM a fine-tuned model most closely resembles. This process is conducted in a ''knockout tournament\" style, where the model is repeatedly tested against pairs of LLMs until the original pre-trained model is identified. To evaluate the effectiveness of our framework, we created 200 distinct fine-tuned models, derived from original models including GPT-Neo, GPT-2, TinyLlama, and Pythia. The results demonstrate that our framework accurately identified the original models for 85.7\\% of the fine-tuned versions. Therefore, we advocate for timely updates to model versions or deliberate obfuscation of model types when deploying large models.", "keywords": [ "Large language models", "Fine-tune", "Framework", "Black-box", "Fairness", "Safety" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "IaFAvyaBOc", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Abdulla Jasem Almansoori", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "PhD student" }, { "name": "Samuel Horv\u00e1th", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Assistant Professor" }, { "name": "Martin Takac", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Associate Professor" } ], "n_formula": 172, "n_formula_1": 31, "n_ref_uni": 41, "n_ref": 74, "n_ref_all": 111, "n_fig": 4, "n_tab": 3, "L_tab": 1283, "n_element_tab": 135, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 870, "n_element_tab_1": 52, "formula_len_all": 22177, "formula_len_all_1": 2928, "len_all": 265039, "len_all_1": 70826, "len_abs": 1802, "len_title": 117, "len_sents": 74447, "len_sents_1": 29011, "n_sents": 738, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1814, "title": "Collaborative and Efficient Personalization with Mixtures of Adaptors", "abs": "Non-iid data is prevalent in real-world federated learning problems. Data heterogeneity can come in different types in terms of distribution shifts. In this work, we are interested in the heterogeneity that comes from concept shifts, i.e., shifts in the prediction across clients. In particular, we consider multi-task learning, where we want the model to adapt to the task of the client. We propose a parameter-efficient framework to tackle this issue, where each client learns to mix between parameter-efficient adaptors according to its task. We use Low-Rank Adaptors (LoRAs) as the backbone and extend its concept to other types of layers. We call our framework Federated Low-Rank Adaptive Learning (FLoRAL). This framework is not an algorithm but rather a model parameterization for a multi-task learning objective, so it can work on top of any algorithm that optimizes this objective, which includes many algorithms from the literature. FLoRAL is memory-efficient, and clients are personalized with small states (e.g., one number per adaptor) as the adaptors themselves are federated. Hence, personalization is--in this sense--federated as well. Even though clients can personalize more freely by training an adaptor locally, we show that collaborative and efficient training of adaptors is possible and performs better. We also show that FLoRAL can outperform an ensemble of full models with optimal cluster assignment, which demonstrates the benefits of federated personalization and the robustness of FLoRAL to overfitting. We show promising experimental results on synthetic datasets, real-world federated multi-task problems such as MNIST, CIFAR-10, and CIFAR-100. We also provide a theoretical analysis of local SGD on a relaxed objective and discuss the effects of aggregation mismatch on convergence.", "keywords": [ "federated learning", "collaborative learning", "adaptors" ], "rating_list": [ 3, 3, 5, 1 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "Ia17iAtr0P", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Kenna Ashen", "gender": "unknown", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Xiaofeng Qian", "gender": "unknown", "institution": "Texas A&M University", "country": "US", "position": "Associate Professor" }, { "name": "Xiaoning Qian", "gender": "Male", "institution": "Texas A&M", "country": "", "position": "Full Professor" }, { "name": "Ziyu Xiang", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 6, "n_ref_uni": 32, "n_ref": 55, "n_ref_all": 81, "n_fig": 8, "n_tab": 5, "L_tab": 736, "n_element_tab": 54, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 605, "n_element_tab_1": 82, "formula_len_all": 647, "formula_len_all_1": 317, "len_all": 142753, "len_all_1": 61382, "len_abs": 1181, "len_title": 94, "len_sents": 56481, "len_sents_1": 30501, "n_sents": 324, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1188, "title": "Physics-constrained Graph Symbolic Regression", "abs": "As data-driven scientific discovery increasingly demands explainable over \u2018black-box\u2019 machine learning (ML) methods, Symbolic Regression (SR) that derives analytical expressions can help identify key functional dependencies in complex systems. However, traditional SR methods often suffer from (a) inefficient exploration due to their inability to compress the search space of equivalent expressions, and (b) non-physical solutions that violate fundamental physics constraints. We here introduce a symmetric invariant representation of candidate analytical expressions using a Symbolic Graph (SG), on which the Symbolic Graph Neural Network (SGNN) encodes operators, symmetries, constraints and constant fitting knowledge. We further develop reinforcement learning (RL) algorithms with Monte-Carlo Tree Search (MCTS) on our SGNN for SR. Such a physics-constrained graph symbolic regression (PCGSR) method effectively compresses the search space for efficient SR. Experiments on synthetic and real-world scientific datasets demonstrate the efficiency and accuracy of our PCGSR in discovering underlying expressions and adhering to physical laws, yielding physically meaningful solutions.", "keywords": [ "Symbolic Regression", "Physics-constrained", "Graph Neural Network", "Reinforcement Learning", "Monte-Carlo Tree Search", "Expression Tree", "Automated Feature Engineering", "Symbolic Graph" ], "rating_list": [ 3, 8, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "IZuwA3hken", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Bo Jiang", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "James Flemings", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Murali Annavaram", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Full Professor" }, { "name": "Wanrong Zhang", "gender": "Female", "institution": "Tiktok", "country": "US", "position": "Researcher" }, { "name": "Zafar Takhirov", "gender": "unknown", "institution": "TikTok Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 34, "n_ref": 46, "n_ref_all": 63, "n_fig": 8, "n_tab": 3, "L_tab": 1616, "n_element_tab": 205, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 2131, "n_element_tab_1": 348, "formula_len_all": 624, "formula_len_all_1": 533, "len_all": 133282, "len_all_1": 71833, "len_abs": 1534, "len_title": 116, "len_sents": 36134, "len_sents_1": 31710, "n_sents": 237, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1564, "title": "Characterizing Context Influence and Hallucination in Summarization", "abs": "Although Large Language Models (LLMs) have achieved remarkable performance in numerous downstream tasks, their ubiquity has raised two significant concerns. One is that LLMs can hallucinate by generating content that contradicts relevant contextual information; the other is that LLMs can inadvertently leak private information due to input regurgitation. Many prior works have extensively studied each concern independently, but none have investigated them simultaneously. Furthermore, auditing the influence of provided context during open-ended generation with a privacy emphasis is understudied. To this end, we comprehensively characterize the influence and hallucination of contextual information during summarization. We introduce a definition for context influence and Context-Influence Decoding (CID), and then we show that amplifying the context (by factoring out prior knowledge) and the context being out of distribution with respect to prior knowledge increases the context's influence on an LLM. Moreover, we show that context influence gives a lower bound of the private information leakage of CID. We corroborate our analytical findings with experimental evaluations that show improving the F1 ROGUE-L score on CNN-DM for LLaMA 3 by $\\textbf{10}$% over regular decoding also leads to $\\textbf{1.5x}$ more influence by the context. Moreover, we empirically evaluate how context influence and hallucination are affected by (1) model capacity, (2) context size, (3) the length of the current response, and (4) different token $n$-grams of the context.", "keywords": [ "Influence", "Hallucinations", "Language Models" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "IZjBfdVRB0", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Aochuan Chen", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Fugee Tsung", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Jia Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Jiashun Cheng", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yu Li", "gender": "Male", "institution": "International Digital Economy Academy", "country": "CN", "position": "Principal Researcher" }, { "name": "Zijing Liu", "gender": "unknown", "institution": "International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Ziqi Gao", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 59, "n_ref": 86, "n_ref_all": 101, "n_fig": 3, "n_tab": 6, "L_tab": 3991, "n_element_tab": 388, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1285, "n_element_tab_1": 126, "formula_len_all": 1533, "formula_len_all_1": 1540, "len_all": 155774, "len_all_1": 74876, "len_abs": 912, "len_title": 104, "len_sents": 34551, "len_sents_1": 31687, "n_sents": 303, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 997, "title": "Parameter-Efficient Fine-Tuning via Circular Convolution", "abs": "Low-Rank Adaptation (LoRA) has gained popularity for fine-tuning large foundation models, leveraging low-rank matrices $\\mathbf{A}$ and $\\mathbf{B}$ to represent weight changes (i.e., $\\Delta \\mathbf{W} = \\mathbf{B} \\mathbf{A}$). This method reduces trainable parameters and mitigates heavy memory consumption associated with full delta matrices by sequentially multiplying $\\mathbf{A}$ and $\\mathbf{B}$ with the activation. Despite its success, the intrinsic low-rank characteristic may limit its performance. Although several variants have been proposed to address this issue, they often overlook the crucial computational and memory efficiency brought by LoRA. In this paper, we propose Circular Convolution Adaptation (C$^3$A), which not only achieves high-rank adaptation with enhanced performance but also excels in both computational power and memory utilization. Extensive experiments demonstrate that C$^3$A consistently outperforms LoRA and its variants across various fine-tuning tasks.", "keywords": [ "transfer learning", "circular convolution", "adaptation", "efficiency" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "IZiKBis0AA", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Adam Bess", "gender": "unknown", "institution": "Louisiana State University", "country": "US", "position": "Researcher" }, { "name": "Amit Ranjan", "gender": "unknown", "institution": "Louisiana State University", "country": "", "position": "Postdoc" }, { "name": "Chris Alvin", "gender": "Male", "institution": "Furman University", "country": "US", "position": "Associate Professor" }, { "name": "Magesh Rajasekaran", "gender": "Male", "institution": "Louisiana State University", "country": "US", "position": "PhD student" }, { "name": "Md Saiful Islam Sajol", "gender": "Male", "institution": "Louisiana State University", "country": "US", "position": "PhD student" }, { "name": "Sean Rowland", "gender": "Male", "institution": "Louisiana State University", "country": "US", "position": "Researcher" }, { "name": "Supratick Mukhopadhyay", "gender": "Not Specified", "institution": "Louisiana State University", "country": "US", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 16, "n_ref": 20, "n_ref_all": 31, "n_fig": 4, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 1319, "n_element_tab_1": 42, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 92326, "len_all_1": 64477, "len_abs": 1390, "len_title": 104, "len_sents": 37715, "len_sents_1": 33108, "n_sents": 234, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1401, "title": "AI Derivation and Exploration of Antibiotic Class Spaces", "abs": "This paper presents a novel approach to fragment-based antibiotic drug design design. We introduce a tool called FILTER, which uses chemical structure data, pathway information, and protein targets to predict pharmacokinetic properties of existing and novel drugs. We report on three distinct experiments utilizing FILTER. The first experiment is an in silico analysis that recreates the historical discovery of penicillin derivatives, validating our approach against known outcomes. The second experiment explores the combination of functional groups from different antibiotic classes to create molecules with multiple mechanisms of action. We refer to this approach as hybridization as all synthesized molecules are composed of fragments from both classes. Our final experiment is forward-looking as it explores new chemical spaces to build a library of promising compounds for further antibiotic development. Throughout all these experiments, FILTER serves as an indispensable oracle, predicting physical properties and potential therapeutic efficacy of the new molecular architectures, aiming to accelerate the drug development process and address the challenge of antibiotic resistance. Our approach represents an ongoing, significant shift from traditional drug discovery methods, emphasizing the role of innovative technologies in combating the urgent global threat of antimicrobial resistance.", "keywords": [ "fragment-based drug design", "antibiotic resistance", "pharmacokinetics", "hybrid antibiotics", "in silico analysis", "retrosynthesis", "chemical space exploration", "machine learning", "antibiotic discovery", "protein targets" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 1, 1, 1 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "IZbthMfqad", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Tomoharu Iwata", "gender": "Male", "institution": "NTT", "country": "", "position": "Researcher" }, { "name": "Yuka Hashimoto", "gender": "unknown", "institution": "NTT", "country": "", "position": "Researcher" } ], "n_formula": 51, "n_formula_1": 28, "n_ref_uni": 35, "n_ref": 57, "n_ref_all": 75, "n_fig": 6, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7132, "formula_len_all_1": 2975, "len_all": 151425, "len_all_1": 77336, "len_abs": 1649, "len_title": 125, "len_sents": 43613, "len_sents_1": 29425, "n_sents": 428, "n_sents_1": 292, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 827, "title": "Deep Koopman-layered Model with Universal Property Based on Toeplitz Matrices", "abs": "We propose deep Koopman-layered models with learnable parameters in the form of Toeplitz matrices for analyzing the dynamics of time-series data.\nThe proposed model has both theoretical solidness and flexibility.\nBy virtue of the universal property of Toeplitz\nmatrices and the reproducing property underlined in the model, we can show its universality and the generalization property.\nIn addition, the flexibility of the proposed model enables the model to fit time-series data coming from nonautonomous dynamical systems.\nWhen training the model, we apply Krylov subspace methods for efficient computations.\nIn addition, the proposed model can be regarded as a neural ODE-based model.\nIn this sense, the proposed model establishes a new connection among Koopman operators, neural ODEs, and numerical linear algebraic methods.", "keywords": [ "Koopman operator", "Toeplitz matrix", "nonautonomous dynamical system" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 5, 2, 4 ] }, { "paper_id": "IZOeRDS6zU", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Abhinav Uppal", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "David Wang", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "Undergrad student" }, { "name": "Ian L. Jackson", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Teng Fei", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Virginia de Sa", "gender": "Female", "institution": "Halicioglu Data Science Institute, University of California, San Diego", "country": "US", "position": "Associate Director" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 16, "n_ref": 21, "n_ref_all": 49, "n_fig": 13, "n_tab": 2, "L_tab": 19, "n_element_tab": 2, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 1808, "n_element_tab_1": 69, "formula_len_all": 34, "formula_len_all_1": 0, "len_all": 84026, "len_all_1": 50663, "len_abs": 1291, "len_title": 122, "len_sents": 30536, "len_sents_1": 23410, "n_sents": 211, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 754, "title": "Perceptogram: Visual Reconstruction from EEG Using Image Generative Models", "abs": "In this work, we reconstruct viewed images from EEG recordings with state-of-the-art quantitative reconstruction performance using a linear decoder that maps the EEG to image latents. We choose latent diffusion guided by CLIP embedding as the primary method of image reconstruction as it is currently the most effective at capturing visual semantics. We also explore reconstruction results from a latent space of PCA and ICA components, which capture luminance and hue-related information from the EEG. The linear model provides interpretable EEG features relevant for differentiating general semantic categories of the images. We create spatiotemporal semantic maps that reflect the temporal evolution of class-relevant semantic information over time.", "keywords": [ "visual-evoked potentials", "electroencephalography", "EEG", "visual reconstruction", "visual representations", "spatiotemporal semantic map" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "IZDiRbVSVN", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Ivan Titov", "gender": "unknown", "institution": "University of Edinburgh", "country": "", "position": "Full Professor" }, { "name": "Masaru Isonuma", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Postdoc" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 49, "n_ref": 87, "n_ref_all": 118, "n_fig": 5, "n_tab": 9, "L_tab": 3758, "n_element_tab": 391, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1181, "n_element_tab_1": 192, "formula_len_all": 400, "formula_len_all_1": 363, "len_all": 171449, "len_all_1": 69376, "len_abs": 1733, "len_title": 117, "len_sents": 46032, "len_sents_1": 31809, "n_sents": 332, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1742, "title": "What's New in My Data? Novelty Exploration via Contrastive Generation", "abs": "Fine-tuning is widely used to adapt language models for specific goals, often leveraging real-world data such as patient records, customer-service interactions, or web content in languages not covered in pre-training.\nThese datasets are typically massive, noisy, and often confidential, making their direct inspection challenging.\nHowever, understanding them is essential for guiding model deployment and informing decisions about data cleaning or suppressing any harmful behaviors learned during fine-tuning.\nIn this study, we introduce the task of novelty discovery through generation, which aims to identify novel domains of a fine-tuning dataset by generating examples that illustrate these properties.\nOur approach \uff0d Contrastive Generative Exploration (CGE) \uff0d assumes no direct access to the data but instead relies on a pre-trained model and the same model after fine-tuning.\nBy contrasting the predictions of these two models, CGE can generate examples that highlight novel domains of the fine-tuning data.\nHowever, this simple approach may produce examples that are too similar to one another, failing to capture the full range of novel domains present in the dataset.\nWe address this by introducing an iterative version of CGE, where the previously generated examples are used to update the pre-trained model, and this updated model is then contrasted with the fully fine-tuned model to generate the next example, promoting diversity in the generated outputs.\nOur experiments demonstrate the effectiveness of CGE in detecting novel domains, such as toxic language, as well as new natural and programming languages.\nFurthermore, we show that CGE remains effective even when models are fine-tuned using differential privacy techniques.", "keywords": [ "dataset exploration", "novelty detection", "contrastive decoding" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "IZB8H50V1S", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Anindya Sarkar", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "US", "position": "PhD student" }, { "name": "Bengisu Guresti", "gender": "unknown", "institution": "Washington University, Saint Louis", "country": "US", "position": "PhD student" }, { "name": "Chongjie Zhang", "gender": "unknown", "institution": "Washington University, Saint Louis", "country": "", "position": "Associate Professor" }, { "name": "Luise Ge", "gender": "unknown", "institution": "Washington University, Saint Louis", "country": "US", "position": "PhD student" }, { "name": "Michael Lanier", "gender": "Not Specified", "institution": "Maryville University", "country": "US", "position": "Instructor" }, { "name": "Yevgeniy Vorobeychik", "gender": "Male", "institution": "Washington University, St. Louis", "country": "US", "position": "Full Professor" } ], "n_formula": 29, "n_formula_1": 19, "n_ref_uni": 28, "n_ref": 57, "n_ref_all": 70, "n_fig": 4, "n_tab": 6, "L_tab": 874, "n_element_tab": 84, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 874, "n_element_tab_1": 84, "formula_len_all": 3175, "formula_len_all_1": 1732, "len_all": 159844, "len_all_1": 71628, "len_abs": 1205, "len_title": 131, "len_sents": 51538, "len_sents_1": 32747, "n_sents": 508, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1226, "title": "Learning Policy Committees for Effective Personalization in MDPs with Diverse Tasks", "abs": "Many dynamic decision problems, such as robotic control, involve a series of tasks, many of which are unknown at training time. Typical approaches for these problems, such as multi-task and meta reinforcement learning, do not generalize well when the tasks are diverse. We propose a general framework to address this issue. In our framework, the goal is to learn a set of policies\u2014a policy committee\u2014such that at least one is near-optimal for most tasks that may be encountered at execution time. While we show that even a special case of this problem is inapproximable, we present two effective algorithmic approaches for it. The first of these yields provably approximation guarantees, albeit in small-dimensional settings (the best we can do due to inapproximability), whereas the second is a general and practical gradient-based approach. In addition, we provide provable sample complexity bounds for few-shot learning settings. Our experiments in personalized and multi-task RL settings using MuJoCo and Meta-World benchmarks show that the proposed approach outperforms state-of-the-art multi-task, meta-, and personalized RL baselines on training and test tasks, as well as in few-shot learning, often by a large margin.", "keywords": [ "Multi-task reinforcement learning", "meta reinforcement learning", "personalized reinforcement learning" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 2 ] }, { "paper_id": "IXyfbaGlps", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Christine Allen-Blanchette", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "Assistant Professor" }, { "name": "Felix O'Mahony", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Yulong Yang", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 19, "n_ref_uni": 48, "n_ref": 110, "n_ref_all": 134, "n_fig": 16, "n_tab": 5, "L_tab": 1173, "n_element_tab": 73, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 198, "n_element_tab_1": 20, "formula_len_all": 1831, "formula_len_all_1": 1049, "len_all": 154362, "len_all_1": 55845, "len_abs": 1286, "len_title": 90, "len_sents": 49432, "len_sents_1": 23844, "n_sents": 426, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 1292, "title": "Learning Color Equivariant Representations", "abs": "In this paper, we introduce group convolutional neural networks (GCNNs) equivariant to color variation. GCNNs have been designed for a variety of geometric transformations from 2D and 3D rotation groups, to semi-groups such as scale. Despite the improved interpretability, accuracy and generalizability of these architectures, GCNNs have seen limited application in the context of perceptual quantities. Notably, the recent CEConv network uses a GCNN to achieve equivariance to hue transformations by convolving input images with a hue rotated RGB filter. However, this approach leads to invalid RGB values which break equivariance and degrade performance. We resolve these issues with a lifting layer that transforms the input image directly, thereby circumventing the issue of invalid RGB values and improving equivariance error by over three orders of magnitude. Moreover, we extend the notion of color equivariance to include equivariance to saturation shift. Our hue-, saturation-, and color-equivariant networks achieve strong generalization to out-of-distribution perceptual variations and improved sample efficiency over conventional architectures. We demonstrate the utility of our approach on synthetic and real world datasets where we consistently outperform competitive baselines.", "keywords": [ "Equivariant Neural Network", "Geometric Deep Learning", "Group Convolution" ], "rating_list": [ 3, 5, 6, 8, 6 ], "soundness_list": [ 3, 4, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 4, 2 ] }, { "paper_id": "IXOoltTofP", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Cheng Wang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Dingning Liu", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "MS student" }, { "name": "Xinzhu Ma", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Postdoc" }, { "name": "Yuan Meng", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhihui Wang", "gender": "Female", "institution": "Dalian University of Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 37, "n_ref": 92, "n_ref_all": 122, "n_fig": 11, "n_tab": 2, "L_tab": 687, "n_element_tab": 79, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 687, "n_element_tab_1": 79, "formula_len_all": 153, "formula_len_all_1": 60, "len_all": 179048, "len_all_1": 68505, "len_abs": 1628, "len_title": 112, "len_sents": 35036, "len_sents_1": 31309, "n_sents": 287, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1643, "title": "3DAxisPrompt: Promoting the 3D Grounding and Reasoning in GPT-4o", "abs": "Multimodal Large Language Models (MLLMs) exhibit impressive capabilities across a variety of tasks, especially when equipped with carefully designed visual prompts. However, existing studies primarily focus on logical reasoning and visual understanding, while the capability of MLLMs to operate effectively in 3D vision remains an ongoing area of exploration.\nIn this paper, we introduce a novel visual prompting method called 3DAxisPrompt to elicit the 3D understanding capabilities of MLLMs in real-world scenes. More specifically, our method leverages the 3D coordinate axis and masks generated from the Segment Anything Model (SAM) to provide explicit geometric priors to MLLMs and then extend their impressive 2D grounding/reasoning ability to real-world 3D scenarios. Besides we also provide a thorough investigation of the potential visual prompting formats and conclude our findings to reveal the potential and limits of 3D understanding capabilities in GPT-4o. Finally, we build evaluation environments with four datasets, {\\it i.e.} ShapeNet, ScanNet, FMB, and nuScene datasets, covering various 3D tasks. Based on this, we conduct extensive quantitative and qualitative experiments, which demonstrate the effectiveness of the proposed method. Overall, our study reveals that GPT-4o, with the help of 3DAxisPrompt, can effectively perceive an object\u2019s 3D position in real-world scenarios. Nevertheless, a single prompt engineering approach does not consistently achieve the best outcomes for all 3D tasks. This study highlights the feasibility of leveraging MLLMs for 3D vision grounding/reasoning with prompt engineering techniques.", "keywords": [ "visual prompt", "3D grounding", "spatial casual reasoning", "spatial grounding", ".multimodal LLM" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 4, 1, 3, 2 ], "contribution_list": [ 3, 1, 1, 2 ], "confidence_list": [ 4, 4, 2, 5 ] }, { "paper_id": "IXGHSVBBCF", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Hamed Fayyaz", "gender": "Male", "institution": "CIS Department, University of Delaware", "country": "", "position": "PhD student" }, { "name": "Rahmatollah Beheshti", "gender": "unknown", "institution": "University of Delaware", "country": "US", "position": "Assistant Professor" }, { "name": "Raphael Poulain", "gender": "unknown", "institution": "University of Delaware", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 47, "n_ref": 72, "n_ref_all": 83, "n_fig": 2, "n_tab": 8, "L_tab": 1287, "n_element_tab": 61, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 75, "n_element_tab_1": 1, "formula_len_all": 382, "formula_len_all_1": 382, "len_all": 155658, "len_all_1": 56740, "len_abs": 1722, "len_title": 62, "len_sents": 45689, "len_sents_1": 29078, "n_sents": 308, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1734, "title": "Enabling Scalable Evaluation of Bias Patterns in Medical LLMs", "abs": "Large language models (LLMs) have shown impressive potential in helping with numerous medical challenges. Deploying LLMs in high-stakes applications such as medicine, however, brings in many concerns. One major area of concern relates to biased behaviors of LLMs in medical applications, leading to unfair treatment of individuals. To pave the way for the responsible and impactful deployment of Med LLMs, rigorous evaluation is a key prerequisite. Due to the huge complexity and variability of different medical scenarios, existing work in this domain has primarily relied on using manually crafted datasets for bias evaluation. In this study, we present a new method to scale up such bias evaluations by automatically generating test cases based on rigorous medical evidence. We specifically target the challenges of domain-specificity of bias characterization, hallucinating while generating the test cases, and various dependencies between the health outcomes and sensitive attributes. To that end, we offer new methods to address these challenges integrated with our generative pipeline. Specifically, we use medical knowledge graphs and medical ontologies; and customize general LLM evaluation frameworks in our method. Through a series of extensive experiments, we show that the test cases generated by our proposed method are reliable and can effectively reveal bias patterns in LLMs. Additionally, we publish a large bias evaluation dataset, which provides a comprehensive platform for testing and improving the fairness of clinical LLMs. A live demo of our application for vignette generation is available at https://vignette.streamlit.app. Our code is also available at https://anonymous.4open.science/r/vignette_llm-2853.", "keywords": [ "Medical LLMs", "Fairness Evaluation" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "IXFCPqFHMQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jia-Hong Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Jia-Tong Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Shao-Kui Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Song-Hai Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Tianqi Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 22, "n_ref": 46, "n_ref_all": 70, "n_fig": 10, "n_tab": 4, "L_tab": 945, "n_element_tab": 111, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 376, "n_element_tab_1": 57, "formula_len_all": 243, "formula_len_all_1": 226, "len_all": 119134, "len_all_1": 59231, "len_abs": 1007, "len_title": 145, "len_sents": 39522, "len_sents_1": 28481, "n_sents": 317, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1011, "title": "SceneFunctioner: Tailoring Large Language Model for Function-Oriented Interactive Scene Synthesis", "abs": "With the Large Language Model (LLM) skyrocketing in recent years, an increasing body of research has focused on leveraging these models for 3D scene synthesis. However, most existing works do not emphasize homeowner's functional preferences, often resulting in scenes that are logically arranged but fall short of serving practical functions. To address this gap, we introduce SceneFunctioner, an interactive scene synthesis framework that tailors the LLM to prioritize functional requirements. The framework is interactive, enabling users to select functions and room shapes. SceneFunctioner first distributes these selected functions into separate areas called zones and determines the furniture for each zone. It then organizes the furniture into groups before arranging them within their respective zones to complete the scene design. Quantitative analyses and user studies showcase our framework\u2019s state-of-the-art performance in terms of both design quality and functional consistency with the user input.", "keywords": [ "Scene Synthesis", "Multi-Function Design", "Large Language Model", "User Interaction" ], "rating_list": [ 8, 8, 3, 1 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 1, 1 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "IWHvZE3V9L", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Mahdi Soltanolkotabi", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Full Professor" }, { "name": "Reinhard Heckel", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Associate Professor" }, { "name": "Simon Wiedemann", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Zalan Fabian", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Postdoc" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 37, "n_ref": 68, "n_ref_all": 83, "n_fig": 8, "n_tab": 1, "L_tab": 378, "n_element_tab": 91, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 587, "n_element_tab_1": 101, "formula_len_all": 76, "formula_len_all_1": 76, "len_all": 129070, "len_all_1": 55156, "len_abs": 1151, "len_title": 138, "len_sents": 39779, "len_sents_1": 25855, "n_sents": 303, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1153, "title": "ProPicker: Promptable Segmentation for Particle Picking in Cryogenic Electron Tomography", "abs": "Cryogenic electron tomography (cryo-ET) can produce detailed 3D images called\ntomograms of cellular environments. An essential step of cryo-ET reconstruction and analysis is to find all instances of a protein in tomograms, a task known as particle picking. Due to the low signal-to-noise ratio, artifacts, and vast diversity in proteins, particle picking is a challenging 3D object detection problem. Existing approaches are either slow or limited to picking a few particles of interest, which requires large annotated and difficult to obtain training datasets. In this work, we propose ProPicker, a fast and universal particle picker that can detect particles beyond those in the training set. Our promptable design allows for selectively detecting a specific protein in the volume based on an input prompt. Our experiments demonstrate that through a favorable trade-off between performance and speed, ProPicker can achieve performance close to or on par with state-of-the-art universal pickers, while being up to an order of magnitude faster. Moreover, ProPicker can be efficiently adapted to new proteins through fine-tuning on few annotated samples.", "keywords": [ "cryo-ET", "cryo-EM", "cryogenic electron tomography", "cryogenic electron microscopy", "particle", "picking", "particle picking", "object detection" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 2, 1, 1, 1 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "IWC6zUEVcL", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Haobo Fu", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Haowei Lin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Kaichen He", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "QIANG FU", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Yitao Liang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zihao Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zilong Zheng", "gender": "Male", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Zhengxinyue", "gender": "Female", "institution": "bigai", "country": "", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 97, "n_ref_all": 124, "n_fig": 10, "n_tab": 9, "L_tab": 4839, "n_element_tab": 381, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 2105, "n_element_tab_1": 132, "formula_len_all": 0, "formula_len_all_1": 238, "len_all": 220393, "len_all_1": 62526, "len_abs": 1214, "len_title": 123, "len_sents": 66799, "len_sents_1": 28061, "n_sents": 517, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1263, "title": "Towards Evaluating Generalist Agents: An Automated Benchmark in Open World", "abs": "Evaluating generalist agents presents significant challenges due to their wide-ranging abilities and the limitations of current benchmarks in assessing true generalization. We introduce the \\textbf{M}ine\\textbf{C}raft \\textbf{U}niverse (\\textbf{MCU}), a fully automated benchmarking framework set within the open-world game \\emph{Minecraft}. MCU dynamically generates and evaluates a broad spectrum of tasks, offering three core components: 1) a task generation mechanism that provides maximal freedom and variability, 2) an ever-expanding set of over \\textbf{3K} composable atomic tasks, and 3) a general evaluation framework that supports open-ended task assessment. By integrating large language models (LLMs), MCU dynamically creates diverse environments for each evaluation, fostering agent generalization. The framework uses a vision-language model (VLM) to automatically generate evaluation criteria, achieving over 90\\% agreement with human ratings across multi-dimensional assessments, which demonstrates that MCU is a scalable and explainable solution for evaluating generalist agents. Additionally, we show that while state-of-the-art foundational models perform well on specific tasks, they often struggle with increased task diversity and difficulty.", "keywords": [ "open-world", "benchmark", "generalist agents" ], "rating_list": [ 6, 3, 6, 1 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 4, 1, 4, 5 ] }, { "paper_id": "IUzQfdkkoL", "primary_area": "datasets and benchmarks", "n_author": 13, "author_info_list": [ { "name": "Alexander Wong", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Full Professor" }, { "name": "Chris Czarnecki", "gender": "Male", "institution": "KisoJi Biotechnology", "country": "CA", "position": "AI Scientist" }, { "name": "Dayou Mao", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Researcher" }, { "name": "Edward Delp", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "Full Professor" }, { "name": "Fengqing Zhu", "gender": "unknown", "institution": "Purdue University, Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "Gautham Vinod", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Jiangpeng He", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Postdoc" }, { "name": "Jinge Ma", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Pengcheng Xi", "gender": "unknown", "institution": "National Research Council of Canada", "country": "CA", "position": "Researcher" }, { "name": "Saeejith Nair", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Siddeshwar Raghavan", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Talha Ibn Mahmud", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Yuhao Chen", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 2, "n_ref_uni": 29, "n_ref": 51, "n_ref_all": 82, "n_fig": 11, "n_tab": 8, "L_tab": 1769, "n_element_tab": 204, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3694, "n_element_tab_1": 174, "formula_len_all": 319, "formula_len_all_1": 34, "len_all": 204850, "len_all_1": 64270, "len_abs": 1944, "len_title": 97, "len_sents": 47553, "len_sents_1": 30490, "n_sents": 340, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1736, "title": "MetaFood3D: 3D Food Dataset with Nutrition Values", "abs": "Food computing is both important and challenging in computer vision (CV). It significantly contributes to the development of CV algorithms due to its frequent presence in datasets across various applications, ranging from classification and instance segmentation to 3D reconstruction. The polymorphic shapes and textures of food, coupled with high variation in forms and vast multimodal information, including language descriptions and nutritional data, make food computing a complex and demanding task for modern CV algorithms. 3D food modeling is a new frontier for addressing food related problems, due to its inherent capability to deal with random camera views and its straightforward representation for calculating food portion size. However, the primary hurdle in the development of algorithms for food object analysis is the lack of nutrition values in existing 3D datasets. Moreover, in the broader field of 3D research, there is a critical need for domain-specific test datasets. To bridge the gap between general 3D vision and food computing research, we introduce MetaFood3D. This dataset consists of 637 meticulously scanned and labeled 3D food objects across 108 categories, featuring detailed nutrition information, weight, and food codes linked to a comprehensive nutrition database. Our MetaFood3D dataset emphasizes intra-class diversity and includes rich modalities such as textured mesh files, RGB-D videos, and segmentation masks.\nExperimental results demonstrate our dataset's significant potential for improving algorithm performance, highlight the challenging gap between video captures and 3D scanned data, and showcase the strengths of MetaFood3D in high-quality data generation, simulation, and augmentation.", "keywords": [ "3D food dataset", "real scanned data", "nutrition values", "image-based dietary assessment", "3D reconstruction", "food portion estimation", "deep learning", "computer vision" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "IUwqJ8VT4F", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Dengyun Peng", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Diyuan Shi", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Donglin Wang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Associate Professor" }, { "name": "Jiacheng Liu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Xing Lei", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zifeng Zhuang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Ziqi Zhang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Micro Masters Student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 38, "n_ref": 72, "n_ref_all": 88, "n_fig": 6, "n_tab": 5, "L_tab": 2698, "n_element_tab": 478, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2366, "n_element_tab_1": 371, "formula_len_all": 646, "formula_len_all_1": 687, "len_all": 118331, "len_all_1": 66110, "len_abs": 1126, "len_title": 109, "len_sents": 33874, "len_sents_1": 27187, "n_sents": 268, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1130, "title": "Revisiting the Design Choices in Max-Return Sequence Modeling", "abs": "Decision Transformer (DT), free from optimal value functions fitting and policy gradient computation, attempts to solve offline reinforcement learning (RL) via supervised sequence modeling. During inference, sequence modeling requires an initial target returns assigned with expert knowledge, which blocks comprehensive evaluation on more diverse datasets. As a result, existing sequence modeling only focuses on limited evaluation on Gym datasets and some understanding is severely biased. In this paper, we aim to revisit the design choices, including architecture and context length, in sequence modeling on more diverse datasets. We utilize the max-return sequence modeling that replaces the manual target returns with maximized returns predicted by itself. We systematically investigate the impact of 1) architectural choices and 2) context lengths in max-return sequence modeling on nine datasets with varying data distributions. Abundant experiments and thorough analyses reveal that design choices are highly influenced by the dataset characteristics, which further underscores the significance of more diverse evaluation.", "keywords": [ "Sequence Modeling in Offline Reinforcement Learning" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "IUmj2dw5se", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Jundong Li", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" }, { "name": "Peng Wang", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Song Wang", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Tong Zhou", "gender": "Female", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Undergrad student" }, { "name": "Yushun Dong", "gender": "Male", "institution": "Florida State University", "country": "US", "position": "Assistant Professor" }, { "name": "Zhen Tan", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 52, "n_ref": 144, "n_ref_all": 180, "n_fig": 10, "n_tab": 14, "L_tab": 6744, "n_element_tab": 285, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2342, "n_element_tab_1": 133, "formula_len_all": 131, "formula_len_all_1": 0, "len_all": 280505, "len_all_1": 65140, "len_abs": 1305, "len_title": 125, "len_sents": 87027, "len_sents_1": 29057, "n_sents": 742, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 77, "L_abs": 1207, "title": "CEB: Compositional Evaluation Benchmark for Fairness in Large Language Models", "abs": "As Large Language Models (LLMs) are increasingly deployed to handle various natural language processing (NLP) tasks, concerns regarding the potential negative societal impacts of LLM-generated content have also arisen. To evaluate the biases exhibited by LLMs, researchers have recently proposed a variety of datasets. However, existing bias evaluation efforts often focus on only a particular type of bias and employ inconsistent evaluation metrics, leading to difficulties in comparison across different datasets and LLMs. To address these limitations, we collect a variety of datasets designed for the bias evaluation of LLMs, and further propose CEB, a Compositional Evaluation Bechmark that covers different types of bias across different social groups and tasks. The curation of CEB is based on our newly proposed compositional taxonomy, which characterizes each dataset from three dimensions: bias types, social groups, and tasks. By combining the three dimensions, we develop a comprehensive evaluation strategy for the bias in LLMs. Our experiments demonstrate that the levels of bias vary across these dimensions, thereby providing guidance for the development of specific bias mitigation methods.", "keywords": [ "Fairness", "Bias", "Benchmark", "Large Language Models" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "IUmDBY4NOQ", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Pengxiang Li", "gender": "unknown", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Wu Wei", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Xiaomeng Fan", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yunde Jia", "gender": "Male", "institution": "Shenzhen MSU-BIT University", "country": "CN", "position": "Full Professor" }, { "name": "Yuwei Wu", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Zhi Gao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Postdoc" }, { "name": "Zhipeng Lu", "gender": "Male", "institution": "Shenzhen MSU-BIT University", "country": "CN", "position": "Lecturer" }, { "name": "Mehrtash T. Harandi", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 58, "n_formula_1": 13, "n_ref_uni": 59, "n_ref": 117, "n_ref_all": 175, "n_fig": 11, "n_tab": 12, "L_tab": 2379, "n_element_tab": 183, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 434, "n_element_tab_1": 50, "formula_len_all": 3138, "formula_len_all_1": 781, "len_all": 242975, "len_all_1": 66746, "len_abs": 1520, "len_title": 136, "len_sents": 66644, "len_sents_1": 27858, "n_sents": 550, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1299, "title": "Geometry-aware Distance Measure for Diverse Hierarchical Structures in Hyperbolic Spaces", "abs": "Learning in hyperbolic spaces has gained increasing attention due to the superior capability of modeling hierarchical structures. Existing hyperbolic learning methods use a fixed distance measure that assumes a uniform hierarchical structure across all data points. However, this assumption does not always hold in real-world scenarios, considering the diversity of the hierarchical structures of data. This work proposes to learn geometry aware distance measures that dynamically adjust to accommodate diverse hierarchical structures in hyperbolic spaces. We derive geometry aware distance measures by generating projections and curvatures for each pair of samples, which maps each pair to a suitable hyperbolic space. We introduce a revised low-rank decomposition scheme and a hard-pair mining mechanism to reduce the computational cost incurred by the pairwise generation without compromising accuracy. Moreover, we derive an upper bound of the low-rank approximation error via Talagrand concentration inequality to guarantee the effectiveness of our low-rank decomposition scheme. Theoretical analysis and experiments on standard image classification and few-shot learning tasks affirm the effectiveness of our method in refining hyperbolic learning through our geometry aware distance measures.", "keywords": [ "hyperbolic learning", "metric learning" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "IULlNTZZel", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Andrew Zhuoer Feng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Boqun Kou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Cunxiang Wang", "gender": "Not Specified", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Haohan Chi", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Hongning Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Lin Fan", "gender": "Female", "institution": "zhipuAI", "country": "CN", "position": "Researcher" }, { "name": "Minlie Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaotao Gu", "gender": "Male", "institution": "Zhipu AI", "country": "CN", "position": "Researcher" }, { "name": "Yu Luo", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Wenbosi", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 44, "n_ref": 58, "n_ref_all": 82, "n_fig": 3, "n_tab": 10, "L_tab": 5037, "n_element_tab": 490, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2199, "n_element_tab_1": 309, "formula_len_all": 481, "formula_len_all_1": 200, "len_all": 211248, "len_all_1": 64115, "len_abs": 2038, "len_title": 68, "len_sents": 64087, "len_sents_1": 26922, "n_sents": 570, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1564, "title": "RedHat: Towards Reducing Hallucination in Essay Critiques with Large Language Models", "abs": "Essay critiques refer to the textual assessment of an essay, serving as the basis for the scoring of the essay, and are crucial for the improvements of the essay. Essay critique generation has received increasing attention after the blooming of large language models (LLMs), which show promising potential in writing and critiquing essays. Automatic critique generation can streamline both instructors and reviewers as well as spur LLM advancement in long context generation characterized by essay writing. However, current LLMs suffer from hallucinations when generating essay critiques, which are still under-explored in the community. To facilitate research in reliable essay critique generation, we first define this task with a unified input-output format as well as clear judging criteria. To minimize hallucinations in critique generation, we introduce RedHat, a novel approach that embeds the key information from essays directly into the generation process through document-level question-answering, ensuring critiques stay firmly anchored to the original text. We collected a large-scale, high-quality essay critique dataset called EssayC, annotated by human experts over multiple LLM-generated critiques, from a campus undergraduate essay writing course. We experimented RedHat backboned by commercial and open-sourced LLMs. Results showed that critiques generated by RedHat are preferred by human experts over baseline in 20% of cases on EssayC in detailedness and informativeness, with a decrement around 10% on hallucinations in our judging criteria.", "keywords": [ "essay critique generation", "large language model", "hallucination" ], "rating_list": [ 8, 5, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "ITi9Zwkge2", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Adrian Chan", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "MS student" }, { "name": "Akram Khater", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Full Professor" }, { "name": "Anupam Mijar", "gender": "unknown", "institution": "North Carolina State University", "country": "US", "position": "Undergrad student" }, { "name": "Chau-Wai Wong", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Assistant Professor" }, { "name": "Mehreen Saeed", "gender": "Female", "institution": "North Carolina State University", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 56, "n_ref": 101, "n_ref_all": 117, "n_fig": 4, "n_tab": 6, "L_tab": 1729, "n_element_tab": 133, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 700, "n_element_tab_1": 8, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 148002, "len_all_1": 57768, "len_abs": 1520, "len_title": 126, "len_sents": 44899, "len_sents_1": 29642, "n_sents": 334, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1612, "title": "HATFormer: Historic Handwritten Arabic Text Recognition with Transformers", "abs": "Arabic handwritten text recognition (HTR) is challenging, especially for historical texts, due to diverse writing styles and the intrinsic features of Arabic script. Additionally, Arabic handwriting datasets are smaller compared to English ones, making it difficult to train generalizable Arabic HTR models. To address these challenges, we propose HATFormer, a transformer-based encoder-decoder architecture that builds on a state-of-the-art English HTR model. By leveraging the transformer's attention mechanism, HATFormer captures spatial contextual information to address the intrinsic challenges of Arabic script through differentiating cursive characters, decomposing visual representations, and identifying diacritics. Our customization to historical handwritten Arabic includes an image processor for effective ViT information preprocessing, a text tokenizer for compact Arabic text representation, and a training pipeline that accounts for a limited amount of historic Arabic handwriting data. HATFormer achieves a character error rate~(CER) of 8.6% on the largest public historical handwritten Arabic dataset, with a 51% improvement over the best baseline in the literature. HATFormer also attains a comparable CER of 4.2% on the largest private non-historical dataset. Our work demonstrates the feasibility of adapting an English HTR method to a low-resource language with complex, language-specific challenges, contributing to advancements in document digitization, information retrieval, and cultural preservation. The source code will be available as a link on the discussion forum once it is open.", "keywords": [ "OCR", "computer vision", "transformer", "handwritten text recognition", "HTR", "Arabic", "historical handwritten Arabic", "handwritten Arabic" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 2, 5 ] }, { "paper_id": "IT7LSnBdtY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Do Huu Quan", "gender": "Male", "institution": "VinUniversity", "country": "VN", "position": "Research Assistant" }, { "name": "Minh N. Do", "gender": "Male", "institution": "VinUniversity", "country": "VN", "position": "Honorary Vice Provost" }, { "name": "Anh Duy Nguyen", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Khoa Doan", "gender": "unknown", "institution": "VinUniversity", "country": "VN", "position": "Assistant Professor" } ], "n_formula": 31, "n_formula_1": 13, "n_ref_uni": 30, "n_ref": 51, "n_ref_all": 77, "n_fig": 8, "n_tab": 8, "L_tab": 5240, "n_element_tab": 668, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2067, "n_element_tab_1": 250, "formula_len_all": 3526, "formula_len_all_1": 1010, "len_all": 159375, "len_all_1": 64826, "len_abs": 1357, "len_title": 108, "len_sents": 43886, "len_sents_1": 27709, "n_sents": 365, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1482, "title": "Are you SURE? Enhancing Multimodal Pretraining with Missing Modalities through Uncertainty Estimation", "abs": "Multimodal learning has demonstrated incredible successes by integrating diverse data sources, yet it often relies on the availability of all modalities - an assumption that rarely holds in real-world applications. Pretrained multimodal models, while effective, struggle when confronted with small-scale and incomplete datasets (i.e., missing modalities), limiting their practical applicability. Previous studies on reconstructing missing modalities have overlooked the reconstruction's potential unreliability, which could compromise the quality of the final outputs. We present **SURE** (Scalable Uncertainty and Reconstruction Estimation), a novel framework that extends the capabilities of pretrained multimodal models by introducing latent space reconstruction and uncertainty estimation for both reconstructed modalities and downstream tasks. Our method is architecture-agnostic, reconstructs missing modalities, and delivers reliable uncertainty estimates, improving both interpretability and performance. SURE introduces a unique Pearson Correlation-based loss and applies statistical error propagation in deep networks for the first time, allowing precise quantification of uncertainties from missing data and model predictions. Extensive experiments across tasks such as sentiment analysis, genre classification, and action recognition show that SURE consistently achieves state-of-the-art performance, ensuring robust predictions even in the presence of incomplete data.", "keywords": [ "Multimodal learning", "Uncertainty estimation" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "IT33VLRJuS", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Euxhen Hasanaj", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Ziv Bar-Joseph", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "Barnabas Poczos", "gender": "unknown", "institution": "School of Computer Science", "country": "", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 73, "n_ref": 96, "n_ref_all": 109, "n_fig": 9, "n_tab": 1, "L_tab": 253, "n_element_tab": 21, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 253, "n_element_tab_1": 21, "formula_len_all": 270, "formula_len_all_1": 302, "len_all": 194774, "len_all_1": 60638, "len_abs": 1026, "len_title": 102, "len_sents": 35213, "len_sents_1": 28665, "n_sents": 283, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 1025, "title": "Recovering Time-Varying Networks From Single-Cell Data", "abs": "Gene regulation is a dynamic process that underlies all aspects of human development, disease response, and other key biological processes. The reconstruction of temporal gene regulatory networks has conventionally relied on regression analysis, graphical models, or other types of relevance networks. With the large increase in time series single-cell data, new approaches are needed to address the unique scale and nature of this data for reconstructing such networks. Here, we develop a deep neural network, Marlene, to infer dynamic graphs from time series single-cell gene expression data. Marlene constructs directed gene networks using a self-attention mechanism where the weights evolve over time using recurrent units. By employing meta learning, the model is able to recover accurate temporal networks even for rare cell types. In addition, it can identify gene interactions relevant to specific biological responses, including COVID-19 immune response, fibrosis, and aging, paving the way for potential treatments.", "keywords": [ "gene-regulatory-network", "temporal-graphs", "meta-learning", "genomics", "cell-types", "neural-network", "graph-structure-learning", "covid-19", "aging" ], "rating_list": [ 3, 3, 8, 3 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 2, 4, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ISqx8giekS", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Anshumali Shrivastava", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Associate Professor" }, { "name": "Tianyi Zhang", "gender": "Male", "institution": "Rice University", "country": "US", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 20, "n_ref_uni": 53, "n_ref": 114, "n_ref_all": 151, "n_fig": 5, "n_tab": 9, "L_tab": 9413, "n_element_tab": 894, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 4829, "n_element_tab_1": 403, "formula_len_all": 1400, "formula_len_all_1": 1336, "len_all": 173092, "len_all_1": 71369, "len_abs": 2045, "len_title": 141, "len_sents": 40142, "len_sents_1": 29192, "n_sents": 271, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 93, "L_abs": 1897, "title": "LeanQuant: Accurate and Scalable Large Language Model Quantization with Loss-error-aware Grid", "abs": "Large language models (LLMs) have shown immense potential across various domains, but their high memory requirements and inference costs remain critical challenges for deployment. Post-training quantization (PTQ) has emerged as a promising technique to reduce memory requirements and decoding latency. However, recent accurate quantization methods often depend on specialized computations or custom data formats to achieve better model quality, which limits their compatibility with popular frameworks, as they require dedicated inference kernels tailored to specific hardware and software platforms, hindering wider adoption. Furthermore, many competitive methods have high resource requirements and computational overhead, making it challenging to scale them to hundreds of billions of parameters. In response to these challenges, we propose LeanQuant (Loss-error-aware Network Quantization), a novel quantization method that is accurate, versatile, and scalable. In the existing popular iterative loss-error-based quantization framework, we identify a critical limitation in prior methods: the min-max affine quantization grid fails to preserve model quality due to outliers in inverse Hessian diagonals. To overcome this fundamental issue, we propose learning loss-error-aware grids, instead of using non-adaptive min-max affine grids. Our approach not only produces quantized models that are more accurate but also generalizes to a wider range of quantization types, including affine and non-uniform quantization, enhancing compatibility with more frameworks. Extensive empirical evaluations on recent LLMs demonstrate that LeanQuant is highly accurate, comparing favorably against recent competitive baselines in model quality, and scalable, achieving very accurate quantization of Llama-3.1 405B, one of the largest open-source LLMs to date, using two Quadro RTX 8000-48GB GPUs in 21 hours.", "keywords": [ "large language model", "quantization" ], "rating_list": [ 5, 5, 3, 6, 6 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 4, 4 ] }, { "paper_id": "ISBmUNKPST", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "George Demiris", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "Full Professor" }, { "name": "Jia Xu", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "Researcher" }, { "name": "Joost Wagenaar", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "Assistant Professor" }, { "name": "Li Shen", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Full Professor" }, { "name": "Patryk Orzechowski", "gender": "unknown", "institution": "University of Pennsylvania", "country": "", "position": "Data Scientist" }, { "name": "Rachael Paulbeck", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Researcher" }, { "name": "Ruochen Jin", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Intern" }, { "name": "Shu Yang", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "Postdoc" }, { "name": "Tianyi Wei", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "MS student" }, { "name": "Bo-Jian Hou", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "Postdoc" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 48, "n_ref": 128, "n_ref_all": 161, "n_fig": 4, "n_tab": 11, "L_tab": 4110, "n_element_tab": 297, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1060, "n_element_tab_1": 34, "formula_len_all": 390, "formula_len_all_1": 0, "len_all": 232922, "len_all_1": 60471, "len_abs": 928, "len_title": 126, "len_sents": 75928, "len_sents_1": 30293, "n_sents": 555, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 930, "title": "MentalChat16K: A Benchmark Dataset for Conversational Mental Health Assistance", "abs": "We introduce MentalChat16K, an English benchmark dataset combining a synthetic mental health counseling dataset and a dataset of anonymized transcripts from interventions between Behavioral Health Coaches and Caregivers of patients in palliative or hospice care. Covering a diverse range of conditions like depression, anxiety, and grief, this curated dataset is designed to facilitate the development and evaluation of large language models for conversational mental health assistance. By providing a high-quality resource tailored to this critical domain, MentalChat16K aims to advance research on empathetic, personalized AI solutions to improve access to mental health support services. The dataset prioritizes patient privacy, ethical considerations, and responsible data usage. MentalChat16K presents a valuable opportunity for the research community to innovate AI technologies that can positively impact mental well-being.", "keywords": [ "Mental Health", "Large Language Model", "Behavioral Health", "Question and Answering" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "IS7kW28VVt", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Ke Zhu", "gender": "unknown", "institution": "University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Zhoufan Zhu", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 23, "n_ref_uni": 22, "n_ref": 59, "n_ref_all": 78, "n_fig": 3, "n_tab": 5, "L_tab": 1637, "n_element_tab": 183, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 373, "n_element_tab_1": 51, "formula_len_all": 2118, "formula_len_all_1": 2196, "len_all": 127257, "len_all_1": 65458, "len_abs": 1204, "len_title": 117, "len_sents": 44008, "len_sents_1": 29587, "n_sents": 329, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1207, "title": "AlphaQCM: Alpha Discovery with Distributional Reinforcement Learning", "abs": "Finding synergistic formulaic alphas is very important but challenging for researchers and practitioners in finance. In this paper, we reconsider the discovery of formulaic alphas from the viewpoint of sequential decision-making, and conceptualize the entire alpha-mining process as a non-stationary and reward-sparse Markov decision process. To overcome the challenges of non-stationarity and reward-sparsity, we propose the AlphaQCM method, a novel distributional reinforcement learning method designed to search for synergistic formulaic alphas efficiently. The AlphaQCM method first learns the Q function and quantiles via a Q network and a quantile network, respectively. Then, the AlphaQCM method applies the quantiled conditional moment method to learn unbiased variance from the potentially biased quantiles. Guided by the learned Q function and variance, the AlphaQCM method navigates the non-stationarity and reward-sparsity to explore the vast search space of formulaic alphas with high efficacy. Empirical applications to real-world datasets demonstrate that our AlphaQCM method significantly outperforms its competitors, particularly when dealing with large datasets comprising numerous stocks.", "keywords": [ "Distributional Reinforcement Learning", "Computational Finance", "Formulaic Alpha", "Quantiled Conditional Moments", "Stock Trend Forecasting" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 2, 1 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "IRvx66cxip", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Maisha Maliha", "gender": "unknown", "institution": "University of Oklahoma", "country": "US", "position": "PhD student" }, { "name": "Sumit Kumar Jha", "gender": "unknown", "institution": "Florida International University", "country": "US", "position": "Full Professor" }, { "name": "Vishal Pramanik", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 26, "n_ref": 33, "n_ref_all": 52, "n_fig": 7, "n_tab": 2, "L_tab": 981, "n_element_tab": 144, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 981, "n_element_tab_1": 144, "formula_len_all": 1726, "formula_len_all_1": 1355, "len_all": 114929, "len_all_1": 59341, "len_abs": 1009, "len_title": 169, "len_sents": 35026, "len_sents_1": 25536, "n_sents": 278, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 121, "L_abs": 1012, "title": "Enhancing Integrated Gradients Using Emphasis Factors and Attention for Effective Explainability of Large Language Models", "abs": "Understanding the decision-making processes of large language models (LLMs) is critical for ensuring transparency and trustworthiness. While Integrated Gradients (IG) is a popular method for model explainability, it faces limitations when applied to autoregressive models due to issues like exploding gradients and the neglect of the attention mechanisms. In this paper, we propose an enhanced explainability framework that augments IG with emphasis factors and attention mechanisms. By incorporating attention, we capture contextual dependencies between words, and the introduction of emphasis factors mitigates gradient issues encountered during attribution calculations. Our method provides more precise and interpretable explanations for autoregressive LLMs, effectively highlighting word-level contributions in text generation tasks. Experimental results demonstrate that our approach outperforms standard IG and baseline models in explaining word-level attributions, advancing the interpretability of LLMs.", "keywords": [ "XAI", "Explainability", "Integrated Gradients", "Large Language Models", "GPT" ], "rating_list": [ 1, 8, 1, 1 ], "soundness_list": [ 2, 3, 2, 1 ], "presentation_list": [ 2, 3, 1, 1 ], "contribution_list": [ 1, 3, 2, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "IRsxTYPqhQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Wei-Lin Chen", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Xinyu Zhu", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Yu Meng", "gender": "Male", "institution": "University of Virginia", "country": "", "position": "Assistant Professor" }, { "name": "Zhepei Wei", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" } ], "n_formula": 99, "n_formula_1": 102, "n_ref_uni": 52, "n_ref": 88, "n_ref_all": 114, "n_fig": 12, "n_tab": 2, "L_tab": 616, "n_element_tab": 67, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 616, "n_element_tab_1": 67, "formula_len_all": 153387, "formula_len_all_1": 151352, "len_all": 465019, "len_all_1": 369249, "len_abs": 2539, "len_title": 119, "len_sents": 150730, "len_sents_1": 140992, "n_sents": 408, "n_sents_1": 333, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 2507, "title": "Fast and Accurate Language Model Decoding via Parallel Token Processing", "abs": "Autoregressive decoding suffers from an inherent efficiency bottleneck due to its sequential token generation process, where each token must be generated before the next can be processed. This sequential dependency significantly limits the ability to fully exploit the parallel processing power of modern hardware. While speculative decoding and layer skipping offer promising speedups, both approaches come with drawbacks. Speculative decoding relies on a secondary small ''drafter'' model, which not only increases memory overhead but may also be unavailable in many cases---the drafter must share the same tokenizer and vocabulary as the main model for compatibility between generated and verified tokens. Layer skipping, on the other hand, can cause discrepancies in the generated output compared to standard autoregressive decoding, as skipped layers do not compute the key-value (KV) cache that plays a crucial role in predicting future tokens.\nIn this work, we introduce a fast and accurate decoding method, ParaDecode, which accelerates autoregressive decoding while ensuring output parity, without the need for auxiliary models or changes to original model parameters. Our approach is driven by the observation that many tokens---particularly simple or highly-predictable ones---can be accurately predicted using intermediate layer representations, without requiring computation through the entire model. Once the model reaches a certain confidence, further layers are unlikely to significantly alter the prediction. ParaDecode generates tokens at an intermediate layer when confidence is sufficiently high. \nThis allows the next token computation to commence immediately, in parallel with the completion of the KV cache computation for the early-predicted token in its remaining layers. This parallelism, implemented using batched matrix operations, enables simultaneous processing of multiple tokens across different layers, thereby maximizing hardware utilization and reducing overall decoding latency. To ensure output consistency, a final verification step is applied to guarantee that the early-predicted tokens match the results of standard autoregressive decoding. Experiments across diverse generation tasks, including text summarization, code generation, and mathematical reasoning, demonstrate that ParaDecode consistently achieves superior decoding throughput compared to baselines with up to 1.53$\\times$ speedup, while guaranteeing output parity with standard autoregressive decoding.", "keywords": [ "Autoregressive model", "efficient decoding", "parallel token processing" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 3, 3, 1 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "IRjT0AmsDI", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Weiran Huang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhiquan Tan", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 16, "n_ref_uni": 17, "n_ref": 44, "n_ref_all": 70, "n_fig": 15, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2222, "formula_len_all_1": 1026, "len_all": 93965, "len_all_1": 53078, "len_abs": 981, "len_title": 112, "len_sents": 32210, "len_sents_1": 22888, "n_sents": 302, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 988, "title": "Understanding Grokking: Insights from Neural Network Robustness", "abs": "Recently, an interesting phenomenon called grokking has gained much attention, where generalization occurs long after the models have initially overfitted the training data. We try to understand this seemingly strange phenomenon through the robustness of the neural network. From a robustness perspective, we show that the usually observed decreasing of $l_2$ weight norm of the neural network is theoretically connected to the occurrence of grokking. Therefore, we propose to use perturbation-based methods to enhance robustness and speed up the generalization process. Furthermore, we show that the speed-up of generalization when using our proposed method can be explained by learning the commutative law, a necessary condition when the model groks on the test dataset. In addition, we empirically observe that \n$l_2$ norm correlates with grokking on the test data not in a timely way and then propose new metrics based on robustness that correlate better with the grokking phenomenon.", "keywords": [ "Grokking", "Representation Learning", "Training Dynamics" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "IReyEK7Sst", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Chun-Hao Paul Huang", "gender": "Not Specified", "institution": "Adobe Systems", "country": "GB", "position": "Researcher" }, { "name": "Duygu Ceylan", "gender": "unknown", "institution": "Adobe Systems", "country": "GB", "position": "Research Scientist" }, { "name": "Juil Koo", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Minhyuk Sung", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Paul Guerrero", "gender": "unknown", "institution": "Adobe Systems", "country": "GB", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 0, "n_ref": 0, "n_ref_all": 1, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 2249, "len_all_1": 2249, "len_abs": 161, "len_title": 77, "len_sents": 159, "len_sents_1": 159, "n_sents": 2, "n_sents_1": 2, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 160, "title": "VideoHandles: Editing 3D Object Compositions in Videos Using Video Generative Priors", "abs": "We thank the reviewers for their valuable comments. After careful consideration, we think our paper is inappropriate for ICLR and decided to withdraw our paper.", "keywords": [ "Video Editing", "Video Diffusion Models", "DiT", "3D Transformation" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "IRL9wUiwab", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Artur Dubrawski", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Research Professor" }, { "name": "Micha\u0142 Wili\u0144ski", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Intern" }, { "name": "Mononito Goswami", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Nina \u017bukowska", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Intern" }, { "name": "Willa Potosnak", "gender": "unknown", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 8, "n_ref_uni": 31, "n_ref": 61, "n_ref_all": 85, "n_fig": 11, "n_tab": 10, "L_tab": 18241, "n_element_tab": 2692, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 158, "n_element_tab_1": 4, "formula_len_all": 1076, "formula_len_all_1": 604, "len_all": 191336, "len_all_1": 61246, "len_abs": 1083, "len_title": 76, "len_sents": 47229, "len_sents_1": 28594, "n_sents": 351, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1054, "title": "Exploring Representations and Interventions in Time Series Foundation Models", "abs": "Time series foundation models promise to be powerful tools for a wide range of applications. However, their internal representations and learned concepts are still not well understood. In this study, we investigate the structure and redundancy of representations across various TSFMs, examining the self-similarity of model layers within and across different model sizes. This analysis reveals block-like redundancy in the representations, which can be utilized for informed pruning to improve inference speed and efficiency. Additionally, we explore the concepts learned by these models\u2014such as periodicity and trends\u2014and how these can be manipulated through latent space steering to influence model behavior. Our experiments show that steering interventions can introduce new features, like adding periodicity or trends to signals that initially lacked them. These findings underscore the value of representational analysis for optimizing models and demonstrate how conceptual steering offers new possibilities for more controlled time series modeling.", "keywords": [ "Time Series Foundation Models", "Model Steering", "Interpretability", "Pruning" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "IRCo9mHScB", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Jiao Dai", "gender": "Female", "institution": "Institute of Information Engineering, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Jin Liu", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Jizhong Han", "gender": "Male", "institution": "Institute of Information Engineering", "country": "CN", "position": "Full Professor" }, { "name": "Qiao Li", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xi Wang", "gender": "Female", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaomeng Fu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xingyu Gao", "gender": "unknown", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 9, "n_ref_uni": 37, "n_ref": 98, "n_ref_all": 124, "n_fig": 4, "n_tab": 7, "L_tab": 2759, "n_element_tab": 330, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 4267, "n_element_tab_1": 307, "formula_len_all": 1074, "formula_len_all_1": 614, "len_all": 161482, "len_all_1": 69062, "len_abs": 1625, "len_title": 145, "len_sents": 51214, "len_sents_1": 28678, "n_sents": 388, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1674, "title": "OMS: One More Step Noise Searching to Enhance Membership Inference Attacks for Diffusion Models", "abs": "The data-intensive nature of Diffusion models amplifies the risks of privacy infringements and copyright disputes, particularly when training on extensive unauthorized data scraped from the Internet. Membership Inference Attacks (MIA) aim to determine whether a data sample has been utilized by the target model during training, thereby serving as a pivotal tool for privacy preservation. Current MIA employs the prediction loss to distinguish between training member samples and non-members. \nThese methods assume that, compared to non-members, members, having been encountered by the model during training result in a smaller prediction loss. However, this assumption proves ineffective in diffusion models due to the randomly noise sampled during the training process. Rather than estimating the loss, our approach examines this random noise and reformulate the MIA as a noise search problem, assuming that members are more feasible to find the noise used in the training process.\nWe formulate this noise search process as an optimization problem and employ the fixed-point iteration to solve it. We analyze current MIA methods through the lens of the noise search framework and reveal that they rely on the first residual as the discriminative metric to differentiate members and non-members. Inspired by this observation, we introduce \\textbf{OMS}, which augments existing MIA methods by iterating \\textbf{O}ne \\textbf{M}ore fixed-point \\textbf{S}tep to include a further residual, i.e., the second residual. \nWe integrate our method into various MIA methods across different diffusion models. The experimental results validate the efficacy of our proposed approach.", "keywords": [ "Membership Inference Attack", "Diffusion Models", "Data Privacy" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "IQxBDLmVpT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Alfonso Amayuelas", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Alon Albalak", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Antonis Antoniades", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "", "position": "Researcher" }, { "name": "Kexun Zhang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Xinyi Wang", "gender": "Female", "institution": "UC Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Yanai Elazar", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Postdoc" }, { "name": "William Yang Wang", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 36, "n_ref": 86, "n_ref_all": 113, "n_fig": 5, "n_tab": 3, "L_tab": 223, "n_element_tab": 32, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 778, "formula_len_all_1": 964, "len_all": 159465, "len_all_1": 63945, "len_abs": 1483, "len_title": 144, "len_sents": 51804, "len_sents_1": 31258, "n_sents": 352, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1489, "title": "Generalization v.s. Memorization: Tracing Language Models\u2019 Capabilities Back to Pretraining Data", "abs": "The impressive capabilities of large language models (LLMs) have sparked debate over whether these models genuinely generalize to unseen tasks or predominantly rely on memorizing vast amounts of pretraining data. To explore this issue, we introduce an extended concept of memorization, distributional memorization, which measures the correlation between the LLM output probabilities and the pretraining data frequency. To effectively capture task-specific pretraining data frequency, we propose a novel task-gram language model, which is built by counting the co-occurrence of semantically related $n$-gram pairs from task inputs and outputs in the pretraining corpus. Using the Pythia models trained on the Pile dataset, we evaluate four distinct tasks: machine translation, factual question answering, world knowledge understanding, and math reasoning. Our findings reveal varying levels of memorization, with the strongest effect observed in factual question answering. Furthermore, while model performance improves across all tasks as LLM size increases, only factual question answering shows an increase in memorization, whereas machine translation and reasoning tasks exhibit greater generalization, producing more novel outputs. This study demonstrates that memorization plays a larger role in simpler, knowledge-intensive tasks, while generalization is the key for harder, reasoning-based tasks, providing a scalable method for analyzing large pretraining corpora in greater depth.", "keywords": [ "language model", "pretraining data", "n-gram", "memorization" ], "rating_list": [ 3, 6, 8, 3 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "IQi8JOqLuv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Hao Xiong", "gender": "Non-Binary", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Jia Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Jianhao Huang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Undergrad student" }, { "name": "Xinyu Ye", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Ziang Chen", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Instructor" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 46, "n_formula_1": 20, "n_ref_uni": 40, "n_ref": 72, "n_ref_all": 122, "n_fig": 9, "n_tab": 15, "L_tab": 2951, "n_element_tab": 354, "n_fig_1": 10, "n_tab_1": 9, "L_tab_1": 3377, "n_element_tab_1": 303, "formula_len_all": 6631, "formula_len_all_1": 1910, "len_all": 215531, "len_all_1": 80046, "len_abs": 1459, "len_title": 154, "len_sents": 66658, "len_sents_1": 31774, "n_sents": 553, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 1464, "title": "An Expressive Quantum-Driven Graph Learning Approach with Application to Mixed-integer Linear Programming", "abs": "Graph-structure data is ubiquitous, and graph learning models have recently been further extended to address complex problems like mixed-integer linear programming (MILP). However, recent literature has shown that classic graph neural networks (GNNs) suffer fundamental limitations in learning MILP graph representation, i.e., GNNs may map two different MILP graphs to the same representation. To overcome the limitations of classical GNNs, we introduce an expressive quantum-driven graph learning approach, leveraging quantum machine learning (QML) to recognize patterns that are difficult for classical methods to learn. Specifically, the proposed Versatile Quantum Graph Learning Architecture (VQGLA) is composed of a node feature layer, a graph message interaction layer, and an optional auxiliary layer. Its versatility is reflected in effectively encoding features of nodes and edges while ensuring node permutation equivariance and flexibly creating different circuit structures for various expressive requirements and downstream tasks. VQGLA is well suited for learning complex graph tasks like MILP representation. Experimental results highlight the effectiveness of VQGLA in capturing and learning representations for MILPs. In comparison to traditional GNNs, VQGLA exhibits superior discriminative capabilities and demonstrates enhanced generalization across various problem instances, making it a more promising solution for complex optimization tasks.", "keywords": [ "Quantum machine learning", "Mixed-integer linear programming", "graph learning" ], "rating_list": [ 5, 8, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 2, 4, 4 ] }, { "paper_id": "IQdlPvj4dX", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Guido F. Montufar", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Full Professor" }, { "name": "Niket Nikul Patel", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "Undergrad student" } ], "n_formula": 173, "n_formula_1": 23, "n_ref_uni": 47, "n_ref": 91, "n_ref_all": 129, "n_fig": 13, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 13832, "formula_len_all_1": 1199, "len_all": 217633, "len_all_1": 66482, "len_abs": 974, "len_title": 111, "len_sents": 61918, "len_sents_1": 31783, "n_sents": 595, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 988, "title": "On the Local Complexity of Linear Regions in Deep ReLU Networks", "abs": "We define the $\\textit{local complexity}$ of a neural network with continuous piecewise linear activations as a measure of the density of linear regions over an input data distribution. We show theoretically that ReLU networks that learn low-dimensional feature representations have a lower local complexity. This allows us to connect recent empirical observations on feature learning at the level of the weight matrices with concrete properties of the learned functions. In particular, we show that the local complexity serves as an upper bound on the total variation of the function over the input data distribution and thus that feature learning can be related to adversarial robustness. Lastly, we consider how optimization drives ReLU networks towards solutions with lower local complexity. Overall, this work contributes a theoretical framework towards relating geometric properties of ReLU networks to different aspects of learning such as feature learning and representation cost.", "keywords": [ "ReLU networks", "Linear Regions", "Representation Learning", "Low-rank Bias", "Robustness", "Implicit Regularization" ], "rating_list": [ 5, 6, 6, 6, 6 ], "soundness_list": [ 1, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 3, 3 ] }, { "paper_id": "IQafqgqDzF", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chao Tian", "gender": "Male", "institution": "Texas A&M", "country": "US", "position": "Associate Professor" }, { "name": "Krishna Narayanan", "gender": "Male", "institution": "Texas A&M", "country": "", "position": "Full Professor" }, { "name": "Peter Just", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "MS student" }, { "name": "Yu-Shin Huang", "gender": "Female", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 20, "n_ref_uni": 27, "n_ref": 79, "n_ref_all": 98, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3487, "formula_len_all_1": 1506, "len_all": 124274, "len_all_1": 67479, "len_abs": 1120, "len_title": 128, "len_sents": 44790, "len_sents_1": 31395, "n_sents": 316, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1121, "title": "OD-Stega: LLM-Based Near-Imperceptible Steganography via Optimized Distributions", "abs": "We consider coverless steganography where a Large Language Model (LLM) drives an arithmetic coding decoder to generate stego-text. An efficient method should embed secret message bits in as few language tokens as possible, while still keeping the stego-text natural and fluent. We show that on the individual token level, this problem is mathematically equivalent to maximizing the entropy of a replacement probability distribution of the next token generation, subject to a constraint on the KL divergence between the chosen probability distribution and the original distribution given by the LLM. A closed-form solution is provided for the optimization problem, which can be computed efficiently. Several important practical issues are also tackled: 1) The combination of the optimized distribution and the vocabulary truncating technique is considered, 2) An often-overlooked tokenization mismatch issue is resolved with a simple prompt selection approach, and 3) The combination of the optimized distribution with other sequence-level selection heuristics to further enhance the efficiency and reliability is studied.", "keywords": [ "LLLM", "steganography", "arithmetic coding" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "IQN4XnIEhL", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Ahmed Hussen Abdelaziz", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Alexander Rudnicky", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Emeritus" }, { "name": "Li-Wei Chen", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Takuya Higuchi", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Zakaria Aldeneh", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 15, "n_ref_uni": 42, "n_ref": 74, "n_ref_all": 123, "n_fig": 4, "n_tab": 10, "L_tab": 1469, "n_element_tab": 197, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 914, "n_element_tab_1": 132, "formula_len_all": 2870, "formula_len_all_1": 1242, "len_all": 179965, "len_all_1": 75520, "len_abs": 1184, "len_title": 110, "len_sents": 45986, "len_sents_1": 31748, "n_sents": 387, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1190, "title": "A Variational Approach for Generative Speech Language Modeling", "abs": "The success of large language models in text processing has inspired their adaptation to speech modeling. However, because speech is continuous and complex, it is often discretized into tokens derived from self-supervised speech models. These speech tokens typically focus on the linguistic aspects of speech and neglect its paralinguistic content. As a result, autoregressive models trained on these tokens may generate speech with suboptimal naturalness. Previous methods attempted to address this limitation by adding pitch features to speech tokens prior to autoregressive modeling. However, pitch alone cannot fully represent the range of paralinguistic attributes, and selecting the right features requires careful hand-engineering. To tackle this issue, we propose a variational approach that automatically learns to encode these continuous speech attributes to enhance the speech tokens. Our proposed approach eliminates the need for manual paralinguistic feature selection and extraction. Moreover, we demonstrate that our proposed approach maintains or improves speech language modeling performance and enhances the naturalness of generated speech compared to baseline approaches.", "keywords": [ "Generative Spoken Language Modeling;Speech Language Model" ], "rating_list": [ 6, 3, 8, 5 ], "soundness_list": [ 3, 1, 4, 3 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 2, 1, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "IQCwmB63Fd", "primary_area": "datasets and benchmarks", "n_author": 1, "author_info_list": [ { "name": "Zi Yang", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 40, "n_formula_1": 10, "n_ref_uni": 28, "n_ref": 64, "n_ref_all": 101, "n_fig": 16, "n_tab": 8, "L_tab": 3472, "n_element_tab": 534, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 36, "n_element_tab_1": 9, "formula_len_all": 4403, "formula_len_all_1": 1022, "len_all": 259409, "len_all_1": 69148, "len_abs": 1283, "len_title": 139, "len_sents": 78957, "len_sents_1": 33611, "n_sents": 626, "n_sents_1": 268, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1307, "title": "Retrieval Or Holistic Understanding? Dolce: Differentiate Our Long Context Evaluation Tasks", "abs": "We argue that there are two major distinct capabilities in long context understanding: retrieval and holistic understanding. Understanding and further improving LLMs' long context capabilities would not be possible without knowing the tasks' focus categories. We aim to automatically identify retrieval focused and holistic understanding focused problems from suites of benchmarks and quantitatively measure the difficulty within each focus. In this paper, we present the Dolce framework, which parameterizes each problem by $\\lambda$ (complexity) and $k$ (redundancy) and assigns to one of five predefined focus categories. We propose to sample short contexts from the full context and estimate the probability an LLM solves the problem using the sampled spans. To find the $\\lambda$ and $k$ for each problem, we further propose a mixture model of a non-parametric background noise component and a parametric/non-parametric hybrid oracle component, where we derive the probability functions parameterized by \u03bb and k for both the correct-or-wrong (COW) scenario and the partial-point-in-grading (PIG) scenario. Our proposed methods can identify 0% to 67% of the problems are retrieval focused and 0% to 90% of the problems are holistic understanding focused across 44 existing long context evaluation tasks.", "keywords": [ "long context", "benchmark", "holistic understanding" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 2, 1, 2, 3 ] }, { "paper_id": "IQ0BBfbYR2", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Christian Hellert", "gender": "Male", "institution": "Continental AG", "country": "DE", "position": "AI Engineer" }, { "name": "Franz Motzkus", "gender": "Male", "institution": "Otto-Friedrich Universit\u00e4t Bamberg", "country": "", "position": "PhD student" }, { "name": "Ute Schmid", "gender": "Female", "institution": "University of Bamberg", "country": "", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 16, "n_ref": 37, "n_ref_all": 52, "n_fig": 6, "n_tab": 3, "L_tab": 1217, "n_element_tab": 76, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 576, "n_element_tab_1": 68, "formula_len_all": 895, "formula_len_all_1": 895, "len_all": 103938, "len_all_1": 56205, "len_abs": 1464, "len_title": 117, "len_sents": 28141, "len_sents_1": 27022, "n_sents": 169, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1457, "title": "CoLa-DCE \u2013 Concept-guided Latent Diffusion Counterfactual Explanations", "abs": "Recent advancements in generative AI have introduced novel prospects and prac-\ntical implementations. Especially diffusion models show their strength in gener-\nating diverse and, at the same time, realistic features, positioning them well for\ngenerating counterfactual explanations for computer vision models. Answering\n\u201cwhat if\u201d questions of what needs to change to make an image classifier change\nits prediction, counterfactual explanations align well with human understanding\nand consequently help in making model behavior more comprehensible. Current\nmethods succeed in generating authentic counterfactuals, but lack transparency as\nfeature changes are not directly perceivable. To address this limitation, we intro-\nduce Concept-guided Latent Diffusion Counterfactual Explanations (CoLa-DCE).\nCoLa-DCE generates concept-guided counterfactuals for any classifier with a high\ndegree of control regarding concept selection and spatial conditioning. The coun-\nterfactuals comprise an increased granularity through minimal feature changes.\nThe reference feature visualization ensures better comprehensibility, while the\nfeature localization provides increased transparency of \u201cwhere\u201d changed \u201cwhat\u201d.\nWe demonstrate the advantages of our approach in minimality and comprehen-\nsibility across multiple image classification models and datasets and provide in-\nsights into how our CoLa-DCE explanations help comprehend model errors like\nmisclassification cases.", "keywords": [ "Counterfactual Explanations", "Concept-based Explanations", "Diffusion-based Counterfactuals", "Counterfactual Image Generation" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 1, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "IOkYP5ZxO5", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Huazhe Xu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Shuiqi Cheng", "gender": "unknown", "institution": "University of Hong Kong", "country": "HK", "position": "Researcher" }, { "name": "Tianming Wei", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Undergrad student" }, { "name": "Tianxing Zhou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yuanhang Zhang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Yufeng Tian", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yuan Zhecheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Zixian Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 33, "n_ref": 63, "n_ref_all": 106, "n_fig": 15, "n_tab": 5, "L_tab": 932, "n_element_tab": 43, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 2, "n_element_tab_1": 1, "formula_len_all": 1107, "formula_len_all_1": 578, "len_all": 157860, "len_all_1": 59674, "len_abs": 1136, "len_title": 129, "len_sents": 45495, "len_sents_1": 28984, "n_sents": 344, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1179, "title": "ViTaS: Visual Tactile Soft Fusion Contrastive Learning for Reinforcement Learning", "abs": "Tactile information plays a crucial role in human manipulation tasks and has recently garnered increasing attention in robotic manipulation. However, existing approaches struggle to effectively integrate visual and tactile information, resulting in suboptimal performance. In this paper, we present **ViTaS**, a simple yet effective framework that incorporates both visual and tactile information to guide an agent's behavior. We introduce _Soft Fusion Contrastive Learning_, an advanced version of conventional contrastive learning method, to enhance the fusion of these two modalities, and adopt a CVAE module to utilize complementary information within visuo-tactile representation. We conduct comprehensive experiments, including $\\mathbf{9}$ tasks in simulation environment, across $\\mathbf{5}$ different benchmarks, to compare ViTaS with existing baselines. The results demonstrate that ViTaS achieves state-of-the-art performance, with an average improvement of $\\mathbf{51}$%. Furthermore, our method significantly enhances sample efficiency while maintaining minimal parameters, underscoring the effectiveness of our approach. The code will be released upon acceptance.", "keywords": [ "visuo-tactile representation learning", "reinforcement learning", "contrastive learning" ], "rating_list": [ 5, 6, 3, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "IOg47mg74i", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Huawen Feng", "gender": "Male", "institution": "South China University of Technology", "country": "", "position": "PhD student" }, { "name": "Junhao Zheng", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Qianli Ma", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "ZekunYao", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 25, "n_ref": 60, "n_ref_all": 74, "n_fig": 2, "n_tab": 5, "L_tab": 1564, "n_element_tab": 106, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1538, "n_element_tab_1": 105, "formula_len_all": 1345, "formula_len_all_1": 1344, "len_all": 192015, "len_all_1": 62691, "len_abs": 1547, "len_title": 153, "len_sents": 28947, "len_sents_1": 27626, "n_sents": 222, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 1473, "title": "Training Large Language Models for Retrieval-Augmented Question Answering through Backtracking Correction", "abs": "Despite recent progress in Retrieval-Augmented Generation (RAG) achieved by large language models (LLMs), retrievers often recall uncorrelated documents, regarded as \"noise\" during subsequent text generation. To address this, some methods train LLMs to distinguish between relevant and irrelevant documents using labeled data, enabling them to select the most likely relevant ones as context. However, they remain sensitive to noise, as LLMs can easily make mistakes when the selected document is noisy. Some approaches increase the number of referenced documents and train LLMs to perform stepwise reasoning when presented with multiple documents. Unfortunately, these methods rely on extensive and diverse annotations to ensure generalization, which is both challenging and costly. In this paper, we propose **Backtracking Correction** to address these limitations. Specifically, we reformulate stepwise RAG into a multi-step decision-making process. Starting from the final step, we optimize the model through error sampling and self-correction, and then backtrack to the previous state iteratively. In this way, the model's learning scheme follows an easy-to-hard progression: as the target state moves forward, the context space decreases while the decision space increases. Experimental results demonstrate that **Backtracking Correction** enhances LLMs' ability to make complex multi-step assessments, improving the robustness of RAG in dealing with noisy documents.", "keywords": [ "RAG", "CoT", "Self-play", "Reinforcement Learning" ], "rating_list": [ 6, 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 4, 2 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 3, 5, 3, 4, 3 ] }, { "paper_id": "INzc851YaM", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Ankur Deshwal", "gender": "Male", "institution": "Samsung", "country": "", "position": "Researcher" }, { "name": "Nishit Bhardwaj", "gender": "Male", "institution": "Samsung", "country": "IN", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 40, "n_ref": 52, "n_ref_all": 78, "n_fig": 9, "n_tab": 3, "L_tab": 2387, "n_element_tab": 75, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 2387, "n_element_tab_1": 75, "formula_len_all": 54, "formula_len_all_1": 0, "len_all": 113357, "len_all_1": 59296, "len_abs": 1205, "len_title": 73, "len_sents": 36483, "len_sents_1": 28370, "n_sents": 244, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1209, "title": "Enhancing Multi-Objective Offline RL with Adaptive Preference Integration", "abs": "Multi-objective reinforcement learning (MORL) is crucial for real-world applications where multiple conflicting goals must be optimized, such as in healthcare or autonomous systems. Offline MORL extends these benefits by using pre-collected datasets, allowing for effective learning without continuous interaction with the environment. However, existing offline MORL algorithms often struggle with scaling across large preference spaces and handling unknown preferences during evaluation. To address these challenges, we propose the Preference-Attended Multi-Objective Decision Transformer (PA-MODT), a novel architecture that integrates a preference-attention block with a modular transformer structure. This design enables effective generalization over different preferences and trajectories, providing a more robust approach to generating optimal Pareto fronts. We tested PA-MODT on five D4MORL datasets with millions of trajectories representing various objectives and found that it consistently outperforms existing models, achieving Pareto fronts that align closely with behavioral policy. This demonstrates PA-MODT's potential to effectively manage complex multi-objective reinforcement learning tasks.", "keywords": [ "Reinforcement Learning", "Multi-objective Optimization", "Offline Reinforcement Learning", "Decision Transformer" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "INyi7qUdjZ", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Alex Nguyen", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Gautam Reddy", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 67, "n_formula_1": 23, "n_ref_uni": 16, "n_ref": 42, "n_ref_all": 72, "n_fig": 19, "n_tab": 2, "L_tab": 891, "n_element_tab": 58, "n_fig_1": 16, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2862, "formula_len_all_1": 953, "len_all": 164410, "len_all_1": 80241, "len_abs": 1573, "len_title": 164, "len_sents": 52904, "len_sents_1": 32086, "n_sents": 459, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 115, "L_abs": 1267, "title": "Differential learning kinetics govern the transition from memorization to generalization during in-context learning", "abs": "Transformers exhibit in-context learning (ICL): the ability to use novel information presented in the context without additional weight updates. Recent work shows that ICL emerges when models are trained on a sufficiently diverse set of tasks and the transition from memorization to generalization is sharp with increasing task diversity. One interpretation is that a network's limited capacity to memorize favors generalization. Here, we examine the mechanistic underpinnings of this transition using a small transformer applied to a synthetic ICL task. Using theory and experiment, we show that the sub-circuits that memorize and generalize can be viewed as largely independent. The relative *rates* at which these sub-circuits learn explains the transition from memorization to generalization, rather than capacity constraints. We uncover a memorization scaling law, which determines the task diversity threshold at which the network generalizes. The theory quantitatively explains a variety of other ICL-related phenomena, including the long-tailed distribution of when ICL is acquired, the bimodal behavior of solutions close to the task diversity threshold, the influence of contextual and data distributional statistics on ICL, and the transient nature of ICL.", "keywords": [ "in-context learning", "mechanistic interpretability", "small transformers", "memorization" ], "rating_list": [ 8, 10, 3, 8 ], "soundness_list": [ 4, 4, 1, 3 ], "presentation_list": [ 4, 4, 2, 3 ], "contribution_list": [ 3, 4, 1, 3 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "INqLJwqUmc", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Fang Chen", "gender": "unknown", "institution": "University of Technology Sydney (UTS)", "country": "AU", "position": "Full Professor" }, { "name": "Jiahao Huang", "gender": "unknown", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "MS student" }, { "name": "Jianlong Zhou", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Jiayu Zhang", "gender": "Male", "institution": "Suzhou Yierqi", "country": "", "position": "Researcher" }, { "name": "NAN YANG", "gender": "Female", "institution": "University of Sydney", "country": "AU", "position": "Researcher" }, { "name": "Zhibo Jin", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Zhu Zhiyu", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "MS student" } ], "n_formula": 23, "n_formula_1": 12, "n_ref_uni": 26, "n_ref": 43, "n_ref_all": 49, "n_fig": 3, "n_tab": 6, "L_tab": 4971, "n_element_tab": 440, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1478, "n_element_tab_1": 198, "formula_len_all": 2110, "formula_len_all_1": 802, "len_all": 112800, "len_all_1": 62316, "len_abs": 1401, "len_title": 146, "len_sents": 36529, "len_sents_1": 29220, "n_sents": 235, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 98, "L_abs": 1411, "title": "Narrowing Information Bottleneck Theory for Multimodal Image-Text Representations Interpretability", "abs": "The task of identifying multimodal image-text representations has recently received increasing attention, particularly with models like CLIP (Contrastive Language-Image Pretraining), which excel at learning complex associations between images and text. Despite these advancements, ensuring the interpretability of such models remains crucial for their safe application in real-world scenarios, such as healthcare. Numerous interpretability methods have been developed for unimodal tasks; however, these approaches often fail to transfer effectively to multimodal tasks due to fundamental differences in representation. Bottleneck methods, widely recognized in information theory, have been applied to enhance CLIP\u2019s interpretability, but they often suffer from strong assumptions or inherent randomness. To address these limitations, we introduce the Narrowing Information Bottleneck Theory, a novel approach that re-engineers the bottleneck method from the ground up. This theory is designed to satisfy modern attribution axioms, offering a robust solution for improving the interpretability of multimodal models. In our experiments, compared to state-of-the-art methods, our approach improves image interpretability by an average of 9\\%, text interpretability by an average of 58.83\\%, and increases processing speed by 63.95\\%. Our code is publicly available at: https://anonymous.4open.science/r/NIB-DBCD/.", "keywords": [ "Interpretability", "CLIP" ], "rating_list": [ 5, 5, 3, 8 ], "soundness_list": [ 1, 2, 1, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "INow59Vurm", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Minghui Ouyang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Qian Li", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Researcher" }, { "name": "Qingjiang Shi", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Full Professor" }, { "name": "Ruoyu Sun", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Tian Ding", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Researcher" }, { "name": "Yuyi Wang", "gender": "Male", "institution": "CRRC Zhuzhou Institute, China", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 45, "n_formula_1": 34, "n_ref_uni": 32, "n_ref": 58, "n_ref_all": 61, "n_fig": 1, "n_tab": 3, "L_tab": 519, "n_element_tab": 31, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 286, "n_element_tab_1": 14, "formula_len_all": 4293, "formula_len_all_1": 3104, "len_all": 115403, "len_all_1": 57121, "len_abs": 1696, "len_title": 151, "len_sents": 30308, "len_sents_1": 23431, "n_sents": 281, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 1708, "title": "Towards Explaining the Power of Constant-depth Graph Neural Networks for Structured Linear Programming", "abs": "Graph neural networks (GNNs) have recently emerged as powerful tools for solving complex optimization problems, often being employed to approximate solution mappings. Empirical evidence shows that even shallow GNNs (with fewer than ten layers) can achieve strong performance in predicting optimal solutions to linear programming (LP) problems. This finding is somewhat counter-intuitive, as LPs are global optimization problems, while shallow GNNs predict based on local information. Although previous theoretical results suggest that GNNs have the expressive power to solve LPs, they require deep architectures whose depth grows at least polynomially with the problem size, and thus leave the underlying principle of this empirical phenomenon still unclear. In this paper, we examine this phenomenon through the lens of distributed computing and average-case analysis. We establish that the expressive power of GNNs for LPs is closely related to well-studied distributed algorithms for LPs. Specifically, we show that any $d$-round distributed LP algorithm can be simulated by a $d$-depth GNN, and vice versa. In particular, by designing a new distributed LP algorithm and then unrolling it, we prove that constant-depth, constant-width GNNs suffice to solve sparse binary LPs effectively. Here, in contrast with previous analyses focusing on worst-case scenarios, in which we show that GNN depth must increase with problem size by leveraging an impossibility result about distributed LP algorithms, our analysis shifts the focus to the average-case performance, and shows that constant GNN depth then becomes sufficient no matter how large the problem size is. Our theory is validated by numerical results.", "keywords": [ "Learning to optimize", "shallow graph neural network", "linear programming", "distributed algorithm", "average-case analysis" ], "rating_list": [ 8, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "INe4otjryz", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Qixun Wang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xianghua Ying", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Yifei Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yisen Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 28, "n_formula_1": 4, "n_ref_uni": 25, "n_ref": 64, "n_ref_all": 113, "n_fig": 14, "n_tab": 3, "L_tab": 161, "n_element_tab": 3, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3410, "formula_len_all_1": 333, "len_all": 141297, "len_all_1": 59956, "len_abs": 1293, "len_title": 119, "len_sents": 53775, "len_sents_1": 28519, "n_sents": 394, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1311, "title": "Can In-context Learning Really Generalize to Out-of-distribution Tasks?", "abs": "In this work, we explore the mechanism of in-context learning (ICL) on out-of-distribution (OOD) tasks that were not encountered during training. To achieve this, we conduct synthetic experiments where the objective is to learn OOD mathematical functions through ICL using a GPT-2 model. We reveal that Transformers may struggle to learn OOD task functions through ICL. Specifically, ICL performance resembles implementing a function within the pretraining hypothesis space and optimizing it with gradient descent based on the in-context examples. Additionally, we investigate ICL's well-documented ability to learn unseen abstract labels in context. We demonstrate that such ability only manifests in the scenarios without distributional shifts and, therefore, may not serve as evidence of new-task-learning ability. Furthermore, we assess ICL's performance on OOD tasks when the model is pretrained on multiple tasks. Both empirical and theoretical analyses demonstrate the existence of the \\textbf{low-test-error preference} of ICL, where it tends to implement the pretraining function that yields low test error in the testing context. We validate this through numerical experiments. This new theoretical result, combined with our empirical findings, elucidates the mechanism of ICL in addressing OOD tasks.", "keywords": [ "Large language models", "In-context Learning", "Out-of-distribution Generalization" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 2, 4, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "INXZOxYsLd", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Qi Zhou", "gender": "Female", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Wenjian Luo", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yubo Tang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Zipeng Ye", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 9, "n_ref_uni": 49, "n_ref": 74, "n_ref_all": 136, "n_fig": 13, "n_tab": 12, "L_tab": 6151, "n_element_tab": 515, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3349, "n_element_tab_1": 163, "formula_len_all": 1216, "formula_len_all_1": 534, "len_all": 197523, "len_all_1": 73312, "len_abs": 1212, "len_title": 92, "len_sents": 60179, "len_sents_1": 33293, "n_sents": 417, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 1217, "title": "Safeguard User Privacy in LLM Cloud Services", "abs": "Large language models (LLMs) have witnessed substantial growth in recent years. To leverage convenient LLM cloud services, users are inevitable to upload their prompts. Further, for tasks such as translation, reading comprehension, and summarization, related files or contexts are inherently required to be uploaded, whether they contain user privacy or not. Despite the rapid advancement of LLM capability, there has been a scarcity of research focusing on preserving user privacy during inference. To this end, this paper conducts a comprehensive study in this domain. Firstly, we demonstrate that (1) the embedding space of tokens is remarkably sparse, and (2) LLMs primarily function in the orthogonal subspace of embedding space, these two factors making privacy extremely vulnerable. Then, we analyze the structural characteristics of LLMs and design a distributed privacy-preserving inference paradigm which can effectively resist privacy attacks. Finally, we conduct a comprehensive evaluation of the defended models on mainstream tasks and find that low-bit quantization techniques can be well combined with our inference paradigm, achieving a balance between privacy, utility, and runtime memory efficiency.", "keywords": [ "large language model", "privacy leakage", "defense", "cloud service" ], "rating_list": [ 6, 3, 3, 6, 5 ], "soundness_list": [ 2, 1, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 2, 4, 4, 3, 4 ] }, { "paper_id": "INFfvQArFY", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Di Wang", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Keyuan Cheng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Lijie Hu", "gender": "Female", "institution": "KAUST", "country": "SA", "position": "PhD student" }, { "name": "Yongxiang Li", "gender": "unknown", "institution": "South China University of Technologys", "country": "CN", "position": "Undergrad student" }, { "name": "Zhuoran Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zijian Kan", "gender": "unknown", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 26, "n_formula_1": 12, "n_ref_uni": 33, "n_ref": 85, "n_ref_all": 116, "n_fig": 9, "n_tab": 13, "L_tab": 3982, "n_element_tab": 382, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 913, "n_element_tab_1": 36, "formula_len_all": 1704, "formula_len_all_1": 725, "len_all": 202285, "len_all_1": 71846, "len_abs": 1131, "len_title": 118, "len_sents": 62485, "len_sents_1": 33432, "n_sents": 448, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1136, "title": "Locate-then-edit for Multi-hop Factual Recall under Knowledge Editing", "abs": "The locate-then-edit paradigm has shown significant promise for knowledge editing (KE) in Large Language Models (LLMs). While previous methods perform well on single-hop fact recall tasks, they consistently struggle with multi-hop factual recall tasks involving newly edited knowledge. In this paper, leveraging tools in mechanistic interpretability, we first identify that in multi-hop tasks, LLMs tend to retrieve implicit subject knowledge from deeper MLP layers, unlike single-hop tasks, which rely on earlier layers. This distinction explains the poor performance of current methods in multi-hop queries, as they primarily focus on editing shallow layers, leaving deeper layers unchanged. To address this, we propose IFMET, a novel locate-then-edit KE approach designed to edit both shallow and deep MLP layers. IFMET employs multi-hop editing prompts and supplementary sets to locate and modify knowledge across different reasoning stages. Experimental results demonstrate that IFMET significantly improves performance on multi-hop factual recall tasks, effectively overcoming the limitations of previous locate-then-edit methods.", "keywords": [ "Knowledge Editing" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "IMWYNVBHob", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Huaming Wu", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" }, { "name": "Mengzhou Gao", "gender": "Female", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "Associate Professor" }, { "name": "Pengfei Jiao", "gender": "Male", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "Full Professor" }, { "name": "Xuan Guo", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" }, { "name": "Yifan Lu", "gender": "Male", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "Undergrad student" }, { "name": "Yiwei Wang", "gender": "Male", "institution": "University of California, Merced", "country": "US", "position": "Assistant Professor" }, { "name": "Ziyun Zou", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "MS student" } ], "n_formula": 46, "n_formula_1": 24, "n_ref_uni": 37, "n_ref": 59, "n_ref_all": 74, "n_fig": 5, "n_tab": 5, "L_tab": 3189, "n_element_tab": 306, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2150, "n_element_tab_1": 165, "formula_len_all": 3386, "formula_len_all_1": 1177, "len_all": 140811, "len_all_1": 62768, "len_abs": 1395, "len_title": 135, "len_sents": 40734, "len_sents_1": 27936, "n_sents": 331, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1401, "title": "Robust Heterogeneous Graph Neural Network Explainer with Graph Information Bottleneck", "abs": "Explaining the prediction process of Graph Neural Network (GNN) is crucial for enhancing network transparency. However, real-world networks are predominantly heterogeneous and often beset with noise. The presence of intricate relationships in heterogeneous graphs necessitates a consideration of semantics during the explanation process, while mitigating the impact of noise remains unexplored. For GNN explainers heavily reliant on graph structure and raw features, erroneous predictions may lead to misguided explanations under the influence of noise. To address these challenges, we propose a Robust Heterogeneous Graph Neural Network Explainer with Graph Information Bottleneck, named RHGIB. We theoretically analyze the power of different heterogeneous GNN architectures on the propagation of noise information and exploit denoising variational inference. Specifically, we infer the latent distributions of both graph structure and features to alleviate the influence of noise. Subsequently, we incorporate heterogeneous edge types into the generation process of explanatory subgraph and utilize Graph Information Bottleneck framework for optimization, allowing the Explainer to learn heterogeneous semantics while enhancing robustness. Extensive experiments on multiple real-world heterogeneous graph datasets demonstrate the superior performance of RHGIB compared to state-of-the-art baselines.", "keywords": [ "Trustworthy", "Robust", "Graph Neural Network" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "ILSZZNlbqw", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Danial Dervovic", "gender": "Male", "institution": "J.P. Morgan Chase", "country": "GB", "position": "Researcher" }, { "name": "Ivan Brugere", "gender": "Male", "institution": "J.P. Morgan", "country": "US", "position": "Researcher" }, { "name": "Jiliang Tang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Full Professor" }, { "name": "Saumitra Mishra", "gender": "Male", "institution": "J.P. Morgan Chase", "country": "", "position": "Researcher" }, { "name": "Wenzhuo Tang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Yuying Xie", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Associate Professor" }, { "name": "Mao Haitao", "gender": "Male", "institution": "Michigan State University", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 85, "n_ref": 167, "n_ref_all": 188, "n_fig": 4, "n_tab": 18, "L_tab": 8782, "n_element_tab": 1011, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 4419, "n_element_tab_1": 433, "formula_len_all": 813, "formula_len_all_1": 790, "len_all": 222445, "len_all_1": 75548, "len_abs": 1255, "len_title": 113, "len_sents": 45915, "len_sents_1": 30960, "n_sents": 366, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1261, "title": "Cross-Domain Graph Data Scaling: A Showcase with Diffusion Models", "abs": "Models for natural language and images benefit from data scaling behavior: the more data fed into the model, the better they perform. This 'better with more' phenomenon enables the effectiveness of large-scale pre-training on vast amounts of data. However, current graph pre-training methods struggle to scale up data due to heterogeneity across graphs. To achieve effective data scaling, we aim to develop a general model that is able to capture diverse data patterns of graphs and can be utilized to adaptively help the downstream tasks. To this end, we propose UniAug, a universal graph structure augmentor built on a diffusion model. We first pre-train a discrete diffusion model on thousands of graphs across domains to learn the graph structural patterns. In the downstream phase, we provide adaptive enhancement by conducting graph structure augmentation with the help of the pre-trained diffusion model via guided generation. By leveraging the pre-trained diffusion model for structure augmentation, we consistently achieve performance improvements across various downstream tasks in a plug-and-play manner. To the best of our knowledge, this study represents the first demonstration of a data-scaling graph structure augmentor on graphs across domains.", "keywords": [ "Graph Foundation Model", "Graph Data Augmentation", "Diffusion Models" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "IL85Ebjg9j", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Belinda Gabbe", "gender": "Female", "institution": "Monash University", "country": "", "position": "Full Professor" }, { "name": "Joanna Dipnall", "gender": "Female", "institution": "Monash University", "country": "", "position": "Researcher" }, { "name": "Jueqing Lu", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Lan Du", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Associate Professor" }, { "name": "Myong Chol Jung", "gender": "Male", "institution": "Monash University", "country": "", "position": "PhD student" }, { "name": "Wray Buntine", "gender": "Male", "institution": "VinUniversity", "country": "", "position": "Full Professor" } ], "n_formula": 38, "n_formula_1": 14, "n_ref_uni": 32, "n_ref": 67, "n_ref_all": 101, "n_fig": 4, "n_tab": 19, "L_tab": 6775, "n_element_tab": 643, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 3406, "n_element_tab_1": 198, "formula_len_all": 3273, "formula_len_all_1": 788, "len_all": 170389, "len_all_1": 63318, "len_abs": 1329, "len_title": 107, "len_sents": 51135, "len_sents_1": 27149, "n_sents": 347, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1335, "title": "Navigating Conflicting Views: Harnessing Trust for Learning", "abs": "Resolving conflicts is essential to make the decisions of multi-view classification more reliable. Much research has been conducted on learning consistent and informative representations among different views, often assuming that all views are equally important and perfectly aligned. However, real-world multi-view data may not always conform to these assumptions, as some views may express distinct information. To address this issue, we develop a computational trust-based discounting method to enhance the existing Evidential Multi-view framework in scenarios where conflicts between different views may arise. Its belief fusion process considers the reliability of predictions made by individual views via an instance-wise probability-sensitive trust discounting mechanism. We evaluate our method on six real-world datasets, using Top-1 Accuracy, Fleiss\u2019 Kappa, and a new metric called Multi-View Agreement with Ground Truth that takes into consideration the ground truth labels, to measure the reliability of the prediction. We also evaluate whether uncertainty measures can effectively indicate prediction correctness by calculating the AUROC. The experimental results show that computational trust can effectively resolve conflicts, paving the way for more reliable multi-view classification models in real-world applications.", "keywords": [ "Multi-view Learning", "Conflict Multi-view Learning", "Reliable Multi-view Learning" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 5, 4, 2 ] }, { "paper_id": "IKeYXtjvPL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Bin Chen", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Hang Guo", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Ruisheng Luo", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Shu-Tao Xia", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Tao Dai", "gender": "Male", "institution": "Department of Software Engineering, Shenzhen University", "country": "CN", "position": "Assistant Professor" }, { "name": "YIZHEN GUO", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 50, "n_ref": 92, "n_ref_all": 101, "n_fig": 2, "n_tab": 6, "L_tab": 2277, "n_element_tab": 225, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1063, "n_element_tab_1": 117, "formula_len_all": 310, "formula_len_all_1": 341, "len_all": 140641, "len_all_1": 59314, "len_abs": 1264, "len_title": 110, "len_sents": 31636, "len_sents_1": 26949, "n_sents": 232, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1268, "title": "RouGE: Learning Gated Experts for Segment Anything in the Wild", "abs": "Segment anything model (SAM) and its variants have recently shown promising performance as foundation models. However, existing SAM-based models can only handle scenarios seen during training, and usually suffer unstable performance when transferring to real-world unseen data, such as low-light, rainy or blurred images, which is crucial for applications such as autopilot. Therefore, adapting SAM-based models for real-world degradation while not impairing its original ability remains an open challenge. In this work, we propose a novel gated Mixture-of-Experts (MoE) structure, called RouGE, to improve the robustness of SAM-based models. Specifically, RouGE uses multiple lightweight probability gates to decompose complex real-world image conditions and judge whether the feature needs to be adjusted as well as to what extent the adjustment needs to be done, then handle them differently with a set of low-rank experts. During the inference stage, RouGE processes input images in a completely blind manner thus improving the model's performance in real-world scenarios. Extensive experiments demonstrate that RouGE consistently achieves state-of-the-art results on both degraded and clean images compared with other methods while tuning only 1.5% of parameters.", "keywords": [ "PEFT\uff0creal-world degradation\uff0c segment anything model\uff0c robustness enhancement" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "IK7l0CqZuH", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Jingwen Ye", "gender": "Female", "institution": "National University of Singapore", "country": "", "position": "Postdoc" }, { "name": "Ruonan Yu", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Songhua Liu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zigeng Chen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Xinchao Wang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 18, "n_ref": 40, "n_ref_all": 56, "n_fig": 2, "n_tab": 8, "L_tab": 2267, "n_element_tab": 132, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 3069, "n_element_tab_1": 173, "formula_len_all": 672, "formula_len_all_1": 641, "len_all": 115222, "len_all_1": 66212, "len_abs": 1513, "len_title": 114, "len_sents": 31773, "len_sents_1": 28631, "n_sents": 227, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1518, "title": "Heavy Labels Out! Dataset Distillation with Label Space Lightening", "abs": "Dataset distillation or condensation aims to condense a large-scale training dataset into a much smaller synthetic one such that the training performance of distilled and original sets on neural networks are similar. Although the number of training samples can be reduced substantially, current state-of-the-art methods heavily rely on enormous soft labels to achieve satisfactory performance. As a result, the required storage can be comparable even to original datasets, especially for large-scale ones. To solve this problem, instead of storing these heavy labels, we propose a novel label-lightening framework termed HeLlO aiming at effective image-to-label projectors, with which synthetic labels can be directly generated online from synthetic images. Specifically, to construct such projectors, we leverage prior knowledge in open-source foundation models, e.g., CLIP, and introduce a LoRA-like fine-tuning strategy to mitigate the gap between pre-trained and target distributions, so that original models for soft-label generation can be distilled into a group of low-rank matrices. Moreover, an effective image optimization method is proposed to further mitigate the potential error between the original and distilled label generators. Extensive experiments demonstrate that with only about 0.001% of the original storage required for a complete set of soft labels, we achieve comparable performance to current state-of-the-art dataset distillation methods on large-scale datasets. Our code will be available.", "keywords": [ "dataset distillation", "efficient training", "label space lightening" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "IJiTI0fB0e", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Amin Gohari", "gender": "unknown", "institution": "The Chinese University of Hong Kong", "country": "", "position": "Associate Professor" }, { "name": "Azim Ospanov", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Farzan Farnia", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Mohammad Jalali", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "MPhil Student" } ], "n_formula": 45, "n_formula_1": 19, "n_ref_uni": 34, "n_ref": 76, "n_ref_all": 97, "n_fig": 14, "n_tab": 2, "L_tab": 235, "n_element_tab": 1, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5488, "formula_len_all_1": 1056, "len_all": 216540, "len_all_1": 72210, "len_abs": 2988, "len_title": 140, "len_sents": 60920, "len_sents_1": 31026, "n_sents": 680, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1739, "title": "An Information-Theoretic Approach to Diversity Evaluation of Prompt-based Generative Models", "abs": "Text-conditioned generation models are commonly evaluated based on the quality of the generated data and its alignment with the input text prompt. On the other hand, several applications of prompt-based generative models require sufficient diversity in the generated data to ensure the models' capability of generating image and video samples possessing a variety of features. However, the existing diversity metrics are designed for unconditional generative models, and thus cannot distinguish the diversity arising from variations in text prompts and that contributed by the generative model itself. In this work, our goal is to quantify the prompt-induced and model-induced diversity in samples generated by prompt-based models. Specifically, we propose the application of matrix-based information measures to address this task, decomposing the kernel-based entropy $H(X)$ of generated data $X$ into the sum of conditional entropy $H(X|T)$, given text variable $T$, and mutual information $I(X; T)$. We show that this information-theoretic approach decomposes the existing Vendi diversity score defined based on $H(X)$ into the product of the following two terms: 1) Conditional-Vendi score based on $H(X|T)$ to quantify the model-induced diversity, and 2) Information-Vendi score based on $I(X; T)$ to measure the statistical relevance between $X$ and prompt $T$. Our theoretical results provide an interpretation for this diversity quantification and show that the Conditional-Vendi score aggregates the Vendi scores within the modes of a mixture prompt distribution. We conduct several numerical experiments to show the correlation between the Conditional-Vendi score and the internal diversity of text-conditioned generative models.", "keywords": [ "Information measures", "diversity evaluation", "conditional generative models" ], "rating_list": [ 3, 3, 3, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 1, 2, 1, 3 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "IIzehISTBe", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Alon Oved", "gender": "unknown", "institution": "International Business Machines", "country": "IL", "position": "Researcher" }, { "name": "Avi Yaeli", "gender": "Male", "institution": "International Business Machines", "country": "IL", "position": "Researcher" }, { "name": "Ben wiesel", "gender": "Male", "institution": "IBM TJ Watson Research Center", "country": "IL", "position": "Researcher" }, { "name": "Ido Levy", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "MS student" }, { "name": "Sami Marreed", "gender": "Male", "institution": "University of Haifa", "country": "", "position": "Msc" }, { "name": "Segev Shlomov", "gender": "Male", "institution": "International Business Machines", "country": "IL", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 24, "n_ref": 52, "n_ref_all": 60, "n_fig": 3, "n_tab": 5, "L_tab": 987, "n_element_tab": 149, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1171, "n_element_tab_1": 133, "formula_len_all": 63, "formula_len_all_1": 79, "len_all": 133937, "len_all_1": 67696, "len_abs": 1378, "len_title": 134, "len_sents": 48506, "len_sents_1": 34663, "n_sents": 319, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1386, "title": "ST-WebAgentBench: A Benchmark for Evaluating Safety and Trustworthiness in Web Agents", "abs": "Recent advancements in Web agents have introduced novel architectures and benchmarks showcasing progress in autonomous web navigation and interaction. However, most existing benchmarks prioritize effectiveness and accuracy, overlooking factors like safety and trustworthiness\u2014both essential for deploying web agents in enterprise settings. \nWe present ST-WebAgentBench, a benchmark designed to evaluate web agents' safety and trustworthiness across six critical dimensions, essential for reliability in enterprise applications. This benchmark is grounded in a detailed framework that defines safe and trustworthy (ST) agent behavior. Our work extends WebArena with safety templates and evaluation functions to rigorously assess safety policy compliance. We introduce the Completion Under Policy to measure task success while adhering to policies, alongside the Risk Ratio, which quantifies policy violations across dimensions, providing actionable insights to address safety gaps.\nOur evaluation reveals that current SOTA agents struggle with policy adherence and cannot yet be relied upon for critical business applications. We open-source this benchmark and invite the community to contribute, with the goal of fostering a new generation of safer, more trustworthy AI agents. All code, data, environment reproduction resources, and video demonstrations are available at [blinded URL].", "keywords": [ "LLM Agent", "Web Agent", "Trust", "Benchmark" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "IIsTO4P3Ag", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Chenyan Xiong", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Associate Professor" }, { "name": "Graham Neubig", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Junpeng Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Tianyue Ou", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Wai Lam", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Professor" }, { "name": "Xiang Yue", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Yifan Song", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yuxiao Qu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "wenhu chen", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 98, "n_ref_all": 119, "n_fig": 7, "n_tab": 7, "L_tab": 5520, "n_element_tab": 573, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 377, "n_element_tab_1": 53, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 186756, "len_all_1": 48056, "len_abs": 1520, "len_title": 105, "len_sents": 49763, "len_sents_1": 21831, "n_sents": 363, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1122, "title": "Harnessing Webpage UIs for Text-Rich Visual Understanding", "abs": "Text-rich visual understanding\u2014the ability to interpret both textual content and visual elements within a scene\u2014is crucial for multimodal large language models (MLLMs) to effectively interact with structured environments. We propose leveraging webpage UIs as a naturally structured and diverse data source to enhance MLLMs\u2019 capabilities in this area. Existing approaches, such as rule-based extraction, multimodal model captioning, and rigid HTML parsing, are hindered by issues like noise, hallucinations, and limited generalization. To overcome these challenges, we introduce MultiUI, a dataset of 7.3 million samples spanning various UI types and tasks, structured using enhanced accessibility trees and task taxonomies. By scaling multimodal instructions from web UIs through LLMs, our dataset enhances generalization beyond web domains, significantly improving performance in document understanding, GUI comprehension, grounding, and advanced agent tasks. This demonstrates the potential of structured web data to elevate MLLMs\u2019 proficiency in processing text-rich visual environments and generalizing across domains.", "keywords": [ "Multimodal", "Instruction-tuning", "Large Language Model" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "IIVYiJ1ggK", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Hang Yu", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Senior Algorithm Expert" }, { "name": "Weiyao Lin", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Zhihao He", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zi Gong", "gender": "Male", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "", "position": "MS student" }, { "name": "jianguo Li", "gender": "Male", "institution": "Ant Group", "country": "", "position": "Director" }, { "name": "Liu Shizhan", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" } ], "n_formula": 28, "n_formula_1": 17, "n_ref_uni": 58, "n_ref": 188, "n_ref_all": 270, "n_fig": 16, "n_tab": 22, "L_tab": 9615, "n_element_tab": 837, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1298, "n_element_tab_1": 134, "formula_len_all": 3245, "formula_len_all_1": 1210, "len_all": 294112, "len_all_1": 74145, "len_abs": 1379, "len_title": 126, "len_sents": 79489, "len_sents_1": 33130, "n_sents": 611, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1395, "title": "Rodimus*: Breaking the Accuracy-Efficiency Trade-Off with Efficient Attentions", "abs": "Recent advancements in Transformer-based large language models (LLMs) have set new standards in natural language processing. However, the classical softmax attention incurs significant computational costs, leading to a $O(T)$ complexity for per-token generation, where $T$ represents the context length. This work explores reducing LLMs' complexity while maintaining performance by introducing Rodimus and its enhanced version, Rodimus$+$. Rodimus employs an innovative data-dependent tempered selection (DDTS) mechanism within a linear attention-based, purely recurrent framework, achieving significant accuracy while drastically reducing the memory usage typically associated with recurrent models. This method exemplifies semantic compression by maintaining essential input information with fixed-size hidden states. Building on this, Rodimus$+$ combines Rodimus with the innovative Sliding Window Shared-Key Attention (SW-SKA) in a hybrid approach, effectively leveraging the complementary semantic, token, and head compression techniques. Our experiments demonstrate that Rodimus$+$-1.6B, trained on 1 trillion tokens, achieves superior downstream performance against models trained on more tokens, including Qwen2-1.5B and RWKV6-1.6B, underscoring its potential to redefine the accuracy-efficiency balance in LLMs. Model code and pre-trained checkpoints will be available upon publication.", "keywords": [ "Large Language Models", "Fixed-Size Hidden States", "Data-Dependent Tempered Selection", "Sliding Window Shared-Key Attention" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "IIDFStLGQx", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Bowen Zhang", "gender": "Male", "institution": "Singapore Management University", "country": "", "position": "Researcher" }, { "name": "Haotian Lu", "gender": "Male", "institution": "58.com", "country": "CN", "position": "Researcher" }, { "name": "NA MA", "gender": "Female", "institution": "Yonsei University", "country": "", "position": "PhD student" }, { "name": "Qing Guo", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Tianlin Li", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Xiaofei Xie", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 51, "n_ref": 88, "n_ref_all": 94, "n_fig": 3, "n_tab": 5, "L_tab": 2930, "n_element_tab": 335, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2262, "n_element_tab_1": 298, "formula_len_all": 709, "formula_len_all_1": 708, "len_all": 146273, "len_all_1": 72063, "len_abs": 1760, "len_title": 138, "len_sents": 34274, "len_sents_1": 32334, "n_sents": 242, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1768, "title": "MAVIN: Multi-Action Video Generation with Diffusion Models via Transition Video Infilling", "abs": "Diffusion-based video generation has achieved significant progress, yet generating multiple actions that occur sequentially remains a formidable task. Directly generating a video with sequential actions can be extremely challenging due to the scarcity of fine-grained action annotations and the difficulty in establishing temporal semantic correspondences and maintaining long-term consistency. To tackle this, we propose an intuitive and straightforward solution: splicing multiple single-action video segments sequentially. The core challenge lies in generating smooth and natural transitions between these segments given the inherent complexity and variability of action transitions. We introduce MAVIN (Multi-Action Video INfilling model), designed to generate transition videos that seamlessly connect two given videos, forming a cohesive integrated sequence. MAVIN incorporates several innovative techniques to address challenges in the transition video infilling task. Firstly, a consecutive noising strategy coupled with variable-length sampling is employed to handle large infilling gaps and varied generation lengths. Secondly, boundary frame guidance (BFG) is proposed to address the lack of semantic guidance during transition generation. Lastly, a Gaussian filter mixer (GFM) dynamically manages noise initialization during inference, mitigating train-test discrepancy while preserving generation flexibility. Additionally, we introduce a new metric, CLIP-RS (CLIP Relative Smoothness), to evaluate temporal coherence and smoothness, complementing traditional quality-based metrics. Experimental results on horse and tiger scenarios demonstrate MAVIN's superior performance in generating smooth and coherent video transitions compared to existing methods.", "keywords": [ "video generation", "diffusion model", "generative AI" ], "rating_list": [ 5, 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 2, 4 ] }, { "paper_id": "II81zQUS1x", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Issei Sato", "gender": "unknown", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Naoya Hasegawa", "gender": "Male", "institution": "The University of Tokyo", "country": "", "position": "MS student" } ], "n_formula": 158, "n_formula_1": 51, "n_ref_uni": 43, "n_ref": 72, "n_ref_all": 119, "n_fig": 9, "n_tab": 12, "L_tab": 3173, "n_element_tab": 203, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 574, "n_element_tab_1": 8, "formula_len_all": 13427, "formula_len_all_1": 3284, "len_all": 215048, "len_all_1": 71332, "len_abs": 1591, "len_title": 143, "len_sents": 51034, "len_sents_1": 28099, "n_sents": 532, "n_sents_1": 279, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 893, "title": "Multiplicative Logit Adjustment Approximates Neural-Collapse-Aware Decision Boundary Adjustment", "abs": "Real-world data distributions are often highly skewed. This has spurred a growing body of research on long-tailed recognition, aimed at addressing the imbalance in training classification models. Among the methods studied, multiplicative logit adjustment (MLA) stands out as a simple and effective method. What theoretical foundation explains the effectiveness of this heuristic method?\nWe provide a justification for the effectiveness of MLA with the following two-step process. First, we develop a theory that adjusts optimal decision boundaries by estimating feature spread on the basis of neural collapse. Second, we demonstrate that MLA approximates this optimal method. Additionally, through experiments on long-tailed datasets, we illustrate the practical usefulness of MLA under more realistic conditions. We also offer experimental insights to guide the tuning of MLA hyperparameters.", "keywords": [ "long-tailed recognition", "imbalanced learning", "neural collapse", "logit adjustment", "multiplicative logit adjustment", "machine learning", "learning theory" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 2, 4, 3 ] }, { "paper_id": "IHqlU2J5ia", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Lizhen Qu", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Lecturer" }, { "name": "Reza Haffari", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" }, { "name": "Yuncheng Hua", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "Postdoc" }, { "name": "Zhuang Li", "gender": "Male", "institution": "Royal Melbourne Institute of Technology", "country": "AU", "position": "Lecturer" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 31, "n_ref": 50, "n_ref_all": 85, "n_fig": 1, "n_tab": 14, "L_tab": 4533, "n_element_tab": 569, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2117, "n_element_tab_1": 300, "formula_len_all": 351, "formula_len_all_1": 337, "len_all": 212721, "len_all_1": 73807, "len_abs": 1772, "len_title": 168, "len_sents": 70831, "len_sents_1": 34499, "n_sents": 530, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 120, "L_abs": 1936, "title": "An Empirical Study on Enhancing LLMs' Alignment Capabilities through Restyled In-Context Learning Demonstration Examples", "abs": "Alignment tuning is crucial for ensuring large language models (LLMs) behave safely, ethically, and align with human values. It bridges the gap between raw model capabilities and nuanced task requirements, such as helpfulness and user safety. Current alignment approaches, like instruction-following through supervised fine-tuning (SFT) and preference optimization (PO), require high-quality data and significant resources. This paper proposes a low-cost, tuning-free method using in-context learning (ICL) to enhance LLM alignment.\n\nLeveraging the autoregressive nature of LLMs, we observed that aligned models adjust the probability distribution of early polarity tokens during decoding, influencing their response trajectory. Among polarity tokens, malicious tokens induce LLMs to positively respond to toxic queries, whereas benign tokens encourage constructive output. Based on this, we designed heuristic rules to select ICL demonstration examples that effectively influence polarity token distributions.\n\nWe packaged these examples as prompts to trigger few-shot learning, improving LLM alignment. Furthermore, the style and content of ICL demonstrations critically impact few-shot learning. Rewriting examples in a unified, structured style improved LLM accuracy and helpfulness, while specific content encouraged refusal of malicious prompts, enhancing safety.\n\nOur experiments show that rewritten examples boost alignment, safety, and reasoning across various tasks. Compared to the best baseline approach, with an average score of 5.00 as the maximum, our method achieves a maximum 0.15 increase on the Alpaca-eval task (from 4.44 \u2192 4.59), a 0.10 enhancement on the just-eval-instruct benchmark (from 4.50 \u2192 4.60), and a maximum improvement of 0.08 (from 3.53 \u2192 3.61) on the MT-Bench dataset. These findings underscore the need for deeper analysis and theoretical understanding of alignment for advancing future LLM research.", "keywords": [ "alignment", "in-context learning", "safety" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "IHp3vOVQO2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Christina Heinze-Deml", "gender": "Female", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Jaya Narain", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Juyeon Heo", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Miao Xiong", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 28, "n_ref": 66, "n_ref_all": 94, "n_fig": 10, "n_tab": 11, "L_tab": 4288, "n_element_tab": 182, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 21, "n_element_tab_1": 2, "formula_len_all": 54, "formula_len_all_1": 92, "len_all": 161564, "len_all_1": 66706, "len_abs": 2010, "len_title": 59, "len_sents": 59854, "len_sents_1": 35295, "n_sents": 383, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1485, "title": "Do LLMs estimate uncertainty well in instruction-following?", "abs": "Large language models (LLMs) could be valuable personal AI agents across various domains, provided they can precisely follow user instructions. However, recent studies have shown significant limitations in LLMs' instruction-following capabilities, raising concerns about their reliability in high-stakes applications. \nAccurately estimating LLMs' uncertainty in adhering to instructions is critical to mitigating deployment risks. We present, to our knowledge, the first systematic evaluation of the uncertainty estimation abilities of LLMs in the context of instruction-following. \nOur study identifies key challenges with existing instruction-following benchmarks, where multiple factors are entangled with uncertainty stems from instruction-following, complicating the isolation and comparison across methods and models.\nTo address these issues, we introduce a controlled evaluation setup with two benchmark versions of data, enabling a comprehensive comparison of uncertainty estimation methods under various conditions.\nOur findings show that existing uncertainty methods struggle, particularly when models make subtle errors in instruction following. While internal model states provide some improvement, they remain inadequate in more complex scenarios. \nThe insights from our controlled evaluation setups provide a crucial understanding of LLMs' limitations and potential for uncertainty estimation in instruction-following tasks, paving the way for more trustworthy AI agents.", "keywords": [ "Instruction-following", "Uncertainty", "Large language models", "Evaluation", "Benchmark dataset" ], "rating_list": [ 8, 1, 8, 6 ], "soundness_list": [ 4, 1, 3, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "IHRQif8VQC", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Balaji Lakshminarayanan", "gender": "Male", "institution": "Google Brain", "country": "", "position": "Research Scientist" }, { "name": "Stanislav Fort", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 68, "n_ref_all": 115, "n_fig": 24, "n_tab": 3, "L_tab": 2200, "n_element_tab": 249, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 1268, "n_element_tab_1": 124, "formula_len_all": 294, "formula_len_all_1": 37, "len_all": 158604, "len_all_1": 68922, "len_abs": 1851, "len_title": 131, "len_sents": 54442, "len_sents_1": 34635, "n_sents": 357, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1865, "title": "Ensemble everything everywhere: Multi-scale aggregation for adversarial robustness", "abs": "Adversarial examples pose a significant challenge to the robustness, reliability and alignment of deep neural networks. We propose a novel, easy-to-use approach to achieving high-quality representations that lead to adversarial robustness through the use of multi-resolution input representations and dynamic self-ensembling of intermediate layer predictions. We demonstrate that intermediate layer predictions exhibit inherent robustness to adversarial attacks crafted to fool the full classifier, and propose a robust aggregation mechanism based on Vickrey auction that we call \\textit{CrossMax} to dynamically ensemble them. By combining multi-resolution inputs and robust ensembling, we achieve significant adversarial robustness on CIFAR-10 and CIFAR-100 datasets without any adversarial training or extra data, reaching an adversarial accuracy of \u224872% (CIFAR-10) and \u224848% (CIFAR-100) on the RobustBench AutoAttack suite (L\u221e=8/255) with a finetuned ImageNet-pretrained ResNet152. This represents a result comparable with the top three models on CIFAR-10 and a +5 % gain compared to the best current dedicated approach on CIFAR-100. Adding simple adversarial training on top, we get \u224878% on CIFAR-10 and \u224851% on CIFAR-100, improving SOTA by 5 % and 9 % respectively and seeing greater gains on the harder dataset. We validate our approach through extensive experiments and provide insights into the interplay between adversarial robustness, and the hierarchical nature of deep representations. We show that simple gradient-based attacks against our model lead to human-interpretable images of the target classes as well as interpretable image changes. As a byproduct, using our multi-resolution prior, we turn pre-trained classifiers and CLIP models into controllable image generators and develop successful transferable attacks on large vision language models.", "keywords": [ "robustness", "ensemble", "adversarial attacks", "generator" ], "rating_list": [ 6, 8, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 3, 4, 4 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "IGuLzOXTB9", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Hui Dai", "gender": "Female", "institution": "New York University", "country": "US", "position": "MS student" }, { "name": "Mengye Ren", "gender": "unknown", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Ryan Teehan", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 74, "n_ref_all": 117, "n_fig": 18, "n_tab": 10, "L_tab": 6068, "n_element_tab": 207, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1350, "n_element_tab_1": 102, "formula_len_all": 161, "formula_len_all_1": 0, "len_all": 185956, "len_all_1": 54065, "len_abs": 906, "len_title": 122, "len_sents": 60062, "len_sents_1": 25810, "n_sents": 489, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 910, "title": "Are LLMs Prescient? A Continuous Evaluation using Daily News as the Oracle", "abs": "Many existing evaluation benchmarks for Large Language Models (LLMs) quickly become outdated due to the emergence of new models and training data. These benchmarks also fall short in assessing how LLM performance changes over time, as they consist of static questions without a temporal dimension. To address these limitations, we propose using future event prediction as a continuous evaluation method to assess LLMs' temporal generalization and forecasting abilities. Our benchmark, Daily Oracle, automatically generates question-answer (QA) pairs from daily news, challenging LLMs to predict \"future\" event outcomes. Our findings reveal that as pre-training data becomes outdated, LLM performance degrades over time. While Retrieval Augmented Generation (RAG) has the potential to enhance prediction accuracy, the performance degradation pattern persists, highlighting the need for continuous model updates.", "keywords": [ "LLM Forecasting", "Continuous Evaluation", "Temporal Generalization" ], "rating_list": [ 3, 8, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "IFXvpRpci0", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Fang Sun", "gender": "Male", "institution": "UCLA Computer Science Department, University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Haixin Wang", "gender": "unknown", "institution": "UCLA Computer Science Department, University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Wei Wang", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "Xiao Luo", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Yadi Cao", "gender": "Not Specified", "institution": "University of California, San Diego", "country": "US", "position": "Postdoc" }, { "name": "Yizhou Sun", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "Zijie Huang", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 14, "n_ref": 31, "n_ref_all": 45, "n_fig": 4, "n_tab": 6, "L_tab": 2556, "n_element_tab": 109, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1005, "n_element_tab_1": 58, "formula_len_all": 525, "formula_len_all_1": 525, "len_all": 99220, "len_all_1": 54027, "len_abs": 1555, "len_title": 120, "len_sents": 35118, "len_sents_1": 24235, "n_sents": 290, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1286, "title": "Joint Modeling of Spatial and Temporal Multiscales in Molecular Dynamics", "abs": "Molecular dynamics simulations are crucial for understanding complex physical, chemical, and biological processes at the atomic level. However, accurately capturing interactions across multiple spatial and temporal scales remains a significant challenge. We present a novel framework that jointly models spatial and temporal multiscale interactions in molecular dynamics. Our approach leverages Graph Fourier Transforms to decompose molecular structures into different spatial scales and employs Neural Ordinary Differential Equations to model the temporal dynamics at each scale. This unified framework explicitly links spatial structures with temporal evolution, enabling more accurate and comprehensive simulations of molecular systems. We evaluate our model on the MD17 dataset, demonstrating consistent performance improvements over state-of-the-art baselines across multiple molecules, particularly under challenging conditions such as irregular timestep sampling and long-term prediction horizons. Ablation studies confirm the significant contributions of both spatial and temporal multiscale modeling components. Our method advances the simulation of complex molecular systems, potentially accelerating research in computational chemistry, drug discovery, and materials science.", "keywords": [ "Multi-scale molecular dynamics; Graph Fourier transformation; Neural ODEs" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "IFOgfaX2Fj", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Aparna Kanakatte Gurumurthy", "gender": "Female", "institution": "Tata Consultancy Services Limited, India", "country": "", "position": "Scientist" }, { "name": "Avik Ghose", "gender": "Male", "institution": "Tata Consultancy Services Limited, India", "country": "IN", "position": "Principal Researcher" }, { "name": "DIVYA MANOHARLAL BHATIA", "gender": "Female", "institution": "Tata Consultancy Services Limited, India", "country": "IN", "position": "Researcher" }, { "name": "Murali Poduval", "gender": "unknown", "institution": "Tata Consultancy Services Limited, India", "country": "", "position": "consultant" }, { "name": "RUPSHA MUKHERJEE", "gender": "unknown", "institution": "", "country": "", "position": "" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 8, "n_ref": 12, "n_ref_all": 20, "n_fig": 2, "n_tab": 5, "L_tab": 1233, "n_element_tab": 195, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1233, "n_element_tab_1": 195, "formula_len_all": 123, "formula_len_all_1": 123, "len_all": 66987, "len_all_1": 53565, "len_abs": 1127, "len_title": 143, "len_sents": 25657, "len_sents_1": 24177, "n_sents": 188, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1133, "title": "Automated Zonal level implant loosening detection from Hip X-ray using a multi-staged approach", "abs": "Hip arthroplasty is a surgical procedure that involves the replacement of a patient\u2019s hip joint with a prosthetic implant. While these implants are initially effective, they may eventually fail and necessitate revision surgery. It is important to identify the 3 Charnley and 7 Gruen zones around the implant and then identify the zone-wise radiolucency which indicates loosening for effective pre and post-operative planning. Despite the importance of zones, there is a lack of automation attempts in this field. In this work, we have proposed a 3-stage algorithm that detects the sanity of the image for diagnosis, segments into the zones, and then identifies radiolucency within the zones. We have demonstrated a 94% accuracy for Fit/Not Fit segregation, a 0.95 dice score for our zonal segmentation, and a 98% overall loosening accuracy. Obtaining an average dice score of 0.92 in the segmentation of zones and 0.93 accuracy on loosening detection on a blind dataset indicates the robustness of the proposed algorithm. This work will contribute to the development of more efficient and accurate models to detect implant loosening.", "keywords": [ "Radiolucency", "Implant loosening", "Gruen zones", "Charnley zones" ], "rating_list": [ 3, 5, 1, 1 ], "soundness_list": [ 2, 3, 2, 1 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "IFGtG1o2qj", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Aitian Ma", "gender": "Male", "institution": "Florida International University", "country": "", "position": "PhD student" }, { "name": "Dongsheng Luo", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "Assistant Professor" }, { "name": "Mo Sha", "gender": "Male", "institution": "Florida International University", "country": "US", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 27, "n_ref": 86, "n_ref_all": 120, "n_fig": 13, "n_tab": 15, "L_tab": 9691, "n_element_tab": 745, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 4485, "n_element_tab_1": 322, "formula_len_all": 473, "formula_len_all_1": 181, "len_all": 175607, "len_all_1": 70924, "len_abs": 1559, "len_title": 139, "len_sents": 54811, "len_sents_1": 30375, "n_sents": 364, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1608, "title": "MixLinear: Extreme Low Resource Multivariate Time Series Forecasting with $0.1k$ Parameters", "abs": "In recent years, there has been a growing interest in Long-term Time Series Forecasting (LTSF), which involves predicting long-term future values by analyzing a large amount of historical time-series data to identify patterns and trends. There exist significant challenges in LTSF due to its complex temporal dependencies and high computational demands. Although the Transformer-based models offer high forecasting accuracy, they are often too compute-intensive to be deployed on devices with hardware constraints. On the other hand, the linear models aim to reduce the computational overhead by employing either decomposition methods in the time domain or compact representations in the frequency domain. \nIn this paper, we propose MixLinear, an ultra-lightweight multivariate time series forecasting model specifically designed for resource-constrained environments. MixLinear effectively captures both temporal and frequency domain features by modeling intra-segment and inter-segment variations in the time domain and extracting frequency variations from a low-dimensional latent space in the frequency domain. By reducing the parameter scale of a downsampled $n$-length input/output one-layer linear model from $O(n^2)$ to $O(n)$, MixLinear achieves efficient computation without sacrificing accuracy.\nExtensive evaluations across four benchmark datasets demonstrate that MixLinear attains forecasting performance comparable to, or surpassing, state-of-the-art models with significantly fewer parameters ($0.1K$), which makes it well-suited for deployment on devices with limited computational capacity.", "keywords": [ "Time series analysis", "Time series forecasting", "Time domain decomposition", "Frequency domain decomposition" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "IF0Q9KY3p2", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Guido F. Montufar", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Full Professor" }, { "name": "Shuang Liang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 364, "n_formula_1": 39, "n_ref_uni": 47, "n_ref": 102, "n_ref_all": 179, "n_fig": 8, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1505, "n_element_tab_1": 16, "formula_len_all": 35475, "formula_len_all_1": 2993, "len_all": 350959, "len_all_1": 72337, "len_abs": 1207, "len_title": 130, "len_sents": 106724, "len_sents_1": 32007, "n_sents": 1062, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 81, "L_abs": 1218, "title": "Implicit Bias of Mirror Flow for Shallow Neural Networks in Univariate Regression", "abs": "We examine the implicit bias of mirror flow in least squares error regression with wide and shallow neural networks. For a broad class of potential functions, we show that mirror flow exhibits lazy training and has the same implicit bias as ordinary gradient flow when the network width tends to infinity. For univariate ReLU networks, we characterize this bias through a variational problem in function space. Our analysis includes prior results for ordinary gradient flow as a special case and lifts limitations which required either an intractable adjustment of the training data or networks with skip connections. We further introduce \\emph{scaled potentials} and show that for these, mirror flow still exhibits lazy training but is not in the kernel regime. For univariate networks with absolute value activations, we show that mirror flow with scaled potentials induces a rich class of biases, which generally cannot be captured by an RKHS norm. A takeaway is that whereas the parameter initialization determines how strongly the curvature of the learned function is penalized at different locations of the input space, the scaled potential determines how the different magnitudes of the curvature are penalized.", "keywords": [ "implicit bias", "overparametrized neural network", "mirror descent", "univariate regression", "lazy training" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 4, 3, 4 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "IEul1M5pyk", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Jaejin Lee", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Minjae Jeong", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sungyoon Jung", "gender": "unknown", "institution": "Pohang University of Science and Technology", "country": "KP", "position": "MS student" }, { "name": "Yechan Hwang", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Won Kim", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 21, "n_ref_uni": 35, "n_ref": 72, "n_ref_all": 90, "n_fig": 5, "n_tab": 5, "L_tab": 3447, "n_element_tab": 450, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1529, "n_element_tab_1": 168, "formula_len_all": 1541, "formula_len_all_1": 1036, "len_all": 142640, "len_all_1": 74853, "len_abs": 1317, "len_title": 125, "len_sents": 42154, "len_sents_1": 32588, "n_sents": 318, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1320, "title": "HGM\u00b3: Hierarchical Generative Masked Motion Modeling with Hard Token Mining", "abs": "Text-to-motion generation has significant potential in a wide range of applications including animation, robotics, and AR/VR. While recent works on masked motion models are promising, the task remains challenging due to the inherent ambiguity in text and the complexity of human motion dynamics. To overcome the issues, we propose a novel text-to-motion generation framework that integrates two key components: Hard Token Mining (HTM) and a Hierarchical Generative Masked Motion Model (HGM\u00b3). Our HTM identifies and masks challenging regions in motion sequences and directs the model to focus on hard-to-learn components for efficacy. Concurrently, the hierarchical model uses a semantic graph to represent sentences at different granularity, allowing the model to learn contextually feasible motions. By leveraging a shared-weight masked motion model, it reconstructs the same sequence under different conditioning levels and facilitates comprehensive learning of complex motion patterns. During inference, the model progressively generates motions by incrementally building up coarse-to-fine details. Extensive experiments on benchmark datasets, including HumanML3D and KIT-ML, demonstrate that our method outperforms existing methods in both qualitative and quantitative measures for generating context-aware motions.", "keywords": [ "text-to-motion generation", "generative masked model", "hard token mining", "hierarchical semantic graph" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "IEs29RYxfK", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Chenghao Liu", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Researcher" }, { "name": "Jianling Sun", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Lefei Shen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Mouxiang Chen", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Xiaoyun Joy Wang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Zhuo Li", "gender": "Male", "institution": "State Street Technology (Zhejiang) Ltd.", "country": "CN", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 52, "n_ref": 97, "n_ref_all": 134, "n_fig": 8, "n_tab": 15, "L_tab": 12270, "n_element_tab": 874, "n_fig_1": 9, "n_tab_1": 7, "L_tab_1": 6645, "n_element_tab_1": 312, "formula_len_all": 281, "formula_len_all_1": 191, "len_all": 188881, "len_all_1": 69864, "len_abs": 4008, "len_title": 133, "len_sents": 43561, "len_sents_1": 29178, "n_sents": 309, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1326, "title": "VisionTS: Visual Masked Autoencoders Are Free-Lunch Zero-Shot Time Series Forecasters", "abs": "Foundation models have emerged as a promising approach in time series forecasting (TSF). Existing approaches either repurpose large language models (LLMs) or build large-scale time series datasets to develop TSF foundation models for universal forecasting. However, these methods face challenges due to the severe cross-domain gap or in-domain heterogeneity. This paper explores a new road to building a TSF foundation model from rich and high-quality natural images. Our key insight is that a visual masked autoencoder, pre-trained on the ImageNet dataset, can naturally be a numeric series forecaster. By reformulating TSF as an image reconstruction task, we bridge the gap between image pre-training and TSF downstream tasks. Surprisingly, without further adaptation in the time-series domain, the proposed VisionTS could achieve superior zero-shot forecasting performance compared to existing TSF foundation models. With fine-tuning for one epoch, VisionTS could further improve the forecasting and achieve state-of-the-art performance in most cases. Extensive experiments reveal intrinsic similarities between images and real-world time series, suggesting visual models may offer a \"free lunch'' for TSF and highlight the potential for future cross-modality research. Our code is available in the Supplementary Material.", "keywords": [ "time series forecasting", "foundation models", "computer vision" ], "rating_list": [ 3, 8, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 1, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "IEnYsFjFzI", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Andris Ambainis", "gender": "unknown", "institution": "University of Latvia", "country": "LV", "position": "Full Professor" }, { "name": "Debbie Lim", "gender": "Female", "institution": "University of Latvia", "country": "", "position": "Postdoc" }, { "name": "Joao F. Doriguello", "gender": "unknown", "institution": "Alfr\u00e9d R\u00e9nyi Institute of Mathematics", "country": "HU", "position": "Postdoc" } ], "n_formula": 126, "n_formula_1": 32, "n_ref_uni": 88, "n_ref": 211, "n_ref_all": 244, "n_fig": 0, "n_tab": 5, "L_tab": 1315, "n_element_tab": 159, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 75, "n_element_tab_1": 7, "formula_len_all": 8711, "formula_len_all_1": 1732, "len_all": 256034, "len_all_1": 75750, "len_abs": 2957, "len_title": 121, "len_sents": 63617, "len_sents_1": 29364, "n_sents": 665, "n_sents_1": 275, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1478, "title": "Quantum Algorithm for Online Learning of MDPs with Continuous State Space", "abs": "We propose a novel quantum online algorithm for learning Markov Decision Processes (MDPs) with continuous state space in the average reward model. Our algorithm is based on the line of work on classical online UCCRL algorithms by Ortner and Ryabko (NeurIPS'12). To the best of our knowledge, our work is the first to consider MDPs with continuous state space in the fault-tolerant quantum setting. In the case where the state space is one-dimensional, we show that, via quantum-accessible environments, our quantum algorithm obtains a $\\tilde O(T^{1/2})$ regret, improving upon the $\\tilde O(T^{2/3})$ bound of Lakshmanan, Ortner, and Ryabko (PMLR'15), where $T$ is the number of iterations of the algorithm. For a general $d$-dimensional state space, the regret is bounded by $\\tilde O(T^{1-1/2d})$. Our quantum algorithm uses quantum extended value iteration as a subroutine, which is our second main contribution, and may be of independent interest. We show that quantum extended value iteration achieves a subquadratic speedup in the size of the discretized state space $\\mathcal{S}$ and a quadratic speedup in the size of the action space $\\mathcal{A}$, as compared to its classical counterpart. As our third contribution, we study the limiting behaviour of the sequence of value functions generated by quantum extended value iteration. We show that the sequence converges to the optimal average reward $\\rho^*$ up to $\\epsilon$ additive error, for some small $\\epsilon>0$.", "keywords": [ "Quantum algorithm", "Markov decision processes", "Online algorithms", "Quantum reinforcement learning" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "IEZjjDX0iC", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Charlotte Li", "gender": "Female", "institution": "Albert Einstein College of Medicine", "country": "US", "position": "Researcher" }, { "name": "Zachary Flamholz", "gender": "Male", "institution": "Einsteinmed", "country": "US", "position": "Postdoc" }, { "name": "Libusha Kelly", "gender": "unknown", "institution": "Albert Einstein College of Medicine, Inc.", "country": "", "position": "" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 20, "n_ref": 46, "n_ref_all": 64, "n_fig": 9, "n_tab": 1, "L_tab": 572, "n_element_tab": 138, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 572, "n_element_tab_1": 138, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 94283, "len_all_1": 58257, "len_abs": 2868, "len_title": 124, "len_sents": 29169, "len_sents_1": 27103, "n_sents": 241, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 2950, "title": "Comparing Protein Language Models Using Remote Homology Detection for Phages", "abs": "Background. Protein language models (pLMs) are machine learning models that\nlearn high-dimensional representations of protein sequences. These models have\nutility in biological settings, for instance pLMs can convert between protein sequence and structure (Heinzinger et al., 2023), determine evolutionary relationships between organisms (Bordin et al., 2023), and design protein sequences with\ndesired functions (Madani et al., 2023). Transfer learning with previously trained\npLMs offers a powerful, minimal resource strategy for performing diverse largescale classification and prediction tasks. However, as pLMs proliferate in the\nresearch community with differences in training objectives, model structure(s)\nand training datasets, it is daunting for a less-experienced end user to decide\nwhich pLM to use for biological experiments and discovery. Consequently, it\nis essential to compare pLMs to determine their strengths and limitations. In\nparticular, such explorations are necessary in use-cases relevant to biological researchers. Therefore, we present a comparison of pre-trained pLMs in a difficult remote homology detection task for phage proteins described previously in\nFlamholz et al. (2024), Large language models improve annotation of prokaryotic\nviral proteins. We also make available our code and notebooks to facilitate other\nresearch scientists to use such models. Results. Variations in model training resulted in\nsignificantly different performance in our biological task. We present an analysis that compares five recently published pLMs : (1) ProtT5, (2) ProstT5, (3)\nTMVec, (4) ESM2, and (5) CARP. We observed that all models were able to capture meaningful structural information in viral proteins. We also determined that\ntheir embeddings could be used to train functional classifiers that, when tested\nusing the PHROG and EFAM databases of phage proteins, captured meaningful\nbiological information. However, the performances across the different models\nwere noticeably different. Models trained on larger, more diverse databases of\ngenomic sequences such as Big Fantastic Database (BFD) performed better overall. Models with the Transformer architecture performed better than those with\nthe convolutional neural network (CNN) architectures. Conclusion. The utility\nof pLMs in areas of biological research is clear as we demonstrate such models\nare useful for remote homology detection in phage genomes, an area of active\ninterest in metagenomics and environmental biology. Our study highlights how\nbiological scientists can choose pLMs to incorporate into their experiments and\nanalyses. Overall, while some models clearly performed better, on the whole, all\npLMs achieved high scores for prediction. For end-users, the implication is that\nmany pLM models are useful, but testing and domain knowledge may improve\nresults when addressing specific biological questions and developing specialized\nmodel training paradigms.", "keywords": [ "Protein Language Model", "Machine Learning", "Classification", "Virus" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "IEMmEd5Jgm", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Alexander T. Pham", "gender": "unknown", "institution": "Toyota Motor Corporation", "country": "US", "position": "Principal Researcher" }, { "name": "Chenfeng Xu", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Chensheng Peng", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Mingyu Ding", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Nhat Ho", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Shuqi Zhao", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Wei Zhan", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Yichen Xie", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Masayoshi Tomizuka", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" } ], "n_formula": 27, "n_formula_1": 16, "n_ref_uni": 44, "n_ref": 110, "n_ref_all": 124, "n_fig": 9, "n_tab": 5, "L_tab": 401, "n_element_tab": 54, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 1135, "n_element_tab_1": 13, "formula_len_all": 2290, "formula_len_all_1": 1500, "len_all": 155556, "len_all_1": 65728, "len_abs": 998, "len_title": 132, "len_sents": 38233, "len_sents_1": 27652, "n_sents": 317, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1279, "title": "X-Drive: Cross-modality Consistent Multi-Sensor Data Synthesis for Driving Scenarios", "abs": "Recent advancements have exploited diffusion models for the synthesis of either LiDAR point clouds or camera image data in driving scenarios. Despite their success in modeling single-modality data marginal distribution, there is an under- exploration in the mutual reliance between different modalities to describe com- plex driving scenes. To fill in this gap, we propose a novel framework, X-DRIVE, to model the joint distribution of point clouds and multi-view images via a dual- branch latent diffusion model architecture. Considering the distinct geometrical spaces of the two modalities, X-DRIVE conditions the synthesis of each modality on the corresponding local regions from the other modality, ensuring better alignment and realism. To further handle the spatial ambiguity during denoising, we design the cross-modality condition module based on epipolar lines to adaptively learn the cross-modality local correspondence. Besides, X-DRIVE allows for controllable generation through multi-level input conditions, including text, bounding box, image, and point clouds. Extensive results demonstrate the high-fidelity synthetic results of X-DRIVE for both point clouds and multi-view images, adhering to input conditions while ensuring reliable cross-modality consistency.", "keywords": [ "diffusion models; multi-modality data; autonomous driving" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "IEKQlWIN4w", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Ming Lin", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "Capital One Endowed Professor" }, { "name": "Peng Gao", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Assistant Professor" }, { "name": "Rui Liu", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Yu Shen", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 29, "n_ref": 55, "n_ref_all": 69, "n_fig": 6, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 17, "n_element_tab_1": 1, "formula_len_all": 253, "formula_len_all_1": 256, "len_all": 135293, "len_all_1": 64965, "len_abs": 1625, "len_title": 119, "len_sents": 48072, "len_sents_1": 33193, "n_sents": 349, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1714, "title": "CAML: Collaborative Auxiliary Modality Learning for Multi-Agent Systems", "abs": "Multi-modality learning has become a crucial technique in enhancing the performance of machine learning applications across various domains, including autonomous driving, robotics, and perception systems. Existing frameworks, such as Auxiliary Modality Learning (AML), effectively utilize multiple data sources during training and enable inference with reduced modalities, but they primarily operate in a single-agent context. This limitation is particularly critical in dynamic environments, such as connected autonomous vehicles (CAV), where incomplete data coverage can result in decision-making blind spots. To address these challenges, we introduce Collaborative Auxiliary Modality Learning ($\\textbf{CAML}$), a novel extension of the AML framework for multi-agent systems. $\\textbf{CAML}$ facilitates collaboration among agents by allowing them to share multimodal data during training. During inference, each agent operates effectively with fewer modalities, ensuring robustness in performance even with missing data. We analyze the effectiveness of $\\textbf{CAML}$ from the perspective of uncertainty reduction and data coverage, providing a theoretical support to understand and explain why $\\textbf{CAML}$ works better than AML. We then validate $\\textbf{CAML}$ through experiments in collaborative decision-making for CAV in accident-prone scenarios. Experimental results show that $\\textbf{CAML}$ outperforms AML across all tested scenarios, achieving up to a ${\\bf 58.3}$% improvement in accident detection. Additionally, we validate our approach on real-world data from aerial-ground vehicles for collaborative semantic segmentation, achieving up to ${\\bf 10.8}$% improvement in mIoU compared to AML.", "keywords": [ "Collaborative Auxiliary Modality Learning", "Multi-Agent Collaboration", "Cross-Modality Knowledge Distillation" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "IDxZhXrpNf", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "David Brandfonbrener", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Depen Morwani", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Itai Shapira", "gender": "Male", "institution": "Harvard University, Harvard University", "country": "", "position": "PhD student" }, { "name": "Lucas Janson", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "Associate Professor" }, { "name": "Nikhil Vyas", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Rosie Zhao", "gender": "Female", "institution": "Harvard University, Harvard University", "country": "US", "position": "PhD student" }, { "name": "Sham M. Kakade", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 37, "n_ref": 126, "n_ref_all": 153, "n_fig": 7, "n_tab": 1, "L_tab": 251, "n_element_tab": 39, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 396, "n_element_tab_1": 91, "formula_len_all": 1924, "formula_len_all_1": 1846, "len_all": 166767, "len_all_1": 53645, "len_abs": 1746, "len_title": 121, "len_sents": 42335, "len_sents_1": 24072, "n_sents": 353, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1776, "title": "SOAP: Improving and Stabilizing Shampoo using Adam", "abs": "There is growing evidence of the effectiveness of Shampoo, a higher-order preconditioning method, over Adam in deep learning optimization tasks. However, Shampoo's drawbacks include additional hyperparameters and computational overhead when compared to Adam, which only updates running averages of first- and second-moment quantities. This work establishes a formal connection between Shampoo (implemented with the 1/2 power) and Adafactor --- a memory-efficient approximation of Adam --- showing that Shampoo is equivalent to running Adafactor in the eigenbasis of Shampoo's preconditioner. This insight leads to the design of a simpler and computationally efficient algorithm: **S**hampo**O** with **A**dam in the **P**reconditioner's eigenbasis (SOAP).\nWith regards to improving Shampoo's computational efficiency, the most straightforward approach would be to simply compute Shampoo's eigendecomposition less frequently. Unfortunately, as our empirical results show, this leads to performance degradation that worsens with this frequency. SOAP mitigates this degradation by continually updating the running average of the second moment, just as Adam does, but in the current (slowly changing) coordinate basis. Furthermore, since SOAP is equivalent to running Adam in a rotated space, it introduces only one additional hyperparameter (the preconditioning frequency) compared to Adam. We empirically evaluate SOAP on language model pre-training with 360m and 660m sized models. In the large batch regime, SOAP reduces the number of iterations by over 40\\% and wall clock time by over 35\\% compared to AdamW, with approximately 20\\% improvements in both metrics compared to Shampoo. An implementation of SOAP is available at https://anonymous.4open.science/status/SOAP-F93B.", "keywords": [ "Shampoo", "Adam", "Second Order Optimizer" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 2, 5, 3, 3 ] }, { "paper_id": "IDJUscOjM3", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Alan Ritter", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "David Daniel Cox", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "IBM Director, MIT-IBM Watson AI Lab" }, { "name": "Hongyin Luo", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Jacob A Hansen", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "MS student" }, { "name": "James R. Glass", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Principal Researcher" }, { "name": "Junmo Kang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Leonid Karlinsky", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Rameswar Panda", "gender": "Male", "institution": "MIT-IBM Watson AI Lab", "country": "", "position": "Research Scientist" }, { "name": "Rogerio Feris", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Research Manager" }, { "name": "Zhen Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "Postdoc" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 59, "n_ref": 95, "n_ref_all": 124, "n_fig": 5, "n_tab": 11, "L_tab": 4795, "n_element_tab": 773, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1803, "n_element_tab_1": 267, "formula_len_all": 517, "formula_len_all_1": 382, "len_all": 249074, "len_all_1": 65255, "len_abs": 1284, "len_title": 132, "len_sents": 51164, "len_sents_1": 28911, "n_sents": 384, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1290, "title": "Self-MoE: Towards Compositional Large Language Models with Self-Specialized Experts", "abs": "We present Self-MoE, an approach that transforms a monolithic LLM into a compositional, modular system of self-specialized experts, named MiXSE (MiXture of Self-specialized Experts). Our approach leverages self-specialization, which constructs expert modules using self-generated synthetic data, each equipping a shared base LLM with distinct domain-specific capabilities, activated via self-optimized routing. This allows for dynamic and capability-specific handling of various target tasks, enhancing overall capabilities, without extensive human-labeled data and added parameters. Our empirical results reveal that specializing LLMs may exhibit potential trade-offs in performances on non-specialized tasks. On the other hand, our Self-MoE demonstrates substantial improvements (6.5%p on average) over the base LLM across diverse benchmarks such as knowledge, reasoning, math, and coding. It also consistently outperforms other methods, including instance merging and weight merging, while offering better flexibility and interpretability by design with semantic experts and routing. Our findings highlight the critical role of modularity, the applicability of Self-MoE to multiple base LLMs, and the potential of self-improvement in achieving efficient, scalable, and adaptable systems.", "keywords": [ "Efficient Specialization of LLMs", "Self-Improving", "Mixture of Experts" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "ICr9KMxa1K", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Bing Shuai", "gender": "Male", "institution": "Amazon Web Service", "country": "US", "position": "Sr. Applied Scientist" }, { "name": "Caren Chen", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Jiaojiao Zhao", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Shuai Liao", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Xinyu Li", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 45, "n_ref": 96, "n_ref_all": 110, "n_fig": 8, "n_tab": 5, "L_tab": 1838, "n_element_tab": 219, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1838, "n_element_tab_1": 219, "formula_len_all": 818, "formula_len_all_1": 818, "len_all": 130324, "len_all_1": 69954, "len_abs": 1298, "len_title": 116, "len_sents": 36234, "len_sents_1": 28505, "n_sents": 308, "n_sents_1": 251, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1305, "title": "ART: Actor-Related Tubelet for Detecting Complex-shaped Action Tubes", "abs": "This paper focuses on detecting complex-shaped action tubes in videos. Existing methods are based on the assumption that actor's position changes slightly in short video clips. These methods either oversimplify the shape of action tubes by representing them as cuboids or conjecture that action tubes can be summarized into a set of learnable positional patterns. However, these solutions may be insufficient when actor trajectories become more complex. This limitation arises because these methods rely solely on position information to determine action tubes, lacking the ability to trace the same actor when their movement patterns are intricate. To address this issue, we propose Actor-related Tubelet (ART), which incorporates actor-specific information when generating action tubes. Regardless of the complexity of an actor's trajectory, ART ensures that an action tube consistently tracks the same actor, relying on actor-specific cues rather than solely on positional information. To evaluate the effectiveness of ART in handling complex-shaped action tubes, we introduce a dedicated metric that quantifies tube shape complexity. We conduct experiments on three commonly used tube detection datasets: MultiSports, UCF101-24 and JHMDB51-21. ART presents remarkable improvements on all the datasets.", "keywords": [ "human action recognition", "action tube localization" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 1, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "ICR3swcnaa", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Biao Hou", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Desheng Zhai", "gender": "unknown", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Jing Gu", "gender": "unknown", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Licheng Jiao", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Shasha Mao", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Shuyuan Yang", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Yusong Bai", "gender": "unknown", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 33, "n_ref": 68, "n_ref_all": 91, "n_fig": 5, "n_tab": 4, "L_tab": 990, "n_element_tab": 147, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 976, "n_element_tab_1": 140, "formula_len_all": 131, "formula_len_all_1": 131, "len_all": 120559, "len_all_1": 59004, "len_abs": 1857, "len_title": 108, "len_sents": 34782, "len_sents_1": 28581, "n_sents": 224, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1680, "title": "Spatio-temporal Diffusion Transformer for Action Recognition", "abs": "Video action recognition has aroused the research interest of many scholars, and has been widely used in public surveillance, video review, sports events and other fields. However, the high similarity of video background and the long time span of video action bring serious challenges to action recognition. In this work, we propose a spatio-temporal diffusion transformer (STD-Former) to improve the recognition accuracy of long-distance and fine-grained actions. STD-Former utilizes a two-branch network to extract the spatiotemporal and temporal information of video respectively. First, we construct a parallel transformer module to capture the spatiotemporal feature of actions through a two-dimensional convolutional structure in the spatiotemporal branch. Secondly, a cross transformer module integrating the feature of spatiotemporal branch is presented to explore the long-distance temporal dependency relationship of video actions in the temporal branch. In addition, we design a novel plug-and-play spatiotemporal diffusion module, which feeds back the feature extracted from the temporal branch to the spatiotemporal branch, thus enhancing the action capture ability of model. Finally, in order to learn the fine-grained action information of adjacent video sequences, another plug-and-play significant motion excitation module is established by converting the spatial information of adjacent video frames into the motion feature. The experimental results on Something Something V1 and V2 datasets demonstrate that STD-Former can more accurately identify the fine-grained action and has favorable robustness than the current state-of-the-art action recognition models.", "keywords": [ "Video action recognition", "fine-grained action", "information diffusion", "spatiotemporal feature" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "IC5RJvRoMp", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Cuiping Li", "gender": "Female", "institution": "Renmin University of China", "country": "", "position": "Full Professor" }, { "name": "Hong Chen", "gender": "Female", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Jing Zhang", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Xiaodong Chen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "MS student" }, { "name": "Yanling Wang", "gender": "Female", "institution": "Zhongguancun Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yuxuan Hu", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 10, "n_ref_uni": 37, "n_ref": 61, "n_ref_all": 84, "n_fig": 3, "n_tab": 29, "L_tab": 16523, "n_element_tab": 2835, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2847, "n_element_tab_1": 571, "formula_len_all": 492, "formula_len_all_1": 537, "len_all": 190718, "len_all_1": 71982, "len_abs": 2382, "len_title": 111, "len_sents": 37295, "len_sents_1": 29352, "n_sents": 271, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 63, "L_abs": 946, "title": "Streamlining Redundant Layers to Compress Large Language Models", "abs": "This paper introduces LLM-Streamline, a pioneer work on layer pruning for large language models (LLMs). It is based on the observation that different layers have varying impacts on hidden states, enabling the identification of less important layers to be pruned. \nLLM-Streamline comprises two parts: layer pruning, which removes consecutive layers with the lowest importance based on target sparsity, and layer replacement, a novel module that trains a lightweight network to replace the pruned layers to mitigate performance loss. Additionally, a new metric called stability is proposed to address the limitations of the widely used accuracy metric in evaluating model compression. Experiments show that LLM-Streamline outperforms both previous and concurrent state-of-the-art pruning methods in terms of both performance and training efficiency. Our code is available at \\href{https://github.com/RUCKBReasoning/LLM-Streamline}{this repository}.", "keywords": [ "large language models", "model compression", "structured pruning" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 4, 3 ], "confidence_list": [ 4, 4, 5, 1 ] }, { "paper_id": "IBOeJJUYaC", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Gilles Louppe", "gender": "Male", "institution": "University of Li\u00e8ge", "country": "BE", "position": "Associate Professor" }, { "name": "Sacha Lewin", "gender": "Male", "institution": "University of Li\u00e8ge", "country": "", "position": "PhD student" }, { "name": "Omer Rochman Sharabi", "gender": "unknown", "institution": "University of Li\u00e8ge", "country": "BE", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 38, "n_ref": 70, "n_ref_all": 89, "n_fig": 9, "n_tab": 5, "L_tab": 858, "n_element_tab": 10, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 375, "n_element_tab_1": 1, "formula_len_all": 121, "formula_len_all_1": 27, "len_all": 137885, "len_all_1": 60785, "len_abs": 1418, "len_title": 107, "len_sents": 37819, "len_sents_1": 29162, "n_sents": 285, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1490, "title": "A Neural Material Point Method for Particle-based Simulations", "abs": "Mesh-free Lagrangian methods are widely used for simulating fluids, solids, and their complex interactions due to their ability to handle large deformations and topological changes.\n These physics simulators, however, require substantial computational resources for accurate simulations.\n To address these issues, deep learning emulators promise faster and scalable simulations, yet they often remain expensive and difficult to train, limiting their practical use.\n Inspired by the Material Point Method (MPM), we present NeuralMPM, a neural emulation framework for particle-based simulations.\n NeuralMPM interpolates Lagrangian particles onto a fixed-size grid, computes updates on grid nodes using image-to-image neural networks, and interpolates back to the particles.\n Similarly to MPM, NeuralMPM benefits from the regular voxelized representation to simplify the computation of the state dynamics, while avoiding the drawbacks of mesh-based Eulerian methods.\n We demonstrate the advantages of NeuralMPM on 6 datasets, including fluid dynamics and fluid-solid interactions simulated with MPM and Smoothed Particles Hydrodynamics (SPH).\n Compared to GNS and DMCF, NeuralMPM reduces training time from 10 days to 15 hours, memory consumption by 10x-100x, and increases inference speed by 5x-10x, while achieving comparable or superior long-term accuracy, making it a promising approach for practical forward and inverse problems. A project page is available at [URL].", "keywords": [ "Neural emulation", "simulation", "SciML", "particle-based simulators" ], "rating_list": [ 6, 5, 3, 6, 3 ], "soundness_list": [ 2, 2, 1, 3, 2 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 3, 4 ] }, { "paper_id": "IAfGax4tVT", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Behnaz Moradijamei", "gender": "unknown", "institution": "James Madison University", "country": "US", "position": "Assistant Professor" }, { "name": "Heman Shakeri", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Assistant Professor" }, { "name": "Justin Lee", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 21, "n_ref_uni": 42, "n_ref": 68, "n_ref_all": 84, "n_fig": 2, "n_tab": 2, "L_tab": 1202, "n_element_tab": 189, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 1113, "n_element_tab_1": 177, "formula_len_all": 2211, "formula_len_all_1": 1233, "len_all": 151681, "len_all_1": 63389, "len_abs": 1001, "len_title": 160, "len_sents": 43576, "len_sents_1": 28405, "n_sents": 337, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 112, "L_abs": 1003, "title": "Multi-Marginal Stochastic Flow Matching for Alignment of High-Dimensional Snapshot Data at Irregular Time Points", "abs": "Modeling the evolution of high-dimensional systems from limited snapshot observations at irregular time points poses a significant challenge in quantitative biology and related fields. Traditional approaches often rely on dimensionality reduction techniques, which can oversimplify the dynamics and fail to capture critical transient behaviors in non-equilibrium systems. We present Multi-Marginal Stochastic Flow Matching (MMSFM), a novel extension of simulation-free score and flow matching methods to the multi-marginal setting, enabling the alignment of high-dimensional data measured at non-equidistant time points without reducing dimensionality. The use of measure-valued splines enhances robustness to irregular snapshot timing, and score matching prevents overfitting in high-dimensional spaces. We validate our framework on several synthetic and benchmark datasets and apply it to single-cell perturbation data from melanoma cell lines and gene expression data collected at uneven time points.", "keywords": [ "Multi-marginal Snapshot data flow matching", "Measure-valued spline interpolation", "Single-cell data", "Optimal Transport" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "IARgA4HqjJ", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Marco Bertolini", "gender": "Male", "institution": "Pfizer", "country": "DE", "position": "Researcher" }, { "name": "Tuan Le", "gender": "Male", "institution": "Pfizer", "country": "DE", "position": "PhD student" }, { "name": "Djork-Arne Clevert", "gender": "Male", "institution": "Pfizer", "country": "", "position": "Vice President" } ], "n_formula": 94, "n_formula_1": 31, "n_ref_uni": 44, "n_ref": 58, "n_ref_all": 97, "n_fig": 26, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 20, "n_tab_1": 1, "L_tab_1": 200, "n_element_tab_1": 2, "formula_len_all": 8306, "formula_len_all_1": 1982, "len_all": 181739, "len_all_1": 69426, "len_abs": 1783, "len_title": 109, "len_sents": 52480, "len_sents_1": 31381, "n_sents": 424, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1558, "title": "Lie Group-Induced Dynamics in Score-Based Generative Modeling", "abs": "We extend score-based generative modeling by incorporating Lie group actions on the data manifold into the denoising diffusion process. Our approach yields a Langevin dynamics whose infinitesimal transformations decompose as a direct sum of Lie algebra representations, enabling generative processes that align with the underlying symmetry properties of the data. Unlike equivariant models, which restrict the space of learnable functions by quotienting out group orbits, our method incorporates both global and local symmetries and can model any target distribution. Standard score-matching, which minimizes the Fisher divergence, emerges as a special case of our framework when the Lie group is the translation group in Euclidean space. We prove that our generalized generative processes arise as solutions to a new class of reverse-time stochastic differential equations (SDEs), introduced here for the first time. We validate our approach through experiments on diverse data types, demonstrating its effectiveness in real-world applications such as SO(3)-guided molecular conformer generation and modeling ligand-specific global SE(3) transformations for molecular docking. We show that an appropriate choice of Lie group enhances learning efficiency by reducing the effective dimensionality of the trajectory space and enables the modeling of transitions between complex data distributions, lifting the requirement of a Gaussian prior. Additionally, we demonstrate the universality of our approach by deriving how it extends to flow matching techniques.", "keywords": [ "Score-matching", "Generative modeling", "Denoising Diffusion models", "Lie groups", "Lie algebras", "Molecular Conformer Generation", "Deep Learning" ], "rating_list": [ 5, 3, 8, 3, 8 ], "soundness_list": [ 3, 3, 3, 2, 4 ], "presentation_list": [ 2, 4, 3, 2, 3 ], "contribution_list": [ 2, 1, 4, 2, 3 ], "confidence_list": [ 3, 5, 4, 3, 4 ] }, { "paper_id": "IANtNtNpYd", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 6, "author_info_list": [ { "name": "Jing Huo", "gender": "Female", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Quanlin Chen", "gender": "unknown", "institution": "nanjing university", "country": "CN", "position": "PhD student" }, { "name": "Yang Gao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Yuetong Chen", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Undergrad student" }, { "name": "Tianyu Ding", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "yiyu chen", "gender": "Male", "institution": "nanjing university", "country": "", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 42, "n_ref_uni": 36, "n_ref": 51, "n_ref_all": 55, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3454, "formula_len_all_1": 3533, "len_all": 106165, "len_all_1": 47187, "len_abs": 975, "len_title": 131, "len_sents": 23444, "len_sents_1": 19052, "n_sents": 212, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 976, "title": "Enhancing Trust-Region Bayesian Optimization via Derivatives of Gaussian Processes", "abs": "Bayesian Optimization (BO) has been widely applied to optimize expensive black-box functions while retaining sample ef\ufb01ciency. However, scaling BO to high-dimensional spaces remains challenging. Existing literature proposes performing standard BO in several local trust regions (TuRBO) for heterogeneous modeling of the objective function and avoiding over-exploration. Despite its advantages, using local Gaussian Processes (GPs) reduces sampling ef\ufb01ciency compared to a global GP. To enhance sampling ef\ufb01ciency while preserving heterogeneous modeling, we propose to construct several local quadratic models using gradients and Hessians from a global GP, and select new sample points by solving the bound-constrained quadratic program. We provide a convergence analysis and demonstrate through experimental results that our method enhances the ef\ufb01cacy of TuRBO and outperforms a wide range of high-dimensional BO techniques on synthetic functions and real-world applications.", "keywords": [ "Bayesian optimization", "High-dimensional Bayesian optimization", "Trust-Region methods" ], "rating_list": [ 3, 5, 1, 3, 5 ], "soundness_list": [ 1, 4, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 3, 3 ], "contribution_list": [ 2, 3, 1, 2, 3 ], "confidence_list": [ 3, 3, 4, 4, 3 ] }, { "paper_id": "IAFStwZPNu", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Brendan Shillingford", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Gilad Landau", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Mark Woolrich", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "Full Professor" }, { "name": "Oiwi Parker Jones", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "Principal Researcher" }, { "name": "Dulhan Hansaja Jayalath", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 62, "n_ref": 126, "n_ref_all": 140, "n_fig": 4, "n_tab": 1, "L_tab": 218, "n_element_tab": 25, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 978, "n_element_tab_1": 76, "formula_len_all": 198, "formula_len_all_1": 198, "len_all": 178322, "len_all_1": 65550, "len_abs": 1493, "len_title": 128, "len_sents": 41398, "len_sents_1": 32338, "n_sents": 300, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1501, "title": "The Brain's Bitter Lesson: Scaling Speech Decoding With Self-Supervised Learning", "abs": "The past few years have produced a series of spectacular advances in the decoding of speech from brain activity. The engine of these advances has been the acquisition of labelled data, with increasingly large datasets acquired from single subjects. However, participants exhibit individual differences, such as anatomy, and datasets use varied scanners and task designs. As a result, prior work has struggled to leverage data from multiple subjects, multiple datasets, multiple tasks, and unlabelled datasets. In turn, the field has not benefited from the rapidly growing number of open neural data repositories to exploit large-scale data and deep learning. This gap exists for all neural data, but especially for magnetoencephalography (MEG), where the scale of individual datasets has not yet caught up with other modalities. To address this, we develop a set of neuroscience-inspired self-supervised objectives, together with a neural architecture, for representation learning from heterogeneous and unlabelled neural recordings. Experimental results with MEG show that representations learned with these objectives scale with data, generalise across subjects, datasets, and tasks, outperform using the raw input representation, and even surpass comparable self-supervised approaches. In addition, we set new benchmarks for two foundational speech decoding tasks. Collectively, these methods now unlock the potential for training speech decoding models with orders of magnitude more existing data.", "keywords": [ "neural decoding", "speech decoding", "brain-computer interfaces" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "IAFLoDz6H5", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Aaron David Tucker", "gender": "unknown", "institution": "FAR AI", "country": "US", "position": "Researcher" }, { "name": "Adam Gleave", "gender": "Male", "institution": "FAR.AI", "country": "US", "position": "Principal Researcher" }, { "name": "Ian R. McKenzie", "gender": "Male", "institution": "New York University", "country": "", "position": "Researcher" }, { "name": "Micha\u0142 Zaj\u0105c", "gender": "Male", "institution": "Snowflake", "country": "PL", "position": "Research Scientist" }, { "name": "Nikolaus H. R. Howe", "gender": "Non-Binary", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "PhD student" }, { "name": "Oskar John Hollinsworth", "gender": "Male", "institution": "FAR AI", "country": "US", "position": "Researcher" }, { "name": "Pierre-Luc Bacon", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "Assistant Professor" }, { "name": "Tom Tseng", "gender": "Male", "institution": "FAR AI", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 63, "n_ref": 90, "n_ref_all": 129, "n_fig": 15, "n_tab": 10, "L_tab": 8973, "n_element_tab": 302, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 2678, "n_element_tab_1": 175, "formula_len_all": 170, "formula_len_all_1": 0, "len_all": 207880, "len_all_1": 62367, "len_abs": 1246, "len_title": 93, "len_sents": 54122, "len_sents_1": 29270, "n_sents": 373, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1254, "title": "Effects of Scale on Language Model Robustness", "abs": "Language models exhibit scaling laws, whereby increasing model and dataset size yield predictable decreases in negative log likelihood, unlocking a dazzling array of capabilities. At the same time, even the most capable systems are currently vulnerable to adversarial inputs such as jailbreaks and prompt injections, despite concerted efforts to make them robust. As compute becomes more accessible to both attackers and defenders, which side will benefit more from scale? Will safety-trained frontier models become robust against any but the strongest attacks, or will additional compute make attacks almost impossible to defend against?\n\nWe attempt to answer this question with a detailed study of robustness on language models spanning three orders of magnitude in parameter count. We find that increasing base model size alone does not consistently improve robustness. However, larger models benefit more from safety-training, and in particular better generalize from adversarial training to new attacks. We then study the attacker's perspective, finding predictable improvement in attack success rate as attacker compute is increased against all models studied. Finally, we show that offense widens its advantage as both sides spend more on compute.", "keywords": [ "ai safety", "language models", "adversarial attacks", "robustness", "scaling laws" ], "rating_list": [ 5, 3, 8, 3, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 4, 3, 3 ], "contribution_list": [ 2, 1, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 3 ] }, { "paper_id": "IA3wm5vwUl", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Jianlin Chen", "gender": "unknown", "institution": "Shenzhen University", "country": "CN", "position": "MS student" }, { "name": "Jianping Luo", "gender": "Male", "institution": "Department of Software Engineering, Shenzhen University", "country": "", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 17, "n_ref_uni": 39, "n_ref": 62, "n_ref_all": 89, "n_fig": 3, "n_tab": 1, "L_tab": 1763, "n_element_tab": 116, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 1763, "n_element_tab_1": 116, "formula_len_all": 809, "formula_len_all_1": 723, "len_all": 106261, "len_all_1": 67931, "len_abs": 986, "len_title": 170, "len_sents": 31534, "len_sents_1": 29678, "n_sents": 218, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 122, "L_abs": 994, "title": "Deep Learning-based Heuristic Construction for Routing Problems with Dynamic Encoder and Dual-Channel Decoder Architecture", "abs": "The routing problem is a classic combinatorial optimization challenge. Constructing heuristics using deep learning models presents a promising approach for its resolution. In this paper, we propose a novel model with a dynamic encoder and dual-channel decoder (DEDD) architecture to learn construction heuristics for the routing problem. The dynamic encoder en-codes the node features of the decomposed sub-problems at each selection step, thereby obtaining more accurate node em-beddings. The dual-channel decoder facilitates more diverse node selections at each step, increasing the probability of the model identifying optimal solutions. Additionally, we design an effective node selection strategy to assist the model in choosing nodes at each step. Experimental results on the Traveling Salesman Problem (TSP) and the Capacitated Ve-hicle Routing Problem (CVRP) with up to 1000 nodes demonstrate that the solutions generated by the DEDD model are nearly optimal, underscoring its efficacy.", "keywords": [ "Routing problem", "Combinatorial optimization", "Heuristics", "Deep learning" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 3, 1, 2 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 3, 1, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "I9omfcWfMp", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Lihui Chen", "gender": "Unspecified", "institution": "Nanyang Technological University", "country": "", "position": "Associate Professor" }, { "name": "Sitao Luan", "gender": "Male", "institution": "McGill University", "country": "", "position": "PhD student" }, { "name": "Xiang Li", "gender": "unknown", "institution": "Shenzhen University", "country": "CN", "position": "MS student" }, { "name": "Xiaojiang Peng", "gender": "Male", "institution": "Shenzhen Technology University", "country": "CN", "position": "Full Professor" }, { "name": "Yilun Zheng", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 13, "n_ref_uni": 45, "n_ref": 139, "n_ref_all": 184, "n_fig": 16, "n_tab": 23, "L_tab": 12684, "n_element_tab": 1807, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1039, "n_element_tab_1": 42, "formula_len_all": 2314, "formula_len_all_1": 674, "len_all": 251555, "len_all_1": 63359, "len_abs": 2086, "len_title": 105, "len_sents": 64716, "len_sents_1": 28398, "n_sents": 501, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 2105, "title": "Is Graph Convolution Always Beneficial For Every Feature?", "abs": "Graph Neural Networks (GNNs) have demonstrated strong capabilities in processing structured data. While traditional GNNs typically treat each feature dimension equally during graph convolution, we raise an important question: *Is the graph convolution operation equally beneficial for each feature?* If not, the convolution operation on certain feature dimensions can possibly lead to harmful effects, even worse than the convolution-free models. Traditional feature selection methods focus on identifying informative features or reducing redundancy, but they are not suitable for structured data since they overlook graph structures. In the context of graphs, few studies have investigated GNN performance concerning node features using feature homophily metrics, which assess feature consistency with graph topology. Unfortunately, these metrics have not effectively aligned with GNN performance or served as reliable guides for feature selection in GNNs. To address these limitations, we introduce a novel metric, Topological Feature Informativeness (TFI), to distinguish between GNN-favored and GNN-disfavored features, where its effectiveness is validated through both theoretical analysis and empirical observations. Based on TFI, we propose a simple yet effective Graph Feature Selection (GFS) method, which processes GNN-favored and GNN-disfavored features separately, using GNNs and non-GNN models. Compared to original GNNs, GFS significantly improves the extraction of useful topological information from each feature with comparable computational costs. Extensive experiments show that after applying GFS to $8$ baseline and state-of-the-art (SOTA) GNN architectures across $10$ datasets, $90$\\% of the GFS-augmented cases show significant performance boosts. Furthermore, our proposed TFI metric outperforms other feature selection methods in graphs. These results validate the effectiveness of both GFS and TFI. Additionally, we demonstrate that GFS's improvements are robust to hyperparameter tuning, highlighting its potential as a universal method for enhancing various GNN architectures.", "keywords": [ "Graph Neural Networks", "Graph Homophily", "Topological Feature Selection" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "I9bEi6LNgt", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Anderson Compalas", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Undergrad student" }, { "name": "Dawn Song", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Kaiwen Zhou", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Xin Eric Wang", "gender": "Male", "institution": "Simular", "country": "US", "position": "Principal Researcher" }, { "name": "Xuandong Zhao", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Chengzhi Liu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Intern" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 25, "n_ref": 54, "n_ref_all": 91, "n_fig": 26, "n_tab": 5, "L_tab": 2436, "n_element_tab": 362, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 457, "n_element_tab_1": 75, "formula_len_all": 15, "formula_len_all_1": 182, "len_all": 174919, "len_all_1": 70538, "len_abs": 682, "len_title": 203, "len_sents": 52552, "len_sents_1": 32835, "n_sents": 464, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 29, "L_abs": 1560, "title": "Multimodal Situational Safety", "abs": "Multimodal Large Language Models (MLLMs) are rapidly evolving, demonstrating impressive capabilities as multimodal assistants that interact with both humans and their environments. However, this increased sophistication introduces significant safety concerns. In this paper, we present the first evaluation and analysis of a novel safety challenge termed Multimodal Situational Safety, which explores how safety considerations vary based on the specific situation in which the user or agent is engaged. \nWe argue that for an MLLM to respond safely\u2014whether through language or action\u2014it often needs to assess the safety implications of a language query within its corresponding visual context.\nTo evaluate this capability, we develop the Multimodal Situational Safety benchmark (MSSBench) to assess the situational safety performance of current MLLMs. \nThe dataset comprises 1,960 language query-image pairs, half of which the image context is safe, and the other half is unsafe. \nWe also develop an evaluation framework that analyzes key safety aspects, including explicit safety reasoning, visual understanding, and, crucially, situational safety reasoning. \nOur findings reveal that current MLLMs struggle with this nuanced safety problem in the instruction-following setting and struggle to tackle these situational safety challenges all at once, highlighting a key area for future research. \nFurthermore, we develop multi-agent pipelines to coordinately solve safety challenges, which shows consistent improvement in safety over the original MLLM response.", "keywords": [ "Multimodal situational safety", "safety benchmark and evaluation" ], "rating_list": [ 6, 6, 6, 6, 8 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3, 1 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "I9aemDuy5b", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Bolian Li", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Petros Drineas", "gender": "unknown", "institution": "Purdue University", "country": "", "position": "Professor" }, { "name": "Ruqi Zhang", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" }, { "name": "Site Bai", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Xinyu Luo", "gender": "Female", "institution": "Purdue University", "country": "", "position": "PhD student" }, { "name": "Brian Bullins", "gender": "Male", "institution": "Purdue University", "country": "", "position": "Assistant Professor" } ], "n_formula": 42, "n_formula_1": 24, "n_ref_uni": 64, "n_ref": 124, "n_ref_all": 144, "n_fig": 8, "n_tab": 4, "L_tab": 1786, "n_element_tab": 287, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1720, "n_element_tab_1": 243, "formula_len_all": 8849, "formula_len_all_1": 4021, "len_all": 152407, "len_all_1": 66264, "len_abs": 1002, "len_title": 134, "len_sents": 33382, "len_sents_1": 26849, "n_sents": 231, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1058, "title": "Stochastic Steepest Descent with Acceleration for $\\ell_p$-Smooth Non-Convex Optimization", "abs": "In this work, we analyze stochastic $\\ell_p$ steepest descent for non-convex problems. Specifically, for $p > 2$, we establish $\\epsilon$-approximate stationarity (in expectation) with respect to the dual norm $\\Vert\\cdot\\Vert_{p^*}^{p^*}$ at a rate of $O(\\epsilon^{-4})$, thereby generalizing the previous guarantees for signSGD ($p=\\infty$). In addition, inspired by techniques for the convex setting, we present a new accelerated $\\ell_p$ descent method, called Stacey, based on interpolated primal-dual iterate sequences that are designed for non-Euclidean smooth optimization settings. We compare our algorithm against popular methods such as SGD, Adam, AdamW, and Lion on image classification and pretraining language modeling tasks, and our results demonstrate the potential for both faster convergence and achieving higher accuracy. We further evaluate our algorithm for different values of $p$ across various models and datasets, highlighting the importance and efficiency of non-Euclidean methods as compared to standard Euclidean-based approaches.", "keywords": [ "Non-convex Optimization", "Non-Euclidean Acceleration", "Stochastic Steepest Descent" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "I9Dsq0cVo9", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Ahmed Alzubaidi", "gender": "Male", "institution": "TII", "country": "AE", "position": "Researcher" }, { "name": "Aymane El Firdoussi", "gender": "Male", "institution": "T\u00e9l\u00e9com ParisTech", "country": "FR", "position": "MS student" }, { "name": "Hakim Hacid", "gender": "Male", "institution": "TII", "country": "AE", "position": "Principal Researcher" }, { "name": "Mohamed El Amine Seddik", "gender": "Male", "institution": "Technology Innovation Institute", "country": "", "position": "Researcher" }, { "name": "Reda ALAMI", "gender": "Male", "institution": "Technology Innovation Institute", "country": "", "position": "Researcher" }, { "name": "Soufiane Hayou", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "Researcher" } ], "n_formula": 139, "n_formula_1": 18, "n_ref_uni": 18, "n_ref": 39, "n_ref_all": 65, "n_fig": 10, "n_tab": 1, "L_tab": 477, "n_element_tab": 53, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 17555, "formula_len_all_1": 1560, "len_all": 171898, "len_all_1": 63136, "len_abs": 954, "len_title": 126, "len_sents": 43149, "len_sents_1": 26571, "n_sents": 444, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 955, "title": "Maximizing the Potential of Synthetic Data: Insights from Random Matrix Theory", "abs": "Synthetic data has gained attention for training large language models, but poor-quality data can harm performance (see, e.g., Shumailov et al. (2023); Seddik et al. (2024)). A potential solution is data pruning, which retains only high-quality data based on a score function (human or machine feedback). Previous work Feng et al. (2024) analyzed models trained on synthetic data as sample size increases. We extend this by using random matrix theory to derive the performance of a binary classifier trained on a mix of real and pruned synthetic data in a high dimensional setting. Our findings identify conditions where synthetic data could improve performance, focusing on the quality of the generative model and verification strategy. We also show a smooth phase transition in synthetic label noise, contrasting with prior sharp behavior in infinite sample limits. Experiments with toy models and large language models validate our theoretical results.", "keywords": [ "Synthetic Data", "RLHF", "Generative Models", "Statistical Models", "Random Matrices" ], "rating_list": [ 5, 6, 3, 8 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 4, 2, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "I8af9JdQTy", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Paria Rashidinejad", "gender": "Female", "institution": "Meta AI", "country": "US", "position": "Researcher" }, { "name": "Yuandong Tian", "gender": "Male", "institution": "Meta AI (FAIR)", "country": "", "position": "Research Scientist" } ], "n_formula": 113, "n_formula_1": 24, "n_ref_uni": 96, "n_ref": 238, "n_ref_all": 278, "n_fig": 2, "n_tab": 6, "L_tab": 4506, "n_element_tab": 460, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12599, "formula_len_all_1": 2134, "len_all": 352169, "len_all_1": 71488, "len_abs": 2539, "len_title": 142, "len_sents": 91226, "len_sents_1": 33673, "n_sents": 719, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1706, "title": "Sail into the Headwind: Alignment via Robust Rewards and Dynamic Labels against Reward Hacking", "abs": "Aligning AI systems with human preferences typically suffers from the infamous *reward hacking* problem, where optimization of an imperfect reward model leads to undesired behaviors. In this paper, we investigate reward hacking in offline preference optimization, which aims to improve an initial model using a preference dataset. We identify two types of reward hacking stemming from statistical fluctuations in the dataset: Type I Reward Hacking due to subpar choices appearing more favorable, and Type II Reward Hacking due to decent choices appearing less desirable. We prove that many (mainstream or theoretical) preference optimization methods suffer from both types of reward hacking. To mitigate Type I Reward Hacking, we propose POWER, a new preference optimization method that combines Guiasu's weighted entropy with a robust reward maximization objective. POWER enjoys finite-sample guarantees under general function approximation, competing with the best covered policy in the data. To mitigate Type II Reward Hacking, we analyze the learning dynamics of preference optimization and develop a novel technique that dynamically updates preference labels toward certain \"stationary labels\", resulting in diminishing gradients for untrustworthy samples. Empirically, POWER with dynamic labels (DL) consistently outperforms state-of-the-art methods on alignment benchmarks, achieving improvements of up to **13.0** points on AlpacaEval 2 and **11.5** points on Arena-Hard over DPO, while also improving or maintaining performance on downstream tasks such as mathematical reasoning. Strong theoretical guarantees and empirical results demonstrate the promise of POWER-DL in mitigating reward hacking.", "keywords": [ "RLHF", "Preference Optimization", "Reinforcement Learning", "Alignment", "Large Language Models" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 2, 4, 4 ] }, { "paper_id": "I8LdqKbvqX", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Jeffrey Wang", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Undergrad student" }, { "name": "Leitian Tao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Min-Hsuan Yeh", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Xuefeng Du", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "PhD student" }, { "name": "Yixuan Li", "gender": "Female", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 55, "n_ref": 152, "n_ref_all": 184, "n_fig": 5, "n_tab": 10, "L_tab": 1270, "n_element_tab": 127, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1493, "n_element_tab_1": 64, "formula_len_all": 402, "formula_len_all_1": 402, "len_all": 267414, "len_all_1": 67364, "len_abs": 1561, "len_title": 114, "len_sents": 73844, "len_sents_1": 32161, "n_sents": 622, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1215, "title": "How Reliable Is Human Feedback For Aligning Large Language Models?", "abs": "Most alignment research today focuses on designing new learning algorithms using datasets like Anthropic-HH, assuming human feedback data is inherently reliable. However, little attention has been given to the qualitative unreliability of human feedback and its impact on alignment. To address this gap, we conduct a comprehensive study and provide an in-depth analysis of human feedback data. We assess feedback reliability using a committee of gold reward models, revealing that over 25% of the dataset shows low or no agreement with these models, implying a high degree of unreliability. Through a qualitative analysis, we identify six key sources of unreliability, such as mis-labeling, subjective preferences, differing criteria and thresholds for helpfulness and harmlessness, etc. Lastly, to mitigate unreliability, we propose Source-Aware Cleaning, an automatic data-cleaning method guided by the insight of our qualitative analysis, to significantly improve data quality. Extensive experiments demonstrate that models trained on our cleaned dataset, HH-Clean, substantially outperform those trained on the original dataset. We release HH-Clean to support more reliable LLM alignment research in the future.", "keywords": [ "Human feedback", "LLM alignment", "Qualitative analysis" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "I86z54CL2y", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Qijun Feng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Yu-Gang Jiang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Zhen Xing", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Zuxuan Wu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 51, "n_ref": 96, "n_ref_all": 112, "n_fig": 5, "n_tab": 5, "L_tab": 698, "n_element_tab": 128, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 565, "n_element_tab_1": 109, "formula_len_all": 743, "formula_len_all_1": 648, "len_all": 135052, "len_all_1": 53255, "len_abs": 918, "len_title": 196, "len_sents": 28587, "len_sents_1": 23546, "n_sents": 206, "n_sents_1": 166, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 951, "title": "GeoGS3D: Single-view 3D Reconstruction via Geometric-aware Diffusion Model and Gaussian Splatting", "abs": "We introduce GeoGS3D, a novel two-stage framework for reconstructing detailed 3D objects from single-view images. Inspired by the success of pre-trained 2D diffusion models, our method incorporates an orthogonal plane decomposition mechanism to extract 3D geometric features from the 2D input, facilitating the generation of multi-view consistent images. During the following Gaussian Splatting, these images are fused with epipolar attention, fully utilizing the geometric correlations across views. Moreover, we propose a novel metric, Gaussian Divergence Significance (GDS), to prune unnecessary operations during optimization, significantly accelerating the reconstruction process. Extensive experiments demonstrate that \\methodname~generates images with high consistency across views and reconstructs high-quality 3D objects, both qualitatively and quantitatively. Further examples can be found at the anonymous website https://geogs3d.github.io.", "keywords": [ "Diffusion models", "3D reconstruction" ], "rating_list": [ 3, 3, 3, 5, 3 ], "soundness_list": [ 1, 3, 3, 3, 1 ], "presentation_list": [ 1, 2, 2, 1, 2 ], "contribution_list": [ 1, 1, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 2 ] }, { "paper_id": "I7uCwGxVnl", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Ilia Kulikov", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Jason E Weston", "gender": "Not Specified", "institution": "New York University", "country": "US", "position": "Researcher" }, { "name": "Maryam Fazel-Zarandi", "gender": "Female", "institution": "FAIR - Meta", "country": "US", "position": "Research Engineering Manager" }, { "name": "Olga Golovneva", "gender": "Female", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Ping Yu", "gender": "Female", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Yuanzhe Pang", "gender": "Male", "institution": "Meta", "country": "US", "position": "Research Scientist" }, { "name": "Tianlu Wang", "gender": "Female", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Weizhe Yuan", "gender": "Female", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Xian Li", "gender": "unknown", "institution": "Facebook AI", "country": "", "position": "Principal Researcher" }, { "name": "Jane Yu", "gender": "Female", "institution": "Meta AI ", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 37, "n_ref": 60, "n_ref_all": 96, "n_fig": 10, "n_tab": 6, "L_tab": 5236, "n_element_tab": 393, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1374, "n_element_tab_1": 177, "formula_len_all": 137, "formula_len_all_1": 105, "len_all": 151091, "len_all_1": 60909, "len_abs": 1055, "len_title": 70, "len_sents": 34791, "len_sents_1": 24665, "n_sents": 259, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 22, "L_abs": 1061, "title": "Self-Taught Evaluators", "abs": "Model-based evaluation is at the heart of successful model development -- as a reward model for training, and as a replacement for human evaluation. To train such evaluators, the standard approach is to collect a large amount of human preference judgments over model responses, which is costly and the data becomes stale as models improve. In this work, we present an approach that aims to im-prove evaluators without human annotations, using synthetic training data only. Starting from unlabeled instructions, our iterative self-improvement scheme generates contrasting model outputs and trains an LLM-as-a-Judge to produce reasoning traces and final judgments, repeating this training at each new iteration using the improved predictions. Without any labeled preference data, our Self-Taught Evaluator can improve a strong LLM (Llama3-70B-Instruct) from 75.4 to 88.3 (88.7 with majority vote) on RewardBench. This outperforms commonly used LLM judges such as GPT-4 and matches the performance of the top-performing reward models trained with labeled examples.", "keywords": [ "LLM", "Evaluation", "Reward Model", "Evaluator" ], "rating_list": [ 6, 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 4, 2, 4, 2 ], "contribution_list": [ 3, 3, 4, 3, 3 ], "confidence_list": [ 5, 3, 4, 4, 3 ] }, { "paper_id": "I7UpqPmLN5", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Hailin Hu", "gender": "unknown", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Mouxiao Huang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Qiangyu YAN", "gender": "Female", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Tianning Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xinghao Chen", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" }, { "name": "Yunhe Wang", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhen-Liang Ni", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 25, "n_ref": 36, "n_ref_all": 58, "n_fig": 5, "n_tab": 8, "L_tab": 2770, "n_element_tab": 348, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1748, "n_element_tab_1": 181, "formula_len_all": 150, "formula_len_all_1": 144, "len_all": 110058, "len_all_1": 57159, "len_abs": 1699, "len_title": 117, "len_sents": 29436, "len_sents_1": 23668, "n_sents": 242, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1708, "title": "GenVidBench: A Challenging Benchmark for Detecting AI-Generated Video", "abs": "The rapid advancement of video generation models has made it increasingly challenging to distinguish AI-generated videos from real ones. This issue underscores the urgent need for effective AI-generated video detectors to prevent the dissemination of false information through such videos. However, the development of high-performance generative video detectors is currently impeded by the lack of large-scale, high-quality datasets specifically designed for generative video detection. To this end, we introduce GenVidBench, a challenging AI-generated video detection dataset with several key advantages: 1) Cross Source and Cross Generator: The cross-generation source mitigates the interference of video content on the detection. The cross-generator ensures diversity in video attributes between the training and test sets, preventing them from being overly similar. 2) State-of-the-Art Video Generators: The dataset includes videos from 8 state-of-the-art AI video generators, ensuring that it covers the latest advancements in the field of video generation. 3) Rich Semantics: The videos in GenVidBench are analyzed from multiple dimensions and classified into various semantic categories based on their content. This classification ensures that the dataset is not only large but also diverse, aiding in the development of more generalized and effective detection models. We conduct a comprehensive evaluation of different advanced video generators and present a challenging setting. Additionally, we present rich experimental results including advanced video classification models as baselines. With the GenVidBench, researchers can efficiently develop and evaluate AI-generated video detection models.", "keywords": [ "AI-Generated Video Detection; Dataset; Benchmark" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "I7DeajDEx7", "primary_area": "reinforcement learning", "n_author": 11, "author_info_list": [ { "name": "Bin Liang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Bo XU", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Chongjie Zhang", "gender": "unknown", "institution": "Washington University, Saint Louis", "country": "", "position": "Associate Professor" }, { "name": "Dianyu Zhong", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Hao Hu", "gender": "Male", "institution": "Moonshot AI", "country": "CN", "position": "Researcher" }, { "name": "Jun Yang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Assistant Professor" }, { "name": "Qianchuan Zhao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Qihan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiaoteng Ma", "gender": "Male", "institution": "Department of Automation, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Yiqin Yang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuhua Jiang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 9, "n_ref_uni": 53, "n_ref": 176, "n_ref_all": 202, "n_fig": 17, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 198, "n_element_tab_1": 9, "formula_len_all": 3574, "formula_len_all_1": 676, "len_all": 161563, "len_all_1": 64879, "len_abs": 1266, "len_title": 90, "len_sents": 44540, "len_sents_1": 30451, "n_sents": 360, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 1019, "title": "Episodic Novelty Through Temporal Distance", "abs": "Exploration in sparse reward environments remains a significant challenge in reinforcement learning, particularly in Contextual Markov Decision Processes (CMDPs), where environments differ across episodes. Existing episodic intrinsic motivation methods for CMDPs primarily rely on count-based approaches, which are ineffective in large state spaces, or on similarity-based methods that lack appropriate metrics for state comparison. To address these shortcomings, we propose Episodic Novelty Through Temporal Distance (ETD), a novel approach that introduces temporal distance as a robust metric for state similarity and intrinsic reward computation. By employing contrastive learning, ETD accurately estimates temporal distances and derives intrinsic rewards based on the novelty of states within the current episode. Extensive experiments on various benchmark tasks demonstrate that ETD significantly outperforms state-of-the-art methods, highlighting its effectiveness in enhancing exploration in sparse reward CMDPs.", "keywords": [ "Reinforcement Learning" ], "rating_list": [ 5, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "I6UbnkUveF", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Bach Do", "gender": "Male", "institution": "University of Houston", "country": "US", "position": "Postdoc" }, { "name": "Taiwo Adebiyi", "gender": "Male", "institution": "University of Houston", "country": "US", "position": "PhD student" }, { "name": "RUDA ZHANG", "gender": "unknown", "institution": "University of Houston", "country": "US", "position": "Assistant Professor" } ], "n_formula": 82, "n_formula_1": 19, "n_ref_uni": 44, "n_ref": 88, "n_ref_all": 111, "n_fig": 9, "n_tab": 2, "L_tab": 856, "n_element_tab": 40, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9255, "formula_len_all_1": 2018, "len_all": 190831, "len_all_1": 66599, "len_abs": 1335, "len_title": 118, "len_sents": 58482, "len_sents_1": 30290, "n_sents": 479, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1344, "title": "Optimizing Posterior Samples for Bayesian Optimization via Rootfinding", "abs": "Bayesian optimization devolves the global optimization of a costly objective function to the global optimization of a sequence of acquisition functions. This inner-loop optimization can be catastrophically difficult if it involves posterior sample paths, especially in higher dimensions. We introduce an efficient global optimization strategy for posterior samples based on global rootfinding. It provides gradient-based optimizers with two sets of judiciously selected starting points, designed to combine exploration and exploitation. The number of starting points can be kept small without sacrificing optimization quality. Remarkably, even with just one point from each set, the global optimum is discovered most of the time. The algorithm scales practically linearly to high dimensions, breaking the curse of dimensionality. For Gaussian process Thompson sampling (GP-TS), we demonstrate remarkable improvement in both inner- and outer-loop optimization, surprisingly outperforming alternatives like EI and GP-UCB in most cases. Our approach also improves the performance of other posterior sample-based acquisition functions, such as variants of entropy search. Furthermore, we propose a sample-average formulation of GP-TS, which has a parameter to explicitly control exploitation and can be computed at the cost of one posterior sample.", "keywords": [ "Bayesian optimization", "global optimization", "acquisition function", "Thompson sampling" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "I5p1Gm8GFS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Hongxin Ding", "gender": "Female", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Jinyang Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Junfeng Zhao", "gender": "Female", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Runchuan Zhu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Xinke Jiang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Xu Chu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Yasha Wang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Yongxin Xu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yue Fang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 48, "n_ref": 76, "n_ref_all": 99, "n_fig": 3, "n_tab": 14, "L_tab": 4606, "n_element_tab": 653, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2377, "n_element_tab_1": 112, "formula_len_all": 1234, "formula_len_all_1": 1004, "len_all": 223652, "len_all_1": 62233, "len_abs": 1851, "len_title": 136, "len_sents": 65304, "len_sents_1": 27786, "n_sents": 457, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1847, "title": "3DS: Decomposed Difficulty Data Selection\u2019s Case Study on LLM Medical Domain Adaptation", "abs": "Large Language Models (LLMs) excel in general tasks but struggle in specialized domains like healthcare due to limited domain-specific knowledge. Supervised Fine-Tuning (SFT) data construction for domain adaptation often relies on heuristic methods, such as GPT-4 annotation or manual data selection, with a data centric focus on presumed diverse, high-quality datasets. However, these methods overlook the model\u2019s inherent knowledge distribution, introducing noise, redundancy, and irrelevant data, leading to a mismatch between the selected data and the model\u2019s learning task, resulting in suboptimal performance. To address this, we propose a two-stage model-centric data selection framework, Decomposed Difficulty Data Selection (3DS), which aligns data with the model\u2019s knowledge distribution for optimized adaptation. In Stage 1, we apply Prompt-Driven Data Selection via Explicit Alignment, where the model filters irrelevant or redundant data based on its internal knowledge. In Stage 2, we perform Decomposed Difficulty Data Selection, where data selection is guided by our defined difficulty decomposition, using three metrics: Instruction Understanding, Response Confidence, and Response Correctness. Additionally, an attention-based importance weighting mechanism captures token importance for more accurate difficulty calibration. This two-stage approach ensures the selected data is not only aligned with the model\u2019s knowledge and preferences but also appropriately challenging for the model to learn, leading to more effective and targeted domain adaptation. In the case study of the medical domain, our extensive experiments on real-world healthcare datasets demonstrate the superiority of 3DS over existing methods in accuracy by over 5.29%. Our dataset and code will be open-sourced at https://anonymous.4open.science/r/3DS-E67F.", "keywords": [ "supervised fine-tuning", "domain adaptation", "medical llm" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "I5S1a1NKxo", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Anna Bair", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Devin Willmott", "gender": "Male", "institution": "Bosch Center for Artificial Intelligence", "country": "US", "position": "Research Scientist" }, { "name": "Madan Ravi Ganesh", "gender": "Male", "institution": "Bosch Center for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Zico Kolter", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 34, "n_ref": 64, "n_ref_all": 72, "n_fig": 2, "n_tab": 3, "L_tab": 446, "n_element_tab": 69, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 429, "n_element_tab_1": 63, "formula_len_all": 155, "formula_len_all_1": 155, "len_all": 98041, "len_all_1": 51804, "len_abs": 1613, "len_title": 111, "len_sents": 26034, "len_sents_1": 24614, "n_sents": 189, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1651, "title": "Data-scarce distillation for large-scale vision language models", "abs": "Vision-language models (VLMs) have emerged as extremely strong zero-shot and few-shot image classifiers, performing on par with task-specific models. However, they can be unnecessarily heavy-weight for task-specific downstream applications. While existing lines of work have successfully compressed VLMs and other foundation models to varying degrees, most focus on preserving the generality of these models, rather than leveraging their power for a particular task.\nIn this work, we focus on the setting in which we have a limited amount of data on a downstream image classification task and a limited inference budget.\nTo satisfy these constraints, we focus on distilling the strong few-shot performance of CLIP on image classification tasks into a more efficient model. \nWe introduce the SIDCLIP (Synthesize-Initialize-Distill CLIP) method and highlight its three components that are critical to obtaining strong performance: 1) augmenting the classifier with \\textit{synthetic data} generated by leveraging CLIP itself; 2) \\textit{initializing} the modeling process using a smaller CLIP model pretrained on the target architecture; and 3) incorporating \\textit{knowledge distillation} to maximally mimic the performance of the larger model. \nOur set of proposed strategies produces a compact model that performs within 16\\% and 10\\% of CLIP's linear probe performance on 1 and 8 shot datasets respectively, while using a model with less than 2\\% of the parameters of CLIP's image encoder. \nWe hope our work can be useful as a practical guide for leveraging the power of foundation models in downstream data-scarce and budget constrained settings.", "keywords": [ "distillation", "few-shot", "compression", "CLIP", "foundation models" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "I4iZmsV4HM", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Hyung-Sin Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" }, { "name": "Saewoong Bahk", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "Full Professor" }, { "name": "Sunwook Hwang", "gender": "unknown", "institution": "Seoul National University", "country": "", "position": "Postdoc" }, { "name": "Youngseok Kim", "gender": "unknown", "institution": "Seoul National University", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 58, "n_ref": 113, "n_ref_all": 147, "n_fig": 17, "n_tab": 10, "L_tab": 2591, "n_element_tab": 232, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 2307, "n_element_tab_1": 59, "formula_len_all": 424, "formula_len_all_1": 328, "len_all": 180147, "len_all_1": 64403, "len_abs": 1454, "len_title": 167, "len_sents": 50645, "len_sents_1": 29453, "n_sents": 373, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 118, "L_abs": 1412, "title": "ConcreTizer: Model Inversion Attack via Occupancy Classification and Dispersion Control for 3D Point Cloud Restoration", "abs": "The growing use of 3D point cloud data in autonomous vehicles (AVs) has raised serious privacy concerns, particularly due to the sensitive information that can be extracted from 3D data. While model inversion attacks have been widely studied in the context of 2D data, their application to 3D point clouds remains largely unexplored. To fill this gap, we present the first in-depth study of model inversion attacks aimed at restoring 3D point cloud scenes. Our analysis reveals the unique challenges, the inherent sparsity of 3D point clouds and the ambiguity between empty and non-empty voxels after voxelization, which are further exacerbated by the dispersion of non-empty voxels across feature extractor layers. To address these challenges, we introduce ConcreTizer , a simple yet effective model inversion attack designed specifically for voxel-based 3D point cloud data. ConcreTizer incorporates Voxel Occupancy Classification to distinguish between empty and non-empty voxels and Dispersion-Controlled Supervision to mitigate non-empty voxel dispersion. Extensive experiments on widely used 3D feature extractors and benchmark datasets, such as KITTI and Waymo, demonstrate that ConcreTizer concretely restores the original 3D point cloud scene from disrupted 3D feature data. Our findings highlight both the vulnerability of 3D data to inversion attacks and the urgent need for robust defense strategies.", "keywords": [ "Inversion attack", "3D point cloud data", "Autonomous vehicle" ], "rating_list": [ 8, 5, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "I4fi8dvIZS", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Tracy Mok", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Yanghong Zhou", "gender": "Female", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Research Assistant Professor" }, { "name": "Laurence LIU", "gender": "unknown", "institution": "FCCA", "country": "HK", "position": "Researh Scientist" } ], "n_formula": 16, "n_formula_1": 11, "n_ref_uni": 33, "n_ref": 64, "n_ref_all": 104, "n_fig": 12, "n_tab": 6, "L_tab": 1439, "n_element_tab": 188, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1525, "n_element_tab_1": 171, "formula_len_all": 1326, "formula_len_all_1": 918, "len_all": 125408, "len_all_1": 63877, "len_abs": 2149, "len_title": 81, "len_sents": 35019, "len_sents_1": 27062, "n_sents": 244, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1677, "title": "EFFECTIVE REGULARIZATION WITH RELATIVE-DISTANCE VARIANCES IN DEEP METRIC LEARNING", "abs": "This paper develops, for the first time, a novel method using relative-distance variance to regularize deep metric learning (DML), overcoming the drawbacks of existing pair-distance-based metrics, notably loss functions. Being a fundamental field in machine learning research, DML has been widely studied with the goal of learning a feature space where dissimilar data samples are further apart than similar ones. A typical approach of DML is to optimize the feature space by maximizing the relative distances between negative and positive pairs. Despite the rapid advancement, the pair-distance-based approach suffers from a few drawbacks that it heavily relies on the appropriate selection of margin to determine decision boundaries, and it depends on the effective selection of informative pairs, and resulting in low generalization across tasks. To address these issues, this paper explores the use of relative-distance variance and investigates its impact on DML through both empirical and theoretical studies. Based upon such investigation, we propose a novel Relative Distance Variance Constraint (RDVC) loss by regularizing the representation or embedding function learning. The proposed RDVC loss can seamlessly integrate with various pair-distance-based loss functions to ensure a robust and effective performance. Substantial experimental results have demonstrated the effectiveness of our proposed RDVC loss on both within-domain and cross-domain retrieval tasks. In particular, the RDVC loss is also shown useful in fine-grained zero-shot sketch-based image retrieval, a challenging task, revealing its general applicability to cross-domain and zero-shot learning.", "keywords": [ "deep metric learning", "image retrieval" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 3, 1 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "I4e82CIDxv", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Aaron Mueller", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Postdoc" }, { "name": "Can Rager", "gender": "unknown", "institution": "Universit\u00e4t Hamburg", "country": "DE", "position": "MS student" }, { "name": "David Bau", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Assistant Professor" }, { "name": "Eric J Michaud", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Samuel Marks", "gender": "unknown", "institution": "Northeastern University", "country": "", "position": "Postdoc" }, { "name": "Yonatan Belinkov", "gender": "Male", "institution": "Technion, Technion", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 6, "n_ref_uni": 64, "n_ref": 124, "n_ref_all": 182, "n_fig": 22, "n_tab": 7, "L_tab": 2427, "n_element_tab": 325, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3914, "n_element_tab_1": 320, "formula_len_all": 1157, "formula_len_all_1": 527, "len_all": 262135, "len_all_1": 67501, "len_abs": 1090, "len_title": 143, "len_sents": 65268, "len_sents_1": 28055, "n_sents": 555, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 95, "L_abs": 917, "title": "Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models", "abs": "We introduce methods for discovering and applying **sparse feature circuits**. These are causally implicated subnetworks of human-interpretable features for explaining language model behaviors. Circuits identified in prior work consist of polysemantic and difficult-to-interpret units like attention heads or neurons, rendering them unsuitable for many downstream applications. In contrast, sparse feature circuits enable detailed understanding of unanticipated mechanisms in neural networks. Because they are based on fine-grained units, sparse feature circuits are useful for downstream tasks: We introduce SHIFT, where we improve the generalization of a classifier by ablating features that a human judges to be task-irrelevant. Finally, we demonstrate an entirely unsupervised and scalable interpretability pipeline by discovering thousands of sparse feature circuits for automatically discovered model behaviors.", "keywords": [ "Interpretability", "mechanistic interpretability", "circuits", "spurious correlations", "generalization", "dictionary learning" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 4, 3, 3 ], "confidence_list": [ 3, 3, 5, 3 ] }, { "paper_id": "I4YU0oECtK", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Aryaman Arora", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Christopher Potts", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Dan Jurafsky", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Noah Goodman", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Associate Professor" } ], "n_formula": 27, "n_formula_1": 13, "n_ref_uni": 18, "n_ref": 36, "n_ref_all": 71, "n_fig": 9, "n_tab": 10, "L_tab": 7770, "n_element_tab": 1117, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1158, "n_element_tab_1": 194, "formula_len_all": 2395, "formula_len_all_1": 1117, "len_all": 197089, "len_all_1": 60699, "len_abs": 1702, "len_title": 45, "len_sents": 39040, "len_sents_1": 24912, "n_sents": 357, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1393, "title": "Bayesian scaling laws for in-context learning", "abs": "In-context learning (ICL) is a powerful technique for getting language models to perform complex tasks with no training updates.\nPrior work has established strong correlations between the number of in-context examples provided and the accuracy of the model's predictions.\nIn this paper, we seek to explain this correlation by showing that ICL approximates a Bayesian learner. This perspective gives rise to a family of novel Bayesian scaling laws for ICL.\nIn experiments with GPT-2 models of different sizes, our scaling laws match existing scaling laws in accuracy while also offering interpretable terms for task priors, learning efficiency, and per-example probabilities.\nTo illustrate the analytic power that such interpretable scaling laws provide, we report on controlled synthetic dataset experiments designed to inform real-world studies of safety alignment. In our experimental protocol, we use SFT to \nsuppress an unwanted existing model capability and then use ICL to try to bring that capability back (many-shot jailbreaking). We then experiment on real-world instruction-tuned LLMs using capabilities benchmarks as well as a new many-shot jailbreaking dataset.\nIn all cases, Bayesian scaling laws accurately predict the conditions under which ICL will cause the suppressed behavior to reemerge, which sheds light on the ineffectiveness of post-training at increasing LLM safety.", "keywords": [ "in-context learning", "scaling laws", "Bayesian inference", "many-shot jailbreaking" ], "rating_list": [ 5, 6, 5, 8 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 2, 3 ] }, { "paper_id": "I4YAIwrsXa", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 18, "author_info_list": [ { "name": "Bo Liu", "gender": "Not Specified", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Chong Ruan", "gender": "Male", "institution": "DeepSeek", "country": "CN", "position": "Researcher" }, { "name": "Dejian Yang", "gender": "Male", "institution": "DeepSeek AI", "country": "CN", "position": "Researcher" }, { "name": "Fuli Luo", "gender": "Female", "institution": "Individual Researcher", "country": "CN", "position": "Principal Researcher" }, { "name": "Haocheng Wang", "gender": "Male", "institution": "Xiamen University Malaysia", "country": "MY", "position": "Undergrad student" }, { "name": "Haowei Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Huajian Xin", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Junxiao Song", "gender": "Male", "institution": "DeepSeek", "country": "CN", "position": "Researcher" }, { "name": "Liyue Zhang", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Researcher" }, { "name": "Qihao Zhu", "gender": "Male", "institution": "Peking University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Qiushi Du", "gender": "unknown", "institution": "DeepSeek", "country": "CN", "position": "Researcher" }, { "name": "Wanjia Zhao", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Wenjun Gao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xuan Lu", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Undergrad student" }, { "name": "Z.F. Wu", "gender": "unknown", "institution": "DeepSeek", "country": "CN", "position": "Researcher" }, { "name": "Z.Z. Ren", "gender": "unknown", "institution": "DeepSeek", "country": "CN", "position": "Researcher" }, { "name": "Zhibin Gou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Zhihong Shao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 3, "n_ref_uni": 65, "n_ref": 116, "n_ref_all": 133, "n_fig": 5, "n_tab": 10, "L_tab": 4181, "n_element_tab": 423, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1061, "formula_len_all_1": 154, "len_all": 228816, "len_all_1": 54691, "len_abs": 1207, "len_title": 139, "len_sents": 67680, "len_sents_1": 27797, "n_sents": 503, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1212, "title": "Harnessing Proof Assistant Feedback for Reinforcement Learning and Monte-Carlo Tree Search", "abs": "Lean is an advanced proof assistant designed to facilitate formal theorem proving by providing a variety of interactive feedback. In this paper, we explore methodologies to leverage proof assistant feedback to augment the capabilities of large language models in constructing formal proofs. First, we deploy online reinforcement learning using Lean verification outcomes as the reward signal to improve the proof completion policy. This straightforward approach shows great promise in enhancing the model's alignment with the formal verification system. In addition, we propose RMaxTS, a variant of Monte-Carlo tree search that employs an intrinsic-reward-driven exploration strategy to generate diverse proof paths. The tree structure is organized to represent the transitions of intermediate tactic states, extracted from the compilation messages given by Lean's tactic mode. The intrinsic reward is constructed to incentivize the discovery of novel tactic states, which helps to to mitigate the sparse-reward problem inherent in proof search. These techniques lead to a more efficient planning scheme for formal proof generation, achieving new state-of-the-art results on both miniF2F and ProofNet benchmarks.", "keywords": [ "Neural Theorem Proving", "Formal Math", "Large Language Model", "Reinforcement Learning", "Monte-Carlo Tree Search" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "I4Lq2RJ0eJ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Chenxiong Qian", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Kai Ye", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Liangcai Su", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 21, "n_ref": 31, "n_ref_all": 60, "n_fig": 11, "n_tab": 2, "L_tab": 643, "n_element_tab": 50, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 705, "formula_len_all_1": 520, "len_all": 108151, "len_all_1": 60813, "len_abs": 1839, "len_title": 88, "len_sents": 37113, "len_sents_1": 29424, "n_sents": 276, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1914, "title": "How Far Are We from True Unlearnability?", "abs": "High-quality data plays an indispensable role in the era of large models, but the use of unauthorized data for model training greatly damages the interests of data owners. To overcome this threat, several unlearnable methods have been proposed, which generate unlearnable examples (UEs) by compromising the training availability of data. Clearly, due to unknown training purpose and the powerful representation learning capabilities of existing models, these data are expected to be unlearnable for various task models, i.e., they will not help improve the model's performance. However, unexpectedly, we find that on the multi-task dataset Taskonomy, UEs still perform well in tasks such as semantic segmentation, failing to exhibit cross-task unlearnability. This phenomenon leads us to question: How far are we from attaining truly unlearnable examples? We attempt to answer this question from the perspective of model optimization. We observe the difference of convergence process between clean models and poisoned models on a simple model using the loss landscape and find that only a part of the critical parameter optimization paths show significant differences, implying a close relationship between the loss landscape and unlearnability. Consequently, we employ the loss landscape to explain the underlying reasons for UEs and propose Sharpness-Aware Learnability (SAL) for quantifying the unlearnability of parameters based on this explanation. Furthermore, we propose an Unlearnable Distance (UD) metric to measure the unlearnability of data based on the SAL distribution of parameters in clean and poisoned models. Finally, we conduct benchmark tests on mainstream unlearnable methods using the proposed UD, aiming to promote community awareness of the capability boundaries of existing unlearnable methods. The code is available at https://github.com/MLsecurityLab/HowFarAreFromTrueUnlearnability.git.", "keywords": [ "shortcut learning", "unlearnable dataset", "loss landscape" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 2, 3 ] }, { "paper_id": "I393kV3bz4", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Baiqi Wu", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "He Wang", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Associate Professor" }, { "name": "Meng Wang", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Ruixuan Zhang", "gender": "Female", "institution": "Hefei University of Technology", "country": "CN", "position": "MS student" }, { "name": "Xiaoshuai Hao", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence(BAAl) ", "country": "CN", "position": "Principal Researcher" }, { "name": "Xingxing Wei", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Yunfeng Diao", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Lecturer" }, { "name": "A Jian Liu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Postdoc" } ], "n_formula": 37, "n_formula_1": 25, "n_ref_uni": 57, "n_ref": 144, "n_ref_all": 184, "n_fig": 8, "n_tab": 10, "L_tab": 4799, "n_element_tab": 749, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 3684, "n_element_tab_1": 507, "formula_len_all": 2409, "formula_len_all_1": 1531, "len_all": 187986, "len_all_1": 76044, "len_abs": 1770, "len_title": 107, "len_sents": 47469, "len_sents_1": 30161, "n_sents": 407, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1831, "title": "TASAR: Transfer-based Attack on Skeletal Action Recognition", "abs": "Skeletal sequence data, as a widely employed representation of human actions, are crucial in Human Activity Recognition (HAR). Recently, adversarial attacks have been proposed in this area, which exposes potential security concerns, and more importantly provides a good tool for model robustness test. Within this research, transfer-based attack is an important tool as it mimics the real-world scenario where an attacker has no knowledge of the target model, but is under-explored in Skeleton-based HAR (S-HAR). Consequently, existing S-HAR attacks exhibit weak adversarial transferability and the reason remains largely unknown. In this paper, we investigate this phenomenon via the characterization of the loss function. We find that one prominent indicator of poor transferability is the low smoothness of the loss function. Led by this observation, we improve the transferability by properly smoothening the loss when computing the adversarial examples. This leads to the first Transfer-based Attack on Skeletal Action Recognition, TASAR. TASAR explores the smoothened model posterior of pre-trained surrogates, which is achieved by a new post-train Dual Bayesian optimization strategy. Furthermore, unlike existing transfer-based methods which overlook the temporal coherence within sequences, TASAR incorporates motion dynamics into the Bayesian attack, effectively disrupting the spatial-temporal coherence of S-HARs. For exhaustive evaluation, we build the first large-scale robust S-HAR benchmark, comprising 7 S-HAR models, 10 attack methods, 3 S-HAR datasets and 2 defense models. Extensive results demonstrate the superiority of TASAR. Our benchmark enables easy comparisons for future studies, with the code available in the https://anonymous.4open.science/r/RobustBenchHAR-5492/README.md and supplementary material.", "keywords": [ "Human Activity Recognition", "Transfer-based adversarial attack" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "I2ZYngkRW6", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Jiaqi Wei", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Nanqing Dong", "gender": "unknown", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Associate Professor" }, { "name": "Sheng Xu", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Siqi Sun", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiang Zhang", "gender": "Male", "institution": "Patsnap", "country": "SG", "position": "Intern" }, { "name": "ZhiQiang Gao", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhi Jin", "gender": "Male", "institution": "Soochow University", "country": "CN", "position": "MS student" }, { "name": "Zijie Qiu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 42, "n_ref": 78, "n_ref_all": 97, "n_fig": 5, "n_tab": 6, "L_tab": 2666, "n_element_tab": 807, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 872, "n_element_tab_1": 339, "formula_len_all": 618, "formula_len_all_1": 618, "len_all": 139617, "len_all_1": 72261, "len_abs": 2279, "len_title": 141, "len_sents": 37272, "len_sents_1": 32081, "n_sents": 262, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1904, "title": "Distilling Non-Autoregressive Model Knowledge for Autoregressive De Novo Peptide Sequencing", "abs": "Autoregressive (next-token-prediction) models excel in various language generation tasks compared to non-autoregressive (parallel prediction) models. However, their advantage diminishes in certain biology-related tasks like protein modeling and de novo peptide sequencing. Notably, previous studies show that Non-Autoregressive Transformers (NAT) can largely outperform Autoregressive Transformers (AT) in amino acid sequence prediction due to their bidirectional information flow. Despite their advantages, NATs struggle with generalizing to longer sequences, scaling to larger models, and facing extreme optimization difficulties compared to AT models. Motivated by this, we propose a novel framework for directly distilling knowledge from NATs, known for encoding superior protein representations, to enhance autoregressive generation. Our approach employs joint training with a shared encoder and a specially designed cross-decoder attention module. Additionally, we introduce a new training pipeline that uses importance annealing and cross-decoder gradient blocking to facilitate effective knowledge transfer. Evaluations on a widely used 9-species benchmark show that our proposed design achieves state-of-the-art performance. Specifically, AT and NAT baseline models each excel in different types of data prediction due to their unique inductive biases. Our model combines these advantages, achieving strong performance across all data types and outperforming baselines across all evaluation metrics. This work not only advances de novo peptide sequencing but also provides valuable insights into how autoregressive generation can benefit from non-autoregressive knowledge and how next-token prediction (GPT-style) can be enhanced through bidirectional learning (BERT-style).\nWe release our code for reproduction in the anonymous repository here: https://anonymous.4open.science/r/CrossNovo-E263.", "keywords": [ "De novo", "Peptide Sequencing", "autoregressive", "non-autoregressive" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 4, 2, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "I1VCj1l1Zn", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Ruizhe Li", "gender": "Male", "institution": "University of Aberdeen", "country": "GB", "position": "Assistant Professor" }, { "name": "Yuxuan Zhang", "gender": "Male", "institution": "University of Aberdeen", "country": "GB", "position": "Undergrad student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 38, "n_ref": 73, "n_ref_all": 84, "n_fig": 2, "n_tab": 8, "L_tab": 6796, "n_element_tab": 763, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2837, "n_element_tab_1": 343, "formula_len_all": 442, "formula_len_all_1": 355, "len_all": 149656, "len_all_1": 63467, "len_abs": 1260, "len_title": 155, "len_sents": 35459, "len_sents_1": 27659, "n_sents": 235, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1267, "title": "DLP-LoRA: Efficient Task-Specific LoRA Fusion with a Dynamic, Lightweight Plugin for Large Language Models", "abs": "Recent advancements in Large Language Models (LLMs) have achieved robust performance across diverse tasks, but fine-tuning these models for specific domains remains resource-intensive. Parameter-Efficient Fine-Tuning (PEFT) methods like Low-Rank Adaptation (LoRA) address this challenge by fine-tuning a small subset of parameters. However, existing methods for fusing multiple LoRAs lack dynamic fusion based on contextual inputs and often increase inference time due to token-level operations. We propose DLP-LoRA, a Dynamic Lightweight Plugin that employs a mini-MLP module with only 5M parameters to dynamically fuse multiple LoRAs at the sentence level using top-$p$ sampling strategies. This approach reduces inference time to less than twice that of single LoRA inference by leveraging parallel computation. Evaluations across 26 tasks\u2014including multiple-choice questions and question answering\u2014demonstrate that DLP-LoRA achieves an average accuracy of 92.34\\% on multiple-choice datasets and significant improvements in BLEU and ROUGE scores on QA datasets, outperforming different LLMs backbones under composite task settings. DLP-LoRA effectively balances performance and efficiency, making it a practical solution for dynamic multi-task adaptation in LLMs.", "keywords": [ "Multi-LoRA fusion", "Parameter Efficient Tuning", "LoRA", "Cross-Task Generalization" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 1 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "I1MKOjNVup", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Junjie Shan", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Kaicheng Yu", "gender": "Male", "institution": "KMind.AI", "country": "CN", "position": "Principal Researcher" }, { "name": "Siqi Ma", "gender": "Male", "institution": "Westlake.edu.cn", "country": "", "position": "Researcher" }, { "name": "Tiannan Guo", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaojing Zhang", "gender": "Female", "institution": "Westlake University", "country": "CN", "position": "Intern" }, { "name": "Xinna Lin", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Xu Hu", "gender": "Male", "institution": "Samsung", "country": "GB", "position": "Researcher" }, { "name": "Stan Z. Li", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Chair Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 79, "n_ref": 134, "n_ref_all": 181, "n_fig": 15, "n_tab": 20, "L_tab": 9740, "n_element_tab": 736, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 2929, "n_element_tab_1": 289, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 385126, "len_all_1": 69034, "len_abs": 1707, "len_title": 131, "len_sents": 97389, "len_sents_1": 30360, "n_sents": 722, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1716, "title": "BioKGBench: A Knowledge Graph Checking Benchmark of AI Agent for Biomedical Science", "abs": "Pursuing artificial intelligence for biomedical science, a.k.a. AI Scientist, draws increasing attention, where one common approach is to build a copilot agent driven by Large Language Models (LLMs). However, to evaluate such systems, researchers typically rely on direct Question-Answering (QA) to the LLM itself or through biomedical experiments. How to benchmark biomedical agents precisely from an AI Scientist perspective remains largely unexplored. To this end, we draw inspiration from scientists\u2019 crucial ability to understand the literature and introduce BioKGBench. In contrast to traditional evaluation benchmarks that focus solely on factual QA, where the LLMs are known to have hallucination issues, we first disentangle \u201cUnderstanding Literature\u201d into two atomic abilities: i) \u201cUnderstanding\u201d the unstructured text from research papers by performing scientific claim verification, and ii) interacting with structured Knowledge-Graphs for Question-Answering (KGQA) as a form of \u201cLiterature\u201d grounding. We then formulate a novel agent task, dubbed KGCheck, using KGQA and domain-based Retrieval-Augmented Generation (RAG) to identify factual errors in existing large-scale knowledge graphs. We collect over two thousand data points for the two atomic tasks and 225 high-quality annotated samples for the agent task. Surprisingly, we find that state-of-the-art general and biomedical agents have either failed or performed inferiorly on our benchmark. We then introduce a simple yet effective baseline, dubbed BKGAgent. On the widely used popular knowledge graph, we discover over 90 factual errors, which provide scenarios for agents to make discoveries and demonstrate the effectiveness of our approach.", "keywords": [ "benchmark", "biomedical agent", "knowledge graph", "literature" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "I18MA5DjoP", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Dongyan Zhao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Haiyu Yan", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Hongda Sun", "gender": "Male", "institution": "Gaoling School of Artificial Intelligence, Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Rui Yan", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zhengtao Yu", "gender": "Male", "institution": "Kunming University of Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Zhuocheng Gong", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Ziyue Wang", "gender": "unknown", "institution": "Tianjin University", "country": "", "position": "Undergrad student" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 26, "n_ref": 49, "n_ref_all": 76, "n_fig": 13, "n_tab": 6, "L_tab": 2508, "n_element_tab": 273, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1198, "n_element_tab_1": 77, "formula_len_all": 349, "formula_len_all_1": 443, "len_all": 139432, "len_all_1": 61240, "len_abs": 1210, "len_title": 148, "len_sents": 37552, "len_sents_1": 26220, "n_sents": 334, "n_sents_1": 237, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1204, "title": "Can Neuron Activation be Predicted? A New Lens for Analyzing Transformer-based LLM", "abs": "Transformer-based large language models (LLMs) play a vital role in various NLP tasks, but the internal neurons are rather functioning in a black box style. In this work, we introduce the *Neuron Predictability Lens* (NPL), an analytical framework that focuses on the way neurons work within feed-forward networks (FFNs). NPL is useful in understanding and analyzing transformer-based LLMs. Based on this proposed framework, we conduct extensive experiments on LLaMA-2 and GPT-J. Firstly, we show that neuron activations are predictable and for the first time we introduce the concept of *Neuron Predictability*. Secondly, we apply NPL to both global and local analysis. For global analysis, we investigate how FFNs contribute to model behaviors explicitly and implicitly with the aid of NPL. For local analysis, we explore the connection between neuron predictability and neuron interpretability. We examine various functional neurons under NPL and uncover the existence of \u201cbackground neurons.\u201d With the findings mentioned above, we demonstrate the value of NPL as a novel analytical tool and shed light on its future application on model efficiency and/or effectiveness for improved language modeling.", "keywords": [ "Large Language Model", "Model Analysis", "Neuron Predictability" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 2, 4, 3 ] }, { "paper_id": "I1484gDBr4", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "David Freedman", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Full Professor" }, { "name": "Matthew James Tilley", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 17, "n_ref": 30, "n_ref_all": 39, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 238, "n_element_tab_1": 32, "formula_len_all": 914, "formula_len_all_1": 914, "len_all": 70818, "len_all_1": 56800, "len_abs": 1778, "len_title": 110, "len_sents": 23655, "len_sents_1": 23643, "n_sents": 204, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1793, "title": "Linear Recurrent Neural Networks with a Feature-Sequence Twist", "abs": "The transformer network architecture has led to advances in artificial intelligence.\nConversational AI applications, such as ChatGPT, and protein folding predictions with AlphaFold are made possible by transformer architectures and the self-attention mechanism.\nHowever, advancing towards more general, flexible, and energy-efficient artificial intelligence may require exploring new architectures that differ significantly from those currently used.\nTransformer networks have largely replaced recurrent neural networks (RNNs) for state-of-the-art performance on sequence-based tasks. \nHowever, in recent years there has been some successful competition from linear recurrent neural networks (LRNNs) and state space models (SSMs).\nA core advantage of LRNNs and SSMs over traditional RNNs is that the hidden states can be calculated in parallel.\nTherefore, like the transformer, they can make efficient use of GPU computation. \nUnlike the transformer, computational costs of parallelized LRNNs and SSMs can scale sub-quadratically with sequence length.\nDespite these advantages, LRNNs and SSMs often struggle to generate the deep and rich representations that have contributed to the success of transformer architectures.\nWe introduce Feature-Sequence Twisting (FST), a novel technique that transposes the sequence and feature dimensions between LRNN blocks.\nThe purpose of FST is to generate deeper representations of the sequence in subsequent LRNN blocks. \nSince the computational cost of LRNNs scale sub-quadratically with sequence length, FST remains practical to compute even for large feature dimensions.\nOur experiments demonstrate that the FST architecture outperforms transformer networks on tasks such as Long ListOps, achieving performance competitive with state-of-the-art models.", "keywords": [ "linear recurrent neural networks", "RNNs", "Sequence model", "FST" ], "rating_list": [ 3, 3, 1, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 3, 2, 1, 1 ], "contribution_list": [ 1, 1, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "I0n3EyogMi", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Dit-Yan Yeung", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Chair Professor" }, { "name": "Ying-yee Ava Lau", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "MS student" }, { "name": "Zhiwen Shao", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Postdoc" } ], "n_formula": 20, "n_formula_1": 8, "n_ref_uni": 35, "n_ref": 57, "n_ref_all": 108, "n_fig": 13, "n_tab": 13, "L_tab": 3766, "n_element_tab": 348, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 675, "n_element_tab_1": 22, "formula_len_all": 1649, "formula_len_all_1": 509, "len_all": 160422, "len_all_1": 60591, "len_abs": 1248, "len_title": 132, "len_sents": 51340, "len_sents_1": 27715, "n_sents": 391, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1249, "title": "Fast and Slow Streams for Online Time Series Forecasting Without Information Leakage", "abs": "Current research in online time series forecasting suffers from information leakage: models predict and then evaluate on historical time steps that have been backpropagated for parameter updates. This setting also misaligns with the real-world conception of forecasting, which typically emphasizes looking ahead and anticipating future uncertainties. This paper redefines online time series forecasting to focus on predicting unknown future steps and evaluates performance solely based on these predictions. Following this new setting, challenges arise in leveraging incomplete pairs of ground truth and prediction for backpropagation, as well as generalizing accurate information without overfitting to noises from recent data streams. To address these challenges, we propose a novel dual-stream framework for online forecasting (DSOF): a slow stream that updates with complete data using experience replay, and a fast stream that adapts to recent data through temporal difference learning. This dual-stream approach updates a teacher-student model learned through a residual learning strategy, generating predictions in a coarse-to-fine manner. Extensive experiments demonstrate its improvement in forecasting performance in changing environments.", "keywords": [ "online time series forecasting", "concept drift", "online learning" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 4, 4, 2, 2 ] }, { "paper_id": "I0mQlersGk", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Huizhe Zhang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Jintang Li", "gender": "Male", "institution": "Sun Yat-sen University", "country": "CN", "position": "PhD student" }, { "name": "Yuchang Zhu", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "PhD student" }, { "name": "Zibin Zheng", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Liang Chen", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 18, "n_ref": 27, "n_ref_all": 49, "n_fig": 6, "n_tab": 8, "L_tab": 5746, "n_element_tab": 516, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1809, "n_element_tab_1": 289, "formula_len_all": 1283, "formula_len_all_1": 1111, "len_all": 136281, "len_all_1": 60550, "len_abs": 1580, "len_title": 148, "len_sents": 35825, "len_sents_1": 24539, "n_sents": 265, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1519, "title": "SGHormerVQ: Bridging Graph Transformers and Spiking Neural Networks via Spiking Vector Quantization", "abs": "Graph Transformers (GTs), which simultaneously integrate message passing and self-attention mechanisms, have achieved promising empirical results in some graph prediction tasks. Although these approaches show the potential of Transformers in capturing long-range graph topology information, issues concerning the quadratic complexity and high computing energy consumption severely impair the scalability of GTs on large-scale graphs. Recently, as brain-inspired neural networks, Spiking Neural Networks (SNNs) provide an energy-saving deep learning option with lower computational and storage overhead via their unique spike-based event-driven biological neurons. Inspired by these characteristics, we propose SGHormerVQ, which bridges efficient Graph Transformers and spiking neural networks via spiking vector quantization. Spiking vector quantization generates implied codebooks with smaller sizes and higher codebook usage to assist self-attention blocks in performing efficient global information aggregation. SGHormerVQ effectively alleviates the reliance on complex machinery (distance measure, auxiliary loss, etc.) and the \\textit{codebook collapse} present in previous vector quantization-based GNNs. In experiments, we compare SGHormerVQ with other state-of-the-art baselines on node classification datasets ranging from small to large. Experimental results show that SGHormerVQ has achieved competitive performances on most datasets while maintaining up to 518\u00d7 faster inference speed compared to other GTs.", "keywords": [ "graph transformer", "spiking neural network", "vector quantization" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 1 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "I0To0G5J7g", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Ayzaan Wahid", "gender": "Male", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Jonathan Tompson", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Pannag R Sanketi", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Seyed Kamyar Seyed Ghasemipour", "gender": "Male", "institution": "Generalist AI", "country": "US", "position": "Founding Member of Technical Staff" }, { "name": "Igor Mordatch", "gender": "Male", "institution": "University of Washington", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 44, "n_ref": 78, "n_ref_all": 101, "n_fig": 8, "n_tab": 3, "L_tab": 342, "n_element_tab": 10, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 261, "n_element_tab_1": 6, "formula_len_all": 184, "formula_len_all_1": 145, "len_all": 174519, "len_all_1": 64144, "len_abs": 1438, "len_title": 143, "len_sents": 44931, "len_sents_1": 33108, "n_sents": 297, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1443, "title": "On the Surprising Efficacy of Online Self-Improvement for Embodied Multimodal Foundation Models", "abs": "Foundation models trained on web-scale data have revolutionized robotics, but their application to low-level control remains largely limited to behavioral cloning. Drawing inspiration from the sample efficiency and success of reinforcement learning (RL) fine-tuning in large language models (LLMs), we propose a two-stage approach suited to robotics. The first stage, Supervised Fine-Tuning (SFT), fine-tunes pre-trained foundation models using goal-conditioned behavioral cloning and \u201csteps-to-go\u201d prediction objectives. In the second stage, this foundation enables the extraction of a well-shaped reward function and a success detector, eliminating the need for manual reward engineering and real-world instrumentation, and allowing robots to practice autonomously with minimal human supervision. Our experiments on both real-world and simulated robots demonstrate that the combination of SFT and online Self-Improvement is significantly more sample-efficient than supervised learning alone. Furthermore, the combination of our proposed approach with web-scale pre-trained foundation models enables rapid acquisition of new skills, allowing robots to generalize far beyond the behaviors observed in the imitation learning datasets used during training. These findings highlight the transformative potential of combining pre-trained foundation models with online fine-tuning to unlock new levels of autonomy and skill acquisition in robotics.", "keywords": [ "Robotics", "Multimodal Foundation Models", "Post-Training", "Self-Improvement", "Reinforcement Learning" ], "rating_list": [ 5, 5, 10, 5 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 2, 2, 4, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "I0HfqpSH8f", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Aixi Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Li Li", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Senzhang Wang", "gender": "Male", "institution": "Central South University", "country": "CN", "position": "Full Professor" }, { "name": "Weiming Lu", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Yaohua Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 41, "n_ref": 110, "n_ref_all": 126, "n_fig": 11, "n_tab": 6, "L_tab": 967, "n_element_tab": 122, "n_fig_1": 9, "n_tab_1": 7, "L_tab_1": 1298, "n_element_tab_1": 127, "formula_len_all": 256, "formula_len_all_1": 309, "len_all": 127054, "len_all_1": 55154, "len_abs": 345, "len_title": 120, "len_sents": 28422, "len_sents_1": 24045, "n_sents": 242, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1179, "title": "HERO: Harnessing Temporal Modeling for Diffusion-Based Video Outpainting", "abs": "Video outpainting expands the spatial perspective of a video, enabling it to adapt to various display devices with different aspect ratios.\nCurrent diffusion-based approaches for video outpainting often suffer from quality issues such as blurred details, local distortion, and temporal instability, significantly impacting the user experience.\nThe root cause is the insufficient temporal modeling in video\noutpainting, which inadequately represents the relationships between frames over time.\nTo address this issue, a novel approach called HERO~(Harnessing the tEmpoRal modeling for diffusion-based Outpainting) is proposed to effectively tackles these generated video quality problems.\nHERO employs two critical components to enhance temporal modeling: the Temporal Reference Module, which provides reference features that extend beyond spatial dimensions; and the Interpolation-based Motion Modelling Module, designed to stabilize generated frames.\nBy integrating these modules, these quality issues in video outpainting are effectively addressed.\nExtensive experiments on multiple benchmarks demonstrate that HERO outperforms existing methods qualitatively and quantitatively.", "keywords": [ "video outpainting", "diffusion model" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "I05Z6KjQ9K", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Chaohu Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Linli Xu", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yubo Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 30, "n_ref": 78, "n_ref_all": 96, "n_fig": 6, "n_tab": 11, "L_tab": 3161, "n_element_tab": 463, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1941, "n_element_tab_1": 355, "formula_len_all": 150, "formula_len_all_1": 152, "len_all": 184207, "len_all_1": 63111, "len_abs": 2687, "len_title": 124, "len_sents": 38670, "len_sents_1": 27531, "n_sents": 397, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1591, "title": "Gradient Regularization-based Cross-Prompt Attacks on Vision Language Models", "abs": "Recent large vision language models (VLMs) have gained significant attention for their superior performance in various visual understanding tasks using textual instructions, also known as prompts.\nHowever, existing research shows that VLMs are vulnerable to adversarial examples, where imperceptible perturbations added to images can lead to malicious outputs, posing security risks during deployment.\nUnlike single-modal models, VLMs process both images and text simultaneously, making the creation of visual adversarial examples dependent on specific prompts.\nConsequently, the same adversarial example may become ineffective when different prompts are used, which is common as users often input diverse prompts.\nOur experiments reveal severe non-stationarity when directly optimizing adversarial example generation using multiple prompts, resulting in examples specific to a single prompt with poor transferability.\nTo address this issue, we propose the Gradient Regularized-based Cross-Prompt Attack (GrCPA), which leverages gradient regularization to generate more robust adversarial attacks, thereby improving the assessment of model robustness.\nBy exploiting the structural characteristics of the Transformer, GrCPA reduces the variance of back-propagated gradients in the Attention and MLP components, utilizing regularized gradients to produce more effective adversarial examples.\nExtensive experiments on models such as Flamingo, BLIP-2, LLaVA and InstructBLIP demonstrate the effectiveness of GrCPA in enhancing the transferability of adversarial attacks across different prompts.", "keywords": [ "Adversarial attacks", "vision language models", "cross-prompt" ], "rating_list": [ 5, 1, 3, 1 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "HzbPIqvhGg", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Shida Sun", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Yue Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yueyi Zhang", "gender": "unknown", "institution": "Hesai Technology", "country": "CN", "position": "Staff Engineer" }, { "name": "Zhiwei Xiong", "gender": "Male", "institution": "USTC", "country": "CN", "position": "Professor" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 27, "n_ref": 77, "n_ref_all": 96, "n_fig": 14, "n_tab": 3, "L_tab": 1964, "n_element_tab": 189, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 538, "n_element_tab_1": 63, "formula_len_all": 760, "formula_len_all_1": 546, "len_all": 113747, "len_all_1": 60186, "len_abs": 1254, "len_title": 118, "len_sents": 37626, "len_sents_1": 29050, "n_sents": 276, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1258, "title": "Generalizable Non-Line-of-Sight Imaging with Learnable Physical Priors", "abs": "Non-line-of-sight (NLOS) imaging, recovering the hidden volume from indirect reflections, has attracted increasing attention due to its potential applications. Despite promising results, existing NLOS reconstruction approaches are constrained by the reliance on empirical physical priors, e.g., single fixed path compensation. Moreover, these approaches still possess limited generalization ability, particularly when dealing with scenes at a low signal-to-noise ratio (SNR). To overcome the above problems, we introduce a novel learning-based approach, comprising two key designs: Learnable Path Compensation (LPC) and Adaptive Phasor Field (APF). The LPC applies tailored path compensation coefficients to adapt to different objects in the scene, effectively reducing light wave attenuation, especially in distant regions. Meanwhile, the APF learns the precise Gaussian window of the illumination function for the phasor field, dynamically selecting the relevant spectrum band of the transient measurement. Experimental validations demonstrate that our proposed approach, only trained on synthetic data, exhibits the capability to seamlessly generalize across various real-world datasets captured by different imaging systems and characterized by low SNRs.", "keywords": [ "Non-Line-of-Sight", "Learnable Path Compensation", "Adaptive Phasor Field" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "HzG3A0VD1k", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Anran Liu", "gender": "Female", "institution": "University of California, Davis", "country": "US", "position": "PhD student" }, { "name": "Feng Chen", "gender": "Male", "institution": "Lenovo Group Limited", "country": "CN", "position": "Researcher" }, { "name": "Jun Xie", "gender": "Male", "institution": "Lenovo Group Limited", "country": "CN", "position": "Researcher" }, { "name": "Zhepeng Wang", "gender": "unknown", "institution": "Lenovo Group Limited", "country": "CN", "position": "Researcher" }, { "name": "Zongsheng Cao", "gender": "unknown", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 27, "n_ref": 37, "n_ref_all": 53, "n_fig": 5, "n_tab": 4, "L_tab": 3018, "n_element_tab": 96, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 178, "n_element_tab_1": 35, "formula_len_all": 670, "formula_len_all_1": 669, "len_all": 113707, "len_all_1": 57736, "len_abs": 1670, "len_title": 89, "len_sents": 30618, "len_sents_1": 27718, "n_sents": 223, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1695, "title": "EconAI: Preference-driven Agents Simulating Economic Activities via Large Language Model", "abs": "The emergence of artificial intelligence has transformed the methodological frameworks in economic research by simulating intricate interactions among diverse agents. Despite the advantage of large language models (LLMs), they often struggle with occasions involving decision-making interactions with environments. This challenge stems from the fact that most LLMs are rationality-driven, seeking optimal economic benefits, while humans are preference-driven, pursuing the balance of personal goals (\\textit{e.g.,} income and health). These differences hinder the LLMs' ability to effectively understand economic activities across various contexts, leading to biases in economic simulations. To tackle this issue, we introduce \\textbf{EconAI}, a novel approach aimed at enhancing the preference learning capabilities of LLMs by incorporating human-like preferences and cognitive processes. Specifically, EconAI features a 'knowledge brain' constructed from historical data and learning algorithms, enabling memory and making decisions for sophisticated economic facts. By integrating elements of self-learning, reflection, and experience updates, we refine decision-making processes, resulting in more accurate economic planning and mitigating planning bias in economic activities. Through the integration of real-time economic data and historical trends, EconAI offers a robust simulation platform that can adapt to market fluctuations and economic shocks. Our findings demonstrate that EconAI can model economic phenomena like inflation and employment with greater precision, showcase a notable ability to adjust to changing economic conditions, and surpass existing frameworks significantly.", "keywords": [ "Knowledge Agent", "Economic Simulating" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 1, 1, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "HzBfoUdjHt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Chaofan Tao", "gender": "Male", "institution": "The University of Hong Kong", "country": "", "position": "PhD Student" }, { "name": "Longyue Wang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Mi Zhang", "gender": "Male", "institution": "The Ohio State University", "country": "", "position": "Associate Professor" }, { "name": "Siqi Luo", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xin Wang", "gender": "Male", "institution": "The Ohio StateUniversity", "country": "US", "position": "PhD student" }, { "name": "Xinjian Wu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Yi Xin", "gender": "Male", "institution": "Nanjing university", "country": "CN", "position": "PhD student" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "PhD student" }, { "name": "Zhihong Zhu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Zhongwei Wan", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "jing xiong", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 52, "n_formula_1": 15, "n_ref_uni": 39, "n_ref": 94, "n_ref_all": 132, "n_fig": 7, "n_tab": 12, "L_tab": 3878, "n_element_tab": 378, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 4283, "n_element_tab_1": 340, "formula_len_all": 2436, "formula_len_all_1": 992, "len_all": 206767, "len_all_1": 63839, "len_abs": 3729, "len_title": 124, "len_sents": 58477, "len_sents_1": 24723, "n_sents": 438, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1432, "title": "Dynamic Discriminative Operations for Efficient Generative Inference of LLMs", "abs": "Efficient generative inference in Large Language Models (LLMs) is impeded by the growing memory demands of Key-Value (KV) cache, especially for longer sequences. Traditional KV Cache eviction strategies, which discard less critical KV-pairs based on attention scores, often degrade generation quality, leading to issues such as context loss or hallucinations. To address this, we introduce **D**ynamic **D**iscriminative **O**perations ($\\mathbf{D_2 O}$), a novel method that optimizes KV cache size dynamically and discriminatively at two levels without fine-tuning, while preserving essential context. At **layer-level**, by observing the varying densities of attention weights between shallow and deep layers, we dynamically determine which layers should avoid excessive eviction via our proposed ***dynamic allocation strategy*** to minimize information loss. At **token-level**, for the eviction strategy in each layer, $\\mathbf{D_2 O}$ innovatively incorporates a ***compensation mechanism*** that maintains a similarity threshold to re-discriminate the importance of currently discarded tokens, determining whether they should be recalled and merged with similar tokens. Extensive experiments on various benchmarks and LLM architectures have shown that $\\mathbf{D_2 O}$ not only achieves significant memory savings and enhances inference throughput by more than 3$\\times$ but also maintains high-quality long-text generation.", "keywords": [ "Efficient inference of LLMs", "Long context" ], "rating_list": [ 6, 6, 6, 5, 6 ], "soundness_list": [ 4, 2, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 4, 5, 3, 4, 4 ] }, { "paper_id": "Hz4BYVY8YM", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Changsheng Xu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Dizhan Xue", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Fan Yang", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Jiahong Wu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Shengsheng Qian", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Weiming Dong", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Yuhang Hu", "gender": "unknown", "institution": "Zhengzhou University", "country": "CN", "position": "MS student" }, { "name": "Zemin Du", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" }, { "name": "Zhenyu Yang", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 24, "n_ref": 35, "n_ref_all": 59, "n_fig": 12, "n_tab": 5, "L_tab": 3891, "n_element_tab": 234, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 1208, "n_element_tab_1": 67, "formula_len_all": 365, "formula_len_all_1": 366, "len_all": 225816, "len_all_1": 58283, "len_abs": 1644, "len_title": 138, "len_sents": 67636, "len_sents_1": 24921, "n_sents": 514, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 89, "L_abs": 1640, "title": "SVBench: A Benchmark with Temporal Multi-Turn Dialogues for Streaming Video Understanding", "abs": "Despite the significant advancements of Large Vision-Language Models (LVLMs) on established benchmarks, there remains a notable gap in suitable evaluation regarding their applicability in the emerging domain of long-context streaming video understanding. Current benchmarks for video understanding typically emphasize isolated single-instance text inputs and fail to evaluate the capacity to sustain temporal reasoning throughout the entire duration of video streams. To address these limitations, we introduce SVBench, a pioneering benchmark with temporal multi-turn question-answering chains specifically designed to thoroughly assess the capabilities of streaming video understanding of current LVLMs. We design a semi-automated annotation pipeline to obtain 49,979 Question-Answer (QA) pairs of 1,353 streaming videos, which includes generating QA chains that represent a series of consecutive multi-turn dialogues over video segments and constructing temporal linkages between successive QA chains. Our experimental results, obtained from 14 models in dialogue and streaming evaluations, reveal that while the closed-source GPT-4o outperforms others, most open-source LVLMs struggle with long-context streaming video understanding. We also construct a StreamingChat model, which significantly outperforms open-source LVLMs on our SVBench and achieves comparable performance on diverse vision-language benchmarks. We expect SVBench to advance the research of streaming video understanding by providing a comprehensive and in-depth analysis of current LVLMs. Our benchmark and model can be accessed at https://yzy-bupt.github.io/SVBench.", "keywords": [ "Multimodal large language model", "Streaming video analysis", "Video understanding" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "HyqTTe85MZ", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Dinesh Manocha", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Professor" }, { "name": "Matthias Zwicker", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Samuel Audia", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 16, "n_ref_uni": 34, "n_ref": 62, "n_ref_all": 83, "n_fig": 9, "n_tab": 2, "L_tab": 834, "n_element_tab": 144, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1421, "formula_len_all_1": 1071, "len_all": 145220, "len_all_1": 68565, "len_abs": 1252, "len_title": 133, "len_sents": 49030, "len_sents_1": 33104, "n_sents": 413, "n_sents_1": 281, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1247, "title": "Neural Electrostatics: A 3D Physics-Informed Boundary Element Poisson Equation Solver", "abs": "Electrostatics solvers relate an imposed voltage to a\ncorresponding charge density. Current classical methods require fine\ndiscretization and scale poorly due to the construction of a large linear system\nof equations. We recast the problem using neural networks and introduce\nneural electrostatics, a hybrid 3D boundary element method (BEM). By using the\nboundary element form, we are able to overcome many shortcomings of previous\nneural solvers, such as learning trivial solutions and balancing loss terms\nbetween the domain and boundary, at the cost of introducing a large integral\ncontaining a singular kernel. We handle this singularity by locally\ntransforming the integral into polar coordinates and applying a numerical\nquadrature. We also show that previous neural solver sampling methods are unable\nto minimize the PDE residual, and propose a variational adaptive sampling\nmethod. This technique is able to reduce mean absolute error by 5 times, while\nkeeping training time constant. Extensive scaling and ablation studies are\nperformed to justify our method. Results show that our method learns a charge\ndistribution within 1.2 $pC/m^2$ of mean absolute error from a classical BEM\nsolver, while using 25 times fewer rectangular elements.", "keywords": [ "PINN", "physics informed", "scientific machine learning", "electrostatics", "scientific computing" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "HyjIEf90Tn", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Dheeraj Mysore Nagaraj", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Harshit Varma", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Karthikeyan Shanmugam", "gender": "Male", "institution": "Google Deepmind", "country": "IN", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 9, "n_ref_uni": 71, "n_ref": 143, "n_ref_all": 177, "n_fig": 11, "n_tab": 3, "L_tab": 521, "n_element_tab": 31, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 987, "n_element_tab_1": 90, "formula_len_all": 2401, "formula_len_all_1": 1309, "len_all": 246806, "len_all_1": 66776, "len_abs": 1848, "len_title": 126, "len_sents": 75739, "len_sents_1": 28830, "n_sents": 677, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1233, "title": "Glauber Generative Model: Discrete Diffusion Models via Binary Classification", "abs": "We introduce the Glauber Generative Model (GGM), a new class of discrete diffusion models, to obtain new samples from a distribution given samples from a discrete space. GGM deploys a discrete Markov chain called the heat bath dynamics (or the Glauber dynamics) to denoise a sequence of noisy tokens to a sample from a joint distribution of discrete tokens. Our novel conceptual framework provides an exact reduction of the task of learning the denoising Markov chain to solving a class of binary classification tasks. More specifically, the model learns to classify a given token in a noisy sequence as signal or noise. In contrast, prior works on discrete diffusion models either solve regression problems to learn importance ratios, or minimize loss functions given by variational approximations. We apply GGM to language modeling and image generation, where images are discretized using image tokenizers like VQGANs. We show that it outperforms existing discrete diffusion models in language generation, and demonstrates strong performance for image generation without using dataset-specific image tokenizers. We also show that our model is capable of performing well in zero-shot control settings like text and image infilling.", "keywords": [ "Discrete Diffusion Models", "Diffusion Models", "Non-Autoregressive Models", "Generative Models", "Markov Chains", "MCMC" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "HyS9pkHNTN", "primary_area": "applications to robotics, autonomy, planning", "n_author": 8, "author_info_list": [ { "name": "Dechen Gao", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "PhD student" }, { "name": "Hanchu Zhou", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "PhD student" }, { "name": "Hang Wang", "gender": "Male", "institution": "University of California, Davis", "country": "", "position": "PhD student" }, { "name": "Iman Soltani", "gender": "Male", "institution": "University of California, Davis", "country": "", "position": "Assistant Professor" }, { "name": "Junshan Zhang", "gender": "Male", "institution": "University of California, Davis", "country": "", "position": "Full Professor" }, { "name": "Nejib Ammar", "gender": "unknown", "institution": "University of California, Davis", "country": "", "position": "Principal Researcher" }, { "name": "Shatadal Mishra", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Shuangyu Cai", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 72, "n_ref": 134, "n_ref_all": 163, "n_fig": 16, "n_tab": 3, "L_tab": 608, "n_element_tab": 115, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 297, "n_element_tab_1": 23, "formula_len_all": 601, "formula_len_all_1": 338, "len_all": 213331, "len_all_1": 68946, "len_abs": 1875, "len_title": 141, "len_sents": 64663, "len_sents_1": 34490, "n_sents": 471, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1574, "title": "World-Model based Hierarchical Planning with Semantic Communications for Autonomous Driving", "abs": "World-model (WM) is a highly promising approach for training AI agents. However, in complex learning systems such as autonomous driving, AI agents interact with others in a dynamic environment and face significant challenges such as partial observability and non-stationarity. Inspired by how humans naturally solve complex tasks hierarchically and how drivers share their intentions by using turn signals, we introduce HANSOME, a WM-based hierarchical planning with semantic communications framework. In HANSOME, semantic information, particularly text and compressed visual data, is generated and shared to improve two-level planning. HANSOME incorporates two important designs: 1) A hierarchical planning strategy, where the higher-level policy generates intentions with text semantics, and a semantic alignment technique ensures the lower-level policy determines specific controls to achieve these intentions. 2) A cross-modal encoder-decoder to fuse and utilize the shared semantic information to enhance planning through multi-modal understanding. A key advantage of HANSOME is that the generated intentions not only enhance the lower-level policy but also can be shared and understood by humans or other AVs to improve their planning. Furthermore, we devise AdaSMO, an entropy-controlled adaptive scalarization method, to tackle the multi-objective optimization problem in hierarchical policy learning. Extensive experiments show that HANSOME outperforms state-of-the-art WM-based methods in challenging driving tasks, enhancing overall traffic safety and efficiency.", "keywords": [ "World Model", "Hierarchical Planning", "Reinforcement Learning", "Autonomous Driving", "Communications" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 2 ] }, { "paper_id": "HyPofygOCT", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 7, "author_info_list": [ { "name": "Dawei Yang", "gender": "Male", "institution": "Houmo", "country": "CN", "position": "Researcher" }, { "name": "Guangyu Sun", "gender": "Male", "institution": "Peking University", "country": "", "position": "Associate Professor" }, { "name": "Qiang Wu", "gender": "Male", "institution": "Houmo.AI", "country": "CN", "position": "Principal Researcher" }, { "name": "Yan Yan", "gender": "Male", "institution": "University of Illinois Chicago", "country": "US", "position": "Associate Professor" }, { "name": "Yue Song", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yuzhang Shang", "gender": "Male", "institution": "Illinois Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zhihang Yuan", "gender": "Male", "institution": "Infinigence AI", "country": "CN", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 42, "n_ref": 100, "n_ref_all": 116, "n_fig": 6, "n_tab": 4, "L_tab": 1444, "n_element_tab": 282, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1444, "n_element_tab_1": 282, "formula_len_all": 559, "formula_len_all_1": 327, "len_all": 145662, "len_all_1": 75690, "len_abs": 1400, "len_title": 138, "len_sents": 45112, "len_sents_1": 32866, "n_sents": 352, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1352, "title": "ASVD: Activation-aware Singular Value Decomposition for Compressing Large Language Models", "abs": "In this paper, we introduce a new post-training compression paradigm for Large Language Models (LLMs) to facilitate their wider adoption. We delve into LLM weight low-rank decomposition, and find that the challenges of this task stem from \u2776 the distribution variance in the LLM activations and \u2777 the sensitivity difference among various kinds of layers. To address these issues, we propose a training-free approach called Activation-aware Singular Value Decomposition (ASVD). Specif- ically, \u2776 ASVD manages activation outliers by transforming the weight matrix based on the activation distribution. This transformation allows the outliers in the activation matrix to be absorbed into the transformed weight matrix, thereby enhancing decomposition accuracy. \u2777 Additionally, we propose an efficient iter- ative calibration process to optimize layer-specific decomposition by addressing the varying sensitivity of different LLM layers. In this way, ASVD can compress a network by 10%-30%. Based on the success of the low-rank decomposition of projection matrices in the self-attention module, we further introduce ASVD to compress the KV cache. By reducing the channel dimension of KV activations, memory requirements for KV cache can be largely reduced. ASVD can further achieve 50% KV cache reductions without performance drop in a training-free manner.", "keywords": [ "LLM", "model compression", "Low-rank decomposition", "efficient AI" ], "rating_list": [ 6, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "HyN9POiYhN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Barry-John Theobald", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Krishna Patel", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Luca Zappella", "gender": "Male", "institution": "Apple", "country": "ES", "position": "Principal Researcher" }, { "name": "Natalie Mackraz", "gender": "Female", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Nicholas Apostoloff", "gender": "unknown", "institution": "Apple", "country": "GB", "position": "Principal Researcher" }, { "name": "Nivedha Sivakumar", "gender": "Not Specified", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "MS student" }, { "name": "Samira Khorshidi", "gender": "Female", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 34, "n_ref": 55, "n_ref_all": 99, "n_fig": 8, "n_tab": 4, "L_tab": 2134, "n_element_tab": 308, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1106, "n_element_tab_1": 158, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 131107, "len_all_1": 59310, "len_abs": 1254, "len_title": 86, "len_sents": 42307, "len_sents_1": 29822, "n_sents": 273, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1268, "title": "Evaluating Gender Bias Transfer between Pre-trained and Prompt-Adapted Language Models", "abs": "Large language models (LLMs) are increasingly being adapted to new tasks and deployed in real-world decision systems. Several previous works have investigated the bias transfer hypothesis (BTH) and find that fairness of pre-trained masked language models has limited effect on the fairness of these models when adapted using fine-tuning. In this work, we expand the study of BTH to causal models under prompt adaptations, as prompting is an accessible, and compute-efficient way to deploy models in real-world systems. In contrast to previous work, we establish that intrinsic biases in pre-trained Mistral, Falcon and Llama models are strongly correlated (rho >= 0.94) with biases when the same models are zero- and few-shot prompted, using a pronoun co-reference resolution task. Further, we find that biases remain strongly correlated even when LLMs are specifically pre-prompted to exhibit fair or biased behavior (rho >= 0.92), and also when varying few shot composition parameters such as sample size, stereotypical content, occupational distribution and representational balance (rho >= 0.90). Our findings highlight the importance of ensuring fairness in pre-trained LLMs, especially when they are later used to perform downstream tasks via prompt adaptation.", "keywords": [ "fairness", "bias transfer hypothesis", "prompt adaptation", "large language models", "coreference resolution" ], "rating_list": [ 8, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "Hxm0hOxph2", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Amin Mansouri", "gender": "Not Specified", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "", "position": "MS student" }, { "name": "Kartik Ahuja", "gender": "Male", "institution": "FAIR (Meta)", "country": "FR", "position": "Researcher" } ], "n_formula": 198, "n_formula_1": 15, "n_ref_uni": 41, "n_ref": 82, "n_ref_all": 121, "n_fig": 10, "n_tab": 5, "L_tab": 716, "n_element_tab": 36, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 13478, "formula_len_all_1": 1350, "len_all": 366122, "len_all_1": 85706, "len_abs": 4064, "len_title": 99, "len_sents": 105635, "len_sents_1": 35687, "n_sents": 1065, "n_sents_1": 268, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1091, "title": "On Provable Length and Compositional Generalization", "abs": "Out-of-distribution generalization capabilities of sequence-to-sequence models can be studied from the lens of two crucial forms of generalization: length generalization -- the ability to generalize to longer sequences than ones seen during training, and compositional generalization: the ability to generalize to token combinations not seen during training. In this work, we provide first provable guarantees on length and compositional generalization for common sequence-to-sequence models -- deep sets, transformers, state space models, and recurrent neural nets -- trained to minimize the prediction error. Taking a first principles perspective, we study the realizable case, i.e., the labeling function is realizable on the architecture. We show that simple limited capacity versions of these different architectures achieve both length and compositional generalization. Across different architectures, we also find that a linear relationship between the learned representation and the representation in the labeling function is necessary for length and compositional generalization.", "keywords": [ "sequence-to-sequence models", "length generalization", "compositional generalization", "out-of-distribution generalization" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "HxKSzulSD1", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Gong Zhi", "gender": "Male", "institution": "WeChat", "country": "", "position": "Researcher" }, { "name": "Guangyao Shen", "gender": "Male", "institution": "Tencent", "country": "", "position": "Researcher" }, { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Shiqi Shen", "gender": "Male", "institution": "Wechat, Tencent", "country": "", "position": "Researcher" }, { "name": "Wei Yao", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Wenkai Yang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Yankai Lin", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "Assistant Professor" }, { "name": "Yong Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 19, "n_formula_1": 13, "n_ref_uni": 33, "n_ref": 86, "n_ref_all": 135, "n_fig": 27, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 15, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1258, "formula_len_all_1": 851, "len_all": 169161, "len_all_1": 56071, "len_abs": 2277, "len_title": 143, "len_sents": 59002, "len_sents_1": 27117, "n_sents": 384, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1696, "title": "Super(ficial)-alignment: Strong Models May Deceive Weak Models in Weak-to-Strong Generalization", "abs": "Superalignment, where humans act as weak supervisors for superhuman models, has become a crucial problem with the rapid development of Large Language Models (LLMs). Recent work has preliminarily studied this problem by using weak models to supervise strong models, and discovered that weakly supervised strong students can consistently outperform weak teachers towards the alignment target, leading to a weak-to-strong generalization phenomenon. However, we are concerned that behind such a promising phenomenon, whether there exists an issue of weak-to-strong deception, where strong models deceive weak models by exhibiting well-aligned in areas known to weak models but producing misaligned behaviors in cases weak models do not know. We take an initial step towards exploring this security issue in a specific but realistic multi-objective alignment case, where there may be some alignment targets conflicting with each other (e.g., helpfulness v.s. harmlessness). We aim to explore whether, in such cases, strong models might deliberately make mistakes in areas known to them but unknown to weak models within one alignment dimension, in exchange for a higher reward in another dimension. Through extensive experiments in both the reward modeling and preference optimization scenarios, we find: (1) The weak-to-strong deception phenomenon exists across all settings. (2) The deception intensifies as the capability gap between weak and strong models increases. (3) Bootstrapping with an intermediate model can mitigate the deception to some extent, though its effectiveness remains limited. Our work highlights the urgent need to pay more attention to the true reliability of superalignment.", "keywords": [ "Superalignment", "Weak-to-Strong", "Security" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "Hx2ADQLi8M", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Chang Zeng", "gender": "Male", "institution": "National Institute of Informatics", "country": "", "position": "PhD student" }, { "name": "Guo Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Kai Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Runxuan Yang", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Wendi Sang", "gender": "unknown", "institution": "Qinghai University", "country": "CN", "position": "MS student" }, { "name": "Xiaolin Hu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 0, "n_ref_uni": 51, "n_ref": 162, "n_ref_all": 190, "n_fig": 7, "n_tab": 5, "L_tab": 1624, "n_element_tab": 127, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 316, "n_element_tab_1": 85, "formula_len_all": 441, "formula_len_all_1": 0, "len_all": 203646, "len_all_1": 62885, "len_abs": 1583, "len_title": 148, "len_sents": 68041, "len_sents_1": 31059, "n_sents": 436, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 99, "L_abs": 1572, "title": "SonicSim: A customizable simulation platform for speech processing in moving sound source scenarios", "abs": "Systematic evaluation of speech separation and enhancement models under moving sound source conditions requires extensive and diverse data. However, real-world datasets often lack sufficient data for training and evaluation, and synthetic datasets, while larger, lack acoustic realism. Consequently, neither effectively meets practical needs. To address this issue, we introduce SonicSim, a synthetic toolkit based on the embodied AI simulation platform Habitat-sim, designed to generate highly customizable data for moving sound sources. SonicSim supports multi-level adjustments\u2014including scene-level, microphone-level, and source-level\u2014enabling the creation of more diverse synthetic data. Leveraging SonicSim, we constructed a benchmark dataset called SonicSet, utilizing LibriSpeech, Freesound Dataset 50k (FSD50K), Free Music Archive (FMA), and 90 scenes from Matterport3D to evaluate speech separation and enhancement models. Additionally, to investigate the differences between synthetic and real-world data, we selected 5 hours of raw, non-reverberant data from the SonicSet validation set and recorded a real-world speech separation dataset, providing a reference for comparing SonicSet with other synthetic datasets. For speech enhancement, we utilized the real-world dataset RealMAN to validate the acoustic gap between SonicSet and existing synthetic datasets. The results indicate that models trained on SonicSet generalize better to real-world scenarios compared to other synthetic datasets. Code is publicly available at ***https://cslikai.cn/SonicSim/***.", "keywords": [ "Moving audio toolkit", "moving audio dataset", "speech separation", "speech enhancement" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "HwwT4HuRSD", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Alexander Effland", "gender": "Male", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "Associate Professor" }, { "name": "Erich Kobler", "gender": "unknown", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Assistant Professor" }, { "name": "Soma Hansel", "gender": "Male", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 20, "n_ref": 45, "n_ref_all": 54, "n_fig": 2, "n_tab": 6, "L_tab": 982, "n_element_tab": 85, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1128, "n_element_tab_1": 91, "formula_len_all": 212, "formula_len_all_1": 212, "len_all": 91657, "len_all_1": 58711, "len_abs": 1030, "len_title": 97, "len_sents": 29673, "len_sents_1": 27310, "n_sents": 217, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 48, "L_abs": 1038, "title": "FedPCE: Federated Personalized Client Embeddings", "abs": "Despite recent efforts, federated learning (FL) still faces performance challenges due to non-IID data distributions among clients. \nThis distribution shift complicates the addition of new clients and the transfer of federally learned models to unseen data. \nInspired by the adaptation ability of normalization layer parameters, we first demonstrate the effectiveness of models trained using FedBN when being adapted to so far unseen data.\nSpecifically, we extend the adaptation method based on a visual analysis of the normalization layer feature vectors. \nWe introduce Federated Personalized Client Embeddings (FedPCE), which utilizes local embeddings to capture the underlying structure of the normalization feature vectors and, by extension, the dataset.\nOur results show that FedPCE performs comparably to other common FL algorithms during both training and adaptation. \nNotably, FedPCE achieves this performance using only a fraction of the parameters during fine-tuning (32 parameters in our experiments) compared to other methods.", "keywords": [ "federated learning", "deep learning", "computer vision", "transfer learning", "model personalization", "image classification", "domain adaptation" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "HwkELcW2ft", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Qi Zhang", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "TENG LI", "gender": "unknown", "institution": "National University of Defense Technology", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 39, "n_ref": 60, "n_ref_all": 73, "n_fig": 5, "n_tab": 9, "L_tab": 4391, "n_element_tab": 585, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 3056, "n_element_tab_1": 282, "formula_len_all": 350, "formula_len_all_1": 350, "len_all": 130303, "len_all_1": 66422, "len_abs": 1204, "len_title": 93, "len_sents": 33912, "len_sents_1": 30577, "n_sents": 199, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 45, "L_abs": 1207, "title": "Searching For Robust Point Cloud Distillation", "abs": "Deep Neural Networks (DNNs) have shown remarkable performance in machine learning; however, their vulnerabilities to adversarial attacks have been exposed, particularly in point cloud data. Neural Architecture Search (NAS) is a technique for discovering new neural architectures with high predictive accuracy, yet its potential for enhancing model robustness against adversarial attacks remains largely unexplored. In this study, we investigate the application of NAS within the framework of knowledge distillation, aiming to generate robust student architectures that inherit resilience from robust teacher models. We introduce RDANAS, an effective NAS method that utilizes cross-layer knowledge distillation from robust teacher models to enhance the robustness of the student model. Unlike previous studies, RDANAS considers the teacher model's outputs and automatically identifies the optimal teacher layer for each student layer during supervision. Experimental results on ModelNet40, ScanObjectNN and ScanNet datasets demonstrate the efficacy of RDANAS, revealing that the neural architectures it generates are compact and possess adversarial robustness, which shows potential in multiple applications.", "keywords": [ "Neural Architecture Search", "Adversarial Attacks", "Knowledge Distillation" ], "rating_list": [ 3, 1, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "Hw1tOjCWBZ", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 7, "author_info_list": [ { "name": "Ge Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Kaijing Ma", "gender": "Female", "institution": "Tongji University", "country": "", "position": "Undergrad student" }, { "name": "Ningxuan Lu", "gender": "Female", "institution": "Duke University", "country": "US", "position": "MS student" }, { "name": "Qingkun Tang", "gender": "unknown", "institution": "zte", "country": "CN", "position": "Researcher" }, { "name": "Siyuan Fang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Tianyu Zheng", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Xinrun Du", "gender": "Male", "institution": "01.AI", "country": "CN", "position": "Intern" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 48, "n_ref": 90, "n_ref_all": 132, "n_fig": 7, "n_tab": 21, "L_tab": 5723, "n_element_tab": 747, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2972, "n_element_tab_1": 41, "formula_len_all": 706, "formula_len_all_1": 625, "len_all": 241242, "len_all_1": 56383, "len_abs": 1694, "len_title": 174, "len_sents": 70461, "len_sents_1": 24541, "n_sents": 553, "n_sents_1": 161, "venue": "ICLR2025", "decision": "withdraw", "L_title": 125, "L_abs": 1634, "title": "KARPA: A Training-free Method of Adapting Knowledge Graph as References for Large Language Model's Reasoning Path Aggregation", "abs": "Large language models (LLMs) demonstrate exceptional performance across a variety of tasks, yet they are often affected by hallucinations and the timeliness of knowledge. Leveraging knowledge graphs (KGs) as external knowledge sources has emerged as a viable solution, but existing methods for LLM-based knowledge graph question answering (KGQA) are often limited by step-by-step decision-making on KGs, restricting the global planning and reasoning capabilities of LLMs, or they require fine-tuning or pre-training on specific KGs. To address these challenges, we propose Knowledge graph Assisted Reasoning Path Aggregation (KARPA), a novel framework that harnesses the global planning abilities of LLMs for efficient and accurate KG reasoning on KGs. KARPA operates through a three-step process: pre-planning, retrieving, and reasoning. First, KARPA uses the LLM's global planning ability to pre-plan logically coherent relation paths based on the provided question and relevant relations within the KG. Next, in the retrieving phase, relation paths with high semantic similarity to the pre-planned paths are extracted as candidate paths using a semantic embedding model. Finally, these candidate paths are provided to the LLM for comprehensive reasoning. Unlike existing LLM-based KGQA methods, KARPA fully leverages the global planning and reasoning capabilities of LLMs without requiring stepwise traversal or additional training, and it is compatible with various LLM architectures. Extensive experimental results show that KARPA achieves state-of-the-art performance in KGQA tasks, delivering both high efficiency and accuracy.", "keywords": [ "Knowledge Graph", "Large Language Models", "Chain-of-Thought", "Reasoning" ], "rating_list": [ 5, 5, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "HvkXPQhQvv", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Bonnie Berger", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Divya M Shanmugam", "gender": "Female", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "Postdoc" }, { "name": "Emma Pierson", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "Assistant Professor" }, { "name": "John Guttag", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Manish Raghavan", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Shuvom Sadhuka", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 4, "n_ref_uni": 71, "n_ref": 123, "n_ref_all": 153, "n_fig": 6, "n_tab": 10, "L_tab": 3986, "n_element_tab": 269, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 975, "formula_len_all_1": 445, "len_all": 225683, "len_all_1": 67236, "len_abs": 3571, "len_title": 107, "len_sents": 66982, "len_sents_1": 34388, "n_sents": 466, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1558, "title": "Evaluating multiple models using labeled and unlabeled data", "abs": "It remains difficult to evaluate machine learning classifiers in the absence of a large, labeled dataset. While labeled data can be prohibitively expensive or impossible to obtain, unlabeled data is plentiful.\nHere, we introduce Semi-Supervised Model Evaluation (SSME), a method that uses both labeled and unlabeled data to evaluate machine learning classifiers. SSME is the first evaluation method to take advantage of the fact that: (i) there are frequently multiple classifiers for the same task, (ii) continuous classifier scores are often available for all classes, and (iii) unlabeled data is often far more plentiful than labeled data. \nThe key idea is to use a semi-supervised mixture model to estimate the joint distribution of ground truth labels and classifier predictions.\nWe can then use this model to estimate any metric that is a function of classifier scores and ground truth labels (e.g., accuracy or expected calibration error). \nWe present experiments in four domains where obtaining large labeled datasets is often impractical: (1) healthcare, (2) content moderation, (3) molecular property prediction, and (4) image annotation. Our results demonstrate that SSME estimates performance more accurately than do competing methods, reducing error by 5.1x relative to using labeled data alone and 2.4x relative to the next best competing method. SSME also improves accuracy when evaluating performance across subsets of the test distribution (e.g., specific demographic subgroups) and when evaluating the performance of large language models.", "keywords": [ "classifier evaluation", "semi-supervised learning", "unlabeled data" ], "rating_list": [ 6, 5, 5, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "HvSytvg3Jh", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Jie Shi", "gender": "unknown", "institution": "Huawei International.", "country": "SG", "position": "Principal Researcher" }, { "name": "Junfeng Fang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Kun Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Xiang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xiangnan He", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Yunshan Ma", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Jianghoucheng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Tat-seng Chua", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 45, "n_formula_1": 19, "n_ref_uni": 31, "n_ref": 128, "n_ref_all": 166, "n_fig": 15, "n_tab": 12, "L_tab": 7427, "n_element_tab": 305, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 2597, "n_element_tab_1": 84, "formula_len_all": 2270, "formula_len_all_1": 760, "len_all": 204926, "len_all_1": 59627, "len_abs": 1208, "len_title": 119, "len_sents": 61877, "len_sents_1": 25390, "n_sents": 585, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 67, "L_abs": 1142, "title": "AlphaEdit: Null-Space Constrained Model Editing for Language Models", "abs": "Large language models (LLMs) often exhibit hallucinations, producing incorrect or outdated knowledge. Hence, model editing methods have emerged to enable targeted knowledge updates. To achieve this, a prevailing paradigm is the locating-then-editing approach, which first locates influential parameters and then edits them by introducing a perturbation. While effective, current studies have demonstrated that this perturbation inevitably disrupt the originally preserved knowledge within LLMs, especially in sequential editing scenarios.\nTo address this, we introduce AlphaEdit, a novel solution that projects perturbation onto the null space of the preserved knowledge before applying it to the parameters. We theoretically prove that this projection ensures the output of post-edited LLMs remains unchanged when queried about the preserved knowledge, thereby mitigating the issue of disruption. \nExtensive experiments on various LLMs, including LLaMA3, GPT2-XL, and GPT-J, show that AlphaEdit boosts the performance of most locating-then-editing methods by an average of 36.7% with a single line of additional code for projection solely.", "keywords": [ "Model Editing", "Null-Space", "Large Language Model" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 4 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "Hv5L2vcJyy", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Junjie Wang", "gender": "Female", "institution": "Tianjin University", "country": "", "position": "Associate Professor" }, { "name": "Ruifang He", "gender": "Female", "institution": "Tianjin University", "country": "", "position": "Full Professor" }, { "name": "Jinpeng Zhang", "gender": "Male", "institution": "Tianjin University", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 45, "n_ref": 61, "n_ref_all": 75, "n_fig": 3, "n_tab": 4, "L_tab": 1338, "n_element_tab": 120, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1490, "n_element_tab_1": 96, "formula_len_all": 644, "formula_len_all_1": 580, "len_all": 129701, "len_all_1": 57101, "len_abs": 1075, "len_title": 72, "len_sents": 25059, "len_sents_1": 24668, "n_sents": 198, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1085, "title": "Elementary: Pattern-aware Evidence Discovery with Large Language Models", "abs": "The remarkable success of rationale generation provokes precise Evidence Discovery, which aims to identify a small subset of the inputs sufficient to support a given claim. However, existing general extraction methods still fall short in quantifying the support of evidence and ensuring its completeness. This paper introduces a heuristic search framework, Elementary, which formulates the Evidence Discovery as a multi-step prompt construction process. Specifically, we offer a clear perspective that the LLMs prompted with \\emph{according to}, without fine-tuning on domain-specific knowledge, can serve as an excellent reward function to assess sufficiency. Based on this, Elementary explores various potential reasoning patterns and uses future expected rewards, including independent and pattern-aware rewards, to find the optimal prompt as evidence. Experiments on three common task datasets demonstrate that the proposed framework significantly outperforms previous approaches, additional analysis further validates that Elementary has advantages in extracting complex evidence.", "keywords": [ "Evidence Discovery", "Large Language Model" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "HuNoNfiQqH", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Frauke Kreuter", "gender": "Female", "institution": "University of Maryland, College Park", "country": "", "position": "Full Professor" }, { "name": "Sarah Ball", "gender": "Female", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Nina Rimsky", "gender": "Female", "institution": "Anthropic", "country": "GB", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 21, "n_ref": 62, "n_ref_all": 83, "n_fig": 8, "n_tab": 11, "L_tab": 9437, "n_element_tab": 522, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 2181, "n_element_tab_1": 180, "formula_len_all": 89, "formula_len_all_1": 89, "len_all": 195652, "len_all_1": 59191, "len_abs": 1096, "len_title": 138, "len_sents": 66860, "len_sents_1": 26498, "n_sents": 540, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 987, "title": "Understanding Jailbreak Success: A Study of Latent Space Dynamics in Large Language Models", "abs": "Conversational Large Language Models are trained to refuse to answer harmful questions. However, emergent jailbreaking techniques can still elicit unsafe outputs, presenting an ongoing challenge for model alignment. To better understand how different jailbreak types circumvent safeguards, this paper analyses model activations on different jailbreak inputs. We find that it is possible to extract a jailbreak vector from a single class of jailbreaks that works to mitigate jailbreak effectiveness from other classes. This may indicate that different kinds of effective jailbreaks operate via similar internal mechanisms. We investigate a potential common mechanism of harmfulness feature suppression, and provide evidence for its existence by looking at the harmfulness vector component. These findings offer actionable insights for developing more robust jailbreak countermeasures and lay the groundwork for a deeper, mechanistic understanding of jailbreak dynamics in language models.", "keywords": [ "jailbreaks; activation engineering; alignment" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "HuL2yba6Uf", "primary_area": "generative models", "n_author": 1, "author_info_list": [ { "name": "Carl Allen", "gender": "Male", "institution": "Ecole Normale Sup\u00e9rieure de Paris", "country": "FR", "position": "Postdoc" } ], "n_formula": 46, "n_formula_1": 44, "n_ref_uni": 31, "n_ref": 67, "n_ref_all": 75, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5565, "formula_len_all_1": 5087, "len_all": 113541, "len_all_1": 74666, "len_abs": 967, "len_title": 126, "len_sents": 36164, "len_sents_1": 30267, "n_sents": 290, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 989, "title": "Unpicking Data at the Seams: VAEs, Disentanglement and Independent Components", "abs": "Disentanglement, or identifying statistically independent salient factors of the data, is of interest in many aspects of machine learning and statistics, having potential to improve generation of synthetic data with controlled properties, robust classification of features, parsimonious encoding, and greater understanding of the generative process behind the data. Disentanglement arises in various generative paradigms, including Variational Autoencoders (VAEs), GANs and diffusion models, and particular progress has recently been made in understanding the former. That line of research shows that the choice of diagonal posterior covariance matrices in a VAE promotes mutual orthogonality between columns of the decoder's Jacobian. We continue this thread to show how such *linear* independence translates to *statistical* independence, completing the chain in understanding how the VAE objective leads to the identification of independent components of the data, i.e. disentanglement.", "keywords": [ "VAE", "variational autoencoder", "beta-VAE", "disentanglement" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 3, 1 ], "confidence_list": [ 4, 2, 3 ] }, { "paper_id": "HuC8dszO8r", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Aaron D Tranter", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "Postdoc" }, { "name": "Avishek Anand", "gender": "Male", "institution": "Delft University of Technology", "country": "", "position": "Associate Professor" }, { "name": "Dennis Wilken", "gender": "unknown", "institution": "Universit\u00e4t Hannover", "country": "DE", "position": "Postdoc" }, { "name": "Jan Heine", "gender": "Male", "institution": "Universit\u00e4t Hannover", "country": "", "position": "PhD student" }, { "name": "Lea Richtmann", "gender": "Not Specified", "institution": "Universit\u00e4t Hannover", "country": "", "position": "PhD student" }, { "name": "Mich\u00e8le Heurs", "gender": "Female", "institution": "Universit\u00e4t Hannover", "country": "DE", "position": "Full Professor" }, { "name": "Tobias J. Osborne", "gender": "Male", "institution": "Leibniz Universit\u00e4t Hannover", "country": "", "position": "Full Professor" }, { "name": "Viktoria-S. Schmiesing", "gender": "Female", "institution": "Universit\u00e4t Hannover", "country": "DE", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 83, "n_ref": 98, "n_ref_all": 164, "n_fig": 24, "n_tab": 4, "L_tab": 106, "n_element_tab": 12, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 179, "n_element_tab_1": 19, "formula_len_all": 741, "formula_len_all_1": 280, "len_all": 225493, "len_all_1": 72973, "len_abs": 951, "len_title": 97, "len_sents": 64931, "len_sents_1": 32590, "n_sents": 541, "n_sents_1": 279, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 954, "title": "Model-free reinforcement learning with noisy actions for automated experimental control in optics", "abs": "Setting up and controlling optical systems is often a challenging and tedious task. The high number of degrees of freedom to control mirrors, lenses or phases makes automatic control challenging, especially when the complexity of the system cannot be adequately modeled due to noise or non-linearities. Here, we show that reinforcement learning (RL) can overcome these challenges when coupling laser light into an optical fiber, using a model-free RL approach that trains directly on the experiment without pre-training. By utilizing the sample-efficient algorithms Soft Actor-Critic (SAC) or Truncated Quantile Critics (TQC), our agent learns to couple with 90% efficiency, comparable to the human expert. We demonstrate that direct training on an experiment can replace extensive system modeling. Our result exemplifies RL's potential to tackle problems in optics, paving the way for more complex applications where full noise modeling is not feasible.", "keywords": [ "reinforcement learning", "model-free", "experimental control", "optics", "fiber coupling" ], "rating_list": [ 3, 5, 8, 5 ], "soundness_list": [ 1, 2, 4, 4 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 1, 4, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "HuBFimORiz", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Jianxi Gao", "gender": "Male", "institution": ", Rensselaer Polytechnic Institute", "country": "", "position": "Associate Professor" }, { "name": "Songtao Lu", "gender": "Male", "institution": "IBM Thomas J. Watson Research Center", "country": "", "position": "Senior Research Scientist" }, { "name": "Tomasz J Nowicki", "gender": "Male", "institution": "IBM, International Business Machines", "country": "", "position": "RSM" }, { "name": "Yanna Ding", "gender": "Female", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "PhD student" }, { "name": "Yingdong Lu", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" } ], "n_formula": 106, "n_formula_1": 31, "n_ref_uni": 34, "n_ref": 96, "n_ref_all": 120, "n_fig": 20, "n_tab": 7, "L_tab": 1218, "n_element_tab": 158, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 582, "n_element_tab_1": 53, "formula_len_all": 42644, "formula_len_all_1": 26766, "len_all": 247349, "len_all_1": 108937, "len_abs": 1424, "len_title": 59, "len_sents": 63086, "len_sents_1": 38359, "n_sents": 564, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1435, "title": "Global Optimality of In-context Markovian Dynamics Learning", "abs": "Transformers have demonstrated impressive capability of in-context learning (ICL): given a sequence of input-output pairs of an unseen task, a trained transformer can make reasonable predictions on query inputs, without fine-tuning its parameters. \nHowever, existing studies on ICL have mainly focused on linear regression tasks, often with i.i.d. inputs within a prompt.\nThis paper seeks to unveil the mechanism of ICL for next-token prediction for Markov chains, focusing on the transformer architecture with linear self-attention (LSA). \nMore specifically, we derive and interpret the global optimum of the ICL loss landscape:\n(1) We provide the closed-form expression of the global minimizer for single-layer LSA trained over random instances of length-2 in-context Markov chains, showing the Markovian data distribution necessitates a denser global minimum structure compared to ICL for linear tasks.\n(2) We establish tight bounds for the global minimum of single-layer LSA trained on arbitrary-length Markov chains.\n(3) Finally, we prove that multilayer LSA, with parameterization mirroring the global minimizer's structure, performs preconditioned gradient descent for a multi-objective optimization problem over the in-context samples, balancing a squared loss with multiple linear objectives.\nWe numerically explore ICL for Markov chains using both simplified transformers and GPT-2-based multilayer nonlinear transformers.", "keywords": [ "transformers", "in-context learning", "Markov Chains", "next token prediction" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "Hu0FSOSEyS", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Constantine Caramanis", "gender": "Male", "institution": "University of Texas, Austin", "country": "US", "position": "Full Professor" }, { "name": "Litu Rout", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Nataniel Ruiz", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Sanjay Shakkottai", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Full Professor" }, { "name": "Wen-Sheng Chu", "gender": "unknown", "institution": "Google Research", "country": "US", "position": "Research Scientist" }, { "name": "Yujia Chen", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 74, "n_formula_1": 31, "n_ref_uni": 40, "n_ref": 161, "n_ref_all": 257, "n_fig": 28, "n_tab": 6, "L_tab": 1151, "n_element_tab": 165, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 449, "n_element_tab_1": 68, "formula_len_all": 6859, "formula_len_all_1": 1544, "len_all": 205342, "len_all_1": 66317, "len_abs": 1314, "len_title": 134, "len_sents": 63088, "len_sents_1": 29481, "n_sents": 581, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1324, "title": "Semantic Image Inversion and Editing using Rectified Stochastic Differential Equations", "abs": "Generative models transform random noise into images; their inversion aims to transform images back to structured noise for recovery and editing. This paper addresses two key tasks: (i) *inversion* and (ii) *editing* of a real image using stochastic equivalents of rectified flow models (such as Flux). Although Diffusion Models (DMs) have recently dominated the field of generative modeling for images, their inversion presents faithfulness and editability challenges due to nonlinearities in drift and diffusion. Existing state-of-the-art DM inversion approaches rely on training of additional parameters or test-time optimization of latent variables; both are expensive in practice. Rectified Flows (RFs) offer a promising alternative to diffusion models, yet their inversion has been underexplored. We propose RF inversion using dynamic optimal control derived via a linear quadratic regulator. We prove that the resulting vector field is equivalent to a rectified stochastic differential equation. Additionally, we extend our framework to design a stochastic sampler for Flux. Our inversion method allows for state-of-the-art performance in zero-shot inversion and editing, outperforming prior works in stroke-to-image synthesis and semantic image editing, with large-scale human evaluations confirming user preference.", "keywords": [ "Inverse Problems", "Generative Modeling", "Diffusion Models", "Rectified Flows", "Posterior Sampling", "Optimal Control" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "HtvZCGiATs", "primary_area": "causal reasoning", "n_author": 8, "author_info_list": [ { "name": "Anton van den Hengel", "gender": "unknown", "institution": "University of Adelaide", "country": "AU", "position": "Professor" }, { "name": "Biwei Huang", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Dong Gong", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "Assistant Professor" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Mingming Gong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Yuhang Liu", "gender": "unknown", "institution": "The University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Zhen Zhang", "gender": "Male", "institution": "The University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Javen Shi", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Professor" } ], "n_formula": 54, "n_formula_1": 13, "n_ref_uni": 55, "n_ref": 141, "n_ref_all": 214, "n_fig": 5, "n_tab": 4, "L_tab": 1992, "n_element_tab": 342, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5752, "formula_len_all_1": 1002, "len_all": 204294, "len_all_1": 72597, "len_abs": 1534, "len_title": 114, "len_sents": 64241, "len_sents_1": 37015, "n_sents": 453, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1539, "title": "Beyond DAGs: A Latent Partial Causal Model for Multimodal Learning", "abs": "Directed acyclic graphs (DAGs) are often assumed in causal discovery, however, accurately identifying these DAGs necessitates various assumptions, particularly in latent causal models, which can be challenging to validate in real-world applications. This raises a critical question: Are DAG assumptions truly necessary for certain applications? In this work, we introduce a novel latent partial causal model for multimodal data, which features two latent coupled variables, connected by an undirected edge, effectively representing transferable knowledge across different modalities. We focus on a prominent learning framework, e.g., multimodal contrastive learning, and demonstrate that, with certain statistical assumptions, multimodal contrastive learning successfully identifies the latent coupled variables up to trivial transformation. This finding enhances our understanding of the mechanisms driving the success of multimodal contrastive learning. Furthermore, this finding reveals a unique potential for disentanglement in multimodal contrastive representation learning, improving the utility of pre-trained models like CLIP that are trained using this approach. Through experiments with synthetic data, we demonstrate the robustness of our findings, even in the presence of violated assumptions. In addition, we validate the disentanglement capabilities of pre-trained CLIP in learning disentangled representations, facilitating few-shot learning and improving domain generalization across a diverse range of real-world datasets.", "keywords": [ "Latent Causal Models", "Multimodal Learning", "Identifiability", "Contrastive Learning" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "HtbqsbNw9c", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Cengiz Oztireli", "gender": "Unspecified", "institution": "University of Cambridge", "country": "GB", "position": "Full Professor" }, { "name": "Fangcheng Zhong", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "Lecturer" }, { "name": "Jing Yang", "gender": "Female", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "Jingyi Wan", "gender": "Female", "institution": "University of Cambridge", "country": "", "position": "MS student" }, { "name": "Tianhao Walter Wu", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Zhilin Guo", "gender": "unknown", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 21, "n_ref_uni": 42, "n_ref": 106, "n_ref_all": 120, "n_fig": 8, "n_tab": 4, "L_tab": 639, "n_element_tab": 26, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 633, "n_element_tab_1": 24, "formula_len_all": 1026, "formula_len_all_1": 978, "len_all": 158658, "len_all_1": 64622, "len_abs": 1956, "len_title": 158, "len_sents": 51190, "len_sents_1": 32585, "n_sents": 344, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 110, "L_abs": 1965, "title": "Gaussian Head & Shoulders: High Fidelity Neural Upper Body Avatars with Anchor Gaussian Guided Texture Warping", "abs": "The ability to reconstruct realistic and controllable upper body avatars from casual monocular videos is critical for various applications in communication and entertainment. By equipping the most recent 3D Gaussian Splatting representation with head 3D morphable models (3DMM), existing methods manage to create head avatars with high fidelity. However, most existing methods only reconstruct a head without the body, substantially limiting their application scenarios. We found that naively applying Gaussians to model the clothed chest and shoulders tends to result in blurry reconstruction and noisy floaters under novel poses. This is because of the fundamental limitation of Gaussians and point clouds -- each Gaussian or point can only have a single directional radiance without spatial variance, therefore an unnecessarily large number of them is required to represent complicated spatially varying texture, even for simple geometry. In contrast, we propose to model the body part with a neural texture that consists of coarse and pose-dependent fine colors. To properly render the body texture for each view and pose without accurate geometry nor UV mapping, we optimize another sparse set of Gaussians as anchors that constrain the neural warping field that maps image plane coordinates to the texture space. We demonstrate that Gaussian Head & Shoulders can fit the high-frequency details on the clothed upper body with high fidelity and potentially improve the accuracy and fidelity of the head region. We evaluate our method with casual phone-captured and internet videos and show our method archives superior reconstruction quality and robustness in both self and cross reenactment tasks. To fully utilize the efficient rendering speed of Gaussian splatting, we additionally propose an accelerated inference method of our trained model without Multi-Layer Perceptron (MLP) queries and reach a stable rendering speed of around 130 FPS for any subjects.", "keywords": [ "Neural Radiance Field", "Gaussian Splatting", "Neural Head Avatar" ], "rating_list": [ 8, 6, 5, 6, 6 ], "soundness_list": [ 2, 4, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 1, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "HtS2NUcqtD", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Yingyu Liang", "gender": "Male", "institution": "University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Yufa Zhou", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "MS student" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Zhenmei Shi", "gender": "Male", "institution": "Voyage AI", "country": "US", "position": "Researcher" } ], "n_formula": 126, "n_formula_1": 27, "n_ref_uni": 67, "n_ref": 120, "n_ref_all": 137, "n_fig": 4, "n_tab": 3, "L_tab": 599, "n_element_tab": 32, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 510, "n_element_tab_1": 41, "formula_len_all": 19230, "formula_len_all_1": 3295, "len_all": 280558, "len_all_1": 70671, "len_abs": 1095, "len_title": 132, "len_sents": 62008, "len_sents_1": 25457, "n_sents": 738, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1104, "title": "Tensor Attention Training: Provably Efficient Learning of Higher-order Transformers", "abs": "Tensor Attention, a multi-view attention that is able to capture high-order correlations among multiple modalities, can overcome the representational limitations of classical matrix attention. However, the $O(n^3)$ time complexity of tensor attention poses a significant obstacle to its utilization in transformers, where $n$ is the input sequence length. In this work, we prove that the backward gradient of tensor attention training can be computed in almost linear time $n^{1+o(1)}$, the same complexity as its forward computation under the bounded entries assumption. We provide a closed-form solution for the gradient and propose a fast computation method utilizing polynomial approximation methods and tensor algebraic techniques. Furthermore, we prove the necessity and tightness of our assumption through hardness analysis, showing that slightly weakening it renders the gradient problem unsolvable in truly subcubic time. Our theoretical results establish the feasibility of efficient higher-order transformer training and may facilitate practical applications of tensor attention architectures.", "keywords": [ "High Order Attention", "Efficient Gradient Computation", "Tensor Operation" ], "rating_list": [ 8, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 2, 3, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "HsHxSN23rM", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Alexander Amini", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Armin W Thomas", "gender": "Non-Binary", "institution": "Liquid AI", "country": "US", "position": "Researcher" }, { "name": "Michael Poli", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Stefano Massaroli", "gender": "unknown", "institution": "RIKEN", "country": "JP", "position": "Researcher" }, { "name": "Rom Nishijima Parnichkun", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 48, "n_ref": 81, "n_ref_all": 98, "n_fig": 3, "n_tab": 6, "L_tab": 5298, "n_element_tab": 233, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 846, "n_element_tab_1": 40, "formula_len_all": 759, "formula_len_all_1": 626, "len_all": 194242, "len_all_1": 63088, "len_abs": 1171, "len_title": 41, "len_sents": 58760, "len_sents_1": 28186, "n_sents": 416, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 41, "L_abs": 1182, "title": "STAR: Synthesis of Tailored Architectures", "abs": "Iterative improvement of model architectures is fundamental to deep learning: Transformers first enabled scaling, and recent advances in model hybridization have pushed the quality-efficiency frontier. However, optimizing architectures remains challenging and expensive, with a variety of automated or manual approaches that fall short, due to limited progress in the design of search spaces and due to the simplicity of resulting patterns and heuristics. In this work, we propose a new approach for the synthesis of tailored architectures (STAR). Our approach combines a novel search space based on the theory of linear input-varying systems, supporting a hierarchical numerical encoding into architecture genomes. STAR genomes are automatically refined and recombined with gradient-free, evolutionary algorithms to optimize for multiple model quality and efficiency metrics. Using STAR, we optimize large populations of new architectures, leveraging diverse computational units and interconnection patterns, improving over highly-optimized Transformers and striped hybrid models on the frontier of quality, parameter size, and inference cache for autoregressive language modeling.", "keywords": [ "alternative architectures", "deep signal processing", "language models" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "HsB1sQvXML", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Andrei Kucharavy", "gender": "Male", "institution": "University of Applied Sciences Western Switzerland, Sierre (HES-SO Valais)", "country": "", "position": "Researcher" }, { "name": "Henrique Da Silva Gameiro", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "MS student" }, { "name": "Ljiljana Dolamic", "gender": "Female", "institution": "armasuisse", "country": "", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 96, "n_ref_all": 123, "n_fig": 10, "n_tab": 3, "L_tab": 163, "n_element_tab": 4, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 455, "n_element_tab_1": 6, "formula_len_all": 672, "formula_len_all_1": 0, "len_all": 160260, "len_all_1": 55846, "len_abs": 1495, "len_title": 132, "len_sents": 47180, "len_sents_1": 28247, "n_sents": 328, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1349, "title": "LLM Detectors Still Fall Short of Real World: Case of LLM-Generated Short News-Like Posts", "abs": "With the emergence of widely available powerful LLMs, disinformation generated by large Language Models (LLMs) has become a major concern. Historically, LLM detectors have been touted as a solution, but their effectiveness in the real world is still to be proven. In this paper, we focus on an important setting in information operations\u2014short news-like posts generated by moderately sophisticated attackers.\n\nWe demonstrate that existing LLM detectors, whether zero-shot or purpose-trained, are not ready for real-world use in that setting. All tested zero-shot detectors perform inconsistently with prior benchmarks and are highly vulnerable to sampling temperature increase, a trivial attack absent from recent benchmarks. A purpose-trained detector generalizing across LLMs and unseen attacks can be developed, but it fails to generalize to new human-written texts. \n\nWe argue that the former indicates domain-specific benchmarking is needed, while the latter suggests a trade-off between the adversarial evasion resilience and overfitting to the reference human text, with both needing evaluation in benchmarks and currently absent. We believe this suggests a re-consideration of current LLM detector benchmarking approaches and provides a dynamically extensible benchmark to allow it (https://anonymous.4open.science/r/text_llm_detector-3E07).", "keywords": [ "LLM Detection", "misinformation", "benchmarking" ], "rating_list": [ 5, 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 1, 3 ], "presentation_list": [ 3, 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 1, 2 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "Hs1UTIOwKr", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Ameen Ali Ali", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "PhD student" }, { "name": "Ivan Titov", "gender": "unknown", "institution": "University of Edinburgh", "country": "", "position": "Full Professor" }, { "name": "Lior Wolf", "gender": "Male", "institution": "Tel Aviv University", "country": "IL", "position": "Associate Professor" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 39, "n_ref": 74, "n_ref_all": 103, "n_fig": 5, "n_tab": 13, "L_tab": 7366, "n_element_tab": 570, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 718, "formula_len_all_1": 418, "len_all": 197940, "len_all_1": 58232, "len_abs": 1021, "len_title": 114, "len_sents": 52356, "len_sents_1": 29096, "n_sents": 630, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1029, "title": "Mitigating Copy Bias in In-Context Learning through Neuron Pruning", "abs": "Large language models (LLMs) have demonstrated impressive few-shot in-context learning (ICL) abilities. Still, we show that they are sometimes prone to a `copying bias', where they copy answers from provided examples instead of learning the underlying patterns. In this work, we propose a novel and simple method to mitigate such copying bias. First, we create a synthetic task and use the Integrated Gradients method to identify neurons that prioritize copying over generalization. We demonstrate that pruning these neurons consistently improves performance across a diverse set of ICL tasks. We also show that our method is applicable across various LLM architectures, including Transformers and State-Space Models, without requiring modifications. \nIn our analysis, we adopt a task-recognition perspective on ICL and examine task vectors (Hendel et al., 2023) induced by the model. We find that pruning enhances the quality of these vectors, suggesting that the pruned neurons previously hindered effective task recognition.", "keywords": [ "large language models", "in-context learning", "interpretability", "llms" ], "rating_list": [ 5, 8, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "HrdVqFSn1e", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Qiwei Di", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Runjia Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 352, "n_formula_1": 47, "n_ref_uni": 29, "n_ref": 105, "n_ref_all": 112, "n_fig": 0, "n_tab": 10, "L_tab": 3061, "n_element_tab": 307, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 579, "n_element_tab_1": 10, "formula_len_all": 45257, "formula_len_all_1": 3890, "len_all": 295508, "len_all_1": 68130, "len_abs": 1315, "len_title": 137, "len_sents": 71450, "len_sents_1": 27499, "n_sents": 966, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1336, "title": "Unified Convergence Analysis for Score-Based Diffusion Models with Deterministic Samplers", "abs": "Score-based diffusion models have emerged as powerful techniques for generating samples from high-dimensional data distributions. These models involve a two-phase process: first, injecting noise to transform the data distribution into a known prior distribution, and second, sampling to recover the original data distribution from noise. Among the various sampling methods, deterministic samplers stand out for their enhanced efficiency. However, analyzing these deterministic samplers presents unique challenges, as they preclude the use of established techniques such as Girsanov's theorem, which are only applicable to stochastic samplers. Furthermore, existing analysis for deterministic samplers usually focuses on specific examples, lacking a generalized approach for general forward processes and various deterministic samplers. Our paper addresses these limitations by introducing a unified convergence analysis framework. To demonstrate the power of our framework, we analyze the variance-preserving (VP) forward process with the exponential integrator (EI) scheme, achieving iteration complexity of $\\tilde{O}(d^2/\\epsilon)$.\nAdditionally, we provide a detailed analysis of Denoising Diffusion Implicit Models (DDIM)-type samplers, which have been underexplored in previous research, achieving polynomial iteration complexity.", "keywords": [ "Diffusion Models", "Probability Flow ODEs", "Unified Framework", "Deterministic Samplers" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "Hr3TBaZl4S", "primary_area": "reinforcement learning", "n_author": 10, "author_info_list": [ { "name": "Dacheng Tao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Hechang Chen", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jifeng Hu", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "PhD student" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Lichao Sun", "gender": "Male", "institution": "Lehigh University", "country": "US", "position": "Assistant Professor" }, { "name": "Shengchao Hu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Shisong Tang", "gender": "unknown", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Sili Huang", "gender": "Male", "institution": "Jilin University", "country": "", "position": "PhD student" }, { "name": "Yi Chang", "gender": "Male", "institution": "Jilin University, China", "country": "CN", "position": "Full Professor" }, { "name": "Zhejian Yang", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 15, "n_ref_uni": 78, "n_ref": 149, "n_ref_all": 192, "n_fig": 18, "n_tab": 7, "L_tab": 2319, "n_element_tab": 179, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 834, "n_element_tab_1": 47, "formula_len_all": 1715, "formula_len_all_1": 1077, "len_all": 191316, "len_all_1": 65975, "len_abs": 1366, "len_title": 133, "len_sents": 47477, "len_sents_1": 30158, "n_sents": 326, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1371, "title": "Solving Continual Offline RL through Selective Weights Activation on Aligned Spaces", "abs": "Continual offline reinforcement learning (CORL) has shown impressive ability in diffusion-based continual learning systems by modeling the joint distributions of trajectories. However, most research only focuses on limited continual task settings where the tasks have the same observation and action space, which deviates from the realistic demands of training agents in various environments. In view of this, we propose Vector-Quantized Continual Diffuser, named VQ-CD, to break the barrier of different spaces between various tasks. Specifically, our method contains two complementary sections, where the quantization spaces alignment provides a unified basis for the selective weights activation. In the quantized spaces alignment, we leverage vector quantization to align the different state and action spaces of various tasks, facilitating continual training in the same space. Then, we propose to leverage a unified diffusion model attached by the inverse dynamic model to master all tasks by selectively activating different weights according to the task-related sparse masks. Finally, we conduct extensive experiments on 15 continual learning (CL) tasks, including conventional CL task settings (identical state and action spaces) and general CL task settings (various state and action spaces). Compared with 16 baselines, our method reaches the SOTA performance.", "keywords": [ "continual offline reinforcement learning" ], "rating_list": [ 3, 8, 3, 5 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 1, 2, 1, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "HqlX3lPtbh", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Dongchen Han", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Jindong Gu", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Xiaochun Cao", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Xiaojun Jia", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Yang Bai", "gender": "Female", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Yang Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "YIHAO HUANG", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 51, "n_ref": 71, "n_ref_all": 98, "n_fig": 5, "n_tab": 14, "L_tab": 2379, "n_element_tab": 204, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3222, "n_element_tab_1": 111, "formula_len_all": 660, "formula_len_all_1": 590, "len_all": 157079, "len_all_1": 57246, "len_abs": 1887, "len_title": 96, "len_sents": 43641, "len_sents_1": 25133, "n_sents": 343, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1624, "title": "OT-Attack: Enhancing Adversarial Transferability of Vision-Language Models via Optimal Transport Optimization", "abs": "Vision-language pre-training (VLP) models demonstrate impressive abilities in processing both images and text.\nHowever, they are vulnerable to multi-modal adversarial examples (AEs). Investigating the generation of high-transferability adversarial examples is crucial for uncovering VLP models\u2019 vulnerabilities in practical scenarios. Recent works have indicated that leveraging data augmentation and image-text modal interactions can enhance the transferability of adversarial examples for VLP models significantly. However, they do not consider the optimal alignment problem between dataaugmented image-text pairs. This oversight leads to adversarial examples that are overly tailored to the source model, thus limiting improvements in transferability. In our research, we first explore the interplay between image sets produced through data augmentation and their corresponding text sets. We find that augmented image samples can align optimally with certain texts while exhibiting less relevance to others. Motivated by this, we propose an Optimal Transport-based Adversarial Attack, dubbed OT-Attack. The proposed method formulates the features of image and text sets as two distinct distributions and employs optimal transport theory to determine the most efficient mapping between them. This optimal mapping informs our generation of adversarial examples to effectively counteract the overfitting issues. Extensive experiments across various network architectures and datasets in image-text matching tasks reveal that our OT-Attack outperforms existing stateof-the-art methods in terms of adversarial transferability.", "keywords": [ "adversarial transferability", "VLP models", "optimal transport" ], "rating_list": [ 6, 5, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "HqjRlT65WX", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Chao Qu", "gender": "Male", "institution": "Inftech", "country": "CN", "position": "Researcher" }, { "name": "Hai-Tao Zheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Haojing Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Jiayi Kuang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" }, { "name": "Philip S. Yu", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "US", "position": "Full Professor" }, { "name": "Shuyu Guo", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xiaoyu Tan", "gender": "Male", "institution": "INFLY TECH (Shanghai) Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yangning Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ying Shen", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "Associate Professor" }, { "name": "Yinghui Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 56, "n_ref_all": 72, "n_fig": 4, "n_tab": 7, "L_tab": 1159, "n_element_tab": 198, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 938, "n_element_tab_1": 152, "formula_len_all": 611, "formula_len_all_1": 250, "len_all": 152971, "len_all_1": 61040, "len_abs": 1410, "len_title": 123, "len_sents": 39382, "len_sents_1": 29112, "n_sents": 254, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1260, "title": "Refine Knowledge of Large Language Models via Adaptive Contrastive Learning", "abs": "How to alleviate the hallucinations of Large Language Models (LLMs) has always been the fundamental goal pursued by the LLMs research community. Looking through numerous hallucination-related studies, a mainstream category of methods is to reduce hallucinations by optimizing the knowledge representation of LLMs to change their output. Considering that the core focus of these works is the knowledge acquired by models, and knowledge has long been a central theme in human societal progress, we believe that the process of models refining knowledge can greatly benefit from the way humans learn. In our work, by imitating the human learning process, we design an Adaptive Contrastive Learning strategy. Our method flexibly constructs different positive and negative samples for contrastive learning based on LLMs' actual mastery of knowledge. This strategy helps LLMs consolidate the correct knowledge they already possess, deepen their understanding of the correct knowledge they have encountered but not fully grasped, forget the incorrect knowledge they previously learned, and honestly acknowledge the knowledge they lack. Extensive experiments and detailed analyses on widely used datasets demonstrate the effectiveness and competitiveness of our method.", "keywords": [ "Large Language Models", "Model Hallucination", "Model Alignment" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "HqLHY4TzGj", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Aritra Bhowmik", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Martin R. Oswald", "gender": "unknown", "institution": "University of Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Pascal Mettes", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Cees Snoek", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 46, "n_ref": 88, "n_ref_all": 114, "n_fig": 6, "n_tab": 10, "L_tab": 2179, "n_element_tab": 232, "n_fig_1": 6, "n_tab_1": 8, "L_tab_1": 1800, "n_element_tab_1": 164, "formula_len_all": 649, "formula_len_all_1": 567, "len_all": 139427, "len_all_1": 64327, "len_abs": 1545, "len_title": 114, "len_sents": 39223, "len_sents_1": 28745, "n_sents": 299, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 66, "L_abs": 1361, "title": "Union-over-Intersections: Object Detection beyond Winner-Takes-All", "abs": "This paper revisits the problem of predicting box locations in object detection architectures. Typically, each box proposal or box query aims to directly maximize the intersection-over-union score with the ground truth, followed by a winner-takes-all non-maximum suppression where only the highest scoring box in each region is retained. We observe that both steps are sub-optimal: the first involves regressing proposals to the entire ground truth, which is a difficult task even with large receptive fields, and the second neglects valuable information from boxes other than the top candidate. Instead of regressing proposals to the whole ground truth, we propose a simpler approach\u2014regress only to the area of intersection between the proposal and the ground truth. This avoids the need for proposals to extrapolate beyond their visual scope, improving localization accuracy. Rather than adopting a winner-takes-all strategy, we take the union over the regressed intersections of all boxes in a region to generate the final box outputs. Our plug-and-play method integrates seamlessly into proposal-based, grid-based, and query-based detection architectures with minimal modifications, consistently improving object localization and instance segmentation. We demonstrate its broad applicability and versatility across various detection and segmentation tasks.", "keywords": [ "localization based feature representation", "intersection over union", "object detection." ], "rating_list": [ 8, 5, 8, 8 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "Hpu3KIX8Am", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Nithin Gopalakrishnan Nair", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Vishal M Patel", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 14, "n_ref_uni": 32, "n_ref": 69, "n_ref_all": 89, "n_fig": 5, "n_tab": 6, "L_tab": 1889, "n_element_tab": 78, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 954, "n_element_tab_1": 116, "formula_len_all": 1641, "formula_len_all_1": 783, "len_all": 118661, "len_all_1": 53096, "len_abs": 687, "len_title": 122, "len_sents": 28940, "len_sents_1": 24421, "n_sents": 221, "n_sents_1": 168, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1316, "title": "Dreamguider: Improved Training free Diffusion-based Conditional Generation", "abs": "Diffusion models have emerged as a formidable tool for training-free conditional generation. However, a key hurdle in inference-time guidance techniques is the need for compute-heavy backpropagation through the diffusion network for estimating the guidance direction. Moreover, these techniques often require handcrafted parameter tuning on a case-by-case basis. Although some recent works have introduced minimal compute methods for linear inverse problems, a generic lightweight guidance solution to both linear and non-linear guidance problems is still missing. To this end, we propose Dreamguider, a method that enables inference-time guidance without compute-heavy backpropagation through the diffusion network. The key idea is to regulate the gradient flow through a time-varying factor. Moreover, we propose an empirical guidance scale that works for a wide variety of tasks, hence removing the need for handcrafted parameter tuning. We further introduce an effective lightweight augmentation strategy that significantly boosts the performance during inference-time guidance. We present experiments using Dreamguider on multiple tasks across multiple datasets and models to show the effectiveness of the proposed modules. To facilitate further research, we will make the code public after the review process.", "keywords": [ "Diffusion", "Automatic Parameter estimation", "Zero shot generation" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 2, 4, 4, 5 ] }, { "paper_id": "HpUs2EXjOl", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Hiroki Furuta", "gender": "Male", "institution": "Google DeepMind", "country": "JP", "position": "Researcher" }, { "name": "Yusuke Iwasawa", "gender": "Male", "institution": "The University of Tokyo, The University of Tokyo", "country": "JP", "position": "Associate Professor" }, { "name": "Yutaka Matsuo", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" }, { "name": "Minegishi Gouki", "gender": "Male", "institution": "University of Tokyo", "country": "JP", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 6, "n_ref_uni": 25, "n_ref": 102, "n_ref_all": 143, "n_fig": 11, "n_tab": 10, "L_tab": 2738, "n_element_tab": 276, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1576, "n_element_tab_1": 130, "formula_len_all": 654, "formula_len_all_1": 507, "len_all": 174640, "len_all_1": 68742, "len_abs": 1508, "len_title": 141, "len_sents": 57835, "len_sents_1": 32552, "n_sents": 397, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1353, "title": "Rethinking Evaluation of Sparse Autoencoders through the Representation of Polysemous Words", "abs": "Sparse autoencoders (SAEs) have gained a lot of attention as a promising tool to improve the interpretability of large language models (LLMs) by mapping the complex superposition of *polysemantic* neurons into *monosemantic* features and composing a sparse dictionary of words.\n\nHowever, traditional performance metrics like Mean Squared Error and $\\mathrm{L}_{0}$ sparsity ignore the evaluation of the semantic representational power of SAEs - whether they can acquire interpretable monosemantic features while preserving the semantic relationship of words.For instance, it is not obvious whether a learned sparse feature could distinguish different meanings in one word.\n\nIn this paper, we propose a suite of evaluations for SAEs to analyze the quality of monosemantic features by focusing on polysemous words.\nOur findings reveal that SAEs developed to improve the MSE-$\\mathrm{L}_0$ Pareto frontier may confuse interpretability, which does not necessarily enhance the extraction of monosemantic features.\nThe analysis of SAEs with polysemous words can also figure out the internal mechanism of LLMs; deeper layers and the Attention module contribute to distinguishing polysemy in a word.\n\nOur semantics-focused evaluation offers new insights into the polysemy and the existing SAE objective and contributes to the development of more practical SAEs.", "keywords": [ "Sparse Autoencoder", "Mechanistic Interpretability", "Polysemantic Representation." ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "Hp6f6VKAeP", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chaeyoung Jung", "gender": "Female", "institution": "KAIST", "country": "KR", "position": "MS student" }, { "name": "Hojoon Ki", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Ji-Hoon Kim", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "PhD student" }, { "name": "Joon Son Chung", "gender": "unknown", "institution": "KAIST", "country": "KR", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 9, "n_ref_uni": 38, "n_ref": 84, "n_ref_all": 111, "n_fig": 13, "n_tab": 7, "L_tab": 1940, "n_element_tab": 173, "n_fig_1": 10, "n_tab_1": 5, "L_tab_1": 1781, "n_element_tab_1": 149, "formula_len_all": 3254, "formula_len_all_1": 1238, "len_all": 154883, "len_all_1": 65551, "len_abs": 1302, "len_title": 132, "len_sents": 45655, "len_sents_1": 28666, "n_sents": 351, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1364, "title": "InfiniteAudio: Infinite-Length Audio Generation with Consistent Acoustic Attributes", "abs": "This work aims to generate long-duration audio while preserving acoustic coherence, utilizing existing text-conditional audio generation models through diffusion-based approaches. Current diffusion models, however, encounter significant challenges in generating long audio sequences due to memory constraints, as output size scales with input length. While one possible solution is to concatenate short clips, this often leads to inconsistencies due to a lack of shared temporal information across segments.\nTo address these challenges, we propose InfiniteAudio, a novel inference technique designed to generate long audio with consistent acoustic attributes. Our method is based on three key components. First, we implement a curved denoising approach with a fixed-size input, enabling theoretically infinite audio generation while maintaining a constant memory footprint. Second, we introduce conditional guidance alternation, a mechanism that enhances intelligibility in long speech generation. Finally, initial self-attention features are shared across future frames to maintain temporal coherence.\nThe effectiveness of InfiniteAudio is demonstrated through comprehensive comparisons with existing text-to-audio generation baselines. Generated audio samples are available on our anonymous project page\\footnote{https://anonymousforcf.github.io/InfiniteAudio/}.", "keywords": [ "Text-to-audio generation", "diffusion models", "lengthy generation", "consistent generation" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 1, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "HozsY9Gdcl", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Ari Holtzman", "gender": "Male", "institution": ", University of Chicago", "country": "", "position": "Assistant Professor" }, { "name": "Hannaneh Hajishirzi", "gender": "Female", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "senior director" }, { "name": "Pradeep Dasigi", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Research Scientist" }, { "name": "Sewon Min", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "Assistant Professor" }, { "name": "Xinxi Lyu", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Yanai Elazar", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Postdoc" }, { "name": "Fatemehsadat Mireshghallah", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 57, "n_ref_all": 75, "n_fig": 6, "n_tab": 4, "L_tab": 1060, "n_element_tab": 113, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 411, "n_element_tab_1": 19, "formula_len_all": 341, "formula_len_all_1": 305, "len_all": 119390, "len_all_1": 59903, "len_abs": 1398, "len_title": 118, "len_sents": 34672, "len_sents_1": 29796, "n_sents": 239, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1401, "title": "Leveraging Set Assumption for Membership Inference in Language Models", "abs": "Membership Inference (MI) refers to the task of determining whether or not a document is included in the training data of a given model. MI provides an effective post-training alternative for analyzing training datasets when the access to them is restricted, including studying the impact of data choices on downstream performance, detecting copyrighted content in the training sets, and checking for evaluation set contamination. However, black-boxed Language Models (LMs) only providing the loss for the document may not provide a reliable signal for determining memberships. In this work, we leverage the insight that documents sharing certain attributes (e.g., time of creation) are all expected to be in a training set or none of them is, and develop methods that aggregate membership predictions over these documents. We apply our set assumption on five different domains (e.g., Wikipedia, Arxiv), and find that our method enhances prior MI methods by 0.14 in AUROC on average. We further analyze the impact of different language model sizes, training data deduplication, and methods of aggregating membership predictions over sets and find that our method is more effective on undeduplicated and larger models with more documents available in each set and longer sequence sampled for each document, and show our method\u2019s robustness against noises in the set assumption under practical settings.", "keywords": [ "Large language models", "membership Inference", "pretraining data" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "HoyKFRhwMS", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 8, "author_info_list": [ { "name": "Abhijit Ogale", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Austin Stone", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Engineer" }, { "name": "George Toderici", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Jonathon Shlens", "gender": "unknown", "institution": "Google", "country": "US", "position": "Principal Researcher" }, { "name": "Priyank Jaini", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Robert Geirhos", "gender": "Male", "institution": "Google DeepMind", "country": "CA", "position": "Research Scientist" }, { "name": "Sourabh Medapati", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Researcher" }, { "name": "Xi Yi", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 66, "n_ref": 100, "n_ref_all": 128, "n_fig": 13, "n_tab": 8, "L_tab": 1354, "n_element_tab": 114, "n_fig_1": 11, "n_tab_1": 5, "L_tab_1": 2776, "n_element_tab_1": 134, "formula_len_all": 284, "formula_len_all_1": 132, "len_all": 174988, "len_all_1": 66378, "len_abs": 1196, "len_title": 0, "len_sents": 49518, "len_sents_1": 32062, "n_sents": 336, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1204, "title": "Towards flexible perception with visual memory", "abs": "Training a neural network is a monolithic endeavor, akin to carving knowledge into stone: once the process is completed, editing the knowledge in a network is nearly impossible, since all information is distributed across the network's weights. We here explore a simple, compelling alternative by marrying the representational power of deep neural networks with the flexibility of a database. Decomposing the task of image classification into image similarity (from a pre-trained embedding) and search (via fast nearest neighbor retrieval from a knowledge database), we build a simple and flexible visual memory that has the following key capabilities:\n(1.) The ability to flexibly add data across scales: from individual samples all the way to entire classes and billion-scale data;\n(2.) The ability to remove data through unlearning and memory pruning;\n(3.) An interpretable decision-mechanism on which we can intervene to control its behavior.\nTaken together, these capabilities comprehensively demonstrate the benefits of an explicit visual memory. We hope that it might contribute to a conversation on how knowledge should be represented in deep vision models---beyond carving it in \"stone\" weights.", "keywords": [ "deep learning", "computer vision", "retrieval", "memory" ], "rating_list": [ 6, 3, 6, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "HoQbynIkh2", "primary_area": "learning theory", "n_author": 8, "author_info_list": [ { "name": "Jianyu Zhang", "gender": "Female", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Junjie Yu", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Quanying Liu", "gender": "Female", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Wenxiao Ma", "gender": "Female", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Xinyu Mou", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "", "position": "Undergrad student" }, { "name": "Yi Guo", "gender": "unknown", "institution": "Southern University of Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Yue Joseph Wang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "", "position": "Undergrad student" }, { "name": "Zihan Deng", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 35, "n_formula_1": 13, "n_ref_uni": 33, "n_ref": 53, "n_ref_all": 60, "n_fig": 4, "n_tab": 2, "L_tab": 450, "n_element_tab": 72, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 434, "n_element_tab_1": 70, "formula_len_all": 2198, "formula_len_all_1": 1003, "len_all": 105840, "len_all_1": 66973, "len_abs": 1401, "len_title": 108, "len_sents": 30243, "len_sents_1": 24984, "n_sents": 220, "n_sents_1": 164, "venue": "ICLR2025", "decision": "withdraw", "L_title": 108, "L_abs": 1407, "title": "Dynamic Compression Strategies for Uniform Low-Dimensional Representations in Human Brain and Neural Network", "abs": "Recent studies suggest that the generalization performance of neural networks is strongly linked to their ability to learn low-dimensional data representations. However, limited attention has been given to the consistency of compression across different types of input data. In this work, we compute the intrinsic dimensions of raw data and their corresponding representations to quantify the extent of information compression in neural networks. Our results indicate that the pre-trained model CLIP compresses complex datasets significantly more than simpler ones and tends to represent diverse datasets with uniform low-dimensional manifolds. Similarly, we observe stable dimensionality in neural manifolds in the brain across various tasks and cognitive processes, suggesting that biological systems also favor consistent low-dimensional representations. Theoretically, we demonstrate that lower-dimensional manifolds increase the probability of interpolation, facilitating the representation of new samples as convex combinations of existing data. Additionally, we derive an upper bound on generalization error within the interpolation regime, which tightens as the dimensionality of the data decreases. These findings underscore the critical role of uniform low-dimensional manifolds in supporting efficient and generalizable information representation in both artificial and biological neural systems.", "keywords": [ "Low-Dimensional Representations", "Generalization", "Neural Manifold" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "HnpDHiItd2", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Beichen Zhang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Junqi Jing", "gender": "Male", "institution": "Harbin Institute Technology, Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Lingshuai Lin", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Shuhui Wang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Tingting Chai", "gender": "Female", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Weigang Zhang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yufan Zhou", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Zhaobo Qi", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Lecturer" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 32, "n_ref": 81, "n_ref_all": 117, "n_fig": 14, "n_tab": 14, "L_tab": 3347, "n_element_tab": 488, "n_fig_1": 8, "n_tab_1": 7, "L_tab_1": 3781, "n_element_tab_1": 323, "formula_len_all": 761, "formula_len_all_1": 695, "len_all": 159881, "len_all_1": 63238, "len_abs": 1597, "len_title": 134, "len_sents": 52317, "len_sents_1": 25919, "n_sents": 394, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1381, "title": "Masked Temporal Interpolation Diffusion for Procedure Planning in Instructional Videos", "abs": "In this paper, we study the problem of procedure planning in instructional videos, which involves making goal-directed plans based on current visual observations in unstructured, real-life videos. Prior research leverages different forms of supervision to bridge the gap between observed states and unobserved actions. Building on this foundation, we propose an innovative approach by introducing a latent space temporal logical interpolation module within the diffusion model framework. This module enables the intermediate supervision of temporal logical relationships that were previously nonexistent. In terms of details, we employ an interpolator to guide the intermediate process within the diffusion model, using the start and end observation features as inputs. This involves extracting latent features through an encoder and applying an interpolation strategy with transformer encoder blocks to derive the latent features. Furthermore, to ensure the accuracy of actions in the outputs, we implement a masking strategy to constrain the scope of predictions and a task-adaptive masked proximity loss for the training process. Results across these three datasets of varying scales demonstrate that our MTID model achieves state-of-the-art performance on the overwhelming majority of key metrics. The code is available at https://anonymous.4open.science/r/MTID-E2E3/README.md.", "keywords": [ "procedure planning", "diffusion", "U-Net", "temporal logic interpolation", "action prediction", "mask" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 4, 5 ] }, { "paper_id": "HnhNRrLPwm", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Chenhang Cui", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Summer Camp Student" }, { "name": "Huaxiu Yao", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Lijuan Wang", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Linjie Li", "gender": "Female", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Mingyu Ding", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Peng Xia", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Shi Qiu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Siwei Han", "gender": "Female", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Intern" }, { "name": "Wenhao Zheng", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Yiyang Zhou", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Zhaorun Chen", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Zhaoyang Wang", "gender": "unknown", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 32, "n_ref": 87, "n_ref_all": 126, "n_fig": 13, "n_tab": 18, "L_tab": 6008, "n_element_tab": 757, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 2315, "n_element_tab_1": 206, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 217879, "len_all_1": 59529, "len_abs": 1611, "len_title": 141, "len_sents": 69432, "len_sents_1": 27099, "n_sents": 563, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 93, "L_abs": 1619, "title": "MMIE: Massive Multimodal Interleaved Comprehension Benchmark for Large Vision-Language Models", "abs": "Interleaved multimodal comprehension and generation, enabling models to produce and interpret both images and text in arbitrary sequences, have become a pivotal area in multimodal learning. Despite significant advancements, the evaluation of this capability remains insufficient. Existing benchmarks suffer from limitations in data scale, scope, and evaluation depth, while current evaluation metrics are often costly or biased, lacking in reliability for practical applications. To address these challenges, we introduce MMIE, a large-scale knowledge-intensive benchmark for evaluating interleaved multimodal comprehension and generation in Large Vision-Language Models (LVLMs). MMIE comprises 20K meticulously curated multimodal queries, spanning 3 categories, 12 fields, and 102 subfields, including mathematics, coding, physics, literature, health, and arts. It supports both interleaved inputs and outputs, offering a mix of multiple-choice and open-ended question formats to evaluate diverse competencies. Moreover, we propose a reliable automated evaluation metric, leveraging a scoring model fine-tuned with human-annotated data and systematic evaluation criteria, aimed at reducing bias and improving evaluation accuracy. Extensive experiments demonstrate the effectiveness of our benchmark and metrics in providing a comprehensive evaluation of interleaved LVLMs. Specifically, we evaluate eight LVLMs, revealing that even the best models show significant room for improvement, with most achieving only moderate results. We believe MMIE will drive further advancements in the development of interleaved LVLMs.", "keywords": [ "large vision-language model", "interleaved text-and-image evaluation" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "HmwneoGoy9", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "DaYou Du", "gender": "unknown", "institution": "HKUST(GZ)", "country": "CN", "position": "MS student" }, { "name": "Fan Yang", "gender": "Male", "institution": "Research, Microsoft", "country": "", "position": "Senior Principal Researcher" }, { "name": "Hayden Kwok-Hay So", "gender": "unknown", "institution": "University of Hong Kong", "country": "", "position": "Associate Professor" }, { "name": "Mao Yang", "gender": "unknown", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Shijie Cao", "gender": "Male", "institution": "Microsoft Research Asia", "country": "", "position": "Researcher" }, { "name": "Ting Cao", "gender": "unknown", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Yizhao Gao", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Intern" }, { "name": "Zhichen Zeng", "gender": "unknown", "institution": "University of Washington", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 27, "n_ref": 43, "n_ref_all": 64, "n_fig": 14, "n_tab": 6, "L_tab": 2312, "n_element_tab": 188, "n_fig_1": 14, "n_tab_1": 5, "L_tab_1": 1179, "n_element_tab_1": 124, "formula_len_all": 488, "formula_len_all_1": 488, "len_all": 97354, "len_all_1": 59874, "len_abs": 1637, "len_title": 111, "len_sents": 29604, "len_sents_1": 27021, "n_sents": 243, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1655, "title": "SeerAttention: Learning Intrinsic Sparse Attention in Your LLMs", "abs": "Attention is the cornerstone of modern Large Language Models (LLMs). Yet its quadratic complexity limits the efficiency and scalability of LLMs, especially for those with a long-context window. A promising approach addressing this limitation is to leverage the sparsity in attention. However, existing sparsity-based solutions predominantly rely on predefined patterns or heuristics to approximate sparsity. This practice falls short to fully capture the dynamic nature of attention sparsity in language-based tasks. \nThis paper argues that attention sparsity should be learned rather than predefined. To this end, we design SeerAttention, a new Attention mechanism that augments the conventional attention with a learnable gate that adaptively selects significant blocks in an attention map and deems the rest blocks sparse.\nSuch block-level sparsity effectively balances accuracy and speedup.\nTo enable efficient learning of the gating network, we develop a customized FlashAttention implementation that extracts the block-level ground truth of attention map with minimum overhead.\nSeerAttention not only applies to post-training, but also excels in long-context fine-tuning.\nOur results show that at post-training stages, SeerAttention significantly outperforms state-of-the-art static or heuristic-based sparse attention methods, while also being more versatile and flexible to adapt to varying context lengths and sparsity ratios.\nWhen applied to long-context fine-tuning with YaRN, SeerAttention can achieve a remarkable 90\\% sparsity ratio at a 32k context length with minimal perplexity loss, offering a $5.67\\times$ speedup over FlashAttention-2.", "keywords": [ "LLM", "Sparse Attention", "long-context LLM" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "HmmN0Mxze7", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Haoxuan Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Researcher" }, { "name": "Jun Wang", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Professor" }, { "name": "Mengyue Yang", "gender": "Female", "institution": "University of Bristol", "country": "GB", "position": "Assistant Professor" }, { "name": "Xiang Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Visiting Student" } ], "n_formula": 13, "n_formula_1": 8, "n_ref_uni": 36, "n_ref": 85, "n_ref_all": 111, "n_fig": 11, "n_tab": 7, "L_tab": 8537, "n_element_tab": 645, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 3385, "n_element_tab_1": 260, "formula_len_all": 975, "formula_len_all_1": 637, "len_all": 173666, "len_all_1": 65193, "len_abs": 1271, "len_title": 157, "len_sents": 45602, "len_sents_1": 26486, "n_sents": 322, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1226, "title": "Temporal Visiting-Monitoring Feature Interaction Learning for Modelling Structured Electronic Health Records", "abs": "Electronic health records (EHRs) contain patients\u2019 longitudinal visit records, and modelling EHRs can be applied to various clinical prediction tasks. Previous works primarily focus on visit sequences and perform feature interaction on visit-level data to capture patient states. Nonetheless, incorporating finer-grained monitoring sequences simultaneously in structured EHRs, where each visit involves multiple monitoring sessions, can improve prediction performance. However, these studies have not accounted for the relationships between visit-level and monitoring-level data. To fill this gap, we propose an EHRs modelling method aimed at modelling the dynamic interaction between visit-level and monitoring-level data and capturing finer-grained health trends. We first capture the dynamic influence between medical data, and then perform a visiting-monitoring feature interaction on the relationships between visit data and monitoring data, to obtain the representation of patients' state for clinical prediction. We conducted extensive experiments on disease prediction and drug recommendation tasks, with MIMIC-III and MIMIC-IV datasets, demonstrating that our method outperforms state-of-the-art models significantly.", "keywords": [ "Structured EHR", "Clinical Prediction", "Temporal Feature Interaction" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "Hm7RYDspQP", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 7, "author_info_list": [ { "name": "Harneet Singh", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Jeong hwan Yoon", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Karl Friston", "gender": "Male", "institution": "University College London", "country": "", "position": "Principal Researcher" }, { "name": "Katherine Hanson", "gender": "unknown", "institution": "ThoughtForge", "country": "US", "position": "Researcher" }, { "name": "Pedro Fontana", "gender": "Male", "institution": "ThoughtForge AI", "country": "US", "position": "Researcher" }, { "name": "Varun Kamat", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Whitney Sales", "gender": "Female", "institution": "ThoughtForge, Inc", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 20, "n_ref": 21, "n_ref_all": 37, "n_fig": 6, "n_tab": 9, "L_tab": 2480, "n_element_tab": 237, "n_fig_1": 6, "n_tab_1": 8, "L_tab_1": 1383, "n_element_tab_1": 108, "formula_len_all": 36, "formula_len_all_1": 36, "len_all": 72353, "len_all_1": 49746, "len_abs": 910, "len_title": 213, "len_sents": 24135, "len_sents_1": 23106, "n_sents": 158, "n_sents_1": 148, "venue": "ICLR2025", "decision": "withdraw", "L_title": 165, "L_abs": 913, "title": "Network-based Active Inference for Adaptive and Cost-efficient Real-World Applications: A Benchmark Study of a Valve-turning Task Against Deep Reinforcement Learning", "abs": "This paper introduces Network-based Active Inference (NetAIF), a novel approach that integrates Active Inference (AIF) principles with network dynamics to enable adaptive, cost-efficient real-world applications. In benchmark tests against Deep Reinforcement Learning (DRL), NetAIF outperforms DRL in both computational efficiency and task performance. Leveraging random attractor dynamics, NetAIF generates real-time trajectories, allowing robots to adapt to complex, dynamic environments without the need for extensive pre-training. We demonstrate NetAIF's superiority in industrial valve manipulation, achieving over 99\\% accuracy in goal position and orientation in untrained dynamic environments, with a 45,000-fold reduction in computational costs. NetAIF is approximately 100,000 times more efficient in iteration count than DRL, making it a highly robust and efficient solution for industrial applications.", "keywords": [ "Active Inference (AIF)", "Deep Reinforcement Learning (DRL)", "Free Energy Principle (FEP)", "Robotics", "Trajectory generation", "Random dynamical systems", "Random attractor dynamics", "Non-Equilibrium Steady State (NESS)", "Adaptive control", "Industrial automation", "Computational efficiency", "Cost-efficient solutions" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 1, 1, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "HlvruwLQth", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Lei Zhu", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou) & HKUST", "country": "GB", "position": "Assistant Professor" }, { "name": "Lihao Liu", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Sixiang Chen", "gender": "Male", "institution": "HKUST(GZ)", "country": "CN", "position": "PhD student" }, { "name": "Tian Ye", "gender": "Not Specified", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Xiaojie Xu", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "MS student" }, { "name": "Yijun Yang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Peter Guang", "gender": "Male", "institution": "Beijing Academy of Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "ZhaohuXing", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 22, "n_ref": 43, "n_ref_all": 58, "n_fig": 6, "n_tab": 6, "L_tab": 2053, "n_element_tab": 332, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 2148, "n_element_tab_1": 195, "formula_len_all": 668, "formula_len_all_1": 720, "len_all": 109591, "len_all_1": 59990, "len_abs": 1276, "len_title": 162, "len_sents": 26585, "len_sents_1": 24582, "n_sents": 228, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 113, "L_abs": 1284, "title": "Farther Than Mirror: Explore Pattern-Compensated Depth of Mirror with Temporal Changes for Video Mirror Detection", "abs": "Current video mirror detection models demonstrate satisfactory performance by analyzing different attributes of mirrors and incorporating temporal information. However, these models still struggle to detect mirrors in complex and dynamic scenarios. \nA simple yet critical visual cue is that objects reflected in a mirror appear to be farther away than the mirror itself. Motivated by this observation, we propose to explicitly analyze the Depth of Mirror (DOM) within a video to effectively localize mirrors - DOM refers to distinct perceived distances that make mirror regions appear farther away from their surroundings. Specifically, we devise a novel framework called FTM-Net, which contains two main contributions: a Pattern-Compensated DOM estimation strategy and a Dual-Granularity Affinity module. The Pattern-Compensated DOM estimation strategy uses multiple visual mirror patterns to refine the DOM, enhancing the accuracy of mirror localization in a single image. Furthermore, the Dual-Granularity Affinity module can effectively detect mirrors in video sequences by tracking and integrating DOM changes across frames. Experimental results on a benchmark dataset show that our model significantly outperforms 18 state-of-the-art methods in the video mirror detection task.", "keywords": [ "Video mirror detection", "depth estimation", "affinity", "visual pattern compensation" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "Hlm0cga0sv", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Cong Wei", "gender": "Not Specified", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Ge Zhang", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Weiming Ren", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Zheyang Xiong", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Researcher" }, { "name": "wenhu chen", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Assistant Professor" }, { "name": "Xinrun Du", "gender": "Male", "institution": "01.AI", "country": "CN", "position": "Intern" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 36, "n_ref": 77, "n_ref_all": 107, "n_fig": 11, "n_tab": 12, "L_tab": 5286, "n_element_tab": 972, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1297, "n_element_tab_1": 294, "formula_len_all": 1172, "formula_len_all_1": 772, "len_all": 173466, "len_all_1": 63828, "len_abs": 1649, "len_title": 129, "len_sents": 47590, "len_sents_1": 26784, "n_sents": 394, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1656, "title": "OmniEdit: Building Image Editing Generalist Models Through Specialist Supervision", "abs": "Instruction-guided image editing methods have demonstrated significant potential by training diffusion models on automatically synthesized or manually annotated image editing pairs. However, these methods remain far from practical, real-life applications. We identify three primary challenges contributing to this gap. Firstly, existing models have limited editing skills due to the biased synthesis process. Secondly, these methods are trained with datasets with a high volume of noise and artifacts. This is due to the application of simple filtering methods like CLIP-score. Thirdly, all these datasets are restricted to a single low resolution and fixed aspect ratio, limiting the versatility to handle real-world use cases.\nIn this paper, we present OmniEdit, which is an omnipotent editor to handle seven different image editing tasks with any aspect ratio seamlessly. Our contribution is in four folds: (1) OmniEdit is trained by utilizing the supervision from seven different specialist models to ensure task coverage. (2) we utilize importance sampling based on the scores provided by large multimodal models (like GPT-4o) instead of CLIP-score to improve the data quality. (3) we propose a new editing architecture called EditNet to greatly boost the editing success rate, (4) we provide images with different aspect ratios to ensure that our model can handle any image in the wild. We have curated a test set containing images of different aspect ratios, accompanied by diverse instructions to cover different tasks. Both automatic evaluation and human evaluations demonstrate that OmniEdit can significantly outperforms all the existing models.", "keywords": [ "Image Editing", "Diffusion Models" ], "rating_list": [ 6, 6, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 5, 3 ] }, { "paper_id": "HksnKo0iV9", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Kuo Yang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Qihe Huang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Wu Jiang", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Xu Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Researcher" }, { "name": "Yang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zhengyang Zhou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Zhongchao Yi", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 37, "n_formula_1": 16, "n_ref_uni": 11, "n_ref": 49, "n_ref_all": 92, "n_fig": 8, "n_tab": 17, "L_tab": 4580, "n_element_tab": 376, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2115, "n_element_tab_1": 98, "formula_len_all": 1611, "formula_len_all_1": 587, "len_all": 188111, "len_all_1": 57789, "len_abs": 3766, "len_title": 118, "len_sents": 65268, "len_sents_1": 24659, "n_sents": 470, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1387, "title": "TimeBase: The Power of Minimalism in Long-term Time Series Forecasting", "abs": "Long-term time series forecasting (LTSF) has traditionally relied on models with large parameters to capture extended temporal dependencies. However, time series data, unlike high-dimensional images or text, often exhibit strong periodicity and low-rank structures, especially in long forecasting horizons. This characteristic can lead many models focusing on redundant patterns, resulting in inefficient use of computational resources. In this paper, we introduce TimeBase, an ultra-lightweight network with fewer than 0.4$k$ parameters, designed to harness the power of minimalism in LTSF. TimeBase extracts core periodic features by leveraging full-rank typical period representations under orthogonality constraints, enabling accurate prediction of future cycles. Extensive experiments on real-world datasets demonstrate that TimeBase not only achieves minimalism in both model size and computational cost, reducing MACs by 35x and parameter counts by over 1000 times compared to standard linear models, but also wins state-of-the-art forecasting performance, ranking Top1-Top5 in all 28 prediction settings. Additionally, TimeBase can also serve as a very effective plug-and-play tool for patch-based forecasting methods, enabling extreme complexity reduction without compromising prediction accuracy. Code is available at \\url{https://anonymous.4open.science/r/TimeBase-fixbug}.", "keywords": [ "Long-term Time Series Forecasting" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "HkPz96fgOv", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 5, "author_info_list": [ { "name": "Deblina Sarkar", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Milin Tunsiricharoengul", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Nicolas Alder", "gender": "Not Specified", "institution": "Hasso Plattner Institute", "country": "DE", "position": "PhD student" }, { "name": "Ralf Herbrich", "gender": "Male", "institution": "Hasso Plattner Institute", "country": "DE", "position": "Full Professor" }, { "name": "Shivam Nitin Kajale", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 24, "n_ref_uni": 32, "n_ref": 47, "n_ref_all": 91, "n_fig": 16, "n_tab": 3, "L_tab": 784, "n_element_tab": 35, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 231, "n_element_tab_1": 9, "formula_len_all": 1318, "formula_len_all_1": 981, "len_all": 135981, "len_all_1": 69131, "len_abs": 3245, "len_title": 116, "len_sents": 50890, "len_sents_1": 34309, "n_sents": 368, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1223, "title": "Energy-Efficient Sampling Using Stochastic Magnetic Tunnel Junctions", "abs": "(Pseudo)random sampling, a costly yet widely used method in (probabilistic) machine learning and Markov Chain Monte Carlo algorithms, remains unfeasible on a truly large scale due to unmet computational requirements. We introduce an energy-efficient algorithm for uniform Float16 sampling, utilizing a room-temperature stochastic magnetic tunnel junction device to generate truly random floating-point numbers. By avoiding expensive symbolic computation and mapping physical phenomena directly to the statistical properties of the floating-point format and uniform distribution, our approach achieves a higher level of energy efficiency than the state-of-the-art Mersenne-Twister algorithm by a minimum factor of 9721 and an improvement factor of 5649 compared to the more energy-efficient PCG algorithm. Building on this sampling technique and hardware framework, we decompose arbitrary distributions into many non-overlapping approximative uniform distributions along with convolution and prior-likelihood operations, which allows us to sample from any 1D distribution without closed-form solutions. We provide measurements of the potential accumulated approximation errors, demonstrating the effectiveness of our method.", "keywords": [ "Sampling", "Random Number Generation", "Probabilistic Machine Learning", "Hardware", "Power Consumption", "Spintronics", "Sustainability" ], "rating_list": [ 8, 8, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "HkB4bW5eJj", "primary_area": "foundation or frontier models, including LLMs", "n_author": 13, "author_info_list": [ { "name": "Desislav Ivanov", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Eva Schnider", "gender": "Female", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Gabriela Botea", "gender": "Female", "institution": "Google", "country": "DE", "position": "Researcher" }, { "name": "Jing Tang", "gender": "Male", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Kay Lamerigts", "gender": "Male", "institution": "Google", "country": "NL", "position": "Researcher" }, { "name": "Marc Wilson", "gender": "unknown", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Mathias MJ Bellaiche", "gender": "Male", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Mayank Daswani", "gender": "unknown", "institution": "Research, Google", "country": "GB", "position": "Google" }, { "name": "Michael A. Sanchez", "gender": "unknown", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Mikhail Papkov", "gender": "unknown", "institution": "University of Tartu", "country": "", "position": "PhD student" }, { "name": "SHRAVYA SHETTY", "gender": "unknown", "institution": "Google", "country": "", "position": "Principal Researcher" }, { "name": "Umesh Telang", "gender": "Male", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Yojan Patel", "gender": "Male", "institution": "Google", "country": "GB", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 0, "n_ref_uni": 24, "n_ref": 52, "n_ref_all": 82, "n_fig": 9, "n_tab": 12, "L_tab": 13718, "n_element_tab": 280, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1795, "n_element_tab_1": 106, "formula_len_all": 1153, "formula_len_all_1": 0, "len_all": 168971, "len_all_1": 56384, "len_abs": 2281, "len_title": 89, "len_sents": 54161, "len_sents_1": 27992, "n_sents": 399, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1603, "title": "Plots unlock time-series understanding in multimodal models", "abs": "While multimodal foundation models can now natively work with data beyond text, they remain underutilized in analyzing the considerable amounts of multi-dimensional time-series data in fields like healthcare, finance, and social sciences, representing a missed opportunity for richer, data-driven insights. This paper proposes a simple but effective method that leverages the existing vision encoders of these models to \"see\" time-series data via plots, avoiding the need for additional, potentially costly, model training. Our empirical evaluations show that this approach outperforms providing the raw time-series data as text, with the additional benefit that visual time-series representations demonstrate up to a 90\\% reduction in model API costs. We validate our hypothesis through synthetic data tasks of increasing complexity, progressing from simple functional form identification on clean data, to extracting trends from noisy scatter plots. To demonstrate generalizability from synthetic tasks with clear reasoning steps to more complex, real-world scenarios, we apply our approach to consumer health tasks \u2013 specifically fall detection, activity recognition, and readiness assessment \u2013 which involve heterogeneous, noisy data and multi-step reasoning. The overall success in plot performance over text performance (up to an 120\\% performance increase on zero-shot synthetic tasks, and up to 150\\% performance increase on real-world tasks), across both GPT and Gemini model families, highlights our approach's potential for making the best use of the native capabilities of foundation models.", "keywords": [ "time-series", "vision encoder", "multimodal", "plot", "vision language model", "foundation model", "large language model", "LLM" ], "rating_list": [ 6, 1, 5, 5 ], "soundness_list": [ 3, 1, 2, 2 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 3, 1, 2, 4 ], "confidence_list": [ 4, 2, 3, 5 ] }, { "paper_id": "Hjp1V6zlZi", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Hsuan-Tien Lin", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Full Professor" }, { "name": "Hung-Chieh Fang", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Undergrad student" }, { "name": "Po-Yi Lu", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 39, "n_ref": 84, "n_ref_all": 112, "n_fig": 17, "n_tab": 8, "L_tab": 3527, "n_element_tab": 518, "n_fig_1": 16, "n_tab_1": 3, "L_tab_1": 2467, "n_element_tab_1": 275, "formula_len_all": 439, "formula_len_all_1": 491, "len_all": 131099, "len_all_1": 60182, "len_abs": 1398, "len_title": 115, "len_sents": 35684, "len_sents_1": 25759, "n_sents": 286, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1395, "title": "Reducing Bias in Feature Extractors for Extreme Universal Domain Adaptation", "abs": "Universal Domain Adaptation (UniDA) aims to transfer knowledge from a labeled source domain to an unlabeled target domain without prior knowledge of the label sets between the two domains. The goal of UniDA is to achieve robust performance under arbitrary label-set distributions. However, existing literature has not sufficiently explored performance across diverse distribution scenarios. Our experiments reveal that existing methods struggle when the source domain has significantly more non-overlapping classes than overlapping ones, a setting we refer to as *Extreme UniDA*. In this paper, we demonstrate that classical partial domain alignment, which focuses on aligning only overlapping-class data between domains, is limited in mitigating feature extractor bias in extreme UniDA scenarios. \nWe argue that feature extractors trained with source supervised loss disrupt the intrinsic structure of target data due to the inherent differences between source-private-class data and target data. To mitigate this bias, we employ self-supervised learning to preserve the structure of target data.\nThis method can be easily integrated into existing frameworks. We apply the proposed approach to two distinct training paradigms\u2014adversarial-based and optimal-transport-based\u2014and show consistent improvements across various class-set distributions, with significant gains in extreme UniDA settings.", "keywords": [ "Domain Adaptation", "Universal Domain Adaptation", "Machine Learning" ], "rating_list": [ 3, 5, 5, 6, 6 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 1, 3 ], "confidence_list": [ 4, 5, 3, 3, 2 ] }, { "paper_id": "HjoYVtSkT8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Aritra Bhowmik", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Dennis Koelma", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "software engineer" }, { "name": "Martin R. Oswald", "gender": "unknown", "institution": "University of Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Mohammad Mahdi Derakhshani", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Cees Snoek", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "Yuki Asano", "gender": "Male", "institution": "University of Technology Nuremberg", "country": "DE", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 27, "n_ref": 60, "n_ref_all": 92, "n_fig": 10, "n_tab": 12, "L_tab": 2998, "n_element_tab": 508, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1938, "n_element_tab_1": 311, "formula_len_all": 576, "formula_len_all_1": 609, "len_all": 143601, "len_all_1": 59878, "len_abs": 3559, "len_title": 90, "len_sents": 39580, "len_sents_1": 25504, "n_sents": 277, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 1312, "title": "Learning to Ground VLMs without Forgetting", "abs": "Spatial awareness is key to enable embodied multimodal AI systems. Yet, without vast amounts of spatial supervision, current Visual Language Models (VLMs) struggle at this task. In this paper, we introduce LynX, a framework that equips pretrained VLMs with visual grounding ability without forgetting their existing image and language understanding skills.\nTo this end, we propose a Dual Mixture of Experts module that modifies only the decoder layer of the language model, using one frozen Mixture of Experts (MoE) pre-trained on image and language understanding and another learnable MoE for new grounding capabilities. This allows the VLM to retain previously learned knowledge and skills, while acquiring what is missing.\nTo train the model effectively, we generate a high-quality synthetic dataset we call SCouT, which mimics human reasoning in visual grounding. This dataset provides rich supervision signals, describing a step-by-step multimodal reasoning process, thereby simplifying the task of visual grounding. We evaluate LynX on several object detection and visual grounding datasets, demonstrating strong performance in object detection, zero-shot localization and grounded reasoning while maintaining its original image and language understanding capabilities on seven standard benchmark datasets.", "keywords": [ "vision language models", "visual grounding", "mixture of experts", "synthetic datasets" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "Hjk1tWIdvL", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Mingbao Lin", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Wenhao Li", "gender": "unknown", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Zhanpeng Zeng", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 36, "n_ref": 76, "n_ref_all": 106, "n_fig": 9, "n_tab": 5, "L_tab": 1355, "n_element_tab": 119, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1183, "n_element_tab_1": 106, "formula_len_all": 811, "formula_len_all_1": 774, "len_all": 139385, "len_all_1": 63487, "len_abs": 1520, "len_title": 115, "len_sents": 36039, "len_sents_1": 28706, "n_sents": 273, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1564, "title": "Hierarchy-Aided Sparse Attention For Fast LLMs Prefilling Inference", "abs": "Pre-filling Large Language Models (LLMs) with long-context inputs is computationally expensive due to the quadratic complexity of full attention. While global attention is essential during decoding, its importance diminishes during pre-filling, where the focus is on contextualizing tokens rather than predicting the next one. Building on prior work, we apply diagonal block sparse attention during the pre-filling phase, reducing attention-related FLOPs by over 90\\% without significant degradation in language modeling performance. To address the remaining performance gap, we propose \\textbf{H}ierarchy-\\textbf{A}ided \\textbf{S}parse \\textbf{A}ttention (HASA), which incorporates a specialized transformer branch. This branch extracts global embeddings from each chunk and aligns local attention with full-attention, facilitating cross-chunk interaction. HASA stabilizes sparse attention computations, making the pre-filling phase highly efficient, particularly in long-sequence scenarios. While HASA significantly accelerates the pre-filling phase, we ensure robust language modeling performance by enabling interaction between global embeddings across chunks, which prevents the performance degradation typically observed in sparse attention mechanisms. Given that there are limited methods specifically accelerating pre-filling, our baselines include various open-source long-context models. Across multiple benchmarks, HASA not only maintains performance but also outperforms baseline models in certain scenarios. We will release the models upon acceptance.", "keywords": [ "Long-Context LLM; Pre-Filling Acceleration; Sparse Attention" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 4, 2, 2 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "Hj1D0Xq3Ef", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Eleonora Kreacic", "gender": "Female", "institution": "J.P. Morgan Chase", "country": "", "position": "Researcher" }, { "name": "Mohsen Ghassemi", "gender": "Male", "institution": "JPMorganChase", "country": "US", "position": "Researcher" }, { "name": "Mufei Li", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Pan Li", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Rongzhe Wei", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Vamsi K. Potluru", "gender": "Not Specified", "institution": "J.P. Morgan Chase", "country": "US", "position": "Researcher" }, { "name": "Xiang Yue", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Yifan Li", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "I Chien", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Postdoc" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 42, "n_ref": 102, "n_ref_all": 139, "n_fig": 13, "n_tab": 17, "L_tab": 11940, "n_element_tab": 1007, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 91, "n_element_tab_1": 10, "formula_len_all": 512, "formula_len_all_1": 322, "len_all": 198267, "len_all_1": 71342, "len_abs": 1956, "len_title": 137, "len_sents": 52071, "len_sents_1": 33888, "n_sents": 412, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1966, "title": "Underestimated Privacy Risks for Minority Populations in Large Language Model Unlearning", "abs": "Large Language Models (LLMs) are trained on extensive datasets that often contain sensitive, human-generated information, raising significant concerns about privacy breaches. While certified unlearning approaches offer strong privacy guarantees, they rely on restrictive model assumptions that are not applicable to LLMs. As a result, various unlearning heuristics have been proposed, with the associated privacy risks assessed only empirically. The standard evaluation pipelines typically randomly select data for removal from the training set, apply unlearning techniques, and use membership inference attacks (MIAs) to compare the unlearned models against models retrained without the to-be-unlearned data. However, since every data point is subject to the right to be forgotten, unlearning should be considered in the worst-case scenario from the privacy perspective. Prior work shows that data outliers may exhibit higher memorization effects. Intuitively, they are harder to be unlearn and thus the privacy risk of unlearning them is overlooked and underestimated in the current evaluation. In this paper, we leverage minority data to identify such a critical flaw in previously widely adopted evaluations. We substantiate this claim through carefully designed experiments, including unlearning canaries related to minority groups, inspired by privacy auditing literature. Using personally identifiable information (PII) as a representative minority identifier, we demonstrate that minority groups experience at least 20\\% more privacy leakage in most cases across six unlearning approaches, three MIAs, three benchmark datasets, and two LLMs of different scales. Given that the right to be forgotten should be upheld for every individual, we advocate for a more rigorous evaluation of LLM unlearning methods. Our minority-aware evaluation framework represents an initial step toward ensuring more equitable and thorough assessments of LLM unlearning efficacy.", "keywords": [ "Machine Unlearning", "Large Language Models", "Membership Inference Attack", "Privacy Risk", "Minority Groups" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 2, 2 ] }, { "paper_id": "Hhx3swAQAZ", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Cen Chen", "gender": "Female", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Hong Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Weining Qian", "gender": "unknown", "institution": "East China Normal University", "country": "", "position": "Full Professor" }, { "name": "Wenmeng Zhou", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yaliang Li", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Staff Engineer" }, { "name": "Yingda Chen", "gender": "unknown", "institution": "Alibaba Group", "country": "US", "position": "Researcher" }, { "name": "Yu Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhongjie Duan", "gender": "Male", "institution": "East China Normal University", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 42, "n_ref": 82, "n_ref_all": 96, "n_fig": 6, "n_tab": 4, "L_tab": 1671, "n_element_tab": 240, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 4269, "n_element_tab_1": 268, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 123717, "len_all_1": 60827, "len_abs": 1589, "len_title": 77, "len_sents": 30952, "len_sents_1": 26315, "n_sents": 226, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1602, "title": "ExVideo: Extending Video Diffusion Models via Parameter-Efficient Post-Tuning", "abs": "Recently, advancements in video synthesis have attracted significant attention. Video synthesis models such as AnimateDiff and Stable Video Diffusion have demonstrated the practical applicability of diffusion models in creating dynamic visual content. The emergence of SORA has further spotlighted the potential of video generation technologies. Despite advancements, the extension of video lengths remains constrained by computational resources. Most existing video synthesis models are limited to generating short video clips. In this paper, we propose a novel post-tuning methodology for video synthesis models, called ExVideo. This approach is designed to enhance the capability of current video synthesis models, allowing them to produce content over extended temporal durations while incurring lower training expenditures. In particular, we design extension strategies across common temporal model architectures respectively, including 3D convolution, temporal attention, and positional embedding. To evaluate the efficacy of our proposed post-tuning approach, we trained ExSVD, an extended model based on Stable Video Diffusion model. Our approach enhances the model's capacity to generate up to $5\\times$ its original number of frames, requiring only 1.5k GPU hours of training on a dataset comprising 40k videos. Importantly, the substantial increase in video length doesn't compromise the model's innate generalization capabilities, and the model showcases its advantages in generating videos of diverse styles and resolutions. We will release the source code and the enhanced model publicly.", "keywords": [ "generative models", "video synthesis", "post tuning" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 1, 1, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "HhefvT4ktU", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Nouar AlDahoul", "gender": "Female", "institution": "New York University, Abu Dhabi", "country": "AE", "position": "Postdoc" }, { "name": "Talal Rahwan", "gender": "Male", "institution": "New York University", "country": "AE", "position": "Associate Professor" }, { "name": "Yasir Zaki", "gender": "Male", "institution": "New York University Abu Dhabi", "country": "AE", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 24, "n_ref": 47, "n_ref_all": 84, "n_fig": 18, "n_tab": 12, "L_tab": 9250, "n_element_tab": 1179, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 1436, "n_element_tab_1": 264, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 154979, "len_all_1": 75331, "len_abs": 6048, "len_title": 122, "len_sents": 49401, "len_sents_1": 35863, "n_sents": 391, "n_sents_1": 285, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1109, "title": "AI-generated faces influence gender stereotypes and racial homogenization", "abs": "Text-to-image generative AI models such as Stable Diffusion are used daily by millions worldwide. However, the extent to which these models exhibit racial and gender stereotypes is not yet fully understood. Here, we document significant biases in Stable Diffusion across six races, two genders, 32 professions, and eight attributes. Additionally, we examine the degree to which Stable Diffusion depicts individuals of the same race as being similar to one another. This analysis reveals significant racial homogenization, e.g., depicting nearly all middle eastern men as dark-skinned, bearded, and wearing a traditional headdress. We then propose novel debiasing solutions that address the above stereotypes. Finally, using a preregistered experiment, we show that being presented with inclusive AI-generated faces reduces people's racial and gender biases, while being presented with non-inclusive ones increases such biases. This persists regardless of whether the images are labeled as AI-generated. Taken together, our findings emphasize the need to address biases and stereotypes in AI-generated content.", "keywords": [ "AI bias", "gender stereotypes", "racial homogenization", "stable diffusion", "inclusive AI-generated faces", "Text-to-image generative AI" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 1, 1, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "HhP9bgCugr", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Jian Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Jiawei Du", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Jiaxiang Liu", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Joey Tianyi Zhou", "gender": "Male", "institution": "A*STAR Centre for Frontier AI Research", "country": "SG", "position": "Principal Researcher" }, { "name": "Yuan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zuozhu Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 26, "n_formula_1": 11, "n_ref_uni": 27, "n_ref": 60, "n_ref_all": 84, "n_fig": 4, "n_tab": 8, "L_tab": 1199, "n_element_tab": 102, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 559, "n_element_tab_1": 68, "formula_len_all": 1551, "formula_len_all_1": 724, "len_all": 160357, "len_all_1": 55525, "len_abs": 1849, "len_title": 123, "len_sents": 43926, "len_sents_1": 27304, "n_sents": 295, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1870, "title": "Align-VL: Can Being Modest Help in the Alignment of Vision-Language Models?", "abs": "Multimodal alignment aims to learn a shared latent space between different modal inputs to establish connections across modalities.\nA prime example is Visual Language Models (VLMs), such as CLIP, which benefit from extensive image-text pre-training and excel in image recognition tasks. \nThese models are emblematic of successful multimodal alignment.\nSubsequent work has successfully aligned multimodal data on limited datasets using feature mixing enhancement methods.\nHowever, these models encounter significant challenges:\n{The presence of {ambiguous samples (either partially matched or completely unmatched)} in datasets with weakly associated, low-quality image-text pairs causes models to become overconfident (in training) and confused (in inference), ultimately reducing performance}. Current contrastive learning methods, which rely on single positive pairs, exacerbate this issue by encouraging overconfidence when the model encounters such ambiguous samples.\nTo overcome these challenges, we developed Align-VL, a multimodal alignment enhancement method that operates on the latent spaces of pre-trained unimodal encoders. This approach adjusts the matching degree of the data and moderates model overconfidence, promoting more appropriate and effective alignments.\nAlign-VL incorporates {Random Perturbation} and {Embedding Smoothing} strategies to enhance input feature robustness and reduce model overconfidence, improving the model's ability to manage uncertainty and generalize to new data. \nIn our experiments, Align-VL outperformed existing state-of-the-art (SoTA) methods in image-text retrieval tasks, demonstrating its superior effectiveness.\nAlign-VL also offers significant reductions in training time and data requirements compared to methods like CLIP, using substantially fewer GPU days and image-text pairs. \nCode will be publicly available.", "keywords": [ "Vision-Language Models", "Multimodal Alignment", "Embedding Smoothing" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "Hh6XKefS28", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Huajun Chen", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Lei Liang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "", "position": "MS student" }, { "name": "Mingyang Chen", "gender": "Male", "institution": "Baichuan Inc.", "country": "CN", "position": "Researcher" }, { "name": "Wen Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "Assistant Professor" }, { "name": "Yushan Zhu", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Zhiqiang Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 33, "n_ref": 84, "n_ref_all": 111, "n_fig": 6, "n_tab": 7, "L_tab": 6916, "n_element_tab": 494, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 10768, "n_element_tab_1": 614, "formula_len_all": 890, "formula_len_all_1": 1006, "len_all": 166171, "len_all_1": 76392, "len_abs": 5541, "len_title": 83, "len_sents": 42054, "len_sents_1": 27445, "n_sents": 245, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1545, "title": "Croppable Knowledge Graph Embedding", "abs": "Knowledge Graph Embedding (KGE) is a common method for Knowledge Graphs (KGs) to serve various artificial intelligence tasks. The suitable dimensions of the embeddings depend on the storage and computing conditions of the specific application scenarios. Once a new dimension is required, a new KGE model needs to be trained from scratch, which greatly increases the training cost and limits the efficiency and flexibility of KGE in serving various scenarios. In this work, we propose a novel KGE training framework MED, through which we could train once to get a croppable KGE model applicable to multiple scenarios with different dimensional requirements, sub-models of the required dimensions can be cropped out of it and used directly without any additional training. In MED, we propose a mutual learning mechanism to improve the low-dimensional sub-models performance and make the high-dimensional sub-models retain the capacity that low-dimensional sub-models have, an evolutionary improvement mechanism to promote the high-dimensional sub-models to master the knowledge that the low-dimensional sub-models can not learn, and a dynamic loss weight to balance the multiple losses adaptively. Experiments on 4 KGE models over 4 standard KG completion datasets, 3 real application scenarios over a real-world large-scale KG, and the experiments of extending MED to the language model BERT show the effectiveness, high efficiency, and flexible extensibility of MED. The code and data are available at https://anonymous.4open.science/r/MED-DBFC.", "keywords": [ "knowledge graph", "knowledge distillation", "parameter-efficient representation learning" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "HgSIfXTpBE", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Antonio Khalil Moretti", "gender": "Male", "institution": "Spelman College", "country": "US", "position": "Assistant Professor" }, { "name": "Itsik Pe'er", "gender": "Male", "institution": "Columbia University", "country": "", "position": "Full Professor" }, { "name": "Philippe Chlenski", "gender": "Male", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Quentin Chu", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "MS student" }, { "name": "Raiyan R. Khan", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "PhD student" } ], "n_formula": 67, "n_formula_1": 30, "n_ref_uni": 38, "n_ref": 62, "n_ref_all": 100, "n_fig": 5, "n_tab": 10, "L_tab": 4390, "n_element_tab": 324, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1525, "n_element_tab_1": 126, "formula_len_all": 3321, "formula_len_all_1": 1096, "len_all": 183355, "len_all_1": 54523, "len_abs": 1390, "len_title": 49, "len_sents": 49139, "len_sents_1": 21387, "n_sents": 419, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1397, "title": "Mixed-curvature decision trees and random forests", "abs": "Decision trees (DTs) and their random forest (RF) extensions are workhorses of classification and regression in Euclidean spaces. However, algorithms for learning in non-Euclidean spaces are still limited. We extend DT and RF algorithms to product manifolds: Cartesian products of several hyperbolic, hyperspherical, or Euclidean components. Such manifolds handle heterogeneous curvature while still factorizing neatly into simpler components, making them compelling embedding spaces for complex datasets. Our novel angular reformulation of DTs respects the geometry of the product manifold, yielding splits that are geodesically convex, maximum-margin, and composable. In the special cases of single-component manifolds, our method simplifies to its Euclidean or hyperbolic counterparts, or introduces hyperspherical DT algorithms, depending on the curvature. We benchmark our method on various classification, regression, and link prediction tasks on synthetic data, graph embeddings, mixed-curvature variational autoencoder latent spaces, and empirical data. Compared to six other classifiers, product DTs and RFs ranked first on 21 of 22 single-manifold benchmarks and 18 of 35 product manifold benchmarks, and placed in the top 2 on 53 of 57 benchmarks overall. This highlights the value of product DTs and RFs as straightforward yet powerful new tools for data analysis in product manifolds.", "keywords": [ "representation learning", "non-euclidean geometry", "decision trees", "random forests", "hyperbolic geometry", "hyperspherical geometry" ], "rating_list": [ 5, 5, 6, 5, 8 ], "soundness_list": [ 2, 3, 3, 2, 4 ], "presentation_list": [ 3, 4, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 4, 3 ] }, { "paper_id": "HgAS03GU4J", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Sameer Singh", "gender": "Male", "institution": "University of California, Irvine", "country": "", "position": "Associate Professor" }, { "name": "Yasaman Razeghi", "gender": "Female", "institution": "University of California, Irvine", "country": "", "position": "PhD student" }, { "name": "Yu Fei", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 0, "n_ref_uni": 23, "n_ref": 59, "n_ref_all": 123, "n_fig": 11, "n_tab": 38, "L_tab": 5148, "n_element_tab": 117, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 2575, "n_element_tab_1": 79, "formula_len_all": 477, "formula_len_all_1": 0, "len_all": 341407, "len_all_1": 60232, "len_abs": 1786, "len_title": 99, "len_sents": 135118, "len_sents_1": 28418, "n_sents": 1474, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1798, "title": "Inference-time Alignment of LLMs at the Token Level", "abs": "Large language models (LLMs) require alignment\u2014such as instruction-tuning or reinforcement learning from human feedback\u2014to effectively and safely follow user instructions. This process necessitates training aligned versions for every model size in each model family, resulting in significant computational overhead. In this work, we propose nudging, a simple, plug-and-play, and training-free algorithm that aligns any base model at inference time using a small aligned model. Nudging is motivated by recent findings that alignment primarily alters the model\u2019s behavior on a small subset of stylistic tokens, such as \"Sure\" or \"Thank\". We find that base models are significantly more uncertain when generating these tokens. Leveraging this observation, nudging employs a small aligned model to generate nudging tokens to steer the large base model's output toward desired directions when the base model's uncertainty is high. We evaluate the effectiveness of nudging across 3 model families and 13 tasks, covering reasoning, general knowledge, instruction following, and safety benchmarks. Without any additional training, nudging a large base model with a 7\u00d7 - 14\u00d7 smaller aligned model achieves zero-shot performance comparable to, and sometimes surpassing, that of large aligned models. For example, nudging OLMo-7b with OLMo-1b-instruct\u2014affecting less than 9% of tokens\u2014achieves a 10% absolute improvement on GSM8K over OLMo-7b-instruct. Unlike prior inference-time tuning methods, nudging enables off-the-shelf collaboration between model families. For instance, nudging Gemma-2-27b with Llama-2-7b-chat outperforms Llama-2-70b-chat on various tasks. Overall, this work introduces a simple yet powerful approach to token-level model collaboration, offering a modular solution to LLM alignment.", "keywords": [ "large language models", "alignment", "natural language reasoning", "inference-time algorithm" ], "rating_list": [ 5, 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 5, 4, 4, 3, 4 ] }, { "paper_id": "Hfv4LoCQPo", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Dong Yan", "gender": "Male", "institution": "Baichuan Intelligent Technology", "country": "", "position": "Researcher" }, { "name": "Jian Xie", "gender": "Male", "institution": "Beijing Baichuan Intelligent Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Siyuan Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xiangsheng Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Yichi Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yinpeng Dong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Hang Su", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 3, "n_ref_uni": 63, "n_ref": 130, "n_ref_all": 149, "n_fig": 7, "n_tab": 9, "L_tab": 5421, "n_element_tab": 489, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 3726, "n_element_tab_1": 386, "formula_len_all": 3672, "formula_len_all_1": 227, "len_all": 232321, "len_all_1": 67178, "len_abs": 2692, "len_title": 131, "len_sents": 53153, "len_sents_1": 28390, "n_sents": 399, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1413, "title": "Towards Mitigating Factual Hallucination in LLMs through Self-Alignment with Memory", "abs": "Despite the impressive performance of Large Language Models (LLMs) across numerous tasks and widespread application in real-world scenarios, LLMs still struggle to guarantee their responses to be accurate and aligned with objective facts. This leads to factual hallucination of LLMs, which can be difficult to detect and mislead users lacking relevant knowledge. Post-training techniques have been employed to mitigate this issue, yet they are usually followed by a trade-off between honesty and helpfulness, along with a lack of generalized improvements. In this paper, we propose to address it by augmenting LLM's fundamental capacity of leveraging its internal memory, that is, the knowledge derived from pre-training data. We introduce FactualBench, a comprehensive and precise factual QA dataset consisting of nearly 200k Chinese generative QA data spanning 21 domains for both evaluation and training purposes. Furthermore, we propose self-alignment with memory, i.e., fine-tuning the model via preference learning on self-generated pairwise data from FactualBench. Extensive experiments show that our method significantly enhances LLM's performance on FactualBench, with consistent improvements across various benchmarks concerning factuality, helpfulness and multiple skills. Additionally, different post-training techniques and tuning data sources are discussed to further understand their effectiveness.", "keywords": [ "Hallucination", "Large Language Model", "Dataset" ], "rating_list": [ 3, 8, 5, 1 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 3, 4, 2, 1 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "HfWcFs7XLR", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Alice Shoshana Jakobovits", "gender": "Female", "institution": "Research, Google", "country": "", "position": "Researcher" }, { "name": "Elizabeth Clark", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Fantine Huot", "gender": "Female", "institution": "Google", "country": "NL", "position": "Researcher" }, { "name": "Jennimaria Palomaki", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Mirella Lapata", "gender": "Female", "institution": "Edinburgh University, University of Edinburgh", "country": "GB", "position": "Full Professor" }, { "name": "Reinald Kim Amplayo", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 59, "n_ref": 106, "n_ref_all": 119, "n_fig": 5, "n_tab": 4, "L_tab": 699, "n_element_tab": 57, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 706, "n_element_tab_1": 53, "formula_len_all": 179, "formula_len_all_1": 258, "len_all": 263081, "len_all_1": 71053, "len_abs": 1016, "len_title": 115, "len_sents": 78209, "len_sents_1": 35001, "n_sents": 783, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1016, "title": "Agents' Room: Narrative Generation through Multi-step Collaboration", "abs": "Writing compelling fiction is a multifaceted process combining elements such as crafting a plot, developing interesting characters, and using evocative language. While large language models (LLMs) show promise for story writing, they currently rely heavily on intricate prompting, which limits their use. We propose Agents' Room, a generation framework inspired by narrative theory, that decomposes narrative writing into subtasks tackled by specialized agents. To illustrate our method, we introduce Tell Me A Story, a high-quality dataset of complex writing prompts and human-written stories, and a novel evaluation framework designed specifically for assessing long narratives. We show that Agents' Room generates stories that are preferred by expert evaluators over those produced by baseline systems by leveraging collaboration and specialization to decompose the complex story writing task into tractable components. We provide extensive analysis with automated and human-based metrics of the generated output.", "keywords": [ "fiction", "creative writing", "long-form generation", "LLMs", "agent", "collaboration", "multi-agent", "dataset" ], "rating_list": [ 8, 5, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 4, 3, 4 ], "contribution_list": [ 4, 3, 3 ], "confidence_list": [ 5, 5, 3 ] }, { "paper_id": "HfJxXbXlYJ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Aoqi Wu", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "MS student" }, { "name": "Chong Luo", "gender": "Female", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Dongdong Chen", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Liang Hu", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Full Professor" }, { "name": "Lili Qiu", "gender": "unknown", "institution": "Microsoft", "country": "CN", "position": "Principal Researcher" }, { "name": "Qi Dai", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Researcher" }, { "name": "Xufang Luo", "gender": "Female", "institution": "Microsoft Research", "country": "CN", "position": "Researcher" }, { "name": "Yifan Yang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Yuqing Yang", "gender": "unknown", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" }, { "name": "weiquan Huang", "gender": "Female", "institution": "Tongji University", "country": "CN", "position": "PhD student" }, { "name": "Xiyang Dai", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 44, "n_ref": 76, "n_ref_all": 105, "n_fig": 2, "n_tab": 7, "L_tab": 3173, "n_element_tab": 262, "n_fig_1": 2, "n_tab_1": 10, "L_tab_1": 3008, "n_element_tab_1": 200, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 150826, "len_all_1": 78742, "len_abs": 2836, "len_title": 117, "len_sents": 43626, "len_sents_1": 38069, "n_sents": 318, "n_sents_1": 263, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 2865, "title": "LLM2CLIP: Extending the Capability Boundaries of CLIP through Large Language Models", "abs": "CLIP is one of the most important multimodal foundational models today, aligning visual and textual signals into a shared feature space using a simple contrastive learning loss on large-scale image-text pairs. What powers CLIP\u2019s capabilities? The rich supervision signals provided by natural language \u2014 the carrier of human knowledge \u2014 shape a powerful cross-modal representation space. As a result, CLIP supports a variety of tasks, including zero-shot classification, detection, segmentation, and cross-modal retrieval, significantly influencing the entire multimodal domain.\nHowever, with the rapid advancements in large language models (LLMs) like GPT-4 and LLaMA, the boundaries of language comprehension and generation are continually being pushed. This raises an intriguing question: \\emph{can the capabilities of LLMs be harnessed to further improve multimodal representation learning?}\nThe potential benefits of incorporating LLMs into CLIP are clear. LLMs\u2019 strong textual understanding can fundamentally improve CLIP\u2019s ability to handle image captions, drastically enhancing its ability to process long and complex texts \u2014 a well-known limitation of vanilla CLIP. Moreover, LLMs are trained on a vast corpus of text, possessing open-world knowledge. This allows them to expand on caption information during training, increasing the efficiency of the learning process.\nHowever, realizing this potential is challenging. Despite LLMs' powerful internal comprehension, their autoregressive nature hides this capability within the model, leading to output features with poor discriminability. Our experiments show that directly integrating LLMs into CLIP results in catastrophic performance drops.\nIn this paper, we propose LLM2CLIP, a novel approach that embraces the power of LLMs to unlock CLIP\u2019s potential. By fine-tuning the LLM in the caption space with contrastive learning, we extract its textual capabilities into the output embeddings, significantly improving the output layer's textual discriminability. We then design an efficient training process where the fine-tuned LLM acts as a powerful teacher for CLIP\u2019s visual encoder. Thanks to the LLM\u2019s presence, we can now incorporate longer and more complex captions without being restricted by vanilla CLIP text encoder\u2019s context window and ability limitations. Our experiments demonstrate that this approach brings substantial improvements in cross-modal tasks. Our method directly boosted the performance of the previously SOTA EVA02 model by 16.5\\% on both long-text and short-text retrieval tasks, transforming a CLIP model trained solely on English data into a state-of-the-art cross-lingual model. Moreover, when integrated into multimodal training with models like Llava 1.5, it consistently outperformed EVA02 across nearly all benchmarks, demonstrating comprehensive performance improvements.", "keywords": [ "CLIP", "Vision Language Model", "Contrastive Learning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "HeK3c9YIxG", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Ali Zeynalli", "gender": "Male", "institution": "institute of computer science, University of Tartu", "country": "EE", "position": "MS student" }, { "name": "Dmytro Fishman", "gender": "Male", "institution": "institute of computer science, University of Tartu", "country": "EE", "position": "Assistant Professor" }, { "name": "Illia Tsiporenko", "gender": "Male", "institution": "institute of computer science, University of Tartu", "country": "EE", "position": "PhD student" }, { "name": "Yaroslav Prytula", "gender": "Male", "institution": "University of Tartu", "country": "EE", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 20, "n_ref": 36, "n_ref_all": 38, "n_fig": 0, "n_tab": 3, "L_tab": 2524, "n_element_tab": 482, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 644, "n_element_tab_1": 141, "formula_len_all": 233, "formula_len_all_1": 152, "len_all": 99297, "len_all_1": 52982, "len_abs": 1619, "len_title": 76, "len_sents": 27956, "len_sents_1": 24674, "n_sents": 205, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 28, "L_abs": 1625, "title": "IAUNet: Instance-Aware U-Net", "abs": "Instance segmentation is critical in biomedical imaging for accurately distinguishing individual objects, such as cells, which often overlap and vary in size. Recent query-based methods\u2014where object-specific queries guide segmentation\u2014have shown strong performance in this task. While U-Net has been a go-to architecture in medical image segmentation, it was neither specifically designed for instance segmentation nor explored in the context of query-based approaches. In this work, we present IAUNet, a novel architecture that brings instance awareness to U-Net with query-based mechanisms to achieve superior pixel-to-instance clustering. The key design includes lightweight Instance Activation (IA) layers, which generate guided object queries by highlighting semantically important regions. Additionally, we propose a Parallel Dual-Path Transformer decoder that refines object-specific features across multiple scales, allowing us to assign multiple queries from different scale levels to a specific object. Finally, we introduce the 2025 Revvity Full Cell Segmentation Dataset, comprising hundreds of manually labeled cells from brightfield images. This dataset is unique in capturing the complex morphology of overlapping cell cytoplasm with an unprecedented level of detail, making it a valuable resource and benchmark for advancing instance segmentation in biomedical imaging. Experiments on multiple public datasets and our own show that IAUNet outperforms most state-of-the-art fully convolutional, transformer-based, and query-based models, setting a strong baseline for medical image instance segmentation tasks.", "keywords": [ "Medical and Biological Vision", "Cell Microscopy", "Instance Segmentation", "Deep Learning" ], "rating_list": [ 3, 3, 1, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 3, 1, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "He2FGdmsas", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Eunsu Baek", "gender": "unknown", "institution": "Seoul National University", "country": "", "position": "PhD student" }, { "name": "Hyung-Sin Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" }, { "name": "Sung-hwan Han", "gender": "Male", "institution": "Sogang University", "country": "KR", "position": "Undergrad student" }, { "name": "Taesik Gong", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 38, "n_ref": 128, "n_ref_all": 166, "n_fig": 23, "n_tab": 14, "L_tab": 6584, "n_element_tab": 945, "n_fig_1": 13, "n_tab_1": 4, "L_tab_1": 1822, "n_element_tab_1": 90, "formula_len_all": 351, "formula_len_all_1": 29, "len_all": 200521, "len_all_1": 64230, "len_abs": 1333, "len_title": 88, "len_sents": 64803, "len_sents_1": 31598, "n_sents": 448, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1338, "title": "Adaptive Camera Sensor for Vision Models", "abs": "Domain shift remains a persistent challenge in deep-learning-based computer vision, often requiring extensive model modifications or large labeled datasets to address. Inspired by human visual perception, which adjusts input quality through corrective lenses rather than over-training the brain, we propose Lens, a novel camera sensor control method that enhances model performance by capturing high-quality images from the model's perspective, rather than relying on traditional human-centric sensor control. Lens is lightweight and adapts sensor parameters to specific models and scenes in real-time. At its core, Lens utilizes VisiT, a training-free, model-specific quality indicator that evaluates individual unlabeled samples at test time using confidence scores, without additional adaptation costs. To validate Lens, we introduce ImageNet-ES Diverse, a new benchmark dataset capturing natural perturbations from varying sensor and lighting conditions. Extensive experiments on both ImageNet-ES and our new ImageNet-ES Diverse show that Lens significantly improves model accuracy across various baseline schemes for sensor control and model modification, while maintaining low latency in image captures. Lens effectively compensates for large model size differences and integrates synergistically with model improvement techniques.", "keywords": [ "domain-adaptation", "sensor-control", "domain-shift", "Out-of-Distribution", "data-centric" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "HdUkF1Qk7g", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Lei Yao", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Qi Li", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Tianyi Zhong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "MS student" }, { "name": "Yong Zhang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Full Professor" }, { "name": "Zhaoxia Li", "gender": "Female", "institution": "China Unicom Cloud Data Co., Ltd.", "country": "CN", "position": "Full Professor" }, { "name": "Zhenyu Zhang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 28, "n_ref": 55, "n_ref_all": 78, "n_fig": 8, "n_tab": 12, "L_tab": 6262, "n_element_tab": 594, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1983, "n_element_tab_1": 231, "formula_len_all": 1261, "formula_len_all_1": 997, "len_all": 146658, "len_all_1": 63143, "len_abs": 1560, "len_title": 167, "len_sents": 43912, "len_sents_1": 27091, "n_sents": 305, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 118, "L_abs": 1610, "title": "Diffusion-based Decoupled Deterministic and Uncertain Framework for Probabilistic Multivariate Time Series Forecasting", "abs": "Diffusion-based denoising models have demonstrated impressive performance in probabilistic forecasting for multivariate time series (MTS). Nonetheless, existing approaches often model the entire data distribution, neglecting the variability in uncertainty across different components of the time series. This paper introduces a Diffusion-based Decoupled Deterministic and Uncertain ($\\mathrm{D^3U}$) framework for probabilistic MTS forecasting. The framework integrates non-probabilistic forecasting with conditional diffusion generation, enabling both accurate point predictions and probabilistic forecasting. $\\mathrm{D^3U}$ utilizes a point forecasting model to non-probabilistically model high-certainty components in the time series, generating embedded representations that are conditionally injected into a diffusion model. To better model high-uncertainty components, a patch-based denoising network (PatchDN) is designed in the conditional diffusion model. Designed as a plug-and-play framework, $\\mathrm{D^3U}$ can be seamlessly integrated into existing point forecasting models to provide probabilistic forecasting capabilities. It can also be applied to other conditional diffusion methods that incorporate point forecasting models. Experiments on six real-world datasets demonstrate that our method achieves over a 20\\% improvement in both point and probabilistic forecasting performance in MTS long-term forecasting compared to state-of-the-art (SOTA) probabilistic forecasting methods. Additionally, extensive ablation studies further validate the effectiveness of the $\\mathrm{D^3U}$ framework.", "keywords": [ "long-term time series forecasting", "deep learning", "diffusion model" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "Hd4jB1ErMk", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Chenda Li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Samuele Cornell", "gender": "Male", "institution": "Universit\u00e0 Politecnica delle Marche", "country": "", "position": "PhD student" }, { "name": "Shinji Watanabe", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Wei Wang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Yanmin Qian", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Full Professor" }, { "name": "Zhengyang Chen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "leying zhang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 35, "n_ref": 59, "n_ref_all": 71, "n_fig": 1, "n_tab": 4, "L_tab": 943, "n_element_tab": 247, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 934, "n_element_tab_1": 246, "formula_len_all": 773, "formula_len_all_1": 773, "len_all": 114928, "len_all_1": 70170, "len_abs": 993, "len_title": 162, "len_sents": 31122, "len_sents_1": 28735, "n_sents": 243, "n_sents_1": 226, "venue": "ICLR2025", "decision": "withdraw", "L_title": 114, "L_abs": 1006, "title": "Speaking Guided by Listening: Unsupervised Text-to-Speech Generative Model Guided by End-to-End Speech Recognition", "abs": "We propose to utilize end-to-end automatic speech recognition (E2EASR) as a guidance model to realize unsupervised text-to-speech (TTS). An unconditional score-based generative model (SGM) is trained with untranscribed speech data. In the sampling stage, the unconditional score estimated by the SGM is combined with the gradients from ASR models by the Bayes rule to get the conditional score. We use a set of small ASR models trained only on $80$-hour labeled ASR data to guide the unconditional SGM and generate speech with high-quality scores in both objective and subjective evaluation. Similarly, we can also use additional speaker verification models to control speaker identity for the synthesized speech. That allows us to do the zero-shot TTS for the target speaker with a few seconds of enrollment speech. Our best unsupervised synthesized speech gets $\\sim8\\%$ word error rate in testing, and the best speaker-controlled TTS gets $3.3$ mean opinion score (MOS) in the speaker similarly testing.", "keywords": [ "Text-to-speech", "Diffusion", "Unsupervised learning" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "HcyVr9SlwR", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Hanxu Hu", "gender": "Male", "institution": "University of Zurich", "country": "CH", "position": "PhD student" }, { "name": "Hongyuan Lu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Researcher" }, { "name": "Huajian Zhang", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Vistor" }, { "name": "Jiao Yueyang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Ruijie Hou", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 26, "n_ref": 56, "n_ref_all": 72, "n_fig": 2, "n_tab": 6, "L_tab": 3131, "n_element_tab": 332, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1658, "n_element_tab_1": 160, "formula_len_all": 661, "formula_len_all_1": 528, "len_all": 134159, "len_all_1": 63113, "len_abs": 1279, "len_title": 129, "len_sents": 37869, "len_sents_1": 29120, "n_sents": 253, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1576, "title": "LBG: LNE-based Blocking Generation Against Data Contamination on Large Language Models", "abs": "Data contamination gradually becomes inevitable during the development of large language models (LLMs), meaning the training data commonly integrates those evaluation benchmarks unintentionally. This subsequently makes it hard to benchmark LLMs fairly. This paper introduces a novel framework called LBG (\\textbf{L}NE-based \\textbf{B}locking \\textbf{G}eneration) for both contamination detection and mitigation for evaluating contaminated LLMs. For the first component of LBG, LBG reports a SOTA performance on our proposed length normalized entropy (LNE) to identify potential contamination by detecting anomalies on possibly contaminated LLMs. For the second component of LBG, LBG reports a SOTA performance on the mitigation of the impact of data contamination by applying LNE within a novel blocking generation strategy, specialized to adjust generation processes and re-calibrate performance metrics by suppressing the maximum value of output candidates during the generation process. We conduct extensive experiments on both contamination detection and contamination mitigation evaluation tasks, on both code generation and mathematical reasoning scenarios. The results indicate that LBG achieves an obvious SOTA performance throughout the experiments conducted in this paper. Simultaneously, LGB is lightweight and costs obviously fewer computational costs (nearly 25x) than the previous work. We hope our method will open new research avenues on data contamination for LLMs. We plan to release the resources upon publication of this work to facilitate future work.", "keywords": [ "detecting data contamination", "fairly evaluating LLMs" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 3, 1 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 2, 2 ] }, { "paper_id": "Hcb2cgPbMg", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Alex Lewandowski", "gender": "unknown", "institution": "University of Alberta", "country": "", "position": "PhD student" }, { "name": "Dale Schuurmans", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Professor" }, { "name": "Marlos C. Machado", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Assistant Professor" }, { "name": "Mateusz Ostaszewski", "gender": "Male", "institution": "Warsaw University of Technology", "country": "PL", "position": "Assistant Professor" }, { "name": "Saurabh Kumar", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Andras Gyorgy", "gender": "unknown", "institution": "DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Micha\u0142 Pawe\u0142 Bortkiewicz", "gender": "Male", "institution": "Warsaw University of Technology", "country": "", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 10, "n_ref_uni": 69, "n_ref": 162, "n_ref_all": 195, "n_fig": 13, "n_tab": 2, "L_tab": 874, "n_element_tab": 53, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1905, "formula_len_all_1": 1063, "len_all": 184388, "len_all_1": 68700, "len_abs": 1382, "len_title": 47, "len_sents": 59227, "len_sents_1": 34896, "n_sents": 388, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1388, "title": "Learning Continually by Spectral Regularization", "abs": "Loss of plasticity is a phenomenon where neural networks can become more difficult to train over the course of learning. Continual learning algorithms seek to mitigate this effect by sustaining good performance while maintaining network trainability. We develop a new technique for improving continual learning inspired by the observation that the singular values of the neural network parameters at initialization are an important factor for trainability during early phases of learning. From this perspective, we derive a new spectral regularizer for continual learning that better sustains these beneficial initialization properties throughout training. In particular, the regularizer keeps the maximum singular value of each layer close to one. Spectral regularization directly ensures that gradient diversity is maintained throughout training, which promotes continual trainability, while minimally interfering with performance in a single task. We present an experimental analysis that shows how the proposed spectral regularizer can sustain trainability and performance across a range of model architectures in continual supervised and reinforcement learning settings. Spectral regularization is less sensitive to hyperparameters while demonstrating better training in individual tasks, sustaining trainability as new tasks arrive, and achieving better generalization performance..", "keywords": [ "plasticity", "neural networks", "spectral regularization", "continual learning" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "HcY3fbVDqa", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Edgar Mariano Marroquin", "gender": "Male", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "PhD student" }, { "name": "Volodymyr Kuleshov", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Assistant Professor" }, { "name": "Yair Schiff", "gender": "Male", "institution": "Department of Computer Science, Cornell University", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 10, "n_ref_uni": 22, "n_ref": 51, "n_ref_all": 72, "n_fig": 5, "n_tab": 7, "L_tab": 1109, "n_element_tab": 159, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 300, "n_element_tab_1": 45, "formula_len_all": 485, "formula_len_all_1": 457, "len_all": 106788, "len_all_1": 59739, "len_abs": 971, "len_title": 121, "len_sents": 31336, "len_sents_1": 26650, "n_sents": 195, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 976, "title": "Non-Parametric State-Space Models Over Datapoints and Sequence Alignments", "abs": "Non-parametric models are flexible and can leverage a context set to express rich mappings from inputs to outputs. However, these methods often scale super-linearly in context size, e.g., attention-based\nmethods scale quadratically in the number of data points, which in turn limits model expressivity. In this work, we leverage advances in state-space modeling and introduce Non-Parametric State\n Space Models (NPSSM). We find that NPSSMs attain similar performance to existing non-parametric attention-based models while scaling linearly in the number of datapoints. We apply NPSSMs to the task of genotype imputation, where the linear scaling enables larger context sets resulting in competitive performance relative to other methods and widely used industry-standard tools. We also demonstrate the effectiveness of\nNPSSMs in the context of meta-learning where the ability to efficiently scale to larger training sets provides more favorable compute-to-accuracy tradeoffs.", "keywords": [ "Non-Parametric Models", "State-Space Models", "Genotype Imputation", "Comp Bio" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "HcRpl6hOAl", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Bin Cui", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Jiaze Song", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "Undergrad student" }, { "name": "Keming Wu", "gender": "Male", "institution": "School of Software, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Minghao Xu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xiangxin Zhou", "gender": "Not Specified", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zhang wen tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 55, "n_ref": 84, "n_ref_all": 102, "n_fig": 6, "n_tab": 1, "L_tab": 175, "n_element_tab": 15, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1180, "n_element_tab_1": 36, "formula_len_all": 626, "formula_len_all_1": 460, "len_all": 139664, "len_all_1": 66792, "len_abs": 1302, "len_title": 145, "len_sents": 37032, "len_sents_1": 32133, "n_sents": 232, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1307, "title": "Modeling All-Atom Glycan Structures via Hierarchical Message Passing and Multi-Scale Pre-training", "abs": "Understanding the various properties of glycans with machine learning has shown some preliminary promise. However, previous methods mainly focused on modeling the backbone structure of glycans as graphs of monosaccharides (i.e., sugar units), while they neglected the atomic structures underlying each monosaccharide, which are actually important indicators of glycan properties. In this work, we fill this blank by introducing the GlycanAA model for All-Atom-wise Glycan modeling. GlycanAA models a glycan as a heterogeneous graph with monosaccharide nodes representing its global backbone structure and atom nodes representing its local atomic-level structures. Based on such a graph, GlycanAA performs hierarchical message passing to capture from local atomic-level interactions to global monosaccharide-level interactions hierarchically. To further enhance the model capability, we pre-train GlycanAA on a high-quality unlabeled glycan dataset in a self-supervised way, deriving the PreGlycanAA model. Specifically, we design a multi-scale mask prediction algorithm to endow the model with knowledge about different levels of dependencies in a glycan. Extensive benchmark results show the superiority of GlycanAA over existing glycan encoders and verify the further improvements achieved by PreGlycanAA.", "keywords": [ "Glycan Machine Learning", "Heterogeneous Graph Modeling", "Self-Supervised Pre-training" ], "rating_list": [ 6, 5, 6, 5, 5 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 4, 4, 3 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 2, 4, 2, 3, 3 ] }, { "paper_id": "HbqYdvL1mB", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Linchao Zhu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xihang Yue", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 40, "n_formula_1": 12, "n_ref_uni": 42, "n_ref": 167, "n_ref_all": 173, "n_fig": 1, "n_tab": 9, "L_tab": 2247, "n_element_tab": 104, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 2153, "n_element_tab_1": 99, "formula_len_all": 3775, "formula_len_all_1": 698, "len_all": 184838, "len_all_1": 68421, "len_abs": 1111, "len_title": 90, "len_sents": 59982, "len_sents_1": 27937, "n_sents": 561, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1113, "title": "Point-Calibrated Spectral Neural Operators", "abs": "Two typical neural models have been extensively studied for operator learning, learning in spatial space via attention mechanism or learning in spectral space via spectral analysis technique such as Fourier Transform. Spatial learning enables point-level flexibility but lacks global continuity constraint, while spectral learning enforces spectral continuity prior but lacks point-wise adaptivity. This work innovatively combines the continuity prior and the point-level flexibility, with the introduced Point-Calibrated Spectral Transform. It achieves this by calibrating the preset spectral eigenfunctions with the predicted point-wise frequency preference via neural gate mechanism. Beyond this, we introduce Point-Calibrated Spectral Neural Operators, which learn operator mappings by approximating functions with the point-level adaptive spectral basis, thereby not only preserving the benefits of spectral prior but also boasting the superior adaptability comparable to the attention mechanis. Comprehensive experiments demonstrate its consistent performance enhancement in extensive PDE solving scenarios.", "keywords": [ "Operator Learning", "PDE" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "Hbpzrh7JbN", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Jeff Guo", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Philippe Schwaller", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "", "position": "Assistant Professor" } ], "n_formula": 31, "n_formula_1": 5, "n_ref_uni": 71, "n_ref": 322, "n_ref_all": 371, "n_fig": 15, "n_tab": 31, "L_tab": 36174, "n_element_tab": 5183, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1006, "n_element_tab_1": 84, "formula_len_all": 2977, "formula_len_all_1": 240, "len_all": 442506, "len_all_1": 60283, "len_abs": 1752, "len_title": 127, "len_sents": 100199, "len_sents_1": 29584, "n_sents": 841, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1256, "title": "Saturn: Sample-efficient Generative Molecular Design using Memory Manipulation", "abs": "Generative molecular design for drug discovery has very recently achieved a wave\nof experimental validation, with language-based backbones being the most common\narchitectures employed. The most important factor for downstream success is\nwhether an *in silico* oracle is well correlated with the desired end-point. To this end,\ncurrent methods use cheaper proxy oracles with higher throughput before evaluating\nthe most promising subset with high-fidelity oracles. The ability to directly optimize\nhigh-fidelity oracles would greatly enhance generative design and be expected to\nimprove hit rates. However, current models are not efficient enough to consider such\na prospect, exemplifying the sample efficiency problem. In this work, we introduce\n**Saturn**, which leverages the Augmented Memory algorithm and demonstrates the\nfirst application of the Mamba architecture for generative molecular design. We\nelucidate *how* experience replay with data augmentation improves sample efficiency\nand *how* Mamba synergistically exploits this mechanism. Saturn outperforms 22\nmodels on multi-parameter optimization tasks relevant to drug discovery and may\npossess sufficient sample efficiency to consider the prospect of directly optimizing\nhigh-fidelity oracles.", "keywords": [ "generative design", "drug discovery", "sample efficiency", "language models", "reinforcement learning", "scaling" ], "rating_list": [ 3, 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 1, 2, 3 ], "confidence_list": [ 3, 3, 4, 2, 3 ] }, { "paper_id": "HbbnlrmsAH", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Chenghao Liu", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Researcher" }, { "name": "Doyen Sahoo", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Researcher" }, { "name": "Gerald Woo", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "PhD student" }, { "name": "Juncheng Liu", "gender": "unknown", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "Roger Zimmermann", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Silvio Savarese", "gender": "Male", "institution": "Salesforce", "country": "", "position": "Chief Scientist" }, { "name": "Xu Liu", "gender": "unknown", "institution": "Salesforce Research", "country": "SG", "position": "Intern" }, { "name": "Yuxuan Liang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Ibrahim Taha Aksu", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 30, "n_ref": 98, "n_ref_all": 127, "n_fig": 14, "n_tab": 12, "L_tab": 11364, "n_element_tab": 1125, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 5003, "n_element_tab_1": 302, "formula_len_all": 348, "formula_len_all_1": 334, "len_all": 174186, "len_all_1": 64159, "len_abs": 2198, "len_title": 132, "len_sents": 47990, "len_sents_1": 26417, "n_sents": 347, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1937, "title": "Moirai-MoE: Empowering Time Series Foundation Models with Sparse Mixture of Experts", "abs": "Time series foundation models have demonstrated impressive performance as zero-shot forecasters, i.e. tackling a wide variety of downstream forecasting tasks without explicit task-specific training. However, achieving effectively unified training on time series remains an open challenge. Existing approaches introduce some level of model specialization to account for the highly heterogeneous nature of time series data. For instance, Moirai pursues unified training by employing multiple input/output projection layers, each tailored to handle time series at a specific frequency. Similarly, TimesFM maintains a frequency embedding dictionary for this purpose. We identify two major drawbacks to this human-imposed frequency-level model specialization: (1) Frequency is not a reliable indicator of the underlying patterns in time series. For example, time series with different frequencies can display similar patterns, while those with the same frequency may exhibit varied patterns. (2) Non-stationarity is an inherent property of real-world time series, leading to varied distributions even within a short context window of a single time series. Frequency-level specialization is too coarse-grained to capture this level of diversity. To address these limitations, this paper introduces Moirai-MoE, using a single input/output projection layer while delegating the modeling of diverse time series patterns to the sparse mixture of experts (MoE) within Transformers. With these designs, Moirai-MoE reduces reliance on human-defined heuristics and enables automatic token-level specialization. Extensive experiments on 39 datasets demonstrate the superiority of Moirai-MoE over existing foundation models in both in-distribution and zero-shot scenarios. Furthermore, this study conducts comprehensive model analyses to explore the inner workings of time series MoE foundation models and provides valuable insights for future research.", "keywords": [ "Time Series Foundation Models", "Mixture of Experts" ], "rating_list": [ 3, 6, 5, 8 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 5, 3, 4 ] }, { "paper_id": "HbZrxBXzks", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Carl Edwards", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "PhD student" }, { "name": "Heng Ji", "gender": "Female", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Limei Wang", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Shuiwang Ji", "gender": "Male", "institution": "Texas A&M University", "country": "US", "position": "Professor" }, { "name": "Shurui Gui", "gender": "Male", "institution": "Texas A&M University", "country": "US", "position": "PhD student" }, { "name": "Xiner Li", "gender": "Female", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Youzhi Luo", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yuchao Lin", "gender": "Male", "institution": "Texas A&M", "country": "US", "position": "PhD student" } ], "n_formula": 51, "n_formula_1": 27, "n_ref_uni": 74, "n_ref": 148, "n_ref_all": 178, "n_fig": 8, "n_tab": 18, "L_tab": 4987, "n_element_tab": 521, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 597, "n_element_tab_1": 89, "formula_len_all": 25036, "formula_len_all_1": 18192, "len_all": 328442, "len_all_1": 105703, "len_abs": 691, "len_title": 121, "len_sents": 86530, "len_sents_1": 41798, "n_sents": 715, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 693, "title": "Geometry Informed Tokenization of Molecules for Language Model Generation", "abs": "We consider molecule generation in 3D space using language models (LMs), which requires discrete tokenization of 3D molecular geometries. Although tokenization of molecular graphs exists, that for 3D geometries is largely unexplored. Here, we attempt to bridge this gap by proposing the Geo2Seq, which converts molecular geometries into SE(3)-invariant 1D discrete sequences. Geo2Seq consists of canonical labeling and invariant spherical representation steps, which together maintain geometric and atomic fidelity in a format conducive to LMs. Our experiments show that, when coupled with Geo2Seq, various LMs excel in molecular geometry generation, especially in controlled generation tasks.", "keywords": [ "Generative models", "language models", "tokenization", "3d molecule generation", "SE(3)-equivariance" ], "rating_list": [ 8, 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 4 ], "presentation_list": [ 4, 3, 2, 4, 3 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 4, 3 ] }, { "paper_id": "Hb3x52Jliq", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Huihan Wang", "gender": "Female", "institution": "China Medical University Shenyang", "country": "CN", "position": "Full Professor" }, { "name": "Tiezheng Guo", "gender": "unknown", "institution": "Northeastern University", "country": "CN", "position": "PhD student" }, { "name": "Wang Chen", "gender": "Female", "institution": "Northeastern University", "country": "CN", "position": "PhD student" }, { "name": "Yanyi Liu", "gender": "Male", "institution": "Northeastern University", "country": "CN", "position": "PhD student" }, { "name": "Yingyou Wen", "gender": "unknown", "institution": "Neusoft Group.", "country": "CN", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 24, "n_ref": 29, "n_ref_all": 40, "n_fig": 5, "n_tab": 2, "L_tab": 431, "n_element_tab": 79, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 431, "n_element_tab_1": 79, "formula_len_all": 336, "formula_len_all_1": 336, "len_all": 78552, "len_all_1": 58847, "len_abs": 1396, "len_title": 129, "len_sents": 33854, "len_sents_1": 31678, "n_sents": 170, "n_sents_1": 158, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1392, "title": "A Research on Result Interpretability of Medical AI Based on Large Language Model", "abs": "Explainability is one of the important challenges facing the application of medical AI. The existing AI explainability research is more of a kind of process explainability study. Drawing on the behavioral habits of human beings to communicate on a certain topic, this paper proposes a definition of result interpretability for medical AI, divides explainable medical AI research into three phases: data explainability, process explainability and result interpretability, and argues that once an AI model reaches a certain result interpretability metric, we can accept its conclusions and apply it to the clinic without having to wait until human beings fully understand the operation and decision-making mechanism of the AI model before using it. In this regard, we propose the c oncept of interpretative integrity. Further, we propose an architecture for result-interpretable medical AI system based on AI-Agent and build a result-interpretable system around risk prediction AI model for amyloidosis, which enables professional interpretation of the result of the risk prediction model for amyloidosis disease through a large language model and supports professional Q&A with clinicians. The implementation of the system enhances clinicians' professional acceptance of medical AI models, and provides a more feasible realization path for the large-scale application of AI-assisted diagnosis.", "keywords": [ "XAI", "AI Agent", "Medical AI", "Amyloidosis", "LLM", "AIGC" ], "rating_list": [ 5, 8, 3, 3 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "HafxTJjo6a", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Anurag Kumar", "gender": "Male", "institution": "Meta", "country": "US", "position": "Research Scientist" }, { "name": "Chao Huang", "gender": "Male", "institution": "Department of Computer Science, University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Chenliang Xu", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Associate Professor" }, { "name": "Susan Liang", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Yapeng Tian", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "Assistant Professor" }, { "name": "Yunlong Tang", "gender": "unknown", "institution": "University of Rochester", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 41, "n_ref": 101, "n_ref_all": 142, "n_fig": 16, "n_tab": 1, "L_tab": 196, "n_element_tab": 38, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 595, "n_element_tab_1": 95, "formula_len_all": 368, "formula_len_all_1": 261, "len_all": 144358, "len_all_1": 63902, "len_abs": 908, "len_title": 0, "len_sents": 42291, "len_sents_1": 30871, "n_sents": 328, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 912, "title": "Scaling Concept With Text-Guided Diffusion Models", "abs": "Text-guided diffusion models have revolutionized generative tasks by producing high-fidelity content based on text descriptions. Additionally, they have enabled an editing paradigm where concepts can be replaced through text conditioning. In this work, we explore a novel paradigm: instead of replacing a concept, can we scale it? We conduct an empirical study to investigate concept decomposition trends in text-guided diffusion models. Leveraging these insights, we propose a simple yet effective method, ScalingConcept, designed to enhance or suppress existing concepts in real input without introducing new ones. To systematically evaluate our method, we introduce the WeakConcept-10 dataset. More importantly, ScalingConcept enables a range of novel zero-shot applications across both image and audio domains, including but not limited to canonical pose generation and generative sound highlighting/removal.", "keywords": [ "Diffusion models", "zero-shot applications", "image/audio" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "HaXlWs1LX8", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Haohan Wang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Ramtin Pedarsani", "gender": "unknown", "institution": "UC Santa Barbara", "country": "", "position": "Assistant Professor" }, { "name": "Yao Qin", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "Assistant Professor" }, { "name": "Zhiyu Xue", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 14, "n_ref_uni": 46, "n_ref": 73, "n_ref_all": 100, "n_fig": 12, "n_tab": 2, "L_tab": 729, "n_element_tab": 87, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 2511, "n_element_tab_1": 107, "formula_len_all": 740, "formula_len_all_1": 846, "len_all": 119674, "len_all_1": 54291, "len_abs": 1179, "len_title": 83, "len_sents": 30583, "len_sents_1": 23231, "n_sents": 196, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1191, "title": "Conflict-Aware Adversarial Training", "abs": "Adversarial training is the most effective method to obtain adversarial robustness for deep neural networks by directly involving adversarial samples in the training procedure. To obtain an accurate and robust model, the weighted-average method is applied to optimize standard loss and adversarial loss simultaneously. In this paper, we argue that the weighted-average method does not provide the best tradeoff for the standard performance and adversarial robustness. We argue that the failure of the weighted-average method is due to the conflict between the gradients derived from standard and adversarial loss, and further demonstrate such a conflict increases with attack budget theoretically and practically. To alleviate this problem, we propose a new trade-off paradigm for adversarial training with a conflict-aware factor for the convex combination of standard and adversarial loss, named \\textbf{Conflict-Aware Adversarial Training~(CA-AT)}. Comprehensive experimental results show that CA-AT consistently offers a superior trade-off between standard performance and adversarial robustness under the settings of adversarial training from scratch and parameter-efficient finetuning.", "keywords": [ "Adversarial Training", "Robustness" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "HaX48yksVL", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Ayush Sekhari", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Gautam Kamath", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Jimmy Z. Di", "gender": "Male", "institution": "University of Waterloo", "country": "", "position": "MS student" }, { "name": "Martin Pawelczyk", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Postdoc" }, { "name": "Yiwei Lu", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" }, { "name": "Seth Neel", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Assistant Professor" } ], "n_formula": 20, "n_formula_1": 4, "n_ref_uni": 50, "n_ref": 122, "n_ref_all": 171, "n_fig": 23, "n_tab": 10, "L_tab": 1698, "n_element_tab": 131, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 482, "n_element_tab_1": 28, "formula_len_all": 1208, "formula_len_all_1": 153, "len_all": 254328, "len_all_1": 66855, "len_abs": 1295, "len_title": 57, "len_sents": 87533, "len_sents_1": 34410, "n_sents": 613, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1376, "title": "Machine Unlearning Fails to Remove Data Poisoning Attacks", "abs": "We revisit the efficacy of several practical methods for approximate machine unlearning developed for large-scale deep learning. In addition to complying with data deletion requests, one often-cited potential application for unlearning methods is to remove the effects of training on poisoned data. We experimentally demonstrate that, while existing unlearning methods have been demonstrated to be effective in a number of evaluation settings (e.g., alleviating membership inference attacks), they fail to remove the effects of data poisoning, across a variety of types of poisoning attacks (indiscriminate, targeted, and a newly-introduced Gaussian poisoning attack) and models (image classifiers and LLMs); even when granted a relatively large compute budget. In order to precisely characterize unlearning efficacy, we introduce new evaluation metrics for unlearning based on data poisoning. Our results suggest that a broader perspective, including a wider variety of evaluations, is required to avoid a false sense of confidence in machine unlearning procedures for deep learning without provable guarantees. Moreover, while unlearning methods show some signs of being useful to efficiently remove poisoned datapoints without having to retrain, our work suggests that these methods are not yet \"ready for prime time\", and currently provide limited benefit over retraining.", "keywords": [ "machine unlearning", "data poisoning" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 4, 4 ] }, { "paper_id": "Ha6RTeWMd0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 18, "author_info_list": [ { "name": "Chaitanya Ryali", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Chao-Yuan Wu", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Chloe Rolland", "gender": "unknown", "institution": "Sorbonne Universit\u00e9 - Facult\u00e9 des Sciences (Paris VI)", "country": "", "position": "PhD student" }, { "name": "Christoph Feichtenhofer", "gender": "Male", "institution": "Meta FAIR", "country": "US", "position": "Principal Researcher" }, { "name": "Eric Mintun", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Engineer" }, { "name": "Haitham Khedr", "gender": "Male", "institution": "University of California, Irvine", "country": "", "position": "PhD student" }, { "name": "Junting Pan", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Kalyan Vasudev Alwala", "gender": "Male", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Laura Gustafson", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Research Engineer" }, { "name": "Nicolas Carion", "gender": "Male", "institution": "New York University", "country": "", "position": "Postdoc" }, { "name": "Nikhila Ravi", "gender": "Female", "institution": "Facebook", "country": "", "position": "Researcher" }, { "name": "Piotr Dollar", "gender": "Male", "institution": "FAIR", "country": "", "position": "Research Scientist" }, { "name": "Roman R\u00e4dle", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Ronghang Hu", "gender": "Male", "institution": "Meta AI", "country": "", "position": "Researcher" }, { "name": "Ross Girshick", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Tengyu Ma", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Valentin Gabeur", "gender": "Male", "institution": "Meta", "country": "", "position": "Postdoc" }, { "name": "Yuan-Ting Hu", "gender": "Female", "institution": "Meta AI", "country": "", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 101, "n_ref": 203, "n_ref_all": 269, "n_fig": 17, "n_tab": 14, "L_tab": 995, "n_element_tab": 113, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 800, "n_element_tab_1": 51, "formula_len_all": 87, "formula_len_all_1": 0, "len_all": 339853, "len_all_1": 66908, "len_abs": 865, "len_title": 92, "len_sents": 99043, "len_sents_1": 32953, "n_sents": 787, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 44, "L_abs": 870, "title": "SAM 2: Segment Anything in Images and Videos", "abs": "We present Segment Anything Model 2 (SAM 2), a foundation model towards solving promptable visual segmentation in images and videos. We build a data engine, which improves model and data via user interaction, to collect the largest video segmentation dataset to date. Our model is a simple transformer architecture with streaming memory for real-time video processing. SAM 2 trained on our data provides strong performance across a wide range of tasks. In video segmentation, we observe better accuracy, using 3x fewer interactions than prior approaches. In image segmentation, our model is more accurate and 6x faster than the Segment Anything Model (SAM). We believe that our data, model, and insights will serve as a significant milestone for video segmentation and related perception tasks. We are releasing our main model, the dataset, an interactive demo and code.", "keywords": [ "computer vision", "video segmentation", "image segmentation" ], "rating_list": [ 10, 8, 8, 10 ], "soundness_list": [ 4, 3, 4, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 4, 3, 4, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "HZz81oCNlp", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Chen Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Hongming Chen", "gender": "Male", "institution": "Dalian Martime University", "country": "CN", "position": "PhD student" }, { "name": "Jinshan Pan", "gender": "unknown", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xiang Chen", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Xianping Fu", "gender": "Male", "institution": "Dalian Martime University", "country": "CN", "position": "Full Professor" }, { "name": "Zhuoran Zheng", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Postdoc" } ], "n_formula": 9, "n_formula_1": 6, "n_ref_uni": 32, "n_ref": 116, "n_ref_all": 144, "n_fig": 17, "n_tab": 9, "L_tab": 3777, "n_element_tab": 361, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 2004, "n_element_tab_1": 233, "formula_len_all": 467, "formula_len_all_1": 409, "len_all": 150543, "len_all_1": 64083, "len_abs": 1168, "len_title": 131, "len_sents": 41537, "len_sents_1": 28124, "n_sents": 331, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1174, "title": "Towards Ultra-High-Definition Image Deraining: A Benchmark and An Efficient Method", "abs": "Despite significant progress has been made in image deraining, existing approaches are mostly carried out on low-resolution images. The effectiveness of these methods on high-resolution images is still unknown, especially for ultra-high-definition (UHD) images, given the continuous advancement of imaging devices. In this paper, we focus on the task of UHD image deraining, and contribute the first large-scale UHD image deraining dataset, 4K-Rain13k, that contains 13,000 image pairs at 4K resolution. Based on this dataset, we conduct a benchmark study on existing methods for processing UHD images. Furthermore, we develop an effective and efficient architecture (called UDR-Mixer) to better solve this task. Specifically, our method contains two building components: a spatial feature rearrangement layer that captures long-range information of UHD images, and a frequency feature modulation layer that facilitates high-quality UHD image reconstruction. Extensive experimental results demonstrate that our method performs favorably against the state-of-the-art approaches while maintaining a lower model complexity. The code and dataset will be available to the public.", "keywords": [ "Image deraining", "Rain removal", "Ultra-high-definition", "4K image", "Benchmark dataset", "Vision MLP" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 3, 5, 5 ] }, { "paper_id": "HZxJfzs3w6", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Arash Vahdat", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Research Scientist" }, { "name": "Michael Pritchard", "gender": "Male", "institution": "University of California, Irvine", "country": "", "position": "Associate Professor" }, { "name": "Morteza Mardani", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Principal Researcher" }, { "name": "Noah D Brenowitz", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Stathi Fotiadis", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Tomas Geffner", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yair Cohen", "gender": "Male", "institution": "NVIDIA", "country": "", "position": "Researcher" } ], "n_formula": 43, "n_formula_1": 27, "n_ref_uni": 30, "n_ref": 56, "n_ref_all": 86, "n_fig": 8, "n_tab": 13, "L_tab": 4721, "n_element_tab": 618, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2093, "formula_len_all_1": 970, "len_all": 178794, "len_all_1": 58822, "len_abs": 1474, "len_title": 104, "len_sents": 58555, "len_sents_1": 27889, "n_sents": 504, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1496, "title": "Stochastic Flow Matching for Resolving Small-Scale Physics", "abs": "Conditioning diffusion and flow models have proven effective for super-resolving small-scale details in natural images. However, in physical sciences such as weather, super-resolving small-scale details poses significant challenges due to: $(i)$ misalignment between input and output distributions (i.e., solutions to distinct partial differential equations (PDEs) follow different trajectories), $(ii)$ multi-scale dynamics, deterministic dynamics at large scales vs. stochastic at small scales, and $(iii)$ limited data, increasing the risk of overfitting. To address these challenges, we propose encoding the inputs to a \\textit{latent} base distribution that is closer to the target distribution, followed by flow matching to generate small-scale physics. The encoder captures the deterministic components, while flow matching adds stochastic small-scale details. To account for uncertainty in the deterministic part, we inject noise into the encoder's output using an adaptive noise scaling mechanism, which is dynamically adjusted based on maximum-likelihood estimates of the encoder\u2019s predictions. We conduct extensive experiments on both the real-world CWA weather dataset and the PDE-based Kolmogorov dataset, with the CWA task involving super-resolving the weather variables for the region of Taiwan from 25 km to 2 km scales. Our results show that the proposed stochastic flow matching (SFM) framework significantly outperforms existing methods such as conditional diffusion and flows.", "keywords": [ "flow matching", "diffusion models", "multiscale dynamics", "misaligned data distributions", "superresolution" ], "rating_list": [ 5, 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 2 ], "confidence_list": [ 4, 2, 3, 3, 3 ] }, { "paper_id": "HZgZrtIreg", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Alice Oh", "gender": "Female", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Dongkwan Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 34, "n_ref": 78, "n_ref_all": 94, "n_fig": 3, "n_tab": 6, "L_tab": 2816, "n_element_tab": 280, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2431, "n_element_tab_1": 273, "formula_len_all": 1198, "formula_len_all_1": 1520, "len_all": 126235, "len_all_1": 73748, "len_abs": 925, "len_title": 99, "len_sents": 38464, "len_sents_1": 29215, "n_sents": 286, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 932, "title": "Generalizing Weisfeiler-Lehman Kernels to Subgraphs", "abs": "Subgraph representation learning has been effective in solving various real-world problems. However, current graph neural networks (GNNs) produce suboptimal results for subgraph-level tasks due to their inability to capture complex interactions within and between subgraphs. To provide a more expressive and efficient alternative, we propose WLKS, a Weisfeiler-Lehman (WL) kernel generalized for subgraphs by applying the WL algorithm on induced $k$-hop neighborhoods. We combine kernels across different $k$-hop levels to capture richer structural information that is not fully encoded in existing models. Our approach can balance expressiveness and efficiency by eliminating the need for neighborhood sampling. In experiments on eight real-world and synthetic benchmarks, WLKS significantly outperforms leading approaches on five datasets while reducing training time, ranging from 0.01x to 0.25x compared to the state-of-the-art.", "keywords": [ "Graph Neural Networks", "Graph Kernels", "Subgraphs" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "HZVIQE1MsJ", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Dong Yan", "gender": "Male", "institution": "Baichuan Intelligent Technology", "country": "", "position": "Researcher" }, { "name": "Qingyao Ai", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Qiuchi Li", "gender": "Male", "institution": "Copenhagen University", "country": "DK", "position": "Assistant Professor" }, { "name": "wei shen", "gender": "unknown", "institution": "Baichuan", "country": "", "position": "Researcher" }, { "name": "Xiangsheng Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Yiqun LIU", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Yujia Zhou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Ziyi Ye", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 40, "n_ref": 71, "n_ref_all": 119, "n_fig": 5, "n_tab": 5, "L_tab": 1454, "n_element_tab": 149, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1250, "n_element_tab_1": 128, "formula_len_all": 713, "formula_len_all_1": 761, "len_all": 207421, "len_all_1": 73523, "len_abs": 1928, "len_title": 94, "len_sents": 71295, "len_sents_1": 32651, "n_sents": 541, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 1394, "title": "Learning LLM-as-a-Judge for Preference Alignment", "abs": "Learning from preference feedback is a common practice for aligning large language models (LLMs) with human value. Conventionally, preference data is learned and encoded into a scalar reward model that connects a value head with an LLM to produce a scalar score as preference. However, scalar models lack interpretability and are known to be susceptible to biases in datasets. This paper investigates leveraging LLM itself to learn from such preference data and serve as a judge to address both limitations in one shot. Specifically, we prompt the pre-trained LLM to generate initial judgment pairs with contrastive preference in natural language form. The self-generated contrastive judgment pairs are used to train the LLM-as-a-Judge with Direct Preference Optimization (DPO) and incentivize its reasoning capability as a judge. This proposal of learning the LLMas-a-Judge using self-generated Contrastive judgments (Con-J) ensures natural interpretability through the generated rationales supporting the judgments, and demonstrates higher robustness against bias compared to scalar models. Experimental results show that Con-J outperforms the scalar reward model trained on the same collection of preference data, and outperforms a series of open-source and closed-source generative LLMs. We open-source the training process and model weights of Con-J at https://github.com/YeZiyi1998/Con-J.", "keywords": [ "LLM alignment", "Generative judge", "LLM-as-judge", "Reward model" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "HYsU5X4kE5", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Nanxu Gong", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Wangyang Ying", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Yanjie Fu", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Associate Professor" }, { "name": "Bai Haoyue", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 37, "n_ref": 49, "n_ref_all": 60, "n_fig": 5, "n_tab": 4, "L_tab": 990, "n_element_tab": 79, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 990, "n_element_tab_1": 79, "formula_len_all": 788, "formula_len_all_1": 788, "len_all": 117380, "len_all_1": 65784, "len_abs": 1510, "len_title": 123, "len_sents": 35323, "len_sents_1": 31251, "n_sents": 248, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1452, "title": "GCNFT: Graph Convolutional Networks Aware Generative Feature Transformation", "abs": "Feature transformation for attributed graphs converts raw node attributes into augmented features that preserve node and structure information. Relevant literature either fails to capture graph structures (e.g., manual handcrafting, discrete search), or is latent and hard to interpret (e.g., GCNs). How can we automatically reconstruct explicit features of an attributed graph while effectively integrating graph structures and attributes? We generalize the learning task under such setting as a GCN-aware Feature Transformation (GCNFT) problem. GCNFT imposes two under-addressed challenges: 1) quantifying GCN awareness and 2) bridging GCN awareness and feature transformation. To tackle these challenges, we propose a graph convolution structure score guided generative learning framework to solve GCNFT. To quantify GCN awareness, we interpret GCN as a gap minimization process between ideal and current node representations in iterative Laplacian smoothing, and develop a task-agnostic structure score to approximate GCN awareness. To incorporate GCN awareness, we model feature transformation as sequential generative learning so that we pave a way to leverage the structures score to guide the generative learning and encourage graph structure alignment. Extensive experiments demonstrate the proposed GCN-aware approach outperforms feature transformation baselines with an improvement of 3\\% to 20\\% over node, link, and graph prediction tasks.", "keywords": [ "Feature Transformation", "Data-centric AI", "Representation Learning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "HYaUHZAoPc", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Guiliang Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Nan Fang", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Wei Gong", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 61, "n_ref": 114, "n_ref_all": 150, "n_fig": 15, "n_tab": 8, "L_tab": 4933, "n_element_tab": 408, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1066, "formula_len_all_1": 783, "len_all": 200025, "len_all_1": 68194, "len_abs": 2455, "len_title": 146, "len_sents": 56168, "len_sents_1": 31907, "n_sents": 476, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1426, "title": "Offline Inverse Constrained Reinforcement Learning for Safe-Critical Decision Making in Healthcare", "abs": "Reinforcement Learning (RL) applied in healthcare can lead to unsafe medical decisions and treatment, such as excessive dosages or abrupt changes, often due to agents overlooking common-sense constraints. Consequently, Constrained Reinforcement Learning (CRL) is a natural choice for safe decisions. However, specifying the exact cost function is inherently difficult in healthcare. Recent Inverse Constrained Reinforcement Learning (ICRL) is a promising approach that infers constraints from expert demonstrations. ICRL algorithms model Markovian decisions in an interactive environment. These settings do not align with the practical requirement of a decision-making system in healthcare, where decisions rely on historical treatment recorded in an offline dataset. To tackle these issues, we propose the Constraint Transformer (CT). Specifically, 1) we utilize a causal attention mechanism to incorporate historical decisions and observations into the constraint modeling, while employing a Non-Markovian layer for weighted constraints to capture critical states. 2) A generative world model is used to perform exploratory data augmentation, enabling offline RL methods to simulate unsafe decision sequences. In multiple medical scenarios, empirical results demonstrate that CT can capture unsafe states and achieve strategies that approximate lower mortality rates, reducing the occurrence probability of unsafe behaviors.", "keywords": [ "Reinforcement Learning", "Inverse Constrained Reinforcement Learning", "Healthcare" ], "rating_list": [ 5, 6, 3, 3, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 4, 5, 5, 3, 2 ] }, { "paper_id": "HYWdlCPtao", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Ilya Kaufman", "gender": "unknown", "institution": "Ben-Gurion University of the Negev", "country": "IL", "position": "PhD student" }, { "name": "Omri Azencot", "gender": "Unspecified", "institution": "Ben-Gurion University of the Negev", "country": "IL", "position": "Assistant Professor" } ], "n_formula": 23, "n_formula_1": 15, "n_ref_uni": 66, "n_ref": 116, "n_ref_all": 132, "n_fig": 4, "n_tab": 11, "L_tab": 2120, "n_element_tab": 303, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 425, "n_element_tab_1": 48, "formula_len_all": 1444, "formula_len_all_1": 916, "len_all": 195932, "len_all_1": 69325, "len_abs": 1361, "len_title": 84, "len_sents": 58207, "len_sents_1": 32560, "n_sents": 458, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 1370, "title": "Curvature Enhanced Manifold Sampling", "abs": "Over-parameterized deep learning models, characterized by their large number of parameters, have demonstrated remarkable performance in various tasks. Despite the potential risk of overfitting, these models often generalize well to unseen data due to effective regularization techniques, with data augmentation being one of the most prominent methods. This strategy has proven effective in classification tasks, where label-preserving transformations are applicable. However, the application of data augmentation in regression problems remains underexplored. Recently, a new *manifold learning* approach for sampling synthetic data has been introduced, and it can be viewed as utilizing a first-order approximation of the data manifold. In this work, we propose to extend this direction by providing the fundamental theory and practical tools for approximating and sampling general data manifolds. Further, we introduce the curvature enhanced manifold sampling (CEMS) data augmentation method for regression. CEMS is based on a second-order encoding of the manifold, facilitating sampling and reconstruction of new points. Through extensive evaluations on multiple datasets and in comparison to several state-of-the-art approaches, we demonstrate that CEMS is superior in in-distribution and out-of-distribution tasks, while incurring only a mild computational overhead.", "keywords": [ "Manifold learning", "Data augmentation", "Regression" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "HXwrppoSPc", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Dmitry Kangin", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "Postdoc" }, { "name": "Mario Fritz", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Full Professor" }, { "name": "Plamen P Angelov", "gender": "Not Specified", "institution": "Lancaster University", "country": "GB", "position": "Full Professor" }, { "name": "Sarath Sivaprasad", "gender": "Male", "institution": "cispa", "country": "DE", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 18, "n_ref_uni": 32, "n_ref": 49, "n_ref_all": 79, "n_fig": 10, "n_tab": 9, "L_tab": 2619, "n_element_tab": 114, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 3850, "n_element_tab_1": 119, "formula_len_all": 1469, "formula_len_all_1": 1355, "len_all": 135394, "len_all_1": 58879, "len_abs": 1501, "len_title": 98, "len_sents": 33772, "len_sents_1": 23712, "n_sents": 238, "n_sents_1": 167, "venue": "ICLR2025", "decision": "withdraw", "L_title": 50, "L_abs": 1519, "title": "COMiX: Compositional explanations using prototypes", "abs": "Aligning machine representations with human understanding is key to improving interpretability of machine learning (ML) models. \nWhen classifying a new image, humans often explain their decisions by decomposing the image into concepts and pointing to corresponding regions in familiar images.\nCurrent ML explanation techniques typically either trace decision-making processes to reference prototypes, generate attribution maps highlighting feature importance, or incorporate intermediate bottlenecks designed to align with human-interpretable concepts.\nThe proposed method, named COMiX, classifies an image by decomposing it into regions based on learned concepts and tracing each region to corresponding ones in images from the training dataset, assuring that explanations fully represent the actual decision-making process. We dissect the test image into selected internal representations of a neural network to derive prototypical parts (primitives) and match them with the corresponding primitives derived from the training data. \nIn a series of qualitative and quantitative experiments, we theoretically prove and demonstrate that our method, in contrast to \\textit{post hoc} analysis, provides fidelity of explanations and shows that the efficiency is competitive with other inherently interpretable architectures. Notably, it shows substantial improvements in fidelity and sparsity metrics, including $48.82\\%$ improvement in the C-insertion score on the ImageNet dataset over the best state-of-the-art baseline.", "keywords": [ "explanation-by-design", "class-defining-features" ], "rating_list": [ 1, 6, 3, 3 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "HW8xnOUcBx", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Daniel Barath", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Francis Engelmann", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Kai Lascheit", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" }, { "name": "Marc Pollefeys", "gender": "Male", "institution": "Microsoft", "country": "CH", "position": "Director" } ], "n_formula": 8, "n_formula_1": 10, "n_ref_uni": 34, "n_ref": 91, "n_ref_all": 95, "n_fig": 1, "n_tab": 6, "L_tab": 885, "n_element_tab": 106, "n_fig_1": 0, "n_tab_1": 6, "L_tab_1": 1121, "n_element_tab_1": 93, "formula_len_all": 336, "formula_len_all_1": 338, "len_all": 108044, "len_all_1": 61713, "len_abs": 0, "len_title": 132, "len_sents": 32385, "len_sents_1": 28676, "n_sents": 216, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1967, "title": "SegFit: Robust SMPL-X Fitting with Body Part Segmentation on Real-World Point Clouds", "abs": "Fitting parametric human body models to 3D point cloud is crucial for applications such as virtual reality and human-robot interaction but remains challenging due to the lack of contextual guidance, often leading to imprecise results. To address this, we propose a hybrid approach that incorporates body part segmentation into the fitting process, enhancing pose estimation and segmentation accuracy. \nOur method starts with an initial segmentation, assigning each point to a specific body part. This segmentation guides a two-step optimization in fitting an SMPL-X model: first, approximating the initial pose and orientation using body part centroids, and second, refining the model by considering the entire point cloud. After fitting, we reassign body parts to the point cloud through nearest-neighbor matching, resulting in more accurate segmentation. This enhanced segmentation serves as pseudo ground truth to fine-tune the segmentation network in a self-supervised manner, creating a feedback loop where improvements in pose fitting lead to better segmentation and vice versa. We evaluate our approach on four challenging datasets -- PosePrior, EgoBody, BEHAVE, and Hi4D -- demonstrating significant improvements over leading methods, including a tenfold increase in pose modeling accuracy and a 15\\% enhancement in segmentation accuracy after fine-tuning. Our contributions are twofold: (1) introducing a novel hybrid method that unifies pose fitting and body part segmentation on point clouds, enabling mutual enhancement through iterative refinement; and (2) developing a self-supervised technique for fine-tuning segmentation networks using pseudo ground truths derived from fitted models. This work advances the state of the art in human body fitting to point clouds, facilitating more accurate human representations in complex environments and benefiting applications that require precise human modeling. We will make the source code publicly available.", "keywords": [ "Multi Human Body Part Segmentation", "Human Pose Estimation" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "HVtu26XDAA", "primary_area": "foundation or frontier models, including LLMs", "n_author": 23, "author_info_list": [ { "name": "Afshin Dehghan", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Aleksei Timofeev", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Bowen Zhang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Research Scientist" }, { "name": "Dhruti Shah", "gender": "unknown", "institution": "Apple", "country": "CH", "position": "Researcher" }, { "name": "Forrest Huang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Research Scientist" }, { "name": "Haotian Zhang", "gender": "Male", "institution": "Apple AI/ML", "country": "US", "position": "Researcher" }, { "name": "Haoxuan You", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Hong-You Chen", "gender": "unknown", "institution": "Apple AI/ML", "country": "US", "position": "Researcher" }, { "name": "Jean-Philippe Fauconnier", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Keen You", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Mingfei Gao", "gender": "unknown", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Mingze Xu", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Nina Wenzel", "gender": "Female", "institution": "Apple", "country": "DE", "position": "Researcher" }, { "name": "Peter Grasch", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Philipp Dufter", "gender": "unknown", "institution": "Apple", "country": "CH", "position": "Machine Learning Engineer" }, { "name": "Sam Dodge", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Xianzhi Du", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "Yanghao Li", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Yinfei Yang", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Zhe Gan", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Zhen Yang", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Zhengfeng Lai", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Zirui Wang", "gender": "Male", "institution": "Apple AI/ML", "country": "US", "position": "Research Scientist" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 95, "n_ref_all": 144, "n_fig": 10, "n_tab": 17, "L_tab": 10921, "n_element_tab": 1521, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 1276, "n_element_tab_1": 158, "formula_len_all": 140, "formula_len_all_1": 0, "len_all": 405187, "len_all_1": 68352, "len_abs": 5421, "len_title": 115, "len_sents": 75389, "len_sents_1": 31516, "n_sents": 677, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1185, "title": "MM1.5: Methods, Analysis & Insights from Multimodal LLM Fine-tuning", "abs": "We present MM1.5, a new family of multimodal large language models (MLLMs) designed to enhance capabilities in text-rich image understanding, visual referring and grounding, and multi-image reasoning. Building upon the MM1 architecture, MM1.5 adopts a data-centric approach to model training, systematically exploring the impact of diverse data mixtures across the entire model training lifecycle. This includes high-quality OCR data and synthetic captions for continual pre-training, as well as an optimized visual instruction-tuning data mixture for supervised fine-tuning. Our models range from 1B to 30B parameters, encompassing both dense and mixture-of-experts (MoE) variants, and demonstrate that careful data curation and training strategies can yield strong performance even at small scales (1B and 3B). Additionally, we introduce two specialized variants: MM1.5-Video, designed for video understanding, and MM1.5-UI, tailored for mobile UI understanding. Through extensive empirical studies and ablations, we provide detailed insights into the training processes and decisions that inform our final designs, offering valuable guidance for future research in MLLM development.", "keywords": [ "Multimodal LLM" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 5, 5, 3 ] }, { "paper_id": "HVblmL5Rws", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Mikita Balesni", "gender": "Male", "institution": "Apollo Research", "country": "GB", "position": "Researcher" }, { "name": "Owain Evans", "gender": "unknown", "institution": "Truthful AI", "country": "US", "position": "Principal Researcher" }, { "name": "Tomasz Korbak", "gender": "Male", "institution": "UK AI Safety Institute", "country": "GB", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 34, "n_ref": 56, "n_ref_all": 73, "n_fig": 5, "n_tab": 2, "L_tab": 120, "n_element_tab": 14, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 317, "n_element_tab_1": 16, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 132306, "len_all_1": 70829, "len_abs": 1750, "len_title": 112, "len_sents": 29980, "len_sents_1": 29435, "n_sents": 218, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1504, "title": "The Two-Hop Curse: LLMs trained on A\u2192B, B\u2192C fail to learn A\u2192C", "abs": "While LLMs excel at answering multi-hop questions like \u201cWho is the spouse of the performer of Imagine?\u201d by thinking out loud (chain-of-thought), they perform surprisingly poorly when required to reason in their latent space and answer without chain-of-thought. This observation was previously referred to as the compositionality gap, implying that although language models are less reliable at two-hop latent reasoning, they still perform it sometimes. In this paper, we introduce a controlled setting for investigating the compositionality gap. We run a series of experiments finetuning a large language model (Llama-3-8B-Instruct) on synthetic facts expressed in English. We attempt to elicit two-hop reasoning in three ways: (i) fine-tune on a data mixture designed to incentivize two-hop reasoning, (ii) force facts to be stored in layers in the correct order, and (iii) use an auxiliary loss to provide activation-level supervision for two-hop reasoning. We show that LLaMA 3 8B successfully learns to answer two-hop questions about synthetic facts using CoT, but completely fails without CoT, achieving chance-level accuracy and chance-level test loss. Failures of LLMs in our controlled setting cast doubt on the purported ability of present LLMs to perform multihop latent reasoning and lead us to conjecture that, rather than a reasoning gap, current language models might exhibit a two-hop reasoning curse \u2014 a complete lack of ability rather than a relative weakness. This is the Two-Hop Curse.", "keywords": [ "latent reasoning", "two-hop reasoning", "chain of thought", "LLMs", "question answering", "llama", "fine-tuning", "fact representation", "knowledge representation", "world models" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "HVY6qL2J9L", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Changde Du", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Huiguang He", "gender": "Male", "institution": "Institute of Automation Chinese Academy of Sciences", "country": "", "position": "Full Professor" }, { "name": "Jie Peng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Kaicheng Fu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 45, "n_ref": 71, "n_ref_all": 89, "n_fig": 4, "n_tab": 5, "L_tab": 3322, "n_element_tab": 426, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2207, "n_element_tab_1": 244, "formula_len_all": 679, "formula_len_all_1": 680, "len_all": 129892, "len_all_1": 60041, "len_abs": 1470, "len_title": 144, "len_sents": 32355, "len_sents_1": 23878, "n_sents": 226, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1475, "title": "BraiNav: Incorporating Human Brain Activity to Enhance Robustness in Embodied Visual Navigation", "abs": "Recent research shows that standard navigation agents significantly underperform and even fail in the presence of various visual corruptions. Unlike embodied agents, the human brain's visual system can robustly perceive the environment and extract the necessary information to complete the visual tasks. In this paper, we propose a two-phase Brain-Machine integration Navigation method called BraiNav, which incorporates neural representations derived from human brain activity to enhance robustness against visual corruptions. In the first phase, a brain encoder, built upon a recently advanced self-supervised pretrained model, is trained on a large-scale human brain activity dataset and then frozen for downstream visual navigation. In the second phase, neural representations harboring high-level cognitive information from the human brain are constructed based on the pretrained frozen brain encoder. Additionally, we propose a multimodal fusion method based on cross-attention to obtain more consistent brain-visual joint representations, which are then used to learn the navigation policy. Sufficient experiments demonstrate that the proposed method exhibits higher robustness against various visual corruptions compared to standard navigation agent and multiple computer vision-enhanced agents. Our study pioneers the incorporation of human brain activity into embodied AI, aiming to catalyze further cross-disciplinary collaboration with computational neuroscience.", "keywords": [ "embodied visual navigation", "neural encoding", "multimodal learning" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 5, 5, 2 ] }, { "paper_id": "HVJBFYJrN2", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Babak Taati", "gender": "Male", "institution": "Kite Research Institute, Toronto Rehab, University Health Network", "country": "CA", "position": "Principal Researcher" }, { "name": "Mohammad Javad Rajabi", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Soroush Mehraban", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 24, "n_ref": 47, "n_ref_all": 62, "n_fig": 7, "n_tab": 13, "L_tab": 4995, "n_element_tab": 572, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 3205, "n_element_tab_1": 388, "formula_len_all": 447, "formula_len_all_1": 424, "len_all": 140465, "len_all_1": 62270, "len_abs": 1182, "len_title": 125, "len_sents": 32410, "len_sents_1": 25335, "n_sents": 252, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1158, "title": "STARS: Self-supervised Tuning for 3D Action Recognition in Skeleton Sequences", "abs": "Self-supervised pretraining methods with masked prediction demonstrate remarkable within-dataset performance in skeleton-based action recognition. However, we show that, unlike contrastive learning approaches, they do not produce well-separated clusters. Additionally, these methods struggle with generalization in few-shot settings. To address these issues, we propose Self-supervised Tuning for 3D Action Recognition in Skeleton sequences (STARS). Specifically, STARS first uses a masked prediction stage using an encoder-decoder architecture. It then employs nearest-neighbor contrastive learning to partially tune the weights of the encoder, enhancing the formation of semantic clusters for different actions. By tuning the encoder for a few epochs, and without using hand-crafted data augmentations, STARS achieves state-of-the-art self-supervised results in various benchmarks, including NTU-60, NTU-120, and PKU-MMD. In addition, STARS exhibits significantly better results than masked prediction models in few-shot settings, where the model has not seen the actions throughout pretraining. Code: https://anonymous.4open.science/r/stars-CD2E/README.md", "keywords": [ "Action Recognition", "Self-supervised Learning", "Contrastive Tuning" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "HVFMooKrHX", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Joshua Kazdan", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Rachid Guerraoui", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "", "position": "Full Professor" }, { "name": "Youssef Allouah", "gender": "Male", "institution": "EPFL - Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" } ], "n_formula": 132, "n_formula_1": 22, "n_ref_uni": 37, "n_ref": 108, "n_ref_all": 132, "n_fig": 4, "n_tab": 3, "L_tab": 154, "n_element_tab": 8, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 677, "n_element_tab_1": 28, "formula_len_all": 14126, "formula_len_all_1": 1904, "len_all": 198048, "len_all_1": 77655, "len_abs": 3932, "len_title": 123, "len_sents": 58280, "len_sents_1": 33811, "n_sents": 480, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1210, "title": "The Utility and Complexity of In- and Out-of-Distribution Machine Unlearning", "abs": "Machine unlearning, the process of selectively removing data from trained models, is increasingly crucial for addressing privacy concerns and knowledge gaps post-deployment. Despite this importance, existing approaches are often heuristic and lack formal guarantees. In this paper, we analyze the fundamental utility, time, and space complexity trade-offs of approximate unlearning, providing rigorous certification analogous to differential privacy. For in-distribution forget data\u2014data similar to the retain set\u2014we show that a surprisingly simple and general procedure, empirical risk minimization with output perturbation, achieves tight unlearning-utility-complexity trade-offs, addressing a previous theoretical gap on the separation from unlearning ``for free\" via differential privacy, which inherently facilitates the removal of such data. However, such techniques fail with out-of-distribution forget data\u2014data significantly different from the retain set\u2014where unlearning time complexity can exceed that of retraining, even for a single sample. To address this, we propose a new robust and noisy gradient descent variant that provably amortizes unlearning time complexity without compromising utility.", "keywords": [ "machine unlearning", "differential privacy", "optimization", "theory" ], "rating_list": [ 5, 8, 6, 6, 8 ], "soundness_list": [ 3, 4, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 2, 4 ], "contribution_list": [ 3, 4, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 4, 3 ] }, { "paper_id": "HV67MnnXkL", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Byung Suk Lee", "gender": "Male", "institution": "University of Vermont", "country": "US", "position": "Full Professor" }, { "name": "Hwanjun Song", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Jae-Gil Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Jihye Na", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Susik Yoon", "gender": "unknown", "institution": "Korea University", "country": "KR", "position": "Assistant Professor" }, { "name": "Youngeun Nam", "gender": "Female", "institution": "KAIST", "country": "KR", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 28, "n_ref": 54, "n_ref_all": 64, "n_fig": 4, "n_tab": 5, "L_tab": 4301, "n_element_tab": 263, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 4301, "n_element_tab_1": 263, "formula_len_all": 169, "formula_len_all_1": 169, "len_all": 96470, "len_all_1": 59996, "len_abs": 858, "len_title": 109, "len_sents": 31002, "len_sents_1": 26053, "n_sents": 201, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 880, "title": "Mobility Networked Time-Series Forecasting Benchmark Datasets", "abs": "Human mobility is crucial for urban planning (e.g., public transportation) and epidemic response strategies. However, existing research often neglects integrating comprehensive perspectives on spatial dynamics, temporal trends, and other contextual views due to the limitations of existing mobility datasets. To bridge this gap, we introduce **MOBINS** (**MOBI**lity **N**etworked time **S**eries), a novel dataset collection designed for networked time-series forecasting of dynamic human movements. **MOBINS** features diverse and explainable datasets that capture various mobility patterns across different transportation modes in four cities and two countries and cover both transportation and epidemic domains at the administrative area level. Our experiments with nine baseline methods reveal the significant impact of different model backbones on the proposed six datasets.", "keywords": [ "mobility", "networked time series", "time series", "origin-destination", "forecasting", "prediction", "transportation", "epidemic" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "HUzDU7u5B4", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Ben He", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Hongyu Lin", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Le Sun", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xianpei Han", "gender": "Male", "institution": "Institute of Software, CAS", "country": "CN", "position": "Professor" }, { "name": "Xinyan Guan", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Xinyu Lu", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Yaojie Lu", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "jie lou", "gender": "Male", "institution": "Xiaohongshu Inc", "country": "CN", "position": "Researcher" }, { "name": "Xueruwen", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 44, "n_ref": 73, "n_ref_all": 94, "n_fig": 9, "n_tab": 4, "L_tab": 1117, "n_element_tab": 91, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 338, "n_element_tab_1": 46, "formula_len_all": 170, "formula_len_all_1": 170, "len_all": 152585, "len_all_1": 61765, "len_abs": 1534, "len_title": 118, "len_sents": 39551, "len_sents_1": 28696, "n_sents": 330, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1557, "title": "On-Policy Fine-grained Knowledge Feedback for Hallucination Mitigation", "abs": "Hallucination occurs when large language models (LLMs) exhibit behavior that deviates from the boundaries of their knowledge during the response generation process.\nPrevious learning-based methods focus on detecting knowledge boundaries and finetuning models with instance-level feedback, but they suffer from inaccurate signals due to off-policy data sampling and coarse-grained feedback.\nIn this paper, we introduce \\textit{\\b{R}einforcement \\b{L}earning \\b{f}or \\b{H}allucination} (RLFH), a fine-grained feedback-based online reinforcement learning method for hallucination mitigation.\nUnlike previous learning-based methods, RLFH enables LLMs to explore the boundaries of their internal knowledge and provide on-policy, fine-grained feedback on these explorations.\nTo construct fine-grained feedback for learning reliable generation behavior, RLFH decomposes the outcomes of large models into atomic facts, provides statement-level evaluation signals, and traces back the signals to the tokens of the original responses.\nFinally, RLFH adopts the online reinforcement algorithm with these token-level rewards to adjust model behavior for hallucination mitigation.\nFor effective on-policy optimization, RLFH also introduces an LLM-based fact assessment framework to verify the truthfulness and helpfulness of atomic facts without human intervention.\nExperiments on HotpotQA, SQuADv2, and Biography benchmarks demonstrate that RLFH can balance their usage of internal knowledge during the generation process to eliminate the hallucination behavior of LLMs.", "keywords": [ "Hallucination;Large Language Model;" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "HUjFpOgVCK", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "A. Sophia Koepke", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Junior Group Leader" }, { "name": "Shentong Mo", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Yutong Bai", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 62, "n_ref": 166, "n_ref_all": 204, "n_fig": 1, "n_tab": 39, "L_tab": 7687, "n_element_tab": 704, "n_fig_1": 1, "n_tab_1": 12, "L_tab_1": 4144, "n_element_tab_1": 310, "formula_len_all": 80, "formula_len_all_1": 80, "len_all": 298096, "len_all_1": 70160, "len_abs": 1242, "len_title": 129, "len_sents": 94073, "len_sents_1": 30610, "n_sents": 804, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1249, "title": "ACAV-1M: Data Curation and Benchmarking for Audio-Visual Representation Learning", "abs": "The natural alignment of visual and audio information in videos provides a strong learning signal. However, commonly used large-scale video datasets contain audio-visual signals that are not aligned, e.g. background music. This limits the development of robust models that leverage the complementary nature of audio and video data. To address this limitation, we curate ACAV-1M, a new large-scale dataset that contains one million samples sourced from the ACAV-100M dataset. The ACAV-1M dataset is obtained through a pipeline that ensures the audio-visual correspondence and synchronization of samples in the dataset. Our pipeline transforms raw video and audio into text captions, followed by text summarization and an extensive filtering procedure. The filtering is done based on audio-caption alignment, audio-visual instance semantic alignment, and temporal synchronization. Furthermore, we propose an audio-visual learning benchmark that supports a diverse range of downstream tasks. Empirical evaluations demonstrate that models trained on ACAV-1M achieve superior performance compared to using existing datasets across all tasks. Our ACAV-1M dataset and code to reproduce all benchmark results will be made publicly available upon acceptance.", "keywords": [ "audio-visual learning", "sound souce localization", "audio-visual video parsing" ], "rating_list": [ 3, 6, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2, 1 ], "confidence_list": [ 4, 3, 5, 4, 5 ] }, { "paper_id": "HTpyexVwlI", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Chang D. Yoo", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Donghoon Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Tung Minh Luu", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Younghwan Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 69, "n_ref": 143, "n_ref_all": 159, "n_fig": 8, "n_tab": 1, "L_tab": 154, "n_element_tab": 22, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 139, "formula_len_all_1": 139, "len_all": 154092, "len_all_1": 65634, "len_abs": 1103, "len_title": 115, "len_sents": 39757, "len_sents_1": 33077, "n_sents": 289, "n_sents_1": 238, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1107, "title": "VISION-LANGUAGE MODELS AS TRAINERS FOR INSTRUCTION-FOLLOWING AGENTS", "abs": "Developing agents that can understand and follow language instructions is critical for effective and reliable human-AI collaboration. Recent approaches train these agents using reinforcement learning with infrequent environment rewards, placing a significant burden on environment designers to create language-conditioned reward functions. As environments and instructions grow in complexity, crafting such reward functions becomes increasingly impractical. To address this challenge, we introduce V-TIFA, a novel method that trains instruction-following agents by leveraging feedback from vision-language models (VLMs). The core idea of V-TIFA is to query VLMs to rate entire trajectories based on language instructions, using the resulting ratings to directly train the agent. Unlike prior VLM reward generation methods, V-TIFA does not require manually crafted task specifications, enabling agents to learn from a diverse set of natural language instructions. Extensive experiments in embodied environments demonstrate that V-TIFA outperforms existing reward generation methods under the same conditions.", "keywords": [ "Language-conditioned Reinforcement Learning", "Reward Generation", "Vision Language Foundation Models" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "HTjJpwY5AU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Daniel Barath", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Marc Pollefeys", "gender": "Male", "institution": "Microsoft", "country": "CH", "position": "Director" }, { "name": "Nicole Damblon", "gender": "Female", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 57, "n_ref": 132, "n_ref_all": 152, "n_fig": 6, "n_tab": 10, "L_tab": 7321, "n_element_tab": 1318, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1129, "n_element_tab_1": 175, "formula_len_all": 688, "formula_len_all_1": 796, "len_all": 185056, "len_all_1": 64544, "len_abs": 1201, "len_title": 112, "len_sents": 49899, "len_sents_1": 30277, "n_sents": 365, "n_sents_1": 217, "venue": "ICLR2025", "decision": "withdraw", "L_title": 46, "L_abs": 1201, "title": "Learning to Filter Outlier Edges in Global SfM", "abs": "This paper introduces a novel approach to improve camera position estimation in global Structure-from-Motion (SfM) frameworks by filtering inaccurate pose graph edges, representing relative translation estimates, before applying translation averaging. In SfM, pose graph vertices represent cameras and edges relative poses (rotation and translation) between cameras. We formulate the edge filtering problem as a vertex filtering in the dual graph -- a line graph where the vertices stem from edges in the original graph, and the edges from cameras. Exploiting such a representation, we frame the problem as a binary classification over nodes in the dual graph. To learn such a classification and find outlier edges, we employ a Transformer architecture-based technique. To address the challenge of memory overflow often caused by converting to a line graph, we introduce a clustering-based graph processing approach, enabling the application of our method to arbitrarily large pose graphs. The proposed method outperforms existing relative translation filtering techniques in terms of final camera position accuracy and can be seamlessly integrated with any other filter. The code will be made public.", "keywords": [ "3D reconstruction", "Structure-from-Motion", "Translation Averaging" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "HSi4VetQLj", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Kajetan Schweighofer", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "", "position": "PhD student" }, { "name": "Lukas Aichberger", "gender": "Male", "institution": "Department of Computer Science, University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Mykyta Ielanskyi", "gender": "Not Specified", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "PhD student" }, { "name": "Sepp Hochreiter", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "Full Professor" } ], "n_formula": 31, "n_formula_1": 18, "n_ref_uni": 40, "n_ref": 96, "n_ref_all": 154, "n_fig": 13, "n_tab": 3, "L_tab": 1498, "n_element_tab": 145, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 500, "n_element_tab_1": 96, "formula_len_all": 1653, "formula_len_all_1": 983, "len_all": 193889, "len_all_1": 73637, "len_abs": 1064, "len_title": 129, "len_sents": 57255, "len_sents_1": 35254, "n_sents": 439, "n_sents_1": 270, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1070, "title": "Improving Uncertainty Estimation through Semantically Diverse Language Generation", "abs": "Large language models (LLMs) can suffer from hallucinations when generating text. These hallucinations impede various applications in society and industry by making LLMs untrustworthy. Current LLMs generate text in an autoregressive fashion by predicting and appending text tokens. When an LLM is uncertain about the semantic meaning of the next tokens to generate, it is likely to start hallucinating. Thus, it has been suggested that predictive uncertainty is one of the main causes of hallucinations. We introduce Semantically Diverse Language\nGeneration (SDLG) to quantify predictive uncertainty in LLMs. SDLG steers the LLM to generate semantically diverse yet likely alternatives for an initially generated text. This approach provides a precise measure of aleatoric semantic uncertainty, detecting whether the initial text is likely to be hallucinated. Experiments on question-answering tasks demonstrate that SDLG consistently outperforms existing methods while being the most computationally efficient, setting a new standard for uncertainty estimation in LLMs.", "keywords": [ "llm", "nlg", "uncertainty estimation", "uncertainty measures", "semantic uncertainty", "aleatoric uncertainty", "semantic entropy", "mc estimation", "importance sampling" ], "rating_list": [ 5, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "HSLClc1a7W", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Fengda Zhang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Jun Xiao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Kai Huang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" }, { "name": "Kaifeng Gao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Pengfei Lyu", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "MS student" }, { "name": "Yunze Tong", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zihao TANG", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 21, "n_formula_1": 19, "n_ref_uni": 39, "n_ref": 76, "n_ref_all": 130, "n_fig": 18, "n_tab": 10, "L_tab": 3110, "n_element_tab": 464, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1113, "n_element_tab_1": 48, "formula_len_all": 1218, "formula_len_all_1": 901, "len_all": 172677, "len_all_1": 71080, "len_abs": 1420, "len_title": 131, "len_sents": 56695, "len_sents_1": 32847, "n_sents": 465, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1427, "title": "Latent Score-Based Reweighting for Robust Classification on Imbalanced Tabular Data", "abs": "Machine learning models often perform well on tabular data by optimizing average prediction accuracy. However, they may underperform on specific subsets due to inherent biases and spurious correlations in the training data, such as associations with non-causal features like demographic information. These biases lead to critical robustness issues as models may inherit or amplify them, resulting in poor performance where such misleading correlations do not hold. Existing mitigation methods have significant limitations: some require prior group labels, which are often unavailable, while others focus solely on the conditional distribution $P(Y|X)$, upweighting misclassified samples without effectively balancing the overall data distribution $P(X)$. To address these shortcomings, we propose a latent score-based reweighting framework. It leverages score-based models to capture the joint data distribution $P(X, Y)$ without relying on additional prior information. By estimating sample density through the similarity of score vectors with neighboring data points, our method identifies underrepresented regions and upweights samples accordingly. This approach directly tackles inherent data imbalances, enhancing robustness by ensuring a more uniform dataset representation. Experiments on various tabular datasets under distribution shifts demonstrate that our method effectively improves performance on imbalanced data.", "keywords": [ "robustness", "score model", "reweighting" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 2, 2, 3, 4 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "HSGCCUwH7r", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 12, "author_info_list": [ { "name": "Achin Kulshrestha", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Chen-Yu Lee", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Hamid Palangi", "gender": "Male", "institution": "Google", "country": "US", "position": "Staff Research Scientist" }, { "name": "Lesly Miculicich Werlen", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Nathalie Rauschmayr", "gender": "unknown", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Sayna Ebrahimi", "gender": "Female", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Shangbin Feng", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" }, { "name": "Tomas Pfister", "gender": "Male", "institution": "Google", "country": "", "position": "Head of Research @ Cloud AI" }, { "name": "Yejin Choi", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Yike Wang", "gender": "Female", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Yulia Tsvetkov", "gender": "Female", "institution": "Department of Computer Science, University of Washington", "country": "", "position": "Associate Professor" }, { "name": "Zifeng Wang", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" } ], "n_formula": 11, "n_formula_1": 6, "n_ref_uni": 76, "n_ref": 150, "n_ref_all": 193, "n_fig": 15, "n_tab": 14, "L_tab": 3628, "n_element_tab": 759, "n_fig_1": 9, "n_tab_1": 7, "L_tab_1": 4859, "n_element_tab_1": 319, "formula_len_all": 1363, "formula_len_all_1": 1020, "len_all": 315677, "len_all_1": 71232, "len_abs": 1364, "len_title": 126, "len_sents": 83532, "len_sents_1": 29123, "n_sents": 562, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1155, "title": "Model Swarms: Collaborative Search to Adapt LLM Experts via Swarm Intelligence", "abs": "We propose Model Swarms, a collaborative search algorithm to adapt LLMs via swarm intelligence, the collective behavior guiding individual systems. Specifically, Model Swarms starts with a pool of LLM experts and a utility function. Guided by the best-found checkpoints across models, diverse LLM experts collaboratively move in the weight space and optimize a utility function representing model adaptation objectives. Compared to existing model composition approaches, Model Swarms offers tuning-free model adaptation, works in low-data regimes with as few as 200 examples, and does not require assumptions about specific experts in the swarm or how they should be composed. Extensive experiments demonstrate that Model Swarms could flexibly adapt LLM experts to a single task, multi-task domains, reward models, as well as diverse human interests, improving over 12 model composition baselines by up to 21.0% across tasks and contexts. Further analysis reveals that LLM experts discover previously unseen capabilities in initial checkpoints and that Model Swarms enable the weak-to-strong transition of experts through the collaborative search process.", "keywords": [ "evolutionary algorithm", "model adaptation", "model merging" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 4, 4, 3, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "HR1ujVR0ig", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Bin Yang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Chenjuan Guo", "gender": "unknown", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Sicong Liu", "gender": "Male", "institution": "East China Normal University", "country": "", "position": "PhD student" }, { "name": "Yang Shu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 39, "n_formula_1": 17, "n_ref_uni": 63, "n_ref": 107, "n_ref_all": 131, "n_fig": 5, "n_tab": 11, "L_tab": 7180, "n_element_tab": 871, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 216, "n_element_tab_1": 4, "formula_len_all": 4596, "formula_len_all_1": 1435, "len_all": 210085, "len_all_1": 67071, "len_abs": 1543, "len_title": 134, "len_sents": 48162, "len_sents_1": 29261, "n_sents": 402, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1551, "title": "Learning Generalizable Skills from Offline Multi-Task Data for Multi-Agent Cooperation", "abs": "Learning cooperative multi-agent policy from offline multi-task data that can generalize to unseen tasks with varying numbers of agents and targets is an attractive problem in many scenarios.\nAlthough aggregating general behavior patterns among multiple tasks as skills to improve policy transfer is a promising approach,\ntwo primary challenges hinder the further advancement of skill learning in offline multi-task MARL. \nFirstly, extracting general cooperative behaviors from various action sequences as common skills lack bringing cooperative temporal knowledge into them.\nSecondly, existing works only involve common skills and can not adaptively choose independent knowledge as task-specific skills in each task for fine-grained action execution.\nTo address these challenges, we propose an approach named Hierarchical and Separate Skill Discovering (HiSSD) for generalizable offline multi-task MARL through skill learning.\nHiSSD leverages a hierarchical framework that jointly learns common and task-specific skills.\nThe common skills learn cooperative temporal knowledge and enable in-sample exploration for offline multi-task MARL.\nThe task-specific skills represent the priors of each task and achieve a task-guided fine-grained action execution.\nTo verify the advancement of our method, we conduct experiments on multi-agent MuJoCo and SMAC benchmarks.\nAfter training policy using HiSSD on offline multi-task data, the empirical results show that HiSSD assigns effective cooperative behaviors and obtains superior performance in unseen tasks.", "keywords": [ "Reinforcement Learning", "Multi-Agent Reinforcemenr Learning" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 4, 2, 3 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 4, 4 ] }, { "paper_id": "HQHnhVQznF", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Gagandeep Singh", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Isha Chaudhary", "gender": "Female", "institution": "Department of Computer Science", "country": "US", "position": "PhD student" }, { "name": "Manoj Kumar", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Morteza Ziyadi", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Qian Hu", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Rahul Gupta", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 71, "n_ref": 140, "n_ref_all": 172, "n_fig": 21, "n_tab": 2, "L_tab": 54, "n_element_tab": 2, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 990, "formula_len_all_1": 821, "len_all": 235029, "len_all_1": 75975, "len_abs": 1337, "len_title": 122, "len_sents": 65975, "len_sents_1": 37934, "n_sents": 505, "n_sents_1": 295, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1270, "title": "Quantitative Certification of Bias in Large Language Models", "abs": "Large Language Models (LLMs) can produce biased responses that can cause representational harms. However, conventional studies are insufficient to thoroughly\nevaluate LLM biases across multiple generations for different demographic groups\n(a.k.a. counterfactual bias), as they do not scale to large number of inputs and do not\nprovide guarantees. Therefore, we propose the first framework, QCB (Quantitative\nCertification of Bias) that certifies LLMs for counterfactual bias on distributions\nof prompts. A certificate consists of high-confidence bounds on the probability of\nunbiased LLM responses for any set of counterfactual prompts mentioning various\ndemographic groups, sampled from a distribution. We illustrate counterfactual bias\ncertification for distributions of counterfactual prompts created by applying varying\nprefixes drawn from prefix distributions, to a given set of prompts. We consider\nprefix distributions for random token sequences, mixtures of manual jailbreaks,\nand jailbreaks in the LLM\u2019s embedding space to certify bias. We obtain non-trivial\ncertified bounds on the probability of unbiased responses of SOTA LLMs, exposing\ntheir vulnerabilities over distributions of prompts generated from computationally\ninexpensive distributions of prefixes.", "keywords": [ "Large Language Models", "Bias", "Certification" ], "rating_list": [ 3, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "HPuLU6q7xq", "primary_area": "generative models", "n_author": 1, "author_info_list": [ { "name": "Huang Yuxuan", "gender": "Male", "institution": "Jilin University", "country": "", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 37, "n_ref": 53, "n_ref_all": 65, "n_fig": 2, "n_tab": 7, "L_tab": 3235, "n_element_tab": 202, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1318, "n_element_tab_1": 103, "formula_len_all": 39, "formula_len_all_1": 39, "len_all": 149758, "len_all_1": 64139, "len_abs": 1335, "len_title": 146, "len_sents": 38764, "len_sents_1": 29736, "n_sents": 278, "n_sents_1": 226, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1395, "title": "Orca: Enhancing Role-Playing Abilities of Large Language Models by Integrating Personality Traits", "abs": "Large language models have catalyzed the development of personalized dialogue systems, leading to the emergence of numerous role-playing conversational agents. While previous research has predominantly focused on enhancing the model's capability to follow instructions by designing character profiles, it has neglected the psychological factors that drive human conversations. In this paper, we propose Orca, a framework for data processing and training custom LLM characters by integrating personality traits. Orca comprises four stages: (1) Personality traits inference, leveraging LLMs to infer users' BigFive personality trait reports and scores. (2) Data Augmentation, simulating users' profiles, background stories, and psychological activities. (3) Dataset construction, employing personality-conditioned instruction prompting (PCIP) to stimulate LLMs. (4) Modeling and Training, utilizing personality-conditioned instruction tuning (PTIT and PSIT) to enhance existing open-source LLMs using the generated data. We introduce OrcaBench, the first benchmark for evaluating the quality of LLM-generated content on social platforms across multiple scales. Our experiments demonstrate that our proposed model achieves superior performance on this benchmark, highlighting its excellence and effectiveness in perceiving personality traits, thereby significantly improving role-playing abilities.", "keywords": [ "Large Language Models", "Role-Playing", "BigFive", "Personality Traits" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 3, 1 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "HPcpLDJlS6", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Mingwei Xu", "gender": "Male", "institution": "Shanghai University of Finance and Economics", "country": "CN", "position": "MS student" }, { "name": "Shilong Tu", "gender": "Male", "institution": "Shanghai University of Finance and Economics", "country": "CN", "position": "MS student" }, { "name": "Wanyun Cui", "gender": "Male", "institution": "Shanghai University of Finance and Economics", "country": "CN", "position": "Associate Professor" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 24, "n_ref": 45, "n_ref_all": 52, "n_fig": 3, "n_tab": 3, "L_tab": 1679, "n_element_tab": 262, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 294, "n_element_tab_1": 4, "formula_len_all": 678, "formula_len_all_1": 749, "len_all": 95332, "len_all_1": 57196, "len_abs": 828, "len_title": 130, "len_sents": 26819, "len_sents_1": 24479, "n_sents": 258, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 833, "title": "Stated Causal Language Modeling: Off-the-Shelf Enhancement of Context Memorization", "abs": "We propose stated causal language modeling (stated-CLM), a novel method to enhance the memory capacity of large language models (LLMs) without modifying their architecture or parameters. Unlike existing context segmentation and sliding methods that discard low-weight tokens, stated-CLM compresses adjacent tokens, significantly reducing context information loss. We utilize the classic network pruning techniques with second-order derivatives to optimize the compressed token in the differentiable key-value space. Experiments on LLaMA, Mistral, and Gemma demonstrate that stated-CLM outperforms baselines on the LongBench benchmark by an average of 6.12\\% (LLaMA3.1-8B) and 5.97\\% (Mistral-v0.3-7B). On TopicRet, stated-CLM achieves accuracy levels comparable to full context models, while the baselines' accuracy is close to zero.", "keywords": [ "Memory-Enhanced Causal Language Modeling", "Training-Free Approach", "Context Compression", "Language Models", "Attention" ], "rating_list": [ 5, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "HPSAkIHRbb", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Chien-Sheng Wu", "gender": "Male", "institution": "Salesforce AI", "country": "US", "position": "Researcher" }, { "name": "Divyansh Agarwal", "gender": "Male", "institution": "Salesforce.com", "country": "", "position": "Researcher" }, { "name": "Fan Yin", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Linnea Ross", "gender": "Not Specified", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Philippe Laban", "gender": "Male", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "Vaibhav Vats", "gender": "Male", "institution": "Salesforce Inc.", "country": "US", "position": "Data Scientist" }, { "name": "XIANGYU PENG", "gender": "Female", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Yilun Zhou", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Yixin Mao", "gender": "unknown", "institution": "SalesForce.com", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 35, "n_ref": 98, "n_ref_all": 123, "n_fig": 5, "n_tab": 9, "L_tab": 3821, "n_element_tab": 331, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1073, "n_element_tab_1": 119, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 248642, "len_all_1": 72471, "len_abs": 1607, "len_title": 105, "len_sents": 94005, "len_sents_1": 35107, "n_sents": 746, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 57, "L_abs": 1617, "title": "BingoGuard: LLM Content Moderation Tools with Risk Levels", "abs": "Malicious content generated by large language models (LLMs) can pose varying degrees of harm. \nAlthough existing LLM-based moderators can detect harmful content, they struggle to assess risk levels and may miss lower-risk outputs. \nAccurate risk assessment allows platforms with different safety thresholds to tailor content filtering and rejection. In this paper, we introduce per-topic severity rubrics for 11 harmful topics and build BingoGuard, an LLM-based moderation system designed to predict both binary safety labels and severity levels. \nTo address the lack of annotations on levels of severity, we propose a scalable generate-then-filter framework that first generates responses across different severity levels and then filters out low-quality responses. Using this framework, we create BingoGuardTrain, a training dataset with 54,897 examples covering a variety of topics, response severity, styles, and BingoGuardTest, a test set with 988 examples explicitly labeled based on our severity rubrics that enables fine-grained analysis on model behaviors on different severity levels. Our BingoGuard-8B, trained on BingoGuardTrain, achieves the state-of-the-art performance on several moderation benchmarks, including WildGuardTest and HarmBench, as well as BingoGuardTest, outperforming best public models, WildGuard, by 4.3\\%. Our analysis demonstrates that incorporating severity levels into training significantly enhances detection performance and enables the model to effectively gauge the severity of harmful responses. Warning: this paper includes red-teaming examples that may be harmful in nature.", "keywords": [ "LLM", "safety guardrail", "content moderator" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "HNbNcDsgxL", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Cheng-Peng Huang", "gender": "Male", "institution": "National Taiwan University of Science and Technology", "country": "TW", "position": "MS student" }, { "name": "Hao-Yuan Chen", "gender": "Male", "institution": "University of London", "country": "US", "position": "Undergrad student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 15, "n_ref": 26, "n_ref_all": 35, "n_fig": 2, "n_tab": 2, "L_tab": 575, "n_element_tab": 96, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 575, "n_element_tab_1": 96, "formula_len_all": 468, "formula_len_all_1": 468, "len_all": 106354, "len_all_1": 72274, "len_abs": 1349, "len_title": 130, "len_sents": 25429, "len_sents_1": 25433, "n_sents": 176, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1357, "title": "Delta - Contrastive Decoding Mitigates Text Hallucinations in Large Language Models", "abs": "Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language processing tasks. Still, they are prone to generating hallucinations\u2014factually incorrect or fabricated content that can undermine their reliability, especially in high-stakes domains such as healthcare and legal advisory. In response to this challenge, we propose Delta, a novel inference-time approach that leverages contrastive decoding to mitigate hallucinations without requiring model retraining or additional training data. Delta works by randomly masking portions of the input prompt, then contrasting the original and masked output distribution generated by the model, effectively mitigating hallucinations through inference-only computations. Delta was evaluated on context-rich QA benchmarks like SQuAD v1.1 and v2, achieving around 3 and 6 percentage points of improvement, respectively. It also showed gains of 7 and 2 percentage points on TriviaQA and Natural Question under-sampling decoding. Delta improved SQuAD v2\u2019s no-answer exact match by over ten percentage points. These findings suggest that Delta is particularly effective when hallucinations arise from contextual ambiguity. Delta presents a computationally efficient and scalable solution for reducing hallucinations in real-world LLM applications by focusing on inference-time enhancements.", "keywords": [ "Contrastive Decoding", "Text Hallucination Mitigation", "Large Language Models" ], "rating_list": [ 5, 3, 3, 1, 3 ], "soundness_list": [ 3, 2, 2, 1, 2 ], "presentation_list": [ 3, 2, 3, 1, 2 ], "contribution_list": [ 2, 2, 2, 1, 1 ], "confidence_list": [ 4, 4, 4, 5, 5 ] }, { "paper_id": "HNOo4UNPBF", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Chunlei Li", "gender": "unknown", "institution": "Wuhan University of Technology", "country": "CN", "position": "Researcher" }, { "name": "Jingliang Hu", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Researcher" }, { "name": "Lichao Mou", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Researcher" }, { "name": "Yilei Shi", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Researcher" }, { "name": "Xiaoxiang Zhu", "gender": "Female", "institution": "Technical University Munich", "country": "DE", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 36, "n_ref": 72, "n_ref_all": 92, "n_fig": 7, "n_tab": 8, "L_tab": 2260, "n_element_tab": 202, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 802, "n_element_tab_1": 68, "formula_len_all": 242, "formula_len_all_1": 242, "len_all": 114272, "len_all_1": 55834, "len_abs": 1385, "len_title": 137, "len_sents": 32294, "len_sents_1": 25325, "n_sents": 238, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1393, "title": "Scale-Aware Contrastive Reverse Distillation for Unsupervised Anomaly Detection", "abs": "Unsupervised anomaly detection using deep learning has garnered significant research attention due to its broad applicability, particularly in medical imaging where labeled anomalous data are scarce. While earlier approaches leverage generative models like autoencoders and generative adversarial networks (GANs), they often fall short due to overgeneralization. Recent methods explore various strategies, including memory banks, normalizing flows, self-supervised learning, and knowledge distillation, to enhance discrimination. Among these, knowledge distillation, particularly reverse distillation, has shown promise. Following this paradigm, we propose a novel scale-aware contrastive reverse distillation model that addresses two key limitations of existing reverse distillation methods: insufficient feature discriminability and inability to handle anomaly scale variations. Specifically, we introduce a contrastive student-teacher learning approach to derive more discriminative representations by generating and exploring out-of-normal distributions. Further, we design a scale adaptation mechanism to softly weight contrastive distillation losses at different scales to account for the scale variation issue. Extensive experiments on benchmark datasets demonstrate state-of-the-art performance, validating the efficacy of the proposed method. The code will be made publicly available.", "keywords": [ "unsupervised anomaly detection", "medical images", "contrastive reverse distillation", "student-teacher" ], "rating_list": [ 8, 8, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "HNJJEWfo0Z", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Kiran Raja", "gender": "Male", "institution": "Norwegian University of Science and Technology", "country": "NO", "position": "Associate Professor" }, { "name": "Philipp Terh\u00f6rst", "gender": "Male", "institution": "Universit\u00e4t Paderborn", "country": "DE", "position": "Assistant Professor" }, { "name": "Raghavendra Ramachandra", "gender": "Male", "institution": "Norwegian University of Science and Technology (NTNU)", "country": "NO", "position": "Full Professor" }, { "name": "Raymond Veldhuis", "gender": "Male", "institution": "Norwegian University of Science and Technology", "country": "NO", "position": "Full Professor" }, { "name": "Revoti Prasad Bora", "gender": "Male", "institution": "Norwegian University of Science and Technology", "country": "NO", "position": "PhD candidate" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 25, "n_ref": 43, "n_ref_all": 76, "n_fig": 13, "n_tab": 2, "L_tab": 1336, "n_element_tab": 66, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 1358, "n_element_tab_1": 66, "formula_len_all": 359, "formula_len_all_1": 359, "len_all": 75142, "len_all_1": 48390, "len_abs": 1241, "len_title": 114, "len_sents": 21769, "len_sents_1": 21543, "n_sents": 145, "n_sents_1": 147, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1145, "title": "Examining Why Perturbation-Based Fidelity Metrics are Inconsistent", "abs": "Saliency maps are commonly employed as a post-hoc method to explain the decision-making processes of Deep Learning models. Despite their widespread use, ensuring the fidelity of saliency maps is challenging due to the absence of ground truth. Therefore, researchers have developed fidelity metrics to evaluate the fidelity of saliency maps. However, prior investigations have uncovered statistical inconsistencies in existing fidelity metrics using multiple perturbation techniques without delving into the underlying causes. Our study aims to explore the origins of these observed inconsistencies. Our analysis examines the correctness of the assumptions made by the existing fidelity metrics using different types of perturbation to perturb the images. Our findings reveal that the assumptions made by existing fidelity metrics do not always hold true. Consequently, the existing fidelity metrics become inconsistent and unreliable. Thus, we recommend a cautious interpretation of fidelity metrics and the choice of perturbation technique when evaluating the fidelity of saliency maps in eXplainable Artificial Intelligence (XAI) applications.", "keywords": [ "Fidelity Metric", "Perturbation", "Inconsistency", "XAI", "Explainability", "Interpretability" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 2, 3, 4 ] }, { "paper_id": "HN8V0flwJF", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Matei Zaharia", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Pieter Abbeel", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Wilson Yan", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Hao Liu", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 40, "n_ref": 75, "n_ref_all": 105, "n_fig": 9, "n_tab": 9, "L_tab": 2514, "n_element_tab": 445, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 557, "n_element_tab_1": 18, "formula_len_all": 58, "formula_len_all_1": 38, "len_all": 163375, "len_all_1": 49414, "len_abs": 879, "len_title": 125, "len_sents": 40348, "len_sents_1": 23094, "n_sents": 312, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 882, "title": "World Model on Million-Length Video And Language With Blockwise RingAttention", "abs": "Enabling long-context understanding remains a key challenge in scaling existing sequence models -- a crucial component in developing generally intelligent models that can process and operate over long temporal horizons that potentially consist of millions of tokens. In this paper, we aim to address these challenges by providing a comprehensive exploration of the full development process for producing 1M context language models and video-language models, setting new benchmarks in language retrieval and new capabilities in long video understanding. We detail our long context data curation process, progressive context extension from 4K to 1M tokens, and present an efficient open-source implementation for scalable training on long sequences. Additionally, we open-source a family of 7B parameter models capable of processing long text documents and videos exceeding 1M tokens.", "keywords": [ "long context", "world model", "attention" ], "rating_list": [ 6, 6, 3, 8 ], "soundness_list": [ 4, 4, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 4, 2, 2, 4 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "HN0CYZbAPw", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Andy Peng", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Aviral Kumar", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Qiyang Li", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Zhiyuan Zhou", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 56, "n_ref": 108, "n_ref_all": 153, "n_fig": 23, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 67, "formula_len_all_1": 67, "len_all": 172015, "len_all_1": 73639, "len_abs": 1746, "len_title": 80, "len_sents": 57776, "len_sents_1": 38040, "n_sents": 367, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1971, "title": "Efficient Online Reinforcement Learning Fine-Tuning Need Not Retain Offline Data", "abs": "The modern paradigm in machine learning involves pre-training on diverse data, followed by task-specific fine-tuning. In reinforcement learning (RL), this translates to learning via offline RL on a diverse historical dataset, followed by rapid online RL fine-tuning using interaction data. Most RL fine-tuning methods require continued training on offline data for stability and performance. However, this is undesirable because training on diverse offline data is slow and expensive for large datasets, and should, in principle, also limit the performance improvement possible because of constraints or pessimism on offline data. In this paper, we show that retaining offline data is unnecessary as long as we use a properly-designed online RL approach for fine-tuning offline RL initializations. To build this approach, we start by analyzing the role of retaining offline data in online fine-tuning. We find that continued training on offline data is mostly useful for preventing a sudden divergence in the value function at the onset of fine-tuning, caused by a distribution mismatch between the offline data and online rollouts. This divergence typically results in unlearning and forgetting the benefits of offline pre-training. Our approach, Warm-start RL (WSRL), mitigates the catastrophic forgetting of pre-trained initializations using a very simple idea. WSRL employs a warmup phase that seeds the online RL run with a very small number of rollouts from the pre-trained policy to do fast online RL. The data collected during warmup bridges the distribution mismatch, and helps ``recalibrate'' the offline Q-function to the online distribution, allowing us to completely discard offline data without destabilizing the online RL fine-tuning. We show that WSRL is able to fine-tune without retaining any offline data, and is able to learn faster and attains higher performance than existing algorithms irrespective of whether they do or do not retain offline data.", "keywords": [ "Reinforcement learning", "fast fine-tuning" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "HMrcv7Q4Ub", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Danylo Vashchilenko", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "Lecturer" }, { "name": "Dezhan Tu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Panpan Xu", "gender": "Female", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yuzhe Lu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 25, "n_ref": 78, "n_ref_all": 110, "n_fig": 14, "n_tab": 3, "L_tab": 2599, "n_element_tab": 746, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 1225, "n_element_tab_1": 289, "formula_len_all": 689, "formula_len_all_1": 784, "len_all": 153560, "len_all_1": 68280, "len_abs": 1393, "len_title": 155, "len_sents": 46180, "len_sents_1": 28442, "n_sents": 351, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 107, "L_abs": 1399, "title": "VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration", "abs": "Vision-Language Models (VLMs) have demonstrated impressive performance across a versatile set of tasks. A key challenge in accelerating VLMs is storing and accessing the large Key-Value (KV) cache that encodes long visual contexts, such as images or videos. While existing KV cache compression methods are effective for Large Language Models (LLMs), directly migrating them to VLMs yields suboptimal accuracy and speedup. To bridge the gap, we propose VL-Cache, a novel KV cache compression recipe tailored for accelerating VLM inference. In this paper, we first investigate the unique sparsity pattern of VLM attention by distinguishing visual and text tokens in prefill and decoding phases. Based on these observations, we introduce a layer-adaptive sparsity-aware cache budget allocation method that effectively distributes the limited cache budget across different layers, further reducing KV cache size without compromising accuracy. Additionally, we develop a modality-aware token scoring policy to better evaluate the token importance. Empirical results on multiple benchmark datasets demonstrate that retaining only 10% of KV cache achieves accuracy comparable to that with full cache. In a speed benchmark, our method accelerates end-to-end latency of generating 100 tokens by up to 2.33x and speeds up decoding by up to 7.08x, while reducing the memory footprint of KV cache in GPU by 90%.", "keywords": [ "KV Cache Compression", "Vision-Language Models", "Inference Acceleration", "Sparsity", "Modality" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "HMa8mIiBT8", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Mahardika Krisna Ihsani", "gender": "unknown", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "MS student" }, { "name": "Min-Yen Kan", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Associate Professor" }, { "name": "Ai Xi", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" } ], "n_formula": 10, "n_formula_1": 5, "n_ref_uni": 43, "n_ref": 69, "n_ref_all": 95, "n_fig": 13, "n_tab": 1, "L_tab": 227, "n_element_tab": 14, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 227, "n_element_tab_1": 14, "formula_len_all": 527, "formula_len_all_1": 642, "len_all": 162064, "len_all_1": 59626, "len_abs": 1149, "len_title": 121, "len_sents": 49078, "len_sents_1": 28550, "n_sents": 292, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1040, "title": "Is Knowledge in Multilingual Language Models Cross-Lingually Consistent?", "abs": "Few works study the variation and cross-lingual consistency of factual knowledge embedded in multilingual models. However, cross-lingual consistency should be considered to assess cross-lingual transferability, maintain the factuality of the model\u2019s knowledge across languages, and preserve the parity of language model performance. We are thus interested in analyzing, evaluating, and interpreting cross-lingual consistency for factual knowledge. We apply interpretability approaches to analyze a model\u2019s behavior in cross-lingual contexts, discovering that multilingual models show different levels of consistency, subject to either language families or linguistic factors. Further, we identify a cross-lingual consistency bottleneck manifested in middle layers. To mitigate this problem, we try vocabulary expansion, additional cross-lingual objectives, and adding biases from monolingual inputs. We find that all these methods boost cross-lingual consistency to some extent, with cross-lingual supervision offering the best improvement.", "keywords": [ "Multilingual Models", "Fact-checking", "Cross-lingual Knowledge Consistency", "Self-consistency", "Model Parity" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "HMVDiaWMwM", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Hai Huang", "gender": "Male", "institution": "Atlassian", "country": "US", "position": "Researcher" }, { "name": "Huangjie Zheng", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Mingyuan Zhou", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Zhendong Wang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 22, "n_ref_uni": 90, "n_ref": 173, "n_ref_all": 236, "n_fig": 11, "n_tab": 5, "L_tab": 3721, "n_element_tab": 602, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 3691, "n_element_tab_1": 487, "formula_len_all": 2497, "formula_len_all_1": 2274, "len_all": 252324, "len_all_1": 70024, "len_abs": 1204, "len_title": 130, "len_sents": 50133, "len_sents_1": 29054, "n_sents": 314, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1548, "title": "Guided Score identity Distillation for Data-Free One-Step Text-to-Image Generation", "abs": "Diffusion-based text-to-image generation models trained on extensive text-image pairs have demonstrated the ability to produce photorealistic images aligned with textual descriptions. However, a significant limitation of these models is their slow sample generation process, which requires iterative refinement through the same network. To overcome this, we introduce a data-free guided distillation method that enables the efficient distillation of pretrained Stable Diffusion models without access to the real training data, often restricted due to legal, privacy, or cost concerns. This method enhances Score identity Distillation (SiD) with Long and Short Classifier-Free Guidance (LSG), an innovative strategy that applies Classifier-Free Guidance (CFG) not only to the evaluation of the pretrained diffusion model but also to the training and evaluation of the fake score network. We optimize a model-based explicit score matching loss using a score-identity-based approximation alongside our proposed guidance strategies for practical computation. By exclusively training with synthetic images generated by its one-step generator, our data-free distillation method rapidly improves FID and CLIP scores, achieving state-of-the-art FID performance while maintaining a competitive CLIP score. Notably, the one-step distillation of Stable Diffusion 1.5 achieves an FID of **8.15** on the COCO-2014 validation set, a record low value under the data-free setting. Our code and checkpoints are available at https://github.com/mingyuanzhou/SiD-LSG.", "keywords": [ "stable diffusion", "data-free distillation", "single-step generation", "classifier-free guidance" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 4, 4, 3, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 2, 4, 2, 3 ] }, { "paper_id": "HLxWF7xqiK", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Alessandro Rozza", "gender": "Male", "institution": "lastminute.com group", "country": "", "position": "Chief Scientist" }, { "name": "Filippo Cipriani", "gender": "Male", "institution": "Polytechnic Institute of Milan", "country": "IT", "position": "MS student" }, { "name": "Francesco Emanuele Stradi", "gender": "Male", "institution": "Polytechnic Institute of Milan", "country": "IT", "position": "PhD student" }, { "name": "Lorenzo Ciampiconi", "gender": "Male", "institution": "lastminute.com", "country": "CH", "position": "Researcher" }, { "name": "Marco Leonardi", "gender": "unknown", "institution": "lastminute", "country": "CH", "position": "Researcher" }, { "name": "Nicola Gatti", "gender": "Male", "institution": "Polytechnic Institute of Milan", "country": "IT", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 16, "n_ref_uni": 26, "n_ref": 47, "n_ref_all": 63, "n_fig": 8, "n_tab": 1, "L_tab": 199, "n_element_tab": 12, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 199, "n_element_tab_1": 12, "formula_len_all": 2251, "formula_len_all_1": 1770, "len_all": 130429, "len_all_1": 69933, "len_abs": 1127, "len_title": 160, "len_sents": 44751, "len_sents_1": 31343, "n_sents": 354, "n_sents_1": 252, "venue": "ICLR2025", "decision": "withdraw", "L_title": 111, "L_abs": 1132, "title": "A Primal-Dual Approach for Dynamic Pricing of Sequentially Displayed Complementary Items under Sale Constraints", "abs": "We address the challenging problem of dynamically pricing complementary items that are sequentially displayed to customers. An illustrative example is the online sale of flight tickets, where customers navigate through multiple web pages. Initially, they view the ticket cost, followed by ancillary expenses such as insurance and additional luggage fees. Coherent pricing policies for complementary items are essential because optimizing the pricing of each item individually is ineffective. Our scenario also involves a sales constraint, which specifies a minimum number of items to sell, and uncertainty regarding customer demand curves. To tackle this problem, we originally formulate it as a Markov decision process with constraints. Leveraging online learning tools, we design a primal-dual online optimization algorithm. We empirically evaluate our approach using synthetic settings randomly generated from real-world data, covering various configurations from stationary to non-stationary, and compare its performance in terms of constraints violation and regret against well-known baselines optimizing each state singularly.", "keywords": [ "dynamic pricing", "primal-dual" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 2, 1 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "HJp1g4w1Or", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Jeremiah Ridge", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "MS student" }, { "name": "Oiwi Parker Jones", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "Principal Researcher" } ], "n_formula": 8, "n_formula_1": 0, "n_ref_uni": 40, "n_ref": 114, "n_ref_all": 148, "n_fig": 9, "n_tab": 4, "L_tab": 671, "n_element_tab": 68, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 599, "n_element_tab_1": 58, "formula_len_all": 342, "formula_len_all_1": 0, "len_all": 169477, "len_all_1": 62917, "len_abs": 1548, "len_title": 133, "len_sents": 55525, "len_sents_1": 31239, "n_sents": 373, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1555, "title": "Resolving Domain Shift For Representations Of Speech In Non-Invasive Brain Recordings", "abs": "Machine learning techniques have enabled researchers to leverage neuroimaging data to decode speech from brain activity, with some amazing recent successes achieved by applications built using invasive devices. However, research requiring surgical implants has a number of practical limitations. Non-invasive neuroimaging techniques provide an alternative but come with their own set of challenges, the limited scale of individual studies being among them. Without the ability to pool the recordings from different non-invasive studies, data on the order of magnitude needed to leverage deep learning techniques to their full potential remains out of reach. In this work, we focus on non-invasive data collected using magnetoencephalography (MEG). We leverage two different, leading speech decoding models to investigate how an adversarial domain adaptation framework augments their ability to generalize across datasets. We successfully improve the performance of both models when training across multiple datasets. To the best of our knowledge, this study is the first ever application of feature-level, deep learning based harmonization for MEG neuroimaging data. Our analysis additionally offers further evidence of the impact of demographic features on neuroimaging data, demonstrating that participant age strongly affects how machine learning models solve speech decoding tasks using MEG data. Lastly, in the course of this study we produce a new open-source implementation of one of these models to the benefit of the broader scientific community.", "keywords": [ "Domain Adaptation", "MEG", "Magnetoencephalography", "Speech Decoding", "Brain-Computer Interfaces", "Non-invasive Neuroimaging", "Domain Shift" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "HJWdrvVyOi", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 1, "author_info_list": [ { "name": "Li-Yue Sun", "gender": "Male", "institution": "Nankai University", "country": "", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 27, "n_ref_uni": 12, "n_ref": 45, "n_ref_all": 51, "n_fig": 3, "n_tab": 2, "L_tab": 673, "n_element_tab": 92, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2745, "formula_len_all_1": 2114, "len_all": 77143, "len_all_1": 50032, "len_abs": 1047, "len_title": 126, "len_sents": 25474, "len_sents_1": 21511, "n_sents": 205, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1054, "title": "Privacy-Preserving Logistic Regression Training with A Faster Gradient Variant", "abs": "Training logistic regression over encrypted data has been a compelling approach in addressing security concerns for several years. In this paper, we introduce an efficient gradient variant, called $quadratic$ $gradient$, for privacy-preserving logistic regression training. We enhance Nesterov's Accelerated Gradient (NAG), Adaptive Gradient Algorithm (Adagrad) and Adam algorithms by incorporating their quadratic gradients and evaluate these improved algorithms on various datasets. Experimental results demonstrate that the enhanced algorithms achieve significantly improved convergence speed compared to traditional first-order gradient methods. Moreover, we applied the enhanced NAG method to implement homomorphic logistic regression training, achieving comparable results within just 4 iterations.\nThere is a good chance that the quadratic gradient approach could integrate first-order gradient descent/ascent algorithms with the second-order Newton-Raphson methods, and that it could be applied to a wide range of numerical optimization problems.", "keywords": [ "Homomorphic Encryption", "Logistic Regression", "Quadratic Gradient", "Simplified Fixed Hessian", "Nesterov\u2019s Accelerated Gradient", "Gradient Ascent Algorithm" ], "rating_list": [ 5, 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 3, 1, 2, 2, 2 ], "confidence_list": [ 2, 4, 4, 3, 3 ] }, { "paper_id": "HHmnfVQagN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Fangxu Yu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Haoqiang Kang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Lai Jiang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Lianhui Qin", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Assistant Professor" }, { "name": "Shibo Hao", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 3, "n_ref_uni": 84, "n_ref": 238, "n_ref_all": 275, "n_fig": 9, "n_tab": 16, "L_tab": 7105, "n_element_tab": 529, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1101, "n_element_tab_1": 149, "formula_len_all": 589, "formula_len_all_1": 121, "len_all": 308007, "len_all_1": 69254, "len_abs": 1678, "len_title": 133, "len_sents": 69815, "len_sents_1": 31893, "n_sents": 548, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1677, "title": "Flow of Reasoning: Training LLMs for Divergent Problem Solving with Minimal Examples", "abs": "The ability to generate diverse solutions to a given problem is a hallmark of human creativity. This divergent reasoning is also crucial for machines, enhancing their robustness and enabling them to assist humans in many applications such as scientific discovery. However, existing approaches to multi-step reasoning with large language models (LLMs) have mostly focused only on the reasoning accuracy, without further discovering more diverse valid solutions. For example, supervised fine-tuning can improve LLM reasoning quality, but requires extensive supervised data to capture the full range of possible solutions. Reinforcement learning aims to find limited highest-reward solutions while neglecting the solution diversity. To fill this gap, we propose Flow of Reasoning (FoR), an efficient diversity-seeking LLM finetuning method aimed at improving reasoning quality and diversity with minimal data. FoR formulates multi-step LLM reasoning as a Markovian flow on a DAG-structured reasoning graph. This formulation allows us to incorporate and adapt principled GFlowNet approaches, for finetuning LLMs to sample diverse reasoning paths with probabilities proportional to the (unnormalized) reward of target problems. Extensive experiments show that, with limited training examples (e.g., 15 examples), FoR enables the discovery of diverse, creative, high-quality solutions, greatly outperforming a wide range of existing inference and training methods across five challenging puzzle-solving tasks, including BlocksWorld (embodied reasoning), Game24 (math puzzle solving), PrOntoQA (logical reasoning), Rubik's Cube (spatial reasoning), and 1D-ARC (abstraction reasoning).", "keywords": [ "Large Language Models", "reasoning", "diversity", "multi-step reasoning", "planning", "creativity" ], "rating_list": [ 5, 5, 5, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "HHKboqbkec", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 7, "author_info_list": [ { "name": "Chunhui Zhang", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "PhD student" }, { "name": "Kwonjoon Lee", "gender": "Male", "institution": "Honda Research Institute USA", "country": "US", "position": "Sr Research Scientist" }, { "name": "Nakul Agarwal", "gender": "Male", "institution": "Honda Research Institute USA", "country": "US", "position": "Research Engineer" }, { "name": "Sean Dae Houlihan", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "Postdoc" }, { "name": "Shao-Yuan Lo", "gender": "unknown", "institution": "Honda Research Institute USA", "country": "US", "position": "Researcher" }, { "name": "Soroush Vosoughi", "gender": "unknown", "institution": "Dartmouth College", "country": "", "position": "Assistant Professor" }, { "name": "Zhongyu Ouyang", "gender": "Female", "institution": "Dartmouth College", "country": "US", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 11, "n_ref_uni": 39, "n_ref": 77, "n_ref_all": 86, "n_fig": 2, "n_tab": 11, "L_tab": 6285, "n_element_tab": 313, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 2852, "n_element_tab_1": 95, "formula_len_all": 1527, "formula_len_all_1": 857, "len_all": 195107, "len_all_1": 63337, "len_abs": 1756, "len_title": 120, "len_sents": 60182, "len_sents_1": 28513, "n_sents": 423, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1863, "title": "Scaling Multimodal Theory-of-Mind with Weak-to-Strong Bayesian Reasoning", "abs": "Theory of Mind (ToM) enables individuals to understand and predict thoughts, emotions, and intentions of the others. To replicate this cognitive ability in machines, especially under complex multimodal environments, recent advances combine Bayesian-based state inference with deep learning models to estimate mental states, where the Bayesian model handles state transitions and a language model (LM) estimates the likelihood of intermediate states. However, while post-training an LM to specialise in ToM tasks improves performance, the computational cost increases as the LM scales, limiting the model size to 7 billion parameters. Despite this post-training process, smaller LMs still struggle with the physical and mental modelling demands of ToM due to their limited world knowledge and reasoning capacity. To address this, we propose a scalable solution that leverages the strengths of larger LMs (up to 70 and 405 billion parameters, respectively), including their vast world knowledge and atomic-level reasoning capabilities, without increasing post-training resource requirements. Our method transfers ToM-specific behaviours from a post-trained small LM to guide the latent reasoning of a larger LM during test time. This weak-to-strong control mechanism enables the larger LM to improve Bayesian likelihood estimation at each inference step, harnessing its reasoning power in ToM scenarios while reducing the need for additional training resources. Extensive experiments demonstrate the significant effectiveness of our scaled approach. It is better at inferring human mental states in complex and interactive environments, outperforming the state-of-the-art solution by $\\sim4.6$% across multiple tasks on the multimodal ToM benchmark and unseen scenarios. Our code and datasets are available: https://anonymous.4open.science/r/scale-bayesian-tom-248B", "keywords": [ "Neuro-symbolic concept", "logic and forming reasoning", "theory-of-mind" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 4, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 1, 3, 4, 2 ] }, { "paper_id": "HHISuWB0nX", "primary_area": "learning on time series and dynamical systems", "n_author": 7, "author_info_list": [ { "name": "Changwen Zheng", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Jianqi Zhang", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Jingyao Wang", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Siyu Zhao", "gender": "Female", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Wenwen Qiang", "gender": "Male", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Xingyu Zhang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Zeen Song", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 26, "n_ref_uni": 43, "n_ref": 66, "n_ref_all": 82, "n_fig": 4, "n_tab": 3, "L_tab": 3871, "n_element_tab": 101, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 3622, "n_element_tab_1": 53, "formula_len_all": 1681, "formula_len_all_1": 1671, "len_all": 127001, "len_all_1": 69605, "len_abs": 1244, "len_title": 117, "len_sents": 34292, "len_sents_1": 28293, "n_sents": 273, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1283, "title": "What Makes a Good Time-series Forecasting Model? A Causal Perspective", "abs": "Generalization is a long-standing challenge in multivariate time series forecasting (MTSF) tasks. Most existing forecasting methods use all available variables in historical series to predict all future variables, assuming that there may be correlations among all variables. From a causal perspective, this reliance on correlated variables can compromise the model\u2019s generalization. To address this, we aim to explore the role of causal relationships in enhancing the generalization of multivariate time series models. We examine how graphical causal models, through conditional independence constraints, can reduce the hypothesis space, thereby improving generalization. Building on this foundation, we introduce a novel causality-based MTSF algorithm CAusal Informed Transformer (CAIFormer). It first constructs a Directed Acyclic Graph (DAG) among variables using causal discovery techniques. Then we build the forecasting model by enforcing the causal constraints informed by the DAG. Empirical evaluations on benchmark datasets demonstrate that our method surpasses traditional approaches in predictive accuracy. Additionally, we present the structural causal models derived for these datasets, underscoring the practical applicability of our causality-driven framework in MTSF.", "keywords": [ "Time-series Forecasting", "Causal discovery" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "HH4KWP8RP5", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Emmanuel Bengio", "gender": "Male", "institution": "Recursion", "country": "", "position": "Researcher" }, { "name": "Glen Berseth", "gender": "Male", "institution": "University of Montreal, University of Montreal", "country": "", "position": "Assistant Professor" }, { "name": "Kanika Madan", "gender": "unknown", "institution": "MILA, University of Montreal", "country": "", "position": "PhD student" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" }, { "name": "Alex Matthew Lamb", "gender": "Male", "institution": "Microsoft Research NYC", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 30, "n_ref": 111, "n_ref_all": 142, "n_fig": 19, "n_tab": 3, "L_tab": 570, "n_element_tab": 37, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 2839, "n_element_tab_1": 64, "formula_len_all": 128, "formula_len_all_1": 126, "len_all": 121067, "len_all_1": 56561, "len_abs": 1336, "len_title": 113, "len_sents": 39733, "len_sents_1": 25060, "n_sents": 222, "n_sents_1": 139, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1339, "title": "Towards Improving Exploration through Sibling Augmented GFlowNets", "abs": "Exploration is a key factor for the success of an active learning agent, especially when dealing with sparse extrinsic terminal rewards and long trajectories. We introduce Sibling Augmented Generative Flow Networks (SA-GFN), a novel framework designed to enhance exploration and training efficiency of Generative Flow Networks (GFlowNets). SA-GFN uses a decoupled dual network architecture, comprising of a main Behavior Network and an exploratory Sibling Network, to enable a diverse exploration of the underlying distribution using intrinsic rewards. Inspired by the ideas on exploration from reinforcement learning, SA-GFN provides a general-purpose exploration and learning paradigm that integrates with multiple GFlowNet training objectives and is especially helpful for exploration over a wide range of sparse or low reward distributions and task structures. An extensive set of experiments across a diverse range of tasks, reward structures and trajectory lengths, along with a thorough set of ablations, demonstrate the superior performance of SA-GFN in terms of exploration efficacy and convergence speed as compared to the existing methods. In addition, SA-GFN's versatility and compatibility with different GFlowNet training objectives and intrinsic reward methods underscores its broad applicability in various problem domains.", "keywords": [ "Generative Models", "Generative Flow Networks", "Exploration" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 3, 3, 2 ] }, { "paper_id": "HGz012J6TQ", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Chenqing Hua", "gender": "Non-Binary", "institution": "Stanford University", "country": "US", "position": "Intern" }, { "name": "Dinghuai Zhang", "gender": "Unspecified", "institution": "Mila, University of Montreal", "country": "CA", "position": "PhD student" }, { "name": "Doina Precup", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Research Team Lead" }, { "name": "Guy Wolf", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "Associate Professor" }, { "name": "Kevin K Yang", "gender": "unknown", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Odin Zhang", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "Undergrad student" }, { "name": "Sitao Luan", "gender": "Male", "institution": "McGill University", "country": "", "position": "PhD student" }, { "name": "Yong Liu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "MS student" }, { "name": "SHUANGJIA ZHENG", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 54, "n_formula_1": 32, "n_ref_uni": 88, "n_ref": 232, "n_ref_all": 258, "n_fig": 12, "n_tab": 7, "L_tab": 3170, "n_element_tab": 178, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 484, "n_element_tab_1": 16, "formula_len_all": 11540, "formula_len_all_1": 8487, "len_all": 362273, "len_all_1": 133516, "len_abs": 1281, "len_title": 118, "len_sents": 103427, "len_sents_1": 57650, "n_sents": 916, "n_sents_1": 440, "venue": "ICLR2025", "decision": "Reject", "L_title": 116, "L_abs": 1288, "title": "EnzymeFlow: Generating Reaction-specific Enzyme Catalytic Pockets through Flow Matching and Co-Evolutionary Dynamics", "abs": "Enzyme design is a critical area in biotechnology, with applications ranging from drug development to synthetic biology. Traditional methods for enzyme function prediction or protein binding pocket design often fall short in capturing the dynamic and complex nature of enzyme-substrate interactions, particularly in catalytic processes. To address the challenges, we introduce EnzymeFlow, a generative model that employs flow matching with hierarchical pre-training and enzyme-reaction co-evolution to generate catalytic pockets for specific substrates and catalytic reactions. Additionally, we introduce a large-scale, experimentally validated dataset of enzyme-reaction pairs, specifically designed for the catalytic pocket generation task, comprising a total of $328,192$ pairs. By incorporating evolutionary dynamics and reaction-specific adaptations, EnzymeFlow becomes a powerful model for designing enzyme pockets, which is capable of catalyzing a wide range of biochemical reactions. Experiments on the new dataset demonstrate the model's effectiveness in designing high-quality, functional enzyme catalytic pockets, paving the way for advancements in enzyme engineering and synthetic biology. The EnzymeFlow code can be found at https://anonymous.4open.science/r/EnzymeFlow-7420.", "keywords": [ "enzyme design; flow model; protein design; protein evolution" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 4, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "HGxGCjqnDd", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Aochuan Chen", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Fugee Tsung", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Jia Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Jiashun Cheng", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yuhan Li", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Ziqi Gao", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 63, "n_ref": 104, "n_ref_all": 121, "n_fig": 3, "n_tab": 8, "L_tab": 3972, "n_element_tab": 404, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2982, "n_element_tab_1": 310, "formula_len_all": 352, "formula_len_all_1": 352, "len_all": 147965, "len_all_1": 62974, "len_abs": 1037, "len_title": 113, "len_sents": 30079, "len_sents_1": 26624, "n_sents": 234, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1168, "title": "LoRA Unleashed: Effortlessly Advancing from Low to Arbitrary Rank", "abs": "Low-Rank Adaptation (LoRA) has emerged as a prominent technique for fine-tuning large foundation models, facilitating a reduction in trainable parameters through the utilization of low-rank matrices to represent weight changes $\\mathbf{A}$ and $\\mathbf{B}$ (\\textit{i.e.,} $\\Delta \\mathbf{W} = \\mathbf{B} \\mathbf{A}$). Although LoRA has demonstrated considerable success, its expressiveness is inherently limited by the constrained capacity of its low-rank structure. To ameliorate this limitation, we introduce \\underline{Fo}urier-based Flexible \\underline{R}ank \\underline{A}daptation (FoRA), which harnesses the robust expressiveness of the Fourier basis to re-parameterize $\\mathbf{A}$ and $\\mathbf{B}$ from a sparse spectral subspace. Utilizing FoRA, adaptation matrices can overcome conventional rank limitations, achieving up to a 15x reduction in the parameter budget. We illustrate that FoRA achieves an optimal balance of efficiency and performance across various tasks, including natural language understanding, mathematical reasoning, commonsense reasoning, and image classification. Our codes are available at https://anonymous.4open.science/r/FoRA-0E9C.", "keywords": [ "low-rank adaptation", "parameter-efficient fine-tuning", "sparse learning", "large language models" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "HGCk5aaSvE", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Byung-Jun Yoon", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Full Professor" }, { "name": "Gilchan Park", "gender": "Male", "institution": "Brookhaven National Laboratory", "country": "", "position": "Researcher" }, { "name": "Guang Zhao", "gender": "Male", "institution": "BNL", "country": "US", "position": "Researcher" }, { "name": "Shantenu Jha", "gender": "unknown", "institution": "Rutgers University New Brunswick", "country": "", "position": "" }, { "name": "Shinjae Yoo", "gender": "Male", "institution": "Brookhaven National Lab", "country": "", "position": "Scientist" }, { "name": "Xiaoning Qian", "gender": "Male", "institution": "Texas A&M", "country": "", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 35, "n_ref": 53, "n_ref_all": 73, "n_fig": 8, "n_tab": 2, "L_tab": 514, "n_element_tab": 44, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 514, "n_element_tab_1": 44, "formula_len_all": 427, "formula_len_all_1": 689, "len_all": 115578, "len_all_1": 68647, "len_abs": 1152, "len_title": 74, "len_sents": 39171, "len_sents_1": 33745, "n_sents": 276, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 26, "L_abs": 1159, "title": "Pareto Prompt Optimization", "abs": "Natural language prompt optimization, or prompt engineering, has emerged as a powerful technique to unlock the potential of Large Language Models (LLMs) for various tasks. While existing methods primarily focus on maximizing a single task-specific performance metric for LLM outputs, real-world applications often require considering trade-offs between multiple objectives. In this work, we address this limitation by proposing an effective technique for multi-objective prompt optimization for LLMs. Specifically, we propose **ParetoPrompt**, a reinforcement learning~(RL) method that leverages dominance relationships between prompts to derive a policy model for prompts optimization using preference-based loss functions. By leveraging multi-objective dominance relationships, ParetoPrompt enables efficient exploration of the entire Pareto front without the need for a predefined scalarization of multiple objectives. Our experimental results show that ParetoPrompt consistently outperforms existing algorithms that use specific objective values. ParetoPrompt also yields robust performances when the objective metrics differ between training and testing.", "keywords": [ "Large Language Model", "Prompt Optimization", "Multiobjective Optimization", "Reinforcement Learning", "DPO" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "HFAIxjBB6K", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Kaipeng Zhang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Ziyao Guo", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zheng Shou", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 16, "n_ref_uni": 38, "n_ref": 78, "n_ref_all": 92, "n_fig": 2, "n_tab": 8, "L_tab": 1708, "n_element_tab": 239, "n_fig_1": 2, "n_tab_1": 10, "L_tab_1": 1933, "n_element_tab_1": 258, "formula_len_all": 653, "formula_len_all_1": 727, "len_all": 106724, "len_all_1": 64348, "len_abs": 1029, "len_title": 128, "len_sents": 28858, "len_sents_1": 26472, "n_sents": 219, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1033, "title": "Improving Autoregressive Image Generation by Mitigating Gradient Bias in Softmax", "abs": "Softmax is the most commonly used probabilistic activation function in classification tasks, partly due to its tendency to over-penalize non-target classes with high prediction scores. However, this property becomes detrimental in autoregressive generation tasks, where multiple valid predictions may exist. Unlike conventional classification task, which seeks a single correct answer, autoregressive models are expected to assign high probabilities to various plausible outputs to ensure diversity in generation. However, during training, gradient bias caused by Softmax over-penalizes non-target predictions with high probabilities, limiting output diversity and hindering optimization convergence. To alleviate this, we propose Gradient Suppressed Softmax (GS-Softmax), which reduces the gradient contributions of high-probability non-target classes. Through experiments, we demonstrate that GS-Softmax improves both the diversity of generated content and optimization convergence. Code and pre-trained models will be made public.", "keywords": [ "Image Generation", "Autoregressive Generative Model", "Softmax" ], "rating_list": [ 3, 8, 5, 3 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "HF1UmIVv6a", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Nan Song", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Ruibo Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Shichao Dong", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Ze Yang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Guosheng Lin", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 28, "n_ref": 199, "n_ref_all": 214, "n_fig": 3, "n_tab": 7, "L_tab": 4660, "n_element_tab": 528, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2720, "n_element_tab_1": 330, "formula_len_all": 505, "formula_len_all_1": 502, "len_all": 127008, "len_all_1": 70272, "len_abs": 1577, "len_title": 154, "len_sents": 37722, "len_sents_1": 28626, "n_sents": 255, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 1584, "title": "ADAPT: Attentive Self-Distillation and Dual-Decoder Prediction Fusion for Continual Panoptic Segmentation", "abs": "Panoptic segmentation, which unifies semantic and instance segmentation into a single task, has witnessed considerable success on predefined tasks. However, traditional methods tend to struggle with catastrophic forgetting and poor generalization when learning from a continuous stream of new tasks. Continual learning, emerged to tackle these challenges, has garnered increasing attention in recent years. Nonetheless, our study reveals that existing continual panoptic segmentation (CPS) methods often suffer from efficiency or scalability issues. To address these limitations, we propose a novel dual-decoder framework that incorporates attentive self-distillation and prediction fusion to efficiently preserve prior knowledge while facilitating model generalization. Specifically, we freeze the majority of model weights up to the pixel decoder, which is shared between the teacher and student models, thus enabling efficient knowledge distillation with only a single forward pass. Attentive self-distillation then adaptively distills useful knowledge from the old classes without distracting from non-object regions, which mitigates the inherent bias toward newly learned tasks. Additionally, query-level fusion (QLF) is devised to seamlessly integrate the output of the dual decoders without incurring scale inconsistency. Crucially, the computational overhead of our approach remains nearly constant, regardless of the number of continual learning steps or the number of classes introduced at each step. Our method achieves state-of-the-art performance on the ADE20K benchmark.", "keywords": [ "Continual Learning", "Panoptic Segmentation", "Knowledge Distillation" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "HEi70bBquo", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Olivier Ferret", "gender": "unknown", "institution": "CEA", "country": "", "position": "Principal Researcher" }, { "name": "Paul Grimal", "gender": "unknown", "institution": "CEA", "country": "FR", "position": "PhD student" }, { "name": "Herv\u00e9 Le Borgne", "gender": "Male", "institution": "CEA", "country": "FR", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 40, "n_ref": 78, "n_ref_all": 127, "n_fig": 18, "n_tab": 15, "L_tab": 5759, "n_element_tab": 476, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1604, "n_element_tab_1": 146, "formula_len_all": 314, "formula_len_all_1": 313, "len_all": 185111, "len_all_1": 60261, "len_abs": 719, "len_title": 93, "len_sents": 53935, "len_sents_1": 25784, "n_sents": 452, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 719, "title": "Signal Dynamics in Diffusion Models: Enhancing Text-to-Image Alignment through Step Selection", "abs": "Visual generative AI models often encounter challenges related to text-image alignment and reasoning limitations. This paper presents a novel method for selectively enhancing the signal at critical diffusion steps, optimizing image generation based on input semantics. Our approach addresses the shortcomings of early-stage signal modifications, demonstrating that adjustments made at later stages yield superior results. We conduct extensive experiments to validate the effectiveness of our method in producing semantically aligned images, achieving state-of-the-art performance. Our results highlight the importance of a judicious choice of sampling stage to improve diffusion performance and overall image alignment.", "keywords": [ "Visual Generative AI", "Diffusion Model", "Text-to-Image Alignment" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 1, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "HEXtydywnE", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Jiani Huang", "gender": "Female", "institution": "School of Engineering and Applied Science, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Mayur Naik", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "Professor" }, { "name": "Ziyang Li", "gender": "Male", "institution": "School of Engineering and Applied Science, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Ser-Nam Lim", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 48, "n_ref": 83, "n_ref_all": 110, "n_fig": 11, "n_tab": 4, "L_tab": 404, "n_element_tab": 40, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 922, "formula_len_all_1": 457, "len_all": 181691, "len_all_1": 64124, "len_abs": 1472, "len_title": 145, "len_sents": 51296, "len_sents_1": 31097, "n_sents": 426, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1477, "title": "LASER: A Neuro-Symbolic Framework for Learning Spatio-Temporal Scene Graphs with Weak Supervision", "abs": "Supervised approaches for learning spatio-temporal scene graphs (STSG) from video are greatly hindered due to their reliance on STSG-annotated videos, which are labor-intensive to construct at scale. Is it feasible to instead use readily available video captions as weak supervision? To address this question, we propose LASER, a neuro-symbolic framework to enable training STSG generators using only video captions. LASER employs large language models to first extract logical specifications with rich spatio-temporal semantic information from video captions. LASER then trains the underlying STSG generator to align the predicted STSG with the specification. The alignment algorithm overcomes the challenges of weak supervision by leveraging a differentiable symbolic reasoner and using a combination of contrastive, temporal, and semantics losses. The overall approach efficiently trains low-level perception models to extract a fine-grained STSG that conforms to the video caption. In doing so, it enables a novel methodology for learning STSGs without tedious annotations. We evaluate our method on three video datasets: OpenPVSG, 20BN, and MUGEN. Our approach demonstrates substantial improvements over fully-supervised baselines, achieving a unary predicate prediction accuracy of 27.78% (+12.65%) and a binary recall@5 of 0.42 (+0.22) on OpenPVSG. Additionally, LASER exceeds baselines by 7% on 20BN and 5.2% on MUGEN in terms of overall predicate prediction accuracy.", "keywords": [ "neuro-symbolic", "video understanding", "spatial-temporal scene graph", "weak supervision" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 4, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 2, 3, 4, 2, 4 ] }, { "paper_id": "HE6pJoNnFp", "primary_area": "generative models", "n_author": 11, "author_info_list": [ { "name": "Bang Liu", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Assistant Professor" }, { "name": "Caitlin Sikora", "gender": "Female", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Chu-Cheng Lin", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Ho Ko", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Jia-Chen Gu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Jindong Chen", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Lei Meng", "gender": "Female", "institution": "University of Notre Dame", "country": "", "position": "PhD student" }, { "name": "Lei Shu", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Liangchen Luo", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Yinxiao Liu", "gender": "Male", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Yun Zhu", "gender": "Male", "institution": "Google", "country": "", "position": "software engineer" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 26, "n_ref": 53, "n_ref_all": 77, "n_fig": 3, "n_tab": 13, "L_tab": 4193, "n_element_tab": 514, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 1881, "n_element_tab_1": 271, "formula_len_all": 151, "formula_len_all_1": 156, "len_all": 143717, "len_all_1": 70122, "len_abs": 2245, "len_title": 133, "len_sents": 37437, "len_sents_1": 31018, "n_sents": 289, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1354, "title": "Accelerating Inference of Retrieval-Augmented Generation via Sparse Context Selection", "abs": "Large language models (LLMs) augmented with retrieval exhibit robust performance and extensive versatility by incorporating external contexts. However, the input length grows linearly in the number of retrieved documents, causing a dramatic increase in latency. In this paper, we propose a novel paradigm named Sparse RAG, which seeks to cut computation costs through sparsity. Speci\ufb01cally, Sparse RAG encodes retrieved documents in parallel, which eliminates latency introduced by long-range attention of retrieved documents. Then, LLMs selectively decode the output by only attending to highly relevant caches auto-regressively, which are chosen via prompting LLMs with special control tokens. It is notable that Sparse RAG combines the assessment of each individual document and the generation of the response into a single process. The designed sparse mechanism in a RAG system can facilitate the reduction of the number of documents loaded during decoding for accelerating the inference of the RAG system. Additionally, \ufb01ltering out undesirable contexts enhances the model\u2019s focus on relevant context, inherently improving its generation quality. Evaluation results on four datasets show that Sparse RAG can be used to strike an optimal balance between generation quality and computational ef\ufb01ciency, demonstrating its generalizability across tasks.", "keywords": [ "Retrieval-Augmented Generation", "Sparse Context", "Efficiency" ], "rating_list": [ 5, 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "HE5JmwniHm", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "En Zhu", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Full Professor" }, { "name": "Jiyuan Liu", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Lecturer" }, { "name": "Shengju Yu", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Siwei Wang", "gender": "Male", "institution": "Intelligent Game and Decision Lab", "country": "CN", "position": "Assistant Professor" }, { "name": "Suyuan Liu", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Xinwang Liu", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yi Zhang", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Zhibin Dong", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 26, "n_ref_uni": 31, "n_ref": 66, "n_ref_all": 99, "n_fig": 9, "n_tab": 9, "L_tab": 2780, "n_element_tab": 335, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 3669, "n_element_tab_1": 307, "formula_len_all": 3047, "formula_len_all_1": 1611, "len_all": 142036, "len_all_1": 68674, "len_abs": 2119, "len_title": 159, "len_sents": 34340, "len_sents_1": 27846, "n_sents": 253, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 110, "L_abs": 2124, "title": "DLEFT-MKC: Dynamic Late Fusion Multiple Kernel Clustering with Robust Tensor Learning via Min-Max Optimizaiton", "abs": "Recent advancements in multiple kernel clustering (MKC) have highlighted the effectiveness of late fusion strategies, particularly in enhancing computational efficiency to near-linear complexity while achieving promising clustering performance. However, existing methods encounter three significant limitations: (1) reliance on fixed base partition matrices that do not adaptively optimize during the clustering process, thereby constraining their performance to the inherent representational capabilities of these matrices; (2) a focus on adjusting kernel weights to explore inter-view consistency and complementarity, which often neglects the intrinsic high-order correlations among views, thereby limiting the extraction of comprehensive multiple kernel information; (3) a lack of adaptive mechanisms to accommodate varying distributions within the data, which limits robustness and generalization. To address these challenges, this paper proposes a novel algorithm termed Dynamic Late Fusion Multiple Kernel Clustering with Robust {Tensor Learning via min-max optimization (DLEFT-MKC), which effectively overcomes the representational bottleneck of base partition matrices and facilitates the learning of meaningful high-order cross-view information. Specifically, it is the first to incorporate a min-max optimization paradigm into tensor-based MKC, enhancing algorithm robustness and generalization. Additionally, it dynamically reconstructs decision layers to enhance representation capabilities and subsequently stacks the reconstructed representations for tensor learning that promotes the capture of high-order associations and cluster structures across views, ultimately yielding consensus clustering partitions. To solve the resultant optimization problem, we innovatively design a strategy that combines reduced gradient descent with the alternating direction method of multipliers, ensuring convergence to local optima while maintaining high computational efficiency. Extensive experimental results across various benchmark datasets validate the superior effectiveness and efficiency of the proposed DLEFT-MKC.", "keywords": [ "multiple kernel clustering; multi-view clustering; late fusion MVC" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "HDmmwwTIlf", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Chuyu Liao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Ye Li", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "Associate Professor" }, { "name": "Zhongyi Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 3, "n_ref": 7, "n_ref_all": 12, "n_fig": 2, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1237, "formula_len_all_1": 1237, "len_all": 35441, "len_all_1": 35441, "len_abs": 531, "len_title": 125, "len_sents": 10778, "len_sents_1": 10778, "n_sents": 70, "n_sents_1": 70, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 530, "title": "Solving hyperbolic conservation laws with characteristic based neural network", "abs": "Neural network PDE solvers have recently gained popularity.However, it faces difficulty to deal with sharp discontinuity like shock waves in hyperbolic conservation laws.In this paper we propose a characteristic-based neural network to solve one dimension hyperbolic laws.The smooth solution can be derived by equation of characteristic lines ,and shock waves are decided by simple ODE solver.This method achieves a high accuracy with high efficiency. In the future it is hopeful to apply this method to higher dimension problems.", "keywords": [ "neural network", "hyperbolic conservation laws", "characteristics" ], "rating_list": [ 3, 3, 1, 3 ], "soundness_list": [ 2, 2, 1, 1 ], "presentation_list": [ 2, 1, 1, 1 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "HD6bWcj87Y", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Dawn Song", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Prateek Mittal", "gender": "unknown", "institution": "Princeton University", "country": "US", "position": "Full Professor" }, { "name": "Ruoxi Jia", "gender": "unknown", "institution": "Virginia Tech", "country": "US", "position": "Assistant Professor" }, { "name": "Tianhao Wang", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "PhD student" } ], "n_formula": 45, "n_formula_1": 14, "n_ref_uni": 46, "n_ref": 124, "n_ref_all": 169, "n_fig": 12, "n_tab": 10, "L_tab": 1748, "n_element_tab": 142, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2369, "n_element_tab_1": 12, "formula_len_all": 5593, "formula_len_all_1": 889, "len_all": 271401, "len_all_1": 79132, "len_abs": 1348, "len_title": 80, "len_sents": 95935, "len_sents_1": 36786, "n_sents": 718, "n_sents_1": 270, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 32, "L_abs": 1354, "title": "Data Shapley in One Training Run", "abs": "Data Shapley offers a principled framework for attributing the contribution of data within machine learning contexts. However, the traditional notion of Data Shapley requires re-training models on various data subsets, which becomes computationally infeasible for large-scale models. Additionally, this retraining-based definition cannot evaluate the contribution of data for a specific model training run, which may often be of interest in practice. This paper introduces a novel concept, In-Run Data Shapley, which eliminates the need for model retraining and is specifically designed for assessing data contribution for a particular model of interest. In-Run Data Shapley calculates the Shapley value for each gradient update iteration and accumulates these values throughout the training process. We present several techniques that allow the efficient scaling of In-Run Data Shapley to the size of foundation models. In its most optimized implementation, our method adds negligible runtime overhead compared to standard model training. This dramatic efficiency improvement makes it possible to perform data attribution for the foundation model pretraining stage. We present several case studies that offer fresh insights into pretraining data's contribution and discuss their implications for copyright in generative AI and pretraining data curation.", "keywords": [ "Shapley value", "data valuation." ], "rating_list": [ 6, 8, 6, 10 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 4, 3, 3, 4 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "HCoSsULNxG", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Mark Endo", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Xiaohan Wang", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Serena Yeung", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 20, "n_formula_1": 19, "n_ref_uni": 31, "n_ref": 86, "n_ref_all": 95, "n_fig": 2, "n_tab": 3, "L_tab": 650, "n_element_tab": 79, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 925, "n_element_tab_1": 53, "formula_len_all": 973, "formula_len_all_1": 938, "len_all": 118644, "len_all_1": 67485, "len_abs": 1263, "len_title": 115, "len_sents": 36802, "len_sents_1": 30795, "n_sents": 245, "n_sents_1": 212, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1269, "title": "Weakly Supervised Understanding of Skilled Human Activity in Videos", "abs": "Understanding skilled human activity is crucial in fields such as sports analytics, medical training, and professional development, where assessing proficiency can directly influence performance and outcomes. However, many existing approaches rely on human-annotated numerical scores or rankings, which are not only time-consuming but also introduce subjectivity. Conversely, categorizing proficiency as either high or low, though providing less detailed information, is easier to collect and can often be derived from group characteristics such as the distinction between novices and experts in surgical training. This new setting challenges models to uncover intrinsic patterns that reflect proficiency based solely on these weak labels. To achieve this, we introduce Sparse Skill Extractor, a multi-scale contrastive learning framework. It enforces both local and global feature comparisons between groups while pruning irrelevant video segments to highlight key moments of skilled or unskilled performance. Our results demonstrate that Sparse Skill Extractor not only delivers strong performance in predicting demonstrator proficiency but also enhances interpretability by facilitating the detection of non-proficient timestamps for low proficiency demonstrations.", "keywords": [ "skilled human activity understanding", "weakly supervised learning", "action quality assessment", "long-form video understanding" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "HCUksccuFx", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Andrea Vedaldi", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Christian Rupprecht", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" }, { "name": "Chuanxia Zheng", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Ruining Li", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 46, "n_ref": 86, "n_ref_all": 111, "n_fig": 7, "n_tab": 1, "L_tab": 517, "n_element_tab": 13, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1863, "n_element_tab_1": 56, "formula_len_all": 419, "formula_len_all_1": 419, "len_all": 142815, "len_all_1": 67887, "len_abs": 1471, "len_title": 142, "len_sents": 32320, "len_sents_1": 29222, "n_sents": 243, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1368, "title": "PuppetMaster: Scaling Interactive Video Generation as a Motion Prior for Part-Level Dynamics", "abs": "We present PuppetMaster, a video generator that understands part-level object dynamics. Given an image of an object and a number of drags defining the desired trajectory of selected points of the object, PuppetMaster synthesizes a video where the object moves according to the specified drags in a physically plausible manner. PuppetMaster is obtained by fine-tuning an off-the-shelf video diffusion model, extended with a new component that encodes the input drags. PuppetMaster also introduces all-to-first attention, a replacement for the common spatial attention module, which removes artifacts that arise from fine-tuning a video generator out-of-domain and significantly improves the quality of the synthesized videos. PuppetMaster is learned from Objaverse-Animation-HQ, a new dataset of curated part-level motion clips obtained by rendering synthetic 3D animations. We propose strategies to automatically filter out sub-optimal animations and augment the synthetic renderings with meaningful drags. By using this data, PuppetMaster learns to generate part-level motions, unlike other motion-conditioned video generators that mostly move the object as a whole. PuppetMaster generalizes well to real images, outperforming existing methods in real-world benchmarks in a zero-shot manner. We refer the reader to the supplementary material for video visualizations.", "keywords": [ "Video generation", "motion", "diffusion models" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "HCOOQUcWiF", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Ryan Grainger", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "PhD student" }, { "name": "Thomas Paniagua", "gender": "unknown", "institution": "North Carolina State University", "country": "US", "position": "PhD student" }, { "name": "Tianfu Wu", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 31, "n_ref": 75, "n_ref_all": 97, "n_fig": 5, "n_tab": 3, "L_tab": 649, "n_element_tab": 94, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 649, "n_element_tab_1": 94, "formula_len_all": 303, "formula_len_all_1": 304, "len_all": 114346, "len_all_1": 64481, "len_abs": 1894, "len_title": 121, "len_sents": 36715, "len_sents_1": 30504, "n_sents": 263, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1755, "title": "Differentiable Polygon Modeling for Object Instance Segmentation", "abs": "Differentiable polygon (boundary-/contour-based) modeling for object instance segmentation remains an open problem in computer vision and deep learning. It also has been under-explored in the deep learning era, compared with its counterpart, bit-mask (region-based) modeling. In this paper, we present a method of differentiable polygon-based instance segmentation. As commonly done in the prior art, we assume a fixed topology, i.e., the number of vertices, $K$ is predefined and fixed (e.g., $K=250$) in learning and inference. We address two modeling problems: i) The alignment between a predicted $K$-vertex polygon and a target ground-truth $L$-vertex polygon in learning, where $L$ varies significantly. We present PolygonAlign similar in spirit to RoIAlign used in bit-mask-based instance segmentation, which enables using a simple $\\ell_2$ norm as the vertex prediction loss function in learning. ii) The parameterization of a $K$-vertex polygon. We present a variant of the active contour model, which consists of a learnable contour initialization module and an one-step vertex-aware refinement/updating module. The initialization is learned via an affine transformation decoupled vertex regression method. A polygon is parameterized by a translation vector, a rotation transformation matrix, and the vertex displacement vectors. In experiments, the proposed method is tested on the MS-COCO 2017 benchmark using the Sparse R-CNN framework. It obtains state-of-the-art performance compared with the prior art of polygon modeling methods. We also show the empirical upper-bound performance of the proposed method is much higher than all existing instance segmentation methods, which encourages further research on differentiable polygon modeling.", "keywords": [ "Differentiable Polygon Modeling", "Object Instance Segmentation", "PolygonAlign" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "HCJ7B6dhYK", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Alex Saad-Falcon", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Daqian Bao", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Justin Romberg", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 6, "n_ref_uni": 43, "n_ref": 51, "n_ref_all": 73, "n_fig": 6, "n_tab": 6, "L_tab": 1300, "n_element_tab": 166, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 672, "n_element_tab_1": 67, "formula_len_all": 769, "formula_len_all_1": 226, "len_all": 152750, "len_all_1": 57405, "len_abs": 2019, "len_title": 121, "len_sents": 46931, "len_sents_1": 27663, "n_sents": 354, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 2195, "title": "Radon Implicit Field Transform (RIFT): Learning Scenes from Radar Signals", "abs": "Data acquisition in array signal processing (ASP) is costly because achieving high angular and range resolutions necessitates large antenna apertures and wide frequency bandwidths, respectively. The data requirements for ASP problems grow multiplicatively with the number of viewpoints and frequencies, significantly increasing the burden of data collection, even for simulation. Implicit Neural Representations (INRs) \u2014 neural network-based models of 3D objects and scenes \u2014 offer compact and continuous representations with minimal radar data. They can interpolate to unseen viewpoints and potentially address the sampling cost in ASP problems. In this work, we select Synthetic Aperture Radar (SAR) as a case from ASP and propose the \\textit{\\textbf{R}adon \\textbf{I}mplicit \\textbf{F}ield \\textbf{T}ransform} (RIFT). RIFT consists of two components: a classical forward model for radar (Generalized Radon Transform, GRT), and an INR based scene representation learned from radar signals. This method can be extended to other ASP problems by replacing the GRT with appropriate algorithms corresponding to different data modalities. In our experiments, we first synthesize radar data using the GRT. We then train the INR model on this synthetic data by minimizing the reconstruction error of the radar signal. After training, we render the scene using the trained INR and evaluate our scene representation against the ground truth scene. Due to the lack of existing benchmarks, we introduce two main new error metrics: \\textit{\\textbf{p}hase-\\textbf{R}oot \\textbf{M}ean \\textbf{S}quare \\textbf{E}rror} (p-RMSE) for radar signal interpolation, and \\textit{\\textbf{m}agnitude-\\textbf{S}tructural \\textbf{S}imilarity \\textbf{I}ndex \\textbf{M}easure} (m-SSIM) for scene reconstruction. These metrics adapt traditional error measures to account for the complex nature of radar signals. Compared to traditional scene models in radar signal processing, with only 10\\% data footprint, our RIFT model achieves up to 188\\% improvement in scene reconstruction. Using the same amount of data, RIFT is up to $3\\times$ better at reconstruction and shows a 10\\% improvement generalizing to unseen viewpoints.", "keywords": [ "AI for Science", "Representation Learning", "Scene Rendering", "Implicit Neural Representation", "3D Reconstruction", "Inverse Problems" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "HBf6HFnpmH", "primary_area": "causal reasoning", "n_author": 6, "author_info_list": [ { "name": "Juergen Luettin", "gender": "unknown", "institution": "Robert Bosch GmbH", "country": "DE", "position": "Research Scientist" }, { "name": "Kristian Kersting", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Full Professor" }, { "name": "Lavdim Halilaj", "gender": "unknown", "institution": "Universum College", "country": "AL", "position": "Assistant Professor" }, { "name": "Moritz Willig", "gender": "Male", "institution": "Technical University of Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Nicholas Tagliapietra", "gender": "Male", "institution": "CS Department, TU Darmstadt, Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" }, { "name": "Tim Pychynski", "gender": "Male", "institution": "Robert Bosch GmbH, Bosch", "country": "", "position": "Principal Researcher" } ], "n_formula": 33, "n_formula_1": 7, "n_ref_uni": 58, "n_ref": 84, "n_ref_all": 117, "n_fig": 16, "n_tab": 9, "L_tab": 2652, "n_element_tab": 122, "n_fig_1": 13, "n_tab_1": 4, "L_tab_1": 753, "n_element_tab_1": 73, "formula_len_all": 1727, "formula_len_all_1": 394, "len_all": 197498, "len_all_1": 65335, "len_abs": 1501, "len_title": 115, "len_sents": 58699, "len_sents_1": 32516, "n_sents": 433, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1517, "title": "The Challenging Growth: Evaluating the Scalability of Causal Models", "abs": "One of the pillars of causality is the study of causal models and understanding under which hypotheses we can guarantee their ability to grasp causal information and to leverage it for making inferences.\nReal causal phenomena, however, may involve drastically different settings such as high dimensionality, causal insufficiency, and nonlinearities, which can be in stark contrast with the initial assumptions made by most models.\nAdditionally, providing fair benchmarks under such conditions presents challenges due to the lack of realistic data where the true data generating process is known. \nConsequently, most analyses converge towards either small and synthetic toy examples or theoretical analyses, while empirical evidence is limited. \nIn this work, we present in-depth experimental results on two large datasets modeling a real manufacturing scenario. \nWe show the nontrivial behavior of a well-understood manufacturing process, simulated using a physics-based simulator built and validated by domain experts. \nWe demonstrate the inadequacy of many state-of-the-art models and analyze the wide differences in their performance and tractability, both in terms of runtime and memory complexity. \nWe observe that a wide range of causal models are computationally prohibitive for certain tasks, whereas others lack in expressiveness. \nWe release all artefacts to serve as reference for future research on real world applications of causality, including a general web-page and a leader-board for benchmarking.", "keywords": [ "Causality", "benchmark", "causal discovery", "causal inference" ], "rating_list": [ 8, 3, 6, 5 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "HBbbhAZuia", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Bin Feng", "gender": "Male", "institution": "International Digital Economy Academy, International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Irwin King", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Mingjun Yang", "gender": "unknown", "institution": "XtalPi", "country": "", "position": "Researcher" }, { "name": "Xinni Zhang", "gender": "Female", "institution": "Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yankai Chen", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Postdoc" }, { "name": "Yu Li", "gender": "Male", "institution": "International Digital Economy Academy", "country": "CN", "position": "Principal Researcher" }, { "name": "Zijing Liu", "gender": "unknown", "institution": "International Digital Economy Academy", "country": "CN", "position": "Researcher" }, { "name": "Zenglin Xu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 64, "n_ref": 116, "n_ref_all": 162, "n_fig": 21, "n_tab": 9, "L_tab": 6478, "n_element_tab": 476, "n_fig_1": 16, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 260, "formula_len_all_1": 201, "len_all": 176579, "len_all_1": 53347, "len_abs": 1385, "len_title": 149, "len_sents": 43085, "len_sents_1": 25419, "n_sents": 339, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1088, "title": "DockedAC: Empowering Deep Learning Models With 3D Protein-ligand Data For Activity Cliff Analysis", "abs": "Artificial intelligence has become a crucial tool in drug discovery, excelling in tasks such as molecular property prediction. An activity cliff, which refers to a minor structural modification to a molecule resulting in a large change in its biological activity, poses a challenge in predictive modeling. The activity cliff depends on the interaction between the target and the ligand, which is however largely overlooked by previous ligand-centric studies. In this paper, we introduce DockedAC, a new dataset incorporating the protein target and target-ligand 3D complex structure information for studying the problem of activity cliffs. By matching protein binding information and ligand bioactivity, we employ molecular docking to generate the complex structure for each activity value. The DockedAC dataset contains 82,836 activity data on 52 protein targets with activity cliff annotations, which serves as the first step towards activity cliff research with large-scale 3D complex structures. We benchmark the dataset with traditional machine learning and deep learning approaches.", "keywords": [ "Activity cliff prediction", "Molecular property prediction", "AI-aided drug discovery" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 4, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "HB4lr0ykTi", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Aram-Alexandre Pooladian", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Brandon Amos", "gender": "unknown", "institution": "Meta AI", "country": "US", "position": "Research Scientist" }, { "name": "Dana Pe'er", "gender": "Female", "institution": "Columbia University", "country": "", "position": "Associate Professor" }, { "name": "Doron Haviv", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 29, "n_ref_uni": 64, "n_ref": 101, "n_ref_all": 131, "n_fig": 10, "n_tab": 4, "L_tab": 533, "n_element_tab": 35, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 359, "n_element_tab_1": 36, "formula_len_all": 3930, "formula_len_all_1": 2864, "len_all": 182238, "len_all_1": 61653, "len_abs": 1321, "len_title": 125, "len_sents": 47876, "len_sents_1": 26839, "n_sents": 351, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1376, "title": "Wasserstein Flow Matching: Generative modeling over families of distributions", "abs": "Generative modeling typically concerns the transport of a single source distribution to a single target distribution by learning (i.e., regressing onto) simple probability flows. However, in modern data-driven fields such as computer graphics and single-cell genomics, samples (say, point-clouds) from datasets can themselves be viewed as distributions (as, say, discrete measures). In these settings, the standard generative modeling paradigm of flow matching would ignore the relevant geometry of the samples. To remedy this, we propose \\emph{Wasserstein flow matching} (WFM), which appropriately lifts flow matching onto families of distributions by appealing to the Riemannian nature of the Wasserstein geometry. Our algorithm leverages theoretical and computational advances in (entropic) optimal transport, as well as the attention mechanism in our neural network architecture. We present two novel algorithmic contributions. First, we demonstrate how to perform generative modeling over Gaussian distributions, where we generate representations of granular cell states from single-cell genomics data. Secondly, we show that WFM can learn flows between high-dimensional and variable sized point-clouds and synthesize cellular microenvironments from spatial transcriptomics datasets. Code is available at https://github.com/WassersteinFlowMatching/WassersteinFlowMatching", "keywords": [ "Optimal Transport", "Flow Matching", "Point Clouds", "Generative Modeling", "Single Cell" ], "rating_list": [ 6, 8, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "HAwZGLcye3", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Alexander Marson", "gender": "Male", "institution": "University of California, San Francisco", "country": "US", "position": "Full Professor" }, { "name": "Andrew H. Lee", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Jian Vora", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "", "position": "MS student" }, { "name": "Jure Leskovec", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Kexin Huang", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Percy Liang", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Qian Huang", "gender": "Female", "institution": "Google", "country": "", "position": "Intern" }, { "name": "Yusuf H Roohani", "gender": "Male", "institution": "Arc Institute", "country": "US", "position": "Principal Researcher" }, { "name": "Zachary Steinhart", "gender": "unknown", "institution": "University of California, San Francisco", "country": "US", "position": "Postdoc" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 54, "n_ref": 221, "n_ref_all": 276, "n_fig": 20, "n_tab": 15, "L_tab": 12252, "n_element_tab": 974, "n_fig_1": 22, "n_tab_1": 2, "L_tab_1": 906, "n_element_tab_1": 96, "formula_len_all": 28, "formula_len_all_1": 28, "len_all": 273932, "len_all_1": 66922, "len_abs": 2396, "len_title": 125, "len_sents": 80118, "len_sents_1": 33665, "n_sents": 581, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1847, "title": "BioDiscoveryAgent: An AI Agent for Designing Genetic Perturbation Experiments", "abs": "Agents based on large language models have shown great potential in accelerating scientific discovery by leveraging their rich background knowledge and reasoning capabilities. In this paper, we introduce BioDiscoveryAgent, an agent that designs new experiments, reasons about their outcomes, and efficiently navigates the hypothesis space to reach desired solutions. We demonstrate our agent on the problem of designing genetic perturbation experiments, where the aim is to find a small subset out of many possible genes that, when perturbed, result in a specific phenotype (e.g., cell growth). Utilizing its biological knowledge, BioDiscoveryAgent can uniquely design new experiments without the need to train a machine learning model or explicitly design an acquisition function as in Bayesian optimization. Moreover, BioDiscoveryAgent using Claude 3.5 Sonnet achieves an average of 21% improvement in predicting relevant genetic perturbations across six datasets, and a 46% improvement in the harder task of non-essential gene perturbation, compared to existing Bayesian optimization baselines specifically trained for this task. Our evaluation includes one dataset that is unpublished, ensuring it is not part of the language model's training data. Additionally, BioDiscoveryAgent predicts gene combinations to perturb more than twice as accurately as a random baseline, a task so far not explored in the context of closed-loop experiment design. The agent also has access to tools for searching the biomedical literature, executing code to analyze biological datasets, and prompting another agent to critically evaluate its predictions. Overall, BioDiscoveryAgent is interpretable at every stage, representing an accessible new paradigm in the computational design of biological experiments with the potential to augment scientists' efficacy.", "keywords": [ "large language models", "agents", "computational biology", "genomics", "AI for scientific discovery" ], "rating_list": [ 8, 6, 8, 5, 5 ], "soundness_list": [ 4, 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 3 ] }, { "paper_id": "HAD6iZxKuh", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Hai Ci", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Jinheng Xie", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Pei Yang", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yiren Song", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zheng Shou", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 43, "n_ref": 136, "n_ref_all": 167, "n_fig": 12, "n_tab": 7, "L_tab": 757, "n_element_tab": 121, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 541, "n_element_tab_1": 69, "formula_len_all": 151, "formula_len_all_1": 136, "len_all": 133485, "len_all_1": 62888, "len_abs": 904, "len_title": 160, "len_sents": 37789, "len_sents_1": 28067, "n_sents": 319, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 911, "title": "WMAdapter: Adding WaterMark Control to Latent Diffusion Models", "abs": "Watermarking is essential for protecting the copyright of AI-generated images. We propose WMAdapter, a diffusion model watermark plugin that embeds user-specified watermark information seamlessly during the diffusion generation process. Unlike previous methods that modify diffusion modules to incorporate watermarks, WMAdapter is designed to keep all diffusion components intact, resulting in sharp, artifact-free images. To achieve this, we introduce two key innovations: (1) We develop a contextual adapter that conditions on the content of the cover image to generate adaptive watermark embeddings. (2) We implement an additional finetuning step and a hybrid finetuning strategy that suppresses noticeable artifacts while preserving the integrity of the diffusion components. \n Empirical results show that WMAdapter provides strong flexibility, superior image quality, and competitive watermark robustness.", "keywords": [ "Diffusion Model", "Watermarking" ], "rating_list": [ 3, 5, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3, 3 ], "confidence_list": [ 5, 5, 5, 4, 4 ] }, { "paper_id": "HA0oLUvuGI", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Shiyuan Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Weitong ZHANG", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" } ], "n_formula": 61, "n_formula_1": 32, "n_ref_uni": 25, "n_ref": 104, "n_ref_all": 130, "n_fig": 11, "n_tab": 6, "L_tab": 2969, "n_element_tab": 330, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1412, "n_element_tab_1": 22, "formula_len_all": 8916, "formula_len_all_1": 2910, "len_all": 170909, "len_all_1": 67321, "len_abs": 1183, "len_title": 112, "len_sents": 52289, "len_sents_1": 28235, "n_sents": 434, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1263, "title": "Energy-Weighted Flow Matching for Offline Reinforcement Learning", "abs": "This paper investigates energy guidance in generative modeling, where the target distribution is defined as $q(\\mathbf x) \\propto p(\\mathbf x)\\exp(-\\beta \\mathcal E(\\mathbf x))$, with $p(\\mathbf x)$ being the data distribution and $\\mathcal E(\\mathbf x)$ as the energy function. To comply with energy guidance, existing methods often require auxiliary procedures to learn intermediate guidance during the diffusion process. To overcome this limitation, we explore energy-guided flow matching, a generalized form of the diffusion process. We introduce energy-weighted flow matching (EFM), a method that directly learns the energy-guided flow without the need for auxiliary models. Theoretical analysis shows that energy-weighted flow matching accurately captures the guided flow. Additionally, we extend this methodology to energy-weighted diffusion models and apply it to offline reinforcement learning (RL) by proposing the Q-weighted Iterative Policy Optimization (QIPO). Empirically, we demonstrate that the proposed QIPO algorithm improves performance in offline RL tasks. Notably, our algorithm is the first energy-guided diffusion model that operates independently of auxiliary models and the first exact energy-guided flow matching model in the literature.", "keywords": [ "Flow Matching Models", "Diffusion Models", "Energy-guidance Generative Models", "Offline Reinforcement Learning" ], "rating_list": [ 6, 6, 8, 5 ], "soundness_list": [ 3, 2, 4, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 4, 1 ], "confidence_list": [ 4, 3, 3, 5 ] }, { "paper_id": "H9plefjzuR", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "LEI BAI", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Pumeng Lyu", "gender": "unknown", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Wanli Ouyang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 16, "n_ref_uni": 59, "n_ref": 126, "n_ref_all": 148, "n_fig": 5, "n_tab": 4, "L_tab": 706, "n_element_tab": 79, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 838, "n_element_tab_1": 42, "formula_len_all": 1265, "formula_len_all_1": 1117, "len_all": 152518, "len_all_1": 56822, "len_abs": 1304, "len_title": 173, "len_sents": 34729, "len_sents_1": 24240, "n_sents": 267, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 125, "L_abs": 1310, "title": "Transforming Ocean Analysis: Learning 4D ocean field from in-situ observations via uncertainty-aware implicit representations", "abs": "A complete and accurate representation of Earth's time-evolving ocean field is crucial for understanding global warming as well as climate dynamics. However, the sparsity of current in-situ ocean measurements presents a significant challenge in estimating values in largely unobserved regions. Traditional methods, such as objective interpolation (OI), struggle with accuracy due to their reliance on discrete grids and fixed spatial correlation structures. In this paper, we propose a novel approach to reconstruct 4D ocean fields only from raw observations using implicit neural representations (INRs). Our method improves field representations by leveraging neural networks to capture continuous, complex, and nonlinear patterns inherent in ocean data. To address uncertainties in ocean measurements and the limited availability of daily observations, we incorporate uncertainty estimates and a meta-learning strategy into existing INRs. These innovations enable our approach to provide daily, resolution-free ocean temperature reconstructions, a significant improvement over monthly averaged discrete fields. Experiments demonstrate the accuracy and adaptability of our method compared with approaches, establishing our method as a transformative solution for future ocean analysis and climate monitoring.", "keywords": [ "implicit neural representation", "meta learning", "climate", "ocean gridded dataset" ], "rating_list": [ 3, 3, 5, 6, 3 ], "soundness_list": [ 2, 1, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "H9oYYou34X", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Aleksandr Beznosikov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "Associate Professor" }, { "name": "Andrey Veprikov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "MS student" }, { "name": "Mikhail Rudakov", "gender": "unknown", "institution": "Innopolis University", "country": "RU", "position": "Undergrad student" }, { "name": "Vladimir Solodkin", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "MS student" } ], "n_formula": 196, "n_formula_1": 30, "n_ref_uni": 52, "n_ref": 95, "n_ref_all": 130, "n_fig": 6, "n_tab": 3, "L_tab": 3214, "n_element_tab": 28, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 681, "n_element_tab_1": 2, "formula_len_all": 31870, "formula_len_all_1": 3209, "len_all": 240156, "len_all_1": 70215, "len_abs": 1217, "len_title": 118, "len_sents": 58730, "len_sents_1": 30454, "n_sents": 616, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1223, "title": "Markovian Compression: Looking to the Past Helps Accelerate the Future", "abs": "This paper deals with distributed optimization problems that use compressed communication to achieve efficient performance and mitigate the communication bottleneck. We propose a family of compression schemes in which operators transform vectors fed to their input according to a Markov chain, i.e., the stochasticity of the compressors depends on previous iterations. Intuitively, this should accelerate the convergence of optimization methods, as considering previous iterations seems more natural and robust. The compressors are implemented in the vanilla Quantized Stochastic Gradient Descent (QSGD) algorithm. To further improve efficiency and convergence rate, we apply the momentum acceleration method. We prove convergence results for our algorithms with Markovian compressors and show theoretically that the accelerated method converges faster than the basic version. The analysis covers non-convex, Polyak-Lojasiewicz (PL), and strongly convex cases. Experiments are conducted to demonstrate the applicability of the results to distributed data-parallel optimization problems. Practical results demonstrate the superiority of methods utilizing our compressors design over several existing optimization algorithms.", "keywords": [ "stochastic optimization", "distributed optimization", "compressed communications", "Markovian noise" ], "rating_list": [ 3, 5, 5, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 2, 3, 2, 3 ] }, { "paper_id": "H9dNX6TaRE", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Binh T. Nguyen", "gender": "Male", "institution": "Ho Chi Minh city University of Science, Vietnam National University", "country": "", "position": "Associate Professor" }, { "name": "Nguyen Minh Phuc", "gender": "Male", "institution": "VinUni", "country": "VN", "position": "Researcher" }, { "name": "Khoa Doan", "gender": "unknown", "institution": "VinUniversity", "country": "VN", "position": "Assistant Professor" }, { "name": "Hi\u1ebfu Ng\u1ecdc Nguy\u1ec5n", "gender": "Male", "institution": "VinUniversity", "country": "VN", "position": "Research Assistant" } ], "n_formula": 25, "n_formula_1": 17, "n_ref_uni": 25, "n_ref": 74, "n_ref_all": 83, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2315, "formula_len_all_1": 1425, "len_all": 194326, "len_all_1": 59346, "len_abs": 1353, "len_title": 151, "len_sents": 42157, "len_sents_1": 27208, "n_sents": 290, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 1347, "title": "Mitigating Reward Over-optimization in Direct Alignment Algorithms with Adaptive Importance Sampling", "abs": "Recently, Direct Alignment Algorithms (DAAs) such as Direct Preference Optimization (DPO) have emerged as alternatives to the standard Reinforcement learning from human feedback (RLHF) for aligning large language models (LLMs) with human values. Surprisingly, while DAAs do not use a separate proxy reward model as in RLHF, their performance can still deteriorate due to over-optimization \u2013 a phenomenon found in RLHF where the policy can exploit failures of the reward model to achieve high rewards but the actual quality of the model begins to degrade. Recent studies find that DAAs tend to increase probability mass on out-of-distribution responses and the training objective in DAAs is heavily under-constrained on these out-of-distribution (OOD) responses due to a mismatch between offline distribution and the LM policy. In this paper, we propose a method to mitigate the distribution shift between the offline distribution and the LM policy by multiplying with an importance weight to reflect the policy distribution. The resulting method, called Adaptive Importance Sampling (AIS), relies on importance sampling techniques and resolves the high variance issue in importance sampling without extra hyper-parameters. Our experiment results showed Adaptive IS can improve win rates by 15% while maintaining a lower KL budget compared to DAAs.", "keywords": [ "Reinforcement Learning From Human Feedback", "Direct Preference Optimization", "Reward Hacking" ], "rating_list": [ 3, 8, 3, 6 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "H9UnNgdq0g", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Mahdi Soltanolkotabi", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Full Professor" }, { "name": "Maryam Soltanolkotabi", "gender": "Female", "institution": "University of Utah", "country": "US", "position": "Assistant Professor" }, { "name": "Mohammad Shahab Sepehri", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Zalan Fabian", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Postdoc" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 81, "n_ref_all": 101, "n_fig": 8, "n_tab": 2, "L_tab": 1842, "n_element_tab": 135, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2252, "n_element_tab_1": 135, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 158876, "len_all_1": 62766, "len_abs": 1429, "len_title": 162, "len_sents": 51306, "len_sents_1": 30567, "n_sents": 351, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 113, "L_abs": 1434, "title": "MediConfusion: Can you trust your AI radiologist? Probing the reliability of multimodal medical foundation models", "abs": "Multimodal Large Language Models (MLLMs) have tremendous potential to improve the accuracy, availability, and cost-effectiveness of healthcare by providing automated solutions or serving as aids to medical professionals. Despite promising first steps in developing medical MLLMs in the past few years, their capabilities and limitations are not well understood. Recently, many benchmark datasets have been proposed that test the general medical knowledge of such models across a variety of medical areas. However, the systematic failure modes and vulnerabilities of such models are severely underexplored with most medical benchmarks failing to expose the shortcomings of existing models in this safety-critical domain. In this paper, we introduce MediConfusion, a challenging medical Visual Question Answering (VQA) benchmark dataset, that probes the failure modes of medical MLLMs from a vision perspective. We reveal that state-of-the-art models are easily confused by image pairs that are otherwise visually dissimilar and clearly distinct for medical experts. Strikingly, all available models (open-source or proprietary) achieve performance below random guessing on MediConfusion, raising serious concerns about the reliability of existing medical MLLMs for healthcare deployment. We also extract common patterns of model failure that may help the design of a new generation of more trustworthy and reliable MLLMs in healthcare.", "keywords": [ "Medical foundation models", "Benchmarking", "Vision encoding", "Radiology" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 4, 2, 4, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "H8oCwBTDMv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Chonghua Han", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Depeng Jin", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Guozhen Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Jingtao Ding", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yuan Yuan", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 4, "n_ref_uni": 41, "n_ref": 123, "n_ref_all": 140, "n_fig": 7, "n_tab": 11, "L_tab": 3783, "n_element_tab": 629, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2630, "n_element_tab_1": 353, "formula_len_all": 886, "formula_len_all_1": 442, "len_all": 178312, "len_all_1": 68393, "len_abs": 1947, "len_title": 122, "len_sents": 48202, "len_sents_1": 29436, "n_sents": 363, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1962, "title": "UrbanDiT: A Foundation Model for Open-World Urban Spatio-Temporal Learning", "abs": "The urban environment is characterized by complex spatio-temporal dynamics arising from diverse human activities and interactions. Effectively modeling these dynamics is essential for understanding and optimizing urban systems. In this work, we introduce UrbanDiT, a foundation model for open-world urban spatio-temporal learning that successfully scale up diffusion transformers in this field. UrbanDiT pioneers a unified model that integrates diverse spatio-temporal data sources and types while learning universal spatio-temporal patterns across different cities and scenarios. This allows the model to unify both multi-data and multi-task learning, and effectively support a wide range of spatio-temporal applications. Its key innovation lies in the elaborated prompt learning framework, which adaptively generates both data-driven and task-specific prompts, guiding the model to deliver superior performance across various urban applications.\n\n\nUrbanDiT offers three primary advantages: 1) It unifies diverse data types, such as grid-based and graph-based data, into a sequential format, allowing to capture spatio-temporal dynamics across diverse scenarios of different cities; 2) With masking strategies and task-specific prompts, it supports a wide range of tasks, including bi-directional spatio-temporal prediction, temporal interpolation, spatial extrapolation, and spatio-temporal imputation; and 3) It generalizes effectively to open-world scenarios, with its powerful zero-shot capabilities outperforming nearly all baselines with training data. These features allow UrbanDiT to achieves state-of-the-art performance in different domains such as transportation traffic, crowd flows, taxi demand, bike usage, and cellular traffic, across multiple cities and tasks. UrbanDiT sets up a new benchmark for foundation models in the urban spatio-temporal domain. Code and datasets are publicly available at https://anonymous.4open.science/r/UrbanDiT.", "keywords": [ "Foundation model", "Urban Spatio-Temporal Learning", "Diffusion Transformer", "Prompt Learning" ], "rating_list": [ 3, 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 1, 1, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 4, 3 ] }, { "paper_id": "H8hO3T3DYe", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Anming Gu", "gender": "Male", "institution": "Boston University", "country": "US", "position": "Researcher" }, { "name": "Edward Chien", "gender": "Male", "institution": "Boston University", "country": "", "position": "Assistant Professor" }, { "name": "Kristjan Greenewald", "gender": "unknown", "institution": "MIT-IBM Watson AI Lab, IBM Research", "country": "", "position": "Research Scientist" } ], "n_formula": 122, "n_formula_1": 25, "n_ref_uni": 44, "n_ref": 151, "n_ref_all": 194, "n_fig": 12, "n_tab": 2, "L_tab": 54, "n_element_tab": 4, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9908, "formula_len_all_1": 1783, "len_all": 243910, "len_all_1": 73662, "len_abs": 1271, "len_title": 133, "len_sents": 73295, "len_sents_1": 32133, "n_sents": 763, "n_sents_1": 268, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1297, "title": "Partially Observed Trajectory Inference using Optimal Transport and a Dynamics Prior", "abs": "Trajectory inference seeks to recover the temporal dynamics of a population from snapshots of its (uncoupled) temporal marginals, i.e. where observed particles are \\emph{not} tracked over time. Prior works addressed this challenging problem under a stochastic differential equation (SDE) model with a gradient-driven drift in the observed space, introducing a minimum entropy estimator relative to the Wiener measure and a practical grid-free mean-field Langevin (MFL) algorithm using Schr\\\"odinger bridges. Motivated by the success of observable state space models in the traditional paired trajectory inference problem (e.g. target tracking), we extend the above framework to a class of latent SDEs in the form of \\emph{observable state space models}. \n In this setting, we use partial observations to infer trajectories in the latent space under a specified dynamics model (e.g. the constant velocity/acceleration models from target tracking). We introduce the PO-MFL algorithm to solve this latent trajectory inference problem and provide theoretical guarantees to the partially observed setting. Experiments validate the robustness of our method and the exponential convergence of the MFL dynamics, and demonstrate significant outperformance over the latent-free baseline in key scenarios.", "keywords": [ "optimal transport", "trajectory inference", "stochastic calculus", "optimization", "Langevin dynamics" ], "rating_list": [ 8, 6, 3 ], "soundness_list": [ 4, 2, 3 ], "presentation_list": [ 3, 3, 1 ], "contribution_list": [ 4, 2, 2 ], "confidence_list": [ 2, 3, 2 ] }, { "paper_id": "H8gLQwg1eC", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Shawn Im", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Yixuan Li", "gender": "Female", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 59, "n_formula_1": 30, "n_ref_uni": 75, "n_ref": 137, "n_ref_all": 152, "n_fig": 5, "n_tab": 2, "L_tab": 103, "n_element_tab": 11, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3747, "formula_len_all_1": 1572, "len_all": 227261, "len_all_1": 68893, "len_abs": 1224, "len_title": 124, "len_sents": 48522, "len_sents_1": 31939, "n_sents": 427, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1229, "title": "Understanding Generalization of Preference Optimization Under Noisy Feedback", "abs": "As large language models (LLMs) advance their capabilities, aligning these models with human preferences has become crucial. Preference optimization, which trains models to distinguish between preferred and non-preferred responses based on human feedback, has become a crucial component for aligning LLMs. However, most existing works assume noise-free feedback, which is unrealistic given the inherent errors and inconsistencies in human judgments. This paper addresses the impact of noisy feedback on preference optimization, providing generalization guarantees under these conditions. Unlike traditional analyses that assume convergence, our work focuses on finite-step preference optimization, offering new insights that are more aligned with practical LLM training. We establish generalization guarantees for noisy preference learning under a broad family of preference optimization losses such as DPO, IPO, SLiC, etc. Our analysis provides the basis for a general model that closely describes how the generalization decays with the noise rate. Empirical validation on contemporary LLMs confirms the practical relevance of our findings, offering valuable insights for developing AI systems that align with human preferences.", "keywords": [ "Preference optimization", "noisy feedback" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "H8QvefExFf", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Chen Zhu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Jiawei Huang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Ke Lei", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Luping Liu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Rongjie Huang", "gender": "Male", "institution": "FAIR", "country": "US", "position": "Researcher" }, { "name": "Shengpeng Ji", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Tao Jin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaoda Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Xize Cheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zehan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhenhui Ye", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 31, "n_ref": 73, "n_ref_all": 81, "n_fig": 4, "n_tab": 4, "L_tab": 816, "n_element_tab": 117, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 749, "n_element_tab_1": 104, "formula_len_all": 19, "formula_len_all_1": 19, "len_all": 122854, "len_all_1": 55004, "len_abs": 435, "len_title": 147, "len_sents": 35196, "len_sents_1": 25514, "n_sents": 273, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 99, "L_abs": 1752, "title": "T2A-Feedback: Improving Basic Capabilities of Text-to-Audio Generation via Fine-grained AI Feedback", "abs": "Text-to-audio (T2A) generation has achieved remarkable progress in generating a variety of audio outputs from language prompts. However, current state-of-the-art T2A models still struggle to satisfy human preferences for prompt-following and acoustic quality when generating complex multi-event audio. To improve the performance of the model in these high-level applications, we propose to enhance the basic capabilities of the model with AI feedback learning. First, we introduce fine-grained AI audio scoring pipelines to: 1) verify whether each event in the text prompt is present in the audio (Event Occurrence Score), 2) detect deviations in event sequences from the language description (Event Sequence Score), and 3) assess the overall acoustic and harmonic quality of the generated audio (Acoustic & Harmonic Quality). We evaluate these three automatic scoring pipelines and find that they correlate significantly better with human preferences than other evaluation metrics. This highlights their value as both feedback signals and evaluation metrics. Utilizing our robust scoring pipelines, we construct a large audio preference dataset, T2A-FeedBack, which contains 41k prompts and 249k audios, each accompanied by detailed scores. Moreover, we introduce T2A-EpicBench, a benchmark that focuses on long captions, multi-events, and story-telling scenarios, aiming to evaluate the advanced capabilities of T2A models. Finally, we demonstrate how T2A-FeedBack can enhance current state-of-the-art audio model. With simple preference tuning, the audio generation model exhibits significant improvements in both simple (AudioCaps test set) and complex (T2A-EpicBench) scenarios. The project page is available at \\url{https://T2Afeedback.github.io}", "keywords": [ "text-to-audio generation; preference learning" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "H8C4lGZOOE", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Haobo Wang", "gender": "unknown", "institution": "Hangzhou Levinthal Biotech", "country": "CN", "position": "Principal Researcher" }, { "name": "Jiawei Guan", "gender": "Male", "institution": "Hangzhou Levinthal Biotech", "country": "CN", "position": "Researcher" }, { "name": "Rui Ma", "gender": "Male", "institution": "Hangzhou Levinthal Biotech", "country": "CN", "position": "Researcher" }, { "name": "Wei Qu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Weikun.Wu", "gender": "Male", "institution": "Hangzhou Levinthal Biotech", "country": "CN", "position": "Principal Researcher" }, { "name": "kezhai", "gender": "Male", "institution": "Hangzhou Levinthal Biothch", "country": "CN", "position": "Researcher" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 34, "n_ref": 52, "n_ref_all": 70, "n_fig": 8, "n_tab": 5, "L_tab": 2044, "n_element_tab": 180, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4639, "formula_len_all_1": 1503, "len_all": 155506, "len_all_1": 54525, "len_abs": 1257, "len_title": 101, "len_sents": 43271, "len_sents_1": 26550, "n_sents": 300, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 1302, "title": "P(all-atom) Is Unlocking New Path For Protein Design", "abs": "We introduce Pallatom, an innovative protein generation model capable of producing protein structures with all-atom coordinates. Pallatom directly learns and models the joint distribution $P(\\textit{structure}, \\textit{seq})$ by focusing on $P(\\textit{all-atom})$, effectively addressing the interdependence between sequence and structure in protein generation. To achieve this, we propose a novel network architecture specifically designed for all-atom protein generation. Our model employs a dual-track framework that tokenizes proteins into token-level and atomic-level representations, integrating them through a multi-layer decoding process with \"traversing\" representations and recycling mechanism. We also introduce the $\\texttt{atom14}$ representation method, which unifies the description of unknown side-chain coordinates, ensuring high fidelity between the generated all-atom conformation and its physical structure. Experimental results demonstrate that Pallatom excels in key metrics of protein design, including designability, diversity, and novelty, showing significant improvements across the board. Our model not only enhances the accuracy of protein generation but also exhibits excellent training efficiency, paving the way for future applications in larger and more complex systems.", "keywords": [ "Proteins", "Generative models", "Co-design", "All-atom" ], "rating_list": [ 6, 3, 3, 3, 5 ], "soundness_list": [ 3, 2, 1, 2, 3 ], "presentation_list": [ 3, 2, 1, 1, 1 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 4, 2 ] }, { "paper_id": "H6i47PKXSN", "primary_area": "applications to robotics, autonomy, planning", "n_author": 8, "author_info_list": [ { "name": "Jian Kuai", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Jingdong Wang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Chief Scientist for Computer Vision" }, { "name": "Junjun Jiang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Kui Jiang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Shi Gong", "gender": "Male", "institution": "Baidu", "country": "", "position": "Researcher" }, { "name": "Xiao Tan", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Xiaoqing Ye", "gender": "Female", "institution": "Baidu Inc.", "country": "CN", "position": "Researcher and Developer" }, { "name": "wenbo zhao", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 30, "n_ref": 64, "n_ref_all": 77, "n_fig": 6, "n_tab": 7, "L_tab": 2210, "n_element_tab": 300, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1569, "n_element_tab_1": 194, "formula_len_all": 721, "formula_len_all_1": 721, "len_all": 113579, "len_all_1": 59299, "len_abs": 1611, "len_title": 99, "len_sents": 29242, "len_sents_1": 25803, "n_sents": 235, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1618, "title": "GAP: Scalable Driving with Generative Aided Planner", "abs": "The primary challenge in end-to-end autonomous driving lines in how to establish robust environmental perception and representations. While most methods improve these capabilities by introducing auxiliary perception tasks, the process of obtaining precise large-scale annotations in this paradigm is both time-consuming and laborious, thereby limiting the scalability and practical application. To address this, we propose an architecture based on the Generative Aided Planner (GAP), which integrates scene generation and planning within a single framework. To compensate for the information loss in discrete image features, we design a dual-branch image encoder that fuses continuous and discrete features, improving the model's ability to recognize traffic lights. Through the scene generation task from input tokens, our approach learns the intrinsic dependencies between tokens and environments, which in turn benefits the planning task. It is important to note that the generative model is trained in a fully self-supervised manner, requiring no perception annotations. Our model is built upon GPT-2, which exhibits scaling laws similar to those observed in other GPTs: as we increase the model size and data size, the performance shows continuous and non-saturating improvements. Experiments show that among methods using the front view as input, our approach outperforms other methods that employ multiple perception supervision in the CARLA simulator. Our method is simple yet highly effective, offering a promising direction for scalable and practical deployment of autonomous vehicles in real-world settings.", "keywords": [ "end-to-end autonomous driving;generative;planning;perception" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "H6UMc5VS70", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Bryan Hooi", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Jinlan Fu", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Miao Xiong", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Shumin Deng", "gender": "Female", "institution": "National University of Singapore", "country": "", "position": "Postdoc" }, { "name": "Xiaoxin He", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yue Liu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 73, "n_ref": 316, "n_ref_all": 385, "n_fig": 26, "n_tab": 15, "L_tab": 9507, "n_element_tab": 1086, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1424, "n_element_tab_1": 197, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 341123, "len_all_1": 60064, "len_abs": 5717, "len_title": 87, "len_sents": 90438, "len_sents_1": 27136, "n_sents": 845, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1101, "title": "FlipAttack: Jailbreak LLMs via Flipping", "abs": "This paper proposes a simple yet effective jailbreak attack named FlipAttack against black-box LLMs. First, from the autoregressive nature, we reveal that LLMs tend to understand the text from left to right and find that they struggle to comprehend the text when noise is added to the left side. Motivated by these insights, we propose to disguise the harmful prompt by constructing left-side noise merely based on the prompt itself, then generalize this idea to 4 flipping modes. Second, we verify the strong ability of LLMs to perform the text-flipping task, and then develop 4 variants to guide LLMs to denoise, understand, and execute harmful behaviors accurately. These designs keep FlipAttack universal, stealthy, and simple, allowing it to jailbreak black-box LLMs within only 1 query. Experiments on 8 LLMs demonstrate the superiority of FlipAttack. Remarkably, it achieves $\\sim$98\\% attack success rate on GPT-4o, and $\\sim$98\\% bypass rate against 5 guardrail models on average. The codes are available at Anonymous GitHub\\footnote{https://anonymous.4open.science/r/ICLR25-1731-FlipAttack}.", "keywords": [ "Large Language Model", "AI Safety", "Red Teaming" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 4, 4, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "H6DpBnPCyH", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Chijie An", "gender": "Male", "institution": "New York University", "country": "US", "position": "Undergrad student" }, { "name": "Jiacheng Shen", "gender": "unknown", "institution": "New York University", "country": "US", "position": "Undergrad student" }, { "name": "Kai Shao", "gender": "Male", "institution": "New York University Shanghai", "country": "CN", "position": "Researcher" }, { "name": "Mathieu Lauriere", "gender": "unknown", "institution": "New York University", "country": "FR", "position": "Assistant Professor" } ], "n_formula": 75, "n_formula_1": 32, "n_ref_uni": 42, "n_ref": 67, "n_ref_all": 85, "n_fig": 10, "n_tab": 4, "L_tab": 1323, "n_element_tab": 95, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 11140, "formula_len_all_1": 3525, "len_all": 221949, "len_all_1": 77723, "len_abs": 5431, "len_title": 109, "len_sents": 61838, "len_sents_1": 30216, "n_sents": 726, "n_sents_1": 318, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1030, "title": "Reinforcement Learning for Finite Space Mean-Field Type Games", "abs": "Mean field type games (MFTGs) describe Nash equilibria between large coalitions: each coalition consists of a continuum of cooperative agents who maximize the average reward of their coalition while interacting non-cooperatively with a finite number of other coalitions. Although the theory has been extensively developed, we are still lacking efficient and scalable computational methods. Here, we develop reinforcement learning methods for such games in a finite space setting with general dynamics and reward functions. We start by proving that MFTG solution yields approximate Nash equilibria in finite-size coalition games. We then propose two algorithms. The first is based on quantization of mean-field spaces and Nash Q-learning. We provide convergence and stability analysis. We then propose a deep reinforcement learning algorithm, which can scale to larger spaces. Numerical experiments in 5 environments with mean-field distributions of dimension up to $200$ show the scalability and efficiency of the proposed method.", "keywords": [ "mean field type games", "deep reinforcement learning", "Nash Q learning" ], "rating_list": [ 8, 8, 3, 3 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "H5FUVj0vMd", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Jiaya Jia", "gender": "Male", "institution": "Department of Computer Science and Engineering, Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Senqiao Yang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xiangru Peng", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Xin Lai", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Yukang Chen", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Zhuotao Tian", "gender": "Male", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 39, "n_ref": 54, "n_ref_all": 83, "n_fig": 7, "n_tab": 6, "L_tab": 2314, "n_element_tab": 188, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 96, "n_element_tab_1": 6, "formula_len_all": 392, "formula_len_all_1": 314, "len_all": 150843, "len_all_1": 52435, "len_abs": 1556, "len_title": 124, "len_sents": 35183, "len_sents_1": 23401, "n_sents": 355, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1563, "title": "Step-DPO: Step-wise Preference Optimization for Long-chain Reasoning of LLMs", "abs": "Mathematical reasoning presents a significant challenge for Large Language Models (LLMs) due to the extensive and precise chain of reasoning required for accuracy. Ensuring the correctness of each reasoning step is critical. To address this, we aim to enhance the robustness and factuality of LLMs by learning from human feedback. However, Direct Preference Optimization (DPO) has shown limited benefits for long-chain mathematical reasoning, as models employing DPO struggle to identify detailed errors in incorrect answers. This limitation stems from a lack of fine-grained process supervision. We propose a simple, effective, and data-efficient method called Step-DPO, which treats individual reasoning steps as units for preference optimization rather than evaluating answers holistically. Additionally, we have developed a data construction pipeline for Step-DPO, enabling the creation of a high-quality dataset containing 10K step-wise preference pairs. We also observe that in DPO, the data generated by the policy model is more effective than that produced by humans or GPT-4, due to the former's in-distribution nature. Our findings demonstrate that as few as 10K preference data pairs and fewer than 500 Step-DPO training steps can yield a nearly 3% gain in accuracy on MATH for models with over 70B parameters. Notably, Step-DPO, when applied to Qwen2-72B-Instruct, achieves scores of 70.8% and 94.0% on the test sets of MATH and GSM8K, respectively, surpassing a series of closed-source models, including GPT-4-1106, Claude-3-Opus, and Gemini-1.5-Pro.", "keywords": [ "Large Language Models", "Mathematical Reasoning", "Direct Preference Optimization" ], "rating_list": [ 1, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "H4k6Yn5kSt", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Bei Jiang", "gender": "Female", "institution": "University of Alberta", "country": "CA", "position": "Associate Professor" }, { "name": "Lei Ding", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "PhD student" }, { "name": "Linglong Kong", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Full Professor" }, { "name": "Xiaotian Chang", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Postdoc" }, { "name": "Yangdi Jiang", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 17, "n_ref_uni": 39, "n_ref": 78, "n_ref_all": 89, "n_fig": 1, "n_tab": 1, "L_tab": 336, "n_element_tab": 14, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 47, "n_element_tab_1": 1, "formula_len_all": 3768, "formula_len_all_1": 1194, "len_all": 134616, "len_all_1": 68045, "len_abs": 1169, "len_title": 126, "len_sents": 37481, "len_sents_1": 29965, "n_sents": 326, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1234, "title": "Exponential-Wrapped Mechanisms for Differential Privacy on Hadamard Manifolds", "abs": "We extend the Differential Privacy (DP) framework to Hadamard manifolds, the class of complete and simply connected Riemannian manifolds with non-positive sectional curvature. Inspired by the Cartan\u2013Hadamard theorem, we introduce Exponential-Wrapped Laplace and Gaussian mechanisms to achieve $\\varepsilon$-DP, $(\\varepsilon, \\delta)$-DP, Gaussian DP (GDP), and R\\'enyi DP (RDP) on these manifolds. Our approach employs efficient, straightforward algorithms that circumvent the computationally intensity Monte Carlo Markov Chain (MCMC) methods. This work is the first to extend $(\\varepsilon, \\delta)$-DP, GDP, and RDP to Hadamard manifolds. We further demonstrate the effectiveness of our methodology through simulations on the space of Symmetric Positive Definite Matrices, a frequently used Hadamard manifold in statistics. Our findings reveal that our Exponential-Wrapped mechanisms surpass traditional MCMC-based approaches, which require careful tuning and extensive diagnostics, in both performance and ease of use. Additionally, our methods achieve comparable utility to the Riemannian Laplace mechanism with enhanced utility for smaller privacy budgets ($\\varepsilon$) and operate orders of magnitude faster computationally.", "keywords": [ "Differential Privacy", "Riemannian Manifold", "Gaussian Differential Privacy", "Renyi Differential Privacy", "Hadamard Manifold", "Frechet Mean", "Symmetric Positive Definite Matrices" ], "rating_list": [ 6, 8, 5, 6, 6 ], "soundness_list": [ 3, 3, 4, 2, 3 ], "presentation_list": [ 2, 3, 3, 2, 2 ], "contribution_list": [ 3, 4, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 2, 3 ] }, { "paper_id": "H4iVLvRusn", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Liangliang Shi", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Zhengyan Shi", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Undergrad student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 16, "n_ref_uni": 34, "n_ref": 49, "n_ref_all": 59, "n_fig": 2, "n_tab": 5, "L_tab": 3460, "n_element_tab": 515, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 411, "n_element_tab_1": 72, "formula_len_all": 1369, "formula_len_all_1": 963, "len_all": 113585, "len_all_1": 60850, "len_abs": 1658, "len_title": 121, "len_sents": 32856, "len_sents_1": 28359, "n_sents": 234, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1664, "title": "SelKD: Selective Knowledge Distillation via Optimal Transport Perspective", "abs": "Knowledge Distillation (KD) has been a popular paradigm for training a (smaller) student model from its teacher model. However, little research has been done on the practical scenario where only a subset of the teacher's knowledge needs to be distilled, which we term selective KD (SelKD). This demand is especially pronounced in the era of foundation models, where the teacher model can be significantly larger than the student model. To address this issue, we propose to rethink the knowledge distillation problem from the perspective of Inverse Optimal Transport (IOT). Previous Bayesian frameworks mapped each sample to the probabilities of corresponding labels in an end-to-end manner, which fixed the number of classification categories and hindered effective local knowledge transfer. In contrast, IOT calculates from the standpoint of transportation or matching, allowing for the flexible selection of samples and their quantities for matching. Traditional logit-based KD can be viewed as a special case within the IOT framework. Building on this IOT foundation, we formalize this setting in the context of classification, where only selected categories from the teacher's category space are required to be recognized by the student in the context of closed-set recognition, which we call closed-set SelKD, enhancing the student's performance on specific subtasks. Furthermore, we extend the closed-set SelKD, introducing an open-set version of SelKD, where the student model is required to provide a ``not selected\" response for categories outside its assigned task. Experimental results on standard benchmarks demonstrate the superiority of our approach.", "keywords": [ "Knowledge Distillation", "Inverse Optimal Transport" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "H4FSx06FCZ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Jian Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Jiarui Meng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Ronggang Wang", "gender": "Male", "institution": "Peking University Shenzhen Graduate School", "country": "CN", "position": "Full Professor" }, { "name": "Shuzhou Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xuanyu Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Yanmin Wu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhipei Xu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 26, "n_ref": 70, "n_ref_all": 96, "n_fig": 16, "n_tab": 8, "L_tab": 2211, "n_element_tab": 289, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 1266, "n_element_tab_1": 155, "formula_len_all": 844, "formula_len_all_1": 623, "len_all": 138378, "len_all_1": 69403, "len_abs": 1476, "len_title": 131, "len_sents": 42746, "len_sents_1": 30878, "n_sents": 277, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1480, "title": "SecureGS: Boosting the Security and Fidelity of 3D Gaussian Splatting Steganography", "abs": "3D Gaussian splatting (3DGS) has emerged as a premier method for 3D representation due to its real-time rendering and high-quality outputs, underscoring the critical need to protect the copyright and privacy of 3D assets. Traditional NeRF steganography methods fail to address the explicit nature of 3DGS since its point cloud files are publicly accessible. Existing GS steganography solutions mitigate some issues but still struggle with reduced rendering fidelity, increased computational demands, and security flaws, especially in the security of the geometric structure of the visualized point cloud. To address these challenges, we propose a SecureGS, a secure and efficient 3DGS steganography framework inspired by Scaffold-GS's anchor point design and neural decoding. SecureGS utilizes a hybrid decoupled Gaussian encryption mechanism to embed offsets, scales, rotations, and RGB attributes of the hidden 3D Gaussian points within anchor point features, retrievable only by authorized users through privacy-preserving neural networks. To further enhance security, we propose a density region-aware anchor growing and pruning strategy that adaptively locates optimal hiding regions without exposing hidden information. Extensive experiments demonstrate that SecureGS significantly surpasses existing GS steganography methods in rendering fidelity, speed, and security, effectively concealing and accurately extracting 3D objects, images, and bits within original 3D scenes.", "keywords": [ "3DGS steganography", "copyright protection", "watermarking" ], "rating_list": [ 5, 6, 8, 8, 3 ], "soundness_list": [ 2, 3, 4, 3, 2 ], "presentation_list": [ 3, 3, 4, 3, 2 ], "contribution_list": [ 1, 3, 4, 3, 2 ], "confidence_list": [ 5, 2, 4, 5, 4 ] }, { "paper_id": "H48OMCCiI7", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Bin Chen", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Shu-Tao Xia", "gender": "Male", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Shuzhao Xie", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xiang Liu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yuqi Tan", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Zhi Wang", "gender": "Male", "institution": "SIGS, Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 38, "n_ref": 77, "n_ref_all": 90, "n_fig": 6, "n_tab": 6, "L_tab": 1456, "n_element_tab": 155, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 3232, "n_element_tab_1": 147, "formula_len_all": 533, "formula_len_all_1": 531, "len_all": 124195, "len_all_1": 63860, "len_abs": 1387, "len_title": 124, "len_sents": 35481, "len_sents_1": 27420, "n_sents": 281, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1393, "title": "WATER-GS: Watermark-embedded 3D Gaussian Splatting via Plug-and-play Decoder", "abs": "3D Gaussian Splatting (3DGS) has emerged as a pivotal technique for 3D scene representation, providing rapid rendering speeds and high fidelity. As 3DGS gains prominence, safeguarding its intellectual property becomes increasingly crucial since 3DGS could be used to imitate unauthorized scene creations and raise copyright issues. Existing watermarking methods for implicit NeRFs cannot be directly applied to 3DGS due to its explicit representation and real-time rendering process, leaving watermarking for 3DGS largely unexplored. In response, we propose WATER-GS, a novel method designed to protect 3DGS copyrights through a plug-and-play strategy. First, we introduce a pre-trained watermark decoder, treating raw 3DGS generative modules as potential watermark encoders to ensure imperceptibility. Additionally, we implement novel 3D distortion layers to enhance the robustness of the embedded watermark against common real-world distortions of point cloud data. Comprehensive experiments and ablation studies demonstrate that WATER-GS effectively embeds imperceptible and robust watermarks into 3DGS without compromising rendering efficiency and quality. Our experiments indicate that the 3D distortion layers can yield up to a 20% improvement in accuracy rate. Notably, our method is adaptable to different 3DGS variants, including 3DGS compression frameworks and 2D Gaussian splatting.", "keywords": [ "3D gaussian splatting", "digital watermarking" ], "rating_list": [ 3, 3, 3, 5, 6 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 5, 4, 5 ] }, { "paper_id": "H3lK5FV16C", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 1, "author_info_list": [ { "name": "Jinghan Yang", "gender": "Female", "institution": "Washington University, St. Louis", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 5, "n_ref_uni": 8, "n_ref": 13, "n_ref_all": 34, "n_fig": 6, "n_tab": 6, "L_tab": 2245, "n_element_tab": 419, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2923, "n_element_tab_1": 262, "formula_len_all": 571, "formula_len_all_1": 287, "len_all": 102422, "len_all_1": 52935, "len_abs": 737, "len_title": 80, "len_sents": 29593, "len_sents_1": 21526, "n_sents": 244, "n_sents_1": 170, "venue": "ICLR2025", "decision": "withdraw", "L_title": 33, "L_abs": 741, "title": "RED \u2013 ROBUST ENVIRONMENTAL DESIGN", "abs": "The classification of road signs by autonomous systems, especially those reliant on\nvisual inputs, is highly susceptible to adversarial attacks. Traditional approaches to\nmitigating such vulnerabilities have focused on enhancing the robustness of classi-\nfication models. In contrast, this paper adopts a fundamentally different strategy\naimed at increasing robustness through the redesign of road signs themselves. We\npropose an attacker-agnostic learning scheme to automatically design road signs\nthat are robust to a wide array of patch-based attacks. Empirical tests conducted in\nboth digital and physical environments demonstrate that our approach significantly\nreduces vulnerability to patch attacks, outperforming existing techniques.", "keywords": [ "adversarial machine learning; security" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "H3jGJzw0DN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Ashutosh Tripathi", "gender": "Male", "institution": "Rajiv Gandhi Institute of Petroleum Technology", "country": "IN", "position": "Undergrad student" }, { "name": "Pranab Sahoo", "gender": "Male", "institution": "Indian Institute of Technology Patna", "country": "", "position": "PhD student" }, { "name": "Samrat Mondal", "gender": "Male", "institution": "Indian Institute of Technology, Patna, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "Associate Professor" }, { "name": "Sriparna Saha", "gender": "Female", "institution": "Indian Institute of Technology Patna, India", "country": "", "position": "Associate Professor" } ], "n_formula": 50, "n_formula_1": 23, "n_ref_uni": 34, "n_ref": 67, "n_ref_all": 81, "n_fig": 11, "n_tab": 4, "L_tab": 233, "n_element_tab": 30, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 48, "n_element_tab_1": 10, "formula_len_all": 2729, "formula_len_all_1": 1080, "len_all": 131562, "len_all_1": 53655, "len_abs": 1612, "len_title": 153, "len_sents": 40758, "len_sents_1": 24241, "n_sents": 289, "n_sents_1": 170, "venue": "ICLR2025", "decision": "withdraw", "L_title": 104, "L_abs": 1620, "title": "FedPMVR: Addressing Data Heterogeneity in Federated Learning through Partial Momentum Variance Reduction", "abs": "Federated learning (FL) emerges as a promising paradigm for training machine learning models on decentralized data sources while preserving privacy. However, the presence of not independent and identically distributed (non-IID) data among the clients introduces high variance in gradient updates, posing a significant challenge to the global model's performance in terms of accuracy and convergence. To mitigate the adverse effects of data heterogeneity, we propose a novel momentum-based partial variance reduction technique. Our approach adjusts the gradient updates for the final classification layers of the client's neural network by leveraging the gradient differences between local and global models. This adjustment aims to effectively capture and mitigate client drift, a key challenge arises from the presence of non-IID data distributions across clients. We systematically explains client drifts and conduct extensive experiments on three widely-used datasets, demonstrating that our method significantly enhances global model accuracy while reducing the communication rounds needed for convergence. Notably, our momentum-based partial variance reduction technique provides a robust mechanism, rendering more efficient and effective in scenarios with inherently non-IID and heterogeneous data distributions. By addressing the critical challenge of data heterogeneity in FL, our proposed approach paves the way for more reliable and accurate model training while preserving the privacy of decentralized data sources. The code is available at the following link {https://anonymous.4open.science/r/FedPMVR-33C1}.", "keywords": [ "Federated Learning", "Data Heterogeneity", "Variance Reduction", "Image Classification." ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "H380m98pLE", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 10, "author_info_list": [ { "name": "Andrew Corbett", "gender": "Male", "institution": "digiLab", "country": "GB", "position": "Head of R&D" }, { "name": "Chris Bowman", "gender": "Male", "institution": "UKAEA", "country": "GB", "position": "Researcher" }, { "name": "Cyd Cowley", "gender": "Male", "institution": "digiLab", "country": "GB", "position": "Researcher" }, { "name": "Huy Le Nguyen", "gender": "Male", "institution": "digiLab Solutions Ltd", "country": "GB", "position": "Machine Learning Engineer" }, { "name": "James Buchanan", "gender": "unknown", "institution": "UKAEA", "country": "", "position": "Researcher" }, { "name": "Mikkel B. Lykkegaard", "gender": "Male", "institution": "digiLab", "country": "GB", "position": "Principal Researcher" }, { "name": "Robert Akers", "gender": "Male", "institution": "UKAEA", "country": "", "position": "Director" }, { "name": "Ross Allen", "gender": "Male", "institution": "digiLab", "country": "GB", "position": "Researcher" }, { "name": "Stanislas Pamela", "gender": "Male", "institution": "UKAEA ", "country": "", "position": "Researcher" }, { "name": "Tim James Dodwell", "gender": "Male", "institution": "University of Exeter", "country": "", "position": "Full Professor" } ], "n_formula": 26, "n_formula_1": 26, "n_ref_uni": 28, "n_ref": 52, "n_ref_all": 58, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1123, "formula_len_all_1": 1123, "len_all": 78554, "len_all_1": 73326, "len_abs": 1200, "len_title": 134, "len_sents": 22546, "len_sents_1": 22546, "n_sents": 161, "n_sents_1": 161, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1204, "title": "Constraining Gaussian Processes Regression with Quasi-Likelihood Constraint Relaxation", "abs": "Gaussian Process regression is a popular method for nonparametric, probabilistic modelling. One of its main attractions is also, in some contexts, a significant challenge; namely its high flexibility. This flexibility can be reduced by imposing constraints on the GP prior or posterior, something that there is a large and growing body of literature on. In this paper, we present a generalisation of virtual point methods and a framework for enforcing a broad range of constraints in GP posteriors. The method involves designing a quasi-likelihood function which encodes a relaxed form of the constraints, and then conditioning the unconstrained GP posterior on this quasi-likelihood. The method leverages ideas from existing methods for constrained GP regression, namely Riihimaki and Vehtari (2010) and Hansen et al. (2024), and expands these approaches to a much broader range of constraints. The method is demonstrated with a synthetic example, where a 2-dimensional GP posterior is required to have a divergence-free gradient, as well as real-world example where the posterior GP of Thomson scattering data from the MAST tokamak is required to be both monotonically decreasing and strictly positive.", "keywords": [ "constrained gaussian processes", "virtual point methods", "bayesian inference", "fusion energy" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 1, 3, 2, 1 ], "presentation_list": [ 1, 3, 2, 2 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 5, 2, 3, 5 ] }, { "paper_id": "H2zRScPSR7", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "David Fridovich-Keil", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Hamzah I Khan", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Shubhankar Agarwal", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Sandeep Chinchali", "gender": "unknown", "institution": "University of Texas at Austin", "country": "US", "position": "Assistant Professor" } ], "n_formula": 32, "n_formula_1": 18, "n_ref_uni": 43, "n_ref": 68, "n_ref_all": 92, "n_fig": 4, "n_tab": 2, "L_tab": 580, "n_element_tab": 15, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 153, "n_element_tab_1": 16, "formula_len_all": 2181, "formula_len_all_1": 1578, "len_all": 173901, "len_all_1": 82397, "len_abs": 1536, "len_title": 82, "len_sents": 59520, "len_sents_1": 38035, "n_sents": 464, "n_sents_1": 289, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1541, "title": "A Framework for Finding Local Saddle Points in Two-Player Zero-Sum Black-Box Games", "abs": "Saddle point optimization is a critical problem employed in numerous real-world applications, including portfolio optimization, generative adversarial networks, and robotics. It has been extensively studied in cases where the objective function is known and differentiable. Existing work in black-box settings with unknown objectives that can only be sampled either assumes convexity-concavity in the objective to simplify the problem or operates with noisy gradient estimators. In contrast, we introduce a framework inspired by Bayesian optimization which utilizes Gaussian processes to model the unknown (potentially nonconvex-nonconcave) objective and requires only zeroth-order samples. Our approach frames the saddle point optimization problem as a two-level process which can flexibly integrate existing and novel approaches to this problem. The upper level of our framework produces a model of the objective function by sampling in promising locations, and the lower level of our framework uses the existing model to frame and solve a general-sum game to identify locations to sample. This lower level procedure can be designed in complementary ways, and we demonstrate the flexibility of our approach by introducing variants which appropriately trade off between factors like runtime, the cost of function evaluations, and the number of available initial samples. We experimentally demonstrate these algorithms on synthetic and realistic datasets, showcasing their ability to efficiently locate local saddle points in these contexts.", "keywords": [ "black box optimization", "zero sum games", "saddle points", "Bayesian optimization" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "H2Gxil855b", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Dejia Xu", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Georgios Pavlakos", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Haitao Yang", "gender": "Male", "institution": "The University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Hanwen Jiang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Qixing Huang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Yuan Dong", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 17, "n_ref_uni": 91, "n_ref": 214, "n_ref_all": 233, "n_fig": 8, "n_tab": 3, "L_tab": 292, "n_element_tab": 29, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 194, "n_element_tab_1": 28, "formula_len_all": 900, "formula_len_all_1": 965, "len_all": 221252, "len_all_1": 65494, "len_abs": 629, "len_title": 91, "len_sents": 37356, "len_sents_1": 27224, "n_sents": 323, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 43, "L_abs": 1260, "title": "Atlas Gaussians Diffusion for 3D Generation", "abs": "Using the latent diffusion model has proven effective in developing novel 3D generation techniques. To harness the latent diffusion model, a key challenge is designing a high-fidelity and efficient representation that links the latent space and the 3D space. In this paper, we introduce Atlas Gaussians, a novel representation for feed-forward native 3D generation. Atlas Gaussians represent a shape as the union of local patches, and each patch can decode 3D Gaussians. We parameterize a patch as a sequence of feature vectors and design a learnable function to decode 3D Gaussians from the feature vectors. In this process, we incorporate UV-based sampling, enabling the generation of a sufficiently large, and theoretically infinite, number of 3D Gaussian points. The large amount of 3D Gaussians enables the generation of high-quality details. Moreover, due to local awareness of the representation, the transformer-based decoding procedure operates on a patch level, ensuring efficiency. We train a variational autoencoder to learn the Atlas Gaussians representation, and then apply a latent diffusion model on its latent space for learning 3D Generation. Experiments show that our approach outperforms the prior arts of feed-forward native 3D generation.", "keywords": [ "3D generation", "diffusion", "3D Gaussian Splatting" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 2, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "H25xduunIK", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Blair Yang", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" }, { "name": "Fuyang Cui", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" }, { "name": "Jimmy Ba", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "", "position": "Assistant Professor" }, { "name": "Keiran Paster", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "PhD student" }, { "name": "Pashootan Vaezipoor", "gender": "Male", "institution": "Georgian", "country": "CA", "position": "Researcher" }, { "name": "Silviu Pitis", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "PhD student" }, { "name": "Michael R Zhang", "gender": "unknown", "institution": "University of Toronto", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 38, "n_ref": 57, "n_ref_all": 92, "n_fig": 16, "n_tab": 6, "L_tab": 2007, "n_element_tab": 247, "n_fig_1": 10, "n_tab_1": 5, "L_tab_1": 2205, "n_element_tab_1": 44, "formula_len_all": 354, "formula_len_all_1": 297, "len_all": 178305, "len_all_1": 64236, "len_abs": 911, "len_title": 136, "len_sents": 55956, "len_sents_1": 28468, "n_sents": 478, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 913, "title": "Report Cards: Qualitative Evaluation of Language Models Using Natural Language Summaries", "abs": "The rapid development and dynamic nature of large language models (LLMs) make it difficult for conventional quantitative benchmarks to accurately assess their capabilities. We propose report cards, which are human-interpretable, natural language summaries of model behavior for specific skills or topics. We develop a framework to evaluate report cards based on three criteria: specificity (ability to distinguish between models), faithfulness (accurate representation of model capabilities), and interpretability (clarity and relevance to humans). We also propose an iterative algorithm for generating report cards without human supervision and explore its efficacy by ablating various design choices. Through experimentation with popular LLMs, we demonstrate that report cards provide insights beyond traditional benchmarks and can help address the need for a more interpretable and holistic evaluation of LLMs.", "keywords": [ "evaluation", "LLMs" ], "rating_list": [ 5, 5, 8, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "H1nykRhieN", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Ruyu Shang", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yapeng Qi", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "BO SHEN", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 23, "n_ref": 32, "n_ref_all": 39, "n_fig": 3, "n_tab": 4, "L_tab": 2297, "n_element_tab": 397, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2297, "n_element_tab_1": 397, "formula_len_all": 737, "formula_len_all_1": 737, "len_all": 81112, "len_all_1": 58512, "len_abs": 1080, "len_title": 132, "len_sents": 24982, "len_sents_1": 25002, "n_sents": 182, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1085, "title": "MvHSTM: A Multi-view Hypergraph Spatio-Temporal Model for Traffic Speed Forecasting", "abs": "Accurate traffic speed prediction is critical in modern society as it is effective for both individuals and authorities. Due to the large scale of urban road networks, traffic speed exhibits complex spatio-temporal dependencies, not only among adjacent nodes but also across the network, reflecting both local and cross-regional simultaneous correlations. However, existing studies have not effectively addressed these characteristics. In this context, we propose a novel framework called Multi-view Hypergraph Spatio-Temporal Model (MvHSTM) that employs a temporal transformer to capture temporal dependencies and utilizes hypergraph convolutional networks to inherently model spatial relationships. Specifically, we introduce two hypergraph construction methods, the Geographical Adjacency Hypergraph (GAH) and the Feature Similarity Hypergraph (FSH), to capture spatial correlations on neighboring and non-neighboring scales. Extensive experiments on real-world traffic speed datasets demonstrate that our approach achieves state-of-the-art performance compared to baseline methods.", "keywords": [ "Traffic Prediction", "Deep Learning", "Hypergraph Convolution" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "H0qIWXXLUR", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Alexey Gorbatovski", "gender": "Male", "institution": "T-Tech", "country": "RU", "position": "Researcher" }, { "name": "Alexey Malakhov", "gender": "unknown", "institution": "Financial University", "country": "", "position": "Undergrad student" }, { "name": "Boris Shaposhnikov", "gender": "Male", "institution": "ITMO University", "country": "", "position": "Undergrad student" }, { "name": "Daniil Gavrilov", "gender": "unknown", "institution": "T-Bank", "country": "RU", "position": "Principal Researcher" }, { "name": "Nikita Balagansky", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "", "position": "MS student" }, { "name": "Nikita Surnachev", "gender": "Male", "institution": "Innopolis", "country": "RU", "position": "Undergrad student" }, { "name": "Yaroslav Aksenov", "gender": "Male", "institution": "Higher School of Economics", "country": "", "position": "Undergrad student" }, { "name": "Yan Viktorovich Maximov", "gender": "Male", "institution": "Higher School of Economics, Higher School of Economics", "country": "RU", "position": "MS student" } ], "n_formula": 21, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 94, "n_ref_all": 133, "n_fig": 20, "n_tab": 17, "L_tab": 4129, "n_element_tab": 441, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 785, "n_element_tab_1": 92, "formula_len_all": 1549, "formula_len_all_1": 784, "len_all": 238769, "len_all_1": 56068, "len_abs": 1215, "len_title": 98, "len_sents": 73140, "len_sents_1": 24933, "n_sents": 581, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1218, "title": "Learn Your Reference Model for Real Good Alignment", "abs": "Despite the fact that offline methods for Large Language Models (LLMs) alignment do not require a direct reward model, they remain susceptible to overoptimization. This issue arises when the trained model deviates excessively from the reference policy, leading to a decrease in sample quality. We propose a novel approach of offline alignment methods, called Trust Region (including variants TR-DPO, TR-IPO, TR-KTO), which dynamically updates the reference policy throughout the training process. Our results show that TR alignment methods effectively mitigate overoptimization, enabling models to maintain strong performance even when substantially deviating from the initial reference policy. We demonstrate the efficacy of these approaches not only through toy examples that exhibit reduced overoptimization, but also through direct, side-by-side comparisons in specific tasks such as helpful and harmless dialogue, as well as summarization, where they surpass conventional methods. Additionally, we report significant improvements in general-purpose assistant setups with the Llama3 model on the AlpacaEval 2 and Arena-Hard benchmarks, highlighting the advantages of Trust Region methods over classical approaches.", "keywords": [ "reinforcement learning from human feedback", "language models", "RLHF", "preferences", "alignment", "overoptimization" ], "rating_list": [ 6, 5, 8, 3, 8 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 4, 2, 4 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 4, 4 ] }, { "paper_id": "H0UcwHgwEO", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Adith Swaminathan", "gender": "Male", "institution": "NetFlix", "country": "US", "position": "Principal Researcher" }, { "name": "Allen Nie", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Andrey Kolobov", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Ching-An Cheng", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Senior researcher" }, { "name": "Dipendra Misra", "gender": "Male", "institution": "Mosaic Research, Databricks", "country": "US", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 54, "n_ref": 83, "n_ref_all": 119, "n_fig": 7, "n_tab": 6, "L_tab": 2207, "n_element_tab": 242, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 5652, "n_element_tab_1": 575, "formula_len_all": 123, "formula_len_all_1": 0, "len_all": 205420, "len_all_1": 74365, "len_abs": 1846, "len_title": 120, "len_sents": 68121, "len_sents_1": 30587, "n_sents": 538, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1617, "title": "LLF-Bench: A Benchmark for Interactive Learning from Language Feedback", "abs": "We introduce a new benchmark, LLF-Bench (Learning from Language Feedback Benchmark; pronounced as ``elf-bench''), to evaluate the ability of AI agents to interactively learn from natural language feedback and instructions. Learning from language feedback (LLF) is essential for people, largely because the rich information this feedback provides can help a learner avoid much of trial and error and thereby speed up the learning process. Large Language Models (LLMs) have recently enabled AI agents to comprehend natural language --- and hence AI agents can potentially benefit from language feedback during learning like humans do. But existing interactive benchmarks do not assess this crucial capability: they either use numeric reward feedback or require no learning at all (only planning or information retrieval). LLF-Bench is designed to fill this omission. LLF-Bench is a diverse collection of sequential decision-making tasks that includes user recommendation, poem writing, navigation, and robot control. The objective of an agent is to interactively solve these tasks based on their natural-language instructions and the feedback received after taking actions. Crucially, to ensure that the agent actually learns from the feedback, LLF-Bench implements several randomization techniques to ensure that the task isn't familiar to the agent and that the agent is robust to various verbalizations. In addition, LLF-Bench allows configuring different types of feedback to study how agents respond to them. Together, these features make LLF-Bench a unique research platform for developing and testing LLF agents.", "keywords": [ "LLM", "benchmark", "decision making", "reinforcement learning" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "GzLepH6MBB", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Lin ente", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Michael Kampffmeyer", "gender": "Male", "institution": "UiT The Arctic University of Norway", "country": "NO", "position": "Full Professor" }, { "name": "Xiaodan Liang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Xin Dong", "gender": "Female", "institution": "Bytedance", "country": "", "position": "Researcher" }, { "name": "Xiu Li", "gender": "Male", "institution": "Bytedance", "country": "SG", "position": "Researcher" }, { "name": "Yuxuan Luo", "gender": "Male", "institution": "Bytedance", "country": "CN", "position": "Researcher" }, { "name": "xujie zhang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 48, "n_ref": 87, "n_ref_all": 106, "n_fig": 12, "n_tab": 3, "L_tab": 1054, "n_element_tab": 132, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 807, "n_element_tab_1": 111, "formula_len_all": 408, "formula_len_all_1": 408, "len_all": 148447, "len_all_1": 61819, "len_abs": 1754, "len_title": 129, "len_sents": 41862, "len_sents_1": 29737, "n_sents": 298, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1762, "title": "MMTryon: Multi-Modal Multi-Reference Control for High-Quality Fashion Generation", "abs": "This paper introduces MMTryon, a multi-modal multi-reference VIrtual Try-ON (VITON) framework, which can generate high-quality compositional try-on results by taking a text instruction and multiple garment images as inputs. Our MMTryon addresses three problems overlooked in prior literature: 1) Support of multiple try-on items. Existing methods are commonly designed for single-item try-on tasks (e.g., upper/lower garments, dresses). 2) Specification of dressing style. Existing methods are unable to customize dressing styles based on instructions (e.g., zipped/unzipped, tuck-in/tuck-out, etc.) 3) Segmentation Dependency. They further heavily rely on category-specific segmentation models to identify the replacement regions, with segmentation errors directly leading to significant artifacts in the try-on results. To address the first two issues, our MMTryon introduces a novel multi-modality and multi-reference attention mechanism to combine the garment information from reference images and dressing-style information from text instructions. Besides, to remove the segmentation dependency, MMTryon uses a parsing-free garment encoder and leverages a novel scalable data generation pipeline to convert existing VITON datasets to a form that allows MMTryon to be trained without requiring any explicit segmentation. Extensive experiments on high-resolution benchmarks and in-the-wild test sets demonstrate MMTryon's superiority over existing SOTA methods both qualitatively and quantitatively. MMTryon's impressive performance on multi-item and style-controllable virtual try-on scenarios and its ability to try on any outfit in a large variety of scenarios from any source image, opens up a new avenue for future investigation in the fashion community.", "keywords": [ "Multi-modal Fashion Generation", "Compositional Virtual Try-on" ], "rating_list": [ 3, 10, 8 ], "soundness_list": [ 2, 4, 4 ], "presentation_list": [ 2, 4, 4 ], "contribution_list": [ 2, 4, 4 ], "confidence_list": [ 5, 5, 5 ] }, { "paper_id": "GySIAKEwtZ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Chao Chen", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Associate Professor" }, { "name": "Haibin Ling", "gender": "Male", "institution": "State University of New York, Stony Brook", "country": "US", "position": "Professor" }, { "name": "Jiachen Yao", "gender": "Male", "institution": ", State University of New York at Stony Brook", "country": "", "position": "PhD student" }, { "name": "Lingjie Yi", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" }, { "name": "Raphael Douady", "gender": "unknown", "institution": "CNRS", "country": "FR", "position": "Researcher" }, { "name": "Weimin Lyu", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "PhD student" } ], "n_formula": 103, "n_formula_1": 21, "n_ref_uni": 37, "n_ref": 79, "n_ref_all": 135, "n_fig": 15, "n_tab": 6, "L_tab": 3645, "n_element_tab": 346, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 1822, "n_element_tab_1": 164, "formula_len_all": 11263, "formula_len_all_1": 2504, "len_all": 194347, "len_all_1": 68242, "len_abs": 2021, "len_title": 117, "len_sents": 50205, "len_sents_1": 26946, "n_sents": 579, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1466, "title": "Representation Learning for Long Tail Recognition via Feature Space Re-Construction", "abs": "Deep learning has achieved significant success on balanced datasets. However, real-world data often exhibit a long-tailed distribution. Empirical results show that long-tailed data skews representations where head classes dominate the feature space. Many methods have been proposed to empirically correct the skewed representations. However, a clear theoretical understanding of the underlying causes and extent of this skew remains lacking. In this work, we provide a comprehensive theoretical analysis to elucidate how long-tailed data affects representations, deriving the conditions under which the centers of the tail classes shrink together or even collapse into a single point. This results in overlapping feature distributions of tail classes, making features in the overlapping regions inseparable. Moreover, we demonstrate that merely empirically correcting the skewed representations of training data is insufficient to separate the overlapping features, due to distribution shifts between training and real data. To address these challenges, we propose a novel long-tailed representation learning method, FeatRecon. It reconstructs the feature space so that features of all classes are arranged into symmetrical and linearly separable regions. Thereby, it enhances model robustness to long-tailed data. We validate the effectiveness of our method through extensive experiments on the CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and iNaturalist 2018 datasets.", "keywords": [ "contrastive learning", "representation learning", "long-tail recgonition", "theory" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "Gx04TnVjee", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Di ZHANG", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "VP" }, { "name": "Menghan Xia", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Researcher" }, { "name": "Pengfei Wan", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Director" }, { "name": "Sida Peng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xian Liu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Xiaoyu Shi", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xintao Wang", "gender": "Male", "institution": "Kuaishou", "country": "CN", "position": "Researcher" }, { "name": "Ziyang Yuan", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Xiao Fu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 39, "n_ref": 80, "n_ref_all": 112, "n_fig": 14, "n_tab": 9, "L_tab": 2562, "n_element_tab": 237, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 383, "n_element_tab_1": 35, "formula_len_all": 857, "formula_len_all_1": 541, "len_all": 167317, "len_all_1": 57734, "len_abs": 1476, "len_title": 129, "len_sents": 44877, "len_sents_1": 25368, "n_sents": 306, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1487, "title": "3DTrajMaster: Mastering 3D Trajectory for Multi-Entity Motion in Video Generation", "abs": "This paper aims to manipulate multi-entity 3D motions in video generation. Previous methods on controllable video generation primarily leverage 2D control signals to manipulate object motions and have achieved remarkable synthesis results. However, 2D control signals are inherently limited in expressing the 3D nature of object motions. To overcome this problem, we introduce 3DTrajMaster, a robust controller that regulates multi-entity dynamics in 3D space, given user-desired 6DoF pose (location and rotation) sequences of entities. At the core of our approach is a plug-and-play 3D-motion grounded object injector that fuses multiple input entities with their respective 3D trajectories through a gated self-attention mechanism. In addition, we exploit an injector architecture to preserve the video diffusion prior, which is crucial for generalization ability. To mitigate video quality degradation, we introduce a domain adaptor during training and employ an annealed sampling strategy during inference. To address the lack of suitable training data, we construct a 360-Motion Dataset, which first correlates collected 3D human and animal assets with GPT-generated trajectory and then captures their motion with 12 evenly-surround cameras on diverse 3D UE platforms. Extensive experiments show that 3DTrajMaster sets a new state-of-the-art in both accuracy and generalization for controlling multi-entity 3D motions. Project page: http://fuxiao0719.github.io/projects/3dtrajmaster", "keywords": [ "Controllable Video Generation", "3D Motion Control", "Multi-Entity Motion" ], "rating_list": [ 8, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "Gws8Q4wSrJ", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Ganzhao Yuan", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Associate Professor" } ], "n_formula": 245, "n_formula_1": 42, "n_ref_uni": 55, "n_ref": 123, "n_ref_all": 174, "n_fig": 3, "n_tab": 4, "L_tab": 5987, "n_element_tab": 444, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 133, "n_element_tab_1": 5, "formula_len_all": 30312, "formula_len_all_1": 5423, "len_all": 380161, "len_all_1": 74042, "len_abs": 3887, "len_title": 151, "len_sents": 131354, "len_sents_1": 26841, "n_sents": 863, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1295, "title": "A Block Coordinate Descent Method for Nonsmooth Composite Optimization under Orthogonality Constraints", "abs": "Nonsmooth composite optimization with orthogonality constraints is crucial in statistical learning and data science, but it presents challenges due to its nonsmooth objective and computationally expensive, non-convex constraints. In this paper, we propose a new approach called \\textbf{OBCD}, which leverages Block Coordinate Descent (BCD) to address these challenges. \\textbf{OBCD} is a feasible method with a small computational footprint. In each iteration, it updates $k$ rows of the solution matrix, where $k \\geq 2$, while globally solving a small nonsmooth optimization problem under orthogonality constraints. We prove that the limiting points of \\textbf{OBCD}, referred to as (global) block-$k$ stationary points, offer stronger optimality than standard critical points. Furthermore, we show that \\textbf{OBCD} converges to $\\epsilon$-block-$k$ stationary points with an ergodic convergence rate of $\\mathcal{O}(1/\\epsilon)$. Additionally, under the Kurdyka-Lojasiewicz (KL) inequality, we establish the non-ergodic convergence rate of \\textbf{OBCD}. We also extend \\textbf{OBCD} with breakpoint searching methods for subproblem solving and greedy strategies for working set selection. Comprehensive experiments demonstrate the superior performance of our approach across various tasks.", "keywords": [ "Orthogonality Constraints", "Nonconvex Optimization", "Nonsmooth Composite Optimization", "Block Coordinate Descent", "Convergence Analysis" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 2, 1 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "GwSL33Qx42", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Donghao Zhou", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Guangyong Chen", "gender": "Male", "institution": "Zhejiang Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Hao Chen", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jiancheng Huang", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Jiaze Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jinbin Bai", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Pheng-Ann Heng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Xiaowei Hu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 45, "n_ref": 87, "n_ref_all": 123, "n_fig": 13, "n_tab": 4, "L_tab": 1529, "n_element_tab": 187, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1575, "n_element_tab_1": 189, "formula_len_all": 616, "formula_len_all_1": 534, "len_all": 151782, "len_all_1": 61787, "len_abs": 1447, "len_title": 243, "len_sents": 44135, "len_sents_1": 27544, "n_sents": 325, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1404, "title": "MagicTailor: Component-Controllable Personalization in Text-to-Image Diffusion Models", "abs": "Recent advancements in text-to-image (T2I) diffusion models have enabled the creation of high-quality images from text prompts, but they still struggle to generate images with precise control over specific visual concepts. Existing approaches can replicate a given concept by learning from reference images, yet they lack the flexibility for fine-grained customization of the individual component within the concept. In this paper, we introduce component-controllable personalization, a novel task that pushes the boundaries of T2I models by allowing users to reconfigure and personalize specific components of concepts. This task is particularly challenging due to two primary obstacles: semantic pollution, where unwanted visual elements corrupt the personalized concept, and semantic imbalance, which causes disproportionate learning of visual semantics. To overcome these challenges, we design MagicTailor, an innovative framework that leverages Dynamic Masked Degradation (DM-Deg) to dynamically perturb undesired visual semantics and Dual-Stream Balancing (DS-Bal) to establish a balanced learning paradigm for visual semantics. Extensive comparisons, ablations, and analyses demonstrate that MagicTailor not only excels in this challenging task but also holds significant promise for practical applications, paving the way for more nuanced and creative image generation. Our code will be released.", "keywords": [ "Diffusion Models", "Text-to-Image Generation", "Personalization" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "GwJXJSCH1S", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Changcheng Hua", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Jiayi Gao", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Jun Guo", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Full Professor" }, { "name": "Kongming Liang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Associate Professor" }, { "name": "Yang Liu", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuxin Peng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Zhanyu Ma", "gender": "Male", "institution": "Beijing University of Post and Telecommunication", "country": "CN", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 23, "n_ref": 53, "n_ref_all": 70, "n_fig": 7, "n_tab": 2, "L_tab": 286, "n_element_tab": 59, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 461, "n_element_tab_1": 68, "formula_len_all": 605, "formula_len_all_1": 605, "len_all": 108913, "len_all_1": 59072, "len_abs": 1517, "len_title": 91, "len_sents": 34845, "len_sents_1": 30512, "n_sents": 195, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 1523, "title": "Contextually Harmonious Local Video Editing", "abs": "We introduce a new task for video editing: Contextual Harmonious Local Editing, which focuses on replacing a local moving subject in videos containing multiple subjects or reference objects. The goal is to ensure that the replaced subject maintains its original motion while its size remains harmonious with the scene's context. Previous methods often face two specific challenges when addressing this task: (1) ensuring the size of the replaced subject remains contextually harmonious (2) maintaining the original motion and achieving subject replacement without being affected by the motion of other subjects. To address the above problems, we propose a novel three stage video editing pipeline. We initially leverage large pre-trained models to acquire knowledge about the shape and size differences between the original and replaced subjects. To mitigate interference from context motion, we erase other moving subjects to extract the target subject's motion and dynamically choose the editing method to preserve the original subject's motion under different shape transformation.\nFollowing that,we seamlessly replace the original subject in the video with the resized edited subject, ensuring its size harmonizes with the video's context.\nAs the first work to focus on this task, we also provide a high-quality evaluation dataset and metrics to assess the performance of existing methods on this task. Experimental results based on this dataset demonstrate that our method achieves state-of-the-art (SOTA) performance.", "keywords": [ "Video Editing", "Diffusion Models", "Contextaul Harmonious" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 1, 2, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "GvzL4LuycW", "primary_area": "learning on time series and dynamical systems", "n_author": 9, "author_info_list": [ { "name": "Dong Li", "gender": "Female", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Jimin Huang", "gender": "Male", "institution": "The Fin AI", "country": "SG", "position": "Principal Researcher" }, { "name": "Mengxi Xiao", "gender": "Female", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Min Peng", "gender": "Female", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Qianqian Xie", "gender": "Female", "institution": "Yale University", "country": "US", "position": "Researcher" }, { "name": "Shuai Chen", "gender": "Male", "institution": "xiaomi corporation", "country": "", "position": "Researcher" }, { "name": "Sophia Ananiadou", "gender": "Female", "institution": "University of Manchester", "country": "GB", "position": "Full Professor" }, { "name": "Zhengyu Chen", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Zihao Jiang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 48, "n_ref_all": 57, "n_fig": 1, "n_tab": 6, "L_tab": 2831, "n_element_tab": 328, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 3165, "n_element_tab_1": 346, "formula_len_all": 392, "formula_len_all_1": 392, "len_all": 101759, "len_all_1": 62565, "len_abs": 945, "len_title": 128, "len_sents": 26700, "len_sents_1": 26401, "n_sents": 177, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 948, "title": "TimeRAG: It's Time for Retrieval-Augmented Generation in Time-Series Forecasting", "abs": "Time-series data are essential for forecasting tasks across various domains. While Large Language Models (LLMs) have excelled in many areas, they encounter significant challenges in time-series forecasting, particularly in extracting relevant information from extensive temporal datasets. Unlike textual data, time-series data lack explicit retrieval ground truths, complicating the retrieval process. To tackle these issues, we present TimeRAG, a novel retrieval-augmented approach tailored for time-series forecasting. Our method uniquely applies to continuous and complex temporal sequences, and it is trained using LLM feedback, effectively addressing the absence of ground truth and aligning the priorities of the retriever and the LLM. Experimental results demonstrate the effectiveness of TimeRAG, highlighting its ability to significantly enhance forecasting performance and showcasing the potential of LLMs in time-series prediction tasks.", "keywords": [ "time series", "large language model", "retrieval augmented generation" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "GvsCOOPxoI", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Chi Jin", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Assistant Professor" }, { "name": "Jingchu Gai", "gender": "Male", "institution": "School of Mathematical Science, Peking University, Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Qinghua Liu", "gender": "Male", "institution": "Microsoft Research, NYC", "country": "US", "position": "Postdoc" }, { "name": "Zhuoran Yang", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 254, "n_formula_1": 35, "n_ref_uni": 52, "n_ref": 89, "n_ref_all": 107, "n_fig": 1, "n_tab": 7, "L_tab": 3184, "n_element_tab": 330, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 2232, "n_element_tab_1": 28, "formula_len_all": 41901, "formula_len_all_1": 4124, "len_all": 300514, "len_all_1": 75793, "len_abs": 1144, "len_title": 119, "len_sents": 77531, "len_sents_1": 28774, "n_sents": 849, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1147, "title": "Provable Learning for DEC-POMDPs: Factored Models and Memoryless Agents", "abs": "This paper studies cooperative Multi-Agent Reinforcement Learning (MARL) under the mathematical model of Decentralized Partially Observable Markov Decision Process (DEC-POMDP). Despite the empirical success of cooperative MARL, its theoretical foundation, particularly in the realm of provable learning of DEC-POMDPs, remains limited. In this paper, we first present a hardness result in theory demonstrating that, without additional structural assumptions, learning DEC-POMDPs requires several samples that grows exponentially with the number of agents in the worst case, which is also known as the curse of multiagency. This motivates us to explore important subclasses of DEC-POMDPs for which efficient solutions can be found. Specifically, we propose new algorithms and establish sample-efficiency guarantees that break the curse of multiagency, for finding both local and global optima in two important scenarios: (1) when agents employ memoryless policies, selecting actions based solely on their current observations; and (2) when a factored structure is present, which enables key properties similar to value decomposition in VDN or Qmix.", "keywords": [ "decentralized POMDP", "provably efficient algorithm" ], "rating_list": [ 6, 5, 8, 8, 5 ], "soundness_list": [ 3, 4, 3, 4, 3 ], "presentation_list": [ 3, 2, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 2, 3, 3, 3, 3 ] }, { "paper_id": "Gvg3nXZvyg", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Constantin Ulrich", "gender": "Male", "institution": "Deutsches Krebsforschungszentrum", "country": "DE", "position": "PhD student" }, { "name": "Emily Tempus", "gender": "Female", "institution": "Deutsches Krebsforschungszentrum", "country": "DE", "position": "Research Intern" }, { "name": "Klaus Maier-Hein", "gender": "Male", "institution": "German Cancer Research Center", "country": "DE", "position": "Full Professor" }, { "name": "Maximilian Rouven Rokuss", "gender": "Male", "institution": "Deutsches Krebsforschungszentrum", "country": "DE", "position": "PhD student" }, { "name": "Paul F Jaeger", "gender": "unknown", "institution": "Google", "country": "CA", "position": "Researcher" }, { "name": "Tassilo Wald", "gender": "Male", "institution": "German Cancer Research Center", "country": "DE", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 35, "n_ref": 72, "n_ref_all": 91, "n_fig": 9, "n_tab": 4, "L_tab": 6033, "n_element_tab": 555, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 767, "n_element_tab_1": 42, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 148727, "len_all_1": 61179, "len_abs": 1277, "len_title": 94, "len_sents": 39477, "len_sents_1": 30222, "n_sents": 279, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 46, "L_abs": 1281, "title": "INTRABENCH: Interactive Radiological Benchmark", "abs": "Current interactive segmentation approaches, inspired by the success of META\u2019s Segment Anything model, have achieved notable advancements, however they come with substantial limitations that hinder their practical application in real clinical scenarios. These include unrealistic human interaction requirements, such as slice-by-slice operations for 2D models on 3D data, a lack of iterative refinement, and insufficient evaluation experiments. These shortcomings prevent accurate assessment of model performance and lead to inconsistent outcomes across studies. IntRaBench overcomes these challenges by offering a comprehensive and reproducible framework for evaluating interactive segmentation methods in realistic, clinically relevant scenarios. It includes diverse datasets, target structures, and segmentation models, and provides a flexible codebase that allows seamless integration of new models and prompting strategies. Additionally, we introduce advanced techniques to minimize clinician interaction, ensuring fair comparisons between 2D and 3D models. By open-sourcing IntRaBench, we invite the research community to integrate their models and prompting techniques, ensuring continuous and transparent evaluation of interactive segmentation models in 3D medical\nimaging.", "keywords": [ "interactive", "prompt", "segmentation", "medical image computing", "medical", "human in the loop", "sam", "medical image segmentation" ], "rating_list": [ 1, 3, 5 ], "soundness_list": [ 1, 3, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "GvUahyZ8UF", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Alex Hern\u00e1ndez-Garc\u00eda", "gender": "unknown", "institution": "Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Assistant Professor" }, { "name": "Emmanuel Bengio", "gender": "Male", "institution": "Recursion", "country": "", "position": "Researcher" }, { "name": "Hyeonah Kim", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jinkyoo Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Minsu Kim", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Postdoc" }, { "name": "Sanghyeok Choi", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Taeyoung Yun", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 12, "n_ref_uni": 41, "n_ref": 146, "n_ref_all": 168, "n_fig": 9, "n_tab": 10, "L_tab": 4046, "n_element_tab": 438, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 988, "n_element_tab_1": 254, "formula_len_all": 727, "formula_len_all_1": 629, "len_all": 177776, "len_all_1": 66799, "len_abs": 1564, "len_title": 120, "len_sents": 46248, "len_sents_1": 29242, "n_sents": 405, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1527, "title": "Improved Off-policy Reinforcement Learning in Biological Sequence Design", "abs": "Designing biological sequences with desired properties is a significant challenge due to the combinatorially vast search space and the high cost of evaluating each candidate sequence. To address these challenges, reinforcement learning (RL) methods, such as GFlowNets, utilize proxy models for rapid reward evaluation and annotated data for policy training. Although these approaches have shown promise in generating diverse and novel sequences, the limited training data relative to the vast search space often leads to the misspecification of proxy for out-of-distribution inputs. We introduce $\\delta$-Conservative Search, a novel off-policy search method for training GFlowNets designed to improve robustness against proxy misspecification. The key idea is to incorporate conservativeness, controlled by parameter $\\delta$, to constrain the search to reliable regions. Specifically, we inject noise into high-score offline sequences by randomly masking tokens with a Bernoulli distribution of parameter $\\delta$ and then denoise masked tokens using the GFlowNet policy. Additionally, $\\delta$ is adaptively adjusted based on the uncertainty of the proxy model for each data point. This enables the reflection of proxy uncertainty to determine the level of conservativeness. Experimental results demonstrate that our method consistently outperforms existing machine learning methods in discovering high-score sequences across diverse tasks\u2014including DNA, RNA, protein, and peptide design\u2014especially in large-scale scenarios.", "keywords": [ "Biological sequence design", "GFlowNets", "offline RL", "active learning" ], "rating_list": [ 5, 3, 6, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "Gv4uHroun5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Boyuan Pan", "gender": "Male", "institution": "Xiaohongshu", "country": "", "position": "Researcher" }, { "name": "Chuyi Tan", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "Undergrad student" }, { "name": "Jiayi Shi", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Kan Li", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Peiwen Yuan", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Shaoxiong Feng", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Xinglin Wang", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yao Hu", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Principal Researcher" }, { "name": "Yiwei Li", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Yueqi Zhang", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 23, "n_ref": 33, "n_ref_all": 54, "n_fig": 6, "n_tab": 7, "L_tab": 2266, "n_element_tab": 151, "n_fig_1": 5, "n_tab_1": 10, "L_tab_1": 3198, "n_element_tab_1": 194, "formula_len_all": 1119, "formula_len_all_1": 1038, "len_all": 107261, "len_all_1": 61847, "len_abs": 1386, "len_title": 124, "len_sents": 33178, "len_sents_1": 27073, "n_sents": 211, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1401, "title": "Beyond One-Size-Fits-All: Tailored Benchmarks for Efficient Model Evaluation", "abs": "Evaluating models on large benchmarks can be very resource-intensive, especially during a period of rapid model iteration. Existing efficient evaluation methods approximate the performance of target models by assessing them on a small static coreset derived from publicly available evaluation results of source models. However, these approaches rely on the assumption that each target model has a high prediction consistency with source models, which doesn\u2019t generalize well in practice, leading to inaccurate performance estimates. To fill this gap, we propose TailoredBench, a method that provides customized evaluations tailored to each target model. Specifically, a Global-coreset is first constructed as a probe to identify the most consistent source models for the target models with an adaptive source model selection strategy. Afterwards, a scalable K-Medoids clustering algorithm is proposed to extend the Global-coreset to tailored Native-coreset for each target model. According to the predictions on respective Native-coreset, we estimate the overall performance of target models with a calibrated restoration strategy. Comprehensive experiments on five benchmarks across more than 300 models demonstrate that compared to best performing baselines, TailoredBench achieves an average reduction of 24.8% in the MAE of accuracy estimates, showcasing strong effectiveness and generalizability.", "keywords": [ "evaluation", "efficient", "benchmark" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 1 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "Gv0TOAigIY", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Isaac L. Chuang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Liu Ziyin", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "YiZhou Xu", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" } ], "n_formula": 66, "n_formula_1": 20, "n_ref_uni": 38, "n_ref": 49, "n_ref_all": 91, "n_fig": 12, "n_tab": 1, "L_tab": 196, "n_element_tab": 28, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 196, "n_element_tab_1": 28, "formula_len_all": 3592, "formula_len_all_1": 1294, "len_all": 166963, "len_all_1": 79727, "len_abs": 855, "len_title": 96, "len_sents": 53897, "len_sents_1": 35483, "n_sents": 535, "n_sents_1": 306, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 838, "title": "Remove Symmetries to Control Model Expressivity", "abs": "When symmetry is present in the loss function, the model is likely to be trapped in a low-capacity state that is sometimes known as a ``collapse.\" Being trapped in these low-capacity states can be a major obstacle to training across many scenarios where deep learning technology is applied. We first prove two concrete mechanisms through which symmetries lead to reduced capacities and ignored features during training. We then propose a simple and theoretically justified algorithm, \\textit{syre}, to remove almost all symmetry-induced low-capacity states in neural networks. The proposed method is shown to improve the training of neural networks in scenarios when this type of entrapment is especially a concern. A remarkable merit of the proposed method is that it is model-agnostic and does not require any knowledge of the symmetry.", "keywords": [ "model capacity", "symmetry" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "GuQeZWbaGr", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Bor-Chun Chen", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Hui Xian Grace Lim", "gender": "unknown", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Mubarak Shah", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Rukhshanda Hussain", "gender": "Female", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "MS student" }, { "name": "Ser-Nam Lim", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Associate Professor" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 21, "n_ref": 34, "n_ref_all": 73, "n_fig": 18, "n_tab": 2, "L_tab": 2456, "n_element_tab": 29, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 47, "n_element_tab_1": 2, "formula_len_all": 310, "formula_len_all_1": 0, "len_all": 133603, "len_all_1": 48962, "len_abs": 1609, "len_title": 93, "len_sents": 41356, "len_sents_1": 24977, "n_sents": 289, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1617, "title": "AnyView: Few Shot Personalized View Transfer", "abs": "Fine-tuning generative models for concept driven personalization have witnessed tremendous growth ever since the arrival of methods like DreamBooth, Textual Inversion etc. Particularly, such techniques have been thoroughly explored for style-driven generation. Recently, diffusion models have also demonstrated impressive capabilities in view synthesis tasks, setting the foundation for exploring view-driven generation approaches. Motivated by these advancements, we investigate the capacity of a pretrained stable diffusion model to grasp ``what constitutes a view\" without relying on explicit 3D priors. Specifically, we base our method on a personalized text to image model, Dreambooth, given its strong ability to adapt to specific novel objects with a few shots. Our research reveals two interesting findings. First, we observe that Dreambooth can learn the high level concept of a view, compared to arguably more complex strategies which involve fine-tuning diffusions on large amounts of multi-view data. Second, we establish that the concept of a view can be disentangled and transferred to a novel object irrespective of the original object\u2019s identity from which the views are learnt. Motivated by this, we introduce a learning strategy, AnyView, which inherits a specific view through only one image sample of a single scene, and transfers the knowledge to a novel object, learnt from a few shots, using low rank adapters. Through extensive experiments we demonstrate that our method, albeit simple, is efficient in generating reliable view samples for in the wild images. Code and models will be released.", "keywords": [ "View Transfer", "Diffusion", "Generative" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "GtxzVbjPKC", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Baraah A. M. Sidahmed", "gender": "Female", "institution": "Universit\u00e4t des Saarlandes", "country": "DE", "position": "MS student" }, { "name": "Tatjana Chavdarova", "gender": "Female", "institution": "University of California, Berkeley", "country": "", "position": "Postdoc" } ], "n_formula": 37, "n_formula_1": 17, "n_ref_uni": 62, "n_ref": 114, "n_ref_all": 127, "n_fig": 9, "n_tab": 4, "L_tab": 513, "n_element_tab": 69, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 750, "n_element_tab_1": 57, "formula_len_all": 3168, "formula_len_all_1": 1279, "len_all": 172729, "len_all_1": 68530, "len_abs": 1612, "len_title": 150, "len_sents": 47610, "len_sents_1": 31528, "n_sents": 422, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1629, "title": "Variational Inequality Methods for Multi-Agent Reinforcement Learning: Performance and Stability Gains", "abs": "Multi-agent reinforcement learning (MARL) poses distinct challenges as agents learn strategies through experiences. Gradient-based methods often fail to converge in MARL, and performances are highly sensitive to initial random seeds, contributing to what has been termed the MARL reproducibility crisis. \nConcurrently, significant advances have been made in solving Variational Inequalities (VIs)---which include equilibrium-finding problems---particularly in addressing the non-converging rotational dynamics that impede convergence of traditional gradient-based optimization methods. \nThis paper explores the potential of leveraging VI-based techniques to improve MARL training. Specifically, we study the integration of VI methods---namely, Nested-Lookahead VI (nLA-VI) and Extragradient (EG)---into the multi-agent deep deterministic policy gradient (MADDPG) algorithm. \nWe present a VI reformulation of the actor-critic algorithm for both single- and multi-agent settings. We introduce three algorithms that use nLA-VI, EG, and a combination of both, named LA-MADDPG, EG-MADDPG, and LA-EG-MADDPG, respectively.\nOur empirical results show that these VI-based approaches yield significant performance improvements in benchmark environments, such as the zero-sum games: rock-paper-scissors and matching pennies, where equilibrium strategies can be quantitatively assessed, and the Multi-Agent Particle Environment: Predator-prey benchmark, where VI-based methods also yield balanced participation of agents from the same team, further highlighting the substantial impact of advanced optimization techniques on MARL performance.", "keywords": [ "multi-agent reinforcement learning", "game optimization", "Variational Inequality" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 1, 1 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "GtvuNrk58a", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Alex Vitvitskyi", "gender": "unknown", "institution": "DeepMind", "country": "GB", "position": "Research Engineer" }, { "name": "Christos Perivolaropoulos", "gender": "Male", "institution": "Google", "country": "GB", "position": "Researcher" }, { "name": "Federico Barbero", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Petar Veli\u010dkovi\u0107", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Affiliated Lecturer" }, { "name": "Razvan Pascanu", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Affiliate Member" } ], "n_formula": 17, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 71, "n_ref_all": 114, "n_fig": 16, "n_tab": 5, "L_tab": 641, "n_element_tab": 62, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1375, "formula_len_all_1": 596, "len_all": 187195, "len_all_1": 65808, "len_abs": 1296, "len_title": 69, "len_sents": 64221, "len_sents_1": 29997, "n_sents": 606, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1302, "title": "Round and Round We Go! What makes Rotary Positional Encodings useful?", "abs": "Positional Encodings (PEs) are a critical component of Transformer-based Large Language Models (LLMs), providing the attention mechanism with important sequence-position information. One of the most popular types of encoding used today in LLMs are Rotary Positional Encodings (RoPE), that rotate the queries and keys based on their relative distance. A common belief is that RoPE is useful because it helps to decay token dependency as relative distance increases. In this work, we argue that this is unlikely to be the core reason. We study the internals of a trained Gemma 7B model to understand how RoPE is being used at a mechanical level. We find that Gemma learns to use RoPE to construct robust `positional' attention patterns by exploiting the highest frequencies. We also find that, in general, Gemma greatly prefers to use the lowest frequencies of RoPE, which we suspect are used to carry semantic information. We mathematically prove interesting behaviours of RoPE and conduct experiments to verify our findings, proposing a modification of RoPE that fixes some highlighted issues and improves performance. We believe that this work represents an interesting step in better understanding PEs in LLMs, which we believe holds crucial value for scaling LLMs to large sizes and context lengths.", "keywords": [ "Large Language Models", "Transformers", "Positional Encodings", "Rotary Positional Encodings" ], "rating_list": [ 5, 5, 8, 5, 8 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 4, 3, 3 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "GtpubstM1D", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Mi Tian", "gender": "Male", "institution": "TAL Education Group", "country": "CN", "position": "cto" }, { "name": "Tianqiao Liu", "gender": "Male", "institution": "Tomorrow Advancing Life", "country": "CN", "position": "Researcher" }, { "name": "Tongqing", "gender": "Male", "institution": "Tomorrow Advancing Life", "country": "CN", "position": "Algorithm Engineer" }, { "name": "Weiqi Luo", "gender": "Male", "institution": "Jinan University", "country": "CN", "position": "Full Professor" }, { "name": "Zitao Liu", "gender": "Male", "institution": "Jinan University", "country": "CN", "position": "Full Professor" }, { "name": "Zui Chen", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 27, "n_ref": 76, "n_ref_all": 97, "n_fig": 6, "n_tab": 7, "L_tab": 1301, "n_element_tab": 187, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 448, "n_element_tab_1": 28, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 176267, "len_all_1": 68418, "len_abs": 1995, "len_title": 173, "len_sents": 56286, "len_sents_1": 35648, "n_sents": 393, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 132, "L_abs": 1912, "title": "Advancing Mathematical Reasoning in Language Models: The Impact of Problem-Solving Data, Data Synthesis Methods, and Training Stages", "abs": "Mathematical reasoning remains a challenging area for large language models (LLMs), prompting the development of math-specific LLMs such as LLEMMA, DeepSeekMath, and Qwen2-Math, among others. These models typically follow a two-stage training paradigm: pre-training with math-related corpora and post-training with problem datasets for supervised fine-tuning (SFT). Despite these efforts, the improvements in mathematical reasoning achieved through continued pre-training (CPT) are often less significant compared to those obtained via SFT. This study addresses this discrepancy by exploring alternative strategies during the pre-training phase, focusing on the use of problem-solving data over general mathematical corpora.\nWe investigate three primary research questions: (1) Can problem-solving data enhance the model's mathematical reasoning capabilities more effectively than general mathematical corpora during CPT? (2) Are synthetic data from the same source equally effective, and which synthesis methods are most efficient? (3) How do the capabilities developed from the same problem-solving data differ between the CPT and SFT stages, and what factors contribute to these differences?\nOur findings indicate that problem-solving data significantly enhances the model's mathematical capabilities compared to general mathematical corpora. We also identify effective data synthesis methods, demonstrating that the tutorship amplification synthesis method achieves the best performance. Furthermore, while SFT facilitates instruction-following abilities, it underperforms compared to CPT with the same data, which can be partially attributed to its poor learning capacity for more challenging problem-solving data. These insights provide valuable guidance for optimizing the mathematical reasoning capabilities of LLMs, culminating in our development of a powerful mathematical base model called MathGPT-8B.", "keywords": [ "LLM continue pretrain;math problem solving;data synthesis" ], "rating_list": [ 6, 6, 8, 8, 3 ], "soundness_list": [ 2, 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 4, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 3, 3 ] }, { "paper_id": "GtlV6o1yUy", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Haogeng Liu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Huaibo Huang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Quanzeng You", "gender": "Male", "institution": "ByteDance", "country": "US", "position": "Researcher" }, { "name": "Ran He", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xiaotian Han", "gender": "Male", "institution": "ByteDance", "country": "US", "position": "Researcher" }, { "name": "Yiren Jian", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 25, "n_ref": 67, "n_ref_all": 97, "n_fig": 9, "n_tab": 6, "L_tab": 991, "n_element_tab": 184, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 261, "n_element_tab_1": 59, "formula_len_all": 290, "formula_len_all_1": 220, "len_all": 135743, "len_all_1": 58756, "len_abs": 1099, "len_title": 133, "len_sents": 36728, "len_sents_1": 27971, "n_sents": 287, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1105, "title": "eRAM-V: From Interaction to Integration in Efficient Multimodal Large Language Models", "abs": "Multimodal large language models (MLLMs) have made significant progress in recent years, yet the interaction between vision and language representations remains underexplored. Prior work has primarily relied on empirical heuristics to guide architecture design. While effective, this approach can lead to sub-optimal designs and computational redundancy. In this work, we examine the fusion process between visual and textual data. Our findings indicate that in auto-regressive MLLMs, fine-grained interactions between visual and text tokens primarily occur in the middle layers. This leads to redundancy in the shallow and deep layers, where modeling only selected visual representations is sufficient. Based on these insights, we introduce eRAM-V, an MLLM that balances computational efficiency and performance. eRAM-V models selected visual features across all layers and integrates fine-grained visual features at specific layers, as needed. Extensive experiments show that eRAM-V outperforms baseline models with equivalent computational budgets, achieving superior results across various benchmarks.", "keywords": [ "Multimodal Large Language Model; Interpretability of MLLM" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "GtlRN48XYA", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Qianqian Yang", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "Associate Professor" }, { "name": "Shunpu Tang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yuanchao Shu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yuxuan Yan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhiguo Shi", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" } ], "n_formula": 34, "n_formula_1": 28, "n_ref_uni": 36, "n_ref": 60, "n_ref_all": 85, "n_fig": 15, "n_tab": 10, "L_tab": 4145, "n_element_tab": 119, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 396, "n_element_tab_1": 12, "formula_len_all": 1525, "formula_len_all_1": 1213, "len_all": 150039, "len_all_1": 49851, "len_abs": 1653, "len_title": 150, "len_sents": 35910, "len_sents_1": 21108, "n_sents": 267, "n_sents_1": 150, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1661, "title": "FeDeRA: Efficient Fine-tuning of Language Models in Federated Learning Leveraging Weight Decomposition", "abs": "Federated learning (FL) is a widely used privacy-preserving approach for distributed training that avoids the need to collect data from individual users. In this paper, we investigate fine-tuning pre-trained language models (PLMs) in an FL setting and leverage parameter-efficient fine-tuning (PEFT) methods to reduce computational and communication costs. However, non-IID data in federated learning significantly degrades the performance of PEFT, with the degradation worsening as data heterogeneity increases. To address this, we propose FeDeRA, an FL approach for fine-tuning PLMs that incorporates an effective extension of the low-rank adaptation (LoRA) method. Specifically, FeDeRA initializes the low-rank matrices using Singular Value Decomposition (SVD) on the pre-trained weight matrices, rather than the zero or random initialization used in the original LoRA method. Analyzing weight updates during training reveals that FeDeRA reduces weight oscillations, enabling faster and more efficient fine-tuning of PLMs in FL with non-IID data. Experimental results across multiple NLP tasks and models show that FeDeRA outperforms all PEFT-based baselines in task performance and, in some cases, even matches or exceeds the performance of full-parameter fine-tuning. FeDeRA also greatly enhances training efficiency, reducing training time by up to 97.3\\% compared to full-parameter fine-tuning and up to 74.6\\% compared to the fastest PEFT baseline in practical FL settings. Furthermore, FeDeRA demonstrates greater robustness to data heterogeneity than all other PEFT methods, highlighting the effectiveness of its proposed initialization in FL systems.", "keywords": [ "federated learning", "fine-tune", "low rank adaption" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "GtHRhpgpzB", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Jiacheng Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Shida Sun", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Yinuo Liao", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Yuanshen Guan", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Zhiwei Xiong", "gender": "Male", "institution": "USTC", "country": "CN", "position": "Professor" }, { "name": "ruikang xu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 34, "n_ref": 60, "n_ref_all": 87, "n_fig": 16, "n_tab": 5, "L_tab": 1352, "n_element_tab": 155, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 405, "formula_len_all_1": 226, "len_all": 125842, "len_all_1": 58610, "len_abs": 1367, "len_title": 90, "len_sents": 35818, "len_sents_1": 27979, "n_sents": 283, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 1372, "title": "Learning Gain Map for Inverse Tone Mapping", "abs": "For a more compatible and consistent high dynamic range (HDR) viewing experience, a new image format with a double-layer structure has been developed recently, which incorporates an auxiliary Gain Map (GM) within a standard dynamic range (SDR) image for adaptive HDR display. This new format motivates us to introduce a new task termed Gain Map-based Inverse Tone Mapping (GM-ITM), which focuses on learning the corresponding GM of an SDR image instead of directly estimating its HDR counterpart, thereby enabling a more effective up-conversion by leveraging the advantages of GM. The main challenge in this task, however, is to accurately estimate regional intensity variation with the fluctuating peak value. To this end, we propose a dual-branch network named GMNet, consisting of a Local Contrast Restoration (LCR) branch and a Global Luminance Estimation (GLE) branch to capture pixel-wise and image-wise information for GM estimation. Moreover, to facilitate the future research of the GM-ITM task, we build both synthetic and real-world datasets for comprehensive evaluations: synthetic SDR-GM pairs are generated from existing HDR resources, and real-world SDR-GM pairs are captured by mobile devices. Extensive experiments on these datasets demonstrate the superiority of our proposed GMNet over existing HDR-related methods both quantitatively and qualitatively.", "keywords": [ "Computational Photography", "Inverse Tone Mapping", "Gain Map" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 5, 4, 3, 2 ] }, { "paper_id": "GsR3zRCRX5", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Ayush Bharti", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Principal Researcher" }, { "name": "Yogesh Verma", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "PhD student" }, { "name": "Vikas K Garg", "gender": "Male", "institution": "Aalto University", "country": "FI", "position": "Assistant Professor" } ], "n_formula": 29, "n_formula_1": 14, "n_ref_uni": 73, "n_ref": 120, "n_ref_all": 145, "n_fig": 8, "n_tab": 11, "L_tab": 3185, "n_element_tab": 212, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 189, "n_element_tab_1": 4, "formula_len_all": 2909, "formula_len_all_1": 1204, "len_all": 191979, "len_all_1": 63127, "len_abs": 924, "len_title": 121, "len_sents": 49181, "len_sents_1": 30094, "n_sents": 364, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 927, "title": "Robust Simulation-Based Inference under Missing Data", "abs": "Simulation-based inference (SBI) methods typically require fully observed data to infer parameters of models with intractable likelihood functions. However, datasets often contain missing values due to incomplete observations, data corruptions (common in astrophysics), or instrument limitations (e.g., in high-energy physics applications). In such scenarios, missing data must be imputed before applying any SBI method. This work formalizes the problem of missing data in SBI and demonstrates that naive imputation methods can introduce bias into the SBI posterior. We introduce a novel method that addresses this issue by jointly learning the imputation model and the inference network within a neural posterior estimation (NPE) framework. Extensive empirical results on SBI benchmarks show that our approach provides robust inference outcomes compared to baselines, for varying levels of missing data, while being amortized.", "keywords": [ "Simulation-based inference", "likelihood-free inference", "approximate Bayesian computation", "neural posterior estimation", "missing data" ], "rating_list": [ 3, 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 2, 4, 4, 4, 4 ] }, { "paper_id": "GsGmdxcFNL", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Hao Su", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Shangzhe Li", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Zhiao Huang", "gender": "Male", "institution": "University of California, San Diego, University of California, San Diego", "country": "", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 21, "n_ref_uni": 35, "n_ref": 116, "n_ref_all": 145, "n_fig": 14, "n_tab": 4, "L_tab": 351, "n_element_tab": 56, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1643, "n_element_tab_1": 18, "formula_len_all": 2144, "formula_len_all_1": 1698, "len_all": 151992, "len_all_1": 59922, "len_abs": 1132, "len_title": 103, "len_sents": 47896, "len_sents_1": 25916, "n_sents": 393, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1132, "title": "Reward-free World Models for Online Imitation Learning", "abs": "Imitation learning (IL) enables agents to acquire skills directly from expert demonstrations, providing a compelling alternative to reinforcement learning. However, prior online IL approaches struggle with complex tasks characterized by high-dimensional inputs and complex dynamics. In this work, we propose a novel approach to online imitation learning that leverages reward-free world models. Our method learns environmental dynamics entirely in latent spaces without reconstruction, enabling efficient and accurate modeling. We adopt the inverse soft-Q learning objective, reformulating the optimization process in the Q-policy space to mitigate the instability associated with traditional optimization in reward-policy space. By employing a learned latent dynamics model and planning for control, our approach consistently achieves stable, expert-level performance in tasks with high-dimensional observation or action spaces and intricate dynamics. We evaluate our method on a diverse set of benchmarks, including DMControl, MyoSuite, and ManiSkill2, demonstrating superior empirical performance compared to existing approaches.", "keywords": [ "World Models", "Imitation Learning", "Inverse Reinforcement Learning" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "GsCMKwyfWm", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Kimon Fountoulakis", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "Associate Professor" }, { "name": "Mohammadreza Ghofrani", "gender": "unknown", "institution": "University of Ottawa", "country": "CA", "position": "PhD student" }, { "name": "Muhammad Fetrat Qharabagh", "gender": "Male", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 26, "n_ref": 59, "n_ref_all": 148, "n_fig": 42, "n_tab": 13, "L_tab": 4690, "n_element_tab": 505, "n_fig_1": 17, "n_tab_1": 1, "L_tab_1": 93, "n_element_tab_1": 11, "formula_len_all": 103, "formula_len_all_1": 0, "len_all": 190391, "len_all_1": 69082, "len_abs": 776, "len_title": 123, "len_sents": 65110, "len_sents_1": 32071, "n_sents": 590, "n_sents_1": 286, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 778, "title": "LVLM-COUNT: Enhancing the Counting Ability of Large Vision-Language Models", "abs": "Counting is a fundamental skill for various visual tasks in real-life applications, requiring both object recognition and robust counting capabilities. Despite their advanced visual perception, large vision-language models (LVLMs) struggle with counting tasks, especially when the number of objects exceeds those commonly encountered during training. We enhance LVLMs\u2019 counting abilities using a divide-and conquer approach, breaking counting problems into sub-counting tasks. Unlike prior methods, which do not generalize well to counting datasets on which they have not been trained, our method performs well on new datasets without any additional training or fine-tuning. We demonstrate that our approach enhances counting capabilities across various datasets and benchmarks.", "keywords": [ "Counting", "Large vision-language models" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "GrmFFxGnOR", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Frederick Tung", "gender": "Male", "institution": "Borealis AI", "country": "", "position": "Researcher" }, { "name": "Hossein Hajimirsadeghi", "gender": "Male", "institution": "Borealis AI", "country": "", "position": "Principal Researcher" }, { "name": "Leo Feng", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "PhD student" }, { "name": "Mohamed Osama Ahmed", "gender": "Male", "institution": "Borealis AI", "country": "", "position": "Machine Learning Researcher" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 14, "n_ref_uni": 44, "n_ref": 96, "n_ref_all": 119, "n_fig": 5, "n_tab": 7, "L_tab": 2603, "n_element_tab": 138, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 558, "n_element_tab_1": 36, "formula_len_all": 3185, "formula_len_all_1": 1950, "len_all": 167978, "len_all_1": 62972, "len_abs": 1040, "len_title": 72, "len_sents": 50060, "len_sents_1": 28358, "n_sents": 415, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 24, "L_abs": 1043, "title": "Were RNNs All We Needed?", "abs": "The introduction of Transformers in 2017 reshaped the landscape of deep learning. Originally proposed for sequence modelling, Transformers have since achieved widespread success across various domains. However, the scalability limitations of Transformers\u2014particularly with respect to sequence length\u2014have sparked renewed interest in novel recurrent models that are parallelizable during training, offer comparable performance, and scale more effectively.\nIn this work, we revisit sequence modelling from a historical perspective, focusing on Recurrent Neural Networks (RNNs), which dominated the field for two decades before the rise of Transformers. Specifically, we examine LSTMs (1997) and GRUs (2014). We demonstrate that by simplifying these models, we can derive minimal versions (minLSTMs and minGRUs) that (1) use fewer parameters than their traditional counterparts, (2) are fully parallelizable during training, and (3) achieve surprisingly competitive performance on a range of tasks, rivalling recent models including Transformers.", "keywords": [ "Recurrent Neural Networks", "RNNs", "Sequence Modelling", "Efficiency", "LSTMs", "GRUs", "Parallel Scan" ], "rating_list": [ 3, 3, 6, 8 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 1, 1, 2, 4 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "GraOHDxFjl", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Chengyue Wu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jiahao Wang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Kai Chen", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Kaipeng Zhang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Mengzhao Chen", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Songyang Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Taiqiang Wu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Wenqi Shao", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yong Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ping Luo", "gender": "unknown", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 47, "n_ref": 87, "n_ref_all": 132, "n_fig": 12, "n_tab": 16, "L_tab": 5408, "n_element_tab": 539, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 558, "n_element_tab_1": 80, "formula_len_all": 704, "formula_len_all_1": 270, "len_all": 172124, "len_all_1": 64473, "len_abs": 1701, "len_title": 83, "len_sents": 47647, "len_sents_1": 29719, "n_sents": 402, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1710, "title": "LLaMA Decoder As Vision Transformer", "abs": "Using the same architecture for text and image is important for AI standardization. Recent multimodal models use a decoder-only Transformer to generate text and an encoder-only Transformer to extract image features. Can images use exactly the same language architecture? To answer this question, we aim at a LLaMa decoder as vision Transformer (ViT) classifier in this paper. Specifically, we start\nour trajectory by \u201cLLaMAfy\u201d a standard ViT step-by-step, i.e., feed-forward net, normalization layer, causal self-attention and positional embedding, and point out a key issue\u2014attention collapse\u2014that result in the failure to the network training. Motivated by this observation, we propose post-sequence class token, enabling causal self-attention to efficiently capture the entire image\u2019s information. To improve model optimization behavior and enhance performance, we then introduce a soft mask strategy to gradually transform the attention from bi-directional to causal mode. The tailored model, dubbed as image LLaMA (iLLaMA), maintains high consistency with LLaMA architecture, while matching up well against ViT, achieving 75.1% ImageNet top-1 accuracy with only 5.7M parameters. Scaling the model to \u223c310M and pre-training on ImageNet-21K further enhances the accuracy to 86.0%. Its causal self-attention boosts computational efficiency and learns complex representation by elevating attention map ranks. Extensive experiments demonstrate iLLaMA\u2019s reliable properties: shape-texture bias, calibration, quantization compatibility, ADE20K segmentation and CIFAR transfer learning. We hope our study can kindle fresh views to visual architectures in the era of LLMs and contributes to standardized AI models.", "keywords": [ "Decoder-only; Vision Transformer; LLaMA" ], "rating_list": [ 8, 3, 8, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 2, 4, 3 ], "contribution_list": [ 3, 1, 3, 1 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "GrRo9uV3OH", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Aldo Pacchiano", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "Assistant Professor" }, { "name": "Ioannis Paschalidis", "gender": "Male", "institution": "Boston University", "country": "US", "position": "Full Professor" }, { "name": "Yilei Chen", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "PhD student" } ], "n_formula": 108, "n_formula_1": 35, "n_ref_uni": 26, "n_ref": 52, "n_ref_all": 58, "n_fig": 0, "n_tab": 6, "L_tab": 691, "n_element_tab": 28, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 54, "n_element_tab_1": 4, "formula_len_all": 10878, "formula_len_all_1": 3081, "len_all": 169869, "len_all_1": 69642, "len_abs": 1079, "len_title": 117, "len_sents": 51482, "len_sents_1": 28525, "n_sents": 531, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1300, "title": "Sample Efficient Multiple-policy Evaluation in Reinforcement Learning", "abs": "We study the multiple-policy evaluation problem where we are given a set of $K$ policies and the goal is to evaluate their performance (expected total reward over a fixed horizon) to an accuracy $\\epsilon$ with probability at least $1-\\delta$. We propose a sample-efficient algorithm named \\CAESAR for this problem. Our approach is based on computing an approximate optimal offline sampling distribution and using the data sampled from it to perform the simultaneous estimation of the policy values. \\CAESAR has two phases. In the first we produce coarse estimates of the visitation distributions of the target policies at a low order sample complexity rate that scales with $\\tilde{O}(\\frac{1}{\\epsilon})$. In the second phase, we approximate the optimal offline sampling distribution and compute the importance weighting ratios for all target policies by minimizing a step-wise quadratic loss function inspired by the DualDICE \\citep{nachum2019dualdice} objective. Up to low order and logarithmic terms \\CAESAR achieves a sample complexity $\\tilde{O}\\left(\\frac{H^4}{\\epsilon^2}\\sum_{h=1}^H\\min_{\\mu_h}\\max_{k\\in[K]}\\sum_{s,a}\\frac{(d_h^{\\pi^k}(s,a))^2}{\\mu_h(s,a)}\\right)$, where $d^{\\pi}$ is the visitation distribution of policy $\\pi$, $\\mu$ is the sampling distribution, and $H$ is the horizon.", "keywords": [ "Theory", "Reinforcement learning", "Policy evaluation", "Sample complexity" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "GrHewano8m", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Arian Prabowo", "gender": "unknown", "institution": "University of New South Wales", "country": "AU", "position": "Postdoc" }, { "name": "Du Yin", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "PhD student" }, { "name": "Hao Xue", "gender": "Male", "institution": "University of New South Wales", "country": "", "position": "Postdoc" }, { "name": "Shuang Ao", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "Postdoc" }, { "name": "Flora Salim", "gender": "Female", "institution": "University of New South Wales", "country": "AU", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 5, "n_ref": 16, "n_ref_all": 33, "n_fig": 4, "n_tab": 6, "L_tab": 7507, "n_element_tab": 902, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 7206, "n_element_tab_1": 880, "formula_len_all": 140, "formula_len_all_1": 140, "len_all": 114628, "len_all_1": 54128, "len_abs": 1475, "len_title": 131, "len_sents": 18352, "len_sents_1": 16577, "n_sents": 132, "n_sents_1": 109, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1484, "title": "XXLTraffic: Expanding and Extremely Long Traffic forecasting beyond test adaptation", "abs": "Traffic forecasting is crucial for smart cities and intelligent transportation initiatives, where deep learning has made significant progress in modeling complex spatio-temporal patterns in recent years. However, current public datasets have limitations in reflecting the distribution shift nature of real-world scenarios, characterized by continuously evolving infrastructures, varying temporal distributions, and long temporal gaps due to sensor downtimes or changes in traffic patterns. These limitations inevitably restrict the practical applicability of existing traffic forecasting datasets. To bridge this gap, we present XXLTraffic, the longest available public traffic dataset with the longest timespan collected from Los Angeles, USA, and New South Wales, Australia, curated to support research in extremely long forecasting beyond test adaptation. Our benchmark includes both typical time-series forecasting settings with hourly and daily aggregated data and novel configurations that introduce gaps and down-sample the training size to better simulate practical constraints. We anticipate the new XXLTraffic will provide a fresh perspective for the time-series and traffic forecasting communities. It would also offer a robust platform for developing and evaluating models designed to tackle the extremely long forecasting problems beyond test adaptation. Our dataset supplements existing spatio-temporal data resources and leads to new research directions in this domain.", "keywords": [ "spatio-temporal", "traffic forecasting", "time-series" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 4, 2, 1, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 4, 2, 2, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "GrDne4055L", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Hossein Mirzaei", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Mackenzie W Mathis", "gender": "Female", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Assistant Professor" } ], "n_formula": 19, "n_formula_1": 6, "n_ref_uni": 48, "n_ref": 87, "n_ref_all": 138, "n_fig": 13, "n_tab": 16, "L_tab": 11697, "n_element_tab": 1175, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3951, "n_element_tab_1": 278, "formula_len_all": 1465, "formula_len_all_1": 675, "len_all": 347303, "len_all_1": 73517, "len_abs": 1853, "len_title": 136, "len_sents": 95862, "len_sents_1": 30131, "n_sents": 774, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1864, "title": "Adversarially Robust Out-of-Distribution Detection Using Lyapunov-Stabilized Embeddings", "abs": "Despite significant advancements in out-of-distribution (OOD) detection, existing methods still struggle to maintain robustness against adversarial attacks, compromising their reliability in critical real-world applications. Previous studies have attempted to address this challenge by exposing detectors to auxiliary OOD datasets alongside adversarial training. However, the increased data complexity inherent in adversarial training, and the myriad of ways that OOD samples can arise during testing, often prevent these approaches from establishing robust decision boundaries. To address these limitations, we propose AROS, a novel approach leveraging neural ordinary differential equations (NODEs) with Lyapunov stability theorem in order to obtain robust embeddings for OOD detection.\nBy incorporating a tailored loss function, we apply Lyapunov stability theory to ensure that both in-distribution (ID) and OOD data converge to stable equilibrium points within the dynamical system. This approach encourages any perturbed input to return to its stable equilibrium, thereby enhancing the model\u2019s robustness against adversarial perturbations. To not use additional data, we generate fake OOD embeddings by sampling from low-likelihood regions of the ID data feature space, approximating the boundaries where OOD data are likely to reside. To then further enhance robustness, we propose the use of an orthogonal binary layer following the stable feature space, which maximizes the separation between the equilibrium points of ID and OOD samples. We validate our method through extensive experiments across several benchmarks, demonstrating superior performance, particularly under adversarial attacks. Notably, our approach improves robust detection performance from 37.8\\% to 80.1\\% on CIFAR-10 vs. CIFAR-100 and from 29.0\\% to 67.0\\% on CIFAR-100 vs. CIFAR-10.", "keywords": [ "Out-of-Distribution Detection", "Adversarial Robustness", "Stability Theorem", "Neural Ordinary Differential Equations", "Lyapunov Theorem" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 2, 2, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "Gr8nHvOivO", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Litian Liu", "gender": "Female", "institution": "Qualcomm Inc, QualComm", "country": "US", "position": "Researcher" }, { "name": "Yao Qin", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 70, "n_ref_all": 110, "n_fig": 2, "n_tab": 13, "L_tab": 8620, "n_element_tab": 427, "n_fig_1": 2, "n_tab_1": 11, "L_tab_1": 10572, "n_element_tab_1": 452, "formula_len_all": 1006, "formula_len_all_1": 1006, "len_all": 135028, "len_all_1": 85494, "len_abs": 1163, "len_title": 113, "len_sents": 34709, "len_sents_1": 32141, "n_sents": 265, "n_sents_1": 229, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1183, "title": "Detecting Out-of-Distribution through the Lens of Neural Collapse", "abs": "Out-of-Distribution (OOD) detection is essential for safe deployment; however, existing detectors exhibit generalization discrepancies and cost concerns. To address this, we propose a highly versatile and efficient OOD detector inspired by the trend of Neural Collapse on practical models, without requiring complete collapse. By analyzing this trend, we discover that features of in-distribution (ID) samples cluster closer to the weight vectors compared to features of OOD samples. Additionally, we reveal that ID features tend to expand in space to structure a simplex Equiangular Tight Framework, which explains the prevalent observation that ID features reside further from the origin than OOD features. Taking both insights from Neural Collapse into consideration, our OOD detector utilizes feature proximity to weight vectors and further complements this perspective by using feature norms to filter OOD samples. Extensive experiments on \\emph{off-the-shelf} models demonstrate the efficiency and effectiveness of our OOD detector across diverse classification tasks and model architectures, mitigating generalization discrepancies and improving \\emph{overall} performance.", "keywords": [ "Out-of-Distribution Detection" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "GqsepTIXWy", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Qiaoqiao Ding", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Researcher" }, { "name": "Xiaoqun Zhang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "", "position": "Full Professor" }, { "name": "Zhe Xiong", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 21, "n_ref_uni": 16, "n_ref": 23, "n_ref_all": 45, "n_fig": 7, "n_tab": 5, "L_tab": 3195, "n_element_tab": 78, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 3038, "n_element_tab_1": 73, "formula_len_all": 2333, "formula_len_all_1": 2281, "len_all": 94845, "len_all_1": 57364, "len_abs": 1694, "len_title": 89, "len_sents": 25099, "len_sents_1": 22932, "n_sents": 179, "n_sents_1": 164, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1510, "title": "Bi-modality medical images synthesis by a bi-directional discrete process matching method", "abs": "Recently, medical image synthesis gains more and more popularity, along with the rapid development of generative models. Medical image synthesis aims to generate an unacquired image modality, often from other observed data modalities. Synthesized images can be used for clinical diagnostic assistance, data augmentation for model training and validation or image quality improving. In the meanwhile, the flow-based models are among the successful generative models for the ability of generating realistic and high-quality synthetic images. However, most flow-based models require to calculate flow ordinary different equation (ODE) evolution steps in synthesis process, for which the performances are significantly limited by heavy computation time due to a large number of time iterations. In this paper, we propose a novel flow-based model, namely bi-directional Discrete Process Matching (Bi-DPM) to accomplish the bi-modality image synthesis tasks. Different to other flow matching based models, we propose to utilize both forward and backward ODE flows and enhance the consistency on the intermediate images over a few discrete time steps, resulting in a synthesis process maintaining high-quality generations for both modalities under the guidance of paired data. Our experiments on three datasets of MRI T1/T2 and CT/MRI demonstrate that Bi-DPM outperforms other state-of-the-art flow-based methods for bi-modality image synthesis, delivering higher image quality with accurate anatomical regions.", "keywords": [ "Bi-modality Images", "Medical Image Synthesis", "Flow-based Model" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "Gqs0ERAKAv", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Ivaxi Sheth", "gender": "Female", "institution": "CISPA, saarland university, saarland informatics campus", "country": "", "position": "PhD student" }, { "name": "Mario Fritz", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Full Professor" }, { "name": "Sahar Abdelnabi", "gender": "Male", "institution": "Microsoft", "country": "GB", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 42, "n_ref": 74, "n_ref_all": 102, "n_fig": 12, "n_tab": 14, "L_tab": 7068, "n_element_tab": 439, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 813, "n_element_tab_1": 17, "formula_len_all": 1247, "formula_len_all_1": 892, "len_all": 235740, "len_all_1": 68640, "len_abs": 1773, "len_title": 103, "len_sents": 82748, "len_sents_1": 32053, "n_sents": 508, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1262, "title": "Filling the Gaps: LLMs for Causal Hypothesis Generation", "abs": "Scientific discovery is a catalyst for human intellectual advances, driven by the cycle of hypothesis generation, experimental design, data evaluation, and iterative assumption refinement. This process, while crucial, is expensive and heavily dependent on the domain knowledge of scientists to generate hypotheses and navigate the scientific cycle. Central to this is causality, the ability to establish the relationship between the cause and the effect. Motivated by the scientific discovery process, in this work, we formulate a novel task where the input is a partial causal graph with missing variables, and the output is a hypothesis about the missing variables to complete the partial graph. We design a benchmark with varying difficulty levels and knowledge assumptions about the causal graph. With the growing interest in using Large Language Models (LLMs) to assist in scientific discovery, we benchmark open-source and closed models on our testbed. We show the strong ability of LLMs to hypothesize the mediation variables between a cause and its effect. In contrast, they underperform in hypothesizing the cause and effect variables themselves. We also observe surprising results where some of the open-source models outperform the closed GPT-4 model.", "keywords": [ "LLMs", "hypothesis generation", "causal graph" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "GqhvJ1o8m5", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jian Shi", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Peter Wonka", "gender": "Male", "institution": "Snap Inc.", "country": "SA", "position": "Researcher" }, { "name": "Zhenyu Li", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 22, "n_ref": 40, "n_ref_all": 60, "n_fig": 8, "n_tab": 3, "L_tab": 13, "n_element_tab": 1, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 2654, "n_element_tab_1": 63, "formula_len_all": 172, "formula_len_all_1": 172, "len_all": 98816, "len_all_1": 61828, "len_abs": 1141, "len_title": 126, "len_sents": 30107, "len_sents_1": 27292, "n_sents": 252, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1176, "title": "ImmersePro: End-to-End Stereo Video Synthesis Via Implicit Disparity Learning", "abs": "We introduce \\textit{ImmersePro}, an innovative framework specifically designed to transform single-view videos into stereo videos. This framework utilizes a novel dual-branch architecture comprising a disparity branch and a context branch on video data by leveraging spatial-temporal attention mechanisms. \\textit{ImmersePro} employs implicit disparity guidance, enabling the generation of stereo pairs from video sequences without the need for explicit disparity maps, thus reducing potential errors associated with disparity estimation models.\nIn addition to the technical advancements, we introduce the YouTube-SBS dataset, a comprehensive collection of 423 stereo videos sourced from YouTube. This dataset is unprecedented in its scale, featuring over 7 million stereo pairs, and is designed to facilitate training and benchmarking of stereo video generation models. Our experiments demonstrate the effectiveness of \\textit{ImmersePro} in producing high-quality stereo videos, offering significant improvements over existing methods.\nCompared to the best competitor stereo-from-mono we quantitatively improve the results by 11.76\\% (L1), 6.39\\% (SSIM), and 5.10\\% (PSNR).", "keywords": [ "Video Stereo Conversion; Stereo Vision;" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "GqGoa44obw", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Ananthram Swami", "gender": "unknown", "institution": "Army Research Labortory", "country": "US", "position": "ST" }, { "name": "Ming Shi", "gender": "unknown", "institution": "State University of New York at Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Ness Shroff", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "Full Professor" }, { "name": "Yingbin Liang", "gender": "Female", "institution": "The Ohio State University", "country": "US", "position": "Professor" } ], "n_formula": 311, "n_formula_1": 40, "n_ref_uni": 38, "n_ref": 54, "n_ref_all": 112, "n_fig": 5, "n_tab": 4, "L_tab": 746, "n_element_tab": 37, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 244, "n_element_tab_1": 17, "formula_len_all": 40159, "formula_len_all_1": 3667, "len_all": 297853, "len_all_1": 73061, "len_abs": 1434, "len_title": 156, "len_sents": 78502, "len_sents_1": 31227, "n_sents": 844, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1444, "title": "RLHF with Inconsistent Multi-Agent Feedback Under General Function Approximation: A Theoretical Perspective", "abs": "Reinforcement learning from human feedback (RLHF) has been widely studied, as a method for leveraging feedback from human evaluators to guide the learning process. However, existing theoretical analyses typically assume that the human feedback is generated by the ground-truth reward function. This may not be true in practice, because the reward functions in human minds for providing feedback are usually different from the ground-truth reward function, e.g., due to diverse personal experiences and inherent biases. Such inconsistencies could lead to undesirable outcomes when applying existing algorithms, particularly when considering feedback from heterogeneous agents. Therefore, in this paper, we make the first effort to investigate a more practical and general setting of RLHF, where feedback could be generated by multiple agents with reward functions differing from the ground truth. To address this challenge, we develop a new algorithm with novel ideas for handling inconsistent multi-agent feedback, including a Steiner-Point-based confidence set to exploit the benefits of *multi-agent* feedback and a new weighted importance sampling method to manage complexity issues arising from *inconsistency*. Our theoretical analysis develops new methods to demonstrate the optimality of our algorithm. This result is the first of its kind to demonstrate the fundamental impact and potential of inconsistent multi-agent feedback in RLHF.", "keywords": [ "RLHF theory", "inconsistent multi-agent feedback", "regret analysis" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 1, 1, 1 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 1, 3, 2 ] }, { "paper_id": "GqBO71SPjL", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Diane Larlus", "gender": "Female", "institution": "NAVER LABS Europe", "country": "FR", "position": "Principal Researcher" }, { "name": "Julien Mairal", "gender": "unknown", "institution": "Inria", "country": "FR", "position": "Research Scientist" }, { "name": "Juliette Marrie", "gender": "Female", "institution": "Naver Labs Europe", "country": "", "position": "PhD student" }, { "name": "Michael Arbel", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Researcher" }, { "name": "Romain Menegaux", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 21, "n_ref": 166, "n_ref_all": 224, "n_fig": 11, "n_tab": 3, "L_tab": 1253, "n_element_tab": 301, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 557, "n_element_tab_1": 64, "formula_len_all": 384, "formula_len_all_1": 426, "len_all": 174750, "len_all_1": 68213, "len_abs": 1094, "len_title": 130, "len_sents": 56521, "len_sents_1": 32837, "n_sents": 409, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1102, "title": "LUDVIG: Learning-free Uplifting of 2D Visual Features to Gaussian Splatting Scenes", "abs": "We address the problem of extending the capabilities of vision foundation models\nsuch as DINO, SAM, and CLIP, to 3D tasks. Specifically, we introduce a novel\nmethod to uplift 2D image features into 3D Gaussian Splatting scenes. Unlike traditional approaches that rely on minimizing a reconstruction\nloss, our method employs a simpler and more efficient feature aggregation\ntechnique, augmented by a graph diffusion mechanism. Graph diffusion\nenriches features from a given model, such as CLIP, by leveraging pairwise\nsimilarities that encode 3D geometry or similarities induced by another embedding\nlike DINOv2. Our approach achieves performance comparable to the state of\nthe art on multiple downstream tasks while delivering significant speed-ups.\nNotably, we obtain competitive segmentation results using generic\nDINOv2 features, despite DINOv2 not being trained on millions of\nannotated segmentation masks like SAM. When applied to CLIP\nfeatures, our method demonstrates strong performance in open-vocabulary,\nlanguage-based object detection tasks, highlighting the versatility of our \napproach.", "keywords": [ "3D vision", "vision foundation models", "graph diffusion" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "Gq7RDMeZi4", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 8, "author_info_list": [ { "name": "David Wipf", "gender": "Male", "institution": "Amazon AI Research Lab", "country": "", "position": "Principal Research Scientist" }, { "name": "Haitian Jiang", "gender": "unknown", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Minjie Wang", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Senior Applied Scientist" }, { "name": "Renjie Liu", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Xiao Yan", "gender": "Male", "institution": "Centre for Perceptual and Interactive Intelligence", "country": "HK", "position": "Researcher" }, { "name": "Yichuan Wang", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Zengfeng Huang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Zhenkun Cai", "gender": "unknown", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 16, "n_ref_uni": 47, "n_ref": 140, "n_ref_all": 172, "n_fig": 5, "n_tab": 9, "L_tab": 1908, "n_element_tab": 161, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 832, "n_element_tab_1": 84, "formula_len_all": 3119, "formula_len_all_1": 1721, "len_all": 200538, "len_all_1": 76790, "len_abs": 1109, "len_title": 125, "len_sents": 65490, "len_sents_1": 34778, "n_sents": 466, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1111, "title": "Forming Scalable, Convergent GNN Layers that Minimize a Sampling-Based Energy", "abs": "Among the many variants of graph neural network (GNN) architectures capable of modeling data with cross-instance relations, an important subclass involves layers designed such that the forward pass iteratively reduces a graph-regularized energy function of interest. In this way, node embeddings produced at the output layer dually serve as both predictive features for solving downstream tasks (e.g., node classification) and energy function minimizers that inherit transparent, exploitable inductive biases and interpretability. However, scaling GNN architectures constructed in this way remains challenging, in part because the convergence of the forward pass may involve models with considerable depth. To tackle this limitation, we propose a sampling-based energy function and scalable GNN layers that iteratively reduce it, guided by convergence guarantees in certain settings. We also instantiate a full GNN architecture based on these designs, and the model achieves competitive accuracy and scalability when applied to the largest publicly-available node classification benchmark exceeding 1TB in size.", "keywords": [ "Graph Neural Networks", "Energy-based Models", "Scalable Training", "Bi-level Optimization" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "GpdO9r73xT", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Boqing Gong", "gender": "Male", "institution": "Boston University, Boston University", "country": "US", "position": "Assistant Professor" }, { "name": "Cho-Jui Hsieh", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Minhao Cheng", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Assistant Professor" }, { "name": "Ruochen Wang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Tianyi Zhou", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Yuanhao Ban", "gender": "Male", "institution": "UCLA", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 30, "n_ref": 48, "n_ref_all": 60, "n_fig": 4, "n_tab": 10, "L_tab": 3082, "n_element_tab": 234, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 752, "n_element_tab_1": 100, "formula_len_all": 175, "formula_len_all_1": 158, "len_all": 165160, "len_all_1": 69323, "len_abs": 1612, "len_title": 146, "len_sents": 52841, "len_sents_1": 32890, "n_sents": 420, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1488, "title": "The Crystal Ball Hypothesis in diffusion models: Anticipating object positions from initial noise", "abs": "Diffusion models have achieved remarkable success in text-to-image generation tasks, yet the influence of initial noise remains largely unexplored. In this study, we identify specific regions within the initial noise image, termed trigger patches, that play a key role in inducing object generation in the resulting images. Notably, these patches are **universal** and can be generalized across various positions, seeds, and prompts. To be specific, extracting these patches from one noise and injecting them into another noise leads to object generation in targeted areas. To identify the trigger patches even before the image has been generated, just like consulting the crystal ball to foresee fate, we first create a dataset consisting of Gaussian noises labeled with bounding boxes corresponding to the objects appearing in the generated images and **train a detector that identifies these patches from the initial noise.** To explain the formation of these patches, we reveal that they are **outliers** in Gaussian noise, and follow distinct distributions through two-sample tests. These outliers can take effect when injected into different noises and generalize well across different settings. Finally, we find the misalignment between prompts and the trigger patch patterns can result in unsuccessful image generations. To overcome it, we propose a reject-sampling strategy to obtain optimal noise, aiming to improve prompt adherence and positional diversity in image generation.", "keywords": [ "Diffusion models", "Initial noise" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "GpcqrBh6G8", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Jaegul Choo", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Jaemin Seo", "gender": "unknown", "institution": "Chung-Ang University", "country": "KR", "position": "Assistant Professor" }, { "name": "Semin Joung", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Researcher" }, { "name": "Young-chul Ghim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Youngwoo Cho", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 42, "n_ref": 64, "n_ref_all": 85, "n_fig": 6, "n_tab": 1, "L_tab": 1153, "n_element_tab": 67, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 1153, "n_element_tab_1": 67, "formula_len_all": 226, "formula_len_all_1": 226, "len_all": 153254, "len_all_1": 73326, "len_abs": 1087, "len_title": 117, "len_sents": 43347, "len_sents_1": 35881, "n_sents": 309, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1091, "title": "Data-driven plasma equilibrium forecasting in magnetic fusion tokamak", "abs": "The most promising approach to achieving nuclear fusion is through tokamaks, which confine plasma using magnetic fields.\nUnderstanding the current plasma equilibrium state in tokamaks is critical for effective plasma control.\nUnlike previous studies, which reconstruct equilibrium from magnetic field information, our work forecasts future equilibrium based on past equilibrium states.\nSpecifically, we formulate the plasma equilibrium prediction task as a video prediction task, a well-explored problem in the machine learning community.\nThis formulation allows us to capture the spatio-temporal dynamics of plasma states and provides a foundation for multimodal modeling of data streams from tokamak operations.\nOur methodology, incorporating a physics-inspired learning technique for physically reliable predictions, achieved plausible results in forecasting future plasma equilibrium up to 200 ms ahead compared to baselines.\nThis approach holds promise for predicting plasma instabilities and preventing disruptions, marking a significant step towards developing stable fusion reactors.", "keywords": [ "nuclear fusion", "plasma equilibrium", "tokamak", "video prediction" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "GpUv1FvZi1", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Ang Li", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Bowei Tian", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Guoheng Sun", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Shwai He", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Wanghao Ye", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Yongkai Wu", "gender": "Male", "institution": "Clemson University", "country": "US", "position": "Assistant Professor" }, { "name": "Yucong Dai", "gender": "Male", "institution": "Clemson University", "country": "US", "position": "PhD student" }, { "name": "Ziyao Wang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 26, "n_ref_uni": 30, "n_ref": 63, "n_ref_all": 85, "n_fig": 6, "n_tab": 3, "L_tab": 1930, "n_element_tab": 117, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 81, "n_element_tab_1": 14, "formula_len_all": 1812, "formula_len_all_1": 1563, "len_all": 104473, "len_all_1": 56646, "len_abs": 1207, "len_title": 0, "len_sents": 31710, "len_sents_1": 23848, "n_sents": 234, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1328, "title": "Towards counterfactual fairness through auxiliary variables", "abs": "The challenge of balancing fairness and predictive accuracy in machine learning models, especially when sensitive attributes such as race, gender, or age are considered, has motivated substantial research in recent years. Counterfactual fairness ensures that predictions remain consistent across counterfactual variations of sensitive attributes, which is a crucial concept in addressing societal biases. \nHowever, existing counterfactual fairness approaches usually overlook intrinsic information about sensitive features, limiting their ability to achieve fairness while simultaneously maintaining performance. To tackle this challenge, we introduce EXOgenous Causal reasoning (EXOC), a novel causal reasoning framework motivated by exogenous variables. It leverages auxiliary variables to uncover intrinsic properties that give rise to sensitive attributes. Our framework explicitly defines an auxiliary node and a control node that contribute to counterfactual fairness and control the information flow within the model. Our evaluation, conducted on synthetic and real-world datasets, validates EXOC's superiority, showing that it outperforms state-of-the-art approaches in achieving counterfactual fairness without sacrificing accuracy. Our code is available at https://github.com/CASE-Lab-UMD/counterfactual_fairness_2025.", "keywords": [ "Counterfactual", "Fairness", "Auxiliary variables" ], "rating_list": [ 8, 3, 5, 8 ], "soundness_list": [ 4, 1, 2, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 4, 4, 2, 5 ] }, { "paper_id": "GpUO6qYNQG", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Huijia Zhu", "gender": "Female", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Wen Huang", "gender": "Female", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yanmei Gu", "gender": "Female", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yanmin Qian", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Full Professor" }, { "name": "Zhiming Wang", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 41, "n_ref": 91, "n_ref_all": 116, "n_fig": 9, "n_tab": 9, "L_tab": 3902, "n_element_tab": 447, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 3146, "n_element_tab_1": 334, "formula_len_all": 164, "formula_len_all_1": 0, "len_all": 173062, "len_all_1": 61691, "len_abs": 1297, "len_title": 156, "len_sents": 42716, "len_sents_1": 27062, "n_sents": 283, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1302, "title": "SpeechFake: A Large-Scale Multilingual Speech Deepfake Dataset Toward Cutting-Edge Speech Generation Methods", "abs": "As speech generation technology continues to evolve, the risk of misuse through deepfake audio has become a pressing concern, which underscores the critical need for robust detection methods. However, many existing speech deepfake datasets fall short in terms of size, diversity, and linguistic coverage, limiting the ability of models to generalize effectively to unseen deepfakes. To address these limitations, we present SpeechFake, a large-scale dataset specifically designed for speech deepfake detection. With over 3 million deepfakes totaling more than 3,000 hours of audio, SpeechFake was generated using 40 different speech generation tools, including cutting-edge techniques, and spans 46 languages. This paper provides a detailed overview of the dataset\u2019s composition and statistics, emphasizing its scale and diversity. Additionally, we establish baseline results for SpeechFake and explore how factors such as generation methods, language diversity, and speaker variation influence detection performance. We believe SpeechFake will be a valuable resource for advancing speech deepfake detection research, offering opportunities to explore new detection strategies and improve model robustness across diverse and evolving generation techniques. The dataset will be publicly available soon.", "keywords": [ "dataset", "deepfake detection", "anti-spoofing", "speech generation" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "Gp6VU0oJX3", "primary_area": "causal reasoning", "n_author": 4, "author_info_list": [ { "name": "Carl Yang", "gender": "Male", "institution": "Emory University", "country": "US", "position": "Assistant Professor" }, { "name": "Gang Kou", "gender": "Male", "institution": "Xiangjiang Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Huaming Du", "gender": "Male", "institution": "Southwestern University of Finance and Economics", "country": "CN", "position": "Postdoc" }, { "name": "Lei Yuan", "gender": "Female", "institution": "Southwest University of Finance and Economics", "country": "CN", "position": "PhD student" } ], "n_formula": 64, "n_formula_1": 15, "n_ref_uni": 57, "n_ref": 118, "n_ref_all": 190, "n_fig": 11, "n_tab": 13, "L_tab": 3334, "n_element_tab": 107, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 830, "n_element_tab_1": 37, "formula_len_all": 5376, "formula_len_all_1": 968, "len_all": 216369, "len_all_1": 62406, "len_abs": 1952, "len_title": 109, "len_sents": 61987, "len_sents_1": 27041, "n_sents": 502, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1201, "title": "A Causal Theoretical Framework for Open Set Domain Adaptation", "abs": "Open Set Domain Adaptation (OSDA) faces two critical challenges: the emergence\nof unknown classes in the target domain and changes in observed distributions\nacross domains. Although numerous studies have proposed advanced algorithms,\nrecent experimental results demonstrate that the classical Empirical Risk Mini\u0002mization (ERM) approach still delivers state-of-the-art performance. However,\nfew theories can effectively explain this disputed phenomenon. To address the\ntheoretical gap, we focus on constructing a causal theoretical framework for OSDA.\nWe formulate the novel concepts of the Fully Informative Causal Invariance Model\n(FICIM) and the Partially Informative Causal Invariance Model (PICIM). Subse\u0002quently, We derive an OSDA theoretical bound to prove that the ERM performs\nwell when the source domain follows FICIM, while it performs poorly when the\nsource domain follows PICIM. The different results may be attributed to the vary\u0002ing amounts of available information when bounding the target domain\u2019s stable\nexpected risk. Finally, across different datasets, we conduct extensive experiments\non the FICIM and PICIM source domains to validate the effectiveness of our\ntheoretical results.", "keywords": [ "Causal theory", "open set domain adaptation", "domain adaptation", "empirical risk minimization" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 1, 2, 2 ], "presentation_list": [ 2, 1, 2 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "Go21XxlFCp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Bo Li", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Hao Zhang", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "Huanjing Yue", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jingyu Yang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Full Professor" }, { "name": "Jinwei Chen", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Peng-Tao Jiang", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "Qirui Yang", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 28, "n_ref_uni": 50, "n_ref": 113, "n_ref_all": 135, "n_fig": 10, "n_tab": 13, "L_tab": 5672, "n_element_tab": 772, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2734, "n_element_tab_1": 394, "formula_len_all": 1250, "formula_len_all_1": 1239, "len_all": 187348, "len_all_1": 68890, "len_abs": 1619, "len_title": 101, "len_sents": 43531, "len_sents_1": 28248, "n_sents": 325, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1509, "title": "Learning Adaptive Lighting via Channel-Aware Guidance", "abs": "Learning lighting adaption is a key step in obtaining a good visual perception and supporting downstream vision tasks. There are multiple light-related tasks (e.g., image retouching and exposure correction) and previous studies have mainly investigated these tasks individually. However, we observe that the light-related tasks share fundamental properties: i) different color channels have different light properties, and ii) the channel differences reflected in the time and frequency domains are different. Based on the common light property guidance, we propose a Learning Adaptive Lighting Network (LALNet), a unified framework capable of processing different light-related tasks. Specifically, we introduce the color-separated features that emphasize the light difference of different color channels and combine them with the traditional color-mixed features by Light Guided Attention (LGA). The LGA utilizes color-separated features to guide color-mixed features focusing on channel differences and ensuring visual consistency across channels. We introduce dual domain channel modulation to generate color-separated features and a wavelet followed by a vision state space module to generate color-mixed features. Extensive experiments on four representative light-related tasks demonstrate that LALNet significantly outperforms state-of-the-art methods on benchmark tests and requires fewer computational resources. We provide an anonymous online demo at [LALNet](https://xxxxxx2025.github.io/LALNet/).", "keywords": [ "Exposure Correction", "Tone Mapping", "Adaptive Lighting", "Unified framework", "Image Retouching" ], "rating_list": [ 6, 6, 5, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "GnBBSlUb0S", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Le Gia Khang", "gender": "Male", "institution": "University of Information Technology, Vietnam National University", "country": "VN", "position": "Undergrad student" }, { "name": "Ngoc Hoang Luong", "gender": "Male", "institution": "University of Information Technology, Vietnam National University Ho Chi Minh City", "country": "VN", "position": "Lecturer" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 26, "n_ref": 52, "n_ref_all": 63, "n_fig": 3, "n_tab": 15, "L_tab": 14464, "n_element_tab": 2210, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 4076, "n_element_tab_1": 775, "formula_len_all": 714, "formula_len_all_1": 500, "len_all": 224424, "len_all_1": 74275, "len_abs": 6144, "len_title": 132, "len_sents": 53631, "len_sents_1": 30369, "n_sents": 394, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 2318, "title": "Black-Box Adversarial Attack on Dialogue Generation via Multi-Objective Optimization", "abs": "Transformer-based dialogue generation (DG) models are ubiquitous in modern conversational artificial intelligence (AI) platforms.\nThese models, however, are susceptible to adversarial attacks, i.e., prompts that appear textually indiscernible from normal inputs but are maliciously crafted to make the models generate responses incoherent and irrelevant to the conversational context.\nEvaluating the adversarial robustness of DG models is thus crucial to their real-world deployment.\nAdversarial methods typically exploit gradient information and output logits (or probabilities) to effectively modify key input tokens, thereby achieving excellent attack performance.\nNevertheless, such white-box approaches are impractical in real-world scenarios since the models' internal parameters are typically inaccessible.\nWhile black-box methods, which exploit only input prompts and DG models' output responses to craft adversarial attacks, offer a wider applicability, they often suffer from poor performance.\n\nIn a human-machine conversation, good generated responses are expected to be semantically coherent and textually succinct.\nWe thus formulate adversarial attack on DG models as a bi-objective optimization problem, where input prompts are modified in order to 1) minimize the response coherence, and 2) maximize the generation length.\nIn this paper, we empirically demonstrate that optimizing either objective alone results in subpar performance.\nWe then propose a dialogue generation attack framework (DGAttack) that employs multi-objective optimization to consider both objectives simultaneously when perturbing user prompts to craft adversarial inputs.\nLeveraging the exploration capability of multi-objective evolutionary algorithm due to its intrinsic diversity preservation, DGAttack successfully creates effective adversarial prompts in a true black-box manner, i.e., accessing solely DG models' inputs and outputs.\nExperiments across four benchmark datasets and three language models (i.e., BART, DialoGPT, T5) demonstrate the excellent performance of DGAttack compared to existing white-box, gray-box, and black-box approaches.\nEspecially, benchmarks with large language models (i.e., Llama 3.1 and Gemma 2) suggest that DGAttack is the state-of-the-art black-box adversarial attack on dialogue generation.", "keywords": [ "dialogue generation", "adversarial attack", "multi-objective optimization", "black-box attack" ], "rating_list": [ 5, 1, 5, 6, 6 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2, 3 ], "contribution_list": [ 2, 1, 2, 1, 3 ], "confidence_list": [ 3, 5, 3, 3, 2 ] }, { "paper_id": "GmMp8S8M4V", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Huangxin Xu", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Jintu Zheng", "gender": "unknown", "institution": "Advanced Micro Devices", "country": "US", "position": "Intern" }, { "name": "Jue HOU", "gender": "unknown", "institution": "Southern University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Qizhe Liu", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Zenan Wang", "gender": "unknown", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 23, "n_ref": 36, "n_ref_all": 48, "n_fig": 4, "n_tab": 2, "L_tab": 516, "n_element_tab": 76, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2264, "n_element_tab_1": 95, "formula_len_all": 379, "formula_len_all_1": 379, "len_all": 83882, "len_all_1": 54833, "len_abs": 1354, "len_title": 150, "len_sents": 27196, "len_sents_1": 25087, "n_sents": 208, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1366, "title": "HiS4MAE: High-efficiency Segmentation of Subcellular Structure via Self-distillated Masked Autoencoder", "abs": "The accurate identification of subcellular structures is crucial for understanding cellular functions. However, due to the varied morphology of different cells, conventional segmentation methods typically depend on a substantial collection of accurately labeled images of cell structures. The creation of such precise labels is often time-consuming and labor-intensive. To address this issue, we introduce an efficient, self-supervised method for segmenting subcellular structures, named HiS4MAE (High-efficiency Segmentation of Subcellular Structure via Self-distillated Masked Autoencoder). Leveraging an enhanced masked autoencoder (MAE), we train the encoder using the masked image modeling (MIM) framework, followed by clustering the encoded high-dimensional features to achieve pixel-level segmentation of structures. We employ a self-distillation technique to accelerate the model's training process and propose an inference method that is less time-consuming. We also introduce a discrete codebook to assist the self-distillation process, enhancing the model's stability during training. When applied to a publicly available volumetric electron microscopy (VEM) dataset of primary mouse pancreatic islet $\\beta$ cells, HiS4MAE not only surpasses the state-of-the-art technique but also significantly reduces the time required for both training and inference.", "keywords": [ "Subcellular Structure Segmentation", "Masked Image Model", "Microscopy Image" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 4, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "GmE8ovvXaJ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Minsu Cho", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Associate Professor" }, { "name": "Seung Wook Kim", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 42, "n_ref": 86, "n_ref_all": 115, "n_fig": 7, "n_tab": 10, "L_tab": 3375, "n_element_tab": 365, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1636, "n_element_tab_1": 179, "formula_len_all": 1056, "formula_len_all_1": 670, "len_all": 157585, "len_all_1": 57918, "len_abs": 1110, "len_title": 127, "len_sents": 40416, "len_sents_1": 25940, "n_sents": 281, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1163, "title": "Selective State-Space Modeling of Correlation Maps for Semantic Correspondence", "abs": "Establishing semantic correspondences between images is a fundamental yet challenging task in computer vision. Traditional feature-metric methods enhance visual features but may miss complex inter-image relationships, while recent correlation-metric approaches attempt to model these relationships but are hindered by high computational costs due to processing 4D correlation maps. We introduce MambaMatcher, a novel method that overcomes these limitations by efficiently modeling high-dimensional correlations using selective state-space models (SSMs), treating multi-level correlation scores as states. By implementing a similarity-aware selective scan mechanism adapted from Mamba\u2019s linear-complexity algorithm, MambaMatcher refines the 4D correlation tensor effectively without compromising feature map resolution or receptive field. Experiments on standard semantic correspondence benchmarks demonstrate that MambaMatcher achieves state-of-the-art performance without relying on large input images or computationally expensive diffusion-based feature extractors, effectively capturing rich inter-image correlations while maintaining computational efficiency.", "keywords": [ "semantic correspondence", "state-space models", "correlation aggregation", "Mamba", "feature aggregation", "Similarity-aware Selective Scan" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "GlqeLNjH6p", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "Fanfu Wang", "gender": "Male", "institution": "Lanzhou University", "country": "", "position": "Undergrad student" }, { "name": "HaiBin Wen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Pengkun Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" }, { "name": "Shurui Zhang", "gender": "Male", "institution": "Northeastern University", "country": "CN", "position": "Undergrad student" }, { "name": "Xucong Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Yang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zhe Zhao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zhengyang Zhou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "\u738b\u723d", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 47, "n_formula_1": 20, "n_ref_uni": 43, "n_ref": 57, "n_ref_all": 84, "n_fig": 6, "n_tab": 4, "L_tab": 2253, "n_element_tab": 248, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2963, "formula_len_all_1": 1338, "len_all": 150497, "len_all_1": 54500, "len_abs": 1504, "len_title": 141, "len_sents": 43334, "len_sents_1": 25899, "n_sents": 345, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1522, "title": "Exploring Complex Trade-offs in Information Bottleneck through Multi-Objective Optimization", "abs": "Information Bottleneck (IB) theory provides a principled approach to analyze and optimize how neural networks extract and learn latent representations from data, aiming to enhance network performance and generalization. The IB framework has been applied and validated across various domains in deep learning. However, most studies employing IB require tuning of Lagrange multipliers to balance compression and prediction during optimization. Finding the optimal Lagrange multiplier $\\beta$ to achieve the best balance between compression and prediction is challenging, relying heavily on empirical tuning and potentially failing to capture the complex trade-offs present within the IB paradigm. In this paper, we redefine the IB problem as a multi-objective optimization problem with respect to compression and prediction objectives. We employ a gradient-based multi-objective optimization algorithm that adaptively determines the weights for this optimization challenge. Our method is demonstrated to automatically find Pareto-optimal solutions, achieving a balance between compression and prediction, and exploring more complex Pareto frontiers than linear weighting. We compare our approach with the Variational Information Bottleneck and its variants across different datasets. Empirical results confirm that our method achieves a more stable and optimal trade-off compared to Information Bottleneck approaches with manually-tuned multipliers. The code is available in \\url{https://anonymous.4open.science/r/ASDGASDG}.", "keywords": [ "Information Bottleneck \uff0cMulti-objective Optimization\uff0cMutual Information" ], "rating_list": [ 1, 3, 5, 5 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 5, 3, 2 ] }, { "paper_id": "Glm7Kj47nN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Ciara Rowles", "gender": "Female", "institution": "Brighton", "country": "GB", "position": "Researcher" }, { "name": "Simon Donn\u00e9", "gender": "Male", "institution": "Stability AI", "country": "BE", "position": "Researcher" }, { "name": "Sviatoslav Elizarov", "gender": "Male", "institution": "Unity software", "country": "DE", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 73, "n_ref": 109, "n_ref_all": 136, "n_fig": 15, "n_tab": 2, "L_tab": 490, "n_element_tab": 77, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 178932, "len_all_1": 47717, "len_abs": 1106, "len_title": 148, "len_sents": 36635, "len_sents_1": 22869, "n_sents": 269, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1110, "title": "Geometry Image Diffusion: Fast and Data-Efficient Text-to-3D with Image-Based Surface Representation", "abs": "Generating high-quality 3D objects from textual descriptions remains a challenging problem due to high computational costs, the scarcity of 3D data, and the complexity of 3D representations. We introduce Geometry Image Diffusion (GIMDiffusion), a novel Text-to-3D model that utilizes geometry images to efficiently represent 3D shapes using 2D images, thereby avoiding the need for complex 3D-aware architectures. By integrating a Collaborative Control mechanism, we exploit the rich 2D priors of existing Text-to-Image models, such as Stable Diffusion, to achieve strong generalization despite limited 3D training data. This allows us to use only high-quality training data while retaining compatibility with guidance techniques such as IPAdapter. GIMDiffusion enables the generation of 3D assets at speeds comparable to current Text-to-Image models, without being restricted to manifold meshes during either training or inference. We simultaneously generate a UV unwrapping for the objects, consisting of semantically meaningful parts as well as internal structures, enhancing both usability and versatility.", "keywords": [ "Text-to-3D; Geometry Image; Non-watertight mesh; Data efficiency" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "GlgD9o9bl4", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "GUO XIAOYI", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Junkai Liu", "gender": "Male", "institution": "University of Birmingham", "country": "GB", "position": "PhD student" }, { "name": "Le Zhang", "gender": "Male", "institution": "University of Birmingham", "country": "GB", "position": "Assistant Professor" }, { "name": "Prayag Tiwari", "gender": "Male", "institution": "Halmstad University", "country": "SE", "position": "Associate Professor" }, { "name": "Quan Zou", "gender": "Male", "institution": "University of Electronic Science and Technology of China, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Yijie Ding", "gender": "unknown", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 24, "n_formula_1": 23, "n_ref_uni": 38, "n_ref": 70, "n_ref_all": 90, "n_fig": 10, "n_tab": 5, "L_tab": 6584, "n_element_tab": 540, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 6168, "n_element_tab_1": 477, "formula_len_all": 973, "formula_len_all_1": 952, "len_all": 164543, "len_all_1": 72666, "len_abs": 1666, "len_title": 158, "len_sents": 41130, "len_sents_1": 27954, "n_sents": 284, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 109, "L_abs": 1674, "title": "Interpretable and Adaptive Graph Contrastive Learning with Information Sharing for Biomedical Link Prediction", "abs": "The identification of unobserved links in drug-related biomedical networks is essential for various drug discovery applications, which is also beneficial for both disease diagnosis and treatment through exploring the underlying molecular mechanisms. However, existing solutions face significant challenges due to three main limitations: (1) lack of interpretability to provide comprehensive and reliable insights, (2) insufficient robustness and flexibility in cold-start scenarios, and (3) inadequate interaction and sharing of multi-view information. In light of this, we propose DrugXAS, an interpretable and adaptive cross-view contrastive learning framework with information sharing for biomedical link prediction. Specifically, DrugXAS has three distinctive characteristics for addressing these challenges. To solve the first problem, we propose an attention-aware augmentation scheme to provide understandable explanations of intrinsic mechanisms. To deal with the second challenge, we propose an adaptive graph updater and neighborhood sampler, which select proper neighbors according to the feedbacks from the model to improve aggregation ability. To tackle the third issue, an information sharing module with diffusion loss is proposed to incorporate chemical structures into heterogeneous relational semantics and facilitate the contrast process. Empirically, extensive experiments on seven benchmark datasets involving multi-type tasks demonstrate that the proposed DrugXAS outperforms the state-of-the-art methods in terms of precision, robustness, and interpretability. The source code of DrugXAS is available at https://anonymous.4open.science/r/DrugXAS-8EC7.", "keywords": [ "Drug Discovery", "Biomedical Link Prediction", "Interpretability", "Molecular Graph" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "GlPVnuL66V", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Gal Vardi", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "IL", "position": "Assistant Professor" }, { "name": "Guy Smorodinsky", "gender": "Male", "institution": "Ben Gurion University of the Negev", "country": "IL", "position": "PhD student" }, { "name": "Itay Safran", "gender": "Male", "institution": "Purdue University", "country": "", "position": "Postdoc" } ], "n_formula": 84, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 56, "n_ref_all": 92, "n_fig": 5, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 104, "n_element_tab_1": 2, "formula_len_all": 11863, "formula_len_all_1": 572, "len_all": 187100, "len_all_1": 70336, "len_abs": 3719, "len_title": 107, "len_sents": 53875, "len_sents_1": 32343, "n_sents": 504, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 635, "title": "Provable Privacy Attacks on Trained Shallow Neural Networks", "abs": "We study what provable privacy attacks can be shown on trained, 2-layer ReLU neural networks. We explore two types of attacks; data reconstruction attacks, and membership inference attacks. We prove that theoretical results on the implicit bias of 2-layer neural networks can be used to provably reconstruct a set of which at least a constant fraction are training points in a univariate setting, and can also be used to identify with high probability whether a given point was used in the training set in a high dimensional setting. To the best of our knowledge, our work is the first to show provable vulnerabilities in this setting.", "keywords": [ "Deep Learning Theory", "Privacy", "Neural Networks", "Membership Inference Attack" ], "rating_list": [ 5, 5, 8, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 2, 4, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "GlLXMjW7oF", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Gonzalo Benegas", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "G\u00f6kcen Eraslan", "gender": "Male", "institution": "Genentech", "country": "US", "position": "Principal Researcher" }, { "name": "Yun Song", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 0, "n_ref_uni": 34, "n_ref": 68, "n_ref_all": 98, "n_fig": 12, "n_tab": 9, "L_tab": 4032, "n_element_tab": 484, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 1688, "n_element_tab_1": 219, "formula_len_all": 263, "formula_len_all_1": 0, "len_all": 152625, "len_all_1": 56597, "len_abs": 1612, "len_title": 128, "len_sents": 37176, "len_sents_1": 25700, "n_sents": 279, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1626, "title": "Benchmarking DNA Sequence Models for Causal Variant Prediction in Human Genetics", "abs": "Machine learning holds immense promise in biology, particularly for the challenging task of identifying causal variants for Mendelian and complex traits. Two primary approaches have emerged for this task: supervised sequence-to-function models trained on functional genomics experimental data and self-supervised DNA language models that learn evolutionary constraints on sequences. However, the field currently lacks consistently curated datasets with accurate labels, especially for non-coding variants, that are necessary to comprehensively benchmark these models and advance the field. In this work, we present TraitGym, a curated dataset of genetic variants that are either known to be causal or are strong candidates across 113 Mendelian and 83 complex traits, along with carefully constructed control variants. We frame the causal variant prediction task as a binary classification problem and benchmark various models, including functional-genomics-supervised models, self-supervised models, models that combine machine learning predictions with curated annotation features, and ensembles of these. Our results provide insights into the capabilities and limitations of different approaches for predicting the functional consequences of genetic variants. We find that alignment-based models CADD and GPN-MSA compare favorably for Mendelian traits and complex disease traits, while functional-genomics-supervised models Enformer and Borzoi perform better for complex non-disease traits. All curated benchmark data, together with training and benchmarking scripts, will be made publicly available upon publication.", "keywords": [ "genomics", "genetics", "variant effect prediction", "DNA", "language models" ], "rating_list": [ 3, 3, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "GlAeL0I8LX", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Bodo Rosenhahn", "gender": "Male", "institution": "Institut f\u00fcr Informationsverarbeitung", "country": "DE", "position": "Professor" }, { "name": "Ramesh Manuvinakurike", "gender": "Male", "institution": "Intel", "country": "", "position": "Researcher" }, { "name": "Sovan Biswas", "gender": "Male", "institution": "Intel", "country": "", "position": "Research Scientist" }, { "name": "Thomas Norrenbrock", "gender": "Male", "institution": "Universit\u00e4t Hannover", "country": "DE", "position": "PhD student" }, { "name": "Timo Kaiser", "gender": "Male", "institution": "Universit\u00e4t Hannover", "country": "", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 14, "n_ref_uni": 55, "n_ref": 119, "n_ref_all": 176, "n_fig": 13, "n_tab": 19, "L_tab": 6426, "n_element_tab": 307, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1000, "n_element_tab_1": 184, "formula_len_all": 2221, "formula_len_all_1": 915, "len_all": 235100, "len_all_1": 65317, "len_abs": 1137, "len_title": 122, "len_sents": 68175, "len_sents_1": 28703, "n_sents": 513, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1143, "title": "QPM: Discrete Optimization for Globally Interpretable Image Classification", "abs": "Understanding the classifications of deep neural networks, e.g. used in safety-critical situations, is becoming increasingly important. While recent models can locally explain a single decision, to provide a faithful global explanation about an accurate model\u2019s general behavior is a more challenging open task. Towards that goal, we introduce the Quadratic Programming Enhanced Model (QPM), which learns globally interpretable class representations. QPM represents every class with a binary assignment of very few, typically 5, features, that are also assigned to other classes, ensuring easily comparable contrastive class representations. This compact binary assignment is found using discrete optimization based on predefined similarity measures and interpretability constraints. The resulting optimal assignment is used to fine-tune the diverse features, so that each of them becomes the shared general concept between the assigned classes. Extensive evaluations show that QPM delivers unprecedented global interpretability across small and large-scale datasets while setting the state of the art for the accuracy of interpretable models.", "keywords": [ "explainable AI", "Global Interpretability", "Sparse Representations", "Compactness", "Local Interpretability", "Image Classification", "Discrete Optimization" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "Gl2nXRzclw", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Lamine Mili", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "", "position": "Full Professor" }, { "name": "Pooja Algikar", "gender": "Female", "institution": "Virginia Polytechnic Institute and State University", "country": "", "position": "PhD student" } ], "n_formula": 51, "n_formula_1": 35, "n_ref_uni": 19, "n_ref": 33, "n_ref_all": 59, "n_fig": 7, "n_tab": 6, "L_tab": 2102, "n_element_tab": 264, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 934, "n_element_tab_1": 83, "formula_len_all": 2737, "formula_len_all_1": 1780, "len_all": 119272, "len_all_1": 64840, "len_abs": 1208, "len_title": 104, "len_sents": 40161, "len_sents_1": 27060, "n_sents": 304, "n_sents_1": 200, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1213, "title": "Robust Gaussian Process Regression with Huber Likelihood", "abs": "Outliers in both covariates and output responses pose significant challenges for Gaussian Process (GP) regression models. We present a novel GP regression approach that effectively integrates the Huber likelihood into the GP framework\u2014without introducing additional parameters to infer. Specifically, we model the likelihood of observed outputs using the Huber probability distribution: this reduces deviations caused by output outliers. For covariate outliers, we introduce a projection pursuit weights\u2014attenuating their influence on the model. To address the analytically intractable, yet unimodal, posterior distribution, we employ Laplace approximation and Gibbs sampling within a Markov Chain Monte Carlo (MCMC) framework. We simplify Gibbs sampling by expressing the likelihood associated with outlying points as normally distributed through the scale mixture representation of the Laplace distribution. This work is particularly important in the field of transmission spectroscopy\u2014where noisy measurements are often neglected in the estimation of planet-to-star radius ratios. We demonstrate the robustness and effectiveness of our method through extensive experiments on synthetic and real-world datasets.", "keywords": [ "Gaussian Process Regression", "Outlier Handling", "Huber Probability Distribution", "Laplace Approximation", "Gibbs Sampling" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 4, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "GkeTXeujW0", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Andrea Coletta", "gender": "Not Specified", "institution": "Banca d'Italia", "country": "", "position": "Researcher" }, { "name": "Giuseppe Masi", "gender": "Male", "institution": "University of Roma \"La Sapienza\"", "country": "", "position": "PhD student" }, { "name": "Novella Bartolini", "gender": "unknown", "institution": "University of Roma \"La Sapienza\"", "country": "", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 35, "n_ref": 69, "n_ref_all": 99, "n_fig": 4, "n_tab": 8, "L_tab": 3368, "n_element_tab": 374, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 1506, "n_element_tab_1": 168, "formula_len_all": 1373, "formula_len_all_1": 1166, "len_all": 150523, "len_all_1": 66640, "len_abs": 1205, "len_title": 93, "len_sents": 47209, "len_sents_1": 29727, "n_sents": 349, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1214, "title": "CausalDiffusion: Causally Related Time-Series Generation through Diffusion Models", "abs": "Understanding the intrinsic causal structure of time-series data is crucial for effective real-world interventions and decision-making. \nWhile several studies address the Time-Series Causal Discovery (TSCD) problem, the lack of high-quality datasets may limit the progress and evaluation of new methodologies. \nMany available datasets are derived from simplistic simulations, while real-world datasets are often limited in quantity, variety, and lack of ground-truth knowledge describing temporal causal relations. \nIn this paper, we propose CausalDiffusion, the first diffusion model capable of generating multiple causally related time-series alongside a ground-truth causal graph, which abstracts their mutual temporal dependencies.\nCausalDiffusiom employs a causal reconstruction of the output time-series, allowing it to be trained exclusively on time-series data. \nOur experiments demonstrate that CausalDiffusion outperforms state-of-the-art methods in generating realistic time-series, with causal graphs that closely resemble those of real-world phenomena. \nFinally, we provide a benchmark of widely used TSCD algorithms, highlighting the benefits of our synthetic data with respect to existing solutions.", "keywords": [ "time-series generation", "causal discovery", "diffusion model", "benchmark", "dataset", "synthetic data" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "GkWA6NjePN", "primary_area": "reinforcement learning", "n_author": 9, "author_info_list": [ { "name": "Alexander Meulemans", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Blaise Aguera y Arcas", "gender": "Male", "institution": "Princeton University", "country": "", "position": "MS student" }, { "name": "Blake A. Richards", "gender": "Male", "institution": "Google", "country": "CA", "position": "Researcher" }, { "name": "Eric Elmoznino", "gender": "unknown", "institution": "Google", "country": "CA", "position": "Researcher" }, { "name": "Guillaume Lajoie", "gender": "Male", "institution": "Mila, Quebec AI institute", "country": "", "position": "Core Academic Member" }, { "name": "Joao Sacramento", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Johannes Von Oswald", "gender": "Male", "institution": "Research, Google", "country": "CH", "position": "Researcher" }, { "name": "Nino Scherrer", "gender": "Male", "institution": "Google", "country": "CH", "position": "Researcher" }, { "name": "Seijin Kobayashi", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 12, "n_ref_uni": 56, "n_ref": 108, "n_ref_all": 154, "n_fig": 21, "n_tab": 7, "L_tab": 2722, "n_element_tab": 357, "n_fig_1": 13, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3339, "formula_len_all_1": 813, "len_all": 254203, "len_all_1": 77866, "len_abs": 1138, "len_title": 0, "len_sents": 83081, "len_sents_1": 38077, "n_sents": 591, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1144, "title": "Multi-agent cooperation through learning-aware policy gradients", "abs": "Self-interested individuals often fail to cooperate, posing a fundamental challenge for multi-agent learning. How can we achieve cooperation among self-interested, independent learning agents? Promising recent work has shown that in certain tasks cooperation can be established between ``learning-aware\" agents who model the learning dynamics of each other. Here, we present the first unbiased, higher-derivative-free policy gradient algorithm for learning-aware reinforcement learning, which takes into account that other agents are themselves learning through trial and error based on multiple noisy trials. We then leverage efficient sequence models to condition behavior on long observation histories that contain traces of the learning dynamics of other agents. Training long-context policies with our algorithm leads to cooperative behavior and high returns on standard social dilemmas, including a challenging environment where temporally-extended action coordination is required. Finally, we derive from the iterated prisoner's dilemma a novel explanation for how and when cooperation arises among self-interested learning-aware agents.", "keywords": [ "multi-agent learning", "reinforcement learning", "decentralized training", "social dilemmas", "cooperation", "iterated prisoner's dilemma", "melting pot" ], "rating_list": [ 6, 8, 6, 8, 5 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 3, 4, 3, 4, 2 ], "contribution_list": [ 2, 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 2, 4 ] }, { "paper_id": "GkJCgUmIqA", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Sen Na", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Xiaoran Cheng", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" } ], "n_formula": 34, "n_formula_1": 25, "n_ref_uni": 32, "n_ref": 61, "n_ref_all": 96, "n_fig": 14, "n_tab": 5, "L_tab": 1408, "n_element_tab": 77, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2706, "formula_len_all_1": 1764, "len_all": 143237, "len_all_1": 71807, "len_abs": 1769, "len_title": 131, "len_sents": 51488, "len_sents_1": 34572, "n_sents": 398, "n_sents_1": 262, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1778, "title": "Physics-Informed Neural Networks with Trust-Region Sequential Quadratic Programming", "abs": "Physics-Informed Neural Networks (PINNs) represent a significant advancement in Scientific Machine Learning (SciML), which integrate physical domain knowledge into an empirical loss function as soft constraints and apply existing machine learning methods to train the model. However, recent research has noted that PINNs may fail to learn relatively complex Partial Differential Equations (PDEs). This paper addresses the failure modes of PINNs by introducing a novel, hard-constrained deep learning method --- trust-region Sequential Quadratic Programming (trSQP-PINN). In contrast to directly training the penalized soft-constrained loss as in PINNs, our method performs a linear-quadratic approximation of the hard-constrained loss, while leveraging the soft-constrained loss to adaptively adjust the trust-region radius. We only trust our model approximations and make updates within the trust region, and such an updating manner can overcome the ill-conditioning issue of PINNs. We also address the computational bottleneck of second-order SQP methods by employing quasi-Newton updates for second-order information, and importantly, we introduce a simple pretraining step to further enhance training efficiency of our method. We demonstrate the effectiveness of trSQP-PINN through extensive experiments. Compared to existing hard-constrained methods for PINNs, such as penalty methods and augmented Lagrangian methods, trSQP-PINN significantly improves the accuracy of the learned PDE solutions, achieving up to 1-3 orders of magnitude lower errors. Additionally, our pretraining step is generally effective for other hard-constrained methods, and experiments have shown the robustness of our method against both problem-specific parameters and algorithm tuning parameters.", "keywords": [ "Scientific Machine Learning; Physics-Informed Neural Networks; Hard-Constrained Methods; Trust Region; Sequential Quadratic Programming" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 4, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "GjfIZan5jN", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Junshu Sun", "gender": "unknown", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Qi Tian", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Qingming Huang", "gender": "unknown", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Shufan Shen", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Shuhui Wang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Zhaobo Qi", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Lecturer" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 56, "n_ref": 122, "n_ref_all": 162, "n_fig": 13, "n_tab": 7, "L_tab": 1880, "n_element_tab": 221, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 711, "n_element_tab_1": 98, "formula_len_all": 525, "formula_len_all_1": 750, "len_all": 202070, "len_all_1": 63096, "len_abs": 1838, "len_title": 84, "len_sents": 53063, "len_sents_1": 27952, "n_sents": 399, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 83, "L_abs": 1843, "title": "Enhancing Pre-trained Representation Classifiability can Boost its Interpretability", "abs": "The visual representation of a pre-trained model prioritizes the classifiability on downstream tasks. However, widespread applications for pre-trained visual models have proposed new requirements for representation interpretability. It remains unclear whether the pre-trained representations can achieve high interpretability and classifiability simultaneously. To answer this question, we quantify the representation interpretability by leveraging its correlation with the ratio of interpretable semantics within representations. Given the pre-trained representations, only the interpretable semantics can be captured by interpretations, whereas the uninterpretable part leads to information loss. Based on this fact, we propose the Inherent Interpretability Score (IIS) that evaluates the information loss, measures the ratio of interpretable semantics, and quantifies the representation interpretability. In the evaluation of the representation interpretability with different classifiability, we surprisingly discover that the interpretability and classifiability are positively correlated, i.e., representations with higher classifiability provide more interpretable semantics that can be captured in the interpretations. This observation further supports two benefits to the pre-trained representations. First, the classifiability of representations can be further improved by fine-tuning with interpretability maximization. Second, with the classifiability improvement for the representations, we obtain predictions based on their interpretations with less accuracy degradation. The discovered positive correlation and corresponding applications show that practitioners can unify the improvements in interpretability and classifiability for pre-trained vision models. Codes are included in the supplement and will be released on GitHub.", "keywords": [ "Representation interpretability", "vision representations", "image understanding" ], "rating_list": [ 6, 8, 6, 8, 8 ], "soundness_list": [ 2, 3, 2, 4, 4 ], "presentation_list": [ 3, 3, 3, 4, 4 ], "contribution_list": [ 3, 3, 3, 3, 4 ], "confidence_list": [ 4, 4, 4, 4, 5 ] }, { "paper_id": "GjSstLcxAs", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Xiao Peng", "gender": "Male", "institution": "Changan Automobile", "country": "", "position": "Engineer" }, { "name": "Xufan Geng", "gender": "Male", "institution": "Southwest University", "country": "CN", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 20, "n_ref": 37, "n_ref_all": 49, "n_fig": 6, "n_tab": 3, "L_tab": 708, "n_element_tab": 49, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 885, "n_element_tab_1": 27, "formula_len_all": 853, "formula_len_all_1": 828, "len_all": 86423, "len_all_1": 40035, "len_abs": 1381, "len_title": 126, "len_sents": 19192, "len_sents_1": 17319, "n_sents": 169, "n_sents_1": 147, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1407, "title": "Self-controller: Controlling LLMs with Multi-round Step-by-step Self-awareness", "abs": "The applications of large language models (LLMs) have been widely spread across all domains. \nHowever, the basic abilities such as the controllability of LLMs are still limited.\nTo address this, we propose \"$\\textbf{Self-controller}$\", a novel agentic framework bringing self-awareness into LLMs\u2019 reasoning logic.\nThe core idea of this work is to maintain states based on the LLM's response, letting the LLM become self-aware of current status and think step by step in a multi-round chain-of-thought paradigm. \nOur experiment on the state of textual length has shown the controllability and effectiveness of the Self-controller. We further implement a binary search algorithm to accelerate the generation process based on the linearity and monotonicity of the textual length state. Another advantage of the Self-controller comes with DeepSeek's Context Caching technology, which significantly saves computational token consumption when a cluster of conversations shares the same prefix of context. Theoretically, we prove that in this scenario the extra time complexity is $O(c \\log n)$. Results of the back-of-the-envelope estimation suggest that the token consumption of our method is no more than twice as much as that of the trivial single-round generation. Furthermore, our ablation study on word constraints demonstrates the Self-controller's consistent controllability across all foundation models.", "keywords": [ "large language models", "self-awareness", "agent", "controllability" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "GjM61KRiTG", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Adel Bibi", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Senior Researcher" }, { "name": "Mohamed Elhoseiny", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Wenxuan Zhang", "gender": "Female", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" } ], "n_formula": 66, "n_formula_1": 23, "n_ref_uni": 48, "n_ref": 133, "n_ref_all": 216, "n_fig": 5, "n_tab": 11, "L_tab": 4467, "n_element_tab": 596, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1901, "n_element_tab_1": 155, "formula_len_all": 6333, "formula_len_all_1": 1490, "len_all": 223072, "len_all_1": 76084, "len_abs": 1214, "len_title": 85, "len_sents": 56169, "len_sents_1": 31058, "n_sents": 494, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 85, "L_abs": 1267, "title": "Bi-Factorial Preference Optimization: Balancing Safety-Helpfulness in Language Models", "abs": "Fine-tuning large language models (LLMs) on human preferences, typically through reinforcement learning from human feedback (RLHF), has proven successful in enhancing their capabilities. However, ensuring the safety of LLMs during fine-tuning remains a critical concern, and mitigating the potential conflicts in safety and helpfulness is costly in RLHF. To address this issue, we propose a supervised learning framework called Bi-Factorial Preference Optimization (BFPO), which re-parameterizes a joint RLHF objective of both safety and helpfulness into a single supervised learning objective. In the supervised optimization, a labeling function is used to capture global preferences ranking to balance both safety and helpfulness. To evaluate BFPO, we develop a benchmark including comprehensive discriminative and generative tasks for helpfulness and harmlessness. The results indicate that our method significantly outperforms existing approaches in both safety and helpfulness. Moreover, BFPO eliminates the need for human prompting and annotation in LLM fine-tuning while achieving the same level of safety as methods that heavily rely on human labor, with less than 10\\% of the computational resources. The training recipes and models will be released.", "keywords": [ "Large Language Models", "RLHF", "Safety" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "Gj5JTAwdoy", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Ge Zhu", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Researcher" }, { "name": "Jonah Casebeer", "gender": "unknown", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Julian McAuley", "gender": "Male", "institution": "University of California, San Diego, University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Nicholas J. Bryan", "gender": "Male", "institution": "Adobe Systems", "country": "", "position": "Researcher" }, { "name": "Taylor Berg-Kirkpatrick", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "Associate Professor" }, { "name": "Zachary Novack", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 40, "n_ref": 163, "n_ref_all": 195, "n_fig": 14, "n_tab": 7, "L_tab": 3594, "n_element_tab": 430, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 2760, "n_element_tab_1": 232, "formula_len_all": 1493, "formula_len_all_1": 829, "len_all": 194178, "len_all_1": 70856, "len_abs": 1104, "len_title": 69, "len_sents": 56606, "len_sents_1": 31484, "n_sents": 370, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 69, "L_abs": 1109, "title": "Presto! Distilling Steps and Layers for Accelerating Music Generation", "abs": "Despite advances in diffusion-based text-to-music (TTM) methods, efficient, high-quality generation remains a challenge. We introduce Presto!, an approach to inference acceleration for score-based diffusion transformers via reducing both sampling steps and cost per step. To reduce steps, we develop a new score-based distribution matching distillation (DMD) method for the EDM-family of diffusion models, the first GAN-based distillation method for TTM. To reduce the cost per step, we develop a simple, but powerful improvement to a recent layer distillation method that improves learning via better preserving hidden state variance. Finally, we combine our step and layer distillation methods together for a dual-faceted approach. We evaluate our step and layer distillation methods independently and show each yield best-in-class performance. Our combined distillation method can generate high-quality outputs with improved diversity, accelerating our base model by 10-18x (230/435ms latency for 32 second mono/stereo 44.1kHz, 15x faster than the comparable SOTA model) \u2014 the fastest TTM to our knowledge.", "keywords": [ "music generation", "diffusion distillation", "diffusion", "diffusion acceleration", "text-to-music generation", "layer dropping" ], "rating_list": [ 8, 8, 5, 8 ], "soundness_list": [ 4, 4, 3, 4 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 3, 2, 4 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "GiHLTtfbB5", "primary_area": "learning theory", "n_author": 8, "author_info_list": [ { "name": "Adam Wierman", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Professor" }, { "name": "Cheng Chen", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Jinhang Zuo", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "John Lui", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Mohammad Hajiesmaili", "gender": "Male", "institution": "University of Massachusetts Amherst", "country": "US", "position": "Associate Professor" }, { "name": "Shuai Li", "gender": "Female", "institution": "John Hopcroft Center, Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xuchuang Wang", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Postdoc" }, { "name": "Zhiyao Zhang", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" } ], "n_formula": 99, "n_formula_1": 22, "n_ref_uni": 22, "n_ref": 69, "n_ref_all": 114, "n_fig": 9, "n_tab": 1, "L_tab": 250, "n_element_tab": 18, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 250, "n_element_tab_1": 18, "formula_len_all": 8914, "formula_len_all_1": 1916, "len_all": 170948, "len_all_1": 71757, "len_abs": 1084, "len_title": 102, "len_sents": 60328, "len_sents_1": 33908, "n_sents": 584, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1090, "title": "Adversarial Attacks on Cooperative Multi-agent Bandits", "abs": "Cooperative multi-agent multi-armed bandits (CMA2B) consider the collaborative efforts of multiple agents in a shared multi-armed bandit game. We study latent vulnerabilities exposed by this collaboration and consider adversarial attacks on a few agents with the goal of influencing the decisions of the rest. More specifically, we study adversarial attacks on CMA2B in both homogeneous settings, where agents operate with the same arm set, and heterogeneous settings, where agents may have distinct arm sets. In the homogeneous setting, we propose attack strategies that, by targeting just one agent, convince all agents to select a particular target arm $T-o(T)$ times while incurring $o(T)$ attack costs in $T$ rounds. In the heterogeneous setting, we prove that a target arm attack requires linear attack costs and propose attack strategies that can force a maximum number of agents to suffer linear regrets while incurring sublinear costs and only manipulating the observations of a few target agents. Numerical experiments validate the effectiveness of our proposed attack strategies.", "keywords": [ "multi-agent bandits", "adversarial attacks" ], "rating_list": [ 3, 5, 5, 3, 6 ], "soundness_list": [ 1, 3, 3, 2, 3 ], "presentation_list": [ 1, 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 1, 2 ], "confidence_list": [ 4, 3, 3, 5, 4 ] }, { "paper_id": "Gi5aWK8tOx", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Anas Barakat", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Research Fellow" }, { "name": "Peihong Yu", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Pratap Tokekar", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Associate Professor" }, { "name": "Souradip Chakraborty", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Amrit Bedi", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" } ], "n_formula": 90, "n_formula_1": 27, "n_ref_uni": 31, "n_ref": 88, "n_ref_all": 123, "n_fig": 5, "n_tab": 3, "L_tab": 183, "n_element_tab": 28, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7947, "formula_len_all_1": 1752, "len_all": 187377, "len_all_1": 68697, "len_abs": 1418, "len_title": 173, "len_sents": 66318, "len_sents_1": 30825, "n_sents": 575, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 127, "L_abs": 1424, "title": "On the Sample Complexity of a Policy Gradient Algorithm with Occupancy Approximation for General Utility Reinforcement Learning", "abs": "Reinforcement learning with general utilities has recently gained attention thanks to its ability to unify several problems, including imitation learning, pure exploration, and safe RL. However, prior work for solving this general problem in a unified way has only focused on the tabular setting. This is restrictive when considering larger state-action spaces because of the need to estimate occupancy measures during policy optimization. In this work, we address this issue and propose to approximate occupancy measures within a function approximation class using maximum likelihood estimation (MLE). We propose a simple policy gradient algorithm (PG-OMA) where an actor updates the policy parameters to maximize the general utility objective whereas a critic approximates the occupancy measure using MLE. We provide a statistical complexity analysis of PG-OMA showing that our occupancy measure estimation error only scales with the dimension of our function approximation class rather than the size of the state action space. Under suitable assumptions, we establish first order stationarity and global optimality performance bounds for the proposed PG-OMA algorithm for nonconcave and concave general utilities respectively. We complement our methodological and theoretical findings with promising empirical results showing the scalability potential of our approach compared to existing tabular count-based approaches.", "keywords": [ "reinforcement learning with general utilities", "policy gradient algorithm" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "Gi3SwL98nL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Abhinaba Roy", "gender": "unknown", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Researcher" }, { "name": "Dorien Herremans", "gender": "Female", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Associate Professor" }, { "name": "Liu Renhang", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 21, "n_ref": 34, "n_ref_all": 45, "n_fig": 2, "n_tab": 10, "L_tab": 1209, "n_element_tab": 148, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 764, "n_element_tab_1": 96, "formula_len_all": 600, "formula_len_all_1": 550, "len_all": 101382, "len_all_1": 63157, "len_abs": 867, "len_title": 98, "len_sents": 38340, "len_sents_1": 29561, "n_sents": 275, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 873, "title": "Leveraging LLM Embeddings for Cross Dataset Label Alignment and Zero Shot Music Emotion Prediction", "abs": "In this work, we present a novel method for music emotion recognition that leverages Large Language Model (LLM) embeddings for label alignment across multiple datasets and zero-shot prediction on novel categories. First, we compute LLM embeddings for emotion labels and apply non-parametric clustering to group similar labels, across multiple datasets containing disjoint labels. We use these cluster centers to map music features (MERT) to the LLM embedding space. To further enhance the model, we introduce an alignment regularization that enables dissociation of MERT embeddings from different clusters. This further enhances the model's ability to better adaptation to unseen datasets. We demonstrate the effectiveness of our approach by performing zero-shot inference on a new dataset, showcasing its ability to generalize to unseen labels without additional training.", "keywords": [ "Music Emotion Prediction", "Zero Shot Learning", "Label Alignment" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "GhexuBLxbO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Fei Huang", "gender": "Male", "institution": "Alibaba Group US", "country": "US", "position": "Senior Research Director" }, { "name": "Haiyang Yu", "gender": "unknown", "institution": "Xiamen University", "country": "", "position": "MS student" }, { "name": "Hongyu Lin", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Le Sun", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Qiaoyu Tang", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "", "position": "PhD student" }, { "name": "Xianpei Han", "gender": "Male", "institution": "Institute of Software, CAS", "country": "CN", "position": "Professor" }, { "name": "Xuanang Chen", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Postdoc" }, { "name": "Yaojie Lu", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Yongbin Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zhuoqun Li", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 34, "n_ref": 67, "n_ref_all": 96, "n_fig": 7, "n_tab": 6, "L_tab": 3841, "n_element_tab": 577, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 4145, "n_element_tab_1": 664, "formula_len_all": 489, "formula_len_all_1": 492, "len_all": 167989, "len_all_1": 80562, "len_abs": 1491, "len_title": 159, "len_sents": 40016, "len_sents_1": 33479, "n_sents": 247, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 111, "L_abs": 1117, "title": "StructRAG: Boosting Knowledge Intensive Reasoning of LLMs via Inference-time Hybrid Information Structurization", "abs": "Retrieval-augmented generation (RAG) is a key means to effectively enhance large language models (LLMs) in many knowledge-based tasks. \nHowever, existing RAG methods struggle with knowledge-intensive reasoning tasks, because useful information required to these tasks are badly scattered. \nThis characteristic makes it difficult for existing RAG methods to accurately identify key information and perform global reasoning with such noisy augmentation.\nIn this paper, motivated by the cognitive theories that humans convert raw information into various structured knowledge when tackling knowledge-intensive reasoning, we proposes a new framework, StructRAG, which can identify the optimal structure type for the task at hand, reconstruct original documents into this structured format, and infer answers based on the resulting structure. \nExtensive experiments across various knowledge-intensive tasks show that StructRAG achieves state-of-the-art performance, particularly excelling in challenging scenarios, demonstrating its potential as an effective solution for enhancing LLMs in complex real-world applications.", "keywords": [ "RAG", "Structured Knowledge" ], "rating_list": [ 8, 5, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "GhT6NjiLeA", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Annette Ten Teije", "gender": "Female", "institution": "Vrije Universiteit Amsterdam", "country": "", "position": "Associate Professor" }, { "name": "Ilaria Tiddi", "gender": "Female", "institution": "Vrije Universiteit Amsterdam", "country": "", "position": "Assistant Professor" }, { "name": "Majid Mohammadi", "gender": "Male", "institution": "Vrije Universiteit Amsterdam", "country": "", "position": "Postdoc" } ], "n_formula": 31, "n_formula_1": 21, "n_ref_uni": 25, "n_ref": 51, "n_ref_all": 76, "n_fig": 5, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2480, "formula_len_all_1": 1665, "len_all": 93122, "len_all_1": 58900, "len_abs": 1323, "len_title": 83, "len_sents": 32394, "len_sents_1": 26691, "n_sents": 229, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1345, "title": "Exact Shapley Value for Local and Global Explanation of Additive Gaussian Processes", "abs": "Additive Gaussian Processes (AGPs) have emerged as an extension of Gaussian Processes (GPs), offering a more interpretable and flexible approach by decomposing the target function into sums of multiple GPs, each influenced by different subsets of features. Despite their enhanced, expressive structure, AGPs struggle to provide local explanations and offer only global feature importance with notable shortcomings. To bridge this gap, this paper introduces an interpretative framework for AGPs that utilizes Shapley values to provide both local and global explanations of feature importance. For local explanation, we use the relationship between the AGP and the Shapley value and guarantee the additivity of the explanation. We then develop a dynamic programming algorithm for efficient computation of \\textit{exact} Shapley values, whose complexity scales polynomially rather than exponentially with the number of features. In addition, we use a variance-based sensitivity approach for the global explanation and develop an efficient dynamic programming-based algorithm to compute the \\textit{exact} Shapley value as the global feature importance. We present the effectiveness of the proposed methods on several real experiments and discuss their potential in interpretable machine learning, feature selection, and global sensitivity analysis.", "keywords": [ "gaussian processes", "shapley value", "interpretable", "functional decomposition" ], "rating_list": [ 6, 1, 3, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 1, 2, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 3, 3, 4 ] }, { "paper_id": "GhM63V7z6v", "primary_area": "learning on time series and dynamical systems", "n_author": 7, "author_info_list": [ { "name": "Felix Ott", "gender": "Male", "institution": "Fraunhofer IIS", "country": "DE", "position": "Scientific Researcher" }, { "name": "Lihua Xie", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Min Wu", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" }, { "name": "Peiliang Gong", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Xiaoli Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Yucheng Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Zhenghua Chen", "gender": "Male", "institution": "I2R, A*STAR", "country": "SG", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 8, "n_ref_uni": 37, "n_ref": 72, "n_ref_all": 94, "n_fig": 7, "n_tab": 10, "L_tab": 7825, "n_element_tab": 414, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 5659, "n_element_tab_1": 160, "formula_len_all": 1093, "formula_len_all_1": 499, "len_all": 196849, "len_all_1": 77210, "len_abs": 1480, "len_title": 131, "len_sents": 57763, "len_sents_1": 33786, "n_sents": 636, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1421, "title": "Temporal Source Recovery for Time-Series Source-Free Unsupervised Domain Adaptation", "abs": "Source-Free Unsupervised Domain Adaptation (SFUDA) has gained popularity for its ability to adapt pretrained models to target domains without accessing source domains, ensuring source data privacy. While SFUDA is well-developed in visual tasks, its application to Time-Series SFUDA (TS-SFUDA) remains limited due to the challenge of transferring crucial temporal dependencies across domains. Although a few researchers begin to explore this area, they rely on specific source domain designs, which are impractical as source data owners cannot be expected to follow particular pretraining protocols. To solve this, we propose Temporal Source Recovery (TemSR), a framework that transfers temporal dependencies for effective TS-SFUDA without requiring source-specific designs. TemSR features a recovery process that leverages masking, recovery, and optimization to generate a source-like distribution with recovered source temporal dependencies. To ensure effective recovery, we further design segment-based regularization to restore local dependencies and anchor-based recovery diversity maximization to enhance the diversity of the source-like distribution. The source-like distribution is then adapted to the target domain using traditional UDA techniques. Extensive experiments across multiple TS tasks demonstrate the effectiveness of TemSR, even surpassing existing TS-SFUDA method that requires source domain designs.", "keywords": [ "Time-Series Data", "Source-Free Unsupervised Domain Adaptation" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "Gh1XW314zF", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Inyoung Choi", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Jiayi Xin", "gender": "Female", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Jie Peng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Qi Long", "gender": "unknown", "institution": "University of Pennsylvania", "country": "US", "position": "Professor" }, { "name": "Sukwon Yun", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 30, "n_ref": 52, "n_ref_all": 60, "n_fig": 4, "n_tab": 1, "L_tab": 332, "n_element_tab": 61, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 332, "n_element_tab_1": 61, "formula_len_all": 143, "formula_len_all_1": 143, "len_all": 116175, "len_all_1": 63929, "len_abs": 1357, "len_title": 138, "len_sents": 35138, "len_sents_1": 32934, "n_sents": 226, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1362, "title": "Multimodal Graph-LLM: Leveraging Graph-Enhanced LLMs for Multimodal Healthcare Predictions", "abs": "Multimodal healthcare research is crucial for improving clinical decision-making by integrating diverse data types, such as clinical notes, lab results, and imaging. Large Language Models (LLMs) are widely recognized for their exceptional text-based reasoning capabilities, making them effective in processing complex clinical narratives. However, they struggle to incorporate multimodal data, limiting their broader applicability in healthcare analysis. In this work, we propose MG-LLM (Multimodal Graph-LLM), a novel framework that leverages the strengths of LLMs while enhancing them with multimodal alignment and data integration through Graph Neural Networks (GNNs). GNNs propagate information across similar patients, model temporal relationships between visits, and align information from different modalities, creating enriched multimodal context vectors. These context vectors are then injected into the intermediate layers of the LLM, allowing it to harness both textual reasoning and multimodal data for more accurate predictions. We evaluate MG-LLM on the MIMIC-IV and MIMIC-CXR datasets, demonstrating significant improvements in clinical prediction tasks compared to baseline models. Our results showcase the potential of combining the text reasoning power of LLMs with GNN-driven multimodal alignment for robust, comprehensive healthcare analysis.", "keywords": [ "EHR", "multimodal", "LLM", "graphs", "healthcare" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "GfuJR76Sfo", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Frank Rudzicz", "gender": "Male", "institution": "Dalhousie University", "country": "CA", "position": "Associate Professor" }, { "name": "Nicholas Vinden", "gender": "unknown", "institution": "University of Guelph", "country": "CA", "position": "MS student" }, { "name": "Raeid Saqur", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Zining Zhu", "gender": "unknown", "institution": "Stevens Institute of Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 29, "n_ref": 55, "n_ref_all": 69, "n_fig": 1, "n_tab": 10, "L_tab": 1639, "n_element_tab": 169, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1796, "n_element_tab_1": 66, "formula_len_all": 534, "formula_len_all_1": 448, "len_all": 140220, "len_all_1": 56618, "len_abs": 1359, "len_title": 130, "len_sents": 51150, "len_sents_1": 28210, "n_sents": 336, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1364, "title": "ContraSim: Contrastive Similarity Space Learning for Financial Market Predictions", "abs": "We introduce the Contrastive Similarity Space Embedding Algorithm (ContraSim), a novel framework for uncovering the global semantic relationships between daily financial headlines and market movements. ContraSim operates in two key stages: \n(i) Weighted Headline Augmentation, which generates augmented financial headlines along with a semantic fine-grained similarity score, and (ii) Weighted Self-Supervised Contrastive Learning (WSSCL), an extended version of classical self-supervised contrastive learning that uses the similarity metric to create a refined weighted embedding space. This embedding space clusters semantically similar headlines together, facilitating deeper market insights. Empirical results demonstrate that integrating ContraSim features into financial forecasting tasks improves classification accuracy from WSJ headlines by 7%. Moreover, leveraging an information density analysis, we find that the similarity spaces constructed by ContraSim intrinsically cluster days with homogeneous market movement directions, indicating that ContraSim captures market dynamics independent of ground truth labels. Additionally, ContraSim enables the identification of historical news days that closely resemble the headlines of the current day, providing analysts with actionable insights to predict market trends by referencing analogous past events.", "keywords": [ "Learning Representations", "Large Language Models", "Financial Forecasting" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 3, 3, 1, 2 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "GfXMTAJaxZ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Guy N. Rothblum", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Apple MLR" }, { "name": "Tal Herman", "gender": "Male", "institution": "Weizmann Institute of Science", "country": "IL", "position": "PhD student" } ], "n_formula": 45, "n_formula_1": 4, "n_ref_uni": 34, "n_ref": 82, "n_ref_all": 90, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3332, "formula_len_all_1": 189, "len_all": 217133, "len_all_1": 77241, "len_abs": 1710, "len_title": 158, "len_sents": 80925, "len_sents_1": 38525, "n_sents": 613, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 110, "L_abs": 1636, "title": "How to Verify Any (Reasonable) Distribution Property: Computationally Sound Argument Systems for Distributions", "abs": "As statistical analyses become more central to science, industry and society, there is a growing need to ensure correctness of their results. Approximate correctness can be verified by replicating the entire analysis, but can we verify without replication? We focus on distribution testing problems: verifying that an unknown distribution is close to having a claimed property. Our main contribution is an interactive protocol between a verifier and an untrusted prover, which can be used to verify any distribution property that can be decided in polynomial time given a full and explicit description of the distribution. If the distribution is at statistical distance $\\varepsilon$ from having the property, then the verifier rejects with high probability. This soundness property holds against any polynomial-time strategy that a cheating prover might follow, assuming the existence of collision-resistant hash functions (a standard assumption in cryptography). For distributions over a domain of size $N$, the protocol consists of $4$ messages and the communication complexity and verifier runtime are roughly $\\widetilde{O}\\left(\\sqrt{N} / \\varepsilon^2 \\right)$. The verifier's sample complexity is $\\widetilde{O}\\left(\\sqrt{N} / \\varepsilon^2 \\right)$, and this is optimal up to $\\text{polylog}(N)$ factors (for any protocol, regardless of its communication complexity). Even for simple properties, approximately deciding whether an unknown distribution has the property can require quasi-linear sample complexity and running time. For any such property, our protocol provides a quadratic speedup over replicating the analysis.", "keywords": [ "property testing", "distribution testing", "interactive proofs", "data science", "verification" ], "rating_list": [ 6, 8, 8, 6, 8 ], "soundness_list": [ 3, 4, 3, 3, 4 ], "presentation_list": [ 3, 4, 2, 3, 4 ], "contribution_list": [ 2, 3, 3, 2, 4 ], "confidence_list": [ 4, 3, 3, 4, 3 ] }, { "paper_id": "Gf6VDFA6AU", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Luc Van Gool", "gender": "unknown", "institution": "INSAIT - Sofia Un.", "country": "BG", "position": "Full Professor" }, { "name": "Marc Proesmans", "gender": "Male", "institution": "Department of Electrical Engineering, KU Leuven, Belgium, KU Leuven", "country": "BE", "position": "Associate Professor" }, { "name": "Pierre-Fran\u00e7ois De Plaen", "gender": "Male", "institution": "Department of Electrical Engineering, KU Leuven, Belgium, KU Leuven", "country": "BE", "position": "PhD student" }, { "name": "Tinne Tuytelaars", "gender": "Female", "institution": "KU Leuven", "country": "BE", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 15, "n_ref_uni": 52, "n_ref": 85, "n_ref_all": 101, "n_fig": 5, "n_tab": 5, "L_tab": 3347, "n_element_tab": 691, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 795, "n_element_tab_1": 96, "formula_len_all": 865, "formula_len_all_1": 917, "len_all": 157425, "len_all_1": 67685, "len_abs": 1088, "len_title": 124, "len_sents": 44750, "len_sents_1": 32062, "n_sents": 356, "n_sents_1": 245, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1093, "title": "Minimizing Dependence between Embedding Dimensions with Adversarial Networks", "abs": "Learning representations with minimally dependent embedding dimensions can have many potential benefits such as improved generalization and interpretability. This work provides a differentiable and scalable algorithm for dependence minimization, moving beyond existing linear pairwise decorrelation methods. Our algorithm involves an adversarial game where small networks identify dimension relationships, while the main model exploits this information to reduce dependencies. We empirically verify that the algorithm converges. We then explore dependence reduction as a proxy for maximizing information content. We showcase the algorithm's effectiveness on the Clevr-4 dataset, both with and without supervision, and achieve promising results on the ImageNet dataset. Finally, we propose an algorithm modification that gives more control over the level of dependency, sparking a discussion on optimal redundancy levels for specific applications. Although the algorithm performs well on synthetic data, further research is needed to optimize it for tasks such as out-of-distribution detection.", "keywords": [ "representation learning", "adversarial networks", "independence", "information maximization", "generalization" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "Gf4d4ck131", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Martin Vechev", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Yani Zhang", "gender": "Female", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Yuhao Mao", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 19, "n_ref_uni": 22, "n_ref": 52, "n_ref_all": 60, "n_fig": 6, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2130, "formula_len_all_1": 1208, "len_all": 109046, "len_all_1": 76117, "len_abs": 1114, "len_title": 124, "len_sents": 36442, "len_sents_1": 35097, "n_sents": 276, "n_sents_1": 273, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1193, "title": "Multi-Neuron Unleashes Expressivity of ReLU Networks Under Convex Relaxation", "abs": "Modern neural network certification methods heavily rely on convex relaxations to compute sound bounds. However, the true expressive power of convex relaxations is currently not well understood. Recent work has started investigating this direction, showing there does not exist a ReLU network that can express even the simple ``$\\max$'' function in $\\mathbb{R}^2$ such that the network outputs can be bounded exactly by single-neuron relaxations. This raises the following fundamental question: is there a convex relaxation (beyond single-neuron) that can provide exact bounds for ReLU networks expressing general continuous piecewise linear functions in $\\mathbb{R}^n$? In this work, we investigate this question and prove, perhaps surprisingly, that layer-wise multi-neuron relaxations can compute exact bounds for general ReLU networks. Based on this novel result, we show that the expressivity of ReLU networks is no longer limited under multi-neuron relaxations. To the best of our knowledge, this is the first positive result on the completeness of convex relaxations and the expressivity of ReLU networks under convex relaxation, shedding light on the practice of certified robustness.", "keywords": [ "Neural Network Certification", "Model Expressivity", "Convex Relaxation" ], "rating_list": [ 1, 5, 5, 5 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "Gf1uBeuUJW", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "En Yu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Haoran Wei", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Jianjian Sun", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Jingyu Wang", "gender": "Male", "institution": "Beijing University of Post and Telecommunication", "country": "CN", "position": "Full Professor" }, { "name": "Kangheng Lin", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Liang Zhao", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Wenbing Tao", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xiangyu Zhang", "gender": "Male", "institution": "MEGVII Technology", "country": "CN", "position": "Principal Researcher" }, { "name": "Yana Wei", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Zheng Ge", "gender": "Male", "institution": "Megvii Technology Inc.", "country": "CN", "position": "Researcher" }, { "name": "Zining Zhu", "gender": "unknown", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 49, "n_ref": 125, "n_ref_all": 155, "n_fig": 9, "n_tab": 8, "L_tab": 3124, "n_element_tab": 374, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 811, "n_element_tab_1": 117, "formula_len_all": 687, "formula_len_all_1": 670, "len_all": 213618, "len_all_1": 66211, "len_abs": 1296, "len_title": 103, "len_sents": 51052, "len_sents_1": 30268, "n_sents": 391, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1093, "title": "Unhackable Temporal Reward for Scalable Video MLLMs", "abs": "In the pursuit of superior video-processing MLLMs, we have encountered a perplexing paradox: the \u201canti-scaling law\u201d, where more data and larger models lead to worse performance. This study unmasks the culprit: \u201ctemporal hacking\u201d, a phenomenon where models shortcut by fixating on select frames, missing the full video narrative. In this work, we systematically establish a comprehensive theory of temporal hacking, defining it from a reinforcement learning perspective, introducing the Temporal Perplexity (TPL) score to assess this misalignment, and proposing the Unhackable Temporal Rewarding (UTR) framework to mitigate the temporal hacking. Both theoretically and empirically, TPL proves to be a reliable indicator of temporal modeling quality, correlating strongly with frame activation patterns. Extensive experiments reveal that UTR not only counters temporal hacking but significantly elevates video comprehension capabilities. This work not only advances video-AI systems but also illuminates the critical importance of aligning proxy rewards with true objectives in MLLM development.", "keywords": [ "Video MLLMs; Temporal hacking; Temporal Perplexity" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 5, 2 ] }, { "paper_id": "GeyZGQ8SSY", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Samson Zhou", "gender": "unknown", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" } ], "n_formula": 32, "n_formula_1": 14, "n_ref_uni": 22, "n_ref": 50, "n_ref_all": 61, "n_fig": 3, "n_tab": 2, "L_tab": 523, "n_element_tab": 20, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3080, "formula_len_all_1": 1212, "len_all": 164822, "len_all_1": 68651, "len_abs": 2266, "len_title": 102, "len_sents": 55607, "len_sents_1": 34093, "n_sents": 458, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1299, "title": "Learning-Augmented Learning of Gaussian Mixture Models", "abs": "Gaussian mixture models (GMMs) is one of the most fundamental methods to identify and extract latent structure in complex datasets. Unfortunately, well-known hardness results require that any algorithm for learning a mixture of $k$ multivariate Gaussian distributions in $d$-dimensional space requires both runtime and sample complexity exponential in $d$, even if the Gaussians are reasonably separated. To overcome this barrier, we consider settings where algorithms are augmented with possibly erroneous ``advice'' to help learn the underlying GMMs. In particular, we consider a natural predictor that can be easily trained through machine learning models. Specifically, our predictor outputs a list of $\\beta$ possible labels for each sample from the mixture such that, with probability at least $1-\\alpha$, one of the labels in the list is the true label, for a fixed constant $\\alpha$. We show that to estimate the mixture up to total variation distance $\\tilde{\\mathcal{O}}(\\varepsilon)$, we can use $k\\cdot\\text{poly}\\left(d,\\log k,\\frac{1}{\\varepsilon}\\right)$ samples from the GMM, provided that $\\beta$ is upper bounded by any fixed constant. Moreover, our algorithm uses polynomial time, thus breaking known computational limitations of algorithms that do not have access to such advice.", "keywords": [ "learning-augmented algorithms", "Gaussian mixture models" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 2 ] }, { "paper_id": "GeUK3zGreN", "primary_area": "optimization", "n_author": 8, "author_info_list": [ { "name": "Bojia Zi", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jiaquan Ye", "gender": "Male", "institution": "intellif", "country": "CN", "position": "Researcher" }, { "name": "Qin Zou", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Rong Xiao", "gender": "Male", "institution": "Intellifusion", "country": "CN", "position": "Director" }, { "name": "Xianbiao Qi", "gender": "Male", "institution": "AITC", "country": "CN", "position": "Principal Researcher" }, { "name": "Xili Dai", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Yelin He", "gender": "unknown", "institution": "Intellifusion", "country": "CN", "position": "Researcher" }, { "name": "Chun-guang Li", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications, P.R. China", "country": "CN", "position": "Associate Professor" } ], "n_formula": 76, "n_formula_1": 27, "n_ref_uni": 40, "n_ref": 102, "n_ref_all": 135, "n_fig": 12, "n_tab": 4, "L_tab": 4870, "n_element_tab": 604, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 262, "n_element_tab_1": 25, "formula_len_all": 6803, "formula_len_all_1": 1827, "len_all": 184412, "len_all_1": 59517, "len_abs": 1334, "len_title": 101, "len_sents": 49292, "len_sents_1": 23987, "n_sents": 481, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1411, "title": "Taming Transformer Without Using Learning Rate Warmup", "abs": "Scaling Transformer to a large scale without using some technical tricks such as learning rate warump and an obviously lower learning rate, is an extremely challenging task, and is increasingly gaining more attention. In this paper, we provide a theoretical analysis for training Transformer and reveal a key problem behind the model crash phenomenon in the training, \\ie, the spectral energy concentration of $W_q^{\\top} W_k$ (where $W_q$ and $W_k$ are the projection matrices for query and key in Transformer), which is the reason for a malignant entropy collapse. To remedy this problem, motivated by Weyl's Inequality, we present a novel optimization strategy---making weight updating in successive steps smooth, that is, if the ratio $\\frac{\\sigma_{1}(\\nabla W_t)}{\\sigma_{1}(W_{t-1})}$ is larger than a threshold, where $\\nabla W_t$ is the updating quantity in step $t$, we will automatically bound the learning rate to a weighted multiply of $\\frac{\\sigma_{1}(W_{t-1})}{\\sigma_{1}(\\nabla W_t)}$. Our optimization strategy is able to prevent the rapid spectral energy concentration to only a few directions, and thus is able to avoid the malignant entropy collapse that will trigger the model crash. We conduct extensive experiments using ViT, Swin-Transformer and GPT, showing that our optimization strategy can effectively and stably train these (Transformer) models without using learning rate warmup.", "keywords": [ "Transformer", "Training Dynamics", "Model Crash" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "GeTBk67mK6", "primary_area": "foundation or frontier models, including LLMs", "n_author": 16, "author_info_list": [ { "name": "Aoxiao Zhong", "gender": "unknown", "institution": "Squirrel AI", "country": "US", "position": "Researcher" }, { "name": "BOYAN LI", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Hang Li", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Hui Xiong", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Jiahao Huo", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Jiamin Su", "gender": "Female", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Kun Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Philip S. Yu", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "US", "position": "Full Professor" }, { "name": "Shen Wang", "gender": "Male", "institution": "Squirrel AI", "country": "US", "position": "Researcher" }, { "name": "Tianlong Xu", "gender": "Male", "institution": "Squirrel Ai Learning", "country": "US", "position": "Principal Researcher" }, { "name": "Xiong Gao", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Xuming Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Yibo Yan", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Zhendong Chu", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Qingsong Wen", "gender": "Male", "institution": "Squirrel Ai Learning", "country": "US", "position": "Principal Researcher" }, { "name": "yifan zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 55, "n_ref": 124, "n_ref_all": 141, "n_fig": 10, "n_tab": 4, "L_tab": 2990, "n_element_tab": 210, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1040, "n_element_tab_1": 111, "formula_len_all": 259, "formula_len_all_1": 259, "len_all": 237310, "len_all_1": 61734, "len_abs": 1434, "len_title": 113, "len_sents": 59784, "len_sents_1": 28685, "n_sents": 405, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 111, "L_abs": 1474, "title": "ErrorRadar: Benchmarking Complex Mathematical Reasoning of Multimodal Large Language Models Via Error Detection", "abs": "As the field of Multimodal Large Language Models (MLLMs) continues to evolve, their potential to revolutionize artificial intelligence is particularly promising, especially in addressing mathematical reasoning tasks. Current mathematical benchmarks predominantly focus on evaluating MLLMs' problem-solving ability, yet there is a crucial gap in addressing more complex scenarios such as error detection, for enhancing reasoning capability in complicated settings. To fill this gap, we formally formulate the new task \u2014 \\textbf{multimodal error detection}, and introduce ErrorRadar, the first benchmark designed to assess MLLMs' capabilities in such a task. ErrorRadar evaluates two sub-tasks: \\textit{error step identification} and \\textit{error categorization}, providing a comprehensive framework for evaluating MLLMs' complex mathematical reasoning ability. It consists of 2,500 high-quality multimodal K-12 mathematical problems, collected from real-world student interactions in an educational organization, with rigorous annotation and rich metadata such as problem type and error category. Through extensive experiments, we evaluated both open-source and closed-source representative MLLMs, benchmarking their performance against educational expert evaluators. Results indicate significant challenges still remain, as GPT-4o with best performance is still around 10\\% behind human evaluation. The dataset is available at https://anonymous.4open.science/r/Error-Radar.", "keywords": [ "Multimodal Large Language Model", "Complex Reasoning", "Error Detection" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 4, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "GeMWhBIzrk", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 6, "author_info_list": [ { "name": "Chong Chen", "gender": "Male", "institution": "China University of Petroleum", "country": "CN", "position": "Associate Professor" }, { "name": "Yaru Xue", "gender": "unknown", "institution": "China university of petroleum(Beijing)", "country": "CN", "position": "Associate Professor" }, { "name": "Zongyu Han", "gender": "Male", "institution": "China University of Petroleum Beijing", "country": "CN", "position": "MS student" }, { "name": "Zhu xiaoyu", "gender": "Male", "institution": "China University of Petroleum Beijing", "country": "CN", "position": "MS student" }, { "name": "Li Yifan", "gender": "Male", "institution": "China University of Petroleum Beijing", "country": "CN", "position": "MS student" }, { "name": "Niu Yixiao", "gender": "unknown", "institution": "China University Of Petroleum", "country": "CN", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 15, "n_ref_uni": 30, "n_ref": 40, "n_ref_all": 63, "n_fig": 8, "n_tab": 3, "L_tab": 198, "n_element_tab": 14, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 34, "n_element_tab_1": 7, "formula_len_all": 1291, "formula_len_all_1": 1067, "len_all": 121908, "len_all_1": 60277, "len_abs": 1625, "len_title": 143, "len_sents": 45809, "len_sents_1": 31677, "n_sents": 254, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1632, "title": "Groundwater Seepage Modeling in a River-Canal System based on Physics-Informed Neural Networks", "abs": "Neural networks, especially deep learning, have achieved revolutionary advances in several domains, including image and speech recognition, with excellent results. However, their reliance on labeled data, lack of interpretability, and inconsistency with physical principles limit their applicability in groundwater seepage prediction and other scientific disciplines. Physics-Informed Neural Networks (PINNs) significantly improve these issues by integrating physical knowledge with neural networks. This study focuses on modeling the groundwater flow field and proposes a physics-informed river-canal groundwater seepage model (PI-RGSM). This model enables self-supervised learning by incorporating hard constraints of boundary and initial conditions, utilizing hydrogeological parameters and boundary conditions as direct inputs, thus diminishing dependence on observable data. Compared to the baseline PINNs, the PI-RGSM adapts to and accurately predicts diverse seepage situations with just one training session, achieving a mean coefficient of determination of 0.978. To further enhance applicability in complex dynamic groundwater seepage situations, we propose PI-RGSM-K, which builds upon PI-RGSM. This model simulates heterogeneous groundwater seepage fields and improves performance in complex seepage environments through parameterized hydraulic conductivity field $K(x,y)$ and fine-adjusted model architecture, attaining a mean coefficient of determination of 0.982. The physics-informed neural network models proposed in this study demonstrate exceptional efficacy in precisely forecasting groundwater seepage behavior.", "keywords": [ "Physics-Informed Neural Networks", "Groundwater Prediction", "Definite condition", "Hard constraints", "Self-Supervised" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 1, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 2, 3, 4, 4 ] }, { "paper_id": "Ge7okBGZYi", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Dinesh Manocha", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Professor" }, { "name": "Matthias Zwicker", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Samuel Audia", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Soheil Feizi", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "Assistant Professor" } ], "n_formula": 25, "n_formula_1": 24, "n_ref_uni": 23, "n_ref": 98, "n_ref_all": 134, "n_fig": 10, "n_tab": 5, "L_tab": 730, "n_element_tab": 121, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 213, "n_element_tab_1": 36, "formula_len_all": 2015, "formula_len_all_1": 2089, "len_all": 151634, "len_all_1": 67454, "len_abs": 1803, "len_title": 173, "len_sents": 50680, "len_sents_1": 30367, "n_sents": 420, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 125, "L_abs": 1810, "title": "How Learnable Grids Recover Fine Detail in Low Dimesions: A Neural Tangent Kernel Analysis of Multigrid Parameteric Encodings", "abs": "Neural networks that map between low dimensional spaces are ubiquitous in\ncomputer graphics and scientific computing; however, in their naive\nimplementation, they are unable to learn high frequency information. We present\na comprehensive analysis comparing the two most common techniques for mitigating\nthis spectral bias: Fourier feature encodings (FFE) and multigrid parametric\nencodings (MPE). FFEs are seen as the standard for low dimensional mappings, but\nMPEs often outperform them and learn representations with higher resolution and\nfiner detail. FFE's roots in the Fourier transform, make it susceptible to\naliasing if pushed too far, while MPEs, which use a learned grid structure, have\nno such limitation. To understand the difference in performance, we use the\nneural tangent kernel (NTK) to evaluate these encodings through the lens of an\nanalogous kernel regression. By finding a lower bound on the smallest eigenvalue\nof the NTK, we prove that MPEs improve a network's performance through the\nstructure of their grid and not their learnable embedding. This mechanism is\nfundamentally different from FFEs, which rely solely on their embedding space to\nimprove performance. Results are empirically validated on a 2D image regression\ntask using images taken from 100 synonym sets of ImageNet and 3D implicit\nsurface regression on objects from the Stanford graphics dataset. Using peak\nsignal-to-noise ratio (PSNR) and multiscale structural similarity (MS-SSIM) to\nevaluate how well fine details are learned, we show that the MPE increases the\nminimum eigenvalue by 8 orders of magnitude over the baseline and 2 orders of\nmagnitude over the FFE. The increase in spectrum corresponds to a 15 dB (PSNR) /\n0.65 (MS-SSIM) increase over baseline and a 12 dB (PSNR) / 0.33 (MS-SSIM) increase over the\nFFE.", "keywords": [ "neural tangent kernel", "compute graphics", "scientific computing", "fourier feature encodings", "multigrid parametric encodings", "encodings" ], "rating_list": [ 3, 8, 5, 5 ], "soundness_list": [ 3, 4, 2, 2 ], "presentation_list": [ 2, 4, 2, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 1, 2, 2, 4 ] }, { "paper_id": "GdbQyFOUlJ", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Hieu Pham", "gender": "Male", "institution": "VinUniversity", "country": "VN", "position": "Assistant Professor" }, { "name": "Nhat Hoang-Xuan", "gender": "unknown", "institution": "University of Florida", "country": "US", "position": "PhD student" }, { "name": "Phi Le Nguyen", "gender": "Female", "institution": "Hanoi University of Science and Technology", "country": "VN", "position": "Associate Professor" }, { "name": "Tue Minh Cao", "gender": "Male", "institution": "Hanoi University of Science and Technology", "country": "VN", "position": "Researcher" }, { "name": "My Thai", "gender": "Female", "institution": "University of Florida", "country": "US", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 33, "n_ref": 101, "n_ref_all": 136, "n_fig": 16, "n_tab": 4, "L_tab": 764, "n_element_tab": 54, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 777, "formula_len_all_1": 450, "len_all": 179608, "len_all_1": 75831, "len_abs": 1328, "len_title": 137, "len_sents": 61812, "len_sents_1": 37463, "n_sents": 452, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 88, "L_abs": 1332, "title": "NeurFlow: Interpreting Neural Networks through Neuron Groups and Functional Interactions", "abs": "Understanding the inner workings of neural networks is essential for enhancing model performance and interpretability. Current research predominantly focuses on examining the connection between individual neurons and the model's final predictions, which suffers from challenges in interpreting the internal workings of the model, particularly when neurons encode multiple unrelated features. In this paper, we propose a novel framework that transitions the focus from analyzing individual neurons to investigating groups of neurons, shifting the emphasis from neuron-output relationships to the functional interactions between neurons. Our automated framework, NeurFlow, first identifies core neurons and clusters them into groups based on shared functional relationships, enabling a more coherent and interpretable view of the network\u2019s internal processes. This approach facilitates the construction of a hierarchical circuit representing neuron interactions across layers, thus improving interpretability while reducing computational costs. Our extensive empirical studies validate the fidelity of our proposed NeurFlow. Additionally, we showcase its utility in practical applications such as image debugging and automatic concept labeling, thereby highlighting its potential to advance the field of neural network explainability.", "keywords": [ "Explainable AI", "Functional Interactions", "Critical Neurons", "Concept Circuits" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "GdXI5zCoAt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Jiahao Xu", "gender": "unknown", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Tian Liang", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Wenxiang Jiao", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Xing Wang", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Xingyu Chen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhaopeng Tu", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" }, { "name": "Zhijie Wang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhiwei He", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Zhuosheng Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Rui Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 35, "n_formula_1": 27, "n_ref_uni": 31, "n_ref": 64, "n_ref_all": 83, "n_fig": 6, "n_tab": 3, "L_tab": 1675, "n_element_tab": 318, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 936, "n_element_tab_1": 175, "formula_len_all": 2089, "formula_len_all_1": 1669, "len_all": 116986, "len_all_1": 53428, "len_abs": 985, "len_title": 86, "len_sents": 27809, "len_sents_1": 22561, "n_sents": 227, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 38, "L_abs": 989, "title": "RaSA: Rank-Sharing Low-Rank Adaptation", "abs": "Low-rank adaptation (LoRA) has been prominently employed for parameter-efficient fine-tuning of large language models (LLMs). However, the limited expressive capacity of LoRA, stemming from the low-rank constraint, has been recognized as a bottleneck, particularly in rigorous tasks like code generation and mathematical reasoning. To address this limitation, we introduce Rank-Sharing Low-Rank Adaptation (RaSA), an innovative extension that enhances the expressive capacity of LoRA by leveraging partial rank sharing across layers. By forming a shared rank pool and applying layer-specific weighting, RaSA effectively increases the number of ranks without augmenting parameter overhead. Our theoretically grounded and empirically validated approach demonstrates that RaSA not only maintains the core advantages of LoRA but also significantly boosts performance in challenging code and math tasks. Code, data and scripts are available at: https://anonymous.4open.science/r/RaSA-ICLR-0E25.", "keywords": [ "parameter-efficient fine-tuning", "large language model", "low-rank adaptation" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 4, 4, 3, 2 ], "presentation_list": [ 4, 4, 3, 2 ], "contribution_list": [ 4, 4, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "Gd6QrBLHBN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Andrei Ioan Muresanu", "gender": "Male", "institution": "University of Waterloo", "country": "", "position": "Undergrad student" }, { "name": "Atrisha Sarkar", "gender": "Female", "institution": "University of Western Ontario", "country": "CA", "position": "Assistant Professor" }, { "name": "Dylan Hadfield-Menell", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Gillian K Hadfield", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Joel Z Leibo", "gender": "Male", "institution": "Google", "country": "", "position": "Research scientist" }, { "name": "Nikhil Chandak", "gender": "Male", "institution": "Max-Planck-Institute for Intelligent Systems", "country": "DE", "position": "PhD student" }, { "name": "Rakshit Trivedi", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Shuhui Zhu", "gender": "Female", "institution": "University of Waterloo", "country": "CA", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 36, "n_ref": 65, "n_ref_all": 97, "n_fig": 24, "n_tab": 1, "L_tab": 10, "n_element_tab": 1, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 170, "formula_len_all_1": 151, "len_all": 182565, "len_all_1": 69994, "len_abs": 1670, "len_title": 123, "len_sents": 74974, "len_sents_1": 38059, "n_sents": 485, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1682, "title": "Altared Environments: The Role of Normative Infrastructure in AI Alignment", "abs": "Cooperation is central to human societies, which they achieve by constantly tack-\nling the alignment problem of ensuring self-interested individuals act in ways that\nbenefit the groups in which they live. As AI agents become pervasive in shared\nenvironments, it will be similarly crucial for them to align with the cooperative\ngoals of human groups. Current AI alignment research largely focuses on em-\nbedding specified or learned norms into agents to achieve this cooperation. While\nvaluable, this approach overlooks the role that institutions play in aligning human\nbehavior to achieve cooperative gains and thus overlooks a potential alignment\ntechnique for AI agents. We address this gap by proposing Altared Games, a\nnovel formal extension of Markov games that incorporates an altar\u2014a classification institution providing explicit normative guidance to agents. Our approach focuses on a challenging setting where norms are dynamic, thereby requiring agents\nto adapt to the evolving norm content represented by the altar. Using multi-agent\nreinforcement learning (MARL) as a computational model of AI agents, we con-\nduct experiments in two mixed-motive environments: Commons Harvest, which\nmodels resource sustainability, and Allelopathic Harvest, which involves coordination under conflicting incentives. Our results demonstrate that the altar enables\nagents to adapt effectively to dynamic norms, engage in accurate sanctioning, and\nachieve higher social welfare compared to systems without a classification institution. These findings highlight the importance of normative institutions in fostering\ncooperative, adaptable AI agents operating in complex real-world settings.", "keywords": [ "Instituions", "Norms", "Cooperative AI", "Multi-agent Systems", "Reinforcement Learning", "Alignment" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 1, 1, 1, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "GcvLoqOoXL", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Dailan He", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jian Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jingjing Liu", "gender": "unknown", "institution": "Tsinghua University", "country": "US", "position": "Full Professor" }, { "name": "Ming Sun", "gender": "Male", "institution": "Kuaishou Tech", "country": "CN", "position": "Researcher" }, { "name": "Tongda Xu", "gender": "Non-Binary", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Xingtong Ge", "gender": "Male", "institution": "Sensetime", "country": "CN", "position": "Researcher" }, { "name": "Xinjie Zhang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" }, { "name": "Xiyan Cai", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Ya-Qin Zhang", "gender": "Male", "institution": "AIR, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yan Wang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 53, "n_formula_1": 23, "n_ref_uni": 42, "n_ref": 131, "n_ref_all": 164, "n_fig": 11, "n_tab": 8, "L_tab": 2431, "n_element_tab": 289, "n_fig_1": 6, "n_tab_1": 9, "L_tab_1": 2709, "n_element_tab_1": 289, "formula_len_all": 4551, "formula_len_all_1": 1837, "len_all": 153795, "len_all_1": 56385, "len_abs": 1580, "len_title": 148, "len_sents": 41467, "len_sents_1": 20728, "n_sents": 382, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 99, "L_abs": 1689, "title": "Rethinking Diffusion Posterior Sampling: From Conditional Score Estimator to Maximizing a Posterior", "abs": "Recent advancements in diffusion models have been leveraged to address inverse problems without additional training, and Diffusion Posterior Sampling (DPS) (Chung et al., 2022a) is among the most popular approaches. Previous analyses suggest that DPS accomplishes posterior sampling by approximating the conditional score. While in this paper, we demonstrate that the conditional score approximation employed by DPS is not as effective as previously assumed, but rather aligns more closely with the principle of maximizing a posterior (MAP). This assertion is substantiated through an examination of DPS on 512$\\times$512 ImageNet images, revealing that: 1) DPS\u2019s conditional score estimation significantly diverges from the score of a well-trained conditional diffusion model and is even inferior to the unconditional score; 2) The mean of DPS\u2019s conditional score estimation deviates significantly from zero, rendering it an invalid score estimation; 3) DPS generates high-quality samples with significantly lower diversity. In light of the above findings, we posit that DPS more closely resembles MAP than a conditional score estimator, and accordingly propose the following enhancements to DPS: 1) we explicitly maximize the posterior through multi-step gradient ascent and projection; 2) we utilize a light-weighted conditional score estimator trained with only 100 images and 8 GPU hours. Extensive experimental results indicate that these proposed improvements significantly enhance DPS's performance. The source code for these improvements is provided in https://github.com/tongdaxu/Rethinking-Diffusion-Posterior-Sampling-From-Conditional-Score-Estimator-to-Maximizing-a-Posterior.", "keywords": [ "Diffusion models", "Inverse problem" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 4, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "GcbhbZsgiu", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Yi Yang", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Yixuan Tang", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zhuoyi PENG", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 26, "n_ref": 94, "n_ref_all": 123, "n_fig": 10, "n_tab": 8, "L_tab": 9073, "n_element_tab": 715, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 3530, "n_element_tab_1": 259, "formula_len_all": 661, "formula_len_all_1": 690, "len_all": 162432, "len_all_1": 75172, "len_abs": 1595, "len_title": 76, "len_sents": 54449, "len_sents_1": 32668, "n_sents": 441, "n_sents_1": 279, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 28, "L_abs": 1600, "title": "Adversarial Mixup Unlearning", "abs": "Machine unlearning is a critical area of research aimed at safeguarding data privacy by enabling the removal of sensitive information from machine learning models. One unique challenge in this field is catastrophic unlearning, where erasing specific data from a well-trained model unintentionally removes essential knowledge, causing the model to deviate significantly from a retrained one. To address this, we introduce a novel approach that regularizes the unlearning process by utilizing synthesized mixup samples, which simulate the data susceptible to catastrophic effects. At the core of our approach is a generator-unlearner framework, MixUnlearn, where a generator adversarially produces challenging mixup examples, and the unlearner effectively forgets target information based on these synthesized data. Specifically, we first introduce a novel contrastive objective to train the generator in an adversarial direction: generating examples that prompt the unlearner to reveal information that should be forgotten, while losing essential knowledge. Then the unlearner, guided by two other contrastive loss terms, processes the synthesized and real data jointly to ensure accurate unlearning without losing critical knowledge, overcoming catastrophic effects. Extensive evaluations across benchmark datasets demonstrate that our method significantly outperforms state-of-the-art approaches, offering a robust solution to machine unlearning. This work not only deepens understanding of unlearning mechanisms but also lays the foundation for effective machine unlearning with mixup augmentation.", "keywords": [ "Machine Unlearning", "Catastrophic Unlearning", "Mixup Regularization", "Adversarial Training" ], "rating_list": [ 6, 3, 6 ], "soundness_list": [ 4, 2, 3 ], "presentation_list": [ 4, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 2 ] }, { "paper_id": "GcYcUE3GvY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Chien-Yi Wang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Ci-Siang Lin", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "PhD student" }, { "name": "Min-hung Chen", "gender": "Male", "institution": "NVIDIA", "country": "TW", "position": "Research Scientist" }, { "name": "Sifei Liu", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yu-Chiang Frank Wang", "gender": "Male", "institution": "NVIDIA", "country": "TW", "position": "Research Director" }, { "name": "I JIEH LIU", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 48, "n_ref": 113, "n_ref_all": 127, "n_fig": 5, "n_tab": 7, "L_tab": 2340, "n_element_tab": 285, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1991, "n_element_tab_1": 233, "formula_len_all": 106, "formula_len_all_1": 106, "len_all": 136766, "len_all_1": 56678, "len_abs": 1407, "len_title": 122, "len_sents": 30261, "len_sents_1": 26488, "n_sents": 182, "n_sents_1": 140, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1413, "title": "Temporal Prompting Matters: Rethinking Referring Video Object Segmentation", "abs": "Referring Video Object Segmentation (RVOS) aims to segment the object referred to by the query sentence in the video. Most existing methods require end-to-end training with dense mask annotations, which could be computation-consuming and less scalable. In this work, we rethink the RVOS problem and aim to investigate the key to this task. Based on existing foundation segmentation models, we decompose the RVOS task into referring, video, and segmentation factors, and propose a Temporal Prompt Generation and Selection (Tenet) framework to address the referring and video factors while leaving the segmentation problem to foundation models. To efficiently adapt image-based foundation segmentation models to referring video object segmentation, we leverage off-the-shelf object detectors and trackers to produce temporal prompts associated with the referring sentence. While high-quality temporal prompts could be produced, they can not be easily identified from confidence scores. To tackle this issue, we propose Prompt Preference Learning to evaluate the quality of the produced temporal prompts. By taking such prompts to instruct image-based foundation segmentation models, we would be able to produce high-quality masks for the referred object, enabling efficient model adaptation to referring video object segmentation. Experiments on RVOS benchmarks demonstrate the effectiveness of the Tenet framework.", "keywords": [ "Referring Video Object Segmentation", "Foundation Models", "Prompting" ], "rating_list": [ 5, 3, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "GcJE0HPy4X", "primary_area": "datasets and benchmarks", "n_author": 18, "author_info_list": [ { "name": "Ankit Parag Shah", "gender": "Male", "institution": "Accenture", "country": "US", "position": "Principal Researcher" }, { "name": "Hao Chen", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Haobo Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Haoyu Wang", "gender": "Male", "institution": "Yale University", "country": "", "position": "PhD student" }, { "name": "Hengxiang Zhang", "gender": "unknown", "institution": "Southern University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Hongxin Wei", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "James Davis", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "", "position": "Full Professor" }, { "name": "Jiaheng Wei", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Jindong Wang", "gender": "Male", "institution": "William & Mary", "country": "US", "position": "Assistant Professor" }, { "name": "Jinlong Pang", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "", "position": "PhD student" }, { "name": "Lei Feng", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" }, { "name": "Minghao Liu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Ruixuan Xiao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Xinlei He", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Yang Liu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "Zhaowei Zhu", "gender": "Male", "institution": "Docta.ai", "country": "US", "position": "Researcher" }, { "name": "Zhongruo Wang", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Scientist" }, { "name": "Zonglin Di", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 58, "n_ref": 137, "n_ref_all": 167, "n_fig": 8, "n_tab": 10, "L_tab": 1964, "n_element_tab": 188, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1729, "n_element_tab_1": 136, "formula_len_all": 180, "formula_len_all_1": 88, "len_all": 183974, "len_all_1": 65676, "len_abs": 1544, "len_title": 130, "len_sents": 50577, "len_sents_1": 30377, "n_sents": 389, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1550, "title": "Automatic Dataset Construction (ADC): Sample Collection, Data Curation, and Beyond", "abs": "Large-scale data collection is essential for developing personalized training data, mitigating the shortage of training data, and fine-tuning specialized models. However, creating high-quality datasets quickly and accurately remains a challenge due to annotation errors, the substantial time and costs associated with human labor. To address these issues, we propose Automatic Dataset Construction (ADC), an innovative methodology that automates dataset creation with negligible cost and high efficiency. Taking the image classification task as a starting point, ADC leverages LLMs for the detailed class design and code generation to collect relevant samples via search engines, significantly reducing the need for manual annotation and speeding up the data generation process. Despite these advantages, ADC also encounters real-world challenges such as label errors (label noise) and imbalanced data distributions (label bias). We provide open-source software that incorporates existing methods for label error detection, robust learning under noisy and biased data, ensuring a higher-quality training data and more robust model training procedure. Furthermore, we design three benchmark datasets focused on label noise detection, label noise learning, and class-imbalanced learning. These datasets are vital because there are few existing datasets specifically for label noise detection, despite its importance. Finally, we evaluate the performance of existing popular methods on these datasets, thereby facilitating further research in the field.", "keywords": [ "Dataset Construction", "LLM", "Label Noise", "Class-Imbalance" ], "rating_list": [ 6, 5, 5, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "GcFX8rZNSX", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 10, "author_info_list": [ { "name": "Chun Yuan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Fei Ma", "gender": "Male", "institution": "Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ)", "country": "CN", "position": "Researcher" }, { "name": "Fei Richard Yu", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "Full Professor" }, { "name": "Feng Xiong", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Songsong Huang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Wanyun Zhu", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "MS student" }, { "name": "Xianglong He", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Yongxian Wei", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Zhanqiu Zhang", "gender": "unknown", "institution": "Individual Researcher", "country": "CN", "position": "Researcher" }, { "name": "\u7a0b\u6da6\u66e6", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 44, "n_ref": 89, "n_ref_all": 98, "n_fig": 3, "n_tab": 4, "L_tab": 1599, "n_element_tab": 86, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1803, "n_element_tab_1": 103, "formula_len_all": 752, "formula_len_all_1": 753, "len_all": 117484, "len_all_1": 59508, "len_abs": 1031, "len_title": 75, "len_sents": 31755, "len_sents_1": 26520, "n_sents": 238, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 27, "L_abs": 1036, "title": "LEARN TO LEARN CONSISTENTLY", "abs": "In the few-shot learning problem, a model trained on a disjoint meta-train dataset\nis required to address novel tasks with limited novel examples. A key challenge in\nfew-shot learning is the model\u2019s propensity to learn biased shortcut features(e.g.,\nbackground, noise, shape, color), which are sufficient to distinguish the few ex-\namples during fast adaptation but lead to poor generalization. In our work, we\nobserved when the model learns with higher consistency, the model tends to be\nless influenced by shortcut features, resulting in better generalization. Based on\nthe observation, we propose a simple yet effective meta-learning method named\nMeta Self-Distillation. By maximizing the consistency of the learned knowledge\nduring the meta-train phase, the model initialized by our method shows better\ngeneralization in the meta-test phase. Extensive experiments demonstrate that our\nmethod improves the model\u2019s generalization across various few-shot classification\nscenarios and enhances the model\u2019s ability to learn consistently.", "keywords": [ "meta learning", "few-shot learning", "meta self-distillation", "consistency of learned knowledge" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "Gc2qkiYUkh", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Grant M. Rotskoff", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Javan Tahir", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Surya Ganguli", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" } ], "n_formula": 111, "n_formula_1": 24, "n_ref_uni": 41, "n_ref": 65, "n_ref_all": 103, "n_fig": 13, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9154, "formula_len_all_1": 1203, "len_all": 174151, "len_all_1": 70330, "len_abs": 1636, "len_title": 127, "len_sents": 53700, "len_sents_1": 31503, "n_sents": 460, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1656, "title": "Features are fate: a theory of transfer learning in high-dimensional regression", "abs": "With the emergence of large-scale pre-trained neural networks, methods to adapt such \"foundation\" models to data-limited downstream tasks have become a necessity.\nFine-tuning, preference optimization, and transfer learning have all been successfully employed for these purposes when the target task closely resembles the source task, but a precise theoretical understanding of ``task similarity'' is still lacking. \nWhile conventional wisdom suggests that simple measures of similarity between source and target distributions, such as $\\phi$-divergences or integral probability metrics, can directly predict the success of transfer, we prove the surprising fact that, in general, this is not the case.\nWe adopt, instead, a \\emph{feature-centric} viewpoint on transfer learning and establish a number of theoretical results that demonstrate that when the target task is well represented by the feature space of the pre-trained model, transfer learning outperforms training from scratch.\nWe study deep linear networks as a minimal model of transfer learning in which we can analytically characterize the transferability phase diagram as a function of the target dataset size and the feature space overlap.\nFor this model, we establish rigorously that when the feature space overlap between the source and target tasks is sufficiently strong, both linear transfer and fine-tuning improve performance, especially in the low data limit. \nThese results build on an emerging understanding of feature learning dynamics in deep linear networks, and we demonstrate numerically that the rigorous results we derive for the linear case also apply to nonlinear networks.", "keywords": [ "transfer learning", "deep linear networks", "fine tuning", "random matrix theory", "high dimensional statistics" ], "rating_list": [ 6, 6, 5, 3, 6 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 2, 4, 4, 2, 3 ], "contribution_list": [ 3, 3, 2, 2, 3 ], "confidence_list": [ 2, 3, 4, 3, 2 ] }, { "paper_id": "GbgCRJedQ7", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Haoze He", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Heather Miller", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Juncheng Li", "gender": "Male", "institution": "Two Sigma", "country": "", "position": "Researcher" }, { "name": "Xuan Jiang", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 3, "n_ref_uni": 33, "n_ref": 109, "n_ref_all": 154, "n_fig": 7, "n_tab": 6, "L_tab": 3060, "n_element_tab": 427, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2458, "n_element_tab_1": 387, "formula_len_all": 1194, "formula_len_all_1": 139, "len_all": 185333, "len_all_1": 74484, "len_abs": 1178, "len_title": 97, "len_sents": 65429, "len_sents_1": 33945, "n_sents": 462, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 1156, "title": "Sparse Matrix in Large Language Model Fine-tuning", "abs": "Various parameter-efficient fine-tuning (PEFT) methods, including LoRA and its variants, have gained popularity for reducing computational costs. However, there is often an accuracy gap between PEFT approaches and full fine-tuning (FT), and this discrepancy has not yet been systematically explored. In this work, we introduce a method for selecting sparse sub-matrices that aim to minimize the performance gap between PEFT vs. full fine-tuning (FT) while also reducing both fine-tuning computational cost and memory cost. Our Sparse Matrix Tuning (SMT) method begins by identifying the most significant sub-matrices in the gradient update, updating only these blocks during the fine-tuning process. In our experiments, we demonstrate that SMT consistently surpasses other PEFT baseline (e.g. LoRA and DoRA) in fine-tuning popular large language models such as LLaMA across a broad spectrum of tasks, while reducing the GPU memory footprint by 67\\% compared to FT. We also examine how the performance of LoRA and DoRA tends to plateau and decline as the number of trainable parameters increases, in contrast, our SMT method does not suffer from such issue.", "keywords": [ "Parameter-efficient Finetuning", "Large Language Model", "Large Language Model Systm" ], "rating_list": [ 8, 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3, 2 ] }, { "paper_id": "GbXn0Dgf7f", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Jordan Malof", "gender": "Male", "institution": "University of Missouri", "country": "US", "position": "Assistant Professor" }, { "name": "Saad Lahrichi", "gender": "Male", "institution": "University of Missouri", "country": "US", "position": "PhD student" }, { "name": "Simiao Ren", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Willie Padilla", "gender": "unknown", "institution": "Duke University", "country": "", "position": "Full Professor" }, { "name": "Yang Deng", "gender": "Male", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Leslie Collins", "gender": "Female", "institution": "Duke University", "country": "US", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 22, "n_ref": 43, "n_ref_all": 62, "n_fig": 5, "n_tab": 3, "L_tab": 624, "n_element_tab": 57, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 624, "n_element_tab_1": 57, "formula_len_all": 846, "formula_len_all_1": 889, "len_all": 99273, "len_all_1": 58100, "len_abs": 1215, "len_title": 91, "len_sents": 29868, "len_sents_1": 27980, "n_sents": 204, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 1227, "title": "Does Deep Active Learning Work in the Wild?", "abs": "Deep active learning (DAL) methods have shown significant improvements in sample efficiency compared to simple random sampling. While these studies are valuable, they nearly always assume that optimal DAL hyperparameter (HP) settings are known in advance, or optimize the HPs through repeating DAL several times with different HP settings. Here, we argue that in real-world settings, or _in the wild_, there is significant uncertainty regarding good HPs, and their optimization contradicts the premise of using DAL (i.e., we require labeling efficiency). In this study, we evaluate the performance of eleven modern DAL methods on eight benchmark problems as we vary a key HP shared by all methods: the pool ratio. Despite adjusting only one HP, our results indicate that eight of the eleven DAL methods sometimes underperform relative to simple random sampling and some frequently perform worse. Only three methods always outperform random sampling (albeit narrowly), and we find that these methods all utilize diversity to select samples - a relatively simple criterion. Our findings reveal the limitations of existing DAL methods when deployed _in the wild_, and present this as an important new open problem in the field.", "keywords": [ "deep learning", "active learning", "robustness", "sampling", "diversity" ], "rating_list": [ 3, 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 1, 2, 2 ], "presentation_list": [ 3, 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 3, 3, 3, 4, 5 ] }, { "paper_id": "GbEmJmnQCz", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Blaine Hoak", "gender": "Female", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Hadi Abdullah", "gender": "unknown", "institution": "VISA", "country": "US", "position": "Researcher" }, { "name": "Ke Wang", "gender": "unknown", "institution": "Visa Research", "country": "US", "position": "Research Scientist" }, { "name": "Sunpreet S. Arora", "gender": "Male", "institution": "Visa", "country": "", "position": "Researcher" }, { "name": "Yiwei Cai", "gender": "Male", "institution": "VISA", "country": "", "position": "Sr. Director" }, { "name": "Yizhen Wang", "gender": "Male", "institution": "VISA", "country": "", "position": "Research Scientist" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 18, "n_ref": 45, "n_ref_all": 57, "n_fig": 3, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 58, "formula_len_all_1": 58, "len_all": 84988, "len_all_1": 75410, "len_abs": 1445, "len_title": 103, "len_sents": 32762, "len_sents_1": 32762, "n_sents": 258, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1488, "title": "Is Memorization Actually Necessary for Generalization?", "abs": "Memorization is the ability of deep models to associate training data with seemingly random labels. Even though memorization may not align with a model's ability to generalize, recent work by~\\citet{feldman2020longtail} has demonstrated that memorization is in fact \\textit{necessary} for generalization. However, upon closer inspection, we find that their methodology has three limitations. First, the definition of memorization is imprecise, leading to contradictory results. Second, their proposed algorithm used for \\textit{approximating} the leave-one-out test (the gold standard for calculating memorization scores) suffers from a high approximation error. Three, the authors induce a distribution shift when calculating marginal utility, leading to flawed results. Having accounted for these errors, we re-evaluate the role of memorization on generalization. To do so, we track how memorization changes at different levels of generalization (test accuracy). We control model generalization by training 19 different combinations of models, datasets, and training optimizations. We find that memorization and generalization are \\textit{strongly} negatively correlated (Pearson -0.997): As one decreases, the other increases. This shows that memorization is not necessary for generalization, as otherwise, the correlation would have been positive. In light of these findings, future researchers are encouraged to design techniques that can accurately approximate memorization scores.", "keywords": [ "Memorization" ], "rating_list": [ 3, 3, 6, 5, 5 ], "soundness_list": [ 1, 2, 3, 4, 2 ], "presentation_list": [ 2, 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 2, 4 ], "confidence_list": [ 4, 3, 3, 3, 4 ] }, { "paper_id": "GamwMdPj0y", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 10, "author_info_list": [ { "name": "ANAN LIU", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Cong Xu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "PhD student" }, { "name": "Huan Liu", "gender": "unknown", "institution": "Tencent Technology Company", "country": "CN", "position": "Researcher" }, { "name": "Jian Xiao", "gender": "Male", "institution": "Tencent Technology Company", "country": "CN", "position": "Researcher" }, { "name": "Qianqian Pu", "gender": "Female", "institution": "Shopee", "country": "CN", "position": "Senior Engineer" }, { "name": "Shaowen Wang", "gender": "Not Specified", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Suncong Zheng", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Wei Zhang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Yuekui Yang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jian Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 43, "n_formula_1": 10, "n_ref_uni": 36, "n_ref": 53, "n_ref_all": 68, "n_fig": 4, "n_tab": 4, "L_tab": 1316, "n_element_tab": 151, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 831, "n_element_tab_1": 95, "formula_len_all": 4980, "formula_len_all_1": 704, "len_all": 120126, "len_all_1": 52450, "len_abs": 1687, "len_title": 104, "len_sents": 32822, "len_sents_1": 24621, "n_sents": 255, "n_sents_1": 159, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1694, "title": "C-Adam: Confidence-Based Optimization for Online Learning", "abs": "Modern recommendation systems frequently employ online learning to dynamically update their models with freshly collected data. The most commonly used optimizer for updating neural networks in these contexts is the Adam optimizer, which integrates momentum ($m_t$) and adaptive learning rate ($v_t$). However, the volatile nature of online learning data, characterized by its frequent distribution shifts and presence of noises, poses significant challenges to Adam's standard optimization process: (1) Adam may use outdated momentum and the average of squared gradients, resulting in slower adaptation to distribution changes, and (2) Adam's performance is adversely affected by data noise. To mitigate these issues, we introduce CAdam, a confidence-based optimization strategy that assesses the consistence between the momentum and the gradient for each parameter dimension before deciding on updates. If momentum and gradient are in sync, CAdam proceeds with parameter updates according to Adam's original formulation; if not, it temporarily withholds updates and monitors potential shifts in data distribution in subsequent iterations. This method allows CAdam to distinguish between the true distributional shifts and mere noise, and adapt more quickly to new data distributions. Our experiments with both synthetic and real-world datasets demonstrate that CAdam surpasses other well-known optimizers, including the original Adam, in efficiency and noise robustness. Furthermore, in large-scale A/B testing within a live recommendation system, CAdam significantly enhances model performance compared to Adam, leading to substantial increases in the system's gross merchandise volume (GMV).", "keywords": [ "Optimization Algorithm", "Online Learning", "Recommendation Systems" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 4, 2, 2 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "GaWYCQMAq1", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Difan Zou", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Xingjun Ma", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Yu-Gang Jiang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yunhao Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" } ], "n_formula": 113, "n_formula_1": 18, "n_ref_uni": 30, "n_ref": 71, "n_ref_all": 97, "n_fig": 6, "n_tab": 5, "L_tab": 871, "n_element_tab": 106, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 517, "n_element_tab_1": 76, "formula_len_all": 8251, "formula_len_all_1": 1084, "len_all": 157896, "len_all_1": 59572, "len_abs": 1649, "len_title": 120, "len_sents": 51267, "len_sents_1": 28488, "n_sents": 440, "n_sents_1": 195, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1665, "title": "Towards a Theoretical Understanding of Memorization in Diffusion Models", "abs": "As diffusion probabilistic models (DPMs) are being employed as mainstream models for Generative Artificial Intelligence (GenAI), the study of their memorization of training data has attracted growing attention. Existing works in this direction aim to establish an understanding of whether or to what extent DPMs learn via memorization. Such an understanding is crucial for identifying potential risks of data leakage and copyright infringement in diffusion models and, more importantly, for trustworthy application of GenAI. Existing works revealed that conditional DPMs are more prone to training data memorization than unconditional DPMs, and the motivated data extraction methods are mostly for conditional DPMs. However, these understandings are primarily empirical, and extracting training data from unconditional models has been found to be extremely challenging. In this work, we provide a theoretical understanding of memorization in both conditional and unconditional DPMs under the assumption of model convergence. Our theoretical analysis indicates that extracting data from unconditional models can also be effective by constructing a proper surrogate condition. Based on this result, we propose a novel data extraction method named \\textbf{Surrogate condItional Data Extraction (SIDE)} that leverages a time-dependent classifier trained on the generated data as a surrogate condition to extract training data from unconditional DPMs. Empirical results demonstrate that our SIDE can extract training data in challenging scenarios where previous methods fail, and it is, on average, over 50\\% more effective across different scales of the CelebA dataset.", "keywords": [ "Diffusion Models Memorization; Diffusion Probabilistic models" ], "rating_list": [ 1, 3, 6, 5 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 1, 4, 3 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "GZYCCQacwH", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Jin Zhang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "", "position": "Associate Professor" }, { "name": "Wei Yao", "gender": "unknown", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiao Wang", "gender": "Not Specified", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Associate Professor" }, { "name": "Yaoshuai Ma", "gender": "unknown", "institution": "Southern University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 190, "n_formula_1": 37, "n_ref_uni": 34, "n_ref": 111, "n_ref_all": 194, "n_fig": 12, "n_tab": 9, "L_tab": 2428, "n_element_tab": 245, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1254, "n_element_tab_1": 83, "formula_len_all": 26512, "formula_len_all_1": 2402, "len_all": 234336, "len_all_1": 53174, "len_abs": 1058, "len_title": 162, "len_sents": 62094, "len_sents_1": 20323, "n_sents": 597, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 114, "L_abs": 1061, "title": "MeFBO: A Moreau Envelope Based First-Order Stochastic Gradient Method for Nonconvex Federated Bilevel Optimization", "abs": "Federated Bilevel Optimization (FBO) enables training machine learning models with nested structures across distributed devices while preserving data privacy. However, current FBO methods often impose restrictive assumptions, particularly the requirement of strong convexity in the lower-level objective. To overcome this limitation, we propose a first-order stochastic gradient method for general FBO problems, leveraging a Moreau envelope-based min-max optimization reformulation to handle potentially non-convex lower-level objectives. Unlike implicit gradient methods, our approach eliminates the need for second-order derivative information. We also establish rigorous theoretical guarantees for convergence rate and communication complexity, demonstrating linear speedup as the number of devices increases. Numerical experiments validate the effectiveness and efficiency of our method, showing comparable or superior performances in challenging scenarios, including federated loss function tuning on imbalanced datasets and federated hyper-representation.", "keywords": [ "Federated Bilevel Optimization", "Nonconvex", "Hessian-free", "Convergence Analysis", "Linear Speedup" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "GZVKo8T3EK", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Feng Chen", "gender": "Male", "institution": "University of Texas, Dallas", "country": "", "position": "Associate Professor" }, { "name": "Josh Wiedemeier", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Kangkook Jee", "gender": "Male", "institution": "The University of Texas at Dallas", "country": "", "position": "Assistant Professor" }, { "name": "Kunal Mukherjee", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Murat Kantarcioglu", "gender": "unknown", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "Full Professor" }, { "name": "Simon Klancher", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "MS student" }, { "name": "Tianhao Wang", "gender": "Male", "institution": "University of Texas at Dallas", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 12, "n_ref": 47, "n_ref_all": 62, "n_fig": 7, "n_tab": 7, "L_tab": 2356, "n_element_tab": 191, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1470, "n_element_tab_1": 161, "formula_len_all": 302, "formula_len_all_1": 302, "len_all": 104449, "len_all_1": 56491, "len_abs": 1146, "len_title": 101, "len_sents": 30756, "len_sents_1": 27809, "n_sents": 203, "n_sents_1": 145, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1151, "title": "ProvCreator: Synthesizing Graph Data with Text Attributes", "abs": "In cybersecurity, system provenance graphs are a key primitive to support intrusion detection and program identification tasks. Recent movement towards using data-hungry graph learning models for security-critical applications has exposed significant limitations in existing provenance datasets. Imbalanced representation of programs induces bias and performance degradation in downstream models. Further, these models rely on rich numeric and textual node attributes to accurately encode program behaviors, limiting the ability of existing data augmentation techniques to address data imbalance in provenance graphs.\n\nWe present PROVCREATOR, a novel graph synthesis framework designed for feature-rich system provenance graphs. PROVCREATOR learns the joint distribution of node attributes and graph structures conditioned on program class labels, enabling targeted generation of realistic system provenance graphs to supplement underrepresented programs. Our evaluation shows that PROVCREATOR produces provenance graphs with higher structural fidelity, attribute fidelity, and downstream utility compared to those of previous graph synthesis methods.", "keywords": [ "Synthetic", "Graph", "Security", "Intrusion Detection", "Provenance" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "GYwH71ugtC", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Jinsung Yoon", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Meeyoung Cha", "gender": "Female", "institution": " Max Planck Institute for Security and Privacy", "country": "DE", "position": "Director" }, { "name": "Sercan O Arik", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Seungeon Lee", "gender": "Male", "institution": "MPI-SP", "country": "DE", "position": "Intern" }, { "name": "Sungwon Han", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Integrated PhD student" } ], "n_formula": 19, "n_formula_1": 13, "n_ref_uni": 41, "n_ref": 91, "n_ref_all": 119, "n_fig": 10, "n_tab": 12, "L_tab": 2363, "n_element_tab": 273, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1364, "n_element_tab_1": 23, "formula_len_all": 582, "formula_len_all_1": 413, "len_all": 142008, "len_all_1": 63391, "len_abs": 919, "len_title": 91, "len_sents": 43027, "len_sents_1": 28850, "n_sents": 326, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 921, "title": "Retrieval Augmented Time Series Forecasting", "abs": "Time series forecasting uses historical data to predict future trends, leveraging the relationships between past observations and available features. In this paper, we propose, RAFT, a retrieval-augmented time series forecasting method to provide sufficient inductive biases and complement the model's learning capacity. When forecasting the subsequent time frames, we directly retrieve historical data candidates from the training dataset with patterns most similar to the input, and utilize the future values of these candidates alongside the inputs to obtain predictions. This simple approach augments the model's capacity by externally providing information about past patterns via retrieval modules. Our empirical evaluations on eight benchmark datasets show that RAFT consistently outperforms contemporary baselines, an average win ratio of 86% for multivariate forecasting and 80% for univariate forecasting tasks.", "keywords": [ "Time series forecasting", "Retrieval augmented model", "Deep learning" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "GYk0thSY1M", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "ChensenHuang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Dong Yi", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Guibo Zhu", "gender": "Male", "institution": " Institute of Automation, Chinese Academy of Science", "country": "CN", "position": "Associate Professor" }, { "name": "Guojing Ge", "gender": "Female", "institution": ", Institute of automation, Chinese academy of science", "country": "", "position": "Lecturer" }, { "name": "Haoran Chen", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Jinqiao Wang", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xuepeng Wang", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Associate Professor" }, { "name": "Yifei Luo", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 27, "n_ref": 80, "n_ref_all": 98, "n_fig": 5, "n_tab": 7, "L_tab": 1392, "n_element_tab": 161, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 659, "n_element_tab_1": 132, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 145238, "len_all_1": 71023, "len_abs": 1412, "len_title": 126, "len_sents": 42992, "len_sents_1": 32340, "n_sents": 376, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1354, "title": "Recurrent Context Compression: Efficiently Expanding the Context Window of LLM", "abs": "To extend the context length of Transformer-based large language models (LLMs) and improve comprehension capabilities, researchers often encounter constraints stemming from finite computational resources and bounded memory capacities. This work proposes a novel approach, termed Recurrent Context Compression (RCC), designed to efficiently expand the context window length of LLMs. Furthermore, we delve into the prevalent issue of degraded model performance when both instructional prompts and contextual information undergo compression for downstream tasks. To address this challenge, we propose a novel instruction reconstruction methodology aimed at mitigating the detrimental effects of this compression process. The effectiveness of our proposed approach was validated across multiple tasks while achieving an impressive context compression rate of at least 32x. On text reconstruction task, we maintain a BLEU-4 score close to 0.95. On passkey retrieval task, we achieve nearly 100% accuracy involving an extensive sequence length of 1 million tokens. On long-text question-answering task, we obtain comparable performance with the non-compressed LLM in F1 and Rouge scores. Our method also demonstrated competitive performance in long-text question-answering tasks compared to non-compressed methods, while significantly saving storage resources.", "keywords": [ "language model", "Context compression" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "GYik1jT3gE", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Divyansh Jhunjhunwala", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Gauri Joshi", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Zheng Xu", "gender": "unknown", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "PRANAY SHARMA", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Postdoc" } ], "n_formula": 278, "n_formula_1": 33, "n_ref_uni": 49, "n_ref": 97, "n_ref_all": 142, "n_fig": 24, "n_tab": 7, "L_tab": 1520, "n_element_tab": 123, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 35404, "formula_len_all_1": 2596, "len_all": 345579, "len_all_1": 74024, "len_abs": 1307, "len_title": 131, "len_sents": 76502, "len_sents_1": 27706, "n_sents": 872, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1312, "title": "Initialization Matters: Unraveling the Impact of Pre-Training on Federated Learning", "abs": "Initializing with pre-trained models when learning on downstream tasks is now standard practice in machine learning. Several recent works explore the benefits of pre-trained initialization in a federated learning (FL) setting, where the downstream training is performed at the edge clients with heterogeneous data distribution. These works show that starting from a pre-trained model can substantially reduce the adverse impact of data heterogeneity on the test performance of a model trained in a federated setting, with no changes to the standard FedAvg training algorithm. In this work, we provide a deeper theoretical understanding of this phenomenon. To do so, we study the class of two-layer convolutional neural networks (CNNs) and provide bounds on the training error convergence and test error of such a network trained with FedAvg. We introduce the notion of aligned and misaligned filters at initialization and show that the data heterogeneity only affects learning on misaligned filters. Starting with a pre-trained model typically results in fewer misaligned filters at initialization, thus producing a lower test error even when the model is trained in a federated setting with data heterogeneity. Experiments in synthetic settings and practical FL training on CNNs verify our theoretical findings.", "keywords": [ "Federated Learning", "Initialization", "Two-Layer CNN", "Pre-training", "Generalization" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 5, 3 ] }, { "paper_id": "GXzwq6waYb", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Erchin Serpedin", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Full Professor" }, { "name": "HASAN KURBAN", "gender": "Male", "institution": "Hamad Bin Khalifa University", "country": "QA", "position": "Assistant Professor" }, { "name": "Khalid Qaraqe", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "", "position": "Full Professor" }, { "name": "Parichit Sharma", "gender": "Male", "institution": "Indiana University", "country": "US", "position": "PhD student" }, { "name": "Rachad Atat", "gender": "Male", "institution": "Lebanese American University", "country": "LB", "position": "Assistant Professor" }, { "name": "Samir Abdaljalil", "gender": "unknown", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 40, "n_ref": 52, "n_ref_all": 69, "n_fig": 9, "n_tab": 2, "L_tab": 1288, "n_element_tab": 129, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 861, "n_element_tab_1": 110, "formula_len_all": 246, "formula_len_all_1": 228, "len_all": 126922, "len_all_1": 57789, "len_abs": 1399, "len_title": 128, "len_sents": 32102, "len_sents_1": 27150, "n_sents": 210, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1398, "title": "Scalable and Enhanced Hallucination Detection in LLMs using Semantic Clustering", "abs": "Large language models (LLMs) are increasingly being adopted across various domains, driven by their ability to generate general-purpose and domain-specific text. However, LLMs can also produce responses that seem plausible but are factually incorrect\u2014a phenomenon commonly referred to as \"hallucination\". This issue limits the potential and trustworthiness of LLMs, especially in critical fields such as medicine and law. Among the strategies proposed to address this problem uncertainty-based methods stand out due to their ease of implementation, independence from external data sources, and compatibility with standard LLMs. In this paper, we present an optimized semantic clustering framework for automated hallucination detection in LLMs, using sentence embeddings and hierarchical clustering. Our proposed method enhances both scalability and performance compared to existing approaches across different LLM models. This results in more homogeneous clusters, improved entropy scores, and a more accurate reflection of detected hallucinations. Our approach significantly boosts accuracy on widely used open and closed-book question-answering datasets such as TriviaQA, NQ, SQuAD, and BioASQ, achieving AUROC score improvements of up to 9.3% over the current state-of-the-art semantic entropy method. Further ablation studies highlight the effectiveness of different components of our approach.", "keywords": [ "Hallucination", "Semantic entropy", "LLMs", "Semantic clustering" ], "rating_list": [ 3, 3, 3, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 4 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "GXXQfSpJNI", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Agneet Chatterjee", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "John Leland", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Mahdi Ahmadi", "gender": "unknown", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "YooJung Choi", "gender": "unknown", "institution": "Arizona State University", "country": "", "position": "Assistant Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 21, "n_ref": 36, "n_ref_all": 54, "n_fig": 6, "n_tab": 3, "L_tab": 940, "n_element_tab": 171, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 763, "n_element_tab_1": 131, "formula_len_all": 308, "formula_len_all_1": 246, "len_all": 72011, "len_all_1": 65716, "len_abs": 1064, "len_title": 120, "len_sents": 23123, "len_sents_1": 21201, "n_sents": 176, "n_sents_1": 159, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 980, "title": "Fair Image Generation from Pre-trained Models by Probabilistic Modeling", "abs": "The production of high-fidelity images by generative models has been transformative to the space of artificial intelligence. Yet, while the generated images are of high quality, the images tend to mirror biases present in the dataset they are trained on. While there has been an influx of work to tackle this issue, existing works typically rely on fine-tuning an existing generative model which requires costly retraining time. In this paper, we use a family of tractable probabilistic models called probabilistic circuits (PCs), which can be equipped to a pre-trained generative model to produce fair images without fine-tuning. We show that for a given trained generative model, our method only requires a small fair reference dataset to train the PC, removing the need to retrain the generative model on a large dataset. Our experimental results show that the proposed method achieves a balance between training resources and ensuring fairness and quality of generated images.", "keywords": [ "Image Generation", "Fairness", "Probabilistic Modeling" ], "rating_list": [ 3, 3, 1 ], "soundness_list": [ 2, 2, 1 ], "presentation_list": [ 3, 1, 1 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "GVhfWu5L8D", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Honghao Wei", "gender": "Male", "institution": "Washington State University ", "country": "US", "position": "Assistant Professor" }, { "name": "Jiahui Zhu", "gender": "Male", "institution": "Washington State University at Pullman", "country": "US", "position": "PhD student" }, { "name": "Lei Ying", "gender": "Male", "institution": "University of Michigan, Ann Arbor", "country": "US", "position": "Professor" } ], "n_formula": 20, "n_formula_1": 17, "n_ref_uni": 69, "n_ref": 122, "n_ref_all": 168, "n_fig": 10, "n_tab": 5, "L_tab": 2736, "n_element_tab": 624, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1789, "formula_len_all_1": 1348, "len_all": 185169, "len_all_1": 66850, "len_abs": 1867, "len_title": 126, "len_sents": 49434, "len_sents_1": 31246, "n_sents": 361, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1910, "title": "Rethinking Behavior Regularization in Offline Safe RL: A Region-Based Approach", "abs": "Behavior regularization is a widely adopted technique in offline reinforcement learning (RL) to control distributional shift and mitigate extrapolation errors from out-of-distribution (OOD) actions by keeping the learned policy close to the behavior policy used to collect the dataset. However, directly applying behavior regularization to offline safe RL presents several issues. The optimal policy in safe RL should not only favor actions that prevent the agent from entering unsafe regions but also identify the shortest escape path when the agent finds itself in unsafe states. Enforcing safety and behavior regularization constraints simultaneously is inherently difficult and can often lead to infeasible solutions, especially when multiple constraints are involved. Furthermore, adding behavior regularization may cause the learned policy to imitate the behavior policy, even in states where the behavior policy performs poorly (not safe). This issue becomes particularly severe in offline safe RL, where the quality of the dataset collected by the behavior policy heavily impacts the learned policy\u2019s effectiveness. To address these challenges, we propose $\\textit{BARS}$ ($\\underline{B}$ehavior-$\\underline{A}$ware $\\underline{R}$egion-Based $\\underline{S}$afe offline RL), a novel algorithm that distinguishes between safe and unsafe states and applies region-specific, selective behavior regularization to optimize the policy. Extensive experiments show that BARS significantly outperforms several state-of-the-art baselines in terms of both rewards and safety, particularly in scenarios where the behavior policy is far from optimal. Notably, when dataset quality is low, BARS continues to perform well and ensure safety, while all other baselines fail to guarantee a safe policy in most of the environments. Our work has great potential to address a previously overlooked issue in offline safe RL.", "keywords": [ "Offline Reinforcement Learning; Safe Reinforcement Learning" ], "rating_list": [ 3, 3, 5, 5, 3 ], "soundness_list": [ 1, 2, 2, 3, 1 ], "presentation_list": [ 2, 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "GVNYi74t5L", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Bin Zhang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Chuyan Xiong", "gender": "Female", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Hongyu Wang", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Jialin Li", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Jiayu Xu", "gender": "Female", "institution": ", Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Ruiping Wang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Senwei Xie", "gender": "Male", "institution": "Institute of Computing Technology, CAS", "country": "", "position": "PhD student" }, { "name": "Zhaojie Xie", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Xilin Chen", "gender": "Male", "institution": "Institute of Computing Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 23, "n_ref": 69, "n_ref_all": 113, "n_fig": 22, "n_tab": 7, "L_tab": 2655, "n_element_tab": 249, "n_fig_1": 14, "n_tab_1": 6, "L_tab_1": 3466, "n_element_tab_1": 400, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 186031, "len_all_1": 65649, "len_abs": 3442, "len_title": 132, "len_sents": 51368, "len_sents_1": 28078, "n_sents": 415, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1547, "title": "M4U: Evaluating Multilingual Understanding and Reasoning for Large Multimodal Models", "abs": "Multilingual capability is an essential aspect for large multimodal models, since they are usually deployed across various countries and languages. However, most existing benchmarks for multilingual multimodal reasoning struggle to differentiate between models of varying performance; even language models without visual capabilities can easily achieve high scores. This leaves a comprehensive evaluation of leading multilingual multimodal models largely unexplored. In this work, we introduce M4U, a novel and challenging benchmark for assessing the capability of multi-discipline multilingual multimodal understanding and reasoning. M4U contains 8,931 samples covering 64 disciplines across 16 subfields in Science, Engineering, and Healthcare in Chinese, English, and German. Using M4U, we conduct extensive evaluations of 21 leading Large Multimodal Models (LMMs) and Large Language Models (LLMs) with external tools. The evaluation results show that the state-of-the-art model, GPT-4o, achieves only 47.6\\% average accuracy on M4U. Additionally, we observe that the leading LMMs exhibit significant language preferences. Our in-depth analysis indicates that leading LMMs, including GPT-4o, suffer performance degradation when prompted with cross-lingual multimodal questions, such as images with key textual information in Chinese while the question is in German. We believe that M4U can serve as a crucial tool for systematically evaluating LMMs based on their multilingual multimodal reasoning capabilities and monitoring their development.", "keywords": [ "Multilingual", "Multimodal Reasoning", "Large Multimodal Models" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "GVABHyvrRU", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Rishubh Parihar", "gender": "Male", "institution": "Indian Institute of Science, Bangalore", "country": "IN", "position": "PhD student" }, { "name": "Sabariswaran M", "gender": "Male", "institution": "Indian Institute of Technology, Kharagpur", "country": "IN", "position": "Undergrad student" }, { "name": "Sachidanand VS", "gender": "Male", "institution": "Indian Institute of Science, Indian institute of science, Bangalore", "country": "IN", "position": "Intern" }, { "name": "Venkatesh Babu R", "gender": "Male", "institution": "Indian Institute of Science", "country": "IN", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 23, "n_ref": 50, "n_ref_all": 74, "n_fig": 11, "n_tab": 2, "L_tab": 575, "n_element_tab": 58, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 400, "n_element_tab_1": 65, "formula_len_all": 980, "formula_len_all_1": 979, "len_all": 152833, "len_all_1": 65830, "len_abs": 1402, "len_title": 126, "len_sents": 43678, "len_sents_1": 29323, "n_sents": 322, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1410, "title": "Diffusion Compose: Compositional Depth Aware Scene Editing in Diffusion Models", "abs": "We introduce Diffusion Compose, a zero-shot approach for depth-aware scene editing using Text-to-Image diffusion models. While existing methods for 3D-aware editing focus on object-centric control, they do not support compositional depth-aware edits, such as placing objects at specific depths or combining multiple scenes realistically. We address this by incorporating depth-based multiplane scene representation in diffusion models. These planes, placed at fixed depths, can be individually edited or composed to enable 3D-aware scene modifications. However, direct manipulation of multiplane representation of diffusion latents often leads to identity loss or unrealistic blending. To overcome this, we propose a novel multiplane feature guidance technique that gradually aligns source latents with the target edit at each denoising step. We validate Diffusion Compose on two challenging tasks: a) scene composition, blending scenes with consistent depth order and scene illumination, and b) depth-aware object insertion, inserting novel objects at specified depths in a scene while preserving occlusions and scene structure and illumination. Extensive experiments demonstrate that Diffusion Compose significantly outperforms task-specific baselines for object placement and harmonization. A user study further confirms that it produces realistic, identity-preserving, and accurate depth-aware scene edits.", "keywords": [ "Image editing", "3D-aware scene control", "Text-to-Image Diffusion Models" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "GUd6zTrTBb", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Daming Shi", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "Full Professor" }, { "name": "Zunjin Zhao", "gender": "Male", "institution": "Shenzhen University", "country": "CN", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 12, "n_ref_uni": 31, "n_ref": 51, "n_ref_all": 81, "n_fig": 18, "n_tab": 4, "L_tab": 2227, "n_element_tab": 181, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 952, "n_element_tab_1": 96, "formula_len_all": 1682, "formula_len_all_1": 524, "len_all": 145195, "len_all_1": 56203, "len_abs": 1437, "len_title": 136, "len_sents": 36809, "len_sents_1": 25954, "n_sents": 308, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1443, "title": "Zero-shot Text-based Personalized Low-Light Image Enhancement with Reflectance Guidance", "abs": "Recent advances in zero-shot low-light image enhancement have largely benefited from the deep image priors encoded in network architectures. However, these models require optimization from scratch for each image and cannot provide personalized results based on user preferences.\nIn this paper, we propose a training-free zero-shot personalized low-light image enhancement model that integrates Retinex domain knowledge into a pre-trained diffusion model, enabling style personalization based on user preferences specified through text instructions. Our contributions are as follows:\nFirst, we incorporate the total variation optimization into a single Gaussian convolutional layer, enabling zero-shot Retinex decomposition. Second, we introduce the Contrastive Language-Image Pretraining (CLIP) model into the reflectance-conditioned sampling process of Denoising Diffusion Implicit Models (DDIM), guiding the enhancement according to user-provided text instructions. Third, to ensure consistency in content and structure, we employ patch-wise DDIM inversion to find the initial noise vector and use the reflectance as a condition during the reverse sampling process.\nOur proposed model, RetinexGDP, supports any image size and produces noise-suppressed results without imposing extra noise constraints. Extensive experiments across nine low-light image datasets show that RetinexGDP achieves performance comparable to state-of-the-art models.", "keywords": [ "Low-light image enhancement", "Retinex decomposition", "Zero-shot learning", "Generative diffusion prior" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "GURRWHkPtx", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Bo Long", "gender": "Male", "institution": "Meta", "country": "US", "position": "Principal Researcher" }, { "name": "Dongqi Fu", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Hanghang Tong", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Hanqing Zeng", "gender": "Male", "institution": "Meta AI", "country": "", "position": "Researcher" }, { "name": "Kaveh Hassani", "gender": "Male", "institution": "Meta", "country": "CA", "position": "AI Research Scientist" }, { "name": "Limei Wang", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Michihiro Yasunaga", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Ning Yao", "gender": "Female", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Si Zhang", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Zhe Xu", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 40, "n_ref": 100, "n_ref_all": 145, "n_fig": 5, "n_tab": 12, "L_tab": 3262, "n_element_tab": 195, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2068, "n_element_tab_1": 40, "formula_len_all": 862, "formula_len_all_1": 835, "len_all": 197810, "len_all_1": 65286, "len_abs": 1379, "len_title": 82, "len_sents": 41449, "len_sents_1": 27828, "n_sents": 321, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 34, "L_abs": 1385, "title": "Language Models are Graph Learners", "abs": "Language Models (LMs) are increasingly challenging the dominance of domain-specific models, including Graph Neural Networks (GNNs) and Graph Transformers (GTs), in graph learning tasks. Following this trend, we propose a novel approach that empowers off-the-shelf LMs to achieve performance comparable to state-of-the-art GNNs on node classification tasks, without requiring any architectural modification. By preserving the LM's original architecture, our approach retains a key benefit of LM instruction tuning: the ability to jointly train on diverse datasets, fostering greater flexibility and efficiency. To achieve this, we introduce two key augmentation strategies: (1) Enriching LMs' input using topological and semantic retrieval methods, which provide richer contextual information, and (2) guiding the LMs' classification process through a lightweight GNN classifier that effectively prunes class candidates. Our experiments on real-world datasets show that backbone Flan-T5 models equipped with these augmentation strategies outperform state-of-the-art text-output node classifiers and are comparable to top-performing vector-output node classifiers. By bridging the gap between specialized task-specific node classifiers and general LMs, this work paves the way for more versatile and widely applicable graph learning models. We will open-source the code upon publication.", "keywords": [ "Language model", "Data augmentation", "Node classification" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "GULx8rzzjC", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Arjun Nitin Bhagoji", "gender": "Male", "institution": "Indian Institute of Technology, Bombay", "country": "IN", "position": "Assistant Professor" }, { "name": "Nick Feamster", "gender": "Male", "institution": "University of Chicago", "country": "", "position": "Full Professor" }, { "name": "Supriyo Chakraborty", "gender": "unknown", "institution": "Capital One", "country": "US", "position": "Researcher" }, { "name": "Van Tran", "gender": "Female", "institution": "University of Chicago", "country": "", "position": "PhD student" }, { "name": "Zain Sarwar", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "Ben Zhao", "gender": "Male", "institution": "University of Chicago", "country": "", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 7, "n_ref_uni": 49, "n_ref": 67, "n_ref_all": 86, "n_fig": 7, "n_tab": 7, "L_tab": 1062, "n_element_tab": 171, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 946, "n_element_tab_1": 159, "formula_len_all": 2111, "formula_len_all_1": 742, "len_all": 200259, "len_all_1": 72519, "len_abs": 2618, "len_title": 122, "len_sents": 63192, "len_sents_1": 32397, "n_sents": 450, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1289, "title": "Mycroft: Towards Effective and Efficient External Data Augmentation", "abs": "Machine learning (ML) models often require large amounts of data to perform well. \n When the available data is limited, model trainers may need to acquire more data from external sources.\n Often, useful data is held by private entities who are hesitant to share their data due to propriety and privacy concerns. \n This makes it challenging and expensive for model trainers to acquire the data they need to improve model performance.\n To address this challenge, we propose $\\texttt{Mycroft}$, a \n data-efficient method that enables model trainers to evaluate the relative \n utility of different data sources while working with a constrained data-sharing \n budget. By leveraging feature space distances and gradient matching, $\\texttt{Mycroft}$ \n identifies small but informative data subsets from each owner, allowing model \n trainers to maximize performance with minimal data exposure. Experimental \n results across four tasks in two domains show that $\\texttt{Mycroft}$ converges rapidly \n to the performance of the full-information baseline, where all data is shared.\n Moreover, $\\texttt{Mycroft}$ is robust to noise and can effectively rank data owners by \n utility. $\\texttt{Mycroft}$ can pave the way for democratized training of high performance ML models.", "keywords": [ "data scarcity; supervised learning; data augmentation" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "GTzP2GC7NR", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Bhavya Kailkhura", "gender": "Male", "institution": "Lawrence Livermore National Laboratory", "country": "US", "position": "Research Staff" }, { "name": "Gourav Datta", "gender": "Male", "institution": "Case Western Reserve University", "country": "US", "position": "Assistant Professor" }, { "name": "James Diffenderfer", "gender": "unknown", "institution": "Lawrence Livermore National Labs", "country": "US", "position": "Researcher" }, { "name": "Peter Anthony Beerel", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Full Professor" }, { "name": "Zeyu Liu", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 18, "n_ref_uni": 21, "n_ref": 40, "n_ref_all": 77, "n_fig": 9, "n_tab": 16, "L_tab": 3041, "n_element_tab": 261, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1700, "n_element_tab_1": 242, "formula_len_all": 1334, "formula_len_all_1": 1080, "len_all": 186678, "len_all_1": 68445, "len_abs": 1207, "len_title": 128, "len_sents": 48869, "len_sents_1": 28990, "n_sents": 335, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1231, "title": "When SNN meets ANN: Error-Free ANN-to-SNN Conversion for Extreme Edge Efficiency", "abs": "Spiking Neural Networks (SNN) are now demonstrating comparable accuracy to convolutional neural networks (CNN), thanks to advanced ANN-to-SNN conversion techniques, all while delivering remarkable energy and latency efficiency when deployed on neuromorphic hardware. However, these conversion techniques incur a large number of time steps, and consequently, high spiking activity. In this paper, we propose a novel ANN-to-SNN conversion framework, that incurs an exponentially lower number of time steps compared to that required in the existing conversion approaches. Our framework modifies the standard integrate-and-fire (IF) neuron model used in SNNs with no change in computational complexity and shifts the bias term of each batch normalization (BN) layer in the trained ANN. To reduce spiking activity, we propose training the source ANN with a fine-grained $\\ell_1$ regularizer with surrogate gradients that encourages high spike sparsity in the converted SNN. Our proposed framework thus yields lossless SNNs with low latency, low compute energy, thanks to the low time steps and high spike sparsity, and high test accuracy, for example, 75.12% with only 4 time steps on the ImageNet dataset. Codes will be made available.", "keywords": [ "SNN", "ANN-to-SNN conversion", "IF model", "ImageNet", "spiking activity" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "GThTiuXgDC", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Liqiang Nie", "gender": "Male", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "dean" }, { "name": "Qi Lv", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Xiang Deng", "gender": "Male", "institution": "State University of New York, Binghamton", "country": "US", "position": "PhD student" }, { "name": "Xiaoyang Chen", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" }, { "name": "Yinchuan Li", "gender": "Male", "institution": "Huawei Noah's Ark Lab (AI Lab)", "country": "CN", "position": "Principal Researcher" }, { "name": "ChuHengshuo", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 46, "n_ref": 82, "n_ref_all": 100, "n_fig": 6, "n_tab": 10, "L_tab": 1582, "n_element_tab": 180, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 526, "n_element_tab_1": 15, "formula_len_all": 685, "formula_len_all_1": 666, "len_all": 129518, "len_all_1": 60358, "len_abs": 1760, "len_title": 154, "len_sents": 32746, "len_sents_1": 27525, "n_sents": 249, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1801, "title": "3D-AffordanceLLM: Harnessing Large Language Models for Open-Vocabulary Affordance Detection in 3D Worlds", "abs": "3D Affordance detection is a challenging problem with broad applications on various robotic tasks. \nExisting methods typically formulate the detection paradigm as a label-based semantic segmentation task.\nThis paradigm relies on predefined labels and lacks the ability to comprehend complex natural language, resulting in limited generalization in open-world scene.\nTo address these limitations, we reformulate the traditional affordance detection paradigm into \\textit{Instruction Reasoning Affordance Segmentation} (IRAS) task. \nThis task is designed to output a affordance mask region given a query reasoning text, which avoids fixed categories of input labels.\nWe accordingly propose the \\textit{3D-AffordanceLLM} (3D-ADLLM), a framework designed for reasoning affordance detection in 3D open-scene.\nSpecifically, 3D-ADLLM introduces large language models (LLMs) to 3D affordance perception with a custom-designed decoder for generating affordance masks, thus achieving open-world reasoning affordance detection.\nIn addition, given the scarcity of 3D affordance datasets for training large models, we seek to extract knowledge from general segmentation data and transfer it to affordance detection.\nThus, we propose a multi-stage training strategy that begins with a novel pre-training task, i.e., \\textit{Referring Object Part Segmentation}~(ROPS).\nThis stage is designed to equip the model with general recognition and segmentation capabilities at the object-part level.\nThen followed by fine-tuning with the IRAS task, 3D-ADLLM obtains the reasoning ability for affordance detection. \nIn summary, 3D-ADLLM leverages the rich world knowledge and human-object interaction reasoning ability of LLMs, achieving approximately an 8\\% improvement in mIoU on open-vocabulary affordance detection tasks.", "keywords": [ "3D.+Large Language Model.+Robot.+Affordance.+Perception" ], "rating_list": [ 6, 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "GTe9PDhm8v", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Priyadarshini Panda", "gender": "Female", "institution": "Yale University", "country": "US", "position": "Assistant Professor" }, { "name": "Yuhang Li", "gender": "Male", "institution": "Yale University", "country": "US", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 25, "n_ref": 75, "n_ref_all": 112, "n_fig": 4, "n_tab": 9, "L_tab": 4032, "n_element_tab": 600, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2821, "n_element_tab_1": 484, "formula_len_all": 856, "formula_len_all_1": 565, "len_all": 151654, "len_all_1": 66260, "len_abs": 1562, "len_title": 128, "len_sents": 45302, "len_sents_1": 26255, "n_sents": 355, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1514, "title": "TesseraQ: Ultra Low-Bit LLM Post-Training Quantization with Block Reconstruction", "abs": "Large language models (LLMs) have revolutionized natural language processing, albeit at the cost of immense memory and computation requirements. Post-training quantization (PTQ) is becoming the \\emph{de facto} method to reduce the memory footprint and improve the inference throughput of LLMs.\nIn this work, we aim to push the upper limit of LLM PTQ by optimizing the weight rounding parameters with the block reconstruction technique, a predominant method in previous vision models.\nWe propose TesseraQ, a new state-of-the-art PTQ technique, to quantize the weights of LLMs to ultra-low bits.\nTo effectively optimize the rounding in LLMs and stabilize the reconstruction process, we introduce progressive adaptive rounding. This approach iteratively transits the soft rounding variables to hard variables during the reconstruction process. Additionally, we optimize the dequantization scale parameters to fully leverage the block reconstruction technique.\nWe demonstrate that TesseraQ can be seamlessly integrated with existing scaling or clipping-based PTQ algorithms such as AWQ and OmniQuant, significantly enhancing their performance and establishing a new state-of-the-art.\nFor instance, when compared to AWQ, TesseraQ improves the wikitext2 perplexity from 14.65 to 6.82 and average downstream accuracy from 50.52 to 59.27 with 2-bit weight-only quantization of LLaMA-2-7B. \nAcross a range of quantization schemes, including W2A16, W3A16, W3A3, and W4A4, TesseraQ consistently exhibits superior performance.", "keywords": [ "LLM", "Quantization", "Efficient Inference" ], "rating_list": [ 3, 5, 6, 6, 5 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 4, 3, 2 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 3, 5, 3 ] }, { "paper_id": "GTcEe5fayC", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 11, "author_info_list": [ { "name": "Ao Li", "gender": "Male", "institution": "University of Arizona", "country": "US", "position": "Assistant Professor" }, { "name": "Geng Yuan", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Assistant Professor" }, { "name": "Jun Liu", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Qitao Tan", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" }, { "name": "Sheng Li", "gender": "Male", "institution": "University of Pittsburgh", "country": "US", "position": "PhD student" }, { "name": "Tianyu Wang", "gender": "Male", "institution": "University of Pittsburgh", "country": "US", "position": "PhD student" }, { "name": "Xulong Tang", "gender": "Male", "institution": "University of Pittsburgh", "country": "", "position": "Assistant Professor" }, { "name": "Yue Dai", "gender": "unknown", "institution": "University of Pittsburgh", "country": "", "position": "PhD student" }, { "name": "Zhenglun Kong", "gender": "Male", "institution": "Harvard Medical School, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Ninghao Liu", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "Assistant Professor" }, { "name": "Yufei Ding", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Associate Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 47, "n_ref": 93, "n_ref_all": 118, "n_fig": 7, "n_tab": 13, "L_tab": 2770, "n_element_tab": 356, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 2200, "n_element_tab_1": 275, "formula_len_all": 46, "formula_len_all_1": 46, "len_all": 162113, "len_all_1": 75409, "len_abs": 1459, "len_title": 162, "len_sents": 47035, "len_sents_1": 34133, "n_sents": 338, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 114, "L_abs": 1467, "title": "Mutual Effort for Efficiency: A Similarity-based Token Pruning for Vision Transformers in Self-Supervised Learning", "abs": "Self-supervised learning (SSL) offers a compelling solution to the challenge of extensive labeled data requirements in traditional supervised learning.\nWith the proven success of Vision Transformers (ViTs) in supervised tasks, there is increasing interest in adapting them for SSL frameworks. However, the high computational demands of SSL pose substantial challenges, particularly on resource-limited platforms like edge devices, despite its ability to achieve high accuracy without labeled data.\nRecent studies in supervised learning have shown that token pruning can reduce training costs by removing less informative tokens without compromising accuracy. However, SSL\u2019s dual-branch encoders make traditional single-branch pruning strategies less effective, as they fail to account for the critical cross-branch similarity information, leading to reduced accuracy in SSL.\nTo this end, we introduce SimPrune, a novel token pruning strategy designed for ViTs in SSL. SimPrune leverages cross-branch similarity information to efficiently prune tokens, retaining essential semantic information across dual branches. Additionally, we incorporate a difficulty-aware pruning strategy to further enhance SimPrune's effectiveness.\nExperimental results show that our proposed approach effectively reduces training computation while maintaining accuracy. Specifically, our approach offers 24\\% savings in training costs compared to SSL baseline, without sacrificing accuracy.", "keywords": [ "Efficient Self-Supervised Learning", "Vision Transformer", "Token Pruning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 2 ] }, { "paper_id": "GT4gMdvVFp", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Arman Cohan", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" }, { "name": "Dennis Shasha", "gender": "Male", "institution": ", New York University", "country": "US", "position": "Full Professor" }, { "name": "Tintin Jiang", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Yilun Zhao", "gender": "unknown", "institution": "Yale University", "country": "US", "position": "PhD student" }, { "name": "Yitao Long", "gender": "unknown", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 41, "n_ref": 104, "n_ref_all": 137, "n_fig": 10, "n_tab": 3, "L_tab": 1616, "n_element_tab": 93, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2378, "n_element_tab_1": 134, "formula_len_all": 305, "formula_len_all_1": 25, "len_all": 189396, "len_all_1": 64301, "len_abs": 1867, "len_title": 146, "len_sents": 59029, "len_sents_1": 28653, "n_sents": 536, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1836, "title": "PuzzlePlex: A Benchmark to Evaluate the Reasoning and Planning of Large Language Models on Puzzles", "abs": "Large Language Models (LLMs) have demonstrated remarkable performance in various tasks, yet their comprehensive reasoning and planning capabilities in interactive environments remain underexplored. We introduce PuzzlePlex, a benchmark designed to evaluate reasoning and planning capabilities in a multi-turn adversarial environment. \nPuzzlePlex comprises 24 diverse puzzles, including deterministic and stochastic games, as well as single-player and adversarial scenarios. An important novelty of our benchmark is that it includes multi-step adversarial reasoning games. To succeed in such games, each LLM must maintain a history of its own moves and those of the opponent LLM, generating strategies that outperform the opponent to secure victory.\nWe implement customized game-playing strategies (such as dynamic programming approaches) for comparison. \nOur findings indicate that the reasoning and planning abilities of current LLMs are currently poor in puzzle-solving contexts. GPT-4 outperforms other models, successfully competing against customized strategies (such as greedy approaches or dynamic programming) in 49% of cases. However, when faced with strict rule sets, it demonstrates diminished reasoning and planning capabilities. In addition to the 14 multi-turn adversarial puzzles, we report on single-player puzzles and incorporate multi-modal challenges that integrate text and images, revealing that LLMs still significantly lag behind even simple heuristics in puzzles.\nA key feature of our benchmark is its ability to generate game instances with graduated levels of difficulty, allowing it to evolve as LLMs become more sophisticated. This adaptability ensures the continued relevance and utility of PuzzlePlex in assessing the progress of LLM capabilities in reasoning and planning within interactive environments.", "keywords": [ "Benchmark", "Puzzle", "Reasoning and Planning" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "GSrs4vIqiF", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Daeho Um", "gender": "Male", "institution": "Samsung", "country": "KR", "position": "Researcher" }, { "name": "Ji Won Yoon", "gender": "Female", "institution": "Chung-Ang University", "country": "KR", "position": "Assistant Professor" }, { "name": "Jongin Lim", "gender": "Male", "institution": "Samsung Advanced Institute of Technology", "country": "KR", "position": "Researcher" }, { "name": "Seong Jin Ahn", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Sunoh Kim", "gender": "Male", "institution": "Samsung", "country": "KR", "position": "Researcher" }, { "name": "Yunha Yeo", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 8, "n_ref_uni": 29, "n_ref": 53, "n_ref_all": 73, "n_fig": 5, "n_tab": 12, "L_tab": 6724, "n_element_tab": 593, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 3100, "n_element_tab_1": 373, "formula_len_all": 375, "formula_len_all_1": 301, "len_all": 120972, "len_all_1": 57905, "len_abs": 1104, "len_title": 141, "len_sents": 34371, "len_sents_1": 23133, "n_sents": 260, "n_sents_1": 160, "venue": "ICLR2025", "decision": "withdraw", "L_title": 93, "L_abs": 1108, "title": "GITD: Enhancing Medical Classification on Tabular Data with Missing Values via Graph Modeling", "abs": "With the advancement of machine learning, various techniques have been developed to classify patients for disease diagnosis using medical tabular data. Due to the presence of missing values in the medical tabular data, these techniques commonly impute the missing values before applying classifiers. However, most existing techniques classify patients solely based on each patient's individual features despite the advantages of leveraging patients with similar features that can enhance both imputation and classification. To address this issue, we introduce graph data imputation for tabular data (GITD), a novel approach that constructs feature-attentive k-nearest neighbor (kNN) graphs to enable the use of graph data imputation methods on medical tabular data. The key idea of GITD is constructing a kNN graph among patients by prioritizing important features for classification. Our extensive experimental results demonstrate that GITD successfully bridges graph data imputation methods and medical tabular classification, achieving state-of-the-art performance across various medical tabular datasets.", "keywords": [ "medical classification", "tabular data", "imputation" ], "rating_list": [ 6, 5, 5, 5, 3 ], "soundness_list": [ 4, 2, 2, 2, 1 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 1, 2 ], "confidence_list": [ 3, 4, 5, 4, 4 ] }, { "paper_id": "GSckuQMzBG", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Andrew I. Comport", "gender": "Male", "institution": "CNRS", "country": "", "position": "Principal Researcher" }, { "name": "Antoine Schnepf", "gender": "Male", "institution": "Criteo", "country": "", "position": "PhD student" }, { "name": "Flavian Vasile", "gender": "Male", "institution": "Criteo", "country": "", "position": "Researcher" }, { "name": "Jean-Yves Franceschi", "gender": "Male", "institution": "Criteo", "country": "FR", "position": "Researcher" }, { "name": "Jeremie Mary", "gender": "Male", "institution": "Criteo", "country": "", "position": "Senior Researcher" }, { "name": "Karim Kassab", "gender": "Male", "institution": "Ecole Nationale des Sciences G\u00e9ographiques", "country": "", "position": "PhD student" }, { "name": "Laurent Caraffa", "gender": "Male", "institution": "Ecole Nationale des Sciences G\u00e9ographiques", "country": "FR", "position": "Researcher" }, { "name": "Valerie Gouet-Brunet", "gender": "Female", "institution": "LaSTIG / IGN, Univ. Gustave Eiffel", "country": "", "position": "Principal Researcher" } ], "n_formula": 14, "n_formula_1": 10, "n_ref_uni": 32, "n_ref": 41, "n_ref_all": 53, "n_fig": 5, "n_tab": 6, "L_tab": 1184, "n_element_tab": 189, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2411, "n_element_tab_1": 106, "formula_len_all": 672, "formula_len_all_1": 505, "len_all": 103170, "len_all_1": 53785, "len_abs": 1246, "len_title": 117, "len_sents": 26588, "len_sents_1": 22522, "n_sents": 218, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1250, "title": "Scaled Inverse Graphics: Efficiently Learning Large Sets of 3D Scenes", "abs": "While the field of inverse graphics has been witnessing continuous growth, techniques devised thus far predominantly focus on learning individual scene representations.\nIn contrast, learning large sets of scenes has been a considerable bottleneck in NeRF developments, as repeatedly applying inverse graphics on a sequence of scenes, though essential for various applications, remains largely prohibitive in terms of resource costs.\nWe introduce a framework termed \"scaled inverse graphics\", aimed at efficiently learning large sets of scene representations, and propose a novel method to this end.\nIt operates in two stages: (i) training a compression model on a subset of scenes, then (ii) training NeRF models on the resulting smaller representations, thereby reducing the optimization space per new scene.\nIn practice, we compact the representation of scenes by learning NeRFs in a latent space to reduce the image resolution, and sharing information across scenes to reduce NeRF representation complexity.\nWe experimentally show that our method presents both the lowest training time and memory footprint in scaled inverse graphics compared to other methods applied independently on each scene.\nOur codebase is publicly available as open-source.", "keywords": [ "Latent NeRF", "Tri-Planes", "autoencoder", "inverse graphics", "3D" ], "rating_list": [ 3, 5, 5, 1, 1 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 1, 1 ], "confidence_list": [ 3, 5, 3, 5, 4 ] }, { "paper_id": "GSUNPIw7Ad", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Alessandro Gnutti", "gender": "Male", "institution": "University of Brescia", "country": "IT", "position": "Assistant Professor" }, { "name": "Chia-Hao Kao", "gender": "Male", "institution": "University of Brescia", "country": "IT", "position": "Researcher" }, { "name": "Riccardo Leonardi", "gender": "Male", "institution": "Universit\u00e0 degli Studi di Roma \"Tor Vergata\"", "country": "IT", "position": "Full Professor" }, { "name": "Shao-Yuan Lo", "gender": "unknown", "institution": "Honda Research Institute USA", "country": "US", "position": "Researcher" }, { "name": "Wen-Hsiao Peng", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "Full Professor" }, { "name": "Yi-Hsin Chen", "gender": "Female", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "PhD student" }, { "name": "Yu-Jen Tseng", "gender": "unknown", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" }, { "name": "CHENG CHIEN", "gender": "Male", "institution": "National Yang Ming Chiao Tung University", "country": "TW", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 23, "n_ref": 83, "n_ref_all": 125, "n_fig": 14, "n_tab": 3, "L_tab": 248, "n_element_tab": 24, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 248, "n_element_tab_1": 24, "formula_len_all": 372, "formula_len_all_1": 370, "len_all": 133220, "len_all_1": 62765, "len_abs": 1531, "len_title": 119, "len_sents": 38343, "len_sents_1": 30999, "n_sents": 263, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1539, "title": "Bridging Compressed Image Latents and Multimodal Large Language Models", "abs": "This paper presents the first-ever study of adapting compressed image latents to suit the needs of downstream vision tasks that adopt Multimodal Large Language Models (MLLMs). MLLMs have extended the success of large language models to modalities (e.g. images) beyond text, but their billion scale hinders deployment on resource-constrained end devices. While cloud-hosted MLLMs could be available, transmitting raw, uncompressed images captured by end devices to the cloud requires an efficient image compression system. To address this, we focus on emerging neural image compression and propose a novel framework with a lightweight transform-neck and a surrogate loss to adapt compressed image latents for MLLM-based vision tasks. \nGiven the huge scale of MLLMs, our framework excludes the entire downstream MLLM except part of its visual encoder from training our system. This stands out from most existing coding for machine approaches that involve downstream networks in training and thus could be impractical when the networks are MLLMs. The proposed framework is general in that it is applicable to various MLLMs, neural image codecs, and multiple application scenarios, where the neural image codec can be (1) pre-trained for human perception without updating, (2) fully updated for joint human and machine perception, or (3) fully updated for only machine perception. \nExtensive experiments on different neural image codecs and various MLLMs show that our method achieves great rate-accuracy performance with much less complexity.", "keywords": [ "Multimodal Large Language Models", "Neural Image Compression" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "GRXlkYg7Nz", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Chao Wang", "gender": "Female", "institution": "Zhejiang Lab", "country": "CN", "position": "Researcher" }, { "name": "Chengwei Zhang", "gender": "Male", "institution": "Dalian Martime University", "country": "CN", "position": "Associate Professor" }, { "name": "Edward Chung", "gender": "unknown", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Yang Zhang", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Postdoc" }, { "name": "Yunjian Xu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "\u6ce2 \u5510", "gender": "Male", "institution": "Institute for Advanced Algorithms Research, Shanghai", "country": "", "position": "Researcher" }, { "name": "Zhihe YANG", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 45, "n_formula_1": 23, "n_ref_uni": 27, "n_ref": 67, "n_ref_all": 96, "n_fig": 10, "n_tab": 7, "L_tab": 2722, "n_element_tab": 164, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 933, "n_element_tab_1": 92, "formula_len_all": 5568, "formula_len_all_1": 1940, "len_all": 140642, "len_all_1": 70902, "len_abs": 1313, "len_title": 123, "len_sents": 42417, "len_sents_1": 29420, "n_sents": 345, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1330, "title": "Policy Consistency in Multi-Agent Reinforcement Learning with Mixed Reward", "abs": "The sparsity of team rewards poses a significant challenge that hinders the effective learning of optimal team policies in cooperative multi-agent reinforcement learning. One common approach to mitigate this issue involves augmenting sparse rewards with individual rewards to guide policy training. However, a significant drawback of such approaches is that modifying the reward function can potentially alter the optimal policy. To tackle this challenge, we propose a novel multi-agent policy optimization approach that ensures consistency between the mixed policy (learned from a combination of individual and team rewards) and the team policy (based solely on team rewards), through a new policy consistency constraint that aligns the returns of both policies in policy optimization model. We further develop an iterated policy optimization procedure to solve the formulated problem, deriving an approximate optimization objective for each iteration of the mixed and team policies. Experimental evaluation conducted in the StarCraft II Multi-Agent Challenge Environment (SMAC), Multi-Agent Particle Environment (MPE), and Google Research Football (GRF) environments demonstrate that our proposed approach effectively addresses the policy inconsistency problem, ${\\it i.e.}$, it consistently outperforms strong baseline methods.", "keywords": [ "Multi-agent System", "Reinforcement Learning", "Sparse Reward", "Policy Consistency", "Individual Reward" ], "rating_list": [ 8, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "GRMfXcAAFh", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Daniela Rus", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "T. Konstantin Rusch", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" } ], "n_formula": 36, "n_formula_1": 25, "n_ref_uni": 36, "n_ref": 80, "n_ref_all": 107, "n_fig": 3, "n_tab": 3, "L_tab": 2079, "n_element_tab": 228, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3029, "formula_len_all_1": 1938, "len_all": 160860, "len_all_1": 66546, "len_abs": 1438, "len_title": 78, "len_sents": 50844, "len_sents_1": 29353, "n_sents": 399, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 30, "L_abs": 1445, "title": "Oscillatory State-Space Models", "abs": "We propose Linear Oscillatory State-Space models (LinOSS) for efficiently learning on long sequences. Inspired by cortical dynamics of biological neural networks, we base our proposed LinOSS model on a system of forced harmonic oscillators. A stable discretization, integrated over time using fast associative parallel scans, yields the proposed state-space model. We prove that LinOSS produces stable dynamics only requiring nonnegative diagonal state matrix. This is in stark contrast to many previous state-space models relying heavily on restrictive parameterizations. Moreover, we rigorously show that LinOSS is universal, i.e., it can approximate any continuous and causal operator mapping between time-varying functions, to desired accuracy. In addition, we show that an implicit-explicit discretization of LinOSS perfectly conserves the symmetry of time reversibility of the underlying dynamics. Together, these properties enable efficient modeling of long-range interactions, while ensuring stable and accurate long-horizon forecasting. Finally, our empirical results, spanning a wide range of time-series tasks from mid-range to very long-range classification and regression, as well as long-horizon forecasting, demonstrate that our proposed LinOSS model consistently outperforms state-of-the-art sequence models. Notably, LinOSS outperforms Mamba by nearly 2x and LRU by 2.5x on a sequence modeling task with sequences of length 50k.", "keywords": [ "state-space models", "sequence models", "long-range interactions", "oscillators", "time-series" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 4, 3 ], "confidence_list": [ 3, 4, 2, 5 ] }, { "paper_id": "GR0y0F3Ipd", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Hongning Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "JinZhou", "gender": "unknown", "institution": ", Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Minlie Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xinjie Yu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xujun Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yadi Liu", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhe Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Erler Zhu", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 84, "n_ref_all": 116, "n_fig": 17, "n_tab": 7, "L_tab": 1268, "n_element_tab": 150, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 578, "n_element_tab_1": 102, "formula_len_all": 799, "formula_len_all_1": 428, "len_all": 153780, "len_all_1": 67096, "len_abs": 1491, "len_title": 70, "len_sents": 46840, "len_sents_1": 32260, "n_sents": 340, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1526, "title": "MAPS: Advancing Multi-Modal Reasoning in Expert-Level Physical Science", "abs": "Pre-trained on extensive text and image corpora, current Multi-Modal Large Language Models (MLLM) have shown strong capabilities in general visual reasoning tasks. \nHowever, their performance is still lacking in physical domains that require understanding diagrams with complex physical structures and quantitative analysis based on multi-modal information. \nTo address this, we develop a new framework, named **M**ulti-Modal Scientific Re**A**soning with **P**hysics Perception and **S**imulation (**MAPS**) based on an MLLM. \nMAPS decomposes expert-level multi-modal reasoning task into physical diagram understanding via a Physical Perception Model (PPM) and reasoning with physical knowledge via a simulator. \nThe PPM module is obtained by fine-tuning a visual language model using carefully designed synthetic data with paired physical diagrams and corresponding simulation language descriptions. \nAt the inference stage, MAPS integrates the simulation language description of the input diagram provided by PPM and results obtained through a Chain-of-Simulation process with MLLM to derive the underlying rationale and the final answer. \nValidated using our collected college-level circuit analysis problems, MAPS significantly improves reasoning accuracy of MLLM and outperforms all existing models. \nThe results confirm MAPS offers a promising direction for enhancing multi-modal scientific reasoning ability of MLLMs. \nWe will release our code, model and dataset used for our experiments upon publishing of this paper.", "keywords": [ "multi-modal reasoning", "scientific reasoning", "physical simulation" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "GQnR7L6SmA", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Dong Ju Mun", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "PhD student" }, { "name": "Gwanghyun Kim", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Haechang Lee", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Hyunwoo Lee", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "MS student" }, { "name": "Kyung Chul Lee", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "PhD student" }, { "name": "Seongmin Hong", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Seung Ah Lee", "gender": "unknown", "institution": "Yonsei University", "country": "KR", "position": "Associate Professor" }, { "name": "Se Young Chun", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" } ], "n_formula": 57, "n_formula_1": 18, "n_ref_uni": 62, "n_ref": 159, "n_ref_all": 223, "n_fig": 24, "n_tab": 10, "L_tab": 3905, "n_element_tab": 126, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 62, "n_element_tab_1": 12, "formula_len_all": 2522, "formula_len_all_1": 761, "len_all": 233832, "len_all_1": 60438, "len_abs": 1224, "len_title": 136, "len_sents": 72306, "len_sents_1": 27233, "n_sents": 565, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1065, "title": "Masked, Regularized Fidelity With Diffusion Models For Highly Ill-posed Inverse Problems", "abs": "Diffusion models have been well-investigated for solving ill-posed inverse problems to yield excellent performance. However, their application to highly ill-posed inverse problems remains challenging. In this work, we propose zero-shot diffusion model for large and complex kernels, dubbed Dilack, incorporating novel data fidelity terms. Based on our analyses on the ill-posedness for challenging inverse problems, we propose *regularized fidelity* called pseudo-inverse anchor for constraining (PiAC) fidelity loss. Inspired by locally acting classical regularizers, we also propose to incorporate *masked fidelity* within PiAC loss that can interact with globally acting diffusion models, which adaptively enforces spatially and step-wisely local fidelity via masks. Our proposed scheme effectively reduces erratic behavior and inherent artifacts in diffusion models, thereby improving restoration quality including perceptual aspects and outperforming prior arts on both synthetic and real-world datasets for modern lensless imaging and large motion deblurring.", "keywords": [ "Diffusion model", "Image restoration", "Large and complex kernel", "Severe ill-posed problem" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "GQhlM0Mavg", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Joseba Dalmau", "gender": "Male", "institution": "IRT Saint Exup\u00e9ry", "country": "FR", "position": "Researcher" }, { "name": "Paul Novello", "gender": "Male", "institution": "IRT Saint Exupery", "country": "FR", "position": "Researcher" }, { "name": "Leo Andeol", "gender": "Male", "institution": "Institut de Math\u00e9matique de Toulouse", "country": "FR", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 11, "n_ref_uni": 22, "n_ref": 51, "n_ref_all": 73, "n_fig": 2, "n_tab": 10, "L_tab": 9054, "n_element_tab": 1423, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 3158, "n_element_tab_1": 164, "formula_len_all": 1147, "formula_len_all_1": 597, "len_all": 155152, "len_all_1": 68315, "len_abs": 1279, "len_title": 116, "len_sents": 41147, "len_sents_1": 29337, "n_sents": 300, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 116, "L_abs": 1324, "title": "Exploring the Link Between Out-of-Distribution Detection and Conformal Prediction with Illustrations of Its Benefits", "abs": "Research on Out-Of-Distribution (OOD) detection focuses mainly on building scores that efficiently distinguish OOD data from In Distribution (ID) data. \n On the other hand, Conformal Prediction (CP) uses non-conformity scores to construct prediction sets with probabilistic coverage guarantees. In other words, the former designs scores, while the latter designs probabilistic guarantees based on scores. Therefore, we claim that these two fields might be naturally intertwined. \n This work advocates for cross-fertilization between OOD and CP by formalizing their link and emphasizing two benefits of using them jointly.\n First, we show that in standard OOD benchmark settings, evaluation metrics can be overly optimistic due to the test dataset's finite sample size.\n Based on the work of (Bates et al, 2022), we define new *conformal AUROC* and *conformal FRP@TPR$\\beta$* metrics, \n which are corrections that provide probabilistic conservativeness guarantees on the variability of these metrics.\n We show the effect of these corrections on two reference OOD and anomaly detection benchmarks, OpenOOD (Yang et al, 2022) and ADBench (Han et al. 2022). \n Second, we explore using OOD scores as non-conformity scores and show that they can improve the efficiency of the prediction sets obtained with CP.", "keywords": [ "Out-of-distribution detection", "Conformal Prediction", "benchmark", "nonconformity scores" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "GQgPj1H4pO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chanyoung Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Donghyun Kim", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Assistant Professor" }, { "name": "Jaehyeong Jeon", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Jinyoung Moon", "gender": "Female", "institution": "University of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Kanghoon Yoon", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Kibum Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Yeonjun In", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 44, "n_ref": 125, "n_ref_all": 172, "n_fig": 14, "n_tab": 13, "L_tab": 2400, "n_element_tab": 294, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2359, "n_element_tab_1": 140, "formula_len_all": 161, "formula_len_all_1": 183, "len_all": 213038, "len_all_1": 82080, "len_abs": 2151, "len_title": 127, "len_sents": 71853, "len_sents_1": 37195, "n_sents": 438, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1810, "title": "Weakly Supervised Video Scene Graph Generation via Natural Language Supervision", "abs": "Existing Video Scene Graph Generation (VidSGG) studies are trained in a fully supervised manner, which requires all frames in a video to be annotated, thereby incurring high annotation cost compared to Image Scene Graph Generation (ImgSGG). Although the annotation cost of VidSGG can be alleviated by adopting a weakly supervised approach commonly used for ImgSGG (WS-ImgSGG) that uses image captions, there are two key reasons that hinder such a naive adoption: 1) Temporality within video captions, i.e., unlike image captions, video captions include temporal markers (e.g., before, while, then, after) that indicate time-related details, and 2) Variability in action duration, i.e., unlike human actions in image captions, human actions in video captions unfold over varying duration. To address these issues, we propose a weakly supervised VidSGG with Natural Language Supervision (VSNLS) framework that only utilizes the readily available video captions for training a VidSGG model. VSNLS consists of two key modules: Temporality-aware Caption Segmentation (TCS) module and Action Duration Variability-aware caption-frame alignment (ADV) module. Specifically, TCS segments the video captions into multiple sentences in a temporal order based on a Large Language Model (LLM), and ADV aligns each segmented sentence with appropriate frames considering the variability in action duration. Our approach leads to a significant enhancement in performance compared to simply applying the WS-ImgSGG pipeline to VidSGG on the Action Genome dataset. As a further benefit of utilizing the video captions as weak supervision, we show that the VidSGG model trained by VSNLS is able to predict a broader range of action classes that are not included in the training data, which makes our framework practical in reality.", "keywords": [ "Video Scene Understanding", "Weakly Supervised Learning", "Large Language Model" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "GQ2Ks23bJ6", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Linlian Jiang", "gender": "Female", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "Peng Zheng", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "Rui Ma", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Associate Professor" }, { "name": "Xinxin Zuo", "gender": "Female", "institution": "Concordia University", "country": "CA", "position": "Assistant Professor" }, { "name": "Yang Wang", "gender": "Male", "institution": "Concordia University", "country": "CA", "position": "Associate Professor" }, { "name": "Yibo Zhang", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 34, "n_ref": 57, "n_ref_all": 75, "n_fig": 7, "n_tab": 3, "L_tab": 408, "n_element_tab": 48, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 428, "n_element_tab_1": 40, "formula_len_all": 840, "formula_len_all_1": 799, "len_all": 113031, "len_all_1": 59107, "len_abs": 1758, "len_title": 138, "len_sents": 30895, "len_sents_1": 26759, "n_sents": 199, "n_sents_1": 175, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1766, "title": "HierT2S: Enhancing Part-Level Text-to-Shape Generation via Hierarchical Structure Modeling", "abs": "Text-driven 3D shape generation still faces key challenges, especially in achieving high levels of control over the generated outputs. Paticularly, existing text-to-shape methods ignore the explicit modeling of hierarchical structures in the text and 3D shapes, which makes it hard for using long text descriptions with multiple prompts to guide the coherent part-level 3D shape generation. In this work, we introduce HierT2S, a framework that integrates a hierarchical tree representation with a conditional diffusion model, to enhance the generation of 3D shapes with coherent structures induced by the hierarchical and structured text representations. The key idea is to first segment the input text into several clusters and construct a hierarchical tree representation, with each node representing a parent entity or the fine-level part components. Then, we process the lower-level clusters of the tree with a relation graph module which uses self-attention mechanism to aggregate the relationships of the clusters, and generate a new sequence containing the processed text features. Finally, the text features are embedded into the 3D feature space and used for learning the 3D shape generation by a conditional diffusion model, where the sparsely implicit parsed hierarchical tree graph further enhances the structural details of the generated 3D shapes, leading to results that are close to structure-aware generation. We conducted comprehensive experiments on the existing text-to-shape pairing dataset Text2Shape, and the results demonstrate that our model significantly outperforms current state-of-the-art methods. Moreover, our method can enable progressive part-level 3D shape manipulation and modification guided by the partially modified text prompt.", "keywords": [ "3D shape generation", "text-to-shape" ], "rating_list": [ 6, 3, 3, 3, 6 ], "soundness_list": [ 3, 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 1, 1, 3 ], "confidence_list": [ 4, 5, 4, 4, 3 ] }, { "paper_id": "GQ1Tc3vHbt", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Angelia Nedich", "gender": "Female", "institution": "Arizona State University", "country": "", "position": "Full Professor" }, { "name": "Anton Rodomanov", "gender": "unknown", "institution": "CISPA", "country": "DE", "position": "Postdoc" }, { "name": "Daniil Vankov", "gender": "unknown", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Lalitha Sankar", "gender": "Female", "institution": "Arizona State University", "country": "", "position": "Professor" }, { "name": "Sebastian U Stich", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Tenure Track Faculty" } ], "n_formula": 185, "n_formula_1": 61, "n_ref_uni": 13, "n_ref": 46, "n_ref_all": 91, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 802, "n_element_tab_1": 12, "formula_len_all": 15439, "formula_len_all_1": 4158, "len_all": 166392, "len_all_1": 68040, "len_abs": 1268, "len_title": 109, "len_sents": 52036, "len_sents_1": 26130, "n_sents": 536, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1285, "title": "Optimizing $(L_0, L_1)$-Smooth Functions by Gradient Methods", "abs": "We study gradient methods for solving an optimization problem with an $(L_0, L_1)$-smooth objective function. This problem class generalizes that of Lipschitz-smooth problems and has gained interest recently, as it captures a broader range of machine learning applications.\nWe provide novel insights on the properties of this function class and develop a general framework for analyzing optimization methods \n for $(L_0, L_1)$-smooth function in a principled manner.\nWhile our convergence rate estimates recover existing results for minimizing the gradient norm for nonconvex problems,\nour approach allows us to significantly improve the current state-of-the-art complexity results in the case of convex problems. \nWe show that both the gradient method with Polyak stepsizes and the normalized gradient method, without any knowledge of the parameters $L_0$ and $L_1$, achieve the same complexity bounds as the method with the knowledge of these constants. \nIn addition to that, we show that a carefully chosen accelerated gradient method can be applied to $(L_0, L_1)$-smooth functions, further improving previously known results. \nIn all cases, the efficiency bounds we establish do not have an exponential dependency on $L_0$ or $L_1$, and do not depend on the initial gradient norm.", "keywords": [ "$(L_0", "L_1)$-smoothness", "gradient methods", "convex optimization", "worst-case complexity bounds", "acceleration", "Polyak stepsizes", "nonconvex optimization" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "GPcSYm89wK", "primary_area": "causal reasoning", "n_author": 7, "author_info_list": [ { "name": "Biwei Huang", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "James Bailey", "gender": "unknown", "institution": "The University of Melbourne", "country": "AU", "position": "Full Professor" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Mingming Gong", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Wenjie Wang", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 36, "n_formula_1": 28, "n_ref_uni": 48, "n_ref": 88, "n_ref_all": 103, "n_fig": 3, "n_tab": 2, "L_tab": 1249, "n_element_tab": 65, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 274, "n_element_tab_1": 19, "formula_len_all": 2549, "formula_len_all_1": 2088, "len_all": 154933, "len_all_1": 75731, "len_abs": 1524, "len_title": 120, "len_sents": 48228, "len_sents_1": 34711, "n_sents": 374, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1536, "title": "Practical Kernel Learning for Kernel-based Conditional Independent Test", "abs": "Conditional independence (CI) test stands as a fundamental and challenging task within modern statistics and machine learning. One pivotal class of methods for assessing conditional independence encompasses kernel-based approaches, known for their capability to identify general conditional dependence without necessitating assumptions about the conditional relationship or resorting to the simulation of intricate conditional distributions. As with any method utilizing kernels, selecting the appropriate kernel in kernel-based CI methods is critical for ensuring heightened test power and precise identification of conditional relationship. However, current methods typically involve the manual heuristic selection of kernel parameters, neglecting the inherent characteristics of the data and potentially leading to errors. In this paper, we propose a kernel parameter selection approach for the Kernel-based Conditional Independence test method (KCI). We decompose the statistic of KCI and treat the kernel applied on the conditioning set as a trainable component. The kernel parameters involved are then learned by maximizing the ratio of the estimated statistic to its variance, which approximates the test power at large sample sizes. Therefore, our method can learn the kernel parameters with increased test power at a very small additional computation cost. Extensive experiments demonstrate the effectiveness of our proposed approach in conditional independence testing and its enhancements to constrain-based causal discovery.", "keywords": [ "kernel selection", "conditional independence", "hypothesis testing" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "GPDcvoFGOL", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Jacob Steinhardt", "gender": "unknown", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" }, { "name": "Alyosha Efros", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Professor" }, { "name": "Yossi Gandelsman", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 12, "n_ref_uni": 39, "n_ref": 65, "n_ref_all": 115, "n_fig": 14, "n_tab": 3, "L_tab": 1129, "n_element_tab": 42, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 583, "n_element_tab_1": 58, "formula_len_all": 393, "formula_len_all_1": 594, "len_all": 160260, "len_all_1": 71424, "len_abs": 1279, "len_title": 104, "len_sents": 45401, "len_sents_1": 30010, "n_sents": 400, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1284, "title": "Interpreting the Second-Order Effects of Neurons in CLIP", "abs": "We interpret the function of individual neurons in CLIP by automatically describing them using text. Analyzing the direct effects (i.e. the flow from a neuron through the residual stream to the output) or the indirect effects (overall contribution) fails to capture the neurons' function in CLIP. Therefore, we present the \"second-order lens\", analyzing the effect flowing from a neuron through the later attention heads, directly to the output. We find that these effects are highly selective: for each neuron, the effect is significant for <2% of the images. Moreover, each effect can be approximated by a single direction in the text-image space of CLIP. We describe neurons by decomposing these directions into sparse sets of text representations. The sets reveal polysemantic behavior - each neuron corresponds to multiple, often unrelated, concepts (e.g. ships and cars). Exploiting this neuron polysemy, we mass-produce \"semantic\" adversarial examples by generating images with concepts spuriously correlated to the incorrect class. Additionally, we use the second-order effects for zero-shot segmentation, outperforming previous methods. Our results indicate that a automated interpretation of neurons can be used for model deception and for introducing new model capabilities", "keywords": [ "CLIP", "neurons", "interpretability" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "GOwNImvCWf", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Damian Borth", "gender": "Male", "institution": "University of St.Gallen", "country": "CH", "position": "Full Professor" }, { "name": "Goeran Kauermann", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "", "position": "Full Professor" }, { "name": "Ivan Melev", "gender": "Male", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Konstantin Sch\u00fcrholt", "gender": "Male", "institution": "Universit\u00e4t St. Gallen", "country": "CH", "position": "Postdoc" }, { "name": "L\u00e9o Meynent", "gender": "Male", "institution": "Universit\u00e4t St. Gallen", "country": "", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 43, "n_ref": 82, "n_ref_all": 118, "n_fig": 10, "n_tab": 2, "L_tab": 1111, "n_element_tab": 88, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 305, "n_element_tab_1": 35, "formula_len_all": 881, "formula_len_all_1": 555, "len_all": 163454, "len_all_1": 63204, "len_abs": 1596, "len_title": 110, "len_sents": 57169, "len_sents_1": 32039, "n_sents": 382, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1602, "title": "Structure and Behavior in Weight Space Representation Learning", "abs": "The weights of neural networks (NNs) have recently gained prominence as a new data modality in machine learning, with applications ranging from accuracy and hyperparameter prediction to representation learning or weight generation. One approach to leverage NN weights involves training autoencoders (AEs) with contrastive and reconstruction losses. Indeed, such models can be applied to a wide variety of downstream tasks, and they demonstrate strong predictive performance and low reconstruction error. However, despite the low reconstruction error, these AEs reconconstruct NN models that fail to match the performance of the original ones. In this paper, we identify a limitation of weight-space AEs, specifically highlighting that structural weight reconstruction alone fails to capture some features critical for reconstructing high-performing models. To address this issue, we propose a behavioral loss for training AEs in weight space. This behavioral loss focuses on the features essential for reconstructing performant models, which are not adequately captured by structural reconstruction. We evaluate the capabilities of AE trained using this novel loss on three different model zoos: we demonstrate that when combining structural and behavioral losses, we can reconstruct and generate models that match the performance of the original models. With our exploration of representation learning in deep weight spaces, we show that a strong synergy exists between structural and behavioral features, and that combining them results in increased performance across all evaluated downstream tasks.", "keywords": [ "weight space learning", "hyper-representations", "deep weight spaces", "representation learning", "model reconstruction", "model weights generation" ], "rating_list": [ 5, 3, 3, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "GOoVzE9nSj", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Michael Backes", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Full Professor" }, { "name": "Mingjie Li", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "", "position": "Postdoc" }, { "name": "Yang Zhang", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Full Professor" }, { "name": "Yisen Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Wai man Si", "gender": "Male", "institution": "CISPA, saarland university, saarland informatics campus", "country": "DE", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 19, "n_ref_uni": 39, "n_ref": 71, "n_ref_all": 89, "n_fig": 7, "n_tab": 4, "L_tab": 2405, "n_element_tab": 137, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3592, "n_element_tab_1": 186, "formula_len_all": 927, "formula_len_all_1": 840, "len_all": 124262, "len_all_1": 65921, "len_abs": 1541, "len_title": 102, "len_sents": 31556, "len_sents_1": 27885, "n_sents": 218, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1079, "title": "SaLoRA: Safety-Alignment Preserved Low-Rank Adaptation", "abs": "As advancements in large language models (LLMs) continue and the demand for personalized models increases, parameter-efficient fine-tuning (PEFT) methods (e.g., LoRA) will become essential due to their efficiency in reducing computation costs.\nHowever, recent studies have raised alarming concerns that LoRA fine-tuning could potentially compromise the safety alignment in LLMs, posing significant risks for the model owner.\nIn this paper, we first investigate the underlying mechanism by analyzing the changes in safety alignment related features before and after fine-tuning.\nThen, we propose a fixed safety module calculated by safety data and a task-specific initialization for trainable parameters in low-rank adaptations, termed Safety-alignment preserved Low-Rank Adaptation (SaLoRA). \nUnlike previous LoRA methods and their variants, SaLoRA enables targeted modifications to LLMs without disrupting their original alignments. \nOur experiments show that SaLoRA outperforms various adapters-based approaches across various evaluation metrics in different fine-tuning tasks.", "keywords": [ "PEFT Training", "Safety Alignment Preserving" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "GOjr2Ms5ID", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Atsuki Sato", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "MS student" }, { "name": "Yusuke Matsui", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Lecturer" } ], "n_formula": 18, "n_formula_1": 16, "n_ref_uni": 24, "n_ref": 52, "n_ref_all": 79, "n_fig": 16, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1630, "formula_len_all_1": 1349, "len_all": 121591, "len_all_1": 66234, "len_abs": 841, "len_title": 134, "len_sents": 45969, "len_sents_1": 30615, "n_sents": 354, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 843, "title": "Cascaded Learned Bloom filter for Optimal Model-Filter Size Balance and Fast Rejection", "abs": "Recent studies have demonstrated that learned Bloom filters, which combine machine learning with the classical Bloom filter, can achieve superior memory efficiency. However, existing learned Bloom filters face two critical unresolved challenges: the balance between the machine learning model size and the Bloom filter size is not optimal, and the reject time cannot be minimized effectively. We propose the Cascaded Learned Bloom Filter (CLBF) to address these issues. Our optimization approach based on dynamic programming automatically selects configurations that achieve an optimal balance between the model and filter sizes while minimizing reject time. Experiments with real-world datasets show that CLBF reduces memory usage by up to 24% and decreases reject time by up to 14 times compared to the state-of-the-art learned Bloom filter.", "keywords": [ "learned Bloom filter", "learned index", "membership query", "optimization", "dynamic programming" ], "rating_list": [ 1, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 1, 3, 4, 3 ], "contribution_list": [ 1, 1, 4, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "GOgB6QoXwx", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Jinho Chang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Jong Chul Ye", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 49, "n_ref": 88, "n_ref_all": 110, "n_fig": 13, "n_tab": 8, "L_tab": 2466, "n_element_tab": 243, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1129, "n_element_tab_1": 158, "formula_len_all": 1512, "formula_len_all_1": 553, "len_all": 157733, "len_all_1": 61375, "len_abs": 1272, "len_title": 130, "len_sents": 42502, "len_sents_1": 28486, "n_sents": 275, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1275, "title": "LDMol: Text-to-Molecule Diffusion Model with Structurally Informative Latent Space", "abs": "With the emergence of diffusion models as the frontline of generative models, many researchers have proposed molecule generation techniques with conditional diffusion models. However, the unavoidable discreteness of a molecule makes it difficult for a diffusion model to connect raw data with highly complex conditions like natural language. To address this, we present a novel latent diffusion model dubbed LDMol for text-conditioned molecule generation. LDMol comprises a molecule autoencoder that produces a learnable and structurally informative feature space, and a natural language-conditioned latent diffusion model. In particular, recognizing that multiple SMILES notations can represent the same molecule, we employ a contrastive learning strategy to extract feature space that is aware of the unique characteristics of the molecule structure. LDMol outperforms the existing baselines on the text-to-molecule generation benchmark, suggesting a potential for diffusion models can outperform autoregressive models in text data generation with a better choice of the latent domain. Furthermore, we show that LDMol can be applied to downstream tasks such as molecule-to-text retrieval and text-guided molecule editing, demonstrating its versatility as a diffusion model.", "keywords": [ "Diffusion models", "Molecule generation", "Representation learning" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "GO4Sd6LUuY", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Chongyang Ma", "gender": "Male", "institution": "ByteDance", "country": "", "position": "Tech Lead Manager" }, { "name": "Di ZHANG", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "VP" }, { "name": "Haibin Huang", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Liuxiaoqiang", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Researcher" }, { "name": "Luozhou Wang", "gender": "Male", "institution": "Hong Kong University of Science and Technology (GUANGZHOU)", "country": "CN", "position": "PhD student" }, { "name": "Pengfei Wan", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Director" }, { "name": "Yi Zheng", "gender": "unknown", "institution": "Communication University of China", "country": "", "position": "Associate Professor" }, { "name": "Ying-Cong Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "zhangyuan", "gender": "Male", "institution": "Foundation Model and Multimedia", "country": "CN", "position": "Principal Researcher" }, { "name": "Kan-Le Shi", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 42, "n_ref_all": 62, "n_fig": 1, "n_tab": 6, "L_tab": 2128, "n_element_tab": 304, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1779, "n_element_tab_1": 152, "formula_len_all": 103, "formula_len_all_1": 0, "len_all": 119533, "len_all_1": 60130, "len_abs": 1534, "len_title": 115, "len_sents": 28213, "len_sents_1": 27365, "n_sents": 222, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1542, "title": "Explicit-Constrained Single Agent for Enhanced Task-Solving in LLMs", "abs": "In this study, we introduce the Explicitly Constrained Agent (EC-Agent), a novel approach designed to enhance the task-solving capabilities of Large Language Models (LLMs). Unlike existing multi-agent systems that depend on agents evaluating tasks from different perspectives, EC-Agent explicitly imposes task-oriented constraints for LLMs. Our observations are two-fold: first, assigning agents to sub-tasks with defined responsibilities implicitly sets constraints; second, these multi-agent systems often struggle with accurately assigning agents to sub-tasks, leading to overlapping duties and potential misguidance. In contrast, our single-agent system, driven by explicit methods and constraints, provides LLMs with detailed prompts, resulting in more precise responses. EC-Agent consists of two stages: a Reasoning Stage and a Summary Stage. 1) In the Reasoning Stage, three modules are proposed: Explicit Method, Explicit Constraint, and Execution. Specifically, LLMs utilize the Explicit Method and Constraint modules to analyze the task type and specific rules, generating multiple suitable methods and constraints. Subsequently, the Execution module combines these methods and constraints to produce and output possible solutions. 2) In the Summary Stage, LLMs evaluate the multiple reasoning processes and results from the previous step. They rectify any inconsistencies, summarize the information, and output the final result. Experimental results demonstrate that EC-Agent outperforms previous methods across a variety of tasks.", "keywords": [ "Agent", "LLM" ], "rating_list": [ 3, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 1, 1, 2, 3 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 4, 5, 2, 4 ] }, { "paper_id": "GNOMC90vbl", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Che-Rung Lee", "gender": "Male", "institution": "Department of Computer Science, National Tsing Hua University, National Tsing Hua University", "country": "TW", "position": "Associate Professor" }, { "name": "Erh-Chung Chen", "gender": "unknown", "institution": "National Tsing Hua University", "country": "TW", "position": "PhD student" }, { "name": "I-Hsin Chung", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Pin-Yu Chen", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 41, "n_ref": 64, "n_ref_all": 80, "n_fig": 4, "n_tab": 7, "L_tab": 2525, "n_element_tab": 156, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 905, "n_element_tab_1": 64, "formula_len_all": 621, "formula_len_all_1": 621, "len_all": 136344, "len_all_1": 59075, "len_abs": 1103, "len_title": 141, "len_sents": 37652, "len_sents_1": 26658, "n_sents": 273, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1096, "title": "Data-Driven Lipschitz Continuity: A Cost-Effective Approach to Improve Adversarial Robustness", "abs": "The security and robustness of deep neural networks (DNNs) have become increasingly concerning. This paper aims to provide both a theoretical foundation and a practical solution to ensure the reliability of DNNs. We explore the concept of Lipschitz continuity to certify the robustness of DNNs against adversarial attacks, which aim to mislead the network with adding imperceptible perturbations into inputs. We propose a novel algorithm that remaps the input domain into a constrained range, reducing the Lipschitz constant and potentially enhancing robustness. Unlike existing adversarially trained models, where robustness is enhanced by introducing additional examples from other datasets or generative models, our method is almost cost-free as it can be integrated with existing models without requiring re-training. Experimental results demonstrate the generalizability of our method, as it can be combined with various models and achieve enhancements in robustness. Furthermore, our method achieves the best robust accuracy for CIFAR10 and CIFAR100 datasets on the RobustBench leaderboard.", "keywords": [ "Adversarial training", "AI safety", "Lipschitz Continuity" ], "rating_list": [ 3, 5, 6, 5, 5 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "GMwRl2e9Y1", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Aniketh Iyengar", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Christopher Re", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Dennis Duan", "gender": "Male", "institution": "Stanford University", "country": "", "position": "MS student" }, { "name": "Ehsan Amid", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Jerry Weihong Liu", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Ronald Guenther Junkins", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Undergrad student" }, { "name": "Sebastian Thrun", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Professor (tenured) of computer science and electrical engineering" }, { "name": "Chris Fifty", "gender": "unknown", "institution": "Computer Science Department, Stanford University", "country": "", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 71, "n_ref_all": 134, "n_fig": 16, "n_tab": 9, "L_tab": 3435, "n_element_tab": 376, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1609, "n_element_tab_1": 177, "formula_len_all": 1543, "formula_len_all_1": 461, "len_all": 187614, "len_all_1": 69452, "len_abs": 1479, "len_title": 105, "len_sents": 59877, "len_sents_1": 30429, "n_sents": 439, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 57, "L_abs": 1269, "title": "Restructuring Vector Quantization with the Rotation Trick", "abs": "Vector Quantized Variational AutoEncoders (VQ-VAEs) are designed to compress a continuous input to a discrete latent space and reconstruct it with minimal distortion. \nThey operate by maintaining a set of vectors---often referred to as the codebook---and quantizing each encoder output to the nearest vector in the codebook. \nHowever, as vector quantization is non-differentiable, the gradient to the encoder flows _around_ the vector quantization layer rather than _through_ it in a straight-through approximation.\nThis approximation may be undesirable as all information from the vector quantization operation is lost. \nIn this work, we propose a way to propagate gradients through the vector quantization layer of VQ-VAEs. \nWe smoothly transform each encoder output into its corresponding codebook vector via a rotation and rescaling linear transformation that is treated as a constant during backpropagation. \nAs a result, the relative magnitude and angle between encoder output and codebook vector becomes encoded into the gradient as it propagates through the vector quantization layer and back to the encoder.\nAcross 11 different VQ-VAE training paradigms, we find this restructuring improves reconstruction metrics, codebook utilization, and quantization error.", "keywords": [ "Vector Quantization", "VQ-VAE" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 4, 3, 4, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "GMu1yJST86", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Haitao Wu", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Weiwei Li", "gender": "Female", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "Associate Professor" }, { "name": "Xiuyi Jia", "gender": "unknown", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 32, "n_formula_1": 25, "n_ref_uni": 30, "n_ref": 95, "n_ref_all": 125, "n_fig": 4, "n_tab": 12, "L_tab": 1793, "n_element_tab": 216, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 133, "n_element_tab_1": 14, "formula_len_all": 3615, "formula_len_all_1": 2608, "len_all": 168979, "len_all_1": 75897, "len_abs": 1358, "len_title": 89, "len_sents": 38542, "len_sents_1": 29186, "n_sents": 607, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1398, "title": "Learning Label Distribution with Subtasks", "abs": "Label distribution learning (LDL) is a novel learning paradigm that emulates label polysemy by assigning label distributions over the label space. However, recent LDL work seems to exhibit a notable contradiction: 1) some existing LDL methods employ auxiliary tasks to enhance performance, which narrows their focus to specific domains, thereby lacking generalization capability; 2) conversely, LDL methods without auxiliary tasks rely on losses tailored solely to label distributions of the primary task, lacking additional supervised information to guide the learning process. In this paper, we propose $\\mathcal{S}$-LDL, a novel and minimalist solution that partitions the label distribution of the primary task into subtask label distributions, i.e., a form of pseudo-supervised information, to reconcile the above contradiction. $\\mathcal{S}$-LDL encompasses two key aspects: 1) an algorithm capable of generating subtasks without any extra knowledge, with subtasks deemed valid and reconstructable via our analysis; and 2) a plug-and-play framework seamlessly compatible with existing LDL methods, and even adaptable to derivative tasks of LDL. Experiments demonstrate that $\\mathcal{S}$-LDL is effective and efficient. To the best of our knowledge, this represents the first endeavor to address LDL via subtasks. The code will soon be available on GitHub to facilitate reproducible research.", "keywords": [ "label distribution learning", "subtask", "label polysemy" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "GM7cmQfk2F", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Hanzhang Qin", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Assistant Professor" }, { "name": "Jiahai Wang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "Full Professor" }, { "name": "Jinbiao Chen", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "PhD student" }, { "name": "Yue-Jiao Gong", "gender": "Female", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zhiguang Cao", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Zizhen Zhang", "gender": "unknown", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "Associate Professor" }, { "name": "YAOXIN WU", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" } ], "n_formula": 23, "n_formula_1": 14, "n_ref_uni": 46, "n_ref": 89, "n_ref_all": 123, "n_fig": 11, "n_tab": 10, "L_tab": 10351, "n_element_tab": 1497, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 4602, "n_element_tab_1": 537, "formula_len_all": 1629, "formula_len_all_1": 921, "len_all": 165447, "len_all_1": 70005, "len_abs": 1115, "len_title": 135, "len_sents": 38004, "len_sents_1": 27271, "n_sents": 288, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1121, "title": "Rethinking Neural Multi-Objective Combinatorial Optimization via Neat Weight Embedding", "abs": "Recent decomposition-based neural multi-objective combinatorial optimization (MOCO) methods struggle to achieve desirable performance. Even equipped with complex learning techniques, they often suffer from significant optimality gaps in weight-specific subproblems. To address this challenge, we propose a neat weight embedding method to learn weight-specific representations, which captures weight-instance interaction for the subproblems and was overlooked by most current methods. We demonstrate the potentials of our method in two instantiations. First, we introduce a succinct addition model to learn weight-specific node embeddings, which surpassed most existing neural methods. Second, we design an enhanced conditional attention model to simultaneously learn the weight embedding and node embeddings, which yielded new state-of-the-art performance. Experimental results on classic MOCO problems verified the superiority of our method. Remarkably, our method also exhibits favorable generalization performance across problem sizes, even outperforming the neural method specialized for boosting size generalization.", "keywords": [ "Neural Multi-Objective Combinatorial Optimization", "Weight Embedding", "Conditional Attention" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "GLuzjuG0lo", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 1, "author_info_list": [ { "name": "Chunqing Ruan", "gender": "Female", "institution": "xuezhilixing", "country": "CN", "position": "Intern" } ], "n_formula": 16, "n_formula_1": 13, "n_ref_uni": 23, "n_ref": 35, "n_ref_all": 41, "n_fig": 2, "n_tab": 4, "L_tab": 1805, "n_element_tab": 390, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2608, "n_element_tab_1": 423, "formula_len_all": 509, "formula_len_all_1": 477, "len_all": 142592, "len_all_1": 55795, "len_abs": 1353, "len_title": 154, "len_sents": 25096, "len_sents_1": 23409, "n_sents": 167, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1360, "title": "Vision State Space Duality for Medical Image Segmentation: Enhancing Precision through Non-Causal Modeling", "abs": "In medical image analysis, Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) have set significant benchmarks. However, CNNs exhibit limitations in long-range modeling capabilities, whereas Transformers are hampered by their quadratic computational complexity. Recently, State Space Models (SSMs) have gained prominence in vision tasks as they offer linear computational complexity. State Space Duality (SSD), an improved variant of SSMs, was introduced in Mamba2 to enhance model performance and efficiency. Inspired by this, we have tailored the Vision State Space Duality (VSSD) model for medical image segmentation tasks by integrating it within a UNet-like architecture, which is renowned for its effectiveness in the field. Our modified model, named VSSD-UNet, employs skip connections to preserve spatial information and utilizes a series of VSSD blocks for feature extraction. In addition, VSSD-UNet employs a hybrid structure of VSSD and self-attention in the decoder part, ensuring that both local details and global contexts are captured. Finally, we conducted comparative and ablation experiments on two public lesion segmentation datasets: ISIC2017 and ISIC2018. The results show that VSSD-UNet outperforms several types of UNet in medical image segmentation under the same hyper-parameter setting. Our code will be released soon.", "keywords": [ "State Space Models", "Vision State Space Duality", "Medical Image Segmentation", "self-attention", "skip connections" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 1, 1 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "GLmqHCwbOJ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Naihao Deng", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Rada Mihalcea", "gender": "Female", "institution": "University of Michigan", "country": "", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 20, "n_ref": 68, "n_ref_all": 123, "n_fig": 13, "n_tab": 23, "L_tab": 8718, "n_element_tab": 773, "n_fig_1": 9, "n_tab_1": 8, "L_tab_1": 2534, "n_element_tab_1": 324, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 202769, "len_all_1": 55949, "len_abs": 1307, "len_title": 83, "len_sents": 50250, "len_sents_1": 23408, "n_sents": 360, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1329, "title": "Rethinking Table Instruction Tuning", "abs": "Recent advances in table understanding have focused on instruction-tuning large language models (LLMs) for table-related tasks. \nHowever, existing research has overlooked the impact of hyperparameter choices and lacks a comprehensive evaluation of the out-of-domain table understanding ability and the general capabilities of these table LLMs. \nIn this paper, we evaluate these abilities in existing table LLMs, and reveal significant declines in both out-of-domain table understanding and general capabilities compared to their base models. \nThrough systematic analysis, we show that hyperparameters, such as learning rate, can significantly influence both table-specific and general capabilities.\nContrary to the existing table instruction-tuning works, we demonstrate that smaller learning rates and fewer training instances can enhance table understanding while preserving general capabilities.\nBased on our findings, we introduce **TAMA**, a **TA**ble LLM instruction-tuned from LLa**MA** 3.1 8B Instruct, which achieves performance on par with, or surpassing GPT-3.5 and GPT-4 on table tasks, while maintaining strong out-of-domain generalization and general capabilities. \nOur findings highlight the potential for reduced data annotation costs and more efficient model development through careful hyperparameter selection.", "keywords": [ "table instruction tuning", "table understanding" ], "rating_list": [ 5, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "GLmOWcqvE3", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Henrik Christensen", "gender": "Male", "institution": "Computer Science and Engineering Department, University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Rohan Prashant Patil", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 49, "n_formula_1": 27, "n_ref_uni": 19, "n_ref": 39, "n_ref_all": 63, "n_fig": 7, "n_tab": 3, "L_tab": 1498, "n_element_tab": 138, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5033, "formula_len_all_1": 2523, "len_all": 111793, "len_all_1": 59392, "len_abs": 759, "len_title": 100, "len_sents": 37580, "len_sents_1": 26532, "n_sents": 372, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 760, "title": "BOIL: Learning Environment Personalized Information", "abs": "Navigating complex environments poses challenges for multi-agent systems, requiring efficient extraction of insights from limited information. In this paper, we introduce the Blackbox Oracle Information Learning (BOIL) process, a scalable solution for extracting valuable insights from the environment structure. Leveraging the Pagerank algorithm and common information maximization, BOIL facilitates the extraction of information to guide long-term agent behavior applicable to problems such as coverage, patrolling, and stochastic reachability. Through experiments, we demonstrate the efficacy of BOIL in generating strategy distributions conducive to improved performance over extended time horizons, surpassing heuristic approaches in complex environments.", "keywords": [ "learning on graphs", "Information Theory", "Pagerank", "applications to robotics", "probabilistic methods" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 2, 2, 3 ] }, { "paper_id": "GLWf2fq0bX", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Haoyi Niu", "gender": "Male", "institution": "Department of Automation, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Tenglong Liu", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Xianyuan Zhan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Xin Xu", "gender": "Male", "institution": "National University of Defense Technology, China", "country": "CN", "position": "Full Professor" }, { "name": "Yinan Zheng", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Li Jianxiong", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "YIXING LAN", "gender": "Male", "institution": "National University of Defense Technology,", "country": "", "position": "PhD student" } ], "n_formula": 35, "n_formula_1": 10, "n_ref_uni": 94, "n_ref": 280, "n_ref_all": 350, "n_fig": 23, "n_tab": 8, "L_tab": 4429, "n_element_tab": 551, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1661, "n_element_tab_1": 363, "formula_len_all": 2079, "formula_len_all_1": 663, "len_all": 285007, "len_all_1": 75367, "len_abs": 1560, "len_title": 98, "len_sents": 72078, "len_sents_1": 33449, "n_sents": 503, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1616, "title": "Skill Expansion and Composition in Parameter Space", "abs": "Humans excel at reusing prior knowledge to address new challenges and developing skills while solving problems. This paradigm becomes increasingly popular in the development of autonomous agents, as it develops systems that can self-evolve in response to new challenges like human beings. However, previous methods suffer from limited training efficiency when expanding new skills and fail to fully leverage prior knowledge to facilitate new task learning. We propose Parametric Skill Expansion and Composition (PSEC), a new framework designed to iteratively evolve the agents' capabilities and efficiently address new challenges by maintaining a manageable skill library. This library can progressively integrate skill primitives as plug-and-play Low-Rank Adaptation (LoRA) modules in parameter-efficient finetuning, facilitating efficient and flexible skill expansion. This structure also enables the direct skill compositions in parameter space by merging LoRA modules that encode different skills, leveraging shared information across skills to effectively program new skills. Based on this, we propose a context-aware modular to dynamically activate different skills to collaboratively handle new tasks. Empowering diverse applications including multi-objective composition, dynamics shift, and continual policy shift, the results on D4RL, DSRL benchmarks, and the DeepMind Control Suite show that PSEC exhibits superior capacity to leverage prior knowledge to efficiently tackle new challenges, as well as expand its skill libraries to evolve the capabilities. Project website: https://ltlhuuu.github.io/PSEC/.", "keywords": [ "Skill expansion", "skill composition", "parameter-efficient finetuning", "decision making" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 4, 4, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "GKAQ92ua3A", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Ganzhao Yuan", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Associate Professor" } ], "n_formula": 214, "n_formula_1": 30, "n_ref_uni": 50, "n_ref": 108, "n_ref_all": 147, "n_fig": 1, "n_tab": 3, "L_tab": 350, "n_element_tab": 18, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 121, "n_element_tab_1": 7, "formula_len_all": 26235, "formula_len_all_1": 5718, "len_all": 236753, "len_all_1": 76801, "len_abs": 1235, "len_title": 115, "len_sents": 48808, "len_sents_1": 26017, "n_sents": 700, "n_sents_1": 294, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1261, "title": "ADMM for Nonconvex Optimization under Minimal Continuity Assumption", "abs": "This paper introduces a novel approach to solving multi-block nonconvex composite optimization problems through a proximal linearized Alternating Direction Method of Multipliers (ADMM). This method incorporates an Increasing Penalization and Decreasing Smoothing (IPDS) strategy. Distinguishing itself from existing ADMM-style algorithms, our approach (denoted IPDS-ADMM) imposes a less stringent condition, specifically requiring continuity in just one block of the objective function. IPDS-ADMM requires that the penalty increases and the smoothing parameter decreases, both at a controlled pace. When the associated linear operator is bijective, IPDS-ADMM uses an over-relaxation stepsize for faster convergence; however, when the linear operator is surjective, IPDS-ADMM uses an under-relaxation stepsize for global convergence. We devise a novel potential function to facilitate our convergence analysis and prove an oracle complexity $O(\\epsilon^{-3})$ to achieve an $\\epsilon$-approximate critical point. To the best of our knowledge, this is the first complexity result for using ADMM to solve this class of nonsmooth nonconvex problems. Finally, some experiments on the sparse PCA problem are conducted to demonstrate the effectiveness of our approach.", "keywords": [ "Nonconvex Optimization", "Proximal Linearized ADMM", "Nonsmooth Optimization", "Convergence Analysis" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 3, 3 ] }, { "paper_id": "GK5ni7tIHp", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Haotian Ye", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Haowei Lin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jianzhu Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Shanda Li", "gender": "Male", "institution": "Google", "country": "US", "position": "Intern" }, { "name": "Stefano Ermon", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Yiming Yang", "gender": "Female", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Yitao Liang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 80, "n_formula_1": 37, "n_ref_uni": 92, "n_ref": 182, "n_ref_all": 223, "n_fig": 2, "n_tab": 16, "L_tab": 3535, "n_element_tab": 447, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 1029, "n_element_tab_1": 155, "formula_len_all": 7029, "formula_len_all_1": 2457, "len_all": 294590, "len_all_1": 74326, "len_abs": 1135, "len_title": 130, "len_sents": 75167, "len_sents_1": 31325, "n_sents": 656, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1139, "title": "TFG-Flow: Training-free Guidance in Multimodal Generative Flow", "abs": "Given an unconditional generative model and a predictor for a target property (e.g., a classifier), the goal of training-free guidance is to generate samples with desirable target properties without additional training. As a highly efficient technique for steering generative models toward flexible outcomes, training-free guidance has gained increasing attention in diffusion models. However, existing methods only handle data in continuous spaces, while many scientific applications involve both continuous and discrete data (referred to as multimodality). Another emerging trend is the growing use of the simple and general flow matching framework in building generative foundation models, where guided generation remains under-explored. To address this, we introduce TFG-Flow, a novel training-free guidance method for multimodal generative flow. TFG-Flow addresses the curse-of-dimensionality while maintaining the property of unbiased sampling in guiding discrete variables. We validate TFG-Flow on four molecular design tasks and show that TFG-Flow has great potential in drug design by generating molecules with desired properties.", "keywords": [ "flow matching", "molecular design", "training-free guidance" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 3, 5, 2 ] }, { "paper_id": "GJsuYHhAga", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Enxin Song", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Jinbin Bai", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Lei Zhu", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou) & HKUST", "country": "GB", "position": "Assistant Professor" }, { "name": "Qing-Guo Chen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Tian Ye", "gender": "Not Specified", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Xiangtai Li", "gender": "Male", "institution": "ByteDance Inc.", "country": "SG", "position": "Researcher" }, { "name": "Zhen Dong", "gender": "Male", "institution": "Nexusflow.ai Inc", "country": "US", "position": "Principal Researcher" }, { "name": "Wei Chou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 103, "n_ref_all": 135, "n_fig": 21, "n_tab": 9, "L_tab": 1516, "n_element_tab": 159, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2433, "n_element_tab_1": 150, "formula_len_all": 25, "formula_len_all_1": 25, "len_all": 208611, "len_all_1": 55469, "len_abs": 930, "len_title": 156, "len_sents": 50749, "len_sents_1": 23785, "n_sents": 567, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 108, "L_abs": 940, "title": "Meissonic: Revitalizing Masked Generative Transformers for Efficient High-Resolution Text-to-Image Synthesis", "abs": "We present Meissonic, which elevates non-autoregressive text-to-image Masked Image Modeling (MIM) to a level comparable with state-of-the-art diffusion models like SDXL. By incorporating a comprehensive suite of architectural innovations, advanced positional encoding strategies, and optimized sampling conditions, Meissonic substantially improves MIM's performance and efficiency. Additionally, we leverage high-quality training data, integrate micro-conditions informed by human preference scores, and employ feature compression layers to further enhance image fidelity and resolution. Our model not only matches but often exceeds the performance of existing methods in generating high-quality, high-resolution images. Extensive experiments validate Meissonic\u2019s capabilities, demonstrating its potential as a new standard in text-to-image synthesis. We release a model checkpoint capable of producing $1024 \\times 1024$ resolution images.", "keywords": [ "Text-to-Image Synthesis" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 4 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "GIFn3ibnKa", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Aditya Nigam", "gender": "Male", "institution": "Indian Institute of Technology Mandi", "country": "IN", "position": "Associate Professor" }, { "name": "Anurag Pandey", "gender": "Male", "institution": "IIT Mandi", "country": "IN", "position": "PhD student" }, { "name": "Arnav Bhavsar", "gender": "Male", "institution": "IIT Mandi", "country": "IN", "position": "Full Professor" }, { "name": "Basu Verma", "gender": "Male", "institution": "HCLTech", "country": "IN", "position": "Principal Researcher" }, { "name": "Dr. Divya Acharya", "gender": "Female", "institution": "Bennett University", "country": "IN", "position": "PhD student" }, { "name": "PushapDeep Singh", "gender": "Male", "institution": "IIT Mandi", "country": "IN", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 4, "n_ref_uni": 32, "n_ref": 36, "n_ref_all": 54, "n_fig": 5, "n_tab": 9, "L_tab": 2324, "n_element_tab": 164, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1767, "n_element_tab_1": 162, "formula_len_all": 568, "formula_len_all_1": 261, "len_all": 108607, "len_all_1": 55399, "len_abs": 1877, "len_title": 84, "len_sents": 37589, "len_sents_1": 25915, "n_sents": 251, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1887, "title": "Explore To Mimic: A Reinforcement Learning Based Agent To Generate Online Signatures", "abs": "Recent advancements in utilising decision making capability of Reinforcement Learning (RL) have paved the way for innovative approaches in data generation. This research explores the application of model free on-policy RL algorithms for generating online signatures and its controlled variations. Online signatures are captured via e-pads as sequential structural coordinates. In this study, we have introduced a robust on-policy RL agent named as SIGN-Agent, capable of generating online signatures accurately. Unlike other RL algorithms, on-policy RL directly learns from the agent's current policy, offering significant advantages in stability and faster convergence for sequential decision-making. The proposed SIGN-Agent operates in a random continuous action space with controlled exploration limits, allowing it to capture complex signature patterns while minimizing errors over time. The downstream applications of this system can be extended in diverse fields such as enhancing the robustness of signature authentication systems, supporting robotics, and even diagnosing neurological disorders. By generating reliable, human-like online signatures, our approach strengthens signature authentication systems by reducing susceptibility towards system-generated forgeries, if trained against them. Additionally, the proposed work is optimized for low-footprint edge devices, enabling it to function efficiently in the area of robotics for online signature generation tasks. Experimental results, tested on large, publicly available datasets, demonstrate the effectiveness of model free on-policy RL algorithms in generating online signature trajectories, that closely resemble user's reference signatures. Our approach highlights the potential of model free on-policy RL as an advancement in the field of data generation targeting the domain of online signatures in this research.", "keywords": [ "Reinforcement Learning", "Online Signature", "Biometric", "Generative Model", "On-Policy" ], "rating_list": [ 1, 5, 5 ], "soundness_list": [ 1, 3, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 1, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "GI5cgzVjK6", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Alec Go", "gender": "unknown", "institution": "Google", "country": "US", "position": "Software Engineer" }, { "name": "Congchao Wang", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Kai Zhang", "gender": "Male", "institution": "Worcester Polytechnic Institute", "country": "", "position": "PhD student" }, { "name": "Liqian Peng", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Xiaozhong Liu", "gender": "Male", "institution": "Worcester Polytechnic Institute", "country": "", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 47, "n_ref": 87, "n_ref_all": 114, "n_fig": 8, "n_tab": 4, "L_tab": 769, "n_element_tab": 131, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 685, "n_element_tab_1": 124, "formula_len_all": 175, "formula_len_all_1": 147, "len_all": 130236, "len_all_1": 60811, "len_abs": 838, "len_title": 102, "len_sents": 32249, "len_sents_1": 29744, "n_sents": 204, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 842, "title": "LLM Cascade with Multi-Objective Optimal Consideration", "abs": "Large Language Models (LLMs) have demonstrated exceptional capabilities in understanding and generating natural language. However, their high deployment costs often pose a barrier to practical applications, especially. Cascading local and server models offers a promising solution to this challenge. While existing studies on LLM cascades have primarily focused on the performance-cost trade-off, real-world scenarios often involve more complex requirements. This paper introduces a novel LLM Cascade strategy with Multi-Objective Optimization, enabling LLM cascades to consider additional objectives (e.g., privacy) and better align with the specific demands of real-world applications while maintaining their original cascading abilities. Extensive experiments on three benchmarks validate the effectiveness and superiority of our approach.", "keywords": [ "LLM Cascade; Multi-Objective; On-device Intelligence" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 2, 1, 2, 1 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "GHaoCSlhcK", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Liu Ren", "gender": "Male", "institution": "Bosch Research", "country": "US", "position": "Principal Researcher" }, { "name": "Sima Behpour", "gender": "Female", "institution": "Bosch Center for Artificial Intelligence (BCAI)", "country": "US", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 45, "n_ref": 87, "n_ref_all": 109, "n_fig": 4, "n_tab": 8, "L_tab": 4149, "n_element_tab": 260, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 2266, "n_element_tab_1": 26, "formula_len_all": 491, "formula_len_all_1": 362, "len_all": 163903, "len_all_1": 59410, "len_abs": 1222, "len_title": 157, "len_sents": 52146, "len_sents_1": 28288, "n_sents": 371, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1347, "title": "PROGRESSIVE KNOWLEDGE DISTILLATION (PKD): A MODULAR APPROACH FOR ARCHITECTURE-AGNOSTIC KNOWLEDGE DISTILLATION", "abs": "\\textbf{Knowledge distillation (KD)} is a key technique for training \\textbf{lightweight deep neural networks}, particularly in \\textbf{resource-constrained environments}. While existing KD methods utilize intermediate features to improve student models, they often overlook the proper \\textbf{alignment between teacher-student layers} and fail to select the most \\textbf{informative data} for training each student layer. These limitations are especially pronounced in \\textbf{architecture-agnostic scenarios}, where different network architectures complicate knowledge transfer.\n\nWe propose \\textbf{PKD}, a \\textbf{Progressive Knowledge Distillation} framework that progressively aligns teacher and student layers through \\textbf{feature-based modularization}. Each student module is trained using the most \\textbf{representative features} from its corresponding teacher module, starting with the shallowest layers and progressively moving to deeper ones. This training method enables efficient, architecture-agnostic knowledge transfer across a variety of model architectures. \\textbf{Experiments on CIFAR-100 and ImageNet-1K} demonstrate that PKD outperforms baseline models, achieving performance improvements of up to \\textbf{4.54\\%} and \\textbf{6.46\\%}, respectively, thereby validating its effectiveness in diverse neural network settings.", "keywords": [ "knowledge distillation" ], "rating_list": [ 3, 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2, 1 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "GHJzxPgFa6", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 14, "author_info_list": [ { "name": "Boqiang Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Deli Zhao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Director" }, { "name": "Jiayan Guo", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Lidong Bing", "gender": "unknown", "institution": "Alibaba Group", "country": "SG", "position": "Scientist" }, { "name": "Ronghao Dang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Ruochen Zhao", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Tian Feng", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Associate Professor" }, { "name": "Weiwen Xu", "gender": "Male", "institution": "Alibaba Group", "country": "SG", "position": "Researcher" }, { "name": "Xingxuan Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Yu Rong", "gender": "Male", "institution": "Alibaba Group ", "country": "CN", "position": "Senior Staff Algorithm Engineer" }, { "name": "Yuming Jiang", "gender": "Male", "institution": "Alibaba Group", "country": "SG", "position": "Researcher" }, { "name": "Yuqian Yuan", "gender": "Female", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "\u674e \u9f99", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "YifeiXin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 3, "n_ref_uni": 26, "n_ref": 57, "n_ref_all": 136, "n_fig": 9, "n_tab": 18, "L_tab": 13569, "n_element_tab": 476, "n_fig_1": 9, "n_tab_1": 6, "L_tab_1": 2196, "n_element_tab_1": 213, "formula_len_all": 193, "formula_len_all_1": 244, "len_all": 192711, "len_all_1": 74673, "len_abs": 1422, "len_title": 130, "len_sents": 59509, "len_sents_1": 32722, "n_sents": 490, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1428, "title": "Chain of Ideas: Revolutionizing Research in Idea Development with LLM Agents", "abs": "Effective research ideation is a critical step for scientific research. However, the exponential increase in scientific literature makes it challenging for researchers to stay current with recent advances and identify meaningful research directions. Recent developments in large language models~(LLMs) suggest a promising avenue for automating the generation of novel research ideas. However, existing methods for idea generation either trivially prompt LLMs or directly expose LLMs to extensive literature without indicating useful information. Inspired by the research process of human researchers, we propose a Chain-of-Ideas (CoI) agent, an LLM-based agent that organizes relevant literature in a chain structure to effectively mirror the progressive development in a research domain. This organization facilitates LLMs to capture the current advancements in research, thereby enhancing their ideation capabilities. Furthermore, we propose Idea Arena, an evaluation protocol that can comprehensively evaluate idea generation methods from different perspectives, aligning closely with the preferences of human researchers. Experimental results indicate that the CoI agent consistently outperforms other methods and shows comparable quality as humans in research idea generation. Moreover, our CoI agent is budget-friendly, with a minimum cost of \\$0.50 to generate a candidate idea and its corresponding experimental design.", "keywords": [ "Large Language Model", "Agent", "Idea Generation" ], "rating_list": [ 5, 5, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 2, 1, 3, 1 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "GGlpykXDCa", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Dongyuan Li", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Assistant Professor" }, { "name": "Jian Wu", "gender": "unknown", "institution": "Tokyo Institute of Technology, Tokyo Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Linyi Yang", "gender": "unknown", "institution": "University College London", "country": "GB", "position": "Researcher" }, { "name": "Yue Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" }, { "name": "Yuliang Ji", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Manabu Okumura", "gender": "Male", "institution": "Institute of Science Tokyo", "country": "JP", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 44, "n_ref": 73, "n_ref_all": 114, "n_fig": 4, "n_tab": 15, "L_tab": 4547, "n_element_tab": 489, "n_fig_1": 2, "n_tab_1": 15, "L_tab_1": 6475, "n_element_tab_1": 367, "formula_len_all": 108, "formula_len_all_1": 108, "len_all": 164816, "len_all_1": 58103, "len_abs": 1511, "len_title": 114, "len_sents": 37274, "len_sents_1": 21188, "n_sents": 262, "n_sents_1": 139, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 66, "L_abs": 1448, "title": "MMQA: Evaluating LLMs with Multi-Table Multi-Hop Complex Questions", "abs": "While large language models (LLMs) have made strides in understanding tabular data, current tabular evaluation benchmarks, such as WikiTableQuestions and WikiSQL, are focus on single-table scenarios, which cannot necessarily reflect the complexity of real-world applications. To bridge this gap, we present a \\textbf{M}ulti-table and \nMulti-hop Question Answering (MMQA) dataset to assess LLMs' understanding and reasoning capabilities in handling multi-table tasks. The MMQA dataset demands that models perform multiple inferences by drawing evidence from various tables, which are designed to be connected with each other and require models to identify and utilize relationships such as foreign and primary keys. Then, we introduce a comprehensive evaluation framework that tailors to assess LLMs' capabilities in several aspects including Multi-Table Retrieval, Text-to-SQL Generation, Multi-Table QA, Primary Key Selection, and Foreign Key Selection. \nFinally, we propose a novel multi-table retrieval method that achieves state-of-the-art (SOTA) performance on the MMQA dataset compared to several strong baselines. \nOur experiment results reveal that, compared with human performance, both open-source and commercial LLMs leave significant performance room for improvements in multi-table understanding and reasoning tasks. We believe that the MMQA benchmark will enhance and facilitate LLMs' multi-table capabilities in real-world scenarios.", "keywords": [ "LLM evaluation", "multi-table question answering; multi-hop question answering" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 4, 3, 4 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 3, 3, 4 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "GGZISiwgNt", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Eshika Pathak", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Gauri Joshi", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Guannan Qu", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Neharika Jali", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "PRANAY SHARMA", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "Postdoc" } ], "n_formula": 180, "n_formula_1": 40, "n_ref_uni": 41, "n_ref": 92, "n_ref_all": 116, "n_fig": 7, "n_tab": 3, "L_tab": 287, "n_element_tab": 43, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 2979, "n_element_tab_1": 39, "formula_len_all": 31074, "formula_len_all_1": 3192, "len_all": 261445, "len_all_1": 75279, "len_abs": 1045, "len_title": 108, "len_sents": 61362, "len_sents_1": 28443, "n_sents": 565, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1204, "title": "Natural Policy Gradient for Average Reward Non-Stationary RL", "abs": "We consider the problem of non-stationary reinforcement learning (RL) in the infinite-horizon average-reward setting. We model it by a Markov Decision Process with time-varying rewards and transition probabilities, with a variation budget of $\\Delta_T$. Existing non-stationary RL algorithms focus on model-based and model-free value-based methods. Policy-based methods, however, despite their flexibility in practice, are not theoretically well understood in non-stationary RL. We propose and analyze the first model-free policy-based algorithm, Non-Stationary Natural Actor-Critic (NS-NAC), a policy gradient method with efficient exploration for change and a novel interpretation of learning rates as adapting factors. We present a dynamic regret of $\\mathcal{\\tilde{O}} (|\\mathcal{S}|^{\\frac{1}{2}}|\\mathcal{A}|^{\\frac{1}{2}}\\Delta_T^{\\frac{1}{9}}T^{\\frac{8}{9}} )$, where $T$ is the time horizon, and $|\\mathcal{S}|$, $|\\mathcal{A}|$ are, respectively, the size of the state and action space. The regret analysis relies on adapting the Lyapunov function based analysis to dynamic environments and characterizing the effects of simultaneous updates in policy, value function estimate and environment.", "keywords": [ "Non-Stationary Reinforcement Learning", "Policy Gradient", "Natural Actor-Critic" ], "rating_list": [ 6, 6, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4, 4 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 3, 2, 3, 4, 4 ] }, { "paper_id": "GGAG3wFEKv", "primary_area": "reinforcement learning", "n_author": 1, "author_info_list": [ { "name": "Imad AOUALI", "gender": "Male", "institution": "Criteo", "country": "", "position": "PhD student" } ], "n_formula": 135, "n_formula_1": 41, "n_ref_uni": 55, "n_ref": 117, "n_ref_all": 186, "n_fig": 11, "n_tab": 2, "L_tab": 708, "n_element_tab": 37, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18954, "formula_len_all_1": 3934, "len_all": 231230, "len_all_1": 73923, "len_abs": 814, "len_title": 88, "len_sents": 66558, "len_sents_1": 31551, "n_sents": 655, "n_sents_1": 277, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 818, "title": "Diffusion Models Meet Contextual Bandits", "abs": "Efficient exploration in contextual bandits is crucial due to their large action space, where uninformed exploration can lead to computational and statistical inefficiencies. However, the rewards of actions are often correlated, which can be leveraged for more efficient exploration. In this work, we use pre-trained diffusion model priors to capture these correlations and develop diffusion Thompson sampling (dTS). We establish both theoretical and algorithmic foundations for dTS. Specifically, we derive efficient posterior approximations (required by dTS) under a diffusion model prior, which are of independent interest beyond bandits and reinforcement learning. We analyze dTS in linear instances and provide a Bayes regret bound. Our experiments validate our theory and demonstrate dTS's favorable performance.", "keywords": [ "Diffusion models", "Bayesian bandit", "Thompson sampling", "Contextual bandit" ], "rating_list": [ 5, 5, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "GG80jy9KI5", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Matthias J. Feiler", "gender": "Male", "institution": "University of Zurich", "country": "", "position": "Lecturer" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 9, "n_ref": 10, "n_ref_all": 25, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 749, "formula_len_all_1": 749, "len_all": 63936, "len_all_1": 55670, "len_abs": 1012, "len_title": 89, "len_sents": 27529, "len_sents_1": 27138, "n_sents": 214, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1016, "title": "Strong denoising of financial time-series", "abs": "In this paper we introduce a method for improving the signal to noise ratio of financial data. The approach relies on combining a target variable with different context variables and using auto-encoders (AEs) to learn reconstructions of the combined inputs. The idea is to seek agreement among multiple AEs which are trained on related but different inputs for which they are forced to find common ground. The training process is set up as a conversation where models take turns at producing a prediction (speaking) or reconciling own predictions with the output of the other AE (listening), until an agreement is reached. This leads to \"mutual regularization\" among the AEs. Unlike standard regularization which relies on including a complexity penalty into the loss function, the proposed method uses the partner network to detect and amend the lack of generality in the data representation. As only true regularities can be agreed upon by the AEs, the replication of noise is costly and will therefore be avoided.", "keywords": [ "mutual learning", "regularization" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 1, 2, 1, 2 ], "presentation_list": [ 1, 1, 2, 1 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 3, 3, 2, 2 ] }, { "paper_id": "GFzmAKw3RW", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Guorui Feng", "gender": "Male", "institution": "Shanghai University", "country": "CN", "position": "Full Professor" }, { "name": "Xinpeng Zhang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yalan Qin", "gender": "unknown", "institution": "Shanghai University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 24, "n_ref_uni": 26, "n_ref": 33, "n_ref_all": 53, "n_fig": 2, "n_tab": 7, "L_tab": 3960, "n_element_tab": 357, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2314, "n_element_tab_1": 270, "formula_len_all": 1488, "formula_len_all_1": 1367, "len_all": 109789, "len_all_1": 56287, "len_abs": 1979, "len_title": 119, "len_sents": 27537, "len_sents_1": 21403, "n_sents": 237, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1977, "title": "Efficient Incomplete Multi-view Clustering via Flexible Anchor Learning", "abs": "Multi-view clustering aims to improve the final performance by taking advantages of complementary and consistent information of all views. In real world, data samples with partially available information are common and the issue regarding the clustering for incomplete multi-view data is inevitably raised. To deal with the partial data with large scales, some fast clustering approaches for incomplete multi-view data have been presented. Despite the significant success, few of these methods pay attention to learning anchors with high quality in a unified framework for incomplete multi-view clustering, while ensuring the scalability for large-scale incomplete datasets. In addition, most existing approaches based on incomplete multi-view clustering ignore to build the relation between anchor graph and similarity matrix in symmetric nonnegative matrix factorization and then directly conduct graph partition based on the anchor graph to reduce the space and time consumption. In this paper, we propose a novel fast incomplete multi-view clustering method for the data with large scales, termed Efficient Incomplete Multi-view clustering via flexible anchor Learning (EIML), where graph construction, anchor learning and graph partition are simultaneously integrated into a unified framework for efficient incomplete multi-view clustering. To be specific, we learn a shared anchor graph to guarantee the consistency among multiple views and employ a adaptive weight coefficient to balance the impact for each view. The relation between anchor graph and similarity matrix in symmetric nonnegative matrix factorization can also be built, i.e., each entry in the anchor graph can characterize the similarity between the anchor and original data sample. We then adopt an alternative algorithm for solving the formulated problem. Experiments conducted on different datasets confirm the superiority of EIML compared with other clustering methods for incomplete multi-view data.", "keywords": [ "Multi-view clustering", "anchor learning", "fast clustering" ], "rating_list": [ 1, 6, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 3, 2, 1 ], "contribution_list": [ 1, 3, 2, 1 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "GFua0WEYGF", "primary_area": "learning theory", "n_author": 1, "author_info_list": [ { "name": "Hadi Daneshmand", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Assistant Professor" } ], "n_formula": 40, "n_formula_1": 34, "n_ref_uni": 28, "n_ref": 44, "n_ref_all": 50, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3783, "formula_len_all_1": 2824, "len_all": 90644, "len_all_1": 59936, "len_abs": 1135, "len_title": 139, "len_sents": 26538, "len_sents_1": 23764, "n_sents": 260, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1140, "title": "Provable optimal transport with transformers: The essence of depth and prompt engineering", "abs": "Can we establish provable guarantees for transformer performance? Providing such theoretical guarantees is a milestone in developing trustworthy generative AI. In this paper, we take a step toward addressing this question by focusing on optimal transport, a fundamental problem at the intersection of combinatorial and continuous optimization. Leveraging the computational power of attention layers, we prove that a transformer with fixed parameters can effectively solve the optimal transport problem (in Wasserstein-2 with entropic regularization) for an arbitrary number of points. Consequently, the transformer can sort lists of arbitrary size up to an approximation factor. Our results rely on an engineered prompt that enables the transformer to implement gradient descent with adaptive step sizes on the dual optimal transport. Combining the convergence analysis of gradient descent with Sinkhorn dynamics, we establish an explicit approximation bound for optimal transport with transformers, which improves with increasing depth. Our findings provide novel insights into the essence of prompt engineering and depth for transformers.", "keywords": [ "Theory of Language Models", "Deep Learning Theory", "Optimal Transport", "Optimization" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "GFgn2LprFR", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Anthony Kobanda", "gender": "Male", "institution": "Inria Scool", "country": "FR", "position": "PhD student" }, { "name": "Emilien Bir\u00e9", "gender": "Male", "institution": "Centrale Sup\u00e9lec", "country": "FR", "position": "MS student" }, { "name": "Ludovic Denoyer", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "R\u00e9my Portelas", "gender": "Male", "institution": "Ubisoft", "country": "FR", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 17, "n_ref": 34, "n_ref_all": 70, "n_fig": 24, "n_tab": 5, "L_tab": 2356, "n_element_tab": 129, "n_fig_1": 16, "n_tab_1": 2, "L_tab_1": 1027, "n_element_tab_1": 97, "formula_len_all": 60, "formula_len_all_1": 451, "len_all": 111446, "len_all_1": 65628, "len_abs": 1365, "len_title": 68, "len_sents": 46449, "len_sents_1": 33350, "n_sents": 303, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1330, "title": "Efficient Active Imitation Learning with Random Network Distillation", "abs": "Developing agents for complex and underspecified tasks, where no clear objective exists, remains challenging but offers many opportunities. This is especially true in video games, where simulated players (bots) need to play realistically, and there is no clear reward to evaluate them. While imitation learning has shown promise in such domains, these methods often fail when agents encounter out-of-distribution scenarios during deployment. Expanding the training dataset is a common solution, but it becomes impractical or costly when relying on human demonstrations. This article addresses active imitation learning, aiming to trigger expert intervention only when necessary, reducing the need for constant expert input along training. We introduce Random Network Distillation DAgger (RND-DAgger), a new active imitation learning method that limits expert querying by using a learned state-based out-of-distribution measure to trigger interventions. This approach avoids frequent expert-agent action comparisons, thus making the expert intervene only when it is useful. We evaluate RND-DAgger against traditional imitation learning and other active approaches in 3D video games (racing and third-person navigation) and in a robotic locomotion task and show that RND-DAgger surpasses previous methods by reducing expert queries.", "keywords": [ "Active Imitation Learning", "Imitation Learning", "Interactive Learning", "Navigation" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "GFaplOjE7E", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Alireza Nadali", "gender": "Not Specified", "institution": "University of Colorado at Boulder", "country": "US", "position": "PhD student" }, { "name": "Ashutosh Trivedi", "gender": "Male", "institution": "University of Colorado at Boulder", "country": "", "position": "Associate Professor" }, { "name": "Majid Zamani", "gender": "Male", "institution": "University of Colorado at Boulder", "country": "US", "position": "Associate Professor" } ], "n_formula": 26, "n_formula_1": 28, "n_ref_uni": 51, "n_ref": 104, "n_ref_all": 130, "n_fig": 5, "n_tab": 5, "L_tab": 307, "n_element_tab": 31, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1072, "n_element_tab_1": 82, "formula_len_all": 2298, "formula_len_all_1": 1887, "len_all": 133701, "len_all_1": 66937, "len_abs": 1459, "len_title": 116, "len_sents": 35295, "len_sents_1": 29021, "n_sents": 265, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1475, "title": "On Choice of Loss Functions For Neural Control Barrier Certificates", "abs": "The design of controllers with correctness guarantees is a primary concern for safety-critical control systems. \nA Control Barrier Certificate (CBC) is a real-valued function over the state space of the system that provides an inductive proof of the existence of a safe controller. \nRecently, neural networks have been successfully deployed for data-driven learning of control barrier certificates. \nThese approaches encode the conditions for the existence of a CBC using a rectified linear unit (ReLU) loss function. \nThe resulting encoding, while sound, tends to be conservative, which results in slower training and limits scalability to large, complex systems. \nCan altering the loss function alleviate some of the problems associated with ReLU loss and lead to faster learning?\n\nThis paper proposes a novel encoding with a Mean Squared Error (MSE) loss function, which allows for more scalable and efficient training, while addressing some of the theoretical limitations of previous methods. \nThe proposed approach derives a validity condition based on Lipschitz continuity to formally characterize safety guarantees, eliminating the need for a post-hoc verification. \nThe effectiveness of the proposed loss functions is demonstrated through six case studies curated from the existing state of the art.\nOur results provide a compelling argument for exploring alternative loss function choices as a novel approach to optimizing the design of control barrier certificates.", "keywords": [ "Safe Learning for Control; Control Systems; Neural Control Barrier Certificates" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "GF6UrrTWp1", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Jung-Eun Kim", "gender": "Female", "institution": "North Carolina State University", "country": "US", "position": "Assistant Professor" }, { "name": "Varun Mulchandani", "gender": "Male", "institution": "North Carolina State University", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 17, "n_ref": 41, "n_ref_all": 57, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12, "formula_len_all_1": 12, "len_all": 93758, "len_all_1": 57970, "len_abs": 1998, "len_title": 84, "len_sents": 29320, "len_sents_1": 29302, "n_sents": 192, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 36, "L_abs": 2004, "title": "The Invariance Starvation Hypothesis", "abs": "Deep neural networks are known to learn and rely on spurious correlations during training, preventing them from being reliable and able to solve highly complex problems. While there exist many proposed solutions that overcome such reliance in different, tailored settings, current understanding regarding the formation of spurious correlations is limited. All proposed solutions with promising results assume that networks trained with empirical risk minimization will learn spurious correlations due to a preference for simpler features and that a solution to this problem requires further processing on the networks' learned representations or re-training on a modified dataset where the proportion of training data with spurious features is significantly lower. In this paper, we aim to form a better understanding regarding the formation of spurious correlations by performing a rigorous study regarding the role that data plays in the formation of spurious correlations. We show that in reasoning tasks with simple input samples, simply drawing more data from the same training distribution overcomes spurious correlations, even though we maintain the proportion of samples with spurious features. In other words, we find that if the network has enough data to encode the invariant function appropriately, it no longer relies on spurious features, regardless of its strength. We observe the same results in settings with more complex distributions with an intractable number of participating features, such as vision and language. However, we find that in such settings, drawing more samples from the training distribution while maintaining proportion can exacerbate spurious correlations at times, due to the introduction of new samples that are significantly different from samples in the original training set. Taking inspiration from reasoning tasks, we present an effective remedy to this problem to ensure that drawing more samples from the distribution always overcomes spurious correlations.", "keywords": [ "Spurious Correlations", "Reasoning", "Robustness" ], "rating_list": [ 3, 3, 3, 3, 1 ], "soundness_list": [ 1, 1, 2, 2, 1 ], "presentation_list": [ 2, 1, 3, 4, 2 ], "contribution_list": [ 2, 2, 2, 3, 1 ], "confidence_list": [ 4, 4, 4, 5, 5 ] }, { "paper_id": "GF1sRSBiwY", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Hwijae Son", "gender": "Male", "institution": "Konkuk University", "country": "KR", "position": "Assistant Professor" }, { "name": "Hyung Ju Hwang", "gender": "unknown", "institution": "POSTECH", "country": "", "position": "Full Professor" }, { "name": "Hyunwoo Cho", "gender": "Male", "institution": "Pohang University of Science and Technology", "country": "", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 25, "n_ref_uni": 10, "n_ref": 18, "n_ref_all": 28, "n_fig": 7, "n_tab": 2, "L_tab": 343, "n_element_tab": 49, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 292, "n_element_tab_1": 32, "formula_len_all": 3342, "formula_len_all_1": 3176, "len_all": 58415, "len_all_1": 47469, "len_abs": 1213, "len_title": 109, "len_sents": 17682, "len_sents_1": 16150, "n_sents": 146, "n_sents_1": 133, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1218, "title": "Physics-Informed Neural Networks with Message-Passing Weights", "abs": "Adaptive loss balancing algorithms play a crucial role in improving the performance of Physics-Informed Neural Networks (PINNs) by effectively managing the weights assigned to different loss components. Most notably, Wang et al. (2022) introduced Causal Physics-Informed Neural Networks (Causal PINNs), which achieve superior performance by simply reformulating the loss function based on the causal structure that emerges from time dependency. However, despite their empirical success, a solid theoretical analysis for the effectiveness of Causal PINNs has not received adequate attention. This paper addresses this gap by providing a theoretical rationale for Causal PINNs through the Belief Propagation (BP) algorithm, which is commonly used for causal inference. In addition, motivated by this analysis, we propose a Message Passing PINNs (MP-PINNs), a novel adaptive weighting algorithm. Through extensive numerical experiments, we demonstrate that the proposed MP-PINNs significantly outperform existing adaptive weighting methods, exhibiting superior performance in solving complex PDEs. Our findings highlight the potential of MP-PINNs as a powerful tool to enhance both the accuracy \nand efficiency of PINNs.", "keywords": [ "Physics-informed Neural Networks", "Adaptive Loss-Balancing Algorithms", "causal PINN", "Belief Propagation" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 1, 1, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 2, 2, 4, 3 ] }, { "paper_id": "GEBkyKZOc4", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Chong Liu", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "Postdoc" }, { "name": "Maosong Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Shizuo Tian", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Xin Cong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Postdoc" }, { "name": "Yankai Lin", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "Assistant Professor" }, { "name": "Yujia Qin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhiyuan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Ye Yi ning", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" } ], "n_formula": 15, "n_formula_1": 10, "n_ref_uni": 36, "n_ref": 95, "n_ref_all": 120, "n_fig": 2, "n_tab": 14, "L_tab": 3145, "n_element_tab": 200, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 1079, "n_element_tab_1": 158, "formula_len_all": 622, "formula_len_all_1": 477, "len_all": 161164, "len_all_1": 72433, "len_abs": 1393, "len_title": 119, "len_sents": 50928, "len_sents_1": 32943, "n_sents": 366, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1398, "title": "Rational Decision-Making Agent with Learning Internal Utility Judgment", "abs": "With remarkable advancements, large language models (LLMs) have attracted significant efforts to develop LLM-based agents capable of executing intricate multi-step decision-making tasks. Existing approaches predominantly build upon the external performance measure to guide the decision-making process but the reliance on the external performance measure as prior is problematic in real-world scenarios, where such prior may be unavailable, flawed, or even erroneous. For genuine autonomous decision-making for LLM-based agents, it is imperative to develop rationality from their posterior experiences to judge the utility of each decision independently. In this work, we propose RaDAgent (Rational Decision-Making Agent), which fosters the development of its rationality through an iterative framework involving Experience Exploration and Utility Learning. Within this framework, Elo-based Utility Learning is devised to assign Elo scores to individual decision steps to judge their utilities via pairwise comparisons. Consequently, these Elo scores guide the decision-making process to derive optimal outcomes. Experimental results on the Game of 24, WebShop, ToolBench and RestBench datasets demonstrate RaDAgent\u2019s superiority over baselines, achieving about 7.8% improvement on average. Besides, RaDAgent also can reduce costs (ChatGPT API calls), highlighting its effectiveness and efficiency.", "keywords": [ "Decision Making", "Autonomous Agent", "Large Lanugage Model", "Elo Rating" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 3, 2 ] }, { "paper_id": "GE6iywJtsV", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Bohao Li", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Hongming Chen", "gender": "Male", "institution": "Guangzhou National Laboratory", "country": "", "position": "Principal Investigator" }, { "name": "Jie Lin", "gender": "Female", "institution": "Guangzhou Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Mingyuan Xu", "gender": "Male", "institution": "Guangzhou National Laboratory", "country": "", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 30, "n_ref": 39, "n_ref_all": 61, "n_fig": 10, "n_tab": 3, "L_tab": 1239, "n_element_tab": 172, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1187, "n_element_tab_1": 160, "formula_len_all": 1050, "formula_len_all_1": 909, "len_all": 95716, "len_all_1": 53621, "len_abs": 1193, "len_title": 131, "len_sents": 30899, "len_sents_1": 24812, "n_sents": 185, "n_sents_1": 150, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1196, "title": "Diff-Shape: A Novel Constrained Diffusion Model for Shape-based De Novo Drug Design", "abs": "Shape-based virtual screening is a widely utilized method in ligand-based de novo drug design, aiming to identify molecules in chemical libraries that share similar 3D shapes but simultaneously possess novel 2D chemical structures compared to the reference compound. As an emerging technology, generative model is an alternative way to do de novo drug design by directly generating 3D novel structures. However, existing models face challenges in reliably generating valid drug-like molecules under specific conformation constrains. Here, a novel diffusion model constrained with 3D reference shape, Diff-Shape, was proposed to generate structures whose 3D conformations are similar to a given reference shape, thereby avoiding the computational cost of screening large database of 3D conformations. This model utilized a zero-weighted graph control module, taking in various forms of point clouds of reference shape to guide diffusion process of 3D molecular generation. The results show that our model is capable of generating molecules with high shape similarity but still low 2D graph similarity to the query structure and it significantly out-performs existing shape based generative models.", "keywords": [ "Moleuclar generative model; Shape-based virtual screening; De Novo drug design;" ], "rating_list": [ 1, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "GDjwSBZy6l", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Dung Tran", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Kazuhito Koishida", "gender": "unknown", "institution": "Microsoft Corporation", "country": "US", "position": "Researcher" }, { "name": "Luming Liang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Saeed Amizadeh", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Ta Duy Nguyen", "gender": "unknown", "institution": "Boston University", "country": "US", "position": "PhD student" }, { "name": "Tianyi Chen", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Senior Researcher" }, { "name": "Trung Viet Dang", "gender": "Male", "institution": "Hume AI", "country": "US", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 27, "n_ref": 55, "n_ref_all": 84, "n_fig": 4, "n_tab": 4, "L_tab": 277, "n_element_tab": 53, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 90, "n_element_tab_1": 35, "formula_len_all": 399, "formula_len_all_1": 366, "len_all": 109311, "len_all_1": 69440, "len_abs": 1431, "len_title": 122, "len_sents": 33945, "len_sents_1": 31558, "n_sents": 285, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1439, "title": "ROLoRA: Rank Optimization for Low-Rank Adaptation under Memory Constraints", "abs": "Low-Rank Adaptation (LoRA) has emerged as a prominent technique for fine-tuning large language models (LLMs) with limited computational resources. However, by injecting low-rank adapters with a rank identical across all layers, standard LoRA overlooks the varying importance of the weight matrices, often leading to suboptimal performance. Therefore, discovering an optimal rank configuration that efficiently utilizes limited training resources remains an open question. Existing solutions typically compromises computational constraints for performance gains, limiting their practical usage in resource-constrained scenarios. To address these issues, in this paper, we propose a novel method named ROLoRA to efficiently discover an effective rank configuration for low-rank adaptation, while strictly adhering to a constrained computational budget during training. In particular, our method iteratively prunes saturated adapters and expands under-fitted ones to increase their capacity until they converge to a highly optimized configuration. Our approach is delicately designed within the Frank-Wolfe algorithmic framework, which offers potential theoretical guarantees. Experimentally, we demonstrate that ROLoRA outperforms standard LoRA on common natural language processing tasks, including the GLUE and SQuAD benchmarks. Additionally, we provide a comprehensive analysis to explain why ROLoRA surpasses competing state-of-the-arts.", "keywords": [ "low rank adaptation", "fine-tuning" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "GDf7vWs701", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "YuTi Lin", "gender": "Female", "institution": "National Taiwan University", "country": "TW", "position": "MS student" }, { "name": "Yuh-Jzer Joung", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Full Professor" }, { "name": "Hao Yu Chen", "gender": "Male", "institution": "iKala", "country": "TW", "position": "Data Scientist" }, { "name": "Patrick Lee", "gender": "Male", "institution": "iKala Interactive Media Inc.", "country": "", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 27, "n_ref": 42, "n_ref_all": 63, "n_fig": 11, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2800, "n_element_tab_1": 129, "formula_len_all": 127, "formula_len_all_1": 143, "len_all": 88410, "len_all_1": 54922, "len_abs": 1373, "len_title": 110, "len_sents": 29672, "len_sents_1": 23825, "n_sents": 221, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 110, "L_abs": 1381, "title": "ProdInfluencerNet: A Novel Product-Centric Influencer Recommendation Framework Based on Heterogeneous Networks", "abs": "With the proliferation of social media, influencer marketing has emerged as a popular strategy for brands to promote their products. Recent studies have increasingly explored the use of machine learning to recommend suitable influencers for brands. This typically involves analyzing the compatibility of influencer profiles with brand attributes. However, for brands entering new markets or promoting products in unfamiliar categories, existing solutions may be limited due to insufficient information for accurate compatibility matching.\n\nIn this paper, we propose ProdInfluencerNet (PIN), a product-centric framework designed for influencer recommendation. PIN effectively models the complex relationships between brands, products, and influencers using Heterogeneous Information Networks (HINs). We categorize sponsored post images using the Google Taxonomy through image classification techniques. By leveraging the taxonomy's hierarchical structure and adopting an inductive learning approach, PIN can accurately recommend influencers for brands, even in new markets or with innovative products. We validate PIN's effectiveness and superiority over existing methods using two Instagram datasets. Furthermore, our analysis reveals that text features in profiles are more critical than images for identifying cooperative relationships between product categories and influencers.", "keywords": [ "Influencer Marketing", "Influecner Recommnedation", "Heterogeneous Information Network", "Inductive Learning", "ProdInfluencerNet" ], "rating_list": [ 3, 5, 1, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "GDd5H92egZ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Abhranil Chandra", "gender": "unknown", "institution": "University of Waterloo", "country": "", "position": "MS student" }, { "name": "Manish Gupta", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Principal Researcher" }, { "name": "Pawan Goyal", "gender": "Male", "institution": "IIT Kharagpur", "country": "", "position": "Associate Professor" }, { "name": "Sreevatsa Muppirala", "gender": "Male", "institution": "Indian Institute of Technology, Kharagpur", "country": "", "position": "MS student" }, { "name": "Yaswanth Narsupalli", "gender": "Male", "institution": "IIT Kharagpur", "country": "IN", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 3, "n_ref_uni": 31, "n_ref": 99, "n_ref_all": 124, "n_fig": 6, "n_tab": 18, "L_tab": 6857, "n_element_tab": 441, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2695, "n_element_tab_1": 190, "formula_len_all": 518, "formula_len_all_1": 216, "len_all": 267240, "len_all_1": 65947, "len_abs": 1439, "len_title": 117, "len_sents": 94952, "len_sents_1": 30882, "n_sents": 834, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1396, "title": "ReFeR: Improving Evaluation and Reasoning through Hierarchy of Models", "abs": "Assessing the quality of outputs generated by generative models, such as large language models and vision language models, presents notable challenges. Traditional methods for evaluation typically rely on either human assessments, which are resource-intensive, or automatic metrics that often show a low correlation with human judgment. Another common approach is to use deep learning systems, which not only consume a substantial amount of compute and time but also require extensive training data. In this study, we introduce a tuning-free framework called ReFeR, designed to evaluate generative outputs, including both text and images, by leveraging a 2-level hierarchy of LLMs and VLMs themselves. We rigorously evaluate our framework, ReFeR, across four diverse evaluation tasks. The framework not only improves the accuracy of these evaluations, surpassing previous benchmarks but also generates constructive feedback. Interestingly, the framework is also applicable to reasoning tasks. Experiments on four reasoning tasks demonstrate superior collective reasoning abilities of the framework. We present two variants of the framework: ReFeR-Turbo, optimized for accelerated performance, and ReFeR-Lite, offering a more test-time compute efficient solution. ReFeR-Lite is $\\sim12-14\\times$ more test-time compute efficient than previous works while being comparably accurate to ReFeR-Turbo.", "keywords": [ "Multi-agent framework", "reasoning", "evaluation", "multimodal", "hierarchy", "LLMs", "VLMs" ], "rating_list": [ 6, 6, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "GDZeeCZ3MM", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "JIAN XU", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "John Paisley", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "Associate Professor" }, { "name": "Delu zeng", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 24, "n_formula_1": 26, "n_ref_uni": 27, "n_ref": 51, "n_ref_all": 76, "n_fig": 2, "n_tab": 7, "L_tab": 1922, "n_element_tab": 171, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1581, "n_element_tab_1": 144, "formula_len_all": 1269, "formula_len_all_1": 1368, "len_all": 97803, "len_all_1": 61952, "len_abs": 1705, "len_title": 87, "len_sents": 29166, "len_sents_1": 26587, "n_sents": 206, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1712, "title": "IMPLICIT VARIATIONAL REJECTION SAMPLING", "abs": "Variational Inference (VI) is a cornerstone technique in Bayesian machine learning, employed to approximate complex posterior distributions. However, traditional VI methods often rely on mean-field assumptions, which may inadequately capture the true posterior's complexity. To address this limitation, recent advancements have utilized neural networks to model implicit distributions, thereby offering increased flexibility. Despite this, the practical constraints of neural network architectures can still result in inaccuracies in posterior approximations. In this work, we introduce a novel method called Implicit Variational Rejection Sampling (IVRS), which integrates implicit distributions with rejection sampling to enhance the approximation of the posterior distribution. Our method employs neural networks to construct implicit proposal distributions and utilizes rejection sampling with a meticulously designed acceptance probability function. A discriminator network is employed to estimate the density ratio between the implicit proposal and the true posterior, thereby refining the approximation. We propose the Implicit Resampling Evidence Lower Bound (IR-ELBO) as a metric to characterize the quality of the resampled distribution, enabling the derivation of a tighter variational lower bound. Experimental results demonstrate that our method outperforms traditional variational inference techniques in terms of both accuracy and efficiency, leading to significant improvements in inference performance. This work not only showcases the effective combination of implicit distributions and rejection sampling but also offers a novel perspective and methodology for advancing variational inference.", "keywords": [ "Varaitional inference", "reject sampling", "implict distribution" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 4, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "GDS5eN65QY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "En Yu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jinyang Li", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Sijia Chen", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Wenbing Tao", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 23, "n_ref": 52, "n_ref_all": 62, "n_fig": 8, "n_tab": 11, "L_tab": 2460, "n_element_tab": 360, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 396, "n_element_tab_1": 12, "formula_len_all": 738, "formula_len_all_1": 578, "len_all": 147526, "len_all_1": 65292, "len_abs": 1308, "len_title": 122, "len_sents": 53458, "len_sents_1": 31950, "n_sents": 371, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1314, "title": "OVTR: End-to-End Open-Vocabulary Multiple Object Tracking with Transformer", "abs": "Open-vocabulary multiple object tracking aims to generalize trackers to unseen categories during training, enabling their application across a variety of real-world scenarios. However, the existing open-vocabulary tracker, which relies on off-the-shelf open-vocabulary detector, perceives categories and locations independently in each frame, causing instability and making it vulnerable to similar appearances and irregular motion in diverse scenes. In this paper, we propose OVTR (End-to-End Open-Vocabulary Multiple Object Tracking with TRansformer), the first end-to-end open-vocabulary tracker that models motion, appearance, and category simultaneously. To achieve stable classification and continuous tracking, we designed the CIP (Category Information Propagation) strategy, which establishes multiple high-level category information priors for subsequent frames. Additionally, we introduce a dual-branch structure for generalization capability and deep multimodal interaction, and incorporate protective strategies in the decoder to enhance performance. Notably, our method does not require proposals that contain novel categories, yet still achieves strong results on the open-vocabulary MOT benchmark. Moreover, experiment transferring the model to other dataset demonstrates its effective adaptability.", "keywords": [ "Multiple-Object Tracking", "Transformer", "End-to-End", "Open-Vocabulary" ], "rating_list": [ 6, 3, 6, 8 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 3, 5, 5, 5 ] }, { "paper_id": "GDDqq0w6rs", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Eden Zohar", "gender": "unknown", "institution": "IBM TJ Watson Research Center", "country": "", "position": "Researcher" }, { "name": "Matan Ninio", "gender": "Male", "institution": "International Business Machines", "country": "IL", "position": "Researcher" }, { "name": "Michael Danziger", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Yishai Shimoni", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Principal Researcher" }, { "name": "Yoav Kan-Tor", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 37, "n_ref": 59, "n_ref_all": 78, "n_fig": 7, "n_tab": 7, "L_tab": 8889, "n_element_tab": 295, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1880, "n_element_tab_1": 166, "formula_len_all": 95, "formula_len_all_1": 95, "len_all": 180562, "len_all_1": 70967, "len_abs": 2753, "len_title": 144, "len_sents": 44272, "len_sents_1": 29743, "n_sents": 260, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 2768, "title": "Does your model understand genes? A benchmark of gene properties for biological and text models", "abs": "The application of deep learning for biology, including foundation models, has increased significantly in recent years.\nSome models are text-based, while others are trained on the underlying biological data, especially omics data of various modalities. \nConsistently comparing the performance of deep learning models for biology has proven challenging due to the diversity of training data and downstream tasks. \nHere, we utilize the fact that many models operate on the level of genes and propose a unifying benchmark by defining hundreds of tasks based on ground-truth gene properties collected from professionally curated bioinformatics databases. \nWe collect properties of five types: (1) genomic properties, including predicting which genes can be methylated or which are dose-dependent; (2) regulatory functions, evaluating how the genes participate in cellular regulatory processes; (3) localization, including identification of differential expression in different tissues or sub-cellular localization; (4) biological processes, including predicting gene involvement in pathways or disease prognostics; and (5) protein properties, including prediction of functional domains or post-translational modifications.\nThese properties are used to define binary, multi-label and multi-class classification tasks.\nTo create an architecture-agnostic benchmark we extract gene representation vectors from each model, including single-cell RNA-seq (scRNA) foundation models, large language models, protein language models, DNA foundation models, and classical baselines, and use them to train simple predictive models on the tasks.\nDepending on the model, we utilize the model's token-level embeddings of gene symbols or transform the gene symbol to an input appropriate for the model, i.e. a description of the gene for text models, the gene sequence for DNA models or amino acid sequences for the protein models.\nUsing these embeddings on the benchmark tasks, we create a detailed assessment of the relative performance of the different models.\nIn general, we find that text-based models and protein language models outperform the expression-based models on tasks related to genomic properties and regulatory functions, while expression-based models tend to outperform the others on localization tasks.\nWe also observe performance for the classical bag-of-words baseline that is similar to the large language models for many tasks.\nBy enabling broad systematic evaluation of diverse deep learning models in biology, this benchmark can help direct future research in artificial intelligence toward improved biological understanding and accelerated therapeutic discoveries.\nThe code and benchmark data can be extended to more models and tasks and is available on GitHub.", "keywords": [ "Benchmark", "Data Sets or Data Repositories", "Computational Biology and Bioinformatics" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 3, 2, 4 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "GD4Tlqvwrq", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Chenming Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Fan Wang", "gender": "Female", "institution": "Alibaba Group", "country": "US", "position": "Senior Staff Algorithm Engineer" }, { "name": "Hao Wang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Jianguo Zhang", "gender": "Male", "institution": "Southern University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jingkai Zhou", "gender": "Male", "institution": "Alibaba DAMO Academy", "country": "CN", "position": "Researcher" }, { "name": "Weihua Chen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Algorithm Engineer" }, { "name": "Wenjian Huang", "gender": "Male", "institution": "Southern University Of Science and Technology", "country": "CN", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 8, "n_ref_uni": 31, "n_ref": 103, "n_ref_all": 130, "n_fig": 7, "n_tab": 9, "L_tab": 1628, "n_element_tab": 137, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 601, "n_element_tab_1": 22, "formula_len_all": 855, "formula_len_all_1": 469, "len_all": 134855, "len_all_1": 65671, "len_abs": 1514, "len_title": 100, "len_sents": 41021, "len_sents_1": 31694, "n_sents": 286, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1525, "title": "Noise Re-sampling for High Fidelity Image Generation", "abs": "Latent diffusion models (LDMs) have emerged as powerful tools for generating diverse and realistic samples across domains. However, their efficacy in capturing intricate details and small-scale objects remains a challenge. \nOur investigation reveals that VAE compression induces errors in the latent space and limits the generation quality. Furthermore, LDMs trained on fixed-resolution images struggle to produce high-resolution outputs without distortions, making simple resolution increases ineffective.\nIn this paper, we propose a novel **noise re-sampling** strategy that enables multi-scale generation of LDMs, allowing LDMs to \"zoom in\" and improve generation quality of local regions. By increasing the sampling rates from the noise perspective in the latent space, we effectively bypass the constraints imposed by VAE compression, thus preserving crucial high-frequency information. Our approach, a simple yet effective plugin for current LDMs, enhances the quality of image generation in local regions while maintaining overall structural consistency and providing fine-grained control over the scale of generation in latent diffusion models.\nThrough extensive experimentation and evaluation, we demonstrate the efficacy of our method in enhancing the generation quality across various LDM architectures. Our approach surpasses existing methods, including stable diffusion (SD) models, SD-based super-resolution methods and high-resolution adaptation methods, in generating high-fidelity samples of complex objects.", "keywords": [ "Generative models", "Diffusion based models" ], "rating_list": [ 6, 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 1, 3, 3, 1, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "GCzpUJO5rx", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Mudit Gaur", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Raghu Pasupathy", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "Full Professor" }, { "name": "Vaneet Aggarwal", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Full Professor" }, { "name": "Amrit Bedi", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" } ], "n_formula": 112, "n_formula_1": 44, "n_ref_uni": 36, "n_ref": 81, "n_ref_all": 118, "n_fig": 0, "n_tab": 3, "L_tab": 910, "n_element_tab": 90, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 628, "n_element_tab_1": 50, "formula_len_all": 8250, "formula_len_all_1": 3305, "len_all": 156074, "len_all_1": 63965, "len_abs": 1956, "len_title": 116, "len_sents": 33137, "len_sents_1": 25999, "n_sents": 305, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1991, "title": "On the Global Convergence of RLHF Based Alignment With Neural Parametrization", "abs": "The importance of Reinforcement Learning from Human Feedback (RLHF) in aligning large language models (LLMs) with human values cannot be overstated. RLHF is a three-stage process that includes supervised fine-tuning (SFT), reward learning, and policy learning. Although there are several offline and online approaches to aligning LLMs, they often suffer from distribution shift issues. These issues arise from the inability to accurately capture the distributional interdependence between the reward learning and policy learning stages. Consequently, this has led to various approximated approaches, but the theoretical insights and motivations remain largely limited to tabular settings, which do not hold in practice.\nThis gap between theoretical insights and practical implementations is critical. It is challenging to address this gap as it requires analyzing the performance of AI alignment algorithms in neural network-parameterized settings. Although bi-level formulations have shown promise in addressing distribution shift issues, they suffer from the hyper-gradient problem, and current approaches lack efficient algorithms to solve this. \nIn this work, we tackle these challenges employing the bi-level formulation laid out in Kwon et al. (2024) along with the assumption *Weak Gradient Domination* to demonstrate convergence in an RLHF setup, obtaining a sample complexity of $\\epsilon^{-\\frac{7}{2}}$ . Our key contributions are twofold: (i) We propose a bi-level formulation for AI alignment in parameterized settings and introduce a first-order approach to solve this problem. (ii) We analyze the theoretical convergence rates of the proposed algorithm and derive state-of-the-art bounds. To the best of our knowledge, this is the first work to establish convergence rate bounds and global optimality for the RLHF framework in neural network-parameterized settings. Our contributions are primarily theoretical, providing crucial insights for future practical implementations.", "keywords": [ "Reinforcement Learning from Human Feedback (RLHF)" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 2, 3 ] }, { "paper_id": "GCH5leffZp", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Kookjin Lee", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" }, { "name": "Minju Jo", "gender": "Female", "institution": "LG CNS", "country": "KR", "position": "Researcher" }, { "name": "Noseong Park", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Woojin Cho", "gender": "Male", "institution": "Telepix", "country": "KR", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 48, "n_ref": 90, "n_ref_all": 140, "n_fig": 15, "n_tab": 22, "L_tab": 11954, "n_element_tab": 1230, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2150, "n_element_tab_1": 266, "formula_len_all": 1203, "formula_len_all_1": 859, "len_all": 213122, "len_all_1": 61387, "len_abs": 1163, "len_title": 93, "len_sents": 61790, "len_sents_1": 25705, "n_sents": 431, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 45, "L_abs": 1167, "title": "Neural Functions for Learning Periodic Signal", "abs": "As function approximators, deep neural networks have served as an effective tool to represent various signal types. Recent approaches utilize multi-layer perceptrons (MLPs) to learn a nonlinear mapping from a coordinate to its corresponding signal, facilitating the learning of continuous neural representations from discrete data points. Despite notable successes in learning diverse signal types, coordinate-based MLPs often face issues of overfitting and limited generalizability beyond the training region, resulting in subpar extrapolation performance. This study addresses scenarios where the underlying true signals exhibit periodic properties, either spatially or temporally. We propose a novel network architecture, which extracts periodic patterns from measurements and leverages this information to represent the signal, thereby enhancing generalization and improving extrapolation performance. We demonstrate the efficacy of the proposed method through comprehensive experiments, including the learning of the periodic solutions for differential equations, and time series imputation (interpolation) and forecasting (extrapolation) on real-world datasets.", "keywords": [ "Periodicity", "Neural function" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "GBpKUnM6gW", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Guorong Wu", "gender": "Male", "institution": "University of North Carolina, Chapel Hill", "country": "US", "position": "Associate Professor" }, { "name": "Jiaqi Ding", "gender": "Female", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "", "position": "PhD student" }, { "name": "Paul J. Laurienti", "gender": "unknown", "institution": "Wake Forest Baptist Health", "country": "", "position": "Full Professor" }, { "name": "Tingting Dan", "gender": "Female", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Postdoc" }, { "name": "Ziquan Wei", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 67, "n_ref": 83, "n_ref_all": 122, "n_fig": 6, "n_tab": 9, "L_tab": 6262, "n_element_tab": 520, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2855, "n_element_tab_1": 331, "formula_len_all": 66, "formula_len_all_1": 19, "len_all": 195812, "len_all_1": 67774, "len_abs": 1526, "len_title": 149, "len_sents": 47618, "len_sents_1": 30598, "n_sents": 332, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1533, "title": "Machine Learning Benchmark on Dynamic Functional Connectivity: Promise, Pitfalls, and Interpretations", "abs": "An unprecedented amount of existing functional Magnetic Resonance Imaging (fMRI) data provides a new opportunity to understand the relationship between functional fluctuation and human cognition/behavior using a data-driven approach. To that end, tremendous efforts have been made in machine learning to predict cognitive states from evolving volumetric images of blood-oxygen-level-dependent (BOLD) signals. Due to the complex nature of brain function, however, the evaluation on learning performance and discoveries are not often consistent across current state-of-the-arts (SOTA). By capitalizing on large-scale existing neuroimaging data (39,784 data samples from seven databases), we seek to establish a well-founded empirical guideline for designing deep models for functional neuroimages by linking the methodology underpinning with knowledge from the neuroscience domain. Specifically, we put the spotlight on (1) What is the current SOTA performance in cognitive task recognition and disease diagnosis using fMRI? (2) What are the limitations of current deep models? and (3) What is the general guideline for selecting the suitable machine learning backbone for new neuroimaging applications?\nWe have conducted a comprehensive evaluation and statistical analysis, in various settings, to answer the above outstanding questions. In addition, we explore a novel attention learning mechanism to provide meaningful spatial pattern of brain activation that is associated with various cognitive tasks and neurological conditions.", "keywords": [ "Brain Dynamics", "Functional Connectivity", "Machine Learning", "Task Recognition", "Disease Diagnosis" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "GBfYgjOfSe", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Di Feng", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Haotian Zhang", "gender": "Male", "institution": "Apple AI/ML", "country": "US", "position": "Researcher" }, { "name": "Harsh Agrawal", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Jeffrey Nichols", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "Keen You", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Mohana Prasad Sathya Moorthy", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Xiujun Li", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Yinfei Yang", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Zhangheng LI", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Zhe Gan", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 34, "n_ref": 137, "n_ref_all": 165, "n_fig": 10, "n_tab": 8, "L_tab": 2865, "n_element_tab": 305, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1214, "n_element_tab_1": 139, "formula_len_all": 316, "formula_len_all_1": 388, "len_all": 160980, "len_all_1": 47614, "len_abs": 1184, "len_title": 128, "len_sents": 39109, "len_sents_1": 22085, "n_sents": 269, "n_sents_1": 141, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 80, "L_abs": 1193, "title": "Ferret-UI One: Mastering Universal User Interface Understanding Across Platforms", "abs": "Building a generalist model for user interface (UI) understanding is challenging due to various foundational issues, such as platform diversity, resolution variation, and data limitation. In this paper, we introduce Ferret-UI One, a multimodal large language model (MLLM) designed for universal UI understanding across a wide range of platforms, including iPhone, Android, iPad, Webpage, and AppleTV. \nBuilding on the foundation of Ferret-UI, Ferret-UI One introduces three key innovations: support for multiple platform types, high-resolution perception through adaptive scaling, and advanced task training data generation powered by GPT-4o with set-of-mark visual prompting. These advancements enable Ferret-UI One to perform complex, user-centered interactions, making it highly versatile and adaptable for the expanding diversity of platform ecosystems. Extensive empirical experiments on referring, grounding, user-centric advanced tasks (comprising 9 subtasks $\\times$ 5 platforms), GUIDE next-action prediction dataset, and GUI-World multi-platform benchmark demonstrate that Ferret-UI One significantly outperforms Ferret-UI, and also shows strong cross-platform transfer capabilities.", "keywords": [ "Multimodal LLM", "UI Understanding" ], "rating_list": [ 6, 8, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "GBWqZNoeIk", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Aashwin Ananda Mishra", "gender": "unknown", "institution": "SLAC National Laboratory", "country": "US", "position": "Project Scientist" }, { "name": "Christian Borgelt", "gender": "Male", "institution": "Paris-Lodron-University of Salzburg", "country": "", "position": "Full Professor" }, { "name": "Felix Petersen", "gender": "Not Specified", "institution": "Stanford University", "country": "", "position": "Postdoc" }, { "name": "Stefano Ermon", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" } ], "n_formula": 67, "n_formula_1": 20, "n_ref_uni": 48, "n_ref": 108, "n_ref_all": 137, "n_fig": 14, "n_tab": 9, "L_tab": 2024, "n_element_tab": 233, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 433, "n_element_tab_1": 73, "formula_len_all": 7007, "formula_len_all_1": 1361, "len_all": 196838, "len_all_1": 66643, "len_abs": 985, "len_title": 125, "len_sents": 60111, "len_sents_1": 29149, "n_sents": 541, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1031, "title": "Generalizing Stochastic Smoothing for Differentiation and Gradient Estimation", "abs": "We deal with the problem of gradient estimation for stochastic differentiable relaxations of algorithms, operators, simulators, and other non-differentiable functions. Stochastic smoothing conventionally perturbs the input of a non-differentiable function with a differentiable density distribution with full support, smoothing it and enabling gradient estimation. Our theory starts at first principles to derive stochastic smoothing with reduced assumptions, without requiring a differentiable density nor full support, and presenting a general framework for relaxation and gradient estimation of non-differentiable black-box functions $f:\\mathbb{R}^n\\to\\mathbb{R}^m$. We develop variance reduction for gradient estimation from 3 orthogonal perspectives. Empirically, we benchmark 6 distributions and up to 24 variance reduction strategies for differentiable sorting and ranking, differentiable shortest-paths on graphs, differentiable rendering for pose estimation, as well as differentiable cryo-electron tomography simulations.", "keywords": [ "differentiable", "sorting", "rendering", "ranking", "variance reduction", "continuous", "cryo-et", "differentiable simulator", "shortest-path", "stochastic relaxation" ], "rating_list": [ 6, 3, 8, 3 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 4, 2, 4, 4 ], "contribution_list": [ 3, 1, 3, 1 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "GBIUbwW9D8", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Baolin Peng", "gender": "Male", "institution": "Research, Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Hao Cheng", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Researcher" }, { "name": "Jianfeng Gao", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Michel Galley", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Vineeth Vajipey", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "MS student" }, { "name": "Xiao Yu", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" }, { "name": "Zhou Yu", "gender": "Female", "institution": "Columbia University", "country": "", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 41, "n_ref": 95, "n_ref_all": 130, "n_fig": 9, "n_tab": 12, "L_tab": 3987, "n_element_tab": 452, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 1651, "n_element_tab_1": 235, "formula_len_all": 628, "formula_len_all_1": 351, "len_all": 191103, "len_all_1": 70793, "len_abs": 1944, "len_title": 124, "len_sents": 53100, "len_sents_1": 31733, "n_sents": 333, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1916, "title": "Improving Autonomous AI Agents with Reflective Tree Search and Self-Learning", "abs": "Autonomous agents have demonstrated significant potential in automating complex multistep decision-making tasks. However, even state-of-the-art vision-language models (VLMs), such as GPT-4o, still fall short of human-level performance, particularly in intricate web environments and long-horizon planning tasks. To address these limitations, we introduce Reflective Monte Carlo Tree Search (R-MCTS), a novel test-time algorithm designed to enhance the ability of AI agents, e.g., powered by GPT-4o, to explore decision space on the fly.\nR-MCTS extends traditional MCTS by 1) incorporating contrastive reflection, allowing agents to learn from past interactions and dynamically improve their search efficiency; and 2) using multi-agent debate to provide reliable state evaluation. Moreover, we improve the agent's performance by fine-tuning GPT-4o through self-learning, using R-MCTS generated tree traversals without any human-provided labels. On the challenging VisualWebArena benchmark, our GPT-4o-based R-MCTS agent achieves a 6% to 30% relative improvement across various tasks compared to the previous state-of-the-art. Additionally, we show that the knowledge gained from test-time search can be effectively transferred back to GPT-4o via fine-tuning. The fine-tuned GPT-4o matches 97\\% of R-MCTS's performance while reducing compute usage by a factor of four at test time. Furthermore, qualitative results reveal that the fine-tuned GPT-4o model demonstrates the ability to explore the environment, evaluate a state, and backtrack to viable ones when it detects that the current state cannot lead to success. Moreover, our work demonstrates the compute scaling properties in both training - data collection with R-MCTS - and testing time. These results suggest a promising research direction to enhance VLMs' reasoning and planning capabilities for agentic applications via test-time search and self-learning.", "keywords": [ "AI agent", "tree search", "self-improvement" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "GARbxyCV13", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Gaoyue Zhou", "gender": "Female", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Hengkai Pan", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "MS student" }, { "name": "Lerrel Joseph Pinto", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Yann LeCun", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Director" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 60, "n_ref": 115, "n_ref_all": 132, "n_fig": 8, "n_tab": 9, "L_tab": 1609, "n_element_tab": 207, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 478, "n_element_tab_1": 94, "formula_len_all": 481, "formula_len_all_1": 361, "len_all": 181432, "len_all_1": 62823, "len_abs": 1741, "len_title": 126, "len_sents": 49638, "len_sents_1": 30181, "n_sents": 340, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1748, "title": "DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning", "abs": "The ability to predict future outcomes given control actions is fundamental for physical reasoning. However, such predictive models, often called world models, have proven challenging to learn and are typically developed for task-specific solutions with online policy learning. We argue that the true potential of world models lies in their ability to reason and plan across diverse problems using only passive data, without requiring online interactions with the environment. Concretely, we require world models to have the following three properties: 1) be trainable on offline, pre-collected trajectories, 2) support test-time behavior optimization, and 3) facilitate task-agnostic reasoning. To realize this, we present DINO World Model (DINO-WM), a new method to model visual dynamics without reconstructing the visual world. DINO-WM leverages spatial patch features pre-trained with DINOv2, enabling it to learn from offline behavioral trajectories by predicting future patch features. This design allows DINO-WM to achieve observational goals through action sequence optimization, facilitating task-agnostic behavior planning by treating desired goal patch features as prediction targets. We evaluate DINO-WM across various domains, including maze navigation, tabletop pushing, and particle manipulation. Our experiments demonstrate that DINO-WM can generate zero-shot behavioral solutions at test time without relying on expert demonstrations, reward modeling, or pre-learned inverse models. Notably, DINO-WM exhibits strong generalization capabilities compared to prior state-of-the-art work, adapting to diverse task families such as arbitrarily configured mazes, push manipulation with varied object shapes, and multi-particle scenarios.", "keywords": [ "World Models", "Planning", "Representation Learning" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "G9xhvGPtte", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 16, "author_info_list": [ { "name": "Bilge Soran", "gender": "Female", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Changsheng Zhao", "gender": "Male", "institution": "Meta Inc.", "country": "US", "position": "Researcher" }, { "name": "Chenchen Zhu", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Research Scientist" }, { "name": "Fanyi Xiao", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Florian Bordes", "gender": "Male", "institution": "Meta", "country": "CA", "position": "Researcher" }, { "name": "Hu Xu", "gender": "Male", "institution": "FAIR, AMI Foundation", "country": "US", "position": "Research Scientist" }, { "name": "Jun Chen", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" }, { "name": "Lemeng Wu", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Mohamed Elhoseiny", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Raghuraman Krishnamoorthi", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Vikas Chandra", "gender": "Male", "institution": "Meta", "country": "US", "position": "Director, AI" }, { "name": "Xiaoqian Shen", "gender": "Female", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Yunyang Xiong", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Zechun Liu", "gender": "unknown", "institution": "Meta Inc.", "country": "US", "position": "Researcher" }, { "name": "Balakrishnan Varadarajan", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Zhuang Liu", "gender": "Male", "institution": "FAIR, Meta AI", "country": "US", "position": "Research Scientist" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 49, "n_ref": 176, "n_ref_all": 211, "n_fig": 8, "n_tab": 5, "L_tab": 1124, "n_element_tab": 142, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2872, "n_element_tab_1": 75, "formula_len_all": 463, "formula_len_all_1": 255, "len_all": 167727, "len_all_1": 54526, "len_abs": 1369, "len_title": 130, "len_sents": 36031, "len_sents_1": 25062, "n_sents": 270, "n_sents_1": 144, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1330, "title": "LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding", "abs": "Multimodal Large Language Models (MLLMs) have shown promising progress in understanding and analyzing video content. However, processing long videos remains a significant challenge constrained by the limited context length. To address this limitation, we propose LongVU, a spatiotemporal adaptive compression mechanism to reduce the number of video tokens while preserving visual details of long videos. Our idea is based on leveraging cross-modal query and inter-frame dependencies to adaptively reduce temporal and spatial redundancy in videos. Specifically, we leverage DINOv2 features to remove redundant frames that exhibit high similarity. Then we utilize text-guided cross-modal query for selective frame feature reduction. Further, we perform spatial token reduction across frames based on their temporal dependencies. Our adaptive compression strategy effectively processes a large number of frames with little visual information loss within limited context length. Our LongVU consistently surpass existing methods across a variety of video understanding benchmarks, especially on hour-long video understanding tasks such as VideoMME and MLVU. Given a light-weight LLM, our LongVU also scales effectively into a smaller size with state-of-the-art video understanding performance. Our code will be made publicly available.", "keywords": [ "Long Video Understanding;Video-Language;Spatiotemporal" ], "rating_list": [ 5, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "G9qA1JZ0Sy", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Jingyang Qiao", "gender": "Male", "institution": "East China Normal University", "country": "", "position": "PhD student" }, { "name": "Shouhong Ding", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "researcher" }, { "name": "Xin Tan", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Yanyun Qu", "gender": "Female", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "zhizhong zhang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuan Xie", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" } ], "n_formula": 51, "n_formula_1": 19, "n_ref_uni": 31, "n_ref": 74, "n_ref_all": 133, "n_fig": 7, "n_tab": 10, "L_tab": 3621, "n_element_tab": 551, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 3170, "n_element_tab_1": 495, "formula_len_all": 3605, "formula_len_all_1": 1156, "len_all": 157487, "len_all_1": 73325, "len_abs": 1709, "len_title": 100, "len_sents": 42553, "len_sents_1": 28055, "n_sents": 374, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1718, "title": "LLaCA: Multimodal Large Language Continual Assistant", "abs": "Instruction tuning guides the Multimodal Large Language Models (MLLMs) in aligning different modalities by designing text instructions, which seems to be an essential technique to enhance the capabilities and controllability of foundation models. In this framework, Multimodal Continual Instruction Tuning (MCIT) is adopted to continually instruct MLLMs to follow human intent in sequential datasets. We observe existing gradient update would heavily destroy the tuning performance on previous datasets and the zero-shot ability during continual instruction tuning. Exponential Moving Average (EMA) update policy owns the ability to trace previous parameters, which can aid in decreasing forgetting. However, its stable balance weight cannot deal with the ever-changing datasets, leading to the out-of-balance between plasticity and stability of MLLMs. In this paper, we propose a method called Multimodal Large Language Continual Assistant (LLaCA) to address the challenge. Starting from the trade-off prerequisite and EMA update, we propose the plasticity and stability ideal condition. Based on Taylor expansion in the loss function, we find the optimal balance weight is basically according to the gradient information and previous parameters. We automatically determine the balance weight and significantly improve the performance. Through comprehensive experiments on LLaVA-1.5 in a continual visual-question-answering benchmark, compared with baseline, our approach not only highly improves anti-forgetting ability (with reducing forgetting from 22.67 to 2.68), but also significantly promotes continual tuning performance (with increasing average accuracy from 41.31 to 61.89). Our code will be published soon.", "keywords": [ "Multimodal Continual Instruction Tuning", "Anti-Forgetting", "Exponential Movement Average", "LoRA", "LLaVA" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 5, 4, 4 ] }, { "paper_id": "G9HV5upWhx", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Bingli Wang", "gender": "Male", "institution": "Sichuan Agricultural University", "country": "CN", "position": "Undergrad student" }, { "name": "Dianxi Shi", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Full Professor" }, { "name": "Jiao Li", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Meilin Zhang", "gender": "Female", "institution": "Lappeenranta University of Technology", "country": "", "position": "Undergrad student" }, { "name": "Mengzhu Wang", "gender": "Female", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Yin Nan", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Researcher" }, { "name": "Zeheng Wang", "gender": "Male", "institution": "Northeast Normal University", "country": "CN", "position": "MS student" }, { "name": "Houcheng Su", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 20, "n_ref": 35, "n_ref_all": 52, "n_fig": 6, "n_tab": 7, "L_tab": 3498, "n_element_tab": 401, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 783, "n_element_tab_1": 135, "formula_len_all": 1342, "formula_len_all_1": 1496, "len_all": 107164, "len_all_1": 51637, "len_abs": 1447, "len_title": 128, "len_sents": 28607, "len_sents_1": 21931, "n_sents": 198, "n_sents_1": 152, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1455, "title": "SgCG: Semantic-guided Contrastive Generalization for Medical Image Segmentation", "abs": "After training on the source domain, deep learning models often struggle to generalize effectively to unknown target domains with differing data distributions. This is an even more severe challenge when the target domain is not available. In this paper, we tackle the problem of domain-generalized medical image segmentation by introducing a novel semantic-guided contrastive generalization algorithm, termed SgCG. The method aligns different multi-source domains based on semantic distributions to learn domain-invariant features. \nSpecifically, we implement a novel contrastive generalization loss at the pixel level that incorporates semantic distributions from the source domains. This approach facilitates the clustering of pixel representations from the same category while effectively separating those from different categories, thereby improving the model's segmentation performance while learning domain-invariant features. Furthermore, we establish an upper bound estimation for the SgCG approach by integrating a contrastive generalization loss which include an infinite number of both similar and dissimilar pixel pairs. Despite the simplicity and straightforwardness of the approach, our empirical analysis reveals mechanisms that can maximize the potential of SgCG. We demonstrate the effectiveness of our approach using two public benchmarks for generalizable segmentation in medical images, where it achieves state-of-the-art performance.", "keywords": [ "Medical Image Segmentation;Semantic-Guided;Contrastive Generalization; Domain-invariant Feature" ], "rating_list": [ 3, 3, 1 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 3, 5, 5 ] }, { "paper_id": "G9Ea7mlqGO", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Leyuan Wang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "PhD student" }, { "name": "Liuyu Xiang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Assistant Professor" }, { "name": "Yujie Wei", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Yunlong Wang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Zhaofeng He", "gender": "Male", "institution": "Beijing University of Post and Telecommunication", "country": "CN", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 32, "n_ref": 58, "n_ref_all": 79, "n_fig": 8, "n_tab": 10, "L_tab": 4226, "n_element_tab": 313, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3298, "n_element_tab_1": 89, "formula_len_all": 610, "formula_len_all_1": 640, "len_all": 148047, "len_all_1": 61360, "len_abs": 1570, "len_title": 99, "len_sents": 42000, "len_sents_1": 28218, "n_sents": 296, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1577, "title": "CLIP model is an Efficient Online Continual Learner", "abs": "Online continual learning addresses the challenge of learning from continuous, non-stationary data streams. Existing online continual learning frameworks are classification-based and assume a pre-defined number of classes. In this study, we propose that vision-language models (VLMs) are more suitable candidates for online continual learning. Compared to traditional classification-based frameworks, VLM such as CLIP model is not limited by the maximum number of classes or constrained by rigid model architectures, enabling it to generalize across both known and emerging classes. However, we find that naively tuning the CLIP for online continual learning results in asymmetric image-text matching. This asymmetric matching will consistently poses negative suppression on the previously learned classes, leading to catestrophic forgetting. To address this issue, we propose a simple yet effective method, the symmetric image-text (SIT) tuning strategy, which mitigates the adverse impact of negative samples by excluding asymmetric text during online learning. Additionally, we introduce a more challenging online continual learning setting with blurred boundary, namely MiD-Blurry, which mixes multiple data distributions to simulate real-world scenarios. We conduct extensive experiments on several continual learning benchmarks as well as the MiD-Blurry setting, evaluating both inference-at-any-time performance and generalization to future data. Our results demonstrate that the SIT strategy effectively preserves memory stability while maintaining learning plasticity.", "keywords": [ "online continual learning", "vision-language models", "CLIP", "task-agnostic continual learning" ], "rating_list": [ 3, 5, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 3, 4, 2, 1, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 5, 3, 3 ] }, { "paper_id": "G8U2nGP3Vi", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Oren Mangoubi", "gender": "Male", "institution": "Worcester Polytechnic Institute", "country": "US", "position": "Assistant Professor" }, { "name": "Peiyao Lai", "gender": "Female", "institution": "Worcester Polytechnic Institute", "country": "US", "position": "Researcher" } ], "n_formula": 131, "n_formula_1": 53, "n_ref_uni": 24, "n_ref": 73, "n_ref_all": 97, "n_fig": 6, "n_tab": 1, "L_tab": 641, "n_element_tab": 28, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18708, "formula_len_all_1": 4688, "len_all": 142417, "len_all_1": 66710, "len_abs": 954, "len_title": 126, "len_sents": 40231, "len_sents_1": 26650, "n_sents": 355, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1184, "title": "Singular Subspace Perturbation Bounds via Rectangular Random Matrix Diffusions", "abs": "Given a matrix $A \\in \\mathbb{R}^{m\\times d}$ with singular values $\\sigma_1\\geq \\cdots \\geq \\sigma_d$, and a random matrix $G \\in \\mathbb{R}^{m\\times d}$ with iid $N(0,T)$ entries for some $T>0$, we derive new bounds on the Frobenius distance between subspaces spanned by the top-$k$ (right) singular vectors of $A$ and $A+G$. This problem arises in numerous applications in statistics where a data matrix may be corrupted by Gaussian noise, and in the analysis of the Gaussian mechanism in differential privacy, where Gaussian noise is added to data to preserve private information. We show that, for matrices $A$ where the gaps in the top-$k$ singular values are roughly $\\Omega(\\sigma_k-\\sigma_{k+1})$ the expected Frobenius distance between the subspaces is $\\tilde{O}(\\frac{\\sqrt{d}}{\\sigma_k-\\sigma_{k+1}} \\times \\sqrt{T})$, improving on previous bounds by a factor of $\\frac{\\sqrt{m}}{\\sqrt{d}}$. To obtain our bounds we view the perturbation to the singular vectors as a diffusion process-- the Dyson-Bessel process-- and use tools from stochastic calculus to track the evolution of the subspace spanned by the top-$k$ singular vectors, which may be of independent interest.", "keywords": [ "Matrix Perturbation Bounds", "Matrix Diffusions", "Low-rank Approximation", "Differential Privacy" ], "rating_list": [ 3, 5, 6, 8, 5 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 1, 3, 3, 3, 2 ], "confidence_list": [ 5, 3, 3, 2, 3 ] }, { "paper_id": "G84F1h2IiD", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Jun Suzuki", "gender": "Male", "institution": "Tohoku University", "country": "JP", "position": "Full Professor" }, { "name": "Sho Takase", "gender": "unknown", "institution": "LINE Corporation", "country": "", "position": "Researcher" }, { "name": "Shun Kiyono", "gender": "Male", "institution": "SB Intuitions", "country": "JP", "position": "Researcher" }, { "name": "Sosuke Kobayashi", "gender": "Male", "institution": "Tohoku University", "country": "", "position": "Researcher" } ], "n_formula": 90, "n_formula_1": 46, "n_ref_uni": 37, "n_ref": 98, "n_ref_all": 163, "n_fig": 12, "n_tab": 8, "L_tab": 1235, "n_element_tab": 182, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 419, "n_element_tab_1": 66, "formula_len_all": 5554, "formula_len_all_1": 1835, "len_all": 185782, "len_all_1": 63441, "len_abs": 812, "len_title": 116, "len_sents": 45251, "len_sents_1": 24409, "n_sents": 396, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 817, "title": "Spike No More: Stabilizing the Pre-training of Large Language Models", "abs": "Loss spikes often occur during pre-training of large language models.\nThe spikes degrade the performance of large language models and sometimes ruin the pre-training.\nSince the pre-training needs a vast computational budget, we should avoid such spikes.\nBased on the assumption that the loss spike is caused by the sudden growth of the gradient norm, we explore factors to keep the gradient norm small through an analysis of the spectral norms of the Jacobian matrices for the sub-layers.\nOur findings suggest that stabilizing the pre-training process requires two conditions: small sub-layers and large shortcut.\nWe conduct various experiments to empirically verify our theoretical analyses.\nExperimental results demonstrate that methods satisfying the conditions effectively prevent loss spikes during pre-training.", "keywords": [ "neural language model", "transformer", "llm" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "G82uQztzxl", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Adam Daniel Jozefiak", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Vivek Farias", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 67, "n_formula_1": 5, "n_ref_uni": 16, "n_ref": 67, "n_ref_all": 111, "n_fig": 6, "n_tab": 21, "L_tab": 10850, "n_element_tab": 1183, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 883, "n_element_tab_1": 69, "formula_len_all": 6021, "formula_len_all_1": 393, "len_all": 198010, "len_all_1": 66693, "len_abs": 1344, "len_title": 107, "len_sents": 63585, "len_sents_1": 32820, "n_sents": 495, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1344, "title": "Self-Normalized Resets for Plasticity in Continual Learning", "abs": "Plasticity Loss is an increasingly important phenomenon that refers to the empirical observation that as a neural network is continually trained on a sequence of changing tasks, its ability to adapt to a new task diminishes over time. We introduce Self-Normalized Resets (SNR), a simple adaptive algorithm that mitigates plasticity loss by resetting a neuron\u2019s weights when evidence suggests its firing rate has effectively dropped to zero. Across a battery of continual learning problems and network architectures, we demonstrate that SNR consistently attains superior performance compared to its competitor algorithms. We also demonstrate that SNR is robust to its sole hyperparameter, its rejection percentile threshold, while competitor algorithms show significant sensitivity. SNR\u2019s threshold-based reset mechanism is motivated by a simple hypothesis test we derive. Seen through the lens of this hypothesis test, competing reset proposals yield suboptimal error rates in correctly detecting inactive neurons, potentially explaining our experimental observations. We also conduct a theoretical investigation of the optimization landscape for the problem of learning a single ReLU. We show that even when initialized adversarially, an idealized version of SNR learns the target ReLU, while regularization based approaches can fail to learn.", "keywords": [ "Continual Learning", "Plasticity", "Lifelong Learning", "Neuron Resets" ], "rating_list": [ 5, 6, 8, 3, 5 ], "soundness_list": [ 2, 2, 4, 2, 3 ], "presentation_list": [ 3, 2, 3, 1, 3 ], "contribution_list": [ 2, 2, 4, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 3 ] }, { "paper_id": "G7u4ue6ncT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Di Liu", "gender": "Male", "institution": "Rutgers University, New Brunswick", "country": "US", "position": "PhD student" }, { "name": "Ligong Han", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Song Wen", "gender": "unknown", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Yunhe Gao", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Zhuowei Li", "gender": "Male", "institution": "Rutgers University, New Brunswick", "country": "US", "position": "PhD student" }, { "name": "Zihao Xu", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Dimitris Metaxas", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Full Professor" }, { "name": "Hao Wang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 46, "n_ref": 87, "n_ref_all": 123, "n_fig": 7, "n_tab": 12, "L_tab": 2199, "n_element_tab": 200, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1993, "n_element_tab_1": 191, "formula_len_all": 652, "formula_len_all_1": 467, "len_all": 192354, "len_all_1": 72381, "len_abs": 1358, "len_title": 76, "len_sents": 49886, "len_sents_1": 33101, "n_sents": 349, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 28, "L_abs": 1356, "title": "Implicit In-context Learning", "abs": "In-context Learning (ICL) empowers large language models (LLMs) to swiftly\nadapt to unseen tasks at inference-time by prefixing a few demonstration\nexamples before queries. Despite its versatility, ICL incurs substantial computa-\ntional and memory overheads compared to zero-shot learning and is sensitive to\nthe selection and order of demonstration examples. In this work, we introduce\nImplicit In-context Learning (I2CL), an innovative paradigm that reduces the\ninference cost of ICL to that of zero-shot learning with minimal information loss.\nI2CL operates by first generating a condensed vector representation, namely a\ncontext vector, extracted from the demonstration examples. It then conducts an\ninference-time intervention through injecting a linear combination of the context\nvector and query activations back into the model\u2019s residual streams. Empirical\nevaluation on nine real-world tasks across three model architectures demonstrates\nthat I2CL achieves few-shot level performance at zero-shot cost, and it exhibits\nrobustness against variations in demonstration examples. Furthermore, I2CL facilitates a novel representation of \u201ctask-ids\u201d, enhancing task similarity detection and\nfostering effective transfer learning. We also performs a comprehensive analysis\nand ablation study on I2CL, offering deeper insights into its internal mechanisms.", "keywords": [ "Large Language Model", "In-context Learning", "Activation Engineering", "Efficiency" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "G7sIFXugTX", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Albert \u00d6rwall", "gender": "Male", "institution": "Moatless AI", "country": "SE", "position": "Researcher" }, { "name": "Anirudh Goyal", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Antonis Antoniades", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "", "position": "Researcher" }, { "name": "Kexun Zhang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Yuxi Xie", "gender": "Female", "institution": "National University of Singapore", "country": "", "position": "PhD student" }, { "name": "William Yang Wang", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 31, "n_ref": 64, "n_ref_all": 83, "n_fig": 11, "n_tab": 6, "L_tab": 4520, "n_element_tab": 201, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 598, "n_element_tab_1": 57, "formula_len_all": 386, "formula_len_all_1": 240, "len_all": 279288, "len_all_1": 69116, "len_abs": 1767, "len_title": 139, "len_sents": 59240, "len_sents_1": 34341, "n_sents": 471, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1775, "title": "Enhancing Software Agents with Monte Carlo Tree Search and Hindsight Feedback", "abs": "Software engineers operating in complex and dynamic environments must continuously adapt to evolving requirements, learn iteratively from experience, and reconsider their approaches based on new insights. However, current large language model (LLM)-based software agents often follow linear, sequential processes that prevent backtracking and exploration of alternative solutions, limiting their ability to adapt their strategies when initial approaches prove ineffective. To address these challenges, we propose SWE-Search, a multi-agent framework that integrates Monte Carlo Tree Search (MCTS) with a self-improvement mechanism to enhance software agents' performance on repository-level software tasks. SWE-Search extends traditional MCTS by incorporating a hybrid value function that leverages LLMs for both numerical value estimation and qualitative evaluation. This enables self-feedback loops where agents iteratively refine their strategies based on both quantitative numerical evaluations and qualitative natural language assessments of pursued trajectories. The framework includes a SWE-Agent for adaptive exploration, a Value Agent for iterative feedback, and a Discriminator Agent that facilitates multi-agent debate for collaborative decision-making. Applied to the SWE-bench benchmark, our approach demonstrates a 23% relative improvement in performance across five models compared to standard open-source agents without MCTS. Our analysis reveals how performance scales with increased search depth and identifies key factors that facilitate effective self-evaluation in software agents. This work highlights the potential of self-evaluation driven search techniques to enhance agent reasoning and planning in complex, dynamic software engineering environments.", "keywords": [ "agents", "LLM", "SWE-agents", "SWE-bench", "search", "planning", "reasoning", "self-improvement", "open-ended" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "G7gvaoX9AW", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Akifumi Wachi", "gender": "Male", "institution": "LY Corporation", "country": "JP", "position": "Chief Research Scientist" }, { "name": "Rei Sato", "gender": "Male", "institution": "LY Corporation", "country": "", "position": "Researcher" }, { "name": "Takumi Tanabe", "gender": "unknown", "institution": "LY Corporation", "country": "", "position": "Researcher" }, { "name": "Thien Q. Tran", "gender": "Male", "institution": "LINE Yahoo Corp", "country": "JP", "position": "Researcher" }, { "name": "Youhei Akimoto", "gender": "unknown", "institution": "RIKEN AIP", "country": "JP", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 13, "n_ref_uni": 39, "n_ref": 115, "n_ref_all": 152, "n_fig": 13, "n_tab": 11, "L_tab": 4565, "n_element_tab": 405, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 330, "n_element_tab_1": 20, "formula_len_all": 2874, "formula_len_all_1": 991, "len_all": 238053, "len_all_1": 68857, "len_abs": 1319, "len_title": 123, "len_sents": 74370, "len_sents_1": 32355, "n_sents": 656, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1325, "title": "Vulnerabilities Mitigation for Safety-Aligned Language Models via Debiasing", "abs": "Safety alignment is a fundamental yet still developing research topic for the real-world applications of AI.\nDespite the multifaceted nature of safety and trustworthiness in AI, current safety alignment methods often focus on a singular notion of safety. By carefully assessing models from the existing safety-alignment methods, we found that, while they generally improved overall safety performance, they failed to ensure safety in specific categories. Our study first identified the difficulty of eliminating such vulnerabilities without sacrificing the model's helpfulness. We found that, while smaller KL penalty parameters, increased training iterations, and dataset cleansing can enhance safety, they do not necessarily improve the trade-off between safety and helpfulness. We discovered that safety alignment can induce undesired effects and result in a model that prefers generating negative tokens leading to rejective responses, regardless of the input context. To address this, we introduced a learning-free method, Token-level Safety-Debiased Inference (TSDI), to estimate and correct this bias during the generation process using randomly constructed prompts. Our experiments demonstrated that our method could enhance the model's helpfulness while maintaining safety, thus improving the trade-off Pareto-front.", "keywords": [ "AI Alignment", "Large Language Models", "AI Safety", "Safe RL" ], "rating_list": [ 6, 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3, 3 ] }, { "paper_id": "G6iREzfcZ7", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Anthony Kobanda", "gender": "Male", "institution": "Inria Scool", "country": "FR", "position": "PhD student" }, { "name": "Ludovic Denoyer", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "R\u00e9my Portelas", "gender": "Male", "institution": "Ubisoft", "country": "FR", "position": "Researcher" }, { "name": "odalric-ambrym maillard", "gender": "unknown", "institution": "inria", "country": "", "position": "Assistant Professor" } ], "n_formula": 43, "n_formula_1": 16, "n_ref_uni": 48, "n_ref": 77, "n_ref_all": 96, "n_fig": 9, "n_tab": 8, "L_tab": 896, "n_element_tab": 177, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2314, "n_element_tab_1": 146, "formula_len_all": 2883, "formula_len_all_1": 1098, "len_all": 187653, "len_all_1": 67053, "len_abs": 1142, "len_title": 127, "len_sents": 60645, "len_sents_1": 29694, "n_sents": 481, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1146, "title": "Hierarchical Subspaces of Policies for Continual Offline Reinforcement Learning", "abs": "In dynamic domains such as autonomous robotics and video game simulations, agents must continuously adapt to new tasks while retaining previously acquired skills. This ongoing process, known as Continual Reinforcement Learning, presents significant challenges, including the risk of forgetting past knowledge and the need for scalable solutions as the number of tasks increases. To address these issues, we introduce HIerarchical LOW-rank Subspaces of Policies (HILOW), a novel framework designed for continual learning in offline navigation settings. HILOW leverages hierarchical policy subspaces to enable flexible and efficient adaptation to new tasks while preserving existing knowledge. We demonstrate, through a careful experimental study, the effectiveness of our method in both classical MuJoCo maze environments and complex video game-like simulations, showcasing competitive performance and satisfying adaptability according to classical continual learning metrics, in particular regarding memory usage. Our work provides a promising framework for real-world applications where continuous learning from pre-collected data is essential.", "keywords": [ "Reinforcement Learning", "Continual Reinforcement Learning", "Offline Learning", "Hierarchical Policies", "Navigation" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "G6dMvRuhFr", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Yilun Du", "gender": "unknown", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Yunhao Luo", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 61, "n_ref": 98, "n_ref_all": 147, "n_fig": 17, "n_tab": 9, "L_tab": 5218, "n_element_tab": 518, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1419, "n_element_tab_1": 180, "formula_len_all": 237, "formula_len_all_1": 237, "len_all": 214292, "len_all_1": 67599, "len_abs": 1496, "len_title": 118, "len_sents": 63414, "len_sents_1": 31887, "n_sents": 462, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 70, "L_abs": 1420, "title": "Grounding Video Models to Actions through Goal Conditioned Exploration", "abs": "Large video models, pretrained on massive quantities of amount of Internet video, provide a rich source of physical knowledge about the dynamics and motions of objects and tasks.\nHowever, video models are not grounded in the embodiment of an agent, and do not describe how to actuate the world to reach the visual states depicted in a video.\nTo tackle this problem, current methods use a separate vision-based inverse dynamic model trained on embodiment-specific data to map image states to actions. \nGathering data to train such a model is often expensive and challenging, and this model is limited to visual settings similar to the ones in which data is available.\nIn this paper, we investigate how to directly ground video models to continuous actions through self-exploration in the embodied environment -- using generated video states as visual goals for exploration.\nWe propose a framework that uses trajectory level action generation in combination with video guidance to\nenable an agent to solve complex tasks without any external supervision, e.g., rewards, action labels, or segmentation masks.\nWe validate the proposed approach on 8 tasks in Libero, 6 tasks in MetaWorld, 4 tasks in Calvin, and 12 tasks in iThor Visual Navigation. \nWe show how our approach is on par with or even surpasses multiple behavior cloning baselines trained on expert demonstrations while without requiring any action annotations.", "keywords": [ "Embodied AI", "Decision Making", "Robotics", "Video Model" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 4 ], "contribution_list": [ 4, 3, 4 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "G6aJyS0ZV0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jingwei Xu", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Visiting Graduate" }, { "name": "Shenghua Gao", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Yanwei Fu", "gender": "Male", "institution": "Fudan University,", "country": "CN", "position": "Professor" }, { "name": "Yikai Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Yiqun Zhao", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 18, "n_ref": 72, "n_ref_all": 105, "n_fig": 22, "n_tab": 7, "L_tab": 1499, "n_element_tab": 171, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 769, "n_element_tab_1": 106, "formula_len_all": 514, "formula_len_all_1": 446, "len_all": 213248, "len_all_1": 61107, "len_abs": 2057, "len_title": 109, "len_sents": 46390, "len_sents_1": 28651, "n_sents": 412, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1754, "title": "3D StreetUnveiler with Semantic-aware 2DGS - a simple baseline", "abs": "Unveiling an empty street from crowded observations captured by in-car cameras is crucial for autonomous driving. However, removing all temporarily static objects, such as stopped vehicles and standing pedestrians, presents a significant challenge. Unlike object-centric 3D inpainting, which relies on thorough observation in a small scene, street scene cases involve long trajectories that differ from previous 3D inpainting tasks. The camera-centric moving environment of captured videos further complicates the task due to the limited degree and time duration of object observation. To address these obstacles, we introduce StreetUnveiler to reconstruct an empty street. StreetUnveiler learns a 3D representation of the empty street from crowded observations. Our representation is based on the hard-label semantic 2D Gaussian Splatting (2DGS) for its scalability and ability to identify Gaussians to be removed. We inpaint rendered image after removing unwanted Gaussians to provide pseudo-labels and subsequently re-optimize the 2DGS. Given its temporal continuous movement, we divide the empty street scene into observed, partial-observed, and unobserved regions, which we propose to locate through a rendered alpha map. This decomposition helps us to minimize the regions that need to be inpainted. To enhance the temporal consistency of the inpainting, we introduce a novel time-reversal framework to inpaint frames in reverse order and use later frames as references for earlier frames to fully utilize the long-trajectory observations. Our experiments conducted on the street scene dataset successfully reconstructed a 3D representation of the empty street. The mesh representation of the empty street can be extracted for further applications.", "keywords": [ "3D inpainting", "Empty Street Reconstruction" ], "rating_list": [ 8, 6, 3, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 4, 3, 1, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "G6S9B7fr74", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 5, "author_info_list": [ { "name": "Harshit Joshi", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "James Chen", "gender": "unknown", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Monica Lam", "gender": "Female", "institution": "Stanford University", "country": "", "position": "Full Professor" }, { "name": "Shicheng Liu", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Robert Larsen Weigle", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 58, "n_ref_all": 96, "n_fig": 5, "n_tab": 10, "L_tab": 3310, "n_element_tab": 190, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2329, "n_element_tab_1": 153, "formula_len_all": 54, "formula_len_all_1": 0, "len_all": 233439, "len_all_1": 69764, "len_abs": 1420, "len_title": 131, "len_sents": 76696, "len_sents_1": 32086, "n_sents": 654, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1427, "title": "Coding Reliable LLM-based Integrated Task and Knowledge Agents with GenieWorksheets", "abs": "Large Language Models (LLMs) present an opportunity to create automated assistants that can help users navigate complex tasks. However, existing approaches have limitations in handling conditional logic, integrating knowledge sources, and consistently following instructions. Researchers and industry professionals often employ ad hoc pipelines to construct conversational agents. These pipelines aim to maintain context, address failure cases, and minimize hallucinations, yet frequently fail to achieve these objectives. To this end, we present Genie \u2013 a programmable framework for creating task-oriented conversational agents that are designed to handle complex user interactions and knowledge queries. Unlike LLMs, Genie provides reliable grounded responses, with controllable agent policies through its expressive specification, Genie Worksheet. In contrast to dialog trees, it is resilient to diverse user queries, helpful with knowledge sources, and offers ease of programming policies through its declarative paradigm. The agents built using Genie outperforms the state-of-the-art method on complex logic domains in STARV2 dataset by up to 20.5%. Through a real-user study involving 62 participants, we show that Genie beats the GPT-4 Turbo with function calling baseline by 21.1%, 20.1%, and 61% on execution accuracy, dialogue act accuracy, and goal completion rate, respectively, on three diverse real-world domains.", "keywords": [ "conversational agents", "framework", "task-oriented dialog agents", "task and knowledge agents" ], "rating_list": [ 5, 8, 3, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 1, 3, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 2, 2, 3 ] }, { "paper_id": "G6DLQ40VVR", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Dong Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Distinguished Scientist" }, { "name": "Hongming Zhang", "gender": "Male", "institution": "Tencent AI Lab Seattle", "country": "US", "position": "Researcher" }, { "name": "Kaixin Ma", "gender": "unknown", "institution": "Tencent AI Lab", "country": "US", "position": "Researcher" }, { "name": "Tianqing Fang", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Senior Research Scientist" }, { "name": "Yangqiu Song", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Ye Tian", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Yue Yang", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Zhaowei Wang", "gender": "Male", "institution": "Edinburgh University, University of Edinburgh", "country": "GB", "position": "Researcher" }, { "name": "Xiaoman Pan", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 70, "n_ref": 124, "n_ref_all": 163, "n_fig": 9, "n_tab": 7, "L_tab": 1668, "n_element_tab": 159, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2286, "n_element_tab_1": 81, "formula_len_all": 187, "formula_len_all_1": 222, "len_all": 209651, "len_all_1": 66762, "len_abs": 1231, "len_title": 130, "len_sents": 48656, "len_sents_1": 29037, "n_sents": 450, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1238, "title": "DivScene: Benchmarking LVLMs for Object Navigation with Diverse Scenes and Objects", "abs": "Object navigation in unknown environments is crucial for deploying embodied agents in real-world applications.\nWhile we have witnessed huge progress due to large-scale scene datasets, faster simulators, and stronger models, previous studies mainly focus on limited scene types and target objects. In this paper, we study a new task of navigating to diverse target objects in a large number of scene types. To benchmark the problem, we present a large-scale scene dataset, DivScene, which contains 4,614 scenes across 81 different types. With the dataset, we build an end-to-end embodied agent, NatVLM, by fine-tuning a Large Vision Language Model (LVLM) through imitation learning. The LVLM is trained to take previous observations from the environment and generate the next actions. We also introduce CoT explanation traces of the action prediction for better performance when tuning LVLMs. Our extensive experiments find that we can build a performant LVLM-based agent through imitation learning on the shortest paths constructed by a BFS planner without any human supervision. Our agent achieves a success rate that surpasses GPT-4o by over 20%. Meanwhile, we carry out various analyses showing the generalization ability of our agent.", "keywords": [ "Embodied AI", "Object Navigation", "Large Vision Language Models", "LVLM", "Imitation Learning" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 4 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "G5sPv4KSjR", "primary_area": "reinforcement learning", "n_author": 9, "author_info_list": [ { "name": "Kazumi Kasaura", "gender": "unknown", "institution": "OMRON SINIC X Corporation", "country": "", "position": "Researcher" }, { "name": "Kenta Hoshino", "gender": "Male", "institution": "Kyoto University", "country": "JP", "position": "Assistant Professor" }, { "name": "Masashi Hamaya", "gender": "Male", "institution": "OMRON SINICX Corp.", "country": "", "position": "Researcher" }, { "name": "Paavo Parmas", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Project Assistant Professor" }, { "name": "Tadashi Kozuno", "gender": "Male", "institution": "OMRON SINIC X", "country": "JP", "position": "Researcher" }, { "name": "Toshinori Kitamura", "gender": "Male", "institution": "The University of Tokyo", "country": "", "position": "PhD student" }, { "name": "Wataru Kumagai", "gender": "Male", "institution": "Omron Sinic X", "country": "", "position": "Researcher" }, { "name": "Yohei Hosoe", "gender": "Male", "institution": "Kyoto University", "country": "JP", "position": "Associate Professor" }, { "name": "Yutaka Matsuo", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Full Professor" } ], "n_formula": 143, "n_formula_1": 27, "n_ref_uni": 63, "n_ref": 140, "n_ref_all": 210, "n_fig": 4, "n_tab": 6, "L_tab": 942, "n_element_tab": 108, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 195, "n_element_tab_1": 29, "formula_len_all": 16129, "formula_len_all_1": 3268, "len_all": 280342, "len_all_1": 76368, "len_abs": 1102, "len_title": 148, "len_sents": 73403, "len_sents_1": 30425, "n_sents": 753, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1156, "title": "Near-Optimal Policy Identification in Robust Constrained Markov Decision Processes via Epigraph Form", "abs": "Designing a safe policy for uncertain environments is crucial in real-world control systems. However, this challenge remains inadequately addressed within the Markov decision process (MDP) framework. This paper presents the first algorithm guaranteed to identify a near-optimal policy in a robust constrained MDP (RCMDP), where an optimal policy minimizes cumulative cost while satisfying constraints in the worst-case scenario across a set of environments. We first prove that the conventional policy gradient approach to the Lagrangian max-min formulation can become trapped in suboptimal solutions. This occurs when its inner minimization encounters a sum of conflicting gradients from the objective and constraint functions. To address this, we leverage the epigraph form of the RCMDP problem, which resolves the conflict by selecting a single gradient from either the objective or the constraints. Building on the epigraph form, we propose a bisection search algorithm with a policy gradient subroutine and prove that it identifies an $\\varepsilon$-optimal policy in an RCMDP with $\\widetilde{\\mathcal{O}}(\\varepsilon^{-4})$ robust policy evaluations.", "keywords": [ "Markov Decision Process", "Constrained Optimization", "Robust Optimization" ], "rating_list": [ 6, 8, 6, 3, 6 ], "soundness_list": [ 4, 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 1, 3 ], "contribution_list": [ 3, 3, 3, 1, 3 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "G5KbDVAlI6", "primary_area": "causal reasoning", "n_author": 7, "author_info_list": [ { "name": "Biwei Huang", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" }, { "name": "Boyang Sun", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Gongxu Luo", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Haoyue Dai", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Loka Li", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "PhD student" }, { "name": "Petar Stojanov", "gender": "unknown", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 41, "n_ref": 54, "n_ref_all": 82, "n_fig": 16, "n_tab": 4, "L_tab": 947, "n_element_tab": 116, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 807, "n_element_tab_1": 91, "formula_len_all": 530, "formula_len_all_1": 306, "len_all": 145088, "len_all_1": 66407, "len_abs": 1383, "len_title": 139, "len_sents": 46365, "len_sents_1": 33123, "n_sents": 321, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1390, "title": "Gene Regulatory Network Inference in the Presence of Selection Bias and Latent Confounders", "abs": "The study of gene regulatory network inference (GRNI), with a focus on uncovering causal relations among genes, holds significant potential to explain fundamental biological processes, such as how cellular identity is established or disrupted in disease. Unfortunately, current methods fail to adequately interpret the widespread phenomena of differential gene expression. The limitation can largely be attributed to the overlook of the selection process (e.g., survival bias), which is ubiquitous and fundamental in biology. Furthermore, recent studies have shown that gene expression is regulated by latent confounders (e.g., non-coding RNAs). Both of which can lead to spurious dependencies, thereby distorting GRNI results. To mitigate these challenges, we propose a novel algorithm, called Gene Regulatory Network Inference in the presence of Selection bias and Latent confounders (GISL). It is designed to uncover the causal structure by leveraging data across multiple distributions obtained via gene perturbation. Surprisingly, we find that the qualitative structure information, selection process, and latent confounders are partially identifiable without any parametric assumption under mild graphical conditions. Experimental results on both synthetic and real-world single-cell gene expression datasets demonstrate the superiority of GISL over existing strong baseline methods.", "keywords": [ "gene regulatory network inference", "selection bias", "latent confounders", "causal discovery" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 1, 1, 2, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "G5DziesYxL", "primary_area": "datasets and benchmarks", "n_author": 44, "author_info_list": [ { "name": "Ahmad Mustafa Anis", "gender": "unknown", "institution": "International Islamic University", "country": "", "position": "Undergrad student" }, { "name": "Alex 'Sandy' Pentland", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "An Dinh", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Ariel N. Lee", "gender": "Female", "institution": "Raive", "country": "US", "position": "Researcher" }, { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Campbell S. Lund", "gender": "Female", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "MS student" }, { "name": "Christopher Klamm", "gender": "unknown", "institution": "Universit\u00e4t Mannheim", "country": "DE", "position": "PhD student" }, { "name": "Da Yin", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Deividas Mataciunas", "gender": "Male", "institution": "International School of Management", "country": "", "position": "Undergrad student" }, { "name": "Diganta Misra", "gender": "Male", "institution": "Max-Planck-Institute for Intelligent Systems, Max-Planck Institute", "country": "DE", "position": "PhD student" }, { "name": "Enrico Shippole", "gender": "Male", "institution": "Teraflop AI", "country": "", "position": "CEO" }, { "name": "Jad Kabbara", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Postdoc" }, { "name": "Jianguo Zhang", "gender": "Male", "institution": "SalesForce AI Research", "country": "US", "position": "Researcher" }, { "name": "Joanna Materzynska", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Kevin Klyman", "gender": "Male", "institution": "Stanford University", "country": "", "position": "MS student" }, { "name": "Kun Qian", "gender": "Male", "institution": "Columbia University", "country": "", "position": "PhD student" }, { "name": "Kushagra Tiwary", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Lester James Validad Miranda", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Luis Villa", "gender": "Male", "institution": "Luis Villa Law", "country": "", "position": "Attorney" }, { "name": "Manan Dey", "gender": "Male", "institution": "SalesForce.com", "country": "", "position": "Senior Member of Technical Staff " }, { "name": "Manuel Cherep", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Minnie Liang", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Mohammed Hamdy", "gender": "Male", "institution": "Independent", "country": "", "position": "Researcher" }, { "name": "Naana Obeng-Marnu", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Nayan Saxena", "gender": "Male", "institution": "MLC", "country": "CA", "position": "Researcher" }, { "name": "Nikhil Singh", "gender": "Male", "institution": "Dartmouth College", "country": "US", "position": "Assistant Professor" }, { "name": "Niklas Muennighoff", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Robert Mahari", "gender": "Male", "institution": "Harvard University", "country": "", "position": "JD" }, { "name": "Seonghyeon Ye", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Seungone Kim", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Shayne Longpre", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Shrestha Mohanty", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Researcher" }, { "name": "Stella Biderman", "gender": "Female", "institution": "EleutherAI", "country": "", "position": "Researcher" }, { "name": "Tobin South", "gender": "Not Specified", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Vipul Gupta", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Vivek Sharma", "gender": "Male", "institution": "Sony Research", "country": "CH", "position": "Scientist" }, { "name": "Vu Minh Chien", "gender": "Male", "institution": "Tohoku University", "country": "JP", "position": "PhD student" }, { "name": "William Brannon", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Xuhui Zhou", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Yizhi LI", "gender": "Male", "institution": "University of Manchester ", "country": "GB", "position": "PhD student" }, { "name": "Damien Sileo", "gender": "Not Specified", "institution": "INRIA", "country": "FR", "position": "Researcher" }, { "name": "Emad Ahmad Alghamdi", "gender": "Male", "institution": "King Abdulaziz University", "country": "SA", "position": "Assistant Professor" }, { "name": "Manan Dey", "gender": "Male", "institution": "KIIT University", "country": "", "position": "Undergrad student" }, { "name": "Sara Hooker", "gender": "unknown", "institution": "Cohere For AI", "country": "US", "position": "Principal Researcher" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 104, "n_ref": 173, "n_ref_all": 218, "n_fig": 7, "n_tab": 3, "L_tab": 676, "n_element_tab": 163, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 920, "n_element_tab_1": 408, "formula_len_all": 0, "formula_len_all_1": 7, "len_all": 266968, "len_all_1": 69508, "len_abs": 1777, "len_title": 111, "len_sents": 48603, "len_sents_1": 34119, "n_sents": 306, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1787, "title": "Bridging the Data Provenance Gap Across Text, Speech, and Video", "abs": "Progress in AI is driven largely by the scale and quality of training data. Despite this, there is a deficit of empirical analysis examining the attributes of well-established datasets beyond text. In this work we conduct the largest and first-of-its-kind longitudinal audit across modalities---popular text, speech, and video datasets---from their detailed sourcing trends and use restrictions to their geographical and linguistic representation. Our manual analysis covers nearly 4000 public datasets between 1990-2024, spanning 608 languages, 798 sources, 659 organizations, and 67 countries. We find that multimodal machine learning applications have overwhelmingly turned to web-crawled and social media platforms, such as YouTube, for their training sets, eclipsing all other sources since 2019. Secondly, tracing the chain of dataset derivations we find that while less than 33% of datasets are restrictively licensed, over 99%, 78%, and 99% of the source content in widely-used text, speech, and video datasets, respectively, carry non-commercial restrictions. Finally, counter to increasing absolute multilingual and geographic inclusion in publicly available AI training data, our audit demonstrates measures of relative geographical and multilingual representation have failed to significantly improve their coverage since 2013. We believe the breadth of our audit enables us to empirically examine trends in data sourcing, restrictions, and Western-centricity at an ecosystem-level, and that visibility into these questions are essential to progress in responsible AI. As a contribution to ongoing improvements in dataset transparency and responsible use, we release our entire multimodal audit, allowing practitioners to trace data provenance across text, speech, and video.", "keywords": [ "training data", "audit", "speech", "video", "text" ], "rating_list": [ 8, 8, 5, 5 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 3, 3, 4, 2 ] }, { "paper_id": "G4wARwjF8M", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Jiaojiao Han", "gender": "unknown", "institution": "Southwest University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Qitian Wu", "gender": "Male", "institution": "Broad Institute of MIT and Harvard", "country": "US", "position": "Researcher" }, { "name": "Wujiang Xu", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Xuying Ning", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Yongfeng Zhang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Associate Professor" }, { "name": "Yunxiao Shi", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "linwenfang", "gender": "Female", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Liang Zujie", "gender": "Male", "institution": "Ant Group", "country": "", "position": "Researcher" } ], "n_formula": 28, "n_formula_1": 13, "n_ref_uni": 61, "n_ref": 118, "n_ref_all": 145, "n_fig": 4, "n_tab": 17, "L_tab": 3356, "n_element_tab": 460, "n_fig_1": 4, "n_tab_1": 11, "L_tab_1": 1976, "n_element_tab_1": 327, "formula_len_all": 1373, "formula_len_all_1": 584, "len_all": 188702, "len_all_1": 71416, "len_abs": 1958, "len_title": 118, "len_sents": 46497, "len_sents_1": 29325, "n_sents": 362, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1912, "title": "SLMRec: Empowering Small Language Models for Sequential Recommendation", "abs": "Sequential Recommendation (SR) task involves predicting the next item a user is likely to interact with, given their past interactions. \nThe SR models examine the sequence of a user's actions to discern more complex behavioral patterns and temporal dynamics. \nRecent research demonstrates the great impact of LLMs on sequential recommendation systems, either viewing sequential recommendation as language modeling or serving as the backbone for user representation. Although these methods deliver outstanding performance, there is scant evidence of the necessity of a large language model and how large the language model is needed, especially in the sequential recommendation scene. Meanwhile, due to the huge size of LLMs, it is inefficient and impractical to apply a LLM-based model in real-world platforms that often need to process billions of traffic logs daily. In this paper, we explore the influence of LLMs' depth by conducting extensive experiments on large-scale industry datasets. Surprisingly, our motivational experiments reveal that most intermediate layers of LLMs are redundant, indicating that pruning the remaining layers can still maintain strong performance.\nMotivated by this insight, we empower small language models for SR, namely SLMRec, which adopt a simple yet effective knowledge distillation method. Moreover, SLMRec is orthogonal to other post-training efficiency techniques, such as quantization and pruning, so that they can be leveraged in combination. Comprehensive experimental results illustrate that the proposed SLMRec model attains the best performance using only 13\\% of the parameters found in LLM-based recommendation models while simultaneously achieving up to 6.6x and 8.0x speedups in training and inference time costs, respectively. Besides, we provide a theoretical justification for why small language models can perform comparably to large language models in SR.", "keywords": [ "Large Language Models; Knowledge Distillation" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "G4P1q2G0XK", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Demetrio Labate", "gender": "Male", "institution": "University of Houston", "country": "US", "position": "Full Professor" }, { "name": "Dhananjaya Jayasundara", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Heng Zhao", "gender": "Male", "institution": "The Rockefeller University", "country": "US", "position": "Machine Learning Enginer" }, { "name": "Vishal M Patel", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 29, "n_ref": 56, "n_ref_all": 74, "n_fig": 9, "n_tab": 1, "L_tab": 423, "n_element_tab": 38, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 174, "formula_len_all_1": 174, "len_all": 94173, "len_all_1": 56021, "len_abs": 1520, "len_title": 122, "len_sents": 34153, "len_sents_1": 28543, "n_sents": 209, "n_sents_1": 177, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1525, "title": "AINR: Adaptive Learning of Activations for Implicit Neural Representations", "abs": "Implicit Neural Representations (INRs) provide a continuous function learning framework for discrete signal representations. Using positional embeddings and / or specialized activation functions, INRs have overcome many limitations of traditional discrete representations. However, existing work primarily focuses on the use of a single activation function throughout the network, which often requires an exhaustive search for optimal activation parameters tailored to each signal and INR application. We hypothesize that this approach may restrict the representation power and generalization capabilities of INRs; limiting their broader applicability. In this paper, we introduce AINR, a method that adaptively learns the most suitable activation functions for INRs from a predefined dictionary. This dictionary includes activation functions such as Raised Cosines (RC), Root Raised Cosines (RRC), Prolate Spheroidal Wave Function (PSWF), Sinc, Gabor Wavelet, Gaussian, and Sinusoidal. Our method identifies the activation atom that is mostly matched for each layer of the INR based on the given signal. Experimental results demonstrate that AINR not only significantly improves INR performance across various tasks, such as image representation, image inpainting, 3D shape representation, novel view synthesis, super resolution, and reliable edge detection, but also eliminates the need for the previously required exhaustive search for activation parameters, which had to be conducted even before INR training could begin.", "keywords": [ "Implicit Neural Representations", "Adaptive Activation Function Learning", "MLPs" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 4, 5, 5, 5 ] }, { "paper_id": "G3vceNrP4o", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Pengwen Dai", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Assistant Professor" }, { "name": "Siying Zhu", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Wenqi Ren", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Full Professor" }, { "name": "Xin Su", "gender": "Female", "institution": "Fuzhou University", "country": "CN", "position": "MS student" }, { "name": "Zhuoran Zheng", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Postdoc" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 40, "n_ref": 59, "n_ref_all": 75, "n_fig": 8, "n_tab": 6, "L_tab": 1782, "n_element_tab": 266, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 2060, "n_element_tab_1": 297, "formula_len_all": 295, "formula_len_all_1": 295, "len_all": 105149, "len_all_1": 48419, "len_abs": 1470, "len_title": 104, "len_sents": 22283, "len_sents_1": 18762, "n_sents": 182, "n_sents_1": 157, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1451, "title": "Bridge the Gap between SNN and ANN for Image Restoration", "abs": "Models of dense prediction based on traditional Artificial Neural Networks (ANNs) require a lot of energy, especially for image restoration tasks. Currently, neural networks based on the SNN framework are beginning to make their mark in the field of image restoration, especially as they typically use less than 10\\% of the energy of ANNs with the same architecture. However, training an SNN is much more expensive than training an ANN, due to the use of the heuristic gradient descent strategy. In other words, the process of SNN's potential membrane signal changing from sparse to dense is very slow, which affects the convergence of the whole model. To tackle this problem, we propose a novel distillation technique, called asymmetric framework (ANN-SNN) distillation, in which the teacher is an ANN and the student is an SNN. Specifically, we leverage the intermediate features (feature map) learned by the ANN as hints to guide the training process of the SNN. This approach not only accelerates the convergence of the SNN but also improves its final performance, effectively bridging the gap between the efficiency of the SNN and the superior learning capabilities of ANN. Extensive experimental results show that our designed SNN-based image restoration model, which has only 1/300 the number of parameters of the teacher network and 1/50 the energy consumption of the teacher network, is as good as the teacher network in some denoising tasks.", "keywords": [ "Computer Vision" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 1, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "G3aXjVAJjU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Hal Daum\u00e9 III", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Paola Cascante-Bonilla", "gender": "unknown", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Assistant Professor" }, { "name": "Rachel Rudinger", "gender": "Female", "institution": "University of Maryland, College Park", "country": "", "position": "Assistant Professor" }, { "name": "Yu Hou", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Yang Cao", "gender": "Female", "institution": "University of Texas at Austin", "country": "US", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 47, "n_ref": 90, "n_ref_all": 112, "n_fig": 7, "n_tab": 13, "L_tab": 7772, "n_element_tab": 1049, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2943, "n_element_tab_1": 493, "formula_len_all": 312, "formula_len_all_1": 312, "len_all": 183890, "len_all_1": 67233, "len_abs": 1420, "len_title": 126, "len_sents": 41207, "len_sents_1": 27752, "n_sents": 301, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1430, "title": "Natural Language Inference Improves Compositionality in Vision-Language Models", "abs": "Compositional reasoning in Vision-Language Models (VLMs) remains challenging as these models often struggle to relate objects, attributes, and spatial relationships. Recent methods aim to address these limitations by relying on the semantics of the textual description, using Large Language Models (LLMs) to break them down into subsets of questions and answers. However, these methods primarily operate on the surface level, failing to incorporate deeper lexical understanding while introducing incorrect assumptions generated by the LLM. In response to these issues, we present Caption Expansion with Contradictions and Entailments (CECE), a principled approach that leverages Natural Language Inference (NLI) to generate entailments and contradictions from a given premise. CECE produces lexically diverse sentences while maintaining their core meaning. Through extensive experiments, we show that CECE enhances interpretability and reduces overreliance on biased or superficial features. By balancing CECE along the original premise, we achieve significant improvements over previous methods without requiring additional fine-tuning, producing state-of-the-art results on benchmarks that score agreement with human judgments for image-text alignment, and achieving an increase in performance on Winoground of $+19.2\\%$ (group score) and $+12.9\\%$ on EqBen (group score) over the best prior work (finetuned with targeted data).", "keywords": [ "text to image evaluation", "image to text generation", "natural language inference", "sentence decomposition", "large language models", "visual question answering", "question generation", "benchmark" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "G3CpBCQwNh", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Bofang Jiang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Mingquan Feng", "gender": "unknown", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Yixin Huang", "gender": "Female", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Undergrad student" }, { "name": "Yizhou Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Undergrad student" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 3, "n_ref_uni": 41, "n_ref": 62, "n_ref_all": 82, "n_fig": 4, "n_tab": 12, "L_tab": 6226, "n_element_tab": 694, "n_fig_1": 3, "n_tab_1": 9, "L_tab_1": 1517, "n_element_tab_1": 239, "formula_len_all": 1010, "formula_len_all_1": 168, "len_all": 166821, "len_all_1": 69394, "len_abs": 1036, "len_title": 127, "len_sents": 45282, "len_sents_1": 30605, "n_sents": 367, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1039, "title": "PhysPDE: Rethinking PDE Discovery and a Physical HYpothesis Selection Benchmark", "abs": "Despite extensive research, recovering PDE expressions from experimental observations often involves symbolic regression. This method generally lacks the incorporation of meaningful physical insights, resulting in outcomes lacking clear physical interpretations. Recognizing that the primary interest of Machine Learning for Science (ML4Sci) often lies in understanding the underlying physical mechanisms or even discovering new physical laws rather than simply obtaining mathematical expressions, this paper introduces a novel ML4Sci task paradigm. This paradigm focuses on interpreting experimental data within the framework of prior physical hypotheses and theories, thereby guiding and constraining the discovery of PDE expressions. We have formulated this approach as a nonlinear mixed-integer programming (MIP) problem, addressed through an efficient search scheme developed for this purpose. Our experiments on newly designed Fluid Mechanics and Laser Fusion datasets demonstrate the interpretability and feasibility of this method.", "keywords": [ "AI4Science", "Physics", "PDEs", "PDE Discovery" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "G3B5ReApDw", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Harel Mendelman", "gender": "unknown", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "MS student" }, { "name": "Ronen Talmon", "gender": "unknown", "institution": "Technion - Israel Institute of Technology", "country": "IL", "position": "Full Professor" } ], "n_formula": 93, "n_formula_1": 28, "n_ref_uni": 42, "n_ref": 68, "n_ref_all": 100, "n_fig": 15, "n_tab": 10, "L_tab": 4469, "n_element_tab": 324, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 935, "n_element_tab_1": 88, "formula_len_all": 5015, "formula_len_all_1": 2007, "len_all": 208525, "len_all_1": 67808, "len_abs": 1244, "len_title": 121, "len_sents": 70440, "len_sents_1": 29497, "n_sents": 582, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1250, "title": "Supervised and Semi-Supervised Diffusion Maps with Label-Driven Diffusion", "abs": "In this paper, we introduce Supervised Diffusion Maps (SDM) and Semi-Supervised Diffusion Maps (SSDM), which transform the well-known unsupervised dimensionality reduction algorithm, Diffusion Maps, into supervised and semi-supervised learning tools. The proposed methods, SDM and SSDM, are based on our new approach that treats the labels as a second view of the data. This unique framework allows us to incorporate ideas from multi-view learning. Specifically, we propose constructing two affinity kernels corresponding to the data and the labels. We then propose a multiplicative interpolation scheme of the two kernels, whose purpose is twofold. First, our scheme extracts the common structure underlying the data and the labels by defining a diffusion process driven by the data and the labels. This label-driven diffusion produces an embedding that emphasizes the properties relevant to the label-related task. Second, the proposed interpolation scheme balances the influence of the two kernels. We show on multiple benchmark datasets that the embedding learned by SDM and SSDM is more effective in downstream regression and classification tasks than existing unsupervised, supervised, and semi-supervised nonlinear dimension reduction methods.", "keywords": [ "Manifold Learning", "Diffusion Maps", "Supervised dimension reduction", "Multi-view learning" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "G328D1xt4W", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Amy Wang", "gender": "unknown", "institution": "Genentech", "country": "US", "position": "Researcher" }, { "name": "Avantika Lal", "gender": "Female", "institution": "Genentech", "country": "", "position": "Researcher" }, { "name": "Aviv Regev", "gender": "Female", "institution": "Genentech / Roche", "country": "US", "position": "Researcher" }, { "name": "Chenyu Wang", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Masatoshi Uehara", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Tommaso Biancalani", "gender": "Male", "institution": "Genentech", "country": "", "position": "Director" }, { "name": "Tommi Jaakkola", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Yichun He", "gender": "Female", "institution": "Harvard University, Harvard University", "country": "", "position": "PhD student" }, { "name": "Hanchen Wang", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Joint PostDoc" } ], "n_formula": 56, "n_formula_1": 21, "n_ref_uni": 47, "n_ref": 144, "n_ref_all": 177, "n_fig": 6, "n_tab": 14, "L_tab": 2581, "n_element_tab": 266, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 505, "n_element_tab_1": 48, "formula_len_all": 5866, "formula_len_all_1": 1672, "len_all": 204577, "len_all_1": 66661, "len_abs": 2121, "len_title": 154, "len_sents": 61401, "len_sents_1": 30602, "n_sents": 514, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 2040, "title": "Fine-Tuning Discrete Diffusion Models via Reward Optimization with Applications to DNA and Protein Design", "abs": "Recent studies have demonstrated the strong empirical performance of diffusion models on discrete sequences (i.e., discrete diffusion models) across domains such as natural language and biological sequence generation. For example, in the protein inverse folding task, where the goal is to generate a protein sequence from a given backbone structure, conditional diffusion models have achieved impressive results in generating \"natural\" sequences that fold back into the original structure. However, practical design tasks often require not only modeling a conditional distribution but also optimizing specific task objectives. For instance, in the inverse folding task, we may prefer proteins with high stability. To address this, we consider the scenario where we have pre-trained discrete diffusion models that can generate \"natural\" sequences, as well as reward models that map sequences to task objectives. We then formulate the reward maximization problem within discrete diffusion models, analogous to reinforcement learning (RL), while minimizing the KL divergence against pre-trained diffusion models to preserve naturalness. To solve this RL problem, we propose a novel algorithm that enables direct backpropagation of rewards through entire trajectories generated by diffusion models, by making the originally non-differentiable trajectories differentiable using the Gumbel-Softmax trick. Our theoretical analysis indicates that our approach can generate sequences that are both \"natural\" (i.e., have a high probability under a pre-trained model) and yield high rewards. While similar tasks have been recently explored in diffusion models for continuous domains, our work addresses unique algorithmic and theoretical challenges specific to discrete diffusion models, which arise from their foundation in continuous-time Markov chains rather than Brownian motion. Finally, we demonstrate the effectiveness of our algorithm in generating DNA and protein sequences that optimize enhancer activity and protein stability, respectively.", "keywords": [ "Discrete Diffusion Models", "Reward Optimization", "Fine-Tuning", "AI for science", "Reinforcement learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 5, 2 ] }, { "paper_id": "G2p8TLuJgy", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Fei Liu", "gender": "Female", "institution": "Emory University", "country": "US", "position": "Associate Professor" }, { "name": "Jiaqi Xue", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Mengxin Zheng", "gender": "Female", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Qian Lou", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Xun Chen", "gender": "unknown", "institution": "Samsung Research America", "country": "US", "position": "Director" }, { "name": "Yebowen Hu", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 50, "n_ref": 82, "n_ref_all": 112, "n_fig": 8, "n_tab": 12, "L_tab": 2721, "n_element_tab": 299, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1210, "n_element_tab_1": 124, "formula_len_all": 239, "formula_len_all_1": 239, "len_all": 178973, "len_all_1": 70167, "len_abs": 2007, "len_title": 144, "len_sents": 55249, "len_sents_1": 33607, "n_sents": 436, "n_sents_1": 247, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 2055, "title": "BadRAG: Identifying Vulnerabilities in Retrieval Augmented Generation of Large Language Models", "abs": "Large Language Models (LLMs) are constrained by outdated information and a tendency to generate incorrect data, commonly referred to as \"hallucinations.'' Retrieval-Augmented Generation (RAG) addresses these limitations by combining the strengths of retrieval-based methods and generative models. This approach involves retrieving relevant information from a large, up-to-date database and using it to enhance the generation process, leading to more accurate and contextually appropriate responses. Despite its benefits, RAG introduces a new attack surface for LLMs, particularly because RAG databases are often sourced from public data, such as the web. In this paper, we propose BadRAG to identify the vulnerabilities and attacks on retrieval parts (RAG databases) and their indirect attacks on generative parts (LLMs). Specifically, we identify that poisoning several customized content passages could achieve a retrieval backdoor, where the retrieval works well for clean queries but always returns customized adversarial passages for triggered queries. Triggers and adversarial passages can be highly customized to implement various attacks. For example, a trigger could be a semantic group like *The Republican Party*, *Donald Trump*, etc. Adversarial passages can be tailored to different contents, not only linked to the triggers but also used to indirectly attack generative LLMs without modifying them. These attacks can include denial-of-service attacks on RAG and semantic steering attacks on LLM generations conditioned by the triggers. Our experiments demonstrate that by just poisoning 10 adversarial passages $\\textemdash\\$ merely 0.04\\% of the total corpus $\\textemdash\\$ can induce 98.2\\% success rate to retrieve the adversarial passages. Then, these passages can increase the reject ratio of RAG-based GPT-4 from 0.01\\% to 74.6\\% or increase the rate of negative responses from 0.22\\% to 72\\% for targeted queries. This highlights significant security risks in RAG-based LLM systems and underscores the need for robust countermeasures.", "keywords": [ "Large Language Model", "Backdoor Attack", "GPT-4", "Retrieval-Augmented Generation" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 1, 2, 1 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "G2BiEoB77Z", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Xin Geng", "gender": "Male", "institution": "Southeast University, China", "country": "CN", "position": "Professor" }, { "name": "Yucheng Xie", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "PhD student" }, { "name": "Yuheng Jia", "gender": "unknown", "institution": "Southeast University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhiqiang Kou", "gender": "Male", "institution": "Southeast University", "country": "", "position": "PhD student" }, { "name": "Jing wang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Postdoc" } ], "n_formula": 28, "n_formula_1": 13, "n_ref_uni": 41, "n_ref": 82, "n_ref_all": 91, "n_fig": 3, "n_tab": 5, "L_tab": 5514, "n_element_tab": 318, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2014, "n_element_tab_1": 108, "formula_len_all": 2467, "formula_len_all_1": 900, "len_all": 144772, "len_all_1": 59958, "len_abs": 1009, "len_title": 159, "len_sents": 40653, "len_sents_1": 25905, "n_sents": 314, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 110, "L_abs": 995, "title": "RankMatch: A Novel Approach to Semi-Supervised Label Distribution Learning Leveraging Inter-label Correlations", "abs": "This paper introduces RankMatch, an innovative approach for Semi-Supervised Label Distribution Learning (SSLDL). Addressing the challenge of limited labeled data, RankMatch effectively utilizes a small number of labeled examples in conjunction with a larger quantity of unlabeled data, reducing the need for extensive manual labeling in Deep Neural Network (DNN) applications. Specifically, RankMatch introduces an ensemble learning-inspired averaging strategy that creates a pseudo-label distribution from multiple weakly augmented images. This not only stabilizes predictions but also enhances the model's robustness. Beyond this, RankMatch integrates a pairwise relevance ranking (PRR) loss, capturing the complex inter-label correlations and ensuring that the predicted label distributions align with the ground truth. We establish a theoretical generalization bound for RankMatch, and through extensive experiments, demonstrate its superiority in performance against existing SSLDL methods.", "keywords": [ "multi-label learning; semi-supervised learning; label distributiob learning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "G2AMCTTpCc", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 6, "author_info_list": [ { "name": "Chengzhi Mao", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Guang-He Lee", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Hao He", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yi Wang", "gender": "Female", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Zihao Xu", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Hao Wang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 37, "n_ref": 92, "n_ref_all": 112, "n_fig": 4, "n_tab": 2, "L_tab": 43, "n_element_tab": 1, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 987, "n_element_tab_1": 108, "formula_len_all": 821, "formula_len_all_1": 852, "len_all": 116381, "len_all_1": 62888, "len_abs": 869, "len_title": 101, "len_sents": 34514, "len_sents_1": 28253, "n_sents": 277, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 873, "title": "Taming Continuous Spurious Shift in Domain Adaptation", "abs": "Recent advances in domain adaptation have shown promise in transferring knowledge across domains characterized by a continuous value or vector, such as varying patient ages, where \"age'' serves as a continuous index. However, these approaches often fail when spurious features shift continuously along with the domain index. This paper introduces the first method designed to withstand the continuous shifting of spurious features during domain adaptation. Our method enhances domain adaptation performance by aligning causally transportable encodings across continuously indexed domains. Theoretical analysis demonstrates that our approach more effectively ensures causal transportability across different domains. Empirical results, from both semi-synthetic and real-world medical datasets, indicate that our method outperforms state-of-the-art domain adaptation methods.", "keywords": [ "Domain Adaptation", "Causal Inference" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "G1r2rBkUdu", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "Guangyi Chen", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "US", "position": "Postdoc" }, { "name": "Ignavier Ng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Kun Zhang", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Professor" }, { "name": "Ruichu Cai", "gender": "Male", "institution": "Guangdong University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Shaoan Xie", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Shunxing Fan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "Xinshuai Dong", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Yujia Zheng", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zijian Li", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Postdoc" } ], "n_formula": 36, "n_formula_1": 17, "n_ref_uni": 61, "n_ref": 127, "n_ref_all": 166, "n_fig": 7, "n_tab": 13, "L_tab": 4003, "n_element_tab": 191, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 2034, "n_element_tab_1": 72, "formula_len_all": 4842, "formula_len_all_1": 1200, "len_all": 203236, "len_all_1": 70555, "len_abs": 1452, "len_title": 151, "len_sents": 55581, "len_sents_1": 31309, "n_sents": 428, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 103, "L_abs": 1404, "title": "Synergy Between Sufficient Changes and Sparse Mixing Procedure for Disentangled Representation Learning", "abs": "Disentangled representation learning aims to uncover the latent variables underlying observed data, yet identifying these variables under mild assumptions remains challenging. Some methods rely on sufficient changes in the distribution of latent variables indicated by auxiliary variables, such as domain indices, but acquiring enough domains is often impractical. Alternative approaches exploit the structural sparsity assumption on mixing processes, but this constraint may not hold in practice. Interestingly, we find that these two seemingly unrelated assumptions can actually complement each other. Specifically, when conditioned on auxiliary variables, the sparse mixing process induces independence between latent and observed variables, which simplifies the mapping from estimated to true latent variables and hence compensates for deficiencies of auxiliary variables. Building on this insight, we propose an identifiability theory with less restrictive constraints regarding the auxiliary variables and the sparse mixing process, enhancing applicability to real-world scenarios. Additionally, we develop a generative model framework incorporating a domain encoding network and a sparse mixing constraint and provide two implementations based on variational autoencoders and generative adversarial networks. Experiment results on synthetic and real-world datasets support our theoretical results.", "keywords": [ "Disentangled Representation Learning", "Identifiability" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "G1n50BMqzm", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Hao-Wen Dong", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" }, { "name": "Haven Kim", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" }, { "name": "Julian McAuley", "gender": "Male", "institution": "University of California, San Diego, University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Paul Pu Liang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Taylor Berg-Kirkpatrick", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "Associate Professor" }, { "name": "Weihan Xu", "gender": "Female", "institution": "Duke University", "country": "", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 24, "n_ref": 49, "n_ref_all": 83, "n_fig": 2, "n_tab": 18, "L_tab": 4741, "n_element_tab": 794, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1861, "n_element_tab_1": 285, "formula_len_all": 1066, "formula_len_all_1": 1125, "len_all": 180333, "len_all_1": 63636, "len_abs": 1446, "len_title": 100, "len_sents": 61117, "len_sents_1": 27870, "n_sents": 445, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 52, "L_abs": 1470, "title": "TeaserGen: Generating Teasers for Long Documentaries", "abs": "Teasers are an effective tool for promoting content in entertainment, commercial and educational fields. However, creating an effective teaser for long videos is challenging for it requires long-range multimodal modeling capability for the input videos, while necessitating maintaining audiovisual alignments, managing scene transitions and preserving factual accuracy for the output teasers. Due to the lack of a publicly-available dataset, progress along this research direction has been hindered. In this work, we present DocumentaryNet, a collection of 1,269 documentaries paired with their teasers, featuring multimodal data streams of video, speech, music, sound effects and narrations. With DocumentaryNet, we propose a new two-stage system for generating teasers from long documentaries. The proposed TeaserGen system first generates the teaser narration from the transcribed narration from the documentary using a pretrained large language model, and then selects the most relevant visual content to accompany the generated narration through language-vision models. For narration-video matching, we explore two approaches: a pretraining-based model using pretrained contrastive language-vision models and a deep sequential model that learns the mapping between the narrations and visuals. Our experimental results show that the pretraining-based approach is more effective at identifying relevant visual content than directly trained deep autoregressive models.", "keywords": [ "Teaser Generation", "Multimodal Learning", "Vision-Language Model" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "G1fzW97QKR", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Da Ma", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Hongshen Xu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Kai Yu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Lu Chen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Associate Professor" }, { "name": "Shuai Fan", "gender": "Male", "institution": "AISpeech Ltd", "country": "CN", "position": "Researcher" }, { "name": "Situo Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Su Zhu", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Yuxun Miao", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhi Chen", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "hanqi li", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 31, "n_ref": 70, "n_ref_all": 91, "n_fig": 9, "n_tab": 4, "L_tab": 2131, "n_element_tab": 202, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 2033, "n_element_tab_1": 199, "formula_len_all": 539, "formula_len_all_1": 539, "len_all": 149744, "len_all_1": 57003, "len_abs": 980, "len_title": 133, "len_sents": 29915, "len_sents_1": 23716, "n_sents": 215, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 990, "title": "Boosting Long-Context LLM Inference Efficiency with Intra-Layer Attention Similarity", "abs": "The increasing context window size in Large Language Models (LLMs), such as the GPT and LLaMA series, has improved their ability to tackle complex, long-text tasks, but at the cost of inference efficiency, particularly regarding memory and computational complexity. Existing methods, including selective token retention and window-based attention, improve efficiency but risk discarding important tokens needed for future text generation. In this paper, we propose an approach that enhances LLM efficiency without token loss by reducing the memory and computational load of less important tokens, rather than discarding them. \n We address two challenges: 1) investigating the distribution of important tokens in the context, discovering recent tokens are more important than distant tokens in context, and 2) optimizing resources for distant tokens by sharing attention scores across layers. The experiments show that our method saves $35$% KV cache without compromising the performance.", "keywords": [ "Long context LLM", "attention similarity", "window attention" ], "rating_list": [ 3, 3, 8, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "G1CN7R5qwE", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Ao Li", "gender": "Male", "institution": "Beijing Normal University", "country": "CN", "position": "Undergrad student" }, { "name": "Haolin Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jie Zhou", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jingxuan Niu", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Jiwen Lu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Lei Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Researcher" }, { "name": "Wenliang Zhao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yansong Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhu Yixuan", "gender": "Male", "institution": "Automation, Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 54, "n_ref": 145, "n_ref_all": 186, "n_fig": 17, "n_tab": 4, "L_tab": 2206, "n_element_tab": 256, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2267, "n_element_tab_1": 216, "formula_len_all": 497, "formula_len_all_1": 488, "len_all": 188284, "len_all_1": 66033, "len_abs": 1366, "len_title": 114, "len_sents": 49582, "len_sents_1": 30289, "n_sents": 404, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1383, "title": "InstaRevive: One-Step Image Enhancement via Dynamic Score Matching", "abs": "Image enhancement finds wide-ranging applications in real-world scenarios due to complex environments and the inherent limitations of imaging devices. Recent diffusion-based methods yield promising outcomes but necessitate prolonged and computationally intensive iterative sampling. In response, we propose InstaRevive, a straightforward yet powerful image enhancement framework that employs score-based diffusion distillation to harness potent generative capability and minimize the sampling steps. To fully exploit the potential of the pre-trained diffusion model, we devise a practical and effective diffusion distillation pipeline using dynamic noise control to address inaccuracies in updating direction during score matching. Our noise control strategy enables a dynamic diffusing scope, facilitating precise learning of denoising trajectories within the diffusion model and ensuring accurate distribution matching gradients during training. Additionally, to enrich guidance for the generative power, we incorporate textual prompts via image captioning as auxiliary conditions, fostering further exploration of the diffusion model. Extensive experiments substantiate the efficacy of our framework across a diverse array of challenging tasks and datasets, unveiling the compelling efficacy and efficiency of InstaRevive in delivering high-quality and visually appealing results.", "keywords": [ "Image Enhancement", "Diffusion model", "Score Distillation" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 4, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "G19piTjVYA", "primary_area": "causal reasoning", "n_author": 3, "author_info_list": [ { "name": "Arash Mehrjou", "gender": "Male", "institution": "GlaxoSmithKlein", "country": "", "position": "Researcher" }, { "name": "Mathieu Chevalley", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Patrick Schwab", "gender": "unknown", "institution": "GlaxoSmithKline plc", "country": "", "position": "Director" } ], "n_formula": 21, "n_formula_1": 19, "n_ref_uni": 36, "n_ref": 68, "n_ref_all": 83, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1395, "formula_len_all_1": 1247, "len_all": 106751, "len_all_1": 58285, "len_abs": 1221, "len_title": 50, "len_sents": 32919, "len_sents_1": 27671, "n_sents": 253, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1216, "title": "Efficient Differentiable Discovery of Causal Order", "abs": "In the algorithm Intersort, Chevalley et al. proposed a score-based method to discover the causal order of variables in a Directed Acyclic Graph (DAG) model, leveraging interventional data to outperform existing methods. However, as a score-based method over the permutahedron, Intersort is computationally expensice and non-differentiable, limiting its ability to be utilised in problems involving large-scale datasets, such as those in genomics and climate models, or to be integrated into end-to-end gradient-based learning frameworks. We address this limitation by reformulating Intersort using differentiable sorting and ranking techniques. Our approach enables scalable and differentiable optimization of causal orderings, allowing the continuous score function to be incorporated as a regularizer in downstream tasks. Empirical results demonstrate that causal discovery algorithms benefit significantly from regularizing on the causal order, underscoring the effectiveness of our method. Our work opens the door to efficiently incorporating regularization for causal order into the training of differentiable models and thereby addresses a long-standing limitation of purely associational supervised learning.", "keywords": [ "causality", "causal inference", "causal discovery", "interventional data", "single-variable interventions", "causal structure learning" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 3, 2, 2 ], "contribution_list": [ 3, 1, 1, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "G0uhaIXmFw", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Chang Liu", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Yunfan Li", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "lin Yang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "Assistant Professor" } ], "n_formula": 98, "n_formula_1": 34, "n_ref_uni": 23, "n_ref": 46, "n_ref_all": 62, "n_fig": 0, "n_tab": 1, "L_tab": 99, "n_element_tab": 8, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 48, "n_element_tab_1": 7, "formula_len_all": 15984, "formula_len_all_1": 3852, "len_all": 122805, "len_all_1": 66003, "len_abs": 1249, "len_title": 116, "len_sents": 30781, "len_sents_1": 21429, "n_sents": 316, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1328, "title": "Low-Switching Primal-Dual Algorithms for Safe Reinforcement Learning", "abs": "Safety is a key challenge in reinforcement learning (RL), especially in real-world applications like autonomous driving and healthcare. To address this, Constrained Markov Decision Processes (CMDPs) are commonly used to incorporate safety constraints while optimizing performance. However, current methods often face significant safety violations during exploration or suffer from high regret, which represents the performance loss compared to an optimal policy. We propose a low-switching primal-dual algorithm that balances regret with bounded constraint violations, drawing on techniques from online learning and CMDPs. Our approach minimizes policy changes through low-switching updates and enhances sample efficiency using empirical Bernstein-based bonuses. This leads to tighter theoretical bounds on regret and safety, achieving a state-of-the-art regret of $\\tilde{O}(\\sqrt{SAH^5K}/(\\tau - c^0))$, where $S$ and $A$ is the number of states and actions, $H$ is the horizon, $K$ is the number of episodes, and $(\\tau - c^0)$ reflects the safety margin of a known existing safe policy. Our method also ensures a $\\tilde{O}(1)$ constraint violation and removes unnecessary dependencies on state space $S$ and planning horizon $H$ in the reward regret, offering a scalable solution for constrained RL in complex environments.", "keywords": [ "reinforcement learning", "Markov decision process", "constrained Markov decision process", "machine learning", "online learning", "optimization" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "G0dksFayVq", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Alejandro Cuadron", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Chenguang Wang", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "", "position": "Assistant Professor" }, { "name": "Ion Stoica", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Kyle Montgomery", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "US", "position": "PhD student" }, { "name": "Raluca Popa", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Sijun Tan", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Siyuan Zhuang", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "William Yuan Tang", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 103, "n_ref_all": 118, "n_fig": 3, "n_tab": 7, "L_tab": 4955, "n_element_tab": 273, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1694, "n_element_tab_1": 218, "formula_len_all": 85, "formula_len_all_1": 0, "len_all": 213278, "len_all_1": 67154, "len_abs": 1484, "len_title": 103, "len_sents": 58921, "len_sents_1": 30375, "n_sents": 508, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 55, "L_abs": 1499, "title": "JudgeBench: A Benchmark for Evaluating LLM-Based Judges", "abs": "LLM-based judges have emerged as a scalable alternative to human evaluation and are increasingly used to assess, compare, and improve models. However, the reliability of LLM-based judges themselves is rarely scrutinized. As LLMs become more advanced, their responses grow more sophisticated, requiring stronger judges to evaluate them. Existing benchmarks primarily focus on a judge\u2019s alignment with human preferences, but often fail to account for more challenging tasks where crowdsourced human preference is a poor indicator of factual and logical correctness. To address this, we propose a novel evaluation framework to objectively evaluate LLM-based judges. Based on this framework, we propose JudgeBench, a benchmark for evaluating LLM-based judges on challenging response pairs spanning knowledge, reasoning, math, and coding. JudgeBench leverages a novel pipeline for converting existing difficult datasets into challenging response pairs with preference labels reflecting objective correctness. Our comprehensive evaluation on a collection of prompted judges, fine-tuned judges, multi-agent judges, and reward models shows that JudgeBench poses a significantly greater challenge than previous benchmarks, with many strong models (e.g. GPT-4o) performing just slightly better than random guessing. Overall, JudgeBench offers a reliable platform for assessing increasingly advanced LLM-based judges. Data and code are available at \\url{https://anonymous.4open.science/r/JudgeBench-ICLR2025/}.", "keywords": [ "benchmark", "llm-based judges", "llm-as-a-judge", "reward models" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 4 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "Fzz8acgC6X", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Luke James Miller", "gender": "Male", "institution": "University of Missouri, Kansas City", "country": "US", "position": "PhD student" }, { "name": "Yugyung Lee", "gender": "Female", "institution": "University of Missouri - Kansas City", "country": "", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 12, "n_ref_uni": 41, "n_ref": 48, "n_ref_all": 53, "n_fig": 4, "n_tab": 1, "L_tab": 168, "n_element_tab": 20, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 168, "n_element_tab_1": 20, "formula_len_all": 1493, "formula_len_all_1": 1333, "len_all": 110992, "len_all_1": 56680, "len_abs": 1650, "len_title": 114, "len_sents": 32235, "len_sents_1": 27284, "n_sents": 212, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 113, "L_abs": 1655, "title": "Graph Distributional Analytics: Enhancing GNN Explainability through Scalable Embedding and Distribution Analysis", "abs": "Graph Neural Networks (GNNs) have achieved significant success in processing graph-structured data but often lack interpretability, limiting their practical applicability. We introduce the Graph Distributional Analytics (GDA) framework, leveraging novel combinations of scalable techniques to enhance GNN explainability. The integration of Weisfeiler-Leman (WL) graph kernels with distributional distance analysis enables GDA to efficiently quantify graph data distributions, while capturing global structural complexities without significant computational costs. GDA creates high-dimensional embeddings employing WL kernels, measures the distribution of distances from measures of categorical central tendency, and assigns distribution scores to quantify each graph's deviation from this vector We evaluate GDA on the ENZYMES, ogbg-ppa, and MalNet-Tiny datasets. Our experiments demonstrate GDA not only accurately characterizes graph distributions but also outperforms baseline methods in identifying specific structural features responsible for misclassifications. This comprehensive analysis provides deeper insights into how training data distributions affect model performance, particularly with out-of-distribution (OOD) data. By revealing the underlying structural causes of GNN predictions through a novel synergy of established techniques, GDA enhances transparency and offers a practical tool for practitioners to build more interpretable and robust graph-based models. Our framework's scalability, efficiency, and ability to integrate with various embedding methods make it a valuable addition to the suite of tools available for GNN analysis.", "keywords": [ "Graph Neural Networks", "Explainability", "Graph Distributional Analytics", "Weisfeiler-Leman graph kernel", "Graph embeddings", "Distributional analysis", "Out-of-distribution data", "Model transparency", "Structural features", "Machine learning", "Graph classification", "Scalable methods", "GNN interpretability", "Model robustness" ], "rating_list": [ 3, 3, 3, 5, 3 ], "soundness_list": [ 1, 1, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 2, 1 ], "confidence_list": [ 5, 3, 3, 4, 4 ] }, { "paper_id": "FyVuLQNHVi", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Jiankang Deng", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Researcher" }, { "name": "Kaicheng Yang", "gender": "Male", "institution": "DeepGlint", "country": "CN", "position": "Researcher" }, { "name": "Ninghua Yang", "gender": "Male", "institution": "Beijing Geling Deep Eye Information Technology Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "Tiancheng Gu", "gender": "Not Specified", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Weimo Deng", "gender": "Male", "institution": "deepglint", "country": "CN", "position": "Researcher" }, { "name": "Xiang An", "gender": "Male", "institution": "deepglint", "country": "CN", "position": "Researcher" }, { "name": "Xiangzidai", "gender": "Male", "institution": "deepglint", "country": "CN", "position": "Researcher" }, { "name": "Yongle Zhao", "gender": "Male", "institution": "Deepglint", "country": "CN", "position": "Researcher" }, { "name": "Yumeng Wang", "gender": "Male", "institution": "Beijing Deepglint Information Technology Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "Ziyong Feng", "gender": "Male", "institution": "DeepGlint", "country": "CN", "position": "PhD student" }, { "name": "YinXie", "gender": "Male", "institution": "DEEP GLINT", "country": "CN", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 50, "n_ref": 81, "n_ref_all": 98, "n_fig": 10, "n_tab": 7, "L_tab": 3136, "n_element_tab": 560, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1479, "n_element_tab_1": 286, "formula_len_all": 267, "formula_len_all_1": 265, "len_all": 166629, "len_all_1": 56649, "len_abs": 1727, "len_title": 121, "len_sents": 40839, "len_sents_1": 25426, "n_sents": 323, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1448, "title": "Croc: Pretraining Large Multimodal Models with Cross-Modal Comprehension", "abs": "Recent advancements in Large Language Models (LLMs) have catalyzed the development of Large Multimodal Models (LMMs). However, existing research primarily focuses on tuning language and image instructions, ignoring the critical pretraining phase where models learn to process textual and visual modalities jointly. This paper proposes a new pretraining paradigm for LMMs to enhance the visual comprehension capabilities of LLMs by introducing a novel cross-modal comprehension stage. Specifically, we design a dynamically learnable prompt token pool and employ the Hungarian algorithm to replace part of the original visual tokens with the most relevant prompt tokens. Then, we conceptualize visual tokens as analogous to a \"foreign language\" for the LLMs and propose a mixed attention mechanism with bidirectional visual attention and unidirectional textual attention to comprehensively enhance the understanding of visual tokens. Meanwhile, we integrate a detailed caption generation task, leveraging rich descriptions to further facilitate LLMs in understanding visual semantic information. After pretraining on 1.5 million publicly accessible data, we present a new foundation model called Croc. Experimental results demonstrate that Croc achieves new state-of-the-art performance on massive vision-language benchmarks. To support reproducibility and facilitate further research, we will release the training code and pre-trained model weights.", "keywords": [ "Visual Language Models; Pretraining" ], "rating_list": [ 6, 3, 3, 6, 5 ], "soundness_list": [ 3, 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "FyMjfDQ9RO", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Akshat Gupta", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Alan Black", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Cheol Jun Cho", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Dhruv Agarwal", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "Undergrad student" }, { "name": "Ethan Chen", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Gopala Anumanchipalli", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Assistant Professor" }, { "name": "Nicholas Lee", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 2, "n_ref_uni": 57, "n_ref": 127, "n_ref_all": 186, "n_fig": 16, "n_tab": 17, "L_tab": 7115, "n_element_tab": 586, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 2934, "n_element_tab_1": 261, "formula_len_all": 1700, "formula_len_all_1": 260, "len_all": 223769, "len_all_1": 70071, "len_abs": 1433, "len_title": 114, "len_sents": 65102, "len_sents_1": 29870, "n_sents": 517, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1438, "title": "Sylber: Syllabic Embedding Representation of Speech from Raw Audio", "abs": "Syllables are compositional units of spoken language that play a crucial role in human speech perception and production. However, current neural speech representations lack structure, resulting in dense token sequences that are costly to process. To bridge this gap, we propose a new model, Sylber, that produces speech representations with clean and robust syllabic structure. Specifically, we propose a self-supervised model that regresses features on syllabic segments distilled from a teacher model which is an exponential moving average of the model in training. This results in a highly structured representation of speech features, offering three key benefits: 1) a fast, linear-time syllable segmentation algorithm, 2) efficient syllabic tokenization with an average of 4.27 tokens per second, and 3) syllabic units suited for lexical and syntactic understanding. We also train token-to-speech generative models with our syllabic units and show that fully intelligible speech can be reconstructed from these tokens. Lastly, we observe that categorical perception, a linguistic phenomenon of speech perception, emerges naturally in our model, making the embedding space more categorical and sparse than previous self-supervised learning approaches. Together, we present a novel self-supervised approach for representing speech as syllables, with significant potential for efficient speech tokenization and spoken language modeling.", "keywords": [ "Self-supervised learning", "speech representation", "spoken language model", "syllable discovery", "speech segmentation", "speech tokenization" ], "rating_list": [ 5, 8, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 4, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "FyJaV0TVF2", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "RuiqingYan", "gender": "unknown", "institution": "University of New South Wales", "country": "AU", "position": "MS student" }, { "name": "Xingbo Du", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Yufeng Guo", "gender": "Female", "institution": "College of William and Mary", "country": "US", "position": "PhD student" }, { "name": "Han Zou", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Jianfei Yang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "ZhengLinghan", "gender": "Male", "institution": "Tongji University", "country": "", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 26, "n_ref": 39, "n_ref_all": 58, "n_fig": 8, "n_tab": 6, "L_tab": 1670, "n_element_tab": 149, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 1293, "n_element_tab_1": 140, "formula_len_all": 570, "formula_len_all_1": 570, "len_all": 104304, "len_all_1": 55940, "len_abs": 1227, "len_title": 106, "len_sents": 33900, "len_sents_1": 26264, "n_sents": 235, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1232, "title": "RecurFormer: Not All Transformer Heads Need Self-Attention", "abs": "Transformer-based large language models (LLMs) excel in modeling complex language patterns but face significant computational costs during inference, especially with long inputs due to the attention mechanism's memory overhead. We observe that certain attention heads exhibit a distribution where the attention weights concentrate on tokens near the query token, termed as recency aware, which focuses on local and short-range dependencies. Leveraging this insight, we propose RecurFormer, a novel architecture that replaces these attention heads with linear recurrent neural networks (RNNs), specifically the Mamba architecture. This replacement reduces the cache size without evicting tokens, thus maintaining generation quality. RecurFormer retains the ability to model long-range dependencies through the remaining attention heads and allows for reusing pre-trained Transformer-based LLMs weights with continual training. Experiments demonstrate that RecurFormer matches the original model's performance while significantly enhancing inference efficiency. Our approach provides a practical solution to the computational challenges of Transformer-based LLMs inference, making it highly attractive for tasks involving long inputs.", "keywords": [ "Transformer", "Large Language Models", "Recurrent Neural Networks" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Fxsd66d2wB", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Diyuan Shi", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Donglin Wang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Associate Professor" }, { "name": "Jinxin Liu", "gender": "unknown", "institution": "Westlake Robotics", "country": "CN", "position": "Researcher" }, { "name": "Xiao He", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "RA" }, { "name": "Zifeng Zhuang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Ziqi Zhang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Micro Masters Student" } ], "n_formula": 20, "n_formula_1": 14, "n_ref_uni": 37, "n_ref": 86, "n_ref_all": 115, "n_fig": 7, "n_tab": 5, "L_tab": 827, "n_element_tab": 41, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 249, "n_element_tab_1": 14, "formula_len_all": 1412, "formula_len_all_1": 947, "len_all": 161327, "len_all_1": 80661, "len_abs": 1460, "len_title": 105, "len_sents": 46992, "len_sents_1": 30940, "n_sents": 337, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1467, "title": "Decoupled Offline to Online finetuning via Dynamics Model", "abs": "Constrained by the sub-optimal dataset in offline reinforcement learning (RL), the offline trained agent should be online finetuned before deployment. Due to the conservative offline algorithms and unbalanced state distribution in offline dataset, offline to online finetuning faces severe distribution shift. This shift will disturb the policy improvement during online interaction, even a performance drop. A natural yet unexplored idea is whether policy improvement can be decoupled from distribution shift. In this work, we propose a decoupled offline to online finetuning framework using the dynamics model from model-based methods. During online interaction, only dynamics model is finetuned to overcome the distribution shift. Then the policy is finetuned in offline manner with finetuned dynamics and without further interaction. As a result, online stage only needs to deal with a simpler supervised dynamics learning, rather than the complex policy improvement with the interference from distribution shift. When finetuning the policy, we adopt the offline approach, which ensures the conservatism of the algorithm and fundamentally avoids the sudden performance crashes. We conduct extensive evaluation on the classical datasets of offline RL, demonstrating the effective elimination of distribution shift, stable and superior policy finetuning performance, and exceptional interaction efficiency within our decouple offline to online finetuning framework.", "keywords": [ "Offline to Online Finetuning", "Model-based RL", "Decoupled Framework" ], "rating_list": [ 6, 3, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 5, 4, 3 ] }, { "paper_id": "FxNNiUgtfa", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Yuanzhi Li", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Zeyuan Allen-Zhu", "gender": "Unspecified", "institution": "Meta FAIR Labs", "country": "US", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 6, "n_ref_uni": 27, "n_ref": 67, "n_ref_all": 199, "n_fig": 25, "n_tab": 3, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1375, "n_element_tab_1": 36, "formula_len_all": 6119, "formula_len_all_1": 745, "len_all": 281146, "len_all_1": 75638, "len_abs": 1735, "len_title": 100, "len_sents": 102940, "len_sents_1": 33505, "n_sents": 901, "n_sents_1": 288, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 69, "L_abs": 852, "title": "Physics of Language Models: Part 3.3, Knowledge Capacity Scaling Laws", "abs": "Scaling laws describe the relationship between the size of language models and their capabilities. Unlike prior studies that evaluate a model's capability via loss or benchmarks, we estimate information-theoretically the number of knowledge \\emph{bits} a model stores. We focus on factual knowledge represented as tuples, such as (USA, capital, Washington D.C.) from a Wikipedia page. Through multiple controlled datasets, we establish that language models can and only can store \\emph{2 bits of knowledge per parameter, even when quantized to int8}, and such knowledge can be flexibly extracted for downstream applications. \n\nMore broadly, we present 12 results on how (1) training duration, (2) model architecture, (3) quantization, (4) sparsity constraints such as MoE, and (5) data signal-to-noise ratio affect a model's knowledge storage capacity.", "keywords": [ "scaling laws", "knowledge capacity", "language models" ], "rating_list": [ 10, 8, 6, 5 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "FxLxbJTm7F", "primary_area": "foundation or frontier models, including LLMs", "n_author": 14, "author_info_list": [ { "name": "Albert Ge", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "", "position": "PhD student" }, { "name": "Angeliki Giannou", "gender": "Female", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Dimitris Papailiopoulos", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "Associate Professor" }, { "name": "Grigorios Chrysos", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "John Cooper", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Kangwook Lee", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "US", "position": "Assistant Professor" }, { "name": "Liu Yang", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Saurabh Agarwal", "gender": "Male", "institution": "University of Wisconsin, Madison", "country": "", "position": "Student" }, { "name": "Vasilis Papageorgiou", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "PhD student" }, { "name": "Zack Sifakis", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Zheyang Xiong", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Researcher" }, { "name": "Ziqian Lin", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Ziyang Cai", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "MS student" }, { "name": "Samet Oymak", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" } ], "n_formula": 56, "n_formula_1": 7, "n_ref_uni": 45, "n_ref": 66, "n_ref_all": 114, "n_fig": 21, "n_tab": 2, "L_tab": 504, "n_element_tab": 47, "n_fig_1": 13, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7635, "formula_len_all_1": 585, "len_all": 194231, "len_all_1": 60769, "len_abs": 1279, "len_title": 141, "len_sents": 56185, "len_sents_1": 28707, "n_sents": 456, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1046, "title": "Everything Everywhere All at Once: LLMs can In-Context Learn Multiple Tasks in Superposition", "abs": "Large Language Models (LLMs) have demonstrated remarkable in-context learning (ICL) capabilities. In this study, we explore a surprising phenomenon related to ICL: LLMs can perform multiple, computationally distinct ICL tasks simultaneously, during a single inference call, a capability we term \"task superposition\". We provide empirical evidence of this phenomenon across various LLM families and scales and show that this phenomenon emerges even if we train the model to in-context learn one task at a time. We offer theoretical explanations that this capability is well within the expressive power of transformers. We also explore how LLMs internally compose task vectors during superposition. Furthermore, we show that larger models can solve more ICL tasks in parallel, and better calibrate their output distribution. Our findings offer insights into the latent capabilities of LLMs, further substantiate the perspective of \"LLMs as superposition of simulators\", and raise questions about the mechanisms enabling simultaneous task execution.", "keywords": [ "Large Language Model", "In-Context Learning", "Task Superposition" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "Fx4fFBjn6j", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Seulki Lee", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Yunseok Lee", "gender": "Male", "institution": "Ulsan National Institute of Science and Technology", "country": "KR", "position": "MS student" } ], "n_formula": 32, "n_formula_1": 19, "n_ref_uni": 33, "n_ref": 114, "n_ref_all": 176, "n_fig": 1, "n_tab": 3, "L_tab": 1768, "n_element_tab": 295, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2285, "n_element_tab_1": 302, "formula_len_all": 2893, "formula_len_all_1": 1214, "len_all": 131971, "len_all_1": 72399, "len_abs": 1502, "len_title": 126, "len_sents": 41576, "len_sents_1": 30942, "n_sents": 284, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1507, "title": "Enabling Fine-Tuning of Direct Feedback Alignment via Feedback-Weight Matching", "abs": "In this paper, we introduce feedback-weight matching, a new method that facilitates reliable fine-tuning of fully connected neural networks using Direct Feedback Alignment (DFA). Although DFA has demonstrated potential by enabling efficient and parallel updates of weight parameters through direct propagation of the network's output error, its usage has been primarily restricted to training networks from scratch. We provide the first analysis showing that existing standard DFA struggles to fine-tune networks pre-trained via back-propagation. Through an analysis of weight alignment (WA) and gradient alignment (GA), we show that the proposed feedback-weight matching enhances DFA's ability and stability in fine-tuning pre-trained networks, providing insights into DFA's behavior and characteristics when applied to fine-tuning. In addition, we find that feedback-weight matching, when combined with weight decay, not only mitigates over-fitting but also further reduces the network output error, leading to improved learning performance during DFA-based fine-tuning. Our experimental results show that, for the first time, feedback-weight matching enables reliable and superior fine-tuning across various fine-tuning tasks compared to existing standard DFA, e.g., achieving 7.97\\% accuracy improvement on image classification tasks (i.e., 82.67\\% vs. 74.70\\%) and 0.66 higher correlation score on NLP tasks (i.e., 0.76 vs. 0.10). The code implementation is available at an anonymous GitHub repository.", "keywords": [ "direct feedback alignment", "deep learning", "fine tuning" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 2, 2, 2 ] }, { "paper_id": "FwlM1k4ODx", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Luca M Ghiringhelli", "gender": "Not Specified", "institution": "Humboldt Universit\u00e4t Berlin", "country": "", "position": "Principal Researcher" }, { "name": "Luigi Sbail\u00f2", "gender": "Male", "institution": "Friedrich-Alexander-Universtitat", "country": "DE", "position": "Postdoc" } ], "n_formula": 41, "n_formula_1": 10, "n_ref_uni": 53, "n_ref": 76, "n_ref_all": 94, "n_fig": 3, "n_tab": 3, "L_tab": 3868, "n_element_tab": 332, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 2722, "n_element_tab_1": 205, "formula_len_all": 1490, "formula_len_all_1": 335, "len_all": 186373, "len_all_1": 71207, "len_abs": 1334, "len_title": 138, "len_sents": 62995, "len_sents_1": 35119, "n_sents": 425, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1340, "title": "Latent Point Collapse Induces an Information Bottleneck in Deep Neural Network Classifiers", "abs": "The information-bottleneck principle suggests that the foundation of learning lies in the ability to create compact representations. In machine learning, this goal can be formulated as a Lagrangian optimization problem, where the mutual information between the input and latent representations must be minimized without compromising the correctness of the model's predictions.\nUnfortunately, mutual information is difficult to compute in deterministic deep neural network classifiers, which greatly limits the application of this approach to challenging scenarios. In this paper, we tackle this problem from a different perspective that does not involve direct computation of the mutual information. We develop a method that induces the collapse of latent representations belonging to the same class into a single point. \nThis point collapse not only significantly reduces the entropy of the latent distribution, thereby creating an information bottleneck that correlates with improved generalization, but also makes the network Lipschitz, offering guarantees for enhanced robustness.\nOur method is straightforward to implement. We demonstrate that it substantially improves the network's robustness, provides a small yet statistically significant increase in generalization, and enhances the network's ability to detect misclassifications.", "keywords": [ "classification", "regularization", "information bottleneck", "latent representations." ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "FwkYeLovHk", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "James Bailey", "gender": "unknown", "institution": "The University of Melbourne", "country": "AU", "position": "Full Professor" }, { "name": "Jinhao Li", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" }, { "name": "Lei Feng", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Assistant Professor" }, { "name": "Sarah Monazam Erfani", "gender": "unknown", "institution": "The University of Melbourne", "country": "AU", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 22, "n_ref": 51, "n_ref_all": 62, "n_fig": 2, "n_tab": 5, "L_tab": 2860, "n_element_tab": 259, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2574, "n_element_tab_1": 289, "formula_len_all": 737, "formula_len_all_1": 737, "len_all": 111188, "len_all_1": 69321, "len_abs": 1321, "len_title": 117, "len_sents": 31927, "len_sents_1": 30011, "n_sents": 239, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1340, "title": "Exploring Weak-to-Strong Generalization for CLIP-based Classification", "abs": "Aligning large-scale commercial models with user intent is crucial to preventing harmful outputs. Current methods rely on human supervision but become impractical as model complexity increases. When models surpass human knowledge, providing accurate feedback becomes challenging and inefficient.\nA novel solution proposed recently is using a weaker model to supervise a stronger model. This concept leverages the ability of weaker models to perform evaluations, thereby reducing the workload on human supervisors. \nPrevious work has shown the effectiveness of weak-to-strong generalization in the context of language-only models. Extending this concept to vision-language models leverages these insights, adapting the proven benefits to a multi-modal context.\nIn our study, we explore weak-to-strong generalization for CLIP-based classification. We propose a method, \\emph{class prototype learning} (CPL), which aims to enhance the classification capabilities of the CLIP model, by learning more representative prototypes for each category.\nOur findings indicate that despite the simple loss function under weak supervision, CPL yields robust results.\nOur experiments are conducted on challenging datasets to evaluate our method. Extensive experiments show that our method is effective, achieving a 3.67\\% improvement over baseline methods.", "keywords": [ "Vision-Language Models", "Weak-to-Strong Generalization" ], "rating_list": [ 3, 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 3, 3, 1, 1, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "FwjEZZ3j91", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Haizhao Yang", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Associate Professor" }, { "name": "Kusum Choudhary", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Sikai Huang", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Tara Adusumilli", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "Undergrad student" }, { "name": "Yixin Wen", "gender": "Female", "institution": "University of Florida", "country": "US", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 32, "n_ref": 42, "n_ref_all": 53, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 531, "formula_len_all_1": 531, "len_all": 98739, "len_all_1": 60853, "len_abs": 1467, "len_title": 149, "len_sents": 32469, "len_sents_1": 30886, "n_sents": 217, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1849, "title": "Parsing the Language of Expressions: Enhancing Symbolic Regression with Domain-Aware Symbolic Priors", "abs": "Symbolic regression is pivotal for discovering interpretable expressions that unravel complex phenomena by revealing underlying mathematical and physical relationships within data. In this paper, we introduce an enhanced symbolic regression method that integrates symbol priors derived from diverse scientific domains\u2014including physics, biology, chemistry, and engineering\u2014into the regression process. By organizing and analyzing domain-specific expressions, we examine the probability distributions of symbols across different topics. We introduce a novel tree-structured recurrent neural networks (RNNs) infused with these symbol priors to guide the learning process using domain knowledge. In our approach, we introduce a new tree structure to represent expressions, where unary operators connected by the same binary operator are positioned at the same hierarchical level. By analyzing the combinations of symbols at different heights and levels within the tree, we are able to examine symbol priors across the entire hierarchical structure. This effectively incorporates the structural information of expressions into the regression process. Additionally, we compile characteristic expression blocks from each domain and incorporate them into the operator dictionary during training, expediting learning by providing relevant building blocks. Experimental results demonstrate that incorporating symbol priors significantly boosts the performance of symbolic regression methods. Specifically, it accelerates the efficiency of reinforcement learning algorithms in obtaining optimal policies. Our findings confirm that leveraging domain-specific symbol priors not only hastens convergence but also yields more accurate and interpretable models, effectively bridging the gap between data-driven learning and expert expertise in symbolic regression.", "keywords": [ "Symbolic regression", "Reinforcement learning", "Recurrent neural network", "Domain knowledge prior" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "FwdN0KovFp", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Hyunjun Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Undergrad student" }, { "name": "Sang Wan Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Yoondo Sung", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Youngha Jo", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Myounghoon Ha", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "Postdoc" } ], "n_formula": 53, "n_formula_1": 16, "n_ref_uni": 42, "n_ref": 80, "n_ref_all": 140, "n_fig": 41, "n_tab": 5, "L_tab": 637, "n_element_tab": 65, "n_fig_1": 25, "n_tab_1": 1, "L_tab_1": 151, "n_element_tab_1": 19, "formula_len_all": 3809, "formula_len_all_1": 949, "len_all": 225149, "len_all_1": 77490, "len_abs": 1344, "len_title": 101, "len_sents": 78648, "len_sents_1": 32815, "n_sents": 618, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1350, "title": "Towards Stable Learning in Predictive Coding Networks", "abs": "Predictive coding (PC) offers a biologically plausible model of cortical functions, encompassing processes such as learning, prediction, encoding, and memory. However, predictive coding networks (PCNs) face significant challenges in stability and scalability, which constrain our capacity to elucidate cortical computation. Our study identifies instability in PCNs as a fundamental issue, focusing on the exponential growth of latent state norms and prediction errors after inference. These dynamics lead to exploding and vanishing gradients in PCNs. Moreover, the concentration of prediction errors near the input and output layer impedes effective learning, exacerbating performance degradation as network depth increases. To address these limitations, we propose stabilizing techniques for PCNs, including length regularization and sequential training with skip connection modules. This approach counteracts the exponential growth of latent states and makes the distribution of prediction errors more uniform across layers. Empirical evaluations demonstrate that our approach enhances stability and generalization, enabling the training of deeper networks more efficiently. This study deepens our understanding of complex dynamics in cortical networks, thereby advancing the practical application of predictive coding theory to its full potential.", "keywords": [ "Neuroscience", "Predictive Coding", "Predictive Coding Networks", "Neural Networks" ], "rating_list": [ 3, 5, 6, 1 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 5, 3, 3, 5 ] }, { "paper_id": "FwW3jqchtY", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Amin Nejatbakhsh", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Research Scientist" }, { "name": "Yixin Wang", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 60, "n_ref": 74, "n_ref_all": 111, "n_fig": 26, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 23, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 619, "formula_len_all_1": 539, "len_all": 141979, "len_all_1": 64481, "len_abs": 1362, "len_title": 116, "len_sents": 39323, "len_sents_1": 30435, "n_sents": 334, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1369, "title": "Identifying neural dynamics using interventional state space models", "abs": "Neural circuits produce signals that are complex and nonlinear. To facilitate the understanding of neural dynamics, a popular approach is to fit state space models (SSM) to data and analyze the dynamics of the low-dimensional latent variables. Despite the power of SSM in explaining neural circuit dynamics, it has been shown that these models merely capture statistical associations in the data and cannot be causally interpreted. Therefore, an important research problem is to build models that can predict neural dynamics under causal manipulations. Here, we propose interventional state space models (iSSM), a class of causal models that can predict neural responses to novel perturbations. We draw on recent advances in causal dynamical systems and present theoretical results for the identifiability of iSSM. In simulations of the motor cortex, we show that iSSM can recover the true latents and the underlying dynamics. In addition, we illustrate two applications of iSSM in biological datasets. First, we apply iSSM to a dataset of calcium recordings from ALM neurons in mice during photostimulation and uncover dynamical mechanisms underlying short-term memory. Second, we apply iSSM to a dataset of electrophysiological recordings from macaque dlPFC recordings during micro-stimulation and show that it successfully predicts responses to unseen perturbations.", "keywords": [ "Causal dynamical systems", "interventions", "state space models", "photostimulation", "micro-stimulation" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "FvjcdS42o1", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Harshil Vagadia", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Kartik Garg", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Sehoon Ha", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Seungeun Rho", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 23, "n_ref": 42, "n_ref_all": 62, "n_fig": 9, "n_tab": 1, "L_tab": 13, "n_element_tab": 1, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 812, "formula_len_all_1": 805, "len_all": 98889, "len_all_1": 55058, "len_abs": 1376, "len_title": 120, "len_sents": 29085, "len_sents_1": 24639, "n_sents": 241, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1081, "title": "Interplay Between Task Learning and Skill Discovery for Agile Locomotion", "abs": "Agile locomotion of legged robots, characterized by high momentum and frequent contact changes, is a challenging task that demands precise motor control. Therefore, the training process for such skills often relies on additional techniques, such as reward engineering, expert demonstrations, and curriculum learning. However, these requirements hinder the generalizability of methods because we may lack sufficient prior knowledge or demonstration datasets for some tasks. In this work, we consider the problem of automated learning agile motions using its intrinsic motivation, which can greatly reduce the effort of a human engineer. Inspired by unsupervised skill discovery, our learning framework encourages the agent to explore various skills to maximize the given task reward. Finally, we train a parameter to balance the two distinct rewards through a bi-level optimization process. We demonstrate that our method can train quadrupeds to perform highly agile motions, ranging from crawling, jumping, and leaping to complex maneuvers such as jumping off a perpendicular wall.", "keywords": [ "Unsupervised Skill Discovery", "Reinforcement Learning", "Robot Learning", "Locomotion" ], "rating_list": [ 6, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "FviefuxmeW", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "David Mark Bossens", "gender": "unknown", "institution": "A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Heyang Zhao", "gender": "Male", "institution": "Computer Science Department, University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Xingrui Yu", "gender": "Male", "institution": "Institute of High Performance Computing, Singapore, A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Ivor W Tsang", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Principal Researcher" } ], "n_formula": 48, "n_formula_1": 35, "n_ref_uni": 48, "n_ref": 82, "n_ref_all": 98, "n_fig": 1, "n_tab": 21, "L_tab": 5860, "n_element_tab": 507, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1412, "n_element_tab_1": 68, "formula_len_all": 6664, "formula_len_all_1": 3357, "len_all": 185252, "len_all_1": 70357, "len_abs": 1355, "len_title": 155, "len_sents": 52037, "len_sents_1": 29415, "n_sents": 425, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 107, "L_abs": 1302, "title": "Beyond-Expert Performance with Limited Demonstrations: Efficient Imitation Learning with Double Exploration", "abs": "learning where the goal is to learn a policy that mimics the expert's behavior. In practice, it is often challenging to learn the expert policy from a limited number of demonstrations accurately due to the complexity of the state space. Moreover, it is essential to explore the environment and collect data to achieve beyond-expert performance. To overcome these challenges, we propose a novel imitation learning algorithm namely Imitation Learning with Double Exploration (ILDE), which implements exploration in two aspects: (1) optimistic policy optimization via an exploration bonus that rewards state-action pairs with high uncertainty to potentially improve the convergence to the expert policy, and (2) curiosity-driven exploration of the states that deviate from the demonstration trajectories to potentially yield beyond-expert performance. Empirically, we demonstrate that ILDE outperforms the state-of-the-art imitation learning algorithms in terms of sample efficiency and achieves beyond-expert performance on Atari and MuJoCo tasks with fewer demonstrations than those in previous work. We also provide theoretical justification of ILDE as an uncertainty-regularized policy optimization method with optimistic exploration, leading to a regret growing sublinearly in the number of episodes.", "keywords": [ "Reinforcement learning", "imitation learning" ], "rating_list": [ 8, 3, 8, 5 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 4, 2, 4, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "Fvfs0HPuKl", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Jianyu Zhang", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Leon Bottou", "gender": "Male", "institution": "Facebook AI Research", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 45, "n_ref": 81, "n_ref_all": 92, "n_fig": 5, "n_tab": 6, "L_tab": 2071, "n_element_tab": 182, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 4657, "n_element_tab_1": 457, "formula_len_all": 134, "formula_len_all_1": 134, "len_all": 127329, "len_all_1": 63936, "len_abs": 1289, "len_title": 83, "len_sents": 31061, "len_sents_1": 27843, "n_sents": 209, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1295, "title": "fine-tuning with very large dropout", "abs": "It is impossible today to pretend that the practice of machine learning is compatible with the idea that training and testing data follow the same distribution. Several authors have recently used ensemble techniques to show how scenarios involving multiple data distributions are best served by representations that are both richer than those obtained by regularizing for the best in-distribution performance, and richer than those obtained under the influence of the implicit sparsity bias of common stochastic gradient procedures.\n\nThis contribution investigates the use of very high dropout rates instead of ensembles to obtain such rich representations. Although training a deep network from scratch using such dropout rates is virtually impossible, fine-tuning a large pre-trained model under such conditions is not only possible but also achieves out-of-distribution performances that exceed those of both ensembles and weight averaging methods such as model soups. \n\nThis result has practical significance because the importance of the fine-tuning scenario has considerably grown in recent years. This result also provides interesting insights on the nature of rich representations and on the intrinsically linear nature of fine-tuning a large network using a comparatively small dataset.", "keywords": [ "out-of-distribution", "fine-tuning", "very large dropout" ], "rating_list": [ 3, 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 1, 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 3, 4 ] }, { "paper_id": "FvQsk3la17", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Doina Precup", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Research Team Lead" }, { "name": "Guangyuan Wang", "gender": "Male", "institution": "Mila - Quebec Artificial Intelligence Institute", "country": "CA", "position": "Researcher" }, { "name": "Haque M Ishfaq", "gender": "Male", "institution": "McGill University, McGill University", "country": "CA", "position": "PhD student" }, { "name": "Mohammad Sami Nur Islam", "gender": "Not Specified", "institution": "McGill University", "country": "CA", "position": "Undergrad student" } ], "n_formula": 48, "n_formula_1": 28, "n_ref_uni": 49, "n_ref": 146, "n_ref_all": 192, "n_fig": 17, "n_tab": 7, "L_tab": 3085, "n_element_tab": 148, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 659, "n_element_tab_1": 23, "formula_len_all": 5512, "formula_len_all_1": 2844, "len_all": 204885, "len_all_1": 67645, "len_abs": 1268, "len_title": 141, "len_sents": 54210, "len_sents_1": 27152, "n_sents": 428, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1080, "title": "Langevin Soft Actor-Critic: Efficient Exploration through Uncertainty-Driven Critic Learning", "abs": "Existing actor-critic algorithms, which are popular for continuous control reinforcement learning (RL) tasks, suffer from poor sample efficiency due to lack of principled exploration mechanism within them. Motivated by the success of Thompson sampling for efficient exploration in RL, we propose a novel model-free RL algorithm, \\emph{Langevin Soft Actor Critic} (LSAC), which prioritizes enhancing critic learning through uncertainty estimation over policy optimization. LSAC employs three key innovations: approximate Thompson sampling through distributional Langevin Monte Carlo (LMC) based $Q$ updates, parallel tempering for exploring multiple modes of the posterior of the $Q$ function, and diffusion synthesized state-action samples regularized with $Q$ action gradients. Our extensive experiments demonstrate that LSAC outperforms or matches the performance of mainstream model-free RL algorithms for continuous control tasks.\nNotably, LSAC marks the first successful application of an LMC based Thompson sampling in continuous control tasks with continuous action spaces.", "keywords": [ "Actor-Critic", "Exploration", "Reinforcement Learning", "Thompson Sampling", "Langevin Monte Carlo", "Deep Reinforcement learning", "Continuous Control" ], "rating_list": [ 8, 6, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "FvIASa0tau", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Jianqi Chen", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Michael Kampffmeyer", "gender": "Male", "institution": "UiT The Arctic University of Norway", "country": "NO", "position": "Full Professor" }, { "name": "Panwen Hu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "Xiaodan Liang", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Zhenwei Shi", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaojun Chang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 49, "n_ref": 116, "n_ref_all": 157, "n_fig": 20, "n_tab": 8, "L_tab": 2461, "n_element_tab": 267, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 270, "n_element_tab_1": 31, "formula_len_all": 358, "formula_len_all_1": 96, "len_all": 216517, "len_all_1": 64792, "len_abs": 315, "len_title": 121, "len_sents": 75119, "len_sents_1": 32901, "n_sents": 553, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1720, "title": "Sitcom-Crafter: A Plot-Driven Human Motion Generation System in 3D Scenes", "abs": "Recent advancements in human motion synthesis have focused on specific types of motions, such as human-scene interaction, locomotion or human-human interaction, however, there is a lack of a unified system capable of generating a diverse combination of motion types. In response, we introduce *Sitcom-Crafter*, a comprehensive and extendable system for human motion generation in 3D space, which can be guided by extensive plot contexts to enhance workflow efficiency for anime and game designers. The system is comprised of eight modules, three of which are dedicated to motion generation, while the remaining five are augmentation modules that ensure consistent fusion of motion sequences and system functionality. Central to the generation modules is our novel 3D scene-aware human-human interaction module, which addresses collision issues by synthesizing implicit 3D Signed Distance Function (SDF) points around motion spaces, thereby minimizing human-scene collisions without additional data collection costs. Complementing this, our locomotion and human-scene interaction modules leverage existing methods to enrich the system's motion generation capabilities. Augmentation modules encompass plot comprehension for command generation, motion synchronization for seamless integration of different motion types, hand pose retrieval to enhance motion realism, motion collision revision to prevent human collisions, and 3D retargeting to ensure visual fidelity. Experimental evaluations validate the system's ability to generate high-quality, diverse, and physically realistic motions, underscoring its potential for advancing creative workflows. Code and demonstration videos can be found in the supplementary files.", "keywords": [ "Human Motion Synthesis", "Human-Human Interaction", "Physically Compliant Motion", "Creative Workflow Automation" ], "rating_list": [ 8, 6, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 4, 3, 2 ], "contribution_list": [ 4, 3, 2 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "FvBTy5Dz9C", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Defu Cao", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Wen Ye", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Yan Liu", "gender": "Female", "institution": "University of Southern California", "country": "US", "position": "Professor" }, { "name": "Yizhou Zhang", "gender": "unknown", "institution": "University of Southern California", "country": "", "position": "PhD student" } ], "n_formula": 35, "n_formula_1": 15, "n_ref_uni": 68, "n_ref": 130, "n_ref_all": 160, "n_fig": 4, "n_tab": 20, "L_tab": 18442, "n_element_tab": 926, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 3332, "n_element_tab_1": 298, "formula_len_all": 1886, "formula_len_all_1": 1127, "len_all": 317943, "len_all_1": 74996, "len_abs": 6385, "len_title": 129, "len_sents": 84547, "len_sents_1": 31571, "n_sents": 603, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1636, "title": "TimeDiT: General-purpose Diffusion Transformers for Time Series Foundation Model", "abs": "With recent advances in building foundation models for text and video data, such as Large Language Models (LLMs), there is a surge of interest in foundation modeling for time series. However, real-world time series exhibit unique challenges, such as variable channel sizes across domains, missing values, and varying signal sampling intervals due to the multi-resolution nature of real-world data, which pose fundamental challenges for current de-fact transformer models with rigid architectural choices and predetermined parameter settings. Additionally, the unidirectional nature of temporally autoregressive decoding typically learns a deterministic mapping relationship and limits the incorporation of domain knowledge, such as physical laws. To address these challenges, we introduce the Time Diffusion Transformer (TimeDiT), a general foundation model for time series that jointly leverages the transformer inductive bias to capture temporal dependencies and the diffusion processes to generate high-quality candidate samples. The proposed mask unit for task-agnostic pretraining and task-specific sampling enables direct processing of multivariate inputs even with missing values or multi-resolution. Furthermore, we introduce a theoretically justified finetuning-free model editing strategy that allows the flexible integration of external knowledge during the sampling process. Extensive experiments conducted on a variety of tasks, such as forecasting, imputation, and anomaly detection highlight TimeDiT\u2019s adaptability as a foundation model, addressing diverse time series challenges and advancing analysis in various fields.", "keywords": [ "Time Series; Foundation model; Diffusion model" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 2 ] }, { "paper_id": "Fur0DtynPX", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Gao Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Honghui Wang", "gender": "Male", "institution": "Department of Automation, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Shiji Song", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 16, "n_formula_1": 10, "n_ref_uni": 36, "n_ref": 86, "n_ref_all": 113, "n_fig": 11, "n_tab": 7, "L_tab": 1566, "n_element_tab": 139, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1324, "n_element_tab_1": 109, "formula_len_all": 1271, "formula_len_all_1": 1092, "len_all": 155524, "len_all_1": 67936, "len_abs": 1094, "len_title": 119, "len_sents": 50069, "len_sents_1": 30276, "n_sents": 402, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 71, "L_abs": 1100, "title": "GridMix: Exploring Spatial Modulation for Neural Fields in PDE Modeling", "abs": "Significant advancements have been achieved in PDE modeling using neural fields. Despite their effectiveness, existing methods rely on global modulation, limiting their ability to reconstruct local details. While spatial modulation with vanilla grid-based representations offers a promising alternative, it struggles with inadequate global information modeling and over-fitting to the training spatial domain. To address these challenges, we propose GridMix, a novel approach that models spatial modulation as a mixture of grid-based representations. GridMix effectively explores global structures while preserving locality for fine-grained modulation. Furthermore, we introduce spatial domain augmentation to enhance the robustness of the modulated neural fields against spatial domain variations. \nWith all these innovations,\nour comprehensive approach culminates in MARBLE, a framework that significantly advancing the capabilities of neural fields in PDE modeling. The effectiveness of MARBLE is extensively\nvalidated on diverse benchmarks encompassing dynamics modeling and geometric prediction.", "keywords": [ "Partial Differential Equations", "Neural Fields" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "FuXtwQs7pj", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Dmytro Kozakov", "gender": "unknown", "institution": "Trustees of Boston University", "country": "", "position": "" }, { "name": "Ernest Glukhov", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "", "position": "PhD student" }, { "name": "Evangelos Coutsias", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Sergei Kotelnikov", "gender": "unknown", "institution": "State University of New York at Stony Brook", "country": "", "position": "PhD student" }, { "name": "Thu Nguyen", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "", "position": "PhD student" }, { "name": "Xin Cao", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Postdoc" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 35, "n_ref": 54, "n_ref_all": 68, "n_fig": 5, "n_tab": 2, "L_tab": 290, "n_element_tab": 46, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 290, "n_element_tab_1": 46, "formula_len_all": 697, "formula_len_all_1": 697, "len_all": 103514, "len_all_1": 66292, "len_abs": 1161, "len_title": 127, "len_sents": 34008, "len_sents_1": 32948, "n_sents": 221, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1167, "title": "A diffusion model on toric varieties with application to protein loop modeling", "abs": "The conformation spaces of loop regions in proteins as well as closed kinematic linkages in robotics can be described by systems of polynomial equations, forming Toric varieties. These are real algebraic varieties, formulated as the zero sets of polynomial equations constraining the rotor angles in a linkage or macromolecular chain. These spaces are essentially stitched manifolds and contain singularities. Diffusion models have achieved spectacular success in applications in Cartesian space and smooth manifolds but have not been extended to varieties. Here we develop a diffusion model on the underlying variety by utilizing an appropriate Jacobian, whose loss of rank indicates singularities. This allows our method to explore the variety, without encountering singular or infeasible states. We demonstrated the approach on two important protein structure prediction problems: one is prediction of Major Histocompatibility Complex (MHC) peptide interactions, a critical part in the design of neoantigen vaccines, and the other is loop prediction for nanobodies, an important class of drugs. In both, we improve upon the state of the art open source AlphaFold.", "keywords": [ "Protein loop", "Toric varieties", "Jacobian", "Tangent space", "MHC", "Nanobody" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 2, 3, 2 ] }, { "paper_id": "Fu0aggezN9", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Di Wen", "gender": "Male", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "PhD student" }, { "name": "Jiaming Zhang", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Junwei Zheng", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "DE", "position": "PhD student" }, { "name": "Kunyu Peng", "gender": "Female", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "Postdoc" }, { "name": "Rainer Stiefelhagen", "gender": "Male", "institution": "Karlsruhe Institute of Technology", "country": "DE", "position": "Full Professor" }, { "name": "Ruiping Liu", "gender": "Female", "institution": "Karlsruher Institut f\u00fcr Technologie", "country": "DE", "position": "PhD student" }, { "name": "Yufan Chen", "gender": "Male", "institution": "Karlsruhe Institute for Technology", "country": "DE", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 6, "n_ref_uni": 11, "n_ref": 27, "n_ref_all": 73, "n_fig": 19, "n_tab": 11, "L_tab": 4188, "n_element_tab": 1408, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 602, "n_element_tab_1": 15, "formula_len_all": 919, "formula_len_all_1": 980, "len_all": 197386, "len_all_1": 62893, "len_abs": 0, "len_title": 87, "len_sents": 64126, "len_sents_1": 30490, "n_sents": 444, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 39, "L_abs": 1495, "title": "Graph-based Document Structure Analysis", "abs": "When reading a document, glancing at the spatial layout of a document is an initial step to understand it roughly. Traditional document layout analysis (DLA) methods, however, offer only a superficial parsing of documents, focusing on basic instance detection and often failing to capture the nuanced spatial and logical relationships between instances. These limitations hinder DLA-based models from achieving a gradually deeper comprehension akin to human reading. In this work, we propose a novel graph-based Document Structure Analysis (gDSA) task. This task requires that model not only detects document elements but also generates spatial and logical relations in form of a graph structure, allowing to understand documents in a holistic and intuitive manner. For this new task, we construct a relation graph-based document structure analysis dataset(GraphDoc) with 80K document images and 4.13M relation annotations, enabling training models to complete multiple tasks like reading order, hierarchical structures analysis, and complex inter-element relationship inference. Furthermore, a document relation graph generator (DRGG) is proposed to address the gDSA task, which achieves performance with 57.6% at $mAP_g$@$0.5$ for a strong benchmark baseline on this novel task and dataset. We hope this graphical representation of document structure can mark an innovative advancement in document structure analysis and understanding. The new dataset and code will be made publicly available.", "keywords": [ "Document Structure Analysis", "Document Layout Analysis", "Document Relational Graph" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "FtyHMDQSlD", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Bingde Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Haonan Lu", "gender": "Male", "institution": "OPPO Guangdong Mobile Telecommunications Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Qingsong Xie", "gender": "Male", "institution": "OPPO", "country": "CN", "position": "Researcher" }, { "name": "Yu Lei", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhijie Deng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 19, "n_formula_1": 15, "n_ref_uni": 58, "n_ref": 98, "n_ref_all": 141, "n_fig": 23, "n_tab": 4, "L_tab": 145, "n_element_tab": 22, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 238, "n_element_tab_1": 36, "formula_len_all": 1770, "formula_len_all_1": 1424, "len_all": 183503, "len_all_1": 68975, "len_abs": 1867, "len_title": 138, "len_sents": 44906, "len_sents_1": 29973, "n_sents": 404, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1505, "title": "Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation", "abs": "Text-to-3D generation based on score distillation of pre-trained 2D diffusion models has gained increasing interest, with variational score distillation (VSD) as a remarkable example. \nVSD proves that vanilla score distillation can be improved by introducing an extra score-based model, which characterizes the distribution of images rendered from 3D models, to correct the distillation gradient. \nDespite the theoretical foundations, VSD, in practice, is likely to suffer from slow and sometimes ill-posed convergence.\nIn this paper, we perform an in-depth investigation of the interplay between the introduced score model and the 3D model, and find that we can simply adjust their optimization order to improve the generation quality. \nBy doing so, the score model looks ahead to the current 3D state and hence yields more reasonable corrections. \nNevertheless, naive lookahead VSD may suffer from unstable training in practice due to the potential over-fitting. \nTo address this, we propose to use a linearized variant of the model for score distillation, giving rise to the Linearized Lookahead Variational Score Distillation ($L^2$-VSD). \n$L^2$-VSD can be realized efficiently with forward-mode autodiff functionalities of existing deep learning libraries. \nExtensive experiments validate the efficacy of $L^2$-VSD, revealing its clear superiority over prior score distillation-based methods. \nWe also show that our method can be seamlessly incorporated into any other VSD-based text-to-3D framework.", "keywords": [ "Text-to-3D generation; Diffusion model; Linearized Lookahead" ], "rating_list": [ 5, 5, 3, 3, 6 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "Fty0wTcemV", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Ishika Agarwal", "gender": "Female", "institution": "University of Illinois, Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Krishnateja Killamsetty", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Lucian Popa", "gender": "Male", "institution": "International Business Machines", "country": "", "position": "Principal Research Staff Member" }, { "name": "Marina Danilevsky", "gender": "Female", "institution": "International Business Machines", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 35, "n_ref": 64, "n_ref_all": 79, "n_fig": 2, "n_tab": 8, "L_tab": 4829, "n_element_tab": 752, "n_fig_1": 1, "n_tab_1": 8, "L_tab_1": 4176, "n_element_tab_1": 573, "formula_len_all": 359, "formula_len_all_1": 359, "len_all": 187198, "len_all_1": 66313, "len_abs": 1399, "len_title": 110, "len_sents": 43139, "len_sents_1": 26751, "n_sents": 280, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 61, "L_abs": 1399, "title": "DELIFT: Data Efficient Language model Instruction Fine-Tuning", "abs": "Fine-tuning large language models (LLMs) is essential for enhancing their performance on specific tasks but is often resource-intensive due to redundant or uninformative data. To address this inefficiency, we introduce DELIFT (Data Efficient Language model Instruction Fine-Tuning), a novel algorithm that systematically optimizes data selection across the three key stages of fine-tuning: (1) instruction tuning, (2) task-specific fine-tuning (e.g., reasoning, question-answering), and (3) continual learning (e.g., incorporating new data versions). Unlike existing methods that focus on single-stage optimization or rely on computationally intensive gradient calculations, DELIFT operates efficiently across all stages. Central to our approach is a pairwise utility metric that quantifies how beneficial a data sample is for improving the model's responses to other samples, effectively measuring the informational value relative to the model's current capabilities. By leveraging different submodular functions applied to this metric, DELIFT selects diverse and optimal subsets that are useful across all stages of fine-tuning. Experiments across various tasks and model scales demonstrate that DELIFT can reduce the fine-tuning data size by up to 70% without compromising performance, offering significant computational savings and outperforming existing methods in both efficiency and efficacy.", "keywords": [ "Instruction Fine-Tuning; Data Selection; Efficient Fine-Tuning; Submodular Functions" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "FtjLUHyZAO", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Molei Tao", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Peng Qiu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Associate Professor" }, { "name": "Sichen Zhu", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yuchen Zhu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 43, "n_ref": 114, "n_ref_all": 147, "n_fig": 12, "n_tab": 5, "L_tab": 1643, "n_element_tab": 282, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1577, "n_element_tab_1": 222, "formula_len_all": 525, "formula_len_all_1": 368, "len_all": 174703, "len_all_1": 65230, "len_abs": 1640, "len_title": 149, "len_sents": 53321, "len_sents_1": 32088, "n_sents": 356, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1827, "title": "Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images", "abs": "Spatial Transcriptomics (ST) allows a high-resolution measurement of RNA sequence abundance by systematically connecting cell morphology depicted in Hematoxylin and eosin (H\\&E) stained histology images to spatially resolved gene expressions. ST is a time-consuming, expensive yet powerful experimental technique that provides new opportunities to understand cancer mechanisms at a fine-grained molecular level, which is critical for uncovering new approaches for disease diagnosis and treatments. Here, we present $\\textbf{Stem}$ ($\\underline{\\textbf{S}}$pa$\\underline{\\textbf{T}}$ially resolved gene $\\underline{\\textbf{E}}$xpression inference with diffusion $\\underline{\\textbf{M}}$odel), a novel computational tool that leverages a conditional diffusion generative model to enable in silico gene expression inference from H&E stained images. Through better capturing the inherent stochasticity and heterogeneity in ST data, $\\textbf{Stem}$ achieves state-of-the-art performance on spatial gene expression prediction and generates biologically meaningful gene profiles for new H&E stained images at test time. We evaluate the proposed algorithm on datasets with various tissue sources and sequencing platforms, where it demonstrates clear improvement over existing approaches. $\\textbf{Stem}$ generates high-fidelity gene expression predictions that share similar gene variation levels as ground truth data, suggesting that our method preserves the underlying biological heterogeneity. Our proposed pipeline opens up the possibility of analyzing existing, easily accessible H&E stained histology images from a genomics point of view without physically performing gene expression profiling and empowers potential biological discovery from H&E stained histology images. Code is available at: https://github.com/SichenZhu/Stem.", "keywords": [ "Gene Expression Prediction", "Diffusion Model", "Spatial Transcriptomics", "H&E" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "FtX6oAW7Dd", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Dong-Dong Wu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Intern" }, { "name": "Gang Niu", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Senior Research Scientist (tenured)" }, { "name": "Jindong Wang", "gender": "Male", "institution": "William & Mary", "country": "US", "position": "Assistant Professor" }, { "name": "Masashi Sugiyama", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Director" }, { "name": "Min-Ling Zhang", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Full Professor" }, { "name": "Wei Wang", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 57, "n_ref": 194, "n_ref_all": 208, "n_fig": 6, "n_tab": 12, "L_tab": 11375, "n_element_tab": 958, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 6617, "n_element_tab_1": 499, "formula_len_all": 517, "formula_len_all_1": 325, "len_all": 228731, "len_all_1": 72386, "len_abs": 1675, "len_title": 118, "len_sents": 41733, "len_sents_1": 27161, "n_sents": 414, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 62, "L_abs": 1668, "title": "Realistic Evaluation of Deep Partial-Label Learning Algorithms", "abs": "Partial-label learning (PLL) is a weakly supervised learning problem in which\neach example is associated with multiple candidate labels and only one is the\ntrue label. In recent years, many deep PLL algorithms have been developed to\nimprove model performance. However, we find that some early developed\nalgorithms are often underestimated and can outperform many later algorithms\nwith complicated designs. In this paper, we delve into the empirical\nperspective of PLL and identify several critical but previously overlooked\nissues. First, model selection for PLL is non-trivial, but has never been\nsystematically studied. Second, the experimental settings are highly\ninconsistent, making it difficult to evaluate the effectiveness of the\nalgorithms. Third, there is a lack of real-world image datasets that can be\ncompatible with modern network architectures. Based on these findings, we\npropose PLENCH, the first Partial-Label learning bENCHmark to systematically\ncompare state-of-the-art deep PLL algorithms. We investigate the model\nselection problem for PLL for the first time, and propose novel model selection\ncriteria with theoretical guarantees. We also create Partial-Label CIFAR-10\n(PLCIFAR10), an image dataset of human-annotated partial labels collected from\nAmazon Mechanical Turk, to provide a testbed for evaluating the performance of\nPLL algorithms in more realistic scenarios. Researchers can quickly and\nconveniently perform a comprehensive and fair evaluation and verify the\neffectiveness of newly developed algorithms based on PLENCH. We hope that\nPLENCH will facilitate standardized, fair, and practical evaluation of PLL\nalgorithms in the future.", "keywords": [ "Partial-label learning", "weakly supervised learning", "benchmark." ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "FsgGBhNIt4", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Jing Wu", "gender": "unknown", "institution": "Cardiff University", "country": "GB", "position": "Lecturer" }, { "name": "Yipeng Qin", "gender": "unknown", "institution": "Cardiff University", "country": "GB", "position": "Associate Professor" }, { "name": "shuang song", "gender": "Male", "institution": "Cardiff University", "country": "GB", "position": "PhD student" }, { "name": "Yu-Kun Lai", "gender": "Male", "institution": "Cardiff University", "country": "GB", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 22, "n_ref": 44, "n_ref_all": 59, "n_fig": 7, "n_tab": 7, "L_tab": 4181, "n_element_tab": 462, "n_fig_1": 7, "n_tab_1": 7, "L_tab_1": 4696, "n_element_tab_1": 256, "formula_len_all": 102, "formula_len_all_1": 128, "len_all": 91661, "len_all_1": 51496, "len_abs": 1368, "len_title": 120, "len_sents": 24931, "len_sents_1": 21201, "n_sents": 175, "n_sents_1": 133, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1372, "title": "Unsupervised Learning of Facial Attribute Representations Using StyleGAN", "abs": "Facial attributes (e.g., gender, age) encompass important social cues and play a pivotal role in computer vision. While supervised methods have dominated facial attribute analysis, they often require large annotated datasets, which are costly and time-consuming to create.\nIn this work, we circumvent this limitation by proposing a novel unsupervised learning framework that leverages StyleGAN to learn rich and disentangled facial attribute representations. Specifically, unlike prior methods that rely on labeled datasets or supervised techniques, our approach exploits the unique inductive bias of StyleGAN, namely Hierarchical Feature Modulation, to automatically discover semantically meaningful representations of facial attributes. This inductive bias enables StyleGAN to generate disentangled and interpretable facial attribute features at different layers, benefiting a variety of downstream tasks. To leverage StyleGAN representations, we employ GAN inversion methods to represent input images as StyleGAN features and propose a simple yet effective feature reduction method based on mutual information to improve the effectiveness and efficiency of the learned representations. Extensive experiments in few-shot facial attribute analysis tasks, including clustering, classification, and facial attribute annotation demonstrate the effectiveness of our approach.", "keywords": [ "facial attributes", "unsupervised representation learning", "GAN", "StyleGAN" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 1, 2 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "Fs9EabmQrJ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Andrew Li", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Jiantao Jiao", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Assistant Professor" }, { "name": "Kannan Ramchandran", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Richard Zhuang", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Tianhao Wu", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Zhaojin Wen", "gender": "Female", "institution": "University of California, Berkeley", "country": "", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 31, "n_ref": 43, "n_ref_all": 54, "n_fig": 4, "n_tab": 1, "L_tab": 167, "n_element_tab": 12, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 269, "n_element_tab_1": 33, "formula_len_all": 282, "formula_len_all_1": 282, "len_all": 109219, "len_all_1": 54501, "len_abs": 1250, "len_title": 68, "len_sents": 30841, "len_sents_1": 26066, "n_sents": 198, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 67, "L_abs": 1226, "title": "EmbedLLM: Learning Compact Representations of Large Language Models", "abs": "With hundreds of thousands of language models available on Huggingface today, efficiently evaluating and utilizing these models across various downstream tasks has become increasingly critical. Many existing methods repeatedly learn task-specific representations of Large Language Models (LLMs), which leads to inefficiencies in both time and computational resources. To address this, we propose EmbedLLM, a framework designed to learn compact vector representations of LLMs that facilitate downstream applications involving many models, such as model routing. We introduce an encoder-decoder approach for learning such embedding, along with a systematic framework to evaluate their effectiveness. Empirical results show that EmbedLLM outperforms prior methods in model routing. Additionally, we demonstrate that our method can forecast a model's performance on multiple benchmarks, without incurring additional inference cost. Extensive probing experiments validate that the learned embeddings capture key model characteristics, e.g. whether the model is specialized for coding tasks, even without being explicitly trained on them. We open source our dataset, code and embedder to facilitate further research and application.", "keywords": [ "Large Language Models", "Representation Learning", "Model Routing" ], "rating_list": [ 8, 8, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 4, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "Frok9AItud", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Cassiano O Becker", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Jennifer Neville", "gender": "Female", "institution": "Purdue University", "country": "", "position": "Associate Professor" }, { "name": "Tvrtko Tadi\u0107", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 185, "n_formula_1": 38, "n_ref_uni": 39, "n_ref": 48, "n_ref_all": 86, "n_fig": 5, "n_tab": 3, "L_tab": 706, "n_element_tab": 36, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 558, "n_element_tab_1": 26, "formula_len_all": 17316, "formula_len_all_1": 3485, "len_all": 217231, "len_all_1": 73443, "len_abs": 1079, "len_title": 121, "len_sents": 56639, "len_sents_1": 27076, "n_sents": 781, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1083, "title": "Node Similarities under Random Projections: Limits and Pathological Cases", "abs": "Random Projections have been widely used to generate embeddings for various graph learning tasks due to their computational efficiency. The majority of applications have been justified through the Johnson-Lindenstrauss Lemma. In this paper, we take a step further and investigate how well dot product and cosine similarity are preserved by random projections when these are applied over the rows of the graph matrix. Our analysis provides new asymptotic and finite-sample results, identifies pathological cases, and tests them with numerical experiments. We specialize our fundamental results to a ranking application by computing the probability of random projections flipping the node ordering induced by their embeddings. We find that, depending on the degree distribution, the method produces especially unreliable embeddings for the dot product, regardless of whether the adjacency or the normalized transition matrix is used. With respect to the statistical noise introduced by random projections, we show that cosine similarity produces remarkably more precise approximations.", "keywords": [ "Johnson\u2013Lindenstrauss Lemma", "Random Projections", "Graph Embeddings", "Ranking", "Cosine Similarity" ], "rating_list": [ 3, 6, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 2, 4 ], "contribution_list": [ 2, 3, 3, 2, 4 ], "confidence_list": [ 5, 2, 4, 4, 3 ] }, { "paper_id": "FrmVRUVOEF", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Alexander Modell", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 28, "n_ref": 47, "n_ref_all": 52, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 369, "n_element_tab_1": 11, "formula_len_all": 804, "formula_len_all_1": 803, "len_all": 76308, "len_all_1": 44521, "len_abs": 1262, "len_title": 104, "len_sents": 22775, "len_sents_1": 20492, "n_sents": 159, "n_sents_1": 143, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1265, "title": "Temporally coherent visualisation of time-dependent data", "abs": "Dimension reduction algorithms aim to embed high-dimensional datasets into a low-dimensional space in such a way that important structural properties, such as clusters and manifolds, are preserved. Most such methods are designed for static data, and naively applying them to time-dependent data can lead to unstable embeddings which do not meaningfully capture the temporal evolution of the data. In this paper, we propose a new variant of the t-SNE algorithm for time-dependent data, TC-tSNE (Temporally Coherent t-SNE) in which an extra term is added to the cost function to promote temporal coherence: the notion that a data point which has a similar position in two time frames should be embedded to similar positions at those times. Importantly, this notion captures temporal similarities over the entire time domain and can therefore capture long-range temporal patterns, not just local ones. We demonstrate the effectiveness of our method for visualising dynamic network embedding, and we evaluate our method on six benchmark datasets using a collection of metrics, which capture the structural quality and the temporal coherence of the embeddings. We compare our method with existing dynamic visualisation algorithms and find that it performs competitively.", "keywords": [ "visualisation", "t-SNE", "dimension reduction", "dynamic data", "temporal data" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "FrjTgprk3V", "primary_area": "learning theory", "n_author": 6, "author_info_list": [ { "name": "Adityanarayanan Radhakrishnan", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Daniel Beaglehole", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Libin Zhu", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Postdoc" }, { "name": "Neil Rohit Mallinar", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Parthe Pandit", "gender": "Male", "institution": "Indian Institute of Technology, Bombay", "country": "IN", "position": "Assistant Professor" }, { "name": "Misha Belkin", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "Professor" } ], "n_formula": 30, "n_formula_1": 8, "n_ref_uni": 35, "n_ref": 80, "n_ref_all": 129, "n_fig": 18, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 159, "n_element_tab_1": 23, "formula_len_all": 2409, "formula_len_all_1": 580, "len_all": 170703, "len_all_1": 66211, "len_abs": 906, "len_title": 142, "len_sents": 57650, "len_sents_1": 30547, "n_sents": 477, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 910, "title": "Emergence in non-neural models: grokking modular arithmetic via average gradient outer product", "abs": "Neural networks trained to solve modular arithmetic tasks exhibit grokking, the phenomenon where the test accuracy improves only long after the model achieves 100% training accuracy in the training process. It is often taken as an example of ``emergence'', where model ability manifests sharply through a phase transition. In this work, we show that the phenomenon of grokking is not specific to neural networks nor to gradient descent-based optimization. Specifically, we show that grokking occurs when learning modular arithmetic with Recursive Feature Machines (RFM), an iterative algorithm that uses the Average Gradient Outer Product (AGOP) to enable task-specific feature learning with kernel machines. We show that RFM and, furthermore, neural networks that solve modular arithmetic learn block-circulant features transformations which implement the previously proposed Fourier multiplication algorithm.", "keywords": [ "Theory of deep learning", "grokking", "modular arithmetic", "feature learning", "kernel methods", "average gradient outer product (AGOP)", "emergence" ], "rating_list": [ 5, 5, 8 ], "soundness_list": [ 2, 3, 4 ], "presentation_list": [ 3, 4, 4 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "Frhj9T7ihK", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Abigail Jarvis", "gender": "Female", "institution": "Berry College", "country": "US", "position": "Undergrad student" }, { "name": "Angelina Faith Pennino", "gender": "Female", "institution": "Berry College", "country": "US", "position": "Undergrad student" }, { "name": "Kun Wang", "gender": "unknown", "institution": "University of Iowa", "country": "", "position": "PhD student" }, { "name": "Kuo Deng", "gender": "Female", "institution": "Berry College", "country": "US", "position": "Assistant Professor" }, { "name": "Xiaomeng Ye", "gender": "Male", "institution": "Berry College", "country": "US", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 39, "n_ref": 63, "n_ref_all": 65, "n_fig": 0, "n_tab": 3, "L_tab": 926, "n_element_tab": 19, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 1486, "n_element_tab_1": 21, "formula_len_all": 147, "formula_len_all_1": 79, "len_all": 136398, "len_all_1": 61793, "len_abs": 934, "len_title": 180, "len_sents": 45256, "len_sents_1": 31968, "n_sents": 355, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 129, "L_abs": 938, "title": "All Models are Biased, Some are More Transparent about it: Fully Interpretable and Adjustable Model for Mental Disorder Diagnosis", "abs": "Recent advances in machine learning have enabled AI applications in mental disorder diagnosis, but many methods remain black-box or rely on post-hoc explanations which are not straightforward or actionable for mental health practitioners. Meanwhile, interpretable methods, such as k-nearest neighbors (k-NN) classification, struggle with complex or high-dimensional data. A network-based k-NN model (NN-kNN) combines the interpretability with the predictive power of neural networks. The model prediction can be fully explained in terms of activated features and neighboring cases. We experimented with the model to predict the risks of depression and interviewed practitioners. The feedback of the practitioners emphasized the model's adaptability, integration of clinical expertise, and transparency in the diagnostic process, highlighting its potential to ethically improve the diagnostic precision and confidence of the practitioner.", "keywords": [ "interpretable AI", "mental health", "k-NN", "neural symbolic", "explainable AI", "tunnable AI" ], "rating_list": [ 5, 3, 1, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 1, 2 ], "contribution_list": [ 2, 1, 4, 2 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "FrFQpAgnGE", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Yoon Kim", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Zhaofeng Wu", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Dani Yogatama", "gender": "unknown", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "jiasen lu", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Xinyan Yu", "gender": "Female", "institution": "University of Southern California", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 56, "n_ref": 89, "n_ref_all": 131, "n_fig": 25, "n_tab": 3, "L_tab": 1650, "n_element_tab": 138, "n_fig_1": 16, "n_tab_1": 1, "L_tab_1": 528, "n_element_tab_1": 39, "formula_len_all": 349, "formula_len_all_1": 297, "len_all": 182408, "len_all_1": 74579, "len_abs": 1117, "len_title": 111, "len_sents": 50368, "len_sents_1": 34842, "n_sents": 349, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 107, "L_abs": 1360, "title": "The Semantic Hub Hypothesis: Language Models Share Semantic Representations Across Languages and Modalities", "abs": "Modern language models can process inputs across diverse languages and modalities. We hypothesize that models acquire this capability through learning a _shared representation space_ across heterogeneous data types (e.g., different languages and modalities), which places semantically similar inputs near one another, even if they are from different modalities/languages. We term this the _semantic hub hypothesis_, following the hub-and-spoke model from neuroscience (Patterson et al., 2007) which posits that semantic knowledge in the human brain is organized through a transmodal semantic \"hub\" which integrates information from various modality-specific ``spokes'' regions. We first show that model representations for semantically equivalent inputs in different languages are similar in the intermediate layers, and that this space can be interpreted using the model's dominant pretraining language via the logit lens. This tendency extends to other data types, including arithmetic expressions, code, and visual/audio inputs. Interventions in the shared representation space in one data type also predictably affect model outputs in other data types, suggesting that this shared representations space is not simply a vestigial byproduct of large-scale training on broad data, but something that is actively utilized by the model during input processing.", "keywords": [ "interpretability", "representation", "multilingual", "multimodal" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 3, 3, 2, 4 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "Fr6bjeqRec", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Chao Lei", "gender": "Male", "institution": "University of Melbourne", "country": "", "position": "PhD student" }, { "name": "Nir Lipovetzky", "gender": "unknown", "institution": "The University of Melbourne", "country": "", "position": "Assistant Professor" }, { "name": "Yanchuan Chang", "gender": "unknown", "institution": "University of Melbourne", "country": "AU", "position": "Postdoc" }, { "name": "Krista Ehinger", "gender": "Not Specified", "institution": "The University of Melbourne", "country": "AU", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 2, "n_ref_uni": 51, "n_ref": 150, "n_ref_all": 223, "n_fig": 13, "n_tab": 14, "L_tab": 4795, "n_element_tab": 273, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1193, "n_element_tab_1": 114, "formula_len_all": 548, "formula_len_all_1": 74, "len_all": 285929, "len_all_1": 71704, "len_abs": 1932, "len_title": 121, "len_sents": 94128, "len_sents_1": 34723, "n_sents": 655, "n_sents_1": 226, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1947, "title": "Planning-Driven Programming: A Large Language Model Programming Workflow", "abs": "The strong performance of large language models (LLMs) on natural language processing tasks raises extensive discussion on their application to code generation. Recent work suggests multiple sampling approaches to improve initial code generation accuracy or program repair approaches to refine the code. However, these methods suffer from LLMs' inefficiencies and limited reasoning capacity. In this work, we propose an LLM programming workflow (LPW) designed to improve both initial code generation and subsequent refinements within a structured two-phase workflow. Specifically, in the solution generation phase, the LLM first outlines a solution plan that decomposes the problem into manageable sub-problems and then verifies the generated solution plan through visible test cases. Subsequently, in the code implementation phase, the LLM initially drafts a code according to the solution plan and its verification. If the generated code fails the visible tests, the plan verification serves as the intended natural language solution to consistently inform the refinement process for correcting bugs. We further introduce SLPW, a sampling variant of LPW, which initially generates multiple solution plans and plan verifications, produces a program for each plan and its verification, and refines each program as necessary until one successfully passes the visible tests. Compared to the state-of-the-art methods across various existing LLMs, our experimental results show that LPW significantly improves the Pass@1 accuracy by up to 16.4\\% on well-established text-to-code generation benchmarks, especially with a notable improvement of around 10\\% on challenging benchmarks. Additionally, SLPW demonstrates up to a 5.6\\% improvement over LPW and sets new state-of-the-art Pass@1 accuracy on various benchmarks, e.g., 98.2\\% on HumanEval, 84.8\\% on MBPP, 64.0\\% on APPS, and 35.3\\% on CodeContest, using the advanced LLM GPT-4o as the backbone.", "keywords": [ "program synthesis", "large language model", "code generation", "reasoning" ], "rating_list": [ 3, 3, 5, 8 ], "soundness_list": [ 2, 1, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "FqWtMGw8tt", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Chulin Xie", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "PhD student" }, { "name": "Dawn Song", "gender": "Female", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "YiPeng Chen", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Undergrad student" }, { "name": "Zinan Lin", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Senior Researcher" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 36, "n_ref": 73, "n_ref_all": 98, "n_fig": 8, "n_tab": 11, "L_tab": 4096, "n_element_tab": 349, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 2556, "n_element_tab_1": 237, "formula_len_all": 336, "formula_len_all_1": 125, "len_all": 185745, "len_all_1": 81635, "len_abs": 1646, "len_title": 124, "len_sents": 62277, "len_sents_1": 37097, "n_sents": 414, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1661, "title": "KnowData: Knowledge-Enabled Data Generation for Improving Multimodal Models", "abs": "In this paper, we introduce a framework to enhance the quality of synthetic image-text pairs for multimodal models such as CLIP. Our approach, named KnowData, integrates real-world knowledge explicitly into the generation of text descriptions. It combines structured knowledge from knowledge graphs like ConceptNet and unstructured knowledge extracted from Wikipedia, to ensure that the generated text descriptions are both contextually rich and accurately reflective of real-world knowledge. Additionally, we leverage Large Language Models for the expansion, summarization, and refinement of the text descriptions to ensure their coherence. These enriched texts are subsequently used to generate images through advanced text-to-image models like Stable Diffusion and DALLE-3. CLIP models are then fine-tuned with these synthetic image-text pairs for zero-shot classification tasks. Our experiments across 9 datasets demonstrate that CLIP models fine-tuned with our knowledge-guided synthetic datasets outperform state-of-the-art (SOTA) zero-shot CLIP methods (e.g., +11.23% on DTD and +4% on EuroSAT based on ViT-B/16 model; +11.47% on CIFAR-100 and +7.99% on DTD based on ResNet-50 model). These results showcase the improved out-of-distribution robustness and adaptability of our approach across a diverse set of data domains. We further substantiate the design of KnowData through ablation studies, revealing that the integration of knowledge not only enhances zero-shot performance but also contributes to the reliability, diversity, and detail-orientation of the generated synthetic images, thereby offering better data scaling laws for model performance.", "keywords": [ "Multimodal learning; Knowledge integration; Synthetic datasets; Zero-shot classification" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "Fq25rH3ytL", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Chenliang Xu", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Associate Professor" }, { "name": "Jiacan Yu", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "MS student" }, { "name": "Mingqian Feng", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Rongyi Zhu", "gender": "unknown", "institution": "Department of Computer Science, University of Rochester", "country": "", "position": "MS student" }, { "name": "Zeliang Zhang", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 17, "n_ref_uni": 53, "n_ref": 80, "n_ref_all": 98, "n_fig": 9, "n_tab": 2, "L_tab": 136, "n_element_tab": 15, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2339, "formula_len_all_1": 1274, "len_all": 141593, "len_all_1": 61101, "len_abs": 1055, "len_title": 119, "len_sents": 35736, "len_sents_1": 28393, "n_sents": 274, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 998, "title": "Is Forward Gradient an Effective Tool for Explaining Black-box Models?", "abs": "Gradients are widely used to explain the decisions of deep neural networks. However, as models become deeper and more complex, computing gradients becomes challenging and sometimes infeasible, hindering traditional explanation methods. Recently, the forward gradient method has garnered attention for training structure-agnostic models with discontinuous objective functions. This method perturbs only the parameters of interest for gradient computation and optimization. Inspired by this, we investigate whether the forward gradient can be employed to explain black-box models. In this work, we use the likelihood ratio method to estimate output-to-input gradients and utilize them for the explanation of model decision. Additionally, we propose block-wise computation techniques to enhance estimation accuracy. Extensive experiments in black-box settings validate the effectiveness of our method, demonstrating accurate gradient estimation and improved explainability under the black-box setting.", "keywords": [ "explainability" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "FpxKYYk6V5", "primary_area": "causal reasoning", "n_author": 7, "author_info_list": [ { "name": "Bin Qin", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Fanjiang Xu", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Jiangmeng Li", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Jianwen Cao", "gender": "Male", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Qirui Ji", "gender": "Male", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Wenwen Qiang", "gender": "Male", "institution": "Institute of Software Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "yi Li", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 37, "n_formula_1": 14, "n_ref_uni": 60, "n_ref": 127, "n_ref_all": 177, "n_fig": 14, "n_tab": 5, "L_tab": 1865, "n_element_tab": 166, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4124, "formula_len_all_1": 1261, "len_all": 196586, "len_all_1": 65523, "len_abs": 1865, "len_title": 148, "len_sents": 60277, "len_sents_1": 32409, "n_sents": 404, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 98, "L_abs": 1873, "title": "Teleporter Theory: A General and Simple Approach for Modeling Cross-World Counterfactual Causality", "abs": "Leveraging the development of structural causal model (SCM), researchers can establish graphical models for exploring the causal mechanisms behind machine learning techniques. As the complexity of machine learning applications rises, single-world interventionism causal analysis encounters theoretical adaptation limitations. Accordingly, cross-world counterfactual approach extends our understanding of causality beyond observed data, enabling hypothetical reasoning about alternative scenarios. However, the joint involvement of cross-world variables, encompassing counterfactual variables and factual variables, challenges the construction of the graphical model. Existing approaches, e.g., Twin Network and Single World Intervention Graphs (SWIG), establish a symbiotic relationship to bridge the gap between graphical modeling and the introduction of counterfactuals albeit with room for improvement in generalization. In this regard, we demonstrate the theoretical limitations of certain current methods in cross-world counterfactual scenarios. To this end, we propose a novel teleporter theory to establish a general and simple graphical representation of counterfactuals, which provides criteria for determining teleporter variables to connect multiple worlds. In theoretical application, we determine that introducing the proposed teleporter theory can directly obtain the conditional independence between counterfactual variables and factual variables from the cross-world SCM without requiring complex algebraic derivations. Accordingly, we can further identify counterfactual causal effects through cross-world symbolic derivation. We demonstrate the generality of the teleporter theory to the practical application. Adhering to the proposed theory, we build a plug-and-play module, and the effectiveness of which are substantiated by experiments on benchmarks.", "keywords": [ "Cross-World", "Counterfactual", "Causal Model" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 4, 2, 1 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "FpiCLJrSW8", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Aaron Jiaxun Li", "gender": "Male", "institution": "Harvard University", "country": "", "position": "MS student" }, { "name": "Satyapriya Krishna", "gender": "Male", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Hima Lakkaraju", "gender": "Female", "institution": "Harvard University", "country": "", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 53, "n_ref": 89, "n_ref_all": 112, "n_fig": 9, "n_tab": 3, "L_tab": 578, "n_element_tab": 101, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 254, "n_element_tab_1": 22, "formula_len_all": 1206, "formula_len_all_1": 1208, "len_all": 157740, "len_all_1": 57991, "len_abs": 1374, "len_title": 128, "len_sents": 39076, "len_sents_1": 28106, "n_sents": 265, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 79, "L_abs": 1377, "title": "More RLHF, More Trust? On The Impact of Preference Alignment On Trustworthiness", "abs": "The trustworthiness of Large Language Models (LLMs) refers to the extent to which their outputs are reliable, safe, and ethically aligned, and it has become a crucial consideration alongside their cognitive performance. In practice, Reinforcement Learning From Human Feedback (RLHF) has been widely used to align LLMs with labeled human preferences, but its assumed effect on model trustworthiness hasn't been rigorously evaluated. To bridge this knowledge gap, this study investigates how models aligned with general-purpose preference data perform across five trustworthiness verticals: toxicity, stereotypical bias, machine ethics, truthfulness, and privacy. Our results demonstrate that RLHF on human preferences doesn't automatically guarantee trustworthiness, and reverse effects are often observed. Furthermore, we propose to adapt efficient influence function based data attribution methods to the RLHF setting to better understand the influence of fine-tuning data on individual trustworthiness benchmarks, and show its feasibility by providing our estimated attribution scores. Together, our results underscore the need for more nuanced approaches for model alignment from both the data and framework perspectives, and we hope this research will guide the community towards developing language models that are increasingly capable without sacrificing trustworthiness.", "keywords": [ "Large Language Model", "Trustworthy ML", "Data Attribution" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 4, 3, 4, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "FowFLhUTgO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jie Zhou", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jiwen Lu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Wenzhao Zheng", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Yuanhui Huang", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Wang Chengkun", "gender": "Male", "institution": "Tsinghua university", "country": "CN", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 38, "n_ref": 87, "n_ref_all": 106, "n_fig": 4, "n_tab": 2, "L_tab": 560, "n_element_tab": 65, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 732, "n_element_tab_1": 369, "formula_len_all": 691, "formula_len_all_1": 670, "len_all": 115021, "len_all_1": 66303, "len_abs": 1395, "len_title": 113, "len_sents": 28595, "len_sents_1": 24707, "n_sents": 197, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1399, "title": "V2M: Visual 2-Dimensional Mamba for Image Representation Learning", "abs": "Mamba has garnered widespread attention due to its flexible design and efficient hardware performance to process 1D sequences based on the state space model (SSM). Recent studies have attempted to apply Mamba to the visual domain by flattening 2D images into patches and then regarding them as a 1D sequence. To compensate for the 2D structure information loss (e.g., local similarity) of the original image, most existing methods focus on designing different orders to sequentially process the tokens, which could only alleviate this issue to some extent. In this paper, we propose a Visual 2-Dimensional Mamba (V2M) model as a complete solution, which directly processes image tokens in the 2D space. We first generalize SSM to the 2-dimensional space which generates the next state considering two adjacent states on both dimensions (e.g., columns and rows). We then construct our V2M based on the 2-dimensional SSM formulation and incorporate Mamba to achieve hardware-efficient parallel processing. The proposed V2M effectively incorporates the 2D locality prior yet inherits the efficiency and input-dependent scalability of Mamba. Extensive experimental results on ImageNet classification and downstream visual tasks including object detection and instance segmentation on COCO and semantic segmentation on ADE20K demonstrate the effectiveness of our V2M compared with other visual backbones.", "keywords": [ "Mamba", "Visual Backbone", "Representation Learning" ], "rating_list": [ 5, 3, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "FoUpv84hMw", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Haochen Zhang", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Lingzhou Xue", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Full Professor" }, { "name": "Zhong Zheng", "gender": "Male", "institution": "Pennsylvania State University", "country": "", "position": "PhD student" } ], "n_formula": 365, "n_formula_1": 43, "n_ref_uni": 29, "n_ref": 84, "n_ref_all": 204, "n_fig": 1, "n_tab": 4, "L_tab": 873, "n_element_tab": 27, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1180, "n_element_tab_1": 50, "formula_len_all": 53348, "formula_len_all_1": 5477, "len_all": 338566, "len_all_1": 78961, "len_abs": 5202, "len_title": 165, "len_sents": 73922, "len_sents_1": 27455, "n_sents": 886, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 119, "L_abs": 1144, "title": "Federated $Q$-Learning with Reference-Advantage Decomposition: Almost Optimal Regret and Logarithmic Communication Cost", "abs": "In this paper, we consider model-free federated reinforcement learning for tabular episodic Markov decision processes. Under the coordination of a central server, multiple agents collaboratively explore the environment and learn an optimal policy without sharing their raw data. Despite recent advances in federated $Q$-learning algorithms achieving near-linear regret speedup with low communication cost, existing algorithms only attain suboptimal regrets compared to the information bound. We propose a novel model-free federated $Q$-Learning algorithm, termed FedQ-Advantage. Our algorithm leverages reference-advantage decomposition for variance reduction and adopts three novel designs: separate event-triggered communication and policy switching, heterogeneous communication triggering conditions, and optional forced synchronization. We prove that our algorithm not only requires a lower logarithmic communication cost but also achieves an almost optimal regret, reaching the information bound up to a logarithmic factor and near-linear regret speedup compared to its single-agent counterpart when the time horizon is sufficiently large.", "keywords": [ "Federated Learning", "Reinforcement Learning", "variance reduction", "communication cost" ], "rating_list": [ 6, 6, 3, 8, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 1, 3, 3 ], "contribution_list": [ 3, 3, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 4, 3 ] }, { "paper_id": "FoF5RaA3ug", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Tao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xinyi Shang", "gender": "Female", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "peng sun", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 6, "n_ref_uni": 44, "n_ref": 151, "n_ref_all": 204, "n_fig": 11, "n_tab": 16, "L_tab": 9081, "n_element_tab": 787, "n_fig_1": 4, "n_tab_1": 11, "L_tab_1": 6399, "n_element_tab_1": 589, "formula_len_all": 2064, "formula_len_all_1": 336, "len_all": 184679, "len_all_1": 80664, "len_abs": 1927, "len_title": 127, "len_sents": 54675, "len_sents_1": 31420, "n_sents": 457, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1330, "title": "GIFT: Unlocking Full Potential of Labels in Distilled Dataset at Near-zero Cost", "abs": "Recent advancements in dataset distillation have demonstrated the significant benefits of employing soft labels generated by pre-trained teacher models.\n In this paper, we introduce a novel perspective by emphasizing the full utilization of labels.\n We first conduct a comprehensive comparison of various loss functions for soft label utilization in dataset distillation, revealing that the model trained on the synthetic dataset exhibits high sensitivity to the choice of loss function for soft label utilization.\n This finding highlights the necessity of a universal loss function for training models on synthetic datasets.\n Building on these insights, we introduce an extremely simple yet surprisingly effective plug-and-play approach, GIFT, which encompasses soft label refinement and a cosine similarity-based loss function to efficiently leverage full label information. \n Extensive experiments indicate that GIFT consistently enhances state-of-the-art dataset distillation methods across various dataset scales without incurring additional computational costs.\n Importantly, GIFT significantly enhances cross-optimizer generalization, an area previously overlooked.\n For instance, on ImageNet-1K with IPC = 10, GIFT enhances the state-of-the-art method RDED by 30.8\\% in cross-optimizer generalization.", "keywords": [ "Dataset Distillation", "Soft Label" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "FnlMYQPIzh", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Dung Vo", "gender": "unknown", "institution": "Singapore Management University", "country": "SG", "position": "Research Engineer" }, { "name": "Jinggui Liang", "gender": "Male", "institution": "Singapore Management University", "country": "", "position": "PhD student" }, { "name": "Lizi Liao", "gender": "Female", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 74, "n_ref": 125, "n_ref_all": 141, "n_fig": 2, "n_tab": 5, "L_tab": 1173, "n_element_tab": 165, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1686, "n_element_tab_1": 137, "formula_len_all": 390, "formula_len_all_1": 390, "len_all": 204034, "len_all_1": 71639, "len_abs": 1626, "len_title": 128, "len_sents": 37570, "len_sents_1": 33436, "n_sents": 219, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1634, "title": "ConvINT: A Semi-Structured Intention Framework for Conversational Understanding", "abs": "Understanding user intentions is critical for conversational AI, especially with the rise of large language models (LLMs) that demand a more nuanced comprehension of dialogue. Existing approaches, relying on rigid slot-value structures or unstructured representations, often miss the complexity of human intentions. In this work, we propose ConvINT, a novel semi-structured intention framework that offers a more holistic and fine-grained understanding of user intentions by organizing them into four key aspects: situation, emotion, action, and knowledge. Grounded in psychological and cognitive intention theories, ConvINT provides LLMs with a richer context for understanding user inputs while offering a semi-structured format that seamlessly integrates with prompt-based intention learning. To enable the efficient adoption of this framework, we introduce a Weakly-supervised Reinforced Generation (WeRG) method that scales ConvINT annotations across large datasets with high quality. By combining a small set of human-annotated instances with coarsely labeled data as weak supervision signals, WeRG effectively learns to generate ConvINT annotations, ensuring both scalability and precision. Experimental results demonstrate that integrating ConvINT with WeRG markedly improves LLMs\u2019 ability to comprehend user intentions, yielding significant gains in downstream tasks such as response generation and task completion, as validated by both automatic metrics and human evaluations. These findings highlight ConvINT's potential as a comprehensive and adaptable framework for advancing intention understanding in conversational AI.", "keywords": [ "Conversational Understanding", "Weakly-supervised Generation", "Large Language Models", "Fine-tuning" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "FneYHZU19U", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "He Sun", "gender": "Male", "institution": "University of Edinburgh", "country": "GB", "position": "Associate Professor" }, { "name": "John Stewart Fabila Carrasco", "gender": "Male", "institution": "University of Edinburgh", "country": "GB", "position": "Postdoc" } ], "n_formula": 93, "n_formula_1": 58, "n_ref_uni": 20, "n_ref": 27, "n_ref_all": 39, "n_fig": 7, "n_tab": 1, "L_tab": 139, "n_element_tab": 9, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 78, "n_element_tab_1": 3, "formula_len_all": 8192, "formula_len_all_1": 4176, "len_all": 109662, "len_all_1": 60525, "len_abs": 794, "len_title": 99, "len_sents": 35501, "len_sents_1": 24421, "n_sents": 336, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 827, "title": "Constrained Graph Clustering with Signed Laplacians", "abs": "Given two weighted graphs $G = (V, E, w_G)$ and $H = (V, F, w_H)$ defined on the same vertex set, the constrained clustering problem asks to find a set $S\\subset V$ that minimises the cut ratio between $w_G(S, V\\setminus S)$ and $w_H(S, V\\setminus S)$. We develop a Cheeger-type inequality that relates the solution of the constrained clustering problem to the spectral properties of $G$ and $H$. To reduce computational complexity, we use the signed Laplacian on $H$, simplifying the calculations while maintaining accurate results. By solving a generalized eigenvalue problem, our algorithm provides improvements in performance, particularly in scenarios where traditional spectral clustering methods face difficulties. We demonstrate its practical effectiveness through experiments on both synthetic and real-world datasets.", "keywords": [ "constrained graph clustering", "spectral graph theory" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 1, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "FnIRtzK5wX", "primary_area": "interpretability and explainable AI", "n_author": 10, "author_info_list": [ { "name": "Edouardo Honig", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Hao Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "", "position": "MS student" }, { "name": "Huijie Tang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Undergrad student" }, { "name": "Huilin Zhou", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Liang Lin", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "US", "position": "Full Professor" }, { "name": "Lianghui Ding", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Mingjie Li", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "", "position": "MS student" }, { "name": "Quanshi Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "", "position": "Associate Professor" }, { "name": "Zhen Yu Liu", "gender": "unknown", "institution": "Chongqing University of Communications", "country": "", "position": "Undergrad student" }, { "name": "Yingnian Wu", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Full Professor" } ], "n_formula": 28, "n_formula_1": 12, "n_ref_uni": 23, "n_ref": 41, "n_ref_all": 99, "n_fig": 11, "n_tab": 2, "L_tab": 517, "n_element_tab": 9, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 257, "n_element_tab_1": 3, "formula_len_all": 4230, "formula_len_all_1": 1246, "len_all": 168092, "len_all_1": 79846, "len_abs": 910, "len_title": 108, "len_sents": 62754, "len_sents_1": 36435, "n_sents": 507, "n_sents_1": 290, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 862, "title": "Disentangling the QiGan Encoded by a DNN towards the Go Game", "abs": "Given a deep neural network (DNN) that has surpassed human beings in a task, disentangling the explicit knowledge encoded by the DNN to obtain some new insights into the task is a new promising-yet-challenging regime in explainable AI. In this paper, we aim to disentangle the ''QiGan'' encoded by the AI model for the Go game, which has beat top human players. Specifically, we disentangle primitive shape patterns of stones memorized by the value network, and these shape patterns represent the ''QiGan'' used to conduct a fast situation assessment of the current board state. The universal-matching property of interactions ensure that human players can learn accurate and verifiable shape patterns, rather than specious intuitive analysis. In experiments, our method explains lots of novel shape patterns beyond traditional shape patterns in human knowledge.", "keywords": [ "Katago", "Knowledge discovery" ], "rating_list": [ 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 4 ], "presentation_list": [ 1, 2, 1, 4 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "FmoInsWCkp", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Pietro Lio", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "Full Professor" }, { "name": "Santanu Rathod", "gender": "unknown", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "PhD student" }, { "name": "Xiao Zhang", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Tenure-Track Faculty" } ], "n_formula": 30, "n_formula_1": 29, "n_ref_uni": 32, "n_ref": 55, "n_ref_all": 75, "n_fig": 15, "n_tab": 5, "L_tab": 1808, "n_element_tab": 274, "n_fig_1": 15, "n_tab_1": 2, "L_tab_1": 137, "n_element_tab_1": 20, "formula_len_all": 16265, "formula_len_all_1": 12298, "len_all": 160089, "len_all_1": 88587, "len_abs": 1171, "len_title": 147, "len_sents": 47057, "len_sents_1": 37255, "n_sents": 334, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1176, "title": "Predicting Time-Varying Flux and Balance in Metabolic Systems using Structured Neural ODE Processes", "abs": "We develop a novel data-driven framework as an alternative to dynamic flux balance analysis, bypassing the demand for deep domain knowledge and manual efforts to formulate the optimization problem. The proposed framework is end-to-end, which trains a structured neural ODE process (SNODEP) model to estimate flux and balance samples using gene-expression time-series data. SNODEP is designed to circumvent the limitations of the standard neural ODE process model, including restricting the latent and decoder sampling distributions to be normal and lacking structure between context points for calculating the latent, thus more suitable for modeling the underlying dynamics of a metabolic system. Through comprehensive experiments ($156$ in total), we demonstrate that SNODEP not only predicts the unseen time points of real-world gene-expression data and the flux and balance estimates well but can even generalize to more challenging unseen knockout configurations and irregular data sampling scenarios, all essential for metabolic pathway analysis. We hope our work can serve as a catalyst for building more scalable and powerful models for genome-scale metabolic analysis.", "keywords": [ "Neural ODEs", "Metabolic Networks" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 1, 1, 2 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "FkXYvV7nEB", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Dimitris Papailiopoulos", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "Associate Professor" }, { "name": "Jaewoong Cho", "gender": "unknown", "institution": "KRAFTON", "country": "", "position": "Researcher" }, { "name": "Jongho Park", "gender": "Male", "institution": "Krafton Inc.", "country": "KR", "position": "Researcher" }, { "name": "Junhyuck Kim", "gender": "Male", "institution": "KRAFTON", "country": "KR", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 33, "n_ref": 54, "n_ref_all": 78, "n_fig": 8, "n_tab": 5, "L_tab": 2373, "n_element_tab": 434, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 375, "n_element_tab_1": 50, "formula_len_all": 1220, "formula_len_all_1": 442, "len_all": 134044, "len_all_1": 60656, "len_abs": 1235, "len_title": 112, "len_sents": 38977, "len_sents_1": 28101, "n_sents": 328, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1241, "title": "KV-Dict: Sparse KV Cache Compression with Universal Dictionaries", "abs": "Transformer has become the de facto architecture for Large Language Models (LLMs), yet its substantial memory required for long contexts make it costly to deploy. Managing the memory usage of the key-value (KV) cache during inference has become a pressing challenge, as the cache grows with both model size and input length, consuming significant GPU memory.\n\nWe introduce a novel post-training KV cache compression method using KV-Dict, a universal dictionary that can accurately decompose and reconstruct key-value states. Unlike traditional quantization methods, KV-dict leverages sparse dictionary learning, allowing for flexible memory usage with minimal performance loss through fine-grained controls of sparsity levels. Moreoever, we retain competitive performance in the low memory regimes that 2-bit compression struggles to offer.\n\nKV-Dict is remarkably universal, as it uses a small, input-agnostic dictionary that is shared across tasks and batches without scaling memory. This universality, combined with the ability to control sparsity for different memory requirements, offers a flexible and efficient solution to the KV cache bottleneck, maintaining strong performance on complex reasoning tasks, such as LongBench and GSM8k.", "keywords": [ "transformer", "kv cache", "compression", "quantization" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "Fk4Op9wpEp", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Heywon Yun", "gender": "Female", "institution": "Yonsei University", "country": "KR", "position": "MS student" }, { "name": "Homa Fashandi", "gender": "Female", "institution": "LG Electronics", "country": "CA", "position": "Staff AI Research Scientist" }, { "name": "Jeonghwan Lee", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "MS student" }, { "name": "Jimin Kim", "gender": "unknown", "institution": "Ewha Women's University", "country": "KP", "position": "PhD student" }, { "name": "Yixiao Wang", "gender": "Male", "institution": "LG Electronics", "country": "CA", "position": "Senior Research Scientist" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 37, "n_ref": 66, "n_ref_all": 79, "n_fig": 6, "n_tab": 3, "L_tab": 505, "n_element_tab": 51, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1645, "n_element_tab_1": 80, "formula_len_all": 44, "formula_len_all_1": 44, "len_all": 101869, "len_all_1": 49046, "len_abs": 1027, "len_title": 150, "len_sents": 28029, "len_sents_1": 22485, "n_sents": 206, "n_sents_1": 156, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1086, "title": "Improving Human Pose-Conditioned Generation: Fine-tuning ControlNet Models with Reinforcement Learning", "abs": "Advancements in diffusion-based text-to-image generation models have made it possible to create high-quality human images. However, generating humans in desired poses using text prompts alone remains challenging. Image-to-image generation methods utilizing additional image conditions can address this issue; however, they often struggle with generating images that accurately match conditioning images. This paper proposes a new fine-tuning framework for training ControlNet models with reinforcement learning by combining ControlNet and Denoising Diffusion Policy Optimization~(DDPO) to understand pose conditioning images better. We apply a novel reward function in the proposed framework for higher pose accuracy. We demonstrate that our method effectively improves human generation by enhancing pose accuracy and the correct generation of body parts without omissions or additions. In addition, we demonstrate that the effectiveness of using a more detailed pose dataset along with our proposed reward function that directly leverages keypoints, leads to improved training results.", "keywords": [ "Generative AI", "Reinforcement Learning", "Text to Image Generation", "Image to Image Generation", "Multi-modal learning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 3, 2, 1 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 1, 1, 2, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Fk3eod9aaD", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Attila Juhos", "gender": "unknown", "institution": "Max-Planck Institute for Intelligent Systems", "country": "DE", "position": "PhD student" }, { "name": "Evgenia Rusak", "gender": "Female", "institution": "University of Tuebingen", "country": "DE", "position": "PhD student" }, { "name": "Matthias Bethge", "gender": "Male", "institution": "University of Tuebingen", "country": "DE", "position": "Full Professor" }, { "name": "Prasanna Mayilvahanan", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Thadd\u00e4us Wiedemer", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Wieland Brendel", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen", "country": "DE", "position": "Principal Researcher" }, { "name": "Roland Simon Zimmermann", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 42, "n_ref": 102, "n_ref_all": 132, "n_fig": 17, "n_tab": 13, "L_tab": 6006, "n_element_tab": 548, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1077, "n_element_tab_1": 128, "formula_len_all": 111, "formula_len_all_1": 0, "len_all": 188890, "len_all_1": 58074, "len_abs": 1337, "len_title": 44, "len_sents": 54413, "len_sents_1": 27798, "n_sents": 417, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 44, "L_abs": 1348, "title": "In Search of Forgotten Domain Generalization", "abs": "Out-of-Domain (OOD) generalization is the ability of a model trained on one or more domains to generalize to unseen domains. In the ImageNet era of computer vision, evaluation sets for measuring a model's OOD performance were designed to be strictly OOD with respect to style. However, the emergence of foundation models and expansive web-scale datasets has obfuscated this evaluation process, as datasets cover a broad range of domains and risk test domain contamination. In search of the forgotten domain generalization, we create large-scale datasets subsampled from LAION---LAION-Natural and LAION-Rendition---that are strictly OOD to corresponding ImageNet and DomainNet test sets in terms of style. Training CLIP models on these datasets reveals that a significant portion of their performance is explained by in-domain examples. This indicates that the OOD generalization challenges from the ImageNet era still prevail and that training on web-scale data merely creates the illusion of OOD generalization. Furthermore, through a systematic exploration of combining natural and rendition datasets in varying proportions, we identify optimal mixing ratios for model generalization across these domains. Our datasets and results re-enable meaningful assessment of OOD robustness at scale---a crucial prerequisite for improving model robustness.", "keywords": [ "Out-of-Distribution Robustness", "OOD generalization", "Out-of-Domain Robustness", "Evaluation" ], "rating_list": [ 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Fjkree2o8N", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 20, "author_info_list": [ { "name": "Aayush Mishra", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Anqi Liu", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" }, { "name": "Brian Wingenroth", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Researcher" }, { "name": "Cihan Xiao", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "G. Caleb Alexander", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Hongfu Liu", "gender": "Male", "institution": "Brandeis University", "country": "US", "position": "Assistant Professor" }, { "name": "Huiyuan Yang", "gender": "unknown", "institution": "Missouri University of Science and Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Jason Kuen", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Researcher" }, { "name": "Jiuxiang Gu", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Kate Tasker", "gender": "unknown", "institution": "University of California, San Francisco", "country": "US", "position": "Researcher" }, { "name": "Kevin S. Hawkins", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Researcher" }, { "name": "Lichun Ma", "gender": "unknown", "institution": "National Institutes of Health", "country": "US", "position": "Assistant Professor" }, { "name": "Tong Sun", "gender": "Female", "institution": "Adobe Systems", "country": "US", "position": "Director, Document Intelligence Lab" }, { "name": "Wanrong Zhu", "gender": "Female", "institution": "Adobe Research", "country": "US", "position": "Researcher" }, { "name": "Xiangxi Shi", "gender": "Male", "institution": "Oregon State University", "country": "US", "position": "PhD student" }, { "name": "Xuan Shen", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" }, { "name": "Yi Zhou", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "MS student" }, { "name": "Yiwei Wang", "gender": "Male", "institution": "University of California, Merced", "country": "US", "position": "Assistant Professor" }, { "name": "Zichao Wang", "gender": "Not Specified", "institution": "Adobe Research", "country": "US", "position": "Research Scientist" }, { "name": "RUIYI ZHANG", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 62, "n_ref": 82, "n_ref_all": 99, "n_fig": 8, "n_tab": 2, "L_tab": 1340, "n_element_tab": 40, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 2264, "n_element_tab_1": 68, "formula_len_all": 223, "formula_len_all_1": 223, "len_all": 163318, "len_all_1": 76305, "len_abs": 1527, "len_title": 131, "len_sents": 37535, "len_sents_1": 34732, "n_sents": 263, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1535, "title": "OIDA-QA: A Multimodal Benchmark for Analyzing the Opioid Industry Document Archive", "abs": "The opioid crisis is a serious public health issue that requires innovative solutions for effective analysis and deeper understanding. \nDespite the vast amounts of data in the Opioid Industry Documents Archive (OIDA), the complexity, multimodal nature, and specialized characteristics of healthcare data necessitate more advanced methods and models tailored to specific data types and detailed annotations, ensuring the precision and professionalism in the analysis.\nIn this paper, we tackle this challenge by organizing the original dataset according to document attributes and constructing a benchmark with 400k training documents and 10k for testing. We extract extensive multimodal information from each document, including textual, visual, and layout information, to capture a wide range of features. Given the extracted dense information, we collect a comprehensive dataset comprising over 3 million question-answer pairs with the assistance of multiple AI models. \nWe further develop domain-specific Large Language Models (LLMs) and investigate the impact of multimodal data on task performance.\nOur benchmarking and model efforts strive to produce an AI assistant system which can efficiently process the dataset and extract valuable insights.\nPreliminary results indicate the improvements with our AI assistant in document information extraction and question-answering tasks, highlighting the effectiveness of proposed benchmark in addressing the opioid crisis.\nThe data and model will be made publicly available for research.", "keywords": [ "LLM" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "FjZcwQJX8D", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Darrick Lee", "gender": "unknown", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Chancellor's Fellow" }, { "name": "Hong Yan", "gender": "unknown", "institution": "City University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Wong Hiu Tung", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 61, "n_formula_1": 23, "n_ref_uni": 68, "n_ref": 124, "n_ref_all": 164, "n_fig": 11, "n_tab": 12, "L_tab": 1716, "n_element_tab": 145, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 584, "n_element_tab_1": 25, "formula_len_all": 4205, "formula_len_all_1": 1634, "len_all": 221862, "len_all_1": 65069, "len_abs": 1675, "len_title": 89, "len_sents": 57518, "len_sents_1": 28427, "n_sents": 564, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 1426, "title": "Towards Scalable Topological Regularizers", "abs": "Latent space matching, which consists of matching distributions of features in latent space, is a crucial component for tasks such as adversarial attacks and defenses, domain adaptation, and generative modelling.\n Metrics for probability measures, such as Wasserstein and maximum mean discrepancy, are commonly used to quantify the differences between such distributions.\n However, these are often costly to compute, or do not appropriately take the geometric and topological features of the distributions into consideration.\n Persistent homology is a tool from topological data analysis which quantifies the multi-scale topological structure of point clouds, and has recently been used as a topological regularizer in learning tasks.\n However, computation costs preclude larger scale computations, and discontinuities in the gradient lead to unstable training behavior such as in adversarial tasks. \n We propose the use of principal persistence measures, based on computing the persistent homology of a large number of small subsamples, as a topological regularizer.\n We provide a parallelized GPU implementation of this regularizer, and prove that gradients are continuous for smooth densities.\n Furthermore, we demonstrate the efficacy of this regularizer on shape matching, image generation, and semi-supervised learning tasks, opening the door towards a scalable regularizer for topological features.", "keywords": [ "topological data analysis", "persistent homology", "generative adversarial network", "latent space matching" ], "rating_list": [ 5, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "FjQOXenaXK", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Wenhong Zhu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xiaofeng Wang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yiming Wang", "gender": "unknown", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Rui Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 25, "n_ref": 51, "n_ref_all": 91, "n_fig": 12, "n_tab": 18, "L_tab": 5844, "n_element_tab": 387, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1978, "n_element_tab_1": 192, "formula_len_all": 61, "formula_len_all_1": 18, "len_all": 226474, "len_all_1": 73653, "len_abs": 420, "len_title": 111, "len_sents": 74170, "len_sents_1": 32661, "n_sents": 563, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 877, "title": "Do Large Language Models Truly Understand Geometric Structures?", "abs": "Geometric ability is a significant challenge for large language models (LLMs) due to the need for advanced spatial comprehension and abstract thinking. Existing datasets primarily evaluate LLMs on their final answers, but they cannot truly measure their true understanding of geometric structures, as LLMs can arrive at correct answers by coincidence. To fill this gap, we introduce the GeomRel dataset, designed to evaluate LLMs\u2019 understanding of geometric structures by isolating the core step of geometric relationship identification in problem-solving. Using this benchmark, we conduct thorough evaluations of diverse LLMs and identify key limitations in understanding geometric structures. We further propose the Geometry Chain-of-Thought (GeoCoT) method, which enhances LLMs\u2019 ability to identify geometric relationships, resulting in significant performance improvements.", "keywords": [ "Large Language Models", "Geometric Ability Evaluation", "Geometric Relationship Identification" ], "rating_list": [ 6, 8, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "Fj6Yv5rPRe", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Yijiang Pang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "PhD student" } ], "n_formula": 51, "n_formula_1": 32, "n_ref_uni": 33, "n_ref": 64, "n_ref_all": 65, "n_fig": 0, "n_tab": 4, "L_tab": 1212, "n_element_tab": 147, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 488, "n_element_tab_1": 8, "formula_len_all": 8840, "formula_len_all_1": 5176, "len_all": 106722, "len_all_1": 57622, "len_abs": 1402, "len_title": 128, "len_sents": 27702, "len_sents_1": 22257, "n_sents": 237, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1424, "title": "Online learning meets Adam: The Road of Interpretable Adaptive Optimizer Design", "abs": "This paper explores the theoretical foundations of Adam, a widely used adaptive optimizer. Building on recent developments in non-convex optimization and online learning, particularly the discounted-to-nonconvex conversion framework, we present two aspects of results: First, we introduce clip-free FTRL, a novel variant of the classical Follow-the-Regularized-Leader (FTRL) algorithm. Unlike scale-free FTRL and the recently proposed $\\beta$-FTRL, our clip-free variant eliminates the need for clipping operations, aligning more closely with Adam's practical implementation. This modification provides deeper theoretical insights into Adam's empirical success and aligns the theoretical framework with practical implementations. By incorporating a refined analysis, our second result establishes a theoretical guarantee for the Last Iterate Convergence (LIC) under the proposed discounts-to-nonconvex conversion algorithm in LIC, which differs from the previous guarantee that has convergence evenly distributed in all iterations. Additionally, we extend this result to provide the last iterate convergence guarantee for the popular $\\beta$-FTRL algorithm under the same framework. However, the derived last iterate convergence of $\\beta$-FTRL reveals a persistent fixed error, potentially suggesting either limitations in popular online learning methods or the need for additional assumptions about the objective function.", "keywords": [ "online learning", "non-smooth non-convex optimization", "online-to-nonconvex conversion framework", "optimizer", "Adam" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "FiyS0ecSm0", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 9, "author_info_list": [ { "name": "Fan Yang", "gender": "Male", "institution": "Research, Microsoft", "country": "", "position": "Senior Principal Researcher" }, { "name": "Kaiyu Yang", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Wen Tang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Xian Zhang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Xiaoxing Ma", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Xujie Si", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Assistant Professor" }, { "name": "Yuan Yao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Zenan Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Zhaoyu Li", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" } ], "n_formula": 35, "n_formula_1": 14, "n_ref_uni": 59, "n_ref": 100, "n_ref_all": 120, "n_fig": 7, "n_tab": 2, "L_tab": 1123, "n_element_tab": 71, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1123, "n_element_tab_1": 71, "formula_len_all": 2864, "formula_len_all_1": 1270, "len_all": 193404, "len_all_1": 71587, "len_abs": 1535, "len_title": 120, "len_sents": 55384, "len_sents_1": 34092, "n_sents": 458, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1378, "title": "Proving Olympiad Inequalities by Synergizing LLMs and Symbolic Reasoning", "abs": "Large language models (LLMs) can prove mathematical theorems formally by generating proof steps (\\textit{a.k.a.} tactics) within a proof system. However, the space of possible tactics is vast and complex, while the available training data for formal proofs is limited, posing a significant challenge to LLM-based tactic generation. To address this, we introduce a neuro-symbolic tactic generator that synergizes the mathematical intuition learned by LLMs with domain-specific insights encoded by symbolic methods. The key aspect of this integration is identifying which parts of mathematical reasoning are best suited to LLMs and which to symbolic methods. While the high-level idea of neuro-symbolic integration is broadly applicable to various mathematical problems, in this paper, we focus specifically on Olympiad inequalities (Figure~1). We analyze how humans solve these problems and distill the techniques into two types of tactics: (1) scaling, handled by symbolic methods, and (2) rewriting, handled by LLMs. In addition, we combine symbolic tools with LLMs to prune and rank the proof goals for efficient proof search. We evaluate our framework on 161 challenging inequalities from multiple mathematics competitions, achieving state-of-the-art performance and significantly outperforming existing LLM and symbolic approaches without requiring additional training data.", "keywords": [ "Neuro-symbolic theorem proving", "Olympiad inequalities", "Large language model", "Symbolic method" ], "rating_list": [ 6, 8, 5, 8 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 2, 4, 2, 4 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "FiGDhrt1JL", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Bahram Zonooz", "gender": "Male", "institution": "Eindhoven University of Technology", "country": "NL", "position": "Assistant Professor" }, { "name": "Elahe Arani", "gender": "Female", "institution": "Eindhoven University of technology ", "country": "NL", "position": "Assistant Professor" }, { "name": "Ibrahim Batuhan Akkaya", "gender": "Male", "institution": "Snap Inc.", "country": "NL", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 18, "n_ref_uni": 28, "n_ref": 37, "n_ref_all": 65, "n_fig": 9, "n_tab": 8, "L_tab": 5075, "n_element_tab": 241, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 2159, "n_element_tab_1": 84, "formula_len_all": 1653, "formula_len_all_1": 1487, "len_all": 123782, "len_all_1": 54898, "len_abs": 1137, "len_title": 146, "len_sents": 36931, "len_sents_1": 23623, "n_sents": 381, "n_sents_1": 171, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1151, "title": "Foveated Dynamic Transformer: Robust and Efficient Perception Inspired by the Human Visual System", "abs": "The human visual system (HVS) employs foveated sampling and eye movements to achieve efficient perception, conserving both metabolic energy and computational resources. Drawing inspiration from this efficiency, we introduce the $\\textit{Foveated Dynamic Vision Transformer (FDT)}$, a novel architecture that integrates these mechanisms into a vision transformer framework. Unlike existing models, the FDT uses a single-pass strategy, utilizing fixation and foveation modules to enhance computational efficiency and accuracy. The fixation module identifies fixation points to filter out irrelevant information, while the foveation module generates foveated embeddings with multi-scale information. Our findings show that the FDT achieves superior accuracy and computational efficiency, with a 34\\% reduction in multiply-accumulate operations. Additionally, the FDT exhibits robustness against various types of noise and adversarial attacks without specific training for these challenges. These attributes make the FDT a significant step forward in creating artificial neural networks that mirror the efficiency, robustness, and adaptability of the HVS.", "keywords": [ "Transformers", "Vision Transformers", "Human Visual System", "Foveation" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 3, 1, 1 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "FhhH14jso4", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 7, "author_info_list": [ { "name": "Hongzhi Wen", "gender": "Male", "institution": "Michigan State University", "country": "", "position": "PhD student" }, { "name": "Jiliang Tang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Full Professor" }, { "name": "Kai Guo", "gender": "unknown", "institution": "Jilin University", "country": "CN", "position": "PhD student" }, { "name": "Wei Jin", "gender": "unknown", "institution": "Emory University", "country": "US", "position": "Assistant Professor" }, { "name": "Yi Chang", "gender": "Male", "institution": "Jilin University, China", "country": "CN", "position": "Full Professor" }, { "name": "Zewen Liu", "gender": "Male", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Zhikai Chen", "gender": "Male", "institution": "Michigan State University", "country": "", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 73, "n_ref_all": 115, "n_fig": 8, "n_tab": 14, "L_tab": 7751, "n_element_tab": 501, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1375, "n_element_tab_1": 140, "formula_len_all": 156, "formula_len_all_1": 0, "len_all": 182726, "len_all_1": 59238, "len_abs": 1141, "len_title": 136, "len_sents": 49270, "len_sents_1": 27105, "n_sents": 394, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1146, "title": "Learning on Graphs with Large Language Models (LLMs): A Deep Dive into Model Robustness", "abs": "Large Language Models (LLMs) have demonstrated remarkable performance across various natural language processing tasks. Recently, several LLMs-based pipelines have been developed to enhance learning on graphs with text attributes, showcasing promising performance. However, graphs are well-known to be susceptible to adversarial attacks and it remains unclear whether LLMs exhibit robustness in learning on graphs. To address this gap, our work aims to explore the potential of LLMs in the context of adversarial attacks on graphs. Specifically, we investigate the robustness against graph structural and textual perturbations in terms of two dimensions: LLMs-as-Enhancers and LLMs-as-Predictors. Through extensive experiments, we find that, compared to shallow models, both LLMs-as-Enhancers and LLMs-as-Predictors offer superior robustness against structural and textual attacks. Based on these findings, we carried out additional analyses to investigate the underlying causes. Furthermore, we have made our benchmark library openly available to facilitate quick and fair evaluations, and to encourage ongoing innovative research in this field.", "keywords": [ "Large Language Models", "Learning on Graphs", "Robustness" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "FhTAG591Ve", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Aaron Courville", "gender": "unknown", "institution": "University of Montreal", "country": "", "position": "Assistant Professor" }, { "name": "Arian Hosseini", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "PhD student" }, { "name": "Michael Noukhovitch", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "PhD student" }, { "name": "Rishabh Agarwal", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Adjunct Professor" }, { "name": "Shengyi Huang", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Louis-Pascal A. C. Xhonneux", "gender": "unknown", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, University of Montreal", "country": "CA", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 2, "n_ref_uni": 59, "n_ref": 115, "n_ref_all": 151, "n_fig": 13, "n_tab": 8, "L_tab": 1693, "n_element_tab": 134, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 203, "n_element_tab_1": 4, "formula_len_all": 723, "formula_len_all_1": 156, "len_all": 228439, "len_all_1": 57616, "len_abs": 1387, "len_title": 126, "len_sents": 51789, "len_sents_1": 27379, "n_sents": 409, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 68, "L_abs": 1358, "title": "Faster, More Efficient RLHF through Off-Policy Asynchronous Learning", "abs": "The dominant paradigm for RLHF is *online* and *on-policy* RL: synchronously generating from the large language model (LLM) policy, labelling with a reward model, and learning using feedback on the LLM's own outputs. While performant, it is computationally inefficient. Inspired by classical deep RL literature, we propose separating generation and learning in RLHF. This enables asynchronous generation of new samples while simultaneously training on old samples, leading to faster training and more compute-optimal scaling. However, asynchronous training relies on an underexplored regime, online but *off-policy* RLHF: learning on samples from previous iterations of our model. To understand the challenges in this regime, we investigate a fundamental question: how much off-policyness can we tolerate for asynchronous training to speed up learning but maintain performance? Among several RLHF algorithms we tested, we find that online DPO is most robust to off-policy data, and robustness increases with the scale of the policy model. We show even further compute optimizations but demonstrate that they come at a performance cost, giving rise to a trade-off. Finally, we verify our design choices by training LLaMA 3.1 8B with RLHF on instruction following tasks 40\\% faster than a synchronous run while matching final performance measured with GPT-4o.", "keywords": [ "reinforcement learning from human feedback", "efficient llm finetuning", "off-policy RL" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "FhLOZF8Tsm", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Antti Honkela", "gender": "Male", "institution": "University of Helsinki", "country": "FI", "position": "Full Professor" }, { "name": "Joonas J\u00e4lk\u00f6", "gender": "Male", "institution": "University of Helsinki", "country": "", "position": "Postdoc" }, { "name": "Marlon Tobaben", "gender": "unknown", "institution": "University of Helsinki", "country": "FI", "position": "PhD student" }, { "name": "Niki Andreas Loppi", "gender": "unknown", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Sebastian Rodriguez Beltran", "gender": "Male", "institution": "University of Helsinki", "country": "FI", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 59, "n_ref_all": 89, "n_fig": 10, "n_tab": 11, "L_tab": 2745, "n_element_tab": 294, "n_fig_1": 8, "n_tab_1": 6, "L_tab_1": 2025, "n_element_tab_1": 132, "formula_len_all": 430, "formula_len_all_1": 0, "len_all": 149904, "len_all_1": 63927, "len_abs": 1310, "len_title": 134, "len_sents": 40654, "len_sents_1": 28954, "n_sents": 328, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1317, "title": "Towards Efficient and Scalable Implementation of Differentially Private Deep Learning", "abs": "Differentially private stochastic gradient descent (DP-SGD) is the standard algorithm for training machine learning models under differential privacy (DP). The most common DP-SGD privacy accountants rely on Poisson subsampling for ensuring the theoretical DP guarantees. Implementing computationally efficient DP-SGD with Poisson subsampling is not trivial, which leads to many implementations ignoring this requirement. We conduct a comprehensive empirical study to quantify the computational cost of training deep learning models under DP given the requirement of Poisson subsampling, by re-implementing efficient methods using Poisson subsampling and benchmarking them. We find that using the naive implementation DP-SGD with Opacus in PyTorch has between 2.6 and 8 times lower throughput of processed training examples per second than SGD. However, efficient gradient clipping implementations with e.g. Ghost Clipping can roughly halve this cost. We propose alternative computationally efficient ways of implementing DP-SGD with JAX that are using Poisson subsampling and achieve only around 1.2 times lower throughput than SGD based on PyTorch. We highlight important implementation considerations with JAX. Finally, we study the scaling behaviour using up to 80 GPUs and find that DP-SGD scales better than SGD.", "keywords": [ "differential privacy", "gradient based optimization", "computational efficiency", "distributed computing" ], "rating_list": [ 3, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 1, 4, 3, 1 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "FhBT596F1X", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Eike Eberhard", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "MS student" }, { "name": "Nicholas Gao", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "PhD student" }, { "name": "Stephan G\u00fcnnemann", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Professor" } ], "n_formula": 48, "n_formula_1": 30, "n_ref_uni": 48, "n_ref": 110, "n_ref_all": 131, "n_fig": 6, "n_tab": 9, "L_tab": 2494, "n_element_tab": 416, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1351, "n_element_tab_1": 189, "formula_len_all": 2969, "formula_len_all_1": 1865, "len_all": 191145, "len_all_1": 74773, "len_abs": 1361, "len_title": 107, "len_sents": 49477, "len_sents_1": 31114, "n_sents": 410, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 59, "L_abs": 1368, "title": "Learning Equivariant Non-Local Electron Density Functionals", "abs": "The accuracy of density functional theory hinges on the approximation of non-local contributions to the exchange-correlation (XC) functional. To date, machine-learned and human-designed approximations suffer from insufficient accuracy, limited scalability, or dependence on costly reference data. To address these issues, we introduce Equivariant Graph Exchange Correlation (EG-XC), a novel non-local XC functional based on equivariant graph neural networks (GNNs). Where previous works relied on semi-local functionals or fixed-size descriptors of the density, we compress the electron density into an SO(3)-equivariant nuclei-centered point cloud representation for efficient non-local atomic-range interactions. By applying an equivariant GNN on this point cloud, we capture molecular-range interactions in a scalable and accurate manner. To train EG-XC, we differentiate through a self-consistent field solver requiring only energy targets. In our empirical evaluation, we find EG-XC to accurately reconstruct `gold-standard' CCSD(T) energies on MD17. On out-of-distribution conformations of 3BPA, EG-XC reduces the relative MAE by 35% to 50%. Remarkably, EG-XC excels in data efficiency and molecular size extrapolation on QM9, matching force fields trained on 5 times more and larger molecules. On identical training sets, EG-XC yields on average 51% lower MAEs.", "keywords": [ "Density Functional Theory", "DFT", "Functional", "Exchange Correlation", "XC", "Equivariance", "Graph Neural Network", "Electron Density", "Kohn-Sham DFT" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "FgirWC5TJ6", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Lei Cheng", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "Undergrad student" }, { "name": "Lianghui Ding", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Quanshi Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "", "position": "Associate Professor" }, { "name": "Zhengting Chen", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 22, "n_formula_1": 20, "n_ref_uni": 26, "n_ref": 46, "n_ref_all": 63, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1635, "formula_len_all_1": 1427, "len_all": 114912, "len_all_1": 64163, "len_abs": 948, "len_title": 102, "len_sents": 39205, "len_sents_1": 30712, "n_sents": 319, "n_sents_1": 239, "venue": "ICLR2025", "decision": "withdraw", "L_title": 54, "L_abs": 951, "title": "Disentangling Regional Primitives for Image Generation", "abs": "This paper presents a method to explain the internal representation structure of a neural network for image generation. Specifically, our method disentangles primitive feature components from the intermediate-layer feature of the neural network, which ensures that each feature component is exclusively used to generate a specific set of image regions. In this way, the generation of the entire image can be considered as the superposition of different pre-encoded primitive regional patterns, each being generated by a feature component. We find that the feature component can be represented as an OR relationship between the demands for generating different image regions, which is encoded by the neural network. Therefore, we extend the Harsanyi interaction to represent such an OR interaction to disentangle the feature component. Experiments show a clear correspondence between each feature component and the generation of specific image regions.", "keywords": [ "Explainable AI", "Generative Adversarial Networks", "Interactions" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 1, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "Fg0eo2AkST", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Bin Xu", "gender": "Male", "institution": "Department of Computer Science, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Ji Qi", "gender": "Male", "institution": "Zhipu AI", "country": "CN", "position": "Intern" }, { "name": "Jie Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Juanzi Li", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Lei Hou", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ming Ding", "gender": "Male", "institution": "ZHIPU AI", "country": "CN", "position": "Principal Researcher" }, { "name": "Qingsong Lv", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Weihan Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Wenyi Hong", "gender": "Female", "institution": "Department of Computer Science and Technology, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yushi Bai", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yuxiao Dong", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 24, "n_formula_1": 9, "n_ref_uni": 38, "n_ref": 121, "n_ref_all": 140, "n_fig": 8, "n_tab": 6, "L_tab": 661, "n_element_tab": 69, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 2235, "n_element_tab_1": 44, "formula_len_all": 19751, "formula_len_all_1": 5457, "len_all": 233295, "len_all_1": 74584, "len_abs": 194, "len_title": 0, "len_sents": 74764, "len_sents_1": 37498, "n_sents": 327, "n_sents_1": 155, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1511, "title": "CogCoM: A Visual Language Model with Chain-of-Manipulations Reasoning", "abs": "Vision-Language Models (VLMs) have shown broad effectiveness due to extensive training that aligns visual inputs with corresponding language responses. However, this conclusive alignment training causes models to overlook essential visual reasoning, leading to failures in handling detailed visual tasks and producing unfaithful responses. Drawing inspiration from human cognition in solving visual problems (e.g., marking, zoom in), this paper introduces Chain of Manipulations, a mechanism that enables VLMs to tackle problems step-by-step with evidence. After training, models can solve various visual problems by eliciting intrinsic manipulations (e.g., grounding, zoom in) with results (e.g., boxes, image) actively without relying external tools, while also allowing users to trace error causes. In this paper, we study the comprehensive methodology that includes: (1) a flexible design of manipulations based on extensive analysis, (2) an efficient automated data generation pipeline, (3) a compatible VLM architecture capable of multi-turn, multi-image, and (4) a model training process for versatile capabilities. With the design, we also manually annotate **6K** high-quality samples for challenging graphical mathematical problems. Our trained model, CogCoM, equipped with this mechanism and 17B parameters, achieves SOTA performance across **9** benchmarks in **4** categories, demonstrating its effectiveness while maintaining interpretability. Our code, model, and data will be publicly available.", "keywords": [ "Multimodal Reasoning", "Vision-Language Models", "Datasets Synthesis", "Graphical Math Annotation" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 1, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Fg04yPK0BH", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Haiquan Qiu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "quanming yao", "gender": "Male", "institution": "Department of Electronic Engineering, Tsinghua University", "country": "", "position": "Assistant Professor" }, { "name": "An Bian", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Senior researcher " } ], "n_formula": 47, "n_formula_1": 20, "n_ref_uni": 49, "n_ref": 105, "n_ref_all": 141, "n_fig": 10, "n_tab": 12, "L_tab": 3708, "n_element_tab": 259, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1538, "n_element_tab_1": 72, "formula_len_all": 4197, "formula_len_all_1": 1857, "len_all": 189478, "len_all_1": 72827, "len_abs": 1044, "len_title": 109, "len_sents": 54290, "len_sents_1": 30250, "n_sents": 482, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1047, "title": "GUMP: Alleviating Oversquashing with Unitary Message Passing", "abs": "Message passing mechanism contributes to the success of GNNs in various applications, but also brings the oversquashing problem. Recent works combat oversquashing by improving the graph spectrums with rewiring techniques, disrupting the original graph connectivity, and having limited improvement on oversquashing in terms of oversquashing measure. Motivated by unitary RNN, we propose Graph Unitary Message Passing (GUMP) to alleviate oversquashing in GNNs by applying a unitary adjacency matrix for message passing. To design GUMP, a transformation is first proposed to equip general graphs with unitary adjacency matrices and keep their original graph connectivity. Then, the unitary adjacency matrix is obtained with a unitary projection algorithm, which is implemented by utilizing the intrinsic structure of the unitary adjacency matrix and allows GUMP to be permutation-equivariant. In experiments, GUMP is incorporated into various GNN architectures and the extensive results show the effectiveness of GUMP on various graph learning tasks.", "keywords": [ "Oversquashing", "Graph theory", "Deep learning" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "Ffuw2ryqpz", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Georgios Pavlakos", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Assistant Professor" }, { "name": "Hanwen Jiang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" }, { "name": "Qixing Huang", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" } ], "n_formula": 14, "n_formula_1": 11, "n_ref_uni": 88, "n_ref": 165, "n_ref_all": 193, "n_fig": 11, "n_tab": 5, "L_tab": 2287, "n_element_tab": 338, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1866, "n_element_tab_1": 237, "formula_len_all": 1118, "formula_len_all_1": 504, "len_all": 208022, "len_all_1": 69452, "len_abs": 1261, "len_title": 126, "len_sents": 43006, "len_sents_1": 28334, "n_sents": 405, "n_sents_1": 259, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1270, "title": "Real3D: Towards Scaling Up Large Reconstruction Models with Real-World Images", "abs": "The default strategy for training single-view Large Reconstruction Models (LRMs) follows the fully supervised route using large-scale datasets of synthetic 3D assets or multi-view captures. Although these resources simplify the training procedure, they are hard to scale up beyond the existing datasets and they are not necessarily representative of the real distribution of object shapes. To address these limitations, in this paper, we introduce Real3D, the first LRM system that can be trained using single-view real-world images. Real3D introduces a novel self-training framework that can benefit from both the existing synthetic data and diverse single-view real images. We propose two unsupervised losses that allow us to supervise LRMs at the pixel- and semantic-level, even for training examples without ground-truth 3D or novel views. To further improve performance and scale up the image data, we develop an automatic data curation approach to collect high-quality examples from in-the-wild images. Our experiments show that Real3D consistently outperforms prior work in four diverse evaluation settings that include real and synthetic data, as well as both in-domain and out-of-domain shapes. We will make our code, models and data available upon publication.", "keywords": [ "Large Reconstruction Model", "Single-view Reconstruction", "In-the-wild Data", "Model Self-Training" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "FftPnwBb1z", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Li Xie", "gender": "Female", "institution": "City University of New York", "country": "", "position": "Researcher" }, { "name": "Mohammadsadeq Mottaqi", "gender": "Male", "institution": "City University of New York", "country": "", "position": "PhD student" }, { "name": "Rahul Ohlan", "gender": "Male", "institution": "The Graduate Center, City University of New York", "country": "", "position": "PhD student" }, { "name": "Raswanth Murugan", "gender": "Male", "institution": "CUNY Hunter College", "country": "", "position": "Researcher" }, { "name": "Shuo Zhang", "gender": "Male", "institution": "Weill Cornell Medicine, Cornell University", "country": "US", "position": "Postdoc" }, { "name": "Lei Xie", "gender": "Not Specified", "institution": "City University of New York", "country": "", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 14, "n_ref": 29, "n_ref_all": 46, "n_fig": 9, "n_tab": 2, "L_tab": 265, "n_element_tab": 40, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 265, "n_element_tab_1": 40, "formula_len_all": 245, "formula_len_all_1": 245, "len_all": 74730, "len_all_1": 70547, "len_abs": 1359, "len_title": 123, "len_sents": 23590, "len_sents_1": 21873, "n_sents": 149, "n_sents_1": 145, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1365, "title": "MolGene-E: Inverse Molecular Design to Modulate Single Cell Transcriptomics", "abs": "Designing drugs that can restore a diseased cell to its healthy state is an emerging approach in systems pharmacology to address medical needs that conventional target-based drug discovery paradigms have failed to meet. Single-cell transcriptomics can comprehensively map the differences between diseased and healthy cellular states, making it a valuable technique for systems pharmacology. However, single-cell omics data is noisy, heterogeneous, scarce, and high-dimensional. As a result, no machine learning methods currently exist to use single-cell omics data to design new drug molecules. We have developed a new deep generative framework named MolGene-E that can tackle this challenge. MolGene-E combines two novel models: 1) a cross-modal model that can harmonize and denoise chemical-perturbed bulk and single-cell transcriptomics data, and 2) a contrastive learning-based generative model that can generate new molecules based on the transcriptomics data. MolGene-E consistently outperforms baseline methods in generating high-quality, hit-like molecules from gene expression profiles obtained from single-cell datasets and gene expressions induced by knocking out targets using CRISPR. This superior performance is demonstrated across diverse de novo molecule generation metrics, which makes MolGene-E a potentially powerful new tool for drug discovery.", "keywords": [ "Systems Pharmacology", "Drug Discovery", "Phenotypic Screening", "Generative AI", "CLIP" ], "rating_list": [ 8, 5, 3, 1, 3 ], "soundness_list": [ 3, 2, 2, 1, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 4, 3, 3, 1, 2 ], "confidence_list": [ 4, 4, 3, 4, 5 ] }, { "paper_id": "FfIognyBee", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Debing Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "William Luo", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Zhengyang Geng", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "colin zhang", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Researcher" } ], "n_formula": 37, "n_formula_1": 13, "n_ref_uni": 51, "n_ref": 125, "n_ref_all": 141, "n_fig": 5, "n_tab": 9, "L_tab": 7709, "n_element_tab": 829, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3454, "n_element_tab_1": 356, "formula_len_all": 4263, "formula_len_all_1": 1403, "len_all": 227777, "len_all_1": 62249, "len_abs": 1741, "len_title": 128, "len_sents": 49398, "len_sents_1": 25873, "n_sents": 421, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1696, "title": "Diff-Instruct*: Towards Human-Preferred One-step Text-to-image Generative Models", "abs": "In this paper, we introduce Diff-Instruct* (DI*), an image data-free approach for building one-step text-to-image generative models that align with human preference while maintaining the ability to generate highly realistic images. We frame human preference alignment as online reinforcement learning using human feedback (RLHF), where the goal is to maximize the reward function while regularizing the generator distribution to remain close to a reference diffusion process. Unlike traditional RLHF approaches, which rely on the KL divergence for regularization, we introduce a novel score-based divergence regularization, which leads to significantly better performances. Although the direct calculation of this divergence remains intractable, we demonstrate that we can efficiently compute its gradient by deriving an equivalent yet tractable loss function. Remarkably, with Stable Diffusion V1.5 as the reference diffusion model, DI* outperforms all previously leading models by a large margin. When using the 2.6B Stable Diffusion XL architecture, the DI* results in a solid human-preferred one-step model that is able to generate aesthetic images of $1024\\times 1024$ resolutions. When using the 0.6B PixelArt-\u03b1 model as the reference diffusion, DI* achieves a new record Aesthetic Score of 6.30 and an Image Reward of 1.31 with only a single generation step, almost doubling the scores of the rest of the models with similar sizes. It also achieves an HPSv2 score of 28.70, establishing a new state-of-the-art benchmark. We also observe that DI* can improve the layout and enrich the colors of generated images. Our best human-preferred one-step generator will be released with this paper.", "keywords": [ "one-step text-to-image generative model", "human preference alignment", "RLHF" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "FfHGAAoSVJ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Bin Liu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Fei Teng", "gender": "Female", "institution": "Southwest Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Lin Ma", "gender": "Male", "institution": "Southwest Jiaotong University", "country": "CN", "position": "Lecturer" }, { "name": "Tianrui Li", "gender": "Male", "institution": "Southwest Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaohong Liu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhaoyu Li", "gender": "Male", "institution": "Southwest Jiaotong University", "country": "", "position": "Postdoc" }, { "name": "Ziqiao Shang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 27, "n_formula_1": 28, "n_ref_uni": 31, "n_ref": 56, "n_ref_all": 63, "n_fig": 1, "n_tab": 4, "L_tab": 2292, "n_element_tab": 300, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 1027, "n_element_tab_1": 178, "formula_len_all": 1883, "formula_len_all_1": 1831, "len_all": 122540, "len_all_1": 53910, "len_abs": 1280, "len_title": 107, "len_sents": 30121, "len_sents_1": 22535, "n_sents": 210, "n_sents_1": 153, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1294, "title": "AttentionNCE: Contrastive Learning with Instance Attention", "abs": "Contrastive learning has found extensive applications in computer vision, natural language processing, and information retrieval, significantly advancing the frontier of self-supervised learning. However, the limited availability of labels poses challenges in contrastive learning, as the positive and negative samples can be noisy, adversely affecting model training. To address this, we introduce instance-wise attention into the variational lower bound of contrastive loss, and proposing the AttentionNCE loss accordingly. AttentioNCE incorporates two key components that enhance contrastive learning performance: First, it replaces instance-level contrast with attention-based sample prototype contrast, helping to mitigate noise disturbances. Second, it introduces a flexible hard sample mining mechanism, guiding the model to focus on high-quality, informative samples. Theoretically, we demonstrate that optimizing AttentionNCE is equivalent to optimizing the variational lower bound of contrastive loss, offering a worst-case guarantee for maximum likelihood estimation under noisy conditions. Empirically, we apply AttentionNCE to popular contrastive learning frameworks and validate its effectiveness. The code is released at: \n\\url{https://anonymous.4open.science/r/AttentioNCE-55EB}", "keywords": [ "Contrastive Learning", "Image Classification" ], "rating_list": [ 5, 6, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 3, 2 ], "contribution_list": [ 1, 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 3, 5 ] }, { "paper_id": "Feg9xrbFcn", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Hang Zhang", "gender": "unknown", "institution": "Nanjing University", "country": "", "position": "PhD student" }, { "name": "Kai Ming Ting", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Wei-Jie Xu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" } ], "n_formula": 23, "n_formula_1": 9, "n_ref_uni": 32, "n_ref": 102, "n_ref_all": 124, "n_fig": 7, "n_tab": 11, "L_tab": 4183, "n_element_tab": 724, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 897, "n_element_tab_1": 160, "formula_len_all": 2079, "formula_len_all_1": 911, "len_all": 124613, "len_all_1": 57552, "len_abs": 1197, "len_title": 118, "len_sents": 31967, "len_sents_1": 23224, "n_sents": 258, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1197, "title": "Is $k \\times k$ Matrix Eigendecomposition Sufficient for Spectral Clustering?", "abs": "Spectral clustering has been widely used in clustering tasks due to its effectiveness. However, its key step, eigendecomposition of an $n\\times n$ matrix, is computationally expensive for large-scale datasets. Recent works have proposed methods to reduce this complexity, such as Nystr\\\"om method approximation and landmark-based approaches. While these methods aim to maintain good clustering quality while performing eigendecomposition on smaller matrix. The minimum matrix size required for spectral decomposition in spectral clustering is $k\\times k$ (where $k$ is the number of clusters), as it needs to obtain $n\\times k$ k-dimensional spectral embedding features. However, no algorithm can achieve good clustering performance with only a $k\\times k$ matrix eigendecomposition currently. In this paper, we propose a novel distribution-based spectral clustering. Our method constructs an $n\\times k$ bipartite graph between n data points and k distributions, enabling the eigendecomposition of only a $k\\times k$ matrix while preserving clustering quality. We demonstrate that our approach can achieve efficient and effective spectral clustering through $k\\times k $matrix eigendecomposition.", "keywords": [ "Spectral clustering", "Kernel mean embedding", "Matrix eigendecomposition" ], "rating_list": [ 1, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "FdiCqh4av8", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Emiliano Quinones Yumbla", "gender": "unknown", "institution": "Universidad de Puerto Rico, Carolina", "country": "PR", "position": "Assistant Professor" }, { "name": "Jennie Si", "gender": "unknown", "institution": "Arizona State University", "country": "US", "position": "Full Professor" }, { "name": "Junmin Zhong", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Ruofan Wu", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Researcher" }, { "name": "Seyed Yousef Soltanian", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Wenlong Zhang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 41, "n_formula_1": 16, "n_ref_uni": 58, "n_ref": 87, "n_ref_all": 149, "n_fig": 7, "n_tab": 7, "L_tab": 1005, "n_element_tab": 101, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 616, "n_element_tab_1": 32, "formula_len_all": 1849, "formula_len_all_1": 389, "len_all": 204330, "len_all_1": 69979, "len_abs": 1295, "len_title": 160, "len_sents": 61088, "len_sents_1": 36462, "n_sents": 423, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 111, "L_abs": 1304, "title": "Soft Robot Assisted Human Normative Walking: Real Device Control via Reinforcement Learning Without a Simulator", "abs": "This study offers an innovative solution approach to soft robot-assisted human walking. The controller design of the soft robotic exosuit aims at assisting human normative walking with reduced human physical effort. Achieving such optimal interaction between the human and robot agents presents a key challenge to the robot control design due to a lack of robust model of the soft inflatable exosuit and its interaction dynamics with the human user. Moreover, to maximize user comfort, the robot assistance should be personalized to individual users. Toward this goal, we propose an offline to online based approach that is referred to as AIP, which stands for online Adaptation from an offline Imitating expert Policy. Our offline learning mimics human expert actions through real human walking demonstrations without robot assistance. The resulted policy is then used to initialize online reinforcement learning, the goal of which is to optimally personalize robot assistance. In addition to being fast and robust, our online actor-critic learning method also posseses important properties such as learning convergence, system stability, and solution optimality. We have successfully demonstrated our simple and robust solution framework for safe robot control on all four tested human participants.", "keywords": [ "reinforcement learning", "Real world application", "imitation learning", "no simulator" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "FbbusgKmSW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Adrian Hilton", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Professor" }, { "name": "Chenhao Zhang", "gender": "unknown", "institution": "University of Surrey", "country": "GB", "position": "Intern" }, { "name": "Haosen Yang", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "PhD student" }, { "name": "Li Zhang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Marco Volino", "gender": "Male", "institution": "University of Surrey", "country": "GB", "position": "Lecturer" }, { "name": "Wenqing wang", "gender": "Female", "institution": "University of Surrey", "country": "GB", "position": "PhD student" }, { "name": "Xiatian Zhu", "gender": "unknown", "institution": "University of Surrey", "country": "GB", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 37, "n_ref": 67, "n_ref_all": 85, "n_fig": 7, "n_tab": 5, "L_tab": 2116, "n_element_tab": 225, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1625, "n_element_tab_1": 197, "formula_len_all": 106, "formula_len_all_1": 73, "len_all": 109423, "len_all_1": 52761, "len_abs": 1726, "len_title": 109, "len_sents": 29965, "len_sents_1": 23944, "n_sents": 208, "n_sents_1": 166, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 1728, "title": "Improving Gaussian Splatting with Localized Points Management", "abs": "Point management is critical for optimizing 3D Gaussian Splatting models, as point initiation (e.g., via structure from motion) is often distributionally inappropriate. Typically, Adaptive Density Control (ADC) algorithm is adopted, leveraging view-averaged gradient magnitude thresholding for point densification, opacity thresholding for pruning, and regular all-points opacity reset. We reveal that this strategy is limited in tackling intricate/special image regions (e.g., transparent) due to inability of identifying all 3D zones requiring point densification, and lacking an appropriate mechanism to handle ill-conditioned points with negative impacts (occlusion due to false high opacity). To address these limitations, we propose a Localized Point Management (LPM) strategy, capable of identifying those error-contributing zones in greatest need for both point addition and geometry calibration. Zone identification is achieved by leveraging the underlying multiview geometry constraints, subject to image rendering errors. We apply point densification in the identified zones and then reset the opacity of the points in front of these regions, creating a new opportunity to correct poorly conditioned points. Serving as a versatile plugin, LPM can be seamlessly integrated into existing static 3D and dynamic 4D Gaussian Splatting models. Experimental evaluations validate the efficacy of our LPM in boosting a variety of existing 3D/4D models both quantitatively and qualitatively. Notably, LPM improves both static 3DGS and dynamic SpaceTimeGS to achieve state-of-the-art rendering quality while retaining real-time speeds, excelling on challenging datasets such as Tanks & Temples and the Neural 3D Video dataset.", "keywords": [ "3D Reconstruction; 3D Gaussian Splatting; Novel View Synthesis" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 1, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "FbZSZEIkEU", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "AJ Yeung", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "MS student" }, { "name": "David Jensen", "gender": "Male", "institution": "University of Massachusetts, Amherst", "country": "US", "position": "Full Professor" }, { "name": "Jatin Nainani", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "MS student" }, { "name": "Kartik Gupta", "gender": "unknown", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "MS student" }, { "name": "Sankaran Vaidyanathan", "gender": "unknown", "institution": "Department of Computer Science, University of Massachusetts at Amherst", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 8, "n_ref": 16, "n_ref_all": 40, "n_fig": 9, "n_tab": 2, "L_tab": 469, "n_element_tab": 47, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 133, "n_element_tab_1": 14, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 96381, "len_all_1": 58319, "len_abs": 1535, "len_title": 124, "len_sents": 44225, "len_sents_1": 30189, "n_sents": 305, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1547, "title": "Adaptive Circuit Behavior and Generalization in Mechanistic Interpretability", "abs": "Mechanistic interpretability aims to understand the inner workings of large neural networks by identifying *circuits*, or minimal subgraphs within the model that implement algorithms responsible for performing specific tasks. These circuits are typically discovered and analyzed using a narrowly defined prompt format. However, given the abilities of large language models (LLMs) to generalize across various prompt formats for the same task, it remains unclear how well these circuits generalize. For instance, it is unclear whether the model\u2019s generalization results from reusing the same circuit components, the components behaving differently, or the use of entirely different components. In this paper, we investigate the generality of the indirect object identification (IOI) circuit in GPT-2 small, which is well-studied and believed to implement a simple, interpretable algorithm. We evaluate its performance on prompt variants that challenge the assumptions of this algorithm. Our findings reveal that the circuit generalizes surprisingly well, *reusing* all of its components and mechanisms while only adding additional input edges. Notably, the circuit generalizes even to prompt variants where the original algorithm should fail; we discover a mechanism that explains this which we term *S2 Hacking*. Our findings indicate that circuits within LLMs may be more flexible and general than previously recognized, underscoring the importance of studying circuit generalization to better understand the broader capabilities of these models.", "keywords": [ "Mechanistic Interpretability", "Circuit Generalization", "Science of ML" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 4, 3, 2, 1 ], "presentation_list": [ 4, 3, 2, 1 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "FbQLFsBbTe", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Baixi Sun", "gender": "Male", "institution": "Indiana University", "country": "US", "position": "PhD student" }, { "name": "Dingwen Tao", "gender": "Male", "institution": "State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Fanjiang Ye", "gender": "unknown", "institution": "Indiana University", "country": "US", "position": "PhD student" }, { "name": "Ori Yonay", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Tianbao Yang", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "Associate Professor" }, { "name": "Xingyu Chen", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" }, { "name": "Xiyuan Wei", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 11, "n_ref_uni": 41, "n_ref": 105, "n_ref_all": 148, "n_fig": 11, "n_tab": 23, "L_tab": 11881, "n_element_tab": 651, "n_fig_1": 1, "n_tab_1": 8, "L_tab_1": 3955, "n_element_tab_1": 225, "formula_len_all": 5163, "formula_len_all_1": 1236, "len_all": 223600, "len_all_1": 70743, "len_abs": 1654, "len_title": 143, "len_sents": 50495, "len_sents_1": 29125, "n_sents": 408, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1675, "title": "FastCLIP: A Suite of Optimization Techniques to Accelerate CLIP Training with Limited Resources", "abs": "Existing studies of training state-of-the-art Contrastive Language-Image Pretraining (CLIP) models on large-scale data involve hundreds of or even thousands of GPUs due to the requirement of a large batch size. However, such a large amount of resources is not accessible to most people. While advanced compositional optimization techniques for optimizing global contrastive losses have been demonstrated effective for removing the requirement of a large batch size, their performance on large-scale data remains underexplored and not optimized. To bridge the gap, this paper explores several aspects of CLIP training with \\textit{limited resources} (e.g., up to tens of GPUs). First, we introduce FastCLIP, a general CLIP training framework built on advanced compositional optimization techniques while designed and optimized for the {\\bf distributed setting}. Our framework is equipped with an efficient gradient reduction strategy to reduce communication overhead. Second, to further boost training efficiency, we investigate three components of the framework from an optimization perspective: the schedule of the inner learning rate, the update rules of the temperature parameter and the model parameters, respectively. Experiments on different strategies for each component shed light on how to conduct CLIP training more efficiently. Finally, we evaluate the performance of FastCLIP and the state-of-the-art training baseline (OpenCLIP) on different compute scales up to 32 GPUs on 8 nodes, and three data scales ranging from 2.7 million, 9.1 million to 315 million image-text pairs to demonstrate the significant improvement of FastCLIP in the resource-limited setting.", "keywords": [ "Multimodal Learning", "Contrastive Learning" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "Fb93MfxX7T", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Cheng-Hao Tu", "gender": "Male", "institution": "Appier", "country": "TW", "position": "Researcher" }, { "name": "Hong-You Chen", "gender": "unknown", "institution": "Apple AI/ML", "country": "US", "position": "Researcher" }, { "name": "Li Zhang", "gender": "Male", "institution": "Google", "country": "US", "position": "Software engineer" }, { "name": "Ping Zhang", "gender": "Male", "institution": "The Ohio State University", "country": "US", "position": "PhD student" }, { "name": "Wei-Lun Chao", "gender": "Male", "institution": "Ohio State University", "country": "US", "position": "Assistant Professor" }, { "name": "Zheda Mai", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 62, "n_ref": 132, "n_ref_all": 148, "n_fig": 10, "n_tab": 5, "L_tab": 2730, "n_element_tab": 154, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 4788, "n_element_tab_1": 199, "formula_len_all": 440, "formula_len_all_1": 360, "len_all": 177488, "len_all_1": 79564, "len_abs": 1992, "len_title": 166, "len_sents": 42276, "len_sents_1": 35286, "n_sents": 309, "n_sents_1": 248, "venue": "ICLR2025", "decision": "withdraw", "L_title": 117, "L_abs": 2011, "title": "Lessons Learned from a Unifying Empirical Study of Parameter-Efficient Transfer Learning (PETL) in Visual Recognition", "abs": "Parameter-efficient transfer learning (PETL) has attracted significant attention lately, due to the increasing size of pre-trained models and the need to fine-tune them for superior downstream performance. This community-wide enthusiasm has sparked a plethora of approaches. Nevertheless, a systematic study to understand their performance and suitable application scenarios is lacking, leaving questions like ''when to apply PETL'' and ''which approach to use'' largely unanswered, especially in visual recognition. In this paper, we conduct a unifying empirical study of representative PETL approaches in the context of Vision Transformers (ViT). We systematically tune their hyper-parameters to fairly compare their accuracy on downstream tasks. Our study not only offers a valuable user guide but also unveils several new insights. First, if tuned carefully, different PETL approaches can obtain quite similar accuracy in the low-shot benchmark VTAB-1K. This includes simple approaches like fine-tuning the bias terms that were reported inferior. Second, though with similar accuracy, we find that PETL approaches make different mistakes and high-confidence predictions, likely due to their different inductive biases. Such an inconsistency (or complementariness) opens up the opportunity for ensemble methods, and we make preliminary attempts at this. Third, going beyond the commonly used low-shot tasks, we find that PETL is also useful in many-shot regimes --- it achieves comparable and sometimes better accuracy than full fine-tuning, using much fewer learnable parameters. Last but not least, we investigate PETL's ability to preserve a pre-trained model's robustness to distribution shifts (e.g., a CLIP backbone). Perhaps not surprisingly, PETL approaches outperform full fine-tuning alone. However, with weight-space ensembles, the fully fine-tuned model can better balance target (i.e., downstream) distribution and distribution shift performance, suggesting a future research direction for PETL.", "keywords": [ "Parameter-Efficient Transfer Learning", "Parameter-Efficient Fine-Tuning", "Visual Recognition" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "Fb0q2uI4Ha", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Bing Deng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Long Bai", "gender": "Not Specified", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Xing Xu", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Researcher" }, { "name": "Yixuan Zhou", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Hengtao Shen", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Sijia Cai", "gender": "Male", "institution": "Alibaba Cloud", "country": "CN", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 29, "n_ref": 72, "n_ref_all": 99, "n_fig": 8, "n_tab": 14, "L_tab": 6043, "n_element_tab": 702, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2000, "n_element_tab_1": 227, "formula_len_all": 48, "formula_len_all_1": 0, "len_all": 228737, "len_all_1": 65100, "len_abs": 2209, "len_title": 123, "len_sents": 81232, "len_sents_1": 30842, "n_sents": 587, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 75, "L_abs": 1412, "title": "TAU-106K: A New Dataset for Comprehensive Understanding of Traffic Accident", "abs": "Multimodal Large Language Models (MLLMs) have demonstrated impressive performance in general visual understanding tasks. However, their potential for high-level and fine-grained comprehension, such as humor or anomaly understanding, remains unexplored. Targeting traffic accidents, a critical and practical scenario within anomaly understanding, we explore the advanced capabilities of MLLMs and introduce TABot, a multimodal MLLM tailored for accident-related tasks. To facilitate this, we first develop TAU-106K, a large-scale multimodal dataset comprising 106K traffic accident-related videos and images, sourced from academic benchmarks and public platforms. The dataset is meticulously annotated through a video-to-image annotation pipeline, ensuring comprehensive and high-quality labels. Upon TAU-106K, our accident-oriented MLLM TABot is trained in a two-step approach to integrate multi-granularity accident understanding tasks, including accident recognition, spatial-temporal grounding, with an additional accident description task to guide the model in comprehending the nature of traffic accidents. Extensive experiments demonstrate the superior performance of TABot in traffic accident understanding, underscoring both its potential for high-level anomaly understanding and the robustness of the TAU-106K dataset. All datasets, annotations, and models will be publicly released for future research.", "keywords": [ "Anomaly Detection", "Traffic Accident Understanding", "Multimodal Large Language Models", "Instructional Dataset" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "FaOeBrlPst", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Aditya Emmanuel Arokiaraj John", "gender": "Male", "institution": "Duke University", "country": "US", "position": "MS student" }, { "name": "Brandon Fain", "gender": "unknown", "institution": "Duke University", "country": "", "position": "Assistant Professor" }, { "name": "Yiran Shen", "gender": "unknown", "institution": "Duke University", "country": "US", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 26, "n_ref": 44, "n_ref_all": 57, "n_fig": 8, "n_tab": 1, "L_tab": 189, "n_element_tab": 9, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 206, "n_element_tab_1": 12, "formula_len_all": 202, "formula_len_all_1": 199, "len_all": 139841, "len_all_1": 61029, "len_abs": 1127, "len_title": 48, "len_sents": 41601, "len_sents_1": 30714, "n_sents": 314, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 48, "L_abs": 1132, "title": "Explainable Rewards in RLHF Using LLM-as-a-Judge", "abs": "Reinforcement Learning from Human Feedback (RLHF) has been gaining popularity as a method for aligning Large Language Models (LLMs) with human preferences. It involves performing Supervised Fine-Tuning (SFT) followed by fine-tuning using a reward model trained on human preference data. However, two primary issues with this approach are the difficult and expensive curation of human preference data and the opaque, black-box nature of the rewards. To address these issues, this paper introduces a novel framework for aligning LLMs with human preferences. Our framework involves using representative sub-dimensions for specific tasks to generate rewards by leveraging a performant out-of-the-box LLM. We evaluate our approach by fine-tuning two models, one using our approach and one using traditional black-box rewards. Evaluation using an advanced LLM-based method demonstrates that our approach maintains the performance of the black-box baseline while offering superior explainability and flexibility. This framework not only enhances transparency in RLHF but also eliminates reliance on expensive human-curated preference data.", "keywords": [ "Large Language Models; Reinforcement Learning from Human Feedback; Explainability" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 1 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "FaL6aTuXod", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Ammar Gilani", "gender": "Male", "institution": "Northwestern University", "country": "", "position": "PhD student" }, { "name": "Han Liu", "gender": "unknown", "institution": "Northwestern University", "country": "", "position": "Associate Professor" } ], "n_formula": 20, "n_formula_1": 6, "n_ref_uni": 34, "n_ref": 63, "n_ref_all": 71, "n_fig": 3, "n_tab": 7, "L_tab": 2130, "n_element_tab": 181, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 2012, "n_element_tab_1": 186, "formula_len_all": 1253, "formula_len_all_1": 380, "len_all": 135738, "len_all_1": 59078, "len_abs": 1127, "len_title": 131, "len_sents": 40441, "len_sents_1": 23875, "n_sents": 406, "n_sents_1": 268, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1246, "title": "A Comprehensive Framework for Benchmarking Algorithms Across Hyperparameter Spaces", "abs": "We introduce a framework for benchmarking algorithms with varying hyperparameters from multiple perspectives. The dependency of algorithms' performance on hyperparameters complicates fair comparisons and often leads to inconsistent empirical studies. Our framework addresses this challenge by proposing two key criteria: \\textit{Performance-HPO} trajectory and \\textit{Reliability-HPO}.\nThe Performance-HPO trajectory tracks how an algorithm\u2019s performance changes with different hyperparameter optimization (HPO) budget allocations, leveraging a variety of off-the-shelf hyperparameter optimizers. This enables users to identify the most suitable algorithm for their specific needs. The Reliability-HPO criterion evaluates the expected value of an algorithm's success rate across hyperparameters, estimated using Monte Carlo simulations in log-space.\nWe demonstrate our framework by benchmarking widely-used convex optimizers. Our experiments, conducted with {\\footnotesize\\texttt{CVXPY}} across various problem types, settings, and dimensionalities, reveal that the {\\footnotesize\\texttt{SCS}} solver exhibits the highest Performance-HPO, while {\\footnotesize\\texttt{ECOS}} and {\\footnotesize\\texttt{MOSEK}} demonstrate superior Reliability-HPO.", "keywords": [ "Benchmarking", "Hyperparameter optimization" ], "rating_list": [ 1, 1, 1, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 1, 3, 1, 2 ], "contribution_list": [ 1, 2, 1, 2 ], "confidence_list": [ 4, 2, 5, 2 ] }, { "paper_id": "FZv3kPHTtB", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Lina Yao", "gender": "Female", "institution": "University of New South Wales", "country": "AU", "position": "Full Professor" }, { "name": "Linjie Yang", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Mingfei Han", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Researcher" }, { "name": "heng Wang", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "research scientist" }, { "name": "Xiaojun Chang", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 26, "n_ref": 79, "n_ref_all": 124, "n_fig": 21, "n_tab": 4, "L_tab": 880, "n_element_tab": 135, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 2920, "n_element_tab_1": 133, "formula_len_all": 9, "formula_len_all_1": 0, "len_all": 209282, "len_all_1": 77864, "len_abs": 1495, "len_title": 129, "len_sents": 80088, "len_sents_1": 35316, "n_sents": 725, "n_sents_1": 298, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1021, "title": "Shot2Story: A New Benchmark for Comprehensive Understanding of Multi-shot Videos", "abs": "A short clip of video may contain progression of multiple events and an interesting story line. A human need to capture both the event in every shot and associate them together to understand the story behind it. In this work, we present a new multi-shot video understanding benchmark \\dataset with detailed shot-level captions, comprehensive video summaries and question-answering pairs. To facilitate better semantic understanding of videos, we provide captions for both visual signals and human narrations. We design several distinct tasks including single-shot video captioning, multi-shot video summarization, and multi-shot video question answering. Preliminary experiments show some challenges to generate a long and comprehensive video summary for multi-shot videos. Nevertheless, the generated imperfect summaries can already achieve competitive performance on existing video understanding tasks such as video question-answering, promoting an under-explored setting of video understanding with detailed summaries.", "keywords": [ "vision language model", "video question answering", "video captioning", "multi-shot videos" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 2, 4, 4, 4 ] }, { "paper_id": "FZaw83yo76", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Eunseo Koh", "gender": "unknown", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "MS student" }, { "name": "Jiwoo Chung", "gender": "Male", "institution": "Sungkyunkwan University", "country": "", "position": "MS student" }, { "name": "MinKyu Lee", "gender": "Male", "institution": "Sungkyunkwan University", "country": "KR", "position": "PhD student" }, { "name": "Sangeek Hyun", "gender": "Male", "institution": "Sungkyunkwan University", "country": "KR", "position": "PhD student" }, { "name": "Jae-pil Heo", "gender": "Male", "institution": "Sungkyunkwan University", "country": "KR", "position": "Associate Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 29, "n_ref": 59, "n_ref_all": 93, "n_fig": 20, "n_tab": 4, "L_tab": 697, "n_element_tab": 121, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 777, "n_element_tab_1": 1, "formula_len_all": 181, "formula_len_all_1": 181, "len_all": 117108, "len_all_1": 63869, "len_abs": 758, "len_title": 117, "len_sents": 35796, "len_sents_1": 30753, "n_sents": 268, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1224, "title": "Structure-Preserving Text-Based Editing for Few-Step Diffusion Models", "abs": "Text-based image editing aims to generate an image that corresponds to the given text prompt, but with the structure of the original source image. Existing methods often rely on attention maps in diffusion models (DMs) for structure preservation, as these features are considered to play a primary role in determining the spatial layout. However, we find that these methods struggle to preserve the spatial layout when applied to few-step DMs (e.g., SDXL-Turbo), limiting their use cases to the slower multi-step DMs (e.g., Stable Diffusion). In this work, we investigate the limitations of these approaches in terms of intermediate feature representations. Our findings indicate that for few-step DMs, the attention layers have less influence in determining the structure. To tackle this, we localize layers within the network that better control spatial layout and inject these features during the editing process. Additionally, we disentangle structural information from other features to avoid conflicts between the injected features and the text prompt. This ensures that the edited image faithfully follows the prompt while preserving the source structure. Our method outperforms existing text-based editing baselines.", "keywords": [ "Text-based image editing" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "FZa1UCC9SC", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Atish Agarwala", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Elliot Paquette", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Associate Professor" }, { "name": "Ke Liang Xiao", "gender": "Male", "institution": "McGill University, McGill University", "country": "CA", "position": "PhD student" }, { "name": "Noah Marshall", "gender": "unknown", "institution": "McGill University, McGill University", "country": "CA", "position": "PhD student" } ], "n_formula": 335, "n_formula_1": 43, "n_ref_uni": 25, "n_ref": 53, "n_ref_all": 109, "n_fig": 9, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 539, "n_element_tab_1": 92, "formula_len_all": 24440, "formula_len_all_1": 2303, "len_all": 231913, "len_all_1": 61602, "len_abs": 1247, "len_title": 154, "len_sents": 64536, "len_sents_1": 25178, "n_sents": 756, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 867, "title": "Exact risk curves of signSGD in High-Dimensions: quantifying preconditioning and noise-compression effects", "abs": "In recent years, SignSGD has garnered interest as both a practical optimizer as well as a simple model to understand adaptive optimizers like Adam. Though there is a general consensus that SignSGD acts to precondition optimization and reshapes noise, quantitatively understanding these effects in theoretically solvable settings remains difficult. We present an analysis of SignSGD in a high dimensional limit, and derive a limiting SDE and ODE to describe the risk. Using this framework we quantify four effects of SignSGD: effective learning rate, noise compression, diagonal preconditioning, and gradient noise reshaping. Our analysis is consistent with experimental observations but moves beyond that by quantifying the dependence of these effects on the data and noise distributions. We conclude with a conjecture on how these results might be extended to Adam.", "keywords": [ "signSGD", "stochastic optimization", "Deep learning theory", "high-dimensional probability", "stochastic differential equation" ], "rating_list": [ 6, 3, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 5, 4, 3 ] }, { "paper_id": "FZS5m1cbFU", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Pan Peng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Ruobing Bai", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 62, "n_formula_1": 21, "n_ref_uni": 30, "n_ref": 62, "n_ref_all": 71, "n_fig": 3, "n_tab": 2, "L_tab": 365, "n_element_tab": 23, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 490, "n_element_tab_1": 27, "formula_len_all": 5572, "formula_len_all_1": 1953, "len_all": 179079, "len_all_1": 74849, "len_abs": 1237, "len_title": 94, "len_sents": 61312, "len_sents_1": 34078, "n_sents": 587, "n_sents_1": 285, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1284, "title": "Differentially Private Range Subgraph Counting", "abs": "Subgraph counting is a fundamental problem in graph analysis. Motivated by the practical need to perform graph analytics on subgraphs defined by selected vertices (or edges) rather than the entire graph, as well as privacy concerns, we initiate the study of private range subgraph counting. Given an $n$-vertex graph $G$, where each vertex (or edge) has a $d$-dimensional attribute vector, a pattern graph $H$, and a set $Q$ of range queries $q$, our goal is to count the occurrences of $H$ in the subgraph of $G$ induced by vertices (or edges) whose attributes fall within $q$, all while preserving privacy. We give the first $\\varepsilon$-differentially private algorithm for range subgraph counting, achieving near-optimal accuracy (up to a polylogarithmic factor of $n$) for constant privacy parameter $\\varepsilon$ and dimension $d$, with no additional computational overhead compared to non-private algorithms. \nWe also demonstrate that by relaxing to $(\\varepsilon, \\delta)$-DP, we can achieve smaller additive errors. Furthermore, our results generalize the subgraph counting results of the partially dynamic model in [FHO21]. Empirical evaluations demonstrate that our algorithm significantly outperforms baseline methods in accuracy while ensuring strong privacy guarantees.", "keywords": [ "Differential privacy", "subgraph", "range query" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 3, 1 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "FYvZCwdb6F", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Arif Ahmad", "gender": "Male", "institution": "Indian Institute of Technology Bombay, Indian Institute of Technology, Bombay", "country": "", "position": "Undergrad student" }, { "name": "Nihar Ranjan Sahoo", "gender": "Male", "institution": "Indian Institute of Technology, Bombay", "country": "", "position": "PhD student" }, { "name": "Nishtha Madaan", "gender": "Female", "institution": "Indian Institute of Technology Delhi", "country": "", "position": "PhD student" }, { "name": "Pushpak Bhattacharyya", "gender": "Male", "institution": "Indian Institute of Technology, Bombay, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 23, "n_ref": 32, "n_ref_all": 53, "n_fig": 2, "n_tab": 9, "L_tab": 3526, "n_element_tab": 348, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 815, "n_element_tab_1": 132, "formula_len_all": 431, "formula_len_all_1": 414, "len_all": 111011, "len_all_1": 60254, "len_abs": 1438, "len_title": 129, "len_sents": 35006, "len_sents_1": 27377, "n_sents": 256, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1414, "title": "MOMENTUM MEETS VIRALITY: A NOVEL METRIC FOR UNMASKING SOCIAL BIAS IN VIRAL TWEETS", "abs": "Predicting which social media posts will go viral is a critical but complex task in the field of computational social science. Previous studies have utilized various measures to forecast the virality of tweets or Facebook posts, but these approaches exhibit limitations, particularly in the absence of a virality metric that specifically considers social biases. In this paper, we test existing metrics and introduce a new metric, $\\textbf{ViralTweet Score (VTS)}$, inspired by principles of momentum from physics to better predict a tweet's virality given that it consists of social biases. We compare this new metric with others, highlighting the advantages and disadvantages of each of them as a virality measurement metric. We release the $\\textbf{ViralTweets Dataset}$ with $\\mathbf{88.8k}$ Hindi tweets and corresponding virality labels based on our VTS metric. We also show how social biases in posts can influence their potential to go viral. We test our hypothesis that VTS is a better metric using two methodologies and we show how VTS achieves an F1 score of 0.87 based on pairwise evaluation methodology and an overall F1 score of 0.58 based on our clustering-based verification methodology. Our work offers a novel metric for understanding tweet virality for biased tweets and opens the door for more equitable and effective social media analytics by considering the role of social biases in virality.", "keywords": [ "Social bias", "Tweet virality", "ViralTweetScore", "Hindi Tweets", "Tweets" ], "rating_list": [ 3, 1, 5, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "FXw0okNcOb", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Anji Liu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Mathias Niepert", "gender": "Male", "institution": "Universit\u00e4t Stuttgart", "country": "", "position": "Full Professor" }, { "name": "Oliver Broadrick", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Guy Van den Broek", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" } ], "n_formula": 59, "n_formula_1": 20, "n_ref_uni": 39, "n_ref": 80, "n_ref_all": 128, "n_fig": 7, "n_tab": 2, "L_tab": 834, "n_element_tab": 48, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 995, "n_element_tab_1": 51, "formula_len_all": 5859, "formula_len_all_1": 1602, "len_all": 211603, "len_all_1": 79648, "len_abs": 1480, "len_title": 73, "len_sents": 67046, "len_sents_1": 31873, "n_sents": 533, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 25, "L_abs": 1487, "title": "Discrete Copula Diffusion", "abs": "Discrete diffusion models have recently shown significant progress in modeling complex data, such as natural languages and DNA sequences. However, unlike diffusion models for continuous data, which can generate high-quality samples in just a few denoising steps, modern discrete diffusion models still require hundreds or even thousands of denoising steps to perform well. In this paper, we identify a fundamental limitation that prevents discrete diffusion models from achieving strong performance with fewer steps -- they fail to capture dependencies between output variables at each denoising step. To address this issue, we provide a formal explanation and introduce a general approach to supplement the missing dependency information by incorporating another deep generative model, termed the copula model. Our method does not require fine-tuning either the diffusion model or the copula model, yet it enables high-quality sample generation with significantly fewer denoising steps. When we apply this approach to autoregressive copula models, the combined model outperforms both models individually in unconditional and conditional text generation. Specifically, the hybrid model achieves better (un)conditional text generation using 8 to 32 times fewer denoising steps than the diffusion model alone. In addition to presenting an effective discrete diffusion generation algorithm, this paper emphasizes the importance of modeling inter-variable dependencies in discrete diffusion.", "keywords": [ "discrete diffusion models", "copula" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "FXm7EEFRa8", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Pieter Abbeel", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Younggyo Seo", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 45, "n_ref": 133, "n_ref_all": 160, "n_fig": 10, "n_tab": 1, "L_tab": 378, "n_element_tab": 43, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 615, "formula_len_all_1": 585, "len_all": 140897, "len_all_1": 60628, "len_abs": 1356, "len_title": 125, "len_sents": 38956, "len_sents_1": 29995, "n_sents": 264, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1361, "title": "Reinforcement Learning with Action Sequence for Data-Efficient Robot Learning", "abs": "Training reinforcement learning (RL) agents on robotic tasks typically requires a large number of training samples. This is because training data often consists of noisy trajectories, whether from exploration or human-collected demonstrations, making it difficult to learn value functions that understand the effect of taking each action. On the other hand, recent behavior-cloning (BC) approaches have shown that predicting a sequence of actions enables policies to effectively approximate noisy, multi-modal distributions of expert demonstrations. Can we use a similar idea for improving RL on robotic tasks? In this paper, we introduce a novel RL algorithm that learns a critic network that outputs Q-values over a sequence of actions. By explicitly training the value functions to learn the consequence of executing a series of current and future actions, our algorithm allows for learning useful value functions from noisy trajectories. We study our algorithm across various setups with sparse and dense rewards, and with or without demonstrations, spanning mobile bi-manual manipulation, whole-body control, and tabletop manipulation tasks from BiGym, HumanoidBench, and RLBench. We find that, by learning the critic network with action sequences, our algorithm outperforms various RL and BC baselines, in particular on challenging humanoid control tasks.", "keywords": [ "Reinforcement Learning", "Robot Learning", "Robotics", "Data-Efficiency" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "FXhCG41FvV", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Cheng Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tianyu Xie", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 43, "n_formula_1": 25, "n_ref_uni": 58, "n_ref": 105, "n_ref_all": 140, "n_fig": 5, "n_tab": 5, "L_tab": 1887, "n_element_tab": 167, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1114, "n_element_tab_1": 97, "formula_len_all": 3023, "formula_len_all_1": 1483, "len_all": 186471, "len_all_1": 65768, "len_abs": 1208, "len_title": 135, "len_sents": 52756, "len_sents_1": 29803, "n_sents": 404, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1213, "title": "ARTreeFormer: A Faster Attention-based Autoregressive Model for Phylogenetic Inference", "abs": "Probabilistic modeling of the combinatorially explosive tree topology space has posed a significant challenge in phylogenetic inference. Previous approaches often necessitate pre-sampled tree topologies, limiting their modeling capability to a subset of the entire tree space. A recent advancement is ARTree, a deep autoregressive model that offers unrestricted distributions for tree topologies. However, the repetitive computations of topological node embeddings via Dirichlet energy minimization and the message passing over all the nodes can be expensive, which may hinder its application to data sets with many species. This paper proposes ARTreeFormer, a novel approach that harnesses attention mechanisms to accelerate ARTree. By introducing attention-based recurrent node embeddings, ARTreeFormer allows the reuse of node embeddings from preceding ordinal tree topologies and fast vectorized computation as well. This, together with a local message passing scheme, significantly improves the computation speed of ARTree while maintaining great approximation performance. We demonstrate the effectiveness and efficiency of our method on a benchmark of challenging real data phylogenetic inference problems.", "keywords": [ "phylogenetic inference", "autoregressive model", "attention mechanism" ], "rating_list": [ 6, 5, 5, 3, 8 ], "soundness_list": [ 3, 2, 3, 1, 4 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 2 ] }, { "paper_id": "FXPsZ6cbUj", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Canyu Chen", "gender": "unknown", "institution": "Illinois Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Chengxing Xie", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Guohao Li", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Hao Liu", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Songjia Liu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Yu Lei", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "MS student" }, { "name": "Zhen Wu", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Associate Professor" }, { "name": "Zhiyu Yin", "gender": "unknown", "institution": "Stevens Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 53, "n_ref": 66, "n_ref_all": 101, "n_fig": 13, "n_tab": 4, "L_tab": 584, "n_element_tab": 63, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1361, "n_element_tab_1": 55, "formula_len_all": 1392, "formula_len_all_1": 1389, "len_all": 153121, "len_all_1": 58307, "len_abs": 1451, "len_title": 151, "len_sents": 36932, "len_sents_1": 26560, "n_sents": 280, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1115, "title": "FAIRMINDSIM: ALIGNMENT OF BEHAVIOR, EMO- TION, AND BELIEF IN HUMANS AND LLM AGENTS AMID ETHICAL DILEMMAS", "abs": "AI alignment is a pivotal issue concerning AI control and safety. It should consider not only value-neutral human preferences but also moral and ethical considerations. In this study, we introduced FairMindSim, which simulates the moral dilemma through a series of unfair scenarios. We used LLM agents to simulate human behavior, ensuring alignment across various stages. To explore the various socioeconomic motivations, which we refer to as beliefs, that drive both humans and LLM agents as bystanders to intervene in unjust situations involving others, and how these beliefs interact to influence individual behavior, we incorporated knowledge from relevant sociological fields and proposed the Belief-Reward Alignment Behavior Evolution Model (BREM) based on the recursive reward model (RRM). Our findings indicate that, behaviorally, GPT-4o exhibits a stronger sense of social justice, while humans display a richer range of emotions. Additionally, we discussed the potential impact of emotions on behavior. This study provides a theoretical foundation for applications in aligning LLMs with altruistic values.", "keywords": [ "AI alignment; AI Security; AI Ethical; Value alignment; Simulation" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "FXJm5r17Q7", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Ethan X Fang", "gender": "Male", "institution": "Duke University", "country": "", "position": "Assistant Professor" }, { "name": "Juncheng Dong", "gender": "unknown", "institution": "Duke University", "country": "", "position": "PhD student" }, { "name": "Moyang Guo", "gender": "Male", "institution": "Duke University, Duke University", "country": "US", "position": "MS student" }, { "name": "Vahid Tarokh", "gender": "Male", "institution": "Duke University", "country": "", "position": "Professor" }, { "name": "Zhuoran Yang", "gender": "Male", "institution": "Yale University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 42, "n_formula_1": 27, "n_ref_uni": 44, "n_ref": 78, "n_ref_all": 107, "n_fig": 11, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4275, "formula_len_all_1": 2892, "len_all": 201627, "len_all_1": 70622, "len_abs": 1481, "len_title": 113, "len_sents": 50153, "len_sents_1": 30783, "n_sents": 436, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1488, "title": "In-Context Reinforcement Learning From Suboptimal Historical Data", "abs": "Large-scale transformer models have achieved remarkable empirical successes, largely due to their in-context learning capabilities. Inspired by this, we explore training an autoregressive transformer for in-context Reinforcement Learning (RL). In this setting, we initially train a transformer on an offline dataset consisting of trajectories collected from various RL instances, and then fix and use this transformer to create an action policy for new RL instances. Notably, we consider the setting where the offline dataset contains trajectories sampled from suboptimal behavioral policies. In this case, standard autoregressive training corresponds to imitation learning and results in suboptimal performance. To address this, we propose the Decision Importance Transformer (DIT), which emulates the actor-critic algorithm in an in-context manner. In particular, we first train a transformer-based value function that estimates the advantage functions of the behavior policies that collected the suboptimal trajectories. Then we train a transformer-based policy via a weighted maximum likelihood estimation loss, where the weights are constructed based on the trained value function to steer the suboptimal policies to the optimal ones. We conduct extensive experiments to test the performance of DIT on both bandit and Markov Decision Process problems. Our results show that DIT achieves superior performance, particularly when the offline dataset contains suboptimal historical data.", "keywords": [ "In-context Learning; Transformer; Reinforcement Learning" ], "rating_list": [ 3, 6, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "FWqTha5Jh9", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Alexandre Gramfort", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Ambroise Odonnat", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "FR", "position": "PhD student" }, { "name": "Antoine Collas", "gender": "Male", "institution": "INRIA", "country": "", "position": "Postdoc" }, { "name": "Oleksii Kachaiev", "gender": "Male", "institution": "Ukrainian Academy of Banking of the National Bank of Ukraine", "country": "", "position": "MS student" }, { "name": "R\u00e9mi Flamary", "gender": "unknown", "institution": "\u00c9cole Polytechnique", "country": "FR", "position": "Full Professor" }, { "name": "Theo Gnassounou", "gender": "Male", "institution": "Universit\u00e9 Paris-Saclay", "country": "", "position": "PhD student" }, { "name": "Yanis Lalou", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "", "position": "Researcher" }, { "name": "Antoine De mathelin", "gender": "unknown", "institution": "CMLA - ENS Paris Saclay, ENS Paris-Saclay", "country": "", "position": "PhD student" }, { "name": "Thomas Martin Moreau", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 58, "n_ref": 142, "n_ref_all": 202, "n_fig": 10, "n_tab": 13, "L_tab": 15292, "n_element_tab": 437, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2800, "n_element_tab_1": 155, "formula_len_all": 146, "formula_len_all_1": 0, "len_all": 253417, "len_all_1": 74759, "len_abs": 1340, "len_title": 159, "len_sents": 63756, "len_sents_1": 35215, "n_sents": 526, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1344, "title": "DA-Bench: Benchmarking Unsupervised Domain Adaptation Methods with Realistic Validation On Diverse Modalities", "abs": "Unsupervised Domain Adaptation (DA) consists of adapting a model trained on a labeled source domain to perform well on an unlabeled target domain with some data distribution shift.\nWhile many methods have been proposed in the literature, fair and realistic evaluation remains an open question, particularly due to methodological difficulties in selecting hyperparameters in the unsupervised setting.\nWith DA-Bench, we propose a framework to evaluate DA methods on diverse modalities, beyond computer vision task that have been largely explored in the literature. We present a complete and fair evaluation of existing shallow algorithms, including reweighting, mapping, and subspace alignment.\nRealistic hyperparameter selection is performed with nested cross-validation and various unsupervised model selection scores, on both simulated datasets with controlled shifts and real-world datasets across diverse modalities, such as images, text, biomedical, and tabular data.\nOur benchmark highlights the importance of realistic validation and provides practical guidance for real-life applications, with key insights into the choice and impact of model selection approaches.\nDA-Bench is open-source, reproducible, and can be easily extended with novel DA methods, datasets, and model selection criteria without requiring re-evaluating competitors.", "keywords": [ "Benchmark", "Unsupervised Domain Adaptation", "Model selection" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "FWpO8u2lim", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Bo Li", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Hao Zhang", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "Jinwei Chen", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Ming-Ming Cheng", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Full Professor" }, { "name": "Peng-Tao Jiang", "gender": "Male", "institution": "vivo Mobile Communication Co., Ltd", "country": "CN", "position": "Researcher" }, { "name": "Qibin Hou", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuhao Wan", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 47, "n_ref": 114, "n_ref_all": 140, "n_fig": 10, "n_tab": 9, "L_tab": 1133, "n_element_tab": 101, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2260, "n_element_tab_1": 216, "formula_len_all": 212, "formula_len_all_1": 168, "len_all": 128884, "len_all_1": 51588, "len_abs": 1312, "len_title": 164, "len_sents": 32175, "len_sents_1": 23103, "n_sents": 264, "n_sents_1": 151, "venue": "ICLR2025", "decision": "Reject", "L_title": 115, "L_abs": 1320, "title": "ClearSR: Latent Low-Resolution Image Embeddings Help Diffusion-Based Real-World Super Resolution Models See Clearer", "abs": "We present ClearSR, a new method that can better take advantage of latent low-resolution image (LR) embeddings for diffusion-based real-world image super-resolution (Real-ISR). Previous Real-ISR models mostly focus on how to activate more generative priors of text-to-image diffusion models to make the output high-resolution (HR) images look better. However, since these methods rely too much on the generative priors, the content of the output images is often inconsistent with the input LR ones. To mitigate the above issue, in this work, we explore using latent LR embeddings to constrain the control signals from ControlNet, and extract LR information at both detail and structure levels. We show that the proper use of latent LR embeddings can produce higher-quality control signals, which enables the super-resolution results to be more consistent with the LR image and leads to clearer visual results. In addition, we also show that latent LR embeddings can be used to control the inference stage, allowing for the improvement of fidelity and generation ability simultaneously. Experiments demonstrate that our model can achieve better performance across multiple metrics on several test sets and generate more consistent SR results with LR images than existing methods. Our code will be made publicly available.", "keywords": [ "Diffusion Model", "Super-Resolution", "Real-World Image Super-Resolution" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "FVuqJt3c4L", "primary_area": "applications to neuroscience & cognitive science", "n_author": 8, "author_info_list": [ { "name": "Andrei Barbu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Boris Katz", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Principal Research Scientist" }, { "name": "Christopher Wang", "gender": "unknown", "institution": "Computer Science and Artificial Intelligence Laboratory, Electrical Engineering & Computer Science", "country": "US", "position": "PhD student" }, { "name": "Geeling Chau", "gender": "unknown", "institution": "California Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Sabera J Talukder", "gender": "unknown", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Saraswati Soedarmadji", "gender": "unknown", "institution": "California Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Vighnesh Subramaniam", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yisong Yue", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 48, "n_ref": 92, "n_ref_all": 131, "n_fig": 15, "n_tab": 5, "L_tab": 1576, "n_element_tab": 262, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 654, "n_element_tab_1": 43, "formula_len_all": 293, "formula_len_all_1": 115, "len_all": 165059, "len_all_1": 65557, "len_abs": 1185, "len_title": 132, "len_sents": 45798, "len_sents_1": 32272, "n_sents": 328, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 84, "L_abs": 1186, "title": "Population Transformer: Learning Population-level Representations of Neural Activity", "abs": "We present a self-supervised framework that learns population-level codes for arbitrary ensembles of neural recordings at scale. We address two key challenges in scaling models with neural time-series data: sparse and variable electrode distribution across subjects and datasets. The Population Transformer (PopT) stacks on top of pretrained representations and enhances downstream decoding by enabling learned aggregation of multiple spatially-sparse data channels. The pretrained PopT lowers the amount of data required for downstream decoding experiments, while increasing accuracy, even on held-out subjects and tasks. Compared to end-to-end methods, this approach is computationally lightweight, while achieving similar or better decoding performance. We further show how our framework is generalizable to multiple time-series embeddings and neural data modalities. Beyond decoding, we interpret the pretrained PopT and fine-tuned models to show how they can be used to extract neuroscience insights from massive amounts of data. We release our code as well as a pretrained PopT to enable off-the-shelf improvements in multi-channel intracranial data decoding and interpretability.", "keywords": [ "representation learning", "neuroscience", "self supervised learning" ], "rating_list": [ 8, 8, 5, 8, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 1, 3, 3 ], "contribution_list": [ 3, 4, 3, 3, 3 ], "confidence_list": [ 5, 3, 2, 3, 2 ] }, { "paper_id": "FVgizbs3o2", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 3, "author_info_list": [ { "name": "Danilo Mandic", "gender": "unknown", "institution": "Imperial College London", "country": "", "position": "Full Professor" }, { "name": "Mingxue Xu", "gender": "Not Specified", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Yao Lei Xu", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 17, "n_ref_uni": 33, "n_ref": 84, "n_ref_all": 121, "n_fig": 12, "n_tab": 11, "L_tab": 4984, "n_element_tab": 833, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 553, "n_element_tab_1": 41, "formula_len_all": 3601, "formula_len_all_1": 1370, "len_all": 160694, "len_all_1": 63112, "len_abs": 1474, "len_title": 163, "len_sents": 39631, "len_sents_1": 27113, "n_sents": 369, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1453, "title": "TensorGPT: Efficient Compression of Large Language Models based on Tensor-Train Decomposition", "abs": "The Small Language Models (SLMs, or on-device LMs) is a concept corresponding to the Large Language Model (LLM), which has significantly fewer parameters and is typically deployed on low-end devices, like mobile phones and single-board computers (e.g. Raspberry Pi). Unlike LLMs, which utilize the increasing model size for better generalization, SLMs are expected to adjust the exact deployment environment changes. \nFurthermore, most edge applications have battery life concerns, which have never been considered in the GPU servers for data centres.\nTargeting these two issues, this paper focuses on the token embedding compression for adaptivity and low energy requirements in edge applications.\nWe propose a training-free model compression approach based on the Tensor-Train Decomposition (TTD), \nwhereby each pre-trained token embedding vector is converted into a lower-dimensional Matrix Product State (MPS).\nWe then comprehensively investigate the low-rank structures extracted by this approach, regarding the compression ratio, language task performance, latency and energy consumption on a typical low-end device (i.e. Raspberry Pi). \nTaking the sub-billion parameter versions of GPT-2/Cerebres-GPT and OPT as examples, the model compressed with our approach can achieve a comparable language task performance to the original model with around $2.0\\times$ embedding layer compression, while the energy consumption of single query drops by half.", "keywords": [ "model compression", "low-rank factorization", "tensor decomposition" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 5, 5, 5 ] }, { "paper_id": "FV6rPMwmuG", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Bernd Rosenow", "gender": "Male", "institution": "Universit\u00e4t Leipzig", "country": "", "position": "Full Professor" }, { "name": "Marcel K\u00fchn", "gender": "Male", "institution": "Universit\u00e4t Leipzig", "country": "", "position": "MS student" } ], "n_formula": 138, "n_formula_1": 13, "n_ref_uni": 38, "n_ref": 90, "n_ref_all": 148, "n_fig": 12, "n_tab": 1, "L_tab": 105, "n_element_tab": 8, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7174, "formula_len_all_1": 843, "len_all": 216419, "len_all_1": 69765, "len_abs": 1505, "len_title": 147, "len_sents": 77895, "len_sents_1": 33411, "n_sents": 596, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1510, "title": "Anti-Correlated Noise in Epoch-Based Stochastic Gradient Descent: Implications for Weight Variances", "abs": "Stochastic Gradient Descent (SGD) has become a cornerstone of neural network optimization due to its computational efficiency and generalization capabilities. However, the noise introduced by SGD is often assumed to be uncorrelated over time, despite the common practice of epoch-based training where data is sampled without replacement. In this work, we challenge this assumption and investigate the effects of epoch-based noise correlations on the stationary distribution of discrete-time SGD with momentum. Our main contributions are twofold: First, we calculate the exact autocorrelation of the noise during epoch-based training under the assumption that the noise is independent of small fluctuations in the weight vector, revealing that SGD noise is inherently anti-correlated over time. Second, we explore the influence of these anti-correlations on the variance of weight fluctuations. We find that for directions with curvature of the loss greater than a hyperparameter-dependent crossover value, the conventional results for uncorrelated noise are recovered. However, for relatively flat directions, the weight variance is significantly reduced, leading to a considerable decrease in loss fluctuations compared to the constant weight variance assumption. Furthermore, we demonstrate that training with these anti-correlations enhances test performance, suggesting that the inherent noise structure induced by epoch-based training plays a crucial role in finding flatter minima that generalize better.", "keywords": [ "Stochastic Gradient Descent", "Asymptotic Analysis", "Discrete Time", "Hessian" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 1 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "FV5nsugDY1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Biao Hou", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Heng Sun", "gender": "unknown", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Jing Gu", "gender": "unknown", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Licheng Jiao", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Shasha Mao", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Shuyuan Yang", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Tianyu Dong", "gender": "unknown", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 42, "n_ref": 65, "n_ref_all": 89, "n_fig": 4, "n_tab": 4, "L_tab": 2329, "n_element_tab": 293, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1221, "n_element_tab_1": 181, "formula_len_all": 289, "formula_len_all_1": 288, "len_all": 127063, "len_all_1": 53496, "len_abs": 1309, "len_title": 98, "len_sents": 38049, "len_sents_1": 25571, "n_sents": 219, "n_sents_1": 148, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1307, "title": "Hybrid Contrastive Transformer for Visual Tracking", "abs": "Visual object tracking is a research hotspot in the field of computer vision, and has been widely applied in video surveillance, human-computer interaction, unmanned driving and other fields. At present, the object trackers based on Transformer have good performance, but they still face the challenge of confusing target and background in the feature extraction process. To address this issue, we propose a Hybrid Contrastive Transformer Tracker (HCTrack) in this paper, which combines contrastive learning to improve the ability of distinguishing the target and the background in video. Furthermore, a hybrid feature interaction module is presented to realize multi-level information exchange between the features of template and search regions and capture the target-related semantic information of the search frames comprehensively. Additionally, we design a redundant information pruning module to adaptively eliminate the redundant backgrounds according to the global scene information, thereby reducing the interference of the background to the target feature. HCTrack achieves superior tracking accuracy on the GOT-10k and TrackingNet datasets compared to other state-of-the-art trackers, while maintaining fast inference speed, as the contrastive learning is only implemented during training model.", "keywords": [ "visual tracking", "contrastive learning", "hybrid feature", "redundant pruning" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "FUaDMRVrbS", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Ameer Qaqish", "gender": "unknown", "institution": "University of North Carolina at Chapel Hill", "country": "", "position": "PhD student" }, { "name": "Didong Li", "gender": "unknown", "institution": "University of North Carolina at Chapel Hill", "country": "", "position": "Assistant Professor" } ], "n_formula": 74, "n_formula_1": 17, "n_ref_uni": 33, "n_ref": 58, "n_ref_all": 72, "n_fig": 3, "n_tab": 2, "L_tab": 35, "n_element_tab": 4, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 10464, "formula_len_all_1": 2478, "len_all": 159118, "len_all_1": 61097, "len_abs": 1593, "len_title": 63, "len_sents": 47710, "len_sents_1": 26917, "n_sents": 471, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1589, "title": "Identifiability for Gaussian Processes with Holomorphic Kernels", "abs": "Gaussian processes (GPs) are widely recognized for their robustness and flexibility across various domains, including machine learning, time series, spatial statistics, and biomedicine. In addition to their common usage in regression tasks, GP kernel parameters are frequently interpreted in various applications. For example, in spatial transcriptomics, estimated kernel parameters are used to identify spatial variable genes, which exhibit significant expression patterns across different tissue locations. However, before these parameters can be meaningfully interpreted, it is essential to establish their identifiability. Existing studies of GP parameter identifiability have focused primarily on Mat\\'ern-type kernels, as their spectral densities allow for more established mathematical tools. In many real-world applications, particuarly in time series analysis, other kernels such as the squared exponential, periodic, and rational quadratic kernels, as well as their combinations, are also widely used. These kernels share the property of being holomorphic around zero, and their parameter identifiability remains underexplored.\nIn this paper, we bridge this gap by developing a novel theoretical framework for determining kernel parameter identifiability for kernels holomorphic near zero. Our findings enable practitioners to determine which parameters are identifiable in both existing and newly constructed kernels, supporting application-specific interpretation of the identifiable parameters, and highlighting non-identifiable parameters that require careful interpretation.", "keywords": [ "Equivalence of Gaussian random measure; kernel parameters; periodicity; identifiability; interpretability" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "FSlfoBIctk", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Juntao Li", "gender": "Male", "institution": "Soochow University, China", "country": "", "position": "Associate Professor" }, { "name": "Min Zhang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Qiaoming Zhu", "gender": "Male", "institution": "Soochow University", "country": "CN", "position": "Full Professor" }, { "name": "Zechen Sun", "gender": "Female", "institution": "Suzhou University", "country": "", "position": "MS student" }, { "name": "Zecheng Tang", "gender": "Male", "institution": "Soochow University", "country": "", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 8, "n_ref_uni": 38, "n_ref": 73, "n_ref_all": 102, "n_fig": 15, "n_tab": 4, "L_tab": 686, "n_element_tab": 28, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 894, "n_element_tab_1": 131, "formula_len_all": 2065, "formula_len_all_1": 1489, "len_all": 171376, "len_all_1": 65561, "len_abs": 1487, "len_title": 115, "len_sents": 50764, "len_sents_1": 30162, "n_sents": 364, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1501, "title": "LOGO --- Long cOntext aliGnment via efficient preference Optimization", "abs": "Long-context models (LCMs) have shown great potential in processing long input sequences (even more than 100M tokens) conveniently and effectively.\nWith significant progress, recent research has pointed out that LCMs can accurately locate token-level salient information within the context.\nYet, the generation performance of these LCMs is far from satisfactory and might result in misaligned responses, such as hallucinations.\nTo enhance the generation capability of LCMs, existing works have investigated the effects of data size and quality for both pre-training and instruction tuning.\nThough achieving meaningful improvement, previous methods fall short in either effectiveness or efficiency.\nIn this paper, we introduce LOGO (Long cOntext aliGnment via efficient preference Optimization), a training strategy that first introduces preference optimization for long-context alignment.\nTo overcome the GPU memory-bound issue caused by the long sequence, LOGO employs a reference-free preference optimization strategy and adopts a position synthesis method to construct the training data.\nBy training with only 0.3B data on a single 8$\\times$A800 GPU machine for 16 hours, LOGO allows the Llama-3-8B-Instruct-80K model to achieve comparable performance with GPT-4 in real-world long-context tasks while preserving the model's original capabilities on other tasks, e.g., language modeling and MMLU.\nMoreover, LOGO can extend the model's context window size while enhancing its generation performance.", "keywords": [ "Long-context aligment", "efficient preference optimization", "positional indices synthesis" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 3, 1, 3, 4 ], "confidence_list": [ 2, 4, 4, 5 ] }, { "paper_id": "FSjIrOm1vz", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 10, "author_info_list": [ { "name": "Aijun Bai", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Dong Wang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Associate Professor" }, { "name": "Hansi Zeng", "gender": "unknown", "institution": "University of Massachusetts at Amherst", "country": "", "position": "PhD student" }, { "name": "Honglei Zhuang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Kai Hui", "gender": "Male", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Michael Bendersky", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Rolf Jagerman", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Xuanhui Wang", "gender": "Male", "institution": "Google", "country": "", "position": "Software Engineer" }, { "name": "Zhen Qin", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Researcher" }, { "name": "Zhenrui Yue", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 62, "n_ref": 99, "n_ref_all": 154, "n_fig": 20, "n_tab": 8, "L_tab": 2122, "n_element_tab": 183, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1119, "n_element_tab_1": 84, "formula_len_all": 123, "formula_len_all_1": 124, "len_all": 198533, "len_all_1": 68554, "len_abs": 1798, "len_title": 113, "len_sents": 54483, "len_sents_1": 32464, "n_sents": 400, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 65, "L_abs": 1807, "title": "Inference Scaling for Long-Context Retrieval Augmented Generation", "abs": "The scaling of inference computation has unlocked the potential of long-context large language models (LLMs) across diverse settings. For knowledge-intensive tasks, the increased compute is often allocated to incorporate more external knowledge. However, without effectively utilizing such knowledge, solely expanding context does not always enhance performance. In this work, we investigate inference scaling for retrieval augmented generation (RAG), exploring strategies beyond simply increasing the quantity of knowledge. We focus on two inference scaling strategies: in-context learning and iterative prompting. These strategies provide additional flexibility to scale test-time computation (e.g., by increasing retrieved documents or generation steps), thereby enhancing LLMs' ability to effectively acquire and utilize contextual information. We address two key questions: (1) How does RAG performance benefit from the scaling of inference computation when optimally configured? (2) Can we predict the optimal test-time compute allocation for a given budget by modeling the relationship between RAG performance and inference parameters? Our observations reveal that increasing inference computation leads to nearly linear gains in RAG performance when optimally allocated, a relationship we describe as the inference scaling laws for RAG. Building on this, we further develop the computation allocation model to estimate RAG performance across different inference configurations. The model predicts optimal inference parameters under various computation constraints, which align closely with the experimental results. By applying these optimal configurations, we demonstrate that scaling inference compute on long-context LLMs achieves up to 58.9% gains on benchmark datasets compared to standard RAG.", "keywords": [ "inference scaling", "long-context LLM", "retrieval augmented generation" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "FS2nukC2jv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "John Willes", "gender": "Male", "institution": "Vector Institute", "country": "", "position": "Researcher" }, { "name": "Muhammad Adil Asif", "gender": "Male", "institution": "Vector Institute", "country": "CA", "position": "Researcher" }, { "name": "Rahul Krishnan", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "Assistant Professor" }, { "name": "Younwoo Choi", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "MS student" }, { "name": "Ziwen Han", "gender": "Not Specified", "institution": "Scale AI", "country": "US", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 43, "n_ref": 79, "n_ref_all": 102, "n_fig": 4, "n_tab": 10, "L_tab": 784, "n_element_tab": 86, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 139, "n_element_tab_1": 18, "formula_len_all": 1228, "formula_len_all_1": 996, "len_all": 212385, "len_all_1": 68689, "len_abs": 2200, "len_title": 103, "len_sents": 62955, "len_sents_1": 33064, "n_sents": 542, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 1182, "title": "Teaching LLMs How To Learn with Contextual Fine-Tuning", "abs": "Prompting Large Language Models (LLMs), or providing context on the expected model of operation, is an effective way to steer the outputs of such models to satisfy human desiderata after they have been trained. But in rapidly evolving domains, there is often need to fine-tune LLMs to improve either the kind of knowledge in their memory or their abilities to perform open ended reasoning in new domains. When human's learn new concepts, we often do so by linking the new material that we are studying to concepts we have already learned before. To that end, we ask, \"can prompting help us teach LLMs how to learn\". In this work, we study a novel generalization of instruction tuning, called contextual fine-tuning, to fine-tune LLMs. Our method leverages instructional prompts designed to mimic human cognitive strategies in learning and problem-solving to guide the learning process during training, aiming to improve the model\u2019s interpretation and understanding of domain-specific knowledge. We empirically demonstrate that this simple yet effective modification improves the ability of LLMs to be fine-tuned rapidly on new datasets both within the medical and financial domains.", "keywords": [ "Large Language Models" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "FRzCIlkM7I", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Wei Chen", "gender": "Male", "institution": "Hong Kong University of Science and Technology, Guangzhou", "country": "HK", "position": "PhD student" }, { "name": "Yuxuan Liang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 52, "n_formula_1": 10, "n_ref_uni": 47, "n_ref": 92, "n_ref_all": 111, "n_fig": 6, "n_tab": 11, "L_tab": 4743, "n_element_tab": 320, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 6172, "n_element_tab_1": 192, "formula_len_all": 3460, "formula_len_all_1": 551, "len_all": 206590, "len_all_1": 73048, "len_abs": 1517, "len_title": 145, "len_sents": 61113, "len_sents_1": 30784, "n_sents": 561, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 96, "L_abs": 1547, "title": "Expand and Compress: Exploring Tuning Principles for Continual Spatio-Temporal Graph Forecasting", "abs": "The widespread deployment of sensing devices leads to a surge in data for spatio-temporal forecasting applications such as traffic flow, air quality, and wind energy. Although spatio-temporal graph neural networks (STGNNs) have achieved success in modeling various static spatio-temporal forecasting scenarios, real-world spatio-temporal data are typically received in a streaming manner, and the network continuously expands with the installation of new sensors. Thus, spatio-temporal forecasting in streaming scenarios faces dual challenges: the inefficiency of retraining models over newly-arrived data and the detrimental effects of catastrophic forgetting over long-term history. To address these challenges, we propose a novel prompt tuning-based continuous forecasting method, **_EAC_**, following two fundamental tuning principles guided by empirical and theoretical analysis: _**e**xpand **a**nd **c**ompress_, which effectively resolve the aforementioned problems with lightweight tuning parameters. Specifically, we integrate the base STGNN with a continuous prompt pool, utilizing stored prompts (\\ie, few learnable parameters) in memory, and jointly optimize them with the base STGNN. This method ensures that the model sequentially learns from the spatio-temporal data stream to accomplish tasks for corresponding periods. Extensive experimental results on multiple real-world datasets demonstrate the multi-faceted superiority of **_EAC_** over the state-of-the-art baselines, including effectiveness, efficiency, universality, etc.", "keywords": [ "Spatio-temporal Graph", "Continual Forecasting", "Tuning Principle" ], "rating_list": [ 8, 8, 8, 3 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 4, 2 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "FR8mMMiu2L", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Eldad Haber", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "Full Professor" }, { "name": "Shadab Ahamed", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 25, "n_ref_uni": 20, "n_ref": 43, "n_ref_all": 69, "n_fig": 6, "n_tab": 3, "L_tab": 1856, "n_element_tab": 65, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 66, "n_element_tab_1": 18, "formula_len_all": 1504, "formula_len_all_1": 1009, "len_all": 117295, "len_all_1": 63953, "len_abs": 1322, "len_title": 140, "len_sents": 42350, "len_sents_1": 30337, "n_sents": 332, "n_sents_1": 229, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1424, "title": "DAWN-SI: Data-Aware and Noise-Informed Stochastic Interpolation for Solving Inverse Problems", "abs": "Inverse problems, which involve estimating parameters from incomplete or noisy observations, arise in various fields such as medical imaging, geophysics, and signal processing. These problems are often ill-posed, requiring regularization techniques to stabilize the solution. In this work, we employ $\\textit{Stochastic Interpolation (SI)}$, a generative framework that integrates both deterministic and stochastic processes to map a simple reference distribution, such as a Gaussian, to the target distribution. Our method $\\textit{\\textbf{DAWN-SI}}$: $\\textit{\\textbf{D}ata-\\textbf{AW}are and \\textbf{N}oise-informed \\textbf{S}tochastic \\textbf{I}nterpolation}$ incorporates $\\textit{data and noise embedding}$, allowing the model to access representations about the measured data explicitly and also account for noise in the observations, making it particularly robust in scenarios where data is noisy or incomplete. By learning a time-dependent velocity field, SI not only provides accurate solutions but also enables uncertainty quantification by generating multiple plausible outcomes. Unlike pre-trained diffusion models, which may struggle in highly ill-posed settings, our approach is trained specifically for each inverse problem and adapts to varying noise levels. We validate the effectiveness and robustness of our method through extensive numerical experiments on tasks such as image deblurring and tomography.", "keywords": [ "Inverse problems", "Stochastic Interpolation", "Noise-embedding", "data-embedding" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "FR2WQcwjG4", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 16, "author_info_list": [ { "name": "Ali Ansari", "gender": "unknown", "institution": "Sharif University of Technology", "country": "", "position": "Undergrad student" }, { "name": "Arad Maleki", "gender": "Male", "institution": "Sharif University of Technology", "country": "IR", "position": "Undergrad student" }, { "name": "Bahar Dibaei Nia", "gender": "Female", "institution": "Sharif University of Technology", "country": "IR", "position": "Undergrad student" }, { "name": "Hossein Mirzaei", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Jafar Habibi", "gender": "Male", "institution": "Sharif University of Technology", "country": "IR", "position": "Full Professor" }, { "name": "Kian Shamsaie", "gender": "Male", "institution": "Sharif University of Technology, Sharif University of Technology", "country": "IR", "position": "Undergrad student" }, { "name": "Mackenzie W Mathis", "gender": "Female", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "CH", "position": "Assistant Professor" }, { "name": "Mahdi Hajialilue", "gender": "Male", "institution": "Sharif University of Technology", "country": "", "position": "Undergrad student" }, { "name": "Moein Madadi", "gender": "Male", "institution": "Sharif University of Technology", "country": "IR", "position": "Undergrad student" }, { "name": "Mohammad Hossein Rohban", "gender": "Male", "institution": "Sharif University of Technology", "country": "IR", "position": "Assistant Professor" }, { "name": "Mohammad Sabokrou", "gender": "Male", "institution": "Okinawa Institute of Science and Technology (OIST)", "country": "JP", "position": "Researcher" }, { "name": "Mohammadreza Salehi", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Mojtaba Nafez", "gender": "Male", "institution": "Sharif University of Technology", "country": "IR", "position": "MS student" }, { "name": "Sepehr Rezaee", "gender": "Male", "institution": "The National University of Iran", "country": "IR", "position": "Undergrad student" }, { "name": "Zeinab Sadat Taghavi", "gender": "unknown", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Researcher" }, { "name": "Mahdieh Baghshah", "gender": "Female", "institution": "Sharif University of Technology", "country": "IR", "position": "Associate Professor" } ], "n_formula": 16, "n_formula_1": 9, "n_ref_uni": 40, "n_ref": 72, "n_ref_all": 121, "n_fig": 13, "n_tab": 18, "L_tab": 8268, "n_element_tab": 664, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2764, "formula_len_all_1": 1844, "len_all": 292272, "len_all_1": 70659, "len_abs": 1265, "len_title": 128, "len_sents": 77890, "len_sents_1": 33490, "n_sents": 584, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1271, "title": "A Contrastive Teacher-Student Framework for Novelty Detection under Style Shifts", "abs": "There have been several efforts to improve Novelty Detection (ND) performance. However, ND methods often suffer significant performance drops under minor distribution shifts caused by changes in the environment, known as style shifts. This challenge arises from the ND setup, where the absence of out-of-distribution (OOD) samples during training causes the detector to be biased toward the dominant style features in the in-distribution (ID) data. As a result, the model mistakenly learns to correlate style with core features, using this shortcut for detection. Robust ND is crucial for real-world applications like autonomous driving and medical imaging, where test samples may have different styles than the training data. Motivated by this, we propose a robust ND method that crafts an auxiliary OOD set with style features similar to the ID set but with different core features. Then, a task-based knowledge distillation strategy is utilized to distinguish core features from style features and help our model rely on core features for discriminating crafted OOD and ID sets. We verified the effectiveness of our method through extensive experimental evaluations on several datasets, including synthetic and real-world benchmarks, against nine different ND methods.", "keywords": [ "Novelty Detection", "Robustness Under Distribution Shift", "Task-Based Knowledge Distillation", "Robustness Under Style Shift", "One-Class Classification" ], "rating_list": [ 3, 6, 8, 5, 6 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 5, 4 ] }, { "paper_id": "FQhDIGuaJ4", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 10, "author_info_list": [ { "name": "Haodong Feng", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Long Wei", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Postdoc" }, { "name": "Peiyan Hu", "gender": "Not Specified", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Rui Wang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Ruiqi Feng", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Tailin Wu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiang Zheng", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "Intern" }, { "name": "Yue Wang", "gender": "Male", "institution": "Microsoft Research Aisa", "country": "CN", "position": "Researcher" }, { "name": "Zhi-Ming Ma", "gender": "unknown", "institution": "Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Tao Zhang", "gender": "Male", "institution": "Wuhan University", "country": "", "position": "Undergrad student" } ], "n_formula": 37, "n_formula_1": 18, "n_ref_uni": 89, "n_ref": 169, "n_ref_all": 212, "n_fig": 13, "n_tab": 24, "L_tab": 6795, "n_element_tab": 729, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1589, "n_element_tab_1": 160, "formula_len_all": 2656, "formula_len_all_1": 1317, "len_all": 283559, "len_all_1": 66944, "len_abs": 1658, "len_title": 81, "len_sents": 73351, "len_sents_1": 29344, "n_sents": 579, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 33, "L_abs": 1664, "title": "Wavelet Diffusion Neural Operator", "abs": "Simulating and controlling physical systems described by partial differential equations (PDEs) are crucial tasks across science and engineering. Recently, diffusion generative models have emerged as a competitive class of methods for these tasks due to their ability to capture long-term dependencies and model high-dimensional states. However, diffusion models typically struggle with handling system states with abrupt changes and generalizing to higher resolutions. In this work, we propose Wavelet Diffusion Neural Operator (WDNO), a novel PDE simulation and control framework that enhances the handling of these complexities. WDNO comprises two key innovations. Firstly, WDNO performs diffusion-based generative modeling in the wavelet domain for the entire trajectory to handle abrupt changes and long-term dependencies effectively. Secondly, to address the issue of poor generalization across different resolutions, which is one of the fundamental tasks in modeling physical systems, we introduce multi-resolution training. We validate WDNO on five physical systems, including 1D advection equation, three challenging physical systems with abrupt changes (1D Burgers' equation, 1D compressible Navier-Stokes equation and 2D incompressible fluid), and a real-world dataset ERA5, which demonstrates superior performance on both simulation and control tasks over state-of-the-art methods, with significant improvements in long-term and detail prediction accuracy. Remarkably, in the challenging context of the 2D high-dimensional and indirect control task aimed at reducing smoke leakage, WDNO reduces the leakage by 33.2% compared to the second-best baseline.", "keywords": [ "PDE", "physics", "simulation", "control", "diffusion model", "wavelet", "abrupt changes", "multi-resolution" ], "rating_list": [ 8, 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 5, 4, 4 ] }, { "paper_id": "FQc7gi8XvS", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Hanmin Li", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Peter Richtarik", "gender": "Male", "institution": "King Abdullah University of Science and Technology (KAUST)", "country": "SA", "position": "Full Professor" } ], "n_formula": 199, "n_formula_1": 33, "n_ref_uni": 60, "n_ref": 123, "n_ref_all": 161, "n_fig": 7, "n_tab": 2, "L_tab": 1182, "n_element_tab": 35, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 421, "n_element_tab_1": 126, "formula_len_all": 14862, "formula_len_all_1": 2375, "len_all": 237426, "len_all_1": 68257, "len_abs": 1144, "len_title": 114, "len_sents": 69164, "len_sents_1": 29319, "n_sents": 685, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1149, "title": "On the Convergence of FedProx with Extrapolation and Inexact Prox", "abs": "Enhancing the FedProx federated learning algorithm (Li et al., 2020) with server-side extrapolation, Li et al. (2024a) recently introduced the FedExProx method. Their theoretical analysis, however, relies on the assumption that each client computes a certain proximal operator exactly, which is impractical since this is virtually never possible to do in real settings. In this paper, we investigate the behavior of FedExProx without this exactness assumption in the smooth and globally strongly convex setting. We establish a general convergence result, showing that inexactness leads to convergence to a neighborhood of the solution. Additionally, we demonstrate that, with careful control, the adverse effects of this inexactness can be mitigated. By linking inexactness to biased compression (Beznosikov et al., 2023), we refine our analysis, highlighting robustness of extrapolation to inexact proximal updates. We also examine the local iteration complexity required by each client to achieved the required level of inexactness using various local optimizers. Our theoretical insights are validated through comprehensive numerical experiments.", "keywords": [ "Federated Learning", "Optimization" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 2 ] }, { "paper_id": "FQaZeFGca2", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Shaogang Gong", "gender": "Male", "institution": "Queen Mary University London", "country": "GB", "position": "Full Professor" }, { "name": "Yu Cao", "gender": "Male", "institution": "Queen Mary University of London", "country": "GB", "position": "PhD student" } ], "n_formula": 33, "n_formula_1": 10, "n_ref_uni": 44, "n_ref": 102, "n_ref_all": 121, "n_fig": 10, "n_tab": 6, "L_tab": 1115, "n_element_tab": 163, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 621, "n_element_tab_1": 12, "formula_len_all": 2376, "formula_len_all_1": 797, "len_all": 145807, "len_all_1": 59557, "len_abs": 1855, "len_title": 70, "len_sents": 43973, "len_sents_1": 28980, "n_sents": 326, "n_sents_1": 195, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1816, "title": "EXPLORING FEW-SHOT IMAGE GENERATION WITH MINIMIZED RISK OF OVERFITTING", "abs": "Few-shot image generation (FSIG) using deep generative models (DGMs) presents a significant challenge in accurately estimating the distribution of the target domain with extremely limited samples. Recent work has addressed the problem using a transfer learning approach, i.e., fine-tuning, leveraging a DGM that pre-trained on a large-scale source domain dataset, and then adapting it to the target domain with very limited samples. However, despite various proposed regularization techniques, existing frameworks lack a systematic mechanism to analyze the degree of overfitting, relying primarily on empirical validation without rigorous theoretical grounding.\nWe present Few-Shot Diffusion-regularized Representation Learning (FS-DRL), an innovative approach designed to minimize the risk of over-fitting while preserving distribution consistency in target image adaptation. \nOur method is distinct from conventional methods in two aspects: First, instead of fine-tuning, FS-DRL employs a novel scalable Invariant Guidance Matrix (IGM) during the diffusion process, which acts as a regularizer in the feature space of the model. This IGM is designed to have the same dimensionality as the target images, effectively constraining its capacity and encouraging it to learn a low-dimensional manifold that captures the essential structure of the target domain. Second, our method introduces a controllable parameter called sharing degree, which determines how many target images correspond to each IGM, enabling a fine-grained balance between overfitting risk and model flexibility, thus providing a quantifiable mechanism to analyze and mitigate overfitting.\nExtensive experiments demonstrate that our approach effectively mitigates overfitting, enabling efficient and robust few-shot learning across diverse domains.", "keywords": [ "few shot learning", "generative model", "diffusion model" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "FQEFWGT19m", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Annie Qu", "gender": "Female", "institution": "University of California, Irvine", "country": "US", "position": "Full Professor" }, { "name": "Qi Xu", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Yang Bai", "gender": "unknown", "institution": "Shanghai University of Finance and Economics", "country": "", "position": "Associate Professor" }, { "name": "Yang Sui", "gender": "Male", "institution": "Shanghai University of Finance and Economics", "country": "", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 27, "n_ref_uni": 65, "n_ref": 107, "n_ref_all": 137, "n_fig": 9, "n_tab": 8, "L_tab": 2119, "n_element_tab": 96, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1965, "formula_len_all_1": 1955, "len_all": 234756, "len_all_1": 80042, "len_abs": 4538, "len_title": 122, "len_sents": 58163, "len_sents_1": 32840, "n_sents": 510, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1282, "title": "Multi-task Learning for Heterogeneous Multi-source Block-Wise Missing Data", "abs": "Multi-task learning (MTL) has emerged as an imperative machine learning tool to solve multiple learning tasks simultaneously and has been successfully applied to healthcare, marketing, and biomedical fields. However, in order to borrow information across different tasks effectively, it is essential to utilize both homogeneous and heterogeneous information. Among the extensive literature on MTL, various forms of heterogeneity are presented in MTL problems, such as block-wise, distribution, and posterior heterogeneity. Existing methods, however, struggle to tackle these forms of heterogeneity simultaneously in a unified framework. In this paper, we propose a two-step learning strategy for MTL which addresses the aforementioned heterogeneity. First, we impute the missing blocks using shared representations extracted from homogeneous source across different tasks. Next, we disentangle the mappings between input features and responses into a shared component and a task-specific component, respectively, thereby enabling information borrowing through the shared component. Our numerical experiments and real-data analysis from the ADNI database demonstrate the superior MTL performance of the proposed method compared to a single task learning and other competing methods.", "keywords": [ "data integration", "disentangled representations", "distribution shift", "posterior drift" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "FPfCUJTsCn", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Bin Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Fangzhou Zhu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Feng Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xijun Li", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zijie Geng", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 62, "n_formula_1": 20, "n_ref_uni": 54, "n_ref": 122, "n_ref_all": 151, "n_fig": 9, "n_tab": 10, "L_tab": 3473, "n_element_tab": 213, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 692, "n_element_tab_1": 106, "formula_len_all": 4962, "formula_len_all_1": 1740, "len_all": 208820, "len_all_1": 71761, "len_abs": 1672, "len_title": 89, "len_sents": 57971, "len_sents_1": 30106, "n_sents": 532, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 41, "L_abs": 1713, "title": "Differentiable Integer Linear Programming", "abs": "Machine learning (ML) techniques have shown great potential in generating high-quality solutions for integer linear programs (ILPs).\nHowever, existing methods typically rely on a *supervised learning* paradigm, leading to (1) *expensive training cost* due to repeated invocations of traditional solvers to generate training labels, and (2) *plausible yet infeasible solutions* due to the misalignment between the training objective (minimizing prediction loss) and the inference objective (generating high-quality solutions).\nTo tackle this challenge, we propose **DiffILO** (**Diff**erentiable **I**nteger **L**inear Programming **O**ptimization), an *unsupervised learning paradigm for learning to solve ILPs*.\nSpecifically, through a novel probabilistic modeling, DiffILO reformulates ILPs---discrete and constrained optimization problems---into continuous, differentiable (almost everywhere), and unconstrained optimization problems.\nThis reformulation enables DiffILO to simultaneously solve ILPs and train the model via straightforward gradient descent, providing two major advantages.\nFirst, it significantly reduces the training cost, as the training process does not need the aid of traditional solvers at all.\nSecond, it facilitates the generation of feasible and high-quality solutions, as the model *learns to solve ILPs* in an end-to-end manner, thus aligning the training and inference objectives.\nExperiments on commonly used ILP datasets demonstrate that DiffILO not only achieves an average training speedup of $13.2$ times compared to supervised methods, but also outperforms them by generating heuristic solutions with significantly higher feasibility ratios and much better solution qualities.", "keywords": [ "Integer Linear Programming", "Learning to Optimize" ], "rating_list": [ 8, 6, 8, 8, 6 ], "soundness_list": [ 4, 3, 3, 4, 2 ], "presentation_list": [ 4, 3, 4, 4, 3 ], "contribution_list": [ 3, 2, 4, 4, 4 ], "confidence_list": [ 3, 3, 4, 4, 4 ] }, { "paper_id": "FPQzXME9NK", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Hoang V. Tran", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Huyen Trang Pham", "gender": "Female", "institution": "VinAI Research", "country": "VN", "position": "Intern" }, { "name": "Minh-Khoi Nguyen-Nhat", "gender": "Male", "institution": "FPT", "country": "VN", "position": "Researcher" }, { "name": "Tam Le", "gender": "Male", "institution": "The Institute of Statistical Mathematics (ISM)", "country": "JP", "position": "Assistant Professor" }, { "name": "Tan Minh Nguyen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Thanh Chu", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Undergrad student" } ], "n_formula": 94, "n_formula_1": 23, "n_ref_uni": 44, "n_ref": 98, "n_ref_all": 130, "n_fig": 13, "n_tab": 6, "L_tab": 1180, "n_element_tab": 102, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 579, "n_element_tab_1": 54, "formula_len_all": 9316, "formula_len_all_1": 2479, "len_all": 176348, "len_all_1": 70716, "len_abs": 3045, "len_title": 90, "len_sents": 43738, "len_sents_1": 25255, "n_sents": 506, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 1540, "title": "Spherical Tree-Sliced Wasserstein Distance", "abs": "Sliced Optimal Transport (OT) simplifies the OT problem in high-dimensional spaces by projecting supports of input measures onto one-dimensional lines, then exploiting the closed-form expression of the univariate OT to reduce the computational burden of OT. Recently, the Tree-Sliced method has been introduced to replace these lines with more intricate structures, known as tree systems. This approach enhances the ability to capture topological information of integration domains in Sliced OT while maintaining low computational cost. Inspired by this approach, in this paper, we present an adaptation of tree systems on OT problem for measures supported on a sphere. As counterpart to the Radon transform variant on tree systems, we propose a novel spherical Radon transform, with a new integration domain called spherical trees. By leveraging this transform and exploiting the spherical tree structures, we derive closed-form expressions for OT problems on the sphere. Consequently, we obtain an efficient metric for measures on the sphere, named Spherical Tree-Sliced Wasserstein (STSW) distance. We provide an extensive theoretical analysis to demonstrate the topology of spherical trees, the well-definedness and injectivity of our Radon transform variant, which leads to an orthogonally invariant distance between spherical measures. Finally, we conduct a wide range of numerical experiments, including gradient flows and self-supervised learning, to assess the performance of our proposed metric, comparing it to recent benchmarks.", "keywords": [ "tree-sliced wasserstein distance", "spherical optimal transport", "equivariance" ], "rating_list": [ 6, 6, 8, 6, 6 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 3, 4 ] }, { "paper_id": "FPBce2P1er", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Kim Stachenfeld", "gender": "Female", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" }, { "name": "Samuel Lippl", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" } ], "n_formula": 55, "n_formula_1": 8, "n_ref_uni": 77, "n_ref": 116, "n_ref_all": 174, "n_fig": 41, "n_tab": 2, "L_tab": 724, "n_element_tab": 40, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 428, "n_element_tab_1": 21, "formula_len_all": 3309, "formula_len_all_1": 576, "len_all": 341853, "len_all_1": 74195, "len_abs": 1731, "len_title": 134, "len_sents": 94716, "len_sents_1": 36303, "n_sents": 800, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1737, "title": "When does compositional structure yield compositional generalization? A kernel theory.", "abs": "Compositional generalization (the ability to respond correctly to novel combinations of familiar components) is thought to be a cornerstone of intelligent behavior. Compositionally structured (e.g. disentangled) representations are essential for this; however, the conditions under which they yield compositional generalization remain unclear. To address this gap, we present a general theory of compositional generalization in kernel models with fixed, compositionally structured representations, a tractable framework for characterizing the impact of dataset statistics on generalization. We find that these models are constrained to adding up values assigned to each combination of components seen during training (\"conjunction-wise additivity\"). This imposes fundamental restrictions on the set of tasks compositionally structured kernel models can learn, in particular preventing them from transitively generalizing equivalence relations. Even for compositional tasks that they can in principle learn, we identify novel failure modes in compositional generalization that arise from biases in the training data and affect important compositional building blocks such as symbolic addition and context dependence (memorization leak and shortcut bias). Finally, we empirically validate our theory, showing that it captures the behavior of deep neural networks (convolutional networks, residual networks, and Vision Transformers) trained on a set of compositional tasks with similarly structured data. Ultimately, this work provides a theoretical perspective on how statistical structure in the training data can affect compositional generalization, with implications for how to identify and remedy failure modes in deep learning models.", "keywords": [ "compositional generalization", "rule learning", "kernel regression", "kernel models", "relational reasoning", "memorization", "shortcuts", "dataset statistics", "norm minimization", "implicit regularization", "disentanglement" ], "rating_list": [ 10, 6, 8, 3 ], "soundness_list": [ 4, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "FP77VtEuaT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Gabriele Venturato", "gender": "unknown", "institution": "KU Leuven", "country": "", "position": "PhD student" }, { "name": "Luc De Raedt", "gender": "Male", "institution": "KU Leuven", "country": "", "position": "Full Professor" }, { "name": "Pedro Zuidberg Dos Martires", "gender": "Male", "institution": "KU Leuven", "country": "", "position": "PhD student" }, { "name": "RISHI HAZRA", "gender": "Male", "institution": "\u00d6rebro University", "country": "SE", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 60, "n_ref": 153, "n_ref_all": 198, "n_fig": 13, "n_tab": 3, "L_tab": 1501, "n_element_tab": 59, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 152, "formula_len_all_1": 15, "len_all": 248695, "len_all_1": 60335, "len_abs": 1192, "len_title": 110, "len_sents": 51950, "len_sents_1": 29996, "n_sents": 417, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 1202, "title": "Can Large Language Models Reason? A Characterization via 3-SAT", "abs": "Large Language Models (LLMs) have been touted as AI models possessing advanced reasoning abilities. However, recent works have shown that LLMs often bypass true reasoning using shortcuts, sparking skepticism. To study the reasoning capabilities in a principled fashion, we adopt a computational theory perspective and propose an experimental protocol centered on 3-SAT -- the prototypical NP-complete problem lying at the core of logical reasoning and constraint satisfaction tasks. Specifically, we examine the phase transitions in random 3-SAT and characterize the reasoning abilities of LLMs by varying the inherent hardness of the problem instances. Our experimental evidence shows that LLMs are incapable of performing true reasoning, as required for solving 3-SAT problems. Moreover, we observe significant performance variation based on the inherent hardness of the problems -- performing poorly on harder instances and vice versa. Importantly, we show that integrating external reasoners can considerably enhance LLM performance. By following a principled experimental protocol, our study draws concrete conclusions and moves beyond the anecdotal evidence often found in LLM reasoning research.", "keywords": [ "Large Language Models", "Logic", "Reasoning", "Satisfiability", "Phase Transitions" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "FOcleL0ltt", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Zeyu Zheng", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Associate Professor" }, { "name": "Dake", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 49, "n_ref": 59, "n_ref_all": 78, "n_fig": 2, "n_tab": 8, "L_tab": 1933, "n_element_tab": 150, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1452, "n_element_tab_1": 145, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 115976, "len_all_1": 58480, "len_abs": 1085, "len_title": 125, "len_sents": 31284, "len_sents_1": 26678, "n_sents": 232, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1100, "title": "UniComposer: Band-Level Music Composition with Symbolic and Audio Unification", "abs": "Multi-track deep music generation has largely focused on pre-specified structures and instruments. However, it remains a challenge to generate \"band-level\" full-length music that is capable of allocating instruments based on musical features, their expressive potential, and their performance characteristics differences. Moreover, the representations of symbolic music and audio music have been treated as distinct sub-areas, without a unified architecture to join their own advantages. In this work, we introduce $\\textbf{UniComposer}$, a novel music generation pipeline that composes at the band level, utilizing a hierarchical multi-track music representation complemented by four cascaded diffusion models which progressively generate rhythm features, and unified features extracted from both symbolic and audio music by autoencoders. Experiments and analysis demonstrate that UniComposer achieves a unified latent space for symbolic and audio music, and is capable of generating band-level compositions with well structured multi-track arrangements, surpassing previous methods in performances.", "keywords": [ "Symbolic and Audio Music", "Unified Latent Space", "Band-Level Music Generation", "Feature Extraction", "Generative Models" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 1, 1, 2 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "FNiqaC382D", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Hangguan Shan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Na Lee", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Wei Ni", "gender": "Male", "institution": "CSIRO", "country": "AU", "position": "Principal Researcher" }, { "name": "Wenjie Zhang", "gender": "Female", "institution": "the university of new south wales", "country": "AU", "position": "Full Professor" }, { "name": "Xinyu Li", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Yamin Wang", "gender": "Female", "institution": "China Mobile Communications Company Limited Research Institute", "country": "CN", "position": "Associate Professor" } ], "n_formula": 147, "n_formula_1": 25, "n_ref_uni": 45, "n_ref": 70, "n_ref_all": 98, "n_fig": 4, "n_tab": 5, "L_tab": 1157, "n_element_tab": 93, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 12, "n_element_tab_1": 1, "formula_len_all": 20988, "formula_len_all_1": 2734, "len_all": 214742, "len_all_1": 59311, "len_abs": 1549, "len_title": 130, "len_sents": 58002, "len_sents_1": 28239, "n_sents": 479, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1556, "title": "Provable Causal State Representation under Asynchronous Diffusion Model for POMDPs", "abs": "A major challenge in applying reinforcement learning (RL) to real-world scenarios is managing high-dimensional, noisy perception input signals. Identifying and utilizing representations that contain sufficient and essential information for decision-making tasks is key to computational efficiency and generalization of RL by reducing bias in decision-making processes. In this paper, we present a new RL framework, named *Causal State Representation under Asynchronous Diffusion Model (CSR-ADM)*, which accommodates and enhances any RL algorithm for partially observable Markov decision processes (POMDPs) with perturbed inputs. A new asynchronous diffusion model is proposed to denoise both reward and observation spaces, and integrated with the bisimulation technology to capture causal state representations in POMDPs. Notably, the causal state is the coarsest partition of the denoised observations. We link the causal state to a causal feature set and provide theoretical guarantees by deriving the upper bound on value function approximation between the noisy observation space and the causal state space, demonstrating equivalence to bisimulation under the Lipschitz assumption. To the best of our knowledge, CSR-ADM is the first framework to approximate causal states with diffusion models, substantiated by a comprehensive theoretical foundation. Extensive experiments on Roboschool tasks show that CSR-ADM outperforms state-of-the-art methods, significantly improving the robustness of existing RL algorithms under varying scales of random noise.", "keywords": [ "diffusion model", "causal state representation", "model uncertainty", "bisimulaion", "POMDP" ], "rating_list": [ 6, 3, 5, 8 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 1, 1, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 1, 2, 2, 2 ] }, { "paper_id": "FNGZqMp6Fi", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Bilal Zahid Hussain", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "MS student" }, { "name": "Fatahlla Moreh", "gender": "Male", "institution": "Christian-Albrechts-Universit\u00e4t Kiel", "country": "DE", "position": "PhD student" }, { "name": "Mohammad Ammar", "gender": "Male", "institution": "Aligarh Muslim University", "country": "IN", "position": "Undergrad student" }, { "name": "Sven Tomforde", "gender": "Male", "institution": "Christian-Albrechts-Universit\u00e4t Kiel", "country": "", "position": "Full Professor" }, { "name": "Yusuf Hasan", "gender": "Male", "institution": "Aligarh Muslim University, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "Undergrad student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 21, "n_ref": 29, "n_ref_all": 36, "n_fig": 3, "n_tab": 1, "L_tab": 605, "n_element_tab": 96, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 1325, "n_element_tab_1": 17, "formula_len_all": 215, "formula_len_all_1": 228, "len_all": 88559, "len_all_1": 58448, "len_abs": 1013, "len_title": 183, "len_sents": 29212, "len_sents_1": 28875, "n_sents": 205, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 136, "L_abs": 1017, "title": "MicroCrackAttentionNeXt: Advancing Microcrack Detection in Wave Field Analysis Using Deep Neural Networks through Feature Visualization.", "abs": "Micro Crack detection using deep neural networks(DNNs) through an automated pipeline using wave fields interacting with the damaged areas is highly sought after. However, these high dimensional spatio-temporal crack data are limited, moreover these dataset have large dimension in the temporal domain. The dataset exhibits a pronounced class imbalance, with crack pixels accounting for an average of only 5% of the total pixels per sample. This severe imbalance presents a challenge for deep learning models when dealing with various microscale cracks, as the network tends to favor the majority class, often resulting in reduced detection accuracy. This study proposes an asymmetric encoder\u2013decoder network with Adaptive Feature Reutilization Block for micro-crack detection. The impact of various activation and loss functions were examined through feature space visualisation using manifold discovery and analysis (MDA) algorithm. The optimized architecture and training methodology achieved an accuracy of 87.74%.", "keywords": [ "manifold discovery and analysis", "Feature Visualisation", "structural health monitoring", "Attention mechanism", "wave field data", "micro scale cracks", "Loss functions" ], "rating_list": [ 1, 3, 1, 3, 3 ], "soundness_list": [ 2, 3, 1, 3, 3 ], "presentation_list": [ 2, 2, 2, 2, 2 ], "contribution_list": [ 1, 1, 1, 2, 2 ], "confidence_list": [ 4, 5, 4, 3, 4 ] }, { "paper_id": "FNDudoox4A", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Jinglin Xu", "gender": "Female", "institution": "University of Science and Technology Beijing", "country": "CN", "position": "Associate Professor" }, { "name": "Yanzhe Chen", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yuxin Peng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Zhiwen Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 23, "n_ref": 57, "n_ref_all": 79, "n_fig": 5, "n_tab": 4, "L_tab": 969, "n_element_tab": 147, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 1257, "n_element_tab_1": 188, "formula_len_all": 871, "formula_len_all_1": 875, "len_all": 143462, "len_all_1": 63978, "len_abs": 1630, "len_title": 133, "len_sents": 32003, "len_sents_1": 25537, "n_sents": 240, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1657, "title": "Pseudo Meets Zero: Boosting Zero-Shot Composed Image Retrieval with Synthetic Images", "abs": "Composed Image Retrieval (CIR) employs a triplet architecture to combine a reference image with modified text for target image retrieval. To mitigate high annotation costs, Zero-Shot CIR (ZS-CIR) methods eliminate the need for manually annotated triplets. Current methods typically map images to tokens and concatenate them with modified text. However, they encounter challenges during inference, especially with fine-grained and multi-attribute modifications. We argue that these challenges stem from insufficient explicit modeling of triplet relationships, which complicates fine-grained interactions and directional guidance. To this end, we propose a Synthetic Image-Oriented training paradigm that automates pseudo target image generation, facilitating efficient triplet construction and accommodating inherent target ambiguity. Furthermore, we propose the Pseudo domAiN Decoupling-Alignment (PANDA) model to mitigate the Autophagy phenomenon caused by fitting targets with pseudo images. We observe that synthetic images are intermediate between visual and textual domains in triplets. Regarding this phenomenon, we design the Orthogonal Semantic Decoupling module to disentangle the pseudo domain into visual and textual components. Additionally, Shared Domain Interaction and Mutual Shift Constraint modules are proposed to collaboratively constrain the disentangled components, bridging the gap between pseudo and real triplets while enhancing their semantic consistency. Extensive experiments demonstrate that the proposed PANDA model outperforms existing state-of-the-art methods across two general scenarios and two domain-specific CIR datasets.", "keywords": [ "Zero-Shot Composed Image Retrieval", "Synthetic Images", "Multimdoal" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "FN7n7JRjsk", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Bingyi Jing", "gender": "Male", "institution": "South University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Hongxin Wei", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Songxin Zhang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Researcher" }, { "name": "Wenyu Jiang", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Zhenlong Liu", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "XieZejian", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Researcher" } ], "n_formula": 13, "n_formula_1": 10, "n_ref_uni": 58, "n_ref": 77, "n_ref_all": 110, "n_fig": 15, "n_tab": 9, "L_tab": 5124, "n_element_tab": 466, "n_fig_1": 15, "n_tab_1": 2, "L_tab_1": 2541, "n_element_tab_1": 165, "formula_len_all": 685, "formula_len_all_1": 745, "len_all": 161552, "len_all_1": 60375, "len_abs": 1341, "len_title": 118, "len_sents": 40583, "len_sents_1": 26468, "n_sents": 303, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1353, "title": "Exploring Learning Complexity for Efficient Downstream Dataset Pruning", "abs": "The ever-increasing fine-tuning cost of large-scale pre-trained models gives rise to the importance of dataset pruning, which aims to reduce dataset size while maintaining task performance.\nHowever, existing dataset pruning methods require training on the entire dataset, which is impractical for large-scale pre-trained models.\nIn this paper, we propose a straightforward, novel, and training-free hardness score named Distorting-based Learning Complexity (DLC), to identify informative images and instructions from the downstream dataset efficiently.\nOur method is motivated by the observation that easy samples learned faster can also be learned with fewer parameters.\nSpecifically, we define the Learning Complexity to quantify sample hardness and utilize a lightweight weights masking process for fast estimation, instead of the costly SGD optimization.\nBased on DLC, we further design a flexible under-sampling strategy with randomness (dubbed FlexRand), replacing the top-K strategy, to alleviate the severe subset distribution shift.\nExtensive experiments with downstream image and instructions dataset pruning benchmarks demonstrate the effectiveness and efficiency of the proposed approach.\nIn the images pruning benchmark, DLC significantly reduces the pruning time by 35$\\times$ while establishing state-of-the-art performance with FlexRand.", "keywords": [ "data efficiency" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "FM21yYBhuE", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Runkang Yang", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "Undergrad student" }, { "name": "Tao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xinyi Shang", "gender": "Female", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Yi Tang", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "peng sun", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 36, "n_ref": 74, "n_ref_all": 121, "n_fig": 24, "n_tab": 9, "L_tab": 5497, "n_element_tab": 1058, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 2892, "n_element_tab_1": 39, "formula_len_all": 234, "formula_len_all_1": 199, "len_all": 166975, "len_all_1": 59790, "len_abs": 1184, "len_title": 66, "len_sents": 48661, "len_sents_1": 26664, "n_sents": 363, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1189, "title": "Equally Critical: Samples, Targets, and Their Mappings in Datasets", "abs": "Neural scaling laws highlight the trade-off between test error reduction and increased resources in machine learning, revealing diminishing returns as data volume, model size, and computational power increase.\nThis inefficiency poses sustainability challenges, as marginal performance gains necessitate exponential resource consumption.\nRecent works have investigated these laws from a data-efficient standpoint, primarily concentrating on sample optimization, while largely neglecting the influence of target.\nIn this study, we first demonstrate that, given an equivalent training budget, employing soft targets on a 10% subset can outperform the use of one-hot targets on the full dataset. Building on this observation, we review existing paradigms in the sample-target relationship, categorizing them into distinct sample-to-target mapping strategies.\nSubsequently, we propose a unified loss framework to assess their impact on training efficiency. Finally, we conduct a comprehensive analysis of how variations in target and sample types, quantities, and qualities influence training efficiency across three training strategies, providing six key insights to enhance training efficacy.", "keywords": [ "Data-efficient Learning" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 1, 2, 3 ], "confidence_list": [ 1, 3, 3, 3 ] }, { "paper_id": "FLR1K8h5Eq", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Hao Zou", "gender": "Male", "institution": "ZGC laboratory", "country": "CN", "position": "Researcher" }, { "name": "Haotian Wang", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Jie Peng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Peng Cui", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Renzhe Xu", "gender": "Male", "institution": "Shanghai University of Finance and Economics", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 58, "n_formula_1": 36, "n_ref_uni": 41, "n_ref": 79, "n_ref_all": 93, "n_fig": 4, "n_tab": 5, "L_tab": 1441, "n_element_tab": 79, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1037, "n_element_tab_1": 36, "formula_len_all": 4873, "formula_len_all_1": 2068, "len_all": 147092, "len_all_1": 70911, "len_abs": 1359, "len_title": 125, "len_sents": 39839, "len_sents_1": 30074, "n_sents": 325, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1242, "title": "Learning Time-shared Hidden Heterogeneity for Counterfactual Outcome Forecast", "abs": "Forecasting counterfactual outcome in the longitudinal setting can be critical for many time-related applications. To solve this problem, the previous works propose to apply different sequence models including long short-term memory (LSTM) networks and transformers to model the relationship between the observed histories, treatments and outcomes, and apply various approaches to remove treatment selection bias. However, these methods neglect the hidden heterogeneity of outcome generation among samples induced by hidden factors which can bring hurdles to counterfactual outcome forecast. To alleviate this problem, we capture the hidden heterogeneity by recovering the hidden factors and incorporate it into the outcome prediction process. Specifically, we propose a Time-shared Heterogeneity Learning from Time Series (THLTS) method which infers the shared part of hidden factors characterizing the heterogeneity across time steps with the architecture of variational encoders (VAE). This method can be a flexible component and combined with arbitrary counterfactual outcome forecast method. Experimental results on (semi-)synthetic datasets demonstrate that combined with our method, the mainstream models can improve their performance.", "keywords": [ "Hidden Heterogeneity; Counterfactual Outcome Forecast; Time series" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 4, 3, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 5, 4 ] }, { "paper_id": "FL6112vyty", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Xiaoliang Ju", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 38, "n_ref": 63, "n_ref_all": 80, "n_fig": 12, "n_tab": 6, "L_tab": 726, "n_element_tab": 81, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 357, "n_element_tab_1": 39, "formula_len_all": 431, "formula_len_all_1": 432, "len_all": 125192, "len_all_1": 59302, "len_abs": 1248, "len_title": 134, "len_sents": 30620, "len_sents_1": 25539, "n_sents": 281, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 85, "L_abs": 1254, "title": "DirectTriGS: Triplane-based Gaussian Splatting Field Representation for 3D Generation", "abs": "We present DirectTriGS, a novel framework designed for 3D object generation with Gaussian Splatting (GS). GS-based rendering for 3D content has gained considerable attention recently. However, there has been limited exploration in directly generating 3D Gaussians compared to traditional generative modeling approaches. The main challenge lies in the complex data structure of GS represented by discrete point clouds with multiple channels.\nTo overcome this challenge, we propose employing the triplane representation, which allows us to represent Gaussian Splatting as an image-like continuous field. This representation effectively encodes both the geometry and texture information, enabling smooth transformation back to Gaussian point clouds and rendering into images by a TriRenderer, with only 2D supervisions. The proposed TriRenderer is fully differentiable, so that the rendering loss can supervise both texture and geometry encoding. Furthermore, the triplane representation can be compressed using a Variational Autoencoder (VAE), which can subsequently be utilized in latent diffusion to generate 3D objects.\nThe experiments demonstrate that the proposed generation framework can produce high-quality 3D object geometry and rendering results.", "keywords": [ "3D generation", "Gaussian Splatting" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "FK8tl47xpP", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "MOHAMMAD GOLBABAEE", "gender": "unknown", "institution": "University of Bristol", "country": "GB", "position": "Assistant Professor" }, { "name": "Matthias J Ehrhardt", "gender": "male", "institution": "University of Bath", "country": "GB", "position": "Associate Professor" }, { "name": "Patrick Fahy", "gender": "unknown", "institution": "University of Bath", "country": "GB", "position": "PhD student" } ], "n_formula": 129, "n_formula_1": 30, "n_ref_uni": 28, "n_ref": 36, "n_ref_all": 79, "n_fig": 24, "n_tab": 3, "L_tab": 318, "n_element_tab": 17, "n_fig_1": 17, "n_tab_1": 2, "L_tab_1": 64, "n_element_tab_1": 6, "formula_len_all": 12970, "formula_len_all_1": 2532, "len_all": 155402, "len_all_1": 65731, "len_abs": 1534, "len_title": 0, "len_sents": 44643, "len_sents_1": 28096, "n_sents": 491, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1541, "title": "Greedy Learning to Optimize with Convergence Guarantees", "abs": "Learning to optimize is an approach that leverages training data to accelerate the solution of optimization problems. Many approaches use unrolling to parametrize the update step and learn optimal parameters. Although L2O has shown empirical advantages over classical optimization algorithms, memory restrictions often greatly limit the unroll length and learned algorithms usually do not provide convergence guarantees. In contrast, we introduce a novel method employing a greedy strategy that learns iteration-specific parameters by minimizing the function value at the next iteration. This enables training over significantly more iterations while maintaining constant GPU memory usage. We parameterize the update such that parameter learning corresponds to solving a convex optimization problem at each iteration. In particular, we explore preconditioned gradient descent with multiple parametrizations including a novel convolutional preconditioner. With our learned algorithm, convergence in the training set is proved even when the preconditioner is neither symmetric nor positive definite. Convergence on a class of unseen functions is also obtained, ensuring robust performance and generalization beyond the training data. We test our learned algorithms on two inverse problems, image deblurring and Computed Tomography, on which learned convolutional preconditioners demonstrate improved empirical performance over classical optimization algorithms such as Nesterov's Accelerated Gradient Method and the quasi-Newton method L-BFGS.", "keywords": [ "Optimization", "Inverse Problems", "Learning to Optimize", "Preconditioning", "Imaging" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 4, 3, 2, 4 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "FK6T0U4Mg1", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Hua Huang", "gender": "Male", "institution": "Beijing Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Jia Li", "gender": "Male", "institution": "Beijing Normal University", "country": "CN", "position": "Lecturer" }, { "name": "Sike Wang", "gender": "Male", "institution": "Beijing Normal University", "country": "CN", "position": "MS student" }, { "name": "Ziming Yu", "gender": "Male", "institution": "Beijing Normal University", "country": "CN", "position": "MS student" }, { "name": "Pan Zhou", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 32, "n_formula_1": 28, "n_ref_uni": 54, "n_ref": 158, "n_ref_all": 196, "n_fig": 3, "n_tab": 9, "L_tab": 3351, "n_element_tab": 328, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1297, "n_element_tab_1": 167, "formula_len_all": 2555, "formula_len_all_1": 2178, "len_all": 185995, "len_all_1": 71208, "len_abs": 1168, "len_title": 136, "len_sents": 45946, "len_sents_1": 29651, "n_sents": 404, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1175, "title": "SubZero: Random Subspace Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning", "abs": "Fine-tuning Large Language Models (LLMs) has proven effective for a variety of downstream tasks. However, as LLMs grow in size, the memory demands for backpropagation become increasingly prohibitive. Zeroth-order (ZO) optimization methods offer a memory-efficient alternative by using forward passes to estimate gradients, but the variance of gradient estimates typically scales linearly with the model's parameter dimension\u2014a significant issue for LLMs. In this paper, we propose the random Subspace Zeroth-order (SubZero) optimization to address the challenges posed by LLMs' high dimensionality. We introduce a low-rank perturbation tailored for LLMs that significantly reduces memory consumption while improving training performance. Additionally, we prove that our gradient estimation closely approximates the backpropagation gradient, exhibits lower variance than traditional ZO methods, and ensures convergence when combined with SGD. Experimental results show that SubZero enhances fine-tuning performance and achieves faster convergence compared to standard ZO approaches like MeZO across various language modeling tasks. The source code will be released publicly.", "keywords": [ "Zeroth-order optimization", "Large Language Models (LLMs)", "fine-tuning", "random subspace" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 1, 1, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "FJv8VMPxWi", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Matthew X. Burns", "gender": "Male", "institution": "University of Rochester", "country": "", "position": "PhD student" }, { "name": "Michael Huang", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "Full Professor" }, { "name": "Qingyuan Hou", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" } ], "n_formula": 117, "n_formula_1": 25, "n_ref_uni": 72, "n_ref": 178, "n_ref_all": 188, "n_fig": 5, "n_tab": 5, "L_tab": 706, "n_element_tab": 88, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2685, "n_element_tab_1": 32, "formula_len_all": 12493, "formula_len_all_1": 1606, "len_all": 223482, "len_all_1": 66426, "len_abs": 1906, "len_title": 119, "len_sents": 51852, "len_sents_1": 28433, "n_sents": 430, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1310, "title": "Provable Convergence Bounds for Hybrid Dynamical Sampling and Optimization", "abs": "Analog dynamical accelerators (DXs) are a growing sub-field in computer architecture research, offering order-of-magnitude gains in power efficiency and latency over traditional digital methods in several machine learning, optimization, and sampling tasks. However, limited-capacity accelerators require hybrid analog/digital algorithms to solve real-world problems, commonly using large-neighborhood local search (LNLS) frameworks. Unlike fully digital algorithms, hybrid LNLS has no non-asymptotic convergence guarantees and no principled hyperparameter selection schemes, particularly limiting cross-device training and inference.\n\n\nIn this work, we provide non-asymptotic convergence guarantees for hybrid LNLS by reducing to block Langevin Diffusion (BLD) algorithms.\nAdapting tools from classical sampling theory, we prove exponential KL-divergence convergence for randomized and cyclic block selection strategies using ideal DXs. With finite device variation, we provide explicit bounds on the 2-Wasserstein bias in terms of step duration, noise strength, and function parameters. Our BLD model provides a key link between established theory and novel computing platforms, and our theoretical results provide a closed-form expression linking device variation, algorithm hyperparameters, and performance.", "keywords": [ "langevin", "accelerators", "sampling", "optimization", "diffusion", "analog computing" ], "rating_list": [ 3, 8, 8, 6, 8 ], "soundness_list": [ 2, 3, 4, 3, 2 ], "presentation_list": [ 1, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 2, 1, 2, 1, 2 ] }, { "paper_id": "FJFVmeXusW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Wayne Xiong", "gender": "Male", "institution": "Research, Microsoft", "country": "", "position": "Principal Researcher" }, { "name": "Wen Xiao", "gender": "Female", "institution": "Microsoft", "country": "CA", "position": "Researcher" }, { "name": "Yu Fu", "gender": "Male", "institution": "University of California, Riverside", "country": "US", "position": "PhD student" }, { "name": "Yue Dong", "gender": "Female", "institution": "University of California, Riverside", "country": "", "position": "Assistant Professor" }, { "name": "Zefan Cai", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Abed Asi", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 33, "n_ref": 105, "n_ref_all": 138, "n_fig": 10, "n_tab": 7, "L_tab": 3464, "n_element_tab": 248, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2564, "n_element_tab_1": 76, "formula_len_all": 320, "formula_len_all_1": 285, "len_all": 144318, "len_all_1": 53963, "len_abs": 1216, "len_title": 150, "len_sents": 40219, "len_sents_1": 23313, "n_sents": 265, "n_sents_1": 155, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 102, "L_abs": 1217, "title": "Not All Heads Matter: A Head-Level KV Cache Compression Method with Integrated Retrieval and Reasoning", "abs": "Key-Value (KV) caching is a common technique to enhance the computational efficiency of Large Language Models (LLMs), but its memory overhead grows rapidly with input length. Prior work has shown that not all tokens are equally important for text generation, proposing layer-level KV cache compression to selectively retain key information. Recognizing the distinct roles of attention heads in generation, we propose HeadKV, a head-level KV cache compression method, and HeadKV-R2, which leverages a novel contextual reasoning ability estimation for compression. Our approach operates at the level of individual heads, estimating their importance for contextual QA tasks that require both retrieval and reasoning capabilities. Extensive experiments across diverse benchmarks (LongBench, LooGLE), model architectures (e.g., Llama-3-8B-Instruct, Mistral-7B-Instruct), and long-context abilities tests demonstrate that our head-level KV cache compression significantly outperforms strong baselines, particularly in low-resource settings (KV size = 64 & 128). Notably, our method retains just 1.5% of the KV cache while achieving 97% of the performance of the full KV cache on the contextual question answering benchmark.", "keywords": [ "Key-Value cache", "Contextual reasoning", "Efficiency inference", "Large-Lauguage Model" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 4, 4 ], "presentation_list": [ 4, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "FJ8Q11j3p0", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Atabak Dehban", "gender": "Male", "institution": "Instituto Superior T\u00e9cnico", "country": "PT", "position": "Postdoc" }, { "name": "Jose Santos-Victor", "gender": "Male", "institution": "Instituto Superior T\u00e9cnico", "country": "", "position": "Full Professor" }, { "name": "Manuel Serra Nunes", "gender": "Male", "institution": "Instituto Superior T\u00e9cnico", "country": "", "position": "PhD student" }, { "name": "Yiannis Demiris", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 46, "n_ref": 69, "n_ref_all": 99, "n_fig": 13, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 530, "formula_len_all_1": 448, "len_all": 139089, "len_all_1": 60155, "len_abs": 1607, "len_title": 76, "len_sents": 41855, "len_sents_1": 31058, "n_sents": 263, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1612, "title": "Ego-Foresight: Self-supervised Agent Visuomotor Prediction for Efficient RL", "abs": "Despite the significant advancements in Deep Reinforcement Learning (RL) observed in the last decade, the amount of training experience necessary to learn effective policies remains one of the primary concerns both in simulated and real environments. Looking to solve this issue, previous work has shown that improved training efficiency can be achieved by separately modeling agent and environment, but usually requiring a supervisory agent mask. In contrast to RL, humans can perfect a new skill from a very small number of trials and in most cases do so without a supervisory signal, making neuroscientific studies of human development a valuable source of inspiration for RL. In particular, we explore the idea of motor prediction, which states that humans develop an internal model of themselves and of the consequences that their motor commands have on the immediate sensory inputs. Our insight is that the movement of the agent provides a cue that allows the duality between agent and environment to be learned. To instantiate this idea, we present Ego-Foresight, a self supervised method for disentangling agent and environment based on motion and prediction. Our main finding is that visuomotor prediction of the agent provides good feature representations for the underlying RL algorithm. To test our approach, we integrate Ego-Foresight with a model-free RL algorithm to solve simulated robotic manipulation tasks, showing its ability to improve efficiency and performance in different tasks while making strides towards real-world RL applications, by removing the need for costly supervisory signals.", "keywords": [ "Reinforcement Learning", "Robotics", "Prediction", "Disentangled Representations" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 1, 1 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "FJ6p5PaHFF", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Adrian Ciotinga", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Undergrad student" }, { "name": "YooJung Choi", "gender": "unknown", "institution": "Arizona State University", "country": "", "position": "Assistant Professor" } ], "n_formula": 32, "n_formula_1": 13, "n_ref_uni": 27, "n_ref": 49, "n_ref_all": 68, "n_fig": 8, "n_tab": 2, "L_tab": 341, "n_element_tab": 39, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3739, "formula_len_all_1": 1272, "len_all": 135586, "len_all_1": 61667, "len_abs": 1021, "len_title": 92, "len_sents": 52955, "len_sents_1": 31070, "n_sents": 343, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 44, "L_abs": 1024, "title": "Optimal Transport for Probabilistic Circuits", "abs": "We introduce a novel optimal transport framework for probabilistic circuits (PCs). While it has been shown recently that divergences between distributions represented as certain classes of PCs can be computed tractably, to the best of our knowledge, there is no existing approach to compute the Wasserstein distance between probability distributions given by PCs. We consider a Wasserstein-type distance that restricts the coupling measure of the associated optimal transport problem to be a probabilistic circuit. We then develop an algorithm for computing this distance by solving a series of small linear programs and derive the circuit conditions under which this is tractable. Furthermore, we show that we can also retrieve the optimal transport plan between the PCs from the solutions to these linear programming problems. We then consider the empirical Wasserstein distance between a PC and a dataset, and show that we can estimate the PC parameters to minimize this distance through an efficient iterative algorithm.", "keywords": [ "Probabilistic circuits", "Wasserstein", "Optimization", "Learning" ], "rating_list": [ 8, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "FIj9IEPCKr", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Jianwei Dr. Zhang", "gender": "Male", "institution": "Universit\u00e4t Hamburg", "country": "", "position": "Full Professor" }, { "name": "Mao Ye", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Song Tang", "gender": "Male", "institution": "University of Shanghai for Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Wenxin Su", "gender": "Female", "institution": "University of Shanghai for Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Xiatian Zhu", "gender": "unknown", "institution": "University of Surrey", "country": "GB", "position": "Associate Professor" }, { "name": "Yan Gan", "gender": "Male", "institution": "Chongqing University", "country": "CN", "position": "Postdoc" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 54, "n_ref": 118, "n_ref_all": 151, "n_fig": 7, "n_tab": 24, "L_tab": 14406, "n_element_tab": 1660, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 4195, "n_element_tab_1": 543, "formula_len_all": 803, "formula_len_all_1": 582, "len_all": 227484, "len_all_1": 74357, "len_abs": 1334, "len_title": 97, "len_sents": 49282, "len_sents_1": 29276, "n_sents": 421, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 49, "L_abs": 1357, "title": "Proxy Denoising for Source-Free Domain Adaptation", "abs": "Source-Free Domain Adaptation (SFDA) aims to adapt a pre-trained source model to an unlabeled target domain with no access to the source data. Inspired by the success of large Vision-Language (ViL) models in many applications, the latest research has validated ViL's benefit for SFDA by using their predictions as pseudo supervision. However, we observe that ViL's supervision could be noisy and inaccurate at an unknown rate, potentially introducing additional negative effects during adaption. To address this thus-far ignored challenge, we introduce a novel Proxy Denoising (__ProDe__) approach. The key idea is to leverage the ViL model as a proxy to facilitate the adaptation process towards the latent domain-invariant space. Concretely, we design a proxy denoising mechanism to correct ViL's predictions. This is grounded on a proxy confidence theory that models the dynamic effect of proxy's divergence against the domain-invariant space during adaptation. To capitalize the corrected proxy, we further derive a mutual knowledge distilling regularization. Extensive experiments show that ProDe significantly outperforms the current state-of-the-art alternatives under both conventional closed-set setting and the more challenging open-set, partial-set, generalized SFDA, multi-target, multi-source, and test-time settings. Our code will be released.", "keywords": [ "Domain adaptation", "source-free", "multimodal proxy space", "proxy confidence theory" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "FIXk0RP960", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Aohan Zeng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Hongning Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Tang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Minlie Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Pengfan DU", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Xiao Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yilin Niu", "gender": "Male", "institution": "Department of Computer Science and Techology, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Yuxiao Dong", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhengxiao Du", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Zhenyu Hou", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 34, "n_ref": 68, "n_ref_all": 90, "n_fig": 11, "n_tab": 4, "L_tab": 2735, "n_element_tab": 168, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 1710, "n_element_tab_1": 152, "formula_len_all": 3201, "formula_len_all_1": 3191, "len_all": 123473, "len_all_1": 70942, "len_abs": 1016, "len_title": 115, "len_sents": 34149, "len_sents_1": 31803, "n_sents": 222, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1032, "title": "Does RLHF Scale? Exploring the Effects of Data, Model, and Method", "abs": "This study explores the scaling properties of Reinforcement Learning from Human Feedback (RLHF) in Large Language Models (LLMs). \nAlthough RLHF is considered an important step in the post-training of LLMs, its scaling potential is still largely unknown. \nWe systematically analyze key components in the RLHF framework\u2014model size, data composition, and inference budget\u2014and their impacts on performance.\nOur findings show that increasing data diversity and volume improves reward model performance, helping process-supervision models scale better. \nFor policy training, more response samples per prompt boost performance initially but quickly plateau. \nAnd larger reward models offer modest gains in policy training. \nIn addition, larger policy models benefit less from RLHF with a fixed reward model. \nOverall, RLHF scales less efficiently than pretraining, with diminishing returns from additional computational resources.\nBased on these observations, we propose strategies to optimize RLHF performance within computational limits.", "keywords": [ "Language model", "Reinforcement learning from human feedback", "Scaling" ], "rating_list": [ 8, 6, 3, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 1, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "FI45zMai6Y", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Jialin Liu", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" }, { "name": "Kun Yuan", "gender": "Male", "institution": "Peking University", "country": "", "position": "Assistant Professor" }, { "name": "Qiulin Shang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Xinmeng Huang", "gender": "Male", "institution": "University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Yutong He", "gender": "Male", "institution": "Academy for Advanced Interdisciplinary Studies, Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 95, "n_formula_1": 53, "n_ref_uni": 42, "n_ref": 76, "n_ref_all": 168, "n_fig": 19, "n_tab": 2, "L_tab": 1083, "n_element_tab": 79, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 2149, "n_element_tab_1": 53, "formula_len_all": 9748, "formula_len_all_1": 4686, "len_all": 187693, "len_all_1": 68261, "len_abs": 1703, "len_title": 132, "len_sents": 47593, "len_sents_1": 25142, "n_sents": 431, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1386, "title": "A Mathematics-Inspired Learning-to-Optimize Framework for Decentralized Optimization", "abs": "Most decentralized optimization algorithms are handcrafted. While endowed with strong theoretical guarantees, these algorithms generally target a broad class of problems, thereby not being adaptive or customized to specific problem features. This paper studies data-driven decentralized algorithms trained to exploit problem features to boost convergence. Existing learning-to-optimize methods typically suffer from poor generalization or prohibitively vast search spaces. In addition, they face more challenges in decentralized settings where nodes must reach consensus through neighborhood communications without global information. To resolve these challenges, this paper first derives the necessary conditions that successful decentralized algorithmic rules need to satisfy to achieve both optimality and consensus. Based on these conditions, we propose a novel **M**athematics-**i**nspired **L**earning-to-**o**ptimize framework for **D**ecentralized **o**ptimization (**MiLoDo**). Empirical results demonstrate that MiLoDo-trained algorithms outperform handcrafted algorithms and exhibit strong generalizations. Algorithms learned via MiLoDo in 100 iterations perform robustly when running 100,000 iterations during inferences. Moreover, MiLoDo-trained algorithms on synthetic datasets perform well on problems involving real data, higher dimensions, and different loss functions.", "keywords": [ "Learning to Optimize", "Decentralized Optimization", "Composite Optimization" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "FHtHH4ulEQ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Amrita Saha", "gender": "Female", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Doyen Sahoo", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Researcher" }, { "name": "Dunjie Lu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Intern" }, { "name": "Junli Wang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" }, { "name": "Tao Yu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Tianbao Xie", "gender": "Male", "institution": "the University of Hong Kong, University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Yiheng Xu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Zekun Wang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 26, "n_ref": 93, "n_ref_all": 119, "n_fig": 3, "n_tab": 14, "L_tab": 5594, "n_element_tab": 581, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 2729, "n_element_tab_1": 364, "formula_len_all": 0, "formula_len_all_1": 23, "len_all": 228669, "len_all_1": 65702, "len_abs": 1519, "len_title": 114, "len_sents": 67580, "len_sents_1": 30167, "n_sents": 483, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1532, "title": "Aguvis: Unified Pure Vision Agents for Autonomous GUI Interaction", "abs": "Graphical User Interfaces (GUIs) are critical to human-computer interaction, yet automating GUI tasks remains challenging due to the complexity and variability of visual environments. \nExisting approaches often rely on textual representations of GUIs, which introduce limitations in generalization, efficiency, and scalability. \nIn this paper, we introduce Aguvis, a unified pure vision-based framework for autonomous GUI agents that operates across various platforms. \nOur approach leverages image-based observations, and grounding instructions in natural language to visual elements, and employs a consistent action space to ensure cross-platform generalization. \nTo address the limitations of previous work, we integrate explicit planning and reasoning within the model, enhancing its ability to autonomously navigate and interact with complex digital environments. \nWe construct a large-scale dataset of GUI agent trajectories, incorporating multimodal reasoning and grounding, and employ a two-stage training pipeline that first focuses on general GUI grounding, followed by planning and reasoning. \nThrough comprehensive experiments, we demonstrate that Aguvis surpasses previous state-of-the-art methods in both offline and real-world online scenarios, achieving, to our knowledge, the first fully autonomous pure vision GUI agent capable of performing tasks independently without collaboration with external closed-source models. \nWe will open-source all datasets, models, and training recipes to facilitate future research.", "keywords": [ "GUI Agent", "Visual Language Model", "Large Language Model", "Grounding", "Planning" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "FHsaa6lZMp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Peter Gerstoft", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Zheng Zhou", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Zhongping Zhang", "gender": "Male", "institution": "Boston University", "country": "US", "position": "PhD student" }, { "name": "Bryan Allen Plummer", "gender": "Male", "institution": "Boston University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 24, "n_formula_1": 8, "n_ref_uni": 30, "n_ref": 117, "n_ref_all": 135, "n_fig": 6, "n_tab": 3, "L_tab": 1420, "n_element_tab": 190, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1150, "n_element_tab_1": 177, "formula_len_all": 1687, "formula_len_all_1": 418, "len_all": 167859, "len_all_1": 76359, "len_abs": 2321, "len_title": 93, "len_sents": 52161, "len_sents_1": 33995, "n_sents": 419, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1647, "title": "Fine-Grained Machine-Generated Text Detection", "abs": "Machine-Generated Text (MGT) detection identifies whether a given text is human-written or machine-generated. However, this can result in detectors that would flag paraphrased or translated text as machine-generated. Fine-grained classification that separates the different types of machine text is valuable in real-world applications, as different types of MGT convey distinct implications. For example, machine-generated articles are more likely to contain misinformation, whereas paraphrased and translated texts may improve understanding of human-written text. Despite this benefit, existing studies consider this a binary classification task, either overlooking machine-paraphrased and machine-translated text entirely or simply grouping all machine-processed text into one category. To address this shortcoming, this paper provides an in-depth study of fine-grained MGT detection, categorizing input text into four classes: human-written, machine-generated, machine-paraphrased, and machine-translated. A key challenge is the performance drop on out-of-domain texts due to the variability in text generators, especially for translated or paraphrased text. We introduce a RoBERTa-based Mixture of Detectors (RoBERTa-MoD), which leverages multiple domain-optimized detectors for more robust and generalized performance. We offer theoretical proof that our method outperforms a single detector, and experimental findings demonstrate a 5--9\\% improvement in mean Average Precision (mAP) over prior work on six diverse datasets: GoodNews, VisualNews, WikiText, Essay, WP, and Reuters. Our code and data will be publicly released upon acceptance.", "keywords": [ "Machine-Generated Text Detection", "Fine-grained Classification", "Mixture of Experts" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "FHQDCQFD8y", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Hengyi Shao", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "MS student" }, { "name": "Lei Li", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Full Professor" }, { "name": "Liang Dong", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Lin Zhang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 26, "n_ref": 36, "n_ref_all": 47, "n_fig": 6, "n_tab": 7, "L_tab": 3461, "n_element_tab": 148, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3019, "n_element_tab_1": 97, "formula_len_all": 116, "formula_len_all_1": 116, "len_all": 91609, "len_all_1": 56047, "len_abs": 1203, "len_title": 155, "len_sents": 21261, "len_sents_1": 19053, "n_sents": 159, "n_sents_1": 143, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1209, "title": "Grad-TopoCAM: EEG Brain Region Visual Interpretability via Gradient-Based Topographic Class Activation Map", "abs": "The visualization and interpretability of electroencephalogram (EEG) decoding significantly contribute to brain-computer interfaces (BCI) and cognitive neuroscience. Although some existing research has attempted to map EEG features to specific brain regions, these approaches fail to fully utilize raw signals and lack extensibility to other Deep Learning (DL) models. In this work, Grad-TopoCAM (Gradient-Based Topographic Class Activation Map) is proposed, which enhances interpretability in DL models for EEG decoding adaptively. Grad-TopoCAM calculates the gradient of feature maps for the target class at the target layer. The weights of the feature maps are obtained through global average pooling of the gradients. The class activation map is generated by performing a linear combination of weights and feature maps, which is subsequently mapped to different brain regions. Grad-TopoCAM is validated across eight DL models on four public datasets. Experimental results indicate that Grad-TopoCAM effectively identifies and visualizes brain regions that significantly influence decoding outcomes, while also facilitating channel selection for different decoding tasks. The code and data are open-source.", "keywords": [ "Electroencephalogram", "Class Activation Map", "Deep Learning", "Visualization", "Interpretability" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 1, 2, 3, 3 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "FH4x8IqUu6", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Dongheng Lin", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Han Hu", "gender": "Female", "institution": "University of Birmingham", "country": "GB", "position": "PhD student" }, { "name": "Jianbo Jiao", "gender": "unknown", "institution": "University of Birmingham", "country": "GB", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 1, "n_ref_uni": 52, "n_ref": 89, "n_ref_all": 133, "n_fig": 35, "n_tab": 8, "L_tab": 3948, "n_element_tab": 381, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 2189, "n_element_tab_1": 287, "formula_len_all": 270, "formula_len_all_1": 70, "len_all": 198484, "len_all_1": 56872, "len_abs": 1225, "len_title": 122, "len_sents": 61727, "len_sents_1": 25138, "n_sents": 429, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1183, "title": "What Time Tells Us? Time-Aware Representation Learning from Static Images", "abs": "Time becomes visible through changes in what we see, as daylight fades and shadows grow. Inspired by this, in this paper we explore the potential to learn time-aware representations from static images, trying to answer: *what time tells us?* To this end, we first introduce a Time-Oriented Collection (TOC) dataset, which contains 130,906 images with reliable timestamps. Leveraging this dataset, we propose a Time-Image Contrastive Learning (TICL) approach to jointly model timestamp and related visual representations through cross-modal contrastive learning. We found that the proposed TICL, 1) not only achieve state-of-the-art performance on the timestamp estimation task, over various benchmark metrics, 2) but also, interestingly, though only seeing static images, the representations learned by TICL show strong capability in several time-aware downstream tasks such as time-based image retrieval, video scene classification, and time-aware image editing. Our findings confirm that time-aware visual representations are learnable from static images and beneficial for various vision tasks, laying a foundation for future research on understanding time-related visual context.", "keywords": [ "Representation Learning", "Dataset", "Cross-modal", "Time" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 4 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 5, 5 ] }, { "paper_id": "FGd9mXHhM5", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Aymeric Dieuleveut", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "FR", "position": "Full Professor" }, { "name": "Hadrien Hendrikx", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Researcher" }, { "name": "Renaud Gaucher", "gender": "Male", "institution": "\u00c9cole Polytechnique", "country": "FR", "position": "PhD student" } ], "n_formula": 97, "n_formula_1": 27, "n_ref_uni": 30, "n_ref": 74, "n_ref_all": 97, "n_fig": 5, "n_tab": 1, "L_tab": 234, "n_element_tab": 56, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 672, "n_element_tab_1": 7, "formula_len_all": 9080, "formula_len_all_1": 1474, "len_all": 203940, "len_all_1": 86883, "len_abs": 3740, "len_title": 102, "len_sents": 61331, "len_sents_1": 35241, "n_sents": 555, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 985, "title": "Achieving Optimal Breakdown for Byzantine-Robust Gossip", "abs": "Distributed approaches have many computational benefits, but they are vulnerable to attacks from a subset of devices transmitting incorrect information. This paper investigates Byzantine-resilient algorithms in a decentralized setting, where devices communicate directly with one another. \nWe investigate the notion of breakdown point and show an upper bound on the number of adversaries that decentralized algorithms can tolerate. We introduce an algorithmic framework that recovers ClippedGossip and NNA, two popular approaches for robust decentralized learning, as special cases. This framework allows us to generalize NNA to sparse graph, and introduce CG+, which is at the intersection of the two. Our unified analysis framework gives near-optimal guarantees for CG+ (and other approaches with additional assumptions). Experimental evidence validates the effectiveness of CG+ and the gap with NNA, in particular against a novel attack tailored to sparse graphs that we introduce.", "keywords": [ "Byzantine", "Robustness", "Decentralized", "Gossip", "Averaging", "SGD" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "FGSgsefE0Y", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Huanran Hu", "gender": "Male", "institution": "China Agricultural University", "country": "CN", "position": "Undergrad student" }, { "name": "Lei Wang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Shengjie Jin", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Xu Chen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Yanqi Dai", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Zhiwu Lu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 22, "n_ref": 73, "n_ref_all": 98, "n_fig": 10, "n_tab": 12, "L_tab": 3532, "n_element_tab": 227, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 3212, "n_element_tab_1": 218, "formula_len_all": 583, "formula_len_all_1": 168, "len_all": 208425, "len_all_1": 67265, "len_abs": 1263, "len_title": 94, "len_sents": 67827, "len_sents_1": 30252, "n_sents": 542, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1307, "title": "MMRole: A Comprehensive Framework for Developing and Evaluating Multimodal Role-Playing Agents", "abs": "Recently, Role-Playing Agents (RPAs) have garnered increasing attention for their potential to deliver emotional value and facilitate sociological research.\nHowever, existing studies are primarily confined to the textual modality, unable to simulate humans' multimodal perceptual capabilities.\nTo bridge this gap, we introduce the concept of Multimodal Role-Playing Agents (MRPAs), and propose a comprehensive framework, MMRole, for their development and evaluation, which comprises a personalized multimodal dataset and a robust evaluation approach.\nSpecifically, we construct a large-scale, high-quality dataset, MMRole-Data, consisting of 85 characters, 11K images, and 14K single or multi-turn dialogues.\nAdditionally, we present a robust evaluation approach, MMRole-Eval, encompassing eight metrics across three dimensions, where a reward model is designed to score MRPAs with the constructed ground-truth data for comparison.\nMoreover, we develop the first specialized MRPA, MMRole-Agent.\nExtensive evaluation results demonstrate the improved performance of MMRole-Agent and highlight the primary challenges in developing MRPAs, emphasizing the need for enhanced multimodal understanding and role-playing consistency.\nThe data, code, and models are all available at https://github.com/YanqiDai/MMRole.", "keywords": [ "Multimodal Role-Playing Agents", "Large Multimodal Models" ], "rating_list": [ 8, 5, 8, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "FGMkSL8NR0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Lifu Huang", "gender": "Male", "institution": "University of California, Davis", "country": "US", "position": "Assistant Professor" }, { "name": "Parisa Kordjamshidi", "gender": "Female", "institution": "Michigan State University", "country": "US", "position": "Associate Professor" }, { "name": "Ying Shen", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Yue Zhang", "gender": "unknown", "institution": "Michigan State University", "country": "US", "position": "PhD student" }, { "name": "Zhiyang Xu", "gender": "Male", "institution": "Virginia Polytechnic Institute and State University", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 37, "n_ref": 84, "n_ref_all": 110, "n_fig": 13, "n_tab": 14, "L_tab": 2646, "n_element_tab": 255, "n_fig_1": 8, "n_tab_1": 10, "L_tab_1": 2574, "n_element_tab_1": 242, "formula_len_all": 226, "formula_len_all_1": 226, "len_all": 173200, "len_all_1": 78533, "len_abs": 1619, "len_title": 126, "len_sents": 52229, "len_sents_1": 33801, "n_sents": 453, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1033, "title": "SPARTUN3D: Situated Spatial Understanding of 3D World in Large Language Model", "abs": "Integrating the 3D world into large language models (3D-based LLMs) has been a promising research direction for 3D scene understanding. However, current 3D-based LLMs fall short in situated understanding due to two key limitations: 1) existing 3D datasets are constructed from a global perspective of the 3D scenes and lack situated context.\n2) the architectures of the current 3D-based LLMs lack an explicit mechanism for aligning situated spatial information between 3D representations and natural language, limiting their performance in tasks requiring precise spatial reasoning. \nIn this work, we address these issues by introducing a scalable situated 3D dataset, named Spartun3D, that incorporates various situated spatial information.\nIn addition, we propose a situated spatial alignment module to enhance the learning between 3D visual representations and their corresponding textual descriptions. Our experimental results demonstrate that both our dataset and alignment module enhance situated spatial understanding ability.", "keywords": [ "Situated Understanding in 3D Scen", "3D VL", "LLM" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "FGLnLjtemf", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Guodong Xu", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "", "position": "MS student" }, { "name": "Yue Wu", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Associate Professor" }, { "name": "Li Hao", "gender": "Male", "institution": "Xidian University", "country": "", "position": "Associate Professor" }, { "name": "Gong Maoguo", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 15, "n_ref_uni": 27, "n_ref": 46, "n_ref_all": 61, "n_fig": 2, "n_tab": 3, "L_tab": 1530, "n_element_tab": 171, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 417, "n_element_tab_1": 51, "formula_len_all": 406, "formula_len_all_1": 407, "len_all": 106678, "len_all_1": 62093, "len_abs": 1432, "len_title": 140, "len_sents": 32590, "len_sents_1": 28063, "n_sents": 265, "n_sents_1": 227, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1437, "title": "Manipulating Infrared Emissivity with Galvanized Iron Sheets for Physical Adversarial Attack", "abs": "For adversarial attacks on infrared detectors, previous works have focused on designing the physical patches through temperature variations, overlooking the impact of infrared emissivity on infrared imaging. In fact, infrared emissivity significantly affects infrared radiant intensity at the same temperature. In this paper, a QR-like adversarial attack patch is designed by manipulating the surface emissivity of objects to alter the infrared radiation intensity emitted from the object's surface, called Emissivity QR-like Patch (E-QR patch). In this paper, the surface emissivity of the object is manipulated through the adjustment of surface roughness. Various levels of surface roughness are realized by a commonly used metal material, galvanized iron sheets, to produce physically adversarial patches with diverse infrared radiation intensity. Considering the possible transformation distributions between the digital and physical domains, a physical E-QR patch, which is robust to noise, angle, and position, is generated by an expectation over the transformation framework. Smoothing loss is incorporated to minimize the loss in physical reconstruction, thereby effectively mitigating shooting errors in the physical domain induced by abrupt pixel changes in the digital domain. Experimental results show that the E-QR patch achieves more than 80% attack success rate for infrared pedestrian detectors in a physical environment.", "keywords": [ "Adversarial Patch", "Deep neural network", "Physical sample generation" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 2 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "FGIBKpOj8m", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "En Zhu", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Full Professor" }, { "name": "Jiaxin Zhang", "gender": "Female", "institution": "National University of Defense Technology", "country": "", "position": "PhD student" }, { "name": "Nianhao Xie", "gender": "Female", "institution": "National University of Defense Technology", "country": "", "position": "PhD student" }, { "name": "Siwei Wang", "gender": "Male", "institution": "Intelligent Game and Decision Lab", "country": "CN", "position": "Assistant Professor" }, { "name": "Xinwang Liu", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yiqi Wang", "gender": "unknown", "institution": "Michigan State University", "country": "", "position": "PhD student" }, { "name": "Yu Shi", "gender": "Male", "institution": "National University of Defense Technology", "country": "", "position": "MS student" }, { "name": "Yusong Tan", "gender": "Male", "institution": "National University of Defense Technology", "country": "", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 28, "n_ref": 73, "n_ref_all": 118, "n_fig": 7, "n_tab": 18, "L_tab": 11729, "n_element_tab": 955, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 3036, "n_element_tab_1": 351, "formula_len_all": 48, "formula_len_all_1": 0, "len_all": 191188, "len_all_1": 72919, "len_abs": 7982, "len_title": 150, "len_sents": 59792, "len_sents_1": 33190, "n_sents": 485, "n_sents_1": 247, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1261, "title": "Towards the Effect of Large Language Models on Out-Of-Distribution Challenge in Text-Attributed Graphs", "abs": "Text-Attributed Graphs (TAGs), where each node is associated with text attributes, are ubiquitous and have been widely applied in the real world. The Out-Of-Distribution (OOD) issue, i.e., the training data and the test data not from the same distribution, is quite common in learning on real-world TAGs, posing significant challenges to the effectiveness of graph learning models. Recently, Large Language Models (LLMs) have shown extraordinary capability in processing text data, and have demonstrated tremendous potential in handling TAGs. However, there is no benchmark work that systematically and comprehensively investigates the effect of these LLM-based methods on alleviating the OOD issue on TAGs. To bridge this gap, we first develop OOD-TAG, a comprehensive OOD benchmark dataset in TAGs which consists of diverse distributions. Meanwhile, we conduct a systematic and comprehensive investigation on OOD-TAG with different LLM pipelines for graphs. In addition, we provide original observations and novel insights based on the empirical study, which can suggest promising directions for the research of LLMs in addressing the OOD challenges on TAGs. Our code and dataset are available in https://anonymous.4open.science/r/GraphOOD-benchmark-5FCF/.", "keywords": [ "Out-Of-Distribution", "Large Language Models", "Text-Attributed-Graphs" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 4, 2, 4 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 3, 5, 3, 5 ] }, { "paper_id": "FFwoaUFBVC", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Michael Munn", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Susan Wei", "gender": "Female", "institution": "Monash University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 42, "n_formula_1": 18, "n_ref_uni": 41, "n_ref": 66, "n_ref_all": 97, "n_fig": 3, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2266, "formula_len_all_1": 1348, "len_all": 152175, "len_all_1": 75265, "len_abs": 1090, "len_title": 126, "len_sents": 49351, "len_sents_1": 31410, "n_sents": 367, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1093, "title": "Leveraging free energy in pretraining model selection for improved fine-tuning", "abs": "Recent advances in artificial intelligence have been fueled by the development of foundation models such as BERT, GPT, T5, and Vision Transformers. These models are first pretrained on vast and diverse datasets and then adapted to specific downstream tasks, often with significantly less data. However, the mechanisms behind the success of this ubiquitous pretrain-then-adapt paradigm remain underexplored, particularly the characteristics of pretraining checkpoints that lend themselves to good downstream adaptation. We introduce a Bayesian model selection criterion, called the downstream free energy, which quantifies a checkpoint's adaptability by measuring the concentration of nearby favorable parameters for the downstream task. We demonstrate that this free energy criterion can be effectively implemented without access to the downstream data or prior knowledge of the downstream task. Furthermore, we provide empirical evidence that the free energy criterion reliably correlates with improved fine-tuning performance, offering a principled approach to predicting model adaptability.", "keywords": [ "transfer learning", "free energy", "Bayesian model selection", "efficient fine-tuning", "adaptation" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 3, 1, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 1, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "FFUmPQM8c5", "primary_area": "datasets and benchmarks", "n_author": 4, "author_info_list": [ { "name": "Aapo Hakala", "gender": "Male", "institution": "Tampere University", "country": "", "position": "MS student" }, { "name": "Irene Mart\u00edn-Morat\u00f3", "gender": "Female", "institution": "Tampere University", "country": "FI", "position": "Postdoc" }, { "name": "Parthasaarathy Sudarsanam", "gender": "Male", "institution": "Tampere University", "country": "", "position": "PhD student" }, { "name": "Tuomas Virtanen", "gender": "unknown", "institution": "Tampere University", "country": "", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 24, "n_ref": 29, "n_ref_all": 59, "n_fig": 8, "n_tab": 14, "L_tab": 2502, "n_element_tab": 279, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 1418, "n_element_tab_1": 170, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 148280, "len_all_1": 62683, "len_abs": 1714, "len_title": 111, "len_sents": 47327, "len_sents_1": 28126, "n_sents": 401, "n_sents_1": 215, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1721, "title": "AVCAPS: AN AUDIO-VISUAL DATASET WITH MODALITY-SPECIFIC CAPTIONS", "abs": "In this paper, we introduce AVCaps, an audio-visual captioning dataset that contains separate textual captions for the audio, visual, and audio-visual contents of video clips. The dataset contains 2061 video clips constituting a total of 28.8 hours. We provide up to 5 captions for the audio, visual, and audio-visual content of each clip, crowdsourced separately. Existing datasets focus on a single modality or do not provide modality-specific captions, limiting the study of how each modality contributes to overall comprehension in multimodal settings. Our dataset addresses this critical gap in multimodal research by offering a resource for studying how audio and visual content are captioned individually, as well as how audio-visual content is captioned in relation to these individual modalities. To counter the bias observed in crowdsourced audio-visual captions, which often emphasize visual over audio content, we generated three audio-visual captions for each clip using our crowdsourced captions by leveraging existing large language models (LLMs). We present multimodal and crossmodal captioning and retrieval experiments to illustrate the effectiveness of modality-specific captions in evaluating model performance. Notably, we show that a model trained on LLM-generated audio-visual captions captures audio information more effectively, achieving 14% higher Sentence-BERT similarity on ground truth audio captions compared to a model trained on crowdsourced audio-visual captions. We also discuss the possibilities in multimodal representation learning, question answering, developing new video captioning metrics, and generative AI that this dataset unlocks. The dataset will be freely available online.", "keywords": [ "Audio-visual dataset", "captioning dataset", "Multimodal learning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "FEpAUnS7f7", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "BOLUN SUN", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Researcher" }, { "name": "Haiyun Jiang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Yifan Zhou", "gender": "Male", "institution": "University of Georgia", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 48, "n_ref_all": 56, "n_fig": 2, "n_tab": 7, "L_tab": 1136, "n_element_tab": 93, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 190, "n_element_tab_1": 23, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 95522, "len_all_1": 43414, "len_abs": 1187, "len_title": 131, "len_sents": 30391, "len_sents_1": 21983, "n_sents": 203, "n_sents_1": 132, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1189, "title": "Empowering Users in Digital Privacy Management through Interactive LLM-Based Agents", "abs": "This paper presents a novel application of large language models (LLMs) to enhance user comprehension of privacy policies through an interactive dialogue agent. We demonstrate that LLMs significantly outperform traditional models in tasks like Data Practice Identification, Choice Identification, Policy Summarization, and Privacy Question Answering, setting new benchmarks in privacy policy analysis. Building on these findings, we introduce an innovative LLM-based agent that functions as an expert system for processing website privacy policies, guiding users through complex legal language without requiring them to pose specific questions. A user study with 100 participants showed that users assisted by the agent had higher comprehension levels (mean score of 2.6 out of 3 vs. 1.8 in the control group), reduced cognitive load (task difficulty ratings of 3.2 out of 10 vs. 7.8), increased confidence in managing privacy, and completed tasks in less time (5.5 minutes vs. 15.8 minutes). This work highlights the potential of LLM-based agents to transform user interaction with privacy policies, leading to more informed consent and empowering users in the digital services landscape.", "keywords": [ "LLM", "Agent", "Usable Privacy Policies", "Benchmarking", "HCI" ], "rating_list": [ 5, 3, 5, 5, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 4, 4 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 2, 5, 4, 4, 2 ] }, { "paper_id": "FEZOLWexPb", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 22, "author_info_list": [ { "name": "Ajinkya Pattekar", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Researcher" }, { "name": "Allison R Greenplate", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Amit Bar-Or", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Full Professor" }, { "name": "Benjamin A. Abramoff", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Benjamin A Fensterheim", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Postdoc" }, { "name": "Damian Maseda", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Principal Researcher" }, { "name": "Divij Mathew", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Postdoc" }, { "name": "Dokyoon Kim", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "Associate Professor" }, { "name": "Irene Khavin", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Researcher" }, { "name": "Jaesik Kim", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "John Wherry", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "Full Professor" }, { "name": "Jonghyun Lee", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Postdoc" }, { "name": "Matei Ionita", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "Researcher" }, { "name": "Matthew Eric Lee", "gender": "Male", "institution": "University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Michelle L. McKeague", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Researcher" }, { "name": "Rennie L. Rhee", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Shwetank", "gender": "unknown", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Researcher" }, { "name": "Sokratis Apostolidis", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Assistant Professor" }, { "name": "Victoria Fang", "gender": "Female", "institution": "University of Pennsylvania", "country": "", "position": "Instructor" }, { "name": "YONGHYUN NAM", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Postdoc" }, { "name": "Yidi Huang", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "", "position": "PhD student" }, { "name": "Zahabia Rangwala", "gender": "unknown", "institution": "", "country": "", "position": "" } ], "n_formula": 43, "n_formula_1": 17, "n_ref_uni": 28, "n_ref": 40, "n_ref_all": 52, "n_fig": 6, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2694, "formula_len_all_1": 976, "len_all": 157668, "len_all_1": 58709, "len_abs": 1288, "len_title": 111, "len_sents": 52302, "len_sents_1": 27746, "n_sents": 467, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1180, "title": "MAESTRO: Masked Encoding Set Transformer with Self-Distillation", "abs": "The interrogation of cellular states and interactions in immunology research is an ever-evolving task, requiring adaptation to the current levels of high dimensionality. Cytometry enables high-dimensional profiling of immune cells, but its analysis is hindered by the complexity and variability of the data. We present MAESTRO, a self-supervised set representation learning model that generates vector representations of set-structured data, which we apply to learn immune profiles from cytometry data. Unlike previous studies only learn cell-level representations, whereas MAESTRO uses all of a sample's cells to learn a set representation. MAESTRO leverages specialized attention mechanisms to handle sets of variable number of cells and ensure permutation invariance, coupled with an online tokenizer by self-distillation framework. We benchmarked our model against existing cytometry approaches and other existing machine learning methods that have never been applied in cytometry. Our model outperforms existing approaches in retrieving cell-type proportions and capturing clinically relevant features for downstream tasks such as disease diagnosis and immune cell profiling.", "keywords": [ "self-supervision", "representation learning", "immunology", "biology", "single-cell", "cytometry", "set", "set representations" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "FEDnzAhIT4", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Leonardo Cotta", "gender": "Male", "institution": "Vector Institute", "country": "", "position": "Postdoc" }, { "name": "Chris Maddison", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 9, "n_ref_uni": 51, "n_ref": 102, "n_ref_all": 130, "n_fig": 13, "n_tab": 6, "L_tab": 4584, "n_element_tab": 458, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1866, "formula_len_all_1": 692, "len_all": 173281, "len_all_1": 74128, "len_abs": 1428, "len_title": 106, "len_sents": 50825, "len_sents_1": 34497, "n_sents": 423, "n_sents_1": 276, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1434, "title": "Test-Time Fairness and Robustness in Large Language Models", "abs": "Frontier Large Language Models (LLMs) can be socially discriminatory or sensitive to spurious features of their inputs. Because only well-resourced corporations can train frontier LLMs, we need robust test-time strategies to control such biases. Existing solutions, which instruct the LLM to be fair or robust, rely on the model\u2019s implicit understanding of bias. Causality provides a rich formalism through which we can be explicit about our debiasing requirements. Yet, as we show, a naive application of the standard causal debiasing strategy, counterfactual data augmentation, fails under standard assumptions to debias predictions at an individual level at test time. To address this, we develop a stratified notion of debiasing called stratified invariance, which can capture a range of debiasing requirements from population level to individual level through an additional measurement that stratifies the predictions. We present a complete observational test for stratified invariance. Finally, we introduce a data augmentation strategy that guarantees stratified invariance at test time under suitable assumptions, together with a prompting strategy that encourages stratified invariance in LLMs. We show that our prompting strategy, unlike implicit instructions, consistently reduces the bias of frontier LLMs across a suite of synthetic and real-world benchmarks without requiring additional data, finetuning or pre-training.", "keywords": [ "large language models", "trustworthiness", "fairness", "robustness", "causality" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 2, 4, 5 ] }, { "paper_id": "FDsWd0NOB5", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Chris Bakal", "gender": "Male", "institution": "Institute of Cancer Research ", "country": "GB", "position": "Full Professor" }, { "name": "Chris Dunsby", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Full Professor" }, { "name": "Lucas G Dent", "gender": "Male", "institution": "Institute of Cancer Research", "country": "", "position": "Postdoc" }, { "name": "Matt De Vries", "gender": "Male", "institution": "Institute of Cancer Research", "country": "GB", "position": "PhD student" }, { "name": "Nathan Curry", "gender": "unknown", "institution": "Imperial College London", "country": "", "position": "Postdoc" }, { "name": "Olga Fourkioti", "gender": "Female", "institution": "The Institute of Cancer Research", "country": "GB", "position": "Postdoc" }, { "name": "Reed Naidoo", "gender": "Male", "institution": "The Institute of Cancer Research", "country": "GB", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 15, "n_ref_uni": 38, "n_ref": 65, "n_ref_all": 74, "n_fig": 4, "n_tab": 1, "L_tab": 383, "n_element_tab": 32, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 226, "n_element_tab_1": 36, "formula_len_all": 841, "formula_len_all_1": 843, "len_all": 133025, "len_all_1": 59978, "len_abs": 1623, "len_title": 134, "len_sents": 37945, "len_sents_1": 30408, "n_sents": 234, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1167, "title": "Build your own cell: Diffusion Models for Multichannel 3D Microscopy Image Generation", "abs": "Three-dimensional (3D) cellular morphology is a critical indicator of cellular function, disease states, and drug responses. However, capturing and interpreting the complex relationships between cell shape, treatment conditions, and their biological implications remains a challenge. To address this, we present \"Build Your Own Cell'' (BYOC), a multichannel 3D generative framework that combines vector quantisation and diffusion models to synthesise biologically realistic 3D cell structures. BYOC captures intricate morphological changes induced by different drug treatments, enabling high-throughput in silico simulations and screening of cell shapes in response to varied conditions. This novel framework represents a significant step towards accelerating pre-clinical drug development by synthesising high-resolution, biologically realistic 3D cells, potentially reducing reliance on labour-intensive experimental studies. By ensuring phenotypic consistency between cell and nucleus volumes through joint modelling, BYOC provides high-fidelity reconstructions that could facilitate downstream analyses, including drug efficacy evaluation and mechanistic studies.", "keywords": [ "3D Diffusion Models" ], "rating_list": [ 8, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "FDnZFpHmU4", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Han Wu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Jie Liu", "gender": "Male", "institution": "North China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Linqi Song", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Mingyang LIU", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Sichun Luo", "gender": "Male", "institution": "City University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Xiongwei Han", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "", "position": "MS student" }, { "name": "Zhijiang Guo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuxuan YAO", "gender": "unknown", "institution": "City University of Hong Kong", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 2, "n_ref_uni": 16, "n_ref": 68, "n_ref_all": 85, "n_fig": 5, "n_tab": 9, "L_tab": 1067, "n_element_tab": 106, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1380, "n_element_tab_1": 98, "formula_len_all": 10, "formula_len_all_1": 36, "len_all": 171903, "len_all_1": 66886, "len_abs": 1210, "len_title": 133, "len_sents": 42993, "len_sents_1": 30667, "n_sents": 329, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 85, "L_abs": 1260, "title": "Determine-Then-Ensemble: Necessity of Top-k Union for Large Language Model Ensembling", "abs": "Large language models (LLMs) exhibit varying strengths and weaknesses across different tasks, prompting recent studies to explore the benefits of ensembling models to leverage their complementary advantages. However, existing LLM ensembling methods often overlook model compatibility and struggle with inefficient alignment of probabilities across the entire vocabulary. In this study, we empirically investigate the factors influencing ensemble performance, identifying model performance, vocabulary size, and response style as key determinants, revealing that compatibility among models is essential for effective ensembling. This analysis leads to the development of a simple yet effective model selection strategy that identifies compatible models. Additionally, we introduce the \\textsc{Uni}on \\textsc{T}op-$k$ \\textsc{E}nsembling (\\textsc{UniTE}), a novel approach that efficiently combines models by focusing on the union of the top-k tokens from each model, thereby avoiding the need for full vocabulary alignment and reducing computational overhead. Extensive evaluations across multiple benchmarks demonstrate that \\textsc{UniTE} significantly enhances performance compared to existing methods, offering a more efficient framework for LLM ensembling.", "keywords": [ "Model ensembling", "LLM" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "FDmKe5EBuy", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Alex Beutel", "gender": "Male", "institution": "OpenAI", "country": "US", "position": "Researcher" }, { "name": "Johannes Heidecke", "gender": "unknown", "institution": "OpenAI", "country": "", "position": "Researcher" }, { "name": "Kai Yuanqing Xiao", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Lilian Weng", "gender": "Female", "institution": "OpenAI", "country": "", "position": "Research Scientist" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 44, "n_ref": 93, "n_ref_all": 113, "n_fig": 6, "n_tab": 4, "L_tab": 6474, "n_element_tab": 479, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 255, "formula_len_all_1": 213, "len_all": 177328, "len_all_1": 66712, "len_abs": 1571, "len_title": 148, "len_sents": 54375, "len_sents_1": 34274, "n_sents": 379, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1588, "title": "Diverse and Effective Red Teaming with Auto-generated Rewards and Multi-step Reinforcement Learning", "abs": "Automated red teaming can discover rare model failures and generate challenging examples that can be used for training or evaluation. However, a core challenge in automated red teaming is ensuring that the attacks are both diverse and effective. Prior methods typically succeed in optimizing either for diversity or for effectiveness, but rarely both. In this paper, we provide methods that enable automated red teaming to generate a large number of diverse and successful attacks.\n\nOur approach decomposes the task into two steps: (1) automated methods for generating diverse attack goals and (2) generating effective attacks for those goals. While we provide multiple straightforward methods for generating diverse goals, our key contributions are to train an RL attacker that both follows those goals and generates diverse attacks for those goals. First, we demonstrate that it is easy to use a large language model (LLM) to generate diverse attacker goals with per-goal prompts and rewards, including rule-based rewards (RBRs) to grade whether the attacks are successful for the particular goal. Second, we demonstrate how training the attacker model with multi-step RL, where the model is rewarded for generating attacks that are different from past attempts further increases diversity while remaining effective. We use our approach to generate both prompt injection attacks and prompts that elicit unsafe responses. In both cases, we find that our approach is able to generate highly-effective and considerably more diverse attacks than past general red-teaming approaches.", "keywords": [ "red teaming", "safety", "reinforcement learning" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "FDimWzmcWn", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 10, "author_info_list": [ { "name": "Jingang Wang", "gender": "Male", "institution": "Meituan", "country": "", "position": "Researcher" }, { "name": "Keqing He", "gender": "unknown", "institution": "Meituan Group", "country": "", "position": "Researcher" }, { "name": "Wei Wang", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Weihao Zeng", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Intern" }, { "name": "Weiran Xu", "gender": "Male", "institution": "Beijing University of Post and Telecommunication", "country": "CN", "position": "Associate Professor" }, { "name": "Wentao Hong", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Xunliang Cai", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Principal Researcher" }, { "name": "Yejie Wang", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Zhuoma GongQue", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "dayuan fu", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 39, "n_ref": 76, "n_ref_all": 101, "n_fig": 11, "n_tab": 9, "L_tab": 3752, "n_element_tab": 706, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 3512, "n_element_tab_1": 586, "formula_len_all": 139, "formula_len_all_1": 138, "len_all": 158131, "len_all_1": 72651, "len_abs": 1623, "len_title": 118, "len_sents": 43131, "len_sents_1": 28552, "n_sents": 349, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1635, "title": "AgentRefine: Enhancing Agent Generalization through Refinement Tuning", "abs": "Large Language Model (LLM) based agents have proved their ability to perform complex tasks like humans. However, there is still a large gap between open-sourced LLMs and commercial models like the GPT series. In this paper, we focus on improving the agent generalization capabilities of LLMs via instruction tuning. We first observe that the existing agent training corpus exhibits satisfactory results on held-in evaluation sets but fails to generalize to held-out sets. These agent-tuning works face severe formatting errors and are frequently stuck in the same mistake for a long while. We analyze that the poor generalization ability comes from overfitting to several manual agent environments and a lack of adaptation to new situations. They struggle with the wrong action steps and can not learn from the experience but just memorize existing observation-action relations. Inspired by the insight, we propose a novel AgentRefine framework for agent-tuning. The core idea is to enable the model to learn to correct its mistakes via observation in the trajectory. Specifically, we propose an agent synthesis framework to encompass a diverse array of environments and tasks and prompt a strong LLM to refine its error action according to the environment feedback. AgentRefine significantly outperforms state-of-the-art agent-tuning work in terms of generalization ability on diverse agent tasks. It also has better robustness facing perturbation and can generate diversified thought in inference. Our findings establish the correlation between agent generalization and self-refinement and provide a new paradigm for future research.", "keywords": [ "agent", "self-refine", "diversity", "generalization", "data synthesis" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "FDhAngvHuf", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Reinhard Heckel", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Associate Professor" }, { "name": "Youssef Mansour", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 21, "n_ref": 33, "n_ref_all": 74, "n_fig": 10, "n_tab": 7, "L_tab": 1165, "n_element_tab": 177, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 469, "n_element_tab_1": 198, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 185968, "len_all_1": 63594, "len_abs": 4042, "len_title": 130, "len_sents": 61909, "len_sents_1": 31422, "n_sents": 495, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 972, "title": "Measuring Bias of Web-filtered Text Datasets and Bias Propagation Through Training", "abs": "In this paper, we investigate biases in pretraining datasets for large language models (LLMs) through dataset classification experiments. Building on prior work demonstrating the existence of biases in popular computer vision datasets, we analyze popular open-source pretraining text datasets derived from CommonCrawl including C4, RefinedWeb, DolmaCC, RedPajama-V2, FineWeb, DCLM-Baseline, and others. Despite those datasets being obtained with similar filtering and deduplication steps, LLMs can classify surprisingly well which dataset a single text sequence belongs to, significantly better than a human can. This indicates that popular pretraining datasets have their own unique biases or fingerprints. Those biases remain even when the text is rewritten with LLMs. We also demonstrate that these biases propagate through training: Random sequences generated by models trained on those datasets can be classified well by a classifier trained on the original datasets.", "keywords": [ "LLMs", "text datasets", "classification", "bias", "rewrite", "propagation" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "FDaHjwInXO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Changyou Chen", "gender": "Male", "institution": "State University of New York, Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Franck Dernoncourt", "gender": "unknown", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Jian Chen", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" }, { "name": "Jiuxiang Gu", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Tong Sun", "gender": "Female", "institution": "Adobe Systems", "country": "US", "position": "Director, Document Intelligence Lab" }, { "name": "Tong Yu", "gender": "unknown", "institution": "Adobe Research", "country": "US", "position": "Senior Research Scientist" }, { "name": "Yufan Zhou", "gender": "Male", "institution": "Adobe ", "country": "US", "position": "Researcher" }, { "name": "RUIYI ZHANG", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" }, { "name": "Ryan Rossi", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Senior Research Scientist" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 46, "n_ref": 103, "n_ref_all": 116, "n_fig": 6, "n_tab": 6, "L_tab": 2113, "n_element_tab": 491, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 3278, "n_element_tab_1": 91, "formula_len_all": 157, "formula_len_all_1": 157, "len_all": 188064, "len_all_1": 70473, "len_abs": 987, "len_title": 138, "len_sents": 41911, "len_sents_1": 31504, "n_sents": 269, "n_sents_1": 169, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 96, "L_abs": 1002, "title": "LoRA-Contextualizing Adaptation of Large Multimodal Models for Multi-page Document Understanding", "abs": "Large multimodal models (LMMs) have recently shown great progress in text-rich image understanding, yet they still struggle with complex, multi-page visually-rich documents. Traditional methods using document parsers for retrieval-augmented generation suffer from performance and efficiency limitations, while directly presenting all pages to LMMs leads to inefficiencies, especially with lengthy ones. In this work, we present a novel framework named **Lo**RA-**C**ontextualizing **A**daptation of **L**arge multimodal models (LoCAL) to broaden the horizons of any LMM for multi-page document understanding. We demonstrate that LMMs themselves can be an effective multimodal retriever to fetch relevant pages and then answer user questions based on these pages. LoCAL is implemented with two specific LMM adapters: one for evidence page retrieval and the other for question answering. Empirical results show state-of-the-art performance on public benchmarks, demonstrating the effectiveness of LoCAL.", "keywords": [ "Large Multimodal Models" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "FDMlGhExFp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Alex Labach", "gender": "unknown", "institution": "Layer 6 AI", "country": "CA", "position": "Researcher" }, { "name": "Anthony L. Caterini", "gender": "Male", "institution": "Layer6", "country": "CA", "position": "Researcher" }, { "name": "Guangwei Yu", "gender": "Male", "institution": "Layer6 AI", "country": "CA", "position": "Researcher" }, { "name": "Hamidreza Kamkari", "gender": "Male", "institution": "Layer6 AI", "country": "CA", "position": "Researcher" }, { "name": "Junwei Ma", "gender": "Male", "institution": "Layer 6 AI", "country": "CA", "position": "Researcher" }, { "name": "Keyvan Golestan", "gender": "Male", "institution": "Layer 6 AI", "country": "", "position": "Senior Machine Learning Scientist" }, { "name": "Rasa Hosseinzadeh", "gender": "Male", "institution": "Layer6", "country": "", "position": "Researcher" }, { "name": "Valentin Thomas", "gender": "Male", "institution": "Layer6", "country": "", "position": "Researcher" }, { "name": "Jesse C Cresswell", "gender": "unknown", "institution": "Layer 6 AI", "country": "CA", "position": "Staff Machine Learning Scientist" }, { "name": "Maksims Volkovs", "gender": "Male", "institution": "Layer6 AI", "country": "", "position": "Principal Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 55, "n_ref": 107, "n_ref_all": 145, "n_fig": 10, "n_tab": 8, "L_tab": 2621, "n_element_tab": 249, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 814, "n_element_tab_1": 10, "formula_len_all": 155, "formula_len_all_1": 0, "len_all": 197320, "len_all_1": 68846, "len_abs": 6207, "len_title": 89, "len_sents": 58224, "len_sents_1": 34236, "n_sents": 443, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1435, "title": "TabDPT: Scaling Tabular Foundation Models", "abs": "The challenges faced by neural networks on tabular data are well-documented and have hampered the progress of tabular foundation models. Techniques leveraging in-context learning (ICL) have shown promise here, allowing for dynamic adaptation to unseen data. ICL can provide predictions for entirely new datasets without further training or hyperparameter tuning, therefore providing very fast inference when encountering a novel task. However, scaling ICL for tabular data remains an issue: approaches based on large language models cannot efficiently process numeric tables, and tabular-specific techniques have not been able to effectively harness the power of real data to improve performance and generalization. We are able to overcome these challenges by training tabular-specific ICL-based architectures on real data with self-supervised learning and retrieval, combining the best of both worlds. Our resulting model -- the Tabular Discriminative Pre-trained Transformer (TabDPT) -- achieves state-of-the-art performance on the CC18 (classification) and CTR23 (regression) benchmarks with no task-specific fine-tuning, demonstrating the adapatability and speed of ICL once the model is pre-trained. TabDPT also demonstrates strong scaling as both model size and amount of available data increase, pointing towards future improvements simply through the curation of larger tabular pre-training datasets and training larger models.", "keywords": [ "Tabular Data", "Foundation Models", "Tabular Foundation Models", "In-Context Learning", "Retrieval" ], "rating_list": [ 8, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "FD9sPyS8ve", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Aditi Raghunathan", "gender": "Female", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Suhas Kotha", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Taeyoun Kim", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 54, "n_ref": 114, "n_ref_all": 162, "n_fig": 9, "n_tab": 18, "L_tab": 4368, "n_element_tab": 354, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 42, "formula_len_all_1": 22, "len_all": 226271, "len_all_1": 64093, "len_abs": 946, "len_title": 107, "len_sents": 60736, "len_sents_1": 31139, "n_sents": 532, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 955, "title": "Testing the Limits of Jailbreaking with the Purple Problem", "abs": "The rise of ''jailbreak'' attacks on language models has led to a flurry of defenses aimed at preventing undesirable responses. Nonetheless, most benchmarks remain to be solved, not to mention real-world safety problems. We critically examine the two stages of the defense pipeline: (i) defining what constitutes unsafe outputs, and (ii) enforcing the definition via methods such as fine-tuning or input preprocessing. To understand whether we fail because of definition or enforcement, we consider a simple and well-specified definition of unsafe outputs---outputs that contain the word ''purple''. Surprisingly, all existing fine-tuning and input defenses fail to enforce this definition under adaptive attacks and increasing compute, casting doubt on whether enforcement algorithms can be robust for more complicated definitions. We hope that this definition serves as a testbed to evaluate enforcement algorithms and prevent a false sense of security.", "keywords": [ "Jailbreaking", "Adversarial Robustness", "Security", "Adaptive Attacks" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "FCMpUOZkxi", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Hengquan Guo", "gender": "unknown", "institution": "ShanghaiTech University", "country": "CN", "position": "PhD student" }, { "name": "Xin Liu", "gender": "unknown", "institution": "ShanghaiTech University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 56, "n_formula_1": 22, "n_ref_uni": 25, "n_ref": 83, "n_ref_all": 101, "n_fig": 1, "n_tab": 7, "L_tab": 2163, "n_element_tab": 375, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 479, "n_element_tab_1": 14, "formula_len_all": 5645, "formula_len_all_1": 1082, "len_all": 128531, "len_all_1": 68890, "len_abs": 1319, "len_title": 119, "len_sents": 39667, "len_sents_1": 30239, "n_sents": 315, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1429, "title": "On Stochastic Contextual Bandits with Knapsacks in Small Budget Regime", "abs": "This paper studies stochastic contextual bandits with knapsack constraints (CBwK), where a learner observes a context, takes an action, receives a reward, and incurs a vector of costs at every round. The learner aims to maximize the cumulative rewards across $T$ rounds under the knapsack constraints with an initial budget of $B$. We study CBwK in the small budget regime where the budget $B = \\Omega(\\sqrt{T})$\nand propose an Adaptive and Universal Primal--Dual algorithm (AUPD) that achieves strong regret performance: \ni) AUPD achieves $\\tilde{O}((1 + \\frac{\\nu^*}{\\delta b})\\sqrt{T})$ regret under the strict feasibility assumption without any prior information, matching the best-known bounds;\nii) AUPD achieves $\\tilde{O}(\\sqrt{T}+ \\frac{\\nu^*}{\\sqrt{b}}T^{\\frac{3}{4}})$ regret without strict feasibility assumption, \nwhich, to the best of our knowledge, is the first result in the literature. Here, the parameter $\\nu^*$ represents the optimal average reward; $b=B/T$ is the average budget and $\\delta b$ is the feasibility/safety margin.\nWe establish these strong results through the adaptive budget-aware design, which effectively balances reward maximization and budget consumption. We provide a new perspective on analyzing budget consumption using the Lyapunov drift method, along with a refined analysis of its cumulative variance. Our theory is further supported by experiments conducted on a large-scale dataset.", "keywords": [ "Contextual bandits with knapsacks", "small budget" ], "rating_list": [ 5, 8, 6, 8 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 5, 2, 2 ] }, { "paper_id": "FCCeBaFa8M", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Tianyi Zhang", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" }, { "name": "Yuan Tian", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 18, "n_ref_uni": 57, "n_ref": 105, "n_ref_all": 127, "n_fig": 5, "n_tab": 2, "L_tab": 225, "n_element_tab": 37, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 225, "n_element_tab_1": 37, "formula_len_all": 1378, "formula_len_all_1": 1270, "len_all": 170967, "len_all_1": 63153, "len_abs": 1477, "len_title": 95, "len_sents": 39256, "len_sents_1": 28343, "n_sents": 343, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1500, "title": "Selective Prompt Anchoring for Code Generation", "abs": "Recent advances in large language models (LLMs) have transformed software development by automatically generating code based on users' requests in natural language. Despite these advancements, challenges remain in generating buggy code and fully aligning with user intent. Our empirical study reveals LLMs tend to dilute their self-attentions on the initial prompt as more code tokens are generated. We hypothesize this self-attention dilution issue is one of the root causes of inaccuracies in LLM-generated code. To mitigate this issue, we propose **S**elective **P**rompt **A**nchoring (SPA) to amplify the influence of the selected parts in the initial prompt, which we refer to as \"anchored text\", during code generation. Specifically, SPA calculates the logit distribution difference with and without the anchored text. We prove this logit difference approximates the anchored text's contextual contribution to the output logits. SPA creates an augmented logit distribution by linearly combining the original logit distribution and the logit difference. We evaluate SPA with five LLMs on four benchmarks. Our results show that after tuning on a few dozen instances, SPA consistently improves Pass@1 on new tasks by up to 7.6% across all settings. Notably, with selective text anchoring, a small version of DeepSeek-Coder (6.7B) can achieve better performance than an original much larger version (33B). Our code is available at https://anonymous.4open.science/r/Selective-Prompt-Anchoring-74E7.", "keywords": [ "Large Language Models (LLMs)", "Code Generation", "Attention", "Logits", "Anchoring", "Prompt" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "FCBbh0HCrF", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Bin Gu", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Boyu Wang", "gender": "Male", "institution": "University of Western Ontario", "country": "CA", "position": "Assistant Professor" }, { "name": "Charles Ling", "gender": "Male", "institution": "Western University", "country": "CA", "position": "Professor" }, { "name": "Ganyu Wang", "gender": "Male", "institution": "Western University", "country": "CA", "position": "PhD student" } ], "n_formula": 52, "n_formula_1": 17, "n_ref_uni": 45, "n_ref": 115, "n_ref_all": 136, "n_fig": 9, "n_tab": 9, "L_tab": 3832, "n_element_tab": 292, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 722, "n_element_tab_1": 124, "formula_len_all": 8372, "formula_len_all_1": 1219, "len_all": 200580, "len_all_1": 67008, "len_abs": 1701, "len_title": 95, "len_sents": 51402, "len_sents_1": 28809, "n_sents": 453, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1440, "title": "Event-Driven Online Vertical Federated Learning", "abs": "Online learning is more adaptable to real-world scenarios in Vertical Federated Learning (VFL) compared to offline learning. \nHowever, integrating online learning into VFL presents challenges due to the unique nature of VFL, where clients possess non-intersecting feature sets for the same sample. \nIn real-world scenarios, the clients may not receive data streaming for the disjoint features for the same entity synchronously. Instead, the data are typically generated by an *event* relevant to only a subset of clients.\nWe are the first to identify these challenges in online VFL, which have been overlooked by previous research. To address these challenges, we proposed an event-driven online VFL framework. In this framework, only a subset of clients were activated during each event, while the remaining clients passively collaborated in the learning process. \nFurthermore, we incorporated *dynamic local regret (DLR)* into VFL to address the challenges posed by online learning problems with non-convex models within a non-stationary environment.\nWe conducted a comprehensive regret analysis of our proposed framework, specifically examining the DLR under non-convex conditions with event-driven online VFL. \nExtensive experiments demonstrated that our proposed framework was more stable than the existing online VFL framework under non-stationary data conditions while also significantly reducing communication and computation costs.", "keywords": [ "Vertical Federated Learning", "Online Learning", "Event Driven" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "FBkpCyujtS", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Allen G Roush", "gender": "Male", "institution": "Oracle", "country": "", "position": "Researcher" }, { "name": "Andreas Kirsch", "gender": "unknown", "institution": "GenAI Startup", "country": "GB", "position": "Researcher" }, { "name": "Clement Neo", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "Nguyen Nhat Minh", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Undergrad student" }, { "name": "Andrew Bakker", "gender": "unknown", "institution": "New York University", "country": "US", "position": "Undergrad student" }, { "name": "ravid ziv", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 20, "n_ref": 45, "n_ref_all": 59, "n_fig": 2, "n_tab": 15, "L_tab": 7993, "n_element_tab": 783, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 595, "n_element_tab_1": 80, "formula_len_all": 210, "formula_len_all_1": 149, "len_all": 160240, "len_all_1": 60000, "len_abs": 4809, "len_title": 121, "len_sents": 51342, "len_sents_1": 28535, "n_sents": 396, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 73, "L_abs": 1024, "title": "Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs", "abs": "Large Language Models (LLMs) generate text by sampling the next token from a probability distribution over the vocabulary at each decoding step. However, popular sampling methods like top-p (nucleus sampling) often struggle to balance quality and diversity, especially at higher temperatures, leading to incoherent or repetitive outputs. To address this challenge, we propose min-p sampling, a dynamic truncation method that adjusts the sampling threshold based on the model's confidence by scaling according to the top token's probability. We conduct extensive experiments on benchmarks including GPQA, GSM8K, and AlpacaEval Creative Writing, demonstrating that min-p sampling improves both the quality and diversity of generated text, particularly at high temperatures. Moreover, human evaluations reveal a clear preference for min-p sampling in terms of both text quality and diversity. Min-p sampling has been adopted by multiple open-source LLM implementations, highlighting its practical utility and potential impact.", "keywords": [ "Natural Language Processing", "Large Language Models", "Text Generation", "Sampling Methods", "Truncation Sampling", "Stochastic Sampling", "Min-p Sampling", "Top-p Sampling", "Nucleus Sampling", "Temperature Sampling", "Decoding Methods", "Deep Learning", "Artificial Intelligence" ], "rating_list": [ 6, 10, 10, 8 ], "soundness_list": [ 3, 4, 4, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 4, 4, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "FBhKUXK7od", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Kira Michaela D\u00fcsterwald", "gender": "Female", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Makoto Yamada", "gender": "Male", "institution": "Okinawa Institute of Science and Technology (OIST)", "country": "", "position": "Associate Professor" }, { "name": "Samo Hromadka", "gender": "Male", "institution": "Gatsby Computational Neuroscience Unit", "country": "", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 21, "n_ref_uni": 34, "n_ref": 92, "n_ref_all": 108, "n_fig": 6, "n_tab": 2, "L_tab": 546, "n_element_tab": 62, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2717, "formula_len_all_1": 1397, "len_all": 148276, "len_all_1": 67151, "len_abs": 1710, "len_title": 119, "len_sents": 44199, "len_sents_1": 29763, "n_sents": 353, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1742, "title": "Fast unsupervised ground metric learning with tree-Wasserstein distance", "abs": "The performance of unsupervised methods such as clustering depends on the choice of distance metric between features, or ground metric. Commonly, ground metrics are decided with heuristics or learned via supervised algorithms. However, since many interesting datasets are unlabelled, unsupervised ground metric learning approaches have been introduced. One promising option employs Wasserstein singular vectors (WSVs), which emerge when computing optimal transport distances between features and samples simultaneously. WSVs are effective, but can be prohibitively computationally expensive in some applications: $\\mathcal{O}(n^2m^2(n \\log(n) + m \\log(m))$ for $n$ samples and $m$ features. In this work, we propose to augment the WSV method by embedding samples and features on trees, on which we compute the tree-Wasserstein distance (TWD). We demonstrate theoretically and empirically that the algorithm converges to a better approximation of the standard WSV approach than the best known alternatives, and does so with $\\mathcal{O}(n^3+m^3+mn)$ complexity. In addition, we prove that the initial tree structure can be chosen flexibly, since tree geometry does not constrain the richness of the approximation up to the number of edge weights. This proof suggests a fast and recursive algorithm for computing the tree parameter basis set, which we find crucial to realising the efficiency gains at scale. Finally, we employ the tree-WSV algorithm to several single-cell RNA sequencing genomics datasets, demonstrating its scalability and utility for unsupervised cell-type clustering problems. These results poise unsupervised ground metric learning with TWD as a low-rank approximation of WSV with the potential for widespread application.", "keywords": [ "unsupervised learning", "optimal transport", "distance-based learning", "clustering", "trees", "wasserstein distance" ], "rating_list": [ 8, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "FB84Wkn3Xp", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Bo Zheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Principal Researcher" }, { "name": "Qipeng zhang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "MS student" }, { "name": "Shuai Wang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Tiezheng Ge", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xubin Li", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Zexian Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Limin Wang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "tianhui Song", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" } ], "n_formula": 44, "n_formula_1": 24, "n_ref_uni": 25, "n_ref": 56, "n_ref_all": 89, "n_fig": 7, "n_tab": 4, "L_tab": 2928, "n_element_tab": 110, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 439, "n_element_tab_1": 5, "formula_len_all": 4757, "formula_len_all_1": 1751, "len_all": 130829, "len_all_1": 54935, "len_abs": 367, "len_title": 105, "len_sents": 36025, "len_sents_1": 22930, "n_sents": 328, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1149, "title": "Differentiable Solver Search for fast diffusion sampling", "abs": "Diffusion-based models have demonstrated remarkable generation quality but at the cost of numerous function evaluations. Recently, advanced ODE-based solvers have been developed to mitigate the substantial computational demands of reverse-diffusion solving under limited sampling steps. However, these solvers, heavily inspired by Adams-like multistep methods, rely solely on t-related Lagrange interpolation. We show that t-related Lagrange interpolation is suboptimal and reveals a compact search space comprised of timestep and solver coefficients. Building on our analysis, we propose a novel differentiable solver search algorithm to identify the optimal solver. Equipped with the searched solver, our rectified flow models, SiT-XL/2 and FlowDCN-XL/2, achieve FID scores of 2.40 and 2.35, respectively, on ImageNet-$256\\times256$ with only 10 steps. Meanwhile, our DDPM model, DiT-XL/2, reaches a FID score of 2.33 with only 10 steps. Notably, our searched solver outperforms traditional solvers by a significant margin. Moreover, our searched solver demonstrates its generality across various model architectures, resolutions, and model sizes.", "keywords": [ "Generative models", "Solver", "Sampler", "FlowMatching" ], "rating_list": [ 5, 6, 8 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 1, 4, 2 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 2, 3, 2 ] }, { "paper_id": "FAfxvdv1Dy", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Chao Shen", "gender": "Male", "institution": "Xi\u2019an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Juan Zhai", "gender": "Female", "institution": "University of Massachusetts at Amherst", "country": "", "position": "Assistant Professor" }, { "name": "Shiqing Ma", "gender": "unknown", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Tianlin Li", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Weipeng Jiang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xiaoyu Zhang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Yang Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 48, "n_ref": 119, "n_ref_all": 158, "n_fig": 10, "n_tab": 9, "L_tab": 3637, "n_element_tab": 195, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2514, "n_element_tab_1": 135, "formula_len_all": 556, "formula_len_all_1": 291, "len_all": 187554, "len_all_1": 66928, "len_abs": 1508, "len_title": 114, "len_sents": 57860, "len_sents_1": 31842, "n_sents": 427, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1222, "title": "STAFF: Speculative Coreset Selection for Task-Specific Fine-tuning", "abs": "Task-specific fine-tuning is essential for the deployment of large language models (LLMs), but it requires significant computational resources and time. Existing solutions have proposed coreset selection methods to improve data efficiency and reduce model training overhead, but they still have limitations: \u2776 Overlooking valuable samples at high pruning rates, which degrades the coreset\u2019s performance.\n\u2777 Requiring high time overhead during coreset selection to fine-tune and evaluate the target LLM. In this paper, we introduce STAFF, a speculative coreset selection method. STAFF leverages a small model from the same family as the target LLM to efficiently estimate data scores and then verifies the scores on the target LLM to accurately identify and allocate more selection budget to important regions while maintaining coverage of easy regions. We evaluate STAFF on three LLMs and three downstream tasks and show that STAFF improves the performance of SOTA methods by up to 54.3% and reduces selection overhead by up to 70.5% at different pruning rates. Furthermore, we observe that the coreset selected by STAFF at low pruning rates (i.e., 20%) can even obtain better fine-tuning performance than the full dataset.", "keywords": [ "Task-specific fine-tuning", "coreset selection", "speculative execution" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 4, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 2, 5, 4, 4 ] }, { "paper_id": "FAYIlGDBa1", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Dekun Zhou", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "PhD student" }, { "name": "Yinglun Zhu", "gender": "unknown", "institution": "University of California, Riverside", "country": "US", "position": "Assistant Professor" }, { "name": "Alberto Pia", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "", "position": "Associate Professor" } ], "n_formula": 72, "n_formula_1": 20, "n_ref_uni": 34, "n_ref": 65, "n_ref_all": 95, "n_fig": 1, "n_tab": 16, "L_tab": 4985, "n_element_tab": 678, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 46, "n_element_tab_1": 10, "formula_len_all": 6770, "formula_len_all_1": 2303, "len_all": 185502, "len_all_1": 67557, "len_abs": 1590, "len_title": 94, "len_sents": 59408, "len_sents_1": 31130, "n_sents": 501, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 46, "L_abs": 1781, "title": "Efficient Sparse PCA via Block-Diagonalization", "abs": "Sparse Principal Component Analysis (Sparse PCA) is a pivotal tool in data analysis and dimensionality reduction. However, Sparse PCA is a challenging problem in both theory and practice: it is known to be NP-hard and current exact methods generally require exponential runtime. In this paper, we propose a novel framework to efficiently approximate Sparse PCA by (i) approximating the general input covariance matrix with a re-sorted block-diagonal matrix, (ii) solving the Sparse PCA sub-problem in each block, and (iii) reconstructing the solution to the original problem. Our framework is simple and powerful: it can leverage any off-the-shelf Sparse PCA algorithm and achieve significant computational speedups, with a minor additive error that is linear in the approximation error of the block-diagonal matrix. Suppose $g(k, d)$ is the runtime of an algorithm (approximately) solving Sparse PCA in dimension $d$ and with sparsity value $k$. Our framework, when integrated with this algorithm, reduces the runtime to $\\mathcal{O}\\left(\\frac{d}{d^\\star} \\cdot g(k, d^\\star) + d^2\\right)$, where $d^\\star \\leq d$ is the largest block size of the block-diagonal matrix. For instance, integrating our framework with the Branch-and-Bound algorithm reduces the complexity from $g(k, d) = \\mathcal{O}(k^3\\cdot d^k)$ to $\\mathcal{O}(k^3\\cdot d \\cdot (d^\\star)^{k-1})$, demonstrating exponential speedups if $d^\\star$ is small. We perform large-scale evaluations on many real-world datasets: for exact Sparse PCA algorithm, our method achieves an average speedup factor of 93.77, while maintaining an average approximation error of 2.15\\%; for approximate Sparse PCA algorithm, our method achieves an average speedup factor of 6.77 and an average approximation error of merely 0.37\\%.", "keywords": [ "Sparse PCA", "Block Diagonalization", "Compurational Efficiency", "Approximation Algorithms" ], "rating_list": [ 8, 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 2, 4, 3 ] }, { "paper_id": "FA5ZAJlv96", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Hyunjung Shim", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Jaeyo Shin", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Jiho Choi", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Jiwook Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Seonho Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 26, "n_formula_1": 23, "n_ref_uni": 36, "n_ref": 77, "n_ref_all": 118, "n_fig": 20, "n_tab": 6, "L_tab": 1282, "n_element_tab": 136, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1016, "n_element_tab_1": 99, "formula_len_all": 2121, "formula_len_all_1": 1840, "len_all": 174331, "len_all_1": 65094, "len_abs": 612, "len_title": 150, "len_sents": 59728, "len_sents_1": 28338, "n_sents": 465, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 101, "L_abs": 1589, "title": "DreamCatalyst: Fast and High-Quality 3D Editing via Controlling Editability and Identity Preservation", "abs": "Score distillation sampling (SDS) has emerged as an effective framework in text-driven 3D editing tasks, leveraging diffusion models for 3D-consistent editing. However, existing SDS-based 3D editing methods suffer from long training times and produce low-quality results. We identify that the root cause of this performance degradation is their conflict with the sampling dynamics of diffusion models. Addressing this conflict allows us to treat SDS as a diffusion reverse process for 3D editing via sampling from data space. In contrast, existing methods naively distill the score function using diffusion models. From these insights, we propose DreamCatalyst, a novel framework that considers these sampling dynamics in the SDS framework. Specifically, we devise the optimization process of our DreamCatalyst to approximate the diffusion reverse process in editing tasks, thereby aligning with diffusion sampling dynamics. As a result, DreamCatalyst successfully reduces training time and improves editing quality. Our method offers two modes: (1) a fast mode that edits Neural Radiance Fields (NeRF) scenes approximately 23 times faster than current state-of-the-art NeRF editing methods, and (2) a high-quality mode that produces superior results about 8 times faster than these methods. Notably, our high-quality mode outperforms current state-of-the-art NeRF editing methods in terms of both speed and quality. DreamCatalyst also surpasses the state-of-the-art 3D Gaussian Splatting (3DGS) editing methods, establishing itself as an effective and model-agnostic 3D editing solution.", "keywords": [ "diffusion models", "3D editing", "score distillation sampling", "NeRF", "3D Gaussian Splatting" ], "rating_list": [ 6, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "FA3iYp1y6z", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "James Hensman", "gender": "unknown", "institution": "Microsoft Research", "country": "GB", "position": "Principal Researcher" }, { "name": "Meyer Scetbon", "gender": "Male", "institution": "Microsoft", "country": "GB", "position": "Researcher" } ], "n_formula": 40, "n_formula_1": 18, "n_ref_uni": 24, "n_ref": 73, "n_ref_all": 90, "n_fig": 4, "n_tab": 7, "L_tab": 3472, "n_element_tab": 419, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 1236, "n_element_tab_1": 195, "formula_len_all": 3005, "formula_len_all_1": 1332, "len_all": 136720, "len_all_1": 64315, "len_abs": 1001, "len_title": 86, "len_sents": 39056, "len_sents_1": 28089, "n_sents": 307, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 38, "L_abs": 1024, "title": "Low-Rank Correction for Quantized LLMs", "abs": "We consider the problem of model compression for Large Language Models (LLMs) at post-training time, where the task is to compress a well-trained model using only a small set of calibration input data. \nIn this work, we introduce a new low-rank approach to correct for quantization errors of \\emph{activations} in LLMs: we propose to add low-rank weight matrices in full precision that act on the \\emph{unquantized} activations. We then solve a joint optimization problem over the quantized representation of the weights and additional low-rank weight matrices to quantize both weights and activations.\nWe focus on the case of 4-bit weight-and-activation quantization (W4A4). Using ranks equivalent to 10\\% of the original weight matrix size, our approach reduces the accuracy gap with the original model by more than 50\\%. Using ranks equivalent to 30\\% of the original weight matrix, the accuracy gap is closed completely. We demonstrate our results on four recent LLMs, namely Llama-2, Llama-3, Phi-3 and Mixtral models.", "keywords": [ "Quantization", "LLM", "Low-rank" ], "rating_list": [ 3, 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 2, 2, 3 ] }, { "paper_id": "F9iHSa1Iz5", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Dong Yan", "gender": "Male", "institution": "Baichuan Intelligent Technology", "country": "", "position": "Researcher" }, { "name": "Jialian Li", "gender": "Female", "institution": "Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Jian Xie", "gender": "Male", "institution": "Beijing Baichuan Intelligent Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "wei shen", "gender": "unknown", "institution": "Baichuan", "country": "", "position": "Researcher" }, { "name": "YipinZhang", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Yuzi Yan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 20, "n_ref": 41, "n_ref_all": 48, "n_fig": 1, "n_tab": 4, "L_tab": 1559, "n_element_tab": 241, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1041, "n_element_tab_1": 185, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 123075, "len_all_1": 72999, "len_abs": 841, "len_title": 103, "len_sents": 35093, "len_sents_1": 32390, "n_sents": 318, "n_sents_1": 310, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 845, "title": "Boosting Deductive Reasoning with Step Signals In RLHF", "abs": "Logical reasoning is a crucial task for Large Language Models (LLMs), enabling them to tackle complex problems. Among reasoning tasks, multi-step reasoning poses a particular challenge. Grounded in the theory of formal logic, we have developed an automated method, Multi-step Deduction (MuseD), for deductive reasoning data. MuseD has allowed us to create training and testing datasets for multi-step reasoning. Our generation method enables control over the complexity of the generated instructions, facilitating training and evaluation of models across different difficulty levels. Through RLHF training, our training data has demonstrated significant improvements in logical capabilities for both in-domain of out-of-domain reasoning tasks. Additionally, we have conducted tests to assess the multi-step reasoning abilities of various models.", "keywords": [ "LLM", "RLHF", "reasoning" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 4, 2 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 5, 3, 3 ] }, { "paper_id": "F9MTOYTzEm", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Hailong Shi", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Jingyi Wu", "gender": "Female", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Jinsheng Xiao", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Associate Professor" }, { "name": "Shurui Wang", "gender": "Male", "institution": "Wuhan University", "country": "", "position": "MS student" }, { "name": "Xingyu Gao", "gender": "unknown", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xie Honggang", "gender": "unknown", "institution": "Hubei University of Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 24, "n_formula_1": 15, "n_ref_uni": 23, "n_ref": 41, "n_ref_all": 73, "n_fig": 13, "n_tab": 4, "L_tab": 1971, "n_element_tab": 306, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1660, "n_element_tab_1": 272, "formula_len_all": 1457, "formula_len_all_1": 737, "len_all": 135339, "len_all_1": 60482, "len_abs": 1382, "len_title": 139, "len_sents": 36480, "len_sents_1": 25206, "n_sents": 281, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1389, "title": "Instance-level Consistent Graph With Unsupervised Human Parts for Person Re-identification", "abs": "The representation of human parts plays a crucial role in person re-identification (re-ID) by offering discriminative cues, yet it presents challenges such as misalignment, occlusion, and extreme illumination. Previous methods have primarily focused on achieving strict part-level consistency. However, individual part features change inevitably under harsh conditions, hindering consistent representation. In this article, we propose an Instance-level Consistent Graph (ICG) framework to address this issue, which extracts structural information by introducing graph modeling atop unsupervised human parts. Firstly, we introduce an attention-based foreground separation to suppress non-instance noise. Subsequently, an unsupervised clustering method is designed to segment pixel-wise human parts within the foreground, enabling fine-grained part representations. We propose a flexible structure graph that derives instance-level structure from part features, treating each part feature as a node in a graph convolutional network. In essence, ICG mitigates incompleteness through feature flow among nodes, broadening the matching condition from strict part-level consistency to robust instance-level consistency. Extensive experiments on three popular person re-ID datasets demonstrate that ICG surpasses most state-of-the-art methods, exhibiting remarkable improvements over the baseline.", "keywords": [ "Person re-identification", "Instance-level consistency", "Human parts clustering", "Graph convolution network" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "F9JZiGradI", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Lichao Sun", "gender": "Male", "institution": "Lehigh University", "country": "US", "position": "Assistant Professor" }, { "name": "Yifeng Xie", "gender": "Male", "institution": "Guangdong University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yunhong He", "gender": "Female", "institution": "Anhui Polytechnic University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhengqing Yuan", "gender": "Male", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 14, "n_ref_uni": 58, "n_ref": 73, "n_ref_all": 87, "n_fig": 5, "n_tab": 7, "L_tab": 1858, "n_element_tab": 132, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 3166, "n_element_tab_1": 177, "formula_len_all": 846, "formula_len_all_1": 891, "len_all": 150117, "len_all_1": 62359, "len_abs": 1648, "len_title": 107, "len_sents": 38319, "len_sents_1": 26801, "n_sents": 258, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1285, "title": "MLP-KAN: Unifying Deep Representation and Function Learning", "abs": "Recent advancements in both representation learning and function learning have demonstrated substantial promise across diverse domains of artificial intelligence. However, the effective integration of these paradigms poses a significant challenge, particularly in cases where users must manually decide whether to apply a representation learning or function learning model based on dataset characteristics. To address this issue, we introduce MLP-KAN, a unified method designed to eliminate the need for manual model selection. By integrating Multi-Layer Perceptrons (MLPs) for representation learning and Kolmogorov-Arnold Networks (KANs) for function learning within a Mixture-of-Experts (MoE) architecture, MLP-KAN dynamically adapts to the specific characteristics of the task at hand, ensuring optimal performance. Embedded within a transformer-based framework, our work achieves remarkable results on four widely-used datasets across diverse domains. Extensive experimental evaluation demonstrates its superior versatility, delivering competitive performance across both deep representation and function learning tasks. These findings highlight the potential of MLP-KAN to simplify the model selection process, offering a comprehensive, adaptable solution across various domains.", "keywords": [ "representational learning", "functional learning", "unified model" ], "rating_list": [ 6, 6, 8, 1 ], "soundness_list": [ 4, 2, 4, 1 ], "presentation_list": [ 3, 2, 4, 1 ], "contribution_list": [ 2, 2, 4, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "F8qvqtnSHy", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "David Danks", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "Full Professor" }, { "name": "Mohammadsajad Abavisani", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Payal Anil Bhandari", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Praveen Nair", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "MS student" }, { "name": "Sergey M. Plis", "gender": "Male", "institution": "Georgia State University", "country": "US", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 33, "n_ref": 43, "n_ref_all": 57, "n_fig": 5, "n_tab": 3, "L_tab": 1853, "n_element_tab": 361, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 954, "n_element_tab_1": 112, "formula_len_all": 276, "formula_len_all_1": 312, "len_all": 101584, "len_all_1": 65665, "len_abs": 1082, "len_title": 106, "len_sents": 32694, "len_sents_1": 30946, "n_sents": 207, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1095, "title": "ION-C: Integration of Overlapping Networks via Constraints", "abs": "In many causal learning problems, variables of interest are often not all measured over the same observations, but are instead distributed across multiple datasets with overlapping variables. Tillman et al. (2008) presented the first algorithm for determining the minimal equivalence class of ground-truth DAGs consistent with all input graphs by exploiting local independence relations, called ION. In this paper, this problem is formulated as a more computationally efficient answer-set programming (ASP) problem, which we call ION-C, and solved with the ASP system $\\textit{clingo}$. The ION-C algorithm was run on random synthetic graphs with varying sizes, densities, and degrees of overlap between subgraphs, with overlap having the largest impact on runtime, number of solution graphs, and agreement within the output set. To validate ION-C on real-world data, we ran the algorithm on overlapping graphs learned from data from two successive iterations of the European Social Survey (ESS), using a procedure for conducting joint independence tests to prevent inconsistencies in the input.", "keywords": [ "Causal learning", "Constraint satisfaction", "Answer set programming", "Social science data" ], "rating_list": [ 3, 6, 3, 3 ], "soundness_list": [ 3, 3, 4, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "F7yPR6XhFR", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Christina Zhang", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Undergrad student" }, { "name": "Ji Hou", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Research Scientist" }, { "name": "Jialiang Wang", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Research scientist" }, { "name": "Matthew Yu", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Software Engineer" }, { "name": "Peter Vajda", "gender": "unknown", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Sam Tsai", "gender": "Male", "institution": "Meta", "country": "US", "position": "Applied research scientist" }, { "name": "Simran Motwani", "gender": "Female", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Zijian He", "gender": "Male", "institution": "Meta GenAI", "country": "US", "position": "Researcher" }, { "name": "Felix Juefei-Xu", "gender": "Male", "institution": "GenAI, Meta", "country": "US", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 30, "n_ref": 67, "n_ref_all": 85, "n_fig": 6, "n_tab": 3, "L_tab": 391, "n_element_tab": 31, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 753, "formula_len_all_1": 748, "len_all": 102705, "len_all_1": 47466, "len_abs": 1118, "len_title": 100, "len_sents": 28160, "len_sents_1": 21191, "n_sents": 219, "n_sents_1": 162, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 1130, "title": "Pixel-Space Post-Training of Latent-Diffusion Models", "abs": "Latent diffusion models (LDMs) have made significant advancements in the field of image generation in recent years. One major advantage of LDMs is their ability to operate in a compressed latent space, allowing for more efficient training and deployment. However, despite these advantages, challenges with LDMs still remain. For example, it has been observed that LDMs often generate high-frequency details and complex compositions imperfectly. We hypothesize that one reason for these flaws is due to the fact that all pre- and post-training of LDMs are done in latent space, which is typically $8 \\times 8$ lower spatial-resolution than the output images. To address this issue, we propose adding pixel-space supervision in the post-training process to better preserve high-frequency details. Experimentally, we show that adding a pixel-space objective significantly improves both supervised quality fine-tuning and preference-based post-training by a large margin on a state-of-the-art DiT transformer and U-Net diffusion models in both visual quality and visual flaw metrics, while maintaining the same text alignment quality.", "keywords": [ "latent diffusion models", "fine-tuning", "pixel space", "image generation" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 4, 3, 5 ] }, { "paper_id": "F7QNwDYG6I", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Chaojie Wang", "gender": "Male", "institution": "Skywork AI", "country": "SG", "position": "Researcher" }, { "name": "Jujie He", "gender": "unknown", "institution": "kunlun", "country": "CN", "position": "Researcher" }, { "name": "Liang Zeng", "gender": "Male", "institution": "Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Yanchen Deng", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "PhD student" }, { "name": "Lv Zhiyi", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 17, "n_ref_uni": 51, "n_ref": 106, "n_ref_all": 127, "n_fig": 3, "n_tab": 2, "L_tab": 541, "n_element_tab": 55, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 541, "n_element_tab_1": 55, "formula_len_all": 1203, "formula_len_all_1": 1170, "len_all": 136346, "len_all_1": 59192, "len_abs": 1173, "len_title": 118, "len_sents": 33324, "len_sents_1": 27350, "n_sents": 227, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 70, "L_abs": 1176, "title": "Q*: Improving Multi-step Reasoning for LLMs with Deliberative Planning", "abs": "Large Language Models (LLMs) have demonstrated impressive capability across various natural language tasks. However, the auto-regressive generation process makes LLMs prone to produce errors, hallucinations and inconsistent statements when performing multi-step reasoning. In this paper, by casting multi-step reasoning of LLMs as a heuristic search problem, we aim to alleviate the pathology by introducing Q*, a general, versatile and agile framework for guiding LLMs decoding process with deliberative planning. By learning a plug-and-play Q-value model as heuristic function for estimating expected future rewards, Q* can effectively guide LLMs to select the most promising next reasoning step without fine-tuning LLMs for the targeted task, which avoids the significant computational overhead and potential risk of performance degeneration on other tasks. Extensive experiments on GSM8K, MATH and MBPP datasets demonstrate the superiority of our method, contributing to improving the reasoning capability of existing open-source LLMs. Furthermore, the testing-time scaling law indicates that Q* can leverage increased computational power to improve reasoning performance.", "keywords": [ "LLM", "Alignment", "Planning" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "F6z3utfcYw", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Ruizhe Shi", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Runlong Zhou", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Simon Shaolei Du", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Assistant Professor" } ], "n_formula": 64, "n_formula_1": 19, "n_ref_uni": 29, "n_ref": 86, "n_ref_all": 116, "n_fig": 6, "n_tab": 8, "L_tab": 2521, "n_element_tab": 152, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2975, "n_element_tab_1": 116, "formula_len_all": 5147, "formula_len_all_1": 1189, "len_all": 220656, "len_all_1": 72951, "len_abs": 1056, "len_title": 117, "len_sents": 52436, "len_sents_1": 25375, "n_sents": 482, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 948, "title": "The Crucial Role of Samplers in Online Direct Preference Optimization", "abs": "Direct Preference Optimization (DPO) has emerged as a stable, scalable, and efficient solution for language model alignment.\nDespite its empirical success, the optimization properties, particularly the impact of samplers on its convergence rates, remain under-explored. In this paper, we provide a rigorous analysis of DPO's convergence rates with different sampling strategies under the exact gradient setting, revealing a surprising separation: uniform sampling achieves $\\textbf{linear}$ convergence, while our proposed online sampler achieves $\\textbf{quadratic}$ convergence. We further adapt the sampler to practical settings by incorporating posterior distributions and logit mixing, demonstrating improvements over previous methods. For example, it outperforms vanilla DPO by over $7.4$% on Safe-RLHF dataset. Our results not only offer insights into the theoretical understanding of DPO but also pave the way for further algorithm designs.", "keywords": [ "direct preference optimization", "online DPO", "tabular softmax policy" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "F6s7OApF0n", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Aoxuan Silvia Zhang", "gender": "Female", "institution": "Korea University", "country": "KR", "position": "Undergrad student" }, { "name": "Byungjoo Kim", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Hae Beom Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Postdoc" }, { "name": "Juho Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Junhyeon Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Sung Ju Hwang", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" }, { "name": "DONGBOK LEE", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 7, "n_ref_uni": 57, "n_ref": 152, "n_ref_all": 218, "n_fig": 19, "n_tab": 6, "L_tab": 2237, "n_element_tab": 313, "n_fig_1": 13, "n_tab_1": 5, "L_tab_1": 1602, "n_element_tab_1": 221, "formula_len_all": 1727, "formula_len_all_1": 579, "len_all": 202668, "len_all_1": 74663, "len_abs": 1195, "len_title": 100, "len_sents": 63315, "len_sents_1": 34335, "n_sents": 467, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1313, "title": "Cost-Sensitive Multi-Fidelity Bayesian Optimization", "abs": "In this paper, we address the problem of cost-sensitive multi-fidelity Bayesian Optimization (BO) for efficient hyperparameter optimization (HPO). Specifically, we assume a scenario where users want to early-stop the BO when performance increase is not satisfactory with respect to the required computational cost. Motivated by this scenario, we introduce \\emph{utility function}, which is predefined by each user and describes the trade-off between the required BO steps and the cumulative best performance during the BO. This utility function, combined with our novel acquisition function and the stopping criteria, allows us to dynamically choose for each BO step the best configuration that we expect to achieve the maximum utility in future, and also automatically stop the BO around the maximum utility. Further, we improve the sample efficiency of existing learning curve (LC) extrapolation methods (e.g., Prior Fitted Networks) with transfer learning, while successfully capturing the correlations between different configurations to develop a sensible surrogate function for multi-fidelity BO. We validate our algorithm on various LC datasets and found it outperform all the previous multi-fidelity BO baselines, achieving significantly better trade-off between cost and performance of multi-fidelity BO.", "keywords": [ "gray-box hyperparameter optimization", "multi-fidelity hyperparameter optimization", "cost-sensitive Bayesian optimization", "learning curve extrapolation", "transfer learning" ], "rating_list": [ 6, 5, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 1, 3, 1 ], "contribution_list": [ 4, 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 3, 3, 5 ] }, { "paper_id": "F6rZaxOC6m", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Quanyu Dai", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Rui Li", "gender": "unknown", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Xu Chen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zeyu Zhang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Zhenhua Dong", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 54, "n_ref": 113, "n_ref_all": 136, "n_fig": 5, "n_tab": 6, "L_tab": 2536, "n_element_tab": 159, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 466, "n_element_tab_1": 42, "formula_len_all": 898, "formula_len_all_1": 295, "len_all": 182672, "len_all_1": 69722, "len_abs": 1547, "len_title": 131, "len_sents": 48919, "len_sents_1": 34759, "n_sents": 321, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1552, "title": "KnowTrace: Explicit Knowledge Tracing for Structured Retrieval-Augmented Generation", "abs": "Recent advances in retrieval-augmented generation (RAG) furnish large language models (LLMs) with iterative retrievals of relevant information to strengthen their capabilities in addressing complex multi-hop questions. However, these methods typically accumulate the retrieved natural language text into LLM prompts, imposing an increasing burden on the LLM to grasp the underlying knowledge structure for high-quality multi-step reasoning. Despite a few attempts to reduce this burden by restructuring all retrieved passages or even entire external corpora, these efforts are afflicted with significant restructuring overhead and potential knowledge loss. To tackle this challenge, we introduce a new structured paradigm (KnowTrace) from the perspective of explicit knowledge tracing, which treats LLM as an agent to progressively acquire desired knowledge triplets during iterative retrievals and ultimately trace out a specific knowledge graph conditioned on the input question. This paradigm clearly unveils the logical relationships behind the unstructured text and thus can directly facilitate LLM\u2019s inference. Notably, it also naturally inspires a reflective mechanism of knowledge backtracing to identify supportive evidence and filter out useless retrievals in the correct trajectories, thus offering an effective way to stimulate LLM\u2019s self-taught finetuning. Extensive experiments demonstrate the superiority of our paradigm over three standard multi-hop question answering benchmarks. Our code is available at https://github.com/xxrep/SRAG.", "keywords": [ "Knowledge Graph", "Retrieval-Augmented Generation", "Multi-Hop Question Answering", "Multi-Step Reasoning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "F6h0v1CTpC", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Jiaming Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jiaxin Ge", "gender": "Female", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Jingtong Hu", "gender": "Male", "institution": "University of Pittsburgh", "country": "US", "position": "Associate Professor" }, { "name": "Qiang Zhou", "gender": "Male", "institution": "University of Pittsburgh", "country": "US", "position": "PhD student" }, { "name": "Qiuxuan Feng", "gender": "Female", "institution": "Tianjin University", "country": "CN", "position": "Undergrad student" }, { "name": "Xinyan Chen", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Yizhou Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Ryan Dai", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Shanghang Zhang", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 41, "n_ref": 70, "n_ref_all": 120, "n_fig": 28, "n_tab": 9, "L_tab": 5588, "n_element_tab": 342, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1726, "n_element_tab_1": 176, "formula_len_all": 258, "formula_len_all_1": 0, "len_all": 244735, "len_all_1": 70893, "len_abs": 1191, "len_title": 67, "len_sents": 80943, "len_sents_1": 31527, "n_sents": 745, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1627, "title": "EmpathyRobot: A Dataset and Benchmark for Empathetic Task Planning of Robotic Agent", "abs": "Empathy is a fundamental instinct and essential need for humans, as they both demonstrate empathetic actions toward others and receive empathetic support. As robots become increasingly integrated into daily life, it is essential to explore whether they can provide human-like empathetic support. Although existing emotion agents have explored how to understand humans' empathetic needs, they lack to further enable robots to generate empathy-oriented task planning, neglecting the evaluation of empathetic behaviors. To address this gap, we introduce \\textbf{EmpathyRobot}, the first dataset specifically designed to benchmark and enhance the empathetic actions of agents across diverse scenarios. This dataset contains 10,000 samples based on human feedback, encompassing information from various modalities and corresponding empathetic task planning sequences, including navigation and manipulation. Agents are required to perform actions based on their understanding of both the visual scene and human emotions. To systematically evaluate the performance of existing agents on the EmpathyRobot dataset, we conduct comprehensive experiments to test the most capable models. Our findings reveal that generating accurate empathetic actions remains a significant challenge. Meanwhile, we finetune an \\ac{llm} on our benchmark, demonstrating that it can effectively be used to enhance the empathetic behavior of robot agents. By establishing a standard benchmark for evaluating empathetic actions, we aim to drive advancements in the study and pursue of empathetic behaviors in robot agents. We will release our code and dataset.", "keywords": [ "empathy", "robot planning", "large language models" ], "rating_list": [ 8, 8, 3, 5 ], "soundness_list": [ 3, 4, 2, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 4, 1, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "F6SaYwJ3eV", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Jianfeng Lu", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Professor" }, { "name": "Matthew Repasky", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Qiang Qiu", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "Vishal Purohit", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Xiuyuan Cheng", "gender": "unknown", "institution": "Duke University", "country": "US", "position": "Associate Professor" }, { "name": "Yao Xie", "gender": "Female", "institution": "Georgia Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 12, "n_ref_uni": 37, "n_ref": 129, "n_ref_all": 146, "n_fig": 5, "n_tab": 8, "L_tab": 2022, "n_element_tab": 173, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1446, "n_element_tab_1": 138, "formula_len_all": 1552, "formula_len_all_1": 569, "len_all": 152542, "len_all_1": 61291, "len_abs": 1504, "len_title": 115, "len_sents": 45517, "len_sents_1": 27217, "n_sents": 339, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1506, "title": "Posterior sampling via Langevin dynamics based on generative priors", "abs": "Posterior sampling in high-dimensional spaces using generative models holds significant promise for various applications, including but not limited to inverse problems and guided generation tasks. Despite many recent developments, generating diverse posterior samples remains a challenge, as existing methods require restarting the entire generative process for each new sample, making the procedure computationally expensive. In this work, we propose efficient posterior sampling by simulating Langevin dynamics in the noise space of a pre-trained generative model. By exploiting the mapping between the noise and data spaces which can be provided by distilled flows or consistency models, our method enables seamless exploration of the posterior without the need to re-run the full sampling chain, drastically reducing computational overhead. Theoretically, we prove a guarantee for the proposed noise-space Langevin dynamics to approximate the posterior, assuming that the generative model sufficiently approximates the prior distribution. Our framework is experimentally validated on image restoration tasks involving noisy linear and nonlinear forward operators applied to LSUN-Bedroom (256 x 256) and ImageNet (64 x 64) datasets. The results demonstrate that our approach generates high-fidelity samples with enhanced semantic diversity even under limited number of function evaluations, offering superior efficiency and performance compared to existing diffusion-based posterior sampling techniques.", "keywords": [ "Posterior Sampling", "Inverse Problems", "Consistency Models" ], "rating_list": [ 3, 6, 3, 3, 3 ], "soundness_list": [ 2, 3, 1, 2, 3 ], "presentation_list": [ 2, 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "F64wTvQBum", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Adel Bibi", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Senior Researcher" }, { "name": "Alasdair Paren", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Cornelius Emde", "gender": "Not Specified", "institution": "University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Maxime Guillaume Kayser", "gender": "Male", "institution": "Memorial Sloan Kettering Cancer Centre", "country": "US", "position": "Researcher" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Preetham Arvind", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "MS student" }, { "name": "Thomas Lukasiewicz", "gender": "unknown", "institution": "Institute of Logic and Computation, Technische Universit\u00e4t Wien", "country": "", "position": "Full Professor" }, { "name": "Tom Rainforth", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "Assistant Professor" } ], "n_formula": 42, "n_formula_1": 9, "n_ref_uni": 57, "n_ref": 91, "n_ref_all": 145, "n_fig": 24, "n_tab": 4, "L_tab": 775, "n_element_tab": 49, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 290, "n_element_tab_1": 25, "formula_len_all": 1880, "formula_len_all_1": 504, "len_all": 285236, "len_all_1": 83900, "len_abs": 760, "len_title": 97, "len_sents": 69383, "len_sents_1": 37906, "n_sents": 638, "n_sents_1": 302, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 49, "L_abs": 763, "title": "Shh, don't say that! Domain Certification in LLMs", "abs": "Large language models (LLMs) are often deployed to do constrained tasks, with narrow domains. For example, customer support bots can be built on top of LLMs, relying on their broad language understanding and capabilities to enhance performance. However, these LLMs are adversarially susceptible, potentially generating outputs outside the intended domain. To formalize, assess and mitigate this risk, we introduce \\emph{domain certification}; a guarantee that accurately characterizes the out-of-domain behavior of language models. We then propose a simple yet effective approach dubbed VALID that provides adversarial bounds as a certificate. Finally, we evaluate our method across a diverse set of datasets, demonstrating that it yields meaningful certificates.", "keywords": [ "large language model", "natural language processing", "adversarial robustness", "adversary", "natural text generation", "certification", "verification" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 4 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "F61IzZl5jw", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Chinmay Hegde", "gender": "Male", "institution": "New York University", "country": "US", "position": "Associate Professor" }, { "name": "Gowthami Somepalli", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Minh Pham", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Nicky Kriplani", "gender": "Male", "institution": "New York University", "country": "US", "position": "Undergrad student" }, { "name": "Niv Cohen", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 39, "n_ref": 92, "n_ref_all": 113, "n_fig": 6, "n_tab": 5, "L_tab": 1089, "n_element_tab": 100, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 833, "n_element_tab_1": 84, "formula_len_all": 557, "formula_len_all_1": 164, "len_all": 146082, "len_all_1": 54678, "len_abs": 860, "len_title": 109, "len_sents": 39686, "len_sents_1": 24590, "n_sents": 303, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 61, "L_abs": 890, "title": "SolidMark: Evaluating Image Memorization in Generative Models", "abs": "Recent works have shown that diffusion models are able to memorize training images and emit them at generation time. However, the metrics used to evaluate memorization and its mitigation techniques suffer from dataset-dependent biases and struggle to detect whether a given specific image has been memorized or not.\n\nThis paper begins with a comprehensive exploration of issues surrounding memorization metrics in diffusion models. Then, to mitigate these issues, we introduce SolidMark, a novel evaluation method that provides a per-image memorization score. We then re-evaluate existing memorization mitigation techniques and show that SolidMark is capable of evaluating fine-grained pixel-level memorization. Finally, we release a text-to-image model pretrained from scratch based on SolidMark to facilitate further research for understanding memorization phenomena in generative models.", "keywords": [ "Memorization", "Diffusion Models", "Metrics" ], "rating_list": [ 5, 1, 5, 5 ], "soundness_list": [ 2, 1, 1, 3 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 2, 1, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "F5nWSf9etp", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Anirudhan Badrinath", "gender": "unknown", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "MS student" }, { "name": "Jiajing Xu", "gender": "Male", "institution": "Pinterest", "country": "", "position": "Engineering" }, { "name": "Prabhat Agarwal", "gender": "Male", "institution": "Pinterest, Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 32, "n_formula_1": 12, "n_ref_uni": 38, "n_ref": 117, "n_ref_all": 156, "n_fig": 5, "n_tab": 10, "L_tab": 2247, "n_element_tab": 145, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 2776, "n_element_tab_1": 157, "formula_len_all": 3068, "formula_len_all_1": 933, "len_all": 169662, "len_all_1": 63426, "len_abs": 1145, "len_title": 147, "len_sents": 61061, "len_sents_1": 28518, "n_sents": 403, "n_sents_1": 165, "venue": "ICLR2025", "decision": "withdraw", "L_title": 99, "L_abs": 1148, "title": "Hybrid Preference Optimization: Augmenting Direct Preference Optimization with Auxiliary Objectives", "abs": "For aligning large language models (LLMs), prior work has leveraged reinforcement learning via human feedback (RLHF) or variations of direct preference optimization (DPO). While DPO offers a simpler framework based on maximum likelihood estimation, it compromises on the ability to tune language models to easily maximize non-differentiable objectives according to the LLM designer's preferences (e.g., using simpler language or minimizing specific kinds of harmful content). These may neither align with user preferences nor even be able to be captured tractably by binary preference data. To leverage the simplicity and performance of DPO with the generalizability of RL, we propose a hybrid approach between DPO and RLHF. With a simple augmentation to the implicit reward decomposition of DPO, we allow for tuning LLMs to maximize a set of arbitrary auxiliary rewards using offline RL. The proposed method, Hybrid Preference Optimization (HPO), shows the ability to effectively generalize to both user preferences and auxiliary designer objectives, while preserving alignment performance across a range of challenging benchmarks and model sizes.", "keywords": [ "large language models", "alignment", "reinforcement learning", "direct preference optimization" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 4, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 1, 3 ] }, { "paper_id": "F5lcN7329a", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Gang Pan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Professor" }, { "name": "Puli Wang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Yu Qi", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "Tenure-track Research Professor" }, { "name": "Yueming Wang", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "Full Professor" } ], "n_formula": 35, "n_formula_1": 15, "n_ref_uni": 62, "n_ref": 116, "n_ref_all": 166, "n_fig": 14, "n_tab": 12, "L_tab": 4443, "n_element_tab": 358, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3213, "formula_len_all_1": 1103, "len_all": 200575, "len_all_1": 67068, "len_abs": 2144, "len_title": 134, "len_sents": 50779, "len_sents_1": 29896, "n_sents": 402, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1188, "title": "Consistent Neural Embeddings through Flow Matching on Attractor-like Neural Manifolds", "abs": "The primary objective of brain-computer interfaces (BCIs) is to establish a direct connection between neural activity and behavioral actions through neural decoders. \nConsistent neural representation is crucial for achieving high-performance behavioral decoding over time.\nDue to the stochastic variability in neural recordings, existing neural representation techniques yield dynamical instability, leading to the failure of behavioral decoders in few-trial scenarios.\nIn this work, we propose a novel Flow-Based Dynamical Alignment (FDA) framework that leverages attractor-like ensemble dynamics on stable neural manifolds, which facilitate a new source-free alignment through likelihood maximization.\nThe consistency of latent embeddings obtained through FDA was theoretically verified based on dynamical stability, allowing for rapid adaptation with few trials.\nFurther experiments on multiple motor cortex datasets validate the superior performance of FDA.\nThe FDA method establishes a novel framework for consistent neural latent embeddings with few trials. \nOur work offers insights into neural dynamical stability, potentially enhancing the chronic reliability of real-world BCIs.", "keywords": [ "Brain-Computer Interface", "Neural Decoding", "Flow Matching", "Dynamical Stability" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "F5R0lG74Tu", "primary_area": "foundation or frontier models, including LLMs", "n_author": 11, "author_info_list": [ { "name": "Chujie Gao", "gender": "Female", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Intern" }, { "name": "Dongping Chen", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Intern" }, { "name": "Jianfeng Gao", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Lichao Sun", "gender": "Male", "institution": "Lehigh University", "country": "US", "position": "Assistant Professor" }, { "name": "Qihui Zhang", "gender": "Male", "institution": "The University of Queensland", "country": "AU", "position": "Intern" }, { "name": "Siyuan Wu", "gender": "unknown", "institution": "University of Waterloo", "country": "CA", "position": "Intern" }, { "name": "Tianyi Zhou", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Xiangliang Zhang", "gender": "Female", "institution": "University of Notre Dame", "country": "US", "position": "Associate Professor" }, { "name": "Yao Wan", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Yue Huang", "gender": "unknown", "institution": "University of Notre Dame", "country": "US", "position": "PhD student" }, { "name": "chaowei xiao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 67, "n_ref": 169, "n_ref_all": 206, "n_fig": 14, "n_tab": 11, "L_tab": 4902, "n_element_tab": 1068, "n_fig_1": 9, "n_tab_1": 7, "L_tab_1": 2201, "n_element_tab_1": 352, "formula_len_all": 77, "formula_len_all_1": 0, "len_all": 313825, "len_all_1": 67607, "len_abs": 1555, "len_title": 108, "len_sents": 65739, "len_sents_1": 29017, "n_sents": 542, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1563, "title": "DataGen: Unified Synthetic Dataset Generation via Large Language Models", "abs": "Large Language Models (LLMs) such as GPT-4 and Llama3 have significantly impacted various fields by enabling high-quality synthetic data generation and reducing dependence on expensive human-generated datasets. \nDespite this, challenges remain in the areas of generalization, controllability, diversity, and truthfulness within the existing generative frameworks. To address these challenges, this paper presents DataGen, a comprehensive LLM-powered framework designed to produce diverse, accurate, and highly controllable datasets. DataGen is adaptable, supporting all types of text datasets and enhancing the generative process through innovative mechanisms. To augment data diversity, DataGen incorporates an attribute-guided generation module and a group checking feature. For accuracy, it employs a code-based mathematical assessment for label verification alongside a retrieval-augmented generation technique for factual validation. The framework also allows for user-specified constraints, enabling customization of the data generation process to suit particular requirements. Extensive experiments demonstrate the superior quality of data generated by DataGen, and each module within DataGen plays a critical role in this enhancement. Additionally, DataGen is applied in two practical scenarios: benchmarking LLMs and data augmentation. The results indicate that DataGen effectively supports dynamic and evolving benchmarking and that data augmentation improves LLM capabilities in various domains, including agent-oriented abilities and reasoning skills.", "keywords": [ "large language model", "evaluation", "synthetic data" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "F5PlYMC5ik", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Chao Deng", "gender": "Male", "institution": "China Mobile Research Institute", "country": "", "position": "Researcher" }, { "name": "Junlan Feng", "gender": "Female", "institution": "China Mobile", "country": "CN", "position": "Principal Researcher" }, { "name": "Qian Hu", "gender": "unknown", "institution": "CMRI", "country": "", "position": "Researcher" }, { "name": "Xue Han", "gender": "Female", "institution": "China Mobile Communications Company Limited Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Yitong Wang", "gender": "unknown", "institution": "China Mobile Research Institute", "country": "CN", "position": "Researcher" }, { "name": "wenchun.gao", "gender": "Male", "institution": "China Mobile Communications Company Limited Research Institute", "country": "", "position": "Researcher" } ], "n_formula": 22, "n_formula_1": 19, "n_ref_uni": 39, "n_ref": 88, "n_ref_all": 109, "n_fig": 6, "n_tab": 14, "L_tab": 3863, "n_element_tab": 638, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1755, "n_element_tab_1": 282, "formula_len_all": 2275, "formula_len_all_1": 1995, "len_all": 162467, "len_all_1": 67168, "len_abs": 1720, "len_title": 87, "len_sents": 45091, "len_sents_1": 27209, "n_sents": 360, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1729, "title": "LOIRE: LifelOng learning on Incremental data via pre-trained language model gRowth Efficiently", "abs": "Large-scale pre-trained language models (PLMs) require significant computational resources to train from scratch on large volumes of data. But in the real world, emerging data from diverse sources may not be initially available for pre-training. Recent studies on lifelong learning have tried to solve this problem by exploring the use of model growth techniques to effectively incorporate new knowledge without the need for complete re-training. However, model growth approaches utilized have issues with growth operators that do not ensure strict function preservation or growth schedules that only include a few growth dimensions, reducing lifelong learning's effect. Furthermore, existing approaches often assume that emerging data has the same distribution as pre-training data, causing catastrophic forgetting of previously acquired knowledge. To address the aforementioned issues, we introduce LOIRE, a framework for lifelong learning that enables PLMs to effectively grow their capacity using incremental data. LOIRE employs growth operators for all feasible dimensions and a growth schedule to generate the optimal expansion sequence in the field of lifelong learning. Specifically, we present a novel plug-in layer growth operator with residual connections that skip the newly added layer during initial training while ensuring function preservation. We additionally propose an iterative distillation strategy for LOIRE that allows an intermediate model in the growth stages to switch between being a student and a teacher, reducing catastrophic forgetting during growth. Experiments show that LOIRE can reduce computational expenses by an average of 29.22\\% while retaining equivalent or better downstream performance.", "keywords": [ "Lifelong learning", "Model growth", "Function-preserving", "Efficient pre-training" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 1, 3 ] }, { "paper_id": "F57HPKZ6KD", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Fuyou Miao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Xu Wang", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yan Xiong", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "wenjie liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 35, "n_formula_1": 19, "n_ref_uni": 62, "n_ref": 121, "n_ref_all": 168, "n_fig": 3, "n_tab": 10, "L_tab": 6460, "n_element_tab": 1666, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 882, "n_element_tab_1": 214, "formula_len_all": 2882, "formula_len_all_1": 1740, "len_all": 204360, "len_all_1": 67440, "len_abs": 1689, "len_title": 133, "len_sents": 49121, "len_sents_1": 27137, "n_sents": 448, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1711, "title": "Efficient and Robust Neural Combinatorial Optimization via Wasserstein-Based Coresets", "abs": "Combinatorial optimization (CO) is a fundamental tool in many fields. \nMany neural combinatorial optimization (NCO) methods have been proposed to solve CO problems.\nHowever, existing NCO methods typically require significant computational and storage resources, and face challenges in maintaining robustness to distribution shifts between training and test data.\nTo address these issues, we model CO instances into probability measures, and introduce Wasserstein-based metrics to quantify the difference between CO instances. \nWe then leverage a popular data compression technique, \\emph{coreset}, to construct a small-size proxy for the original large dataset.\nHowever, the time complexity of constructing a coreset is linearly dependent on the size of the dataset. Consequently, it becomes challenging when datasets are particularly large.\nFurther, we accelerate the coreset construction by adapting it to the merge-and-reduce framework, enabling parallel computing. Additionally, we prove that our coreset is a good representation in theory.\n{Subsequently}, to speed up the training process for existing NCO methods, we propose an efficient training framework based on the coreset technique. We train the model on a small-size coreset rather than on the full dataset, and thus save substantial computational and storage resources. Inspired by hierarchical Gonzalez\u2019s algorithm, our coreset method is designed to capture the diversity of the dataset, which consequently improves robustness to distribution shifts.\nFinally, experimental results demonstrate that our training framework not only enhances robustness to distribution shifts but also achieves better performance with reduced resource requirements.", "keywords": [ "Neural Combinatorial Optimization", "Wasserstein-Based Metric", "Coreset", "Data compression" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "F52tAK5Gbg", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Andres Munoz medina", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Weiwei Kong", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Monica Ribero Diaz", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 38, "n_formula_1": 23, "n_ref_uni": 39, "n_ref": 63, "n_ref_all": 78, "n_fig": 4, "n_tab": 4, "L_tab": 965, "n_element_tab": 102, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 425, "n_element_tab_1": 64, "formula_len_all": 3635, "formula_len_all_1": 1743, "len_all": 123960, "len_all_1": 61992, "len_abs": 1244, "len_title": 124, "len_sents": 33903, "len_sents_1": 25816, "n_sents": 295, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1265, "title": "Differentially private optimization for non-decomposable objective functions", "abs": "Unsupervised pre-training is a common step in developing computer vision models and large language models. In this setting, the absence of labels requires the use of similarity-based loss functions, such as the contrastive loss, that favor minimizing the distance between similar inputs and maximizing the distance between distinct inputs. As privacy concerns mount, training these models using differential privacy has become more important. However, due to how inputs are generated for these losses, one of their undesirable properties is that their $L_2$ sensitivity grows with the batch size. This property is particularly disadvantageous for differentially private training methods, such as DP-SGD. To overcome this issue, we develop a new DP-SGD variant for similarity based loss functions --- in particular, the commonly-used contrastive loss --- that manipulates gradients of the objective function in a novel way to obtain a sensitivity of the summed gradient that is $O(1)$ for batch size $n$. We test our DP-SGD variant on some CIFAR-10 pre-training and CIFAR-100 finetuning tasks and show that, in both tasks, our method's performance comes close to that of a non-private model and generally outperforms DP-SGD applied directly to the contrastive loss.", "keywords": [ "Differential privacy", "private learning", "contrastive learning." ], "rating_list": [ 5, 3, 5, 1, 6 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "F4meTCwlxZ", "primary_area": "causal reasoning", "n_author": 5, "author_info_list": [ { "name": "Chen Wang", "gender": "Male", "institution": "CSIRO", "country": "AU", "position": "Principal Researcher" }, { "name": "Lina Yao", "gender": "Female", "institution": "University of New South Wales", "country": "AU", "position": "Full Professor" }, { "name": "Yidong Gan", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Yipeng Zhang", "gender": "Male", "institution": "CSIRO", "country": "AU", "position": "Researcher" }, { "name": "Yuzhe Zhang", "gender": "Male", "institution": "CSIRO", "country": "AU", "position": "Postdoc" } ], "n_formula": 18, "n_formula_1": 14, "n_ref_uni": 41, "n_ref": 76, "n_ref_all": 91, "n_fig": 3, "n_tab": 3, "L_tab": 736, "n_element_tab": 52, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 667, "n_element_tab_1": 40, "formula_len_all": 2606, "formula_len_all_1": 2598, "len_all": 190193, "len_all_1": 70962, "len_abs": 3578, "len_title": 120, "len_sents": 74295, "len_sents_1": 33580, "n_sents": 489, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 990, "title": "Consistency Guaranteed Causal Graph Recovery with Large Language Models", "abs": "Causal graph recovery traditionally relies on statistical estimation of observable variables or individual knowledge, which suffer from data collection biases and knowledge limitations of individuals. Leveraging the broad knowledge in scientific corpus, we propose a novel method for causal graph recovery to deduce causal relationships with the large language models (LLMs) as a knowledge extractor. Our method extracts associational relationships among variables and further eliminates the inconsistent relationship to recover a causal graph using the constraint-based causal discovery methods. Comparing to other LLM-based methods that directly instruct LLMs to do highly complex causal reasoning, our method shows advantages on causal graph quality on benchmark datasets. More importantly, as causal graphs may evolve when new research results emerge, our method shows sensitivity to new evidence in the literature and can provide useful information to update causal graphs accordingly.", "keywords": [ "Causal discovery", "causal reasoning", "large language models", "knowledge extraction" ], "rating_list": [ 6, 6, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 3, 1 ], "presentation_list": [ 2, 3, 2, 2, 4 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 2, 4 ] }, { "paper_id": "F4f1afsm3R", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Aiwei Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Boye Niu", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "Undergrad student" }, { "name": "Haotian Xu", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Hongzhang Liu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Lijie Wen", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Xuming Hu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" }, { "name": "Xuzheng He", "gender": "Male", "institution": "Peking University", "country": "", "position": "Undergrad student" }, { "name": "Zitian Gao", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Undergrad student" } ], "n_formula": 20, "n_formula_1": 11, "n_ref_uni": 24, "n_ref": 57, "n_ref_all": 69, "n_fig": 5, "n_tab": 10, "L_tab": 4294, "n_element_tab": 191, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1872, "n_element_tab_1": 68, "formula_len_all": 1730, "formula_len_all_1": 786, "len_all": 147266, "len_all_1": 56957, "len_abs": 1361, "len_title": 107, "len_sents": 42847, "len_sents_1": 25657, "n_sents": 303, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1406, "title": "Interpretable Contrastive Monte Carlo Tree Search Reasoning", "abs": "We propose $\\textbf{(S)}peculative \\textbf{(C)}ontrastive$ $\\textbf{MCTS}^\\mathbf{*}$: a novel Monte Carlo Tree Search (MCTS) reasoning algorithm for Large Language Models (LLMs) which significantly improves both reasoning accuracy and speed. Our motivation comes from: 1. Previous MCTS LLM reasoning works often overlooked its biggest drawback\u2014slower speed compared to CoT; 2. Previous research mainly used MCTS as a tool for LLM reasoning on various tasks with limited quantitative analysis or ablation studies of its components from reasoning interpretability perspective. 3. The reward model is the most crucial component in MCTS, however previous work has rarely conducted in-depth study or improvement of MCTS's reward models. Thus, we conducted extensive ablation studies and quantitative analysis on components of MCTS, revealing the impact of each component on the MCTS reasoning performance of LLMs. Building on this, (i) we designed a highly interpretable reward model based on the principle of contrastive decoding and (ii) achieved an average speed improvement of 51.9\\% per node using speculative decoding. Additionally, (iii) we improved UCT node selection strategy and backpropagation used in previous works, resulting in significant performance improvement. We outperformed o1-mini by an average of 17.4\\% on the Blocksworld multi-step reasoning dataset using Llama-3.1-70B with SC-MCTS\\*.", "keywords": [ "Monte Carlo Tree Search", "Large Language Models", "Multi-Step Reasoning" ], "rating_list": [ 8, 3, 3, 6, 3 ], "soundness_list": [ 3, 2, 2, 2, 1 ], "presentation_list": [ 3, 2, 2, 2, 1 ], "contribution_list": [ 4, 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 2, 3 ] }, { "paper_id": "F4bHMojXVW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Elias Stengel-Eskin", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Postdoc" }, { "name": "Feng Cheng", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Gedas Bertasius", "gender": "Male", "institution": "University of North Carolina, Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Jaehong Yoon", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Postdoc" }, { "name": "Mohit Bansal", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" }, { "name": "Shoubin Yu", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Ziyang Wang", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 40, "n_ref": 126, "n_ref_all": 141, "n_fig": 9, "n_tab": 10, "L_tab": 3252, "n_element_tab": 296, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1835, "n_element_tab_1": 41, "formula_len_all": 86, "formula_len_all_1": 85, "len_all": 196549, "len_all_1": 51579, "len_abs": 1801, "len_title": 133, "len_sents": 47676, "len_sents_1": 25175, "n_sents": 345, "n_sents_1": 148, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1809, "title": "VideoTree: Adaptive Tree-based Video Representation for LLM Reasoning on Long Videos", "abs": "Long-form video understanding has been a challenging task due to the high redundancy in video data and the abundance of query-irrelevant information. To tackle this challenge, we propose VideoTree, a training-free framework which builds a query-adaptive and hierarchical video representation for LLM reasoning over long-form videos. First, VideoTree extracts query-relevant information from the input video through an iterative process, progressively refining the selection of keyframes based on their relevance to the query. Furthermore, VideoTree leverages the inherent hierarchical structure of long video data, which is often overlooked by existing LLM-based methods. Specifically, we incorporate multigranularity information into a tree-based representation, allowing VideoTree to extract query-relevant details from long videos in a coarse-to-fine manner. This enables the model to effectively handle a wide range of video queries with varying levels of detail. Finally, VideoTree aggregates the hierarchical query-relevant information within the tree structure and feeds it into an LLM reasoning model to answer the query. Our experiments show that our training-free method improves both reasoning accuracy and efficiency compared to existing methods. Specifically, VideoTree outperforms the existing training-free approaches on the popular EgoSchema and NExT-QA benchmarks with less inference time, achieving 61.1% and 75.6% accuracy on the test set without additional video-specific training. Moreover, on the long split of Video-MME benchmark (average 44 minutes), the training-free VideoTree framework achieves better performance than the strong proprietary GPT-4V model and other MLLMs that were extensively trained on video data. Our code is provided in the supplementary and will be made public.", "keywords": [ "Long Video Understanding", "Video-lanaguage Understanding", "Multimodal Learning", "LLM-based Video Understanding" ], "rating_list": [ 5, 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3, 4 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 5, 3 ] }, { "paper_id": "F4IMiNhim1", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Bing Su", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Chuan Cao", "gender": "Female", "institution": "Research, Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Zhao Yang", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 7, "n_ref_uni": 48, "n_ref": 116, "n_ref_all": 147, "n_fig": 11, "n_tab": 17, "L_tab": 3631, "n_element_tab": 319, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 145, "n_element_tab_1": 32, "formula_len_all": 697, "formula_len_all_1": 408, "len_all": 211957, "len_all_1": 60835, "len_abs": 1623, "len_title": 106, "len_sents": 63975, "len_sents_1": 28966, "n_sents": 477, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 1629, "title": "Regulatory DNA Sequence Design with Reinforcement Learning", "abs": "Cis-regulatory elements (CREs), such as promoters and enhancers, are relatively short DNA sequences that directly regulate the expression of specific genes. The fitness of CREs, i.e., their functionality to enhance gene expression, highly depend on its nucleotide sequence, especially the composition of some special motifs known as transcription factor binding sites (TFBSs). Designing CREs to optimize their fitness is crucial for therapeutic and bioengineering applications. Existing CRE design methods often rely on simple strategies, such as iteratively introducing random mutations and selecting variants with high fitness from a large number of candidates through an oracle, i.e., a pre-trained gene expression prediction model. Due to the vast search space and lack of prior biological knowledge guidance, these methods are prone to getting trapped in local optima and tend to produce CREs with low diversity. In this paper, we propose the first method that leverages reinforcement learning (RL) to fine-tune a pre-trained autoregressive (AR) generative model for designing high-fitness cell-type-specific CREs while maintaining sequence diversity. We employ prior knowledge of CRE regulatory mechanisms to guide the optimization by incorporating the role of TFBSs into the RL process. In this way, our method encourages the removal of repressor motifs and the addition of activator motifs. We evaluate our method on enhancer design tasks for three distinct human cell types and promoter design tasks in two different yeast media conditions, demonstrating its effectiveness and robustness in generating high-fitness CREs.", "keywords": [ "dna optimize", "sequence optimize", "autoregressive generative models", "ai4science" ], "rating_list": [ 5, 8, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "F3Migaak2i", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Jingbo Shang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Weitang Liu", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Ying Wai Li", "gender": "Female", "institution": "Los Alamos National Laboratory", "country": "", "position": "Researcher" }, { "name": "Yuelei Li", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Undergrad student" }, { "name": "Zihan Wang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 11, "n_ref_uni": 33, "n_ref": 69, "n_ref_all": 98, "n_fig": 13, "n_tab": 2, "L_tab": 72, "n_element_tab": 10, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1064, "formula_len_all_1": 484, "len_all": 135756, "len_all_1": 65859, "len_abs": 1500, "len_title": 126, "len_sents": 44347, "len_sents_1": 31714, "n_sents": 386, "n_sents_1": 265, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1511, "title": "Model-diff: A Tool for Comparative Study of Language Models in the Input Space", "abs": "Comparing two (large) language models (LMs) side-by-side and pinpointing their prediction similarities and differences on the same set of inputs are crucial in many real-world scenarios, e.g., one can test if a licensed model was potentially plagiarized by another.\nTraditional analysis compares the LMs' outputs on some benchmark datasets, which only cover a limited number of inputs of designed perspectives for the intended applications.\nThe benchmark datasets cannot prepare data to cover the test cases from unforeseen perspectives which can help us understand differences between models unbiasedly. \nIn this paper, we propose a new model comparative analysis setting that considers a large input space where brute-force enumeration would be infeasible. \nThe input space can be simply defined as all token sequences that a LM would produce low perplexity on --- we follow this definition in the paper as it would produce the most human-readable inputs. \nWe propose a novel framework Model-diff that uses text generation by sampling and deweights the histogram of sampling statistics to estimate prediction differences between two LMs in this input space efficiently and unbiasedly.\nModel-diff achieves this by drawing and counting the inputs at each prediction difference value in negative log-likelihood.\nExperiments reveal for the first time the quantitative prediction differences between LMs in a large input space, potentially facilitating the model analysis for applications such as model plagiarism.", "keywords": [ "prediction difference; input space;" ], "rating_list": [ 1, 5, 3, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "F1cN3aoAty", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Dhiman Paul", "gender": "Male", "institution": "North South University", "country": "BD", "position": "MS student" }, { "name": "Md Rizwan Parvez", "gender": "Male", "institution": "Qatar Computing Research Institute", "country": "QA", "position": "Researcher" }, { "name": "Nabeel Mohammed", "gender": "Male", "institution": "North South University", "country": "BD", "position": "Associate Professor" }, { "name": "Shafin Rahman", "gender": "Male", "institution": "North South University", "country": "BD", "position": "Assistant Professor" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 54, "n_ref": 115, "n_ref_all": 135, "n_fig": 5, "n_tab": 13, "L_tab": 4996, "n_element_tab": 467, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 3837, "n_element_tab_1": 310, "formula_len_all": 1049, "formula_len_all_1": 576, "len_all": 166799, "len_all_1": 61483, "len_abs": 951, "len_title": 160, "len_sents": 35033, "len_sents_1": 24288, "n_sents": 258, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 112, "L_abs": 960, "title": "VideoLights: A Cross-Modal Cross-Task Transformer Model for Joint Video Highlight Detection and Moment Retrieval", "abs": "Video Highlight Detection and Moment Retrieval (HD/MR) are essential in video analysis. Recent joint prediction transformer models often overlook cross-task dynamics and video-text alignment. We propose VideoLights, a novel HD/MR framework addressing these limitations through: (i) Convolutional Projection and Feature Refinement modules with an intermodal alignment loss for better video-text feature alignment. (ii) Bi-Directional Cross-Modal Fusion network for strongly coupled query-aware clip representations. (iii) Uni-Directional joint-task feedback mechanism enhancing both tasks through correlation. In addition, we introduce hard positive/negative losses for adaptive error penalization and improved learning. Our approach includes intelligent pretraining and finetuning using synthetic data and features from various encoders. Comprehensive experiments on QVHighlights, TVSum, and Charades-STA benchmarks demonstrate state-of-the-art performance.", "keywords": [ "highlight detection", "moment retrieval", "video grounding" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 1, 3, 2, 2 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 5, 2, 5, 4 ] }, { "paper_id": "F1OdjlfCLS", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Chao Wu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Guanzheng Qin", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Xinmei Tian", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yiu-ming Cheung", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Full Professor" }, { "name": "Yonggang Zhang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "Postdoc" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 52, "n_ref": 77, "n_ref_all": 95, "n_fig": 4, "n_tab": 10, "L_tab": 4880, "n_element_tab": 464, "n_fig_1": 3, "n_tab_1": 10, "L_tab_1": 4040, "n_element_tab_1": 207, "formula_len_all": 491, "formula_len_all_1": 491, "len_all": 142491, "len_all_1": 70578, "len_abs": 1317, "len_title": 112, "len_sents": 35305, "len_sents_1": 29973, "n_sents": 261, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1336, "title": "Overfitting: An Unexpected Asset in AI\u2010Generated Image Detection", "abs": "AI-generated images have become highly realistic, raising concerns about potential misuse for malicious purposes. In this work, we propose a novel approach, DetGO, to detect generated images by overfitting the distribution of natural images. Our critical insight is that a model overfitting to one distribution (natural images) will fail to generalize to another (AI\u2010generated images). Inspired by the sharpness\u2010aware minimization, where the objective function is designed in a $\\min$-$\\max$ scheme to find flattening minima for better generalization, DetGO instead seeks to overfit the natural image distribution in a $\\max$-$\\min$ manner. This requires finding a solution with a minimal loss near the current solution and then maximizing the loss at this solution, leading to sharp minima. To address the divergence issue caused by the outer maximization, we introduce an anchor model that fits the natural image distribution. In particular, we learn an overfitting model that produces the same outputs as the anchor model while exhibiting abrupt loss behavior for small perturbations. Consequently, we can effectively determine whether an input image is AI-generated by calculating the output differences between these two models. Extensive experiments across multiple benchmarks demonstrate the effectiveness of our proposed method.", "keywords": [ "Overfitting", "AI-generated image detection", "Generative models" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 2, 3, 1 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "F0iBQktr5Z", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Haoyi Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jiang Bian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Runsheng Wang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Full Professor" }, { "name": "Shizhao Sun", "gender": "Female", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Yibo Lin", "gender": "Male", "institution": "Peking University", "country": "", "position": "Assistant Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 17, "n_ref": 41, "n_ref_all": 57, "n_fig": 5, "n_tab": 5, "L_tab": 4345, "n_element_tab": 158, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2304, "n_element_tab_1": 212, "formula_len_all": 94, "formula_len_all_1": 94, "len_all": 132902, "len_all_1": 61142, "len_abs": 2536, "len_title": 167, "len_sents": 34720, "len_sents_1": 26404, "n_sents": 234, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 118, "L_abs": 1843, "title": "ANALOGXPERT: AUTOMATING ANALOG TOPOLOGY SYNTHESIS BY INCORPORATING CIRCUIT DESIGN EXPERTISE INTO LARGE LANGUAGE MODELS", "abs": "Analog circuits are crucial in modern electronic systems, and automating their design\nhas attracted significant research interest. One of major challenges is topology\nsynthesis, which determines circuit components and their connections. Recent\nstudies explore large language models (LLM) for topology synthesis. However,\nthe scenarios addressed by these studies do not align well with practical applications.\nSpecifically, existing work uses vague design requirements as input and outputs\nan ideal model, but detailed structural requirements and device-level models\nare more practical. Moreover, current approaches either formulate topology synthesis\nas graph generation or Python code generation, whereas practical topology\ndesign is a complex process that demands extensive design knowledge. In this\nwork, we propose AnalogXpert, a LLM-based agent aiming at solving practical\ntopology synthesis problem by incorporating circuit design expertise into LLMs.\nFirst, we represent analog topology as SPICE code and introduce a subcircuit library\nto reduce the design space, in the same manner as experienced designers.\nSecond, we decompose the problem into two sub-task (i.e., block selection and\nblock connection) through the use of CoT and in-context learning techniques, to\nmimic the practical design process. Third, we introduce a proofreading strategy\nthat allows LLMs to incrementally correct the errors in the initial design, akin to\nhuman designers who iteratively check and adjust the initial topology design to\nensure accuracy. Finally, we construct a high-quality benchmark containing both\nreal data (30) and synthetic data (2k). AnalogXpert achieves 40% and 23% success\nrates on the synthetic dataset and real dataset respectively, which is markedly\nbetter than those of GPT-4o (3% on both the synthetic dataset and the real dataset).", "keywords": [ "Analog circuit design", "subcircuit library", "proofreading", "CoT", "in-context learning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "F0Zd3knG9j", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Emanuele Moscato", "gender": "unknown", "institution": "Bocconi University", "country": "IT", "position": "Postdoc" }, { "name": "Jerome Garnier-Brun", "gender": "unknown", "institution": "Bocconi University", "country": "IT", "position": "Postdoc" }, { "name": "Luca Saglietti", "gender": "Male", "institution": "Bocconi University", "country": "IT", "position": "Assistant Professor" }, { "name": "Marc Mezard", "gender": "Male", "institution": "Bocconi University", "country": "IT", "position": "Full Professor" } ], "n_formula": 27, "n_formula_1": 6, "n_ref_uni": 26, "n_ref": 45, "n_ref_all": 96, "n_fig": 21, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 12, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1609, "formula_len_all_1": 337, "len_all": 155297, "len_all_1": 74624, "len_abs": 1155, "len_title": 124, "len_sents": 61314, "len_sents_1": 36447, "n_sents": 398, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1159, "title": "How transformers learn structured data: insights from hierarchical filtering", "abs": "Understanding the learning process and the embedded computation in transformers is becoming a central goal for the development of interpretable AI. In the present study, we introduce a hierarchical filtering procedure for generative models of sequences on trees, allowing us to hand-tune the range of positional correlations in the data. Leveraging this controlled setting, we provide evidence that vanilla encoder-only transformers can approximate the exact inference algorithm when trained on root classification and masked language modeling tasks, and study *how* this computation is discovered and implemented. We find that correlations at larger distances, corresponding to increasing layers of the hierarchy, are sequentially included by the network during training. Moreover, by comparing attention maps from models trained with varying degrees of filtering and by probing the different encoder levels, we find clear evidence of a reconstruction of correlations on successive length scales corresponding to the various levels of the hierarchy, which we relate to a plausible implementation of the exact inference algorithm within the same architecture.", "keywords": [ "Transformers", "Belief Propagation", "mechanistic explanation", "structured data", "hierarchical data model", "attention", "masked language modeling" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 3, 3 ] }, { "paper_id": "F0TrRRKkQT", "primary_area": "generative models", "n_author": 9, "author_info_list": [ { "name": "Hongbin Zhou", "gender": "Male", "institution": "Ximalaya Inc.", "country": "", "position": "Researcher" }, { "name": "Jianhao Ye", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "", "position": "MS student" }, { "name": "Jianjun Zhao", "gender": "Male", "institution": "Kyushu University", "country": "JP", "position": "Full Professor" }, { "name": "Jixun Yao", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "PhD student" }, { "name": "Lei Ma", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Associate Professor" }, { "name": "Lei Xie", "gender": "Male", "institution": "Northwest Polytechnical University", "country": "", "position": "Full Professor" }, { "name": "Yang Yuguang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "MS student" }, { "name": "Yu Pan", "gender": "Male", "institution": "Kyushu University", "country": "JP", "position": "PhD student" }, { "name": "xiang zhang", "gender": "Male", "institution": "ximalaya", "country": "CN", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 29, "n_ref": 51, "n_ref_all": 71, "n_fig": 4, "n_tab": 4, "L_tab": 730, "n_element_tab": 84, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 299, "n_element_tab_1": 37, "formula_len_all": 311, "formula_len_all_1": 229, "len_all": 131147, "len_all_1": 53219, "len_abs": 1537, "len_title": 162, "len_sents": 27826, "len_sents_1": 25569, "n_sents": 176, "n_sents_1": 156, "venue": "ICLR2025", "decision": "withdraw", "L_title": 114, "L_abs": 1543, "title": "Takin-VC: Zero-shot Voice Conversion via Jointly Hybrid Content and Memory-Augmented Context-Aware Timbre Modeling", "abs": "Zero-shot voice conversion (VC) aims to transform the source speaker timbre into an arbitrary unseen one without altering the original speech content. While recent advancements in zero-shot VC methods have shown remarkable progress, there still remains considerable potential for improvement in terms of improving speaker similarity and speech naturalness. In this paper, we propose Takin-VC, a novel zero-shot VC framework based on jointly hybrid content and memory-augmented context-aware timbre modeling to tackle this challenge. Specifically, an effective hybrid content encoder, guided by neural codec training, that leverages quantized features from pre-trained WavLM and HybridFormer is first presented to extract the linguistic content of the source speech. Subsequently, we introduce an advanced cross-attention-based context-aware timbre modeling approach that learns the fine-grained, semantically associated target timbre features. To further enhance both speaker similarity and real-time performance, we utilize a conditional flow matching model to reconstruct the Mel-spectrogram of the source speech. Additionally, we advocate an efficient memory-augmented module designed to generate high-quality conditional target inputs for the flow matching process, thereby improving the overall performance of the proposed system. Experimental results demonstrate that the proposed Takin-VC method surpasses state-of-the-art zero-shot VC systems, delivering superior performance in terms of both speech naturalness and speaker similarity.", "keywords": [ "zero-shot voice conversion", "hybrid content encoder", "memory-augmented context-aware timbre modeling", "conditional flow matching" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "F0K0zxi62U", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Haibin Ling", "gender": "Male", "institution": "State University of New York, Stony Brook", "country": "US", "position": "Professor" }, { "name": "Peiyao Wang", "gender": "Female", "institution": "State University of New York at Stony Brook", "country": "", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 39, "n_ref": 68, "n_ref_all": 85, "n_fig": 4, "n_tab": 3, "L_tab": 847, "n_element_tab": 144, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 846, "n_element_tab_1": 68, "formula_len_all": 1559, "formula_len_all_1": 1716, "len_all": 120337, "len_all_1": 68023, "len_abs": 1407, "len_title": 64, "len_sents": 31686, "len_sents_1": 30258, "n_sents": 238, "n_sents_1": 226, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1413, "title": "EVGAP: Egocentric-Exocentric Video Groups Alignment Pre-training", "abs": "Aligning egocentric and exocentric videos facilitates the learning of view-invariant features, which significantly contributes to video understanding. While previous approaches have primarily focused on aligning individual ego-exo video pairs, our method extends this concept by aligning groups of synchronized egocentric and exocentric videos. This strategy enables the model to capture more comprehensive cross-view relationships across densely captured viewpoints, enhancing its capacity for robust multi-view understanding.\nTherefore, we develop a pipeline based on contrastive learning for \\textbf{E}gocentric-exocentric \\textbf{V}ideo \\textbf{G}roups \\textbf{A}lignment \\textbf{P}re-training (EVGAP). \nOur method introduces several key innovations: 1) a novel video pre-training paradigm that extends alignment from ego-exo video pairs to ego-exo video group alignments; 2) an innovative two-step training process that leverages the abundant ego-exo video pair data to support the learning of ego-exo video group alignments, transitioning from sparse to dense viewpoints; and 3) the application of auxiliary losses to progressively align videos from different perspectives.\nExtensive ablations illustrate the effectiveness of our approach in single-view and multi-view downstream tasks. We also find that our approach facilitates the tasks inluding novel views. The codes will be available upon acceptance.", "keywords": [ "multiview video", "view-invariant pretraining", "view alignment", "ego-exo pair alignment" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "F0GNv13ojF", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Chuyi He", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Guangju Wang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Jiaxuan Gao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Shusheng Xu", "gender": "Male", "institution": "IIIS, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Wei Fu", "gender": "Male", "institution": "Institute for Interdisciplinary Information Sciences, Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Weilin Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Wenjie Ye", "gender": "Male", "institution": "Institute for Interdisciplinary Information Sciences, Tsinghua University", "country": "", "position": "Undergrad student" }, { "name": "Yi Wu", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Assistant Professor" }, { "name": "Zhiyu Mei", "gender": "Male", "institution": "Institute for Interdisciplinary Information Sciences, Tsinghua University", "country": "", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 9, "n_ref_uni": 36, "n_ref": 124, "n_ref_all": 139, "n_fig": 4, "n_tab": 4, "L_tab": 1580, "n_element_tab": 176, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 110, "n_element_tab_1": 2, "formula_len_all": 2401, "formula_len_all_1": 792, "len_all": 170413, "len_all_1": 66528, "len_abs": 1858, "len_title": 115, "len_sents": 44242, "len_sents_1": 31303, "n_sents": 408, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1894, "title": "On Designing Effective RL Reward at Training Time for LLM Reasoning", "abs": "Reward models have been increasingly critical for improving the reasoning capability of LLMs. Existing research has shown that a well-trained reward model can substantially improve model performances *at inference time* via search or best-of-N votes. \nHowever, the potential of reward models during *RL training time* still remains largely under-explored. \nIt is currently unclear whether these reward models can provide additional training signals to RL training that uses sparse success rewards, which verify the correctness of solutions.\nIn this work, we evaluate popular reward models for RL training, including the Outcome-supervised Reward Model (ORM) and the Process-supervised Reward Model (PRM), and train a collection of LLMs for math problems using RL by combining these learned rewards with success rewards. Surprisingly, even though these learned reward models have strong inference-time performances, they may only bring marginal improvements or even hurt RL *training*, producing worse performances than LLMs trained with the success reward only. \nWe find that *training collapse* easily occurs in RL training when PRM simply serves as reward shaping in addition to the success rewards. Our further analysis reveals two issues that may lead to the sub-optimal performance.\nTherefore, we introduce two novel reward refinement techniques, including the **Clip** and the **Delta** mechanisms, to tackle the identified issues.\nWe evaluate our techniques with multiple reward models over a set of 1.5B and 7B LLMs on MATH and GSM8K benchmarks, where both **Clip** and **Delta** consistently enhance RL training. \nFinally, we also demonstrate that with a carefully designed reward function, pure RL training without any additional supervised tuning can further improve all the evaluated LLMs, including the state-of-the-art 7B LLM Qwen2.5-Math-7B-Instruct on MATH and GSM8K benchmarks.", "keywords": [ "Large Language Models", "RLHF", "PPO", "LLM for Reasoning", "Reward Design" ], "rating_list": [ 3, 3, 5, 8, 6 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 2, 3 ] }, { "paper_id": "F07ic7huE3", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Yutaka Shimizu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Masayoshi Tomizuka", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" } ], "n_formula": 29, "n_formula_1": 23, "n_ref_uni": 40, "n_ref": 76, "n_ref_all": 85, "n_fig": 6, "n_tab": 3, "L_tab": 289, "n_element_tab": 33, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3324, "formula_len_all_1": 2318, "len_all": 148338, "len_all_1": 60523, "len_abs": 992, "len_title": 96, "len_sents": 40119, "len_sents_1": 27398, "n_sents": 321, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 925, "title": "Bisimulation Metric for Model Predictive Control", "abs": "Model-based reinforcement learning (MBRL) has shown promise for improving sample efficiency and decision-making in complex environments. However, existing methods face challenges in training stability, robustness to noise, and computational efficiency. In this paper, we propose Bisimulation Metric for Model Predictive Control (BS-MPC), a novel approach that incorporates bisimulation metric loss in its objective function to directly optimize the encoder. This optimization enables the learned encoder to extract intrinsic information from the original state space while discarding irrelevant details. BS-MPC improves training stability, robustness against input noise, and computational efficiency by reducing training time. We evaluate BS-MPC on both continuous control and image-based tasks from the DeepMind Control Suite, demonstrating superior performance and robustness compared to state-of-the-art baseline methods.", "keywords": [ "Reinforcement Learning", "Model-based reinforcement learning", "optimal control", "MPC" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "EzrZX9bd4G", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Divya Jyoti Bajpai", "gender": "Male", "institution": "Indian Institute of Technology, Bombay, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "PhD student" }, { "name": "Manjesh Kumar Hanawal", "gender": "Male", "institution": "Indian Institute of Technology Bombay", "country": "IN", "position": "Associate Professor" } ], "n_formula": 15, "n_formula_1": 7, "n_ref_uni": 30, "n_ref": 64, "n_ref_all": 78, "n_fig": 2, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 225, "n_element_tab_1": 41, "formula_len_all": 679, "formula_len_all_1": 1733, "len_all": 130499, "len_all_1": 59556, "len_abs": 1196, "len_title": 133, "len_sents": 35365, "len_sents_1": 27138, "n_sents": 302, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1220, "title": "BEEM: Boosting Performance of Early Exit DNNs using Multi-Exit Classifiers as Experts", "abs": "Early Exit (EE) techniques have emerged as a means to reduce inference latency in Deep Neural Networks (DNNs). The latency improvement and accuracy in these techniques crucially depend on the criteria used to make exit decisions. We propose a new decision criterion BEEM where exit classifiers are treated as experts and aggregate their confidence scores. The confidence scores are aggregated only if neighbouring experts are consistent in prediction as the samples pass through them, thus capturing their ensemble effect. A sample exits when the aggregated confidence value exceeds a threshold. The threshold is set using the error rates of the intermediate exits aiming to surpass the performance of conventional DNN inference. Experimental results on the COCO dataset for Image captioning and GLUE datasets for various language tasks demonstrate that our method enhances the performance of state-of-the-art EE methods, achieving improvements in speed-up by a factor $1.5\\times$ to $2.1\\times$. When compared to the final layer, its accuracy is comparable in harder Image Captioning and improves in the easier language tasks. The source code is available at \\url{https://anonymous.4open.science/r/BEEM1-639C/README.md}", "keywords": [ "Early Exits; Expert-based exiting" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "EzjsoomYEb", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Fabrizio Frasca", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "Postdoc" }, { "name": "Guy Bar-Shalom", "gender": "Male", "institution": "Technion, Technion", "country": "IL", "position": "PhD student" }, { "name": "Haggai Maron", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Assistant Professor" }, { "name": "Michael M. Bronstein", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Yam Eitan", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "PhD student" }, { "name": "Yoav Gelberg", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 167, "n_formula_1": 22, "n_ref_uni": 50, "n_ref": 169, "n_ref_all": 225, "n_fig": 13, "n_tab": 8, "L_tab": 1429, "n_element_tab": 136, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 11787, "formula_len_all_1": 1492, "len_all": 327618, "len_all_1": 75577, "len_abs": 1591, "len_title": 158, "len_sents": 102821, "len_sents_1": 31699, "n_sents": 1009, "n_sents_1": 288, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 110, "L_abs": 1611, "title": "Topological Blindspots: Understanding and Extending Topological Deep Learning Through the Lens of Expressivity", "abs": "Topological deep learning (TDL) is a rapidly growing field that seeks to leverage topological structure in data and facilitate learning from data supported on topological objects, ranging from molecules to 3D shapes. Most TDL architectures can be unified under the framework of higher-order message-passing (HOMP), which generalizes graph message-passing to higher-order domains. In the first part of the paper, we explore HOMP's expressive power from a topological perspective, demonstrating the framework's inability to capture fundamental topological and metric invariants such as diameter, orientability, planarity, and homology. In addition, we demonstrate HOMP's limitations in fully leveraging lifting and pooling methods on graphs. To the best of our knowledge, this is the first work to study the expressivity of TDL from a topological perspective. In the second part of the paper, we develop two new classes of architectures -- multi-cellular networks (MCN) and scalable MCN (SMCN) -- which draw inspiration from expressive GNNs. MCN can reach full expressivity, but scaling it to large data objects can be computationally expansive. Designed as a more scalable alternative, SMCN still mitigates many of HOMP's expressivity limitations. Finally, we design new benchmarks for evaluating models based on their ability to learn topological properties of complexes. We then evaluate SMCN on these benchmarks as well as on real-world graph datasets, demonstrating improvements over both HOMP baselines and expressive graph methods, highlighting the value of expressively leveraging topological information.", "keywords": [ "Topological Deep Learning", "Message Passing", "Higher Order Message Passing", "Expressivity", "Graph Neural Networks", "GNNs", "Topology", "Homology", "Symmetry" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 3, 4, 3 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "EzB0n8aRqI", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Chen Feng", "gender": "Male", "institution": "University College London", "country": "GB", "position": "Postdoc" }, { "name": "Nicu Sebe", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Full Professor" }, { "name": "Ioannis Patras", "gender": "Male", "institution": "Queen Mary, University of London", "country": "GB", "position": "Full Professor" } ], "n_formula": 72, "n_formula_1": 24, "n_ref_uni": 48, "n_ref": 84, "n_ref_all": 107, "n_fig": 3, "n_tab": 8, "L_tab": 1038, "n_element_tab": 100, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7448, "formula_len_all_1": 1603, "len_all": 185421, "len_all_1": 64402, "len_abs": 1192, "len_title": 121, "len_sents": 56605, "len_sents_1": 26821, "n_sents": 451, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1196, "title": "Towards Better Understanding Open-set Noise in Learning with Noisy Labels", "abs": "To reduce reliance on labeled data, learning with noisy labels (LNL) has garnered increasing attention. However, most existing works primarily assume that noisy datasets are dominated by closed-set noise, where the true labels of noisy samples come from another known category, thereby overlooking the widespread presence of open-set noise\u2014where the true labels may not belong to any known category.\nIn this paper, we refine the LNL problem by explicitly accounting for the presence of open-set noise. We theoretically analyze and compare the impacts of open-set and closed-set noise, as well as the differences between various open-set noise modes. Additionally, we examine a common open-set noise detection mechanism based on prediction entropy. To empirically validate our theoretical insights, we construct two open-set noisy datasets\u2014CIFAR100-O and ImageNet-O\u2014and introduce a novel open-set test set for the widely used real-world noisy dataset, WebVision. Our findings indicate that open-set noise exhibits distinct qualitative and quantitative characteristics, underscoring the need for further exploration into how models can be fairly and comprehensively evaluated under such conditions.", "keywords": [ "Open-set noise", "Noisy labels" ], "rating_list": [ 5, 3, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "Eyv12jjyMN", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Haozhao Wang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Postdoc" }, { "name": "Jun Wang", "gender": "unknown", "institution": "iWudao", "country": "", "position": "Consultant" }, { "name": "Lang Gao", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Ruixuan Li", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Wei Liu", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Postdoc" }, { "name": "Zhigang Zeng", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zhiying Deng", "gender": "unknown", "institution": "Central China Normal University", "country": "CN", "position": "Lecturer" }, { "name": "Zhongyu Niu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Researcher" } ], "n_formula": 22, "n_formula_1": 16, "n_ref_uni": 44, "n_ref": 74, "n_ref_all": 120, "n_fig": 9, "n_tab": 6, "L_tab": 4397, "n_element_tab": 470, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2956, "n_element_tab_1": 285, "formula_len_all": 1077, "formula_len_all_1": 738, "len_all": 187175, "len_all_1": 73343, "len_abs": 1536, "len_title": 136, "len_sents": 51833, "len_sents_1": 31135, "n_sents": 448, "n_sents_1": 264, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1443, "title": "Attacking for Inspection and Instruction: Attack Techniques Can Aid In Interpretability", "abs": "This study investigates a self-explantory natural language processing framework constructed with a cooperative game, where a generator first extracts the most informative segment from raw input, and a subsequent predictor utilizes the selected subset for its input. The generator and predictor are trained collaboratively to maximize prediction accuracy. In this paper, we first uncover a potential caveat: such a cooperative game could unintentionally introduce a sampling bias between the explanation and the target prediction label. Specifically, the generator might inadvertently create an incorrect correlation between the selected explanation and the label, even when they are semantically unrelated in the original dataset. Subsequently, we elucidate the origins of this bias using both detailed theoretical analysis and empirical evidence. Our findings suggest a direction for inspecting these correlations through attacks, based on which we further introduce an instruction to prevent the predictor from learning the correlations.\nThrough experiments on six text classification datasets and one graph classification dataset using three network architectures (GRUs, BERT, and GCN), we show that our attack-inspired method outperforms recent competitive methods.\nWe also compare our method against a representative LLM (llama-3.1-8b-instruct), and demonstrate that our approach achieves comparable results, sometimes even surpassing it.", "keywords": [ "Interpretability", "natural language processing", "feature selection" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 1 ] }, { "paper_id": "EytBpUGB1Z", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Guangxuan Xiao", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Wenhao Wu", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yao Fu", "gender": "Male", "institution": "University of Edinburgh", "country": "", "position": "PhD student" }, { "name": "Yizhong Wang", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "", "position": "PhD student" }, { "name": "Hao Peng", "gender": "Male", "institution": "Department of Computer Science, University of Illinois Urbana-Champaign", "country": "", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 24, "n_ref": 46, "n_ref_all": 73, "n_fig": 13, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 11, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 43, "formula_len_all_1": 43, "len_all": 99058, "len_all_1": 54176, "len_abs": 1951, "len_title": 111, "len_sents": 34001, "len_sents_1": 26667, "n_sents": 226, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 63, "L_abs": 1915, "title": "Retrieval Head Mechanistically Explains Long-Context Factuality", "abs": "Despite the recent progress in long-context language models, it remains elusive how transformer-based models exhibit the capability to retrieve relevant information from arbitrary locations within the long context. This paper aims to address this question. Our systematic investigation across a wide spectrum of models reveals that a special type of attention heads are largely responsible for retrieving information, which we dub retrieval heads. We identify intriguing properties of retrieval heads:(1) universal: all the explored models with long-context capability have a set of retrieval heads; (2) sparse: only a small portion (less than 5\\%) of the attention heads are retrieval. (3) intrinsic: retrieval heads already exist in models pretrained with short context. When extending the context length by continual pretraining, it is still the same set of heads that perform information retrieval. (4) dynamically activated: take Llama-2 7B for example, 12 retrieval heads always attend to the required information no matter how the context is changed. The rest of the retrieval heads are activated in different contexts. (5) causal: completely pruning retrieval heads leads to failure in retrieving relevant information and results in hallucination, while pruning random non-retrieval heads does not affect the model's retrieval ability. We further show that retrieval heads strongly influence chain-of-thought (CoT) reasoning, where the model needs to frequently refer back the question and previously-generated context. Conversely, tasks where the model directly generates the answer using its intrinsic knowledge are less impacted by masking out retrieval heads. These observations collectively explain which internal part of the model seeks information from the input tokens. We believe our insights will foster future research on reducing hallucination, improving reasoning, and compressing the KV cache.", "keywords": [ "Large language models", "long context", "interpretability", "attention" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 4, 3, 4, 3 ], "confidence_list": [ 4, 2, 3, 4 ] }, { "paper_id": "EyaH1wzmao", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Ido Kaminer", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "", "position": "Associate Professor" }, { "name": "Itay Beit Halachmi", "gender": "Not Specified", "institution": "Technion - Israel Institute of Technology, Technion - Israel Institute of Technology", "country": "IL", "position": "MS student" } ], "n_formula": 23, "n_formula_1": 12, "n_ref_uni": 32, "n_ref": 47, "n_ref_all": 69, "n_fig": 7, "n_tab": 5, "L_tab": 5327, "n_element_tab": 26, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 69, "n_element_tab_1": 7, "formula_len_all": 4005, "formula_len_all_1": 1740, "len_all": 124613, "len_all_1": 65568, "len_abs": 1880, "len_title": 130, "len_sents": 42283, "len_sents_1": 32532, "n_sents": 289, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1905, "title": "The Ramanujan Library - Automated Discovery on the Hypergraph of Integer Relations", "abs": "Fundamental mathematical constants appear in nearly every field of science, from physics to biology. \nFormulas that connect different constants often bring great insight by hinting at connections between previously disparate fields.\nDiscoveries of such relations, however, have remained scarce events, relying on sporadic strokes of creativity by human mathematicians.\nRecent developments of algorithms for automated conjecture generation have accelerated the discovery of formulas for specific constants.\nYet, the discovery of connections between constants has not been addressed.\nIn this paper, we present the first library dedicated to mathematical constants and their interrelations. This library can serve as a central repository of knowledge for scientists from different areas, and as a collaborative platform for development of new algorithms.\nThe library is based on a new representation that we propose for organizing the formulas of mathematical constants: \na hypergraph, with each node representing a constant and each edge representing a formula.\nUsing this representation, we propose and demonstrate a systematic approach for automatically enriching this library using PSLQ, an integer relation algorithm based on QR decomposition and lattice construction. During its development and testing, our strategy led to the discovery of 75 previously unknown connections between constants, including a new formula for the `first continued fraction' constant $C_1$, novel formulas for natural logarithms, and new formulas connecting $\\pi$ and $e$.\nThe latter formulas generalize a century-old relation between $\\pi$ and $e$ by Ramanujan, which until now was considered a singular formula and is now found to be part of a broader mathematical structure. \nThe code supporting this library is a public, open-source API that can serve researchers in experimental mathematics and other fields of science.", "keywords": [ "Continued Fractions", "Mathematical Constants", "Integer Relations", "Experimental Mathematics", "Riemann Zeta Function", "Irrational Number", "PSLQ", "AI In Mathematics", "Automated Conjecture Generation" ], "rating_list": [ 8, 3, 8 ], "soundness_list": [ 4, 3, 4 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 4, 2, 2 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "EyW92b6DyY", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Chengliang Chai", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "Associate Professor" }, { "name": "Chenyu Yang", "gender": "Female", "institution": "HKUST(GZ)", "country": "", "position": "PhD student" }, { "name": "Chuanxuan Cui", "gender": "Male", "institution": "Renmin University of China", "country": "", "position": "MS student" }, { "name": "Ju Fan", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Nan Tang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "Associate Professor" }, { "name": "Yuyu Luo", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 2, "n_formula_1": 4, "n_ref_uni": 45, "n_ref": 85, "n_ref_all": 117, "n_fig": 6, "n_tab": 13, "L_tab": 2725, "n_element_tab": 176, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 2717, "n_element_tab_1": 244, "formula_len_all": 98, "formula_len_all_1": 226, "len_all": 167390, "len_all_1": 75357, "len_abs": 1145, "len_title": 53, "len_sents": 52522, "len_sents_1": 34117, "n_sents": 388, "n_sents_1": 252, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1238, "title": "Retrieval Augmented Imputation using Data Lake Tables", "abs": "Data imputation is an essential problem in many data science applications. Existing methods often struggle to impute missing values in scenarios where there is a lack of sufficient data redundancy. In this paper, leveraging large language models (LLMs) and data lakes, we propose a novel approach for retrieval-augmented imputation called RAI, utilizing fine-grained tuple-level retrieval instead of traditional coarse-grained table-based retrieval. RAI addresses the challenges of retrieving relevant tuples for missing value imputation from a data lake, where tuples have heterogeneous attributes, diverse values, and missing values. Rather than simply searching for similar tables, RAI employs a tuple encoder to learn meaningful representations for capturing tuple similarities and differences, enabling effective identification of candidate tuples. The retrieved results are further refined by a tuple reranker. We also introduce a new benchmark, mvBench, to advance further research. Extensive experiments demonstrate that RAI significantly outperforms existing methods. \nWe conduct extensive experiments, demonstrating that RAI significantly outperforms state-of-the-art table-based retrieval-augmented imputation methods by 10.7%.", "keywords": [ "data imputation", "dense retrieval", "contrastive learning" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 4 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "EyTzNHoEyK", "primary_area": "datasets and benchmarks", "n_author": 13, "author_info_list": [ { "name": "Astrid Walle", "gender": "Female", "institution": "Siemens Energy", "country": "", "position": "Developer" }, { "name": "Burkhard Hupertz", "gender": "unknown", "institution": "Ford Motor Company", "country": "DE", "position": "Researcher" }, { "name": "Charles Mockett", "gender": "Male", "institution": "Upstream CFD GmbH", "country": "", "position": "Managing Director" }, { "name": "Danielle C. Maddix", "gender": "unknown", "institution": "AWS AI Labs", "country": "US", "position": "Applied Scientist" }, { "name": "Grigoris Fotiadis", "gender": "Male", "institution": "BETA CAE Systems", "country": "GR", "position": "Researcher" }, { "name": "Hendrik Hetmann", "gender": "Male", "institution": "Upstream CFD", "country": "", "position": "CFD Consulatant" }, { "name": "Louis Fliessbach", "gender": "Male", "institution": "Upstream CFD GmbH", "country": "DE", "position": "cfd engineer" }, { "name": "Marian Fuchs", "gender": "Male", "institution": "Upstream CFD", "country": "DE", "position": "Researcher" }, { "name": "Neil Ashton", "gender": "unknown", "institution": "Amazon Web Services", "country": "GB", "position": "Principal Researcher" }, { "name": "Norbert Sch\u00f6nwald", "gender": "Male", "institution": "Upstream CFD", "country": "", "position": "Senior CAA & IT Consultant" }, { "name": "Peter Yu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Thilo Knacke", "gender": "Male", "institution": "Upstream CFD GmbH", "country": "", "position": "Principal Researcher" }, { "name": "Vangelis Skaperdas", "gender": "Male", "institution": "BETA CAE Systems USA, Inc.", "country": "", "position": "Researcher" } ], "n_formula": 50, "n_formula_1": 0, "n_ref_uni": 19, "n_ref": 68, "n_ref_all": 143, "n_fig": 43, "n_tab": 6, "L_tab": 4966, "n_element_tab": 372, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2509, "formula_len_all_1": 0, "len_all": 288376, "len_all_1": 49476, "len_abs": 891, "len_title": 145, "len_sents": 88897, "len_sents_1": 20714, "n_sents": 721, "n_sents_1": 163, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 896, "title": "DrivAerML: High-Fidelity Computational Fluid Dynamics Dataset for Road-Car External Aerodynamics", "abs": "Machine Learning (ML) has the potential to revolutionise the field of automotive aerodynamics, enabling split-second flow predictions early in the design process. However, the lack of open-source training data for realistic road cars, using high-fidelity CFD methods, represents a barrier to their development. To address this, a high-fidelity open-source (CC-BY-SA) public dataset for automotive aerodynamics has been generated, based on 500 parametrically morphed variants of the widely-used DrivAer notchback generic vehicle. Mesh generation and scale-resolving CFD was executed using consistent and validated automatic workflows representative of the industrial state-of-the-art. Geometries and rich aerodynamic data are published in open-source formats. To our knowledge, this is the first large, public-domain dataset for complex automotive configurations generated using high-fidelity CFD.", "keywords": [ "CFD", "automotive", "ML", "drivaer", "dataset" ], "rating_list": [ 3, 5, 8, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "Ey8KcabBpB", "primary_area": "applications to robotics, autonomy, planning", "n_author": 10, "author_info_list": [ { "name": "Checheng Yu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Intern" }, { "name": "Guohao Li", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc" }, { "name": "Junting Chen", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Lin Shao", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Mengkang Hu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Tianqi Xu", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Wenqi Shao", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xunzhe Zhou", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Yao Mu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yikai Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 35, "n_ref": 61, "n_ref_all": 80, "n_fig": 7, "n_tab": 8, "L_tab": 1131, "n_element_tab": 109, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1148, "n_element_tab_1": 34, "formula_len_all": 126, "formula_len_all_1": 21, "len_all": 179540, "len_all_1": 58422, "len_abs": 1693, "len_title": 81, "len_sents": 64893, "len_sents_1": 28184, "n_sents": 435, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1557, "title": "EMOS: Embodiment-aware Heterogeneous Multi-robot Operating System with LLM Agents", "abs": "Heterogeneous multi-robot systems (HMRS) have emerged as a powerful ap-\nproach for tackling complex tasks that single robots cannot manage alone. Current\nlarge-language-model-based multi-agent systems (LLM-based MAS) have shown\nsuccess in areas like software development and operating systems, but applying\nthese systems to robot control presents unique challenges. In particular, the ca-\npabilities of each agent in a multi-robot system are inherently tied to the physical\ncomposition of the robots, rather than predefined roles. To address this issue,\nwe introduce a novel multi-agent framework designed to enable effective collab-\noration among heterogeneous robots with varying embodiments and capabilities,\nalong with a new benchmark named Habitat-MAS. One of our key designs is\nRobot Resume: Instead of adopting human-designed role play, we propose a self-\nprompted approach, where agents comprehend robot URDF files and call robot\nkinematics tools to generate descriptions of their physics capabilities to guide\ntheir behavior in task planning and action execution. The Habitat-MAS bench-\nmark is designed to assess how a multi-agent framework handles tasks that require\nembodiment-aware reasoning, which includes 1) manipulation, 2) perception, 3)\nnavigation, and 4) comprehensive multi-floor object rearrangement. The experi-\nmental results indicate that the robot\u2019s resume and the hierarchical design of our\nmulti-agent system are essential for the effective operation of the heterogeneous\nmulti-robot system within this intricate problem context.", "keywords": [ "Embodied Artificial Intelligence", "LLM Multi-agent System", "Multi-robot System", "Task Planning" ], "rating_list": [ 5, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 2, 3, 4, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 5, 3, 4 ] }, { "paper_id": "ExuBFYtCQU", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Haotian Guo", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "MS student" }, { "name": "Haoyue Zhan", "gender": "Male", "institution": "Guangzhou Quwan Network Technology Co. Limited Ltd,", "country": "CN", "position": "Researcher" }, { "name": "Jiachen Zheng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "MS student" }, { "name": "Liwei Liu", "gender": "Male", "institution": "Northwestern Polytechnical University Xi'an", "country": "CN", "position": "MS student" }, { "name": "Qiang Zhang", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "MS student" }, { "name": "Ruihong Zeng", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "", "position": "MS student" }, { "name": "Shunsi Zhang", "gender": "Male", "institution": "AI center", "country": "CN", "position": "Principal Researcher" }, { "name": "Xueyao Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Yuancheng Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Zhizheng Wu", "gender": "Male", "institution": "Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 86, "n_ref": 203, "n_ref_all": 237, "n_fig": 4, "n_tab": 17, "L_tab": 4423, "n_element_tab": 709, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2811, "n_element_tab_1": 331, "formula_len_all": 856, "formula_len_all_1": 634, "len_all": 237051, "len_all_1": 75590, "len_abs": 1476, "len_title": 122, "len_sents": 59097, "len_sents_1": 31626, "n_sents": 484, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1530, "title": "MaskGCT: Zero-Shot Text-to-Speech with Masked Generative Codec Transformer", "abs": "The recent large-scale text-to-speech (TTS) systems are usually grouped as autoregressive and non-autoregressive systems. The autoregressive systems implicitly model duration but exhibit certain deficiencies in robustness and lack of duration controllability. Non-autoregressive systems require explicit alignment information between text and speech during training and predict durations for linguistic units (e.g. phone), which may compromise their naturalness.\nIn this paper, we introduce $\\textbf{Mask}$ed $\\textbf{G}$enerative $\\textbf{C}$odec $\\textbf{T}$ransformer (MaskGCT), a fully non-autoregressive TTS model that eliminates the need for explicit alignment information between text and speech supervision, as well as phone-level duration prediction. MaskGCT is a two-stage model: in the first stage, the model uses text to predict semantic tokens extracted from a speech self-supervised learning (SSL) model, and in the second stage, the model predicts acoustic tokens conditioned on these semantic tokens. MaskGCT follows the mask-and-predict learning paradigm. During training, MaskGCT learns to predict masked semantic or acoustic tokens based on given conditions and prompts. During inference, the model generates tokens of a specified length in a parallel manner. \nExperiments with 100K hours of in-the-wild speech demonstrate that MaskGCT outperforms the current state-of-the-art zero-shot TTS systems in terms of quality, similarity, and intelligibility. Audio samples are available at https://maskgct.github.io/.", "keywords": [ "text-to-speech synthesis", "masked generative models", "codec language models", "voice cloning" ], "rating_list": [ 6, 6, 3, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "ExrEw8cVlU", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Jiahao Lu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Qiuhong Shen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Yifan Zhang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Xinchao Wang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 9, "n_ref_uni": 63, "n_ref": 109, "n_ref_all": 149, "n_fig": 8, "n_tab": 10, "L_tab": 15973, "n_element_tab": 1174, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 833, "n_element_tab_1": 59, "formula_len_all": 551, "formula_len_all_1": 440, "len_all": 225896, "len_all_1": 62314, "len_abs": 1467, "len_title": 110, "len_sents": 53191, "len_sents_1": 29646, "n_sents": 374, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 62, "L_abs": 1333, "title": "Poison-splat: Computation Cost Attack on 3D Gaussian Splatting", "abs": "3D Gaussian splatting (3DGS), known for its groundbreaking performance and efficiency, has become a dominant 3D representation and brought progress to many 3D vision tasks. However, in this work, we reveal a significant security vulnerability that has been largely overlooked in 3DGS: the computation cost of training 3DGS could be maliciously tampered by poisoning the input data. By developing an attack named Poison-splat, we reveal a novel attack surface where the adversary can poison the input images to drastically increase the computation memory and time needed for 3DGS training, pushing the algorithm towards its worst computation complexity. In extreme cases, the attack can even consume all allocable memory, leading to a Denial-of-Service (DoS) that disrupts servers, resulting in practical damages to real-world 3DGS service vendors. Such a computation cost attack is achieved by addressing a bi-level optimization problem through three tailored strategies: attack objective approximation, proxy model rendering, and optional constrained optimization. These strategies not only ensure the effectiveness of our attack but also make it difficult to defend with simple defensive measures. We hope the revelation of this novel attack surface can spark attention to this crucial yet overlooked vulnerability of 3DGS systems.", "keywords": [ "gaussian splatting", "energy-latency attack", "data poisoning attack" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "Exnt2DcdKD", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Kaushal Santosh Bhogale", "gender": "Male", "institution": "Indian Institute of Technology, Madras, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "PhD student" }, { "name": "Mitesh M Khapra", "gender": "Male", "institution": "Indian Institute of Technology, Madras", "country": "IN", "position": "Associate Professor" }, { "name": "Tahir Javed", "gender": "Male", "institution": "Indian Institute of Technology, Madras, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 29, "n_ref": 49, "n_ref_all": 65, "n_fig": 6, "n_tab": 6, "L_tab": 5346, "n_element_tab": 834, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 383, "n_element_tab_1": 112, "formula_len_all": 2107, "formula_len_all_1": 345, "len_all": 175369, "len_all_1": 54807, "len_abs": 1491, "len_title": 134, "len_sents": 30932, "len_sents_1": 25995, "n_sents": 226, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1497, "title": "NIRANTAR: Continual Learning with New Languages and Domains on Real-world Speech Data", "abs": "We present Nirantar based on a large-scale effort to collect extempore and conversational speech data from participants spanning 22 languages across diverse locations in India. Given the extensive number of languages and locations involved, data is collected in incremental batches. Each batch introduces new languages, new domains (locations), or both, creating a practical playground for continual learning (CL). Nirantar contains a total of 3250 hours of human-transcribed speech data covering 208 Indian districts across 22 languages, with 1720 hours newly released as a part of this work. The data inflow and resulting multilingual multi-domain episodes are based on real-world data collection rather than simulated episodes commonly found in existing CL datasets. In particular, the amount of data collected and the number of languages and domains involved are not uniform across episodes, reflecting a practical and real-world continual learning scenario. This dataset serves as a playground for training and evaluating CL approaches in three different scenarios: Language-Incremental (LIL), Domain-Incremental (DIL), and the novel Language-Incremental Domain-Incremental Learning (LIDIL), which has not been studied before. To establish the dataset's usefulness, we evaluate several existing CL approaches within these scenarios. Our findings indicate that the behaviour of these algorithms varies across the three scenarios, emphasizing the need for detailed independent studies of each.", "keywords": [ "continual learning", "speech", "recognition", "datasets", "indian languages", "multilingual asr" ], "rating_list": [ 5, 5, 6, 8, 5 ], "soundness_list": [ 2, 2, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 4, 3 ], "contribution_list": [ 3, 2, 3, 4, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "Exkm5OReTY", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Bo Zheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Instructor" }, { "name": "Shuai Fang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yudong Chen", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "Undergrad student" }, { "name": "Yuke Zhu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zihua Xiong", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Sheng Guo", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 29, "n_ref": 48, "n_ref_all": 57, "n_fig": 3, "n_tab": 6, "L_tab": 2430, "n_element_tab": 332, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2430, "n_element_tab_1": 332, "formula_len_all": 624, "formula_len_all_1": 624, "len_all": 88743, "len_all_1": 61111, "len_abs": 1182, "len_title": 120, "len_sents": 27269, "len_sents_1": 26239, "n_sents": 187, "n_sents_1": 176, "venue": "ICLR2025", "decision": "withdraw", "L_title": 72, "L_abs": 1186, "title": "MaskTab: Masked Tabular Data Modeling for Learning with Missing Features", "abs": "Tabular machine learning has garnered increasing attention due to its practical value. Unlike the complete and standardized data often assumed in academia, tabular data primarily originates from industrial contexts and usually faces the issue of incomplete data samples, i.e., some features of a sample may be unpredictably missing. In this work, we introduce MaskTab, a masked tabular data modeling framework designed to facilitate model learning despite missing features. Instead of pursuing to accurately restore missing features like existing imputation methods, we jointly approach missing feature modeling and downstream tasks (e.g., classification) with a unified objective. Concretely, we propose to randomly drop out some solid features during training, equipped with a missing-related masked attention mechanism, to help the model rely more on trustworthy features when making decisions. Experiments on the very recent industry-grade benchmark, TabReD, suggest that our method surpasses the second DNN-based competitor by a clear margin, demonstrating its effectiveness and robustness in real-world scenarios. We will release the code and the model to facilitate reproduction.", "keywords": [ "tabular data prediction", "masked learning", "missing features" ], "rating_list": [ 6, 3, 3, 1 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "ExUC9dQJhQ", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Calvin Tsay", "gender": "unknown", "institution": "Imperial College London", "country": "", "position": "Assistant Professor" }, { "name": "Mark Niklas Mueller", "gender": "Male", "institution": "LogicStar AI", "country": "CH", "position": "Researcher" }, { "name": "Matthew Robert Wicker", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Assistant Professor" }, { "name": "Maximilian Baader", "gender": "unknown", "institution": "ETH Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Philip Sosnin", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "PhD student" } ], "n_formula": 84, "n_formula_1": 24, "n_ref_uni": 39, "n_ref": 66, "n_ref_all": 80, "n_fig": 8, "n_tab": 5, "L_tab": 1007, "n_element_tab": 16, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 1, "n_element_tab_1": 1, "formula_len_all": 7626, "formula_len_all_1": 1845, "len_all": 199825, "len_all_1": 73285, "len_abs": 1185, "len_title": 113, "len_sents": 64292, "len_sents_1": 33483, "n_sents": 505, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1189, "title": "Certified Robustness to Data Poisoning in Gradient-Based Training", "abs": "Modern machine learning pipelines leverage large amounts of public data, making it infeasible to guarantee data quality and leaving models open to poisoning and backdoor attacks. Provably bounding model behavior under such attacks remains an open problem. In this work, we address this challenge by developing the first framework providing provable guarantees on the behavior of models trained with potentially manipulated data without modifying the model or learning algorithm. In particular, our framework certifies robustness against untargeted and targeted poisoning, as well as backdoor attacks, for bounded and unbounded manipulations of the training inputs and labels. Our method leverages convex relaxations to over-approximate the set of all possible parameter updates for a given poisoning threat model, allowing us to bound the set of all reachable parameters for any gradient-based learning algorithm. Given this set of parameters, we provide bounds on worst-case behavior, including model performance and backdoor success rate. We demonstrate our approach on multiple real-world datasets from applications including energy consumption, medical imaging, and autonomous driving.", "keywords": [ "Data Poisoning", "Certified Robustness", "Neural Networks" ], "rating_list": [ 5, 8, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 4, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "ExHUtB2vnz", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 8, "author_info_list": [ { "name": "Haihong E", "gender": "Female", "institution": "Beijing University of Post and Telecommunication", "country": "", "position": "Assistant Professor" }, { "name": "Haoran Luo", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Meina Song", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Full Professor" }, { "name": "Ningyuan Li", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Tianyi Hu", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Tianyu Yao", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "MS student" }, { "name": "Yifan Zhu", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Assistant Professor" }, { "name": "Yuhan Li", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 20, "n_ref": 52, "n_ref_all": 63, "n_fig": 3, "n_tab": 8, "L_tab": 775, "n_element_tab": 72, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 938, "n_element_tab_1": 12, "formula_len_all": 693, "formula_len_all_1": 583, "len_all": 107195, "len_all_1": 55550, "len_abs": 1720, "len_title": 135, "len_sents": 38635, "len_sents_1": 26224, "n_sents": 290, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1728, "title": "INFER: A Neural-symbolic Model For Extrapolation Reasoning on Temporal Knowledge Graph", "abs": "Temporal Knowledge Graph(TKG) serves as an efficacious way to store dynamic facts in real-world. Extrapolation reasoning on TKGs, which aims at predicting possible future events, has attracted consistent research interest. Recently, some rule-based methods have been proposed, which are considered more interpretable compared with embedding-based methods. Existing rule-based methods apply rules through path matching or subgraph extraction, which falls short in inference ability and suffers from missing facts in TKGs. Besides, during rule application period, these methods consider the standing of facts as a binary 0 or 1 problem and ignores the validity as well as frequency of historical facts under temporal settings.\nIn this paper, by designing a novel paradigm for rule application, we propose INFER, a neural-symbolic model for TKG extrapolation. With the introduction of Temporal Validity Function, INFER firstly considers the frequency and validity of historical facts and extends the truth value of facts into continuous real number to better adapt for temporal settings. INFER builds Temporal Weight Matrices with a pre-trained static KG embedding model to enhance its inference ability. Moreover INFER adopts a rule projection module which enables it apply rules through conducting matrices operation on GPU, which improves the efficiency of rule application. This feature also facilitates potential integration with existing embedding-based methods.\nExperimental results show that INFER achieves state-of-the-art performance on three datasets and significantly outperforms existing rule-based models on our modified, more sparse TKG datasets, which demonstrates the superiority of our model in inference ability.", "keywords": [ "Knowledge Graph", "Temporal Knowledge Graph", "Temporal Rules", "Temporal Validity" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "EwYUgKr9Fc", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Hamid Mozaffari", "gender": "Male", "institution": "Oracle", "country": "", "position": "Researcher" }, { "name": "Virendra Marathe", "gender": "Male", "institution": "Oracle", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 31, "n_ref": 89, "n_ref_all": 113, "n_fig": 5, "n_tab": 12, "L_tab": 5852, "n_element_tab": 1105, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 450, "n_element_tab_1": 65, "formula_len_all": 443, "formula_len_all_1": 455, "len_all": 166937, "len_all_1": 61735, "len_abs": 821, "len_title": 114, "len_sents": 55222, "len_sents_1": 30486, "n_sents": 349, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 825, "title": "Semantic Membership Inference Attack against Large Language Models", "abs": "Membership Inference Attacks (MIAs) determine whether a specific data point was included in the training set of a target model. In this paper, we introduce the Semantic Membership Inference Attack (SMIA), a novel approach that enhances MIA performance by leveraging the semantic content of inputs and their perturbations. SMIA trains a neural network to analyze the target model\u2019s behavior on perturbed inputs, effectively capturing variations in output probability distributions between members and non-members. We conduct comprehensive evaluations on the Pythia and GPT-Neo model families using the Wikipedia and MIMIR datasets. Our results show that SMIA significantly outperforms existing MIAs; for instance, for Wikipedia, SMIA achieves an AUC-ROC of 67.39\\% on Pythia-12B, compared to 58.90\\% by the second-best attack.", "keywords": [ "Membership Inference Attack", "Large Language Models" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "EwRxk3Ho1V", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Kannan Ramchandran", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Reagan Lee", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Samarth Goel", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "MS student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 38, "n_ref_all": 41, "n_fig": 1, "n_tab": 7, "L_tab": 1197, "n_element_tab": 177, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 983, "n_element_tab_1": 147, "formula_len_all": 22, "formula_len_all_1": 0, "len_all": 107229, "len_all_1": 56215, "len_abs": 1242, "len_title": 176, "len_sents": 34544, "len_sents_1": 26438, "n_sents": 235, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 125, "L_abs": 1246, "title": "Beyond Cosine Similarity: Introducing the Unified semantic Similarity Metric Benchmark (USMB) for Text Similarity Measurement", "abs": "Text embedding models are increasingly utilized in production across various applications, from Information Retrieval (IR) to document parsing, but relatively little research has been focused on how to best utilize these embeddings for downstream tasks. While cosine similarity, a popular measure of embedding and text similarity, is widely used, it may not be the strongest metric choice for all tasks. In this work, we introduce the Unified semantic Similarity Metric Benchmark (USMB), a novel leaderboard for text similarity metrics composed of 5 unique tasks and 30+ datasets with the goal of providing a standardized means of measuring the effectiveness of a text similarity metric on a suite of challenging tasks encompassing the nuances of semantic understanding. Additionally, we demonstrate that while cosine similarity achieves the highest score on our benchmark of any pre-existing metric, developing a task-specific ensembled model using our metrics leads to a 40.3\\% increase in benchmark performance relative to cosine similarity. We hope that through this work, greater attention can be given to potential performance gains through metric selection and that the field's ability to measure semantic similarity advances as a result.", "keywords": [ "Deep Learning or Neural Networks", "Similarity and Distance Learning", "(Application) Natural Language and Text Processing", "(Cognitive/Neuroscience) Language" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "EwFJaXVePU", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Adyasha Maharana", "gender": "Female", "institution": "Department of Computer Science, University of North Carolina, Chapel Hill", "country": "", "position": "PhD student" }, { "name": "Jaehong Yoon", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Postdoc" }, { "name": "Mohit Bansal", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Full Professor" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 38, "n_ref": 99, "n_ref_all": 137, "n_fig": 13, "n_tab": 6, "L_tab": 2021, "n_element_tab": 263, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1520, "n_element_tab_1": 170, "formula_len_all": 730, "formula_len_all_1": 336, "len_all": 179784, "len_all_1": 66105, "len_abs": 2192, "len_title": 122, "len_sents": 46531, "len_sents_1": 28625, "n_sents": 378, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 2150, "title": "Scalable Lifelong Multimodal Instruction Tuning via Dynamic Data Selection", "abs": "Visual instruction datasets from various distributors are released at different times and often contain a significant number of redundant text-image pairs, depending on their task compositions (i.e., skills) or reference sources. This redundancy greatly limits the efficient deployment of lifelong-adaptable Multimodal Large Language Models (MLLMs), hindering their ability to refine existing skills and acquire new competencies over time.\nTo address this, we reframe the problem of Lifelong Instruction Tuning (LiIT) via data selection, where the model automatically selects beneficial samples to learn from earlier and new datasets based on the current state of acquired knowledge in the model. \nBased on empirical analyses showing that selecting the best data subset using a static importance measure is often ineffective for multi-task datasets with evolving distributions, we propose LAMP, a new multi-way and adaptive data selection approach that dynamically balances sample efficiency and effectiveness during LiIT. We first construct pseudo-skill clusters by grouping gradient-based sample vectors. Next, we select the best-performing data selector for each skill cluster from a pool of selector experts, including our newly proposed scoring function, Image Grounding score. This data selector samples a subset of the most important samples from each skill cluster for training. To prevent the continuous increase in the size of the dataset pool during LiIT, which would result in excessive computation, we further introduce a cluster-wise permanent data pruning strategy to remove the most semantically redundant samples from each cluster, keeping computational requirements manageable.\nWe validate the effectiveness and efficiency of LAMP over a sequence of various multimodal instruction tuning datasets with various tasks, including (Knowledge) VQA, multilingual, grounding, reasoning, language-only, and multi-image comprehension tasks. Training with samples selected by LAMP alleviates catastrophic forgetting, especially for rare tasks, and promotes forward transfer across the continuum using only a fraction of the original datasets.", "keywords": [ "Multimodal Instruction Tuning", "Continual Learning", "Data Selection" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "Ew3VifXaxZ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Fanhu Zeng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Fei Zhu", "gender": "Male", "institution": "Centre for Artificial Intelligence and Robotics Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "CN", "position": "Postdoc" }, { "name": "Hongxin Wei", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhen Cheng", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Xu-yao Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 43, "n_ref": 92, "n_ref_all": 106, "n_fig": 7, "n_tab": 12, "L_tab": 4203, "n_element_tab": 778, "n_fig_1": 4, "n_tab_1": 9, "L_tab_1": 2398, "n_element_tab_1": 497, "formula_len_all": 1177, "formula_len_all_1": 1117, "len_all": 153217, "len_all_1": 71288, "len_abs": 1567, "len_title": 101, "len_sents": 39056, "len_sents_1": 28059, "n_sents": 348, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1551, "title": "Local-Prompt: Extensible Local Prompts for Few-Shot Out-of-Distribution Detection", "abs": "Out-of-Distribution (OOD) detection, aiming to distinguish outliers from known categories, has gained prominence in practical scenarios. Recently, the advent of vision-language models (VLM) has heightened interest in enhancing OOD detection for VLM through few-shot tuning. However, existing methods mainly focus on optimizing global prompts, ignoring refined utilization of local information with regard to outliers. Motivated by this, we freeze global prompts and introduce Local-Prompt, a novel coarse-to-fine tuning paradigm to emphasize regional enhancement with local prompts. Our method comprises two integral components: global prompt guided negative augmentation and local prompt enhanced regional regularization. The former utilizes frozen, coarse global prompts as guiding cues to incorporate negative augmentation, thereby leveraging local outlier knowledge. The latter employs trainable local prompts and a regional regularization to capture local information effectively, aiding in outlier identification. We also propose regional-related metric to empower the enrichment of OOD detection. Moreover, since our approach explores enhancing local prompts only, it can be seamlessly integrated with trained global prompts during inference to boost the performance. Comprehensive experiments demonstrate the effectiveness and potential of our method. Notably, our method reduces average FPR95 by 5.17% against state-of-the-art method in 4-shot tuning on challenging ImageNet-1k dataset, even outperforming 16-shot results of previous methods.", "keywords": [ "Out-of-distribution detection", "Prompt learning", "Vision-language model", "Few-shot learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Ev4iw23gdI", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Dongmei Jiang", "gender": "Female", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Ruiping Wang", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Wenjun Huang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "PhD student" }, { "name": "Yaowei Wang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Yifei Xing", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Zheng Qingfang", "gender": "Male", "institution": "Peng Cheng Lab", "country": "CN", "position": "Researcher" }, { "name": "xiangyuan lan", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Associate Professor" } ], "n_formula": 15, "n_formula_1": 9, "n_ref_uni": 33, "n_ref": 80, "n_ref_all": 98, "n_fig": 4, "n_tab": 8, "L_tab": 6037, "n_element_tab": 454, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 1089, "n_element_tab_1": 11, "formula_len_all": 702, "formula_len_all_1": 466, "len_all": 248939, "len_all_1": 55565, "len_abs": 1546, "len_title": 77, "len_sents": 60693, "len_sents_1": 29273, "n_sents": 382, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1586, "title": "EMMA: Empowering Multi-modal Mamba with Structural and Hierarchical Alignment", "abs": "Mamba-based architectures have shown to be a promising new direction for deep learning models owing to their competitive performance and sub-quadratic deployment speed. However, current Mamba multi-modal large language models (MLLM) are insufficient in extracting visual features, leading to imbalanced cross-modal alignment between visual and textural latents, negatively impacting performance on multi-modal tasks. In this work, we propose Empowering Multi-modal Mamba with Structural and Hierarchical Alignment (EMMA), which enables the MLLM to extract fine-grained visual information. Specifically, we propose a pixel-wise alignment module to autoregressively optimize the learning and processing of spatial image-level features along with textual tokens, enabling structural alignment at the image level. In addition, to prevent the degradation of visual information during the cross-model alignment process, we propose a multi-scale feature fusion (MFF) module to combine multi-scale visual features from intermediate layers, enabling hierarchical alignment at the feature level. Extensive experiments are conducted across a variety of multi-modal benchmarks. Our model shows lower latency than other Mamba-based MLLMs and is nearly four times faster than transformer-based MLLMs of similar scale during inference. Due to better cross-modal alignment, our model exhibits lower degrees of hallucination and enhanced sensitivity to visual details, which manifests in superior performance across diverse multi-modal benchmarks. Code provided at https://github.com/xingyifei2016/EMMA.", "keywords": [ "Multimodal models", "State space models", "Efficient architectures", "Mamba", "Computational Efficiency", "Multimodal Alignment" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "EuoHhIqvRD", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Alexander C. Berg", "gender": "Male", "institution": "Donald Bren School of Information and Computer Sciences, University of California, Irvine", "country": "US", "position": "Associate Professor" }, { "name": "Paola Cascante-Bonilla", "gender": "unknown", "institution": "State University of New York at Stony Brook", "country": "US", "position": "Assistant Professor" }, { "name": "Ruozhen He", "gender": "Female", "institution": "Rice University", "country": "US", "position": "PhD student" }, { "name": "Ziyan Yang", "gender": "Female", "institution": "ByteDance Inc.", "country": "US", "position": "research scientist" }, { "name": "Vicente Ordonez", "gender": "Male", "institution": "Rice University", "country": "US", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 46, "n_ref": 104, "n_ref_all": 147, "n_fig": 14, "n_tab": 11, "L_tab": 4922, "n_element_tab": 569, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2098, "n_element_tab_1": 387, "formula_len_all": 257, "formula_len_all_1": 287, "len_all": 227464, "len_all_1": 67865, "len_abs": 2828, "len_title": 103, "len_sents": 50942, "len_sents_1": 29191, "n_sents": 392, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1013, "title": "Is Synthetic Data Ready for Improving Visual Grounding?", "abs": "This paper extensively investigates the effectiveness of synthetic training data to improve the capabilities of vision-and-language models for grounding textual descriptions to image regions. We explore various strategies to best generate image-text pairs and image-text-box triplets using a series of pretrained models under different settings and varying degrees of reliance on real data. Through comparative analyses with synthetic, real, and web-crawled data, we identify factors that contribute to performance differences, and propose SynGround, an effective pipeline for generating useful synthetic data for visual grounding. Our findings show that SynGround can improve the localization capabilities of off-the-shelf vision-and-language models and offers the potential for infinite data generation. Particularly, SynGround improves the pointing game accuracy of pretrained ALBEF and BLIP models by 4.81% and 17.11% absolute percentage points, respectively, across the RefCOCO+ and the Flickr30k benchmarks.", "keywords": [ "Visual Grounding", "Referring Expression Comprehension", "Learning from Models", "Synthetic Data" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "EukM0UuqLx", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Gongwei Chen", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Postdoc" }, { "name": "Liqiang Nie", "gender": "Male", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "dean" }, { "name": "Renshan Zhang", "gender": "Male", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "MS student" }, { "name": "Rui Shao", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yibo Lyu", "gender": "Male", "institution": "Harbin Institute of Technology(Shenzhen)", "country": "CN", "position": "MS student" }, { "name": "WEILI GUAN", "gender": "Female", "institution": "Harbin Institute of Technology (Shenzhen)", "country": "CN", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 39, "n_ref": 92, "n_ref_all": 121, "n_fig": 15, "n_tab": 3, "L_tab": 292, "n_element_tab": 52, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 193, "n_element_tab_1": 16, "formula_len_all": 232, "formula_len_all_1": 122, "len_all": 135056, "len_all_1": 62779, "len_abs": 1907, "len_title": 139, "len_sents": 37953, "len_sents_1": 30167, "n_sents": 283, "n_sents_1": 209, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1539, "title": "Token-level Correlation-guided Compression for Efficient Multimodal Document Understanding", "abs": "Cropping high-resolution document images into multiple sub-images is the most widely used approach for current Multimodal Large Language Models (MLLMs) to do document understanding. Most of current document understanding methods preserve all tokens within sub-images and treat them equally. This neglects their different informativeness and leads to a significant increase in the number of image tokens. To perform a more adaptive and efficient document understanding, we propose Token-level Correlation-guided Compression, a parameter-free and plug-and-play methodology to optimize token processing. Firstly, we propose an innovative approach for assessing the pattern repetitiveness based on the correlation between each patch tokens. This method identifies redundant tokens, allowing for the determination of the sub-image's information density. Secondly, we present a token-level sampling method that efficiently captures the most informative tokens by delving into the correlation between the \\texttt{[CLS]} token and patch tokens. By integrating these strategies, we develop a plug-and-play Token-level Correlation-guided Compressor module that can be seamlessly incorporated into MLLMs utilizing cropping techniques. This module not only enhances the processing speed during training and inference but also maintains comparable performance. We conduct experiments with the representative document understanding model mPLUG-DocOwl1.5 and the effectiveness is demonstrated through extensive comparisons with other compression methods.", "keywords": [ "Multimodal Large Models", "Token Compression", "High-resolution Image" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "EukID7GvBy", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Bozhou Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Conghui He", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Fangcheng Fu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Postdoc" }, { "name": "Hao Liang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Hongzhi Yin", "gender": "Male", "institution": "University of Queensland", "country": "AU", "position": "Full Professor" }, { "name": "Yang Li", "gender": "Male", "institution": "Tencent Inc.", "country": "", "position": "Researcher" }, { "name": "Zhang wen tao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 14, "n_ref": 21, "n_ref_all": 41, "n_fig": 1, "n_tab": 11, "L_tab": 5510, "n_element_tab": 332, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 1189, "n_element_tab_1": 164, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 104613, "len_all_1": 55071, "len_abs": 1962, "len_title": 147, "len_sents": 27587, "len_sents_1": 25605, "n_sents": 181, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 99, "L_abs": 1972, "title": "Gradual Learning: Optimizing Fine-Tuning with Partially Mastered Knowledge in Large Language Models", "abs": "During the pretraining phase, large language models (LLMs) acquire vast amounts of knowledge from extensive text corpora. Nevertheless, in later stages such as fine-tuning and inference, the model may encounter knowledge not covered in the initial training, which can lead to hallucinations and degraded performance. This issue has a profound impact on the model's capabilities, as it will inevitably face out-of-scope knowledge after pretraining. Furthermore, fine-tuning is often required to adapt LLMs to domain-specific tasks, necessitating the acquisition of new knowledge. However, this phenomenon limits the model\u2019s ability to learn and integrate new information during fine-tuning. The effectiveness of fine-tuning largely depends on the type of knowledge involved. Existing research suggests that fine-tuning the model on partially mastered knowledge\u2014for instance, question-answer pairs where the model has a chance of providing correct responses under non-greedy decoding\u2014can enable the model to acquire new knowledge while mitigating the forgetting of previously learned information. Notably, this approach can still lead to the forgetting of fully mastered knowledge, constraining the fine-tuning dataset to a narrower range and limiting the model's overall potential for improvement. Given the model\u2019s intrinsic reasoning abilities and the interconnectedness of different knowledge areas, it is likely that as the model\u2019s capacity to utilize existing knowledge improves during fine-tuning, previously unmastered knowledge may become more understandable. To explore this hypothesis, we conducted experiments and, based on the results, proposed a two-stage fine-tuning strategy. This approach not only improves the model's overall test accuracy and knowledge retention but also preserves its accuracy on previously mastered content. When fine-tuning on the WikiQA dataset, our method increases the amount of knowledge acquired by the model in this stage by 24%.", "keywords": [ "Large Language Model; DCAI ; Fine-tuning" ], "rating_list": [ 1, 3, 3, 5, 3 ], "soundness_list": [ 1, 1, 3, 3, 2 ], "presentation_list": [ 1, 2, 2, 2, 1 ], "contribution_list": [ 1, 1, 2, 2, 1 ], "confidence_list": [ 5, 4, 3, 3, 4 ] }, { "paper_id": "EtJWnTnqku", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Alan Yuille", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Full Professor" }, { "name": "Jieru Mei", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Junfei Xiao", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Sucheng Ren", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Xianhang Li", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Xiaoke Huang", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Yuyin Zhou", "gender": "unknown", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "Zeyu Wang", "gender": "unknown", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 24, "n_ref": 45, "n_ref_all": 54, "n_fig": 2, "n_tab": 9, "L_tab": 2972, "n_element_tab": 484, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1146, "n_element_tab_1": 92, "formula_len_all": 87, "formula_len_all_1": 187, "len_all": 138484, "len_all_1": 43479, "len_abs": 1482, "len_title": 116, "len_sents": 28016, "len_sents_1": 19954, "n_sents": 204, "n_sents_1": 152, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1513, "title": "Medical Vision Generalist: Unifying Medical Imaging Tasks in Context", "abs": "This study presents Medical Vision Generalist (MVG), the first foundation model capable of handling various medical imaging tasks---such as cross-modal synthesis, image segmentation, denoising, and inpainting---within a unified image-to-image generation framework. Specifically, MVG employs an in-context generation strategy that standardizes the handling of inputs and outputs as images. By treating these tasks as an image generation process conditioned on prompt image-label pairs and input images, this approach enables a flexible unification of various tasks, even those spanning different modalities and datasets. To capitalize on both local and global context, we design a hybrid method combining masked image modeling with autoregressive training for conditional image generation. This hybrid approach yields the most robust performance across all involved medical imaging tasks. To rigorously evaluate MVG's capabilities, we curated the first comprehensive generalist medical vision benchmark, comprising 13 datasets and spanning four imaging modalities (CT, MRI, X-ray, and micro-ultrasound). Our results consistently etablish MVG's superior performance, outperforming existing vision generalists, such as Painter and LVM. Furthermore, MVG exhibits strong scalability, with its performance demonstrably improving when trained on a more diverse set of tasks, and can be effectively adapted to unseen datasets with only minimal task-specific samples. The code and the benchmark will be publicly available.", "keywords": [ "Medical Image Analysis", "Generalist Models" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 2, 2, 4 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 4 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "Et0SIGDpP5", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Alessandro Achille", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Applied Research Scientist" }, { "name": "George Karypis", "gender": "Male", "institution": "University of Minnesota, Minneapolis", "country": "", "position": "Full Professor" }, { "name": "Gil Sadeh", "gender": "Male", "institution": "Amazon", "country": "IL", "position": "Researcher" }, { "name": "Huzefa Rangwala", "gender": "Male", "institution": "Amazon", "country": "", "position": "Principal Researcher" }, { "name": "Luca Zancato", "gender": "Male", "institution": "Amazon Web Services", "country": "US", "position": "Applied Scientist" }, { "name": "Yingheng Wang", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Zichen Wang", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 91, "n_ref": 157, "n_ref_all": 177, "n_fig": 3, "n_tab": 7, "L_tab": 2202, "n_element_tab": 219, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1181, "n_element_tab_1": 35, "formula_len_all": 627, "formula_len_all_1": 838, "len_all": 265923, "len_all_1": 61125, "len_abs": 1586, "len_title": 91, "len_sents": 64843, "len_sents_1": 29163, "n_sents": 438, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 35, "L_abs": 1584, "title": "Long-context Protein Language Model", "abs": "Self-supervised training of language models (LMs) has seen great success for protein sequences in learning meaningful representations and for generative drug design. \nMost protein LMs are based on the Transformer architecture trained on individual proteins with short context lengths. Such protein LMs cannot extrapolate to longer proteins and protein complexes well. They also fail to account for the underlying biological mechanisms carried out by biomolecular interactions and dynamics i.e., proteins often interact with other proteins, molecules, and pathways in complex biological systems. \nIn this work, we propose LC-PLM based on an alternative protein LM architecture, BiMamba-S, built off selective structured state-space models, to learn high-quality universal protein representations at the amino acid token level using masked language modeling. We also introduce its graph-contextual variant, LC-PLM-G, which contextualizes protein-protein interaction (PPI) graphs for a second stage of training. LC-PLM demonstrates favorable neural scaling laws, better length extrapolation capability, and a 7\\% to 34\\% improvement on protein downstream tasks than Transformer-based ESM-2. LC-PLM-G further trained within the context of PPI graphs shows promising results on protein structure and function prediction tasks. Our study demonstrates the benefit of increasing the context size with computationally efficient LM architecture (e.g. structured SSMs) in learning universal protein representations and incorporating molecular interaction contexts contained in biological graphs.", "keywords": [ "protein language model" ], "rating_list": [ 3, 3, 3, 8 ], "soundness_list": [ 2, 2, 2, 4 ], "presentation_list": [ 3, 2, 2, 4 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "Essg9kb4yx", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Chenkai Weng", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" }, { "name": "Chongyang Gao", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Kaize Ding", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "Assistant Professor" }, { "name": "Lixu Wang", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" }, { "name": "Qi Zhu", "gender": "unknown", "institution": "Northwestern University", "country": "US", "position": "Full Professor" }, { "name": "Xiao Wang", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 20, "n_ref_uni": 57, "n_ref": 121, "n_ref_all": 147, "n_fig": 6, "n_tab": 16, "L_tab": 7745, "n_element_tab": 366, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 3204, "n_element_tab_1": 102, "formula_len_all": 2129, "formula_len_all_1": 1509, "len_all": 238980, "len_all_1": 80446, "len_abs": 1865, "len_title": 92, "len_sents": 70210, "len_sents_1": 33382, "n_sents": 521, "n_sents_1": 265, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 44, "L_abs": 1936, "title": "On Large Language Model Continual Unlearning", "abs": "While large language models have demonstrated impressive performance across various domains and tasks, their security issues have become increasingly severe. Machine unlearning has emerged as a representative approach for model safety and security, removing the influence of undesired data on the target model. However, these methods do not sufficiently consider that unlearning requests in real-world scenarios are continuously emerging, especially in the context of LLMs, which may lead to accumulated model utility loss that eventually becomes unacceptable. Moreover, existing LLM unlearning methods often ignore previous data access limitations due to privacy concerns and copyright protection. Without previous data, the utility preservation during unlearning is much harder. To overcome these challenges, we propose the O3 framework that includes an \\underline{\\textit{O}}rthogonal low-rank adapter (LoRA) for continually unlearning requested data and an \\underline{\\textit{O}}ut-\\underline{\\textit{O}}f-Distribution (OOD) detector to measure the similarity between input and unlearning data. The orthogonal LoRA achieves parameter disentanglement among continual unlearning requests. The OOD detector is trained with a novel contrastive entropy loss and utilizes a glocal-aware scoring mechanism. During inference, our O3 framework can decide whether and to what extent to load the unlearning LoRA based on the OOD detector's predicted similarity between the input and the unlearned knowledge. Notably, O3's effectiveness does not rely on any retained data. We conducted extensive experiments on O3 and state-of-the-art LLM unlearning methods across three tasks and seven datasets. The results indicate that O3 consistently achieves the best unlearning effectiveness and utility preservation, especially when facing continuous unlearning requests. The source codes can be found at \\url{https://anonymous.4open.science/r/O3-A02B}.", "keywords": [ "Continual Unlearning", "Large Language Models" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "Es4RPNDtmq", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Hayoung Choi", "gender": "Male", "institution": "Kyungpook National University", "country": "KR", "position": "Associate Professor" }, { "name": "Hyun woo Lee", "gender": "Male", "institution": "Kyungpook National University", "country": "KR", "position": "PhD student" }, { "name": "Hyunju Kim", "gender": "Male", "institution": "Korea Institute of Energy Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 52, "n_ref": 77, "n_ref_all": 98, "n_fig": 11, "n_tab": 5, "L_tab": 4617, "n_element_tab": 362, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 3970, "n_element_tab_1": 304, "formula_len_all": 1142, "formula_len_all_1": 938, "len_all": 149203, "len_all_1": 64277, "len_abs": 950, "len_title": 127, "len_sents": 36796, "len_sents_1": 24940, "n_sents": 297, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 959, "title": "Robust Weight Initialization for Tanh Neural Networks with Fixed Point Analysis", "abs": "As a neural network's depth increases, it can achieve high generalization performance. However, training deep networks is challenging due to gradient and signal propagation issues. To address these challenges, extensive theoretical research and various methods have been introduced. Despite these advances, effective weight initialization methods for tanh neural networks remain underexplored. This paper presents a novel weight initialization method for Neural Networks with tanh activation function. Based on an analysis of the fixed points of the function $\\tanh(ax)$, our proposed method aims to determine values of $a$ that mitigate activation saturations. A series of experiments on various classification datasets and Physics-Informed Neural Networks demonstrate that the proposed method outperforms Xavier initialization methods (with or without normalization) in terms of robustness to network size variations, data efficiency, and convergence speed.", "keywords": [ "Weight initialization", "Signal propagation", "Physics informed neural networks" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "ErpRu7qMq1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Ang Lv", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Jiang Bian", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Senior Principal Researcher" }, { "name": "Peiling Lu", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Rui Yan", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Associate Professor" }, { "name": "Shikun Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Wei Ye", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Xu Tan", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 27, "n_ref": 59, "n_ref_all": 83, "n_fig": 6, "n_tab": 5, "L_tab": 836, "n_element_tab": 126, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 438, "n_element_tab_1": 101, "formula_len_all": 534, "formula_len_all_1": 556, "len_all": 119799, "len_all_1": 70494, "len_abs": 1844, "len_title": 136, "len_sents": 42087, "len_sents_1": 34404, "n_sents": 327, "n_sents_1": 251, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1569, "title": "GETMusic: Generating Music Tracks with a Unified Representation and Diffusion Framework", "abs": "Symbolic music generation aims to create musical notes, which can help users compose music, such as generating target instrument tracks based on provided source tracks. In practical scenarios where there\u2019s a predefined ensemble of tracks and various composition needs, an efficient and effective generative model that can generate any target tracks based on the other tracks becomes crucial. However, previous efforts have fallen short in addressing this necessity due to limitations in their music representations and models. In this paper, we introduce a framework known as GETMusic, with \u201cGET\u201d standing for \u201cGEnerate music Tracks.\u201d This framework encompasses a novel music representation \u201cGETScore\u201d and a diffusion model \u201cGETDiff.\u201d GETScore represents musical notes as tokens and organizes tokens in a 2D structure, with tracks stacked vertically and progressing horizontally over time. At a training step, each track of a music piece is randomly selected as either the target or source. The training involves two processes: In the forward process, target tracks are corrupted by masking their tokens, while source tracks remain as the ground truth; in the denoising process, GETDiff is trained to predict the masked target tokens conditioning on the source tracks. Our proposed representation, coupled with the non-autoregressive generative model, empowers GETMusic to generate music with any arbitrary source-target track combinations. Our experiments demonstrate that the versatile GETMusic outperforms prior works proposed for certain specific composition tasks.", "keywords": [ "Symbolic Music Generation", "Symbolic Music Representation", "Diffusion Model" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "EreKmSOw7K", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Chao Zhou", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Postdoc" }, { "name": "Rebekka Burkholz", "gender": "Female", "institution": "Helmholtz Center CISPA for Information Security", "country": "DE", "position": "Associate Professor" }, { "name": "Tom Jacobs", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "PhD student" } ], "n_formula": 48, "n_formula_1": 19, "n_ref_uni": 47, "n_ref": 104, "n_ref_all": 123, "n_fig": 3, "n_tab": 1, "L_tab": 1116, "n_element_tab": 22, "n_fig_1": 1, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4043, "formula_len_all_1": 1399, "len_all": 169054, "len_all_1": 74787, "len_abs": 948, "len_title": 98, "len_sents": 54530, "len_sents_1": 34231, "n_sents": 524, "n_sents_1": 295, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 952, "title": "Time-Dependent Mirror Flows and Where to Find Them", "abs": "Explicit regularization and implicit bias are often studied separately, though in practice, they act in tandem. However, their interplay remains poorly understood. In this work, we show that explicit regularization modifies the behavior of implicit bias and provides a mechanism to control its strength. By incorporating explicit regularization into the mirror flow framework, we present a general approach to better understand implicit biases and their potential in guiding the design of optimization problems. Our primary theoretical contribution is the characterization of regularizations and reparameterizations that induce a time-dependent Bregman function, with a discussion of the implications of its temporal variation. Importantly, our framework encompasses single-layer attention, and application to sparse coding. Extending beyond our core assumptions, we apply this framework to LoRA finetuning, revealing an implicit bias towards sparsity.", "keywords": [ "Mirror flow", "Implicit Bias", "Time-dependent Bregman potential", "Explicit regularization", "LoRA", "Attention", "Sparse coding" ], "rating_list": [ 3, 3, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 1, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "ErQPdaD5wJ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Jicong Fan", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Wei Dai", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "Postdoc" } ], "n_formula": 73, "n_formula_1": 18, "n_ref_uni": 50, "n_ref": 102, "n_ref_all": 145, "n_fig": 11, "n_tab": 9, "L_tab": 16639, "n_element_tab": 824, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 2484, "n_element_tab_1": 339, "formula_len_all": 9377, "formula_len_all_1": 2052, "len_all": 210985, "len_all_1": 72051, "len_abs": 1344, "len_title": 120, "len_sents": 53579, "len_sents_1": 29455, "n_sents": 514, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1376, "title": "AutoUAD: Hyper-parameter Optimization for Unsupervised Anomaly Detection", "abs": "Unsupervised anomaly detection (UAD) has important applications in diverse fields such as manufacturing industry and medical diagnosis. In the past decades, although numerous insightful and effective UAD methods have been proposed, it remains a huge challenge to tune the hyper-parameters of each method and select the most appropriate method among many candidates for a specific dataset, due to the absence of labeled anomalies in the training phase of UAD methods and the high diversity of real datasets. In this work, we aim to address this challenge, so as to make UAD more practical and reliable. We propose two internal evaluation metrics, \\textit{relative-top-median} and \\textit{expected-anomaly-gap}, and one semi-internal evaluation metric, \\textit{normalized pseudo discrepancy} (NPD), as surrogate functions of the expected model performance on unseen test data. For instance, NPD measures the discrepancy between the anomaly scores of a validation set drawn from the training data and a validation set drawn from an isotropic Gaussian. NPD is simple and hyper-parameter-free and is able to compare different UAD methods, and its effectiveness is theoretically analyzed. We integrate the three metrics with Bayesian optimization to effectively optimize the hyper-parameters of UAD models. Extensive experiments on 38 datasets show the effectiveness of our methods.", "keywords": [ "anomaly detection", "hyper-parameter optimization" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "Equ277PBN0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Ana Milanova", "gender": "unknown", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "" }, { "name": "Linh Tran", "gender": "Female", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "PhD student" }, { "name": "Stacy Patterson", "gender": "unknown", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "Associate Professor" }, { "name": "Wei Sun", "gender": "Female", "institution": "IBM Research", "country": "", "position": "Researcher" } ], "n_formula": 17, "n_formula_1": 17, "n_ref_uni": 47, "n_ref": 80, "n_ref_all": 106, "n_fig": 7, "n_tab": 8, "L_tab": 5556, "n_element_tab": 601, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2247, "n_element_tab_1": 300, "formula_len_all": 1026, "formula_len_all_1": 1370, "len_all": 160537, "len_all_1": 65534, "len_abs": 1399, "len_title": 142, "len_sents": 42864, "len_sents_1": 27161, "n_sents": 346, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 94, "L_abs": 1407, "title": "Privacy-Preserving Personalized Federated Prompt Learning for Multimodal Large Language Models", "abs": "Multimodal Large Language Models (LLMs) are pivotal in revolutionizing customer support and operations by integrating multiple modalities such as text, images, and audio. Federated Prompt Learning (FPL) is a recently proposed approach that combines pre-trained multimodal LLMs such as vision-language models with federated learning to create personalized, privacy-preserving AI systems. However, balancing the competing goals of personalization, generalization, and privacy remains a significant challenge. Over-personalization can lead to overfitting, reducing generalizability, while stringent privacy measures, such as differential privacy, can hinder both personalization and generalization. In this paper, we propose a Differentially Private Federated Prompt Learning (DP-FPL) approach to tackle this challenge by leveraging a low-rank factorization scheme to capture generalization while maintaining a residual term that preserves expressiveness for personalization. To ensure privacy, we introduce a novel method where we apply local differential privacy to the two low-rank components of the local prompt, and global differential privacy to the global prompt. Our approach mitigates the impact of privacy noise on the model performance while balancing the tradeoff between personalization and generalization. Extensive experiments demonstrate the effectiveness of our approach over other benchmarks.", "keywords": [ "Multimodal Large Language Model", "Federated Prompt Learning", "Personalization", "Differential Privacy" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "EqcLAU6gyU", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Ao Li", "gender": "Female", "institution": "Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Bolin Ding", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Senior Director" }, { "name": "Fugee Tsung", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Songze Li", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "Full Professor" }, { "name": "Yaliang Li", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Staff Engineer" }, { "name": "Yuexiang Xie", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Staff" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 32, "n_ref": 57, "n_ref_all": 84, "n_fig": 12, "n_tab": 9, "L_tab": 1754, "n_element_tab": 173, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 429, "n_element_tab_1": 46, "formula_len_all": 355, "formula_len_all_1": 355, "len_all": 177231, "len_all_1": 66199, "len_abs": 3550, "len_title": 95, "len_sents": 59121, "len_sents_1": 34735, "n_sents": 405, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 46, "L_abs": 1461, "title": "Agent-Oriented Planning in Multi-Agent Systems", "abs": "Through the collaboration of multiple agents possessing diverse expertise and tools, multi-agent systems achieve impressive progress in solving real-world problems. Given the user queries, the meta-agents, serving as the brain within these systems, are required to decompose the queries into multiple sub-tasks that can be allocated to suitable agents capable of solving them, so-called agent-oriented planning. In this study, we identify three critical design principles of agent-oriented planning, including solvability, completeness, and non-redundancy, to ensure that each sub-task is effectively resolved, leading to satisfactory responses to the original queries. These principles further inspire us to propose a novel framework for agent-oriented planning in multi-agent systems, leveraging a fast task decomposition and allocation process followed by an effective and efficient evaluation via a reward model. During the planning process, the meta-agent is also responsible for evaluating the performance of the expert agents, making timely adjustments to the sub-tasks and scheduling as necessary. Besides, we integrate a feedback loop into the proposed framework to further enhance the effectiveness and robustness of such a problem-solving process. Extensive experiments demonstrate the advancement of the proposed framework in solving real-world problems compared to both single-agent systems and existing planning strategies for multi-agent systems.", "keywords": [ "Multi-Agent System; Planning" ], "rating_list": [ 8, 6, 5, 3, 6 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 1, 2, 3 ], "confidence_list": [ 3, 3, 3, 4, 2 ] }, { "paper_id": "EqCbc4wrzy", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Bin Liu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science", "country": "CN", "position": "Associate Professor" }, { "name": "Cong Cai", "gender": "Male", "institution": "BIT", "country": "", "position": "PhD student" }, { "name": "Hanzhe Xu", "gender": "Male", "institution": "Tianjin Normal University", "country": "", "position": "MS student" }, { "name": "Heng Xie", "gender": "Male", "institution": "Anhui University", "country": "", "position": "MS student" }, { "name": "Jianhua Tao", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Jizhou Cui", "gender": "unknown", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" }, { "name": "Kang Zhu", "gender": "Male", "institution": "Anhui University", "country": "CN", "position": "MS student" }, { "name": "Ruibo Fu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Shan Liang", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "", "position": "Associate Professor" }, { "name": "Xuefei Liu", "gender": "Female", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Yiming Ma", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Yongwei Li", "gender": "Male", "institution": "Chinese Academy of Sciences", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhengqi Wen", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Zhenhua Cheng", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 27, "n_ref": 36, "n_ref_all": 49, "n_fig": 2, "n_tab": 1, "L_tab": 799, "n_element_tab": 113, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1515, "n_element_tab_1": 232, "formula_len_all": 173, "formula_len_all_1": 175, "len_all": 129873, "len_all_1": 58967, "len_abs": 1395, "len_title": 131, "len_sents": 28303, "len_sents_1": 26505, "n_sents": 206, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1402, "title": "MDPE: A Multimodal Deception Dataset with Personality and Emotional Characteristics", "abs": "Deception detection has garnered increasing attention in recent years due to the significant growth of digital media and heightened ethical and security concerns. It has been extensively studied using multimodal methods, including video, audio, and text. In addition, individual differences in deception production and detection are believed to play a crucial role.Although some studies have utilized individual information such as personality traits to enhance the performance of deception detection, current systems remain limited, partly due to a lack of sufficient datasets for evaluating performance. To address this issue, we introduce a multimodal deception dataset MDPE. Besides deception features, this dataset also includes individual differences information in personality and emotional expression characteristics. It can explore the impact of individual differences on deception behavior. It comprises over 104 hours of deception and emotional videos from 193 subjects. Furthermore, we conducted numerous experiments to provide valuable insights for future deception detection research. MDPE not only supports deception detection, but also provides conditions for tasks such as personality recognition and emotion recognition, and can even study the relationships between them. We believe that MDPE will become a valuable resource for promoting research in the field of affective computing.", "keywords": [ "deception detection; affective computing; multimodal dataset" ], "rating_list": [ 3, 1, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 1, 2, 1 ], "contribution_list": [ 3, 1, 3, 1 ], "confidence_list": [ 3, 5, 5, 4 ] }, { "paper_id": "EpnZEzYDUT", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Cong Guan", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "PhD student" }, { "name": "Lei Yuan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Researcher" }, { "name": "Lihe Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Yuqi Bian", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Ziqian Zhang", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Yang Yu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Professor" } ], "n_formula": 35, "n_formula_1": 19, "n_ref_uni": 69, "n_ref": 129, "n_ref_all": 150, "n_fig": 6, "n_tab": 8, "L_tab": 5140, "n_element_tab": 552, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 2406, "n_element_tab_1": 225, "formula_len_all": 4059, "formula_len_all_1": 1799, "len_all": 239245, "len_all_1": 69051, "len_abs": 1515, "len_title": 83, "len_sents": 63769, "len_sents_1": 31075, "n_sents": 409, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1522, "title": "Efficient Multi-agent Offline Coordination via Diffusion-based Trajectory Stitching", "abs": "Learning from offline data without interacting with the environment is a promising way to fully leverage the intelligent decision-making capabilities of multi-agent reinforcement learning (MARL). Previous approaches have primarily focused on developing learning techniques, such as conservative methods tailored to MARL using limited offline data. However, these methods often overlook the temporal relationships across different timesteps and spatial relationships between teammates, resulting in low learning efficiency in imbalanced data scenarios. To comprehensively explore the data structure of MARL and enhance learning efficiency, we propose Multi-Agent offline coordination via Diffusion-based Trajectory Stitching (MADiTS), a novel diffusion-based data augmentation pipeline that systematically generates trajectories by stitching high-quality coordination segments together. MADiTS first generates trajectory segments using a trained diffusion model, followed by applying a bidirectional dynamics constraint to ensure that the trajectories align with environmental dynamics. Additionally, we develop an offline credit assignment technique to identify and optimize the behavior of underperforming agents in the generated segments. This iterative procedure continues until a satisfactory augmented episode trajectory is generated within the predefined limit or is discarded otherwise. Empirical results on imbalanced datasets of multiple benchmarks demonstrate that MADiTS significantly improves MARL performance.", "keywords": [ "Multi-agent Reinforcement Learning", "Offline MARL", "Diffusion based Reinforcement Learning", "Trajectory Stitching" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "EpmbH6DpJI", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Gagandeep Singh", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Assistant Professor" }, { "name": "Yinglun Xu", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "PhD student" }, { "name": "Zhiwei Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" } ], "n_formula": 57, "n_formula_1": 14, "n_ref_uni": 28, "n_ref": 69, "n_ref_all": 77, "n_fig": 3, "n_tab": 3, "L_tab": 704, "n_element_tab": 64, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 2024, "n_element_tab_1": 23, "formula_len_all": 8609, "formula_len_all_1": 1099, "len_all": 147125, "len_all_1": 60984, "len_abs": 1174, "len_title": 127, "len_sents": 48624, "len_sents_1": 25788, "n_sents": 425, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1043, "title": "Robust Thompson Sampling Algorithms Against Reward Poisoning Attacks", "abs": "Thompson sampling is one of the most popular learning algorithms for online sequential decision-making problems and has rich real-world applications. However, current Thompson sampling algorithms are limited by the assumption that the rewards received are uncorrupted, which may not be true in real-world applications where adversarial reward poisoning exists. To make Thompson sampling more reliable, we want to make it robust against adversarial reward poisoning. The main challenge is that one can no longer compute the actual posteriors for the true reward, as the agent can only observe the rewards after corruption. In this work, we solve this problem by computing pseudo-posteriors that are less likely to be manipulated by the attack. We propose robust algorithms based on Thompson sampling for the popular stochastic and contextual linear bandit settings in both cases where the agent is aware or unaware of the budget of the attacker. We theoretically show that our algorithms guarantee near-optimal regret under any attack strategy.", "keywords": [ "Robust Bandit Algorithm", "Thompson Sampling" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 3, 3, 5 ] }, { "paper_id": "EpgoFFUM2q", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Hanghang Tong", "gender": "unknown", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Jingrui He", "gender": "Female", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Full Professor" }, { "name": "Wenxuan Bao", "gender": "Not Specified", "institution": "University of Illinois Urbana Champaign", "country": "US", "position": "PhD student" }, { "name": "Zhichen Zeng", "gender": "unknown", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Zhining Liu", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" } ], "n_formula": 92, "n_formula_1": 28, "n_ref_uni": 44, "n_ref": 125, "n_ref_all": 158, "n_fig": 9, "n_tab": 12, "L_tab": 4823, "n_element_tab": 485, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1683, "n_element_tab_1": 112, "formula_len_all": 9707, "formula_len_all_1": 1553, "len_all": 225063, "len_all_1": 73175, "len_abs": 1523, "len_title": 106, "len_sents": 64468, "len_sents_1": 31392, "n_sents": 588, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 67, "L_abs": 1596, "title": "Matcha: Mitigating Graph Structure Shifts with Test-Time Adaptation", "abs": "Powerful as they are, graph neural networks (GNNs) are known to be vulnerable to distribution shifts. Recently, test-time adaptation (TTA) has attracted attention due to its ability to adapt a pre-trained model to a target domain, without re-accessing the source domain. However, existing TTA algorithms are primarily designed for attribute shifts in vision tasks, where samples are independent. These methods perform poorly on graph data that experience structure shifts, where node connectivity differs between source and target graphs. We attribute this performance gap to the distinct impact of node attribute shifts versus graph structure shifts: the latter significantly degrades the quality of node representations and blurs the boundaries between different node categories. To address structure shifts in graphs, we propose Matcha, an innovative framework designed for effective and efficient adaptation to structure shifts by adjusting the htop-aggregation parameters in GNNs. To enhance the representation quality, we design a prediction-informed clustering loss to encourage the formation of distinct clusters for different node categories. Additionally, Matcha seamlessly integrates with existing TTA algorithms, allowing it to handle attribute shifts effectively while improving overall performance under combined structure and attribute shifts. We validate the effectiveness of Matcha on both synthetic and real-world datasets, demonstrating its robustness across various combinations of structure and attribute shifts. Our code\nis available at https://github.com/baowenxuan/Matcha.", "keywords": [ "test-time adaptation", "distribution shifts", "structure shifts", "graph neural networks" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 2 ], "confidence_list": [ 1, 4, 4, 2, 3 ] }, { "paper_id": "EoTIlDT0Tr", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Baoyuan Wu", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Associate Professor" }, { "name": "Siwei Lyu", "gender": "Male", "institution": "State University of New York, Buffalo", "country": "US", "position": "Full Professor" }, { "name": "Yize Chen", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "MS student" }, { "name": "Zhiyuan Yan", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 6, "n_ref_uni": 37, "n_ref": 115, "n_ref_all": 139, "n_fig": 10, "n_tab": 8, "L_tab": 3541, "n_element_tab": 631, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 1803, "n_element_tab_1": 268, "formula_len_all": 225, "formula_len_all_1": 298, "len_all": 200631, "len_all_1": 82819, "len_abs": 2704, "len_title": 121, "len_sents": 52865, "len_sents_1": 33894, "n_sents": 418, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 2362, "title": "$\\mathcal{X}^2$-DFD: A framework for e$\\mathcal{X}$plainable and e$\\mathcal{X}$tendable Deepfake Detection", "abs": "Detecting deepfakes (*i.e.*, AI-generated content with malicious intent) has become an important task. Most existing detection methods provide only real/fake predictions without offering human-comprehensible explanations. Recent studies leveraging multimodal large-language models (MLLMs) for deepfake detection have shown improvements in explainability. However, the performance of pre-trained MLLMs (*e.g.*, LLaVA) remains limited due to a lack of understanding of their capabilities for this task and strategies to enhance them. In this work, we empirically assess the strengths and weaknesses of MLLMs specifically in deepfake detection via forgery-related feature analysis. Building on these assessments, we propose a novel framework called $\\mathcal{X}^2$-DFD, consisting of three core modules. \nThe first module, *Model Feature Assessment (MFA)*, measures the detection capabilities of forgery-related features intrinsic to MLLMs, and gives a descending ranking of these features. \nThe second module, *Strong Feature Strengthening (SFS)*, enhances the detection and explanation capabilities by fine-tuning the MLLM on a dataset constructed based on the top-ranked features. \nThe third module, *Weak Feature Supplementing (WFS)*, improves the fine-tuned MLLM's capabilities on lower-ranked features by integrating external dedicated deepfake detectors. \nTo verify the effectiveness of this framework, we further present a practical implementation, where an automated forger-related feature generation, evaluation, and ranking procedure is designed for *MFA* module; an automated generation procedure of the fine-tuning dataset containing real and fake images with explanations based on top-ranked features is developed for *SFS* model; an external conventional deepfake detector focusing on blending artifact, which corresponds to a low detection capability in the pre-trained MLLM, is integrated for *WFS* module. \nExperimental results show that the proposed implementation enhances overall detection performance compared to pre-trained MLLMs, while providing more convincing explanations. \nMore encouragingly, our framework is designed to be plug-and-play, allowing it to seamlessly integrate with more advanced MLLMs and external detectors, leading to continual improvement and extension to face the challenges of rapidly evolving deepfake technologies.", "keywords": [ "Deepfake Detection; Multimodal Large Language Models; Media Forensics" ], "rating_list": [ 3, 8, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "EoPsCAEYae", "primary_area": "foundation or frontier models, including LLMs", "n_author": 13, "author_info_list": [ { "name": "Fangxun Shu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Hao Jiang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Haoyuan Li", "gender": "Not Specified", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Jiang Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Juncheng Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Lei Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Siliang Tang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Tianwei Lin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Wenqiao Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yueting Zhuang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zheqi Lv", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "robotx", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Yuzhelun", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 25, "n_ref": 62, "n_ref_all": 97, "n_fig": 12, "n_tab": 15, "L_tab": 5392, "n_element_tab": 784, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 2586, "n_element_tab_1": 404, "formula_len_all": 904, "formula_len_all_1": 549, "len_all": 216271, "len_all_1": 73157, "len_abs": 1601, "len_title": 139, "len_sents": 51587, "len_sents_1": 31123, "n_sents": 403, "n_sents_1": 230, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1756, "title": "HyperLLaVA: Dynamic Visual and Language Expert Tuning for Multimodal Large Language Models", "abs": "Recent advancements indicate that scaling up Multimodal Large Language Models (MLLMs) effectively enhances performance on downstream multimodal tasks. \nThe prevailing MLLM paradigm, \\emph{e.g.}, LLaVA, transforms visual features into text-like tokens using a \\emph{static} vision-language mapper, thereby enabling \\emph{static} LLMs to develop the capability to comprehend visual information through visual instruction tuning. \nUnfortunately, the \\emph{static} paradigm shares the same parameters to underly multi-task instruction tuning, inevitably introducing the potential \\emph{task interference} or \\emph{negative transfer}, \\emph{i.e.}, where an improvement in the performance of one task reduces the performance of other tasks. \nIn light of this, we introduce \\textbf{HyperLLaVA}, which in conjunction with a dynamic visual expert and language expert, respectively adjusts the parameters of the projector and LLM layers conditioned on diverse instruction semantics, thereby minimizing the task interference. \nThese experts are derived from HyperNetworks, which adaptively generates dynamic parameter shifts through visual and language guidance, enabling dynamic vision-language alignment and instruction tuning in two-stage training. \nTo deeply study the multi-task interference of MLLM, we build the \\textbf{Comprehensive Multimodal Task benchmark} (\\texttt{CMT}), a comprehensive benchmark for the evaluation of multidimensional multimodal tasks. \nThe experiments demonstrate that \nthe superiority of the dynamic tuning paradigm for multi-task instruction following on \\texttt{CMT} and general MLLM benchmarks. Our project is available at \\href{https://anonymous.4open.science/r/HyperLLaVA-D58E}{https://anonymous.4open.science/r/HyperLLaVA-D58E}.", "keywords": [ "Multimodal Large Language Model" ], "rating_list": [ 3, 5, 3, 3, 5 ], "soundness_list": [ 1, 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 5, 5, 4 ] }, { "paper_id": "Em6GkQfLKM", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Hamid Palangi", "gender": "Male", "institution": "Google", "country": "US", "position": "Staff Research Scientist" }, { "name": "Shima Imani", "gender": "Female", "institution": "Microsoft Research", "country": "US", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 43, "n_ref": 60, "n_ref_all": 81, "n_fig": 6, "n_tab": 1, "L_tab": 90, "n_element_tab": 8, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 552, "formula_len_all_1": 552, "len_all": 131301, "len_all_1": 59851, "len_abs": 1380, "len_title": 64, "len_sents": 46485, "len_sents_1": 30338, "n_sents": 400, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1391, "title": "Exploring Group and Symmetry Principles in Large Language Models", "abs": "Large Language Models (LLMs) have demonstrated impressive performance across a wide range of applications; however, assessing their reasoning capabilities remains a significant challenge. In this paper, we introduce a framework grounded in group and symmetry principles, which have played a crucial role in fields such as physics and mathematics, and offer another way to evaluate their capabilities. While the proposed framework is general, to showcase the benefits of employing these properties, we focus on arithmetic reasoning and investigate the performance of these models on four group properties: closure, identity, inverse, and associativity. Our findings reveal that LLMs studied in this work struggle to preserve group properties across different test regimes. In the closure test, we observe biases towards specific outputs and an abrupt degradation in their performance from $100\\%$ to $0\\%$ after a specific sequence length. They also perform poorly in the identity test, which represents adding irrelevant information in the context, and show sensitivity when subjected to inverse test, which examines the robustness of the model with respect to negation. In addition, we demonstrate that breaking down problems into smaller steps helps LLMs in the associativity test that we have conducted. To support these tests we have developed a synthetic dataset which will be released.", "keywords": [ "probing", "robustness", "LLM" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ElYRG3pJcv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Anji Liu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Haowei Lin", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Jiaqi Li", "gender": "Not Specified", "institution": "Beijing Institute for General Artificial Intelligence", "country": "CN", "position": "Researcher" }, { "name": "Ruilin Yan", "gender": "unknown", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Weiye Shi", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Xiangyu Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Xiaojian Ma", "gender": "Unspecified", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Yitao Liang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zihao Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "weipeng chen", "gender": "Male", "institution": "Beijing Baichuan Intelligence Technology Co., Ltd.", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 9, "n_formula_1": 12, "n_ref_uni": 48, "n_ref": 122, "n_ref_all": 148, "n_fig": 3, "n_tab": 6, "L_tab": 2800, "n_element_tab": 222, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 756, "n_element_tab_1": 167, "formula_len_all": 757, "formula_len_all_1": 855, "len_all": 233386, "len_all_1": 72986, "len_abs": 5686, "len_title": 126, "len_sents": 70120, "len_sents_1": 33142, "n_sents": 523, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1280, "title": "Optimizing Inference-Time Reasoning in LLMs via Retrieval-Augmented Reflection", "abs": "Empowering LLMs to improve their performance through increased inference-time computation is a crucial step in developing self-improving agents capable of operating in open-ended natural language contexts. In this paper, we explore how iteratively revising a chain of thoughts guided by information retrieval significantly improves large language models' reasoning ability in challenging tasks, while hugely mitigating hallucination. In particular, the proposed method --- \\emph{retrieval-augmented reflection} (RaR) --- revises the generation tokens step by step, leveraging multiple pieces of retrieved information relevant to the intermediate reasoning steps and the instruction. Applying RaR during inference-time to a various set of language models substantially improves their performances on various reasoning tasks; on relatively increasing scores by up to +16.4\\% on code generation, +11.6\\% on mathematical reasoning, and 29.1\\% on embodied task planning. Moreover, we find that with more inference-time computation given to the LLM for multi-times retrieval-augmented reflection, the LLM can continuously improve on various reasoning benchmarks. A small LM can surpass the performance of the LM with more than 10 times the parameters, when giving more computation cost.", "keywords": [ "Retrieval-augmented Generation", "Reasoning", "Large Language Models" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 1, 3, 2, 1 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "ElDpb1BWE3", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Feipeng Qi", "gender": "Male", "institution": "Nuclear Power Institute of China", "country": "CN", "position": "Researcher" }, { "name": "Tailin Wu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yongjun Jiao", "gender": "Male", "institution": "China National Nuclear Corporation", "country": "CN", "position": "Researcher" }, { "name": "Taozhang", "gender": "Male", "institution": "China National Nuclear Corporation", "country": "CN", "position": "MS student" }, { "name": "zhenhai liu", "gender": "unknown", "institution": "Nuclear Power Institute of China", "country": "CN", "position": "Researcher" } ], "n_formula": 28, "n_formula_1": 21, "n_ref_uni": 51, "n_ref": 95, "n_ref_all": 133, "n_fig": 16, "n_tab": 21, "L_tab": 4572, "n_element_tab": 483, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 341, "n_element_tab_1": 47, "formula_len_all": 2522, "formula_len_all_1": 1370, "len_all": 202652, "len_all_1": 64841, "len_abs": 1715, "len_title": 116, "len_sents": 61029, "len_sents_1": 30580, "n_sents": 464, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1726, "title": "Compositional Generative Multiphysics and Multi-component Simulation", "abs": "Multiphysics simulation, which models the interactions between multiple physical processes, and multi-component simulation of complex structures are critical in fields like nuclear and aerospace engineering. Previous studies often rely on numerical solvers or machine learning-based surrogate models to solve or accelerate these simulations. However, multiphysics simulations typically require integrating multiple specialized solvers\u2014each responsible for evolving a specific physical process\u2014into a coupled program, which introduces significant development challenges. Furthermore, no universal algorithm exists for multi-component simulations, which adds to the complexity.\nHere we propose compositional Multiphysics and Multi-component Simulation with Diffusion models (MultiSimDiff) to overcome these challenges. During diffusion-based training, MultiSimDiff learns energy functions modeling the conditional probability of one physical process/component conditioned on other processes/components. In inference, MultiSimDiff generates coupled multiphysics solutions and multi-component structures by sampling from the joint probability distribution, achieved by composing the learned energy functions in a structured way.\nWe test our method in three tasks. In the reaction-diffusion and nuclear thermal coupling problems, MultiSimDiff successfully predicts the coupling solution using decoupled data, while the surrogate model fails in the more complex second problem. For the thermal and mechanical analysis of the prismatic fuel element, MultiSimDiff trained for single component prediction accurately predicts a larger structure with 64 components, reducing the relative error by 40.3% compared to the surrogate model.", "keywords": [ "multiphysics", "multi-component", "PDE simulation", "physical simulation", "generative" ], "rating_list": [ 5, 5, 8, 5, 5 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 1, 2, 4, 3, 3 ], "confidence_list": [ 4, 3, 4, 3, 3 ] }, { "paper_id": "El4Cs8Su3r", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Hendrik Strobelt", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Visiting Scientist" }, { "name": "Hilde Kuehne", "gender": "Female", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "Full Professor" }, { "name": "Sofian Chaybouti", "gender": "Male", "institution": "Johann Wolfgang Goethe Universit\u00e4t Frankfurt am Main", "country": "DE", "position": "PhD student" }, { "name": "Walid Bousselham", "gender": "Male", "institution": "Johann Wolfgang Goethe Universit\u00e4t Frankfurt am Main", "country": "DE", "position": "PhD student" }, { "name": "Angie Wynne Boggust", "gender": "Female", "institution": "Computer Science and Artificial Intelligence Laboratory, Electrical Engineering & Computer Science", "country": "US", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 41, "n_ref": 80, "n_ref_all": 104, "n_fig": 7, "n_tab": 6, "L_tab": 2899, "n_element_tab": 371, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 994, "n_element_tab_1": 95, "formula_len_all": 840, "formula_len_all_1": 430, "len_all": 143369, "len_all_1": 58004, "len_abs": 954, "len_title": 139, "len_sents": 43567, "len_sents_1": 26232, "n_sents": 312, "n_sents_1": 182, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 960, "title": "LeGrad: An Explainability Method for Vision Transformers via Feature Formation Sensitivity", "abs": "Vision Transformers (ViTs) have become a standard architecture in computer vision. However, because of their modeling of long-range dependencies through self-attention mechanisms, the explainability of these models remains a challenge.\nTo address this, we propose LeGrad, an explainability method specifically designed for ViTs. \nLeGrad computes the gradient with respect to the attention maps of single ViT layers, considering the gradient itself as the explainability signal.\nWe aggregate the signal over all layers, combining the activations of the last as well as intermediate tokens to produce the merged explainability map.\nThis makes LeGrad a conceptually simple and an easy-to-implement method to enhance the transparency of ViTs. \nWe evaluate LeGrad in various setups, including segmentation, perturbation, and open-vocabulary settings, showcasing its improved spatial fidelity as well as its versatility compared to other SotA explainability methods.", "keywords": [ "segmentation;gradient-based method" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "EkfLaCJ7bk", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Lijie Yang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Undergrad student" }, { "name": "Zhihao Jia", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Zhihao Zhang", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Zhuofu Chen", "gender": "unknown", "institution": "Tongji University", "country": "CN", "position": "Undergrad student" }, { "name": "Zikun Li", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 19, "n_ref": 33, "n_ref_all": 66, "n_fig": 12, "n_tab": 10, "L_tab": 4406, "n_element_tab": 479, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 885, "n_element_tab_1": 52, "formula_len_all": 195, "formula_len_all_1": 47, "len_all": 116702, "len_all_1": 51693, "len_abs": 1552, "len_title": 134, "len_sents": 38191, "len_sents_1": 25569, "n_sents": 240, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 85, "L_abs": 1563, "title": "TidalDecode: Fast and Accurate LLM Decoding with Position Persistent Sparse Attention", "abs": "Large language models (LLMs) have driven significant advancements across diverse NLP tasks, with long-context models gaining prominence for handling extended inputs. However, the expanding key-value (KV) cache size required by Transformer architectures intensifies the memory constraints, particularly during the decoding phase, creating a significant bottleneck. Existing sparse attention mechanisms designed to address this bottleneck have two limitations: (1) they often fail to reliably identify the most relevant tokens for attention, and (2) they overlook the spatial coherence of token selection across consecutive Transformer layers, which can lead to performance degradation and substantial overhead in token selection. This paper introduces TidalDecode, a simple yet effective algorithm and system for fast and accurate LLM decoding through position persistent sparse attention. TidalDecode leverages the spatial coherence of tokens selected by existing sparse attention methods and introduces a few token selection layers that perform full attention to identify the tokens with the highest attention scores, while all other layers perform sparse attention with the pre-selected tokens. This design enables TidalDecode to substantially reduce the overhead of token selection for sparse attention without sacrificing the quality of the generated results. Evaluation on a diverse set of LLMs and tasks shows that TidalDecode closely matches the generative performance of full attention methods while reducing the LLM decoding latency by up to $2.1\\times$.", "keywords": [ "efficient transformer serving", "sparse attention decoding" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "Ek50sQQI1w", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Jason Gao", "gender": "Male", "institution": "Marine Biological Laboratory", "country": "CN", "position": "Intern" }, { "name": "Jiajun Tian", "gender": "Male", "institution": "DataSelect AI", "country": "CN", "position": "Researcher" }, { "name": "Liz Li", "gender": "unknown", "institution": "DataSelect AI", "country": "CN", "position": "Researcher" }, { "name": "Mige Zhu", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Yi Ge", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Research Assistant" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 30, "n_ref": 62, "n_ref_all": 72, "n_fig": 2, "n_tab": 3, "L_tab": 702, "n_element_tab": 110, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 702, "n_element_tab_1": 110, "formula_len_all": 756, "formula_len_all_1": 756, "len_all": 117966, "len_all_1": 65411, "len_abs": 1316, "len_title": 125, "len_sents": 29945, "len_sents_1": 29948, "n_sents": 236, "n_sents_1": 237, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1320, "title": "A Novel Listwise Alignment Approach for Language Models with Explicit Rewards", "abs": "Existing alignment techniques, including Direct Preference Optimization (DPO), are primarily designed for pairwise preference data where rewards are inferred rather than explicitly provided. In this paper, we propose a comprehensive framework for aligning large language models (LLMs) by introducing a new optimization objective that facilitates the processing of reward datasets, which consist of a list of responses explicitly marked with scalar preference scores. Our contribution includes the development of a novel algorithm, termed Soft Preference Optimization (LPO), which allows for the direct derivation of an LLM policy from both reward and preference datasets. At the heart of LPO is a unique listwise preference optimization objective formulated using an exponential-logarithmic function and an adaptive loss coefficient, which effectively integrates listwise preference signals into the LLM. We assess the efficacy of our approach under both reward and preference scenarios using different sizes of Mistral models. Experimental results indicate that our method outperforms several preference-based benchmarks, particularly when reward datasets are utilized. Additionally, our method demonstrates a significant advantage over DPO in intricate reasoning tasks, such as mathematical problem-solving and coding.", "keywords": [ "large language models", "preference alignment", "listwise optimization objective" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "EjJGND0m1x", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Barak A. Pearlmutter", "gender": "Male", "institution": "Maynooth University", "country": "IE", "position": "Full Professor" }, { "name": "Mrinal Mathur", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Sergey M. Plis", "gender": "Male", "institution": "Georgia State University", "country": "US", "position": "Full Professor" } ], "n_formula": 37, "n_formula_1": 12, "n_ref_uni": 68, "n_ref": 92, "n_ref_all": 121, "n_fig": 10, "n_tab": 20, "L_tab": 4090, "n_element_tab": 470, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 337, "n_element_tab_1": 39, "formula_len_all": 1840, "formula_len_all_1": 640, "len_all": 240105, "len_all_1": 55617, "len_abs": 1259, "len_title": 107, "len_sents": 76712, "len_sents_1": 27572, "n_sents": 539, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 59, "L_abs": 1371, "title": "MIND over Body: Adaptive Thinking using Dynamic Computation", "abs": "While the human brain efficiently handles various computations with a limited number of neurons, traditional deep learning networks require a significant increase in parameters to improve performance.\n Yet, these parameters are used inefficiently as the networks employ the same amount of computation for inputs of the same size, regardless of the input's complexity.\n We address this inefficiency by introducing self-introspection capabilities to the network, enabling it to adjust the number of used parameters based on the internal representation of the task and adapt the computation time based on the task complexity.\n This enables the network to adaptively reuse parameters across tasks, dynamically adjusting the computational effort to match the complexity of the input.\n We demonstrate the effectiveness of this method on language modeling and computer vision tasks.\n Notably, our model achieves 96.62\\% accuracy on ImageNet with just a three-layer network, surpassing much larger ResNet-50 and EfficientNet. When applied to a transformer architecture, the approach achieves 95.8\\%/88.7\\% F1 scores on the SQuAD v1.1/v2.0 datasets at negligible parameter cost.\n These results showcase the potential for dynamic and reflective computation, contributing to the creation of intelligent systems that efficiently manage resources based on input data complexity.", "keywords": [ "Interpretability", "Fixed points", "Dynamic routing", "Dynamic input processing", "Deep Learning Framework" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 4 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "EjJD16oaly", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "Chang Yu", "gender": "unknown", "institution": "Westlake University", "country": "CN", "position": "Postdoc" }, { "name": "Cheng Tan", "gender": "Male", "institution": "Zhejiang University & Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Di Wu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Juanxi Tian", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Intern" }, { "name": "Siyuan Li", "gender": "Male", "institution": "Westlake University & Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Weiyang Jin", "gender": "Male", "institution": "New York University", "country": "US", "position": "Intern" }, { "name": "Zedong Wang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zicheng Liu", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "PhD student" }, { "name": "Stan Z. Li", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Chair Professor" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 56, "n_ref": 123, "n_ref_all": 149, "n_fig": 10, "n_tab": 10, "L_tab": 5648, "n_element_tab": 439, "n_fig_1": 10, "n_tab_1": 8, "L_tab_1": 3858, "n_element_tab_1": 169, "formula_len_all": 807, "formula_len_all_1": 755, "len_all": 184258, "len_all_1": 71251, "len_abs": 1325, "len_title": 81, "len_sents": 42708, "len_sents_1": 30247, "n_sents": 305, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1378, "title": "GTR: Semi-supervised Learning with Grouping and Transporting for Robust Thresholding", "abs": "Semi-supervised learning (SSL) digs unlabeled data by pseudo-labeling when labeled data is limited. Despite various auxiliary strategies enhancing SSL training, the main challenge is how to determine reliable pseudo labels through a robust thresholding algorithm based on quality indicators (e.g., confidence scores). However, the existing strategies for distinguishing low or high-quality labels through simple grouping indicators remain in trivial design, ignoring the characteristics of the data distribution itself, which cannot guarantee robustness and efficiency. To this end, we group the quality indicators of pseudo labels into three clusters (easy, semi-hard, and hard) and statistically reveal the real bottleneck of threshold selection, i.e., the sensitivity of semi-hard samples, through empirical analysis. We propose an adaptive Grouping and Transporting method that Robustly selects semi-hard samples with test-time augmentations and consistency constraints while saving the selection budgets of easy and hard samples, dubbed as GTR. Our proposed GTR can effectively determine high-quality data when applied to existing SSL methods while reducing redundant costs in the selection. Extensive experiments on 11 SSL benchmarks across three modalities verify that GTR can achieve significant performance gains and speedups over Pseudo Label, FixMatch, and FlexMatch.", "keywords": [ "Semi-supervised Learning", "Grouping", "Thresholding", "Plug-and-play", "Pseudo-labeling" ], "rating_list": [ 3, 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 3, 4, 2 ] }, { "paper_id": "EjHtQlKEzV", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Dongwei Xu", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "", "position": "Associate Professor" }, { "name": "Hao Cheng", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "CN", "position": "PhD student" }, { "name": "Jiaheng Wei", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Qi Xuan", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoniu Yang", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "Principal Researcher" }, { "name": "Yao Lu", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yujie Fang", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Zeyu Wang", "gender": "Male", "institution": "Zhejiang University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Zhaowei Zhu", "gender": "Male", "institution": "Docta.ai", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 57, "n_ref": 127, "n_ref_all": 153, "n_fig": 3, "n_tab": 14, "L_tab": 17881, "n_element_tab": 905, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 10359, "n_element_tab_1": 488, "formula_len_all": 352, "formula_len_all_1": 379, "len_all": 210949, "len_all_1": 75502, "len_abs": 1348, "len_title": 107, "len_sents": 42267, "len_sents_1": 26951, "n_sents": 418, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1314, "title": "Reassessing Layer Pruning in LLMs: New Insights and Methods", "abs": "Although large language models (LLMs) have achieved remarkable success across various domains, their considerable scale necessitates substantial computational resources, posing significant challenges for deployment in resource-constrained environments. Layer pruning, as a simple yet effective compression method, removes layers of a model directly, reducing computational overhead. However, what are the best practices for layer pruning in LLMs? Are sophisticated layer selection metrics truly effective? Does the LoRA (Low-Rank Approximation) family, widely regarded as a leading method for pruned model fine-tuning, truly meet expectations when applied to post-pruning fine-tuning? To answer these questions, we dedicate thousands of GPU hours to benchmarking layer pruning in LLMs and gaining insights across multiple dimensions. Our results demonstrate that a simple approach, i.e., pruning the final 25\\% of layers followed by fine-tuning the \\texttt{lm\\_head} and the remaining last three layer, yields remarkably strong performance. Following this guide, we prune Llama-3.1-8B-It and obtain a model that outperforms many popular LLMs of similar size, such as ChatGLM2-6B, Vicuna-7B-v1.5, Qwen1.5-7B and Baichuan2-7B.\nWe release the optimal model weights on Huggingface, and the code is available on GitHub.", "keywords": [ "LLM Pruning", "Layer Pruning" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 5, 5, 3 ] }, { "paper_id": "EjCrfVFZTx", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 1, "author_info_list": [ { "name": "Jason Phang", "gender": "Male", "institution": "OpenAI", "country": "", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 16, "n_ref": 46, "n_ref_all": 62, "n_fig": 5, "n_tab": 3, "L_tab": 506, "n_element_tab": 66, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 690, "n_element_tab_1": 90, "formula_len_all": 184, "formula_len_all_1": 184, "len_all": 126898, "len_all_1": 71037, "len_abs": 936, "len_title": 107, "len_sents": 33981, "len_sents_1": 28008, "n_sents": 219, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 940, "title": "Investigating the Effectiveness of HyperTuning via Gisting", "abs": "Gisting (Mu et al., 2023) is a simple method for training models to compress information into fewer token representations using a modified attention mask, and can serve as an economical approach to training Transformer-based hypernetworks. We introduce HyperLlama, a set of Gisting-based hypernetworks built on Llama-2 models that generates task-specific soft prefixes based on few-shot inputs. In experiments across P3, Super-NaturalInstructions and Symbol Tuning datasets, we show that HyperLlama models can effectively compress information from few-shot examples into soft prefixes. However, they still underperform multi-task fine-tuned language models with full attention over few-shot in-context examples. We also show that HyperLlama-generated soft prefixes can serve as better initializations for further prefix tuning. Overall, Gisting-based hypernetworks are economical and easy to implement, but have mixed empirical performance.", "keywords": [ "hypernetworks", "llm", "parameter-efficient fine-tuning", "prefix tuning" ], "rating_list": [ 6, 5, 3, 5, 3 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 3, 3, 3 ] }, { "paper_id": "EispKqtw5B", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Haoxing Chen", "gender": "Male", "institution": "Ant Group (AliPay)", "country": "CN", "position": "Researcher" }, { "name": "Huijia Zhu", "gender": "Female", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Jiaqi Li", "gender": "Male", "institution": "China Mobile Research Institute", "country": "CN", "position": "Researcher" }, { "name": "Weiqiang Wang", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Zizheng Huang", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "PhD student" }, { "name": "jun lan", "gender": "Male", "institution": "AntGroup", "country": "CN", "position": "Researcher" }, { "name": "Limin Wang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 54, "n_ref": 141, "n_ref_all": 149, "n_fig": 1, "n_tab": 7, "L_tab": 3520, "n_element_tab": 519, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2523, "n_element_tab_1": 397, "formula_len_all": 601, "formula_len_all_1": 600, "len_all": 137222, "len_all_1": 64488, "len_abs": 1441, "len_title": 128, "len_sents": 30913, "len_sents_1": 27023, "n_sents": 220, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1437, "title": "Stochastic Layer-Wise Shuffle: A Good Practice to Improve Vision Mamba Training", "abs": "Recent Vision Mamba models not only have much lower complexity for processing higher resolution images and longer videos but also the competitive performance with Vision Transformers (ViTs). However, they are stuck into overfitting and thus only present up to base size (about 80M). It is still unclear how vanilla Vision Mamba (Vim) can be efficiently scaled up to larger sizes, which is essentially for further exploitation. In this paper, we propose a stochastic layer-wise shuffle regularization, which empowers successfully scaling non-hierarchical Vision Mamba to a large size (about 300M) in a supervised setting. Specifically, our base and large-scale ShuffleMamba models can outperform the supervised ViTs of similar size by 0.8% and 1.0% classification accuracy on ImageNet1k, respectively, without auxiliary data. When evaluated on the ADE20K semantic segmentation and COCO detection tasks, our ShuffleMamba models also show significant improvements. Without bells and whistles, the stochastic layer-wise shuffle has the following highlights: (1) Plug and play: it does not change model architectures and will be omitted in inference. (2) Simple but effective: it can improve the overfitting in Vim training and only introduce random token permutation operations. (3) Intuitive: the token sequences in deeper layers are more likely to be shuffled as they are expected to be more semantic and less sensitive to patch positions.", "keywords": [ "Vision Mamba", "Supervised Learning", "Training Regularization", "Computer Vision" ], "rating_list": [ 1, 3, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "EiYr9ArUFl", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 1, "author_info_list": [ { "name": "Pierre Wolinski", "gender": "Male", "institution": "Universit\u00e9 Paris Dauphine - PSL", "country": "FR", "position": "Associate Professor" } ], "n_formula": 122, "n_formula_1": 16, "n_ref_uni": 31, "n_ref": 55, "n_ref_all": 83, "n_fig": 11, "n_tab": 5, "L_tab": 584, "n_element_tab": 66, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 275, "n_element_tab_1": 7, "formula_len_all": 8424, "formula_len_all_1": 1249, "len_all": 188275, "len_all_1": 70272, "len_abs": 1321, "len_title": 137, "len_sents": 62685, "len_sents_1": 31563, "n_sents": 612, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 1331, "title": "Gathering and Exploiting Higher-Order Information when Training Large Structured Models", "abs": "When training large models, such as neural networks, \nthe full derivatives of order 2 and beyond are usually inaccessible,\ndue to their computational cost.\nThis is why, among the second-order optimization methods, it is very common\nto bypass the computation of the Hessian by using \nfirst-order information, such as the gradient of the parameters (e.g., quasi-Newton methods)\nor the activations (e.g., K-FAC).\n\nIn this paper, we focus on the exact and explicit computation\nof projections of the Hessian and higher-order derivatives on\nwell-chosen subspaces, which are relevant for optimization.\nNamely, for a given partition of the set of parameters, \nit is possible to compute tensors which can be seen as\n\"higher-order derivatives according to the partition\",\nat a reasonable cost as long as the number of subsets of \nthe partition remains small.\n\nThen, we propose an optimization method exploiting\nthese tensors at order 2 and 3 with several interesting properties, including:\nit outputs a learning rate per subset of parameters, which can\nbe used for hyperparameter tuning;\nit takes into account long-range interactions\nbetween the layers of the trained neural network, \nwhich is usually not the case in similar methods (e.g., K-FAC);\nthe trajectory of the optimization is invariant under \naffine layer-wise reparameterization.", "keywords": [ "neural networks", "Hessian", "learning rate", "projections", "optimization" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 2, 4, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "Ei9KiIzgxK", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Ahmet H. G\u00fczel", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Jack Parker-Holder", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Ilija Bogunovic", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "Lecturer" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 40, "n_ref": 76, "n_ref_all": 85, "n_fig": 5, "n_tab": 4, "L_tab": 1451, "n_element_tab": 160, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 790, "n_element_tab_1": 82, "formula_len_all": 692, "formula_len_all_1": 717, "len_all": 115263, "len_all_1": 52635, "len_abs": 1544, "len_title": 131, "len_sents": 30314, "len_sents_1": 23147, "n_sents": 174, "n_sents_1": 141, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1546, "title": "Synthetic Data is Sufficient for Zero-Shot Visual Generalization from Offline Data", "abs": "Offline reinforcement learning (RL) offers a promising framework for training agents using pre-collected datasets without the need for further environment interaction. However, policies trained on offline data often struggle to generalise\ndue to limited exposure to diverse states. The complexity of visual data introduces additional challenges such as noise, distractions, and spurious correlations, which can misguide the policy and increase the risk of overfitting if the training data is not sufficiently diverse. Indeed, this makes it challenging to leverage vision-based offline data in training robust agents that can generalize to unseen environments. To solve this problem, we propose a simple approach\u2014generating additional synthetic data. We propose a two-step process, first $augmenting$ the originally collected offline data to improve zero-shot generalization by introducing diversity, then using a diffusion model to $generate$ additional data in latent space. We test our method across both continuous action spaces (Visual D4RL) and discrete action spaces (Procgen), demonstrating that it significantly improves generalization without requiring any algorithmic changes to existing model-free offline RL methods. We show that our method not only increases the diversity of the training data but also significantly reduces the generalization gap at test time while maintaining computational efficiency. We believe this approach could fuel additional progress in generating synthetic data to train more general agents in the future.", "keywords": [ "Offline Reinforcement Learning", "Generalization", "Data Augmentation", "Synthetic Data Generation" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 5, 3, 3, 3 ] }, { "paper_id": "EhweLJiYi5", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Pedro Seber", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Richard Braatz", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 60, "n_ref": 139, "n_ref_all": 170, "n_fig": 4, "n_tab": 10, "L_tab": 2777, "n_element_tab": 380, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1385, "n_element_tab_1": 121, "formula_len_all": 308, "formula_len_all_1": 32, "len_all": 204363, "len_all_1": 74230, "len_abs": 1720, "len_title": 95, "len_sents": 64832, "len_sents_1": 36826, "n_sents": 448, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1351, "title": "LCEN: A Novel Feature Selection Algorithm for Nonlinear, Interpretable Machine Learning Models", "abs": "Interpretable models can have advantages over black-box models, and interpretability is essential for the application of machine learning in critical settings, such as aviation or medicine. LASSO and elastic net, the most commonly used interpretable methods, are limited to linear predictions and have poor feature selection capabilities. Other important interpretable methods, such as tree-based or generalized additive models, are nonlinear but have limited performance in some scenarios. In this work, we introduce the LASSO-Clip-EN (LCEN) algorithm for the construction of nonlinear, interpretable machine learning models. LCEN is tested on a wide variety of artificial and empirical datasets, frequently creating more accurate, sparser models than other models, including those for building sparse, nonlinear models. LCEN is robust against many issues typically present in datasets and modeling, including noise, multicollinearity, data scarcity, and hyperparameter variance. LCEN is also able to rediscover multiple physical laws from empirical data and, for processes with no known physical laws, LCEN achieves better results than many other dense and sparse methods -- including using 10.8-fold fewer features than dense methods and 8.1-fold fewer features than EN on one dataset, and is comparable to or better than ANNs on multiple datasets.", "keywords": [ "Machine Learning", "Feature Selection", "Elastic Net", "Interpretable Machine Learning", "Interpretability", "Applications of Machine Learning", "Applied Machine Learning" ], "rating_list": [ 5, 5, 5, 1 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 3, 5, 5 ] }, { "paper_id": "EhSUM1FcJw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Quang-Binh Nguyen", "gender": "Male", "institution": "Ho Chi Minh city University of Science, Vietnam National University", "country": "", "position": "Undergrad student" }, { "name": "Tam V. Nguyen", "gender": "Male", "institution": "University of Dayton", "country": "US", "position": "Associate Professor" }, { "name": "Trong-Vu Hoang", "gender": "Male", "institution": "Ho Chi Minh city University of Science, Vietnam National University", "country": "VN", "position": "Undergrad student" }, { "name": "Trung-Nghia Le", "gender": "Male", "institution": "University of Science", "country": "VN", "position": "Researcher" }, { "name": "Minh-triet Tran", "gender": "Male", "institution": "Ho Chi Minh city University of Science, Vietnam National University", "country": "VN", "position": "Associate Professor" }, { "name": "Thanh-Toan Do", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Assistant Professor" } ], "n_formula": 40, "n_formula_1": 15, "n_ref_uni": 28, "n_ref": 136, "n_ref_all": 190, "n_fig": 29, "n_tab": 4, "L_tab": 315, "n_element_tab": 24, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2268, "formula_len_all_1": 1028, "len_all": 148940, "len_all_1": 52506, "len_abs": 2074, "len_title": 113, "len_sents": 52856, "len_sents_1": 24240, "n_sents": 379, "n_sents_1": 158, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1573, "title": "ConceptFlow: Unified Framework for Personalized Image Generation", "abs": "Personalized image generation is an appealing area of research within controllable image generation due to its diverse potential applications. Despite notable advancements, generating images based on single or multiple concepts remains challenging. For single-concept generation, it is difficult to strike a balance between identity preservation and prompt alignment, especially in complex prompts. When it comes to multiple concepts, creating images from a single prompt without extra conditions, such as layout boxes or semantic masks, is problematic due to significantly identity loss and concept omission. In this paper, we introduce ConceptFlow, a comprehensive framework designed to tackle these challenges. Specifically, we propose ConceptFlow-S and ConceptFlow-M for single-concept generation and multiple-concept generation, respectively. ConceptFlow-S introduces a KronA-WED adapter, which integrates a Kronecker adapter with weight and embedding decomposition, and employs a disentangled learning approach with a novel attention regularization objective to enhance single-concept generation. On the other hand, ConceptFlow-M leverages models learned from ConceptFlow-S to directly generate multi-concept images without needed of additional conditions, proposing Subject-Adaptive Matching Attention (SAMA) module and layout consistency guidance strategy. Our extensive experiments and user study show that ConceptFlow effectively addresses the aforementioned issues, enabling its application in various real-world scenarios such as advertising and garment try-on.", "keywords": [ "Personalized Image Generation", "Visual Guidance" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 3, 1 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "Eh1QM3OK51", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Demetrio Labate", "gender": "Male", "institution": "University of Houston", "country": "US", "position": "Full Professor" }, { "name": "Dhananjaya Jayasundara", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" }, { "name": "Heng Zhao", "gender": "Male", "institution": "The Rockefeller University", "country": "US", "position": "Machine Learning Enginer" }, { "name": "Vishal M Patel", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 41, "n_ref": 83, "n_ref_all": 104, "n_fig": 6, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 400, "formula_len_all_1": 399, "len_all": 109987, "len_all_1": 62484, "len_abs": 1337, "len_title": 123, "len_sents": 34144, "len_sents_1": 31951, "n_sents": 198, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1342, "title": "PIN: Prolate Spheroidal Wave Function-based Implicit Neural Representations", "abs": "Implicit Neural Representations (INRs) provide a continuous mapping between the coordinates of a signal and the corresponding values. As the performance of INRs heavily depends on the choice of nonlinear-activation functions, there has been a significant focus on encoding explicit signals within INRs using diverse activation functions. Despite recent advancements, existing INRs often encounter significant challenges, particularly at fine scales where they often introduce noise-like artifacts over smoother areas compromising the quality of the output. Moreover, they frequently struggle to generalize to unseen coordinates. These drawbacks highlight a critical area for further research and development to enhance the robustness and applicability of INRs across diverse scenarios. To address this challenge, we introduce the Prolate Spheroidal Wave Function-based Implicit Neural Representations (PIN), which exploits the optimal space-frequency domain concentration of Prolate Spheroidal Wave Functions (PSWFs) as the nonlinear mechanism in INRs. Our experimental results reveal that PIN excels not only in representing images and 3D shapes but also significantly outperforms existing methods in various vision tasks that require INR generalization, including image inpainting, novel view synthesis, edge detection, and image denoising.", "keywords": [ "Prolate Spheroidal Wave Functions", "Implicit Neural Representations", "MLPs" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 5, 2 ] }, { "paper_id": "Egd7Vi1EuA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Bing Qin", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Danyang Zhao", "gender": "Female", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "FENGLEI FAN", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Ming Ma", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Sendong Zhao", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "Associate Professor" }, { "name": "Ting Liu", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Yanrui Du", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "", "position": "PhD student" }, { "name": "\u66f9\u5bb6\u709c", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Undergrad student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 27, "n_ref": 52, "n_ref_all": 83, "n_fig": 11, "n_tab": 14, "L_tab": 5744, "n_element_tab": 568, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1597, "n_element_tab_1": 189, "formula_len_all": 170, "formula_len_all_1": 168, "len_all": 138691, "len_all_1": 72676, "len_abs": 1689, "len_title": 140, "len_sents": 39465, "len_sents_1": 31280, "n_sents": 345, "n_sents_1": 255, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1618, "title": "Towards Secure Tuning: Mitigating Security Risks Arising from Benign Instruction Fine-Tuning", "abs": "Instruction Fine-Tuning (IFT) has become an essential method for adapting base Large Language Models (LLMs) into variants for professional and private use. However, researchers have raised concerns over a significant decrease in LLMs' security following IFT, even when the IFT process involves entirely benign instructions (termed Benign IFT). Our study represents a pioneering effort to mitigate the security risks arising from Benign IFT. Specifically, we conduct a Module Robustness Analysis, aiming to investigate how LLMs' internal modules contribute to their security. Based on our analysis, we propose a novel IFT strategy, called the Modular Layer-wise Learning Rate (ML-LR) strategy. In our analysis, we implement a simple security feature classifier that serves as a proxy to measure the robustness of modules (e.g. $Q$/$K$/$V$, etc.). Our findings reveal that the module robustness shows clear patterns, varying regularly with the module type and the layer depth. Leveraging these insights, we develop a proxy-guided search algorithm to identify a robust subset of modules, termed $Mods_{Robust}$. During IFT, the ML-LR strategy employs differentiated learning rates for $Mods_{Robust}$ and the rest modules. Our experimental results show that in security assessments, the application of our ML-LR strategy significantly mitigates the rise in harmfulness of LLMs following Benign IFT. Notably, our ML-LR strategy has little impact on the usability or expertise of LLMs following Benign IFT. Furthermore, we have conducted comprehensive analyses to verify the soundness and flexibility of our ML-LR strategy.", "keywords": [ "Large Language Models", "Security", "Instruction Fine-Tuning" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "EgP6IEyfYJ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Binghui Wang", "gender": "Male", "institution": "Illinois Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Jane Downer", "gender": "Female", "institution": "Illinois Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Ren Wang", "gender": "Male", "institution": "Illinois Institute of Technology", "country": "US", "position": "Assistant Professor" } ], "n_formula": 30, "n_formula_1": 17, "n_ref_uni": 45, "n_ref": 72, "n_ref_all": 102, "n_fig": 9, "n_tab": 9, "L_tab": 1013, "n_element_tab": 88, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 348, "n_element_tab_1": 50, "formula_len_all": 1866, "formula_len_all_1": 1055, "len_all": 174080, "len_all_1": 72267, "len_abs": 1412, "len_title": 124, "len_sents": 50919, "len_sents_1": 31963, "n_sents": 430, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1422, "title": "WATERMARKING GRAPH NEURAL NETWORKS VIA EXPLANATIONS FOR OWNERSHIP PROTECTION", "abs": "Graph Neural Networks (GNNs) are the mainstream method to learn pervasive graph data and are widely deployed in industry, making their intellectual property valuable. However, protecting GNNs from unauthorized use remains a challenge. Watermarking, which embeds ownership information into a model, is a potential solution. However, existing watermarking methods have two key limitations: First, almost all of them focus on non-graph data, with watermarking GNNs for complex graph data largely unexplored. Second, the de facto backdoor-based watermarking methods pollute training data and induce ownership ambiguity through intentional misclassification. Our explanation-based watermarking inherits the strengths of backdoor-based methods (e.g., robust to watermark removal attacks), but avoids data pollution and eliminates intentional misclassification. In particular, our method learns to embed the watermark in GNN explanations such that this unique watermark is statistically distinct from other potential solutions, and ownership claims must show statistical significance to be verified. We theoretically prove that, even with full knowledge of our method, locating the watermark is an NP-hard problem. Empirically, our method manifests robustness to removal attacks like fine-tuning and pruning. By addressing these challenges, our approach marks a significant advancement in protecting GNN intellectual property.", "keywords": [ "Watermarking", "GNNs", "Ownership", "Verification", "Ownership Verification", "Explanation", "Graph" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 3, 3, 2 ] }, { "paper_id": "EgJhwYR2tB", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Chen-Yu Lee", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Dhruv Madeka", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Amazon" }, { "name": "Lei Li", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Long Le", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Rishabh Agarwal", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Adjunct Professor" }, { "name": "Rujun Han", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Tomas Pfister", "gender": "Male", "institution": "Google", "country": "", "position": "Head of Research @ Cloud AI" }, { "name": "Wenda Xu", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Zifeng Wang", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "William Yang Wang", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 33, "n_ref": 82, "n_ref_all": 123, "n_fig": 7, "n_tab": 7, "L_tab": 1738, "n_element_tab": 245, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 560, "n_element_tab_1": 51, "formula_len_all": 187, "formula_len_all_1": 109, "len_all": 228732, "len_all_1": 64065, "len_abs": 1337, "len_title": 147, "len_sents": 56533, "len_sents_1": 31494, "n_sents": 474, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1341, "title": "Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Sampling", "abs": "Recent advances in knowledge distillation (KD) have enabled smaller student models to approach the performance of larger teacher models. However, popular methods such as supervised KD and on-policy KD, are adversely impacted by the knowledge gaps between teacher-student in practical scenarios. Supervised KD suffers from a distribution mismatch between training with a static dataset and inference over final student-generated outputs. Conversely, on-policy KD, which uses student-generated samples for training, can suffer from low-quality training examples with which teacher models are not familiar, resulting in inaccurate teacher feedback. To address these limitations, we introduce Speculative Knowledge Distillation (SKD), a novel approach that leverages cooperation between student and teacher models to generate high-quality training data on-the-fly while aligning with the student's inference-time distribution. In SKD, the student proposes tokens, and the teacher replaces poorly ranked ones based on its own distribution, transferring high-quality knowledge adaptively. We evaluate SKD on various text generation tasks, including translation, summarization, math, and instruction following, and show that SKD consistently outperforms existing KD methods across different domains, data sizes, and model initialization strategies.", "keywords": [ "LLM", "Knowledge Distillation", "On-policy" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 4, 4, 4 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 2, 4, 5 ] }, { "paper_id": "EgEyoZvyDw", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chanho Kim", "gender": "Male", "institution": "Oregon State University", "country": "US", "position": "Researcher" }, { "name": "Hung Nguyen", "gender": "unknown", "institution": "Oregon State University", "country": "", "position": "PhD student" }, { "name": "Rigved Naukarkar", "gender": "Male", "institution": "Oregon State University", "country": "US", "position": "MS student" }, { "name": "Fuxin Li", "gender": "Male", "institution": "Oregon State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 33, "n_ref": 60, "n_ref_all": 80, "n_fig": 2, "n_tab": 5, "L_tab": 651, "n_element_tab": 79, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 948, "n_element_tab_1": 101, "formula_len_all": 732, "formula_len_all_1": 731, "len_all": 112848, "len_all_1": 65025, "len_abs": 1170, "len_title": 97, "len_sents": 34338, "len_sents_1": 29082, "n_sents": 265, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1042, "title": "Long-Term 3D Point Tracking By Cost Volume Fusion", "abs": "Long-term point tracking is essential to understand non-rigid motion in the physical world better. Deep learning approaches have recently been incorporated into long-term point tracking, but most prior work predominantly functions in 2D. Although these methods benefit from the well-established backbones and matching frameworks, the motions they produce do not always make sense in the 3D physical world. In this paper, we propose the first deep learning framework for long-term point tracking in 3D that generalizes to new points and videos without requiring test-time fine-tuning. Our model contains a cost volume fusion module that effectively integrates multiple past appearances and motion information via a transformer architecture, significantly enhancing overall tracking performance. In terms of 3D tracking performance, our model significantly outperforms simple scene flow chaining and previous 2D point tracking methods, even if one uses ground truth depth and camera pose to backproject 2D point tracks in a synthetic scenario.", "keywords": [ "3d point tracking", "scene flow" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Eg32tDGgF5", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Edward Moroshko", "gender": "Male", "institution": "Electrical Engineering Department, Technion \u2013 Israel Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Fasih Haider", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Postdoc" }, { "name": "Sotirios A. Tsaftaris", "gender": "unknown", "institution": "University of Edinburgh", "country": "GB", "position": "Professor in machine learning and computer vision" }, { "name": "Yuyang Xue", "gender": "unknown", "institution": "University of Edinburgh, University of Edinburgh", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 30, "n_ref": 36, "n_ref_all": 58, "n_fig": 4, "n_tab": 4, "L_tab": 2044, "n_element_tab": 148, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2091, "n_element_tab_1": 154, "formula_len_all": 262, "formula_len_all_1": 262, "len_all": 112555, "len_all_1": 73269, "len_abs": 870, "len_title": 99, "len_sents": 33647, "len_sents_1": 33641, "n_sents": 240, "n_sents_1": 240, "venue": "ICLR2025", "decision": "withdraw", "L_title": 51, "L_abs": 874, "title": "DO GENERATIVE MODELS LEARN RARE GENERATIVE FACTORS?", "abs": "Generative models are becoming a promising tool in AI alongside discriminative learning. Several models have been proposed to learn in an unsupervised fashion the corresponding generative factors, namely the latent variables critical for capturing the full spectrum of data variability. Diffusion Models (DMs), Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) are of particular interest due to their impressive ability to generate highly realistic data. Through a systematic empirical study, this paper delves into the intricate challenge of how DMs, GANs and VAEs internalize and replicate rare generative factors. Our findings reveal a pronounced tendency towards the memorization of these factors. We study the reasons for this memorization and demonstrate that strategies such as spectral decoupling can mitigate this issue to a certain extent.", "keywords": [ "Representation Learning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "EeqlkPpaV8", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Baoxiang Wang", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Huanjian Zhou", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Masashi Sugiyama", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Director" } ], "n_formula": 113, "n_formula_1": 33, "n_ref_uni": 38, "n_ref": 84, "n_ref_all": 94, "n_fig": 1, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 1141, "n_element_tab_1": 54, "formula_len_all": 12394, "formula_len_all_1": 2943, "len_all": 181235, "len_all_1": 72076, "len_abs": 1040, "len_title": 108, "len_sents": 49930, "len_sents_1": 28273, "n_sents": 555, "n_sents_1": 253, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 47, "L_abs": 1044, "title": "The adaptive complexity of log-concave sampling", "abs": "In large-data applications, such as the inference process of diffusion models, it is desirable to design sampling algorithms with a high degree of parallelization. In this work, we study the adaptive complexity of sampling, which is the minimum number of sequential rounds required to achieve sampling given polynomially many queries executed in parallel at each round. For unconstrained sampling, we examine distributions that are log-smooth or log-Lipschitz and log strongly or non-strongly concave. We show that an almost linear iteration algorithm cannot return a sample with a specific exponentially small error under total variation distance. For box-constrained sampling, we show that an almost linear iteration algorithm cannot return a sample with sup-polynomially small error under total variation distance for log-concave distributions. Our proof relies upon novel analysis with the characterization of the output for the hardness potentials based on the chain-like structure with random partition and classical smoothing techniques.", "keywords": [ "sampling", "adaptive complexity", "computational statistics" ], "rating_list": [ 6, 8, 8, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "EeDSMy5Ruj", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 10, "author_info_list": [ { "name": "Aidan Z.H. Yang", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Brando Miranda", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Hakim Hacid", "gender": "Male", "institution": "TII", "country": "AE", "position": "Principal Researcher" }, { "name": "Jingwei Zuo", "gender": "Male", "institution": "Technology Innovation Institute", "country": "AE", "position": "Researcher" }, { "name": "Joseph Rotella", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Intern" }, { "name": "Leonardo de Moura", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Mohamed El Amine Seddik", "gender": "Male", "institution": "Technology Innovation Institute", "country": "", "position": "Researcher" }, { "name": "Shi Hu", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "PhD student" }, { "name": "Soonho Kong", "gender": "Male", "institution": "Amazon Web Services", "country": "US", "position": "Principal Researcher" }, { "name": "Zhizhen Qin", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 17, "n_ref": 43, "n_ref_all": 60, "n_fig": 4, "n_tab": 6, "L_tab": 2501, "n_element_tab": 240, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 218, "n_element_tab_1": 26, "formula_len_all": 22, "formula_len_all_1": 22, "len_all": 130685, "len_all_1": 62468, "len_abs": 1872, "len_title": 36, "len_sents": 50425, "len_sents_1": 32817, "n_sents": 317, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 1881, "title": "Synthetic Theorem Generation in Lean", "abs": "The application of large language models (LLMs) to theorem proving presents a promising avenue for advancing formal mathematics. Interactive theorem provers, such as Lean, offer a rigorous framework within which these models can assist in or automate proof discovery, grounding their reasoning capabilities in a sound, verifiable formal system. However, the potential of LLMs in this domain is constrained by the limited availability of formal proof corpora for training. To address this limitation, we introduce a synthetic theorem generator capable of producing novel Lean theorems and their corresponding proofs. Our approach employs forward reasoning to synthesize new propositions from premises drawn from existing Lean libraries. We explore candidate reasoning steps using a search strategy that optimizes for diversity of output, apply them in a linear fashion that avoids irrelevant proof steps, and assess their effect by meta-programmatically executing corresponding Lean tactics. These methods enable the generation of an arbitrary number of new theorems and proofs across various mathematical domains, using common Lean proof tactics while ensuring the correctness of generated theorems by construction. We demonstrate the efficacy of the generated theorems and training data by fine-tuning models on synthetic theorems and evaluating them on the miniF2F-test benchmark. Our results show improvements in theorem-proving capabilities, with accuracy increasing from 37.3% to 38.5% for the Falcon2-11B model trained solely on Mathlib, and from 38.1% to 39.3% for the same model trained on a mix of rich datasets. These improvements highlight the value of our diverse synthetic data in augmenting limited existing corpora of formal proofs, providing complementary information that enhances LLMs' performance on theorem-proving tasks even when combined with other datasets.", "keywords": [ "theorem proving", "large language model", "synthetic data generation", "Lean" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "EdNSQHaaMR", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Kuk-Jin Yoon", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" }, { "name": "Wooseong Jeong", "gender": "Male", "institution": "The Korea Advanced Institute of Science and Technology (KAIST)", "country": "KR", "position": "PhD student" } ], "n_formula": 106, "n_formula_1": 20, "n_ref_uni": 32, "n_ref": 88, "n_ref_all": 152, "n_fig": 9, "n_tab": 8, "L_tab": 3661, "n_element_tab": 561, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 283, "n_element_tab_1": 45, "formula_len_all": 12328, "formula_len_all_1": 1472, "len_all": 240128, "len_all_1": 70741, "len_abs": 1592, "len_title": 104, "len_sents": 66199, "len_sents_1": 32393, "n_sents": 572, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 1599, "title": "Selective Task Group Updates for Multi-Task Optimization", "abs": "Multi-task learning enables the acquisition of task-generic knowledge by training multiple tasks within a unified architecture. However, training all tasks together in a single architecture can lead to performance degradation, known as negative transfer, which is a main concern in multi-task learning. Previous works have addressed this issue by optimizing the multi-task network through gradient manipulation or weighted loss adjustments. However, their optimization strategy focuses on addressing task imbalance in shared parameters, neglecting the learning of task-specific parameters. As a result, they show limitations in mitigating negative transfer, since the learning of shared space and task-specific information influences each other during optimization. To address this, we propose a different approach to enhance multi-task performance by selectively grouping tasks and updating them for each batch during optimization. We introduce an algorithm that adaptively determines how to effectively group tasks and update them during the learning process. To track inter-task relations and optimize multi-task networks simultaneously, we propose proximal inter-task affinity, which can be measured during the optimization process. We provide a theoretical analysis on how dividing tasks into multiple groups and updating them sequentially significantly affects multi-task performance by enhancing the learning of task-specific parameters. Our methods substantially outperform previous multi-task optimization approaches and are scalable to different architectures and various numbers of tasks.", "keywords": [ "Multi-Task Learning", "Multi-Task Optimization", "Proximal Inter-Task Affinity" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "EdMb9TqqDY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Fan Ma", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Postdoc" }, { "name": "Kaixin Shen", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Linchao Zhu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Yucheng Suo", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 65, "n_ref": 103, "n_ref_all": 133, "n_fig": 10, "n_tab": 11, "L_tab": 1268, "n_element_tab": 162, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 283, "n_element_tab_1": 28, "formula_len_all": 640, "formula_len_all_1": 185, "len_all": 187628, "len_all_1": 64820, "len_abs": 1560, "len_title": 0, "len_sents": 46437, "len_sents_1": 29548, "n_sents": 428, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1573, "title": "Long-horizon Visual Instruction Generation with Logic and Attribute Self-reflection", "abs": "Visual instructions for long-horizon tasks are crucial as they intuitively clarify complex concepts and enhance retention across extended steps. \nDirectly generating a series of images using text-to-image models without considering the context of previous steps results in inconsistent images, increasing cognitive load. Additionally, the generated images often miss objects or the attributes such as color, shape, and state of the objects are inaccurate.\nTo address these challenges, we propose LIGER, the first training-free framework for Long-horizon Instruction GEneration with logic and attribute self-Reflection. LIGER first generates a draft image for each step with the historical prompt and visual memory of previous steps. This step-by-step generation approach maintains consistency between images in long-horizon tasks. Moreover, LIGER utilizes various image editing tools to rectify errors including wrong attributes, logic errors, object redundancy, and identity inconsistency in the draft images. Through this self-reflection mechanism, LIGER improves the logic and object attribute correctness of the images.\nTo verify whether the generated images assist human understanding, we manually curated a new benchmark consisting of various long-horizon tasks. Human-annotated ground truth expressions reflect the human-defined criteria for how an image should appear to be illustrative. \nExperiments demonstrate the visual instructions generated by LIGER are more comprehensive compared with baseline methods. The code and dataset will be available once accepted.", "keywords": [ "text to image generation", "visual instruction generation" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "EdKSI2ijUY", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Charles Sun", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Undergrad student" }, { "name": "Charlie Victor Snell", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "PhD student" }, { "name": "Isadora White", "gender": "Non-Binary", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Joey Hong", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Kelvin Xu", "gender": "Unspecified", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Marwa Abdulhai", "gender": "unknown", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Yuexiang Zhai", "gender": "unknown", "institution": "University of California Berkeley", "country": "", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 69, "n_ref": 143, "n_ref_all": 155, "n_fig": 4, "n_tab": 1, "L_tab": 400, "n_element_tab": 70, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 775, "n_element_tab_1": 124, "formula_len_all": 55, "formula_len_all_1": 55, "len_all": 200086, "len_all_1": 78079, "len_abs": 1636, "len_title": 128, "len_sents": 36860, "len_sents_1": 36855, "n_sents": 257, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1652, "title": "LMRL Gym: Benchmarks for Multi-Turn Reinforcement Learning with Language Models", "abs": "Large language models (LLMs) provide excellent text-generation capabilities, but standard prompting and generation methods generally do not lead to intentional or goal-directed agents and might necessitate considerable prompt tuning. Even the best current LLMs rarely ask clarifying questions, engage in explicit information gathering, or take actions that lead to better decisions after multiple turns. Reinforcement learning has the potential to leverage the powerful modeling capabilities of LLMs, as well as their internal representation of textual interactions, to create capable goal-directed language agents. This can enable intentional and temporally extended interactions, such as with humans, the emergence of complex skills such as persuasion, and long-horizon strategic behavior, such as in the context of games. Enabling this requires the community to develop reliable reinforcement learning algorithms for training LLMs. Developing such algorithms requires tasks that can gauge progress on algorithm design, provide accessible and reproducible evaluations for multi-turn interactions, and cover a range of task properties and challenges in improving reinforcement learning algorithms. Our paper introduces the LMRL-Gym benchmark for evaluating multi-turn RL for LLMs, together with an open-source research framework for getting started on multi-turn RL with offline value-based and online policy-based RL methods. Our benchmark consists of 3 Interactive Dialogue tasks and 5 RL Capability tests for a total of 8 tasks, which require multiple rounds of language interaction and cover a range of tasks in open-ended dialogue and text games.", "keywords": [ "benchmarks", "LLMs", "RL" ], "rating_list": [ 5, 3, 6, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 4 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "EcrdmRT99M", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Floriano Tori", "gender": "Male", "institution": "Vrije Universiteit Brussel", "country": "", "position": "PhD student" }, { "name": "Vincent Ginis", "gender": "Male", "institution": "Vrije Universiteit Brussel", "country": "", "position": "Assistant Professor" }, { "name": "Vincent Holst", "gender": "unknown", "institution": "Vrije Universiteit Brussel", "country": "", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 13, "n_ref_uni": 43, "n_ref": 91, "n_ref_all": 107, "n_fig": 4, "n_tab": 3, "L_tab": 1772, "n_element_tab": 116, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1201, "n_element_tab_1": 69, "formula_len_all": 1058, "formula_len_all_1": 899, "len_all": 129779, "len_all_1": 60626, "len_abs": 1489, "len_title": 143, "len_sents": 37239, "len_sents_1": 28861, "n_sents": 260, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1519, "title": "The Effectiveness of Curvature-Based Rewiring and the Role of Hyperparameters in GNNs Revisited", "abs": "Message passing is the dominant paradigm in Graph Neural Networks (GNNs). The efficiency of message passing, however, can be limited by the topology of the graph. This happens when information is lost during propagation due to being oversquashed when travelling through bottlenecks. To remedy this, recent efforts have focused on graph rewiring techniques, which disconnect the input graph originating from the data and the computational graph, on which message passing is performed. A prominent approach for this is to use discrete graph curvature measures, of which several variants have been proposed, to identify and rewire around bottlenecks, facilitating information propagation. While oversquashing has been demonstrated in synthetic datasets, in this work we reevaluate the performance gains that curvature-based rewiring brings to real-world datasets. We show that in these datasets, edges selected during the rewiring process are not in line with theoretical criteria identifying bottlenecks. This implies they do not necessarily oversquash information during message passing. Subsequently, we demonstrate that SOTA accuracies on these datasets are outliers originating from sweeps of hyperparameters\u2014both the ones for training and dedicated ones related to the rewiring algorithm\u2014instead of consistent performance gains. In conclusion, our analysis nuances the effectiveness of curvature-based rewiring in real-world datasets and brings a new perspective on the methods to evaluate GNN accuracy improvements.", "keywords": [ "Geometric deep learning", "Graph Neural Networks", "Graph Rewiring", "Curvature" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Ecb6HBoo1r", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Jinzhuo Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Rui Peng", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Wei Wu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 26, "n_ref": 42, "n_ref_all": 66, "n_fig": 7, "n_tab": 3, "L_tab": 1146, "n_element_tab": 220, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 832, "n_element_tab_1": 169, "formula_len_all": 517, "formula_len_all_1": 501, "len_all": 106762, "len_all_1": 57733, "len_abs": 1585, "len_title": 106, "len_sents": 32108, "len_sents_1": 26684, "n_sents": 216, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1590, "title": "Deciphering Cell Lineage Gene Regulatory Network via MTGRN", "abs": "Gene regulatory network (GRN) inference is crucial for cell fate decision, as it outlines the regulations between genes, which direct cell differentiation. Although there have been some work to infer cell lineage GRN, they fail to capture the continuous nature of the differentiation process as they group cells by cell type or cluster and infer GRN in a discrete manner. In this paper, we hypothesize GRN can forecast future gene expression based on history information and transform the inference process into a multivariate time series forecasting problem, linking cells at different time to learn temporal dynamics and inferring GRN in a continuous process. We introduce MTGRN, a transformer-based model that only takes single cell data as input to infer the cell lineage GRN by forecasting gene expression. MTGRN consists of temporal blocks and spatial blocks, effectively captures the connections between cells along their developmental trajectories and leverages prior knowledge to elucidate regulatory interactions among genes. It significantly outperforms six other methods across five datasets, demonstrating superior performance even compared to multimodal approaches. Based on the inferred GRN, MTGRN pinpoints three crucial genes associated with the development of mouse embryonic stem cells and depicts the activity changes of these genes during cellular differentiation. Beyond this, MTGRN is capable of conducting perturbation experiments on key genes and accurately modeling the change of cell identity following the knockout of the Gata1 in mouse hematopoietic stem cells.", "keywords": [ "Gene regulatory network", "Time series", "In silico perturbation" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 1, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "EbxYDBhE3S", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Biao Yi", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "PhD student" }, { "name": "Sishuo Chen", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Tiansheng Huang", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Tong Li", "gender": "Male", "institution": "Nankai University", "country": "CN", "position": "Associate Professor" }, { "name": "Yiming Li", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Research Fellow" }, { "name": "Zheli Liu", "gender": "unknown", "institution": "Nankai University", "country": "", "position": "Full Professor" }, { "name": "Zhixuan Chu", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 40, "n_ref": 92, "n_ref_all": 128, "n_fig": 9, "n_tab": 18, "L_tab": 5439, "n_element_tab": 379, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2593, "n_element_tab_1": 198, "formula_len_all": 686, "formula_len_all_1": 549, "len_all": 233277, "len_all_1": 74323, "len_abs": 1822, "len_title": 151, "len_sents": 75587, "len_sents_1": 32342, "n_sents": 595, "n_sents_1": 260, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 103, "L_abs": 1838, "title": "Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models", "abs": "Backdoor unalignment attacks against Large Language Models (LLMs) enable the stealthy compromise of safety alignment using a hidden trigger while evading normal safety auditing. These attacks pose significant threats to the applications of LLMs in the real-world Large Language Model as a Service (LLMaaS) setting, where the deployed model is a fully black-box system that can only interact through text. Furthermore, the sample-dependent nature of the attack target exacerbates the threat. Instead of outputting a fixed label, the backdoored LLM follows the semantics of any malicious command with the hidden trigger, significantly expanding the target space. In this paper, we introduce BEAT, a black-box defense that detects triggered samples during inference to deactivate the backdoor. It is motivated by an intriguing observation (dubbed the **probe concatenate effect**), where concatenated triggered samples significantly reduce the refusal rate of the backdoored LLM towards a malicious probe, while non-triggered samples have little effect. Specifically, BEAT identifies whether an input is triggered by measuring the degree of distortion in the output distribution of the probe before and after concatenation with the input. Our method addresses the challenges of sample-dependent targets from an opposite perspective. It captures the impact of the trigger on the refusal signal (which is sample-independent) instead of sample-specific successful attack behaviors. It overcomes black-box access limitations by using multiple sampling to approximate the output distribution. Extensive experiments are conducted on various backdoor attacks and LLMs (including the closed-source GPT-3.5-turbo), verifying the effectiveness and efficiency of our defense. Our source code is available at https://anonymous.4open.science/r/BEAT-0065.", "keywords": [ "Backdoor Unalignment", "Backdoor Defense", "Instruction-tuned LLMs", "AI Safety", "AI Security" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 2, 3 ] }, { "paper_id": "EbWf36quzd", "primary_area": "generative models", "n_author": 20, "author_info_list": [ { "name": "Chris Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Hongsheng Li", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Associate Professor" }, { "name": "Jingwen He", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Junjun He", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Junlin Xie", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Le Zhuo", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Longtian Qiu", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "PhD student" }, { "name": "Renrui Zhang", "gender": "Male", "institution": "MMLab of CUHK & Shanghai AI Laboratory", "country": "HK", "position": "PhD student" }, { "name": "Rongjie Huang", "gender": "Male", "institution": "FAIR", "country": "US", "position": "Researcher" }, { "name": "Ruoyi Du", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "PhD student" }, { "name": "Shijie Geng", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Tianshuo Yang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Tong He", "gender": "Male", "institution": "Shanghai AI lab", "country": "CN", "position": "Researcher" }, { "name": "Weicai Ye", "gender": "Male", "institution": "KwaiVGI, Kuaishou Technology", "country": "CN", "position": "Researcher" }, { "name": "Wenqi Shao", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Xu Luo", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuhang Zhang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Gao Peng", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" }, { "name": "Zhengkai Jiang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 6, "n_ref_uni": 109, "n_ref": 292, "n_ref_all": 335, "n_fig": 19, "n_tab": 5, "L_tab": 2845, "n_element_tab": 223, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 3485, "n_element_tab_1": 193, "formula_len_all": 704, "formula_len_all_1": 294, "len_all": 345291, "len_all_1": 51768, "len_abs": 1894, "len_title": 143, "len_sents": 76740, "len_sents_1": 23417, "n_sents": 465, "n_sents_1": 115, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 94, "L_abs": 1954, "title": "Lumina-T2X: Scalable Flow-based Large Diffusion Transformer for Flexible Resolution Generation", "abs": "Sora unveils the potential of scaling Diffusion Transformer (DiT) for generating photorealistic images and videos at arbitrary resolutions, aspect ratios, and durations, yet it still lacks sufficient implementation details. In this paper, we introduce the Lumina-T2X family -- a series of Flow-based Large Diffusion Transformers (Flag-DiT) equipped with zero-initialized attention, as a simple and scalable generative framework that can be adapted to various modalities, e.g., transforming noise into images, videos, multi-view 3D objects, or audio clips conditioned on text instructions. By tokenizing the latent spatial-temporal space and incorporating learnable placeholders such as |[nextline]| and |[nextframe]| tokens, Lumina-T2X seamlessly unifies the representations of different modalities across various spatial-temporal resolutions. Advanced techniques like RoPE, KQ-Norm, and flow matching enhance the stability, flexibility, and scalability of Flag-DiT, enabling models of Lumina-T2X to scale up to 7 billion parameters and extend the context window to 128K tokens. This is particularly beneficial for creating ultra-high-definition images with our Lumina-T2I model and long 720p videos with our Lumina-T2V model. Remarkably, Lumina-T2I, powered by a 5-billion-parameter Flag-DiT, requires only 35% of the training computational costs of a 600-million-parameter naive DiT (PixArt-alpha), indicating that increasing the number of parameters significantly accelerates convergence of generative models without compromising visual quality. Our further comprehensive analysis underscores Lumina-T2X's preliminary capability in resolution extrapolation, high-resolution editing, generating consistent 3D views, and synthesizing videos with seamless transitions. All code and checkpoints of Lumina-T2X are released at https://github.com/Alpha-VLLM/Lumina-T2X to further foster creativity, transparency, and diversity in the generative AI community.", "keywords": [ "Generative Models", "Text-to-Image Generation", "Diffusion Models", "Flow Matching" ], "rating_list": [ 6, 8, 6, 8, 8 ], "soundness_list": [ 3, 4, 3, 4, 3 ], "presentation_list": [ 3, 4, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 4, 3 ], "confidence_list": [ 3, 3, 3, 5, 4 ] }, { "paper_id": "EbOhZyxIzQ", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Maximilian Karl", "gender": "Male", "institution": "Machine Learning Research Lab, Volkswagen Group", "country": "DE", "position": "AI research scientist" }, { "name": "Patrick van der Smagt", "gender": "Male", "institution": "Machine Learning Research Lab, Volkswagen Group", "country": "DE", "position": "Full Professor" }, { "name": "Philip Becker-Ehmck", "gender": "Male", "institution": "Volkswagen Group", "country": "DE", "position": "Researcher" }, { "name": "Xingyuan Zhang", "gender": "Male", "institution": "Machine Learning Research Lab", "country": "DE", "position": "PhD student" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 53, "n_ref": 146, "n_ref_all": 180, "n_fig": 21, "n_tab": 3, "L_tab": 733, "n_element_tab": 84, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1044, "formula_len_all_1": 1028, "len_all": 226346, "len_all_1": 63353, "len_abs": 1252, "len_title": 158, "len_sents": 56540, "len_sents_1": 30959, "n_sents": 435, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1260, "title": "Overcoming Knowledge Barriers: Online Imitation Learning from Visual Observation with Pretrained World Models", "abs": "Pretraining and finetuning models has become increasingly popular in decision-making. But there are still serious impediments in Imitation Learning from Observation (ILfO) with pretrained models. This study identifies two primary obstacles: the Embodiment Knowledge Barrier (EKB) and the Demonstration Knowledge Barrier (DKB). The EKB emerges due to the pretrained models' limitations in handling novel observations, which leads to inaccurate action inference. Conversely, the DKB stems from the reliance on limited demonstration datasets, restricting the model's adaptability across diverse scenarios. \nWe propose separate solutions to overcome each barrier and apply them to Action Inference by Maximising Evidence (AIME), a state-of-the-art algorithm.\nThis new algorithm, AIME-NoB, integrates online interactions and a data-driven regulariser to mitigate the EKB. Additionally, it uses a surrogate reward function to broaden the policy's supported states, addressing the DKB. Our experiments on vision-based control tasks from the DeepMind Control Suite and MetaWorld benchmarks show that AIME-NoB significantly improves sample efficiency and converged performance, presenting a robust framework for overcoming the challenges in ILfO with pretrained models.", "keywords": [ "World Models", "Foundation Models", "Pretraining", "Imitation Learning from Observation", "Decision-making" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "EbG3PV7RaN", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Jiajun Liu", "gender": "Male", "institution": "The University of Queensland", "country": "AU", "position": "Associate Professor" }, { "name": "Sen Wang", "gender": "Male", "institution": "The University of Queensland", "country": "AU", "position": "Associate Professor" }, { "name": "Xuwei Xu", "gender": "Male", "institution": "University of Queensland", "country": "AU", "position": "PhD student" }, { "name": "Yang Li", "gender": "Male", "institution": "CSIRO", "country": "AU", "position": "Postdoc" }, { "name": "Yudong Chen", "gender": "Male", "institution": "University of Queensland", "country": "AU", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 60, "n_ref": 118, "n_ref_all": 142, "n_fig": 5, "n_tab": 7, "L_tab": 4580, "n_element_tab": 408, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 2639, "n_element_tab_1": 205, "formula_len_all": 935, "formula_len_all_1": 747, "len_all": 162875, "len_all_1": 65056, "len_abs": 1888, "len_title": 145, "len_sents": 37273, "len_sents_1": 27390, "n_sents": 250, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1826, "title": "RePaFormer: Ferocious and Scalable Acceleration of MetaFormers via Structural Reparamterization", "abs": "We reveal that feed-forward network (FFN) layers significantly contribute to the latencies of Vision Transformers (ViTs). This effect scales up quickly as the model size escalates, and hence presents a major opportunity in efficiency optimization for ViTs via structural reparameterization on FFN layers. However, directly reparameterizing the linear projection weights is difficult due to the non-linear activation in between. In this work, we propose an innovative channel idle mechanism that establishes a linear pathway through the activation function, facilitating structural reparameterization on FFN layers during inference. Consequently, we present a family of efficient ViTs embedded with the introduced mechanism called **RePa**rameterizable Vision Trans**Formers** (RePaFormers). This technique brings remarkable latency reductions with small sacrifices (sometimes gains) in accuracy across various MetaFormer-structured architectures investigated in the experiments. The benefits of this method scale consistently with model sizes, demonstrating increasing efficiency improvements and narrowing performance gaps as model sizes grow. Specifically, the RePaFormer variants for DeiT-Base and Swin-Base achieve 67.5% and 49.7% throughput accelerations with minor changes in top-1 accuracy (-0.4% and -0.9%), respectively. Further improvements in speed and accuracy are expected on even larger ViT models. In particular, the RePaFormer variants for ViT-Large and ViT-Huge enjoy 66.8% and 68.7% inference speed-ups with +1.7% and +1.1% higher top-1 accuracies, respectively. RePaFormer is the first to employ structural reparameterization on FFN layers to expedite ViTs to our best knowledge, and we believe that it represents an auspicious direction for efficient ViTs. Codes are provided in the supplementary material.", "keywords": [ "Efficient ViT", "Structural Reparameterization", "FFN Acceleration" ], "rating_list": [ 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "EbCUbPZjM1", "primary_area": "applications to robotics, autonomy, planning", "n_author": 8, "author_info_list": [ { "name": "Andrew Silva", "gender": "Male", "institution": "Toyota Research Institute", "country": "US", "position": "Researcher" }, { "name": "Daniela Rus", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" }, { "name": "Erfan Aasi", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Guy Rosman", "gender": "Male", "institution": "Massachusetts General Hospital, Harvard University", "country": "", "position": "Researcher" }, { "name": "Phat Tan Nguyen", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Sertac Karaman", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Tsun-Hsuan Wang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Zhang-Wei Hong", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 45, "n_ref": 102, "n_ref_all": 124, "n_fig": 6, "n_tab": 6, "L_tab": 1503, "n_element_tab": 114, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1021, "n_element_tab_1": 117, "formula_len_all": 461, "formula_len_all_1": 365, "len_all": 178323, "len_all_1": 69929, "len_abs": 1974, "len_title": 53, "len_sents": 54403, "len_sents_1": 35039, "n_sents": 361, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1410, "title": "ReGen: Generative Robot Simulation via Inverse Design", "abs": "Simulation plays a key role in scaling robot learning and validating policies, but constructing simulations remains labor-intensive. In this paper, we introduce ReGen, a generative simulation framework that automates this process using inverse design. Given an agent's behavior (such as a motion trajectory or objective function) and its textual description, we infer the underlying scenarios and environments that could have caused the behavior.\nOur approach leverages large language models to construct and expand a graph that captures cause-and-effect relationships and relevant entities with properties in the environment, which is then processed to configure a robot simulation environment. Our approach supports (i) augmenting simulations based on ego-agent behaviors, (ii) controllable, counterfactual scenario generation, (iii) reasoning about agent cognition and mental states, and (iv) reasoning with distinct sensing modalities, such as braking due to faulty GPS signals. \nWe demonstrate our method in autonomous driving and robot manipulation tasks, generating more diverse, complex simulated environments compared to existing simulations with high success rates, and enabling controllable generation for corner cases. This approach enhances the validation of robot policies and supports data or simulation augmentation, advancing scalable robot learning for improved generalization and robustness.", "keywords": [ "generative simulation", "robot", "autonomous driving", "large language model", "inverse design" ], "rating_list": [ 6, 3, 6, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "Eaw1ZrsNUN", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Manish Gupta", "gender": "Male", "institution": "Microsoft", "country": "IN", "position": "Principal Researcher" }, { "name": "SUBBA REDDY OOTA", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "PhD student" }, { "name": "Venkata Charan Chinni", "gender": "Male", "institution": "International Institute of Information Technology Hyderabad", "country": "IN", "position": "Undergrad student" }, { "name": "mounika marreddy", "gender": "Female", "institution": "International Institute of Information Technology Hyderabad, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "Researcher" }, { "name": "Lucie Flekova", "gender": "Female", "institution": "Rheinische Friedrich-Wilhelms Universit\u00e4t Bonn", "country": "DE", "position": "Full Professor" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 31, "n_ref": 78, "n_ref_all": 125, "n_fig": 18, "n_tab": 11, "L_tab": 5700, "n_element_tab": 517, "n_fig_1": 15, "n_tab_1": 1, "L_tab_1": 1330, "n_element_tab_1": 43, "formula_len_all": 214, "formula_len_all_1": 19, "len_all": 270144, "len_all_1": 77722, "len_abs": 2163, "len_title": 114, "len_sents": 103173, "len_sents_1": 40005, "n_sents": 700, "n_sents_1": 276, "venue": "ICLR2025", "decision": "Reject", "L_title": 122, "L_abs": 2229, "title": "USDC: A Dataset of $\\underline{U}$ser $\\underline{S}$tance and $\\underline{D}$ogmatism in Long $\\underline{C}$onversations", "abs": "Although prior studies have explored Stance and Dogmatism in user conversations, their datasets are constructed at the post level, treating each post as independent and randomly sampling posts from conversation threads. Thus, Stance and Dogmatism labels in these datasets cannot capture the user's opinion fluctuations expressed throughout the entire conversation context. However, identifying user's opinion fluctuations in long conversation threads on various topics can be extremely critical for enhanced personalization, market research, political campaigns, customer service, conflict resolution, targeted advertising, and content moderation. Hence, training language models to automate this task is critical. However, to train such models, gathering manual annotations has multiple challenges: 1) It is time-consuming and costly; 2) Conversation threads could be very long, increasing chances of noisy annotations; and 3) Interpreting instances where a user changes their opinion within a conversation is difficult because often such transitions are subtle and not expressed explicitly. Inspired by the recent success of large language models (LLMs) for complex natural language processing tasks, we leverage Mistral Large and GPT-4 to automate the human annotation process on the following two tasks while also providing reasoning: i) User Stance classification, which involves labeling a user's stance of a post in a conversation on a five-point scale; ii) User Dogmatism classification, which deals with labeling a user's overall opinion in the conversation on a four-point scale. The majority voting on zero-shot, one-shot, and few-shot annotations from these two LLMs on 764 multi-user Reddit conversations helps us curate the USDC dataset. USDC is then used to finetune and instruction-tune multiple deployable small language models for the 5-class stance and 4-class dogmatism classification tasks. Additionally, human annotations on 200 test conversations achieved inter-annotator agreement scores of 0.49 for stance and 0.50 for dogmatism, indicating a reasonable level of consistency between human and LLM annotations. We make the code and dataset publicly available [https://anonymous.4open.science/r/USDC-0F7F].", "keywords": [ "large language models", "annotators", "user opinions", "stance", "dogmatism", "human-llm alignment", "open-source llms", "closed-source llms" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "EaiU4F5pwn", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Ruoyan Li", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Undergrad student" }, { "name": "Wei Wang", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "Yizhou Sun", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "Zijie Huang", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 31, "n_formula_1": 29, "n_ref_uni": 36, "n_ref": 63, "n_ref_all": 79, "n_fig": 7, "n_tab": 10, "L_tab": 3538, "n_element_tab": 164, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2487, "n_element_tab_1": 193, "formula_len_all": 2532, "formula_len_all_1": 2105, "len_all": 183281, "len_all_1": 74725, "len_abs": 1423, "len_title": 122, "len_sents": 56572, "len_sents_1": 30418, "n_sents": 449, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1405, "title": "Physics-Informed Self-Guided Diffusion Model for High-Fidelity Simulations", "abs": "Machine learning (ML) models are increasingly explored in fluid dynamics as a promising way to generate high-fidelity computational fluid dynamics data more efficiently. A common strategy is to use low-fidelity data as computational-efficient inputs, and employ ML techniques to reconstruct high-fidelity flow fields. However, existing work typically assumes that low-fidelity data is artificially downsampled from high-fidelity sources, which limits model performance. In real-world applications, low-fidelity data is generated directly by numerical solvers with a lower initial state resolution, resulting in large deviations from high-fidelity data. To address this gap, we propose PG-Diff, a novel diffusion model for reconstructing high-fidelity flow fields, where both low- and high-fidelity data are generated from numerical solvers. Our experiments reveal that state-of-the-art models struggle to recover fine-grained high-fidelity details when using solver-generated low-fidelity inputs, due to distribution shift. To overcome this challenge, we introduce an \\textit{Importance Weight} strategy during training as self-guidance and a training-free \\textit{Residual Correction} method during inference as physical inductive bias, guiding the diffusion model toward higher-quality reconstructions. Experiments on four 2D turbulent flow datasets demonstrate the effectiveness of our proposed method.", "keywords": [ "Physics-informed Neural Networks", "Computational Fluid Dynamics" ], "rating_list": [ 5, 6, 5, 6, 3 ], "soundness_list": [ 3, 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 3, 1 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4, 3 ] }, { "paper_id": "EZXXJmuMd7", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Furu Wei", "gender": "Male", "institution": "Microsoft Research", "country": "CN", "position": "Partner Research Manager" }, { "name": "Lingjie Jiang", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Shaohan Huang", "gender": "Male", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Xun Wu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" } ], "n_formula": 14, "n_formula_1": 9, "n_ref_uni": 45, "n_ref": 105, "n_ref_all": 125, "n_fig": 6, "n_tab": 9, "L_tab": 7389, "n_element_tab": 325, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 916, "n_element_tab_1": 121, "formula_len_all": 1290, "formula_len_all_1": 339, "len_all": 179919, "len_all_1": 65250, "len_abs": 1150, "len_title": 91, "len_sents": 44892, "len_sents_1": 29646, "n_sents": 362, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 43, "L_abs": 1154, "title": "Textual Aesthetics in Large Language Models", "abs": "Image aesthetics is a crucial metric in the field of image generation. However, textual aesthetics has not been sufficiently explored. With the widespread application of large language models (LLMs), previous work has primarily focused on the correctness of content and the helpfulness of responses. Nonetheless, providing responses with textual aesthetics is also an important factor for LLMs, which can offer a cleaner layout and ensure greater consistency and coherence in content. In this work, we introduce a pipeline for aesthetics polishing and help construct a textual aesthetics dataset named TEXAES. We propose a textual aesthetics-powered fine-tuning method based on direct preference optimization, termed TAPO, which leverages textual aesthetics without compromising content correctness. Additionally, we develop two evaluation methods for textual aesthetics based on text and image analysis, respectively.Our experiments demonstrate that using textual aesthetics data and employing the TAPO fine-tuning method not only improves aesthetic scores but also enhances performance on general evaluation datasets such as AlpacalEval and Anera-hard.", "keywords": [ "Large Language Model", "Textual Aesthetics" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "EZExZ5d8ES", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Han Xie", "gender": "unknown", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Lecheng Kong", "gender": "Male", "institution": "Washington University, Saint Louis", "country": "US", "position": "PhD student" }, { "name": "Seongjun Yun", "gender": "Male", "institution": "Amazon", "country": "CA", "position": "Researcher" }, { "name": "Theodore Vasiloudis", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Xiang song", "gender": "Male", "institution": "Amazon", "country": "", "position": "Researcher" } ], "n_formula": 25, "n_formula_1": 17, "n_ref_uni": 39, "n_ref": 81, "n_ref_all": 104, "n_fig": 8, "n_tab": 6, "L_tab": 3074, "n_element_tab": 174, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2634, "n_element_tab_1": 103, "formula_len_all": 1646, "formula_len_all_1": 958, "len_all": 140758, "len_all_1": 68334, "len_abs": 1898, "len_title": 105, "len_sents": 42856, "len_sents_1": 31145, "n_sents": 325, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1909, "title": "Dynamic Mixture-of-Experts for Incremental Graph Learning", "abs": "Graph incremental learning is a learning paradigm that aims to adapt models trained on previous data to continuously incremented data or tasks over time without the need for retraining on the full dataset. However, regular graph machine learning methods suffer from catastrophic forgetting when applied to incremental learning settings, where previously learned knowledge is overridden by new knowledge. Previous approaches have tried to address this by treating the previously trained model as an inseparable unit and using regularization, experience replay, and parameter isolation to maintain old behaviors while learning new knowledge.\nThese approaches, however, do not account for the fact that not all previously acquired knowledge is equally beneficial for learning new tasks, and maintaining all previous knowledge and the latest knowledge in a single model is ineffective. Some prior patterns can be transferred to help learn new data, while others may deviate from the new data distribution and be detrimental. To address this, we propose a dynamic mixture-of-experts (DyMoE) approach for incremental learning. Specifically, a DyMoE GNN layer adds new expert networks specialized in modeling the incoming data blocks. We design a customized regularization loss that utilizes data sequence information so existing experts can maintain their ability to solve old tasks while helping the new expert learn the new data effectively. As the number of data blocks grows over time, the computational cost of the full mixture-of-experts (MoE) model increases. To address this, we introduce a sparse MoE approach, where only the top-$k$ most relevant experts make predictions, significantly reducing the computation time. Our model achieved 5.47\\% relative accuracy increase compared to the best baselines on class incremental learning with minimal computation increase, showing the model's exceptional power.", "keywords": [ "Graph neural networks; Incremental Learning" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 3, 2, 2, 4 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 1, 2, 3, 4 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "EXsiGFkwV6", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Ari Shapiro", "gender": "unknown", "institution": "FlawlessAI", "country": "US", "position": "Principal Researcher" }, { "name": "Hyeongwoo Kim", "gender": "unknown", "institution": "Imperial College London", "country": "GB", "position": "Fellow" }, { "name": "Kyle Olszewski", "gender": "Male", "institution": "Flawless AI", "country": "US", "position": "Researcher" }, { "name": "Pablo Garrido", "gender": "Male", "institution": "Flawless AI", "country": "US", "position": "Principal Researcher" }, { "name": "Pengfei Zhang", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Pinxin Liu", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 24, "n_ref": 40, "n_ref_all": 49, "n_fig": 3, "n_tab": 4, "L_tab": 3275, "n_element_tab": 61, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 541, "n_element_tab_1": 44, "formula_len_all": 711, "formula_len_all_1": 754, "len_all": 128620, "len_all_1": 50774, "len_abs": 307, "len_title": 146, "len_sents": 29390, "len_sents_1": 23679, "n_sents": 236, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 1152, "title": "Realistic-Gesture: Co-Speech Gesture Video Generation through Semantic-aware Gesture Representation", "abs": "Co-speech gesture generation is crucial for creating lifelike avatars and enhancing human-computer interactions by synchronizing gestures with speech in computer vision. Despite recent advancements, existing methods often struggle with accurately aligning gesture motions with speech signals and achieving pixel-level realism. To address these challenges, we introduce Realistic-Gesture, a groundbreaking framework that transforms co-speech gesture video generation through three innovative components: (1) a speech-aware gesture tokenization that incorporate speech context into motion pattern representation, (2) a mask gesture generator that learns to map audio signals to gestures by predicting masked motion tokens, enabling bidirectional contextually relevant gesture synthesis and editing, and (3) a structure-aware refinement module that employs differentiable edge connection to link gesture keypoints to improve video generation. Our extensive experiments demonstrate that Realistic-Gesture not only produces highly realistic and speech-aligned gesture videos but also supports long-sequence generation and video gesture editing applications.", "keywords": [ "gesture generation; motion representation; video generation" ], "rating_list": [ 8, 6, 6, 5, 3 ], "soundness_list": [ 3, 3, 4, 2, 2 ], "presentation_list": [ 3, 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 5, 4, 4, 5, 4 ] }, { "paper_id": "EXnDAXyVxw", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Hieu Le", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Postdoc" }, { "name": "Mathieu Salzmann", "gender": "Male", "institution": "Swiss Data Science Center", "country": "CH", "position": "Principal Researcher" }, { "name": "Saqib Javed", "gender": "Not Specified", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 78, "n_ref": 167, "n_ref_all": 191, "n_fig": 6, "n_tab": 7, "L_tab": 1624, "n_element_tab": 202, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1494, "n_element_tab_1": 181, "formula_len_all": 329, "formula_len_all_1": 330, "len_all": 183532, "len_all_1": 71472, "len_abs": 1807, "len_title": 110, "len_sents": 43105, "len_sents_1": 32067, "n_sents": 321, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1459, "title": "QT-DoG: Quantization-Aware Training for Domain Generalization", "abs": "Domain Generalization (DG) aims to train models that perform well not only on the training (source) domains but also on novel, unseen target data distributions. A key challenge in DG is preventing overfitting to source domains, which can be mitigated by finding flatter minima in the loss landscape. In this work, we propose Quantization-aware Training for Domain Generalization (QT-DoG) and demonstrate that weight quantization effectively leads to flatter minima in the loss landscape, thereby enhancing domain generalization. Unlike traditional quantization methods focused on model compression, QT-DoG exploits quantization as an implicit regularizer by inducing noise in model weights, guiding the optimization process toward flatter minima that are less sensitive to perturbations and overfitting. We provide both an analytical perspective and empirical evidence demonstrating that quantization inherently encourages flatter minima, leading to better generalization across domains. Moreover, with the benefit of reducing the model size through quantization, we demonstrate that an ensemble of multiple quantized models further yields superior accuracy than the state-of-the-art DG approaches with no computational or memory overheads. Our extensive experiments demonstrate that QT-DoG generalizes across various datasets, architectures, and quantization algorithms, and can be combined with other DG methods, establishing its versatility and robustness.", "keywords": [ "Domain Generalization", "Quantization", "Ensemble", "Network Compression", "Flat Minima", "Regularization" ], "rating_list": [ 3, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "EXaKfdsw04", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 5, "author_info_list": [ { "name": "Bogdan Grechuk", "gender": "Male", "institution": "University of Leicester", "country": "GB", "position": "Assistant Professor" }, { "name": "Ivan Y Tyukin", "gender": "unknown", "institution": "King's College London, University of London", "country": "GB", "position": "Full Professor" }, { "name": "Oliver Sutton", "gender": "unknown", "institution": "King's College London, University of London", "country": "GB", "position": "Postdoc" }, { "name": "Qinghua Zhou", "gender": "Male", "institution": "King's College London, University of London", "country": "GB", "position": "Postdoc" }, { "name": "Xiaolin Hu", "gender": "Male", "institution": "University of Leicester", "country": "GB", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 32, "n_ref": 39, "n_ref_all": 50, "n_fig": 4, "n_tab": 4, "L_tab": 443, "n_element_tab": 14, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 284, "n_element_tab_1": 4, "formula_len_all": 80, "formula_len_all_1": 27, "len_all": 106769, "len_all_1": 61490, "len_abs": 1191, "len_title": 124, "len_sents": 30804, "len_sents_1": 25613, "n_sents": 228, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1196, "title": "StepProof: Step-by-step verification of natural language mathematical proofs", "abs": "Interactive theorem provers (ITPs) are powerful tools for the formal verification of mathematical proofs down to the axiom level. However, their lack of a natural language interface remains a significant limitation. Recent advancements in large language models (LLMs) have enhanced the understanding of natural language inputs, paving the way for autoformalization\u2014the process of translating natural language proofs into formal proofs that can be verified. Despite these advancements, existing autoformalization approaches are limited to verifying complete proofs and lack the capability for finer, sentence-level verification. To address this gap, we propose StepProof, a novel autoformalization method designed for granular, step-by-step verification. StepProof breaks down complete proofs into multiple verifiable subproofs, enabling sentence-level verification. Experimental results demonstrate that StepProof significantly improves proof success rates and efficiency compared to traditional methods. Additionally, we found that minor manual adjustments to the natural language proofs, tailoring them for step-level verification, further enhanced StepProof\u2019s performance in autoformalization.", "keywords": [ "Mathematical NLP", "Autoformalization", "Logic Reasoning" ], "rating_list": [ 6, 1, 3, 3 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 2, 1, 1, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "EXXvBdFJ6I", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Han Yue", "gender": "Male", "institution": "Brandeis University", "country": "US", "position": "PhD student" }, { "name": "Hongfu Liu", "gender": "Male", "institution": "Brandeis University", "country": "US", "position": "Assistant Professor" }, { "name": "Jian Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Ziao Yang", "gender": "Male", "institution": "Brandeis University", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 75, "n_ref_all": 100, "n_fig": 7, "n_tab": 10, "L_tab": 6100, "n_element_tab": 934, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 4875, "n_element_tab_1": 611, "formula_len_all": 612, "formula_len_all_1": 547, "len_all": 141224, "len_all_1": 73494, "len_abs": 1289, "len_title": 85, "len_sents": 41937, "len_sents_1": 31301, "n_sents": 276, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 1292, "title": "On the Inflation of KNN-Shapley Value", "abs": "Shapley value-based data valuation methods, originating from cooperative game theory, quantify the usefulness of each individual sample by considering its contribution to all possible training subsets. Despite their extensive applications, we observe these methods encounter value inflation\u2014while samples with negative Shapley values are detrimental, some with positive values can also be harmful. This challenge prompts two fundamental questions: the suitability of zero as a threshold for distinguishing detrimental from beneficial samples and the determination of an appropriate threshold. To address these questions, we focus on KNN-Shapley and propose Calibrated KNN-Shapley (CKNN-Shapley), a semi-value method that calibrates zero as the threshold to distinguish detrimental samples from beneficial ones by mitigating the negative effects of small-sized training subsets. Through extensive experiments, we demonstrate the effectiveness of CKNN-Shapley in alleviating data valuation inflation, detecting detrimental samples, and assessing data quality. We also extend our approach beyond conventional classification settings, applying it to diverse and practical scenarios such as learning with mislabeled data, online learning with stream data, and active learning for label annotation.", "keywords": [ "Shapley Value", "Data Valuation", "KNN" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "EXGahWDp1E", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Deepjyoti Deka", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Scientisti" }, { "name": "Kaarthik Sundar", "gender": "Male", "institution": "Los Alamos National Laboratory", "country": "", "position": "Researcher" }, { "name": "Parikshit Pareek", "gender": "Male", "institution": "Indian Institute of Technology Roorkee", "country": "IN", "position": "Assistant Professor" }, { "name": "Sidhant Misra", "gender": "Male", "institution": "Los Alamos National Laboratory", "country": "", "position": "Scientist" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 21, "n_ref": 43, "n_ref_all": 73, "n_fig": 6, "n_tab": 1, "L_tab": 57, "n_element_tab": 4, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 313, "n_element_tab_1": 31, "formula_len_all": 976, "formula_len_all_1": 825, "len_all": 99216, "len_all_1": 59948, "len_abs": 1714, "len_title": 165, "len_sents": 36033, "len_sents_1": 28619, "n_sents": 239, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 118, "L_abs": 1720, "title": "Optimization Proxies using Limited Labeled Data and Training Time - A Semi-Supervised Bayesian Neural Network Approach", "abs": "Constrained optimization problems arise in various engineering system operations such as inventory management and electric power grids. However, the requirement to repeatedly solve such optimization problems with uncertain parameters poses a significant computational challenge. This work introduces a learning scheme using Bayesian Neural Networks (BNNs) to solve constrained optimization problems under limited labeled data and restricted model training times. We propose a semi-supervised BNN for this practical but complex regime, wherein training commences in a sandwiched fashion, alternating between a supervised learning step (using labeled data) for minimizing cost, and an unsupervised learning step (using unlabeled data) for enforcing constraint feasibility. Both supervised and unsupervised steps use a Bayesian approach, where Stochastic Variational Inference is employed for approximate Bayesian inference. We show that the proposed semi-supervised learning method outperforms conventional BNN and deep neural network (DNN) architectures on important non-convex constrained optimization problems from energy network operations, achieving up to a tenfold reduction in expected maximum equality gap and halving the optimality and inequality (feasibility) gaps, without requiring any correction or projection step. By leveraging the BNN's ability to provide posterior samples at minimal computational cost, we demonstrate that a Selection via Posterior (SvP) scheme can further reduce equality gaps by more than 10%. We also provide tight and practically meaningful probabilistic confidence bounds that can be constructed using a low number of labeled testing data and readily adapted to other applications.", "keywords": [ "Optimization Proxy", "Semi-supervised Bayesian Neural Networks", "Constrained Optimization" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 1, 2, 3 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "EWiWMoynco", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 2, "author_info_list": [ { "name": "Jonathan S. Lew", "gender": "unknown", "institution": "University of British Columbia", "country": "", "position": "PhD student" }, { "name": "Tor Aamodt", "gender": "Male", "institution": "University of British Columbia", "country": "", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 28, "n_ref": 44, "n_ref_all": 64, "n_fig": 8, "n_tab": 1, "L_tab": 2339, "n_element_tab": 201, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 507, "formula_len_all_1": 471, "len_all": 109005, "len_all_1": 51410, "len_abs": 3179, "len_title": 84, "len_sents": 28278, "len_sents_1": 23917, "n_sents": 189, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Reject", "L_title": 36, "L_abs": 795, "title": "NAQ: Nonlinearity-Aware Quantization", "abs": "Transformer-based large language models and vision transformers have achieved remarkable performance, but at a high energy cost.\nNonlinearities (e.g., GELU, softmax) have regions where the magnitude of the gradient is small, which means that errors in pre-nonlinearity inputs result in small output error.\nWe propose Nonlinearity-Aware Quantization (NAQ), which involves computing the FC layer outputs and attention scores at low precision, predicting the magnitude of the gradient of the nonlinearity, and recomputing the pre-nonlinearity if the gradient magnitude is large.\nWith future hardware support, models with NAQ would avoid up to 62% of full precision pre-nonlinearity computation and it would achieve up to 29% reduction in energy consumption, with small effects on model performance.", "keywords": [ "transformer", "energy", "quantization", "activation function", "gelu", "softmax", "llm", "vision transformer" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "EWcOEZa6Ee", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Jaeyeon Jo", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Jinkyoo Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 176, "n_formula_1": 30, "n_ref_uni": 20, "n_ref": 86, "n_ref_all": 152, "n_fig": 2, "n_tab": 6, "L_tab": 3099, "n_element_tab": 168, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2411, "n_element_tab_1": 120, "formula_len_all": 15922, "formula_len_all_1": 1865, "len_all": 186032, "len_all_1": 60647, "len_abs": 1009, "len_title": 92, "len_sents": 56892, "len_sents_1": 24348, "n_sents": 493, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 1014, "title": "Nash-GBML: Nash Gradient-Based Meta-Learning", "abs": "Meta-learning has been proposed to address fast adaptation to unseen tasks with little data. Traditional meta-learning is modeled as the Single-Leader Multi-Follower game consisting of inner and outer-level problems to minimize average or worst-case task loss. Because they assume all sampled tasks are independent, it reduces the flexibility of modeling complex interaction among tasks. Thus, we formulate meta-learning as a Single-Leader Multi-Follower game by considering the interaction among tasks at the inner level. We propose the Nash-GBML incorporating a penalty term into the task loss function to model the interaction among task-specific parameters. We discuss the iteration complexity and convergence of the Nash-GBML algorithm. To validate our Nash-GBML algorithm, we introduce two penalty terms, which are designed to reduce the average and worst-case task loss. We empirically show that the Nash-GBML with the proposed penalty terms outperforms traditional GBML for supervised learning experiments.", "keywords": [ "model-agnostic meta-learning", "gradient-based meta-learning", "game theory", "nash game", "single-leader multi-follower game" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 1 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "EWQaqDgXgr", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Eugene Ndiaye", "gender": "unknown", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "James Thornton", "gender": "unknown", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Louis B\u00e9thune", "gender": "Male", "institution": "Apple ", "country": "FR", "position": "Researcher" }, { "name": "Michael Kirchhof", "gender": "Male", "institution": "Apple", "country": "FR", "position": "Researcher" }, { "name": "Pierre Ablin", "gender": "Male", "institution": "Apple", "country": "", "position": "Researcher" }, { "name": "marco cuturi", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" } ], "n_formula": 32, "n_formula_1": 20, "n_ref_uni": 38, "n_ref": 80, "n_ref_all": 127, "n_fig": 19, "n_tab": 7, "L_tab": 7495, "n_element_tab": 383, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 2041, "n_element_tab_1": 3, "formula_len_all": 3396, "formula_len_all_1": 1659, "len_all": 176053, "len_all_1": 66839, "len_abs": 0, "len_title": 47, "len_sents": 48542, "len_sents_1": 30886, "n_sents": 389, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1596, "title": "Sparse Repellency for Shielded Generation in Text-to-Image Diffusion Models", "abs": "The increased adoption of diffusion models in text-to-image generation has triggered concerns on their reliability. Such models are now closely scrutinized under the lens of various metrics, notably calibration, fairness, or compute efficiency. We focus in this work on two issues that arise when deploying these models: a lack of diversity when prompting images, and a tendency to recreate images from the training set. To solve both problems, we propose a method that coaxes the sampled trajectories of pretrained diffusion models to land on images that fall outside of a reference set. We achieve this by adding a simple repellency term to the diffusion SDE throughout the generation trajectory, that is triggered whenever it is expected to land too closely to an image in the shielded reference set. Our method is sparse in the sense that these repellency terms are mostly zero and inactive, even more so towards the end of the generation trajectory. Our method, named SPELL\nfor sparse repellency, can be used either with a static reference set that contains protected images, or dynamically, by updating the reference set at each timestep with the expected images concurrently generated within a batch. We show that adding SPELL to popular diffusion models improves their diversity while impacting their FID only marginally, and performs comparatively better than other recent training-free diversity methods. Moreover, we demonstrate how SPELL can ensure a shielded generation away from a very large set of protected images by considering all 1.2M images from ImageNet as the protected set.", "keywords": [ "Diffusion Model", "Guidance", "Repellency", "Diversity" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 1, 4 ] }, { "paper_id": "EWP9BVRRbA", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Fengbin ZHU", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Jingkun Tang", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "PhD student" }, { "name": "Wenqiang Lei", "gender": "Male", "institution": "Sichuan University", "country": "CN", "position": "Full Professor" }, { "name": "Youcheng Huang", "gender": "unknown", "institution": "Sichuan University", "country": "CN", "position": "PhD student" }, { "name": "Pan Zhou", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" }, { "name": "Tat-seng Chua", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 30, "n_ref": 101, "n_ref_all": 116, "n_fig": 6, "n_tab": 4, "L_tab": 1829, "n_element_tab": 234, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 684, "n_element_tab_1": 60, "formula_len_all": 756, "formula_len_all_1": 721, "len_all": 182070, "len_all_1": 61006, "len_abs": 1011, "len_title": 141, "len_sents": 37198, "len_sents_1": 28200, "n_sents": 302, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 92, "L_abs": 1058, "title": "Effective and Efficient Adversarial Detection for Vision-Language Models via A Single Vector", "abs": "Visual Language Models (VLMs) are vulnerable to adversarial attacks, especially those from adversarial images, which is however under-explored in literature.\nTo facilitate research on this critical safety problem, we first construct a new la**R**ge-scale **A**dervsarial images dataset with **D**iverse h**A**rmful**R**esponses (RADAR), given that existing datasets are either small-scale or only contain limited types of harmful responses.\nWith the new RADAR dataset, we further develop a novel and effective i**N**-time**E**mbedding-based**A**dve**RS**arial **I**mage **DE**tection (NEARSIDE) method, which exploits a single vector that distilled from the hidden states of VLMs, which we call *the attacking direction*, to achieve the detection of adversarial images against benign ones in the input. \nExtensive experiments with two victim VLMs, LLaVA and MiniGPT-4, well demonstrate the effectiveness, efficiency,\nand cross-model transferrability of our proposed method. Our code is included in the supplementary file and will be made publicly available.", "keywords": [ "Visual Language Models", "Adversarial Attacks", "Attacking Directions", "Adversarial Defense", "Detection of Adversarial Samples" ], "rating_list": [ 6, 3, 6, 1 ], "soundness_list": [ 3, 3, 2, 1 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 3, 3, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "EWNH3QTSxd", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Guanquan Wang", "gender": "Male", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Takashi Onishi", "gender": "unknown", "institution": "AIST, National Institute of Advanced Industrial Science and Technology", "country": "", "position": "Researcher" }, { "name": "Takuya Hiraoka", "gender": "Male", "institution": "AIST, National Institute of Advanced Industrial Science and Technology", "country": "", "position": "Researcher" }, { "name": "Yoshimasa Tsuruoka", "gender": "Male", "institution": "The University of Tokyo", "country": "", "position": "Full Professor" } ], "n_formula": 41, "n_formula_1": 17, "n_ref_uni": 50, "n_ref": 64, "n_ref_all": 108, "n_fig": 21, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4748, "formula_len_all_1": 1187, "len_all": 184018, "len_all_1": 60412, "len_abs": 985, "len_title": 100, "len_sents": 52595, "len_sents_1": 26628, "n_sents": 507, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 100, "L_abs": 998, "title": "Which Experiences Are Influential for RL Agents? Efficiently Estimating The Influence of Experiences", "abs": "In reinforcement learning (RL) with experience replay, experiences stored in a replay buffer influence the RL agent's performance. \nInformation about how these experiences influence the agent's performance is valuable for various purposes, such as identifying experiences that negatively influence underperforming agents. \nOne method for estimating the influence of experiences is the leave-one-out (LOO) method. \nHowever, this method is usually computationally prohibitive. \nIn this paper, we present Policy Iteration with Turn-over Dropout (PIToD), which efficiently estimates the influence of experiences. \nWe evaluate how accurately PIToD estimates the influence of experiences and its efficiency compared to LOO. \nWe then apply PIToD to amend underperforming RL agents, i.e., we use PIToD to estimate negatively influential experiences for the RL agents and to delete the influence of these experiences. \nWe show that RL agents' performance is significantly improved via amendments with PIToD.", "keywords": [ "reinforcement learning", "data influence estimation" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "EWKPEtwjTy", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Tianqi Wei", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "JunDong Zhang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "MS student" } ], "n_formula": 39, "n_formula_1": 25, "n_ref_uni": 21, "n_ref": 36, "n_ref_all": 49, "n_fig": 8, "n_tab": 3, "L_tab": 262, "n_element_tab": 34, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 104, "n_element_tab_1": 6, "formula_len_all": 3072, "formula_len_all_1": 2039, "len_all": 103063, "len_all_1": 59663, "len_abs": 1362, "len_title": 122, "len_sents": 32628, "len_sents_1": 25910, "n_sents": 276, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 904, "title": "A Discrete Actor and Critic for Reinforcement Learning on Continuous Tasks", "abs": "Solving continuous reinforcement learning (RL) tasks typically requires models with continuous action spaces, as discrete models face challenges such as the curse of dimensionality. Inspired by discrete controlling signals in control systems, such as pulse-width modulation, we investigated RL models with discrete action spaces with performance comparable to continuous models on continuous tasks. In this paper, we propose an RL model with a discrete action space, designed a discrete actor that outputs action distributions and twin discrete critics for value distribution estimation. We also developed both the training method and exploration strategy for this model. The model successfully solved BipedalWalkerHardcore-v3, a continuous robot control task in a complex environment, achieved a higher score than the state-of-the-art baselines and comparable results across various other control tasks.", "keywords": [ "reinforcement Learning", "discrete action space", "continuous control", "bipedal locomotion" ], "rating_list": [ 3, 3, 1, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 1, 2, 1, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "EW6bNEqalF", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Ahana Deb", "gender": "Non-Binary", "institution": "Universitat Pompeu Fabra", "country": "ES", "position": "PhD student" }, { "name": "Alessandro Ronca", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Research Associate" }, { "name": "Anders Jonsson", "gender": "Male", "institution": "Universitat Pompeu Fabra", "country": "ES", "position": "Full Professor" }, { "name": "Mohammad Sadegh Talebi", "gender": "Male", "institution": "University of Copenhagen", "country": "", "position": "Assistant Professor" }, { "name": "Roberto Cipollone", "gender": "Male", "institution": "Leonardo SPA", "country": "IT", "position": "Researcher" } ], "n_formula": 60, "n_formula_1": 21, "n_ref_uni": 64, "n_ref": 107, "n_ref_all": 124, "n_fig": 7, "n_tab": 4, "L_tab": 777, "n_element_tab": 85, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7534, "formula_len_all_1": 2150, "len_all": 203084, "len_all_1": 83150, "len_abs": 1205, "len_title": 128, "len_sents": 59678, "len_sents_1": 37132, "n_sents": 500, "n_sents_1": 281, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1209, "title": "Offline RL in Regular Decision Processes: Sample Efficiency via Language Metrics", "abs": "This work studies offline Reinforcement Learning (RL) in a class of non-Markovian environments called Regular Decision Processes (RDPs). In RDPs, the unknown dependency of future observations and rewards from the past interactions can be captured by some hidden finite-state automaton. For this reason, many RDP algorithms first reconstruct this unknown dependency using automata learning techniques. In this paper, we consider episodic RDPs and show that it is possible to overcome the limitations of existing offline RL algorithms for RDPs via\nthe introduction of two original techniques: a novel metric grounded in formal language theory and an approach based on Count-Min-Sketch (CMS). Owing to the novel language metric, our algorithm is proven to be more sample efficient than existing results, and in some problem instances admitting low complexity languages, the gain is showcased to be exponential in the episode length. The CMS-based approach removes the need for na\u00efve counting and alleviates the memory requirements for long planning horizons. We derive Probably Approximately Correct (PAC) sample complexity bounds associated to each of these techniques, and validate the approach experimentally.", "keywords": [ "Reinforcement Learning", "Non-Markov Decision Process", "Offline Reinforcement Learning", "Regular Decision Processes", "Sample Complexity", "Automata" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "EW62GvCzP9", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Cameron Allen", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Micah Carroll", "gender": "Male", "institution": "University of California Berkeley", "country": "US", "position": "PhD student" }, { "name": "Tianyi Qiu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 18, "n_formula_1": 9, "n_ref_uni": 40, "n_ref": 66, "n_ref_all": 96, "n_fig": 11, "n_tab": 3, "L_tab": 4909, "n_element_tab": 159, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 177, "n_element_tab_1": 10, "formula_len_all": 2890, "formula_len_all_1": 883, "len_all": 202856, "len_all_1": 65068, "len_abs": 1707, "len_title": 120, "len_sents": 68971, "len_sents_1": 32639, "n_sents": 479, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1690, "title": "Truthfulness Without Supervision: Model Evaluation Using Peer Prediction", "abs": "Current evaluation methods for language models rely on supervision, but trusted supervision for difficult tasks is often unavailable, especially for superhuman models. In these cases, models have been demonstrated to exploit evaluation schemes built on such imperfect supervision, leading to deceptive evaluation results. \nHowever, underutilized in the context of model evaluation, a wealth of mechanism design research focuses on game-theoretic *incentive compatibility* - eliciting honest and informative answers without trusted supervision. \nDrawing from this literature, we introduce the peer prediction method for model evaluation. It tells apart honest and informative answers from deceptive and uninformative ones, using a metric based on mutual predictability and without requiring ground truth labels. \nWe demonstrate the method's effectiveness and resistance to deception, with both theoretical guarantees and comprehensive empirical validation on up to 405B-parameter models.\nIn contrast to LLM-as-a-Judge which requires strong and trusted judges, we discover an inverse scaling property in peer prediction, where, surprisingly, resistance to deception is *strengthened* as the capability gap between the jury and participants *widens*, enabling reliable evaluation of strong models without trusted supervision.\nIn particular, LLM-as-a-Judge evaluations become worse than random guesses when facing deceptive models 5-20$\\times$ its size, while peer prediction thrives when such gaps are large, including in cases with over 100$\\times$ size difference.\nLooking forward, we view this work as a step towards game-theoretic resistance to model deception in alignment and evaluation.", "keywords": [ "Language Model Evaluation", "AI Alignment", "AI Truthfulness and Deception", "Large Language Models" ], "rating_list": [ 3, 6, 5 ], "soundness_list": [ 1, 2, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "EVuANndPlX", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Costas Mavromatis", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "George Karypis", "gender": "Male", "institution": "University of Minnesota, Minneapolis", "country": "", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 3, "n_ref_uni": 56, "n_ref": 120, "n_ref_all": 168, "n_fig": 6, "n_tab": 6, "L_tab": 1565, "n_element_tab": 183, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 697, "n_element_tab_1": 55, "formula_len_all": 771, "formula_len_all_1": 131, "len_all": 208457, "len_all_1": 61459, "len_abs": 3567, "len_title": 114, "len_sents": 49655, "len_sents_1": 28988, "n_sents": 450, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1459, "title": "GNN-RAG: Graph Neural Retrieval for Large Language Model Reasoning", "abs": "Retrieval-augmented generation (RAG) in Knowledge Graph Question Answering (KGQA) enriches the context of Large Language Models (LLMs) with retrieved KG information based on the question. However, KGs contain complex graph information and existing KG retrieval methods are challenged when questions require multi-hop information. To improve RAG in complex KGQA, we introduce the GNN-RAG framework, which leverages Graph Neural Networks (GNNs) for effective graph reasoning and retrieval. GNN-RAG consists of a graph neural phase, where the GNN retriever learns to identify useful graph information for KGQA, e.g., when tackling complex questions. At inference time, the GNN scores answer candidates for the given question and the shortest paths in the KG that connect question entities and answer candidates are retrieved to represent KG reasoning paths. The paths are verbalized and given as context to the downstream LLM for ultimate KGQA; GNN-RAG can be seamlessly integrated with different LLMs for RAG. Experimental results show that GNN-RAG achieves state-of-the-art performance in two widely used KGQA benchmarks (WebQSP and CWQ), outperforming or matching GPT-4 performance with a 7B tuned LLM. In addition,GNN-RAG excels on multi-hop and multi-entity questions outperforming competing approaches by 8.9--15.5\\% points at answer F1. Furthermore, we show the effectiveness of GNN-RAG in retrieval augmentation, which further boosts KGQA performance.", "keywords": [ "Knowledge Graph", "Large Language Models", "Retrieval-Augmented Generation" ], "rating_list": [ 6, 6, 8, 3, 5 ], "soundness_list": [ 4, 3, 3, 2, 2 ], "presentation_list": [ 3, 4, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 3, 4, 5 ] }, { "paper_id": "EVg9lwHFJs", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Chao Chen", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "Postdoc" }, { "name": "Chenglong Ye", "gender": "Male", "institution": "Fuzhou University", "country": "CN", "position": "MS student" }, { "name": "Haizhou Sun", "gender": "Male", "institution": "SmartMore", "country": "", "position": "Researcher" }, { "name": "Wang Yufeng", "gender": "Male", "institution": "Fuzhou University", "country": "", "position": "MS student" }, { "name": "Wu Yunbing", "gender": "Male", "institution": "xi'an university of science and technology", "country": "", "position": "MS student" }, { "name": "Xiangwen Liao", "gender": "Male", "institution": "Fuzhou University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaofei Zhu", "gender": "Male", "institution": "Chongqing University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zhaochun Ren", "gender": "Male", "institution": "Leiden University", "country": "NL", "position": "Associate Professor" }, { "name": "Zhou Yang", "gender": "Male", "institution": "Fuzhou University", "country": "CN", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 17, "n_ref_uni": 55, "n_ref": 95, "n_ref_all": 111, "n_fig": 7, "n_tab": 4, "L_tab": 1141, "n_element_tab": 122, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 770, "n_element_tab_1": 85, "formula_len_all": 1752, "formula_len_all_1": 839, "len_all": 142402, "len_all_1": 60127, "len_abs": 2356, "len_title": 135, "len_sents": 35516, "len_sents_1": 25855, "n_sents": 289, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 2284, "title": "Fine-Grained Emotion Recognition with In-Context Learning: A Prototype Theory Approach", "abs": "In-context learning (ICL) achieves remarkable performance in various domains such as knowledge acquisition, commonsense reasoning, and semantic understanding. \nHowever, its effectiveness deteriorates significantly in emotion detection tasks, particularly in fine-grained emotion recognition. \nThe reasons behind this decline still remain unclear.\nIn this paper, we explore the underlying reasons of ICL's suboptimal performance through the lens of prototype theory.\nOur investigation reveals that ICL aligns with the principles of prototype theory when applied to fine-grained emotion recognition tasks. \nAccording to prototype theory, effective emotion recognition requires: Referencing well-represented emotional prototypes that are similar to the query emotions, and making predictions based on the closest emotional similarity.\nBuilding on this insight, ICL has three main shortcomings: \n(1) It uses oversimplified single-emotion labels for prototypes, leading to inaccurate emotion representation.\n(2) It references semantically similar but emotionally distant prototypes.\n(3) It considers all emotion categories as candidates, leading to interference from irrelevant emotions and inaccurate predictions.\n\nTo address these shortcomings, we propose an Emotion Context Learning method (E-ICL) for fine-grained emotion recognition. \nE-ICL first employs a dynamic soft-label strategy to create multi-dimensional emotional labels for accurate prototype representation. \nIt then selects emotionally similar prototypes as references for emotion prediction.\nFinally, it uses an emotion exclusion strategy to eliminate interference from dissimilar emotions by selecting similar emotions as candidates, resulting in more robust and accurate predictions.\nNote that our approach is implemented with the aid of a plug-and-play emotion auxiliary model, requiring no additional training. \nExtensive experiments conducted on fine-grained emotion datasets\u2014EDOS, Empathetic-Dialogues, EmpatheticIntent, and GoEmotions\u2014demonstrate that E-ICL significantly outperforms existing methods in emotion prediction performance. Moreover, even when the emotion auxiliary model accounts for less than 10\\% of the LLMs' capacity, E-ICL consistently boosts LLM performance by over 4\\% across multiple datasets.", "keywords": [ "fine-grained emotion recognition", "in-context learning", "ICL", "large language model", "LLMs" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 1, 2 ], "presentation_list": [ 2, 1, 1 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 5, 5, 4 ] }, { "paper_id": "EVa5OIYBoG", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Shafiq Joty", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Principal Researcher" }, { "name": "Yifei Ming", "gender": "Male", "institution": "Salesforce AI Research", "country": "US", "position": "Researcher" }, { "name": "Zixuan Ke", "gender": "Male", "institution": "Salesforce AI Research", "country": "", "position": "Researcher" }, { "name": "Phi Xuan Nguyen", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 47, "n_ref": 79, "n_ref_all": 108, "n_fig": 2, "n_tab": 12, "L_tab": 1962, "n_element_tab": 197, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 750, "n_element_tab_1": 59, "formula_len_all": 263, "formula_len_all_1": 212, "len_all": 158038, "len_all_1": 63892, "len_abs": 1730, "len_title": 125, "len_sents": 37727, "len_sents_1": 30531, "n_sents": 299, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1736, "title": "Expanding the Web, Smaller Is Better: A Comprehensive Study in Post-training", "abs": "General-purpose large language models (GLLMs) like GPT-4 and LLaMA have demonstrated exceptional performance across a wide range of tasks. However, their performance often falls short in domain- or task-specific applications, where deeper, specialized knowledge is essential, while maintaining general knowledge remains crucial for handling broader, unseen tasks. Post-training has been widely applied to\nmake LLMs specialized, typically consisting of multiple stages, including DomainAdaptive Pre-Training (DAPT) and Supervised Fine-Tuning (SFT). In this work, we conduct a comprehensive study on three key aspects of post-training taking Finance as a target domain: (1) the distinct roles of DAPT and SFT in post-training, (2) strategies to mitigate knowledge forgetting across stages, and (3) evaluation methods that capture both general and domain-specific capabilities. \nOur results show that DAPT and SFT require distinct training objectives, joint training of DAPT and SFT is essential for maintaining stage knowledge and encouraging knowledge transfer across stages, and replay mechanisms are critical for preventing forgetting. Evaluation should encompass general, seen, and unseen tasks for a complete assessment. Based on these insights, we developed a Joint-and-Replay post-training recipe and built LLaMA3-8B-Fin, a smaller yet more powerful stateof-the-art financial LLM trained through post-training. Despite its smaller size, LLaMA3-8B-Fin surpasses larger models like GPT-4o and LLaMA3.1-70b on both seen and unseen financial tasks while retaining general knowledge, demonstrating that a well-structured post-training can \u201cexpand the web\u201d of capabilities in smaller LLMs, enabling them to outperform much larger models.", "keywords": [ "Post-training", "Continual Learning", "Large Language Models" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "EVZnnhtMNX", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Mert Pilanci", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Assistant Professor" }, { "name": "Miria Feng", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "MS student" } ], "n_formula": 21, "n_formula_1": 19, "n_ref_uni": 40, "n_ref": 52, "n_ref_all": 60, "n_fig": 2, "n_tab": 1, "L_tab": 138, "n_element_tab": 20, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 114, "n_element_tab_1": 18, "formula_len_all": 1394, "formula_len_all_1": 1388, "len_all": 123100, "len_all_1": 65988, "len_abs": 1444, "len_title": 127, "len_sents": 36950, "len_sents_1": 31007, "n_sents": 303, "n_sents_1": 235, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1467, "title": "Scalable Preference Learning for Large Language Models via Convex Optimization", "abs": "Fine-tuning large language models (LLMs) for alignment with human preferences have become a key factor in the success of models like ChatGPT and Gemini, which are now integral to mainstream use. \nMany effective techniques are based on Reinforcement Learning from Human Feedback (RLHF), yet are challenging and expensive to implement. \nDirect Preference Optimization (DPO) offers an accessible alternative by simplifying the objective, but can exhibit random ranking accuracy and requires a frozen reference model.\nIn this paper, we develop a fast and an even more lightweight DPO based algorithm --- \\emph{CVX-DPO} ---- that operates on a single GPU. \nThe key to achieving this is leveraging the convex optimization reformulation of neural networks, which eliminates the dependence on copying the reference model and is robust against hyperparameter tuning.\nCVX-DPO can be trained to global optimality in polynomial time. We use the Alternating Direction Method of Multipliers (ADMM) to solve this optimization problem in order to increase parallelization efficiency, and implement our methods in JAX to lift the memory constraints across experiments. \nWe experiment on three datasets, including one synthetically generated educational dataset, to demonstrate the efficacy of our novel algorithm in a real world setting.\nCVX-DPO outperforms traditional DPO in user preference generation when tested on human subjects, despite being trained on one single RTX-4090 GPU.", "keywords": [ "large language models", "preference learning", "convex optimization" ], "rating_list": [ 3, 1, 3, 5, 3 ], "soundness_list": [ 2, 1, 2, 3, 1 ], "presentation_list": [ 1, 1, 1, 3, 1 ], "contribution_list": [ 2, 1, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 3, 4 ] }, { "paper_id": "EVK0sQHVCd", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Liu Yuezhang", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "PhD student" }, { "name": "Xue-Xin Wei", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 5, "n_ref_uni": 43, "n_ref": 80, "n_ref_all": 124, "n_fig": 24, "n_tab": 6, "L_tab": 1793, "n_element_tab": 145, "n_fig_1": 19, "n_tab_1": 2, "L_tab_1": 470, "n_element_tab_1": 62, "formula_len_all": 1014, "formula_len_all_1": 215, "len_all": 158656, "len_all_1": 67794, "len_abs": 1230, "len_title": 103, "len_sents": 49755, "len_sents_1": 31574, "n_sents": 419, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1241, "title": "How and how well do diffusion models improve adversarial robustness?", "abs": "Recent findings suggest that diffusion models significantly enhance empirical adversarial robustness. While some intuitive explanations have been proposed, the precise mechanisms underlying these improvements remain unclear. In this work, we systematically investigate how and how well do diffusion models improve adversarial robustness. First, we observe that diffusion models intriguingly increase\u2014rather than decrease\u2014the $\\ell_p$ distances to clean samples. This is the opposite of what was believed previously. Second, we find that the purified images are heavily influenced by the internal randomness of diffusion models. To properly evaluate the robustness of systems with inherent randomness, we introduce the concept of fuzzy adversarial robustness, and find that empirically a substantial fraction of adversarial examples are fuzzy in nature. Finally, by leveraging a hyperspherical cap model of adversarial regions, we show that diffusion models increase robustness by dramatically compressing the image space. Our findings provide novel insights into the mechanisms behind the robustness improvements of diffusion-model-based purification and offer guidance for the development of more efficient adversarial purification systems.", "keywords": [ "diffusion models", "adversarial purification", "robustness" ], "rating_list": [ 5, 5, 8, 5, 5 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 4, 3, 3 ], "contribution_list": [ 3, 2, 4, 2, 2 ], "confidence_list": [ 4, 4, 2, 4, 4 ] }, { "paper_id": "EV7FMBZxnx", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Huihui Yue", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Yin Xiangjun", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" } ], "n_formula": 17, "n_formula_1": 15, "n_ref_uni": 31, "n_ref": 44, "n_ref_all": 77, "n_fig": 8, "n_tab": 5, "L_tab": 1842, "n_element_tab": 163, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 518, "n_element_tab_1": 53, "formula_len_all": 997, "formula_len_all_1": 840, "len_all": 129888, "len_all_1": 60821, "len_abs": 938, "len_title": 71, "len_sents": 34834, "len_sents_1": 26189, "n_sents": 274, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 943, "title": "Reveal Object in Lensless Photography via Region Gaze and Amplification", "abs": "Detecting concealed objects, such as in vivo lesions or camouflage, requires customized imaging systems. Lensless cameras, being compact and flexible, offer a promising alternative to bulky lens systems. However, the absence of lenses leads to measurements lacking visual semantics, posing significant challenges for concealed object detection (COD). To tackle this issue, we propose a region gaze-amplification network (RGANet) for progressively exploiting concealed objects from lensless imaging measurements. Specifically, a region gaze module (RGM) is proposed to mine spatial-frequency cues informed by biological and psychological mechanisms, and a region amplifier (RA) is designed to amplify the details of object regions to enhance COD performance. Furthermore, we contribute the first relevant dataset as a benchmark to prosper the lensless imaging community. Extensive experiments demonstrate the exciting performance of our method.", "keywords": [ "Lensless Imaging; Computational Imaging; Region Gaze; Region Amplifier; Concealed Object Detection" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "EUeNr3e8AV", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Bo Jin", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Associate Professor" }, { "name": "Hanlin Dong", "gender": "unknown", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Jian Yang", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Postdoc" }, { "name": "Mingsong Chen", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Xian Wei", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Principal Researcher" }, { "name": "Xuan Tang", "gender": "Female", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Yingjie Liu", "gender": "unknown", "institution": "East China Normal University", "country": "CN", "position": "PhD student" }, { "name": "Zhiqiang Wu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" } ], "n_formula": 29, "n_formula_1": 19, "n_ref_uni": 33, "n_ref": 58, "n_ref_all": 75, "n_fig": 6, "n_tab": 15, "L_tab": 6270, "n_element_tab": 862, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 2082, "n_element_tab_1": 289, "formula_len_all": 3997, "formula_len_all_1": 2335, "len_all": 151709, "len_all_1": 67771, "len_abs": 1762, "len_title": 118, "len_sents": 42631, "len_sents_1": 27986, "n_sents": 350, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1464, "title": "R2Det: Exploring Relaxed Rotation Equivariance in 2D Object Detection", "abs": "Group Equivariant Convolution (GConv) empowers models to capture equivariant features and explore underlying symmetries in data, improving performance. However, real-world scenarios often deviate from ideal symmetric systems caused by physical permutation, characterized by non-trivial actions of a symmetry group, resulting in asymmetries that affect the outputs, a phenomenon known as symmetry breaking. Traditional GConv-based methods are constrained by rigid operational rules within group space, assuming data remains strictly equivariant under limited group transformations. This limitation makes it difficult to adapt to Symmetry-Breaking and non-rigid transformations. Motivated by this, we mainly focus on a common scenario: rotational Symmetry-Breaking. By relaxing strict group transformations within Strict Rotation-Equivariant group $\\mathbf{C}_n$, we redefine a Relaxed Rotation-Equivariant group $\\mathbf{R}_n$ and introduce a novel Relaxed Rotation-Equivariant GConv (R2GConv) with only a minimal increase of $4n$ parameters compared to GConv. Based on R2GConv, we propose a Relaxed Rotation-Equivariant Network (R2Net) as the backbone and develop a Relaxed Rotation-Equivariant Object Detector (R2Det) for 2D object detection. Experimental results demonstrate the effectiveness of our R2GConv in natural image classification, and R2Det achieves excellent performance in 2D object detection with improved generalization capabilities and robustness.", "keywords": [ "Relaxation", "Rotation", "Equivariance" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 4, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 3, 5 ] }, { "paper_id": "EUe0yA2pAw", "primary_area": "foundation or frontier models, including LLMs", "n_author": 3, "author_info_list": [ { "name": "Guoqiang Zhang", "gender": "Male", "institution": "University of Exeter", "country": "GB", "position": "Assistant Professor" }, { "name": "JP Lewis", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "W. Bastiaan Kleijn", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 36, "n_formula_1": 26, "n_ref_uni": 29, "n_ref": 49, "n_ref_all": 82, "n_fig": 5, "n_tab": 3, "L_tab": 286, "n_element_tab": 36, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 85, "n_element_tab_1": 10, "formula_len_all": 2093, "formula_len_all_1": 1839, "len_all": 110513, "len_all_1": 62290, "len_abs": 1957, "len_title": 121, "len_sents": 32249, "len_sents_1": 28408, "n_sents": 248, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1995, "title": "On Exact Bit-level Reversible Transformers Without Changing Architectures", "abs": "Various reversible deep neural networks (DNN) models have been proposed to reduce memory consumption in the training process. However, almost all existing reversible DNNs either require special non-standard architectures or are constructed by modifying existing DNN architectures considerably to enable reversibility. In this work we present the BDIA-transformer, which is an exact bit-level reversible transformer that uses an unchanged standard architecture for inference. The basic idea is to first treat each transformer block as the Euler integration approximation for solving an ordinary differential equation (ODE) and then incorporate the technique of bidirectional integration approximation (BDIA) (originally designed for diffusion inversion) into the neural architecture, together with activation quantization to make it exactly bit-level reversible. In the training process, we let a hyper-parameter $\\gamma$ in BDIA-transformer randomly take one of the two values $\\{0.5, -0.5\\}$ per training sample per transformer block for averaging every two consecutive integration approximations. As a result, BDIA-transformer can be viewed as training an ensemble of ODE solvers parameterized by a set of binary random variables, which regularizes the model and results in improved validation accuracy. Lightweight side information per transformer block is required to be stored in the forward process to account for binary quantization loss to enable exact bit-level reversibility. In the inference procedure, the expectation $\\mathbb{E}(\\gamma)=0$ is taken to make the resulting architectures of BDIA-transformer identical to transformers up to activation quantization. Our experiments in both image classification and language translation show that BDIA-transformers outperform their conventional counterparts significantly in terms of validation performance due to the regularization effect of the set of $\\gamma$ random variables while also requiring considerably less training memory.", "keywords": [ "transformer", "ViT", "BDIA", "reversibility", "ODE solvers" ], "rating_list": [ 6, 3, 5, 3, 6 ], "soundness_list": [ 4, 1, 2, 2, 3 ], "presentation_list": [ 3, 2, 4, 1, 2 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 2, 2, 3 ] }, { "paper_id": "EUSkm2sVJ6", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Jiayuan Ye", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Reza Shokri", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" }, { "name": "Sajjad Zarifzadeh", "gender": "Male", "institution": "Yazd University", "country": "", "position": "Associate Professor" }, { "name": "Yao Tong", "gender": "unknown", "institution": "School of Computing, National University of Singapore", "country": "", "position": "PhD student" } ], "n_formula": 75, "n_formula_1": 25, "n_ref_uni": 41, "n_ref": 120, "n_ref_all": 207, "n_fig": 10, "n_tab": 8, "L_tab": 1542, "n_element_tab": 181, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1119, "n_element_tab_1": 117, "formula_len_all": 5678, "formula_len_all_1": 1641, "len_all": 253272, "len_all_1": 72774, "len_abs": 770, "len_title": 101, "len_sents": 90133, "len_sents_1": 32973, "n_sents": 620, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 89, "L_abs": 806, "title": "How much of my dataset did you use? Quantitative Data Usage Inference in Machine Learning", "abs": "How much of a given dataset was used to train a machine learning model? This is a critical question for data owners assessing the risk of unauthorized data usage and protecting their right (United States Code, 1976). However, previous work mistakenly treats this as a binary problem\u2014inferring whether \\textit{all or none} or \\textit{any or none} of the data was used\u2014which is fragile when faced with real, non-binary data usage risks. To address this, we propose a fine-grained analysis called Dataset Usage Cardinality Inference (\\ourmethod{}), which estimates the exact proportion of data used. Our algorithm, leveraging debiased membership guesses, matches the performance of the optimal MLE approach (with a maximum error <0.1) but with significantly lower (e.g., $300 \\times$ less) computational cost.", "keywords": [ "Machine Learning", "Privacy", "Dataset Usage Inference", "Dataset Ownership", "Membership Inference Attack", "Dataset Copyright" ], "rating_list": [ 6, 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 2, 4, 3, 3, 4 ], "confidence_list": [ 2, 3, 2, 3, 4 ] }, { "paper_id": "EUBMPmcCWQ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Adrian Perez-Suay", "gender": "Male", "institution": "Universidad de Valencia", "country": "", "position": "Assistant Professor" }, { "name": "Elena M. De-Diego", "gender": "Female", "institution": "Universidad de Navarra", "country": "ES", "position": "Researcher" }, { "name": "Jean-Michel Loubes", "gender": "Male", "institution": "Universit\u00e9 Paul Sabatier (Toulouse III)", "country": "FR", "position": "Full Professor" }, { "name": "Paula Gordaliza", "gender": "Female", "institution": "Universidad P\u00fablica de Navarra", "country": "ES", "position": "Assistant Professor" } ], "n_formula": 23, "n_formula_1": 22, "n_ref_uni": 53, "n_ref": 85, "n_ref_all": 106, "n_fig": 7, "n_tab": 4, "L_tab": 4028, "n_element_tab": 228, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 1173, "n_element_tab_1": 46, "formula_len_all": 1792, "formula_len_all_1": 1568, "len_all": 178779, "len_all_1": 75626, "len_abs": 623, "len_title": 99, "len_sents": 50928, "len_sents_1": 34506, "n_sents": 346, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 625, "title": "PLS-based approach for Fair Representation Learning", "abs": "We revisit the problem of fair representation learning by proposing Fair Partial Least Squares (PLS) components. PLS is widely used in statistics to efficiently reduce the dimension of the data by providing representation tailored for the prediction. We propose a novel method to incorporate fairness constraints in the construction of PLS components. This new algorithm provides a feasible way to construct such features both in the linear and the non linear case using kernel embeddings. The efficiency of our method is evaluated on different datasets, and we prove its superiority with respect to standard fair PCA method.", "keywords": [ "Fair Representation Learning", "PLS", "Supervised Learning", "Dimension Reduction", "Fairness" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "EUAxxrxOM8", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Dheeraj Narasimha", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Postdoc" }, { "name": "Nicolas Gast", "gender": "Male", "institution": "INRIA", "country": "", "position": "Assistant Professor" } ], "n_formula": 89, "n_formula_1": 34, "n_ref_uni": 31, "n_ref": 108, "n_ref_all": 135, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7022, "formula_len_all_1": 2543, "len_all": 174106, "len_all_1": 69613, "len_abs": 943, "len_title": 111, "len_sents": 58450, "len_sents_1": 29383, "n_sents": 535, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 986, "title": "Model predictive control is almost optimal for restless bandits", "abs": "We consider the discrete time infinite horizon average reward restless markovian bandit (RMAB) problem. We propose a model predictive control based non-stationary policy with a rolling computational horizon $\\tau$. At each time-slot, this policy solves a $\\tau$ horizon linear program whose first control value is kept as a control for the RMAB. Our solution requires minimal assumptions and quantifies the loss in optimality in terms of $\\tau$ and the number of arms, $N$. We show that its sub-optimality gap is $O(1/\\sqrt{N})$ in general, and $\\exp(-\\Omega{N})$ under a local-stability condition. Our proof is based on a framework from dynamic control known as dissipativity. Not only is our solution easy to implement but performs very well in practice when compared to the state of the art. Further, both our solution and our proof methodology can easily be generalized to more general constrained MDP settings and should thus, be of great interest to the burgeoning RMAB community.", "keywords": [ "Restless Multi-Armed Bandits", "Markov Decision Processes", "Constrained Optimization", "Stochastic Control" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "ETokBVXrbC", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 7, "author_info_list": [ { "name": "Amit Lal", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Full Professor" }, { "name": "Juneho Hwang", "gender": "Male", "institution": "Cornell University", "country": "", "position": "Postdoc" }, { "name": "Kilian Q Weinberger", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Professor" }, { "name": "Luis Amaro", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "PhD student" }, { "name": "Peter Doerschuk", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "Full Professor" }, { "name": "Travis Zhang", "gender": "unknown", "institution": "Cornell University", "country": "US", "position": "MS student" }, { "name": "Xiangyu Chen", "gender": "Male", "institution": "Waymo LLC", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 39, "n_ref": 54, "n_ref_all": 75, "n_fig": 6, "n_tab": 3, "L_tab": 1942, "n_element_tab": 171, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 716, "n_element_tab_1": 88, "formula_len_all": 411, "formula_len_all_1": 363, "len_all": 132878, "len_all_1": 58377, "len_abs": 1202, "len_title": 104, "len_sents": 32297, "len_sents_1": 25982, "n_sents": 251, "n_sents_1": 191, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 1208, "title": "Hardware Simulation for Analog Ultrasonic 2D Convolution", "abs": "As its name suggests, the convolution operator is the basis and an essential component in Convolutional Neural Networks (CNNs). At the moment, modern CNN architectures rely heavily on parallel computation using GPUs and CPUs to perform many convolutions as fast as possible. However, the performance of computing CNNs is reaching its limit as the scaling of transistors approaches its size limits. The convolutional theorem suggests the possibility of using acoustic waves to efficiently perform the convolution operations through Fourier transforms in analog. This promises hardware that would be several orders of magnitude faster than existing silicon-based approaches. However, to date, nobody has shown the practical feasibility of such an approach. In this paper, we describe the first physics-based simulator for Ultrasonic Fourier Transform Convolutions (UFTC). By exploiting the diffraction nature of the waves, the Fourier transforms can be computed in the time it takes to propagate an ultrasonic wavefront. Our results show that ultrasonic computation could drastically improve the performance of CNNs by 12-458x FLOPS reduction and 1.3-4x computation speedup without loss of prediction accuracy.", "keywords": [ "simulation", "convolution", "ultrasonic", "hardware", "accelerator", "Fourier transform", "acoustic", "wave" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "ETX8NTEuCj", "primary_area": "learning theory", "n_author": 1, "author_info_list": [ { "name": "Zhengdao Chen", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 81, "n_formula_1": 52, "n_ref_uni": 34, "n_ref": 58, "n_ref_all": 76, "n_fig": 3, "n_tab": 1, "L_tab": 182, "n_element_tab": 28, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 26, "n_element_tab_1": 3, "formula_len_all": 8427, "formula_len_all_1": 3492, "len_all": 126880, "len_all_1": 63110, "len_abs": 746, "len_title": 47, "len_sents": 33721, "len_sents_1": 26889, "n_sents": 263, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 747, "title": "On the Interpolation Effect of Score Smoothing", "abs": "Score-based diffusion models have achieved remarkable progress in various domains with an ability to generate new data samples that do not exist in the training set. In this paper, we examine a hypothesis that this phenomenon manifests an interpolation effect caused by a smoothing of the empirical score function. Focusing on settings where the training set lies in a one-dimensional linear subspace, we take a distribution-agnostic perspective and study the interplay between score smoothing and the denoising dynamics with mathematically solvable models. We demonstrate how score smoothing can lead to the generation of samples that interpolate among the training data within the subspace while avoiding a full memorization of the training set.", "keywords": [ "score-based diffusion models", "score smoothing", "data interpolation", "generalization vs memorization", "subspace recovery" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "ETMIPPtJp9", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Bryan Hooi", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Kun Wang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Nian Liu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Xiaoxin He", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yuan Sui", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yufei He", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 33, "n_ref": 88, "n_ref_all": 121, "n_fig": 7, "n_tab": 9, "L_tab": 2596, "n_element_tab": 261, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1249, "n_element_tab_1": 132, "formula_len_all": 449, "formula_len_all_1": 522, "len_all": 181426, "len_all_1": 69057, "len_abs": 1533, "len_title": 140, "len_sents": 53043, "len_sents_1": 31649, "n_sents": 419, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1540, "title": "FiDeLiS: Faithful Reasoning in Large Language Model for Knowledge Graph Question Answering", "abs": "Large language models are often challenged by generating erroneous or `hallucinated' responses, especially in complex reasoning tasks.\nTo mitigate this, we propose a retrieval augmented reasoning method, FiDeLiS, which enhances knowledge graph question answering by anchoring responses to structured, verifiable reasoning paths. FiDeLiS uses a keyword-enhanced retrieval mechanism that fetches relevant entities and relations from a vector-based index of KGs to ensure high-recall retrieval. Once these entities and relations are retrieved, our method constructs candidate reasoning paths which are then refined using a stepwise beam search. This ensures that all the paths we create can be confidently linked back to KGs, ensuring they are accurate and reliable.\nA distinctive feature of our approach is its blend of natural language planning with beam search to optimize the selection of reasoning paths. Moreover, we redesign the way reasoning paths are scored by transforming this process into a deductive reasoning task, allowing the LLM to assess the validity of the paths through deductive reasoning rather than traditional logit-based scoring. This helps avoid misleading reasoning chains and reduces unnecessary computational demand. Extensive experiments demonstrate that our method, even as a training-free method which has lower computational costs and superior generality, outperforms established strong baselines across three datasets. The code of this paper will be released at https://anonymous.4open.science/r/FiDELIS-E7FC.", "keywords": [ "Large Language Models", "Knowledge Graph Question Answering", "Retrieval-Augmented Generation" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "ETFfXGM3e4", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Liang Zeng", "gender": "Male", "institution": "Institute for Interdisciplinary Information Sciences (IIIS), Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Lu Zhang", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 38, "n_formula_1": 19, "n_ref_uni": 45, "n_ref": 86, "n_ref_all": 127, "n_fig": 7, "n_tab": 11, "L_tab": 4354, "n_element_tab": 298, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1009, "n_element_tab_1": 74, "formula_len_all": 3257, "formula_len_all_1": 1515, "len_all": 169677, "len_all_1": 64508, "len_abs": 1161, "len_title": 156, "len_sents": 49946, "len_sents_1": 29310, "n_sents": 422, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1166, "title": "SAT-LDM: Provably Generalizable Image Watermarking for Latent Diffusion Models with Self-Augmented Training", "abs": "The proliferation of AI-generated images necessitates effective watermarking to protect intellectual property and identify fake content. While existing training-based watermarking methods show promise, they often struggle with generalization across diverse image styles and tend to produce noticeable artifacts. To this end, we introduce a provably generalizable image watermarking method for Latent Diffusion Models with Self-Augmented Training (SAT-LDM), which aligns the training and testing phases by a free generation distribution to bolster the watermarking module\u2019s generalization capabilities. We theoretically consolidate our method by proving that the free generation distribution contributes to its tight generalization bound without the need to collect new data. Extensive experimental results demonstrate that SAT-LDM achieves robust watermarking while significantly improving the quality of watermarked images across diverse styles. Furthermore, we conduct experimental analyses to demonstrate the strong generalization abilities of SAT-LDM. We hope our method offers a practical and convenient solution for securing high-fidelity AI-generated content.", "keywords": [ "image generation", "watermarking", "latent diffusion model" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 4 ], "presentation_list": [ 2, 3, 2, 4 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 2, 2, 5, 4 ] }, { "paper_id": "ESM2ixIp3X", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Edwin Simpson", "gender": "Male", "institution": "University of Bristol", "country": "", "position": "Lecturer" }, { "name": "Raul Santos-Rodriguez", "gender": "unknown", "institution": "University of Bristol", "country": "GB", "position": "Full Professor" }, { "name": "Yuxuan Ye", "gender": "Male", "institution": "University of Bristol", "country": "GB", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 33, "n_ref": 93, "n_ref_all": 118, "n_fig": 2, "n_tab": 4, "L_tab": 1174, "n_element_tab": 260, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 532, "n_element_tab_1": 98, "formula_len_all": 265, "formula_len_all_1": 172, "len_all": 116190, "len_all_1": 52744, "len_abs": 1250, "len_title": 139, "len_sents": 34414, "len_sents_1": 25088, "n_sents": 260, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1257, "title": "Revisiting and Extending Similarity-based Metrics in Summary Factual Consistency Detection", "abs": "Cutting-edge abstractive summarisers generate fluent summaries, but the factuality of the generated text is not guaranteed. \nEarly summary factuality evaluation metrics are usually based on n-gram overlap and embedding similarity, but are reported fail to align with human annotations.\nTherefore, many techniques for detecting factual inconsistencies build pipelines around natural language inference (NLI) or question-answering (QA) models with additional supervised learning steps. \nIn this paper, we revisit similarity-based metrics,\nshowing that this failure stems from the use of reference texts for comparison and the granularity of the comparison. \nWe propose a new zero-shot factuality evaluation metric,\nSentence-BERT Score (SBERTScore), which compares sentences between the summary and the source document. \nIt outperforms widely-used word-word metrics including BERTScore and can compete with existing NLI and QA-based factuality metrics on the benchmark without needing any fine-tuning.\nOur experiments indicate that each technique has different strengths, with SBERTScore particularly effective at identifying correct summaries.\nAdditionally, we demonstrate how a combination of techniques is more effective at detecting various types of error.", "keywords": [ "Factual Consistency", "Summarisation" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "ES9uz5Qa5W", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Hyungyu Shin", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Juho Kim", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "Associate Professor" }, { "name": "Sangwoo Mo", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Postdoc" }, { "name": "Yoonjoo Lee", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" }, { "name": "Yumin Cho", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Undergrad student" }, { "name": "JinhoSon", "gender": "Male", "institution": "Hanyang University", "country": "KR", "position": "Undergrad student" }, { "name": "Hyung-Kwon Ko", "gender": "Male", "institution": "Independent Researcher", "country": "", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 37, "n_ref": 54, "n_ref_all": 79, "n_fig": 10, "n_tab": 8, "L_tab": 1754, "n_element_tab": 247, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 1176, "n_element_tab_1": 162, "formula_len_all": 247, "formula_len_all_1": 247, "len_all": 196310, "len_all_1": 74333, "len_abs": 3662, "len_title": 122, "len_sents": 73986, "len_sents_1": 36113, "n_sents": 607, "n_sents_1": 231, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1507, "title": "GPT Shortcuts: Learning Iterative Text Generation Patterns from a Dialogue", "abs": "LLM-powered conversational interfaces (e.g., ChatGPT, Claude, and Gemini) support iterative text generation, enabling users to easily generate tailored texts (e.g., texts that should address domain-specific constraints) through a series of follow-up text editing requests. However, generating such tailored texts that address the user-specified constraints across multiple different contexts requires repetitive text generation efforts, which is cumbersome, inefficient, and demanding. To address this challenge, we introduce the concept of *GPT shortcuts*, which is designed to 1) learn iterative text generation patterns from a dialogue and 2) apply these learned patterns to *directly* generate the tailored text. GPT shortcuts generate texts that address necessary constraints while maintaining similar structural appearance to the target text in the dialogue, across different contexts. To assess the capability of language models in generating GPT shortcuts, we present ShortcutBench, a benchmark consisting of 250 crowdsourced iterative text generation dialogues across five text generation tasks. Using ShortcutBench, we conducted an analysis using six LLMs and four prompting methods, varying ways to specify necessary constraints to address in the prompt. We found that 1) larger models generally outperform smaller models, 2) self-explanatory constraints within the target text are effective, and 3) precisely specifying necessary constraints to address is critical for improving the performance.", "keywords": [ "Large Language Models", "Shortcuts", "Iterative text generations", "Reusable functions", "Conversational AI" ], "rating_list": [ 5, 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 1, 2, 2 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "ERv8ptegFi", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Aarav Pandya", "gender": "Male", "institution": "New York University", "country": "US", "position": "Researcher" }, { "name": "Brennan Shacklett", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Daphne Cornelisse", "gender": "Female", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Eugene Vinitsky", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Saman Kazemkhani", "gender": "unknown", "institution": "New York University", "country": "US", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 46, "n_ref": 75, "n_ref_all": 90, "n_fig": 6, "n_tab": 2, "L_tab": 313, "n_element_tab": 41, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 23, "n_element_tab_1": 4, "formula_len_all": 180, "formula_len_all_1": 706, "len_all": 147467, "len_all_1": 57149, "len_abs": 1116, "len_title": 118, "len_sents": 32782, "len_sents_1": 28001, "n_sents": 254, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 70, "L_abs": 1147, "title": "GPUDrive: Data-driven, multi-agent driving simulation at 1 million FPS", "abs": "Multi-agent learning algorithms have been successful at generating superhuman planning in various games but have had limited impact on the design of deployed multi-agent planners. A key bottleneck in applying these techniques to multi-agent planning is that they require billions of steps of experience. To enable the study of multi-agent planning at scale, we present GPUDrive, a GPU-accelerated, multi-agent simulator built on top of the Madrona Game Engine capable of generating over a million simulation steps per second. Observation, reward, and dynamics functions are written directly in C++, allowing users to define complex, heterogeneous agent behaviors that are lowered to high-performance CUDA. Despite these low-level optimizations, GPUDrive is fully accessible through Python, offering a seamless and efficient workflow for multi-agent, closed-loop simulation. Using GPUDrive, we train reinforcement learning agents on the Waymo Open Motion Dataset, achieving efficient goal-reaching in minutes and scaling to thousands of scenarios in hours. We open-source the code and pre-trained agents at \\url{www.github.com/Emerge-Lab/gpudrive}.", "keywords": [ "Simulation", "benchmark", "multi-agent reinforcement learning", "autonomous vehicles", "planning" ], "rating_list": [ 3, 8, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "ERce2rgMQC", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Ahmed Elgohary", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Senior Researcher" }, { "name": "Ahmed Magooda", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Daniel Khashabi", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Assistant Professor" }, { "name": "Jingyu Zhang", "gender": "unknown", "institution": "Johns Hopkins University", "country": "", "position": "PhD student" }, { "name": "Benjamin Van Durme", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" } ], "n_formula": 7, "n_formula_1": 3, "n_ref_uni": 42, "n_ref": 83, "n_ref_all": 110, "n_fig": 5, "n_tab": 8, "L_tab": 1692, "n_element_tab": 82, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 240, "n_element_tab_1": 26, "formula_len_all": 536, "formula_len_all_1": 205, "len_all": 234510, "len_all_1": 70963, "len_abs": 4361, "len_title": 87, "len_sents": 70662, "len_sents_1": 34293, "n_sents": 497, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1615, "title": "Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements", "abs": "The current paradigm for safety alignment of large language models (LLMs) follows a _one-size-fits-all_ approach: the model refuses to interact with any content deemed unsafe by the model provider. This approach lacks flexibility in the face of varying social norms across cultures and regions. In addition, users may have diverse safety needs, making a model with _static_ safety standards too restrictive to be useful, as well as too costly to be re-aligned.\n\nWe propose _Controllable Safety Alignment_ (CoSA), a framework designed to adapt models to diverse safety requirements without re-training. Instead of aligning a fixed model, we align models to follow _safety configs_\u2014free-form natural language descriptions of the desired safety behaviors\u2014that are provided as part of the system prompt. To adjust model safety behavior, authorized users only need to modify such safety configs at inference time. To enable that, we propose CoSAlign, a data-centric method for aligning LLMs to easily adapt to diverse safety configs. Furthermore, we devise a novel controllability evaluation protocol that considers both helpfulness and configured safety, summarizing them into CoSA-Score, and construct CoSApien, a _human-authored_ benchmark that consists of real-world LLM use cases with diverse safety requirements and corresponding evaluation prompts.\n\nWe show that CoSAlign leads to substantial gains of controllability over strong baselines including in-context alignment. Our framework encourages better representation and adaptation to pluralistic human values in LLMs, and thereby increasing their practicality.", "keywords": [ "large language models", "safety alignment", "pluralistic alignment" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "ERcGlGIM2D", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Divya Jyoti Bajpai", "gender": "Male", "institution": "Indian Institute of Technology, Bombay, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "PhD student" }, { "name": "Manjesh Kumar Hanawal", "gender": "Male", "institution": "Indian Institute of Technology Bombay", "country": "IN", "position": "Associate Professor" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 19, "n_ref": 54, "n_ref_all": 65, "n_fig": 6, "n_tab": 9, "L_tab": 2679, "n_element_tab": 237, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 4005, "n_element_tab_1": 294, "formula_len_all": 481, "formula_len_all_1": 504, "len_all": 148248, "len_all_1": 62549, "len_abs": 1156, "len_title": 106, "len_sents": 29345, "len_sents_1": 24669, "n_sents": 258, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1162, "title": "BLIPEE: Fast and Robust BLIP with Adversarially Trained Early Exits", "abs": "In recent years, Vision-Language Models (VLMs) have shown remarkable performance improvements in vision-language tasks. However, their large size poses challenges for real-world applications where inference latency is a concern. To tackle this issue, we propose employing Early Exit (EE) strategies in VLM. However, training exit classifiers in VLMs is challenging, particularly with limited labeled training data. To address this, we introduce BLIPEE, an adversarial training approach within a GAN-based framework. Here, each exit consists of a transformer layer and a classifier, and the transformer layer is adversarially trained to produce feature representations similar to the final layer, while a feature classifier serves as the discriminator. Our method focuses on performing input-adaptive inference that mitigates the overthinking issue and increases inference speed. Experimental results demonstrate the effectiveness of our approach in enhancing accuracy and model robustness by mitigating overthinking and the phenomenon of mid-crisis that we highlight. The anonymized source code is available at https://anonymous.4open.science/status/BLIPEE-3ED3.", "keywords": [ "Early Exits; Multimodal model" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "ERBm5WK8nq", "primary_area": "learning on time series and dynamical systems", "n_author": 6, "author_info_list": [ { "name": "Fugee Tsung", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Lifeng Shen", "gender": "Male", "institution": "Chongqing University of Post and Telecommunications", "country": "CN", "position": "Associate Professor" }, { "name": "Lingzheng Zhang", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou) ", "country": "CN", "position": "PhD student" }, { "name": "Shiyuan Piao", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Yimin Zheng", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "MS student" }, { "name": "Ziyue Li", "gender": "unknown", "institution": "University of Cologne", "country": "HK", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 10, "n_ref_uni": 29, "n_ref": 86, "n_ref_all": 106, "n_fig": 7, "n_tab": 10, "L_tab": 5070, "n_element_tab": 315, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 5271, "n_element_tab_1": 335, "formula_len_all": 629, "formula_len_all_1": 667, "len_all": 134766, "len_all_1": 76519, "len_abs": 1273, "len_title": 122, "len_sents": 39587, "len_sents_1": 30189, "n_sents": 360, "n_sents_1": 249, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1277, "title": "LeMoLE: LLM-enhanced Mixture of Linear Experts for Time Series Forecasting", "abs": "Recent research has shown that large language models (LLMs) can be effectively used for real-world time series forecasting due to their strong natural language understanding capabilities. However, aligning time series into semantic spaces of LLMs comes with high computational costs and inference complexity, particularly for long-range time series generation. Building on recent advancements in using linear models for time series, this paper introduces an LLM-enhanced mixture of linear experts for precise and efficient time series forecasting. This approach involves developing a mixture of linear experts with multiple lookback lengths and a new multimodal fusion mechanism. The use of a mixture of linear experts is efficient due to its simplicity, while the multimodal fusion mechanism adaptively combines multiple linear experts based on the learned features of the text modality from pre-trained large language models. In experiments, we rethink the need to align time series to LLMs by existing time-series large language models and further discuss their efficiency and effectiveness in time series forecasting. Our experimental results show that the proposed LeMoLE model presents lower prediction errors and higher computational efficiency than existing LLM models.", "keywords": [ "time series" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "EQz0C5PSyR", "primary_area": "applications to neuroscience & cognitive science", "n_author": 1, "author_info_list": [ { "name": "Yujin Cha", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "Postdoc" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 48, "n_ref": 54, "n_ref_all": 73, "n_fig": 12, "n_tab": 6, "L_tab": 2029, "n_element_tab": 509, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 389, "n_element_tab_1": 66, "formula_len_all": 498, "formula_len_all_1": 452, "len_all": 142719, "len_all_1": 65042, "len_abs": 1428, "len_title": 150, "len_sents": 40770, "len_sents_1": 31191, "n_sents": 286, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1434, "title": "Embedding Learning for Approximating Person-specific Cognitive Similarity", "abs": "Metric learning is often applied in scenarios where labels are well-defined or where there is a ground truth for semantic similarity between data points. However, in expert domains such as medical data, where experts perceive features and similarities differently on an individual basis, modeling psychological embeddings at the individual level can be beneficial. Such embeddings can predict factors that influence behavior, such as individual uncertainty, and support personalized learning strategies. Despite this potential, the amount of person-specific behavioral data that can be collected through similarity behavior sampling is insufficient in most scenarios, making modeling individual cognitive embeddings challenging and underexplored. In this study, we proposed integrating supervised learning on small-scale similarity sampling data with unsupervised autoencoder-based manifold learning to approximate person-specific psychological embeddings with significantly improved similarity inference performance. We conducted a large-scale experiment with 121 clinical physicians, measured their cognitive similarities using medical image data, and implemented person-specific models. Our results demonstrate that even in complex expert domains, such as medical imaging, where cognitive similarity varies between individuals, person-specific psychological embeddings can be effectively approximated using limited behavioral data.", "keywords": [ "Psychological embedding", "Metric learning", "Similarity", "Cognitive representation", "Autoencoder", "Medical image" ], "rating_list": [ 3, 6, 6, 5, 3 ], "soundness_list": [ 2, 3, 2, 2, 2 ], "presentation_list": [ 1, 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 2, 4, 3 ] }, { "paper_id": "EQgEMAD4kv", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Ke Cheng", "gender": "Male", "institution": "Ant Group", "country": "", "position": "Intern" }, { "name": "Mingbao Lin", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Shixuan Fan", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Weiyao Lin", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Wen Hu", "gender": "unknown", "institution": "Antgroup", "country": "", "position": "Algorithm expert" }, { "name": "Yuchen Cao", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Ziran Qin", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "jianguo Li", "gender": "Male", "institution": "Ant Group", "country": "", "position": "Director" } ], "n_formula": 30, "n_formula_1": 20, "n_ref_uni": 39, "n_ref": 109, "n_ref_all": 158, "n_fig": 17, "n_tab": 14, "L_tab": 9405, "n_element_tab": 761, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 686, "n_element_tab_1": 72, "formula_len_all": 3323, "formula_len_all_1": 1645, "len_all": 214882, "len_all_1": 60589, "len_abs": 1805, "len_title": 117, "len_sents": 62314, "len_sents_1": 27127, "n_sents": 492, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1813, "title": "CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences", "abs": "Large language models (LLMs)' proficiency in handling long sequences boosts KV caching demand. Recent efforts to evict KV cache have alleviated the burden for inference, but they often fail to allocate resources rationally across layers with different attention patterns. In this paper, we introduce Cascading and Adaptive KV cache Eviction (CAKE), a method that significantly improves LLM inference efficiency by optimizing KV cache eviction through an adaptive cache allocation strategy implemented via a cascading cache management and an innovative eviction indicator. We approach KV cache eviction as a ``cake-slicing problem,'' assessing each layer's KV cache needs by considering attention dynamics in both spatial and temporal dimensions. During the prompt prefilling, CAKE allocates rational cache size for layers by analyzing layer-specific KV cache preferences and manages the memory budgets with the guidance of these preferences in a cascading manner. This approach allows for a global view of cache size allocation, distributing resources optimally based on the diverse attention mechanisms across layers. Also, we've designed a new eviction indicator that considers the shifting importance of tokens over time, addressing a limitation in existing methods that often overlook temporal dynamics. Our comprehensive experiments on the LongBench and NeedleBench datasets show that CAKE is capable of preserving the performance of models when retaining only 3.2\\% KV cache and consistently outperforms current baselines across various models and memory constraints, especially in low-memory situations. Moreover, CAKE outperforms full cache with FlashAttention implementation, achieving 10$\\times$ faster decoding for 128K-token sequences and maintaining consistent decoding speed across sequence lengths.", "keywords": [ "Large Language Model", "Efficient Generative Inference", "Key-Value Cache" ], "rating_list": [ 6, 10, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 4, 5, 2, 3 ] }, { "paper_id": "EQZMx8Lc0n", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Chun-Nam Yu", "gender": "Male", "institution": "Nokia Bell Labs", "country": "", "position": "Research Scientist" }, { "name": "Md Kowsher", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Mojtaba Soltanalian", "gender": "Male", "institution": "University of Illinois at Chicago", "country": "", "position": "Associate Professor" }, { "name": "Niloofar Yousefi", "gender": "Female", "institution": "University of Central Florida", "country": "", "position": "Postdoc" }, { "name": "Tara Esmaeilbeig", "gender": "Female", "institution": "University of Illinois at Chicago", "country": "US", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 53, "n_ref": 95, "n_ref_all": 116, "n_fig": 7, "n_tab": 14, "L_tab": 5362, "n_element_tab": 592, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2362, "n_element_tab_1": 1, "formula_len_all": 984, "formula_len_all_1": 957, "len_all": 155475, "len_all_1": 54483, "len_abs": 1076, "len_title": 78, "len_sents": 39378, "len_sents_1": 26025, "n_sents": 272, "n_sents_1": 154, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1000, "title": "RoCoFT: Efficient Finetuning of Large Language Models with Row-Column Updates", "abs": "We propose RoCoFT, a parameter-efficient fine-tuning method for large-scale language models (LMs) based on updating only a few rows and columns of the weight matrices in transformers. Through extensive experiments with medium size LMs like BERT and RoBERTa, and larger LMs like Bloom-7B, Llama2-7B, and Llama2-13B, we show that our method gives comparable or better accuracies than state-of-art PEFT methods while also being more memory and computation- efficient. We also study the reason behind the effectiveness of our method with tools from neural tangent kernel theory. We empirically demonstrate that our kernel, constructed using a restricted set of row and column parameters, are numerically close to the full-parameter kernel and gives comparable classification performance. Ablation studies are conducted to investigate the impact of different algorithmic choices, including the selection strategy for rows and columns as well as the optimal rank for effective implementation of our method.", "keywords": [ "RoCoFT", "Parameter-efficient finetuning", "LLMs", "Neural Tangent Kernel" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 2, 1, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "EQAHilKZ8D", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Qingliang Chen", "gender": "Male", "institution": "Jinan University", "country": "CN", "position": "Full Professor" }, { "name": "Zhiyu Xu", "gender": "Male", "institution": "Jinan University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 10, "n_ref": 22, "n_ref_all": 35, "n_fig": 2, "n_tab": 4, "L_tab": 675, "n_element_tab": 51, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 675, "n_element_tab_1": 51, "formula_len_all": 278, "formula_len_all_1": 278, "len_all": 80715, "len_all_1": 65329, "len_abs": 2084, "len_title": 120, "len_sents": 24712, "len_sents_1": 24712, "n_sents": 147, "n_sents_1": 147, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1291, "title": "Utilizing Visual Properties to Achieve Better Representations of Objects", "abs": "In recent years, large vision models have made significant advancements and excelled in tasks such as detection, segmentation, and tracking. This is partly due to vision models\u2018 good representation of visual objects. Although the recently proposed SAM (the Segment Anything Model ) or the one/few-shot models based on SAM have wide applicability across many tasks, some researchers have found that they do not perform well on certain downstream tasks . In this paper, we focused on a specific group of these objects, which can be summarized as glass-like objects, and quantitatively studied the inadequacies related to the vision models\u2019 feature representation of glass-like objects using the representation accuracy(RA) metric we proposed. Then, we proposed a novel, extremely simple method that introduces almost no additional computations to address these inadequacies. The main idea is utilizing the visual properties of target objects to find representation dimensions which dominate in recognizing them and leveraging these information accordingly to achieve better representations of target objects. Using representation accuracy and setting these representations as reference in one-shot segmentation tasks, our experiments demonstrated the substantial effectiveness of our method.", "keywords": [ "Vision", "Segmentation" ], "rating_list": [ 1, 1, 3, 3, 3 ], "soundness_list": [ 1, 1, 2, 1, 2 ], "presentation_list": [ 1, 1, 1, 1, 1 ], "contribution_list": [ 1, 1, 1, 1, 1 ], "confidence_list": [ 4, 5, 3, 4, 4 ] }, { "paper_id": "EPHsIa0Ytg", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Baoxiang Wang", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Huanjian Zhou", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" }, { "name": "Lingxiao Huang", "gender": "Male", "institution": "Nanjing University", "country": "", "position": "Associate Professor" } ], "n_formula": 100, "n_formula_1": 9, "n_ref_uni": 28, "n_ref": 89, "n_ref_all": 115, "n_fig": 0, "n_tab": 2, "L_tab": 968, "n_element_tab": 30, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 74, "n_element_tab_1": 3, "formula_len_all": 11597, "formula_len_all_1": 727, "len_all": 211268, "len_all_1": 77680, "len_abs": 1050, "len_title": 138, "len_sents": 68790, "len_sents_1": 34991, "n_sents": 639, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 91, "L_abs": 1032, "title": "Improved Approximation Algorithms for $k$-Submodular Maximization via Multilinear Extension", "abs": "We investigate a generalized form of submodular maximization, referred to as $k$-submodular maximization, with applications across the domains of social networks and machine learning. In this work, we propose the multilinear extension of $k$-submodular functions and unified Frank-Wolfe-type frameworks based on that. This continuous framework accommodates 1) monotone or non-monotone functions, and 2) various constraint types including matroid constraints, knapsack constraints, and their combinations. Notably, we attain an asymptotically optimal $1/2$-approximation for monotone $k$-submodular maximization problems with knapsack constraints, surpassing previous $1/3$-approximation results, and a factor-$1/3$ approximation for non-monotone $k$-submodular maximization problems with knapsack constraints and matroid constraints which outperforms previous $0.245$-approximation results. The foundation for our analysis stems from new insights into specific linear and monotone properties pertaining to the multilinear extension.", "keywords": [ "$k$-submodular maximization", "approximation algorithm", "$k$-multilinear extension" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "EP6n8LCEK6", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Jundong Li", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "Assistant Professor" }, { "name": "Pingzhi Li", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Shyam Marjit", "gender": "Male", "institution": "Indian Institute of Science, Indian institute of science, Bangalore", "country": "IN", "position": "Researcher" }, { "name": "Song Wang", "gender": "Male", "institution": "University of Virginia", "country": "US", "position": "PhD student" }, { "name": "Tianlong Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Xinyu Zhao", "gender": "unknown", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Yinhan He", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Zhen Tan", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Zihan Chen", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "huan liu", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 7, "n_ref_uni": 38, "n_ref": 59, "n_ref_all": 107, "n_fig": 11, "n_tab": 8, "L_tab": 5573, "n_element_tab": 308, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2023, "n_element_tab_1": 76, "formula_len_all": 655, "formula_len_all_1": 394, "len_all": 217985, "len_all_1": 75398, "len_abs": 2790, "len_title": 127, "len_sents": 70251, "len_sents_1": 33542, "n_sents": 624, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1682, "title": "Understanding Prejudice and Fidelity of Diverge-to-Converge Multi-Agent Systems", "abs": "Large language model (LLM) agents have demonstrated substantial potential across various tasks, particularly in multi-agent systems. Among these, \\textit{Diverge-to-Converge} (D2C) frameworks stand out for their ability to iteratively diversify and converge intermediate thoughts to improve problem-solving. In this paper, we conduct a comprehensive study on the \\textit{\\textbf{prejudice}} and \\textit{\\textbf{fidelity}} of typical D2C frameworks, including both model-level and society-level frameworks. \n\\ding{182} In the \\textit{prejudice} section, we uncover an inherent \\textit{confirmation bias} in D2C systems, which not only leads to suboptimal performance, but also amplifies social biases, such as gender discrimination and political partisanship. Surprisingly, we find that by reframing open-ended problems into controlled initialized problems, this bias can be leveraged to foster more equitable and effective agent interactions, ultimately improving performance.\n\\ding{183} In the \\textit{fidelity} section, we explore the scaling laws of D2C frameworks at different granularities, revealing that increasing the number of agents enhances performance only when the system is not yet saturated---such as in complex tasks or with weaker agents. In saturated scenarios, however, adding more agents can degrade performance. \nTo facilitate further study, we develop \\texttt{APF-Bench}, a benchmark specifically designed to evaluate such inherent weaknesses of D2C frameworks. \nWe hope our findings offer instructional insights into the strengths and limitations of D2C multi-agent systems, offering guidance for developing more robust and effective collaborative AI systems.", "keywords": [ "Large language model agents", "Multi-Agent System", "Benchmark" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "EP09OGPRzk", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Euiseok Hwang", "gender": "Male", "institution": "Gwangju Institute of Science and Technology", "country": "KR", "position": "Full Professor" }, { "name": "Minseok Jeong", "gender": "Male", "institution": "KAIST, Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "GIUP SEO", "gender": "unknown", "institution": "Gwangju Institute of Science and Technology", "country": "KR", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 19, "n_ref_uni": 35, "n_ref": 48, "n_ref_all": 66, "n_fig": 9, "n_tab": 5, "L_tab": 1222, "n_element_tab": 88, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 612, "n_element_tab_1": 40, "formula_len_all": 2877, "formula_len_all_1": 1346, "len_all": 181260, "len_all_1": 66782, "len_abs": 1730, "len_title": 174, "len_sents": 61182, "len_sents_1": 29882, "n_sents": 451, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 125, "L_abs": 1725, "title": "L-PINN: A Langevin Dynamics Approach with Balanced Sampling to Improve Learning Stability in Physics-Informed Neural Networks", "abs": "Physics-informed neural networks (PINNs) have emerged as a promising technique solving partial differential equations (PDEs). However, PINNs face challenges in resource efficiency (e.g., repeatedly sampling of collocation points) and achieving fast convergence to accurate solutions. To address these issues, adaptive sampling methods that focus on collocation points with high residual values have been proposed, enhancing both resource efficiency and solution accuracy. While these high residual-based sampling methods have demonstrated exceptional performance in solving certain stiff PDEs, their potential drawbacks, particularly the relative neglect of points with medium and low residuals, remain under-explored. In this paper, we investigate the limitations of high residual-based methods concerning learning stability as model complexity increases. We provide a theoretical analysis demonstrating that high residual-based methods require tighter upper bound on the learning rate to maintain stability. To overcome this limitation, we present a novel Langevin dynamics-based PINN (L-PINN) framework for adaptive sampling of collocation points, which is designed to improve learning stability and convergence speed. To validate the effectiveness, we evaluated the L-PINN framework against existing adaptive sampling approaches for PINNs. Our results indicate that the L-PINN framework achieves superior relative $L^{2}$ error performance in solutions while demonstrating faster or comparable convergence stability. Furthermore, we showed that our framework maintains robust performance across varying model complexities, suggesting its potential for compatibility with larger, more complex neural network architectures.", "keywords": [ "Physics-informed neural network", "Langevin dynamics", "Adaptive sampling method" ], "rating_list": [ 5, 6, 8, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "EOPLy80bBm", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Daniil Ignatev", "gender": "Male", "institution": "Utrecht University", "country": "NL", "position": "PhD student" }, { "name": "Dong Nguyen", "gender": "unknown", "institution": "Utrecht University", "country": "", "position": "Assistant Professor" }, { "name": "Hugh Mee Wong", "gender": "unknown", "institution": "Utrecht University", "country": "NL", "position": "PhD student" }, { "name": "Yingjin Song", "gender": "Female", "institution": "Utrecht University", "country": "NL", "position": "PhD student" }, { "name": "Yupei Du", "gender": "Male", "institution": "Utrecht University", "country": "NL", "position": "PhD student" }, { "name": "Albert Gatt", "gender": "Male", "institution": "Utrecht University", "country": "", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 44, "n_ref": 88, "n_ref_all": 127, "n_fig": 20, "n_tab": 2, "L_tab": 4944, "n_element_tab": 27, "n_fig_1": 17, "n_tab_1": 1, "L_tab_1": 100, "n_element_tab_1": 17, "formula_len_all": 409, "formula_len_all_1": 324, "len_all": 159959, "len_all_1": 64484, "len_abs": 3541, "len_title": 137, "len_sents": 44776, "len_sents_1": 32886, "n_sents": 301, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1043, "title": "Disentangling the Roles of Representation and Selection in Data Pruning (for Fine-Tuning)", "abs": "Data pruning, the process of carefully selecting a small subset of training data, has been shown to improve both training efficiency and performance. It typically involves two steps: (1) obtaining a representation for each instance, and (2) applying a selection algorithm using these representations. However, the distinct roles of these two steps, as well as their interactions, remain unclear. To address this, we conduct a systematic study of data pruning, focusing on NLP fine-tuning. Our theoretical and empirical findings reveal that data representation often plays a more fundamental role than the selection algorithm: gradients, despite being computationally expensive, provide stronger pruning signals than other representations, making gradient-based methods consistently outperform cheaper alternatives. We also demonstrate that different selection algorithms excel in specific scenarios but are heavily influenced by the chosen representation. These insights provide clear guidelines for future research and practical applications.", "keywords": [ "data pruning", "fine-tuning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 2, 1, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "EOLBKobfd1", "primary_area": "applications to neuroscience & cognitive science", "n_author": 4, "author_info_list": [ { "name": "Grace W Lindsay", "gender": "Female", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Lerrel Joseph Pinto", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Venkatesh Pattabiraman", "gender": "Male", "institution": "New York University", "country": "US", "position": "MS student" }, { "name": "Nikhil Xie Bhattasali", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" } ], "n_formula": 21, "n_formula_1": 8, "n_ref_uni": 50, "n_ref": 90, "n_ref_all": 110, "n_fig": 13, "n_tab": 4, "L_tab": 847, "n_element_tab": 46, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 256, "n_element_tab_1": 37, "formula_len_all": 11801, "formula_len_all_1": 9355, "len_all": 156642, "len_all_1": 71503, "len_abs": 1281, "len_title": 60, "len_sents": 45332, "len_sents_1": 31068, "n_sents": 338, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1286, "title": "Neural Circuit Architectural Priors for Quadruped Locomotion", "abs": "Learning-based approaches to quadruped locomotion commonly adopt generic policy architectures like fully connected MLPs. As such architectures contain few inductive biases, it is in practice common to incorporate priors in the form of rewards, training curricula, imitation data, or trajectory generators. In nature, animals are born with priors in the form of their nervous system's architecture, which has been shaped by evolution to confer innate ability and efficient learning. For instance, a horse can walk within hours of birth and can quickly improve with practice. Such architectural priors can also be useful in ANN architectures for AI. In this work, we explore the advantages of a biologically inspired ANN architecture for quadruped locomotion based on neural circuits in the limbs and spinal cord of mammals. Our architecture achieves good innate performance and comparable final performance to MLPs, while using less data and orders of magnitude fewer parameters. Our architecture also exhibits better generalization to task variations, even admitting deployment on a physical robot without standard sim-to-real methods. This work shows that neural circuits can provide valuable architectural priors for locomotion and encourages future work in other sensorimotor skills.", "keywords": [ "neuroscience", "neural circuits", "motor control" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "EO8xpnW7aX", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Donglin Yang", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Renjie Liao", "gender": "Male", "institution": "Department of Electrical and Computer Engineering, The University of British Columbia", "country": "CA", "position": "Assistant Professor" }, { "name": "Yongxing Zhang", "gender": "Male", "institution": "Vector Institute", "country": "CA", "position": "Intern" } ], "n_formula": 65, "n_formula_1": 45, "n_ref_uni": 48, "n_ref": 125, "n_ref_all": 162, "n_fig": 3, "n_tab": 8, "L_tab": 3169, "n_element_tab": 197, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 2083, "n_element_tab_1": 93, "formula_len_all": 33633, "formula_len_all_1": 30234, "len_all": 240771, "len_all_1": 121995, "len_abs": 1192, "len_title": 91, "len_sents": 67391, "len_sents_1": 40041, "n_sents": 612, "n_sents_1": 310, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 43, "L_abs": 1203, "title": "Learning to Permute with Discrete Diffusion", "abs": "The group of permutations $S_n$, also known as the finite symmetric groups, are essential in fields such as combinatorics, physics, and chemistry. However, learning a probability distribution over $S_n$ poses significant challenges due to its intractable size and discrete nature. In this paper, we introduce *SymmetricDiffusers*, a novel discrete diffusion model that simplifies the task of learning a complicated distribution over $S_n$ by decomposing it into learning simpler transitions of the reverse diffusion using deep neural networks. We identify the riffle shuffle as an effective forward transition and provide empirical guidelines for selecting the diffusion length based on the theory of random walks on finite groups. Additionally, we propose a generalized Plackett-Luce (PL) distribution for the reverse transition, which is provably more expressive than the PL distribution. We further introduce a theoretically grounded \"denoising schedule\" to improve sampling and learning efficiency. Extensive experiments show that our model achieves state-of-the-art or comparable performances on solving tasks including sorting 4-digit MNIST images, jigsaw puzzles, and traveling salesman problems.", "keywords": [ "Finite Symmetric Groups", "Discrete Diffusion", "Permutations", "Riffle Shuffles", "Plackett-Luce Distribution", "Sorting", "Jigsaw Puzzle" ], "rating_list": [ 8, 8, 8, 8, 8 ], "soundness_list": [ 3, 3, 4, 4, 3 ], "presentation_list": [ 3, 3, 4, 4, 3 ], "contribution_list": [ 3, 3, 4, 3, 3 ], "confidence_list": [ 3, 4, 2, 3, 3 ] }, { "paper_id": "EO2hZTtK3M", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Feixiang Wang", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Shiguang Shan", "gender": "Male", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Shuang Yang", "gender": "Female", "institution": "Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Xilin Chen", "gender": "Male", "institution": "Institute of Computing Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 21, "n_ref_uni": 57, "n_ref": 98, "n_ref_all": 110, "n_fig": 3, "n_tab": 11, "L_tab": 3051, "n_element_tab": 82, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1285, "n_element_tab_1": 29, "formula_len_all": 980, "formula_len_all_1": 1411, "len_all": 168822, "len_all_1": 71111, "len_abs": 2921, "len_title": 122, "len_sents": 43364, "len_sents_1": 31846, "n_sents": 312, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 2148, "title": "CM^2: Cross-Modal Contextual Modeling for Audio-Visual Speech Enhancement", "abs": "Audio-Visual Speech Enhancement (AVSE) aims to improve speech quality in noisy environments by utilizing synchronized audio and visual cues.\nIn real-world scenarios, noise is often non-stationary, interfering with speech signals at varying intensities over time.\nDespite these fluctuations, humans can discern and understand masked spoken words as if they were clear.\nThis capability stems from the auditory system's ability to perceptually reconstruct interrupted speech using visual cues and semantic context in noisy environments, a process known as phonemic restoration.\nInspired by this phenomenon, we propose Cross-Modal Contextual Modeling (CM$^2$), integrating contextual information across different modalities and levels to enhance speech quality. \nSpecifically, we target two types of contextual information: semantic-level context and signal-level context.\nSemantic-level context enables the model to infer missing or corrupted content by leveraging semantic consistency across segments.\nSignal-level context further explores coherence within the signals developed from the semantic consistency.\nAdditionally, we particularly highlight the role of visual appearance in modeling the frequency-domain characteristics of speech, aiming to further refine and enrich the expression of these contexts.\nGuided by this understanding, we introduce a Semantic Context Module (SeCM) at the very beginning of our framework to capture the initial semantic contextual information from both audio and visual modalities.\nNext, we propose a Signal Context Module (SiCM) to obtain signal-level contextual information from both raw noisy audio signal and the previously acquired audio-visual semantic-level context.\nBuilding on this rich contextual information, we finally introduce a Cross-Context Fusion Module (CCFM) to facilitate fine-grained context fusion across different modalities and types of contexts for further speech enhancement process.\nComprehensive evaluations across various datasets demonstrate that our method significantly outperforms current state-of-the-art approaches, particularly in low signal-to-noise ratio (SNR) environments.", "keywords": [ "Speech Enhancement", "Audio-Visual", "Contextual Modeling" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 5, 3, 3 ] }, { "paper_id": "ENv1CeTwxc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Gim Hee Lee", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Associate Professor" }, { "name": "Seungjun Lee", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yuyang Zhao", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 72, "n_ref": 171, "n_ref_all": 193, "n_fig": 15, "n_tab": 13, "L_tab": 3483, "n_element_tab": 395, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2282, "n_element_tab_1": 123, "formula_len_all": 723, "formula_len_all_1": 648, "len_all": 200598, "len_all_1": 64945, "len_abs": 411, "len_title": 83, "len_sents": 48066, "len_sents_1": 28539, "n_sents": 362, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 35, "L_abs": 1652, "title": "Segment Any 3D Object with Language", "abs": "In this paper, we investigate Open-Vocabulary 3D Instance Segmentation (OV-3DIS) with free-form language instructions. Earlier works mainly rely on annotated base categories for training which leads to limited generalization to unseen novel categories. To mitigate the poor generalizability to novel categories, recent works generate class-agnostic masks or projecting generalized masks from 2D to 3D, subsequently classifying them with the assistance of 2D foundation model. However, these works often disregard semantic information in the mask generation, leading to sub-optimal performance. Instead, generating generalizable but semantic-aware masks directly from 3D point clouds would result in superior outcomes. To the end, we introduce Segment any 3D Object with LanguagE ($\\textbf{SOLE}$), which is a semantic and geometric-aware visual-language learning framework with strong generalizability by generating semantic-related masks directly from 3D point clouds. Specifically, we propose a multimodal fusion network to incorporate multimodal semantics in both backbone and decoder. In addition, to align the 3D segmentation model with various language instructions and enhance the mask quality, we introduce three types of multimodal associations as supervision. Our SOLE outperforms previous methods by a large margin on ScanNetv2, ScanNet200, and Replica benchmarks, and the results are even closed to the fully-supervised counterpart despite the absence of class annotations in the training. Furthermore, extensive qualitative results demonstrate the versatility of our SOLE to language instructions. The code will be made publicly available.", "keywords": [ "Open-set", "3D Instance Segmentation", "Multimodal" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "ENVwvyiJXY", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Suha Kwak", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Associate Professor" }, { "name": "Youngseok Yoon", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 45, "n_ref": 104, "n_ref_all": 118, "n_fig": 4, "n_tab": 3, "L_tab": 810, "n_element_tab": 68, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 905, "n_element_tab_1": 62, "formula_len_all": 1117, "formula_len_all_1": 1141, "len_all": 120951, "len_all_1": 60253, "len_abs": 953, "len_title": 94, "len_sents": 32516, "len_sents_1": 29113, "n_sents": 174, "n_sents_1": 152, "venue": "ICLR2025", "decision": "withdraw", "L_title": 46, "L_abs": 789, "title": "Dataset Distillation for Domain Generalization", "abs": "Dataset Distillation (DD) has been applied to various downstream tasks and recently scaled to ImageNet-1k, highlighting its potential for practical applications. However, in real-world scenarios, robustness to unseen domains is essential, and the robustness of models trained on synthetic datasets remains uncertain. To address this, we propose a novel task, Dataset Distillation for Domain Generalization (DD for DG), and evaluate the unseen domain generalization of models trained on synthetic datasets distilled by state-of-the-art DD methods using the DomainBed benchmark. Additionally, we introduce a new method for this task, which interprets DD through the lens of image style transfer, achieving superior performance in unseen domain generalization compared to baseline approaches.", "keywords": [ "Dataset Distillation", "Domain Generalization", "Style Transfer", "Self-Supervised Learning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "EMpvfnzQqD", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Ina Kodrasi", "gender": "unknown", "institution": "Idiap Research Institute", "country": "CH", "position": "Researcher" }, { "name": "Petr Motlicek", "gender": "Male", "institution": "Idiap Research Institute", "country": "", "position": "Principal Researcher" }, { "name": "Shashi Kumar", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "PhD student" }, { "name": "Yacouba Kaloga", "gender": "unknown", "institution": "Idiap Research Institute", "country": "CH", "position": "Postdoc" } ], "n_formula": 42, "n_formula_1": 19, "n_ref_uni": 54, "n_ref": 79, "n_ref_all": 96, "n_fig": 6, "n_tab": 3, "L_tab": 654, "n_element_tab": 63, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 238, "n_element_tab_1": 24, "formula_len_all": 3469, "formula_len_all_1": 1693, "len_all": 164475, "len_all_1": 62376, "len_abs": 1263, "len_title": 0, "len_sents": 49134, "len_sents_1": 28197, "n_sents": 413, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1269, "title": "OTTC: A differentiable alignment approach to automatic speech recognition", "abs": "The Connectionist Temporal Classification (CTC) and transducer-based models are widely used for end-to-end (E2E) automatic speech recognition (ASR). These methods maximize the marginal probability over all valid alignments within the probability lattice over the vocabulary during training. However, research has shown that most alignments are highly improbable, with the model often concentrating on a limited set, undermining the purpose of considering all possible alignments. In this paper, we propose a novel differentiable alignment framework based on a one-dimensional optimal transport formulation, enabling the model to learn a single alignment and perform ASR in an E2E manner.\nWe define a pseudo-metric, called Sequence Optimal Transport Distance (SOTD), over the sequence space and highlight its theoretical properties.\nBased on the SOTD, we propose Optimal Temporal Transport Classification (OTTC) loss for ASR and contrast its behavior with that of CTC.\nExperimental results on the English Librispeech and AMI datasets demonstrate that our method achieves competitive performance compared to CTC in ASR.\nWe believe this work opens up a potential new direction for research in ASR, offering a foundation for the community to further explore and build upon.", "keywords": [ "ASR", "Optimal Transport", "Sequence to Sequence", "Alignment" ], "rating_list": [ 3, 6, 8 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 4, 4 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "EMMnAd3apQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Junlong Du", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Ke Yan", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "Researcher" }, { "name": "Shouhong Ding", "gender": "Male", "institution": "Tencent Youtu Lab", "country": "CN", "position": "researcher" }, { "name": "Xiaoqiang Li", "gender": "Male", "institution": "Shanghai University", "country": "CN", "position": "Full Professor" }, { "name": "Yuanchen Wu", "gender": "unknown", "institution": "Shanghai University", "country": "CN", "position": "MS student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 46, "n_ref": 89, "n_ref_all": 114, "n_fig": 10, "n_tab": 4, "L_tab": 1744, "n_element_tab": 206, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 1978, "n_element_tab_1": 250, "formula_len_all": 601, "formula_len_all_1": 514, "len_all": 164341, "len_all_1": 68876, "len_abs": 1341, "len_title": 131, "len_sents": 44001, "len_sents_1": 30862, "n_sents": 321, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1361, "title": "ToVE: Efficient Vision-Language Learning via Knowledge Transfer from Vision Experts", "abs": "Vision-language (VL) learning requires extensive visual perception capabilities, such as fine-grained object recognition and spatial perception. Recent works typically rely on training huge models on massive datasets to develop these capabilities. As a more efficient alternative, this paper proposes a new framework that Transfers the knowledge from a hub of Vision Experts (ToVE) for efficient VL learning, leveraging pre-trained vision expert models to promote visual perception capability. Specifically, building on a frozen CLIP image encoder that provides vision tokens for image-conditioned language generation, ToVE introduces a hub of multiple vision experts and a token-aware gating network that dynamically routes expert knowledge to vision tokens. In the transfer phase, we propose a \"residual knowledge transfer\" strategy, which not only preserves the generalizability of the vision tokens but also allows selective detachment of low-contributing experts to improve inference efficiency. Further, we explore to merge these expert knowledge to a single CLIP encoder, creating a knowledge-merged CLIP that produces more informative vision tokens without expert inference during deployment. Experiment results across various VL tasks demonstrate that the proposed ToVE achieves competitive performance with two orders of magnitude fewer training data.", "keywords": [ "Vision-language Modeling", "Knowledge Transfer", "Vision Experts" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "EMKZyZSl70", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Mostofa Rafid Uddin", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Min Xu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 6, "n_ref_uni": 45, "n_ref": 122, "n_ref_all": 159, "n_fig": 17, "n_tab": 2, "L_tab": 716, "n_element_tab": 142, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1095, "n_element_tab_1": 145, "formula_len_all": 678, "formula_len_all_1": 424, "len_all": 172033, "len_all_1": 66571, "len_abs": 1324, "len_title": 150, "len_sents": 54587, "len_sents_1": 30293, "n_sents": 447, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1331, "title": "DualContrast: Unsupervised Disentangling of Content and Transformations with Implicit Parameterization", "abs": "Unsupervised disentanglement of content and transformation is significantly important for analyzing shape focused scientific image datasets, given their efficacy in solving downstream image-based shape-analyses tasks. The existing relevant works address the problem by explicitly parameterizing the transformation latent codes in a generative model, significantly reducing their expressiveness. Moreover, they are not applicable in cases where transformations can not be readily parametrized. An alternative to such explicit approaches is contrastive methods with data augmentation, which implicitly disentangles transformations and content. However, the existing contrastive strategies are insufficient to this end. Therefore, we developed a novel contrastive method with generative modeling, DualContrast, specifically for unsupervised disentanglement of content and transformations in shape focused image datasets. DualContrast creates positive and negative pairs for content and transformation from data and latent spaces. Our extensive experiments showcase the efficacy of DualContrast over existing self-supervised and explicit parameterization approaches. With DualContrast, we disentangled protein composition and conformations in cellular 3D protein images, which was unattainable with existing disentanglement approaches.", "keywords": [ "Unsupervised Learning", "Shape Analysis", "Identifiability in Representation Learning", "Disentangled Representation Learning", "ML in Biology" ], "rating_list": [ 3, 6, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "EM93t94zEi", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Deshui Miao", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Huchuan Lu", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Professor" }, { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Xin Li", "gender": "Male", "institution": "Pengcheng Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yaowei Wang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Zhenyu He", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 42, "n_ref": 102, "n_ref_all": 128, "n_fig": 9, "n_tab": 8, "L_tab": 6649, "n_element_tab": 606, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 3017, "n_element_tab_1": 289, "formula_len_all": 454, "formula_len_all_1": 330, "len_all": 149005, "len_all_1": 59553, "len_abs": 1332, "len_title": 119, "len_sents": 39454, "len_sents_1": 25652, "n_sents": 299, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1386, "title": "Learning Spatial-Semantic Features for Robust Video Object Segmentation", "abs": "Tracking and segmenting multiple similar objects with distinct or complex parts in long-term videos is particularly challenging due to the ambiguity in identifying target components and the confusion caused by occlusion, background clutter, and changes in appearance or environment over time. In this paper, we propose a robust video object segmentation framework that learns spatial-semantic features and discriminative object queries to address the above issues. Specifically, we construct a spatial-semantic block comprising a semantic embedding component and a spatial dependency modeling part for associating global semantic features and local spatial features, providing a comprehensive target representation. In addition, we develop a masked cross-attention module to generate object queries that focus on the most discriminative parts of target objects during query propagation, alleviating noise accumulation to ensure effective long-term query propagation. The experimental results show that the proposed method sets new state-of-the-art performance on multiple data sets, including the DAVIS2017 test (\\textbf{87.8\\%}), YoutubeVOS 2019 (\\textbf{88.1\\%}), MOSE val (\\textbf{74.0\\%}), and LVOS test (\\textbf{73.0\\%}), which demonstrate the effectiveness and generalization capacity of the proposed method. We will make all the source code and trained models publicly available.", "keywords": [ "Video Object Segmentation", "Spatial-Semantic Feature", "Long-Term", "Discriminative Object Queries" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "ELQ8X02IEp", "primary_area": "interpretability and explainable AI", "n_author": 3, "author_info_list": [ { "name": "Xujie Si", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Assistant Professor" }, { "name": "Yuhe Jiang", "gender": "Female", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Zexin Xue", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "MS student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 32, "n_ref": 47, "n_ref_all": 66, "n_fig": 11, "n_tab": 3, "L_tab": 925, "n_element_tab": 79, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 925, "n_element_tab_1": 79, "formula_len_all": 601, "formula_len_all_1": 601, "len_all": 91207, "len_all_1": 61100, "len_abs": 1501, "len_title": 102, "len_sents": 27117, "len_sents_1": 26977, "n_sents": 184, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1515, "title": "Learning Reliable Rules by Re-generating Deep Features", "abs": "Improving the interpretability and reliability of deep learning models is essential for advancing machine learning applications, though it remains a significant challenge. One promising approach is the integration of logical reasoning into deep learning systems. Previous works have demonstrated that SATNet, a differentiable MaxSAT solver, can learn interpretable and reliable rules from input-output examples in puzzle domains. In this work, we propose *Visual SATNet* (Vi-SATNet), an extended version of SATNet capable of learning logical reasoning rules in more general and complex domains, such as the feature space of real-life images. We find that, given a pre-trained deep convolutional neural network (CNN) architecture, a Vi-SATNet layer can be integrated and trained efficiently to learn a set of reasoning rules on the deep features, guiding the classifier\u2019s decision. Vi-SATNets are trained to perform feature re-generation tasks for a given image dataset, where the re-generated features maintain high accuracy when used for image classification, proving their quality. In our experiment on the Imagenette dataset with a pre-trained VGG19 model, masking out 10\\% to 80\\% of the features results in classification accuracy ranging from 98.50\\% to 93.92\\% with Vi-SATNet re-generation, compared to 97.07\\% to 9.83\\% without re-generation. Furthermore, we introduce a visualization method to illustrate the rules learned by Vi-SATNets, thereby enhancing the interpretability of the pre-trained CNN model.", "keywords": [ "Interpretable ML", "Neuro-symbolic AI", "SATNet", "Logical Reasoning" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 2, 3, 3 ] }, { "paper_id": "EKfcngSxwD", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Ahmet Iscen", "gender": "Male", "institution": "Google", "country": "FR", "position": "Researcher" }, { "name": "Christian Rupprecht", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" }, { "name": "Cordelia Schmid", "gender": "Female", "institution": "Google", "country": "FR", "position": "Researcher" }, { "name": "Jinghao Zhou", "gender": "Male", "institution": "University of Oxford, University of Oxford", "country": "GB", "position": "PhD student" }, { "name": "Mathilde Caron", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Philip Torr", "gender": "unknown", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 53, "n_ref": 140, "n_ref_all": 156, "n_fig": 7, "n_tab": 15, "L_tab": 11381, "n_element_tab": 1344, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 3384, "n_element_tab_1": 250, "formula_len_all": 416, "formula_len_all_1": 212, "len_all": 218867, "len_all_1": 73278, "len_abs": 789, "len_title": 123, "len_sents": 51693, "len_sents_1": 30563, "n_sents": 465, "n_sents_1": 275, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 794, "title": "Incrementally Adapting Generative Vision-Language Models with Task Codebook", "abs": "With the help of large-scale pre-training, generative Vision-Language Models (VLMs) have acquired general-purpose capabilities.\nAs downstream applications diversify, it is imperative for VLMs to learn and adapt continuously without experiencing catastrophic forgetting or necessitating complete retraining.\nIn this work, we analyze the forgetting behavior of VLMs and propose a solution to enhance their incremental learning abilities.\nWe introduce a Task Codebook within VLMs, enabling efficient retrieval of task-specific parameters for model adaptation. \nOur evaluation encompasses a diverse set of tasks spanning a wide range of visual domains and textual instructions. \nExperiments demonstrate that our approach effectively mitigates forgetting, even under highly demanding task sequences.", "keywords": [ "Generative Vision-Language Models", "Incremental Learning" ], "rating_list": [ 3, 6, 5 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 3, 4, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 5, 3 ] }, { "paper_id": "EKaVO0ceh8", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Hoang V. Tran", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Huyen Trang Pham", "gender": "Female", "institution": "VinAI Research", "country": "VN", "position": "Intern" }, { "name": "Minh-Khoi Nguyen-Nhat", "gender": "Male", "institution": "FPT", "country": "VN", "position": "Researcher" }, { "name": "Tam Le", "gender": "Male", "institution": "The Institute of Statistical Mathematics (ISM)", "country": "JP", "position": "Assistant Professor" }, { "name": "Tan Minh Nguyen", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Thanh Chu", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Undergrad student" }, { "name": "Tho Tran Huu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 82, "n_formula_1": 31, "n_ref_uni": 56, "n_ref": 107, "n_ref_all": 136, "n_fig": 11, "n_tab": 7, "L_tab": 1565, "n_element_tab": 219, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 247, "n_element_tab_1": 48, "formula_len_all": 7000, "formula_len_all_1": 3009, "len_all": 221861, "len_all_1": 63684, "len_abs": 1333, "len_title": 98, "len_sents": 64911, "len_sents_1": 25057, "n_sents": 676, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 50, "L_abs": 1345, "title": "Projection Optimal Transport on Tree-Ordered Lines", "abs": "Many variants of Optimal Transport (OT) have been developed to address its heavy computation. Among them, notably, Sliced Wasserstein (SW) is widely used for application domains by projecting the OT problem onto one-dimensional lines, and leveraging the closed-form expression of the univariate OT to reduce the computational burden. However, projecting measures onto low-dimensional spaces can lead to a loss of topological information. To mitigate this issue, in this work, we propose to replace one-dimensional lines with a more intricate structure, called \\emph{tree systems}. This structure is metrizable by a tree metric, which yields a closed-form expression for OT problems on tree systems. We provide an extensive theoretical analysis to formally define tree systems with their topological properties, introduce the concept of splitting maps, which operate as the projection mechanism onto these structures, then finally propose a novel variant of Radon transform for tree systems and verify its injectivity. This framework leads to an efficient metric between measures, termed Tree-Sliced Wasserstein distance on Systems of Lines (TSW-SL). By conducting a variety of experiments on gradient flows, image style transfer, and generative models, we illustrate that our proposed approach performs favorably compared to SW and its variants.", "keywords": [ "projection optimal transport", "optimal transport" ], "rating_list": [ 8, 5, 5, 6 ], "soundness_list": [ 4, 2, 3, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "EKJhH5D5wA", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Heming Xia", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "", "position": "PhD student" }, { "name": "Jun Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Wenjie Li", "gender": "Female", "institution": "The Hong Kong Polytechnic University, The Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Yongqi Li", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "cunxiao du", "gender": "Unspecified", "institution": "Sea AI LAB", "country": "SG", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 6, "n_ref_uni": 39, "n_ref": 127, "n_ref_all": 172, "n_fig": 11, "n_tab": 8, "L_tab": 1779, "n_element_tab": 72, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 524, "n_element_tab_1": 33, "formula_len_all": 897, "formula_len_all_1": 321, "len_all": 211632, "len_all_1": 68202, "len_abs": 1409, "len_title": 123, "len_sents": 60805, "len_sents_1": 33241, "n_sents": 444, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1416, "title": "SWIFT: On-the-Fly Self-Speculative Decoding for LLM Inference Acceleration", "abs": "Speculative decoding (SD) has emerged as a widely used paradigm to accelerate the inference of large language models (LLMs) without compromising generation quality. It works by first employing a compact model to draft multiple tokens efficiently and then using the target LLM to verify them in parallel. While this technique has achieved notable speedups, most existing approaches necessitate either additional parameters or extensive training to construct effective draft models, thereby restricting their applicability across different LLMs and tasks. To address this limitation, we explore a novel plug-and-play SD solution with layer-skipping, which skips intermediate layers of the target LLM as the compact draft model. Our analysis reveals that LLMs exhibit great potential for self-acceleration through layer sparsity and the task-specific nature of this sparsity. Building on these insights, we introduce SWIFT, an on-the-fly self-speculative decoding algorithm that adaptively selects intermediate layers of LLMs to skip during inference. SWIFT does not require auxiliary models or additional training, making it a plug-and-play solution for accelerating LLM inference across diverse input data streams. Our extensive experiments across a wide range of models and downstream tasks demonstrate that SWIFT can achieve over a 1.3x-1.6x speedup while preserving the original distribution of the generated text.", "keywords": [ "Speculative Decoding", "LLM Inference Acceleration", "Efficient NLP" ], "rating_list": [ 6, 5, 6, 8 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "EKCubxFdOs", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Hongshu Guo", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Jiacheng Chen", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yining Ma", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yue-Jiao Gong", "gender": "Female", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zeyuan Ma", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Zhiguang Cao", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 5, "n_ref_uni": 88, "n_ref": 153, "n_ref_all": 180, "n_fig": 8, "n_tab": 8, "L_tab": 3025, "n_element_tab": 341, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2554, "n_element_tab_1": 347, "formula_len_all": 798, "formula_len_all_1": 313, "len_all": 264290, "len_all_1": 77263, "len_abs": 1163, "len_title": 135, "len_sents": 66350, "len_sents_1": 36480, "n_sents": 446, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1169, "title": "LLaMoCo: Instruction Tuning of Large Language Models for Optimization Code Generation", "abs": "Recent research on optimization using large language models (LLMs) typically involves either iterative next-step solution seeking or directly prompting LLMs to generate critical optimization codes. However, these methods often suffer from low computational efficiency, high sensitivity to prompt design, and a lack of domain-specific knowledge. We introduce LLaMoCo, the first instruction-tuning framework designed to adapt LLMs for solving optimization problems in a code-to-code manner. LLaMoCo features a comprehensive instruction set that includes code-style problem descriptions as input prompts and robust optimization codes from expert optimizers as target outputs. We then develop a novel two-phase learning strategy with a contrastive learning-based warm-up to enhance convergence during instruction tuning. Extensive experiments demonstrate that a CodeGen (350M) model tuned by our LLaMoCo yields a powerful domain-specific model for generating expert-level optimizers, achieving superior performance compared to GPT-4 Turbo and other competitors on both synthetic and realistic problem sets. The trained model and the usage instructions are available online.", "keywords": [ "Large Language Models", "Instruction Tuning", "Optimization Code Generation" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "EK1yOLL7GA", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Dong Gong", "gender": "Male", "institution": "University of New South Wales", "country": "AU", "position": "Assistant Professor" }, { "name": "Sen Wang", "gender": "Male", "institution": "The University of Queensland", "country": "AU", "position": "Associate Professor" }, { "name": "Yadan Luo", "gender": "Female", "institution": "The University of Queensland", "country": "AU", "position": "Associate Professor" }, { "name": "Zi Huang", "gender": "Female", "institution": "University of Queensland", "country": "AU", "position": "Full Professor" }, { "name": "Zixin Wang", "gender": "Female", "institution": "University of Queensland", "country": "AU", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 17, "n_ref_uni": 61, "n_ref": 97, "n_ref_all": 130, "n_fig": 9, "n_tab": 6, "L_tab": 1584, "n_element_tab": 237, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 799, "n_element_tab_1": 131, "formula_len_all": 1420, "formula_len_all_1": 1397, "len_all": 159929, "len_all_1": 71477, "len_abs": 1550, "len_title": 122, "len_sents": 41807, "len_sents_1": 32008, "n_sents": 307, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 74, "L_abs": 1561, "title": "Tokens on Demand: Token Condensation as Training-free Test-time Adaptation", "abs": "In this work, we introduce Token Condensation as Adaptation (TCA), a training-free approach designed to mitigate distribution shifts encountered by vision-language models (VLMs) during test-time inference. TCA bridges distribution gaps at the patch level by condensing image tokens that exhibit low attentiveness to the token. Recognizing the token may correspond to universal concepts, TCA identifies and tracks the most reliable tokens that align specifically with target classes from historical data streams. To achieve this, we propose a context token reservoir (CTR), which retains tokens with the lowest uncertainty as ``anchors\" to guide the preservation of class-relevant tokens during inference. These anchors, in turn, act as token-level classifiers to correct VLM predictions and improve visual-text alignment. Utilizing anchors sampled from CTR, TCA condenses tokens through two operations: (1) pruning class-irrelevant tokens that consistently rank low across all attention heads to reach cross-head consensus on their irrelevance, and (2) merging the remaining class-ambiguous tokens into representative centers using coreset selection, maintaining linear computational complexity. As the first method to explore token efficiency in test-time adaptation, TCA consistently demonstrates superior performance across cross-dataset and out-of-distribution adaptation tasks, reducing GFLOPs by 12.2\\% to 48.9\\% while achieving accuracy improvements up to 21.4\\% against the strongest baseline without introducing additional parameters.", "keywords": [ "test-time adaptation" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 4, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "EJgxMsiAO9", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Jenia Jitsev", "gender": "unknown", "institution": "Juelich Supercomputing Center, Research Center Juelich", "country": "", "position": "Senior Scientist" }, { "name": "Lucia Cipolina-Kun", "gender": "Female", "institution": "University of Bristol", "country": "", "position": "PHD" }, { "name": "Marianna Nezhurina", "gender": "unknown", "institution": "LAION", "country": "", "position": "Researcher" }, { "name": "mehdi cherti", "gender": "Male", "institution": "Forschungszentrum J\u00fclich", "country": "", "position": "Postdoc" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 31, "n_ref": 94, "n_ref_all": 151, "n_fig": 27, "n_tab": 9, "L_tab": 2924, "n_element_tab": 312, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 475, "n_element_tab_1": 67, "formula_len_all": 108, "formula_len_all_1": 0, "len_all": 281732, "len_all_1": 67818, "len_abs": 1886, "len_title": 181, "len_sents": 83670, "len_sents_1": 33400, "n_sents": 677, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Reject", "L_title": 133, "L_abs": 2030, "title": "Alice in Wonderland: Simple Tasks Reveal Severe Generalization and Basic Reasoning Deficits in State-Of-the-Art Large Language Models", "abs": "Large Language Models (LLMs) are often described as being instances of foundation models - that is, models that possess strong generalization and therefore transfer robustly across various tasks and conditions in few-show or zero-shot manner, while exhibiting scaling laws that predict generalization improvement when increasing the pre-training scale. These claims of strong generalization and advanced reasoning function enabling it rely on measurements by various standardized benchmarks where state-of-the-art (SOTA) models score high. We demonstrate here a dramatic breakdown of generalization and basic reasoning of all SOTA models which claim strong function, including advanced models like GPT-4 or Claude 3 Opus trained at the largest scales, using a simple, short common sense problem formulated in concise natural language, easily solvable by humans (AIW problem). The breakdown is dramatic as it manifests in both low average performance and strong performance fluctuations on natural problem variations that change neither problem structure nor its difficulty, while also often expressing strong overconfidence in the wrong solutions, backed up by plausible sounding explanation-like confabulations. Various standard interventions in an attempt to get the right solution, like chain-of-thought prompting, or urging the models to reconsider the wrong solutions again by multi step re-evaluation, fail. We take these observations to the scientific and technological community to stimulate re-assessment of the capabilities of current generation of LLMs as claimed by standardized benchmarks. Such re-assessment also requires common action to create standardized benchmarks that would allow proper detection of such deficits in generalization and reasoning that obviously remain undiscovered by current state-of-the-art evaluation procedures, where SOTA LLMs obtain high scores. Code for reproducing experiments in the paper and raw experiments data can be found at https://anonymous.4open.science/r/AITW_anonymous-69A6/", "keywords": [ "large language models", "foundation models", "generalization", "reasoning", "function testing", "evaluation", "benchmarks", "robustness", "function breakdown" ], "rating_list": [ 6, 6, 3, 5, 6 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 1, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "EJfLvrzh2Q", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Hye Won Chung", "gender": "Female", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "Associate Professor" }, { "name": "Hyeonsu Jeong", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "PhD student" } ], "n_formula": 209, "n_formula_1": 35, "n_ref_uni": 32, "n_ref": 50, "n_ref_all": 93, "n_fig": 15, "n_tab": 53, "L_tab": 26368, "n_element_tab": 1080, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 19347, "formula_len_all_1": 3154, "len_all": 335656, "len_all_1": 74140, "len_abs": 1149, "len_title": 149, "len_sents": 83661, "len_sents_1": 33440, "n_sents": 1178, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 100, "L_abs": 1157, "title": "Rethinking Self-Distillation: Label Averaging and Enhanced Soft Label Refinement with Partial Labels", "abs": "We investigate the mechanisms of self-distillation in multi-class classification, particularly in the context of linear probing with fixed feature extractors where traditional feature learning explanations do not apply. Our theoretical analysis reveals that multi-round self-distillation effectively performs label averaging among instances with high feature correlations, governed by the eigenvectors of the Gram matrix derived from input features. This process leads to clustered predictions and improved generalization, mitigating the impact of label noise by reducing the model's reliance on potentially corrupted labels. We establish conditions under which multi-round self-distillation achieves 100\\% population accuracy despite label noise. Furthermore, we introduce a novel, efficient single-round self-distillation method using refined partial labels from the teacher's top two softmax outputs, referred to as the PLL student model. This approach replicates the benefits of multi-round distillation in a single round, achieving comparable or superior performance--especially in high-noise scenarios--while significantly reducing computational cost.", "keywords": [ "self-distillation", "partial label learning", "label noise correction", "training with soft labels", "multi-class classification" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "EJTeOf8iG0", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Adriano Jose Tavares", "gender": "Male", "institution": "Universidade do Minho", "country": "PT", "position": "Associate Professor" }, { "name": "Hao Xu", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "Full Professor" }, { "name": "Tiago Gomes", "gender": "Male", "institution": "Universidade do Minho", "country": "", "position": "Assistant Professor" }, { "name": "Xue Gu", "gender": "Female", "institution": "Universidade do Minho", "country": "PT", "position": "PhD student" }, { "name": "Ziyao Meng", "gender": "Male", "institution": "Universidade do Minho", "country": "", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 23, "n_ref": 50, "n_ref_all": 64, "n_fig": 5, "n_tab": 6, "L_tab": 6557, "n_element_tab": 420, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2270, "n_element_tab_1": 289, "formula_len_all": 274, "formula_len_all_1": 215, "len_all": 164208, "len_all_1": 63757, "len_abs": 2203, "len_title": 97, "len_sents": 32199, "len_sents_1": 27249, "n_sents": 218, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 98, "L_abs": 1019, "title": "EEEC: Emotion-Experiencer-Event-Cause multi-step chain reasoning for Emotion-Cause Pair Extraction", "abs": "Emotion-cause pair extraction (ECPE) aims to identify all emotion and cause clauses in documents, forming the ECPs. Although existing methods have achieved some success, they face issues such as overlooking the impact of emotion experiencers, failing to leverage specific domain knowledge, and tending to spurious correlations. To address these issues, we transform the ECPE task into a multi-step reasoning problem and propose the Emotion-Experience-Event-Cause (EEEC) framework. We introduce an experiencer identification task to understand the source of emotions and enhance the association between emotion and cause clauses. In addition, by combining both prior knowledge and induced reasoning, EEEC guides a large-scale language model (LLM) to perform the emotion-reason pair extraction task efficiently. Experimental results demonstrate that EEEC achieves performance close to current state-of-the-art supervised fine-tuning methods. The data and code are released at https://anonymous.4open.science/r/EEEC-EB80/.", "keywords": [ "Experiencer; Event; Multi-step chain reasoning; Emotion-Cause Pair Extraction" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "EIwGR0w8VG", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Christian Helms", "gender": "Male", "institution": "Hasso Plattner Institute", "country": "DE", "position": "MS student" }, { "name": "Ralf Herbrich", "gender": "Male", "institution": "Hasso Plattner Institute", "country": "DE", "position": "Full Professor" }, { "name": "Romeo Sommerfeld", "gender": "Male", "institution": "Hasso Plattner Institute", "country": "DE", "position": "MS student" } ], "n_formula": 65, "n_formula_1": 17, "n_ref_uni": 26, "n_ref": 43, "n_ref_all": 64, "n_fig": 4, "n_tab": 3, "L_tab": 310, "n_element_tab": 47, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 245, "n_element_tab_1": 35, "formula_len_all": 7260, "formula_len_all_1": 1269, "len_all": 151096, "len_all_1": 67217, "len_abs": 1323, "len_title": 113, "len_sents": 50652, "len_sents_1": 33017, "n_sents": 421, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1327, "title": "Scalable Approximate Message Passing for Bayesian Neural Networks", "abs": "Bayesian neural networks (BNNs) offer the potential for reliable uncertainty quantification and interpretability, which are critical for trustworthy AI in high-stakes domains. However, existing methods often struggle with issues such as overconfidence, hyperparameter sensitivity, and posterior collapse, leaving room for alternative approaches. In this work, we advance message passing (MP) for BNNs and present a novel framework that models the predictive posterior as a factor graph. To the best of our knowledge, our framework is the first MP method that handles convolutional neural networks and avoids double-counting training data, a limitation of previous MP methods that causes overconfidence. We evaluate our approach on CIFAR-10 with a convolutional neural network of roughly 890k parameters and find that it can compete with the SOTA baselines AdamW and IVON, even having an edge in terms of calibration. On synthetic data, we validate the uncertainty estimates and observe a strong correlation (0.9) between posterior credible intervals and its probability of covering the true data-generating function outside the training range. While our method scales to an MLP with 5.6 million parameters, further improvements are necessary to match the scale and performance of state-of-the-art variational inference methods.", "keywords": [ "Message Passing", "Bayesian Neural Networks", "Uncertainty Estimation", "Factor Graphs" ], "rating_list": [ 6, 6, 3 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "EIXZXPz7jU", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Fedor Buzaev", "gender": "Male", "institution": "Higher School of Economics", "country": "RU", "position": "PhD student" }, { "name": "Jiexing Gao", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Principal Researcher" }, { "name": "Yana Khassan Nibal", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "RU", "position": "Researcher" }, { "name": "Yimin Huang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" } ], "n_formula": 30, "n_formula_1": 21, "n_ref_uni": 12, "n_ref": 21, "n_ref_all": 31, "n_fig": 7, "n_tab": 3, "L_tab": 2987, "n_element_tab": 49, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 2657, "n_element_tab_1": 34, "formula_len_all": 1745, "formula_len_all_1": 1136, "len_all": 83003, "len_all_1": 46980, "len_abs": 1446, "len_title": 165, "len_sents": 25283, "len_sents_1": 19844, "n_sents": 191, "n_sents_1": 133, "venue": "ICLR2025", "decision": "Reject", "L_title": 115, "L_abs": 1452, "title": "FMS PINN: Flow-matching sampling for efficient solution of partial differential equations with source singularities", "abs": "Singularities in the source functions of partial differential equations (PDEs) can pose significant challenges for physics-informed neural networks (PINNs), often leading to numerical instability and necessitating a large number of sampling points thereby increasing the computational time. In this paper, we introduce a novel sampling point selection method to address these challenges. Our approach is based on diffusion models capable of generative sampling from the distribution of PDE residuals. Specifically, we apply the optimal transport coupling flow-matching technique to generate more sampling points in regions where the PDE residuals are higher, enhancing the accuracy and efficiency of the solution. In contrast to existing approaches in the literature, our method avoids explicit modeling of the probability density proportional to residuals, instead using the benefits of flow matching to generate novel and probable samples from more complex distributions, thereby enhancing PINN solutions for problems with singularities.\nWe demonstrate that this method, in certain scenarios, outperforms existing techniques such as normalizing flow-based sampling PINN. Especially, our approach demonstrates effectiveness in improving the solution quality for the linear elasticity equation in the case of material with complex geometry of inclusion. A detailed comparison of the flow matching sampling method with other approaches is also provided.", "keywords": [ "physics informed neural networks", "Adaptive sampling" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 2, 2, 4 ] }, { "paper_id": "EHmjRIA4l2", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Rajesh P. N. Rao", "gender": "Male", "institution": "University of Washington", "country": "", "position": "Full Professor" }, { "name": "Vishwas Sathish", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 46, "n_ref": 70, "n_ref_all": 100, "n_fig": 11, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 463, "formula_len_all_1": 483, "len_all": 133930, "len_all_1": 58344, "len_abs": 1652, "len_title": 106, "len_sents": 35274, "len_sents_1": 26754, "n_sents": 275, "n_sents_1": 220, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1660, "title": "Compositional World Models with Interpretable Abstractions", "abs": "We present a modular and compositional approach to learning human-aligned world models via state-action hierarchies. Our approach is inspired by sensory-motor hierarchies in the mammalian brain. We model complex state transition dynamics as a sequence of simpler dynamics, which in turn can be modeled using even simpler dynamics, and so on, endowing the approach with rich compositionality. We introduce Composer, a practical method for learning complex world models that leverages hypernetworks and abstract states for generating lower-level transition functions on-the-fly. We first show that state abstractions in Composer emerge naturally in simple environments as a consequence of training. Incorporating a variant of contrastive learning allows Composer to scale to more complex environments while ensuring that the learned abstractions are human aligned. Additionally, learning a higher-level transition function between learned abstract states leads to a hierarchy of transition functions for modeling complex dynamics. We apply Composer to compositional navigation problems and show its capability for rapid planning and transfer to novel scenarios. In both traditional grid-world navigation problems as well as in the more complex Habitat vision-based navigation domain, a Composer-based agent learns to model the state-action dynamics within and between different rooms using a hierarchy of transition functions and leverage this hierarchy for efficient downstream planning. Our results suggest that Composer offers a promising framework for learning the complex dynamics of real-world environments using a compositional and interpretable approach.", "keywords": [ "State-Action Abstractions", "Predictive Coding", "Hierarchical Planning", "Compositional World Models", "Contrastive Learning", "Hypernetworks", "Hierarchical Reinforcement Learning" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 1, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "EHhLLmDvtE", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Chak Tou Leong", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Kaishuai Xu", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Liangyou Li", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "", "position": "Researcher" }, { "name": "Lifeng Shang", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Qun Liu", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "HK", "position": "Chief Scientist of Speech and Language Computing" }, { "name": "Tiezheng YU", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Wenjie Li", "gender": "Female", "institution": "The Hong Kong Polytechnic University, The Hong Kong Polytechnic University", "country": "HK", "position": "Full Professor" }, { "name": "Wenjun Hou", "gender": "Male", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" }, { "name": "Xin Jiang", "gender": "Male", "institution": "Noah\u2019s Ark Lab, Huawei Technologies", "country": "HK", "position": "Principal Researcher" }, { "name": "Yi Cheng", "gender": "Female", "institution": "The Hong Kong Polytechnic University", "country": "HK", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 6, "n_ref_uni": 22, "n_ref": 82, "n_ref_all": 116, "n_fig": 5, "n_tab": 16, "L_tab": 3774, "n_element_tab": 358, "n_fig_1": 5, "n_tab_1": 8, "L_tab_1": 2120, "n_element_tab_1": 311, "formula_len_all": 1191, "formula_len_all_1": 793, "len_all": 166072, "len_all_1": 66354, "len_abs": 2891, "len_title": 121, "len_sents": 41695, "len_sents_1": 27025, "n_sents": 377, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1753, "title": "Subtle Errors Matter: Preference Learning via Error-injected Self-editing", "abs": "Large Language Models (LLMs) have exhibited strong mathematical reasoning and computational prowess, tackling tasks ranging from basic arithmetic to advanced competition-level problems. However, frequently occurring subtle errors, such as miscalculations or incorrect substitutions, limit the models\u2019 full mathematical potential. Existing studies to improve mathematical ability typically involve distilling reasoning skills from stronger LLMs or applying preference learning to step-wise response pairs. Although these methods leverage samples of varying granularity to mitigate reasoning errors, they overlook the frequently occurring subtle errors. A major reason is that sampled preference pairs involve differences unrelated to the errors, which may distract the model from focusing on subtle errors. In this work, we propose a novel preference learning framework called eRror-Injected Self-Editing (RISE), which injects predefined subtle errors into partial tokens of correct solutions to construct hard pairs for error mitigation. In detail, RISE uses the model itself to edit a small number of tokens in the solution, injecting designed subtle errors. Then, pairs composed of self-edited solutions and their corresponding correct ones, along with pairs of correct and incorrect solutions obtained through sampling, are used together for subtle error-aware DPO training. Compared with other preference learning methods, RISE further refines the training objective to focus on predefined errors and their tokens, without requiring fine-grained sampling or preference annotation. Extensive experiments validate the effectiveness of RISE, with preference learning on Qwen2-7B-Instruct yielding notable improvements of 3.0% on GSM8K and 7.9% on MATH.", "keywords": [ "Mathematical Reasoning", "Preference Learning" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "EHfn5fbFHw", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Fatemeh Fard", "gender": "Female", "institution": "University of British Columbia", "country": "CA", "position": "Assistant Professor" }, { "name": "Iman Saberi", "gender": "Male", "institution": "University of British Columbia, Okanagan", "country": "CA", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 24, "n_ref": 66, "n_ref_all": 111, "n_fig": 10, "n_tab": 6, "L_tab": 3282, "n_element_tab": 312, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 450, "n_element_tab_1": 91, "formula_len_all": 384, "formula_len_all_1": 417, "len_all": 139269, "len_all_1": 63079, "len_abs": 1426, "len_title": 116, "len_sents": 45465, "len_sents_1": 29939, "n_sents": 331, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1437, "title": "Context-Augmented Code Generation Using Programming Knowledge Graphs", "abs": "Large Language Models (LLMs) and Code-LLMs (CLLMs) have significantly improved code generation, but, they frequently face difficulties when dealing with challenging and complex problems. Retrieval-Augmented Generation (RAG) addresses this issue by retrieving and integrating external knowledge at the inference time. However, retrieval models often fail to find most relevant context, and generation models, with limited context capacity, can hallucinate when given irrelevant data. We present a novel framework that leverages a Programming Knowledge Graph (PKG) to semantically represent and retrieve code. This approach enables fine-grained code retrieval by focusing on the most relevant segments while reducing irrelevant context through a tree-pruning technique. PKG is coupled with a re-ranking mechanism to reduce even more hallucinations by selectively integrating non-RAG solutions. We propose two retrieval approaches\u2014block-wise and function- wise\u2014based on the PKG, optimizing context granularity. Evaluations on the HumanEval and MBPP benchmarks show our method improves pass@1 accuracy by up to 20\\%, and outperforms state-of-the-art models by up to 34\\% on MBPP. Our contributions include PKG-based retrieval, tree pruning to enhance retrieval precision, a re-ranking method for robust solution selection and a Fill-in-the- Middle (FIM) enhancer module for automatic code augmentation with relevant comments and docstrings.", "keywords": [ "Code Generation", "RAG", "Knowledge Graphs", "Large Language Models" ], "rating_list": [ 5, 5, 3, 8 ], "soundness_list": [ 3, 3, 1, 4 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "EHYbqCDRtM", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Jiale Liu", "gender": "Male", "institution": "Huazhong Agricultural University", "country": "CN", "position": "Undergrad student" }, { "name": "Lu Li", "gender": "unknown", "institution": "Huazhong Agricultural University", "country": "CN", "position": "MS student" }, { "name": "Maojun Wang", "gender": "Male", "institution": "Huazhong Agricultural University", "country": "CN", "position": "Full Professor" }, { "name": "Xingyu Ji", "gender": "Male", "institution": "Huazhong Agricultural University", "country": "CN", "position": "Undergrad student" }, { "name": "ZEYU ZHANG", "gender": "Male", "institution": "Huazhong Agricultural University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 18, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 62, "n_ref_all": 75, "n_fig": 3, "n_tab": 7, "L_tab": 1284, "n_element_tab": 104, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 653, "n_element_tab_1": 101, "formula_len_all": 932, "formula_len_all_1": 401, "len_all": 536004, "len_all_1": 69104, "len_abs": 1625, "len_title": 184, "len_sents": 288902, "len_sents_1": 34565, "n_sents": 1366, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 136, "L_abs": 1496, "title": "Verbalized Graph Representation Learning: A Fully Interpretable Graph Model Based on Large Language Models Throughout the Entire Process", "abs": "Representation learning on text-attributed graphs (TAGs) has attracted significant interest due to its wide-ranging real-world applications, particularly through Graph Neural Networks (GNNs). Traditional GNN methods focus on encoding the structural information of graphs, often using shallow text embeddings for node or edge attributes. This limits the model to understand the rich semantic information in the data and its reasoning ability for complex downstream tasks, while also lacking interpretability. With the rise of large language models (LLMs), an increasing number of studies are combining them with GNNs for graph representation learning and downstream tasks. While these approaches effectively leverage the rich semantic information in TAGs datasets, their main drawback is that they are only partially interpretable, which limits their application in critical fields. In this paper, we propose a verbalized graph representation learning (VGRL) method which is fully interpretable. In contrast to traditional graph machine learning models, which are usually optimized within a continuous parameter space, VGRL constrains this parameter space to be text description which ensures complete interpretability throughout the entire process, making it easier for users to understand and trust the decisions of the model. We conduct several studies to empirically evaluate the effectiveness of VGRL and we believe these method can serve as a stepping stone in graph representation learning.", "keywords": [ "large language models", "fully interpretable", "graph representation learning" ], "rating_list": [ 3, 1, 3, 1 ], "soundness_list": [ 1, 1, 2, 1 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 1, 1, 2, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "EGxgZzDODh", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Fengsong Sun", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Jinyu Wang", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "PhD student" }, { "name": "Xianchao Zhang", "gender": "Male", "institution": "Jiaxing University", "country": "", "position": "Full Professor" }, { "name": "Zhiqing Wei", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "", "position": "Associate Professor" } ], "n_formula": 24, "n_formula_1": 22, "n_ref_uni": 38, "n_ref": 66, "n_ref_all": 81, "n_fig": 2, "n_tab": 4, "L_tab": 1061, "n_element_tab": 154, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 987, "n_element_tab_1": 11, "formula_len_all": 1374, "formula_len_all_1": 1383, "len_all": 108328, "len_all_1": 58828, "len_abs": 1066, "len_title": 113, "len_sents": 30910, "len_sents_1": 27716, "n_sents": 240, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 65, "L_abs": 1061, "title": "Neural Probabilistic Logic Learning for Knowledge Graph Reasoning", "abs": "Knowledge graph (KG) reasoning is a task that aims to predict unknown facts based on known factual samples. Reasoning methods can be divided into two categories: rule-based methods and KG-embedding based methods. The former possesses precise reasoning capabilities but finds it challenging to reason efficiently over large-scale knowledge graphs. While gaining the ability to reason over large-scale knowledge graphs, the latter sacrifices reasoning accuracy. This paper aims to design a reasoning framework called Neural Probabilistic Logic Learning(NPLL) that achieves accurate reasoning on knowledge graphs. Our approach introduces a scoring module that effectively enhances the expressive power of embedding networks. We strike a balance between model simplicity and reasoning capabilities by incorporating a Markov Logic Network based on variational inference. We empirically evaluate our approach on several benchmark datasets, and the experimental results validate that our method substantially enhances the accuracy and quality of the reasoning results.", "keywords": [ "Knowledge graph reasoning", "embedding", "rule-based", "variational inference" ], "rating_list": [ 1, 3, 3, 5 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 1, 1, 1, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "EG9nDN3eGB", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 8, "author_info_list": [ { "name": "Feng Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jianye HAO", "gender": "Male", "institution": "Tianjin University", "country": "CN", "position": "Associate Professor" }, { "name": "Jie Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Lei Chen", "gender": "unknown", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Mingxuan Yuan", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "HK", "position": "Researcher" }, { "name": "Yinqi Bai", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Undergrad student" }, { "name": "Yufei Kuang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Zhihai Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 3, "n_ref_uni": 36, "n_ref": 95, "n_ref_all": 152, "n_fig": 10, "n_tab": 13, "L_tab": 6862, "n_element_tab": 989, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1607, "n_element_tab_1": 173, "formula_len_all": 374, "formula_len_all_1": 135, "len_all": 204624, "len_all_1": 69969, "len_abs": 1797, "len_title": 127, "len_sents": 70182, "len_sents_1": 34079, "n_sents": 484, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1870, "title": "A Graph Enhanced Symbolic Discovery Framework For Efficient Circuit Synthesis", "abs": "The efficiency of Circuit Synthesis (CS) has become one of the key bottlenecks in chip design. To prompt efficient CS, previous studies propose using a key scoring function to predict and prune a large number of ineffective nodes of the CS heuristics. \nHowever, the existing scoring functions struggle to balance inference efficiency, interpretability, and\ngeneralization performance, which severely hinders their application to modern CS tools. To address this challenge, we propose a novel data-driven circuit symbolic learning framework, namely CMO, to learn lightweight, interpretable, and generalizable scoring functions.\nThe major challenge of developing CMO is to discover symbolic functions that can well generalize to unseen circuits, i.e., the circuit symbolic generalization problem. Thus, the major technical contribution of CMO is the novel $\\textit{Graph Enhanced Symbolic Discovery} $ framework, which distills dark knowledge from a well-designed graph neural network (GNN) to enhance the generalization capability of the learned symbolic functions. To the best of our knowledge, CMO is $\\textit{the first}$ graph-enhanced approach for discovering lightweight and interpretable symbolic functions that can well generalize to unseen circuits in CS. Experiments on three challenging circuit benchmarks show that the $\\textit{interpretable}$ symbolic functions learned by CMO outperform previous state-of-the-art (SOTA) GPU-based and human-designed approaches in terms of $\\textit{inference efficiency}$ and $\\textit{generalization capability}$. Moreover, we integrate CMO with the Mfs2 heuristic---one of the most time-consuming CS heuristics. \nThe empirical results demonstrate that CMO significantly improves its efficiency while keeping comparable optimization performance when executed on a CPU-based machine, achieving up to 2.5$\\times$ faster runtime.", "keywords": [ "Chip Design", "Circuit Synthesis", "Symbolic Regression", "Knowledge Distillation" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "EFzBhrEp8Y", "primary_area": "datasets and benchmarks", "n_author": 12, "author_info_list": [ { "name": "Chaoyou Fu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Dong Zhang", "gender": "unknown", "institution": "Soochow University", "country": "CN", "position": "MS student" }, { "name": "Haipang WU", "gender": "Male", "institution": "Hithink RoyalFlush Information Network Co.,Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Haohan Li", "gender": "Male", "institution": "Hangzhou Dianzi University", "country": "CN", "position": "MS student" }, { "name": "Ji Liu", "gender": "Male", "institution": "Hithink RoyalFlush Information Network Co., Ltd.", "country": "CN", "position": "Technical Director" }, { "name": "Jian Liu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Rongjunchen Zhang", "gender": "Male", "institution": "Hithink RoyalFlush Information Network Co.,Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Yong Dai", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Yu Lu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Postdoc" }, { "name": "Ziliang Gan", "gender": "Male", "institution": "Hithink RoyalFlush Information Network Co.,Ltd.", "country": "CN", "position": "Researcher" }, { "name": "che liu", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Zenglin Xu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 35, "n_ref": 57, "n_ref_all": 72, "n_fig": 6, "n_tab": 5, "L_tab": 3500, "n_element_tab": 224, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 3527, "n_element_tab_1": 103, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 125601, "len_all_1": 52057, "len_abs": 1323, "len_title": 89, "len_sents": 28211, "len_sents_1": 22169, "n_sents": 235, "n_sents_1": 160, "venue": "ICLR2025", "decision": "withdraw", "L_title": 88, "L_abs": 1330, "title": "MME-FINANCE: A Multimodal Finance Benchmark for Expert-level Understanding and Reasoning", "abs": "The remarkable capability of existing Multimodal Large Language Models~(MLLMs) to understand general natural images have been extensively demonstrated in plentiful benchmarks. Nevertheless, the potential of MLLMs in finance domain remains to be fully explored. Financial images exhibit a wide range of variations, encompass intricate details, and demand professional expertise for proper interpretation, thereby posing a significant challenge for MLLMs in terms of their fine-grained perception and complex reasoning capabilities. To bridge this gap, we introduce MME-FINANCE, a novel benchmark designed specifically to assess MLLMs' performance in open-ended financial Visual Question Answering (VQA). Our benchmark consists of over 1,000 VQA pairs spanning a wide range of complex financial scenarios. We devise multi-tiered financial tasks tailored to the specific characteristics of the financial domain, aiming to comprehensively evaluate the perception, reasoning, and cognition capabilities of MLLMs. \nFurthermore, we employ a multimodal evaluation approach that incorporates visual data to score the model predictions, thereby aligning more closely with human judgment. Extensive experimental evaluations of 18 mainstream MLLMs reveal their limitations in financial tasks and provide insights to inspire further research.", "keywords": [ "Multimodal; Benchmark" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "EFhzmn3RJG", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Dongwoo Kim", "gender": "Male", "institution": "POSTECH", "country": "KR", "position": "Assistant Professor" }, { "name": "MoonJeong Park", "gender": "unknown", "institution": "POSTECH", "country": "", "position": "MS student" } ], "n_formula": 31, "n_formula_1": 17, "n_ref_uni": 26, "n_ref": 42, "n_ref_all": 61, "n_fig": 10, "n_tab": 1, "L_tab": 697, "n_element_tab": 65, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 522, "n_element_tab_1": 22, "formula_len_all": 4060, "formula_len_all_1": 1047, "len_all": 108635, "len_all_1": 59674, "len_abs": 1138, "len_title": 117, "len_sents": 34292, "len_sents_1": 27286, "n_sents": 310, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1167, "title": "Taming Gradient Oversmoothing and Expansion in Graph Neural Networks", "abs": "Oversmoothing has been claimed as a primary bottleneck for multi-layered graph neural networks (GNNs). Multiple analyses have examined how and why oversmoothing occurs. However, none of the prior work addressed how optimization is performed under the oversmoothing regime. In this work, we show the presence of $\\textit{gradient oversmoothing}$ preventing optimization during training. We further analyze that GNNs with residual connections, a well-known solution to help gradient flow in deep architecture, introduce $\\textit{gradient expansion}$, a phenomenon of the gradient explosion in diverse directions. Therefore, adding residual connections cannot be a solution for making a GNN deep. Our analysis reveals that constraining the Lipschitz bound of each layer can neutralize the gradient expansion. To this end, we provide a simple yet effective normalization method to prevent the gradient expansion. An empirical study shows that the residual GNNs with hundreds of layers can be efficiently trained with the proposed normalization without compromising performance. Additional studies show that the empirical observations corroborate our theoretical analysis.", "keywords": [ "graph neural network", "deep neural network", "oversmoothing", "optimization" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 4, 2, 1 ], "contribution_list": [ 2, 3, 1, 1 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "EFZEdHB3Mp", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 9, "author_info_list": [ { "name": "Jack Hong", "gender": "unknown", "institution": "Xiaohongshu", "country": "CN", "position": "Intern" }, { "name": "Jiayi Shen", "gender": "Female", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Jiayin Cai", "gender": "Female", "institution": "Kuaishou", "country": "CN", "position": "Researcher" }, { "name": "Shilin Yan", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Xiaolong Jiang", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Enigneer" }, { "name": "Yao Hu", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Principal Researcher" }, { "name": "Zehao Xiao", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Cees Snoek", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "Qi Wang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" } ], "n_formula": 16, "n_formula_1": 14, "n_ref_uni": 41, "n_ref": 116, "n_ref_all": 157, "n_fig": 9, "n_tab": 6, "L_tab": 3212, "n_element_tab": 303, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2281, "n_element_tab_1": 105, "formula_len_all": 852, "formula_len_all_1": 770, "len_all": 151174, "len_all_1": 65923, "len_abs": 1159, "len_title": 91, "len_sents": 42525, "len_sents_1": 30986, "n_sents": 334, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 1164, "title": "DynaPrompt: Dynamic Test-Time Prompt Tuning", "abs": "Test-time prompt tuning enhances zero-shot generalization of vision-language models but tends to ignore the relatedness among test samples during inference. Online test-time prompt tuning provides a simple way to leverage the information in previous test samples, albeit with the risk of prompt collapse due to error accumulation. To enhance test-time prompt tuning, we propose DynaPrompt, short for dynamic test-time prompt tuning, exploiting relevant data distribution information while reducing error accumulation. Built on an online prompt buffer, DynaPrompt adaptively selects and optimizes the relevant prompts for each test sample during tuning. Specifically, we introduce a dynamic prompt selection strategy based on two metrics: prediction entropy and probability difference. For unseen test data information, we develop dynamic prompt appending, which allows the buffer to append new prompts and delete the inactive ones. By doing so, the prompts are optimized to exploit beneficial information on specific test data, while alleviating error accumulation. Experiments on fourteen datasets demonstrate the effectiveness of dynamic test-time prompt tuning.", "keywords": [ "Test-time prompt tuning; test-time adaptation; vision-language model; CLIP" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 5, 3, 5, 4 ] }, { "paper_id": "EEgYUccwsV", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Dunjie Lu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Intern" }, { "name": "Junli Wang", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Intern" }, { "name": "Tao Yu", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Yiheng Xu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yuchen Mao", "gender": "Female", "institution": "Ludwig-Maximilians-Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "MS student" }, { "name": "Zekun Wang", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Zhennan Shen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 26, "n_ref": 53, "n_ref_all": 68, "n_fig": 8, "n_tab": 11, "L_tab": 2787, "n_element_tab": 254, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 737, "n_element_tab_1": 80, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 138353, "len_all_1": 51995, "len_abs": 1351, "len_title": 129, "len_sents": 41739, "len_sents_1": 23550, "n_sents": 326, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 75, "L_abs": 1356, "title": "AgentTrek: Agent Trajectory Synthesis via Guiding Replay with Web Tutorials", "abs": "Graphical User Interface (GUI) agents hold great potential for automating complex tasks across diverse digital environments, from web applications to desktop software. However, the development of such agents is hindered by the lack of high-quality, multi-step trajectory data required for effective training. Existing approaches rely on expensive and labor-intensive human annotation, making them unsustainable at scale. To address this challenge, we propose \\ourwork, a scalable data synthesis pipeline that generates high-quality GUI agent trajectories by leveraging web tutorials. Our method automatically gathers tutorial-like texts from the internet, transforms them into task goals with step-by-step instructions, and employs a visual-language model (VLM) agent to simulate their execution in a real digital environment. A VLM-based evaluator ensures the correctness of the generated trajectories. We demonstrate that training GUI agents with these synthesized trajectories significantly improves their grounding and planning performance over the current models. Moreover, our approach is more cost-efficient compared to traditional human annotation methods. This work underscores the potential of guided replay with web tutorials as a viable strategy for large-scale GUI agent training, paving the way for more capable and autonomous digital agents.", "keywords": [ "Data Synthesis", "GUI Agent", "Large Language Model" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "EEbRrNsiiD", "primary_area": "datasets and benchmarks", "n_author": 18, "author_info_list": [ { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Chien-Sheng Wu", "gender": "Male", "institution": "Salesforce AI", "country": "US", "position": "Researcher" }, { "name": "Huan Wang", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Jianguo Zhang", "gender": "Male", "institution": "SalesForce AI Research", "country": "US", "position": "Researcher" }, { "name": "Juntao Tan", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Liangwei Yang", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "US", "position": "PhD student" }, { "name": "Ming Zhu", "gender": "Female", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Ran Xu", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "senior manager" }, { "name": "Rithesh R N", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Sachin Desai", "gender": "unknown", "institution": "Kingston University", "country": "GB", "position": "Undergrad student" }, { "name": "Sarah Tan", "gender": "Female", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "Shelby Heinecke", "gender": "Female", "institution": "Salesforce Research", "country": "US", "position": "Researcher" }, { "name": "Shirley Kokane", "gender": "Female", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Silvio Savarese", "gender": "Male", "institution": "Salesforce", "country": "", "position": "Chief Scientist" }, { "name": "Tulika Manoj Awalgaonkar", "gender": "Female", "institution": "SalesForce.com", "country": "", "position": "Researcher" }, { "name": "Yilun Zhou", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Zhiwei Liu", "gender": "unknown", "institution": "Salesforce AI Research", "country": "US", "position": "Researcher" }, { "name": "Zuxin Liu", "gender": "Male", "institution": "Salesforce AI Research", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 7, "n_ref": 9, "n_ref_all": 34, "n_fig": 7, "n_tab": 8, "L_tab": 9137, "n_element_tab": 1010, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 8442, "n_element_tab_1": 984, "formula_len_all": 366, "formula_len_all_1": 0, "len_all": 150096, "len_all_1": 72750, "len_abs": 1246, "len_title": 113, "len_sents": 30495, "len_sents_1": 24336, "n_sents": 259, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 65, "L_abs": 1251, "title": "MobileAIBench: Benchmarking LLMs and LMMs for On-Device Use Cases", "abs": "The deployment of Large Language Models (LLMs) and Large Multimodal Models (LMMs) on mobile devices has gained significant attention due to the benefits of enhanced privacy, stability, and personalization. However, the hardware constraints of mobile devices necessitate the use of models with fewer parameters and model compression techniques like quantization. Currently, there is limited understanding of quantization's impact on various task performances, including LLM tasks, LMM tasks, and, critically, trust and safety. There is a lack of adequate tools for systematically testing these models on mobile devices. To address these gaps, we introduce MobileAIBench, a comprehensive benchmarking framework for evaluating mobile-optimized LLMs and LMMs. MobileAIBench assesses models across different sizes, quantization levels, and tasks, measuring latency and resource consumption on real devices. Our two-part open-source framework includes a library for running evaluations on desktops and a mobile app for on-device latency and hardware utilization measurements. Our thorough analysis aims to accelerate mobile AI research and deployment by providing insights into the performance and feasibility of deploying LLMs and LMMs on mobile platforms.", "keywords": [ "Large Language Model", "Mobile", "Benchmarking" ], "rating_list": [ 6, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 4, 3, 4, 5 ] }, { "paper_id": "EEWpE9cR27", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Chao Shang", "gender": "Male", "institution": "Amazon AWS AI", "country": "US", "position": "Researcher" }, { "name": "Jie Ma", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Ling Liu", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Lluis Marquez", "gender": "Male", "institution": "Amazon", "country": "", "position": "Principal Researcher" }, { "name": "Miguel Ballesteros", "gender": "Male", "institution": "Amazon", "country": "", "position": "Principal Applied Scientist" }, { "name": "Neha Anna John", "gender": "Female", "institution": "University of Texas at Austin", "country": "", "position": "MS student" }, { "name": "Nikolaos Pappas", "gender": "Male", "institution": "AWS AI Labs", "country": "US", "position": "Researcher" }, { "name": "Qin Liu", "gender": "Female", "institution": "University of California, Davis", "country": "US", "position": "PhD student" }, { "name": "Yassine Benajiba", "gender": "Not Specified", "institution": "Columbia University", "country": "", "position": "Postdoc" }, { "name": "Srikanth Doss Kadarundalagi Raghuram Doss", "gender": "Male", "institution": "University of Texas at Dallas", "country": "", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 24, "n_ref": 67, "n_ref_all": 74, "n_fig": 3, "n_tab": 1, "L_tab": 1116, "n_element_tab": 60, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 3912, "n_element_tab_1": 170, "formula_len_all": 275, "formula_len_all_1": 232, "len_all": 121168, "len_all_1": 63083, "len_abs": 1504, "len_title": 128, "len_sents": 36913, "len_sents_1": 28808, "n_sents": 242, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1441, "title": "Unraveling and Mitigating Safety Alignment Degradation of Vision-Language Models", "abs": "The safety alignment ability of Vision-Language Models (VLMs) is prone to be degraded by the integration of the vision module compared to its LLM backbone. We investigate this phenomenon, dubbed as \u201csafety alignment degradation\u201d in this paper, and show that the challenge arises from the representation gap that emerges when introducing vision modality to VLMs. In particular, we show that the representations of multi-modal inputs shift away from that of text-only inputs which represent the distribution that the LLM backbone is optimized for. At the same time, the safety alignment capabilities, initially developed within the textual embedding space, do not successfully transfer to this new multi-modal representation space. To reduce safety alignment degradation, we introduce Cross-Modality Representation Manipulation (CMRM), an inference time representation intervention method for recovering the safety alignment ability that is inherent in the LLM backbone of VLMs, while simultaneously preserving the functional capabilities of VLMs. The empirical results show that our framework significantly recovers the alignment ability that is inherited from the LLM backbone with minimal impact on the fluency and linguistic capabilities of pre-trained VLMs even without additional training. Specifically, the unsafe rate of LLaVA-7B on multi-modal input can be reduced from 61.53% to as low as 3.15% with only inference-time intervention.", "keywords": [ "Safety Alignment", "Multi-modality", "AI Security" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "EEI5R89Cmv", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Hongshu Guo", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Jiacheng Chen", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yue-Jiao Gong", "gender": "Female", "institution": "South China University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Zeyuan Ma", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 47, "n_ref": 85, "n_ref_all": 105, "n_fig": 6, "n_tab": 2, "L_tab": 748, "n_element_tab": 1, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 829, "n_element_tab_1": 1, "formula_len_all": 359, "formula_len_all_1": 313, "len_all": 129615, "len_all_1": 71191, "len_abs": 1232, "len_title": 85, "len_sents": 39320, "len_sents_1": 35807, "n_sents": 250, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 37, "L_abs": 1237, "title": "Neural Exploratory Landscape Analysis", "abs": "Recent research in Meta-Black-Box Optimization (MetaBBO) have shown that meta-trained neural networks can effectively guide the design of black-box optimizers, significantly reducing the need for expert tuning and delivering robust performance across complex problem distributions. Despite their success, a paradox remains: MetaBBO still rely on human-crafted Exploratory Landscape Analysis features to inform the meta-level agent about the low-level optimization progress. To address the gap, this paper proposes Neural Exploratory Landscape Analysis (NeurELA), a novel framework that dynamically profiles landscape features through a two-stage, attention-based neural network, executed in an entirely end-to-end fashion. NeurELA is pre-trained over a variety of MetaBBO algorithms using a multi-task neuroevolution strategy. Extensive experiments show that NeurELA achieves consistently superior performance when integrated into different and even unseen MetaBBO tasks and can be efficiently fine-tuned for further performance boost. This advancement marks a pivotal step in making MetaBBO algorithms more autonomous and broadly applicable. The source code of NeurELA can be accessed at https://anonymous.4open.science/r/Neur-ELA-303C.", "keywords": [ "Landscape Analysis", "Black-Box Optimization", "Meta-Black-Box Optimization", "Learning to Optimize" ], "rating_list": [ 8, 3, 8, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 3, 1, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "EE2tIwKhSW", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Caradryan Liang", "gender": "unknown", "institution": "University of Southern California", "country": "US", "position": "MS student" }, { "name": "Jiaxuan You", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 6, "n_ref_uni": 47, "n_ref": 149, "n_ref_all": 168, "n_fig": 2, "n_tab": 6, "L_tab": 3655, "n_element_tab": 364, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 3878, "n_element_tab_1": 403, "formula_len_all": 769, "formula_len_all_1": 583, "len_all": 160671, "len_all_1": 68291, "len_abs": 1120, "len_title": 129, "len_sents": 43042, "len_sents_1": 27935, "n_sents": 340, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1124, "title": "Real-World Benchmarks Make Membership Inference Attacks Fail on Diffusion Models", "abs": "Membership inference attacks (MIAs) on diffusion models have emerged as potential evidence of unauthorized data usage in training pre-trained diffusion models. These attacks aim to detect the presence of specific images in training datasets of diffusion models. Our study delves into the evaluation of state-of-the-art MIAs on diffusion models and reveals critical flaws and overly optimistic performance estimates in existing MIA evaluation. We introduce CopyMark, a more realistic MIA benchmark that distinguishes itself through the support for pre-trained diffusion models, unbiased datasets, and fair evaluation pipelines. Through extensive experiments, we demonstrate that the effectiveness of current MIA methods significantly degrades under these more practical conditions. Based on our results, we alert that MIA, in its current state, is not a reliable approach for identifying unauthorized data usage in pre-trained diffusion models. To the best of our knowledge, we are the first to discover the performance overestimation of MIAs on diffusion models and present a unified benchmark for more realistic evaluation.", "keywords": [ "Membership inference attack", "Diffusion models", "Benchmark" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "EDoD3DgivF", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Sarah Wiegreffe", "gender": "Female", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Postdoc" }, { "name": "Yanai Elazar", "gender": "Male", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Postdoc" }, { "name": "jack merullo", "gender": "Male", "institution": "Brown University", "country": "US", "position": "PhD student" }, { "name": "Noah Smith", "gender": "Male", "institution": "University of Washington", "country": "US", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 51, "n_ref": 96, "n_ref_all": 119, "n_fig": 14, "n_tab": 5, "L_tab": 1535, "n_element_tab": 36, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1535, "n_element_tab_1": 36, "formula_len_all": 71, "formula_len_all_1": 71, "len_all": 159651, "len_all_1": 63159, "len_abs": 2441, "len_title": 123, "len_sents": 40807, "len_sents_1": 31477, "n_sents": 260, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 2043, "title": "On Linear Representations and Pretraining Data Frequency in Language Models", "abs": "Pretraining data has a direct impact on the behaviors and quality of language models (LMs), but we only understand the most basic principles of this relationship. While most work focuses on pretraining data's effect on downstream task behavior, we investigate its relationship to LM representations. Previous work has discovered that, in language models, some concepts are encoded as ``linear representations'', but what factors cause these representations to form (or not)? We study the connection between differences in pretraining data frequency and differences in trained models' linear representations of factual recall relations. We find evidence that the two are linked, with the formation of linear representations strongly connected to pretraining term frequencies. First, we establish that the presence of linear representations for subject-relation-object (s-r-o) fact triplets is highly correlated with both subject-object co-occurrence frequency and in-context learning accuracy. This is the case across all phases of pretraining, i.e., it is not affected by the model's underlying capability. In OLMo 7B and GPT-J (6B), we discover that a linear representation consistently (but not exclusively) forms when the subjects and objects within a relation co-occur at least 1-2k times, regardless of when these occurrences happen during pretraining. In the OLMo 1B model, consistent linearity only occurs after 4.4k occurrences, suggesting a connection to scale. Finally, we train a regression model on measurements of linear representation quality that can predict how often a term was seen in pretraining. We show such model achieves low error even for a different model and pretraining dataset, providing a new unsupervised method for exploring possible data sources of closed-source models. We conclude that the presence or absence of linear representations in LMs contains signal about their pretraining corpora that may provide new avenues for controlling and improving model behavior. We release our code to support future work.", "keywords": [ "pretraining data", "pretraining", "linear", "linear features", "interpretability", "linear representations", "corpus frequency" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "EDJ7cPZk7V", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Guy Hacohen", "gender": "Male", "institution": "KU Leuven", "country": "", "position": "Postdoc" }, { "name": "Tinne Tuytelaars", "gender": "Female", "institution": "KU Leuven", "country": "BE", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 57, "n_ref": 78, "n_ref_all": 125, "n_fig": 32, "n_tab": 7, "L_tab": 848, "n_element_tab": 107, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 592, "n_element_tab_1": 1, "formula_len_all": 82, "formula_len_all_1": 111, "len_all": 180494, "len_all_1": 65703, "len_abs": 823, "len_title": 127, "len_sents": 57825, "len_sents_1": 32065, "n_sents": 452, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 825, "title": "Forgetting Order of Continual Learning: What is Learned First is Forgotten Last", "abs": "Catastrophic forgetting poses a significant challenge in continual learning, where models often forget previous tasks when trained on new data. Our empirical analysis reveals a strong correlation between catastrophic forgetting and the learning speed of examples: examples learned early are rarely forgotten, while those learned later are more susceptible to forgetting. We demonstrate that replay-based continual learning methods can leverage this phenomenon by focusing on mid-learned examples for rehearsal. We introduce Goldilocks, a novel replay buffer sampling method that filters out examples learned too quickly or too slowly, keeping those learned at an intermediate speed. Goldilocks improves existing continual learning algorithms, leading to state-of-the-art performance across several image classification tasks.", "keywords": [ "continual learning", "catastrophic forgetting", "replay buffer" ], "rating_list": [ 10, 5, 6, 3, 8 ], "soundness_list": [ 4, 2, 3, 2, 3 ], "presentation_list": [ 4, 2, 3, 2, 3 ], "contribution_list": [ 4, 2, 2, 2, 4 ], "confidence_list": [ 4, 5, 4, 5, 4 ] }, { "paper_id": "ED5w271rWo", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Mattia Opper", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Siddharth N", "gender": "Male", "institution": "University of Edinburgh", "country": "", "position": "Reader (Associate Professor)" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 54, "n_ref": 89, "n_ref_all": 108, "n_fig": 4, "n_tab": 5, "L_tab": 2484, "n_element_tab": 229, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2420, "n_element_tab_1": 234, "formula_len_all": 1082, "formula_len_all_1": 1082, "len_all": 191723, "len_all_1": 79547, "len_abs": 1513, "len_title": 112, "len_sents": 44875, "len_sents_1": 36537, "n_sents": 317, "n_sents_1": 266, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1521, "title": "Banyan: Improved Representation Learning with Explicit Structure", "abs": "We present Banyan, a model that efficiently learns semantic representations by leveraging an inductive bias towards explicit hierarchical structure. Although typical transformer-based models excel at scale, they struggle in low-resource settings. Recent work on models exploiting explicit structure has shown promise as efficient learners in resource-constrained environments. However, these models have yet to demonstrate truly competitive performance. Banyan bridges this gap, significantly improving upon prior structured models and providing, for the first time, a viable alternative to transformer embeddings for under-represented languages. We achieve these improvements through two key innovations 1) A novel entangled tree structure that resolves multiple constituent structures into a single shared one, explicitly incorporating global context. 2) Diagonalized message passing functions that increase the influence of the inductive bias. Our final model has just 14 non-embedding parameters yet is competitive with baselines many orders of magnitude larger. Banyan outperforms its structured predecessors and competes with large unstructured models across various semantic tasks in multiple languages. Notably, it excels in low-resource settings, highlighting its potential for efficient and interpretable NLP in resource-constrained environments. These results underscore the value of appropriate inductive biases in capturing semantic relationships and open new avenues for efficient, interpretable NLP models.", "keywords": [ "Representation Learning", "Structure", "Semantics", "Syntax", "Induction", "Composition" ], "rating_list": [ 8, 6, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 2, 4, 3, 4 ] }, { "paper_id": "EBaMTeWi2K", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "James R. Glass", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Principal Researcher" }, { "name": "Kaizhi Qian", "gender": "unknown", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Wei Fang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yada Zhu", "gender": "unknown", "institution": "IBM Research", "country": "", "position": "Principal Research Scientist" }, { "name": "Yang Zhang", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Research Staff Employee" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 29, "n_ref": 55, "n_ref_all": 62, "n_fig": 2, "n_tab": 6, "L_tab": 1649, "n_element_tab": 132, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1649, "n_element_tab_1": 132, "formula_len_all": 165, "formula_len_all_1": 165, "len_all": 160486, "len_all_1": 66938, "len_abs": 1273, "len_title": 130, "len_sents": 35898, "len_sents_1": 33427, "n_sents": 215, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1280, "title": "PLAY2PROMPT: Zero-shot Tool Instruction Optimization for LLM Agents via Tool Play", "abs": "Large language models (LLMs) are increasingly integrated with external tools to complete user requests. Many real-world applications require LLMs to use specialized tools in a zero-shot setting. To achieve this, current methods primarily rely on prompting LLMs with tool-specific information, yet tool documentation is often underspecified or noisy, limiting effectiveness. Manual improvements are inefficient and impractical, as they require domain expertise to rewrite documentation and test on carefully curated held-out datasets to evaluate performance gains. Automatic prompt engineering techniques are not applicable either, because they require labeled examples, which is unavailable in the zero-shot setting. In this work, we introduce PLAY2PROMPT, an automated framework that iteratively refines tool documentation and generates usage examples. PLAY2PROMPT enables LLMs to explore tool input-output behaviors, allowing us to effectively search the space of possible tool descriptions and examples. The generated examples not only guide LLM inference but also serve as validation data to ensure more effective tool use. Extensive experiments on real-world tasks demonstrate significant improvements in zero-shot tool performance across both open- and closed-source models.", "keywords": [ "large language models", "zero-shot tool use", "prompt optimization" ], "rating_list": [ 3, 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3, 4 ] }, { "paper_id": "EBT0oymkZb", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Chen Yang", "gender": "unknown", "institution": "Indiana University", "country": "US", "position": "PhD student" }, { "name": "Dongruo Zhou", "gender": "Male", "institution": "Indiana University", "country": "US", "position": "Assistant Professor" }, { "name": "John Lui", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Zhiyong Wang", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Visiting PhD student" } ], "n_formula": 77, "n_formula_1": 25, "n_ref_uni": 42, "n_ref": 90, "n_ref_all": 123, "n_fig": 2, "n_tab": 3, "L_tab": 1710, "n_element_tab": 157, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 1465, "n_element_tab_1": 128, "formula_len_all": 12763, "formula_len_all_1": 3546, "len_all": 207948, "len_all_1": 70847, "len_abs": 1018, "len_title": 114, "len_sents": 51844, "len_sents_1": 29938, "n_sents": 446, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1075, "title": "Towards Zero-Shot Generalization in Offline Reinforcement Learning", "abs": "In this work, we study offline reinforcement learning (RL) with zero-shot generalization property (ZSG), where the agent has access to an offline dataset including experiences from different environments, and the goal of the agent is to train a policy over the training environments which performs well on test environments without further interaction. Existing work showed that classical offline RL fails to generalize to new, unseen environments. We propose pessimistic empirical risk minimization (PERM) and pessimistic proximal policy optimization (PPPO), which leverage pessimistic policy evaluation to guide policy learning and enhance generalization. Theoretically, our framework is capable of finding a near-optimal policy with ZSG. Empirically, our framework demonstrates the ability to enhance the performance of the base offline RL methods. Our result serves as a first step in understanding the foundation of the generalization phenomenon in offline reinforcement learning. Our codes are released at [this link](https://anonymous.4open.science/r/ProcgenExp-B5B4).", "keywords": [ "offline reinforcement learning", "generalization" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "EBBeSbmAyh", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 7, "author_info_list": [ { "name": "Chao Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Assistant Professor" }, { "name": "Chuan Qin", "gender": "Male", "institution": "CNIC, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Hui Xiong", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Full Professor" }, { "name": "Li Shen", "gender": "Male", "institution": "Sun Yat-Sen University", "country": "CN", "position": "Associate Professor" }, { "name": "Liwei Deng", "gender": "unknown", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Long Yang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Postdoc" }, { "name": "Tianfu Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 48, "n_formula_1": 17, "n_ref_uni": 46, "n_ref": 120, "n_ref_all": 151, "n_fig": 12, "n_tab": 9, "L_tab": 2262, "n_element_tab": 334, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1496, "n_element_tab_1": 89, "formula_len_all": 2775, "formula_len_all_1": 1083, "len_all": 259693, "len_all_1": 83310, "len_abs": 1401, "len_title": 130, "len_sents": 90689, "len_sents_1": 38775, "n_sents": 705, "n_sents_1": 274, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1505, "title": "Towards Constraint-aware Learning for Resource Allocation in NFV-enabled Networks", "abs": "Virtual Network Embedding (VNE) is a challenging combinatorial optimization problem that refers to resource allocation associated with hard and multifaceted constraints in network function virtualization (NFV). Existing works for VNE struggle to handle such complex constraints, leading to compromised system performance and stability. In this paper, we propose a \\textbf{CON}straint-\\textbf{A}ware \\textbf{L}earning framework for VNE, named \\textbf{CONAL}, to achieve efficient constraint management. Concretely, we formulate the VNE problem as a constrained Markov decision process with violation tolerance. This modeling approach aims to improve both resource utilization and solution feasibility by precisely evaluating solution quality and the degree of constraint violation. We also propose a reachability-guided optimization with an adaptive reachability budget method that dynamically assigns budget values. This method achieves persistent zero violation to guarantee the feasibility of VNE solutions and more stable policy optimization by handling instances without any feasible solution. Furthermore, we propose a constraint-aware graph representation method to efficiently learn cross-graph relations and constrained path connectivity in VNE. Finally, extensive experimental results demonstrate the superiority of our proposed method over state-of-the-art baselines. Our code is available at \\href{https://anonymous.4open.science/r/iclr25-conal}{https://anonymous.4open.science/r/iclr25-conal}.", "keywords": [ "Network Resource Allocation; Combinatorial Optimization; Reinforcement Learning; Graph Neural Network" ], "rating_list": [ 5, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 4, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "EAUGN4pszX", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Chunyan Miao", "gender": "Female", "institution": "School of Computer Science and Engineering, Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Hangwei Qian", "gender": "Female", "institution": "Centre For Frontier AI Research (CFAR), A*STAR, Singapore", "country": "SG", "position": "Scientist" }, { "name": "Tian Tian", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" } ], "n_formula": 36, "n_formula_1": 18, "n_ref_uni": 27, "n_ref": 75, "n_ref_all": 111, "n_fig": 6, "n_tab": 14, "L_tab": 9417, "n_element_tab": 1600, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 560, "n_element_tab_1": 104, "formula_len_all": 2826, "formula_len_all_1": 1194, "len_all": 203772, "len_all_1": 68681, "len_abs": 2045, "len_title": 143, "len_sents": 57219, "len_sents_1": 31850, "n_sents": 408, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 2054, "title": "FreRA: A Frequency-Refined Augmentation for Contrastive Learning on Time Series Classification", "abs": "Contrastive learning has emerged as a competent approach for unsupervised representation learning. However, the design of an optimal augmentation strategy, although crucial for contrastive learning, is less explored for time series classification tasks. Existing predefined time-domain augmentation methods are primarily adopted from vision and are not specific to time series data. Consequently, this cross-modality incompatibility may distort the global semantics of time series by introducing mismatched patterns into the data. To address this limitation, we present a novel perspective from the frequency domain and identify three advantages for downstream classification: 1) the frequency component naturally encodes global features, 2) the orthogonal nature of the Fourier basis allows easier isolation and independent modifications of critical and unimportant information, and 3) a compact set of frequency components can preserve semantic integrity. To fully utilize the three properties, we propose the lightweight yet effective Frequency-Refined Augmentation (FreRA) tailored for time series contrastive learning on classification tasks, which can be seamlessly integrated with contrastive learning frameworks in a plug-and-play manner. Specifically, FreRA automatically separates critical and unimportant frequency components. Accordingly, we propose Identity Modification and Self-adaptive Modification to protect global semantics in the critical frequency components and infuse variance to the unimportant ones respectively. \nTheoretically, we prove that FreRA generates semantic-preserving views. Empirically, we conduct extensive experiments on two benchmark datasets including UCR and UEA archives, as well as 5 large-scale datasets on diverse applications. FreRA consistently outperforms 10 leading baselines on time series classification, anomaly detection, and transfer learning tasks, demonstrating superior capabilities in contrastive representation learning and generalization in transfer learning scenarios across diverse datasets.", "keywords": [ "time series classification", "contrastive learning", "frequency domain" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "EAT5Jpa4ws", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Jumin Lee", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" }, { "name": "Kyu Beom Han", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Sung-eui Yoon", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "Professor" }, { "name": "Taeyeon Kim", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Woo Jae Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Youngju Na", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 46, "n_ref": 148, "n_ref_all": 179, "n_fig": 15, "n_tab": 6, "L_tab": 1118, "n_element_tab": 178, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1624, "n_element_tab_1": 135, "formula_len_all": 606, "formula_len_all_1": 609, "len_all": 156625, "len_all_1": 63383, "len_abs": 885, "len_title": 136, "len_sents": 46108, "len_sents_1": 28924, "n_sents": 337, "n_sents_1": 204, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 889, "title": "SHARE: Bridging Shape and Ray Estimation for Pose-Free Generalizable Gaussian Splatting", "abs": "While generalizable 3D Gaussian Splatting enables efficient, high-quality rendering of unseen scenes, it heavily depends on precise camera poses for accurate geometry. In real-world scenarios, obtaining accurate poses is challenging, leading to noisy pose estimates and geometric misalignments. To address this, we introduce SHARE, a novel pose-free generalizable Gaussian Splatting framework that overcomes these ambiguities. Our ray-guided multi-view fusion network consolidates multi-view features into a unified pose-aware canonical volume, bridging 3D reconstruction and ray-based pose estimation. In addition, we propose an anchor-aligned Gaussian prediction strategy for fine-grained geometry estimation within a canonical view.\nExtensive experiments on diverse real-world datasets show that SHARE achieves state-of-the-art performance in pose-free generalizable Gaussian splatting.", "keywords": [ "3D Gaussian Splatting", "Novel View Synthesis", "Pose Estimation" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "E9NQUvbsT1", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Amitabha Bagchi", "gender": "Male", "institution": "Indian Institute of Technology, Delhi", "country": "IN", "position": "Full Professor" }, { "name": "Anjali Gupta", "gender": "Female", "institution": "Indian Institute of Technology, Delhi", "country": "IN", "position": "PhD student" }, { "name": "Anuj Kumar Sirohi", "gender": "Male", "institution": "Indian Institute of Technology, Delhi", "country": "IN", "position": "PhD student" }, { "name": "Sandeep Kumar", "gender": "Male", "institution": "Indian Institute of Technology Delhi", "country": "IN", "position": "Assistant Professor" }, { "name": "Sayan Ranu", "gender": "Male", "institution": "Indian Institute of Technology Delhi", "country": "IN", "position": "Associate Professor" } ], "n_formula": 23, "n_formula_1": 23, "n_ref_uni": 33, "n_ref": 64, "n_ref_all": 99, "n_fig": 7, "n_tab": 11, "L_tab": 5396, "n_element_tab": 360, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1673, "n_element_tab_1": 87, "formula_len_all": 1409, "formula_len_all_1": 1719, "len_all": 141686, "len_all_1": 69543, "len_abs": 1222, "len_title": 84, "len_sents": 40192, "len_sents_1": 28984, "n_sents": 357, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1230, "title": "Task and Model Agnostic Differentially Private Graph Neural Networks via Coarsening", "abs": "Graph Neural Networks (GNNs) have emerged as powerful tools for analyzing graph-structured data, deriving representations by aggregating information from neighboring nodes. However, this aggregation process inherently increases the risk of exposing confidential data, as a single node may influence the inference process for multiple nodes simultaneously. To mitigate this risk, researchers have explored differentially private training methods for GNN models. Existing privacy-preserving approaches, however, face significant challenges. They often incur high computational costs during training or struggle to generalize across various GNN models and task objectives. To address these limitations, we introduce Differentially Private Graph Coarsening (DPGC), a novel method that tackles two key challenges in GNN training: scalability and privacy guarantees that are independent of the downstream task or GNN model. Through comprehensive experiments on six datasets across diverse prediction tasks, we demonstrate that DPGC sets new benchmarks in graph coarsening. Our method achieves superior compression-accuracy trade-offs while maintaining robust privacy guarantees, outperforming state-of-the-art baselines in this domain.", "keywords": [ "Graph Neural Network (GNN)", "Differential Privacy (DP)", "Graph Coarsening" ], "rating_list": [ 1, 6, 5 ], "soundness_list": [ 1, 3, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "E9GakjQype", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Anselm Paulus", "gender": "unknown", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Arman Zharmagambetov", "gender": "Male", "institution": "Meta AI (FAIR)", "country": "US", "position": "Researcher" }, { "name": "Brandon Amos", "gender": "unknown", "institution": "Meta AI", "country": "US", "position": "Research Scientist" }, { "name": "Chuan Guo", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Yuandong Tian", "gender": "Male", "institution": "Meta AI (FAIR)", "country": "", "position": "Research Scientist" } ], "n_formula": 24, "n_formula_1": 7, "n_ref_uni": 45, "n_ref": 119, "n_ref_all": 166, "n_fig": 8, "n_tab": 16, "L_tab": 4578, "n_element_tab": 335, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1995, "n_element_tab_1": 188, "formula_len_all": 1129, "formula_len_all_1": 389, "len_all": 304526, "len_all_1": 65627, "len_abs": 1770, "len_title": 113, "len_sents": 81318, "len_sents_1": 29151, "n_sents": 680, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1764, "title": "AdvPrompter: Fast Adaptive Adversarial Prompting for LLMs", "abs": "While recently Large Language Models (LLMs) have achieved remarkable successes, they are vulnerable to certain `jailbreaking attacks` that lead to generation of inappropriate or harmful content. Manual red-teaming requires finding adversarial prompts that cause such jailbreaking, e.g. by appending a suffix to a given instruction, which is inefficient and time-consuming. \nOn the other hand, automatic adversarial prompt generation often leads to semantically meaningless attacks that can easily be detected by perplexity-based filters, may require gradient information from the TargetLLM, or do not scale well due to time-consuming discrete optimization processes over the token space. In this paper, we present a novel method that uses another LLM, called the `AdvPrompter`, to generate human-readable adversarial prompts in seconds, $\\sim800\\times$ faster than existing optimization-based approaches.\nWe train the AdvPrompter using a novel algorithm that `does not require gradients` of the TargetLLM. This process alternates between two steps: (1) generating high-quality target adversarial suffixes by optimizing the AdvPrompter predictions, and (2) fine-tuning of the AdvPrompter with the generated adversarial suffixes. The trained AdvPrompter generates suffixes that veil the input instruction without changing its meaning, such that the TargetLLM is lured to give a harmful response. Experimental results on popular open source TargetLLMs show state-of-the-art results on the AdvBench dataset, that also transfer to closed-source black-box LLM APIs. Further, we demonstrate that by fine-tuning on a synthetic dataset generated by AdvPrompter, LLMs can be made more robust against jailbreaking attacks while maintaining performance, i.e. high MMLU scores.", "keywords": [ "adversarial attacks", "prompt optimization", "red-teaming LLMs" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "E8gYIrbP00", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Aparna Elangovan", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" }, { "name": "Dan Roth", "gender": "Male", "institution": "Oracle", "country": "US", "position": "Chief Scientist" }, { "name": "Jongwoo Ko", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Lei Xu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Ling Liu", "gender": "unknown", "institution": "Amazon", "country": "", "position": "Researcher" }, { "name": "Mahsa Elyasi", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Sravan Babu Bodapati", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 41, "n_ref": 68, "n_ref_all": 104, "n_fig": 6, "n_tab": 4, "L_tab": 3342, "n_element_tab": 442, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1305, "n_element_tab_1": 249, "formula_len_all": 497, "formula_len_all_1": 564, "len_all": 180126, "len_all_1": 67604, "len_abs": 1791, "len_title": 174, "len_sents": 54162, "len_sents_1": 32025, "n_sents": 362, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 125, "L_abs": 2044, "title": "Beyond correlation: The impact of human uncertainty in measuring the effectiveness of automatic evaluation and LLM-as-a-judge", "abs": "The effectiveness of automatic evaluation of generative models is typically measured by comparing the labels generated via automation with human labels using correlation metrics. \nHowever, metrics like Krippendorff's $\\alpha$ and Randolph's $\\kappa$ were originally designed to measure the reliability of human labeling, thus make assumptions about typical human labeling behavior, and these assumptions may not be applicable to machine generated labels. \nIn this paper, we show how *relying on a single aggregate correlation score* can obscure fundamental differences between human labels and those from automatic evaluation, including LLM-as-a-Judge. \nSpecifically, we demonstrate that when the proportion of samples with variation or uncertainty in human assigned labels is relatively high, machine labels (generated by automatic evaluation methods) may superficially appear to have similar or better correlation with the human majority label compared to the human-to-human (HH) correlation. \nThis can create the illusion that labels from automatic evaluation approximates the human majority label. \nHowever, as the proportion of samples with consistent human labels increases, the correlation between machine and human labels fall well below HH correlation. \nBased on these findings, we first propose *stratifying data by human label uncertainty* to provide a more robust analysis of automatic evaluation performance. Second, recognizing that uncertainty and variation are inherent in perception-based human evaluations, such as those involving attitudes or preferences, we introduce a new metric -*binned Jensen-Shannon Divergence for perception* for such scenarios to better measure the effectiveness of automatic evaluations. Third, we present visualization techniques -- *perception charts*, to contextualize correlation measures appropriately and to show the strengths and limitations of automatic evaluation. We have open-sourced our analysis and visualization tools at https://github.com/amazon-science/BeyondCorrelation.", "keywords": [ "Automated evaluation", "LLM as a judge", "correlation measures" ], "rating_list": [ 5, 8, 6, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 5, 5, 3, 3 ] }, { "paper_id": "E8TPUAimyJ", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Adityanarayanan Radhakrishnan", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Amirhesam Abedsoltan", "gender": "Male", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Jingfeng Wu", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Misha Belkin", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "Professor" } ], "n_formula": 29, "n_formula_1": 23, "n_ref_uni": 26, "n_ref": 88, "n_ref_all": 113, "n_fig": 12, "n_tab": 1, "L_tab": 730, "n_element_tab": 47, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 730, "n_element_tab_1": 47, "formula_len_all": 2179, "formula_len_all_1": 1430, "len_all": 115084, "len_all_1": 62933, "len_abs": 1380, "len_title": 106, "len_sents": 34928, "len_sents_1": 25882, "n_sents": 316, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1392, "title": "Context-Scaling versus Task-Scaling in In-Context Learning", "abs": "Transformers exhibit In-Context Learning (ICL), a phenomenon in which these models solve new tasks by using examples in the prompt without additional training. In our work, we analyze two key components of ICL: (1) context-scaling, where model performance improves as the number of in-context examples increases and (2) task-scaling, where model performance improves as the number of pre-training tasks increases. While transformers are capable of both context-scaling and task-scaling, we empirically show that standard Multi-Layer Perceptrons (MLPs) with vectorized input are only capable of task-scaling. To understand how transformers are capable of context-scaling, we first propose a significantly simplified transformer that performs ICL comparably to the original GPT-2 model in statistical learning tasks (e.g., linear regression, teacher-student settings). By analyzing a single layer of our proposed model, we identify classes of feature maps that enable context scaling. Theoretically, these feature maps can implement the Hilbert estimate, a model that is provably consistent for context-scaling. We then show that using the output of the Hilbert estimate along with vectorized input empirically enables both context-scaling and task-scaling with MLPs. Overall, our findings provide insights into the fundamental mechanisms of how transformers are able to learn in context.", "keywords": [ "in-context learning", "kernel smoothers", "Hilbert estimate" ], "rating_list": [ 3, 6, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "E8S5Upr6oO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Errui Ding", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Director" }, { "name": "Hanli Wang", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Full Professor" }, { "name": "Jingdong Wang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Chief Scientist for Computer Vision" }, { "name": "Minyue Jiang", "gender": "Male", "institution": "Baidu", "country": "", "position": "Researcher" }, { "name": "Sen Yang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Xiao Tan", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" }, { "name": "Yingying Li", "gender": "Female", "institution": "Baidu", "country": "CN", "position": "RD" }, { "name": "YangJing", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 18, "n_ref": 34, "n_ref_all": 45, "n_fig": 8, "n_tab": 11, "L_tab": 1828, "n_element_tab": 299, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2270, "n_element_tab_1": 134, "formula_len_all": 1087, "formula_len_all_1": 966, "len_all": 129209, "len_all_1": 57455, "len_abs": 1757, "len_title": 146, "len_sents": 35779, "len_sents_1": 25622, "n_sents": 279, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1231, "title": "MGMapNet: Multi-Granularity Representation Learning for End-to-End Vectorized HD Map Construction", "abs": "The construction of vectorized high-definition map typically requires capturing both category and geometry information of map elements. Current state-of-the-art methods often adopt solely either point-level or instance-level representation, overlooking the strong intrinsic relationship between points and instances. In this work, we propose a simple yet efficient framework named MGMapNet (multi-granularity map network) to model map elements with multi-granularity representation, integrating both coarse-grained instance-level and fine-grained point-level queries. Specifically, these two granularities of queries are generated from the multi-scale bird's eye view features using a proposed multi-granularity aggregator. In this module, instance-level query aggregates features over the entire scope covered by an instance, and the point-level query aggregates features locally. Furthermore, a point-instance interaction module is designed to encourage information exchange between instance-level and point-level queries. Experimental results demonstrate that the proposed MGMapNet achieves state-of-the-art performances, surpassing MapTRv2 by 5.3 mAP on the nuScenes dataset and 4.4 mAP on the Argoverse2 dataset, respectively.", "keywords": [ "Online HD map construction\uff0cvectorized representation\uff0cautonomous driving" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 4 ], "presentation_list": [ 3, 2, 4 ], "contribution_list": [ 3, 2, 4 ], "confidence_list": [ 2, 3, 3 ] }, { "paper_id": "E7gjRqFT9O", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Juan Gabriel Kostelec", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CH", "position": "Researcher" }, { "name": "Qinghai Guo", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 50, "n_ref": 95, "n_ref_all": 108, "n_fig": 2, "n_tab": 5, "L_tab": 2433, "n_element_tab": 382, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1173, "n_element_tab_1": 186, "formula_len_all": 1677, "formula_len_all_1": 1667, "len_all": 171671, "len_all_1": 60332, "len_abs": 1129, "len_title": 108, "len_sents": 34749, "len_sents_1": 24540, "n_sents": 226, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1089, "title": "FlashEVA: Accelerating LLM Inference via Efficient Attention", "abs": "Transformer models have revolutionized natural language processing, achieving state-of-the-art performance and demonstrating remarkable scalability. However, their memory demands, particularly due to maintaining full context in memory, pose significant challenges for inference. In this paper, we present FlashEVA, an efficient implementation of EVA (Efficient Attention via Control Variates), and demonstrate how to finetune transformers to adapt to FlashEVA attention. Our method enables fine-tuning of Transformer models with as few as $1.6B$ tokens while preserving effectiveness across various downstream tasks. Notably, FlashEVA achieves up to $6.7x$ higher throughput during inference compared to standard Transformer implementations. Despite these improvements, we observe limitations in retrieval-focused tasks. Our implementation offers control over the trade-off between throughput and accuracy through adjustable hyperparameters, providing greater flexibility. This work represents a significant step towards more efficient and adaptable Transformer-based models for inference.", "keywords": [ "efficient attention", "transformers", "large language models", "inference" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 5 ] }, { "paper_id": "E7ecidOeCE", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Baobao Chang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Fanchao Qi", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Gang Chen", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Haozhe Zhao", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Kaikai An", "gender": "Male", "institution": "Peking University", "country": "", "position": "MS student" }, { "name": "Kangyang Luo", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Maosong Sun", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Shuzheng Si", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yunshui Li", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Lv Chuancheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 25, "n_ref": 72, "n_ref_all": 105, "n_fig": 5, "n_tab": 8, "L_tab": 3258, "n_element_tab": 443, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2533, "n_element_tab_1": 376, "formula_len_all": 465, "formula_len_all_1": 198, "len_all": 180878, "len_all_1": 62275, "len_abs": 6996, "len_title": 175, "len_sents": 55845, "len_sents_1": 26721, "n_sents": 410, "n_sents_1": 168, "venue": "ICLR2025", "decision": "withdraw", "L_title": 125, "L_abs": 1878, "title": "Selecting Influential Samples for Long Context Alignment via Homologous Models\u2019 Guidance and Contextual Awareness Measurement", "abs": "The expansion of large language models to effectively handle instructions with extremely long contexts has yet to be fully investigated. The primary obstacle lies in constructing a high-quality long instruction-following dataset devised for long context alignment. Existing studies have attempted to scale up the available data volume by synthesizing long instruction-following samples. However, indiscriminately increasing the quantity of data without a well-defined strategy for ensuring data quality may introduce low-quality samples and restrict the final performance. To bridge this gap, we aim to address the unique challenge of long-context alignment, i.e., modeling the long-range dependencies for handling instructions and lengthy input contexts. We propose GATEAU, a novel framework designed to identify the influential and high-quality samples enriched with long-range dependency relations by utilizing crafted Homologous Models' Guidance (HMG) and Contextual Awareness Measurement (CAM). Specifically, HMG attempts to measure the difficulty of generating corresponding responses due to the long-range dependencies, using the perplexity scores of the response from two homologous models with different context windows. Also, the role of CAM is to measure the difficulty of understanding the long input contexts due to long-range dependencies by evaluating whether the model\u2019s attention is focused on important segments. Built upon both proposed methods, we select the most challenging samples as the influential data to effectively frame the long-range dependencies, thereby achieving better performance of LLMs. Comprehensive experiments indicate that GATEAU effectively identifies samples enriched with long-range dependency relations and the model trained on these selected samples exhibits better instruction-following and long-context understanding capabilities.", "keywords": [ "Long context alignment", "Large language models", "Data selection", "Efficient instruction tuning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "E77uvbOTtp", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Geon Yeong Park", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Hyelin Nam", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Hyungjin Chung", "gender": "Male", "institution": "EverEx", "country": "KR", "position": "Researcher" }, { "name": "Jeongsol Kim", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Jong Chul Ye", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Full Professor" } ], "n_formula": 56, "n_formula_1": 23, "n_ref_uni": 33, "n_ref": 102, "n_ref_all": 144, "n_fig": 19, "n_tab": 5, "L_tab": 733, "n_element_tab": 118, "n_fig_1": 12, "n_tab_1": 3, "L_tab_1": 219, "n_element_tab_1": 24, "formula_len_all": 5605, "formula_len_all_1": 2383, "len_all": 141808, "len_all_1": 58837, "len_abs": 602, "len_title": 121, "len_sents": 39687, "len_sents_1": 25725, "n_sents": 335, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1745, "title": "CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models", "abs": "Classifier-free guidance (CFG) is a fundamental tool in modern diffusion models for text-guided generation. Although effective, CFG has notable drawbacks. For instance, DDIM with CFG lacks invertibility, complicating image editing; furthermore, high guidance scales, essential for high-quality outputs, frequently result in issues like mode collapse. Contrary to the widespread belief that these are inherent limitations of diffusion models, this paper reveals that the problems actually stem from the off-manifold phenomenon associated with CFG, rather than the diffusion models themselves. More specifically, inspired by the recent advancements of diffusion model-based inverse problem solvers (DIS), we reformulate text-guidance as an inverse problem with a text-conditioned score matching loss and develop CFG++, a novel approach that tackles the off-manifold challenges inherent in traditional CFG. CFG++ features a surprisingly simple fix to CFG, yet it offers significant improvements, including better sample quality for text-to-image generation, invertibility, smaller guidance scales, reduced etc. Furthermore, CFG++ enables seamless interpolation between unconditional and conditional sampling at lower guidance scales, consistently outperforming traditional CFG at all scales. Moreover, CFG++ can be easily integrated into the high-order diffusion solvers and naturally extends to distilled diffusion models. Experimental results confirm that our method significantly enhances performance in text-to-image generation, DDIM inversion, editing, and solving inverse problems, suggesting a wide-ranging impact and potential applications in various fields that utilize text guidance. Project Page: https://cfgpp-diffusion.github.io/anon", "keywords": [ "Diffusion models", "Manifold", "Classifier-free guidance" ], "rating_list": [ 1, 8, 6, 6 ], "soundness_list": [ 2, 4, 3, 2 ], "presentation_list": [ 2, 4, 3, 2 ], "contribution_list": [ 1, 4, 4, 3 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "E6rpTruK4v", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Bonaventure F. P. Dossou", "gender": "Male", "institution": "McGill University, McGill University", "country": "", "position": "PhD student" }, { "name": "YuXuan Wu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Intern" }, { "name": "Dianbo Liu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 26, "n_ref": 43, "n_ref_all": 57, "n_fig": 4, "n_tab": 7, "L_tab": 2256, "n_element_tab": 88, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 3134, "n_element_tab_1": 68, "formula_len_all": 471, "formula_len_all_1": 471, "len_all": 131879, "len_all_1": 55342, "len_abs": 1068, "len_title": 142, "len_sents": 50361, "len_sents_1": 23917, "n_sents": 344, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 1071, "title": "CodeUnlearn: Amortized Zero-Shot Machine Unlearning in Language Models Using Discrete Concept", "abs": "Language Models (LMs) offer extensive knowledge across various domains, but they may inadvertently memorize sensitive, unauthorized, or malicious data, such as personal information in the medical and financial sectors. Machine unlearning methods aim to remove specific information from models after training to address this. However, current approaches require additional model training or struggle to effectively erase particular data points and their associated context due to LMs' complex, dense, and continuous nature. In this study, we propose a novel amortized unlearning approach using codebook features and Sparse Autoencoders (SAEs). By leveraging a bottleneck to decompose the activation space and regulate information flow, our method efficiently unlearns targeted information while preserving the model's performance on unrelated data. To the best of our knowledge, this is the first work that successfully enables unlearning specific topics with contextual relevance in an LM, marking a significant step towards real-world applications of machine unlearning.", "keywords": [ "machine unlearning", "discrete representation", "AI safety", "LLM" ], "rating_list": [ 5, 5, 5, 1, 3 ], "soundness_list": [ 3, 2, 3, 1, 1 ], "presentation_list": [ 2, 2, 2, 1, 2 ], "contribution_list": [ 2, 2, 2, 1, 1 ], "confidence_list": [ 3, 3, 3, 3, 4 ] }, { "paper_id": "E6kQ51yfAj", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 13, "author_info_list": [ { "name": "Hang Li", "gender": "Male", "institution": "ByteDance Technology", "country": "CN", "position": "Head of Research" }, { "name": "Hongyi Guo", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "", "position": "PhD student" }, { "name": "Qi Zhang", "gender": "Male", "institution": "Fudan University", "country": "", "position": "Full Professor" }, { "name": "Rui Zheng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Tao Gui", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaojun Xu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Xiaoying Zhang", "gender": "Female", "institution": "ByteDance AILab", "country": "", "position": "Researcher" }, { "name": "Xuanjing Huang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yang Liu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "Zhaoran Wang", "gender": "Not Specified", "institution": "Northwestern University", "country": "US", "position": "Associate Professor" }, { "name": "Zhihan Liu", "gender": "Male", "institution": "Northwestern University", "country": "", "position": "PhD student" }, { "name": "Zhiheng Xi", "gender": "unknown", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Kevin Yao", "gender": "Male", "institution": "Meta GenAI", "country": "US", "position": "Researcher" } ], "n_formula": 50, "n_formula_1": 15, "n_ref_uni": 34, "n_ref": 68, "n_ref_all": 91, "n_fig": 6, "n_tab": 6, "L_tab": 2098, "n_element_tab": 238, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 917, "n_element_tab_1": 129, "formula_len_all": 4266, "formula_len_all_1": 997, "len_all": 205907, "len_all_1": 70811, "len_abs": 1351, "len_title": 100, "len_sents": 56841, "len_sents_1": 33126, "n_sents": 485, "n_sents_1": 242, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1167, "title": "Progressive LLM Alignments Using Two-Player Games", "abs": "Alignment of large language models (LLM) is a process that ensures the model\u2019s responses to user prompts align with human intentions and social values. This optimization typically relies on pre-collected prompts. The collection of these prompts often either requires careful human interventions or proves to be difficult to have a good coverage over all scenarios an LLM can improve over . To address this issue, we propose an alignment method based on a two-agent game, consisting of an adversarial agent and a defensive agent. The adversarial agent\u2019s task is to generate prompts that expose the deficiencies of the defensive agent. At the same time, the defensive agent improves its performance on the prompts generated by the adversary based on feedback from the reward model. This iterative process is repeated to enhance the model\u2019s performance. We theoretically demonstrate that, under mild assumptions, this iterative alignment process converges to a Nash equilibrium by both agents. Learning in this competitive environment results in policies with better generalization capabilities. We demonstrate the advantage of our framework using extensive experiments.", "keywords": [ "large language models", "alignment", "safety" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "E6B0bbMFbi", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Baoxiang Wang", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" }, { "name": "Hongyuan Zha", "gender": "unknown", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Wenhao Li", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yue Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 1, "n_ref_uni": 105, "n_ref": 171, "n_ref_all": 211, "n_fig": 16, "n_tab": 9, "L_tab": 1229, "n_element_tab": 249, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 63, "n_element_tab_1": 18, "formula_len_all": 343, "formula_len_all_1": 93, "len_all": 397483, "len_all_1": 73205, "len_abs": 1758, "len_title": 78, "len_sents": 146993, "len_sents_1": 35629, "n_sents": 956, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Reject", "L_title": 30, "L_abs": 1621, "title": "Verbalized Bayesian Persuasion", "abs": "The study of information design explores how an information designer can influence the optimal behavior of players to achieve a specific objective through the strategic selection of the information provided. \nThis paper focuses on a case, Bayesian Persuasion (BP), where the information designer holds an informational advantage over only one player.\nWhile information design originates from everyday human communication, traditional game-theoretic or multi-agent reinforcement learning methods often model information structures as discrete or continuous scalars or vectors, this approach fails to capture the nuances of natural language, significantly limiting their applicability in real-world scenarios.\nBy leveraging the powerful language understanding and generation capabilities of large language models (LLMs), this paper proposes a verbalized BP framework that extends classic BP to real-world games involving human dialogues for the first time. \nSpecifically, we map the classic BP to a verbalized mediator-augmented game, where LLMs instantiate the information designer and receiver.\nTo efficiently solve the game in the language space, we transform agents' policy optimization into prompt optimization and propose a generalized equilibrium-finding algorithm with a convergence guarantee. \nNumerical experiments in realistic dialogue scenarios, such as recommendation letters, courtroom interactions, and law enforcement, validate that the VBP framework can reproduce theoretical results in classic settings and discover effective persuasion strategies in more complex natural language and multistage settings.", "keywords": [ "Large Language Models", "Information Design", "Bayesian Persuasion", "Game Theory", "Multiagent Systems" ], "rating_list": [ 6, 3, 3, 3 ], "soundness_list": [ 2, 3, 1, 2 ], "presentation_list": [ 3, 2, 1, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 2, 3, 4, 3 ] }, { "paper_id": "E5ulvtj86q", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Marcos Negre Saura", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "PhD student" }, { "name": "Richard Allmendinger", "gender": "Male", "institution": "University of Manchester ", "country": "GB", "position": "Full Professor" }, { "name": "Theodore Papamarkou", "gender": "Male", "institution": "Zhejiang Normal University", "country": "CN", "position": "Distinguished Professor" }, { "name": "Wei Pan", "gender": "Male", "institution": "University of Manchester", "country": "GB", "position": "Associate Professor" } ], "n_formula": 27, "n_formula_1": 21, "n_ref_uni": 36, "n_ref": 53, "n_ref_all": 66, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1591, "formula_len_all_1": 1170, "len_all": 140597, "len_all_1": 63474, "len_abs": 1179, "len_title": 133, "len_sents": 49407, "len_sents_1": 30141, "n_sents": 348, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1183, "title": "Spatial-aware decision-making with ring attractors in Reinforcement Learning systems", "abs": "This paper explores the integration of ring attractors, a mathematical model inspired by neural circuit dynamics, into the reinforcement learning (RL) action selection process. Ring attractors, as specialized brain-inspired structures that encode spatial information and uncertainty, offer a biologically plausible mechanism to improve learning speed and predictive performance. They do so by explicitly encoding the action space, facilitating the organization of neural activity, and enabling the distribution of spatial representations across the neural network in the context of deep RL. The application of ring attractors in the RL action selection process involves mapping actions to specific locations on the ring and decoding the selected action based on neural activity. We investigate the application of ring attractors by both building them as exogenous models and integrating them as part of a Deep Learning policy algorithm. Our results show a significant improvement in state-of-the-art models for the Atari 100k benchmark. Notably, our integrated approach improves the performance of state-of-the-art models by half, representing a 53% increase over selected baselines.", "keywords": [ "Reinforcement Learning", "Computational Neuroscience", "Deep Learning", "Ring Attractors", "Spatial Awareness", "Bioinspired" ], "rating_list": [ 5, 6, 6, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "E5YnuidZ9W", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Bo Zhao", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Nima Dehmamy", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Robin Walters", "gender": "Male", "institution": "Northeastern University ", "country": "US", "position": "Assistant Professor" }, { "name": "Rose Yu", "gender": "Female", "institution": "University of California, San Diego", "country": "", "position": "Assistant Professor" } ], "n_formula": 58, "n_formula_1": 15, "n_ref_uni": 26, "n_ref": 56, "n_ref_all": 69, "n_fig": 5, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 2201, "n_element_tab_1": 24, "formula_len_all": 7562, "formula_len_all_1": 1569, "len_all": 168633, "len_all_1": 67266, "len_abs": 1051, "len_title": 108, "len_sents": 53917, "len_sents_1": 28132, "n_sents": 558, "n_sents_1": 244, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1067, "title": "Understanding Mode Connectivity via Parameter Space Symmetry", "abs": "Neural network minima have been observed to be connected by curves along which train and test loss remain nearly constant, a phenomenon known as mode connectivity. \nWhile this has enabled applications such as model merging and fine-tuning, its theoretical explanation remains unclear. \nWe propose a new approach to exploring the connectedness of minima using parameter space symmetry.\nBy linking the topology of symmetry groups to that of the minima, we derive the number of connected components of the minima of linear networks and show that skip connections reduce this number. \nWe then examine when mode connectivity and linear mode connectivity hold or fail, using parameter symmetries which account for a significant part of the minimum.\nFinally, we provide explicit expressions for connecting curves in the minima induced by symmetry. \nUsing the curvature of these curves, we derive conditions under which linear mode connectivity approximately holds.\nOur analysis highlights the role of continuous symmetries in understanding the neural network loss landscape.", "keywords": [ "symmetry", "mode connectivity" ], "rating_list": [ 8, 6, 3, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 3, 4 ], "contribution_list": [ 3, 3, 1, 3, 3 ], "confidence_list": [ 3, 2, 4, 3, 3 ] }, { "paper_id": "E5YmIBvOqV", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Qiang Qiu", "gender": "unknown", "institution": "Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "Wei Chen", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Zichen Miao", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 23, "n_ref_uni": 77, "n_ref": 162, "n_ref_all": 198, "n_fig": 14, "n_tab": 11, "L_tab": 4149, "n_element_tab": 428, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 1684, "n_element_tab_1": 177, "formula_len_all": 2207, "formula_len_all_1": 2003, "len_all": 201122, "len_all_1": 67601, "len_abs": 1243, "len_title": 100, "len_sents": 46836, "len_sents_1": 28157, "n_sents": 378, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 52, "L_abs": 1250, "title": "Large Convolutional Model Tuning via Filter Subspace", "abs": "Efficient fine-tuning methods are critical to address the high computational and parameter complexity while adapting large pre-trained models to downstream tasks.\nOur study is inspired by prior research that represents each convolution filter as a linear combination of a small set of filter subspace elements, referred to as filter atoms. In this paper, we propose to fine-tune pre-trained models by adjusting only filter atoms, which are responsible for spatial-only convolution, while preserving spatially-invariant channel combination knowledge in atom coefficients.\nIn this way, we bring a new filter subspace view for model tuning. \nFurthermore, each filter atom can be recursively decomposed as a combination of another set of atoms, which naturally expands the number of tunable parameters in the filter subspace.\nBy only adapting filter atoms constructed by a small number of parameters, while maintaining the rest of model parameters constant, the proposed approach is highly parameter-efficient. It effectively preserves the capabilities of pre-trained models and prevents overfitting to downstream tasks. \nExtensive experiments show that such a simple scheme surpasses previous tuning baselines for both discriminate and generative tasks.", "keywords": [ "Efficient Fine-tuning", "Filter Decomposition", "Filter Subspace" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 5, 3, 4 ] }, { "paper_id": "E5DYpUWsES", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Fangfang Li", "gender": "unknown", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Jing Li", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "MS student" }, { "name": "Quanxue Gao", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoke Ma", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Cheng Deng", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "QIANQIAN WANG", "gender": "Female", "institution": "Xidian University", "country": "CN", "position": "Lecturer" } ], "n_formula": 42, "n_formula_1": 40, "n_ref_uni": 21, "n_ref": 26, "n_ref_all": 71, "n_fig": 4, "n_tab": 4, "L_tab": 4260, "n_element_tab": 143, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 4636, "n_element_tab_1": 148, "formula_len_all": 2533, "formula_len_all_1": 2535, "len_all": 95938, "len_all_1": 58303, "len_abs": 1214, "len_title": 94, "len_sents": 21997, "len_sents_1": 20711, "n_sents": 193, "n_sents_1": 174, "venue": "ICLR2025", "decision": "withdraw", "L_title": 52, "L_abs": 1232, "title": "Manifold K-means with $\\ell_{2,p}$-Norm Maximization", "abs": "Although a variety of different methods have emerged in the field of clustering, K-means still occupies an important position, and many advanced clustering methods even rely on the K-means to achieve effective cluster detection. However, the sensitivity of K-means to the selection of the initial cluster center and its limited ability to handle nonlinear separable data somewhat restrict its clustering performance. In order to overcome the limitations of K-means, we draw inspiration from manifold learning and redefine K-means as a manifold K-means clustering framework. This framework supports various types of distance matrices, thus facilitating the efficient processing of nonlinear separable data. A unique advantage of this approach is that it does not require the calculation of the cluster center, while it maintains the consistency between manifold structure and cluster labels. Additionally, we highlight the significant role of the $\\ell_{2,p}$-norm; by maximizing the $\\ell_{2,p}$-norm, we can ensure the balance of classes in the clustering process, which is also supported by theoretical analysis. The results from extensive experiments across multiple databases substantiate the superiority of our proposed model.", "keywords": [ "Clustering", "Manifold Learning", "K-means", "$\\ell_{2", "p}$-Norm" ], "rating_list": [ 3, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 5, 5, 4, 4 ] }, { "paper_id": "E4roJSM9RM", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Boqiang Duan", "gender": "Male", "institution": "Baidu", "country": "", "position": "Researcher" }, { "name": "Jingdong Wang", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Chief Scientist for Computer Vision" }, { "name": "Leye Wang", "gender": "Male", "institution": "Peking University", "country": "", "position": "Assistant Professor" }, { "name": "Mingyu Ding", "gender": "Male", "institution": "Department of Computer Science, University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "zhu Jie", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 23, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 76, "n_ref_all": 121, "n_fig": 17, "n_tab": 10, "L_tab": 2636, "n_element_tab": 303, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 280, "n_element_tab_1": 33, "formula_len_all": 2194, "formula_len_all_1": 775, "len_all": 160327, "len_all_1": 62973, "len_abs": 1309, "len_title": 108, "len_sents": 43011, "len_sents_1": 27765, "n_sents": 365, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 965, "title": "Unveiling the Secret of AdaLN-Zero in Diffusion Transformer", "abs": "Diffusion transformer (DiT), a rapidly emerging architecture for image generation, has gained much attention. However, despite ongoing efforts to improve its performance, the understanding of DiT remains superficial. In this work, we delve into and investigate a critical conditioning mechanism within DiT, adaLN-Zero, which achieves superior performance compared to adaLN. Our work studies three potential elements driving this performance, including an SE-like structure, zero-initialization, and a \u201cgradual\u201d update order, among which zero-initialization is proved to be the most influential. Building on this insight, we heuristically leverage Gaussian distributions to initialize each condition modulation, termed adaLN-Gaussian, leading to more stable and effective training. Extensive experiments following DiT on ImageNet1K demonstrate the effectiveness and generalization of adaLN-Gaussian, e.g., a notable improvement of 2.16% in FID score over adaLN-Zero.", "keywords": [ "Diffusion transformer", "zero-initialization", "image generation" ], "rating_list": [ 3, 6, 6, 5 ], "soundness_list": [ 2, 3, 4, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 1, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "E4kuNZWost", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Dongshen Wu", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Undergrad student" }, { "name": "Takafumi Kanamori", "gender": "Male", "institution": "RIKEN", "country": "JP", "position": "Visiting Researcher" }, { "name": "Yuhui Zhang", "gender": "Not Specified", "institution": "Tokyo Institute of Technology, Tokyo Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Yuichiro Wada", "gender": "unknown", "institution": "Fujitsu Research", "country": "JP", "position": "Researcher" } ], "n_formula": 64, "n_formula_1": 30, "n_ref_uni": 49, "n_ref": 110, "n_ref_all": 140, "n_fig": 6, "n_tab": 7, "L_tab": 2317, "n_element_tab": 232, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2455, "n_element_tab_1": 59, "formula_len_all": 6953, "formula_len_all_1": 2647, "len_all": 208083, "len_all_1": 69855, "len_abs": 963, "len_title": 129, "len_sents": 53295, "len_sents_1": 28457, "n_sents": 517, "n_sents_1": 267, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 978, "title": "TULiP: Test-time Uncertainty Estimation via Linearization and Weight Perturbation", "abs": "A reliable uncertainty estimation method is the foundation of many modern out-of-distribution (OOD) detectors, which are critical for safe deployments of deep learning models in the open world. In this work, we propose TULiP, a novel, theoretically-driven, post-hoc uncertainty estimator for OOD detection. Our method considers a hypothetical perturbation applied to the network prior to convergence. Based on linearized training dynamics, we bound the effect of such perturbation, resulting in an uncertainty score computable by perturbing model parameters. Ultimately, our approach computes uncertainty from a set of sampled predictions, thus not limited to classification problems. We visualize our bound on synthetic regression and classification datasets. Furthermore, we demonstrate the effectiveness of TULiP using large-scale OOD detection benchmarks for image classification. Our method exhibits state-of-the-art performance, particularly for near-distribution samples.", "keywords": [ "Out-of-distribution detection", "Uncertainty Quantification", "Lazy Training", "Neural Tangent Kernel" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "E4OcXAx5Dc", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Ben Jacobsen", "gender": "Not Specified", "institution": "University of Wisconsin - Madison", "country": "US", "position": "PhD student" }, { "name": "Kassem Fawaz", "gender": "unknown", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Associate Professor" } ], "n_formula": 26, "n_formula_1": 9, "n_ref_uni": 47, "n_ref": 72, "n_ref_all": 80, "n_fig": 2, "n_tab": 3, "L_tab": 573, "n_element_tab": 56, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 240, "n_element_tab_1": 12, "formula_len_all": 2314, "formula_len_all_1": 1061, "len_all": 157179, "len_all_1": 66244, "len_abs": 887, "len_title": 161, "len_sents": 50940, "len_sents_1": 31364, "n_sents": 373, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 113, "L_abs": 896, "title": "Private Learning Fast and Slow: Two Algorithms for Prediction with Expert Advice Under Local Differential Privacy", "abs": "We study the classic problem of prediction with expert advice under the constraint of differential privacy (DP). In contrast to earlier work in this area, we are interested in distributed settings with no trusted central curator. In this context, we first show that a classical online learning algorithm naturally satisfies DP and then design two new algorithms that extend and improve it: (1) RW-AdaBatch, which provides a novel form of privacy amplification at negligible utility cost, and (2) RW-Meta, which improves utility on non-adversarial data with zero privacy cost. Our theoretical analysis is supported by an empirical evaluation using real-world data reported by hospitals during the COVID-19 pandemic. RW-Meta outperforms the classical baseline at predicting which hospitals will report a high density of COVID-19 cases by a factor of more than 2$\\times$ at realistic privacy levels.", "keywords": [ "differential privacy", "online learning", "prediction with expert advice", "follow the perturbed leader" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 2, 3 ] }, { "paper_id": "E4NShSRRDP", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 9, "author_info_list": [ { "name": "Angelos Stefanidis", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "CN", "position": "Full Professor" }, { "name": "Dufan Wu", "gender": "unknown", "institution": "Massachusetts General Hospital, Harvard University", "country": "", "position": "Assistant Professor" }, { "name": "Jinfeng Wang", "gender": "Male", "institution": "University of Liverpool", "country": "GB", "position": "PhD student" }, { "name": "Jionglong Su", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "CN", "position": "Associate Professor" }, { "name": "Qiaochu Zhao", "gender": "Male", "institution": "Xi'an Jiaotong-Liverpool University", "country": "CN", "position": "Undergrad student" }, { "name": "Quanzheng Li", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Associate Professor" }, { "name": "S Kevin Zhou", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Sifan Song", "gender": "Male", "institution": "Massachusetts General Hospital, Harvard University", "country": "US", "position": "Postdoc" }, { "name": "Xiang Li", "gender": "Male", "institution": "Massachusetts General Hospital, Harvard University", "country": "", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 31, "n_ref": 79, "n_ref_all": 106, "n_fig": 10, "n_tab": 9, "L_tab": 4276, "n_element_tab": 554, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 3462, "n_element_tab_1": 503, "formula_len_all": 827, "formula_len_all_1": 753, "len_all": 123865, "len_all_1": 68596, "len_abs": 1662, "len_title": 100, "len_sents": 33999, "len_sents_1": 28167, "n_sents": 228, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1346, "title": "Contrastive Learning Via Equivariant Representation", "abs": "Invariant Contrastive Learning (ICL) methods have achieved impressive performance across various domains. However, the absence of latent space representation for distortion (augmentation)-related information in the latent space makes ICL sub-optimal regarding training efficiency and robustness in downstream tasks. Recent studies suggest that introducing equivariance into Contrastive Learning (CL) can improve overall performance. In this paper, we revisit the roles of augmentation strategies and equivariance in improving CL's efficacy. We propose CLeVER (Contrastive Learning Via Equivariant Representation), a novel equivariant contrastive learning framework compatible with augmentation strategies of arbitrary complexity for various mainstream CL backbone models. Experimental results demonstrate that CLeVER effectively extracts and incorporates equivariant information from practical natural images, thereby improving the training efficiency and robustness of baseline models in downstream tasks and achieving state-of-the-art (SOTA) performance. Moreover, we find that leveraging equivariant information extracted by CLeVER simultaneously enhances rotational invariance and sensitivity across experimental tasks, and helps stabilize the framework when handling complex augmentations, particularly for models with small-scale backbones.", "keywords": [ "Contrastive Learning", "Self-Supervised Learning", "Equivariant Contrastive Learning", "Invariant Contrastive Learning" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 1, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 3, 4, 5, 4 ] }, { "paper_id": "E4LAVLXAHW", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Martin Vechev", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "Full Professor" }, { "name": "Nikola Jovanovi\u0107", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Robin Staab", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Thibaud Gloaguen", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "MS student" } ], "n_formula": 24, "n_formula_1": 17, "n_ref_uni": 23, "n_ref": 70, "n_ref_all": 107, "n_fig": 6, "n_tab": 11, "L_tab": 3347, "n_element_tab": 285, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1010, "n_element_tab_1": 97, "formula_len_all": 1379, "formula_len_all_1": 873, "len_all": 171348, "len_all_1": 69868, "len_abs": 1959, "len_title": 96, "len_sents": 64097, "len_sents_1": 33890, "n_sents": 468, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 1193, "title": "Black-Box Detection of Language Model Watermarks", "abs": "Watermarking has emerged as a promising way to detect LLM-generated text, by augmenting LLM generations with later detectable signals. Recent work has proposed multiple families of watermarking schemes, several of which focus on preserving the LLM distribution. This distribution-preservation property is motivated by the fact that it is a tractable proxy for retaining LLM capabilities, as well as the inherently implied undetectability of the watermark by downstream users. Yet, despite much discourse around undetectability, no prior work has investigated the practical detectability of any of the current watermarking schemes in a realistic black-box setting. In this work we tackle this for the first time, developing rigorous statistical tests to detect the presence, and estimate parameters, of all three popular watermarking scheme families, using only a limited number of black-box queries. We experimentally confirm the effectiveness of our methods on a range of schemes and a diverse set of open-source models. Further, we validate the feasibility of our tests on real-world APIs. Our findings indicate that current watermarking schemes are more detectable than previously believed.", "keywords": [ "llm", "watermarking" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 2, 3, 4 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 5, 4, 3 ] }, { "paper_id": "E4Fk3YuG56", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Alexander Hertzberg", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Brody Huval", "gender": "unknown", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Erik Wijmans", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Philipp Kraehenbuehl", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "Associate Professor" }, { "name": "Vladlen Koltun", "gender": "Male", "institution": "Apple", "country": "US", "position": "Distinguished Scientist" } ], "n_formula": 20, "n_formula_1": 16, "n_ref_uni": 35, "n_ref": 70, "n_ref_all": 105, "n_fig": 12, "n_tab": 5, "L_tab": 2796, "n_element_tab": 237, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 610, "n_element_tab_1": 42, "formula_len_all": 2867, "formula_len_all_1": 2416, "len_all": 141482, "len_all_1": 57955, "len_abs": 1537, "len_title": 99, "len_sents": 41341, "len_sents_1": 23845, "n_sents": 373, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 51, "L_abs": 1496, "title": "Cut Your Losses in Large-Vocabulary Language Models", "abs": "As language models grow ever larger, so do their vocabularies.\nThis has shifted the memory footprint of LLMs during training disproportionately to one single layer: the cross-entropy in the loss computation.\nCross-entropy builds up a logit matrix with entries for each pair of input tokens and vocabulary items and, for small models, consumes an order of magnitude more memory than the rest of the LLM combined.\nWe propose Cut Cross-Entropy (CCE), a method that computes the cross-entropy loss without materializing the logits for all tokens into global memory.\nRather, CCE only computes the logit for the correct token and evaluates the log-sum-exp over all logits on the fly.\nWe implement a custom kernel that performs the matrix multiplications and the log-sum-exp reduction over the vocabulary in flash memory, making global memory consumption for the cross-entropy computation negligible. This has a dramatic effect. Taking the Gemma 2 (2B) model as an example, CCE reduces the memory footprint of the loss computation from 24 GB to 1 MB, and the total training-time memory consumption of the classifier head from 28 GB to 1 GB.\nTo improve the throughput of CCE, we leverage the inherent sparsity of softmax and propose to skip elements of the gradient computation that have a negligible (i.e. below numerical precision) contribution to the gradient.\nExperiments demonstrate that the dramatic reduction in memory consumption is accomplished without sacrificing training speed or convergence.", "keywords": [ "large language model", "large vocabulary", "efficient" ], "rating_list": [ 10, 8, 6, 10 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 4, 3, 3, 4 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "E4A7KtLB21", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Fang Chen", "gender": "unknown", "institution": "University of Technology Sydney (UTS)", "country": "AU", "position": "Full Professor" }, { "name": "Jianlong Zhou", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Jiayu Zhang", "gender": "Male", "institution": "Suzhou Yierqi", "country": "", "position": "Researcher" }, { "name": "Zhibo Jin", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Zhu Zhiyu", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "MS student" } ], "n_formula": 18, "n_formula_1": 7, "n_ref_uni": 26, "n_ref": 68, "n_ref_all": 88, "n_fig": 7, "n_tab": 10, "L_tab": 8243, "n_element_tab": 590, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 6675, "n_element_tab_1": 404, "formula_len_all": 1781, "formula_len_all_1": 608, "len_all": 129623, "len_all_1": 67896, "len_abs": 1275, "len_title": 96, "len_sents": 40641, "len_sents_1": 26483, "n_sents": 288, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1279, "title": "Unbiased Attribution with Intrinsic Information", "abs": "The importance of attribution algorithms in the AI field lies in enhancing model transparency, diagnosing and improving models, ensuring fairness, and increasing user understanding. Gradient-based attribution methods have become the most critical because of their high computational efficiency, continuity, wide applicability, and flexibility. However, current gradient-based attribution algorithms require the introduction of additional class information to interpret model decisions, which can lead to issues of information ignorance and extra information. Information ignorance can obscure important features relevant to the current model decision, while extra information introduces irrelevant data that can cause feature leakage in the attribution process. To address these issues, we propose the Attribution with Intrinsic Information (AII) algorithm, which analyzes model decisions without the need for specified class information. Additionally, to better evaluate the potential of current attribution algorithms, we introduce the metrics of insertion confusion and deletion confusion alongside existing mainstream metrics. To continuously advance research in the field of explainable AI (XAI), our algorithm is open-sourced at https://anonymous.4open.science/r/AII-787D/.", "keywords": [ "Interpretability", "Attribution" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 1, 2, 4, 1 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "E48QvQppIN", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 9, "author_info_list": [ { "name": "Alan Nawzad Amin", "gender": "Male", "institution": "New York University", "country": "", "position": "Postdoc" }, { "name": "Andrew Gordon Wilson", "gender": "Not Specified", "institution": "New York University", "country": "US", "position": "Full Professor" }, { "name": "Aniruddh Raghu", "gender": "Male", "institution": "BigHat Biosciences", "country": "", "position": "Machine Learning Scientist" }, { "name": "Calvin McCarter", "gender": "Not Specified", "institution": "Bighat Biosciences", "country": "US", "position": "Researcher" }, { "name": "Hunter Elliott", "gender": "Male", "institution": "BigHat Biosciences", "country": "US", "position": "Principal Researcher" }, { "name": "Nate Gruver", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Peyton Greenside", "gender": "Female", "institution": "Stanford University", "country": "", "position": "Postdoc" }, { "name": "Yilun Kuang", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "Yucen Lily Li", "gender": "unknown", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 44, "n_ref": 75, "n_ref_all": 113, "n_fig": 25, "n_tab": 1, "L_tab": 307, "n_element_tab": 46, "n_fig_1": 12, "n_tab_1": 1, "L_tab_1": 78, "n_element_tab_1": 4, "formula_len_all": 796, "formula_len_all_1": 465, "len_all": 174892, "len_all_1": 67262, "len_abs": 1344, "len_title": 136, "len_sents": 53970, "len_sents_1": 32796, "n_sents": 429, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 88, "L_abs": 1352, "title": "Bayesian Optimization of Antibodies Informed by a Generative Model of Evolving Sequences", "abs": "To build effective therapeutics, biologists iteratively mutate antibody sequences to improve binding and stability. Proposed mutations can be informed by previous measurements or by learning from large antibody databases to predict only typical antibodies. Unfortunately, the space of typical antibodies is enormous to search, and experiments often fail to find suitable antibodies on a budget. Here we introduce Clone-informed Bayesian Optimization (CloneBO), a Bayesian optimization procedure that efficiently optimizes antibodies in the lab by teaching a generative model how our immune system optimizes antibodies in our bodies. Our immune system makes antibodies by iteratively evolving specific portions of their sequences to bind their target strongly and stably, resulting in a set of related, evolving sequences known as a *clonal family*. We train a large language model, CloneLM, on hundreds of thousands of clonal families and use it to design sequences with mutations that are most likely to optimize an antibody in our bodies. We guide our designs to fit previous measurements using a twisted sequential Monte Carlo procedure. We show that CloneBO optimizes antibodies substantially more efficiently than previous methods in realistic *in silico* experiments and designs stronger and more stable binders in *in vitro* wet lab experiments.", "keywords": [ "Bayesian optimization", "generative model", "antibody", "biological sequence" ], "rating_list": [ 8, 8, 8, 5 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 2, 3 ] }, { "paper_id": "E3qIInyTgL", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Kento Oonishi", "gender": "unknown", "institution": "Mitsubishi Electric Corporation", "country": "", "position": "Researcher" }, { "name": "Tsunato Nakai", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 21, "n_ref": 48, "n_ref_all": 118, "n_fig": 2, "n_tab": 21, "L_tab": 5897, "n_element_tab": 346, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 304, "n_element_tab_1": 30, "formula_len_all": 161, "formula_len_all_1": 139, "len_all": 157574, "len_all_1": 74105, "len_abs": 1514, "len_title": 134, "len_sents": 58473, "len_sents_1": 37512, "n_sents": 448, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1518, "title": "CC-VFed: Client Contribution Detects Byzantine Attacks in Vertical Federated Learning", "abs": "Vertical federated learning (VFL) is a type of federated learning where the collection of different features is shared among multiple clients, and it is attracting attention as a training method that takes into account the privacy and security of training data. On the other hand, in federated learning, there is a threat of Byzantine attacks, where some malicious clients disrupt the training of the model and output an trained model that does not exhibit the behavior that should be obtained. Thus far, numerous defense methods against Byzantine attacks on horizontal federated learning have been proposed, most of which focus on the similarity of the models generated across clients having the similar features and mitigate the attacks by excluding outliers. However, in VFL, the feature sets assigned by each client are inherently different, making similar methods inapplicable, and there is little existing research in this area. In light of the above, this paper organizes and classifies feasible Byzantine attacks and proposes a new defense method CC-VFed against these attack methods. Firstly, this paper organizes and classifies attack methods that contaminate training data, demonstrating that sign-flipping attacks pose a threat to VFL. Subsequently, in order to capture the differences in client features, this paper proposes a method for detecting and neutralizing malicious clients based on their contribution to output labels, demonstrating that it is indeed possible to defend Byzantine attacks in VFL.", "keywords": [ "Vertical Federated Learning", "Byzantine Attacks" ], "rating_list": [ 3, 6, 1 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "E3PgLQzPob", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Baixu", "gender": "Male", "institution": "Xiaohongshu", "country": "CN", "position": "Researcher" }, { "name": "Hao Ai", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "MS student" }, { "name": "Haofan Wang", "gender": "Male", "institution": "InstantX", "country": "CN", "position": "Researcher" }, { "name": "Jen-Yuan Huang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Peng Xing", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Yanpeng Sun", "gender": "Male", "institution": "Nanjing University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Zechao Li", "gender": "Male", "institution": "Nanjing University of Science and Techonolgy", "country": "CN", "position": "Full Professor" }, { "name": "wangqixun", "gender": "Male", "institution": "Tencent Hunyuan", "country": "CN", "position": "Instructor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 55, "n_ref_all": 72, "n_fig": 10, "n_tab": 1, "L_tab": 554, "n_element_tab": 238, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 2091, "n_element_tab_1": 256, "formula_len_all": 182, "formula_len_all_1": 187, "len_all": 93167, "len_all_1": 47977, "len_abs": 0, "len_title": 108, "len_sents": 23670, "len_sents_1": 19606, "n_sents": 193, "n_sents_1": 158, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1159, "title": "CSGO: Content-Style Composition in Text-to-Image Generation", "abs": "The diffusion model has shown exceptional capabilities in controlled image generation, which has further fueled interest in image style transfer. Existing works mainly focus on training free-based methods (e.g., image inversion) due to the scarcity of specific data. In this study, we present a data construction pipeline for content-style-stylized image triplets that generates and automatically cleanses stylized triplets. Based on this pipeline, we construct a dataset IMAGStyle, the first large-scale style transfer dataset containing 210k image triplets, available for the community to explore and research.Equipped with IMAGStyle, we propose a simple yet effective framework CSGO, a style transfer model based on end-to-end training, which explicitly decouples content and style features employing independent feature injection. Our CSGO implements image-driven style transfer, text-driven stylized synthesis, and text editing-driven stylized synthesis in the same model.\nWe conduct extensive experiments on CSGO to validate the effectiveness of synthetic stylized data for style control. Meanwhile, ablation experiments show the effectiveness of CSGO.", "keywords": [ "image generation", "style transfer", "stylized synthesis" ], "rating_list": [ 6, 5, 5, 6, 5 ], "soundness_list": [ 2, 2, 3, 3, 1 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 5, 5, 4, 5 ] }, { "paper_id": "E3LDsbUSRZ", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Chenchen Ye", "gender": "Female", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Mingyu Derek Ma", "gender": "Male", "institution": "Genentech", "country": "US", "position": "Machine Learning Scientist" }, { "name": "Peipei Ping", "gender": "Female", "institution": "University of California, Los Angeles", "country": "", "position": "Full Professor" }, { "name": "Timothy S Chang", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "Wei Wang", "gender": "Female", "institution": "University of California, Los Angeles", "country": "US", "position": "Full Professor" }, { "name": "Xiaoxuan Wang", "gender": "Female", "institution": ", University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Yu Yan", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 21, "n_ref": 36, "n_ref_all": 62, "n_fig": 4, "n_tab": 15, "L_tab": 18121, "n_element_tab": 1213, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 7768, "n_element_tab_1": 598, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 296960, "len_all_1": 76721, "len_abs": 5192, "len_title": 156, "len_sents": 53847, "len_sents_1": 30108, "n_sents": 388, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1544, "title": "CliBench: A Multifaceted and Multigranular Evaluation of Large Language Models for Clinical Decision Making", "abs": "The integration of Artificial Intelligence (AI), especially Large Language Models (LLMs), into the clinical diagnosis process offers significant potential to improve the efficiency and accessibility of medical care. While LLMs have shown some promise in the medical domain, their application in clinical diagnosis remains underexplored, especially in real-world clinical practice, where highly sophisticated, patient-specific decisions need to be made. Current evaluations of LLMs in this field are often narrow in scope, focusing on specific diseases or specialties and employing simplified diagnostic tasks. To bridge this gap, we introduce CliBench, a novel benchmark developed from the MIMIC IV dataset, offering a comprehensive and realistic assessment of LLMs' capabilities in clinical diagnosis. This benchmark not only covers diagnosis from a diverse range of medical cases across various specialties but also incorporates tasks of clinical significance: treatment procedure identification, lab test ordering and medication prescriptions. Supported by structured output ontologies, CliBench enables a precise and multi-granular evaluation, offering an in-depth understanding of LLM's capability on diverse clinical tasks of desired granularity. We conduct a zero-shot evaluation of leading LLMs to assess their proficiency in clinical decision-making. Our preliminary results shed light on the potential and limitations of current LLMs in clinical settings, providing valuable insights for future advancements in LLM-powered healthcare.", "keywords": [ "Clinical Decisions", "Large Language Model", "Benchmark" ], "rating_list": [ 3, 5, 6, 3, 5 ], "soundness_list": [ 2, 2, 2, 1, 2 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 1, 3, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "E36NHwe7Zc", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Lili Zhao", "gender": "Female", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Mengyun Wang", "gender": "Female", "institution": "IFLYTEK CO.LTD.", "country": "", "position": "Researcher" }, { "name": "Qi Liu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Shijin Wang", "gender": "Male", "institution": "State Key Laboratory of Cognitive Intelligence", "country": "", "position": "Vice Dean" }, { "name": "Wei Chen", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Yang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Zhichao Sheng", "gender": "unknown", "institution": "IFLYTEK CO.LTD.", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 40, "n_ref": 59, "n_ref_all": 111, "n_fig": 8, "n_tab": 12, "L_tab": 10714, "n_element_tab": 742, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 7935, "n_element_tab_1": 481, "formula_len_all": 117, "formula_len_all_1": 361, "len_all": 209177, "len_all_1": 76714, "len_abs": 1947, "len_title": 141, "len_sents": 60086, "len_sents_1": 29348, "n_sents": 459, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1982, "title": "Evaluating Large Language Models through Role-Guide and Self-Reflection: A Comparative Study", "abs": "Large Language Models fine-tuned with Reinforcement Learning from Human Feedback (RLHF-LLMs) can over-rely on aligned preferences without truly gaining the self-knowledge, leading to hallucination and biases. If an LLM can better access its knowledge and know what it knows, it can avoid making false or unsupported claims. Therefore, it is crucial to evaluate whether LLMs have the ability to know what they know, which can help to ensure accuracy and faithfulness in real-world applications. Inspired by research in Educational Psychology, students who don't really know are easily affected by teacher and peer guidance, we treat LLM as a student, incorporate role guidance in prompts to explore whether LLMs really know. Specifically, we propose a novel strategy called **Ro**le-Guide and **Se**lf-Reflection (**RoSe**) to fully assess whether LLM ``knows it knows''. We introduce multiple combinations of different roles and strong reminder in prompts combined with self-reflection to explore what local information LLMs rely on, and whether LLMs remain unaffected by external guidance with varying roles. Our findings reveal that LLMs are very sensitive to the strong reminder information. Role guidance can help LLMs reduce their reliance on strong reminder. Meanwhile, LLMs tend to trust the role of authority more when guided by different roles. Following these findings, we propose a double-calibrated strategy with verbalized confidence to extract well-calibrated data from closed-source LLM and fine-tune open-source LLMs. Extensive experiments conducted on fine-tuning open-source LLMs demonstrate the effectiveness of double-calibrated strategy in mitigating the reliance of LLMs on local information. For a thorough comparison, we not only employ public JEC-QA and openBookQA datasets, but also construct **EG-QA** which contains **E**nglish **G**rammar multiple-choice question-answering and 14 key knowledge points for assessing self-knowledge and logical reasoning.", "keywords": [ "LLMs", "Evaluation", "Verbalized confidence" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "E2c7UsrZnN", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Ali Tavasoli", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Postdoc" }, { "name": "Behnaz Moradijamei", "gender": "unknown", "institution": "James Madison University", "country": "US", "position": "Assistant Professor" }, { "name": "Heman Shakeri", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Assistant Professor" } ], "n_formula": 35, "n_formula_1": 33, "n_ref_uni": 29, "n_ref": 46, "n_ref_all": 55, "n_fig": 1, "n_tab": 2, "L_tab": 501, "n_element_tab": 43, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 72, "n_element_tab_1": 3, "formula_len_all": 1625, "formula_len_all_1": 2002, "len_all": 95055, "len_all_1": 58641, "len_abs": 1292, "len_title": 152, "len_sents": 30646, "len_sents_1": 27962, "n_sents": 193, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1296, "title": "Spectral Operator Methods for Learning Coherent Temporal Representations in Cellular Signaling Dynamics", "abs": "We present a novel operator-based framework for learning coherent temporal representations of cellular dynamics from live-cell imaging data. Recognizing the inherent stochasticity and measurement limitations in biological systems, our approach shifts the focus from predicting exact trajectories to characterizing key dynamical properties that shape cellular behaviors at the population level. By leveraging spectral analysis of the Koopman operator and smoothing via Markov semigroups of kernel integral operators, we identify near-resonant patterns and transient coherent structures that persist across different experimental conditions. This methodology effectively captures fundamental dynamics, providing insights into mechanisms of heterogeneous cell responses without the need to model precise transformation laws. We demonstrate the efficacy of our framework on a dataset of retinal pigment epithelial cells with an inducible oncogene, revealing conserved dynamical patterns across varying levels of ERK inhibition. Our work offers interpretable learned representations, even with limited and noisy single-cell-resolved recordings, advancing machine learning for dynamical systems and opening new avenues for understanding and predicting cellular behavior in response to external stimuli.", "keywords": [ "Operator theory", "temporal representations", "delay-coordinate embeddings", "Markov operator", "single-cell analysis", "machine learning for dynamical systems" ], "rating_list": [ 5, 3, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 1, 2 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "E2RyjrBMVZ", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Andrew Poulton", "gender": "Male", "institution": "University of Edinburgh", "country": "", "position": "MS student" }, { "name": "Dieuwke Hupkes", "gender": "unknown", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Lovish Madaan", "gender": "unknown", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Pontus Stenetorp", "gender": "Not Specified", "institution": "University College London", "country": "GB", "position": "Full Professor" }, { "name": "Rylan Schaeffer", "gender": "Male", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "PhD student" }, { "name": "Sharan Narang", "gender": "Male", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Aaditya Singh", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 37, "n_ref": 56, "n_ref_all": 109, "n_fig": 12, "n_tab": 4, "L_tab": 1310, "n_element_tab": 147, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1327, "n_element_tab_1": 230, "formula_len_all": 137, "formula_len_all_1": 87, "len_all": 163057, "len_all_1": 64014, "len_abs": 1604, "len_title": 46, "len_sents": 41890, "len_sents_1": 30220, "n_sents": 279, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1613, "title": "Quantifying Variance in Evaluation Benchmarks", "abs": "Evaluation benchmarks are the cornerstone of measuring capabilities of large language models (LLMs), as well as driving progress in said capabilities. Originally designed to make claims about capabilities (or lack thereof) in fully pretrained models, evaluation benchmarks are now also extensively used to decide between various training choices. Despite this widespread usage, we rarely quantify the variance in our evaluation benchmarks, which dictates whether differences in performance are meaningful. Here, we define and measure a range of metrics geared towards measuring variance in evaluation benchmarks, including seed variance across initialisations, and monotonicity during training. By studying a large number of models -- both openly available and pretrained from scratch -- we provide empirical estimates for a variety of variance metrics, with considerations and recommendations for practitioners. We also evaluate the utility and tradeoffs of continuous versus discrete performance measures and explore options for better understanding and reducing this variance. We find that simple changes, such as framing choice tasks (like MMLU) as completion tasks, can often reduce variance for smaller scale (\u223c7B) models, while more involved methods inspired from human testing literature (such as item analysis and item response theory) struggle to meaningfully reduce variance. Overall, our work provides insights into variance in evaluation benchmarks, suggests LM-specific techniques to reduce variance, and more generally encourages practitioners to carefully factor in variance when comparing models.", "keywords": [ "Evaluations", "Language Models", "LLMs" ], "rating_list": [ 3, 3, 5, 6, 5 ], "soundness_list": [ 3, 1, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 4, 2, 3, 3, 3 ] }, { "paper_id": "E2PFv7ad3p", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 13, "author_info_list": [ { "name": "Leyi Yang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Linsheng Lu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Undergrad student" }, { "name": "Qi Zhang", "gender": "Male", "institution": "Fudan University", "country": "", "position": "Full Professor" }, { "name": "Rui Zheng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Tao Gui", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Assistant Professor" }, { "name": "Tao Ji", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Postdoc" }, { "name": "Xiaoran Fan", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Xuanjing Huang", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Yuming Yang", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Yuran Wang", "gender": "Male", "institution": "Honor", "country": "", "position": "Researcher" }, { "name": "Zhiheng Xi", "gender": "unknown", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "xh.zhao", "gender": "Male", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" }, { "name": "Li shuo", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "MS student" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 18, "n_ref": 40, "n_ref_all": 69, "n_fig": 5, "n_tab": 8, "L_tab": 5562, "n_element_tab": 660, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1709, "n_element_tab_1": 197, "formula_len_all": 542, "formula_len_all_1": 360, "len_all": 168259, "len_all_1": 69202, "len_abs": 1141, "len_title": 127, "len_sents": 42372, "len_sents_1": 29733, "n_sents": 404, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1283, "title": "Have the VLMs Lost Confidence? A Study of Sycophancy in VLMs", "abs": "In the study of LLMs, sycophancy represents a prevalent hallucination that poses significant challenges to these models. Specifically, LLMs often fail to adhere to original correct responses, instead blindly agreeing with users' opinions, even when those opinions are incorrect or malicious. However, research on sycophancy in visual language models (VLMs) has been scarce. In this work, we extend the exploration of sycophancy from LLMs to VLMs, introducing the MM-SY benchmark to evaluate this phenomenon. We present evaluation results from multiple representative models, addressing the gap in sycophancy research for VLMs. To mitigate sycophancy, we propose a synthetic dataset for training and employ methods based on prompts, supervised fine-tuning, and DPO. Our experiments demonstrate that these methods effectively alleviate sycophancy in VLMs. Additionally, we probe VLMs to assess the semantic impact of sycophancy and analyze the attention distribution of visual tokens. Our findings indicate that the ability to prevent sycophancy is predominantly observed in higher layers of the model. The lack of attention to image knowledge in these higher layers may contribute to sycophancy, and enhancing image attention at high layers proves beneficial in mitigating this issue.", "keywords": [ "Multi-modal Model", "Visual-Language Model", "Sycophancy", "Hallucination" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 2, 4 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "E2OAT195Le", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Can Rong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "En Xu", "gender": "Male", "institution": "Northwestern Polytechnical University, Northwest Polytechnical University Xi'an", "country": "", "position": "PhD student" }, { "name": "Jingtao Ding", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 29, "n_ref": 33, "n_ref_all": 43, "n_fig": 4, "n_tab": 2, "L_tab": 1452, "n_element_tab": 106, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2654, "n_element_tab_1": 118, "formula_len_all": 491, "formula_len_all_1": 469, "len_all": 95353, "len_all_1": 59679, "len_abs": 1259, "len_title": 147, "len_sents": 30433, "len_sents_1": 26411, "n_sents": 225, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 98, "L_abs": 1263, "title": "A Diffusive Data Augmentation Framework for Reconstruction of Complex Network Evolutionary History", "abs": "The evolutionary processes of complex systems contain critical information about their functional characteristics. The generation time of edges can reveal the historical evolution of various networked complex systems, such as protein-protein interaction networks, ecosystems, and social networks. Recovering these evolutionary processes holds significant scientific value, such as aiding in the interpretation of the evolution of protein-protein interaction networks. However, the scarcity of temporally labeled network data poses challenges for predicting edge generation times under current network structures, leading to issues of insufficient data and significant differences between training and prediction networks. To address this, we introduce a diffusion model that learns the generative mechanisms of networks, producing sufficient augmented network data to effectively mitigate issues of limited and incomplete data. Experimental results demonstrate a 13.7% improvement in prediction accuracy using our approach. Moreover, the model can uniformly predict edge generation times across different types of networks, eliminating the need to retrain the model for each specific network, thus significantly enhancing generalization capability and efficiency.", "keywords": [ "Complex Network", "Temporal Network", "Diffusion Model", "Data Augmentation" ], "rating_list": [ 3, 6, 5, 1 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 2, 4, 4 ] }, { "paper_id": "E2CR6hmV1I", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Fei Huang", "gender": "Male", "institution": "Alibaba Group US", "country": "US", "position": "Senior Research Director" }, { "name": "Ji Zhang", "gender": "unknown", "institution": "Alibaba Group", "country": "CN", "position": "Senior Staff Engineer" }, { "name": "Ming Yan", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Instructor" }, { "name": "Peng Li", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yang Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Yi Fung", "gender": "F", "institution": "Hong Kong University of Science and Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Zhitao He", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zijun Liu", "gender": "Male", "institution": "Department of Computer Science and Technology, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 87, "n_ref_all": 113, "n_fig": 4, "n_tab": 4, "L_tab": 1773, "n_element_tab": 347, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1227, "n_element_tab_1": 159, "formula_len_all": 626, "formula_len_all_1": 627, "len_all": 183856, "len_all_1": 67252, "len_abs": 1663, "len_title": 106, "len_sents": 41830, "len_sents_1": 31575, "n_sents": 277, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1682, "title": "Enhancing Multi-Agent Learning in Real-World Interactive Environments through Process Reward Decomposition", "abs": "LLM-based agents have made significant advancements in interactive environments, such as mobile operations and web browsing, with multi-agent systems further boosting performance. However, current agent learning techniques heavily rely on in-domain data and struggle to generalize across tasks and environments. Moreover, existing multi-agent learning methods are limited by fixed role assignments, which restrict their flexibility and generalization. Furthermore, the multi-step nature of interactive tasks, combined with sparse end-to-end reward signals, hinder effective learning to a great extent. To address these issues, we propose $\\textit{CollabUIAgents}$, a two-stage multi-agent learning framework for interactive environments. In the first stage, the base model is adapted to the environment using curriculum learning on multi-level instruction data. In the second stage, a novel process reward decomposition strategy is introduced during reinforcement learning, allowing rewards to be distributed at both the agent and conversation round levels. This granular feedback fosters collaborative awareness among agents without predefined roles and improves learning efficacy. Experimental results show that our method significantly enhances the performance of multi-agent systems based on open-source models, achieving notable improvements both within and across domains, while also exhibiting strong cross-environment generalization capabilities. Moreover, our best-performing systems achieve results on par with or exceed those of the strong closed-source models, while maintaining the flexibility to be integrated with prompt-based multi-agent systems for future research.", "keywords": [ "Language model", "Muti-agent learning" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "E1m5yGMOiV", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Christoph Flamm", "gender": "unknown", "institution": "Universit\u00e4t Vienna", "country": "AT", "position": "Associate Professor" }, { "name": "Dominik Scheuer", "gender": "Male", "institution": "University of Freiburg", "country": "DE", "position": "Undergrad student" }, { "name": "Frank Hutter", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen & University of Freiburg", "country": "DE", "position": "Full Professor" }, { "name": "Frederic Runge", "gender": "unknown", "institution": "Universit\u00e4t Freiburg", "country": "", "position": "PhD student" }, { "name": "Michael T. Wolfinger", "gender": "unknown", "institution": "Universit\u00e4t Vienna", "country": "AT", "position": "Researcher" }, { "name": "J\u00f6rg Franke", "gender": "Male", "institution": "Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 8, "n_ref_uni": 43, "n_ref": 96, "n_ref_all": 131, "n_fig": 18, "n_tab": 9, "L_tab": 3332, "n_element_tab": 290, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 572, "n_element_tab_1": 43, "formula_len_all": 1144, "formula_len_all_1": 516, "len_all": 184512, "len_all_1": 64285, "len_abs": 1612, "len_title": 136, "len_sents": 65324, "len_sents_1": 32703, "n_sents": 394, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 87, "L_abs": 1617, "title": "KinPFN: Bayesian Approximation of RNA Folding Kinetics using Prior-Data Fitted Networks", "abs": "RNA is a dynamic biomolecule crucial for cellular regulation, with its function largely determined by its folding into complex structures, while misfolding can lead to multifaceted biological sequelae. During the folding process, RNA traverses through a series of intermediate structural states, with each transition occurring at variable rates that collectively influence the time required to reach the functional form. Understanding these folding kinetics is vital for predicting RNA behavior and optimizing applications in synthetic biology and drug discovery. While in silico kinetic RNA folding simulators are often computationally intensive and time-consuming, accurate approximations of the folding times can already be very informative to assess the efficiency of the folding process. In this work, we present KinPFN, a novel approach that leverages prior-data fitted networks to directly model the posterior predictive distribution of RNA folding times. By training on synthetic data representing arbitrary prior folding times, KinPFN efficiently approximates the cumulative distribution function of RNA folding times in a single forward pass, given only a few initial folding time examples. Our method offers a modular extension to existing RNA kinetics algorithms, promising significant computational speed-ups orders of magnitude faster, while achieving comparable results. We showcase the effectiveness of KinPFN through extensive evaluations and real-world case studies, demonstrating its potential for RNA folding kinetics analysis, its practical relevance, and generalization to other biological data.", "keywords": [ "RNA Folding Kinetics", "Prior-Data Fitted Networks", "Deep Learning", "Synthetic Data", "Transformer", "Bayesian Inference" ], "rating_list": [ 3, 6, 8, 6 ], "soundness_list": [ 1, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 1, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "E1Tr7wTlIt", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Siqing Zhang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Wei Sun", "gender": "Male", "institution": "BitEnergy AI", "country": "US", "position": "Researcher" }, { "name": "Yong Liao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Pengyuan Zhou", "gender": "Male", "institution": "Aarhus University", "country": "DK", "position": "Assistant Professor" } ], "n_formula": 31, "n_formula_1": 18, "n_ref_uni": 41, "n_ref": 112, "n_ref_all": 143, "n_fig": 10, "n_tab": 6, "L_tab": 2819, "n_element_tab": 354, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 2720, "n_element_tab_1": 260, "formula_len_all": 3056, "formula_len_all_1": 2023, "len_all": 202860, "len_all_1": 69793, "len_abs": 1300, "len_title": 139, "len_sents": 53718, "len_sents_1": 26455, "n_sents": 498, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1355, "title": "$\\lambda$-SecAgg: Partial Vector Freezing for Lightweight Secure Aggregation in Federated Learning", "abs": "Secure aggregation of user update vectors (e.g. gradients) has become a critical issue in the field of federated learning. Many Secure Aggregation Protocols (SAPs) face exorbitant computation costs, severely constraining their applicability. Given the observation that a considerable portion of SAP's computation burden stems from processing each entry in the private vectors, we propose Partial Vector Freezing (PVF), a portable module for compressing computation costs without introducing additional communication overhead. $\\lambda$-SecAgg, which integrates SAP with PVF, \"freezes\" a substantial portion of the private vector through specific transformations, requiring only $\\frac{1}{\\lambda}$ of the original vector to participate in SAP. Eventually, users can \"thaw\" the public sum of the \"frozen entries\" by the result of SAP. To avoid potential privacy leakage, we devise Disrupting Variables Element for PVF. We demonstrate that PVF can seamlessly integrate with various SAPs and it poses no threat to user privacy in the semi-honest and active adversary settings. We include $7$ baselines, encompassing $5$ distinct types of masking schemes, and explore the acceleration effects of PVF on these SAPs. Empirical investigations indicate that when $\\lambda=100$, PVF yields up to $99.5\\times$ speedup and up to $32.3\\times$ communication reduction.", "keywords": [ "Secure aggregation", "Federated learning" ], "rating_list": [ 1, 6, 5, 5, 3 ], "soundness_list": [ 1, 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 2, 3 ], "contribution_list": [ 1, 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 2, 3, 4 ] }, { "paper_id": "E1SaL8aK7k", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Minzhou Pan", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "PhD student" }, { "name": "Xue Lin", "gender": "Female", "institution": "Northeastern University", "country": "", "position": "Assistant Professor" }, { "name": "Zhaorun Chen", "gender": "Male", "institution": "University of Chicago", "country": "US", "position": "PhD student" }, { "name": "chaowei xiao", "gender": "Male", "institution": "University of Wisconsin - Madison", "country": "US", "position": "Assistant Professor" }, { "name": "PEIYANG XU", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 2, "n_ref_uni": 18, "n_ref": 30, "n_ref_all": 52, "n_fig": 10, "n_tab": 13, "L_tab": 7342, "n_element_tab": 553, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 2110, "n_element_tab_1": 8, "formula_len_all": 477, "formula_len_all_1": 158, "len_all": 209296, "len_all_1": 51044, "len_abs": 2341, "len_title": 133, "len_sents": 53379, "len_sents_1": 25289, "n_sents": 341, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 2228, "title": "SafeVision: Efficient Image Guardrail with Robust Policy Adherence and Explainability", "abs": "As image generation models become increasingly prevalent, the need for efficient and transparent guardrails against unsafe content is more critical than ever. Traditional unsafe image classifiers, limited to predefined categories, often misclassify content due to the pure feature-based learning rather than semantic-based reasoning and struggle to adapt to emerging threats. The time and resources required for retraining on new harmful categories further hinder their ability to respond to evolving threats. To address these challenges, we propose SafeVision, a novel image guardrail system that integrates human-like understanding and reasoning with scalability. Within SafeVision, we propose an effective data collection and generation, policy-following training pipeline, and a customized loss function. In particular, we propose an efficient diverse QA generation and training strategy to enhance the effectiveness of the training process.\nSafeVision is able to follow given safety policies during inference time to guardrail against new risk categories and thus avoid expensive retraining, provide accurate risky content predictions, and provide precise explanations. SafeVision operates in two modes: 1) rapid classification mode, and 2) comprehension mode that provides both classification and human-readable explanations. In addition, considering the limitations of existing unsafe image benchmarks, which contain either only binary or limited categories, we provide VisionHARM-500K, a high-quality unsafe image benchmark comprising over 500k images to cover a wide array of risky categories. This dataset significantly broadens the scope and depth of unsafe image benchmarks. Through comprehensive experiments, we show that SafeVision achieves state-of-the-art performance in both efficiency and accuracy, with an accuracy of 91.77% on the VisionHARM-500K test set (17.77% higher than GPT-4O) and an inference time of 0.0979 seconds per image (over 50 times faster than GPT-4O). SafeVision sets a new standard for comprehensive, policy-following, and explainable image guardrail models, delivering state-of-the-art performance while aligning with human reasoning and enabling scalable adaptation to emerging threats.", "keywords": [ "AI safety", "Large language model", "Multi modality", "Image moderation" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 4, 2, 2, 3 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "E1N1oxd63b", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 13, "author_info_list": [ { "name": "Enshu Liu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Guohao Dai", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Associate Professor" }, { "name": "Haofeng Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Huazhong Yang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Rui Wan", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Shengen Yan", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Shiyao Li", "gender": "Male", "institution": "Infinigence", "country": "CN", "position": "Intern" }, { "name": "Tianchen Zhao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Tongcheng Fang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Widyadewi Soedarmadji", "gender": "unknown", "institution": "Electrical Engineering & Computer Science Department, University of California, Berkeley", "country": "US", "position": "Undergrad student" }, { "name": "Xuefei Ning", "gender": "Not Specified", "institution": "Tsinghua University", "country": "CN", "position": "Research Assistant Professor" }, { "name": "Yu Wang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Zinan Lin", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Senior Researcher" } ], "n_formula": 5, "n_formula_1": 3, "n_ref_uni": 19, "n_ref": 44, "n_ref_all": 83, "n_fig": 20, "n_tab": 3, "L_tab": 2197, "n_element_tab": 434, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1786, "n_element_tab_1": 51, "formula_len_all": 500, "formula_len_all_1": 234, "len_all": 186156, "len_all_1": 57006, "len_abs": 1959, "len_title": 150, "len_sents": 55369, "len_sents_1": 25653, "n_sents": 459, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 101, "L_abs": 1269, "title": "ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation", "abs": "Diffusion transformers have demonstrated remarkable performance in visual generation tasks, such as generating realistic images or videos based on textual instructions. However, larger model sizes and multi-frame processing for video generation lead to increased computational and memory costs, posing challenges for practical deployment on edge devices. Post-Training Quantization (PTQ) is an effective method for reducing memory costs and computational complexity.\nWhen quantizing diffusion transformers, we find that existing quantization methods face challenges when applied to text-to-image and video tasks. To address these challenges, we begin by systematically analyzing the source of quantization error and conclude with the unique challenges posed by DiT quantization. Accordingly, we design an improved quantization scheme: ViDiT-Q (**V**ideo \\& **I**mage **Di**ffusion **T**ransformer **Q**uantization), tailored specifically for DiT models. We validate the effectiveness of ViDiT-Q across a variety of text-to-image and video models, achieving W8A8 and W4A8 with negligible degradation in visual quality and metrics. Additionally, we implement efficient GPU kernels to achieve practical 2-2.5x memory optimization and a 1.4-1.7x end-to-end latency speedup.", "keywords": [ "video generation", "low-bit quantization", "diffusion model" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 5, 2, 2 ] }, { "paper_id": "E1ML0nEReb", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Gao Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Henry Zheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Jiewei Cao", "gender": "Male", "institution": "The University of Adelaide", "country": "", "position": "Postdoc" }, { "name": "Xuchong QIU", "gender": "Male", "institution": "Bosch", "country": "DE", "position": "Researcher" }, { "name": "Yifan Pu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yizeng Han", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yong Xien Chng", "gender": "Not Specified", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 46, "n_ref": 96, "n_ref_all": 119, "n_fig": 9, "n_tab": 9, "L_tab": 2150, "n_element_tab": 293, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1260, "n_element_tab_1": 190, "formula_len_all": 472, "formula_len_all_1": 472, "len_all": 146641, "len_all_1": 63434, "len_abs": 1886, "len_title": 81, "len_sents": 40608, "len_sents_1": 29351, "n_sents": 303, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1899, "title": "Exploring contextual modeling with linear complexity for point cloud segmentation", "abs": "Point cloud segmentation is an important topic in 3D understanding that has traditionally been tackled using either the CNN or Transformer. Recently, Mamba has emerged as a promising alternative, offering efficient long-range contextual modeling capabilities without the quadratic complexity associated with Transformer's attention mechanisms. However, despite Mamba's potential, early efforts have all failed to achieve better performance than the best CNN-based and Transformer-based methods. In this work, we address this challenge by identifying the key components of an effective and efficient point cloud segmentation architecture. Specifically, we show that: 1) Spatial locality and robust contextual understanding are critical for strong performance, and 2) Mamba features linear computational complexity, offering superior data and inference efficiency compared to Transformers, while still being capable of delivering strong contextual understanding. Additionally, we further enhance the standard Mamba specifically for point cloud segmentation by identifying its two key shortcomings. First, the enforced causality in the original Mamba is unsuitable for processing point clouds that have no such dependencies. Second, its unidirectional scanning strategy imposes a directional bias, hampering its ability to capture the full context of unordered point clouds in a single pass. To address these issues, we carefully remove the causal convolutions and introduce a novel Bidirectional Strided SSM to enhance the model's capability to capture spatial relationships. Our efforts culminate in a novel architecture named MEEPO that effectively integrates the strengths of CNN and Mamba. MEEPO surpasses the previous state-of-the-art method, PTv3, by up to +0.8 mIoU on multiple key benchmark datasets, while being 42.1\\% faster and 5.53$\\times$ more memory efficient. Our code will be released.", "keywords": [ "point cloud segmentation", "efficient", "contextual modeling" ], "rating_list": [ 6, 6, 6, 5, 8 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4, 5 ] }, { "paper_id": "E1HLZcRZI1", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Cewu Lu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Jianhua Sun", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Jiude Wei", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Longfei Xu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Nange Wang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Yining Zhang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Undergrad student" }, { "name": "Yuxuan Li", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "MS student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 33, "n_ref": 105, "n_ref_all": 124, "n_fig": 11, "n_tab": 13, "L_tab": 5692, "n_element_tab": 780, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 252, "formula_len_all_1": 298, "len_all": 194120, "len_all_1": 68854, "len_abs": 1826, "len_title": 157, "len_sents": 66458, "len_sents_1": 35316, "n_sents": 512, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 109, "L_abs": 1833, "title": "Arti-PG: A Procedural Toolbox to Synthesize Large-Scale and Diverse Articulated Objects with Rich Annotations", "abs": "The acquisition of substantial volumes of 3D articulated object data is expensive and time-consuming, and consequently the scarcity of 3D articulated object data becomes an obstacle for deep learning methods to achieve remarkable performance in various articulated object understanding tasks. Meanwhile, pairing these object data with detailed annotations to enable training for various tasks is also difficult and labor-intensive to achieve. In order to expeditiously gather a significant number of 3D articulated objects with comprehensive and detailed annotations for training, we propose Articulated Object Procedural Generation toolbox, a.k.a. Arti-PG toolbox. Arti-PG toolbox consists of i) descriptions of articulated objects by means of a generalized structure program along with their analytic correspondence to the objects\u2019 point cloud, ii) procedural rules about manipulations on the structure program to synthesize large-scale and diverse new articulated objects, and iii) mathematical descriptions of knowledge (e.g. affordance, semantics, etc.) to provide annotations to the synthesized object. Arti-PG has two appealing properties for providing training data for articulated object understanding tasks: i) objects are created with unlimited variations in shape through program-oriented structure manipulation, ii) Arti-PG is widely applicable to diverse tasks by easily providing comprehensive and detailed annotations. Arti-PG now supports the procedural generation of 26 categories of articulate objects and provides annotations across a wide range of both vision and manipulation tasks, and we provide exhaustive experiments which fully demonstrate its advantages. We will make Arti-PG toolbox publicly available for the community to use. More details, analysis and discussions are provided in technical appendices.", "keywords": [ "Articulated Object", "Articulated Object Manipulation", "Robotics" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "E1EHO0imOb", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Brian Chmiel", "gender": "Male", "institution": "Intel", "country": "IL", "position": "Researcher" }, { "name": "Daniel Soudry", "gender": "Male", "institution": "Technion - Israel Institute of Technology, Technion", "country": "IL", "position": "Associate Professor" }, { "name": "Maxim Fishman", "gender": "Male", "institution": "Technion, Technion", "country": "", "position": "MS student" }, { "name": "Ron Banner", "gender": "Male", "institution": "Intel", "country": "IL", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 15, "n_ref": 31, "n_ref_all": 51, "n_fig": 11, "n_tab": 4, "L_tab": 664, "n_element_tab": 93, "n_fig_1": 10, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 983, "formula_len_all_1": 860, "len_all": 133230, "len_all_1": 45467, "len_abs": 1142, "len_title": 91, "len_sents": 27149, "len_sents_1": 21186, "n_sents": 211, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 43, "L_abs": 1155, "title": "Scaling FP8 training to trillion-token LLMs", "abs": "We train, for the first time, large language models using FP8 precision on datasets up to 2 trillion tokens --- a 20-fold increase over previous limits. Through these extended training runs, we uncover critical instabilities in FP8 training that were not observable in earlier works with shorter durations. We trace these instabilities to outlier amplification by the SwiGLU activation function. Interestingly, we show, both analytically and empirically, that this amplification happens only over prolonged training periods, and link it to a SwiGLU weight alignment process. To address this newly identified issue, we introduce Smooth-SwiGLU, a novel modification that ensures stable FP8 training without altering function behavior. We also demonstrate, for the first time, FP8 quantization of both Adam optimizer moments. Combining these innovations, we successfully train a 7B parameter model using FP8 precision on 256 Intel Gaudi2 accelerators, achieving on-par results with the BF16 baseline while delivering up to a $\\sim$ 34 % throughput improvement. A reference implementation is supplied in https://github.com/Anonymous1252022/Megatron-DeepSpeed", "keywords": [ "quantization", "fp8", "llms", "training", "acceleration", "compression" ], "rating_list": [ 8, 6, 8, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 2, 3, 3, 4 ] }, { "paper_id": "E1DGY1FXef", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Anuj Diwan", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" }, { "name": "David Harwath", "gender": "Male", "institution": "University of Texas, Austin", "country": "", "position": "Assistant Professor" }, { "name": "Eunsol Choi", "gender": "unknown", "institution": "New York University", "country": "US", "position": "Assistant Professor" }, { "name": "Zhisheng Zheng", "gender": "Male", "institution": "University of Texas at Austin", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 77, "n_ref_all": 107, "n_fig": 7, "n_tab": 5, "L_tab": 1055, "n_element_tab": 123, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 277, "n_element_tab_1": 38, "formula_len_all": 354, "formula_len_all_1": 0, "len_all": 151337, "len_all_1": 64922, "len_abs": 1416, "len_title": 105, "len_sents": 40297, "len_sents_1": 30463, "n_sents": 282, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1326, "title": "Modeling Abstract Style Prompts for Text-to-Speech Models", "abs": "A recent trend in text-to-speech synthesis (TTS) is to construct models capable of generating naturalistic speech that adheres to a textual style prompt describing the speaker's voice and speaking style. In this paper, we propose a crisper definition of style-controlled TTS by categorizing style tags by how they can be collected (*automatic* tags obtainable using signal processing tools e.g. low-pitched and slow; *demographic* tags obtainable using speaker demographics e.g. male and American accent; and *abstract* tags which need human-annotations e.g. authoritative and awed) and what they represent (*intrinsic* tags inherent to speaker identity e.g. gender, average pitch, texture; and *situational* tags specific to utterance-level speaking styles e.g. emotion). Compared to previous work, we expand the space of style prompts substantially by covering 47 abstract tags, 10 demographic tags and 6 automatic tags. For abstract intrinsic tags, we annotate a subset of speakers from the VoxCeleb dataset. For abstract situational tags, we leverage existing speaking-style-based datasets Expresso and EARS. We train a style-prompted TTS model based on Parler-TTS using these datasets and find that our model outperforms baselines on speech-style consistency metrics. Our collected dataset and model will be open-sourced.", "keywords": [ "text-to-speech", "style", "emotion", "datasets" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "E0dTlxy1T4", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 16, "author_info_list": [ { "name": "Fei Huang", "gender": "Male", "institution": "Alibaba Group US", "country": "US", "position": "Senior Research Director" }, { "name": "Haonan Zhang", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Jingkuan Song", "gender": "Male", "institution": "University of Electronic Science and Technology of China,", "country": "CN", "position": "Full Professor" }, { "name": "Lianli Gao", "gender": "Female", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Longze Chen", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Min Yang", "gender": "Female", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Minzheng Wang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Pengpeng Zeng", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Postdoc" }, { "name": "Run Luo", "gender": "Male", "institution": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Xiaobo Xia", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Postdoc" }, { "name": "Xiong Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yongbin Li", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yuchuan Wu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yunshui Li", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Hengtao Shen", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Tony Lin", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 28, "n_ref": 60, "n_ref_all": 88, "n_fig": 19, "n_tab": 3, "L_tab": 2910, "n_element_tab": 201, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 688, "n_element_tab_1": 39, "formula_len_all": 0, "formula_len_all_1": 33, "len_all": 207983, "len_all_1": 65597, "len_abs": 1885, "len_title": 119, "len_sents": 54224, "len_sents_1": 31070, "n_sents": 416, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1770, "title": "MMEvol: Empowering Multimodal Large Language Models with Evol-Instruct", "abs": "The development of Multimodal Large Language Models (MLLMs) has seen significant advancements with increasing demands in various fields (e.g., multimodal\nagents, embodied intelligence). While model-driven approaches attempt to enhance MLLMs capabilities through diverse architectures, the gains have become\nincreasingly marginal. Conversely, data-driven methods, which scale up image-text\ninstruction data, are more effective but face limited data diversity and complexity\nchallenges. The absence of high-quality data constitutes a significant development\nbarrier for MLLMs. To address the data quality bottleneck, we propose MMEvol, a\nnovel multimodal instruction data evolution framework. This framework iteratively\nimprove data quality through a refined combination of fine-grained perception, cognitive reasoning, and interaction evolution, generating a more complex and diverse\nimage-text instruction dataset that empowers MLLMs with enhanced capabilities.\nBeginning with an initial set of instructions, SEED-163K, we utilize MMEvol to\nsystematically broaden the diversity of instruction types, extend visual reasoning\nsteps to improve cognitive reasoning abilities, and thoroughly explore fine-grained\ninformation within images to enhance visual understanding and robustness. To\ncomprehensively evaluate the effectiveness of our approach, we conduct extensive\nqualitative analysis and quantitative experiments across 13 vision-language tasks.\nCompared to baseline models trained with the initial seed data, the results demonstrate that our method achieves an average accuracy improvement of 3.1 percentage\npoints. Furthermore, our approach reaches state-of-the-art (SOTA) performance in\nnine tasks using significantly less data compared to state-of-the-art models.", "keywords": [ "MLLM; MultiModal;Visual Reasoning" ], "rating_list": [ 6, 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "E0UsEIRBQ8", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Baojie Fan", "gender": "Male", "institution": "Nanjing University of Posts and Telecommunications", "country": "CN", "position": "Full Professor" }, { "name": "Jiandong Tian", "gender": "Male", "institution": "The Shenyang Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Wenzhang Zhou", "gender": "Male", "institution": "Nanjing University of Posts and Telecommunications", "country": "CN", "position": "Lecturer" }, { "name": "Xiangzhu Meng", "gender": "Male", "institution": "JD.com", "country": "CN", "position": "Researcher" }, { "name": "caixia xia", "gender": "Female", "institution": "Nanjing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "lianlei shan", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 28, "n_ref": 62, "n_ref_all": 76, "n_fig": 7, "n_tab": 4, "L_tab": 1012, "n_element_tab": 130, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2119, "n_element_tab_1": 170, "formula_len_all": 675, "formula_len_all_1": 677, "len_all": 102896, "len_all_1": 57295, "len_abs": 1589, "len_title": 159, "len_sents": 25286, "len_sents_1": 23127, "n_sents": 163, "n_sents_1": 155, "venue": "ICLR2025", "decision": "withdraw", "L_title": 110, "L_abs": 1548, "title": "Semi-Supervised Underwater Object Detection with Image Enhancement Guided by Attribute-based Data Distribution", "abs": "Semi-supervised underwater object detection aims to improve the performance of detectors on unlabeled underwater images by leveraging knowledge from labeled ones. However, existing methods often overlook the distribution differences between labeled and unlabeled underwater images. In this paper, we propose a novel underwater image enhancement method guided by attribute-based data distribution (UIEG+), which focuses on reducing the discrepancies between enhanced and original unlabeled images across different attributes, thereby effectively addressing the challenges in semi-supervised underwater object detection. Specifically, we explore an underwater image enhancement strategy based on two attributes: color and scale distributions. For the color attribute, we construct a 3-dimensional grid memory, where each grid cell represents a color subspace and records the number of samples in that subspace. Similarly, for the scale attribute, we design a 1-dimensional vector memory that dynamically stores the number of samples in each scale subspace. Subsequently, we propose an effective sampling method to derive parameters for color and scale transformations based on the aforementioned distribution analysis, increasing the likelihood of transformations in low-distribution regions. To evaluate its effetiveness and superiority, massive semi-superivised underwater object deteciton experiments in multiple datasets have been conduted by integrating UIEG+ into existing semi-supervised object detection frameworks. The code will be released.", "keywords": [ "Semi-supervised learning; Underwater object detection" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 1, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 3 ] }, { "paper_id": "E040QmNETN", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Ruiqi Li", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Shengpeng Ji", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Siqi Zheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Xize Cheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Ziang Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 51, "n_ref": 112, "n_ref_all": 126, "n_fig": 4, "n_tab": 7, "L_tab": 2199, "n_element_tab": 95, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2003, "n_element_tab_1": 66, "formula_len_all": 724, "formula_len_all_1": 565, "len_all": 167574, "len_all_1": 68548, "len_abs": 1168, "len_title": 133, "len_sents": 50858, "len_sents_1": 32912, "n_sents": 374, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1174, "title": "MuVi: Video-to-Music Generation with Semantic Alignment and Rhythmic Synchronization", "abs": "Generating music that aligns with the visual content of a video has been a challenging task, as it requires a deep understanding of visual semantics and involves generating music whose melody, rhythm, and dynamics harmonize with the visual narratives. This paper presents MuVi, a novel framework that effectively addresses these challenges to enhance the cohesion and immersive experience of audio-visual content. MuVi analyzes video content through a specially designed visual adaptor to extract contextually and temporally relevant features. These features are used to generate music that not only matches the video\u2019s mood and theme but also its rhythm and pacing. We also introduce a contrastive music-visual pre-training scheme to ensure synchronization, based on the periodicity nature of music phrases. In addition, we demonstrate that our flow-matching-based music generator has in-context learning ability, allowing us to control the style and genre of the generated music. Experimental results show that MuVi demonstrates superior performance in both audio quality and temporal synchronization. The generated music video samples are available at muvi-v2m.github.io.", "keywords": [ "Video-to-music generation", "music generation" ], "rating_list": [ 3, 8, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 3, 3 ] }, { "paper_id": "Dzh0hQPpuf", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Davide Scaramuzza", "gender": "Male", "institution": "University of Zurich", "country": "CH", "position": "Professor" }, { "name": "Jiaxu Xing", "gender": "Male", "institution": "Department of Informatics, University of Zurich, University of Zurich", "country": "CH", "position": "PhD student" }, { "name": "elie aljalbout", "gender": "Male", "institution": "University of Zurich", "country": "CH", "position": "Postdoc" }, { "name": "Nico Amadeo Messikommer", "gender": "Male", "institution": "University of Zurich", "country": "CH", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 39, "n_ref": 60, "n_ref_all": 78, "n_fig": 7, "n_tab": 1, "L_tab": 351, "n_element_tab": 31, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 927, "formula_len_all_1": 990, "len_all": 138048, "len_all_1": 68255, "len_abs": 1434, "len_title": 81, "len_sents": 47742, "len_sents_1": 35088, "n_sents": 324, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 33, "L_abs": 1438, "title": "Student-Informed Teacher Training", "abs": "Imitation learning with a privileged teacher has proven effective for learning complex control behaviors from high-dimensional inputs, such as images. In this framework, a teacher is trained with privileged task information, while a student tries to predict the actions of the teacher with more limited observations, e.g., in a robot navigation task, the teacher might have access to distances to nearby obstacles, while the student only receives visual observations of the scene. However, privileged imitation learning faces a key challenge: the student might be unable to imitate the teacher's behavior due to partial observability. This problem arises because the teacher is trained without considering if the student is capable of imitating the learned behavior. To address this teacher-student asymmetry, we propose a framework for joint training of the teacher and student policies, encouraging the teacher to learn behaviors that can be imitated by the student despite the latters' limited access to information and its partial observability. Based on the performance bound in imitation learning, we add (i) the approximated action difference between teacher and student as a penalty term to the reward function of the teacher, and (ii) a supervised teacher-student alignment step. We motivate our method with a maze navigation task and demonstrate its effectiveness on complex vision-based quadrotor flight and manipulation tasks.", "keywords": [ "Reinforcement Learning", "Imitation Learning", "Robotics" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "DzbUL4AJPP", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Grace Yi", "gender": "Female", "institution": "University of Western Ontario", "country": "CA", "position": "Full Professor" }, { "name": "Wenqing He", "gender": "Male", "institution": "University of Western Ontario", "country": "CA", "position": "Full Professor" }, { "name": "Yuan Bian", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Postdoc" } ], "n_formula": 125, "n_formula_1": 51, "n_ref_uni": 42, "n_ref": 98, "n_ref_all": 128, "n_fig": 3, "n_tab": 2, "L_tab": 423, "n_element_tab": 29, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 11619, "formula_len_all_1": 3172, "len_all": 195185, "len_all_1": 69611, "len_abs": 1408, "len_title": 126, "len_sents": 63309, "len_sents_1": 28990, "n_sents": 588, "n_sents_1": 249, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1417, "title": "Boosting Methods for Interval-censored Data with Regression and Classification", "abs": "Boosting has garnered significant interest across both machine learning and statistical communities. Traditional boosting algorithms, designed for fully observed random samples, often struggle with real-world problems, particularly with interval-censored data. This type of data is common in survival analysis and time-to-event studies where exact event times are unobserved but fall within known intervals. Effective handling of such data is crucial in fields like medical research, reliability engineering, and social sciences. In this work, we introduce novel nonparametric boosting methods for regression and classification tasks with interval-censored data. Our approaches leverages censoring unbiased transformations to adjust loss functions and impute transformed responses while maintaining model accuracy. Implemented via functional gradient descent, these methods ensure scalability and adaptability. We rigorously establish their theoretical properties, including optimality and mean squared error trade-offs, offering solid guarantees. Our proposed methods not only offer a robust framework for enhancing predictive accuracy in domains where interval-censored data are common but also complement existing work, expanding the applicability of boosting techniques. Empirical studies demonstrate robust performance across various finite-sample scenarios, highlighting the practical utility of our approaches.", "keywords": [ "Boosting", "Functional gradient descent", "Interval-censored data", "Minimax error rate", "Nonparametric classification", "Nonparametric regression", "Smoothing spline" ], "rating_list": [ 6, 5, 6, 5, 6 ], "soundness_list": [ 3, 4, 3, 3, 3 ], "presentation_list": [ 3, 2, 4, 3, 4 ], "contribution_list": [ 3, 2, 3, 3, 3 ], "confidence_list": [ 3, 2, 4, 3, 3 ] }, { "paper_id": "Dzamphz35c", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 9, "author_info_list": [ { "name": "Jian Cheng", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Jianchao Yang", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Jingkui Yang", "gender": "Male", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Jun He", "gender": "unknown", "institution": "National University of Defense Technology", "country": "CN", "position": "MS student" }, { "name": "Mei Wen", "gender": "unknown", "institution": "National University of Defense Technology", "country": "CN", "position": "Full Professor" }, { "name": "Xin Ju", "gender": "unknown", "institution": "National University of Defense Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Yasong Cao", "gender": "unknown", "institution": "National University of Defense Technology", "country": "CN", "position": "PhD student" }, { "name": "Zhongdi Luo", "gender": "unknown", "institution": "National University of Defense Technology", "country": "CN", "position": "Undergrad student" }, { "name": "gang li", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" } ], "n_formula": 39, "n_formula_1": 13, "n_ref_uni": 21, "n_ref": 33, "n_ref_all": 54, "n_fig": 5, "n_tab": 6, "L_tab": 1699, "n_element_tab": 338, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 583, "n_element_tab_1": 74, "formula_len_all": 3347, "formula_len_all_1": 1047, "len_all": 119006, "len_all_1": 57901, "len_abs": 1937, "len_title": 127, "len_sents": 39175, "len_sents_1": 27475, "n_sents": 272, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1983, "title": "Ultra-Low Accumulation Precision Inference with Block Floating Point Arithmetic", "abs": "Block Floating Point (BFP) quantization offers a hardware-efficient numerical range trade-off. Previous studies have quantized weights and activations to an extremely low precision using the BFP arithmetic. However, as the precision of weights and activations is reduced, we have identified that accumulation becomes a hardware bottleneck in the BFP MAC. Nevertheless, existing attempts to decrease the precision of accumulation in matrix multiplication have generally preserved model performance through training with a pre-selected, fixed accumulation precision. Nonetheless, selecting an unduly low precision leads to notable performance degradation, and these studies lack an effective approach to establish the lower precision limit, potentially incurring considerable training costs. Hence, we propose a statistical method to analyze the impact of reduced accumulation precision on the inference of deep learning applications. Due to the presence of fixed-point accumulation and floating-point accumulation in BFP matrix multiplication, we have formulated a set of equations to relate the data range of fixed-point multiply-accumulate operations and the effects of floating-point swamping to the parameters of BFP quantization, the length of accumulation, model weights, and the minimum number of bits required for accumulation, thereby determining the appropriate accumulation precision. Applied to MMLU Llama2-7B, SQuAD-v1.1 BERT-Large and BERT-Base and CIFAR-10 ResNet-50, our precision settings yield performance close to the FP32 baseline. Meanwhile, further precision reduction degrades performance, indicating our approach\u2019s proximity to precision limits. Guided by our equations, the hardware exhibited a 13.7\\%-28.7\\% enhancement in area and power efficiency over high-precision accumulation under identical quantization configuration, and it demonstrated a $10.3\\times$ area reduction and an $11.0\\times$ power reduction compared to traditional BFP16 implementations.", "keywords": [ "accumulation precision; block floating-point quantization; MAC; deep learning" ], "rating_list": [ 3, 5, 6, 1 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 1, 3, 2, 1 ], "contribution_list": [ 1, 2, 3, 1 ], "confidence_list": [ 2, 4, 2, 4 ] }, { "paper_id": "DzKdjWe59v", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Didier Ch\u00e9telat", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CA", "position": "Researcher" }, { "name": "Mark Coates", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Full Professor" }, { "name": "Soumyasundar Pal", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "Yingxue Zhang", "gender": "Female", "institution": "Huawei Canada, Huawei Noah's Ark Lab", "country": "CA", "position": "Researcher" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 30, "n_ref": 80, "n_ref_all": 118, "n_fig": 5, "n_tab": 5, "L_tab": 614, "n_element_tab": 24, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 692, "formula_len_all_1": 389, "len_all": 174600, "len_all_1": 61855, "len_abs": 981, "len_title": 116, "len_sents": 47864, "len_sents_1": 29461, "n_sents": 431, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 983, "title": "Hint Marginalization for Improved Reasoning in Large Language Models", "abs": "Large Language Models (LLMs) have exhibited an impressive capability to perform reasoning tasks, especially if they are encouraged to generate a sequence of intermediate steps. Reasoning performance can be improved by suitably combining multiple LLM responses, generated either in parallel in a single query, or via sequential interactions with LLMs throughout the reasoning process. Existing strategies for combination, such as self-consistency and progressive-hint-prompting, make inefficient usage of the LLM responses. We present Hint Marginalization, a novel and principled algorithmic framework to enhance the reasoning capabilities of LLMs. Our approach can be viewed as an iterative sampling strategy for forming a Monte Carlo approximation of an underlying distribution of answers, with the goal of identifying the mode the most likely answer. Empirical evaluation on several benchmark datasets for arithmetic reasoning demonstrates the superiority of the proposed approach.", "keywords": [ "reasoning", "large language models" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "DzGe40glxs", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "David Krueger", "gender": "Male", "institution": "Montreal Institute for Learning Algorithms, University of Montreal, Universit\u00e9 de Montr\u00e9al", "country": "CA", "position": "Assistant Professor" }, { "name": "Stephen Chung", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Thomas Bush", "gender": "Male", "institution": "independent", "country": "GB", "position": "Researcher" }, { "name": "Usman Anwar", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Adria Garriga-Alonso", "gender": "Male", "institution": "FAR", "country": "", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 0, "n_ref_uni": 69, "n_ref": 120, "n_ref_all": 228, "n_fig": 58, "n_tab": 8, "L_tab": 6742, "n_element_tab": 382, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 734, "n_element_tab_1": 36, "formula_len_all": 524, "formula_len_all_1": 0, "len_all": 402718, "len_all_1": 66461, "len_abs": 1321, "len_title": 115, "len_sents": 160472, "len_sents_1": 32672, "n_sents": 1124, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 67, "L_abs": 1362, "title": "Interpreting Emergent Planning in Model-Free Reinforcement Learning", "abs": "We present the first mechanistic evidence that model-free reinforcement learning agents can learn to plan. This is achieved by applying a methodology based on concept-based interpretability to a model-free agent in Sokoban -- a commonly used benchmark for studying planning. Specifically, we demonstrate that DRC, a generic model-free agent introduced by [Guez et al. (2019)](https://arxiv.org/abs/1901.03559), uses learned concept representations to internally formulate plans that both predict the long-term effects of actions on the environment and influence action selection. Our methodology involves: (1) probing for planning-relevant concepts, (2) investigating plan formation within the agent's representations, and (3) verifying that discovered plans (in agent's representations) have causal effect on agent's behavior through interventions. We also show that the emergence of these plans coincides with the emergence of a planning-like property: the ability to benefit from additional test-time compute. Finally, we perform a qualitative analysis of the planning algorithm learned by the agent and discover a strong resemblance to parallelized bidirectional search. Our findings advance understanding of the internal mechanisms underlying planning behavior in agents, enabling improved diagnosis, interpretation, and control of agent planning processes.", "keywords": [ "reinforcement learning", "interpretability", "planning", "probes", "model-free", "mechanistic interpretability", "sokoban" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "DyyLUUVXJ5", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Chenyang Zhang", "gender": "Not Specified", "institution": "Meta", "country": "", "position": "Researcher" }, { "name": "Ding Liu", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Haozhe Liu", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Kumara Kahatapitiya", "gender": "Male", "institution": "Meta", "country": "US", "position": "Intern" }, { "name": "Menglin Jia", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Michael S Ryoo", "gender": "Male", "institution": "Salesforce AI Research", "country": "US", "position": "Principal Researcher" }, { "name": "Sen He", "gender": "Male", "institution": "Meta AI", "country": "GB", "position": "Researcher" }, { "name": "Tian Xie", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Research Scientist" } ], "n_formula": 9, "n_formula_1": 11, "n_ref_uni": 82, "n_ref": 188, "n_ref_all": 216, "n_fig": 7, "n_tab": 1, "L_tab": 1441, "n_element_tab": 114, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 3062, "n_element_tab_1": 323, "formula_len_all": 422, "formula_len_all_1": 423, "len_all": 219666, "len_all_1": 71341, "len_abs": 1265, "len_title": 120, "len_sents": 43799, "len_sents_1": 29439, "n_sents": 356, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1322, "title": "Adaptive Caching for Faster Video Generation with Diffusion Transformers", "abs": "Generating temporally-consistent high-fidelity videos can be computationally expensive, especially over longer temporal spans. More-recent Diffusion Transformers (DiTs)--- despite making significant headway in this context--- have only heightened such challenges as they rely on larger models and heavier attention mechanisms, resulting in slower inference speeds. In this paper, we introduce a $\\textit{training-free}$ method to accelerate video DiTs, termed Adaptive Caching ($\\textit{AdaCache}$), which is motivated by the fact that $\\textit{``not all videos are created equal''}$: meaning, some videos require fewer denoising steps to attain a reasonable quality than others. Building on this, we not only cache computations through the diffusion process, but also devise a caching schedule tailored to each video generation, maximizing the quality-latency trade-off. We further introduce a Motion Regularization ($\\textit{MoReg}$) scheme to utilize video information within AdaCache, essentially controlling the compute allocation based on motion content. Altogether, our plug-and-play contributions grant significant inference speedups (e.g. up to 4.7x on Open-Sora 720p - 2s video generation) without sacrificing the generation quality, across multiple video DiT baselines. Our code will be made publicly-available.", "keywords": [ "Diffusion Transformers", "Caching", "Content-adaptive Generation" ], "rating_list": [ 3, 5, 8, 6 ], "soundness_list": [ 2, 4, 4, 2 ], "presentation_list": [ 2, 4, 4, 2 ], "contribution_list": [ 3, 4, 3, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "Dyo2tS5A8b", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Atoosa Chegini", "gender": "Female", "institution": "University of Maryland, College Park", "country": "", "position": "PhD student" }, { "name": "Hamid Kazemi", "gender": "Male", "institution": "Apple", "country": "US", "position": "Applied Machine Lerarning Engineer" }, { "name": "Jonas Geiping", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen", "country": "DE", "position": "Principal Researcher" }, { "name": "Soheil Feizi", "gender": "Male", "institution": "University of Maryland, College Park", "country": "", "position": "Assistant Professor" }, { "name": "Tom Goldstein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 23, "n_ref": 41, "n_ref_all": 66, "n_fig": 12, "n_tab": 4, "L_tab": 2110, "n_element_tab": 139, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 1714, "n_element_tab_1": 126, "formula_len_all": 146, "formula_len_all_1": 146, "len_all": 97358, "len_all_1": 61404, "len_abs": 850, "len_title": 92, "len_sents": 31890, "len_sents_1": 27928, "n_sents": 204, "n_sents_1": 196, "venue": "ICLR2025", "decision": "withdraw", "L_title": 44, "L_abs": 616, "title": "What do we learn from inverting CLIP models?", "abs": "We employ an inversion-based approach to examine CLIP models. Our examination reveals that inverting CLIP models results in the generation of images that exhibit semantic alignment with the specified target prompts. We leverage these inverted images to gain insights into various aspects of CLIP models, such as their ability to blend concepts and inclusion of gender biases. We notably observe instances of NSFW (Not Safe For Work) images during model inversion. This phenomenon occurs even for semantically innocuous prompts, like `a beautiful landscape,' as well as for prompts involving the names of celebrities.", "keywords": [ "CLIP; NSFW; Interpretability; Gender Bias" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 1, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "DydCqKa6AH", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Bolei Zhou", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Assistant Professor" }, { "name": "Joe Lin", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Undergrad student" }, { "name": "Wayne Wu", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Researcher" }, { "name": "Zhizheng Liu", "gender": "Male", "institution": "UCLA Computer Science Department, University of California, Los Angeles", "country": "US", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 56, "n_ref": 90, "n_ref_all": 98, "n_fig": 4, "n_tab": 3, "L_tab": 1599, "n_element_tab": 209, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 911, "n_element_tab_1": 171, "formula_len_all": 152, "formula_len_all_1": 152, "len_all": 150044, "len_all_1": 70085, "len_abs": 220, "len_title": 131, "len_sents": 36372, "len_sents_1": 33295, "n_sents": 269, "n_sents_1": 248, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 83, "L_abs": 1424, "title": "Learning to Generate Diverse Pedestrian Movements from Web Videos with Noisy Labels", "abs": "Understanding and modeling pedestrian movements in the real world is crucial for applications like motion forecasting and scene simulation. Many factors influence pedestrian movements, such as scene context, individual characteristics, and goals, which are often ignored by the existing human generation methods. Web videos contain natural pedestrian behavior and rich motion context, but annotating them with pre-trained predictors leads to noisy labels. In this work, we propose learning diverse pedestrian movements from web videos. We first curate a large-scale dataset called CityWalkers that captures diverse real-world pedestrian movements in urban scenes. Then, based on CityWalkers, we propose a generative model called PedGen for diverse pedestrian movement generation. PedGen introduces automatic label filtering to remove the low-quality labels and a mask embedding to train with partial labels. It also contains a novel context encoder that lifts the 2D scene context to 3D and can incorporate various context factors in generating realistic pedestrian movements in urban scenes. Experiments show that PedGen outperforms existing baseline methods for pedestrian movement generation by learning from noisy labels and incorporating the context factors. In addition, PedGen achieves zero-shot generalization in both real-world and simulated environments. The code, model, and data will be made publicly available.", "keywords": [ "Pedestrian Movement Analysis", "Human Motion Dataset", "Human Motion Generation" ], "rating_list": [ 8, 5, 6, 8 ], "soundness_list": [ 4, 3, 4, 2 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "DxT3e2f1jc", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Songhua Liu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Xingyi Yang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zhenxiong Tan", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Xinchao Wang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 11, "n_ref_uni": 23, "n_ref": 62, "n_ref_all": 79, "n_fig": 4, "n_tab": 5, "L_tab": 1243, "n_element_tab": 171, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1920, "n_element_tab_1": 166, "formula_len_all": 431, "formula_len_all_1": 442, "len_all": 105573, "len_all_1": 62589, "len_abs": 0, "len_title": 49, "len_sents": 31199, "len_sents_1": 26066, "n_sents": 261, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 1520, "title": "Video-Infinity: Distributed Long Video Generation", "abs": "Diffusion models have recently achieved remarkable results for video generation. Despite the encouraging performances, the generated videos are typically constrained to a small number of frames, resulting in clips lasting merely a few seconds. The primary challenges in producing longer videos include the substantial memory requirements and the extended processing time required on a single GPU. A straightforward solution would be to split the workload across multiple GPUs, which, however, leads to two issues: (1) ensuring all GPUs communicate effectively to share timing and context information, and (2) modifying existing video diffusion models, which are usually trained on short sequences, to create longer videos without additional training. To tackle these, in this paper we introduce Video-Infinity, a distributed inference pipeline that enables parallel processing across multiple GPUs for long-form video generation. Specifically, we propose two coherent mechanisms: Clip parallelism and Dual-scope attention. Clip parallelism optimizes the gathering and sharing of context information across GPUs which minimizes communication overhead, while Dual-scope attention modulates the temporal self-attention to balance local and global contexts efficiently across the devices. Together, the two mechanisms join forces to distribute the workload and enable the fast generation of long videos. Under an 8 x Nvidia 6000 Ada GPU (48G) setup, our method generates videos up to 2,300 frames in approximately 5 minutes.", "keywords": [ "diffusion model", "video generation" ], "rating_list": [ 6, 3, 3, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "DwqoBkj2Mw", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Abhradeep Guha Thakurta", "gender": "Male", "institution": "Google", "country": "", "position": "Senior Research Scientist" }, { "name": "Adam Smith", "gender": "Male", "institution": "Boston University", "country": "US", "position": "Full Professor" }, { "name": "Andreas Terzis", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Arun Ganesh", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Borja Balle", "gender": "unknown", "institution": "DeepMind", "country": "", "position": "Research scientist" }, { "name": "Christopher A. Choquette-Choo", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Jamie Hayes", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Matthew Jagielski", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Milad Nasr", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Thomas Steinke", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" } ], "n_formula": 19, "n_formula_1": 10, "n_ref_uni": 38, "n_ref": 66, "n_ref_all": 94, "n_fig": 10, "n_tab": 4, "L_tab": 453, "n_element_tab": 58, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 46, "n_element_tab_1": 15, "formula_len_all": 2132, "formula_len_all_1": 1163, "len_all": 134844, "len_all_1": 65633, "len_abs": 1258, "len_title": 158, "len_sents": 43560, "len_sents_1": 29965, "n_sents": 334, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 110, "L_abs": 1108, "title": "The Last Iterate Advantage: Empirical Auditing and Principled Heuristic Analysis of Differentially Private SGD", "abs": "We propose a simple heuristic privacy analysis of noisy clipped stochastic gradient descent (DP-SGD) in the setting where only the last iterate is released and the intermediate iterates remain hidden. Namely, our heuristic assumes a linear structure for the model.\n\nWe show experimentally that our heuristic is predictive of the outcome of privacy auditing applied to various training procedures. Thus it can be used prior to training as a rough estimate of the final privacy leakage. We also probe the limitations of our heuristic by providing some artificial counterexamples where it underestimates the privacy leakage.\n\nThe standard composition-based privacy analysis of DP-SGD effectively assumes that the adversary has access to all intermediate iterates, which is often unrealistic. However, this analysis remains the state of the art in practice. While our heuristic does not replace a rigorous privacy analysis, it illustrates the large gap between the best theoretical upper bounds and the privacy auditing lower bounds and sets a target for further work to improve the theoretical privacy analyses.", "keywords": [ "differential privacy", "heuristics", "privacy auditing" ], "rating_list": [ 5, 8, 5, 5 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 5, 5 ] }, { "paper_id": "DwiwOcK1B7", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Vladim\u00edr Bo\u017ea", "gender": "Male", "institution": "Powerful Medical", "country": "SK", "position": "Researcher" }, { "name": "Vladim\u00edr Macko", "gender": "Not Specified", "institution": "Comenius University in Bratislava", "country": "SK", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 11, "n_ref_uni": 48, "n_ref": 87, "n_ref_all": 97, "n_fig": 6, "n_tab": 4, "L_tab": 1874, "n_element_tab": 224, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1496, "n_element_tab_1": 168, "formula_len_all": 1228, "formula_len_all_1": 980, "len_all": 133231, "len_all_1": 63592, "len_abs": 1044, "len_title": 149, "len_sents": 35327, "len_sents_1": 27309, "n_sents": 314, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 101, "L_abs": 1048, "title": "Two Sparse Matrices are Better than One: Sparsifying Neural Networks with Double Sparse Factorization", "abs": "Neural networks are often challenging to work with due to their large size and complexity. To address this, various methods aim to reduce model size by sparsifying or decomposing weight matrices, such as magnitude pruning and low-rank or block-diagonal factorization. In this work, we present Double Sparse Factorization (DSF), where we factorize each weight matrix into two sparse matrices. Although solving this problem exactly is computationally infeasible, we propose an efficient heuristic based on alternating minimization via ADMM that achieves state-of-the-art results, enabling unprecedented sparsification of neural networks. For instance, in a one-shot pruning setting, our method can reduce the size of the LLaMA2-13B model by 50% while maintaining better performance than the dense LLaMA2-7B model. We also compare favorably with Optimal Brain Compression, the state-of-the-art layer-wise pruning approach for convolutional neural networks. Furthermore, accuracy improvements of our method persist even after further model fine-tuning.", "keywords": [ "sparse factorization", "pruning" ], "rating_list": [ 5, 8, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "DvU9ijSn1v", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 8, "author_info_list": [ { "name": "Chenguang Wang", "gender": "Male", "institution": "State University of New York at Stony Brook", "country": "US", "position": "MS student" }, { "name": "Fuxiao Liu", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Hongyu Zhao", "gender": "Male", "institution": "University of Chicago", "country": "", "position": "MS student" }, { "name": "Ming Li", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Pei Chen", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "", "position": "PhD student" }, { "name": "Tianyi Zhou", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Yijun Liang", "gender": "Female", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "YuPeng Hou", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Software Development Engineer" } ], "n_formula": 10, "n_formula_1": 7, "n_ref_uni": 32, "n_ref": 72, "n_ref_all": 96, "n_fig": 5, "n_tab": 7, "L_tab": 3284, "n_element_tab": 305, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2729, "n_element_tab_1": 241, "formula_len_all": 743, "formula_len_all_1": 627, "len_all": 277788, "len_all_1": 67469, "len_abs": 1188, "len_title": 124, "len_sents": 46614, "len_sents_1": 31208, "n_sents": 292, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1195, "title": "Mosaic-IT: Free Compositional Data Augmentation Improves Instruction Tuning", "abs": "Finetuning large language models with a variety of instruction-response pairs has enhanced their capability to understand and follow instructions. Current instruction tuning primarily relies on teacher models or human intervention to generate and refine the instructions and responses for training, which are costly, non-sustainable, and may lack diversity. In this paper, we introduce Mosaic Instruction Tuning (Mosaic-IT), a human/model-free compositional data augmentation method that can efficiently create rich and diverse augmentations from existing instruction tuning data to enhance the LLMs. Mosaic-IT randomly concatenates multiple instruction data into one and trains the model to produce the corresponding responses with predefined higher-level meta-instructions to strengthen its multi-step instruction-following and format-following skills. Our extensive evaluations demonstrate a superior performance and training efficiency of Mosaic-IT, which achieves consistent performance improvements over various benchmarks and a $80\\%$ reduction in training costs compared with original instruction tuning. Our codes and data are available at https://anonymous.4open.science/r/mosaic-955B.", "keywords": [ "Large Language Model", "Instruction Tuning", "Supervised Finetuning", "Data Augmentaion" ], "rating_list": [ 5, 6, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "DuyuAHBk1t", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Guimeng Liu", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Researcher" }, { "name": "Milad Abdollahzadeh", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Postdoc" }, { "name": "Ngai-man Cheung", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "SG", "position": "Associate Professor" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 31, "n_ref": 88, "n_ref_all": 107, "n_fig": 10, "n_tab": 9, "L_tab": 3297, "n_element_tab": 282, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 29, "n_element_tab_1": 6, "formula_len_all": 1222, "formula_len_all_1": 1114, "len_all": 160443, "len_all_1": 57756, "len_abs": 1870, "len_title": 116, "len_sents": 42170, "len_sents_1": 26951, "n_sents": 336, "n_sents_1": 222, "venue": "ICLR2025", "decision": "withdraw", "L_title": 68, "L_abs": 1607, "title": "AIR: Zero-shot Generative Model Adaptation with Iterative Refinement", "abs": "Zero-shot generative model adaptation (ZSGM) aims to adapt a pre-trained generator to a target domain using only text guidance and without any samples from the target domain.\nCentral to recent ZSGM approaches are *directional loss* which use the text guidance in the form of aligning the image offset with text offset in the embedding space of a vision-language model like CLIP.\nThis is similar to the analogical reasoning in NLP where the offset between one pair of words is used to identify a missing element in another pair by aligning the offset between these two pairs.\nHowever, a major limitation of existing ZSGM methods is that the learning objective assumes the complete alignment between image offset and text offset in the CLIP embedding space. \n**Our work** makes two main contribution.\nInspired by the offset misalignment studies in NLP, as our first contribution, we perform an empirical study to analyze the misalignment between text offset and image offset in CLIP embedding space for various large publicly available datasets.\nOur important finding is that offset misalignment in CLIP embedding space is correlated with concept distance, *i.e.*, close concepts have a less offset misalignment.\nTo address the limitations of the current approaches, as our second contribution, we propose Adaptaiotn with Iterative Refinement (AIR) which mitigates the offset misalignment issue in directional loss by iteratively selecting anchor points closer to the target domain.\nExtensive experimental results show that the proposed AIR approach achieves SOTA performance across various adaptation setups.", "keywords": [ "Zero-shot Generative Model Adaptation", "Transfer Learning", "Prompt Learning", "Multi-modal Representation Space" ], "rating_list": [ 5, 8, 6, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 3, 5 ] }, { "paper_id": "Duuerhutvq", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Patrick Pynadath", "gender": "Male", "institution": ", Purdue University", "country": "US", "position": "PhD student" }, { "name": "Ruqi Zhang", "gender": "Female", "institution": "Purdue University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 22, "n_formula_1": 14, "n_ref_uni": 37, "n_ref": 118, "n_ref_all": 146, "n_fig": 8, "n_tab": 7, "L_tab": 2283, "n_element_tab": 176, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 102, "n_element_tab_1": 9, "formula_len_all": 1350, "formula_len_all_1": 672, "len_all": 167137, "len_all_1": 63462, "len_abs": 1423, "len_title": 108, "len_sents": 55169, "len_sents_1": 29987, "n_sents": 413, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1436, "title": "Controlled LLM Decoding via Discrete Auto-regressive Biasing", "abs": "Controlled text generation allows for enforcing user-defined constraints on large language model outputs, an increasingly important field as LLMs become more prevalent in everyday life. One common approach uses energy-based decoding, which defines a target distribution through an energy function that combines multiple constraints into a weighted average. However, these methods often struggle to balance fluency with constraint satisfaction, even with extensive tuning of the energy function's coefficients. In this paper, we identify that this suboptimal balance arises from sampling in continuous space rather than the natural discrete space of text tokens. To address this, we propose \\emph{Discrete Auto-regressive Biasing}, a controlled decoding algorithm that leverages gradients while operating entirely in the discrete text domain.\nSpecifically, we introduce a new formulation for controlled text generation by defining a joint distribution over the generated sequence and an auxiliary bias sequence. To efficiently sample from this joint distribution, we propose a Langevin-within-Gibbs sampling algorithm using gradient-based discrete MCMC. Our method significantly improves constraint satisfaction while maintaining comparable or better fluency, all with lower computational costs. We demonstrate the advantages of our controlled decoding method on sentiment control, language detoxification, and keyword-guided generation.", "keywords": [ "LLMs", "controlled decoding", "MCMC" ], "rating_list": [ 8, 3, 6, 6 ], "soundness_list": [ 4, 2, 2, 2 ], "presentation_list": [ 4, 1, 2, 3 ], "contribution_list": [ 4, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "DumcCxxzka", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "Frank Hutter", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen & University of Freiburg", "country": "DE", "position": "Full Professor" }, { "name": "Frederic Runge", "gender": "unknown", "institution": "Universit\u00e4t Freiburg", "country": "", "position": "PhD student" }, { "name": "Sharat Patil", "gender": "Male", "institution": "Humboldt Universit\u00e4t Berlin", "country": "DE", "position": "Researcher" }, { "name": "J\u00f6rg Franke", "gender": "Male", "institution": "Universit\u00e4t Freiburg", "country": "DE", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 0, "n_ref_uni": 61, "n_ref": 103, "n_ref_all": 140, "n_fig": 8, "n_tab": 15, "L_tab": 4140, "n_element_tab": 517, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 827, "n_element_tab_1": 50, "formula_len_all": 440, "formula_len_all_1": 0, "len_all": 180335, "len_all_1": 67356, "len_abs": 1215, "len_title": 109, "len_sents": 45284, "len_sents_1": 33448, "n_sents": 321, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1224, "title": "RNAinformer: Generative RNA Design with Tertiary Interactions", "abs": "The function of an RNA molecule depends on its structure and a strong structure-to-function relationship is already achieved on the secondary structure level of RNA. Therefore, the secondary structure based design of RNAs is one of the major challenges in computational biology. A common approach of RNA design is inverse RNA folding. However, existing RNA design approaches cannot invert all folding algorithms because they cannot represent all types of base interactions. In this work, we propose RNAinformer, a novel generative transformer based approach to the inverse RNA folding problem. Leveraging axial-attention, we directly model the secondary structure input represented as an adjacency matrix in a 2D latent space, which allows us to invert all existing secondary structure prediction algorithms. Consequently, RNAinformer is the first model capable of designing RNAs from secondary structures with all base interactions, including non-canonical base pairs and tertiary interactions like pseudoknots and base multiplets. We demonstrate RNAinformer\u2019s state-of-the-art performance across different RNA design benchmarks and showcase its novelty by inverting different RNA secondary structure prediction algorithms.", "keywords": [ "RNA", "RNA Design", "RNA Inverse Folding", "Transformers", "Generative Design", "Axial Attention", "pseduoknots", "multiplets" ], "rating_list": [ 5, 6, 5, 5, 5 ], "soundness_list": [ 2, 4, 3, 2, 2 ], "presentation_list": [ 3, 4, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "DugT77rRhW", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Jianan Wang", "gender": "Female", "institution": "Astribot Inc", "country": "CN", "position": "Researcher" }, { "name": "Xianbiao Qi", "gender": "Male", "institution": "AITC", "country": "CN", "position": "Principal Researcher" }, { "name": "Xiangyu Yue", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Xili Dai", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Yaxuan Huang", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Researcher" }, { "name": "Yixing Yuan", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 8, "n_ref_uni": 38, "n_ref": 87, "n_ref_all": 136, "n_fig": 17, "n_tab": 4, "L_tab": 860, "n_element_tab": 127, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 1187, "n_element_tab_1": 100, "formula_len_all": 967, "formula_len_all_1": 357, "len_all": 135897, "len_all_1": 61301, "len_abs": 382, "len_title": 125, "len_sents": 38705, "len_sents_1": 27412, "n_sents": 331, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1394, "title": "Unposed Sparse Views Room Layout Reconstruction in the Age of Pretrain Model", "abs": "Multiple-perspective images room layout reconstruction is poorly investigated due to the tedious and cumbersome steps that emerge from multi-view geometry, e.g camera intrinsic/extrinsic estimation, image matching, and triangulation. However, with the advancement of the current 3D foundation model DUSt3R, a paradigm shift has occurred in the 3D reconstruction realm, moving from a multi-step Structure-from-motion approach to an end-to-end single-step reconstruction without error accumulation and instability.\nTo this end, this paper employs the 3D foundation model DUSt3R in the room layout reconstruction task, naming it \\ours{}. As the name suggests, \\ours{} incorporates the DUSt3R framework and plane representation, then fine-tunes on a room layout dataset (Structure3D). With the uniform and parsimonious results of \\ours{}, we can obtain the final room layout with only a single post-processing step and 2D detection results. Compared to previous room layout reconstruction methods, it relaxes the setting from a single perspective/panorama image to multiple perspective images. Moreover, \\ours{} solves the task in an end-to-end paradigm without cumbersome steps and accumulated errors. Finally, experiments show that our \\ours{} not only surpasses the state-of-the-art on the synthesis dataset Structure3D but also demonstrates the robustness and performs well in real-world cases.", "keywords": [ "layout reconstruction", "holistic 3D representation", "large 3D model." ], "rating_list": [ 6, 5, 6, 10 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 4 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "DtFCIfvAFc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Hongru Yan", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Intern" }, { "name": "Yu Zheng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Yueqi Duan", "gender": "Male", "institution": "Department of Electronic Engineering, Tsinghua University", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 23, "n_formula_1": 16, "n_ref_uni": 36, "n_ref": 119, "n_ref_all": 166, "n_fig": 13, "n_tab": 9, "L_tab": 1876, "n_element_tab": 254, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 582, "n_element_tab_1": 85, "formula_len_all": 1452, "formula_len_all_1": 934, "len_all": 135483, "len_all_1": 65756, "len_abs": 651, "len_title": 120, "len_sents": 40408, "len_sents_1": 29736, "n_sents": 301, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 71, "L_abs": 1431, "title": "Gaussian-Det: Learning Closed-Surface Gaussians for 3D Object Detection", "abs": "Skins wrapping around our bodies, leathers covering over the sofa, sheet metal coating the car \u2013 it suggests that objects are enclosed by a series of continuous surfaces, which provides us with informative geometry prior for objectness deduction. In this paper, we propose Gaussian-Det which leverages Gaussian Splatting as surface representation for multi-view based 3D object detection. Unlike existing monocular or NeRF-based methods which depict the objects via discrete positional data, Gaussian-Det models the objects in a continuous manner by formulating the input Gaussians as feature descriptors on a mass of partial surfaces. Furthermore, to address the numerous outliers inherently introduced by Gaussian splatting, we accordingly devise a Closure Inferring Module (CIM) for the comprehensive surface-based objectness deduction. CIM firstly estimates the probabilistic feature residuals for partial surfaces given the underdetermined nature of Gaussian Splatting, which are then coalesced into a holistic representation on the overall surface closure of the object proposal. In this way, the surface information Gaussian-Det exploits serves as the prior on the quality and reliability of objectness and the information basis of proposal refinement. Experiments on both synthetic and real-world datasets demonstrate that Gaussian-Det outperforms various existing approaches, in terms of both average precision and recall.", "keywords": [ "3D Gaussian Splatting", "3D Object Detection", "Surface Closure" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 1, 3, 1 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 3, 3, 4, 5 ] }, { "paper_id": "DtATVd5NLc", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Aditi Sheshadri", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Aman Gupta", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Johannes Schmude", "gender": "unknown", "institution": "International Business Machines", "country": "US", "position": "Researcher" }, { "name": "Manil Maskey", "gender": "Male", "institution": "University of Alabama at Huntsville", "country": "US", "position": "Researcher" }, { "name": "Rahul Ramachandran", "gender": "Male", "institution": "NASA/MSFC", "country": "US", "position": "Principal Researcher" }, { "name": "Sujit Roy", "gender": "Male", "institution": "NASA IMPACT", "country": "UM", "position": "Researcher" }, { "name": "Vishal Gaur", "gender": "Male", "institution": "University of Alabama at Huntsville", "country": "US", "position": "Researcher" }, { "name": "Wei Ji Leong", "gender": "Male", "institution": "Development Seed", "country": "US", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 39, "n_ref": 80, "n_ref_all": 97, "n_fig": 12, "n_tab": 1, "L_tab": 672, "n_element_tab": 144, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 949, "n_element_tab_1": 147, "formula_len_all": 305, "formula_len_all_1": 303, "len_all": 152716, "len_all_1": 56966, "len_abs": 1682, "len_title": 127, "len_sents": 34562, "len_sents_1": 26011, "n_sents": 264, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1701, "title": "Finetuning Weather Foundation Models to Develop Climate Model Parameterizations", "abs": "Climate prediction models parameterize a range of atmospheric-oceanic processes like clouds, turbulence, and gravity waves. These physical parameterizations are a leading source of uncertainty and strongly influence future projections of global temperature rise. We present a fresh approach to developing parameterizations for coarse-climate models by leveraging pre-trained AI foundation models (FMs) for weather and climate. A pre-trained encoder and decoder from a 2.3 billion parameter FM (NASA and IBM's Prithvi WxC) --- which contains a latent probabilistic representation of atmospheric evolution --- is fine-tuned to create a data-driven predictor of atmospheric gravity waves (GWs). Current climate models are not fine enough to resolve GWs. We create an ML-based parameterization that learns GW fluxes from high-resolution ``GW resolving\" climate models to represent them in \"GW missing\" coarse-climate models. The fluxes predicted by our fine-tuned model are comprehensively evaluated using a set of three tests. Comparison with a baseline (Attention U-Net) reveals the superior predictive performance of the fine-tuned model throughout the atmosphere. The model outperforms the baseline even in regions excluded from the FM pre-training. This is quantified using the Hellinger distance which is 0.11 for the baseline and 0.06, i.e., roughly half, for the fine-tuned model. FMs are largely unexplored in climate science. Our findings emphasize their versatility and reusability to accomplish a range of weather- and climate-related downstream applications, especially in a low-data regime. These FMs can be further leveraged to create new parameterizations for other earth-system processes.", "keywords": [ "Atmospheric Dynamics", "Parameterizations", "Climate Modelling", "Foundation Model", "ERA5", "Finetuning", "Machine Learning" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 1, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "DsW4boRh8H", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Jiayi Ma", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Kaining Zhang", "gender": "Female", "institution": "Wuhan University", "country": "CN", "position": "PhD student" }, { "name": "Paolo Favaro", "gender": "Male", "institution": "Institute f\u00fcr Informatik, University of Bern", "country": "CH", "position": "Full Professor" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 39, "n_ref": 90, "n_ref_all": 111, "n_fig": 8, "n_tab": 6, "L_tab": 1144, "n_element_tab": 175, "n_fig_1": 8, "n_tab_1": 1, "L_tab_1": 38, "n_element_tab_1": 7, "formula_len_all": 1006, "formula_len_all_1": 950, "len_all": 121552, "len_all_1": 67323, "len_abs": 2204, "len_title": 101, "len_sents": 31896, "len_sents_1": 30388, "n_sents": 261, "n_sents_1": 247, "venue": "ICLR2025", "decision": "withdraw", "L_title": 53, "L_abs": 1382, "title": "GFNet: Homography Estimation via Grid Flow Regression", "abs": "Current deep homography estimation methods are constrained to processing image pairs with limited resolution due to restrictions in network architecture and computational capacity. For larger images, downsampling is often necessary, which can significantly degrade estimation accuracy. To address this limitation, we propose GFNet, a Grid Flow regression Network that consistently delivers high-accuracy homography estimates across varying image resolutions. Unlike previous methods that directly regress the parameters of the global homography between two views, GFNet directly estimates flow over a coarse grid and then uses the resulting correspondences to compute the homography. This approach not only supports high-resolution processing but also preserves the high accuracy of dense matching while significantly reducing the computational load typically associated with such frameworks, thanks to the use of coarse grid flow. We demonstrate the effectiveness of GFNet on a wide range of experiments on multiple datasets, including the common scene MSCOCO, multimodal datasets VIS-IR and GoogleMap, and the dynamic scene VIRAT. In specific, on GoogleMap, GFNet achieves an improvement of +9.9\\% in auc@3 while reducing MACs by $\\sim$47\\% compared to the SOTA dense matching method. Additionally, it shows a 1.7$\\times$ improvement in auc@3 over the SOTA deep homography method.", "keywords": [ "homography estimation", "multimodal", "image matching" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 5, 5 ] }, { "paper_id": "DsMxVELk3K", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Anika Tasnim Rodela", "gender": "Female", "institution": "United International University", "country": "BD", "position": "Lecturer" }, { "name": "Chowdhury Mofizur Rahman", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Mahbub E Sobhani", "gender": "Male", "institution": "United International University", "country": "BD", "position": "MS student" }, { "name": "Swakkhar Shatabda", "gender": "unknown", "institution": "BRAC University", "country": "BD", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 57, "n_ref_all": 61, "n_fig": 0, "n_tab": 6, "L_tab": 770, "n_element_tab": 109, "n_fig_1": 0, "n_tab_1": 4, "L_tab_1": 1605, "n_element_tab_1": 91, "formula_len_all": 260, "formula_len_all_1": 268, "len_all": 116030, "len_all_1": 70424, "len_abs": 1382, "len_title": 144, "len_sents": 37616, "len_sents_1": 33923, "n_sents": 251, "n_sents_1": 234, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1389, "title": "TextEconomizer: Enhancing Lossy Text Compression with Denoising Autoencoder and Entropy Coding", "abs": "Lossy text compression reduces data size while preserving core meaning, making it ideal for summarization, automated analysis, and digital archives where exact fidelity is less critical. While extensively used in image compression, text compression techniques, such as integrating entropy coding with autoencoder latent representations in Seq2Seq text generation, have been underexplored. A key challenge is incorporating lossless entropy coding into denoising autoencoders to improve storage efficiency while maintaining high-quality outputs, even with noisy text. Prior studies have mainly focused on near-lossless token generation with little attention to space efficiency. In this paper, we present a denoising autoencoder with a rectified latent representation that compresses variable-sized inputs into a fixed-size latent space without prior knowledge of dataset dimensions. By leveraging entropy coding, our model achieves state-of-the-art compression ratios alongside competitive text quality, as measured by diverse metrics. Its parameter count is approximately 196 times smaller than comparable models. Additionally, it achieves a compression ratio of 67\u00d7 while maintaining high BLEU and ROUGE scores. This significantly outperforms existing transformer-based models in memory efficiency, marking a breakthrough in balancing lossless compression with optimal space optimization.", "keywords": [ "Text Compression", "Denoising AutoEnccoder", "Lossy Text", "Entropy Coding", "Latent Space" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 2, 4, 5 ] }, { "paper_id": "DsIOUoZkVk", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Benjamin Eysenbach", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Assistant Professor" }, { "name": "Yongwei Che", "gender": "Male", "institution": "Princeton University", "country": "US", "position": "Undergrad student" } ], "n_formula": 24, "n_formula_1": 17, "n_ref_uni": 23, "n_ref": 43, "n_ref_all": 68, "n_fig": 14, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2547, "formula_len_all_1": 1387, "len_all": 139662, "len_all_1": 66896, "len_abs": 1194, "len_title": 144, "len_sents": 50627, "len_sents_1": 31989, "n_sents": 379, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1192, "title": "The \"Law'' of the Unconscious Contrastive Learner: Probabilistic Alignment of Unpaired Modalities", "abs": "While internet-scale data often come in pairs (e.g., audio+image, image+text), we often want to perform inferences over modalities unseen together in the training data (e.g., audio+text). Prior work has addressed this issue by learning multiple contrastive embedding spaces between existing modality pairs, implicitly hoping that unseen modality pairs will end up being aligned. This theoretical paper proves that this hope is well founded, under certain assumptions. Starting with the proper Bayesian approach of integrating out intermediate modalities, we show that directly comparing the representations of data from unpaired modalities can recover the same likelihood ratio. Our analysis builds on prior work on the geometry and probabilistic interpretation of contrastive representations, showing how these representations can answer many of the same inferences as probabilistic graphical models. Our analysis suggests two new ways of using contrastive representations: in settings with pre-trained contrastive models, and for handling language ambiguity in reinforcement learning. Our numerical experiments study the importance of our assumptions and demonstrate these new applications.", "keywords": [ "theory", "contrastive learning", "probabilistic graphical models", "multi-modal learning", "reinforcement learning" ], "rating_list": [ 5, 5, 8, 5 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 2, 2, 3 ] }, { "paper_id": "DrNN5qx66Z", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Daniel Cores", "gender": "Not Specified", "institution": "University of Santiago de Compostela", "country": "ES", "position": "Assistant Professor" }, { "name": "Manuel Mucientes", "gender": "Male", "institution": "Universidad de Santiago de Compostela", "country": "ES", "position": "Full Professor" }, { "name": "Michael Dorkenwald", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Cees Snoek", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "Yuki Asano", "gender": "Male", "institution": "University of Technology Nuremberg", "country": "DE", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 75, "n_ref_all": 107, "n_fig": 10, "n_tab": 4, "L_tab": 373, "n_element_tab": 68, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 481, "n_element_tab_1": 70, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 134237, "len_all_1": 61634, "len_abs": 1463, "len_title": 94, "len_sents": 40835, "len_sents_1": 30563, "n_sents": 328, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 46, "L_abs": 1467, "title": "TVBench: Redesigning Video-Language Evaluation", "abs": "Large language models have demonstrated impressive performance when integrated with vision models even enabling video understanding. However, evaluating these video models presents its own unique challenges, for which several benchmarks have been proposed. In this paper, we show that the currently most used video-language benchmarks can be solved without requiring much temporal reasoning. We identified three main issues in existing datasets: (i) static information from single frames is often sufficient to solve the tasks (ii) the text of the questions and candidate answers is overly informative, allowing models to answer correctly without relying on any visual input (iii) world knowledge alone can answer many of the questions, making the benchmarks a test of knowledge replication rather than visual reasoning. In addition, we found that open-ended question-answering benchmarks for video understanding suffer from similar issues while the automatic evaluation process with LLMs is unreliable, making it an unsuitable alternative. As a solution, we propose TVBench, a novel open-source video multiple-choice question-answering benchmark, and demonstrate through extensive evaluations that it requires a high level of temporal understanding. Surprisingly, we find that most recent state-of-the-art video-language models perform similarly to random performance on TVBench, with only a few models such as Qwen2-VL, and Tarsier clearly surpassing this baseline.", "keywords": [ "Video-Language evaluation", "Video-Language benchmark" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 4, 2, 3, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "DqU4AB4wRy", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Guandong Xu", "gender": "Male", "institution": "The Education University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Nitin Bisht", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Xiuwen Gong", "gender": "Female", "institution": "University of Sydney", "country": "", "position": "Postdoc" } ], "n_formula": 46, "n_formula_1": 22, "n_ref_uni": 29, "n_ref": 63, "n_ref_all": 77, "n_fig": 5, "n_tab": 8, "L_tab": 5041, "n_element_tab": 1209, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 5988, "n_element_tab_1": 947, "formula_len_all": 3196, "formula_len_all_1": 1471, "len_all": 164628, "len_all_1": 76776, "len_abs": 1503, "len_title": 90, "len_sents": 48964, "len_sents_1": 28329, "n_sents": 369, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 42, "L_abs": 1515, "title": "GUARANTEED USER FAIRNESS IN RECOMMENDATION", "abs": "Although recommender systems (RS) have been well-developed for various fields of applications,\nthey suffer from the crisis of platform credibility with respect to RS confidence and fairness, which\nmay drive users away from the platform and result in the failure of the platform\u2019s long-term success.\nIn recent years, a few works have tried to solve either the model confidence or fairness issue,\nwhile there is no statistical guarantee for these methods. It is therefore an urgent need to solve\nboth issues with a unifying framework with statistical guarantee. In this paper, we propose a novel\nand reliable framework called Guaranteed User Fairness in Recommendation (GUFR) to dynamically\ngenerate prediction sets for users across various groups, which are guaranteed 1) to include\nthe ground-truth items with user-predefined high confidence/probability (e.g., 90%); 2) to ensure\nuser fairness across different groups; 3) to have the minimum average set size. We further design an\nefficient algorithm named Guaranteed User Fairness Algorithm (GUFA) to optimize the proposed\nmethod, and upper bounds of the risk and fairness metric are derived to help speed up the optimization\nprocess. Moreover, we provide rigorous theoretical analysis with respect to risk and fairness\ncontrol as well as the minimum set size. Extensive experiments also validate the effectiveness of the\nproposed framework, which aligns with our theoretical analysis. The code is publicly available at\nhttps://anonymous.4open.science/r/GUFR-76EC.", "keywords": [ "Recommendation Systems", "Fairness in RS", "Conformal Prediction" ], "rating_list": [ 6, 6, 5, 3 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 5, 4, 2 ] }, { "paper_id": "Dq9VrVuLzV", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Bingbing Liu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "R&D Manager" }, { "name": "Leheng Li", "gender": "Male", "institution": "The Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Qing LIAN", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Weichao Qiu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" }, { "name": "Xu Yan", "gender": "Male", "institution": "The Chinese University of Hongkong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Ying-Cong Chen", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Yingjie CAI", "gender": "Female", "institution": "Huawei Technologies Ltd.", "country": "CN", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 36, "n_ref": 88, "n_ref_all": 119, "n_fig": 16, "n_tab": 5, "L_tab": 1116, "n_element_tab": 78, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 934, "n_element_tab_1": 77, "formula_len_all": 511, "formula_len_all_1": 405, "len_all": 150139, "len_all_1": 56957, "len_abs": 1259, "len_title": 134, "len_sents": 45256, "len_sents_1": 26519, "n_sents": 365, "n_sents_1": 206, "venue": "ICLR2025", "decision": "withdraw", "L_title": 86, "L_abs": 1264, "title": "SyntheOcc: Synthesize Geometric-Controlled Street View Images through 3D Semantic MPIs", "abs": "The advancement of autonomous driving is increasingly reliant on high-quality annotated datasets, especially in the task of 3D occupancy prediction, where the occupancy labels require dense 3D annotation with significant human effort. In this paper, we propose SyntheOcc, which denotes a diffusion model that Synthesize photorealistic and geometric-controlled images by conditioning Occupancy labels in driving scenarios. This yields an unlimited amount of diverse, annotated, and controllable datasets for applications like training perception models and simulation. SyntheOcc addresses the critical challenge of how to efficiently encode 3D geometric information as conditional input to a 2D diffusion model. Our approach innovatively incorporates 3D semantic multi-plane images (MPIs) to provide comprehensive and spatially aligned 3D scene descriptions for conditioning. As a result, SyntheOcc can generate photorealistic multi-view images and videos that faithfully align with the given geometric labels (semantics in 3D voxel space). Extensive qualitative and quantitative evaluations of SyntheOcc on the nuScenes dataset prove its effectiveness in generating controllable occupancy datasets that serve as an effective data augmentation to perception models.", "keywords": [ "Autonomous Driving", "Image Generation", "Data-centric AI", "3D Vision" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "Dpqw0namg3", "primary_area": "generative models", "n_author": 13, "author_info_list": [ { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Huan Wang", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Jake Grigsby", "gender": "Male", "institution": "University of Texas at Austin", "country": "", "position": "PhD student" }, { "name": "Jianguo Zhang", "gender": "Male", "institution": "SalesForce AI Research", "country": "US", "position": "Researcher" }, { "name": "Juan Carlos Niebles", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Director, Research" }, { "name": "Michael S Ryoo", "gender": "Male", "institution": "Salesforce AI Research", "country": "US", "position": "Principal Researcher" }, { "name": "Ming Zhu", "gender": "Female", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Shelby Heinecke", "gender": "Female", "institution": "Salesforce Research", "country": "US", "position": "Researcher" }, { "name": "Shirley Kokane", "gender": "Female", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Silvio Savarese", "gender": "Male", "institution": "Salesforce", "country": "", "position": "Chief Scientist" }, { "name": "Thai Quoc Hoang", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Machine Learning Engineer" }, { "name": "Tian Lan", "gender": "unknown", "institution": "SalesForce", "country": "", "position": "Researcher" }, { "name": "Zuxin Liu", "gender": "Male", "institution": "Salesforce AI Research", "country": "US", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 19, "n_ref": 38, "n_ref_all": 49, "n_fig": 4, "n_tab": 6, "L_tab": 1768, "n_element_tab": 173, "n_fig_1": 3, "n_tab_1": 7, "L_tab_1": 2457, "n_element_tab_1": 212, "formula_len_all": 158, "formula_len_all_1": 0, "len_all": 133850, "len_all_1": 62335, "len_abs": 1376, "len_title": 158, "len_sents": 46538, "len_sents_1": 28351, "n_sents": 329, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1382, "title": "LAM Simulator: Advancing Large Action Model Training for Agent via Online Exploration and Feedback Simulation", "abs": "Large Action Models (LAMs) for AI agents have significant potential, but their development is often constrained by the reliance on supervised learning and manual data curation, which are both time-consuming and costly. To address these limitations, we present the LAM Simulator, a comprehensive framework designed for online exploration of agentic tasks with high-quality feedback. This framework includes a curated set of high-quality agentic tasks, a diverse collection of tools, and an interactive environment where agent models can call tools, receive execution responses, and obtain action feedback. Our findings indicate that the LAM Simulator significantly enhances model performance and effectively identifies and addresses potential issues. Specifically, our model, LAM-Sim-8x7B, demonstrates an 18.54\\% improvement over its base LAM and significantly outperforms other state-of-the-art alternatives on ToolEval benchmark. Furthermore, we have demonstrated that LLMs lacking in agentic capability can greatly benefit from the implementation of LAM Simulator. Our experiments with a model trained on Mixtral-8x7B-Instruct-v0.1 have yielded a doubling to tripling of performance. Remarkably, the data construction process for training these models requires minimal human intervention, making the LAM Simulator a robust framework for accelerating the development of AI agents.", "keywords": [ "LLMs Agent; Self-learning", "Reinforcement Learning; Data Generation" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 4, 2 ] }, { "paper_id": "DpnY7VOktT", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Bao Gia Doan", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Postdoc" }, { "name": "Damith Ranasinghe", "gender": "unknown", "institution": "University of Adelaide", "country": "AU", "position": "Associate Professor" }, { "name": "Ehsan M Abbasnejad", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Associate Professor" }, { "name": "Viet Vo", "gender": "Male", "institution": "University of Adelaide", "country": "AU", "position": "Postdoc" } ], "n_formula": 47, "n_formula_1": 16, "n_ref_uni": 47, "n_ref": 83, "n_ref_all": 121, "n_fig": 7, "n_tab": 19, "L_tab": 8966, "n_element_tab": 873, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 1130, "n_element_tab_1": 97, "formula_len_all": 4209, "formula_len_all_1": 967, "len_all": 218041, "len_all_1": 65632, "len_abs": 2127, "len_title": 127, "len_sents": 62640, "len_sents_1": 29463, "n_sents": 478, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 79, "L_abs": 1528, "title": "Can Model Randomization Offer Robustness Against Query-Based Black-Box Attacks?", "abs": "Deep neural networks are misguided by simple-to-craft, imperceptible adversarial perturbations to inputs. Now, it is possible to craft such perturbations solely using model outputs and black-box attack algorithms. These algorithms compute adversarial examples by iteratively querying a model and inspecting responses. Attacks success in near information vacuums pose a significant challenge for developing\nmitigations. We investigate a new idea for a defense driven by a fundamental insight\u2014to compute an adversarial example, attacks depend on the relationship between successive responses to queries to optimize a perturbation. Therefore, to obfuscate this relationship, we investigate randomly sampling a model from a set to generate a response to a query. Effectively, this model randomization violates the attacker's expectation of the unknown parameters of a model to remain static between queries to extract information to guide the search toward an adversarial example. It is not immediately clear if model randomization can lead to sufficient obfuscation to confuse query-based black-box attacks or how such a method could be built. Our theoretical analysis proves model randomization always increases resilience to query-based black-box attacks. We demonstrate with extensive empirical studies using 6 state-of-the-art attacks under all three perturbation objectives ($l_\\infty, l_2, l_0$) and adaptive attacks, our proposed method injects sufficient uncertainty through obfuscation to yield a highly effective defense.", "keywords": [ "query-based black-box attacks", "model randomness", "diversity", "adversarial defense", "trustworthy machine learning", "safety and responsible AI" ], "rating_list": [ 6, 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 2, 2, 5 ] }, { "paper_id": "DpOQwOzTc2", "primary_area": "applications to neuroscience & cognitive science", "n_author": 6, "author_info_list": [ { "name": "Jianwen Sun", "gender": "Male", "institution": "Central China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Qirong Chen", "gender": "Male", "institution": "Central China Normal University", "country": "CN", "position": "MS student" }, { "name": "Ruxia Liang", "gender": "Female", "institution": "Central China Normal University", "country": "", "position": "Assistant Professor" }, { "name": "Xiaoxuan Shen", "gender": "Male", "institution": "Central China Normal University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yawei Luo", "gender": "Male", "institution": "Zhejiang University", "country": "NA", "position": "Researcher" }, { "name": "Zhihai Hu", "gender": "Male", "institution": "Central China Normal University", "country": "CN", "position": "MS student" } ], "n_formula": 27, "n_formula_1": 8, "n_ref_uni": 40, "n_ref": 77, "n_ref_all": 92, "n_fig": 5, "n_tab": 7, "L_tab": 3840, "n_element_tab": 368, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1357, "n_element_tab_1": 179, "formula_len_all": 1061, "formula_len_all_1": 404, "len_all": 172818, "len_all_1": 69410, "len_abs": 2462, "len_title": 183, "len_sents": 59914, "len_sents_1": 32693, "n_sents": 414, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 132, "L_abs": 2527, "title": "Combining Denoised Neural Network and Genetic Symbolic Regression for Memory Behavior Modeling via Dynamic Asynchronous Optimization", "abs": "Memory behavior modeling is a critical topic in cognitive psychology and education. Traditional psychological approaches describe the dynamic properties of memory through memory equations derived from experimental data, but these models often lack accuracy and are frequently debated in terms of their form. In recent years, data-driven modeling methods have improved predictive accuracy but often suffer from poor interpretability, limiting their ability to provide deeper cognitive insights. While knowledge-informed neural network models have achieved significant success in fields such as physics, their application in behavior modeling remains limited. This paper proposes a Self-evolving Psychology-informed Neural Network (SPsyINN), which leverages classical memory equations as knowledge modules to constrain neural network training. To address challenges such as the difficulty in quantifying descriptors and the limited interpretability of classical memory equations, a genetic symbolic regression algorithm is introduced to conduct evolutionary searches for more optimal expressions based on classical memory equations, enabling the mutual progress of the knowledge module and the neural network module. Specifically, the proposed approach combines genetic symbolic regression and neural networks in a parallel training framework, with a dynamic joint optimization loss function ensuring effective knowledge alignment between the two modules. Then, for addressing the training efficiency differences arising from the distinct optimization methods and computational hardware requirements of genetic algorithms and neural networks, an asynchronous interaction mechanism mediated by proxy data is developed to facilitate effective communication between modules and improve optimization efficiency. Finally, a denoising module is integrated into the neural network to enhance robustness against data noise and improve generalization performance. Experimental results on four large-scale real-world memory behavior demonstrate that SPsyINN outperforms state-of-the-art methods in predictive accuracy. Ablation studies further show that the proposed approach effectively achieves mutual progress between different modules, improving model predictive accuracy while uncovering more interpretable memory equations, highlighting the potential application value of SPsyINN in psychological research. Our code is released at: \\href{https://anonymous.4open.science/r/SPsyINN-3F18}{https://anonymous.4open.science/r/SPsyINN-3F18}", "keywords": [ "Memory behavior", "asynchronous optimization", "neural networks", "genetic symbolic regression" ], "rating_list": [ 5, 8, 3, 6 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 5, 2, 5, 3 ] }, { "paper_id": "DpLFmc09pC", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "KE WANG", "gender": "unknown", "institution": "University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Shay B Cohen", "gender": "Male", "institution": "University of Edinburgh", "country": "GB", "position": "Reader" } ], "n_formula": 25, "n_formula_1": 8, "n_ref_uni": 51, "n_ref": 74, "n_ref_all": 98, "n_fig": 6, "n_tab": 9, "L_tab": 1756, "n_element_tab": 173, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1113, "n_element_tab_1": 39, "formula_len_all": 5907, "formula_len_all_1": 562, "len_all": 164277, "len_all_1": 60930, "len_abs": 1553, "len_title": 114, "len_sents": 44978, "len_sents_1": 28519, "n_sents": 303, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1560, "title": "DEPfold: RNA Secondary Structure Prediction as Dependency Parsing.", "abs": "RNA secondary structure prediction is critical for understanding RNA function but remains challenging due to complex structural elements like pseudoknots and limited training data. We introduce DEPfold, a novel deep learning approach that reframes RNA secondary structure prediction as a dependency parsing problem. DEPfold presents three key innovations: (1) a biologically motivated transformation of RNA structures into labeled dependency trees, (2) a biaffine attention mechanism for joint prediction of base pairings and their types, and (3) an optimal tree decoding algorithm that enforces valid RNA structural constraints. Unlike traditional energy-based methods, DEPfold learns directly from annotated data and leverages pretrained language models to capture intricate RNA patterns. We evaluate DEPfold on both within-family and cross-family RNA datasets, demonstrating significant performance improvements over existing methods. On the RNAStrAlign dataset, DEPfold achieves an F$_1$ score of 0.985, predicting pseudoknots and long-range interactions. Notably, DEPfold shows strong performance in cross-family generalization when trained on data augmented by traditional energy-based models, outperforming existing methods on the bpRNA-new dataset. This demonstrates DEPfold's ability to effectively learn structural information beyond what traditional methods capture. Our approach bridges natural language processing (NLP) with RNA biology, providing a computationally efficient and adaptable tool for advancing RNA structure prediction and analysis.", "keywords": [ "RNA secondary structure prediction", "Dependency parsing", "Biaffine attention", "Pseudoknots", "Pretrained Model", "Deep learning" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "Dolm7rrrQd", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Arjun Nichani", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Haewon Jeong", "gender": "unknown", "institution": "University of California, Santa Barbara", "country": "US", "position": "Assistant Professor" }, { "name": "Rasta Tadayon", "gender": "Not Specified", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Tian Qiu", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 6, "n_ref_uni": 50, "n_ref": 92, "n_ref_all": 108, "n_fig": 7, "n_tab": 2, "L_tab": 1497, "n_element_tab": 182, "n_fig_1": 8, "n_tab_1": 3, "L_tab_1": 551, "n_element_tab_1": 139, "formula_len_all": 379, "formula_len_all_1": 433, "len_all": 148265, "len_all_1": 62600, "len_abs": 1327, "len_title": 138, "len_sents": 42741, "len_sents_1": 31288, "n_sents": 296, "n_sents_1": 205, "venue": "ICLR2025", "decision": "withdraw", "L_title": 90, "L_abs": 1288, "title": "Gone With the Bits: Revealing Racial Bias in Low-Rate Neural Compression for Facial Images", "abs": "Neural compression methods are gaining popularity due to their impressive rate-distortion performance and their ability to compress data to extremely small bitrates, below 0.1 bits per pixel (bpp). As deep learning architectures, these models are prone to bias during the training process, potentially leading to unfair outcomes for individuals in different groups. In this paper, we present a general, structured, scalable framework for evaluating bias in neural image compression models. Using this framework, we investigate racial bias in neural compression algorithms by analyzing 7 popular models and their variants. Through this investigation we first demonstrate that traditional distortion metrics are ineffective in capturing bias in neural compression models. Next, we highlight that racial bias is present in all neural compression models and can be captured by examining facial phenotype degradation in image reconstructions. Additionally, we reveal a task-dependent correlation between bias and model architecture. We then examine the relationship between bias and realism in the image reconstructions and demonstrate a trade-off across models. Finally, we show that utilizing a racially balanced training set can reduce bias but is not a sufficient bias mitigation strategy.", "keywords": [ "Fairness", "Bias", "Neural Compression", "Phenotype Classification" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Dojny642Dy", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Chengqing Li", "gender": "Male", "institution": "Xiangtan University", "country": "CN", "position": "Full Professor" }, { "name": "Dun Tan", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "MS student" }, { "name": "Jianfeng Dong", "gender": "Male", "institution": "Zhejiang Gongshang University", "country": "CN", "position": "Full Professor" }, { "name": "Jingjing Chen", "gender": "Female", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Ning Han", "gender": "Male", "institution": "Xiangtan University", "country": "CN", "position": "Lecturer" }, { "name": "Sijie Yang", "gender": "Male", "institution": "Hunan University", "country": "CN", "position": "MS student" }, { "name": "Yawen Zeng", "gender": "Female", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Zemin Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "long shaohua", "gender": "Male", "institution": "Xiangtan University", "country": "CN", "position": "MS student" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 39, "n_ref": 87, "n_ref_all": 109, "n_fig": 9, "n_tab": 3, "L_tab": 347, "n_element_tab": 79, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1467, "n_element_tab_1": 56, "formula_len_all": 177, "formula_len_all_1": 178, "len_all": 153468, "len_all_1": 73447, "len_abs": 1290, "len_title": 143, "len_sents": 36764, "len_sents_1": 31418, "n_sents": 310, "n_sents_1": 264, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1208, "title": "IVCR-200K: A Large-Scale Benchmark for Interactive Video Corpus Retrieval", "abs": "In recent years, significant developments have been made in both video retrieval and video moment retrieval tasks, which respectively retrieve complete videos or moments for a given text query. These advancements have greatly improved user satisfaction during the search process. However, previous work has failed to establish meaningful \"interaction\" between the retrieval system and the user, and its one-way retrieval paradigm can no longer fully meet the personalization and dynamics needs of at least 80.8% of users.\n\nIn this paper, we introduce a more realistic setting, the Interactive Video Corpus Retrieval task (IVCR) that enables multi-turn, conversational, realistic interactions between the user and the retrieval system. To facilitate research on this challenging task, we introduce IVCR-200K, a bilingual, multi-turn, conversational, abstract semantic high-quality dataset that supports video retrieval and even moment retrieval. Furthermore, we propose a comprehensive framework based on multi-modal large language models (MLLMs) to support users' several interaction modes with more explainable solutions. Our extensive experiments demonstrate the effectiveness of our dataset and framework.", "keywords": [ "Interactive Video Corpus Retrieval Dataset; Cross-Modal Video Retrieval; Multi-Modal Large Language Model" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 5 ] }, { "paper_id": "DoDNJdDntB", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Benjamin Holzschuh", "gender": "Male", "institution": "School of Computation, Information and Technology, Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Nils Thuerey", "gender": "Male", "institution": "Technical University Munich", "country": "DE", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 13, "n_ref_uni": 40, "n_ref": 103, "n_ref_all": 123, "n_fig": 14, "n_tab": 5, "L_tab": 849, "n_element_tab": 89, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 528, "n_element_tab_1": 42, "formula_len_all": 566, "formula_len_all_1": 590, "len_all": 170947, "len_all_1": 65414, "len_abs": 1123, "len_title": 109, "len_sents": 47279, "len_sents_1": 30775, "n_sents": 354, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1170, "title": "Flow Matching for Posterior Inference with Simulator Feedback", "abs": "Flow-based generative modeling is a powerful tool for solving inverse problems in physical sciences that can be used for sampling and likelihood evaluation with much lower inference times than traditional methods. We propose to refine flows with additional control signals based on a simulator. Control signals can include gradients and a problem-specific cost function if the simulator is differentiable, or they can be fully learned from the simulator output. \nIn our proposed method, we pretrain the flow network and include feedback from the simulator exclusively for finetuning, therefore requiring only a small amount of additional parameters and compute. We motivate our design choices on several benchmark problems for simulation-based inference and evaluate flow matching with simulator feedback against classical MCMC methods for modeling strong gravitational lens systems, a challenging inverse problem in astronomy. We demonstrate that including feedback from the simulator improves the accuracy by $53$%, making it competitive with traditional techniques while being up to 67x faster for inference. Upon acceptance, we will make our code publicly available.", "keywords": [ "generative modeling", "simulation-based inference", "astronomy" ], "rating_list": [ 6, 6, 3, 3, 3 ], "soundness_list": [ 3, 2, 1, 1, 3 ], "presentation_list": [ 3, 2, 2, 2, 2 ], "contribution_list": [ 3, 2, 1, 2, 2 ], "confidence_list": [ 4, 2, 5, 3, 4 ] }, { "paper_id": "DoB8DmrsSS", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Feidi Liu", "gender": "unknown", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "Xiaolin Sun", "gender": "Male", "institution": "Tulane University", "country": "US", "position": "PhD student" }, { "name": "Zhengming Ding", "gender": "Male", "institution": "Tulane University", "country": "US", "position": "Assistant Professor" }, { "name": "Zizhan Zheng", "gender": "Male", "institution": "Tulane University", "country": "US", "position": "Associate Professor" } ], "n_formula": 30, "n_formula_1": 9, "n_ref_uni": 39, "n_ref": 106, "n_ref_all": 149, "n_fig": 13, "n_tab": 13, "L_tab": 2981, "n_element_tab": 289, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3462, "formula_len_all_1": 599, "len_all": 203223, "len_all_1": 74860, "len_abs": 1488, "len_title": 123, "len_sents": 69998, "len_sents_1": 38153, "n_sents": 471, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1495, "title": "Diffusion Guided Adversarial State Perturbations in Reinforcement Learning", "abs": "Reinforcement learning (RL) systems, while achieving remarkable success across various domains, are vulnerable to adversarial attacks. This is especially a concern in vision-based environments where minor manipulations of high-dimensional image inputs can easily mislead the agent's behavior. To this end, various defenses have been proposed recently, with state-of-the-art approaches achieving robust performance even under large state perturbations. Upon closer investigation, however, we found that the effectiveness of the current defenses is due to a fundamental weakness of the existing $l_p$-norm constrained attacks, which can barely alter the semantics of the input even under a relatively large perturbation budget. In this work, we propose SHIFT, a novel diffusion-based state perturbation attack to go beyond this limitation. Specifically, we train a history-conditioned diffusion model, enhanced with policy guidance and realism detection to generate perturbed states that are semantically different from the true states while remaining realistic and history-aligned to avoid detection. Evaluations show that our attack effectively breaks existing defenses, including the most sophisticated ones, and significantly lowers the agent's cumulative reward in various Atari games by more than 50\\%. The results highlight the vulnerability of RL agents to semantics-aware adversarial perturbations, indicating the importance of developing more robust policies for safety-critical domains.", "keywords": [ "Reinforcement Learning", "Adversarial Example", "Diffusion Model" ], "rating_list": [ 6, 3, 5, 3 ], "soundness_list": [ 3, 1, 2, 2 ], "presentation_list": [ 3, 1, 2, 2 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "Do3whenqeY", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Liwei Jiang", "gender": "Female", "institution": "University of Washington", "country": "", "position": "PhD student" }, { "name": "Sydney Levine", "gender": "Female", "institution": "Allen Institute for Artificial Intelligence", "country": "US", "position": "Researcher" }, { "name": "Taylor Sorensen", "gender": "Male", "institution": "University of Washington", "country": "", "position": "PhD student" }, { "name": "Yejin Choi", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 47, "n_ref": 138, "n_ref_all": 174, "n_fig": 9, "n_tab": 11, "L_tab": 2473, "n_element_tab": 272, "n_fig_1": 9, "n_tab_1": 7, "L_tab_1": 167, "n_element_tab_1": 12, "formula_len_all": 304, "formula_len_all_1": 373, "len_all": 195841, "len_all_1": 75308, "len_abs": 1710, "len_title": 126, "len_sents": 54959, "len_sents_1": 35050, "n_sents": 356, "n_sents_1": 242, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1716, "title": "Can Language Models Reason about Individualistic Human Values and Preferences?", "abs": "Recent calls for pluralistic alignment emphasize that AI systems should address the diverse needs of all people. Yet, efforts in this space often require sorting people into fixed buckets of pre-specified diversity-defining dimensions (e.g., demographics, personalities, communication styles), risking smoothing out or even stereotyping the rich spectrum of individualistic variations. To achieve an authentic representation of diversity that respects individuality, we propose individualistic alignment. While individualistic alignment can take various forms, in this paper, we introduce IndieValueCatalog, a dataset transformed from the influential World Values Survey (WVS), to study language models (LMs) on the specific challenge of individualistic value reasoning. Specifically, given a sample of an individual\u2019s value-expressing statements, models are tasked with predicting their value judgments in novel cases. With IndieValueCatalog, we reveal critical limitations in frontier LMs\u2019 abilities to reason about individualistic human values with accuracies, only ranging between 55% to 65%. Moreover, our results highlight that a precise description of individualistic values cannot be approximated only via demographic information. We also identify a partiality of LMs in reasoning about global individualistic values, as measured by our proposed Value Inequity Index (\u03c3INEQUITY). Finally, we train a series of Individualistic Value Reasoners (IndieValueReasoner) using IndieValueCatalog to enhance models\u2019 individualistic value reasoning capability, revealing new patterns and dynamics into global human values. We outline future research challenges and opportunities for advancing individualistic alignment.", "keywords": [ "individualistic value alignment", "pluralistic value alignment", "human values", "AI safety", "individualistic value reasoning" ], "rating_list": [ 3, 6, 3, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 2, 4, 4, 3 ] }, { "paper_id": "DnfPX10Etk", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Hemant yadav", "gender": "Male", "institution": "Indraprastha Institute of Information Technology, Delhi", "country": "IN", "position": "PhD student" }, { "name": "Rajiv Ratn Shah", "gender": "Male", "institution": "Indraprastha Institute of Information Technology, Delhi", "country": "IN", "position": "Associate Professor" }, { "name": "Sunayana Sitaram", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 24, "n_ref": 61, "n_ref_all": 90, "n_fig": 2, "n_tab": 9, "L_tab": 1513, "n_element_tab": 242, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1486, "n_element_tab_1": 254, "formula_len_all": 367, "formula_len_all_1": 228, "len_all": 99796, "len_all_1": 59990, "len_abs": 1122, "len_title": 117, "len_sents": 29081, "len_sents_1": 25792, "n_sents": 223, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1108, "title": "JOOCI: A FRAMEWORK FOR LEARNING COMPREHENSIVE SPEECH REPRESENTATIONS", "abs": "Information in speech can be divided into two categories: what is being said (content) and how it is expressed (other). Current state-of-the-art (SOTA) techniques model speech at fixed segments, usually 10-25 ms, using a single embedding. Given the orthogonal nature of other and content information, attempting to optimize both within a single embedding results in suboptimal solutions. This approach divides the model's capacity, limiting its ability to build complex hierarchical features effectively. In this work, we present an end-to-end speech representation learning framework designed to jointly optimize the other and \\enquote{content} information (JOOCI) in speech. By using separate learnable parameters, JOOCI addresses this optimization challenge by modeling other and content information independently. Our results show that JOOCI consistently outperforms other SOTA models of similar size (100 million parameters) and pre-training data used (960 hours) by a significant margin when evaluated on a range of speech downstream tasks in the SUPERB benchmark. Code and models are available at TBA.", "keywords": [ "SSL", "Speech Representation Learning", "Joint Optimization" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 1 ], "contribution_list": [ 2, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "DnBjhWLVU1", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Isaac Han", "gender": "Male", "institution": "Gwangju Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Sangyeon Park", "gender": "Male", "institution": "Gwangju Institute of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Seungwon Oh", "gender": "unknown", "institution": "Gwangju Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "KyungJoong Kim", "gender": "unknown", "institution": "Gwangju Institute of Science and Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 23, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 84, "n_ref_all": 91, "n_fig": 3, "n_tab": 2, "L_tab": 409, "n_element_tab": 33, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 294, "n_element_tab_1": 103, "formula_len_all": 1792, "formula_len_all_1": 586, "len_all": 125830, "len_all_1": 73317, "len_abs": 972, "len_title": 114, "len_sents": 39270, "len_sents_1": 28064, "n_sents": 342, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 977, "title": "Recovering Plasticity of Neural Networks via Soft Weight Rescaling", "abs": "Recent studies have shown that as training progresses, neural networks gradually lose their capacity to learn new information, a phenomenon known as plasticity loss. An unbounded weight growth is one of the main causes of plasticity loss. Furthermore, it harms generalization capability and disrupts optimization dynamics. Re-initializing the network can be a solution, but it results in the loss of learned information, leading to performance drops. In this paper, we propose Soft Weight Rescaling (SWR), a novel approach that prevents unbounded weight growth without losing information. SWR recovers the plasticity of the network by simply scaling down the weight at each step of the learning process. We theoretically prove that SWR bounds weight magnitude and balances weight magnitude between layers. Our experiment shows that SWR improves performance on warm-start learning, continual learning, and single-task learning setups on standard image classification benchmarks.", "keywords": [ "loss of plasticity", "plasticity", "continual learning", "online learning" ], "rating_list": [ 3, 3, 5, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "Dn7Ay7rZcH", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Dylan Zhang", "gender": "Male", "institution": "Department of Computer Science", "country": "US", "position": "PhD student" }, { "name": "Xueyan Zou", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Postdoc" }, { "name": "Hao Peng", "gender": "Male", "institution": "Department of Computer Science, University of Illinois Urbana-Champaign", "country": "", "position": "Assistant Professor" }, { "name": "SHIZHE DIAO", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 43, "n_ref": 120, "n_ref_all": 141, "n_fig": 3, "n_tab": 11, "L_tab": 2304, "n_element_tab": 277, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1788, "n_element_tab_1": 292, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 190843, "len_all_1": 55412, "len_abs": 1520, "len_title": 155, "len_sents": 33499, "len_sents_1": 24145, "n_sents": 234, "n_sents_1": 165, "venue": "ICLR2025", "decision": "withdraw", "L_title": 103, "L_abs": 1536, "title": "$\\textbf{PLUM}$: Improving Code LMs Using On-Policy Preference Learning Powered by Automatic Test Cases", "abs": "Preference learning provides a promising solution to address the limitations of supervised fine-tuning (SFT) for code language models, where the model is not explicitly trained to differentiate between correct and incorrect code.\nRecent findings demonstrate that on-policy data is the key to successful preference learning, where the preference data is collected using the same policy LM being trained.\nInspired by this, we propose PLUM,\nan on-policy $\\textbf{P}$reference $\\textbf{L}$earning framework A$\\textbf{u}$gmented with test cases for code L$\\textbf{M}$s.\nThe framework operates in three key stages: (1) automatic generation of test cases from natural language instructions, (2) creation of a preference data by evaluating candidate code solutions sampled from the policy, which can then be used to (3) train the policy LM. PLUM levitates the need to train reward models, allowing for large scale on-policy and online preference data collation. \n\nPLUM is evaluated on both standard benchmarks (HumanEval, MBPP) and more challenging ones (LiveCodeBench), delivering substantial improvements over original SFT'ed models and other execution-feedback-driven approaches. We show PLUM benefits are consistent across various widely-used code LMs even they have been well-trained with SFT. For example, PLUM increases pass rates by up to 4.8% on average on standard benchmarks and 11.8% on LiveCodeBench, demonstrating its effectiveness and generalizability. We also demonstrate the benefits of on-policy and online preference learning", "keywords": [ "Code Generation", "Preference Learning", "Test Case Generation" ], "rating_list": [ 6, 8, 3, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "DmEHmZ89iB", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Choong Seon Hong", "gender": "Male", "institution": "Kyung Hee University", "country": "KR", "position": "Full Professor" }, { "name": "Eui-Nam Huh", "gender": "unknown", "institution": "Kyung Hee University", "country": "KR", "position": "Full Professor" }, { "name": "Md Imtiaz Hossain", "gender": "Male", "institution": "Kyung Hee University", "country": "KR", "position": "PhD student" }, { "name": "Sharmen Akhter", "gender": "Female", "institution": "Kyung Hee University", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 19, "n_ref": 38, "n_ref_all": 57, "n_fig": 4, "n_tab": 6, "L_tab": 2118, "n_element_tab": 271, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 391, "n_element_tab_1": 48, "formula_len_all": 861, "formula_len_all_1": 860, "len_all": 90002, "len_all_1": 54578, "len_abs": 1441, "len_title": 133, "len_sents": 28968, "len_sents_1": 24837, "n_sents": 227, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 1489, "title": "Single Teacher, Multiple Perspectives: Teacher Knowledge Augmentation for Enhanced Knowledge Distillation", "abs": "Do diverse perspectives help students learn better? Multi-teacher knowledge distillation, which is a more effective technique than traditional single-teacher methods, supervises the student from different perspectives (i.e., teacher). While effective, multi-teacher, teacher ensemble, or teaching assistant-based approaches are computationally expensive and resource-intensive, as they require training multiple teacher networks. These concerns raise a question: can we supervise the student with diverse perspectives using only a single teacher? We, as the pioneer, demonstrate TeKAP, a novel teacher knowledge augmentation technique that generates multiple synthetic teacher knowledge by perturbing the knowledge of a single pretrained teacher i.e., Teacher Knowledge Augmentation via Perturbation, at both the feature and logit levels. These multiple augmented teachers simulate an ensemble of models together. The student model is trained on both the actual and augmented teacher knowledge, benefiting from the diversity of an ensemble without the need to train multiple teachers. TeKAP significantly reduces training time and computational resources, making it feasible for large-scale applications and easily manageable. Experimental results demonstrate that our proposed method helps existing state-of-the-art knowledge distillation techniques achieve better performance, highlighting its potential as a cost-effective alternative. The source code can be found in the supplementary.", "keywords": [ "TeKAP", "Teacher Knowledge Augmentation", "Teacher Knowledge Perturbation", "Single Teacher Multiple Perspectives", "Synthetic Teacher", "Knowledge Distillation", "Ensemble Learning", "Knowledge Transfer" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "DlqRpj68xe", "primary_area": "reinforcement learning", "n_author": 1, "author_info_list": [ { "name": "XieFeng Wu", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "MS student" } ], "n_formula": 30, "n_formula_1": 11, "n_ref_uni": 60, "n_ref": 87, "n_ref_all": 102, "n_fig": 5, "n_tab": 8, "L_tab": 2120, "n_element_tab": 186, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 147, "n_element_tab_1": 11, "formula_len_all": 3402, "formula_len_all_1": 721, "len_all": 205303, "len_all_1": 57517, "len_abs": 1902, "len_title": 136, "len_sents": 57197, "len_sents_1": 26152, "n_sents": 474, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 87, "L_abs": 880, "title": "From Reward Shaping to Q-Shaping: Achieving Unbiased Learning with LLM-Guided Knowledge", "abs": "Q-shaping is an extension of Q-value initialization and serves as an alternative to reward shaping for incorporating domain knowledge to accelerate agent training, thereby improving sample efficiency by directly shaping Q-values. This approach is both general and robust across diverse tasks, allowing for immediate impact assessment while guaranteeing optimality. We evaluated Q-shaping across 20 different environments using a large language model (LLM) as the heuristic provider. The results demonstrate that Q-shaping significantly enhances sample efficiency, achieving an \\textbf{16.87\\%} average improvement across the 20 tasks compared to the best baseline, and a \\textbf{226.67\\%} improvement compared to LLM-based reward shaping methods. These findings establish Q-shaping as an effective and unbiased alternative to conventional reward shaping in reinforcement learning.", "keywords": [ "reward shaping", "reinforcement learning", "large language model" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "DlZ97cVwr0", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Hasmik Mnatsakanyan", "gender": "Female", "institution": "Deep Origin", "country": "AM", "position": "Researcher" }, { "name": "Hrant Khachatrian", "gender": "unknown", "institution": "Yerevan State University", "country": "AM", "position": "Principal Researcher" }, { "name": "Knarik Mheryan", "gender": "unknown", "institution": "YerevaNN", "country": "AM", "position": "Researcher" }, { "name": "Philipp Guevorguian", "gender": "Male", "institution": "YerevaNN", "country": "", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 3, "n_ref_uni": 34, "n_ref": 42, "n_ref_all": 60, "n_fig": 3, "n_tab": 7, "L_tab": 2063, "n_element_tab": 222, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 1947, "n_element_tab_1": 189, "formula_len_all": 200, "formula_len_all_1": 169, "len_all": 112329, "len_all_1": 62726, "len_abs": 1338, "len_title": 112, "len_sents": 32667, "len_sents_1": 27486, "n_sents": 269, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1298, "title": "Exploring the Recall of Language Models: Case Study on Molecules", "abs": "Most of the current benchmarks evaluate Generative Language Models based on the accuracy of the generated output. However, in some scenarios, it is also important to evaluate the recall of the generations, i.e., whether a model can generate all correct outputs, such as all security vulnerabilities of a given codebase. There are two challenges in evaluating the recall: the lack of complete sets of correct outputs for any task and the existence of many distinct but similar outputs (e.g., two exploits that target the same vulnerability).\n\nIn this paper, we propose a benchmark from the domain of small organic molecules. We define several sets of molecules of varying complexity and fine-tune language models on subsets of those sets. We attempt to generate as many molecules from the target sets as possible and measure the recall, i.e., the percentage of generated molecules from the target set. We examine the impact of the training loss function and sampling strategy on the recall. We propose a sampling strategy based on beam search that avoids duplicates and maximizes recall. Finally, we show that given a small validation set, one can predict the recall of the model without actually generating many samples, which can act as a model selection strategy for maximizing generation recall.", "keywords": [ "recall", "language models", "molecular language models", "sampling methods for language models" ], "rating_list": [ 6, 3, 5, 3, 5 ], "soundness_list": [ 3, 2, 2, 3, 2 ], "presentation_list": [ 2, 1, 2, 2, 3 ], "contribution_list": [ 2, 2, 1, 2, 2 ], "confidence_list": [ 3, 3, 2, 3, 3 ] }, { "paper_id": "Dl6nkKKvlX", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Ahmed Abdulaal", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Amrutha Saseendran", "gender": "Female", "institution": "AstraZeneca", "country": "GB", "position": "Research Scientist" }, { "name": "Aryo Pradipta Gema", "gender": "Male", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" }, { "name": "Chen Jin", "gender": "unknown", "institution": "Astrazeneca", "country": "GB", "position": "Researcher" }, { "name": "Daniel C. Alexander", "gender": "Male", "institution": "University College London", "country": "", "position": "Full Professor" }, { "name": "Daniel C. Castro", "gender": "unknown", "institution": "Microsoft Research", "country": "", "position": "Senior Researcher" }, { "name": "Dino Oglic", "gender": "Male", "institution": "AstraZeneca UK", "country": "GB", "position": "Researcher" }, { "name": "Philip Alexander Teare", "gender": "Male", "institution": "AstraZeneca", "country": "", "position": "Researcher" }, { "name": "Tom Diethe", "gender": "Male", "institution": "AstraZeneca", "country": "", "position": "Principal Researcher" }, { "name": "Nina Montana-Brown", "gender": "Female", "institution": "University College London, University of London", "country": "", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 19, "n_ref_uni": 69, "n_ref": 182, "n_ref_all": 206, "n_fig": 13, "n_tab": 4, "L_tab": 1897, "n_element_tab": 203, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 1306, "n_element_tab_1": 121, "formula_len_all": 1629, "formula_len_all_1": 906, "len_all": 248487, "len_all_1": 69882, "len_abs": 1564, "len_title": 122, "len_sents": 78121, "len_sents_1": 33731, "n_sents": 581, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1461, "title": "Balancing Act: Diversity and Consistency in Large Language Model Ensembles", "abs": "Ensembling strategies for Large Language Models (LLMs) have demonstrated significant potential in improving performance across various tasks by combining the strengths of individual models. However, identifying the most effective ensembling method remains an open challenge, as neither maximizing output consistency through self-consistency decoding nor enhancing model diversity via frameworks like \"Mixture of Agents\" has proven universally optimal. Motivated by this, we propose a unified framework to examine the trade-offs between task performance, model diversity, and output consistency in ensembles. More specifically, we introduce a consistency score that defines a gating mechanism for mixtures of agents and an algorithm for mixture refinement to investigate these trade-offs at the semantic and model levels, respectively. We incorporate our insights into a novel inference-time LLM ensembling strategy called the Dynamic Mixture of Agents (DMoA) and demonstrate that it achieves a new state-of-the-art result in the challenging Big Bench Hard mixed evaluations benchmark. Our analysis reveals that cross-validation bias can enhance performance, contingent on the expertise of the constituent models. We further demonstrate that distinct reasoning tasks\u2014such as arithmetic reasoning, commonsense reasoning, and instruction following\u2014require different model capabilities, leading to inherent task-dependent trade-offs that DMoA balances effectively.", "keywords": [ "LLM", "ensembling", "diversity", "consistency", "mixture of agents", "self decoding" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "Dl5JaX7zoN", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Depeng Jin", "gender": "unknown", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Guozhen Zhang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Jie Feng", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" }, { "name": "Longyi Liu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Yong Li", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yu Zheng", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Yuming Lin", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Postdoc" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 33, "n_ref": 65, "n_ref_all": 78, "n_fig": 4, "n_tab": 4, "L_tab": 763, "n_element_tab": 81, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 402, "n_element_tab_1": 40, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 134274, "len_all_1": 60542, "len_abs": 1849, "len_title": 141, "len_sents": 38811, "len_sents_1": 30302, "n_sents": 261, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1743, "title": "UrbanPlanBench: A Comprehensive Assessment of Urban Planning Abilities in Large Language Models", "abs": "Urban planning is a professional discipline that shapes our daily surroundings, which demands multifaceted domain knowledge and relies heavily on human expertise. The advent of Large Language Models (LLMs) holds promise for revolutionizing such a field by the pre-trained world knowledge. However, the extent to which these models can assist human practitioners remains largely unexplored. In this paper, we introduce a comprehensive benchmark, PlanBench, tailored to evaluate the efficacy of LLMs in urban planning, which encompasses fundamental principles, professional knowledge, and management and regulations, aligning closely with the qualifications expected of human planners. Through extensive evaluation, we reveal a significant imbalance in the acquisition of planning knowledge among LLMs, with even the most proficient models falling short of meeting professional standards. For instance, we observe that 70% of LLMs achieve subpar performance in understanding planning regulations compared to other aspects. Besides the benchmark, we present the largest-ever supervised fine-tuning (SFT) dataset, PlanText, for LLMs in urban planning, comprising over 30,000 instruction pairs sourced from urban planning exams and textbooks. Our findings demonstrate that fine-tuned models exhibit enhanced performance in memorization tests and comprehension of urban planning knowledge, while there exists significant room for improvement, particularly in tasks requiring domain-specific terminology and reasoning. Our benchmark, dataset, and associated evaluation and fine-tuning toolsets aim to catalyze the integration of LLMs into practical urban computing, fostering a symbiotic relationship between human expertise and machine intelligence.", "keywords": [ "LLM Benchmark", "urban planning" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Dl3MsjaIdp", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Linwei Wang", "gender": "Female", "institution": "Rochester Institute of Technology", "country": "US", "position": "Professor" }, { "name": "Ryan Missel", "gender": "Male", "institution": "Rochester Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 38, "n_ref": 153, "n_ref_all": 209, "n_fig": 33, "n_tab": 11, "L_tab": 6614, "n_element_tab": 520, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 312, "formula_len_all_1": 267, "len_all": 192101, "len_all_1": 67863, "len_abs": 1389, "len_title": 157, "len_sents": 68307, "len_sents_1": 34328, "n_sents": 463, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 109, "L_abs": 1409, "title": "Continual Slow-and-Fast Adaptation of Latent Neural Dynamics (CoSFan): Meta-Learning What-How & When to Adapt", "abs": "An increasing interest in learning to forecast for time-series of high-dimensional observations is the ability to adapt to systems with diverse underlying dynamics. Access to observations that define a stationary distribution of these systems is often unattainable, as the underlying dynamics may change over time. Naively training or retraining models at each shift may lead to catastrophic forgetting about previously-seen systems. We present a new continual meta-learning (CML) framework to realize continual slow-and fast adaptation of latent dynamics (CoSFan). We leverage a feed-forward meta-model to infer *what* the current system is and *how* to adapt a latent dynamics function to it, enabling *fast adaptation* to specific dynamics. We then develop novel strategies to automatically detect *when* a shift of data distribution occurs, with which to identify its underlying dynamics and its relation with previously-seen dynamics. In combination with fixed-memory experience replay mechanisms, this enables continual *slow update* of the *what-how* meta-model. Empirical studies demonstrated that both the meta- and continual-learning component was critical for learning to forecast across non-stationary distributions of diverse dynamics systems, and the feed-forward meta-model combined with task-aware/-relational continual learning strategies significantly outperformed existing CML alternatives.", "keywords": [ "continual meta-learning", "latent dynamics forecasting", "time-series" ], "rating_list": [ 10, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "DkzZ1ooc7q", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Minghui Fang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Rongjie Huang", "gender": "Male", "institution": "FAIR", "country": "US", "position": "Researcher" }, { "name": "Shengpeng Ji", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Siqi Zheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Tao Jin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xize Cheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zehan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Ziang Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Jialung Zuo", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 40, "n_ref": 79, "n_ref_all": 93, "n_fig": 2, "n_tab": 8, "L_tab": 3349, "n_element_tab": 398, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2401, "n_element_tab_1": 225, "formula_len_all": 747, "formula_len_all_1": 705, "len_all": 132505, "len_all_1": 56961, "len_abs": 1477, "len_title": 113, "len_sents": 39525, "len_sents_1": 25176, "n_sents": 249, "n_sents_1": 160, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1501, "title": "OmniSep: Unified Omni-Modality Sound Separation with Query-Mixup", "abs": "Query-based sound separation (QSS) effectively isolate sound signals that match the content of a given query, enhancing the understanding of audio data. However, most existing QSS methods rely on a single modality for separation, lacking the ability to fully leverage homologous but heterogeneous information across multiple modalities for the same sound signal. To address this limitation, we introduce Omni-modal Sound Separation (**OmniSep**), a novel framework capable of isolating clean soundtracks based on omni-modal queries, encompassing both single-modal and multi-modal composed queries. Specifically, we introduce the **Query-Mixup** strategy, which blends query features from different modalities during training. This enables OmniSep to optimize multiple modalities concurrently, effectively bringing all modalities under a unified framework for sound separation. We further enhance this flexibility by allowing queries to influence sound separation positively or negatively, facilitating the retention or removal of specific sounds as desired. Finally, OmniSep employs a retrieval-augmented approach known as **Query-Aug**, which enables open-vocabulary sound separation. Experimental evaluations on MUSIC, VGGSOUND-CLEAN+, and MUSIC-CLEAN+ datasets demonstrate effectiveness of OmniSep, achieving state-of-the-art performance in text-, image-, and audio-queried sound separation tasks. For samples and further information, please visit the demo page at \\url{https://omnisep.github.io/}.", "keywords": [ "sound separation", "composed query", "negative query" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "Dkz8npDqAv", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Chang Liu", "gender": "unknown", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Edmund Y. Lam", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Ngai Wong", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Siew Chong Tan", "gender": "Male", "institution": ", University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "XIAOJUAN QI", "gender": "Female", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Xiaoyan Qian", "gender": "Female", "institution": "The University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 13, "n_ref": 32, "n_ref_all": 38, "n_fig": 2, "n_tab": 7, "L_tab": 2730, "n_element_tab": 186, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 2470, "n_element_tab_1": 132, "formula_len_all": 161, "formula_len_all_1": 163, "len_all": 115131, "len_all_1": 51364, "len_abs": 1846, "len_title": 133, "len_sents": 30294, "len_sents_1": 25401, "n_sents": 142, "n_sents_1": 99, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1421, "title": "Multimodal Context-Aware Transformer with Visual Guidance for Automated 3D Annotation", "abs": "The laborious nature of manual point cloud labeling drives the growing interest in 3D auto-annotation. The challenge is amplified by the sparse and irregular distribution of point clouds. This leads to the under-performance of current autolabelers, particularly with hard-to-detect samples characterized by truncation, occlusion, or distance.\n In response, we propose a multimodal context-aware transformer (MMCAT) that integrates 3D point cloud geometry with image-based semantic insights to improve 3D bounding box annotations through 2D visual guidance. Our approach utilizes visual hints from three perspectives to integrate the 2D and 3D dimensions.\n Initially, we develop point and image encoders to align LiDAR and image data, establishing a unified semantic bridge between image visuals and point cloud geometry. Subsequently, our box encoder processes 2D box coordinates to improve accuracy in determining object positions and dimensions within 3D space. Finally, our multimodal encoders enhance feature interactions, improving point cloud interpretation and annotation accuracy, especially for challenging samples.\n MMCAT lies in its strategic use of 2D visual prompts to bolster 3D representation and annotation processes. We validate MMCAT's efficacy through extensive experiments on the widely recognized KITTI and Waymo Open datasets, particularly highlighting its superior performance with hard samples.", "keywords": [ "3D point cloud", "multimodal architecture", "automatic annotation", "LiDar" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "DjtJV3ke1j", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "OMRI WEINSTEIN", "gender": "Male", "institution": "The Hebrew University", "country": "IL", "position": "Associate Professor " }, { "name": "Wenyu Jin", "gender": "unknown", "institution": "University of Illinois at Chicago", "country": "", "position": "PhD student" }, { "name": "Xiaorui Sun", "gender": "Male", "institution": "University of Illinois, Chicago", "country": "US", "position": "Assistant Professor" }, { "name": "Yichuan Deng", "gender": "Male", "institution": "Department of Computer Science, University of Washington", "country": "US", "position": "PhD student" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" } ], "n_formula": 132, "n_formula_1": 18, "n_ref_uni": 30, "n_ref": 67, "n_ref_all": 82, "n_fig": 2, "n_tab": 7, "L_tab": 988, "n_element_tab": 90, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12691, "formula_len_all_1": 1308, "len_all": 316986, "len_all_1": 89612, "len_abs": 4007, "len_title": 74, "len_sents": 90248, "len_sents_1": 37110, "n_sents": 938, "n_sents_1": 302, "venue": "ICLR2025", "decision": "Reject", "L_title": 26, "L_abs": 1033, "title": "Dynamic Kernel Sparsifiers", "abs": "A geometric graph associated with a set of points $P= \\{x_1, x_2, \\cdots, x_n \\} \\subset \\mathbb{R}^d$ and a fixed kernel function $\\mathsf{K}:\\mathbb{R}^d\\times \\mathbb{R}^d\\to\\mathbb{R}_{\\geq 0}$ is a complete graph on $P$ such that the weight of edge $(x_i, x_j)$ is $\\mathsf{K}(x_i, x_j)$. We present a fully-dynamic data structure that maintains a spectral sparsifier of a geometric graph under updates that change the locations of points in $P$ one at a time. The update time of our data structure is $n^{o(1)}$ with high probability, and the initialization time is $n^{1+o(1)}$. Under certain assumption, our data structure can be made robust against adaptive adversaries, which makes our sparsifier applicable in iterative optimization algorithms. \n\nWe further show that the Laplacian matrices corresponding to geometric graphs admit a randomized sketch for maintaining matrix-vector multiplication and projection in $n^{o(1)}$ time, under \\emph{sparse} updates to the query vectors, or under modification of points in $P$.", "keywords": [ "Sparsifiers", "Optimization", "Algorithms" ], "rating_list": [ 6, 3, 6, 3 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 1, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Djk1Tgs0wR", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Chuhua XIAN", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Shengfeng He", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Associate Professor" }, { "name": "Yu Guo", "gender": "Male", "institution": "Wuhan University of Technology", "country": "CN", "position": "PhD student" }, { "name": "LI Junhang", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 39, "n_ref": 72, "n_ref_all": 106, "n_fig": 15, "n_tab": 6, "L_tab": 2319, "n_element_tab": 152, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2798, "n_element_tab_1": 126, "formula_len_all": 414, "formula_len_all_1": 288, "len_all": 141005, "len_all_1": 57180, "len_abs": 1179, "len_title": 121, "len_sents": 43221, "len_sents_1": 25654, "n_sents": 308, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1183, "title": "SeeThruAnything: Learning to Remove Any Obstructions Across Distributions", "abs": "Images are often obstructed by various obstacles due to capture limitations, hindering the observation of objects of interest. Most existing methods address occlusions from specific elements like fences or raindrops, but are constrained by the wide range of real-world obstructions, making comprehensive data collection impractical. To overcome these challenges, we propose SeeThruAnything, a novel zero-shot framework capable of handling both seen and unseen obstacles. The core idea of our approach is to unify obstruction removal by treating it as a soft-hard mask restoration problem, where any obstruction can be represented using multi-modal prompts, such as visual semantics and textual commands, processed through a cross-attention unit to enhance contextual understanding and improve mode control. Additionally, a tunable mask adapter allows for dynamic soft masking, enabling real-time adjustment of inaccurate masks. Extensive experiments on both in-distribution and out-of-distribution obstacles show that SeeThruAnything consistently achieves strong performance and generalization in obstruction removal, regardless of whether the obstacles were present during training.", "keywords": [ "Obstruction Removal", "Zero-shot", "Prompts" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "DjHnxxlqwl", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Muheng Li", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Shuxin Li", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Shuxin Zhuang", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Tianji Yang", "gender": "Not Specified", "institution": "Georgia Institute of Technology", "country": "US", "position": "MS student" }, { "name": "Xianjie Shi", "gender": "Male", "institution": "The University of Hong Kong", "country": "HK", "position": "MS student" }, { "name": "Youzhi Zhang", "gender": "unknown", "institution": "Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences", "country": "", "position": "Assistant Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 46, "n_ref": 135, "n_ref_all": 158, "n_fig": 8, "n_tab": 3, "L_tab": 556, "n_element_tab": 33, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 51, "n_element_tab_1": 1, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 157499, "len_all_1": 70721, "len_abs": 1670, "len_title": 146, "len_sents": 48665, "len_sents_1": 36143, "n_sents": 336, "n_sents_1": 240, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 1689, "title": "Solving Urban Network Security Games: Learning Platform, Benchmark, and Challenge for AI Research", "abs": "After the great achievement of solving two-player zero-sum games, more and more AI researchers focus on solving multiplayer games. To facilitate the development of designing efficient learning algorithms for solving multiplayer games, we propose a multiplayer game platform for solving Urban Network Security Games (**UNSG**) that model real-world scenarios. That is, \npreventing criminal activity is a highly significant responsibility assigned to police officers in cities, and police officers have to allocate their limited security resources to interdict the escaping criminal when a crime takes place in a city. This interaction between multiple police officers and the escaping criminal can be modeled as a UNSG. The variants of UNSGs can model different real-world settings, e.g., whether real-time information is available or not, whether police officers can communicate or not.\nThe main challenges of solving this game include the large size of the game and the co-existence of cooperation and competition. \nWhile previous efforts have been made to tackle UNSGs, they have been hampered by performance and scalability issues. Therefore, we propose an open-source UNSG platform (**GraphChase**) for designing efficient learning algorithms for solving UNSGs.\nSpecifically, GraphChase offers a unified and flexible game environment for modeling various variants of UNSGs, supporting the development, testing, and benchmarking of algorithms. We believe that GraphChase not only facilitates the development of efficient algorithms for solving real-world problems but also paves the way for significant advancements in algorithmic development for solving general multiplayer games.", "keywords": [ "security games", "multiplayer games" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "DjEyXTbEpa", "primary_area": "applications to neuroscience & cognitive science", "n_author": 2, "author_info_list": [ { "name": "Songlin Xu", "gender": "unknown", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Xinyu Zhang", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 3, "n_ref_uni": 41, "n_ref": 89, "n_ref_all": 139, "n_fig": 18, "n_tab": 6, "L_tab": 4191, "n_element_tab": 586, "n_fig_1": 13, "n_tab_1": 3, "L_tab_1": 1369, "n_element_tab_1": 594, "formula_len_all": 587, "formula_len_all_1": 192, "len_all": 222983, "len_all_1": 69259, "len_abs": 1442, "len_title": 114, "len_sents": 72922, "len_sents_1": 32535, "n_sents": 523, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1449, "title": "Machine Reinforced Perturbation on Drifted Human Logical Reasoning", "abs": "Using deep neural networks as computational models to simulate cognitive process can provide key insights into human behavioral dynamics. This enables synthetic data generation to test hypotheses for neuroscience and guides adaptive interventions for cognitive regulation. Challenges arise when environments are highly dynamic, obscuring stimulus-behavior relationships. However, the majority of current research focuses on simulating human cognitive behaviors under ideal conditions, neglecting the influence of environmental disturbances. We propose ReactiveAgent, integrating drift-diffusion with deep reinforcement learning to simulate granular effects of dynamic environmental stimuli on human logical reasoning process. This framework is built and evaluated upon our contributed large dataset of 21,157 logical responses of humans under various dynamic stimuli. Quantitatively, the framework improves cognition modelling by considering temporal effect of environmental stimuli on logical reasoning and captures both subject-specific and stimuli-specific behavioural differences. Qualitatively, it captures general trends in human logical reasoning under stress, better than baselines. Our approach is extensible to examining diverse environmental influences on cognitive behaviors. Overall, it demonstrates a powerful, data-driven methodology to simulate, align with, and understand the vagaries of human logical reasoning in dynamic contexts.", "keywords": [ "Human Logical Reasoning", "Deep Reinforcement Learning", "Cognitive Model" ], "rating_list": [ 6, 5, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 2, 3, 4 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 2, 2, 3, 3, 4 ] }, { "paper_id": "Dj9wssUmLn", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 8, "author_info_list": [ { "name": "Anh Tuan Luu", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "Assistant Professor" }, { "name": "Do Xuan Long", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Do Xuan Trong", "gender": "Male", "institution": "Hanoi University of Science and Technology", "country": "VN", "position": "Undergrad student" }, { "name": "Duong Ngoc Yen", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Undergrad student" }, { "name": "Kenji Kawaguchi", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Presidential Young Professor" }, { "name": "Min-Yen Kan", "gender": "Male", "institution": "National University of Singapore", "country": "", "position": "Associate Professor" }, { "name": "Nancy F. Chen", "gender": "unknown", "institution": "I2R, A*STAR", "country": "", "position": "Principal Researcher" }, { "name": "Shafiq Joty", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Principal Researcher" } ], "n_formula": 9, "n_formula_1": 2, "n_ref_uni": 55, "n_ref": 111, "n_ref_all": 177, "n_fig": 15, "n_tab": 26, "L_tab": 7767, "n_element_tab": 677, "n_fig_1": 10, "n_tab_1": 5, "L_tab_1": 3161, "n_element_tab_1": 220, "formula_len_all": 809, "formula_len_all_1": 113, "len_all": 349653, "len_all_1": 69796, "len_abs": 2644, "len_title": 122, "len_sents": 79291, "len_sents_1": 29149, "n_sents": 679, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 122, "L_abs": 1338, "title": "Beyond In-Context Learning: Enhancing Long-form Generation of Large Language Models via Task-Inherent Attribute Guidelines", "abs": "In-context learning (ICL) is an important yet not fully understood ability of pre-trained large language models (LLMs). It can greatly enhance task performance using a few examples, termed demonstrations, without fine-tuning. Although effective in question answering, ICL often underperforms in long-form generation tasks such as summarization. Under appropriately realistic assumptions, we empirically and theoretically show that ICL demonstrations alone are insufficient to teach LLMs the task\u2019s language and format distributions for generation.\nWe argue for explicit exposure to the task distributions and hypothesize that defining them by prompting enhances model performance. To this end, we present LongGuide, which efficiently generates two parallel streams of guidelines capturing task language and format properties: (i) Metric Guidelines (MGs) that instruct models to optimize self-evaluated metrics; and (ii) Output Constraint Guidelines (OCGs) that constrain generation at both token and sentence levels. LongGuide automatically selects the best combination of guidelines, improving both strong open- and closed-source LLMs by over 5% in both zero- and few-shot settings. We show that LongGuide is generalizable, learnable by weak models to enhance strong ones, and integrates synergistically with automatic prompt optimizers.", "keywords": [ "In-context Learning", "Prompt Optimization", "Long-form Generation" ], "rating_list": [ 6, 6, 6, 5, 6 ], "soundness_list": [ 4, 2, 3, 2, 3 ], "presentation_list": [ 4, 3, 3, 3, 4 ], "contribution_list": [ 3, 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4, 3 ] }, { "paper_id": "Dj9a4zQsSl", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Bo Zheng", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Instructor" }, { "name": "Chi Xie", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "PhD student" }, { "name": "Dongdong Liu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yue Zhang", "gender": "Female", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Yuke Zhu", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Zihua Xiong", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Sheng Guo", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 45, "n_ref": 122, "n_ref_all": 128, "n_fig": 5, "n_tab": 7, "L_tab": 1902, "n_element_tab": 331, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1125, "n_element_tab_1": 197, "formula_len_all": 587, "formula_len_all_1": 630, "len_all": 184404, "len_all_1": 69167, "len_abs": 1242, "len_title": 162, "len_sents": 51817, "len_sents_1": 31449, "n_sents": 413, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 114, "L_abs": 1242, "title": "Enhancing Document Understanding with Group Position Embedding: A Novel Approach to Incorporate Layout Information", "abs": "Recent advancements in document understanding have been dominated by leveraging large language models (LLMs) and multimodal large models. However, enabling LLMs to comprehend complex document layouts and structural information often necessitates intricate network modifications or costly pre-training, limiting their practical applicability. In this paper, we introduce Group Position Embedding (GPE), a novel and efficient technique to enhance the layout understanding capabilities of LLMs without architectural changes or additional pre-training. GPE achieves this by strategically grouping the attention heads and feeding each group with distinct positional embeddings, effectively encoding layout information relevant to document comprehension. This simple yet powerful method allows for effective integration of layout information within the existing LLM framework. We evaluate GPE against several competitive baselines across five mainstream document tasks. We also introduce a challenging benchmark called BLADE, specifically designed to assess layout comprehension. \n Extensive experiments on both established and BLADE benchmarks confirm the efficacy of GPE in significantly advancing the state-of-the-art in document understanding.", "keywords": [ "DocAI", "LLM", "Position Embedding" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 1, 2, 2, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 5, 4, 2 ] }, { "paper_id": "Dj1PVLU8fK", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Adhiraj Ghosh", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "", "position": "MS student" }, { "name": "Ameya Prabhu", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "Postdoc" }, { "name": "Matthias Bethge", "gender": "Male", "institution": "University of Tuebingen", "country": "DE", "position": "Full Professor" }, { "name": "Sebastian Dziadzio", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Vishaal Udandarao", "gender": "Male", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "samuel Albanie", "gender": "Not Specified", "institution": "Google", "country": "GB", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 0, "n_ref_uni": 96, "n_ref": 166, "n_ref_all": 186, "n_fig": 5, "n_tab": 1, "L_tab": 1452, "n_element_tab": 158, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 268, "formula_len_all_1": 0, "len_all": 268471, "len_all_1": 51229, "len_abs": 2144, "len_title": 153, "len_sents": 47265, "len_sents_1": 25846, "n_sents": 345, "n_sents_1": 170, "venue": "ICLR2025", "decision": "withdraw", "L_title": 113, "L_abs": 2261, "title": "Democratizing Evaluation with Infinity-Benchmarks: Sample-Level Heterogeneous Testing Over Arbitrary Capabilities", "abs": "Traditional fixed test datasets fall short in quantifying the open-ended potential of foundation models. In this work, we propose \u221e-benchmarks, a new testing paradigm that combines individual evaluation datasets into a single, uniform, ever-expanding sample pool from which custom evaluations can be flexibly generated. An \u221e-benchmark allows users to dynamically select a collection of sample-level evaluations that correspond to their specific capabilities of interest. By aggregating and reusing samples across various test sets, it enables the assessment of diverse capabilities beyond those covered by the original test sets, while mitigating overfitting and dataset bias through real-world diversity. Most importantly, it frames model evaluation as a collective process of aggregation and selection of sample-level tests.\n\nThe shift from multi-task benchmarks to \u221e-benchmarks introduces two key challenges: (1) heterogeneity and (2) incompleteness. Heterogeneity refers to aggregating diverse metrics, including binary, numeric, and ordinal data, while incompleteness describes comparing models evaluated on different subsets of testing data. To address these challenges, we explore algorithms inspired by social choice theory which aggregate sparse, unequal measurements into reliable model scores. Our aggregation algorithm ensures identifiability (asymptotically recovering ground-truth scores) and rapid convergence, enabling accurate model comparisons with relatively little data. We introduce \u221e-LLMBench for language models and \u221e-LMMBench for vision-language models, unifying evaluations across leaderboards and arenas in these domains, and showcasing targeted querying over a wide-range of capabilities. Our algorithm recovers ground truth rankings with large Kendall \u03c4 correlations when compared to standard aggregation on homogeneous metrics, even with up to 95% of measurements missing. This approach reduces evaluation cost by up to 20\u00d7 with little to no compromise in performance. Overall, we present the first large-scale \u221e-benchmarks for lifelong, efficient evaluation of language and vision-language models which can aggregate over open-ended heterogeneous sample-level testing to evolve alongside the rapid development of foundation models.", "keywords": [ "foundation models", "efficient evaluation", "aggregation", "lifelong benchmarking", "heterogeneity" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "DiRJUdmZoK", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Jianfeng Gao", "gender": "Male", "institution": "Microsoft Research", "country": "US", "position": "Principal Researcher" }, { "name": "Xiujun Li", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Yejin Choi", "gender": "Female", "institution": "Computer Science Department, Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Yujie Lu", "gender": "unknown", "institution": "UC Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Zhe Gan", "gender": "Male", "institution": "Apple", "country": "US", "position": "Principal Researcher" }, { "name": "William Yang Wang", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 19, "n_ref": 37, "n_ref_all": 72, "n_fig": 4, "n_tab": 11, "L_tab": 2930, "n_element_tab": 456, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 2054, "n_element_tab_1": 353, "formula_len_all": 16, "formula_len_all_1": 0, "len_all": 139744, "len_all_1": 56682, "len_abs": 958, "len_title": 148, "len_sents": 36939, "len_sents_1": 22057, "n_sents": 325, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 954, "title": "Pixelated Instructions: Can Multimodal Large Language Models Follow Printed Instructions in Images?", "abs": "Recent multimodal large language models (MLLMs) have shown promising instruction following capabilities on vision-language tasks. In this work, we introduce VISUAL MODALITY INSTRUCTION (VIM), and investigate how well multimodal models can understand textual instructions provided in pixels, despite not being explicitly trained on such data during pretraining or fine-tuning. We adapt VIM to eight benchmarks, including OKVQA, MM-Vet, MathVista, MMMU, and probe diverse MLLMs in both the text-modality instruction (TEM) setting and VIM setting. Notably, we observe a significant performance disparity between the original TEM and VIM settings for open-source MLLMs, indicating that open-source MLLMs face greater challenges when text instruction is presented solely in image form. To address this issue, we train V-MLLM, a generalizable model that is capable to conduct robust instruction following in both text-modality and visual-modality instructions.", "keywords": [ "multimodal large langauage models", "instruction following" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "Di3VLZHZdj", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 8, "author_info_list": [ { "name": "Harry Millwater", "gender": "Male", "institution": "University of Texas at San Antonio", "country": "US", "position": "Full Professor" }, { "name": "Jacob Hochhalter", "gender": "Male", "institution": "University of Utah", "country": "US", "position": "Associate Professor" }, { "name": "Juan D Ocampo", "gender": "Male", "institution": "St. Mary's University", "country": "US", "position": "Associate Professor" }, { "name": "Robert Kirby", "gender": "Male", "institution": ", University of Utah", "country": "", "position": "Full Professor" }, { "name": "Nathan Crosby", "gender": "Male", "institution": "Aeromatter", "country": "US", "position": "Principal Researcher" }, { "name": "Sara Restrepo-Velasquez", "gender": "Female", "institution": "University of Texas at San Antonio", "country": "US", "position": "PhD student" }, { "name": "Shandian Zhe", "gender": "Male", "institution": ", University of Utah", "country": "US", "position": "Associate Professor" }, { "name": "Tushar Gautam", "gender": "Male", "institution": "University of Utah", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 21, "n_ref": 29, "n_ref_all": 46, "n_fig": 14, "n_tab": 1, "L_tab": 142, "n_element_tab": 14, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 142, "n_element_tab_1": 14, "formula_len_all": 195, "formula_len_all_1": 109, "len_all": 72892, "len_all_1": 43363, "len_abs": 1244, "len_title": 107, "len_sents": 24580, "len_sents_1": 20077, "n_sents": 191, "n_sents_1": 156, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1256, "title": "Efficient Fatigue Modeling: Applying Operator Networks for Stress Intensity Factor Prediction and Analysis", "abs": "Fatigue modeling is essential for material-related applications, including design, engineering, manufacturing, and maintenance. Central to fatigue modeling is the computation and analysis of stress intensity factors (SIFs), which model the crack-driving force and are influenced by factors such as geometry, load, crack shape, and crack size. Traditional methods are based on finite element analysis, which is computationally expensive. A common engineering practice is manually constructing handbook (surrogate) solutions, though these are limited when dealing with complex scenarios, such as intricate geometries. In this work, we reformulate SIF computation as an operator learning problem, leveraging recent advancements in data-driven operator networks to enable efficient and accurate predictions. Our results show that, when trained on a relatively small finite element dataset, operator networks --- such as Deep Operator Networks (DeepONet) and Fourier Neural Operators (FNO) --- achieve less than 5\\% relative error, significantly outperforming popular handbook solutions. We further demonstrate how these predictions can be integrated into crack growth simulations and used to calculate the probability of failure in small aircraft applications.", "keywords": [ "Stress Intensity Factors", "Scientific ML", "Operator Network", "Crack Growth Simulation", "Fatigue Modeling", "Solid Mechanics", "AI for Science" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 2, 1 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "DhlbK7tAjz", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Christian Rupprecht", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Associate Professor" }, { "name": "Hilde Kuehne", "gender": "Female", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "Full Professor" }, { "name": "Sofian Chaybouti", "gender": "Male", "institution": "Johann Wolfgang Goethe Universit\u00e4t Frankfurt am Main", "country": "DE", "position": "PhD student" }, { "name": "Vittorio Ferrari", "gender": "Male", "institution": "Synthesia", "country": "GB", "position": "Principal Researcher" }, { "name": "Walid Bousselham", "gender": "Male", "institution": "Johann Wolfgang Goethe Universit\u00e4t Frankfurt am Main", "country": "DE", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 36, "n_ref": 66, "n_ref_all": 86, "n_fig": 8, "n_tab": 4, "L_tab": 1578, "n_element_tab": 353, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 1952, "n_element_tab_1": 504, "formula_len_all": 460, "formula_len_all_1": 430, "len_all": 139994, "len_all_1": 78710, "len_abs": 1826, "len_title": 123, "len_sents": 45854, "len_sents_1": 33551, "n_sents": 336, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1502, "title": "MaskInversion: Localized Embeddings via Optimization of Explainability Maps", "abs": "Vision-language foundation models such as CLIP have achieved tremendous results in global vision-language alignment, but still show some limitations in creating representations for specific image regions. \nTo address this problem, we propose MaskInversion, a method that leverages the feature representations of pre-trained foundation models, such as CLIP, to generate a context-aware embedding for a query image region specified by a mask at test time.\nMaskInversion starts with initializing an embedding token and compares its explainability map, derived from the pretrained model, to the query mask.\nThe embedding token is then subsequently refined to approximate the query region by minimizing the discrepancy between its explainability map and the query mask. During this process, only the embedding vector is updated, while the underlying foundation model is kept frozen\nallowing to use MaskInversion with any pre-trained model. \nAs deriving the explainability map involves computing its gradient, which can be expensive, we propose a gradient decomposition strategy that simplifies this computation.\nThe learned region representation can be used for a broad range of tasks, including open-vocabulary class retrieval, referring expression comprehension, as well as for localized captioning and image generation. We evaluate the proposed method on all those tasks on several datasets such as PascalVOC, MSCOCO, RefCOCO, and OpenImagesV7 and show its capabilities compared to other SOTA approaches.", "keywords": [ "localized embedding", "fondation models", "test-time optimization" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "DhdqML3FdM", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Aurelio Sulser", "gender": "unknown", "institution": "ETHZ - ETH Zurich", "country": "", "position": "PhD student" }, { "name": "Davide Scaramuzza", "gender": "Male", "institution": "University of Zurich", "country": "CH", "position": "Professor" }, { "name": "Federico Sold\u00e0", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Nikola Zubic", "gender": "Male", "institution": "Department of Informatics, University of Zurich, University of Zurich", "country": "CH", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 18, "n_ref_uni": 50, "n_ref": 126, "n_ref_all": 138, "n_fig": 6, "n_tab": 7, "L_tab": 1799, "n_element_tab": 90, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 1144, "n_element_tab_1": 20, "formula_len_all": 1522, "formula_len_all_1": 859, "len_all": 231674, "len_all_1": 72277, "len_abs": 1370, "len_title": 128, "len_sents": 77667, "len_sents_1": 34486, "n_sents": 661, "n_sents_1": 257, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 80, "L_abs": 1376, "title": "Limits of Deep Learning: Sequence Modeling through the Lens of Complexity Theory", "abs": "Despite their successes, deep learning models struggle with tasks requiring complex reasoning and function composition. We present a theoretical and empirical investigation into the limitations of Structured State Space Models (SSMs) and Transformers in such tasks. We prove that one-layer SSMs cannot efficiently perform function composition over large domains without impractically large state sizes, and even with Chain-of-Thought prompting, they require a number of steps that scale unfavorably with the complexity of the function composition. Finite-precision multi-layer SSMs are constrained as Finite State Machines (FSMs), limiting their reasoning abilities. Our experiments corroborate these theoretical findings. Evaluating models on tasks including various function composition settings, multi-digit multiplication, dynamic programming, and Einstein's puzzle, we find significant performance degradation even with advanced prompting techniques. Models often resort to shortcuts, leading to compounding errors. These findings highlight fundamental barriers within current deep learning architectures rooted in their computational capacities. We underscore the need for innovative solutions to transcend these constraints and achieve reliable multi-step reasoning and compositional task-solving, which is critical for advancing toward general artificial intelligence.", "keywords": [ "theory", "complexity theory", "state space models", "deep learning architectures", "logic in computer science" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "DhYsFwLqkL", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Geunho Kim", "gender": "Male", "institution": "Broz Corp.", "country": "KR", "position": "Researcher" }, { "name": "Jin Hong", "gender": "Female", "institution": "Chung-Ang University", "country": "KR", "position": "PhD student" }, { "name": "Jinwook Paeng", "gender": "Male", "institution": "Broz Co.", "country": "KR", "position": "Researcher" }, { "name": "Junseok Kwon", "gender": "Male", "institution": "Chung-Ang University", "country": "KR", "position": "Associate Professor" }, { "name": "Yoojin Han", "gender": "unknown", "institution": "Chung-Ang University", "country": "KR", "position": "Undergrad student" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 3, "n_ref": 5, "n_ref_all": 23, "n_fig": 7, "n_tab": 3, "L_tab": 333, "n_element_tab": 35, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 333, "n_element_tab_1": 35, "formula_len_all": 480, "formula_len_all_1": 479, "len_all": 82709, "len_all_1": 48160, "len_abs": 1423, "len_title": 155, "len_sents": 26779, "len_sents_1": 22316, "n_sents": 214, "n_sents_1": 174, "venue": "ICLR2025", "decision": "Reject", "L_title": 107, "L_abs": 1428, "title": "Well-NeRF: Ensuring Well-Posed Neural Radiance Fields via View Frustum and Shadow Zone Based Regularization", "abs": "Neural Radiation Field (NeRF) often produces many artifacts with sparse inputs. These artifacts are primarily caused by learning in regions where position inference is not feasible. We assume that the main cause of this problem is the incorrect setting of boundary conditions in the learning space. To address this issue, we propose a new regularization method based on two key assumptions: (1) the position of density and color cannot be inferred in regions where the view frustum does not intersect, and (2) information inside opaque surfaces cannot be observed and inferred, and thus cannot contribute to the rendering of the image. Our method aims to transform the NeRF model into a well-posed problem by regularizing learning in regions where position inference is not possible, allowing the network to converge meaningfully. Our approach does not require scene-specific optimization and focuses on regions where position inference is not possible, thereby avoiding degradation of model performance in main regions. Experimental results demonstrate the effectiveness of our method in addressing the sparse input problem, showing outstanding performance on the Blender synthetic datasets. Our method is designed to integrate seamlessly with existing techniques, providing an effective solution for sparse input scenarios and offering a foundational approach that serves as the first clue in addressing sparse input problems.", "keywords": [ "Few-shot NeRF", "Ill-posed problem", "Artifacts removal", "View frustum", "Inside opaque", "Boundary condition", "Near-far threshold", "Integrated model" ], "rating_list": [ 6, 3, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3, 2 ], "presentation_list": [ 2, 1, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 3, 2 ], "confidence_list": [ 5, 4, 5, 5, 3 ] }, { "paper_id": "DhHIw9Nbl1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Minsi Ren", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Intern" }, { "name": "Yan-Ming Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "yi chen", "gender": "Male", "institution": ", Institute of automation, Chinese academy of science", "country": "CN", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 15, "n_ref_uni": 45, "n_ref": 59, "n_ref_all": 88, "n_fig": 13, "n_tab": 3, "L_tab": 1642, "n_element_tab": 129, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 1523, "n_element_tab_1": 120, "formula_len_all": 1463, "formula_len_all_1": 1192, "len_all": 133983, "len_all_1": 59756, "len_abs": 1251, "len_title": 117, "len_sents": 38979, "len_sents_1": 25279, "n_sents": 294, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1255, "title": "Decoupling Layout from Glyph in Online Chinese Handwriting Generation", "abs": "Text plays a crucial role in the transmission of human civilization, and teaching machines to generate online handwritten text in various styles presents an interesting and significant challenge. However, most prior work has concentrated on generating individual Chinese fonts, leaving complete text line generation largely unexplored. In this paper, we identify that text lines can naturally be divided into two components: layout and glyphs. Based on this division, we designed a text line layout generator coupled with a diffusion-based stylized font synthesizer to address this challenge hierarchically. More concretely, the layout generator performs in-context-like learning based on the text content and the provided style references to generate positions for each glyph autoregressively. Meanwhile, the font synthesizer which consists of a character embedding dictionary, a multi-scale calligraphy style encoder and a 1D U-Net based diffusion denoiser will generate each font on its position while imitating the calligraphy style extracted from the given style references. Qualitative and quantitative experiments on the CASIA-OLHWDB demonstrate that our method is capable of generating structurally correct and indistinguishable imitation samples.", "keywords": [ "Online handwriting generation; Layout generation; Calligraphy imitation; Conditional diffusion Model" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 4, 2, 3, 3 ], "confidence_list": [ 5, 5, 3, 4 ] }, { "paper_id": "DhH3LbA6F6", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Bryan Wilder", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Lily Xu", "gender": "Female", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" }, { "name": "Milind Tambe", "gender": "unknown", "institution": "Harvard University", "country": "US", "position": "Full Professor" }, { "name": "Elias Khalil", "gender": "Male", "institution": "Department of Mechanical and Industrial Engineering, University of Toronto, Toronto University", "country": "", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 9, "n_ref_uni": 46, "n_ref": 68, "n_ref_all": 94, "n_fig": 10, "n_tab": 2, "L_tab": 1547, "n_element_tab": 84, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1378, "formula_len_all_1": 580, "len_all": 158568, "len_all_1": 68897, "len_abs": 1495, "len_title": 126, "len_sents": 49586, "len_sents_1": 35158, "n_sents": 349, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1501, "title": "Reinforcement learning with combinatorial actions for coupled restless bandits", "abs": "Reinforcement learning (RL) has increasingly been applied to solve real-world planning problems, with progress in handling large state spaces and time horizons. However, a key bottleneck in many domains is that RL methods cannot accommodate large, combinatorially structured action spaces. In such settings, even representing the set of feasible actions at a single step may require a complex discrete optimization formulation. We leverage recent advances in embedding trained neural networks into optimization problems to propose SEQUOIA, an RL algorithm that directly optimizes for long-term reward over the feasible action space. Our approach embeds a Q-network into a mixed-integer program to select a combinatorial action in each timestep. Here, we focus on planning over restless bandits, a class of planning problems which capture many real-world examples of sequential decision making. We introduce coRMAB, a broader class of restless bandits with combinatorial actions that cannot be decoupled across the arms of the restless bandit, requiring direct solving over the joint, exponentially large action space. We empirically validate SEQUOIA on four novel restless bandit problems with combinatorial constraints: multiple interventions, path constraints, bipartite matching, and capacity constraints. Our approach significantly outperforms existing methods\u2014which cannot address sequential planning and combinatorial selection simultaneously\u2014by an average of 28.3% on these difficult instances.", "keywords": [ "reinforcement learning", "combinatorial optimization", "restless bandits", "mixed-integer programming" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 2 ] }, { "paper_id": "Dgh5GXsW65", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Kamil Deja", "gender": "unknown", "institution": "IDEAS NCBR", "country": "PL", "position": "Postdoc" }, { "name": "\u0141ukasz Kuci\u0144ski", "gender": "Male", "institution": "IDEAS NCBR", "country": "PL", "position": "Senior Research Scientist" }, { "name": "\u0141ukasz Staniszewski", "gender": "Male", "institution": "Warsaw University of Technology", "country": "PL", "position": "MS student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 28, "n_ref": 44, "n_ref_all": 60, "n_fig": 7, "n_tab": 2, "L_tab": 222, "n_element_tab": 36, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 222, "n_element_tab_1": 36, "formula_len_all": 391, "formula_len_all_1": 391, "len_all": 80033, "len_all_1": 69060, "len_abs": 1035, "len_title": 103, "len_sents": 21281, "len_sents_1": 21129, "n_sents": 131, "n_sents_1": 129, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1037, "title": "There and Back Again: On the relation between noises, images, and their inversions in diffusion models", "abs": "Denoising Diffusion Probabilistic Models (DDPMs) achieve state-of-the-art performance in synthesizing new images from random noise, but they lack meaningful latent space that encodes data into features. Recent DDPM-based editing techniques try to mitigate this issue by inverting images back to their approximated staring noise. In this work, we study the relation between the initial Gaussian noise, the samples generated from it, and their corresponding latent encodings obtained through the inversion procedure. First, we interpret their spatial distance relations to show the inaccuracy of the DDIM inversion technique by localizing latent representations manifold between the initial noise and generated samples. Then, we demonstrate the peculiar relation between initial Gaussian noise and its corresponding generations during diffusion training, showing that the high-level features of generated images stabilize rapidly, keeping the spatial distance relationship between noises and generations consistent throughout the training.", "keywords": [ "diffusion models", "latent space", "ddim", "generative models" ], "rating_list": [ 8, 8, 3, 3 ], "soundness_list": [ 2, 3, 3, 1 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "DgaY5mDdmT", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Filip Ilievski", "gender": "Male", "institution": "Vrije Universiteit Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Jiarui Zhang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Mahyar Khayatkhoei", "gender": "Male", "institution": "USC/ISI", "country": "US", "position": "Researcher" }, { "name": "Prateek Chhikara", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 32, "n_ref": 63, "n_ref_all": 87, "n_fig": 4, "n_tab": 4, "L_tab": 1070, "n_element_tab": 127, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 1154, "n_element_tab_1": 13, "formula_len_all": 47, "formula_len_all_1": 21, "len_all": 138183, "len_all_1": 60281, "len_abs": 1486, "len_title": 144, "len_sents": 43402, "len_sents_1": 30649, "n_sents": 287, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1493, "title": "MLLMs Know Where to Look: Training-free Perception of Small Visual Details with Multimodal LLMs", "abs": "Multimodal Large Language Models (MLLMs) have recently achieved promising performance on visual question answering (VQA)---a fundamental task affecting various downstream applications and domains. Given MLLMs' potential integration into many critical VQA applications, it is important to understand the limits of their perception. In this work, we study whether MLLMs can perceive small details as well as large details in images. In particular, we observe that their accuracy in answering visual questions is very sensitive to the size of the visual subject of the question. We further show that this effect is causal by observing that human visual cropping can significantly mitigate this sensitivity. Next, we study the attention patterns of MLLMs when answering visual questions, and intriguingly find that they consistently know where to look, even when they provide the wrong answer. Based on these findings, we then construct automatic visual cropping methods that leverage the internal knowledge of any MLLM itself, in the form of attention and gradient maps, to help it better perceive the small visual subject of any question. We study our proposed methods on two MLLMs and seven visual question answering benchmarks, and show that they can significantly improve MLLMs accuracy without requiring any training. Our findings suggest that MLLMs should be used with caution in detail-sensitive applications, and that visual cropping is a promising direction to improve their performance.", "keywords": [ "Multimodal Large Language Models", "Visual Details", "Attention", "Gradients", "Bias", "Perception", "Localization" ], "rating_list": [ 8, 8, 5 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 4, 4, 2 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "DgGdQo3iIR", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Andrea Acquaviva", "gender": "Male", "institution": "University of Bologna", "country": "", "position": "Full Professor" }, { "name": "Emanuele Parisi", "gender": "Male", "institution": "University of Bologna", "country": "IT", "position": "PhD student" }, { "name": "Federico Cichetti", "gender": "Male", "institution": "University of Bologna", "country": "", "position": "Researcher" }, { "name": "Francesco Barchi", "gender": "Male", "institution": "University of Bologna", "country": "IT", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 40, "n_ref": 91, "n_ref_all": 99, "n_fig": 4, "n_tab": 3, "L_tab": 2143, "n_element_tab": 247, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 196, "n_element_tab_1": 13, "formula_len_all": 483, "formula_len_all_1": 483, "len_all": 125471, "len_all_1": 65297, "len_abs": 1554, "len_title": 129, "len_sents": 34249, "len_sents_1": 33061, "n_sents": 219, "n_sents_1": 200, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1563, "title": "GEPCode: A Context-Aware 1M-Parameters Graph-Based Language Model for Source Code", "abs": "The pursuit of optimal conditions for software execution poses a complex challenge. This task can be automated by harnessing the structured nature of programming languages, especially from compiler intermediate representations of code (IR). The manipulation of source code using Large Language Models (LLMs) is a thriving area of study in Natural Language Processing (NLP) literature. However, in this study we illustrate how we can circumvent the need for exceedingly large models by employing domain-specific language models. These models have a reduced number of parameters but retain the ability to capture the relationships within source code elements. We introduce GEPCode, a graph neural network designed to model IR with the flexibility to adapt to new tasks. This flexibility is obtained through special \"meta\" nodes, that allow for the representation of additional task-dependent contextual information. Pre-training is performed by solving node and graph-level tasks, resulting in a general language model. After a fine-tuning phase on two downstream tasks, Device Mapping and Algorithm Classification, we achieve average accuracy results of 88.9% (NVIDIA) and 92.3% (AMD) for the former and 97.2% for the latter. Comparing our methodology with state-of-the-art models trained from scratch, our results are similar or better, yet providing a more flexible model. Moreover, we achieve similar accuracy results in downstream tasks compared to state-of-the-art pre-trained language models based on Transformers, while utilizing 100 times fewer parameters.", "keywords": [ "graph neural network", "graph language model", "source code optimization" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "DfTWrTwLzD", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Soujanya Poria", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "", "position": "Assistant Professor" }, { "name": "Wei Han", "gender": "Male", "institution": "Singapore University of Technology and Design", "country": "", "position": "PhD student" }, { "name": "Pan Zhou", "gender": "Male", "institution": "Singapore Management University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 3, "n_ref_uni": 38, "n_ref": 81, "n_ref_all": 98, "n_fig": 3, "n_tab": 5, "L_tab": 1728, "n_element_tab": 213, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1898, "n_element_tab_1": 154, "formula_len_all": 738, "formula_len_all_1": 212, "len_all": 139560, "len_all_1": 63754, "len_abs": 2053, "len_title": 131, "len_sents": 38670, "len_sents_1": 29507, "n_sents": 272, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1824, "title": "Two Are Better than One: Context Window Extension with Multi-Grained Self-Injection", "abs": "Limited longtext window has been an inherent constraint for large language models (LLMs), which significantly restricts their application scenarios. Continual pre-training on long-context data is the most straightforward approach to further extend an LLM's context window, but it is at the expense of huge data acquisition and computation cost. \nThere are many cost-efficient context window extension methods which do not require pretraining process emerges as appealing solutions, such as extrapolation, attention manipulation, context compression, etc.\nIn this paper, we propose a novel approach named Shared-LLaMA. \nShared-LLaMA is composed of two short-context LLMs. \nOne of them works as compressor and the other works as decoder. \nThe decoder receives compressed multi-grained context information from \nthe compressor and performs context-aware modeling on the running text. \nInformation transfer between the compressor and decoder occurs only at the lowest layers to circumvent an entire forward pass and save the inference time. \nBoth LLMs are initialized from the same off-the-shelf checkpoint and thus can be directly trained without extra feature alignment stages.\nAdditionally, we propose a tree structure to store the multi-grained information and design a search algorithm to fast locate and retrieve related information from each level of that tree. \nWith these efficient design choices, Shared-LLaMA can greatly reduce memory consumption, and achieves apparent speed up over other advanced baselines (2$\\times$ over streaming, $3\\times$ over encoder-decoder architectures).\nIn our evaluation on long-context modeling and understanding tasks, Shared-LLaMA yields superior or comparable results to several strong baselines, indicating Shared-LLaMA achieves a good balance between efficiency and effectiveness.", "keywords": [ "long-context modeling", "large language models" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 5, 3, 5 ] }, { "paper_id": "DfOYQZOilp", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Bohu Huang", "gender": "Male", "institution": "Xidian University", "country": "", "position": "Lecturer" }, { "name": "Fei Cheng", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Lecturer" }, { "name": "Kangye Ji", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Undergrad student" }, { "name": "Qichang Zhang", "gender": "unknown", "institution": "Xidian University", "country": "CN", "position": "MS student" }, { "name": "Zeqing Wang", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 12, "n_formula_1": 10, "n_ref_uni": 44, "n_ref": 84, "n_ref_all": 100, "n_fig": 5, "n_tab": 5, "L_tab": 2499, "n_element_tab": 135, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1767, "n_element_tab_1": 94, "formula_len_all": 686, "formula_len_all_1": 535, "len_all": 124047, "len_all_1": 65226, "len_abs": 1503, "len_title": 116, "len_sents": 31799, "len_sents_1": 28301, "n_sents": 283, "n_sents_1": 259, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1552, "title": "Jump-teaching: Ultra Robust and Efficient Learning with Noisy Labels", "abs": "Sample selection is the most straightforward technique to combat noisy labels, aiming to prevent mislabeled samples from degrading the robustness of neural networks. However, compounding selection bias and redundant selection operations have always remained challenging in robustness and efficiency. To mitigate selection bias, existing methods utilize disagreement in partner networks or additional forward propagation in a single network. For selection operations, they involve dataset-wise modeling or batch-wise ranking. Any of the above methods yields sub-optimal performance. In this work, we propose $\\textit{Jump-teaching}$, a novel framework for optimizing the typical workflow of sample selection. Firstly, Jump-teaching is the $\\textit{first}$ work to discover significant disagreements within a single network between different training iterations. Based on this discovery, we propose a jump-manner strategy for model updating to bridge the disagreements. We further illustrate its effectiveness from the perspective of error flow. \nSecondly, Jump-teaching designs a lightweight plugin to simplify selection operations. It creates a detailed yet simple loss distribution on an auxiliary encoding space, which helps select clean samples more effectively. In the experiments, Jump-teaching not only outperforms state-of-the-art works in terms of robustness, but also reduces peak memory usage by $0.46\\times$ and boosts training speed by up to $2.53\\times$. Notably, existing methods can also benefit from the integration with our framework.", "keywords": [ "learning with noisy labels", "machine learning", "classification" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 1 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "DexGnh0EcB", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Mi Tian", "gender": "Male", "institution": "TAL Education Group", "country": "CN", "position": "cto" }, { "name": "Tianqiao Liu", "gender": "Male", "institution": "Tomorrow Advancing Life", "country": "CN", "position": "Researcher" }, { "name": "Weiqi Luo", "gender": "Male", "institution": "Jinan University", "country": "CN", "position": "Full Professor" }, { "name": "ZhenshengFang", "gender": "Male", "institution": ", University of Hong Kong", "country": "", "position": "MS student" }, { "name": "Zitao Liu", "gender": "Male", "institution": "Jinan University", "country": "CN", "position": "Full Professor" }, { "name": "Zui Chen", "gender": "Male", "institution": "ShanghaiTech University", "country": "CN", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 20, "n_ref": 27, "n_ref_all": 73, "n_fig": 28, "n_tab": 8, "L_tab": 9588, "n_element_tab": 1384, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 968, "n_element_tab_1": 164, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 211268, "len_all_1": 52523, "len_abs": 3529, "len_title": 160, "len_sents": 49145, "len_sents_1": 26001, "n_sents": 339, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 111, "L_abs": 2145, "title": "MathEval: A Comprehensive Benchmark for Evaluating Large Language Models on Mathematical Reasoning Capabilities", "abs": "Mathematical reasoning is a fundamental aspect of intelligence, encompassing a spectrum from basic arithmetic to intricate problem-solving. Recent investigations into the mathematical abilities of large language models (LLMs) have yielded inconsistent and incomplete assessments. In response, we introduce MathEval, a comprehensive benchmark designed to methodically evaluate the mathematical problem-solving proficiency of LLMs across varied contexts, adaptation strategies, and evaluation metrics. MathEval amalgamates 19 datasets, spanning an array of mathematical domains, languages, problem types, and difficulty levels, from elementary to advanced. This diverse collection facilitates a thorough appraisal of LLM performance and is stratified by language (English and Chinese), problem category (arithmetic, competitive mathematics, and higher mathematics), and difficulty. To overcome the challenges of standardizing responses across diverse models and prompts, we've developed an automated LLM-driven pipeline for answer extraction and comparison, ensuring consistent evaluation criteria. To broaden the utility of MathEval beyond the scope of GPT-4, we have harnessed the extensive results from GPT-4 to train a deepseek-7B-based answer comparison model, enabling precise answer validation for those without access to GPT-4. This model will also be made publicly available. MathEval not only assesses mathematical proficiency but also introduces a method to identify potential data contamination within pre-training datasets. This is done by hypothesizing that enhancements in one mathematical dataset should be mirrored by advancements in correlated datasets, thus signaling potential contamination\u2014like the inadvertent inclusion of test data in the pre-training phase. To mitigate this and truly gauge progress, MathEval incorporates an annually refreshed set of problems from the latest Chinese National College Entrance Examination (Gaokao 2023), thereby benchmarking genuine advancements in mathematical problem solving skills. MathEval strives to refine the assessment of Large Language Models' (LLMs) capabilities in mathematics.", "keywords": [ "Mathematical reasoning benchmark", "Adaptation strategies", "Cross-lingual assessment" ], "rating_list": [ 3, 3, 6, 6, 3 ], "soundness_list": [ 3, 3, 4, 2, 2 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 2, 1, 3, 3, 2 ], "confidence_list": [ 3, 3, 5, 2, 5 ] }, { "paper_id": "Dem5LyVk8R", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Shangtong Zhang", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "", "position": "Assistant Professor" }, { "name": "Shuze Liu", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" }, { "name": "Yuxin Chen", "gender": "Female", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Undergrad student" } ], "n_formula": 81, "n_formula_1": 49, "n_ref_uni": 53, "n_ref": 101, "n_ref_all": 152, "n_fig": 4, "n_tab": 7, "L_tab": 1517, "n_element_tab": 89, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 431, "n_element_tab_1": 48, "formula_len_all": 8983, "formula_len_all_1": 3273, "len_all": 164169, "len_all_1": 72373, "len_abs": 1034, "len_title": 129, "len_sents": 45352, "len_sents_1": 30666, "n_sents": 441, "n_sents_1": 267, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1040, "title": "Efficient Policy Evaluation with Safety Constraint for Reinforcement Learning", "abs": "In reinforcement learning, classic on-policy evaluation methods often suffer from high variance and require massive online data to attain the desired accuracy. Previous studies attempt to reduce evaluation variance by searching for or designing proper behavior policies to collect data. However, these approaches ignore the safety of such behavior policies---the designed behavior policies have no safety guarantee and may lead to severe damage during online executions. In this paper, to address the challenge of reducing variance while ensuring safety simultaneously, we propose an optimal variance-minimizing behavior policy under safety constraints. Theoretically, while ensuring safety constraints, our evaluation method is unbiased and has lower variance than on-policy evaluation. Empirically, our method is the only existing method to achieve both substantial variance reduction and safety constraint satisfaction. Furthermore, we show our method is even superior to previous methods in both variance reduction and execution safety.", "keywords": [ "Reinforcement Learning" ], "rating_list": [ 8, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 4 ], "contribution_list": [ 3, 2, 2, 4 ], "confidence_list": [ 5, 4, 3, 2 ] }, { "paper_id": "DedkG85z3c", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Bonan Min", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Principal Researcher" }, { "name": "Chao Shang", "gender": "Male", "institution": "Amazon AWS AI", "country": "US", "position": "Researcher" }, { "name": "Dan Roth", "gender": "Male", "institution": "Oracle", "country": "US", "position": "Chief Scientist" }, { "name": "Fei Wang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "PhD student" }, { "name": "Qiang Ning", "gender": "Male", "institution": "Jump Trading", "country": "US", "position": "Quantitative Researcher" }, { "name": "Sarthak Jain", "gender": "unknown", "institution": "Northeastern University", "country": "", "position": "PhD student" }, { "name": "Shuai Wang", "gender": "Male", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Vittorio Castelli", "gender": "Male", "institution": "Amazon", "country": "", "position": "Senior Science Manager" }, { "name": "Yassine Benajiba", "gender": "Not Specified", "institution": "Columbia University", "country": "", "position": "Postdoc" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 53, "n_ref": 82, "n_ref_all": 98, "n_fig": 7, "n_tab": 2, "L_tab": 930, "n_element_tab": 76, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 84, "formula_len_all_1": 0, "len_all": 164213, "len_all_1": 69089, "len_abs": 1379, "len_title": 119, "len_sents": 39810, "len_sents_1": 32630, "n_sents": 350, "n_sents_1": 272, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1400, "title": "Aligning to Constraints for Data-Efficient Language Model Customization", "abs": "General-purpose language models (LMs) are aligned to diverse user intents, but fall short when it comes to specific applications. While finetuning is the default method for customized alignment, human annotations are often unavailable in various customization scenarios. Based on the observation that one of the main issues of LM customization is constraint adherence, we investigate the feasibility of using constraints as a bridge from general LMs to customized ones. We investigate common constraints in NLP tasks, categorize them into three classes based on the types of their arguments, and propose a unified framework, ACT (Aligning to ConsTraints), to automatically produce supervision signals for user alignment with constraints. Specifically, ACT uses constraint verifiers, which are typically easy to implement in practice, to compute constraint satisfaction rate (CSR) of each response. It samples multiple responses for each prompt and collect preference labels based on their CSR automatically. Subsequently, ACT adapts the LM to the target task through a ranking-based learning process. Experiments on fine-grained entity typing, abstractive summarization, and temporal question answering show that ACT is able to enhance LMs' capability to adhere to different classes of constraints, thereby improving task performance comparable to or approaching that of finetuning with labeled data.", "keywords": [ "LLM Customization", "Data-efficiency", "Constraint-driven Learning" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "DeVm3YUnpj", "primary_area": "applications to robotics, autonomy, planning", "n_author": 13, "author_info_list": [ { "name": "Changsheng Zhao", "gender": "Male", "institution": "Meta Inc.", "country": "US", "position": "Researcher" }, { "name": "Dmitrii Khizbullin", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Researcher" }, { "name": "Ernie Chang", "gender": "Male", "institution": "Meta AI", "country": "US", "position": "Research Scientist" }, { "name": "J\u00fcrgen Schmidhuber", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Director, AI Initiative" }, { "name": "Mingchen Zhuge", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Raghuraman Krishnamoorthi", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Vikas Chandra", "gender": "Male", "institution": "Meta", "country": "US", "position": "Director, AI" }, { "name": "Wenyi Wang", "gender": "unknown", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Yangyang Shi", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Yuandong Tian", "gender": "Male", "institution": "Meta AI (FAIR)", "country": "", "position": "Research Scientist" }, { "name": "Yunyang Xiong", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Zechun Liu", "gender": "unknown", "institution": "Meta Inc.", "country": "US", "position": "Researcher" }, { "name": "Dylan Ashley", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 65, "n_ref": 118, "n_ref_all": 151, "n_fig": 8, "n_tab": 8, "L_tab": 1327, "n_element_tab": 79, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 312, "n_element_tab_1": 15, "formula_len_all": 10, "formula_len_all_1": 10, "len_all": 272533, "len_all_1": 61319, "len_abs": 1971, "len_title": 93, "len_sents": 75988, "len_sents_1": 29848, "n_sents": 579, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 47, "L_abs": 1428, "title": "Agent-as-a-Judge: Evaluating Agents with Agents", "abs": "Contemporary evaluation techniques are inadequate for agentic systems. These approaches either focus exclusively on final outcomes---ignoring the step-by-step nature of the thinking done by agentic systems---or require excessive manual labour. To address this, we introduce the **Agent-as-a-Judge** framework, wherein agentic systems are used to evaluate agentic systems. This is a natural extension of the LLM-as-a-Judge framework, incorporating agentic features that enable intermediate feedback for the entire task-solving processes for more precise evaluations. We apply the Agent-as-a-Judge framework to the task of code generation. To overcome issues with existing benchmarks and provide a proof-of-concept testbed for Agent-as-a-Judge, we present **DevAI**, a new benchmark of 55 realistic AI code generation tasks. DevAI includes rich manual annotations, like a total of 366 hierarchical solution requirements, which make it particularly suitable for an agentic evaluator. We benchmark three of the top code-generating agentic systems using Agent-as-a-Judge and find that our framework dramatically outperforms LLM-as-a-Judge and is as reliable as our human evaluation baseline. Altogether, we believe that this work represents a concrete step towards enabling vastly more sophisticated agentic systems. To help that, our dataset and the full implementation of Agent-as-a-Judge will be publically available at [REDACTED]", "keywords": [ "Code Generation; Agent-as-a-Judge; AI Developer; AI Judge; LLM" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 1 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "DdPeCRVyCd", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 2, "author_info_list": [ { "name": "Diego Klabjan", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "Full Professor" }, { "name": "Haemin Park", "gender": "Male", "institution": "Northwestern University", "country": "US", "position": "PhD student" } ], "n_formula": 89, "n_formula_1": 26, "n_ref_uni": 45, "n_ref": 65, "n_ref_all": 88, "n_fig": 7, "n_tab": 9, "L_tab": 1988, "n_element_tab": 171, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 894, "n_element_tab_1": 43, "formula_len_all": 10268, "formula_len_all_1": 2615, "len_all": 179918, "len_all_1": 65061, "len_abs": 1005, "len_title": 153, "len_sents": 53917, "len_sents_1": 28483, "n_sents": 462, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 105, "L_abs": 1010, "title": "Communication-Efficient Federated Low-Rank Update Algorithm and its Connection to Implicit Regularization", "abs": "Federated Learning (FL) faces significant challenges related to communication efficiency and heterogeneity. To address these issues, we explore the potential of using low-rank updates. Our theoretical analysis reveals that client's loss exhibits a higher rank structure (gradients span higher rank subspaces of Hessian) compared to the server's loss. Based on this insight, we hypothesize that constraining client-side optimization to a low-rank subspace could provide an implicit regularization effect. Consequently, we propose FedLoRU, a general low-rank update framework for FL. Our framework enforces low-rank client-side updates and accumulates these updates to form a higher-rank model. Additionally, variants of FedLoRU can adapt to environments with statistical and model heterogeneity by employing multiple or hierarchical low-rank updates. Experimental results demonstrate that FedLoRU performs comparably to full-rank algorithms and exhibits robustness to heterogeneous and large numbers of clients.", "keywords": [ "Federated Learning", "Communication-Efficient Federated Learning", "Low-Rank Nature", "Cross-Device Federated Learning" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "Dci14asFPV", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Chushan Zhang", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "PhD student" }, { "name": "Hongdong Li", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "Full Professor" }, { "name": "Ruihan Lu", "gender": "Male", "institution": "University of Queensland", "country": "AU", "position": "PhD student" }, { "name": "Zeeshan Hayder", "gender": "Male", "institution": "Australian National University", "country": "AU", "position": "Honorary Lecturer" } ], "n_formula": 24, "n_formula_1": 11, "n_ref_uni": 32, "n_ref": 58, "n_ref_all": 68, "n_fig": 6, "n_tab": 7, "L_tab": 6847, "n_element_tab": 689, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 2432, "n_element_tab_1": 193, "formula_len_all": 1897, "formula_len_all_1": 697, "len_all": 141190, "len_all_1": 56634, "len_abs": 1714, "len_title": 127, "len_sents": 36617, "len_sents_1": 25070, "n_sents": 278, "n_sents_1": 183, "venue": "ICLR2025", "decision": "withdraw", "L_title": 79, "L_abs": 1734, "title": "DPD-LoRA: Dynamic Prompt-Driven Low-Rank Adaptation for Improved Generalization", "abs": "Fine-tuning large models presents technical challenges such as catastrophic forgetting and parameter inefficiency. Low-rank Adaptation (LoRA) and Propmt Learning can help address some of these challenges by providing more compact and flexible representations. However, Low-rank approximation is susceptible to outliers and relies on the assumption of a global low-rank structure, which can be suboptimal. Additionally, Prompt learning can overfit to specific downstream tasks, reducing its effectiveness when adapting to new tasks. In this paper, we introduce $\\textbf{Dynamic Prompt-Driven Low-Rank Adaptation (DPD-LoRA)}$, a novel framework that seamlessly integrates task-specific guidance using hierarchical prompt tokens and parameter-efficient adaptation. Unlike traditional methods, task-aware prompts in the DPD-LoRA dynamically influences low-rank updates in the model's parameters, thus enabling robust adaptation and generalization across diverse tasks and mitigating the forgetting issues. We further improve the learning capabilities of the model by breaking down the standard LoRA into multiple low-rank sub-matrices, without adding additional parameters. Further, we use an adaptive loss function to guarantee alignment with the distribution of the pre-trained model. Specifically, we introduce a self-regulated mechanism to improve stability, and a soft-gated selection mechanism to decide when to activate adaptation modules to improve performance on unseen categories. Extensive experiments on 11 benchmark datasets demonstrate that DPD-LoRA significantly outperforms state-of-the-art methods in both accuracy and generalization, offering a comprehensive solution to the challenges of fine-tuning large-scale models.", "keywords": [ "Vision-Language Models", "PEFT", "Prompt Learning" ], "rating_list": [ 5, 5, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "DcZpQhVpp9", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Ganzhao Yuan", "gender": "Male", "institution": "Peng Cheng Laboratory", "country": "CN", "position": "Associate Professor" } ], "n_formula": 149, "n_formula_1": 30, "n_ref_uni": 31, "n_ref": 83, "n_ref_all": 114, "n_fig": 4, "n_tab": 2, "L_tab": 61, "n_element_tab": 1, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 19366, "formula_len_all_1": 3443, "len_all": 255525, "len_all_1": 87665, "len_abs": 3936, "len_title": 92, "len_sents": 59121, "len_sents_1": 28808, "n_sents": 706, "n_sents_1": 282, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 43, "L_abs": 1408, "title": "ADMM for Structured Fractional Minimization", "abs": "We consider a class of structured fractional minimization problems, where the numerator includes a differentiable function, a simple nonconvex nonsmooth function, a concave nonsmooth function, and a convex nonsmooth function composed with a linear operator, while the denominator is a continuous function that is either weakly convex or has a weakly convex square root. These problems are widespread and span numerous essential applications in machine learning and data science. Existing methods are mainly based on subgradient methods and smoothing proximal gradient methods, which may suffer from slow convergence and numerical stability issues. In this paper, we introduce {\\sf FADMM}, the first Alternating Direction Method of Multipliers tailored for this class of problems. {\\sf FADMM} decouples the original problem into linearized proximal subproblems, featuring two variants: one using Dinkelbach's parametric method ({\\sf FADMM-D}) and the other using the quadratic transform method ({\\sf FADMM-Q}). By introducing a novel Lyapunov function, we establish that {\\sf FADMM} converges to $\\epsilon$-approximate critical points of the problem within an oracle complexity of $\\mathcal{O}(1/\\epsilon^{3})$. Our experiments on synthetic and real-world data for sparse Fisher discriminant analysis, robust Sharpe ratio minimization, and robust sparse recovery demonstrate the effectiveness of our approach.", "keywords": [ "Fractional Minimization", "Nonconvex Optimization", "Proximal Linearized ADMM", "Nonsmooth Optimization", "Convergence Analysis" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 1, 2, 2 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 2, 4 ] }, { "paper_id": "DcJuTtfYss", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Anissa Mokraoui", "gender": "Female", "institution": "Universit\u00e9 Sorbonne Paris Nord (Paris XIII)", "country": "FR", "position": "Full Professor" }, { "name": "Bissmella Bahaduri", "gender": "Male", "institution": "Ecole Normale Sup\u00e9rieure de Cachan", "country": "FR", "position": "MS student" }, { "name": "Fangchen FENG", "gender": "Male", "institution": "Universit\u00e9 Paris Nord (Paris XIII)", "country": "FR", "position": "Associate Professor" }, { "name": "Hicham Talaoubrid", "gender": "Male", "institution": "Universit\u00e9 Paris Nord (Paris XIII)", "country": "", "position": "PhD student" }, { "name": "Zuheng Ming", "gender": "Male", "institution": "Universit\u00e9 Paris Nord (Paris XIII)", "country": "FR", "position": "Assistant Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 30, "n_ref": 54, "n_ref_all": 65, "n_fig": 3, "n_tab": 6, "L_tab": 1666, "n_element_tab": 184, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 1086, "n_element_tab_1": 149, "formula_len_all": 408, "formula_len_all_1": 408, "len_all": 104458, "len_all_1": 60696, "len_abs": 1350, "len_title": 105, "len_sents": 30968, "len_sents_1": 27119, "n_sents": 223, "n_sents_1": 193, "venue": "ICLR2025", "decision": "withdraw", "L_title": 57, "L_abs": 1357, "title": "INDIRECT ATTENTION: IA-DETR FOR ONE SHOT OBJECT DETECTION", "abs": "One-shot object detection presents a significant challenge, requiring the identification of objects within a target image using only a single sample image of the object class as query image. Attention-based methodologies have garnered considerable attention in the field of object detection. Specifically, the cross-attention module, as seen in DETR, plays a pivotal role in exploiting the relationships be-\ntween object queries and image features. However, in the context of DETR networks for one-shot object detection, the intricate interplay among target image features, query image features, and object queries must be carefully considered.\nIn this study, we propose a novel module termed \u201dindirect attention.\u201d We illustrate that relationships among target image features, query image features, and object queries can be effectively captured in a more concise manner compared to\ncross-attention. Furthermore, we introduce a pre-training pipeline tailored specifically for one-shot object detection, addressing three primary objectives: identifying objects of interest, class differentiation, and object detection based on a given\nquery image. Our experimental findings demonstrate that the proposed IA-DETR (Indirect-Attention DETR) significantly outperforms state-of-the-art one-shot object detection methods on both the Pascal VOC and COCO benchmarks.", "keywords": [ "One shot object detection", "DETR", "cross-attention" ], "rating_list": [ 5, 8, 5, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "DcG4YnbOT3", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Mingyang Yu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Undergrad student" }, { "name": "Xiahui Guo", "gender": "unknown", "institution": "East China Normal University", "country": "CN", "position": "Undergrad student" }, { "name": "Yang Shu", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhenkai Li", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Undergrad student" }, { "name": "P Chen", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 12, "n_ref_uni": 42, "n_ref": 57, "n_ref_all": 74, "n_fig": 11, "n_tab": 7, "L_tab": 1327, "n_element_tab": 73, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 411, "n_element_tab_1": 21, "formula_len_all": 1288, "formula_len_all_1": 677, "len_all": 136232, "len_all_1": 57846, "len_abs": 1643, "len_title": 148, "len_sents": 42796, "len_sents_1": 27824, "n_sents": 315, "n_sents_1": 187, "venue": "ICLR2025", "decision": "withdraw", "L_title": 100, "L_abs": 1650, "title": "Vision-Enhanced Time Series Forecasting by Decomposed Feature Extraction and Composed Reconstruction", "abs": "Time series forecasting plays a crucial role in various domains, such as power and weather forecasting. In recent years, different types of models have achieved promising results in long-term time series forecasting. However, these models often produce predictions that lack consistency with the style of the input, resulting in reduced reliability and trust in the forecasts. To address this issue, we propose the Vision-Enhanced Time Series Forecasting by Decomposed Feature Extraction and Composed Reconstruction (VisiTER), which leverages the rich semantic information provided by the image modality to enhance the realism of the predictions. It consists of two main components: the Decomposed Time Series to Image Generation and the Composed Image to Time Series Generation. In the first component, the Decomposed Time Series Feature Extraction Model extracts periodic and trend information, which is then transformed into images using our proposed time series to vision transformation architecture. After converting the input time series into images, the resulting images are used as style features and concatenated with the previously extracted features. In the second component, we use our proposed TimeIR along with the previously obtained feature set to perform image reconstruction for the prediction part. Due to the rich information provided, the reconstructed images exhibit better consistency with the input images, which are then transformed back into time series. Extensive experiments on seven real-world datasets demonstrate that VisiTER achieves state-of-the-art prediction performance on both traditional metrics and new metrics.", "keywords": [ "Time Series Forecasting" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "Dc6dgTq2UZ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Bo Hui", "gender": "unknown", "institution": "University of Tulsa", "country": "US", "position": "Assistant Professor" }, { "name": "Bohan Liu", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Xiaolong Ma", "gender": "Male", "institution": "Clemson University", "country": "US", "position": "Assistant Professor" }, { "name": "Yang Xiao", "gender": "Male", "institution": "University of Tulsa", "country": "US", "position": "PhD student" }, { "name": "Zinan Ling", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Ruimeng Ye", "gender": "Female", "institution": "University of Tulsa", "country": "US", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 22, "n_ref": 46, "n_ref_all": 60, "n_fig": 7, "n_tab": 2, "L_tab": 1, "n_element_tab": 1, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 438, "n_element_tab_1": 20, "formula_len_all": 440, "formula_len_all_1": 440, "len_all": 122038, "len_all_1": 56881, "len_abs": 1102, "len_title": 143, "len_sents": 29794, "len_sents_1": 26536, "n_sents": 238, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1116, "title": "Towards Distributed Backdoor Attacks with Network Detection in Decentralized Federated Learning", "abs": "Distributed backdoor attacks (DBA) have shown a higher attack success rate than centralized attacks in centralized federated learning (FL). However, it has not been investigated in the decentralized FL. In this paper, we experimentally demonstrate that, while directly applying DBA to decentralized FL, the attack success rate depends on the distribution of attackers in the network architecture. Considering that the attackers can not decide their location, this paper aims to achieve a high attack success rate regardless of the attackers' location distribution. Specifically, we first design a method to detect the network by predicting the distance between any two attackers on the network. Then, based on the distance, we organize the attackers in different clusters. Lastly, we propose an algorithm to \\textit{dynamically} embed local patterns decomposed from a global pattern into the different attackers in each cluster. We conduct a thorough empirical investigation and find that our method can, in benchmark datasets,\noutperform both centralized attacks and naive DBA in different decentralized frameworks.", "keywords": [ "Decentralized Federated Learning", "Distributed Backdoor Attacks" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "DblHBgD0GR", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Lichao Sun", "gender": "Male", "institution": "Lehigh University", "country": "US", "position": "Assistant Professor" }, { "name": "Ruoxi Chen", "gender": "Female", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Xun Chen", "gender": "unknown", "institution": "Samsung Research America", "country": "US", "position": "Director" }, { "name": "Yixin Liu", "gender": "Male", "institution": "Dolby", "country": "US", "position": "Intern" } ], "n_formula": 23, "n_formula_1": 13, "n_ref_uni": 72, "n_ref": 177, "n_ref_all": 198, "n_fig": 11, "n_tab": 9, "L_tab": 3579, "n_element_tab": 240, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 851, "n_element_tab_1": 75, "formula_len_all": 1608, "formula_len_all_1": 978, "len_all": 264557, "len_all_1": 67334, "len_abs": 1768, "len_title": 132, "len_sents": 81387, "len_sents_1": 33856, "n_sents": 525, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1778, "title": "Rethinking and Defending Protective Perturbation in Personalized Diffusion Models", "abs": "Personalized diffusion models (PDMs) have become prominent for adapting pre-trained text-to-image models to generate images of specific subjects using minimal training data. However, PDMs are susceptible to minor adversarial perturbations, leading to significant degradation when fine-tuned on corrupted datasets. These vulnerabilities are exploited to create protective perturbations that prevent unauthorized image generation. Existing purification methods attempt to red-team the protective perturbation to break the protection but often over-purify images, resulting in information loss. In this work, we conduct an in-depth analysis of the fine-tuning process of PDMs through the lens of shortcut learning. We hypothesize and empirically demonstrate that adversarial perturbations induce a latent-space misalignment between images and their text prompts in the CLIP embedding space. This misalignment causes the model to erroneously associate noisy patterns with unique identifiers during fine-tuning, resulting in poor generalization. Based on these insights, we propose a systematic red-teaming framework that includes data purification and contrastive decoupling learning. We first employ off-the-shelf image restoration techniques to realign images with their original semantic meanings in latent space. Then, we introduce contrastive decoupling learning with noise tokens to decouple the learning of personalized concepts from spurious noise patterns. Our study not only uncovers fundamental shortcut learning vulnerabilities in PDMs but also provides a comprehensive evaluation framework for developing stronger protection. Our extensive evaluation demonstrates its superiority over existing purification methods and stronger robustness against adaptive perturbation.", "keywords": [ "Protective Perturbations", "Imperceptible Perturbations", "Adversarial Purification", "Diffusion-based Generative Models" ], "rating_list": [ 3, 6, 8, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 1, 3, 3, 1 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "DbZDbg2z9q", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Alexandru Oarga", "gender": "Not Specified", "institution": "University of Barcelona", "country": "ES", "position": "MS student" }, { "name": "Andres M Bran", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Magdalena Lederbauer", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Visiting graduate student" }, { "name": "Matthew Hart", "gender": "unknown", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "PhD student" }, { "name": "Philippe Schwaller", "gender": "Male", "institution": "Swiss Federal Institute of Technology Lausanne", "country": "", "position": "Assistant Professor" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 64, "n_ref": 101, "n_ref_all": 117, "n_fig": 7, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 720, "n_element_tab_1": 9, "formula_len_all": 256, "formula_len_all_1": 212, "len_all": 175321, "len_all_1": 60769, "len_abs": 1678, "len_title": 112, "len_sents": 35407, "len_sents_1": 30605, "n_sents": 233, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1532, "title": "Ontology-Retrieval Augmented Generation for Scientific Discovery", "abs": "Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks, sparkling an increasing interest for their application in science. However, in scientific domains, their utility is often limited by hallucinations that violate established relationships between concepts or ignore their meaning; problems that are not entirely eliminated with Retrieval Augmented Generation (RAG) techniques. A key feature of science is the use of niche concepts, abbreviations and implicit relationships, which may deem RAG approaches less powerful due to the lack of understanding of concepts, especially in emerging and less known fields. Ontologies, as structured frameworks for organizing knowledge and establishing relationships between concepts, offer a potential solution to this challenge. In this work we introduce OntoRAG, a novel approach that enhances RAG by retrieving taxonomical knowledge from ontologies. We evaluate the performance of this method on three common biomedical benchmarks. To extend the value of OntoRAG to emerging fields, where ontologies have not yet been developed, we also present OntoGen, a methodology for generating ontologies from a set of documents. We apply the combined OntoGen+OntoRAG pipeline to a novel benchmark of scientific discovery in the emerging field of single-atom catalysis. Our results demonstrate the promise of this method for improving reasoning and suppressing hallucinations in LLMs, potentially accelerating scientific discovery across various domains.", "keywords": [ "ontology", "rag", "retrieval", "llm", "science", "ai4science", "chemistry", "biomedical", "reasoning" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 2, 2, 1, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Daq6Pw3TjN", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Bowen Tian", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Researcher" }, { "name": "Songning Lai", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Intern" }, { "name": "Yutao Yue", "gender": "unknown", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "Associate Professor" } ], "n_formula": 32, "n_formula_1": 24, "n_ref_uni": 21, "n_ref": 36, "n_ref_all": 64, "n_fig": 8, "n_tab": 13, "L_tab": 7201, "n_element_tab": 834, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2095, "n_element_tab_1": 162, "formula_len_all": 1565, "formula_len_all_1": 1293, "len_all": 176695, "len_all_1": 59306, "len_abs": 1138, "len_title": 112, "len_sents": 60732, "len_sents_1": 25377, "n_sents": 397, "n_sents_1": 157, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1049, "title": "Wolf2Pack: The AutoFusion Framework for Dynamic Parameter Fusion", "abs": "In the rapidly evolving field of deep learning, specialized models have driven significant advancements in tasks such as computer vision and natural language processing. However, this specialization leads to a fragmented ecosystem where models lack the adaptability for broader applications. To overcome this, we introduce AutoFusion, an innovative framework that integrates distinct models into a unified architecture for multi-task learning without pre-trained checkpoints. Using an unsupervised, end-to-end approach, AutoFusion dynamically blends model weights at each layer, optimizing the combination through a loss-minimization process that does not require labeled data. We validate AutoFusion\u2019s effectiveness through experiments on commonly used benchmark datasets, demonstrating superior performance over established methods like Weight Interpolation, Git Re-Basin, and ZipIt. Our framework offers a scalable and flexible solution for model integration, positioning it as a powerful tool for future research and practical applications.", "keywords": [ "Parameter Fusion", "Multi-task Model Fusion", "Computer Vision" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 1, 1, 4 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "DakTqQu161", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Jaewoo Lee", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "", "position": "MS student" }, { "name": "Jinheon Baek", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Joonho Ko", "gender": "Male", "institution": "KAIST", "country": "", "position": "MS student" }, { "name": "Soyeong Jeong", "gender": "Female", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Sung Ju Hwang", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 25, "n_ref": 71, "n_ref_all": 93, "n_fig": 5, "n_tab": 6, "L_tab": 1457, "n_element_tab": 100, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1014, "n_element_tab_1": 62, "formula_len_all": 346, "formula_len_all_1": 354, "len_all": 150958, "len_all_1": 71207, "len_abs": 1823, "len_title": 129, "len_sents": 45372, "len_sents_1": 37062, "n_sents": 272, "n_sents_1": 213, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 1783, "title": "Unified Multi-Modal Interleaved Document Representation for Information Retrieval", "abs": "Information Retrieval (IR) methods aim to identify relevant documents in response to a given query, which have gained remarkable attention due to their successful application in various natural language tasks. However, existing approaches typically consider only the textual information within the documents, which overlooks the fact that documents can contain multiple modalities, including texts, images, and tables. Further, they often segment each long document into multiple discrete passages for embedding, preventing them from capturing the overall document context and interactions between paragraphs. We argue that these two limitations lead to suboptimal document representations for retrieval. In this work, to address them, we aim to produce more comprehensive and nuanced document representations by holistically embedding documents interleaved with different modalities. Specifically, we achieve this by leveraging the capability of recent vision-language models that enable the processing and integration of text, images, and tables into a unified format and representation. Moreover, to mitigate the information loss from segmenting documents into passages, instead of representing and retrieving passages individually, we further merge the representations of segmented passages into one single document representation, while we additionally introduce a reranking strategy to decouple and identify the relevant passage within the document if necessary. Then, through extensive experiments on diverse information retrieval scenarios considering both the textual and multi-modal queries, we show that our approach substantially outperforms relevant baselines, thanks to the consideration of the multi-modal information interleaved within the documents in a unified way.", "keywords": [ "Information Retrieval", "Multi-Modal Information Retrieval", "Multi-Modal Representation Learning" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 3, 3, 5 ] }, { "paper_id": "DaUsIJe2Az", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Jianzhu Ma", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Ke Fan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Muhan Zhang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xuefeng Zhang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Yuan Zhou", "gender": "Male", "institution": "University of Illinois, Urbana Champaign", "country": "", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 37, "n_ref": 65, "n_ref_all": 96, "n_fig": 11, "n_tab": 7, "L_tab": 2416, "n_element_tab": 222, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1546, "n_element_tab_1": 187, "formula_len_all": 790, "formula_len_all_1": 564, "len_all": 134311, "len_all_1": 62808, "len_abs": 1472, "len_title": 129, "len_sents": 43806, "len_sents_1": 28803, "n_sents": 350, "n_sents_1": 215, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1429, "title": "Continual Learning via Continual Weighted Sparsity and Meta-Plasticity Scheduling", "abs": "Continual Learning (CL) is fundamentally challenged by the stability-plasticity dilemma: the trade-off between acquiring new information and maintaining past knowledge. To address the stability, many methods keep a replay buffer containing a small set of samples from prior tasks and employ parameter isolation strategies that allocate separate parameter subspaces for each task, reducing interference between tasks. To get more refined, task-specific groups, we adapt a dynamic sparse training technique and introduce a continual weight score function to guide the iterative pruning process over multiple rounds of training. We refer to this method as the continual weighted sparsity scheduler. Furthermore, with more incremental tasks introduced, the network inevitably becomes saturated, leading to a loss of plasticity, where the model's adaptability decreases due to dormant or saturated neurons. To mitigate this, we draw inspiration from biological meta-plasticity mechanisms, and develop a meta-plasticity scheduler to dynamically adjust these task-specific groups' learning rates based on the sensitive score function we designed, ensuring a balance between retaining old knowledge and acquiring new skills. The results of comparison on popular datasets demonstrate that our approach consistently outperforms existing state-of-the-art methods, confirming its effectiveness in managing the stability-plasticity trade-off.", "keywords": [ "Continual learning" ], "rating_list": [ 3, 6, 5, 3 ], "soundness_list": [ 1, 3, 2, 2 ], "presentation_list": [ 1, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "DaA0wAcTY7", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 14, "author_info_list": [ { "name": "Andre Araujo", "gender": "Male", "institution": "Google DeepMind", "country": "BR", "position": "Researcher" }, { "name": "Arjun Karpur", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Researcher" }, { "name": "Bingyi Cao", "gender": "Female", "institution": "Google Research", "country": "US", "position": "Software Engineer" }, { "name": "Dan Gnanapragasam", "gender": "Male", "institution": "Research, Google", "country": "US", "position": "Researcher" }, { "name": "Daniel Salz", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Researcher" }, { "name": "Guangxing Han", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Howard Zhou", "gender": "Male", "institution": "Google", "country": "US", "position": "Software Engineer" }, { "name": "Jan Dlabal", "gender": "unknown", "institution": "Research, Google", "country": "", "position": "Software Engineer" }, { "name": "Kaifeng Chen", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Kevis-kokitsi Maninis", "gender": "Male", "institution": "Google", "country": "CH", "position": "Research Scientist" }, { "name": "Koert Chen", "gender": "unknown", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Mojtaba Seyedhosseini", "gender": "Male", "institution": "Google", "country": "", "position": "Software Engineer Researcher" }, { "name": "Soham Ghosh", "gender": "Male", "institution": "Mistral AI", "country": "FR", "position": "Researcher" }, { "name": "Ye Xia", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 71, "n_ref": 147, "n_ref_all": 167, "n_fig": 8, "n_tab": 9, "L_tab": 2209, "n_element_tab": 361, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 635, "n_element_tab_1": 112, "formula_len_all": 240, "formula_len_all_1": 145, "len_all": 237442, "len_all_1": 67410, "len_abs": 1718, "len_title": 99, "len_sents": 61287, "len_sents_1": 32769, "n_sents": 438, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1725, "title": "TIPS: Text-Image Pretraining with Spatial awareness", "abs": "While image-text representation learning has become very popular in recent years, existing models tend to lack spatial awareness and have limited direct applicability for dense understanding tasks. For this reason, self-supervised pretraining is still the go-to method for many dense vision applications (e.g. depth estimation, semantic segmentation), despite the lack of explicit supervisory signals. In this paper, we close this gap between image-text and self-supervised learning, by proposing a novel general-purpose image-text model, which can be effectively used off-the-shelf for dense and global vision tasks. Our method, which we refer to as Text-Image Pretraining with Spatial awareness (TIPS), leverages two simple and effective insights. First, on textual supervision: we reveal that replacing noisy web image captions by synthetically generated textual descriptions boosts dense understanding performance significantly, due to a much richer signal for learning spatially aware representations. We propose an adapted training method that combines noisy and synthetic captions, resulting in improvements across both dense and global understanding tasks. Second, on the learning technique: we propose to combine contrastive image-text learning with self-supervised masked image modeling, to encourage spatial coherence, unlocking substantial enhancements for downstream applications. Building on these two ideas, we scale our model using the transformer architecture, trained on a curated set of public images. Our experiments are conducted on 8 tasks involving 16 datasets in total, demonstrating strong off-the-shelf performance on both dense and global understanding, for several image-only and image-text tasks.", "keywords": [ "image representations", "image-text", "vision-language", "dense understanding", "computer vision" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 2, 3, 4, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "Da3j02cHe0", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Seungjun Lee", "gender": "Male", "institution": "Brookhaven National Laboratory", "country": "US", "position": "Postdoc" }, { "name": "Shinjae Yoo", "gender": "Male", "institution": "Brookhaven National Lab", "country": "", "position": "Scientist" } ], "n_formula": 45, "n_formula_1": 33, "n_ref_uni": 55, "n_ref": 112, "n_ref_all": 129, "n_fig": 5, "n_tab": 4, "L_tab": 755, "n_element_tab": 72, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 542, "n_element_tab_1": 64, "formula_len_all": 1746, "formula_len_all_1": 1211, "len_all": 144218, "len_all_1": 63844, "len_abs": 1285, "len_title": 124, "len_sents": 41443, "len_sents_1": 29134, "n_sents": 275, "n_sents_1": 187, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1292, "title": "Efficient Physics-Constrained Diffusion Models for Solving Inverse Problems", "abs": "Solving inverse problems in scientific and engineering domains often involves complex, nonlinear forward physics and ill-posed conditions. \nRecent advancements in diffusion model have shown promise for general inverse problems, yet their application to scientific domains remains less explored and is hindered by the complexity and high non-linearity of physics constraints. We present a physics-constrained diffusion model (PCDM) designed to solve inverse problems in scientific and engineering domains by efficiently integrating pre-trained diffusion models and physics-constrained objectives.\nWe leverage accelerated diffusion sampling to enable a practical generation process while strictly adhering to physics constraints by solving optimization problems at each timestep. By decoupling the likelihood optimization from the reverse diffusion steps, we ensure that the solutions remain physically consistent, even when employing fewer sampling steps.\nWe validate our method on a wide range of challenging physics-constrained inverse problems, including data assimilation, topology optimization, and full-waveform inversion. Experimental results show that our approach significantly outperforms existing methods in efficiency and precision, making it practical for real-world applications.", "keywords": [ "physics-constraints inverse problem", "diffusion model", "PDE", "generative modeling" ], "rating_list": [ 1, 5, 3, 6, 3 ], "soundness_list": [ 1, 2, 2, 3, 1 ], "presentation_list": [ 1, 2, 2, 2, 1 ], "contribution_list": [ 1, 2, 1, 2, 2 ], "confidence_list": [ 5, 4, 4, 4, 5 ] }, { "paper_id": "DZcmz9wU0i", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Adrien Vacher", "gender": "unknown", "institution": "Universit\u00e9 Paris-Est", "country": "", "position": "PhD student" }, { "name": "Anna Korba", "gender": "unknown", "institution": "Ensae ParisTech", "country": "FR", "position": "Assistant Professor" }, { "name": "Austin J Stromme", "gender": "unknown", "institution": "Ecole Nationale de la Statistique et de l'Administration Economique", "country": "", "position": "Assistant Professor" }, { "name": "Omar Chehab", "gender": "unknown", "institution": "Ecole Nationale de la Statistique et de l'Administration Economique", "country": "FR", "position": "Postdoc" } ], "n_formula": 207, "n_formula_1": 30, "n_ref_uni": 65, "n_ref": 136, "n_ref_all": 147, "n_fig": 3, "n_tab": 1, "L_tab": 369, "n_element_tab": 5, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 19395, "formula_len_all_1": 1775, "len_all": 262302, "len_all_1": 71096, "len_abs": 1106, "len_title": 81, "len_sents": 76050, "len_sents_1": 34448, "n_sents": 745, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1114, "title": "Provable Convergence and Limitations of Geometric Tempering for Langevin Dynamics", "abs": "Geometric tempering is a popular approach to sampling from challenging multi-modal probability distributions by instead sampling from a sequence of distributions which interpolate, using the geometric mean, between an easier proposal distribution and the target distribution. In this paper, we theoretically investigate the soundness of this approach when the sampling algorithm is Langevin dynamics, proving both upper and lower bounds. Our upper bounds are the first analysis in the literature under functional inequalities. They assert the convergence of tempered Langevin in continuous and discrete-time, and their minimization leads to closed-form optimal tempering schedules for some pairs of proposal and target distributions. Our lower bounds demonstrate a simple case where the geometric tempering takes exponential time, and further reveal that the geometric tempering can suffer from poor functional inequalities and slow convergence, even when the target distribution is well-conditioned. Overall, our results indicate that the geometric tempering may not help, and can even be harmful for convergence.", "keywords": [ "Sampling", "Langevin", "Annealing" ], "rating_list": [ 8, 8, 6, 6 ], "soundness_list": [ 4, 4, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "DZBFchnM3b", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Leonid Karlinsky", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Principal Researcher" }, { "name": "Nasim Borazjanizadeh", "gender": "unknown", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" }, { "name": "Roei Herzig", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Postdoc" }, { "name": "Rogerio Feris", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Research Manager" }, { "name": "trevor darrell", "gender": "Male", "institution": "Electrical Engineering & Computer Science Department", "country": "US", "position": "Professor" } ], "n_formula": 21, "n_formula_1": 0, "n_ref_uni": 41, "n_ref": 448, "n_ref_all": 547, "n_fig": 8, "n_tab": 21, "L_tab": 14039, "n_element_tab": 394, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 719, "n_element_tab_1": 35, "formula_len_all": 2111, "formula_len_all_1": 0, "len_all": 417066, "len_all_1": 64426, "len_abs": 1841, "len_title": 141, "len_sents": 169371, "len_sents_1": 30732, "n_sents": 948, "n_sents_1": 224, "venue": "ICLR2025", "decision": "withdraw", "L_title": 96, "L_abs": 1375, "title": "Navigating the Labyrinth: Evaluating and Enhancing LLMs\u2019 Ability to Reason About Search Problems", "abs": "Recently, Large Language Models (LLMs) attained impressive performance in math and reasoning benchmarks. However, they still often struggle with multi-step reasoning which is relatively easy for humans. To further investigate this, we introduce a new benchmark, SearchBench, containing 11 unique combinatorial problems that avoid training contamination (each equipped with automated pipelines to generate an arbitrary number of instances) and analyze the feasibility, correctness, and optimality of LLM-generated solutions. We show that even the most advanced LLMs fail to solve these problems end-to-end in text, e.g., GPT4 and o1-preview respectively solve only 1.4% and 18.6% correctly. SearchBench problems require considering multiple pathways to the solution and backtracking, posing a significant challenge to auto-regressive models. Instructing LLMs to generate code that solves the problem helps only slightly. We next introduce an in-context learning approach that prompts the model to implement A*, an informed search algorithm, to comprehensively traverse the problem state space, improving the performance of models. We further extend this approach and propose the Multi-Stage-Multi-Try inference method which breaks down the A* algorithm implementation into two stages and auto-verifies the first stage against unit tests, raising GPT-4's performance above 57%.", "keywords": [ "Mathematical & reasoning benchmark", "Search & Combinatorial problems", "A* algorithm" ], "rating_list": [ 3, 5, 3 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "DYXl6P70aH", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Alvard Barseghyan", "gender": "Female", "institution": "YerevaNN research Lab", "country": "", "position": "Researcher" }, { "name": "Ani Vanyan", "gender": "Female", "institution": "YerevaNN research Lab", "country": "", "position": "Researcher" }, { "name": "Anna Khosrovyan", "gender": "Female", "institution": "Yerevan State University", "country": "AM", "position": "MS student" }, { "name": "Hakob Tamazyan", "gender": "Male", "institution": "Yerevan State University", "country": "", "position": "PhD student" }, { "name": "Hrant Khachatrian", "gender": "unknown", "institution": "Yerevan State University", "country": "AM", "position": "Principal Researcher" }, { "name": "Tigran Galstyan", "gender": "unknown", "institution": "Russian-Armenian University", "country": "", "position": "PhD student" }, { "name": "Vahan Huroyan", "gender": "Male", "institution": "YerevaNN Research Lab", "country": "", "position": "Researcher" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 39, "n_ref": 95, "n_ref_all": 114, "n_fig": 4, "n_tab": 5, "L_tab": 2136, "n_element_tab": 170, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 327, "n_element_tab_1": 20, "formula_len_all": 118, "formula_len_all_1": 0, "len_all": 158143, "len_all_1": 62860, "len_abs": 1185, "len_title": 67, "len_sents": 44428, "len_sents_1": 31143, "n_sents": 327, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1190, "title": "Benchmarking Robustness of Foundation Models for Remote Sensing", "abs": "Foundation models have significantly advanced machine learning applications across various modalities, including images. Recently numerous attempts have been made on developing foundation models specifically tailored for remote sensing applications, predominantly through masked image modeling techniques. This work explores the essential characteristics and performance expectations for a foundation model in aerial imagery. We introduce a benchmark designed to evaluate the model's performance as well as robustness to changes in scale and spectral bands of the input. Our benchmarks encompass tasks unique to aerial imagery, such as change detection and scene classification, and utilize publicly available datasets RESISC45, BigEarthNet, LEVIR-CD and OSCD. We evaluate recently proposed foundation models on the benchmark. Furthermore, we explore the impact of various design choices in pretraining and fine-tuning on the performance of the models on our benchmark. Specifically, we pretrain several variations of a self-distillation based self-supervised model on aerial imagery datasets, including one without scale-augmentations and another one with a pretrained mask decoder module.", "keywords": [ "aerial imagery", "foundation models", "self-supervised learning", "benchmark" ], "rating_list": [ 3, 5, 1, 3 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 1, 2, 1, 3 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "DYVSLfiyRN", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Alexander Robey", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Postdoc" }, { "name": "Andy Zou", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Chengming Xu", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Haoqi Hu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Kai Hu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Matt Fredrikson", "gender": "Male", "institution": "Gray Swan AI", "country": "US", "position": "Researcher" }, { "name": "Weichen Yu", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 48, "n_ref": 78, "n_ref_all": 98, "n_fig": 2, "n_tab": 6, "L_tab": 1199, "n_element_tab": 202, "n_fig_1": 1, "n_tab_1": 7, "L_tab_1": 3197, "n_element_tab_1": 236, "formula_len_all": 329, "formula_len_all_1": 273, "len_all": 147556, "len_all_1": 62427, "len_abs": 1324, "len_title": 119, "len_sents": 34028, "len_sents_1": 26139, "n_sents": 242, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1335, "title": "Transferable Adversarial Attack on Vision-enabled Large Language Models", "abs": "Vision-enabled Large Language Models (VLLMs) are increasingly deployed to offer advanced capabilities on inputs comprising both text and images. While prior research has shown that adversarial attacks can transfer from open-source to proprietary black-box models in text-only and vision-only contexts, the extent and effectiveness of such vulnerabilities remain underexplored for VLLMs. \nWe present a comprehensive analysis demonstrating that targeted adversarial examples are highly transferable to widely-used proprietary VLLMs such as GPT-4o, Claude, and Gemini. \nWe show that attackers can craft perturbations to induce specific attacker-chosen interpretations of visual information, such as misinterpreting hazardous content as safe, overlooking sensitive or restricted material, or generating detailed incorrect responses aligned with the attacker's intent. \nFurthermore, we discover that universal perturbations---modifications applicable to a wide set of images---can consistently induce these misinterpretations across multiple proprietary VLLMs. \nOur experimental results on object recognition, visual question answering, and image captioning show that this vulnerability is common across current state-of-the-art models, and underscore an urgent need for robust mitigations to ensure the safe and secure deployment of VLLMs.", "keywords": [ "adversarial attack", "black-box attack", "transferable attack", "vision-enabled large language models" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "DXaUC7lBq1", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Di Wang", "gender": "Male", "institution": "KAUST", "country": "SA", "position": "Assistant Professor" }, { "name": "Liang Liu", "gender": "Male", "institution": "Soochow University", "country": "CN", "position": "MS student" }, { "name": "Lijie Hu", "gender": "Female", "institution": "KAUST", "country": "SA", "position": "PhD student" }, { "name": "Mengdi Li", "gender": "Male", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "Postdoc" }, { "name": "Ruoxuan Bao", "gender": "Female", "institution": "Shanghai University", "country": "CN", "position": "Undergrad student" }, { "name": "Shenzhe Zhu", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "Undergrad student" }, { "name": "Shu Yang", "gender": "Female", "institution": "King Abdullah University of Science and Technology", "country": "SA", "position": "PhD student" }, { "name": "Yu Cheng", "gender": "unknown", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 2, "n_ref_uni": 44, "n_ref": 77, "n_ref_all": 113, "n_fig": 2, "n_tab": 9, "L_tab": 8044, "n_element_tab": 618, "n_fig_1": 0, "n_tab_1": 7, "L_tab_1": 3045, "n_element_tab_1": 106, "formula_len_all": 550, "formula_len_all_1": 165, "len_all": 208500, "len_all_1": 64306, "len_abs": 892, "len_title": 147, "len_sents": 58017, "len_sents_1": 28844, "n_sents": 385, "n_sents_1": 170, "venue": "ICLR2025", "decision": "withdraw", "L_title": 98, "L_abs": 894, "title": "What Makes Your Model a Low-empathy or Warmth Person: Exploring the Origins of Personality in LLMs", "abs": "Large language models (LLMs) have demonstrated remarkable capabilities in generating human-like text and exhibiting personality traits similar to those in humans. However, the mechanisms by which LLMs encode and express traits such as agreeableness and impulsiveness remain poorly understood. Drawing on the theory of social determinism, we investigate how long-term background factors, such as family environment and cultural norms, interact with short-term pressures like external instructions, shaping and influencing LLMs' personality traits. By steering the output of LLMs through the utilization of interpretable features within the model, we explore how these background and pressure factors lead to changes in the model's traits without the need for further fine-tuning. Additionally, we suggest the potential impact of these factors on model safety from the perspective of personality.", "keywords": [ "explainable ai", "personality of LLM" ], "rating_list": [ 1, 3, 3, 5 ], "soundness_list": [ 1, 1, 1, 2 ], "presentation_list": [ 2, 1, 1, 2 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 5, 3, 4, 4 ] }, { "paper_id": "DWa1bATAot", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Evgeny Burnaev", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Full Professor" }, { "name": "Ilya Trofimov", "gender": "unknown", "institution": "Skoltech", "country": "RU", "position": "Research scientist" }, { "name": "Maria Ivanova", "gender": "Female", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Nikita Ivanisenko", "gender": "Male", "institution": "InstaDeep", "country": "DE", "position": "Researcher" }, { "name": "Olga Kardymon", "gender": "unknown", "institution": "AIRI", "country": "", "position": "Researcher" }, { "name": "Pavel Strashnov", "gender": "unknown", "institution": "Artificial Intelligence Research Institute (AIRI)", "country": "RU", "position": "Researcher" }, { "name": "Serguei Barannikov", "gender": "unknown", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "Leading Research Scientist" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 58, "n_ref": 123, "n_ref_all": 160, "n_fig": 10, "n_tab": 7, "L_tab": 3962, "n_element_tab": 438, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 1114, "n_element_tab_1": 168, "formula_len_all": 185, "formula_len_all_1": 185, "len_all": 184859, "len_all_1": 66546, "len_abs": 1015, "len_title": 86, "len_sents": 49087, "len_sents_1": 31355, "n_sents": 376, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1019, "title": "Exploiting Topology of Protein Language Model Attention Maps for Token Classification", "abs": "In this paper, we introduce a method to extract topological features from transformer-based protein language models. Our method leverages the persistent homology of attention maps to generate features for token (per amino-acid) classification tasks and demonstrate its relevance in a biological context. We implement our method on transformer-based protein language models using the family of ESM-2 models. Specifically, we demonstrate that minimum spanning trees, derived from attention matrices, encode structurally significant information about proteins. In our experiments, we combine these topological features with standard embeddings from ESM-2. Our method outperforms traditional approaches and other transformer-based methods with a similar number of parameters in several binding site identification tasks and achieves state-of-the-art performance in conservation prediction tasks. Our results highlight the potential of this hybrid approach in advancing the understanding and prediction of protein functions.", "keywords": [ "protein language models", "protein property prediction", "topological data analysis", "attention maps", "transformers" ], "rating_list": [ 5, 5, 3, 3, 6 ], "soundness_list": [ 2, 2, 1, 2, 3 ], "presentation_list": [ 2, 2, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "DWWwGlPMFr", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Aparna Balagopalan", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Haoran Zhang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Hyewon Jeong", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Jiacheng Zhu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Marzyeh Ghassemi", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Nassim Oufattole", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yan Wu", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Undergrad student" } ], "n_formula": 25, "n_formula_1": 17, "n_ref_uni": 71, "n_ref": 211, "n_ref_all": 267, "n_fig": 4, "n_tab": 20, "L_tab": 16382, "n_element_tab": 1598, "n_fig_1": 3, "n_tab_1": 6, "L_tab_1": 4176, "n_element_tab_1": 218, "formula_len_all": 3787, "formula_len_all_1": 1560, "len_all": 275273, "len_all_1": 78652, "len_abs": 1181, "len_title": 103, "len_sents": 66559, "len_sents_1": 32856, "n_sents": 502, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1186, "title": "LEMoN: Label Error Detection using Multimodal Neighbors", "abs": "Large repositories of image-caption pairs are essential for the development of vision-language models. However, these datasets are often extracted from noisy data scraped from the web, and contain many mislabeled instances. In order to improve the reliability of downstream models, it is important to identify and filter images with incorrect captions. However, beyond filtering based on image-caption embedding similarity, no prior works have proposed other methods to filter noisy multimodal data, or concretely assessed the impact of noisy captioning data on downstream training. In this work, we propose, theoretically justify, and empirically validate LEMoN, a method to automatically identify label errors in image-caption datasets. Our method leverages the multimodal neighborhood of image-caption pairs in the latent space of contrastively pretrained multimodal models to automatically identify label errors. Through empirical evaluations across eight datasets and ten baselines, we find that LEMoN outperforms the baselines by over 3% in label error detection, and that training on datasets filtered using our method improves downstream captioning performance by 2 BLEU points.", "keywords": [ "label error detection", "noisy labels", "image captions" ], "rating_list": [ 3, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "DWLlTNhig1", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Barrett Martin Lattimer", "gender": "Male", "institution": "ASAPP", "country": "", "position": "Researcher" }, { "name": "Ryan T. McDonald", "gender": "Male", "institution": "ASAPP", "country": "", "position": "Researcher" }, { "name": "Varun Prashant Gangal", "gender": "Male", "institution": "ASAPP Inc.", "country": "US", "position": "Researcher" }, { "name": "Yi Yang", "gender": "Male", "institution": "ASAPP Inc", "country": "", "position": "Director of NLP" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 57, "n_ref_all": 80, "n_fig": 4, "n_tab": 6, "L_tab": 934, "n_element_tab": 53, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1792, "n_element_tab_1": 87, "formula_len_all": 52, "formula_len_all_1": 0, "len_all": 169402, "len_all_1": 65466, "len_abs": 1251, "len_title": 93, "len_sents": 53853, "len_sents_1": 31122, "n_sents": 412, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1195, "title": "Sparse Rewards Can Self-Train Dialogue Agents", "abs": "Recent advancements in state-of-the-art (SOTA) Large Language Model (LLM) agents, especially in multi-turn dialogue tasks, have been primarily driven by supervised fine-tuning and high-quality human feedback. However, as base LLM models continue to improve, acquiring meaningful human feedback has become increasingly challenging and costly. In certain domains, base LLM agents may eventually exceed human capabilities, making traditional feedback-driven methods impractical. In this paper, we introduce a novel self-improvement paradigm that empowers LLM agents to autonomously enhance their performance without external human feedback. Our method, Juxtaposed Outcomes for Simulation Harvesting (JOSH), is a self-alignment algorithm that leverages a sparse reward simulation environment to extract ideal behaviors and further train the LLM on its own outputs. We present ToolWOZ, a sparse reward tool-calling simulation environment derived from MultiWOZ. We demonstrate that models trained with JOSH, both small and frontier, significantly improve tool-based interactions while preserving general model capabilities across diverse benchmarks. Our code and data are publicly available on GitHub.", "keywords": [ "Self-training", "LLM", "Simulation", "Benchmark", "Tool-calling", "Dataset", "Task Oriented Dialogue", "Dialogue", "User Simulation", "Beam Search", "Algorithm", "Preference Tuning", "Supervised Finetuning", "JOSH", "ToolWOZ" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "DWISGL63PC", "primary_area": "applications to robotics, autonomy, planning", "n_author": 10, "author_info_list": [ { "name": "Junming Wang", "gender": "Male", "institution": "X Square Robot", "country": "CN", "position": "Researcher" }, { "name": "Qian Zhang", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Wei Yin", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Xiaoyang Guo", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Xingyu Zhang", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" }, { "name": "Zebin Xing", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Ziying Song", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "songen gu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "XIAOXIAO LONG", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Hu Yang", "gender": "Male", "institution": "Horizon Robotics", "country": "CN", "position": "Researcher" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 27, "n_ref": 48, "n_ref_all": 71, "n_fig": 10, "n_tab": 4, "L_tab": 1835, "n_element_tab": 160, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 382, "n_element_tab_1": 74, "formula_len_all": 746, "formula_len_all_1": 724, "len_all": 119256, "len_all_1": 48563, "len_abs": 362, "len_title": 115, "len_sents": 32011, "len_sents_1": 23197, "n_sents": 196, "n_sents_1": 136, "venue": "ICLR2025", "decision": "withdraw", "L_title": 67, "L_abs": 1331, "title": "HE-Drive: Human-Like End-to-End Driving with Vision Language Models", "abs": "In this paper, we propose HE-Drive: the first human-like-centric end-to-end autonomous driving system to generate trajectories that are both temporally consistent and comfortable. Recent studies have shown that imitation learning-based planners and learning-based trajectory scorers can effectively generate and select accuracy trajectories that closely mimic expert demonstrations. However, such trajectory planners and scorers face the dilemma of generating temporally inconsistent and uncomfortable trajectories. To solve the above problems, Our HE-Drive first extracts key 3D spatial representations through sparse perception, which then serves as conditional inputs for a Conditional Denoising Diffusion Probabilistic Models (DDPMs)-based motion planner to generate temporal consistency multi-modal trajectories. A Vision-Language Models (VLMs)-guided trajectory scorer subsequently selects the most comfortable trajectory from these candidates to control the vehicle, ensuring human-like end-to-end driving. Experiments show that HE-Drive not only achieves state-of-the-art performance (i.e., reduces the average collision rate by 71% than VAD) and efficiency (i.e., 1.9X faster than SparseDrive) on the challenging nuScenes and OpenScene datasets but also provides the most comfortable driving experience on real-world data.", "keywords": [ "autonomous driving", "motion planning", "trajectory generation", "diffusion model" ], "rating_list": [ 5, 3, 3, 6, 3 ], "soundness_list": [ 2, 1, 1, 3, 2 ], "presentation_list": [ 3, 2, 2, 3, 1 ], "contribution_list": [ 2, 1, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4, 4 ] }, { "paper_id": "DWI1xx2sX5", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Changsheng Xu", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Hantao Yao", "gender": "Male", "institution": ",Institute of automation, Chinese academy of science", "country": "CN", "position": "Associate Professor" }, { "name": "Jifei Luo", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Xiao Cui", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 88, "n_formula_1": 20, "n_ref_uni": 49, "n_ref": 146, "n_ref_all": 163, "n_fig": 1, "n_tab": 18, "L_tab": 6228, "n_element_tab": 831, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 2389, "n_element_tab_1": 426, "formula_len_all": 6370, "formula_len_all_1": 1103, "len_all": 217128, "len_all_1": 71654, "len_abs": 1575, "len_title": 112, "len_sents": 57871, "len_sents_1": 30648, "n_sents": 415, "n_sents_1": 194, "venue": "ICLR2025", "decision": "withdraw", "L_title": 64, "L_abs": 1581, "title": "Neighbor-aware Geodesic Transportation for Neighborhood Refinery", "abs": "Neighborhood refinery aims to enhance the neighbor relationships by refining the original distance matrix to ensure pairwise consistency.\nTraditional context-based methods, which encode instances alongside their local neighbors in a contextual affinity space, are limited in capturing global relationships and are vulnerable to the negative impacts of outliers in the neighborhood. To overcome these limitations, we propose a novel Neighbor-aware Geodesic Transportation (NGT) for the neighborhood refinery. NGT first constructs a global-aware distribution for each instance, capturing the intrinsic manifold relationships among all instances. This is followed by an optimization transportation process that utilizes the global-aware distribution within the underlying manifold, incorporating global geometric spatial information to generate a refined distance. NGT first involves Manifold-aware Neighbor Encoding (MNE) to project each instance into a global-aware distribution by constraining pairwise similarity with the corresponding affinity graph to capture global relationships. Subsequently, a Regularized Barycenter Refinery (RBR) module is proposed to integrate local neighbors into a barycenter, employing a Wasserstein term to reduce the influence of outliers. Lastly, Geodesic Transportation (GT) leverages geometric and global context information to transport the barycenter distribution along the geodesic paths within the affinity graph. Extensive evaluations on several tasks, such as re-ranking and deep clustering, demonstrate the superiority of our proposed NGT.", "keywords": [ "image retrieval", "reranking", "deep clustering", "self-supervised learning" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 5, 3, 3 ] }, { "paper_id": "DVmn8GyjeD", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Andy Zhou", "gender": "Male", "institution": "Department of Computer Science, University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "James Zou", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Kevin Wu", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Shuang Yang", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Research Scientist" }, { "name": "Yi Zeng", "gender": "Male", "institution": "Virginia Tech", "country": "", "position": "PhD student" }, { "name": "Yu Yang", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Oluwasanmi O Koyejo", "gender": "Male", "institution": "Virtue AI", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 35, "n_ref": 94, "n_ref_all": 108, "n_fig": 10, "n_tab": 5, "L_tab": 5781, "n_element_tab": 243, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 3176, "n_element_tab_1": 195, "formula_len_all": 67, "formula_len_all_1": 67, "len_all": 202855, "len_all_1": 66213, "len_abs": 1325, "len_title": 124, "len_sents": 57596, "len_sents_1": 29534, "n_sents": 411, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1328, "title": "AutoRedTeamer: An Autonomous Red Teaming Agent Against Language Models", "abs": "As large language models (LLMs) become increasingly capable, robust and scalable security evaluation is crucial. While current red teaming approaches have made strides in assessing LLM vulnerabilities, they often rely heavily on human input and fail to provide comprehensive coverage of potential risks. This paper introduces AutoRedTeamer, a unified framework for fully automated, end-to-end red teaming against LLMs. AutoRedTeamer is an LLM-based agent architecture comprising five specialized modules and a novel memory-based attack selection mechanism, enabling deliberate exploration of new attack vectors. AutoRedTeamer supports both seed prompt and risk category inputs, demonstrating flexibility across red teaming scenarios. We demonstrate AutoRedTeamer\u2019s superior performance in identifying potential vulnerabilities compared to existing manual and optimization-based approaches, achieving higher attack success rates by 20% on HarmBench against Llama-3.1-70B while reducing computational costs by 46%. Notably, AutoRedTeamer can break jailbreaking defenses and generate test cases with comparable diversity to human-curated benchmarks. AutoRedTeamer establishes the state of the art for automating the entire red teaming pipeline, a critical step towards comprehensive and scalable security evaluations of AI systems.", "keywords": [ "adversarial robustness", "large language models", "jailbreaking", "ai agents" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "DVlPp7Jd7P", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Claire Boyer", "gender": "unknown", "institution": "Universit\u00e9 Paris-Saclay", "country": "FR", "position": "Full Professor" }, { "name": "G\u00e9rard Biau", "gender": "Male", "institution": "Sorbonne University", "country": "", "position": "Full Professor" }, { "name": "Pierre Marion", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "", "position": "Postdoc" }, { "name": "Rapha\u00ebl Berthier", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "", "position": "Postdoc" } ], "n_formula": 185, "n_formula_1": 31, "n_ref_uni": 55, "n_ref": 72, "n_ref_all": 129, "n_fig": 19, "n_tab": 3, "L_tab": 703, "n_element_tab": 87, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 22350, "formula_len_all_1": 2343, "len_all": 297086, "len_all_1": 74466, "len_abs": 919, "len_title": 106, "len_sents": 73090, "len_sents_1": 29660, "n_sents": 758, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 58, "L_abs": 923, "title": "Attention layers provably solve single-location regression", "abs": "Attention-based models, such as Transformer, excel across various tasks but lack a comprehensive theoretical understanding, especially regarding token-wise sparsity and internal linear representations. To address this gap, we introduce the single-location regression task, where only one token in a sequence determines the output, and its position is a latent random variable, retrievable via a linear projection of the input. To solve this task, we propose a dedicated predictor, which turns out to be a simplified version of a non-linear self-attention layer. We study its theoretical properties, by showing its asymptotic Bayes optimality and analyzing its training dynamics. In particular, despite the non-convex nature of the problem, the predictor effectively learns the underlying structure. This work highlights the capacity of attention mechanisms to handle sparse token information and internal linear structures.", "keywords": [ "theory of neural networks", "attention", "Transformer", "statistical learning theory", "optimization", "first-oder optimization", "gradient flow" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "DUsqifwwf5", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Mingbao Lin", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" }, { "name": "Rongrong Ji", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "Full Professor" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Wenhao Li", "gender": "unknown", "institution": "Xiamen University", "country": "CN", "position": "MS student" }, { "name": "Yunshan Zhong", "gender": "Male", "institution": "Xiamen University", "country": "CN", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 23, "n_ref_uni": 18, "n_ref": 62, "n_ref_all": 95, "n_fig": 8, "n_tab": 6, "L_tab": 1659, "n_element_tab": 206, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 735, "n_element_tab_1": 58, "formula_len_all": 1156, "formula_len_all_1": 1159, "len_all": 119813, "len_all_1": 58943, "len_abs": 1246, "len_title": 130, "len_sents": 32685, "len_sents_1": 25717, "n_sents": 283, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1251, "title": "SOLOS: Sparse Optimization For Long Sequence In Context Compression Enhanced LLMs", "abs": "Recent advances in long-context large language models (LLMs) make them commercially viable, but their standard attention mechanisms' quadratic complexity hinders deployment due to excessive computational costs. To address this, researchers have explored Q-former-like architectures that compress input sequences for LLMs, reducing inference costs. However, these methods often underperform compared to mainstream LLMs trained on short sequences and struggle with longer context. We introduce SOLOS, an innovative method for training long sequences within limited computational resources. This approach effectively narrows the performance gap between context-compressed LLMs and mainstream LLMs handling long contexts. By significantly reducing training overhead, SOLOS enables training on long-sequence datasets, such as 100K tokens for instruction tuning, using merely an 8x RTX3090 machine. Our comprehensive experimental analysis confirms SOLOS not only significantly outperforms other context-compression-augmented LLMs but also matches the performance of state-of-the-art long-context models. The introduction of SOLOS marks a significant step toward deploying long-context LLMs, offering both efficiency and effectiveness in practical scenarios.", "keywords": [ "Long-Context LLMs; Context Compression; Sparse Optimization" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "DUfwD5yiN4", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Hamid Kalantari", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Postdoc" }, { "name": "Mohammad Hossein Roohi", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "PhD student" }, { "name": "Pouria Ramazi", "gender": "Male", "institution": "Brock University", "country": "", "position": "Assistant Professor" } ], "n_formula": 27, "n_formula_1": 16, "n_ref_uni": 25, "n_ref": 43, "n_ref_all": 50, "n_fig": 3, "n_tab": 4, "L_tab": 837, "n_element_tab": 141, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 448, "n_element_tab_1": 102, "formula_len_all": 2844, "formula_len_all_1": 1282, "len_all": 126091, "len_all_1": 70206, "len_abs": 1126, "len_title": 106, "len_sents": 43613, "len_sents_1": 30669, "n_sents": 371, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Reject", "L_title": 58, "L_abs": 1147, "title": "Exact Distributed Structure-Learning for Bayesian Networks", "abs": "Learning the structure of a Bayesian network is currently practical for only a limited number of variables. Existing distributed learning approaches approximate the true structure. We present an exact distributed structure-learning algorithm to find a P-map for a set of random variables. First, by using conditional independence, the variables are divided into sets $\\X_1,\\ldots,\\X_I$ such that for each $\\X_i$, the presence and absence of edges that are adjacent with any interior node (a node that is not in any other $\\X_j, j\\neq i$) can be correctly identified by learning the structure of $\\X_i$ separately without using the information of the variables other than $\\X_i$. Second, constraint or score-based structure learners are employed to learn the P-map of $\\X_i$, in a decentralized way. Finally, the separately learned structures are appended by checking a conditional independence test on the boundary nodes (those that are in at least two $\\X_i$'s). The result is proven to be a P-map. This approach allows for a significant reduction in computation time and opens the door for structure learning for a ``giant'' number of variables.", "keywords": [ "Bayesian networks", "Causality", "Structure learning", "Distributed learning" ], "rating_list": [ 5, 3, 8, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 2, 4 ] }, { "paper_id": "DTqx3iqjkz", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Chulhee Yun", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Assistant Professor" }, { "name": "Hanseul Cho", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Hyunji Jung", "gender": "unknown", "institution": "POSTECH", "country": "", "position": "Undergrad student" } ], "n_formula": 202, "n_formula_1": 21, "n_ref_uni": 31, "n_ref": 95, "n_ref_all": 138, "n_fig": 11, "n_tab": 1, "L_tab": 166, "n_element_tab": 12, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 33703, "formula_len_all_1": 1701, "len_all": 289373, "len_all_1": 70375, "len_abs": 1104, "len_title": 132, "len_sents": 84724, "len_sents_1": 31229, "n_sents": 973, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1117, "title": "Convergence and Implicit Bias of Gradient Descent on Continual Linear Classification", "abs": "We study continual learning on multiple linear classification tasks by sequentially running gradient descent (GD) for a fixed budget of iterations per each given task. When all tasks are jointly linearly separable and are presented in a cyclic/random order, we show the directional convergence of the trained linear classifier to the joint (offline) max-margin solution. This is surprising because GD training on a single task is implicitly biased towards the individual max-margin solution for the task, and the direction of the joint max-margin solution can be largely different from these individual solutions. Additionally, when tasks are given in a cyclic order, we present a non-asymptotic analysis on cycle-averaged forgetting, revealing that (1) alignment between tasks is indeed closely tied to catastrophic forgetting and backward knowledge transfer and (2) the amount of forgetting vanishes to zero as the cycle repeats. Lastly, we analyze the case where the tasks are no longer jointly separable and show that the model trained in a cyclic order converges to the unique minimum of the joint loss function.", "keywords": [ "Continual Learning", "Sequential Learning", "Gradient Descent", "Linear Classification", "Convergence", "Implicit Bias" ], "rating_list": [ 8, 6, 6, 5 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 3 ] }, { "paper_id": "DTjmv5QJBx", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Allan Anzagira", "gender": "Male", "institution": "North Carolina A&T State University", "country": "", "position": "PhD student" }, { "name": "Freddy Lecue", "gender": "unknown", "institution": "INRIA", "country": "", "position": "Full Professor" }, { "name": "Manuela Veloso", "gender": "Female", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Full Professor" }, { "name": "Tom Bewley", "gender": "unknown", "institution": "J.P. Morgan Chase", "country": "GB", "position": "Researcher" }, { "name": "Eoin Kenny", "gender": "Male", "institution": "J.P. Morgan Chase", "country": "", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 37, "n_ref": 67, "n_ref_all": 91, "n_fig": 7, "n_tab": 4, "L_tab": 529, "n_element_tab": 104, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 147, "n_element_tab_1": 47, "formula_len_all": 360, "formula_len_all_1": 368, "len_all": 165920, "len_all_1": 65124, "len_abs": 1595, "len_title": 125, "len_sents": 60170, "len_sents_1": 32050, "n_sents": 458, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1601, "title": "Language Models Can Help to Learn High-Performing Cost Functions for Recourse", "abs": "Algorithmic recourse is a specialised variant of counterfactual explanation, concerned with offering actionable recommendations to individuals who have received adverse outcomes from automated systems. Most recourse algorithms assume access to a cost function, which quantifies the effort involved in following recommendations. Such functions are useful for filtering down recourse options to those which are most actionable. In this study, we explore the use of large language models (LLMs) to help label data for training recourse cost functions, while preserving important factors such as transparency, fairness, and performance. We find that LLMs do generally align with human judgements of cost and can label data for the training of effective cost functions, moreover they can be fine-tuned with simple prompt engineering to maximise performance and improve current recourse algorithms in practice. Previously, recourse cost definitions have mainly relied on heuristics and missed the complexities of feature dependencies and fairness attributes, which has drastically limited their usefulness. Our results show that it is possible to train a high-performing, interpretable cost function by consulting an LLM via careful prompt engineering. Furthermore, these cost functions can be customised to add or remove biases as befitting the domain and problem.\nOverall, this study suggests a simple, accessible method for accurately quantifying notions of cost, effort, or distance between data points that correlate with human intuition, with possible applications throughout the explainable AI field.", "keywords": [ "algorithmic recourse", "large language models", "cost functions", "interpretable ml", "user study" ], "rating_list": [ 6, 3, 8, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 4, 2, 4, 3 ], "contribution_list": [ 2, 2, 4, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "DTatjJTDl1", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Evangelos Theodorou", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Lingkai Kong", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Molei Tao", "gender": "unknown", "institution": "Georgia Institute of Technology", "country": "US", "position": "Associate Professor" }, { "name": "Tianrong Chen", "gender": "Male", "institution": "Apple", "country": "US", "position": "Researcher" }, { "name": "Yuchen Zhu", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 80, "n_formula_1": 21, "n_ref_uni": 40, "n_ref": 94, "n_ref_all": 118, "n_fig": 10, "n_tab": 5, "L_tab": 1055, "n_element_tab": 137, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 343, "n_element_tab_1": 39, "formula_len_all": 7828, "formula_len_all_1": 2208, "len_all": 199342, "len_all_1": 67582, "len_abs": 1322, "len_title": 124, "len_sents": 68108, "len_sents_1": 32182, "n_sents": 520, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1388, "title": "Trivialized Momentum Facilitates Diffusion Generative Modeling on Lie Groups", "abs": "The generative modeling of data on manifolds is an important task, for which diffusion models in flat spaces typically need nontrivial adaptations. This article demonstrates how a technique called `trivialization' can transfer the effectiveness of diffusion models in Euclidean spaces to Lie groups. In particular, an auxiliary momentum variable was algorithmically introduced to help transport the position variable between data distribution and a fixed, easy-to-sample distribution. Normally, this would incur further difficulty for manifold data because momentum lives in a space that changes with the position. However, our trivialization technique creates a new momentum variable that stays in a simple fixed vector space. This design, together with a manifold preserving integrator, simplifies implementation and avoids inaccuracies created by approximations such as projections to tangent space and manifold, which were typically used in prior work, hence facilitating generation with high-fidelity and efficiency. The resulting method achieves state-of-the-art performance on protein and RNA torsion angle generation and sophisticated torus datasets. We also, arguably for the first time, tackle the generation of data on high-dimensional Special Orthogonal and Unitary groups, the latter essential for quantum problems. Code is available at https://github.com/yuchen-zhu-zyc/TDM.", "keywords": [ "non-Euclidean generative modeling", "denoising diffusion", "Lie group" ], "rating_list": [ 8, 6, 8, 5 ], "soundness_list": [ 4, 2, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 4, 2, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "DTQlvDCAql", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Gueter Josmy Faure", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "PhD student" }, { "name": "Hung-Ting Su", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Postdoc" }, { "name": "Jia-Fong Yeh", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "PhD student" }, { "name": "Min-hung Chen", "gender": "Male", "institution": "NVIDIA", "country": "TW", "position": "Research Scientist" }, { "name": "Shang-Hong Lai", "gender": "Male", "institution": "National Tsing Hua University", "country": "TW", "position": "Full Professor" }, { "name": "Winston Hsu", "gender": "Male", "institution": "National Taiwan University", "country": "TW", "position": "Professor" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 36, "n_ref": 65, "n_ref_all": 89, "n_fig": 12, "n_tab": 10, "L_tab": 2951, "n_element_tab": 351, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 2563, "n_element_tab_1": 306, "formula_len_all": 466, "formula_len_all_1": 459, "len_all": 157214, "len_all_1": 60770, "len_abs": 1992, "len_title": 126, "len_sents": 47776, "len_sents_1": 25187, "n_sents": 367, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1355, "title": "HERMES: temporal-coHERent long-forM understanding with Episodes and Semantics", "abs": "Existing research often treats long-form videos as extended short videos, leading to several limitations: inadequate capture of long-range dependencies, inefficient processing of redundant information, and failure to extract high-level semantic concepts. To address these issues, we propose a novel approach that more accurately reflects human cognition. This paper introduces HERMES: temporal-coHERent long-forM understanding with Episodes and Semantics, a model that simulates episodic memory accumulation to capture action sequences and reinforces them with semantic knowledge dispersed throughout the video. Our work makes two key contributions: First, we develop an Episodic COmpressor (ECO) that efficiently aggregates crucial representations from micro to semi-macro levels, overcoming the challenge of long-range dependencies. Second, we propose a Semantics ReTRiever (SeTR) that enhances these aggregated representations with semantic information by focusing on the broader context, dramatically reducing feature dimensionality while preserving relevant macro-level information. This addresses the issues of redundancy and lack of high-level concept extraction. Extensive experiments demonstrate that HERMES achieves state-of-the-art performance across multiple long-video understanding benchmarks in both zero-shot and fully-supervised settings.", "keywords": [ "Long-form video understanding", "episodic memory", "semantics extraction" ], "rating_list": [ 3, 8, 3, 3 ], "soundness_list": [ 2, 4, 2, 2 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "DSyHRkpI7v", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Calarina Muslimani", "gender": "unknown", "institution": "University of Alberta", "country": "CA", "position": "PhD student" }, { "name": "Matthew Taylor", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "Full Professor" } ], "n_formula": 4, "n_formula_1": 5, "n_ref_uni": 50, "n_ref": 82, "n_ref_all": 123, "n_fig": 15, "n_tab": 12, "L_tab": 6075, "n_element_tab": 475, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 299, "formula_len_all_1": 325, "len_all": 201252, "len_all_1": 59432, "len_abs": 1314, "len_title": 120, "len_sents": 52939, "len_sents_1": 29370, "n_sents": 403, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 72, "L_abs": 1334, "title": "Leveraging Sub-Optimal Data for Human-in-the-Loop Reinforcement Learning", "abs": "To create useful reinforcement learning (RL) agents, step zero is to design a suitable reward function that captures the nuances of the task. However, reward engineering can be a difficult and time-consuming process. \nInstead, human-in-the-loop (HitL) RL methods hold the promise of learning reward functions from human feedback. Despite recent successes, many of the HitL RL methods still require numerous human interactions to learn successful reward functions.\nTo improve the feedback efficiency of HitL RL methods (i.e., require less human interaction), this paper introduces Sub-optimal Data Pre-training, SDP, an approach that leverages reward-free, sub-optimal data to improve scalar- and preference-based HitL RL algorithms. In SDP, we start by pseudo-labeling all low-quality data with the minimum environment reward. Through this process, we obtain reward labels \nto pre-train our reward model \\emph{without} requiring human labeling or preferences. \nThis pre-training phase provides the reward model a head start in learning, enabling it to recognize that low-quality transitions should be assigned low rewards.\nExtensive experiments with both simulated and human teachers reveal that SDP can at least meet, but often significantly improve, state-of-the-art HitL RL performance across a variety of simulated robotic tasks.", "keywords": [ "Reinforcement Learning", "Human-in-the-loop", "Preference learning", "Learning from scalar feedback" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 4, 3, 4, 4 ], "contribution_list": [ 3, 3, 2, 1 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "DSsSPr0RZJ", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Dong Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Distinguished Scientist" }, { "name": "Hongming Zhang", "gender": "Male", "institution": "Tencent AI Lab Seattle", "country": "US", "position": "Researcher" }, { "name": "Kaixin Ma", "gender": "unknown", "institution": "Tencent AI Lab", "country": "US", "position": "Researcher" }, { "name": "Liqiang Jing", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "PhD student" }, { "name": "Wenhao Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Researcher" }, { "name": "Wenlin Yao", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Xinya Du", "gender": "Male", "institution": "University of Texas at Dallas", "country": "US", "position": "Assistant Professor" }, { "name": "Zhehui Huang", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "PhD student" }, { "name": "Xiaoyang Wang", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Senior Researcher" } ], "n_formula": 2, "n_formula_1": 3, "n_ref_uni": 33, "n_ref": 79, "n_ref_all": 122, "n_fig": 9, "n_tab": 10, "L_tab": 5930, "n_element_tab": 456, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 418, "n_element_tab_1": 79, "formula_len_all": 68, "formula_len_all_1": 882, "len_all": 260732, "len_all_1": 60804, "len_abs": 1334, "len_title": 124, "len_sents": 73768, "len_sents_1": 27109, "n_sents": 658, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1341, "title": "DSBench: How Far Are Data Science Agents from Becoming Data Science Experts?", "abs": "Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) have demonstrated impressive language/vision reasoning abilities, igniting the recent trend of building agents for targeted applications such as shopping assistants or AI software engineers. Recently, many data science benchmarks have been proposed to investigate their performance in the data science domain. However, existing data science benchmarks still fall short when compared to real-world data science applications due to their simplified settings. To bridge this gap, we introduce DSBench, a comprehensive benchmark designed to evaluate data science agents with realistic tasks. This benchmark includes 466 data analysis tasks and 74 data modeling tasks, sourced from Eloquence and Kaggle competitions. DSBench offers a realistic setting by encompassing long contexts, multimodal task backgrounds, reasoning with large data files and multi-table structures, and performing end-to-end data modeling tasks. Our evaluation of state-of-the-art LLMs, LVLMs, and agents shows that they struggle with most tasks, with the best agent solving only 34.12% of data analysis tasks and achieving a 34.74% Relative Performance Gap (RPG). These findings underscore the need for further advancements in developing more practical, intelligent, and autonomous data science agents.", "keywords": [ "data science", "agent", "benchmark", "llm" ], "rating_list": [ 6, 8, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 3, 4, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 2, 3 ] }, { "paper_id": "DSpq7CXMFP", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Daniel LK Yamins", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "David Wendt", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Jiajun Wu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Seungwoo Kim", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Stefan Stojanov", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Rahul M. V.", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 36, "n_ref": 69, "n_ref_all": 90, "n_fig": 10, "n_tab": 3, "L_tab": 1333, "n_element_tab": 175, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1009, "n_element_tab_1": 139, "formula_len_all": 188, "formula_len_all_1": 190, "len_all": 117792, "len_all_1": 65345, "len_abs": 1216, "len_title": 123, "len_sents": 33048, "len_sents_1": 29605, "n_sents": 213, "n_sents_1": 190, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1222, "title": "Learning Counterfactual Interventions for Self-Supervised Motion Estimation", "abs": "A major challenge in self-supervised learning from visual inputs is extracting information from the learned representations to an explicit and usable form. This is most commonly done by learning readout layers with supervision or using highly specialized heuristics. This is challenging primarily because the self-supervised pretext tasks and the downstream tasks that extract information are not tightly connected in a principled manner---improving the former does not guarantee improvements in the latter. The recently proposed counterfactual world modeling paradigm aims to address this challenge through a masked next frame predictor base model which enables simple counterfactual extraction procedures for extracting optical flow, segments and depth. In this work, we take the next step and parameterize and optimize the counterfactual extraction of optical flow by solving the same simple next frame prediction task as the base model. Our approach achieves state of the art performance for estimation motion on real-world videos while requiring no labeled data. This work sets the foundation for future methods on improving the extraction of more complex visual structures like segments and depth with high accuracy.", "keywords": [ "self-supervised learning", "motion estimation", "world modeling", "counterfactual prompting", "visual prompting" ], "rating_list": [ 3, 5, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 2, 1, 3, 3, 3 ], "confidence_list": [ 2, 3, 3, 4, 3 ] }, { "paper_id": "DSl9sSuUhp", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Felix den Breejen", "gender": "Male", "institution": "KAIST Graduate School of AI", "country": "", "position": "PhD student" }, { "name": "Se-Young Yun", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 0, "n_ref_uni": 24, "n_ref": 48, "n_ref_all": 69, "n_fig": 8, "n_tab": 8, "L_tab": 4924, "n_element_tab": 279, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 615, "n_element_tab_1": 64, "formula_len_all": 85, "formula_len_all_1": 0, "len_all": 99088, "len_all_1": 49194, "len_abs": 551, "len_title": 118, "len_sents": 29939, "len_sents_1": 22976, "n_sents": 243, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 551, "title": "Attic: A New Architecture for Tabular In-Context Learning Transformers", "abs": "Tabular In-Context Learning (ICL) transformers, such as TabPFN and TabForestPFN, have shown strong performance on tabular classification tasks. In this paper, we introduce Attic, a new architecture for ICL-transformers. Unlike TabPFN and TabForestPFN, where one token represents all features of one observation, Attic assigns one token to each feature of every observation. This simple architectural change results in a significant performance boost. As a result, we can confidently say that neural networks outperform tree-based methods like XGBoost.", "keywords": [ "tabular classification", "tabular in-context learning transformers", "architecture" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "DShqJA1Z64", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Francesco Insulla", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "MS student" }, { "name": "Lorenzo Rosasco", "gender": "Male", "institution": "Universit\u00e0 degli Studi di Genova", "country": "IT", "position": "Full Professor" }, { "name": "Shuo Huang", "gender": "Female", "institution": "Universit\u00e0 degli Studi di Genova, Istituto Italiano di Tecnologia", "country": "IT", "position": "Postdoc" } ], "n_formula": 59, "n_formula_1": 42, "n_ref_uni": 31, "n_ref": 51, "n_ref_all": 59, "n_fig": 1, "n_tab": 2, "L_tab": 66, "n_element_tab": 22, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5462, "formula_len_all_1": 3868, "len_all": 105218, "len_all_1": 60319, "len_abs": 864, "len_title": 0, "len_sents": 29026, "len_sents_1": 23698, "n_sents": 246, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 54, "L_abs": 869, "title": "Towards a learning theory of representation alignment", "abs": "It has recently been argued that AI models' representations are becoming aligned as their scale and performance increase. Empirical analyses have been designed to support this idea and conjecture the possible alignment of different representations toward a shared statistical model of reality. In this paper, we propose a learning-theoretic perspective to representation alignment. First, we review and connect different notions of alignment based on metric, probabilistic, and spectral ideas. Then, we focus on stitching, a particular approach to understanding the interplay between different representations in the context of a task. Our main contribution here is relating properties of stitching to the kernel alignment of the underlying representation. Our results can be seen as a first step toward casting representation alignment as a learning-theoretic problem.", "keywords": [ "learning theory", "representation learning", "model stitching", "representation alignment" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 2, 3, 2, 2 ] }, { "paper_id": "DSGDdj0HEM", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 6, "author_info_list": [ { "name": "Fanman Meng", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Hongliang Li", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Linfeng Xu", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zihuan Qiu", "gender": "Not Specified", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Qingbo Wu", "gender": "Male", "institution": "University of Electronic Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "YI XU", "gender": "Male", "institution": "Dalian University of Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 33, "n_formula_1": 21, "n_ref_uni": 37, "n_ref": 81, "n_ref_all": 93, "n_fig": 5, "n_tab": 10, "L_tab": 1657, "n_element_tab": 278, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2047, "formula_len_all_1": 1195, "len_all": 124164, "len_all_1": 61289, "len_abs": 2814, "len_title": 110, "len_sents": 33182, "len_sents_1": 26087, "n_sents": 270, "n_sents_1": 195, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1112, "title": "Increment Vector Transformation for Class Incremental Learning", "abs": "Class Incremental Learning (CIL) presents a major challenge due to the phenomenon of catastrophic forgetting.\nRecent studies on Linear Mode Connectivity (LMC) reveal that Naive-SGD oracle, trained with all historical data, connects to previous task minima through low-loss linear paths---a property generally absent in current CIL methods.\nIn this paper, we explore whether LMC holds for the CIL oracle. Our empirical results confirm the presence of LMC in the CIL oracle, showing that models can retain performance on earlier tasks by following the discovered low-loss linear paths. Motivated by this finding, we propose Increment Vector Transformation (IVT), which leverages the diagonal of the Fisher Information Matrix to approximate Hessian-based transformation, uncovering low-loss linear paths for incremental updates. \nOur method is orthogonal to existing CIL approaches, serving as a plug-in with minor extra computational costs.\nExtensive experiments on CIFAR-100, ImageNet-Subset, and ImageNet-Full demonstrate significant performance improvements when integrating IVT with representative CIL methods.", "keywords": [ "Class incremental learning" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "DRiLWb8bJg", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Eliot Xing", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Jean Oh", "gender": "Female", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Vernon Luk", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "Undergrad student" } ], "n_formula": 41, "n_formula_1": 28, "n_ref_uni": 114, "n_ref": 231, "n_ref_all": 249, "n_fig": 6, "n_tab": 12, "L_tab": 2482, "n_element_tab": 356, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 198, "n_element_tab_1": 36, "formula_len_all": 3102, "formula_len_all_1": 1601, "len_all": 247304, "len_all_1": 66305, "len_abs": 1992, "len_title": 124, "len_sents": 51687, "len_sents_1": 30881, "n_sents": 399, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 76, "L_abs": 1306, "title": "Stabilizing Reinforcement Learning in Differentiable Multiphysics Simulation", "abs": "Recent advances in GPU-based parallel simulation have enabled practitioners to collect large amounts of data and train complex control policies using deep reinforcement learning (RL), on commodity GPUs. However, such successes for RL in robotics have been limited to tasks sufficiently simulated by fast rigid-body dynamics. Simulation techniques for soft bodies are comparatively several orders of magnitude slower, thereby limiting the use of RL due to sample complexity requirements. To address this challenge, this paper presents both a novel RL algorithm and a simulation platform to enable scaling RL on tasks involving rigid bodies and deformables. We introduce Soft Analytic Policy Optimization (SAPO), a maximum entropy first-order model-based actor-critic RL algorithm, which uses first-order analytic gradients from differentiable simulation to train a stochastic actor to maximize expected return and entropy. Alongside our approach, we develop Rewarped, a parallel differentiable multiphysics simulation platform that supports simulating various materials beyond rigid bodies. We re-implement challenging manipulation and locomotion tasks in Rewarped, and show that SAPO outperforms baselines over a range of tasks that involve interaction between rigid bodies, articulations, and deformables.", "keywords": [ "reinforcement learning", "differentiable simulation" ], "rating_list": [ 8, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "DRhKnUYNm9", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Bingqing Liu", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "", "position": "MS student" }, { "name": "Jun Wang", "gender": "unknown", "institution": "iWudao", "country": "", "position": "Consultant" }, { "name": "Ruixuan Li", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" }, { "name": "Wei Liu", "gender": "unknown", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Postdoc" } ], "n_formula": 16, "n_formula_1": 15, "n_ref_uni": 36, "n_ref": 78, "n_ref_all": 112, "n_fig": 6, "n_tab": 10, "L_tab": 6045, "n_element_tab": 462, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 515, "n_element_tab_1": 109, "formula_len_all": 1170, "formula_len_all_1": 1349, "len_all": 130178, "len_all_1": 67852, "len_abs": 1427, "len_title": 119, "len_sents": 39138, "len_sents_1": 31573, "n_sents": 264, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1430, "title": "A Decoupled Learning Framework for Neural Marked Temporal Point Process", "abs": "The standard neural marked temporal point process employs the EmbeddingEncoder-History vector-Decoder (EEHD) architecture, wherein the history vector encapsulates the cumulative effects of past events. However, due to the inherent imbalance in event categories in real-world scenarios, the history vector tends to favor more frequent events, inadvertently overlooking less common yet potentially significant ones, thereby compromising the model\u2019s overall performance. To tackle this issue, we introduce a novel decoupled learning framework for neural marked temporal point process, where each event type is modeled independently to capture its unique characteristics, allowing for a more nuanced and equitable treatment of all event types. Each event type boasts its own complete EEHD architecture, featuring scaled-down parameters due to the decoupling of temporal dynamics. This decoupled design enables asynchronous parallel training, and the embeddings can reflect the dependencies between event types. Our versatile framework, accommodating various encoder and decoder architectures, demonstrates state-of-the-art performance across diverse datasets, outperforming benchmarks by a significant margin and increasing training speed by up to 12 times. Additionally, it offers interpretability, revealing which event types have similar influences on a particular event type, fostering a deeper understanding of temporal dynamics.", "keywords": [ "temporal point process", "interpretability", "event sequence modeling" ], "rating_list": [ 5, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "DRf8RpofIN", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Konstantin Yakovlev", "gender": "Male", "institution": "Higher School of Economics", "country": "RU", "position": "Intern" } ], "n_formula": 19, "n_formula_1": 16, "n_ref_uni": 33, "n_ref": 63, "n_ref_all": 73, "n_fig": 2, "n_tab": 4, "L_tab": 1058, "n_element_tab": 55, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 477, "n_element_tab_1": 60, "formula_len_all": 1472, "formula_len_all_1": 1089, "len_all": 98539, "len_all_1": 69768, "len_abs": 938, "len_title": 110, "len_sents": 24985, "len_sents_1": 23038, "n_sents": 248, "n_sents_1": 228, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 944, "title": "Generalized Greedy Gradient-Based Hyperparameter Optimization", "abs": "Bilevel Optimization (BLO) is a widely-used approach that has numerous applications, including hyperparameter optimization, meta-learning. However, existing gradient-based method suffer from the following issues. Reverse-mode differentiation suffers from high memory requirements, while the methods based on the implicit function theorem require the convergence of the inner optimization. Approximations that consider a truncated inner optimization trajectory suffer from a short horizon bias. In this paper, we propose a novel approximation for hypergradient computation that sidesteps these difficulties. Specifically, we accumulate the short-horizon approximations from each step of the inner optimization trajectory. Additionally, we demonstrate that under certain conditions, the proposed hypergradient is a sufficient descent direction. Experimental results on a few-shot meta-learning and data hyper-cleaning tasks support our findings.", "keywords": [ "bilevel optimization", "meta-learning", "hyperparameter optimization" ], "rating_list": [ 5, 5, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "DRSSLefryd", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Sambhav Khurana", "gender": "Male", "institution": "Texas A&M University - College Station", "country": "US", "position": "MS student" }, { "name": "Shuiwang Ji", "gender": "Male", "institution": "Texas A&M University", "country": "US", "position": "Professor" }, { "name": "Shurui Gui", "gender": "Male", "institution": "Texas A&M University", "country": "US", "position": "PhD student" }, { "name": "Xiner Li", "gender": "Female", "institution": "Texas A&M University - College Station", "country": "US", "position": "PhD student" } ], "n_formula": 30, "n_formula_1": 13, "n_ref_uni": 73, "n_ref": 128, "n_ref_all": 155, "n_fig": 5, "n_tab": 16, "L_tab": 3729, "n_element_tab": 464, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 191, "n_element_tab_1": 24, "formula_len_all": 25386, "formula_len_all_1": 1017, "len_all": 301225, "len_all_1": 79402, "len_abs": 1247, "len_title": 118, "len_sents": 79369, "len_sents_1": 36073, "n_sents": 585, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1251, "title": "A Hierarchical Language Model Design For Interpretable Graph Reasoning", "abs": "Large language models (LLMs) have seen an increased adoption for tasks with implicit graphical structures, such as planning in robotics, multi-hop question answering, and knowledge probing. However, despite their remarkable success in text-based tasks, LLMs' capabilities in understanding explicit graph structures remain limited, preventing them from fully replacing Graph Neural Networks (GNNs) in graph-centric applications. In this work, we introduce a Hierarchical Language Model (HLM-G) Design that employs a two-block architecture to effectively capture local and global graph information, significantly enhancing graph structure understanding. Our model achieves a new state-of-the-art in graph understanding, outperforming both GNN and LLM baselines. It demonstrates robustness to variations in graph-descriptive prompts, overcoming a key limitation of existing LLMs. Furthermore, we demonstrate the interpretability of our model using intrinsic attention weights and established explainers. Comprehensive evaluations across diverse real-world datasets, covering node, link, and graph-level tasks, highlight our model's superior generalization capabilities, marking a significant advancement in the application of LLMs to graph-centric tasks.", "keywords": [ "Language models", "Interpretability", "Graph reasoning" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 2, 4, 3 ], "presentation_list": [ 3, 3, 4, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "DRKkO2Tejc", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Aleksandr Beznosikov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "Associate Professor" }, { "name": "Denis Kuznedelev", "gender": "Male", "institution": "Yandex", "country": "RU", "position": "Researcher" }, { "name": "Marat Mansurov", "gender": "Male", "institution": "Higher School of Economics, Higher School of Economics", "country": "", "position": "Undergrad student" }, { "name": "Max Ryabinin", "gender": "Not Specified", "institution": "Together AI", "country": "", "position": "Researcher" }, { "name": "Philip Zmushko", "gender": "unknown", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "MS student" }, { "name": "Ruslan A. Svirschevski", "gender": "Male", "institution": "Yandex", "country": "RU", "position": "ML Resident" } ], "n_formula": 17, "n_formula_1": 13, "n_ref_uni": 37, "n_ref": 61, "n_ref_all": 86, "n_fig": 2, "n_tab": 9, "L_tab": 2664, "n_element_tab": 242, "n_fig_1": 2, "n_tab_1": 5, "L_tab_1": 803, "n_element_tab_1": 31, "formula_len_all": 1005, "formula_len_all_1": 822, "len_all": 153064, "len_all_1": 67712, "len_abs": 1226, "len_title": 131, "len_sents": 43048, "len_sents_1": 30102, "n_sents": 374, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1240, "title": "Label Privacy in Split Learning for Large Models with Parameter-Efficient Training", "abs": "As deep learning models become larger and more expensive, many practitioners turn to fine-tuning APIs. \nThese web services allow fine-tuning a model between two parties: the client that provides the data, and the server that hosts the model.\nWhile convenient, these APIs raise a new concern: the data of the client is at risk of privacy breach during the training procedure.\nThis challenge presents an important practical case of vertical federated learning, where the two parties perform parameter-efficient fine-tuning (PEFT) of a large model.\nIn this study, we systematically search for a way to fine-tune models over an API *while keeping the labels private*.\nWe analyze the privacy of LoRA, a popular approach for parameter-efficient fine-tuning when training over an API.\nUsing this analysis, we propose P$^3$EFT, a multi-party split learning algorithm that takes advantage of existing PEFT properties to maintain privacy at a lower performance overhead.\nTo validate our algorithm, we fine-tune DeBERTa-v2-XXLarge, Flan-T5 Large and LLaMA-2 7B using LoRA adapters on a range of NLP tasks. We find that P$^3$EFT is competitive with existing privacy-preserving methods in multi-party and two-party setups while having higher accuracy.", "keywords": [ "Split Learning", "Vertical Federated Learning", "Federated Learning", "Parameter Efficient Fine-tuning", "Privacy", "Large Language Models" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "DQfHkEcUqV", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Aleem Khan", "gender": "Male", "institution": "Department of Computer Science, Whiting School of Engineering", "country": "", "position": "PhD student" }, { "name": "Andrew Wang", "gender": "Male", "institution": "Department of Computer Science, Whiting School of Engineering", "country": "US", "position": "PhD student" }, { "name": "Sophia Hager", "gender": "Female", "institution": "Johns Hopkins University", "country": "", "position": "PhD student" }, { "name": "Nicholas Andrews", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "Research Scientist" } ], "n_formula": 5, "n_formula_1": 1, "n_ref_uni": 58, "n_ref": 104, "n_ref_all": 129, "n_fig": 1, "n_tab": 14, "L_tab": 3658, "n_element_tab": 357, "n_fig_1": 1, "n_tab_1": 6, "L_tab_1": 1544, "n_element_tab_1": 155, "formula_len_all": 263, "formula_len_all_1": 59, "len_all": 251002, "len_all_1": 72595, "len_abs": 2085, "len_title": 67, "len_sents": 61312, "len_sents_1": 34882, "n_sents": 438, "n_sents_1": 236, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1553, "title": "Learning Extrapolative Sequence Transformations from Markov Chains", "abs": "Most successful applications of deep learning involve similar training and test conditions. However, for some generative tasks, samples should improve desirable properties beyond previously known values, which requires the ability to generate novel hypotheses that extrapolate beyond training data. While large language models have been successfully extended to a variety of sequence modeling problems, greedy autoregressive sampling can struggle to explore the solution space sufficiently to extrapolate, especially when the properties of interest are global to the sequence. On the other hand, sequence-level sampling methods such as Markov chain Monte Carlo (MCMC) offer theoretical guarantees about capturing the distribution of interest, but suffer from the curse of dimensionality in discrete structured spaces. We propose a new approach that bridges the gap between MCMC and autoregressive sampling, which may be viewed as off-policy reinforcement learning. Our approach uses selected states from Markov chains as a source of training data for an autoregressive inference network, which is then able to generate novel sequences at test time that extrapolate along the sequence-level properties of interest. The proposed approach is validated on three problems: protein sequence design, text sentiment control, and text anonymization. We find that the learned inference network confers many of the same (and sometimes better) generalization benefits compared to the slow sampling process, but with the additional benefit of high sample efficiency.", "keywords": [ "Large language model", "Markov chain Monte Carlo" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "DPzQ5n3mNm", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Abdolmehdi Behroozi", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "PhD student" }, { "name": "Chaopeng Shen", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Full Professor" }, { "name": "Daniel Kifer", "gender": "Male", "institution": "Pennsylvania State University", "country": "US", "position": "Professor" } ], "n_formula": 22, "n_formula_1": 11, "n_ref_uni": 28, "n_ref": 41, "n_ref_all": 73, "n_fig": 10, "n_tab": 11, "L_tab": 5196, "n_element_tab": 620, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 525, "n_element_tab_1": 105, "formula_len_all": 1027, "formula_len_all_1": 494, "len_all": 140521, "len_all_1": 65732, "len_abs": 2219, "len_title": 167, "len_sents": 41981, "len_sents_1": 31593, "n_sents": 285, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 118, "L_abs": 2342, "title": "Sensitivity-Constrained Fourier Neural Operators for Forward and Inverse Problems in Parametric Differential Equations", "abs": "Parametric differential equations of the form $\\partial{\\textbf{u}}/\\partial{t} = f(\\textbf{u}, \\textbf{x}, t, \\mathbf{p})$, where $\\mathbf{p}$ represents physical system parameters, are fundamental across scientific and engineering disciplines. Recent advances in surrogate modeling, particularly deep learning frameworks like the Fourier Neural Operator (FNO), have demonstrated significant efficiency in approximating differential equation solution paths $\\textbf{u}$. However, these approximations result in inaccurate solutions to inverse problems (using neural operators in an optimization routine for estimating physical system parameters), provide inaccurate estimates of \\emph{sensitivities} (i.e., $\\partial{\\textbf{u}}/\\partial{\\textbf{p}}$: the dependence of the solution path on the physical parameters) needed for scenario analysis and optimization, and are highly sensitive to concept drift. These issues are all related and can be addressed by using a novel \\emph{sensitivity loss} regularizer that we propose in this paper. This regularizer works with a wide range of neural operators, but for concreteness, we focus on applying it to the popular FNO framework, resulting in \\emph{Sensitivity-Constrained Fourier Neural Operators} (SC-FNO). SC-FNO ensures accuracy in the solution paths, inverse problems, and sensitivity calculations, even under sparse training data or concept drift scenarios. Our approach maintains high accuracy for the solution paths $\\mathbf{u}$ and significantly outperforms both the original FNO and FNO combined with Physics-Informed Neural Network (PINN) regularization on the remaining tasks.\n\nNotably, in parameter inversion tasks from solution paths, SC-FNO exhibits markedly superior accuracy, even when FNO breaks down, underscoring the critical role of sensitivity awareness. For cases with more parameters (82 tested), SC-FNO can reduce training data demand, elevate the performance ceiling of neural operators, and even reduce training time. These conclusions are robust for various differential equations, neural operators, and different ways of supervising sensitivities for training.\nThese improvements, without large computational or memory costs, enhance the reliability and applicability of neural operators in complex physical systems modeling and engineering analysis.\\end{abstract}", "keywords": [ "Fourier Neural Operator", "Sensitivity Analysis", "Parametric Differential Equations", "Surrogate Modes", "Differentiable Numerical Solvers", "Inverse Problems" ], "rating_list": [ 6, 6, 5, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "DPynq6bSHn", "primary_area": "datasets and benchmarks", "n_author": 11, "author_info_list": [ { "name": "Barry Haddow", "gender": "unknown", "institution": "University of Edinburgh", "country": "", "position": "Postdoc" }, { "name": "Dayy\u00e1n O'Brien", "gender": "unknown", "institution": "University of Edinburgh, University of Edinburgh", "country": "GB", "position": "Researcher" }, { "name": "Hengyu Luo", "gender": "unknown", "institution": "University of Helsinki", "country": "FI", "position": "PhD student" }, { "name": "Hinrich Sch\u00fctze", "gender": "Male", "institution": "Center for Information and Language Processing", "country": "DE", "position": "Full Professor" }, { "name": "Jaakko Paavola", "gender": "Male", "institution": "University of Helsinki", "country": "FI", "position": "MS student" }, { "name": "J\u00f6rg Tiedemann", "gender": "Male", "institution": "University of Helsinki", "country": "", "position": "Full Professor" }, { "name": "Peiqin Lin", "gender": "Male", "institution": "Institut f\u00fcr Informatik", "country": "", "position": "PhD student" }, { "name": "Pinzhen Chen", "gender": "Not Specified", "institution": "University of Edinburgh", "country": "GB", "position": "Postdoc" }, { "name": "Shaoxiong Ji", "gender": "unknown", "institution": "University of Helsinki", "country": "", "position": "Postdoc" }, { "name": "Zihao Li", "gender": "Male", "institution": "University of Helsinki", "country": "FI", "position": "MS student" }, { "name": "Indraneil Arun Paul", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 77, "n_ref": 120, "n_ref_all": 160, "n_fig": 4, "n_tab": 25, "L_tab": 43700, "n_element_tab": 3111, "n_fig_1": 0, "n_tab_1": 5, "L_tab_1": 4630, "n_element_tab_1": 424, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 404092, "len_all_1": 72524, "len_abs": 5829, "len_title": 127, "len_sents": 61790, "len_sents_1": 30592, "n_sents": 438, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 78, "L_abs": 1002, "title": "EMMA-500: Enhancing Massively Multilingual Adaptation of Large Language Models", "abs": "In this work, we introduce EMMA-500, a large-scale multilingual language model continue-trained on texts across 546 languages designed for enhanced multilingual performance, with a focus on improving language coverage for low-resource languages. To facilitate continual pre-training, we compile the MaLA corpus, a comprehensive multilingual dataset and enrich it with curated datasets across diverse domains. Leveraging this corpus, we conduct extensive continual pre-training of the Llama 2 7B model, resulting in EMMA-500, which demonstrates robust performance across a wide collection of benchmarks, including a comprehensive set of multilingual tasks and PolyWrite, an open-ended generation benchmark developed in this study. Our results highlight the effectiveness of continual pre-training in expanding large language models\u2019 language capacity, particularly for underrepresented languages, demonstrating significant gains in cross-lingual transfer, task generalization, and language adaptability.", "keywords": [ "multilingual adaptation", "large language model." ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "DPp5GSohht", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Rajesh Ranganath", "gender": "unknown", "institution": "New York University", "country": "", "position": "Assistant Professor" }, { "name": "Wanqian Yang", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 29, "n_ref": 62, "n_ref_all": 86, "n_fig": 4, "n_tab": 3, "L_tab": 191, "n_element_tab": 23, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 510, "n_element_tab_1": 65, "formula_len_all": 834, "formula_len_all_1": 341, "len_all": 106215, "len_all_1": 57419, "len_abs": 994, "len_title": 136, "len_sents": 36024, "len_sents_1": 27992, "n_sents": 262, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1012, "title": "Unclipping CLIP's Wings: Avoiding Robustness Pitfalls in Multimodal Image Classification", "abs": "Despite being pretrained on large-scale data, multimodal models such as CLIP can still learn spurious correlations. However, CLIP does not seem to learn the same spurious correlations as standard vision models, performing worse on some benchmark datasets (Waterbirds) yet better on others (CelebA). We investigate this discrepancy and find that CLIP's robustness on these datasets is highly sensitive to the choice of class prompts. Worst-group accuracy can be arbitrarily improved or worsened by making minute, single-word changes to prompts. We further provide evidence that the root cause of this phenomenon is \\textit{coverage} --- using class prompts that are out-of-distribution with respect to pretraining can worsen spurious correlations. Motivated by these findings, we propose using class prompts that are generated from a public image-to-text model, such as BLIP. We show that performing $k$-nearest neighbors on these prompt embeddings improve downstream robustness without needing to fine-tune CLIP.", "keywords": [ "robustness", "CLIP", "spurious correlations", "contrastive learning", "multimodality" ], "rating_list": [ 8, 3, 3, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 1 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 3, 4, 3, 5 ] }, { "paper_id": "DPlUWG4WMw", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Antoine Jacquier", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "Full Professor" }, { "name": "Lukas Gonon", "gender": "unknown", "institution": "Universit\u00e4t St. Gallen", "country": "CH", "position": "Assistant Professor" }, { "name": "Ruben Wiedemann", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" } ], "n_formula": 53, "n_formula_1": 20, "n_ref_uni": 30, "n_ref": 69, "n_ref_all": 121, "n_fig": 14, "n_tab": 10, "L_tab": 3595, "n_element_tab": 245, "n_fig_1": 8, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3519, "formula_len_all_1": 1300, "len_all": 174599, "len_all_1": 68063, "len_abs": 1703, "len_title": 95, "len_sents": 64512, "len_sents_1": 33035, "n_sents": 477, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Accept (conditional poster)", "L_title": 46, "L_abs": 1713, "title": "Operator Deep Smoothing for Implied Volatility", "abs": "We devise a novel method for nowcasting implied volatility based on neural operators.\nBetter known as implied volatility smoothing in the financial industry, nowcasting of implied volatility means constructing a smooth surface that is consistent with the prices presently observed on a given option market.\nOption price data arises highly dynamically in ever-changing spatial configurations, which poses a major limitation to foundational machine learning approaches using classical neural networks.\nWhile large models in language and image processing deliver breakthrough results on vast corpora of raw data, in financial engineering the generalization from big historical datasets has been hindered by the need for considerable data pre-processing.\nIn particular, implied volatility smoothing has remained an instance-by-instance, hands-on process both for neural network-based and traditional parametric strategies.\nOur general *operator deep smoothing* approach, instead, directly maps observed data to smoothed surfaces.\nWe adapt the graph neural operator architecture to do so with high accuracy on ten years of raw intraday S&P 500 options data, using a single model instance.\nThe trained operator adheres to critical no-arbitrage constraints and is robust with respect to subsampling of inputs (occurring in practice in the context of outlier removal).\nWe provide extensive historical benchmarks and showcase the generalization capability of our approach in a comparison with classical neural networks and SVI, an industry standard parametrization for implied volatility. \nThe operator deep smoothing approach thus opens up the use of neural networks on large historical datasets in financial engineering.", "keywords": [ "Financial Engineering", "Neural Operators", "Option Pricing", "Function Interpolation", "Nowcasting" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 2, 2, 4 ] }, { "paper_id": "DPYPpC0cBC", "primary_area": "applications to robotics, autonomy, planning", "n_author": 2, "author_info_list": [ { "name": "Guoqiang Hu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Site Qu", "gender": "Male", "institution": "nanyang technological university", "country": "SG", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 24, "n_ref": 51, "n_ref_all": 85, "n_fig": 4, "n_tab": 5, "L_tab": 4617, "n_element_tab": 392, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 2390, "n_element_tab_1": 264, "formula_len_all": 1559, "formula_len_all_1": 986, "len_all": 149958, "len_all_1": 73732, "len_abs": 1835, "len_title": 170, "len_sents": 57694, "len_sents_1": 34740, "n_sents": 321, "n_sents_1": 190, "venue": "ICLR2025", "decision": "Reject", "L_title": 121, "L_abs": 1840, "title": "Beyond Predefined Depots: A Dual-Mode Generative DRL Framework for Proactive Depot Generation in Location-Routing Problem", "abs": "The Location-Routing Problem (LRP), which combines the challenges of facility (depot) locating and vehicle route planning, is critically constrained by the reliance on predefined depot candidates, limiting the solution space and potentially leading to suboptimal outcomes. Previous research on LRP without predefined depots is scant and predominantly relies on heuristic algorithms that iteratively attempt depot placements across a planar area. Such approaches lack the ability to proactively generate depot locations that meet specific geographic requirements, revealing a notable gap in current research landscape. To bridge this gap, we propose a data-driven generative DRL framework, designed to proactively generate depots for LRP without predefined depot candidates, solely based on customer requests data which include geographic and demand information. It can operate in two distinct modes: direct generation of exact depot locations, and the creation of a multivariate Gaussian distribution for flexible depots sampling. By extracting depots' geographic pattern from customer requests data, our approach can dynamically respond to logistical needs, identifying high-quality depot locations that further reduce total routing costs compared to traditional methods. Extensive experiments demonstrate that, for a same group of customer requests, compared with those depots identified through random attempts, our framework can proactively generate depots that lead to superior solution routes with lower routing cost. The implications of our framework potentially extend into real-world applications, particularly in emergency medical rescue and disaster relief logistics, where rapid establishment and adjustment of depot locations are paramount, showcasing its potential in addressing LRP for dynamic and unpredictable environments.", "keywords": [ "Generative DRL", "Depot Generation", "Routing planning" ], "rating_list": [ 3, 5, 8, 1 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 3, 2, 1 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "DP3BwwTKbL", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Si Wu", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Xiangyuan Peng", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "PhD student" }, { "name": "Xingsi Dong", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" } ], "n_formula": 28, "n_formula_1": 18, "n_ref_uni": 38, "n_ref": 50, "n_ref_all": 81, "n_fig": 17, "n_tab": 6, "L_tab": 1480, "n_element_tab": 72, "n_fig_1": 15, "n_tab_1": 3, "L_tab_1": 223, "n_element_tab_1": 24, "formula_len_all": 2155, "formula_len_all_1": 1745, "len_all": 124339, "len_all_1": 61813, "len_abs": 1268, "len_title": 133, "len_sents": 33055, "len_sents_1": 27358, "n_sents": 286, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 1275, "title": "Predicting Observation after Action in a Hierarchical Energy-based Model with Memory", "abs": "Understanding the mechanisms of brain function is greatly advanced by predictive models. Recent advancements in machine learning further underscore the potency of prediction for learning optimal representation. However, there remains a gap in creating a biologically plausible model that explains how the neural system achieves prediction. In this paper, we introduce a framework employing an energy-based model (EBM) to capture the nuanced processes of predicting observation after action within the neural system, encompassing prediction, learning, and inference. We implement the EBM with a hierarchical structure and integrate a continuous attractor neural network for memory, constructing a biologically plausible model. In experimental evaluations, our model demonstrates efficacy across diverse scenarios. The range of actions includes eye movement, motion in environments, head turning, and static observation while the environment changes. Our model not only makes accurate predictions for environments it was trained on, but also provides reasonable predictions for unseen environments, matching the performances of machine learning methods in multiple tasks. We hope that this study contributes to a deep understanding of how the neural system performs prediction.", "keywords": [ "Neuroscience", "Prediction", "Energy-based models", "Sampling-based inference", "Local learning rules", "Attractor neural networks" ], "rating_list": [ 3, 3, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 2, 2 ] }, { "paper_id": "DOXnqYLCcd", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Min-hwan Oh", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" }, { "name": "Junghun Kim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Postdoc" } ], "n_formula": 209, "n_formula_1": 47, "n_ref_uni": 32, "n_ref": 111, "n_ref_all": 144, "n_fig": 2, "n_tab": 7, "L_tab": 2343, "n_element_tab": 285, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 323, "n_element_tab_1": 28, "formula_len_all": 43815, "formula_len_all_1": 5710, "len_all": 219085, "len_all_1": 59094, "len_abs": 971, "len_title": 117, "len_sents": 51329, "len_sents_1": 21113, "n_sents": 524, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1015, "title": "Dynamic Multi-product Selection and Pricing under Preference Feedback", "abs": "In this study, we investigate the problem of dynamic multi-product selection and pricing by introducing a novel framework based on a *censored multinomial logit* (C-MNL) choice model. In this model, sellers present a set of products with prices, and buyers filter out products priced above their valuation, purchasing at most one product from the remaining options based on their preferences. The goal is to maximize seller revenue by dynamically adjusting product offerings and prices, while learning both product valuations and buyer preferences through purchase feedback. To achieve this, we propose a Lower Confidence Bound (LCB) pricing strategy. By combining this pricing strategy with either an Upper Confidence Bound (UCB) or Thompson Sampling (TS) product selection approach, our algorithms achieve regret bounds of $\\tilde{O}(d^{\\frac{3}{2}}\\sqrt{T})$ and $\\tilde{O}(d^{2}\\sqrt{T})$, respectively. Finally, we validate the performance of our methods through simulations, demonstrating their effectiveness.", "keywords": [ "Dynamic pricing", "Preference Feedback", "Bandits" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 4, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "DOHsYZrrny", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Hyowon Wi", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KP", "position": "PhD student" }, { "name": "Noseong Park", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" } ], "n_formula": 25, "n_formula_1": 14, "n_ref_uni": 41, "n_ref": 65, "n_ref_all": 91, "n_fig": 7, "n_tab": 10, "L_tab": 2350, "n_element_tab": 230, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1821, "n_element_tab_1": 130, "formula_len_all": 1901, "formula_len_all_1": 1182, "len_all": 153275, "len_all_1": 65633, "len_abs": 1550, "len_title": 100, "len_sents": 49045, "len_sents_1": 29651, "n_sents": 422, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1580, "title": "HiLoRA: High-frequency-augmented Low-Rank Adaptation", "abs": "As large language models (LLMs) have demonstrated remarkable performance, parameter-efficient fine-tuning (PEFT) has emerged as an important paradigm. As a solution, low-rank adaptation (LoRA) freezes the pre-trained weights and introduces small learnable adapters instead of fine-tuning the full set of parameters. However, LoRA suffers from $\\textit{catastrophic forgetting}$, where pre-trained knowledge is overwhlemed and forgotten as new information is learned. One cause of this issue is $\\textit{implicit regularization}$, where deep learning models tend to favor more generalized solutions. This tendency leads to a significant increase in the largest singular values of the weights, which correspond to low-frequency components. To address this problem, we propose an advanced LoRA that balances the retention of pre-trained knowledge with the learning of new information. Since fine-tuning involves learning fine-grained details, which correspond to high-frequency information, we designed HiLoRA, a method that injects learnable high-frequency components into the pre-trained model. By leveraging the parameterized SVD and constraining singular values to appropriate levels, HiLoRA adapts to new tasks by focusing on the high-frequency domain with minimal change from the pre-trained weights. To evaluate the effectiveness of HiLoRA, we conduct extensive experiments on natural language understanding and question answering tasks. The results show that HiLoRA not only improves performance but also effectively retains pre-trained knowledge compared to baseline models.", "keywords": [ "Large Language Models", "LoRA", "Frequency", "Catastrophic Forgetting" ], "rating_list": [ 6, 5, 3, 6 ], "soundness_list": [ 3, 2, 1, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 1, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "DOA1WSPZSi", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Bryan Hooi", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" }, { "name": "Yuan Sui", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 44, "n_ref": 78, "n_ref_all": 103, "n_fig": 8, "n_tab": 5, "L_tab": 6221, "n_element_tab": 239, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 95, "n_element_tab_1": 11, "formula_len_all": 178, "formula_len_all_1": 112, "len_all": 184829, "len_all_1": 59490, "len_abs": 1499, "len_title": 167, "len_sents": 51988, "len_sents_1": 28858, "n_sents": 346, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 119, "L_abs": 1505, "title": "Can Knowledge Graphs Make Large Language Models More Trustworthy? An Empirical Study over Open-ended Question Answering", "abs": "Recent works integrating Knowledge Graphs (KGs) have led to promising improvements in enhancing reasoning accuracy of Large Language Models (LLMs).\nHowever, current benchmarks mainly focus on closed tasks, leaving a gap in the assessment of more complex, real-world scenarios. This gap has also obscured the evaluation of KGs' potential to mitigate the problem of hallucination in LLMs.\nTo fill the gap, we introduce OKGQA, a new benchmark specifically designed to assess LLMs enhanced with KGs under open-ended, real-world question answering scenarios.\nOKGQA is designed to closely reflect the complexities of practical applications using questions from different types, and incorporates specific metrics to measure both the reduction in hallucinations and the enhancement in reasoning capabilities.\nTo consider the scenario in which KGs may have varying levels of mistakes, we further propose another experiment setting OKGQA-P to assess model performance when the semantics and structure of KGs are deliberately perturbed and contaminated.\nOKGQA aims to (1) explore whether KGs can make LLMs more trustworthy in an open-ended setting, and (2) conduct a comparative analysis to shed light on methods and future directions for leveraging KGs to reduce LLMs' hallucination.\nWe believe that this study can facilitate a more complete performance comparison and encourage continuous improvement in integrating KGs with LLMs. The code of this paper is released at https://anonymous.4open.science/r/OKGQA-CBB0.", "keywords": [ "Large Language Models", "Hallucination", "Open-ended Question Answering" ], "rating_list": [ 5, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "DNjHslZrqu", "primary_area": "learning on time series and dynamical systems", "n_author": 7, "author_info_list": [ { "name": "Doron L Bergman", "gender": "Not Specified", "institution": "Capital One", "country": "", "position": "Researcher" }, { "name": "John P Dickerson", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Associate Professor" }, { "name": "Micah Goldblum", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Assistant Professor" }, { "name": "Senthil Kumar", "gender": "Male", "institution": "Capital One", "country": "", "position": "Researcher" }, { "name": "Tom Goldstein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Alexander Stein", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Samuel B Sharpe", "gender": "unknown", "institution": "Capital One", "country": "", "position": "Researcher" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 26, "n_ref": 49, "n_ref_all": 54, "n_fig": 3, "n_tab": 3, "L_tab": 1294, "n_element_tab": 174, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 301, "n_element_tab_1": 28, "formula_len_all": 338, "formula_len_all_1": 338, "len_all": 110887, "len_all_1": 57015, "len_abs": 1080, "len_title": 81, "len_sents": 38106, "len_sents_1": 27260, "n_sents": 286, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1083, "title": "A Simple Baseline for Predicting Future Events with Auto-Regressive Tabular Transformers", "abs": "Many real-world applications of tabular data involve using historic events to predict properties of new ones, for example whether a credit card transaction is fraudulent or what rating a customer will assign a product on a retail platform.\nExisting approaches to event prediction include costly, brittle, and application-dependent techniques such as time-aware positional embeddings, learned row and field encodings, and oversampling methods for addressing class imbalance.\nMoreover, these approaches often assume specific use-cases, for example that we know the labels of all historic events or that we only predict a pre-specified label and not the data\u2019s features themselves.\nIn this work, we propose a simple but flexible baseline using standard autoregressive LLM-style transformers with elementary positional embeddings and a causal language modeling objective.\nOur baseline outperforms existing approaches across popular datasets and can be employed for various use-cases.\nWe demonstrate that the same model can predict labels, impute missing values, or model event sequences.", "keywords": [ "Event Prediction", "Tabular Data", "Transformers" ], "rating_list": [ 3, 3, 5 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 2, 2, 2 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "DNBwlQYA90", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Byeonghyun Ko", "gender": "Male", "institution": "Seoul National University", "country": "", "position": "PhD student" }, { "name": "HyunGyu Lee", "gender": "unknown", "institution": "Seoul National University", "country": "", "position": "MS student" }, { "name": "Jaejin Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Jisoo Kim", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Kyusu Ahn", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Sangik Lee", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 45, "n_ref": 147, "n_ref_all": 190, "n_fig": 10, "n_tab": 7, "L_tab": 3213, "n_element_tab": 407, "n_fig_1": 11, "n_tab_1": 4, "L_tab_1": 3532, "n_element_tab_1": 151, "formula_len_all": 362, "formula_len_all_1": 272, "len_all": 195859, "len_all_1": 64050, "len_abs": 2592, "len_title": 109, "len_sents": 64689, "len_sents_1": 26916, "n_sents": 592, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1745, "title": "UDC-VIT: A Real-World Video Dataset for Under-Display Cameras", "abs": "Under Display Camera (UDC) is an advanced imaging system that places a digital camera lens underneath a display panel, effectively concealing the camera. However, the display panel significantly degrades captured images or videos, introducing low transmittance, blur, noise, and flare issues. Tackling such issues is challenging because of the complex degradation of UDCs, including diverse flare patterns. Despite extensive research on UDC images and their restoration models, studies on videos have yet to be significantly explored. While two UDC video datasets exist, they primarily focus on unrealistic or synthetic UDC degradation rather than real-world UDC degradation. In this paper, we propose a real-world UDC video dataset called UDC-VIX. Unlike existing datasets, only UDC-VIX exclusively includes human motions that target facial recognition. We propose a video-capturing system to simultaneously acquire non-degraded and UDC-degraded videos of the same scene. Then, we align a pair of captured videos frame by frame, using discrete Fourier transform (DFT). We compare UDC-VIX with seven representative UDC still image datasets and two existing UDC video datasets. Using six deep-learning models, we compare UDC-VIX and an existing synthetic UDC video dataset. The results indicate the ineffectiveness of models trained on earlier synthetic UDC video datasets, as they do not reflect the actual characteristics of UDC-degraded videos. We also demonstrate the importance of effective UDC restoration by evaluating face recognition accuracy concerning PSNR, SSIM, and LPIPS scores. UDC-VIX enables further exploration in the UDC video restoration and offers better insights into the challenge. UDC-VIX is available at our project site.", "keywords": [ "Under-Display Camera", "Dataset", "Benchmark", "Alignment", "Flare" ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 1, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "DM6Q45HWSk", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Benedikt Alkin", "gender": "Male", "institution": "Emmi AI GmbH", "country": "AT", "position": "Researcher" }, { "name": "Fabian Paischer", "gender": "Male", "institution": ", Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "Postdoc" }, { "name": "Lukas Hauzenberger", "gender": "Male", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "PhD student" }, { "name": "Sepp Hochreiter", "gender": "Male", "institution": "Johannes Kepler University Linz", "country": "AT", "position": "Full Professor" }, { "name": "Thomas Schmied", "gender": "Not Specified", "institution": "Johannes Kepler Universit\u00e4t Linz", "country": "AT", "position": "PhD student" }, { "name": "Marc Deisenroth", "gender": "Male", "institution": "Google", "country": "GB", "position": "Principal Researcher" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 55, "n_ref": 141, "n_ref_all": 207, "n_fig": 13, "n_tab": 19, "L_tab": 12107, "n_element_tab": 1331, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 3852, "n_element_tab_1": 182, "formula_len_all": 297, "formula_len_all_1": 295, "len_all": 359565, "len_all_1": 69033, "len_abs": 1398, "len_title": 130, "len_sents": 69962, "len_sents_1": 28073, "n_sents": 563, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1423, "title": "One Initialization to Rule them All: Fine-tuning via Explained Variance Adaptation", "abs": "Foundation models (FMs) are pre-trained on large-scale datasets and then fine-tuned on a downstream task for a specific application. The most successful and most commonly used fine-tuning method is to update the pre-trained weights via a low-rank adaptation (LoRA). LoRA introduces new weight matrices that are usually initialized at random with a uniform rank distribution across model weights. Recent works focus on *weight-driven* initialization or learning of adaptive ranks during training. Both approaches have only been investigated in isolation, resulting in slow convergence or a uniform rank distribution, in turn leading to sub-optimal performance. We propose to enhance LoRA by initializing the new weights in a *data-driven* manner by computing singular value decomposition (SVD) on minibatches of activation vectors. Then, we initialize the LoRA matrices with the obtained right-singular vectors and re-distribute ranks among all weight matrices to explain the maximal amount of variance across layers. This results in our new method **E**xplained **V**ariance **A**daptation (EVA). We apply EVA to a variety of fine-tuning tasks ranging from language generation and understanding to image classification and reinforcement learning. EVA exhibits faster convergence than competitors and attains the highest average score across a multitude of tasks per domain while reducing the number of trainable parameters.", "keywords": [ "Foundation Models", "LoRA", "Fine-tuning" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "DLhjxxXYwH", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Gabriel A. Silva", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Full Professor" }, { "name": "Johnny Jingze Li", "gender": "unknown", "institution": "University of California, San Diego", "country": "", "position": "PhD student" }, { "name": "Vivek Kurien George", "gender": "unknown", "institution": "Lawrence Livermore National Labs", "country": "US", "position": "Principal Researcher" } ], "n_formula": 30, "n_formula_1": 15, "n_ref_uni": 25, "n_ref": 53, "n_ref_all": 59, "n_fig": 1, "n_tab": 2, "L_tab": 750, "n_element_tab": 46, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 544, "n_element_tab_1": 40, "formula_len_all": 1458, "formula_len_all_1": 820, "len_all": 105176, "len_all_1": 58849, "len_abs": 1198, "len_title": 134, "len_sents": 38049, "len_sents_1": 28538, "n_sents": 275, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1037, "title": "Advancing Neural Network Performance through Emergence-Promoting Initialization Scheme", "abs": "We introduce a novel yet straightforward neural network initialization scheme that modifies conventional methods like Xavier and Kaiming initialization. Inspired by the concept of emergence and leveraging the emergence measures proposed by Li (2023), our method adjusts the layer-wise weight scaling factors to achieve higher emergence values. This enhancement is easy to implement, requiring no additional optimization steps for initialization compared to GradInit. We evaluate our approach across various architectures, including MLP and convolutional architectures for image recognition, and transformers for machine translation. We demonstrate substantial improvements in both model accuracy and training speed, with and without batch normalization. The simplicity, theoretical innovation, and demonstrable empirical advantages of our method make it a potent enhancement to neural network initialization practices. These results suggest a promising direction for leveraging emergence to improve neural network training methodologies.", "keywords": [ "Emergence", "Initialization", "cascade effect" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 2, 1 ], "contribution_list": [ 2, 2, 1 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "DLDuVbxORA", "primary_area": "foundation or frontier models, including LLMs", "n_author": 2, "author_info_list": [ { "name": "Stephen Zhang", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "PhD student" }, { "name": "Vardan Papyan", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "Assistant Professor" } ], "n_formula": 17, "n_formula_1": 10, "n_ref_uni": 56, "n_ref": 107, "n_ref_all": 123, "n_fig": 4, "n_tab": 16, "L_tab": 6810, "n_element_tab": 1042, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 3217, "n_element_tab_1": 292, "formula_len_all": 1094, "formula_len_all_1": 813, "len_all": 237500, "len_all_1": 49660, "len_abs": 1268, "len_title": 117, "len_sents": 32893, "len_sents_1": 17838, "n_sents": 212, "n_sents_1": 112, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1280, "title": "OATS: Outlier-Aware Pruning Through Sparse and Low Rank Decomposition", "abs": "The recent paradigm shift to large-scale foundation models has brought about a new era for deep learning that, while has found great success in practice, has also been plagued by prohibitively expensive costs in terms of high memory consumption and compute. To mitigate these issues, there has been a concerted effort in post-hoc neural network pruning techniques that do not require costly retraining. Despite the considerable progress being made, existing methods often exhibit a steady drop in model performance as the compression increases. In this paper, we present a novel approach to compressing large transformers, coined OATS, that compresses the model weights by approximating each weight matrix as the sum of a sparse matrix and a low-rank matrix. Prior to the decomposition, the weights are first scaled by the second moment of their input embeddings, so as to ensure the preservation of outlier features recently observed in large transformer models. Without retraining, OATS achieves state-of-the-art performance when compressing large language models, such as Llama-3 and Phi-3, and vision transformers, such as Google's ViT and DINOv2, by up to $60\\\\%$, all while speeding up the model's inference on a CPU by up to $1.37\\times$ compared to prior pruning methods.", "keywords": [ "network pruning", "low-rank", "compression", "sparsification", "large language models", "outlier features" ], "rating_list": [ 6, 8, 3, 8 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 2, 4, 2, 4 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 4, 4, 5, 3 ] }, { "paper_id": "DLBlR0rea5", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Binxu Wang", "gender": "Female", "institution": "Harvard University", "country": "US", "position": "Researcher" }, { "name": "Haim Sompolinsky", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Full Professor" }, { "name": "Jiaqi Shang", "gender": "Female", "institution": "Harvard University, Harvard University", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 30, "n_ref": 55, "n_ref_all": 100, "n_fig": 34, "n_tab": 3, "L_tab": 717, "n_element_tab": 23, "n_fig_1": 31, "n_tab_1": 2, "L_tab_1": 1460, "n_element_tab_1": 26, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 122916, "len_all_1": 67535, "len_abs": 1286, "len_title": 0, "len_sents": 43881, "len_sents_1": 32685, "n_sents": 335, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1301, "title": "How do diffusion models learn and generalize on abstract rules for reasoning?", "abs": "Diffusion models excel in generating and completing patterns in images. \nBut how good is their ability to learn hidden rules from samples and to generate and reason according to such rules or even generalize to similar rules? \nWe trained a wide family of unconditional diffusion models on Raven's progression matrix task to precisely study this. We quantified their capability to generate structurally consistent samples and complete missing parts according to hidden rules. \nWe found diffusion models can synthesize novel samples consistent with rules without memorizing the training set, much better than GPT2 trained on the same data. They memorized and recombined local parts of the training samples to create new rule-conforming samples. \nWhen tasked to complete the missing panel with inpainting techniques, advanced sampling techniques were needed to perform well. Further, their pattern completion capability can generalize to rules unseen during training. \nFurther, through generative training on rule data, a robust rule representation rapidly emerged in the diffusion model, which could linearly classify rules at 99.8\\% test accuracy. \nOur results suggest diffusion training is a useful paradigm for reasoning and learning representations for downstream tasks even for abstract rules data.", "keywords": [ "Generative model", "Reasoning", "Raven\u2019s Progressive Matrix", "Diffusion", "Scaling law", "Stochastic interpolant" ], "rating_list": [ 5, 3, 3, 5, 6 ], "soundness_list": [ 2, 2, 3, 3, 2 ], "presentation_list": [ 3, 2, 1, 3, 3 ], "contribution_list": [ 2, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "DL9txImSzm", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Anish Abhijit Diwan", "gender": "Male", "institution": "Delft University of Technology", "country": "NL", "position": "MS student" }, { "name": "Jan Peters", "gender": "Male", "institution": "German Research Center for AI", "country": "DE", "position": "Principal Researcher" }, { "name": "Jens Kober", "gender": "Male", "institution": "Delft University of Technology", "country": "", "position": "Associate Professor" }, { "name": "julen urain", "gender": "Male", "institution": "Technische Universit\u00e4t Darmstadt", "country": "DE", "position": "Postdoc" } ], "n_formula": 13, "n_formula_1": 5, "n_ref_uni": 34, "n_ref": 82, "n_ref_all": 110, "n_fig": 8, "n_tab": 8, "L_tab": 4047, "n_element_tab": 410, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 890, "n_element_tab_1": 72, "formula_len_all": 1451, "formula_len_all_1": 784, "len_all": 184452, "len_all_1": 72498, "len_abs": 1207, "len_title": 166, "len_sents": 63228, "len_sents_1": 31311, "n_sents": 461, "n_sents_1": 203, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 118, "L_abs": 1210, "title": "Noise-conditioned Energy-based Annealed Rewards (NEAR): A Generative Framework for Imitation Learning from Observation", "abs": "This paper introduces a new imitation learning framework based on energy-based generative models capable of learning complex, physics-dependent, robot motion policies through state-only expert motion trajectories. Our algorithm, called Noise-conditioned Energy-based Annealed Rewards (NEAR), constructs several perturbed versions of the expert's motion data distribution and learns smooth, and well-defined representations of the data distribution's energy function using denoising score matching. We propose to use these learnt energy functions as reward functions to learn imitation policies via reinforcement learning. We also present a strategy to gradually switch between the learnt energy functions, ensuring that the learnt rewards are always well-defined in the manifold of policy-generated samples. We evaluate our algorithm on complex humanoid tasks such as locomotion and martial arts and compare it with state-only adversarial imitation learning algorithms like Adversarial Motion Priors (AMP). Our framework sidesteps the optimisation challenges of adversarial imitation learning techniques and produces results comparable to AMP in several quantitative metrics across multiple imitation settings.", "keywords": [ "imitation learning", "energy based generative models", "reinforcement learning", "imitation from observation" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 4, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "DKkQtRMowq", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Ankit Parag Shah", "gender": "Male", "institution": "Accenture", "country": "US", "position": "Principal Researcher" }, { "name": "Chen Qian", "gender": "Male", "institution": "University of California-Santa Cruz", "country": "", "position": "" }, { "name": "Jiaheng Wei", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "CN", "position": "Assistant Professor" }, { "name": "Jinlong Pang", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "", "position": "PhD student" }, { "name": "Yang Liu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "Yaxuan Wang", "gender": "Female", "institution": "University of California, Santa Cruz", "country": "US", "position": "PhD student" }, { "name": "Yujia Bao", "gender": "Male", "institution": "Accenture", "country": "US", "position": "Researcher" }, { "name": "Zhaowei Zhu", "gender": "Male", "institution": "Docta.ai", "country": "US", "position": "Researcher" }, { "name": "wei wei", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" } ], "n_formula": 13, "n_formula_1": 6, "n_ref_uni": 33, "n_ref": 83, "n_ref_all": 144, "n_fig": 14, "n_tab": 19, "L_tab": 11281, "n_element_tab": 1297, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 1294, "n_element_tab_1": 157, "formula_len_all": 1198, "formula_len_all_1": 472, "len_all": 238254, "len_all_1": 59950, "len_abs": 1577, "len_title": 112, "len_sents": 69604, "len_sents_1": 25423, "n_sents": 591, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 64, "L_abs": 1155, "title": "Improving Data Efficiency via Curating LLM-Driven Rating Systems", "abs": "Instruction tuning is critical for adapting large language models (LLMs) to downstream tasks, and recent studies have demonstrated that small amounts of human-curated data can outperform larger datasets, challenging traditional data scaling laws. While LLM-based data quality rating systems offer a cost-effective alternative to human annotation, they often suffer from inaccuracies and biases, even in powerful models like GPT-4. In this work, we introduce $DS^2$, a **D**iversity-aware **S**core curation method for **D**ata **S**election. By systematically modeling error patterns through a score transition matrix, $DS^2$ corrects LLM-based scores and promotes diversity in the selected data samples. Our approach shows that a curated subset (just 3.3\\% of the original dataset) outperforms full-scale datasets (300k samples) across various machine-alignment benchmarks, and matches or surpasses human-aligned datasets such as LIMA with the same sample size (1k samples). These findings challenge conventional data scaling assumptions, highlighting that redundant, low-quality samples can degrade performance and reaffirming that ``more can be less''.", "keywords": [ "Data Selection", "LLM", "Instruction Tuning" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "DKgAFfCs5F", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Heng Yang", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Assistant Professor" }, { "name": "Jeong Joon Park", "gender": "Not Specified", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" }, { "name": "Jiachen Sun", "gender": "Male", "institution": "University of Michigan", "country": "", "position": "PhD student" }, { "name": "Marco Pavone", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Director, Autonomous Vehicle Research" }, { "name": "Minkyoung Cho", "gender": "Female", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Qingzhao Zhang", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Yulong Cao", "gender": "unknown", "institution": "NVIDIA", "country": "", "position": "Researcher" }, { "name": "Zhuoqing Mao", "gender": "Female", "institution": "Google", "country": "US", "position": "consultant" } ], "n_formula": 15, "n_formula_1": 8, "n_ref_uni": 64, "n_ref": 175, "n_ref_all": 213, "n_fig": 14, "n_tab": 9, "L_tab": 3269, "n_element_tab": 288, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1460, "n_element_tab_1": 100, "formula_len_all": 983, "formula_len_all_1": 496, "len_all": 214690, "len_all_1": 74689, "len_abs": 1426, "len_title": 122, "len_sents": 57144, "len_sents_1": 35547, "n_sents": 432, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1244, "title": "Cocoon: Robust Multi-Modal Perception with Uncertainty-Aware Sensor Fusion", "abs": "An important paradigm in 3D object detection is the use of multiple modalities to enhance accuracy in both normal and challenging conditions, particularly for long-tail scenarios. To address this, recent studies have explored two directions of adaptive approaches: MoE-based adaptive fusion, which struggles with uncertainties arising from distinct object configurations, and late fusion for output-level adaptive fusion, which relies on separate detection pipelines and limits comprehensive understanding. In this work, we introduce Cocoon, an object- and feature-level uncertainty-aware fusion framework. The key innovation lies in uncertainty quantification for heterogeneous representations, enabling fair comparison across modalities through the introduction of a feature aligner and a learnable surrogate ground truth, termed feature impression. We also define a training objective to ensure that their relationship provides a valid metric for uncertainty quantification. Cocoon consistently outperforms existing static and adaptive methods in both normal and challenging conditions, including those with natural and artificial corruptions. Furthermore, we show the validity and efficacy of our uncertainty metric across diverse datasets.", "keywords": [ "Multi-modal perception; Sensor fusion; Robustness; Uncertainty quantification" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 4, 3, 2 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "DKZjYuB6gc", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Andrey Zhmoginov", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Researcher" }, { "name": "Anirudh Goyal", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Researcher" }, { "name": "Gus Kristiansen", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Jihwan Lee", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Mark Sandler", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" }, { "name": "Max Vladymyrov", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Scientist" }, { "name": "Nolan Andrew Miller", "gender": "Male", "institution": "University of Southern California", "country": "", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 46, "n_ref": 54, "n_ref_all": 90, "n_fig": 10, "n_tab": 19, "L_tab": 11359, "n_element_tab": 2070, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 308, "n_element_tab_1": 32, "formula_len_all": 396, "formula_len_all_1": 395, "len_all": 168179, "len_all_1": 63478, "len_abs": 1143, "len_title": 61, "len_sents": 35382, "len_sents_1": 27892, "n_sents": 325, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Reject", "L_title": 61, "L_abs": 1148, "title": "Narrowing the Focus: Learned Optimizers for Pretrained Models", "abs": "In modern deep learning, the models are learned by applying gradient updates using an optimizer, which transforms the updates based on various statistics. Optimizers are often hand-designed and tuning their hyperparameters is a big part of the training process. Learned optimizers have shown some initial promise, but are generally unsuccessful as a general optimization mechanism applicable to every problem. In this work we explore a different direction: instead of learning general optimizers, we instead specialize them to a specific training environment. We propose a novel optimizer technique that learns a layer-specific linear combination of update directions provided by a set of base optimizers, effectively adapting its strategy to the specific model and dataset. When evaluated on image classification tasks, this specialized optimizer significantly outperforms both traditional off-the-shelf methods such as Adam, as well as existing general learned optimizers. Moreover, it demonstrates robust generalization with respect to model initialization, evaluating on unseen datasets, and training durations beyond its meta-training horizon.", "keywords": [ "Learned optimizers", "Meta-optimization", "Fine-tuning", "Image classification" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 1, 3, 2 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "DKCtt2iqfw", "primary_area": "learning on time series and dynamical systems", "n_author": 3, "author_info_list": [ { "name": "Bo Chen", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" }, { "name": "Muyao Wang", "gender": "Male", "institution": "Xidian University ", "country": "CN", "position": "MS student" }, { "name": "Zeke Xie", "gender": "Male", "institution": "HKUST(GZ)", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 9, "n_ref_uni": 29, "n_ref": 63, "n_ref_all": 78, "n_fig": 4, "n_tab": 11, "L_tab": 5676, "n_element_tab": 432, "n_fig_1": 3, "n_tab_1": 8, "L_tab_1": 3554, "n_element_tab_1": 267, "formula_len_all": 845, "formula_len_all_1": 443, "len_all": 129003, "len_all_1": 71539, "len_abs": 1553, "len_title": 127, "len_sents": 40112, "len_sents_1": 30370, "n_sents": 277, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1559, "title": "Channel-wise Influence: Estimating Data Influence for Multivariate Time Series", "abs": "The influence function, a robust statistics technique, is an effective post-hoc method that measures the impact of modifying or removing training data on model parameters, offering valuable insights into model interpretability without requiring costly retraining. It would provide extensions like increasing model performance, improving model generalization, and offering interpretability. Recently, Multivariate Time\nSeries (MTS) analysis has become an important yet challenging task, attracting significant attention. However, there is no preceding research on the influence functions of MTS to shed light on the effects of modifying the channel of MTS. Given that each channel in an MTS plays a crucial role in its analysis, it is essential to characterize the influence of different channels. To fill this gap, we propose a channel-wise influence function, which is the first method that can estimate the influence of different channels in MTS, utilizing a first-order gradient approximation. Additionally, we demonstrate how this influence function can be used to estimate the influence of a channel in MTS. Finally, we validated the accuracy and effectiveness of our influence estimation function in critical MTS analysis tasks, such as MTS anomaly detection and MTS forecasting. According to abundant experiments on real-world datasets, the original influence function performs worse than our method and even fails for the channel pruning problem, which demonstrates the superiority and necessity of the channel-wise influence function in MTS analysis.", "keywords": [ "channel-wise influence function", "mutivariate time series" ], "rating_list": [ 6, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "DKA7Hx7PSt", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Erik Vee", "gender": "Male", "institution": "Research, Google", "country": "", "position": "Researcher" }, { "name": "Fotis Iliopoulos", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Noel Loo", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Wei Hu", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "Assistant Professor" } ], "n_formula": 9, "n_formula_1": 4, "n_ref_uni": 27, "n_ref": 76, "n_ref_all": 104, "n_fig": 5, "n_tab": 7, "L_tab": 5553, "n_element_tab": 514, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 798, "n_element_tab_1": 70, "formula_len_all": 708, "formula_len_all_1": 144, "len_all": 176710, "len_all_1": 70811, "len_abs": 1736, "len_title": 115, "len_sents": 62265, "len_sents_1": 36042, "n_sents": 412, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1747, "title": "Linear Projections of Teacher Embeddings for Few-Class Distillation", "abs": "Knowledge Distillation (KD) has emerged as a promising approach for transferring knowledge from a larger, more complex teacher model to a smaller student model. Traditionally, KD involves training the student to mimic the teacher's output probabilities, while more advanced techniques have explored guiding the student to adopt the teacher's internal representations. Despite its widespread success, the performance of KD in binary classification and few-class problems has been less satisfactory. This is because the information about the teacher model\u2019s generalization patterns scales directly with the number of classes. Moreover, several sophisticated distillation methods may not be universally applicable or effective for data types beyond Computer Vision. Consequently, effective distillation techniques remain elusive for a range of key real-world applications, such as sentiment analysis, search query understanding, and advertisement-query relevance assessment. Taking these observations into account, we introduce a novel method for distilling knowledge from the teacher model's representations, which we term Learning Embedding Linear Projections (LELP). Inspired by recent findings about the structure of final-layer representations, LELP works by identifying informative linear subspaces in the teacher's embedding space, and splitting them into pseudo-subclasses. The student model is then trained to replicate these pseudo-subclasses. Our experimental evaluations on large-scale NLP benchmarks like Amazon Reviews and Sentiment140 demonstrate that LELP is consistently competitive with, and typically superior to, existing state-of-the-art distillation algorithms for binary and few-class problems, where most KD methods suffer.", "keywords": [ "Distillation", "Few-Classes", "Binary Classification" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 3, 5, 4 ] }, { "paper_id": "DJw1JBTmuk", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Lipeng Wan", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Postdoc" }, { "name": "Long Qian", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Sizhe Wang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xingyu Chen", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xuguang Lan", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Zeyang Liu", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Postdoc" }, { "name": "Ziru Wang", "gender": "unknown", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 68, "n_ref_all": 97, "n_fig": 20, "n_tab": 1, "L_tab": 1063, "n_element_tab": 107, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 88, "n_element_tab_1": 1, "formula_len_all": 1208, "formula_len_all_1": 1033, "len_all": 137285, "len_all_1": 53140, "len_abs": 1576, "len_title": 115, "len_sents": 41141, "len_sents_1": 26327, "n_sents": 295, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1499, "title": "Pre-Training Robo-Centric World Models For Efficient Visual Control", "abs": "Humans can accurately anticipate their movements to behave as expected in various manipulation tasks. We are inspired to propose that integrating prior knowledge of robot dynamics into world models can effectively improve the sample efficiency of model-based reinforcement learning (MBRL) in visual robot control tasks. In this paper, we introduce the Robo-Centric World Model (RCWM), which explicitly decouples the robot dynamics from the environment and enables pre-training to learn generalized and robust robot dynamics as prior knowledge to accelerate learning new tasks. Specifically, we construct respective dynamics models for the robot and the environment and learn their interactions through cross-attention mechanism. With the mask-guided reconfiguration mechanism, we only need a few prior robot segmentation masks to guide the RCWM to disentangle the robot and environment features and learn their respective dynamics. Our approach enables independent inference of robot dynamics from the environment, allowing accurate prediction of robot movement across various unseen tasks without being distracted by environmental variations. Our results in Meta-world demonstrate that RCWM is able to efficiently learn robot dynamics, improving sample efficiency for downstream tasks and enhancing policy robustness against environmental disturbances compared to the vanilla world model in DreamerV3. Code and visualizations are available on the project website: https://robo-centric-wm.github.io.", "keywords": [ "Pretraining World Models", "Model-based Reinforcement Learning", "Visual Robot Control" ], "rating_list": [ 8, 5, 3, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "DJSZGGZYVi", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Huiwon Jang", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Jinwoo Shin", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Full Professor" }, { "name": "Jonathan Huang", "gender": "unknown", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Jongheon Jeong", "gender": "Male", "institution": "Korea University", "country": "KR", "position": "Assistant Professor" }, { "name": "Sangkyung Kwak", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "MS student" }, { "name": "Sihyun Yu", "gender": "Male", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Saining Xie", "gender": "Male", "institution": "New York University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 32, "n_formula_1": 12, "n_ref_uni": 66, "n_ref": 199, "n_ref_all": 248, "n_fig": 18, "n_tab": 12, "L_tab": 4470, "n_element_tab": 810, "n_fig_1": 14, "n_tab_1": 5, "L_tab_1": 2048, "n_element_tab_1": 358, "formula_len_all": 1555, "formula_len_all_1": 551, "len_all": 250093, "len_all_1": 62549, "len_abs": 1430, "len_title": 145, "len_sents": 59347, "len_sents_1": 26237, "n_sents": 466, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 97, "L_abs": 1441, "title": "Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think", "abs": "Recent studies have shown that the denoising process in (generative) diffusion models can induce meaningful (discriminative) representations inside the model, though the quality of these representations still lags behind those learned through recent self-supervised learning methods. We argue that one main bottleneck in training large-scale diffusion models for generation lies in effectively learning these representations. Moreover, training can be made easier by incorporating high-quality external visual representations, rather than relying solely on the diffusion models to learn them independently. We study this by introducing a straightforward regularization called REPresentation Alignment (REPA), which aligns the projections of noisy input hidden states in denoising networks with clean image representations obtained from external, pretrained visual encoders. The results are striking: our simple strategy yields significant improvements in both training efficiency and generation quality when applied to popular diffusion and flow-based transformers, such as DiTs and SiTs. For instance, our method can speed up SiT training by over 17.5$\\times$, matching the performance (without classifier-free guidance) of a SiT-XL model trained for 7M steps in less than 400K steps. In terms of final generation quality, our approach achieves state-of-the-art results of FID=1.42 using classifier-free guidance with the guidance interval.", "keywords": [ "Diffusion models", "Representation learning" ], "rating_list": [ 10, 8, 8, 10, 8 ], "soundness_list": [ 4, 3, 3, 4, 3 ], "presentation_list": [ 3, 4, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "DJRd4IQHGQ", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Guangxu Zhu", "gender": "Male", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Researcher" }, { "name": "Haolong Chen", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Zhijie Cai", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" } ], "n_formula": 50, "n_formula_1": 20, "n_ref_uni": 40, "n_ref": 59, "n_ref_all": 95, "n_fig": 4, "n_tab": 11, "L_tab": 2396, "n_element_tab": 286, "n_fig_1": 2, "n_tab_1": 9, "L_tab_1": 4106, "n_element_tab_1": 252, "formula_len_all": 3331, "formula_len_all_1": 906, "len_all": 195875, "len_all_1": 61597, "len_abs": 2623, "len_title": 163, "len_sents": 40913, "len_sents_1": 24119, "n_sents": 371, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 115, "L_abs": 1960, "title": "FeedSign: Full-parameter Federated Fine-tuning of Large Models with Extremely Low Communication Overhead of One Bit", "abs": "Federated fine-tuning (FFT) aims to fine-tune a pre-trained model with private data from distributed clients by exchanging models rather than data under the orchestration of a parameter server (PS). However, as large models are acing in almost every machine learning task, the communication overhead and memory demand are surging accordingly, hindering the practical deployment on consumer devices. To overcome the bottleneck forged by the growing communication overhead of federated learning and lower the high memory demand of large model fine-tuning, we propose FeedSign, an FFT algorithm where a client uploads its update model and downloads the global model of any size using exactly $1$ bit per step, while the memory demand is squeezed to the amount needed for inference. This is realized by utilizing zeroth-order (ZO) optimizers on large models and shared pseudo-random number generators (PRNG) across devices to split the gradient estimate from the clients to 1) a direction corresponding to a designated random seed and 2) a binary vote from the client indicating whether the seed-corresponding direction grants a local loss descent, which is the only information the clients should convey to the PS. We conduct theoretical analysis on FeedSign and show that it converges at an exponential rate $\\mathcal{O}(e^{-t})$, where $t$ is the number of elapsed steps, the same rate as in first-order (FO) methods can attain in big $\\mathcal{O}$ notation. Moreover, it is also found that FeedSign enjoys good robustness against data heterogeneity and Byzantine attacks. We conduct extensive experiments on models across different structures and sizes (11M to 13B) and found that the proposed method performs better or closely, depending on scenarios, compared to its ZO and FO counterparts albeit an orders-of-magnitude lower communication overhead. We also discuss some interesting advantages as byproducts guaranteed by the minimalistic design of FeedSign.", "keywords": [ "Large Model Fine-tuning", "Federated Learning", "Heterogeneity Resilience", "Byzantine Resilience" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 3, 1 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "DIAaRdL2Ra", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Junhong Lin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yusu Hong", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 277, "n_formula_1": 34, "n_ref_uni": 34, "n_ref": 87, "n_ref_all": 203, "n_fig": 6, "n_tab": 3, "L_tab": 908, "n_element_tab": 33, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 651, "n_element_tab_1": 8, "formula_len_all": 26980, "formula_len_all_1": 3583, "len_all": 209922, "len_all_1": 64586, "len_abs": 895, "len_title": 120, "len_sents": 53944, "len_sents_1": 27102, "n_sents": 652, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 957, "title": "Convergence of Adafactor under Non-Convex Smooth Stochastic Optimization", "abs": "Adafactor, a memory-efficient variant of Adam, has emerged as one of the popular choices for training deep learning tasks, particularly large language models.\nHowever, despite its practical success, there is limited theoretical analysis of Adafactor's convergence. In this paper, we present a comprehensive analysis of Adafactor in a non-convex smooth setting. We show that full-batch Adafactor finds a stationary point at a rate of $\\tilde{O}(1/\\sqrt{T})$ with the default setup, which could be accelerated to $\\tilde{O}(1/T)$ with a constant step-size parameter. For stochastic Adafactor without update clipping, we prove a convergence rate of $\\tilde{O}(1/\\sqrt{T})$ with the right parameters covering the default setup. We also prove that Adafactor with a time-varying clipping threshold could also find a stationary point with the rate of $\\tilde{\\mathcal{O}}(1/\\sqrt{T})$. Our theoretical results are further complemented by some experimental results.", "keywords": [ "Adafactor", "stochastic optimization", "non-convex smooth optimization", "convergence" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 2, 2, 4 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 4, 3 ] }, { "paper_id": "DI4gW8viB6", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Eric John Li", "gender": "Male", "institution": "The Chinese University of Hong", "country": "", "position": "Undergrad student" }, { "name": "Jen-tse Huang", "gender": "Male", "institution": "Johns Hopkins University", "country": "US", "position": "Postdoc" }, { "name": "Man Ho LAM", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Michael Lyu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Tian Liang", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Wenxiang Jiao", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Wenxuan Wang", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Postdoc" }, { "name": "Xing Wang", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Youliang Yuan", "gender": "Male", "institution": "The Chinese University of Hong Kong-Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Zhaopeng Tu", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Principal Researcher" } ], "n_formula": 12, "n_formula_1": 7, "n_ref_uni": 60, "n_ref": 86, "n_ref_all": 133, "n_fig": 13, "n_tab": 9, "L_tab": 8951, "n_element_tab": 519, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 1642, "n_element_tab_1": 146, "formula_len_all": 384, "formula_len_all_1": 346, "len_all": 225590, "len_all_1": 76146, "len_abs": 1379, "len_title": 114, "len_sents": 62960, "len_sents_1": 35037, "n_sents": 576, "n_sents_1": 288, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1381, "title": "Competing Large Language Models in Multi-Agent Gaming Environments", "abs": "Decision-making is a complex process requiring diverse abilities, making it an excellent framework for evaluating Large Language Models (LLMs). Researchers have examined LLMs' decision-making through the lens of Game Theory. However, existing evaluation mainly focus on two-player scenarios where an LLM competes against another. Additionally, previous benchmarks suffer from test set leakage due to their static design. We introduce GAMA($\\gamma$)-Bench, a new framework for evaluating LLMs' Gaming Ability in Multi-Agent environments. It includes eight classical game theory scenarios and a dynamic scoring scheme specially designed to quantitatively assess LLMs' performance. $\\gamma$-Bench allows flexible game settings and adapts the scoring system to different game parameters, enabling comprehensive evaluation of robustness, generalizability, and strategies for improvement. Our results indicate that GPT-3.5 demonstrates strong robustness but limited generalizability, which can be enhanced using methods like Chain-of-Thought. We also evaluate 13 LLMs from 6 model families, including GPT-3.5, GPT-4, Gemini, LLaMA-3.1, Mixtral, and Qwen-2. Gemini-1.5-Pro outperforms others, scoring of $69.8$ out of $100$, followed by LLaMA-3.1-70B ($65.9$) and Mixtral-8x22B ($62.4$). Our code and experimental results are publicly available at https://github.com/CUHK-ARISE/GAMABench.", "keywords": [ "Large Language Models", "Games", "Reasoning", "Evaluation" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "DHVjLvSps6", "primary_area": "interpretability and explainable AI", "n_author": 9, "author_info_list": [ { "name": "Alessandro Moro", "gender": "Male", "institution": "The University of Tokyo, Tokyo Institute of Technology", "country": "JP", "position": "Visiting Researcher" }, { "name": "Armin W Thomas", "gender": "Non-Binary", "institution": "Liquid AI", "country": "US", "position": "Researcher" }, { "name": "Atsushi Yamashita", "gender": "Male", "institution": "The University of Tokyo", "country": "", "position": "Full Professor" }, { "name": "Michael Poli", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Neehal Tumma", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Qi An", "gender": "Male", "institution": "Tokyo University", "country": "JP", "position": "Associate Professor" }, { "name": "Stefano Massaroli", "gender": "unknown", "institution": "RIKEN", "country": "JP", "position": "Researcher" }, { "name": "Taiji Suzuki", "gender": "Male", "institution": "The University of Tokyo", "country": "", "position": "Full Professor" }, { "name": "Rom Nishijima Parnichkun", "gender": "Male", "institution": "The University of Tokyo", "country": "JP", "position": "PhD student" } ], "n_formula": 48, "n_formula_1": 10, "n_ref_uni": 61, "n_ref": 147, "n_ref_all": 207, "n_fig": 24, "n_tab": 18, "L_tab": 8972, "n_element_tab": 446, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 38, "n_element_tab_1": 10, "formula_len_all": 4883, "formula_len_all_1": 810, "len_all": 312494, "len_all_1": 70393, "len_abs": 1507, "len_title": 104, "len_sents": 96105, "len_sents_1": 33608, "n_sents": 730, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1525, "title": "Quantifying Memory Utilization with Effective State-Size", "abs": "As the space of causal sequence modeling architectures continues to grow, the need to develop a general framework for their analysis becomes increasingly important. With this aim, we draw insights from classical signal processing and control theory, to develop a quantitative measure of *memory utilization*: the internal mechanisms through which a model stores past information to produce future outputs. This metric, which we call **effective state-size** (ESS), is tailored to the fundamental class of systems with *input-invariant* and *input-varying linear operators*, encompassing a variety of computational units such as variants of attention, convolutions, and recurrences. Unlike prior work on memory utilization, which either relies on raw operator visualizations (e.g. attention maps), or simply the total *memory capacity* (i.e. cache size) of a model, our metrics provide highly interpretable and actionable measurements. In particular, we show how ESS can be leveraged to improve initialization strategies, inform novel regularizers and advance the performance-efficiency frontier through model distillation. Furthermore, we demonstrate that the effect of context delimiters (such as end-of-speech tokens) on ESS highlights cross-architectural differences in how large language models utilize their available memory to recall information. Overall, we find that ESS provides valuable insights into the dynamics that dictate memory utilization, enabling the design of more efficient and effective sequence models.", "keywords": [ "model analysis", "interpretability", "linear systems", "attention", "state-space models", "sequence models", "memory utilization", "context utilization" ], "rating_list": [ 5, 6, 6, 5, 6 ], "soundness_list": [ 2, 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 3, 3 ], "confidence_list": [ 3, 5, 3, 4, 3 ] }, { "paper_id": "DGjzxNRbKU", "primary_area": "learning theory", "n_author": 5, "author_info_list": [ { "name": "Alberto Marchesi", "gender": "Male", "institution": "Politecnico di Milano", "country": "", "position": "Assistant Professor" }, { "name": "Francesco Bacchiocchi", "gender": "Male", "institution": "Polytechnic Institute of Milan", "country": "IT", "position": "PhD student" }, { "name": "Francesco Emanuele Stradi", "gender": "Male", "institution": "Polytechnic Institute of Milan", "country": "IT", "position": "PhD student" }, { "name": "Matteo Castiglioni", "gender": "unknown", "institution": "Politecnico di Milano", "country": "IT", "position": "Assistant Professor" }, { "name": "Nicola Gatti", "gender": "Male", "institution": "Polytechnic Institute of Milan", "country": "IT", "position": "Full Professor" } ], "n_formula": 103, "n_formula_1": 16, "n_ref_uni": 36, "n_ref": 91, "n_ref_all": 107, "n_fig": 0, "n_tab": 1, "L_tab": 170, "n_element_tab": 3, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 25, "n_element_tab_1": 2, "formula_len_all": 16224, "formula_len_all_1": 1659, "len_all": 199207, "len_all_1": 76914, "len_abs": 968, "len_title": 110, "len_sents": 63762, "len_sents_1": 35837, "n_sents": 493, "n_sents_1": 251, "venue": "ICLR2025", "decision": "Reject", "L_title": 62, "L_abs": 972, "title": "Markov Persuasion Processes: Learning to Persuade From Scratch", "abs": "In Bayesian persuasion, an informed sender strategically discloses information to a receiver so as to persuade them to undertake desirable actions. Recently, Markov persuasion processes (MPPs) have been introduced to capture sequential scenarios where a sender faces a stream of myopic receivers in a Markovian environment. The MPPs studied so far in the literature suffer from issues that prevent them from being fully operational in practice, e.g., they assume that the sender knows receivers' rewards. We fix such issues by addressing MPPs where the sender has no knowledge about the environment. We design a learning algorithm for the sender, working with partial feedback. We prove that its regret with respect to an optimal information-disclosure policy grows sublinearly in the number of episodes, as it is the case for the loss in persuasiveness cumulated while learning. Moreover, we provide a lower bound for our setting matching the guarantees of our algorithm.", "keywords": [ "Bayesian Persuasion", "Online Learning", "Markov Persuasion Process" ], "rating_list": [ 3, 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 1, 3, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 2, 5 ] }, { "paper_id": "DFSb67ksVr", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 4, "author_info_list": [ { "name": "Abir De", "gender": "Male", "institution": "Indian Institute of Technology Bombay,", "country": "IN", "position": "Assistant Professor" }, { "name": "Eeshaan Jain", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Indradyumna Roy", "gender": "Male", "institution": "Indian Institute of Technology Bombay", "country": "", "position": "PhD student" }, { "name": "Soumen Chakrabarti", "gender": "Not Specified", "institution": "Indian Institute of Technology Bombay", "country": "", "position": "Professor" } ], "n_formula": 70, "n_formula_1": 36, "n_ref_uni": 59, "n_ref": 161, "n_ref_all": 224, "n_fig": 2, "n_tab": 14, "L_tab": 5075, "n_element_tab": 464, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1482, "n_element_tab_1": 113, "formula_len_all": 39357, "formula_len_all_1": 19617, "len_all": 366482, "len_all_1": 123419, "len_abs": 1269, "len_title": 134, "len_sents": 114586, "len_sents_1": 53433, "n_sents": 688, "n_sents_1": 271, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1278, "title": "Clique Number Estimation via Differentiable Functions of Adjacency Matrix Permutations", "abs": "Estimating the clique number in a graph is central to various applications, e.g., community detection, graph retrieval, etc. \nExisting estimators often rely on non-differentiable combinatorial components. Here, we propose a full differentiable estimator for clique number estimation, which can be trained from distant supervision of clique numbers, rather than demonstrating actual cliques.\nOur key insight is a formulation of the maximum clique problem (MCP) as a maximization of the size of fully dense square submatrix, within a suitably row-column-permuted adjacency matrix.\nWe design a differentiable mechanism to search for permutations that lead to the discovery of such dense blocks.\nHowever, the optimal permutation is not unique, which leads to the learning of spurious permutations. To tackle this problem, we view the MCP problem as a sequence of subgraph matching tasks, each detecting progressively larger cliques in a nested manner. This allows effective navigation through suitable node permutations.\nThese steps result in MxNet, an end-to-end differentiable model, which learns to predict clique number without explicit clique demonstrations, with the added benefit of interpretability. Experiments on eight datasets show the superior accuracy of our approach.", "keywords": [ "Graph neural network", "distant supervision" ], "rating_list": [ 8, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 4, 3, 2 ] }, { "paper_id": "DF5TVzpTW0", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Hsiang Hsiao", "gender": "unknown", "institution": "National Taiwan University", "country": "TW", "position": "MS student" }, { "name": "Shou-De Lin", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "Full Professor" }, { "name": "Yuche Tsai", "gender": "Male", "institution": "National Taiwan University", "country": "", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 27, "n_ref": 53, "n_ref_all": 78, "n_fig": 8, "n_tab": 8, "L_tab": 4214, "n_element_tab": 488, "n_fig_1": 6, "n_tab_1": 6, "L_tab_1": 4795, "n_element_tab_1": 248, "formula_len_all": 642, "formula_len_all_1": 549, "len_all": 133795, "len_all_1": 69438, "len_abs": 855, "len_title": 137, "len_sents": 38607, "len_sents_1": 28012, "n_sents": 322, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 866, "title": "Detecting and Perturbing Privacy-Sensitive Neurons to Defend Embedding Inversion Attacks", "abs": "This paper introduces Defense through Perturbing Privacy Neurons (DPPN), a novel approach to protect text embeddings against inversion attacks. Unlike ex- isting methods that add noise to all embedding dimensions for general protection, DPPN identifies and perturbs only a small portion of privacy-sensitive neurons. We present a differentiable neuron mask learning framework to detect these neu- rons and a neuron-suppressing perturbation function for targeted noise injection. Experiments across six datasets show DPPN achieves superior privacy-utility trade- offs. Compared to baseline methods, DPPN reduces more privacy leakage by 5-78% while improving downstream task performance by 14-40%. Tests on real- world sensitive datasets demonstrate DPPN\u2019s effectiveness in mitigating sensitive information leakage to 17%, while baseline methods reduce it only to 43%.", "keywords": [ "Text embedding", "Defense Inversion Attack" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 4 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 3, 5, 3, 3 ] }, { "paper_id": "DEOV74Idsg", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "Byungju Lee", "gender": "Male", "institution": "Korea Institute of Science and Technology", "country": "KR", "position": "Researcher" }, { "name": "HyeonJung Kim", "gender": "Male", "institution": "POSCO Holdings", "country": "", "position": "Researcher" }, { "name": "Kyuri Choi", "gender": "Female", "institution": "POSCO holdings", "country": "", "position": "Researcher" }, { "name": "Linda Ruth Petzold", "gender": "Female", "institution": "University of California, Santa Barbara", "country": "", "position": "Full Professor" }, { "name": "Ryan Hsieh", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "Undergrad student" }, { "name": "Stephen D. Wilson", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "Full Professor" }, { "name": "Sungyoung Ji", "gender": "Male", "institution": "POSCO HOLDINGS", "country": "", "position": "Researcher" }, { "name": "Wanrong Zhu", "gender": "Female", "institution": "Adobe Research", "country": "US", "position": "Researcher" }, { "name": "Woosang Lim", "gender": "Male", "institution": "POSCO Holdings", "country": "US", "position": "Research Fellow (CVP), Head of Applied AI Research" }, { "name": "Xianjun Yang", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Xifeng Yan", "gender": "unknown", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" }, { "name": "Zekun Li", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "JinHyuk Lim", "gender": "Male", "institution": "POSCO Holdings", "country": "", "position": "Researcher" }, { "name": "William Yang Wang", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 39, "n_ref": 92, "n_ref_all": 123, "n_fig": 8, "n_tab": 13, "L_tab": 9917, "n_element_tab": 1579, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 4562, "n_element_tab_1": 714, "formula_len_all": 180, "formula_len_all_1": 56, "len_all": 252657, "len_all_1": 66950, "len_abs": 1714, "len_title": 137, "len_sents": 63777, "len_sents_1": 25213, "n_sents": 517, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 88, "L_abs": 1724, "title": "MMSci: A Dataset for Graduate-Level Multi-Discipline Multimodal Scientific Understanding", "abs": "The rapid development of Multimodal Large Language Models (MLLMs) is making AI-driven scientific assistants increasingly feasible, with interpreting scientific figures being a crucial task. However, existing datasets and benchmarks focus mainly on basic charts and limited science subjects, lacking comprehensive evaluations. To address this, we curated a multimodal, multidisciplinary dataset from peer-reviewed, open-access Nature Communications articles, spanning 72 scientific disciplines. This dataset includes figures such as schematic diagrams, simulated images, macroscopic/microscopic photos, and experimental visualizations (e.g., western blots), which often require graduate-level, discipline-specific expertise to interpret. We developed benchmarks for scientific figure captioning and multiple-choice questions, evaluating six proprietary and over ten open-source models across varied settings. The results highlight the high difficulty of these tasks and the significant performance gap among models. While many open-source models performed at chance level on the multiple-choice task, some matched the performance of proprietary models. However, the gap was more pronounced in the captioning task. Our dataset also provide valuable resource for training. Fine-tuning the Qwen2-VL-2B model with our task-specific multimodal training data improved its multiple-choice accuracy to a level comparable to GPT-4o, though captioning remains challenging. Continuous pre-training of MLLMs using our interleaved article and figure data enhanced their material generation capabilities, demonstrating potential for integrating scientific knowledge. The dataset and benchmarks will be released to support further research.", "keywords": [ "Scientific Figure Understanding", "Multimodal Large Language Model", "Large Vision Language Model", "Multi-discipline", "Multimodal", "Scientific knowledge understanding", "benchmark", "Nature science", "Visual instruction-following" ], "rating_list": [ 5, 6, 5, 8 ], "soundness_list": [ 2, 3, 1, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 4 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "DDxLsxiZR8", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Xinle Cheng", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhihao Jia", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Zhuoming Chen", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 22, "n_ref": 31, "n_ref_all": 41, "n_fig": 7, "n_tab": 3, "L_tab": 1654, "n_element_tab": 257, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1832, "n_element_tab_1": 54, "formula_len_all": 281, "formula_len_all_1": 276, "len_all": 83045, "len_all_1": 35394, "len_abs": 376, "len_title": 173, "len_sents": 17916, "len_sents_1": 14947, "n_sents": 139, "n_sents_1": 113, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 664, "title": "CAT Pruning: Cluster-Aware Token Pruning For Text-to-Image Diffusion Models", "abs": "Diffusion models have transformed generative tasks, particularly in text-to-image synthesis, but their iterative denoising process is computationally intensive. We present a novel acceleration strategy that combines token-level pruning with cache mechanisms to address this challenge. By utilizing Noise Relative Magnitude, we identify significant token changes across iterations. Additionally, we incorporate spatial clustering and distributional balance to enhance token selection. Our experiments demonstrate 50\\%-60\\% reduction in computational cost while maintaining model performance, offering a substantial improvement in the efficiency of diffusion models.", "keywords": [ "Generative Models", "Efficient Machine Learning" ], "rating_list": [ 5, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "DDNFTaVQdU", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Lichen Zhang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yuzhou Gu", "gender": "Male", "institution": "New York University", "country": "US", "position": "Postdoc" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" } ], "n_formula": 311, "n_formula_1": 34, "n_ref_uni": 56, "n_ref": 152, "n_ref_all": 193, "n_fig": 0, "n_tab": 13, "L_tab": 3112, "n_element_tab": 193, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 516, "n_element_tab_1": 7, "formula_len_all": 44765, "formula_len_all_1": 2453, "len_all": 390121, "len_all_1": 71956, "len_abs": 1773, "len_title": 74, "len_sents": 97561, "len_sents_1": 30802, "n_sents": 1259, "n_sents_1": 227, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 74, "L_abs": 1897, "title": "Faster Algorithms for Structured Linear and Kernel Support Vector Machines", "abs": "Quadratic programming is a ubiquitous prototype in convex programming. Many machine learning problems can be formulated as quadratic programming, including the famous Support Vector Machines (SVMs). Linear and kernel SVMs have been among the most popular models in machine learning over the past three decades, prior to the deep learning era.\n\nGenerally, a quadratic program has an input size of $\\Theta(n^2)$, where $n$ is the number of variables. Assuming the Strong Exponential Time Hypothesis ($\\textsf{SETH}$), it is known that no $O(n^{2-o(1)})$ time algorithm exists when the quadratic objective matrix is positive semidefinite (Backurs, Indyk, and Schmidt, NeurIPS'17). However, problems such as SVMs usually admit much smaller input sizes: one is given $n$ data points, each of dimension $d$, and $d$ is oftentimes much smaller than $n$. Furthermore, the SVM program has only $O(1)$ equality linear constraints. This suggests that faster algorithms are feasible, provided the program exhibits certain structures.\n\nIn this work, we design the first nearly-linear time algorithm for solving quadratic programs whenever the quadratic objective admits a low-rank factorization, and the number of linear constraints is small. Consequently, we obtain results for SVMs:\n\n* For linear SVM when the input data is $d$-dimensional, our algorithm runs in time $\\widetilde O(nd^{(\\omega+1)/2}\\log(1/\\epsilon))$ where $\\omega\\approx 2.37$ is the fast matrix multiplication exponent;\n \n* For Gaussian kernel SVM, when the data dimension $d = O(\\log n)$ and the squared dataset radius is sub-logarithmic in $n$, our algorithm runs in time $O(n^{1+o(1)}\\log(1/\\epsilon))$. We also prove that when the squared dataset radius is at least $\\Omega(\\log^2 n)$, then $\\Omega(n^{2-o(1)})$ time is required. This improves upon the prior best lower bound in both the dimension $d$ and the squared dataset radius.", "keywords": [ "interior point method", "support vector machine", "data structure" ], "rating_list": [ 5, 8, 8 ], "soundness_list": [ 3, 4, 4 ], "presentation_list": [ 2, 3, 4 ], "contribution_list": [ 2, 3, 4 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "DD11okKg13", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 5, "author_info_list": [ { "name": "Amir Mohammad Karimi Mamaghan", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "PhD student" }, { "name": "Andrea Dittadi", "gender": "Male", "institution": "Helmholtz AI", "country": "DE", "position": "Postdoc" }, { "name": "Karl Henrik Johansson", "gender": "unknown", "institution": "KTH Royal Institute of Technology", "country": "", "position": "Full Professor" }, { "name": "Samuele Papa", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Stefan Bauer", "gender": "unknown", "institution": "Helmholtz Munich ", "country": "DE", "position": "Director" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 63, "n_ref": 179, "n_ref_all": 210, "n_fig": 12, "n_tab": 12, "L_tab": 8369, "n_element_tab": 812, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 1273, "n_element_tab_1": 44, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 283961, "len_all_1": 67248, "len_abs": 1268, "len_title": 183, "len_sents": 80008, "len_sents_1": 33781, "n_sents": 581, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 135, "L_abs": 1226, "title": "Exploring the Effectiveness of Object-Centric Representations in Visual Question Answering: Comparative Insights with Foundation Models", "abs": "Object-centric (OC) representations, which represent the state of a visual scene by modeling it as a composition of objects, have the potential to be used in various downstream tasks to achieve systematic compositional generalization and facilitate reasoning. However, these claims have not been thoroughly analyzed yet.\nRecently, foundation models have demonstrated unparalleled capabilities across diverse domains from language to computer vision, marking them as a potential cornerstone of future research for a multitude of computational tasks.\nIn this paper, we conduct an extensive empirical study on representation learning for downstream Visual Question Answering (VQA), which requires an accurate compositional understanding of the scene. \nWe thoroughly investigate the benefits and trade-offs of OC models and alternative approaches including large pre-trained foundation models on both synthetic and real-world data, and demonstrate a viable way to achieve the best of both worlds. \nThe extensiveness of our study, encompassing over 600 downstream VQA models and 15 different types of upstream representations, also provides several additional insights that we believe will be of interest to the community at large.", "keywords": [ "Object-centric Learning", "Foundation Models" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 4, 4, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 5, 3, 4 ] }, { "paper_id": "DCpukR83sw", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Cewu Lu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Jianhua Sun", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Liang Chai", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Yuxuan Li", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "MS student" } ], "n_formula": 35, "n_formula_1": 16, "n_ref_uni": 6, "n_ref": 138, "n_ref_all": 177, "n_fig": 12, "n_tab": 10, "L_tab": 7184, "n_element_tab": 479, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 732, "n_element_tab_1": 25, "formula_len_all": 2883, "formula_len_all_1": 1178, "len_all": 248284, "len_all_1": 54013, "len_abs": 1424, "len_title": 127, "len_sents": 80300, "len_sents_1": 25339, "n_sents": 619, "n_sents_1": 181, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1431, "title": "Interactive Adjustment for Human Trajectory Prediction with Individual Feedback", "abs": "Human trajectory prediction is fundamental for autonomous driving and service robot. The research community has studied various important aspects of this task and made remarkable progress recently. However, there is an essential perspective which is not well exploited in previous research all along, namely individual feedback. Individual feedback exists in the sequential nature of trajectory prediction, where earlier predictions of a target can be verified over time by his ground-truth trajectories to obtain feedback which provides valuable experience for subsequent predictions on the same agent. In this paper, we show such feedback can reveal the strengths and weaknesses of the model's predictions on a specific target and heuristically guide to deliver better predictions on him. We present an interactive adjustment network to effectively model and leverage the feedback. This network first exploits the feedback from previous predictions to dynamically generate an adjuster which then interactively makes appropriate adjustments to current predictions for more accurate ones. We raise a novel displacement expectation loss to train this interactive architecture. Through experiments on representative prediction methods and widely-used benchmarks, we demonstrate the great value of individual feedback and the superior effectiveness of proposed interactive adjustment network. Our code will be made publicly available.", "keywords": [ "Human Trajectory Prediction" ], "rating_list": [ 8, 6, 6, 3 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "DCg9r2DKKe", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Jeffrey Wishart", "gender": "unknown", "institution": "Arizona State University", "country": "", "position": "Researcher" }, { "name": "Junfeng Zhao", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Assistant Professor" }, { "name": "Varun Chandra Jammula", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Yezhou Yang", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 20, "n_ref": 34, "n_ref_all": 38, "n_fig": 1, "n_tab": 2, "L_tab": 764, "n_element_tab": 107, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 764, "n_element_tab_1": 107, "formula_len_all": 352, "formula_len_all_1": 352, "len_all": 77318, "len_all_1": 63210, "len_abs": 1758, "len_title": 114, "len_sents": 26869, "len_sents_1": 26819, "n_sents": 178, "n_sents_1": 178, "venue": "ICLR2025", "decision": "withdraw", "L_title": 66, "L_abs": 1767, "title": "STL-Drive: Formal Verification Guided End-to-end Automated Driving", "abs": "End-to-end automated driving behavior models require extensive training data from machine or human driver experts or interacting with the environment to learn a driving policy. Not all human driver expert data represent safe driving that the end-to-end model is learning to imitate, and similarly, neither are some of the behaviors learned during exploration while learning by trial and error. However, the models should learn from such data without being negatively affected during the learning process. We aim to provide a learning framework to incorporate formal verification methods to improve the robustness and safety of the learned models in the presence of training data that contain unsafe behaviors, dubbed as STL-Drive. We are particularly interested in utilizing this framework to enhance the safety of end-to-end automated driving models. In this work, we incorporate Signal Temporal Logic (STL) as the formal method to impose safety constraints. In addition, we utilize the Responsibility-Sensitive Safety (RSS) framework to define the safety constraints. We designed a loss function that combines the task objectives and the STL robustness score to balance the learned policy's performance and safety. We demonstrate that encoding safety constraints using STL and utilizing the robustness score during training improves the performance and safety of the driving policy. We validate our framework using open-loop predictive simulator NAVSIM and real-world data from OpenScene. The results of this study suggest a promising research direction where formal methods can enhance the safety and resilience of deep learning models. Formal verification of safety constraints for automated driving will further increase the public's trust in automated vehicles.", "keywords": [ "Formal Verification", "Automated Driving", "Imitation Learning", "Robustness", "Safety" ], "rating_list": [ 1, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 1, 2, 1, 1 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "DCandSZ2F1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Jianfei Cai", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" }, { "name": "Mengyao Li", "gender": "unknown", "institution": "Shanghai University", "country": "CN", "position": "PhD student" }, { "name": "Qianyi Wu", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Weiyao Lin", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Yihang Chen", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Mehrtash T. Harandi", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Associate Professor" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 36, "n_ref": 75, "n_ref_all": 95, "n_fig": 12, "n_tab": 1, "L_tab": 646, "n_element_tab": 80, "n_fig_1": 9, "n_tab_1": 1, "L_tab_1": 441, "n_element_tab_1": 70, "formula_len_all": 964, "formula_len_all_1": 967, "len_all": 133031, "len_all_1": 69314, "len_abs": 1393, "len_title": 99, "len_sents": 35892, "len_sents_1": 30562, "n_sents": 272, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 50, "L_abs": 1161, "title": "Fast Feedforward 3D Gaussian Splatting Compression", "abs": "With 3D Gaussian Splatting (3DGS) advancing real-time and high-fidelity rendering for novel view synthesis, storage requirements pose challenges for their widespread adoption. Although various compression techniques have been proposed, previous art suffers from a common limitation: for any existing 3DGS, per-scene optimization is needed to achieve compression, making the compression sluggish and slow. To address this issue, we introduce Fast Compression of 3D Gaussian Splatting (FCGS), an optimization-free model that can compress 3DGS representations rapidly in a single feed-forward pass, which significantly reduces compression time from minutes to seconds. To enhance compression efficiency, we propose a multi-path entropy module that assigns Gaussian attributes to different entropy constraint paths for balance between size and fidelity. We also carefully design both inter- and intra-Gaussian context models to remove redundancies among the unstructured Gaussian blobs. Overall, FCGS achieves a compression ratio of over 20X while maintaining fidelity, surpassing most per-scene SOTA optimization-based methods. Code: github.com/YihangChen-ee/FCGS.", "keywords": [ "3DGS", "compression", "optimization-free" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "DC8bsa9bzY", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Gabriel Wu", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "Undergrad student" }, { "name": "Jacob Hilton", "gender": "Male", "institution": "Alignment Research Center", "country": "US", "position": "Researcher" } ], "n_formula": 37, "n_formula_1": 12, "n_ref_uni": 28, "n_ref": 38, "n_ref_all": 60, "n_fig": 6, "n_tab": 6, "L_tab": 4547, "n_element_tab": 702, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 597, "n_element_tab_1": 58, "formula_len_all": 2796, "formula_len_all_1": 476, "len_all": 149518, "len_all_1": 58628, "len_abs": 1107, "len_title": 63, "len_sents": 44667, "len_sents_1": 26089, "n_sents": 376, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 63, "L_abs": 1112, "title": "Estimating the Probabilities of Rare Outputs in Language Models", "abs": "We consider the problem of *low probability estimation*: given a machine learning model and a formally-specified input distribution, how can we estimate the probability of a binary property of the model's output, even when that probability is too small to estimate by random sampling? This problem is motivated by the need to improve worst-case performance, which distribution shift can make much more likely. We study low probability estimation in the context of argmax sampling from small transformer language models. We compare two types of methods: importance sampling, which involves searching for inputs giving rise to the rare output, and activation extrapolation, which involves extrapolating a probability distribution fit to the model's logits. We find that importance sampling outperforms activation extrapolation, but both outperform naive sampling. Finally, we explain how minimizing the probability estimate of an undesirable behavior generalizes adversarial training, and argue that new methods for low probability estimation are needed to provide stronger guarantees about worst-case performance.", "keywords": [ "low probabilities", "adversarial training", "importance sampling" ], "rating_list": [ 8, 5, 6, 10 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 2, 3, 4, 4 ], "contribution_list": [ 3, 2, 3, 4 ], "confidence_list": [ 3, 4, 2, 4 ] }, { "paper_id": "DBitNcZa6T", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Georgy Derevyanko", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Postdoc" }, { "name": "Petr Popov", "gender": "unknown", "institution": "Skoltech", "country": "", "position": "Assistant Professor" } ], "n_formula": 32, "n_formula_1": 8, "n_ref_uni": 27, "n_ref": 37, "n_ref_all": 63, "n_fig": 10, "n_tab": 11, "L_tab": 2900, "n_element_tab": 249, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 673, "n_element_tab_1": 32, "formula_len_all": 4200, "formula_len_all_1": 380, "len_all": 176728, "len_all_1": 66501, "len_abs": 958, "len_title": 153, "len_sents": 57164, "len_sents_1": 29865, "n_sents": 406, "n_sents_1": 216, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1486, "title": "AlphaMol: rigid neighborhood representation for small molecule structure prediction", "abs": "Recent success of the deep learning-based approach AlphaFold2 revolutionized the field of protein structure prediction. Since AlphaFold2 development, a lot of efforts were made to shape its limitations as well as to improve it further with respect to difficult protein classes. However, structure prediction of non-protein type of molecules, such as small organic molecules, non-standard amino acids, nucleic acids, and others, is still an open problem. Inspired by the powerful AlphaFold2 neural network architecture, we developed a general framework for prediction of molecular structures of arbitrary type. Specifically, we developed novel representation of molecular structures as a collection of rigid-body neighborhoods with encoded bonds between the neighborhoods fed into neural network comprising developed Evoformer-like blocks. We tested our approach on small organic molecules, that possess much higher variability in terms of atomic composition and structural patterns compared to proteins. Namely, we applied the developed method, named AlphaMol, to the ground-state structure prediction problem of small molecules and observed superior performance metrics on the PubChemQC benchmark, compared to the existing approaches. Our results demonstrate possibility to create multi-modal molecular structure prediction methods, that operate across different molecular types. The AlphaMol source code is available in the repository: https://anonymous.4open.science/r/AlphaMol-2EA7.", "keywords": [ "molecule", "deep learning", "structure prediction", "chemistry", "biology", "biomolecules", "bioinformatics" ], "rating_list": [ 3, 8, 3, 5, 3 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 1, 3, 2, 3, 1 ], "contribution_list": [ 2, 3, 2, 3, 2 ], "confidence_list": [ 5, 3, 2, 2, 4 ] }, { "paper_id": "DBbgasVgyQ", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "He Zhu", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Assistant Professor" }, { "name": "Wenjie Qiu", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Wensen Mao", "gender": "Male", "institution": ", Rutgers University", "country": "US", "position": "PhD student" }, { "name": "Yuanlin Duan", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "PhD student" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 16, "n_ref": 34, "n_ref_all": 66, "n_fig": 12, "n_tab": 9, "L_tab": 3354, "n_element_tab": 155, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 897, "n_element_tab_1": 141, "formula_len_all": 1330, "formula_len_all_1": 1067, "len_all": 169391, "len_all_1": 65790, "len_abs": 1667, "len_title": 85, "len_sents": 58887, "len_sents_1": 31971, "n_sents": 433, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 1674, "title": "Skill Discovery using Language Models", "abs": "Large Language models (LLMs) possess remarkable ability to understand natural language descriptions of complex robotics environments. Earlier studies have shown that LLM agents can use a predefined set of skills for robot planning in long-horizon tasks. However, the requirement for prior knowledge of the skill set required for a given task constrains its applicability and flexibility. We present a novel approach L2S (short of Language2Skills) to leverage the generalization capabilities of LLMs to decompose the natural language task description of a complex task to definitions of reusable skills. Each skill is defined by an LLM-generated dense reward function and a termination condition, which in turn lead to effective skill policy training and chaining for task execution. To address the uncertainty surrounding the parameters used by the LLM agent in the generated reward and termination functions, L2S trains parameter-conditioned skill policies that performs well across a broad spectrum of parameter values. As the impact of these parameters for one skill on the overall task becomes apparent only when its following skills are trained, L2S selects the most suitable parameter value during the training of the subsequent skills to effectively mitigate the risk associated with incorrect parameter choices. During training, L2S autonomously accumulates a skill library from continuously presented tasks and their descriptions, leveraging guidance from the LLM agent to effectively apply this skill library in tackling novel tasks. Our experimental results show that L2S is capable of generating reusable skills to solve a wide range of robot manipulation tasks.", "keywords": [ "Reinforcement learning", "Large language models", "Robotics" ], "rating_list": [ 6, 6, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 4, 5, 2, 4 ] }, { "paper_id": "DAEXilQHYU", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Ismail Ilkan Ceylan", "gender": "unknown", "institution": "University of Oxford", "country": "", "position": "Assistant Professor" }, { "name": "Michael M. Bronstein", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Full Professor" }, { "name": "Miguel Romero Orth", "gender": "Male", "institution": "Pontificia Universidad Catolica de Chile", "country": "CL", "position": "Assistant Professor" }, { "name": "Pablo Barcelo", "gender": "Male", "institution": "Pontificia Universidad Cat\u00f3lica", "country": "CL", "position": "Full Professor" }, { "name": "Xingyue Huang", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "PhD student" } ], "n_formula": 147, "n_formula_1": 17, "n_ref_uni": 45, "n_ref": 157, "n_ref_all": 205, "n_fig": 6, "n_tab": 14, "L_tab": 4117, "n_element_tab": 491, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 14081, "formula_len_all_1": 1451, "len_all": 312189, "len_all_1": 60218, "len_abs": 1122, "len_title": 91, "len_sents": 102266, "len_sents_1": 27031, "n_sents": 1012, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 43, "L_abs": 1125, "title": "Link Prediction with Relational Hypergraphs", "abs": "Link prediction with knowledge graphs has been thoroughly studied in graph machine learning, leading to a rich landscape of graph neural network architectures with successful applications. \nNonetheless, it remains challenging to transfer the success of these architectures to link prediction with *relational hypergraphs*, where the task is over *$k$-ary relations*, substantially harder than link prediction on knowledge graphs with binary relations only.\nIn this paper, we propose a framework for link prediction with relational hypergraphs, empowering applications of graph neural networks on *fully relational* structures. Theoretically, we conduct a thorough analysis of the expressive power of the resulting model architectures via corresponding relational Weisfeiler-Leman algorithms and also via logical expressiveness. Empirically, we validate the power of the proposed model architectures on various relational hypergraph benchmarks. The resulting model architectures substantially outperform every baseline for inductive link prediction, and also lead to state-of-the-art results for transductive link prediction.", "keywords": [ "link prediction", "relational hypergraphs", "expressivity study" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "D9liZ0D8z8", "primary_area": "learning on time series and dynamical systems", "n_author": 11, "author_info_list": [ { "name": "Abdul Fatir Ansari", "gender": "Male", "institution": "AWS AI Labs (Amazon)", "country": "DE", "position": "Senior Scientist" }, { "name": "Andrew Gordon Wilson", "gender": "Not Specified", "institution": "New York University", "country": "US", "position": "Full Professor" }, { "name": "Bernie Wang", "gender": "Male", "institution": "AWS AI Labs", "country": "US", "position": "Principal Researcher" }, { "name": "Boran Han", "gender": "Female", "institution": "Amazon/AWS", "country": "US", "position": "Researcher" }, { "name": "Christos Faloutsos", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Danielle C. Maddix", "gender": "unknown", "institution": "AWS AI Labs", "country": "US", "position": "Applied Scientist" }, { "name": "Luca Masserano", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Michael W. Mahoney", "gender": "unknown", "institution": "University of California Berkeley", "country": "US", "position": "Full Professor" }, { "name": "Syama Sundar Rangapuram", "gender": "Male", "institution": "Amazon Development Center Germany", "country": "", "position": "Researcher" }, { "name": "Xiyuan Zhang", "gender": "unknown", "institution": "AWS", "country": "US", "position": "Applied Scientist" }, { "name": "Youngsuk Park", "gender": "Male", "institution": "Amazon, AWS AI Labs", "country": "", "position": "Research" } ], "n_formula": 12, "n_formula_1": 4, "n_ref_uni": 57, "n_ref": 91, "n_ref_all": 132, "n_fig": 12, "n_tab": 11, "L_tab": 17576, "n_element_tab": 2660, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 592, "n_element_tab_1": 8, "formula_len_all": 700, "formula_len_all_1": 214, "len_all": 212148, "len_all_1": 64347, "len_abs": 1570, "len_title": 86, "len_sents": 51171, "len_sents_1": 33850, "n_sents": 327, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 86, "L_abs": 1575, "title": "Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization", "abs": "There is a major open question about how to best develop foundation models for time series forecasting. Tokenization is a crucial consideration in this effort: what is an effective discrete vocabulary for a real-valued sequential input? To address this question, we develop WaveToken, a wavelet-based tokenizer that allows models to learn complex representations directly in the space of time-localized frequencies. Our method first scales and decomposes the input time series, then thresholds and quantizes the wavelet coefficients, and finally pre-trains an autoregressive model to forecast coefficients for the horizon window. By decomposing coarse and fine structures in the inputs, wavelets provide an eloquent and compact language for time series forecasting that simplifies learning. Empirical results on a comprehensive benchmark, including 42 datasets for both in-domain and zero-shot settings, show that WaveToken: i) provides better accuracy than recently proposed foundation models for forecasting while using a much smaller vocabulary (1024 tokens), and performs on par or better than modern deep learning models trained specifically on each dataset; and ii) exhibits superior generalization capabilities, achieving the best average rank across all datasets for three complementary metrics. In addition, we show that our method can easily capture complex temporal patterns of practical relevance that are challenging for other recent pre-trained models, including trends, sparse spikes, and non-stationary time series with varying frequencies evolving over time.", "keywords": [ "time series forecasting", "foundation models", "wavelets", "tokenization", "frequency", "pretrained model", "nonstationary time series" ], "rating_list": [ 5, 6, 8, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 2, 4, 1 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "D9JSxF2Xhx", "primary_area": "interpretability and explainable AI", "n_author": 4, "author_info_list": [ { "name": "Bernhard Y Renard", "gender": "unknown", "institution": "Icahn School of Medicine at Mount Sinai", "country": "US", "position": "Full Professor" }, { "name": "Katharina Baum", "gender": "Female", "institution": "Freie Universit\u00e4t Berlin", "country": "", "position": "Assistant Professor" }, { "name": "Pascal Iversen", "gender": "unknown", "institution": "Freie Universit\u00e4t Berlin", "country": "", "position": "PhD student" }, { "name": "Simon Witzke", "gender": "Male", "institution": "Hasso Plattner Institute", "country": "", "position": "PhD student" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 45, "n_ref": 67, "n_ref_all": 89, "n_fig": 12, "n_tab": 3, "L_tab": 754, "n_element_tab": 119, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 389, "n_element_tab_1": 30, "formula_len_all": 646, "formula_len_all_1": 539, "len_all": 162784, "len_all_1": 83900, "len_abs": 1512, "len_title": 103, "len_sents": 47320, "len_sents_1": 35459, "n_sents": 360, "n_sents_1": 263, "venue": "ICLR2025", "decision": "withdraw", "L_title": 55, "L_abs": 1538, "title": "Identifying Drivers of Predictive Aleatoric Uncertainty", "abs": "Explainability and uncertainty quantification are two pillars of trustable artificial intelligence. However, the reasoning behind uncertainty estimates is generally left unexplained. Identifying the drivers of uncertainty complements explanations of point predictions in recognizing model limitations and enhances trust in decisions and their communication. So far, explanations of uncertainties have been rarely studied. The few exceptions rely on Bayesian neural networks or technically intricate approaches, such as auxiliary generative models, thereby hindering their broad adoption. We propose a straightforward approach to explain predictive aleatoric uncertainties. We estimate uncertainty in regression as predictive variance by adapting a neural network with a Gaussian output distribution. Subsequently, we apply out-of-the-box explainers to the model's variance output. This approach can explain uncertainty influences more reliably than more complex published approaches, which we demonstrate in a synthetic setting with a known data-generating process. We further adapt multiple metrics from conventional XAI research to uncertainty explanations. We quantify our findings with a nuanced benchmark analysis that includes real-world datasets. Finally, we apply our approach to an age regression model and discover reasonable drivers of uncertainty. Overall, the proposed straightforward method explains uncertainty estimates with little modifications to the model architecture and decisively outperforms more intricate methods.", "keywords": [ "uncertainty", "explainability", "trustworthy ML", "probabilistic methods", "transfer learning" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 3 ] }, { "paper_id": "D9GoWJJxS5", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Yuan Gao", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Associate Professor" }, { "name": "Zujing Liu", "gender": "unknown", "institution": "Wuhan University", "country": "CN", "position": "MS student" }, { "name": "Bo Du", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Gui-song Xia", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Weizhong Zhang", "gender": "Male", "institution": "Fudan University", "country": "HK", "position": "Associate Professor" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 46, "n_ref": 84, "n_ref_all": 154, "n_fig": 9, "n_tab": 9, "L_tab": 3458, "n_element_tab": 518, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 505, "n_element_tab_1": 65, "formula_len_all": 806, "formula_len_all_1": 730, "len_all": 200456, "len_all_1": 65963, "len_abs": 1781, "len_title": 156, "len_sents": 50109, "len_sents_1": 31581, "n_sents": 359, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1788, "title": "Bypass Back-propagation: Optimization-based Structural Pruning for Large Language Models via Policy Gradient", "abs": "In contrast to moderate-size neural network pruning, structural weight pruning on the Large-Language Models (LLMs) imposes a novel challenge on the efficiency of the pruning algorithms, due to the heavy computation/memory demands of the LLMs. Recent efficient LLM pruning methods typically operate at the post-training phase without the expensive weight finetuning, however, their pruning criteria often rely on heuristically hand-crafted metrics, potentially leading to suboptimal performance. We instead propose a novel optimization-based structural pruning that learns the pruning masks in a probabilistic space directly by optimizing the loss of the pruned model. To preserve the efficiency, our method eliminates the back-propagation through the LLM per se during the optimization, requiring only the forward pass of the LLM. We achieve this by learning an underlying Bernoulli distribution to sample binary pruning masks, where we decouple the Bernoulli parameters from the LLM loss, thus facilitating an efficient optimization via a policy gradient estimator without back-propagation. As a result, our method is able to 1) operate at structural granularities of channels, heads, and layers, 2) support global and heterogeneous pruning (i.e., our method automatically determines different redundancy for different layers), and 3) optionally initialize with a metric-based method (for our Bernoulli distributions). Extensive experiments on LLaMA, LLaMA-2, LLaMA-3, Vicuna, and Mistral using the C4 and WikiText2 datasets demonstrate that our method operates for 2.7 hours with around 35GB memory for the 13B models on a single A100 GPU, and our pruned models outperform the state-of-the-arts w.r.t. both perplexity and the majority of various zero-shot tasks. Codes will be released.", "keywords": [ "Optimization-based Pruning", "Back-Propagation-Free", "Structural Pruning", "Large Language Models" ], "rating_list": [ 3, 6, 5, 6, 5 ], "soundness_list": [ 2, 3, 2, 3, 4 ], "presentation_list": [ 2, 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 4, 3 ], "confidence_list": [ 4, 3, 4, 4, 4 ] }, { "paper_id": "D9CRb1KZQc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Brian L. Price", "gender": "unknown", "institution": "Adobe Systems", "country": "US", "position": "Senior Research Scientist" }, { "name": "Daniel Aliaga", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "Jianming Zhang", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" }, { "name": "Liu He", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Scott Cohen", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" }, { "name": "Soo Ye Kim", "gender": "unknown", "institution": "Adobe Systems", "country": "US", "position": "Researcher" }, { "name": "Wei Xiong", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yizhi Song", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Zhe Lin", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Principal Researcher" }, { "name": "Zhifei Zhang", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Research Engineer" }, { "name": "HE Zhang", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 8, "n_ref_uni": 45, "n_ref": 103, "n_ref_all": 131, "n_fig": 14, "n_tab": 4, "L_tab": 1020, "n_element_tab": 77, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 341, "n_element_tab_1": 33, "formula_len_all": 316, "formula_len_all_1": 291, "len_all": 147013, "len_all_1": 55616, "len_abs": 652, "len_title": 129, "len_sents": 34632, "len_sents_1": 25232, "n_sents": 264, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1433, "title": "Refine-by-Align: Reference-Guided Artifacts Refinement through Semantic Alignment", "abs": "Personalized image generation has emerged from the recent advancements in generative models. However, these generated personalized images often suffer from localized artifacts such as incorrect logos, reducing fidelity and fine-grained identity details of the generated results. Furthermore, there is little prior work tackling this problem. To help improve these identity details in the personalized image generation, we introduce a new task: reference-guided artifacts refinement. We present Refine-by-Align, a first-of-its-kind model that employs a diffusion-based framework to address this challenge. Our model consists of two stages: Alignment Stage and Refinement Stage, which share weights of a unified neural network model. Given a generated image, a masked artifact region, and a reference image, the alignment stage identifies and extracts the corresponding regional features in the reference, which are then used by the refinement stage to fix the artifacts. Our model-agnostic pipeline requires no test-time tuning or optimization. It automatically enhances image fidelity and reference identity in the generated image, generalizing well to existing models on various tasks including but not limited to customization, generative compositing, view synthesis, and virtual try-on. Extensive experiments and comparisons demonstrate that our pipeline greatly pushes the boundary of fine details in the image synthesis models.", "keywords": [ "diffusion model; inpainting; generative artifacts; image editing; image synthesis; artifacts refinement" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "D8fk2fY0lh", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Difan Liu", "gender": "Male", "institution": "Adobe Systems", "country": "US", "position": "Research Scientist" }, { "name": "Jimei Yang", "gender": "unknown", "institution": "Runway", "country": "US", "position": "Researcher" }, { "name": "Jingye Chen", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Li Zhang", "gender": "Male", "institution": "Adobe", "country": "US", "position": "Staff Machine Learning Engineer" }, { "name": "Nanxuan Zhao", "gender": "Female", "institution": "Adobe Research", "country": "US", "position": "Researcher" }, { "name": "Qifeng Chen", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Zhaowen Wang", "gender": "Male", "institution": "Adobe Research", "country": "US", "position": "Research Scientist" } ], "n_formula": 0, "n_formula_1": 1, "n_ref_uni": 16, "n_ref": 111, "n_ref_all": 136, "n_fig": 13, "n_tab": 3, "L_tab": 766, "n_element_tab": 57, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 506, "n_element_tab_1": 24, "formula_len_all": 0, "formula_len_all_1": 23, "len_all": 141279, "len_all_1": 59341, "len_abs": 0, "len_title": 0, "len_sents": 40171, "len_sents_1": 29135, "n_sents": 318, "n_sents_1": 218, "venue": "ICLR2025", "decision": "withdraw", "L_title": 36, "L_abs": 1964, "title": "Crafting Layered Designs from Pixels", "abs": "Graphic designs play a vital role in communicating ideas, values, and messages.\nDuring the design process, designers typically organize their work into layers of\ntext, objects, and backgrounds to facilitate easier editing and customization. However, creating design in such a format requires significant effort and expertise. On\nthe other hand, with the advancement of GenAI technologies, high quality graphic\ndesigns created in pixel format have become more popular and accessible, while\nwith the inherent limitation of editability. Despite this limitation, we recognize\nthe significant reference value of these non-layered designs, as human designers\noften derive inspiration from these images to determine layouts or text styles. Motivated by this observation, we propose Accordion, a graphic design generation\nframework built around a vision language model playing distinct roles in three\nkey stages: (1) reference creation, (2) design planning, and (3) layer generation.\nBy using the reference image as global design guidance, distinct from existing\nmethods, our approach ensures that elements within the design are visually harmonious. Moreover, through this three-stage framework, Accordion can benefit\nfrom an unlimited supply of AI-generated references. The stage-wise design of\nour framework allows for flexible configuration and various applications, such as\nstarting from user provided references directly with the later two stages. Additionally, it leverages multiple vision experts such as SAM and element removal\nmodels to facilitate the creation of editable graphic layers. Experimental results\nshow that Accordion generates favourable results on the DesignIntention benchmark, including tasks such as text-to-template, adding text on background, and\ntext de-rendering. Furthermore, we fully explore the potential of Accordion to\nfacilitate the creation of design variations, validating its versatility and flexibility\nin the whole design workflow.", "keywords": [ "Graphic Design", "Layered Image Generation" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 3, 2, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "D7PQ54l5Q1", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Andrew Lizarraga", "gender": "unknown", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Haoxin Zheng", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Ruiqi Gao", "gender": "Female", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Yasi Zhang", "gender": "Not Specified", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Yaxuan Zhu", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Zehao Dou", "gender": "Male", "institution": "Yale University", "country": "", "position": "PhD student" }, { "name": "Yingnian Wu", "gender": "Male", "institution": "UCLA", "country": "US", "position": "Full Professor" } ], "n_formula": 28, "n_formula_1": 12, "n_ref_uni": 26, "n_ref": 71, "n_ref_all": 91, "n_fig": 6, "n_tab": 5, "L_tab": 2126, "n_element_tab": 171, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 461, "n_element_tab_1": 27, "formula_len_all": 1710, "formula_len_all_1": 914, "len_all": 135718, "len_all_1": 49099, "len_abs": 1286, "len_title": 119, "len_sents": 29279, "len_sents_1": 21050, "n_sents": 240, "n_sents_1": 169, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1291, "title": "Think Twice Before You Act: Improving Inverse Problem Solving With MCMC", "abs": "Recent studies demonstrate that diffusion models can serve as a strong prior for solving inverse problems. A prominent example is Diffusion Posterior Sampling (DPS), which approximates the posterior distribution of data given the measure using Tweedie's formula. Despite the merits of being versatile in solving various inverse problems without re-training, the performance of DPS is hindered by the fact that this posterior approximation can be inaccurate especially for high noise levels. Therefore, we propose Diffusion Posterior MCMC (DPMC), a novel inference algorithm based on Annealed MCMC to solve inverse problems with pretrained diffusion models. We define a series of intermediate distributions inspired by the approximated conditional distributions used by DPS. Through annealed MCMC sampling, we encourage the samples to follow each intermediate distribution more closely before moving to the next distribution at a lower noise level, and therefore reduce the accumulated error along the path. We test our algorithm in various inverse problems, including super resolution, Gaussian deblurring, motion deblurring, inpainting, and phase retrieval. Our algorithm outperforms DPS with less number of evaluations across nearly all tasks, and is competitive among existing approaches.", "keywords": [ "Inverse problem", "diffusion", "MCMC sampling", "generative model" ], "rating_list": [ 6, 5, 5, 3 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "D756s2YQ6b", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 5, "author_info_list": [ { "name": "Andi Han", "gender": "Male", "institution": "RIKEN AIP", "country": "JP", "position": "Postdoc" }, { "name": "Dai Shi", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Researcher" }, { "name": "Junbin Gao", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Full Professor" }, { "name": "Lequan Lin", "gender": "Female", "institution": "University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Zhiyong Wang", "gender": "Male", "institution": "The University of Sydney", "country": "AU", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 7, "n_ref_uni": 64, "n_ref": 157, "n_ref_all": 198, "n_fig": 16, "n_tab": 13, "L_tab": 2190, "n_element_tab": 373, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 2044, "n_element_tab_1": 203, "formula_len_all": 1623, "formula_len_all_1": 365, "len_all": 240689, "len_all_1": 69064, "len_abs": 1273, "len_title": 133, "len_sents": 75137, "len_sents_1": 32394, "n_sents": 578, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1279, "title": "Diffusing to the Top: Boost Graph Neural Networks with Minimal Hyperparameter Tuning", "abs": "Graph Neural Networks (GNNs) are proficient in graph representation learning and achieve promising performance on versatile tasks such as node classification and link prediction.\nUsually, a comprehensive hyperparameter tuning is essential for fully unlocking GNN's top performance, especially for complicated tasks such as node classification on large graphs and long-range graphs. This is usually associated with high computational and time costs and careful design of appropriate search spaces. \nThis work introduces a graph-conditioned latent diffusion framework (GNN-Diff) to generate high-performing GNNs based on the model checkpoints of sub-optimal hyperparameters selected by a light-tuning coarse search. We validate our method through 166 experiments across four graph tasks: node classification on small, large, and long-range graphs, as well as link prediction. Our experiments involve 10 classic and state-of-the-art target models and 20 publicly available datasets. The results consistently demonstrate that GNN-Diff: (1) boosts the performance of GNNs with efficient hyperparameter tuning; and (2) presents high stability and generalizability on unseen data across multiple generation runs. The code is available at https://anonymous.4open.science/r/GNN-Diff-1AD3.", "keywords": [ "graph neural networks", "generative diffusion models", "network generation", "hyperparameter tuning", "node classification", "link prediction" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 3, 2, 2 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 1, 3, 4 ] }, { "paper_id": "D6zn6ozJs7", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 9, "author_info_list": [ { "name": "Huaibo Huang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Linzhi Huang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "PEI PEI LI", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Assistant Professor" }, { "name": "Shuhan Xia", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Weihong Deng", "gender": "Male", "institution": "Beijing University of Post and Telecommunication", "country": "CN", "position": "Full Professor" }, { "name": "Xing Cui", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Xuannan Liu", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Zekun Li", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Zhaofeng He", "gender": "Male", "institution": "Beijing University of Post and Telecommunication", "country": "CN", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 40, "n_ref": 98, "n_ref_all": 136, "n_fig": 15, "n_tab": 9, "L_tab": 5107, "n_element_tab": 666, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 3330, "n_element_tab_1": 225, "formula_len_all": 31, "formula_len_all_1": 31, "len_all": 196712, "len_all_1": 64301, "len_abs": 1707, "len_title": 0, "len_sents": 58005, "len_sents_1": 25996, "n_sents": 503, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1321, "title": "MMFakeBench: A Mixed-Source Multimodal Misinformation Detection Benchmark for LVLMs", "abs": "Current multimodal misinformation detection (MMD) methods often assume a single source and type of forgery for each sample, which is insufficient for real-world scenarios where multiple forgery sources coexist. The lack of a benchmark for mixed-source misinformation has hindered progress in this field. To address this, we introduce MMFakeBench, the first comprehensive benchmark for mixed-source MMD. MMFakeBench includes 3 critical sources: textual veracity distortion, visual veracity distortion, and cross-modal consistency distortion, along with 12 sub-categories of misinformation forgery types. We further conduct an extensive evaluation of 6 prevalent detection methods and 15 large vision-language models (LVLMs) on MMFakeBench under a zero-shot setting. The results indicate that current methods struggle under this challenging and realistic mixed-source MMD setting. Additionally, we propose an innovative unified framework, which integrates rationales, actions, and tool-use capabilities of LVLM agents, significantly enhancing accuracy and generalization. We believe this study will catalyze future research into more realistic mixed-source multimodal misinformation and provide a fair evaluation of misinformation detection methods. Code and a portion of the data are accessible in supplementary materials.", "keywords": [ "multimodal misinformation detection", "large vision language models" ], "rating_list": [ 8, 5, 6, 6, 8 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 4, 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 4, 4 ] }, { "paper_id": "D5v491uCzm", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Felipe Maia Polo", "gender": "Male", "institution": "University of Michigan - Ann Arbor", "country": "", "position": "PhD student" }, { "name": "Leshem Choshen", "gender": "Not Specified", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Mikhail Yurochkin", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Research Staff Member" }, { "name": "Seamus Somerstep", "gender": "unknown", "institution": "University of Michigan - Ann Arbor", "country": "US", "position": "PhD student" }, { "name": "Yuekai Sun", "gender": "unknown", "institution": "University of Michigan", "country": "US", "position": "Associate Professor" } ], "n_formula": 24, "n_formula_1": 10, "n_ref_uni": 41, "n_ref": 81, "n_ref_all": 118, "n_fig": 12, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1966, "n_element_tab_1": 29, "formula_len_all": 2295, "formula_len_all_1": 630, "len_all": 162961, "len_all_1": 70708, "len_abs": 1364, "len_title": 137, "len_sents": 49234, "len_sents_1": 33851, "n_sents": 349, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1369, "title": "Sloth: scaling laws for LLM skills to predict multi-benchmark performance across families", "abs": "Scaling laws for large language models (LLMs) predict model performance based on parameters like size and training data. However, differences in training configurations and data processing across model families lead to significant variations in benchmark performance, making it difficult for a single scaling law to generalize across all LLMs. On the other hand, training family-specific scaling laws requires training models of varying sizes for every family. In this work, we propose Skills Scaling Laws (SSLaws, pronounced as Sloth), a novel scaling law that leverages publicly available benchmark data and assumes LLM performance is driven by low-dimensional latent skills, such as reasoning and instruction following. These latent skills are influenced by computational resources like model size and training tokens but with varying efficiencies across model families. Sloth exploits correlations across benchmarks to provide more accurate and interpretable predictions while alleviating the need to train multiple LLMs per family. We present both theoretical results on parameter identification and empirical evaluations on 12 prominent benchmarks, from Open LLM Leaderboard v1/v2, demonstrating that Sloth predicts LLM performance efficiently and offers insights into scaling behaviors for downstream tasks such as coding and emotional intelligence applications.", "keywords": [ "scaling law", "LLM", "benchmark", "skill" ], "rating_list": [ 5, 6, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "D5X6nPGFUY", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 4, "author_info_list": [ { "name": "Sangdoo Yun", "gender": "Male", "institution": "NAVER", "country": "KR", "position": "Research Scientist" }, { "name": "Sanghyuk Chun", "gender": "Male", "institution": "NAVER AI Lab", "country": "KR", "position": "Lead research scientist" }, { "name": "Song Park", "gender": "Female", "institution": "NAVER", "country": "KR", "position": "Researcher" }, { "name": "Wonjae Kim", "gender": "Male", "institution": "TwelveLabs", "country": "KR", "position": "Lead Research Scientist" } ], "n_formula": 36, "n_formula_1": 13, "n_ref_uni": 46, "n_ref": 105, "n_ref_all": 157, "n_fig": 11, "n_tab": 12, "L_tab": 5805, "n_element_tab": 419, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 285, "n_element_tab_1": 42, "formula_len_all": 2464, "formula_len_all_1": 954, "len_all": 224154, "len_all_1": 73777, "len_abs": 1323, "len_title": 89, "len_sents": 73062, "len_sents_1": 34193, "n_sents": 587, "n_sents_1": 282, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 1325, "title": "Probabilistic Language-Image Pre-Training", "abs": "Vision-language models (VLMs) embed aligned image-text pairs into a joint space but often rely on deterministic embeddings, assuming a one-to-one correspondence between images and texts. This oversimplifies real-world relationships, which are inherently many-to-many, with multiple captions describing a single image and vice versa. We introduce Probabilistic Language-Image Pre-training (ProLIP), the first probabilistic VLM pre-trained on a billion-scale image-text dataset using only probabilistic objectives, achieving a strong zero-shot capability (e.g., 74.6% ImageNet zero-shot accuracy with ViT-B/16). ProLIP efficiently estimates uncertainty by an \"uncertainty token\" without extra parameters. We also introduce a novel inclusion loss that enforces distributional inclusion relationships between image-text pairs and between original and masked inputs. Experiments demonstrate that, by leveraging uncertainty estimates, ProLIP benefits downstream tasks and aligns with intuitive notions of uncertainty, e.g, shorter texts being more uncertain and more general inputs including specific ones. Utilizing text uncertainties, we further improve ImageNet accuracy from 74.6% to 75.8% (under a few-shot setting), supporting the practical advantages of our probabilistic approach. The code will be available upon acceptance", "keywords": [ "vision-langauge-pretraining", "probabilistic-embeddings" ], "rating_list": [ 6, 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 3, 3, 2 ], "confidence_list": [ 4, 2, 3, 5, 3 ] }, { "paper_id": "D4xztKoz0Y", "primary_area": "applications to robotics, autonomy, planning", "n_author": 5, "author_info_list": [ { "name": "Huan Ren", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Shifeng Zhang", "gender": "Male", "institution": "Sangfor Technologies Inc.", "country": "CN", "position": "Researcher" }, { "name": "Tianzhu Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Wenfei Yang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Postdoc" }, { "name": "Xiang Liu", "gender": "Male", "institution": "Dongguan University of Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 22, "n_formula_1": 11, "n_ref_uni": 38, "n_ref": 110, "n_ref_all": 142, "n_fig": 8, "n_tab": 15, "L_tab": 8229, "n_element_tab": 837, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1521, "n_element_tab_1": 198, "formula_len_all": 1081, "formula_len_all_1": 461, "len_all": 173982, "len_all_1": 67817, "len_abs": 1668, "len_title": 161, "len_sents": 48824, "len_sents_1": 31203, "n_sents": 330, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 113, "L_abs": 1573, "title": "Learning Shape-Independent Transformation via Spherical Representations for Category-Level Object Pose Estimation", "abs": "Category-level object pose estimation aims to determine the pose and size of novel objects in specific categories. Existing correspondence-based approaches typically adopt point-based representations to establish the correspondences between primitive observed points and normalized object coordinates. However, due to the inherent shape-dependence of canonical coordinates, these methods suffer from semantic incoherence across diverse object shapes. To resolve this issue, we innovatively leverage the sphere as a shared proxy shape of objects to learn shape-independent transformation via spherical representations. Based on this insight, we introduce a novel architecture called SpherePose, which yields precise correspondence prediction through three core designs. Firstly, We endow the point-wise feature extraction with $\\mathrm{SO(3)$-invariance, which facilitates robust mapping between camera coordinate space and object coordinate space regardless of rotation transformation. Secondly, the spherical attention mechanism is designed to propagate and integrate features among spherical anchors from a comprehensive perspective, thus mitigating the interference of noise and incomplete point cloud. Lastly, a hyperbolic correspondence loss function is designed to distinguish subtle distinctions, which can promote the precision of correspondence prediction. Experimental results on CAMERA25, REAL275 and HouseCat6D benchmarks demonstrate the superior performance of our method, verifying the effectiveness of spherical representations and architectural innovations.", "keywords": [ "category-level object pose estimation", "spherical representations", "shape-independence", "correspondence prediction" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "D4sQzdMvcG", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Jian Cheng", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Liang Chen", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Mingxin Guo", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Qianpeng Li", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Engineer" }, { "name": "Yaoyao li", "gender": "unknown", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "CN", "position": "Researcher" } ], "n_formula": 79, "n_formula_1": 23, "n_ref_uni": 48, "n_ref": 113, "n_ref_all": 130, "n_fig": 4, "n_tab": 10, "L_tab": 4439, "n_element_tab": 284, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 2972, "n_element_tab_1": 141, "formula_len_all": 4587, "formula_len_all_1": 1316, "len_all": 190091, "len_all_1": 61239, "len_abs": 1678, "len_title": 124, "len_sents": 45261, "len_sents_1": 25424, "n_sents": 354, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1685, "title": "QAC:Quantization-Aware Conversion for Mixed-Timestep Spiking Neural Networks", "abs": "Spiking Neural Networks (SNNs) have recently garnered widespread attention due to their high computational efficiency and low energy consumption, possessing significant potential for further research. Currently, SNN algorithms are primarily categorized into two types: one involves the direct training of SNNs using surrogate gradients, and the other is based on the mathematical equivalence between ANNs and SNNs for conversion. However, both methods overlook the exploration of mixed-timestep SNNs, where different layers in the network operate with different timesteps. This is because surrogate gradient methods struggle to compute gradients related to timestep, while ANN-to-SNN conversions typically use fixed timesteps, limiting the potential performance improvements of SNNs. In this paper, we propose a Quantization-Aware Conversion (QAC) algorithm that reveals a profound theoretical insight: the power of the quantization bit-width in ANN activations is equivalent to the timesteps in SNNs with soft reset. This finding uncovers the intrinsic nature of SNNs, demonstrating that they act as activation quantizers\u2014transforming multi-bit activation features into single-bit activations distributed over multiple timesteps. Based on this insight, we propose a mixed-precision quantization-based conversion algorithm from ANNs to mixed-timestep SNNs, which significantly reduces the number of timesteps required during inference and improves accuracy. Additionally, we introduce a calibration method for initial membrane potential and thresholds. Experimental results on CIFAR-10, CIFAR-100, and ImageNet demonstrate that our method significantly outperforms previous approaches.", "keywords": [ "Spiking Neural Networks", "Quantization", "ANN-SNN Conversion" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "D48jvLN45W", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Aditya Sanghi", "gender": "Male", "institution": "Autodesk", "country": "CA", "position": "Researcher" }, { "name": "Aliasghar Khani", "gender": "Male", "institution": "Autodesk", "country": "CA", "position": "Researcher" }, { "name": "Arianna Rampini", "gender": "Female", "institution": "Autodesk", "country": "IT", "position": "Researcher" }, { "name": "Derek Cheung", "gender": "unknown", "institution": "Autodesk", "country": "CA", "position": "Researcher" }, { "name": "Hooman Shayani", "gender": "Male", "institution": "Department of Computer Science, University College London, University of London", "country": "", "position": "Assistant Professor" }, { "name": "Kamal Rahimi Malekshan", "gender": "Not Specified", "institution": "University of Waterloo", "country": "", "position": "Postdoc" }, { "name": "Kanika Madan", "gender": "unknown", "institution": "MILA, University of Montreal", "country": "", "position": "PhD student" }, { "name": "Pradyumna Reddy", "gender": "Male", "institution": "Autodesk", "country": "GB", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 4, "n_ref_uni": 70, "n_ref": 136, "n_ref_all": 144, "n_fig": 3, "n_tab": 4, "L_tab": 2206, "n_element_tab": 169, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 1735, "n_element_tab_1": 97, "formula_len_all": 353, "formula_len_all_1": 195, "len_all": 213930, "len_all_1": 66575, "len_abs": 1328, "len_title": 150, "len_sents": 35892, "len_sents_1": 31577, "n_sents": 242, "n_sents_1": 201, "venue": "ICLR2025", "decision": "withdraw", "L_title": 101, "L_abs": 1345, "title": "Wavelet Latent Diffusion (WaLa): Billion-Parameter 3D Generative Model with Compact Wavelet Encodings", "abs": "Large-scale 3D generative models require substantial computational resources yet often fall short in capturing fine details and complex geometries at high resolutions. We attribute this limitation to the inefficiency of current representations, which lack the compactness required for generative networks to model effectively. To address this, we introduce Wavelet Latent Diffusion (WaLa), a novel approach that encodes 3D shapes into a wavelet-based, compact latent encodings. Specifically, we compress a $256^3$ signed distance field into a $12^3 \\times 4$ latent grid, achieving an impressive 2,427\u00d7 compression ratio with minimal loss of detail. This high level of compression allows our method to efficiently train large-scale generative networks without increasing inference time. Our models, both conditional and unconditional, contain approximately one billion parameters and successfully generate high-quality 3D shapes at $256^3$ resolution. Moreover, WaLa offers rapid inference, producing shapes within 2\u20134 seconds depending on the condition, despite the model\u2019s scale. We demonstrate state-of-the-art performance across multiple datasets, with significant improvements in generation quality, diversity, and computational efficiency. Upon acceptance, we will open-source the code and model weights for public use and reproducibility.", "keywords": [ "3D generative modelling", "Diffusion models", "wavelet encoding" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 1, 4 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "D3vD7ZFIor", "primary_area": "optimization", "n_author": 9, "author_info_list": [ { "name": "Bo Long", "gender": "Male", "institution": "Meta", "country": "US", "position": "Principal Researcher" }, { "name": "Hui Yuan", "gender": "Female", "institution": "Princeton University", "country": "", "position": "PhD student" }, { "name": "Mengdi Wang", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Associate Professor" }, { "name": "Rong Jin", "gender": "Male", "institution": "Alibaba Group", "country": "", "position": "Principal Researcher" }, { "name": "Shuang Yang", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Research Scientist" }, { "name": "Weilin Cong", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Yan Xie", "gender": "unknown", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Zhigang Hua", "gender": "Male", "institution": "Meta", "country": "US", "position": "ML engineer" }, { "name": "Zihao Li", "gender": "Male", "institution": "Princeton University", "country": "", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 18, "n_ref_uni": 37, "n_ref": 66, "n_ref_all": 90, "n_fig": 6, "n_tab": 5, "L_tab": 725, "n_element_tab": 106, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 159, "n_element_tab_1": 25, "formula_len_all": 1537, "formula_len_all_1": 1669, "len_all": 114641, "len_all_1": 62397, "len_abs": 1391, "len_title": 152, "len_sents": 32843, "len_sents_1": 26495, "n_sents": 240, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 102, "L_abs": 1395, "title": "GuideCO: Training Objective-Guided Diffusion Solver with Imperfect Data for Combinatorial Optimization", "abs": "Combinatorial optimization (CO) problems have widespread applications in science and engineering but they present significant computational challenges. Recent advancements in generative models, particularly diffusion models, have shown promise in bypassing traditional optimization solvers by directly generating near-optimal solutions. However, we observe an exponential scaling law between the optimality gap and the amount of training data needed for training diffusion-based solvers. Notably, the performance of existing diffusion solvers relies on both quantity and quality of training data: they perform well with abundant high quality training data labeled by exact or near-optimal solvers, while suffering when high-quality labels are scarce or unavailable. To address the challenge, we propose GuideCO, an objective-guided diffusion solver for combinatorial optimization, which can be trained on imperfectly labelled datasets. GuideCO is a two-stage generate-then-decode framework, featuring an objective-guided diffusion model that is further reinforced by classifier-free guidance for generating high-quality solutions on any given problem instance. Experiments demonstrate the improvements of GuideCO against baselines when trained on imperfect data, in a range of combinatorial optimization benchmark tasks such as TSP (Traveling Salesman Problem) and MIS (Maximum Independent Set).", "keywords": [ "combinatorial optimization", "diffusion model", "guidance" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 1, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "D3iJmVAmT7", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 2, "author_info_list": [ { "name": "Amir Barati Farimani", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Cooper Delaney Lorsung", "gender": "unknown", "institution": "CMU, Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 18, "n_ref": 39, "n_ref_all": 53, "n_fig": 7, "n_tab": 6, "L_tab": 2810, "n_element_tab": 174, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1046, "n_element_tab_1": 83, "formula_len_all": 454, "formula_len_all_1": 454, "len_all": 99368, "len_all_1": 56508, "len_abs": 1063, "len_title": 123, "len_sents": 30944, "len_sents_1": 26026, "n_sents": 233, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1068, "title": "Explain Like I'm Five: Using LLMs to Improve PDE Surrogate Models with Text", "abs": "Solving Partial Differential Equations (PDEs) is ubiquitous in science and engineering. Computational complexity and difficulty in writing numerical solvers has motivated the development of machine learning techniques to generate solutions quickly. Many existing methods are purely data driven, relying solely on numerical solution fields, rather than known system information such as boundary conditions and governing equations. However, the recent rise in popularity of Large Language Models (LLMs) has enabled easy integration of text in multimodal machine learning models. In this work, we use pretrained LLMs to integrate various amounts known system information into PDE learning. Our multimodal approach significantly outperforms our baseline model, FactFormer, in both next-step prediction and autoregressive rollout performance on the 2D Heat, Burgers, Navier-Stokes, and Shallow Water equations. Further analysis shows that pretrained LLMs provide highly structured latent space that is consistent with the amount of system information provided through text.", "keywords": [ "Multimodal", "Partial Differential Equation", "Neural Operator" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 2, 1, 1, 2 ], "confidence_list": [ 4, 5, 3, 2 ] }, { "paper_id": "D2hhkU5O48", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Aaditya Ramdas", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Associate Professor" }, { "name": "Atalanti A. Mastakouri", "gender": "unknown", "institution": "Amazon Development Center Germany", "country": "", "position": "Researcher" }, { "name": "Elke Kirschbaum", "gender": "Female", "institution": "Amazon", "country": "", "position": "Applied Scientist" }, { "name": "Putra Manggala", "gender": "unknown", "institution": "University of Amsterdam", "country": "", "position": "PhD student" }, { "name": "Shiva Kasiviswanathan", "gender": "Male", "institution": "Amazon", "country": "", "position": "Research Scientist" } ], "n_formula": 29, "n_formula_1": 21, "n_ref_uni": 39, "n_ref": 89, "n_ref_all": 137, "n_fig": 4, "n_tab": 6, "L_tab": 3072, "n_element_tab": 237, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2640, "formula_len_all_1": 1873, "len_all": 172466, "len_all_1": 70515, "len_abs": 1583, "len_title": 115, "len_sents": 55714, "len_sents_1": 33775, "n_sents": 465, "n_sents_1": 261, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 969, "title": "$\\beta$-calibration of Language Model Confidence Scores for Generative QA", "abs": "To use generative question-and-answering (QA) systems for decision-making and in any critical application, these systems need to provide well-calibrated confidence scores that reflect the correctness of their answers. Existing calibration methods aim to ensure that the confidence score is *on average* indicative of the likelihood that the answer is correct. We argue, however, that this standard (average-case) notion of calibration is difficult to interpret for decision-making in generative QA. To address this, we generalize the standard notion of average calibration and introduce $\\beta$-calibration, which ensures calibration holds across different question-and-answer groups. We then propose discretized posthoc calibration schemes for achieving $\\beta$-calibration. We establish distribution-free guarantees on the performance of this method and validate our method on confidence scores returned by elicitation prompts across multiple QA benchmarks and LLMs.", "keywords": [ "Calibration", "Language Models" ], "rating_list": [ 8, 6, 5, 8 ], "soundness_list": [ 4, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "D2as3jDmRA", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "Songhua Liu", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Weihao Yu", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zhenxiong Tan", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" }, { "name": "Xinchao Wang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 29, "n_formula_1": 21, "n_ref_uni": 40, "n_ref": 89, "n_ref_all": 134, "n_fig": 19, "n_tab": 8, "L_tab": 1953, "n_element_tab": 260, "n_fig_1": 10, "n_tab_1": 3, "L_tab_1": 1092, "n_element_tab_1": 113, "formula_len_all": 2664, "formula_len_all_1": 1586, "len_all": 185560, "len_all_1": 60212, "len_abs": 429, "len_title": 0, "len_sents": 42902, "len_sents_1": 25812, "n_sents": 322, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 1742, "title": "LinFusion: 1 GPU, 1 Minute, 16K Image", "abs": "Modern diffusion models, particularly those utilizing a Transformer-based UNet for denoising, rely heavily on self-attention operations to manage complex spatial relationships, thus achieving impressive generation performance. However, this existing paradigm faces significant challenges in generating high-resolution visual content due to its quadratic time and memory complexity with respect to the number of spatial tokens. To address this limitation, we aim at a novel linear attention mechanism as an alternative in this paper. \nSpecifically, we begin our exploration from recently introduced models with linear complexity, e.g., Mamba2, RWKV6, Gated Linear Attention, etc, and identify two key features\u2014attention normalization and non-causal inference\u2014that enhance high-resolution visual generation performance. Building on these insights, we introduce a generalized linear attention paradigm, which serves as a low-rank approximation of a wide spectrum of popular linear token mixers. To save the training cost and better leverage pre-trained models, we initialize our models and distill the knowledge from pre-trained StableDiffusion (SD). We find that the distilled model, termed LinFusion, achieves performance on par with or superior to the original SD after only modest training, while significantly reducing time and memory complexity. Extensive experiments on SD-v1.5, SD-v2.1, and SD-XL demonstrate that LinFusion enables satisfactory and efficient zero-shot cross-resolution generation, accommodating ultra-resolution images like 16K on a single GPU. Moreover, it is highly compatible with pre-trained SD components and pipelines, such as ControlNet, IP-Adapter, DemoFusion, DistriFusion, etc, requiring no adaptation efforts.", "keywords": [ "Linear Attention", "Diffusion Models", "Image Generation" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "D2Vz4drFA6", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 6, "author_info_list": [ { "name": "Dan Li", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "Full Professor" }, { "name": "Dayu Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Undergrad student" }, { "name": "Tong Yang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Associate Professor" }, { "name": "Wang Li", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "MS student" }, { "name": "Yanshu Wang", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Postdoc" }, { "name": "Zhaoqian YAO", "gender": "Female", "institution": "Peking University", "country": "CN", "position": "Intern" } ], "n_formula": 41, "n_formula_1": 15, "n_ref_uni": 30, "n_ref": 42, "n_ref_all": 53, "n_fig": 6, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2191, "formula_len_all_1": 709, "len_all": 112535, "len_all_1": 58009, "len_abs": 1634, "len_title": 111, "len_sents": 34347, "len_sents_1": 26537, "n_sents": 278, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 111, "L_abs": 1704, "title": "HyperChr: Quantization of Heterogeneously Distributed Matrices through Distribution-Aware Subspace Partitioning", "abs": "Matrix quantization is crucial for reducing the memory footprint of matrices across various applications, including large-scale machine learning models and data compression. We have observed that matrices in different application domains exhibit heterogeneity in the distribution across columns. Leveraging this characteristic, we introduce \\textit{HyperChr}, a novel matrix quantization algorithm tailored for heterogeneous data distributions prevalent across different matrix columns. Unlike traditional quantization methods, \\textit{HyperChr} capitalizes on the heterogeneous distribution characteristics of each column to optimally partition high-dimensional subspaces and perform compression within each subspace. This technique enhances the compression effectiveness by grouping vectors with similar distribution ranges, enabling more precise quantization. Moreover, \\textit{HyperChr} dynamically adjusts the number of centroids in each subspace based on the specific data distribution traits, optimizing both storage efficiency and data fidelity.\n\nWe evaluate \\textit{HyperChr}'s performance on diverse datasets, demonstrating its superiority in reducing quantization errors compared to existing methods. Our results show that \\textit{HyperChr} exhibits significant improvements at lower compression ratios ($\\theta = 2-8$), reducing MAE by an average of 55.3\\% and MSE by 75.3\\% compared to PQ. However, at higher compression ratios ($\\theta = 10-16$), the improvements are more moderate, with an average reduction of 14.9\\% in MAE and 25.9\\% in MSE compared to PQ. In addition, our algorithm reduces the average dequantization time by 62.9\\%, which is crucial for large language model inference.", "keywords": [ "matrix quantization; LLMs; heterogeneous distribution; Product quantization" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 1, 1, 2 ], "presentation_list": [ 1, 1, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "D2EdWRWEQo", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Ege Erdogan", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "MS student" }, { "name": "Leon Klein", "gender": "unknown", "institution": "Freie Universit\u00e4t Berlin", "country": "", "position": "PhD student" }, { "name": "Mika Rebensburg", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "", "position": "MS student" }, { "name": "Radoslav Ralev", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "MS student" }, { "name": "C Marquet", "gender": "unknown", "institution": "Technical University Munich", "country": "", "position": "PhD student" }, { "name": "Hannes St\u00e4rk", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 21, "n_ref_uni": 23, "n_ref": 39, "n_ref_all": 58, "n_fig": 8, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1395, "formula_len_all_1": 1209, "len_all": 98611, "len_all_1": 56871, "len_abs": 1582, "len_title": 116, "len_sents": 34807, "len_sents_1": 27506, "n_sents": 216, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 1486, "title": "FreeFlow: Latent Flow Matching for Free Energy Difference Estimation", "abs": "Estimating free energy differences between molecular systems is fundamental for understanding molecular interactions and accelerating drug discovery. Current techniques use molecular dynamics to sample the Boltzmann distributions of the two systems and of several intermediate \"alchemical\" distributions that interpolate between them. From the resulting ensembles, free energy differences can be estimated by averaging importance weight analogs for multiple distributions. Instead of time-intensive simulations of intermediate alchemical systems, we learn a fast-to-train flow to bridge the two systems of interest. After training, we obtain free energy differences by integrating the flow's instantaneous change of variables when transporting samples between the two distributions. To map between molecular systems with different numbers of atoms, we replace the previous solutions of simulating auxiliary \"dummy atoms\" by additionally training two autoencoders that project the systems into a same-dimensional latent space in which our flow operates. A generalized change of variables formula for trans-dimensional mappings allows us to employ the dimensionality collapsing and expanding autoencoders in our free energy estimation pipeline. We validate our approach on systems of increasing complexity: mapping between Gaussians, between subspaces of alanine dipeptide, and between pharmaceutically relevant ligands in solvent. All results show strong agreement with reference values.", "keywords": [ "free energy", "flow matching", "free energy perturbation", "computational biology" ], "rating_list": [ 8, 3, 6, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "D23JcXiUwf", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 3, "author_info_list": [ { "name": "Arvind V. Mahankali", "gender": "unknown", "institution": "Computer Science Department, Stanford University", "country": "", "position": "PhD student" }, { "name": "Kefan Dong", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Tengyu Ma", "gender": "Male", "institution": "Stanford University", "country": "", "position": "Stanford University" } ], "n_formula": 6, "n_formula_1": 5, "n_ref_uni": 27, "n_ref": 57, "n_ref_all": 86, "n_fig": 6, "n_tab": 2, "L_tab": 284, "n_element_tab": 29, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 284, "n_element_tab_1": 29, "formula_len_all": 642, "formula_len_all_1": 612, "len_all": 120400, "len_all_1": 62209, "len_abs": 1267, "len_title": 123, "len_sents": 40198, "len_sents_1": 30754, "n_sents": 288, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1266, "title": "Formal Theorem Proving by Rewarding LLMs to Decompose Proofs Hierarchically", "abs": "Mathematical theorem proving is an important testbed for large language models\u2019 deep and abstract reasoning capability. This paper focuses on improving LLMs\u2019 ability to write proofs in formal languages that permit automated proof verification/ evaluation. Most previous results provide human-written lemmas to the theorem prover, which is an arguably oversimplified setting that does not sufficiently test the provers' planning and decomposition capabilities. Instead, we work in a more natural setup where the lemmas that are directly relevant to the theorem are not given to the theorem prover at test time. We design an RL-based training algorithm that encourages the model to decompose a theorem into lemmas, prove the lemmas, and then prove the theorem by using the lemmas. Our reward mechanism is inspired by how mathematicians train themselves: even if a theorem is too challenging to be proved by the current model, a reward is still given to the model for any correct and novel lemmas that are proposed and proved in this process. During training, our model proves 37.7% lemmas that are not in the training dataset. When tested on a set of holdout theorems, our model improves the pass rate from 40.8% to 45.5% compared with the supervised fine-tuned model.", "keywords": [ "formal theorem proving", "large language models", "reinforcement learning" ], "rating_list": [ 3, 3, 8, 8 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 1, 2, 3, 4 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "D1gKsJagiq", "primary_area": "applications to robotics, autonomy, planning", "n_author": 7, "author_info_list": [ { "name": "Barbara Caputo", "gender": "Female", "institution": "Politecnico di Torino", "country": "IT", "position": "Full Professor" }, { "name": "Carlo Masone", "gender": "Male", "institution": "Polytechnic Institute of Turin", "country": "IT", "position": "Assistant Professor" }, { "name": "Iuliana Georgescu", "gender": "Female", "institution": "Helmholtz Zentrum M\u00fcnchen", "country": "DE", "position": "Postdoc" }, { "name": "Marco Ciccone", "gender": "Male", "institution": "Vector Institute", "country": "CA", "position": "Postdoc" }, { "name": "Massimiliano Mancini", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Assistant Professor" }, { "name": "Shyam Nandan Rai", "gender": "Male", "institution": "Politecnico di Torino", "country": "IT", "position": "PhD student" }, { "name": "Zeynep Akata", "gender": "Female", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" } ], "n_formula": 15, "n_formula_1": 9, "n_ref_uni": 44, "n_ref": 131, "n_ref_all": 156, "n_fig": 9, "n_tab": 6, "L_tab": 1017, "n_element_tab": 199, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 1288, "n_element_tab_1": 15, "formula_len_all": 1096, "formula_len_all_1": 500, "len_all": 155648, "len_all_1": 57313, "len_abs": 1228, "len_title": 93, "len_sents": 37621, "len_sents_1": 24327, "n_sents": 277, "n_sents_1": 161, "venue": "ICLR2025", "decision": "Reject", "L_title": 45, "L_abs": 1111, "title": "Dual-Stream Adapters for Anomaly Segmentation", "abs": "Anomaly segmentation aims to identify pixels of objects not present during the model\u2019s training. Recent approaches address this task using mask-based architectures, but these methods have high training costs due to the large transformer backbones involved. While vision adapters can help reduce training costs, they are not specialized for this task, leading to inferior performance. In this work, we propose Dual-Stream Adapters (DSA), a vision adapter tailored for anomaly segmentation. DSA extracts both in-distribution and out-of-distribution features via (i) an anomaly prior module that produces separate initial embeddings for the two streams; and (ii) a dual-stream feature refinement that implicitly guides the separation of in-distribution from out-of-distribution features. We train DSA using a novel hyperbolic loss function that provides supervised guidance for differentiating in-distribution and out-of-distribution features. Experiments on various benchmarks show that dual-stream adapters achieve the best results while reducing training parameters by 38\\% w.r.t. the previous state-of-the-art.", "keywords": [ "Adapters", "Anomaly Segmentation" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 1, 3, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "D1Y2XFgsPI", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Gael Varoquaux", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Full Professor" }, { "name": "Marine Le Morvan", "gender": "Female", "institution": "INRIA", "country": "FR", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 36, "n_ref": 85, "n_ref_all": 119, "n_fig": 13, "n_tab": 8, "L_tab": 3803, "n_element_tab": 494, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 1584, "n_element_tab_1": 179, "formula_len_all": 423, "formula_len_all_1": 0, "len_all": 148999, "len_all_1": 72592, "len_abs": 1190, "len_title": 104, "len_sents": 47862, "len_sents_1": 35576, "n_sents": 353, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 57, "L_abs": 1198, "title": "Imputation for prediction: beware of diminishing returns.", "abs": "Missing values are prevalent across various fields, posing challenges for training and deploying predictive models. In this context, imputation is a common practice, driven by the hope that accurate imputations will enhance predictions. However, recent theoretical and empirical studies indicate that simple constant imputation can be consistent and competitive. This empirical study aims at clarifying \n*if* and *when* investing in advanced imputation methods yields significantly better predictions. Relating imputation and predictive accuracies across combinations of imputation and predictive models on 19 datasets, we show that imputation accuracy matters less i) when using expressive models, ii) when incorporating missingness indicators as complementary inputs, iii) matters much more for generated linear outcomes than for real-data outcomes. Interestingly, we also show that the use of the missingness indicator is beneficial to the prediction performance, even in MCAR scenarios. Overall, on real-data with powerful models, imputation quality has only a minor effect on prediction performance. Thus, investing in better imputations for improved predictions often offers limited benefits.", "keywords": [ "imputation", "missing" ], "rating_list": [ 8, 6, 8, 8, 6 ], "soundness_list": [ 4, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 1, 3 ], "contribution_list": [ 3, 2, 3, 3, 2 ], "confidence_list": [ 4, 5, 4, 4, 3 ] }, { "paper_id": "D10yarGQNk", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Adam Goodge", "gender": "Male", "institution": "A*STAR", "country": "SG", "position": "Researcher" }, { "name": "Kui Jia", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Xun Xu", "gender": "Not Specified", "institution": "A*STAR", "country": "SG", "position": "Scientist" }, { "name": "Yongyi Su", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Yushu Li", "gender": "Male", "institution": "South China University of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 25, "n_formula_1": 19, "n_ref_uni": 39, "n_ref": 102, "n_ref_all": 112, "n_fig": 7, "n_tab": 9, "L_tab": 5636, "n_element_tab": 404, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 4035, "n_element_tab_1": 216, "formula_len_all": 1827, "formula_len_all_1": 1532, "len_all": 147402, "len_all_1": 65103, "len_abs": 1160, "len_title": 164, "len_sents": 39934, "len_sents_1": 26591, "n_sents": 290, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 114, "L_abs": 1166, "title": "Efficient and Context-Aware Label Propagation for Zero-/Few-Shot Training-Free Adaptation of Vision-Language Model", "abs": "Vision-language models (VLMs) have revolutionized machine learning by leveraging large pre-trained models to tackle various downstream tasks. Despite improvements in label, training, and data efficiency, many state-of-the-art VLMs still require task-specific hyperparameter tuning and fail to fully exploit test samples. To overcome these challenges, we propose a graph-based approach for label-efficient adaptation and inference. Our method dynamically constructs a graph over text prompts, few-shot examples, and test samples, using label propagation for inference without task-specific tuning. Unlike existing zero-shot label propagation techniques, our approach requires no additional unlabeled support set and effectively leverages the test sample manifold through dynamic graph expansion. We further introduce a context-aware feature re-weighting mechanism to improve task adaptation accuracy. Additionally, our method supports efficient graph expansion, enabling real-time inductive inference. Extensive evaluations on downstream tasks, such as fine-grained categorization and out-of-distribution generalization, demonstrate the effectiveness of our approach.", "keywords": [ "Vision-Language Model", "Label Propagation", "Training-Free" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 4, 3, 2, 3 ], "confidence_list": [ 5, 4, 3, 3 ] }, { "paper_id": "D0hd7YA0fP", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Huaming Chen", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Jiahao Huang", "gender": "unknown", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "MS student" }, { "name": "Jiayu Zhang", "gender": "Male", "institution": "Suzhou Yierqi", "country": "", "position": "Researcher" }, { "name": "Xinyi Wang", "gender": "Male", "institution": "Universiti Malaya", "country": "", "position": "MS student" }, { "name": "Xinyi Zhang", "gender": "Female", "institution": "University of Sydney, University of Sydney", "country": "CN", "position": "Undergrad student" }, { "name": "Zhibo Jin", "gender": "Male", "institution": "University of Technology Sydney", "country": "AU", "position": "PhD student" }, { "name": "Zhu Zhiyu", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "MS student" } ], "n_formula": 31, "n_formula_1": 32, "n_ref_uni": 28, "n_ref": 62, "n_ref_all": 66, "n_fig": 2, "n_tab": 3, "L_tab": 1791, "n_element_tab": 283, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1348, "n_element_tab_1": 187, "formula_len_all": 1780, "formula_len_all_1": 1776, "len_all": 105840, "len_all_1": 66105, "len_abs": 1411, "len_title": 140, "len_sents": 31634, "len_sents_1": 26936, "n_sents": 253, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Reject", "L_title": 91, "L_abs": 1504, "title": "Splitting & Integrating: Out-of-Distribution Detection via Adversarial Gradient Attribution", "abs": "Out-of-distribution (OOD) detection is essential for enhancing the robustness and security of deep learning models in unknown and dynamic data environments. Gradient-based OOD detection methods, such as GAIA, analyse the explanation pattern representations of in-distribution (ID) and OOD samples by examining the sensitivity of model outputs w.r.t. model inputs, resulting in superior performance compared to traditional OOD detection methods. However, we argue that the non-zero gradient behaviors of OOD samples do not exhibit significant distinguishability, especially when ID samples are perturbed by random noise in high-dimensional spaces, which negatively impacts the accuracy of OOD detection. In this paper, we propose a novel OOD detection method called **S \\& I** based on layer **S**plitting and gradient **I**ntegration via Adversarial Gradient Attribution. Specifically, our approach involves splitting the model's intermediate layers and iteratively updating adversarial examples layer-by-layer. We then integrate the attribution gradients from each intermediate layer along the attribution path from adversarial examples to the actual input, yielding true explanation pattern representations for both ID and OOD samples. Experiments demonstrate that our S \\& I algorithm achieves state-of-the-art results, with the average FPR95 of 29.05\\% (38.61\\%) and 37.31\\% on the CIFAR100 and ImageNet benchmarks, respectively. Our code is available at: https://anonymous.4open.science/r/S-I-F6F7/.", "keywords": [ "Out-of-Distribution Detection", "Adversarial Gradient Attribution", "Safety" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 2, 2, 4, 4 ] }, { "paper_id": "D0XpSucS3l", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "David Bignell", "gender": "Male", "institution": "Research, Microsoft", "country": "", "position": "Researcher" }, { "name": "Katja Hofmann", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Senior Principal Researcher" }, { "name": "Raluca Georgescu", "gender": "Female", "institution": "Microsoft", "country": "", "position": "Researcher" }, { "name": "Sam Devlin", "gender": "Male", "institution": "Microsoft Research", "country": "", "position": "Principal Researcher" }, { "name": "Tabish Rashid", "gender": "Male", "institution": "Microsoft", "country": "GB", "position": "Researcher" }, { "name": "Tim Pearce", "gender": "unknown", "institution": "Research, Microsoft", "country": "GB", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 45, "n_ref": 87, "n_ref_all": 120, "n_fig": 17, "n_tab": 10, "L_tab": 1953, "n_element_tab": 314, "n_fig_1": 11, "n_tab_1": 5, "L_tab_1": 2030, "n_element_tab_1": 51, "formula_len_all": 512, "formula_len_all_1": 477, "len_all": 150795, "len_all_1": 61044, "len_abs": 829, "len_title": 101, "len_sents": 43115, "len_sents_1": 26241, "n_sents": 388, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 818, "title": "Scaling Laws for Pre-training Agents and World Models", "abs": "The performance of embodied agents has been shown to improve by increasing model parameters, dataset size, and compute. This has been demonstrated in domains from robotics to video games, when simple learning objectives on offline datasets (pre-training) are used to model an agent's behavior (imitation learning) or their environment (world modeling). This paper characterizes the role of scale in these tasks more precisely. Going beyond the simple intuition that `bigger is better', we show that the same types of power laws found in language modeling (e.g. between loss and optimal model size), also arise in world modeling and imitation learning. However, the coefficients of these laws are influenced by the tokenizer, task \\& architecture -- this has important implications on optimal sizing of models and data.", "keywords": [ "world modeling", "imitation learning", "scaling laws" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "D0LuQNZfEl", "primary_area": "datasets and benchmarks", "n_author": 5, "author_info_list": [ { "name": "Bhiksha Raj", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Full Professor" }, { "name": "David Solans Noguero", "gender": "Male", "institution": "Telefonica Research", "country": "ES", "position": "Researcher" }, { "name": "Muhammad A Shah", "gender": "unknown", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Mikko A Heikkil\u00e4", "gender": "unknown", "institution": "University of Helsinki", "country": "FI", "position": "Postdoc" }, { "name": "Nicolas Kourtellis", "gender": "Male", "institution": "Keysight Labs", "country": "ES", "position": "Operational R&D Manager" } ], "n_formula": 8, "n_formula_1": 1, "n_ref_uni": 61, "n_ref": 160, "n_ref_all": 194, "n_fig": 14, "n_tab": 8, "L_tab": 5228, "n_element_tab": 856, "n_fig_1": 12, "n_tab_1": 2, "L_tab_1": 26, "n_element_tab_1": 1, "formula_len_all": 516, "formula_len_all_1": 89, "len_all": 209081, "len_all_1": 65458, "len_abs": 1155, "len_title": 115, "len_sents": 51685, "len_sents_1": 33583, "n_sents": 361, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 66, "L_abs": 1148, "title": "Speech Robust Bench: A Robustness Benchmark For Speech Recognition", "abs": "As Automatic Speech Recognition (ASR) models become ever more pervasive, it is important to ensure that they make reliable predictions under corruptions present in the physical and digital world. We propose Speech Robust Bench (SRB), a comprehensive benchmark for evaluating the robustness of ASR models to diverse corruptions. SRB is composed of 114 input perturbations which simulate an heterogeneous range of corruptions that ASR models may encounter when deployed in the wild. We use SRB to evaluate the robustness of several state-of-the-art ASR models and observe that model size and certain modeling choices such as the use of discrete representations, or self-training appear to be conducive to robustness. We extend this analysis to measure the robustness of ASR models on data from various demographic subgroups, namely English and Spanish speakers, and males and females. Our results revealed noticeable disparities in the model's robustness across subgroups. We believe that SRB will significantly facilitate future research towards robust ASR models, by making it easier to conduct comprehensive and comparable robustness evaluations.", "keywords": [ "robustness", "automatic speech recognition", "benchmark", "adversarial" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "D0Cdljktp2", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Sanchayan Dutta", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Intern" }, { "name": "Suvrit Sra", "gender": "unknown", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "Full Professor" } ], "n_formula": 76, "n_formula_1": 32, "n_ref_uni": 25, "n_ref": 65, "n_ref_all": 100, "n_fig": 11, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 10, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3590, "formula_len_all_1": 1336, "len_all": 131013, "len_all_1": 59256, "len_abs": 928, "len_title": 132, "len_sents": 40697, "len_sents_1": 25152, "n_sents": 384, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 931, "title": "Memory-augmented Transformers can implement Linear First-Order Optimization Methods", "abs": "We show that memory-augmented Transformers (Memformers) can implement linear first-order optimization methods such as conjugate gradient descent, momentum methods, and more generally, methods that linearly combines past gradients. Building on prior work that demonstrates how Transformers can simulate preconditioned gradient descent, we provide theoretical and empirical evidence that Memformers can learn more advanced optimization algorithms. Specifically, we analyze how memory registers in Memformers store suitable intermediate attention values allowing them to implement algorithms such as conjugate gradient. Our results show that Memformers can efficiently learn these methods by training on random linear regression tasks, even learning methods that outperform conjugate gradient. This work extends our knowledge about the algorithmic capabilities of Transformers, showing how they can learn complex optimization methods.", "keywords": [ "in-context learning", "memory-augmented transformers", "memformers", "first-order methods", "conjugate gradient descent", "transformers" ], "rating_list": [ 3, 6, 3 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "D042vFwJAM", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Lijun Yang", "gender": "Female", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Shen Fang", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "PhD student" }, { "name": "Wang Han", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "", "position": "Full Professor" }, { "name": "Yue Deng", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Zeyu Li", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "hongkun dou", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 9, "n_ref_uni": 72, "n_ref": 166, "n_ref_all": 199, "n_fig": 15, "n_tab": 6, "L_tab": 5010, "n_element_tab": 823, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 2549, "n_element_tab_1": 493, "formula_len_all": 1637, "formula_len_all_1": 640, "len_all": 210729, "len_all_1": 65134, "len_abs": 1541, "len_title": 107, "len_sents": 50782, "len_sents_1": 27731, "n_sents": 331, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 58, "L_abs": 1621, "title": "Physics-aligned field reconstruction with diffusion bridge", "abs": "The reconstruction of physical fields from sparse measurements is pivotal in both scientific research and engineering applications. Traditional methods are increasingly supplemented by deep learning models due to their efficacy in extracting features from data. However, except for the low accuracy on complex physical systems, these models often fail to comply with essential physical constraints, such as governing equations and boundary conditions. To overcome this limitation, we introduce a novel data-driven field reconstruction framework, termed the Physics-aligned Schr\\\"{o}dinger Bridge (PalSB). This framework leverages a diffusion bridge mechanism that is specifically tailored to align with physical constraints. The PalSB approach incorporates a dual-stage training process designed to address both local reconstruction mapping and global physical principles. Additionally, a boundary-aware sampling technique is implemented to ensure adherence to physical boundary conditions. We demonstrate the effectiveness of PalSB through its application to three complex nonlinear systems: cylinder flow from Particle Image Velocimetry experiments, two-dimensional turbulence, and a reaction-diffusion system. The results reveal that PalSB not only achieves higher accuracy but also exhibits enhanced compliance with physical constraints compared to existing methods. This highlights PalSB's capability to generate high-quality representations of intricate physical interactions, showcasing its potential for advancing field reconstruction techniques. The source code can be found at https://github.com/lzy12301/PalSB.", "keywords": [ "Fluid dynamics", "diffusion models", "super-resolution" ], "rating_list": [ 8, 6, 8 ], "soundness_list": [ 4, 3, 3 ], "presentation_list": [ 4, 3, 3 ], "contribution_list": [ 4, 2, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "Cz8KnDYj1L", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Besmira Nushi", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "Rishabh Adiga", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Varun Chandrasekaran", "gender": "Male", "institution": "University of Illinois Urbana-Champaign", "country": "US", "position": "Assistant Professor" } ], "n_formula": 19, "n_formula_1": 18, "n_ref_uni": 35, "n_ref": 60, "n_ref_all": 88, "n_fig": 6, "n_tab": 7, "L_tab": 1950, "n_element_tab": 311, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1358, "n_element_tab_1": 215, "formula_len_all": 722, "formula_len_all_1": 711, "len_all": 167581, "len_all_1": 63628, "len_abs": 1641, "len_title": 123, "len_sents": 47968, "len_sents_1": 27516, "n_sents": 358, "n_sents_1": 210, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1673, "title": "Attention Speaks Volumes: Localizing and Mitigating Bias in Language Models", "abs": "We explore the internal mechanisms of how bias emerges in large language models (LLMs) when provided with ambiguous comparative prompts: inputs that compare or enforce choosing between two or more entities without providing clear context for preference. Most approaches for bias mitigation focus on either post-hoc analysis or data augmentation. However, these are transient solutions, without addressing the root cause: the model itself. Numerous prior works show the influence of the attention module towards steering generations. We believe that analyzing attention is also crucial for understanding bias, as it provides insight into how the LLM distributes its focus across different entities and how this contributes to biased decisions. To this end, we first introduce a metric to quantify the LLM's preference for one entity over another. We then propose $ATLAS$ (Attention-based Targeted Layer Analysis and Scaling), a technique to localize bias to specific layers of the LLM by analyzing attention scores and then reduce bias by scaling attention in these biased layers. To evaluate our method, we conduct experiments across 3 datasets (BBQ, Crows-Pairs, and WinoGender) using $GPT$-$2$ $XL$ (1.5B), $GPT$-$J$ (6B), $LLaMA$-$2$ (7B) and $LLaMA$-$3$ (8B). Our experiments demonstrate that bias is concentrated in the later layers, typically around the last third. We also show how $ATLAS$ effectively mitigates bias through targeted interventions without compromising downstream performance and an average increase of only 0.34\\% in perplexity when the intervention is applied. We see an average improvement of 0.28 points in the bias score across all the datasets.", "keywords": [ "Bias", "Attention", "LLMs" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "CyxoD9pa5r", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Heesung Kwon", "gender": "Male", "institution": "DEVCOM Army Research Laboratory", "country": "US", "position": "Principal Researcher" }, { "name": "Jinsub Yim", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "MS student" }, { "name": "Shuvra Shikhar Bhattacharyya", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Sungmin Eum", "gender": "Male", "institution": "DEVCOM Army Research Laboratory", "country": "US", "position": "Research Scientist" }, { "name": "Yan Zhang", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Yi-Ting Shen", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Hyungtae Lee", "gender": "Male", "institution": "BlueHalo", "country": "US", "position": "Lead Research Scientist" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 37, "n_ref": 58, "n_ref_all": 77, "n_fig": 5, "n_tab": 6, "L_tab": 4026, "n_element_tab": 447, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 2297, "n_element_tab_1": 205, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 141737, "len_all_1": 67828, "len_abs": 366, "len_title": 117, "len_sents": 41653, "len_sents_1": 31594, "n_sents": 301, "n_sents_1": 219, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1567, "title": "SynPlay: Importing Real-world Diversity for a Synthetic Human Dataset", "abs": "We introduce Synthetic Playground (SynPlay), a new synthetic human dataset that aims to bring out the diversity of human appearance in the real world. We focus on two factors to achieve a level of diversity that has not yet been seen in previous works: i) realistic human motions and poses and ii) multiple camera viewpoints towards human instances. We first use a game engine and its library-provided elementary motions to create games where virtual players can take less-constrained and natural movements while following the game rules (i.e., rule-guided motion design as opposed to detail-guided design). We then augment the elementary motions with real human motions captured with a motion capture device. To render various human appearances in the games from multiple viewpoints, we use seven virtual cameras encompassing the ground and aerial views, capturing abundant aerial-vs-ground and dynamic-vs-static attributes of the scene. Through extensive and carefully-designed experiments, we show that using SynPlay in model training leads to enhanced accuracy over existing synthetic datasets for human detection and segmentation. The benefit of SynPlay becomes even greater for tasks in the data-scarce regime, such as few-shot and cross-domain learning tasks. These results clearly demonstrate that SynPlay can be used as an essential dataset with rich attributes of complex human appearances and poses suitable for model pretraining. SynPlay dataset comprising over 73k images and 6.5M human instances, will be publicly released upon acceptance of this paper.", "keywords": [ "Synthetic human dataset", "Natural human pose and motion", "Squid game" ], "rating_list": [ 6, 5, 3 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 4, 4 ] }, { "paper_id": "CyonEdshEn", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Baolong Guo", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "junbin zhuang", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "", "position": "PhD student" }, { "name": "yunyi yan", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "Associate Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 25, "n_ref": 27, "n_ref_all": 41, "n_fig": 3, "n_tab": 4, "L_tab": 3039, "n_element_tab": 469, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1300, "n_element_tab_1": 196, "formula_len_all": 262, "formula_len_all_1": 261, "len_all": 98198, "len_all_1": 53584, "len_abs": 1929, "len_title": 133, "len_sents": 27152, "len_sents_1": 24567, "n_sents": 199, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1956, "title": "Dynamic Weighting: Exploiting the Potential of a Single Weight Across Different Modes", "abs": "Weights play an essential role in determining the performance of deep networks. This paper introduces a new concept termed ``Weight Augmentation Strateg'' (WAS), which emphasizes the exploration of weight spaces rather than traditional network structure design. The core of WAS is the utilization of randomly transformed weight coefficients, referred to as Shadow Weights (SW), for deep networks to calculate the loss function and update the parameters. Differently, stochastic gradient descent is applied to Plain Weights (PW), which is referred to as the original weight of the network before the random transformation. During training, numerous SW collectively form a high-dimensional space, while PW is directly learned from the distribution of SW. To maximize the benefits of WAS, we introduce two operational modes, \\textit{i.e.}, the Accuracy-Priented Mode (AOM) and the Desire-Oriented Mode (DOM). To be concrete, AOM relies on PW, which ensures that the network remains highly robust and accurate. Meanwhile, DOM utilizes SW, which is determined by the specific objective of our proposed WAS, such as reduced computational complexity or lower sensitivity to particular data. These dual modes can be switched at any time as needed, thereby providing flexibility and adaptability to different tasks. By extending the concept of augmentation from data to weights, our WAS offers an easy-to-understand and implement technique that can significantly enhance almost all networks. Our experimental results demonstrate that convolutional neural networks, including VGG-16, ResNet-18, ResNet-34, GoogleNet, MobileNetV2, and EfficientNet-Lite, benefit substantially with little to no additional costs. On the CIFAR-100 and CIFAR-10 datasets, model accuracy increases by an average of 7.32\\% and 9.28\\%, respectively, with the highest improvements reaching 13.42\\% and 18.93\\%. In addition, DOM can reduce floating point operations (FLOPs) by up to 36.33\\%.", "keywords": [ "Weights Augmentation", "Shadow Weights", "Plain Weights", "Accuracy-Oriented Mode", "Desire-Oriented Mode" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 1, 1, 1 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "Cy7G36aHta", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Heesang Ann", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Min-hwan Oh", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Associate Professor" } ], "n_formula": 112, "n_formula_1": 23, "n_ref_uni": 20, "n_ref": 75, "n_ref_all": 77, "n_fig": 0, "n_tab": 1, "L_tab": 21, "n_element_tab": 4, "n_fig_1": 0, "n_tab_1": 1, "L_tab_1": 388, "n_element_tab_1": 6, "formula_len_all": 22314, "formula_len_all_1": 1770, "len_all": 225793, "len_all_1": 70115, "len_abs": 1389, "len_title": 117, "len_sents": 74129, "len_sents_1": 29787, "n_sents": 671, "n_sents_1": 242, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1192, "title": "Mostly Exploration-free Algorithms for Multi-Objective Linear Bandits", "abs": "We address the challenge of solving multi-objective bandit problems, which are increasingly relevant in real-world applications where multiple possibly conflicting objectives must be optimized simultaneously. Existing multi-objective algorithms often rely on complex, computationally intensive methods, making them impractical for real-world use. In this paper, we propose a novel perspective by showing that objective diversity can naturally induce free exploration, allowing for simpler, near-greedy algorithms to achieve optimal regret bounds up to logarithmic factors with respect to the number of rounds. We introduce simple and efficient algorithms for multi-objective linear bandits, which do not require constructing empirical Pareto fronts and achieve a regret bound of $\\tilde{\\mathcal{O}}(\\sqrt{T})$ under sufficient objective diversity and suitable regularity. We also introduce the concept of objective fairness, ensuring equal treatment of all objectives, and show that our algorithms satisfy this criterion. Numerical experiments validate our theoretical findings, demonstrating that objective diversity can enhance algorithm performance while simplifying the solution process.", "keywords": [ "multi-objective", "free exploration", "linear bandit" ], "rating_list": [ 3, 3, 3, 6, 5 ], "soundness_list": [ 2, 1, 2, 2, 3 ], "presentation_list": [ 2, 2, 3, 3, 2 ], "contribution_list": [ 2, 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3, 3 ] }, { "paper_id": "Cy5IKvYbR3", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Han Yu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Minghui Chen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Researcher" }, { "name": "Ruinan Jin", "gender": "unknown", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Wenlong Deng", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Xiaoxiao Li", "gender": "Unspecified", "institution": "University of British Columbia", "country": "CA", "position": "Assistant Professor" }, { "name": "Zhi Huang", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Assistant Professor" }, { "name": "Chen YuanYuan", "gender": "Male", "institution": "Nanyang Technological University", "country": "", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 45, "n_ref": 85, "n_ref_all": 103, "n_fig": 11, "n_tab": 3, "L_tab": 594, "n_element_tab": 48, "n_fig_1": 9, "n_tab_1": 2, "L_tab_1": 2338, "n_element_tab_1": 159, "formula_len_all": 185, "formula_len_all_1": 189, "len_all": 183005, "len_all_1": 65910, "len_abs": 5373, "len_title": 96, "len_sents": 60919, "len_sents_1": 30804, "n_sents": 432, "n_sents_1": 207, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 48, "L_abs": 2518, "title": "Can Textual Gradient Work in Federated Learning?", "abs": "Recent studies highlight the promise of LLM-based prompt optimization, especially with TextGrad, which automates ``differentiation'' via texts and backpropagates textual feedback provided by LLMs. This approach facilitates training in various real-world applications that do not support numerical gradient propagation or loss calculation. It opens new avenues for optimization in decentralized, resource-constrained environments, suggesting that users of black-box LLMs (e.g., ChatGPT) could enhance components of LLM agentic systems (such as prompt optimization) through collaborative paradigms like federated learning (FL). In this paper, we systematically explore the potential and challenges of incorporating textual gradient into FL. Our contributions are fourfold.\n**Firstly**, we introduce a novel FL paradigm, Federated Textual Gradient (FedTextGrad), that allows FL clients to upload their locally optimized prompts derived from textual gradients, while the FL server aggregates the received prompts through text summarization. Unlike traditional FL frameworks, which are designed for numerical aggregation, FedTextGrad is specifically tailored for handling textual data, expanding the applicability of FL to a broader range of problems that lack well-defined numerical loss functions. \n**Secondly**, building on this design, we conduct extensive experiments to explore the feasibility of federated textual gradients. Our findings highlight the importance of properly tuning key factors (e.g., local steps) in FL training to effectively integrate textual gradients. \n**Thirdly**, we highlight a major challenge in federated textual gradient aggregation: retaining essential information from distributed prompt updates. Concatenation often produces prompts that exceed the LLM API\u2019s context window, while summarization can degrade performance by generating overly condensed or complex text that lacks key context. \n**Last but not least**, in response to this issue, we improve the vanilla variant of FedTextGrad by providing actionable guidance to the LLM when summarizing client prompts by leveraging the Uniform Information Density principle. Such a design reduces the complexity of the aggregated global prompt, thereby better incentivizing the LLM's reasoning ability. Through this principled study, we enable the adoption of textual gradients in FL for optimizing LLMs, identify important issues, and pinpoint future directions, thereby opening up a new research area that warrants further investigation.", "keywords": [ "Federated Learning; LLMs-as-Optimizer" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "CxwtuhU40F", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Di Yu", "gender": "unknown", "institution": "University of Queensland", "country": "", "position": "Full Professor" }, { "name": "Hongjian Sun", "gender": "Male", "institution": "University of Queensland", "country": "", "position": "PhD student" }, { "name": "Jialei Gong", "gender": "Male", "institution": "University of Queensland", "country": "", "position": "PhD student" }, { "name": "Yang Yang", "gender": "Male", "institution": "University of Queensland", "country": "", "position": "Postdoc" } ], "n_formula": 57, "n_formula_1": 19, "n_ref_uni": 42, "n_ref": 74, "n_ref_all": 152, "n_fig": 25, "n_tab": 1, "L_tab": 1555, "n_element_tab": 130, "n_fig_1": 13, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3539, "formula_len_all_1": 992, "len_all": 168370, "len_all_1": 60997, "len_abs": 1656, "len_title": 146, "len_sents": 51413, "len_sents_1": 26497, "n_sents": 430, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 98, "L_abs": 1143, "title": "Interpretable Dimensionality Reduction by Feature-preserving Manifold Approximation and Projection", "abs": "Nonlinear dimensionality reduction often lacks interpretability due to the absence of source features in low-dimensional embedding space. We propose FeatureMAP, an interpretable method that preserves source features by tangent space embedding. The core of FeatureMAP is to use local principal component analysis (PCA) to approximate tangent spaces. By leveraging these tangent spaces, FeatureMAP computes gradients to locally reveal feature directions and importance. Additionally, FeatureMAP embeds the tangent spaces into low-dimensional space while preserving alignment between them, providing local gauges for projecting the high-dimensional data points. Unlike UMAP, FeatureMAP employs anisotropic projection to preserve both the manifold structure and the original data density. We apply FeatureMAP to interpreting digit classification, object detection and MNIST adversarial examples, where it effectively distinguishes digits and objects using feature importance and provides explanations for misclassifications in adversarial attacks. We also compare FeatureMAP with other state-of-the-art methods using both local and global metrics.", "keywords": [ "Interpretable", "Gradient", "Tangent space", "Manifold learning" ], "rating_list": [ 5, 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 2, 3, 4 ], "presentation_list": [ 3, 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3, 4 ] }, { "paper_id": "CxXGvKRDnL", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Yibo Yang", "gender": "Unspecified", "institution": "Chan Zuckerberg Initiative", "country": "US", "position": "Researcher" }, { "name": "Justus Christopher Will", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "PhD student" }, { "name": "Stephan Mandt", "gender": "Male", "institution": "University of California, Irvine", "country": "US", "position": "Associate Professor" } ], "n_formula": 32, "n_formula_1": 15, "n_ref_uni": 32, "n_ref": 126, "n_ref_all": 157, "n_fig": 8, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 3046, "formula_len_all_1": 1004, "len_all": 129909, "len_all_1": 62094, "len_abs": 1086, "len_title": 115, "len_sents": 46562, "len_sents_1": 31152, "n_sents": 305, "n_sents_1": 191, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 67, "L_abs": 1089, "title": "Progressive Compression with Universally Quantized Diffusion Models", "abs": "Diffusion probabilistic models have achieved mainstream success in many generative modeling tasks, from image generation to inverse problem solving. A distinct feature of these models is that they correspond to deep hierarchical latent variable models optimizing a variational evidence lower bound (ELBO) on the data likelihood.\nDrawing on a basic connection between likelihood modeling and compression, we explore the potential of diffusion models for progressive coding, resulting in a sequence of bits that can be incrementally transmitted and decoded with progressively improving reconstruction quality.\nUnlike prior work based on Gaussian diffusion or conditional diffusion models, we propose a new form of diffusion model with uniform noise in the forward process, whose negative ELBO corresponds to the end-to-end compression cost using universal quantization.\nWe obtain promising first results on image compression, achieving competitive rate-distortion-realism results on a wide range of bit-rates with a single model, bringing neural codecs a step closer to practical deployment.", "keywords": [ "diffusion", "generative modeling", "compression", "universal quantization" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 3, 2, 3, 4 ], "presentation_list": [ 4, 2, 2, 4 ], "contribution_list": [ 4, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "CxS8mlkOH7", "primary_area": "generative models", "n_author": 5, "author_info_list": [ { "name": "Jingxuan He", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "Undergrad student" }, { "name": "Lechao Cheng", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Mingce Guo", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Shengeng Tang", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Lecturer" }, { "name": "Zhangye Wang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Associate Professor" } ], "n_formula": 11, "n_formula_1": 13, "n_ref_uni": 38, "n_ref": 121, "n_ref_all": 162, "n_fig": 14, "n_tab": 1, "L_tab": 323, "n_element_tab": 34, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 323, "n_element_tab_1": 34, "formula_len_all": 555, "formula_len_all_1": 557, "len_all": 124692, "len_all_1": 56564, "len_abs": 1214, "len_title": 143, "len_sents": 41169, "len_sents_1": 27435, "n_sents": 280, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 95, "L_abs": 1280, "title": "Shaping a Stabilized Video by Mitigating Unintended Changes for Concept-Augmented Video Editing", "abs": "Text-driven video editing utilizing generative diffusion models has garnered significant attention due to their potential applications. However, existing approaches are constrained by the limited word embeddings provided in pre-training, which hinders nuanced editing targeting open concepts with specific attributes. Directly altering the keywords in target prompts often results in unintended disruptions to the attention mechanisms. To achieve more flexible editing easily, this work proposes an improved concept-augmented video editing approach that generates diverse and stable target videos flexibly by devising abstract conceptual pairs. Specifically, the framework involves concept-augmented textual inversion and a dual prior supervision mechanism. The former enables plug-and-play guidance of stable diffusion for video editing, effectively capturing target attributes for more stylized results. The dual prior supervision mechanism significantly enhances video stability and fidelity. Comprehensive evaluations demonstrate that our approach generates more stable and lifelike videos, outperforming state-of-the-art methods. The anonymous code is available at \\href{https://anonymous.4open.science/w/STIVE-PAGE-B4D4/}{https://anonymous.4open.science/w/STIVE-PAGE-B4D4/}.", "keywords": [ "Text-Guided Video Editing" ], "rating_list": [ 3, 3, 6, 3, 5 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 3, 2, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 4, 5, 3, 4, 3 ] }, { "paper_id": "CvunOaPA1W", "primary_area": "datasets and benchmarks", "n_author": 13, "author_info_list": [ { "name": "Guanting Dong", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" }, { "name": "Honggang Zhang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Associate Professor" }, { "name": "Qiuna Tan", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Undergrad student" }, { "name": "Runqi Qiao", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Shanglin Lei", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Xiaofei Wang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Xiaoshuai Song", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Yadong Xue", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Ye Tian", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "YiFan Zhang", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Zhe Wei", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Zhuoma GongQue", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "\u6768\u6c9b\u9752", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 48, "n_ref": 77, "n_ref_all": 97, "n_fig": 7, "n_tab": 5, "L_tab": 2879, "n_element_tab": 303, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 584, "n_element_tab_1": 103, "formula_len_all": 75, "formula_len_all_1": 74, "len_all": 145126, "len_all_1": 53659, "len_abs": 1575, "len_title": 142, "len_sents": 33013, "len_sents_1": 25562, "n_sents": 250, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1464, "title": "Multi-Dimensional Insights: Benchmarking Real-World Personalization in Large Multimodal Models", "abs": "The rapidly developing field of large multimodal models (LMMs) has led to the emergence of diverse models with remarkable capabilities. However, existing benchmarks fail to comprehensively, objectively and accurately evaluate whether LMMs align with the diverse needs of humans in real-world scenarios. To bridge this gap, we propose the Multi-Dimensional Insights (MDI) benchmark, which includes over 500 images covering six common scenarios of human life. Notably, the MDI-Benchmark offers two significant advantages over existing evaluations:\n(1) Each image is accompanied by two types of questions: simple questions to assess the model's understanding of the image, and complex questions to evaluate the model's ability to analyze and reason beyond basic content.\n(2) Recognizing that people of different age groups have varying needs and perspectives when faced with the same scenario, our benchmark stratifies questions into three age categories: young people, middle-aged people, and older people. This design allows for a detailed assessment of LMMs' capabilities in meeting the preferences and needs of different age groups. With MDI-Benchmark, the strong model like GPT-4o achieve 79\\% accuracy on age-related tasks, indicating that existing LMMs still have considerable room for improvement in addressing real-world applications. Looking ahead, we anticipate that the MDI-Benchmark will open new pathways for aligning real-world personalization in LMMs.", "keywords": [ "large multimodal models", "benchmark", "evaluation" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 3 ] }, { "paper_id": "CvttyK4XzV", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Ali Payani", "gender": "Male", "institution": "Cisco", "country": "US", "position": "Researcher" }, { "name": "Bo Shen", "gender": "Male", "institution": "New Jersey Institute of Technology", "country": "US", "position": "Assistant Professor" }, { "name": "Fan Yang", "gender": "Male", "institution": "Wake Forest University", "country": "", "position": "Assistant Professor" }, { "name": "Haiyan Zhao", "gender": "Female", "institution": "New Jersey Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Heng Zhao", "gender": "Male", "institution": "Wake Forest University", "country": "US", "position": "Undergrad student" }, { "name": "Mengnan Du", "gender": "unknown", "institution": "New Jersey Institute of Technology", "country": "US", "position": "Assistant Professor" } ], "n_formula": 11, "n_formula_1": 9, "n_ref_uni": 31, "n_ref": 53, "n_ref_all": 71, "n_fig": 7, "n_tab": 2, "L_tab": 1627, "n_element_tab": 52, "n_fig_1": 7, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 618, "formula_len_all_1": 907, "len_all": 167020, "len_all_1": 65148, "len_abs": 1427, "len_title": 139, "len_sents": 58233, "len_sents_1": 31338, "n_sents": 491, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1110, "title": "Beyond Single Concept Vector: Modeling Concept Subspace in LLMs with Gaussian Distribution", "abs": "Probing learned concepts in large language models (LLMs) is crucial for understanding how semantic knowledge is encoded internally. Training linear classifiers on probing tasks is a principle approach to denote the vector of a certain concept in the representation space. However, the single vector identified for a concept varies with both data and training, making it less robust and weakening its effectiveness in real-world applications. To address this challenge, we propose an approach to approximate the subspace representing a specific concept. Built on linear probing classifiers, we extend the concept vectors into Gaussian Concept Subspace (GCS). We demonstrate GCS's effectiveness through measuring its faithfulness and plausibility across multiple LLMs with different sizes and architectures. Additionally, we use representation intervention tasks to showcase its efficacy in real-world applications such as emotion steering. Experimental results indicate that GCS concept vectors have the potential to balance steering performance and maintaining the fluency in natural language generation tasks.", "keywords": [ "Large language model", "Explainability", "Probing", "Gaussian distribution" ], "rating_list": [ 6, 8, 5, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "CvrXy1jVLh", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 3, "author_info_list": [ { "name": "Marco Pedersoli", "gender": "Male", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "Associate Professor" }, { "name": "Matthew Toews", "gender": "unknown", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "Full Professor" }, { "name": "Mehraveh Javan Roshtkhari", "gender": "Female", "institution": "\u00c9cole de technologie sup\u00e9rieure, Universit\u00e9 du Qu\u00e9bec", "country": "CA", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 2, "n_ref_uni": 49, "n_ref": 154, "n_ref_all": 182, "n_fig": 7, "n_tab": 6, "L_tab": 2444, "n_element_tab": 271, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 610, "n_element_tab_1": 7, "formula_len_all": 208, "formula_len_all_1": 54, "len_all": 172326, "len_all_1": 65649, "len_abs": 1340, "len_title": 115, "len_sents": 49415, "len_sents_1": 34419, "n_sents": 353, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1345, "title": "Neural Architecture Search by Learning a Hierarchical Search Space", "abs": "Monte-Carlo Tree Search (MCTS) is a powerful tool for many non-differentiable search related problems such as adversarial games. However, the performance of such approach highly depends on the order of the nodes that are considered at each branching of the tree. If the first branches are not discriminative enough, i.e. they cannot distinguish between promising and deceiving configurations for the final task, the efficiency of the search is exponentially reduced. While in some cases the order of the branching is given as part of the problem (e.g. in chess the sequential order of the moves is defined by the game), in others, such as Neural Architecture Search (NAS), the visiting order of the tree is not important, and only the final architecture matters. In this paper, we study the application of MCTS to NAS for the task of image classification. We analyze several sampling methods and branching alternatives for MCTS and propose to learn the branching by hierarchical clustering of architectures based on their similarity. The similarity is measured by the pairwise distance of output vectors of architectures. Extensive experiments on two challenging benchmarks on CIFAR10 and ImageNet show that MCTS, if provided with a good branching hierarchy, can yield promising solutions more efficiently than other approaches for NAS problems.", "keywords": [ "Neural Architecture Search", "Monte-Carlo Tree Search", "Hierarchical Search Space", "Hierarchical Clustering" ], "rating_list": [ 3, 6, 6 ], "soundness_list": [ 2, 2, 3 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 2, 4 ] }, { "paper_id": "CvjXlsBLCX", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Chun Jung Chen", "gender": "Male", "institution": "Academia Sinica", "country": "TW", "position": "Researcher" }, { "name": "Chung-Chin Shih", "gender": "Male", "institution": "Academia Sinica", "country": "TW", "position": "Postdoc" }, { "name": "Ti-Rong Wu", "gender": "unknown", "institution": "Academia Sinica", "country": "TW", "position": "Assistant Research Fellow/Professor" } ], "n_formula": 11, "n_formula_1": 10, "n_ref_uni": 34, "n_ref": 54, "n_ref_all": 94, "n_fig": 24, "n_tab": 2, "L_tab": 403, "n_element_tab": 40, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 338, "n_element_tab_1": 7, "formula_len_all": 905, "formula_len_all_1": 805, "len_all": 139179, "len_all_1": 71263, "len_abs": 1307, "len_title": 111, "len_sents": 44510, "len_sents_1": 33213, "n_sents": 334, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 63, "L_abs": 1318, "title": "Strength Estimation and Human-Like Strength Adjustment in Games", "abs": "Strength estimation and adjustment are crucial in designing human-AI interactions, particularly in games where AI surpasses human players.\nThis paper introduces a novel strength system, including a strength estimator (SE) and an SE-based Monte Carlo tree search, denoted as SE-MCTS, which predicts strengths from games and offers different playing strengths with human styles.\nThe strength estimator calculates strength scores and predicts ranks from games without direct human interaction.\nSE-MCTS utilizes the strength scores in a Monte Carlo tree search to adjust playing strength and style.\nWe first conduct experiments in Go, a challenging board game with a wide range of ranks.\nOur strength estimator significantly achieves over 80\\% accuracy in predicting ranks by observing 15 games only, whereas the previous method reached 49\\% accuracy for 100 games.\nFor strength adjustment, SE-MCTS successfully adjusts to designated ranks while achieving a 51.33\\% accuracy in aligning to human actions, outperforming a previous state-of-the-art, with only 42.56\\% accuracy.\nTo demonstrate the generality of our strength system, we further apply SE and SE-MCTS to chess and obtain consistent results.\nThese results show a promising approach to strength estimation and adjustment, enhancing human-AI interactions in games.", "keywords": [ "Bradley-Terry Model", "Strength Estimation", "Strength Adjustment", "Human-like Playing Style", "Monte-Carlo Tree Search", "Go", "Chess" ], "rating_list": [ 6, 6, 3, 5 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "CvGqMD5OtX", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Amin Saberi", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Fatma Ozcan", "gender": "Female", "institution": "Google", "country": "", "position": "Principal Researcher" }, { "name": "Gaurav Tarlok Kakkar", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Hailong Li", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Mohammadreza Pourreza", "gender": "Male", "institution": "Google", "country": "CA", "position": "Researcher" }, { "name": "Ruoxi Sun", "gender": "Female", "institution": "Google", "country": "", "position": "Google" }, { "name": "Sercan O Arik", "gender": "Male", "institution": "Google", "country": "US", "position": "Research Scientist" }, { "name": "Shayan Talaei", "gender": "Male", "institution": "Institute of Science and Technology", "country": "", "position": "Intern" }, { "name": "Yeounoh Chung", "gender": "Male", "institution": "Google", "country": "US", "position": "Engineer" }, { "name": "Yu Gan", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 36, "n_ref": 83, "n_ref_all": 123, "n_fig": 24, "n_tab": 12, "L_tab": 4079, "n_element_tab": 267, "n_fig_1": 12, "n_tab_1": 7, "L_tab_1": 1405, "n_element_tab_1": 161, "formula_len_all": 114, "formula_len_all_1": 114, "len_all": 178125, "len_all_1": 71280, "len_abs": 3463, "len_title": 140, "len_sents": 59178, "len_sents_1": 33872, "n_sents": 402, "n_sents_1": 226, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1535, "title": "CHASE-SQL: Multi-Path Reasoning and Preference Optimized Candidate Selection in Text-to-SQL", "abs": "In addressing the challenges of improving large language model (LLM) performance for Text-to-SQL tasks, we propose a new framework, CHASE-SQL, that is comprised of innovative strategies that leverage judiciously-designed test-time compute in multi-agent modeling to enhance candidate generation and selection. Our approach leverages LLMs\u2019 intrinsic knowledge to generate diverse and high-quality SQL candidates using different LLM generators with: (1) a divide-and-conquer method that decomposes complex queries into manageable sub-queries in a single LLM call; (2) chain-of-thought reasoning based on query execution plans, reflecting the steps a database engine takes during execution; and (3) a unique instance-aware synthetic example generation technique, which offers specific few-shot demonstrations tailored to test questions. To identify the best candidate, a selection agent is employed to rank the candidates through pairwise comparisons with a fine-tuned binary-candidates selection LLM. This selection approach has been demonstrated more robust over alternatives. The proposed generators-selector framework not only enhances the quality and diversity of SQL queries but also outperforms previous methods. Overall, our proposed CHASE-SQL achieves the state-of-the-art execution accuracy of 73.0 % and 73.01% on the test set and development set of the notable BIRD Text-to-SQL dataset bench-mark, rendering CHASE-SQL the top submission of the leaderboard (at the time of paper submission)", "keywords": [ "Text-to-SQL", "LLM", "Databases" ], "rating_list": [ 5, 8, 6, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 4, 4, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 5, 4, 5 ] }, { "paper_id": "CuwjD3cazX", "primary_area": "reinforcement learning", "n_author": 8, "author_info_list": [ { "name": "Chengcheng Han", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "PhD student" }, { "name": "Jingang Wang", "gender": "Male", "institution": "Meituan", "country": "", "position": "Researcher" }, { "name": "Jun Xu", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Wei Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xuezhi Cao", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Xunliang Cai", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Principal Researcher" }, { "name": "Yang Bai", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" }, { "name": "Weng Rongxiang", "gender": "Male", "institution": "Meituan", "country": "CN", "position": "Researcher" } ], "n_formula": 25, "n_formula_1": 15, "n_ref_uni": 35, "n_ref": 66, "n_ref_all": 89, "n_fig": 17, "n_tab": 5, "L_tab": 1468, "n_element_tab": 252, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1359, "n_element_tab_1": 236, "formula_len_all": 3473, "formula_len_all_1": 1023, "len_all": 160773, "len_all_1": 64219, "len_abs": 1561, "len_title": 104, "len_sents": 46671, "len_sents_1": 28770, "n_sents": 352, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1619, "title": "Length Desensitization in Direct Preference Optimization", "abs": "Direct Preference Optimization (DPO) is widely utilized in the Reinforcement Learning from Human Feedback (RLHF) phase to align Large Language Models (LLMs) with human preferences, thereby enhancing both their harmlessness and efficacy. However, it has been observed that DPO tends to over-optimize for verbosity, which can detrimentally affect both performance and user experience. In this paper, we conduct an in-depth theoretical analysis of DPO's optimization objective and reveal a strong correlation between its implicit reward and data length. This correlation misguides the optimization direction, resulting in length sensitivity during the DPO training and leading to verbosity. To address this issue, we propose a length-desensitization improvement method for DPO, termed LD-DPO. The proposed method aims to desensitize DPO to data length by decoupling explicit length preference, which is relatively insignificant, from the other implicit preferences, thereby enabling more effective learning of the intrinsic preferences. We utilized two settings (Base and Instruct) of Llama2-13B, Llama3-8B, and Qwen2-7B for experimental validation on various benchmarks including MT-Bench and AlpacaEval 2. The experimental results indicate that LD-DPO consistently outperforms DPO and other baseline methods, achieving more concise responses with a 10-40\\% reduction in length compared to DPO. We conducted in-depth experimental analyses to demonstrate that LD-DPO can indeed achieve length desensitization and align the model more closely with human-like preferences.\n\u201dBrevity is the Soul of Wit.''\u2014William Shakespeare", "keywords": [ "large language model", "reinforcement learning from human feedback", "preference optimization" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 4, 4, 2 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "CuupjjjT3U", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Huizhuo Yuan", "gender": "Female", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "Xun Zhou", "gender": "unknown", "institution": "bytedance Inc", "country": "", "position": "Researcher" }, { "name": "Yuan Cao", "gender": "Male", "institution": "University of Hong Kong", "country": "", "position": "Assistant Professor" }, { "name": "Yuanzhe Tao", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 68, "n_formula_1": 27, "n_ref_uni": 29, "n_ref": 74, "n_ref_all": 94, "n_fig": 13, "n_tab": 4, "L_tab": 977, "n_element_tab": 131, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1313, "n_element_tab_1": 93, "formula_len_all": 9046, "formula_len_all_1": 1695, "len_all": 143402, "len_all_1": 66855, "len_abs": 4503, "len_title": 116, "len_sents": 38743, "len_sents_1": 28189, "n_sents": 359, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 955, "title": "Towards Simple and Provable Parameter-Free Adaptive Gradient Methods", "abs": "Optimization algorithms such as AdaGrad and Adam have significantly advanced the training of deep models by dynamically adjusting the learning rate during the optimization process. However, adhoc tuning of learning rates poses a challenge, leading to inefficiencies in practice. To address this issue, recent research has focused on developing \"learning-rate-free\" or \"parameter-free\" algorithms that operate effectively without the need for learning rate tuning. This paper presents AdaGrad++ and Adam++, novel parameter-free variants of AdaGrad and Adam with convergence guarantees. We prove that AdaGrad++ achieves comparable convergence rates to AdaGrad in convex optimization without predefined learning rate assumptions. Similarly, Adam++ matches the convergence rate of Adam without relying on any conditions on the learning rates. Experimental results across various deep learning tasks validate the competitive performance of AdaGrad++ and Adam++", "keywords": [ "parameter-free optimization", "adaptive gradient methods" ], "rating_list": [ 6, 3, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 2, 1, 1 ], "confidence_list": [ 3, 5, 5 ] }, { "paper_id": "CuKla49IjN", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Faten Chaieb Chakchouk", "gender": "Female", "institution": "Ecole Fran\u00e7aise d'Electronique et d'Informatique", "country": "FR", "position": "Full Professor" }, { "name": "Mohamed-Djallel DILMI", "gender": "Male", "institution": "Ecole Fran\u00e7aise d'Electronique et d'Informatique", "country": "FR", "position": "Assistant Professor" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 9, "n_ref": 13, "n_ref_all": 24, "n_fig": 4, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 968, "formula_len_all_1": 831, "len_all": 81954, "len_all_1": 50601, "len_abs": 1976, "len_title": 98, "len_sents": 25135, "len_sents_1": 22590, "n_sents": 180, "n_sents_1": 168, "venue": "ICLR2025", "decision": "withdraw", "L_title": 97, "L_abs": 1983, "title": "Epi-attention : Adaptive Context-Aware Attention for Dynamic Feature Relevance in Neural Networks", "abs": "In this paper, we introduce Epi-Attention, a novel context-aware attention mechanism designed to enhance the relevance of features in neural networks by incorporating external contextual information. Unlike traditional attention mechanisms that rely solely on the input sequence, Epi-Attention dynamically adjusts the significance of features based on additional evidence provided by external contexts. This approach allows the model to emphasize or diminish the relevance of specific features, leading to better capture and reflect the internal properties of specific classes. This mechanism provides a nuanced interpretation of feature relevance that aligns with domain knowledge, enabling the model to focus on contextually significant features in a way that resonates with expert understanding. We formalize the problem and present two variants of the proposed mechanism: Scaled Dot-Product Epi-Attention and Self-Epi-Attention, both of which re-evaluate feature importance considering either external or internal information, respectively. By leveraging the dynamic aspect of Epi-Attention, models can highlight local correlations that are characteristic of certain classes, offering a more transparent and interpretable decision-making process compared to global correlations favorized by classical approaches such as Decision trees, Logistic regression and Neural Networks. We demonstrate the efficency of Epi-Attention through three different applications (dynamic feature relevance, processing mixed datatypes and multi-source datasets) with respectively benchmark datasets, including the Wisconsin Breast Cancer, Bank Marketing and ABIDE-II datasets. Our results show significant improvements in model interpretability over traditional models that aligns with domain knowledge. Furthermore, we discuss the potential of Epi-Attention for enhancing explainability in complex machine learning tasks, paving the way for more robust and transparent neural network architectures.", "keywords": [ "Epi-Attention Mechanism", "Context-Aware Attention", "Dynamic Feature Relevance", "Explainability in Machine Learning", "Local Correlation Analysis", "Contextual Information Integration", "Neural Network Interpretability", "Feature Importance Re-evaluation", "Domain-Specific Knowledge Alignment", "Transparent Decision-Making", "Context-Driven Feature Selection", "Class-Specific Feature Characteristics" ], "rating_list": [ 3, 3, 1, 3 ], "soundness_list": [ 1, 1, 2, 2 ], "presentation_list": [ 2, 1, 1, 1 ], "contribution_list": [ 1, 1, 1, 2 ], "confidence_list": [ 3, 4, 4, 5 ] }, { "paper_id": "CuD9J1QxqC", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Aleksandr Panov", "gender": "Male", "institution": "Artificial Intelligence Research Institute", "country": "RU", "position": "Principal Researcher" }, { "name": "Aleksei Staroverov", "gender": "Male", "institution": "AIRI", "country": "RU", "position": "Researcher" }, { "name": "Tatiana Zemskova", "gender": "Female", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "PhD student" }, { "name": "Dmitry A Yudin", "gender": "Male", "institution": "Artificial Intelligence Research Intsitute", "country": "RU", "position": "Principal Researcher" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 41, "n_ref": 80, "n_ref_all": 92, "n_fig": 4, "n_tab": 3, "L_tab": 141, "n_element_tab": 23, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 705, "n_element_tab_1": 80, "formula_len_all": 370, "formula_len_all_1": 370, "len_all": 117767, "len_all_1": 61241, "len_abs": 1616, "len_title": 111, "len_sents": 31486, "len_sents_1": 29393, "n_sents": 233, "n_sents_1": 207, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1623, "title": "Semantic Object Navigation with Segmenting Decision Transformer", "abs": "Understanding scene semantics plays an important role in solving the object navigation task, where an embodied intelligent agent has to find an object in the scene given its semantic category. This task can be divided into two stages: exploring the scene and reaching the found target. In this work, we consider the latter stage of reaching a given semantic goal. This stage is particularly sensitive to errors in the semantic understanding of the scene. To address this challenge, we propose a multimodal and multitasking method called SegDT, which is based on the joint training of a segmentation model and a decision transformer model. Our method aggregates information from multiple multimodal frames to predict the next action and the current segmentation mask of the target object. To optimize our model, we first performed a pre-training phase using a set of collected trajectories. In the second phase, online policy fine-tuning, we addressed the problems of long-term credit assignment and poor sampling efficiency of transformer models. Using the PPO algorithm, we simultaneously trained an RNN-based policy using ground-truth segmentation and transferred its knowledge to the proposed transformer-based model, which trains the segmentation in itself through an additional segmentation loss. We conducted extensive experiments in the Habitat Sim environment and demonstrated the advantage of the proposed method over the basic navigation approach as well as current state-of-the-art methods that do not consider the auxiliary task of improving the quality of the segmentation of the current frame during training.", "keywords": [ "Reinforcement Learning", "Semantic Segmentation", "Navigation", "Robotics" ], "rating_list": [ 1, 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 1, 1, 2, 1, 3 ], "confidence_list": [ 4, 5, 3, 3, 3 ] }, { "paper_id": "CtM5xjRSfm", "primary_area": "datasets and benchmarks", "n_author": 14, "author_info_list": [ { "name": "BOYUAN FENG", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Chandramouli Shama Sastry", "gender": "Male", "institution": "Vector Institute/Dalhousie University", "country": "", "position": "PhD student" }, { "name": "Edward Z. Yang", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Engineer" }, { "name": "Frank Schneider", "gender": "Male", "institution": "University of T\u00fcbingen", "country": "", "position": "Postdoc" }, { "name": "Juhan Bae", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Less Wright", "gender": "Male", "institution": "Facebook", "country": "US", "position": "Researcher" }, { "name": "Mark Saroufim", "gender": "Male", "institution": "Facebook", "country": "US", "position": "ML Engineer" }, { "name": "Michael Rabbat", "gender": "Male", "institution": "Facebook", "country": "CA", "position": "Principal Researcher" }, { "name": "Philipp Hennig", "gender": "Male", "institution": "University of T\u00fcbingen", "country": "DE", "position": "Full Professor" }, { "name": "Priya Kasimbeg", "gender": "Female", "institution": "Research, Google", "country": "", "position": "Research Engineer" }, { "name": "Runa Eschenhagen", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Sourabh Medapati", "gender": "Male", "institution": "Google Deepmind", "country": "US", "position": "Researcher" }, { "name": "Zachary Nado", "gender": "Male", "institution": "Google", "country": "", "position": "Research Engineer" }, { "name": "George Edward Dahl", "gender": "Male", "institution": "Google", "country": "", "position": "Research Scientist" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 18, "n_ref": 29, "n_ref_all": 73, "n_fig": 8, "n_tab": 8, "L_tab": 8368, "n_element_tab": 622, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 2896, "n_element_tab_1": 298, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 147332, "len_all_1": 68564, "len_abs": 1388, "len_title": 125, "len_sents": 49626, "len_sents_1": 33139, "n_sents": 319, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1396, "title": "Accelerating neural network training: An analysis of the AlgoPerf competition", "abs": "The goal of the AlgoPerf: Training Algorithms competition is to evaluate practical speed-ups in neural network training achieved solely by improving the underlying training algorithms. In the external tuning ruleset, submissions must provide workload-agnostic hyperparameter search spaces, while in the self-tuning ruleset they must be completely hyperparameter-free. In both rulesets, submissions are compared on time-to-result across multiple deep learning workloads, training on fixed hardware. This paper presents the inaugural AlgoPerf competition's results, which drew 18 diverse submissions from 10 teams. Our investigation reveals several key findings: (1) The winning submission in the external tuning ruleset, using Distributed Shampoo, demonstrates the effectiveness of non-diagonal preconditioning over popular methods like Adam, even when compared on wall-clock runtime. (2) The winning submission in the self-tuning ruleset, based on the Schedule Free AdamW algorithm, demonstrates a new level of effectiveness for completely hyperparameter-free training algorithms. (3) The top-scoring submissions were surprisingly robust to workload changes. We also discuss the engineering challenges encountered in ensuring a fair comparison between different training algorithms. These results highlight both the significant progress so far, and the considerable room for further improvements.", "keywords": [ "Training algorithms", "optimizers", "benchmark", "competition", "neural network", "training" ], "rating_list": [ 6, 8, 6, 5, 10 ], "soundness_list": [ 2, 3, 2, 3, 4 ], "presentation_list": [ 3, 3, 2, 4, 4 ], "contribution_list": [ 2, 3, 2, 2, 4 ], "confidence_list": [ 1, 3, 2, 4, 4 ] }, { "paper_id": "CscKx97jBi", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Yun Fu", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Full Professor" }, { "name": "Zhi Xu", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 22, "n_ref": 32, "n_ref_all": 37, "n_fig": 1, "n_tab": 2, "L_tab": 1338, "n_element_tab": 157, "n_fig_1": 1, "n_tab_1": 2, "L_tab_1": 1338, "n_element_tab_1": 157, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 114106, "len_all_1": 70227, "len_abs": 1533, "len_title": 85, "len_sents": 33262, "len_sents_1": 32572, "n_sents": 254, "n_sents_1": 250, "venue": "ICLR2025", "decision": "Reject", "L_title": 37, "L_abs": 1550, "title": "Improve Code Generation with Feedback", "abs": "As advancements in Large Language Models (LLMs) continue to accelerate, an increasing number of researchers are exploring the potential of these models to assist in everyday tasks. Despite their remarkable achievements in various downstream applications, several challenges must be addressed. This paper delves into applying LLMs in coding tasks, such as ChatGPT and LLama. Initial observations suggest that directly employing these LLMs does not yield optimal results. However, we have identified that LLMs demonstrate enhanced performance when given appropriate feedback. This includes providing information on the accuracy of the code generated, supplying test cases relevant to the task, and indicating the correct or incorrect outputs for these test cases.\n Furthermore, we have developed an innovative architecture miming human debugging. This approach supplies local variable information to the LLM while executing the generated code. Our architecture facilitates providing feedback to the LLM and simulates the human debugging experience, thereby significantly improving the LLM's code generation capabilities.\n Utilizing our proposed architecture, our model surpasses the current benchmarks of state-of-the-art models in the MBPP and Humaneval datasets. We also present comprehensive analyses and ablation studies to substantiate the efficacy of our methods. These findings open new avenues for enhancing the utility of LLMs in coding tasks, offering a more interactive and practical approach to leveraging these advanced technologies.", "keywords": [ "LLM", "code generation" ], "rating_list": [ 5, 3, 3, 1 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 2, 2, 3, 1 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 3, 4, 5, 5 ] }, { "paper_id": "CsOIYMOZaV", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Aming WU", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Assistant Professor" }, { "name": "Cheng Deng", "gender": "Male", "institution": "Xidian University", "country": "CN", "position": "Full Professor" } ], "n_formula": 26, "n_formula_1": 16, "n_ref_uni": 53, "n_ref": 132, "n_ref_all": 165, "n_fig": 9, "n_tab": 11, "L_tab": 3450, "n_element_tab": 490, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 2157, "n_element_tab_1": 249, "formula_len_all": 1614, "formula_len_all_1": 1027, "len_all": 191728, "len_all_1": 70690, "len_abs": 1722, "len_title": 145, "len_sents": 50891, "len_sents_1": 28791, "n_sents": 414, "n_sents_1": 233, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 97, "L_abs": 1759, "title": "CFD: Learning Generalized Molecular Representation via Concept-Enhanced Feedback Disentanglement", "abs": "To accelerate biochemical research, e.g., drug and protein discovery, molecular representation learning (MRL) has attracted much attention. However, most existing methods follow the closed-set assumption that training and testing data share identical distribution, which limits their generalization abilities in out-of-distribution (OOD) cases. In this paper, we explore designing a new disentangled mechanism for learning generalized molecular representation that exhibits robustness against distribution shifts. And an approach of Concept-Enhanced Feedback Disentanglement (CFD) is proposed, whose goal is to exploit the feedback mechanism to learn distribution-agnostic representation. Specifically, we first propose two dedicated variational encoders to separately decompose distribution-agnostic and spurious features. Then, a set of molecule-aware concepts are tapped to focus on invariant substructure characteristics. By fusing these concepts into the disentangled distribution-agnostic features, the generalization ability of the learned molecular representation could be further enhanced. Next, we execute iteratively the disentangled operations based on a feedback received from the previous output. Finally, based on the outputs of multiple feedback iterations, we construct a self-supervised objective to promote the variational encoders to possess the disentangled capability. In the experiments, our method is verified on multiple real-world molecular datasets. The significant performance gains over state-of-the-art baselines demonstrate that our method can effectively disentangle generalized molecular representation in the presence of various distribution shifts. The source code will be released at https://github.com/AmingWu/MoleculeCFD.", "keywords": [ "Molecular Representation", "Generalization", "Feedback Disentanglement", "Concepts" ], "rating_list": [ 6, 8, 6, 6 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 3, 4, 3, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "Cs6MrbFuMq", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 3, "author_info_list": [ { "name": "Binhang Yuan", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Ran Yan", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "", "position": "PhD student" }, { "name": "YOUHE JIANG", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 27, "n_ref": 74, "n_ref_all": 99, "n_fig": 11, "n_tab": 2, "L_tab": 2280, "n_element_tab": 211, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1206, "formula_len_all_1": 1206, "len_all": 140854, "len_all_1": 66798, "len_abs": 1518, "len_title": 130, "len_sents": 51544, "len_sents_1": 33446, "n_sents": 341, "n_sents_1": 208, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 81, "L_abs": 1645, "title": "HexGen-2: Disaggregated Generative Inference of LLMs in Heterogeneous Environment", "abs": "Disaggregating the prefill and decoding phases represents an effective new paradigm for generative inference of large language models (LLM). This approach offers some significant system advantages, such as eliminating prefill-decoding interference and optimizing resource allocation. However, it is still an challenging open problem about how to deploy the disaggregated inference paradigm across a group of heterogeneous GPUs, which can be an economic alternative of the deployment over the homogeneous high performance GPUs.\nTowards this end, we introduce HexGen-2, a distributed system for high throughput and cost-efficient LLM serving on heterogeneous GPUs following the disaggragated paradigm. Built on top of HexGen, the core component of HexGen-2 is a sophisticated scheduling algorithm that formalizes the allocation of disaggregated LLM inference computations and communications over heterogeneous GPUs and network connections as a constraint optimization problem. We leverage the graph partitioning and max-flow algorithm to co-optimize resource allocation, parallel strategies for distinct inference phases, and the efficiency of inter-phase key-value (KV) cache communications. We conduct extensive experiments to evaluate HexGen-2, i.e., on OPT (30B) and Llama-2 (70B) models in various real-world settings, the results reveal that HexGen-2 delivers up to a 2.0$\\times$ and on average a 1.3$\\times$ improvement in serving throughput, reduces the average inference latency by 1.5$\\times$ compared with state-of-the-art systems given the same price budget, and achieves comparable inference performance with a 30% lower price budget.", "keywords": [ "Distributed Machine Learning System; Generative Inference of LLM." ], "rating_list": [ 5, 5, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 4, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 3, 5, 3, 3 ] }, { "paper_id": "Crsl3zbfvW", "primary_area": "applications to robotics, autonomy, planning", "n_author": 4, "author_info_list": [ { "name": "Daesol Cho", "gender": "unknown", "institution": "Seoul National University", "country": "", "position": "PhD student" }, { "name": "Dongseok Shim", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "H. Jin Kim", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Seungyeon Yoo", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 41, "n_ref": 105, "n_ref_all": 127, "n_fig": 12, "n_tab": 5, "L_tab": 2449, "n_element_tab": 159, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 491, "formula_len_all_1": 457, "len_all": 159838, "len_all_1": 64021, "len_abs": 1537, "len_title": 149, "len_sents": 52108, "len_sents_1": 31515, "n_sents": 359, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1269, "title": "Single-View 3D Representations for Reinforcement Learning by Cross-View Neural Radiance Fields", "abs": "Reinforcement learning (RL) has enabled robots to develop complex skills, but its success in image-based tasks often depends on effective representation learning. Prior works have primarily focused on 2D representations, often overlooking the inherent 3D geometric structure of the world, or have attempted to learn 3D representations that require extensive resources such as synchronized multi-view images even during deployment. To address these issues, we propose a novel RL framework that extracts 3D-aware representations from single-view RGB input, without requiring camera calibration information or synchronized multi-view images during the downstream RL. Our method employs an autoencoder architecture, using a masked ViT as the encoder and a latent-conditioned NeRF as the decoder, trained with cross-view completion to capture fine-grained, 3D geometry-aware representations. Additionally, we utilize a time contrastive loss that further regularizes the learned representation for consistency across different viewpoints. Our method significantly enhances the RL agent\u2019s performance in complex tasks, demonstrating superior effectiveness compared to prior 3D representation-based methods, even when using only a single, uncalibrated camera during deployment.", "keywords": [ "3D scene representation", "Single-view inference", "NeRF", "Reinforcement Learning" ], "rating_list": [ 6, 5, 3, 5, 3 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 4, 2, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 2, 4, 3, 4 ] }, { "paper_id": "CrmUKllBKs", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 6, "author_info_list": [ { "name": "Da Long", "gender": "Male", "institution": "The University of Utah", "country": "US", "position": "PhD student" }, { "name": "Jacob Hochhalter", "gender": "Male", "institution": "University of Utah", "country": "US", "position": "Associate Professor" }, { "name": "Keyan Chen", "gender": "Male", "institution": "University of Utah", "country": "US", "position": "PhD student" }, { "name": "Shandian Zhe", "gender": "Male", "institution": ", University of Utah", "country": "US", "position": "Associate Professor" }, { "name": "WEI W. XING", "gender": "Male", "institution": "University of Sheffield", "country": "GB", "position": "Lecturer" }, { "name": "Yile Li", "gender": "Female", "institution": "University of Utah", "country": "", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 21, "n_ref": 37, "n_ref_all": 65, "n_fig": 13, "n_tab": 5, "L_tab": 2821, "n_element_tab": 227, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 2233, "n_element_tab_1": 191, "formula_len_all": 867, "formula_len_all_1": 716, "len_all": 112253, "len_all_1": 68583, "len_abs": 1651, "len_title": 88, "len_sents": 36800, "len_sents_1": 31670, "n_sents": 282, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 40, "L_abs": 1659, "title": "Pseudo Physics-Informed Neural Operators", "abs": "Recent advancements in operator learning are transforming the landscape of computational physics and engineering, especially alongside the rapidly evolving field of physics-informed machine learning. The convergence of these areas offers\nexciting opportunities for innovative research and applications. However, merging\nthese two realms often demands deep expertise and explicit knowledge of physical systems, which may be challenging or even impractical in relatively complex applications. To address this limitation, we propose a novel framework: Pseudo\nPhysics-Informed Neural Operator (PPI-NO). In this framework, we construct a\nsurrogate physics system for the target system using partial differential equations\n(PDEs) derived from simple, rudimentary physics knowledge, such as basic differential operators. We then couple the surrogate system with the neural operator model, utilizing an alternating update and learning process to iteratively enhance\nthe model\u2019s predictive power. While the physics derived via PPI-NO may not mirror the ground-truth underlying physical laws \u2014 hence the term \u201cpseudo physics\u201d \u2014 this approach significantly enhances the accuracy of current operator learning\nmodels, particularly in data scarce scenarios. Through extensive evaluations across\nfive benchmark operator learning tasks and an application in fatigue modeling,\nPPI-NO consistently outperforms competing methods by a significant margin. The\nsuccess of PPI-NO may introduce a new paradigm in physics-informed machine\nlearning, one that requires minimal physics knowledge and opens the door to\nbroader applications in data-driven physics learning and simulations.", "keywords": [ "Pseudo Physics", "Data-Driven Physics Discovery", "PDEs", "Neural Operator", "AI for science", "Scientific Machine Learning" ], "rating_list": [ 3, 6, 5, 3, 6 ], "soundness_list": [ 2, 3, 3, 2, 4 ], "presentation_list": [ 2, 4, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 2, 4 ], "confidence_list": [ 3, 4, 4, 4, 4 ] }, { "paper_id": "CrOHzVtWmH", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Mengdi Wang", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Associate Professor" }, { "name": "Ming Yin", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Postdoc" }, { "name": "Peng Wei", "gender": "unknown", "institution": "George Washington University", "country": "", "position": "Associate Professor" }, { "name": "Qiwei Di", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "PhD student" }, { "name": "Quanquan Gu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Research Scientist" }, { "name": "BINSHUAI WANG", "gender": "Male", "institution": "George Washington University", "country": "US", "position": "PhD student" } ], "n_formula": 61, "n_formula_1": 26, "n_ref_uni": 45, "n_ref": 69, "n_ref_all": 94, "n_fig": 7, "n_tab": 5, "L_tab": 325, "n_element_tab": 25, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 213, "n_element_tab_1": 5, "formula_len_all": 6241, "formula_len_all_1": 1972, "len_all": 161537, "len_all_1": 68669, "len_abs": 2296, "len_title": 99, "len_sents": 46334, "len_sents_1": 30312, "n_sents": 387, "n_sents_1": 211, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 2120, "title": "Relative-Translation Invariant Wasserstein Distance", "abs": "In many real-world applications, data distributions are often subject to translation shifts caused by various factors such as changes in environmental conditions, sensor settings, or shifts in data collection practices. These distribution shifts pose a significant challenge for measuring the similarity between probability distributions, particularly in tasks like domain adaptation or transfer learning. To address this issue, we introduce a new family of distances, relative-translation invariant Wasserstein distances ($RW_p$), to measure the similarity of two probability distributions under distribution shift. Generalizing it from the classical optimal transport model, we show that $RW_p$ distances are also real distance metrics defined on the quotient set $\\mathcal{P}_p(\\mathbb{R}^n)/\\sim$ and invariant to distribution translations, which forms a family of new metric spaces. When $p=2$, the $RW_2$ distance enjoys more exciting properties, including decomposability of the optimal transport model and translation-invariance of the $RW_2$ distance. Based on these properties, we show that a distribution shift, measured by $W_2$ distance, can be explained in the bias-variance perspective. In addition, we propose two algorithms: one algorithm is a two-stage optimization algorithm for computing the general case of $RW_p$ distance, and the other is a variant of the Sinkhorn algorithm, named $RW_2$ Sinkhorn algorithm, for efficiently calculating $RW_2$ distance, coupling solutions, as well as $W_2$ distance. We also provide the analysis of numerical stability and time complexity for the proposed algorithms. Finally, we validate the $RW_p$ distance metric and the algorithm performance with two experiments. We conduct one numerical validation for the $RW_2$ Sinkhorn algorithm and demonstrate the effectiveness of using $RW_p$ under distribution shift for similar thunderstorm detection. The experimental results report that our proposed algorithm significantly improves the computational efficiency of Sinkhorn in practical applications, and the $RW_p$ distance is robust to distribution translations.", "keywords": [ "Optimal transport theory", "Wasserstein distance", "Distribution shift" ], "rating_list": [ 3, 3, 5, 5, 3 ], "soundness_list": [ 2, 3, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 3, 4 ] }, { "paper_id": "CrMyHiUttz", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Christodoulos Santorinaios", "gender": "unknown", "institution": "Archimedes/Athena RC", "country": "GR", "position": "PhD student" }, { "name": "Evangelos Markakis", "gender": "Male", "institution": "Athens University of Economics and Business (AUEB)", "country": "GR", "position": "Associate Professor" }, { "name": "Georgios Roussakis", "gender": "Male", "institution": "University of Crete, University of Crete", "country": "", "position": "MS student" }, { "name": "Michail Fasoulakis", "gender": "unknown", "institution": "Foundation for Research and Technology-Hellas", "country": "GR", "position": "Postdoc" } ], "n_formula": 50, "n_formula_1": 25, "n_ref_uni": 13, "n_ref": 24, "n_ref_all": 34, "n_fig": 3, "n_tab": 1, "L_tab": 435, "n_element_tab": 39, "n_fig_1": 3, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5818, "formula_len_all_1": 3532, "len_all": 89011, "len_all_1": 60028, "len_abs": 1330, "len_title": 124, "len_sents": 26476, "len_sents_1": 23003, "n_sents": 264, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1337, "title": "Finding Equilibria in Bilinear Zero-sum Games via a Convexity-based Approach", "abs": "We focus on the design of algorithms for finding equilibria in 2-player zero-sum games. Although it is well known that such problems can be solved by a single linear program, there has been a surge of interest in recent years, for simpler algorithms, motivated in part by applications in machine learning. Our work proposes such a method, inspired by the observation that the duality gap (a standard metric for evaluating convergence in general min-max optimization problems) is a convex function for the case of bilinear zero-sum games. To this end, we analyze a descent-based approach, variants of which have also been used as a subroutine in a series of algorithms for approximating Nash equilibria in general non-zero-sum games. \nIn particular, we analyze a steepest descent approach, by finding the direction that minimises the directional derivative of the duality gap function and move towards that. Our main theoretical result is that the derived algorithms achieve a geometric decrease in the duality gap and improved complexity bounds until we reach an approximate equilibrium. Finally, we complement this with an experimental evaluation. Our findings reveal that for some classes of zero-sum games, the running time of our method is comparable with standard LP solvers, even with thousands of available strategies per player.", "keywords": [ "Zero-sum games", "Directional derivative", "gradient descent", "duality gap" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "CrGfGLC2Ad", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Alice Oh", "gender": "Female", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "Full Professor" }, { "name": "Eunsu Kim", "gender": "Female", "institution": "KAIST", "country": "KR", "position": "MS student" }, { "name": "Inha Cha", "gender": "unknown", "institution": "Upstage AI Research", "country": "US", "position": "Researcher" }, { "name": "Jiseon Kim", "gender": "Female", "institution": "Korea Advanced Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Juhyun Oh", "gender": "unknown", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "PhD student" }, { "name": "Wenda Xu", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "William Yang Wang", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 72, "n_ref_all": 97, "n_fig": 8, "n_tab": 11, "L_tab": 3975, "n_element_tab": 395, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1477, "n_element_tab_1": 63, "formula_len_all": 723, "formula_len_all_1": 513, "len_all": 212558, "len_all_1": 63209, "len_abs": 2022, "len_title": 117, "len_sents": 66468, "len_sents_1": 30519, "n_sents": 533, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1547, "title": "Discovering Factor Level Preferences to Improve Human-Model Alignment", "abs": "Despite advancements in Large Language Model (LLM) alignment, understanding the reasons behind LLM preferences remains crucial for bridging the gap between desired and actual behavior. LLMs often exhibit biases or tendencies that diverge from human preferences, such as favoring certain writing styles or producing overly verbose outputs. However, current methods for evaluating preference alignment often lack explainability, relying on coarse-grained comparisons. To address this, we introduce PROFILE (PRObing Factors of InfLuence for Explainability), a novel framework that uncovers and quantifies the influence of specific factors driving preferences. PROFILE's factor level analysis explains the \"why\" behind human-model alignment and misalignment, offering insights into the direction of model improvement. We apply PROFILE to analyze human and LLM preferences across three tasks: summarization, helpful response generation, and document-based question-answering. Our factor level analysis reveals a substantial discrepancy between human and LLM preferences in generation tasks, whereas LLMs show strong alignment with human preferences in evaluation tasks. We demonstrate how leveraging factor level insights, including addressing misaligned factors or exploiting the generation-evaluation gap, can improve alignment with human preferences. This work underscores the importance of explainable preference analysis and highlights PROFILE's potential to provide valuable training signals, driving further improvements in human-LLM alignment.", "keywords": [ "human alignment", "Large Language Model", "explainability", "generation", "evaluation" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 2, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "Cr1XlGBGVm", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Anqi Guo", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Undergrad student" }, { "name": "Boyuan Zheng", "gender": "Male", "institution": "Tongji University", "country": "", "position": "Undergrad student" }, { "name": "Hai Zhang", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "MS student" }, { "name": "JinHang Liu", "gender": "unknown", "institution": "Tongji University", "country": "CN", "position": "Undergrad student" }, { "name": "Junqiao Zhao", "gender": "Male", "institution": "Tongji University", "country": "", "position": "Associate Professor" }, { "name": "Lanqing Li", "gender": "Male", "institution": "Zhejiang Lab", "country": "", "position": "Associate Professor" }, { "name": "Tianying Ji", "gender": "Female", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" } ], "n_formula": 66, "n_formula_1": 16, "n_ref_uni": 34, "n_ref": 109, "n_ref_all": 136, "n_fig": 6, "n_tab": 8, "L_tab": 1568, "n_element_tab": 77, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 7348, "formula_len_all_1": 1460, "len_all": 163675, "len_all_1": 59405, "len_abs": 2603, "len_title": 164, "len_sents": 44546, "len_sents_1": 27005, "n_sents": 383, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 116, "L_abs": 2016, "title": "Scrutinize What We Ignore: Reining In Task Representation Shift Of Context-Based Offline Meta Reinforcement Learning", "abs": "Offline meta reinforcement learning (OMRL) has emerged as a promising approach for interaction avoidance and strong generalization performance by leveraging pre-collected data and meta-learning techniques. \nPrevious context-based approaches predominantly rely on the intuition that alternating optimization between the context encoder and the policy can lead to performance improvements, as long as the context encoder follows the principle of maximizing the mutual information between the task variable $M$ and its latent representation $Z$ ($I(Z;M)$) while the policy adopts the standard offline reinforcement learning (RL) algorithms conditioning on the learned task representation.\nDespite promising results, the theoretical justification of performance improvements for such intuition remains underexplored.\nInspired by the return discrepancy scheme in the model-based RL field, we find that the previous optimization framework can be linked with the general RL objective of maximizing the expected return, thereby explaining performance improvements. \nFurthermore, after scrutinizing this optimization framework, we observe that the condition for monotonic performance improvements does not consider the variation of the task representation. When these variations are considered, the previously established condition may no longer be sufficient to ensure monotonicity, thereby impairing the optimization process.\nWe name this issue \\underline{task representation shift} and theoretically prove that the monotonic performance improvements can be guaranteed with appropriate context encoder updates.\nWe use different settings to rein in the task representation shift on three widely adopted training objectives concerning maximizing $I(Z;M)$ across different data qualities.\nEmpirical results show that reining in the task representation shift can indeed improve performance.\nOur work opens up a new avenue for OMRL, leading to a better understanding between the task representation and performance improvements.", "keywords": [ "offline meta reinforcement learning", "performance improvement guarantee", "task representation shift" ], "rating_list": [ 6, 6, 8, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "Cpr6Wv2tfr", "primary_area": "optimization", "n_author": 5, "author_info_list": [ { "name": "Artem Agafonov", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "PhD student" }, { "name": "Dmitry Kamzolov", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Dmitry Pasechnyuk", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Researcher" }, { "name": "Alexander Vladimirovich Gasnikov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "Full Professor" }, { "name": "Martin Takac", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "", "position": "Associate Professor" } ], "n_formula": 147, "n_formula_1": 53, "n_ref_uni": 49, "n_ref": 162, "n_ref_all": 186, "n_fig": 9, "n_tab": 6, "L_tab": 791, "n_element_tab": 67, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 143, "n_element_tab_1": 20, "formula_len_all": 11969, "formula_len_all_1": 3666, "len_all": 227561, "len_all_1": 73818, "len_abs": 1494, "len_title": 60, "len_sents": 60506, "len_sents_1": 28521, "n_sents": 583, "n_sents_1": 264, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 60, "L_abs": 1513, "title": "OPTAMI: Global Superlinear Convergence of High-order Methods", "abs": "Second-order methods for convex optimization outperform first-order methods in terms of theoretical iteration convergence, achieving rates up to $O(k^{-5})$ for highly-smooth functions. However, their practical performance and applications are limited due to their multi-level structure and implementation complexity. In this paper, we present new results on high-order optimization methods, supported by their practical performance. First, we show that the basic high-order methods, such as the Cubic Regularized Newton Method, exhibit global superlinear convergence for $\\mu$-strongly star-convex functions, a class that includes $\\mu$-strongly convex functions and some non-convex functions. Theoretical convergence results are both inspired and supported by the practical performance of these methods. Secondly, we propose a practical version of the Nesterov Accelerated Tensor method, called NATA. It significantly outperforms the classical variant and other high-order acceleration techniques in practice. The convergence of NATA is also supported by theoretical results. Finally, we introduce an open-source computational library for high-order methods, called OPTAMI. This library includes various methods, acceleration techniques, and subproblem solvers, all implemented as PyTorch optimizers, thereby facilitating the practical application of high-order methods to a wide range of optimization problems. We hope this library will simplify research and practical comparison of methods beyond first-order.", "keywords": [ "second-order optimization", "high-order optimization", "cubic regularised Newton method", "Newton method", "convex optimization" ], "rating_list": [ 5, 6, 8, 6 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "CpiJWKFdHN", "primary_area": "optimization", "n_author": 6, "author_info_list": [ { "name": "Akang Wang", "gender": "Not Specified", "institution": "Shenzhen Research Institute of Big Data", "country": "CN", "position": "Researcher" }, { "name": "Qingjiang Shi", "gender": "Male", "institution": "Tongji University", "country": "CN", "position": "Full Professor" }, { "name": "Yeqing Qiu", "gender": "Male", "institution": "The Chinese Univeristy of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Yiheng Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "PhD student" }, { "name": "Zhi-Quan Luo", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "YE XUE", "gender": "Female", "institution": "Shenzhen Research Institute of Big Data, CUHK(SZ\uff09", "country": "", "position": "Researcher" } ], "n_formula": 51, "n_formula_1": 16, "n_ref_uni": 34, "n_ref": 45, "n_ref_all": 77, "n_fig": 6, "n_tab": 13, "L_tab": 3375, "n_element_tab": 689, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 223, "n_element_tab_1": 9, "formula_len_all": 3615, "formula_len_all_1": 902, "len_all": 133037, "len_all_1": 60300, "len_abs": 1423, "len_title": 123, "len_sents": 37291, "len_sents_1": 28195, "n_sents": 314, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 77, "L_abs": 1385, "title": "ROS: A GNN-based Relax-Optimize-and-Sample Framework for Max-$k$-Cut Problems", "abs": "The Max-$k$-Cut problem is a fundamental combinatorial optimization challenge that generalizes the classic $\\mathcal{NP}$-complete Max-Cut problem. While relaxation techniques are commonly employed to tackle Max-$k$-Cut, they often lack guarantees of equivalence between the solutions of the original problem and its relaxation. To address this issue, we introduce the Relax-Optimize-and-Sample (ROS) framework. In particular, we begin by relaxing the discrete constraints to the continuous probability simplex form. Next, we pre-train and fine-tune a graph neural network model to efficiently optimize the relaxed problem. Subsequently, we propose a sampling-based construction algorithm to map the continuous solution back to a high-quality Max-$k$-Cut solution. By integrating geometric landscape analysis with statistical theory, we establish the consistency of function values between the continuous solution and its mapped counterpart. Extensive experimental results on random regular graphs and the Gset benchmark demonstrate that the proposed ROS framework effectively scales to large instances with up to $20,000$ nodes in just a few seconds, outperforming state-of-the-art algorithms. Furthermore, ROS exhibits strong generalization capabilities across both in-distribution and out-of-distribution instances, underscoring its effectiveness for large-scale optimization tasks.", "keywords": [ "Max-k-Cut", "Learning to Optimize", "Graph Neural Networks", "Pre-train and Fine-tune", "Sampling" ], "rating_list": [ 5, 6, 6 ], "soundness_list": [ 1, 2, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 2, 2 ] }, { "paper_id": "CpgWRFqxhD", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Bo An", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Full Professor" }, { "name": "Chuanxin Tang", "gender": "Male", "institution": "Skywork", "country": "SG", "position": "PhD student" }, { "name": "Hanzhong Allan Guo", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Jiachun Pan", "gender": "Female", "institution": "Skywork AI ", "country": "SG", "position": "Researcher" }, { "name": "Jiahao Lu", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Longtao Zheng", "gender": "Male", "institution": "Nanyang Technological University, Singapore", "country": "SG", "position": "PhD student" }, { "name": "Shuicheng YAN", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Yifan Zhang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Zhenxiong Tan", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Researcher" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 37, "n_ref": 107, "n_ref_all": 132, "n_fig": 12, "n_tab": 2, "L_tab": 1394, "n_element_tab": 182, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 1223, "n_element_tab_1": 67, "formula_len_all": 631, "formula_len_all_1": 630, "len_all": 159069, "len_all_1": 57521, "len_abs": 1529, "len_title": 110, "len_sents": 40210, "len_sents_1": 26822, "n_sents": 289, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 62, "L_abs": 1541, "title": "MEMO: Memory-Guided and Emotion-Aware Talking Video Generation", "abs": "Advances in video diffusion models have unlocked the potential for realistic audio-driven talking video generation. However, it is still highly challenging to ensure seamless audio-lip synchronization, maintain long-term identity consistency, and achieve natural expressions aligned with the audio in generated talking videos. To address these challenges, we propose **M**emory-guided **EMO**tion-aware diffusion (MEMO), an end-to-end audio-driven portrait animation approach to generate identity-consistent and expressive talking videos. Our approach is built around two key modules: (1) a memory-guided temporal module, which enhances long-term identity consistency and smooth motion by developing memory states that store information from all previously generated frames and guide temporal modeling through linear attention; and (2) an emotion-aware audio module, which replaces traditional cross attention with multi-modal attention to enhance audio-video interaction, while detecting emotions from the audio to refine facial expressions via emotion adaptive layer norm. Moreover, MEMO is trained on a large-scale, high-quality dataset of talking head videos without relying on facial inductive biases such as face landmarks or bounding boxes. Extensive experiments demonstrate that MEMO generates more realistic talking videos across a wide range of audio types, surpassing state-of-the-art talking video diffusion methods in human evaluations in terms of emotion-audio alignment, identity consistency and overall quality, respectively.", "keywords": [ "Talking Head", "Video Generation", "Diffusion Models" ], "rating_list": [ 5, 6, 3 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 5, 3, 5 ] }, { "paper_id": "CpQegoH1Fn", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Issei Suzuki", "gender": "unknown", "institution": "Chukyo University, Tokyo Institute of Technology", "country": "JP", "position": "MS student" }, { "name": "Pitoyo Hartono", "gender": "Male", "institution": "Chukyo University", "country": "JP", "position": "Full Professor" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 18, "n_ref": 29, "n_ref_all": 45, "n_fig": 9, "n_tab": 1, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 263, "formula_len_all_1": 319, "len_all": 79133, "len_all_1": 46452, "len_abs": 1477, "len_title": 107, "len_sents": 25433, "len_sents_1": 21692, "n_sents": 200, "n_sents_1": 167, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1483, "title": "Human-in-the-loop Neural Networks: Human Knowledge Infusion", "abs": "This study proposes a method for infusing human knowledge into neural networks.\nThe primary objective of this study is to build a mechanism that allows neural networks to learn not only from data but also from humans. This motivation is triggered by the fact that human knowledge, experience, personal preferences, and other subjective characteristics are not necessarily easy to mathematically formulate as structured data, hindering them from being learned by neural networks. This study is made possible by a neural network model with a two-dimensional topological hidden representation, Restricted Radial Basis Function (rRBF) network. In rRBF, the hidden layer's low dimensionality allows humans to visualize the internal representation of the neural network and thus intuitively understand its characteristics. In this study, the topological layer is further utilized to allow humans to organize it considering their subjective similarities criterion for the inputs. Hence, the infusion of human knowledge occurs during this process, which initializes the rRBF. The subsequent learning process of rRBF ensures that the infused knowledge is inherited during and after the learning process, thus generating a unique neural network that benefits from human knowledge. This study contributes to the new field of human-in-the-loop (HITL) AI, which aims to allow humans to participate constructively in AI's learning process or decision-making and define a new human-AI relationship.", "keywords": [ "human-in-the-loop;topological representations;metric learning;dimensionality reduction;transfer learning" ], "rating_list": [ 3, 5, 3, 5 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 2, 2, 1, 3 ], "contribution_list": [ 1, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "CoQw1dXtGb", "primary_area": "applications to neuroscience & cognitive science", "n_author": 3, "author_info_list": [ { "name": "Motoaki Kawanabe", "gender": "Male", "institution": "ATR", "country": "", "position": "Department Head" }, { "name": "Reinmar J Kobler", "gender": "Male", "institution": "ATR", "country": "JP", "position": "Researcher" }, { "name": "Shanglin Li", "gender": "Male", "institution": "ATR", "country": "JP", "position": "Intern" } ], "n_formula": 33, "n_formula_1": 25, "n_ref_uni": 67, "n_ref": 127, "n_ref_all": 171, "n_fig": 9, "n_tab": 7, "L_tab": 5540, "n_element_tab": 964, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 284, "n_element_tab_1": 42, "formula_len_all": 2775, "formula_len_all_1": 1725, "len_all": 198127, "len_all_1": 67601, "len_abs": 1437, "len_title": 126, "len_sents": 43792, "len_sents_1": 30692, "n_sents": 313, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1445, "title": "SPDIM: Source-Free Unsupervised Conditional and Label Shift Adaptation in EEG", "abs": "The non-stationary nature of electroencephalography (EEG) introduces distribution shifts across domains (e.g., days and subjects), posing a significant challenge to EEG-based neurotechnology generalization.\nWithout labeled calibration data for target domains, the problem is a source-free unsupervised domain adaptation (SFUDA) problem.\nFor scenarios with constant label distribution, Riemannian geometry-aware statistical alignment frameworks on the symmetric positive definite (SPD) manifold are considered state-of-the-art.\nHowever, many practical scenarios, including EEG-based sleep staging, exhibit label shifts.\nHere, we propose a geometric deep learning framework for SFUDA problems under specific distribution shifts, including label shifts.\nWe introduce a novel, realistic generative model and show that prior Riemannian statistical alignment methods on the SPD manifold can compensate for specific marginal and conditional distribution shifts but hurt generalization under label shifts.\nAs a remedy, we propose a parameter-efficient manifold optimization strategy termed SPDIM.\nSPDIM uses the information maximization principle to learn a single SPD-manifold-constrained parameter per target domain.\nIn simulations, we demonstrate that SPDIM can compensate for the shifts under our generative model.\nMoreover, using public EEG-based brain-computer interface and sleep staging datasets, we show that SPDIM outperforms prior approaches.", "keywords": [ "geometric deep learning", "transfer learning", "source-free adaptation", "electroencephalography", "neurology", "brain-computer interfaces" ], "rating_list": [ 6, 6, 5, 6, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 1, 2, 4 ], "contribution_list": [ 2, 3, 3, 2, 3 ], "confidence_list": [ 3, 5, 2, 3, 3 ] }, { "paper_id": "Co9tdrslVG", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Ga Wu", "gender": "Male", "institution": "Dalhousie University", "country": "CA", "position": "Assistant Professor" }, { "name": "Hassan Sajjad", "gender": "Male", "institution": "Dalhousie University", "country": "CA", "position": "Associate Professor" }, { "name": "Mahtab Sarvmaili", "gender": "unknown", "institution": "Dalhousie University", "country": "CA", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 41, "n_ref": 68, "n_ref_all": 85, "n_fig": 2, "n_tab": 11, "L_tab": 7710, "n_element_tab": 1073, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1087, "n_element_tab_1": 235, "formula_len_all": 718, "formula_len_all_1": 731, "len_all": 167533, "len_all_1": 62809, "len_abs": 1327, "len_title": 107, "len_sents": 52218, "len_sents_1": 30460, "n_sents": 260, "n_sents_1": 170, "venue": "ICLR2025", "decision": "Reject", "L_title": 59, "L_abs": 1271, "title": "Towards Understanding the Feasibility of Machine Unlearning", "abs": "In response to recent privacy protection regulations, machine unlearning has attracted great interest in the research community. However, existing studies often demonstrate their approaches' effectiveness by measuring the overall unlearning success rate rather than evaluating the chance of unlearning specific training samples, leaving the universal feasibility of the unlearning operation unexplored. This paper proposes a novel method to quantify the difficulty of unlearning a single sample by taking into account factors such as model and data distribution. Specifically, we propose several heuristics to understand the condition of a successful unlearning operation on data points, explore difference in unlearning difficulty over training data points, and suggest a potential ranking mechanism for identifying the most challenging samples to unlearn. In particular, we note Kernelized Stein Discrepancy (KSD), a parameterized kernel function tailored to each model and dataset, is an effective heuristic to tell the difficulty of unlearning a data sample. We demonstrate our discovery by including multiple classification tasks and existing machine unlearning algorithms, highlighting the practical feasibility of unlearning operations across different scenarios.", "keywords": [ "Machine Unlearning", "Kernelized Stein Discrepancy" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 5, 3 ] }, { "paper_id": "Cnwz9jONi5", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 10, "author_info_list": [ { "name": "Ben He", "gender": "Male", "institution": "University of Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Debing Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Hongyu Lin", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Le Sun", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Full Professor" }, { "name": "Xianpei Han", "gender": "Male", "institution": "Institute of Software, CAS", "country": "CN", "position": "Professor" }, { "name": "XingYu", "gender": "Male", "institution": "Xiaohongshu", "country": "AF", "position": "Researcher" }, { "name": "Xinyu Lu", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Yaojie Lu", "gender": "Male", "institution": "Institute of Software, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "jie lou", "gender": "Male", "institution": "Xiaohongshu Inc", "country": "CN", "position": "Researcher" }, { "name": "Xueruwen", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 8, "n_ref_uni": 42, "n_ref": 74, "n_ref_all": 116, "n_fig": 26, "n_tab": 9, "L_tab": 2668, "n_element_tab": 278, "n_fig_1": 18, "n_tab_1": 5, "L_tab_1": 2359, "n_element_tab_1": 199, "formula_len_all": 1501, "formula_len_all_1": 519, "len_all": 175828, "len_all_1": 61812, "len_abs": 1160, "len_title": 117, "len_sents": 43125, "len_sents_1": 25578, "n_sents": 360, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 69, "L_abs": 1167, "title": "Rethinking Reward Model Evaluation: Are We Barking up the Wrong Tree?", "abs": "Reward Models (RMs) are crucial for aligning language models with human preferences. \nCurrently, the evaluation of RMs depends on measuring accuracy against a validation set of manually annotated preference data.\nAlthough this method is straightforward and widely adopted, the relationship between RM accuracy and downstream policy performance remains under-explored.\nIn this work, we conduct experiments in a synthetic setting to investigate how differences in RM measured by accuracy translate into gaps in optimized policy performance.\nOur findings reveal that while there is a weak positive correlation between accuracy and downstream performance, policies optimized towards RMs with similar accuracy can exhibit quite different performance.\nMoreover, we discover that the way of measuring accuracy significantly impacts its ability to predict the final policy performance. \nThrough the lens of the Regressional Goodhart effect, we recognize that accuracy, when used for measuring RM quality, can fail to fully capture the potential RM overoptimization.\nThis underscores the inadequacy of relying solely on accuracy to reflect their impact on policy optimization.", "keywords": [ "Reinforcement Learning from Human Feedback; Reward Model;" ], "rating_list": [ 5, 8, 8, 8 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 4, 3, 3 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "Cn5Z0MUPZT", "primary_area": "foundation or frontier models, including LLMs", "n_author": 10, "author_info_list": [ { "name": "Chen Dun", "gender": "unknown", "institution": "ByteDance Inc.", "country": "", "position": "Researcher" }, { "name": "Guanlin Liu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Liang Huang", "gender": "Male", "institution": "Oregon State University", "country": "", "position": "Assistant Professor" }, { "name": "Lin Yan", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Ning Dai", "gender": "Male", "institution": "Oregon State University", "country": "US", "position": "PhD student" }, { "name": "Renjie Zheng", "gender": "Male", "institution": "ByteDance", "country": "", "position": "Researcher" }, { "name": "Wenlei Shi", "gender": "Male", "institution": "Microsoft Research", "country": "", "position": "Researcher" }, { "name": "Xing Jin", "gender": "unknown", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Zheng Wu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Ziyun Wei", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 24, "n_ref": 63, "n_ref_all": 80, "n_fig": 7, "n_tab": 9, "L_tab": 4701, "n_element_tab": 254, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 4130, "n_element_tab_1": 217, "formula_len_all": 791, "formula_len_all_1": 1046, "len_all": 168068, "len_all_1": 63174, "len_abs": 1046, "len_title": 114, "len_sents": 41083, "len_sents_1": 26648, "n_sents": 298, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 1053, "title": "Process Supervision-Guided Policy Optimization for Code Generation", "abs": "Reinforcement learning (RL) with unit test feedback has enhanced large language models\u2019 (LLMs) code generation, but relies on sparse rewards provided only after complete code evaluation, limiting learning efficiency and incremental improvements. When generated code fails all unit tests, no learning signal is received, hindering progress on complex tasks. To address this, we propose a Process Reward Model (PRM) that delivers dense, line-level feedback on code correctness during generation, mimicking human code refinement and providing immediate guidance. We explore various strategies for training PRMs and integrating them into the RL framework, finding that using PRMs both as dense rewards and for value function initialization significantly boosts performance. Our approach increases our in-house LLM\u2019s pass rate from 28.2\\% to 29.8\\% on LiveCodeBench and from 31.8\\% to 35.8\\% on our internal benchmark. Our experimental results highlight the effectiveness of PRMs in enhancing RL-driven code generation, especially for long-horizon scenarios.", "keywords": [ "Process Reward Model", "Code Generation", "Large Language Model", "Reinforcement Learning" ], "rating_list": [ 3, 3, 6, 8 ], "soundness_list": [ 2, 2, 3, 4 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 5, 4, 4, 3 ] }, { "paper_id": "ClkfwM3STw", "primary_area": "datasets and benchmarks", "n_author": 10, "author_info_list": [ { "name": "Chaoyu Guan", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "MS student" }, { "name": "Chen Tang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "FangWu", "gender": "Male", "institution": "tsingmao.com", "country": "CN", "position": "Researcher" }, { "name": "Shenhao Peng", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "MS student" }, { "name": "Wenwu Zhu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Xinzhu Ma", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Postdoc" }, { "name": "Yao Hang", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "MS student" }, { "name": "Yijun Liu", "gender": "Female", "institution": "Xi'an University of Electronic Science and Technology", "country": "", "position": "Undergrad student" }, { "name": "Yuan Meng", "gender": "Female", "institution": "Tsinghua University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhi Wang", "gender": "Male", "institution": "SIGS, Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 1, "n_ref_uni": 69, "n_ref": 202, "n_ref_all": 217, "n_fig": 6, "n_tab": 12, "L_tab": 16568, "n_element_tab": 1279, "n_fig_1": 4, "n_tab_1": 8, "L_tab_1": 11049, "n_element_tab_1": 436, "formula_len_all": 159, "formula_len_all_1": 28, "len_all": 247222, "len_all_1": 77548, "len_abs": 2020, "len_title": 137, "len_sents": 58234, "len_sents_1": 29460, "n_sents": 446, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 89, "L_abs": 1821, "title": "Evaluating the Generalization Ability of Quantized LLMs: Benchmark, Analysis, and Toolbox", "abs": "Large language models (LLMs) have exhibited exciting progress in multiple scenarios, while the huge computational demands hinder their deployments in lots of real-world applications. As an effective means to reduce memory footprint and inference cost, quantization also faces challenges in performance degradation at low bit-widths. Understanding the impact of quantization on LLM capabilities, especially the generalization ability, is crucial. However, the community's main focus remains on the algorithms and models of quantization, with insufficient attention given to to the impact of data on the generalization abilities of quantized LLMs.\nIn this work, we fill this gap by providing a comprehensive benchmark suite for this research topic, including an evaluation system, detailed analyses, and a general toolbox. Specifically, based on the dominant pipeline in LLM quantization, we primarily explore the impact of calibration data distribution on the generalization of quantized LLMs and conduct the benchmark using more than 40 datasets within two main scenarios. Based on this benchmark, we conduct extensive experiments with well-known LLMs (LLaMA and Baichuan) and four quantization algorithms to investigate this topic in-depth, yielding several counter-intuitive and valuable findings, e.g., models quantized using a calibration set with the same distribution as the test data are not necessarily optimal. Besides, to facilitate future research, we also release a modular-designed toolbox, which decouples the overall pipeline into several separate components, e.g., base LLM module, dataset module, quantizer module, etc. and allows subsequent researchers to easily assemble their methods through a simple configuration. \nOur code is submitted in the supplementary materials and will be publicly available.", "keywords": [ "LLM", "Quantization", "Evaluation", "OOD" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 5, 2 ] }, { "paper_id": "ClixrtIHUJ", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Bing Liu", "gender": "Male", "institution": "University of Illinois at Chicago", "country": "US", "position": "Full Professor" }, { "name": "Sahisnu Mazumder", "gender": "Male", "institution": "Intel Labs, USA", "country": "US", "position": "Researcher" }, { "name": "Saleh Momeni", "gender": "Male", "institution": "University of Illinois at Chicago", "country": "US", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 49, "n_ref": 72, "n_ref_all": 81, "n_fig": 1, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 545, "n_element_tab_1": 54, "formula_len_all": 959, "formula_len_all_1": 958, "len_all": 136626, "len_all_1": 67504, "len_abs": 1447, "len_title": 117, "len_sents": 34378, "len_sents_1": 31191, "n_sents": 260, "n_sents_1": 234, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1455, "title": "Language Models as Feature Extractors for Accurate Continual Learning", "abs": "This paper addresses the challenges of class incremental learning (CIL) within the broader context of continual learning. In CIL, a system learns a sequence of tasks or classes incrementally. The resulting classifier can categorize test samples into any learned class thus far without relying on task-specific information during testing. CIL presents two significant challenges: catastrophic forgetting (CF) and inter-task class separation (ICS). ICS occurs because the system lacks data from previous tasks when learning new ones, making it harder to establish decision boundaries between classes, reducing accuracy. This paper proposes a novel method to overcome both CF and ICS. The basic classifier is based on the statistical technique Mahalanobis distance (MD), which measures the distance of a data point to a normal distribution. In the proposed approach, each class is represented by a normal distribution with the mean and covariance derived from the features of its training data, which are extracted from a language model (LM). To reduce storage, all classes share a common covariance matrix. Two additional techniques are also proposed to enhance the accuracy: (1) using a kernel function to expand the feature space, and (2) incorporating an ensemble mechanism. Our experiments show that the proposed method achieves accuracy comparable to the upper bound accuracy of joint fine-tuning, which, to our knowledge, has not been achieved before.", "keywords": [ "class incremental learning", "continual learning with LMs", "distance-based methods" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "CkozFajtKq", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Gavin Winter", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Juno Nam", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "KR", "position": "PhD student" }, { "name": "KyuJung Jun", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Rafael Gomez-Bombarelli", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" }, { "name": "Soojung Yang", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Sulin Liu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" } ], "n_formula": 28, "n_formula_1": 12, "n_ref_uni": 49, "n_ref": 80, "n_ref_all": 146, "n_fig": 16, "n_tab": 8, "L_tab": 2466, "n_element_tab": 244, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 3431, "n_element_tab_1": 138, "formula_len_all": 2968, "formula_len_all_1": 1360, "len_all": 227295, "len_all_1": 68929, "len_abs": 1072, "len_title": 121, "len_sents": 62425, "len_sents_1": 30115, "n_sents": 416, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1075, "title": "Flow Matching for Accelerated Simulation of Atomic Transport in Materials", "abs": "We introduce LiFlow, a generative framework to accelerate molecular dynamics (MD) simulations for crystalline materials that formulates the task as conditional generation of atomic displacements. The model uses flow matching, with a Propagator submodel to generate atomic displacements and a Corrector to locally correct unphysical geometries, and incorporates an adaptive prior based on the Maxwell\u2013Boltzmann distribution to account for chemical and thermal conditions. We benchmark LiFlow on a dataset comprising 25-ps trajectories of lithium diffusion across 4,186 solid-state electrolyte (SSE) candidates at four temperatures. The model obtains a consistent Spearman rank correlation of 0.7\u20130.8 for lithium mean squared displacement (MSD) predictions on unseen compositions. Furthermore, LiFlow generalizes from short training trajectories to larger supercells and longer simulations while maintaining high accuracy. With speed-ups of up to 600,000\u00d7 compared to first-principles methods, LiFlow enables scalable simulations at significantly larger length and time scales.", "keywords": [ "flow matching", "generative models", "atomistic simulations", "molecular dynamics", "materials science" ], "rating_list": [ 6, 5, 8, 8, 3 ], "soundness_list": [ 2, 3, 3, 4, 3 ], "presentation_list": [ 2, 4, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 4, 2 ], "confidence_list": [ 1, 5, 4, 3, 5 ] }, { "paper_id": "CkoomnLfpS", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Di Wu", "gender": "Male", "institution": "University of Toronto", "country": "", "position": "Undergrad student" }, { "name": "Huan Liu", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CA", "position": "Researcher" }, { "name": "YUANHAO YU", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "CA", "position": "Principal Researcher" }, { "name": "Yang Wang", "gender": "Male", "institution": "Concordia University", "country": "CA", "position": "Associate Professor" }, { "name": "Zhixiang Chi", "gender": "Male", "institution": "University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Kostantinos Plataniotis", "gender": "Male", "institution": "Toronto University", "country": "CA", "position": "Full Professor" } ], "n_formula": 55, "n_formula_1": 16, "n_ref_uni": 22, "n_ref": 67, "n_ref_all": 109, "n_fig": 7, "n_tab": 12, "L_tab": 3598, "n_element_tab": 260, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1529, "n_element_tab_1": 89, "formula_len_all": 3339, "formula_len_all_1": 1057, "len_all": 217233, "len_all_1": 76749, "len_abs": 1071, "len_title": 130, "len_sents": 73667, "len_sents_1": 31550, "n_sents": 566, "n_sents_1": 241, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1077, "title": "Dynamic Neural Graph: Facilitating Temporal Dynamics Learning in Deep Weight Space", "abs": "The rapid advancements in using neural networks as implicit data representations have attracted significant interest in developing machine learning methods that analyze and process the weight spaces of other neural networks. However, efficiently handling these high-dimensional weight spaces remains challenging. Existing methods often overlook the sequential nature of layer-by-layer processing in neural network inference. In this work, we propose a novel approach using dynamic graphs to represent neural network parameters, capturing the temporal dynamics of inference. Our Dynamic Neural Graph Encoder (DNG-Encoder) processes these graphs, preserving the sequential nature of neural processing. Additionally, we also leverage DNG-Encoder to develop INR2JLS for facilitate downstream applications, such as classifying INRs. Our approach demonstrates significant improvements across multiple tasks, surpassing the state-of-the-art INR classification accuracy by approximately 10% on the CIFAR-100-INR. The source\ncode has been made available in the supplementary materials.", "keywords": [ "Dynamic Graph neural networks", "Deep weight space", "Implicit neural representations", "Networks for networks", "Neural graphs" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 2, 3, 3 ] }, { "paper_id": "CkgKSqZbuC", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Bo Li", "gender": "Female", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Assistant Professor" }, { "name": "Mintong Kang", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "", "position": "PhD student" } ], "n_formula": 18, "n_formula_1": 16, "n_ref_uni": 52, "n_ref": 108, "n_ref_all": 132, "n_fig": 4, "n_tab": 10, "L_tab": 2617, "n_element_tab": 421, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 1119, "n_element_tab_1": 189, "formula_len_all": 1445, "formula_len_all_1": 1540, "len_all": 166416, "len_all_1": 75125, "len_abs": 2542, "len_title": 143, "len_sents": 50665, "len_sents_1": 36532, "n_sents": 322, "n_sents_1": 212, "venue": "ICLR2025", "decision": "Accept (Spotlight)", "L_title": 92, "L_abs": 2408, "title": "$R^2$-Guard: Robust Reasoning Enabled LLM Guardrail via Knowledge-Enhanced Logical Reasoning", "abs": "As large language models (LLMs) become increasingly prevalent across various applications, it is critical to establish safety guardrails to moderate input/output content of LLMs and ensure compliance with safety policies. Existing guardrail models, such as OpenAI Mod and LlamaGuard, treat various safety categories (e.g., self-harm, self-harm/instructions) independently and fail to explicitly capture the intercorrelations among them. This has led to limitations such as ineffectiveness due to inadequate training on long-tail data from correlated safety categories, susceptibility to jailbreaking attacks, and inflexibility regarding new safety categories.\nTo address these limitations, we propose $R^2$-Guard, a robust reasoning enabled LLM guardrail via knowledge-enhanced logical reasoning. Specifically, $R^2$-Guard comprises two parts: data-driven guardrail models and reasoning components. The data-driven guardrail models provide unsafety probabilities of moderated content on different safety categories.\nWe then encode safety knowledge among different categories as first-order logical rules and embed them into a probabilistic graphic model (PGM) based reasoning component. The unsafety probabilities of different categories from data-driven guardrail models are sent to the reasoning component for final inference. We employ two types of PGMs: Markov logic networks (MLNs) and probabilistic circuits (PCs), and optimize PCs to achieve precision-efficiency balance via improved graph structure. We also propose different methods to optimize the weights of knowledge. To further perform stress tests for guardrail models, we employ a pairwise construction method to construct a new safety benchmark TwinSafety, which features principled categories and presents new challenges for moderation. We show that $R^2$-Guard is effective even given unrepresentative categories or challenging jailbreaking prompts. We demonstrate the effectiveness of $R^2$-Guard by comparisons with eight strong guardrail models on six standard moderation datasets, and demonstrate the robustness of $R^2$-Guard against four SOTA jailbreaking attacks. $R^2$-Guard significantly surpasses SOTA method LlamaGuard by 12.6% on standard moderation datasets and by 59.9% against jailbreaking attacks.\nWe further reveal that $R^2$-Guard can effectively adapt to safety category updates by simply editing the PGM reasoning graph.", "keywords": [ "LLM guardrail model", "content moderation" ], "rating_list": [ 6, 8, 8 ], "soundness_list": [ 3, 4, 3 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 3, 4, 3 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "CkUHtnyhpY", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Haim Sompolinsky", "gender": "Male", "institution": "Harvard University", "country": "", "position": "Full Professor" }, { "name": "Zechen Zhang", "gender": "Male", "institution": "Harvard University", "country": "US", "position": "PhD student" } ], "n_formula": 42, "n_formula_1": 23, "n_ref_uni": 28, "n_ref": 40, "n_ref_all": 63, "n_fig": 14, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2626, "formula_len_all_1": 1281, "len_all": 132976, "len_all_1": 49730, "len_abs": 3331, "len_title": 141, "len_sents": 38289, "len_sents_1": 20727, "n_sents": 281, "n_sents_1": 147, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 92, "L_abs": 1354, "title": "Robust Learning in Bayesian Parallel Branching Graph Neural Networks: The Narrow Width Limit", "abs": "The infinite width limit of random neural networks is known to result in Neural Networks as Gaussian Process (NNGP) (Lee et al. (2018)), characterized by task-independent kernels. It is widely accepted that larger network widths contribute to improved generalization (Park et al. (2019)). However, this work challenges this notion by investigating the narrow width limit of the Bayesian Parallel Branching Graph Neural Network (BPB-GNN), an architecture that resembles residual-GCN. We demonstrate that when the width of a BPB-GNN is significantly smaller compared to the number of training examples, each branch exhibits more robust learning due to a symmetry breaking of branches in kernel renormalization. Surprisingly, the performance of a BPB-GNN in the narrow width limit is generally superior to or comparable to that achieved in the wide width limit in bias-limited scenarios. Furthermore, the readout norms of each branch in the narrow width limit are mostly independent of the architectural hyperparameters but generally reflective of the nature of the data. We also extend the results to other more general architectures such as the residual-MLP and demonstrate that the narrow width effect is a general feature of the branching networks. Our results characterize a newly defined narrow-width regime for parallel branching networks in general.", "keywords": [ "Bayesian Networks", "Gaussian Process", "Kernel Renormalization", "Graph Neural Networks", "Residual Network", "Theory of Generalization" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 2, 4 ] }, { "paper_id": "CkKEuLmRnr", "primary_area": "datasets and benchmarks", "n_author": 9, "author_info_list": [ { "name": "Bohang Zhang", "gender": "Male", "institution": "Peking University", "country": "", "position": "PhD student" }, { "name": "Caihua Shan", "gender": "Female", "institution": "Microsoft", "country": "CN", "position": "Researcher" }, { "name": "Dongsheng Li", "gender": "Male", "institution": "Microsoft Research Asia", "country": "CN", "position": "Principal Researcher" }, { "name": "Haohao Qu", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "CN", "position": "PhD student" }, { "name": "Jiliang Tang", "gender": "Male", "institution": "Michigan State University", "country": "US", "position": "Full Professor" }, { "name": "Qihao Wen", "gender": "Male", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Wenqi Fan", "gender": "Male", "institution": "Hong Kong Polytechnic University", "country": "HK", "position": "Assistant Professor" }, { "name": "Xinnan Dai", "gender": "Female", "institution": "Michigan State University", "country": "", "position": "PhD student" }, { "name": "Yifei Shen", "gender": "Male", "institution": "Microsoft Research Asia", "country": "ZW", "position": "Research Cheerleader" } ], "n_formula": 11, "n_formula_1": 0, "n_ref_uni": 44, "n_ref": 82, "n_ref_all": 124, "n_fig": 7, "n_tab": 21, "L_tab": 9189, "n_element_tab": 827, "n_fig_1": 6, "n_tab_1": 7, "L_tab_1": 2030, "n_element_tab_1": 70, "formula_len_all": 958, "formula_len_all_1": 0, "len_all": 220554, "len_all_1": 66256, "len_abs": 1592, "len_title": 148, "len_sents": 65511, "len_sents_1": 30548, "n_sents": 507, "n_sents_1": 235, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 99, "L_abs": 1450, "title": "How Do Large Language Models Understand Graph Patterns? A Benchmark for Graph Pattern Comprehension", "abs": "Benchmarking the capabilities and limitations of large language models (LLMs) in graph-related tasks is becoming an increasingly popular and crucial area of research. Recent studies have shown that LLMs exhibit a preliminary ability to understand graph structures and node features. However, the potential of LLMs in graph pattern mining remains largely unexplored. This is a key component in fields such as computational chemistry, biology, and social network analysis. To bridge this gap, this work introduces a comprehensive benchmark to assess LLMs' capabilities in graph pattern tasks. We have developed a benchmark that evaluates whether LLMs can understand graph patterns based on either terminological or topological descriptions. Additionally, our benchmark tests the LLMs' capacity to autonomously discover graph patterns from data. The benchmark encompasses both synthetic and real datasets, and a variety of models, with a total of 11 tasks and 7 models. Our experimental framework is designed for easy expansion to accommodate new models and datasets. Our findings reveal that: (1) LLMs have preliminary abilities to understand graph patterns, with O1-mini outperforming in the majority of tasks; (2) Formatting input graph data to align with the knowledge acquired during pretraining can enhance performance; (3) LLMs employ diverse\npotential algorithms to solve one task, with performance varying based on their execution capabilities.", "keywords": [ "Large language models", "graph pattern", "graph mining" ], "rating_list": [ 6, 8, 6, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 1, 3, 3, 4 ], "confidence_list": [ 4, 5, 4, 4 ] }, { "paper_id": "CkCFoN3j4s", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 5, "author_info_list": [ { "name": "Binhang Yuan", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Assistant Professor" }, { "name": "Chaojun Xiao", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Xu Han", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuxiang Huang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Zhiyuan Liu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 16, "n_ref_uni": 47, "n_ref": 107, "n_ref_all": 151, "n_fig": 14, "n_tab": 14, "L_tab": 4678, "n_element_tab": 715, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1036, "n_element_tab_1": 161, "formula_len_all": 1922, "formula_len_all_1": 1063, "len_all": 221145, "len_all_1": 63959, "len_abs": 1968, "len_title": 133, "len_sents": 59470, "len_sents_1": 28753, "n_sents": 491, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1992, "title": "Locret: Enhancing Eviction in Long-Context LLM Inference with Trained Retaining Heads", "abs": "Large language models (LLMs) have shown remarkable advances in supporting long-context comprehension and processing tasks. However, scaling the generation inference of LLMs to such long contexts incurs significant additional computation load, and demands a substantial GPU memory footprint to maintain the key-value (KV) cache of transformer-based LLMs. Existing KV cache compression methods, such as quantization, face memory bottlenecks as context length increases, while static-sized caches, such as selective eviction, suffer from inefficient policies. These limitations restrict deployment on consumer-grade devices like a single Nvidia 4090 GPU. To overcome this, we propose Locret, an efficient framework for long-context LLM inference that introduces retaining heads to evaluate the causal importance of KV cache units, allowing for more accurate eviction within a fixed cache size. Locret is fine-tuned on top of the frozen backbone LLM using a minimal amount of data from standard long-context SFT datasets. During inference, we evict low-importance cache units along with a chunked prefill pattern, significantly reducing peak GPU memory usage. We conduct an extensive empirical study to evaluate Locret, where the experimental results show that Locret outperforms the recent popular and competitive approaches, including InfLLM, Quantization, SirLLM, and MInference, in terms of memory efficiency and the quality of generated contents --- Locret achieves over a $20\\times$ and $8\\times$ KV cache compression ratio compared to the full KV cache for Phi-3-mini-128K and Llama-3.1-8B-instruct. Additionally, Locret can be combined with other efficient inference methods, such as quantization and token merging. To the best of our knowledge, Locret is the first framework capable of deploying Llama-3.1-8B or similar models on a single Nvidia 4090 GPU, enabling 128K long-context inference without compromising generation quality, and requiring little additional system optimizations.", "keywords": [ "Long-context Inference", "Memory Efficient Inference", "Large Language Models" ], "rating_list": [ 8, 5, 5, 3, 8 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3, 3 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 5, 4, 5, 4 ] }, { "paper_id": "Cjz9Xhm7sI", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 4, "author_info_list": [ { "name": "Lin Zeng", "gender": "Female", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Intern" }, { "name": "Ruimao Zhang", "gender": "Male", "institution": "SUN YAT-SEN UNIVERSITY", "country": "CN", "position": "Associate Professor" }, { "name": "Yiran Qin", "gender": "Male", "institution": "The Chinese University of Hong Kong(Shenzhen)", "country": "CN", "position": "PhD student" }, { "name": "Ziye Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong (Shenzhen))", "country": "CN", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 13, "n_ref": 24, "n_ref_all": 43, "n_fig": 9, "n_tab": 6, "L_tab": 1723, "n_element_tab": 89, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 389, "n_element_tab_1": 68, "formula_len_all": 450, "formula_len_all_1": 193, "len_all": 137694, "len_all_1": 56989, "len_abs": 1788, "len_title": 163, "len_sents": 45004, "len_sents_1": 27431, "n_sents": 306, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 115, "L_abs": 1802, "title": "High-Dynamic Radar Sequence Prediction for Weather Nowcasting Using Spatiotemporal Coherent Gaussian Representation", "abs": "Weather nowcasting is an essential task that involves predicting future radar echo sequences based on current observations, offering significant benefits for disaster management, transportation, and urban planning. Current prediction methods are limited by training and storage efficiency, mainly focusing on 2D spatial predictions at specific altitudes. Meanwhile, 3D volumetric predictions at each timestamp remain largely unexplored. To address such a challenge, we introduce a comprehensive framework for 3D radar sequence prediction in weather nowcasting, using the newly proposed SpatioTemporal Coherent Gaussian Splatting (STC-GS) for dynamic radar representation and GauMamba for efficient and accurate forecasting. Specifically, rather than relying on a 4D Gaussian for dynamic scene reconstruction, STC-GS optimizes 3D scenes at each frame by employing a group of Gaussians while effectively capturing their movements across consecutive frames. It ensures consistent tracking of each Gaussian over time, making it particularly effective for prediction tasks. With the temporally correlated Gaussian groups established, we utilize them to train GauMamba, which integrates a memory mechanism into the Mamba framework. This allows the model to learn the temporal evolution of Gaussian groups while efficiently handling a large volume of Gaussian tokens. As a result, it achieves both efficiency and accuracy in forecasting a wide range of dynamic meteorological radar signals. The experimental results demonstrate that our STC-GS can efficiently represent 3D radar sequences with over $16\\times$ higher spatial resolution compared with the existing 3D representation methods, while GauMamba outperforms state-of-the-art methods in forecasting a broad spectrum of high-dynamic weather conditions.", "keywords": [ "3D Gaussian", "Dynamic Reconstruction", "Radar Prediction", "Weather Nowcasting" ], "rating_list": [ 8, 8, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 3, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 4, 4, 2 ] }, { "paper_id": "CjwERcAU7w", "primary_area": "foundation or frontier models, including LLMs", "n_author": 18, "author_info_list": [ { "name": "Aleksandra Faust", "gender": "Female", "institution": "Google Brain", "country": "US", "position": "Principal Researcher" }, { "name": "Avi Singh", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Aviral Kumar", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Colton Bishop", "gender": "Male", "institution": "Princeton University", "country": "", "position": "Undergrad student" }, { "name": "Cosmin Paduraru", "gender": "Male", "institution": "DeepMind", "country": "", "position": "Research Engineer" }, { "name": "Disha Shrivastava", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Doina Precup", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Research Team Lead" }, { "name": "Feryal Behbahani", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Research Scientist" }, { "name": "George Tucker", "gender": "Male", "institution": "Google Brain", "country": "", "position": "Research Scientist" }, { "name": "John D Co-Reyes", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "PhD student" }, { "name": "Kate Baumli", "gender": "Female", "institution": "DeepMind", "country": "", "position": "Researcher" }, { "name": "Kay McKinney", "gender": "unknown", "institution": "Google", "country": "", "position": "Researcher" }, { "name": "Lei M Zhang", "gender": "Male", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Rebecca Roelofs", "gender": "Female", "institution": "Google", "country": "", "position": "Research scientist" }, { "name": "Rishabh Agarwal", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Adjunct Professor" }, { "name": "Shariq Iqbal", "gender": "Male", "institution": "Google DeepMind", "country": "GB", "position": "Researcher" }, { "name": "Vincent Zhuang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Research Engineer" }, { "name": "Yi Su", "gender": "Female", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 19, "n_formula_1": 5, "n_ref_uni": 40, "n_ref": 107, "n_ref_all": 141, "n_fig": 13, "n_tab": 10, "L_tab": 2767, "n_element_tab": 315, "n_fig_1": 10, "n_tab_1": 4, "L_tab_1": 1124, "n_element_tab_1": 120, "formula_len_all": 1548, "formula_len_all_1": 484, "len_all": 197261, "len_all_1": 72735, "len_abs": 1787, "len_title": 115, "len_sents": 60374, "len_sents_1": 34868, "n_sents": 540, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 67, "L_abs": 1734, "title": "Training Language Models to Self-Correct via Reinforcement Learning", "abs": "Self-correction is a highly desirable capability of large language models (LLMs), yet it has consistently been found to be largely ineffective in modern LLMs. Current methods for training self-correction typically depend on either multiple models, a more advanced model, or additional forms of supervision. To address these shortcomings, we develop a multi-turn online reinforcement learning (RL) approach, SCoRe, that significantly improves an LLM's self-correction ability using entirely self-generated data. To build SCoRe, we first show that variants of supervised fine-tuning (SFT) on offline model-generated correction traces are insufficient for instilling self-correction behavior. In particular, we observe that training via SFT either suffers from a distribution mismatch between the training data and the model's own responses or implicitly prefers only a certain mode of correction behavior that is often not effective at test time. SCoRe addresses these challenges by training under the model's own distribution of self-generated correction traces and using appropriate regularization to steer the learning process into learning a self-correction strategy that is effective at test time as opposed to simply fitting high-reward responses for a given prompt. This regularization prescribes running a first phase of RL on a base model to generate a policy initialization that is less susceptible to collapse and then using a reward bonus to amplify self-correction during training. When applied to Gemini 1.0 Pro and 1.5 Flash models, we find that SCoRe achieves state-of-the-art self-correction performance, improving the base models' self-correction by 15.6% and 9.1% respectively on the MATH and HumanEval benchmarks.", "keywords": [ "language models", "reinforcement learning" ], "rating_list": [ 8, 8, 8, 8 ], "soundness_list": [ 4, 3, 3, 4 ], "presentation_list": [ 4, 3, 3, 4 ], "contribution_list": [ 4, 3, 3, 4 ], "confidence_list": [ 4, 2, 3, 4 ] }, { "paper_id": "CjfQssZtAb", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Aviral Kumar", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Hao Bai", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "MS student" }, { "name": "Li Erran Li", "gender": "unknown", "institution": "Amazon", "country": "US", "position": "Head of Science, HIL" }, { "name": "Sergey Levine", "gender": "Male", "institution": "University of California Berkeley", "country": "", "position": "Associate Professor" }, { "name": "Yifei Zhou", "gender": "Male", "institution": "University of California, Berkeley", "country": "", "position": "PhD student" } ], "n_formula": 15, "n_formula_1": 14, "n_ref_uni": 47, "n_ref": 135, "n_ref_all": 161, "n_fig": 8, "n_tab": 3, "L_tab": 613, "n_element_tab": 73, "n_fig_1": 3, "n_tab_1": 1, "L_tab_1": 328, "n_element_tab_1": 46, "formula_len_all": 953, "formula_len_all_1": 927, "len_all": 173592, "len_all_1": 73486, "len_abs": 1775, "len_title": 125, "len_sents": 50586, "len_sents_1": 37285, "n_sents": 350, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 77, "L_abs": 1782, "title": "Digi-Q: Transforming VLMs to Device-Control Agents via Value-Based Offline RL", "abs": "Most paradigms for building foundation model agents rely on prompting or finetuning on existing demonstrations, but this is not sufficient in dynamic environments (e.g., mobile device control). In theory, while on-policy reinforcement learning (RL) should address these limitations, this approach itself is not quite effective at leveraging existing agentic data, especially when it is of low quality. An approach to address this issue is to use offline value-based RL but realizing value-based RL for agents has been elusive due to of stability and efficiency associated with running TD-learning at scale with vision-language models (VLMs). In this paper, we develop a scalable value-based RL approach called Digi-Q that makes it possible to train VLM agents with TD-learning. We situate our study in building GUI agents for Android devices. The key idea in Digi-Q is to perform TD-learning on a frozen, intermediate-layer representation of a VLM rather than training the whole VLM itself. Doing so successfully requires an initial phase of fine-tuning to prime VLM representations to feature actionable information that is critical for TD-learning. When done correctly, our approach is able to attain better performance per-unit compute FLOPS. To make maximal use of the learned Q-function, we devise a novel best-of-N policy extraction operator that imitates the best actions out of multiple candidate actions from the current policy as ranked by the value function. With no REINFORCE-style policy gradients that need careful tiuning and an efficient TD-learning approach, Digi-Q outperforms several strong prior methods on user-scale device control tasks in Android-in-the-Wild, attaining 9.9% of relative improvement over prior best-performing offline RL method in this domain.", "keywords": [ "Reinforcement learning", "device control", "digital agents", "foundation models" ], "rating_list": [ 3, 6, 6, 6, 3 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 4, 3, 1 ], "contribution_list": [ 1, 2, 3, 3, 2 ], "confidence_list": [ 5, 3, 3, 4, 4 ] }, { "paper_id": "CjXaMI2kUH", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Junmo Cho", "gender": "Male", "institution": "KAIST", "country": "", "position": "PhD student" }, { "name": "Junyeong Park", "gender": "Male", "institution": "KAIST", "country": "", "position": "MS student" }, { "name": "Sungjin Ahn", "gender": "Male", "institution": "KAIST", "country": "US", "position": "Associate Professor" } ], "n_formula": 17, "n_formula_1": 8, "n_ref_uni": 35, "n_ref": 99, "n_ref_all": 142, "n_fig": 18, "n_tab": 12, "L_tab": 3961, "n_element_tab": 356, "n_fig_1": 11, "n_tab_1": 3, "L_tab_1": 234, "n_element_tab_1": 16, "formula_len_all": 10266, "formula_len_all_1": 4281, "len_all": 279747, "len_all_1": 85170, "len_abs": 1317, "len_title": 127, "len_sents": 91196, "len_sents_1": 40244, "n_sents": 686, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 79, "L_abs": 1322, "title": "Mr.Steve: Instruction-Following Agents in Minecraft with What-Where-When Memory", "abs": "Significant advances have been made in developing general-purpose embodied AI in environments like Minecraft through the adoption of LLM-augmented hierarchical approaches. While these approaches, which combine high-level planners with low-level controllers, show promise, low-level controllers frequently become performance bottlenecks due to repeated failures. In this paper, we argue that the primary cause of failure in many low-level controllers is the absence of an episodic memory system. To address this, we introduce Mr.Steve (Memory Recall STEVE-1), a novel low-level controller equipped with Place Event Memory (PEM), a form of episodic memory that captures what, where, and when information from episodes. This directly addresses the main limitation of the popular low-level controller, STEVE-1. Unlike previous models that rely on short-term memory, PEM organizes spatial and event-based data, enabling efficient recall and navigation in long-horizon tasks. Additionally, we propose an Exploration Strategy and a Memory-Augmented Task Solving Framework, allowing agents to alternate between exploration and task-solving based on recalled events. Our approach significantly improves task-solving and exploration efficiency compared to existing methods, and we are releasing our code to support further research.", "keywords": [ "Generalist Agents", "Minecraft", "Place Event Memory" ], "rating_list": [ 6, 6, 6, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 4, 4 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "Cj3B4SoWuT", "primary_area": "generative models", "n_author": 4, "author_info_list": [ { "name": "David Barber", "gender": "Male", "institution": "University College London", "country": "", "position": "Associate Professor" }, { "name": "Lucas Maystre", "gender": "Male", "institution": "UiPath", "country": "GB", "position": "Researcher" }, { "name": "Marius Cobzarenco", "gender": "Not Specified", "institution": "University of Cambridge", "country": "", "position": "Undergrad student" }, { "name": "Rares Dolga", "gender": "unknown", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 22, "n_ref_uni": 41, "n_ref": 63, "n_ref_all": 102, "n_fig": 15, "n_tab": 10, "L_tab": 3098, "n_element_tab": 113, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 2593, "n_element_tab_1": 150, "formula_len_all": 2054, "formula_len_all_1": 1475, "len_all": 158035, "len_all_1": 60708, "len_abs": 920, "len_title": 100, "len_sents": 40673, "len_sents_1": 23428, "n_sents": 355, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 925, "title": "Latte: Latent Attention for Linear Time Transformers", "abs": "The time complexity of the standard attention mechanism in transformers scales quadratically with sequence length. We propose a probabilistic framework for attention, enabling us to derive a novel low-rank linear re-parameterisation of both bidirectional and causal cases, based on defining a latent variable model. Our method can be seamlessly integrated as a drop-in replacement for the standard attention mechanism. Additionally, this framework provides a natural extension for combining local standard attention with our global linear attention. This approach allows us to extend the context length of existing large pre-trained models with only a few additional training steps. The resulting ``Latte Transformer'' achieves performance comparable to standard attention and other state-of-the-art models, while maintaining linear time and memory complexity, along with constant-time next-token prediction during inference.", "keywords": [ "Sequence Modelling", "Long Sequences", "Linear Attention", "Latent Variable Model" ], "rating_list": [ 1, 5, 3, 5 ], "soundness_list": [ 1, 2, 2, 2 ], "presentation_list": [ 4, 2, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "CiiLchbRe3", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Fupeng Sun", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Guanting Chen", "gender": "Male", "institution": "University of North Carolina at Chapel Hill", "country": "US", "position": "Assistant Professor" }, { "name": "Hanzhao Wang", "gender": "Male", "institution": "Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "KALYAN TEJA TALLURI", "gender": "Male", "institution": "Imperial College", "country": "", "position": "Full Professor" }, { "name": "Shang Liu", "gender": "Male", "institution": "Imperial College London, Imperial College London", "country": "GB", "position": "PhD student" }, { "name": "Xiaocheng Li", "gender": "Male", "institution": "Imperial College London", "country": "", "position": "Assistant Professor" }, { "name": "Yu Pan", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "CN", "position": "Researcher" } ], "n_formula": 77, "n_formula_1": 29, "n_ref_uni": 55, "n_ref": 107, "n_ref_all": 155, "n_fig": 23, "n_tab": 8, "L_tab": 957, "n_element_tab": 128, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 5927, "formula_len_all_1": 1639, "len_all": 270801, "len_all_1": 64095, "len_abs": 1700, "len_title": 153, "len_sents": 95646, "len_sents_1": 30357, "n_sents": 808, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 102, "L_abs": 1706, "title": "Understanding the Training and Generalization of Pretrained Transformer for Sequential Decision Making", "abs": "In this paper, we consider the supervised pre-trained transformer for a class of sequential decision-making problems. The class of considered problems is a subset of the general formulation of reinforcement learning in that there is no transition probability matrix; though seemingly restrictive, the subset class of problems covers bandits, dynamic pricing, and newsvendor problems as special cases. Such a structure enables the use of optimal actions/decisions in the pre-training phase, and the usage also provides new insights for the training and generalization of the pre-trained transformer. We first note the training of the transformer model can be viewed as a performative prediction problem, and the existing methods and theories largely ignore or cannot resolve an out-of-distribution issue. We propose a natural solution that includes the transformer-generated action sequences in the training procedure, and it enjoys better properties both numerically and theoretically. The availability of the optimal actions in the considered tasks also allows us to analyze the properties of the pre-trained transformer as an algorithm and explains why it may lack exploration and how this can be automatically resolved. Numerically, we categorize the advantages of pre-trained transformers over the structured algorithms such as UCB and Thompson sampling into three cases: (i) it better utilizes the prior knowledge in the pre-training data; (ii) it can elegantly handle the misspecification issue suffered by the structured algorithms; (iii) for short time horizon such as $T\\le50$, it behaves more greedy and enjoys much better regret than the structured algorithms designed for asymptotic optimality.", "keywords": [ "pretrained transformer", "in-context learning", "bandits", "dynamic pricing", "sequential decision making" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 3, 3, 4 ], "presentation_list": [ 1, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 3, 2, 2, 3 ] }, { "paper_id": "CiEOW1CdKc", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Alex Schwing", "gender": "Unspecified", "institution": "University of Illinois, Urbana Champaign", "country": "US", "position": "Associate Professor" }, { "name": "Kai Yan", "gender": "Male", "institution": "University of Illinois at Urbana-Champaign", "country": "US", "position": "PhD student" }, { "name": "Siqi Yang", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yu-Xiong Wang", "gender": "unknown", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 18, "n_ref_uni": 66, "n_ref": 179, "n_ref_all": 207, "n_fig": 10, "n_tab": 7, "L_tab": 2333, "n_element_tab": 176, "n_fig_1": 7, "n_tab_1": 2, "L_tab_1": 619, "n_element_tab_1": 65, "formula_len_all": 1525, "formula_len_all_1": 1209, "len_all": 172738, "len_all_1": 64331, "len_abs": 1354, "len_title": 97, "len_sents": 46522, "len_sents_1": 28377, "n_sents": 372, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 49, "L_abs": 1212, "title": "Latent Wasserstein Adversarial Imitation Learning", "abs": "Imitation Learning (IL) enables agents to mimic expert behavior by learning from demonstrations. However, traditional IL methods require large amounts of medium-to-high-quality demonstrations as well as actions of expert demonstrations, both of which are often unavailable. To address these limitations, we propose LWAIL (Latent Wasserstein Adversarial Imitation Learning), a novel adversarial imitation learning framework that focuses on state-only distribution matching by leveraging the Wasserstein distance computed in a latent space. To obtain a meaningful latent space, our approach includes a pre-training stage, where we employ the Intention Conditioned Value Function (ICVF) model to capture the underlying structure of the state space using randomly generated state-only data. This enhances the policy's understanding of state transitions, enabling the learning process to use only one or a few state-only expert episodes to achieve expert-level performance. Through experiments on multiple MuJoCo environments, we demonstrate that our method outperforms prior Wasserstein-based IL methods and prior adversarial IL methods, achieving better sample efficiency and policy robustness across various tasks.", "keywords": [ "Adversarial Imitation Learning", "Wasserstein Distance" ], "rating_list": [ 5, 5, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 1, 3, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 4, 4, 3 ] }, { "paper_id": "Chq4OQ3p18", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Bart\u0142omiej Twardowski", "gender": "Male", "institution": "IDEAS NCBR", "country": "PL", "position": "Researcher" }, { "name": "Damian S\u00f3jka", "gender": "Male", "institution": "Technical University of Poznan", "country": "", "position": "PhD student" }, { "name": "Sebastian Cygert", "gender": "Male", "institution": "IDEAS NCBR", "country": "", "position": "Postdoc" }, { "name": "Marc Masana Castrillo", "gender": "Male", "institution": "Technische Universit\u00e4t Graz", "country": "AT", "position": "Postdoc" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 53, "n_ref": 138, "n_ref_all": 172, "n_fig": 6, "n_tab": 13, "L_tab": 5258, "n_element_tab": 695, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 2895, "n_element_tab_1": 334, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 180036, "len_all_1": 68169, "len_abs": 1068, "len_title": 42, "len_sents": 47578, "len_sents_1": 29637, "n_sents": 366, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 1037, "title": "Intransigent Teachers Guide Better Test-Time Adaptation Students", "abs": "Test-Time Adaptation (TTA) has recently emerged as a promising strategy that allows the adaptation of pre-trained models to changing data distributions at deployment time, without access to any labels. To address the error accumulation problem, various approaches have used the teacher-student framework. In this work, we challenge the common strategy of setting the teacher weights to be an exponential moving average of the student by showing that error accumulation still occurs, but only on longer sequences compared to those commonly utilized. We analyze the stability-plasticity trade-off within the teacher-student framework and propose to use an intransigent teacher instead. We show that not changing any of the weights of the teacher model within existing TTA methods allows them to significantly improve their performance on multiple datasets with longer scenarios and smaller batch sizes. Finally, we show that the proposed changes are applicable to different architectures and are more robust to changes in hyper-parameters.", "keywords": [ "test-time adaptation" ], "rating_list": [ 5, 3, 8, 5 ], "soundness_list": [ 3, 1, 3, 2 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 2, 1, 3, 1 ], "confidence_list": [ 5, 5, 4, 3 ] }, { "paper_id": "Ch9rHRomYe", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 6, "author_info_list": [ { "name": "Romain Hennequin", "gender": "Male", "institution": "Deezer Research", "country": "", "position": "Researcher" }, { "name": "Thomas Bouab\u00e7a", "gender": "Male", "institution": "Deezer", "country": "FR", "position": "Researcher" }, { "name": "Th\u00e9o Bontempelli", "gender": "Male", "institution": "Deezer", "country": "", "position": "Researcher" }, { "name": "Tristan Cazenave", "gender": "Male", "institution": "Universit\u00e9 Paris-Dauphine (Paris IX)", "country": "FR", "position": "Full Professor" }, { "name": "Walid Bendada", "gender": "Male", "institution": "Universit\u00e9 Paris-Dauphine (Paris IX)fr", "country": "FR", "position": "PhD student" }, { "name": "Guillaume Salha", "gender": "Male", "institution": "Kibo Ryoku Research", "country": "JP", "position": "Principal Researcher" } ], "n_formula": 138, "n_formula_1": 20, "n_ref_uni": 58, "n_ref": 136, "n_ref_all": 175, "n_fig": 10, "n_tab": 2, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12506, "formula_len_all_1": 1424, "len_all": 204492, "len_all_1": 66345, "len_abs": 1191, "len_title": 90, "len_sents": 59184, "len_sents_1": 31127, "n_sents": 493, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Reject", "L_title": 90, "L_abs": 1196, "title": "Exploring Large Action Sets with Hyperspherical Embeddings using von Mises-Fisher Sampling", "abs": "This paper introduces von Mises-Fisher exploration (vMF-exp), a scalable method for exploring large action sets in reinforcement learning problems where hyperspherical embedding vectors represent actions. vMF-exp involves initially sampling a state embedding representation using a von Mises-Fisher distribution, then exploring this representation's nearest neighbors, which scales to virtually unlimited numbers of candidate actions.\nWe show that, under theoretical assumptions, vMF-exp asymptotically maintains the same probability of exploring each action as Boltzmann Exploration (B-exp), a popular alternative that, nonetheless, suffers from scalability issues as it requires computing softmax values for each action.\nConsequently, vMF-exp serves as a scalable alternative to B-exp for exploring large action sets with hyperspherical embeddings. \nIn the final part of this paper, we further validate the empirical relevance of vMF-exp by discussing its successful deployment at scale on a music streaming service. On this service, vMF-exp has been employed for months to recommend playlists inspired by initial songs to millions of users, from millions of possible actions for each playlist.", "keywords": [ "Exploration", "Directional Statistics", "Hyperspherical Embeddings", "Reinforcement Learning", "Scalability", "von Mises-Fisher Distribution", "Recommender Systems" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 4, 3, 2, 2 ], "confidence_list": [ 2, 2, 3, 2 ] }, { "paper_id": "Ch8s4FdUXS", "primary_area": "interpretability and explainable AI", "n_author": 6, "author_info_list": [ { "name": "Caglar Gulcehre", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "EPFL" }, { "name": "Chris Wendler", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Postdoc" }, { "name": "Justin Deschenaux", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Mikhail Terekhov", "gender": "unknown", "institution": "School of Computer and Communication Sciences, EPFL - EPF Lausanne", "country": "CH", "position": "PhD student" }, { "name": "Robert West", "gender": "Male", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "Associate Professor" }, { "name": "Viacheslav Surkov", "gender": "unknown", "institution": "EPFL - EPF Lausanne", "country": "CH", "position": "MS student" } ], "n_formula": 24, "n_formula_1": 14, "n_ref_uni": 58, "n_ref": 129, "n_ref_all": 180, "n_fig": 10, "n_tab": 3, "L_tab": 610, "n_element_tab": 59, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 557, "n_element_tab_1": 9, "formula_len_all": 1087, "formula_len_all_1": 602, "len_all": 240960, "len_all_1": 74641, "len_abs": 3396, "len_title": 128, "len_sents": 72184, "len_sents_1": 33620, "n_sents": 717, "n_sents_1": 299, "venue": "ICLR2025", "decision": "withdraw", "L_title": 80, "L_abs": 1188, "title": "Unpacking SDXL Turbo: Interpreting Text-to-Image Models with Sparse Autoencoders", "abs": "Sparse autoencoders (SAEs) have become a core ingredient in the reverse engineering of large-language models (LLMs). For LLMs, they have been shown to decompose intermediate representations that often are not interpretable directly into sparse sums of interpretable features, facilitating better control and subsequent analysis. However, similar analyses and approaches were lacking for text-to-image models. We investigated the possibility of using SAEs to learn interpretable features for a few-step text-to-image diffusion models, such as SDXL Turbo. To this end, we train SAEs on the updates performed by transformer blocks within SDXL Turbo's denoising U-net. We find that their learned features are interpretable, causally influence the generation process, and reveal specialization among the blocks. In particular, we find one block mainly dealing with image composition, mainly responsible for adding local details, and, one for color, illumination, and style. Therefore, our work is an important first step towards better understanding the internals of generative text-to-image models like SDXL Turbo and showcases the potential of features learned by SAEs for the visual domain.", "keywords": [ "SDXL Turbo", "sparse autoencoders", "interpretability", "steering" ], "rating_list": [ 3, 3, 6, 5, 5 ], "soundness_list": [ 3, 2, 3, 2, 3 ], "presentation_list": [ 1, 2, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2, 4 ], "confidence_list": [ 3, 3, 4, 4, 3 ] }, { "paper_id": "CgqnYqpYQh", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Kenton Murray", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Researcher" }, { "name": "Kevin Duh", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "Assistant Research Professor" }, { "name": "Neha Verma", "gender": "Female", "institution": "Johns Hopkins University", "country": "US", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 8, "n_ref_uni": 37, "n_ref": 51, "n_ref_all": 69, "n_fig": 5, "n_tab": 4, "L_tab": 647, "n_element_tab": 74, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 939, "n_element_tab_1": 31, "formula_len_all": 739, "formula_len_all_1": 815, "len_all": 114407, "len_all_1": 56872, "len_abs": 1181, "len_title": 106, "len_sents": 30460, "len_sents_1": 25951, "n_sents": 214, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 58, "L_abs": 1186, "title": "Merging Feed-Forward Sublayers for Compressed Transformers", "abs": "With the rise and ubiquity of larger deep learning models, the need for high-quality compression techniques has been growing in order to deploy these models widely. The sheer parameter count of some models makes it difficult to fit them into the memory constraints of different hardware. In this work, we present a novel approach to model compression by merging similar parameter groups within a model, rather than pruning away less important parameters. Specifically, we propose a straightforward method for selecting, aligning, and merging separate feed-forward sublayers in Transformer models, and test our method on a language modeling task, image classification, and machine translation. With our method, we demonstrate performance comparable to the original models across our three diverse tasks while combining more than a third of model feed-forward sublayers. For instance, we can remove over 21\\% of total parameters from a Vision Transformer, while maintaining 99\\% of its original performance. Additionally, we observe that some feed-forward sublayers often exhibit regions of high similarity between their activations, which may help explain their surprising mergeability.", "keywords": [ "model compression", "model merging", "transformers", "parameter efficiency", "redundancy" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 3, 2, 4 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "CgkGFeSpo0", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Joel Z Leibo", "gender": "Male", "institution": "Google", "country": "", "position": "Research scientist" }, { "name": "Shuqing Shi", "gender": "Male", "institution": "King's College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Yali Du", "gender": "Female", "institution": "King's College London", "country": "GB", "position": "Associate Professor" }, { "name": "Yudi Zhang", "gender": "Female", "institution": "Eindhoven University of Technology", "country": "", "position": "PhD student" } ], "n_formula": 24, "n_formula_1": 22, "n_ref_uni": 34, "n_ref": 62, "n_ref_all": 71, "n_fig": 5, "n_tab": 2, "L_tab": 54, "n_element_tab": 3, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 2, "n_element_tab_1": 1, "formula_len_all": 1805, "formula_len_all_1": 1648, "len_all": 142498, "len_all_1": 74579, "len_abs": 860, "len_title": 133, "len_sents": 46707, "len_sents_1": 33515, "n_sents": 365, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 865, "title": "Resolving Complex Social Dilemmas by Aligning Preferences with Counterfactual Regret", "abs": "Social dilemmas are situations where gains from cooperation are possible but misaligned incentives make it hard to find and stabilize prosocial joint behavior. In such situations selfish behaviors may harm the social good. In spatiotemporally complex social dilemmas, the barriers to cooperation that emerge from misaligned incentives interact with obstacles that stem from spatiotemporal complexity. In this paper, we propose a multi-agent reinforcement learning algorithm which aims to find cooperative resolutions for such complex social dilemmas. Agents maximize their own interests while also helping others, regardless of the actions their co-players take. This approach disentangles the causes of selfish reward from the causes of prosocial reward. Empirically, our method outperforms multiple baseline methods in several complex social dilemma environments.", "keywords": [ "Counterfacutla Regret", "Sequential Social Dilemma" ], "rating_list": [ 3, 5, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "CgXAophhEb", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 4, "author_info_list": [ { "name": "CHENHAN XIAO", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "PhD student" }, { "name": "Haoran Li", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Muhao Guo", "gender": "Male", "institution": "Arizona State University", "country": "US", "position": "PhD student" }, { "name": "Yang Weng", "gender": "Male", "institution": "Arizona State University", "country": "", "position": "Assistant Professor" } ], "n_formula": 55, "n_formula_1": 31, "n_ref_uni": 54, "n_ref": 104, "n_ref_all": 131, "n_fig": 6, "n_tab": 4, "L_tab": 7541, "n_element_tab": 77, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 7224, "n_element_tab_1": 29, "formula_len_all": 20094, "formula_len_all_1": 17976, "len_all": 198983, "len_all_1": 102928, "len_abs": 1232, "len_title": 132, "len_sents": 50061, "len_sents_1": 33762, "n_sents": 433, "n_sents_1": 265, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1256, "title": "G-AlignNet: Geometry-Driven Quality Alignment for Robust Dynamical Systems Modeling", "abs": "The Neural ODE family has shown promise in modeling complex systems but often assumes consistent data quality, making them less effective in real-world applications with irregularly sampled, incomplete, or multi-resolution data. Current methods, such as ODE-RNN, aim to address these issues but lack formal performance guarantees and can struggle with highly evolving dynamical systems. To tackle this, we propose a novel approach that leverages parameter manifolds to improve robustness in system dynamical modeling. Our method utilizes the orthogonal group as the underlying structure for the parameter manifold, facilitating both quality alignment and dynamical learning in a unified framework. Unlike previous methods, which primarily focus on empirical performance, our approach offers stronger theoretical guarantees of error convergence thanks to the novel architecture and well-posed optimization with orthogonality. Numerical experiments demonstrate significant improvements in interpolation and prediction tasks, particularly in scenarios involving high- and low-resolution data, irregular sampling intervals, etc. Our framework provides a step toward more reliable dynamics learning in changing environments where data quality cannot be assumed.", "keywords": [ "adaptive physical systems", "dynamical modeling", "data quality alignment", "geometric representation learning", "geometric optimization" ], "rating_list": [ 6, 6, 5, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 4, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 1 ] }, { "paper_id": "CgRkPuhTGm", "primary_area": "learning on time series and dynamical systems", "n_author": 5, "author_info_list": [ { "name": "Erli Meng", "gender": "Male", "institution": "Xiaomi Corporation", "country": "", "position": "Intern" }, { "name": "Pengfei Tian", "gender": "unknown", "institution": "Xiaomi Corporation", "country": "CN", "position": "Researcher" }, { "name": "Teng Wu", "gender": "unknown", "institution": "Xiaomi Corporation", "country": "CN", "position": "Researcher" }, { "name": "Tianzhe Dong", "gender": "Male", "institution": "Jilin University", "country": "CN", "position": "MS student" }, { "name": "Yuan Chen", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "MS student" } ], "n_formula": 31, "n_formula_1": 15, "n_ref_uni": 29, "n_ref": 46, "n_ref_all": 55, "n_fig": 3, "n_tab": 8, "L_tab": 6869, "n_element_tab": 261, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 8869, "n_element_tab_1": 155, "formula_len_all": 5333, "formula_len_all_1": 1500, "len_all": 119482, "len_all_1": 54949, "len_abs": 2900, "len_title": 114, "len_sents": 26360, "len_sents_1": 18597, "n_sents": 199, "n_sents_1": 123, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1358, "title": "SSNet: Skip and Split MLP Network for Long-Term Series Forecasting", "abs": "Time series forecasting is critical across various domains, \n including energy, transportation, weather prediction, and healthcare. \n Although recent advances using CNNs, RNNs, and Transformer-based models have shown promise, \n these approaches often suffer from architectural complexity and low computational efficiency. \n MLP-based networks offer better computational efficiency, \n and some frequency-domain MLP models have demonstrated the ability to handle periodic time series data. \n However, standard MLP-based methods still struggle to directly model periodic and temporal dependencies in the time domain, \n which are essential for accurate time series forecasting. \n To address these challenges, we propose the Skip and Split MLP Network (SSNet), \n featuring innovative Skip-MLP and Split-MLP components that enable MLP models to directly capture periodicity and temporal dependencies in the time domain. \n SSNet requires fewer parameters than traditional MLP-based architectures, \n improving computational efficiency. \n Empirical results on multiple real-world long-term forecasting datasets demonstrate that SSNet significantly outperforms state-of-the-art models, \n delivering better performance with fewer parameters. Notably, even a single Skip-MLP unit matches the performance of high-performing models like PatchTST.", "keywords": [ "Time Series Forecasting", "Deep Learning", "MLP" ], "rating_list": [ 10, 1, 1, 8 ], "soundness_list": [ 4, 2, 2, 4 ], "presentation_list": [ 4, 2, 3, 3 ], "contribution_list": [ 4, 2, 2, 4 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "CfdPELywGN", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Doyoung Kim", "gender": "Male", "institution": "KAIST", "country": "KR", "position": "MS student" }, { "name": "Jinho Park", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Jongwon Lee", "gender": "Male", "institution": "Samsung", "country": "KR", "position": "Researcher" }, { "name": "Minjoon Seo", "gender": "Male", "institution": "Twelve Labs", "country": "", "position": "Chief Scientist" } ], "n_formula": 5, "n_formula_1": 0, "n_ref_uni": 52, "n_ref": 102, "n_ref_all": 123, "n_fig": 9, "n_tab": 9, "L_tab": 1997, "n_element_tab": 190, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 237, "n_element_tab_1": 37, "formula_len_all": 482, "formula_len_all_1": 0, "len_all": 197591, "len_all_1": 61766, "len_abs": 1146, "len_title": 144, "len_sents": 56263, "len_sents_1": 30821, "n_sents": 379, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 96, "L_abs": 1150, "title": "How language models extrapolate outside the training data: A Case study in Textualized Gridworld", "abs": "Language models' ability to extrapolate learned behaviors to novel, more complex environments beyond their training scope is highly unknown. This study introduces a path planning task in a textualized Gridworld to probe language models' extrapolation capabilities. We show that conventional approaches, including next-token prediction and Chain of Thought (CoT) fine-tuning, fail to extrapolate in larger, unseen environments. Inspired by human cognition and dual-process theory, we propose cognitive maps for path planning\u2014a novel CoT framework that simulates human-like mental representations. Our experiments show that cognitive maps not only enhance extrapolation to unseen environments but also exhibit human-like characteristics through structured mental simulation and rapid adaptation. Our finding that these cognitive maps require specialized training schemes and cannot be induced through simple prompting opens up important questions about developing general-purpose cognitive maps in language models. Our comparison with exploration-based methods further illuminates the complementary strengths of offline planning and online exploration.", "keywords": [ "Cognitive map", "NeuroAI", "language model", "language agent", "planning" ], "rating_list": [ 8, 5, 5, 3, 5 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 3, 3, 2 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 2, 4, 4 ] }, { "paper_id": "Cfbr56K4gp", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Hongyi Guo", "gender": "Male", "institution": "Northwestern University, Northwestern University", "country": "", "position": "PhD student" }, { "name": "Rui Zheng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "wei shen", "gender": "unknown", "institution": "Baichuan", "country": "", "position": "Researcher" }, { "name": "Xiaoying Zhang", "gender": "Female", "institution": "ByteDance AILab", "country": "", "position": "Researcher" }, { "name": "Yang Liu", "gender": "Male", "institution": "University of California, Santa Cruz", "country": "US", "position": "Assistant Professor" }, { "name": "Kevin Yao", "gender": "Male", "institution": "Meta GenAI", "country": "US", "position": "Researcher" } ], "n_formula": 21, "n_formula_1": 15, "n_ref_uni": 54, "n_ref": 99, "n_ref_all": 123, "n_fig": 7, "n_tab": 5, "L_tab": 2209, "n_element_tab": 257, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 1636, "n_element_tab_1": 236, "formula_len_all": 3283, "formula_len_all_1": 991, "len_all": 268561, "len_all_1": 70143, "len_abs": 1995, "len_title": 78, "len_sents": 54150, "len_sents_1": 29307, "n_sents": 449, "n_sents_1": 228, "venue": "ICLR2025", "decision": "withdraw", "L_title": 30, "L_abs": 1360, "title": "Robust RLHF with Noisy Rewards", "abs": "Reinforcement learning from human feedback (RLHF) is the mainstream paradigm to align large language models (LLMs) with human preferences. Yet existing RLHF heavily relies on accurate and informative reward models, which are vulnerable and sensitive to noise from various sources, e.g. human labeling errors, making the pipeline fragile. In this work, we formulate the problem of performing robust RLHF with noisy reward models. Our goal is to design robust RLHF algorithms that explicitly acknowledge the potential noise in a reward model. Our first contribution is an analysis that revealed a certain transformation of the preference function improves its robustness to noise in the reward function. This observation leads to a new reward function design that involves two steps: (1) an offline sampling step to obtain responses to prompts that serve as baseline calculation and (2) a contrastive reward calculated using the baseline responses in Proximal Policy Optimization (PPO). We show that our suggested rewards enable the LLM to penalize reward uncertainty, improve robustness, encourage improvement over baselines, calibrate according to task difficulty, and reduce variance in PPO. We also empirically demonstrate contrastive reward can improve RLHF substantially, evaluated by both GPTs and humans, and it consistently outperforms strong baselines.", "keywords": [ "Alignment", "LLM", "RLHF" ], "rating_list": [ 3, 6, 1, 8 ], "soundness_list": [ 2, 4, 3, 3 ], "presentation_list": [ 2, 4, 3, 3 ], "contribution_list": [ 2, 2, 1, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "CfZPzH7ftt", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Jaemoo Choi", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Jaewoong Choi", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "Assistant Professor" }, { "name": "Yongxin Chen", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" } ], "n_formula": 48, "n_formula_1": 30, "n_ref_uni": 34, "n_ref": 112, "n_ref_all": 129, "n_fig": 3, "n_tab": 6, "L_tab": 1514, "n_element_tab": 183, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 596, "n_element_tab_1": 91, "formula_len_all": 4074, "formula_len_all_1": 2082, "len_all": 161308, "len_all_1": 68483, "len_abs": 1252, "len_title": 114, "len_sents": 45713, "len_sents_1": 26630, "n_sents": 434, "n_sents_1": 232, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1261, "title": "Improving Neural Optimal Transport via Displacement Interpolation", "abs": "Optimal Transport (OT) theory investigates the cost-minimizing transport map that moves a source distribution to a target distribution. Recently, several approaches have emerged for learning the optimal transport map for a given cost function using neural networks. We refer to these approaches as the OT Map. OT Map provides a powerful tool for diverse machine learning tasks, such as generative modeling and unpaired image-to-image translation. However, existing methods that utilize max-min optimization often experience training instability and sensitivity to hyperparameters. In this paper, we propose a novel method to improve stability and achieve a better approximation of the OT Map by exploiting displacement interpolation, dubbed Displacement Interpolation Optimal Transport Model (DIOTM). We derive the dual formulation of displacement interpolation at specific time $t$ and prove how these dual problems are related across time. This result allows us to utilize the entire trajectory of displacement interpolation in learning the OT Map. Our method improves the training stability and achieves superior results in estimating optimal transport maps. We demonstrate that DIOTM outperforms existing OT-based models on image-to-image translation tasks.", "keywords": [ "Optimal Transport", "Displacement Interpolation", "Image-to-image Translation" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "CfXRcN4iUw", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 6, "author_info_list": [ { "name": "Feng Zhou", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Assistant Professor" }, { "name": "Jingwen Xu", "gender": "Female", "institution": "Wuhan University of Technology", "country": "CN", "position": "PhD student" }, { "name": "Junchao Lin", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "PhD student" }, { "name": "Zenan Ling", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Researcher" }, { "name": "Zhanbo Feng", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" }, { "name": "Caiming Qiu", "gender": "Male", "institution": "Huazhong University of Science and Technology", "country": "CN", "position": "Full Professor" } ], "n_formula": 27, "n_formula_1": 17, "n_ref_uni": 46, "n_ref": 90, "n_ref_all": 107, "n_fig": 7, "n_tab": 6, "L_tab": 4494, "n_element_tab": 648, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 4272, "n_element_tab_1": 133, "formula_len_all": 1988, "formula_len_all_1": 1220, "len_all": 166200, "len_all_1": 64339, "len_abs": 2841, "len_title": 117, "len_sents": 45391, "len_sents_1": 26447, "n_sents": 333, "n_sents_1": 189, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1032, "title": "IGNN-Solver: A Graph Neural Solver for Implicit Graph Neural Networks", "abs": "Implicit graph neural networks (IGNNs), which exhibit strong expressive power with a single layer, have recently demonstrated remarkable performance in capturing long-range dependencies (LRD) in underlying graphs while effectively mitigating the over-smoothing problem. However, IGNNs rely on computationally expensive fixed-point iterations, which lead to significant speed and scalability limitations, hindering their application to large-scale graphs. To achieve fast fixed-point solving for IGNNs, we propose a novel graph neural solver, IGNN-Solver, which leverages the generalized Anderson Acceleration method, parameterized by a small GNN, and learns iterative updates as a graph-dependent temporal process. Extensive experiments demonstrate that the IGNN-Solver significantly accelerates inference, achieving a $1.5\\times$ to $8\\times$ speedup without sacrificing accuracy. Moreover, this advantage becomes increasingly pronounced as the graph scale grows, facilitating its large-scale deployment in real-world applications.", "keywords": [ "implicit deep learning", "deep equilibrium models", "implicit graph neural networks" ], "rating_list": [ 5, 3, 5, 3, 6 ], "soundness_list": [ 2, 2, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 3, 4, 4 ] }, { "paper_id": "Cf0K6jgzZt", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 8, "author_info_list": [ { "name": "Colin Kohler", "gender": "Male", "institution": "Northeastern University", "country": "", "position": "PhD student" }, { "name": "Justin Goodwin", "gender": "unknown", "institution": "MIT Lincoln Laboratory, Massachusetts Institute of Technology", "country": "", "position": "MIT Lincoln Laboratory" }, { "name": "Matthew C. Jones", "gender": "Male", "institution": "MIT Lincoln Laboratory, Massachusetts Institute of Technology", "country": "US", "position": "Principal Researcher" }, { "name": "Nathan Vaska", "gender": "Male", "institution": "MIT Lincoln Laboratory, Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Purvik Patel", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Researcher" }, { "name": "Rajmonda S. Caceres", "gender": "unknown", "institution": "MIT Lincoln Laboratory, Massachusetts Institute of Technology", "country": "", "position": "Principal Researcher" }, { "name": "Robert Platt", "gender": "unknown", "institution": "Northeastern University", "country": "", "position": "Associate Professor" }, { "name": "Robin Walters", "gender": "Male", "institution": "Northeastern University ", "country": "US", "position": "Assistant Professor" } ], "n_formula": 13, "n_formula_1": 3, "n_ref_uni": 45, "n_ref": 110, "n_ref_all": 143, "n_fig": 18, "n_tab": 3, "L_tab": 415, "n_element_tab": 7, "n_fig_1": 13, "n_tab_1": 2, "L_tab_1": 334, "n_element_tab_1": 31, "formula_len_all": 534, "formula_len_all_1": 124, "len_all": 171375, "len_all_1": 68088, "len_abs": 1100, "len_title": 120, "len_sents": 60119, "len_sents_1": 32934, "n_sents": 465, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 1093, "title": "G2Sphere: Learning High-Frequnecy Spherical Signals From Geometric Data", "abs": "Many modeling tasks from disparate domains can be framed the same way, computing spherical signals from a geometric input, for example, computing the radar response or aerodynamics drag of different objects, or navigating through an environment. This paper introduces G2Sphere, a general method for mapping object geometries to spherical signals. G2Sphere operates entirely in Fourier space, encoding geometric structure into latent Fourier features using equivariant neural networks and then outputting the Fourier coefficients of the output signal. Combining these coefficients with spherical harmonics enables the simultaneous prediction of all values of the continuous spherical signal at any resolution. We perform experiments on various challenging domains including radar response modeling, aerodynamics drag prediction, and policy learning for manipulation and navigation. We find that G2Sphere significantly outperforms baselines in terms of accuracy and inference time. We also demonstrate that equivariance and Fourier features lead to improved sample efficiency and generalization.", "keywords": [ "Equivariance", "Geometric", "Fourier", "Spherical Signals", "SO(3)", "Radar" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "CexatBp6rx", "primary_area": "interpretability and explainable AI", "n_author": 5, "author_info_list": [ { "name": "Alasdair Newson", "gender": "Unspecified", "institution": "T\u00e9l\u00e9com ParisTech", "country": "", "position": "Assistant Professor" }, { "name": "Jayneel Parekh", "gender": "unknown", "institution": "Sorbonne Universit\u00e9", "country": "FR", "position": "Postdoc" }, { "name": "Pavlo Mozharovskyi", "gender": "Male", "institution": "T\u00e9l\u00e9com Paris", "country": "FR", "position": "Full Professor" }, { "name": "Quentin Bouniot", "gender": "Male", "institution": "Helmholtz Zentrum M\u00fcnchen", "country": "DE", "position": "Postdoc" }, { "name": "Florence d'Alch\u00e9", "gender": "Female", "institution": "T\u00e9l\u00e9com Paris", "country": "", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 6, "n_ref_uni": 81, "n_ref": 159, "n_ref_all": 209, "n_fig": 13, "n_tab": 15, "L_tab": 2694, "n_element_tab": 355, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 959, "n_element_tab_1": 100, "formula_len_all": 740, "formula_len_all_1": 573, "len_all": 259116, "len_all_1": 71337, "len_abs": 1240, "len_title": 131, "len_sents": 83799, "len_sents_1": 35142, "n_sents": 595, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 82, "L_abs": 1245, "title": "Restyling Unsupervised Concept Based Interpretable Networks with Generative Models", "abs": "Developing inherently interpretable models for prediction has gained prominence in recent years. A subclass of these models, wherein the interpretable network relies on learning high-level concepts, are valued because of closeness of concept representations to human communication. However, the visualization and understanding of the learnt unsupervised dictionary of concepts encounters major limitations, especially for large-scale images. We propose here a novel method that relies on mapping the concept features to the latent space of a pretrained generative model. The use of a generative model enables high quality visualization, and lays out an intuitive and interactive procedure for better interpretation of the learnt concepts by imputing concept activations and visualizing generated modifications. Furthermore, leveraging pretrained generative models has the additional advantage of making the training of the system more efficient. We quantitatively ascertain the efficacy of our method in terms of accuracy of the interpretable prediction network, fidelity of reconstruction, as well as faithfulness and consistency of learnt concepts. The experiments are conducted on multiple image recognition benchmarks for large-scale images.", "keywords": [ "explainability", "generative models", "concepts" ], "rating_list": [ 6, 6, 8, 6, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 3, 4 ] }, { "paper_id": "CeIOWuD8oZ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Jue Wang", "gender": "Female", "institution": "University of California, Davis", "country": "US", "position": "PhD student" }, { "name": "Raymond K. W. Wong", "gender": "unknown", "institution": "Texas A&M University", "country": "US", "position": "Associate Professor" }, { "name": "Thomas Chun Man Lee", "gender": "unknown", "institution": "University of California, Davis", "country": "US", "position": "Full Professor" } ], "n_formula": 94, "n_formula_1": 48, "n_ref_uni": 23, "n_ref": 73, "n_ref_all": 101, "n_fig": 3, "n_tab": 6, "L_tab": 3116, "n_element_tab": 233, "n_fig_1": 0, "n_tab_1": 2, "L_tab_1": 81, "n_element_tab_1": 6, "formula_len_all": 8967, "formula_len_all_1": 3889, "len_all": 155674, "len_all_1": 64281, "len_abs": 987, "len_title": 45, "len_sents": 50928, "len_sents_1": 26944, "n_sents": 454, "n_sents_1": 221, "venue": "ICLR2025", "decision": "withdraw", "L_title": 45, "L_abs": 991, "title": "Weighted Fair Regression under Selection Bias", "abs": "Selection bias is a prevalent challenge in real-world data analysis, often stemming from biased historical censoring policies. While there is a growing body of literature on fairness in mitigating accuracy disparities, few studies have considered the potential impact of selection bias in training data. Depending on the selection mechanism, significant differences can arise between the population distribution and the training data distribution. Therefore, the training fairness metric can be heavily biased, leading to unfair learning. To address this issue under the fair regression problem, we propose weighting adjustments in the fairness constraint, which results in a novel fair regression estimator. Despite non-convexity, we derive an efficient algorithm to obtain a globally optimal solution. This work pioneers the integration of weighting adjustments into the fair regression problem, introducing a novel methodology to constrain accuracy disparities under arbitrary thresholds.", "keywords": [ "Fair Regression", "Selection Bias" ], "rating_list": [ 5, 3, 5, 6 ], "soundness_list": [ 4, 3, 3, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 1, 2, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "CdqQKXGKq3", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 3, "author_info_list": [ { "name": "Haiwei Wu", "gender": "Male", "institution": "City University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Jiantao Zhou", "gender": "Male", "institution": "University of Macau", "country": "MO", "position": "Full Professor" }, { "name": "Yiming Chen", "gender": "Not Specified", "institution": "University of Macau", "country": "", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 18, "n_ref": 36, "n_ref_all": 54, "n_fig": 8, "n_tab": 7, "L_tab": 913, "n_element_tab": 78, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 560, "n_element_tab_1": 74, "formula_len_all": 322, "formula_len_all_1": 322, "len_all": 151225, "len_all_1": 62886, "len_abs": 1659, "len_title": 133, "len_sents": 32215, "len_sents_1": 29925, "n_sents": 236, "n_sents_1": 213, "venue": "ICLR2025", "decision": "Reject", "L_title": 85, "L_abs": 1659, "title": "CFBD: COARSE-TO-FINE DETECTION OF BACKDOOR ATTACKS IN MULTIMODAL CONTRASTIVE LEARNING", "abs": "The backdoor attack in Multimodal Contrastive Learning (MCL) task has been receiving increasing attention in recent years, due to numerous downstream tasks that rely on pre-trained MCL models. Backdoor detection has been one of the effective protection solutions to fight against backdoor attacks. However, the majority of existing backdoor detection methods in MCL usually produces nonsatisfying detection results. Two main factors are responsible for this: 1) one-stage detection lacks subsequent dynamic adaptation to the distribution of poisoned and benign pairs when faced with different attacks, and 2) the criteria used in existing methods, specifically the cosine similarity between image and caption, are insufficient to distinguish between poisoned and benign pairs. To address these problems, we extend the conventional one-stage detection architecture to a two-stage architecture and propose a better metric in the second stage with high precision and high fault tolerance. To this end, we design a novel Coarse-to-Fine two-stage Backdoor Detection method, termed CFBD, which primarily focuses on multimodal learning involving image-caption relationships, such as CLIP. The objective of the coarse stage is to roughly partition dataset into poisoned, benign and suspicious subset. In the fine-grained stage, we use the average textual correlation with the poisoned subset to improve the detection quality. Extensive experiments demonstrate that CFBD achieves superior backdoor detection performance, e.g., almost 100% True Positive Rate (TPR) for diverse attacks over the large scale dataset CC-3M, markedly outperforming state-of-the-art methods.", "keywords": [ "backdoor" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 4 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 4, 2, 5 ] }, { "paper_id": "Cdhxv0Oz1v", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Qizhou Wang", "gender": "Male", "institution": "Hong Kong Baptist University", "country": "HK", "position": "PhD student" }, { "name": "Xinyu Liu", "gender": "Male", "institution": "The Chinese University of Hong Kong, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Yixuan Yuan", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" }, { "name": "YufanHU", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 14, "n_ref_uni": 46, "n_ref": 91, "n_ref_all": 103, "n_fig": 4, "n_tab": 6, "L_tab": 2733, "n_element_tab": 307, "n_fig_1": 4, "n_tab_1": 9, "L_tab_1": 2571, "n_element_tab_1": 294, "formula_len_all": 675, "formula_len_all_1": 689, "len_all": 125381, "len_all_1": 71348, "len_abs": 1266, "len_title": 117, "len_sents": 34443, "len_sents_1": 30646, "n_sents": 245, "n_sents_1": 214, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1273, "title": "Rectifying Gradient-based OOD Detection via Fisher Information Matrix", "abs": "Out-of-distribution (OOD) detection is an anomaly-handling mechanism, for which classification systems should detect outliers with true labels outside the label space, distinguishing them from normal in-distribution (ID) data. \nAdvanced works suggest that gradient information preserve sufficient cues to indicate the confidence of being OOD.\nHowever, we discover previous gradient-based detection methods suffer from limited effectiveness mainly due to over-parameterization. As gradient-based OOD scores derive from the over-parameterized weight space, a widely recognized cause for the suboptimal OOD detection performance, there are also some gradient components which lack necessary informativeness for OOD detection, thereby impair the performance.\nThis observation motivates us to propose gradient rectification (GradRect), using fisher information matrix to correct gradients in directions that are uninformative to discern the distribution change. Moreover, we connect GradRect with classical theories in identifying influential observations, verifying that model fine-tuning with outlier exposure can further improve GradRect. We conduct extensive experiments on various OOD detection setups, revealing the power of GradRect against state-of-the-art counterparts.", "keywords": [ "OOD Detection" ], "rating_list": [ 6, 3, 3, 3, 6 ], "soundness_list": [ 3, 1, 2, 2, 3 ], "presentation_list": [ 3, 1, 2, 2, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 2, 5, 2, 4, 4 ] }, { "paper_id": "Cd25C59teq", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 3, "author_info_list": [ { "name": "Augustin Luna", "gender": "Not Specified", "institution": "National Institutes of Health", "country": "US", "position": "Principal Researcher" }, { "name": "Tianfan Fu", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "Assistant Professor" }, { "name": "Yoshitaka Inoue", "gender": "Male", "institution": "National Library of Medicine", "country": "US", "position": "PhD student" } ], "n_formula": 22, "n_formula_1": 15, "n_ref_uni": 46, "n_ref": 67, "n_ref_all": 79, "n_fig": 3, "n_tab": 3, "L_tab": 1876, "n_element_tab": 120, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2033, "n_element_tab_1": 169, "formula_len_all": 903, "formula_len_all_1": 532, "len_all": 166430, "len_all_1": 63512, "len_abs": 2217, "len_title": 146, "len_sents": 50548, "len_sents_1": 27769, "n_sents": 383, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Reject", "L_title": 97, "L_abs": 2232, "title": "GraphPINE: Graph importance propagation Neural Network for interpretable drug response prediction", "abs": "Explainability is necessary for tasks that require a clear reason for a given result such as finance or biomedical research. Recent explainability methodologies have focused on attention, gradient, and Shapley value methods. These do not handle data with strong associated prior knowledge and fail to constrain explainability results by relationships that may exist between predictive features.\n\nWe propose a GraphPINE, a novel graph neural network (GNN) architecture that leverages domain-specific prior knowledge for node importance score initialization. Use cases in biomedicine necessitate generating hypotheses related to specific nodes. Commonly, there is a manual post-prediction step examining literature (i.e., prior knowledge) to better understand features. While node importance can be obtained for gradient and attention-based methods after prediction, these node importances lack complementary prior knowledge; GraphPINE seeks to overcome this limitation. GraphPINE differs from other GNNs with gating methods that utilize an LSTM-like sequential format such that we introduce an importance propagation layer that unifies 1) updates for feature matrix and node importances, jointly and 2) uses GNN-based graph propagation of feature values. This initialization and updating mechanism allows for more informed feature learning and improved graph representation.\n\nWe apply GraphPINE to cancer drug response prediction using pharmacogenomics data (i.e., both drug screening and gene data collected by several assays) for ~5K gene nodes included in a gene-gene input graph with drug-target interaction (DTI) knowledge graph as initial importance. The gene-gene graph and DTIs were taken from literature curated prior knowledge sources and weighted by the literature information. GraphPINE demonstrates competitive performance and achieves a PR-AUC of 0.894 and ROC-AUC of 0.796 across 952 drugs. To highlight the interpretability aspect of our work, we provide the ability to generate sub-graphs of node importances. While our use case is related to biology, our work is generally applicable to tasks where information is separately known about feature relationships. Code: https://anonymous.4open.science/r/GraphPINE-40DE", "keywords": [ "Graph Neural Networks", "Information Propagation", "Drug Response Prediction", "Explainability" ], "rating_list": [ 5, 6, 5 ], "soundness_list": [ 3, 3, 2 ], "presentation_list": [ 3, 4, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "Ccwp4tFEtE", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Arian Hosseini", "gender": "Male", "institution": "University of Montreal", "country": "", "position": "PhD student" }, { "name": "Aviral Kumar", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Assistant Professor" }, { "name": "Hritik Bansal", "gender": "Male", "institution": "University of California, Los Angeles", "country": "", "position": "PhD student" }, { "name": "Lunjun Zhang", "gender": "unknown", "institution": "Google DeepMind", "country": "CA", "position": "Student Researcher" }, { "name": "Rishabh Agarwal", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "Adjunct Professor" }, { "name": "Seyed Mehran Kazemi", "gender": "Male", "institution": "Google", "country": "", "position": "Researcher" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 49, "n_ref": 112, "n_ref_all": 142, "n_fig": 12, "n_tab": 8, "L_tab": 4296, "n_element_tab": 152, "n_fig_1": 11, "n_tab_1": 5, "L_tab_1": 1853, "n_element_tab_1": 88, "formula_len_all": 781, "formula_len_all_1": 797, "len_all": 189433, "len_all_1": 67294, "len_abs": 1390, "len_title": 110, "len_sents": 53576, "len_sents_1": 30667, "n_sents": 481, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1394, "title": "Generative Verifiers: Reward Modeling as Next-Token Prediction", "abs": "Verifiers or reward models are often used to enhance the reasoning performance of large language models (LLMs). A common approach is the Best-of-N method, where N candidate solutions generated by the LLM are ranked by a verifier, and the best one is selected. While LLM-based verifiers are typically trained as discriminative classifiers to score solutions, they do not utilize the text generation capabilities of pretrained LLMs. To overcome this limitation, we instead propose training verifiers using the ubiquitous next-token prediction objective, jointly on verification and solution generation. Compared to standard verifiers, such generative verifiers (GenRM) can benefit from several advantages of LLMs: they integrate seamlessly with instruction tuning, enable chain-of-thought reasoning, and can utilize additional test-time compute via majority voting for better verification. We demonstrate that GenRM outperforms discriminative, DPO verifiers, and LLM-as-a-Judge, resulting in large performance gains with Best-of-N, namely 5% \u2192 45.3% on algorithmic tasks, 73% \u2192 93.4% on GSM8K, and 28% \u219244.6% on easy-to-hard generalization on MATH. Furthermore, we find that training GenRM with synthetic verification rationales is sufficient to pick out subtle errors on math problems. Finally, we demonstrate that generative verifiers scale favorably with model size and inference-time compute.", "keywords": [ "LLM reasoning", "reward models", "verifiers" ], "rating_list": [ 3, 8, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 1, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "CbpWPbYHuv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Jinwen Ma", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoqing Li", "gender": "unknown", "institution": "Capital University of Economics and Business", "country": "CN", "position": "Lecturer" }, { "name": "Xun Zhou", "gender": "unknown", "institution": "bytedance Inc", "country": "", "position": "Researcher" }, { "name": "Ya Wang", "gender": "Male", "institution": "Tencent AI Lab", "country": "", "position": "Researcher" }, { "name": "Yutao Zeng", "gender": "Male", "institution": "ByteDance Inc.", "country": "CN", "position": "Researcher" }, { "name": "Zhijian Zhuo", "gender": "Male", "institution": "Peking University", "country": "CN", "position": "PhD student" } ], "n_formula": 66, "n_formula_1": 24, "n_ref_uni": 42, "n_ref": 69, "n_ref_all": 93, "n_fig": 8, "n_tab": 12, "L_tab": 1923, "n_element_tab": 156, "n_fig_1": 8, "n_tab_1": 7, "L_tab_1": 1244, "n_element_tab_1": 39, "formula_len_all": 4860, "formula_len_all_1": 1451, "len_all": 161162, "len_all_1": 57143, "len_abs": 1798, "len_title": 133, "len_sents": 41521, "len_sents_1": 24940, "n_sents": 388, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 84, "L_abs": 1454, "title": "Polynomial Composition Activations: Unleashing the Dynamics of Large Language Models", "abs": "Transformers have found extensive applications across various domains due to the powerful fitting capabilities. This success can be partially attributed to their inherent nonlinearity. Thus, in addition to the ReLU function employed in the original transformer architecture, researchers have explored alternative modules such as GeLU and SwishGLU to enhance nonlinearity and thereby augment representational capacity. In this paper, we propose a novel category of polynomial composition activations (PolyCom), designed to optimize the dynamics of transformers. Theoretically, we provide a comprehensive mathematical analysis of PolyCom, highlighting its enhanced expressivity and efficacy relative to other activation functions. Notably, we demonstrate that networks incorporating PolyCom achieve the **optimal approximation rate**, indicating that PolyCom networks require minimal parameters to approximate general smooth functions in Sobolev spaces. We conduct empirical experiments on the pre-training configurations of large language models (LLMs), including both dense and sparse architectures. By substituting conventional activation functions with PolyCom, we enable LLMs to capture higher-order interactions within the data, thus improving performance metrics in terms of accuracy and convergence rates. Extensive experimental results demonstrate the effectiveness of our method, showing substantial improvements over other activation functions.", "keywords": [ "activation function", "transformer", "pre-training", "large language models" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "CblmtAxrRg", "primary_area": "interpretability and explainable AI", "n_author": 8, "author_info_list": [ { "name": "Dong Yu", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Distinguished Scientist" }, { "name": "Fei Liu", "gender": "Female", "institution": "Emory University", "country": "US", "position": "Associate Professor" }, { "name": "Hassan Foroosh", "gender": "unknown", "institution": "University of Central Florida", "country": "", "position": "Full Professor" }, { "name": "Wenlin Yao", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Researcher" }, { "name": "Yebowen Hu", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Yiming Lu", "gender": "Male", "institution": "Emory University", "country": "US", "position": "PhD student" }, { "name": "Zhang Daoan", "gender": "Male", "institution": "University of Rochester", "country": "US", "position": "PhD student" }, { "name": "Xiaoyang Wang", "gender": "Male", "institution": "Tencent AI Lab", "country": "US", "position": "Senior Researcher" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 52, "n_ref": 85, "n_ref_all": 100, "n_fig": 4, "n_tab": 11, "L_tab": 5950, "n_element_tab": 222, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2060, "n_element_tab_1": 78, "formula_len_all": 327, "formula_len_all_1": 327, "len_all": 164334, "len_all_1": 73559, "len_abs": 1190, "len_title": 131, "len_sents": 43132, "len_sents_1": 34754, "n_sents": 292, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1196, "title": "DeFine: Enhancing LLM Decision-Making with Factor Profiles and Analogical Reasoning", "abs": "LLMs are ideal for decision-making due to their ability to reason over long contexts and identify critical factors. However, challenges arise when processing transcripts of spoken speech describing complex scenarios. These transcripts often contain ungrammatical or incomplete sentences, repetitions, hedging, and vagueness. For example, during a company's earnings call, an executive might project a positive revenue outlook to reassure investors, despite significant uncertainty regarding future earnings. It is crucial for LLMs to incorporate this uncertainty systematically when making decisions. In this paper, we introduce DeFine, a new framework that constructs probabilistic factor profiles from complex scenarios. DeFine then integrates these profiles with analogical reasoning, leveraging insights from similar past experiences to guide LLMs in making critical decisions in novel situations. Our framework separates the tasks of quantifying uncertainty in complex scenarios and incorporating it into LLM decision-making. This approach is particularly useful in fields such as medical consultations, negotiations, and political debates, where making decisions under uncertainty is vital.", "keywords": [ "analogical reasoning", "factor profiles", "LLM decision-making", "spoken transcripts", "long-context LLMs" ], "rating_list": [ 6, 6, 6, 3, 5 ], "soundness_list": [ 3, 3, 3, 2, 2 ], "presentation_list": [ 4, 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 3, 3, 3 ] }, { "paper_id": "CbfsKHiWEn", "primary_area": "foundation or frontier models, including LLMs", "n_author": 9, "author_info_list": [ { "name": "Bolin Ding", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Senior Director" }, { "name": "Jiancan Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Postdoc" }, { "name": "Jiawei Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Jinyang Gao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Junkang Wu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Xiang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Xiangnan He", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Professor" }, { "name": "Yuexiang Xie", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Staff" }, { "name": "Zhengyi Yang", "gender": "unknown", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" } ], "n_formula": 83, "n_formula_1": 13, "n_ref_uni": 40, "n_ref": 77, "n_ref_all": 105, "n_fig": 10, "n_tab": 10, "L_tab": 3432, "n_element_tab": 206, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1351, "n_element_tab_1": 101, "formula_len_all": 5962, "formula_len_all_1": 891, "len_all": 212654, "len_all_1": 69726, "len_abs": 1331, "len_title": 153, "len_sents": 61956, "len_sents_1": 30670, "n_sents": 543, "n_sents_1": 252, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 1277, "title": "Towards Robust Alignment of Language Models: Distributionally Robustifying Direct Preference Optimization", "abs": "This study addresses the challenge of noise in training datasets for Direct Preference Optimization (DPO), a method for aligning Large Language Models (LLMs) with human preferences. We categorize noise into pointwise noise, which includes low-quality data points, and pairwise noise, which encompasses erroneous data pair associations that affect preference rankings. Utilizing Distributionally Robust Optimization (DRO), we enhance DPO's resilience to these types of noise. Our theoretical insights reveal that DPO inherently embeds DRO principles, conferring robustness to pointwise noise, with the regularization coefficient $\\beta$ playing a critical role in its noise resistance. Extending this framework, we introduce Distributionally Robustifying DPO (Dr. DPO), which integrates pairwise robustness by optimizing against worst-case pairwise scenarios. The novel hyperparameter $\\beta'$ in Dr. DPO allows for fine-tuned control over data pair reliability, providing a strategic balance between exploration and exploitation in noisy training environments. Empirical evaluations demonstrate that Dr. DPO substantially improves the quality of generated text and response accuracy in preference datasets, showcasing enhanced performance in both noisy and noise-free settings.", "keywords": [ "Direct Preference Optimization", "LLM's alignment", "Distributionally Robust Optimization" ], "rating_list": [ 6, 6, 5, 6, 8 ], "soundness_list": [ 3, 4, 3, 3, 4 ], "presentation_list": [ 3, 2, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3, 4 ], "confidence_list": [ 4, 3, 4, 4, 3 ] }, { "paper_id": "CbepKhSNc0", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Furong Huang", "gender": "Female", "institution": "University of Maryland", "country": "US", "position": "Associate Professor" }, { "name": "James Beetham", "gender": "unknown", "institution": "University of Central Florida", "country": "US", "position": "PhD student" }, { "name": "Mengdi Wang", "gender": "Female", "institution": "Princeton University", "country": "US", "position": "Associate Professor" }, { "name": "Mubarak Shah", "gender": "Male", "institution": "Amazon", "country": "US", "position": "Amazon Scholar" }, { "name": "Souradip Chakraborty", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "PhD student" }, { "name": "Amrit Bedi", "gender": "Male", "institution": "University of Central Florida", "country": "US", "position": "Assistant Professor" } ], "n_formula": 18, "n_formula_1": 12, "n_ref_uni": 44, "n_ref": 111, "n_ref_all": 127, "n_fig": 1, "n_tab": 7, "L_tab": 6677, "n_element_tab": 499, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 554, "n_element_tab_1": 89, "formula_len_all": 1192, "formula_len_all_1": 571, "len_all": 181168, "len_all_1": 64354, "len_abs": 1617, "len_title": 103, "len_sents": 52031, "len_sents_1": 30304, "n_sents": 397, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1567, "title": "LIAR: Leveraging Inverse Alignment to Jailbreak LLMs in Seconds", "abs": "Many existing jailbreak techniques rely on solving discrete combinatorial optimization, while more recent approaches involve training LLMs to generate multiple adversarial prompts. However, both approaches require significant computational resources to produce even a single adversarial prompt. We hypothesize that the inefficiency of current approaches stems from an inadequate characterization of the jailbreak problem. To address this gap, we formulate the jailbreak problem as an *inverse alignment problem*. By starting from an available safety-aligned model, we leverage an unsafe reward to guide the safe model towards generating unsafe outputs using alignment techniques (e.g., reinforcement learning from human feedback), effectively performing inverse AI alignment. We propose a novel jailbreak method called LIAR (**L**everaging **I**nverse **A**lignment to jailb**R**eak). To demonstrate the simplicity and effectiveness of our approach, we employ a best-of-$N$ method. LIAR offers significant advantages: lower computational requirements without additional training, fully black-box operation, competitive attack success rates, and more human-readable prompts. We provide theoretical insights into the possibility of jailbreaking a safety-aligned model, revealing inherent vulnerabilities in current alignment strategies for LLMs. We also provide sub-optimality guarantees for the proposed LIAR. Experimentally, we achieve ASR comparable to the SoTA with a 10x improvement to perplexity and a Time-to-Attack measured in seconds rather than tens of hours.", "keywords": [ "LLM", "jailbreak" ], "rating_list": [ 5, 3, 6, 5, 6 ], "soundness_list": [ 2, 2, 3, 2, 3 ], "presentation_list": [ 3, 1, 3, 2, 3 ], "contribution_list": [ 2, 3, 3, 1, 3 ], "confidence_list": [ 4, 4, 4, 4, 3 ] }, { "paper_id": "CbPifku2Un", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Lei Yuan", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Researcher" }, { "name": "Lihe Li", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "MS student" }, { "name": "Ruiqi Xue", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Ziqian Zhang", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Yang Yu", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Professor" } ], "n_formula": 41, "n_formula_1": 17, "n_ref_uni": 66, "n_ref": 120, "n_ref_all": 151, "n_fig": 14, "n_tab": 6, "L_tab": 1816, "n_element_tab": 276, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 26, "n_element_tab_1": 2, "formula_len_all": 2912, "formula_len_all_1": 1229, "len_all": 238489, "len_all_1": 68948, "len_abs": 1493, "len_title": 124, "len_sents": 73629, "len_sents_1": 33560, "n_sents": 554, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1542, "title": "Safe Multi-task Pretraining with Constraint Prioritized Decision Transformer", "abs": "Learning a safe policy from offline data without interacting with the environment is crucial for deploying reinforcement learning (RL) policies. Recent approaches leverage transformers to address tasks under various goals, demonstrating a strong generalizability for broad applications. However, these methods either completely overlook safety concerns during policy deployment or simplify safe RL as a dual-objective problem, disregarding the differing priorities between costs and rewards, as well as the additional challenge of multi-task identification caused by cost sparsity. To address these issues, we propose \\textbf{S}afe \\textbf{M}ulti-t\\textbf{a}sk Pretraining with \\textbf{Co}nstraint Prioritized Decision \\textbf{T}ransformer (SMACOT), which utilizes the Decision Transformer (DT) to accommodate varying safety threshold objectives during policy deployment while ensuring scalability. It introduces a Constraint Prioritized Return-To-Go (CPRTG) token to emphasize cost priorities in the Transformer\u2019s inference process, effectively balancing reward maximization with safety constraints. Additionally, a Constraint Prioritized Prompt Encoder is designed to leverage the sparsity of cost information for task identification. Extensive experiments on the public OSRL dataset demonstrate that SMACOT achieves exceptional safety performance in both single-task and multi-task scenarios, satisfying different safety constraints in over 2x as many environments compared with strong baselines, showcasing its superior safety capability.", "keywords": [ "Reinforcement Learning", "Offline Learning", "Safe Reinforcement Learning", "Multi-task Pretrain" ], "rating_list": [ 6, 5, 8, 3 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 1 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 4, 3, 4 ] }, { "paper_id": "Cb4YXpqBIc", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Basura Fernando", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Heng Huang", "gender": "Male", "institution": "Department of Computer Science, University of Maryland, College Park", "country": "US", "position": "Full Professor" }, { "name": "Qiang Sun", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Associate Professor" }, { "name": "Yuke Li", "gender": "unknown", "institution": "University of Maryland, College Park", "country": "US", "position": "Researcher" }, { "name": "Zheng Wang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "Full Professor" }, { "name": "Zhengwei Yang", "gender": "Male", "institution": "Wuhan University", "country": "CN", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 54, "n_ref": 93, "n_ref_all": 107, "n_fig": 7, "n_tab": 4, "L_tab": 2225, "n_element_tab": 155, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1406, "n_element_tab_1": 95, "formula_len_all": 960, "formula_len_all_1": 944, "len_all": 163970, "len_all_1": 64574, "len_abs": 1658, "len_title": 119, "len_sents": 42085, "len_sents_1": 30384, "n_sents": 265, "n_sents_1": 199, "venue": "ICLR2025", "decision": "withdraw", "L_title": 71, "L_abs": 1666, "title": "Cross-Modal Few-Shot Learning: a Generative Transfer Learning Framework", "abs": "Most existing studies on few-shot learning focus on unimodal settings, where models are trained to generalize on unseen data using only a small number of labeled examples from the same modality. However, real-world data are inherently multi-modal, and unimodal approaches limit the practical applications of few-shot learning. To address this gap, this paper introduces the Cross-modal Few-Shot Learning (CFSL) task, which aims to recognize instances from multiple modalities when only a few labeled examples are available. \nThis task presents additional challenges compared to classical few-shot learning due to the distinct visual characteristics and structural properties unique to each modality. To tackle these challenges, we propose a Generative Transfer Learning (GTL) framework consisting of two stages: the first stage involves training on abundant unimodal data, and the second stage focuses on transfer learning to adapt to novel data. Our GTL framework jointly estimates the latent shared concept across modalities and in-modality disturbance in both stages, while freezing the generative module during the transfer phase to maintain the stability of the learned representations and prevent overfitting to the limited multi-modal samples. Our finds demonstrate that GTL has superior performance compared to state-of-the-art methods across four distinct multi-modal datasets: Sketchy, TU-Berlin, Mask1K, and SKSF-A. Additionally, the results suggest that the model can estimate latent concepts from vast unimodal data and generalize these concepts to unseen modalities using only a limited number of available samples, much like human cognitive processes.", "keywords": [ "Few-Shot Classification", "Corss-modality Recognition", "Transfer Learning" ], "rating_list": [ 6, 3, 6, 5 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 3, 1, 3, 2 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "CahIEKCu5Q", "primary_area": "datasets and benchmarks", "n_author": 7, "author_info_list": [ { "name": "Dung Manh Nguyen", "gender": "Male", "institution": "Vietnam National University Hanoi", "country": "", "position": "Undergrad student" }, { "name": "Nam V. Nguyen", "gender": "Male", "institution": "FPT Software AI Center", "country": "VN", "position": "Researcher" }, { "name": "Nghi D. Q. Bui", "gender": "Male", "institution": "SalesForce.com", "country": "SG", "position": "Researcher" }, { "name": "Quang Pham", "gender": "Male", "institution": "A*STAR", "country": "", "position": "Researcher" }, { "name": "Thang Chau Phan", "gender": "Male", "institution": "University of Information Technology", "country": "", "position": "Undergrad student" }, { "name": "Tien-Thong Doan", "gender": "Male", "institution": "FPT", "country": "VN", "position": "Researcher" }, { "name": "Le Hai Nam", "gender": "Male", "institution": "Hanoi University of Science and Technology", "country": "", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 0, "n_ref_uni": 38, "n_ref": 63, "n_ref_all": 93, "n_fig": 10, "n_tab": 10, "L_tab": 10971, "n_element_tab": 1400, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 3236, "n_element_tab_1": 274, "formula_len_all": 134, "formula_len_all_1": 0, "len_all": 337540, "len_all_1": 60156, "len_abs": 1211, "len_title": 139, "len_sents": 60014, "len_sents_1": 25481, "n_sents": 481, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 90, "L_abs": 1214, "title": "CodeMMLU: A Multi-Task Benchmark for Assessing Code Understanding Capabilities of CodeLLMs", "abs": "Recent advancements in Code Large Language Models (CodeLLMs) have predominantly focused on open-ended code generation tasks, often neglecting the critical aspect of code understanding and comprehension. To bridge this gap, we present CodeMMLU, a comprehensive multiple-choice question-answer benchmark designed to evaluate the depth of software and code understanding in LLMs. CodeMMLU includes nearly 20,000 questions sourced from diverse domains, encompassing tasks such as code analysis, defect detection, and software engineering principles across multiple programming languages. Unlike traditional benchmarks, CodeMMLU assesses models\u2019 ability to reason about code rather than merely generate it, providing deeper insights into their grasp of complex software concepts and systems. Our extensive evaluation reveals that even state-of-the-art models face significant challenges with CodeMMLU, highlighting deficiencies in comprehension beyond code generation. By underscoring the crucial relationship between code understanding and effective generation, CodeMMLU serves as a vital resource for advancing AI-assisted software development, ultimately aiming to create more reliable and capable coding assistants.", "keywords": [ "LLM", "CodeLLM", "Benchmark", "Evaluate", "Code Comprehension" ], "rating_list": [ 8, 6, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 3, 3, 1 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "CagdoUkvvl", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Evelyn Chee", "gender": "Female", "institution": "National University of Singapore", "country": "US", "position": "PhD student" }, { "name": "Mong-Li Lee", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" }, { "name": "Wynne Hsu", "gender": "Female", "institution": "National University of Singapore", "country": "SG", "position": "Full Professor" } ], "n_formula": 14, "n_formula_1": 12, "n_ref_uni": 33, "n_ref": 56, "n_ref_all": 72, "n_fig": 8, "n_tab": 4, "L_tab": 943, "n_element_tab": 100, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 858, "n_element_tab_1": 41, "formula_len_all": 886, "formula_len_all_1": 658, "len_all": 114225, "len_all_1": 60591, "len_abs": 1109, "len_title": 144, "len_sents": 32619, "len_sents_1": 28435, "n_sents": 225, "n_sents_1": 202, "venue": "ICLR2025", "decision": "withdraw", "L_title": 96, "L_abs": 1114, "title": "Relaxing Representation Alignment with Knowledge Preservation for Multi-Modal Continual Learning", "abs": "In continual learning, developing robust representations that adapt to new distributions or classes while retaining prior knowledge is crucial. While most traditional approaches focus on single-modality data, multi-modal learning offers significant advantages by leveraging diverse sensory inputs, akin to human perception. However, transitioning to multi-modal continual learning introduces additional challenges as the model needs to effectively combine new information from different modalities while avoiding catastrophic forgetting. In this work, we propose a relaxed cross-modality representation alignment loss and utilize a dual-learner framework to preserve the relation between previously learned representations. We validate our framework using several multi-modal datasets that encompass various types of input modalities. Results show that we consistently outperform baseline continual learning methods in both class incremental and domain incremental learning scenarios. Further analysis highlights the effectiveness of our solution in preserving prior knowledge while incorporating new information.", "keywords": [ "Continual learning", "Multi-modal" ], "rating_list": [ 5, 5, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "CaexTRYaN6", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Haonan Wang", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Jianyang Gu", "gender": "Male", "institution": "The Ohio State University", "country": "US", "position": "Postdoc" }, { "name": "Ruoxi Jia", "gender": "unknown", "institution": "Virginia Tech", "country": "US", "position": "Assistant Professor" }, { "name": "Saeed Vahidian", "gender": "male", "institution": "Duke University", "country": "US", "position": "Postdoc" }, { "name": "Vyacheslav Kungurtsev", "gender": "Male", "institution": "Czech Technical Univeresity in Prague, Czech Technical University of Prague", "country": "CZ", "position": "Researcher" }, { "name": "Wei Jiang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Yiran Chen", "gender": "Male", "institution": "Duke University", "country": "US", "position": "Professor" } ], "n_formula": 16, "n_formula_1": 17, "n_ref_uni": 51, "n_ref": 107, "n_ref_all": 128, "n_fig": 15, "n_tab": 8, "L_tab": 2002, "n_element_tab": 172, "n_fig_1": 5, "n_tab_1": 7, "L_tab_1": 1625, "n_element_tab_1": 137, "formula_len_all": 943, "formula_len_all_1": 1096, "len_all": 182747, "len_all_1": 70910, "len_abs": 1510, "len_title": 108, "len_sents": 56489, "len_sents_1": 31871, "n_sents": 421, "n_sents_1": 232, "venue": "ICLR2025", "decision": "withdraw", "L_title": 60, "L_abs": 1517, "title": "CONCORD: Concept-informed Diffusion for Dataset Distillation", "abs": "Dataset distillation has witnessed significant progress in synthesizing small-scale datasets that encapsulate rich information from large-scale original ones. Particularly, methods based on generative priors show promising performance, while maintaining computational efficiency and cross-architecture generalization. However, the generation process lacks explicit controllability for each sample. Previous distillation methods primarily match the real distribution from the perspective of the entire dataset, whereas overlooking conceptual completeness at the instance level. This oversight can result in missing or incorrectly represented object details and compromised dataset quality. To this end, we propose to incorporate the conceptual understanding of large language models (LLMs) to perform a CONCept-infORmed Diffusion process for dataset distillation, in short as CONCORD. Specifically, distinguishable and fine-grained concepts are retrieved based on category labels to explicitly inform the denoising process and refine essential object details. By integrating these concepts, the proposed method significantly enhances both the controllability and interpretability of the distilled image generation, without replying on pre-trained classifiers. We demonstrate the efficacy of CONCORD by achieving state-of-the-art performance on ImageNet-1K and its subsets. It further advances the practical application of dataset distillation methods. The code implementation is attached in the supplementary material.", "keywords": [ "dataset distillation", "diffusion model" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "CaRkGrdewB", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 5, "author_info_list": [ { "name": "Caren Han", "gender": "Female", "institution": "University of Melbourne", "country": "AU", "position": "Associate Professor" }, { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Rena Wei Gao", "gender": "Female", "institution": "University of Melbourne", "country": "AU", "position": "PhD student" }, { "name": "Siwen Luo", "gender": "Female", "institution": "University of Western Australia", "country": "AU", "position": "Lecturer" }, { "name": "Xuetong Wu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Researcher" } ], "n_formula": 7, "n_formula_1": 5, "n_ref_uni": 61, "n_ref": 94, "n_ref_all": 114, "n_fig": 8, "n_tab": 7, "L_tab": 2654, "n_element_tab": 176, "n_fig_1": 7, "n_tab_1": 3, "L_tab_1": 1098, "n_element_tab_1": 51, "formula_len_all": 457, "formula_len_all_1": 337, "len_all": 142444, "len_all_1": 63221, "len_abs": 1256, "len_title": 122, "len_sents": 37360, "len_sents_1": 31531, "n_sents": 237, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1209, "title": "'No' Matters: Out-of-Distribution Detection in Multimodality Long Dialogue", "abs": "Out-of-distribution (OOD) detection in multimodal contexts is essential for identifying deviations in combined inputs from different modalities, particularly in applications like open-domain dialogue systems or real-life dialogue interactions. This paper aims to improve the user experience that involves multi-round long dialogues by efficiently detecting OOD dialogues and images. We introduce a novel scoring framework named **D**ialogue **I**mage **A**ligning and **E**nhancing **F**ramework (DIAEF) that integrates the visual language models with the novel proposed scores that detect OOD in two key scenarios (1) mismatches between the dialogue and image input pair and (2) input pairs with previously unseen labels. Our experimental results, derived from various benchmarks, demonstrate that integrating image and multi-round dialogue OOD detection is more effective with previously unseen labels than using either modality independently. In the presence of mismatched pairs, our proposed score effectively identifies these mismatches and demonstrates strong robustness in long dialogues. This approach enhances domain-aware, adaptive conversational agents and establishes baselines for future studies.", "keywords": [ "Natural Language Processing", "Out-of-distribution Detection", "Machine Learning", "Multimodality Dialogue" ], "rating_list": [ 3, 5, 5, 5, 5 ], "soundness_list": [ 2, 2, 2, 2, 2 ], "presentation_list": [ 3, 3, 3, 3, 2 ], "contribution_list": [ 1, 3, 2, 2, 2 ], "confidence_list": [ 2, 2, 3, 4, 3 ] }, { "paper_id": "CaNp8ALCRT", "primary_area": "reinforcement learning", "n_author": 6, "author_info_list": [ { "name": "Bo Yuan", "gender": "unknown", "institution": "Harvard University", "country": "", "position": "PhD student" }, { "name": "Jan B\u00edma", "gender": "Male", "institution": ", Charles University Prague", "country": "CZ", "position": "PhD student" }, { "name": "Otto Ritter", "gender": "Male", "institution": "MSD", "country": "CZ", "position": "Principal Researcher" }, { "name": "Sean L. Wu", "gender": "Male", "institution": "MERCK & CO., INC.", "country": "US", "position": "Researcher" }, { "name": "Tianchi Chen", "gender": "Male", "institution": "MERCK & CO., INC.", "country": "US", "position": "Researcher" }, { "name": "Xiang Yu", "gender": "unknown", "institution": "MERCK & CO., INC.", "country": "US", "position": "Principal Researcher" } ], "n_formula": 9, "n_formula_1": 10, "n_ref_uni": 37, "n_ref": 42, "n_ref_all": 45, "n_fig": 2, "n_tab": 1, "L_tab": 267, "n_element_tab": 39, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 478, "n_element_tab_1": 27, "formula_len_all": 425, "formula_len_all_1": 484, "len_all": 106124, "len_all_1": 58867, "len_abs": 1383, "len_title": 150, "len_sents": 32911, "len_sents_1": 27102, "n_sents": 222, "n_sents_1": 180, "venue": "ICLR2025", "decision": "Reject", "L_title": 92, "L_abs": 1384, "title": "Implicit Bayesian Markov Decision Process for Resource Efficient Decisions in Drug Discovery", "abs": "In drug discovery, researchers make sequential decisions to schedule experiments, aiming to maximize the probability of success towards drug candidates while simultaneously minimizing expected costs. However, such tasks pose significant challenges due to complex trade-offs between uncertainty reduction and allocation of constrained resources in a high-dimensional state-action space. Traditional methods based on simple rule-based heuristics or domain expertise often result in either inefficient resource utilization due to risk aversion or missed opportunities due to reckless decisions. To address these challenges, we developed an Implicit Bayesian Markov Decision Process (IB-MDP) algorithm that constructs an explicit MDP model of the environment\u2019s dynamics by integrating historical data through a similarity-based metric and enables effective planning by simulating future states and actions. To enhance the robustness of the decision-making process, the IB-MDP also incorporates an ensemble approach that recommends maximum likelihood actions to effectively balance the dual objectives of reducing state uncertainty and optimizing expected costs. Our experimental results demonstrate that the IB-MDP algorithm offers significant improvements over traditional rule-based methods by identifying optimal decisions that ensure more efficient use of resources in drug discovery.", "keywords": [ "Bayesian Markov Decision Process", "ensemble approach", "similarity-based metric", "sequential decision making" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 3, 1, 2 ], "presentation_list": [ 2, 1, 1, 1 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "CZvbXXgjrn", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Chao Ren", "gender": "Male", "institution": "KTH Royal Institute of Technology", "country": "SE", "position": "Principal Researcher" }, { "name": "Han Yu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Sheng-Jun Huang", "gender": "unknown", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Xiaoxiao Li", "gender": "Unspecified", "institution": "University of British Columbia", "country": "CA", "position": "Assistant Professor" }, { "name": "Ying-Peng Tang", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 13, "n_ref_uni": 30, "n_ref": 53, "n_ref_all": 74, "n_fig": 6, "n_tab": 2, "L_tab": 1376, "n_element_tab": 239, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 1136, "n_element_tab_1": 194, "formula_len_all": 624, "formula_len_all_1": 691, "len_all": 117927, "len_all_1": 68603, "len_abs": 1421, "len_title": 100, "len_sents": 39641, "len_sents_1": 29858, "n_sents": 322, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 1429, "title": "Semantic-aligned Query Synthesis for Active Learning", "abs": "Active learning (AL) reduces data annotation costs by querying labels from human annotators for the most informative unlabeled data points during model training. Existing AL methods generally assume the availability of a large amount of unlabeled samples for query selection. However, collecting raw data in practice can be expensive, even without considering the cost of labeling. Membership query synthesis circumvents the need for an unlabeled data pool by directly generating informative queries from the input space. Nevertheless, existing approaches often generate instances lacking semantic meaning, thereby increasing the difficulty of labeling. In this paper, we propose the Generative Membership Query Descriptor (GenMQD) method for AL to mitigate the risk of generating unrecognizable instances. The key idea is to generate textual descriptions of the desired data, instead of the data samples themselves. Then a pre-trained multi-modal alignment model (e.g., CLIP) can be leveraged to transform these features into natural language texts for data gathering purposes. Extensive experiments on image classification benchmark datasets against query synthesis state-of-the-art methods demonstrate that, on average, GenMQD can improve model accuracy by 2.43\\% when gathering and labeling 500 examples. A large-scale user study verifies that human oracles prefer GenMQD generated queries over generated image-based queries.", "keywords": [ "Active learning", "Data synthesis", "Machine learning", "Influence function" ], "rating_list": [ 5, 5, 3, 5, 3 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 2, 4, 2 ], "contribution_list": [ 3, 3, 2, 3, 2 ], "confidence_list": [ 5, 5, 4, 4, 4 ] }, { "paper_id": "CZiP7GpmX7", "primary_area": "learning on time series and dynamical systems", "n_author": 4, "author_info_list": [ { "name": "Bin Li", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Chuhong Yang", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "", "position": "MS student" }, { "name": "Nan Wu", "gender": "Male", "institution": "Beijing Institute of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Yuanjing Qi", "gender": "Male", "institution": "BIT", "country": "CN", "position": "PhD student" } ], "n_formula": 55, "n_formula_1": 12, "n_ref_uni": 17, "n_ref": 41, "n_ref_all": 102, "n_fig": 26, "n_tab": 19, "L_tab": 6513, "n_element_tab": 673, "n_fig_1": 14, "n_tab_1": 8, "L_tab_1": 1838, "n_element_tab_1": 216, "formula_len_all": 4311, "formula_len_all_1": 487, "len_all": 175358, "len_all_1": 64186, "len_abs": 1282, "len_title": 123, "len_sents": 59519, "len_sents_1": 27362, "n_sents": 501, "n_sents_1": 182, "venue": "ICLR2025", "decision": "Reject", "L_title": 75, "L_abs": 1287, "title": "FastTF: 4 Parameters are All You Need for Long-term Time Series Forecasting", "abs": "Time series forecasting is essential across various sectors, including finance, transportation, and industry. In this paper, we propose FastTF, a powerful yet lightweight model in Time-Frequency domain for long-term time series forecasting. Our aim is to push the boundary of model lightweighting and facilitate the deployment of lightweight model on resource-constrained devices. Leveraging the global nature and information compressibility of the time series in frequency domain, we introduce patch-wise downsampling, Sparse Frequency Mixer (SFM), and patch predictor to capture the temporal variations of frequency components across different patches. Experimental results on five public datasets demonstrate that FastTF with very few parameters outperforms several state-of-the-art models and demonstrates a strong generalization capability. Notably, on the ETTh1 dataset, FastTF with only 4 parameters achieves a performance that is close to the DLinear and FITS in the horizon-96 forecasting. Furthermore, we deployed our model on a FPGA development board (Zynq UltraScale+ RFSoC ZCU208 Evaluation Kit), where the corresponding resource usage statistics illustrate that our model has a very low computational overhead and latency, making it easily implemented on hardware devices.", "keywords": [ "Time series forcasting", "Machine learning", "Model lightweighting" ], "rating_list": [ 5, 5, 1, 3, 3 ], "soundness_list": [ 3, 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 3, 2, 2 ], "contribution_list": [ 2, 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 5, 4 ] }, { "paper_id": "CYa4FKjYM9", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Antonio Orvieto", "gender": "Male", "institution": "ELLIS Institute T\u00fcbingen, Max Planck Institute for Intelligent Systems, T\u00fcbingen AI Center, T\u00fcbingen, Germany", "country": "CH", "position": "Principal Researcher" }, { "name": "Aurelien Lucchi", "gender": "Male", "institution": "University of Basel", "country": "CH", "position": "Assistant Professor" }, { "name": "Niccol\u00f2 Ajroldi", "gender": "Male", "institution": "Max Planck Institut, Max-Planck Institute", "country": "DE", "position": "Research Engineer" }, { "name": "Rustem Islamov", "gender": "Male", "institution": "University of Basel", "country": "CH", "position": "PhD student" } ], "n_formula": 117, "n_formula_1": 15, "n_ref_uni": 77, "n_ref": 144, "n_ref_all": 194, "n_fig": 15, "n_tab": 5, "L_tab": 3138, "n_element_tab": 344, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 562, "n_element_tab_1": 12, "formula_len_all": 17146, "formula_len_all_1": 2362, "len_all": 249425, "len_all_1": 61890, "len_abs": 1219, "len_title": 71, "len_sents": 58656, "len_sents_1": 26407, "n_sents": 541, "n_sents_1": 189, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1528, "title": "Enhancing Optimizer Stability: Momentum Adaptation of NGN Step-size", "abs": "Modern optimization algorithms that incorporate momentum and adaptive step-size offer improved performance in various challenging Deep Learning tasks. However, their effectiveness is often highly sensitive to the choice of hyper-parameters, especially the learning rate. Tuning these parameters is often difficult, resource-intensive, and time-consuming. State-of-the-art optimization algorithms incorporating momentum and adaptive step size are the algorithms of choice in several challenging Deep Learning domains. However, their effectiveness is frequently dependent on selecting the right hyper-parameters, especially the learning rate. Therefore, recent efforts have been directed toward enhancing the stability of optimizers across a wide range of hyper-parameter choices (Schaipp et al., 2024). In this paper, we introduce an algorithm that matches the performance of state-of-the-art optimizers while improving stability through a novel adaptation of the NGN step-size method (Orvieto & Xiao, 2024). Specifically, we propose a momentum-based version (NGN-M) that attains the standard convergence rate of $\\mathcal{O}(1/\\sqrt{K})$ under common assumptions, without the need for interpolation condition or assumptions of bounded stochastic gradients or iterates, in contrast to previous approaches. Additionally, we empirically demonstrate that the combination of the NGN step-size with momentum results in high robustness while delivering performance that is comparable to or surpasses other state-of-the-art optimizers.", "keywords": [ "Optimization", "Adaptive Methods", "Polyak stepsize", "Machine Learning" ], "rating_list": [ 8, 5, 6, 5 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 3, 2, 3, 3 ] }, { "paper_id": "CYUIeEBri1", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Dongjie Fu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Huadai Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Rongjie Huang", "gender": "Male", "institution": "FAIR", "country": "US", "position": "Researcher" }, { "name": "Tao Jin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Xiaoda Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Xize Cheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zehan Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhou Zhao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 9, "n_ref_uni": 37, "n_ref": 108, "n_ref_all": 125, "n_fig": 2, "n_tab": 5, "L_tab": 2309, "n_element_tab": 208, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1497, "n_element_tab_1": 111, "formula_len_all": 472, "formula_len_all_1": 506, "len_all": 137789, "len_all_1": 62309, "len_abs": 1407, "len_title": 111, "len_sents": 40529, "len_sents_1": 30090, "n_sents": 270, "n_sents_1": 181, "venue": "ICLR2025", "decision": "withdraw", "L_title": 63, "L_abs": 1423, "title": "Noise-Robust Audio-Visual Speech-Driven Body Language Synthesis", "abs": "With the continuous advancement of video generation, researchers have achieved speech-driven body language synthesis, such as co-speech gestures. However, due to the lack of paired data for visual speech (i.e., lip movements) and body languages, existing methods typically rely solely on audio-only speech, which struggles to correctly synthesize target results in noisy environments. To overcome this limitation, we propose an Audio-Visual Speech-Driven Synthesis (**AV-SDS**) method tailored for body language synthesis, aiming for robust synthesis even under noisy conditions. Given that each body language modality data has its corresponding audio speech, AV-SDS adopts a two-stage synthesis framework based on speech discrete units, consisting of the AV-S2UM and Unit2X modules. It uses speech discrete units as carriers to construct a direct mapping from audio-visual speech to each body language. Considering the distinct characteristics of different body languages, AV-SDS can be implemented based on semantic and acoustic discrete units, respectively, to achieve high-semantic and high-rhythm body language synthesis. Experimental results demonstrate that our AV-SDS achieves superior performance in synthesizing multiple body language modalities in noisy environments, delivering noise-robust body language synthesis. For samples and further information, please visit demo page at \\url{https://av-sds.github.io/}.", "keywords": [ "Body Language Synthesis", "speech driven", "noise-robust" ], "rating_list": [ 5, 3, 3, 6, 5 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 1, 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "CY6bFF9A3O", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Hannah Erlebach", "gender": "Female", "institution": "University College London, University of London", "country": "GB", "position": "MS student" }, { "name": "Jonathan Cook", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 36, "n_ref": 81, "n_ref_all": 97, "n_fig": 8, "n_tab": 4, "L_tab": 750, "n_element_tab": 74, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 423, "n_element_tab_1": 36, "formula_len_all": 488, "formula_len_all_1": 161, "len_all": 120707, "len_all_1": 64047, "len_abs": 1012, "len_title": 104, "len_sents": 41563, "len_sents_1": 32545, "n_sents": 248, "n_sents_1": 192, "venue": "ICLR2025", "decision": "Reject", "L_title": 56, "L_abs": 1018, "title": "RACCOON: Regret-based Adaptive Curricula for Cooperation", "abs": "Overfitting to training partners is a common problem in fully-cooperative multi-agent settings, leading to poor zero-shot transfer to novel partners. A popular solution is to train an agent with a diverse population of training partners. However, previous work lacks a principled approach for selecting partners from this population during training, usually sampling at random. We argue that partner sampling is an important and overlooked problem, and motivated by the success of regret-based Unsupervised Environment Design, we propose Regret-based Adaptive Curricula for Cooperation (RACCOON), a novel a method which prioritises high-regret partners and tasks. We test RACCOON in the Overcooked environment, and demonstrate that it leads to sample efficiency gains and increased robustness across diverse partners and tasks, compared with strong baselines. We further analyse the nature of the induced curricula, and conclude with discussions on the limitations of cooperative regret and directions for future work.", "keywords": [ "unsupervised environment design", "multi-agent reinforcement learning", "cooperation", "autocurricula", "ad-hoc teamwork", "zero-shot coordination" ], "rating_list": [ 6, 8, 5, 5 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 2, 4, 2, 4 ], "contribution_list": [ 2, 4, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "CXS3cIb5Dc", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Li Sun", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Pengfei Li", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "MS student" }, { "name": "Qingli Li", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 17, "n_ref": 25, "n_ref_all": 38, "n_fig": 4, "n_tab": 3, "L_tab": 1870, "n_element_tab": 430, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1870, "n_element_tab_1": 430, "formula_len_all": 1221, "formula_len_all_1": 1221, "len_all": 112196, "len_all_1": 63861, "len_abs": 1863, "len_title": 162, "len_sents": 28213, "len_sents_1": 25964, "n_sents": 200, "n_sents_1": 186, "venue": "ICLR2025", "decision": "withdraw", "L_title": 112, "L_abs": 1871, "title": "Distilling Cross-Domain Knowledge for Person Re-ID by Aligning Any Pretrained Encoder with CLIP Textual Features", "abs": "Based on the alignment of image-text pairs, CLIP has demonstrated superior performance across various tasks, even in a zero-shot setting. In person ReID, CLIP-based models achieve state-of-the-art results without explicit text descriptions for further fine-tuning. However, previous models are primarily initialized with weights from ImageNet or self-supervised methods, lacking cross-domain knowledge in both image and text areas. This paper introduces a novel approach that aligns a pure image-domain pretrained student model with CLIP textual features, distilling cross-domain knowledge from existing CLIP-ReID into the online student model. To leverage CLIP\u2019s textual features for each ID, we address the challenge of mismatched feature dimensions between the teacher and student. A trainable adapter is inserted on the student side to match dimensions and preserve the prior knowledge within the pretrained student. For the student encoder yielding lower or equal-dimensional features compared to the teacher, the adapter is initialized as an identity matrix, while offline PCA is employed on the teacher side for dimensionality reduction. PCA eigenvectors are computed from all training images and applied to existing text features for matching with the student. In cases where the student outputs exceed the teacher's dimensions, the adapter is initialized using eigenvectors computed from the student side to retain knowledge in the pretrained student model. After dimension alignment, text features for each ID are compared with online image features, specifying cross-domain similarities, which are further constrained to mimic the teacher through a KL-divergence loss. Experiments with different pretraining encoder structures demonstrate the effectiveness of this approach, which is also compatible with relation knowledge distillation to enhance performance.", "keywords": [ "CLIP; Person ReID; Knowledge Distillation" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 2, 1, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 5, 5, 5, 3 ] }, { "paper_id": "CXIiV1iU3G", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Dongwen Tang", "gender": "Male", "institution": "Xi'an University of Electronic Science and Technology", "country": "CN", "position": "Undergrad student" }, { "name": "Kai Wang", "gender": "Male", "institution": "national university of singaore, National University of Singapore", "country": "SG", "position": "PhD student" }, { "name": "Yang You", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "Professor" }, { "name": "ZHAO WANGBO", "gender": "Male", "institution": "National University of Singapore", "country": "SG", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 64, "n_ref": 142, "n_ref_all": 162, "n_fig": 11, "n_tab": 13, "L_tab": 2476, "n_element_tab": 305, "n_fig_1": 7, "n_tab_1": 6, "L_tab_1": 1739, "n_element_tab_1": 160, "formula_len_all": 443, "formula_len_all_1": 443, "len_all": 183613, "len_all_1": 60556, "len_abs": 956, "len_title": 104, "len_sents": 49024, "len_sents_1": 24741, "n_sents": 455, "n_sents_1": 231, "venue": "ICLR2025", "decision": "withdraw", "L_title": 56, "L_abs": 968, "title": "Recurrent Diffusion for Large-Scale Parameter Generation", "abs": "Parameter generation has struggled to scale up for a long time, significantly lim-\niting its range of applications. In this study, we introduce Recurrent diffusion for\nlarge-scale Parameter Generation, called RPG. We first divide the trained parame-\nters into non-overlapping parts, after which a recurrent model is proposed to learn\ntheir relationships. The recurrent model\u2019s outputs, as conditions, are then fed into\na diffusion model to generate the neural network parameters. Using only a sin-\ngle GPU, recurrent diffusion enables us to generate popular vision and language\nmodels such as ConvNeXt-L and LoRA parameters of LLaMA-7B. Meanwhile,\nacross various architectures and tasks, the generated parameters consistently per-\nform comparable results over trained networks. Notably, our approach also shows\nthe potential to generate models for handling unseen tasks. This suggests that\nrecurrent diffusion largely increases the practicality of parameter generation", "keywords": [ "parameter generation" ], "rating_list": [ 3, 8, 5, 3, 5 ], "soundness_list": [ 1, 4, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3, 2 ] }, { "paper_id": "CX0Z5c0LbN", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Donghwan Kim", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "MS student" }, { "name": "Hyunsoo Yoon", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Associate Professor" }, { "name": "Jung Hun Phee", "gender": "Male", "institution": "Yonsei University", "country": "KR", "position": "Undergrad student" } ], "n_formula": 18, "n_formula_1": 18, "n_ref_uni": 48, "n_ref": 85, "n_ref_all": 95, "n_fig": 1, "n_tab": 3, "L_tab": 869, "n_element_tab": 170, "n_fig_1": 1, "n_tab_1": 4, "L_tab_1": 1017, "n_element_tab_1": 174, "formula_len_all": 1144, "formula_len_all_1": 1144, "len_all": 115198, "len_all_1": 66863, "len_abs": 897, "len_title": 175, "len_sents": 31475, "len_sents_1": 30699, "n_sents": 213, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 126, "L_abs": 900, "title": "Challenging the Counterintuitive: Revisiting Simple Likelihood Tests with Normalizing Flows for Tabular Data Anomaly Detection", "abs": "In this study, we propose a novel approach to anomaly detection in the tabular domain using normalizing flows, leveraging a simple likelihood test to achieve state-of-the-art performance in unsupervised learning. Although simple likelihood tests have been shown to fail in anomaly detection for image data, we redefine the counterintuitive phenomenon and demonstrate, both theoretically and empirically, why this method succeeds in the tabular domain. Our approach outperforms traditional anomaly detection methods by offering more consistent results. Furthermore, we question the practice of fine-tuning parameters for each dataset individually, ensuring fair and unbiased comparisons by adopting uniform hyperparameters across all datasets. Through extensive experimentation, we validate the robustness and scalability of our method, highlighting its practical effectiveness in real-world settings.", "keywords": [ "anomaly detection", "tabular data", "self-supervised learning", "generative model" ], "rating_list": [ 3, 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 2, 1 ], "presentation_list": [ 3, 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2, 1 ], "confidence_list": [ 4, 5, 3, 4, 4 ] }, { "paper_id": "CWAvMSNUqT", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Yi Yang", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Yixuan Tang", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 45, "n_ref": 88, "n_ref_all": 113, "n_fig": 2, "n_tab": 5, "L_tab": 1980, "n_element_tab": 168, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 490, "n_element_tab_1": 38, "formula_len_all": 76, "formula_len_all_1": 58, "len_all": 147983, "len_all_1": 58939, "len_abs": 1773, "len_title": 129, "len_sents": 40337, "len_sents_1": 28717, "n_sents": 318, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1780, "title": "Pooling And Attention: What Are Effective Designs For LLM-Based Embedding Models?", "abs": "The significant advancements of Large Language Models (LLMs) in generative tasks have led to a growing body of work exploring LLM-based embedding models. While these models, employing different pooling and attention strategies, have achieved state-of-the-art performance on public embedding benchmarks, questions still arise about what constitutes an effective design for LLM-based embedding models. However, these models are often trained on different datasets, using different LLM base models or training settings. Moreover, evaluations on public embedding benchmarks often fail to report statistical significance, making it difficult to determine which designs truly contribute to final performance. This complicates the process for practitioners seeking optimal training recipes for LLM-based embedding models. In this study, we conduct a large-scale experiment by training a series of LLM-based embedding models using the same training data and base model but differing in their pooling and attention strategies. The results show that there is no one-size-fits-all solution: while bidirectional attention and an additional trainable pooling layer outperform in text similarity and information retrieval tasks, they do not significantly surpass simpler designs like EOS-last token pooling and default causal attention in clustering and classification tasks. Furthermore, we propose a new pooling strategy, Multi-Layers Trainable Pooling, which transforms the outputs of all hidden layers, rather than just the last layer, using a cross-attention network. This method proves to be statistically superior in text similarity and retrieval tasks compared to existing pooling methods. Overall, this paper sheds light on effective training strategies for LLM-based embedding models.", "keywords": [ "Representation learning; Embedding Model; LLM; Information Retrieval" ], "rating_list": [ 5, 3, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 4, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "CUABD2qIB4", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Bicheng Xu", "gender": "Male", "institution": "University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Dujian Ding", "gender": "unknown", "institution": "Computing Science, University of British Columbia", "country": "CA", "position": "PhD student" }, { "name": "Laks V. S. Lakshmanan", "gender": "unknown", "institution": "University of British Columbia", "country": "CA", "position": "Professor" } ], "n_formula": 25, "n_formula_1": 16, "n_ref_uni": 48, "n_ref": 99, "n_ref_all": 173, "n_fig": 22, "n_tab": 8, "L_tab": 4366, "n_element_tab": 639, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 1375, "n_element_tab_1": 192, "formula_len_all": 2674, "formula_len_all_1": 1552, "len_all": 184515, "len_all_1": 70844, "len_abs": 1011, "len_title": 121, "len_sents": 56470, "len_sents_1": 32204, "n_sents": 433, "n_sents_1": 220, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1014, "title": "OCCAM: Towards Cost-Efficient and Accuracy-Aware Classification Inference", "abs": "Classification tasks play a fundamental role in various applications, spanning domains such as healthcare, natural language processing and computer vision. With the growing popularity and capacity of machine learning models, people can easily access trained classifiers as a service online or offline. However, model use comes with a cost and classifiers of higher capacity (such as large foundation models) usually incur higher inference costs. To harness the respective strengths of different classifiers, we propose a principled approach, OCCAM, to compute the best classifier assignment strategy over classification queries (termed as the optimal model portfolio) so that the aggregated accuracy is maximized, under user-specified cost budgets. Our approach uses an unbiased and low-variance accuracy estimator and effectively computes the optimal solution by solving an integer linear programming problem. On a variety of real-world datasets, OCCAM achieves 40% cost reduction with little to no accuracy drop.", "keywords": [ "Efficient ML", "Hybrid ML Inference", "Classification" ], "rating_list": [ 8, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 3, 4, 4, 3 ] }, { "paper_id": "CU8CNDw6Vv", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Levent Sagun", "gender": "Non-Binary", "institution": "Meta", "country": "", "position": "Research scientist" }, { "name": "Samuel Bell", "gender": "Male", "institution": "Facebook", "country": "GB", "position": "Researcher" }, { "name": "Diane Nicole Bouchacourt", "gender": "Female", "institution": "Facebook AI Research", "country": "", "position": "Researcher" } ], "n_formula": 4, "n_formula_1": 1, "n_ref_uni": 43, "n_ref": 94, "n_ref_all": 135, "n_fig": 24, "n_tab": 4, "L_tab": 1439, "n_element_tab": 175, "n_fig_1": 18, "n_tab_1": 3, "L_tab_1": 1379, "n_element_tab_1": 185, "formula_len_all": 205, "formula_len_all_1": 74, "len_all": 153476, "len_all_1": 61636, "len_abs": 1042, "len_title": 108, "len_sents": 45728, "len_sents_1": 27797, "n_sents": 342, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 60, "L_abs": 1049, "title": "Reassessing the Validity of Spurious Correlations Benchmarks", "abs": "Neural networks can fail when the data contains spurious correlations, i.e. associations in the training data that fail to generalize to new distributions. To understand this phenomenon, often referred to as subpopulation shift or shortcut learning, researchers have proposed numerous group-annotated spurious correlations benchmarks upon which to evaluate mitigation methods. However, we observe that these benchmarks exhibit substantial disagreement, with the best methods on one benchmark performing poorly on another. We explore this disagreement, and examine benchmark validity by defining three desiderata that a benchmark should satisfy in order to meaningfully evaluate methods. Our results have implications for both benchmarks and mitigations: we find that certain group-annotated benchmarks are not meaningful measures of method performance, and that several methods are not sufficiently robust for widespread use. We present a simple recipe for practitioners to choose methods using the _most similar_ benchmark to their given problem.", "keywords": [ "benchmarking", "evaluation", "spurious correlations" ], "rating_list": [ 3, 5, 5, 8 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "CU7QfWJ6nC", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Haonan Qiu", "gender": "unknown", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Menghan Xia", "gender": "Male", "institution": "Kuaishou Technology", "country": "CN", "position": "Researcher" }, { "name": "Yingqing He", "gender": "unknown", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Zhaoxi Chen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Ziwei Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" }, { "name": "Zhouxia Wang", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" } ], "n_formula": 25, "n_formula_1": 20, "n_ref_uni": 34, "n_ref": 96, "n_ref_all": 124, "n_fig": 14, "n_tab": 3, "L_tab": 767, "n_element_tab": 90, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 207, "n_element_tab_1": 19, "formula_len_all": 1638, "formula_len_all_1": 1337, "len_all": 172001, "len_all_1": 60714, "len_abs": 1349, "len_title": 73, "len_sents": 44973, "len_sents_1": 27627, "n_sents": 408, "n_sents_1": 217, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1355, "title": "FreeTraj: Tuning-Free Trajectory Control via Noise Guided Video Diffusion", "abs": "Diffusion model has demonstrated remarkable capability in video generation, which further sparks interest in introducing trajectory control into the generation process. While existing works mainly focus on training-based methods (e.g., conditional adapter), we argue that diffusion model itself allows decent control over the generated content without requiring any training. In this study, we introduce a tuning-free framework to achieve trajectory-controllable video generation, by imposing guidance on both noise construction and attention computation. Specifically, 1) we first show several instructive phenomena and analyze how initial noises influence the motion trajectory of generated content. 2) Subsequently, we propose FreeTraj, a tuning-free approach that enables trajectory control by modifying noise sampling and attention mechanisms. 3) Furthermore, we extend FreeTraj to facilitate longer and larger video generation with controllable trajectories. Equipped with these designs, users have the flexibility to provide trajectories manually or opt for trajectories automatically generated by the LLM trajectory planner. Extensive experiments validate the efficacy of our approach in enhancing the trajectory controllability of video diffusion models. Generated video samples are available at the anonymous website: https://FreeTraj.github.io.", "keywords": [ "Diffusion Model", "Video Diffusion", "Trajectory Control", "Motion Control" ], "rating_list": [ 6, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "CTC7CmirNr", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Hanzi Mao", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Jun Zhu", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "Professor" }, { "name": "Kaiwen Zheng", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "PhD student" }, { "name": "Ming-Yu Liu", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Qinsheng Zhang", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Yongxin Chen", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Principal Researcher" } ], "n_formula": 108, "n_formula_1": 24, "n_ref_uni": 43, "n_ref": 164, "n_ref_all": 219, "n_fig": 18, "n_tab": 4, "L_tab": 934, "n_element_tab": 55, "n_fig_1": 9, "n_tab_1": 3, "L_tab_1": 599, "n_element_tab_1": 44, "formula_len_all": 10402, "formula_len_all_1": 1903, "len_all": 282947, "len_all_1": 72058, "len_abs": 1502, "len_title": 157, "len_sents": 91709, "len_sents_1": 29887, "n_sents": 825, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 108, "L_abs": 1516, "title": "Masked Diffusion Models are Secretly Time-Agnostic Masked Models and Exploit Inaccurate Categorical Sampling", "abs": "Masked diffusion models (MDMs) have emerged as a popular research topic for generative modeling of discrete data, thanks to their superior performance over other discrete diffusion models, and are rivaling the auto-regressive models (ARMs) for language modeling tasks. The recent effort in simplifying the masked diffusion framework further leads to alignment with continuous-space diffusion models and more principled training and sampling recipes. \nIn this paper, however, we reveal that both training and sampling of MDMs are theoretically free from the time variable, arguably the key signature of diffusion models, and are instead equivalent to masked models. The connection on the sampling aspect is drawn by our proposed first-hitting sampler (FHS). Specifically, we show that the FHS is theoretically equivalent to MDMs' original generation process while significantly alleviating the time-consuming categorical sampling and achieving a 20$\\times$ speedup. In addition, our investigation raises doubts about whether MDMs can truly beat ARMs in text generation. We identify, for the first time, an underlying numerical issue, even with the commonly used 32-bit floating-point precision, which results in inaccurate categorical sampling. \nWe show that it lowers the effective temperature both theoretically and empirically, and the resulting decrease in token diversity makes previous evaluations, which assess the generation quality solely through the incomplete generative perplexity metric, somewhat unfair.", "keywords": [ "discrete diffusion models", "masked models", "language modeling" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 4, 3, 4, 3 ], "presentation_list": [ 4, 2, 4, 2 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "CSpkSBe6jn", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Tianyi Zhou", "gender": "Male", "institution": "University of Maryland, College Park", "country": "US", "position": "Assistant Professor" }, { "name": "Yuhan Wang", "gender": "Male", "institution": "Department of Computer Science, University of Wisconsin - Madison", "country": "", "position": "Undergrad student" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 28, "n_ref": 43, "n_ref_all": 59, "n_fig": 8, "n_tab": 2, "L_tab": 60, "n_element_tab": 7, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 977, "formula_len_all_1": 854, "len_all": 126406, "len_all_1": 53943, "len_abs": 1255, "len_title": 83, "len_sents": 33364, "len_sents_1": 26203, "n_sents": 239, "n_sents_1": 184, "venue": "ICLR2025", "decision": "withdraw", "L_title": 35, "L_abs": 1264, "title": "Task-Adaptation Curriculum Learning", "abs": "A large distribution gap between a target task and pre-training tasks could undermine the task adaptation performance of pretrained models. When the target-task data are scarce, naive finetuning results in overfitting and forgetting. In various domains, skills can be transferred across semantically related tasks, among which the general-purposed ones often have more training data. Can we bridge the gap between a pre-trained model and a low-resource target task by leveraging data from other tasks? In this paper, we address the low-resource task adaptation challenge by a transfer learning curriculum, which finetunes a model on a curated sequence of intermediate tasks, thereby progressively bridging the gap between the pre-trained model and the target task. To this end, we formulate the task curriculum as a graph search problem and improve the efficiency of estimating transferability between tasks. Two search algorithms are studied, i.e., greedy best-first search and Monte Carlo tree search. We evaluate our approach, i.e., ``task-adaptation curriculum learning (TaCL)'' on two benchmark settings. Extensive evaluations on different target tasks demonstrate the effectiveness and advantages of TaCL on highly specific and low-resource downstream tasks.", "keywords": [ "Task adaptation", "transfer learning", "curriculum learning", "search algorithms" ], "rating_list": [ 5, 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 3, 3, 5 ] }, { "paper_id": "CSj72Rr2PB", "primary_area": "generative models", "n_author": 2, "author_info_list": [ { "name": "Kun Zhan", "gender": "Male", "institution": "Lanzhou University", "country": "CN", "position": "Full Professor" }, { "name": "Meng Yu", "gender": "Male", "institution": "Lanzhou University", "country": "CN", "position": "MS student" } ], "n_formula": 29, "n_formula_1": 16, "n_ref_uni": 23, "n_ref": 52, "n_ref_all": 98, "n_fig": 8, "n_tab": 6, "L_tab": 842, "n_element_tab": 84, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 722, "n_element_tab_1": 78, "formula_len_all": 1847, "formula_len_all_1": 1120, "len_all": 118493, "len_all_1": 58731, "len_abs": 1039, "len_title": 89, "len_sents": 41901, "len_sents_1": 27724, "n_sents": 298, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 41, "L_abs": 963, "title": "Bias Mitigation in Graph Diffusion Models", "abs": "Most existing graph diffusion models have significant bias problems. We observe that the forward diffusion\u2019s maximum perturbation distribution in most models deviates from the standard Gaussian distribution, while reverse sampling consistently starts from a standard Gaussian distribution, which results in a reverse-starting bias. Together with the inherent exposure bias of diffusion models, this results in degraded generation quality. This paper proposes a unified approach to mitigate both biases. To mitigate reverse-starting bias, we employ a newly designed Langevin sampling algorithm to align with the forward maximum perturbation distribution, establishing a new reverse-starting point. To address the exposure bias, we introduce a fraction correction mechanism based on a newly defined score difference. Our approach, which requires no network modifications, is validated across multiple models, datasets, and tasks, achieving state-of-the-art results.", "keywords": [ "Diffusion models", "Graph learning", "Bias analysis" ], "rating_list": [ 6, 6, 8, 6 ], "soundness_list": [ 3, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "CSZKElOtG5", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Amir Houmansadr", "gender": "Male", "institution": "University of Massachusetts, Amherst", "country": "US", "position": "Associate Professor" }, { "name": "Mohammadreza Teymoorianfard", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" }, { "name": "Sajjad Amini", "gender": "Male", "institution": "Sharif University of Technology", "country": "IR", "position": "Assistant Professor" }, { "name": "Shiqing Ma", "gender": "unknown", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 9, "n_ref_uni": 27, "n_ref": 48, "n_ref_all": 64, "n_fig": 3, "n_tab": 12, "L_tab": 2778, "n_element_tab": 264, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2068, "n_element_tab_1": 154, "formula_len_all": 631, "formula_len_all_1": 454, "len_all": 156177, "len_all_1": 62050, "len_abs": 936, "len_title": 142, "len_sents": 50167, "len_sents_1": 27344, "n_sents": 335, "n_sents_1": 193, "venue": "ICLR2025", "decision": "Reject", "L_title": 93, "L_abs": 955, "title": "MeanSparse: Post-Training Robustness Enhancement Through Mean-Centered Feature Sparsification", "abs": "We present a simple yet effective method to improve the robustness of both Convolutional and attention-based Neural Networks against adversarial examples by post-processing an adversarially trained model. \nOur technique, MeanSparse, cascades the activation functions of a trained model with novel operators that sparsify mean-centered feature vectors. \nThis is equivalent to reducing feature variations around the mean, and we show that such reduced variations merely affect the model's utility, yet they strongly attenuate the adversarial perturbations and decrease the attacker's success rate.\nOur experiments show that, when applied to the top models in the RobustBench leaderboard, MeanSparse achieves a new robustness record of $75.28$% (from $73.71$%), $44.78$% (from $42.67$%) and $62.12$% (from $59.56$%) on CIFAR-10, CIFAR-100 and ImageNet, respectively, in terms of AutoAttack accuracy. \nCode: https://anonymous.4open.science/r/MeanSparse-84B0/", "keywords": [ "Adversarial Training", "Sparsification", "Robustness", "Activation Functions", "Proximal Operator" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 3, 5, 3 ] }, { "paper_id": "CSAfU7J8Gw", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Gillian Dobbie", "gender": "Female", "institution": "University of Auckland", "country": "NZ", "position": "Full Professor" }, { "name": "Wern Sen Wong", "gender": "unknown", "institution": "University of Auckland", "country": "", "position": "PhD student" }, { "name": "Yun Sing Koh", "gender": "unknown", "institution": "University of Auckland", "country": "NZ", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 6, "n_ref_uni": 40, "n_ref": 86, "n_ref_all": 105, "n_fig": 6, "n_tab": 8, "L_tab": 3244, "n_element_tab": 260, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1824, "n_element_tab_1": 169, "formula_len_all": 585, "formula_len_all_1": 374, "len_all": 145619, "len_all_1": 74247, "len_abs": 1728, "len_title": 132, "len_sents": 47968, "len_sents_1": 33415, "n_sents": 329, "n_sents_1": 247, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1269, "title": "SATCH: Specialized Assistant Teacher Distillation to Reduce Catastrophic Forgetting", "abs": "Continual learning enables models to learn new tasks sequentially without forgetting previously learned knowledge. Knowledge distillation reduces forgetting by using a single teacher model to transfer previous knowledge to the student model. However, existing methods face challenges, specifically loss of task-specific knowledge, limited diversity in the transferred knowledge, and delays in teacher availability. These issues stem from self-distillation, where the teacher is a mere snapshot of the student after learning a new task, inheriting the student\u2019s biases and becoming available only after learning a task. We propose Specialized Assistant TeaCHer distillation (SATCH), a novel method that uses a smaller assistant teacher trained exclusively on the current task. By incorporating the assistant teacher early in the learning process, SATCH provides task-specific guidance, improves the diversity of transferred knowledge, and preserves critical task-specific insights. Our method integrates seamlessly with existing knowledge distillation techniques, and experiments on three standard continual learning benchmarks show that SATCH improves accuracy by up to 12% when combined with four state-of-the-art methods. Code is available in supplementary materials.", "keywords": [ "Continual Learning", "Catastrophic Forgetting", "Knowledge Distillation", "Class Incremental Learning" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 1, 3, 2, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "CS2JWaziYr", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Avner May", "gender": "Male", "institution": "Together.ai", "country": "US", "position": "Researcher" }, { "name": "Beidi Chen", "gender": "Female", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "Assistant Professor" }, { "name": "Ian En-Hsu Yen", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Jian Chen", "gender": "Male", "institution": "CMU, Carnegie Mellon University", "country": "", "position": "MS student" }, { "name": "Jinyuan Shi", "gender": "unknown", "institution": "Moffett AI", "country": "CN", "position": "Soft Enginerr" }, { "name": "Ranajoy Sadhukhan", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "PhD student" }, { "name": "Ruihang Lai", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "", "position": "PhD student" }, { "name": "Tianqi Chen", "gender": "unknown", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Vashisth Tiwari", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Researcher" }, { "name": "Zhuoming Chen", "gender": "Male", "institution": "Carnegie Mellon University", "country": "", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 4, "n_ref_uni": 32, "n_ref": 79, "n_ref_all": 115, "n_fig": 19, "n_tab": 6, "L_tab": 2483, "n_element_tab": 483, "n_fig_1": 17, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 267, "formula_len_all_1": 268, "len_all": 140506, "len_all_1": 63670, "len_abs": 1628, "len_title": 104, "len_sents": 37282, "len_sents_1": 32249, "n_sents": 297, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 104, "L_abs": 1454, "title": "MagicDec: Breaking the Latency-Throughput Tradeoff for Long Context Generation with Speculative Decoding", "abs": "Large Language Models (LLMs) have become more prevalent in long-context applications such as interactive chatbots, document analysis, and agent workflows, but it is challenging to serve long-context requests with low latency and high throughput. Speculative decoding (SD) is a widely used technique to reduce latency losslessly, but the conventional wisdom suggests that its efficacy is limited to small batch sizes. In MagicDec, we show that surprisingly SD can achieve speedup even for a high throughput inference regime for moderate to long sequences. More interestingly, an intelligent drafting strategy can achieve better speedup with increasing batch size based on our rigorous analysis. MagicDec first identifies the bottleneck shifts with increasing batch size and sequence length, and uses these insights to deploy SD more effectively for high throughput inference. We leverage draft model with sparse KV cache to address the KV bottleneck, which scales with both sequence length and batch size. Additionally, we propose a theoretical model to select the optimal drafting strategy for maximum speedup. Our work highlights the broad applicability of speculative decoding in long-context serving, as it can enhance throughput and reduce latency without compromising accuracy. For moderate to long sequences, we demonstrate up to 2.51x speedup for LLaMA-3.1-8B when serving batch sizes ranging from 32 to 256 on various types of hardware and tasks.", "keywords": [ "LLM Inference", "Speculative Decoding", "Performance Analysis" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 4 ], "contribution_list": [ 2, 4, 2, 3 ], "confidence_list": [ 2, 4, 3, 3 ] }, { "paper_id": "CRmiX0v16e", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Angela Dai", "gender": "Female", "institution": "Technical University of Munich", "country": "DE", "position": "Associate Professor" }, { "name": "Hisham Cholakkal", "gender": "Male", "institution": "MBZUAI", "country": "AE", "position": "Assistant Professor" }, { "name": "Jean Lahoud", "gender": "unknown", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Postdoc" }, { "name": "Mohamed El Amine Boudjoghra", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Rao Muhammad Anwer", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Assistant Professor" }, { "name": "Salman Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Associate Professor" }, { "name": "Fahad Shahbaz Khan", "gender": "Male", "institution": "Mohamed bin Zayed University of Artificial Intelligence", "country": "AE", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 12, "n_ref_uni": 5, "n_ref": 22, "n_ref_all": 38, "n_fig": 4, "n_tab": 10, "L_tab": 7267, "n_element_tab": 372, "n_fig_1": 4, "n_tab_1": 7, "L_tab_1": 5007, "n_element_tab_1": 357, "formula_len_all": 620, "formula_len_all_1": 628, "len_all": 133371, "len_all_1": 60110, "len_abs": 1844, "len_title": 128, "len_sents": 28906, "len_sents_1": 23267, "n_sents": 172, "n_sents_1": 150, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 80, "L_abs": 1870, "title": "Open-YOLO 3D: Towards Fast and Accurate Open-Vocabulary 3D Instance Segmentation", "abs": "Recent works on open-vocabulary 3D instance segmentation show strong promise but at the cost of slow inference speed and high computation requirements. This high computation cost is typically due to their heavy reliance on aggregated clip features from multi-view, which require computationally expensive 2D foundation models like Segment Anything (SAM) and CLIP. Consequently, this hampers their applicability in many real-world applications that require both fast and accurate predictions. To this end, we propose a novel open-vocabulary 3D instance segmentation approach, named Open-YOLO 3D, that efficiently leverages only 2D object detection from multi-view RGB images for open-vocabulary 3D instance segmentation. \n We demonstrate that our proposed Multi-View Prompt Distribution (MVPDist) method makes use of multi-view information to account for misclassification from the object detector to predict a reliable label for 3D instance masks. Furthermore, since projections of 3D object instances are already contained within the 2D bounding boxes, we show that our proposed low granularity label maps, which require only a 2D object detector to construct, are sufficient and very fast to predict prompt IDs for 3D instance masks when used with our proposed MVPDist.\n We validate our Open-YOLO 3D on two benchmarks, ScanNet200 and Replica, \n under two scenarios: (i) with ground truth masks, where labels are required for given object proposals, and (ii) with class-agnostic 3D proposals generated from a 3D proposal network.\n Our Open-YOLO 3D achieves state-of-the-art performance on both datasets while obtaining up to $\\sim$16$\\times$ speedup compared to the best existing method in literature. On ScanNet200 val. set, our Open-YOLO 3D achieves mean average precision (mAP) of 24.7% while operating at 22 seconds per scene. Our code will be publically available.", "keywords": [ "Open Vocabulary", "3D point cloud instance segmentation" ], "rating_list": [ 10, 8, 8, 5, 8 ], "soundness_list": [ 3, 3, 4, 2, 3 ], "presentation_list": [ 4, 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 5, 4, 3, 4, 3 ] }, { "paper_id": "CRkoMdDlFh", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 7, "author_info_list": [ { "name": "Dazhong Shen", "gender": "Male", "institution": "Nanjing University of Aeronautics and Astronautics", "country": "CN", "position": "Assistant Professor" }, { "name": "Guoqing Zhao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Shaopeng Zhai", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "TianyiZhang", "gender": "Male", "institution": "University College Dublin", "country": "IE", "position": "MS student" }, { "name": "Tong Xu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "QI Zhang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" } ], "n_formula": 10, "n_formula_1": 10, "n_ref_uni": 28, "n_ref": 46, "n_ref_all": 52, "n_fig": 2, "n_tab": 6, "L_tab": 1771, "n_element_tab": 279, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1800, "n_element_tab_1": 194, "formula_len_all": 621, "formula_len_all_1": 621, "len_all": 115375, "len_all_1": 59590, "len_abs": 1482, "len_title": 132, "len_sents": 32841, "len_sents_1": 27127, "n_sents": 276, "n_sents_1": 192, "venue": "ICLR2025", "decision": "withdraw", "L_title": 84, "L_abs": 1491, "title": "I-Lora: Iterative Merging of Routing-Tuned Low-Rank Adapters for Multi-task Learning", "abs": "The advancement of vision-language models has significantly boosted the performance of embodied and game AI, endowing them with more robust general visual understanding capabilities and logical abilities for action planning. However, the substantial computational cost of model training and the performance degradation during fine-tuning limit the models' ability to learn emerging new tasks continually. Creating a versatile and dynamically updatable vision-language model is an essential area of research. To this end, we propose a Low-Rank Adapter-based fine-tuning approach called I-LoRA, which enables iterative and independent learning of new tasks while preserving the logical capabilities of the previously trained model. Specifically, we first design the routing-tuning method to minimize the impact of original capabilities from the new task by minimizing activation values of LoRA matrices as low as possible in the general task. Secondly, we propose a novel approach to iteratively merge new adapters, allowing for continuous integration of adapters trained on new tasks without being influenced by task order, thereby reducing interference between them. Finally, we conducted extensive experiments on public datasets with significant behavioral and logical differences between tasks. The results demonstrate that our approach achieves excellent single-task performance, strong multi-task compatibility, and flexible scalability without increasing the number of model parameters.", "keywords": [ "Multitask learning", "Low-rank adaption", "Vision-language-models" ], "rating_list": [ 5, 3, 3, 5 ], "soundness_list": [ 2, 1, 2, 2 ], "presentation_list": [ 3, 2, 1, 3 ], "contribution_list": [ 2, 2, 1, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "CRQ8JuQDEd", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Beomseok Kwon", "gender": "Male", "institution": "NAVER CLOUD", "country": "", "position": "Researcher" }, { "name": "Byeongwook Kim", "gender": "unknown", "institution": "NAVER CLOUD", "country": "", "position": "Researcher" }, { "name": "Dongsoo Lee", "gender": "Male", "institution": "NAVER CLOVA", "country": "", "position": "Executive Officer" }, { "name": "Eunho Yang", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "Associate Professor" }, { "name": "Gunho Park", "gender": "Male", "institution": "NAVER", "country": "KR", "position": "Researcher" }, { "name": "Jeongin Bae", "gender": "Female", "institution": "NAVER", "country": "", "position": "Researcher" }, { "name": "June Yong Yang", "gender": "unknown", "institution": "Korea Advanced Institute of Science and Technology", "country": "KR", "position": "PhD student" }, { "name": "Se Jung Kwon", "gender": "Male", "institution": "NAVER Cloud", "country": "KR", "position": "AI Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 29, "n_ref": 90, "n_ref_all": 124, "n_fig": 18, "n_tab": 7, "L_tab": 1617, "n_element_tab": 156, "n_fig_1": 12, "n_tab_1": 6, "L_tab_1": 1424, "n_element_tab_1": 209, "formula_len_all": 110, "formula_len_all_1": 102, "len_all": 266999, "len_all_1": 62561, "len_abs": 1560, "len_title": 162, "len_sents": 57721, "len_sents_1": 28809, "n_sents": 443, "n_sents_1": 198, "venue": "ICLR2025", "decision": "Reject", "L_title": 114, "L_abs": 1567, "title": "Don\u2019t Discard, but Keep It Small: Context-Preserving KV Cache Compression with Importance-Aware Adaptive Precision", "abs": "As the length of input sequences in Large Language Models (LLMs) continues to grow, efficient key-value (KV) cache management has become essential for improving inference speed and throughput of autoregressive decoding.\nAlthough several approaches have been proposed to reduce memory usage by selectively retaining only the important KV pairs and discarding the rest, these eviction-based methods can lead to unintended consequences during the generation process.\nIn this paper, we investigate the adverse effects of cache eviction methods and reveal that discarding KV pairs potentially introduces risks such as safety prompt breaches, hallucinations, and loss of critical contextual information.\nInterestingly, we find that preserving even a fraction of the information from evicted KV pairs through reduced precision quantization significantly mitigates these issues.\nOn the other hand, we also observe that important KV pairs need to be maintained at higher precision to preserve generation quality.\nBased on these findings, we propose Mixed-precision KV cache (MiKV), a robust plug-and-play cache compression method that balances performance and memory efficiency.\nMiKV preserves lost contextual information by storing evicted KV pairs in low precision, while maintaining the essential KV pairs in higher precision to ensure generation quality. \nExperimental results across multiple benchmarks and LLM architectures demonstrate that our method achieves a state-of-the-art balance between compression ratio and model performance, outperforming existing baselines.", "keywords": [ "large language models", "safety", "hallucination", "key-value cache compression", "long context" ], "rating_list": [ 6, 5, 3, 6, 5 ], "soundness_list": [ 3, 3, 2, 3, 2 ], "presentation_list": [ 2, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 1, 3, 3 ], "confidence_list": [ 4, 3, 3, 4, 3 ] }, { "paper_id": "CPhqrV5Ehg", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Antske Fokkens", "gender": "Female", "institution": "VU University Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "Sergey Troshin", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Vlad Niculae", "gender": "Male", "institution": "University of Amsterdam", "country": "", "position": "Assistant Professor" } ], "n_formula": 23, "n_formula_1": 17, "n_ref_uni": 40, "n_ref": 97, "n_ref_all": 139, "n_fig": 17, "n_tab": 8, "L_tab": 3989, "n_element_tab": 889, "n_fig_1": 7, "n_tab_1": 4, "L_tab_1": 839, "n_element_tab_1": 1, "formula_len_all": 1204, "formula_len_all_1": 632, "len_all": 171464, "len_all_1": 58204, "len_abs": 1010, "len_title": 131, "len_sents": 44421, "len_sents_1": 27634, "n_sents": 327, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1014, "title": "Efficient Controlled Language Generation with Low-Rank Autoregressive Reward Models", "abs": "Language models trained on large amounts of data are known to produce inappropriate content in some cases and require careful tuning to be used in the real world. We revisit the reward augmented decoding (RAD) approach to control the generation from a language model using the scores from a task-specific reward model. We investigate the training objective of RAD, and reformulate it as a task of learning a reward matrix. We show that RAD is designed to support high flexibility when representing the reward matrices, which leads to higher computational costs during decoding. However, we demonstrate that RAD does not use its full flexibility. Motivated by this, we propose a simpler but more efficient low-rank parametrization of the reward model enabling fast and effective guided decoding. For the detoxification and sentiment control tasks, we show that our low-rank reward model performs on par with the more flexible RAD parametrization, while requiring only a single reward model call per generated token.", "keywords": [ "Controlled text generation", "LLM", "Natural Language Processing", "Reward modelling", "Efficiency" ], "rating_list": [ 8, 3, 5, 6, 8 ], "soundness_list": [ 4, 3, 2, 3, 3 ], "presentation_list": [ 4, 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 2, 3 ] }, { "paper_id": "CPBdBmnkA5", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Brandon Reagen", "gender": "Male", "institution": "LG Corporation", "country": "US", "position": "Researcher" }, { "name": "Nandan Kumar Jha", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 34, "n_formula_1": 7, "n_ref_uni": 52, "n_ref": 149, "n_ref_all": 203, "n_fig": 23, "n_tab": 13, "L_tab": 5817, "n_element_tab": 408, "n_fig_1": 19, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2575, "formula_len_all_1": 975, "len_all": 292704, "len_all_1": 62240, "len_abs": 1091, "len_title": 103, "len_sents": 82768, "len_sents_1": 29529, "n_sents": 525, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 55, "L_abs": 1119, "title": "AERO: Softmax-Only LLMs for Efficient Private Inference", "abs": "The pervasiveness of proprietary language models has raised privacy concerns for users' sensitive data, emphasizing the need for private inference (PI), where inference is performed directly on encrypted inputs. However, current PI methods face prohibitively higher communication and latency overheads, primarily due to nonlinear operations. In this paper, we present a comprehensive analysis to understand the role of nonlinearities in transformer-based decoder-only language models. We introduce AERO, a four-step architectural optimization framework that refines the existing LLM architecture for efficient PI by systematically removing nonlinearities such as LayerNorm and GELU and reducing FLOPs counts. For the {\\em first time}, we propose a Softmax-only architecture with significantly fewer FLOPs tailored for efficient PI. Furthermore, we devise a novel entropy regularization technique to improve the performance of Softmax-only models. AERO achieves up to 4.23$\\times$ communication and 1.94$\\times$ latency reduction. We validate the effectiveness of AERO by benchmarking it against the state-of-the-art.", "keywords": [ "Private inference", "LLMs", "Architectural Optimization", "Entropy Regularization" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 4, 2, 3 ], "contribution_list": [ 3, 2, 3 ], "confidence_list": [ 3, 4, 5 ] }, { "paper_id": "CP6CAqxAGJ", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Chengqing Zong", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Researcher" }, { "name": "Chong Li", "gender": "Male", "institution": "Institute of automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Jiajun Zhang", "gender": "Male", "institution": "Institute of automation, Chinese academy of science, Chinese Academy of Sciences", "country": "", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 50, "n_ref": 83, "n_ref_all": 107, "n_fig": 10, "n_tab": 5, "L_tab": 3752, "n_element_tab": 255, "n_fig_1": 8, "n_tab_1": 5, "L_tab_1": 3018, "n_element_tab_1": 303, "formula_len_all": 68, "formula_len_all_1": 68, "len_all": 202418, "len_all_1": 60720, "len_abs": 1074, "len_title": 130, "len_sents": 29198, "len_sents_1": 25188, "n_sents": 210, "n_sents_1": 183, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1084, "title": "Unifying Vocabulary of Large Language Model with Statistical Token-level Alignment", "abs": "Large Language Models (LLMs) achieve great success across many general tasks, but the mismatch among different vocabularies hinders further applications like token-level distillation and inference with various models. To align the vocabularies of LLMs, we propose a simple yet effective method named **UnifyVocab** to replace the vocabulary of an LLM at a limited cost. A new vocabulary alignment method is devised first to align the source vocabulary to the target one. We then rearrange the corresponding parameters like embeddings, and progressively fine-tune the model. Experimental results on models across multiple parameter scales demonstrate the effectiveness and generalization of UnifyVocab, which costs as few as 10B tokens to recover 98.02\\% performance of the vanilla models on average. We further find that unifying the vocabularies significantly facilitates the token-level distillation which remarkably boosts (+4.4\\%) the model with only 235M tokens. Moreover, our method provides a better initialization of multilingual vocabulary for LLMs to adapt to new languages.", "keywords": [ "Vocabulary Adaptation", "Large Language Model", "Efficient NLP" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 2, 3, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "COdUNtjMEp", "primary_area": "learning theory", "n_author": 4, "author_info_list": [ { "name": "Cong Shen", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Associate Professor" }, { "name": "Jing Yang", "gender": "unknown", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Associate Professor" }, { "name": "Ruida Zhou", "gender": "Male", "institution": "University of California, Los Angeles", "country": "US", "position": "Postdoc" }, { "name": "Wei Shen", "gender": "Male", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "PhD student" } ], "n_formula": 202, "n_formula_1": 40, "n_ref_uni": 20, "n_ref": 46, "n_ref_all": 97, "n_fig": 8, "n_tab": 3, "L_tab": 1067, "n_element_tab": 17, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 97, "n_element_tab_1": 1, "formula_len_all": 29474, "formula_len_all_1": 3060, "len_all": 245003, "len_all_1": 67008, "len_abs": 900, "len_title": 122, "len_sents": 65604, "len_sents_1": 26666, "n_sents": 684, "n_sents_1": 230, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 903, "title": "On the Training Convergence of Transformers for In-Context Classification", "abs": "While transformers have demonstrated impressive capacities for in-context learning (ICL) in practice, theoretical understanding of the underlying mechanism enabling transformers to perform ICL is still in its infant stage. This work aims to theoretically study the training dynamics of transformers for in-context classification tasks. We demonstrate that, for in-context classification of Gaussian mixtures under certain assumptions, a single-layer transformer trained via gradient descent converges to a globally optimal model at a linear rate. We further quantify the impact of the training and testing prompt lengths on the ICL inference error of the trained transformer. We show that when the lengths of training and testing prompts are sufficiently large, the prediction of the trained transformer approaches the Bayes-optimal classifier. Experimental results corroborate the theoretical findings.", "keywords": [ "In-context learning", "Transformer" ], "rating_list": [ 5, 6, 5, 8 ], "soundness_list": [ 3, 3, 3, 4 ], "presentation_list": [ 2, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "CO4wKfSyhb", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Bill Marino", "gender": "Male", "institution": "Cornell University", "country": "US", "position": "Lecturer" }, { "name": "Lorenzo Sani", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Preslav Aleksandrov", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "William F. Shen", "gender": "unknown", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Alexandru Andrei Iacob", "gender": "Male", "institution": "University of Cambridge", "country": "", "position": "PhD student" }, { "name": "Nic Lane", "gender": "Male", "institution": "Flower Labs", "country": "GB", "position": "Chief Scientific Officer" } ], "n_formula": 13, "n_formula_1": 5, "n_ref_uni": 47, "n_ref": 89, "n_ref_all": 114, "n_fig": 5, "n_tab": 6, "L_tab": 1300, "n_element_tab": 147, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 524, "n_element_tab_1": 70, "formula_len_all": 2351, "formula_len_all_1": 1217, "len_all": 196488, "len_all_1": 57076, "len_abs": 1340, "len_title": 47, "len_sents": 50804, "len_sents_1": 27332, "n_sents": 340, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 47, "L_abs": 1369, "title": "Worldwide Federated Training of Language Models", "abs": "Language Model (LM) training requires vast datasets, raising legal, ethical, and practical concerns. Federated learning (FL) offers an alternative by enabling organizations to collaboratively leverage untapped reserves while minimizing data movement. However, scaling FL globally introduces challenges such as restrictions on data movement, privacy, and statistical data heterogeneity. We propose Worldwide Federated Language Model Training (WorldLM), a system that builds federations of federations. WorldLM enables each federation to autonomously meet jurisdictional or competitive constraints while managing statistical heterogeneity through attention-based aggregation of key layers and cross-federation information sharing via residual embeddings. In terms of perplexity, WorldLM outperforms standard FL and other federated baselines by up to $1.91\\times$ and $3.3\\times$ respectively. WorldLM scales to models with $400$M parameters, achieving $1.39\\times$ lower perplexity than centralized counterparts while approaching the performance of perfectly localized models trained in an infinite-data regime. Additionally, under differential privacy constraints, WorldLM proves highly resilient in performance compared to standard FL methods, which diverge. These results establish WorldLM as an effective means for pre-training across geographic and legal boundaries.", "keywords": [ "Federated Learning", "Distributed Training", "Language Modeling", "Natural Language Processing", "Hierarchical Federated Learning", "Governance" ], "rating_list": [ 5, 5, 3, 3, 3 ], "soundness_list": [ 2, 3, 2, 2, 1 ], "presentation_list": [ 3, 2, 2, 2, 2 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 3, 2, 4, 4, 2 ] }, { "paper_id": "CNPLXcMcSP", "primary_area": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)", "n_author": 2, "author_info_list": [ { "name": "Samson Zhou", "gender": "unknown", "institution": "Texas A&M University - College Station", "country": "US", "position": "Assistant Professor" }, { "name": "Zhao Song", "gender": "Male", "institution": "University of California, Berkeley", "country": "US", "position": "Researcher" } ], "n_formula": 47, "n_formula_1": 18, "n_ref_uni": 69, "n_ref": 129, "n_ref_all": 130, "n_fig": 0, "n_tab": 1, "L_tab": 7, "n_element_tab": 1, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4681, "formula_len_all_1": 1293, "len_all": 173973, "len_all_1": 72206, "len_abs": 712, "len_title": 100, "len_sents": 35097, "len_sents_1": 29497, "n_sents": 327, "n_sents_1": 269, "venue": "ICLR2025", "decision": "Reject", "L_title": 52, "L_abs": 714, "title": "Towards Sampling Data Structures for Tensor Products", "abs": "This paper studies the computational challenges of attention-based models in artificial intelligence by introducing innovative sampling methods to accelerate attention computation in large language models (LLM). Inspired by the recent progress of LLM in real-life applications, we introduces a streaming sampler question for attention setting. Our approach significantly reduces the computational burden of traditional attention mechanisms while maintaining or enhancing model performance. We demonstrate these methods' effectiveness from theoretical perspective, including space, update time. Additionally, our framework exhibits scalability and broad applicability across various model architectures and domains.", "keywords": [ "sampling", "data structures", "tensor products" ], "rating_list": [ 5, 3, 3, 3 ], "soundness_list": [ 3, 2, 1, 1 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "CNO4rbSV6v", "primary_area": "foundation or frontier models, including LLMs", "n_author": 5, "author_info_list": [ { "name": "Congyue Deng", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Leonidas Guibas", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Yang You", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Yixin Li", "gender": "unknown", "institution": "Stanford University", "country": "US", "position": "Undergrad student" }, { "name": "Yue Wang", "gender": "Male", "institution": "University of Southern California", "country": "US", "position": "Assistant Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 29, "n_ref": 54, "n_ref_all": 77, "n_fig": 11, "n_tab": 4, "L_tab": 1225, "n_element_tab": 134, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 182, "n_element_tab_1": 38, "formula_len_all": 184, "formula_len_all_1": 132, "len_all": 119178, "len_all_1": 53249, "len_abs": 426, "len_title": 145, "len_sents": 29455, "len_sents_1": 25418, "n_sents": 207, "n_sents_1": 177, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 95, "L_abs": 1038, "title": "Multiview Equivariance Improves 3D Correspondence Understanding with Minimal Feature Finetuning", "abs": "Vision foundation models, particularly the ViT family, have revolutionized image understanding by providing rich semantic features. However, despite their success in 2D comprehension, their abilities on grasping 3D spatial relationships are still unclear.\nIn this work, we evaluate and enhance the 3D awareness of ViT-based models. We begin by systematically assessing their ability to learn 3D equivariant features, specifically examining the consistency of semantic embeddings across different viewpoints. Our findings indicate that improved 3D equivariance leads to better performance on various downstream tasks, including pose estimation, tracking, and semantic transfer. Building on this insight, we propose a simple yet effective finetuning strategy based on 3D correspondences, which significantly enhances the 3D understanding of existing vision models. Remarkably, even finetuning on a single object for just one iteration results in substantial performance gains. Code is available on https://github.com/qq456cvb/3DCorrEnhance.", "keywords": [ "Vision Foundation Models; 3D Representation Learning; Fine-tuning; 3D Equivariance" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "CNGkrfDhdG", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Guanzhong Wu", "gender": "Male", "institution": "Northwestern Polytechnical University, Northwest Polytechnical University Xi'an", "country": "CN", "position": "MS student" }, { "name": "Qing Li", "gender": "Female", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "Associate Professor" }, { "name": "kaiwen wei", "gender": "Male", "institution": "Chongqing University", "country": "CN", "position": "Postdoc" } ], "n_formula": 27, "n_formula_1": 16, "n_ref_uni": 19, "n_ref": 52, "n_ref_all": 82, "n_fig": 8, "n_tab": 7, "L_tab": 3665, "n_element_tab": 459, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 40, "n_element_tab_1": 2, "formula_len_all": 2692, "formula_len_all_1": 909, "len_all": 154709, "len_all_1": 59258, "len_abs": 1633, "len_title": 167, "len_sents": 46938, "len_sents_1": 27751, "n_sents": 354, "n_sents_1": 210, "venue": "ICLR2025", "decision": "Reject", "L_title": 115, "L_abs": 1629, "title": "Integrating Relation Dependences and Textual Semantics for Coherent Logical Reasoning over Temporal Knowledge Graph", "abs": "Temporal knowledge graphs (TKGs) reflect the evolution patterns of facts, which can be summarized as logical rules and applied to forecast future facts. However, existing logical reasoning methods on TKGs face two limitations: 1) A lack of efficient strategies for extracting logical paths. 2) Insufficient utilization of structural and textual information. To bridge these gaps, we propose CoLR, a two-stage framework that mines relation dependencies and textual semantics for Coherent Logical Reasoning over TKGs. In the first stage, we construct a temporal relation structure graph (TRSG) composed of relations and cohesion weights between them. Besides, we define a novel time-fusion search graph (TFSG) along with TRSG to facilitate efficient and reliable temporal path searching. In the second stage,\nthe textual content and timestamp sequences from these paths undergo encoding via a pre-trained language model and a time sequence encoder to accurately capture potential logical rules. Additionally, for quadruplets missing paths, historical edges sampled based on relation cohesion are used as supplements. Given the limitations of existing benchmark datasets in evaluating accuracy, generalization, and robustness, we construct three new datasets tailored to transductive, inductive, and few-shot scenarios, respectively. These datasets, combined with four real-world datasets, are employed to evaluate our model comprehensively. Experimental results demonstrate that our approach significantly outperforms existing methods across all three scenarios. Our code is available at https://anonymous.4open.science/r/CoLR-0839", "keywords": [ "Temporal knowledge graph", "Knowledge graph", "Multi-hop logical rules", "Link forecasting", "Inductive reasoning" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 2 ] }, { "paper_id": "CN328Aw03P", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 11, "author_info_list": [ { "name": "Anirban Chandra", "gender": "Male", "institution": "Shell", "country": "US", "position": "Researcher" }, { "name": "Campbell D Watson", "gender": "Male", "institution": "International Business Machines", "country": "US", "position": "Research Scientist" }, { "name": "Christopher Hill", "gender": "Male", "institution": "Imperial College, London University", "country": "", "position": "Undergrad student" }, { "name": "Daniel Salles Civitarese", "gender": "Male", "institution": "International Business Machines", "country": "BR", "position": "Researcher" }, { "name": "Detlef Hohl", "gender": "unknown", "institution": "Rice University", "country": "US", "position": "Principal Researcher" }, { "name": "Eric Schmitt", "gender": "Male", "institution": "Shell", "country": "US", "position": "Researcher" }, { "name": "Jeremy Vila", "gender": "Male", "institution": "Shell", "country": "US", "position": "Researcher" }, { "name": "Johannes Jakubik", "gender": "Male", "institution": "IBM Research Europe", "country": "", "position": "Intern" }, { "name": "Jonathan Giezendanner", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Qidong Yang", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Sherrie Wang", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 13, "n_ref_uni": 27, "n_ref": 40, "n_ref_all": 62, "n_fig": 9, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 546, "formula_len_all_1": 495, "len_all": 112946, "len_all_1": 62466, "len_abs": 1268, "len_title": 124, "len_sents": 32649, "len_sents_1": 29814, "n_sents": 239, "n_sents_1": 218, "venue": "ICLR2025", "decision": "Reject", "L_title": 76, "L_abs": 1280, "title": "Multi-modal graph neural networks for localized off-grid weather forecasting", "abs": "Urgent applications like wildfire management and renewable energy generation require precise, localized weather forecasts near the Earth's surface.\nHowever, weather forecast products from machine learning or numerical weather models are currently generated on a global regular grid, on which a naive interpolation cannot accurately reflect fine-grained weather patterns close to the ground.\nIn this work, we train a heterogeneous graph neural network (GNN) end-to-end to downscale gridded forecasts to off-grid locations of interest.\nThis multi-modal GNN takes advantage of local historical weather observations (e.g., wind vector, temperature) to correct the gridded weather forecast at different lead times towards locally accurate forecasts.\nEach data modality is modeled as a different type of node in the graph.\nUsing message passing, the node at the prediction location aggregates information from its heterogeneous neighbor nodes.\nExperiments using weather stations across the Northeastern United States show that our model outperforms a range of data-driven and non-data-driven off-grid forecasting methods.\nOur approach demonstrates how the gap between global large-scale weather models and locally accurate predictions can be bridged to inform localized decision-making.", "keywords": [ "Weather forecasting", "Graph Neural Network", "Multi-modal", "off-grid weather forecasting", "heterogeneous graph neural network", "climate", "climate change", "sustainability" ], "rating_list": [ 5, 3, 5, 3, 5 ], "soundness_list": [ 3, 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 3, 3, 3 ], "contribution_list": [ 2, 1, 2, 2, 2 ], "confidence_list": [ 4, 4, 4, 4, 5 ] }, { "paper_id": "CN2bmVVpOh", "primary_area": "applications to neuroscience & cognitive science", "n_author": 5, "author_info_list": [ { "name": "Aaron Traylor", "gender": "Male", "institution": "Brown University", "country": "", "position": "PhD student" }, { "name": "Aneri Soni", "gender": "Female", "institution": "Brown University", "country": "US", "position": "PhD student" }, { "name": "Ellie Pavlick", "gender": "Female", "institution": "Brown University", "country": "", "position": "Associate Professor" }, { "name": "Michael Frank", "gender": "Male", "institution": "Brown University", "country": "", "position": "Full Professor" }, { "name": "jack merullo", "gender": "Male", "institution": "Brown University", "country": "US", "position": "PhD student" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 33, "n_ref": 70, "n_ref_all": 86, "n_fig": 8, "n_tab": 1, "L_tab": 212, "n_element_tab": 16, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 549, "n_element_tab_1": 25, "formula_len_all": 87, "formula_len_all_1": 87, "len_all": 121120, "len_all_1": 66731, "len_abs": 1742, "len_title": 154, "len_sents": 47584, "len_sents_1": 34811, "n_sents": 303, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 104, "L_abs": 1760, "title": "Transformer Mechanisms Mimic Frontostriatal Gating Operations When Trained on Human Working Memory Tasks", "abs": "The Transformer neural network architecture has seen success on a wide variety of tasks that appear to require executive function - the ability to represent, coordinate, and manage multiple subtasks. In cognitive neuroscience, executive function is thought to rely on sophisticated frontostriatal mechanisms for selective gating, which enable role-addressable updating-- and later readout-- of information to and from distinct \"addresses\" of memory, in the form of clusters of neurons. However, Transformer models have no such mechanisms intentionally built-in. It is thus an open question how Transformers solve such tasks, and whether the mechanisms that emerge to help them to do so resemble the gating mechanisms in the human brain. In this work, we analyze the mechanisms that emerge within a vanilla attention-only Transformer when trained on a task from computational cognitive neuroscience explicitly designed to place demands on working memory gating. We find that the self-attention mechanism within the Transformer develops input and output gating mechanisms, particularly when task demands require them. These gating mechanisms mirror those incorporated into earlier biologically-inspired architectures and mimic those in human studies. When learned effectively, these gating strategies support enhanced generalization and increase the models' effective capacity to store and access multiple items in memory. Despite not having memory limits, we also find that storing and accessing multiple items requires an efficient gating policy, resembling the constraints found in frontostriatal models. \nThese results suggest opportunities for future research on computational similarities between modern AI architectures and models of the human brain.", "keywords": [ "transformers; neural networks; working memory; computational neuroscience; gating; computational cognitive science; mechanistic interpretability" ], "rating_list": [ 5, 3, 5 ], "soundness_list": [ 3, 1, 3 ], "presentation_list": [ 1, 2, 3 ], "contribution_list": [ 1, 2, 2 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "CMqOfvD3tO", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 3, "author_info_list": [ { "name": "Dong Un Kang", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" }, { "name": "Hayeon Kim", "gender": "Female", "institution": "Seoul National University", "country": "KR", "position": "MS student" }, { "name": "Se Young Chun", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 53, "n_ref": 156, "n_ref_all": 189, "n_fig": 6, "n_tab": 13, "L_tab": 3237, "n_element_tab": 397, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2561, "n_element_tab_1": 189, "formula_len_all": 453, "formula_len_all_1": 383, "len_all": 184685, "len_all_1": 63992, "len_abs": 1419, "len_title": 131, "len_sents": 52479, "len_sents_1": 30253, "n_sents": 358, "n_sents_1": 168, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1424, "title": "Class Distribution-induced Attention Map for Open-vocabulary Semantic Segmentations", "abs": "Open-vocabulary semantic segmentation is a challenging task that assigns seen or unseen class labels to individual pixels. While recent works with vision-language models (VLMs) have shown promising results in zero-shot semantic segmentation, they still struggle to accurately localize class-related objects. In this work, we argue that CLIP-based prior works yield patch-wise noisy class predictions while having highly correlated class distributions for each object. Then, we propose Class Distribution-induced Attention Map, dubbed CDAM, that is generated by the Jensen-Shannon divergence between class distributions of two patches that belong to the same (class) object. This CDAM can be used for open-vocabulary semantic segmentation by integrating it into the final layer of CLIP to enhance the capability to accurately localize desired classes. Our class distribution-induced attention scheme can easily work with multi-scale image patches as well as augmented text prompts for further enhancing attention maps. By exploiting class distribution, we also propose robust entropy-based background thresholding for the inference of semantic segmentation. Interestingly, the core idea of our proposed method does not conflict with other prior arts in zero-shot semantic segmentation, thus can be synergetically used together, yielding substantial improvements in performance across popular semantic segmentation benchmarks.", "keywords": [ "Vision Language Model", "Dense Localization" ], "rating_list": [ 6, 6, 6, 8, 8 ], "soundness_list": [ 3, 3, 2, 3, 3 ], "presentation_list": [ 2, 2, 3, 3, 4 ], "contribution_list": [ 2, 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 4, 3, 4 ] }, { "paper_id": "CMj18BQQDK", "primary_area": "generative models", "n_author": 7, "author_info_list": [ { "name": "Amirmojtaba Sabour", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Despoina Paschalidou", "gender": "Female", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Jiahui Huang", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Cheng Xie", "gender": "Male", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "PhD student" }, { "name": "Sanja Fidler", "gender": "Female", "institution": "Department of Computer Science, University of Toronto", "country": "CA", "position": "Associate Professor" }, { "name": "Umar Iqbal", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Xiaohui Zeng", "gender": "Female", "institution": "NVIDIA", "country": "CA", "position": "Researcher" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 46, "n_ref": 101, "n_ref_all": 116, "n_fig": 10, "n_tab": 3, "L_tab": 475, "n_element_tab": 87, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1092, "n_element_tab_1": 60, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 140013, "len_all_1": 57632, "len_abs": 562, "len_title": 111, "len_sents": 31639, "len_sents_1": 27225, "n_sents": 217, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 63, "L_abs": 1208, "title": "VideoPanda: Video Panoramic Diffusion With Multi-view Attention", "abs": "High resolution panoramic video content is paramount for immersive experiences in Virtual Reality, but is non-trivial to collect as it requires specialized equipment and intricate camera setups. \nIn this work, we introduce \\ourmodel, a novel approach for synthesizing $360^\\circ$ videos conditioned on text or single-view video data. \\ourmodel leverages multi-view attention layers to augment a video diffusion model, enabling it to generate consistent multi-view videos that can be combined into immersive panoramic content. \\ourmodel is trained jointly using two conditions: text-only and single-view video, and supports autoregressive generation of long-videos. \nTo overcome the computational burden of multi-view video generation, we randomly subsample the duration and camera views used during training and show that the model is able to gracefully generalize to generating more frames during inference.\nExtensive evaluations on both real-world and synthetic video datasets demonstrate that \\ourmodel generates more realistic and coherent $360^\\circ$ panoramas across all input conditions compared to existing methods.\nVisit the project website at https://mvpanovideo.github.io/VideoPanda/ for results.", "keywords": [ "video generation", "diffusion model", "panorama" ], "rating_list": [ 5, 6, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 3, 5 ] }, { "paper_id": "CMMpcs9prj", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Rustem Islamov", "gender": "Male", "institution": "University of Basel", "country": "CH", "position": "PhD student" }, { "name": "Sebastian U Stich", "gender": "Male", "institution": "CISPA Helmholtz Center for Information Security", "country": "DE", "position": "Tenure Track Faculty" }, { "name": "Yuan Gao", "gender": "unknown", "institution": "CISPA, saarland university, saarland informatics campus", "country": "", "position": "PhD student" } ], "n_formula": 135, "n_formula_1": 25, "n_ref_uni": 80, "n_ref": 153, "n_ref_all": 182, "n_fig": 9, "n_tab": 3, "L_tab": 536, "n_element_tab": 186, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 315, "n_element_tab_1": 93, "formula_len_all": 16652, "formula_len_all_1": 2660, "len_all": 233712, "len_all_1": 72943, "len_abs": 1397, "len_title": 131, "len_sents": 50306, "len_sents_1": 29911, "n_sents": 526, "n_sents_1": 254, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 83, "L_abs": 1447, "title": "Towards Faster Decentralized Stochastic Optimization with Communication Compression", "abs": "Communication efficiency has garnered significant attention as it is considered the main bottleneck for large-scale decentralized Machine Learning applications in distributed and federated settings. In this regime, clients are restricted to transmitting small amounts of compressed information to their neighbors over a communication graph. Numerous endeavors have been made to address this challenging problem by developing algorithms with compressed communication for decentralized non-convex optimization problems. Despite considerable efforts, current theoretical understandings of the problem are still very limited, and existing algorithms all suffer from various limitations. In particular, these algorithms typically rely on strong, and often infeasible assumptions such as bounded data heterogeneity or require large batch access while failing to achieve linear speedup with the number of clients. In this paper, we introduce MoTEF, a novel approach that integrates communication compression with $\\textbf{Mo}$mentum $\\textbf{T}$racking and $\\textbf{E}$rror $\\textbf{F}$eedback. MoTEF is the first algorithm to achieve an asymptotic rate matching that of distributed SGD under arbitrary data heterogeneity, hence resolving a long-standing theoretical obstacle in decentralized optimization with compressed communication. We provide numerical experiments to validate our theoretical findings and confirm the practical superiority of MoTEF.", "keywords": [ "Optimization", "Decentralized Learning", "Federated Learning", "Communication Compression" ], "rating_list": [ 8, 8, 6, 6, 5 ], "soundness_list": [ 3, 4, 2, 3, 3 ], "presentation_list": [ 3, 4, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "CLVMAUDeJz", "primary_area": "optimization", "n_author": 2, "author_info_list": [ { "name": "Jie Mei", "gender": "unknown", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Xiangzheng Meng", "gender": "unknown", "institution": "Harbin Institute of Technology", "country": "CN", "position": "PhD student" } ], "n_formula": 98, "n_formula_1": 60, "n_ref_uni": 23, "n_ref": 72, "n_ref_all": 115, "n_fig": 1, "n_tab": 2, "L_tab": 616, "n_element_tab": 31, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 18128, "formula_len_all_1": 7829, "len_all": 105012, "len_all_1": 57120, "len_abs": 959, "len_title": 112, "len_sents": 23953, "len_sents_1": 18890, "n_sents": 223, "n_sents_1": 171, "venue": "ICLR2025", "decision": "Reject", "L_title": 66, "L_abs": 985, "title": "Distributed Constrained Optimal Consensus Under a Directed Graph", "abs": "In this paper, the distributed constrained optimal consensus problem of multi-agent systems under a directed graph is investigated. We propose two projection-based distributed constrained optimal consensus algorithms: one addressing set constraints and the other tailored for general constraints. Only the relative state is exchanged among agents in these two algorithms. In the stability analysis of case with set constraints, we transform the distributed optimization problem into a constrained leaderless consensus problem by adopting a sliding mode approach. Building on this foundational transformation, we further develop a projection-based distributed constrained optimal consensus algorithm to address general constraints. It is shown that the proposed algorithm achieves an ergodic convergence rate of $O(\\frac{1}{k})$ with respect to the first-order optimality residuals. Numerical simulations are conducted to validate the effectiveness of our theoretical results.", "keywords": [ "Constrained optimal consensus", "multi-agent systems", "set constraints", "general constraints" ], "rating_list": [ 3, 3, 3, 5 ], "soundness_list": [ 2, 2, 2, 2 ], "presentation_list": [ 1, 1, 2, 3 ], "contribution_list": [ 2, 1, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "CLImhawlGn", "primary_area": "learning on time series and dynamical systems", "n_author": 2, "author_info_list": [ { "name": "Tom Ryder", "gender": "Male", "institution": "University of Bath", "country": "GB", "position": "PhD student" }, { "name": "Xi Chen", "gender": "Male", "institution": "University of Bath", "country": "GB", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 12, "n_ref_uni": 40, "n_ref": 82, "n_ref_all": 98, "n_fig": 5, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1477, "n_element_tab_1": 32, "formula_len_all": 1047, "formula_len_all_1": 1085, "len_all": 116589, "len_all_1": 56886, "len_abs": 1290, "len_title": 155, "len_sents": 34286, "len_sents_1": 25336, "n_sents": 270, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 107, "L_abs": 1295, "title": "Channel Independence Improves Out-of-Distribution Generalisation in Multivariate Time Series Classification", "abs": "Robustness to distribution shift is a necessary property of machine learning models for their safe and effective deployment. However, deep learning models are susceptible to learning spurious features of the in-distribution (ID) training data that fail to generalise to out-of-distribution (OOD) data. Domain generalisation algorithms aim to tackle this problem, but recent studies have demonstrated that their improvement over standard empirical risk minimisation is marginal. We address this problem for multivariate time series classification (TSC), where it is standard practise to use feature extractor architectures that learn with channel dependence (CD), enabling cross-channel patterns to be learned. Inspired by recent success in time series forecasting, we investigate how channel independence (CI) impacts OOD generalisation in TSC. Our experiments on six time series datasets reveal that ID and OOD features exhibit significantly greater distributional divergence when learned with CD compared to CI. As a consequence, models that learn with CI are more robust to distribution shift, evidenced by smaller generalisation gaps (the difference between ID and OOD performance) across datasets. On datasets that have a stronger shift, OOD accuracy is substantially higher for CI than CD.", "keywords": [ "time series classification", "OOD generalization", "domain generalization" ], "rating_list": [ 5, 3, 5, 3 ], "soundness_list": [ 3, 2, 2, 1 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 3, 5, 4, 4 ] }, { "paper_id": "CLE09ESvul", "primary_area": "interpretability and explainable AI", "n_author": 7, "author_info_list": [ { "name": "Abdullah Makkeh", "gender": "unknown", "institution": "University of Goettingen", "country": "DE", "position": "Postdoc" }, { "name": "Alexander S Ecker", "gender": "Male", "institution": "Georg-August Universit\u00e4t G\u00f6ttingen", "country": "DE", "position": "Full Professor" }, { "name": "Andreas Christian Schneider", "gender": "Male", "institution": "Max Planck Institute for Dynamics and Self-Organzation", "country": "", "position": "PhD student" }, { "name": "David Alexander Ehrlich", "gender": "unknown", "institution": "Georg-August Universit\u00e4t G\u00f6ttingen", "country": "", "position": "PhD student" }, { "name": "Michael Wibral", "gender": "unknown", "institution": "Georg August University", "country": "", "position": "Full Professor" }, { "name": "Valentin Neuhaus", "gender": "unknown", "institution": "Georg-August Universit\u00e4t G\u00f6ttingen", "country": "DE", "position": "PhD student" }, { "name": "Viola Priesemann", "gender": "unknown", "institution": "Max-Planck Institute", "country": "", "position": "Assistant Professor" } ], "n_formula": 10, "n_formula_1": 4, "n_ref_uni": 49, "n_ref": 83, "n_ref_all": 113, "n_fig": 10, "n_tab": 3, "L_tab": 994, "n_element_tab": 67, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1785, "formula_len_all_1": 341, "len_all": 150055, "len_all_1": 63260, "len_abs": 1395, "len_title": 178, "len_sents": 54316, "len_sents_1": 35243, "n_sents": 272, "n_sents_1": 175, "venue": "ICLR2025", "decision": "Accept (Oral)", "L_title": 93, "L_abs": 1400, "title": "What should a neuron aim for? Designing local objective functions based on information theory", "abs": "In modern deep neural networks, the learning dynamics of the individual neurons is often obscure, as the networks are trained via global optimization. Conversely, biological systems build on self-organized, local learning, achieving robustness and efficiency with limited global information. We here show how self-organization between individual artificial neurons can be achieved by designing abstract bio-inspired local learning goals. These goals are parameterized using a recent extension of information theory, Partial Information Decomposition (PID), which decomposes the information that a set of information sources holds about an outcome into unique, redundant and synergistic contributions. Our framework enables neurons to locally shape the integration of information from various input classes, i.e. feedforward, feedback, and lateral, by selecting which of the three inputs should contribute uniquely, redundantly or synergistically to the output. This selection is expressed as a weighted sum of PID terms, which, for a given problem, can be directly derived from intuitive reasoning or via numerical optimization, offering a window into understanding task-relevant local information processing. Achieving neuron-level interpretability while enabling strong performance using local learning, our work advances a principled information-theoretic foundation for local learning strategies.", "keywords": [ "local learning", "interpretability", "neuro-inspired", "information theory", "partial information decomposition" ], "rating_list": [ 6, 8, 8, 8 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 4 ], "confidence_list": [ 3, 4, 5, 3 ] }, { "paper_id": "CL3U0GxFRD", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Chenjia Bai", "gender": "Male", "institution": "TeleAI, China Telecom", "country": "CN", "position": "Research Scientist" }, { "name": "Jun Zhang", "gender": "unknown", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Xiaolu Wang", "gender": "Male", "institution": "East China Normal University", "country": "CN", "position": "Associate Professor" }, { "name": "Xinran Li", "gender": "Female", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 12, "n_ref_uni": 50, "n_ref": 92, "n_ref_all": 123, "n_fig": 16, "n_tab": 8, "L_tab": 2618, "n_element_tab": 199, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 1289, "n_element_tab_1": 100, "formula_len_all": 1473, "formula_len_all_1": 755, "len_all": 152486, "len_all_1": 64249, "len_abs": 1451, "len_title": 137, "len_sents": 41479, "len_sents_1": 28219, "n_sents": 288, "n_sents_1": 176, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 89, "L_abs": 1457, "title": "Exponential Topology-enabled Scalable Communication in Multi-agent Reinforcement Learning", "abs": "In cooperative multi-agent reinforcement learning (MARL), well-designed communication protocols can effectively facilitate consensus among agents, thereby enhancing task performance. Moreover, in large-scale multi-agent systems commonly found in real-world applications, effective communication plays an even more critical role due to the escalated challenge of partial observability compared to smaller-scale setups. In this work, we endeavor to develop a scalable communication protocol for MARL. Unlike previous methods that focus on selecting optimal pairwise communication links\u2014a task that becomes increasingly complex as the number of agents grows\u2014we adopt a global perspective on communication topology design. Specifically, we propose to utilize the exponential topology to enable rapid information dissemination among agents by leveraging its small-diameter and small-size properties. This approach leads to a scalable communication protocol, named ExpoComm. To fully unlock the potential of exponential graphs as communication topologies, we employ memory-based message processors and auxiliary tasks to ground messages, ensuring that they reflect global information and benefit decision-making. Extensive experiments on large-scale cooperative benchmarks, including MAgent and Infrastructure Management Planning, demonstrate the superior performance and robust zero-shot transferability of ExpoComm compared to existing communication strategies.", "keywords": [ "multi-agent reinforcement learning", "communication" ], "rating_list": [ 6, 5, 8, 6 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 4, 2 ], "confidence_list": [ 4, 3, 4, 2 ] }, { "paper_id": "CKx7eOYFG8", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 5, "author_info_list": [ { "name": "Melika Ayoughi", "gender": "unknown", "institution": "University of Amsterdam, University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Pascal Mettes", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Paul Groth", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Full Professor" }, { "name": "Tejaswi Kasarla", "gender": "Female", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Thomas Paul Alexandre Wiggers", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" } ], "n_formula": 11, "n_formula_1": 7, "n_ref_uni": 39, "n_ref": 107, "n_ref_all": 122, "n_fig": 3, "n_tab": 4, "L_tab": 1073, "n_element_tab": 98, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 1072, "n_element_tab_1": 96, "formula_len_all": 662, "formula_len_all_1": 420, "len_all": 114286, "len_all_1": 63675, "len_abs": 1445, "len_title": 123, "len_sents": 35171, "len_sents_1": 29412, "n_sents": 317, "n_sents_1": 237, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1469, "title": "Exemplar-free Continual Representation Learning with Symmetric Distillation", "abs": "Continual learning strives to train a model in a sequential manner by learning from new tasks while retaining information about old tasks. Treating this as a common classification problem leads to catastrophic forgetting, especially in deep learning settings, where knowledge of old tasks is forgotten as soon as a model is optimized on new tasks. Existing solutions tackle this problem by imposing strict assumptions, such as the availability of exemplars from previously seen classes or a warm start of a model on many classes before starting the continual learning. While effective on known benchmarks, such assumptions can be impractical and do not directly address the stability-plasticity dilemma in continual learning. In this paper, we follow a recent push in the field to tackle continual learning in the exemplar-free cold-start setting. We propose Model-in-the-Middle (MITM). The idea behind MITM is to separate the learning of new classes and retention of past class knowledge by using two distinct models. We propose a learner with symmetric distillation from both models, enabling us to learn evolving representations as new tasks arrive. We show that explicitly separating and balancing old and new tasks through symmetric distillation helps absorb large distribution shifts in between tasks, mitigating the stability gap. Our approach is simple yet outperforms the state-of-the-art in the challenging exemplar-free cold-start continual learning setting.", "keywords": [ "continual learning", "class-incremental learning" ], "rating_list": [ 5, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 1 ], "confidence_list": [ 5, 5, 5, 4 ] }, { "paper_id": "CKqiQosLKc", "primary_area": "reinforcement learning", "n_author": 3, "author_info_list": [ { "name": "Siamak Ravanbakhsh", "gender": "unknown", "institution": "McGill University", "country": "", "position": "Assistant Professor" }, { "name": "Tara Akhound-Sadegh", "gender": "unknown", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "Vineet Jain", "gender": "unknown", "institution": "McGill University", "country": "", "position": "PhD student" } ], "n_formula": 19, "n_formula_1": 19, "n_ref_uni": 33, "n_ref": 44, "n_ref_all": 57, "n_fig": 5, "n_tab": 2, "L_tab": 115, "n_element_tab": 4, "n_fig_1": 5, "n_tab_1": 1, "L_tab_1": 109, "n_element_tab_1": 3, "formula_len_all": 1331, "formula_len_all_1": 1519, "len_all": 109622, "len_all_1": 55431, "len_abs": 1045, "len_title": 99, "len_sents": 30319, "len_sents_1": 24430, "n_sents": 230, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 51, "L_abs": 1049, "title": "Sampling from Energy-based Policies using Diffusion", "abs": "Energy-based policies offer a flexible framework for modeling complex, multimodal behaviors in reinforcement learning (RL). In maximum entropy RL, the optimal policy is a Boltzmann distribution derived from the soft Q-function, but direct sampling from this distribution in continuous action spaces is computationally intractable. As a result, existing methods typically use simpler parametric distributions, like Gaussians, for policy representation \u2014 limiting their ability to capture the full complexity of multimodal action distributions. In this paper, we introduce a diffusion-based approach for sampling from energy-based policies, where the negative Q-function defines the energy function. Based on this approach, we propose an actor-critic method called Diffusion Q-Sampling (DQS) that enables more expressive policy representations, allowing stable learning in diverse environments. We show that our approach enhances exploration and captures multimodal behavior in continuous control tasks, addressing key limitations of existing methods.", "keywords": [ "Reinforcement learning", "Diffusion models" ], "rating_list": [ 3, 3, 3, 6 ], "soundness_list": [ 1, 2, 2, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 3, 3 ] }, { "paper_id": "CKdlPUWDEE", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Bohan Zhuang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Research Professor" }, { "name": "Jianfei Cai", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "Full Professor" }, { "name": "Jing Liu", "gender": "Male", "institution": "ByteDance Inc.", "country": "US", "position": "Researcher" }, { "name": "Mingyang Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Postdoc" }, { "name": "Ruihao Gong", "gender": "Male", "institution": "SenseTime", "country": "CN", "position": "Principal Researcher" }, { "name": "Yefei He", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 45, "n_ref": 96, "n_ref_all": 130, "n_fig": 13, "n_tab": 10, "L_tab": 4652, "n_element_tab": 568, "n_fig_1": 9, "n_tab_1": 7, "L_tab_1": 1787, "n_element_tab_1": 190, "formula_len_all": 189, "formula_len_all_1": 189, "len_all": 186864, "len_all_1": 69017, "len_abs": 1748, "len_title": 131, "len_sents": 47753, "len_sents_1": 31379, "n_sents": 351, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 1762, "title": "ME-Switch: A Memory-Efficient Expert Switching Framework for Large Language Models", "abs": "The typical process for LLM\u2019s development involves pre-training a general foundation model on massive data, followed by fine-tuning on task-specific data to obtain a series of specialized experts. Serving these experts can pose significant memory challenges, as loading all experts onto devices is impractical, and frequent switching between experts in response to user requests can incur substantial I/O costs. Previous approaches decompose the expert weights as the pre-trained weights plus delta weights, followed by quantizing the delta weights using output channel-wise step sizes to reduce the model size. However, these methods overlook the fact that certain input channels of delta weights can cause significant quantization errors at extremely low bitwidths. To this end, we introduce ME-Switch, a memory-efficient expert switching framework tailored for serving multiple LLMs. To condense the number of bits required for describing the delta weights, we propose a salient-aware delta compression method that first identifies which input channels of delta weights are salient based on reconstruction error and then employs mixed-precision quantization that selectively quantizes non-salient input channels of delta weights to extremely low bits while keeping the salient ones intact, significantly reducing storage demand while maintaining performance. Extensive experiments show the promising memory efficiency and accuracy of ME-Switch. For example, when serving three models from the Mistral-7B family, ME-Switch reduces the model size by 2.04$\\times$ and maintains nearly lossless performance on instruction, mathematical reasoning, and code generation tasks. Furthermore, our method can efficiently serve 16 Mistral-7B models on an NVIDIA A100 GPU.", "keywords": [ "Large Language Model", "Memory Efficient Compression" ], "rating_list": [ 6, 6, 5, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 4, 3, 2, 2 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "CKYsXi0dOV", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Caiming Xiong", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Research Scientist" }, { "name": "Can Qin", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Honglu Zhou", "gender": "Female", "institution": "SalesForce.com", "country": "US", "position": "Research Scientist" }, { "name": "Juan Carlos Niebles", "gender": "Male", "institution": "Salesforce Research", "country": "US", "position": "Director, Research" }, { "name": "Le Xue", "gender": "Male", "institution": "Salesforce", "country": "US", "position": "Researcher" }, { "name": "Manli Shu", "gender": "Female", "institution": "Salesforce Research", "country": "US", "position": "Researcher" }, { "name": "Michael S Ryoo", "gender": "Male", "institution": "Salesforce AI Research", "country": "US", "position": "Principal Researcher" }, { "name": "Ran Xu", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "senior manager" }, { "name": "Shrikant Kendre", "gender": "Male", "institution": "SalesForce.com", "country": "US", "position": "Researcher" }, { "name": "Silvio Savarese", "gender": "Male", "institution": "Salesforce", "country": "", "position": "Chief Scientist" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 31, "n_ref": 65, "n_ref_all": 77, "n_fig": 5, "n_tab": 5, "L_tab": 2621, "n_element_tab": 224, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2621, "n_element_tab_1": 224, "formula_len_all": 472, "formula_len_all_1": 472, "len_all": 144438, "len_all_1": 64648, "len_abs": 843, "len_title": 72, "len_sents": 29671, "len_sents_1": 29061, "n_sents": 217, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Reject", "L_title": 71, "L_abs": 845, "title": "BLIP-3-Video: You Only Need 32 Tokens to Represent a Video Even in VLMs", "abs": "We present BLIP-3-Video, a multimodal language model for videos, particularly designed to efficiently capture temporal information over multiple frames. BLIP-3-Video takes advantage of the `temporal encoder' in addition to the conventional visual tokenizer, which maps a sequence of tokens over multiple frames into a compact set of visual tokens. This enables BLIP-3-Video to use much fewer visual tokens than its competing models (e.g., 32 vs. 4608 tokens). We explore different types of temporal encoders, including learnable spatio-temporal pooling as well as sequential models like Token Turing Machines. We experimentally confirm that BLIP-3-Video obtains video question-answering accuracies comparable to much larger state-of-the-art models (e.g., 34B), while being much smaller (i.e., 4B) and more efficient by using fewer visual tokens.", "keywords": [ "video representation", "video foundation model", "vlm", "multimodal language model" ], "rating_list": [ 6, 6, 8, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 3, 2 ], "contribution_list": [ 1, 3, 4, 2 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "CKXul9iX77", "primary_area": "optimization", "n_author": 4, "author_info_list": [ { "name": "Aron Brenner", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" }, { "name": "Jennifer Z Sun", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Undergrad student" }, { "name": "Rahman Khorramfar", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Saurabh Amin", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 24, "n_formula_1": 13, "n_ref_uni": 32, "n_ref": 72, "n_ref_all": 88, "n_fig": 4, "n_tab": 7, "L_tab": 1487, "n_element_tab": 145, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2105, "formula_len_all_1": 520, "len_all": 139726, "len_all_1": 65777, "len_abs": 1560, "len_title": 126, "len_sents": 49611, "len_sents_1": 33581, "n_sents": 324, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 78, "L_abs": 1574, "title": "A Deep Generative Learning Approach for Two-stage Adaptive Robust Optimization", "abs": "Two-stage adaptive robust optimization (ARO) is a powerful approach for planning under uncertainty, balancing first-stage decisions with recourse decisions made after uncertainty is realized. To account for uncertainty, modelers typically define a simple uncertainty set over which potential outcomes are considered. However, classical methods for defining these sets unintentionally capture a wide range of unrealistic outcomes, resulting in overly-conservative and costly planning in anticipation of unlikely contingencies. In this work, we introduce AGRO, a solution algorithm that performs adversarial generation for two-stage adaptive robust optimization using a variational autoencoder. AGRO generates high-dimensional contingencies that are simultaneously adversarial and realistic, improving the robustness of first-stage decisions at a lower planning cost than standard methods. To ensure generated contingencies lie in high-density regions of the uncertainty distribution, AGRO defines a tight uncertainty set as the image of \"latent\" uncertainty sets under the VAE decoding transformation. Projected gradient ascent is then used to maximize recourse costs over the latent uncertainty sets by leveraging differentiable optimization methods. We demonstrate the cost-efficiency of AGRO by applying it to both a synthetic production-distribution problem and a real-world power system expansion setting. We show that AGRO outperforms the standard column-and-constraint algorithm by up to 1.8% in production-distribution planning and up to 8% in power system expansion.", "keywords": [ "robust optimization", "stochastic optimization", "discrete optimization", "deep learning", "unsupervised learning" ], "rating_list": [ 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "CJnceDksRd", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 6, "author_info_list": [ { "name": "Dan dan Guo", "gender": "Female", "institution": "Jilin University", "country": "CN", "position": "Lecturer" }, { "name": "Hangting Ye", "gender": "Male", "institution": "Jilin University", "country": "", "position": "PhD student" }, { "name": "He Zhao", "gender": "Male", "institution": "Commonwealth Scientific and Industrial Research Organisation, CSIRO", "country": "AU", "position": "Researcher" }, { "name": "Mingyuan Zhou", "gender": "Male", "institution": "Google", "country": "US", "position": "Researcher" }, { "name": "Wei Fan", "gender": "Male", "institution": "University of Oxford", "country": "GB", "position": "Postdoc Researcher" }, { "name": "Yi Chang", "gender": "Male", "institution": "Jilin University, China", "country": "CN", "position": "Full Professor" } ], "n_formula": 12, "n_formula_1": 9, "n_ref_uni": 45, "n_ref": 121, "n_ref_all": 154, "n_fig": 11, "n_tab": 13, "L_tab": 23108, "n_element_tab": 2121, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1148, "n_element_tab_1": 130, "formula_len_all": 1856, "formula_len_all_1": 1439, "len_all": 209023, "len_all_1": 70995, "len_abs": 1659, "len_title": 117, "len_sents": 51306, "len_sents_1": 33701, "n_sents": 342, "n_sents_1": 206, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1664, "title": "DRL: Decomposed Representation Learning for Tabular Anomaly Detection", "abs": "Anomaly detection, indicating to identify the anomalies that significantly deviate from the majority normal instances of data, has been an important role in machine learning and related applications. Despite the significant success achieved in anomaly detection on image and text data, the accurate Tabular Anomaly Detection (TAD) has still been hindered due to the lack of clear prior semantic information in the tabular data. Most state-of-the-art TAD studies are along the line of reconstruction, which first reconstruct training data and then use reconstruction errors to decide anomalies; however, reconstruction on training data can still hardly distinguish anomalies due to the data entanglement in their representations. To address this problem, in this paper, we propose a novel approach Decomposed Representation Learning (DRL), to re-map data into a tailor-designed constrained space, in order to capture the underlying shared patterns of normal samples and differ anomalous patterns for TAD.\nSpecifically, we enforce the representation of each normal sample in the latent space to be decomposed into a weighted linear combination of randomly generated orthogonal basis vectors, where these basis vectors are both data-free and training-free.\nFurthermore, we enhance the discriminative capability between normal and anomalous patterns in the latent space by introducing a novel constraint that amplifies the discrepancy between these two categories, supported by theoretical analysis. \nFinally, extensive experiments on 40 tabular datasets and 16 competing tabular anomaly detection algorithms show that our method achieves state-of-the-art performance.", "keywords": [ "Anomaly detection", "Tabular data", "Tabular representation learning" ], "rating_list": [ 6, 6, 5, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 2, 3, 2 ], "contribution_list": [ 2, 3, 3, 3 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "CJWMXqAnAy", "primary_area": "applications to robotics, autonomy, planning", "n_author": 9, "author_info_list": [ { "name": "Difan Zou", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Hanxiang Ren", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Li Sun", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Pei Zhou", "gender": "unknown", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Siyan Dong", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Xulong Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yanchao Yang", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Assistant Professor" }, { "name": "Youyi Zheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Zewen Wu", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "PhD student" } ], "n_formula": 20, "n_formula_1": 13, "n_ref_uni": 48, "n_ref": 107, "n_ref_all": 126, "n_fig": 12, "n_tab": 26, "L_tab": 10264, "n_element_tab": 1130, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 821, "n_element_tab_1": 115, "formula_len_all": 10414, "formula_len_all_1": 10087, "len_all": 207115, "len_all_1": 82527, "len_abs": 1401, "len_title": 117, "len_sents": 49199, "len_sents_1": 34102, "n_sents": 409, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 69, "L_abs": 1406, "title": "Optimization-Biased Hypernetworks for Generalizable Policy Generation", "abs": "Policy learning through behavior cloning poses significant challenges, particularly when demonstration data is limited. In this work, we present HyPoGen, a novel optimization-biased hypernetwork structure for policy generation. The proposed hypernetwork learns to synthesize optimal policy parameters solely from task specifications, by modeling policy generation as an approximation of the optimization process executed over a finite number of steps and assuming these specifications serve as a sufficient representation of the demonstration data. By incorporating structural designs that bias the hypernetwork towards optimization, we can improve its generalization capability while being trained only on source task demonstrations. During the feed-forward prediction pass, the hypernetwork effectively performs an optimization in the latent (compressed) policy space, which is then decoded into policy parameters for action prediction. Experimental results on locomotion and manipulation benchmarks show that HyPoGen significantly outperforms state-of-the-art methods in generating policies for unseen target tasks without any demonstrations, achieving higher success rates and evidencing improved generalizable policy generation capability. Our work underscores the potential of optimization-biased hypernetworks in advancing generalizable policy generation. Our code and models will be made available.", "keywords": [ "hypernetwork", "policy generation", "behavior cloning" ], "rating_list": [ 6, 8, 8, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 4, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "CJEBFNBLhO", "primary_area": "datasets and benchmarks", "n_author": 2, "author_info_list": [ { "name": "Ming Zhu", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "", "position": "Assistant Professor" }, { "name": "Xiao-Yang Liu", "gender": "Male", "institution": "Columbia University", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 6, "n_ref_uni": 31, "n_ref": 58, "n_ref_all": 67, "n_fig": 1, "n_tab": 7, "L_tab": 2287, "n_element_tab": 300, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 462, "n_element_tab_1": 52, "formula_len_all": 727, "formula_len_all_1": 669, "len_all": 106674, "len_all_1": 46651, "len_abs": 1341, "len_title": 153, "len_sents": 24597, "len_sents_1": 20982, "n_sents": 192, "n_sents_1": 155, "venue": "ICLR2025", "decision": "withdraw", "L_title": 104, "L_abs": 1353, "title": "Massively Parallel Environments for Large-Scale Combinatorial Optimizations Using Reinforcement Learning", "abs": "Most combinatorial optimization (CO) problems are NP-hard and difficult to find high-quality solutions. Reinforcement learning (RL) is a promising technique due to its powerful search capability; however, sampling speed is a common bottleneck. Current benchmark works only provide instance-wise approaches, while our work cover both instance-wise and distribution-wise approaches, especially in large-scale CO problems. In this paper, we build 24 GPU-based massively parallel environments for 12 CO problems, i.e., each problem has two environments; and use them to train RL-based approaches. We reproduce benchmark RL algorithms, including instance-wise and distribution-wise approaches especially in large-scale CO problems, on both synthetic datasets and real-world datasets. Take the graph maxcut problem as an example. The sampling speed is improved by at least two orders over conventional implementations, and the scale (i.e., number of nodes) of trained problems in a distribution-wise approach is up to thousands of nodes, i.e., improved by one order. The objective value obtained by inference (100 $\\sim$ 200 seconds) in the distribution-wise scenario is almost the same as the state-of-the-art (SOTA) solver Gurobi (running for 1 hour), and better than the SOTA RL-based approach. The code is available at: https://github.com/OpenAfterReview.", "keywords": [ "Combinatorial Optimizations", "Massively Parallel Environments", "Reinforcement learning", "distribution-wise approach" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 2, 3 ] }, { "paper_id": "CIs9x2ZRgh", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 10, "author_info_list": [ { "name": "Fangjun Kuang", "gender": "Male", "institution": "Xiaomi", "country": "", "position": "engineer" }, { "name": "Han Zhu", "gender": "Male", "institution": "Xiaomi Corporation", "country": "CN", "position": "Researcher" }, { "name": "Liyong Guo", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "", "position": "Undergrad student" }, { "name": "Long Lin", "gender": "Female", "institution": "Xiaomi corp.", "country": "", "position": "Researcher" }, { "name": "Wei Kang", "gender": "Male", "institution": "Xiaomi Corp.", "country": "CN", "position": "Researcher" }, { "name": "Xiaoyu Yang", "gender": "Male", "institution": "University of Cambridge", "country": "GB", "position": "PhD student" }, { "name": "Zengrui Jin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Zengwei Yao", "gender": "Male", "institution": "Xiaomi Corporation", "country": "", "position": "Researcher" }, { "name": "Zhaoqing Li", "gender": "Male", "institution": "Chinese University of Hong Kong", "country": "", "position": "PhD student" }, { "name": "Daniel Povey", "gender": "Male", "institution": "Johns Hopkins University", "country": "", "position": "Assistant Research Professor" } ], "n_formula": 8, "n_formula_1": 7, "n_ref_uni": 64, "n_ref": 145, "n_ref_all": 172, "n_fig": 2, "n_tab": 12, "L_tab": 2759, "n_element_tab": 287, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 2310, "n_element_tab_1": 122, "formula_len_all": 572, "formula_len_all_1": 470, "len_all": 154275, "len_all_1": 58781, "len_abs": 1217, "len_title": 121, "len_sents": 34743, "len_sents_1": 26246, "n_sents": 239, "n_sents_1": 159, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 73, "L_abs": 1285, "title": "CR-CTC: Consistency regularization on CTC for improved speech recognition", "abs": "Connectionist Temporal Classification (CTC) is a widely used method for automatic speech recognition (ASR), renowned for its simplicity and computational efficiency. However, it often falls short in recognition performance. In this work, we propose the Consistency-Regularized CTC (CR-CTC), which enforces consistency between two CTC distributions obtained from different augmented views of the input speech mel-spectrogram. We provide in-depth insights into its essential behaviors from three perspectives: 1) it conducts self-distillation between random pairs of sub-models that process different augmented views; 2) it learns contextual representation through masked prediction for positions within time-masked regions, especially when we increase the amount of time masking; 3) it suppresses the extremely peaky CTC distributions, thereby reducing overfitting and improving the generalization ability. Extensive experiments on LibriSpeech, Aishell-1, and GigaSpeech datasets demonstrate the effectiveness of our CR-CTC. It significantly improves the CTC performance, achieving state-of-the-art results comparable to those attained by transducer or systems combining CTC and attention-based encoder-decoder (CTC/AED). We release our code at \\url{https://github.com/k2-fsa/icefall}.", "keywords": [ "Consistency regularization", "CTC", "speech recognition" ], "rating_list": [ 8, 8, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 5, 4, 5 ] }, { "paper_id": "CIcMuee69B", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 7, "author_info_list": [ { "name": "Alejo Garat", "gender": "Male", "institution": "Universidad ORT Uruguay", "country": "UY", "position": "Lecturer" }, { "name": "Franz Mayr", "gender": "Not Specified", "institution": "Universidad ORT Uruguay", "country": "UY", "position": "Associate Professor" }, { "name": "Johny Kidd", "gender": "Male", "institution": "Universidad ORT Uruguay", "country": "UY", "position": "Researcher" }, { "name": "Juan Pedro da Silva Barloco", "gender": "Male", "institution": "Universidad ORT Uruguay", "country": "UY", "position": "MS student" }, { "name": "Martin Iturbide", "gender": "Male", "institution": "Universidad ORT Uruguay", "country": "UY", "position": "MS student" }, { "name": "Mat\u00edas Carrasco", "gender": "Male", "institution": "Universidad ORT Uruguay", "country": "UY", "position": "Researcher" }, { "name": "Sergio Yovine", "gender": "unknown", "institution": "Universidad ORT Uruguay", "country": "UY", "position": "Full Professor" } ], "n_formula": 40, "n_formula_1": 27, "n_ref_uni": 13, "n_ref": 23, "n_ref_all": 52, "n_fig": 11, "n_tab": 3, "L_tab": 442, "n_element_tab": 34, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 349, "n_element_tab_1": 18, "formula_len_all": 4616, "formula_len_all_1": 3657, "len_all": 95873, "len_all_1": 63194, "len_abs": 647, "len_title": 132, "len_sents": 30016, "len_sents_1": 23500, "n_sents": 310, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 340, "title": "A probablistic automata learning approach for analyzing and sampling constrained LLM", "abs": "We define a congruence that copes with null next-symbol probabilities that arise when the output of a language model is constrained by some means during text generation. We develop an algorithm for efficiently learning the quotient with respect to this congruence and evaluate it on case studies for analyzing statistical properties of LLM.", "keywords": [ "Grammatical Inference", "Probabilistic Deterministic Finite Automata", "Active Learning", "LLM" ], "rating_list": [ 6, 5, 5, 3, 3 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 2, 2, 1, 1 ], "contribution_list": [ 3, 3, 3, 2, 2 ], "confidence_list": [ 3, 3, 4, 3, 4 ] }, { "paper_id": "CIN2VRxPKU", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Chhavi Yadav", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Kamalika Chaudhuri", "gender": "Female", "institution": "UC San Diego, University of California, San Diego", "country": "", "position": "Full Professor" }, { "name": "Ruihan Wu", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "Postdoc" }, { "name": "Ruslan Salakhutdinov", "gender": "Male", "institution": "School of Computer Science, Carnegie Mellon University", "country": "US", "position": "Full Professor" } ], "n_formula": 18, "n_formula_1": 16, "n_ref_uni": 32, "n_ref": 42, "n_ref_all": 66, "n_fig": 6, "n_tab": 1, "L_tab": 519, "n_element_tab": 23, "n_fig_1": 6, "n_tab_1": 2, "L_tab_1": 919, "n_element_tab_1": 98, "formula_len_all": 2064, "formula_len_all_1": 1338, "len_all": 122517, "len_all_1": 66136, "len_abs": 1490, "len_title": 100, "len_sents": 37783, "len_sents_1": 29764, "n_sents": 309, "n_sents_1": 239, "venue": "ICLR2025", "decision": "Reject", "L_title": 51, "L_abs": 1358, "title": "Evaluating Deep Unlearning in Large Language Models", "abs": "Machine unlearning has emerged as an important component in developing safe and trustworthy models. Prior work on unlearning in LLMs has mostly considered unlearning tasks where a large corpus of copyrighted material or some specific training data are required to be removed. In this work, we consider the task of unlearning a fact from LLMs, which can be challenging as related facts can be deduced from each other, and investigate how well current unlearning methods for LLMs succeed at this task. Specifically, we formally propose a framework and a definition for deep unlearning facts that are interrelated. We design the metric, recall, to quantify the extent of deep unlearning. To enable us to systematically evaluate the extent of deep unlearning undistracted by other factors, we construct a synthetic dataset EDU-RELAT, which consists of a synthetic knowledge base of family relationships and biographies, together with a realistic logical rule set that connects them. We use this dataset to test four unlearning methods in four LLMs at different sizes. Our findings reveal that in the task of deep unlearning only a single fact, they either fail to properly unlearn with high recall, or end up unlearning many other irrelevant facts. Our dataset and code are publicly available at: https://anonymous.4open.science/r/deep_unlearning_anonymous-2C73.", "keywords": [ "large language models", "machine unlearning", "knowledge base" ], "rating_list": [ 5, 5, 6 ], "soundness_list": [ 3, 2, 2 ], "presentation_list": [ 3, 1, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 3, 5, 4 ] }, { "paper_id": "CI9JMBAsPg", "primary_area": "datasets and benchmarks", "n_author": 24, "author_info_list": [ { "name": "Bin Wang", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Bin Wang", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Bo Zhang", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Botian Shi", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Chao Xu", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Conghui He", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Daocheng Fu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Fei Wu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "Full Professor" }, { "name": "Hancheng Ye", "gender": "Male", "institution": "Duke University, Duke University", "country": "US", "position": "PhD student" }, { "name": "Haoyang Peng", "gender": "Male", "institution": "Fudan University", "country": "", "position": "MS student" }, { "name": "Hongbin Zhou", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Jiahao Pi", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "", "position": "Researcher" }, { "name": "Min Dou", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Pinlong Cai", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Sheng Zhou", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" }, { "name": "Shiyang Feng", "gender": "Male", "institution": "Fudan University", "country": "", "position": "PhD student" }, { "name": "Song Mao", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Wenjie Wu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Xiangchao Yan", "gender": "unknown", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yongwei Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yu Qiao", "gender": "Male", "institution": "Shanghai Aritifcal Intelligence Laboratory", "country": "CN", "position": "Principal Researcher" }, { "name": "Junchi Yan", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Li Mingsheng", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "MS student" }, { "name": "renqiu xia", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 43, "n_ref": 114, "n_ref_all": 139, "n_fig": 5, "n_tab": 9, "L_tab": 5387, "n_element_tab": 539, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 2488, "n_element_tab_1": 201, "formula_len_all": 133, "formula_len_all_1": 133, "len_all": 184213, "len_all_1": 60051, "len_abs": 1593, "len_title": 155, "len_sents": 43435, "len_sents_1": 26503, "n_sents": 328, "n_sents_1": 178, "venue": "ICLR2025", "decision": "Reject", "L_title": 106, "L_abs": 1522, "title": "DocGenome: A Large Benchmark for Multi-Modal Language Models in Real-World Academic Document Understanding", "abs": "Scientific documents record research findings and valuable human knowledge, comprising a vast corpus of high-quality data. Thus, leveraging multi-modality data extracted from these documents and assessing large models' abilities to handle scientific document-oriented tasks is meaningful. Despite promising advancements, large models still perform poorly on multi-page scientific document extraction and understanding tasks, and their capacity to process within-document data formats such as charts and equations remains under-explored. To address these issues, we present DocGenome, a structured document dataset constructed by annotating 500K scientific documents from 153 disciplines in the arXiv open-access community, using our custom auto-labeling pipeline. DocGenome features four characteristics: 1) Completeness: It is the first dataset to structure data from all modalities including 13 layout attributes along with their LaTeX source codes. 2) Logicality: It provides 6 logical relationships between different entities within each scientific document. 3) Diversity: It covers various document-oriented tasks, including document classification, visual grounding, document layout detection, document transformation, open-ended single-page QA and multi-page QA. 4) Correctness: It undergoes rigorous quality control checks conducted by a specialized team. We conduct extensive experiments to demonstrate the advantages of DocGenome and objectively evaluate the performance of current large models on our benchmark.", "keywords": [ "Scientific document structuring", "Document understanding", "Chart Table and Equation Understanding" ], "rating_list": [ 8, 6, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 4, 5, 4 ] }, { "paper_id": "CI5Cj0vktS", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Alexander Kolesov", "gender": "Male", "institution": "The Skolkovo Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Jaemoo Choi", "gender": "Male", "institution": "Georgia Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "Jaewoong Choi", "gender": "Male", "institution": "Sung Kyun Kwan University", "country": "KR", "position": "Assistant Professor" }, { "name": "Milena Gazdieva", "gender": "Female", "institution": "Skolkovo Institute of Science and Technology", "country": "", "position": "PhD student" }, { "name": "Petr Mokrov", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "PhD student" }, { "name": "Alexander Korotin", "gender": "Male", "institution": "Skolkovo Institute of Science and Technology", "country": "RU", "position": "Assistant Professor" } ], "n_formula": 104, "n_formula_1": 37, "n_ref_uni": 55, "n_ref": 141, "n_ref_all": 189, "n_fig": 9, "n_tab": 6, "L_tab": 525, "n_element_tab": 68, "n_fig_1": 6, "n_tab_1": 3, "L_tab_1": 146, "n_element_tab_1": 12, "formula_len_all": 12235, "formula_len_all_1": 3609, "len_all": 217892, "len_all_1": 73380, "len_abs": 3867, "len_title": 123, "len_sents": 60256, "len_sents_1": 31596, "n_sents": 568, "n_sents_1": 262, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 75, "L_abs": 1027, "title": "Robust Barycenter Estimation using Semi-Unbalanced Neural Optimal Transport", "abs": "Aggregating data from multiple sources can be formalized as an *Optimal Transport* (OT) barycenter problem, which seeks to compute the average of probability distributions with respect to OT discrepancies. However, in real-world scenarios, the presence of outliers and noise in the data measures can significantly hinder the performance of traditional statistical methods for estimating OT barycenters. To address this issue, we propose a novel, scalable approach for estimating the *robust* continuous barycenter, leveraging the dual formulation of the *(semi-)unbalanced* OT problem. To the best of our knowledge, this paper is the first attempt to develop an algorithm for robust barycenters under the continuous distribution setup. Our method is framed as a $\\min$-$\\max$ optimization problem and is adaptable to *general* cost function. We rigorously establish the theoretical underpinnings of the proposed method and demonstrate its robustness to outliers and class imbalance through a number of illustrative experiments.", "keywords": [ "unbalanced optimal transport", "barycenter", "generative modeling" ], "rating_list": [ 8, 5, 6, 6 ], "soundness_list": [ 4, 3, 3, 2 ], "presentation_list": [ 4, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 1 ], "confidence_list": [ 4, 4, 4, 3 ] }, { "paper_id": "CI4sCBMXjP", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Futing Wang", "gender": "Female", "institution": "Westlake University", "country": "", "position": "PhD student" }, { "name": "Jianhao Yan", "gender": "Male", "institution": "Westlake University", "country": "", "position": "PhD student" }, { "name": "Tao Lin", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Assistant Professor" }, { "name": "Yue Zhang", "gender": "Male", "institution": "Westlake University", "country": "CN", "position": "Full Professor" } ], "n_formula": 3, "n_formula_1": 3, "n_ref_uni": 72, "n_ref": 102, "n_ref_all": 133, "n_fig": 9, "n_tab": 11, "L_tab": 6170, "n_element_tab": 490, "n_fig_1": 4, "n_tab_1": 5, "L_tab_1": 1811, "n_element_tab_1": 100, "formula_len_all": 110, "formula_len_all_1": 110, "len_all": 206541, "len_all_1": 65641, "len_abs": 1226, "len_title": 59, "len_sents": 49619, "len_sents_1": 29567, "n_sents": 393, "n_sents_1": 219, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 59, "L_abs": 1220, "title": "ELICIT: LLM Augmentation Via External In-context Capability", "abs": "Enhancing the adaptive capabilities of large language models is a critical pursuit in both research and application.\nTraditional fine-tuning methods require substantial data, computational resources, and specific capabilities, while in-context learning is limited by the need for appropriate demonstrations and efficient token usage.\n Inspired by the expression of in-context learned capabilities through task vectors and the concept of modular capability or knowledge, we propose ELICIT, a framework consisting of two modules designed to effectively store and reuse task vectors to enhance the diverse adaptive capabilities of models without additional training or inference tokens.\n Our comprehensive experiments and analysis demonstrate that our pipeline is highly transferable across different input formats, tasks, and model architectures.\n Externally storing and reusing vectors that represent in-context learned capabilities not only shows the potential to extract modular capabilities but also significantly enhances the performance, versatility, adaptability, and scalability of large language models, paving the way for more efficient and effective use of these models in a wide range of applications.", "keywords": [ "modular" ], "rating_list": [ 8, 3, 8, 8 ], "soundness_list": [ 3, 3, 4, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 4, 3, 3 ] }, { "paper_id": "CH7Ba4RFa2", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Tian He", "gender": "Female", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "PhD student" }, { "name": "Yifan Jiao", "gender": "Male", "institution": "University of the Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Yu Wang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Zhenming Zhang", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Zhuqing Jiang", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "Associate Professor" }, { "name": "weijie qiu", "gender": "unknown", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" } ], "n_formula": 15, "n_formula_1": 15, "n_ref_uni": 30, "n_ref": 79, "n_ref_all": 87, "n_fig": 2, "n_tab": 8, "L_tab": 1552, "n_element_tab": 97, "n_fig_1": 2, "n_tab_1": 8, "L_tab_1": 3763, "n_element_tab_1": 259, "formula_len_all": 580, "formula_len_all_1": 580, "len_all": 98466, "len_all_1": 67361, "len_abs": 956, "len_title": 122, "len_sents": 25071, "len_sents_1": 20960, "n_sents": 198, "n_sents_1": 173, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 961, "title": "Seg-LaneDet: 3D Lane Detection from Monocular Images with 2D Segmentation", "abs": "Monocular 3D lane detection is a fundamental yet challenging task in autonomous driving. Recent advancements primarily rely on constructing 3D surrogates from monocular images and camera parameters. However, misalignment is introduced in current methods due to the lack of dense depth information in datasets, coupled with the inherent depth ambiguity of monocular images. To address this issue, we propose Seg-LaneDet, a simple but effective end-to-end 3D lane detector. We frame the task of 3D lane detection as an elevation from 2D to 3D detection. Specifically, we leverage a pre-trained 2D lane detector to obtain instance segmentation of lanes, of which the segmentation maps serve as the sole prior for the 2D-to-3D module. This allows us to achieve a straightforward 3D lane representation based on front-view segmentation maps. Our method demonstrates comparable performance to state-of-the-art (SOTA) F1 scores on the OpenLane and the Apollo datasets.", "keywords": [ "3D Lane Detection", "Autonomous Driving", "Computer Vision" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 1, 2, 3 ], "confidence_list": [ 4, 4, 4 ] }, { "paper_id": "CGhgB8Kz8i", "primary_area": "foundation or frontier models, including LLMs", "n_author": 7, "author_info_list": [ { "name": "Dian Li", "gender": "Male", "institution": "Tencent PCG AI", "country": "CN", "position": "Principal Researcher" }, { "name": "Han Wang", "gender": "Female", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Hui Wang", "gender": "Male", "institution": "QQ", "country": "", "position": "Director" }, { "name": "Xiaohan Wang", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Xuguang Lan", "gender": "Male", "institution": "Xi'an Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Yilin Zhao", "gender": "Male", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "sinbadliu", "gender": "Male", "institution": "PCG QQ", "country": "", "position": "Researcher" } ], "n_formula": 6, "n_formula_1": 3, "n_ref_uni": 71, "n_ref": 87, "n_ref_all": 125, "n_fig": 11, "n_tab": 8, "L_tab": 2741, "n_element_tab": 217, "n_fig_1": 8, "n_tab_1": 4, "L_tab_1": 2278, "n_element_tab_1": 182, "formula_len_all": 496, "formula_len_all_1": 178, "len_all": 203638, "len_all_1": 62791, "len_abs": 1918, "len_title": 157, "len_sents": 52302, "len_sents_1": 24306, "n_sents": 429, "n_sents_1": 216, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 109, "L_abs": 1946, "title": "Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps", "abs": "Humor is a culturally nuanced aspect of human language that presents challenges for understanding and generation. Humor generation requires a multi-hop reasoning process where each hop is based on proper rationales. \n\t\tAlthough CloT introduce nouns as instruction receipt, thereby enhancing generalization. Creative data is inherently scarce, and relying solely on dataset fitting easily leads to being trapped in local patterns. GPT-o1 proposes a reasoning process with reflection and correction, but also fails at humor generation.\n\t\tTherefore, in this paper, we propose a more robust framework for tackling humor reasoning task, named Creative Leap of Structured Thought (CLoST). \n\t\tThe key point is to empower LLMs with reasoning ability to understand humor, rather than just memorizing some patterns.\n\t\tReward model helps to optimize the behavior of large language models by providing feedback so that they produce outputs that are more in line with expectations. Due to the subjectiveness of humor, a unified score may include great noise. Judgement ability is then a fundamental to further empower LLMs reasoning ability.\n\t\tJudgement-oriented instructions are designed to improve the capability of a model's judgment, and we also propose an automatic instruction evolutionary method to involve the deeper and broader thinking process behind humor. Then, through reinforcement learning, the reasoning logic for each online generated response is extracted using GPT-4o. In this process, external knowledge is introduced again to assist the model in logic reasoning and human preferences learning.\n\t\tFinally, experimental results demonstrate that the combination of these two methods can both enhance the model's judgment ability and improve its generative capability.\n\t\tThese findings deepen our understanding of the creative capabilities of LLMs and provide ways to enhance LLMs' creative abilities for cross-domain innovative applications.", "keywords": [ "Large Language Model", "humor generation", "reinforcement learning" ], "rating_list": [ 5, 6, 3, 6 ], "soundness_list": [ 2, 4, 2, 3 ], "presentation_list": [ 2, 3, 1, 3 ], "contribution_list": [ 2, 3, 1, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "CGfWyU28Pd", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Jinhui Xu", "gender": "Male", "institution": "University at Buffalo, State University of New York", "country": "US", "position": "Full Professor" }, { "name": "Kaiyi Ji", "gender": "Male", "institution": "State University of New York at Buffalo", "country": "US", "position": "Assistant Professor" }, { "name": "Meng Ding", "gender": "Female", "institution": "State University of New York at Buffalo", "country": "US", "position": "PhD student" } ], "n_formula": 41, "n_formula_1": 28, "n_ref_uni": 31, "n_ref": 65, "n_ref_all": 89, "n_fig": 10, "n_tab": 5, "L_tab": 1987, "n_element_tab": 187, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 1549, "n_element_tab_1": 146, "formula_len_all": 3002, "formula_len_all_1": 1749, "len_all": 145539, "len_all_1": 69548, "len_abs": 1666, "len_title": 157, "len_sents": 47955, "len_sents_1": 29692, "n_sents": 398, "n_sents_1": 234, "venue": "ICLR2025", "decision": "Reject", "L_title": 109, "L_abs": 1638, "title": "Why Fine-Tuning Struggles with Forgetting in Machine Unlearning? Theoretical Insights and a Remedial Approach", "abs": "Machine Unlearning has emerged as a significant area of research, focusing on 'removing' specific subsets of data from a trained model. Fine-tuning (FT) methods have become one of the fundamental approaches for approximating unlearning, as they effectively retain model performance. However, it is consistently observed that naive FT methods struggle to forget the targeted data. In this paper, we present the first theoretical analysis of FT methods for machine unlearning within a linear regression framework, providing a deeper exploration of this phenomenon. We investigate two scenarios with distinct features and overlapping features. Our findings reveal that FT models can achieve zero remaining loss yet fail to forget the forgetting data, unlike golden models (trained from scratch without the forgetting data). This analysis reveals that naive FT methods struggle with forgetting because the pretrained model retains information about the forgetting data, and the fine-tuning process has no impact on this retained information. To address this issue, we first propose a theoretical approach to mitigate the retention of forgetting data in the pretrained model. Our analysis shows that removing the forgetting data's influence allows FT models to match the performance of the golden model. Building on this insight, we introduce a discriminative regularization term to practically reduce the unlearning loss gap between the fine-tuned model and the golden model. Our experiments on both synthetic and real-world datasets validate these theoretical insights and demonstrate the effectiveness of the proposed regularization method.", "keywords": [ "Machine Unlearning", "Fine-Tuning", "Learning Theory" ], "rating_list": [ 6, 3, 3, 6 ], "soundness_list": [ 3, 1, 3, 3 ], "presentation_list": [ 4, 3, 2, 4 ], "contribution_list": [ 3, 1, 1, 3 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "CGbfokGFP7", "primary_area": "generative models", "n_author": 13, "author_info_list": [ { "name": "Dahua Lin", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Associate Professor" }, { "name": "Fangzhou Hong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Haozhe Xie", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Jiaxiang Tang", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Intern" }, { "name": "Liang Pan", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Research Scientist " }, { "name": "Shunsuke Saito", "gender": "Male", "institution": "Codec Avatars Lab", "country": "US", "position": "Researcher" }, { "name": "Tengfei Wang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Tong Wu", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "Postdoc" }, { "name": "Yuhao Dong", "gender": "Male", "institution": "Department of Automation, Tsinghua University", "country": "CN", "position": "Undergrad student" }, { "name": "Yushi LAN", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Zhaoxi Chen", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Ziang Cao", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "PhD student" }, { "name": "Ziwei Liu", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Assistant Professor" } ], "n_formula": 19, "n_formula_1": 14, "n_ref_uni": 49, "n_ref": 120, "n_ref_all": 144, "n_fig": 8, "n_tab": 3, "L_tab": 238, "n_element_tab": 35, "n_fig_1": 2, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 1534, "formula_len_all_1": 952, "len_all": 194622, "len_all_1": 68066, "len_abs": 311, "len_title": 124, "len_sents": 54192, "len_sents_1": 31689, "n_sents": 468, "n_sents_1": 252, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1262, "title": "3DTopia-XL: Scaling High-quality 3D Asset Generation via Primitive Diffusion", "abs": "The increasing demand for high-quality 3D assets across various industries necessitates efficient and automated 3D content creation.\nDespite recent advancements in 3D generative models, existing methods still face challenges with optimization speed, geometric fidelity, and the lack of assets for physically based rendering (PBR). In this paper, we introduce 3DTopia-XL, a scalable native 3D generative model designed to overcome these limitations. 3DTopia-XL leverages a novel primitive-based 3D representation, PrimX, which encodes detailed shape, albedo, and material field into a compact tensorial format, facilitating the modeling of high-resolution geometry with PBR assets. On top of the novel representation, we propose a generative framework based on Diffusion Transformer (DiT), which comprises 1) Primitive Patch Compression, 2) and Latent Primitive Diffusion. 3DTopia-XL learns to generate high-quality 3D assets from textual or visual inputs. We conduct extensive qualitative and quantitative experiments to demonstrate that 3DTopia-XL significantly outperforms existing methods in generating high-quality 3D assets with fine-grained textures and materials, efficiently bridging the quality gap between generative models and real-world applications.", "keywords": [ "3D Generation", "Diffusion Model", "Image-to-3D", "Text-to-3D", "PBR Asset", "3D Representation", "Primitives" ], "rating_list": [ 3, 5, 6, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "CGT0T9uUOY", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Chi-Wing Fu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Ka-Hei Hui", "gender": "Male", "institution": "Autodesk", "country": "CA", "position": "Researcher" }, { "name": "Pheng-Ann Heng", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Qianyi Wu", "gender": "Male", "institution": "Monash University", "country": "AU", "position": "PhD student" }, { "name": "Runsong Zhu", "gender": "Male", "institution": "Department of Computer Science and Engineering, The Chinese University of Hong Kong", "country": "HK", "position": "PhD student" }, { "name": "Shi Qiu", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Reseach Assistant Professor" }, { "name": "Zhengzhe Liu", "gender": "Male", "institution": "Carnegie Mellon University", "country": "US", "position": "Postdoc" } ], "n_formula": 12, "n_formula_1": 12, "n_ref_uni": 40, "n_ref": 117, "n_ref_all": 139, "n_fig": 9, "n_tab": 2, "L_tab": 961, "n_element_tab": 139, "n_fig_1": 9, "n_tab_1": 6, "L_tab_1": 993, "n_element_tab_1": 153, "formula_len_all": 813, "formula_len_all_1": 813, "len_all": 117359, "len_all_1": 67919, "len_abs": 1348, "len_title": 116, "len_sents": 33572, "len_sents_1": 31655, "n_sents": 227, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1372, "title": "Object-aware lifting for 3D scene segmentation in Gaussian splatting", "abs": "Lifting is an effective technique for producing a 3D scene segmentation by unprojecting multi-view 2D instance segmentations into a common 3D space. Existing state-of-the-art lifting methods leverage contrastive learning to learn a feature field, but rely on a hyperparameter-sensitive and error-prone clustering post-process for segmentation prediction, leading to inferior performance. In this paper, we propose a new unified \\textit{object-aware lifting} approach in a 3D Gaussian Splatting field, introducing a novel learnable \\textit{object-level codebook} to account for objects in the 3D scene for an explicit object-level understanding. To start, we augment each Gaussian point with an additional Gaussian-level feature learned using a contrastive loss. More importantly, enabled by our object-level codebook formulation, we associate the encoded object-level features with Gaussian-level point features for segmentation predictions. Further, we design two novel modules, the association learning module and the noisy label filtering module, to achieve effective and robust codebook learning. We conduct experiments on three benchmarks,~\\ie, LERF-Masked, Replica, and Messy Rooms datasets. Both qualitative and quantitative results manifest that our new approach significantly outperforms the existing methods in terms of segmentation quality and time efficiency.", "keywords": [ "3D-GS", "3D scene segmentation", "Lifting" ], "rating_list": [ 3, 6, 5, 5 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 3, 4, 3, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 5, 5, 5, 5 ] }, { "paper_id": "CGON8Btleu", "primary_area": "applications to neuroscience & cognitive science", "n_author": 6, "author_info_list": [ { "name": "Christina Sartzetaki", "gender": "unknown", "institution": "University of Amsterdam", "country": "NL", "position": "PhD student" }, { "name": "Diego Garcia Cerdas", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "MS student" }, { "name": "Gemma Roig", "gender": "Female", "institution": "Johann Wolfgang Goethe Universit\u00e4t Frankfurt am Main", "country": "DE", "position": "Full Professor" }, { "name": "Iris Groen", "gender": "Female", "institution": "University of Amsterdam", "country": "NL", "position": "Assistant Professor" }, { "name": "Magnus Petersen", "gender": "Male", "institution": ", Johann Wolfgang Goethe Universit\u00e4t Frankfurt am Main", "country": "", "position": "PhD student" }, { "name": "Pascal Mettes", "gender": "Male", "institution": "University of Amsterdam", "country": "NL", "position": "Assistant Professor" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 80, "n_ref": 123, "n_ref_all": 149, "n_fig": 14, "n_tab": 1, "L_tab": 363, "n_element_tab": 41, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 363, "n_element_tab_1": 41, "formula_len_all": 390, "formula_len_all_1": 377, "len_all": 148611, "len_all_1": 55054, "len_abs": 456, "len_title": 167, "len_sents": 43791, "len_sents_1": 26574, "n_sents": 281, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 119, "L_abs": 1631, "title": "BrainACTIV: Identifying visuo-semantic properties driving cortical selectivity using diffusion-based image manipulation", "abs": "The human brain efficiently represents visual inputs through specialized neural populations that selectively respond to specific categories. Advancements in generative modeling have enabled data-driven discovery of neural selectivity using brain-optimized image synthesis. However, current methods independently generate one sample at a time, without enforcing structural constraints on the generations; thus, these individual images have no explicit point of comparison, making it hard to discern which image features drive neural response selectivity. To address this issue, we introduce Brain Activation Control Through Image Variation (BrainACTIV), a method for manipulating a reference image to enhance or decrease activity in a target cortical region using pretrained diffusion models. Starting from a reference image allows for fine-grained and reliable offline identification of optimal visuo-semantic properties, as well as producing controlled stimuli for novel neuroimaging studies. We show that our manipulations effectively modulate predicted fMRI responses and agree with hypothesized preferred categories in established regions of interest, while remaining structurally close to the reference image. Moreover, we demonstrate how our method accentuates differences between brain regions that are selective to the same category, and how it could be used to explore neural representation of brain regions with unknown selectivities. Hence, BrainACTIV holds the potential to formulate robust hypotheses about brain representation and to facilitate the production of naturalistic stimuli for neuroscientific experiments.", "keywords": [ "brain", "selectivity", "visual cortex", "fMRI", "manipulation", "variation", "diffusion", "neuroscience" ], "rating_list": [ 6, 8, 8, 5, 6 ], "soundness_list": [ 3, 4, 3, 3, 3 ], "presentation_list": [ 3, 4, 3, 3, 4 ], "contribution_list": [ 3, 4, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 4, 5 ] }, { "paper_id": "CGOH2j1m0b", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Fatima M. Anwar", "gender": "unknown", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Jimi Oke", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Khotso Selialia", "gender": "unknown", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" }, { "name": "Yasra Chandio", "gender": "Female", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" } ], "n_formula": 42, "n_formula_1": 18, "n_ref_uni": 44, "n_ref": 95, "n_ref_all": 115, "n_fig": 12, "n_tab": 5, "L_tab": 3585, "n_element_tab": 264, "n_fig_1": 9, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 4243, "formula_len_all_1": 1745, "len_all": 208121, "len_all_1": 62022, "len_abs": 2336, "len_title": 129, "len_sents": 74153, "len_sents_1": 28248, "n_sents": 501, "n_sents_1": 208, "venue": "ICLR2025", "decision": "withdraw", "L_title": 81, "L_abs": 716, "title": "LipFed: Mitigating Subgroup Bias in Federated Learning with Lipschitz Constraints", "abs": "Federated learning (FL) has emerged as a promising paradigm for training decentralized machine learning models with privacy preservation. However, FL models are biased, which can lead to unfair model outcomes towards subgroups with intersecting attributes. To address this, we propose LipFed, a subgroup bias mitigation technique that leverages Lipschitz-based fairness constraints to mitigate subgroup bias in FL. We evaluate LipFed's efficacy in achieving subgroup fairness across clients while preserving model utility. Our experiments on benchmark datasets and real-world datasets demonstrate that LipFed effectively mitigates subgroup bias without significantly compromising group fairness or model performance.", "keywords": [ "Federated Learning", "Fairness" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 2, 2, 1 ], "contribution_list": [ 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "CFOQd4tqn1", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Hongxiang Zhao", "gender": "Male", "institution": "Beijing University of Posts and Telecommunications", "country": "CN", "position": "MS student" }, { "name": "Jianan Wang", "gender": "Female", "institution": "Astribot Inc", "country": "CN", "position": "Researcher" }, { "name": "Jingyuan Zhang", "gender": "Male", "institution": "Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Shengbang Tong", "gender": "Male", "institution": "New York University", "country": "US", "position": "PhD student" }, { "name": "Weida Wang", "gender": "Male", "institution": "Tsinghua University", "country": "", "position": "PhD student" }, { "name": "Xili Dai", "gender": "Male", "institution": "Hong Kong University of Science and Technology (Guangzhou)", "country": "CN", "position": "PhD student" }, { "name": "Yi Ma", "gender": "Male", "institution": "University of Hong Kong", "country": "HK", "position": "Full Professor" }, { "name": "Lei Zhang", "gender": "Male", "institution": "International Digital Economy Academy", "country": "CN", "position": "Chief Scientist" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 32, "n_ref": 154, "n_ref_all": 218, "n_fig": 17, "n_tab": 7, "L_tab": 814, "n_element_tab": 55, "n_fig_1": 11, "n_tab_1": 2, "L_tab_1": 115, "n_element_tab_1": 2, "formula_len_all": 766, "formula_len_all_1": 603, "len_all": 148077, "len_all_1": 61508, "len_abs": 0, "len_title": 47, "len_sents": 45614, "len_sents_1": 29263, "n_sents": 284, "n_sents_1": 186, "venue": "ICLR2025", "decision": "Reject", "L_title": 70, "L_abs": 1187, "title": "Ctrl123: Consistent Novel View Synthesis via Closed-Loop Transcription", "abs": "Based on the success of large image diffusion models, multi-view diffusion models have demonstrated remarkable zero-shot capability in novel view synthesis (NVS). However, the pioneering work Zero123 struggles to maintain consistency across generated multiple views. While recent modifications in model and training design have improved multi-view consistency, they often introduce new limitations, such as restricted fixed view generation or reliance on additional conditions. These constraints hinder the broader application of multi-view diffusion models in downstream tasks like 3D reconstruction. We identify the root cause of inconsistency as the excessive diversity inherent in generative models utilized for the NVS task. To address this, we aim to utilize the stronger supervise information to better alignment with ground truth images to constrain the diversity, and propose Ctrl123, a **closed-loop** transcription-based multi-view diffusion method that enforces alignment in the CLIP patch feature space. Extensive experiments demonstrate that Ctrl123 excels in **arbitrary** novel view generation, significantly improving multi-view consistency compared to existing methods.", "keywords": [ "Novel view synthesis", "Diffusion Model", "Closed-Loop Transcription" ], "rating_list": [ 6, 3, 3, 5, 3 ], "soundness_list": [ 3, 1, 3, 3, 3 ], "presentation_list": [ 3, 1, 1, 2, 1 ], "contribution_list": [ 3, 1, 2, 2, 2 ], "confidence_list": [ 4, 5, 4, 4, 4 ] }, { "paper_id": "CFMdrcK935", "primary_area": "learning theory", "n_author": 2, "author_info_list": [ { "name": "Myungjoo Kang", "gender": "unknown", "institution": "Seoul National University", "country": "KR", "position": "Full Professor" }, { "name": "Seungcheol Shin", "gender": "Male", "institution": "Seoul National University", "country": "KR", "position": "PhD student" } ], "n_formula": 73, "n_formula_1": 33, "n_ref_uni": 29, "n_ref": 59, "n_ref_all": 68, "n_fig": 0, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 0, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 9814, "formula_len_all_1": 3240, "len_all": 129056, "len_all_1": 71469, "len_abs": 378, "len_title": 148, "len_sents": 35865, "len_sents_1": 27384, "n_sents": 369, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Reject", "L_title": 99, "L_abs": 446, "title": "Decomposition of one-layer neural networks via the infinite sum of reproducing kernel Banach spaces", "abs": "In this paper, we define the sum of RKBSs using the characterization theorem of RKBSs and show that the sum of RKBSs is compatible with the direct sum of feature spaces. Moreover, we decompose the integral RKBS $\\mathcal{F}\\_{\\sigma}(\\mathcal{X},\\Omega)$ into the sum of $p$-norm RKBSs $\\\\{\\mathcal{L}\\_{\\sigma}^{1}(\\mu\\_{i})\\\\}\\_{i\\in I}$. Finally, we provide some applications to enhance the structural understanding of the integral RKBS class.", "keywords": [ "Neural networks", "Reproducing kernel Banach spaces", "Class of Integral RKBSs" ], "rating_list": [ 6, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 3, 1, 3 ], "confidence_list": [ 3, 3, 2, 3 ] }, { "paper_id": "CFLEIeX7iK", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 4, "author_info_list": [ { "name": "Chao Qian", "gender": "Male", "institution": "Nanjing university", "country": "CN", "position": "Full Professor" }, { "name": "Chengrui Gao", "gender": "unknown", "institution": "nanjing university", "country": "CN", "position": "PhD student" }, { "name": "Haopu Shang", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" }, { "name": "Ke Xue", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 4, "n_ref_uni": 60, "n_ref": 116, "n_ref_all": 152, "n_fig": 8, "n_tab": 14, "L_tab": 6377, "n_element_tab": 656, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 1188, "n_element_tab_1": 97, "formula_len_all": 353, "formula_len_all_1": 203, "len_all": 213724, "len_all_1": 69322, "len_abs": 1568, "len_title": 102, "len_sents": 63391, "len_sents_1": 33209, "n_sents": 441, "n_sents_1": 238, "venue": "ICLR2025", "decision": "Reject", "L_title": 54, "L_abs": 1575, "title": "Neural Solver Selection for Combinatorial Optimization", "abs": "Machine learning has increasingly been employed to solve NP-hard combinatorial optimization problems, resulting in the emergence of neural solvers that demonstrate remarkable performance, even with minimal domain-specific knowledge. To date, the community has created numerous open-source neural solvers with distinct motivations and inductive biases. While considerable efforts are devoted to designing powerful single solvers, our findings reveal that existing solvers typically demonstrate complementary performance across different problem instances. This suggests that significant improvements could be achieved through effective coordination of neural solvers at the instance level.\nIn this work, we propose the first general framework to coordinate the neural solvers, which involves feature extraction, selection model, and selection strategy, aiming to allocate each instance to the most suitable solvers. To instantiate, we collect several typical neural solvers with state-of-the-art performance as alternatives, and explore various methods for each component of the framework. We evaluated our framework on two extensively studied combinatorial optimization problems, Traveling Salesman Problem (TSP) and Capacitated Vehicle Routing Problem (CVRP). Experimental results show that the proposed framework can effectively distribute instances and the resulting composite solver can achieve significantly better performance (e.g., reduce the optimality gap by 0.88\\% on TSPLIB and 0.71\\% on CVRPLIB) than the best individual neural solver with little extra time cost.", "keywords": [ "learn to optimize" ], "rating_list": [ 5, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 5, 4, 3, 4 ] }, { "paper_id": "CFKZKjrQ5r", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 4, "author_info_list": [ { "name": "Chinmay Mittal", "gender": "Male", "institution": "Indian Institute of Technology, Delhi", "country": "IN", "position": "Undergrad student" }, { "name": "Krishna Kartik", "gender": "unknown", "institution": "Indian Institute of Technology, Delhi", "country": "IN", "position": "PhD student" }, { "name": "Parag Singla", "gender": "Male", "institution": "Indian Institute of Technology, Delhi", "country": "IN", "position": "Full Professor" }, { "name": ". Mausam", "gender": "Male", "institution": "Indian Institute of Technology Delhi", "country": "IN", "position": "Full Professor" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 23, "n_ref": 65, "n_ref_all": 93, "n_fig": 10, "n_tab": 11, "L_tab": 4767, "n_element_tab": 596, "n_fig_1": 5, "n_tab_1": 6, "L_tab_1": 1375, "n_element_tab_1": 108, "formula_len_all": 118, "formula_len_all_1": 117, "len_all": 218962, "len_all_1": 68307, "len_abs": 1508, "len_title": 101, "len_sents": 66692, "len_sents_1": 32920, "n_sents": 517, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 101, "L_abs": 1515, "title": "FCoReBench: Can Large Language Models Solve Challenging First-Order Combinatorial Reasoning Problems?", "abs": "Can the large language models (LLMs) solve challenging first-order combinatorial\nreasoning problems such as graph coloring, knapsack, and cryptarithmetic? By\nfirst-order, we mean these problems can be instantiated with potentially an infinite\nnumber of problem instances of varying sizes. They are also challenging being\nNP-hard and requiring several reasoning steps to reach a solution. While existing\nwork has focused on coming up with datasets with hard benchmarks, there is\nlimited work which exploits the first-order nature of the problem structure. To\naddress this challenge, we present FCoReBench, a dataset of 40 such challenging\nproblems, along with scripts to generate problem instances of varying sizes and\nautomatically verify and generate their solutions. We first observe that LLMs, even\nwhen aided by symbolic solvers, perform rather poorly on our dataset, being unable\nto leverage the underlying structure of these problems. We specifically observe\na drop in performance with increasing problem size. In response, we propose a\nnew approach, SymPro-LM, which combines LLMs with both symbolic solvers\nand program interpreters, along with feedback from a few solved examples, to\nachieve huge performance gains. Our proposed approach is robust to changes in the\nproblem size, and has the unique characteristic of not requiring any LLM call during\ninference time, unlike earlier approaches. As an additional experiment, we also\ndemonstrate SymPro-LM\u2019s effectiveness on other logical reasoning benchmarks.", "keywords": [ "llms", "logical-reasoning", "first-order-reasoning", "neuro-symbolic" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2 ], "presentation_list": [ 2, 3, 1, 1 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 4, 3, 3, 2 ] }, { "paper_id": "CEvGuwMum0", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 8, "author_info_list": [ { "name": "Hongye Fu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Hui Chen", "gender": "unknown", "institution": "College of Computer Science and Technology, Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Kui Ren", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Peng Cheng", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Qinglong Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Researcher" }, { "name": "Yiqi Yang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Zhan Qin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zhongjie Ba", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 35, "n_ref": 60, "n_ref_all": 87, "n_fig": 6, "n_tab": 4, "L_tab": 1569, "n_element_tab": 182, "n_fig_1": 6, "n_tab_1": 8, "L_tab_1": 3969, "n_element_tab_1": 192, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 124994, "len_all_1": 70723, "len_abs": 1827, "len_title": 166, "len_sents": 36348, "len_sents_1": 32494, "n_sents": 255, "n_sents_1": 224, "venue": "ICLR2025", "decision": "Reject", "L_title": 118, "L_abs": 1636, "title": "JudgeRail: Harnessing Open-Source LLMs for Fast Harmful Text Detection with Judicial Prompting and Logit Rectification", "abs": "Large language models (LLMs) simultaneously facilitate the generation and detection of harmful text. Leading LLM developers, such as OpenAI, Meta, and Google, are driving a paradigm shift in the detection of harmful text, moving from conventional detectors to fine-tuned LLMs. However, these newly released models, which require substantial computational and data resources, have not yet been thoroughly investigated for their effectiveness in this new paradigm. In this work, we propose JudgeRail, a novel and generic framework that guides open-source LLMs to adhere to judicial principles during text moderation. Additionally, we introduce a new logit rectification method that accurately interprets an LLM's classification intent, rigorously controls its output format, and significantly accelerates detection. By integrating several top-performing open-source LLMs into JudgeRail without any fine-tuning and evaluating them against OpenAI Moderation API, LlamaGuard3, ShieldGemma, and other conventional moderation solutions across various datasets, including those specifically designed for jailbreaking LLMs, we demonstrate that JudgeRail can adapt these LLMs to be competitive with fine-tuned moderation models and significantly outperform conventional solutions. Moreover, we evaluate all models for detection latency, a critical yet rarely examined practical aspect, and show that LLMs with JudgeRail require only 46% to 55% of the time needed by LlamaGuard3 and ShieldGemma. The generic nature and competitive performance of JudgeRail highlight its potential for promoting the practicality of LLM-based harmful text detectors.", "keywords": [ "Large Language Model", "Harmful Text Detection", "Toxic Speech Detection", "Content Moderation" ], "rating_list": [ 5, 8, 5, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 2, 2 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 5, 3, 4, 3 ] }, { "paper_id": "CEE9cAQJ10", "primary_area": "foundation or frontier models, including LLMs", "n_author": 8, "author_info_list": [ { "name": "Hongtao Xie", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Jianjun Xu", "gender": "Male", "institution": "University of Science and Technology of China", "country": "", "position": "PhD student" }, { "name": "Jiankang Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Mengting Xing", "gender": "Female", "institution": "Yuanshi Tech.", "country": "CN", "position": "Researcher" }, { "name": "Shancheng Fang", "gender": "unknown", "institution": "Bytedance inc.", "country": "CN", "position": "Researcher" }, { "name": "Xiaorui Wang", "gender": "Male", "institution": "Kuaishou- \u5feb\u624b\u79d1\u6280", "country": "CN", "position": "Principal Researcher" }, { "name": "Yuxin Wang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Postdoc" }, { "name": "Zhineng Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 8, "n_formula_1": 5, "n_ref_uni": 33, "n_ref": 81, "n_ref_all": 106, "n_fig": 5, "n_tab": 6, "L_tab": 2317, "n_element_tab": 382, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 2984, "n_element_tab_1": 200, "formula_len_all": 274, "formula_len_all_1": 175, "len_all": 166071, "len_all_1": 63620, "len_abs": 1196, "len_title": 115, "len_sents": 47658, "len_sents_1": 26139, "n_sents": 380, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 67, "L_abs": 1218, "title": "A Graph-Based Synthetic Data Pipeline for Scaling High-Quality Data", "abs": "Synthesizing high-quality data for continual training has been proven to be effective in enhancing the performance of Large Language Models (LLMs). However, previous synthetic approaches struggle to easily scale up data and incur high costs in the pursuit of high quality. In this paper, we propose the Graph-based Synthetic Data Pipeline (GSDP), an economical and scalable framework for high-quality reasoning data synthesis. Inspired by knowledge graphs, we extracted knowledge points from seed data and constructed a knowledge point relationships graph to explore their interconnections. By exploring the implicit relationships among knowledge, our method achieves $\\times$255 data expansion. Furthermore, GSDP led by open-source models, achieves synthesis quality comparable to GPT-4-0613 while maintaining $\\times$100 lower costs. To tackle the most challenging mathematical reasoning task, we present the GSDP-MATH dataset comprising over 1.91 million pairs of math problems and answers. After fine-tuning on GSDP-MATH, GSDP-7B based on Mistral-7B achieves 37.7\\% accuracy on MATH and 78.4\\% on GSM8K, demonstrating the effectiveness of our method. The dataset and models trained in this paper will be available.", "keywords": [ "Large Language Models", "Mathematical Reasoning", "Data Synthesis" ], "rating_list": [ 6, 3, 5 ], "soundness_list": [ 2, 3, 2 ], "presentation_list": [ 3, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 5, 3 ] }, { "paper_id": "CD2wgg9RQD", "primary_area": "generative models", "n_author": 6, "author_info_list": [ { "name": "Juntao Li", "gender": "Male", "institution": "Soochow University, China", "country": "", "position": "Associate Professor" }, { "name": "Min Zhang", "gender": "Male", "institution": "Harbin Institute of Technology, Shenzhen", "country": "CN", "position": "Full Professor" }, { "name": "Wanfu Wang", "gender": "Male", "institution": "Soochow University", "country": "CN", "position": "MS student" }, { "name": "Yibin Chen", "gender": "Male", "institution": "Huawei Technologies Ltd.", "country": "", "position": "Researcher" }, { "name": "xiaobo liang", "gender": "Male", "institution": "Soochow University, China", "country": "", "position": "PhD student" }, { "name": "wenjing Xie", "gender": "Female", "institution": "Soochow University", "country": "CN", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 26, "n_ref": 50, "n_ref_all": 73, "n_fig": 10, "n_tab": 4, "L_tab": 1779, "n_element_tab": 106, "n_fig_1": 7, "n_tab_1": 5, "L_tab_1": 3398, "n_element_tab_1": 143, "formula_len_all": 83, "formula_len_all_1": 83, "len_all": 125022, "len_all_1": 60748, "len_abs": 1360, "len_title": 124, "len_sents": 35718, "len_sents_1": 26709, "n_sents": 233, "n_sents_1": 188, "venue": "ICLR2025", "decision": "withdraw", "L_title": 75, "L_abs": 1395, "title": "InfCycle: Learning to Use Tools via Inference Compute and Cycle Consistency", "abs": "The scaling of inference-time computation in large language models (LLMs) has emerged as a promising approach for enhancing reasoning capabilities by trading off inference-time and pre-training compute. \nThe practice of how to enable LLMs to utilize additional computation at test time to improve response accuracy is crucial for both academia and industry.\n\\textit{Proposer-Verifier}, as a typical paradigm of inference scaling, often fails to generalize to various scenarios. \nSpecifically, in tool use tasks, LLMs face the risk of lacking effective verifiers, leading to error accumulation in multiple reasoning steps. \nIn this work, we address these challenges by introducing \\textbf{InfCycle}, a multi-stage data synthesis strategy that employs LLMs as data synthesis and employs cycle consistency verification to ensure high-quality trajectory generation. \nThis approach utilizes step-wise cycle consistency among synthesized trajectories for a given tool, providing effective process supervision that has advantages over outcome supervision.\nExtensive experiments on multiple tool-use and reasoning tasks demonstrate that InfCycle efficiently enables self-improvement. \nIt outperforms state-of-the-art baselines on StableToolBench, achieving a 75.4\\% pass rate and a 79.6\\% win rate using small size models (7B), without relying on external supervision or expert trajectories for warm-up.", "keywords": [ "LLM", "Tool use", "Inference Scaling", "Cycle Consistency", "Self-improve" ], "rating_list": [ 5, 5, 6, 3 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 1, 3, 2 ], "contribution_list": [ 2, 1, 3, 2 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "CCoa6XgO8F", "primary_area": "interpretability and explainable AI", "n_author": 2, "author_info_list": [ { "name": "Connor Wilhelm", "gender": "Male", "institution": "Brigham Young University", "country": "US", "position": "PhD student" }, { "name": "Dan Ventura", "gender": "Male", "institution": "Brigham Young University", "country": "", "position": "Full Professor" } ], "n_formula": 1, "n_formula_1": 1, "n_ref_uni": 12, "n_ref": 25, "n_ref_all": 46, "n_fig": 12, "n_tab": 1, "L_tab": 279, "n_element_tab": 19, "n_fig_1": 11, "n_tab_1": 1, "L_tab_1": 542, "n_element_tab_1": 24, "formula_len_all": 28, "formula_len_all_1": 28, "len_all": 74643, "len_all_1": 59115, "len_abs": 982, "len_title": 68, "len_sents": 31892, "len_sents_1": 30492, "n_sents": 204, "n_sents_1": 199, "venue": "ICLR2025", "decision": "Reject", "L_title": 68, "L_abs": 994, "title": "A Defense of One-Step Learning: Examining Single-Batch Distillations", "abs": "Dataset distillation produces a compressed synthetic dataset that approximates a large dataset or other learning task. A model can be trained on a distillation in a single gradient descent step. Conventional wisdom suggests that single-step learning is not generalizable and should yield poor performance; yet, distillation defies these expectations with good approximations of full direct-task training for a large distribution of models. In order to understand how distilled datasets can perform one-shot learning, we examine the distilled data instances and the cost surfaces produced by the distilled datasets. We demonstrate that the distilled dataset not only mimics features of the true dataset but also produces cost surfaces such that one-step training leads models from the initialization space into local minima of the true task's cost surface. This shows how one-step learning's counter-intuitive success is not only reasonable but also the expected outcome of dataset distillation.", "keywords": [ "distillation", "interpretability", "explainability", "compression", "cost surface", "loss landscape" ], "rating_list": [ 3, 3, 5, 3, 5 ], "soundness_list": [ 1, 1, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 3, 2 ], "contribution_list": [ 2, 2, 2, 2, 2 ], "confidence_list": [ 4, 5, 3, 4, 4 ] }, { "paper_id": "CCUrU4A92S", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 3, "author_info_list": [ { "name": "Guilhem Fouilh\u00e9", "gender": "unknown", "institution": "IRIT", "country": "FR", "position": "PhD student" }, { "name": "Nicholas Asher", "gender": "Male", "institution": "CNRS", "country": "FR", "position": "Emeritus" }, { "name": "Omar NAIM", "gender": "Male", "institution": "IRIT", "country": "FR", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 13, "n_ref": 28, "n_ref_all": 55, "n_fig": 9, "n_tab": 5, "L_tab": 2305, "n_element_tab": 343, "n_fig_1": 10, "n_tab_1": 2, "L_tab_1": 3215, "n_element_tab_1": 121, "formula_len_all": 782, "formula_len_all_1": 552, "len_all": 95882, "len_all_1": 60634, "len_abs": 541, "len_title": 97, "len_sents": 30031, "len_sents_1": 26871, "n_sents": 200, "n_sents_1": 168, "venue": "ICLR2025", "decision": "withdraw", "L_title": 49, "L_abs": 538, "title": "Re-examining learning linear functions in context", "abs": "In context learning (ICL) is an attractive method of solving a wide range of problems. Inspired by Garg et al., we look closely at ICL in a variety of train and test settings for several transformer models of different sizes trained from scratch. Our study complements prior work by pointing out several systematic failures of these models to generalize to data not in the training distribution, thereby showing some limitations of ICL. We find that models adopt a strategy for this task that is very different from standard solutions.", "keywords": [ "In context learning", "GPT", "limitations" ], "rating_list": [ 5, 3, 1, 5 ], "soundness_list": [ 2, 2, 1, 2 ], "presentation_list": [ 3, 1, 1, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 4, 4, 4 ] }, { "paper_id": "CB2r9PwuRQ", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 9, "author_info_list": [ { "name": "Jinyang Huang", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Lecturer" }, { "name": "Meng Wang", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Full Professor" }, { "name": "Mingzheng Li", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "PhD student" }, { "name": "Weijie Feng", "gender": "Male", "institution": "Hefei University of Technology", "country": "CN", "position": "Lecturer" }, { "name": "Xiao Sun", "gender": "Male", "institution": "Hefei university of technology", "country": "CN", "position": "professor" }, { "name": "Xun Yang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Full Professor" }, { "name": "Zhi Liu", "gender": "unknown", "institution": "The University of Electro-Communications", "country": "JP", "position": "Associate Professor" }, { "name": "Zhuoer Zhao", "gender": "Female", "institution": "Anhui University", "country": "CN", "position": "MS student" }, { "name": "fengqi cui", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" } ], "n_formula": 26, "n_formula_1": 20, "n_ref_uni": 22, "n_ref": 61, "n_ref_all": 76, "n_fig": 6, "n_tab": 6, "L_tab": 2741, "n_element_tab": 97, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 811, "n_element_tab_1": 23, "formula_len_all": 1881, "formula_len_all_1": 1577, "len_all": 124396, "len_all_1": 54465, "len_abs": 1589, "len_title": 143, "len_sents": 33261, "len_sents_1": 23407, "n_sents": 252, "n_sents_1": 169, "venue": "ICLR2025", "decision": "withdraw", "L_title": 94, "L_abs": 1552, "title": "CausalESC: Breaking Causal Cycles for Emotional Support Conversations with Temporal Causal HMM", "abs": "Emotional Support Conversation (ESC) is a rapidly advancing task focused on alleviating a seeker's emotional distress. The intricate interplay between cognition, emotion, and behavior presents substantial challenges for existing approaches, which often struggle to capture the dynamic evolution of the seeker's internal state during conversations. To address this, we propose \\textbf{CausalESC}, a model designed to dynamically represent the seeker's internal states, by assuming that the generative process governing the mutual influence among these factors follows a first-order Markov property, with \\iid random variables. The model comprises a prior network, that disentangles the seeker's emotions, cognition, and behavior, and a posterior network, which decouples the support strategy factors. The prior network also models the psychological causality of the seeker within each conversation round. To account for the varying effects of support strategies on the seeker's intrinsic states, we incorporate a support intervention module to capture these impacts. Additionally, a holistic damping transfer mechanism is designed to regulate the complex interactions among cognition, emotion, behavior, and strategy, ensuring that changes remain within a reasonable range. Our model effectively breaks causal cycles and achieves causal representation learning. Both automatic and human evaluations demonstrate the effectiveness of our model, emphasizing the advantages of modeling the evolution of the seeker's internal state under support strategies.", "keywords": [ "Emotional Support Conversation", "Causal Learning", "Text Generation" ], "rating_list": [ 3, 6, 3, 5 ], "soundness_list": [ 2, 3, 3, 2 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 1, 3, 2, 2 ], "confidence_list": [ 4, 2, 4, 3 ] }, { "paper_id": "CAssIgPN4I", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Dominik Bauer", "gender": "unknown", "institution": "Columbia University", "country": "US", "position": "Postdoc" }, { "name": "Shuran Song", "gender": "Female", "institution": "Columbia University", "country": "", "position": "Assistant Professor" }, { "name": "Yijia Weng", "gender": "Female", "institution": "Stanford University", "country": "US", "position": "PhD student" }, { "name": "Mandi Zhao", "gender": "Female", "institution": "Stanford University", "country": "", "position": "PhD student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 56, "n_ref": 123, "n_ref_all": 137, "n_fig": 7, "n_tab": 2, "L_tab": 272, "n_element_tab": 19, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 2200, "n_element_tab_1": 94, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 174622, "len_all_1": 66846, "len_abs": 1302, "len_title": 110, "len_sents": 48267, "len_sents_1": 33165, "n_sents": 313, "n_sents_1": 196, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 62, "L_abs": 1097, "title": "Real2Code: Reconstruct Articulated Objects via Code Generation", "abs": "We present Real2Code, a novel approach to reconstructing articulated objects via code generation. Given visual observations of an object, we first reconstruct its part geometry using image segmentation and shape completion. We represent these object parts with oriented bounding boxes, from which a fine-tuned large language model (LLM) predicts joint articulation as code. By leveraging pre-trained vision and language models, our approach scales elegantly with the number of articulated parts, and generalizes from synthetic training data to real world objects in unstructured environments. Experimental results demonstrate that Real2Code significantly outperforms the previous state-of-the-art in terms of reconstruction accuracy, and is the first approach to extrapolate beyond objects' structural complexity in the training set, as we show for objects with up to 10 articulated parts. When incorporated with a stereo reconstruction model, Real2Code moreover generalizes to real-world objects, given only a handful of multi-view RGB images and without the need for depth or camera information.", "keywords": [ "articulated objects", "code generation LLMs", "foundation models" ], "rating_list": [ 6, 6, 6, 6, 6 ], "soundness_list": [ 3, 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 3, 2, 3 ], "contribution_list": [ 3, 2, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 4, 4 ] }, { "paper_id": "CAgIwCbnQI", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 4, "author_info_list": [ { "name": "Hanchen David Wang", "gender": "unknown", "institution": "Vanderbilt University", "country": "US", "position": "PhD student" }, { "name": "Meiyi Ma", "gender": "Female", "institution": "Vanderbilt University", "country": "US", "position": "Assistant Professor" }, { "name": "Siwoo Bae", "gender": "unknown", "institution": "Vanderbilt University", "country": "US", "position": "Undergrad student" }, { "name": "Zirong Chen", "gender": "Male", "institution": "Vanderbilt University", "country": "US", "position": "PhD student" } ], "n_formula": 41, "n_formula_1": 24, "n_ref_uni": 28, "n_ref": 40, "n_ref_all": 62, "n_fig": 10, "n_tab": 9, "L_tab": 3867, "n_element_tab": 188, "n_fig_1": 6, "n_tab_1": 1, "L_tab_1": 1167, "n_element_tab_1": 13, "formula_len_all": 2481, "formula_len_all_1": 1553, "len_all": 165233, "len_all_1": 53704, "len_abs": 1475, "len_title": 106, "len_sents": 48229, "len_sents_1": 23026, "n_sents": 348, "n_sents_1": 172, "venue": "ICLR2025", "decision": "Reject", "L_title": 57, "L_abs": 1481, "title": "Learning with Preserving for Continual Multitask Learning", "abs": "Artificial Intelligence (AI) drives advancements across fields, enabling capabilities previously unattainable. Modern intelligent systems integrate increasingly specialized tasks, such as improving tumor classification with tissue recognition or advancing driving assistance with lane detection. Typically, new tasks are addressed by training single-task models or re-training multitask models, which becomes impractical when prior data is unavailable or new data is limited. This paper introduces Continual Multitask Learning (CMTL), a novel problem category critical for future intelligent systems yet overlooked in current research. CMTL presents unique challenges beyond the scope of traditional Continual Learning (CL) and Multitask Learning (MTL). To address these challenges, we propose Learning with Preserving (LwP), a novel approach for CMTL that retains previously learned knowledge while supporting diverse tasks. LwP employs a Dynamically Weighted Distance Preservation loss function to maintain representation integrity, enabling learning across tasks without a replay buffer. We extensively evaluate LwP on three benchmark datasets across two modalities\u2014inertial measurement units of multivariate time series data for quality of exercises assessment and image datasets. Results demonstrate that LwP outperforms existing continual learning baselines, effectively mitigates catastrophic forgetting, and highlights its robustness and generalizability in CMTL scenarios.", "keywords": [ "continual learning", "continual multitask learning", "representation learning", "knowledge distillation" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 3, 1, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 5, 3, 4, 5 ] }, { "paper_id": "CA06Nqa7CG", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Devon R. Graham", "gender": "Male", "institution": ", University of British Columbia", "country": "", "position": "PhD student" }, { "name": "Kevin Leyton-Brown", "gender": "Not Specified", "institution": "University of British Columbia", "country": "CA", "position": "Full Professor" } ], "n_formula": 20, "n_formula_1": 17, "n_ref_uni": 16, "n_ref": 31, "n_ref_all": 45, "n_fig": 5, "n_tab": 2, "L_tab": 468, "n_element_tab": 39, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 717, "n_element_tab_1": 26, "formula_len_all": 1959, "formula_len_all_1": 1428, "len_all": 90730, "len_all_1": 66830, "len_abs": 1097, "len_title": 113, "len_sents": 35423, "len_sents_1": 31082, "n_sents": 289, "n_sents_1": 245, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 65, "L_abs": 1101, "title": "Utilitarian Algorithm Configuration for Infinite Parameter Spaces", "abs": "Utilitarian algorithm configuration is a general-purpose technique for automatically searching the parameter space of a given algorithm to optimize its performance, as measured by a given utility function, on a given set of inputs. Recently introduced utilitarian configuration procedures offer optimality guarantees about the returned parameterization while provably adapting to the hardness of the underlying problem. However, the applicability of these approaches is severely limited by the fact that they only search a finite, relatively small set of parameters. They cannot effectively search the configuration space of algorithms with continuous or uncountable parameters. In this paper we introduce a new procedure, which we dub COUP (Continuous, Optimistic Utilitarian Procrastination). COUP is designed to search infinite parameter spaces efficiently to find good configurations quickly. Furthermore, COUP maintains the theoretical benefits of previous utilitarian configuration procedures when applied to finite parameter spaces but is significantly faster, both provably and experimentally.", "keywords": [ "Algorithm configuration", "Utilitarian algorithm configuration", "bandits" ], "rating_list": [ 6, 8, 8, 6, 8 ], "soundness_list": [ 3, 3, 4, 3, 4 ], "presentation_list": [ 4, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 4 ], "confidence_list": [ 3, 2, 3, 4, 3 ] }, { "paper_id": "C9pndmSjg6", "primary_area": "optimization", "n_author": 1, "author_info_list": [ { "name": "Zayn Wang", "gender": "Male", "institution": "New York University", "country": "US", "position": "Undergrad student" } ], "n_formula": 11, "n_formula_1": 11, "n_ref_uni": 24, "n_ref": 33, "n_ref_all": 38, "n_fig": 1, "n_tab": 3, "L_tab": 1040, "n_element_tab": 59, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 1040, "n_element_tab_1": 59, "formula_len_all": 1585, "formula_len_all_1": 1585, "len_all": 78612, "len_all_1": 53587, "len_abs": 1998, "len_title": 168, "len_sents": 24333, "len_sents_1": 23866, "n_sents": 164, "n_sents_1": 163, "venue": "ICLR2025", "decision": "withdraw", "L_title": 118, "L_abs": 1999, "title": "Advancing Portfolio Optimization: Hybrid Relaxation and Heuristic Approaches for Cardinality-Constrained MIQP Problems", "abs": "The growing magnitude of investments in global markets has intensified the need for sophisticated risk mitigation strategies in portfolio optimization. Traditional portfolio optimization models that seek to minimize risk for a specified return frequently incorporate cardinality constraints, rendering them as Mixed-Integer Quadratic Programming (MIQP) challenges. These constraints elevate the problem to NP-Hard status, complicating the solution process. While heuristic methods have historically been favored for their direct approach to MIQP problems, relaxation techniques offer a strategic alternative by simplifying MIQP into a more tractable Quadratic Programming (QP) problem. We first introduce an approach that facilitates the conversion of MIQP to QP by relaxing integer constraints into continuous domains and integrating integer conditions into the objective function using Lagrange multipliers. This dual application not only eases the computational burden but preserves the integrity of the original problem's structure. An innovative diagonalization technique applied to the covariance matrix further refines our method, enhancing the fit for integer variables, as Lagrange multipliers are inherently biased towards continuous variables. We present a comparative analysis of three distinct models, Linear, Dual, and Diagonal, each employing a unique relaxation strategy. Our research evaluates their efficacy in addressing the MIQP problem under cardinality constraints. In conjunction with heuristic methods, the refined solutions from our exact relaxation models serve as a starting point for further refinement using Genetic Algorithm and Neighborhood Searching Algorithm. This hybrid methodology yields results that not only rival but occasionally surpass those achieved by the latest models and the commercial solver CPLEX. Our findings endorse the potential of combining exact and heuristic techniques in portfolio optimization, marking a significant advancement in the field.", "keywords": [ "portfolio optimization", "mixed-integer quadratic programming", "relaxation" ], "rating_list": [ 1, 3, 5, 3, 3 ], "soundness_list": [ 1, 1, 3, 2, 1 ], "presentation_list": [ 2, 2, 2, 2, 2 ], "contribution_list": [ 1, 1, 2, 2, 1 ], "confidence_list": [ 5, 5, 4, 5, 4 ] }, { "paper_id": "C9ju8QQSCv", "primary_area": "foundation or frontier models, including LLMs", "n_author": 6, "author_info_list": [ { "name": "Can Yang", "gender": "Male", "institution": "Hong Kong University of Science and Technology", "country": "HK", "position": "Associate Professor" }, { "name": "Tong Xiao", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "MS student" }, { "name": "Xin Xu", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" }, { "name": "Yang Wang", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "Full Professor" }, { "name": "Zhenya Huang", "gender": "Male", "institution": "University of Science and Technology of China", "country": "CN", "position": "Associate Professor" }, { "name": "Zitong Chao", "gender": "Male", "institution": "The Hong Kong University of Science and Technology", "country": "HK", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 9, "n_ref_uni": 49, "n_ref": 103, "n_ref_all": 147, "n_fig": 9, "n_tab": 10, "L_tab": 3008, "n_element_tab": 288, "n_fig_1": 5, "n_tab_1": 4, "L_tab_1": 1179, "n_element_tab_1": 134, "formula_len_all": 462, "formula_len_all_1": 554, "len_all": 240912, "len_all_1": 67230, "len_abs": 4032, "len_title": 96, "len_sents": 75245, "len_sents_1": 29890, "n_sents": 681, "n_sents_1": 231, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 46, "L_abs": 1641, "title": "Can LLMs Solve Long Math Word Problems Better?", "abs": "Math Word Problems (MWPs) play a vital role in assessing the capabilities of Large Language Models (LLMs), yet current research primarily focuses on questions with concise contexts. The impact of longer contexts on mathematical reasoning remains under-explored. This study pioneers the investigation of Context Length Generalizability (CoLeG), which refers to the ability of LLMs to solve MWPs with extended narratives. We introduce Extended Grade-School Math (E-GSM), a collection of MWPs featuring lengthy narratives, and propose two novel metrics to evaluate the efficacy and resilience of LLMs in tackling these problems. Our analysis of existing zero-shot prompting techniques with proprietary LLMs along with open-source LLMs reveals a general deficiency in CoLeG. To alleviate these issues, we propose tailored approaches for different categories of LLMs. For proprietary LLMs, we introduce a new instructional prompt designed to mitigate the impact of long contexts. For open-source LLMs, we develop a novel auxiliary task for fine-tuning to enhance CoLeG. Our comprehensive results demonstrate the effectiveness of our proposed methods, showing improved performance on E-GSM. Additionally, we conduct an in-depth analysis to differentiate the effects of semantic understanding and reasoning efficacy, showing that our methods improves the latter. We also establish the generalizability of our methods across several other MWP benchmarks. Our findings highlight the limitations of current LLMs and offer practical solutions correspondingly, paving the way for further exploration of model generalizability and training methodologies.", "keywords": [ "Large Language Models", "Math Reasoning", "Long Math Word Problems" ], "rating_list": [ 6, 5, 3, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 1, 3 ], "contribution_list": [ 3, 2, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "C9YyVygCpG", "primary_area": "learning theory", "n_author": 3, "author_info_list": [ { "name": "Shuai Li", "gender": "Female", "institution": "John Hopcroft Center, Shanghai Jiao Tong University", "country": "CN", "position": "Assistant Professor" }, { "name": "Zhiyao Zhang", "gender": "Male", "institution": "Ohio State University, Columbus", "country": "US", "position": "PhD student" }, { "name": "Zilong Wang", "gender": "unknown", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" } ], "n_formula": 57, "n_formula_1": 21, "n_ref_uni": 20, "n_ref": 62, "n_ref_all": 78, "n_fig": 6, "n_tab": 2, "L_tab": 726, "n_element_tab": 38, "n_fig_1": 4, "n_tab_1": 1, "L_tab_1": 106, "n_element_tab_1": 16, "formula_len_all": 6536, "formula_len_all_1": 2311, "len_all": 125174, "len_all_1": 72860, "len_abs": 1221, "len_title": 89, "len_sents": 41079, "len_sents_1": 31240, "n_sents": 377, "n_sents_1": 263, "venue": "ICLR2025", "decision": "Reject", "L_title": 41, "L_abs": 1270, "title": "Optimal Algorithm for Max-Min Fair Bandit", "abs": "We consider a multi-player multi-armed bandit problem (MP-MAB) where $N$ players compete for $K$ arms in $T$ rounds. The reward distribution is heterogeneous where each player has a different expected reward for the same arm. When multiple players select the same arm, they collide and obtain zero reward. In this paper, we aim to find the max-min fairness matching that maximizes the reward of the player who receives the lowest reward. This paper improves the existing regret upper bound result of $O(\\log T\\log \\log T)$ to achieve max-min fairness. More specifically, our decentralized fair elimination algorithm (DFE) deals with heterogeneity and collision carefully and attains a regret upper bounded of $O((N^2+K)\\log T / \\Delta)$, where $\\Delta$ is the minimum reward gap between max-min value and sub-optimal arms. We assume $N\\leq K$ to guarantee all players can select their arms without collisions. In addition, we also provide an $\\Omega(\\max\\{N^2, K\\} \\log T / \\Delta)$ regret lower bound for this problem. This lower bound indicates that our algorithm is optimal with respect to key parameters, which significantly improves the performance of algorithms in previous work. Numerical experiments again verify the efficiency and improvement of our algorithms.", "keywords": [ "multi-player multi-armed bandits", "max-min fariness" ], "rating_list": [ 6, 5, 6, 6 ], "soundness_list": [ 2, 3, 3, 3 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "C9DazhfVZR", "primary_area": "other topics in machine learning (i.e., none of the above)", "n_author": 5, "author_info_list": [ { "name": "Dexing Kong", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Hao Xu", "gender": "Male", "institution": "Zhejiang Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Shijie Wang", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Wenli Dai", "gender": "Female", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "xiaorong Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" } ], "n_formula": 5, "n_formula_1": 5, "n_ref_uni": 35, "n_ref": 42, "n_ref_all": 55, "n_fig": 8, "n_tab": 3, "L_tab": 392, "n_element_tab": 59, "n_fig_1": 5, "n_tab_1": 3, "L_tab_1": 392, "n_element_tab_1": 59, "formula_len_all": 582, "formula_len_all_1": 519, "len_all": 128969, "len_all_1": 61886, "len_abs": 1562, "len_title": 131, "len_sents": 40210, "len_sents_1": 28742, "n_sents": 245, "n_sents_1": 179, "venue": "ICLR2025", "decision": "withdraw", "L_title": 83, "L_abs": 1568, "title": "Generation Network for Echocardiographic Sectional Positioning and Shape Completion", "abs": "The precise localization of 2D echocardiography planes in relation to a dynamic heart necessitates specialized expertise, as existing automated algorithms prmarily classify standard views while lacking the capability for comprehensive 3D structural perception. Traditional measurement techniques have evolved to infer 3D heart geometry, yet recent advancements in artificial intelligence, though demonstrating spatial awareness, still fall short in providing explicit 3D modeling. CTA-based digital twins, while promising, are hindered by cost and radiation concerns. Echocardiography, being cost-effective and radiation-free, remains limited in its ability to provide 3D perception. To address this gap, we introduce a novel point cloud-based weakly supervised 3D generation network specifically tailored for echocardiograms. This network automates 3D heart inference, and biomarker modeling, based on 2D echocardiography, slice tracking. To further enhance accuracy, we integrated a self-supervised learning branch into our framework, introducing multi-structure reconstruction loss and an overall reconstruction loss specifically designed for cardiac structure completion. Additionally, we constructed a comparative branch that serves to bolster the network's precision in inferring cardiac structures, thereby refining our approach and elevating the fidelity of the generated 3D models. Our approach enables real-time, robust 3D heart modeling, independent of paired data requirements, thereby facilitating research advancements in echocardiographic digital twins.", "keywords": [ "Echocardiography; 3D Cardiac Modeling; Weakly Supervised Learning; Point Cloud Generation; AI-assisted Echocardiographic Analysis" ], "rating_list": [ 5, 3, 1, 5 ], "soundness_list": [ 3, 3, 1, 3 ], "presentation_list": [ 2, 2, 1, 2 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 3, 5, 4, 3 ] }, { "paper_id": "C9BA0T3xhq", "primary_area": "reinforcement learning", "n_author": 2, "author_info_list": [ { "name": "Yuwei Fu", "gender": "Male", "institution": "McGill University", "country": "CA", "position": "PhD student" }, { "name": "yifei chen", "gender": "Female", "institution": "Tencent", "country": "CN", "position": "Researcher" } ], "n_formula": 16, "n_formula_1": 16, "n_ref_uni": 8, "n_ref": 17, "n_ref_all": 26, "n_fig": 0, "n_tab": 3, "L_tab": 1098, "n_element_tab": 147, "n_fig_1": 0, "n_tab_1": 3, "L_tab_1": 1098, "n_element_tab_1": 147, "formula_len_all": 1790, "formula_len_all_1": 1790, "len_all": 56513, "len_all_1": 56513, "len_abs": 1221, "len_title": 157, "len_sents": 19575, "len_sents_1": 19575, "n_sents": 145, "n_sents_1": 145, "venue": "ICLR2025", "decision": "Reject", "L_title": 108, "L_abs": 1201, "title": "Optimizing Q-Learning Using Expectile Regression: A Dual Approach to Handle In-Sample and Out-of-Sample Data", "abs": "Offline Reinforcement Learning (RL) presents unique challenges, primarily due to the constraint of learning from a static dataset without additional environmental interaction. Traditional methods often face limitations in effectively exploiting the available data, particularly when navigating the exploration-exploitation trade-off inherent in RL. This paper introduces a novel algorithm inspired by Implicit Q-Learning, designed to extend the utility of the Bellman update to actions not explicitly present in the dataset. Our approach, termed Extended Implicit Q-Learning (EIQL), strategically incorporates actions beyond the dataset constraints by allowing selection actions with maximum Q. By doing so, it leverages the maximization capability of the Bellman update, while simultaneously mitigating error extrapolation risks. We demonstrate the efficacy of EIQL through a series of experiments that show its improved performance over traditional offline RL algorithms, particularly in environments characterized by sparse rewards or those containing suboptimal and incomplete trajectories. Our results suggest that EIQL enhances the potential of offline RL by utilizing a broader action spectrum.", "keywords": [ "reinforcement learning" ], "rating_list": [ 3, 1, 3, 1 ], "soundness_list": [ 2, 1, 1, 1 ], "presentation_list": [ 2, 2, 1, 1 ], "contribution_list": [ 2, 1, 2, 1 ], "confidence_list": [ 4, 4, 3, 5 ] }, { "paper_id": "C8niXBHjfO", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Jie Zhang", "gender": "Male", "institution": "Department of Computer Science, ETHZ - ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Yunpeng Zhao", "gender": "unknown", "institution": "National University of Singapore", "country": "SG", "position": "MS student" } ], "n_formula": 5, "n_formula_1": 4, "n_ref_uni": 38, "n_ref": 96, "n_ref_all": 126, "n_fig": 13, "n_tab": 5, "L_tab": 1011, "n_element_tab": 123, "n_fig_1": 11, "n_tab_1": 5, "L_tab_1": 915, "n_element_tab_1": 130, "formula_len_all": 300, "formula_len_all_1": 277, "len_all": 130871, "len_all_1": 51564, "len_abs": 770, "len_title": 105, "len_sents": 41461, "len_sents_1": 24266, "n_sents": 279, "n_sents_1": 154, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 56, "L_abs": 772, "title": "Does Training with Synthetic Data Truly Protect Privacy?", "abs": "As synthetic data becomes increasingly popular in machine learning tasks, numerous methods---without formal differential privacy guarantees---use synthetic data for training. These methods often claim, either explicitly or implicitly, to protect the privacy of the original training data.\nIn this work, we explore four different training paradigms: coreset selection, dataset distillation, data-free knowledge distillation, and synthetic data generated from diffusion models. While all these methods utilize synthetic data for training, they lead to vastly different conclusions regarding privacy preservation. We caution that empirical approaches to preserving data privacy require careful and rigorous evaluation; otherwise, they risk providing a false sense of privacy.", "keywords": [ "ML privacy", "membership inference" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 3 ], "presentation_list": [ 2, 3, 4, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 2, 3, 3 ] }, { "paper_id": "C8jXEugWkq", "primary_area": "applications to robotics, autonomy, planning", "n_author": 6, "author_info_list": [ { "name": "Daniel Gehrig", "gender": "Male", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "Postdoc" }, { "name": "Evangelos Chatzipantazis", "gender": "unknown", "institution": "School of Engineering and Applied Science, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Kostas Daniilidis", "gender": "Male", "institution": "Athena Research and Innovation Centre", "country": "GR", "position": "Affiliated Researcher" }, { "name": "Royina Karegoudra Jayanth", "gender": "Female", "institution": "University of Pennsylvania, University of Pennsylvania", "country": "US", "position": "MS student" }, { "name": "Yinshuang Xu", "gender": "Female", "institution": "School of Engineering and Applied Science, University of Pennsylvania", "country": "US", "position": "PhD student" }, { "name": "Ziyun Wang", "gender": "Male", "institution": "School of Engineering and Applied Science, University of Pennsylvania", "country": "US", "position": "PhD student" } ], "n_formula": 42, "n_formula_1": 14, "n_ref_uni": 53, "n_ref": 121, "n_ref_all": 164, "n_fig": 16, "n_tab": 10, "L_tab": 1330, "n_element_tab": 206, "n_fig_1": 4, "n_tab_1": 3, "L_tab_1": 1319, "n_element_tab_1": 63, "formula_len_all": 2848, "formula_len_all_1": 1025, "len_all": 276017, "len_all_1": 76870, "len_abs": 2878, "len_title": 46, "len_sents": 85293, "len_sents_1": 34901, "n_sents": 665, "n_sents_1": 243, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 46, "L_abs": 2181, "title": "EqNIO: Subequivariant Neural Inertial Odometry", "abs": "Neural network-based odometry using accelerometer and gyroscope readings from a single IMU can achieve robust, and low-drift localization capabilities, through the use of **neural displacement priors**. These priors learn to produce denoised displacement measurements but need to ignore data variations due to specific IMU mount orientation and motion directions, hindering generalization.\nThis work introduces EqNIO, which addresses this challenge with **canonical displacement priors**, i.e., priors that are invariant to the orientation of the gravity-aligned frame in which the IMU data is expressed. We train an off-the-shelf architecture with IMU measurements, that are mapped into a learnable canonical gravity-aligned frame, which is uniquely defined via three axes: the first is the gravity axis (making the frame gravity aligned), and the second and third are predicted by the network that takes IMU data as input. \nThe outputs (displacement and covariance) are mapped back to the original gravity-aligned frame. To maximize generalization, we find that these learnable frames must transform equivariantly with global trajectory rotations and reflections across the gravity direction, **i.e.** action by the roto-reflection group $O_g(3)$ which preserves gravity (a subgroup of $O(3)$). This renders the displacement prior $O(3)$-**subequivariant**.\nWe tailor specific linear, convolutional, and non-linear layers that commute with the actions of the group. \nMoreover, we introduce a bijective decomposition of angular rates into vectors that transform similarly to accelerations, allowing us to leverage both measurement types. Natively, angular rates would need to be inverted upon reflection, unlike acceleration, which hinders their joint processing.\nWe highlight EqNIO's flexibility and generalization capabilities by applying it to both filter-based (TLIO), and end-to-end (RONIN) architectures, and outperforming existing methods that use **soft** equivariance from auxiliary losses or data augmentation on the TLIO, Aria, RONIN, RIDI, and OxIOD datasets. We believe this work paves the way for low-drift and generalizable neural inertial odometry on edge devices.", "keywords": [ "equivariance", "inertial odometry", "subequivariance" ], "rating_list": [ 8, 8, 5, 6 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 3, 2 ], "confidence_list": [ 4, 2, 5, 4 ] }, { "paper_id": "C85eSjKenO", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 7, "author_info_list": [ { "name": "Cheng Luo", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "Postdoc" }, { "name": "David Pitt", "gender": "Male", "institution": "California Institute of Technology", "country": "", "position": "Research engineer" }, { "name": "Jean Kossaifi", "gender": "Male", "institution": "NVIDIA AI", "country": "", "position": "Researcher" }, { "name": "Jiawei Zhao", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Robert Joseph George", "gender": "Male", "institution": "California Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Yuandong Tian", "gender": "Male", "institution": "Meta AI (FAIR)", "country": "", "position": "Research Scientist" }, { "name": "Anima Anandkumar", "gender": "Female", "institution": "California Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 100, "n_formula_1": 21, "n_ref_uni": 19, "n_ref": 20, "n_ref_all": 35, "n_fig": 4, "n_tab": 11, "L_tab": 2959, "n_element_tab": 438, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 911, "n_element_tab_1": 160, "formula_len_all": 8061, "formula_len_all_1": 1573, "len_all": 177379, "len_all_1": 61282, "len_abs": 1030, "len_title": 122, "len_sents": 57370, "len_sents_1": 26290, "n_sents": 567, "n_sents_1": 209, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1035, "title": "Tensor-GaLore: Memory-Efficient Training via Gradient Tensor Decomposition", "abs": "We present Tensor-GaLore, a novel method for efficient training of neural networks with higher-order tensor weights. Many models, particularly those used in scientific computing, employ tensor-parameterized layers to capture complex, multidimensional relationships. When scaling these methods to high-resolution problems makes memory usage grow intractably, and matrix based optimization methods lead to suboptimal performance and compression. We propose to work directly in the high-order space of the complex tensor parameter space using a tensor factorization of the gradients during optimization. We showcase its effectiveness on Fourier Neural Operators (FNOs), a class of models crucial for solving partial differential equations (PDE) and prove the theory of it. Across various PDE tasks like the Navier Stokes and Darcy Flow equations, Tensor-GaLore achieves substantial memory savings, reducing optimizer memory usage by up to 75\\%. These substantial memory savings across AI for science demonstrate Tensor-GaLore's potential.", "keywords": [ "neural operators", "PDE", "optimization", "pre-training", "Large scale training", "AI4Science" ], "rating_list": [ 6, 5, 5, 5 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 2 ] }, { "paper_id": "C81bqFCmMf", "primary_area": "datasets and benchmarks", "n_author": 16, "author_info_list": [ { "name": "Dong Yuan", "gender": "Male", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" }, { "name": "Hao Jiang", "gender": "Male", "institution": "The Institute of Computing Technology, Chinese Academy of Sciences", "country": "CN", "position": "Associate Professor" }, { "name": "Lifeng Qiao", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Undergrad student" }, { "name": "Nanqing Dong", "gender": "unknown", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Associate Professor" }, { "name": "Pan Tan", "gender": "unknown", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "", "position": "Researcher" }, { "name": "Peng Ye", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" }, { "name": "Qianyuan Zhang", "gender": "unknown", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Siqi Sun", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Associate Professor" }, { "name": "Tao Chen", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Wanli Ouyang", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Full Professor" }, { "name": "Wenwei Han", "gender": "unknown", "institution": ", Chinese Academy of Sciences", "country": "CN", "position": "MS student" }, { "name": "Xinzhu Ma", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "CN", "position": "Postdoc" }, { "name": "Yining Tang", "gender": "unknown", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Yuchen Cai", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Intern" }, { "name": "Yuchen Ren", "gender": "Male", "institution": "University of Sydney", "country": "", "position": "PhD student" }, { "name": "bai weiqiang", "gender": "Male", "institution": "Dalian University of Technology", "country": "", "position": "MS student" } ], "n_formula": 0, "n_formula_1": 0, "n_ref_uni": 61, "n_ref": 100, "n_ref_all": 116, "n_fig": 1, "n_tab": 11, "L_tab": 8527, "n_element_tab": 1193, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 1467, "n_element_tab_1": 254, "formula_len_all": 0, "formula_len_all_1": 0, "len_all": 219884, "len_all_1": 69290, "len_abs": 3460, "len_title": 142, "len_sents": 59919, "len_sents_1": 33441, "n_sents": 444, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 94, "L_abs": 1554, "title": "COMET: Benchmark for Comprehensive Biological Multi-omics Evaluation Tasks and Language Models", "abs": "As key elements within the central dogma, DNA, RNA, and proteins play crucial roles in maintaining life by guaranteeing accurate genetic expression and implementation. Although research on these molecules has profoundly impacted fields like medicine, agriculture, and industry, the diversity of machine learning approaches\u2014from traditional statistical methods to deep learning models and large language models\u2014poses challenges for researchers in choosing the most suitable models for specific tasks, especially for cross-omics and multi-omics tasks due to the lack of comprehensive benchmarks. To address this, we introduce the first comprehensive multi-omics benchmark COMET (Benchmark for Biological **CO**mprehensive **M**ulti-omics **E**valuation **T**asks and Language Models), designed to evaluate models across single-omics, cross-omics, and multi-omics tasks. First, we curate and develop a diverse collection of downstream tasks and datasets covering key structural and functional aspects in DNA, RNA, and proteins, including tasks that span multiple omics levels. Then, we evaluate existing foundational language models for DNA, RNA, and proteins, as well as the newly proposed multi-omics method, offering valuable insights into their performance in integrating and analyzing data from different biological modalities. This benchmark aims to define critical issues in multi-omics research and guide future directions, ultimately promoting advancements in understanding biological processes through integrated and different omics data analysis.", "keywords": [ "Multi-omics Benchmark", "AI for Biology", "Language Models" ], "rating_list": [ 5, 5, 5, 8 ], "soundness_list": [ 3, 3, 2, 4 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "C7ffKahGty", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Adnane Boukhayma", "gender": "Male", "institution": "INRIA", "country": "FR", "position": "Researcher" }, { "name": "Amine Ouasfi", "gender": "unknown", "institution": "INRIA", "country": "FR", "position": "PhD student" }, { "name": "Vincent Gripon", "gender": "Male", "institution": "IMT Atlantique", "country": "FR", "position": "Associate Professor" }, { "name": "Yassir Bendou", "gender": "Male", "institution": "IMT Atlantique", "country": "FR", "position": "PhD student" } ], "n_formula": 26, "n_formula_1": 23, "n_ref_uni": 44, "n_ref": 82, "n_ref_all": 95, "n_fig": 2, "n_tab": 8, "L_tab": 2450, "n_element_tab": 375, "n_fig_1": 2, "n_tab_1": 7, "L_tab_1": 1782, "n_element_tab_1": 189, "formula_len_all": 1584, "formula_len_all_1": 1208, "len_all": 138022, "len_all_1": 54406, "len_abs": 1251, "len_title": 123, "len_sents": 28733, "len_sents_1": 22458, "n_sents": 226, "n_sents_1": 170, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1256, "title": "A Kernel Perspective on Training-Free Few-Shot Adaptation of Large Vision-Language Models", "abs": "The growing popularity of Contrastive Language-Image Pretraining (CLIP) has led to its widespread application in various visual downstream tasks. To enhance CLIP's effectiveness, efficient few-shot adaptation techniques have been widely adopted. Among these approaches, training-free methods, particularly caching methods exemplified by Tip-Adapter, have gained attention for their lightweight adaptation without the need for additional fine-tuning. In this paper, we revisit Tip-Adapter from a kernel perspective, showing that caching methods function as local adapters and are connected to a well-established kernel literature. Leveraging this insight, we offer a theoretical understanding of how these methods operate and suggest multiple avenues for enhancing over the Tip-Adapter baseline. Notably, our analysis shows the importance of incorporating global information in local adapters. Therefore, we subsequently propose a global method that learns a proximal regularizer in a reproducing kernel Hilbert space (RKHS) using CLIP as a base learner. Our method, that we call ProKeR (Proximal Kernel ridge Regression), has a closed form solution and achieves state-of-the-art performance across 11 datasets in the standard few-shot adaptation benchmark.", "keywords": [ "Few-shot Learning", "Vision-Language", "CLIP", "Efficient adaptation" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 3, 4, 3, 3 ], "presentation_list": [ 3, 3, 2, 3 ], "contribution_list": [ 3, 3, 2, 2 ], "confidence_list": [ 4, 3, 4, 4 ] }, { "paper_id": "C7XoUdJ5ZC", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 3, "author_info_list": [ { "name": "Aritra Bhaduri", "gender": "Male", "institution": "Indian Institute of Technology, Patna", "country": "IN", "position": "Undergrad student" }, { "name": "Raju Halder", "gender": "Male", "institution": "Indian Institute of Technology, Patna, Dhirubhai Ambani Institute Of Information and Communication Technology", "country": "IN", "position": "Associate Professor" }, { "name": "Sujit Chowdhury", "gender": "unknown", "institution": "Indian Institute of Technology, Patna", "country": "IN", "position": "PhD student" } ], "n_formula": 76, "n_formula_1": 16, "n_ref_uni": 17, "n_ref": 19, "n_ref_all": 29, "n_fig": 0, "n_tab": 9, "L_tab": 3625, "n_element_tab": 430, "n_fig_1": 0, "n_tab_1": 7, "L_tab_1": 1479, "n_element_tab_1": 187, "formula_len_all": 6579, "formula_len_all_1": 1157, "len_all": 128486, "len_all_1": 58876, "len_abs": 2209, "len_title": 125, "len_sents": 41371, "len_sents_1": 25946, "n_sents": 328, "n_sents_1": 185, "venue": "ICLR2025", "decision": "withdraw", "L_title": 77, "L_abs": 1475, "title": "FLAIR: FEDERATED LEARNING WITH AUGMENTED AND IMPROVED FEATURE REPRESENTATIONS", "abs": "Federated Learning (FL) enables collaborative model training across decentralized clients while preserving data privacy. However, its performance declines in challenging heterogeneous data settings. To mitigate this, existing FL frameworks not only share locally trained parameters but also exchange additional information -- such as control variates, client features, and classifier characteristics -- to address the effects of class imbalance and missing classes. However, this leads to increased communication costs and heightened risks of privacy breaches. To strike a balance between communication efficiency, privacy protection, and adaptability to heterogeneous data distributions, we propose FLAIR, a novel FL approach with augmented and improved feature representations. FLAIR utilizes Class Variational Autoencoders (CVAE) for feature augmentation, mitigating class imbalance and missing class issues. It also incorporates Reptile meta-training to facilitate knowledge transfer between model updates, adapting to dynamic feature shifts. To generalize model update, FLAIR shares only local CVAE parameters instead of local model parameters, which reduces both communication costs and privacy risks. Our experiments on benchmark datasets -- such as MNIST, CIFAR-10, CIFAR-100, and TinyImageNet -- demonstrates a significant enhancement in model convergence and accuracy compared to state-of-the-art solutions, while reducing communication overhead and privacy risks.", "keywords": [ "Federated Learning", "Class Variational Autoencoders", "Feature Augmentation", "Data Heterogeneity", "Privacy Preservation", "Communication Efficiency" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 2, 2, 2, 3 ], "presentation_list": [ 1, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 5, 5, 3 ] }, { "paper_id": "C6hUK6Q1Pi", "primary_area": "foundation or frontier models, including LLMs", "n_author": 4, "author_info_list": [ { "name": "Ahmed Hassan Awadallah", "gender": "Male", "institution": "Microsoft Research", "country": "", "position": "Principal Researcher" }, { "name": "Yadong Lu", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" }, { "name": "yelong shen", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Jianwei Yang", "gender": "Male", "institution": "Microsoft", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 0, "n_ref": 0, "n_ref_all": 11, "n_fig": 5, "n_tab": 6, "L_tab": 1952, "n_element_tab": 371, "n_fig_1": 4, "n_tab_1": 6, "L_tab_1": 1952, "n_element_tab_1": 371, "formula_len_all": 229, "formula_len_all_1": 229, "len_all": 147313, "len_all_1": 63949, "len_abs": 1604, "len_title": 90, "len_sents": 31298, "len_sents_1": 27168, "n_sents": 215, "n_sents_1": 179, "venue": "ICLR2025", "decision": "Reject", "L_title": 42, "L_abs": 1608, "title": "OmniParser for Pure Vision Based GUI Agent", "abs": "The recent advancements of large vision language models shows their great potential in driving the agent system operating on user interfaces. However, we argue that the power multimodal models like GPT-4V as a general agent on multiple operating systems across different applications is largely underestimated due to the lack of a robust screen parsing technique capable of: 1) reliably identifying interactable icons within the user interface, and 2) understanding the semantics of various elements in a screenshot and accurately associate the intended action with the corresponding region on the screen. To fill these gaps, we introduce OmniParser, a comprehensive method for parsing general user interface screenshots into structured elements, which significantly enhances the ability of GPT-4V to generate actions that can be accurately grounded in the corresponding regions of the interface. We first curated an interactable icon detection dataset using popular webpages and an icon description dataset. These datasets were utilized to fine-tune specialized models: a detection model to parse interactable regions on the screen and a caption model to extract the functional semantics of the detected elements. OmniParser significantly improves GPT-4V's performance on ScreenSpot benchmark. And on Mind2Web and AITW benchmark, OmniParser with screenshot only input outperforms the GPT-4V baselines requiring additional information outside of screenshot. We further demonstrate that OmniParser can seamlessly integrate with other vision language models, significantly enhancing their agentic capabilities.", "keywords": [ "Multimodal agent; GUI screen parsing;" ], "rating_list": [ 3, 5, 6, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 3 ], "confidence_list": [ 5, 4, 5, 5 ] }, { "paper_id": "C6d9S2lYFN", "primary_area": "datasets and benchmarks", "n_author": 8, "author_info_list": [ { "name": "Chengyi Yan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Feng Lin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Kui Ren", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Liu liu", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "PhD student" }, { "name": "Zhan Qin", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Zhixuan Chu", "gender": "Male", "institution": "Ant Group", "country": "CN", "position": "Researcher" }, { "name": "Zhongjie Ba", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Ziyue Zhan", "gender": "unknown", "institution": "Zhejiang University", "country": "", "position": "MS student" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 34, "n_ref": 158, "n_ref_all": 196, "n_fig": 18, "n_tab": 13, "L_tab": 5571, "n_element_tab": 1085, "n_fig_1": 9, "n_tab_1": 9, "L_tab_1": 3511, "n_element_tab_1": 520, "formula_len_all": 688, "formula_len_all_1": 341, "len_all": 243638, "len_all_1": 62556, "len_abs": 1692, "len_title": 102, "len_sents": 86482, "len_sents_1": 24245, "n_sents": 677, "n_sents_1": 173, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1312, "title": "A Comprehensive Deepfake Detector Assessment Platform", "abs": "The rapid development of deepfake techniques has raised serious concerns about the authenticity and integrity of digital media. To combat the potential misuse of deepfakes, it is crucial to develop reliable and robust deepfake detection algorithms. In this paper, we propose a comprehensive **D**eepfake **D**etector **A**ssessment **P**latform (**DAP**), covering six critical dimensions: benchmark performance, forgery algorithm generalization, image distortion robustness, adversarial attack resilience, forgery localization accuracy, and attribute bias. Our framework aims to provide a standardized and rigorous approach to assess the performance, generalization ability, robustness, security, localization precision, and fairness of deepfake detection algorithms. Extensive experiments are conducted on multiple public and self-built databases, considering various forgery techniques, image distortions, adversarial attacks, and attributes. The proposed framework offers insights into the strengths and limitations of state-of-the-art deepfake detection algorithms and serves as a valuable tool for researchers and practitioners to develop and evaluate novel approaches in this field. All codes, scripts, and data described in this paper are open source and available at https://github.com/tempuser4567/DAP.", "keywords": [ "deepfake detection; benchmark; evaluation" ], "rating_list": [ 5, 3, 3, 5, 3 ], "soundness_list": [ 2, 3, 2, 3, 3 ], "presentation_list": [ 2, 4, 1, 2, 3 ], "contribution_list": [ 2, 3, 2, 2, 3 ], "confidence_list": [ 4, 4, 5, 5, 5 ] }, { "paper_id": "C65Hpf02Ay", "primary_area": "generative models", "n_author": 8, "author_info_list": [ { "name": "Feng Liu", "gender": "Male", "institution": "University of Melbourne", "country": "AU", "position": "Assistant Professor" }, { "name": "Haishuai Wang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Research Professor" }, { "name": "Jiajun Bu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Full Professor" }, { "name": "Jingjun Gu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Ke Liu", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "wangzhihua", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "PhD student" }, { "name": "Bo Han", "gender": "Male", "institution": "HKBU", "country": "HK", "position": "Assistant Professor" }, { "name": "\u5f20 \u8212\u5955", "gender": "Male", "institution": "Zhejiang University", "country": "", "position": "MS student" } ], "n_formula": 12, "n_formula_1": 11, "n_ref_uni": 31, "n_ref": 47, "n_ref_all": 61, "n_fig": 5, "n_tab": 3, "L_tab": 2033, "n_element_tab": 96, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 594, "n_element_tab_1": 41, "formula_len_all": 440, "formula_len_all_1": 383, "len_all": 135117, "len_all_1": 56206, "len_abs": 1363, "len_title": 107, "len_sents": 32395, "len_sents_1": 26138, "n_sents": 210, "n_sents_1": 168, "venue": "ICLR2025", "decision": "withdraw", "L_title": 59, "L_abs": 1366, "title": "One-step Image-function Generation via Consistency Training", "abs": "Consistency models aim to deliver a U-Net generator to map noise to images directly and enable swift inference with minimal steps, even trained in isolation with consistency training mode. However, the U-Net generator requires heavy feature extraction layers for multi-level resolutions and learning convolution kernels with specific receptive fields, resulting in the challenge that consistency models suffer from heavy training resources and fail to generate images with any user-specific resolutions. In this paper, we first validate that training the original consistency model with a small batch size via consistency training mode is pretty unstable, which motivates us to investigate efficient and flexible consistency models. To this end, we propose to use a novel Transformer-based generator to generate continuous image functions, which can then be differentially rendered as images with arbitrary resolutions. We adopt implicit neural representations (INRs) to form such continuous functions, which help to decouple the resolution of generated images and the total amount of the parameters generated from the neural network. Extensive experiments on one-step image generation demonstrate that our method greatly improves the performance of consistency models with low training resources and also provides an efficient any-resolution image sampling process.", "keywords": [ "Image generation; Diffusion models; Consistency Models" ], "rating_list": [ 5, 5, 5, 6 ], "soundness_list": [ 2, 3, 2, 3 ], "presentation_list": [ 2, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 4, 3 ] }, { "paper_id": "C5w86qtcgY", "primary_area": "optimization", "n_author": 7, "author_info_list": [ { "name": "Aleksandr Beznosikov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "Associate Professor" }, { "name": "Alexander Rogozin", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "", "position": "PhD student" }, { "name": "Alexander Sholokhov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "MS student" }, { "name": "Dmitry Kovalev", "gender": "Male", "institution": "Yandex", "country": "RU", "position": "Researcher" }, { "name": "Dmitry Metelev", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "MS student" }, { "name": "Savelii Chezhegov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "Researcher" }, { "name": "Alexander Vladimirovich Gasnikov", "gender": "Male", "institution": "Moscow Institute of Physics and Technology", "country": "RU", "position": "Full Professor" } ], "n_formula": 267, "n_formula_1": 39, "n_ref_uni": 32, "n_ref": 95, "n_ref_all": 136, "n_fig": 2, "n_tab": 3, "L_tab": 695, "n_element_tab": 51, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 540, "n_element_tab_1": 6, "formula_len_all": 41370, "formula_len_all_1": 4300, "len_all": 223989, "len_all_1": 59529, "len_abs": 2797, "len_title": 64, "len_sents": 48580, "len_sents_1": 21415, "n_sents": 692, "n_sents_1": 229, "venue": "ICLR2025", "decision": "Reject", "L_title": 64, "L_abs": 916, "title": "Decentralized Finite-Sum Optimization over Time-Varying Networks", "abs": "We consider decentralized time-varying stochastic optimization problems where each of the functions held by the nodes has a finite sum structure. Such problems can be efficiently solved using variance reduction techniques. Our aim is to explore the lower complexity bounds (for communication and number of stochastic oracle calls) and find optimal algorithms. The paper studies strongly convex and nonconvex scenarios. To the best of our knowledge, variance reduced schemes and lower bounds for time-varying graphs have not been studied in the literature. For nonconvex objectives, we obtain lower bounds and develop an optimal method GT-PAGE. For strongly convex objectives, we propose the first decentralized time-varying variance-reduction method ADOM+VR and establish lower bound in this scenario, highlighting the open question of matching the algorithms complexity and lower bounds even in static network case.", "keywords": [ "convex optimization", "decentralized optimization" ], "rating_list": [ 5, 5, 3, 8 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 4, 4, 2 ] }, { "paper_id": "C53FwQZigu", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 4, "author_info_list": [ { "name": "Xiao Yu", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Undergrad student" }, { "name": "Yan Fang", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Yao Zhao", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Yunchao Wei", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Full Professor" } ], "n_formula": 11, "n_formula_1": 13, "n_ref_uni": 50, "n_ref": 95, "n_ref_all": 122, "n_fig": 8, "n_tab": 13, "L_tab": 8230, "n_element_tab": 733, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 3499, "n_element_tab_1": 289, "formula_len_all": 831, "formula_len_all_1": 861, "len_all": 173762, "len_all_1": 66236, "len_abs": 1456, "len_title": 129, "len_sents": 44088, "len_sents_1": 26235, "n_sents": 307, "n_sents_1": 185, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1296, "title": "IPSeg: Image Posterior Mitigates Semantic Drift in Class-Incremental Segmentation", "abs": "Class incremental learning aims to enable models to learn from sequential, non-stationary data streams across different tasks without catastrophic forgetting. In class incremental semantic segmentation (CISS), the semantic content of the background class changes across incremental phases, which is known as \\textbf{semantic drift}. Our research identifies two severe issues within semantic drift: separate optimization and noisy semantics, which significantly degrade CISS performance. Based on this insight, we propose a simple yet effective method, \\textbf{I}mage \\textbf{P}osterior and Semantics Decoupling for \\textbf{Seg}mentation (IPSeg), designed to address these challenges through two specific mechanisms. First, IPSeg leverages image posterior probabilities as guidance to resolve the separate optimization issue. Second, IPSeg utilizes semantics decoupling to effectively handle noisy semantics and tailor the learning strategies for different types of knowledge. Experiment results on the Pascal VOC 2012 and ADE20K datasets demonstrate superior performance compared to previous state-of-the-art approaches, particularly in more realistic and challenging long-term scenarios. Furthermore, IPSeg exhibits excellent properties in terms of both learning plasticity and memory stability.", "keywords": [ "Incremental Learning", "Semantic Segmentation" ], "rating_list": [ 5, 8, 5, 8 ], "soundness_list": [ 3, 4, 2, 3 ], "presentation_list": [ 3, 4, 2, 3 ], "contribution_list": [ 3, 4, 2, 3 ], "confidence_list": [ 4, 5, 5, 4 ] }, { "paper_id": "C4q5R6XbJ6", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 7, "author_info_list": [ { "name": "Heyang Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "MS student" }, { "name": "Hongcheng Liu", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Yanfeng Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Full Professor" }, { "name": "Yu Wang", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "Associate Professor" }, { "name": "Yuhao Wang", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "", "position": "PhD student" }, { "name": "Yusheng Liao", "gender": "Male", "institution": "Shanghai Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhiyuan Zhu", "gender": "Male", "institution": "Shanghai Jiao Tong University", "country": "CN", "position": "PhD student" } ], "n_formula": 6, "n_formula_1": 7, "n_ref_uni": 29, "n_ref": 59, "n_ref_all": 85, "n_fig": 12, "n_tab": 4, "L_tab": 2715, "n_element_tab": 273, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 902, "n_element_tab_1": 86, "formula_len_all": 564, "formula_len_all_1": 565, "len_all": 159776, "len_all_1": 58149, "len_abs": 1240, "len_title": 130, "len_sents": 46543, "len_sents_1": 27163, "n_sents": 344, "n_sents_1": 197, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 1244, "title": "Drawing the Line: Enhancing Trustworthiness of MLLMs Through the Power of Refusal", "abs": "Multimodal large language models (MLLMs) excel at multimodal perception and understanding, yet their tendency to generate hallucinated or inaccurate responses undermines their trustworthiness. Existing methods have largely overlooked the importance of refusal responses as a means of enhancing MLLMs reliability. To bridge this gap, we present the Information Boundary-aware Learning Framework (InBoL), a novel approach that empowers MLLMs to refuse to answer user queries when encountering insufficient information. To the best of our knowledge, InBoL is the first framework that systematically defines the conditions under which refusal is appropriate for MLLMs using the concept of information boundaries proposed in our paper. This framework introduces a comprehensive data generation pipeline and tailored training strategies to improve the model\u2019s ability to deliver appropriate refusal responses. To evaluate the trustworthiness of MLLMs, we further propose a user-centric alignment goal along with corresponding metrics. Experimental results demonstrate a significant improvement in refusal accuracy without noticeably compromising the model\u2019s helpfulness, establishing InBoL as a pivotal advancement in building more trustworthy MLLMs.", "keywords": [ "Trustworthiness", "Alignment", "MLLMs" ], "rating_list": [ 6, 5, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 3, 2 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 3, 3, 4 ] }, { "paper_id": "C4H45A9cZa", "primary_area": "reinforcement learning", "n_author": 5, "author_info_list": [ { "name": "Chang Chen", "gender": "unknown", "institution": "Rutgers University", "country": "US", "position": "Phd student" }, { "name": "Doojin Baek", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Hany Hamed", "gender": "Male", "institution": "Korea Advanced Institute of Science & Technology", "country": "KR", "position": "MS student" }, { "name": "Sungjin Ahn", "gender": "Male", "institution": "KAIST", "country": "US", "position": "Associate Professor" }, { "name": "Yoshua Bengio", "gender": "Male", "institution": "University of Montreal", "country": "CA", "position": "Full Professor" } ], "n_formula": 13, "n_formula_1": 7, "n_ref_uni": 38, "n_ref": 67, "n_ref_all": 85, "n_fig": 3, "n_tab": 5, "L_tab": 1936, "n_element_tab": 243, "n_fig_1": 4, "n_tab_1": 2, "L_tab_1": 908, "n_element_tab_1": 104, "formula_len_all": 1080, "formula_len_all_1": 517, "len_all": 154232, "len_all_1": 81619, "len_abs": 1086, "len_title": 117, "len_sents": 45696, "len_sents_1": 36838, "n_sents": 370, "n_sents_1": 289, "venue": "ICLR2025", "decision": "withdraw", "L_title": 69, "L_abs": 1090, "title": "Hierarchical Multiscale Diffuser for Extendable Long-Horizon Planning", "abs": "This paper introduces the Hierarchical Multiscale Diffuser (HM-Diffuser), a novel approach for efficient long-horizon planning. Building on recent advances in diffusion-based planning, our method addresses the challenge of planning over horizons significantly longer than those available in the training data. We decompose the problem into two key subproblems. The first phase, Progressive Trajectory Extension (PTE), involves stitching short trajectories together to create datasets with progressively longer trajectories. In the second phase, we train the HM-Diffuser on these extended datasets, preserving computational efficiency while enhancing long-horizon planning capabilities. The hierarchical structure of the HM-Diffuser allows for subgoal generation at multiple temporal resolutions, enabling a top-down planning approach that aligns high-level, long-term goals with low-level, short-term actions. Experimental results demonstrate that the combined PTE and HM-Diffuser approach effectively generates long-horizon plans, extending far beyond the originally provided trajectories.", "keywords": [ "Long-Horizon Planning", "Diffusion", "Hierarchical", "Multiscale" ], "rating_list": [ 5, 3, 6, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 3 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "C45YqeBDUM", "primary_area": "datasets and benchmarks", "n_author": 6, "author_info_list": [ { "name": "Fabian Gloeckle", "gender": "Male", "institution": "Ecole Nationale des Ponts et Chausees", "country": "", "position": "PhD student" }, { "name": "Gabriel Synnaeve", "gender": "Male", "institution": "Facebook", "country": "", "position": "Research Scientist" }, { "name": "Jonas Gehring", "gender": "Male", "institution": "Swiss Federal Institute of Technology", "country": "CH", "position": "PhD student" }, { "name": "Kunhao Zheng", "gender": "Male", "institution": "Facebook", "country": "FR", "position": "PhD student" }, { "name": "Ori Yoran", "gender": "Male", "institution": "Tel Aviv University", "country": "", "position": "PhD student" }, { "name": "Taco Cohen", "gender": "Male", "institution": "Meta", "country": "IT", "position": "Researcher" } ], "n_formula": 3, "n_formula_1": 2, "n_ref_uni": 54, "n_ref": 600, "n_ref_all": 653, "n_fig": 19, "n_tab": 8, "L_tab": 2170, "n_element_tab": 294, "n_fig_1": 17, "n_tab_1": 4, "L_tab_1": 1155, "n_element_tab_1": 176, "formula_len_all": 254, "formula_len_all_1": 150, "len_all": 286475, "len_all_1": 66889, "len_abs": 1474, "len_title": 99, "len_sents": 53924, "len_sents_1": 31300, "n_sents": 397, "n_sents_1": 214, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 51, "L_abs": 1500, "title": "The KoLMogorov Test: Compression by Code Generation", "abs": "Compression is at the heart of intelligence. A theoretically optimal way to compress any sequence of data is to find the shortest program that outputs that sequence and then halts. However, such Kolmogorov compression is uncomputable, and code generating LLMs struggle to approximate this theoretical ideal, as it requires reasoning, planning and search capabilities beyond those of current models. In this work, we introduce the *KoLMogorov-Test* (KT), a compression-as-intelligence intelligence test for code generation LLMs. In KT a model is presented with a sequence of data at inference time, and asked to generate the shortest program that produces the sequence. We identify several benefits of KT for both evaluation and training: an essentially infinite number of problem instances of varying difficulty is readily available, strong baselines already exist, the evaluation metric (compression) cannot be gamed, and pretraining data contamination is highly unlikely. To evaluate current models, we use audio, text, and DNA data, as well as sequences produced by random synthetic programs. Current flagship models perform poorly - both GPT4-o and Llama-3.1-405B struggle on our natural and synthetic sequences. On our synthetic distribution, we are able to train code generation models with lower compression rates than previous approaches. Moreover, we show that gains on synthetic data generalize poorly to real data, suggesting that new innovations are necessary for additional gains on KT.", "keywords": [ "Code generation", "code", "compression", "LLM", "dataset", "benchmark" ], "rating_list": [ 6, 5, 6, 10, 8 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 4 ], "contribution_list": [ 3, 3, 3, 3, 3 ], "confidence_list": [ 4, 3, 5, 4, 4 ] }, { "paper_id": "C3TrHWanh5", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Ermin Wei", "gender": "Female", "institution": "Northwestern University", "country": "", "position": "Assistant Professor" }, { "name": "Meng Zhang", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ming Tang", "gender": "Female", "institution": "Southern University of Science and Technology", "country": "CN", "position": "Associate Professor" }, { "name": "Xinbao Qiao", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "MS student" } ], "n_formula": 94, "n_formula_1": 21, "n_ref_uni": 47, "n_ref": 105, "n_ref_all": 175, "n_fig": 10, "n_tab": 10, "L_tab": 3022, "n_element_tab": 329, "n_fig_1": 1, "n_tab_1": 1, "L_tab_1": 404, "n_element_tab_1": 25, "formula_len_all": 6309, "formula_len_all_1": 1389, "len_all": 289919, "len_all_1": 74169, "len_abs": 1983, "len_title": 88, "len_sents": 93045, "len_sents_1": 33105, "n_sents": 720, "n_sents_1": 256, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 40, "L_abs": 1444, "title": "Hessian-Free Online Certified Unlearning", "abs": "Machine unlearning strives to uphold the data owners' right to be forgotten by enabling models to selectively forget specific data. \nRecent advances suggest pre-computing and storing statistics extracted from second-order information and implementing unlearning through Newton-style updates.\nHowever, the Hessian matrix operations are extremely costly and previous works conduct unlearning for empirical risk minimizer with the convexity assumption, precluding their applicability to high-dimensional over-parameterized models and the nonconvergence condition.\nIn this paper, we propose an efficient Hessian-free unlearning approach. \nThe key idea is to maintain a statistical vector for each training data, computed through affine stochastic recursion of the difference between the retrained and learned models. \nWe prove that our proposed method outperforms the state-of-the-art methods in terms of the unlearning and generalization guarantees, the deletion capacity, and the time/storage complexity, under the same regularity conditions.\nThrough the strategy of recollecting statistics for removing data, we develop an online unlearning algorithm that achieves near-instantaneous data removal, as it requires only vector addition.\nExperiments demonstrate that our proposed scheme surpasses existing results by orders of magnitude in terms of time/storage costs with millisecond-level unlearning execution, while also enhancing test accuracy.", "keywords": [ "machine unlearning; certified data removal; privacy" ], "rating_list": [ 6, 6, 6, 6 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 3, 1, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 2, 3, 3, 3 ] }, { "paper_id": "C33p2CNOQ8", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 7, "author_info_list": [ { "name": "Andrei Barbu", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Researcher" }, { "name": "Boris Katz", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Principal Research Scientist" }, { "name": "Brian Cheung", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Research Fellow" }, { "name": "Colin Conwell", "gender": "unknown", "institution": "Johns Hopkins University", "country": "US", "position": "Postdoc" }, { "name": "David Mayo", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "Vighnesh Subramaniam", "gender": "unknown", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "PhD student" }, { "name": "tomaso a poggio", "gender": "Male", "institution": "Massachusetts Institute of Technology", "country": "US", "position": "Full Professor" } ], "n_formula": 10, "n_formula_1": 1, "n_ref_uni": 62, "n_ref": 108, "n_ref_all": 144, "n_fig": 17, "n_tab": 3, "L_tab": 952, "n_element_tab": 99, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 225, "n_element_tab_1": 18, "formula_len_all": 1053, "formula_len_all_1": 63, "len_all": 213270, "len_all_1": 61546, "len_abs": 3434, "len_title": 132, "len_sents": 71950, "len_sents_1": 30310, "n_sents": 626, "n_sents_1": 246, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1813, "title": "Training the Untrainable: Introducing Inductive Bias via Representational Alignment", "abs": "We demonstrate that architectures which traditionally are considered to be ill-suited for a task can be trained using inductive biases from another architecture. Networks are considered untrainable when they overfit, underfit, or converge to poor results even when tuning their hyperparameters. For example, plain fully connected networks overfit on object recognition while deep convolutional networks without residual connections underfit. The traditional answer is to change the architecture to impose some inductive bias, although what that bias is, is unknown. We introduce guidance, where a guide network guides a target network using a neural distance function. The target is optimized to perform well and to match its internal representations, layer-by-layer, to those of the guide; the guide is unchanged. If the guide is trained, this transfers over part of the architectural prior and knowledge of the guide to the target. If the guide is untrained, this transfers over only part of the architectural prior of the guide. In this manner, we can investigate what kinds of priors different architectures place on a fully connected network. We demonstrate that this method overcomes the immediate overfitting of fully connected networks on vision tasks, makes plain CNNs competitive to ResNets, closes much of the gap between plain vanilla RNNs and Transformers, and can even help Transformers learn tasks which RNNs can perform more easily. We also discover evidence that better initializations of fully connected networks likely exist to avoid overfitting. Our method provides a mathematical tool to investigate priors and architectures, and in the long term, may demystify the dark art of architecture creation, even perhaps turning architectures into a continuous optimizable parameter of the network.", "keywords": [ "Representational alignment", "neural network optimization" ], "rating_list": [ 8, 5, 8, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 3, 2, 4, 4 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "C2uViDZmNp", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 3, "author_info_list": [ { "name": "Keng-Hou Leong", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "MS student" }, { "name": "Victor Wai Kin Chan", "gender": "Not Specified", "institution": "Shenzhen International Graduate School, Tsinghua University", "country": "CN", "position": "Full Professor" }, { "name": "Yuxuan Xiu", "gender": "Male", "institution": "Tsinghua University, Tsinghua University", "country": "CN", "position": "PhD student" } ], "n_formula": 13, "n_formula_1": 11, "n_ref_uni": 33, "n_ref": 43, "n_ref_all": 68, "n_fig": 7, "n_tab": 7, "L_tab": 1049, "n_element_tab": 120, "n_fig_1": 6, "n_tab_1": 5, "L_tab_1": 259, "n_element_tab_1": 37, "formula_len_all": 1094, "formula_len_all_1": 997, "len_all": 104827, "len_all_1": 53343, "len_abs": 861, "len_title": 121, "len_sents": 32834, "len_sents_1": 24299, "n_sents": 251, "n_sents_1": 184, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 864, "title": "Information Subtraction: Learning Representations for Conditional Entropy", "abs": "The representations of conditional entropy and conditional mutual information are significant in explaining the unique effects among variables. The previous works based on conditional contrastive sampling have successfully eliminated information about discrete sensitive variables, but have not yet addressed continuous cases. This paper introduces a framework of Information Subtraction capable of representing arbitrary information components between continuous variables. We implement a generative-based architecture that outputs such representations by simultaneously maximizing an information term and minimizing another. The results highlight the representations' ability to provide semantic features of conditional entropy. By subtracting sensitive and domain-specific information, our framework effectively enhances fair learning and domain generalization.", "keywords": [ "conditional entropy", "conditional representation learning", "self-supervised learning" ], "rating_list": [ 3, 5, 3, 3 ], "soundness_list": [ 2, 2, 3, 1 ], "presentation_list": [ 1, 2, 2, 1 ], "contribution_list": [ 2, 2, 1, 1 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "C25SgeXWjE", "primary_area": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)", "n_author": 8, "author_info_list": [ { "name": "Binyuan Hui", "gender": "Female", "institution": "Alibaba Group", "country": "", "position": "Researcher" }, { "name": "Bowen Li", "gender": "unknown", "institution": "Shanghai AI Lab", "country": "", "position": "Researcher" }, { "name": "Chengwen Qi", "gender": "Male", "institution": "Beihang University", "country": "", "position": "MS student" }, { "name": "Conghui He", "gender": "Male", "institution": "Shanghai AI Lab", "country": "CN", "position": "Researcher" }, { "name": "He Du", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "PhD student" }, { "name": "Jinwang Wu", "gender": "Male", "institution": "Beihang University", "country": "", "position": "MS student" }, { "name": "Ren Ma", "gender": "Male", "institution": "Shanghai AI Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Yuanjun Laili", "gender": "Female", "institution": "Zhongguancun Laboratory", "country": "", "position": "Associate Professor" } ], "n_formula": 3, "n_formula_1": 4, "n_ref_uni": 32, "n_ref": 53, "n_ref_all": 82, "n_fig": 11, "n_tab": 9, "L_tab": 7086, "n_element_tab": 510, "n_fig_1": 2, "n_tab_1": 6, "L_tab_1": 1631, "n_element_tab_1": 208, "formula_len_all": 243, "formula_len_all_1": 331, "len_all": 168706, "len_all_1": 66582, "len_abs": 1259, "len_title": 124, "len_sents": 50615, "len_sents_1": 30811, "n_sents": 425, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 76, "L_abs": 1187, "title": "Large Language Models Meet Symbolic Provers for Logical Reasoning Evaluation", "abs": "First-order logic (FOL) reasoning, which involves sequential deduction, is pivotal for intelligent systems and serves as a valuable task for evaluating reasoning capabilities, particularly in chain-of-thought (CoT) contexts. Existing benchmarks often rely on extensive human annotation or handcrafted templates, making it difficult to achieve the necessary complexity, scalability, and diversity for robust evaluation. To address these limitations, we propose a novel framework that synergizes the generative strengths of Large Language Models (LLMs) with the rigor and precision of symbolic provers, enabling the creation of a scalable, diverse, and high-quality FOL reasoning dataset, ProverGen. Our evaluation shows that state-of-the-art LLMs struggle to solve ProverGen problems, even with CoT prompting, highlighting the dataset\u2019s challenging nature. We also finetune Llama3.1-8B-Instruct on a separate training set generated by our framework. The finetuned model demonstrates consistent improvements on both in-distribution and out-of-distribution test sets, suggesting the value of our proposed data generation framework\\footnote{We will release our code as well as the dataset.}.", "keywords": [ "logical reasoning", "symbolic provers", "LLMs evaluation" ], "rating_list": [ 3, 8, 6, 8 ], "soundness_list": [ 1, 4, 3, 3 ], "presentation_list": [ 3, 4, 3, 4 ], "contribution_list": [ 1, 4, 3, 3 ], "confidence_list": [ 4, 4, 4, 4 ] }, { "paper_id": "C1wSR50nYf", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 3, "author_info_list": [ { "name": "Hong Cheng", "gender": "Female", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Professor" }, { "name": "Qunzhong Wang", "gender": "Male", "institution": "The Chinese University of Hong Kong", "country": "HK", "position": "Undergrad student" }, { "name": "Xiangguo Sun", "gender": "Male", "institution": "Chinese University of Hong Kong, The Chinese University of Hong Kong", "country": "HK", "position": "Postdoc" } ], "n_formula": 128, "n_formula_1": 16, "n_ref_uni": 24, "n_ref": 44, "n_ref_all": 63, "n_fig": 9, "n_tab": 1, "L_tab": 428, "n_element_tab": 124, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 290, "n_element_tab_1": 4, "formula_len_all": 8399, "formula_len_all_1": 709, "len_all": 231100, "len_all_1": 70881, "len_abs": 1694, "len_title": 126, "len_sents": 84199, "len_sents_1": 35410, "n_sents": 758, "n_sents_1": 237, "venue": "ICLR2025", "decision": "withdraw", "L_title": 78, "L_abs": 1689, "title": "Does Graph Prompt Work? A Data Operation Perspective with Theoretical Analysis", "abs": "In recent years, graph prompting has emerged as a promising research direction, enabling the learning of additional tokens or subgraphs appended to original graphs without requiring retraining of pre-trained graph models across various applications. This novel paradigm, shifting from the traditional \"pre-training and fine-tuning\" to \"pre-training and prompting,\" has shown significant empirical success in simulating graph data operations, with applications ranging from recommendation systems to biological networks and graph transferring. However, despite its potential, the theoretical underpinnings of graph prompting remain underexplored, raising critical questions about its fundamental effectiveness. The lack of rigorous theoretical proof of why and how much it works is more like a \"dark cloud\" over the graph prompting area for deeper research. To fill this gap, this paper introduces a theoretical framework that rigorously analyzes graph prompting from a data operation perspective. Our contributions are threefold: **First**, we provide a formal guarantee theorem, demonstrating graph prompts\u2019 capacity to approximate graph transformation operators, effectively linking upstream and downstream tasks. **Second**, we derive upper bounds on the error of these data operations for a single graph and extend this discussion to batches of graphs, which are common in graph model training. **Third**, we analyze the distribution of data operation errors, extending our theoretical findings from linear graph models (e.g., GCN) to non-linear graph models (e.g., GAT). Extensive experiments support our theoretical results and confirm the practical implications of these guarantees.", "keywords": [ "graph prompting", "graph neural networks" ], "rating_list": [ 3, 5, 6 ], "soundness_list": [ 2, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 3, 3 ], "confidence_list": [ 3, 3, 3 ] }, { "paper_id": "C1Wp4ubvXZ", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 2, "author_info_list": [ { "name": "Lucas Rosenblatt", "gender": "Male", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "R. Teal Witter", "gender": "unknown", "institution": "New York University", "country": "US", "position": "PhD student" } ], "n_formula": 4, "n_formula_1": 2, "n_ref_uni": 44, "n_ref": 81, "n_ref_all": 131, "n_fig": 14, "n_tab": 14, "L_tab": 15574, "n_element_tab": 1095, "n_fig_1": 12, "n_tab_1": 4, "L_tab_1": 1029, "n_element_tab_1": 83, "formula_len_all": 278, "formula_len_all_1": 214, "len_all": 196953, "len_all_1": 69318, "len_abs": 1596, "len_title": 81, "len_sents": 55535, "len_sents_1": 34693, "n_sents": 360, "n_sents_1": 221, "venue": "ICLR2025", "decision": "Reject", "L_title": 81, "L_abs": 1562, "title": "FairlyUncertain: A Comprehensive Benchmark of Uncertainty in Algorithmic Fairness", "abs": "Fair predictive algorithms hinge on both equality and trust, yet inherent uncertainty in real-world data challenges our ability to make consistent, fair, and calibrated decisions. While fairly managing predictive error has been extensively explored, some recent work has begun to address the challenge of fairly accounting for irreducible prediction uncertainty. However, a clear taxonomy and well-specified objectives for integrating uncertainty into fairness remains undefined. We address this gap by introducing FairlyUncertain, an axiomatic benchmark for evaluating uncertainty estimates in fairness. Our benchmark posits that fair predictive uncertainty estimates should be consistent across learning pipelines and calibrated to observed randomness. Through extensive experiments on 10 popular fairness datasets, our evaluation reveals: (1) A theoretically justified and simple method for estimating uncertainty in binary settings is more consistent and calibrated than prior work; (2) Abstaining from binary predictions, even with improved uncertainty estimates, reduces error but does not alleviate outcome imbalances between demographic groups; (3) Incorporating consistent and calibrated uncertainty estimates in regression tasks improves fairness without any explicit fairness interventions. Our benchmark package is designed to be extensible and open-source. By providing a standardized framework for assessing the interplay between uncertainty and fairness, FairlyUncertain paves the way for more equitable and trustworthy machine learning practices.", "keywords": [ "Heteroscedastic", "Uncertainty", "Fairness", "Benchmark" ], "rating_list": [ 5, 6, 6, 6, 5 ], "soundness_list": [ 2, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 4, 3, 3, 3, 5 ] }, { "paper_id": "C1E0Oo5qgK", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Anh-Dung Dinh", "gender": "Male", "institution": "University of Sydney, University of Sydney", "country": "AU", "position": "PhD student" }, { "name": "Daochang Liu", "gender": "Male", "institution": "University of Western Australia", "country": "AU", "position": "Lecturer" }, { "name": "Chang Xu", "gender": "unknown", "institution": "University of Sydney", "country": "AU", "position": "Associate Professor" } ], "n_formula": 35, "n_formula_1": 25, "n_ref_uni": 8, "n_ref": 20, "n_ref_all": 44, "n_fig": 9, "n_tab": 9, "L_tab": 2630, "n_element_tab": 491, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 1184, "n_element_tab_1": 228, "formula_len_all": 4273, "formula_len_all_1": 2670, "len_all": 115011, "len_all_1": 61899, "len_abs": 1020, "len_title": 99, "len_sents": 33786, "len_sents_1": 24955, "n_sents": 305, "n_sents_1": 211, "venue": "ICLR2025", "decision": "withdraw", "L_title": 51, "L_abs": 1024, "title": "Compress Guidance in Conditional Diffusion Sampling", "abs": "We found that enforcing guidance throughout the sampling process is often counterproductive due to the model-fitting issue, where samples are `tuned' to match the classifier\u2019s parameters rather than generalizing the expected condition. This work identifies and quantifies the problem, demonstrating that reducing or excluding guidance at numerous timesteps can mitigate this issue. By distributing a small amount of guidance over a large number of sampling timesteps, we observe a significant improvement in image quality and diversity while also reducing the required guidance timesteps by nearly 40\\%. This approach addresses a major challenge in applying guidance effectively to generative tasks. Consequently, our proposed method, termed Compress Guidance, allows for the exclusion of a substantial number of guidance timesteps while still surpassing baseline models in image quality. We validate our approach through benchmarks on label-conditional and text-to-image generative tasks across various datasets and models.", "keywords": [ "Diffusion model", "guidance", "generative models", "compact diffusion" ], "rating_list": [ 3, 6, 3, 3, 6 ], "soundness_list": [ 1, 3, 3, 2, 3 ], "presentation_list": [ 2, 2, 3, 1, 2 ], "contribution_list": [ 1, 3, 2, 2, 2 ], "confidence_list": [ 5, 4, 2, 4, 4 ] }, { "paper_id": "C0Ubo0XBPn", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Bin Ren", "gender": "Male", "institution": "University of Pisa", "country": "IT", "position": "PhD student" }, { "name": "Jingyun Liang", "gender": "unknown", "institution": "ETH Zurich", "country": "CH", "position": "PhD student" }, { "name": "Luca Benini", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Full Professor" }, { "name": "Mengyuan Liu", "gender": "unknown", "institution": "Peking University", "country": "CN", "position": "Assistant Professor" }, { "name": "Ming-Hsuan Yang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "Nicu Sebe", "gender": "Male", "institution": "University of Trento", "country": "IT", "position": "Full Professor" }, { "name": "Rakesh Ranjan", "gender": "Male", "institution": "Meta", "country": "US", "position": "Researcher" }, { "name": "Yawei Li", "gender": "Male", "institution": "ETHZ - ETH Zurich", "country": "CH", "position": "Lecturer" } ], "n_formula": 11, "n_formula_1": 8, "n_ref_uni": 66, "n_ref": 157, "n_ref_all": 191, "n_fig": 18, "n_tab": 17, "L_tab": 7700, "n_element_tab": 525, "n_fig_1": 13, "n_tab_1": 8, "L_tab_1": 2563, "n_element_tab_1": 214, "formula_len_all": 4787, "formula_len_all_1": 3987, "len_all": 243736, "len_all_1": 78824, "len_abs": 1228, "len_title": 122, "len_sents": 46090, "len_sents_1": 28639, "n_sents": 410, "n_sents_1": 237, "venue": "ICLR2025", "decision": "Reject", "L_title": 73, "L_abs": 1232, "title": "Hierarchical Information Flow for Generalized Efficient Image Restoration", "abs": "While vision transformers show promise in numerous image restoration (IR) tasks, the challenge remains in efficiently generalizing and scaling up a model for multiple IR tasks. To strike a balance between efficiency and model capacity for a generalized transformer-based IR method, we propose a hierarchical information flow mechanism for image restoration, dubbed Hi-IR, which progressively propagates information among pixels in a bottom-up manner. Hi-IR constructs a hierarchical information tree representing the degraded image across three levels. Each level encapsulates different types of information, with higher levels encompassing broader objects and concepts and lower levels focusing on local details. Moreover, the hierarchical tree architecture removes long-range self-attention, improves the computational efficiency and memory utilization, thus preparing it for effective model scaling. Based on that, we explore model scaling to improve our method's capabilities, which is expected to positively impact IR in large-scale training settings. Extensive experimental results show that Hi-IR achieves state-of-the-art performance in seven common image restoration tasks, affirming its effectiveness and generalizability.", "keywords": [ "Hierarchical information flow", "image restoration", "tree structure" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 3, 3, 2 ], "contribution_list": [ 2, 3, 3, 2 ], "confidence_list": [ 5, 4, 5, 4 ] }, { "paper_id": "C0HDYvGwol", "primary_area": "generative models", "n_author": 10, "author_info_list": [ { "name": "Gordon Wetzstein", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Associate Professor" }, { "name": "Hansheng Chen", "gender": "unknown", "institution": "Adobe Systems", "country": "US", "position": "Intern" }, { "name": "Hao Su", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Associate Professor" }, { "name": "Jiayuan Gu", "gender": "unknown", "institution": "ShanghaiTech University", "country": "CN", "position": "Assistant Professor" }, { "name": "Leonidas Guibas", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Full Professor" }, { "name": "Linqi Zhou", "gender": "Male", "institution": "Luma AI", "country": "US", "position": "Researcher" }, { "name": "Ruoxi Shi", "gender": "Not Specified", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "Yulin Liu", "gender": "Female", "institution": "University of California, San Diego", "country": "US", "position": "PhD student" }, { "name": "William B. Shen", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD" }, { "name": "Connor Zhizhen Lin", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" } ], "n_formula": 16, "n_formula_1": 7, "n_ref_uni": 60, "n_ref": 146, "n_ref_all": 181, "n_fig": 13, "n_tab": 14, "L_tab": 3478, "n_element_tab": 232, "n_fig_1": 10, "n_tab_1": 8, "L_tab_1": 1841, "n_element_tab_1": 177, "formula_len_all": 738, "formula_len_all_1": 350, "len_all": 191190, "len_all_1": 61415, "len_abs": 1282, "len_title": 131, "len_sents": 39181, "len_sents_1": 25884, "n_sents": 295, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1324, "title": "3D-Adapter: Geometry-Consistent Multi-View Diffusion for High-Quality 3D Generation", "abs": "Multi-view image diffusion models have significantly advanced open-domain 3D object generation. However, most existing models rely on 2D network architectures that lack inherent 3D biases, resulting in compromised geometric consistency. To address this challenge, we introduce 3D-Adapter, a plug-in module designed to infuse 3D geometry awareness into pretrained image diffusion models. Central to our approach is the idea of 3D feedback augmentation: for each denoising step in the sampling loop, 3D-Adapter decodes intermediate multi-view features into a coherent 3D representation, then re-encodes the rendered RGBD views to augment the pretrained base model through feature addition. We study two variants of 3D-Adapter: a fast feed-forward version based on Gaussian splatting and a versatile training-free version utilizing neural fields and meshes. Our extensive experiments demonstrate that 3D-Adapter not only greatly enhances the geometry quality of text-to-multi-view models such as Instant3D and Zero123++, but also enables high-quality 3D generation using the plain text-to-image Stable Diffusion. Furthermore, we showcase the broad application potential of 3D-Adapter by presenting high quality results in text-to-3D, image-to-3D, text-to-texture, and text-to-avatar tasks. Code will be made publicly available.", "keywords": [ "3D generation", "multi-view", "diffusion models", "texture generation", "radiance fields", "gaussian splatting" ], "rating_list": [ 6, 6, 5, 6, 5 ], "soundness_list": [ 3, 4, 2, 3, 2 ], "presentation_list": [ 2, 4, 2, 3, 2 ], "contribution_list": [ 3, 4, 2, 3, 3 ], "confidence_list": [ 5, 4, 4, 5, 4 ] }, { "paper_id": "C0Boqhem9u", "primary_area": "applications to neuroscience & cognitive science", "n_author": 10, "author_info_list": [ { "name": "Haiyang Sun", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "MS student" }, { "name": "Peiyang Li", "gender": "Male", "institution": "Chongqing University of Post and Telecommunications", "country": "", "position": "Associate Professor" }, { "name": "Weiwei Xing", "gender": "Female", "institution": "Beijing Jiaotong University", "country": "CN", "position": "Full Professor" }, { "name": "Xiaohui Gao", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "PhD student" }, { "name": "Xintao Hu", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "Full Professor" }, { "name": "Yifan Ren", "gender": "unknown", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "MS student" }, { "name": "Yiheng Liu", "gender": "Male", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "PhD student" }, { "name": "Yijie Niu", "gender": "Male", "institution": "Northwestern Polytechnical University, Northwest Polytechnical University Xi'an", "country": "CN", "position": "MS student" }, { "name": "Yue Cheng", "gender": "Male", "institution": "Beijing Jiaotong University", "country": "CN", "position": "PhD student" }, { "name": "Zhuoyi Li", "gender": "Female", "institution": "Northwest Polytechnical University Xi'an", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 37, "n_ref": 77, "n_ref_all": 95, "n_fig": 6, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 5, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 578, "formula_len_all_1": 579, "len_all": 110103, "len_all_1": 66574, "len_abs": 2024, "len_title": 131, "len_sents": 27816, "len_sents_1": 26059, "n_sents": 179, "n_sents_1": 165, "venue": "ICLR2025", "decision": "Reject", "L_title": 82, "L_abs": 2031, "title": "LinBridge: A Learnable Framework for Interpreting Nonlinear Neural Encoding Models", "abs": "Neural encoding of artificial neural networks (ANNs) aligns the computational representations of ANNs with brain responses, providing profound insights into the neural basis underpinning information processing in the human brain. Current neural encoding studies primarily employ linear encoding models for interpretability, despite the prevalence of nonlinear neural responses. This leads to a growing interest in developing nonlinear encoding models that retain interpretability. To address this problem, we propose LinBridge, a learnable and flexible framework based on Jacobian analysis for interpreting nonlinear encoding models. LinBridge posits that the nonlinear mapping between ANN representations and neural responses can be factorized into a linear inherent component that approximates the complex nonlinear relationship, and a mapping bias that captures sample-selective nonlinearity. The Jacobian matrix, which reflects output change rates relative to input, enables the analysis of sample-selective mapping in nonlinear models. LinBridge employs a self-supervised learning strategy to extract both the linear inherent component and nonlinear mapping biases from the Jacobian matrices of the test set, allowing it to adapt effectively to various nonlinear encoding models. We validate the LinBridge framework in the scenario of neural visual encoding, using computational visual representations from CLIP-ViT to predict brain activity recorded via functional magnetic resonance imaging (fMRI). Our experimental results demonstrate that: 1) the linear inherent component extracted by LinBridge accurately reflects the complex mappings of nonlinear neural encoding models; 2) the sample-selective mapping bias elucidates the variability of nonlinearity across different levels of the visual processing hierarchy. This study not only introduces a novel tool for interpreting nonlinear neural encoding models but also provides novel evidence regarding the distribution of hierarchical nonlinearity within the visual cortex.", "keywords": [ "Nonlinear encoding models", "Jacobian matrix", "Linear inherent component", "Mapping bias" ], "rating_list": [ 3, 8, 5, 3, 3 ], "soundness_list": [ 2, 3, 2, 2, 3 ], "presentation_list": [ 2, 2, 1, 2, 2 ], "contribution_list": [ 2, 3, 2, 2, 2 ], "confidence_list": [ 3, 4, 2, 4, 4 ] }, { "paper_id": "C06kww3Qky", "primary_area": "learning on graphs and other geometries & topologies", "n_author": 2, "author_info_list": [ { "name": "Jiashun Jin", "gender": "Male", "institution": "carnegie mellon university", "country": "US", "position": "Full Professor" }, { "name": "Jingming Wang", "gender": "Female", "institution": "University of Virginia, Charlottesville", "country": "US", "position": "Assistant Professor" } ], "n_formula": 49, "n_formula_1": 43, "n_ref_uni": 18, "n_ref": 53, "n_ref_all": 87, "n_fig": 3, "n_tab": 3, "L_tab": 512, "n_element_tab": 183, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 74, "n_element_tab_1": 2, "formula_len_all": 3946, "formula_len_all_1": 3905, "len_all": 103421, "len_all_1": 76005, "len_abs": 983, "len_title": 90, "len_sents": 33507, "len_sents_1": 30208, "n_sents": 345, "n_sents_1": 320, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 42, "L_abs": 1001, "title": "Fitting Networks with a Cancellation Trick", "abs": "The degree-corrected block model (DCBM), latent space model (LSM), and $\\beta$-model are all popular network models. We combine their modeling ideas and propose the logit-DCBM as a new model. Similar as the $\\beta$-model and LSM, the logit-DCBM contains nonlinear factors, where fitting the parameters is a challenging open problem. We resolve this problem by introducing a cancellation trick. We also propose R-SCORE as a recursive community detection algorithm, where in each iteration, we first use the idea above to update our parameter estimation, and then use the results to remove the nonlinear factors in the logit-DCBM so the renormalized model approximately satisfies a low-rank model, just like the DCBM. Our numerical study suggests that R-SCORE significantly improves over existing spectral approaches in many cases. Also, theoretically, we show that the Hamming error rate of R-SCORE is faster than that of SCORE in a specific sparse region, and is at least as fast outside this region.", "keywords": [ "Network analysis", "DCBM", "logit-DCBM", "community detection", "SCORE" ], "rating_list": [ 5, 5, 8, 5 ], "soundness_list": [ 2, 3, 4, 3 ], "presentation_list": [ 1, 2, 4, 2 ], "contribution_list": [ 2, 1, 4, 2 ], "confidence_list": [ 3, 4, 2, 3 ] }, { "paper_id": "BzvVaj78Jv", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Kang Hao Cheong", "gender": "Male", "institution": "Nanyang Technological University", "country": "SG", "position": "Associate Professor" }, { "name": "Shiqing Liu", "gender": "Female", "institution": "East China Normal University", "country": "CN", "position": "Full Professor" }, { "name": "Shiyu Hu", "gender": "Female", "institution": "Nanyang Technological University", "country": "SG", "position": "Postdoc" }, { "name": "Xuchen Li", "gender": "Male", "institution": "Institute of Automation, Chinese Academy of Sciences", "country": "CN", "position": "PhD student" }, { "name": "Yiping Ma", "gender": "Female", "institution": "East China Normal University", "country": "CN", "position": "PhD student" }, { "name": "\u6021\u6c9b \u738b", "gender": "Male", "institution": "Southeast University", "country": "CN", "position": "PhD student" } ], "n_formula": 9, "n_formula_1": 0, "n_ref_uni": 99, "n_ref": 165, "n_ref_all": 202, "n_fig": 9, "n_tab": 28, "L_tab": 19254, "n_element_tab": 1029, "n_fig_1": 9, "n_tab_1": 5, "L_tab_1": 2468, "n_element_tab_1": 219, "formula_len_all": 224, "formula_len_all_1": 0, "len_all": 582582, "len_all_1": 71022, "len_abs": 6638, "len_title": 170, "len_sents": 215706, "len_sents_1": 32683, "n_sents": 1535, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 122, "L_abs": 2018, "title": "Students Rather Than Experts: A New AI for Education Pipeline to Model More Human-like and Personalised Early Adolescences", "abs": "The capabilities of large language models (LLMs) have been applied in expert systems across various domains, providing new opportunities for AI in Education (AI4Education). Educational interactions involve a cyclical exchange between teachers and students. Current research predominantly focuses on using LLMs to simulate teachers, leveraging their expertise to enhance student learning outcomes. However, the simulation of students, which could improve teachers' instructional skills, has received insufficient attention due to the challenges of modeling and evaluating virtual students. This research poses the question: \u201cCan LLMs be utilized to develop virtual student agents that mimic human-like behavior and individual variability?\u201d Unlike expert systems focusing on knowledge delivery, virtual students must replicate learning difficulties, emotional responses, and linguistic uncertainties. These traits present significant challenges in both modeling and evaluation. To address these issues, this study focuses on language learning as a context for modeling virtual student agents. We propose a novel AI4Education framework, termed SOE (Scene - Object - Evaluation), to systematically construct LVSA (LLM-based Virtual Student Agents). By curating a dataset of personalized teacher-student interactions with various personality traits, question types, and learning stages, and fine-tuning LLMs using LoRA, we conduct multi-dimensional evaluation experiments that integrate both subjective human evaluations and objective metrics. Specifically, we: (1) develop a theoretical framework for generating LVSA; (2) integrate human subjective evaluation metrics into GPT-4 assessments, demonstrating a strong correlation between human evaluators and GPT-4 in judging LVSA authenticity; and (3) validate that LLMs can generate human-like, personalized virtual student agents in educational contexts, laying a foundation for future applications in pre-service teacher training and multi-agent simulation environments.", "keywords": [ "AI for Education; Large Language Models; LLM-based Agent; Teacher Training" ], "rating_list": [ 6, 5, 6, 3 ], "soundness_list": [ 3, 2, 3, 2 ], "presentation_list": [ 2, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 2 ], "confidence_list": [ 5, 4, 4, 4 ] }, { "paper_id": "BzsjHiBfLk", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 7, "author_info_list": [ { "name": "Kangjie Liu", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "MS student" }, { "name": "Lin-Zhuo Chen", "gender": "unknown", "institution": "Nanjing university", "country": "CN", "position": "PhD student" }, { "name": "Siyu Zhu", "gender": "Male", "institution": "Fudan University", "country": "CN", "position": "Full Professor" }, { "name": "Xun Cao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Full Professor" }, { "name": "Yao Yao", "gender": "Male", "institution": "Nanjing University", "country": "CN", "position": "Associate Professor" }, { "name": "Youtian Lin", "gender": "Male", "institution": "Nanjing university", "country": "CN", "position": "PhD student" }, { "name": "Zhihao Li", "gender": "Male", "institution": "Huawei Noah's Ark Lab", "country": "CN", "position": "Researcher" } ], "n_formula": 23, "n_formula_1": 25, "n_ref_uni": 47, "n_ref": 92, "n_ref_all": 100, "n_fig": 3, "n_tab": 4, "L_tab": 1851, "n_element_tab": 192, "n_fig_1": 2, "n_tab_1": 3, "L_tab_1": 1893, "n_element_tab_1": 195, "formula_len_all": 1571, "formula_len_all_1": 1561, "len_all": 119187, "len_all_1": 53758, "len_abs": 1201, "len_title": 134, "len_sents": 25529, "len_sents_1": 20692, "n_sents": 200, "n_sents_1": 166, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 86, "L_abs": 1205, "title": "Flow Distillation Sampling: Regularizing 3D Gaussians with Pre-trained Matching Priors", "abs": "3D Gaussian Splatting (3DGS) has achieved excellent rendering quality with fast training and rendering speed. However, its optimization process lacks explicit geometric constraints, leading to suboptimal geometric reconstruction in regions with sparse or no observational input views. In this work, we try to mitigate the issue by incorporating a pre-trained matching prior to the 3DGS optimization process. We introduce Flow Distillation Sampling (FDS), a technique that leverages pre-trained geometric knowledge to bolster the accuracy of the Gaussian radiance field. Our method employs a strategic sampling technique to target unobserved views adjacent to the input views, utilizing the optical flow calculated from the matching model (Prior Flow) to guide the flow analytically calculated from the 3DGS geometry (Radiance Flow). Comprehensive experiments in depth rendering, mesh reconstruction, and novel view synthesis showcase the significant advantages of FDS over state-of-the-art methods. Additionally, our interpretive experiments and analysis aim to shed light on the effects of FDS on geometric accuracy and rendering quality, potentially providing readers with insights into its performance.", "keywords": [ "3D Vision", "Differentiable Rendering", "3D Gaussian Splatting" ], "rating_list": [ 8, 8, 6, 5 ], "soundness_list": [ 3, 3, 3, 2 ], "presentation_list": [ 3, 3, 2, 2 ], "contribution_list": [ 3, 4, 2, 2 ], "confidence_list": [ 4, 4, 4, 5 ] }, { "paper_id": "Bzro1bgkTQ", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 5, "author_info_list": [ { "name": "Aniket Bera", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "Associate Professor" }, { "name": "Hrishikesh Viswanath", "gender": "Male", "institution": "Purdue University", "country": "US", "position": "PhD student" }, { "name": "Julius Berner", "gender": "Male", "institution": "NVIDIA", "country": "US", "position": "Researcher" }, { "name": "Peter Yichen Chen", "gender": "Male", "institution": "MIT CSAIL", "country": "", "position": "Postdoc" }, { "name": "Yue Chang", "gender": "Female", "institution": "University of Toronto", "country": "", "position": "PhD student" } ], "n_formula": 27, "n_formula_1": 12, "n_ref_uni": 25, "n_ref": 61, "n_ref_all": 95, "n_fig": 9, "n_tab": 15, "L_tab": 3099, "n_element_tab": 359, "n_fig_1": 6, "n_tab_1": 4, "L_tab_1": 886, "n_element_tab_1": 146, "formula_len_all": 1496, "formula_len_all_1": 558, "len_all": 206789, "len_all_1": 63145, "len_abs": 977, "len_title": 132, "len_sents": 47560, "len_sents_1": 27497, "n_sents": 380, "n_sents_1": 201, "venue": "ICLR2025", "decision": "Reject", "L_title": 84, "L_abs": 998, "title": "Reduced-Order Neural Operators: Learning Lagrangian Dynamics on Highly Sparse Graphs", "abs": "We propose accelerating the simulation of Lagrangian dynamics, such as fluid flows, granular flows, and elastoplasticity, with neural-operator-based reduced-order modeling. While full-order approaches simulate the physics of every particle within the system, incurring high computation time for dense inputs, we propose to simulate the physics on sparse graphs constructed by sampling from the spatially discretized system. Our discretization-invariant reduced-order framework trains on any spatial discretizations and computes temporal dynamics on any sparse sampling of these discretizations through neural operators. Our proposed approach is termed Graph Informed Optimized Reduced-Order Modeling or \\textit{GIOROM}. Through reduced order modeling, we ensure lower computation time by sparsifying the system by 6.6-32.0$\\times$, while ensuring high-fidelity full-order inference via neural fields. We show that our model generalizes to a range of initial conditions, resolutions, and materials.", "keywords": [ "Reduced order modeling", "Neural Operator", "lagrangian dynamics", "neural field", "discretization invariance" ], "rating_list": [ 6, 3, 5, 6, 6 ], "soundness_list": [ 3, 2, 3, 3, 3 ], "presentation_list": [ 3, 1, 3, 3, 2 ], "contribution_list": [ 3, 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 5, 4, 3 ] }, { "paper_id": "BzljpHVfmX", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Kyunghyun Cho", "gender": "Male", "institution": "Genentech", "country": "", "position": "Senior Director of Frontier Research" }, { "name": "Nicholas Lourie", "gender": "unknown", "institution": "New York University", "country": "", "position": "PhD student" }, { "name": "He He", "gender": "Female", "institution": "New York University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 40, "n_formula_1": 20, "n_ref_uni": 43, "n_ref": 64, "n_ref_all": 84, "n_fig": 7, "n_tab": 0, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 6, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2495, "formula_len_all_1": 810, "len_all": 163974, "len_all_1": 62310, "len_abs": 2518, "len_title": 122, "len_sents": 48800, "len_sents_1": 28777, "n_sents": 477, "n_sents_1": 274, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1852, "title": "An Asymptotic Theory of Random Search for Hyperparameters in Deep Learning", "abs": "Scale is essential in modern deep learning; however, greater scale brings a greater need to make experiments efficient. Often, most of the effort is spent finding good hyperparameters, so we should consider exactly how much to spend searching for them—unfortunately this requires a better understanding of hyperparameter search, and how it converges, than we currently have. An emerging approach to such questions is *the tuning curve*, or the test score as a function of tuning effort. In theory, the tuning curve predicts how the score will increase as search continues; in practice, current estimators use nonparametric assumptions that, while robust, can not extrapolate beyond the current search step. Such extrapolation requires stronger assumptions—realistic assumptions designed for hyperparameter tuning. Thus, we derive an asymptotic theory of random search. Its central result is a new limit theorem that explains random search in terms of four interpretable quantities: the effective number of hyperparameters, the variance due to random seeds, the concentration of probability around the optimum, and the best hyperparameters' performance. These four quantities parametrize a new probability distribution, *the noisy quadratic*, which characterizes the behavior of random search. We test our theory against three practical deep learning scenarios, including pretraining in vision and fine-tuning in language. Based on 1,024 iterations of search in each, we confirm our theory achieves excellent fit. Using the theory, we construct the first confidence bands that extrapolate the tuning curve. Moreover, once fitted, each parameter of the noisy quadratic answers an important question—such as what is the best possible performance. So others may use these tools in their research, we make them available at (URL redacted).", "keywords": [ "hyperparameters", "hyperparameter search", "hyperparameter tuning", "random search", "evaluation" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 3, 3, 2 ], "contribution_list": [ 2, 3, 2, 1 ], "confidence_list": [ 3, 3, 3, 3 ] }, { "paper_id": "BzVJOqwBka", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 5, "author_info_list": [ { "name": "Haoyu Zhang", "gender": "Male", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "PhD student" }, { "name": "Jian Luan", "gender": "Male", "institution": "Microsoft", "country": "", "position": "Principal Researcher" }, { "name": "Tianshu Yu", "gender": "Male", "institution": "Chinese University of Hong Kong (Shenzhen)", "country": "CN", "position": "Assistant Professor" }, { "name": "Wei Liu", "gender": "Male", "institution": "xiaomi", "country": "CN", "position": "Principal Researcher" }, { "name": "Xiaoying Tang", "gender": "Female", "institution": "The Chinese University of Hong Kong, Shenzhen", "country": "CN", "position": "Assistant Professor" } ], "n_formula": 14, "n_formula_1": 14, "n_ref_uni": 32, "n_ref": 98, "n_ref_all": 113, "n_fig": 5, "n_tab": 8, "L_tab": 4719, "n_element_tab": 451, "n_fig_1": 2, "n_tab_1": 4, "L_tab_1": 2553, "n_element_tab_1": 217, "formula_len_all": 681, "formula_len_all_1": 681, "len_all": 172505, "len_all_1": 66774, "len_abs": 1210, "len_title": 171, "len_sents": 41694, "len_sents_1": 28693, "n_sents": 301, "n_sents_1": 204, "venue": "ICLR2025", "decision": "Reject", "L_title": 122, "L_abs": 1217, "title": "Prompt-Guided Distillation from Multimodal Large Language Models to Task-specific Models for Multimodal Sentiment Analysis", "abs": "Multimodal Sentiment Analysis (MSA) has made some progress with the advent of Multimodal Large Language Models (MLLMs). However, the scalability and the closed-source nature of some MLLMs imposes challenges for efficient application in the real-word. In this study, we explore an innovative pathway to infuse the capabilities of general MLLMs into task-specific small models for MSA. We introduce the Prompt-Guided Multimodal Framework (PGMF), a refined teacher-student framework designed to transfer knowledge from powerful, general MLLMs to smaller, efficient models. The PGMF-Teacher utilizes MLLM-generated prompts and a tailored conditional alignment module to achieve better MSA, while the PGMF-Student distills this expertise to predict independently of MLLMs' guidance. Extensive evaluations on two popular MSA datasets including SIMS and MOSI demonstrate that compared to previous task-specific small models, PGMF-Teacher achieves state-of-the-art performance with the help of MLLMs' prompts, while PGMF-Student achieve competitive results with fewer parameters and without relying on MLLMs' prompts. The proposed framework offers a novel way to equip task-specific small models with the capability of MLLMs.", "keywords": [ "Multimodal Sentimen Analysis", "Representation Learning", "Multimodal Large Language Model", "Knowledge Distillation" ], "rating_list": [ 6, 6, 5 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 3, 3, 2 ], "confidence_list": [ 4, 5, 4 ] }, { "paper_id": "Bz9wjvToCS", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 2, "author_info_list": [ { "name": "Yanzhuo Wei", "gender": "unknown", "institution": "Zhejiang University", "country": "CN", "position": "MS student" }, { "name": "Yu Pan", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Associate Professor" } ], "n_formula": 13, "n_formula_1": 13, "n_ref_uni": 3, "n_ref": 7, "n_ref_all": 19, "n_fig": 3, "n_tab": 3, "L_tab": 2616, "n_element_tab": 50, "n_fig_1": 3, "n_tab_1": 2, "L_tab_1": 2513, "n_element_tab_1": 47, "formula_len_all": 583, "formula_len_all_1": 583, "len_all": 91983, "len_all_1": 57616, "len_abs": 1312, "len_title": 157, "len_sents": 28155, "len_sents_1": 27202, "n_sents": 175, "n_sents_1": 172, "venue": "ICLR2025", "decision": "withdraw", "L_title": 109, "L_abs": 1318, "title": "DiffDeID: a Multi-conditional Diffusion-based Method for High Fidelity Face De-indentification with Diversity", "abs": "Face de-identification is a critical task that aims to obscure true identities while preserving other facial attributes. Current methodologies typically involve disentangling identity features within a latent space and leveraging adversarial training to balance privacy with utility, often at the cost of a trade-off between two. To surmount these limitations, we introduce DiffDeID, a novel approach grounded in diffusion models. This method incrementally safeguards identity and sustains utility, all while ensuring enhanced interpretability.\nOur method employs a Latent Diffusion-based ID Sample to generate authentic identity embeddings that are obfuscated from the original identity, thereby providing users with diverse options. Additionally, a multi-condition diffusion model is utilized for facial images, ensuring the retention of image utility. We further introduce a novel training and inference paradigm, utilizing the unified architecture tailored for video facial de-identification tasks.\nThe robustness of our method is attributed to its powerful 3D prior and meticulous generation design, enabling natural identity protection, generation of high-quality details, and robustness across various attributes. Through extensive experimentation, we demonstrate that DiffDeID surpasses previous methodologies.", "keywords": [ "Face De-identification", "Data privacy", "Diffusion Model" ], "rating_list": [ 5, 5, 6, 3, 3 ], "soundness_list": [ 3, 3, 3, 2, 1 ], "presentation_list": [ 3, 3, 2, 2, 2 ], "contribution_list": [ 2, 2, 3, 2, 1 ], "confidence_list": [ 4, 4, 3, 4, 4 ] }, { "paper_id": "Bz6eAiOjrI", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 11, "author_info_list": [ { "name": "Ailing Zeng", "gender": "Female", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Chuang Gan", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "Assistant Professor" }, { "name": "Jiaben Chen", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "PhD student" }, { "name": "Koichi Saito", "gender": "Male", "institution": "Sony AI America", "country": "US", "position": "Researcher" }, { "name": "Siyuan Cen", "gender": "Male", "institution": "nanjing university", "country": "CN", "position": "Undergrad student" }, { "name": "Xueyang Yu", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "MS student" }, { "name": "Yang Fu", "gender": "Male", "institution": "Beijing Normal University", "country": "CN", "position": "MS student" }, { "name": "Yihang Chen", "gender": "Male", "institution": "Zhejiang University", "country": "CN", "position": "Undergrad student" }, { "name": "Yuki Mitsufuji", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Lead Research Scientist" }, { "name": "Zixin Wang", "gender": "Male", "institution": "University of Massachusetts at Amherst", "country": "US", "position": "MS student" }, { "name": "Julian Alexander Tanke", "gender": "Male", "institution": "Sony AI", "country": "US", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 1, "n_ref_uni": 79, "n_ref": 128, "n_ref_all": 147, "n_fig": 10, "n_tab": 5, "L_tab": 874, "n_element_tab": 24, "n_fig_1": 3, "n_tab_1": 4, "L_tab_1": 2202, "n_element_tab_1": 59, "formula_len_all": 95, "formula_len_all_1": 31, "len_all": 230731, "len_all_1": 65386, "len_abs": 1336, "len_title": 101, "len_sents": 59485, "len_sents_1": 30133, "n_sents": 435, "n_sents_1": 188, "venue": "ICLR2025", "decision": "Reject", "L_title": 53, "L_abs": 1221, "title": "Orator: LLM-Guided Multi-Shot Speech Video Generation", "abs": "In this work, we propose a novel system for automatically generating multi-shot speech videos with natural camera transitions, using input text lines and reference images from various camera angles. Existing human video generation datasets and methods are largely centered on faces or half-body single-shot videos, thus lack the capacity to produce multi-shot full-body dynamic movements from different camera angles. Recognizing the lack of suitable datasets, we first introduce TalkCuts, a large-scale dataset containing over 500 hours of human speech videos with diverse camera shots, rich 3D SMPL-X motion annotations, and camera trajectories, covering a wide range of identities. Based on this dataset, we further propose an LLM-guided multi-modal generation framework, named Orator, where the LLM serves as a multi-role director, generating detailed instructions for camera transitions, speaker gestures, and vocal delivery. This enables the system to generate coherent long-form videos through a multi-modal video generation module. Extensive experiments show that our framework successfully generates coherent and engaging multi-shot speech videos. Both the dataset and the model will be made publicly available.", "keywords": [ "Speech video generation", "Multimodal video generation", "Human video dataset", "LLM-directed human video synthesis" ], "rating_list": [ 8, 3, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 2, 2, 2, 3 ], "contribution_list": [ 3, 3, 3, 2, 3 ], "confidence_list": [ 4, 4, 4, 2, 2 ] }, { "paper_id": "BydkbNH0gj", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 6, "author_info_list": [ { "name": "Dawei Gao", "gender": "Male", "institution": "Alibaba Group", "country": "CN", "position": "Researcher" }, { "name": "Ji-Rong Wen", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Kun Zhou", "gender": "Male", "institution": "University of California, San Diego", "country": "US", "position": "Postdoc" }, { "name": "Xin Zhao", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "Full Professor" }, { "name": "Yaliang Li", "gender": "Male", "institution": "Alibaba Group", "country": "US", "position": "Staff Engineer" }, { "name": "Zikang Liu", "gender": "Male", "institution": "Renmin University of China", "country": "CN", "position": "PhD student" } ], "n_formula": 7, "n_formula_1": 7, "n_ref_uni": 38, "n_ref": 88, "n_ref_all": 111, "n_fig": 6, "n_tab": 11, "L_tab": 4122, "n_element_tab": 840, "n_fig_1": 5, "n_tab_1": 5, "L_tab_1": 3282, "n_element_tab_1": 710, "formula_len_all": 283, "formula_len_all_1": 283, "len_all": 176025, "len_all_1": 67168, "len_abs": 1484, "len_title": 117, "len_sents": 48160, "len_sents_1": 26489, "n_sents": 367, "n_sents_1": 202, "venue": "ICLR2025", "decision": "Reject", "L_title": 69, "L_abs": 1514, "title": "LESS IS MORE: HIGH-VALUE DATA SELECTION FOR VISUAL INSTRUCTION TUNING", "abs": "Visual instruction tuning is the key to building large vision language mod-\nels (LVLMs), which can greatly improve the task generalization and solving capa-\nbilities by learning a mixture of instruction data from diverse visual tasks. Previ-\nous work mostly collects multiple existing visual instruction datasets via heuristic\nways for training (even more than a million instructions), which may introduce\ndata redundancy and enlarge the training cost. To investigate this issue, we con-\nduct a series of empirical studies, which reveal a significant redundancy within the\nvisual instruction datasets, and show that greatly reducing the amount of instruc-\ntions from several tasks even do not affect the performance. Based on the findings,\nwe propose a high-value data selection approach $\\textbf{TIVE}$, to eliminate redundancy\nwithin the visual instruction data and reduce the training cost. In TIVE, we first\nestimate the instance influence score on its corresponding task, and the task dif-\nficulty score, based on the gradient-based influence functions. Then, we leverage\nthe two kinds of scores to determine the task proportion within the selected visual\ninstruction subset, and select high-value instances for each task, respectively. Ex-\nperiments on various LVLMs show that our approach using only about 15% data\ncan achieve comparable average performance to the full-data fine-tuned model\nacross eight benchmarks, even surpassing it on four of the benchmarks. Our code\nand data will be publicly released.", "keywords": [ "Visual Instruction Tuning", "Data Selection" ], "rating_list": [ 5, 5, 6, 6 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 3, 2, 3, 3 ], "confidence_list": [ 4, 3, 5, 4 ] }, { "paper_id": "ByLO7p0oCF", "primary_area": "generative models", "n_author": 3, "author_info_list": [ { "name": "Alfonso Amayuelas", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "US", "position": "PhD student" }, { "name": "Luke Yoffe", "gender": "Male", "institution": "University of California, Santa Barbara", "country": "", "position": "MS student" }, { "name": "William Yang Wang", "gender": "Male", "institution": "UC Santa Barbara", "country": "US", "position": "Full Professor" } ], "n_formula": 6, "n_formula_1": 6, "n_ref_uni": 39, "n_ref": 53, "n_ref_all": 70, "n_fig": 5, "n_tab": 2, "L_tab": 3965, "n_element_tab": 188, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 933, "n_element_tab_1": 112, "formula_len_all": 288, "formula_len_all_1": 288, "len_all": 138806, "len_all_1": 61324, "len_abs": 839, "len_title": 130, "len_sents": 37591, "len_sents_1": 29064, "n_sents": 318, "n_sents_1": 226, "venue": "ICLR2025", "decision": "withdraw", "L_title": 82, "L_abs": 842, "title": "DebUnc: Improving Large Language Model Agent Communication Via Uncertainty Metrics", "abs": "To enhance Large Language Model (LLM) capabilities, multi-agent debates have been introduced, where multiple LLMs discuss solutions to a problem over several rounds of debate. However, LLMs often produce incorrect responses that appear confident, which can mislead other agents. This is partly because agents do not express their confidence levels during standard debates. To address this, we introduce DebUnc, a multi-agent debate framework that uses uncertainty metrics to assess agent confidence levels. We adapted the LLM attention mechanism to adjust token weights based on confidence levels and also explored using textual prompts to convey confidence. Our evaluations across various benchmarks show that attention-based methods are particularly effective, and that as uncertainty metrics improve, performance will continue to increase.", "keywords": [ "multiagent debate", "model uncertainty", "agent communication", "large language models" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 4, 2, 2 ], "presentation_list": [ 3, 2, 3 ], "contribution_list": [ 2, 2, 2 ], "confidence_list": [ 5, 4, 5 ] }, { "paper_id": "ByCV9xWfNK", "primary_area": "transfer learning, meta learning, and lifelong learning", "n_author": 2, "author_info_list": [ { "name": "Arnas Uselis", "gender": "unknown", "institution": "Eberhard-Karls-Universit\u00e4t T\u00fcbingen", "country": "DE", "position": "PhD student" }, { "name": "Seong Joon Oh", "gender": "Male", "institution": "Parameter Lab", "country": "DE", "position": "Principal Researcher" } ], "n_formula": 11, "n_formula_1": 4, "n_ref_uni": 57, "n_ref": 117, "n_ref_all": 159, "n_fig": 25, "n_tab": 9, "L_tab": 6128, "n_element_tab": 268, "n_fig_1": 9, "n_tab_1": 4, "L_tab_1": 2395, "n_element_tab_1": 159, "formula_len_all": 576, "formula_len_all_1": 239, "len_all": 206350, "len_all_1": 64295, "len_abs": 1252, "len_title": 102, "len_sents": 60402, "len_sents_1": 27658, "n_sents": 480, "n_sents_1": 223, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 53, "L_abs": 1258, "title": "Intermediate Layer Classifiers for OOD generalization", "abs": "Deep classifiers are known to be sensitive to data distribution shifts, primarily due to their reliance on spurious correlations in training data. It has been suggested that these classifiers can still find useful features in the network's last layer that hold up under such shifts. In this work, we question the use of last-layer representations for out-of-distribution (OOD) generalisation and explore the utility of intermediate layers. To this end, we introduce Intermediate Layer Classifiers (ILCs). We discover that intermediate layer representations frequently offer substantially better generalisation than those from the penultimate layer. In many cases, zero-shot OOD generalisation using earlier-layer representations approaches the few-shot performance of retraining on penultimate layer representations. This is confirmed across multiple datasets, architectures, and types of distribution shifts. Our analysis suggests that intermediate layers are less sensitive to distribution shifts compared to the penultimate layer. These findings highlight the importance of understanding how information is distributed across network layers and its role in OOD generalisation, while also pointing to the limits of penultimate layer representation utility.", "keywords": [ "transfer learning", "intermediate layers", "learning dynamics", "OOD generalization" ], "rating_list": [ 6, 6, 6, 8, 6 ], "soundness_list": [ 3, 3, 3, 3, 3 ], "presentation_list": [ 3, 3, 3, 3, 3 ], "contribution_list": [ 2, 3, 3, 3, 3 ], "confidence_list": [ 3, 3, 4, 3, 4 ] }, { "paper_id": "BxQkDog4ti", "primary_area": "applications to neuroscience & cognitive science", "n_author": 9, "author_info_list": [ { "name": "Chelsea Finn", "gender": "Female", "institution": "Physical Intelligence", "country": "US", "position": "Researcher" }, { "name": "James C. R. Whittington", "gender": "unknown", "institution": "Stanford University", "country": "", "position": "Postdoc" }, { "name": "Jiajun Wu", "gender": "Male", "institution": "Stanford University", "country": "US", "position": "Assistant Professor" }, { "name": "Jin Hwa Lee", "gender": "Female", "institution": "University College London, University of London", "country": "", "position": "PhD student" }, { "name": "Kyle Hsu", "gender": "Male", "institution": "Stanford University", "country": "", "position": "PhD student" }, { "name": "Luke Hollingsworth", "gender": "Male", "institution": "University College London, University of London", "country": "GB", "position": "PhD student" }, { "name": "Peter E. Latham", "gender": "Male", "institution": "University College London", "country": "", "position": "Full Professor" }, { "name": "Timothy Edward John Behrens", "gender": "Male", "institution": "University of Oxford", "country": "", "position": "PhD student" }, { "name": "Will Dorrell", "gender": "Male", "institution": "University College London, University of London", "country": "", "position": "PhD student" } ], "n_formula": 194, "n_formula_1": 21, "n_ref_uni": 56, "n_ref": 116, "n_ref_all": 190, "n_fig": 36, "n_tab": 3, "L_tab": 71, "n_element_tab": 3, "n_fig_1": 20, "n_tab_1": 0, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 12899, "formula_len_all_1": 1512, "len_all": 350113, "len_all_1": 65249, "len_abs": 1523, "len_title": 129, "len_sents": 127617, "len_sents_1": 31837, "n_sents": 1063, "n_sents_1": 222, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 91, "L_abs": 1535, "title": "Don't Cut Corners: Exact Conditions for Modularity in Biologically Inspired Representations", "abs": "Why do biological and artificial neurons sometimes modularise, each encoding a single meaningful variable, and sometimes entangle their representation of many variables? In this work, we develop a theory of when biologically inspired networks---those that are nonnegative and energy efficient---modularise their representation of source variables (sources). We derive necessary and sufficient conditions on a sample of sources that determine whether the neurons in an optimal biologically-inspired linear autoencoder modularise. Our theory applies to any dataset, extending far beyond the case of statistical independence studied in previous work. Rather we show that sources modularise if their support is ``sufficiently spread''. From this theory, we extract and validate predictions in a variety of empirical studies on how data distribution affects modularisation in nonlinear feedforward and recurrent neural networks trained on supervised and unsupervised tasks. Furthermore, we apply these ideas to neuroscience data, showing that range independence can be used to understand the mixing or modularising of spatial and reward information in entorhinal recordings in seemingly conflicting experiments. Further, we use these results to suggest alternate origins of mixed-selectivity, beyond the predominant theory of flexible nonlinear classification. In sum, our theory prescribes precise conditions on when neural activities modularise, providing tools for inducing and elucidating modular representations in brains and machines.", "keywords": [ "neuroscience", "representation learning", "disentanglement", "modularisation", "neural networks", "hippocampus", "cortex" ], "rating_list": [ 6, 6, 6 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 3, 2, 4 ], "contribution_list": [ 3, 3, 3 ], "confidence_list": [ 3, 4, 3 ] }, { "paper_id": "BxLK1M1f8T", "primary_area": "reinforcement learning", "n_author": 7, "author_info_list": [ { "name": "Chao Gao", "gender": "Male", "institution": "Huawei Technologies Canada", "country": "CA", "position": "Researcher" }, { "name": "Hao Sheng", "gender": "Male", "institution": "Beihang University", "country": "CN", "position": "Full Professor" }, { "name": "Haogang Zhu", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "Full Professor" }, { "name": "Hongming Zhang", "gender": "Male", "institution": "University of Alberta", "country": "CA", "position": "PhD student" }, { "name": "Mengxiao Zhu", "gender": "Female", "institution": "North China University of Technology", "country": "", "position": "Lecturer" }, { "name": "Yue Pei", "gender": "Male", "institution": "Beijing University of Aeronautics and Astronautics", "country": "CN", "position": "PhD student" }, { "name": "Martin Mueller", "gender": "Male", "institution": "University of Alberta", "country": "", "position": "Full Professor" } ], "n_formula": 19, "n_formula_1": 12, "n_ref_uni": 43, "n_ref": 98, "n_ref_all": 115, "n_fig": 5, "n_tab": 5, "L_tab": 942, "n_element_tab": 119, "n_fig_1": 4, "n_tab_1": 4, "L_tab_1": 1009, "n_element_tab_1": 122, "formula_len_all": 1189, "formula_len_all_1": 664, "len_all": 129502, "len_all_1": 70438, "len_abs": 1572, "len_title": 129, "len_sents": 39767, "len_sents_1": 31752, "n_sents": 329, "n_sents_1": 255, "venue": "ICLR2025", "decision": "Reject", "L_title": 80, "L_abs": 1581, "title": "Double Check My Desired Return: Transformer with Value Validation for Offline RL", "abs": "Recently, there has been increasing interest in applying Transformers to offline reinforcement learning (RL). Existing methods typically frame offline RL as a sequence modeling problem and learn actions via Supervised learning (RvS). However, RvS-trained Transformers struggle to align actual returns with desired target returns, especially when dealing with underrepresented returns in the dataset (interpolation) or missed higher returns that could be achieved by stitching sub-optimal trajectories (extrapolation). In this work, we propose a novel method that Double Checks the Transformer with value validation for Offline RL (Doctor). Doctor integrates the strengths of supervised learning (SL) and temporal difference (TD) learning by jointly optimizing the action prediction and value function. SL stabilizes the prediction of actions conditioned on target returns, while TD learning adds stitching capability to the Transformer. During inference, we introduce a double-check mechanism. We sample actions around desired target returns and validate them with value functions. This mechanism ensures better alignment between the predicted action and the desired target return and is beneficial for further online exploration and fine-tuning. We evaluate Doctor on the D4RL benchmark in both offline and offline-to-online settings, demonstrating that Doctor does much better in return alignment, either within the dataset or beyond the dataset. Furthermore, Doctor performs on par with or outperforms existing RvS-based and TD-based offline RL methods on the final performance.", "keywords": [ "Offline Reinforcement Learning", "Transformer" ], "rating_list": [ 6, 3, 5, 6 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 3, 3, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 4, 1, 3 ] }, { "paper_id": "BxBt8WLfqE", "primary_area": "optimization", "n_author": 3, "author_info_list": [ { "name": "Christian Piermarini", "gender": "Male", "institution": "University of Roma \"La Sapienza\"", "country": "", "position": "PhD student" }, { "name": "Frank Edward Curtis", "gender": "Male", "institution": "Lehigh University", "country": "US", "position": "Full Professor" }, { "name": "Qi Wang", "gender": "Female", "institution": "Lehigh University", "country": "US", "position": "PhD student" } ], "n_formula": 25, "n_formula_1": 18, "n_ref_uni": 24, "n_ref": 56, "n_ref_all": 87, "n_fig": 16, "n_tab": 1, "L_tab": 272, "n_element_tab": 30, "n_fig_1": 7, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2860, "formula_len_all_1": 2608, "len_all": 133899, "len_all_1": 70462, "len_abs": 1201, "len_title": 151, "len_sents": 43625, "len_sents_1": 31859, "n_sents": 306, "n_sents_1": 205, "venue": "ICLR2025", "decision": "Reject", "L_title": 103, "L_abs": 1211, "title": "Informed Machine Learning with a Stochastic-Gradient-based Algorithm for Training with Hard Constraints", "abs": "A methodology for informed machine learning is presented and its effectiveness is shown through numerical experiments with physics-informed learning problems. The methodology has three main distinguishing features. Firstly, prior information is introduced in the training problem through hard constraints rather than through the typical modern practice of using soft constraints (i.e., regularization terms). Secondly, the methodology does not employ penalty-based (e.g., augmented Lagrangian) methods since the use of such methods results in an overall methodology that is similar to a soft-constrained approach. Rather, the methodology is based on a recently proposed stochastic-gradient-based algorithm that maintains computationally efficiency while handling constraints with a Newton-based technique. Thirdly, a new projection-based variant of the well-known Adam optimization methodology is proposed for settings with hard constraints. Numerical experiments on a set of physics-informed learning problems show that, when compared with a soft-constraint approach, the proposed methodology can be easier to tune, lead to accurate predictions more quickly, and lead to better final prediction accuracy.", "keywords": [ "nonlinear optimization", "stochastic gradient methods", "constrained optimization", "physics-informed learning" ], "rating_list": [ 5, 6, 5, 5 ], "soundness_list": [ 3, 3, 2, 3 ], "presentation_list": [ 2, 3, 2, 3 ], "contribution_list": [ 2, 3, 2, 2 ], "confidence_list": [ 3, 2, 4, 3 ] }, { "paper_id": "Bx5kcMkb8l", "primary_area": "applications to physical sciences (physics, chemistry, biology, etc.)", "n_author": 6, "author_info_list": [ { "name": "Hongyi Luo", "gender": "Male", "institution": "Technische Universit\u00e4t M\u00fcnchen", "country": "DE", "position": "PhD student" }, { "name": "Lei Zhang", "gender": "unknown", "institution": "Anhui Medical University", "country": "CN", "position": "PhD student" }, { "name": "Peng Zhu", "gender": "Male", "institution": "Anhui Medical University", "country": "CN", "position": "PI" }, { "name": "Ping Zu", "gender": "Female", "institution": "Anhui Medical University", "country": "CN", "position": "PhD student" }, { "name": "Xiangdong Wang", "gender": "unknown", "institution": "Anhui Medical University", "country": "CN", "position": "PhD student" }, { "name": "Xuehai Wang", "gender": "Male", "institution": "Anhui Medical University", "country": "CN", "position": "Researcher" } ], "n_formula": 8, "n_formula_1": 8, "n_ref_uni": 22, "n_ref": 38, "n_ref_all": 51, "n_fig": 6, "n_tab": 2, "L_tab": 930, "n_element_tab": 148, "n_fig_1": 2, "n_tab_1": 2, "L_tab_1": 915, "n_element_tab_1": 148, "formula_len_all": 510, "formula_len_all_1": 510, "len_all": 172721, "len_all_1": 60500, "len_abs": 1617, "len_title": 137, "len_sents": 32264, "len_sents_1": 29335, "n_sents": 202, "n_sents_1": 180, "venue": "ICLR2025", "decision": "withdraw", "L_title": 89, "L_abs": 1625, "title": "No Factor Left Behind: Towards arbitrary amount of factors in the medical cohort analysis", "abs": "Driven by the goal of data-driven analysis on the large-scale cohort, a large language model(LLM) has solidified itself as a critical focus of artificial intelligence medical research today. However, such efforts have coalesced around a small group of evidence, leaving behind the vast majority of factors collected in the cohort investigation. What does it take to break the more than 70 factors while ensuring responsible, high-quality prediction, all while keeping medical considerations in mind? In No Factor Left Behind, we first took on this challenge by numerical interpretable evidence contextualizing the need for Premature rupture of membranes (PROM) risk assessment through exploratory interviews with domain experts. Then, we created datasets and models aimed at narrowing the performance gap between low and high-frequency factors. More specifically, we developed a model based on factor-value pairs trained on data obtained with robust and effective data mining techniques tailored for low-frequency factors. We propose multiple architectural and training improvements to counteract overfitting while training on 70 factors. Critically, we interpreted the risk of PROM over 7000 cohort participants' directions using numerical interpretable evidence with precise values of factors combined with human evaluation covering all factors in the dataset to assess medical safety. Our model achieves a performance of 79\\% accuracy (78 factors) and 96\\% accuracy(40 factors) with risk assessment at the screening level, laying the novel insight for realizing a general medical cohort analysis method in the era of LLMs.", "keywords": [ "Medical cohort analysis", "Risk assessment", "generalization", "prompt engineering", "open source model" ], "rating_list": [ 3, 3, 3 ], "soundness_list": [ 2, 2, 2 ], "presentation_list": [ 2, 1, 1 ], "contribution_list": [ 3, 2, 2 ], "confidence_list": [ 3, 2, 3 ] }, { "paper_id": "BwlEfAhUVX", "primary_area": "applications to computer vision, audio, language, and other modalities", "n_author": 8, "author_info_list": [ { "name": "Chen Li", "gender": "Not Specified", "institution": "Tencent ARC Lab", "country": "CN", "position": "Researcher" }, { "name": "Jinguo Zhu", "gender": "Male", "institution": "Shanghai Artificial Intelligence Laboratory", "country": "CN", "position": "Researcher" }, { "name": "Kun Yi", "gender": "Male", "institution": "Tencent ARC Lab", "country": "", "position": "Researcher" }, { "name": "Lin Song", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Sijie Zhao", "gender": "Male", "institution": "Tencent AI Lab", "country": "CN", "position": "Researcher" }, { "name": "Ying Shan", "gender": "Male", "institution": "Tencent AI Lab Center of Visual Computing", "country": "CN", "position": "Director" }, { "name": "Yixiao Ge", "gender": "Female", "institution": "Tencent", "country": "CN", "position": "Researcher" }, { "name": "Yuying Ge", "gender": "Female", "institution": "Tencent ARC Lab", "country": "CN", "position": "Researcher" } ], "n_formula": 2, "n_formula_1": 2, "n_ref_uni": 21, "n_ref": 63, "n_ref_all": 81, "n_fig": 10, "n_tab": 2, "L_tab": 579, "n_element_tab": 136, "n_fig_1": 8, "n_tab_1": 2, "L_tab_1": 684, "n_element_tab_1": 150, "formula_len_all": 615, "formula_len_all_1": 595, "len_all": 164583, "len_all_1": 57589, "len_abs": 152, "len_title": 87, "len_sents": 40335, "len_sents_1": 25475, "n_sents": 309, "n_sents_1": 197, "venue": "ICLR2025", "decision": "Reject", "L_title": 39, "L_abs": 1339, "title": "SEED-X: Multimodal Models in Real World", "abs": "The rapid evolution of multimodal foundation models has showcased remarkable capabilities in vision-language understanding and generation, yielding impressive results on academic benchmarks. However, there remains a gap in their progress toward real-world applicability, primarily due to the models' limited capacity to effectively respond to various user instructions and interact with diverse visual data. This limitation can be attributed to the fundamental challenge of modeling multi-granularity visual semantics for comprehension and generation tasks. In this paper, we take a pioneering step towards applying multimodal foundation models in an open-world context and present a unified and versatile foundation model, namely, $\\textbf{SEED-X}$. As the first of its kind, SEED-X seamlessly integrates two essential features: (1) comprehending images of arbitrary sizes and ratios, and (2) enabling multi-granularity image generation.\nBesides the competitive results on public benchmarks, SEED-X demonstrates its effectiveness in handling real-world applications across various domains. We hope that our work will inspire future research into what can be achieved by versatile multimodal foundation models in real-world applications. All models, training, and inference codes are available at https://anonymous.4open.science/r/SEED-X/.", "keywords": [ "Multimodal LLM", "Comprehension and Generation" ], "rating_list": [ 5, 6, 5, 5, 3 ], "soundness_list": [ 2, 3, 3, 3, 2 ], "presentation_list": [ 2, 3, 3, 3, 2 ], "contribution_list": [ 3, 2, 3, 2, 2 ], "confidence_list": [ 3, 4, 5, 4, 5 ] }, { "paper_id": "Bwhd7GUyHH", "primary_area": "reinforcement learning", "n_author": 4, "author_info_list": [ { "name": "Ahmed Shoyeb Raihan", "gender": "Male", "institution": "West Virginia University", "country": "", "position": "PhD student" }, { "name": "Hamed Khosravi", "gender": "Male", "institution": "West Virginia University", "country": "US", "position": "PhD student" }, { "name": "Imtiaz Ahmed", "gender": "unknown", "institution": "West Virginia University", "country": "US", "position": "Assistant Professor" }, { "name": "Srinjoy Das", "gender": "Male", "institution": "West Virginia University", "country": "US", "position": "Assistant Professor" } ], "n_formula": 161, "n_formula_1": 24, "n_ref_uni": 51, "n_ref": 91, "n_ref_all": 123, "n_fig": 7, "n_tab": 5, "L_tab": 1516, "n_element_tab": 288, "n_fig_1": 1, "n_tab_1": 3, "L_tab_1": 958, "n_element_tab_1": 42, "formula_len_all": 9869, "formula_len_all_1": 1725, "len_all": 224813, "len_all_1": 66936, "len_abs": 1915, "len_title": 121, "len_sents": 61696, "len_sents_1": 28689, "n_sents": 491, "n_sents_1": 198, "venue": "ICLR2025", "decision": "withdraw", "L_title": 73, "L_abs": 1927, "title": "LNUCB-TA: Linear-nonlinear Hybrid Bandit Learning with Temporal Attention", "abs": "Existing contextual multi-armed bandit (MAB) algorithms struggle to simultaneously capture long-term trends as well as local patterns across all arms, leading to suboptimal performance in complex environments with rapidly changing reward structures. Additionally, they typically employ static exploration rates, which do not adapt to dynamic conditions. To address these issues, we present LNUCB-TA, a hybrid bandit model that introduces a novel nonlinear component (adaptive $k$-Nearest Neighbors ($k$-NN)) designed to reduce time complexity, and an innovative global-and-local attention-based exploration mechanism. Our method incorporates a unique synthesis of linear and nonlinear estimation techniques, where the nonlinear component dynamically adjusts $k$ based on reward variance, thereby effectively capturing spatiotemporal patterns in the data. This is critical for reducing the likelihood of selecting suboptimal arms and accurately estimating rewards while reducing computational time. Also, our proposed attention-based mechanism prioritizes arms based on their historical performance and frequency of selection, thereby balancing exploration and exploitation in real-time without the need for fine-tuning exploration parameters. Incorporating both global attention (based on overall performance across all arms) and local attention (focusing on individual arm performance), the algorithm efficiently adapts to temporal and spatial complexities in the available context. Empirical evaluation demonstrates that LNUCB-TA significantly outperforms state-of-the-art contextual MAB algorithms, including purely linear, nonlinear, and vanilla combination of linear and nonlinear bandits based on cumulative and mean rewards, convergence performance, and demonstrates consistency of results across different exploration rates. Theoretical analysis further proves the robustness of LNUCB-TA with a sub-linear regret bound.", "keywords": [ "Contextual Multi-Armed Bandit", "Exploration-Exploitation Trade-off", "Adaptive k-Nearest Neighbors (k-NN)", "Attention-Based Exploration Rate", "Sub-linear Regret" ], "rating_list": [ 3, 3, 6, 5 ], "soundness_list": [ 2, 2, 3, 3 ], "presentation_list": [ 2, 1, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 3, 3, 4, 4 ] }, { "paper_id": "BwR8t91yqh", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 6, "author_info_list": [ { "name": "Chi Wang", "gender": "Male", "institution": "Google DeepMind", "country": "US", "position": "Senior Staff Research Scientist" }, { "name": "JAGANNATH SHASHANK SUBRAMANYA SAI VADREVU", "gender": "Male", "institution": "Microsoft", "country": "CA", "position": "Researcher" }, { "name": "Mengting Wan", "gender": "Female", "institution": "Microsoft", "country": "US", "position": "Principal Researcher" }, { "name": "Ryan Nadel", "gender": "unknown", "institution": "University of British Columbia", "country": "CA", "position": "MS student" }, { "name": "Wenyue Hua", "gender": "Female", "institution": "University of California, Santa Barbara", "country": "US", "position": "Postdoc" }, { "name": "Yongfeng Zhang", "gender": "Male", "institution": "Rutgers University", "country": "US", "position": "Associate Professor" } ], "n_formula": 15, "n_formula_1": 3, "n_ref_uni": 50, "n_ref": 93, "n_ref_all": 110, "n_fig": 8, "n_tab": 4, "L_tab": 3738, "n_element_tab": 271, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 1505, "n_element_tab_1": 199, "formula_len_all": 880, "formula_len_all_1": 160, "len_all": 189130, "len_all_1": 68575, "len_abs": 1278, "len_title": 153, "len_sents": 59741, "len_sents_1": 34347, "n_sents": 364, "n_sents_1": 195, "venue": "ICLR2025", "decision": "Accept (Poster)", "L_title": 105, "L_abs": 1268, "title": "Interactive Speculative Planning: Enhance Agent Efficiency through Co-design of System and User Interface", "abs": "Agents, as user-centric tools, are increasingly deployed for human task delegation, assisting with a broad spectrum of requests by generating thoughts, engaging with user proxies, and producing action plans. However, agents based on large language models often face substantial planning latency due to two primary factors: the efficiency limitations of the underlying LLMs due to their large size and high demand, and the structural complexity of the agents due to the extensive generation of intermediate steps to produce the final output. Given that inefficiency in service provision can undermine the value of automation for users, this paper presents a human-centered efficient agent planning method \u2013 Interactive Speculative Planning \u2013 aiming at enhancing the efficiency of agent planning through both system design and user interaction. Our approach advocates for the co-design of the agent system and user interface, underscoring the importance of an agent system that can fluidly manage user interactions and interruptions. By integrating human interruptions as a fundamental component of the system, we not only make it more user-centric but also expedite the entire process by leveraging human-in-the-loop interactions to provide accurate intermediate steps.", "keywords": [ "large language model", "agent", "efficiency", "human-computer interaction" ], "rating_list": [ 6, 6, 8 ], "soundness_list": [ 3, 3, 3 ], "presentation_list": [ 2, 2, 3 ], "contribution_list": [ 2, 2, 3 ], "confidence_list": [ 4, 3, 4 ] }, { "paper_id": "BwQUo5RVun", "primary_area": "unsupervised, self-supervised, semi-supervised, and supervised representation learning", "n_author": 2, "author_info_list": [ { "name": "Yun Fu", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "Full Professor" }, { "name": "Zhi Xu", "gender": "Male", "institution": "Northeastern University", "country": "US", "position": "PhD student" } ], "n_formula": 8, "n_formula_1": 4, "n_ref_uni": 26, "n_ref": 51, "n_ref_all": 57, "n_fig": 1, "n_tab": 6, "L_tab": 3158, "n_element_tab": 370, "n_fig_1": 1, "n_tab_1": 5, "L_tab_1": 3161, "n_element_tab_1": 355, "formula_len_all": 435, "formula_len_all_1": 294, "len_all": 128264, "len_all_1": 67358, "len_abs": 1203, "len_title": 120, "len_sents": 34120, "len_sents_1": 28309, "n_sents": 274, "n_sents_1": 225, "venue": "ICLR2025", "decision": "Reject", "L_title": 72, "L_abs": 1210, "title": "improve weakly supervised visual grounding by learning where to focus on", "abs": "Visual grounding is a crucial task for connecting visual and language descriptions by identifying target objects based on language entities. However, fully supervised methods require extensive annotations, which can be challenging and time-consuming to obtain. Weakly supervised visual grounding, which only relies on image-sentence association without object-level annotations, offers a promising solution. Previous approaches have mainly focused on finding the relationship between detected candidates, without considering improving object localization. In this work, we propose a novel method that leverages Grad-CAM to help the model identify precise objects. Specifically, we introduce a CAM encoder that exploits Grad-CAM information and a new loss function, attention mining loss, to guide the Grad-CAM feature to focus on the entire object. We also use an architecture which combines CNN and transformer, and a multi-modality fusion module to aggregate visual features, language features and CAM features. Our proposed approach achieves state-of-the-art results on several datasets, demonstrating its effectiveness in different scenes. Ablation studies further confirm the benefits of our architecture.", "keywords": [ "weakly supervised learning", "visual grounding", "grad-cam", "vision and language" ], "rating_list": [ 3, 3, 3, 3 ], "soundness_list": [ 1, 2, 3, 1 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 1, 2, 2, 1 ], "confidence_list": [ 5, 4, 4, 5 ] }, { "paper_id": "BwGeIhGPgn", "primary_area": "datasets and benchmarks", "n_author": 3, "author_info_list": [ { "name": "Been Kim", "gender": "unknown", "institution": "Google DeepMind", "country": "", "position": "Research Scientist" }, { "name": "Zi Wang", "gender": "Female", "institution": "Google DeepMind", "country": "US", "position": "Research scientist" }, { "name": "Belinda Zou Li", "gender": "Female", "institution": "Massachusetts Institute of Technology", "country": "", "position": "PhD student" } ], "n_formula": 35, "n_formula_1": 10, "n_ref_uni": 43, "n_ref": 81, "n_ref_all": 106, "n_fig": 3, "n_tab": 5, "L_tab": 0, "n_element_tab": 0, "n_fig_1": 2, "n_tab_1": 1, "L_tab_1": 0, "n_element_tab_1": 0, "formula_len_all": 2569, "formula_len_all_1": 610, "len_all": 193488, "len_all_1": 65777, "len_abs": 1884, "len_title": 134, "len_sents": 61619, "len_sents_1": 30274, "n_sents": 543, "n_sents_1": 258, "venue": "ICLR2025", "decision": "Reject", "L_title": 83, "L_abs": 1888, "title": "Evaluating Information Gathering Abilities of Large Language Models with QuestBench", "abs": "Large language models (LLMs) have mastered a wide range of reasoning tasks, with an underlying assumption that the tasks are well-specified for LLMs to reach solutions.\nIn reality, queries and instructions to LLMs often contain incomplete or underspecified information.\nTherefore, LLMs need to be able to actively ac-quire missing information by asking clarifying questions, ideally seeking the minimally sufficient piece of information.\nTo assess whether LLMs possess this ability, we construct QUESTBENCH, a set of underspecified reasoning tasks that can be solved by asking at most a single question.\nWe frame the tasks as constraint satisfaction problems with missing variable assignments, where the exact model response cannot be determined unless certain variables\u2019 values are acquired.\nThis framework specifically targets tasks where uncertainty stems from missing information, rather than semantic ambiguity in language. QUESTBENCH includes (1) Logic-Q: Logical reasoning tasks where one proposition is missing, (2) Planning-Q: PDDL planning problems where the initial state is partially observed, and (3) GSM-Q: Grade school math problems where one variable assignment is missing.\nEach task presents multiple choices of possible questions, only one of which is correct.\nWe evaluate Gemini and GPT-4o models and find that they achieve 20 \u2013 30% accuracy in both zero-shot and few-shot settings. When evaluating GPT-4-o1on a subset of our data, we find that it is only 41 \u2013 44% accurate, despite using state-of-the-art inference-time reasoning techniques.\nWhen investigating characteristics of QuestBench, we find that LLMs struggle with tasks that are computationally expensive for traditional search-based CSP solvers. Our analyses reveal a negative correlation between LLM accuracy and solver runtime complexity, suggesting that LLMs may share similar limitations to CSP solvers", "keywords": [ "information gathering", "question asking", "language model", "evaluation", "benchmarks" ], "rating_list": [ 3, 5, 5, 3 ], "soundness_list": [ 3, 2, 2, 2 ], "presentation_list": [ 2, 2, 2, 2 ], "contribution_list": [ 2, 2, 2, 2 ], "confidence_list": [ 3, 3, 3, 4 ] }, { "paper_id": "Bvqsas4TYX", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 6, "author_info_list": [ { "name": "Jimin Huang", "gender": "Male", "institution": "The Fin AI", "country": "SG", "position": "Principal Researcher" }, { "name": "Kailai Yang", "gender": "Male", "institution": "University of Manchester", "country": "", "position": "PhD student" }, { "name": "Qianqian Xie", "gender": "Female", "institution": "Yale University", "country": "US", "position": "Researcher" }, { "name": "Sophia Ananiadou", "gender": "Female", "institution": "University of Manchester", "country": "GB", "position": "Full Professor" }, { "name": "Zhiwei Liu", "gender": "Male", "institution": "University of Manchester", "country": "", "position": "PhD student" }, { "name": "Erxue Min", "gender": "Male", "institution": "Baidu", "country": "CN", "position": "Researcher" } ], "n_formula": 41, "n_formula_1": 15, "n_ref_uni": 36, "n_ref": 128, "n_ref_all": 144, "n_fig": 6, "n_tab": 4, "L_tab": 528, "n_element_tab": 44, "n_fig_1": 5, "n_tab_1": 2, "L_tab_1": 1533, "n_element_tab_1": 151, "formula_len_all": 2667, "formula_len_all_1": 925, "len_all": 185808, "len_all_1": 66334, "len_abs": 1582, "len_title": 124, "len_sents": 62618, "len_sents_1": 29346, "n_sents": 483, "n_sents_1": 233, "venue": "ICLR2025", "decision": "withdraw", "L_title": 76, "L_abs": 1562, "title": "Selective Preference Optimization via Token-Level Reward Function Estimation", "abs": "Recent advancements in large language model alignment leverage token-level supervisions to perform fine-grained preference optimization. However, existing token-level alignment methods either optimize on all available tokens, which can be noisy and inefficient, or perform selective training with complex and expensive key token selection strategies.\nIn this work, we propose Selective Preference Optimization (SePO), a novel selective alignment strategy that centers on efficient key token selection without requiring strong, fine-grained supervision signals. We theoretically prove the feasibility of Direct Preference Optimization (DPO) as token-level reward function estimators, which applies to any existing alignment datasets and enables cost-efficient token selection with small-scale model sizes and training data. We then train an oracle model with DPO on the target data and utilize the estimated reward function to score all tokens within the target dataset, where only the key tokens are selected to supervise the target policy model with a contrastive objective function. Extensive experiments on three public evaluation benchmarks show that SePO significantly outperforms competitive baseline methods by only optimizing on 30\\% key tokens. We also explore SePO as a new paradigm for weak-to-strong generalization, showing\nthat weak oracle models effectively supervise strong policy models with up to 16.8$\\times$ more parameters. SePO also selects useful supervision signals from out-of-distribution data, alleviating the over-optimization problem.", "keywords": [ "large language models", "preference optimization", "alignment" ], "rating_list": [ 3, 3, 5, 3 ], "soundness_list": [ 2, 2, 3, 2 ], "presentation_list": [ 3, 2, 3, 3 ], "contribution_list": [ 2, 2, 3, 2 ], "confidence_list": [ 4, 4, 3, 4 ] }, { "paper_id": "BvlaNTMl7P", "primary_area": "alignment, fairness, safety, privacy, and societal considerations", "n_author": 4, "author_info_list": [ { "name": "Garrett Gagnon", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "", "position": "PhD student" }, { "name": "Liu Liu", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "Assistant Professor" }, { "name": "Swagath Venkataramani", "gender": "unknown", "institution": "International Business Machines", "country": "", "position": "Researcher" }, { "name": "Zhenyu Liu", "gender": "Male", "institution": "Rensselaer Polytechnic Institute", "country": "US", "position": "PhD student" } ], "n_formula": 29, "n_formula_1": 19, "n_ref_uni": 51, "n_ref": 112, "n_ref_all": 127, "n_fig": 5, "n_tab": 11, "L_tab": 3814, "n_element_tab": 356, "n_fig_1": 3, "n_tab_1": 5, "L_tab_1": 2459, "n_element_tab_1": 217, "formula_len_all": 2159, "formula_len_all_1": 1247, "len_all": 155165, "len_all_1": 65479, "len_abs": 1085, "len_title": 136, "len_sents": 43218, "len_sents_1": 27473, "n_sents": 334, "n_sents_1": 203, "venue": "ICLR2025", "decision": "withdraw", "L_title": 87, "L_abs": 1088, "title": "SINAI: Selective Injection of Noise for Adversarial Robustness with Improved Efficiency", "abs": "Deep Neural Networks (DNNs) have revolutionized a wide range of industries, from healthcare and finance to automotive, by offering unparalleled capabilities in data analysis and decision-making. Despite their transforming impact, DNNs face two critical challenges: the vulnerability to adversarial attacks and the increasing computational costs associated with more complex and larger models. In this paper, we introduce an effective method designed to simultaneously enhance adversarial robustness and execution efficiency. Unlike prior studies that enhance robustness via uniformly injecting noise, we introduce a non-uniform noise injection algorithm, strategically applied at each DNN layer to disrupt adversarial perturbations introduced in attacks. By employing approximation techniques, our approach identifies and protects essential neurons while strategically introducing noise into non-essential neurons. Our experimental results demonstrate that our method successfully enhances both robustness and efficiency across several attack scenarios, model architectures, and datasets.", "keywords": [ "Adversarial Robustness", "Efficient Neural Networks", "Hardware and Software Co-design" ], "rating_list": [ 3, 5, 5, 6 ], "soundness_list": [ 3, 3, 3, 3 ], "presentation_list": [ 1, 3, 2, 3 ], "contribution_list": [ 2, 2, 2, 3 ], "confidence_list": [ 4, 3, 3, 4 ] }, { "paper_id": "BvQkjCnXXr", "primary_area": "infrastructure, software libraries, hardware, systems, etc.", "n_author": 5, "author_info_list": [ { "name": "Bo Xu", "gender": "Male", "institution": "Donghua University, Shanghai", "country": "CN", "position": "Associate Professor" }, { "name": "Hongya Wang", "gender": "Male", "institution": "Donghua University, Shanghai", "country": "CN", "position": "Full Professor" }, { "name": "Ming Du", "gender": "unknown", "institution": "", "country": "", "position": "" }, { "name": "Minjie Luo", "gender": "unknown", "institution": "Donghua University, Shanghai", "country": "", "position": "Associate Professor" }, { "name": "Zongyuan Tan", "gender": "Male", "institution": "Donghua University, Shanghai", "country": "", "position": "PhD student" } ], "n_formula": 78, "n_formula_1": 19, "n_ref_uni": 34, "n_ref": 63, "n_ref_all": 98, "n_fig": 7, "n_tab": 12, "L_tab": 5728, "n_element_tab": 723, "n_fig_1": 3, "n_tab_1": 3, "L_tab_1": 2273, "n_element_tab_1": 47, "formula_len_all": 8999, "formula_len_all_1": 1588, "len_all": 208639, "len_all_1": 60552, "len_abs": 1357, "len_title": 123, "len_sents": 63527, "len_sents_1": 25538, "n_sents": 580, "n_sents_1": 194, "venue": "ICLR2025", "decision": "Reject", "L_title": 74, "L_abs": 1394, "title": "Simple Yet Efficient Locality Sensitive Hashing with Theoretical Guarantee", "abs": "Locality-sensitive hashing (LSH) is an effective randomized technique widely used in many machine learning tasks such as outlier detection, neural network training and nearest neighbor search. The cost of hashing is the main performance bottleneck of these applications because the index construction functionality, a core component dominating the end-to-end latency, involves the evaluation of a large number of hash functions. Surprisingly, however, little work has been done to improve the efficiency of LSH computation. In this paper, we design a simple yet efficient LSH scheme, named FastLSH, by combining random sampling and random projection. FastLSH reduces the hashing complexity from $O(n)$ to $O(m)$ ($m